diff --git a/.gitattributes b/.gitattributes index 2ab943c2b330763602d4a531e40c481d48ceb667..f4b031806a3ddbdbb92625f2b882a68bf72bcdb1 100644 --- a/.gitattributes +++ b/.gitattributes @@ -3833,3 +3833,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2023/Self-supervised[[:space:]]Learning[[:space:]]to[[:space:]]Bring[[:space:]]Dual[[:space:]]Reversed[[:space:]]Rolling[[:space:]]Shutter[[:space:]]Images[[:space:]]Alive/0f7c9f43-96dc-4fd0-a054-e650a421a3a0_origin.pdf filter=lfs diff=lfs merge=lfs -text 2023/Self-supervised[[:space:]]Monocular[[:space:]]Depth[[:space:]]Estimation_[[:space:]]Let's[[:space:]]Talk[[:space:]]About[[:space:]]The[[:space:]]Weather/7432c306-22bf-4ee4-b730-8425345482aa_origin.pdf filter=lfs diff=lfs merge=lfs -text 2023/Self-supervised[[:space:]]Monocular[[:space:]]Underwater[[:space:]]Depth[[:space:]]Recovery,[[:space:]]Image[[:space:]]Restoration,[[:space:]]and[[:space:]]a[[:space:]]Real-sea[[:space:]]Video[[:space:]]Dataset/e11d4111-32f8-407a-a327-9e6fbe49fac5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Self-supervised[[:space:]]Pre-training[[:space:]]for[[:space:]]Mirror[[:space:]]Detection/9faec277-19c4-4583-a3c8-84bf53084f2f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SemARFlow_[[:space:]]Injecting[[:space:]]Semantics[[:space:]]into[[:space:]]Unsupervised[[:space:]]Optical[[:space:]]Flow[[:space:]]Estimation[[:space:]]for[[:space:]]Autonomous[[:space:]]Driving/fa7524e1-a76d-4844-9361-43c8b8046bf9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Semantic[[:space:]]Attention[[:space:]]Flow[[:space:]]Fields[[:space:]]for[[:space:]]Monocular[[:space:]]Dynamic[[:space:]]Scene[[:space:]]Decomposition/68ea1ef8-9cb5-44da-894d-92e02632af66_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Semantic[[:space:]]Information[[:space:]]in[[:space:]]Contrastive[[:space:]]Learning/c3144656-62d8-4975-97d6-ef4630f410df_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Semantic-Aware[[:space:]]Dynamic[[:space:]]Parameter[[:space:]]for[[:space:]]Video[[:space:]]Inpainting[[:space:]]Transformer/706f36ef-c7c5-4ea7-8481-e3a3bb79a94c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Semantic-Aware[[:space:]]Implicit[[:space:]]Template[[:space:]]Learning[[:space:]]via[[:space:]]Part[[:space:]]Deformation[[:space:]]Consistency/840dd92d-8817-48f0-ac2c-586a3d27151d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Semantically[[:space:]]Structured[[:space:]]Image[[:space:]]Compression[[:space:]]via[[:space:]]Irregular[[:space:]]Group-Based[[:space:]]Decoupling/e3cccf23-d23a-4b30-90e1-8c04059b69dd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Semantics[[:space:]]Meets[[:space:]]Temporal[[:space:]]Correspondence_[[:space:]]Self-supervised[[:space:]]Object-centric[[:space:]]Learning[[:space:]]in[[:space:]]Videos/0da0877f-f78d-4f97-9a49-55c50b2292de_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Semantics-Consistent[[:space:]]Feature[[:space:]]Search[[:space:]]for[[:space:]]Self-Supervised[[:space:]]Visual[[:space:]]Representation[[:space:]]Learning/fd85c784-c77b-49cb-91b7-f3a6f94e50c3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Semantify_[[:space:]]Simplifying[[:space:]]the[[:space:]]Control[[:space:]]of[[:space:]]3D[[:space:]]Morphable[[:space:]]Models[[:space:]]Using[[:space:]]CLIP/ff2d3341-5fbc-47bb-9ce8-0e16c61108bc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Semi-Supervised[[:space:]]Learning[[:space:]]via[[:space:]]Weight-Aware[[:space:]]Distillation[[:space:]]under[[:space:]]Class[[:space:]]Distribution[[:space:]]Mismatch/ab302b84-1623-4a04-b35b-84d3ca9ad536_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Semi-Supervised[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]under[[:space:]]Label[[:space:]]Noise[[:space:]]via[[:space:]]Diverse[[:space:]]Learning[[:space:]]Groups/8e5c15a5-f270-46c8-829c-81448a8fde6f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Semi-supervised[[:space:]]Semantics-guided[[:space:]]Adversarial[[:space:]]Training[[:space:]]for[[:space:]]Robust[[:space:]]Trajectory[[:space:]]Prediction/487dfb22-6094-4f31-976c-310213635421_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Semi-supervised[[:space:]]Speech-driven[[:space:]]3D[[:space:]]Facial[[:space:]]Animation[[:space:]]via[[:space:]]Cross-modal[[:space:]]Encoding/ebc67249-8be2-4ca5-975d-54a569184984_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Sensitivity-Aware[[:space:]]Visual[[:space:]]Parameter-Efficient[[:space:]]Fine-Tuning/219bbdca-cfe8-433f-a6db-97c4773a1a10_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Sentence[[:space:]]Attention[[:space:]]Blocks[[:space:]]for[[:space:]]Answer[[:space:]]Grounding/0622d151-144e-4e04-b267-1cccf0dba507_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Sequential[[:space:]]Texts[[:space:]]Driven[[:space:]]Cohesive[[:space:]]Motions[[:space:]]Synthesis[[:space:]]with[[:space:]]Natural[[:space:]]Transitions/2c5b9c34-9105-42fb-b593-c16dc3db7c25_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Set-level[[:space:]]Guidance[[:space:]]Attack_[[:space:]]Boosting[[:space:]]Adversarial[[:space:]]Transferability[[:space:]]of[[:space:]]Vision-Language[[:space:]]Pre-training[[:space:]]Models/4d2215bf-24a1-45b3-bcea-24623e459934_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Shape[[:space:]]Analysis[[:space:]]of[[:space:]]Euclidean[[:space:]]Curves[[:space:]]under[[:space:]]Frenet-Serret[[:space:]]Framework/235fdcf4-691b-437b-8ee8-c7b76395f3ab_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Shape[[:space:]]Anchor[[:space:]]Guided[[:space:]]Holistic[[:space:]]Indoor[[:space:]]Scene[[:space:]]Understanding/4f878220-d6cc-48e2-9d1e-82c1cf350095_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/ShapeScaffolder_[[:space:]]Structure-Aware[[:space:]]3D[[:space:]]Shape[[:space:]]Generation[[:space:]]from[[:space:]]Text/3212f743-a520-438c-b4e2-ff41229e0a86_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Shatter[[:space:]]and[[:space:]]Gather_[[:space:]]Learning[[:space:]]Referring[[:space:]]Image[[:space:]]Segmentation[[:space:]]with[[:space:]]Text[[:space:]]Supervision/eba869e5-3a94-47e8-afcc-2f1d3dc46bfb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Shift[[:space:]]from[[:space:]]Texture-bias[[:space:]]to[[:space:]]Shape-bias_[[:space:]]Edge[[:space:]]Deformation-based[[:space:]]Augmentation[[:space:]]for[[:space:]]Robust[[:space:]]Object[[:space:]]Recognition/e071fbaf-3036-4ac7-9bfd-fb6983ccf03f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/ShiftNAS_[[:space:]]Improving[[:space:]]One-shot[[:space:]]NAS[[:space:]]via[[:space:]]Probability[[:space:]]Shift/9d0cd6fb-8343-4e85-b19e-ff450eb2e5e4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Shortcut-V2V_[[:space:]]Compression[[:space:]]Framework[[:space:]]for[[:space:]]Video-to-Video[[:space:]]Translation[[:space:]]Based[[:space:]]on[[:space:]]Temporal[[:space:]]Redundancy[[:space:]]Reduction/99da4ac0-6f5c-433f-bc7b-caa55aa672d5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Shrinking[[:space:]]Class[[:space:]]Space[[:space:]]for[[:space:]]Enhanced[[:space:]]Certainty[[:space:]]in[[:space:]]Semi-Supervised[[:space:]]Learning/34917a9e-765e-47a5-bd3d-8044ab9de911_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SiLK_[[:space:]]Simple[[:space:]]Learned[[:space:]]Keypoints/545a50e4-a6f6-42d1-849e-c613959c21fd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Sigmoid[[:space:]]Loss[[:space:]]for[[:space:]]Language[[:space:]]Image[[:space:]]Pre-Training/85b123ac-a2cf-49ce-9282-93b24961a80b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Sign[[:space:]]Language[[:space:]]Translation[[:space:]]with[[:space:]]Iterative[[:space:]]Prototype/f2c1bf6d-f55b-40d7-9372-42bff493c6d3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SimFIR_[[:space:]]A[[:space:]]Simple[[:space:]]Framework[[:space:]]for[[:space:]]Fisheye[[:space:]]Image[[:space:]]Rectification[[:space:]]with[[:space:]]Self-supervised[[:space:]]Representation[[:space:]]Learning/e93fe221-1f98-47dd-bbe6-409ea24f23b2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SimMatchV2_[[:space:]]Semi-Supervised[[:space:]]Learning[[:space:]]with[[:space:]]Graph[[:space:]]Consistency/4de24035-8115-4285-8e43-089c94dd354f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SimNP_[[:space:]]Learning[[:space:]]Self-Similarity[[:space:]]Priors[[:space:]]Between[[:space:]]Neural[[:space:]]Points/54ca6d2a-b8a1-4a4e-a6f5-57eb0fddb4dd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Similarity[[:space:]]Min-Max_[[:space:]]Zero-Shot[[:space:]]Day-Night[[:space:]]Domain[[:space:]]Adaptation/d2d1b76d-0c03-43ab-96b8-c3ce7c667abf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Simoun_[[:space:]]Synergizing[[:space:]]Interactive[[:space:]]Motion-appearance[[:space:]]Understanding[[:space:]]for[[:space:]]Vision-based[[:space:]]Reinforcement[[:space:]]Learning/17dc049b-d1dc-4e17-921e-1317f662d966_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Simple[[:space:]]Baselines[[:space:]]for[[:space:]]Interactive[[:space:]]Video[[:space:]]Retrieval[[:space:]]with[[:space:]]Questions[[:space:]]and[[:space:]]Answers/149533a1-79f1-428b-885a-6c528f5bd0a8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Simple[[:space:]]and[[:space:]]Effective[[:space:]]Out-of-Distribution[[:space:]]Detection[[:space:]]via[[:space:]]Cosine-based[[:space:]]Softmax[[:space:]]Loss/081c8308-e453-453c-8d9f-1e2b9b04b642_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SimpleClick_[[:space:]]Interactive[[:space:]]Image[[:space:]]Segmentation[[:space:]]with[[:space:]]Simple[[:space:]]Vision[[:space:]]Transformers/2b87aeaa-0b31-4e94-84d5-d9bfaea7e328_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Simulating[[:space:]]Fluids[[:space:]]in[[:space:]]Real-World[[:space:]]Still[[:space:]]Images/d51a02b7-7fc4-4f85-a456-fd173559877f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Single[[:space:]]Depth-image[[:space:]]3D[[:space:]]Reflection[[:space:]]Symmetry[[:space:]]and[[:space:]]Shape[[:space:]]Prediction/fcff96a5-45a1-4f7e-ab9f-0c309af710cd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Single[[:space:]]Image[[:space:]]Deblurring[[:space:]]with[[:space:]]Row-dependent[[:space:]]Blur[[:space:]]Magnitude/05f13a7b-32f2-44f9-b585-5bd4c1beae6a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Single[[:space:]]Image[[:space:]]Defocus[[:space:]]Deblurring[[:space:]]via[[:space:]]Implicit[[:space:]]Neural[[:space:]]Inverse[[:space:]]Kernels/bee3d863-7fc6-4e0b-a339-87a4afbe8bb4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Single[[:space:]]Image[[:space:]]Reflection[[:space:]]Separation[[:space:]]via[[:space:]]Component[[:space:]]Synergy/6169238f-a2a5-4863-b901-a9125cd1dec8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Single-Stage[[:space:]]Diffusion[[:space:]]NeRF_[[:space:]]A[[:space:]]Unified[[:space:]]Approach[[:space:]]to[[:space:]]3D[[:space:]]Generation[[:space:]]and[[:space:]]Reconstruction/06f68dbb-5c23-4704-ae55-c667675f6d9f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Size[[:space:]]Does[[:space:]]Matter_[[:space:]]Size-aware[[:space:]]Virtual[[:space:]]Try-on[[:space:]]via[[:space:]]Clothing-oriented[[:space:]]Transformation[[:space:]]Try-on[[:space:]]Network/2e5e4941-f171-4e8c-be33-35affe248d0e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SkeleTR_[[:space:]]Towards[[:space:]]Skeleton-based[[:space:]]Action[[:space:]]Recognition[[:space:]]in[[:space:]]the[[:space:]]Wild/27665a84-18aa-44bc-b25b-e6429d7734c6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SkeletonMAE_[[:space:]]Graph-based[[:space:]]Masked[[:space:]]Autoencoder[[:space:]]for[[:space:]]Skeleton[[:space:]]Sequence[[:space:]]Pre-training/dcc19c04-02ae-4f8d-8f44-a86cdd91e8a4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Sketch[[:space:]]and[[:space:]]Text[[:space:]]Guided[[:space:]]Diffusion[[:space:]]Model[[:space:]]for[[:space:]]Colored[[:space:]]Point[[:space:]]Cloud[[:space:]]Generation/6a47e5bb-0a8f-4500-ae70-d8f17ed82ef9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Skill[[:space:]]Transformer_[[:space:]]A[[:space:]]Monolithic[[:space:]]Policy[[:space:]]for[[:space:]]Mobile[[:space:]]Manipulation/e273608a-9944-4287-91dc-881cf0220c1e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Skip-Plan_[[:space:]]Procedure[[:space:]]Planning[[:space:]]in[[:space:]]Instructional[[:space:]]Videos[[:space:]]via[[:space:]]Condensed[[:space:]]Action[[:space:]]Space[[:space:]]Learning/c1a3a0ec-3885-4f2f-93dd-62793c50f314_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SlaBins_[[:space:]]Fisheye[[:space:]]Depth[[:space:]]Estimation[[:space:]]using[[:space:]]Slanted[[:space:]]Bins[[:space:]]on[[:space:]]Road[[:space:]]Environments/2f72858d-bf2e-45dd-b589-9c5d6578e810_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Small[[:space:]]Object[[:space:]]Detection[[:space:]]via[[:space:]]Coarse-to-fine[[:space:]]Proposal[[:space:]]Generation[[:space:]]and[[:space:]]Imitation[[:space:]]Learning/231f8e4e-8ec6-40cc-a455-514ae44e6fa1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Smoothness[[:space:]]Similarity[[:space:]]Regularization[[:space:]]for[[:space:]]Few-Shot[[:space:]]GAN[[:space:]]Adaptation/e521c479-803a-4c4d-bf48-e946e8872ca9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Snow[[:space:]]Removal[[:space:]]in[[:space:]]Video_[[:space:]]A[[:space:]]New[[:space:]]Dataset[[:space:]]and[[:space:]]A[[:space:]]Novel[[:space:]]Method/5ae50b63-d59f-4dc4-bf7f-4d6ac34210b4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SoDaCam_[[:space:]]Software-defined[[:space:]]Cameras[[:space:]]via[[:space:]]Single-Photon[[:space:]]Imaging/ce1d43a9-f5c5-4d36-93a7-015029e926c0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Social[[:space:]]Diffusion_[[:space:]]Long-term[[:space:]]Multiple[[:space:]]Human[[:space:]]Motion[[:space:]]Anticipation/e840bc61-69a6-4f9d-9e0d-f25143903533_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Sound[[:space:]]Localization[[:space:]]from[[:space:]]Motion_[[:space:]]Jointly[[:space:]]Learning[[:space:]]Sound[[:space:]]Direction[[:space:]]and[[:space:]]Camera[[:space:]]Rotation/ebfa790a-8a91-4c3c-9e52-be53fe87e14f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Sound[[:space:]]Source[[:space:]]Localization[[:space:]]is[[:space:]]All[[:space:]]about[[:space:]]Cross-Modal[[:space:]]Alignment/7b57a83c-7e70-4280-a6ee-f0f550384305_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Source-free[[:space:]]Depth[[:space:]]for[[:space:]]Object[[:space:]]Pop-out/1ac228ed-431b-40e3-b444-e5a574a99abb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Source-free[[:space:]]Domain[[:space:]]Adaptive[[:space:]]Human[[:space:]]Pose[[:space:]]Estimation/6db7e049-3b2d-4942-adce-fb8455dd18b2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Space[[:space:]]Engage_[[:space:]]Collaborative[[:space:]]Space[[:space:]]Supervision[[:space:]]for[[:space:]]Contrastive-Based[[:space:]]Semi-Supervised[[:space:]]Semantic[[:space:]]Segmentation/5953e602-104c-4dc8-b226-b90796937563_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Space-time[[:space:]]Prompting[[:space:]]for[[:space:]]Video[[:space:]]Class-incremental[[:space:]]Learning/23f1a364-1b5a-487e-9a38-2d1ef0112aa7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SpaceEvo_[[:space:]]Hardware-Friendly[[:space:]]Search[[:space:]]Space[[:space:]]Design[[:space:]]for[[:space:]]Efficient[[:space:]]INT8[[:space:]]Inference/7d8d9b36-b64d-45d0-8aa5-d17e8abbee45_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Spacetime[[:space:]]Surface[[:space:]]Regularization[[:space:]]for[[:space:]]Neural[[:space:]]Dynamic[[:space:]]Scene[[:space:]]Reconstruction/11a08db4-fb69-460a-a705-16e05f3f48ac_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Sparse[[:space:]]Instance[[:space:]]Conditioned[[:space:]]Multimodal[[:space:]]Trajectory[[:space:]]Prediction/5a87e2e4-dac0-41f3-9966-85985369c6c6_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2023/Self-supervised Pre-training for Mirror Detection/9faec277-19c4-4583-a3c8-84bf53084f2f_content_list.json b/2023/Self-supervised Pre-training for Mirror Detection/9faec277-19c4-4583-a3c8-84bf53084f2f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..869d910ceeca6ce0619b3bd788c443414b8a121a --- /dev/null +++ b/2023/Self-supervised Pre-training for Mirror Detection/9faec277-19c4-4583-a3c8-84bf53084f2f_content_list.json @@ -0,0 +1,1586 @@ +[ + { + "type": "text", + "text": "Self-supervised Pre-training for Mirror Detection", + "text_level": 1, + "bbox": [ + 233, + 130, + 736, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiaying Lin* Rynson W.H. Lau* City University of Hong Kong", + "bbox": [ + 341, + 181, + 625, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jiayinlin5, rynson.lau}@cityu.edu.hk", + "bbox": [ + 323, + 219, + 647, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing mirror detection methods require supervised ImageNet pre-training to obtain good general-purpose image features. However, supervised ImageNet pre-training focuses on category-level discrimination and may not be suitable for downstream tasks like mirror detection, due to the overfitting upstream tasks (e.g., supervised image classification). We observe that mirror reflection is crucial to how people perceive the presence of mirrors, and such mid-level features can be better transferred from self-supervised pretrained models. Inspired by this observation, in this paper we aim to improve mirror detection methods by proposing a new self-supervised learning (SSL) pre-training framework for modeling the representation of mirror reflection progressively in the pre-training process. Our framework consists of three pre-training stages at different levels: 1) an image-level pre-training stage to globally incorporate mirror reflection features into the pre-trained model; 2) a patch-level pre-training stage to spatially simulate and learn local mirror reflection from image patches; and 3) a pixel-level pre-training stage to pixel-wisely capture mirror reflection via reconstructing corrupted mirror images based on the relationship between the inside and outside of mirrors. Extensive experiments show that our SSL pre-training framework significantly outperforms previous state-of-the-art CNN-based SSL pre-training frameworks and even outperforms supervised ImageNet pre-training when transferred to the mirror detection task. Code and models are available at https://jiaying.link/iccv2023-sslmirror/", + "bbox": [ + 73, + 301, + 473, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 756, + 209, + 771 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mirrors are prevalent in our daily lives. As their appearances are largely determined by their surroundings, they generally lack a consistent appearance, making it difficult to separate them from their surroundings. This may affect many computer vision tasks such as object detection [2], vision-language navigation [1] and depth estimation [29].", + "bbox": [ + 75, + 782, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2053fb7e33566aa3d281812f2faf69bdad556d6356cac893b22a3c6a21bc08bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 267, + 583, + 340 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/97624ca6d1125c6f4390b003b014011a351f9c1950c7ccbc8c0f5ed0ce450959.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 267, + 658, + 340 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a3a1688cab67f58dac910e5e2accc57c4180feddaabc8e18cc4be9edbfbf4b69.jpg", + "image_caption": [ + "(b) MirrorNet on ImageNet" + ], + "image_footnote": [], + "bbox": [ + 586, + 342, + 658, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9c79b20acd129e22e321256d41d612df587f83459c20148214c01a2b2a65136b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 267, + 735, + 340 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/21f231f0b30b96f0819c87b4f82bdc5afed12d33bcd81d816967afb479a37ca7.jpg", + "image_caption": [ + "(c)VCNet on ImageNet" + ], + "image_footnote": [], + "bbox": [ + 661, + 342, + 735, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2407085591717f6000629a100ebbac3458bd38d6891f0485e3353e069cd334da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 267, + 810, + 340 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d85123e767780ec94c3702dff904630533d38df76179b68b715bd5920b81888.jpg", + "image_caption": [ + "Figure 1. State-of-the-art mirror detection methods [37, 30] are based on costly supervised ImageNet pre-training. They may fail even in obvious cases, e.g., (a) when the mirror clearly reflects a real object outside of the mirror. (b) and (c) are MirrorNet [37] and VCNet [30], respectively, pre-trained on ImageNet with full supervision. (c) is MirrorNet with our proposed SSL framework and without supervised ImageNet pre-training. Our SSL scheme leverages the mirror reflection cue and avoids feature redundancy in the pre-training stage. It outperforms those pre-trained on ImageNet with full supervision (i.e., (b) and (c))." + ], + "image_footnote": [], + "bbox": [ + 736, + 342, + 810, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/172cb65753d1987a963d6fe5c399475d5ab9a638956f1ba00e1256d1d7e58382.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 813, + 267, + 887, + 340 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db1d2d05c41cec15303148981a731908325a4eedbf1269610181f34c7e074c40.jpg", + "image_caption": [ + "1) Ours (MirrorNet on our SSL)", + "(e) GT" + ], + "image_footnote": [], + "bbox": [ + 813, + 342, + 887, + 415 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "It is therefore crucial to design high-performance mirror detection methods to facilitate computer vision applications.", + "bbox": [ + 498, + 613, + 890, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, a few methods [37, 23, 30] have been proposed for mirror detection, but they all require to be initialized with supervised ImageNet [8] pre-trained weights and then fine-tuned on mirror detection datasets. While such a transfer learning approach is common in general computer vision tasks (e.g., object detection [24] and semantic segmentation [7]) to utilize category-oriented features from large-scale image datasets, we challenge if this strategy is necessary for the mirror detection task for two reasons. First, the supervised ImageNet pre-training process involves high labelling costs, as it is very time-consuming and labor-intensive to label such a large-scale image dataset ( $\\sim$ 1.3M images). Second, unlike general vision tasks, mirror detection does not require category-level object understanding. Instead, it requires an understanding of the relationship between inside and outside of mirrors (e.g., contrast [37], similarity [23], and visual chirality [30]). Thus,", + "bbox": [ + 496, + 643, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 94, + 886, + 227, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12227", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "while pre-training on ImageNet requires expensive labels, it causes feature redundancy for the mirror detection task.", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Figure 1 shows that existing models pre-trained on ImageNet tend to over-detect the mirror regions. Although the two SOTA methods [37, 30] can correctly locate the mirror regions, they are unable to distinguish between the mirrored objects inside the mirrors and the real objects outside, even though these methods contain well-designed modules to learn to address this problem via the fine-tuning stage. This motivates us to investigate whether the current pre-training approach may affect the final detection performance, and whether we may incorporate some intrinsic mirror properties in the pre-training process to assist mirror detection.", + "bbox": [ + 75, + 121, + 467, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An interesting observation is that humans do not need to be \"trained\" to recognize mirrors. Instead, we learn to recognize them implicitly from young. This aligns with the key idea of self-supervised learning (SSL). Works from neuroscience [31, 20] have shown that humans use mid-level visual cortex to recognize mirror reflection. According to [44], mid-level features are hard to be transferred from supervised pre-training on ImageNet, and SSL is suitable for learning mid-level representations, which inspires us to develop an SSL pre-training framework for modeling mirror reflection.", + "bbox": [ + 75, + 287, + 467, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a new SSL pre-training framework for mirror detection, which explicitly considers mirror reflection during the pre-training process. Our framework does not require human-annotated labels from large-scale image datasets. It consists of three stages to progressively pre-train the backbone network from global to local: 1) an image-level pre-training stage to obtain the representation of mirror reflection globally by recognizing the geometric transformation applied to the image; 2) a patch-level pretraining stage to mimic patch-wise mirror reflection and then learn the spatial correlation between the original object patches and the corresponding mirrored object patches; and 3) a pixel-level pre-training stage to extract the pixel-to-pixel relationship of mirror reflection by image reconstruction; Under such progressive pre-training scheme, our SSL pre-training framework can learn the representation of mirror reflection, and then effectively transfer this knowledge to the subsequent mirror detection process for better detection performances. We conduct comprehensive experiments to demonstrate the effectiveness of our SSL pretraining framework. We show that it can significantly boost the performances of existing mirror detection methods, and outperform other CNN-based SSL pre-training frameworks on the mirror detection task.", + "bbox": [ + 75, + 454, + 467, + 813 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To conclude, this paper makes three key contributions:", + "bbox": [ + 96, + 816, + 455, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- To the best of our knowledge, we are the first to investigate how existing SSL pre-training frameworks perform on the mirror detection task, compared with supervised ImageNet pre-training.", + "bbox": [ + 94, + 839, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new SSL pre-training framework that consists of three stages at different levels to progressively learn the representation of mirror reflection. Compared with the features from a supervised ImageNet pre-trained model, our representation is better due to the reduced gap between the pre-training task and the target downstream task (i.e., mirror detection).", + "- Extensive experiments show that our SSL pre-training framework performs the best among all state-of-the-art SSL methods, and even better than models with supervised ImageNet pre-training." + ], + "bbox": [ + 517, + 90, + 890, + 270 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 285, + 640, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Mirror Detection", + "text_level": 1, + "bbox": [ + 500, + 310, + 666, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mirror detection is an essential task in computer vision, as mirrors are everywhere these days. It is also a challenging task, since mirrors do not possess a consistent appearance, but reflect those of their surroundings. Other vision tasks such as generic object detection and segmentation would tend to detect and segment objects outside as well as inside of the mirror, providing incorrect information on the mirror region, and a depth prediction task may return the depths of the reflected objects inside the mirror region instead of the depth of the mirror.", + "bbox": [ + 496, + 335, + 890, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the mirror detection problem, Yang et al. [37] propose the first large-scale mirror detection dataset and the first deep-learning model, MirrorNet, for mirror detection by learning the contextual contrast between the inside and outside of mirrors. Lin et al. [23] further propose a new model with similarity and edge learning to progressively detect mirrors. Most recently, Guan et al. [15] employ graph convolutional networks to model contextual associations for mirror detection, while Tan et al. propose VCNet [30] to exploit the visual chirality cue at a feature level.", + "bbox": [ + 496, + 486, + 890, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, existing mirror detection methods are all fully-supervised. They also directly adopt the supervised ImageNet [8] pre-training to obtain generic image features. Unlike recent works [19, 18, 22] that propose novel network architectures for mirror detection, we propose in this paper to tackle the mirror detection problem from a learning-based perspective, offering the first self-supervised pre-training framework for mirror detection.", + "bbox": [ + 496, + 638, + 890, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Self-supervised Learning (SSL)", + "text_level": 1, + "bbox": [ + 500, + 770, + 774, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SSL is a popular research topic in computer vision [38, 39], and has attracted a lot of research interests. It aims to learn good feature representations without needing labeled data. Early SSL methods focus on proposing novel pretext tasks, such as relative location prediction [10] and rotation prediction [21]. Recently, most methods are based on the utilization of contrastive learning [16, 4, 14, 3, 32, 5, 41,", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12228", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/908689b51dc7553c0ccc26fde83f0bd65eb6d802e9a616f0f113c004d699d59b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 88, + 459, + 183 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/855243fd694ad6ad2d7363759b0686b9465af05eab1878a3b52b8f769d8e9a32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 464, + 88, + 890, + 183 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d955a6d1249ed61c65caf126c7f5730a39a2581f3573b017e7bde9fc5f13c937.jpg", + "image_caption": [ + "Figure 2. Our SSL pre-training framework. It consists of three pre-training stages at different levels: an image-level pre-training stage, a patch-level pre-training stage, and a pixel-level pre-training stage. It pre-trains the backbone network from global to local (image-level $\\rightarrow$ patch-level $\\rightarrow$ pixel-level) progressively." + ], + "image_footnote": [], + "bbox": [ + 163, + 183, + 834, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "40]. Most recently, [44] shows that some SSL methods perform even better than supervised ImageNet pre-training in downstream tasks, such as object detection [11, 24] and semantic segmentation [7, 34], with the help of instance discrimination during SSL pre-training. Recent transformer-based SSL frameworks [36] also involve masked image modeling to obtain better self-supervised representations.", + "bbox": [ + 75, + 412, + 468, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we observe that unlike general vision tasks, mirror detection is not benefited from supervised ImageNet pre-training since it does not require category-level object understanding. Hence, we propose a novel SSL pre-training framework specifically for the mirror detection task. The proposed framework aims to simulate and learn the relationship between mirrors and their surrounding during the pre-training stage.", + "bbox": [ + 75, + 518, + 470, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 656, + 168, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 shows the overall architecture of the proposed SSL pre-training framework. Our SSL pre-training framework consists of three pre-training stages at different levels: an image-level pre-training, a patch-level pre-training, and a pixel-level pre-training. The pre-training at each level is conducted sequentially and independently. The progressive and sequential training strategy can help the pre-trained model obtain a global-to-local representation of mirrors.", + "bbox": [ + 75, + 681, + 468, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Image-level Pre-training", + "text_level": 1, + "bbox": [ + 76, + 816, + 302, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the definition of visual chirality [26] and its potential applications [46, 30] in various computer vision problems, we formulate the image-level pre-training as a binary classification problem: Given an input image and its", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "corresponding horizontal flipped image, can the backbone network tell which one is the original image? To achieve this objective, we attach a classification head to predict the probability that the input image is flipped. The classification head consists of a convolution layer followed by a sigmoid activation function to project the output value into $[0,1]$ . In this pre-training stage, we use binary cross-entropy (BCE) as our loss function $\\mathcal{L}_f$ to optimize the backbone network. Formally, we have:", + "bbox": [ + 496, + 412, + 892, + 549 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f} = - \\left[ y \\log (\\hat {y}) + (1 - y) \\log (1 - \\hat {y}) \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 556, + 890, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $y$ is the ground truth label and $\\hat{y}$ is the predicted probability that the input image is horizontally flipped.", + "bbox": [ + 496, + 582, + 890, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Patch-level Pre-training", + "text_level": 1, + "bbox": [ + 500, + 619, + 720, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The image-level pre-training pretext task alone cannot bring much help in locating the mirror inside an image since mirror detection is a pixel-level problem. However, directly connecting the image-level and the pixel-level pretext tasks is challenging due to the gap between the scales of these two pretext tasks. Thus, we design a patch-level pre-training stage as a bridge to cover the feature representations of mirrors at different scales. Besides, the image-level pre-training does not benefit relational spatial understanding, which has been proven important in mirror detection [23]. Our patch-level pre-training consists of two sub-tasks to both obtain the patch-wise relationship of mirror reflection and simulate the mirrored region with pseudo labels. We first split the input image into a grid of $3 \\times 3$ image patches and randomly shuffle them. After that, we randomly apply a flipping operation to one of the image patches, to simulate a mirrored region. The first sub-task", + "bbox": [ + 496, + 643, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12229", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is to predict the location index from the shuffled grid by the backbone network. We append a location predictor head to the backbone network to predict the location index of input patches. We apply cross-entropy loss as $\\mathcal{L}_{loc}$ to optimize this sub-task. The second sub-task is to predict the flipped patch from the nine randomly shuffled image patches. Similar to our image-level pre-training, we use BCE loss as our loss function $\\mathcal{L}_{cls}$ in this sub-task.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l o c} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} y _ {i} \\log \\left(\\hat {y} _ {i}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 223, + 468, + 263 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} = - \\frac {1}{N} \\sum_ {j = 1} ^ {N} y _ {j} \\log \\left(\\hat {y _ {j}}\\right) + (1 - y _ {j}) \\log \\left(1 - \\hat {y _ {j}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 268, + 468, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{y}_i$ is the predicted location index, and $y_i$ is the corresponding ground truth label. $\\hat{y}_j$ is the predicted probability that the input patch is horizontally flipped and $y_j$ is the corresponding ground truth probability. Finally, we jointly optimize these two sub-tasks and the final loss function is:", + "bbox": [ + 75, + 321, + 468, + 398 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p a t c h}} = \\mathcal {L} _ {\\text {l o c}} + \\mathcal {L} _ {\\text {c l s}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 425, + 468, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Pixel-level Pre-training", + "text_level": 1, + "bbox": [ + 76, + 453, + 292, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Extracting the pixel-level relationship of mirror reflection is a challenging task, as studied by previous related work [12] for symmetry detection. It requires point-level GT labels to represent symmetry centers. However, obtaining such labels requires huge human annotation effort and they are not applicable for SSL pre-training. Instead of generating pseudo labels for symmetry detection, our solution is that we formulate this task as an image reconstruction problem on images containing mirrors. An ideal pre-trained network should be able to reconstruct the input image containing mirrors with the pixel-level understanding of mirror reflection. We randomly mask out the input image by cutout augmentation [9] with the ratio of 0.3 and then feed it to the backbone network with a reconstruction head. The details of the reconstruction head we used are listed in Table 1. The objective of our pixel-level pre-training is defined as:", + "bbox": [ + 75, + 477, + 472, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p i x} = \\| x - \\hat {x} \\| _ {1}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 732, + 468, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{pix}$ is an L1 loss. $x$ and $\\hat{x}$ are the input image and the reconstructed image, respectively.", + "bbox": [ + 75, + 761, + 468, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 806, + 209, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 830, + 284, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dataset and Evaluation Metrics. The image-level and patch-level pre-training experiments are conducted on the training set of ImageNet-100 [28] and MS COCO [25],", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. The architecture of the reconstruction head used in the pixel-level pre-training. Note that each \"conv-IR\" corresponds to a sequence of convolution layer, InstanceNorm layer and ReLU activation. $K$ , $S$ and $P$ denote the number of kernels, the number of strides and the padding size, respectively, used in the convolution layer.", + "bbox": [ + 496, + 85, + 893, + 170 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/d5f1df4fcdc5de47ce0dcf08320a1def3b98b1b81fdc75f45f072bacb7572ba9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Layer Details
3 × 3 conv-IR, K = 1024, S = 1, P = 1
2 × Upsample
3 × 3 conv-IR, K = 512, S = 1, P = 1
2 × Upsample
3 × 3 conv-IR, K = 256, S = 1, P = 1
2 × Upsample
3 × 3 conv-IR, K = 128, S = 1, P = 1
2 × Upsample
3 × 3 conv-IR, K = 128, S = 1, P = 1
2 × Upsample
1 × 1 conv-IR, K = 3, S = 1, P = 1
", + "bbox": [ + 558, + 171, + 834, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which contains $130\\mathrm{K}$ and $\\sim 118\\mathrm{k}$ images respectively. In the pixel-level pre-training and fine-tuning stage, our experiments are conducted on MSD [37], the first large-scale dataset for mirror detection. MSD [37] provides 4,018 mirror images, which are divided into 3,063 images for training and 955 images for testing. We adopt two popular metrics namely, F-measure $(\\mathbf{F}_{\\beta})$ and mean absolute error (MAE) to evaluate the performance of our SSL pre-training framework. F-measure can reflect the overall model performance and is calculated by the weighted harmonic mean of the precision and recall: $F_{\\beta} = \\frac{(1 + \\beta^{2})\\cdot\\text{precision}\\cdot\\text{recall}}{\\beta^{2}\\cdot\\text{precision} + \\text{recall}}$ , where $\\beta^2$ is set to 0.3. MAE can measure the average pixel-wise absolute disparity between the predicted mirror maps $M$ and the binary ground truth $G$ .", + "bbox": [ + 496, + 393, + 890, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implementation Details. We use Pytorch to implement our proposed SSL pre-training framework. All experiments are conducted on a GPU server with eight NVIDIA RTX 3090 GPUs. Following existing SSL pre-training frameworks, we adopt ResNet-50 [17] as our backbone network. All pre-trained model weights of existing SSL pretraining frameworks are obtained from the released models of mmselfsup [6] for fair comparisons. For each pretext task in the pre-training stage, all images are resized to $256 \\times 256$ . The batch size is set to 352. We use AdamW [27] with an initial learning rate of 1e-4 to optimize our SSL pre-training framework and stop pre-training after 20,000 iterations. We only adopt random color jittering for data augmentation. In the fine-tuning stage, we use the same hyperparameters reported in the original papers of the adopted", + "bbox": [ + 496, + 611, + 892, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1It is worth noting that the datasets that we used (about 250K training images) are much smaller than the pre-trained dataset ImageNet (14 million images) used by other baseline SSL methods. We adopted ImageNet-100 to pre-train our network due to the limited computational resources.", + "bbox": [ + 500, + 849, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12230", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fb1c2b369ed7aed8c0640a1bc2c105e7f93e01f49c8804770f246f9699221c5a.jpg", + "image_caption": [ + "Figure 3. Qualitative Comparison of our proposed method with NPID [33], MocoV2 [4], SwAV [3], Barlow Twins [41], ImageNet pretraining (Supervised). We use the released pre-trained models of these existing SSL methods from mmselfsup [6] for fair comparison." + ], + "image_footnote": [], + "bbox": [ + 84, + 87, + 890, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "base methods MirrorNet [37] and VCNet [30].", + "bbox": [ + 76, + 657, + 383, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Arts", + "text_level": 1, + "bbox": [ + 76, + 686, + 382, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare against 11 state-of-the-art CNN-based frameworks in self-supervised learning, including Relative Location (Rel. Loc.) [10], Rotation Prediction (Rot. Pred.) [21], NPID [33], ODC [43], Moco V2 [16], SimCLR [4], BYOL [14], SwAV [3], DenseCL [32], SimSiam [5] and Barlow Twins (Bar. Twins) [41]. We select MirrorNet [37] and VCNet [30], which are two representative mirror detection methods, as base methods to conduct experiments in the fine-tuning stage.", + "bbox": [ + 75, + 710, + 468, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative Evaluation. We further show the performance of our SSL pre-training framework in Figure 3. We compare our framework with four state-of-the-art SSL pre-training", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "frameworks (including the two best-performing frameworks NPID [33] and Moco V2 [16] based on MirrorNet and the two best-performing frameworks SwAV [3] and Barlow Twins [41] based on VCNet, according to their performance in Table 2) and supervised ImageNet pre-training due to the limitation of space. Our framework can correctly detect the mirror regions by utilizing the relationship of mirror reflection caused by mirrors, even though the reflected regions are pure texture (i.e., lack of object keypoints in the second row) and relatively small in the image (e.g., the last three rows). We attribute the superior performance of our framework to the utilization of mirror reflection in the pre-training stage.", + "bbox": [ + 496, + 657, + 890, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative Evaluation. Table 2 shows the experimentation results comparing the proposed SSL pre-training framework with the state-of-the-arts. Besides, we also re", + "bbox": [ + 498, + 854, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12231", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2f5fd76e72a5d854ad1405ab978ab00692d8bc861f3d00afd8860386de67b80f.jpg", + "table_caption": [ + "Table 2. Quantitative comparison between our SSL pre-training framework and existing CNN-based SSL methods. We also compare it with random initialization (Random) and ImageNet pretraining (Supervised)." + ], + "table_footnote": [], + "table_body": "
MethodsVenueMirrorNetVCNet
Fβ↑MAE↓Fβ↑MAE↓
Rel. Loc.ICCV '150.5820.2540.8420.072
Rot. Pred.ICLR '180.5370.2120.8510.069
NPIDCVPR '180.7390.1320.8320.076
ODCCVPR '200.5580.2230.8360.077
Moco V2CVPR '200.7070.1770.8350.072
SimCLRICML '200.5940.2440.8370.071
BYOLNeurIPS '200.7040.1580.8320.077
SwAVNeurIPS '200.5780.2240.8560.071
DenseCLCVPR '210.6940.2400.8440.072
SimSiamCVPR '210.6720.1980.8180.080
Bar. TwinsICML '210.6550.1780.8570.069
Random0.5860.2000.4000.103
Supervised20.7270.1700.8710.062
Ours0.7630.1160.8860.057
", + "bbox": [ + 76, + 142, + 490, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "port the results from models adopting Xavier initialization [13] (Random) or supervised ImageNet pre-training (Supervised). Our SSL pre-training framework (Ours) achieves the best performance across both metrics when applied on MirrorNet and VCNet and shows a good generalization of different mirror detection methods. It is worth noting that our framework even outperforms the supervised ImageNet pre-training.", + "bbox": [ + 75, + 436, + 468, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 566, + 228, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 demonstrates the effectiveness of each pretraining stage in our framework. We analyze its results point by point as follows.", + "bbox": [ + 75, + 590, + 468, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image-level Pre-training vs. Random Initialization. To evaluate the effectiveness of the proposed image-level pretraining, we compare the model only adopting image-level pre-training (B2) with the model initialized randomly without any pre-training (B1). We can see that only adopting image-level pre-training (B2) will even cause a performance drop compared with the one without pre-training (B1). This somehow supports the finding from VCNet [30] that image-level flipping cannot provide sufficient information to pixelwisely locate the mirror inside an image.", + "bbox": [ + 75, + 642, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2f7cfdf59eca0a1b160578d6d2e9538726cbed7619ecb9b41733fca790aad8b2.jpg", + "table_caption": [ + "Table 3. Ablation study of the proposed SSL pre-training framework. We use MirrorNet [37] trained on the MSD dataset as our base method to conduct the experiments in this ablation study. Best results are shown in bold." + ], + "table_footnote": [], + "table_body": "
Rand. Init.Image-lev. PretrainPatch-lev. PretrainPixel-lev. PretrainFβ↑MAE↓
B1✓0.6010.195
B2✓0.6230.191
B3✓0.6890.162
B4✓0.6200.198
B5✓✓0.7230.166
B6✓✓0.7580.124
B7✓✓0.6660.178
Ours✓✓✓0.7630.116
", + "bbox": [ + 500, + 141, + 898, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Patch-level Pre-training vs. Random Initialization. We compare our patch-level pre-training (B3) with random initialization (B1) to analyze if patch-level pre-training is useful in our SSL pre-training framework. Our patch-level pretraining outperforms the model without pre-training on both two metrics (14.64% improvement on $F_{\\beta}$ and 16.32% improvement on MAE), which indicates the effectiveness of our patch-level pre-training. It also demonstrates the importance of modeling the spatial relationship of mirror reflection at a patch level.", + "bbox": [ + 498, + 311, + 893, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pixel-level Pre-training vs. Random Initialization. Similarly, we conduct the ablation experiment on our pixel-level pre-training. The model adopting pixel-level pre-training performs better than the model with image pre-training. However, the performance gain is not as much as the model with patch-level pre-training (B3). One possible reason is that directly optimizing the backbone network for the representation of mirrors at a pixel level is challenging and may not produce satisfactory results for mirror detection.", + "bbox": [ + 498, + 473, + 890, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Location Predictor in Patch-level Pre-training. The location predictor in our patch-level pre-training is designed to capture the relational spatial information, which has been proven important in mirror detection [23]. We conduct an ablation experiment to remove the location predictor and the performance of our methods drops from 0.763 to 0.731. This shows the importance of learning spatial information in the mirror detection task.", + "bbox": [ + 498, + 618, + 890, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Combinations of Different Pre-training Stages. Based on the above three comparisons, we find that the patch-level pre-training contributes the most improvement. It demonstrates the importance of modeling the spatial relationship of mirror reflection at a patch level. We then analyze the results from the model with different combinations of our pretraining stages. The model with image-level pre-training and patch-level pre-training (B5) performs better than the model that only adopts patch-level pre-training (B3), which shows that the global information provided by image-level", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2Note that the original backbone networks reported in their papers are ResNext-101 [35]. However, the model weights of ResNext-101 pre-trained by our baseline SSL frameworks are not available in mmselfsup [6]. Thus, we switch the backbone network of these two methods to ResNet-50 [17], which is a common backbone network used in evaluating the performances of SSL frameworks. We also directly use the released ResNet-50 SSL pre-trained weights in our experiments for fair comparisons.", + "bbox": [ + 75, + 801, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12232", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c94e799004490cc55ca02611f6e1d3d0a18c07e817a95c713108ccb4eda94f38.jpg", + "image_caption": [ + "Image" + ], + "image_footnote": [], + "bbox": [ + 107, + 103, + 215, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/adb43e744b866dd7ca54d2d708480002a84938a4c0f338e7cd4663448de446fd.jpg", + "image_caption": [ + "Image-level Pretraining" + ], + "image_footnote": [], + "bbox": [ + 217, + 103, + 323, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef97824b33362b7b6edd9ca6e0ece5d9cda25eeb9aba72ad576370e87fd8e77c.jpg", + "image_caption": [ + "Patch-level Pretraining" + ], + "image_footnote": [], + "bbox": [ + 325, + 103, + 433, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6155df1df3d0122b4be72a7609d83ada630161ef92e3acf130d06f647cb0f2fd.jpg", + "image_caption": [ + "Pixel-level Pretraining" + ], + "image_footnote": [], + "bbox": [ + 434, + 103, + 540, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db77c22435167185c10f3cc363d5ff96d45d33e0f592430360e6dbc4e66205d5.jpg", + "image_caption": [ + "Figure 4. Visualization of the effects of the proposed pre-training stages." + ], + "image_footnote": [], + "bbox": [ + 542, + 103, + 648, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/20de36026fd80a4e087ead7e427f1f5f25a35580a4efbb87fc1d63e9b09084d1.jpg", + "image_caption": [ + "(Image+Patch)-level Pretraining", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 651, + 103, + 756, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2247c9baf0804f7f3164b3547a535be04149800f888792e44bb69438800898a8.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 759, + 103, + 867, + 207 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pre-training can benefit the model with patch-level pretraining. Similarly, as shown in B7, B4 (i.e., only pixel-level pretraining) can boost the performance of the pre-trained model after image-level pre-training. These two comparisons indicate that image-level pre-training should be used with other pre-training stages to benefit the overall performance and it should not be solely adopted to our SSL pretraining framework. Besides, we find that combining patch-level and pixel-level pre-training (B6) has superior performance when compared with B3, especially on MAE with about $23.4\\%$ improvement. This combination also outperforms supervised ImageNet pre-training, according to the results from Table 2. We attribute its superior performance to the pixel-level representation of mirror reflection. Our final model with all pre-training stages outperforms other ablation models, showing the effectiveness of adopting pretraining stages progressively.", + "bbox": [ + 75, + 285, + 472, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual Comparison of Our Pre-training Stages. Figure 4 shows a visual example of our ablation study. We can see that only adopting image-level pre-training on our SSL pre-training framework can approximately locate the mirror region but fails to produce a precise boundary for the predicted mask of the mirror. Patch-level pre-training can help predict mirrors more accurately with clearer boundaries but over-predict the non-mirror regions as mirrors. In comparison to predictions from all other ablation models, the prediction from the model with pixel-level pretraining has the clearest boundary but contains the largest over-predicted regions, primarily because of the insufficient global-level learning on mirrors. Our framework can significantly reduce over-predictions when image-level and patch-level pre-training are used, as opposed to when patch-level pre-training is used only, and performs the best with the help of the global-to-local pre-training process.", + "bbox": [ + 75, + 549, + 468, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Discussions", + "text_level": 1, + "bbox": [ + 76, + 816, + 202, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impacts of Pre-training Strategy. Some previous SSL works [42, 45], especially for those focusing on self-supervised pre-training for downstream tasks like segmentation [42] and salient object detection [45], would adopt", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9739fadea13fe3780833f9878be510cf87ad0a19c2692b950d139ef45f588e77.jpg", + "table_caption": [ + "Table 4. The results of MirrorNet [37] when using different pretraining strategies. Best results are shown in bold." + ], + "table_footnote": [], + "table_body": "
Fβ↑MAE↓
Random Initialized0.5860.200
ImageNet Supervised0.7270.170
Ours (In parallel)0.7080.331
Ours (Sequential)0.7630.116
", + "bbox": [ + 563, + 308, + 828, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "different pretext tasks sequentially to pre-train their backbone networks. We also adopt this common strategy in our proposed SSL framework. To verify if employing different pre-training strategies may affect the performance of the proposed SSL framework, we have tried training the three stages sequentially as well as in parallel. We find that when pre-training the three stages in parallel, the backbone network fails to converge and its final fine-tuning results are much worse than those using sequential pre-training. Table 4 shows the results of this experiment. We can see that the MAE performance of pre-training the different stages in parallel is even worse than that of random initialization. A possible reason for this is that directly pre-training different pretext tasks in parallel is difficult to transfer the learned knowledge from the pre-trained network to the target downstream task (i.e., pixel-level mirror detection).", + "bbox": [ + 496, + 435, + 890, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Apart from the superior performance made by the sequential pre-training strategy, we also notice that the number of pre-training iterations for each pretext task can heavily affect the performance of our SSL framework. To avoid heavy hyperparameter tuning and increase the reproducibility of our pre-training framework, we adopt the same number of pre-training iterations for all stages. Although this strategy of using the same number of pre-training iterations is likely not the optimal design for individual stages, we believe that it can better reflect the real performance of our framework with less stochasticity and more reactivity. We provide a more detailed discussion on this issue next.", + "bbox": [ + 496, + 680, + 890, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impacts of Pre-training Iterations. SSL pre-training frameworks usually require huge pre-training iterations to", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12233", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb08feac410142ebb0149c6e0bde878812dd728c60a1ad1abf429cb263c5f165.jpg", + "image_caption": [ + "Figure 5. Ablation study on the number of pre-training iterations for our SSL pre-training framework. We find that our SSL pretraining framework continues to improve in performance until $\\sim 20,000$ pre-training iterations. After that, the performance starts to decline, mainly due to overfitting." + ], + "image_footnote": [], + "bbox": [ + 133, + 90, + 410, + 275 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "converge. For example, SimSiam [5] requires about 250K iterations to pre-train their model on ImageNet, which is very time-consuming. To validate how this issue may affect our SSL pre-training framework, we conduct the ablation study on the number of pre-training iterations for our SSL pre-training framework, and the results are shown in Figure 5. Note that we use the same number of iterations for all pre-training stages for convenience and to get rid of heavy hyperparameter tuning.", + "bbox": [ + 75, + 372, + 467, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From the results, we find that the performance of our SSL pre-training framework improves rapidly as the number of pre-training iterations increases, up to around 20,000, which is a relatively small number compared with those used in existing SSL pre-training frameworks. The performance of our SSL pre-training framework gradually drops as the number of pre-training iterations exceeds 20,000, mainly due to overfitting. One possible explanation of this is that the model pre-trained by our SSL pre-training framework can efficiently learn the representation of mirror reflection under a short pre-training period. Another possible explanation is that unlike generic SSL pre-training frameworks that attempt to extract general-propose image representation, our SSL pre-training framework is developed specially for a specific task, i.e., mirror detection. While generic SSL pre-training may cause feature redundancy for our task, having too many pre-training iterations (i.e., over 20,000) causes overfitting.", + "bbox": [ + 75, + 511, + 467, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Failure Cases. Our SSL pre-training framework does have limitations. It continues to struggle with the limitations of the base method [37], as the base method adopted can also significantly affect the final detection performances. For example, as shown in first row of Figure 6, MirrorNet with our SSL pre-training may still fail to handle some mirror-like regions (e.g., the top-right region). It may also fail to de", + "bbox": [ + 75, + 794, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5e0fa381c7da1010282d6fc2a226bfa379b1d1dfb405d95dcf84c3caa7ac4974.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 87, + 635, + 165 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a21f8058477846417d25926fda83d6439c42b732b24f2a0af94409428e39c431.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 88, + 759, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5739f80c78e3835cac5bc18142f5c73c76e0a05db49d5b10faeba1c9b4c9370a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 88, + 883, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7d9e1db27b0ef5915888bd41c21fdb0bf768ae131bbe9dc9f456e9e524bf3e5c.jpg", + "image_caption": [ + "Image" + ], + "image_footnote": [], + "bbox": [ + 511, + 167, + 635, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/646ac3372c6f1b8574c145ef0ee9f0430b5de75bf66388c8b1ba8e482a373c38.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 637, + 167, + 759, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ef33e88d90d37b68820d4534a4f8cef4e29bf268f405ad1c5d3162e47e1c4c44.jpg", + "image_caption": [ + "GT", + "Figure 6. Failure cases of our SSL pre-training framework based on MirrorNet. Our SSL pre-training framework fails to tackle the intrinsic limitation of existing mirror detection methods. The model pre-trained by our SSL pre-training framework still has wrong predictions when the input images are challenging (e.g., containing mirror-like regions in the first row, and insufficient relationship between mirror/non-mirror regions in the second row." + ], + "image_footnote": [], + "bbox": [ + 761, + 167, + 883, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tect mirrors in some challenging cases when the input image contains ambiguous mirrors without the relationship of mirror reflection and sufficient contextual contrast, as shown in the second row of Figure 6.", + "bbox": [ + 498, + 441, + 890, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 521, + 617, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have investigated how self-supervised learning (SSL) works with the mirror detection task. To the best of our knowledge, we are the first to explore SSL pre-training frameworks applied to mirror detection. We have found that the supervised ImageNet pre-training might not be the ideal way to extract backbone image features for mirror detection due to the discrepancy between general-propose representation and mirror-specific representation. We have also proposed a new SSL pre-training framework to pre-train the backbone network for mirror detection. Our SSL pre-training framework does not require any labeled data. It progressively (image-, patch- and pixel-level pre-training) models the relationship of mirror reflection, and obtains a better representation of mirrors in the pre-training stage. Experimental results show that the proposed SSL pre-training framework outperforms CNN-based state-of-the-art SSL pre-training frameworks, and even achieves better results compared with supervised ImageNet pre-training.", + "bbox": [ + 496, + 547, + 890, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: The work described in this paper was partially supported by a GRF grant from the Research Grants Council of Hong Kong (Project No.: CityU 11211223).", + "bbox": [ + 498, + 839, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12234", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, pages 3674-3683, 2018.", + "[2] M. Braun, S. Krebs, F. Flohr, and D. M. Gavrila. Eurocity persons: A novel benchmark for person detection in traffic scenes. IEEE TPAMI, 41(8):1844-1861, Aug 2019.", + "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020.", + "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020.", + "[5] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021.", + "[6] MMSelfSup Contributors. MMSelfSup: Openmmlab self-supervised learning toolbox and benchmark. https://github.com/open-mmlab/mmselfsup, 2021.", + "[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016.", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009.", + "[9] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017.", + "[10] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In ICCV, 2015.", + "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010.", + "[12] Christopher Funk and Yanxi Liu. Beyond planar symmetry: Modeling human perception of reflection and rotation symmetries in the wild. In ICCV, pages 793-803, 2017.", + "[13] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249–256. JMLR Workshop and Conference Proceedings, 2010.", + "[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS, 2020.", + "[15] Huankang Guan, Jiaying Lin, and Rynson W.H. Lau. Learning semantic associations for mirror detection. In CVPR, pages 5941-5950, June 2022." + ], + "bbox": [ + 78, + 116, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020.", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016.", + "[18] Ruozhen He, Jiaying Lin, and Rynson WH Lau. Efficient mirror detection via multi-level heterogeneous learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 790-798, 2023.", + "[19] Tianyu Huang, Bowen Dong, Jiaying Lin, Xiaohui Liu, Rynson WH Lau, and Wangmeng Zuo. Symmetry-aware transformer-based mirror detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 935–943, 2023.", + "[20] Peter Kohler, Alexandra Yakovleva, Alasdair Clarke, Yanxi Liu, and Anthony Norcia. Parametric responses to rotation symmetry in mid-level visual cortex. Journal of Vision, 15(12):1122-1122, 2015.", + "[21] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In ICLR, 2018.", + "[22] Jiaying Lin, Xin Tan, and Rynson WH Lau. Learning to detect mirrors from videos via dual correspondences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9109-9118, 2023.", + "[23] Jiaying Lin, Guodong Wang, and Rynson W. H. Lau. Progressive mirror detection. In CVPR, 2020.", + "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014.", + "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014.", + "[26] Zhiqiu Lin, Jin Sun, Abe Davis, and Noah Snavely. Visual chirality. In CVPR, pages 12295-12303, 2020.", + "[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2018.", + "[28] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015.", + "[29] Jiaqi Tan, Weijie Lin, Angel X Chang, and Manolis Savva. Mirror3d: Depth refinement for mirror surfaces. In CVPR, pages 15990-15999, 2021.", + "[30] Xin Tan, Jiaying Lin, Ke Xu, Chen Pan, Lizhuang Ma, and Rynson W. H. Lau. Mirror detection with the visual chirality cue. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022.", + "[31] Christopher W Tyler. Human symmetry perception and its computational analysis. Psychology Press, 2003.", + "[32] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In CVPR, 2021." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12235", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018.", + "[34] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, pages 418-434, 2018.", + "[35] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, pages 1492-1500, 2017.", + "[36] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, pages 9653-9663, 2022.", + "[37] Xin Yang, Haiyang Mei, Ke Xu, Xiaopeng Wei, Baocai Yin, and Rynson W.H. Lau. Where is my mirror? In ICCV, 2019.", + "[38] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. Learning with noisy labels for robust point cloud segmentation. In ICCV, pages 6443-6452, 2021.", + "[39] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. Robust point cloud segmentation with noisy annotations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7696-7710, 2022.", + "[40] Shuquan Ye, Yujia Xie, Dongdong Chen, Yichong Xu, Lu Yuan, Chenguang Zhu, and Jing Liao. Improving commonsense in vision-language models via knowledge graph riddles. In CVPR, pages 2634-2645, 2023.", + "[41] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021.", + "[42] Xiaohang Zhan, Ziwei Liu, Ping Luo, Xiaou Tang, and Chen Loy. Mix-and-match tuning for self-supervised semantic segmentation. In AAAI, volume 32, 2018.", + "[43] Xiaohang Zhan, Jiahao Xie, Ziwei Liu, Yew-Soon Ong, and Chen Change Loy. Online deep clustering for unsupervised representation learning. In CVPR, 2020.", + "[44] Nanxuan Zhao, Zhirong Wu, Rynson WH Lau, and Stephen Lin. What makes instance discrimination good for transfer learning? In ICLR, 2021.", + "[45] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, , Hutchuan Lu, and Xiang Ruan. Self-supervised pretraining for rgb-d salient object detection. In AAAI, 2022.", + "[46] Ying Zheng, Yiyi Zhang, Xiaogang Xu, Jun Wang, and Hongxun Yao. Visual chirality meets freehand sketches. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1544-1548. IEEE, 2021." + ], + "bbox": [ + 78, + 90, + 467, + 729 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12236", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Self-supervised Pre-training for Mirror Detection/9faec277-19c4-4583-a3c8-84bf53084f2f_model.json b/2023/Self-supervised Pre-training for Mirror Detection/9faec277-19c4-4583-a3c8-84bf53084f2f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1c86bced3c9ce88496432c58e56d0fbc733c3da2 --- /dev/null +++ b/2023/Self-supervised Pre-training for Mirror Detection/9faec277-19c4-4583-a3c8-84bf53084f2f_model.json @@ -0,0 +1,2200 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.131, + 0.737, + 0.154 + ], + "angle": 0, + "content": "Self-supervised Pre-training for Mirror Detection" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.182, + 0.626, + 0.218 + ], + "angle": 0, + "content": "Jiaying Lin* Rynson W.H. Lau* City University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.324, + 0.22, + 0.648, + 0.234 + ], + "angle": 0, + "content": "{jiayinlin5, rynson.lau}@cityu.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.302, + 0.474, + 0.726 + ], + "angle": 0, + "content": "Existing mirror detection methods require supervised ImageNet pre-training to obtain good general-purpose image features. However, supervised ImageNet pre-training focuses on category-level discrimination and may not be suitable for downstream tasks like mirror detection, due to the overfitting upstream tasks (e.g., supervised image classification). We observe that mirror reflection is crucial to how people perceive the presence of mirrors, and such mid-level features can be better transferred from self-supervised pretrained models. Inspired by this observation, in this paper we aim to improve mirror detection methods by proposing a new self-supervised learning (SSL) pre-training framework for modeling the representation of mirror reflection progressively in the pre-training process. Our framework consists of three pre-training stages at different levels: 1) an image-level pre-training stage to globally incorporate mirror reflection features into the pre-trained model; 2) a patch-level pre-training stage to spatially simulate and learn local mirror reflection from image patches; and 3) a pixel-level pre-training stage to pixel-wisely capture mirror reflection via reconstructing corrupted mirror images based on the relationship between the inside and outside of mirrors. Extensive experiments show that our SSL pre-training framework significantly outperforms previous state-of-the-art CNN-based SSL pre-training frameworks and even outperforms supervised ImageNet pre-training when transferred to the mirror detection task. Code and models are available at https://jiaying.link/iccv2023-sslmirror/" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.757, + 0.21, + 0.772 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.783, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Mirrors are prevalent in our daily lives. As their appearances are largely determined by their surroundings, they generally lack a consistent appearance, making it difficult to separate them from their surroundings. This may affect many computer vision tasks such as object detection [2], vision-language navigation [1] and depth estimation [29]." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.268, + 0.584, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.268, + 0.659, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.343, + 0.659, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.418, + 0.655, + 0.439 + ], + "angle": 0, + "content": "(b) MirrorNet on ImageNet" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.268, + 0.736, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.343, + 0.736, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.418, + 0.73, + 0.439 + ], + "angle": 0, + "content": "(c)VCNet on ImageNet" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.268, + 0.812, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.343, + 0.812, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.737, + 0.418, + 0.822, + 0.439 + ], + "angle": 0, + "content": "1) Ours (MirrorNet on our SSL)" + }, + { + "type": "image", + "bbox": [ + 0.815, + 0.268, + 0.888, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.815, + 0.343, + 0.888, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.835, + 0.424, + 0.867, + 0.434 + ], + "angle": 0, + "content": "(e) GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.45, + 0.893, + 0.59 + ], + "angle": 0, + "content": "Figure 1. State-of-the-art mirror detection methods [37, 30] are based on costly supervised ImageNet pre-training. They may fail even in obvious cases, e.g., (a) when the mirror clearly reflects a real object outside of the mirror. (b) and (c) are MirrorNet [37] and VCNet [30], respectively, pre-trained on ImageNet with full supervision. (c) is MirrorNet with our proposed SSL framework and without supervised ImageNet pre-training. Our SSL scheme leverages the mirror reflection cue and avoids feature redundancy in the pre-training stage. It outperforms those pre-trained on ImageNet with full supervision (i.e., (b) and (c))." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.614, + 0.892, + 0.644 + ], + "angle": 0, + "content": "It is therefore crucial to design high-performance mirror detection methods to facilitate computer vision applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Recently, a few methods [37, 23, 30] have been proposed for mirror detection, but they all require to be initialized with supervised ImageNet [8] pre-trained weights and then fine-tuned on mirror detection datasets. While such a transfer learning approach is common in general computer vision tasks (e.g., object detection [24] and semantic segmentation [7]) to utilize category-oriented features from large-scale image datasets, we challenge if this strategy is necessary for the mirror detection task for two reasons. First, the supervised ImageNet pre-training process involves high labelling costs, as it is very time-consuming and labor-intensive to label such a large-scale image dataset (\\(\\sim\\)1.3M images). Second, unlike general vision tasks, mirror detection does not require category-level object understanding. Instead, it requires an understanding of the relationship between inside and outside of mirrors (e.g., contrast [37], similarity [23], and visual chirality [30]). Thus," + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.228, + 0.9 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12227" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "while pre-training on ImageNet requires expensive labels, it causes feature redundancy for the mirror detection task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.287 + ], + "angle": 0, + "content": "Figure 1 shows that existing models pre-trained on ImageNet tend to over-detect the mirror regions. Although the two SOTA methods [37, 30] can correctly locate the mirror regions, they are unable to distinguish between the mirrored objects inside the mirrors and the real objects outside, even though these methods contain well-designed modules to learn to address this problem via the fine-tuning stage. This motivates us to investigate whether the current pre-training approach may affect the final detection performance, and whether we may incorporate some intrinsic mirror properties in the pre-training process to assist mirror detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.468, + 0.452 + ], + "angle": 0, + "content": "An interesting observation is that humans do not need to be \"trained\" to recognize mirrors. Instead, we learn to recognize them implicitly from young. This aligns with the key idea of self-supervised learning (SSL). Works from neuroscience [31, 20] have shown that humans use mid-level visual cortex to recognize mirror reflection. According to [44], mid-level features are hard to be transferred from supervised pre-training on ImageNet, and SSL is suitable for learning mid-level representations, which inspires us to develop an SSL pre-training framework for modeling mirror reflection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.468, + 0.814 + ], + "angle": 0, + "content": "In this paper, we propose a new SSL pre-training framework for mirror detection, which explicitly considers mirror reflection during the pre-training process. Our framework does not require human-annotated labels from large-scale image datasets. It consists of three stages to progressively pre-train the backbone network from global to local: 1) an image-level pre-training stage to obtain the representation of mirror reflection globally by recognizing the geometric transformation applied to the image; 2) a patch-level pretraining stage to mimic patch-wise mirror reflection and then learn the spatial correlation between the original object patches and the corresponding mirrored object patches; and 3) a pixel-level pre-training stage to extract the pixel-to-pixel relationship of mirror reflection by image reconstruction; Under such progressive pre-training scheme, our SSL pre-training framework can learn the representation of mirror reflection, and then effectively transfer this knowledge to the subsequent mirror detection process for better detection performances. We conduct comprehensive experiments to demonstrate the effectiveness of our SSL pretraining framework. We show that it can significantly boost the performances of existing mirror detection methods, and outperform other CNN-based SSL pre-training frameworks on the mirror detection task." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.817, + 0.457, + 0.832 + ], + "angle": 0, + "content": "To conclude, this paper makes three key contributions:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- To the best of our knowledge, we are the first to investigate how existing SSL pre-training frameworks perform on the mirror detection task, compared with supervised ImageNet pre-training." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "- We propose a new SSL pre-training framework that consists of three stages at different levels to progressively learn the representation of mirror reflection. Compared with the features from a supervised ImageNet pre-trained model, our representation is better due to the reduced gap between the pre-training task and the target downstream task (i.e., mirror detection)." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.21, + 0.892, + 0.271 + ], + "angle": 0, + "content": "- Extensive experiments show that our SSL pre-training framework performs the best among all state-of-the-art SSL methods, and even better than models with supervised ImageNet pre-training." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.092, + 0.892, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.286, + 0.642, + 0.302 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.311, + 0.668, + 0.326 + ], + "angle": 0, + "content": "2.1. Mirror Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.336, + 0.892, + 0.486 + ], + "angle": 0, + "content": "Mirror detection is an essential task in computer vision, as mirrors are everywhere these days. It is also a challenging task, since mirrors do not possess a consistent appearance, but reflect those of their surroundings. Other vision tasks such as generic object detection and segmentation would tend to detect and segment objects outside as well as inside of the mirror, providing incorrect information on the mirror region, and a depth prediction task may return the depths of the reflected objects inside the mirror region instead of the depth of the mirror." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.637 + ], + "angle": 0, + "content": "To address the mirror detection problem, Yang et al. [37] propose the first large-scale mirror detection dataset and the first deep-learning model, MirrorNet, for mirror detection by learning the contextual contrast between the inside and outside of mirrors. Lin et al. [23] further propose a new model with similarity and edge learning to progressively detect mirrors. Most recently, Guan et al. [15] employ graph convolutional networks to model contextual associations for mirror detection, while Tan et al. propose VCNet [30] to exploit the visual chirality cue at a feature level." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.892, + 0.759 + ], + "angle": 0, + "content": "However, existing mirror detection methods are all fully-supervised. They also directly adopt the supervised ImageNet [8] pre-training to obtain generic image features. Unlike recent works [19, 18, 22] that propose novel network architectures for mirror detection, we propose in this paper to tackle the mirror detection problem from a learning-based perspective, offering the first self-supervised pre-training framework for mirror detection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.776, + 0.787 + ], + "angle": 0, + "content": "2.2. Self-supervised Learning (SSL)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "SSL is a popular research topic in computer vision [38, 39], and has attracted a lot of research interests. It aims to learn good feature representations without needing labeled data. Early SSL methods focus on proposing novel pretext tasks, such as relative location prediction [10] and rotation prediction [21]. Recently, most methods are based on the utilization of contrastive learning [16, 4, 14, 3, 32, 5, 41," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12228" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.46, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.089, + 0.891, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.164, + 0.184, + 0.835, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.347, + 0.895, + 0.391 + ], + "angle": 0, + "content": "Figure 2. Our SSL pre-training framework. It consists of three pre-training stages at different levels: an image-level pre-training stage, a patch-level pre-training stage, and a pixel-level pre-training stage. It pre-trains the backbone network from global to local (image-level \\(\\rightarrow\\) patch-level \\(\\rightarrow\\) pixel-level) progressively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.47, + 0.52 + ], + "angle": 0, + "content": "40]. Most recently, [44] shows that some SSL methods perform even better than supervised ImageNet pre-training in downstream tasks, such as object detection [11, 24] and semantic segmentation [7, 34], with the help of instance discrimination during SSL pre-training. Recent transformer-based SSL frameworks [36] also involve masked image modeling to obtain better self-supervised representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.471, + 0.643 + ], + "angle": 0, + "content": "In this paper, we observe that unlike general vision tasks, mirror detection is not benefited from supervised ImageNet pre-training since it does not require category-level object understanding. Hence, we propose a novel SSL pre-training framework specifically for the mirror detection task. The proposed framework aims to simulate and learn the relationship between mirrors and their surrounding during the pre-training stage." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.657, + 0.169, + 0.673 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.47, + 0.806 + ], + "angle": 0, + "content": "Figure 2 shows the overall architecture of the proposed SSL pre-training framework. Our SSL pre-training framework consists of three pre-training stages at different levels: an image-level pre-training, a patch-level pre-training, and a pixel-level pre-training. The pre-training at each level is conducted sequentially and independently. The progressive and sequential training strategy can help the pre-trained model obtain a global-to-local representation of mirrors." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.817, + 0.303, + 0.833 + ], + "angle": 0, + "content": "3.1. Image-level Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Inspired by the definition of visual chirality [26] and its potential applications [46, 30] in various computer vision problems, we formulate the image-level pre-training as a binary classification problem: Given an input image and its" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.893, + 0.55 + ], + "angle": 0, + "content": "corresponding horizontal flipped image, can the backbone network tell which one is the original image? To achieve this objective, we attach a classification head to predict the probability that the input image is flipped. The classification head consists of a convolution layer followed by a sigmoid activation function to project the output value into \\([0,1]\\). In this pre-training stage, we use binary cross-entropy (BCE) as our loss function \\(\\mathcal{L}_f\\) to optimize the backbone network. Formally, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.558, + 0.892, + 0.576 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f} = - \\left[ y \\log (\\hat {y}) + (1 - y) \\log (1 - \\hat {y}) \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.614 + ], + "angle": 0, + "content": "where \\(y\\) is the ground truth label and \\(\\hat{y}\\) is the predicted probability that the input image is horizontally flipped." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.621, + 0.722, + 0.637 + ], + "angle": 0, + "content": "3.2. Patch-level Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.893, + 0.902 + ], + "angle": 0, + "content": "The image-level pre-training pretext task alone cannot bring much help in locating the mirror inside an image since mirror detection is a pixel-level problem. However, directly connecting the image-level and the pixel-level pretext tasks is challenging due to the gap between the scales of these two pretext tasks. Thus, we design a patch-level pre-training stage as a bridge to cover the feature representations of mirrors at different scales. Besides, the image-level pre-training does not benefit relational spatial understanding, which has been proven important in mirror detection [23]. Our patch-level pre-training consists of two sub-tasks to both obtain the patch-wise relationship of mirror reflection and simulate the mirrored region with pseudo labels. We first split the input image into a grid of \\( 3 \\times 3 \\) image patches and randomly shuffle them. After that, we randomly apply a flipping operation to one of the image patches, to simulate a mirrored region. The first sub-task" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12229" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.213 + ], + "angle": 0, + "content": "is to predict the location index from the shuffled grid by the backbone network. We append a location predictor head to the backbone network to predict the location index of input patches. We apply cross-entropy loss as \\(\\mathcal{L}_{loc}\\) to optimize this sub-task. The second sub-task is to predict the flipped patch from the nine randomly shuffled image patches. Similar to our image-level pre-training, we use BCE loss as our loss function \\(\\mathcal{L}_{cls}\\) in this sub-task." + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.224, + 0.47, + 0.265 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l o c} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} y _ {i} \\log \\left(\\hat {y} _ {i}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.269, + 0.47, + 0.311 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} = - \\frac {1}{N} \\sum_ {j = 1} ^ {N} y _ {j} \\log \\left(\\hat {y _ {j}}\\right) + (1 - y _ {j}) \\log \\left(1 - \\hat {y _ {j}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.47, + 0.399 + ], + "angle": 0, + "content": "where \\(\\hat{y}_i\\) is the predicted location index, and \\(y_i\\) is the corresponding ground truth label. \\(\\hat{y}_j\\) is the predicted probability that the input patch is horizontally flipped and \\(y_j\\) is the corresponding ground truth probability. Finally, we jointly optimize these two sub-tasks and the final loss function is:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.426, + 0.47, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p a t c h}} = \\mathcal {L} _ {\\text {l o c}} + \\mathcal {L} _ {\\text {c l s}}. \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.454, + 0.294, + 0.471 + ], + "angle": 0, + "content": "3.3. Pixel-level Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.473, + 0.721 + ], + "angle": 0, + "content": "Extracting the pixel-level relationship of mirror reflection is a challenging task, as studied by previous related work [12] for symmetry detection. It requires point-level GT labels to represent symmetry centers. However, obtaining such labels requires huge human annotation effort and they are not applicable for SSL pre-training. Instead of generating pseudo labels for symmetry detection, our solution is that we formulate this task as an image reconstruction problem on images containing mirrors. An ideal pre-trained network should be able to reconstruct the input image containing mirrors with the pixel-level understanding of mirror reflection. We randomly mask out the input image by cutout augmentation [9] with the ratio of 0.3 and then feed it to the backbone network with a reconstruction head. The details of the reconstruction head we used are listed in Table 1. The objective of our pixel-level pre-training is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.733, + 0.47, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p i x} = \\| x - \\hat {x} \\| _ {1}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.762, + 0.47, + 0.793 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{pix}\\) is an L1 loss. \\(x\\) and \\(\\hat{x}\\) are the input image and the reconstructed image, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.807, + 0.21, + 0.824 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.285, + 0.849 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Dataset and Evaluation Metrics. The image-level and patch-level pre-training experiments are conducted on the training set of ImageNet-100 [28] and MS COCO [25]," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.087, + 0.894, + 0.171 + ], + "angle": 0, + "content": "Table 1. The architecture of the reconstruction head used in the pixel-level pre-training. Note that each \"conv-IR\" corresponds to a sequence of convolution layer, InstanceNorm layer and ReLU activation. \\( K \\), \\( S \\) and \\( P \\) denote the number of kernels, the number of strides and the padding size, respectively, used in the convolution layer." + }, + { + "type": "table", + "bbox": [ + 0.559, + 0.172, + 0.835, + 0.368 + ], + "angle": 0, + "content": "
Layer Details
3 × 3 conv-IR, K = 1024, S = 1, P = 1
2 × Upsample
3 × 3 conv-IR, K = 512, S = 1, P = 1
2 × Upsample
3 × 3 conv-IR, K = 256, S = 1, P = 1
2 × Upsample
3 × 3 conv-IR, K = 128, S = 1, P = 1
2 × Upsample
3 × 3 conv-IR, K = 128, S = 1, P = 1
2 × Upsample
1 × 1 conv-IR, K = 3, S = 1, P = 1
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.394, + 0.892, + 0.612 + ], + "angle": 0, + "content": "which contains \\(130\\mathrm{K}\\) and \\(\\sim 118\\mathrm{k}\\) images respectively. In the pixel-level pre-training and fine-tuning stage, our experiments are conducted on MSD [37], the first large-scale dataset for mirror detection. MSD [37] provides 4,018 mirror images, which are divided into 3,063 images for training and 955 images for testing. We adopt two popular metrics namely, F-measure \\((\\mathbf{F}_{\\beta})\\) and mean absolute error (MAE) to evaluate the performance of our SSL pre-training framework. F-measure can reflect the overall model performance and is calculated by the weighted harmonic mean of the precision and recall: \\(F_{\\beta} = \\frac{(1 + \\beta^{2})\\cdot\\text{precision}\\cdot\\text{recall}}{\\beta^{2}\\cdot\\text{precision} + \\text{recall}}\\), where \\(\\beta^2\\) is set to 0.3. MAE can measure the average pixel-wise absolute disparity between the predicted mirror maps \\(M\\) and the binary ground truth \\(G\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.893, + 0.84 + ], + "angle": 0, + "content": "Implementation Details. We use Pytorch to implement our proposed SSL pre-training framework. All experiments are conducted on a GPU server with eight NVIDIA RTX 3090 GPUs. Following existing SSL pre-training frameworks, we adopt ResNet-50 [17] as our backbone network. All pre-trained model weights of existing SSL pretraining frameworks are obtained from the released models of mmselfsup [6] for fair comparisons. For each pretext task in the pre-training stage, all images are resized to \\(256 \\times 256\\). The batch size is set to 352. We use AdamW [27] with an initial learning rate of 1e-4 to optimize our SSL pre-training framework and stop pre-training after 20,000 iterations. We only adopt random color jittering for data augmentation. In the fine-tuning stage, we use the same hyperparameters reported in the original papers of the adopted" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.851, + 0.894, + 0.901 + ], + "angle": 0, + "content": "1It is worth noting that the datasets that we used (about 250K training images) are much smaller than the pre-trained dataset ImageNet (14 million images) used by other baseline SSL methods. We adopted ImageNet-100 to pre-train our network due to the limited computational resources." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12230" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.088, + 0.891, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.605, + 0.893, + 0.634 + ], + "angle": 0, + "content": "Figure 3. Qualitative Comparison of our proposed method with NPID [33], MocoV2 [4], SwAV [3], Barlow Twins [41], ImageNet pretraining (Supervised). We use the released pre-trained models of these existing SSL methods from mmselfsup [6] for fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.658, + 0.385, + 0.673 + ], + "angle": 0, + "content": "base methods MirrorNet [37] and VCNet [30]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.687, + 0.383, + 0.703 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.711, + 0.47, + 0.848 + ], + "angle": 0, + "content": "We compare against 11 state-of-the-art CNN-based frameworks in self-supervised learning, including Relative Location (Rel. Loc.) [10], Rotation Prediction (Rot. Pred.) [21], NPID [33], ODC [43], Moco V2 [16], SimCLR [4], BYOL [14], SwAV [3], DenseCL [32], SimSiam [5] and Barlow Twins (Bar. Twins) [41]. We select MirrorNet [37] and VCNet [30], which are two representative mirror detection methods, as base methods to conduct experiments in the fine-tuning stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Qualitative Evaluation. We further show the performance of our SSL pre-training framework in Figure 3. We compare our framework with four state-of-the-art SSL pre-training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.84 + ], + "angle": 0, + "content": "frameworks (including the two best-performing frameworks NPID [33] and Moco V2 [16] based on MirrorNet and the two best-performing frameworks SwAV [3] and Barlow Twins [41] based on VCNet, according to their performance in Table 2) and supervised ImageNet pre-training due to the limitation of space. Our framework can correctly detect the mirror regions by utilizing the relationship of mirror reflection caused by mirrors, even though the reflected regions are pure texture (i.e., lack of object keypoints in the second row) and relatively small in the image (e.g., the last three rows). We attribute the superior performance of our framework to the utilization of mirror reflection in the pre-training stage." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Quantitative Evaluation. Table 2 shows the experimentation results comparing the proposed SSL pre-training framework with the state-of-the-arts. Besides, we also re" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12231" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.47, + 0.143 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison between our SSL pre-training framework and existing CNN-based SSL methods. We also compare it with random initialization (Random) and ImageNet pretraining (Supervised)." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.143, + 0.491, + 0.412 + ], + "angle": 0, + "content": "
MethodsVenueMirrorNetVCNet
Fβ↑MAE↓Fβ↑MAE↓
Rel. Loc.ICCV '150.5820.2540.8420.072
Rot. Pred.ICLR '180.5370.2120.8510.069
NPIDCVPR '180.7390.1320.8320.076
ODCCVPR '200.5580.2230.8360.077
Moco V2CVPR '200.7070.1770.8350.072
SimCLRICML '200.5940.2440.8370.071
BYOLNeurIPS '200.7040.1580.8320.077
SwAVNeurIPS '200.5780.2240.8560.071
DenseCLCVPR '210.6940.2400.8440.072
SimSiamCVPR '210.6720.1980.8180.080
Bar. TwinsICML '210.6550.1780.8570.069
Random0.5860.2000.4000.103
Supervised20.7270.1700.8710.062
Ours0.7630.1160.8860.057
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.438, + 0.47, + 0.558 + ], + "angle": 0, + "content": "port the results from models adopting Xavier initialization [13] (Random) or supervised ImageNet pre-training (Supervised). Our SSL pre-training framework (Ours) achieves the best performance across both metrics when applied on MirrorNet and VCNet and shows a good generalization of different mirror detection methods. It is worth noting that our framework even outperforms the supervised ImageNet pre-training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.568, + 0.23, + 0.584 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.47, + 0.637 + ], + "angle": 0, + "content": "Table 3 demonstrates the effectiveness of each pretraining stage in our framework. We analyze its results point by point as follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.794 + ], + "angle": 0, + "content": "Image-level Pre-training vs. Random Initialization. To evaluate the effectiveness of the proposed image-level pretraining, we compare the model only adopting image-level pre-training (B2) with the model initialized randomly without any pre-training (B1). We can see that only adopting image-level pre-training (B2) will even cause a performance drop compared with the one without pre-training (B1). This somehow supports the finding from VCNet [30] that image-level flipping cannot provide sufficient information to pixelwisely locate the mirror inside an image." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.088, + 0.894, + 0.142 + ], + "angle": 0, + "content": "Table 3. Ablation study of the proposed SSL pre-training framework. We use MirrorNet [37] trained on the MSD dataset as our base method to conduct the experiments in this ablation study. Best results are shown in bold." + }, + { + "type": "table", + "bbox": [ + 0.5, + 0.142, + 0.9, + 0.281 + ], + "angle": 0, + "content": "
Rand. Init.Image-lev. PretrainPatch-lev. PretrainPixel-lev. PretrainFβ↑MAE↓
B1✓0.6010.195
B2✓0.6230.191
B3✓0.6890.162
B4✓0.6200.198
B5✓✓0.7230.166
B6✓✓0.7580.124
B7✓✓0.6660.178
Ours✓✓✓0.7630.116
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.313, + 0.894, + 0.466 + ], + "angle": 0, + "content": "Patch-level Pre-training vs. Random Initialization. We compare our patch-level pre-training (B3) with random initialization (B1) to analyze if patch-level pre-training is useful in our SSL pre-training framework. Our patch-level pretraining outperforms the model without pre-training on both two metrics (14.64% improvement on \\( F_{\\beta} \\) and 16.32% improvement on MAE), which indicates the effectiveness of our patch-level pre-training. It also demonstrates the importance of modeling the spatial relationship of mirror reflection at a patch level." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.474, + 0.892, + 0.61 + ], + "angle": 0, + "content": "Pixel-level Pre-training vs. Random Initialization. Similarly, we conduct the ablation experiment on our pixel-level pre-training. The model adopting pixel-level pre-training performs better than the model with image pre-training. However, the performance gain is not as much as the model with patch-level pre-training (B3). One possible reason is that directly optimizing the backbone network for the representation of mirrors at a pixel level is challenging and may not produce satisfactory results for mirror detection." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.619, + 0.892, + 0.74 + ], + "angle": 0, + "content": "Location Predictor in Patch-level Pre-training. The location predictor in our patch-level pre-training is designed to capture the relational spatial information, which has been proven important in mirror detection [23]. We conduct an ablation experiment to remove the location predictor and the performance of our methods drops from 0.763 to 0.731. This shows the importance of learning spatial information in the mirror detection task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Combinations of Different Pre-training Stages. Based on the above three comparisons, we find that the patch-level pre-training contributes the most improvement. It demonstrates the importance of modeling the spatial relationship of mirror reflection at a patch level. We then analyze the results from the model with different combinations of our pretraining stages. The model with image-level pre-training and patch-level pre-training (B5) performs better than the model that only adopts patch-level pre-training (B3), which shows that the global information provided by image-level" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.803, + 0.47, + 0.901 + ], + "angle": 0, + "content": "2Note that the original backbone networks reported in their papers are ResNext-101 [35]. However, the model weights of ResNext-101 pre-trained by our baseline SSL frameworks are not available in mmselfsup [6]. Thus, we switch the backbone network of these two methods to ResNet-50 [17], which is a common backbone network used in evaluating the performances of SSL frameworks. We also directly use the released ResNet-50 SSL pre-trained weights in our experiments for fair comparisons." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12232" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.104, + 0.216, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.217, + 0.179, + 0.23 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.104, + 0.324, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.211, + 0.304, + 0.235 + ], + "angle": 0, + "content": "Image-level Pretraining" + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.104, + 0.434, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.211, + 0.41, + 0.235 + ], + "angle": 0, + "content": "Patch-level Pretraining" + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.104, + 0.542, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.211, + 0.517, + 0.235 + ], + "angle": 0, + "content": "Pixel-level Pretraining" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.104, + 0.65, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.211, + 0.65, + 0.235 + ], + "angle": 0, + "content": "(Image+Patch)-level Pretraining" + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.104, + 0.758, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.218, + 0.719, + 0.228 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.104, + 0.868, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.803, + 0.218, + 0.823, + 0.227 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.247, + 0.699, + 0.262 + ], + "angle": 0, + "content": "Figure 4. Visualization of the effects of the proposed pre-training stages." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.286, + 0.473, + 0.543 + ], + "angle": 0, + "content": "pre-training can benefit the model with patch-level pretraining. Similarly, as shown in B7, B4 (i.e., only pixel-level pretraining) can boost the performance of the pre-trained model after image-level pre-training. These two comparisons indicate that image-level pre-training should be used with other pre-training stages to benefit the overall performance and it should not be solely adopted to our SSL pretraining framework. Besides, we find that combining patch-level and pixel-level pre-training (B6) has superior performance when compared with B3, especially on MAE with about \\(23.4\\%\\) improvement. This combination also outperforms supervised ImageNet pre-training, according to the results from Table 2. We attribute its superior performance to the pixel-level representation of mirror reflection. Our final model with all pre-training stages outperforms other ablation models, showing the effectiveness of adopting pretraining stages progressively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.47, + 0.807 + ], + "angle": 0, + "content": "Visual Comparison of Our Pre-training Stages. Figure 4 shows a visual example of our ablation study. We can see that only adopting image-level pre-training on our SSL pre-training framework can approximately locate the mirror region but fails to produce a precise boundary for the predicted mask of the mirror. Patch-level pre-training can help predict mirrors more accurately with clearer boundaries but over-predict the non-mirror regions as mirrors. In comparison to predictions from all other ablation models, the prediction from the model with pixel-level pretraining has the clearest boundary but contains the largest over-predicted regions, primarily because of the insufficient global-level learning on mirrors. Our framework can significantly reduce over-predictions when image-level and patch-level pre-training are used, as opposed to when patch-level pre-training is used only, and performs the best with the help of the global-to-local pre-training process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.203, + 0.832 + ], + "angle": 0, + "content": "4.4. Discussions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Impacts of Pre-training Strategy. Some previous SSL works [42, 45], especially for those focusing on self-supervised pre-training for downstream tasks like segmentation [42] and salient object detection [45], would adopt" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.282, + 0.892, + 0.309 + ], + "angle": 0, + "content": "Table 4. The results of MirrorNet [37] when using different pretraining strategies. Best results are shown in bold." + }, + { + "type": "table", + "bbox": [ + 0.564, + 0.309, + 0.83, + 0.407 + ], + "angle": 0, + "content": "
Fβ↑MAE↓
Random Initialized0.5860.200
ImageNet Supervised0.7270.170
Ours (In parallel)0.7080.331
Ours (Sequential)0.7630.116
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.436, + 0.892, + 0.679 + ], + "angle": 0, + "content": "different pretext tasks sequentially to pre-train their backbone networks. We also adopt this common strategy in our proposed SSL framework. To verify if employing different pre-training strategies may affect the performance of the proposed SSL framework, we have tried training the three stages sequentially as well as in parallel. We find that when pre-training the three stages in parallel, the backbone network fails to converge and its final fine-tuning results are much worse than those using sequential pre-training. Table 4 shows the results of this experiment. We can see that the MAE performance of pre-training the different stages in parallel is even worse than that of random initialization. A possible reason for this is that directly pre-training different pretext tasks in parallel is difficult to transfer the learned knowledge from the pre-trained network to the target downstream task (i.e., pixel-level mirror detection)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.681, + 0.892, + 0.862 + ], + "angle": 0, + "content": "Apart from the superior performance made by the sequential pre-training strategy, we also notice that the number of pre-training iterations for each pretext task can heavily affect the performance of our SSL framework. To avoid heavy hyperparameter tuning and increase the reproducibility of our pre-training framework, we adopt the same number of pre-training iterations for all stages. Although this strategy of using the same number of pre-training iterations is likely not the optimal design for individual stages, we believe that it can better reflect the real performance of our framework with less stochasticity and more reactivity. We provide a more detailed discussion on this issue next." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Impacts of Pre-training Iterations. SSL pre-training frameworks usually require huge pre-training iterations to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12233" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.091, + 0.411, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.282, + 0.47, + 0.351 + ], + "angle": 0, + "content": "Figure 5. Ablation study on the number of pre-training iterations for our SSL pre-training framework. We find that our SSL pretraining framework continues to improve in performance until \\(\\sim 20,000\\) pre-training iterations. After that, the performance starts to decline, mainly due to overfitting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.373, + 0.468, + 0.509 + ], + "angle": 0, + "content": "converge. For example, SimSiam [5] requires about 250K iterations to pre-train their model on ImageNet, which is very time-consuming. To validate how this issue may affect our SSL pre-training framework, we conduct the ablation study on the number of pre-training iterations for our SSL pre-training framework, and the results are shown in Figure 5. Note that we use the same number of iterations for all pre-training stages for convenience and to get rid of heavy hyperparameter tuning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.468, + 0.786 + ], + "angle": 0, + "content": "From the results, we find that the performance of our SSL pre-training framework improves rapidly as the number of pre-training iterations increases, up to around 20,000, which is a relatively small number compared with those used in existing SSL pre-training frameworks. The performance of our SSL pre-training framework gradually drops as the number of pre-training iterations exceeds 20,000, mainly due to overfitting. One possible explanation of this is that the model pre-trained by our SSL pre-training framework can efficiently learn the representation of mirror reflection under a short pre-training period. Another possible explanation is that unlike generic SSL pre-training frameworks that attempt to extract general-propose image representation, our SSL pre-training framework is developed specially for a specific task, i.e., mirror detection. While generic SSL pre-training may cause feature redundancy for our task, having too many pre-training iterations (i.e., over 20,000) causes overfitting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Failure Cases. Our SSL pre-training framework does have limitations. It continues to struggle with the limitations of the base method [37], as the base method adopted can also significantly affect the final detection performances. For example, as shown in first row of Figure 6, MirrorNet with our SSL pre-training may still fail to handle some mirror-like regions (e.g., the top-right region). It may also fail to de" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.088, + 0.637, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.089, + 0.761, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.089, + 0.885, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.168, + 0.636, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.293, + 0.59, + 0.306 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.168, + 0.761, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.293, + 0.712, + 0.305 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.168, + 0.885, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.814, + 0.293, + 0.834, + 0.304 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.318, + 0.892, + 0.415 + ], + "angle": 0, + "content": "Figure 6. Failure cases of our SSL pre-training framework based on MirrorNet. Our SSL pre-training framework fails to tackle the intrinsic limitation of existing mirror detection methods. The model pre-trained by our SSL pre-training framework still has wrong predictions when the input images are challenging (e.g., containing mirror-like regions in the first row, and insufficient relationship between mirror/non-mirror regions in the second row." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.443, + 0.892, + 0.504 + ], + "angle": 0, + "content": "tect mirrors in some challenging cases when the input image contains ambiguous mirrors without the relationship of mirror reflection and sufficient contextual contrast, as shown in the second row of Figure 6." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.522, + 0.618, + 0.537 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.836 + ], + "angle": 0, + "content": "In this paper, we have investigated how self-supervised learning (SSL) works with the mirror detection task. To the best of our knowledge, we are the first to explore SSL pre-training frameworks applied to mirror detection. We have found that the supervised ImageNet pre-training might not be the ideal way to extract backbone image features for mirror detection due to the discrepancy between general-propose representation and mirror-specific representation. We have also proposed a new SSL pre-training framework to pre-train the backbone network for mirror detection. Our SSL pre-training framework does not require any labeled data. It progressively (image-, patch- and pixel-level pre-training) models the relationship of mirror reflection, and obtains a better representation of mirrors in the pre-training stage. Experimental results show that the proposed SSL pre-training framework outperforms CNN-based state-of-the-art SSL pre-training frameworks, and even achieves better results compared with supervised ImageNet pre-training." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgments: The work described in this paper was partially supported by a GRF grant from the Research Grants Council of Hong Kong (Project No.: CityU 11211223)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12234" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.47, + 0.186 + ], + "angle": 0, + "content": "[1] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, pages 3674-3683, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.188, + 0.47, + 0.231 + ], + "angle": 0, + "content": "[2] M. Braun, S. Krebs, F. Flohr, and D. M. Gavrila. Eurocity persons: A novel benchmark for person detection in traffic scenes. IEEE TPAMI, 41(8):1844-1861, Aug 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.233, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.291, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.334, + 0.47, + 0.363 + ], + "angle": 0, + "content": "[5] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.365, + 0.47, + 0.407 + ], + "angle": 0, + "content": "[6] MMSelfSup Contributors. MMSelfSup: Openmmlab self-supervised learning toolbox and benchmark. https://github.com/open-mmlab/mmselfsup, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.409, + 0.47, + 0.478 + ], + "angle": 0, + "content": "[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.47, + 0.523 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.525, + 0.47, + 0.567 + ], + "angle": 0, + "content": "[9] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.569, + 0.469, + 0.61 + ], + "angle": 0, + "content": "[10] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In ICCV, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.47, + 0.669 + ], + "angle": 0, + "content": "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.469, + 0.713 + ], + "angle": 0, + "content": "[12] Christopher Funk and Yanxi Liu. Beyond planar symmetry: Modeling human perception of reflection and rotation symmetries in the wild. In ICCV, pages 793-803, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.715, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[13] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249–256. JMLR Workshop and Conference Proceedings, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Huankang Guan, Jiaying Lin, and Rynson W.H. Lau. Learning semantic associations for mirror detection. In CVPR, pages 5941-5950, June 2022." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[18] Ruozhen He, Jiaying Lin, and Rynson WH Lau. Efficient mirror detection via multi-level heterogeneous learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 790-798, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[19] Tianyu Huang, Bowen Dong, Jiaying Lin, Xiaohui Liu, Rynson WH Lau, and Wangmeng Zuo. Symmetry-aware transformer-based mirror detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 935–943, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[20] Peter Kohler, Alexandra Yakovleva, Alasdair Clarke, Yanxi Liu, and Anthony Norcia. Parametric responses to rotation symmetry in mid-level visual cortex. Journal of Vision, 15(12):1122-1122, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.361, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[21] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In ICLR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[22] Jiaying Lin, Xin Tan, and Rynson WH Lau. Learning to detect mirrors from videos via dual correspondences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9109-9118, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[23] Jiaying Lin, Guodong Wang, and Rynson W. H. Lau. Progressive mirror detection. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.603, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[26] Zhiqiu Lin, Jin Sun, Abe Davis, and Noah Snavely. Visual chirality. In CVPR, pages 12295-12303, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[28] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[29] Jiaqi Tan, Weijie Lin, Angel X Chang, and Manolis Savva. Mirror3d: Depth refinement for mirror surfaces. In CVPR, pages 15990-15999, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[30] Xin Tan, Jiaying Lin, Ke Xu, Chen Pan, Lizhuang Ma, and Rynson W. H. Lau. Mirror detection with the visual chirality cue. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[31] Christopher W Tyler. Human symmetry perception and its computational analysis. Psychology Press, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In CVPR, 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12235" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.133 + ], + "angle": 0, + "content": "[33] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.177 + ], + "angle": 0, + "content": "[34] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, pages 418-434, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.468, + 0.22 + ], + "angle": 0, + "content": "[35] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, pages 1492-1500, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.222, + 0.468, + 0.275 + ], + "angle": 0, + "content": "[36] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, pages 9653-9663, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.278, + 0.468, + 0.305 + ], + "angle": 0, + "content": "[37] Xin Yang, Haiyang Mei, Ke Xu, Xiaopeng Wei, Baocai Yin, and Rynson W.H. Lau. Where is my mirror? In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.307, + 0.468, + 0.348 + ], + "angle": 0, + "content": "[38] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. Learning with noisy labels for robust point cloud segmentation. In ICCV, pages 6443-6452, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.35, + 0.468, + 0.404 + ], + "angle": 0, + "content": "[39] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. Robust point cloud segmentation with noisy annotations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7696-7710, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.407, + 0.468, + 0.46 + ], + "angle": 0, + "content": "[40] Shuquan Ye, Yujia Xie, Dongdong Chen, Yichong Xu, Lu Yuan, Chenguang Zhu, and Jing Liao. Improving commonsense in vision-language models via knowledge graph riddles. In CVPR, pages 2634-2645, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.463, + 0.468, + 0.503 + ], + "angle": 0, + "content": "[41] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.506, + 0.468, + 0.546 + ], + "angle": 0, + "content": "[42] Xiaohang Zhan, Ziwei Liu, Ping Luo, Xiaou Tang, and Chen Loy. Mix-and-match tuning for self-supervised semantic segmentation. In AAAI, volume 32, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.549, + 0.468, + 0.589 + ], + "angle": 0, + "content": "[43] Xiaohang Zhan, Jiahao Xie, Ziwei Liu, Yew-Soon Ong, and Chen Change Loy. Online deep clustering for unsupervised representation learning. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.592, + 0.468, + 0.632 + ], + "angle": 0, + "content": "[44] Nanxuan Zhao, Zhirong Wu, Rynson WH Lau, and Stephen Lin. What makes instance discrimination good for transfer learning? In ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.635, + 0.468, + 0.674 + ], + "angle": 0, + "content": "[45] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, , Hutchuan Lu, and Xiang Ruan. Self-supervised pretraining for rgb-d salient object detection. In AAAI, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.677, + 0.468, + 0.731 + ], + "angle": 0, + "content": "[46] Ying Zheng, Yiyi Zhang, Xiaogang Xu, Jun Wang, and Hongxun Yao. Visual chirality meets freehand sketches. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1544-1548. IEEE, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12236" + } + ] +] \ No newline at end of file diff --git a/2023/Self-supervised Pre-training for Mirror Detection/9faec277-19c4-4583-a3c8-84bf53084f2f_origin.pdf b/2023/Self-supervised Pre-training for Mirror Detection/9faec277-19c4-4583-a3c8-84bf53084f2f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..24694d9a3cc103a2305eb6df4f42eccc456f9a7e --- /dev/null +++ b/2023/Self-supervised Pre-training for Mirror Detection/9faec277-19c4-4583-a3c8-84bf53084f2f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a188edcbb1d8c7cbd200d7f57d0fd77c56d829b51bf4014d19bce8299d48cedb +size 1583664 diff --git a/2023/Self-supervised Pre-training for Mirror Detection/full.md b/2023/Self-supervised Pre-training for Mirror Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fa3f4d7cc829f03c5fb1acadf49f8303a350b006 --- /dev/null +++ b/2023/Self-supervised Pre-training for Mirror Detection/full.md @@ -0,0 +1,303 @@ +# Self-supervised Pre-training for Mirror Detection + +Jiaying Lin* Rynson W.H. Lau* City University of Hong Kong + +{jiayinlin5, rynson.lau}@cityu.edu.hk + +# Abstract + +Existing mirror detection methods require supervised ImageNet pre-training to obtain good general-purpose image features. However, supervised ImageNet pre-training focuses on category-level discrimination and may not be suitable for downstream tasks like mirror detection, due to the overfitting upstream tasks (e.g., supervised image classification). We observe that mirror reflection is crucial to how people perceive the presence of mirrors, and such mid-level features can be better transferred from self-supervised pretrained models. Inspired by this observation, in this paper we aim to improve mirror detection methods by proposing a new self-supervised learning (SSL) pre-training framework for modeling the representation of mirror reflection progressively in the pre-training process. Our framework consists of three pre-training stages at different levels: 1) an image-level pre-training stage to globally incorporate mirror reflection features into the pre-trained model; 2) a patch-level pre-training stage to spatially simulate and learn local mirror reflection from image patches; and 3) a pixel-level pre-training stage to pixel-wisely capture mirror reflection via reconstructing corrupted mirror images based on the relationship between the inside and outside of mirrors. Extensive experiments show that our SSL pre-training framework significantly outperforms previous state-of-the-art CNN-based SSL pre-training frameworks and even outperforms supervised ImageNet pre-training when transferred to the mirror detection task. Code and models are available at https://jiaying.link/iccv2023-sslmirror/ + +# 1. Introduction + +Mirrors are prevalent in our daily lives. As their appearances are largely determined by their surroundings, they generally lack a consistent appearance, making it difficult to separate them from their surroundings. This may affect many computer vision tasks such as object detection [2], vision-language navigation [1] and depth estimation [29]. + +![](images/2053fb7e33566aa3d281812f2faf69bdad556d6356cac893b22a3c6a21bc08bc.jpg) + +![](images/97624ca6d1125c6f4390b003b014011a351f9c1950c7ccbc8c0f5ed0ce450959.jpg) + +![](images/a3a1688cab67f58dac910e5e2accc57c4180feddaabc8e18cc4be9edbfbf4b69.jpg) +(b) MirrorNet on ImageNet + +![](images/9c79b20acd129e22e321256d41d612df587f83459c20148214c01a2b2a65136b.jpg) + +![](images/21f231f0b30b96f0819c87b4f82bdc5afed12d33bcd81d816967afb479a37ca7.jpg) +(c)VCNet on ImageNet + +![](images/2407085591717f6000629a100ebbac3458bd38d6891f0485e3353e069cd334da.jpg) + +![](images/1d85123e767780ec94c3702dff904630533d38df76179b68b715bd5920b81888.jpg) +Figure 1. State-of-the-art mirror detection methods [37, 30] are based on costly supervised ImageNet pre-training. They may fail even in obvious cases, e.g., (a) when the mirror clearly reflects a real object outside of the mirror. (b) and (c) are MirrorNet [37] and VCNet [30], respectively, pre-trained on ImageNet with full supervision. (c) is MirrorNet with our proposed SSL framework and without supervised ImageNet pre-training. Our SSL scheme leverages the mirror reflection cue and avoids feature redundancy in the pre-training stage. It outperforms those pre-trained on ImageNet with full supervision (i.e., (b) and (c)). + +![](images/172cb65753d1987a963d6fe5c399475d5ab9a638956f1ba00e1256d1d7e58382.jpg) + +![](images/db1d2d05c41cec15303148981a731908325a4eedbf1269610181f34c7e074c40.jpg) +1) Ours (MirrorNet on our SSL) +(e) GT + +It is therefore crucial to design high-performance mirror detection methods to facilitate computer vision applications. + +Recently, a few methods [37, 23, 30] have been proposed for mirror detection, but they all require to be initialized with supervised ImageNet [8] pre-trained weights and then fine-tuned on mirror detection datasets. While such a transfer learning approach is common in general computer vision tasks (e.g., object detection [24] and semantic segmentation [7]) to utilize category-oriented features from large-scale image datasets, we challenge if this strategy is necessary for the mirror detection task for two reasons. First, the supervised ImageNet pre-training process involves high labelling costs, as it is very time-consuming and labor-intensive to label such a large-scale image dataset ( $\sim$ 1.3M images). Second, unlike general vision tasks, mirror detection does not require category-level object understanding. Instead, it requires an understanding of the relationship between inside and outside of mirrors (e.g., contrast [37], similarity [23], and visual chirality [30]). Thus, + +while pre-training on ImageNet requires expensive labels, it causes feature redundancy for the mirror detection task. + +Figure 1 shows that existing models pre-trained on ImageNet tend to over-detect the mirror regions. Although the two SOTA methods [37, 30] can correctly locate the mirror regions, they are unable to distinguish between the mirrored objects inside the mirrors and the real objects outside, even though these methods contain well-designed modules to learn to address this problem via the fine-tuning stage. This motivates us to investigate whether the current pre-training approach may affect the final detection performance, and whether we may incorporate some intrinsic mirror properties in the pre-training process to assist mirror detection. + +An interesting observation is that humans do not need to be "trained" to recognize mirrors. Instead, we learn to recognize them implicitly from young. This aligns with the key idea of self-supervised learning (SSL). Works from neuroscience [31, 20] have shown that humans use mid-level visual cortex to recognize mirror reflection. According to [44], mid-level features are hard to be transferred from supervised pre-training on ImageNet, and SSL is suitable for learning mid-level representations, which inspires us to develop an SSL pre-training framework for modeling mirror reflection. + +In this paper, we propose a new SSL pre-training framework for mirror detection, which explicitly considers mirror reflection during the pre-training process. Our framework does not require human-annotated labels from large-scale image datasets. It consists of three stages to progressively pre-train the backbone network from global to local: 1) an image-level pre-training stage to obtain the representation of mirror reflection globally by recognizing the geometric transformation applied to the image; 2) a patch-level pretraining stage to mimic patch-wise mirror reflection and then learn the spatial correlation between the original object patches and the corresponding mirrored object patches; and 3) a pixel-level pre-training stage to extract the pixel-to-pixel relationship of mirror reflection by image reconstruction; Under such progressive pre-training scheme, our SSL pre-training framework can learn the representation of mirror reflection, and then effectively transfer this knowledge to the subsequent mirror detection process for better detection performances. We conduct comprehensive experiments to demonstrate the effectiveness of our SSL pretraining framework. We show that it can significantly boost the performances of existing mirror detection methods, and outperform other CNN-based SSL pre-training frameworks on the mirror detection task. + +To conclude, this paper makes three key contributions: + +- To the best of our knowledge, we are the first to investigate how existing SSL pre-training frameworks perform on the mirror detection task, compared with supervised ImageNet pre-training. + +- We propose a new SSL pre-training framework that consists of three stages at different levels to progressively learn the representation of mirror reflection. Compared with the features from a supervised ImageNet pre-trained model, our representation is better due to the reduced gap between the pre-training task and the target downstream task (i.e., mirror detection). +- Extensive experiments show that our SSL pre-training framework performs the best among all state-of-the-art SSL methods, and even better than models with supervised ImageNet pre-training. + +# 2. Related Work + +# 2.1. Mirror Detection + +Mirror detection is an essential task in computer vision, as mirrors are everywhere these days. It is also a challenging task, since mirrors do not possess a consistent appearance, but reflect those of their surroundings. Other vision tasks such as generic object detection and segmentation would tend to detect and segment objects outside as well as inside of the mirror, providing incorrect information on the mirror region, and a depth prediction task may return the depths of the reflected objects inside the mirror region instead of the depth of the mirror. + +To address the mirror detection problem, Yang et al. [37] propose the first large-scale mirror detection dataset and the first deep-learning model, MirrorNet, for mirror detection by learning the contextual contrast between the inside and outside of mirrors. Lin et al. [23] further propose a new model with similarity and edge learning to progressively detect mirrors. Most recently, Guan et al. [15] employ graph convolutional networks to model contextual associations for mirror detection, while Tan et al. propose VCNet [30] to exploit the visual chirality cue at a feature level. + +However, existing mirror detection methods are all fully-supervised. They also directly adopt the supervised ImageNet [8] pre-training to obtain generic image features. Unlike recent works [19, 18, 22] that propose novel network architectures for mirror detection, we propose in this paper to tackle the mirror detection problem from a learning-based perspective, offering the first self-supervised pre-training framework for mirror detection. + +# 2.2. Self-supervised Learning (SSL) + +SSL is a popular research topic in computer vision [38, 39], and has attracted a lot of research interests. It aims to learn good feature representations without needing labeled data. Early SSL methods focus on proposing novel pretext tasks, such as relative location prediction [10] and rotation prediction [21]. Recently, most methods are based on the utilization of contrastive learning [16, 4, 14, 3, 32, 5, 41, + +![](images/908689b51dc7553c0ccc26fde83f0bd65eb6d802e9a616f0f113c004d699d59b.jpg) + +![](images/855243fd694ad6ad2d7363759b0686b9465af05eab1878a3b52b8f769d8e9a32.jpg) + +![](images/d955a6d1249ed61c65caf126c7f5730a39a2581f3573b017e7bde9fc5f13c937.jpg) +Figure 2. Our SSL pre-training framework. It consists of three pre-training stages at different levels: an image-level pre-training stage, a patch-level pre-training stage, and a pixel-level pre-training stage. It pre-trains the backbone network from global to local (image-level $\rightarrow$ patch-level $\rightarrow$ pixel-level) progressively. + +40]. Most recently, [44] shows that some SSL methods perform even better than supervised ImageNet pre-training in downstream tasks, such as object detection [11, 24] and semantic segmentation [7, 34], with the help of instance discrimination during SSL pre-training. Recent transformer-based SSL frameworks [36] also involve masked image modeling to obtain better self-supervised representations. + +In this paper, we observe that unlike general vision tasks, mirror detection is not benefited from supervised ImageNet pre-training since it does not require category-level object understanding. Hence, we propose a novel SSL pre-training framework specifically for the mirror detection task. The proposed framework aims to simulate and learn the relationship between mirrors and their surrounding during the pre-training stage. + +# 3. Method + +Figure 2 shows the overall architecture of the proposed SSL pre-training framework. Our SSL pre-training framework consists of three pre-training stages at different levels: an image-level pre-training, a patch-level pre-training, and a pixel-level pre-training. The pre-training at each level is conducted sequentially and independently. The progressive and sequential training strategy can help the pre-trained model obtain a global-to-local representation of mirrors. + +# 3.1. Image-level Pre-training + +Inspired by the definition of visual chirality [26] and its potential applications [46, 30] in various computer vision problems, we formulate the image-level pre-training as a binary classification problem: Given an input image and its + +corresponding horizontal flipped image, can the backbone network tell which one is the original image? To achieve this objective, we attach a classification head to predict the probability that the input image is flipped. The classification head consists of a convolution layer followed by a sigmoid activation function to project the output value into $[0,1]$ . In this pre-training stage, we use binary cross-entropy (BCE) as our loss function $\mathcal{L}_f$ to optimize the backbone network. Formally, we have: + +$$ +\mathcal {L} _ {f} = - \left[ y \log (\hat {y}) + (1 - y) \log (1 - \hat {y}) \right], \tag {1} +$$ + +where $y$ is the ground truth label and $\hat{y}$ is the predicted probability that the input image is horizontally flipped. + +# 3.2. Patch-level Pre-training + +The image-level pre-training pretext task alone cannot bring much help in locating the mirror inside an image since mirror detection is a pixel-level problem. However, directly connecting the image-level and the pixel-level pretext tasks is challenging due to the gap between the scales of these two pretext tasks. Thus, we design a patch-level pre-training stage as a bridge to cover the feature representations of mirrors at different scales. Besides, the image-level pre-training does not benefit relational spatial understanding, which has been proven important in mirror detection [23]. Our patch-level pre-training consists of two sub-tasks to both obtain the patch-wise relationship of mirror reflection and simulate the mirrored region with pseudo labels. We first split the input image into a grid of $3 \times 3$ image patches and randomly shuffle them. After that, we randomly apply a flipping operation to one of the image patches, to simulate a mirrored region. The first sub-task + +is to predict the location index from the shuffled grid by the backbone network. We append a location predictor head to the backbone network to predict the location index of input patches. We apply cross-entropy loss as $\mathcal{L}_{loc}$ to optimize this sub-task. The second sub-task is to predict the flipped patch from the nine randomly shuffled image patches. Similar to our image-level pre-training, we use BCE loss as our loss function $\mathcal{L}_{cls}$ in this sub-task. + +$$ +\mathcal {L} _ {l o c} = - \frac {1}{N} \sum_ {i = 1} ^ {N} y _ {i} \log \left(\hat {y} _ {i}\right), \tag {2} +$$ + +$$ +\mathcal {L} _ {c l s} = - \frac {1}{N} \sum_ {j = 1} ^ {N} y _ {j} \log \left(\hat {y _ {j}}\right) + (1 - y _ {j}) \log \left(1 - \hat {y _ {j}}\right), \tag {3} +$$ + +where $\hat{y}_i$ is the predicted location index, and $y_i$ is the corresponding ground truth label. $\hat{y}_j$ is the predicted probability that the input patch is horizontally flipped and $y_j$ is the corresponding ground truth probability. Finally, we jointly optimize these two sub-tasks and the final loss function is: + +$$ +\mathcal {L} _ {\text {p a t c h}} = \mathcal {L} _ {\text {l o c}} + \mathcal {L} _ {\text {c l s}}. \tag {4} +$$ + +# 3.3. Pixel-level Pre-training + +Extracting the pixel-level relationship of mirror reflection is a challenging task, as studied by previous related work [12] for symmetry detection. It requires point-level GT labels to represent symmetry centers. However, obtaining such labels requires huge human annotation effort and they are not applicable for SSL pre-training. Instead of generating pseudo labels for symmetry detection, our solution is that we formulate this task as an image reconstruction problem on images containing mirrors. An ideal pre-trained network should be able to reconstruct the input image containing mirrors with the pixel-level understanding of mirror reflection. We randomly mask out the input image by cutout augmentation [9] with the ratio of 0.3 and then feed it to the backbone network with a reconstruction head. The details of the reconstruction head we used are listed in Table 1. The objective of our pixel-level pre-training is defined as: + +$$ +\mathcal {L} _ {p i x} = \| x - \hat {x} \| _ {1}, \tag {5} +$$ + +where $\mathcal{L}_{pix}$ is an L1 loss. $x$ and $\hat{x}$ are the input image and the reconstructed image, respectively. + +# 4. Experiments + +# 4.1. Experimental Settings + +Dataset and Evaluation Metrics. The image-level and patch-level pre-training experiments are conducted on the training set of ImageNet-100 [28] and MS COCO [25], + +Table 1. The architecture of the reconstruction head used in the pixel-level pre-training. Note that each "conv-IR" corresponds to a sequence of convolution layer, InstanceNorm layer and ReLU activation. $K$ , $S$ and $P$ denote the number of kernels, the number of strides and the padding size, respectively, used in the convolution layer. + +
Layer Details
3 × 3 conv-IR, K = 1024, S = 1, P = 1
2 × Upsample
3 × 3 conv-IR, K = 512, S = 1, P = 1
2 × Upsample
3 × 3 conv-IR, K = 256, S = 1, P = 1
2 × Upsample
3 × 3 conv-IR, K = 128, S = 1, P = 1
2 × Upsample
3 × 3 conv-IR, K = 128, S = 1, P = 1
2 × Upsample
1 × 1 conv-IR, K = 3, S = 1, P = 1
+ +which contains $130\mathrm{K}$ and $\sim 118\mathrm{k}$ images respectively. In the pixel-level pre-training and fine-tuning stage, our experiments are conducted on MSD [37], the first large-scale dataset for mirror detection. MSD [37] provides 4,018 mirror images, which are divided into 3,063 images for training and 955 images for testing. We adopt two popular metrics namely, F-measure $(\mathbf{F}_{\beta})$ and mean absolute error (MAE) to evaluate the performance of our SSL pre-training framework. F-measure can reflect the overall model performance and is calculated by the weighted harmonic mean of the precision and recall: $F_{\beta} = \frac{(1 + \beta^{2})\cdot\text{precision}\cdot\text{recall}}{\beta^{2}\cdot\text{precision} + \text{recall}}$ , where $\beta^2$ is set to 0.3. MAE can measure the average pixel-wise absolute disparity between the predicted mirror maps $M$ and the binary ground truth $G$ . + +Implementation Details. We use Pytorch to implement our proposed SSL pre-training framework. All experiments are conducted on a GPU server with eight NVIDIA RTX 3090 GPUs. Following existing SSL pre-training frameworks, we adopt ResNet-50 [17] as our backbone network. All pre-trained model weights of existing SSL pretraining frameworks are obtained from the released models of mmselfsup [6] for fair comparisons. For each pretext task in the pre-training stage, all images are resized to $256 \times 256$ . The batch size is set to 352. We use AdamW [27] with an initial learning rate of 1e-4 to optimize our SSL pre-training framework and stop pre-training after 20,000 iterations. We only adopt random color jittering for data augmentation. In the fine-tuning stage, we use the same hyperparameters reported in the original papers of the adopted + +![](images/fb1c2b369ed7aed8c0640a1bc2c105e7f93e01f49c8804770f246f9699221c5a.jpg) +Figure 3. Qualitative Comparison of our proposed method with NPID [33], MocoV2 [4], SwAV [3], Barlow Twins [41], ImageNet pretraining (Supervised). We use the released pre-trained models of these existing SSL methods from mmselfsup [6] for fair comparison. + +base methods MirrorNet [37] and VCNet [30]. + +# 4.2. Comparison with State-of-the-Arts + +We compare against 11 state-of-the-art CNN-based frameworks in self-supervised learning, including Relative Location (Rel. Loc.) [10], Rotation Prediction (Rot. Pred.) [21], NPID [33], ODC [43], Moco V2 [16], SimCLR [4], BYOL [14], SwAV [3], DenseCL [32], SimSiam [5] and Barlow Twins (Bar. Twins) [41]. We select MirrorNet [37] and VCNet [30], which are two representative mirror detection methods, as base methods to conduct experiments in the fine-tuning stage. + +Qualitative Evaluation. We further show the performance of our SSL pre-training framework in Figure 3. We compare our framework with four state-of-the-art SSL pre-training + +frameworks (including the two best-performing frameworks NPID [33] and Moco V2 [16] based on MirrorNet and the two best-performing frameworks SwAV [3] and Barlow Twins [41] based on VCNet, according to their performance in Table 2) and supervised ImageNet pre-training due to the limitation of space. Our framework can correctly detect the mirror regions by utilizing the relationship of mirror reflection caused by mirrors, even though the reflected regions are pure texture (i.e., lack of object keypoints in the second row) and relatively small in the image (e.g., the last three rows). We attribute the superior performance of our framework to the utilization of mirror reflection in the pre-training stage. + +Quantitative Evaluation. Table 2 shows the experimentation results comparing the proposed SSL pre-training framework with the state-of-the-arts. Besides, we also re + +Table 2. Quantitative comparison between our SSL pre-training framework and existing CNN-based SSL methods. We also compare it with random initialization (Random) and ImageNet pretraining (Supervised). + +
MethodsVenueMirrorNetVCNet
Fβ↑MAE↓Fβ↑MAE↓
Rel. Loc.ICCV '150.5820.2540.8420.072
Rot. Pred.ICLR '180.5370.2120.8510.069
NPIDCVPR '180.7390.1320.8320.076
ODCCVPR '200.5580.2230.8360.077
Moco V2CVPR '200.7070.1770.8350.072
SimCLRICML '200.5940.2440.8370.071
BYOLNeurIPS '200.7040.1580.8320.077
SwAVNeurIPS '200.5780.2240.8560.071
DenseCLCVPR '210.6940.2400.8440.072
SimSiamCVPR '210.6720.1980.8180.080
Bar. TwinsICML '210.6550.1780.8570.069
Random0.5860.2000.4000.103
Supervised20.7270.1700.8710.062
Ours0.7630.1160.8860.057
+ +port the results from models adopting Xavier initialization [13] (Random) or supervised ImageNet pre-training (Supervised). Our SSL pre-training framework (Ours) achieves the best performance across both metrics when applied on MirrorNet and VCNet and shows a good generalization of different mirror detection methods. It is worth noting that our framework even outperforms the supervised ImageNet pre-training. + +# 4.3. Ablation Study + +Table 3 demonstrates the effectiveness of each pretraining stage in our framework. We analyze its results point by point as follows. + +Image-level Pre-training vs. Random Initialization. To evaluate the effectiveness of the proposed image-level pretraining, we compare the model only adopting image-level pre-training (B2) with the model initialized randomly without any pre-training (B1). We can see that only adopting image-level pre-training (B2) will even cause a performance drop compared with the one without pre-training (B1). This somehow supports the finding from VCNet [30] that image-level flipping cannot provide sufficient information to pixelwisely locate the mirror inside an image. + +Table 3. Ablation study of the proposed SSL pre-training framework. We use MirrorNet [37] trained on the MSD dataset as our base method to conduct the experiments in this ablation study. Best results are shown in bold. + +
Rand. Init.Image-lev. PretrainPatch-lev. PretrainPixel-lev. PretrainFβ↑MAE↓
B1✓0.6010.195
B2✓0.6230.191
B3✓0.6890.162
B4✓0.6200.198
B5✓✓0.7230.166
B6✓✓0.7580.124
B7✓✓0.6660.178
Ours✓✓✓0.7630.116
+ +Patch-level Pre-training vs. Random Initialization. We compare our patch-level pre-training (B3) with random initialization (B1) to analyze if patch-level pre-training is useful in our SSL pre-training framework. Our patch-level pretraining outperforms the model without pre-training on both two metrics (14.64% improvement on $F_{\beta}$ and 16.32% improvement on MAE), which indicates the effectiveness of our patch-level pre-training. It also demonstrates the importance of modeling the spatial relationship of mirror reflection at a patch level. + +Pixel-level Pre-training vs. Random Initialization. Similarly, we conduct the ablation experiment on our pixel-level pre-training. The model adopting pixel-level pre-training performs better than the model with image pre-training. However, the performance gain is not as much as the model with patch-level pre-training (B3). One possible reason is that directly optimizing the backbone network for the representation of mirrors at a pixel level is challenging and may not produce satisfactory results for mirror detection. + +Location Predictor in Patch-level Pre-training. The location predictor in our patch-level pre-training is designed to capture the relational spatial information, which has been proven important in mirror detection [23]. We conduct an ablation experiment to remove the location predictor and the performance of our methods drops from 0.763 to 0.731. This shows the importance of learning spatial information in the mirror detection task. + +Combinations of Different Pre-training Stages. Based on the above three comparisons, we find that the patch-level pre-training contributes the most improvement. It demonstrates the importance of modeling the spatial relationship of mirror reflection at a patch level. We then analyze the results from the model with different combinations of our pretraining stages. The model with image-level pre-training and patch-level pre-training (B5) performs better than the model that only adopts patch-level pre-training (B3), which shows that the global information provided by image-level + +![](images/c94e799004490cc55ca02611f6e1d3d0a18c07e817a95c713108ccb4eda94f38.jpg) +Image + +![](images/adb43e744b866dd7ca54d2d708480002a84938a4c0f338e7cd4663448de446fd.jpg) +Image-level Pretraining + +![](images/ef97824b33362b7b6edd9ca6e0ece5d9cda25eeb9aba72ad576370e87fd8e77c.jpg) +Patch-level Pretraining + +![](images/6155df1df3d0122b4be72a7609d83ada630161ef92e3acf130d06f647cb0f2fd.jpg) +Pixel-level Pretraining + +![](images/db77c22435167185c10f3cc363d5ff96d45d33e0f592430360e6dbc4e66205d5.jpg) +Figure 4. Visualization of the effects of the proposed pre-training stages. + +![](images/20de36026fd80a4e087ead7e427f1f5f25a35580a4efbb87fc1d63e9b09084d1.jpg) +(Image+Patch)-level Pretraining +Ours + +![](images/2247c9baf0804f7f3164b3547a535be04149800f888792e44bb69438800898a8.jpg) +GT + +pre-training can benefit the model with patch-level pretraining. Similarly, as shown in B7, B4 (i.e., only pixel-level pretraining) can boost the performance of the pre-trained model after image-level pre-training. These two comparisons indicate that image-level pre-training should be used with other pre-training stages to benefit the overall performance and it should not be solely adopted to our SSL pretraining framework. Besides, we find that combining patch-level and pixel-level pre-training (B6) has superior performance when compared with B3, especially on MAE with about $23.4\%$ improvement. This combination also outperforms supervised ImageNet pre-training, according to the results from Table 2. We attribute its superior performance to the pixel-level representation of mirror reflection. Our final model with all pre-training stages outperforms other ablation models, showing the effectiveness of adopting pretraining stages progressively. + +Visual Comparison of Our Pre-training Stages. Figure 4 shows a visual example of our ablation study. We can see that only adopting image-level pre-training on our SSL pre-training framework can approximately locate the mirror region but fails to produce a precise boundary for the predicted mask of the mirror. Patch-level pre-training can help predict mirrors more accurately with clearer boundaries but over-predict the non-mirror regions as mirrors. In comparison to predictions from all other ablation models, the prediction from the model with pixel-level pretraining has the clearest boundary but contains the largest over-predicted regions, primarily because of the insufficient global-level learning on mirrors. Our framework can significantly reduce over-predictions when image-level and patch-level pre-training are used, as opposed to when patch-level pre-training is used only, and performs the best with the help of the global-to-local pre-training process. + +# 4.4. Discussions + +Impacts of Pre-training Strategy. Some previous SSL works [42, 45], especially for those focusing on self-supervised pre-training for downstream tasks like segmentation [42] and salient object detection [45], would adopt + +Table 4. The results of MirrorNet [37] when using different pretraining strategies. Best results are shown in bold. + +
Fβ↑MAE↓
Random Initialized0.5860.200
ImageNet Supervised0.7270.170
Ours (In parallel)0.7080.331
Ours (Sequential)0.7630.116
+ +different pretext tasks sequentially to pre-train their backbone networks. We also adopt this common strategy in our proposed SSL framework. To verify if employing different pre-training strategies may affect the performance of the proposed SSL framework, we have tried training the three stages sequentially as well as in parallel. We find that when pre-training the three stages in parallel, the backbone network fails to converge and its final fine-tuning results are much worse than those using sequential pre-training. Table 4 shows the results of this experiment. We can see that the MAE performance of pre-training the different stages in parallel is even worse than that of random initialization. A possible reason for this is that directly pre-training different pretext tasks in parallel is difficult to transfer the learned knowledge from the pre-trained network to the target downstream task (i.e., pixel-level mirror detection). + +Apart from the superior performance made by the sequential pre-training strategy, we also notice that the number of pre-training iterations for each pretext task can heavily affect the performance of our SSL framework. To avoid heavy hyperparameter tuning and increase the reproducibility of our pre-training framework, we adopt the same number of pre-training iterations for all stages. Although this strategy of using the same number of pre-training iterations is likely not the optimal design for individual stages, we believe that it can better reflect the real performance of our framework with less stochasticity and more reactivity. We provide a more detailed discussion on this issue next. + +Impacts of Pre-training Iterations. SSL pre-training frameworks usually require huge pre-training iterations to + +![](images/eb08feac410142ebb0149c6e0bde878812dd728c60a1ad1abf429cb263c5f165.jpg) +Figure 5. Ablation study on the number of pre-training iterations for our SSL pre-training framework. We find that our SSL pretraining framework continues to improve in performance until $\sim 20,000$ pre-training iterations. After that, the performance starts to decline, mainly due to overfitting. + +converge. For example, SimSiam [5] requires about 250K iterations to pre-train their model on ImageNet, which is very time-consuming. To validate how this issue may affect our SSL pre-training framework, we conduct the ablation study on the number of pre-training iterations for our SSL pre-training framework, and the results are shown in Figure 5. Note that we use the same number of iterations for all pre-training stages for convenience and to get rid of heavy hyperparameter tuning. + +From the results, we find that the performance of our SSL pre-training framework improves rapidly as the number of pre-training iterations increases, up to around 20,000, which is a relatively small number compared with those used in existing SSL pre-training frameworks. The performance of our SSL pre-training framework gradually drops as the number of pre-training iterations exceeds 20,000, mainly due to overfitting. One possible explanation of this is that the model pre-trained by our SSL pre-training framework can efficiently learn the representation of mirror reflection under a short pre-training period. Another possible explanation is that unlike generic SSL pre-training frameworks that attempt to extract general-propose image representation, our SSL pre-training framework is developed specially for a specific task, i.e., mirror detection. While generic SSL pre-training may cause feature redundancy for our task, having too many pre-training iterations (i.e., over 20,000) causes overfitting. + +Failure Cases. Our SSL pre-training framework does have limitations. It continues to struggle with the limitations of the base method [37], as the base method adopted can also significantly affect the final detection performances. For example, as shown in first row of Figure 6, MirrorNet with our SSL pre-training may still fail to handle some mirror-like regions (e.g., the top-right region). It may also fail to de + +![](images/5e0fa381c7da1010282d6fc2a226bfa379b1d1dfb405d95dcf84c3caa7ac4974.jpg) + +![](images/a21f8058477846417d25926fda83d6439c42b732b24f2a0af94409428e39c431.jpg) + +![](images/5739f80c78e3835cac5bc18142f5c73c76e0a05db49d5b10faeba1c9b4c9370a.jpg) + +![](images/7d9e1db27b0ef5915888bd41c21fdb0bf768ae131bbe9dc9f456e9e524bf3e5c.jpg) +Image + +![](images/646ac3372c6f1b8574c145ef0ee9f0430b5de75bf66388c8b1ba8e482a373c38.jpg) +Ours + +![](images/ef33e88d90d37b68820d4534a4f8cef4e29bf268f405ad1c5d3162e47e1c4c44.jpg) +GT +Figure 6. Failure cases of our SSL pre-training framework based on MirrorNet. Our SSL pre-training framework fails to tackle the intrinsic limitation of existing mirror detection methods. The model pre-trained by our SSL pre-training framework still has wrong predictions when the input images are challenging (e.g., containing mirror-like regions in the first row, and insufficient relationship between mirror/non-mirror regions in the second row. + +tect mirrors in some challenging cases when the input image contains ambiguous mirrors without the relationship of mirror reflection and sufficient contextual contrast, as shown in the second row of Figure 6. + +# 5. Conclusion + +In this paper, we have investigated how self-supervised learning (SSL) works with the mirror detection task. To the best of our knowledge, we are the first to explore SSL pre-training frameworks applied to mirror detection. We have found that the supervised ImageNet pre-training might not be the ideal way to extract backbone image features for mirror detection due to the discrepancy between general-propose representation and mirror-specific representation. We have also proposed a new SSL pre-training framework to pre-train the backbone network for mirror detection. Our SSL pre-training framework does not require any labeled data. It progressively (image-, patch- and pixel-level pre-training) models the relationship of mirror reflection, and obtains a better representation of mirrors in the pre-training stage. Experimental results show that the proposed SSL pre-training framework outperforms CNN-based state-of-the-art SSL pre-training frameworks, and even achieves better results compared with supervised ImageNet pre-training. + +Acknowledgments: The work described in this paper was partially supported by a GRF grant from the Research Grants Council of Hong Kong (Project No.: CityU 11211223). + +# References + +[1] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, pages 3674-3683, 2018. +[2] M. Braun, S. Krebs, F. Flohr, and D. M. Gavrila. Eurocity persons: A novel benchmark for person detection in traffic scenes. IEEE TPAMI, 41(8):1844-1861, Aug 2019. +[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. +[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. +[5] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. +[6] MMSelfSup Contributors. MMSelfSup: Openmmlab self-supervised learning toolbox and benchmark. https://github.com/open-mmlab/mmselfsup, 2021. +[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. +[9] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. +[10] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In ICCV, 2015. +[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. +[12] Christopher Funk and Yanxi Liu. Beyond planar symmetry: Modeling human perception of reflection and rotation symmetries in the wild. In ICCV, pages 793-803, 2017. +[13] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249–256. JMLR Workshop and Conference Proceedings, 2010. +[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS, 2020. +[15] Huankang Guan, Jiaying Lin, and Rynson W.H. Lau. Learning semantic associations for mirror detection. In CVPR, pages 5941-5950, June 2022. + +[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. +[18] Ruozhen He, Jiaying Lin, and Rynson WH Lau. Efficient mirror detection via multi-level heterogeneous learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 790-798, 2023. +[19] Tianyu Huang, Bowen Dong, Jiaying Lin, Xiaohui Liu, Rynson WH Lau, and Wangmeng Zuo. Symmetry-aware transformer-based mirror detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 935–943, 2023. +[20] Peter Kohler, Alexandra Yakovleva, Alasdair Clarke, Yanxi Liu, and Anthony Norcia. Parametric responses to rotation symmetry in mid-level visual cortex. Journal of Vision, 15(12):1122-1122, 2015. +[21] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In ICLR, 2018. +[22] Jiaying Lin, Xin Tan, and Rynson WH Lau. Learning to detect mirrors from videos via dual correspondences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9109-9118, 2023. +[23] Jiaying Lin, Guodong Wang, and Rynson W. H. Lau. Progressive mirror detection. In CVPR, 2020. +[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. +[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. +[26] Zhiqiu Lin, Jin Sun, Abe Davis, and Noah Snavely. Visual chirality. In CVPR, pages 12295-12303, 2020. +[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2018. +[28] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015. +[29] Jiaqi Tan, Weijie Lin, Angel X Chang, and Manolis Savva. Mirror3d: Depth refinement for mirror surfaces. In CVPR, pages 15990-15999, 2021. +[30] Xin Tan, Jiaying Lin, Ke Xu, Chen Pan, Lizhuang Ma, and Rynson W. H. Lau. Mirror detection with the visual chirality cue. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. +[31] Christopher W Tyler. Human symmetry perception and its computational analysis. Psychology Press, 2003. +[32] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In CVPR, 2021. + +[33] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. +[34] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, pages 418-434, 2018. +[35] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, pages 1492-1500, 2017. +[36] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, pages 9653-9663, 2022. +[37] Xin Yang, Haiyang Mei, Ke Xu, Xiaopeng Wei, Baocai Yin, and Rynson W.H. Lau. Where is my mirror? In ICCV, 2019. +[38] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. Learning with noisy labels for robust point cloud segmentation. In ICCV, pages 6443-6452, 2021. +[39] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. Robust point cloud segmentation with noisy annotations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7696-7710, 2022. +[40] Shuquan Ye, Yujia Xie, Dongdong Chen, Yichong Xu, Lu Yuan, Chenguang Zhu, and Jing Liao. Improving commonsense in vision-language models via knowledge graph riddles. In CVPR, pages 2634-2645, 2023. +[41] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021. +[42] Xiaohang Zhan, Ziwei Liu, Ping Luo, Xiaou Tang, and Chen Loy. Mix-and-match tuning for self-supervised semantic segmentation. In AAAI, volume 32, 2018. +[43] Xiaohang Zhan, Jiahao Xie, Ziwei Liu, Yew-Soon Ong, and Chen Change Loy. Online deep clustering for unsupervised representation learning. In CVPR, 2020. +[44] Nanxuan Zhao, Zhirong Wu, Rynson WH Lau, and Stephen Lin. What makes instance discrimination good for transfer learning? In ICLR, 2021. +[45] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, , Hutchuan Lu, and Xiang Ruan. Self-supervised pretraining for rgb-d salient object detection. In AAAI, 2022. +[46] Ying Zheng, Yiyi Zhang, Xiaogang Xu, Jun Wang, and Hongxun Yao. Visual chirality meets freehand sketches. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1544-1548. IEEE, 2021. \ No newline at end of file diff --git a/2023/Self-supervised Pre-training for Mirror Detection/images.zip b/2023/Self-supervised Pre-training for Mirror Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d3dbef6c9f3535a1bf8aa419cdbc29bb527c8778 --- /dev/null +++ b/2023/Self-supervised Pre-training for Mirror Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b294b6b18d975d2749654590fca9fde243f2fc9872987b9bd9676078393ce52 +size 421414 diff --git a/2023/Self-supervised Pre-training for Mirror Detection/layout.json b/2023/Self-supervised Pre-training for Mirror Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0c38fc6e79ba49b3eda7bca9b9c05e22c4746351 --- /dev/null +++ b/2023/Self-supervised Pre-training for Mirror Detection/layout.json @@ -0,0 +1,7575 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 103, + 451, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 103, + 451, + 121 + ], + "spans": [ + { + "bbox": [ + 143, + 103, + 451, + 121 + ], + "type": "text", + "content": "Self-supervised Pre-training for Mirror Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 209, + 144, + 383, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 144, + 383, + 172 + ], + "spans": [ + { + "bbox": [ + 209, + 144, + 383, + 172 + ], + "type": "text", + "content": "Jiaying Lin* Rynson W.H. Lau* City University of Hong Kong" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 198, + 174, + 396, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 174, + 396, + 185 + ], + "spans": [ + { + "bbox": [ + 198, + 174, + 396, + 185 + ], + "type": "text", + "content": "{jiayinlin5, rynson.lau}@cityu.edu.hk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 239, + 290, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 239, + 290, + 574 + ], + "spans": [ + { + "bbox": [ + 45, + 239, + 290, + 574 + ], + "type": "text", + "content": "Existing mirror detection methods require supervised ImageNet pre-training to obtain good general-purpose image features. However, supervised ImageNet pre-training focuses on category-level discrimination and may not be suitable for downstream tasks like mirror detection, due to the overfitting upstream tasks (e.g., supervised image classification). We observe that mirror reflection is crucial to how people perceive the presence of mirrors, and such mid-level features can be better transferred from self-supervised pretrained models. Inspired by this observation, in this paper we aim to improve mirror detection methods by proposing a new self-supervised learning (SSL) pre-training framework for modeling the representation of mirror reflection progressively in the pre-training process. Our framework consists of three pre-training stages at different levels: 1) an image-level pre-training stage to globally incorporate mirror reflection features into the pre-trained model; 2) a patch-level pre-training stage to spatially simulate and learn local mirror reflection from image patches; and 3) a pixel-level pre-training stage to pixel-wisely capture mirror reflection via reconstructing corrupted mirror images based on the relationship between the inside and outside of mirrors. Extensive experiments show that our SSL pre-training framework significantly outperforms previous state-of-the-art CNN-based SSL pre-training frameworks and even outperforms supervised ImageNet pre-training when transferred to the mirror detection task. Code and models are available at https://jiaying.link/iccv2023-sslmirror/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 599, + 128, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 128, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 128, + 611 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 620, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 620, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 620, + 287, + 693 + ], + "type": "text", + "content": "Mirrors are prevalent in our daily lives. As their appearances are largely determined by their surroundings, they generally lack a consistent appearance, making it difficult to separate them from their surroundings. This may affect many computer vision tasks such as object detection [2], vision-language navigation [1] and depth estimation [29]." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 212, + 357, + 270 + ], + "blocks": [ + { + "bbox": [ + 312, + 212, + 357, + 270 + ], + "lines": [ + { + "bbox": [ + 312, + 212, + 357, + 270 + ], + "spans": [ + { + "bbox": [ + 312, + 212, + 357, + 270 + ], + "type": "image", + "image_path": "2053fb7e33566aa3d281812f2faf69bdad556d6356cac893b22a3c6a21bc08bc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 359, + 212, + 403, + 270 + ], + "blocks": [ + { + "bbox": [ + 359, + 212, + 403, + 270 + ], + "lines": [ + { + "bbox": [ + 359, + 212, + 403, + 270 + ], + "spans": [ + { + "bbox": [ + 359, + 212, + 403, + 270 + ], + "type": "image", + "image_path": "97624ca6d1125c6f4390b003b014011a351f9c1950c7ccbc8c0f5ed0ce450959.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 359, + 271, + 403, + 329 + ], + "blocks": [ + { + "bbox": [ + 359, + 271, + 403, + 329 + ], + "lines": [ + { + "bbox": [ + 359, + 271, + 403, + 329 + ], + "spans": [ + { + "bbox": [ + 359, + 271, + 403, + 329 + ], + "type": "image", + "image_path": "a3a1688cab67f58dac910e5e2accc57c4180feddaabc8e18cc4be9edbfbf4b69.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 331, + 400, + 347 + ], + "lines": [ + { + "bbox": [ + 359, + 331, + 400, + 347 + ], + "spans": [ + { + "bbox": [ + 359, + 331, + 400, + 347 + ], + "type": "text", + "content": "(b) MirrorNet on ImageNet" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 405, + 212, + 450, + 270 + ], + "blocks": [ + { + "bbox": [ + 405, + 212, + 450, + 270 + ], + "lines": [ + { + "bbox": [ + 405, + 212, + 450, + 270 + ], + "spans": [ + { + "bbox": [ + 405, + 212, + 450, + 270 + ], + "type": "image", + "image_path": "9c79b20acd129e22e321256d41d612df587f83459c20148214c01a2b2a65136b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 405, + 271, + 450, + 329 + ], + "blocks": [ + { + "bbox": [ + 405, + 271, + 450, + 329 + ], + "lines": [ + { + "bbox": [ + 405, + 271, + 450, + 329 + ], + "spans": [ + { + "bbox": [ + 405, + 271, + 450, + 329 + ], + "type": "image", + "image_path": "21f231f0b30b96f0819c87b4f82bdc5afed12d33bcd81d816967afb479a37ca7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 331, + 446, + 347 + ], + "lines": [ + { + "bbox": [ + 409, + 331, + 446, + 347 + ], + "spans": [ + { + "bbox": [ + 409, + 331, + 446, + 347 + ], + "type": "text", + "content": "(c)VCNet on ImageNet" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 451, + 212, + 496, + 270 + ], + "blocks": [ + { + "bbox": [ + 451, + 212, + 496, + 270 + ], + "lines": [ + { + "bbox": [ + 451, + 212, + 496, + 270 + ], + "spans": [ + { + "bbox": [ + 451, + 212, + 496, + 270 + ], + "type": "image", + "image_path": "2407085591717f6000629a100ebbac3458bd38d6891f0485e3353e069cd334da.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 451, + 271, + 496, + 329 + ], + "blocks": [ + { + "bbox": [ + 451, + 271, + 496, + 329 + ], + "lines": [ + { + "bbox": [ + 451, + 271, + 496, + 329 + ], + "spans": [ + { + "bbox": [ + 451, + 271, + 496, + 329 + ], + "type": "image", + "image_path": "1d85123e767780ec94c3702dff904630533d38df76179b68b715bd5920b81888.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 356, + 546, + 467 + ], + "lines": [ + { + "bbox": [ + 305, + 356, + 546, + 467 + ], + "spans": [ + { + "bbox": [ + 305, + 356, + 546, + 467 + ], + "type": "text", + "content": "Figure 1. State-of-the-art mirror detection methods [37, 30] are based on costly supervised ImageNet pre-training. They may fail even in obvious cases, e.g., (a) when the mirror clearly reflects a real object outside of the mirror. (b) and (c) are MirrorNet [37] and VCNet [30], respectively, pre-trained on ImageNet with full supervision. (c) is MirrorNet with our proposed SSL framework and without supervised ImageNet pre-training. Our SSL scheme leverages the mirror reflection cue and avoids feature redundancy in the pre-training stage. It outperforms those pre-trained on ImageNet with full supervision (i.e., (b) and (c))." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 498, + 212, + 543, + 270 + ], + "blocks": [ + { + "bbox": [ + 498, + 212, + 543, + 270 + ], + "lines": [ + { + "bbox": [ + 498, + 212, + 543, + 270 + ], + "spans": [ + { + "bbox": [ + 498, + 212, + 543, + 270 + ], + "type": "image", + "image_path": "172cb65753d1987a963d6fe5c399475d5ab9a638956f1ba00e1256d1d7e58382.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 498, + 271, + 543, + 329 + ], + "blocks": [ + { + "bbox": [ + 451, + 331, + 503, + 347 + ], + "lines": [ + { + "bbox": [ + 451, + 331, + 503, + 347 + ], + "spans": [ + { + "bbox": [ + 451, + 331, + 503, + 347 + ], + "type": "text", + "content": "1) Ours (MirrorNet on our SSL)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 498, + 271, + 543, + 329 + ], + "lines": [ + { + "bbox": [ + 498, + 271, + 543, + 329 + ], + "spans": [ + { + "bbox": [ + 498, + 271, + 543, + 329 + ], + "type": "image", + "image_path": "db1d2d05c41cec15303148981a731908325a4eedbf1269610181f34c7e074c40.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 511, + 335, + 530, + 343 + ], + "lines": [ + { + "bbox": [ + 511, + 335, + 530, + 343 + ], + "spans": [ + { + "bbox": [ + 511, + 335, + 530, + 343 + ], + "type": "text", + "content": "(e) GT" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 486, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 486, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 486, + 545, + 510 + ], + "type": "text", + "content": "It is therefore crucial to design high-performance mirror detection methods to facilitate computer vision applications." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 510, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 714 + ], + "type": "text", + "content": "Recently, a few methods [37, 23, 30] have been proposed for mirror detection, but they all require to be initialized with supervised ImageNet [8] pre-trained weights and then fine-tuned on mirror detection datasets. While such a transfer learning approach is common in general computer vision tasks (e.g., object detection [24] and semantic segmentation [7]) to utilize category-oriented features from large-scale image datasets, we challenge if this strategy is necessary for the mirror detection task for two reasons. First, the supervised ImageNet pre-training process involves high labelling costs, as it is very time-consuming and labor-intensive to label such a large-scale image dataset (" + }, + { + "bbox": [ + 304, + 510, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 304, + 510, + 546, + 714 + ], + "type": "text", + "content": "1.3M images). Second, unlike general vision tasks, mirror detection does not require category-level object understanding. Instead, it requires an understanding of the relationship between inside and outside of mirrors (e.g., contrast [37], similarity [23], and visual chirality [30]). Thus," + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12227" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "while pre-training on ImageNet requires expensive labels, it causes feature redundancy for the mirror detection task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 227 + ], + "type": "text", + "content": "Figure 1 shows that existing models pre-trained on ImageNet tend to over-detect the mirror regions. Although the two SOTA methods [37, 30] can correctly locate the mirror regions, they are unable to distinguish between the mirrored objects inside the mirrors and the real objects outside, even though these methods contain well-designed modules to learn to address this problem via the fine-tuning stage. This motivates us to investigate whether the current pre-training approach may affect the final detection performance, and whether we may incorporate some intrinsic mirror properties in the pre-training process to assist mirror detection." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 228, + 286, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 286, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 286, + 357 + ], + "type": "text", + "content": "An interesting observation is that humans do not need to be \"trained\" to recognize mirrors. Instead, we learn to recognize them implicitly from young. This aligns with the key idea of self-supervised learning (SSL). Works from neuroscience [31, 20] have shown that humans use mid-level visual cortex to recognize mirror reflection. According to [44], mid-level features are hard to be transferred from supervised pre-training on ImageNet, and SSL is suitable for learning mid-level representations, which inspires us to develop an SSL pre-training framework for modeling mirror reflection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 360, + 286, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 286, + 644 + ], + "type": "text", + "content": "In this paper, we propose a new SSL pre-training framework for mirror detection, which explicitly considers mirror reflection during the pre-training process. Our framework does not require human-annotated labels from large-scale image datasets. It consists of three stages to progressively pre-train the backbone network from global to local: 1) an image-level pre-training stage to obtain the representation of mirror reflection globally by recognizing the geometric transformation applied to the image; 2) a patch-level pretraining stage to mimic patch-wise mirror reflection and then learn the spatial correlation between the original object patches and the corresponding mirrored object patches; and 3) a pixel-level pre-training stage to extract the pixel-to-pixel relationship of mirror reflection by image reconstruction; Under such progressive pre-training scheme, our SSL pre-training framework can learn the representation of mirror reflection, and then effectively transfer this knowledge to the subsequent mirror detection process for better detection performances. We conduct comprehensive experiments to demonstrate the effectiveness of our SSL pretraining framework. We show that it can significantly boost the performances of existing mirror detection methods, and outperform other CNN-based SSL pre-training frameworks on the mirror detection task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 647, + 279, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 647, + 279, + 658 + ], + "spans": [ + { + "bbox": [ + 59, + 647, + 279, + 658 + ], + "type": "text", + "content": "To conclude, this paper makes three key contributions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 665, + 286, + 713 + ], + "type": "text", + "content": "- To the best of our knowledge, we are the first to investigate how existing SSL pre-training frameworks perform on the mirror detection task, compared with supervised ImageNet pre-training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 72, + 545, + 214 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 317, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 545, + 156 + ], + "type": "text", + "content": "- We propose a new SSL pre-training framework that consists of three stages at different levels to progressively learn the representation of mirror reflection. Compared with the features from a supervised ImageNet pre-trained model, our representation is better due to the reduced gap between the pre-training task and the target downstream task (i.e., mirror detection)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 166, + 545, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 166, + 545, + 214 + ], + "spans": [ + { + "bbox": [ + 317, + 166, + 545, + 214 + ], + "type": "text", + "content": "- Extensive experiments show that our SSL pre-training framework performs the best among all state-of-the-art SSL methods, and even better than models with supervised ImageNet pre-training." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 226, + 392, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 226, + 392, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 226, + 392, + 239 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 246, + 408, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 246, + 408, + 258 + ], + "spans": [ + { + "bbox": [ + 306, + 246, + 408, + 258 + ], + "type": "text", + "content": "2.1. Mirror Detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 266, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 384 + ], + "type": "text", + "content": "Mirror detection is an essential task in computer vision, as mirrors are everywhere these days. It is also a challenging task, since mirrors do not possess a consistent appearance, but reflect those of their surroundings. Other vision tasks such as generic object detection and segmentation would tend to detect and segment objects outside as well as inside of the mirror, providing incorrect information on the mirror region, and a depth prediction task may return the depths of the reflected objects inside the mirror region instead of the depth of the mirror." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 385, + 545, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 504 + ], + "type": "text", + "content": "To address the mirror detection problem, Yang et al. [37] propose the first large-scale mirror detection dataset and the first deep-learning model, MirrorNet, for mirror detection by learning the contextual contrast between the inside and outside of mirrors. Lin et al. [23] further propose a new model with similarity and edge learning to progressively detect mirrors. Most recently, Guan et al. [15] employ graph convolutional networks to model contextual associations for mirror detection, while Tan et al. propose VCNet [30] to exploit the visual chirality cue at a feature level." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 506, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 601 + ], + "type": "text", + "content": "However, existing mirror detection methods are all fully-supervised. They also directly adopt the supervised ImageNet [8] pre-training to obtain generic image features. Unlike recent works [19, 18, 22] that propose novel network architectures for mirror detection, we propose in this paper to tackle the mirror detection problem from a learning-based perspective, offering the first self-supervised pre-training framework for mirror detection." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 610, + 474, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 474, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 474, + 623 + ], + "type": "text", + "content": "2.2. Self-supervised Learning (SSL)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "SSL is a popular research topic in computer vision [38, 39], and has attracted a lot of research interests. It aims to learn good feature representations without needing labeled data. Early SSL methods focus on proposing novel pretext tasks, such as relative location prediction [10] and rotation prediction [21]. Recently, most methods are based on the utilization of contrastive learning [16, 4, 14, 3, 32, 5, 41," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12228" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 281, + 145 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 281, + 145 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 281, + 145 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 281, + 145 + ], + "type": "image", + "image_path": "908689b51dc7553c0ccc26fde83f0bd65eb6d802e9a616f0f113c004d699d59b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 284, + 70, + 545, + 145 + ], + "blocks": [ + { + "bbox": [ + 284, + 70, + 545, + 145 + ], + "lines": [ + { + "bbox": [ + 284, + 70, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 284, + 70, + 545, + 145 + ], + "type": "image", + "image_path": "855243fd694ad6ad2d7363759b0686b9465af05eab1878a3b52b8f769d8e9a32.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 100, + 145, + 511, + 270 + ], + "blocks": [ + { + "bbox": [ + 100, + 145, + 511, + 270 + ], + "lines": [ + { + "bbox": [ + 100, + 145, + 511, + 270 + ], + "spans": [ + { + "bbox": [ + 100, + 145, + 511, + 270 + ], + "type": "image", + "image_path": "d955a6d1249ed61c65caf126c7f5730a39a2581f3573b017e7bde9fc5f13c937.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 274, + 547, + 309 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 547, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 547, + 309 + ], + "type": "text", + "content": "Figure 2. Our SSL pre-training framework. It consists of three pre-training stages at different levels: an image-level pre-training stage, a patch-level pre-training stage, and a pixel-level pre-training stage. It pre-trains the backbone network from global to local (image-level " + }, + { + "bbox": [ + 46, + 274, + 547, + 309 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 274, + 547, + 309 + ], + "type": "text", + "content": " patch-level " + }, + { + "bbox": [ + 46, + 274, + 547, + 309 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 274, + 547, + 309 + ], + "type": "text", + "content": " pixel-level) progressively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "text", + "content": "40]. Most recently, [44] shows that some SSL methods perform even better than supervised ImageNet pre-training in downstream tasks, such as object detection [11, 24] and semantic segmentation [7, 34], with the help of instance discrimination during SSL pre-training. Recent transformer-based SSL frameworks [36] also involve masked image modeling to obtain better self-supervised representations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 411, + 288, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 288, + 509 + ], + "type": "text", + "content": "In this paper, we observe that unlike general vision tasks, mirror detection is not benefited from supervised ImageNet pre-training since it does not require category-level object understanding. Hence, we propose a novel SSL pre-training framework specifically for the mirror detection task. The proposed framework aims to simulate and learn the relationship between mirrors and their surrounding during the pre-training stage." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 520, + 103, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 103, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 103, + 533 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 540, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 638 + ], + "type": "text", + "content": "Figure 2 shows the overall architecture of the proposed SSL pre-training framework. Our SSL pre-training framework consists of three pre-training stages at different levels: an image-level pre-training, a patch-level pre-training, and a pixel-level pre-training. The pre-training at each level is conducted sequentially and independently. The progressive and sequential training strategy can help the pre-trained model obtain a global-to-local representation of mirrors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 647, + 185, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 185, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 185, + 659 + ], + "type": "text", + "content": "3.1. Image-level Pre-training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Inspired by the definition of visual chirality [26] and its potential applications [46, 30] in various computer vision problems, we formulate the image-level pre-training as a binary classification problem: Given an input image and its" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 327, + 546, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 546, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 546, + 435 + ], + "type": "text", + "content": "corresponding horizontal flipped image, can the backbone network tell which one is the original image? To achieve this objective, we attach a classification head to predict the probability that the input image is flipped. The classification head consists of a convolution layer followed by a sigmoid activation function to project the output value into " + }, + { + "bbox": [ + 304, + 327, + 546, + 435 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 304, + 327, + 546, + 435 + ], + "type": "text", + "content": ". In this pre-training stage, we use binary cross-entropy (BCE) as our loss function " + }, + { + "bbox": [ + 304, + 327, + 546, + 435 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 304, + 327, + 546, + 435 + ], + "type": "text", + "content": " to optimize the backbone network. Formally, we have:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 345, + 441, + 545, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 441, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 345, + 441, + 545, + 456 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f} = - \\left[ y \\log (\\hat {y}) + (1 - y) \\log (1 - \\hat {y}) \\right], \\tag {1}", + "image_path": "f14ca48a44efac8e48dddd08344aa2b7ce25d850431c1e2f55f021523a5875f2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 461, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 461, + 545, + 486 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 461, + 545, + 486 + ], + "type": "text", + "content": " is the ground truth label and " + }, + { + "bbox": [ + 304, + 461, + 545, + 486 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 304, + 461, + 545, + 486 + ], + "type": "text", + "content": " is the predicted probability that the input image is horizontally flipped." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 491, + 441, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 441, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 441, + 504 + ], + "type": "text", + "content": "3.2. Patch-level Pre-training" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 510, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 714 + ], + "type": "text", + "content": "The image-level pre-training pretext task alone cannot bring much help in locating the mirror inside an image since mirror detection is a pixel-level problem. However, directly connecting the image-level and the pixel-level pretext tasks is challenging due to the gap between the scales of these two pretext tasks. Thus, we design a patch-level pre-training stage as a bridge to cover the feature representations of mirrors at different scales. Besides, the image-level pre-training does not benefit relational spatial understanding, which has been proven important in mirror detection [23]. Our patch-level pre-training consists of two sub-tasks to both obtain the patch-wise relationship of mirror reflection and simulate the mirrored region with pseudo labels. We first split the input image into a grid of " + }, + { + "bbox": [ + 304, + 510, + 546, + 714 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 510, + 546, + 714 + ], + "type": "text", + "content": " image patches and randomly shuffle them. After that, we randomly apply a flipping operation to one of the image patches, to simulate a mirrored region. The first sub-task" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12229" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "is to predict the location index from the shuffled grid by the backbone network. We append a location predictor head to the backbone network to predict the location index of input patches. We apply cross-entropy loss as " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{loc}" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " to optimize this sub-task. The second sub-task is to predict the flipped patch from the nine randomly shuffled image patches. Similar to our image-level pre-training, we use BCE loss as our loss function " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " in this sub-task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 177, + 287, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 177, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 58, + 177, + 287, + 209 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l o c} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} y _ {i} \\log \\left(\\hat {y} _ {i}\\right), \\tag {2}", + "image_path": "1b9b5fca78b123015f5d7760399ec58338d01990a01dabac3fca4edaa79a77f3.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 213, + 287, + 246 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 213, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 58, + 213, + 287, + 246 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} = - \\frac {1}{N} \\sum_ {j = 1} ^ {N} y _ {j} \\log \\left(\\hat {y _ {j}}\\right) + (1 - y _ {j}) \\log \\left(1 - \\hat {y _ {j}}\\right), \\tag {3}", + "image_path": "860c613f6201780d54c0d0132173369ad0cf33d5b5d21fe034b237ca4910f509.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 255, + 287, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 316 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 255, + 287, + 316 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 46, + 255, + 287, + 316 + ], + "type": "text", + "content": " is the predicted location index, and " + }, + { + "bbox": [ + 46, + 255, + 287, + 316 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 46, + 255, + 287, + 316 + ], + "type": "text", + "content": " is the corresponding ground truth label. " + }, + { + "bbox": [ + 46, + 255, + 287, + 316 + ], + "type": "inline_equation", + "content": "\\hat{y}_j" + }, + { + "bbox": [ + 46, + 255, + 287, + 316 + ], + "type": "text", + "content": " is the predicted probability that the input patch is horizontally flipped and " + }, + { + "bbox": [ + 46, + 255, + 287, + 316 + ], + "type": "inline_equation", + "content": "y_j" + }, + { + "bbox": [ + 46, + 255, + 287, + 316 + ], + "type": "text", + "content": " is the corresponding ground truth probability. Finally, we jointly optimize these two sub-tasks and the final loss function is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 120, + 337, + 287, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 337, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 120, + 337, + 287, + 352 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p a t c h}} = \\mathcal {L} _ {\\text {l o c}} + \\mathcal {L} _ {\\text {c l s}}. \\tag {4}", + "image_path": "a181f23cdb99d24df34eae9138062359cd71b99dcfd77043cc9c883deb686961.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 359, + 179, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 359, + 179, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 359, + 179, + 373 + ], + "type": "text", + "content": "3.3. Pixel-level Pre-training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 378, + 289, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 289, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 289, + 571 + ], + "type": "text", + "content": "Extracting the pixel-level relationship of mirror reflection is a challenging task, as studied by previous related work [12] for symmetry detection. It requires point-level GT labels to represent symmetry centers. However, obtaining such labels requires huge human annotation effort and they are not applicable for SSL pre-training. Instead of generating pseudo labels for symmetry detection, our solution is that we formulate this task as an image reconstruction problem on images containing mirrors. An ideal pre-trained network should be able to reconstruct the input image containing mirrors with the pixel-level understanding of mirror reflection. We randomly mask out the input image by cutout augmentation [9] with the ratio of 0.3 and then feed it to the backbone network with a reconstruction head. The details of the reconstruction head we used are listed in Table 1. The objective of our pixel-level pre-training is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 129, + 580, + 287, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 580, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 129, + 580, + 287, + 594 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p i x} = \\| x - \\hat {x} \\| _ {1}, \\tag {5}", + "image_path": "7768b0ebe2be45c399964c8130e893c73017b269640f1952923a65e8b984cb76.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pix}" + }, + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "type": "text", + "content": " is an L1 loss. " + }, + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "type": "text", + "content": " are the input image and the reconstructed image, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 639, + 128, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 128, + 652 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 128, + 652 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 658, + 174, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 174, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 174, + 672 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "Dataset and Evaluation Metrics. The image-level and patch-level pre-training experiments are conducted on the training set of ImageNet-100 [28] and MS COCO [25]," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 68, + 547, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 68, + 547, + 135 + ], + "spans": [ + { + "bbox": [ + 304, + 68, + 547, + 135 + ], + "type": "text", + "content": "Table 1. The architecture of the reconstruction head used in the pixel-level pre-training. Note that each \"conv-IR\" corresponds to a sequence of convolution layer, InstanceNorm layer and ReLU activation. " + }, + { + "bbox": [ + 304, + 68, + 547, + 135 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 68, + 547, + 135 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 68, + 547, + 135 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 68, + 547, + 135 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 68, + 547, + 135 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 68, + 547, + 135 + ], + "type": "text", + "content": " denote the number of kernels, the number of strides and the padding size, respectively, used in the convolution layer." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 342, + 136, + 511, + 291 + ], + "blocks": [ + { + "bbox": [ + 342, + 136, + 511, + 291 + ], + "lines": [ + { + "bbox": [ + 342, + 136, + 511, + 291 + ], + "spans": [ + { + "bbox": [ + 342, + 136, + 511, + 291 + ], + "type": "table", + "html": "
Layer Details
3 × 3 conv-IR, K = 1024, S = 1, P = 1
2 × Upsample
3 × 3 conv-IR, K = 512, S = 1, P = 1
2 × Upsample
3 × 3 conv-IR, K = 256, S = 1, P = 1
2 × Upsample
3 × 3 conv-IR, K = 128, S = 1, P = 1
2 × Upsample
3 × 3 conv-IR, K = 128, S = 1, P = 1
2 × Upsample
1 × 1 conv-IR, K = 3, S = 1, P = 1
", + "image_path": "d5f1df4fcdc5de47ce0dcf08320a1def3b98b1b81fdc75f45f072bacb7572ba9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 312, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 484 + ], + "type": "text", + "content": "which contains " + }, + { + "bbox": [ + 304, + 312, + 545, + 484 + ], + "type": "inline_equation", + "content": "130\\mathrm{K}" + }, + { + "bbox": [ + 304, + 312, + 545, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 312, + 545, + 484 + ], + "type": "inline_equation", + "content": "\\sim 118\\mathrm{k}" + }, + { + "bbox": [ + 304, + 312, + 545, + 484 + ], + "type": "text", + "content": " images respectively. In the pixel-level pre-training and fine-tuning stage, our experiments are conducted on MSD [37], the first large-scale dataset for mirror detection. MSD [37] provides 4,018 mirror images, which are divided into 3,063 images for training and 955 images for testing. We adopt two popular metrics namely, F-measure " + }, + { + "bbox": [ + 304, + 312, + 545, + 484 + ], + "type": "inline_equation", + "content": "(\\mathbf{F}_{\\beta})" + }, + { + "bbox": [ + 304, + 312, + 545, + 484 + ], + "type": "text", + "content": " and mean absolute error (MAE) to evaluate the performance of our SSL pre-training framework. F-measure can reflect the overall model performance and is calculated by the weighted harmonic mean of the precision and recall: " + }, + { + "bbox": [ + 304, + 312, + 545, + 484 + ], + "type": "inline_equation", + "content": "F_{\\beta} = \\frac{(1 + \\beta^{2})\\cdot\\text{precision}\\cdot\\text{recall}}{\\beta^{2}\\cdot\\text{precision} + \\text{recall}}" + }, + { + "bbox": [ + 304, + 312, + 545, + 484 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 312, + 545, + 484 + ], + "type": "inline_equation", + "content": "\\beta^2" + }, + { + "bbox": [ + 304, + 312, + 545, + 484 + ], + "type": "text", + "content": " is set to 0.3. MAE can measure the average pixel-wise absolute disparity between the predicted mirror maps " + }, + { + "bbox": [ + 304, + 312, + 545, + 484 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 312, + 545, + 484 + ], + "type": "text", + "content": " and the binary ground truth " + }, + { + "bbox": [ + 304, + 312, + 545, + 484 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 312, + 545, + 484 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 484, + 546, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 546, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 546, + 665 + ], + "type": "text", + "content": "Implementation Details. We use Pytorch to implement our proposed SSL pre-training framework. All experiments are conducted on a GPU server with eight NVIDIA RTX 3090 GPUs. Following existing SSL pre-training frameworks, we adopt ResNet-50 [17] as our backbone network. All pre-trained model weights of existing SSL pretraining frameworks are obtained from the released models of mmselfsup [6] for fair comparisons. For each pretext task in the pre-training stage, all images are resized to " + }, + { + "bbox": [ + 304, + 484, + 546, + 665 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 484, + 546, + 665 + ], + "type": "text", + "content": ". The batch size is set to 352. We use AdamW [27] with an initial learning rate of 1e-4 to optimize our SSL pre-training framework and stop pre-training after 20,000 iterations. We only adopt random color jittering for data augmentation. In the fine-tuning stage, we use the same hyperparameters reported in the original papers of the adopted" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 673, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 673, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 673, + 547, + 713 + ], + "type": "text", + "content": "1It is worth noting that the datasets that we used (about 250K training images) are much smaller than the pre-trained dataset ImageNet (14 million images) used by other baseline SSL methods. We adopted ImageNet-100 to pre-train our network due to the limited computational resources." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12230" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 69, + 545, + 469 + ], + "blocks": [ + { + "bbox": [ + 52, + 69, + 545, + 469 + ], + "lines": [ + { + "bbox": [ + 52, + 69, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 52, + 69, + 545, + 469 + ], + "type": "image", + "image_path": "fb1c2b369ed7aed8c0640a1bc2c105e7f93e01f49c8804770f246f9699221c5a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 479, + 546, + 502 + ], + "lines": [ + { + "bbox": [ + 46, + 479, + 546, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 546, + 502 + ], + "type": "text", + "content": "Figure 3. Qualitative Comparison of our proposed method with NPID [33], MocoV2 [4], SwAV [3], Barlow Twins [41], ImageNet pretraining (Supervised). We use the released pre-trained models of these existing SSL methods from mmselfsup [6] for fair comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 521, + 235, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 235, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 235, + 533 + ], + "type": "text", + "content": "base methods MirrorNet [37] and VCNet [30]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 544, + 234, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 234, + 556 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 234, + 556 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Arts" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 563, + 287, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 671 + ], + "type": "text", + "content": "We compare against 11 state-of-the-art CNN-based frameworks in self-supervised learning, including Relative Location (Rel. Loc.) [10], Rotation Prediction (Rot. Pred.) [21], NPID [33], ODC [43], Moco V2 [16], SimCLR [4], BYOL [14], SwAV [3], DenseCL [32], SimSiam [5] and Barlow Twins (Bar. Twins) [41]. We select MirrorNet [37] and VCNet [30], which are two representative mirror detection methods, as base methods to conduct experiments in the fine-tuning stage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Qualitative Evaluation. We further show the performance of our SSL pre-training framework in Figure 3. We compare our framework with four state-of-the-art SSL pre-training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": "frameworks (including the two best-performing frameworks NPID [33] and Moco V2 [16] based on MirrorNet and the two best-performing frameworks SwAV [3] and Barlow Twins [41] based on VCNet, according to their performance in Table 2) and supervised ImageNet pre-training due to the limitation of space. Our framework can correctly detect the mirror regions by utilizing the relationship of mirror reflection caused by mirrors, even though the reflected regions are pure texture (i.e., lack of object keypoints in the second row) and relatively small in the image (e.g., the last three rows). We attribute the superior performance of our framework to the utilization of mirror reflection in the pre-training stage." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "content": "Quantitative Evaluation. Table 2 shows the experimentation results comparing the proposed SSL pre-training framework with the state-of-the-arts. Besides, we also re" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12231" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 113, + 300, + 326 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "type": "text", + "content": "Table 2. Quantitative comparison between our SSL pre-training framework and existing CNN-based SSL methods. We also compare it with random initialization (Random) and ImageNet pretraining (Supervised)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 113, + 300, + 326 + ], + "lines": [ + { + "bbox": [ + 47, + 113, + 300, + 326 + ], + "spans": [ + { + "bbox": [ + 47, + 113, + 300, + 326 + ], + "type": "table", + "html": "
MethodsVenueMirrorNetVCNet
Fβ↑MAE↓Fβ↑MAE↓
Rel. Loc.ICCV '150.5820.2540.8420.072
Rot. Pred.ICLR '180.5370.2120.8510.069
NPIDCVPR '180.7390.1320.8320.076
ODCCVPR '200.5580.2230.8360.077
Moco V2CVPR '200.7070.1770.8350.072
SimCLRICML '200.5940.2440.8370.071
BYOLNeurIPS '200.7040.1580.8320.077
SwAVNeurIPS '200.5780.2240.8560.071
DenseCLCVPR '210.6940.2400.8440.072
SimSiamCVPR '210.6720.1980.8180.080
Bar. TwinsICML '210.6550.1780.8570.069
Random0.5860.2000.4000.103
Supervised20.7270.1700.8710.062
Ours0.7630.1160.8860.057
", + "image_path": "2f5fd76e72a5d854ad1405ab978ab00692d8bc861f3d00afd8860386de67b80f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 346, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 441 + ], + "type": "text", + "content": "port the results from models adopting Xavier initialization [13] (Random) or supervised ImageNet pre-training (Supervised). Our SSL pre-training framework (Ours) achieves the best performance across both metrics when applied on MirrorNet and VCNet and shows a good generalization of different mirror detection methods. It is worth noting that our framework even outperforms the supervised ImageNet pre-training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 449, + 140, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 140, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 140, + 462 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 468, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 504 + ], + "type": "text", + "content": "Table 3 demonstrates the effectiveness of each pretraining stage in our framework. We analyze its results point by point as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 628 + ], + "type": "text", + "content": "Image-level Pre-training vs. Random Initialization. To evaluate the effectiveness of the proposed image-level pretraining, we compare the model only adopting image-level pre-training (B2) with the model initialized randomly without any pre-training (B1). We can see that only adopting image-level pre-training (B2) will even cause a performance drop compared with the one without pre-training (B1). This somehow supports the finding from VCNet [30] that image-level flipping cannot provide sufficient information to pixelwisely locate the mirror inside an image." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 306, + 112, + 550, + 222 + ], + "blocks": [ + { + "bbox": [ + 306, + 69, + 547, + 112 + ], + "lines": [ + { + "bbox": [ + 306, + 69, + 547, + 112 + ], + "spans": [ + { + "bbox": [ + 306, + 69, + 547, + 112 + ], + "type": "text", + "content": "Table 3. Ablation study of the proposed SSL pre-training framework. We use MirrorNet [37] trained on the MSD dataset as our base method to conduct the experiments in this ablation study. Best results are shown in bold." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 112, + 550, + 222 + ], + "lines": [ + { + "bbox": [ + 306, + 112, + 550, + 222 + ], + "spans": [ + { + "bbox": [ + 306, + 112, + 550, + 222 + ], + "type": "table", + "html": "
Rand. Init.Image-lev. PretrainPatch-lev. PretrainPixel-lev. PretrainFβ↑MAE↓
B1✓0.6010.195
B2✓0.6230.191
B3✓0.6890.162
B4✓0.6200.198
B5✓✓0.7230.166
B6✓✓0.7580.124
B7✓✓0.6660.178
Ours✓✓✓0.7630.116
", + "image_path": "2f7cfdf59eca0a1b160578d6d2e9538726cbed7619ecb9b41733fca790aad8b2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 247, + 547, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 247, + 547, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 547, + 369 + ], + "type": "text", + "content": "Patch-level Pre-training vs. Random Initialization. We compare our patch-level pre-training (B3) with random initialization (B1) to analyze if patch-level pre-training is useful in our SSL pre-training framework. Our patch-level pretraining outperforms the model without pre-training on both two metrics (14.64% improvement on " + }, + { + "bbox": [ + 305, + 247, + 547, + 369 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 305, + 247, + 547, + 369 + ], + "type": "text", + "content": " and 16.32% improvement on MAE), which indicates the effectiveness of our patch-level pre-training. It also demonstrates the importance of modeling the spatial relationship of mirror reflection at a patch level." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 375, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 375, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 305, + 375, + 545, + 483 + ], + "type": "text", + "content": "Pixel-level Pre-training vs. Random Initialization. Similarly, we conduct the ablation experiment on our pixel-level pre-training. The model adopting pixel-level pre-training performs better than the model with image pre-training. However, the performance gain is not as much as the model with patch-level pre-training (B3). One possible reason is that directly optimizing the backbone network for the representation of mirrors at a pixel level is challenging and may not produce satisfactory results for mirror detection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 490, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 490, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 305, + 490, + 545, + 586 + ], + "type": "text", + "content": "Location Predictor in Patch-level Pre-training. The location predictor in our patch-level pre-training is designed to capture the relational spatial information, which has been proven important in mirror detection [23]. We conduct an ablation experiment to remove the location predictor and the performance of our methods drops from 0.763 to 0.731. This shows the importance of learning spatial information in the mirror detection task." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "Combinations of Different Pre-training Stages. Based on the above three comparisons, we find that the patch-level pre-training contributes the most improvement. It demonstrates the importance of modeling the spatial relationship of mirror reflection at a patch level. We then analyze the results from the model with different combinations of our pretraining stages. The model with image-level pre-training and patch-level pre-training (B5) performs better than the model that only adopts patch-level pre-training (B3), which shows that the global information provided by image-level" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 635, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 287, + 713 + ], + "type": "text", + "content": "2Note that the original backbone networks reported in their papers are ResNext-101 [35]. However, the model weights of ResNext-101 pre-trained by our baseline SSL frameworks are not available in mmselfsup [6]. Thus, we switch the backbone network of these two methods to ResNet-50 [17], which is a common backbone network used in evaluating the performances of SSL frameworks. We also directly use the released ResNet-50 SSL pre-trained weights in our experiments for fair comparisons." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12232" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 82, + 132, + 164 + ], + "blocks": [ + { + "bbox": [ + 66, + 82, + 132, + 164 + ], + "lines": [ + { + "bbox": [ + 66, + 82, + 132, + 164 + ], + "spans": [ + { + "bbox": [ + 66, + 82, + 132, + 164 + ], + "type": "image", + "image_path": "c94e799004490cc55ca02611f6e1d3d0a18c07e817a95c713108ccb4eda94f38.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 171, + 109, + 182 + ], + "lines": [ + { + "bbox": [ + 86, + 171, + 109, + 182 + ], + "spans": [ + { + "bbox": [ + 86, + 171, + 109, + 182 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 133, + 82, + 198, + 164 + ], + "blocks": [ + { + "bbox": [ + 133, + 82, + 198, + 164 + ], + "lines": [ + { + "bbox": [ + 133, + 82, + 198, + 164 + ], + "spans": [ + { + "bbox": [ + 133, + 82, + 198, + 164 + ], + "type": "image", + "image_path": "adb43e744b866dd7ca54d2d708480002a84938a4c0f338e7cd4663448de446fd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 167, + 186, + 186 + ], + "lines": [ + { + "bbox": [ + 145, + 167, + 186, + 186 + ], + "spans": [ + { + "bbox": [ + 145, + 167, + 186, + 186 + ], + "type": "text", + "content": "Image-level Pretraining" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 199, + 82, + 265, + 164 + ], + "blocks": [ + { + "bbox": [ + 199, + 82, + 265, + 164 + ], + "lines": [ + { + "bbox": [ + 199, + 82, + 265, + 164 + ], + "spans": [ + { + "bbox": [ + 199, + 82, + 265, + 164 + ], + "type": "image", + "image_path": "ef97824b33362b7b6edd9ca6e0ece5d9cda25eeb9aba72ad576370e87fd8e77c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 213, + 167, + 250, + 186 + ], + "lines": [ + { + "bbox": [ + 213, + 167, + 250, + 186 + ], + "spans": [ + { + "bbox": [ + 213, + 167, + 250, + 186 + ], + "type": "text", + "content": "Patch-level Pretraining" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 266, + 82, + 331, + 164 + ], + "blocks": [ + { + "bbox": [ + 266, + 82, + 331, + 164 + ], + "lines": [ + { + "bbox": [ + 266, + 82, + 331, + 164 + ], + "spans": [ + { + "bbox": [ + 266, + 82, + 331, + 164 + ], + "type": "image", + "image_path": "6155df1df3d0122b4be72a7609d83ada630161ef92e3acf130d06f647cb0f2fd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 167, + 316, + 186 + ], + "lines": [ + { + "bbox": [ + 280, + 167, + 316, + 186 + ], + "spans": [ + { + "bbox": [ + 280, + 167, + 316, + 186 + ], + "type": "text", + "content": "Pixel-level Pretraining" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 332, + 82, + 397, + 164 + ], + "blocks": [ + { + "bbox": [ + 332, + 82, + 397, + 164 + ], + "lines": [ + { + "bbox": [ + 332, + 82, + 397, + 164 + ], + "spans": [ + { + "bbox": [ + 332, + 82, + 397, + 164 + ], + "type": "image", + "image_path": "db77c22435167185c10f3cc363d5ff96d45d33e0f592430360e6dbc4e66205d5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 195, + 427, + 207 + ], + "lines": [ + { + "bbox": [ + 164, + 195, + 427, + 207 + ], + "spans": [ + { + "bbox": [ + 164, + 195, + 427, + 207 + ], + "type": "text", + "content": "Figure 4. Visualization of the effects of the proposed pre-training stages." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 399, + 82, + 463, + 164 + ], + "blocks": [ + { + "bbox": [ + 331, + 167, + 397, + 186 + ], + "lines": [ + { + "bbox": [ + 331, + 167, + 397, + 186 + ], + "spans": [ + { + "bbox": [ + 331, + 167, + 397, + 186 + ], + "type": "text", + "content": "(Image+Patch)-level Pretraining" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 82, + 463, + 164 + ], + "lines": [ + { + "bbox": [ + 399, + 82, + 463, + 164 + ], + "spans": [ + { + "bbox": [ + 399, + 82, + 463, + 164 + ], + "type": "image", + "image_path": "20de36026fd80a4e087ead7e427f1f5f25a35580a4efbb87fc1d63e9b09084d1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 172, + 440, + 180 + ], + "lines": [ + { + "bbox": [ + 422, + 172, + 440, + 180 + ], + "spans": [ + { + "bbox": [ + 422, + 172, + 440, + 180 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 465, + 82, + 531, + 164 + ], + "blocks": [ + { + "bbox": [ + 465, + 82, + 531, + 164 + ], + "lines": [ + { + "bbox": [ + 465, + 82, + 531, + 164 + ], + "spans": [ + { + "bbox": [ + 465, + 82, + 531, + 164 + ], + "type": "image", + "image_path": "2247c9baf0804f7f3164b3547a535be04149800f888792e44bb69438800898a8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 491, + 172, + 503, + 179 + ], + "lines": [ + { + "bbox": [ + 491, + 172, + 503, + 179 + ], + "spans": [ + { + "bbox": [ + 491, + 172, + 503, + 179 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 226, + 289, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 226, + 289, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 289, + 430 + ], + "type": "text", + "content": "pre-training can benefit the model with patch-level pretraining. Similarly, as shown in B7, B4 (i.e., only pixel-level pretraining) can boost the performance of the pre-trained model after image-level pre-training. These two comparisons indicate that image-level pre-training should be used with other pre-training stages to benefit the overall performance and it should not be solely adopted to our SSL pretraining framework. Besides, we find that combining patch-level and pixel-level pre-training (B6) has superior performance when compared with B3, especially on MAE with about " + }, + { + "bbox": [ + 46, + 226, + 289, + 430 + ], + "type": "inline_equation", + "content": "23.4\\%" + }, + { + "bbox": [ + 46, + 226, + 289, + 430 + ], + "type": "text", + "content": " improvement. This combination also outperforms supervised ImageNet pre-training, according to the results from Table 2. We attribute its superior performance to the pixel-level representation of mirror reflection. Our final model with all pre-training stages outperforms other ablation models, showing the effectiveness of adopting pretraining stages progressively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 435, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 639 + ], + "type": "text", + "content": "Visual Comparison of Our Pre-training Stages. Figure 4 shows a visual example of our ablation study. We can see that only adopting image-level pre-training on our SSL pre-training framework can approximately locate the mirror region but fails to produce a precise boundary for the predicted mask of the mirror. Patch-level pre-training can help predict mirrors more accurately with clearer boundaries but over-predict the non-mirror regions as mirrors. In comparison to predictions from all other ablation models, the prediction from the model with pixel-level pretraining has the clearest boundary but contains the largest over-predicted regions, primarily because of the insufficient global-level learning on mirrors. Our framework can significantly reduce over-predictions when image-level and patch-level pre-training are used, as opposed to when patch-level pre-training is used only, and performs the best with the help of the global-to-local pre-training process." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 647, + 124, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 124, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 124, + 658 + ], + "type": "text", + "content": "4.4. Discussions" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "Impacts of Pre-training Strategy. Some previous SSL works [42, 45], especially for those focusing on self-supervised pre-training for downstream tasks like segmentation [42] and salient object detection [45], would adopt" + } + ] + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 345, + 244, + 507, + 322 + ], + "blocks": [ + { + "bbox": [ + 305, + 223, + 545, + 244 + ], + "lines": [ + { + "bbox": [ + 305, + 223, + 545, + 244 + ], + "spans": [ + { + "bbox": [ + 305, + 223, + 545, + 244 + ], + "type": "text", + "content": "Table 4. The results of MirrorNet [37] when using different pretraining strategies. Best results are shown in bold." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 345, + 244, + 507, + 322 + ], + "lines": [ + { + "bbox": [ + 345, + 244, + 507, + 322 + ], + "spans": [ + { + "bbox": [ + 345, + 244, + 507, + 322 + ], + "type": "table", + "html": "
Fβ↑MAE↓
Random Initialized0.5860.200
ImageNet Supervised0.7270.170
Ours (In parallel)0.7080.331
Ours (Sequential)0.7630.116
", + "image_path": "9739fadea13fe3780833f9878be510cf87ad0a19c2692b950d139ef45f588e77.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 345, + 545, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 345, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 304, + 345, + 545, + 537 + ], + "type": "text", + "content": "different pretext tasks sequentially to pre-train their backbone networks. We also adopt this common strategy in our proposed SSL framework. To verify if employing different pre-training strategies may affect the performance of the proposed SSL framework, we have tried training the three stages sequentially as well as in parallel. We find that when pre-training the three stages in parallel, the backbone network fails to converge and its final fine-tuning results are much worse than those using sequential pre-training. Table 4 shows the results of this experiment. We can see that the MAE performance of pre-training the different stages in parallel is even worse than that of random initialization. A possible reason for this is that directly pre-training different pretext tasks in parallel is difficult to transfer the learned knowledge from the pre-trained network to the target downstream task (i.e., pixel-level mirror detection)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 539, + 545, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 539, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 304, + 539, + 545, + 682 + ], + "type": "text", + "content": "Apart from the superior performance made by the sequential pre-training strategy, we also notice that the number of pre-training iterations for each pretext task can heavily affect the performance of our SSL framework. To avoid heavy hyperparameter tuning and increase the reproducibility of our pre-training framework, we adopt the same number of pre-training iterations for all stages. Although this strategy of using the same number of pre-training iterations is likely not the optimal design for individual stages, we believe that it can better reflect the real performance of our framework with less stochasticity and more reactivity. We provide a more detailed discussion on this issue next." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "Impacts of Pre-training Iterations. SSL pre-training frameworks usually require huge pre-training iterations to" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12233" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 72, + 251, + 218 + ], + "blocks": [ + { + "bbox": [ + 82, + 72, + 251, + 218 + ], + "lines": [ + { + "bbox": [ + 82, + 72, + 251, + 218 + ], + "spans": [ + { + "bbox": [ + 82, + 72, + 251, + 218 + ], + "type": "image", + "image_path": "eb08feac410142ebb0149c6e0bde878812dd728c60a1ad1abf429cb263c5f165.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 223, + 287, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 223, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 287, + 277 + ], + "type": "text", + "content": "Figure 5. Ablation study on the number of pre-training iterations for our SSL pre-training framework. We find that our SSL pretraining framework continues to improve in performance until " + }, + { + "bbox": [ + 46, + 223, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\sim 20,000" + }, + { + "bbox": [ + 46, + 223, + 287, + 277 + ], + "type": "text", + "content": " pre-training iterations. After that, the performance starts to decline, mainly due to overfitting." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 295, + 286, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 295, + 286, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 286, + 403 + ], + "type": "text", + "content": "converge. For example, SimSiam [5] requires about 250K iterations to pre-train their model on ImageNet, which is very time-consuming. To validate how this issue may affect our SSL pre-training framework, we conduct the ablation study on the number of pre-training iterations for our SSL pre-training framework, and the results are shown in Figure 5. Note that we use the same number of iterations for all pre-training stages for convenience and to get rid of heavy hyperparameter tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 405, + 286, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 286, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 286, + 622 + ], + "type": "text", + "content": "From the results, we find that the performance of our SSL pre-training framework improves rapidly as the number of pre-training iterations increases, up to around 20,000, which is a relatively small number compared with those used in existing SSL pre-training frameworks. The performance of our SSL pre-training framework gradually drops as the number of pre-training iterations exceeds 20,000, mainly due to overfitting. One possible explanation of this is that the model pre-trained by our SSL pre-training framework can efficiently learn the representation of mirror reflection under a short pre-training period. Another possible explanation is that unlike generic SSL pre-training frameworks that attempt to extract general-propose image representation, our SSL pre-training framework is developed specially for a specific task, i.e., mirror detection. While generic SSL pre-training may cause feature redundancy for our task, having too many pre-training iterations (i.e., over 20,000) causes overfitting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "Failure Cases. Our SSL pre-training framework does have limitations. It continues to struggle with the limitations of the base method [37], as the base method adopted can also significantly affect the final detection performances. For example, as shown in first row of Figure 6, MirrorNet with our SSL pre-training may still fail to handle some mirror-like regions (e.g., the top-right region). It may also fail to de" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 313, + 69, + 389, + 131 + ], + "blocks": [ + { + "bbox": [ + 313, + 69, + 389, + 131 + ], + "lines": [ + { + "bbox": [ + 313, + 69, + 389, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 69, + 389, + 131 + ], + "type": "image", + "image_path": "5e0fa381c7da1010282d6fc2a226bfa379b1d1dfb405d95dcf84c3caa7ac4974.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 390, + 70, + 465, + 132 + ], + "blocks": [ + { + "bbox": [ + 390, + 70, + 465, + 132 + ], + "lines": [ + { + "bbox": [ + 390, + 70, + 465, + 132 + ], + "spans": [ + { + "bbox": [ + 390, + 70, + 465, + 132 + ], + "type": "image", + "image_path": "a21f8058477846417d25926fda83d6439c42b732b24f2a0af94409428e39c431.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 466, + 70, + 541, + 132 + ], + "blocks": [ + { + "bbox": [ + 466, + 70, + 541, + 132 + ], + "lines": [ + { + "bbox": [ + 466, + 70, + 541, + 132 + ], + "spans": [ + { + "bbox": [ + 466, + 70, + 541, + 132 + ], + "type": "image", + "image_path": "5739f80c78e3835cac5bc18142f5c73c76e0a05db49d5b10faeba1c9b4c9370a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 313, + 133, + 389, + 228 + ], + "blocks": [ + { + "bbox": [ + 313, + 133, + 389, + 228 + ], + "lines": [ + { + "bbox": [ + 313, + 133, + 389, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 389, + 228 + ], + "type": "image", + "image_path": "7d9e1db27b0ef5915888bd41c21fdb0bf768ae131bbe9dc9f456e9e524bf3e5c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 232, + 361, + 242 + ], + "lines": [ + { + "bbox": [ + 339, + 232, + 361, + 242 + ], + "spans": [ + { + "bbox": [ + 339, + 232, + 361, + 242 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 390, + 133, + 465, + 228 + ], + "blocks": [ + { + "bbox": [ + 390, + 133, + 465, + 228 + ], + "lines": [ + { + "bbox": [ + 390, + 133, + 465, + 228 + ], + "spans": [ + { + "bbox": [ + 390, + 133, + 465, + 228 + ], + "type": "image", + "image_path": "646ac3372c6f1b8574c145ef0ee9f0430b5de75bf66388c8b1ba8e482a373c38.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 232, + 435, + 241 + ], + "lines": [ + { + "bbox": [ + 419, + 232, + 435, + 241 + ], + "spans": [ + { + "bbox": [ + 419, + 232, + 435, + 241 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 466, + 133, + 541, + 228 + ], + "blocks": [ + { + "bbox": [ + 466, + 133, + 541, + 228 + ], + "lines": [ + { + "bbox": [ + 466, + 133, + 541, + 228 + ], + "spans": [ + { + "bbox": [ + 466, + 133, + 541, + 228 + ], + "type": "image", + "image_path": "ef33e88d90d37b68820d4534a4f8cef4e29bf268f405ad1c5d3162e47e1c4c44.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 498, + 232, + 510, + 240 + ], + "lines": [ + { + "bbox": [ + 498, + 232, + 510, + 240 + ], + "spans": [ + { + "bbox": [ + 498, + 232, + 510, + 240 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 251, + 545, + 328 + ], + "lines": [ + { + "bbox": [ + 305, + 251, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 305, + 251, + 545, + 328 + ], + "type": "text", + "content": "Figure 6. Failure cases of our SSL pre-training framework based on MirrorNet. Our SSL pre-training framework fails to tackle the intrinsic limitation of existing mirror detection methods. The model pre-trained by our SSL pre-training framework still has wrong predictions when the input images are challenging (e.g., containing mirror-like regions in the first row, and insufficient relationship between mirror/non-mirror regions in the second row." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 350, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 350, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 350, + 545, + 399 + ], + "type": "text", + "content": "tect mirrors in some challenging cases when the input image contains ambiguous mirrors without the relationship of mirror reflection and sufficient contextual contrast, as shown in the second row of Figure 6." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 413, + 378, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 413, + 378, + 425 + ], + "spans": [ + { + "bbox": [ + 306, + 413, + 378, + 425 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 434, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 662 + ], + "type": "text", + "content": "In this paper, we have investigated how self-supervised learning (SSL) works with the mirror detection task. To the best of our knowledge, we are the first to explore SSL pre-training frameworks applied to mirror detection. We have found that the supervised ImageNet pre-training might not be the ideal way to extract backbone image features for mirror detection due to the discrepancy between general-propose representation and mirror-specific representation. We have also proposed a new SSL pre-training framework to pre-train the backbone network for mirror detection. Our SSL pre-training framework does not require any labeled data. It progressively (image-, patch- and pixel-level pre-training) models the relationship of mirror reflection, and obtains a better representation of mirrors in the pre-training stage. Experimental results show that the proposed SSL pre-training framework outperforms CNN-based state-of-the-art SSL pre-training frameworks, and even achieves better results compared with supervised ImageNet pre-training." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgments: The work described in this paper was partially supported by a GRF grant from the Research Grants Council of Hong Kong (Project No.: CityU 11211223)." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12234" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 92, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 147 + ], + "type": "text", + "content": "[1] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, pages 3674-3683, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 287, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 287, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 287, + 182 + ], + "type": "text", + "content": "[2] M. Braun, S. Krebs, F. Flohr, and D. M. Gavrila. Eurocity persons: A novel benchmark for person detection in traffic scenes. IEEE TPAMI, 41(8):1844-1861, Aug 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 184, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 184, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 184, + 287, + 228 + ], + "type": "text", + "content": "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 230, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 230, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 230, + 287, + 262 + ], + "type": "text", + "content": "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 264, + 287, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 287, + 287 + ], + "type": "text", + "content": "[5] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 289, + 287, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 289, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 53, + 289, + 287, + 322 + ], + "type": "text", + "content": "[6] MMSelfSup Contributors. MMSelfSup: Openmmlab self-supervised learning toolbox and benchmark. https://github.com/open-mmlab/mmselfsup, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 323, + 287, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 323, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 323, + 287, + 378 + ], + "type": "text", + "content": "[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 414 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 415, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 287, + 449 + ], + "type": "text", + "content": "[9] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 450, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 450, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 450, + 287, + 483 + ], + "type": "text", + "content": "[10] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In ICCV, 2015." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 485, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 287, + 529 + ], + "type": "text", + "content": "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 531, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 531, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 287, + 564 + ], + "type": "text", + "content": "[12] Christopher Funk and Yanxi Liu. Beyond planar symmetry: Modeling human perception of reflection and rotation symmetries in the wild. In ICCV, pages 793-803, 2017." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 566, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 287, + 621 + ], + "type": "text", + "content": "[13] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249–256. JMLR Workshop and Conference Proceedings, 2010." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 623, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 287, + 678 + ], + "type": "text", + "content": "[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[15] Huankang Guan, Jiaying Lin, and Rynson W.H. Lau. Learning semantic associations for mirror detection. In CVPR, pages 5941-5950, June 2022." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "text", + "content": "[18] Ruozhen He, Jiaying Lin, and Rynson WH Lau. Efficient mirror detection via multi-level heterogeneous learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 790-798, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "type": "text", + "content": "[19] Tianyu Huang, Bowen Dong, Jiaying Lin, Xiaohui Liu, Rynson WH Lau, and Wangmeng Zuo. Symmetry-aware transformer-based mirror detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 935–943, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 241, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 284 + ], + "type": "text", + "content": "[20] Peter Kohler, Alexandra Yakovleva, Alasdair Clarke, Yanxi Liu, and Anthony Norcia. Parametric responses to rotation symmetry in mid-level visual cortex. Journal of Vision, 15(12):1122-1122, 2015." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 285, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 545, + 318 + ], + "type": "text", + "content": "[21] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In ICLR, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 319, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 364 + ], + "type": "text", + "content": "[22] Jiaying Lin, Xin Tan, and Rynson WH Lau. Learning to detect mirrors from videos via dual correspondences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9109-9118, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 365, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 386 + ], + "type": "text", + "content": "[23] Jiaying Lin, Guodong Wang, and Rynson W. H. Lau. Progressive mirror detection. In CVPR, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 388, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 432 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 433, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 476 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 477, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 477, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 477, + 545, + 499 + ], + "type": "text", + "content": "[26] Zhiqiu Lin, Jin Sun, Abe Davis, and Noah Snavely. Visual chirality. In CVPR, pages 12295-12303, 2020." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 500, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 521 + ], + "type": "text", + "content": "[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2018." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 523, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 577 + ], + "type": "text", + "content": "[28] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 578, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 612 + ], + "type": "text", + "content": "[29] Jiaqi Tan, Weijie Lin, Angel X Chang, and Manolis Savva. Mirror3d: Depth refinement for mirror surfaces. In CVPR, pages 15990-15999, 2021." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "type": "text", + "content": "[30] Xin Tan, Jiaying Lin, Ke Xu, Chen Pan, Lizhuang Ma, and Rynson W. H. Lau. Mirror detection with the visual chirality cue. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 658, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 680 + ], + "type": "text", + "content": "[31] Christopher W Tyler. Human symmetry perception and its computational analysis. Psychology Press, 2003." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[32] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In CVPR, 2021." + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12235" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 578 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "type": "text", + "content": "[33] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "type": "text", + "content": "[34] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, pages 418-434, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 286, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 286, + 174 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 286, + 174 + ], + "type": "text", + "content": "[35] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, pages 1492-1500, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 175, + 286, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 286, + 217 + ], + "type": "text", + "content": "[36] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, pages 9653-9663, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 220, + 286, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 286, + 241 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 286, + 241 + ], + "type": "text", + "content": "[37] Xin Yang, Haiyang Mei, Ke Xu, Xiaopeng Wei, Baocai Yin, and Rynson W.H. Lau. Where is my mirror? In ICCV, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 243, + 286, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 243, + 286, + 275 + ], + "spans": [ + { + "bbox": [ + 49, + 243, + 286, + 275 + ], + "type": "text", + "content": "[38] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. Learning with noisy labels for robust point cloud segmentation. In ICCV, pages 6443-6452, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 277, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 277, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 277, + 286, + 319 + ], + "type": "text", + "content": "[39] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. Robust point cloud segmentation with noisy annotations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7696-7710, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 322, + 286, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 322, + 286, + 364 + ], + "spans": [ + { + "bbox": [ + 49, + 322, + 286, + 364 + ], + "type": "text", + "content": "[40] Shuquan Ye, Yujia Xie, Dongdong Chen, Yichong Xu, Lu Yuan, Chenguang Zhu, and Jing Liao. Improving commonsense in vision-language models via knowledge graph riddles. In CVPR, pages 2634-2645, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 366, + 286, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 366, + 286, + 398 + ], + "spans": [ + { + "bbox": [ + 49, + 366, + 286, + 398 + ], + "type": "text", + "content": "[41] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 400, + 286, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 400, + 286, + 432 + ], + "spans": [ + { + "bbox": [ + 49, + 400, + 286, + 432 + ], + "type": "text", + "content": "[42] Xiaohang Zhan, Ziwei Liu, Ping Luo, Xiaou Tang, and Chen Loy. Mix-and-match tuning for self-supervised semantic segmentation. In AAAI, volume 32, 2018." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 434, + 286, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 434, + 286, + 466 + ], + "spans": [ + { + "bbox": [ + 49, + 434, + 286, + 466 + ], + "type": "text", + "content": "[43] Xiaohang Zhan, Jiahao Xie, Ziwei Liu, Yew-Soon Ong, and Chen Change Loy. Online deep clustering for unsupervised representation learning. In CVPR, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 468, + 286, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 468, + 286, + 500 + ], + "spans": [ + { + "bbox": [ + 49, + 468, + 286, + 500 + ], + "type": "text", + "content": "[44] Nanxuan Zhao, Zhirong Wu, Rynson WH Lau, and Stephen Lin. What makes instance discrimination good for transfer learning? In ICLR, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 502, + 286, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 502, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 49, + 502, + 286, + 533 + ], + "type": "text", + "content": "[45] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, , Hutchuan Lu, and Xiang Ruan. Self-supervised pretraining for rgb-d salient object detection. In AAAI, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 536, + 286, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 536, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 49, + 536, + 286, + 578 + ], + "type": "text", + "content": "[46] Ying Zheng, Yiyi Zhang, Xiaogang Xu, Jun Wang, and Hongxun Yao. Visual chirality meets freehand sketches. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1544-1548. IEEE, 2021." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12236" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/fa7524e1-a76d-4844-9361-43c8b8046bf9_content_list.json b/2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/fa7524e1-a76d-4844-9361-43c8b8046bf9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f3591f393f0e269fd486b3621457dd1055df0662 --- /dev/null +++ b/2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/fa7524e1-a76d-4844-9361-43c8b8046bf9_content_list.json @@ -0,0 +1,1876 @@ +[ + { + "type": "text", + "text": "SemARFlow: Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving", + "text_level": 1, + "bbox": [ + 94, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shuai Yuan, Shuzhi Yu, Hannah Kim, and Carlo Tomasi \nDuke University", + "bbox": [ + 261, + 203, + 707, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{shuai, shuzhiyu, hannah, tomasi}@cs.duke.edu", + "bbox": [ + 285, + 241, + 683, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unsupervised optical flow estimation is especially hard near occlusions and motion boundaries and in low-texture regions. We show that additional information such as semantics and domain knowledge can help better constrain this problem. We introduce SemARFlow, an unsupervised optical flow network designed for autonomous driving data that takes estimated semantic segmentation masks as additional inputs. This additional information is injected into the encoder and into a learned upsampler that refines the flow output. In addition, a simple yet effective semantic augmentation module provides self-supervision when learning flow and its boundaries for vehicles, poles, and sky. Together, these injections of semantic information improve the KITTI-2015 optical flow test error rate from $11.80\\%$ to $8.38\\%$ . We also show visible improvements around object boundaries as well as a greater ability to generalize across datasets. Code is available at https://github.com/duke-vision/semantic-unsup-flow-release.", + "bbox": [ + 75, + 324, + 473, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 643, + 209, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Optical flow estimation, i.e., pixel-level motion tracking across video frames, has broad applications in many computer vision tasks that include object tracking [61], video editing [14, 31], and autonomous driving [6, 58].", + "bbox": [ + 75, + 671, + 468, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thanks to the success of deep convolutional neural networks [34, 17, 36] and transformer networks [33, 69, 88, 28] in computer vision, many top-performing supervised optical flow networks have been proposed in recent years [10, 65, 22, 68, 85], in which ground-truth labels supervise training. However, real optical flow is hard to label, so most supervised methods train (or at least pre-train) on synthetic datasets [10, 44, 4, 56] which makes them hard to adapt to real applications due to the significant gap between synthetic and real data [70, 29].", + "bbox": [ + 75, + 733, + 468, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to label scarcity, unsupervised training of optical", + "bbox": [ + 96, + 885, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9e65a6cce7dfec507031a1407f7e49721073b9180ae490256c7b968ba3697ddc.jpg", + "image_caption": [ + "Figure 1. An example on KITTI-2015 test set (sample #63). Only one input frame is shown for conciseness. Our SemARFlow takes additional semantic segmentation inputs (estimated by an off-the-shelf model) and outputs much sharper flow around semantic boundaries." + ], + "image_footnote": [], + "bbox": [ + 501, + 290, + 671, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/343f36ac10bb396cab4abe94b28fbb130cf086e612f0ae4d6d44fd01bbc59452.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 290, + 727, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/81f42782b8b7e34d6a081d106bfccb2d41438995034e911df052f59146cd50c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 291, + 779, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/85dc6b3c2243ce70f0fa7c131cc53de504af6f1b5e13f883b0520d0b57d3ac48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 781, + 291, + 834, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b36e7fc92d3acc591e035ba081195e897be82f0c0d8276200009904df42750ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 834, + 291, + 888, + 414 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "flow estimators [81] instead uses loss terms based on the assumptions of constant brightness and smooth flow [45, 27, 42]. Some self-supervision techniques have also been studied to enhance model performance [38, 39, 37]. Unsupervised training makes it possible to train flow networks directly on large real datasets from the target domain.", + "bbox": [ + 496, + 529, + 893, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Even so, unsupervised optical flow estimation is a poorly constrained problem. Brightness is not constant in regions with occlusions [72] or on shiny surfaces [43], nor is it smooth across motion boundaries [30, 82]. Moreover, points in regions with poor texture content [59] or in dark shadows [58] are difficult to track as they are easily confused with neighboring points (the so-called aperture problem). This dearth of reliable constraints makes the unsupervised training of flow networks especially challenging.", + "bbox": [ + 496, + 623, + 893, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A natural way to address this issue is to inject additional constraints in the form of semantics and domain knowledge. For example, in autonomous driving applications [84, 6], we have clear expectations about object types and layouts of the scene, as well as prior knowledge of how each type of object typically moves. We focus on this application domain and show that this additional information helps the network achieve better flow results.", + "bbox": [ + 496, + 761, + 893, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To make an autonomous driving system work well in", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9566", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "reality, it is best to train it on real data. However, annotating real driving video with optical flow labels is expensive [83], as it requires careful synchronization and calibration of diverse sensors including cameras, LiDAR, and GPS/IMU [16], aided by some manual annotation and curation based on CAD models of moving objects [46].", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, annotating semantic labels seems much more feasible, and indeed semantic labels are available in most (if not all) existing driving datasets. We consider semantic segmentation because it provides semantics at the pixel level, the same level as optical flow. As one of the most popular and well-studied tasks in modern computer vision, semantic segmentation [40, 47, 87] has been extensively adopted for autonomous driving systems. In this paper, we show that adding semantic segmentation inputs helps improve unsupervised optical flow performance significantly.", + "bbox": [ + 75, + 184, + 470, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we first infer semantic segmentation maps using an off-the-shelf model [89], which of course is trained with semantic labels. An encoder with semantic map input is used to aggregate image and semantic features (Sec. 3.1), and a learned upsampler is added into the iterative decoder to refine flow around object boundaries given semantic inputs (Sec. 3.2). We also propose a simple yet effective semantic augmentation module for self-supervision, which provides realistic augmentations specific to the vehicles, poles, and sky classes based on domain knowledge and segmentation maps (Sec. 3.3). An occluder cache is implemented to improve efficiency (Sec. 3.4). Semantic augmentation provides challenging samples for self-supervision, which help train flow better in occluded regions and around foreground object boundaries.", + "bbox": [ + 75, + 337, + 470, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, by injecting semantic segmentation inputs, our SemARFlow network achieves significantly better unsupervised flow results both quantitatively (Sec. 4.3) and qualitatively (Sec. 4.4). Adapted from ARFlow [37], SemARFlow reduces KITTI-2015 [46] test error from $11.80\\%$ to $8.38\\%$ and out-performs the current unsupervised state-of-the-art UPFlow [42] $(9.38\\%)$ by a clear margin. The example in Fig. 1 also demonstrates visible improvements around object boundaries. The effectiveness of each module is justified through an extensive ablation study (Sec. 4.5) and improvement analysis (Sec. 4.6). In addition to performance boost, unsupervised flow networks with additional semantic inputs generalize better across different datasets (Sec. 4.7).", + "bbox": [ + 75, + 566, + 468, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our research is essentially novel compared to previous approaches. Some early work incorporates semantics in traditional energy-minimization methods [41, 18] for optical flow through geometric constraints such as piece-wise rigid motion and planar surface motion [57, 21, 1, 75]. In comparison, to the best of our knowledge, we are the first to inject semantics into the unsupervised training of recent optical flow networks. Some research also trains a network for segmentation and optical flow jointly [9]. In contrast, we", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "leverage existing, separately trained segmentation systems both because they are available and because modularity—separating segmentation from flow estimation—is important for the development of large real-application systems such as autonomous driving.", + "bbox": [ + 496, + 90, + 890, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows.", + "bbox": [ + 517, + 167, + 818, + 181 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, we are the first to explore adding semantic inputs to assist the unsupervised training of deep optical flow networks.", + "- We propose a simple yet effective network called SemARFlow that achieves state-of-the-art results both quantitatively and qualitatively. Our model works well on real-life occlusions and yields sharp motion boundaries around objects.", + "- We provide full training and inference code as well as trained models to encourage follow-up research." + ], + "bbox": [ + 517, + 190, + 890, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 375, + 637, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Supervised optical flow Since the introduction of the pioneering deep optical flow network FlowNet [10], more and more top-performing CNN-based flow estimators have been proposed over the years [23, 49, 65, 20, 22, 68]. Recently, vision transformer networks and attention mechanism have also been applied to this problem, and these have achieved state-of-the-art performance on benchmark datasets [76, 26, 85, 64, 77]. The supervised methods are often pre-trained on large synthetic datasets such as FlyingChairs [10] and FlyingThings3D [44] before fine-tuning on the target dataset. However, there is a clear gap between the artificially generated data and real scenarios.", + "bbox": [ + 496, + 398, + 890, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised optical flow Due to the lack of ground-truth labels, unsupervised optical flow estimation uses surrogate losses such as photometric loss and smoothness loss to supervise training [81, 55]. To tackle the issues around occlusion regions, various methods have been proposed including estimated occlusion masks [72], forward-backward consistency [45], and multi-frame fusion [25, 53]. To better upsample the flow in the decoder, UPFlow [42] additionally predicts a confidence map and an interpolation flow to guide flow refinement and has become the state-of-the-art unsupervised flow method.", + "bbox": [ + 496, + 598, + 890, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some latest research has also explored the use of self-supervision to enhance flow prediction. Early methods apply the knowledge distillation technique to train a two-stage teacher-student network [38, 39]. ARFlow [37] further improves this idea by generating reliable self-supervision signals from data transformations, while merging the two training stages into single-stage training with one added loss term. SimFlow [24] replaces the handcrafted features with deep self-supervised features to measure similarity in the", + "bbox": [ + 496, + 763, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9567", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "unsupervised losses. SMURF [63] utilizes a RAFT-like structure [68] and applies multi-frame self-supervised training with many technical improvements. Our SemARFlow also uses self-supervision but with the guidance of semantic segmentation, which is much more realistic than guidance without semantics.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semantic segmentation Semantic segmentation classifies each pixel of the given image into semantic objects. Fully Convolutional Network (FCN) [40] is one of the early CNN-based segmentation methods. It takes inputs of arbitrary sizes and outputs dense pixel-level predictions, becoming one of the popular backbone architectures for follow-up work. Deconvolution network [47] is also proposed to better recover low-level details of the prediction. One main challenge for these systems is that they lack global scene information. Subsequent work addresses this issue by enlarging the receptive field of the network with global pyramid pooling layers as in PSPNet [86], hybrid dilated convolutions [71], and a fast-down-sampling strategy [80, 79]. Attention modules have also shown to help in semantic segmentation by capturing full-image dependencies of all pixels [19, 87] or the semantic inter-dependencies across spatial and channel dimensions [12].", + "bbox": [ + 75, + 200, + 468, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There also exists extensive work on semantic segmentation in the context of autonomous driving [2, 67, 13, 5, 89], thanks to the publication of large-scale driving datasets [8, 46]. To better train the network on a coarsely labeled dataset like KITTI [46], Zhu et al. use a video prediction network SDCNet [52] to synthesize new training samples with relaxed label propagation [89]. Due to their good performance on KITTI [46], we utilize their network models to infer semantic inputs for all our experiments.", + "bbox": [ + 75, + 458, + 468, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Combining semantics and optical flow Though there has been much progress on both semantic segmentation and optical flow estimation, the semantic optical flow problem (how to exploit semantics to help optical flow estimation) has received limited attention in recent years, and the current best results are thus much outdated. Some early methods incorporate semantics through geometric constraints to refine flow on various semantic regions, such as planar regions (using homographies) [57], static regions (using rigid camera motion and epipolar constraints) [21, 75], and rigid objects (estimating rigid motion for each object instance) [1]. However, most methods are traditional flow methods based on energy minimization, where an initial flow estimate is usually needed and semantics is mostly used for refinement. In comparison, we explore adapting latest unsupervised optical flow networks to leverage semantic inputs in one single stage of estimation.", + "bbox": [ + 75, + 613, + 468, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Apart from using semantics to help flow, some research has also explored using optical flow to help seman", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tic segmentation [51, 35], or to train both tasks jointly [9]. There are also studies on exploiting semantics on some other correspondence matching tasks such as stereo matching [74, 78] and 3D scene flow estimation, where some additional depth cues such as stereo camera inputs [54, 11] or point clouds [60] are needed.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 196, + 589, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our network is adapted from the two-frame version of ARFlow [37], which uses a light-weight PwCNet [65] as its backbone. The inputs are two consecutive frames $I_1, I_2 \\in \\mathbb{R}^{H \\times W \\times 3}$ as well as their semantic segmentation maps $S_1, S_2 \\in \\{0, 1, \\dots, c\\}^{H \\times W}$ , where the number of classes $c$ is 19 as we use the Cityscapes format [8]. A detailed diagram of the structure of our network can be found in the Appendix.", + "bbox": [ + 496, + 222, + 890, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Semantic encoder", + "text_level": 1, + "bbox": [ + 500, + 354, + 671, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first inject semantic information into the encoder. As shown in Fig. 2(a), shared by each frame $i \\in \\{1,2\\}$ , separate convolutional layers extract features from image $I_{i}$ and semantics $S_{i}$ (one-hot encoded). Features from the two pipelines are concatenated and fed to additional layers. Features at different resolutions $(H / 2^{l}, W / 2^{l})$ form a pyramid $\\{f_{i}^{(l)} \\mid 2 \\leq l \\leq 6\\}$ . The semantic information in these features helps delineate objects in dark shadows, where appearance is more homogeneous.", + "bbox": [ + 496, + 377, + 890, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Iterative decoder with a learned upsampler", + "text_level": 1, + "bbox": [ + 500, + 526, + 867, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following [22, 37], an iterative residual refinement decoder starts from zero estimate $F_{1\\rightarrow 2}^{(7)} = 0$ . For iteration $l \\in \\{6,5,4,3,2\\}$ , the decoder refines feature map $F_{1\\rightarrow 2}^{(l + 1)}$ into $F_{1\\rightarrow 2}^{(l)}$ based on features $f_{1}^{(l)}, f_{2}^{(l)}$ (Fig. 2(b)).", + "bbox": [ + 496, + 550, + 890, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More specifically, $F_{1\\rightarrow 2}^{(l + 1)}$ is upsampled to match the resolution and used to warp $f_2^{(l)}$ to yield warped feature $\\hat{f}_1^{(l)}$ . Correlation volumes are computed between $f_1^{(l)}$ and $\\hat{f}_1^{(l)}$ . A one-by-one convolutional layer $C^{(l)}$ compresses the number of channels to a fixed number so that the same layer can be reused across all iterations as proposed by [22]. A flow estimator network predicts a flow residual to be added to the current estimate, and a context network then aggregates flow information spatially and refines the current flow again.", + "bbox": [ + 496, + 619, + 890, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A learned upsampler network upsamples the final output $F_{1\\rightarrow 2}^{(2)}$ in our system. This is different from ARFlow [37], which simply uses four-fold bilinear interpolation, making the final flow boundaries blurry. In contrast, our model learns to sharpen flow boundaries based on the semantic inputs, which have clear boundaries around moving objects. To this end, we add a convex upsampler network similar to the one in RAFT [68]. Different from UPFlow [42],", + "bbox": [ + 496, + 777, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9568", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1f578db99d3ccda59acd52464574df776530c318c409b21f7fd6dae0bee7b43c.jpg", + "image_caption": [ + "(a) shared pyramidal encoder $(i\\in \\{1,2\\})$" + ], + "image_footnote": [], + "bbox": [ + 94, + 97, + 343, + 223 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c30e1a5c14116a469523b3f00f526371d8f8a59ad87c073d3cadb07e3d7d3548.jpg", + "image_caption": [ + "(b) one iteration (at the $l$ th-level) of the iterative decoder $(l\\in \\{6,5,4,3,2\\})$" + ], + "image_footnote": [], + "bbox": [ + 375, + 98, + 870, + 219 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4b0eca463840993e1e56b7ce31b8aa97cb087ccf3349b518dc7346f392bc4ee8.jpg", + "image_caption": [ + "Figure 2. Network structure. See text in Sec. 3.1 and Sec. 3.2 for explanations. More detailed diagrams are in appendix.", + "Figure 3. Illustration of semantic augmentation as self-supervision. See text in Sec. 3.3 for details." + ], + "image_footnote": [], + "bbox": [ + 84, + 284, + 888, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "our learned upsampler is only used to upsample from internal estimate $F_{1\\rightarrow 2}^{(l)}$ to output flow $U_{1\\rightarrow 2}^{(l)}$ , but not when we upsample $F_{1\\rightarrow 2}^{(l + 1)}$ at the first step in the decoder iteration. The upsampled $U_{1\\rightarrow 2}^{(l)}$ has resolution $(H / 2^{l - 2},W / 2^{l - 2})$ , so $U_{1\\rightarrow 2} = U_{1\\rightarrow 2}^{(2)}$ (with the original resolution) is the final flow prediction of the network.", + "bbox": [ + 75, + 604, + 468, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Semantic augmentation as self-supervision", + "text_level": 1, + "bbox": [ + 76, + 720, + 439, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ARFlow has a very effective in-network augmentation module that samples random transformations $\\mathcal{T}_{\\theta_1},\\mathcal{T}_{\\theta_2}$ of the flow prediction $U_{1\\rightarrow 2}$ in the first pass of the network and then uses the transformed images $\\hat{I}_1 = \\mathcal{T}_{\\theta_1}(I_1),\\hat{I}_2 =$ $\\mathcal{T}_{\\theta_2}(I_2)$ in a second pass. The prediction $U_{1\\rightarrow 2}$ is also transformed accordingly and used to self-supervise the output of the second pass. See [37] for details.", + "bbox": [ + 75, + 747, + 468, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We retain the augmentation module but make a third pass of the network using semantics-transformed inputs for self-supervision in addition to the ARFlow augmentations of ap", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pearance (e.g., color jitter, random noise) and spatial transformations (e.g., random rotation, random rescaling). The idea behind semantic augmentation is to blend in real object motions across samples.", + "bbox": [ + 496, + 604, + 890, + 665 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To this end (see Fig. 3), we carve objects from other samples based on their semantic maps and paste them as moving foreground objects into the current sample $I_{1}, I_{2}$ , thereby producing augmented images $\\tilde{I}_{1}, \\tilde{I}_{2}$ . The segmentation maps $S_{1}, S_{2}$ and the first pass output flow $U_{1 \\to 2}$ are transformed accordingly, and the transformed $\\tilde{U}_{1 \\to 2}$ self-supervises the third-pass network outputs $\\tilde{U}_{1 \\to 2}' = \\mathcal{F}(\\tilde{I}_{1}, \\tilde{I}_{2}, \\tilde{S}_{1}, \\tilde{S}_{2})$ , where $\\mathcal{F}$ is the flow network.", + "bbox": [ + 496, + 667, + 892, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to occlusion hallucination in SelFlow [39], our semantic augmentation also creates new occlusions and uses the reliable non-occluded flow from the first pass to self-supervise the flow on those newly occluded pixels. However, rather than superpixels filled with noise, we create occlusions with real objects, which are much more realistic.", + "bbox": [ + 496, + 791, + 890, + 882 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Moreover, since we determine the motion of each oc", + "bbox": [ + 517, + 885, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9569", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cluder, we can use its known optical flow for motion self-supervision. As illustrated in Fig. 3, the occluder motion can be very different from that of the background. This trains the model to assume less smoothness around highly-dynamic objects and to output sharper motion estimates near motion boundaries. These hallucinated but realistic occluders also provide more examples of actively moving objects, which are relatively scarce in the original datasets.", + "bbox": [ + 75, + 90, + 472, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Vehicles** We augment vehicle classes (car, truck, bus, train) as they are major sources of errors in previous models (see Tab. 5 for statistics). We group these four classes because they are easily confused with each other in the semantic input. To carve out a single car instance with high probability, we find connected components of vehicle segmentation regions using OpenCV [3] and accept masks whose bounding boxes are 50-300 pixels wide and 50-150 pixels high. We also drop masks that occupy less than $60\\%$ area of their bounding boxes to remove largely occluded vehicles.", + "bbox": [ + 75, + 244, + 472, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Poles We augment pole classes (pole, traffic light, traffic sign) not only because they are common foreground objects that cause occlusions, but also because their motions are usually poorly estimated due to their thin shapes. We use a 200-pixel wide, image-height sliding window to scan the segmentation of these pole classes and find the window with the most pole pixels. If more than $10\\%$ pole pixels occur in it, that object mask is stored. This produces groups of nearby poles rather than just individual poles, which may not provide enough augmentation by themselves.", + "bbox": [ + 75, + 426, + 472, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sky We also augment the sky through prior knowledge that flow in the sky should be very small. We shrink the sky flow estimate by half before using it for self-supervision. As shown in Fig. 3, the middle tree part in $U_{1\\rightarrow 2}$ is very blurry, with a large part of sky mixed in there. After shrinking sky flow by half, the tree in $\\tilde{U}_{1\\rightarrow 2}$ now has sharper flow around the boundary between tree and sky. Using this as self-supervision helps the model sharpen the motion boundaries between sky and other regions.", + "bbox": [ + 75, + 611, + 468, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Another way to apply this prior knowledge is to postprocess, i.e., shrink the final output flow in the sky region, instead of using the shrunk flow to self-supervise. However, the latter approach is preferable because it allows the network to balance prior knowledge (self-supervision) with the current sample observation (photometric loss). Our experiments do show cases where some small objects (such as some overhead electric power lines) are misclassified as part of the sky in the semantic input. In these cases, the network learns to rely less on prior knowledge.", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.Occluder cache", + "text_level": 1, + "bbox": [ + 500, + 90, + 656, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, we maintain a cache of occluding objects for semantic augmentation. We sample objects from the cache to augment the current batch and then push any new objects found in the batch into the cache, with random replacement if the cache is full.", + "bbox": [ + 496, + 113, + 893, + 188 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, after we load a sample for training, we first use the current model to infer its optical flow and then search within the sample for any cars (or poles) that meet the standards defined in Sec. 3.3 as \"cutouts\". For each cutout, we first find the flow vectors on all pixels that belong to that cutout and then compute the mean flow for later reference. We store the current sample, the cutout mask, and the mean flow together in the cache.", + "bbox": [ + 496, + 189, + 893, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Later on, when we randomly select a cutout to augment another sample, we also retrieve the stored mean flow of this cutout, which is then augmented by random rescaling (by 0.8-1.5 times) and reversing (with $50\\%$ probability). We use the augmented mean flow to translate the entire cutout object as an occluder to generate a new sample. For realism, occluders are pasted to the same location they occupied in their image of origin. Holding occluders in a cache provides a random mixture of samples across multiple batches and makes semantic augmentation very efficient.", + "bbox": [ + 496, + 310, + 893, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Loss functions", + "text_level": 1, + "bbox": [ + 500, + 469, + 648, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Photometric loss In the first pass, forward and backward flow $U_{1\\rightarrow 2}^{(l)}, U_{2\\rightarrow 1}^{(l)}$ at each level are predicted and occlusion masks $O_{1\\rightarrow 2}^{(l)}, O_{2\\rightarrow 1}^{(l)}$ are computed by forward-backward consistency [45]. Frames are warped by", + "bbox": [ + 496, + 492, + 893, + 559 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {i} ^ {\\prime (l)} (\\pmb {p}) = I _ {j} ^ {(l)} (\\pmb {p} + U _ {i \\rightarrow j} ^ {(l)} (\\pmb {p})), \\quad (i, j \\in \\{1, 2 \\})\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 565, + 854, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I_{i}^{(l)}$ is $I_{i}$ down-sampled to the $l$ -th scale, and $\\pmb{p}$ denotes pixel coordinates at that scale. Following [37], three measures, namely $L_{1}$ -distance $(\\rho_{1})$ , structural similarity (SSIM) [73] $(\\rho_{2})$ , and census loss [45] $(\\rho_{3})$ are linearly combined to measure photometric differences between $I_{i}^{(l)}$ and $I_{i}^{\\prime(l)}$ . Occlusion regions are masked out and both forward and backward directions are taken into account at each level. The final photometric loss is", + "bbox": [ + 496, + 595, + 893, + 726 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\mathrm {p h}} = \\frac {1}{2} \\sum_ {(i, j) \\in \\{(1, 2), (2, 1) \\}} \\sum_ {l = 2} ^ {6} a _ {l} \\sum_ {k = 1} ^ {3} c _ {k} \\rho_ {k} \\left(I _ {i} ^ {(l)}, I _ {i} ^ {\\prime (l)}, O _ {i \\rightarrow j} ^ {(l)}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 732, + 890, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Smoothness loss Unlike most previous work, we do not include a smoothness loss because we find it to conflict with our learned upsampler. See ablation study in Sec. 4.5.", + "bbox": [ + 496, + 806, + 893, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Augmentation loss As in ARFlow [37], our second forward pass computes the $\\mathrm{L}_1$ -distance between the trans-", + "bbox": [ + 496, + 869, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9570", + "bbox": [ + 480, + 944, + 517, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a0fa49c82b26e757e90ee9676028000104af60fc54e15a78edce27fa6d41aa86.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. KITTI benchmark results (EPE/px and Fl/%). Metrics evaluated at 'all' (all pixels, default for EPE), 'noc' (non-occlusions), 'bg' (background), and 'fg' (foreground). Key metrics (used in official ranking) are underlined. Our 'baseline' is an adapted ARFlow with added learned upsampler and no smoothness loss. +enc means adding semantic encoder; +aug means adding semantic augmentation. () means evaluation data used in training. - means unavailable. $\\dagger$ denotes models with semantic inputs. For all metrics, lower is better." + ], + "table_body": "
MethodTrainTest
2012 EPE2015 EPE2012 Fl-nocEPEFl-all2015 Fl-nocFl-bgFl-fg
supervisedPWC-Net+ [66]-(1.50)3.361.47.724.917.697.88
IRR-PWC [22]-(1.63)3.211.67.654.867.687.52
RAFT [68]-(0.63)--5.103.074.746.87
Separable Flow [85]-(0.69)--4.532.784.255.92
unsupervisedSelFlow [39]1.694.844.312.214.199.6512.6821.74
SimFlow [24]-5.19--13.388.2112.6017.27
ARFlow [37]1.442.85-1.811.80---
UFlow [27]1.682.714.261.911.138.419.7817.87
UPFlow [42]1.272.45-1.49.38---
Ours (baseline)1.392.614.301.79.896.988.8215.21
Ours (+enc)†1.292.423.971.58.993.978.1913.01
Ours (+enc +aug)†1.282.183.901.58.383.907.4812.91
", + "bbox": [ + 163, + 87, + 807, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "formed flow $\\hat{U}_{1\\rightarrow 2}$ and the second pass output $\\hat{U}_{1\\rightarrow 2}^{\\prime} = \\mathcal{F}(\\hat{I}_1,\\hat{I}_2,\\hat{S}_1,\\hat{S}_2)$ for each pixel $\\pmb{p}$ as", + "bbox": [ + 75, + 409, + 468, + 441 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {D} (\\boldsymbol {p}) = \\left\\| \\hat {U} _ {1 \\rightarrow 2} (\\boldsymbol {p}) - \\hat {U} _ {1 \\rightarrow 2} ^ {\\prime} (\\boldsymbol {p}) \\right\\| _ {1}.\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 449, + 388, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This loss is then averaged over the transformed non-occluded region where self-supervision $\\hat{U}_{1\\rightarrow 2}$ is accurate", + "bbox": [ + 75, + 478, + 468, + 508 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\mathrm {a r}} = \\frac {\\sum_ {\\boldsymbol {p}} (1 - \\hat {O} _ {1 \\rightarrow 2} (\\boldsymbol {p})) \\hat {D} (\\boldsymbol {p})}{\\sum_ {\\boldsymbol {p}} (1 - \\hat {O} _ {1 \\rightarrow 2} (\\boldsymbol {p}))}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 518, + 468, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For semantic augmentation (third pass), since we are also doing self-supervision with a new generated sample, we use a similar loss definition", + "bbox": [ + 75, + 569, + 468, + 613 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\mathrm {a u g}} = \\frac {\\sum_ {\\boldsymbol {p}} \\left(1 - \\tilde {O} _ {1 \\rightarrow 2} (\\boldsymbol {p})\\right) \\tilde {D} (\\boldsymbol {p})}{\\sum_ {\\boldsymbol {p}} \\left(1 - \\tilde {O} _ {1 \\rightarrow 2} (\\boldsymbol {p})\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 621, + 468, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tilde{D} (\\pmb {p})$ is the distance between the semanticaugmented flow $\\tilde{U}_{1\\rightarrow 2}$ and the third-pass output $\\tilde{U}_{1\\rightarrow 2}^{\\prime}$", + "bbox": [ + 75, + 672, + 468, + 705 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {D} (\\boldsymbol {p}) = \\left\\| \\tilde {U} _ {1 \\rightarrow 2} (\\boldsymbol {p}) - \\tilde {U} _ {1 \\rightarrow 2} ^ {\\prime} (\\boldsymbol {p}) \\right\\| _ {1},\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 713, + 388, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and the augmented mask $\\tilde{O}_{1\\rightarrow 2}$ is now computed by", + "bbox": [ + 75, + 741, + 423, + 758 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {O} _ {1 \\rightarrow 2} = 1 - \\max (1 - O _ {1 \\rightarrow 2}, M)\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 766, + 387, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "because we penalize on both the originally non-occluded region $1 - O_{1\\rightarrow 2}$ and the pasted foreground object mask $M$ (of which we know the true motion).", + "bbox": [ + 75, + 795, + 468, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Final loss (with $\\lambda = 0.02$ as in ARFlow [37]):", + "bbox": [ + 75, + 858, + 398, + 875 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\ell = \\ell_ {\\mathrm {p h}} + \\lambda \\left(\\ell_ {\\mathrm {a r}} + \\ell_ {\\mathrm {a u g}}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 883, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 409, + 632, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 433, + 601, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We mainly use KITTI [15, 46] and Cityscapes [8] datasets for experiments. Following [37], we train our model first on KITTI raw sequences [46] (55.7k samples) and then fine-tune on KITTI-2015 multi-view extension [46] (11.8k samples). We validate our model using KITTI-2015 train [46] (200 samples) and KITTI-2012 train set [15] (194 samples) since they are the only sets that have optical flow labels. We also train on Cityscapes [8] sequences (83.3k samples) to test model generalization ability, and we sample every other frame to match its frame rate with KITTI. All semantic segmentation inputs are estimated using an off-the-shelf model [89] with a DeepLabV3Plus [7] backbone, which achieves $83.45\\%$ mean IoU on Cityscapes and $72.82\\%$ mean IoU on KITTI.", + "bbox": [ + 496, + 455, + 890, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation details", + "text_level": 1, + "bbox": [ + 500, + 676, + 714, + 693 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement the model in PyTorch [48] $^{1}$ . We use the Adam optimizer [32] with $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ and batch size 4. We first train on KITTI raw sequences [46] for 100k iterations with a fixed learning rate 0.0002 and then train on KITTI multi-view extension set [15, 46] for another 100k iterations using OneCycleLR schedule [62] with maximum learning rate 0.0004 and linear annealing.", + "bbox": [ + 496, + 699, + 890, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For data augmentation, we include random horizontal flipping and swapping of the input frames. We resize the inputs to $256 \\times 832$ before feeding into the network. The photometric loss weight for each scale $a_{l}$ ( $2 \\leq l \\leq 6$ ) in", + "bbox": [ + 496, + 805, + 890, + 867 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code and instructions are available at https://github.com/duke-vision/semantic-unsup-flow-release.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9571", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1837df44ffbaed2651e485dc2dbce8ffd2943490d301b3338d54190c2c707589.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFl-allFl-nocFl-bgFl-fg
JFS [21]17.079.8115.922.92
SOF [57]16.8110.8614.6327.73
MRFlow [75]12.198.8610.1322.52
Bai et al. [1]11.628.758.6126.69
Ours (final)8.383.907.4812.91
", + "bbox": [ + 101, + 88, + 444, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Eq. (1) are set as 1, 1, 1, 1, 0. The weights for three photometric distance measures $\\rho_{k}$ ( $1 \\leq k \\leq 3$ ) in Eq. (1) are set as 0.15, 0.85, 0 for the first 50k iterations and 0, 0, 1 afterwards. We start the appearance and spatial augmentation (second pass as in ARFlow [37]) at 50k iterations, and semantic augmentation (third pass) after 150k iterations.", + "bbox": [ + 76, + 267, + 467, + 357 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Benchmark testing results", + "text_level": 1, + "bbox": [ + 76, + 378, + 313, + 395 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As common practice, we evaluate optical flow predictions based on two error measurements, Fl (error rate) and EPE (mean $\\mathrm{L}_2$ distance). When computing Fl, the estimate of each pixel is considered correct if the error is smaller than 3 pixels or $5\\%$ of the magnitude of ground-truth flow [46].", + "bbox": [ + 76, + 406, + 467, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 1, our semantic modules (both semantic encoder and semantic augmentation) improve performance on both KITTI-2012 and KITTI-2015 sets on all metrics. Our final model achieves $8.38\\%$ F1-all error rate on KITTI-2015 test set, which is significantly better than ARFlow [37] $(11.80\\%)$ , from which we adapt. We also outperform the current state-of-the-art UPFlow [42] $(9.38\\%)$ by a clear margin. All these results strongly suggest that adding semantic inputs can help improve unsupervised flow estimation significantly.", + "bbox": [ + 75, + 484, + 467, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, Tab. 2 shows that our test results significantly outperform all previous semantic optical flow methods, most of which are based on traditional energy minimization. We are the first to apply semantic inputs to recent unsupervised flow networks, so we are able to push the state-of-the-art by a clear margin.", + "bbox": [ + 75, + 640, + 467, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Qualitative results", + "text_level": 1, + "bbox": [ + 76, + 752, + 253, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Some qualitative results are shown in Fig. 4. We can see that our flow outputs have very sharp boundaries and very clean object motion, especially around cars and poles. Our model successfully learns to adapt flow estimation to the semantic map input. Moreover, our model is able to handle very challenging samples where the foreground car motion is drastically different from the background due to the challenging samples we create in semantic augmentation.", + "bbox": [ + 75, + 779, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8e9b8866ad2171ccb810955a26d057356e880f8be6848a0e401c5ed180d3de35.jpg", + "table_caption": [ + "Table 2. KITTI-2015 test results (Fl/\\%) compared with other semantic optical flow methods. Metrics evaluated at 'all' (all pixels), 'noc' (non-occlusions), 'bg' (background), and 'fg' (foreground)." + ], + "table_footnote": [], + "table_body": "
upno smencaugFl-allEPE-allEPE-nocEPE-occ
10.362.902.076.97
19.922.691.896.57
29.832.651.866.40
39.752.611.856.54
49.752.641.856.55
✓10.222.831.976.75
✓✓8.872.611.856.04
✓✓38.262.421.735.79
✓✓✓8.802.481.606.64
✓✓3✓7.792.181.405.64
", + "bbox": [ + 506, + 88, + 885, + 244 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d6899810fa46e2fa8b8110cd20f6e09521321eab8a3028129ff4d8c259abe500.jpg", + "table_caption": [ + "Table 3. Ablation study on KITTI-2015 train set (EPE/px and F1%). 'up' means the learned upsampler; 'no sm' means turning off smoothness loss; 'enc' means the level at which we merge semantic features and image features in the encoder; 'aug' means semantic augmentation." + ], + "table_footnote": [], + "table_body": "
Options of augFl-allEPE-allEPE-nocEPE-occ
Ours (final)7.792.181.405.64
start from 100k7.922.191.405.41
vehicles only7.942.211.445.78
loss on new occ8.152.271.425.62
", + "bbox": [ + 501, + 348, + 890, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation study of different semantic augmentation options on KITTI-2015 train (EPE/px and F1/%). See text for explanations.", + "bbox": [ + 498, + 439, + 890, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 503, + 648, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also do ablation study to show the effectiveness of each of our proposed modules. We can see in Tab. 3 that adding a semantic encoder to the vanilla ARFlow does help, and the optimal number of encoder layer groups added is 3. Also, our learned upsampler improves results significantly once we turn off the smoothness loss. This makes sense because the upsampler learns to adapt to boundaries, which can be distracted by the smoothness loss. Moreover, both our semantic encoder and semantic augmentation module help improve the results further, which suggests the effectiveness of adding semantic inputs.", + "bbox": [ + 496, + 532, + 890, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Another ablation study is on different options for the semantic augmentation (Tab. 4). We try starting augmentation earlier from 100k iterations, which works slightly worse. There are mainly two reasons: (1) we use the first forward output to self-supervise our augmented output, so if we start early, the model will use poor output to self-supervise the augmented pass; and (2) our semantic augmentation creates very challenging samples with objects moving very differently from the background, so exposing these hard samples to the model too early may make it hard to train. Apart from this, we also try only using vehicles to augment as well as focusing loss on the newly occluded region, which are both inferior to our final version.", + "bbox": [ + 496, + 704, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9572", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/07c59636c83be6777bb7d26b345ec42e13f6245ab9cf5eed0a798a8bc2146132.jpg", + "image_caption": [ + "Image Input" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 241, + 284 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cbb5321d0383d9cd403ec9c926150713f4acbaadab3bca47ced4520392966d5e.jpg", + "image_caption": [ + "Semantics Input" + ], + "image_footnote": [], + "bbox": [ + 241, + 88, + 401, + 284 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b18ba7a222c78a7ae580df67155847eef12a2edad21db45d5541879811b4cb0a.jpg", + "image_caption": [ + "ARFlow", + "Figure 4. Qualitative results on KITTI test set (sample #7, 20, 38, 112, 183) compared with ARFlow [37] and UPFlow [42]. We only show first frame input for conciseness. Our outputs have much clearer boundaries around objects. For more qualitative examples, see Appendix." + ], + "image_footnote": [], + "bbox": [ + 403, + 88, + 563, + 284 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/81bafe3a42b8dceeb4b7fce071a2fb656aab904f459b663294b6eca2000eff17.jpg", + "image_caption": [ + "UPFlow" + ], + "image_footnote": [], + "bbox": [ + 565, + 88, + 727, + 284 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/02089204af00bcd73238d308a144146cad31925f78d7c8c61735a06124c2374c.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 727, + 88, + 890, + 284 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/be5ecba44ee485cd5a6314673f47cc7873dfc49f9180708a710e28cff251dc48.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
roadcarterrainvegetationsidewalkbuildingwallpole
Proportion42.4%17.6%14.0%11.6%6.2%4.1%1.3%1.1%
ARFlow [37]4.5715.799.4215.344.616.0016.3410.75
Ours (final)3.6710.178.7413.473.094.2711.189.44
Relative improvement19.6%35.6%7.3%12.2%33.0%28.7%31.6%12.2%
Reweighed contribution19.0%49.4%4.8%10.9%4.7%3.5%3.4%0.7%
", + "bbox": [ + 122, + 356, + 848, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Improvement analysis", + "text_level": 1, + "bbox": [ + 76, + 515, + 284, + 531 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It may be interesting to understand where our improvements come from in terms of semantic classes, so we compute the KITTI-2015 train set error for each class. As shown in Tab. 5, our model improves the flow on every class. After reweighing the absolute improvement of each class based on their proportion in the evaluated pixels, we find that car flow accounts for nearly half of our improvement on Fl-all, indicating the effectiveness of augmenting vehicle objects.", + "bbox": [ + 75, + 539, + 468, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Generalization ability", + "text_level": 1, + "bbox": [ + 76, + 670, + 282, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also investigate the generalization ability of our semantic-aided flow models. We train flow models on Cityscapes and directly test them on KITTI-2015 train set without fine-tuning. Following [50], we crop the bottom $25\\%$ of the frame to remove the car logo and resize the frame to size $256 \\times 704$ . All other experiment settings are exactly the same as for KITTI. As we can see from Tab. 6, adding semantic inputs significantly helps our unsupervised flow model generalize and adapt better across datasets.", + "bbox": [ + 75, + 694, + 468, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and future work", + "text_level": 1, + "bbox": [ + 76, + 844, + 336, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we show that adding semantic segmentation inputs can help significantly improve the performance", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4685a30e710e8fd8d831d674fe3ac4d8359dba10f7cdf0baff8260804b356c97.jpg", + "table_caption": [ + "Table 5. KITTI-2015 flow error (Fl-all/%) for each semantic class. The first row shows the proportion of each class among evaluated pixels (KITTI flow does not evaluate on every pixel). We only show classes that account for at least $1\\%$ of the evaluated pixels." + ], + "table_footnote": [], + "table_body": "
MethodFl-allEPE-allEPE-nocEPE-occ
ARFlow (our impl.)13.214.082.889.40
Ours (baseline)12.273.812.439.91
Ours (+enc)†11.283.332.128.75
Ours (+enc +aug)†10.322.641.567.12
", + "bbox": [ + 503, + 512, + 888, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Generalization results (train on Cityscapes, and test on KITTI-2015 train). We use our implementation of ARFlow [37] to run this experiment. Our 'baseline' is an adapted ARFlow with added learned upsampler and no smoothness loss. $^+$ enc means adding semantic encoder; $^+$ aug means adding semantic augmentation. $\\dagger$ denotes models with semantic inputs.", + "bbox": [ + 496, + 603, + 892, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of unsupervised optical flow networks on autonomous driving datasets. We propose a novel network model called SemARFlow, a semantic adaptation of ARFlow [37], with a semantic encoder, a learned upsampler, and a semantic augmentation module, where some domain-specific motion prior has been taken into account. Our network better predicts flow at occlusion regions and effectively sharpens flow estimates around object boundaries, even for very challenging samples. The additional semantic inputs also make our network generalize better across datasets.", + "bbox": [ + 496, + 715, + 890, + 867 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "One direction for future work is motivated by the example illustrated in Fig. 5. Our model does output sharp flow", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9573", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1163bd8c232875b98653a17d74378498bfe4ef69b40c6cd4a3f95ac58e27a07a.jpg", + "image_caption": [ + "Figure 5. An example on KITTI-2015 train set (sample #164). See the discussion in Sec. 5 for details." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 207, + 148 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d3f5dfb78ff75b20c7534879e6a07823a7a3f0014c676ba3b095e4f7f671dee7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 90, + 336, + 148 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a7bdac7bbcf7508d90aa410a55955595fee7f51f1cf2bf6e69324c96fcd7f87f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 90, + 465, + 148 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "boundaries between the cars and the background, thanks to good boundary information from the semantic maps. However, the boundary between these two fast-moving cars is inferred from image information only, because their semantic maps merge into one. As a result, that part of the flow boundary is less accurate. An interesting direction for future work is to use instance-level semantics as input for more detailed object masks.", + "bbox": [ + 76, + 218, + 470, + 338 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgments. This research is based upon work supported by the National Science Foundation under Grant No. 1909821.", + "bbox": [ + 76, + 359, + 468, + 405 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 419, + 173, + 435 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Min Bai, Wenjie Luo, Kaustav Kundu, and Raquel Urtasun. Exploiting semantic information and deep matching for optical flow. In Proceedings of the European Conference on Computer Vision, pages 154-170. Springer, 2016. 2, 3, 7", + "[2] Shubhankar Borse, Ying Wang, Yizhe Zhang, and Fatih Porikli. Inverseform: A loss function for structured boundary-aware segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5901-5911, 2021. 3", + "[3] Gary Bradski. The opencv library. Dr. Dobb's Journal: Software Tools for the Professional Programmer, 25(11):120-123, 2000. 5", + "[4] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Proceedings of the European Conference on Computer Vision, pages 611-625. Springer, 2012. 1", + "[5] Yingfeng Cai, Lei Dai, Hai Wang, and Zhixiong Li. Multi-target pan-class intrinsic relevance driven model for improving semantic segmentation in autonomous driving. IEEE Transactions on Image Processing, 30:9069-9084, 2021. 3", + "[6] Linda Capito, Umit Ozguner, and Keith Redmill. Optical flow based visual potential field for autonomous driving. In IEEE Intelligent Vehicles Symposium, pages 885-891. IEEE, 2020. 1", + "[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European Conference on Computer Vision, pages 801-818, 2018. 6", + "[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes" + ], + "bbox": [ + 86, + 445, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "dataset for semantic urban scene understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3213-3223, 2016. 3, 6", + "[9] Mingyu Ding, Zhe Wang, Bolei Zhou, Jianping Shi, Zhiwu Lu, and Ping Luo. Every frame counts: Joint learning of video segmentation and optical flow. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10713-10720, 2020. 2, 3", + "[10] Alexey Dosovitskiy, Philipp Fischery, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick van der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE International Conference on Computer Vision, ICCV '15, pages 2758-2766, Washington, DC, USA, 2015. IEEE Computer Society. 1, 2", + "[11] Cheng Feng, Long Ma, Congxuan Zhang, Zhen Chen, Liyue Ge, and Shaofeng Jiang. Ss-sf: Piecewise 3d scene flow estimation with semantic segmentation. IEEE Access, 9:22745-22759, 2021. 3", + "[12] Jun Fu, Jing Liu, Haijie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3146-3154, 2019. 3", + "[13] Aditya Ganeshan, Alexis Vallet, Yasunori Kudo, Shin-ichi Maeda, Tommi Kerola, Rares Ambrus, Dennis Park, and Adrien Gaidon. Warp-refine propagation: Semi-supervised auto-labeling via cycle-consistency. In Proceedings of the IEEE International Conference on Computer Vision, pages 15499–15509, 2021. 3", + "[14] Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In Proceedings of the European Conference on Computer Vision, pages 713-729. Springer, 2020. 1", + "[15] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research, 32(11):1231-1237, 2013. 6", + "[16] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 2", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1", + "[18] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial Intelligence, 17(1-3):185–203, 1981. 2", + "[19] Zilong Huang, Xinggang Wang, Lichao Huang, Chang Huang, Yunchao Wei, and Wenyu Liu. Ccnet: Criss-cross attention for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 603-612, 2019. 3", + "[20] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. Liteflownet: A lightweight convolutional neural network for optical flow estimation. In Proceedings of the IEEE Conference" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9574", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 8981-8989, 2018. 2", + "[21] Junhwa Hur and Stefan Roth. Joint optical flow and temporally consistent semantic segmentation. In Proceedings of the European Conference on Computer Vision, pages 163-177. Springer, 2016. 2, 3, 7", + "[22] Junhwa Hur and Stefan Roth. Iterative residual refinement for joint optical flow and occlusion estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5754-5763, 2019. 1, 2, 3, 6", + "[23] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1647-1655, 2017. 2", + "[24] Woobin Im, Tae-Kyun Kim, and Sung-Eui Yoon. Unsupervised learning of optical flow with deep feature similarity. In Proceedings of the European Conference on Computer Vision, pages 172-188. Springer, 2020. 2, 6", + "[25] Joel Janai, Fatma Guney, Anurag Ranjan, Michael Black, and Andreas Geiger. Unsupervised learning of multi-frame optical flow with occlusions. In Proceedings of the European Conference on Computer Vision, pages 690-706, 2018. 2", + "[26] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In Proceedings of the IEEE International Conference on Computer Vision, pages 9772-9781, 2021. 2", + "[27] Rico Jonschkowski, Austin Stone, Jonathan T Barron, Ariel Gordon, Kurt Konolige, and Anelia Angelova. What matters in unsupervised optical flow. In Proceedings of the European Conference on Computer Vision, pages 557-572. Springer, 2020. 1, 6", + "[28] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM computing surveys, 54(10s):1-41, 2022. 1", + "[29] Hannah H Kim, Celia Cintas, Girmaw Abebe Tadesse, and Skyler Speakman. Spatially constrained adversarial attack detection and localization in the representation space of optical flow networks. In International Joint Conference on Artificial Intelligence, 2023. 1", + "[30] Hannah Halin Kim, Shuzhi Yu, and Carlo Tomasi. Joint detection of motion boundaries and occlusions. In Proceedings of the British Machine Vision Conference, 2021. 1", + "[31] Hannah Halin Kim, Shuzhi Yu, Shuai Yuan, and Carlo Tomasi. Cross-attention transformer for video interpolation. In Proceedings of the Asian Conference on Computer Vision Workshops, pages 320–337, 2022. 1", + "[32] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. Proceedings of the International Conference on Learning Representations, 2014. 6", + "[33] Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weissenborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Sylvain Gelly, Thomas Unterthiner, and Xiaohua Zhai. An image is worth 16x16 words: Transformers for image recogni" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion at scale. In Proceedings of the International Conference on Learning Representations, 2021. 1", + "[34] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 1", + "[35] Xiangtai Li, Jiangang Bai, Kuiyuan Yang, and Yunhai Tong. Flow2seg: Motion-aided semantic segmentation. In Proceedings of the International Conference on Artificial Neural Networks, pages 225–237. Springer, 2019. 3", + "[36] Zewen Li, Fan Liu, Wenjie Yang, Shouheng Peng, and Jun Zhou. A survey of convolutional neural networks: analysis, applications, and prospects. IEEE transactions on neural networks and learning systems, 2021. 1", + "[37] Liang Liu, Jiangning Zhang, Ruifei He, Yong Liu, Yabiao Wang, Ying Tai, Donghao Luo, Chengjie Wang, Jilin Li, and Feiyue Huang. Learning by analogy: Reliable supervision from transformations for unsupervised optical flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6489-6498, 2020. 1, 2, 3, 4, 5, 6, 7, 8", + "[38] Pengpeng Liu, Irwin King, Michael R Lyu, and Jia Xu. Ddflow: Learning optical flow with unlabeled data distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8770-8777, 2019. 1, 2", + "[39] Pengpeng Liu, Michael Lyu, Irwin King, and Jia Xu. Selfflow: Self-supervised learning of optical flow. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4571-4580, 2019. 1, 2, 4, 6", + "[40] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3431-3440, 2015. 2, 3", + "[41] Bruce D Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In International Joint Conference on Artificial Intelligence, volume 2, pages 674-679, 1981. 2", + "[42] Kunming Luo, Chuan Wang, Shuaicheng Liu, Haoqiang Fan, Jue Wang, and Jian Sun. Upflow: Upsampling pyramid for unsupervised optical flow learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1045-1054, 2021. 1, 2, 3, 6, 7, 8", + "[43] Rémi Marsal, Florian Chabot, Angélique Loesch, and Hichem Sahbi. Brightflow: Brightness-change-aware unsupervised learning of optical flow. In Proceedings of the IEEE winter conference on applications of computer vision, pages 2061-2070, 2023. 1", + "[44] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4040-4048, 2016. 1, 2", + "[45] Simon Meister, Junhwa Hur, and Stefan Roth. Unflow: Unsupervised learning of optical flow with a bidirectional census loss. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 1, 2, 5", + "[46] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE Conference" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9575", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, 2015. 2, 3, 6, 7", + "[47] Hyeonwoo Noh, Seunghoon Hong, and Bohyung Han. Learning deconvolution network for semantic segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1520-1528, 2015. 2, 3", + "[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, pages 8024-8035. Curran Associates, Inc., 2019. 6", + "[49] Anurag Ranjan and Michael J. Black. Optical flow estimation using a spatial pyramid network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2720-2729, 2017. 2", + "[50] Anurag Ranjan, Varun Jampani, Lukas Balles, Kihwan Kim, Deqing Sun, Jonas Wulff, and Michael J Black. Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12240–12249, 2019. 8", + "[51] Hazem Rashed, Senthil Yogamani, Ahmad El-Sallab, Pavel Krizek, and Mohamed El-Helw. Optical flow augmented semantic segmentation networks for automated driving. In International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, pages 165-172, 2019. 3", + "[52] Fitsum A Reda, Guilin Liu, Kevin J Shih, Robert Kirby, Jon Barker, David Tarjan, Andrew Tao, and Bryan Catanzaro. Sdc-net: Video prediction using spatially-displaced convolution. In Proceedings of the European Conference on Computer Vision, pages 718-733, 2018. 3", + "[53] Zhile Ren, Orazio Gallo, Deqing Sun, Ming-Hsuan Yang, Erik B Sudderth, and Jan Kautz. A fusion approach for multiframe optical flow estimation. In Proceedings of the IEEE winter conference on applications of computer vision, pages 2077-2086. IEEE, 2019. 2", + "[54] Zhile Ren, Deqing Sun, Jan Kautz, and Erik Sudderth. Cascaded scene flow prediction using semantic segmentation. In Proceedings of the International Conference on 3D Vision, pages 225-233. IEEE, 2017. 3", + "[55] Zhe Ren, Junchi Yan, Bingbing Ni, Bin Liu, Xiaokang Yang, and Hongyuan Zha. Unsupervised deep learning for optical flow estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2017. 2", + "[56] Stephan R. Richter, Zeeshan Hayder, and Vladlen Koltun. Playing for benchmarks. In Proceedings of the IEEE International Conference on Computer Vision, pages 2232-2241, 2017. 1", + "[57] Laura Sevilla-Lara, Deqing Sun, Varun Jampani, and Michael J Black. Optical flow with semantic segmentation and localized layers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3889-3898, 2016. 2, 3, 7" + ], + "bbox": [ + 78, + 90, + 467, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Shihao Shen, Louis Kerofsky, and Senthil Yogamani. Optical flow for autonomous driving: Applications, challenges and improvements. arXiv preprint arXiv:2301.04422, 2023. 1", + "[59] Jianbo Shi and Carlo Tomasi. Good features to track. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 593-600, 1994. 1", + "[60] Yukang Shi and Kaiheng Ma. Safit: Segmentation-aware scene flow with improved transformer. In Proceedings of the International Conference on Robotics and Automation, pages 10648-10655. IEEE, 2022. 3", + "[61] Jeongho Shin, Sangjin Kim, Sangkyu Kang, Seong-Won Lee, Joonki Paik, Besma Abidi, and Mongi Abidi. Optical flow-based real-time object tracking using non-prior training active feature model. Real-time imaging, 11(3):204-218, 2005. 1", + "[62] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, volume 11006, pages 369-386. SPIE, 2019. 6", + "[63] Austin Stone, Daniel Maurer, Alper Ayvaci, Anelia Angelova, and Rico Jonschkowski. Smurf: Self-teaching multiframe unsupervised raft with full-image warping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3887-3896, 2021. 3", + "[64] Xiuchao Sui, Shaohua Li, Xue Geng, Yan Wu, Xinxing Xu, Yong Liu, Rick Goh, and Hongyuan Zhu. Craft: Cross-attentional flow transformer for robust optical flow. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 17602-17611, 2022. 2", + "[65] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8934-8943, 2018. 1, 2, 3", + "[66] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Models matter, so does training: An empirical study of cnns for optical flow estimation. IEEE transactions on pattern analysis and machine intelligence, 42(6):1408-1423, 2019. 6", + "[67] Andrew Tao, Karan Sapra, and Bryan Catanzaro. Hierarchical multi-scale attention for semantic segmentation. arXiv preprint arXiv:2005.10821, 2020. 3", + "[68] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proceedings of the European Conference on Computer Vision, pages 402-419. Springer, 2020. 1, 2, 3, 6", + "[69] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Proceedings of the International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1", + "[70] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE Con" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9576", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on Computer Vision and Pattern Recognition Workshops, pages 969-977, 2018. 1", + "[71] Panqu Wang, Pengfei Chen, Ye Yuan, Ding Liu, Zehua Huang, Xiaodi Hou, and Garrison Cottrell. Understanding convolution for semantic segmentation. In Proceedings of the IEEE winter conference on applications of computer vision, pages 1451-1460. IEEE, 2018. 3", + "[72] Yang Wang, Yi Yang, Zhenheng Yang, Liang Zhao, Peng Wang, and Wei Xu. Occlusion aware unsupervised learning of optical flow. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4884-4893, 2018. 1, 2", + "[73] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5", + "[74] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Song Wang, and Lili Ju. Semantic stereo matching with pyramid cost volumes. In Proceedings of the IEEE International Conference on Computer Vision, pages 7484-7493, 2019. 3", + "[75] Jonas Wulff, Laura Sevilla-Lara, and Michael J Black. Optical flow in mostly rigid scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4671-4680, 2017. 2, 3, 7", + "[76] Haofei Xu, Jiaolong Yang, Jianfei Cai, Juyong Zhang, and Xin Tong. High-resolution optical flow from 1d attention and correlation. In Proceedings of the IEEE International Conference on Computer Vision, pages 10498-10507, 2021. 2", + "[77] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8121-8130, 2022. 2", + "[78] Guorun Yang, Hengshuang Zhao, Jianping Shi, Zhidong Deng, and Jiaya Jia. Segstereo: Exploiting semantic information for disparity estimation. In Proceedings of the European Conference on Computer Vision, pages 636-651, 2018. 3", + "[79] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. Bisenet v2: Bilateral network with guided aggregation for real-time semantic segmentation. International Journal of Computer Vision, 129:3051-3068, 2021. 3", + "[80] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceedings of the European Conference on Computer Vision, pages 325-341, 2018. 3", + "[81] Jason J. Yu, Adam W. Harley, and Konstantinos G. Derpanis. Back to basics: Unsupervised learning of optical flow via brightness constancy and motion smoothness. In Gang Hua and Hervé Jégou, editors, Proceedings of the European Conference on Computer Vision Workshops, pages 3-10, Cham, 2016. Springer International Publishing. 1, 2", + "[82] Shuzhi Yu, Hannah H Kim, Shuai Yuan, and Carlo Tomasi. Unsupervised flow refinement near motion boundaries. In" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the British Machine Vision Conference. BMVA Press, 2022. 1", + "[83] Shuai Yuan, Xian Sun, Hannah Kim, Shuzhi Yu, and Carlo Tomasi. Optical flow training under limited label budget via active learning. In Proceedings of the European Conference on Computer Vision, pages 410-427, 2022. 2", + "[84] Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Common practices and emerging technologies. IEEE access, 8:58443-58469, 2020. 1", + "[85] Feihu Zhang, Oliver J Woodford, Victor Adrian Prisacariu, and Philip HS Torr. Separable flow: Learning motion cost volumes for optical flow estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 10807-10817, 2021. 1, 2, 6", + "[86] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2881-2890, 2017. 3", + "[87] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6881-6890, 2021. 2, 3", + "[88] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In Proceedings of the International Conference on Learning Representations, 2021. 1", + "[89] Yi Zhu, Karan Sapra, Fitsum A Reda, Kevin J Shih, Shawn Newsam, Andrew Tao, and Bryan Catanzaro. Improving semantic segmentation via video propagation and label relaxation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8856-8865, 2019. 2, 3, 6" + ], + "bbox": [ + 503, + 92, + 890, + 612 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "9577", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/fa7524e1-a76d-4844-9361-43c8b8046bf9_model.json b/2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/fa7524e1-a76d-4844-9361-43c8b8046bf9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..566a595edd1c1818c3dc042f83979d02ce53a615 --- /dev/null +++ b/2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/fa7524e1-a76d-4844-9361-43c8b8046bf9_model.json @@ -0,0 +1,2930 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.13, + 0.877, + 0.177 + ], + "angle": 0, + "content": "SemARFlow: Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.204, + 0.709, + 0.24 + ], + "angle": 0, + "content": "Shuai Yuan, Shuzhi Yu, Hannah Kim, and Carlo Tomasi \nDuke University" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.242, + 0.684, + 0.256 + ], + "angle": 0, + "content": "{shuai, shuzhiyu, hannah, tomasi}@cs.duke.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.474, + 0.613 + ], + "angle": 0, + "content": "Unsupervised optical flow estimation is especially hard near occlusions and motion boundaries and in low-texture regions. We show that additional information such as semantics and domain knowledge can help better constrain this problem. We introduce SemARFlow, an unsupervised optical flow network designed for autonomous driving data that takes estimated semantic segmentation masks as additional inputs. This additional information is injected into the encoder and into a learned upsampler that refines the flow output. In addition, a simple yet effective semantic augmentation module provides self-supervision when learning flow and its boundaries for vehicles, poles, and sky. Together, these injections of semantic information improve the KITTI-2015 optical flow test error rate from \\(11.80\\%\\) to \\(8.38\\%\\). We also show visible improvements around object boundaries as well as a greater ability to generalize across datasets. Code is available at https://github.com/duke-vision/semantic-unsup-flow-release." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.645, + 0.21, + 0.66 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Optical flow estimation, i.e., pixel-level motion tracking across video frames, has broad applications in many computer vision tasks that include object tracking [61], video editing [14, 31], and autonomous driving [6, 58]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.47, + 0.884 + ], + "angle": 0, + "content": "Thanks to the success of deep convolutional neural networks [34, 17, 36] and transformer networks [33, 69, 88, 28] in computer vision, many top-performing supervised optical flow networks have been proposed in recent years [10, 65, 22, 68, 85], in which ground-truth labels supervise training. However, real optical flow is hard to label, so most supervised methods train (or at least pre-train) on synthetic datasets [10, 44, 4, 56] which makes them hard to adapt to real applications due to the significant gap between synthetic and real data [70, 29]." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Due to label scarcity, unsupervised training of optical" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.291, + 0.673, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.291, + 0.728, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.292, + 0.781, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.292, + 0.835, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.836, + 0.292, + 0.89, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.42, + 0.895, + 0.49 + ], + "angle": 0, + "content": "Figure 1. An example on KITTI-2015 test set (sample #63). Only one input frame is shown for conciseness. Our SemARFlow takes additional semantic segmentation inputs (estimated by an off-the-shelf model) and outputs much sharper flow around semantic boundaries." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.894, + 0.622 + ], + "angle": 0, + "content": "flow estimators [81] instead uses loss terms based on the assumptions of constant brightness and smooth flow [45, 27, 42]. Some self-supervision techniques have also been studied to enhance model performance [38, 39, 37]. Unsupervised training makes it possible to train flow networks directly on large real datasets from the target domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.895, + 0.761 + ], + "angle": 0, + "content": "Even so, unsupervised optical flow estimation is a poorly constrained problem. Brightness is not constant in regions with occlusions [72] or on shiny surfaces [43], nor is it smooth across motion boundaries [30, 82]. Moreover, points in regions with poor texture content [59] or in dark shadows [58] are difficult to track as they are easily confused with neighboring points (the so-called aperture problem). This dearth of reliable constraints makes the unsupervised training of flow networks especially challenging." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.762, + 0.895, + 0.883 + ], + "angle": 0, + "content": "A natural way to address this issue is to inject additional constraints in the form of semantics and domain knowledge. For example, in autonomous driving applications [84, 6], we have clear expectations about object types and layouts of the scene, as well as prior knowledge of how each type of object typically moves. We focus on this application domain and show that this additional information helps the network achieve better flow results." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "To make an autonomous driving system work well in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9566" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.183 + ], + "angle": 0, + "content": "reality, it is best to train it on real data. However, annotating real driving video with optical flow labels is expensive [83], as it requires careful synchronization and calibration of diverse sensors including cameras, LiDAR, and GPS/IMU [16], aided by some manual annotation and curation based on CAD models of moving objects [46]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.185, + 0.471, + 0.336 + ], + "angle": 0, + "content": "In contrast, annotating semantic labels seems much more feasible, and indeed semantic labels are available in most (if not all) existing driving datasets. We consider semantic segmentation because it provides semantics at the pixel level, the same level as optical flow. As one of the most popular and well-studied tasks in modern computer vision, semantic segmentation [40, 47, 87] has been extensively adopted for autonomous driving systems. In this paper, we show that adding semantic segmentation inputs helps improve unsupervised optical flow performance significantly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.471, + 0.565 + ], + "angle": 0, + "content": "Specifically, we first infer semantic segmentation maps using an off-the-shelf model [89], which of course is trained with semantic labels. An encoder with semantic map input is used to aggregate image and semantic features (Sec. 3.1), and a learned upsampler is added into the iterative decoder to refine flow around object boundaries given semantic inputs (Sec. 3.2). We also propose a simple yet effective semantic augmentation module for self-supervision, which provides realistic augmentations specific to the vehicles, poles, and sky classes based on domain knowledge and segmentation maps (Sec. 3.3). An occluder cache is implemented to improve efficiency (Sec. 3.4). Semantic augmentation provides challenging samples for self-supervision, which help train flow better in occluded regions and around foreground object boundaries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.764 + ], + "angle": 0, + "content": "Overall, by injecting semantic segmentation inputs, our SemARFlow network achieves significantly better unsupervised flow results both quantitatively (Sec. 4.3) and qualitatively (Sec. 4.4). Adapted from ARFlow [37], SemARFlow reduces KITTI-2015 [46] test error from \\(11.80\\%\\) to \\(8.38\\%\\) and out-performs the current unsupervised state-of-the-art UPFlow [42] \\((9.38\\%)\\) by a clear margin. The example in Fig. 1 also demonstrates visible improvements around object boundaries. The effectiveness of each module is justified through an extensive ablation study (Sec. 4.5) and improvement analysis (Sec. 4.6). In addition to performance boost, unsupervised flow networks with additional semantic inputs generalize better across different datasets (Sec. 4.7)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Our research is essentially novel compared to previous approaches. Some early work incorporates semantics in traditional energy-minimization methods [41, 18] for optical flow through geometric constraints such as piece-wise rigid motion and planar surface motion [57, 21, 1, 75]. In comparison, to the best of our knowledge, we are the first to inject semantics into the unsupervised training of recent optical flow networks. Some research also trains a network for segmentation and optical flow jointly [9]. In contrast, we" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "leverage existing, separately trained segmentation systems both because they are available and because modularity—separating segmentation from flow estimation—is important for the development of large real-application systems such as autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.168, + 0.819, + 0.182 + ], + "angle": 0, + "content": "In summary, our contributions are as follows." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.191, + 0.892, + 0.237 + ], + "angle": 0, + "content": "- To the best of our knowledge, we are the first to explore adding semantic inputs to assist the unsupervised training of deep optical flow networks." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.245, + 0.892, + 0.322 + ], + "angle": 0, + "content": "- We propose a simple yet effective network called SemARFlow that achieves state-of-the-art results both quantitatively and qualitatively. Our model works well on real-life occlusions and yields sharp motion boundaries around objects." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.331, + 0.892, + 0.362 + ], + "angle": 0, + "content": "- We provide full training and inference code as well as trained models to encourage follow-up research." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.191, + 0.892, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.375, + 0.638, + 0.39 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.892, + 0.581 + ], + "angle": 0, + "content": "Supervised optical flow Since the introduction of the pioneering deep optical flow network FlowNet [10], more and more top-performing CNN-based flow estimators have been proposed over the years [23, 49, 65, 20, 22, 68]. Recently, vision transformer networks and attention mechanism have also been applied to this problem, and these have achieved state-of-the-art performance on benchmark datasets [76, 26, 85, 64, 77]. The supervised methods are often pre-trained on large synthetic datasets such as FlyingChairs [10] and FlyingThings3D [44] before fine-tuning on the target dataset. However, there is a clear gap between the artificially generated data and real scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.764 + ], + "angle": 0, + "content": "Unsupervised optical flow Due to the lack of ground-truth labels, unsupervised optical flow estimation uses surrogate losses such as photometric loss and smoothness loss to supervise training [81, 55]. To tackle the issues around occlusion regions, various methods have been proposed including estimated occlusion masks [72], forward-backward consistency [45], and multi-frame fusion [25, 53]. To better upsample the flow in the decoder, UPFlow [42] additionally predicts a confidence map and an interpolation flow to guide flow refinement and has become the state-of-the-art unsupervised flow method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Some latest research has also explored the use of self-supervision to enhance flow prediction. Early methods apply the knowledge distillation technique to train a two-stage teacher-student network [38, 39]. ARFlow [37] further improves this idea by generating reliable self-supervision signals from data transformations, while merging the two training stages into single-stage training with one added loss term. SimFlow [24] replaces the handcrafted features with deep self-supervised features to measure similarity in the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9567" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.471, + 0.182 + ], + "angle": 0, + "content": "unsupervised losses. SMURF [63] utilizes a RAFT-like structure [68] and applies multi-frame self-supervised training with many technical improvements. Our SemARFlow also uses self-supervision but with the guidance of semantic segmentation, which is much more realistic than guidance without semantics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.202, + 0.47, + 0.457 + ], + "angle": 0, + "content": "Semantic segmentation Semantic segmentation classifies each pixel of the given image into semantic objects. Fully Convolutional Network (FCN) [40] is one of the early CNN-based segmentation methods. It takes inputs of arbitrary sizes and outputs dense pixel-level predictions, becoming one of the popular backbone architectures for follow-up work. Deconvolution network [47] is also proposed to better recover low-level details of the prediction. One main challenge for these systems is that they lack global scene information. Subsequent work addresses this issue by enlarging the receptive field of the network with global pyramid pooling layers as in PSPNet [86], hybrid dilated convolutions [71], and a fast-down-sampling strategy [80, 79]. Attention modules have also shown to help in semantic segmentation by capturing full-image dependencies of all pixels [19, 87] or the semantic inter-dependencies across spatial and channel dimensions [12]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.47, + 0.595 + ], + "angle": 0, + "content": "There also exists extensive work on semantic segmentation in the context of autonomous driving [2, 67, 13, 5, 89], thanks to the publication of large-scale driving datasets [8, 46]. To better train the network on a coarsely labeled dataset like KITTI [46], Zhu et al. use a video prediction network SDCNet [52] to synthesize new training samples with relaxed label propagation [89]. Due to their good performance on KITTI [46], we utilize their network models to infer semantic inputs for all our experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Combining semantics and optical flow Though there has been much progress on both semantic segmentation and optical flow estimation, the semantic optical flow problem (how to exploit semantics to help optical flow estimation) has received limited attention in recent years, and the current best results are thus much outdated. Some early methods incorporate semantics through geometric constraints to refine flow on various semantic regions, such as planar regions (using homographies) [57], static regions (using rigid camera motion and epipolar constraints) [21, 75], and rigid objects (estimating rigid motion for each object instance) [1]. However, most methods are traditional flow methods based on energy minimization, where an initial flow estimate is usually needed and semantics is mostly used for refinement. In comparison, we explore adapting latest unsupervised optical flow networks to leverage semantic inputs in one single stage of estimation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Apart from using semantics to help flow, some research has also explored using optical flow to help seman" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "tic segmentation [51, 35], or to train both tasks jointly [9]. There are also studies on exploiting semantics on some other correspondence matching tasks such as stereo matching [74, 78] and 3D scene flow estimation, where some additional depth cues such as stereo camera inputs [54, 11] or point clouds [60] are needed." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.197, + 0.591, + 0.212 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.223, + 0.892, + 0.344 + ], + "angle": 0, + "content": "Our network is adapted from the two-frame version of ARFlow [37], which uses a light-weight PwCNet [65] as its backbone. The inputs are two consecutive frames \\( I_1, I_2 \\in \\mathbb{R}^{H \\times W \\times 3} \\) as well as their semantic segmentation maps \\( S_1, S_2 \\in \\{0, 1, \\dots, c\\}^{H \\times W} \\), where the number of classes \\( c \\) is 19 as we use the Cityscapes format [8]. A detailed diagram of the structure of our network can be found in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.355, + 0.673, + 0.37 + ], + "angle": 0, + "content": "3.1. Semantic encoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.892, + 0.517 + ], + "angle": 0, + "content": "We first inject semantic information into the encoder. As shown in Fig. 2(a), shared by each frame \\( i \\in \\{1,2\\} \\), separate convolutional layers extract features from image \\( I_{i} \\) and semantics \\( S_{i} \\) (one-hot encoded). Features from the two pipelines are concatenated and fed to additional layers. Features at different resolutions \\( (H / 2^{l}, W / 2^{l}) \\) form a pyramid \\( \\{f_{i}^{(l)} \\mid 2 \\leq l \\leq 6\\} \\). The semantic information in these features helps delineate objects in dark shadows, where appearance is more homogeneous." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.527, + 0.868, + 0.543 + ], + "angle": 0, + "content": "3.2. Iterative decoder with a learned upsampler" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.891, + 0.62 + ], + "angle": 0, + "content": "Following [22, 37], an iterative residual refinement decoder starts from zero estimate \\( F_{1\\rightarrow 2}^{(7)} = 0 \\). For iteration \\( l \\in \\{6,5,4,3,2\\} \\), the decoder refines feature map \\( F_{1\\rightarrow 2}^{(l + 1)} \\) into \\( F_{1\\rightarrow 2}^{(l)} \\) based on features \\( f_{1}^{(l)}, f_{2}^{(l)} \\) (Fig. 2(b))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.892, + 0.777 + ], + "angle": 0, + "content": "More specifically, \\( F_{1\\rightarrow 2}^{(l + 1)} \\) is upsampled to match the resolution and used to warp \\( f_2^{(l)} \\) to yield warped feature \\( \\hat{f}_1^{(l)} \\). Correlation volumes are computed between \\( f_1^{(l)} \\) and \\( \\hat{f}_1^{(l)} \\). A one-by-one convolutional layer \\( C^{(l)} \\) compresses the number of channels to a fixed number so that the same layer can be reused across all iterations as proposed by [22]. A flow estimator network predicts a flow residual to be added to the current estimate, and a context network then aggregates flow information spatially and refines the current flow again." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.901 + ], + "angle": 0, + "content": "A learned upsampler network upsamples the final output \\( F_{1\\rightarrow 2}^{(2)} \\) in our system. This is different from ARFlow [37], which simply uses four-fold bilinear interpolation, making the final flow boundaries blurry. In contrast, our model learns to sharpen flow boundaries based on the semantic inputs, which have clear boundaries around moving objects. To this end, we add a convex upsampler network similar to the one in RAFT [68]. Different from UPFlow [42]," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9568" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.098, + 0.344, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.228, + 0.335, + 0.242 + ], + "angle": 0, + "content": "(a) shared pyramidal encoder \\((i\\in \\{1,2\\})\\)" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.099, + 0.871, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.424, + 0.228, + 0.83, + 0.242 + ], + "angle": 0, + "content": "(b) one iteration (at the \\(l\\)th-level) of the iterative decoder \\((l\\in \\{6,5,4,3,2\\})\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.257, + 0.837, + 0.273 + ], + "angle": 0, + "content": "Figure 2. Network structure. See text in Sec. 3.1 and Sec. 3.2 for explanations. More detailed diagrams are in appendix." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.285, + 0.889, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.553, + 0.774, + 0.568 + ], + "angle": 0, + "content": "Figure 3. Illustration of semantic augmentation as self-supervision. See text in Sec. 3.3 for details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.47, + 0.707 + ], + "angle": 0, + "content": "our learned upsampler is only used to upsample from internal estimate \\( F_{1\\rightarrow 2}^{(l)} \\) to output flow \\( U_{1\\rightarrow 2}^{(l)} \\), but not when we upsample \\( F_{1\\rightarrow 2}^{(l + 1)} \\) at the first step in the decoder iteration. The upsampled \\( U_{1\\rightarrow 2}^{(l)} \\) has resolution \\( (H / 2^{l - 2},W / 2^{l - 2}) \\), so \\( U_{1\\rightarrow 2} = U_{1\\rightarrow 2}^{(2)} \\) (with the original resolution) is the final flow prediction of the network." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.722, + 0.44, + 0.74 + ], + "angle": 0, + "content": "3.3. Semantic augmentation as self-supervision" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.47, + 0.854 + ], + "angle": 0, + "content": "ARFlow has a very effective in-network augmentation module that samples random transformations \\(\\mathcal{T}_{\\theta_1},\\mathcal{T}_{\\theta_2}\\) of the flow prediction \\(U_{1\\rightarrow 2}\\) in the first pass of the network and then uses the transformed images \\(\\hat{I}_1 = \\mathcal{T}_{\\theta_1}(I_1),\\hat{I}_2 =\\) \\(\\mathcal{T}_{\\theta_2}(I_2)\\) in a second pass. The prediction \\(U_{1\\rightarrow 2}\\) is also transformed accordingly and used to self-supervise the output of the second pass. See [37] for details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We retain the augmentation module but make a third pass of the network using semantics-transformed inputs for self-supervision in addition to the ARFlow augmentations of ap" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.892, + 0.666 + ], + "angle": 0, + "content": "pearance (e.g., color jitter, random noise) and spatial transformations (e.g., random rotation, random rescaling). The idea behind semantic augmentation is to blend in real object motions across samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.893, + 0.79 + ], + "angle": 0, + "content": "To this end (see Fig. 3), we carve objects from other samples based on their semantic maps and paste them as moving foreground objects into the current sample \\( I_{1}, I_{2} \\), thereby producing augmented images \\( \\tilde{I}_{1}, \\tilde{I}_{2} \\). The segmentation maps \\( S_{1}, S_{2} \\) and the first pass output flow \\( U_{1 \\to 2} \\) are transformed accordingly, and the transformed \\( \\tilde{U}_{1 \\to 2} \\) self-supervises the third-pass network outputs \\( \\tilde{U}_{1 \\to 2}' = \\mathcal{F}(\\tilde{I}_{1}, \\tilde{I}_{2}, \\tilde{S}_{1}, \\tilde{S}_{2}) \\), where \\( \\mathcal{F} \\) is the flow network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.792, + 0.892, + 0.883 + ], + "angle": 0, + "content": "Similar to occlusion hallucination in SelFlow [39], our semantic augmentation also creates new occlusions and uses the reliable non-occluded flow from the first pass to self-supervise the flow on those newly occluded pixels. However, rather than superpixels filled with noise, we create occlusions with real objects, which are much more realistic." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Moreover, since we determine the motion of each oc" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9569" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.214 + ], + "angle": 0, + "content": "cluder, we can use its known optical flow for motion self-supervision. As illustrated in Fig. 3, the occluder motion can be very different from that of the background. This trains the model to assume less smoothness around highly-dynamic objects and to output sharper motion estimates near motion boundaries. These hallucinated but realistic occluders also provide more examples of actively moving objects, which are relatively scarce in the original datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.245, + 0.473, + 0.398 + ], + "angle": 0, + "content": "**Vehicles** We augment vehicle classes (car, truck, bus, train) as they are major sources of errors in previous models (see Tab. 5 for statistics). We group these four classes because they are easily confused with each other in the semantic input. To carve out a single car instance with high probability, we find connected components of vehicle segmentation regions using OpenCV [3] and accept masks whose bounding boxes are 50-300 pixels wide and 50-150 pixels high. We also drop masks that occupy less than \\(60\\%\\) area of their bounding boxes to remove largely occluded vehicles." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.428, + 0.473, + 0.581 + ], + "angle": 0, + "content": "Poles We augment pole classes (pole, traffic light, traffic sign) not only because they are common foreground objects that cause occlusions, but also because their motions are usually poorly estimated due to their thin shapes. We use a 200-pixel wide, image-height sliding window to scan the segmentation of these pole classes and find the window with the most pole pixels. If more than \\(10\\%\\) pole pixels occur in it, that object mask is stored. This produces groups of nearby poles rather than just individual poles, which may not provide enough augmentation by themselves." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.748 + ], + "angle": 0, + "content": "Sky We also augment the sky through prior knowledge that flow in the sky should be very small. We shrink the sky flow estimate by half before using it for self-supervision. As shown in Fig. 3, the middle tree part in \\( U_{1\\rightarrow 2} \\) is very blurry, with a large part of sky mixed in there. After shrinking sky flow by half, the tree in \\( \\tilde{U}_{1\\rightarrow 2} \\) now has sharper flow around the boundary between tree and sky. Using this as self-supervision helps the model sharpen the motion boundaries between sky and other regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Another way to apply this prior knowledge is to postprocess, i.e., shrink the final output flow in the sky region, instead of using the shrunk flow to self-supervise. However, the latter approach is preferable because it allows the network to balance prior knowledge (self-supervision) with the current sample observation (photometric loss). Our experiments do show cases where some small objects (such as some overhead electric power lines) are misclassified as part of the sky in the semantic input. In these cases, the network learns to rely less on prior knowledge." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.658, + 0.107 + ], + "angle": 0, + "content": "3.4.Occluder cache" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.894, + 0.189 + ], + "angle": 0, + "content": "During training, we maintain a cache of occluding objects for semantic augmentation. We sample objects from the cache to augment the current batch and then push any new objects found in the batch into the cache, with random replacement if the cache is full." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.19, + 0.895, + 0.31 + ], + "angle": 0, + "content": "Specifically, after we load a sample for training, we first use the current model to infer its optical flow and then search within the sample for any cars (or poles) that meet the standards defined in Sec. 3.3 as \"cutouts\". For each cutout, we first find the flow vectors on all pixels that belong to that cutout and then compute the mean flow for later reference. We store the current sample, the cutout mask, and the mean flow together in the cache." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.895, + 0.463 + ], + "angle": 0, + "content": "Later on, when we randomly select a cutout to augment another sample, we also retrieve the stored mean flow of this cutout, which is then augmented by random rescaling (by 0.8-1.5 times) and reversing (with \\(50\\%\\) probability). We use the augmented mean flow to translate the entire cutout object as an occluder to generate a new sample. For realism, occluders are pasted to the same location they occupied in their image of origin. Holding occluders in a cache provides a random mixture of samples across multiple batches and makes semantic augmentation very efficient." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.47, + 0.649, + 0.485 + ], + "angle": 0, + "content": "3.5. Loss functions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.894, + 0.56 + ], + "angle": 0, + "content": "Photometric loss In the first pass, forward and backward flow \\( U_{1\\rightarrow 2}^{(l)}, U_{2\\rightarrow 1}^{(l)} \\) at each level are predicted and occlusion masks \\( O_{1\\rightarrow 2}^{(l)}, O_{2\\rightarrow 1}^{(l)} \\) are computed by forward-backward consistency [45]. Frames are warped by" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.566, + 0.856, + 0.589 + ], + "angle": 0, + "content": "\\[\nI _ {i} ^ {\\prime (l)} (\\pmb {p}) = I _ {j} ^ {(l)} (\\pmb {p} + U _ {i \\rightarrow j} ^ {(l)} (\\pmb {p})), \\quad (i, j \\in \\{1, 2 \\})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.895, + 0.727 + ], + "angle": 0, + "content": "where \\( I_{i}^{(l)} \\) is \\( I_{i} \\) down-sampled to the \\( l \\)-th scale, and \\( \\pmb{p} \\) denotes pixel coordinates at that scale. Following [37], three measures, namely \\( L_{1} \\)-distance \\( (\\rho_{1}) \\), structural similarity (SSIM) [73] \\( (\\rho_{2}) \\), and census loss [45] \\( (\\rho_{3}) \\) are linearly combined to measure photometric differences between \\( I_{i}^{(l)} \\) and \\( I_{i}^{\\prime(l)} \\). Occlusion regions are masked out and both forward and backward directions are taken into account at each level. The final photometric loss is" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.733, + 0.892, + 0.791 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\mathrm {p h}} = \\frac {1}{2} \\sum_ {(i, j) \\in \\{(1, 2), (2, 1) \\}} \\sum_ {l = 2} ^ {6} a _ {l} \\sum_ {k = 1} ^ {3} c _ {k} \\rho_ {k} \\left(I _ {i} ^ {(l)}, I _ {i} ^ {\\prime (l)}, O _ {i \\rightarrow j} ^ {(l)}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.808, + 0.894, + 0.854 + ], + "angle": 0, + "content": "Smoothness loss Unlike most previous work, we do not include a smoothness loss because we find it to conflict with our learned upsampler. See ablation study in Sec. 4.5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Augmentation loss As in ARFlow [37], our second forward pass computes the \\( \\mathrm{L}_1 \\)-distance between the trans-" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "9570" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.164, + 0.088, + 0.808, + 0.322 + ], + "angle": 0, + "content": "
MethodTrainTest
2012 EPE2015 EPE2012 Fl-nocEPEFl-all2015 Fl-nocFl-bgFl-fg
supervisedPWC-Net+ [66]-(1.50)3.361.47.724.917.697.88
IRR-PWC [22]-(1.63)3.211.67.654.867.687.52
RAFT [68]-(0.63)--5.103.074.746.87
Separable Flow [85]-(0.69)--4.532.784.255.92
unsupervisedSelFlow [39]1.694.844.312.214.199.6512.6821.74
SimFlow [24]-5.19--13.388.2112.6017.27
ARFlow [37]1.442.85-1.811.80---
UFlow [27]1.682.714.261.911.138.419.7817.87
UPFlow [42]1.272.45-1.49.38---
Ours (baseline)1.392.614.301.79.896.988.8215.21
Ours (+enc)†1.292.423.971.58.993.978.1913.01
Ours (+enc +aug)†1.282.183.901.58.383.907.4812.91
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.33, + 0.892, + 0.388 + ], + "angle": 0, + "content": "Table 1. KITTI benchmark results (EPE/px and Fl/%). Metrics evaluated at 'all' (all pixels, default for EPE), 'noc' (non-occlusions), 'bg' (background), and 'fg' (foreground). Key metrics (used in official ranking) are underlined. Our 'baseline' is an adapted ARFlow with added learned upsampler and no smoothness loss. +enc means adding semantic encoder; +aug means adding semantic augmentation. () means evaluation data used in training. - means unavailable. \\(\\dagger\\) denotes models with semantic inputs. For all metrics, lower is better." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.47, + 0.443 + ], + "angle": 0, + "content": "formed flow \\(\\hat{U}_{1\\rightarrow 2}\\) and the second pass output \\(\\hat{U}_{1\\rightarrow 2}^{\\prime} = \\mathcal{F}(\\hat{I}_1,\\hat{I}_2,\\hat{S}_1,\\hat{S}_2)\\) for each pixel \\(\\pmb{p}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.45, + 0.389, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\hat {D} (\\boldsymbol {p}) = \\left\\| \\hat {U} _ {1 \\rightarrow 2} (\\boldsymbol {p}) - \\hat {U} _ {1 \\rightarrow 2} ^ {\\prime} (\\boldsymbol {p}) \\right\\| _ {1}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.47, + 0.51 + ], + "angle": 0, + "content": "This loss is then averaged over the transformed non-occluded region where self-supervision \\(\\hat{U}_{1\\rightarrow 2}\\) is accurate" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.519, + 0.47, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\mathrm {a r}} = \\frac {\\sum_ {\\boldsymbol {p}} (1 - \\hat {O} _ {1 \\rightarrow 2} (\\boldsymbol {p})) \\hat {D} (\\boldsymbol {p})}{\\sum_ {\\boldsymbol {p}} (1 - \\hat {O} _ {1 \\rightarrow 2} (\\boldsymbol {p}))}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.47, + 0.614 + ], + "angle": 0, + "content": "For semantic augmentation (third pass), since we are also doing self-supervision with a new generated sample, we use a similar loss definition" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.622, + 0.469, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\mathrm {a u g}} = \\frac {\\sum_ {\\boldsymbol {p}} \\left(1 - \\tilde {O} _ {1 \\rightarrow 2} (\\boldsymbol {p})\\right) \\tilde {D} (\\boldsymbol {p})}{\\sum_ {\\boldsymbol {p}} \\left(1 - \\tilde {O} _ {1 \\rightarrow 2} (\\boldsymbol {p})\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.469, + 0.707 + ], + "angle": 0, + "content": "where \\(\\tilde{D} (\\pmb {p})\\) is the distance between the semanticaugmented flow \\(\\tilde{U}_{1\\rightarrow 2}\\) and the third-pass output \\(\\tilde{U}_{1\\rightarrow 2}^{\\prime}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.714, + 0.389, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\tilde {D} (\\boldsymbol {p}) = \\left\\| \\tilde {U} _ {1 \\rightarrow 2} (\\boldsymbol {p}) - \\tilde {U} _ {1 \\rightarrow 2} ^ {\\prime} (\\boldsymbol {p}) \\right\\| _ {1},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.424, + 0.759 + ], + "angle": 0, + "content": "and the augmented mask \\(\\tilde{O}_{1\\rightarrow 2}\\) is now computed by" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.767, + 0.388, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\tilde {O} _ {1 \\rightarrow 2} = 1 - \\max (1 - O _ {1 \\rightarrow 2}, M)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.841 + ], + "angle": 0, + "content": "because we penalize on both the originally non-occluded region \\(1 - O_{1\\rightarrow 2}\\) and the pasted foreground object mask \\(M\\) (of which we know the true motion)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.859, + 0.399, + 0.875 + ], + "angle": 0, + "content": "Final loss (with \\(\\lambda = 0.02\\) as in ARFlow [37]):" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.885, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\ell = \\ell_ {\\mathrm {p h}} + \\lambda \\left(\\ell_ {\\mathrm {a r}} + \\ell_ {\\mathrm {a u g}}\\right). \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.41, + 0.633, + 0.427 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.434, + 0.602, + 0.449 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.67 + ], + "angle": 0, + "content": "We mainly use KITTI [15, 46] and Cityscapes [8] datasets for experiments. Following [37], we train our model first on KITTI raw sequences [46] (55.7k samples) and then fine-tune on KITTI-2015 multi-view extension [46] (11.8k samples). We validate our model using KITTI-2015 train [46] (200 samples) and KITTI-2012 train set [15] (194 samples) since they are the only sets that have optical flow labels. We also train on Cityscapes [8] sequences (83.3k samples) to test model generalization ability, and we sample every other frame to match its frame rate with KITTI. All semantic segmentation inputs are estimated using an off-the-shelf model [89] with a DeepLabV3Plus [7] backbone, which achieves \\(83.45\\%\\) mean IoU on Cityscapes and \\(72.82\\%\\) mean IoU on KITTI." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.678, + 0.715, + 0.694 + ], + "angle": 0, + "content": "4.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.806 + ], + "angle": 0, + "content": "We implement the model in PyTorch [48] \\(^{1}\\). We use the Adam optimizer [32] with \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999\\) and batch size 4. We first train on KITTI raw sequences [46] for 100k iterations with a fixed learning rate 0.0002 and then train on KITTI multi-view extension set [15, 46] for another 100k iterations using OneCycleLR schedule [62] with maximum learning rate 0.0004 and linear annealing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.806, + 0.892, + 0.868 + ], + "angle": 0, + "content": "For data augmentation, we include random horizontal flipping and swapping of the input frames. We resize the inputs to \\(256 \\times 832\\) before feeding into the network. The photometric loss weight for each scale \\(a_{l}\\) (\\(2 \\leq l \\leq 6\\)) in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)Code and instructions are available at https://github.com/duke-vision/semantic-unsup-flow-release." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9571" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.445, + 0.183 + ], + "angle": 0, + "content": "
MethodFl-allFl-nocFl-bgFl-fg
JFS [21]17.079.8115.922.92
SOF [57]16.8110.8614.6327.73
MRFlow [75]12.198.8610.1322.52
Bai et al. [1]11.628.758.6126.69
Ours (final)8.383.907.4812.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.193, + 0.468, + 0.234 + ], + "angle": 0, + "content": "Table 2. KITTI-2015 test results (Fl/\\%) compared with other semantic optical flow methods. Metrics evaluated at 'all' (all pixels), 'noc' (non-occlusions), 'bg' (background), and 'fg' (foreground)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.268, + 0.468, + 0.358 + ], + "angle": 0, + "content": "Eq. (1) are set as 1, 1, 1, 1, 0. The weights for three photometric distance measures \\(\\rho_{k}\\) (\\(1 \\leq k \\leq 3\\)) in Eq. (1) are set as 0.15, 0.85, 0 for the first 50k iterations and 0, 0, 1 afterwards. We start the appearance and spatial augmentation (second pass as in ARFlow [37]) at 50k iterations, and semantic augmentation (third pass) after 150k iterations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.38, + 0.314, + 0.396 + ], + "angle": 0, + "content": "4.3. Benchmark testing results" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.407, + 0.468, + 0.483 + ], + "angle": 0, + "content": "As common practice, we evaluate optical flow predictions based on two error measurements, Fl (error rate) and EPE (mean \\( \\mathrm{L}_2 \\) distance). When computing Fl, the estimate of each pixel is considered correct if the error is smaller than 3 pixels or \\( 5\\% \\) of the magnitude of ground-truth flow [46]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.468, + 0.637 + ], + "angle": 0, + "content": "As shown in Tab. 1, our semantic modules (both semantic encoder and semantic augmentation) improve performance on both KITTI-2012 and KITTI-2015 sets on all metrics. Our final model achieves \\(8.38\\%\\) F1-all error rate on KITTI-2015 test set, which is significantly better than ARFlow [37] \\((11.80\\%)\\), from which we adapt. We also outperform the current state-of-the-art UPFlow [42] \\((9.38\\%)\\) by a clear margin. All these results strongly suggest that adding semantic inputs can help improve unsupervised flow estimation significantly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.468, + 0.732 + ], + "angle": 0, + "content": "In addition, Tab. 2 shows that our test results significantly outperform all previous semantic optical flow methods, most of which are based on traditional energy minimization. We are the first to apply semantic inputs to recent unsupervised flow networks, so we are able to push the state-of-the-art by a clear margin." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.753, + 0.254, + 0.768 + ], + "angle": 0, + "content": "4.4. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Some qualitative results are shown in Fig. 4. We can see that our flow outputs have very sharp boundaries and very clean object motion, especially around cars and poles. Our model successfully learns to adapt flow estimation to the semantic map input. Moreover, our model is able to handle very challenging samples where the foreground car motion is drastically different from the background due to the challenging samples we create in semantic augmentation." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.089, + 0.887, + 0.245 + ], + "angle": 0, + "content": "
upno smencaugFl-allEPE-allEPE-nocEPE-occ
10.362.902.076.97
19.922.691.896.57
29.832.651.866.40
39.752.611.856.54
49.752.641.856.55
✓10.222.831.976.75
✓✓8.872.611.856.04
✓✓38.262.421.735.79
✓✓✓8.802.481.606.64
✓✓3✓7.792.181.405.64
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.256, + 0.892, + 0.326 + ], + "angle": 0, + "content": "Table 3. Ablation study on KITTI-2015 train set (EPE/px and F1%). 'up' means the learned upsampler; 'no sm' means turning off smoothness loss; 'enc' means the level at which we merge semantic features and image features in the encoder; 'aug' means semantic augmentation." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.349, + 0.891, + 0.43 + ], + "angle": 0, + "content": "
Options of augFl-allEPE-allEPE-nocEPE-occ
Ours (final)7.792.181.405.64
start from 100k7.922.191.405.41
vehicles only7.942.211.445.78
loss on new occ8.152.271.425.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.44, + 0.892, + 0.468 + ], + "angle": 0, + "content": "Table 4. Ablation study of different semantic augmentation options on KITTI-2015 train (EPE/px and F1/%). See text for explanations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.504, + 0.649, + 0.52 + ], + "angle": 0, + "content": "4.5. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.699 + ], + "angle": 0, + "content": "We also do ablation study to show the effectiveness of each of our proposed modules. We can see in Tab. 3 that adding a semantic encoder to the vanilla ARFlow does help, and the optimal number of encoder layer groups added is 3. Also, our learned upsampler improves results significantly once we turn off the smoothness loss. This makes sense because the upsampler learns to adapt to boundaries, which can be distracted by the smoothness loss. Moreover, both our semantic encoder and semantic augmentation module help improve the results further, which suggests the effectiveness of adding semantic inputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Another ablation study is on different options for the semantic augmentation (Tab. 4). We try starting augmentation earlier from 100k iterations, which works slightly worse. There are mainly two reasons: (1) we use the first forward output to self-supervise our augmented output, so if we start early, the model will use poor output to self-supervise the augmented pass; and (2) our semantic augmentation creates very challenging samples with objects moving very differently from the background, so exposing these hard samples to the model too early may make it hard to train. Apart from this, we also try only using vehicles to augment as well as focusing loss on the newly occluded region, which are both inferior to our final version." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9572" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.242, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.286, + 0.203, + 0.3 + ], + "angle": 0, + "content": "Image Input" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.089, + 0.403, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.286, + 0.378, + 0.3 + ], + "angle": 0, + "content": "Semantics Input" + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.089, + 0.565, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.286, + 0.512, + 0.298 + ], + "angle": 0, + "content": "ARFlow" + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.089, + 0.728, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.286, + 0.674, + 0.298 + ], + "angle": 0, + "content": "UPFlow" + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.089, + 0.892, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.286, + 0.826, + 0.298 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.304, + 0.894, + 0.334 + ], + "angle": 0, + "content": "Figure 4. Qualitative results on KITTI test set (sample #7, 20, 38, 112, 183) compared with ARFlow [37] and UPFlow [42]. We only show first frame input for conciseness. Our outputs have much clearer boundaries around objects. For more qualitative examples, see Appendix." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.357, + 0.849, + 0.455 + ], + "angle": 0, + "content": "
roadcarterrainvegetationsidewalkbuildingwallpole
Proportion42.4%17.6%14.0%11.6%6.2%4.1%1.3%1.1%
ARFlow [37]4.5715.799.4215.344.616.0016.3410.75
Ours (final)3.6710.178.7413.473.094.2711.189.44
Relative improvement19.6%35.6%7.3%12.2%33.0%28.7%31.6%12.2%
Reweighed contribution19.0%49.4%4.8%10.9%4.7%3.5%3.4%0.7%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.463, + 0.893, + 0.493 + ], + "angle": 0, + "content": "Table 5. KITTI-2015 flow error (Fl-all/%) for each semantic class. The first row shows the proportion of each class among evaluated pixels (KITTI flow does not evaluate on every pixel). We only show classes that account for at least \\(1\\%\\) of the evaluated pixels." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.516, + 0.285, + 0.532 + ], + "angle": 0, + "content": "4.6. Improvement analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.47, + 0.661 + ], + "angle": 0, + "content": "It may be interesting to understand where our improvements come from in terms of semantic classes, so we compute the KITTI-2015 train set error for each class. As shown in Tab. 5, our model improves the flow on every class. After reweighing the absolute improvement of each class based on their proportion in the evaluated pixels, we find that car flow accounts for nearly half of our improvement on Fl-all, indicating the effectiveness of augmenting vehicle objects." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.671, + 0.283, + 0.687 + ], + "angle": 0, + "content": "4.7. Generalization ability" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.695, + 0.47, + 0.831 + ], + "angle": 0, + "content": "We also investigate the generalization ability of our semantic-aided flow models. We train flow models on Cityscapes and directly test them on KITTI-2015 train set without fine-tuning. Following [50], we crop the bottom \\(25\\%\\) of the frame to remove the car logo and resize the frame to size \\(256 \\times 704\\). All other experiment settings are exactly the same as for KITTI. As we can see from Tab. 6, adding semantic inputs significantly helps our unsupervised flow model generalize and adapt better across datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.338, + 0.861 + ], + "angle": 0, + "content": "5. Conclusion and future work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In this paper, we show that adding semantic segmentation inputs can help significantly improve the performance" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.513, + 0.89, + 0.594 + ], + "angle": 0, + "content": "
MethodFl-allEPE-allEPE-nocEPE-occ
ARFlow (our impl.)13.214.082.889.40
Ours (baseline)12.273.812.439.91
Ours (+enc)†11.283.332.128.75
Ours (+enc +aug)†10.322.641.567.12
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.893, + 0.687 + ], + "angle": 0, + "content": "Table 6. Generalization results (train on Cityscapes, and test on KITTI-2015 train). We use our implementation of ARFlow [37] to run this experiment. Our 'baseline' is an adapted ARFlow with added learned upsampler and no smoothness loss. \\(^+\\) enc means adding semantic encoder; \\(^+\\) aug means adding semantic augmentation. \\(\\dagger\\) denotes models with semantic inputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.892, + 0.868 + ], + "angle": 0, + "content": "of unsupervised optical flow networks on autonomous driving datasets. We propose a novel network model called SemARFlow, a semantic adaptation of ARFlow [37], with a semantic encoder, a learned upsampler, and a semantic augmentation module, where some domain-specific motion prior has been taken into account. Our network better predicts flow at occlusion regions and effectively sharpens flow estimates around object boundaries, even for very challenging samples. The additional semantic inputs also make our network generalize better across datasets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "One direction for future work is motivated by the example illustrated in Fig. 5. Our model does output sharp flow" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9573" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.209, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.092, + 0.338, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.091, + 0.467, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.154, + 0.47, + 0.182 + ], + "angle": 0, + "content": "Figure 5. An example on KITTI-2015 train set (sample #164). See the discussion in Sec. 5 for details." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.219, + 0.471, + 0.339 + ], + "angle": 0, + "content": "boundaries between the cars and the background, thanks to good boundary information from the semantic maps. However, the boundary between these two fast-moving cars is inferred from image information only, because their semantic maps merge into one. As a result, that part of the flow boundary is less accurate. An interesting direction for future work is to use instance-level semantics as input for more detailed object masks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.36, + 0.47, + 0.406 + ], + "angle": 0, + "content": "Acknowledgments. This research is based upon work supported by the National Science Foundation under Grant No. 1909821." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.42, + 0.174, + 0.436 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.446, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[1] Min Bai, Wenjie Luo, Kaustav Kundu, and Raquel Urtasun. Exploiting semantic information and deep matching for optical flow. In Proceedings of the European Conference on Computer Vision, pages 154-170. Springer, 2016. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.504, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[2] Shubhankar Borse, Ying Wang, Yizhe Zhang, and Fatih Porikli. Inverseform: A loss function for structured boundary-aware segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5901-5911, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.575, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[3] Gary Bradski. The opencv library. Dr. Dobb's Journal: Software Tools for the Professional Programmer, 25(11):120-123, 2000. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.618, + 0.469, + 0.672 + ], + "angle": 0, + "content": "[4] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Proceedings of the European Conference on Computer Vision, pages 611-625. Springer, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.675, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[5] Yingfeng Cai, Lei Dai, Hai Wang, and Zhixiong Li. Multi-target pan-class intrinsic relevance driven model for improving semantic segmentation in autonomous driving. IEEE Transactions on Image Processing, 30:9069-9084, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.732, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[6] Linda Capito, Umit Ozguner, and Keith Redmill. Optical flow based visual potential field for autonomous driving. In IEEE Intelligent Vehicles Symposium, pages 885-891. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.789, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European Conference on Computer Vision, pages 801-818, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.446, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "dataset for semantic urban scene understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3213-3223, 2016. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.137, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[9] Mingyu Ding, Zhe Wang, Bolei Zhou, Jianping Shi, Zhiwu Lu, and Ping Luo. Every frame counts: Joint learning of video segmentation and optical flow. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10713-10720, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Philipp Fischery, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick van der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE International Conference on Computer Vision, ICCV '15, pages 2758-2766, Washington, DC, USA, 2015. IEEE Computer Society. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.306, + 0.892, + 0.358 + ], + "angle": 0, + "content": "[11] Cheng Feng, Long Ma, Congxuan Zhang, Zhen Chen, Liyue Ge, and Shaofeng Jiang. Ss-sf: Piecewise 3d scene flow estimation with semantic segmentation. IEEE Access, 9:22745-22759, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.43 + ], + "angle": 0, + "content": "[12] Jun Fu, Jing Liu, Haijie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3146-3154, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.514 + ], + "angle": 0, + "content": "[13] Aditya Ganeshan, Alexis Vallet, Yasunori Kudo, Shin-ichi Maeda, Tommi Kerola, Rares Ambrus, Dennis Park, and Adrien Gaidon. Warp-refine propagation: Semi-supervised auto-labeling via cycle-consistency. In Proceedings of the IEEE International Conference on Computer Vision, pages 15499–15509, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.518, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[14] Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In Proceedings of the European Conference on Computer Vision, pages 713-729. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.575, + 0.892, + 0.628 + ], + "angle": 0, + "content": "[15] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research, 32(11):1231-1237, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[16] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.703, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[18] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial Intelligence, 17(1-3):185–203, 1981. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[19] Zilong Huang, Xinggang Wang, Lichao Huang, Chang Huang, Yunchao Wei, and Wenyu Liu. Ccnet: Criss-cross attention for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 603-612, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[20] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. Liteflownet: A lightweight convolutional neural network for optical flow estimation. In Proceedings of the IEEE Conference" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9574" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 8981-8989, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[21] Junhwa Hur and Stefan Roth. Joint optical flow and temporally consistent semantic segmentation. In Proceedings of the European Conference on Computer Vision, pages 163-177. Springer, 2016. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[22] Junhwa Hur and Stefan Roth. Iterative residual refinement for joint optical flow and occlusion estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5754-5763, 2019. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[23] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1647-1655, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.362 + ], + "angle": 0, + "content": "[24] Woobin Im, Tae-Kyun Kim, and Sung-Eui Yoon. Unsupervised learning of optical flow with deep feature similarity. In Proceedings of the European Conference on Computer Vision, pages 172-188. Springer, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.419 + ], + "angle": 0, + "content": "[25] Joel Janai, Fatma Guney, Anurag Ranjan, Michael Black, and Andreas Geiger. Unsupervised learning of multi-frame optical flow with occlusions. In Proceedings of the European Conference on Computer Vision, pages 690-706, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[26] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In Proceedings of the IEEE International Conference on Computer Vision, pages 9772-9781, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[27] Rico Jonschkowski, Austin Stone, Jonathan T Barron, Ariel Gordon, Kurt Konolige, and Anelia Angelova. What matters in unsupervised optical flow. In Proceedings of the European Conference on Computer Vision, pages 557-572. Springer, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.617 + ], + "angle": 0, + "content": "[28] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM computing surveys, 54(10s):1-41, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[29] Hannah H Kim, Celia Cintas, Girmaw Abebe Tadesse, and Skyler Speakman. Spatially constrained adversarial attack detection and localization in the representation space of optical flow networks. In International Joint Conference on Artificial Intelligence, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[30] Hannah Halin Kim, Shuzhi Yu, and Carlo Tomasi. Joint detection of motion boundaries and occlusions. In Proceedings of the British Machine Vision Conference, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[31] Hannah Halin Kim, Shuzhi Yu, Shuai Yuan, and Carlo Tomasi. Cross-attention transformer for video interpolation. In Proceedings of the Asian Conference on Computer Vision Workshops, pages 320–337, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[32] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. Proceedings of the International Conference on Learning Representations, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[33] Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weissenborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Sylvain Gelly, Thomas Unterthiner, and Xiaohua Zhai. An image is worth 16x16 words: Transformers for image recogni" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "tion at scale. In Proceedings of the International Conference on Learning Representations, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[34] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[35] Xiangtai Li, Jiangang Bai, Kuiyuan Yang, and Yunhai Tong. Flow2seg: Motion-aided semantic segmentation. In Proceedings of the International Conference on Artificial Neural Networks, pages 225–237. Springer, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[36] Zewen Li, Fan Liu, Wenjie Yang, Shouheng Peng, and Jun Zhou. A survey of convolutional neural networks: analysis, applications, and prospects. IEEE transactions on neural networks and learning systems, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.275, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[37] Liang Liu, Jiangning Zhang, Ruifei He, Yong Liu, Yabiao Wang, Ying Tai, Donghao Luo, Chengjie Wang, Jilin Li, and Feiyue Huang. Learning by analogy: Reliable supervision from transformations for unsupervised optical flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6489-6498, 2020. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[38] Pengpeng Liu, Irwin King, Michael R Lyu, and Jia Xu. Ddflow: Learning optical flow with unlabeled data distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8770-8777, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[39] Pengpeng Liu, Michael Lyu, Irwin King, and Jia Xu. Selfflow: Self-supervised learning of optical flow. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4571-4580, 2019. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[40] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3431-3440, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[41] Bruce D Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In International Joint Conference on Artificial Intelligence, volume 2, pages 674-679, 1981. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.596, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[42] Kunming Luo, Chuan Wang, Shuaicheng Liu, Haoqiang Fan, Jue Wang, and Jian Sun. Upflow: Upsampling pyramid for unsupervised optical flow learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1045-1054, 2021. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.732 + ], + "angle": 0, + "content": "[43] Rémi Marsal, Florian Chabot, Angélique Loesch, and Hichem Sahbi. Brightflow: Brightness-change-aware unsupervised learning of optical flow. In Proceedings of the IEEE winter conference on applications of computer vision, pages 2061-2070, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.734, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[44] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4040-4048, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[45] Simon Meister, Junhwa Hur, and Stefan Roth. Unflow: Unsupervised learning of optical flow with a bidirectional census loss. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[46] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE Conference" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "9575" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, 2015. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.176 + ], + "angle": 0, + "content": "[47] Hyeonwoo Noh, Seunghoon Hong, and Bohyung Han. Learning deconvolution network for semantic segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1520-1528, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.468, + 0.3 + ], + "angle": 0, + "content": "[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, pages 8024-8035. Curran Associates, Inc., 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.468, + 0.357 + ], + "angle": 0, + "content": "[49] Anurag Ranjan and Michael J. Black. Optical flow estimation using a spatial pyramid network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2720-2729, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.468, + 0.44 + ], + "angle": 0, + "content": "[50] Anurag Ranjan, Varun Jampani, Lukas Balles, Kihwan Kim, Deqing Sun, Jonas Wulff, and Michael J Black. Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12240–12249, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.468, + 0.522 + ], + "angle": 0, + "content": "[51] Hazem Rashed, Senthil Yogamani, Ahmad El-Sallab, Pavel Krizek, and Mohamed El-Helw. Optical flow augmented semantic segmentation networks for automated driving. In International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, pages 165-172, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.468, + 0.594 + ], + "angle": 0, + "content": "[52] Fitsum A Reda, Guilin Liu, Kevin J Shih, Robert Kirby, Jon Barker, David Tarjan, Andrew Tao, and Bryan Catanzaro. Sdc-net: Video prediction using spatially-displaced convolution. In Proceedings of the European Conference on Computer Vision, pages 718-733, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.468, + 0.662 + ], + "angle": 0, + "content": "[53] Zhile Ren, Orazio Gallo, Deqing Sun, Ming-Hsuan Yang, Erik B Sudderth, and Jan Kautz. A fusion approach for multiframe optical flow estimation. In Proceedings of the IEEE winter conference on applications of computer vision, pages 2077-2086. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.468, + 0.719 + ], + "angle": 0, + "content": "[54] Zhile Ren, Deqing Sun, Jan Kautz, and Erik Sudderth. Cascaded scene flow prediction using semantic segmentation. In Proceedings of the International Conference on 3D Vision, pages 225-233. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.468, + 0.775 + ], + "angle": 0, + "content": "[55] Zhe Ren, Junchi Yan, Bingbing Ni, Bin Liu, Xiaokang Yang, and Hongyuan Zha. Unsupervised deep learning for optical flow estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.777, + 0.468, + 0.829 + ], + "angle": 0, + "content": "[56] Stephan R. Richter, Zeeshan Hayder, and Vladlen Koltun. Playing for benchmarks. In Proceedings of the IEEE International Conference on Computer Vision, pages 2232-2241, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[57] Laura Sevilla-Lara, Deqing Sun, Varun Jampani, and Michael J Black. Optical flow with semantic segmentation and localized layers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3889-3898, 2016. 2, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[58] Shihao Shen, Louis Kerofsky, and Senthil Yogamani. Optical flow for autonomous driving: Applications, challenges and improvements. arXiv preprint arXiv:2301.04422, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[59] Jianbo Shi and Carlo Tomasi. Good features to track. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 593-600, 1994. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[60] Yukang Shi and Kaiheng Ma. Safit: Segmentation-aware scene flow with improved transformer. In Proceedings of the International Conference on Robotics and Automation, pages 10648-10655. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[61] Jeongho Shin, Sangjin Kim, Sangkyu Kang, Seong-Won Lee, Joonki Paik, Besma Abidi, and Mongi Abidi. Optical flow-based real-time object tracking using non-prior training active feature model. Real-time imaging, 11(3):204-218, 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[62] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, volume 11006, pages 369-386. SPIE, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.378, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[63] Austin Stone, Daniel Maurer, Alper Ayvaci, Anelia Angelova, and Rico Jonschkowski. Smurf: Self-teaching multiframe unsupervised raft with full-image warping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3887-3896, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[64] Xiuchao Sui, Shaohua Li, Xue Geng, Yan Wu, Xinxing Xu, Yong Liu, Rick Goh, and Hongyuan Zhu. Craft: Cross-attentional flow transformer for robust optical flow. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 17602-17611, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.52, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[65] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8934-8943, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[66] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Models matter, so does training: An empirical study of cnns for optical flow estimation. IEEE transactions on pattern analysis and machine intelligence, 42(6):1408-1423, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[67] Andrew Tao, Karan Sapra, and Bryan Catanzaro. Hierarchical multi-scale attention for semantic segmentation. arXiv preprint arXiv:2005.10821, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[68] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proceedings of the European Conference on Computer Vision, pages 402-419. Springer, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[69] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Proceedings of the International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[70] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE Con" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9576" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "ference on Computer Vision and Pattern Recognition Workshops, pages 969-977, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[71] Panqu Wang, Pengfei Chen, Ye Yuan, Ding Liu, Zehua Huang, Xiaodi Hou, and Garrison Cottrell. Understanding convolution for semantic segmentation. In Proceedings of the IEEE winter conference on applications of computer vision, pages 1451-1460. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.468, + 0.261 + ], + "angle": 0, + "content": "[72] Yang Wang, Yi Yang, Zhenheng Yang, Liang Zhao, Peng Wang, and Wei Xu. Occlusion aware unsupervised learning of optical flow. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4884-4893, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.468, + 0.319 + ], + "angle": 0, + "content": "[73] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.468, + 0.376 + ], + "angle": 0, + "content": "[74] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Song Wang, and Lili Ju. Semantic stereo matching with pyramid cost volumes. In Proceedings of the IEEE International Conference on Computer Vision, pages 7484-7493, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.468, + 0.433 + ], + "angle": 0, + "content": "[75] Jonas Wulff, Laura Sevilla-Lara, and Michael J Black. Optical flow in mostly rigid scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4671-4680, 2017. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.468, + 0.503 + ], + "angle": 0, + "content": "[76] Haofei Xu, Jiaolong Yang, Jianfei Cai, Juyong Zhang, and Xin Tong. High-resolution optical flow from 1d attention and correlation. In Proceedings of the IEEE International Conference on Computer Vision, pages 10498-10507, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.468, + 0.573 + ], + "angle": 0, + "content": "[77] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8121-8130, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.468, + 0.644 + ], + "angle": 0, + "content": "[78] Guorun Yang, Hengshuang Zhao, Jianping Shi, Zhidong Deng, and Jiaya Jia. Segstereo: Exploiting semantic information for disparity estimation. In Proceedings of the European Conference on Computer Vision, pages 636-651, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.468, + 0.716 + ], + "angle": 0, + "content": "[79] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. Bisenet v2: Bilateral network with guided aggregation for real-time semantic segmentation. International Journal of Computer Vision, 129:3051-3068, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.468, + 0.785 + ], + "angle": 0, + "content": "[80] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceedings of the European Conference on Computer Vision, pages 325-341, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[81] Jason J. Yu, Adam W. Harley, and Konstantinos G. Derpanis. Back to basics: Unsupervised learning of optical flow via brightness constancy and motion smoothness. In Gang Hua and Hervé Jégou, editors, Proceedings of the European Conference on Computer Vision Workshops, pages 3-10, Cham, 2016. Springer International Publishing. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[82] Shuzhi Yu, Hannah H Kim, Shuai Yuan, and Carlo Tomasi. Unsupervised flow refinement near motion boundaries. In" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.89, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the British Machine Vision Conference. BMVA Press, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[83] Shuai Yuan, Xian Sun, Hannah Kim, Shuzhi Yu, and Carlo Tomasi. Optical flow training under limited label budget via active learning. In Proceedings of the European Conference on Computer Vision, pages 410-427, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[84] Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Common practices and emerging technologies. IEEE access, 8:58443-58469, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[85] Feihu Zhang, Oliver J Woodford, Victor Adrian Prisacariu, and Philip HS Torr. Separable flow: Learning motion cost volumes for optical flow estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 10807-10817, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[86] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2881-2890, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.458 + ], + "angle": 0, + "content": "[87] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6881-6890, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.46, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[88] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In Proceedings of the International Conference on Learning Representations, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.531, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[89] Yi Zhu, Karan Sapra, Fitsum A Reda, Kevin J Shih, Shawn Newsam, Andrew Tao, and Bryan Catanzaro. Improving semantic segmentation via video propagation and label relaxation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8856-8865, 2019. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9577" + } + ] +] \ No newline at end of file diff --git a/2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/fa7524e1-a76d-4844-9361-43c8b8046bf9_origin.pdf b/2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/fa7524e1-a76d-4844-9361-43c8b8046bf9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7602c01b26b9f24f6b4188541f778d9bd2b0e797 --- /dev/null +++ b/2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/fa7524e1-a76d-4844-9361-43c8b8046bf9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc15db8dd049638629bf26f8c28502c4c186f10218a672b3100206f81601f834 +size 5219511 diff --git a/2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/full.md b/2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..34775138fa7b328a93ada5cc366e775e4cf10ad0 --- /dev/null +++ b/2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/full.md @@ -0,0 +1,395 @@ +# SemARFlow: Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving + +Shuai Yuan, Shuzhi Yu, Hannah Kim, and Carlo Tomasi +Duke University + +{shuai, shuzhiyu, hannah, tomasi}@cs.duke.edu + +# Abstract + +Unsupervised optical flow estimation is especially hard near occlusions and motion boundaries and in low-texture regions. We show that additional information such as semantics and domain knowledge can help better constrain this problem. We introduce SemARFlow, an unsupervised optical flow network designed for autonomous driving data that takes estimated semantic segmentation masks as additional inputs. This additional information is injected into the encoder and into a learned upsampler that refines the flow output. In addition, a simple yet effective semantic augmentation module provides self-supervision when learning flow and its boundaries for vehicles, poles, and sky. Together, these injections of semantic information improve the KITTI-2015 optical flow test error rate from $11.80\%$ to $8.38\%$ . We also show visible improvements around object boundaries as well as a greater ability to generalize across datasets. Code is available at https://github.com/duke-vision/semantic-unsup-flow-release. + +# 1. Introduction + +Optical flow estimation, i.e., pixel-level motion tracking across video frames, has broad applications in many computer vision tasks that include object tracking [61], video editing [14, 31], and autonomous driving [6, 58]. + +Thanks to the success of deep convolutional neural networks [34, 17, 36] and transformer networks [33, 69, 88, 28] in computer vision, many top-performing supervised optical flow networks have been proposed in recent years [10, 65, 22, 68, 85], in which ground-truth labels supervise training. However, real optical flow is hard to label, so most supervised methods train (or at least pre-train) on synthetic datasets [10, 44, 4, 56] which makes them hard to adapt to real applications due to the significant gap between synthetic and real data [70, 29]. + +Due to label scarcity, unsupervised training of optical + +![](images/9e65a6cce7dfec507031a1407f7e49721073b9180ae490256c7b968ba3697ddc.jpg) +Figure 1. An example on KITTI-2015 test set (sample #63). Only one input frame is shown for conciseness. Our SemARFlow takes additional semantic segmentation inputs (estimated by an off-the-shelf model) and outputs much sharper flow around semantic boundaries. + +![](images/343f36ac10bb396cab4abe94b28fbb130cf086e612f0ae4d6d44fd01bbc59452.jpg) + +![](images/81f42782b8b7e34d6a081d106bfccb2d41438995034e911df052f59146cd50c3.jpg) + +![](images/85dc6b3c2243ce70f0fa7c131cc53de504af6f1b5e13f883b0520d0b57d3ac48.jpg) + +![](images/b36e7fc92d3acc591e035ba081195e897be82f0c0d8276200009904df42750ff.jpg) + +flow estimators [81] instead uses loss terms based on the assumptions of constant brightness and smooth flow [45, 27, 42]. Some self-supervision techniques have also been studied to enhance model performance [38, 39, 37]. Unsupervised training makes it possible to train flow networks directly on large real datasets from the target domain. + +Even so, unsupervised optical flow estimation is a poorly constrained problem. Brightness is not constant in regions with occlusions [72] or on shiny surfaces [43], nor is it smooth across motion boundaries [30, 82]. Moreover, points in regions with poor texture content [59] or in dark shadows [58] are difficult to track as they are easily confused with neighboring points (the so-called aperture problem). This dearth of reliable constraints makes the unsupervised training of flow networks especially challenging. + +A natural way to address this issue is to inject additional constraints in the form of semantics and domain knowledge. For example, in autonomous driving applications [84, 6], we have clear expectations about object types and layouts of the scene, as well as prior knowledge of how each type of object typically moves. We focus on this application domain and show that this additional information helps the network achieve better flow results. + +To make an autonomous driving system work well in + +reality, it is best to train it on real data. However, annotating real driving video with optical flow labels is expensive [83], as it requires careful synchronization and calibration of diverse sensors including cameras, LiDAR, and GPS/IMU [16], aided by some manual annotation and curation based on CAD models of moving objects [46]. + +In contrast, annotating semantic labels seems much more feasible, and indeed semantic labels are available in most (if not all) existing driving datasets. We consider semantic segmentation because it provides semantics at the pixel level, the same level as optical flow. As one of the most popular and well-studied tasks in modern computer vision, semantic segmentation [40, 47, 87] has been extensively adopted for autonomous driving systems. In this paper, we show that adding semantic segmentation inputs helps improve unsupervised optical flow performance significantly. + +Specifically, we first infer semantic segmentation maps using an off-the-shelf model [89], which of course is trained with semantic labels. An encoder with semantic map input is used to aggregate image and semantic features (Sec. 3.1), and a learned upsampler is added into the iterative decoder to refine flow around object boundaries given semantic inputs (Sec. 3.2). We also propose a simple yet effective semantic augmentation module for self-supervision, which provides realistic augmentations specific to the vehicles, poles, and sky classes based on domain knowledge and segmentation maps (Sec. 3.3). An occluder cache is implemented to improve efficiency (Sec. 3.4). Semantic augmentation provides challenging samples for self-supervision, which help train flow better in occluded regions and around foreground object boundaries. + +Overall, by injecting semantic segmentation inputs, our SemARFlow network achieves significantly better unsupervised flow results both quantitatively (Sec. 4.3) and qualitatively (Sec. 4.4). Adapted from ARFlow [37], SemARFlow reduces KITTI-2015 [46] test error from $11.80\%$ to $8.38\%$ and out-performs the current unsupervised state-of-the-art UPFlow [42] $(9.38\%)$ by a clear margin. The example in Fig. 1 also demonstrates visible improvements around object boundaries. The effectiveness of each module is justified through an extensive ablation study (Sec. 4.5) and improvement analysis (Sec. 4.6). In addition to performance boost, unsupervised flow networks with additional semantic inputs generalize better across different datasets (Sec. 4.7). + +Our research is essentially novel compared to previous approaches. Some early work incorporates semantics in traditional energy-minimization methods [41, 18] for optical flow through geometric constraints such as piece-wise rigid motion and planar surface motion [57, 21, 1, 75]. In comparison, to the best of our knowledge, we are the first to inject semantics into the unsupervised training of recent optical flow networks. Some research also trains a network for segmentation and optical flow jointly [9]. In contrast, we + +leverage existing, separately trained segmentation systems both because they are available and because modularity—separating segmentation from flow estimation—is important for the development of large real-application systems such as autonomous driving. + +In summary, our contributions are as follows. + +- To the best of our knowledge, we are the first to explore adding semantic inputs to assist the unsupervised training of deep optical flow networks. +- We propose a simple yet effective network called SemARFlow that achieves state-of-the-art results both quantitatively and qualitatively. Our model works well on real-life occlusions and yields sharp motion boundaries around objects. +- We provide full training and inference code as well as trained models to encourage follow-up research. + +# 2. Related work + +Supervised optical flow Since the introduction of the pioneering deep optical flow network FlowNet [10], more and more top-performing CNN-based flow estimators have been proposed over the years [23, 49, 65, 20, 22, 68]. Recently, vision transformer networks and attention mechanism have also been applied to this problem, and these have achieved state-of-the-art performance on benchmark datasets [76, 26, 85, 64, 77]. The supervised methods are often pre-trained on large synthetic datasets such as FlyingChairs [10] and FlyingThings3D [44] before fine-tuning on the target dataset. However, there is a clear gap between the artificially generated data and real scenarios. + +Unsupervised optical flow Due to the lack of ground-truth labels, unsupervised optical flow estimation uses surrogate losses such as photometric loss and smoothness loss to supervise training [81, 55]. To tackle the issues around occlusion regions, various methods have been proposed including estimated occlusion masks [72], forward-backward consistency [45], and multi-frame fusion [25, 53]. To better upsample the flow in the decoder, UPFlow [42] additionally predicts a confidence map and an interpolation flow to guide flow refinement and has become the state-of-the-art unsupervised flow method. + +Some latest research has also explored the use of self-supervision to enhance flow prediction. Early methods apply the knowledge distillation technique to train a two-stage teacher-student network [38, 39]. ARFlow [37] further improves this idea by generating reliable self-supervision signals from data transformations, while merging the two training stages into single-stage training with one added loss term. SimFlow [24] replaces the handcrafted features with deep self-supervised features to measure similarity in the + +unsupervised losses. SMURF [63] utilizes a RAFT-like structure [68] and applies multi-frame self-supervised training with many technical improvements. Our SemARFlow also uses self-supervision but with the guidance of semantic segmentation, which is much more realistic than guidance without semantics. + +Semantic segmentation Semantic segmentation classifies each pixel of the given image into semantic objects. Fully Convolutional Network (FCN) [40] is one of the early CNN-based segmentation methods. It takes inputs of arbitrary sizes and outputs dense pixel-level predictions, becoming one of the popular backbone architectures for follow-up work. Deconvolution network [47] is also proposed to better recover low-level details of the prediction. One main challenge for these systems is that they lack global scene information. Subsequent work addresses this issue by enlarging the receptive field of the network with global pyramid pooling layers as in PSPNet [86], hybrid dilated convolutions [71], and a fast-down-sampling strategy [80, 79]. Attention modules have also shown to help in semantic segmentation by capturing full-image dependencies of all pixels [19, 87] or the semantic inter-dependencies across spatial and channel dimensions [12]. + +There also exists extensive work on semantic segmentation in the context of autonomous driving [2, 67, 13, 5, 89], thanks to the publication of large-scale driving datasets [8, 46]. To better train the network on a coarsely labeled dataset like KITTI [46], Zhu et al. use a video prediction network SDCNet [52] to synthesize new training samples with relaxed label propagation [89]. Due to their good performance on KITTI [46], we utilize their network models to infer semantic inputs for all our experiments. + +Combining semantics and optical flow Though there has been much progress on both semantic segmentation and optical flow estimation, the semantic optical flow problem (how to exploit semantics to help optical flow estimation) has received limited attention in recent years, and the current best results are thus much outdated. Some early methods incorporate semantics through geometric constraints to refine flow on various semantic regions, such as planar regions (using homographies) [57], static regions (using rigid camera motion and epipolar constraints) [21, 75], and rigid objects (estimating rigid motion for each object instance) [1]. However, most methods are traditional flow methods based on energy minimization, where an initial flow estimate is usually needed and semantics is mostly used for refinement. In comparison, we explore adapting latest unsupervised optical flow networks to leverage semantic inputs in one single stage of estimation. + +Apart from using semantics to help flow, some research has also explored using optical flow to help seman + +tic segmentation [51, 35], or to train both tasks jointly [9]. There are also studies on exploiting semantics on some other correspondence matching tasks such as stereo matching [74, 78] and 3D scene flow estimation, where some additional depth cues such as stereo camera inputs [54, 11] or point clouds [60] are needed. + +# 3. Method + +Our network is adapted from the two-frame version of ARFlow [37], which uses a light-weight PwCNet [65] as its backbone. The inputs are two consecutive frames $I_1, I_2 \in \mathbb{R}^{H \times W \times 3}$ as well as their semantic segmentation maps $S_1, S_2 \in \{0, 1, \dots, c\}^{H \times W}$ , where the number of classes $c$ is 19 as we use the Cityscapes format [8]. A detailed diagram of the structure of our network can be found in the Appendix. + +# 3.1. Semantic encoder + +We first inject semantic information into the encoder. As shown in Fig. 2(a), shared by each frame $i \in \{1,2\}$ , separate convolutional layers extract features from image $I_{i}$ and semantics $S_{i}$ (one-hot encoded). Features from the two pipelines are concatenated and fed to additional layers. Features at different resolutions $(H / 2^{l}, W / 2^{l})$ form a pyramid $\{f_{i}^{(l)} \mid 2 \leq l \leq 6\}$ . The semantic information in these features helps delineate objects in dark shadows, where appearance is more homogeneous. + +# 3.2. Iterative decoder with a learned upsampler + +Following [22, 37], an iterative residual refinement decoder starts from zero estimate $F_{1\rightarrow 2}^{(7)} = 0$ . For iteration $l \in \{6,5,4,3,2\}$ , the decoder refines feature map $F_{1\rightarrow 2}^{(l + 1)}$ into $F_{1\rightarrow 2}^{(l)}$ based on features $f_{1}^{(l)}, f_{2}^{(l)}$ (Fig. 2(b)). + +More specifically, $F_{1\rightarrow 2}^{(l + 1)}$ is upsampled to match the resolution and used to warp $f_2^{(l)}$ to yield warped feature $\hat{f}_1^{(l)}$ . Correlation volumes are computed between $f_1^{(l)}$ and $\hat{f}_1^{(l)}$ . A one-by-one convolutional layer $C^{(l)}$ compresses the number of channels to a fixed number so that the same layer can be reused across all iterations as proposed by [22]. A flow estimator network predicts a flow residual to be added to the current estimate, and a context network then aggregates flow information spatially and refines the current flow again. + +A learned upsampler network upsamples the final output $F_{1\rightarrow 2}^{(2)}$ in our system. This is different from ARFlow [37], which simply uses four-fold bilinear interpolation, making the final flow boundaries blurry. In contrast, our model learns to sharpen flow boundaries based on the semantic inputs, which have clear boundaries around moving objects. To this end, we add a convex upsampler network similar to the one in RAFT [68]. Different from UPFlow [42], + +![](images/1f578db99d3ccda59acd52464574df776530c318c409b21f7fd6dae0bee7b43c.jpg) +(a) shared pyramidal encoder $(i\in \{1,2\})$ + +![](images/c30e1a5c14116a469523b3f00f526371d8f8a59ad87c073d3cadb07e3d7d3548.jpg) +(b) one iteration (at the $l$ th-level) of the iterative decoder $(l\in \{6,5,4,3,2\})$ + +![](images/4b0eca463840993e1e56b7ce31b8aa97cb087ccf3349b518dc7346f392bc4ee8.jpg) +Figure 2. Network structure. See text in Sec. 3.1 and Sec. 3.2 for explanations. More detailed diagrams are in appendix. +Figure 3. Illustration of semantic augmentation as self-supervision. See text in Sec. 3.3 for details. + +our learned upsampler is only used to upsample from internal estimate $F_{1\rightarrow 2}^{(l)}$ to output flow $U_{1\rightarrow 2}^{(l)}$ , but not when we upsample $F_{1\rightarrow 2}^{(l + 1)}$ at the first step in the decoder iteration. The upsampled $U_{1\rightarrow 2}^{(l)}$ has resolution $(H / 2^{l - 2},W / 2^{l - 2})$ , so $U_{1\rightarrow 2} = U_{1\rightarrow 2}^{(2)}$ (with the original resolution) is the final flow prediction of the network. + +# 3.3. Semantic augmentation as self-supervision + +ARFlow has a very effective in-network augmentation module that samples random transformations $\mathcal{T}_{\theta_1},\mathcal{T}_{\theta_2}$ of the flow prediction $U_{1\rightarrow 2}$ in the first pass of the network and then uses the transformed images $\hat{I}_1 = \mathcal{T}_{\theta_1}(I_1),\hat{I}_2 =$ $\mathcal{T}_{\theta_2}(I_2)$ in a second pass. The prediction $U_{1\rightarrow 2}$ is also transformed accordingly and used to self-supervise the output of the second pass. See [37] for details. + +We retain the augmentation module but make a third pass of the network using semantics-transformed inputs for self-supervision in addition to the ARFlow augmentations of ap + +pearance (e.g., color jitter, random noise) and spatial transformations (e.g., random rotation, random rescaling). The idea behind semantic augmentation is to blend in real object motions across samples. + +To this end (see Fig. 3), we carve objects from other samples based on their semantic maps and paste them as moving foreground objects into the current sample $I_{1}, I_{2}$ , thereby producing augmented images $\tilde{I}_{1}, \tilde{I}_{2}$ . The segmentation maps $S_{1}, S_{2}$ and the first pass output flow $U_{1 \to 2}$ are transformed accordingly, and the transformed $\tilde{U}_{1 \to 2}$ self-supervises the third-pass network outputs $\tilde{U}_{1 \to 2}' = \mathcal{F}(\tilde{I}_{1}, \tilde{I}_{2}, \tilde{S}_{1}, \tilde{S}_{2})$ , where $\mathcal{F}$ is the flow network. + +Similar to occlusion hallucination in SelFlow [39], our semantic augmentation also creates new occlusions and uses the reliable non-occluded flow from the first pass to self-supervise the flow on those newly occluded pixels. However, rather than superpixels filled with noise, we create occlusions with real objects, which are much more realistic. + +Moreover, since we determine the motion of each oc + +cluder, we can use its known optical flow for motion self-supervision. As illustrated in Fig. 3, the occluder motion can be very different from that of the background. This trains the model to assume less smoothness around highly-dynamic objects and to output sharper motion estimates near motion boundaries. These hallucinated but realistic occluders also provide more examples of actively moving objects, which are relatively scarce in the original datasets. + +**Vehicles** We augment vehicle classes (car, truck, bus, train) as they are major sources of errors in previous models (see Tab. 5 for statistics). We group these four classes because they are easily confused with each other in the semantic input. To carve out a single car instance with high probability, we find connected components of vehicle segmentation regions using OpenCV [3] and accept masks whose bounding boxes are 50-300 pixels wide and 50-150 pixels high. We also drop masks that occupy less than $60\%$ area of their bounding boxes to remove largely occluded vehicles. + +Poles We augment pole classes (pole, traffic light, traffic sign) not only because they are common foreground objects that cause occlusions, but also because their motions are usually poorly estimated due to their thin shapes. We use a 200-pixel wide, image-height sliding window to scan the segmentation of these pole classes and find the window with the most pole pixels. If more than $10\%$ pole pixels occur in it, that object mask is stored. This produces groups of nearby poles rather than just individual poles, which may not provide enough augmentation by themselves. + +Sky We also augment the sky through prior knowledge that flow in the sky should be very small. We shrink the sky flow estimate by half before using it for self-supervision. As shown in Fig. 3, the middle tree part in $U_{1\rightarrow 2}$ is very blurry, with a large part of sky mixed in there. After shrinking sky flow by half, the tree in $\tilde{U}_{1\rightarrow 2}$ now has sharper flow around the boundary between tree and sky. Using this as self-supervision helps the model sharpen the motion boundaries between sky and other regions. + +Another way to apply this prior knowledge is to postprocess, i.e., shrink the final output flow in the sky region, instead of using the shrunk flow to self-supervise. However, the latter approach is preferable because it allows the network to balance prior knowledge (self-supervision) with the current sample observation (photometric loss). Our experiments do show cases where some small objects (such as some overhead electric power lines) are misclassified as part of the sky in the semantic input. In these cases, the network learns to rely less on prior knowledge. + +# 3.4.Occluder cache + +During training, we maintain a cache of occluding objects for semantic augmentation. We sample objects from the cache to augment the current batch and then push any new objects found in the batch into the cache, with random replacement if the cache is full. + +Specifically, after we load a sample for training, we first use the current model to infer its optical flow and then search within the sample for any cars (or poles) that meet the standards defined in Sec. 3.3 as "cutouts". For each cutout, we first find the flow vectors on all pixels that belong to that cutout and then compute the mean flow for later reference. We store the current sample, the cutout mask, and the mean flow together in the cache. + +Later on, when we randomly select a cutout to augment another sample, we also retrieve the stored mean flow of this cutout, which is then augmented by random rescaling (by 0.8-1.5 times) and reversing (with $50\%$ probability). We use the augmented mean flow to translate the entire cutout object as an occluder to generate a new sample. For realism, occluders are pasted to the same location they occupied in their image of origin. Holding occluders in a cache provides a random mixture of samples across multiple batches and makes semantic augmentation very efficient. + +# 3.5. Loss functions + +Photometric loss In the first pass, forward and backward flow $U_{1\rightarrow 2}^{(l)}, U_{2\rightarrow 1}^{(l)}$ at each level are predicted and occlusion masks $O_{1\rightarrow 2}^{(l)}, O_{2\rightarrow 1}^{(l)}$ are computed by forward-backward consistency [45]. Frames are warped by + +$$ +I _ {i} ^ {\prime (l)} (\pmb {p}) = I _ {j} ^ {(l)} (\pmb {p} + U _ {i \rightarrow j} ^ {(l)} (\pmb {p})), \quad (i, j \in \{1, 2 \}) +$$ + +where $I_{i}^{(l)}$ is $I_{i}$ down-sampled to the $l$ -th scale, and $\pmb{p}$ denotes pixel coordinates at that scale. Following [37], three measures, namely $L_{1}$ -distance $(\rho_{1})$ , structural similarity (SSIM) [73] $(\rho_{2})$ , and census loss [45] $(\rho_{3})$ are linearly combined to measure photometric differences between $I_{i}^{(l)}$ and $I_{i}^{\prime(l)}$ . Occlusion regions are masked out and both forward and backward directions are taken into account at each level. The final photometric loss is + +$$ +\ell_ {\mathrm {p h}} = \frac {1}{2} \sum_ {(i, j) \in \{(1, 2), (2, 1) \}} \sum_ {l = 2} ^ {6} a _ {l} \sum_ {k = 1} ^ {3} c _ {k} \rho_ {k} \left(I _ {i} ^ {(l)}, I _ {i} ^ {\prime (l)}, O _ {i \rightarrow j} ^ {(l)}\right). \tag {1} +$$ + +Smoothness loss Unlike most previous work, we do not include a smoothness loss because we find it to conflict with our learned upsampler. See ablation study in Sec. 4.5. + +Augmentation loss As in ARFlow [37], our second forward pass computes the $\mathrm{L}_1$ -distance between the trans- + +
MethodTrainTest
2012 EPE2015 EPE2012 Fl-nocEPEFl-all2015 Fl-nocFl-bgFl-fg
supervisedPWC-Net+ [66]-(1.50)3.361.47.724.917.697.88
IRR-PWC [22]-(1.63)3.211.67.654.867.687.52
RAFT [68]-(0.63)--5.103.074.746.87
Separable Flow [85]-(0.69)--4.532.784.255.92
unsupervisedSelFlow [39]1.694.844.312.214.199.6512.6821.74
SimFlow [24]-5.19--13.388.2112.6017.27
ARFlow [37]1.442.85-1.811.80---
UFlow [27]1.682.714.261.911.138.419.7817.87
UPFlow [42]1.272.45-1.49.38---
Ours (baseline)1.392.614.301.79.896.988.8215.21
Ours (+enc)†1.292.423.971.58.993.978.1913.01
Ours (+enc +aug)†1.282.183.901.58.383.907.4812.91
+ +Table 1. KITTI benchmark results (EPE/px and Fl/%). Metrics evaluated at 'all' (all pixels, default for EPE), 'noc' (non-occlusions), 'bg' (background), and 'fg' (foreground). Key metrics (used in official ranking) are underlined. Our 'baseline' is an adapted ARFlow with added learned upsampler and no smoothness loss. +enc means adding semantic encoder; +aug means adding semantic augmentation. () means evaluation data used in training. - means unavailable. $\dagger$ denotes models with semantic inputs. For all metrics, lower is better. + +formed flow $\hat{U}_{1\rightarrow 2}$ and the second pass output $\hat{U}_{1\rightarrow 2}^{\prime} = \mathcal{F}(\hat{I}_1,\hat{I}_2,\hat{S}_1,\hat{S}_2)$ for each pixel $\pmb{p}$ as + +$$ +\hat {D} (\boldsymbol {p}) = \left\| \hat {U} _ {1 \rightarrow 2} (\boldsymbol {p}) - \hat {U} _ {1 \rightarrow 2} ^ {\prime} (\boldsymbol {p}) \right\| _ {1}. +$$ + +This loss is then averaged over the transformed non-occluded region where self-supervision $\hat{U}_{1\rightarrow 2}$ is accurate + +$$ +\ell_ {\mathrm {a r}} = \frac {\sum_ {\boldsymbol {p}} (1 - \hat {O} _ {1 \rightarrow 2} (\boldsymbol {p})) \hat {D} (\boldsymbol {p})}{\sum_ {\boldsymbol {p}} (1 - \hat {O} _ {1 \rightarrow 2} (\boldsymbol {p}))}. \tag {2} +$$ + +For semantic augmentation (third pass), since we are also doing self-supervision with a new generated sample, we use a similar loss definition + +$$ +\ell_ {\mathrm {a u g}} = \frac {\sum_ {\boldsymbol {p}} \left(1 - \tilde {O} _ {1 \rightarrow 2} (\boldsymbol {p})\right) \tilde {D} (\boldsymbol {p})}{\sum_ {\boldsymbol {p}} \left(1 - \tilde {O} _ {1 \rightarrow 2} (\boldsymbol {p})\right)}, \tag {3} +$$ + +where $\tilde{D} (\pmb {p})$ is the distance between the semanticaugmented flow $\tilde{U}_{1\rightarrow 2}$ and the third-pass output $\tilde{U}_{1\rightarrow 2}^{\prime}$ + +$$ +\tilde {D} (\boldsymbol {p}) = \left\| \tilde {U} _ {1 \rightarrow 2} (\boldsymbol {p}) - \tilde {U} _ {1 \rightarrow 2} ^ {\prime} (\boldsymbol {p}) \right\| _ {1}, +$$ + +and the augmented mask $\tilde{O}_{1\rightarrow 2}$ is now computed by + +$$ +\tilde {O} _ {1 \rightarrow 2} = 1 - \max (1 - O _ {1 \rightarrow 2}, M) +$$ + +because we penalize on both the originally non-occluded region $1 - O_{1\rightarrow 2}$ and the pasted foreground object mask $M$ (of which we know the true motion). + +Final loss (with $\lambda = 0.02$ as in ARFlow [37]): + +$$ +\ell = \ell_ {\mathrm {p h}} + \lambda \left(\ell_ {\mathrm {a r}} + \ell_ {\mathrm {a u g}}\right). \tag {4} +$$ + +# 4. Experiments + +# 4.1. Datasets + +We mainly use KITTI [15, 46] and Cityscapes [8] datasets for experiments. Following [37], we train our model first on KITTI raw sequences [46] (55.7k samples) and then fine-tune on KITTI-2015 multi-view extension [46] (11.8k samples). We validate our model using KITTI-2015 train [46] (200 samples) and KITTI-2012 train set [15] (194 samples) since they are the only sets that have optical flow labels. We also train on Cityscapes [8] sequences (83.3k samples) to test model generalization ability, and we sample every other frame to match its frame rate with KITTI. All semantic segmentation inputs are estimated using an off-the-shelf model [89] with a DeepLabV3Plus [7] backbone, which achieves $83.45\%$ mean IoU on Cityscapes and $72.82\%$ mean IoU on KITTI. + +# 4.2. Implementation details + +We implement the model in PyTorch [48] $^{1}$ . We use the Adam optimizer [32] with $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ and batch size 4. We first train on KITTI raw sequences [46] for 100k iterations with a fixed learning rate 0.0002 and then train on KITTI multi-view extension set [15, 46] for another 100k iterations using OneCycleLR schedule [62] with maximum learning rate 0.0004 and linear annealing. + +For data augmentation, we include random horizontal flipping and swapping of the input frames. We resize the inputs to $256 \times 832$ before feeding into the network. The photometric loss weight for each scale $a_{l}$ ( $2 \leq l \leq 6$ ) in + +
MethodFl-allFl-nocFl-bgFl-fg
JFS [21]17.079.8115.922.92
SOF [57]16.8110.8614.6327.73
MRFlow [75]12.198.8610.1322.52
Bai et al. [1]11.628.758.6126.69
Ours (final)8.383.907.4812.91
+ +Eq. (1) are set as 1, 1, 1, 1, 0. The weights for three photometric distance measures $\rho_{k}$ ( $1 \leq k \leq 3$ ) in Eq. (1) are set as 0.15, 0.85, 0 for the first 50k iterations and 0, 0, 1 afterwards. We start the appearance and spatial augmentation (second pass as in ARFlow [37]) at 50k iterations, and semantic augmentation (third pass) after 150k iterations. + +# 4.3. Benchmark testing results + +As common practice, we evaluate optical flow predictions based on two error measurements, Fl (error rate) and EPE (mean $\mathrm{L}_2$ distance). When computing Fl, the estimate of each pixel is considered correct if the error is smaller than 3 pixels or $5\%$ of the magnitude of ground-truth flow [46]. + +As shown in Tab. 1, our semantic modules (both semantic encoder and semantic augmentation) improve performance on both KITTI-2012 and KITTI-2015 sets on all metrics. Our final model achieves $8.38\%$ F1-all error rate on KITTI-2015 test set, which is significantly better than ARFlow [37] $(11.80\%)$ , from which we adapt. We also outperform the current state-of-the-art UPFlow [42] $(9.38\%)$ by a clear margin. All these results strongly suggest that adding semantic inputs can help improve unsupervised flow estimation significantly. + +In addition, Tab. 2 shows that our test results significantly outperform all previous semantic optical flow methods, most of which are based on traditional energy minimization. We are the first to apply semantic inputs to recent unsupervised flow networks, so we are able to push the state-of-the-art by a clear margin. + +# 4.4. Qualitative results + +Some qualitative results are shown in Fig. 4. We can see that our flow outputs have very sharp boundaries and very clean object motion, especially around cars and poles. Our model successfully learns to adapt flow estimation to the semantic map input. Moreover, our model is able to handle very challenging samples where the foreground car motion is drastically different from the background due to the challenging samples we create in semantic augmentation. + +Table 2. KITTI-2015 test results (Fl/\%) compared with other semantic optical flow methods. Metrics evaluated at 'all' (all pixels), 'noc' (non-occlusions), 'bg' (background), and 'fg' (foreground). + +
upno smencaugFl-allEPE-allEPE-nocEPE-occ
10.362.902.076.97
19.922.691.896.57
29.832.651.866.40
39.752.611.856.54
49.752.641.856.55
✓10.222.831.976.75
✓✓8.872.611.856.04
✓✓38.262.421.735.79
✓✓✓8.802.481.606.64
✓✓3✓7.792.181.405.64
+ +Table 3. Ablation study on KITTI-2015 train set (EPE/px and F1%). 'up' means the learned upsampler; 'no sm' means turning off smoothness loss; 'enc' means the level at which we merge semantic features and image features in the encoder; 'aug' means semantic augmentation. + +
Options of augFl-allEPE-allEPE-nocEPE-occ
Ours (final)7.792.181.405.64
start from 100k7.922.191.405.41
vehicles only7.942.211.445.78
loss on new occ8.152.271.425.62
+ +Table 4. Ablation study of different semantic augmentation options on KITTI-2015 train (EPE/px and F1/%). See text for explanations. + +# 4.5. Ablation study + +We also do ablation study to show the effectiveness of each of our proposed modules. We can see in Tab. 3 that adding a semantic encoder to the vanilla ARFlow does help, and the optimal number of encoder layer groups added is 3. Also, our learned upsampler improves results significantly once we turn off the smoothness loss. This makes sense because the upsampler learns to adapt to boundaries, which can be distracted by the smoothness loss. Moreover, both our semantic encoder and semantic augmentation module help improve the results further, which suggests the effectiveness of adding semantic inputs. + +Another ablation study is on different options for the semantic augmentation (Tab. 4). We try starting augmentation earlier from 100k iterations, which works slightly worse. There are mainly two reasons: (1) we use the first forward output to self-supervise our augmented output, so if we start early, the model will use poor output to self-supervise the augmented pass; and (2) our semantic augmentation creates very challenging samples with objects moving very differently from the background, so exposing these hard samples to the model too early may make it hard to train. Apart from this, we also try only using vehicles to augment as well as focusing loss on the newly occluded region, which are both inferior to our final version. + +![](images/07c59636c83be6777bb7d26b345ec42e13f6245ab9cf5eed0a798a8bc2146132.jpg) +Image Input + +![](images/cbb5321d0383d9cd403ec9c926150713f4acbaadab3bca47ced4520392966d5e.jpg) +Semantics Input + +![](images/b18ba7a222c78a7ae580df67155847eef12a2edad21db45d5541879811b4cb0a.jpg) +ARFlow +Figure 4. Qualitative results on KITTI test set (sample #7, 20, 38, 112, 183) compared with ARFlow [37] and UPFlow [42]. We only show first frame input for conciseness. Our outputs have much clearer boundaries around objects. For more qualitative examples, see Appendix. + +![](images/81bafe3a42b8dceeb4b7fce071a2fb656aab904f459b663294b6eca2000eff17.jpg) +UPFlow + +![](images/02089204af00bcd73238d308a144146cad31925f78d7c8c61735a06124c2374c.jpg) +Ours + +
roadcarterrainvegetationsidewalkbuildingwallpole
Proportion42.4%17.6%14.0%11.6%6.2%4.1%1.3%1.1%
ARFlow [37]4.5715.799.4215.344.616.0016.3410.75
Ours (final)3.6710.178.7413.473.094.2711.189.44
Relative improvement19.6%35.6%7.3%12.2%33.0%28.7%31.6%12.2%
Reweighed contribution19.0%49.4%4.8%10.9%4.7%3.5%3.4%0.7%
+ +# 4.6. Improvement analysis + +It may be interesting to understand where our improvements come from in terms of semantic classes, so we compute the KITTI-2015 train set error for each class. As shown in Tab. 5, our model improves the flow on every class. After reweighing the absolute improvement of each class based on their proportion in the evaluated pixels, we find that car flow accounts for nearly half of our improvement on Fl-all, indicating the effectiveness of augmenting vehicle objects. + +# 4.7. Generalization ability + +We also investigate the generalization ability of our semantic-aided flow models. We train flow models on Cityscapes and directly test them on KITTI-2015 train set without fine-tuning. Following [50], we crop the bottom $25\%$ of the frame to remove the car logo and resize the frame to size $256 \times 704$ . All other experiment settings are exactly the same as for KITTI. As we can see from Tab. 6, adding semantic inputs significantly helps our unsupervised flow model generalize and adapt better across datasets. + +# 5. Conclusion and future work + +In this paper, we show that adding semantic segmentation inputs can help significantly improve the performance + +Table 5. KITTI-2015 flow error (Fl-all/%) for each semantic class. The first row shows the proportion of each class among evaluated pixels (KITTI flow does not evaluate on every pixel). We only show classes that account for at least $1\%$ of the evaluated pixels. + +
MethodFl-allEPE-allEPE-nocEPE-occ
ARFlow (our impl.)13.214.082.889.40
Ours (baseline)12.273.812.439.91
Ours (+enc)†11.283.332.128.75
Ours (+enc +aug)†10.322.641.567.12
+ +Table 6. Generalization results (train on Cityscapes, and test on KITTI-2015 train). We use our implementation of ARFlow [37] to run this experiment. Our 'baseline' is an adapted ARFlow with added learned upsampler and no smoothness loss. $^+$ enc means adding semantic encoder; $^+$ aug means adding semantic augmentation. $\dagger$ denotes models with semantic inputs. + +of unsupervised optical flow networks on autonomous driving datasets. We propose a novel network model called SemARFlow, a semantic adaptation of ARFlow [37], with a semantic encoder, a learned upsampler, and a semantic augmentation module, where some domain-specific motion prior has been taken into account. Our network better predicts flow at occlusion regions and effectively sharpens flow estimates around object boundaries, even for very challenging samples. The additional semantic inputs also make our network generalize better across datasets. + +One direction for future work is motivated by the example illustrated in Fig. 5. Our model does output sharp flow + +![](images/1163bd8c232875b98653a17d74378498bfe4ef69b40c6cd4a3f95ac58e27a07a.jpg) +Figure 5. An example on KITTI-2015 train set (sample #164). See the discussion in Sec. 5 for details. + +![](images/d3f5dfb78ff75b20c7534879e6a07823a7a3f0014c676ba3b095e4f7f671dee7.jpg) + +![](images/a7bdac7bbcf7508d90aa410a55955595fee7f51f1cf2bf6e69324c96fcd7f87f.jpg) + +boundaries between the cars and the background, thanks to good boundary information from the semantic maps. However, the boundary between these two fast-moving cars is inferred from image information only, because their semantic maps merge into one. As a result, that part of the flow boundary is less accurate. An interesting direction for future work is to use instance-level semantics as input for more detailed object masks. + +Acknowledgments. This research is based upon work supported by the National Science Foundation under Grant No. 1909821. + +# References + +[1] Min Bai, Wenjie Luo, Kaustav Kundu, and Raquel Urtasun. Exploiting semantic information and deep matching for optical flow. In Proceedings of the European Conference on Computer Vision, pages 154-170. Springer, 2016. 2, 3, 7 +[2] Shubhankar Borse, Ying Wang, Yizhe Zhang, and Fatih Porikli. Inverseform: A loss function for structured boundary-aware segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5901-5911, 2021. 3 +[3] Gary Bradski. The opencv library. Dr. Dobb's Journal: Software Tools for the Professional Programmer, 25(11):120-123, 2000. 5 +[4] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Proceedings of the European Conference on Computer Vision, pages 611-625. Springer, 2012. 1 +[5] Yingfeng Cai, Lei Dai, Hai Wang, and Zhixiong Li. Multi-target pan-class intrinsic relevance driven model for improving semantic segmentation in autonomous driving. IEEE Transactions on Image Processing, 30:9069-9084, 2021. 3 +[6] Linda Capito, Umit Ozguner, and Keith Redmill. Optical flow based visual potential field for autonomous driving. In IEEE Intelligent Vehicles Symposium, pages 885-891. IEEE, 2020. 1 +[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European Conference on Computer Vision, pages 801-818, 2018. 6 +[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes + +dataset for semantic urban scene understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3213-3223, 2016. 3, 6 +[9] Mingyu Ding, Zhe Wang, Bolei Zhou, Jianping Shi, Zhiwu Lu, and Ping Luo. Every frame counts: Joint learning of video segmentation and optical flow. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10713-10720, 2020. 2, 3 +[10] Alexey Dosovitskiy, Philipp Fischery, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick van der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE International Conference on Computer Vision, ICCV '15, pages 2758-2766, Washington, DC, USA, 2015. IEEE Computer Society. 1, 2 +[11] Cheng Feng, Long Ma, Congxuan Zhang, Zhen Chen, Liyue Ge, and Shaofeng Jiang. Ss-sf: Piecewise 3d scene flow estimation with semantic segmentation. IEEE Access, 9:22745-22759, 2021. 3 +[12] Jun Fu, Jing Liu, Haijie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3146-3154, 2019. 3 +[13] Aditya Ganeshan, Alexis Vallet, Yasunori Kudo, Shin-ichi Maeda, Tommi Kerola, Rares Ambrus, Dennis Park, and Adrien Gaidon. Warp-refine propagation: Semi-supervised auto-labeling via cycle-consistency. In Proceedings of the IEEE International Conference on Computer Vision, pages 15499–15509, 2021. 3 +[14] Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In Proceedings of the European Conference on Computer Vision, pages 713-729. Springer, 2020. 1 +[15] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research, 32(11):1231-1237, 2013. 6 +[16] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 2 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1 +[18] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial Intelligence, 17(1-3):185–203, 1981. 2 +[19] Zilong Huang, Xinggang Wang, Lichao Huang, Chang Huang, Yunchao Wei, and Wenyu Liu. Ccnet: Criss-cross attention for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 603-612, 2019. 3 +[20] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. Liteflownet: A lightweight convolutional neural network for optical flow estimation. In Proceedings of the IEEE Conference + +on Computer Vision and Pattern Recognition, pages 8981-8989, 2018. 2 +[21] Junhwa Hur and Stefan Roth. Joint optical flow and temporally consistent semantic segmentation. In Proceedings of the European Conference on Computer Vision, pages 163-177. Springer, 2016. 2, 3, 7 +[22] Junhwa Hur and Stefan Roth. Iterative residual refinement for joint optical flow and occlusion estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5754-5763, 2019. 1, 2, 3, 6 +[23] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1647-1655, 2017. 2 +[24] Woobin Im, Tae-Kyun Kim, and Sung-Eui Yoon. Unsupervised learning of optical flow with deep feature similarity. In Proceedings of the European Conference on Computer Vision, pages 172-188. Springer, 2020. 2, 6 +[25] Joel Janai, Fatma Guney, Anurag Ranjan, Michael Black, and Andreas Geiger. Unsupervised learning of multi-frame optical flow with occlusions. In Proceedings of the European Conference on Computer Vision, pages 690-706, 2018. 2 +[26] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In Proceedings of the IEEE International Conference on Computer Vision, pages 9772-9781, 2021. 2 +[27] Rico Jonschkowski, Austin Stone, Jonathan T Barron, Ariel Gordon, Kurt Konolige, and Anelia Angelova. What matters in unsupervised optical flow. In Proceedings of the European Conference on Computer Vision, pages 557-572. Springer, 2020. 1, 6 +[28] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM computing surveys, 54(10s):1-41, 2022. 1 +[29] Hannah H Kim, Celia Cintas, Girmaw Abebe Tadesse, and Skyler Speakman. Spatially constrained adversarial attack detection and localization in the representation space of optical flow networks. In International Joint Conference on Artificial Intelligence, 2023. 1 +[30] Hannah Halin Kim, Shuzhi Yu, and Carlo Tomasi. Joint detection of motion boundaries and occlusions. In Proceedings of the British Machine Vision Conference, 2021. 1 +[31] Hannah Halin Kim, Shuzhi Yu, Shuai Yuan, and Carlo Tomasi. Cross-attention transformer for video interpolation. In Proceedings of the Asian Conference on Computer Vision Workshops, pages 320–337, 2022. 1 +[32] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. Proceedings of the International Conference on Learning Representations, 2014. 6 +[33] Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weissenborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Sylvain Gelly, Thomas Unterthiner, and Xiaohua Zhai. An image is worth 16x16 words: Transformers for image recogni + +tion at scale. In Proceedings of the International Conference on Learning Representations, 2021. 1 +[34] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 1 +[35] Xiangtai Li, Jiangang Bai, Kuiyuan Yang, and Yunhai Tong. Flow2seg: Motion-aided semantic segmentation. In Proceedings of the International Conference on Artificial Neural Networks, pages 225–237. Springer, 2019. 3 +[36] Zewen Li, Fan Liu, Wenjie Yang, Shouheng Peng, and Jun Zhou. A survey of convolutional neural networks: analysis, applications, and prospects. IEEE transactions on neural networks and learning systems, 2021. 1 +[37] Liang Liu, Jiangning Zhang, Ruifei He, Yong Liu, Yabiao Wang, Ying Tai, Donghao Luo, Chengjie Wang, Jilin Li, and Feiyue Huang. Learning by analogy: Reliable supervision from transformations for unsupervised optical flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6489-6498, 2020. 1, 2, 3, 4, 5, 6, 7, 8 +[38] Pengpeng Liu, Irwin King, Michael R Lyu, and Jia Xu. Ddflow: Learning optical flow with unlabeled data distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8770-8777, 2019. 1, 2 +[39] Pengpeng Liu, Michael Lyu, Irwin King, and Jia Xu. Selfflow: Self-supervised learning of optical flow. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4571-4580, 2019. 1, 2, 4, 6 +[40] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3431-3440, 2015. 2, 3 +[41] Bruce D Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In International Joint Conference on Artificial Intelligence, volume 2, pages 674-679, 1981. 2 +[42] Kunming Luo, Chuan Wang, Shuaicheng Liu, Haoqiang Fan, Jue Wang, and Jian Sun. Upflow: Upsampling pyramid for unsupervised optical flow learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1045-1054, 2021. 1, 2, 3, 6, 7, 8 +[43] Rémi Marsal, Florian Chabot, Angélique Loesch, and Hichem Sahbi. Brightflow: Brightness-change-aware unsupervised learning of optical flow. In Proceedings of the IEEE winter conference on applications of computer vision, pages 2061-2070, 2023. 1 +[44] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4040-4048, 2016. 1, 2 +[45] Simon Meister, Junhwa Hur, and Stefan Roth. Unflow: Unsupervised learning of optical flow with a bidirectional census loss. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 1, 2, 5 +[46] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE Conference + +on Computer Vision and Pattern Recognition, 2015. 2, 3, 6, 7 +[47] Hyeonwoo Noh, Seunghoon Hong, and Bohyung Han. Learning deconvolution network for semantic segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1520-1528, 2015. 2, 3 +[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, pages 8024-8035. Curran Associates, Inc., 2019. 6 +[49] Anurag Ranjan and Michael J. Black. Optical flow estimation using a spatial pyramid network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2720-2729, 2017. 2 +[50] Anurag Ranjan, Varun Jampani, Lukas Balles, Kihwan Kim, Deqing Sun, Jonas Wulff, and Michael J Black. Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12240–12249, 2019. 8 +[51] Hazem Rashed, Senthil Yogamani, Ahmad El-Sallab, Pavel Krizek, and Mohamed El-Helw. Optical flow augmented semantic segmentation networks for automated driving. In International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, pages 165-172, 2019. 3 +[52] Fitsum A Reda, Guilin Liu, Kevin J Shih, Robert Kirby, Jon Barker, David Tarjan, Andrew Tao, and Bryan Catanzaro. Sdc-net: Video prediction using spatially-displaced convolution. In Proceedings of the European Conference on Computer Vision, pages 718-733, 2018. 3 +[53] Zhile Ren, Orazio Gallo, Deqing Sun, Ming-Hsuan Yang, Erik B Sudderth, and Jan Kautz. A fusion approach for multiframe optical flow estimation. In Proceedings of the IEEE winter conference on applications of computer vision, pages 2077-2086. IEEE, 2019. 2 +[54] Zhile Ren, Deqing Sun, Jan Kautz, and Erik Sudderth. Cascaded scene flow prediction using semantic segmentation. In Proceedings of the International Conference on 3D Vision, pages 225-233. IEEE, 2017. 3 +[55] Zhe Ren, Junchi Yan, Bingbing Ni, Bin Liu, Xiaokang Yang, and Hongyuan Zha. Unsupervised deep learning for optical flow estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2017. 2 +[56] Stephan R. Richter, Zeeshan Hayder, and Vladlen Koltun. Playing for benchmarks. In Proceedings of the IEEE International Conference on Computer Vision, pages 2232-2241, 2017. 1 +[57] Laura Sevilla-Lara, Deqing Sun, Varun Jampani, and Michael J Black. Optical flow with semantic segmentation and localized layers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3889-3898, 2016. 2, 3, 7 + +[58] Shihao Shen, Louis Kerofsky, and Senthil Yogamani. Optical flow for autonomous driving: Applications, challenges and improvements. arXiv preprint arXiv:2301.04422, 2023. 1 +[59] Jianbo Shi and Carlo Tomasi. Good features to track. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 593-600, 1994. 1 +[60] Yukang Shi and Kaiheng Ma. Safit: Segmentation-aware scene flow with improved transformer. In Proceedings of the International Conference on Robotics and Automation, pages 10648-10655. IEEE, 2022. 3 +[61] Jeongho Shin, Sangjin Kim, Sangkyu Kang, Seong-Won Lee, Joonki Paik, Besma Abidi, and Mongi Abidi. Optical flow-based real-time object tracking using non-prior training active feature model. Real-time imaging, 11(3):204-218, 2005. 1 +[62] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, volume 11006, pages 369-386. SPIE, 2019. 6 +[63] Austin Stone, Daniel Maurer, Alper Ayvaci, Anelia Angelova, and Rico Jonschkowski. Smurf: Self-teaching multiframe unsupervised raft with full-image warping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3887-3896, 2021. 3 +[64] Xiuchao Sui, Shaohua Li, Xue Geng, Yan Wu, Xinxing Xu, Yong Liu, Rick Goh, and Hongyuan Zhu. Craft: Cross-attentional flow transformer for robust optical flow. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 17602-17611, 2022. 2 +[65] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8934-8943, 2018. 1, 2, 3 +[66] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Models matter, so does training: An empirical study of cnns for optical flow estimation. IEEE transactions on pattern analysis and machine intelligence, 42(6):1408-1423, 2019. 6 +[67] Andrew Tao, Karan Sapra, and Bryan Catanzaro. Hierarchical multi-scale attention for semantic segmentation. arXiv preprint arXiv:2005.10821, 2020. 3 +[68] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proceedings of the European Conference on Computer Vision, pages 402-419. Springer, 2020. 1, 2, 3, 6 +[69] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Proceedings of the International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1 +[70] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE Con + +ference on Computer Vision and Pattern Recognition Workshops, pages 969-977, 2018. 1 +[71] Panqu Wang, Pengfei Chen, Ye Yuan, Ding Liu, Zehua Huang, Xiaodi Hou, and Garrison Cottrell. Understanding convolution for semantic segmentation. In Proceedings of the IEEE winter conference on applications of computer vision, pages 1451-1460. IEEE, 2018. 3 +[72] Yang Wang, Yi Yang, Zhenheng Yang, Liang Zhao, Peng Wang, and Wei Xu. Occlusion aware unsupervised learning of optical flow. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4884-4893, 2018. 1, 2 +[73] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5 +[74] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Song Wang, and Lili Ju. Semantic stereo matching with pyramid cost volumes. In Proceedings of the IEEE International Conference on Computer Vision, pages 7484-7493, 2019. 3 +[75] Jonas Wulff, Laura Sevilla-Lara, and Michael J Black. Optical flow in mostly rigid scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4671-4680, 2017. 2, 3, 7 +[76] Haofei Xu, Jiaolong Yang, Jianfei Cai, Juyong Zhang, and Xin Tong. High-resolution optical flow from 1d attention and correlation. In Proceedings of the IEEE International Conference on Computer Vision, pages 10498-10507, 2021. 2 +[77] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8121-8130, 2022. 2 +[78] Guorun Yang, Hengshuang Zhao, Jianping Shi, Zhidong Deng, and Jiaya Jia. Segstereo: Exploiting semantic information for disparity estimation. In Proceedings of the European Conference on Computer Vision, pages 636-651, 2018. 3 +[79] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. Bisenet v2: Bilateral network with guided aggregation for real-time semantic segmentation. International Journal of Computer Vision, 129:3051-3068, 2021. 3 +[80] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceedings of the European Conference on Computer Vision, pages 325-341, 2018. 3 +[81] Jason J. Yu, Adam W. Harley, and Konstantinos G. Derpanis. Back to basics: Unsupervised learning of optical flow via brightness constancy and motion smoothness. In Gang Hua and Hervé Jégou, editors, Proceedings of the European Conference on Computer Vision Workshops, pages 3-10, Cham, 2016. Springer International Publishing. 1, 2 +[82] Shuzhi Yu, Hannah H Kim, Shuai Yuan, and Carlo Tomasi. Unsupervised flow refinement near motion boundaries. In + +Proceedings of the British Machine Vision Conference. BMVA Press, 2022. 1 +[83] Shuai Yuan, Xian Sun, Hannah Kim, Shuzhi Yu, and Carlo Tomasi. Optical flow training under limited label budget via active learning. In Proceedings of the European Conference on Computer Vision, pages 410-427, 2022. 2 +[84] Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Common practices and emerging technologies. IEEE access, 8:58443-58469, 2020. 1 +[85] Feihu Zhang, Oliver J Woodford, Victor Adrian Prisacariu, and Philip HS Torr. Separable flow: Learning motion cost volumes for optical flow estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 10807-10817, 2021. 1, 2, 6 +[86] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2881-2890, 2017. 3 +[87] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6881-6890, 2021. 2, 3 +[88] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In Proceedings of the International Conference on Learning Representations, 2021. 1 +[89] Yi Zhu, Karan Sapra, Fitsum A Reda, Kevin J Shih, Shawn Newsam, Andrew Tao, and Bryan Catanzaro. Improving semantic segmentation via video propagation and label relaxation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8856-8865, 2019. 2, 3, 6 \ No newline at end of file diff --git a/2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/images.zip b/2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5cfd934d6230ea01b6906fa32b07c7919f21b8b3 --- /dev/null +++ b/2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ebb38f954c58a16387c5b5fae12db4f4d8f8b030471693e8c4b95bdb3c8e6628 +size 551527 diff --git a/2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/layout.json b/2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..77d451123cf7b72d1aead5743828c6ebd19dc969 --- /dev/null +++ b/2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/layout.json @@ -0,0 +1,10867 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 536, + 140 + ], + "type": "text", + "content": "SemARFlow: Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 161, + 433, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 161, + 433, + 190 + ], + "spans": [ + { + "bbox": [ + 160, + 161, + 433, + 190 + ], + "type": "text", + "content": "Shuai Yuan, Shuzhi Yu, Hannah Kim, and Carlo Tomasi \nDuke University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 175, + 191, + 418, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 191, + 418, + 202 + ], + "spans": [ + { + "bbox": [ + 175, + 191, + 418, + 202 + ], + "type": "text", + "content": "{shuai, shuzhiyu, hannah, tomasi}@cs.duke.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 257, + 290, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 290, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 290, + 485 + ], + "type": "text", + "content": "Unsupervised optical flow estimation is especially hard near occlusions and motion boundaries and in low-texture regions. We show that additional information such as semantics and domain knowledge can help better constrain this problem. We introduce SemARFlow, an unsupervised optical flow network designed for autonomous driving data that takes estimated semantic segmentation masks as additional inputs. This additional information is injected into the encoder and into a learned upsampler that refines the flow output. In addition, a simple yet effective semantic augmentation module provides self-supervision when learning flow and its boundaries for vehicles, poles, and sky. Together, these injections of semantic information improve the KITTI-2015 optical flow test error rate from " + }, + { + "bbox": [ + 46, + 257, + 290, + 485 + ], + "type": "inline_equation", + "content": "11.80\\%" + }, + { + "bbox": [ + 46, + 257, + 290, + 485 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 257, + 290, + 485 + ], + "type": "inline_equation", + "content": "8.38\\%" + }, + { + "bbox": [ + 46, + 257, + 290, + 485 + ], + "type": "text", + "content": ". We also show visible improvements around object boundaries as well as a greater ability to generalize across datasets. Code is available at https://github.com/duke-vision/semantic-unsup-flow-release." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 510, + 128, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 128, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 128, + 522 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 532, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 581 + ], + "type": "text", + "content": "Optical flow estimation, i.e., pixel-level motion tracking across video frames, has broad applications in many computer vision tasks that include object tracking [61], video editing [14, 31], and autonomous driving [6, 58]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 581, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 700 + ], + "type": "text", + "content": "Thanks to the success of deep convolutional neural networks [34, 17, 36] and transformer networks [33, 69, 88, 28] in computer vision, many top-performing supervised optical flow networks have been proposed in recent years [10, 65, 22, 68, 85], in which ground-truth labels supervise training. However, real optical flow is hard to label, so most supervised methods train (or at least pre-train) on synthetic datasets [10, 44, 4, 56] which makes them hard to adapt to real applications due to the significant gap between synthetic and real data [70, 29]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": "Due to label scarcity, unsupervised training of optical" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 230, + 411, + 329 + ], + "blocks": [ + { + "bbox": [ + 307, + 230, + 411, + 329 + ], + "lines": [ + { + "bbox": [ + 307, + 230, + 411, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 411, + 329 + ], + "type": "image", + "image_path": "9e65a6cce7dfec507031a1407f7e49721073b9180ae490256c7b968ba3697ddc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 332, + 547, + 388 + ], + "lines": [ + { + "bbox": [ + 305, + 332, + 547, + 388 + ], + "spans": [ + { + "bbox": [ + 305, + 332, + 547, + 388 + ], + "type": "text", + "content": "Figure 1. An example on KITTI-2015 test set (sample #63). Only one input frame is shown for conciseness. Our SemARFlow takes additional semantic segmentation inputs (estimated by an off-the-shelf model) and outputs much sharper flow around semantic boundaries." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 413, + 230, + 445, + 328 + ], + "blocks": [ + { + "bbox": [ + 413, + 230, + 445, + 328 + ], + "lines": [ + { + "bbox": [ + 413, + 230, + 445, + 328 + ], + "spans": [ + { + "bbox": [ + 413, + 230, + 445, + 328 + ], + "type": "image", + "image_path": "343f36ac10bb396cab4abe94b28fbb130cf086e612f0ae4d6d44fd01bbc59452.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 446, + 231, + 477, + 328 + ], + "blocks": [ + { + "bbox": [ + 446, + 231, + 477, + 328 + ], + "lines": [ + { + "bbox": [ + 446, + 231, + 477, + 328 + ], + "spans": [ + { + "bbox": [ + 446, + 231, + 477, + 328 + ], + "type": "image", + "image_path": "81f42782b8b7e34d6a081d106bfccb2d41438995034e911df052f59146cd50c3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 478, + 231, + 511, + 328 + ], + "blocks": [ + { + "bbox": [ + 478, + 231, + 511, + 328 + ], + "lines": [ + { + "bbox": [ + 478, + 231, + 511, + 328 + ], + "spans": [ + { + "bbox": [ + 478, + 231, + 511, + 328 + ], + "type": "image", + "image_path": "85dc6b3c2243ce70f0fa7c131cc53de504af6f1b5e13f883b0520d0b57d3ac48.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 511, + 231, + 544, + 328 + ], + "blocks": [ + { + "bbox": [ + 511, + 231, + 544, + 328 + ], + "lines": [ + { + "bbox": [ + 511, + 231, + 544, + 328 + ], + "spans": [ + { + "bbox": [ + 511, + 231, + 544, + 328 + ], + "type": "image", + "image_path": "b36e7fc92d3acc591e035ba081195e897be82f0c0d8276200009904df42750ff.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 419, + 547, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 547, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 547, + 492 + ], + "type": "text", + "content": "flow estimators [81] instead uses loss terms based on the assumptions of constant brightness and smooth flow [45, 27, 42]. Some self-supervision techniques have also been studied to enhance model performance [38, 39, 37]. Unsupervised training makes it possible to train flow networks directly on large real datasets from the target domain." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 494, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 547, + 602 + ], + "type": "text", + "content": "Even so, unsupervised optical flow estimation is a poorly constrained problem. Brightness is not constant in regions with occlusions [72] or on shiny surfaces [43], nor is it smooth across motion boundaries [30, 82]. Moreover, points in regions with poor texture content [59] or in dark shadows [58] are difficult to track as they are easily confused with neighboring points (the so-called aperture problem). This dearth of reliable constraints makes the unsupervised training of flow networks especially challenging." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 603, + 547, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 603, + 547, + 699 + ], + "spans": [ + { + "bbox": [ + 304, + 603, + 547, + 699 + ], + "type": "text", + "content": "A natural way to address this issue is to inject additional constraints in the form of semantics and domain knowledge. For example, in autonomous driving applications [84, 6], we have clear expectations about object types and layouts of the scene, as well as prior knowledge of how each type of object typically moves. We focus on this application domain and show that this additional information helps the network achieve better flow results." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "To make an autonomous driving system work well in" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9566" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "reality, it is best to train it on real data. However, annotating real driving video with optical flow labels is expensive [83], as it requires careful synchronization and calibration of diverse sensors including cameras, LiDAR, and GPS/IMU [16], aided by some manual annotation and curation based on CAD models of moving objects [46]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 146, + 288, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 146, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 288, + 266 + ], + "type": "text", + "content": "In contrast, annotating semantic labels seems much more feasible, and indeed semantic labels are available in most (if not all) existing driving datasets. We consider semantic segmentation because it provides semantics at the pixel level, the same level as optical flow. As one of the most popular and well-studied tasks in modern computer vision, semantic segmentation [40, 47, 87] has been extensively adopted for autonomous driving systems. In this paper, we show that adding semantic segmentation inputs helps improve unsupervised optical flow performance significantly." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 267, + 288, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 288, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 288, + 447 + ], + "type": "text", + "content": "Specifically, we first infer semantic segmentation maps using an off-the-shelf model [89], which of course is trained with semantic labels. An encoder with semantic map input is used to aggregate image and semantic features (Sec. 3.1), and a learned upsampler is added into the iterative decoder to refine flow around object boundaries given semantic inputs (Sec. 3.2). We also propose a simple yet effective semantic augmentation module for self-supervision, which provides realistic augmentations specific to the vehicles, poles, and sky classes based on domain knowledge and segmentation maps (Sec. 3.3). An occluder cache is implemented to improve efficiency (Sec. 3.4). Semantic augmentation provides challenging samples for self-supervision, which help train flow better in occluded regions and around foreground object boundaries." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 449, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 605 + ], + "type": "text", + "content": "Overall, by injecting semantic segmentation inputs, our SemARFlow network achieves significantly better unsupervised flow results both quantitatively (Sec. 4.3) and qualitatively (Sec. 4.4). Adapted from ARFlow [37], SemARFlow reduces KITTI-2015 [46] test error from " + }, + { + "bbox": [ + 46, + 449, + 287, + 605 + ], + "type": "inline_equation", + "content": "11.80\\%" + }, + { + "bbox": [ + 46, + 449, + 287, + 605 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 449, + 287, + 605 + ], + "type": "inline_equation", + "content": "8.38\\%" + }, + { + "bbox": [ + 46, + 449, + 287, + 605 + ], + "type": "text", + "content": " and out-performs the current unsupervised state-of-the-art UPFlow [42] " + }, + { + "bbox": [ + 46, + 449, + 287, + 605 + ], + "type": "inline_equation", + "content": "(9.38\\%)" + }, + { + "bbox": [ + 46, + 449, + 287, + 605 + ], + "type": "text", + "content": " by a clear margin. The example in Fig. 1 also demonstrates visible improvements around object boundaries. The effectiveness of each module is justified through an extensive ablation study (Sec. 4.5) and improvement analysis (Sec. 4.6). In addition to performance boost, unsupervised flow networks with additional semantic inputs generalize better across different datasets (Sec. 4.7)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Our research is essentially novel compared to previous approaches. Some early work incorporates semantics in traditional energy-minimization methods [41, 18] for optical flow through geometric constraints such as piece-wise rigid motion and planar surface motion [57, 21, 1, 75]. In comparison, to the best of our knowledge, we are the first to inject semantics into the unsupervised training of recent optical flow networks. Some research also trains a network for segmentation and optical flow jointly [9]. In contrast, we" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": "leverage existing, separately trained segmentation systems both because they are available and because modularity—separating segmentation from flow estimation—is important for the development of large real-application systems such as autonomous driving." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 133, + 501, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 133, + 501, + 144 + ], + "spans": [ + { + "bbox": [ + 317, + 133, + 501, + 144 + ], + "type": "text", + "content": "In summary, our contributions are as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 151, + 545, + 286 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 317, + 151, + 545, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 545, + 187 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 545, + 187 + ], + "type": "text", + "content": "- To the best of our knowledge, we are the first to explore adding semantic inputs to assist the unsupervised training of deep optical flow networks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 194, + 545, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 194, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 317, + 194, + 545, + 255 + ], + "type": "text", + "content": "- We propose a simple yet effective network called SemARFlow that achieves state-of-the-art results both quantitatively and qualitatively. Our model works well on real-life occlusions and yields sharp motion boundaries around objects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 262, + 545, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 262, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 317, + 262, + 545, + 286 + ], + "type": "text", + "content": "- We provide full training and inference code as well as trained models to encourage follow-up research." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 297, + 390, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 297, + 390, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 390, + 308 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 316, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 460 + ], + "type": "text", + "content": "Supervised optical flow Since the introduction of the pioneering deep optical flow network FlowNet [10], more and more top-performing CNN-based flow estimators have been proposed over the years [23, 49, 65, 20, 22, 68]. Recently, vision transformer networks and attention mechanism have also been applied to this problem, and these have achieved state-of-the-art performance on benchmark datasets [76, 26, 85, 64, 77]. The supervised methods are often pre-trained on large synthetic datasets such as FlyingChairs [10] and FlyingThings3D [44] before fine-tuning on the target dataset. However, there is a clear gap between the artificially generated data and real scenarios." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": "Unsupervised optical flow Due to the lack of ground-truth labels, unsupervised optical flow estimation uses surrogate losses such as photometric loss and smoothness loss to supervise training [81, 55]. To tackle the issues around occlusion regions, various methods have been proposed including estimated occlusion masks [72], forward-backward consistency [45], and multi-frame fusion [25, 53]. To better upsample the flow in the decoder, UPFlow [42] additionally predicts a confidence map and an interpolation flow to guide flow refinement and has become the state-of-the-art unsupervised flow method." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": "Some latest research has also explored the use of self-supervision to enhance flow prediction. Early methods apply the knowledge distillation technique to train a two-stage teacher-student network [38, 39]. ARFlow [37] further improves this idea by generating reliable self-supervision signals from data transformations, while merging the two training stages into single-stage training with one added loss term. SimFlow [24] replaces the handcrafted features with deep self-supervised features to measure similarity in the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9567" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "unsupervised losses. SMURF [63] utilizes a RAFT-like structure [68] and applies multi-frame self-supervised training with many technical improvements. Our SemARFlow also uses self-supervision but with the guidance of semantic segmentation, which is much more realistic than guidance without semantics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 159, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 159, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 287, + 361 + ], + "type": "text", + "content": "Semantic segmentation Semantic segmentation classifies each pixel of the given image into semantic objects. Fully Convolutional Network (FCN) [40] is one of the early CNN-based segmentation methods. It takes inputs of arbitrary sizes and outputs dense pixel-level predictions, becoming one of the popular backbone architectures for follow-up work. Deconvolution network [47] is also proposed to better recover low-level details of the prediction. One main challenge for these systems is that they lack global scene information. Subsequent work addresses this issue by enlarging the receptive field of the network with global pyramid pooling layers as in PSPNet [86], hybrid dilated convolutions [71], and a fast-down-sampling strategy [80, 79]. Attention modules have also shown to help in semantic segmentation by capturing full-image dependencies of all pixels [19, 87] or the semantic inter-dependencies across spatial and channel dimensions [12]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 363, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 471 + ], + "type": "text", + "content": "There also exists extensive work on semantic segmentation in the context of autonomous driving [2, 67, 13, 5, 89], thanks to the publication of large-scale driving datasets [8, 46]. To better train the network on a coarsely labeled dataset like KITTI [46], Zhu et al. use a video prediction network SDCNet [52] to synthesize new training samples with relaxed label propagation [89]. Due to their good performance on KITTI [46], we utilize their network models to infer semantic inputs for all our experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 486, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 689 + ], + "type": "text", + "content": "Combining semantics and optical flow Though there has been much progress on both semantic segmentation and optical flow estimation, the semantic optical flow problem (how to exploit semantics to help optical flow estimation) has received limited attention in recent years, and the current best results are thus much outdated. Some early methods incorporate semantics through geometric constraints to refine flow on various semantic regions, such as planar regions (using homographies) [57], static regions (using rigid camera motion and epipolar constraints) [21, 75], and rigid objects (estimating rigid motion for each object instance) [1]. However, most methods are traditional flow methods based on energy minimization, where an initial flow estimate is usually needed and semantics is mostly used for refinement. In comparison, we explore adapting latest unsupervised optical flow networks to leverage semantic inputs in one single stage of estimation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Apart from using semantics to help flow, some research has also explored using optical flow to help seman" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "tic segmentation [51, 35], or to train both tasks jointly [9]. There are also studies on exploiting semantics on some other correspondence matching tasks such as stereo matching [74, 78] and 3D scene flow estimation, where some additional depth cues such as stereo camera inputs [54, 11] or point clouds [60] are needed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 156, + 361, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 156, + 361, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 361, + 167 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 176, + 545, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 176, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 304, + 176, + 545, + 272 + ], + "type": "text", + "content": "Our network is adapted from the two-frame version of ARFlow [37], which uses a light-weight PwCNet [65] as its backbone. The inputs are two consecutive frames " + }, + { + "bbox": [ + 304, + 176, + 545, + 272 + ], + "type": "inline_equation", + "content": "I_1, I_2 \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 176, + 545, + 272 + ], + "type": "text", + "content": " as well as their semantic segmentation maps " + }, + { + "bbox": [ + 304, + 176, + 545, + 272 + ], + "type": "inline_equation", + "content": "S_1, S_2 \\in \\{0, 1, \\dots, c\\}^{H \\times W}" + }, + { + "bbox": [ + 304, + 176, + 545, + 272 + ], + "type": "text", + "content": ", where the number of classes " + }, + { + "bbox": [ + 304, + 176, + 545, + 272 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 176, + 545, + 272 + ], + "type": "text", + "content": " is 19 as we use the Cityscapes format [8]. A detailed diagram of the structure of our network can be found in the Appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 281, + 411, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 281, + 411, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 411, + 293 + ], + "type": "text", + "content": "3.1. Semantic encoder" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 299, + 545, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 545, + 409 + ], + "type": "text", + "content": "We first inject semantic information into the encoder. As shown in Fig. 2(a), shared by each frame " + }, + { + "bbox": [ + 304, + 299, + 545, + 409 + ], + "type": "inline_equation", + "content": "i \\in \\{1,2\\}" + }, + { + "bbox": [ + 304, + 299, + 545, + 409 + ], + "type": "text", + "content": ", separate convolutional layers extract features from image " + }, + { + "bbox": [ + 304, + 299, + 545, + 409 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 304, + 299, + 545, + 409 + ], + "type": "text", + "content": " and semantics " + }, + { + "bbox": [ + 304, + 299, + 545, + 409 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 304, + 299, + 545, + 409 + ], + "type": "text", + "content": " (one-hot encoded). Features from the two pipelines are concatenated and fed to additional layers. Features at different resolutions " + }, + { + "bbox": [ + 304, + 299, + 545, + 409 + ], + "type": "inline_equation", + "content": "(H / 2^{l}, W / 2^{l})" + }, + { + "bbox": [ + 304, + 299, + 545, + 409 + ], + "type": "text", + "content": " form a pyramid " + }, + { + "bbox": [ + 304, + 299, + 545, + 409 + ], + "type": "inline_equation", + "content": "\\{f_{i}^{(l)} \\mid 2 \\leq l \\leq 6\\}" + }, + { + "bbox": [ + 304, + 299, + 545, + 409 + ], + "type": "text", + "content": ". The semantic information in these features helps delineate objects in dark shadows, where appearance is more homogeneous." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 417, + 531, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 417, + 531, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 417, + 531, + 430 + ], + "type": "text", + "content": "3.2. Iterative decoder with a learned upsampler" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 436, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 491 + ], + "type": "text", + "content": "Following [22, 37], an iterative residual refinement decoder starts from zero estimate " + }, + { + "bbox": [ + 304, + 436, + 545, + 491 + ], + "type": "inline_equation", + "content": "F_{1\\rightarrow 2}^{(7)} = 0" + }, + { + "bbox": [ + 304, + 436, + 545, + 491 + ], + "type": "text", + "content": ". For iteration " + }, + { + "bbox": [ + 304, + 436, + 545, + 491 + ], + "type": "inline_equation", + "content": "l \\in \\{6,5,4,3,2\\}" + }, + { + "bbox": [ + 304, + 436, + 545, + 491 + ], + "type": "text", + "content": ", the decoder refines feature map " + }, + { + "bbox": [ + 304, + 436, + 545, + 491 + ], + "type": "inline_equation", + "content": "F_{1\\rightarrow 2}^{(l + 1)}" + }, + { + "bbox": [ + 304, + 436, + 545, + 491 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 436, + 545, + 491 + ], + "type": "inline_equation", + "content": "F_{1\\rightarrow 2}^{(l)}" + }, + { + "bbox": [ + 304, + 436, + 545, + 491 + ], + "type": "text", + "content": " based on features " + }, + { + "bbox": [ + 304, + 436, + 545, + 491 + ], + "type": "inline_equation", + "content": "f_{1}^{(l)}, f_{2}^{(l)}" + }, + { + "bbox": [ + 304, + 436, + 545, + 491 + ], + "type": "text", + "content": " (Fig. 2(b))." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 491, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 615 + ], + "type": "text", + "content": "More specifically, " + }, + { + "bbox": [ + 304, + 491, + 545, + 615 + ], + "type": "inline_equation", + "content": "F_{1\\rightarrow 2}^{(l + 1)}" + }, + { + "bbox": [ + 304, + 491, + 545, + 615 + ], + "type": "text", + "content": " is upsampled to match the resolution and used to warp " + }, + { + "bbox": [ + 304, + 491, + 545, + 615 + ], + "type": "inline_equation", + "content": "f_2^{(l)}" + }, + { + "bbox": [ + 304, + 491, + 545, + 615 + ], + "type": "text", + "content": " to yield warped feature " + }, + { + "bbox": [ + 304, + 491, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\hat{f}_1^{(l)}" + }, + { + "bbox": [ + 304, + 491, + 545, + 615 + ], + "type": "text", + "content": ". Correlation volumes are computed between " + }, + { + "bbox": [ + 304, + 491, + 545, + 615 + ], + "type": "inline_equation", + "content": "f_1^{(l)}" + }, + { + "bbox": [ + 304, + 491, + 545, + 615 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 491, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\hat{f}_1^{(l)}" + }, + { + "bbox": [ + 304, + 491, + 545, + 615 + ], + "type": "text", + "content": ". A one-by-one convolutional layer " + }, + { + "bbox": [ + 304, + 491, + 545, + 615 + ], + "type": "inline_equation", + "content": "C^{(l)}" + }, + { + "bbox": [ + 304, + 491, + 545, + 615 + ], + "type": "text", + "content": " compresses the number of channels to a fixed number so that the same layer can be reused across all iterations as proposed by [22]. A flow estimator network predicts a flow residual to be added to the current estimate, and a context network then aggregates flow information spatially and refines the current flow again." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 616, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 713 + ], + "type": "text", + "content": "A learned upsampler network upsamples the final output " + }, + { + "bbox": [ + 304, + 616, + 545, + 713 + ], + "type": "inline_equation", + "content": "F_{1\\rightarrow 2}^{(2)}" + }, + { + "bbox": [ + 304, + 616, + 545, + 713 + ], + "type": "text", + "content": " in our system. This is different from ARFlow [37], which simply uses four-fold bilinear interpolation, making the final flow boundaries blurry. In contrast, our model learns to sharpen flow boundaries based on the semantic inputs, which have clear boundaries around moving objects. To this end, we add a convex upsampler network similar to the one in RAFT [68]. Different from UPFlow [42]," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9568" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 77, + 210, + 177 + ], + "blocks": [ + { + "bbox": [ + 58, + 77, + 210, + 177 + ], + "lines": [ + { + "bbox": [ + 58, + 77, + 210, + 177 + ], + "spans": [ + { + "bbox": [ + 58, + 77, + 210, + 177 + ], + "type": "image", + "image_path": "1f578db99d3ccda59acd52464574df776530c318c409b21f7fd6dae0bee7b43c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 180, + 205, + 191 + ], + "lines": [ + { + "bbox": [ + 67, + 180, + 205, + 191 + ], + "spans": [ + { + "bbox": [ + 67, + 180, + 205, + 191 + ], + "type": "text", + "content": "(a) shared pyramidal encoder " + }, + { + "bbox": [ + 67, + 180, + 205, + 191 + ], + "type": "inline_equation", + "content": "(i\\in \\{1,2\\})" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 230, + 78, + 533, + 174 + ], + "blocks": [ + { + "bbox": [ + 230, + 78, + 533, + 174 + ], + "lines": [ + { + "bbox": [ + 230, + 78, + 533, + 174 + ], + "spans": [ + { + "bbox": [ + 230, + 78, + 533, + 174 + ], + "type": "image", + "image_path": "c30e1a5c14116a469523b3f00f526371d8f8a59ad87c073d3cadb07e3d7d3548.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 259, + 180, + 507, + 191 + ], + "lines": [ + { + "bbox": [ + 259, + 180, + 507, + 191 + ], + "spans": [ + { + "bbox": [ + 259, + 180, + 507, + 191 + ], + "type": "text", + "content": "(b) one iteration (at the " + }, + { + "bbox": [ + 259, + 180, + 507, + 191 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 259, + 180, + 507, + 191 + ], + "type": "text", + "content": "th-level) of the iterative decoder " + }, + { + "bbox": [ + 259, + 180, + 507, + 191 + ], + "type": "inline_equation", + "content": "(l\\in \\{6,5,4,3,2\\})" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 52, + 225, + 544, + 430 + ], + "blocks": [ + { + "bbox": [ + 79, + 203, + 512, + 216 + ], + "lines": [ + { + "bbox": [ + 79, + 203, + 512, + 216 + ], + "spans": [ + { + "bbox": [ + 79, + 203, + 512, + 216 + ], + "type": "text", + "content": "Figure 2. Network structure. See text in Sec. 3.1 and Sec. 3.2 for explanations. More detailed diagrams are in appendix." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 225, + 544, + 430 + ], + "lines": [ + { + "bbox": [ + 52, + 225, + 544, + 430 + ], + "spans": [ + { + "bbox": [ + 52, + 225, + 544, + 430 + ], + "type": "image", + "image_path": "4b0eca463840993e1e56b7ce31b8aa97cb087ccf3349b518dc7346f392bc4ee8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 437, + 473, + 449 + ], + "lines": [ + { + "bbox": [ + 117, + 437, + 473, + 449 + ], + "spans": [ + { + "bbox": [ + 117, + 437, + 473, + 449 + ], + "type": "text", + "content": "Figure 3. Illustration of semantic augmentation as self-supervision. See text in Sec. 3.3 for details." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 479, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 559 + ], + "type": "text", + "content": "our learned upsampler is only used to upsample from internal estimate " + }, + { + "bbox": [ + 46, + 479, + 287, + 559 + ], + "type": "inline_equation", + "content": "F_{1\\rightarrow 2}^{(l)}" + }, + { + "bbox": [ + 46, + 479, + 287, + 559 + ], + "type": "text", + "content": " to output flow " + }, + { + "bbox": [ + 46, + 479, + 287, + 559 + ], + "type": "inline_equation", + "content": "U_{1\\rightarrow 2}^{(l)}" + }, + { + "bbox": [ + 46, + 479, + 287, + 559 + ], + "type": "text", + "content": ", but not when we upsample " + }, + { + "bbox": [ + 46, + 479, + 287, + 559 + ], + "type": "inline_equation", + "content": "F_{1\\rightarrow 2}^{(l + 1)}" + }, + { + "bbox": [ + 46, + 479, + 287, + 559 + ], + "type": "text", + "content": " at the first step in the decoder iteration. The upsampled " + }, + { + "bbox": [ + 46, + 479, + 287, + 559 + ], + "type": "inline_equation", + "content": "U_{1\\rightarrow 2}^{(l)}" + }, + { + "bbox": [ + 46, + 479, + 287, + 559 + ], + "type": "text", + "content": " has resolution " + }, + { + "bbox": [ + 46, + 479, + 287, + 559 + ], + "type": "inline_equation", + "content": "(H / 2^{l - 2},W / 2^{l - 2})" + }, + { + "bbox": [ + 46, + 479, + 287, + 559 + ], + "type": "text", + "content": ", so " + }, + { + "bbox": [ + 46, + 479, + 287, + 559 + ], + "type": "inline_equation", + "content": "U_{1\\rightarrow 2} = U_{1\\rightarrow 2}^{(2)}" + }, + { + "bbox": [ + 46, + 479, + 287, + 559 + ], + "type": "text", + "content": " (with the original resolution) is the final flow prediction of the network." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 571, + 269, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 571, + 269, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 571, + 269, + 586 + ], + "type": "text", + "content": "3.3. Semantic augmentation as self-supervision" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "type": "text", + "content": "ARFlow has a very effective in-network augmentation module that samples random transformations " + }, + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\theta_1},\\mathcal{T}_{\\theta_2}" + }, + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "type": "text", + "content": " of the flow prediction " + }, + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "type": "inline_equation", + "content": "U_{1\\rightarrow 2}" + }, + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "type": "text", + "content": " in the first pass of the network and then uses the transformed images " + }, + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\hat{I}_1 = \\mathcal{T}_{\\theta_1}(I_1),\\hat{I}_2 =" + }, + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\theta_2}(I_2)" + }, + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "type": "text", + "content": " in a second pass. The prediction " + }, + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "type": "inline_equation", + "content": "U_{1\\rightarrow 2}" + }, + { + "bbox": [ + 46, + 592, + 287, + 676 + ], + "type": "text", + "content": " is also transformed accordingly and used to self-supervise the output of the second pass. See [37] for details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "We retain the augmentation module but make a third pass of the network using semantics-transformed inputs for self-supervision in addition to the ARFlow augmentations of ap" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 479, + 545, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 527 + ], + "type": "text", + "content": "pearance (e.g., color jitter, random noise) and spatial transformations (e.g., random rotation, random rescaling). The idea behind semantic augmentation is to blend in real object motions across samples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 529, + 546, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 546, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 546, + 625 + ], + "type": "text", + "content": "To this end (see Fig. 3), we carve objects from other samples based on their semantic maps and paste them as moving foreground objects into the current sample " + }, + { + "bbox": [ + 304, + 529, + 546, + 625 + ], + "type": "inline_equation", + "content": "I_{1}, I_{2}" + }, + { + "bbox": [ + 304, + 529, + 546, + 625 + ], + "type": "text", + "content": ", thereby producing augmented images " + }, + { + "bbox": [ + 304, + 529, + 546, + 625 + ], + "type": "inline_equation", + "content": "\\tilde{I}_{1}, \\tilde{I}_{2}" + }, + { + "bbox": [ + 304, + 529, + 546, + 625 + ], + "type": "text", + "content": ". The segmentation maps " + }, + { + "bbox": [ + 304, + 529, + 546, + 625 + ], + "type": "inline_equation", + "content": "S_{1}, S_{2}" + }, + { + "bbox": [ + 304, + 529, + 546, + 625 + ], + "type": "text", + "content": " and the first pass output flow " + }, + { + "bbox": [ + 304, + 529, + 546, + 625 + ], + "type": "inline_equation", + "content": "U_{1 \\to 2}" + }, + { + "bbox": [ + 304, + 529, + 546, + 625 + ], + "type": "text", + "content": " are transformed accordingly, and the transformed " + }, + { + "bbox": [ + 304, + 529, + 546, + 625 + ], + "type": "inline_equation", + "content": "\\tilde{U}_{1 \\to 2}" + }, + { + "bbox": [ + 304, + 529, + 546, + 625 + ], + "type": "text", + "content": " self-supervises the third-pass network outputs " + }, + { + "bbox": [ + 304, + 529, + 546, + 625 + ], + "type": "inline_equation", + "content": "\\tilde{U}_{1 \\to 2}' = \\mathcal{F}(\\tilde{I}_{1}, \\tilde{I}_{2}, \\tilde{S}_{1}, \\tilde{S}_{2})" + }, + { + "bbox": [ + 304, + 529, + 546, + 625 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 529, + 546, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 529, + 546, + 625 + ], + "type": "text", + "content": " is the flow network." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 627, + 545, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 627, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 304, + 627, + 545, + 699 + ], + "type": "text", + "content": "Similar to occlusion hallucination in SelFlow [39], our semantic augmentation also creates new occlusions and uses the reliable non-occluded flow from the first pass to self-supervise the flow on those newly occluded pixels. However, rather than superpixels filled with noise, we create occlusions with real objects, which are much more realistic." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 701, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 712 + ], + "type": "text", + "content": "Moreover, since we determine the motion of each oc" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9569" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "content": "cluder, we can use its known optical flow for motion self-supervision. As illustrated in Fig. 3, the occluder motion can be very different from that of the background. This trains the model to assume less smoothness around highly-dynamic objects and to output sharper motion estimates near motion boundaries. These hallucinated but realistic occluders also provide more examples of actively moving objects, which are relatively scarce in the original datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 289, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 289, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 289, + 315 + ], + "type": "text", + "content": "**Vehicles** We augment vehicle classes (car, truck, bus, train) as they are major sources of errors in previous models (see Tab. 5 for statistics). We group these four classes because they are easily confused with each other in the semantic input. To carve out a single car instance with high probability, we find connected components of vehicle segmentation regions using OpenCV [3] and accept masks whose bounding boxes are 50-300 pixels wide and 50-150 pixels high. We also drop masks that occupy less than " + }, + { + "bbox": [ + 46, + 194, + 289, + 315 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 194, + 289, + 315 + ], + "type": "text", + "content": " area of their bounding boxes to remove largely occluded vehicles." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 338, + 289, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 289, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 289, + 460 + ], + "type": "text", + "content": "Poles We augment pole classes (pole, traffic light, traffic sign) not only because they are common foreground objects that cause occlusions, but also because their motions are usually poorly estimated due to their thin shapes. We use a 200-pixel wide, image-height sliding window to scan the segmentation of these pole classes and find the window with the most pole pixels. If more than " + }, + { + "bbox": [ + 46, + 338, + 289, + 460 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 338, + 289, + 460 + ], + "type": "text", + "content": " pole pixels occur in it, that object mask is stored. This produces groups of nearby poles rather than just individual poles, which may not provide enough augmentation by themselves." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 484, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 592 + ], + "type": "text", + "content": "Sky We also augment the sky through prior knowledge that flow in the sky should be very small. We shrink the sky flow estimate by half before using it for self-supervision. As shown in Fig. 3, the middle tree part in " + }, + { + "bbox": [ + 46, + 484, + 287, + 592 + ], + "type": "inline_equation", + "content": "U_{1\\rightarrow 2}" + }, + { + "bbox": [ + 46, + 484, + 287, + 592 + ], + "type": "text", + "content": " is very blurry, with a large part of sky mixed in there. After shrinking sky flow by half, the tree in " + }, + { + "bbox": [ + 46, + 484, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\tilde{U}_{1\\rightarrow 2}" + }, + { + "bbox": [ + 46, + 484, + 287, + 592 + ], + "type": "text", + "content": " now has sharper flow around the boundary between tree and sky. Using this as self-supervision helps the model sharpen the motion boundaries between sky and other regions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Another way to apply this prior knowledge is to postprocess, i.e., shrink the final output flow in the sky region, instead of using the shrunk flow to self-supervise. However, the latter approach is preferable because it allows the network to balance prior knowledge (self-supervision) with the current sample observation (photometric loss). Our experiments do show cases where some small objects (such as some overhead electric power lines) are misclassified as part of the sky in the semantic input. In these cases, the network learns to rely less on prior knowledge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 402, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 402, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 402, + 84 + ], + "type": "text", + "content": "3.4.Occluder cache" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 90, + 547, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 547, + 149 + ], + "type": "text", + "content": "During training, we maintain a cache of occluding objects for semantic augmentation. We sample objects from the cache to augment the current batch and then push any new objects found in the batch into the cache, with random replacement if the cache is full." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 150, + 547, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 150, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 304, + 150, + 547, + 245 + ], + "type": "text", + "content": "Specifically, after we load a sample for training, we first use the current model to infer its optical flow and then search within the sample for any cars (or poles) that meet the standards defined in Sec. 3.3 as \"cutouts\". For each cutout, we first find the flow vectors on all pixels that belong to that cutout and then compute the mean flow for later reference. We store the current sample, the cutout mask, and the mean flow together in the cache." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 246, + 547, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 547, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 547, + 366 + ], + "type": "text", + "content": "Later on, when we randomly select a cutout to augment another sample, we also retrieve the stored mean flow of this cutout, which is then augmented by random rescaling (by 0.8-1.5 times) and reversing (with " + }, + { + "bbox": [ + 304, + 246, + 547, + 366 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 246, + 547, + 366 + ], + "type": "text", + "content": " probability). We use the augmented mean flow to translate the entire cutout object as an occluder to generate a new sample. For realism, occluders are pasted to the same location they occupied in their image of origin. Holding occluders in a cache provides a random mixture of samples across multiple batches and makes semantic augmentation very efficient." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 372, + 397, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 372, + 397, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 397, + 384 + ], + "type": "text", + "content": "3.5. Loss functions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 390, + 547, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 547, + 443 + ], + "type": "text", + "content": "Photometric loss In the first pass, forward and backward flow " + }, + { + "bbox": [ + 304, + 390, + 547, + 443 + ], + "type": "inline_equation", + "content": "U_{1\\rightarrow 2}^{(l)}, U_{2\\rightarrow 1}^{(l)}" + }, + { + "bbox": [ + 304, + 390, + 547, + 443 + ], + "type": "text", + "content": " at each level are predicted and occlusion masks " + }, + { + "bbox": [ + 304, + 390, + 547, + 443 + ], + "type": "inline_equation", + "content": "O_{1\\rightarrow 2}^{(l)}, O_{2\\rightarrow 1}^{(l)}" + }, + { + "bbox": [ + 304, + 390, + 547, + 443 + ], + "type": "text", + "content": " are computed by forward-backward consistency [45]. Frames are warped by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 328, + 448, + 523, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 448, + 523, + 466 + ], + "spans": [ + { + "bbox": [ + 328, + 448, + 523, + 466 + ], + "type": "interline_equation", + "content": "I _ {i} ^ {\\prime (l)} (\\pmb {p}) = I _ {j} ^ {(l)} (\\pmb {p} + U _ {i \\rightarrow j} ^ {(l)} (\\pmb {p})), \\quad (i, j \\in \\{1, 2 \\})", + "image_path": "406ef424747832b2393238e17e68a010bc5299c9cbe94cbe1ee7dd8f8bcfa186.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "inline_equation", + "content": "I_{i}^{(l)}" + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "text", + "content": " down-sampled to the " + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "text", + "content": "-th scale, and " + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "text", + "content": " denotes pixel coordinates at that scale. Following [37], three measures, namely " + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "text", + "content": "-distance " + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "inline_equation", + "content": "(\\rho_{1})" + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "text", + "content": ", structural similarity (SSIM) [73] " + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "inline_equation", + "content": "(\\rho_{2})" + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "text", + "content": ", and census loss [45] " + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "inline_equation", + "content": "(\\rho_{3})" + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "text", + "content": " are linearly combined to measure photometric differences between " + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "inline_equation", + "content": "I_{i}^{(l)}" + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "inline_equation", + "content": "I_{i}^{\\prime(l)}" + }, + { + "bbox": [ + 304, + 472, + 547, + 575 + ], + "type": "text", + "content": ". Occlusion regions are masked out and both forward and backward directions are taken into account at each level. The final photometric loss is" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 580, + 545, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 580, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 306, + 580, + 545, + 626 + ], + "type": "interline_equation", + "content": "\\ell_ {\\mathrm {p h}} = \\frac {1}{2} \\sum_ {(i, j) \\in \\{(1, 2), (2, 1) \\}} \\sum_ {l = 2} ^ {6} a _ {l} \\sum_ {k = 1} ^ {3} c _ {k} \\rho_ {k} \\left(I _ {i} ^ {(l)}, I _ {i} ^ {\\prime (l)}, O _ {i \\rightarrow j} ^ {(l)}\\right). \\tag {1}", + "image_path": "a6afb54eb74f89590c0e7094752d1ae70de326b616e2bd3a945e9bc39102a194.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 639, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 547, + 676 + ], + "type": "text", + "content": "Smoothness loss Unlike most previous work, we do not include a smoothness loss because we find it to conflict with our learned upsampler. See ablation study in Sec. 4.5." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "content": "Augmentation loss As in ARFlow [37], our second forward pass computes the " + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1" + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "content": "-distance between the trans-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "text", + "content": "9570" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 69, + 494, + 255 + ], + "blocks": [ + { + "bbox": [ + 100, + 69, + 494, + 255 + ], + "lines": [ + { + "bbox": [ + 100, + 69, + 494, + 255 + ], + "spans": [ + { + "bbox": [ + 100, + 69, + 494, + 255 + ], + "type": "table", + "html": "
MethodTrainTest
2012 EPE2015 EPE2012 Fl-nocEPEFl-all2015 Fl-nocFl-bgFl-fg
supervisedPWC-Net+ [66]-(1.50)3.361.47.724.917.697.88
IRR-PWC [22]-(1.63)3.211.67.654.867.687.52
RAFT [68]-(0.63)--5.103.074.746.87
Separable Flow [85]-(0.69)--4.532.784.255.92
unsupervisedSelFlow [39]1.694.844.312.214.199.6512.6821.74
SimFlow [24]-5.19--13.388.2112.6017.27
ARFlow [37]1.442.85-1.811.80---
UFlow [27]1.682.714.261.911.138.419.7817.87
UPFlow [42]1.272.45-1.49.38---
Ours (baseline)1.392.614.301.79.896.988.8215.21
Ours (+enc)†1.292.423.971.58.993.978.1913.01
Ours (+enc +aug)†1.282.183.901.58.383.907.4812.91
", + "image_path": "a0fa49c82b26e757e90ee9676028000104af60fc54e15a78edce27fa6d41aa86.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 261, + 545, + 307 + ], + "lines": [ + { + "bbox": [ + 46, + 261, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 545, + 307 + ], + "type": "text", + "content": "Table 1. KITTI benchmark results (EPE/px and Fl/%). Metrics evaluated at 'all' (all pixels, default for EPE), 'noc' (non-occlusions), 'bg' (background), and 'fg' (foreground). Key metrics (used in official ranking) are underlined. Our 'baseline' is an adapted ARFlow with added learned upsampler and no smoothness loss. +enc means adding semantic encoder; +aug means adding semantic augmentation. () means evaluation data used in training. - means unavailable. " + }, + { + "bbox": [ + 46, + 261, + 545, + 307 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 261, + 545, + 307 + ], + "type": "text", + "content": " denotes models with semantic inputs. For all metrics, lower is better." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 324, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 350 + ], + "type": "text", + "content": "formed flow " + }, + { + "bbox": [ + 46, + 324, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\hat{U}_{1\\rightarrow 2}" + }, + { + "bbox": [ + 46, + 324, + 287, + 350 + ], + "type": "text", + "content": " and the second pass output " + }, + { + "bbox": [ + 46, + 324, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\hat{U}_{1\\rightarrow 2}^{\\prime} = \\mathcal{F}(\\hat{I}_1,\\hat{I}_2,\\hat{S}_1,\\hat{S}_2)" + }, + { + "bbox": [ + 46, + 324, + 287, + 350 + ], + "type": "text", + "content": " for each pixel " + }, + { + "bbox": [ + 46, + 324, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 46, + 324, + 287, + 350 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 356, + 238, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 356, + 238, + 372 + ], + "spans": [ + { + "bbox": [ + 96, + 356, + 238, + 372 + ], + "type": "interline_equation", + "content": "\\hat {D} (\\boldsymbol {p}) = \\left\\| \\hat {U} _ {1 \\rightarrow 2} (\\boldsymbol {p}) - \\hat {U} _ {1 \\rightarrow 2} ^ {\\prime} (\\boldsymbol {p}) \\right\\| _ {1}.", + "image_path": "d0a013b791e1e1158fc8b30028f02d6e692855989a3ce3ec0f889a2202ba3be2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 379, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 403 + ], + "type": "text", + "content": "This loss is then averaged over the transformed non-occluded region where self-supervision " + }, + { + "bbox": [ + 46, + 379, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\hat{U}_{1\\rightarrow 2}" + }, + { + "bbox": [ + 46, + 379, + 287, + 403 + ], + "type": "text", + "content": " is accurate" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 411, + 287, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 411, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 100, + 411, + 287, + 444 + ], + "type": "interline_equation", + "content": "\\ell_ {\\mathrm {a r}} = \\frac {\\sum_ {\\boldsymbol {p}} (1 - \\hat {O} _ {1 \\rightarrow 2} (\\boldsymbol {p})) \\hat {D} (\\boldsymbol {p})}{\\sum_ {\\boldsymbol {p}} (1 - \\hat {O} _ {1 \\rightarrow 2} (\\boldsymbol {p}))}. \\tag {2}", + "image_path": "755a434f1c91303c70b8b936674a1ed78864015e2ec936d8c03e3fb945501ecb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 451, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 486 + ], + "type": "text", + "content": "For semantic augmentation (third pass), since we are also doing self-supervision with a new generated sample, we use a similar loss definition" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 98, + 492, + 287, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 492, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 98, + 492, + 287, + 525 + ], + "type": "interline_equation", + "content": "\\ell_ {\\mathrm {a u g}} = \\frac {\\sum_ {\\boldsymbol {p}} \\left(1 - \\tilde {O} _ {1 \\rightarrow 2} (\\boldsymbol {p})\\right) \\tilde {D} (\\boldsymbol {p})}{\\sum_ {\\boldsymbol {p}} \\left(1 - \\tilde {O} _ {1 \\rightarrow 2} (\\boldsymbol {p})\\right)}, \\tag {3}", + "image_path": "e7b867ff771d63c67562a542231955dae320ecc0da9bb69e3ce6e78fa0ef4129.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 533, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 533, + 287, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{D} (\\pmb {p})" + }, + { + "bbox": [ + 46, + 533, + 287, + 559 + ], + "type": "text", + "content": " is the distance between the semanticaugmented flow " + }, + { + "bbox": [ + 46, + 533, + 287, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{U}_{1\\rightarrow 2}" + }, + { + "bbox": [ + 46, + 533, + 287, + 559 + ], + "type": "text", + "content": " and the third-pass output " + }, + { + "bbox": [ + 46, + 533, + 287, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{U}_{1\\rightarrow 2}^{\\prime}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 96, + 565, + 238, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 565, + 238, + 581 + ], + "spans": [ + { + "bbox": [ + 96, + 565, + 238, + 581 + ], + "type": "interline_equation", + "content": "\\tilde {D} (\\boldsymbol {p}) = \\left\\| \\tilde {U} _ {1 \\rightarrow 2} (\\boldsymbol {p}) - \\tilde {U} _ {1 \\rightarrow 2} ^ {\\prime} (\\boldsymbol {p}) \\right\\| _ {1},", + "image_path": "58edadc5828b883c12c02c6e7a728caa9d8613c135f1f0c1d5d43a624bc27365.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 587, + 259, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 259, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 259, + 601 + ], + "type": "text", + "content": "and the augmented mask " + }, + { + "bbox": [ + 46, + 587, + 259, + 601 + ], + "type": "inline_equation", + "content": "\\tilde{O}_{1\\rightarrow 2}" + }, + { + "bbox": [ + 46, + 587, + 259, + 601 + ], + "type": "text", + "content": " is now computed by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 96, + 607, + 237, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 607, + 237, + 622 + ], + "spans": [ + { + "bbox": [ + 96, + 607, + 237, + 622 + ], + "type": "interline_equation", + "content": "\\tilde {O} _ {1 \\rightarrow 2} = 1 - \\max (1 - O _ {1 \\rightarrow 2}, M)", + "image_path": "0bac4f3c002b62759385f570e3d9d5a931e760053d6e1311c100a81da33ff773.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 630, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 666 + ], + "type": "text", + "content": "because we penalize on both the originally non-occluded region " + }, + { + "bbox": [ + 46, + 630, + 287, + 666 + ], + "type": "inline_equation", + "content": "1 - O_{1\\rightarrow 2}" + }, + { + "bbox": [ + 46, + 630, + 287, + 666 + ], + "type": "text", + "content": " and the pasted foreground object mask " + }, + { + "bbox": [ + 46, + 630, + 287, + 666 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 630, + 287, + 666 + ], + "type": "text", + "content": " (of which we know the true motion)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 680, + 244, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 680, + 244, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 680, + 244, + 693 + ], + "type": "text", + "content": "Final loss (with " + }, + { + "bbox": [ + 46, + 680, + 244, + 693 + ], + "type": "inline_equation", + "content": "\\lambda = 0.02" + }, + { + "bbox": [ + 46, + 680, + 244, + 693 + ], + "type": "text", + "content": " as in ARFlow [37]):" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 116, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\ell = \\ell_ {\\mathrm {p h}} + \\lambda \\left(\\ell_ {\\mathrm {a r}} + \\ell_ {\\mathrm {a u g}}\\right). \\tag {4}", + "image_path": "270319af14411486b6b76c4b17a7782448c2321c632bfdafbfe35774c57f0e98.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 324, + 387, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 324, + 387, + 338 + ], + "spans": [ + { + "bbox": [ + 306, + 324, + 387, + 338 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 343, + 368, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 343, + 368, + 355 + ], + "spans": [ + { + "bbox": [ + 306, + 343, + 368, + 355 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 361, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 530 + ], + "type": "text", + "content": "We mainly use KITTI [15, 46] and Cityscapes [8] datasets for experiments. Following [37], we train our model first on KITTI raw sequences [46] (55.7k samples) and then fine-tune on KITTI-2015 multi-view extension [46] (11.8k samples). We validate our model using KITTI-2015 train [46] (200 samples) and KITTI-2012 train set [15] (194 samples) since they are the only sets that have optical flow labels. We also train on Cityscapes [8] sequences (83.3k samples) to test model generalization ability, and we sample every other frame to match its frame rate with KITTI. All semantic segmentation inputs are estimated using an off-the-shelf model [89] with a DeepLabV3Plus [7] backbone, which achieves " + }, + { + "bbox": [ + 304, + 361, + 545, + 530 + ], + "type": "inline_equation", + "content": "83.45\\%" + }, + { + "bbox": [ + 304, + 361, + 545, + 530 + ], + "type": "text", + "content": " mean IoU on Cityscapes and " + }, + { + "bbox": [ + 304, + 361, + 545, + 530 + ], + "type": "inline_equation", + "content": "72.82\\%" + }, + { + "bbox": [ + 304, + 361, + 545, + 530 + ], + "type": "text", + "content": " mean IoU on KITTI." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 536, + 437, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 437, + 549 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 437, + 549 + ], + "type": "text", + "content": "4.2. Implementation details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 554, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 545, + 638 + ], + "type": "text", + "content": "We implement the model in PyTorch [48] " + }, + { + "bbox": [ + 304, + 554, + 545, + 638 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 304, + 554, + 545, + 638 + ], + "type": "text", + "content": ". We use the Adam optimizer [32] with " + }, + { + "bbox": [ + 304, + 554, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 304, + 554, + 545, + 638 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 554, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 304, + 554, + 545, + 638 + ], + "type": "text", + "content": " and batch size 4. We first train on KITTI raw sequences [46] for 100k iterations with a fixed learning rate 0.0002 and then train on KITTI multi-view extension set [15, 46] for another 100k iterations using OneCycleLR schedule [62] with maximum learning rate 0.0004 and linear annealing." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 638, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 638, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 638, + 545, + 687 + ], + "type": "text", + "content": "For data augmentation, we include random horizontal flipping and swapping of the input frames. We resize the inputs to " + }, + { + "bbox": [ + 304, + 638, + 545, + 687 + ], + "type": "inline_equation", + "content": "256 \\times 832" + }, + { + "bbox": [ + 304, + 638, + 545, + 687 + ], + "type": "text", + "content": " before feeding into the network. The photometric loss weight for each scale " + }, + { + "bbox": [ + 304, + 638, + 545, + 687 + ], + "type": "inline_equation", + "content": "a_{l}" + }, + { + "bbox": [ + 304, + 638, + 545, + 687 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 638, + 545, + 687 + ], + "type": "inline_equation", + "content": "2 \\leq l \\leq 6" + }, + { + "bbox": [ + 304, + 638, + 545, + 687 + ], + "type": "text", + "content": ") in" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "Code and instructions are available at https://github.com/duke-vision/semantic-unsup-flow-release." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9571" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 272, + 144 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 272, + 144 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 272, + 144 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 272, + 144 + ], + "type": "table", + "html": "
MethodFl-allFl-nocFl-bgFl-fg
JFS [21]17.079.8115.922.92
SOF [57]16.8110.8614.6327.73
MRFlow [75]12.198.8610.1322.52
Bai et al. [1]11.628.758.6126.69
Ours (final)8.383.907.4812.91
", + "image_path": "1837df44ffbaed2651e485dc2dbce8ffd2943490d301b3338d54190c2c707589.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 212, + 286, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 212, + 286, + 283 + ], + "spans": [ + { + "bbox": [ + 47, + 212, + 286, + 283 + ], + "type": "text", + "content": "Eq. (1) are set as 1, 1, 1, 1, 0. The weights for three photometric distance measures " + }, + { + "bbox": [ + 47, + 212, + 286, + 283 + ], + "type": "inline_equation", + "content": "\\rho_{k}" + }, + { + "bbox": [ + 47, + 212, + 286, + 283 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 212, + 286, + 283 + ], + "type": "inline_equation", + "content": "1 \\leq k \\leq 3" + }, + { + "bbox": [ + 47, + 212, + 286, + 283 + ], + "type": "text", + "content": ") in Eq. (1) are set as 0.15, 0.85, 0 for the first 50k iterations and 0, 0, 1 afterwards. We start the appearance and spatial augmentation (second pass as in ARFlow [37]) at 50k iterations, and semantic augmentation (third pass) after 150k iterations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 300, + 192, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 192, + 313 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 192, + 313 + ], + "type": "text", + "content": "4.3. Benchmark testing results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 322, + 286, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 322, + 286, + 382 + ], + "spans": [ + { + "bbox": [ + 47, + 322, + 286, + 382 + ], + "type": "text", + "content": "As common practice, we evaluate optical flow predictions based on two error measurements, Fl (error rate) and EPE (mean " + }, + { + "bbox": [ + 47, + 322, + 286, + 382 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_2" + }, + { + "bbox": [ + 47, + 322, + 286, + 382 + ], + "type": "text", + "content": " distance). When computing Fl, the estimate of each pixel is considered correct if the error is smaller than 3 pixels or " + }, + { + "bbox": [ + 47, + 322, + 286, + 382 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 47, + 322, + 286, + 382 + ], + "type": "text", + "content": " of the magnitude of ground-truth flow [46]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 384, + 286, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 286, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 286, + 504 + ], + "type": "text", + "content": "As shown in Tab. 1, our semantic modules (both semantic encoder and semantic augmentation) improve performance on both KITTI-2012 and KITTI-2015 sets on all metrics. Our final model achieves " + }, + { + "bbox": [ + 46, + 384, + 286, + 504 + ], + "type": "inline_equation", + "content": "8.38\\%" + }, + { + "bbox": [ + 46, + 384, + 286, + 504 + ], + "type": "text", + "content": " F1-all error rate on KITTI-2015 test set, which is significantly better than ARFlow [37] " + }, + { + "bbox": [ + 46, + 384, + 286, + 504 + ], + "type": "inline_equation", + "content": "(11.80\\%)" + }, + { + "bbox": [ + 46, + 384, + 286, + 504 + ], + "type": "text", + "content": ", from which we adapt. We also outperform the current state-of-the-art UPFlow [42] " + }, + { + "bbox": [ + 46, + 384, + 286, + 504 + ], + "type": "inline_equation", + "content": "(9.38\\%)" + }, + { + "bbox": [ + 46, + 384, + 286, + 504 + ], + "type": "text", + "content": " by a clear margin. All these results strongly suggest that adding semantic inputs can help improve unsupervised flow estimation significantly." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 507, + 286, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 286, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 286, + 579 + ], + "type": "text", + "content": "In addition, Tab. 2 shows that our test results significantly outperform all previous semantic optical flow methods, most of which are based on traditional energy minimization. We are the first to apply semantic inputs to recent unsupervised flow networks, so we are able to push the state-of-the-art by a clear margin." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 596, + 155, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 155, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 155, + 608 + ], + "type": "text", + "content": "4.4. Qualitative results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": "Some qualitative results are shown in Fig. 4. We can see that our flow outputs have very sharp boundaries and very clean object motion, especially around cars and poles. Our model successfully learns to adapt flow estimation to the semantic map input. Moreover, our model is able to handle very challenging samples where the foreground car motion is drastically different from the background due to the challenging samples we create in semantic augmentation." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 310, + 70, + 542, + 194 + ], + "blocks": [ + { + "bbox": [ + 47, + 152, + 286, + 185 + ], + "lines": [ + { + "bbox": [ + 47, + 152, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 286, + 185 + ], + "type": "text", + "content": "Table 2. KITTI-2015 test results (Fl/\\%) compared with other semantic optical flow methods. Metrics evaluated at 'all' (all pixels), 'noc' (non-occlusions), 'bg' (background), and 'fg' (foreground)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 70, + 542, + 194 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 542, + 194 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 542, + 194 + ], + "type": "table", + "html": "
upno smencaugFl-allEPE-allEPE-nocEPE-occ
10.362.902.076.97
19.922.691.896.57
29.832.651.866.40
39.752.611.856.54
49.752.641.856.55
✓10.222.831.976.75
✓✓8.872.611.856.04
✓✓38.262.421.735.79
✓✓✓8.802.481.606.64
✓✓3✓7.792.181.405.64
", + "image_path": "8e9b8866ad2171ccb810955a26d057356e880f8be6848a0e401c5ed180d3de35.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 276, + 545, + 340 + ], + "blocks": [ + { + "bbox": [ + 305, + 202, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 305, + 202, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 305, + 202, + 545, + 258 + ], + "type": "text", + "content": "Table 3. Ablation study on KITTI-2015 train set (EPE/px and F1%). 'up' means the learned upsampler; 'no sm' means turning off smoothness loss; 'enc' means the level at which we merge semantic features and image features in the encoder; 'aug' means semantic augmentation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 276, + 545, + 340 + ], + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 340 + ], + "type": "table", + "html": "
Options of augFl-allEPE-allEPE-nocEPE-occ
Ours (final)7.792.181.405.64
start from 100k7.922.191.405.41
vehicles only7.942.211.445.78
loss on new occ8.152.271.425.62
", + "image_path": "d6899810fa46e2fa8b8110cd20f6e09521321eab8a3028129ff4d8c259abe500.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 348, + 545, + 370 + ], + "lines": [ + { + "bbox": [ + 305, + 348, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 305, + 348, + 545, + 370 + ], + "type": "text", + "content": "Table 4. Ablation study of different semantic augmentation options on KITTI-2015 train (EPE/px and F1/%). See text for explanations." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 399, + 397, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 397, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 397, + 411 + ], + "type": "text", + "content": "4.5. Ablation study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 422, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 553 + ], + "type": "text", + "content": "We also do ablation study to show the effectiveness of each of our proposed modules. We can see in Tab. 3 that adding a semantic encoder to the vanilla ARFlow does help, and the optimal number of encoder layer groups added is 3. Also, our learned upsampler improves results significantly once we turn off the smoothness loss. This makes sense because the upsampler learns to adapt to boundaries, which can be distracted by the smoothness loss. Moreover, both our semantic encoder and semantic augmentation module help improve the results further, which suggests the effectiveness of adding semantic inputs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 558, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 712 + ], + "type": "text", + "content": "Another ablation study is on different options for the semantic augmentation (Tab. 4). We try starting augmentation earlier from 100k iterations, which works slightly worse. There are mainly two reasons: (1) we use the first forward output to self-supervise our augmented output, so if we start early, the model will use poor output to self-supervise the augmented pass; and (2) our semantic augmentation creates very challenging samples with objects moving very differently from the background, so exposing these hard samples to the model too early may make it hard to train. Apart from this, we also try only using vehicles to augment as well as focusing loss on the newly occluded region, which are both inferior to our final version." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9572" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 148, + 225 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 148, + 225 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 148, + 225 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 148, + 225 + ], + "type": "image", + "image_path": "07c59636c83be6777bb7d26b345ec42e13f6245ab9cf5eed0a798a8bc2146132.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 226, + 124, + 237 + ], + "lines": [ + { + "bbox": [ + 75, + 226, + 124, + 237 + ], + "spans": [ + { + "bbox": [ + 75, + 226, + 124, + 237 + ], + "type": "text", + "content": "Image Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 148, + 70, + 246, + 225 + ], + "blocks": [ + { + "bbox": [ + 148, + 70, + 246, + 225 + ], + "lines": [ + { + "bbox": [ + 148, + 70, + 246, + 225 + ], + "spans": [ + { + "bbox": [ + 148, + 70, + 246, + 225 + ], + "type": "image", + "image_path": "cbb5321d0383d9cd403ec9c926150713f4acbaadab3bca47ced4520392966d5e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 226, + 231, + 237 + ], + "lines": [ + { + "bbox": [ + 165, + 226, + 231, + 237 + ], + "spans": [ + { + "bbox": [ + 165, + 226, + 231, + 237 + ], + "type": "text", + "content": "Semantics Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 247, + 70, + 345, + 225 + ], + "blocks": [ + { + "bbox": [ + 247, + 70, + 345, + 225 + ], + "lines": [ + { + "bbox": [ + 247, + 70, + 345, + 225 + ], + "spans": [ + { + "bbox": [ + 247, + 70, + 345, + 225 + ], + "type": "image", + "image_path": "b18ba7a222c78a7ae580df67155847eef12a2edad21db45d5541879811b4cb0a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 226, + 313, + 236 + ], + "lines": [ + { + "bbox": [ + 280, + 226, + 313, + 236 + ], + "spans": [ + { + "bbox": [ + 280, + 226, + 313, + 236 + ], + "type": "text", + "content": "ARFlow" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 240, + 547, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 240, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 547, + 264 + ], + "type": "text", + "content": "Figure 4. Qualitative results on KITTI test set (sample #7, 20, 38, 112, 183) compared with ARFlow [37] and UPFlow [42]. We only show first frame input for conciseness. Our outputs have much clearer boundaries around objects. For more qualitative examples, see Appendix." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 346, + 70, + 445, + 225 + ], + "blocks": [ + { + "bbox": [ + 346, + 70, + 445, + 225 + ], + "lines": [ + { + "bbox": [ + 346, + 70, + 445, + 225 + ], + "spans": [ + { + "bbox": [ + 346, + 70, + 445, + 225 + ], + "type": "image", + "image_path": "81bafe3a42b8dceeb4b7fce071a2fb656aab904f459b663294b6eca2000eff17.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 226, + 412, + 236 + ], + "lines": [ + { + "bbox": [ + 380, + 226, + 412, + 236 + ], + "spans": [ + { + "bbox": [ + 380, + 226, + 412, + 236 + ], + "type": "text", + "content": "UPFlow" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 445, + 70, + 545, + 225 + ], + "blocks": [ + { + "bbox": [ + 445, + 70, + 545, + 225 + ], + "lines": [ + { + "bbox": [ + 445, + 70, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 445, + 70, + 545, + 225 + ], + "type": "image", + "image_path": "02089204af00bcd73238d308a144146cad31925f78d7c8c61735a06124c2374c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 226, + 505, + 236 + ], + "lines": [ + { + "bbox": [ + 484, + 226, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 484, + 226, + 505, + 236 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 75, + 282, + 519, + 360 + ], + "blocks": [ + { + "bbox": [ + 75, + 282, + 519, + 360 + ], + "lines": [ + { + "bbox": [ + 75, + 282, + 519, + 360 + ], + "spans": [ + { + "bbox": [ + 75, + 282, + 519, + 360 + ], + "type": "table", + "html": "
roadcarterrainvegetationsidewalkbuildingwallpole
Proportion42.4%17.6%14.0%11.6%6.2%4.1%1.3%1.1%
ARFlow [37]4.5715.799.4215.344.616.0016.3410.75
Ours (final)3.6710.178.7413.473.094.2711.189.44
Relative improvement19.6%35.6%7.3%12.2%33.0%28.7%31.6%12.2%
Reweighed contribution19.0%49.4%4.8%10.9%4.7%3.5%3.4%0.7%
", + "image_path": "be5ecba44ee485cd5a6314673f47cc7873dfc49f9180708a710e28cff251dc48.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 408, + 174, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 408, + 174, + 421 + ], + "spans": [ + { + "bbox": [ + 47, + 408, + 174, + 421 + ], + "type": "text", + "content": "4.6. Improvement analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 427, + 287, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 287, + 523 + ], + "type": "text", + "content": "It may be interesting to understand where our improvements come from in terms of semantic classes, so we compute the KITTI-2015 train set error for each class. As shown in Tab. 5, our model improves the flow on every class. After reweighing the absolute improvement of each class based on their proportion in the evaluated pixels, we find that car flow accounts for nearly half of our improvement on Fl-all, indicating the effectiveness of augmenting vehicle objects." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 531, + 173, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 173, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 173, + 544 + ], + "type": "text", + "content": "4.7. Generalization ability" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 550, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 550, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 550, + 287, + 658 + ], + "type": "text", + "content": "We also investigate the generalization ability of our semantic-aided flow models. We train flow models on Cityscapes and directly test them on KITTI-2015 train set without fine-tuning. Following [50], we crop the bottom " + }, + { + "bbox": [ + 46, + 550, + 287, + 658 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 550, + 287, + 658 + ], + "type": "text", + "content": " of the frame to remove the car logo and resize the frame to size " + }, + { + "bbox": [ + 46, + 550, + 287, + 658 + ], + "type": "inline_equation", + "content": "256 \\times 704" + }, + { + "bbox": [ + 46, + 550, + 287, + 658 + ], + "type": "text", + "content": ". All other experiment settings are exactly the same as for KITTI. As we can see from Tab. 6, adding semantic inputs significantly helps our unsupervised flow model generalize and adapt better across datasets." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 669, + 206, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 206, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 206, + 681 + ], + "type": "text", + "content": "5. Conclusion and future work" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "In this paper, we show that adding semantic segmentation inputs can help significantly improve the performance" + } + ] + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 308, + 406, + 544, + 470 + ], + "blocks": [ + { + "bbox": [ + 46, + 366, + 546, + 390 + ], + "lines": [ + { + "bbox": [ + 46, + 366, + 546, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 546, + 390 + ], + "type": "text", + "content": "Table 5. KITTI-2015 flow error (Fl-all/%) for each semantic class. The first row shows the proportion of each class among evaluated pixels (KITTI flow does not evaluate on every pixel). We only show classes that account for at least " + }, + { + "bbox": [ + 46, + 366, + 546, + 390 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 366, + 546, + 390 + ], + "type": "text", + "content": " of the evaluated pixels." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 406, + 544, + 470 + ], + "lines": [ + { + "bbox": [ + 308, + 406, + 544, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 406, + 544, + 470 + ], + "type": "table", + "html": "
MethodFl-allEPE-allEPE-nocEPE-occ
ARFlow (our impl.)13.214.082.889.40
Ours (baseline)12.273.812.439.91
Ours (+enc)†11.283.332.128.75
Ours (+enc +aug)†10.322.641.567.12
", + "image_path": "4685a30e710e8fd8d831d674fe3ac4d8359dba10f7cdf0baff8260804b356c97.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 478, + 546, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 478, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 478, + 546, + 544 + ], + "type": "text", + "content": "Table 6. Generalization results (train on Cityscapes, and test on KITTI-2015 train). We use our implementation of ARFlow [37] to run this experiment. Our 'baseline' is an adapted ARFlow with added learned upsampler and no smoothness loss. " + }, + { + "bbox": [ + 304, + 478, + 546, + 544 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 304, + 478, + 546, + 544 + ], + "type": "text", + "content": " enc means adding semantic encoder; " + }, + { + "bbox": [ + 304, + 478, + 546, + 544 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 304, + 478, + 546, + 544 + ], + "type": "text", + "content": " aug means adding semantic augmentation. " + }, + { + "bbox": [ + 304, + 478, + 546, + 544 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 304, + 478, + 546, + 544 + ], + "type": "text", + "content": " denotes models with semantic inputs." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 567, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 687 + ], + "type": "text", + "content": "of unsupervised optical flow networks on autonomous driving datasets. We propose a novel network model called SemARFlow, a semantic adaptation of ARFlow [37], with a semantic encoder, a learned upsampler, and a semantic augmentation module, where some domain-specific motion prior has been taken into account. Our network better predicts flow at occlusion regions and effectively sharpens flow estimates around object boundaries, even for very challenging samples. The additional semantic inputs also make our network generalize better across datasets." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "One direction for future work is motivated by the example illustrated in Fig. 5. Our model does output sharp flow" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9573" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 127, + 118 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 127, + 118 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 127, + 118 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 127, + 118 + ], + "type": "image", + "image_path": "1163bd8c232875b98653a17d74378498bfe4ef69b40c6cd4a3f95ac58e27a07a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 121, + 287, + 144 + ], + "lines": [ + { + "bbox": [ + 47, + 121, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 287, + 144 + ], + "type": "text", + "content": "Figure 5. An example on KITTI-2015 train set (sample #164). See the discussion in Sec. 5 for details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 72, + 206, + 118 + ], + "blocks": [ + { + "bbox": [ + 129, + 72, + 206, + 118 + ], + "lines": [ + { + "bbox": [ + 129, + 72, + 206, + 118 + ], + "spans": [ + { + "bbox": [ + 129, + 72, + 206, + 118 + ], + "type": "image", + "image_path": "d3f5dfb78ff75b20c7534879e6a07823a7a3f0014c676ba3b095e4f7f671dee7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 208, + 72, + 285, + 118 + ], + "blocks": [ + { + "bbox": [ + 208, + 72, + 285, + 118 + ], + "lines": [ + { + "bbox": [ + 208, + 72, + 285, + 118 + ], + "spans": [ + { + "bbox": [ + 208, + 72, + 285, + 118 + ], + "type": "image", + "image_path": "a7bdac7bbcf7508d90aa410a55955595fee7f51f1cf2bf6e69324c96fcd7f87f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 173, + 288, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 173, + 288, + 268 + ], + "spans": [ + { + "bbox": [ + 47, + 173, + 288, + 268 + ], + "type": "text", + "content": "boundaries between the cars and the background, thanks to good boundary information from the semantic maps. However, the boundary between these two fast-moving cars is inferred from image information only, because their semantic maps merge into one. As a result, that part of the flow boundary is less accurate. An interesting direction for future work is to use instance-level semantics as input for more detailed object masks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 285, + 287, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 287, + 321 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 287, + 321 + ], + "type": "text", + "content": "Acknowledgments. This research is based upon work supported by the National Science Foundation under Grant No. 1909821." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 106, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 106, + 345 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 106, + 345 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 353, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 353, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 353, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 353, + 287, + 396 + ], + "type": "text", + "content": "[1] Min Bai, Wenjie Luo, Kaustav Kundu, and Raquel Urtasun. Exploiting semantic information and deep matching for optical flow. In Proceedings of the European Conference on Computer Vision, pages 154-170. Springer, 2016. 2, 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 399, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 287, + 453 + ], + "type": "text", + "content": "[2] Shubhankar Borse, Ying Wang, Yizhe Zhang, and Fatih Porikli. Inverseform: A loss function for structured boundary-aware segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5901-5911, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 455, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 455, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 455, + 287, + 485 + ], + "type": "text", + "content": "[3] Gary Bradski. The opencv library. Dr. Dobb's Journal: Software Tools for the Professional Programmer, 25(11):120-123, 2000. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 489, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 489, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 53, + 489, + 287, + 532 + ], + "type": "text", + "content": "[4] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Proceedings of the European Conference on Computer Vision, pages 611-625. Springer, 2012. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 534, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 534, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 53, + 534, + 287, + 577 + ], + "type": "text", + "content": "[5] Yingfeng Cai, Lei Dai, Hai Wang, and Zhixiong Li. Multi-target pan-class intrinsic relevance driven model for improving semantic segmentation in autonomous driving. IEEE Transactions on Image Processing, 30:9069-9084, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 579, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 579, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 53, + 579, + 287, + 621 + ], + "type": "text", + "content": "[6] Linda Capito, Umit Ozguner, and Keith Redmill. Optical flow based visual potential field for autonomous driving. In IEEE Intelligent Vehicles Symposium, pages 885-891. IEEE, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 53, + 624, + 287, + 678 + ], + "type": "text", + "content": "[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European Conference on Computer Vision, pages 801-818, 2018. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 681, + 287, + 713 + ], + "type": "text", + "content": "[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "dataset for semantic urban scene understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3213-3223, 2016. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 108, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 108, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 312, + 108, + 545, + 161 + ], + "type": "text", + "content": "[9] Mingyu Ding, Zhe Wang, Bolei Zhou, Jianping Shi, Zhiwu Lu, and Ping Luo. Every frame counts: Joint learning of video segmentation and optical flow. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10713-10720, 2020. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 163, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 239 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Philipp Fischery, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick van der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE International Conference on Computer Vision, ICCV '15, pages 2758-2766, Washington, DC, USA, 2015. IEEE Computer Society. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 242, + 545, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 242, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 308, + 242, + 545, + 283 + ], + "type": "text", + "content": "[11] Cheng Feng, Long Ma, Congxuan Zhang, Zhen Chen, Liyue Ge, and Shaofeng Jiang. Ss-sf: Piecewise 3d scene flow estimation with semantic segmentation. IEEE Access, 9:22745-22759, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 286, + 545, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 340 + ], + "type": "text", + "content": "[12] Jun Fu, Jing Liu, Haijie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3146-3154, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 342, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 407 + ], + "type": "text", + "content": "[13] Aditya Ganeshan, Alexis Vallet, Yasunori Kudo, Shin-ichi Maeda, Tommi Kerola, Rares Ambrus, Dennis Park, and Adrien Gaidon. Warp-refine propagation: Semi-supervised auto-labeling via cycle-consistency. In Proceedings of the IEEE International Conference on Computer Vision, pages 15499–15509, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 410, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 410, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 410, + 545, + 453 + ], + "type": "text", + "content": "[14] Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In Proceedings of the European Conference on Computer Vision, pages 713-729. Springer, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 455, + 545, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 455, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 545, + 497 + ], + "type": "text", + "content": "[15] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research, 32(11):1231-1237, 2013. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 500, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 553 + ], + "type": "text", + "content": "[16] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 556, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 556, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 556, + 545, + 599 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 601, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 622 + ], + "type": "text", + "content": "[18] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial Intelligence, 17(1-3):185–203, 1981. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "type": "text", + "content": "[19] Zilong Huang, Xinggang Wang, Lichao Huang, Chang Huang, Yunchao Wei, and Wenyu Liu. Ccnet: Criss-cross attention for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 603-612, 2019. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[20] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. Liteflownet: A lightweight convolutional neural network for optical flow estimation. In Proceedings of the IEEE Conference" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9574" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 8981-8989, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[21] Junhwa Hur and Stefan Roth. Joint optical flow and temporally consistent semantic segmentation. In Proceedings of the European Conference on Computer Vision, pages 163-177. Springer, 2016. 2, 3, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "type": "text", + "content": "[22] Junhwa Hur and Stefan Roth. Iterative residual refinement for joint optical flow and occlusion estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5754-5763, 2019. 1, 2, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 241 + ], + "type": "text", + "content": "[23] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1647-1655, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 287, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 286 + ], + "type": "text", + "content": "[24] Woobin Im, Tae-Kyun Kim, and Sung-Eui Yoon. Unsupervised learning of optical flow with deep feature similarity. In Proceedings of the European Conference on Computer Vision, pages 172-188. Springer, 2020. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 287, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 331 + ], + "type": "text", + "content": "[25] Joel Janai, Fatma Guney, Anurag Ranjan, Michael Black, and Andreas Geiger. Unsupervised learning of multi-frame optical flow with occlusions. In Proceedings of the European Conference on Computer Vision, pages 690-706, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 386 + ], + "type": "text", + "content": "[26] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In Proceedings of the IEEE International Conference on Computer Vision, pages 9772-9781, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "type": "text", + "content": "[27] Rico Jonschkowski, Austin Stone, Jonathan T Barron, Ariel Gordon, Kurt Konolige, and Anelia Angelova. What matters in unsupervised optical flow. In Proceedings of the European Conference on Computer Vision, pages 557-572. Springer, 2020. 1, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "type": "text", + "content": "[28] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM computing surveys, 54(10s):1-41, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "text", + "content": "[29] Hannah H Kim, Celia Cintas, Girmaw Abebe Tadesse, and Skyler Speakman. Spatially constrained adversarial attack detection and localization in the representation space of optical flow networks. In International Joint Conference on Artificial Intelligence, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "type": "text", + "content": "[30] Hannah Halin Kim, Shuzhi Yu, and Carlo Tomasi. Joint detection of motion boundaries and occlusions. In Proceedings of the British Machine Vision Conference, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 579, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 624 + ], + "type": "text", + "content": "[31] Hannah Halin Kim, Shuzhi Yu, Shuai Yuan, and Carlo Tomasi. Cross-attention transformer for video interpolation. In Proceedings of the Asian Conference on Computer Vision Workshops, pages 320–337, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "text", + "content": "[32] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. Proceedings of the International Conference on Learning Representations, 2014. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[33] Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weissenborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Sylvain Gelly, Thomas Unterthiner, and Xiaohua Zhai. An image is worth 16x16 words: Transformers for image recogni" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "tion at scale. In Proceedings of the International Conference on Learning Representations, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "text", + "content": "[34] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "type": "text", + "content": "[35] Xiangtai Li, Jiangang Bai, Kuiyuan Yang, and Yunhai Tong. Flow2seg: Motion-aided semantic segmentation. In Proceedings of the International Conference on Artificial Neural Networks, pages 225–237. Springer, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 173, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 217 + ], + "type": "text", + "content": "[36] Zewen Li, Fan Liu, Wenjie Yang, Shouheng Peng, and Jun Zhou. A survey of convolutional neural networks: analysis, applications, and prospects. IEEE transactions on neural networks and learning systems, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 217, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 293 + ], + "type": "text", + "content": "[37] Liang Liu, Jiangning Zhang, Ruifei He, Yong Liu, Yabiao Wang, Ying Tai, Donghao Luo, Chengjie Wang, Jilin Li, and Feiyue Huang. Learning by analogy: Reliable supervision from transformations for unsupervised optical flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6489-6498, 2020. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "type": "text", + "content": "[38] Pengpeng Liu, Irwin King, Michael R Lyu, and Jia Xu. Ddflow: Learning optical flow with unlabeled data distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8770-8777, 2019. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "text", + "content": "[39] Pengpeng Liu, Michael Lyu, Irwin King, and Jia Xu. Selfflow: Self-supervised learning of optical flow. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4571-4580, 2019. 1, 2, 4, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "text", + "content": "[40] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3431-3440, 2015. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "type": "text", + "content": "[41] Bruce D Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In International Joint Conference on Artificial Intelligence, volume 2, pages 674-679, 1981. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 472, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 525 + ], + "type": "text", + "content": "[42] Kunming Luo, Chuan Wang, Shuaicheng Liu, Haoqiang Fan, Jue Wang, and Jian Sun. Upflow: Upsampling pyramid for unsupervised optical flow learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1045-1054, 2021. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 526, + 545, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 579 + ], + "type": "text", + "content": "[43] Rémi Marsal, Florian Chabot, Angélique Loesch, and Hichem Sahbi. Brightflow: Brightness-change-aware unsupervised learning of optical flow. In Proceedings of the IEEE winter conference on applications of computer vision, pages 2061-2070, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 581, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 581, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 581, + 545, + 646 + ], + "type": "text", + "content": "[44] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4040-4048, 2016. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "text", + "content": "[45] Simon Meister, Junhwa Hur, and Stefan Roth. Unflow: Unsupervised learning of optical flow with a bidirectional census loss. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 1, 2, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[46] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE Conference" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9575" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, 2015. 2, 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 139 + ], + "type": "text", + "content": "[47] Hyeonwoo Noh, Seunghoon Hong, and Bohyung Han. Learning deconvolution network for semantic segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1520-1528, 2015. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 286, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 286, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 286, + 237 + ], + "type": "text", + "content": "[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, pages 8024-8035. Curran Associates, Inc., 2019. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 239, + 286, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 286, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 286, + 282 + ], + "type": "text", + "content": "[49] Anurag Ranjan and Michael J. Black. Optical flow estimation using a spatial pyramid network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2720-2729, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 283, + 286, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 286, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 286, + 348 + ], + "type": "text", + "content": "[50] Anurag Ranjan, Varun Jampani, Lukas Balles, Kihwan Kim, Deqing Sun, Jonas Wulff, and Michael J Black. Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12240–12249, 2019. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 350, + 286, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 286, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 286, + 413 + ], + "type": "text", + "content": "[51] Hazem Rashed, Senthil Yogamani, Ahmad El-Sallab, Pavel Krizek, and Mohamed El-Helw. Optical flow augmented semantic segmentation networks for automated driving. In International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, pages 165-172, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 415, + 286, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 286, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 286, + 470 + ], + "type": "text", + "content": "[52] Fitsum A Reda, Guilin Liu, Kevin J Shih, Robert Kirby, Jon Barker, David Tarjan, Andrew Tao, and Bryan Catanzaro. Sdc-net: Video prediction using spatially-displaced convolution. In Proceedings of the European Conference on Computer Vision, pages 718-733, 2018. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 471, + 286, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 286, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 286, + 524 + ], + "type": "text", + "content": "[53] Zhile Ren, Orazio Gallo, Deqing Sun, Ming-Hsuan Yang, Erik B Sudderth, and Jan Kautz. A fusion approach for multiframe optical flow estimation. In Proceedings of the IEEE winter conference on applications of computer vision, pages 2077-2086. IEEE, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 526, + 286, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 286, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 286, + 569 + ], + "type": "text", + "content": "[54] Zhile Ren, Deqing Sun, Jan Kautz, and Erik Sudderth. Cascaded scene flow prediction using semantic segmentation. In Proceedings of the International Conference on 3D Vision, pages 225-233. IEEE, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 571, + 286, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 286, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 286, + 613 + ], + "type": "text", + "content": "[55] Zhe Ren, Junchi Yan, Bingbing Ni, Bin Liu, Xiaokang Yang, and Hongyuan Zha. Unsupervised deep learning for optical flow estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 615, + 286, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 615, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 615, + 286, + 656 + ], + "type": "text", + "content": "[56] Stephan R. Richter, Zeeshan Hayder, and Vladlen Koltun. Playing for benchmarks. In Proceedings of the IEEE International Conference on Computer Vision, pages 2232-2241, 2017. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "text", + "content": "[57] Laura Sevilla-Lara, Deqing Sun, Varun Jampani, and Michael J Black. Optical flow with semantic segmentation and localized layers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3889-3898, 2016. 2, 3, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[58] Shihao Shen, Louis Kerofsky, and Senthil Yogamani. Optical flow for autonomous driving: Applications, challenges and improvements. arXiv preprint arXiv:2301.04422, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 108, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 139 + ], + "type": "text", + "content": "[59] Jianbo Shi and Carlo Tomasi. Good features to track. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 593-600, 1994. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 141, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 185 + ], + "type": "text", + "content": "[60] Yukang Shi and Kaiheng Ma. Safit: Segmentation-aware scene flow with improved transformer. In Proceedings of the International Conference on Robotics and Automation, pages 10648-10655. IEEE, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "type": "text", + "content": "[61] Jeongho Shin, Sangjin Kim, Sangkyu Kang, Seong-Won Lee, Joonki Paik, Besma Abidi, and Mongi Abidi. Optical flow-based real-time object tracking using non-prior training active feature model. Real-time imaging, 11(3):204-218, 2005. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 243, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 297 + ], + "type": "text", + "content": "[62] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, volume 11006, pages 369-386. SPIE, 2019. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 299, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 299, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 299, + 545, + 353 + ], + "type": "text", + "content": "[63] Austin Stone, Daniel Maurer, Alper Ayvaci, Anelia Angelova, and Rico Jonschkowski. Smurf: Self-teaching multiframe unsupervised raft with full-image warping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3887-3896, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 355, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 409 + ], + "type": "text", + "content": "[64] Xiuchao Sui, Shaohua Li, Xue Geng, Yan Wu, Xinxing Xu, Yong Liu, Rick Goh, and Hongyuan Zhu. Craft: Cross-attentional flow transformer for robust optical flow. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 17602-17611, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "type": "text", + "content": "[65] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8934-8943, 2018. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 468, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 520 + ], + "type": "text", + "content": "[66] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Models matter, so does training: An empirical study of cnns for optical flow estimation. IEEE transactions on pattern analysis and machine intelligence, 42(6):1408-1423, 2019. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 523, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 555 + ], + "type": "text", + "content": "[67] Andrew Tao, Karan Sapra, and Bryan Catanzaro. Hierarchical multi-scale attention for semantic segmentation. arXiv preprint arXiv:2005.10821, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 557, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 600 + ], + "type": "text", + "content": "[68] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proceedings of the European Conference on Computer Vision, pages 402-419. Springer, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 602, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 657 + ], + "type": "text", + "content": "[69] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Proceedings of the International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[70] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE Con" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9576" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "type": "text", + "content": "ference on Computer Vision and Pattern Recognition Workshops, pages 969-977, 2018. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[71] Panqu Wang, Pengfei Chen, Ye Yuan, Ding Liu, Zehua Huang, Xiaodi Hou, and Garrison Cottrell. Understanding convolution for semantic segmentation. In Proceedings of the IEEE winter conference on applications of computer vision, pages 1451-1460. IEEE, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 286, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 286, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 286, + 206 + ], + "type": "text", + "content": "[72] Yang Wang, Yi Yang, Zhenheng Yang, Liang Zhao, Peng Wang, and Wei Xu. Occlusion aware unsupervised learning of optical flow. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4884-4893, 2018. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 286, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 286, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 286, + 252 + ], + "type": "text", + "content": "[73] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 286, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 286, + 297 + ], + "type": "text", + "content": "[74] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Song Wang, and Lili Ju. Semantic stereo matching with pyramid cost volumes. In Proceedings of the IEEE International Conference on Computer Vision, pages 7484-7493, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 299, + 286, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 286, + 342 + ], + "type": "text", + "content": "[75] Jonas Wulff, Laura Sevilla-Lara, and Michael J Black. Optical flow in mostly rigid scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4671-4680, 2017. 2, 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 344, + 286, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 286, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 286, + 398 + ], + "type": "text", + "content": "[76] Haofei Xu, Jiaolong Yang, Jianfei Cai, Juyong Zhang, and Xin Tong. High-resolution optical flow from 1d attention and correlation. In Proceedings of the IEEE International Conference on Computer Vision, pages 10498-10507, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 286, + 453 + ], + "type": "text", + "content": "[77] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8121-8130, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 286, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 286, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 286, + 510 + ], + "type": "text", + "content": "[78] Guorun Yang, Hengshuang Zhao, Jianping Shi, Zhidong Deng, and Jiaya Jia. Segstereo: Exploiting semantic information for disparity estimation. In Proceedings of the European Conference on Computer Vision, pages 636-651, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 286, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 286, + 567 + ], + "type": "text", + "content": "[79] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. Bisenet v2: Bilateral network with guided aggregation for real-time semantic segmentation. International Journal of Computer Vision, 129:3051-3068, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 286, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 286, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 286, + 621 + ], + "type": "text", + "content": "[80] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceedings of the European Conference on Computer Vision, pages 325-341, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 286, + 689 + ], + "type": "text", + "content": "[81] Jason J. Yu, Adam W. Harley, and Konstantinos G. Derpanis. Back to basics: Unsupervised learning of optical flow via brightness constancy and motion smoothness. In Gang Hua and Hervé Jégou, editors, Proceedings of the European Conference on Computer Vision Workshops, pages 3-10, Cham, 2016. Springer International Publishing. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "text", + "content": "[82] Shuzhi Yu, Hannah H Kim, Shuai Yuan, and Carlo Tomasi. Unsupervised flow refinement near motion boundaries. In" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 485 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "text", + "content": "Proceedings of the British Machine Vision Conference. BMVA Press, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[83] Shuai Yuan, Xian Sun, Hannah Kim, Shuzhi Yu, and Carlo Tomasi. Optical flow training under limited label budget via active learning. In Proceedings of the European Conference on Computer Vision, pages 410-427, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "text", + "content": "[84] Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Common practices and emerging technologies. IEEE access, 8:58443-58469, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "type": "text", + "content": "[85] Feihu Zhang, Oliver J Woodford, Victor Adrian Prisacariu, and Philip HS Torr. Separable flow: Learning motion cost volumes for optical flow estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 10807-10817, 2021. 1, 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 545, + 285 + ], + "type": "text", + "content": "[86] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2881-2890, 2017. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 286, + 545, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 362 + ], + "type": "text", + "content": "[87] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6881-6890, 2021. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 364, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 364, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 308, + 364, + 545, + 418 + ], + "type": "text", + "content": "[88] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In Proceedings of the International Conference on Learning Representations, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 420, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 420, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 308, + 420, + 545, + 485 + ], + "type": "text", + "content": "[89] Yi Zhu, Karan Sapra, Fitsum A Reda, Kevin J Shih, Shawn Newsam, Andrew Tao, and Bryan Catanzaro. Improving semantic segmentation via video propagation and label relaxation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8856-8865, 2019. 2, 3, 6" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9577" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/68ea1ef8-9cb5-44da-894d-92e02632af66_content_list.json b/2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/68ea1ef8-9cb5-44da-894d-92e02632af66_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c63fe776595e3288ace0060fbe157630d4593fe2 --- /dev/null +++ b/2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/68ea1ef8-9cb5-44da-894d-92e02632af66_content_list.json @@ -0,0 +1,2254 @@ +[ + { + "type": "text", + "text": "Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition", + "text_level": 1, + "bbox": [ + 96, + 102, + 870, + 125 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiqing Liang", + "bbox": [ + 114, + 143, + 222, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Eliot Laidlaw", + "bbox": [ + 243, + 143, + 354, + 161 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alexander Meyerowitz Brown University", + "bbox": [ + 377, + 143, + 560, + 180 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Srinath Sridhar", + "bbox": [ + 581, + 143, + 705, + 160 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "James Tompkin", + "bbox": [ + 727, + 143, + 854, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 209, + 313, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "From video, we reconstruct a neural volume that captures time-varying color, density, scene flow, semantics, and attention information. The semantics and attention let us identify salient foreground objects separately from the background across spacetime. To mitigate low resolution semantic and attention features, we compute pyramids that trade detail with whole-image context. After optimization, we perform a saliency-aware clustering to decompose the scene. To evaluate real-world scenes, we annotate object masks in the NVIDIA Dynamic Scene and DyCheck datasets. We demonstrate that this method can decompose dynamic scenes in an unsupervised way with competitive performance to a supervised method, and that it improves foreground/background segmentation over recent static/dynamic split methods. Project webpage: https://visual.cs.brown.edu/saff", + "bbox": [ + 75, + 238, + 473, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 479, + 207, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given a casually-captured monocular RGB video of a dynamic scene, decomposing it into foreground objects and background is an important task in computer vision, with downstream applications in segmentation and video editing. An ideal method would also reconstruct the geometry and appearance over time including frame correspondences.", + "bbox": [ + 75, + 503, + 470, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous methods have made great progress but there are many limitations. Some works assume that there is no object motion in the scene [32, 46, 31, 16, 4], take input from multi-view cameras [16, 46], or do not explicitly reconstruct the underlying 3D structure of the scene [4, 7]. For objects, some works rely on masks or user input to aid segmentation [42, 15, 7], or use task-specific training datasets [7]. Sometimes, works assume the number of foreground objects [4, 20]. Given the challenges, many works train and test on synthetic data [14, 38].", + "bbox": [ + 75, + 597, + 470, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present Semantic Attention Flow Fields (SAFF): A method to overcome these limitations by integrating low-level reconstruction cues with high-level pretrained information—both bottom-up and top-down—into a neural volume. With this, we demonstrate that embedded semantic and saliency (attention) information is useful for unsupervised dynamic scene decomposition. SAFF builds upon neural scene flow fields [18], an approach that reconstructs appearance, geometry, and motion. This uses frame interpolation rather than explicit canonicalization [35] or a latent hyperspace [26],", + "bbox": [ + 75, + 750, + 472, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/01fefb4a9e45aeceefff10089d5bae4b9533e86d532a1f53df7e2cc944385a37.jpg", + "image_caption": [ + "Density $\\sigma$ , Saliency $a$ , Semantics $s$ , Color $c$" + ], + "image_footnote": [], + "bbox": [ + 511, + 198, + 879, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17f14ffcf60fce11f834a50b881e8ff330d777688d5b719bbc157a87104f5d21.jpg", + "image_caption": [ + "Novel view renderings with effects (depth of field, color adjustment)", + "Figure 1: We decompose a dynamic 3D scene from monocular input into time-varying color, density, scene flow, semantics, and attention information. This could be used to provide precise editing of objects or effects focused around objects." + ], + "image_footnote": [], + "bbox": [ + 513, + 382, + 694, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/41eff40600d65b7aa1eeea11fea2b17c3386a06b748cc4ca9b57cb24afc10184.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 382, + 875, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "which lets it more easily apply to casual videos. For optimization, we supervise two network heads with pretrained DINO-ViT [6] semantic features and attention. Naively supervising high-resolution SAFF with low-resolution DINO-ViT output reduces reconstruction quality. To mitigate the mismatch, we build a semantic attention pyramid that trades detail with whole-image context. Having optimized a SAFF representation for a dynamic scene, we perform a saliency-aware clustering both in 3D and on rendered feature images to describe objects and their background. Given the volume reconstruction, the clustering generalizes to novel spacetime views.", + "bbox": [ + 496, + 551, + 893, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To evaluate SAFF's dynamic scene decomposition capacity, we expand the NVIDIA Dynamic Scene [45] and DyCheck [10] datasets by manually annotating object masks across input and hold-out views. We demonstrate that SAFF outperforms 2D DINO-ViT baselines and is comparable to a state-of-the-art video segmentation method ProposeReduce[19] on our data. Existing monocular video dynamic volume reconstruction methods typically separate static and dynamic parts, but these often do not represent meaningful foreground. We show improved foreground segmentation over NSFF and the current $\\mathrm{D}^2\\mathrm{NeRF}$ [40] method for downstream tasks like editing.", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21797", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 89, + 217, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Decomposing a scene into regions of interest is a long-studied task in computer vision [5], including class, instance, and panoptic segmentation in high-level or top-down vision, and use of low-level or bottom-up cues like motion. Recent progress has considered images [12, 4, 20], videos [14, 15], layer decomposition [44, 24], and in-the-wild databases using deep generative models [25]. One example, $\\mathrm{SAVi}++[7]$ , uses slot attention [20] to define 2D objects in real-world videos. Providing first-frame segmentation masks achieves stable performance, with validation on the driving Waymo Open Dataset [33]. Our work attempts 3D scene decomposition for a casual monocular video without initial masks.", + "bbox": [ + 75, + 112, + 470, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scene Decomposition with NeRFs Neural Radiance Fields (NeRF) [41] have spurred new scene decomposition research through volumes. ObSuRF [32] and uORF [46] are unsupervised slot attention works that bind a latent code to each object. Unsupervised decomposition is also possible on light fields [31]. For dynamic scenes, works like NeuralDiff [37] and $\\mathrm{D}^2\\mathrm{NeRF}$ [40] focus on foreground separation, where foreground is defined to contain moving objects. Other works like N3F [36] and occlusions-4d [13] also decompose foregrounds into individual objects. N3F requires user input to specify which object to segment, and occlusions-4d takes RGB point clouds as input. Our work attempts to recover a segmented dynamic scene from a monocular RGB video without added markup or masks.", + "bbox": [ + 75, + 310, + 470, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Fields Beyond Radiance Research has begun to add non-color information to neural volumes to aid decomposition via additional feature heads on the MLP. iLabel [47] adds a semantic head to propagate user-provided segmentations in the volume. PNF [17] and Panoptic-NeRF [9] attempt panoptic segmentation within neural fields, and Object-NeRF integrates instance segmentation masks into the field during optimization [42]. Research also investigates how to apply generic pretrained features to neural fields, like DINO-ViT.", + "bbox": [ + 75, + 537, + 468, + 675 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DINO-ViT for Semantics and Saliency DINO-ViT is a self-supervised transformer that, after pretraining, extracts generic semantic information [6]. Amir et al. [1] use DINO-ViT features with $k$ -means clustering to achieve co-segmentation across a video. Seitzer et al. [29] combine slot attention and DINO-ViT features for object-centric learning on real-world 2D data. TokenCut [39] performs normalized cuts on DINO-ViT features for foreground segmentation on natural images. Deep Spectral Segmentation [22] show that graph Laplacian processing of DINO-ViT features provides unsupervised foreground segmentation, and Selfmask [30] shows that these features can provide object saliency masks. Our approach considers these clustering and saliency findings for the setting of 3D decomposition from monocular video.", + "bbox": [ + 75, + 691, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DINO-ViT Fields Concurrent works have integrated DINO-ViT features into neural fields. DFF [16] distills features for dense multi-view static scenes with user input for segmentation. N3F [36] expands NeuralDiff to dynamic scenes, and relies on user input for segmentation. AutoLabel [3] uses DINO-ViT features to accelerate segmentation in static scenes given a ground truth segmentation mask. Other works use DINO-ViT differently. FeatureRealistic-Fusion [21] uses DINO-ViT in an online feature fusion task, focusing on propagating user input, and NeRF-SOS [8] uses 2D DINO-ViT to process NeRF-rendered multi-view RGB images of a static scene. In contrast to these works, we consider real-world casual monocular videos, recover and decompose a 3D scene, then explore whether saliency can avoid the need for masks or user input in segmenting objects.", + "bbox": [ + 496, + 90, + 893, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 329, + 589, + 344 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For a baseline dynamic scene reconstruction method, we begin with NSFF from Li et al. [18] (Sec. 3.1), which builds upon NeRF [23]. NSFF's low-level scene flow frame-to-frame approach provides better reconstructions for real-world casual monocular videos than deformation-based methods [35, 26]. We modify the architecture to integrate higher-level semantic and saliency (or attention) features (Sec. 3.2). After optimizing a SAFF for each scene, we perform saliency-aware clustering of the field (Sec. 3.4). All implementation details are in our supplemental material.", + "bbox": [ + 496, + 349, + 893, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Input Our method takes in a single RGB video over time $i$ as an ordered set of images $I \\in \\mathcal{I}$ and camera poses. We use COLMAP to recover camera poses [28]. From all poses, we define an NDC-like space that bounds the scene, and a set of rays $\\mathbf{r} \\in \\mathcal{R}$ , one per image pixel with color $\\hat{\\mathbf{c}}^\\dagger$ . Here, $\\hat{\\mathbf{c}}$ denotes a 2D pixel value in contrast to a 3D field value, and $\\hat{\\mathbf{c}}^\\dagger$ denotes an input value in contrast to an estimated value.", + "bbox": [ + 496, + 512, + 893, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From pretrained networks, we estimate single-frame monocular depth $\\hat{d}^{\\dagger}$ (MiDaSv2 [27]), optical flow $\\hat{\\mathbf{p}}_i^\\dagger$ (RAFT [34]), and semantic features $\\hat{\\mathbf{s}}^\\dagger$ and attention $\\hat{\\mathbf{a}}^\\dagger$ (DINO-ViT [6]) after important preprocessing (Sec. 3.3).", + "bbox": [ + 496, + 619, + 890, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Initial Dynamic Neural Volume Representation", + "text_level": 1, + "bbox": [ + 498, + 686, + 890, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The initial representation comprises a static NeRF $F_{\\theta}^{\\mathrm{st}}$ and a dynamic NeRF $F_{\\theta}^{\\mathrm{dy}}$ . The static network predicts a color $\\mathbf{c}$ , density $\\sigma$ , and blending weight $v$ (Eq. (1)), and the dynamic network predicts time-varying color $\\mathbf{c}_i$ , density $\\sigma_i$ , scene flow $\\mathbf{f}_i$ , and occlusion weights $w_i$ (Eq. (2)). In both network architectures, other than position $\\mathbf{x}$ , we add direction $\\boldsymbol{\\omega}$ to a late separate head to only condition the estimation of color $\\mathbf{c}$ .", + "bbox": [ + 496, + 707, + 893, + 814 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\boldsymbol {\\theta}} ^ {\\mathrm {s t}}: (\\mathbf {x}, \\omega) \\rightarrow \\left(\\mathbf {c} ^ {\\mathrm {s t}}, \\sigma^ {\\mathrm {s t}}, v\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 823, + 890, + 840 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\boldsymbol {\\theta}} ^ {\\mathrm {d y}}: (\\mathbf {x}, \\boldsymbol {\\omega}, i) \\rightarrow \\left(\\mathbf {c} _ {i} ^ {\\mathrm {d y}}, \\sigma_ {i} ^ {\\mathrm {d y}}, \\mathbf {f} _ {i}, w _ {i}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 843, + 890, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To produce a pixel's color, we sample points at distances $t$ along the ray $\\mathbf{x}_t = \\mathbf{x} - \\omega t$ between near and far planes $t_n$", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21798", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f11cce526596a99c53cb277dfee5c4ba6fe1f8e2e53d703a83c76fedd620f2ad.jpg", + "image_caption": [ + "(a) SAFF video inputs" + ], + "image_footnote": [], + "bbox": [ + 86, + 106, + 349, + 303 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/08f48acf6d9bc6c16bccd6de7bb8553c7cb55375f61b01b26ae38212d33b08ce.jpg", + "image_caption": [ + "(b) SAFF model" + ], + "image_footnote": [], + "bbox": [ + 375, + 112, + 573, + 300 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8f5ce03d1d46c9edbba3fab841f598fd38e818f838815b8436f0ed4d1f04ab7e.jpg", + "image_caption": [ + "(c) Volume rendered outputs", + "Figure 2: Overview. From monocular video, SAFF builds a neural field of scene-flowed 3D density, radiance, semantics, and attention (b). This is guided by semantic attention pyramids that increase resolution, plus depth and optical flow priors (a). We can render new spacetime views of any channel, and use saliency-aware clustering to decompose objects and background (c)." + ], + "image_footnote": [], + "bbox": [ + 598, + 108, + 879, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to $t_f$ , query each network, then integrate transmittance $T$ , density $\\sigma$ , and color along the ray from these samples [23]. For brevity, we omit evaluating at $\\mathbf{x}_t, \\boldsymbol{\\omega}$ , e.g., $\\sigma^{\\mathrm{st}}(\\mathbf{x}_t)$ is simply $\\sigma^{\\mathrm{st}}$ ; $\\mathbf{c}^{\\mathrm{st}}(\\mathbf{x}_t, \\boldsymbol{\\omega})$ is simply $\\mathbf{c}^{\\mathrm{st}}$ .", + "bbox": [ + 75, + 364, + 470, + 424 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We produce a combined color from the static and dynamic colors by multiplication with their densities:", + "bbox": [ + 76, + 424, + 468, + 454 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {i} \\mathbf {c} _ {i} = v \\sigma^ {\\mathrm {s t}} \\mathbf {c} ^ {\\mathrm {s t}} + (1 - v) \\sigma_ {i} ^ {\\mathrm {d y}} \\mathbf {c} _ {i} ^ {\\mathrm {d y}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 460, + 468, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given that transmittance integrates density up to the current sampled point under Beer-Lambert volume attenuation, the rendered pixel color for the ray is computed as:", + "bbox": [ + 76, + 491, + 468, + 537 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {c}} _ {i} = \\int_ {t _ {n}} ^ {t _ {f}} T _ {i} \\sigma_ {i} \\mathbf {c} _ {i} d t \\text {w h e r e} T _ {i} = \\exp \\left(- \\int_ {t _ {n}} ^ {t} \\sigma_ {i} d t\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 542, + 468, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To optimize the volume to reconstruct input images, we compute a photometric loss $L_{\\hat{\\mathbf{c}}}$ between rendered and input colors:", + "bbox": [ + 75, + 589, + 468, + 633 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\hat {\\mathbf {c}}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\left\\| \\hat {\\mathbf {c}} _ {i} \\left(\\mathbf {r} _ {i}\\right) - \\hat {\\mathbf {c}} _ {i} ^ {\\dagger} \\left(\\mathbf {r} _ {i}\\right) \\right\\| _ {2} ^ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 642, + 468, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scene Flow Defining correspondence over time is important for monocular input as it lets us penalize a reprojection error with neighboring frames $j \\in \\mathcal{N}$ , e.g., where $j = i + 1$ or $j = i - 1$ . We denote $i \\rightarrow j$ for the projection of frame $i$ onto frame $j$ by scene flow. $F_{\\theta}^{\\mathrm{dy}}$ estimates both forwards and backwards scene flow at every point to penalize a bi-direction loss.", + "bbox": [ + 75, + 693, + 468, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thus, we approximate color output at time step $j$ by flowing the queried network values at time step $i$ :", + "bbox": [ + 76, + 786, + 468, + 816 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {c}} _ {i \\rightarrow j} = \\int_ {t _ {n}} ^ {t _ {f}} T _ {i \\rightarrow j} \\sigma_ {i \\rightarrow j} \\mathbf {c} _ {i \\rightarrow j} d t \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 825, + 468, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reprojection must account for occlusion and disocclusion by motion. As such, $F_{\\theta}^{\\mathrm{dy}}$ also predicts forwards and backwards", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "scene occlusion weights $w_{i + 1}$ and $w_{i - 1} \\in [0,1]$ , where a point with $w_{i + 1} = 0$ means that occlusion status is changed one step forwards in time. We can integrate $w$ to a pixel:", + "bbox": [ + 496, + 364, + 892, + 410 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {w} _ {i \\rightarrow j} = \\int_ {t _ {n}} ^ {t _ {f}} T _ {i \\rightarrow j} \\sigma_ {i \\rightarrow j} w _ {j} d t \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 415, + 892, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, this pixel weight modulates the color loss such that occluded pixels are ignored. (8)", + "bbox": [ + 496, + 457, + 892, + 491 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\hat {\\mathbf {c}} _ {i \\rightarrow j}} = \\frac {1}{| \\mathcal {R} | | \\mathcal {N} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\sum_ {j \\in \\mathcal {N}} \\hat {w} _ {i \\rightarrow j} (\\mathbf {r} _ {i}) | | \\hat {\\mathbf {c}} _ {i \\rightarrow j} (\\mathbf {r} _ {i}) - \\hat {\\mathbf {c}} _ {j} ^ {\\dagger} (\\mathbf {r} _ {j}) | | _ {2} ^ {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 493, + 892, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prior losses We use the pretrained depth and optical flow map losses to help overcome the ill-posed monocular reconstruction problem. These losses decay as optimization progresses to rely more and more on the optimized self-consistent geometry and scene flow. For geometry, we estimate a depth $\\hat{d}_i$ for each ray $\\mathbf{r}_i$ by replacing $\\mathbf{c}_i$ in Eq. (4) by the distance $t$ along the ray. Transform $z$ estimates a scale and shift as the pretrained network produces only relative depth.", + "bbox": [ + 496, + 545, + 893, + 667 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\hat {d}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\left| \\left| \\hat {d} _ {i} - z \\left(\\hat {d} _ {i} ^ {\\dagger}\\right) \\right| \\right| _ {1} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 670, + 892, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For motion, projecting scene flow to a camera lets us compare to the estimated optical flow. Each sample point along a ray $\\mathbf{x}_i$ is advected to a point in the neighboring frame $\\mathbf{x}_{i\\rightarrow j}$ , then integrated to the neighboring camera plane to produce a 2D point offset $\\hat{\\mathbf{p}}_i(\\mathbf{r}_i)$ . Then, we expect the difference in the start and end positions to match the prior:", + "bbox": [ + 496, + 714, + 892, + 806 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\hat {\\mathbf {p}}} = \\frac {1}{| \\mathcal {R} | | \\mathcal {N} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\sum_ {j \\in \\mathcal {N} (i)} | | \\hat {\\mathbf {p}} _ {i} (\\mathbf {r} _ {i}) - \\hat {\\mathbf {p}} _ {i} ^ {\\dagger} (\\mathbf {r} _ {i}) | | _ {1} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 809, + 892, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additional regularizations encourage occlusion weights to be close to one, scene flow to be small, locally constant, and cyclically consistent, and blending weight $v$ to be sparse.", + "bbox": [ + 496, + 854, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21799", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Semantic Attention Flow Fields", + "text_level": 1, + "bbox": [ + 76, + 90, + 354, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Beyond low-level or bottom-up features, high-level or top-down features are also useful to define objects and help down-stream tasks like segmentation. For example, methods like NSFF or $\\mathrm{D}^2\\mathrm{NeRF}$ struggle to provide useful separation of static and dynamic parts because blend weight $v$ estimates whether the volume appears to be occupied by some moving entity. This is not the same as objectness; tasks like video editing could benefit from accurate dynamic object masks.", + "bbox": [ + 75, + 113, + 468, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As such, we extract 2D semantic features and attention (or saliency) values from a pretrained DINO-ViT network, then optimize the SAFF such that unknown 3D semantic and attention features over time can be projected to recreate their 2D complements. This helps us to ascribe semantic meaning to the volume and to identify objects. As semantics/attention are integrated into the 4D volume, we can render them from novel spacetime views without further DINO-ViT computation.", + "bbox": [ + 75, + 234, + 470, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To estimate semantic features $\\mathbf{s}$ and attention $\\mathbf{a}$ at 3D points in the volume at time $i$ , we add two new heads to both the static $F_{\\theta}^{\\mathrm{st}}$ and the dynamic $F_{\\theta}^{\\mathrm{dy}}$ networks:", + "bbox": [ + 76, + 354, + 468, + 402 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\boldsymbol {\\theta}} ^ {\\mathrm {s t}}: (\\mathbf {x}, \\omega) \\quad \\rightarrow \\left(\\mathbf {c} ^ {\\mathrm {s t}}, \\sigma^ {\\mathrm {s t}}, v, \\mathbf {s} ^ {\\mathrm {s t}}, \\mathbf {a} ^ {\\mathrm {s t}}\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 407, + 468, + 425 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\boldsymbol {\\theta}} ^ {\\mathrm {d y}}: (\\mathbf {x}, \\boldsymbol {\\omega}, i) \\rightarrow \\left(\\mathbf {c} _ {i} ^ {\\mathrm {d y}}, \\sigma_ {i} ^ {\\mathrm {d y}}, \\mathbf {f} _ {i}, w _ {i}, \\mathbf {s} _ {i} ^ {\\mathrm {d y}}, \\mathbf {a} _ {i} ^ {\\mathrm {d y}}\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 428, + 468, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As semantic features have been demonstrated to be somewhat robust to view dependence [1], in our architectures both heads for s, a are taken off the backbone before $\\omega$ is injected.", + "bbox": [ + 76, + 454, + 468, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To render semantics from the volume, we replace the color term $\\mathbf{c}$ in Eq. (4) with $\\mathbf{s}$ and equivalently for $\\mathbf{a}$ :", + "bbox": [ + 76, + 500, + 468, + 529 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {i} \\mathbf {s} _ {i} = v \\sigma^ {\\mathrm {s t}} \\mathbf {s} ^ {\\mathrm {s t}} + (1 - v) \\sigma_ {i} ^ {\\mathrm {d y}} \\mathbf {s} _ {i} ^ {\\mathrm {d y}}, \\hat {\\mathbf {s}} _ {i} = \\int_ {t _ {n}} ^ {t _ {f}} T _ {i} \\sigma_ {i} \\mathbf {s} _ {i} d t \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 534, + 468, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To encourage scene flow to respect semantics over time, we penalize complementary losses on s and a (showing s only):", + "bbox": [ + 76, + 575, + 468, + 604 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\hat {\\mathbf {s}} _ {i \\rightarrow j}} = \\frac {1}{| \\mathcal {R} | | \\mathcal {N} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\sum_ {j \\in \\mathcal {N}} \\hat {w} _ {i \\rightarrow j} (\\mathbf {r} _ {i}) \\| \\hat {\\mathbf {s}} _ {i \\rightarrow j} (\\mathbf {r} _ {i}) - \\hat {\\mathbf {s}} _ {j} ^ {\\dagger} (\\mathbf {r} _ {j}) \\| _ {2} ^ {2} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 607, + 468, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, as supervision, we add respective losses on the reconstruction of the 2D semantic and attention features from projected 3D volume points (showing s only):", + "bbox": [ + 76, + 660, + 468, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\hat {\\mathbf {s}}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\left\\| \\hat {\\mathbf {s}} _ {i} \\left(\\mathbf {r} _ {i}\\right) - \\hat {\\mathbf {s}} _ {i} ^ {\\dagger} \\left(\\mathbf {r} _ {i}\\right) \\right\\| _ {2} ^ {2} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 708, + 468, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike depth and scene flow priors, these are not priors—there is no self-consistency for semantics to constrain their values. Thus, we do not decay their contribution. While decaying avoids disagreements between semantic and attention features and color-enforced scene geometry, it also leads to a loss of useful meaning (please see supplemental).", + "bbox": [ + 76, + 751, + 470, + 842 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus our final loss becomes:", + "bbox": [ + 96, + 842, + 284, + 856 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {S A F F}} = \\mathcal {L} _ {\\hat {\\mathbf {c}}} + \\lambda_ {\\hat {\\mathbf {c}} _ {i \\rightarrow j}} \\mathcal {L} _ {\\hat {\\mathbf {c}} _ {i \\rightarrow j}} + \\lambda_ {\\hat {d}} \\mathcal {L} _ {\\hat {d}} + \\lambda_ {\\hat {\\mathbf {p}}} \\mathcal {L} _ {\\hat {\\mathbf {p}}} \\tag {16} \\\\ + \\lambda_ {\\hat {\\mathbf {s}} _ {i \\rightarrow j}} \\mathcal {L} _ {\\hat {\\mathbf {s}} _ {i \\rightarrow j}} + \\lambda_ {\\hat {\\mathbf {a}} _ {i \\rightarrow j}} \\mathcal {L} _ {\\hat {\\mathbf {a}} _ {i \\rightarrow j}} + \\lambda_ {\\hat {\\mathbf {s}}} \\mathcal {L} _ {\\hat {\\mathbf {s}}} + \\lambda_ {\\hat {\\mathbf {a}}} \\mathcal {L} _ {\\hat {\\mathbf {a}}} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 862, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Semantic Attention Pyramids", + "text_level": 1, + "bbox": [ + 500, + 90, + 763, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When thinking about scenes, we might argue that semantics from an ideal extractor should be scale invariant, as distant objects have the same class as close objects. We might also argue that saliency (or attention features) may not be scale invariant, as small details in a scene should only be salient when viewed close up. In practice, both extracted features vary across scale and have limited resolution, e.g., DINO-ViT [6] produces one output for each $8 \\times 8$ patch. But, from this, we want semantic features and saliency for every RGB pixel that still respects scene boundaries.", + "bbox": [ + 496, + 113, + 893, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus far, work on static scenes has ignored the input/feature resolution mismatch [16] as multi-view constraints provide improved localization within the volume. For monocular video, this approach has limitations [36]. Forming many constraints on dynamic objects requires long-term motion correspondence—a tricky task—and so we want to maximize the resolution of any input features where possible without changing their meaning.", + "bbox": [ + 496, + 265, + 893, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One way may be through a pyramid of semantic and attention features that uses a sliding window approach at finer resolutions. Averaging features could increase detail around edges, but we must overcome the practical limit that these features are not stable across scales. This is especially important for saliency: unlike typical RGB pyramids that must preserve energy in an alias-free way [2], saliency changes significantly over scales and does not preserve energy.", + "bbox": [ + 496, + 385, + 893, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consider a feature pyramid $\\mathcal{P}$ with loss weights per level:", + "bbox": [ + 517, + 506, + 892, + 521 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {P} \\hat {\\mathbf {s}}} = \\sum_ {i \\in \\mathcal {P}} \\lambda_ {\\hat {\\mathbf {s}}} ^ {i} \\mathcal {L} _ {\\hat {\\mathbf {s}}} ^ {i} \\quad \\mathcal {L} _ {\\mathcal {P} \\hat {\\mathbf {a}}} = \\sum_ {i \\in \\mathcal {P}} \\lambda_ {\\hat {\\mathbf {a}}} ^ {i} \\mathcal {L} _ {\\hat {\\mathbf {a}}} ^ {i} \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 541, + 892, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Naively encouraging scale-consistent semantics and whole-image saliency, e.g., $\\lambda_{\\hat{\\mathbf{s}}} = \\{1/3, 1/3, 1/3\\}$ with $\\lambda_{\\hat{\\mathbf{a}}} = \\{1, 0, 0\\}$ , empirically leads to poor recovered object edges because the balanced semantics and coarse saliency compete over where the underlying geometry is. Instead, we weight both equally $\\lambda_{\\hat{\\mathbf{s}}} = \\lambda_{\\hat{\\mathbf{a}}} = \\{1/9, 4/9, 4/9\\}$ . Even though the coarse layer has smaller weight, it is sufficient to guide the overall result. This balances high resolution edges from fine layers and whole object features from coarse layers while reducing geometry conflicts, and leads to improved features (Fig. 3).", + "bbox": [ + 496, + 583, + 893, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Of course, any sliding window must contain an object to extract reliable features for that object. At coarse levels, an object is always in view. At fine levels, an object is only captured in some windows. Objects of interest tend to be near the middle of the frame, meaning that boundary windows at finer pyramid levels contain features that less reliably capture those objects. This can cause spurious connections in clustering. To cope with this, we relatively decrease finer level boundary window weights: We upsample all levels to the finest level, then increase the coarsest level weight towards the frame boundary to $\\lambda_{\\hat{\\mathbf{s}}} = \\lambda_{\\hat{\\mathbf{a}}} = \\{1/3, 1/3, 1/3\\}$ .", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21800", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f178d6f08ab968f5f60529102575cf05b82ce1c9023a855566a217b585eb38d3.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 109, + 99, + 258, + 174 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9b0a274bb513cdbaff27d2e781ba72fa72821d8b34d8ed9cb154a712427de7bd.jpg", + "image_caption": [ + "(b) Semantics" + ], + "image_footnote": [], + "bbox": [ + 259, + 99, + 408, + 174 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ca4d6cc4b08415640f4ab2ce014bf87ff955f0bdf83c5692ffb80dd75522697d.jpg", + "image_caption": [ + "(c) Semantics (volume)" + ], + "image_footnote": [], + "bbox": [ + 410, + 99, + 558, + 174 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/286f833094b4816d0f26ea379f6c302a9a40006555badeca5a9734de5a57cb54.jpg", + "image_caption": [ + "(d) Saliency" + ], + "image_footnote": [], + "bbox": [ + 560, + 99, + 707, + 174 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6073298e54a3d420c359a7e00e8a065e7a90a15f4532e6dd95c82ee33325699a.jpg", + "image_caption": [ + "(e) Saliency (volume)" + ], + "image_footnote": [], + "bbox": [ + 709, + 99, + 856, + 174 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9a772ca97f6d34a23e449345bc01ca9b625edcc4b2d9d0c0e4c5afca1890d8ad.jpg", + "image_caption": [ + "(f) Pyr. semantics" + ], + "image_footnote": [], + "bbox": [ + 259, + 186, + 408, + 261 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b1d357ef4844ecadfd229b10f4610f2f43a0e7ebec2c3b41c59afbd833772873.jpg", + "image_caption": [ + "(g) Our pyr. semantics (volume)", + "Figure 3: Semantics and saliency improve by both volume integration and by our pyramid. On Balloon NBoard, blocky artifacts are removed. Semantics are visualized as most significant three PCA dimensions; specific colors are less meaningful." + ], + "image_footnote": [], + "bbox": [ + 410, + 186, + 557, + 261 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/db2932d48e70bef0de273f03533759cd8762fef47e784615bda46b636c792bcb.jpg", + "image_caption": [ + "(h) Pyr. saliency" + ], + "image_footnote": [], + "bbox": [ + 560, + 186, + 707, + 261 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d9cad011c3b648a0848faa5ed6182fbf8a504b14c318a601d52790f011237b06.jpg", + "image_caption": [ + "(i) Our pyr. saliency (volume)" + ], + "image_footnote": [], + "bbox": [ + 710, + 186, + 856, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Using SAFF for Saliency-aware Clustering", + "text_level": 1, + "bbox": [ + 76, + 314, + 439, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We now wish to isolate salient objects. Even in dynamic scenes, relevant objects may not move, so analyzing dynamic elements is insufficient (cf. [40]). One approach predicts segmentation end-to-end [19]. However, end-to-end learning requires priors about the scene provided by supervision, and even large-scale pretraining might fail given unseen test scenes. To achieve scene-specific decompositions from per-video semantics, inspired by Amir et al. [1], we design clustering for spacetime volumes that allows segmenting novel spatio-temporal views. While DINO-ViT is trained on images, its features are loosely temporally consistent [6].", + "bbox": [ + 75, + 338, + 468, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Some works optimize a representation with a fixed number of clusters, e.g., via slot attention [20] in NeRFs [32, 46]. Instead, we cluster using elbow $k$ -means, letting us adaptively find the number of clusters after optimization. This is more flexible than baking an anticipated number of slots (sometimes with fixed semantics), and lets us cluster and segment in novel spatio-temporal viewpoints.", + "bbox": [ + 75, + 506, + 470, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We demonstrate clustering results in both 3D over time and on rendered volume 2D projections over time. Given the volume reconstruction, we might think that clustering directly in 3D would be better. But, monocular input with narrow baselines makes it challenging to precisely reconstruct geometries: Consider that depth integrated a long a ray can still be accurate even though geometry at specific 3D points may be inaccurate or 'fluffy'. As such, we use the 2D volume projection clustering results in 2D comparisons.", + "bbox": [ + 75, + 612, + 468, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Method For 3D over time, we sample points from the SAFF uniformly along input rays $(128\\times H\\times W)$ and treat each pixel as a feature point, e.g., semantics are 64 dim. and saliency is 1 dim. For volume projection, we render to $N$ input poses $(N\\times H\\times W)$ , and treat each pixel as a feature point instead. In either case, we cluster all feature points together using elbow $k$ -means to produce an initial set of separate regions. For each cluster $c$ , for each image, we calculate the mean attention of all feature points within the cluster $\\bar{\\mathbf{a}}_c$ . If $\\bar{\\mathbf{a}}_c > 0.07$ , then this", + "bbox": [ + 75, + 763, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cluster is salient for this image. Finally, all feature points vote on saliency: if more than $70\\%$ agree, the cluster is salient.", + "bbox": [ + 496, + 315, + 890, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Salient objects may still be split into semantic parts: e.g., in Fig. 4, the person's head/body are separated. Plus, unwanted background saliency may exist, e.g., input $\\hat{\\mathbf{a}}^\\dagger$ is high for the teal graphic on the wall. As such, before saliency voting, we merge clusters whose centroids have a cosine similarity $>0.5$ . This reduces the first problem as heads and bodies are similar, and reduces the second problem as merging the graphic cluster into the background reduces its average saliency (Fig. 4).", + "bbox": [ + 496, + 345, + 892, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To extract an object from the 3D volume, we sample 3D points along each input ray, then ascribe the label from the semantically-closest centroid. All clusters not similar to the stored salient clusters are marked as background with zero density. For novel space-time views, we render feature images from the volume, then assign cluster labels to each pixel according to its similarity with stored input view centroids.", + "bbox": [ + 496, + 467, + 893, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 587, + 630, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We show the impact of adding semantic and saliency features through scene decomposition and foreground experiments. Our website contains supplemental videos.", + "bbox": [ + 496, + 613, + 890, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data: NVIDIA Dynamic Scene Dataset (Masked) This data [45] has 8 sequences of 24 time steps formed from 12 cameras simultaneously capturing video. We manually annotate object masks for view and time step splits; we will release this data publicly. We define three data splits per sequence:", + "bbox": [ + 496, + 670, + 893, + 747 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Input: A monocular camera that moves position for every timestep is simulated from the input sequences; we use Yoon et al.'s input sequences [45].", + "2. Fix Cam 0 (hold out): We fix the camera at position 0 as time plays, requiring novel view and time synthesis. $\\{(\\mathrm{cam}_0,\\mathrm{time}_i),i\\in [1,2,\\dots 23]\\}$", + "3. Fix Time 0 (hold out): We fix time at step 0 as the camera moves, requiring novel view and time synthesis. $\\{(\\mathrm{cam}_i,\\mathrm{time}_0),i\\in [1,2,\\ldots 11]\\}$" + ], + "bbox": [ + 500, + 752, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21801", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0916a3dac95539f8c91aff80d1f8a07ab4e1ba67931550ee0b2cae13ffc89eda.jpg", + "image_caption": [ + "(i) Input frame", + "(ii) Input semantic/attention" + ], + "image_footnote": [], + "bbox": [ + 104, + 99, + 251, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c5ce24e922c13f4821e3240ca0f0021f73e48b716f94b30e793395bc9e3a343b.jpg", + "image_caption": [ + "(a) Our pyr. semantic (volume)", + "(d) Our pyr. saliency (volume)" + ], + "image_footnote": [], + "bbox": [ + 253, + 99, + 401, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5d567c20160ac212f558b1a90c90f18dce7c52124b47c9276e2fdd6d4ca11381.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 89, + 563, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f462e2415f3c4072ff3dd65adef4b2717184e0446ea7a0119253bc18cc96ab9d.jpg", + "image_caption": [ + "(e) Our merged clustering", + "Figure 4: Saliency-aware clustering improves decomposition. On Dynamic Face, the head and body are semantically and saliently different, but are mutually different from the background. This allows us to extract a time-varying 3D field of the object." + ], + "image_footnote": [], + "bbox": [ + 410, + 184, + 558, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/83fe39f2429a8082a84ae5b94ad37683b9859938142cb82fa7d83523de749201.jpg", + "image_caption": [ + "(c) Salient clusters (pre-merge)", + "(f) Our final foreground", + "(iii) Extraction (fixed time)" + ], + "image_footnote": [], + "bbox": [ + 566, + 99, + 715, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5cc523c6ad484004e2d5657e5123e5dbf9683ce2a9478a19116031c1ea14a0d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 185, + 715, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5cbf3a9ab54ff469420060cb15eb9ca62dcd439ff5e73fb19525cc2d36915a10.jpg", + "image_caption": [ + "(iv) Extraction (fixed view)" + ], + "image_footnote": [], + "bbox": [ + 718, + 119, + 866, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d2014859e5f8596c4f049fd7a5542a994321f44e6e401f6f045a53168d008f72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 190, + 866, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e20bdacede3a7a5868b7cef4fcdcf593b23fa2655dfd89336a1d4e0a92985e35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 227, + 866, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/09c1891a830a23fbc620a22ac5e9c58856b78f5efaea2cf8f1da9005971890f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 301, + 867, + 378 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7def0160a5af8dfbc2da02f54623763411d8065e9bb925784ad0a2b1dd18007b.jpg", + "image_caption": [ + "(a) Input", + "Figure 5: SAFF object segmentations show balanced quality while recovering a volumetric scene representation (e). Basic DINO-ViT produces low-quality segmentations and misses objects. A state-of-the-art 2D video learning method [19] sometimes has edge detail (Umbrella, legs) but othertimes misses detail and objects (Balloon NBoard)." + ], + "image_footnote": [], + "bbox": [ + 96, + 316, + 261, + 508 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b13ac5c9c77ea60444552e5524c3e891235297addb0747db1403a807f583b6e.jpg", + "image_caption": [ + "(b) DINO-ViT (2D)" + ], + "image_footnote": [], + "bbox": [ + 264, + 316, + 411, + 508 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6e89c0f7d9b9ee8d30bfbc5bec0ae618274ab0497d01755188af532b00ba8da6.jpg", + "image_caption": [ + "(c) ProposeReduce [19]" + ], + "image_footnote": [], + "bbox": [ + 416, + 316, + 563, + 508 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a9ed21c21fb0dcc4c27f0d7b78c66d1cf4381b857645386272644700ab7fdba2.jpg", + "image_caption": [ + "(d) SAFF (Ours; input)" + ], + "image_footnote": [], + "bbox": [ + 568, + 316, + 715, + 508 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8810bee4ce7e1f93f0f11ff2df60cc4844cb666696cdb3af51a551bec96b38da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 383, + 866, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data: DyCheck Dataset (Masked) For single-camera casual monocular data, we select a subset of Gao et al.'s DyCheck dataset [10]. We remove scenes without an obvious subject, e.g., where telling foreground from background is hard even for a human. We select haru-sit, space-out, spin, paper-windmill, and mochi-high-five. We uniformly sample 48 frames and manually annotate object masks. We take even frames as the input set and odd frames as the hold out set. We use the same hyperparameters between both datasets.", + "bbox": [ + 75, + 587, + 468, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics To assess clustering performance, we use the Adjusted Rand Index (ARI; $[-1, 1]$ ). This compares the similarity of two assignments without label matching, where random assignment would score $\\approx 0$ . For foreground segmentation, we compute IoU (Jaccard), and for RGB quality we use PSNR, SSIM, and LPIPS.", + "bbox": [ + 75, + 731, + 468, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparisons including ablations", + "text_level": 1, + "bbox": [ + 76, + 832, + 366, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SAFF (ours) We optimize upon the input split of each scene, and perform clustering to obtain object segmentations. To produce a foreground, we merge all salient objects.", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SAFF (3D) The same as above, but processed in 3D rather than on 2D volume projections.", + "bbox": [ + 498, + 587, + 890, + 618 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $\\mathbf{w} / \\mathbf{pyr}\\lambda_{\\hat{\\mathbf{a}}} = \\{1,0,0\\}$ Pyramid with only coarse saliency (Sec. 3.3) and balanced semantic weight across levels.", + "—w/o pyr No pyramid (Sec. 3.3); we optimize with features and saliency extracted from the input image only.", + "— w/o merge With pyramid, but we remove cluster merging inside the saliency-aware clustering algorithm.", + "- w/blend $v$ To compare generic dynamic segmentation to saliency segmentation, we use the static/dynamic weight instead of volume saliency to segment foreground objects. We set every pixel below the $80\\%$ $v$ quantile in each image to be background, or otherwise foreground.", + "— w/post process We add a step after the saliency-aware clustering to refine edges using a conditional random field (please see supplemental material for details). This gains significantly from the depth estimated via volume reconstruction, producing sharp and detailed edges." + ], + "bbox": [ + 496, + 625, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21802", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/36a63135f22edbeb9eb8b71e22f73a3ba4d8b599837e8409cfe50dd8f4c40fc9.jpg", + "image_caption": [ + "(a) Input", + "Figure 6: Saliency improves foreground segmentation. Static/dynamic separations are not foreground segmentations, leading to limited use of dynamic NeRF models for downstream tasks. Minor improvements to dynamic blending weight $v$ are seen in some sequences (Jumping) by adding the saliency head to the shared backbone." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 259, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2cc18da6cca93a8b5c3df31bac7ab4722ea3c3d49eb0f8b46f3dc1a1f05dfb45.jpg", + "image_caption": [ + "(b) NSFF [18] blend $v$ (c) $\\mathrm{D}^2\\mathrm{NeRF}$ [40] blend $v$" + ], + "image_footnote": [], + "bbox": [ + 261, + 90, + 411, + 222 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "c) $\\mathrm{D}^2\\mathrm{NeRF}$ [40] blend $v$", + "bbox": [ + 416, + 75, + 573, + 89 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/21cc1dd0431ac7f807f9d4cb3470d324905107b22fdef2f00b6fcc51388a164d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 90, + 566, + 151 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e957d9085e676a91bfd7c9c567126cb875382a76823f9a1b0396fca531b832e.jpg", + "image_caption": [ + "(d) SAFF blend $v$" + ], + "image_footnote": [], + "bbox": [ + 573, + 90, + 720, + 151 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2fa19228ccb41766b7bf0778ab8bac6cf30332c73aef241eaefc80e232489399.jpg", + "image_caption": [ + "(e) SAFF foregrounds" + ], + "image_footnote": [], + "bbox": [ + 723, + 90, + 870, + 151 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "NSFF [18] This method cannot produce semantic clusterings. While saliency and blend weight $v$ have different meanings, if we compare our $v$ to NSFF's, then we can see any impact of a shared backbone with attention heads upon the static/dynamic separation. We disable the supervised motion mask initialization in NSFF as our method does not use such information.", + "bbox": [ + 75, + 286, + 470, + 377 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathbf{D}^{2}\\mathbf{N}\\mathbf{e}\\mathbf{R}\\mathbf{F}$ [40] This method also cannot produce semantic clusterings. Over HyperNeRF [26], it adds a shadow field network and further losses to try to isolate objects into the dynamic NeRF over the separate static one. The paper also compares to NSFF without motion mask initialization.", + "bbox": [ + 75, + 381, + 470, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DINO-ViT (2D) [1] We ignore the volume and pass 2D semantic and attention features into the clustering algorithm. This cannot apply to novel viewpoints. Instead, we evaluate the approach upon all multi-view color images—input and hold-out—whereas other methods must render hold-out views. With our added pyramid processing (Sec. 3.3).", + "bbox": [ + 75, + 465, + 470, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- w/o pyr No pyramid; upsample to input RGB size.", + "bbox": [ + 76, + 563, + 429, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ProposeReduce (2D) [19] As a general comparison point, we apply this state-of-the-art 2D video segmentation method. For object segmentation, we use a ProposeReduce network that was pretrained with supervision on YouTube-VIS 2019 [43] for instance segmentation. For foreground segmentation, we use weights pretrained on UVOS [5] intended specifically for unsupervised foreground segmentation. As ProposeReduce is only a 2D method, we provide it with hold-out images for splits with novel views rather than our approach that must render novel views at hold-out poses.", + "bbox": [ + 75, + 585, + 470, + 737 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Findings", + "text_level": 1, + "bbox": [ + 76, + 746, + 181, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dynamic scene decomposition We separate the background and each foreground object individually (Tab. 1). The baseline 2D DINO-ViT method is improved by our pyramid approach. But, being only 2D, this fails to produce a consistent decomposition across novel spacetime views even when given ground truth hold-out images. This shows the value of a volume integration. Next, supervised ProposeReduce produces good results (Fig. 5), but sometimes misses salient objects or fails to join them behind occluding objects, and only", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/350ded1779f0cc54162f52260ead563617104562086209c2929a161f8df6657e.jpg", + "table_caption": [ + "Table 1: Spacetime volume integration improves dynamic scene decomposition. Pyramid construction and cluster merging help quantitatively, and ours is comparable to SOTA supervised 2D video segmentation network ProposeReduce.", + "Metric: Adjusted Rand Index $[-1, 1]$ , higher is better." + ], + "table_footnote": [], + "table_body": "
InputFix Cam 0Fix Time 0
ProposeReduce (2D)0.7250.7360.742
DINO-ViT (2D)0.5010.4950.321
w/o pyr帽子,hat0.4700.4640.346
SAFF (3D)0.5940.5780.566
SAFF (ours)0.6530.6340.625
w/ pyr λa = {1,0,0}0.6200.5980.592
w/o pyr帽子,hat0.5450.5320.521
w/o merge cluster0.5930.5740.563
w/post process0.7590.7330.735
w/oracle0.8340.8060.800
w/oracle + post process0.9220.8900.880
", + "bbox": [ + 504, + 359, + 888, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sometimes produces better edges than our method without post-processing as it can oversmooth edges. ProposeReduce also receives ground truth images in hold-out sets.", + "bbox": [ + 496, + 556, + 892, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Instead, our approach must render hold-out views via the volume reconstruction. This produces more consistent segmentations through spacetime manipulations—this is the added value of volume integration through learned saliency and attention heads. Ablated components show the value of our pyramid step, its coarse-saliency-only variant, and the cluster merge and image post processing steps. Qualitatively, we see good detail (Fig. 5); post-processing additionally improves edge quality and removes small unwanted regions.", + "bbox": [ + 496, + 604, + 893, + 741 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Oracle saliency We also include an oracle experiment where saliency voting is replaced by cluster selection based on ground truth masks. This experiment tells us what part of the performance gap lies with saliency itself, and what remains due to volume integration and cluster boundary issues. With oracle clusters, our decomposition performance is 0.8 ARI (Tab. 1) even in hold-out views. This shows that our existing cluster boundaries are accurate, and that accurate saliency for object selection is the larger remaining problem.", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21803", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cc037c42e5797e3aadb14db85e9ca407d4c278629fbb159978018f9ae1fcf99d.jpg", + "table_caption": [ + "Table 2: Saliency improves foreground segmentation. Adding saliency also slightly aids how much static/dynamic blend weight $v$ represents the foreground (cf. NSFF blend $v$ to SAFF's). Here, ProposeReduce uses unsupervised training and data [5]. Metric: IoU/Jaccard ([0, 1], higher is better)." + ], + "table_footnote": [], + "table_body": "
InputFix Cam 0Fix Time 0
ProposeReduce (2D)0.6090.4640.591
DINO-ViT (2D)0.3810.3820.357
NSFF — blend v0.3220.3090.268
D2NeRF— blend v0.4700.3340.269
SAFF — saliency0.6090.5890.572
— blend v0.3880.3800.329
— post process v0.7200.6940.679
", + "bbox": [ + 94, + 169, + 449, + 297 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/38078e65d1eed18ee574ddb69b341f59d29e0a4b8545ee092f7d422d3faf9930.jpg", + "table_caption": [ + "Table 3: SAFF semantics generalize to DyCheck Dataset (Masked) Oracle saliency. ProposeReduce is given ground truth test frames, while SAFF must render them. SAFF is comparable to SOTA supervised 2D video segmentation network ProposeReduce. Metric: Adjusted Rand Index $[-1, 1]$ , higher is better); IoU/Jaccard index $[-1, 1]$ , higher is better)." + ], + "table_footnote": [], + "table_body": "
ARIIoU
InputTestInputTest
ProposeReduce (2D)0.7610.7620.7610.762
DINO-ViT (2D)0.8010.8000.7970.797
SAFF (3D)0.9020.8200.9100.812
", + "bbox": [ + 81, + 407, + 465, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Foreground segmentation We simplify the problem and consider all objects as simply 'foreground' to compare to methods that do not produce per object masks. Here, the same trend continues (Tab. 2). We note more subtle improvements to static/dynamic blending weights when adding our additional feature heads to the backbone MLP, and overall show that adding top-down information helps produce more useful object masks. Qualitative results show whole objects in the foreground rather than broad regions of possible dynamics (NSFF) or broken objects $(\\mathrm{D}^2\\mathrm{NeRF};$ Fig. 6).", + "bbox": [ + 75, + 511, + 468, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DyCheck evaluation DyCheck often has close-up objects and, even with our selected sequences, saliency struggles to find foreground objects. Thus, we use oracle saliency. DINOViT and ProposeReduce are given test views while SAFF must render them. Quantitative (Tab. 3) and qualitative (Fig. 7) experiments show similar trends as before: ProposeReduce is good but may miss objects and fine details; SAFF may produce finer details and is more geometrically consistent.", + "bbox": [ + 75, + 671, + 470, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion and Limitations", + "text_level": 1, + "bbox": [ + 76, + 806, + 326, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DINO-ViT features are not instance-aware (Fig. 8). This is in contrast to object-centric learning approaches that aim to identify individual objects. To represent these different approaches, we compare to a result from slot-attention based $\\mathrm{SAVi}++$ [7]. This method trains on thousands of supervised", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cdd7b3c6f9d1e1e1425caa90ae6f392fa17fe43167b54091676d627962928cfa.jpg", + "image_caption": [ + "Input ProposeReduce [19] SAFF (Ours; input) SAFF (Ours; novel)" + ], + "image_footnote": [], + "bbox": [ + 504, + 99, + 598, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/273f11278c90df905bb04230318be2162cc45f2dd6ab17510c1db4b067fbc468.jpg", + "image_caption": [ + "Figure 7: SAFF can apply to DyCheck Dataset (Masked). SAFF is comparable to supervised large-scale 2D video learning [19]. Some fine details are improved (bottom, fingers) and novel views maintain their geometry (bottom, gap between legs occluded by head at input timestep)." + ], + "image_footnote": [], + "bbox": [ + 504, + 196, + 599, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e822e330344030109a8f27f3b32ab14a8a9a04c53f39ea6bf92763540d178a21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 99, + 697, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f118d418e44feb85ea722ffb224829fa95d77cb850e703874a4d2dcaedeb8466.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 196, + 697, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/41e9bf5dedef7a9aa4433f49ac7d51e411c80fd64d4ff68c6fd4a6c02c4dc620.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 99, + 795, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d63c51f4f565fe03ce474325545808eb92b7f3beafc4605e703a8166445f1dad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 196, + 795, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/662699dfb3e9528689d2f33db4a4c8da2c57f93c1e930c0ca44483f2db93a7b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 799, + 99, + 895, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c894697ca2b290957c8e39c04bb2ec7be403cd2d62bcad781bbbe2c4acfe88a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 799, + 196, + 895, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6729bdccffeb5d82405d77c40eecf02828d678f74eb3d84859773e7e5f4a06ba.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 504, + 393, + 627, + 463 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/988c7a0f5473de6d3d76cb7af8e50e1eb2966cae1592553828303c6e8169f2ae.jpg", + "image_caption": [ + "(b) $\\mathrm{SAV_i + + [7]}$" + ], + "image_footnote": [], + "bbox": [ + 632, + 393, + 756, + 463 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e8844aa33ab7268e2befb4420a8bdff4f44183a0e86e2b9956937aa9eb001cd7.jpg", + "image_caption": [ + "(c) SAFF" + ], + "image_footnote": [], + "bbox": [ + 761, + 393, + 885, + 462 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cc9c34111c72b5cd2873e5abb927357374f13c87078284374297bf3b4b5a6620.jpg", + "image_caption": [ + "(d) Input", + "Figure 8: Limitations. DINO-ViT is not instance-aware, causing unwanted grouping (c). Unwanted static objects may be salient; assuming salient objects are dynamic fixes this (f)." + ], + "image_footnote": [], + "bbox": [ + 504, + 481, + 627, + 532 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d0b34e16db9794f6f8c60fa715c6dfee06a9337e01c9fc2539e48867a77d9b04.jpg", + "image_caption": [ + "(e) SAFF salient" + ], + "image_footnote": [], + "bbox": [ + 629, + 481, + 753, + 532 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bfdbd6632afcaeb6b3475a898124129ca7c741b11ecfe7f74ba05018b82b489f.jpg", + "image_caption": [ + "(f) Salient+dynamic" + ], + "image_footnote": [], + "bbox": [ + 758, + 481, + 883, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MOVi [11] sequences with per-object masks, whereas we use generic pre-trained features and gain better edges from volume integration. Combining these two approaches could give accurate instance-level scene objects.", + "bbox": [ + 498, + 594, + 890, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DINO-ViT saliency may attend to unwanted regions. In Figure 8d-e, the static pillars could be isolated using scene flow. But, often our desired subjects do not move (cf. people in Umbrella or Balloon NBoard). For tasks or data that can assume that salient objects are dynamic, we use SAFF's 4D scene reconstruction to reject static-but-salient objects by merging clusters via scene flow: First, we project $\\mathbf{f}$ over each timestep into each input camera pose—this simulates optical flow with a static camera. Clusters are marked as salient per image if mean flow magnitude per cluster $|\\bar{\\mathbf{p}}| > 0.07$ and mean attention $\\bar{\\mathbf{a}}_c > 0.07$ . Finally, as before, a cluster is globally salient if $70\\%$ of images agree (Fig. 8f).", + "bbox": [ + 496, + 655, + 892, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements The CV community in New England for feedback, and funding from NSF CNS-2038897 and an Amazon Research Award. Eliot thanks a Randy F. Pausch '82 Undergraduate Summer Research Award at Brown CS.", + "bbox": [ + 498, + 839, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21804", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2021. 2, 4, 5, 7", + "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 4", + "[3] Kenneth Blomqvist, Lionel Ott, Jen Jen Chung, and Roland Y. Siegwart. Baking in the feature: Accelerating volumetric segmentation by rendering feature maps. ArXiv, abs/2209.12744, 2022. 2", + "[4] Christopher P. Burgess, Loic Matthews, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matthew M. Botvinick, and Alexander Lerchner. Monet: Unsupervised scene decomposition and representation. ArXiv, abs/1901.11390, 2019. 1, 2", + "[5] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv:1905.00737, 2019. 2, 7, 8", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 1, 2, 4, 5", + "[7] Gamaeldin F. Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael C. Mozer, and Thomas Kipf. $\\mathrm{SAVi}++$ : Towards end-to-end object-centric learning from real-world videos. In Advances in Neural Information Processing Systems, 2022. 1, 2, 8", + "[8] Zhiwen Fan, Peihao Wang, Yifan Jiang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation on complex scenes. ArXiv, abs/2209.08776, 2022. 2", + "[9] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In International Conference on 3D Vision (3DV), 2022. 2", + "[10] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. In NeurIPS, 2022. 1, 6", + "[11] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam H. Laradji, Hsueh-Ti Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: A scalable dataset generator. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3739-3751, 2022. 8", + "[12] Klaus Greff, Raphael Lopez Kaufman, Rishabh Kabra, Nicholas Watters, Christopher P. Burgess, Daniel Zoran, Loic" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Matthew, Matthew M. Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative variational inference. In ICML, 2019. 2", + "[13] Basile Van Hoorick, Purva Tendulka, Dídac Surís, Dennis Park, Simon Stent, and Carl Vondrick. Revealing occlusions with 4d neural fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3001-3011, 2022. 2", + "[14] Rishabh Kabra, Daniel Zoran, Goker Erdogan, Loic Matthey, Antonia Creswell, Matthew M. Botvinick, Alexander Lerchner, and Christopher P. Burgess. Simone: View-invariant, temporally-abstracted object representations via unsupervised video decomposition. ArXiv, abs/2106.03849, 2021. 1, 2", + "[15] Thomas Kipf, Gamaleldin F. Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional Object-Centric Learning from Video. In International Conference on Learning Representations (ICLR), 2022. 1, 2", + "[16] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. In Advances in Neural Information Processing Systems, volume 35, 2022. 1, 2, 4", + "[17] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J. Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12871-12881, June 2022. 2", + "[18] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 7", + "[19] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1719-1728, 2021. 1, 5, 6, 7, 8", + "[20] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. arXiv preprint arXiv:2006.15055, 2020.1,2,5", + "[21] Kirill Mazur, Edgar Sucar, and Andrew J. Davison. Feature-realistic neural fusion for real-time, open set scene understanding. ArXiv, abs/2210.03043, 2022. 2", + "[22] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In CVPR, 2022. 2", + "[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3", + "[24] Tom Monnier, Elliot Vincent, Jean Ponce, and Mathieu Aubry. Unsupervised layered image decomposition into object prototypes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8640-8650, October 2021. 2" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21805", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[26] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), dec 2021. 1, 2, 7", + "[27] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022. 2", + "[28] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2", + "[29] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, and Francesco Locatello. Bridging the gap to real-world object-centric learning. ArXiv, abs/2209.14860, 2022. 2", + "[30] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In CVPRW, 2022. 2", + "[31] Cameron Smith, Hong-Xing Yu, Sergey Zakharov, Frédo Durand, Joshua B. Tenenbaum, Jiajun Wu, and Vincent Sitzmann. Unsupervised discovery and composition of object light fields. ArXiv, abs/2205.03923, 2022. 1, 2", + "[32] Karl Stelzner, Kristian Kersting, and Adam R. Kosiorek. Decomposing 3d scenes into objects via unsupervised volume segmentation. ArXiv, abs/2104.01148, 2021. 1, 2, 5", + "[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2", + "[34] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 2", + "[35] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12939-12950, 2021. 1, 2", + "[36] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural Feature Fusion Fields: 3D distillation of self-supervised 2D image representations. In Proceedings of the International Conference on 3D Vision (3DV), 2022. 2, 4", + "[37] Vadim Tschernezki, Diane Larlus, and Andrea Vedaldi. NeuralDiff: Segmenting 3D objects that move in egocentric videos. In Proceedings of the International Conference on 3D Vision (3DV), 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Rishi Veerapaneni, John D. Co-Reyes, Michael Chang, Michael Janner, Chelsea Finn, Jiajun Wu, Joshua Tenenbaum, and Sergey Levine. Entity abstraction in visual model-based reinforcement learning. In Leslie Pack Kaelbling, Danica Kragic, and Komei Sugiura, editors, Proceedings of the Conference on Robot Learning, volume 100 of Proceedings of Machine Learning Research, pages 1439-1456. PMLR, 30 Oct-01 Nov 2020. 1", + "[39] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L. Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[40] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Öz Tireli. D2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. ArXiv, abs/2205.15838, 2022. 1, 2, 5, 7", + "[41] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 2022. 2", + "[42] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for editable scene rendering. In International Conference on Computer Vision (ICCV), October 2021. 1, 2", + "[43] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Oct 2019. 7", + "[44] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snavely. Deformable sprites for unsupervised video decomposition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 2", + "[45] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 5", + "[46] Hong-Xing Yu, Leonidas J. Guibas, and Jiajun Wu. Unsupervised discovery of object radiance fields. In International Conference on Learning Representations, 2022. 1, 2, 5", + "[47] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 2" + ], + "bbox": [ + 501, + 92, + 893, + 726 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21806", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/68ea1ef8-9cb5-44da-894d-92e02632af66_model.json b/2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/68ea1ef8-9cb5-44da-894d-92e02632af66_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4ffa38b38f7f3f119f99058300c64739ae946771 --- /dev/null +++ b/2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/68ea1ef8-9cb5-44da-894d-92e02632af66_model.json @@ -0,0 +1,3146 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.103, + 0.872, + 0.126 + ], + "angle": 0, + "content": "Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.144, + 0.223, + 0.163 + ], + "angle": 0, + "content": "Yiqing Liang" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.144, + 0.356, + 0.162 + ], + "angle": 0, + "content": "Eliot Laidlaw" + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.144, + 0.562, + 0.181 + ], + "angle": 0, + "content": "Alexander Meyerowitz Brown University" + }, + { + "type": "text", + "bbox": [ + 0.583, + 0.144, + 0.706, + 0.161 + ], + "angle": 0, + "content": "Srinath Sridhar" + }, + { + "type": "text", + "bbox": [ + 0.728, + 0.144, + 0.856, + 0.163 + ], + "angle": 0, + "content": "James Tompkin" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.21, + 0.314, + 0.227 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.239, + 0.474, + 0.466 + ], + "angle": 0, + "content": "From video, we reconstruct a neural volume that captures time-varying color, density, scene flow, semantics, and attention information. The semantics and attention let us identify salient foreground objects separately from the background across spacetime. To mitigate low resolution semantic and attention features, we compute pyramids that trade detail with whole-image context. After optimization, we perform a saliency-aware clustering to decompose the scene. To evaluate real-world scenes, we annotate object masks in the NVIDIA Dynamic Scene and DyCheck datasets. We demonstrate that this method can decompose dynamic scenes in an unsupervised way with competitive performance to a supervised method, and that it improves foreground/background segmentation over recent static/dynamic split methods. Project webpage: https://visual.cs.brown.edu/saff" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.481, + 0.208, + 0.497 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.471, + 0.596 + ], + "angle": 0, + "content": "Given a casually-captured monocular RGB video of a dynamic scene, decomposing it into foreground objects and background is an important task in computer vision, with downstream applications in segmentation and video editing. An ideal method would also reconstruct the geometry and appearance over time including frame correspondences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.471, + 0.748 + ], + "angle": 0, + "content": "Previous methods have made great progress but there are many limitations. Some works assume that there is no object motion in the scene [32, 46, 31, 16, 4], take input from multi-view cameras [16, 46], or do not explicitly reconstruct the underlying 3D structure of the scene [4, 7]. For objects, some works rely on masks or user input to aid segmentation [42, 15, 7], or use task-specific training datasets [7]. Sometimes, works assume the number of foreground objects [4, 20]. Given the challenges, many works train and test on synthetic data [14, 38]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.902 + ], + "angle": 0, + "content": "We present Semantic Attention Flow Fields (SAFF): A method to overcome these limitations by integrating low-level reconstruction cues with high-level pretrained information—both bottom-up and top-down—into a neural volume. With this, we demonstrate that embedded semantic and saliency (attention) information is useful for unsupervised dynamic scene decomposition. SAFF builds upon neural scene flow fields [18], an approach that reconstructs appearance, geometry, and motion. This uses frame interpolation rather than explicit canonicalization [35] or a latent hyperspace [26]," + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.199, + 0.88, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.367, + 0.747, + 0.38 + ], + "angle": 0, + "content": "Density \\(\\sigma\\), Saliency \\(a\\), Semantics \\(s\\), Color \\(c\\)" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.383, + 0.695, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.383, + 0.876, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.457, + 0.874, + 0.47 + ], + "angle": 0, + "content": "Novel view renderings with effects (depth of field, color adjustment)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.477, + 0.895, + 0.539 + ], + "angle": 0, + "content": "Figure 1: We decompose a dynamic 3D scene from monocular input into time-varying color, density, scene flow, semantics, and attention information. This could be used to provide precise editing of objects or effects focused around objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.895, + 0.719 + ], + "angle": 0, + "content": "which lets it more easily apply to casual videos. For optimization, we supervise two network heads with pretrained DINO-ViT [6] semantic features and attention. Naively supervising high-resolution SAFF with low-resolution DINO-ViT output reduces reconstruction quality. To mitigate the mismatch, we build a semantic attention pyramid that trades detail with whole-image context. Having optimized a SAFF representation for a dynamic scene, we perform a saliency-aware clustering both in 3D and on rendered feature images to describe objects and their background. Given the volume reconstruction, the clustering generalizes to novel spacetime views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "To evaluate SAFF's dynamic scene decomposition capacity, we expand the NVIDIA Dynamic Scene [45] and DyCheck [10] datasets by manually annotating object masks across input and hold-out views. We demonstrate that SAFF outperforms 2D DINO-ViT baselines and is comparable to a state-of-the-art video segmentation method ProposeReduce[19] on our data. Existing monocular video dynamic volume reconstruction methods typically separate static and dynamic parts, but these often do not represent meaningful foreground. We show improved foreground segmentation over NSFF and the current \\(\\mathrm{D}^2\\mathrm{NeRF}\\) [40] method for downstream tasks like editing." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "21797" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.218, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.113, + 0.471, + 0.293 + ], + "angle": 0, + "content": "Decomposing a scene into regions of interest is a long-studied task in computer vision [5], including class, instance, and panoptic segmentation in high-level or top-down vision, and use of low-level or bottom-up cues like motion. Recent progress has considered images [12, 4, 20], videos [14, 15], layer decomposition [44, 24], and in-the-wild databases using deep generative models [25]. One example, \\(\\mathrm{SAVi}++[7]\\), uses slot attention [20] to define 2D objects in real-world videos. Providing first-frame segmentation masks achieves stable performance, with validation on the driving Waymo Open Dataset [33]. Our work attempts 3D scene decomposition for a casual monocular video without initial masks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.471, + 0.522 + ], + "angle": 0, + "content": "Scene Decomposition with NeRFs Neural Radiance Fields (NeRF) [41] have spurred new scene decomposition research through volumes. ObSuRF [32] and uORF [46] are unsupervised slot attention works that bind a latent code to each object. Unsupervised decomposition is also possible on light fields [31]. For dynamic scenes, works like NeuralDiff [37] and \\(\\mathrm{D}^2\\mathrm{NeRF}\\) [40] focus on foreground separation, where foreground is defined to contain moving objects. Other works like N3F [36] and occlusions-4d [13] also decompose foregrounds into individual objects. N3F requires user input to specify which object to segment, and occlusions-4d takes RGB point clouds as input. Our work attempts to recover a segmented dynamic scene from a monocular RGB video without added markup or masks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.47, + 0.676 + ], + "angle": 0, + "content": "Neural Fields Beyond Radiance Research has begun to add non-color information to neural volumes to aid decomposition via additional feature heads on the MLP. iLabel [47] adds a semantic head to propagate user-provided segmentations in the volume. PNF [17] and Panoptic-NeRF [9] attempt panoptic segmentation within neural fields, and Object-NeRF integrates instance segmentation masks into the field during optimization [42]. Research also investigates how to apply generic pretrained features to neural fields, like DINO-ViT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.692, + 0.471, + 0.903 + ], + "angle": 0, + "content": "DINO-ViT for Semantics and Saliency DINO-ViT is a self-supervised transformer that, after pretraining, extracts generic semantic information [6]. Amir et al. [1] use DINO-ViT features with \\( k \\)-means clustering to achieve co-segmentation across a video. Seitzer et al. [29] combine slot attention and DINO-ViT features for object-centric learning on real-world 2D data. TokenCut [39] performs normalized cuts on DINO-ViT features for foreground segmentation on natural images. Deep Spectral Segmentation [22] show that graph Laplacian processing of DINO-ViT features provides unsupervised foreground segmentation, and Selfmask [30] shows that these features can provide object saliency masks. Our approach considers these clustering and saliency findings for the setting of 3D decomposition from monocular video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.32 + ], + "angle": 0, + "content": "DINO-ViT Fields Concurrent works have integrated DINO-ViT features into neural fields. DFF [16] distills features for dense multi-view static scenes with user input for segmentation. N3F [36] expands NeuralDiff to dynamic scenes, and relies on user input for segmentation. AutoLabel [3] uses DINO-ViT features to accelerate segmentation in static scenes given a ground truth segmentation mask. Other works use DINO-ViT differently. FeatureRealistic-Fusion [21] uses DINO-ViT in an online feature fusion task, focusing on propagating user input, and NeRF-SOS [8] uses 2D DINO-ViT to process NeRF-rendered multi-view RGB images of a static scene. In contrast to these works, we consider real-world casual monocular videos, recover and decompose a 3D scene, then explore whether saliency can avoid the need for masks or user input in segmenting objects." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.33, + 0.59, + 0.345 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.894, + 0.502 + ], + "angle": 0, + "content": "For a baseline dynamic scene reconstruction method, we begin with NSFF from Li et al. [18] (Sec. 3.1), which builds upon NeRF [23]. NSFF's low-level scene flow frame-to-frame approach provides better reconstructions for real-world casual monocular videos than deformation-based methods [35, 26]. We modify the architecture to integrate higher-level semantic and saliency (or attention) features (Sec. 3.2). After optimizing a SAFF for each scene, we perform saliency-aware clustering of the field (Sec. 3.4). All implementation details are in our supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.513, + 0.894, + 0.62 + ], + "angle": 0, + "content": "Input Our method takes in a single RGB video over time \\(i\\) as an ordered set of images \\(I \\in \\mathcal{I}\\) and camera poses. We use COLMAP to recover camera poses [28]. From all poses, we define an NDC-like space that bounds the scene, and a set of rays \\(\\mathbf{r} \\in \\mathcal{R}\\), one per image pixel with color \\(\\hat{\\mathbf{c}}^\\dagger\\). Here, \\(\\hat{\\mathbf{c}}\\) denotes a 2D pixel value in contrast to a 3D field value, and \\(\\hat{\\mathbf{c}}^\\dagger\\) denotes an input value in contrast to an estimated value." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.892, + 0.68 + ], + "angle": 0, + "content": "From pretrained networks, we estimate single-frame monocular depth \\(\\hat{d}^{\\dagger}\\) (MiDaSv2 [27]), optical flow \\(\\hat{\\mathbf{p}}_i^\\dagger\\) (RAFT [34]), and semantic features \\(\\hat{\\mathbf{s}}^\\dagger\\) and attention \\(\\hat{\\mathbf{a}}^\\dagger\\) (DINO-ViT [6]) after important preprocessing (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.688, + 0.892, + 0.705 + ], + "angle": 0, + "content": "3.1. Initial Dynamic Neural Volume Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.708, + 0.894, + 0.815 + ], + "angle": 0, + "content": "The initial representation comprises a static NeRF \\( F_{\\theta}^{\\mathrm{st}} \\) and a dynamic NeRF \\( F_{\\theta}^{\\mathrm{dy}} \\). The static network predicts a color \\( \\mathbf{c} \\), density \\( \\sigma \\), and blending weight \\( v \\) (Eq. (1)), and the dynamic network predicts time-varying color \\( \\mathbf{c}_i \\), density \\( \\sigma_i \\), scene flow \\( \\mathbf{f}_i \\), and occlusion weights \\( w_i \\) (Eq. (2)). In both network architectures, other than position \\( \\mathbf{x} \\), we add direction \\( \\boldsymbol{\\omega} \\) to a late separate head to only condition the estimation of color \\( \\mathbf{c} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.824, + 0.892, + 0.841 + ], + "angle": 0, + "content": "\\[\nF _ {\\boldsymbol {\\theta}} ^ {\\mathrm {s t}}: (\\mathbf {x}, \\omega) \\rightarrow \\left(\\mathbf {c} ^ {\\mathrm {s t}}, \\sigma^ {\\mathrm {s t}}, v\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.844, + 0.892, + 0.863 + ], + "angle": 0, + "content": "\\[\nF _ {\\boldsymbol {\\theta}} ^ {\\mathrm {d y}}: (\\mathbf {x}, \\boldsymbol {\\omega}, i) \\rightarrow \\left(\\mathbf {c} _ {i} ^ {\\mathrm {d y}}, \\sigma_ {i} ^ {\\mathrm {d y}}, \\mathbf {f} _ {i}, w _ {i}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To produce a pixel's color, we sample points at distances \\(t\\) along the ray \\(\\mathbf{x}_t = \\mathbf{x} - \\omega t\\) between near and far planes \\(t_n\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "21798" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.09, + 0.303, + 0.105 + ], + "angle": 0, + "content": "(a) SAFF video inputs" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.107, + 0.351, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.417, + 0.09, + 0.532, + 0.105 + ], + "angle": 0, + "content": "(b) SAFF model" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.113, + 0.574, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.09, + 0.846, + 0.106 + ], + "angle": 0, + "content": "(c) Volume rendered outputs" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.109, + 0.88, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.305, + 0.894, + 0.351 + ], + "angle": 0, + "content": "Figure 2: Overview. From monocular video, SAFF builds a neural field of scene-flowed 3D density, radiance, semantics, and attention (b). This is guided by semantic attention pyramids that increase resolution, plus depth and optical flow priors (a). We can render new spacetime views of any channel, and use saliency-aware clustering to decompose objects and background (c)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.471, + 0.425 + ], + "angle": 0, + "content": "to \\( t_f \\), query each network, then integrate transmittance \\( T \\), density \\( \\sigma \\), and color along the ray from these samples [23]. For brevity, we omit evaluating at \\( \\mathbf{x}_t, \\boldsymbol{\\omega} \\), e.g., \\( \\sigma^{\\mathrm{st}}(\\mathbf{x}_t) \\) is simply \\( \\sigma^{\\mathrm{st}} \\); \\( \\mathbf{c}^{\\mathrm{st}}(\\mathbf{x}_t, \\boldsymbol{\\omega}) \\) is simply \\( \\mathbf{c}^{\\mathrm{st}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.425, + 0.47, + 0.455 + ], + "angle": 0, + "content": "We produce a combined color from the static and dynamic colors by multiplication with their densities:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.461, + 0.47, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {i} \\mathbf {c} _ {i} = v \\sigma^ {\\mathrm {s t}} \\mathbf {c} ^ {\\mathrm {s t}} + (1 - v) \\sigma_ {i} ^ {\\mathrm {d y}} \\mathbf {c} _ {i} ^ {\\mathrm {d y}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.492, + 0.47, + 0.538 + ], + "angle": 0, + "content": "Given that transmittance integrates density up to the current sampled point under Beer-Lambert volume attenuation, the rendered pixel color for the ray is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.544, + 0.47, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {c}} _ {i} = \\int_ {t _ {n}} ^ {t _ {f}} T _ {i} \\sigma_ {i} \\mathbf {c} _ {i} d t \\text {w h e r e} T _ {i} = \\exp \\left(- \\int_ {t _ {n}} ^ {t} \\sigma_ {i} d t\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.47, + 0.634 + ], + "angle": 0, + "content": "To optimize the volume to reconstruct input images, we compute a photometric loss \\( L_{\\hat{\\mathbf{c}}} \\) between rendered and input colors:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.643, + 0.47, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\hat {\\mathbf {c}}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\left\\| \\hat {\\mathbf {c}} _ {i} \\left(\\mathbf {r} _ {i}\\right) - \\hat {\\mathbf {c}} _ {i} ^ {\\dagger} \\left(\\mathbf {r} _ {i}\\right) \\right\\| _ {2} ^ {2} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.694, + 0.47, + 0.787 + ], + "angle": 0, + "content": "Scene Flow Defining correspondence over time is important for monocular input as it lets us penalize a reprojection error with neighboring frames \\( j \\in \\mathcal{N} \\), e.g., where \\( j = i + 1 \\) or \\( j = i - 1 \\). We denote \\( i \\rightarrow j \\) for the projection of frame \\( i \\) onto frame \\( j \\) by scene flow. \\( F_{\\theta}^{\\mathrm{dy}} \\) estimates both forwards and backwards scene flow at every point to penalize a bi-direction loss." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.787, + 0.469, + 0.817 + ], + "angle": 0, + "content": "Thus, we approximate color output at time step \\(j\\) by flowing the queried network values at time step \\(i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.826, + 0.47, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {c}} _ {i \\rightarrow j} = \\int_ {t _ {n}} ^ {t _ {f}} T _ {i \\rightarrow j} \\sigma_ {i \\rightarrow j} \\mathbf {c} _ {i \\rightarrow j} d t \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Reprojection must account for occlusion and disocclusion by motion. As such, \\( F_{\\theta}^{\\mathrm{dy}} \\) also predicts forwards and backwards" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.893, + 0.411 + ], + "angle": 0, + "content": "scene occlusion weights \\( w_{i + 1} \\) and \\( w_{i - 1} \\in [0,1] \\), where a point with \\( w_{i + 1} = 0 \\) means that occlusion status is changed one step forwards in time. We can integrate \\( w \\) to a pixel:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.416, + 0.893, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\hat {w} _ {i \\rightarrow j} = \\int_ {t _ {n}} ^ {t _ {f}} T _ {i \\rightarrow j} \\sigma_ {i \\rightarrow j} w _ {j} d t \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.458, + 0.893, + 0.492 + ], + "angle": 0, + "content": "Then, this pixel weight modulates the color loss such that occluded pixels are ignored. (8)" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.494, + 0.893, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\hat {\\mathbf {c}} _ {i \\rightarrow j}} = \\frac {1}{| \\mathcal {R} | | \\mathcal {N} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\sum_ {j \\in \\mathcal {N}} \\hat {w} _ {i \\rightarrow j} (\\mathbf {r} _ {i}) | | \\hat {\\mathbf {c}} _ {i \\rightarrow j} (\\mathbf {r} _ {i}) - \\hat {\\mathbf {c}} _ {j} ^ {\\dagger} (\\mathbf {r} _ {j}) | | _ {2} ^ {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.895, + 0.669 + ], + "angle": 0, + "content": "Prior losses We use the pretrained depth and optical flow map losses to help overcome the ill-posed monocular reconstruction problem. These losses decay as optimization progresses to rely more and more on the optimized self-consistent geometry and scene flow. For geometry, we estimate a depth \\(\\hat{d}_i\\) for each ray \\(\\mathbf{r}_i\\) by replacing \\(\\mathbf{c}_i\\) in Eq. (4) by the distance \\(t\\) along the ray. Transform \\(z\\) estimates a scale and shift as the pretrained network produces only relative depth." + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.671, + 0.893, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\hat {d}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\left| \\left| \\hat {d} _ {i} - z \\left(\\hat {d} _ {i} ^ {\\dagger}\\right) \\right| \\right| _ {1} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.893, + 0.807 + ], + "angle": 0, + "content": "For motion, projecting scene flow to a camera lets us compare to the estimated optical flow. Each sample point along a ray \\(\\mathbf{x}_i\\) is advected to a point in the neighboring frame \\(\\mathbf{x}_{i\\rightarrow j}\\), then integrated to the neighboring camera plane to produce a 2D point offset \\(\\hat{\\mathbf{p}}_i(\\mathbf{r}_i)\\). Then, we expect the difference in the start and end positions to match the prior:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.81, + 0.893, + 0.849 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\hat {\\mathbf {p}}} = \\frac {1}{| \\mathcal {R} | | \\mathcal {N} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\sum_ {j \\in \\mathcal {N} (i)} | | \\hat {\\mathbf {p}} _ {i} (\\mathbf {r} _ {i}) - \\hat {\\mathbf {p}} _ {i} ^ {\\dagger} (\\mathbf {r} _ {i}) | | _ {1} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Additional regularizations encourage occlusion weights to be close to one, scene flow to be small, locally constant, and cyclically consistent, and blending weight \\( v \\) to be sparse." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "21799" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.355, + 0.106 + ], + "angle": 0, + "content": "3.2. Semantic Attention Flow Fields" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.235 + ], + "angle": 0, + "content": "Beyond low-level or bottom-up features, high-level or top-down features are also useful to define objects and help down-stream tasks like segmentation. For example, methods like NSFF or \\(\\mathrm{D}^2\\mathrm{NeRF}\\) struggle to provide useful separation of static and dynamic parts because blend weight \\(v\\) estimates whether the volume appears to be occupied by some moving entity. This is not the same as objectness; tasks like video editing could benefit from accurate dynamic object masks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.236, + 0.471, + 0.356 + ], + "angle": 0, + "content": "As such, we extract 2D semantic features and attention (or saliency) values from a pretrained DINO-ViT network, then optimize the SAFF such that unknown 3D semantic and attention features over time can be projected to recreate their 2D complements. This helps us to ascribe semantic meaning to the volume and to identify objects. As semantics/attention are integrated into the 4D volume, we can render them from novel spacetime views without further DINO-ViT computation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.356, + 0.47, + 0.403 + ], + "angle": 0, + "content": "To estimate semantic features \\(\\mathbf{s}\\) and attention \\(\\mathbf{a}\\) at 3D points in the volume at time \\(i\\), we add two new heads to both the static \\(F_{\\theta}^{\\mathrm{st}}\\) and the dynamic \\(F_{\\theta}^{\\mathrm{dy}}\\) networks:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.408, + 0.47, + 0.426 + ], + "angle": 0, + "content": "\\[\nF _ {\\boldsymbol {\\theta}} ^ {\\mathrm {s t}}: (\\mathbf {x}, \\omega) \\quad \\rightarrow \\left(\\mathbf {c} ^ {\\mathrm {s t}}, \\sigma^ {\\mathrm {s t}}, v, \\mathbf {s} ^ {\\mathrm {s t}}, \\mathbf {a} ^ {\\mathrm {s t}}\\right) \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.429, + 0.47, + 0.449 + ], + "angle": 0, + "content": "\\[\nF _ {\\boldsymbol {\\theta}} ^ {\\mathrm {d y}}: (\\mathbf {x}, \\boldsymbol {\\omega}, i) \\rightarrow \\left(\\mathbf {c} _ {i} ^ {\\mathrm {d y}}, \\sigma_ {i} ^ {\\mathrm {d y}}, \\mathbf {f} _ {i}, w _ {i}, \\mathbf {s} _ {i} ^ {\\mathrm {d y}}, \\mathbf {a} _ {i} ^ {\\mathrm {d y}}\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.455, + 0.47, + 0.5 + ], + "angle": 0, + "content": "As semantic features have been demonstrated to be somewhat robust to view dependence [1], in our architectures both heads for s, a are taken off the backbone before \\(\\omega\\) is injected." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.501, + 0.47, + 0.53 + ], + "angle": 0, + "content": "To render semantics from the volume, we replace the color term \\(\\mathbf{c}\\) in Eq. (4) with \\(\\mathbf{s}\\) and equivalently for \\(\\mathbf{a}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.535, + 0.47, + 0.569 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {i} \\mathbf {s} _ {i} = v \\sigma^ {\\mathrm {s t}} \\mathbf {s} ^ {\\mathrm {s t}} + (1 - v) \\sigma_ {i} ^ {\\mathrm {d y}} \\mathbf {s} _ {i} ^ {\\mathrm {d y}}, \\hat {\\mathbf {s}} _ {i} = \\int_ {t _ {n}} ^ {t _ {f}} T _ {i} \\sigma_ {i} \\mathbf {s} _ {i} d t \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.576, + 0.47, + 0.606 + ], + "angle": 0, + "content": "To encourage scene flow to respect semantics over time, we penalize complementary losses on s and a (showing s only):" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.608, + 0.47, + 0.653 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\hat {\\mathbf {s}} _ {i \\rightarrow j}} = \\frac {1}{| \\mathcal {R} | | \\mathcal {N} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\sum_ {j \\in \\mathcal {N}} \\hat {w} _ {i \\rightarrow j} (\\mathbf {r} _ {i}) \\| \\hat {\\mathbf {s}} _ {i \\rightarrow j} (\\mathbf {r} _ {i}) - \\hat {\\mathbf {s}} _ {j} ^ {\\dagger} (\\mathbf {r} _ {j}) \\| _ {2} ^ {2} \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.661, + 0.469, + 0.707 + ], + "angle": 0, + "content": "Finally, as supervision, we add respective losses on the reconstruction of the 2D semantic and attention features from projected 3D volume points (showing s only):" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.709, + 0.47, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\hat {\\mathbf {s}}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\left\\| \\hat {\\mathbf {s}} _ {i} \\left(\\mathbf {r} _ {i}\\right) - \\hat {\\mathbf {s}} _ {i} ^ {\\dagger} \\left(\\mathbf {r} _ {i}\\right) \\right\\| _ {2} ^ {2} \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.752, + 0.471, + 0.843 + ], + "angle": 0, + "content": "Unlike depth and scene flow priors, these are not priors—there is no self-consistency for semantics to constrain their values. Thus, we do not decay their contribution. While decaying avoids disagreements between semantic and attention features and color-enforced scene geometry, it also leads to a loss of useful meaning (please see supplemental)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.843, + 0.285, + 0.857 + ], + "angle": 0, + "content": "Thus our final loss becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.863, + 0.47, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {S A F F}} = \\mathcal {L} _ {\\hat {\\mathbf {c}}} + \\lambda_ {\\hat {\\mathbf {c}} _ {i \\rightarrow j}} \\mathcal {L} _ {\\hat {\\mathbf {c}} _ {i \\rightarrow j}} + \\lambda_ {\\hat {d}} \\mathcal {L} _ {\\hat {d}} + \\lambda_ {\\hat {\\mathbf {p}}} \\mathcal {L} _ {\\hat {\\mathbf {p}}} \\tag {16} \\\\ + \\lambda_ {\\hat {\\mathbf {s}} _ {i \\rightarrow j}} \\mathcal {L} _ {\\hat {\\mathbf {s}} _ {i \\rightarrow j}} + \\lambda_ {\\hat {\\mathbf {a}} _ {i \\rightarrow j}} \\mathcal {L} _ {\\hat {\\mathbf {a}} _ {i \\rightarrow j}} + \\lambda_ {\\hat {\\mathbf {s}}} \\mathcal {L} _ {\\hat {\\mathbf {s}}} + \\lambda_ {\\hat {\\mathbf {a}}} \\mathcal {L} _ {\\hat {\\mathbf {a}}} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.764, + 0.107 + ], + "angle": 0, + "content": "3.3. Semantic Attention Pyramids" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.894, + 0.265 + ], + "angle": 0, + "content": "When thinking about scenes, we might argue that semantics from an ideal extractor should be scale invariant, as distant objects have the same class as close objects. We might also argue that saliency (or attention features) may not be scale invariant, as small details in a scene should only be salient when viewed close up. In practice, both extracted features vary across scale and have limited resolution, e.g., DINO-ViT [6] produces one output for each \\(8 \\times 8\\) patch. But, from this, we want semantic features and saliency for every RGB pixel that still respects scene boundaries." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.894, + 0.386 + ], + "angle": 0, + "content": "Thus far, work on static scenes has ignored the input/feature resolution mismatch [16] as multi-view constraints provide improved localization within the volume. For monocular video, this approach has limitations [36]. Forming many constraints on dynamic objects requires long-term motion correspondence—a tricky task—and so we want to maximize the resolution of any input features where possible without changing their meaning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.894, + 0.507 + ], + "angle": 0, + "content": "One way may be through a pyramid of semantic and attention features that uses a sliding window approach at finer resolutions. Averaging features could increase detail around edges, but we must overcome the practical limit that these features are not stable across scales. This is especially important for saliency: unlike typical RGB pyramids that must preserve energy in an alias-free way [2], saliency changes significantly over scales and does not preserve energy." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.507, + 0.893, + 0.522 + ], + "angle": 0, + "content": "Consider a feature pyramid \\(\\mathcal{P}\\) with loss weights per level:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.542, + 0.893, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {P} \\hat {\\mathbf {s}}} = \\sum_ {i \\in \\mathcal {P}} \\lambda_ {\\hat {\\mathbf {s}}} ^ {i} \\mathcal {L} _ {\\hat {\\mathbf {s}}} ^ {i} \\quad \\mathcal {L} _ {\\mathcal {P} \\hat {\\mathbf {a}}} = \\sum_ {i \\in \\mathcal {P}} \\lambda_ {\\hat {\\mathbf {a}}} ^ {i} \\mathcal {L} _ {\\hat {\\mathbf {a}}} ^ {i} \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.894, + 0.735 + ], + "angle": 0, + "content": "Naively encouraging scale-consistent semantics and whole-image saliency, e.g., \\(\\lambda_{\\hat{\\mathbf{s}}} = \\{1/3, 1/3, 1/3\\}\\) with \\(\\lambda_{\\hat{\\mathbf{a}}} = \\{1, 0, 0\\}\\), empirically leads to poor recovered object edges because the balanced semantics and coarse saliency compete over where the underlying geometry is. Instead, we weight both equally \\(\\lambda_{\\hat{\\mathbf{s}}} = \\lambda_{\\hat{\\mathbf{a}}} = \\{1/9, 4/9, 4/9\\}\\). Even though the coarse layer has smaller weight, it is sufficient to guide the overall result. This balances high resolution edges from fine layers and whole object features from coarse layers while reducing geometry conflicts, and leads to improved features (Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Of course, any sliding window must contain an object to extract reliable features for that object. At coarse levels, an object is always in view. At fine levels, an object is only captured in some windows. Objects of interest tend to be near the middle of the frame, meaning that boundary windows at finer pyramid levels contain features that less reliably capture those objects. This can cause spurious connections in clustering. To cope with this, we relatively decrease finer level boundary window weights: We upsample all levels to the finest level, then increase the coarsest level weight towards the frame boundary to \\(\\lambda_{\\hat{\\mathbf{s}}} = \\lambda_{\\hat{\\mathbf{a}}} = \\{1/3, 1/3, 1/3\\}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "21800" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.09, + 0.207, + 0.099 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.1, + 0.259, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.09, + 0.368, + 0.098 + ], + "angle": 0, + "content": "(b) Semantics" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.1, + 0.409, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.432, + 0.09, + 0.539, + 0.099 + ], + "angle": 0, + "content": "(c) Semantics (volume)" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.1, + 0.559, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.09, + 0.662, + 0.099 + ], + "angle": 0, + "content": "(d) Saliency" + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.1, + 0.708, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.177, + 0.671, + 0.187 + ], + "angle": 0, + "content": "(h) Pyr. saliency" + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.09, + 0.834, + 0.099 + ], + "angle": 0, + "content": "(e) Saliency (volume)" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.1, + 0.857, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.177, + 0.375, + 0.187 + ], + "angle": 0, + "content": "(f) Pyr. semantics" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.187, + 0.409, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.412, + 0.177, + 0.558, + 0.187 + ], + "angle": 0, + "content": "(g) Our pyr. semantics (volume)" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.187, + 0.558, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.187, + 0.708, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.716, + 0.177, + 0.852, + 0.187 + ], + "angle": 0, + "content": "(i) Our pyr. saliency (volume)" + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.187, + 0.857, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.268, + 0.892, + 0.299 + ], + "angle": 0, + "content": "Figure 3: Semantics and saliency improve by both volume integration and by our pyramid. On Balloon NBoard, blocky artifacts are removed. Semantics are visualized as most significant three PCA dimensions; specific colors are less meaningful." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.315, + 0.441, + 0.332 + ], + "angle": 0, + "content": "3.4. Using SAFF for Saliency-aware Clustering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.47, + 0.506 + ], + "angle": 0, + "content": "We now wish to isolate salient objects. Even in dynamic scenes, relevant objects may not move, so analyzing dynamic elements is insufficient (cf. [40]). One approach predicts segmentation end-to-end [19]. However, end-to-end learning requires priors about the scene provided by supervision, and even large-scale pretraining might fail given unseen test scenes. To achieve scene-specific decompositions from per-video semantics, inspired by Amir et al. [1], we design clustering for spacetime volumes that allows segmenting novel spatio-temporal views. While DINO-ViT is trained on images, its features are loosely temporally consistent [6]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.471, + 0.612 + ], + "angle": 0, + "content": "Some works optimize a representation with a fixed number of clusters, e.g., via slot attention [20] in NeRFs [32, 46]. Instead, we cluster using elbow \\( k \\)-means, letting us adaptively find the number of clusters after optimization. This is more flexible than baking an anticipated number of slots (sometimes with fixed semantics), and lets us cluster and segment in novel spatio-temporal viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.75 + ], + "angle": 0, + "content": "We demonstrate clustering results in both 3D over time and on rendered volume 2D projections over time. Given the volume reconstruction, we might think that clustering directly in 3D would be better. But, monocular input with narrow baselines makes it challenging to precisely reconstruct geometries: Consider that depth integrated a long a ray can still be accurate even though geometry at specific 3D points may be inaccurate or 'fluffy'. As such, we use the 2D volume projection clustering results in 2D comparisons." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Method For 3D over time, we sample points from the SAFF uniformly along input rays \\((128\\times H\\times W)\\) and treat each pixel as a feature point, e.g., semantics are 64 dim. and saliency is 1 dim. For volume projection, we render to \\(N\\) input poses \\((N\\times H\\times W)\\), and treat each pixel as a feature point instead. In either case, we cluster all feature points together using elbow \\(k\\)-means to produce an initial set of separate regions. For each cluster \\(c\\), for each image, we calculate the mean attention of all feature points within the cluster \\(\\bar{\\mathbf{a}}_c\\). If \\(\\bar{\\mathbf{a}}_c > 0.07\\), then this" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.316, + 0.892, + 0.346 + ], + "angle": 0, + "content": "cluster is salient for this image. Finally, all feature points vote on saliency: if more than \\(70\\%\\) agree, the cluster is salient." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.347, + 0.893, + 0.468 + ], + "angle": 0, + "content": "Salient objects may still be split into semantic parts: e.g., in Fig. 4, the person's head/body are separated. Plus, unwanted background saliency may exist, e.g., input \\(\\hat{\\mathbf{a}}^\\dagger\\) is high for the teal graphic on the wall. As such, before saliency voting, we merge clusters whose centroids have a cosine similarity \\(>0.5\\). This reduces the first problem as heads and bodies are similar, and reduces the second problem as merging the graphic cluster into the background reduces its average saliency (Fig. 4)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.894, + 0.574 + ], + "angle": 0, + "content": "To extract an object from the 3D volume, we sample 3D points along each input ray, then ascribe the label from the semantically-closest centroid. All clusters not similar to the stored salient clusters are marked as background with zero density. For novel space-time views, we render feature images from the volume, then assign cluster labels to each pixel according to its similarity with stored input view centroids." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.588, + 0.631, + 0.605 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.659 + ], + "angle": 0, + "content": "We show the impact of adding semantic and saliency features through scene decomposition and foreground experiments. Our website contains supplemental videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.894, + 0.748 + ], + "angle": 0, + "content": "Data: NVIDIA Dynamic Scene Dataset (Masked) This data [45] has 8 sequences of 24 time steps formed from 12 cameras simultaneously capturing video. We manually annotate object masks for view and time step splits; we will release this data publicly. We define three data splits per sequence:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.753, + 0.892, + 0.798 + ], + "angle": 0, + "content": "1. Input: A monocular camera that moves position for every timestep is simulated from the input sequences; we use Yoon et al.'s input sequences [45]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.804, + 0.894, + 0.851 + ], + "angle": 0, + "content": "2. Fix Cam 0 (hold out): We fix the camera at position 0 as time plays, requiring novel view and time synthesis. \\(\\{(\\mathrm{cam}_0,\\mathrm{time}_i),i\\in [1,2,\\dots 23]\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.894, + 0.902 + ], + "angle": 0, + "content": "3. Fix Time 0 (hold out): We fix time at step 0 as the camera moves, requiring novel view and time synthesis. \\(\\{(\\mathrm{cam}_i,\\mathrm{time}_0),i\\in [1,2,\\ldots 11]\\}\\)" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.753, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.518, + 0.937 + ], + "angle": 0, + "content": "21801" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.09, + 0.213, + 0.099 + ], + "angle": 0, + "content": "(i) Input frame" + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.1, + 0.252, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.175, + 0.244, + 0.185 + ], + "angle": 0, + "content": "(ii) Input semantic/attention" + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.09, + 0.398, + 0.099 + ], + "angle": 0, + "content": "(a) Our pyr. semantic (volume)" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.1, + 0.403, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.175, + 0.397, + 0.185 + ], + "angle": 0, + "content": "(d) Our pyr. saliency (volume)" + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.09, + 0.565, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.428, + 0.175, + 0.545, + 0.185 + ], + "angle": 0, + "content": "(e) Our merged clustering" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.185, + 0.56, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.09, + 0.715, + 0.099 + ], + "angle": 0, + "content": "(c) Salient clusters (pre-merge)" + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.1, + 0.717, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.175, + 0.696, + 0.185 + ], + "angle": 0, + "content": "(f) Our final foreground" + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.186, + 0.716, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.09, + 0.854, + 0.099 + ], + "angle": 0, + "content": "(iii) Extraction (fixed time)" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.121, + 0.867, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.175, + 0.854, + 0.185 + ], + "angle": 0, + "content": "(iv) Extraction (fixed view)" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.191, + 0.867, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.228, + 0.867, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.265, + 0.892, + 0.296 + ], + "angle": 0, + "content": "Figure 4: Saliency-aware clustering improves decomposition. On Dynamic Face, the head and body are semantically and saliently different, but are mutually different from the background. This allows us to extract a time-varying 3D field of the object." + }, + { + "type": "image_caption", + "bbox": [ + 0.161, + 0.302, + 0.219, + 0.316 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.302, + 0.405, + 0.316 + ], + "angle": 0, + "content": "(b) DINO-ViT (2D)" + }, + { + "type": "image_caption", + "bbox": [ + 0.417, + 0.302, + 0.567, + 0.316 + ], + "angle": 0, + "content": "(c) ProposeReduce [19]" + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.302, + 0.718, + 0.316 + ], + "angle": 0, + "content": "(d) SAFF (Ours; input)" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.302, + 0.868, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.317, + 0.263, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.265, + 0.317, + 0.413, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.317, + 0.565, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.317, + 0.717, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.384, + 0.867, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.516, + 0.892, + 0.561 + ], + "angle": 0, + "content": "Figure 5: SAFF object segmentations show balanced quality while recovering a volumetric scene representation (e). Basic DINO-ViT produces low-quality segmentations and misses objects. A state-of-the-art 2D video learning method [19] sometimes has edge detail (Umbrella, legs) but othertimes misses detail and objects (Balloon NBoard)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.588, + 0.47, + 0.724 + ], + "angle": 0, + "content": "Data: DyCheck Dataset (Masked) For single-camera casual monocular data, we select a subset of Gao et al.'s DyCheck dataset [10]. We remove scenes without an obvious subject, e.g., where telling foreground from background is hard even for a human. We select haru-sit, space-out, spin, paper-windmill, and mochi-high-five. We uniformly sample 48 frames and manually annotate object masks. We take even frames as the input set and odd frames as the hold out set. We use the same hyperparameters between both datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.47, + 0.822 + ], + "angle": 0, + "content": "Metrics To assess clustering performance, we use the Adjusted Rand Index (ARI; \\([-1, 1]\\)). This compares the similarity of two assignments without label matching, where random assignment would score \\(\\approx 0\\). For foreground segmentation, we compute IoU (Jaccard), and for RGB quality we use PSNR, SSIM, and LPIPS." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.367, + 0.849 + ], + "angle": 0, + "content": "4.1. Comparisons including ablations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "SAFF (ours) We optimize upon the input split of each scene, and perform clustering to obtain object segmentations. To produce a foreground, we merge all salient objects." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.588, + 0.892, + 0.619 + ], + "angle": 0, + "content": "SAFF (3D) The same as above, but processed in 3D rather than on 2D volume projections." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.626, + 0.892, + 0.657 + ], + "angle": 0, + "content": "- \\(\\mathbf{w} / \\mathbf{pyr}\\lambda_{\\hat{\\mathbf{a}}} = \\{1,0,0\\}\\) Pyramid with only coarse saliency (Sec. 3.3) and balanced semantic weight across levels." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.665, + 0.892, + 0.696 + ], + "angle": 0, + "content": "—w/o pyr No pyramid (Sec. 3.3); we optimize with features and saliency extracted from the input image only." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.703, + 0.892, + 0.734 + ], + "angle": 0, + "content": "— w/o merge With pyramid, but we remove cluster merging inside the saliency-aware clustering algorithm." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.741, + 0.894, + 0.818 + ], + "angle": 0, + "content": "- w/blend \\( v \\) To compare generic dynamic segmentation to saliency segmentation, we use the static/dynamic weight instead of volume saliency to segment foreground objects. We set every pixel below the \\( 80\\% \\) \\( v \\) quantile in each image to be background, or otherwise foreground." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "— w/post process We add a step after the saliency-aware clustering to refine edges using a conditional random field (please see supplemental material for details). This gains significantly from the depth estimated via volume reconstruction, producing sharp and detailed edges." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.626, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "21802" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.076, + 0.216, + 0.09 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.091, + 0.26, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.076, + 0.574, + 0.09 + ], + "angle": 0, + "content": "(b) NSFF [18] blend \\(v\\) (c) \\(\\mathrm{D}^2\\mathrm{NeRF}\\) [40] blend \\(v\\)" + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.091, + 0.412, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.076, + 0.574, + 0.09 + ], + "angle": 0, + "content": "c) \\(\\mathrm{D}^2\\mathrm{NeRF}\\) [40] blend \\(v\\)" + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.091, + 0.568, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.076, + 0.708, + 0.089 + ], + "angle": 0, + "content": "(d) SAFF blend \\(v\\)" + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.091, + 0.722, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.726, + 0.076, + 0.871, + 0.09 + ], + "angle": 0, + "content": "(e) SAFF foregrounds" + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.091, + 0.872, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.227, + 0.894, + 0.273 + ], + "angle": 0, + "content": "Figure 6: Saliency improves foreground segmentation. Static/dynamic separations are not foreground segmentations, leading to limited use of dynamic NeRF models for downstream tasks. Minor improvements to dynamic blending weight \\( v \\) are seen in some sequences (Jumping) by adding the saliency head to the shared backbone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.287, + 0.471, + 0.378 + ], + "angle": 0, + "content": "NSFF [18] This method cannot produce semantic clusterings. While saliency and blend weight \\( v \\) have different meanings, if we compare our \\( v \\) to NSFF's, then we can see any impact of a shared backbone with attention heads upon the static/dynamic separation. We disable the supervised motion mask initialization in NSFF as our method does not use such information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.471, + 0.459 + ], + "angle": 0, + "content": "\\(\\mathbf{D}^{2}\\mathbf{N}\\mathbf{e}\\mathbf{R}\\mathbf{F}\\) [40] This method also cannot produce semantic clusterings. Over HyperNeRF [26], it adds a shadow field network and further losses to try to isolate objects into the dynamic NeRF over the separate static one. The paper also compares to NSFF without motion mask initialization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.471, + 0.557 + ], + "angle": 0, + "content": "DINO-ViT (2D) [1] We ignore the volume and pass 2D semantic and attention features into the clustering algorithm. This cannot apply to novel viewpoints. Instead, we evaluate the approach upon all multi-view color images—input and hold-out—whereas other methods must render hold-out views. With our added pyramid processing (Sec. 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.564, + 0.431, + 0.58 + ], + "angle": 0, + "content": "- w/o pyr No pyramid; upsample to input RGB size." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.587, + 0.472, + 0.738 + ], + "angle": 0, + "content": "ProposeReduce (2D) [19] As a general comparison point, we apply this state-of-the-art 2D video segmentation method. For object segmentation, we use a ProposeReduce network that was pretrained with supervision on YouTube-VIS 2019 [43] for instance segmentation. For foreground segmentation, we use weights pretrained on UVOS [5] intended specifically for unsupervised foreground segmentation. As ProposeReduce is only a 2D method, we provide it with hold-out images for splits with novel views rather than our approach that must render novel views at hold-out poses." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.747, + 0.182, + 0.763 + ], + "angle": 0, + "content": "4.2. Findings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Dynamic scene decomposition We separate the background and each foreground object individually (Tab. 1). The baseline 2D DINO-ViT method is improved by our pyramid approach. But, being only 2D, this fails to produce a consistent decomposition across novel spacetime views even when given ground truth hold-out images. This shows the value of a volume integration. Next, supervised ProposeReduce produces good results (Fig. 5), but sometimes misses salient objects or fails to join them behind occluding objects, and only" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.285, + 0.895, + 0.345 + ], + "angle": 0, + "content": "Table 1: Spacetime volume integration improves dynamic scene decomposition. Pyramid construction and cluster merging help quantitatively, and ours is comparable to SOTA supervised 2D video segmentation network ProposeReduce." + }, + { + "type": "table_caption", + "bbox": [ + 0.502, + 0.346, + 0.86, + 0.361 + ], + "angle": 0, + "content": "Metric: Adjusted Rand Index \\([-1, 1]\\), higher is better." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.361, + 0.89, + 0.546 + ], + "angle": 0, + "content": "
InputFix Cam 0Fix Time 0
ProposeReduce (2D)0.7250.7360.742
DINO-ViT (2D)0.5010.4950.321
w/o pyr帽子,hat0.4700.4640.346
SAFF (3D)0.5940.5780.566
SAFF (ours)0.6530.6340.625
w/ pyr λa = {1,0,0}0.6200.5980.592
w/o pyr帽子,hat0.5450.5320.521
w/o merge cluster0.5930.5740.563
w/post process0.7590.7330.735
w/oracle0.8340.8060.800
w/oracle + post process0.9220.8900.880
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.558, + 0.893, + 0.604 + ], + "angle": 0, + "content": "sometimes produces better edges than our method without post-processing as it can oversmooth edges. ProposeReduce also receives ground truth images in hold-out sets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.894, + 0.742 + ], + "angle": 0, + "content": "Instead, our approach must render hold-out views via the volume reconstruction. This produces more consistent segmentations through spacetime manipulations—this is the added value of volume integration through learned saliency and attention heads. Ablated components show the value of our pyramid step, its coarse-saliency-only variant, and the cluster merge and image post processing steps. Qualitatively, we see good detail (Fig. 5); post-processing additionally improves edge quality and removes small unwanted regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Oracle saliency We also include an oracle experiment where saliency voting is replaced by cluster selection based on ground truth masks. This experiment tells us what part of the performance gap lies with saliency itself, and what remains due to volume integration and cluster boundary issues. With oracle clusters, our decomposition performance is 0.8 ARI (Tab. 1) even in hold-out views. This shows that our existing cluster boundaries are accurate, and that accurate saliency for object selection is the larger remaining problem." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "21803" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.471, + 0.166 + ], + "angle": 0, + "content": "Table 2: Saliency improves foreground segmentation. Adding saliency also slightly aids how much static/dynamic blend weight \\( v \\) represents the foreground (cf. NSFF blend \\( v \\) to SAFF's). Here, ProposeReduce uses unsupervised training and data [5]. Metric: IoU/Jaccard ([0, 1], higher is better)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.17, + 0.45, + 0.299 + ], + "angle": 0, + "content": "
InputFix Cam 0Fix Time 0
ProposeReduce (2D)0.6090.4640.591
DINO-ViT (2D)0.3810.3820.357
NSFF — blend v0.3220.3090.268
D2NeRF— blend v0.4700.3340.269
SAFF — saliency0.6090.5890.572
— blend v0.3880.3800.329
— post process v0.7200.6940.679
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.309, + 0.471, + 0.4 + ], + "angle": 0, + "content": "Table 3: SAFF semantics generalize to DyCheck Dataset (Masked) Oracle saliency. ProposeReduce is given ground truth test frames, while SAFF must render them. SAFF is comparable to SOTA supervised 2D video segmentation network ProposeReduce. Metric: Adjusted Rand Index \\([-1, 1]\\), higher is better); IoU/Jaccard index \\([-1, 1]\\), higher is better)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.409, + 0.466, + 0.495 + ], + "angle": 0, + "content": "
ARIIoU
InputTestInputTest
ProposeReduce (2D)0.7610.7620.7610.762
DINO-ViT (2D)0.8010.8000.7970.797
SAFF (3D)0.9020.8200.9100.812
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.47, + 0.664 + ], + "angle": 0, + "content": "Foreground segmentation We simplify the problem and consider all objects as simply 'foreground' to compare to methods that do not produce per object masks. Here, the same trend continues (Tab. 2). We note more subtle improvements to static/dynamic blending weights when adding our additional feature heads to the backbone MLP, and overall show that adding top-down information helps produce more useful object masks. Qualitative results show whole objects in the foreground rather than broad regions of possible dynamics (NSFF) or broken objects \\((\\mathrm{D}^2\\mathrm{NeRF};\\) Fig. 6)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.471, + 0.793 + ], + "angle": 0, + "content": "DyCheck evaluation DyCheck often has close-up objects and, even with our selected sequences, saliency struggles to find foreground objects. Thus, we use oracle saliency. DINOViT and ProposeReduce are given test views while SAFF must render them. Quantitative (Tab. 3) and qualitative (Fig. 7) experiments show similar trends as before: ProposeReduce is good but may miss objects and fine details; SAFF may produce finer details and is more geometrically consistent." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.807, + 0.327, + 0.822 + ], + "angle": 0, + "content": "5. Discussion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "DINO-ViT features are not instance-aware (Fig. 8). This is in contrast to object-centric learning approaches that aim to identify individual objects. To represent these different approaches, we compare to a result from slot-attention based \\(\\mathrm{SAVi}++\\) [7]. This method trains on thousands of supervised" + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.089, + 0.895, + 0.099 + ], + "angle": 0, + "content": "Input ProposeReduce [19] SAFF (Ours; input) SAFF (Ours; novel)" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.1, + 0.599, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.197, + 0.6, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.1, + 0.698, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.197, + 0.698, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.1, + 0.797, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.197, + 0.797, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.8, + 0.1, + 0.896, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.8, + 0.197, + 0.896, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.296, + 0.894, + 0.372 + ], + "angle": 0, + "content": "Figure 7: SAFF can apply to DyCheck Dataset (Masked). SAFF is comparable to supervised large-scale 2D video learning [19]. Some fine details are improved (bottom, fingers) and novel views maintain their geometry (bottom, gap between legs occluded by head at input timestep)." + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.38, + 0.598, + 0.394 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.395, + 0.628, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.468, + 0.597, + 0.482 + ], + "angle": 0, + "content": "(d) Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.38, + 0.746, + 0.394 + ], + "angle": 0, + "content": "(b) \\(\\mathrm{SAV_i + + [7]}\\)" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.395, + 0.757, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.468, + 0.746, + 0.482 + ], + "angle": 0, + "content": "(e) SAFF salient" + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.38, + 0.858, + 0.392 + ], + "angle": 0, + "content": "(c) SAFF" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.394, + 0.887, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.468, + 0.887, + 0.482 + ], + "angle": 0, + "content": "(f) Salient+dynamic" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.482, + 0.629, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.482, + 0.754, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.759, + 0.482, + 0.884, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.54, + 0.894, + 0.586 + ], + "angle": 0, + "content": "Figure 8: Limitations. DINO-ViT is not instance-aware, causing unwanted grouping (c). Unwanted static objects may be salient; assuming salient objects are dynamic fixes this (f)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.595, + 0.892, + 0.656 + ], + "angle": 0, + "content": "MOVi [11] sequences with per-object masks, whereas we use generic pre-trained features and gain better edges from volume integration. Combining these two approaches could give accurate instance-level scene objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.893, + 0.838 + ], + "angle": 0, + "content": "DINO-ViT saliency may attend to unwanted regions. In Figure 8d-e, the static pillars could be isolated using scene flow. But, often our desired subjects do not move (cf. people in Umbrella or Balloon NBoard). For tasks or data that can assume that salient objects are dynamic, we use SAFF's 4D scene reconstruction to reject static-but-salient objects by merging clusters via scene flow: First, we project \\(\\mathbf{f}\\) over each timestep into each input camera pose—this simulates optical flow with a static camera. Clusters are marked as salient per image if mean flow magnitude per cluster \\(|\\bar{\\mathbf{p}}| > 0.07\\) and mean attention \\(\\bar{\\mathbf{a}}_c > 0.07\\). Finally, as before, a cluster is globally salient if \\(70\\%\\) of images agree (Fig. 8f)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements The CV community in New England for feedback, and funding from NSF CNS-2038897 and an Amazon Research Award. Eliot thanks a Randy F. Pausch '82 Undergraduate Summer Research Award at Brown CS." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "21804" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2021. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.158, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.472, + 0.282 + ], + "angle": 0, + "content": "[3] Kenneth Blomqvist, Lionel Ott, Jen Jen Chung, and Roland Y. Siegwart. Baking in the feature: Accelerating volumetric segmentation by rendering feature maps. ArXiv, abs/2209.12744, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.472, + 0.34 + ], + "angle": 0, + "content": "[4] Christopher P. Burgess, Loic Matthews, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matthew M. Botvinick, and Alexander Lerchner. Monet: Unsupervised scene decomposition and representation. ArXiv, abs/1901.11390, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.471, + 0.396 + ], + "angle": 0, + "content": "[5] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv:1905.00737, 2019. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.398, + 0.471, + 0.466 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.468, + 0.47, + 0.536 + ], + "angle": 0, + "content": "[7] Gamaeldin F. Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael C. Mozer, and Thomas Kipf. \\(\\mathrm{SAVi}++\\): Towards end-to-end object-centric learning from real-world videos. In Advances in Neural Information Processing Systems, 2022. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.538, + 0.471, + 0.592 + ], + "angle": 0, + "content": "[8] Zhiwen Fan, Peihao Wang, Yifan Jiang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation on complex scenes. ArXiv, abs/2209.08776, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.594, + 0.471, + 0.662 + ], + "angle": 0, + "content": "[9] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In International Conference on 3D Vision (3DV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.47, + 0.705 + ], + "angle": 0, + "content": "[10] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. In NeurIPS, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[11] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam H. Laradji, Hsueh-Ti Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: A scalable dataset generator. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3739-3751, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[12] Klaus Greff, Raphael Lopez Kaufman, Rishabh Kabra, Nicholas Watters, Christopher P. Burgess, Daniel Zoran, Loic" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "Matthew, Matthew M. Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative variational inference. In ICML, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.894, + 0.192 + ], + "angle": 0, + "content": "[13] Basile Van Hoorick, Purva Tendulka, Dídac Surís, Dennis Park, Simon Stent, and Carl Vondrick. Revealing occlusions with 4d neural fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3001-3011, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.194, + 0.894, + 0.262 + ], + "angle": 0, + "content": "[14] Rishabh Kabra, Daniel Zoran, Goker Erdogan, Loic Matthey, Antonia Creswell, Matthew M. Botvinick, Alexander Lerchner, and Christopher P. Burgess. Simone: View-invariant, temporally-abstracted object representations via unsupervised video decomposition. ArXiv, abs/2106.03849, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.894, + 0.332 + ], + "angle": 0, + "content": "[15] Thomas Kipf, Gamaleldin F. Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional Object-Centric Learning from Video. In International Conference on Learning Representations (ICLR), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.894, + 0.389 + ], + "angle": 0, + "content": "[16] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. In Advances in Neural Information Processing Systems, volume 35, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.894, + 0.475 + ], + "angle": 0, + "content": "[17] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J. Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12871-12881, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.894, + 0.545 + ], + "angle": 0, + "content": "[18] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.894, + 0.603 + ], + "angle": 0, + "content": "[19] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1719-1728, 2021. 1, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.894, + 0.672 + ], + "angle": 0, + "content": "[20] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. arXiv preprint arXiv:2006.15055, 2020.1,2,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.894, + 0.716 + ], + "angle": 0, + "content": "[21] Kirill Mazur, Edgar Sucar, and Andrew J. Davison. Feature-realistic neural fusion for real-time, open set scene understanding. ArXiv, abs/2210.03043, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.894, + 0.773 + ], + "angle": 0, + "content": "[22] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.894, + 0.83 + ], + "angle": 0, + "content": "[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[24] Tom Monnier, Elliot Vincent, Jean Ponce, and Mathieu Aubry. Unsupervised layered image decomposition into object prototypes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8640-8650, October 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "21805" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[25] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[26] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), dec 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.287 + ], + "angle": 0, + "content": "[27] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.471, + 0.328 + ], + "angle": 0, + "content": "[28] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[29] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, and Francesco Locatello. Bridging the gap to real-world object-centric learning. ArXiv, abs/2209.14860, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.439 + ], + "angle": 0, + "content": "[30] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In CVPRW, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.471, + 0.496 + ], + "angle": 0, + "content": "[31] Cameron Smith, Hong-Xing Yu, Sergey Zakharov, Frédo Durand, Joshua B. Tenenbaum, Jiajun Wu, and Vincent Sitzmann. Unsupervised discovery and composition of object light fields. ArXiv, abs/2205.03923, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.471, + 0.538 + ], + "angle": 0, + "content": "[32] Karl Stelzner, Kristian Kersting, and Adam R. Kosiorek. Decomposing 3d scenes into objects via unsupervised volume segmentation. ArXiv, abs/2104.01148, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.471, + 0.663 + ], + "angle": 0, + "content": "[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.471, + 0.706 + ], + "angle": 0, + "content": "[34] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[35] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12939-12950, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[36] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural Feature Fusion Fields: 3D distillation of self-supervised 2D image representations. In Proceedings of the International Conference on 3D Vision (3DV), 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[37] Vadim Tschernezki, Diane Larlus, and Andrea Vedaldi. NeuralDiff: Segmenting 3D objects that move in egocentric videos. In Proceedings of the International Conference on 3D Vision (3DV), 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.203 + ], + "angle": 0, + "content": "[38] Rishi Veerapaneni, John D. Co-Reyes, Michael Chang, Michael Janner, Chelsea Finn, Jiajun Wu, Joshua Tenenbaum, and Sergey Levine. Entity abstraction in visual model-based reinforcement learning. In Leslie Pack Kaelbling, Danica Kragic, and Komei Sugiura, editors, Proceedings of the Conference on Robot Learning, volume 100 of Proceedings of Machine Learning Research, pages 1439-1456. PMLR, 30 Oct-01 Nov 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.894, + 0.274 + ], + "angle": 0, + "content": "[39] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L. Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.275, + 0.893, + 0.33 + ], + "angle": 0, + "content": "[40] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Öz Tireli. D2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. ArXiv, abs/2205.15838, 2022. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.332, + 0.894, + 0.387 + ], + "angle": 0, + "content": "[41] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.389, + 0.893, + 0.457 + ], + "angle": 0, + "content": "[42] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for editable scene rendering. In International Conference on Computer Vision (ICCV), October 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.459, + 0.893, + 0.5 + ], + "angle": 0, + "content": "[43] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Oct 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.502, + 0.893, + 0.557 + ], + "angle": 0, + "content": "[44] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snavely. Deformable sprites for unsupervised video decomposition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.558, + 0.893, + 0.627 + ], + "angle": 0, + "content": "[45] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.629, + 0.893, + 0.67 + ], + "angle": 0, + "content": "[46] Hong-Xing Yu, Leonidas J. Guibas, and Jiajun Wu. Unsupervised discovery of object radiance fields. In International Conference on Learning Representations, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.894, + 0.727 + ], + "angle": 0, + "content": "[47] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "21806" + } + ] +] \ No newline at end of file diff --git a/2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/68ea1ef8-9cb5-44da-894d-92e02632af66_origin.pdf b/2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/68ea1ef8-9cb5-44da-894d-92e02632af66_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9d06d81fc9c9269e5eb8dd08d5d5c2e04246ae00 --- /dev/null +++ b/2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/68ea1ef8-9cb5-44da-894d-92e02632af66_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f077e0f38009353dabe84ae3054bdcfb093a393d4411d5040ebf97c7968ddb5 +size 7001137 diff --git a/2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/full.md b/2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..610d8022d30b3ee89d322be31f3674ce4446c56c --- /dev/null +++ b/2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/full.md @@ -0,0 +1,468 @@ +# Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition + +Yiqing Liang + +Eliot Laidlaw + +Alexander Meyerowitz Brown University + +Srinath Sridhar + +James Tompkin + +# Abstract + +From video, we reconstruct a neural volume that captures time-varying color, density, scene flow, semantics, and attention information. The semantics and attention let us identify salient foreground objects separately from the background across spacetime. To mitigate low resolution semantic and attention features, we compute pyramids that trade detail with whole-image context. After optimization, we perform a saliency-aware clustering to decompose the scene. To evaluate real-world scenes, we annotate object masks in the NVIDIA Dynamic Scene and DyCheck datasets. We demonstrate that this method can decompose dynamic scenes in an unsupervised way with competitive performance to a supervised method, and that it improves foreground/background segmentation over recent static/dynamic split methods. Project webpage: https://visual.cs.brown.edu/saff + +# 1. Introduction + +Given a casually-captured monocular RGB video of a dynamic scene, decomposing it into foreground objects and background is an important task in computer vision, with downstream applications in segmentation and video editing. An ideal method would also reconstruct the geometry and appearance over time including frame correspondences. + +Previous methods have made great progress but there are many limitations. Some works assume that there is no object motion in the scene [32, 46, 31, 16, 4], take input from multi-view cameras [16, 46], or do not explicitly reconstruct the underlying 3D structure of the scene [4, 7]. For objects, some works rely on masks or user input to aid segmentation [42, 15, 7], or use task-specific training datasets [7]. Sometimes, works assume the number of foreground objects [4, 20]. Given the challenges, many works train and test on synthetic data [14, 38]. + +We present Semantic Attention Flow Fields (SAFF): A method to overcome these limitations by integrating low-level reconstruction cues with high-level pretrained information—both bottom-up and top-down—into a neural volume. With this, we demonstrate that embedded semantic and saliency (attention) information is useful for unsupervised dynamic scene decomposition. SAFF builds upon neural scene flow fields [18], an approach that reconstructs appearance, geometry, and motion. This uses frame interpolation rather than explicit canonicalization [35] or a latent hyperspace [26], + +![](images/01fefb4a9e45aeceefff10089d5bae4b9533e86d532a1f53df7e2cc944385a37.jpg) +Density $\sigma$ , Saliency $a$ , Semantics $s$ , Color $c$ + +![](images/17f14ffcf60fce11f834a50b881e8ff330d777688d5b719bbc157a87104f5d21.jpg) +Novel view renderings with effects (depth of field, color adjustment) +Figure 1: We decompose a dynamic 3D scene from monocular input into time-varying color, density, scene flow, semantics, and attention information. This could be used to provide precise editing of objects or effects focused around objects. + +![](images/41eff40600d65b7aa1eeea11fea2b17c3386a06b748cc4ca9b57cb24afc10184.jpg) + +which lets it more easily apply to casual videos. For optimization, we supervise two network heads with pretrained DINO-ViT [6] semantic features and attention. Naively supervising high-resolution SAFF with low-resolution DINO-ViT output reduces reconstruction quality. To mitigate the mismatch, we build a semantic attention pyramid that trades detail with whole-image context. Having optimized a SAFF representation for a dynamic scene, we perform a saliency-aware clustering both in 3D and on rendered feature images to describe objects and their background. Given the volume reconstruction, the clustering generalizes to novel spacetime views. + +To evaluate SAFF's dynamic scene decomposition capacity, we expand the NVIDIA Dynamic Scene [45] and DyCheck [10] datasets by manually annotating object masks across input and hold-out views. We demonstrate that SAFF outperforms 2D DINO-ViT baselines and is comparable to a state-of-the-art video segmentation method ProposeReduce[19] on our data. Existing monocular video dynamic volume reconstruction methods typically separate static and dynamic parts, but these often do not represent meaningful foreground. We show improved foreground segmentation over NSFF and the current $\mathrm{D}^2\mathrm{NeRF}$ [40] method for downstream tasks like editing. + +# 2. Related Work + +Decomposing a scene into regions of interest is a long-studied task in computer vision [5], including class, instance, and panoptic segmentation in high-level or top-down vision, and use of low-level or bottom-up cues like motion. Recent progress has considered images [12, 4, 20], videos [14, 15], layer decomposition [44, 24], and in-the-wild databases using deep generative models [25]. One example, $\mathrm{SAVi}++[7]$ , uses slot attention [20] to define 2D objects in real-world videos. Providing first-frame segmentation masks achieves stable performance, with validation on the driving Waymo Open Dataset [33]. Our work attempts 3D scene decomposition for a casual monocular video without initial masks. + +Scene Decomposition with NeRFs Neural Radiance Fields (NeRF) [41] have spurred new scene decomposition research through volumes. ObSuRF [32] and uORF [46] are unsupervised slot attention works that bind a latent code to each object. Unsupervised decomposition is also possible on light fields [31]. For dynamic scenes, works like NeuralDiff [37] and $\mathrm{D}^2\mathrm{NeRF}$ [40] focus on foreground separation, where foreground is defined to contain moving objects. Other works like N3F [36] and occlusions-4d [13] also decompose foregrounds into individual objects. N3F requires user input to specify which object to segment, and occlusions-4d takes RGB point clouds as input. Our work attempts to recover a segmented dynamic scene from a monocular RGB video without added markup or masks. + +Neural Fields Beyond Radiance Research has begun to add non-color information to neural volumes to aid decomposition via additional feature heads on the MLP. iLabel [47] adds a semantic head to propagate user-provided segmentations in the volume. PNF [17] and Panoptic-NeRF [9] attempt panoptic segmentation within neural fields, and Object-NeRF integrates instance segmentation masks into the field during optimization [42]. Research also investigates how to apply generic pretrained features to neural fields, like DINO-ViT. + +DINO-ViT for Semantics and Saliency DINO-ViT is a self-supervised transformer that, after pretraining, extracts generic semantic information [6]. Amir et al. [1] use DINO-ViT features with $k$ -means clustering to achieve co-segmentation across a video. Seitzer et al. [29] combine slot attention and DINO-ViT features for object-centric learning on real-world 2D data. TokenCut [39] performs normalized cuts on DINO-ViT features for foreground segmentation on natural images. Deep Spectral Segmentation [22] show that graph Laplacian processing of DINO-ViT features provides unsupervised foreground segmentation, and Selfmask [30] shows that these features can provide object saliency masks. Our approach considers these clustering and saliency findings for the setting of 3D decomposition from monocular video. + +DINO-ViT Fields Concurrent works have integrated DINO-ViT features into neural fields. DFF [16] distills features for dense multi-view static scenes with user input for segmentation. N3F [36] expands NeuralDiff to dynamic scenes, and relies on user input for segmentation. AutoLabel [3] uses DINO-ViT features to accelerate segmentation in static scenes given a ground truth segmentation mask. Other works use DINO-ViT differently. FeatureRealistic-Fusion [21] uses DINO-ViT in an online feature fusion task, focusing on propagating user input, and NeRF-SOS [8] uses 2D DINO-ViT to process NeRF-rendered multi-view RGB images of a static scene. In contrast to these works, we consider real-world casual monocular videos, recover and decompose a 3D scene, then explore whether saliency can avoid the need for masks or user input in segmenting objects. + +# 3. Method + +For a baseline dynamic scene reconstruction method, we begin with NSFF from Li et al. [18] (Sec. 3.1), which builds upon NeRF [23]. NSFF's low-level scene flow frame-to-frame approach provides better reconstructions for real-world casual monocular videos than deformation-based methods [35, 26]. We modify the architecture to integrate higher-level semantic and saliency (or attention) features (Sec. 3.2). After optimizing a SAFF for each scene, we perform saliency-aware clustering of the field (Sec. 3.4). All implementation details are in our supplemental material. + +Input Our method takes in a single RGB video over time $i$ as an ordered set of images $I \in \mathcal{I}$ and camera poses. We use COLMAP to recover camera poses [28]. From all poses, we define an NDC-like space that bounds the scene, and a set of rays $\mathbf{r} \in \mathcal{R}$ , one per image pixel with color $\hat{\mathbf{c}}^\dagger$ . Here, $\hat{\mathbf{c}}$ denotes a 2D pixel value in contrast to a 3D field value, and $\hat{\mathbf{c}}^\dagger$ denotes an input value in contrast to an estimated value. + +From pretrained networks, we estimate single-frame monocular depth $\hat{d}^{\dagger}$ (MiDaSv2 [27]), optical flow $\hat{\mathbf{p}}_i^\dagger$ (RAFT [34]), and semantic features $\hat{\mathbf{s}}^\dagger$ and attention $\hat{\mathbf{a}}^\dagger$ (DINO-ViT [6]) after important preprocessing (Sec. 3.3). + +# 3.1. Initial Dynamic Neural Volume Representation + +The initial representation comprises a static NeRF $F_{\theta}^{\mathrm{st}}$ and a dynamic NeRF $F_{\theta}^{\mathrm{dy}}$ . The static network predicts a color $\mathbf{c}$ , density $\sigma$ , and blending weight $v$ (Eq. (1)), and the dynamic network predicts time-varying color $\mathbf{c}_i$ , density $\sigma_i$ , scene flow $\mathbf{f}_i$ , and occlusion weights $w_i$ (Eq. (2)). In both network architectures, other than position $\mathbf{x}$ , we add direction $\boldsymbol{\omega}$ to a late separate head to only condition the estimation of color $\mathbf{c}$ . + +$$ +F _ {\boldsymbol {\theta}} ^ {\mathrm {s t}}: (\mathbf {x}, \omega) \rightarrow \left(\mathbf {c} ^ {\mathrm {s t}}, \sigma^ {\mathrm {s t}}, v\right) \tag {1} +$$ + +$$ +F _ {\boldsymbol {\theta}} ^ {\mathrm {d y}}: (\mathbf {x}, \boldsymbol {\omega}, i) \rightarrow \left(\mathbf {c} _ {i} ^ {\mathrm {d y}}, \sigma_ {i} ^ {\mathrm {d y}}, \mathbf {f} _ {i}, w _ {i}\right) \tag {2} +$$ + +To produce a pixel's color, we sample points at distances $t$ along the ray $\mathbf{x}_t = \mathbf{x} - \omega t$ between near and far planes $t_n$ + +![](images/f11cce526596a99c53cb277dfee5c4ba6fe1f8e2e53d703a83c76fedd620f2ad.jpg) +(a) SAFF video inputs + +![](images/08f48acf6d9bc6c16bccd6de7bb8553c7cb55375f61b01b26ae38212d33b08ce.jpg) +(b) SAFF model + +![](images/8f5ce03d1d46c9edbba3fab841f598fd38e818f838815b8436f0ed4d1f04ab7e.jpg) +(c) Volume rendered outputs +Figure 2: Overview. From monocular video, SAFF builds a neural field of scene-flowed 3D density, radiance, semantics, and attention (b). This is guided by semantic attention pyramids that increase resolution, plus depth and optical flow priors (a). We can render new spacetime views of any channel, and use saliency-aware clustering to decompose objects and background (c). + +to $t_f$ , query each network, then integrate transmittance $T$ , density $\sigma$ , and color along the ray from these samples [23]. For brevity, we omit evaluating at $\mathbf{x}_t, \boldsymbol{\omega}$ , e.g., $\sigma^{\mathrm{st}}(\mathbf{x}_t)$ is simply $\sigma^{\mathrm{st}}$ ; $\mathbf{c}^{\mathrm{st}}(\mathbf{x}_t, \boldsymbol{\omega})$ is simply $\mathbf{c}^{\mathrm{st}}$ . + +We produce a combined color from the static and dynamic colors by multiplication with their densities: + +$$ +\sigma_ {i} \mathbf {c} _ {i} = v \sigma^ {\mathrm {s t}} \mathbf {c} ^ {\mathrm {s t}} + (1 - v) \sigma_ {i} ^ {\mathrm {d y}} \mathbf {c} _ {i} ^ {\mathrm {d y}} \tag {3} +$$ + +Given that transmittance integrates density up to the current sampled point under Beer-Lambert volume attenuation, the rendered pixel color for the ray is computed as: + +$$ +\hat {\mathbf {c}} _ {i} = \int_ {t _ {n}} ^ {t _ {f}} T _ {i} \sigma_ {i} \mathbf {c} _ {i} d t \text {w h e r e} T _ {i} = \exp \left(- \int_ {t _ {n}} ^ {t} \sigma_ {i} d t\right) \tag {4} +$$ + +To optimize the volume to reconstruct input images, we compute a photometric loss $L_{\hat{\mathbf{c}}}$ between rendered and input colors: + +$$ +\mathcal {L} _ {\hat {\mathbf {c}}} = \frac {1}{| \mathcal {R} |} \sum_ {\mathbf {r} _ {i} \in \mathcal {R}} \left\| \hat {\mathbf {c}} _ {i} \left(\mathbf {r} _ {i}\right) - \hat {\mathbf {c}} _ {i} ^ {\dagger} \left(\mathbf {r} _ {i}\right) \right\| _ {2} ^ {2} \tag {5} +$$ + +Scene Flow Defining correspondence over time is important for monocular input as it lets us penalize a reprojection error with neighboring frames $j \in \mathcal{N}$ , e.g., where $j = i + 1$ or $j = i - 1$ . We denote $i \rightarrow j$ for the projection of frame $i$ onto frame $j$ by scene flow. $F_{\theta}^{\mathrm{dy}}$ estimates both forwards and backwards scene flow at every point to penalize a bi-direction loss. + +Thus, we approximate color output at time step $j$ by flowing the queried network values at time step $i$ : + +$$ +\hat {\mathbf {c}} _ {i \rightarrow j} = \int_ {t _ {n}} ^ {t _ {f}} T _ {i \rightarrow j} \sigma_ {i \rightarrow j} \mathbf {c} _ {i \rightarrow j} d t \tag {6} +$$ + +Reprojection must account for occlusion and disocclusion by motion. As such, $F_{\theta}^{\mathrm{dy}}$ also predicts forwards and backwards + +scene occlusion weights $w_{i + 1}$ and $w_{i - 1} \in [0,1]$ , where a point with $w_{i + 1} = 0$ means that occlusion status is changed one step forwards in time. We can integrate $w$ to a pixel: + +$$ +\hat {w} _ {i \rightarrow j} = \int_ {t _ {n}} ^ {t _ {f}} T _ {i \rightarrow j} \sigma_ {i \rightarrow j} w _ {j} d t \tag {7} +$$ + +Then, this pixel weight modulates the color loss such that occluded pixels are ignored. (8) + +$$ +\mathcal {L} _ {\hat {\mathbf {c}} _ {i \rightarrow j}} = \frac {1}{| \mathcal {R} | | \mathcal {N} |} \sum_ {\mathbf {r} _ {i} \in \mathcal {R}} \sum_ {j \in \mathcal {N}} \hat {w} _ {i \rightarrow j} (\mathbf {r} _ {i}) | | \hat {\mathbf {c}} _ {i \rightarrow j} (\mathbf {r} _ {i}) - \hat {\mathbf {c}} _ {j} ^ {\dagger} (\mathbf {r} _ {j}) | | _ {2} ^ {2} +$$ + +Prior losses We use the pretrained depth and optical flow map losses to help overcome the ill-posed monocular reconstruction problem. These losses decay as optimization progresses to rely more and more on the optimized self-consistent geometry and scene flow. For geometry, we estimate a depth $\hat{d}_i$ for each ray $\mathbf{r}_i$ by replacing $\mathbf{c}_i$ in Eq. (4) by the distance $t$ along the ray. Transform $z$ estimates a scale and shift as the pretrained network produces only relative depth. + +$$ +\mathcal {L} _ {\hat {d}} = \frac {1}{| \mathcal {R} |} \sum_ {\mathbf {r} _ {i} \in \mathcal {R}} \left| \left| \hat {d} _ {i} - z \left(\hat {d} _ {i} ^ {\dagger}\right) \right| \right| _ {1} \tag {9} +$$ + +For motion, projecting scene flow to a camera lets us compare to the estimated optical flow. Each sample point along a ray $\mathbf{x}_i$ is advected to a point in the neighboring frame $\mathbf{x}_{i\rightarrow j}$ , then integrated to the neighboring camera plane to produce a 2D point offset $\hat{\mathbf{p}}_i(\mathbf{r}_i)$ . Then, we expect the difference in the start and end positions to match the prior: + +$$ +\mathcal {L} _ {\hat {\mathbf {p}}} = \frac {1}{| \mathcal {R} | | \mathcal {N} |} \sum_ {\mathbf {r} _ {i} \in \mathcal {R}} \sum_ {j \in \mathcal {N} (i)} | | \hat {\mathbf {p}} _ {i} (\mathbf {r} _ {i}) - \hat {\mathbf {p}} _ {i} ^ {\dagger} (\mathbf {r} _ {i}) | | _ {1} \tag {10} +$$ + +Additional regularizations encourage occlusion weights to be close to one, scene flow to be small, locally constant, and cyclically consistent, and blending weight $v$ to be sparse. + +# 3.2. Semantic Attention Flow Fields + +Beyond low-level or bottom-up features, high-level or top-down features are also useful to define objects and help down-stream tasks like segmentation. For example, methods like NSFF or $\mathrm{D}^2\mathrm{NeRF}$ struggle to provide useful separation of static and dynamic parts because blend weight $v$ estimates whether the volume appears to be occupied by some moving entity. This is not the same as objectness; tasks like video editing could benefit from accurate dynamic object masks. + +As such, we extract 2D semantic features and attention (or saliency) values from a pretrained DINO-ViT network, then optimize the SAFF such that unknown 3D semantic and attention features over time can be projected to recreate their 2D complements. This helps us to ascribe semantic meaning to the volume and to identify objects. As semantics/attention are integrated into the 4D volume, we can render them from novel spacetime views without further DINO-ViT computation. + +To estimate semantic features $\mathbf{s}$ and attention $\mathbf{a}$ at 3D points in the volume at time $i$ , we add two new heads to both the static $F_{\theta}^{\mathrm{st}}$ and the dynamic $F_{\theta}^{\mathrm{dy}}$ networks: + +$$ +F _ {\boldsymbol {\theta}} ^ {\mathrm {s t}}: (\mathbf {x}, \omega) \quad \rightarrow \left(\mathbf {c} ^ {\mathrm {s t}}, \sigma^ {\mathrm {s t}}, v, \mathbf {s} ^ {\mathrm {s t}}, \mathbf {a} ^ {\mathrm {s t}}\right) \tag {11} +$$ + +$$ +F _ {\boldsymbol {\theta}} ^ {\mathrm {d y}}: (\mathbf {x}, \boldsymbol {\omega}, i) \rightarrow \left(\mathbf {c} _ {i} ^ {\mathrm {d y}}, \sigma_ {i} ^ {\mathrm {d y}}, \mathbf {f} _ {i}, w _ {i}, \mathbf {s} _ {i} ^ {\mathrm {d y}}, \mathbf {a} _ {i} ^ {\mathrm {d y}}\right) \tag {12} +$$ + +As semantic features have been demonstrated to be somewhat robust to view dependence [1], in our architectures both heads for s, a are taken off the backbone before $\omega$ is injected. + +To render semantics from the volume, we replace the color term $\mathbf{c}$ in Eq. (4) with $\mathbf{s}$ and equivalently for $\mathbf{a}$ : + +$$ +\sigma_ {i} \mathbf {s} _ {i} = v \sigma^ {\mathrm {s t}} \mathbf {s} ^ {\mathrm {s t}} + (1 - v) \sigma_ {i} ^ {\mathrm {d y}} \mathbf {s} _ {i} ^ {\mathrm {d y}}, \hat {\mathbf {s}} _ {i} = \int_ {t _ {n}} ^ {t _ {f}} T _ {i} \sigma_ {i} \mathbf {s} _ {i} d t \tag {13} +$$ + +To encourage scene flow to respect semantics over time, we penalize complementary losses on s and a (showing s only): + +$$ +\mathcal {L} _ {\hat {\mathbf {s}} _ {i \rightarrow j}} = \frac {1}{| \mathcal {R} | | \mathcal {N} |} \sum_ {\mathbf {r} _ {i} \in \mathcal {R}} \sum_ {j \in \mathcal {N}} \hat {w} _ {i \rightarrow j} (\mathbf {r} _ {i}) \| \hat {\mathbf {s}} _ {i \rightarrow j} (\mathbf {r} _ {i}) - \hat {\mathbf {s}} _ {j} ^ {\dagger} (\mathbf {r} _ {j}) \| _ {2} ^ {2} \tag {14} +$$ + +Finally, as supervision, we add respective losses on the reconstruction of the 2D semantic and attention features from projected 3D volume points (showing s only): + +$$ +\mathcal {L} _ {\hat {\mathbf {s}}} = \frac {1}{| \mathcal {R} |} \sum_ {\mathbf {r} _ {i} \in \mathcal {R}} \left\| \hat {\mathbf {s}} _ {i} \left(\mathbf {r} _ {i}\right) - \hat {\mathbf {s}} _ {i} ^ {\dagger} \left(\mathbf {r} _ {i}\right) \right\| _ {2} ^ {2} \tag {15} +$$ + +Unlike depth and scene flow priors, these are not priors—there is no self-consistency for semantics to constrain their values. Thus, we do not decay their contribution. While decaying avoids disagreements between semantic and attention features and color-enforced scene geometry, it also leads to a loss of useful meaning (please see supplemental). + +Thus our final loss becomes: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {S A F F}} = \mathcal {L} _ {\hat {\mathbf {c}}} + \lambda_ {\hat {\mathbf {c}} _ {i \rightarrow j}} \mathcal {L} _ {\hat {\mathbf {c}} _ {i \rightarrow j}} + \lambda_ {\hat {d}} \mathcal {L} _ {\hat {d}} + \lambda_ {\hat {\mathbf {p}}} \mathcal {L} _ {\hat {\mathbf {p}}} \tag {16} \\ + \lambda_ {\hat {\mathbf {s}} _ {i \rightarrow j}} \mathcal {L} _ {\hat {\mathbf {s}} _ {i \rightarrow j}} + \lambda_ {\hat {\mathbf {a}} _ {i \rightarrow j}} \mathcal {L} _ {\hat {\mathbf {a}} _ {i \rightarrow j}} + \lambda_ {\hat {\mathbf {s}}} \mathcal {L} _ {\hat {\mathbf {s}}} + \lambda_ {\hat {\mathbf {a}}} \mathcal {L} _ {\hat {\mathbf {a}}} \\ \end{array} +$$ + +# 3.3. Semantic Attention Pyramids + +When thinking about scenes, we might argue that semantics from an ideal extractor should be scale invariant, as distant objects have the same class as close objects. We might also argue that saliency (or attention features) may not be scale invariant, as small details in a scene should only be salient when viewed close up. In practice, both extracted features vary across scale and have limited resolution, e.g., DINO-ViT [6] produces one output for each $8 \times 8$ patch. But, from this, we want semantic features and saliency for every RGB pixel that still respects scene boundaries. + +Thus far, work on static scenes has ignored the input/feature resolution mismatch [16] as multi-view constraints provide improved localization within the volume. For monocular video, this approach has limitations [36]. Forming many constraints on dynamic objects requires long-term motion correspondence—a tricky task—and so we want to maximize the resolution of any input features where possible without changing their meaning. + +One way may be through a pyramid of semantic and attention features that uses a sliding window approach at finer resolutions. Averaging features could increase detail around edges, but we must overcome the practical limit that these features are not stable across scales. This is especially important for saliency: unlike typical RGB pyramids that must preserve energy in an alias-free way [2], saliency changes significantly over scales and does not preserve energy. + +Consider a feature pyramid $\mathcal{P}$ with loss weights per level: + +$$ +\mathcal {L} _ {\mathcal {P} \hat {\mathbf {s}}} = \sum_ {i \in \mathcal {P}} \lambda_ {\hat {\mathbf {s}}} ^ {i} \mathcal {L} _ {\hat {\mathbf {s}}} ^ {i} \quad \mathcal {L} _ {\mathcal {P} \hat {\mathbf {a}}} = \sum_ {i \in \mathcal {P}} \lambda_ {\hat {\mathbf {a}}} ^ {i} \mathcal {L} _ {\hat {\mathbf {a}}} ^ {i} \tag {17} +$$ + +Naively encouraging scale-consistent semantics and whole-image saliency, e.g., $\lambda_{\hat{\mathbf{s}}} = \{1/3, 1/3, 1/3\}$ with $\lambda_{\hat{\mathbf{a}}} = \{1, 0, 0\}$ , empirically leads to poor recovered object edges because the balanced semantics and coarse saliency compete over where the underlying geometry is. Instead, we weight both equally $\lambda_{\hat{\mathbf{s}}} = \lambda_{\hat{\mathbf{a}}} = \{1/9, 4/9, 4/9\}$ . Even though the coarse layer has smaller weight, it is sufficient to guide the overall result. This balances high resolution edges from fine layers and whole object features from coarse layers while reducing geometry conflicts, and leads to improved features (Fig. 3). + +Of course, any sliding window must contain an object to extract reliable features for that object. At coarse levels, an object is always in view. At fine levels, an object is only captured in some windows. Objects of interest tend to be near the middle of the frame, meaning that boundary windows at finer pyramid levels contain features that less reliably capture those objects. This can cause spurious connections in clustering. To cope with this, we relatively decrease finer level boundary window weights: We upsample all levels to the finest level, then increase the coarsest level weight towards the frame boundary to $\lambda_{\hat{\mathbf{s}}} = \lambda_{\hat{\mathbf{a}}} = \{1/3, 1/3, 1/3\}$ . + +![](images/f178d6f08ab968f5f60529102575cf05b82ce1c9023a855566a217b585eb38d3.jpg) +(a) Input + +![](images/9b0a274bb513cdbaff27d2e781ba72fa72821d8b34d8ed9cb154a712427de7bd.jpg) +(b) Semantics + +![](images/ca4d6cc4b08415640f4ab2ce014bf87ff955f0bdf83c5692ffb80dd75522697d.jpg) +(c) Semantics (volume) + +![](images/286f833094b4816d0f26ea379f6c302a9a40006555badeca5a9734de5a57cb54.jpg) +(d) Saliency + +![](images/6073298e54a3d420c359a7e00e8a065e7a90a15f4532e6dd95c82ee33325699a.jpg) +(e) Saliency (volume) + +![](images/9a772ca97f6d34a23e449345bc01ca9b625edcc4b2d9d0c0e4c5afca1890d8ad.jpg) +(f) Pyr. semantics + +![](images/b1d357ef4844ecadfd229b10f4610f2f43a0e7ebec2c3b41c59afbd833772873.jpg) +(g) Our pyr. semantics (volume) +Figure 3: Semantics and saliency improve by both volume integration and by our pyramid. On Balloon NBoard, blocky artifacts are removed. Semantics are visualized as most significant three PCA dimensions; specific colors are less meaningful. + +![](images/db2932d48e70bef0de273f03533759cd8762fef47e784615bda46b636c792bcb.jpg) +(h) Pyr. saliency + +![](images/d9cad011c3b648a0848faa5ed6182fbf8a504b14c318a601d52790f011237b06.jpg) +(i) Our pyr. saliency (volume) + +# 3.4. Using SAFF for Saliency-aware Clustering + +We now wish to isolate salient objects. Even in dynamic scenes, relevant objects may not move, so analyzing dynamic elements is insufficient (cf. [40]). One approach predicts segmentation end-to-end [19]. However, end-to-end learning requires priors about the scene provided by supervision, and even large-scale pretraining might fail given unseen test scenes. To achieve scene-specific decompositions from per-video semantics, inspired by Amir et al. [1], we design clustering for spacetime volumes that allows segmenting novel spatio-temporal views. While DINO-ViT is trained on images, its features are loosely temporally consistent [6]. + +Some works optimize a representation with a fixed number of clusters, e.g., via slot attention [20] in NeRFs [32, 46]. Instead, we cluster using elbow $k$ -means, letting us adaptively find the number of clusters after optimization. This is more flexible than baking an anticipated number of slots (sometimes with fixed semantics), and lets us cluster and segment in novel spatio-temporal viewpoints. + +We demonstrate clustering results in both 3D over time and on rendered volume 2D projections over time. Given the volume reconstruction, we might think that clustering directly in 3D would be better. But, monocular input with narrow baselines makes it challenging to precisely reconstruct geometries: Consider that depth integrated a long a ray can still be accurate even though geometry at specific 3D points may be inaccurate or 'fluffy'. As such, we use the 2D volume projection clustering results in 2D comparisons. + +Method For 3D over time, we sample points from the SAFF uniformly along input rays $(128\times H\times W)$ and treat each pixel as a feature point, e.g., semantics are 64 dim. and saliency is 1 dim. For volume projection, we render to $N$ input poses $(N\times H\times W)$ , and treat each pixel as a feature point instead. In either case, we cluster all feature points together using elbow $k$ -means to produce an initial set of separate regions. For each cluster $c$ , for each image, we calculate the mean attention of all feature points within the cluster $\bar{\mathbf{a}}_c$ . If $\bar{\mathbf{a}}_c > 0.07$ , then this + +cluster is salient for this image. Finally, all feature points vote on saliency: if more than $70\%$ agree, the cluster is salient. + +Salient objects may still be split into semantic parts: e.g., in Fig. 4, the person's head/body are separated. Plus, unwanted background saliency may exist, e.g., input $\hat{\mathbf{a}}^\dagger$ is high for the teal graphic on the wall. As such, before saliency voting, we merge clusters whose centroids have a cosine similarity $>0.5$ . This reduces the first problem as heads and bodies are similar, and reduces the second problem as merging the graphic cluster into the background reduces its average saliency (Fig. 4). + +To extract an object from the 3D volume, we sample 3D points along each input ray, then ascribe the label from the semantically-closest centroid. All clusters not similar to the stored salient clusters are marked as background with zero density. For novel space-time views, we render feature images from the volume, then assign cluster labels to each pixel according to its similarity with stored input view centroids. + +# 4. Experiments + +We show the impact of adding semantic and saliency features through scene decomposition and foreground experiments. Our website contains supplemental videos. + +Data: NVIDIA Dynamic Scene Dataset (Masked) This data [45] has 8 sequences of 24 time steps formed from 12 cameras simultaneously capturing video. We manually annotate object masks for view and time step splits; we will release this data publicly. We define three data splits per sequence: + +1. Input: A monocular camera that moves position for every timestep is simulated from the input sequences; we use Yoon et al.'s input sequences [45]. +2. Fix Cam 0 (hold out): We fix the camera at position 0 as time plays, requiring novel view and time synthesis. $\{(\mathrm{cam}_0,\mathrm{time}_i),i\in [1,2,\dots 23]\}$ +3. Fix Time 0 (hold out): We fix time at step 0 as the camera moves, requiring novel view and time synthesis. $\{(\mathrm{cam}_i,\mathrm{time}_0),i\in [1,2,\ldots 11]\}$ + +![](images/0916a3dac95539f8c91aff80d1f8a07ab4e1ba67931550ee0b2cae13ffc89eda.jpg) +(i) Input frame +(ii) Input semantic/attention + +![](images/c5ce24e922c13f4821e3240ca0f0021f73e48b716f94b30e793395bc9e3a343b.jpg) +(a) Our pyr. semantic (volume) +(d) Our pyr. saliency (volume) + +![](images/5d567c20160ac212f558b1a90c90f18dce7c52124b47c9276e2fdd6d4ca11381.jpg) + +![](images/f462e2415f3c4072ff3dd65adef4b2717184e0446ea7a0119253bc18cc96ab9d.jpg) +(e) Our merged clustering +Figure 4: Saliency-aware clustering improves decomposition. On Dynamic Face, the head and body are semantically and saliently different, but are mutually different from the background. This allows us to extract a time-varying 3D field of the object. + +![](images/83fe39f2429a8082a84ae5b94ad37683b9859938142cb82fa7d83523de749201.jpg) +(c) Salient clusters (pre-merge) +(f) Our final foreground +(iii) Extraction (fixed time) + +![](images/5cc523c6ad484004e2d5657e5123e5dbf9683ce2a9478a19116031c1ea14a0d0.jpg) + +![](images/5cbf3a9ab54ff469420060cb15eb9ca62dcd439ff5e73fb19525cc2d36915a10.jpg) +(iv) Extraction (fixed view) + +![](images/d2014859e5f8596c4f049fd7a5542a994321f44e6e401f6f045a53168d008f72.jpg) + +![](images/e20bdacede3a7a5868b7cef4fcdcf593b23fa2655dfd89336a1d4e0a92985e35.jpg) + +![](images/09c1891a830a23fbc620a22ac5e9c58856b78f5efaea2cf8f1da9005971890f1.jpg) + +![](images/7def0160a5af8dfbc2da02f54623763411d8065e9bb925784ad0a2b1dd18007b.jpg) +(a) Input +Figure 5: SAFF object segmentations show balanced quality while recovering a volumetric scene representation (e). Basic DINO-ViT produces low-quality segmentations and misses objects. A state-of-the-art 2D video learning method [19] sometimes has edge detail (Umbrella, legs) but othertimes misses detail and objects (Balloon NBoard). + +![](images/4b13ac5c9c77ea60444552e5524c3e891235297addb0747db1403a807f583b6e.jpg) +(b) DINO-ViT (2D) + +![](images/6e89c0f7d9b9ee8d30bfbc5bec0ae618274ab0497d01755188af532b00ba8da6.jpg) +(c) ProposeReduce [19] + +![](images/a9ed21c21fb0dcc4c27f0d7b78c66d1cf4381b857645386272644700ab7fdba2.jpg) +(d) SAFF (Ours; input) + +![](images/8810bee4ce7e1f93f0f11ff2df60cc4844cb666696cdb3af51a551bec96b38da.jpg) + +Data: DyCheck Dataset (Masked) For single-camera casual monocular data, we select a subset of Gao et al.'s DyCheck dataset [10]. We remove scenes without an obvious subject, e.g., where telling foreground from background is hard even for a human. We select haru-sit, space-out, spin, paper-windmill, and mochi-high-five. We uniformly sample 48 frames and manually annotate object masks. We take even frames as the input set and odd frames as the hold out set. We use the same hyperparameters between both datasets. + +Metrics To assess clustering performance, we use the Adjusted Rand Index (ARI; $[-1, 1]$ ). This compares the similarity of two assignments without label matching, where random assignment would score $\approx 0$ . For foreground segmentation, we compute IoU (Jaccard), and for RGB quality we use PSNR, SSIM, and LPIPS. + +# 4.1. Comparisons including ablations + +SAFF (ours) We optimize upon the input split of each scene, and perform clustering to obtain object segmentations. To produce a foreground, we merge all salient objects. + +SAFF (3D) The same as above, but processed in 3D rather than on 2D volume projections. + +- $\mathbf{w} / \mathbf{pyr}\lambda_{\hat{\mathbf{a}}} = \{1,0,0\}$ Pyramid with only coarse saliency (Sec. 3.3) and balanced semantic weight across levels. +—w/o pyr No pyramid (Sec. 3.3); we optimize with features and saliency extracted from the input image only. +— w/o merge With pyramid, but we remove cluster merging inside the saliency-aware clustering algorithm. +- w/blend $v$ To compare generic dynamic segmentation to saliency segmentation, we use the static/dynamic weight instead of volume saliency to segment foreground objects. We set every pixel below the $80\%$ $v$ quantile in each image to be background, or otherwise foreground. +— w/post process We add a step after the saliency-aware clustering to refine edges using a conditional random field (please see supplemental material for details). This gains significantly from the depth estimated via volume reconstruction, producing sharp and detailed edges. + +![](images/36a63135f22edbeb9eb8b71e22f73a3ba4d8b599837e8409cfe50dd8f4c40fc9.jpg) +(a) Input +Figure 6: Saliency improves foreground segmentation. Static/dynamic separations are not foreground segmentations, leading to limited use of dynamic NeRF models for downstream tasks. Minor improvements to dynamic blending weight $v$ are seen in some sequences (Jumping) by adding the saliency head to the shared backbone. + +![](images/2cc18da6cca93a8b5c3df31bac7ab4722ea3c3d49eb0f8b46f3dc1a1f05dfb45.jpg) +(b) NSFF [18] blend $v$ (c) $\mathrm{D}^2\mathrm{NeRF}$ [40] blend $v$ + +c) $\mathrm{D}^2\mathrm{NeRF}$ [40] blend $v$ + +![](images/21cc1dd0431ac7f807f9d4cb3470d324905107b22fdef2f00b6fcc51388a164d.jpg) + +![](images/9e957d9085e676a91bfd7c9c567126cb875382a76823f9a1b0396fca531b832e.jpg) +(d) SAFF blend $v$ + +![](images/2fa19228ccb41766b7bf0778ab8bac6cf30332c73aef241eaefc80e232489399.jpg) +(e) SAFF foregrounds + +NSFF [18] This method cannot produce semantic clusterings. While saliency and blend weight $v$ have different meanings, if we compare our $v$ to NSFF's, then we can see any impact of a shared backbone with attention heads upon the static/dynamic separation. We disable the supervised motion mask initialization in NSFF as our method does not use such information. + +$\mathbf{D}^{2}\mathbf{N}\mathbf{e}\mathbf{R}\mathbf{F}$ [40] This method also cannot produce semantic clusterings. Over HyperNeRF [26], it adds a shadow field network and further losses to try to isolate objects into the dynamic NeRF over the separate static one. The paper also compares to NSFF without motion mask initialization. + +DINO-ViT (2D) [1] We ignore the volume and pass 2D semantic and attention features into the clustering algorithm. This cannot apply to novel viewpoints. Instead, we evaluate the approach upon all multi-view color images—input and hold-out—whereas other methods must render hold-out views. With our added pyramid processing (Sec. 3.3). + +- w/o pyr No pyramid; upsample to input RGB size. + +ProposeReduce (2D) [19] As a general comparison point, we apply this state-of-the-art 2D video segmentation method. For object segmentation, we use a ProposeReduce network that was pretrained with supervision on YouTube-VIS 2019 [43] for instance segmentation. For foreground segmentation, we use weights pretrained on UVOS [5] intended specifically for unsupervised foreground segmentation. As ProposeReduce is only a 2D method, we provide it with hold-out images for splits with novel views rather than our approach that must render novel views at hold-out poses. + +# 4.2. Findings + +Dynamic scene decomposition We separate the background and each foreground object individually (Tab. 1). The baseline 2D DINO-ViT method is improved by our pyramid approach. But, being only 2D, this fails to produce a consistent decomposition across novel spacetime views even when given ground truth hold-out images. This shows the value of a volume integration. Next, supervised ProposeReduce produces good results (Fig. 5), but sometimes misses salient objects or fails to join them behind occluding objects, and only + +Table 1: Spacetime volume integration improves dynamic scene decomposition. Pyramid construction and cluster merging help quantitatively, and ours is comparable to SOTA supervised 2D video segmentation network ProposeReduce. +Metric: Adjusted Rand Index $[-1, 1]$ , higher is better. + +
InputFix Cam 0Fix Time 0
ProposeReduce (2D)0.7250.7360.742
DINO-ViT (2D)0.5010.4950.321
w/o pyr帽子,hat0.4700.4640.346
SAFF (3D)0.5940.5780.566
SAFF (ours)0.6530.6340.625
w/ pyr λa = {1,0,0}0.6200.5980.592
w/o pyr帽子,hat0.5450.5320.521
w/o merge cluster0.5930.5740.563
w/post process0.7590.7330.735
w/oracle0.8340.8060.800
w/oracle + post process0.9220.8900.880
+ +sometimes produces better edges than our method without post-processing as it can oversmooth edges. ProposeReduce also receives ground truth images in hold-out sets. + +Instead, our approach must render hold-out views via the volume reconstruction. This produces more consistent segmentations through spacetime manipulations—this is the added value of volume integration through learned saliency and attention heads. Ablated components show the value of our pyramid step, its coarse-saliency-only variant, and the cluster merge and image post processing steps. Qualitatively, we see good detail (Fig. 5); post-processing additionally improves edge quality and removes small unwanted regions. + +Oracle saliency We also include an oracle experiment where saliency voting is replaced by cluster selection based on ground truth masks. This experiment tells us what part of the performance gap lies with saliency itself, and what remains due to volume integration and cluster boundary issues. With oracle clusters, our decomposition performance is 0.8 ARI (Tab. 1) even in hold-out views. This shows that our existing cluster boundaries are accurate, and that accurate saliency for object selection is the larger remaining problem. + +Table 2: Saliency improves foreground segmentation. Adding saliency also slightly aids how much static/dynamic blend weight $v$ represents the foreground (cf. NSFF blend $v$ to SAFF's). Here, ProposeReduce uses unsupervised training and data [5]. Metric: IoU/Jaccard ([0, 1], higher is better). + +
InputFix Cam 0Fix Time 0
ProposeReduce (2D)0.6090.4640.591
DINO-ViT (2D)0.3810.3820.357
NSFF — blend v0.3220.3090.268
D2NeRF— blend v0.4700.3340.269
SAFF — saliency0.6090.5890.572
— blend v0.3880.3800.329
— post process v0.7200.6940.679
+ +Table 3: SAFF semantics generalize to DyCheck Dataset (Masked) Oracle saliency. ProposeReduce is given ground truth test frames, while SAFF must render them. SAFF is comparable to SOTA supervised 2D video segmentation network ProposeReduce. Metric: Adjusted Rand Index $[-1, 1]$ , higher is better); IoU/Jaccard index $[-1, 1]$ , higher is better). + +
ARIIoU
InputTestInputTest
ProposeReduce (2D)0.7610.7620.7610.762
DINO-ViT (2D)0.8010.8000.7970.797
SAFF (3D)0.9020.8200.9100.812
+ +Foreground segmentation We simplify the problem and consider all objects as simply 'foreground' to compare to methods that do not produce per object masks. Here, the same trend continues (Tab. 2). We note more subtle improvements to static/dynamic blending weights when adding our additional feature heads to the backbone MLP, and overall show that adding top-down information helps produce more useful object masks. Qualitative results show whole objects in the foreground rather than broad regions of possible dynamics (NSFF) or broken objects $(\mathrm{D}^2\mathrm{NeRF};$ Fig. 6). + +DyCheck evaluation DyCheck often has close-up objects and, even with our selected sequences, saliency struggles to find foreground objects. Thus, we use oracle saliency. DINOViT and ProposeReduce are given test views while SAFF must render them. Quantitative (Tab. 3) and qualitative (Fig. 7) experiments show similar trends as before: ProposeReduce is good but may miss objects and fine details; SAFF may produce finer details and is more geometrically consistent. + +# 5. Discussion and Limitations + +DINO-ViT features are not instance-aware (Fig. 8). This is in contrast to object-centric learning approaches that aim to identify individual objects. To represent these different approaches, we compare to a result from slot-attention based $\mathrm{SAVi}++$ [7]. This method trains on thousands of supervised + +![](images/cdd7b3c6f9d1e1e1425caa90ae6f392fa17fe43167b54091676d627962928cfa.jpg) +Input ProposeReduce [19] SAFF (Ours; input) SAFF (Ours; novel) + +![](images/273f11278c90df905bb04230318be2162cc45f2dd6ab17510c1db4b067fbc468.jpg) +Figure 7: SAFF can apply to DyCheck Dataset (Masked). SAFF is comparable to supervised large-scale 2D video learning [19]. Some fine details are improved (bottom, fingers) and novel views maintain their geometry (bottom, gap between legs occluded by head at input timestep). + +![](images/e822e330344030109a8f27f3b32ab14a8a9a04c53f39ea6bf92763540d178a21.jpg) + +![](images/f118d418e44feb85ea722ffb224829fa95d77cb850e703874a4d2dcaedeb8466.jpg) + +![](images/41e9bf5dedef7a9aa4433f49ac7d51e411c80fd64d4ff68c6fd4a6c02c4dc620.jpg) + +![](images/d63c51f4f565fe03ce474325545808eb92b7f3beafc4605e703a8166445f1dad.jpg) + +![](images/662699dfb3e9528689d2f33db4a4c8da2c57f93c1e930c0ca44483f2db93a7b7.jpg) + +![](images/c894697ca2b290957c8e39c04bb2ec7be403cd2d62bcad781bbbe2c4acfe88a0.jpg) + +![](images/6729bdccffeb5d82405d77c40eecf02828d678f74eb3d84859773e7e5f4a06ba.jpg) +(a) Input + +![](images/988c7a0f5473de6d3d76cb7af8e50e1eb2966cae1592553828303c6e8169f2ae.jpg) +(b) $\mathrm{SAV_i + + [7]}$ + +![](images/e8844aa33ab7268e2befb4420a8bdff4f44183a0e86e2b9956937aa9eb001cd7.jpg) +(c) SAFF + +![](images/cc9c34111c72b5cd2873e5abb927357374f13c87078284374297bf3b4b5a6620.jpg) +(d) Input +Figure 8: Limitations. DINO-ViT is not instance-aware, causing unwanted grouping (c). Unwanted static objects may be salient; assuming salient objects are dynamic fixes this (f). + +![](images/d0b34e16db9794f6f8c60fa715c6dfee06a9337e01c9fc2539e48867a77d9b04.jpg) +(e) SAFF salient + +![](images/bfdbd6632afcaeb6b3475a898124129ca7c741b11ecfe7f74ba05018b82b489f.jpg) +(f) Salient+dynamic + +MOVi [11] sequences with per-object masks, whereas we use generic pre-trained features and gain better edges from volume integration. Combining these two approaches could give accurate instance-level scene objects. + +DINO-ViT saliency may attend to unwanted regions. In Figure 8d-e, the static pillars could be isolated using scene flow. But, often our desired subjects do not move (cf. people in Umbrella or Balloon NBoard). For tasks or data that can assume that salient objects are dynamic, we use SAFF's 4D scene reconstruction to reject static-but-salient objects by merging clusters via scene flow: First, we project $\mathbf{f}$ over each timestep into each input camera pose—this simulates optical flow with a static camera. Clusters are marked as salient per image if mean flow magnitude per cluster $|\bar{\mathbf{p}}| > 0.07$ and mean attention $\bar{\mathbf{a}}_c > 0.07$ . Finally, as before, a cluster is globally salient if $70\%$ of images agree (Fig. 8f). + +Acknowledgements The CV community in New England for feedback, and funding from NSF CNS-2038897 and an Amazon Research Award. Eliot thanks a Randy F. Pausch '82 Undergraduate Summer Research Award at Brown CS. + +# References + +[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2021. 2, 4, 5, 7 +[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 4 +[3] Kenneth Blomqvist, Lionel Ott, Jen Jen Chung, and Roland Y. Siegwart. Baking in the feature: Accelerating volumetric segmentation by rendering feature maps. ArXiv, abs/2209.12744, 2022. 2 +[4] Christopher P. Burgess, Loic Matthews, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matthew M. Botvinick, and Alexander Lerchner. Monet: Unsupervised scene decomposition and representation. ArXiv, abs/1901.11390, 2019. 1, 2 +[5] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv:1905.00737, 2019. 2, 7, 8 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 1, 2, 4, 5 +[7] Gamaeldin F. Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael C. Mozer, and Thomas Kipf. $\mathrm{SAVi}++$ : Towards end-to-end object-centric learning from real-world videos. In Advances in Neural Information Processing Systems, 2022. 1, 2, 8 +[8] Zhiwen Fan, Peihao Wang, Yifan Jiang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation on complex scenes. ArXiv, abs/2209.08776, 2022. 2 +[9] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In International Conference on 3D Vision (3DV), 2022. 2 +[10] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. In NeurIPS, 2022. 1, 6 +[11] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam H. Laradji, Hsueh-Ti Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: A scalable dataset generator. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3739-3751, 2022. 8 +[12] Klaus Greff, Raphael Lopez Kaufman, Rishabh Kabra, Nicholas Watters, Christopher P. Burgess, Daniel Zoran, Loic + +Matthew, Matthew M. Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative variational inference. In ICML, 2019. 2 +[13] Basile Van Hoorick, Purva Tendulka, Dídac Surís, Dennis Park, Simon Stent, and Carl Vondrick. Revealing occlusions with 4d neural fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3001-3011, 2022. 2 +[14] Rishabh Kabra, Daniel Zoran, Goker Erdogan, Loic Matthey, Antonia Creswell, Matthew M. Botvinick, Alexander Lerchner, and Christopher P. Burgess. Simone: View-invariant, temporally-abstracted object representations via unsupervised video decomposition. ArXiv, abs/2106.03849, 2021. 1, 2 +[15] Thomas Kipf, Gamaleldin F. Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional Object-Centric Learning from Video. In International Conference on Learning Representations (ICLR), 2022. 1, 2 +[16] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. In Advances in Neural Information Processing Systems, volume 35, 2022. 1, 2, 4 +[17] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J. Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12871-12881, June 2022. 2 +[18] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 7 +[19] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1719-1728, 2021. 1, 5, 6, 7, 8 +[20] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. arXiv preprint arXiv:2006.15055, 2020.1,2,5 +[21] Kirill Mazur, Edgar Sucar, and Andrew J. Davison. Feature-realistic neural fusion for real-time, open set scene understanding. ArXiv, abs/2210.03043, 2022. 2 +[22] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In CVPR, 2022. 2 +[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3 +[24] Tom Monnier, Elliot Vincent, Jean Ponce, and Mathieu Aubry. Unsupervised layered image decomposition into object prototypes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8640-8650, October 2021. 2 + +[25] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[26] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), dec 2021. 1, 2, 7 +[27] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022. 2 +[28] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2 +[29] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, and Francesco Locatello. Bridging the gap to real-world object-centric learning. ArXiv, abs/2209.14860, 2022. 2 +[30] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In CVPRW, 2022. 2 +[31] Cameron Smith, Hong-Xing Yu, Sergey Zakharov, Frédo Durand, Joshua B. Tenenbaum, Jiajun Wu, and Vincent Sitzmann. Unsupervised discovery and composition of object light fields. ArXiv, abs/2205.03923, 2022. 1, 2 +[32] Karl Stelzner, Kristian Kersting, and Adam R. Kosiorek. Decomposing 3d scenes into objects via unsupervised volume segmentation. ArXiv, abs/2104.01148, 2021. 1, 2, 5 +[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2 +[34] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 2 +[35] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12939-12950, 2021. 1, 2 +[36] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural Feature Fusion Fields: 3D distillation of self-supervised 2D image representations. In Proceedings of the International Conference on 3D Vision (3DV), 2022. 2, 4 +[37] Vadim Tschernezki, Diane Larlus, and Andrea Vedaldi. NeuralDiff: Segmenting 3D objects that move in egocentric videos. In Proceedings of the International Conference on 3D Vision (3DV), 2021. 2 + +[38] Rishi Veerapaneni, John D. Co-Reyes, Michael Chang, Michael Janner, Chelsea Finn, Jiajun Wu, Joshua Tenenbaum, and Sergey Levine. Entity abstraction in visual model-based reinforcement learning. In Leslie Pack Kaelbling, Danica Kragic, and Komei Sugiura, editors, Proceedings of the Conference on Robot Learning, volume 100 of Proceedings of Machine Learning Research, pages 1439-1456. PMLR, 30 Oct-01 Nov 2020. 1 +[39] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L. Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In Conference on Computer Vision and Pattern Recognition, 2022. 2 +[40] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Öz Tireli. D2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. ArXiv, abs/2205.15838, 2022. 1, 2, 5, 7 +[41] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 2022. 2 +[42] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for editable scene rendering. In International Conference on Computer Vision (ICCV), October 2021. 1, 2 +[43] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Oct 2019. 7 +[44] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snavely. Deformable sprites for unsupervised video decomposition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 2 +[45] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 5 +[46] Hong-Xing Yu, Leonidas J. Guibas, and Jiajun Wu. Unsupervised discovery of object radiance fields. In International Conference on Learning Representations, 2022. 1, 2, 5 +[47] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 2 \ No newline at end of file diff --git a/2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/images.zip b/2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..41cc4f91d89ea9436dabd3ad8a9bbe5bd33387cc --- /dev/null +++ b/2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2fc7d860933a12e1a8ce3b81aa3c2f6be10b5151457d0988f0cbd17074d573cd +size 672479 diff --git a/2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/layout.json b/2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d1b6e3b5c0c63c6e8b100ff5e669dce6c1f05f99 --- /dev/null +++ b/2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/layout.json @@ -0,0 +1,12398 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 81, + 533, + 99 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 81, + 533, + 99 + ], + "spans": [ + { + "bbox": [ + 59, + 81, + 533, + 99 + ], + "type": "text", + "content": "Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 114, + 136, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 114, + 136, + 129 + ], + "spans": [ + { + "bbox": [ + 70, + 114, + 136, + 129 + ], + "type": "text", + "content": "Yiqing Liang" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 114, + 217, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 114, + 217, + 128 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 217, + 128 + ], + "type": "text", + "content": "Eliot Laidlaw" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 231, + 114, + 343, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 114, + 343, + 143 + ], + "spans": [ + { + "bbox": [ + 231, + 114, + 343, + 143 + ], + "type": "text", + "content": "Alexander Meyerowitz Brown University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 356, + 114, + 432, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 114, + 432, + 127 + ], + "spans": [ + { + "bbox": [ + 356, + 114, + 432, + 127 + ], + "type": "text", + "content": "Srinath Sridhar" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 445, + 114, + 523, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 445, + 114, + 523, + 129 + ], + "spans": [ + { + "bbox": [ + 445, + 114, + 523, + 129 + ], + "type": "text", + "content": "James Tompkin" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 166, + 192, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 166, + 192, + 179 + ], + "spans": [ + { + "bbox": [ + 143, + 166, + 192, + 179 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 189, + 290, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 189, + 290, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 290, + 369 + ], + "type": "text", + "content": "From video, we reconstruct a neural volume that captures time-varying color, density, scene flow, semantics, and attention information. The semantics and attention let us identify salient foreground objects separately from the background across spacetime. To mitigate low resolution semantic and attention features, we compute pyramids that trade detail with whole-image context. After optimization, we perform a saliency-aware clustering to decompose the scene. To evaluate real-world scenes, we annotate object masks in the NVIDIA Dynamic Scene and DyCheck datasets. We demonstrate that this method can decompose dynamic scenes in an unsupervised way with competitive performance to a supervised method, and that it improves foreground/background segmentation over recent static/dynamic split methods. Project webpage: https://visual.cs.brown.edu/saff" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 380, + 127, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 127, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 127, + 393 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 399, + 288, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 288, + 472 + ], + "type": "text", + "content": "Given a casually-captured monocular RGB video of a dynamic scene, decomposing it into foreground objects and background is an important task in computer vision, with downstream applications in segmentation and video editing. An ideal method would also reconstruct the geometry and appearance over time including frame correspondences." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 473, + 288, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 288, + 592 + ], + "type": "text", + "content": "Previous methods have made great progress but there are many limitations. Some works assume that there is no object motion in the scene [32, 46, 31, 16, 4], take input from multi-view cameras [16, 46], or do not explicitly reconstruct the underlying 3D structure of the scene [4, 7]. For objects, some works rely on masks or user input to aid segmentation [42, 15, 7], or use task-specific training datasets [7]. Sometimes, works assume the number of foreground objects [4, 20]. Given the challenges, many works train and test on synthetic data [14, 38]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": "We present Semantic Attention Flow Fields (SAFF): A method to overcome these limitations by integrating low-level reconstruction cues with high-level pretrained information—both bottom-up and top-down—into a neural volume. With this, we demonstrate that embedded semantic and saliency (attention) information is useful for unsupervised dynamic scene decomposition. SAFF builds upon neural scene flow fields [18], an approach that reconstructs appearance, geometry, and motion. This uses frame interpolation rather than explicit canonicalization [35] or a latent hyperspace [26]," + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 313, + 157, + 538, + 289 + ], + "blocks": [ + { + "bbox": [ + 313, + 157, + 538, + 289 + ], + "lines": [ + { + "bbox": [ + 313, + 157, + 538, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 538, + 289 + ], + "type": "image", + "image_path": "01fefb4a9e45aeceefff10089d5bae4b9533e86d532a1f53df7e2cc944385a37.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 290, + 457, + 300 + ], + "lines": [ + { + "bbox": [ + 316, + 290, + 457, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 290, + 457, + 300 + ], + "type": "text", + "content": "Density " + }, + { + "bbox": [ + 316, + 290, + 457, + 300 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 316, + 290, + 457, + 300 + ], + "type": "text", + "content": ", Saliency " + }, + { + "bbox": [ + 316, + 290, + 457, + 300 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 316, + 290, + 457, + 300 + ], + "type": "text", + "content": ", Semantics " + }, + { + "bbox": [ + 316, + 290, + 457, + 300 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 316, + 290, + 457, + 300 + ], + "type": "text", + "content": ", Color " + }, + { + "bbox": [ + 316, + 290, + 457, + 300 + ], + "type": "inline_equation", + "content": "c" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 314, + 303, + 425, + 361 + ], + "blocks": [ + { + "bbox": [ + 314, + 303, + 425, + 361 + ], + "lines": [ + { + "bbox": [ + 314, + 303, + 425, + 361 + ], + "spans": [ + { + "bbox": [ + 314, + 303, + 425, + 361 + ], + "type": "image", + "image_path": "17f14ffcf60fce11f834a50b881e8ff330d777688d5b719bbc157a87104f5d21.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 361, + 534, + 372 + ], + "lines": [ + { + "bbox": [ + 314, + 361, + 534, + 372 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 534, + 372 + ], + "type": "text", + "content": "Novel view renderings with effects (depth of field, color adjustment)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 377, + 547, + 426 + ], + "lines": [ + { + "bbox": [ + 305, + 377, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 377, + 547, + 426 + ], + "type": "text", + "content": "Figure 1: We decompose a dynamic 3D scene from monocular input into time-varying color, density, scene flow, semantics, and attention information. This could be used to provide precise editing of objects or effects focused around objects." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 425, + 303, + 536, + 361 + ], + "blocks": [ + { + "bbox": [ + 425, + 303, + 536, + 361 + ], + "lines": [ + { + "bbox": [ + 425, + 303, + 536, + 361 + ], + "spans": [ + { + "bbox": [ + 425, + 303, + 536, + 361 + ], + "type": "image", + "image_path": "41eff40600d65b7aa1eeea11fea2b17c3386a06b748cc4ca9b57cb24afc10184.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 437, + 547, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 569 + ], + "type": "text", + "content": "which lets it more easily apply to casual videos. For optimization, we supervise two network heads with pretrained DINO-ViT [6] semantic features and attention. Naively supervising high-resolution SAFF with low-resolution DINO-ViT output reduces reconstruction quality. To mitigate the mismatch, we build a semantic attention pyramid that trades detail with whole-image context. Having optimized a SAFF representation for a dynamic scene, we perform a saliency-aware clustering both in 3D and on rendered feature images to describe objects and their background. Given the volume reconstruction, the clustering generalizes to novel spacetime views." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "To evaluate SAFF's dynamic scene decomposition capacity, we expand the NVIDIA Dynamic Scene [45] and DyCheck [10] datasets by manually annotating object masks across input and hold-out views. We demonstrate that SAFF outperforms 2D DINO-ViT baselines and is comparable to a state-of-the-art video segmentation method ProposeReduce[19] on our data. Existing monocular video dynamic volume reconstruction methods typically separate static and dynamic parts, but these often do not represent meaningful foreground. We show improved foreground segmentation over NSFF and the current " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " [40] method for downstream tasks like editing." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "21797" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 133, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 133, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 133, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 89, + 288, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 89, + 288, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 89, + 288, + 232 + ], + "type": "text", + "content": "Decomposing a scene into regions of interest is a long-studied task in computer vision [5], including class, instance, and panoptic segmentation in high-level or top-down vision, and use of low-level or bottom-up cues like motion. Recent progress has considered images [12, 4, 20], videos [14, 15], layer decomposition [44, 24], and in-the-wild databases using deep generative models [25]. One example, " + }, + { + "bbox": [ + 46, + 89, + 288, + 232 + ], + "type": "inline_equation", + "content": "\\mathrm{SAVi}++[7]" + }, + { + "bbox": [ + 46, + 89, + 288, + 232 + ], + "type": "text", + "content": ", uses slot attention [20] to define 2D objects in real-world videos. Providing first-frame segmentation masks achieves stable performance, with validation on the driving Waymo Open Dataset [33]. Our work attempts 3D scene decomposition for a casual monocular video without initial masks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 246, + 288, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 288, + 413 + ], + "type": "text", + "content": "Scene Decomposition with NeRFs Neural Radiance Fields (NeRF) [41] have spurred new scene decomposition research through volumes. ObSuRF [32] and uORF [46] are unsupervised slot attention works that bind a latent code to each object. Unsupervised decomposition is also possible on light fields [31]. For dynamic scenes, works like NeuralDiff [37] and " + }, + { + "bbox": [ + 46, + 246, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 46, + 246, + 288, + 413 + ], + "type": "text", + "content": " [40] focus on foreground separation, where foreground is defined to contain moving objects. Other works like N3F [36] and occlusions-4d [13] also decompose foregrounds into individual objects. N3F requires user input to specify which object to segment, and occlusions-4d takes RGB point clouds as input. Our work attempts to recover a segmented dynamic scene from a monocular RGB video without added markup or masks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 426, + 287, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 535 + ], + "type": "text", + "content": "Neural Fields Beyond Radiance Research has begun to add non-color information to neural volumes to aid decomposition via additional feature heads on the MLP. iLabel [47] adds a semantic head to propagate user-provided segmentations in the volume. PNF [17] and Panoptic-NeRF [9] attempt panoptic segmentation within neural fields, and Object-NeRF integrates instance segmentation masks into the field during optimization [42]. Research also investigates how to apply generic pretrained features to neural fields, like DINO-ViT." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 548, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 288, + 715 + ], + "type": "text", + "content": "DINO-ViT for Semantics and Saliency DINO-ViT is a self-supervised transformer that, after pretraining, extracts generic semantic information [6]. Amir et al. [1] use DINO-ViT features with " + }, + { + "bbox": [ + 46, + 548, + 288, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 548, + 288, + 715 + ], + "type": "text", + "content": "-means clustering to achieve co-segmentation across a video. Seitzer et al. [29] combine slot attention and DINO-ViT features for object-centric learning on real-world 2D data. TokenCut [39] performs normalized cuts on DINO-ViT features for foreground segmentation on natural images. Deep Spectral Segmentation [22] show that graph Laplacian processing of DINO-ViT features provides unsupervised foreground segmentation, and Selfmask [30] shows that these features can provide object saliency masks. Our approach considers these clustering and saliency findings for the setting of 3D decomposition from monocular video." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "text", + "content": "DINO-ViT Fields Concurrent works have integrated DINO-ViT features into neural fields. DFF [16] distills features for dense multi-view static scenes with user input for segmentation. N3F [36] expands NeuralDiff to dynamic scenes, and relies on user input for segmentation. AutoLabel [3] uses DINO-ViT features to accelerate segmentation in static scenes given a ground truth segmentation mask. Other works use DINO-ViT differently. FeatureRealistic-Fusion [21] uses DINO-ViT in an online feature fusion task, focusing on propagating user input, and NeRF-SOS [8] uses 2D DINO-ViT to process NeRF-rendered multi-view RGB images of a static scene. In contrast to these works, we consider real-world casual monocular videos, recover and decompose a 3D scene, then explore whether saliency can avoid the need for masks or user input in segmenting objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 261, + 361, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 261, + 361, + 273 + ], + "spans": [ + { + "bbox": [ + 306, + 261, + 361, + 273 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 277, + 547, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 547, + 397 + ], + "type": "text", + "content": "For a baseline dynamic scene reconstruction method, we begin with NSFF from Li et al. [18] (Sec. 3.1), which builds upon NeRF [23]. NSFF's low-level scene flow frame-to-frame approach provides better reconstructions for real-world casual monocular videos than deformation-based methods [35, 26]. We modify the architecture to integrate higher-level semantic and saliency (or attention) features (Sec. 3.2). After optimizing a SAFF for each scene, we perform saliency-aware clustering of the field (Sec. 3.4). All implementation details are in our supplemental material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 406, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 406, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 547, + 491 + ], + "type": "text", + "content": "Input Our method takes in a single RGB video over time " + }, + { + "bbox": [ + 304, + 406, + 547, + 491 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 406, + 547, + 491 + ], + "type": "text", + "content": " as an ordered set of images " + }, + { + "bbox": [ + 304, + 406, + 547, + 491 + ], + "type": "inline_equation", + "content": "I \\in \\mathcal{I}" + }, + { + "bbox": [ + 304, + 406, + 547, + 491 + ], + "type": "text", + "content": " and camera poses. We use COLMAP to recover camera poses [28]. From all poses, we define an NDC-like space that bounds the scene, and a set of rays " + }, + { + "bbox": [ + 304, + 406, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{r} \\in \\mathcal{R}" + }, + { + "bbox": [ + 304, + 406, + 547, + 491 + ], + "type": "text", + "content": ", one per image pixel with color " + }, + { + "bbox": [ + 304, + 406, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}}^\\dagger" + }, + { + "bbox": [ + 304, + 406, + 547, + 491 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 406, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}}" + }, + { + "bbox": [ + 304, + 406, + 547, + 491 + ], + "type": "text", + "content": " denotes a 2D pixel value in contrast to a 3D field value, and " + }, + { + "bbox": [ + 304, + 406, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}}^\\dagger" + }, + { + "bbox": [ + 304, + 406, + 547, + 491 + ], + "type": "text", + "content": " denotes an input value in contrast to an estimated value." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "text", + "content": "From pretrained networks, we estimate single-frame monocular depth " + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\hat{d}^{\\dagger}" + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "text", + "content": " (MiDaSv2 [27]), optical flow " + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}_i^\\dagger" + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "text", + "content": " (RAFT [34]), and semantic features " + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{s}}^\\dagger" + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "text", + "content": " and attention " + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}^\\dagger" + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "text", + "content": " (DINO-ViT [6]) after important preprocessing (Sec. 3.3)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 544, + 545, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 544, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 305, + 544, + 545, + 558 + ], + "type": "text", + "content": "3.1. Initial Dynamic Neural Volume Representation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "text", + "content": "The initial representation comprises a static NeRF " + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "inline_equation", + "content": "F_{\\theta}^{\\mathrm{st}}" + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "text", + "content": " and a dynamic NeRF " + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "inline_equation", + "content": "F_{\\theta}^{\\mathrm{dy}}" + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "text", + "content": ". The static network predicts a color " + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "text", + "content": ", density " + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "text", + "content": ", and blending weight " + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "text", + "content": " (Eq. (1)), and the dynamic network predicts time-varying color " + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "text", + "content": ", density " + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "inline_equation", + "content": "\\sigma_i" + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "text", + "content": ", scene flow " + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "text", + "content": ", and occlusion weights " + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "text", + "content": " (Eq. (2)). In both network architectures, other than position " + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "text", + "content": ", we add direction " + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\omega}" + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "text", + "content": " to a late separate head to only condition the estimation of color " + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 560, + 547, + 645 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 356, + 652, + 545, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 652, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 356, + 652, + 545, + 666 + ], + "type": "interline_equation", + "content": "F _ {\\boldsymbol {\\theta}} ^ {\\mathrm {s t}}: (\\mathbf {x}, \\omega) \\rightarrow \\left(\\mathbf {c} ^ {\\mathrm {s t}}, \\sigma^ {\\mathrm {s t}}, v\\right) \\tag {1}", + "image_path": "ed14993eeada3c690ff711f95c74140258763e4192313cdce24d0835d5a8f59d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 356, + 668, + 545, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 668, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 356, + 668, + 545, + 683 + ], + "type": "interline_equation", + "content": "F _ {\\boldsymbol {\\theta}} ^ {\\mathrm {d y}}: (\\mathbf {x}, \\boldsymbol {\\omega}, i) \\rightarrow \\left(\\mathbf {c} _ {i} ^ {\\mathrm {d y}}, \\sigma_ {i} ^ {\\mathrm {d y}}, \\mathbf {f} _ {i}, w _ {i}\\right) \\tag {2}", + "image_path": "4eb1c18c232ade8b57e043ec0524d5d2edba25b0b84ffefd03655a24791d092e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "To produce a pixel's color, we sample points at distances " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " along the ray " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = \\mathbf{x} - \\omega t" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " between near and far planes " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "t_n" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "21798" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 84, + 214, + 240 + ], + "blocks": [ + { + "bbox": [ + 86, + 71, + 185, + 83 + ], + "lines": [ + { + "bbox": [ + 86, + 71, + 185, + 83 + ], + "spans": [ + { + "bbox": [ + 86, + 71, + 185, + 83 + ], + "type": "text", + "content": "(a) SAFF video inputs" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 84, + 214, + 240 + ], + "lines": [ + { + "bbox": [ + 53, + 84, + 214, + 240 + ], + "spans": [ + { + "bbox": [ + 53, + 84, + 214, + 240 + ], + "type": "image", + "image_path": "f11cce526596a99c53cb277dfee5c4ba6fe1f8e2e53d703a83c76fedd620f2ad.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 230, + 89, + 351, + 238 + ], + "blocks": [ + { + "bbox": [ + 255, + 71, + 325, + 83 + ], + "lines": [ + { + "bbox": [ + 255, + 71, + 325, + 83 + ], + "spans": [ + { + "bbox": [ + 255, + 71, + 325, + 83 + ], + "type": "text", + "content": "(b) SAFF model" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 230, + 89, + 351, + 238 + ], + "lines": [ + { + "bbox": [ + 230, + 89, + 351, + 238 + ], + "spans": [ + { + "bbox": [ + 230, + 89, + 351, + 238 + ], + "type": "image", + "image_path": "08f48acf6d9bc6c16bccd6de7bb8553c7cb55375f61b01b26ae38212d33b08ce.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 366, + 86, + 538, + 240 + ], + "blocks": [ + { + "bbox": [ + 388, + 71, + 517, + 83 + ], + "lines": [ + { + "bbox": [ + 388, + 71, + 517, + 83 + ], + "spans": [ + { + "bbox": [ + 388, + 71, + 517, + 83 + ], + "type": "text", + "content": "(c) Volume rendered outputs" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 366, + 86, + 538, + 240 + ], + "lines": [ + { + "bbox": [ + 366, + 86, + 538, + 240 + ], + "spans": [ + { + "bbox": [ + 366, + 86, + 538, + 240 + ], + "type": "image", + "image_path": "8f5ce03d1d46c9edbba3fab841f598fd38e818f838815b8436f0ed4d1f04ab7e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 241, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 241, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 547, + 277 + ], + "type": "text", + "content": "Figure 2: Overview. From monocular video, SAFF builds a neural field of scene-flowed 3D density, radiance, semantics, and attention (b). This is guided by semantic attention pyramids that increase resolution, plus depth and optical flow priors (a). We can render new spacetime views of any channel, and use saliency-aware clustering to decompose objects and background (c)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 289, + 288, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 288, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 288, + 336 + ], + "type": "text", + "content": "to " + }, + { + "bbox": [ + 46, + 289, + 288, + 336 + ], + "type": "inline_equation", + "content": "t_f" + }, + { + "bbox": [ + 46, + 289, + 288, + 336 + ], + "type": "text", + "content": ", query each network, then integrate transmittance " + }, + { + "bbox": [ + 46, + 289, + 288, + 336 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 289, + 288, + 336 + ], + "type": "text", + "content": ", density " + }, + { + "bbox": [ + 46, + 289, + 288, + 336 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 289, + 288, + 336 + ], + "type": "text", + "content": ", and color along the ray from these samples [23]. For brevity, we omit evaluating at " + }, + { + "bbox": [ + 46, + 289, + 288, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t, \\boldsymbol{\\omega}" + }, + { + "bbox": [ + 46, + 289, + 288, + 336 + ], + "type": "text", + "content": ", e.g., " + }, + { + "bbox": [ + 46, + 289, + 288, + 336 + ], + "type": "inline_equation", + "content": "\\sigma^{\\mathrm{st}}(\\mathbf{x}_t)" + }, + { + "bbox": [ + 46, + 289, + 288, + 336 + ], + "type": "text", + "content": " is simply " + }, + { + "bbox": [ + 46, + 289, + 288, + 336 + ], + "type": "inline_equation", + "content": "\\sigma^{\\mathrm{st}}" + }, + { + "bbox": [ + 46, + 289, + 288, + 336 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 46, + 289, + 288, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^{\\mathrm{st}}(\\mathbf{x}_t, \\boldsymbol{\\omega})" + }, + { + "bbox": [ + 46, + 289, + 288, + 336 + ], + "type": "text", + "content": " is simply " + }, + { + "bbox": [ + 46, + 289, + 288, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^{\\mathrm{st}}" + }, + { + "bbox": [ + 46, + 289, + 288, + 336 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 336, + 287, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 287, + 360 + ], + "type": "text", + "content": "We produce a combined color from the static and dynamic colors by multiplication with their densities:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 102, + 365, + 287, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 365, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 102, + 365, + 287, + 380 + ], + "type": "interline_equation", + "content": "\\sigma_ {i} \\mathbf {c} _ {i} = v \\sigma^ {\\mathrm {s t}} \\mathbf {c} ^ {\\mathrm {s t}} + (1 - v) \\sigma_ {i} ^ {\\mathrm {d y}} \\mathbf {c} _ {i} ^ {\\mathrm {d y}} \\tag {3}", + "image_path": "d9360b47c7e2bbda74a4439dac0805a5cad4d32fc3623e1e5987618d23658550.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 389, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 389, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 389, + 287, + 426 + ], + "type": "text", + "content": "Given that transmittance integrates density up to the current sampled point under Beer-Lambert volume attenuation, the rendered pixel color for the ray is computed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 430, + 287, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 287, + 459 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {c}} _ {i} = \\int_ {t _ {n}} ^ {t _ {f}} T _ {i} \\sigma_ {i} \\mathbf {c} _ {i} d t \\text {w h e r e} T _ {i} = \\exp \\left(- \\int_ {t _ {n}} ^ {t} \\sigma_ {i} d t\\right) \\tag {4}", + "image_path": "0e9b970813e1d0a0f23ab67bb23e14257af7fe73abd1802e2b3f5fe5d3314135.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 467, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 287, + 502 + ], + "type": "text", + "content": "To optimize the volume to reconstruct input images, we compute a photometric loss " + }, + { + "bbox": [ + 46, + 467, + 287, + 502 + ], + "type": "inline_equation", + "content": "L_{\\hat{\\mathbf{c}}}" + }, + { + "bbox": [ + 46, + 467, + 287, + 502 + ], + "type": "text", + "content": " between rendered and input colors:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 95, + 509, + 287, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 509, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 95, + 509, + 287, + 538 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\hat {\\mathbf {c}}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\left\\| \\hat {\\mathbf {c}} _ {i} \\left(\\mathbf {r} _ {i}\\right) - \\hat {\\mathbf {c}} _ {i} ^ {\\dagger} \\left(\\mathbf {r} _ {i}\\right) \\right\\| _ {2} ^ {2} \\tag {5}", + "image_path": "02a8896ce1eddd0f7b9d44be43127defa400a312f1695a0e03b05bcb04c3f8f2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 549, + 287, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 549, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 46, + 549, + 287, + 623 + ], + "type": "text", + "content": "Scene Flow Defining correspondence over time is important for monocular input as it lets us penalize a reprojection error with neighboring frames " + }, + { + "bbox": [ + 46, + 549, + 287, + 623 + ], + "type": "inline_equation", + "content": "j \\in \\mathcal{N}" + }, + { + "bbox": [ + 46, + 549, + 287, + 623 + ], + "type": "text", + "content": ", e.g., where " + }, + { + "bbox": [ + 46, + 549, + 287, + 623 + ], + "type": "inline_equation", + "content": "j = i + 1" + }, + { + "bbox": [ + 46, + 549, + 287, + 623 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 549, + 287, + 623 + ], + "type": "inline_equation", + "content": "j = i - 1" + }, + { + "bbox": [ + 46, + 549, + 287, + 623 + ], + "type": "text", + "content": ". We denote " + }, + { + "bbox": [ + 46, + 549, + 287, + 623 + ], + "type": "inline_equation", + "content": "i \\rightarrow j" + }, + { + "bbox": [ + 46, + 549, + 287, + 623 + ], + "type": "text", + "content": " for the projection of frame " + }, + { + "bbox": [ + 46, + 549, + 287, + 623 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 549, + 287, + 623 + ], + "type": "text", + "content": " onto frame " + }, + { + "bbox": [ + 46, + 549, + 287, + 623 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 549, + 287, + 623 + ], + "type": "text", + "content": " by scene flow. " + }, + { + "bbox": [ + 46, + 549, + 287, + 623 + ], + "type": "inline_equation", + "content": "F_{\\theta}^{\\mathrm{dy}}" + }, + { + "bbox": [ + 46, + 549, + 287, + 623 + ], + "type": "text", + "content": " estimates both forwards and backwards scene flow at every point to penalize a bi-direction loss." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 623, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 287, + 647 + ], + "type": "text", + "content": "Thus, we approximate color output at time step " + }, + { + "bbox": [ + 47, + 623, + 287, + 647 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 623, + 287, + 647 + ], + "type": "text", + "content": " by flowing the queried network values at time step " + }, + { + "bbox": [ + 47, + 623, + 287, + 647 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 623, + 287, + 647 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 102, + 654, + 287, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 654, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 102, + 654, + 287, + 681 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {c}} _ {i \\rightarrow j} = \\int_ {t _ {n}} ^ {t _ {f}} T _ {i \\rightarrow j} \\sigma_ {i \\rightarrow j} \\mathbf {c} _ {i \\rightarrow j} d t \\tag {6}", + "image_path": "ca3504cb6da00ddd335bd52ae969892e9b18efe88de6b04ca7bd3aecf96211c3.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Reprojection must account for occlusion and disocclusion by motion. As such, " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "F_{\\theta}^{\\mathrm{dy}}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " also predicts forwards and backwards" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 289, + 546, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 546, + 325 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 546, + 325 + ], + "type": "text", + "content": "scene occlusion weights " + }, + { + "bbox": [ + 304, + 289, + 546, + 325 + ], + "type": "inline_equation", + "content": "w_{i + 1}" + }, + { + "bbox": [ + 304, + 289, + 546, + 325 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 289, + 546, + 325 + ], + "type": "inline_equation", + "content": "w_{i - 1} \\in [0,1]" + }, + { + "bbox": [ + 304, + 289, + 546, + 325 + ], + "type": "text", + "content": ", where a point with " + }, + { + "bbox": [ + 304, + 289, + 546, + 325 + ], + "type": "inline_equation", + "content": "w_{i + 1} = 0" + }, + { + "bbox": [ + 304, + 289, + 546, + 325 + ], + "type": "text", + "content": " means that occlusion status is changed one step forwards in time. We can integrate " + }, + { + "bbox": [ + 304, + 289, + 546, + 325 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 289, + 546, + 325 + ], + "type": "text", + "content": " to a pixel:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 365, + 329, + 546, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 329, + 546, + 357 + ], + "spans": [ + { + "bbox": [ + 365, + 329, + 546, + 357 + ], + "type": "interline_equation", + "content": "\\hat {w} _ {i \\rightarrow j} = \\int_ {t _ {n}} ^ {t _ {f}} T _ {i \\rightarrow j} \\sigma_ {i \\rightarrow j} w _ {j} d t \\tag {7}", + "image_path": "dd77865e18000d84b15d3ce6aeb7916ba747f25c3ec0f94dc514c1717a9d8200.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 362, + 546, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 362, + 546, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 546, + 389 + ], + "type": "text", + "content": "Then, this pixel weight modulates the color loss such that occluded pixels are ignored. (8)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 391, + 546, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 391, + 546, + 420 + ], + "spans": [ + { + "bbox": [ + 306, + 391, + 546, + 420 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\hat {\\mathbf {c}} _ {i \\rightarrow j}} = \\frac {1}{| \\mathcal {R} | | \\mathcal {N} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\sum_ {j \\in \\mathcal {N}} \\hat {w} _ {i \\rightarrow j} (\\mathbf {r} _ {i}) | | \\hat {\\mathbf {c}} _ {i \\rightarrow j} (\\mathbf {r} _ {i}) - \\hat {\\mathbf {c}} _ {j} ^ {\\dagger} (\\mathbf {r} _ {j}) | | _ {2} ^ {2}", + "image_path": "df8897fb6d339812ef7fb051a92cceb39873dfd1638de4286a534cc6d7ffd194.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 432, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 547, + 529 + ], + "type": "text", + "content": "Prior losses We use the pretrained depth and optical flow map losses to help overcome the ill-posed monocular reconstruction problem. These losses decay as optimization progresses to rely more and more on the optimized self-consistent geometry and scene flow. For geometry, we estimate a depth " + }, + { + "bbox": [ + 304, + 432, + 547, + 529 + ], + "type": "inline_equation", + "content": "\\hat{d}_i" + }, + { + "bbox": [ + 304, + 432, + 547, + 529 + ], + "type": "text", + "content": " for each ray " + }, + { + "bbox": [ + 304, + 432, + 547, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i" + }, + { + "bbox": [ + 304, + 432, + 547, + 529 + ], + "type": "text", + "content": " by replacing " + }, + { + "bbox": [ + 304, + 432, + 547, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 304, + 432, + 547, + 529 + ], + "type": "text", + "content": " in Eq. (4) by the distance " + }, + { + "bbox": [ + 304, + 432, + 547, + 529 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 432, + 547, + 529 + ], + "type": "text", + "content": " along the ray. Transform " + }, + { + "bbox": [ + 304, + 432, + 547, + 529 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 432, + 547, + 529 + ], + "type": "text", + "content": " estimates a scale and shift as the pretrained network produces only relative depth." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 363, + 531, + 546, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 531, + 546, + 561 + ], + "spans": [ + { + "bbox": [ + 363, + 531, + 546, + 561 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\hat {d}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\left| \\left| \\hat {d} _ {i} - z \\left(\\hat {d} _ {i} ^ {\\dagger}\\right) \\right| \\right| _ {1} \\tag {9}", + "image_path": "81e24d816964187bd18192cbc1afabd03e6ad12f7f5462f233f014a01dc7c1e5.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 566, + 546, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 546, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 546, + 639 + ], + "type": "text", + "content": "For motion, projecting scene flow to a camera lets us compare to the estimated optical flow. Each sample point along a ray " + }, + { + "bbox": [ + 304, + 566, + 546, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 566, + 546, + 639 + ], + "type": "text", + "content": " is advected to a point in the neighboring frame " + }, + { + "bbox": [ + 304, + 566, + 546, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i\\rightarrow j}" + }, + { + "bbox": [ + 304, + 566, + 546, + 639 + ], + "type": "text", + "content": ", then integrated to the neighboring camera plane to produce a 2D point offset " + }, + { + "bbox": [ + 304, + 566, + 546, + 639 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}_i(\\mathbf{r}_i)" + }, + { + "bbox": [ + 304, + 566, + 546, + 639 + ], + "type": "text", + "content": ". Then, we expect the difference in the start and end positions to match the prior:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 323, + 641, + 546, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 641, + 546, + 672 + ], + "spans": [ + { + "bbox": [ + 323, + 641, + 546, + 672 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\hat {\\mathbf {p}}} = \\frac {1}{| \\mathcal {R} | | \\mathcal {N} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\sum_ {j \\in \\mathcal {N} (i)} | | \\hat {\\mathbf {p}} _ {i} (\\mathbf {r} _ {i}) - \\hat {\\mathbf {p}} _ {i} ^ {\\dagger} (\\mathbf {r} _ {i}) | | _ {1} \\tag {10}", + "image_path": "142cd64397112021e03ca89a1261cb2e4476dc776affc1a5931447b5a487b05c.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "content": "Additional regularizations encourage occlusion weights to be close to one, scene flow to be small, locally constant, and cyclically consistent, and blending weight " + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "content": " to be sparse." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "21799" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 217, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 217, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 217, + 83 + ], + "type": "text", + "content": "3.2. Semantic Attention Flow Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 186 + ], + "type": "text", + "content": "Beyond low-level or bottom-up features, high-level or top-down features are also useful to define objects and help down-stream tasks like segmentation. For example, methods like NSFF or " + }, + { + "bbox": [ + 46, + 90, + 287, + 186 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 46, + 90, + 287, + 186 + ], + "type": "text", + "content": " struggle to provide useful separation of static and dynamic parts because blend weight " + }, + { + "bbox": [ + 46, + 90, + 287, + 186 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 90, + 287, + 186 + ], + "type": "text", + "content": " estimates whether the volume appears to be occupied by some moving entity. This is not the same as objectness; tasks like video editing could benefit from accurate dynamic object masks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 186, + 288, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 288, + 281 + ], + "type": "text", + "content": "As such, we extract 2D semantic features and attention (or saliency) values from a pretrained DINO-ViT network, then optimize the SAFF such that unknown 3D semantic and attention features over time can be projected to recreate their 2D complements. This helps us to ascribe semantic meaning to the volume and to identify objects. As semantics/attention are integrated into the 4D volume, we can render them from novel spacetime views without further DINO-ViT computation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 281, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 287, + 319 + ], + "type": "text", + "content": "To estimate semantic features " + }, + { + "bbox": [ + 47, + 281, + 287, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 47, + 281, + 287, + 319 + ], + "type": "text", + "content": " and attention " + }, + { + "bbox": [ + 47, + 281, + 287, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 47, + 281, + 287, + 319 + ], + "type": "text", + "content": " at 3D points in the volume at time " + }, + { + "bbox": [ + 47, + 281, + 287, + 319 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 281, + 287, + 319 + ], + "type": "text", + "content": ", we add two new heads to both the static " + }, + { + "bbox": [ + 47, + 281, + 287, + 319 + ], + "type": "inline_equation", + "content": "F_{\\theta}^{\\mathrm{st}}" + }, + { + "bbox": [ + 47, + 281, + 287, + 319 + ], + "type": "text", + "content": " and the dynamic " + }, + { + "bbox": [ + 47, + 281, + 287, + 319 + ], + "type": "inline_equation", + "content": "F_{\\theta}^{\\mathrm{dy}}" + }, + { + "bbox": [ + 47, + 281, + 287, + 319 + ], + "type": "text", + "content": " networks:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 323, + 287, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 323, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 82, + 323, + 287, + 337 + ], + "type": "interline_equation", + "content": "F _ {\\boldsymbol {\\theta}} ^ {\\mathrm {s t}}: (\\mathbf {x}, \\omega) \\quad \\rightarrow \\left(\\mathbf {c} ^ {\\mathrm {s t}}, \\sigma^ {\\mathrm {s t}}, v, \\mathbf {s} ^ {\\mathrm {s t}}, \\mathbf {a} ^ {\\mathrm {s t}}\\right) \\tag {11}", + "image_path": "c019fffff6af56a7b9a00e3cd8005760e1622185301d216796eeb1361835174f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 339, + 287, + 355 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 339, + 287, + 355 + ], + "spans": [ + { + "bbox": [ + 80, + 339, + 287, + 355 + ], + "type": "interline_equation", + "content": "F _ {\\boldsymbol {\\theta}} ^ {\\mathrm {d y}}: (\\mathbf {x}, \\boldsymbol {\\omega}, i) \\rightarrow \\left(\\mathbf {c} _ {i} ^ {\\mathrm {d y}}, \\sigma_ {i} ^ {\\mathrm {d y}}, \\mathbf {f} _ {i}, w _ {i}, \\mathbf {s} _ {i} ^ {\\mathrm {d y}}, \\mathbf {a} _ {i} ^ {\\mathrm {d y}}\\right) \\tag {12}", + "image_path": "1474fcce044c879679dbff499d4969743c0c55ccd580b5ebdc8537053269b3c2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 360, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 287, + 396 + ], + "type": "text", + "content": "As semantic features have been demonstrated to be somewhat robust to view dependence [1], in our architectures both heads for s, a are taken off the backbone before " + }, + { + "bbox": [ + 47, + 360, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 47, + 360, + 287, + 396 + ], + "type": "text", + "content": " is injected." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 396, + 287, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 287, + 419 + ], + "type": "text", + "content": "To render semantics from the volume, we replace the color term " + }, + { + "bbox": [ + 47, + 396, + 287, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 47, + 396, + 287, + 419 + ], + "type": "text", + "content": " in Eq. (4) with " + }, + { + "bbox": [ + 47, + 396, + 287, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 47, + 396, + 287, + 419 + ], + "type": "text", + "content": " and equivalently for " + }, + { + "bbox": [ + 47, + 396, + 287, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 47, + 396, + 287, + 419 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 423, + 287, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 423, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 53, + 423, + 287, + 450 + ], + "type": "interline_equation", + "content": "\\sigma_ {i} \\mathbf {s} _ {i} = v \\sigma^ {\\mathrm {s t}} \\mathbf {s} ^ {\\mathrm {s t}} + (1 - v) \\sigma_ {i} ^ {\\mathrm {d y}} \\mathbf {s} _ {i} ^ {\\mathrm {d y}}, \\hat {\\mathbf {s}} _ {i} = \\int_ {t _ {n}} ^ {t _ {f}} T _ {i} \\sigma_ {i} \\mathbf {s} _ {i} d t \\tag {13}", + "image_path": "60a5ad14cbb70bcc594fdd349dc7e8487c5e2249a1c3702707aff3e915024eca.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 456, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 287, + 479 + ], + "type": "text", + "content": "To encourage scene flow to respect semantics over time, we penalize complementary losses on s and a (showing s only):" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 481, + 287, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 481, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 481, + 287, + 517 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\hat {\\mathbf {s}} _ {i \\rightarrow j}} = \\frac {1}{| \\mathcal {R} | | \\mathcal {N} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\sum_ {j \\in \\mathcal {N}} \\hat {w} _ {i \\rightarrow j} (\\mathbf {r} _ {i}) \\| \\hat {\\mathbf {s}} _ {i \\rightarrow j} (\\mathbf {r} _ {i}) - \\hat {\\mathbf {s}} _ {j} ^ {\\dagger} (\\mathbf {r} _ {j}) \\| _ {2} ^ {2} \\tag {14}", + "image_path": "05111168127741f2d1dc72ba52b580c1f917dc06aba4ff8abf064d6f9c059a8b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 523, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 523, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 523, + 287, + 559 + ], + "type": "text", + "content": "Finally, as supervision, we add respective losses on the reconstruction of the 2D semantic and attention features from projected 3D volume points (showing s only):" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 96, + 561, + 287, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 561, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 96, + 561, + 287, + 590 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\hat {\\mathbf {s}}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\left\\| \\hat {\\mathbf {s}} _ {i} \\left(\\mathbf {r} _ {i}\\right) - \\hat {\\mathbf {s}} _ {i} ^ {\\dagger} \\left(\\mathbf {r} _ {i}\\right) \\right\\| _ {2} ^ {2} \\tag {15}", + "image_path": "6c33f28cb832c67e0c70332708148d6d2905cdb43f5e3ce542c895617c67b64c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 595, + 288, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 595, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 47, + 595, + 288, + 667 + ], + "type": "text", + "content": "Unlike depth and scene flow priors, these are not priors—there is no self-consistency for semantics to constrain their values. Thus, we do not decay their contribution. While decaying avoids disagreements between semantic and attention features and color-enforced scene geometry, it also leads to a loss of useful meaning (please see supplemental)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 667, + 174, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 667, + 174, + 678 + ], + "spans": [ + { + "bbox": [ + 59, + 667, + 174, + 678 + ], + "type": "text", + "content": "Thus our final loss becomes:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 683, + 287, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 683, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 683, + 287, + 712 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {S A F F}} = \\mathcal {L} _ {\\hat {\\mathbf {c}}} + \\lambda_ {\\hat {\\mathbf {c}} _ {i \\rightarrow j}} \\mathcal {L} _ {\\hat {\\mathbf {c}} _ {i \\rightarrow j}} + \\lambda_ {\\hat {d}} \\mathcal {L} _ {\\hat {d}} + \\lambda_ {\\hat {\\mathbf {p}}} \\mathcal {L} _ {\\hat {\\mathbf {p}}} \\tag {16} \\\\ + \\lambda_ {\\hat {\\mathbf {s}} _ {i \\rightarrow j}} \\mathcal {L} _ {\\hat {\\mathbf {s}} _ {i \\rightarrow j}} + \\lambda_ {\\hat {\\mathbf {a}} _ {i \\rightarrow j}} \\mathcal {L} _ {\\hat {\\mathbf {a}} _ {i \\rightarrow j}} + \\lambda_ {\\hat {\\mathbf {s}}} \\mathcal {L} _ {\\hat {\\mathbf {s}}} + \\lambda_ {\\hat {\\mathbf {a}}} \\mathcal {L} _ {\\hat {\\mathbf {a}}} \\\\ \\end{array}", + "image_path": "7bd88bce1b23b67ee7a083194a8a096dea322403d9445e4b7d0fe1d5c14e36d1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 72, + 467, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 467, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 467, + 84 + ], + "type": "text", + "content": "3.3. Semantic Attention Pyramids" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 90, + 547, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 547, + 209 + ], + "type": "text", + "content": "When thinking about scenes, we might argue that semantics from an ideal extractor should be scale invariant, as distant objects have the same class as close objects. We might also argue that saliency (or attention features) may not be scale invariant, as small details in a scene should only be salient when viewed close up. In practice, both extracted features vary across scale and have limited resolution, e.g., DINO-ViT [6] produces one output for each " + }, + { + "bbox": [ + 304, + 90, + 547, + 209 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 304, + 90, + 547, + 209 + ], + "type": "text", + "content": " patch. But, from this, we want semantic features and saliency for every RGB pixel that still respects scene boundaries." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 210, + 547, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 547, + 305 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 547, + 305 + ], + "type": "text", + "content": "Thus far, work on static scenes has ignored the input/feature resolution mismatch [16] as multi-view constraints provide improved localization within the volume. For monocular video, this approach has limitations [36]. Forming many constraints on dynamic objects requires long-term motion correspondence—a tricky task—and so we want to maximize the resolution of any input features where possible without changing their meaning." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 305, + 547, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 547, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 547, + 401 + ], + "type": "text", + "content": "One way may be through a pyramid of semantic and attention features that uses a sliding window approach at finer resolutions. Averaging features could increase detail around edges, but we must overcome the practical limit that these features are not stable across scales. This is especially important for saliency: unlike typical RGB pyramids that must preserve energy in an alias-free way [2], saliency changes significantly over scales and does not preserve energy." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 401, + 546, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 401, + 546, + 413 + ], + "spans": [ + { + "bbox": [ + 317, + 401, + 546, + 413 + ], + "type": "text", + "content": "Consider a feature pyramid " + }, + { + "bbox": [ + 317, + 401, + 546, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 317, + 401, + 546, + 413 + ], + "type": "text", + "content": " with loss weights per level:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 348, + 429, + 546, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 429, + 546, + 455 + ], + "spans": [ + { + "bbox": [ + 348, + 429, + 546, + 455 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {P} \\hat {\\mathbf {s}}} = \\sum_ {i \\in \\mathcal {P}} \\lambda_ {\\hat {\\mathbf {s}}} ^ {i} \\mathcal {L} _ {\\hat {\\mathbf {s}}} ^ {i} \\quad \\mathcal {L} _ {\\mathcal {P} \\hat {\\mathbf {a}}} = \\sum_ {i \\in \\mathcal {P}} \\lambda_ {\\hat {\\mathbf {a}}} ^ {i} \\mathcal {L} _ {\\hat {\\mathbf {a}}} ^ {i} \\tag {17}", + "image_path": "20770d818a6554451c46d3d3af083d7d8718fdd2154f836c9826bfe3339e0708.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 462, + 547, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 582 + ], + "type": "text", + "content": "Naively encouraging scale-consistent semantics and whole-image saliency, e.g., " + }, + { + "bbox": [ + 304, + 462, + 547, + 582 + ], + "type": "inline_equation", + "content": "\\lambda_{\\hat{\\mathbf{s}}} = \\{1/3, 1/3, 1/3\\}" + }, + { + "bbox": [ + 304, + 462, + 547, + 582 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 462, + 547, + 582 + ], + "type": "inline_equation", + "content": "\\lambda_{\\hat{\\mathbf{a}}} = \\{1, 0, 0\\}" + }, + { + "bbox": [ + 304, + 462, + 547, + 582 + ], + "type": "text", + "content": ", empirically leads to poor recovered object edges because the balanced semantics and coarse saliency compete over where the underlying geometry is. Instead, we weight both equally " + }, + { + "bbox": [ + 304, + 462, + 547, + 582 + ], + "type": "inline_equation", + "content": "\\lambda_{\\hat{\\mathbf{s}}} = \\lambda_{\\hat{\\mathbf{a}}} = \\{1/9, 4/9, 4/9\\}" + }, + { + "bbox": [ + 304, + 462, + 547, + 582 + ], + "type": "text", + "content": ". Even though the coarse layer has smaller weight, it is sufficient to guide the overall result. This balances high resolution edges from fine layers and whole object features from coarse layers while reducing geometry conflicts, and leads to improved features (Fig. 3)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Of course, any sliding window must contain an object to extract reliable features for that object. At coarse levels, an object is always in view. At fine levels, an object is only captured in some windows. Objects of interest tend to be near the middle of the frame, meaning that boundary windows at finer pyramid levels contain features that less reliably capture those objects. This can cause spurious connections in clustering. To cope with this, we relatively decrease finer level boundary window weights: We upsample all levels to the finest level, then increase the coarsest level weight towards the frame boundary to " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{\\hat{\\mathbf{s}}} = \\lambda_{\\hat{\\mathbf{a}}} = \\{1/3, 1/3, 1/3\\}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "21800" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 79, + 158, + 138 + ], + "blocks": [ + { + "bbox": [ + 100, + 71, + 126, + 78 + ], + "lines": [ + { + "bbox": [ + 100, + 71, + 126, + 78 + ], + "spans": [ + { + "bbox": [ + 100, + 71, + 126, + 78 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 79, + 158, + 138 + ], + "lines": [ + { + "bbox": [ + 67, + 79, + 158, + 138 + ], + "spans": [ + { + "bbox": [ + 67, + 79, + 158, + 138 + ], + "type": "image", + "image_path": "f178d6f08ab968f5f60529102575cf05b82ce1c9023a855566a217b585eb38d3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 159, + 79, + 250, + 138 + ], + "blocks": [ + { + "bbox": [ + 186, + 71, + 225, + 77 + ], + "lines": [ + { + "bbox": [ + 186, + 71, + 225, + 77 + ], + "spans": [ + { + "bbox": [ + 186, + 71, + 225, + 77 + ], + "type": "text", + "content": "(b) Semantics" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 79, + 250, + 138 + ], + "lines": [ + { + "bbox": [ + 159, + 79, + 250, + 138 + ], + "spans": [ + { + "bbox": [ + 159, + 79, + 250, + 138 + ], + "type": "image", + "image_path": "9b0a274bb513cdbaff27d2e781ba72fa72821d8b34d8ed9cb154a712427de7bd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 251, + 79, + 342, + 138 + ], + "blocks": [ + { + "bbox": [ + 264, + 71, + 329, + 78 + ], + "lines": [ + { + "bbox": [ + 264, + 71, + 329, + 78 + ], + "spans": [ + { + "bbox": [ + 264, + 71, + 329, + 78 + ], + "type": "text", + "content": "(c) Semantics (volume)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 251, + 79, + 342, + 138 + ], + "lines": [ + { + "bbox": [ + 251, + 79, + 342, + 138 + ], + "spans": [ + { + "bbox": [ + 251, + 79, + 342, + 138 + ], + "type": "image", + "image_path": "ca4d6cc4b08415640f4ab2ce014bf87ff955f0bdf83c5692ffb80dd75522697d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 343, + 79, + 433, + 138 + ], + "blocks": [ + { + "bbox": [ + 371, + 71, + 405, + 78 + ], + "lines": [ + { + "bbox": [ + 371, + 71, + 405, + 78 + ], + "spans": [ + { + "bbox": [ + 371, + 71, + 405, + 78 + ], + "type": "text", + "content": "(d) Saliency" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 343, + 79, + 433, + 138 + ], + "lines": [ + { + "bbox": [ + 343, + 79, + 433, + 138 + ], + "spans": [ + { + "bbox": [ + 343, + 79, + 433, + 138 + ], + "type": "image", + "image_path": "286f833094b4816d0f26ea379f6c302a9a40006555badeca5a9734de5a57cb54.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 434, + 79, + 524, + 138 + ], + "blocks": [ + { + "bbox": [ + 449, + 71, + 510, + 78 + ], + "lines": [ + { + "bbox": [ + 449, + 71, + 510, + 78 + ], + "spans": [ + { + "bbox": [ + 449, + 71, + 510, + 78 + ], + "type": "text", + "content": "(e) Saliency (volume)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 434, + 79, + 524, + 138 + ], + "lines": [ + { + "bbox": [ + 434, + 79, + 524, + 138 + ], + "spans": [ + { + "bbox": [ + 434, + 79, + 524, + 138 + ], + "type": "image", + "image_path": "6073298e54a3d420c359a7e00e8a065e7a90a15f4532e6dd95c82ee33325699a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 159, + 148, + 250, + 207 + ], + "blocks": [ + { + "bbox": [ + 181, + 140, + 229, + 148 + ], + "lines": [ + { + "bbox": [ + 181, + 140, + 229, + 148 + ], + "spans": [ + { + "bbox": [ + 181, + 140, + 229, + 148 + ], + "type": "text", + "content": "(f) Pyr. semantics" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 148, + 250, + 207 + ], + "lines": [ + { + "bbox": [ + 159, + 148, + 250, + 207 + ], + "spans": [ + { + "bbox": [ + 159, + 148, + 250, + 207 + ], + "type": "image", + "image_path": "9a772ca97f6d34a23e449345bc01ca9b625edcc4b2d9d0c0e4c5afca1890d8ad.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 251, + 148, + 341, + 207 + ], + "blocks": [ + { + "bbox": [ + 252, + 140, + 341, + 148 + ], + "lines": [ + { + "bbox": [ + 252, + 140, + 341, + 148 + ], + "spans": [ + { + "bbox": [ + 252, + 140, + 341, + 148 + ], + "type": "text", + "content": "(g) Our pyr. semantics (volume)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 251, + 148, + 341, + 207 + ], + "lines": [ + { + "bbox": [ + 251, + 148, + 341, + 207 + ], + "spans": [ + { + "bbox": [ + 251, + 148, + 341, + 207 + ], + "type": "image", + "image_path": "b1d357ef4844ecadfd229b10f4610f2f43a0e7ebec2c3b41c59afbd833772873.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 212, + 545, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 212, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 545, + 236 + ], + "type": "text", + "content": "Figure 3: Semantics and saliency improve by both volume integration and by our pyramid. On Balloon NBoard, blocky artifacts are removed. Semantics are visualized as most significant three PCA dimensions; specific colors are less meaningful." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 343, + 148, + 433, + 207 + ], + "blocks": [ + { + "bbox": [ + 365, + 140, + 410, + 148 + ], + "lines": [ + { + "bbox": [ + 365, + 140, + 410, + 148 + ], + "spans": [ + { + "bbox": [ + 365, + 140, + 410, + 148 + ], + "type": "text", + "content": "(h) Pyr. saliency" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 343, + 148, + 433, + 207 + ], + "lines": [ + { + "bbox": [ + 343, + 148, + 433, + 207 + ], + "spans": [ + { + "bbox": [ + 343, + 148, + 433, + 207 + ], + "type": "image", + "image_path": "db2932d48e70bef0de273f03533759cd8762fef47e784615bda46b636c792bcb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 435, + 148, + 524, + 207 + ], + "blocks": [ + { + "bbox": [ + 438, + 140, + 521, + 148 + ], + "lines": [ + { + "bbox": [ + 438, + 140, + 521, + 148 + ], + "spans": [ + { + "bbox": [ + 438, + 140, + 521, + 148 + ], + "type": "text", + "content": "(i) Our pyr. saliency (volume)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 148, + 524, + 207 + ], + "lines": [ + { + "bbox": [ + 435, + 148, + 524, + 207 + ], + "spans": [ + { + "bbox": [ + 435, + 148, + 524, + 207 + ], + "type": "image", + "image_path": "d9cad011c3b648a0848faa5ed6182fbf8a504b14c318a601d52790f011237b06.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 249, + 269, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 249, + 269, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 249, + 269, + 262 + ], + "type": "text", + "content": "3.4. Using SAFF for Saliency-aware Clustering" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 268, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 287, + 400 + ], + "type": "text", + "content": "We now wish to isolate salient objects. Even in dynamic scenes, relevant objects may not move, so analyzing dynamic elements is insufficient (cf. [40]). One approach predicts segmentation end-to-end [19]. However, end-to-end learning requires priors about the scene provided by supervision, and even large-scale pretraining might fail given unseen test scenes. To achieve scene-specific decompositions from per-video semantics, inspired by Amir et al. [1], we design clustering for spacetime volumes that allows segmenting novel spatio-temporal views. While DINO-ViT is trained on images, its features are loosely temporally consistent [6]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 401, + 288, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 288, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 288, + 484 + ], + "type": "text", + "content": "Some works optimize a representation with a fixed number of clusters, e.g., via slot attention [20] in NeRFs [32, 46]. Instead, we cluster using elbow " + }, + { + "bbox": [ + 46, + 401, + 288, + 484 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 401, + 288, + 484 + ], + "type": "text", + "content": "-means, letting us adaptively find the number of clusters after optimization. This is more flexible than baking an anticipated number of slots (sometimes with fixed semantics), and lets us cluster and segment in novel spatio-temporal viewpoints." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 594 + ], + "type": "text", + "content": "We demonstrate clustering results in both 3D over time and on rendered volume 2D projections over time. Given the volume reconstruction, we might think that clustering directly in 3D would be better. But, monocular input with narrow baselines makes it challenging to precisely reconstruct geometries: Consider that depth integrated a long a ray can still be accurate even though geometry at specific 3D points may be inaccurate or 'fluffy'. As such, we use the 2D volume projection clustering results in 2D comparisons." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": "Method For 3D over time, we sample points from the SAFF uniformly along input rays " + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "inline_equation", + "content": "(128\\times H\\times W)" + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": " and treat each pixel as a feature point, e.g., semantics are 64 dim. and saliency is 1 dim. For volume projection, we render to " + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": " input poses " + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "inline_equation", + "content": "(N\\times H\\times W)" + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": ", and treat each pixel as a feature point instead. In either case, we cluster all feature points together using elbow " + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": "-means to produce an initial set of separate regions. For each cluster " + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": ", for each image, we calculate the mean attention of all feature points within the cluster " + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{a}}_c" + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{a}}_c > 0.07" + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": ", then this" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 250, + 545, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 250, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 304, + 250, + 545, + 274 + ], + "type": "text", + "content": "cluster is salient for this image. Finally, all feature points vote on saliency: if more than " + }, + { + "bbox": [ + 304, + 250, + 545, + 274 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 304, + 250, + 545, + 274 + ], + "type": "text", + "content": " agree, the cluster is salient." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 274, + 546, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 546, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 546, + 370 + ], + "type": "text", + "content": "Salient objects may still be split into semantic parts: e.g., in Fig. 4, the person's head/body are separated. Plus, unwanted background saliency may exist, e.g., input " + }, + { + "bbox": [ + 304, + 274, + 546, + 370 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}^\\dagger" + }, + { + "bbox": [ + 304, + 274, + 546, + 370 + ], + "type": "text", + "content": " is high for the teal graphic on the wall. As such, before saliency voting, we merge clusters whose centroids have a cosine similarity " + }, + { + "bbox": [ + 304, + 274, + 546, + 370 + ], + "type": "inline_equation", + "content": ">0.5" + }, + { + "bbox": [ + 304, + 274, + 546, + 370 + ], + "type": "text", + "content": ". This reduces the first problem as heads and bodies are similar, and reduces the second problem as merging the graphic cluster into the background reduces its average saliency (Fig. 4)." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 370, + 547, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 547, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 547, + 454 + ], + "type": "text", + "content": "To extract an object from the 3D volume, we sample 3D points along each input ray, then ascribe the label from the semantically-closest centroid. All clusters not similar to the stored salient clusters are marked as background with zero density. For novel space-time views, we render feature images from the volume, then assign cluster labels to each pixel according to its similarity with stored input view centroids." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 465, + 386, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 465, + 386, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 465, + 386, + 479 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 486, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 521 + ], + "type": "text", + "content": "We show the impact of adding semantic and saliency features through scene decomposition and foreground experiments. Our website contains supplemental videos." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 531, + 547, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 547, + 592 + ], + "type": "text", + "content": "Data: NVIDIA Dynamic Scene Dataset (Masked) This data [45] has 8 sequences of 24 time steps formed from 12 cameras simultaneously capturing video. We manually annotate object masks for view and time step splits; we will release this data publicly. We define three data splits per sequence:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 596, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 306, + 596, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 596, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 306, + 596, + 545, + 632 + ], + "type": "text", + "content": "1. Input: A monocular camera that moves position for every timestep is simulated from the input sequences; we use Yoon et al.'s input sequences [45]." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 636, + 547, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 636, + 547, + 673 + ], + "spans": [ + { + "bbox": [ + 306, + 636, + 547, + 673 + ], + "type": "text", + "content": "2. Fix Cam 0 (hold out): We fix the camera at position 0 as time plays, requiring novel view and time synthesis. " + }, + { + "bbox": [ + 306, + 636, + 547, + 673 + ], + "type": "inline_equation", + "content": "\\{(\\mathrm{cam}_0,\\mathrm{time}_i),i\\in [1,2,\\dots 23]\\}" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 306, + 677, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 547, + 714 + ], + "type": "text", + "content": "3. Fix Time 0 (hold out): We fix time at step 0 as the camera moves, requiring novel view and time synthesis. " + }, + { + "bbox": [ + 306, + 677, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\{(\\mathrm{cam}_i,\\mathrm{time}_0),i\\in [1,2,\\ldots 11]\\}" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "21801" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 64, + 79, + 154, + 137 + ], + "blocks": [ + { + "bbox": [ + 88, + 71, + 130, + 78 + ], + "lines": [ + { + "bbox": [ + 88, + 71, + 130, + 78 + ], + "spans": [ + { + "bbox": [ + 88, + 71, + 130, + 78 + ], + "type": "text", + "content": "(i) Input frame" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 64, + 79, + 154, + 137 + ], + "lines": [ + { + "bbox": [ + 64, + 79, + 154, + 137 + ], + "spans": [ + { + "bbox": [ + 64, + 79, + 154, + 137 + ], + "type": "image", + "image_path": "0916a3dac95539f8c91aff80d1f8a07ab4e1ba67931550ee0b2cae13ffc89eda.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 138, + 149, + 146 + ], + "lines": [ + { + "bbox": [ + 70, + 138, + 149, + 146 + ], + "spans": [ + { + "bbox": [ + 70, + 138, + 149, + 146 + ], + "type": "text", + "content": "(ii) Input semantic/attention" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 155, + 79, + 246, + 137 + ], + "blocks": [ + { + "bbox": [ + 157, + 71, + 243, + 78 + ], + "lines": [ + { + "bbox": [ + 157, + 71, + 243, + 78 + ], + "spans": [ + { + "bbox": [ + 157, + 71, + 243, + 78 + ], + "type": "text", + "content": "(a) Our pyr. semantic (volume)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 155, + 79, + 246, + 137 + ], + "lines": [ + { + "bbox": [ + 155, + 79, + 246, + 137 + ], + "spans": [ + { + "bbox": [ + 155, + 79, + 246, + 137 + ], + "type": "image", + "image_path": "c5ce24e922c13f4821e3240ca0f0021f73e48b716f94b30e793395bc9e3a343b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 138, + 242, + 146 + ], + "lines": [ + { + "bbox": [ + 158, + 138, + 242, + 146 + ], + "spans": [ + { + "bbox": [ + 158, + 138, + 242, + 146 + ], + "type": "text", + "content": "(d) Our pyr. saliency (volume)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 247, + 71, + 345, + 137 + ], + "blocks": [ + { + "bbox": [ + 247, + 71, + 345, + 137 + ], + "lines": [ + { + "bbox": [ + 247, + 71, + 345, + 137 + ], + "spans": [ + { + "bbox": [ + 247, + 71, + 345, + 137 + ], + "type": "image", + "image_path": "5d567c20160ac212f558b1a90c90f18dce7c52124b47c9276e2fdd6d4ca11381.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 251, + 146, + 342, + 205 + ], + "blocks": [ + { + "bbox": [ + 261, + 138, + 333, + 146 + ], + "lines": [ + { + "bbox": [ + 261, + 138, + 333, + 146 + ], + "spans": [ + { + "bbox": [ + 261, + 138, + 333, + 146 + ], + "type": "text", + "content": "(e) Our merged clustering" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 251, + 146, + 342, + 205 + ], + "lines": [ + { + "bbox": [ + 251, + 146, + 342, + 205 + ], + "spans": [ + { + "bbox": [ + 251, + 146, + 342, + 205 + ], + "type": "image", + "image_path": "f462e2415f3c4072ff3dd65adef4b2717184e0446ea7a0119253bc18cc96ab9d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 209, + 545, + 234 + ], + "lines": [ + { + "bbox": [ + 47, + 209, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 545, + 234 + ], + "type": "text", + "content": "Figure 4: Saliency-aware clustering improves decomposition. On Dynamic Face, the head and body are semantically and saliently different, but are mutually different from the background. This allows us to extract a time-varying 3D field of the object." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 347, + 79, + 438, + 137 + ], + "blocks": [ + { + "bbox": [ + 349, + 71, + 437, + 78 + ], + "lines": [ + { + "bbox": [ + 349, + 71, + 437, + 78 + ], + "spans": [ + { + "bbox": [ + 349, + 71, + 437, + 78 + ], + "type": "text", + "content": "(c) Salient clusters (pre-merge)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 347, + 79, + 438, + 137 + ], + "lines": [ + { + "bbox": [ + 347, + 79, + 438, + 137 + ], + "spans": [ + { + "bbox": [ + 347, + 79, + 438, + 137 + ], + "type": "image", + "image_path": "83fe39f2429a8082a84ae5b94ad37683b9859938142cb82fa7d83523de749201.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 138, + 425, + 146 + ], + "lines": [ + { + "bbox": [ + 360, + 138, + 425, + 146 + ], + "spans": [ + { + "bbox": [ + 360, + 138, + 425, + 146 + ], + "type": "text", + "content": "(f) Our final foreground" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 446, + 71, + 522, + 78 + ], + "lines": [ + { + "bbox": [ + 446, + 71, + 522, + 78 + ], + "spans": [ + { + "bbox": [ + 446, + 71, + 522, + 78 + ], + "type": "text", + "content": "(iii) Extraction (fixed time)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 347, + 147, + 438, + 205 + ], + "blocks": [ + { + "bbox": [ + 347, + 147, + 438, + 205 + ], + "lines": [ + { + "bbox": [ + 347, + 147, + 438, + 205 + ], + "spans": [ + { + "bbox": [ + 347, + 147, + 438, + 205 + ], + "type": "image", + "image_path": "5cc523c6ad484004e2d5657e5123e5dbf9683ce2a9478a19116031c1ea14a0d0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 440, + 95, + 530, + 137 + ], + "blocks": [ + { + "bbox": [ + 440, + 95, + 530, + 137 + ], + "lines": [ + { + "bbox": [ + 440, + 95, + 530, + 137 + ], + "spans": [ + { + "bbox": [ + 440, + 95, + 530, + 137 + ], + "type": "image", + "image_path": "5cbf3a9ab54ff469420060cb15eb9ca62dcd439ff5e73fb19525cc2d36915a10.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 138, + 522, + 146 + ], + "lines": [ + { + "bbox": [ + 447, + 138, + 522, + 146 + ], + "spans": [ + { + "bbox": [ + 447, + 138, + 522, + 146 + ], + "type": "text", + "content": "(iv) Extraction (fixed view)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 440, + 151, + 530, + 175 + ], + "blocks": [ + { + "bbox": [ + 440, + 151, + 530, + 175 + ], + "lines": [ + { + "bbox": [ + 440, + 151, + 530, + 175 + ], + "spans": [ + { + "bbox": [ + 440, + 151, + 530, + 175 + ], + "type": "image", + "image_path": "d2014859e5f8596c4f049fd7a5542a994321f44e6e401f6f045a53168d008f72.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 440, + 180, + 530, + 205 + ], + "blocks": [ + { + "bbox": [ + 440, + 180, + 530, + 205 + ], + "lines": [ + { + "bbox": [ + 440, + 180, + 530, + 205 + ], + "spans": [ + { + "bbox": [ + 440, + 180, + 530, + 205 + ], + "type": "image", + "image_path": "e20bdacede3a7a5868b7cef4fcdcf593b23fa2655dfd89336a1d4e0a92985e35.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 441, + 239, + 531, + 300 + ], + "blocks": [ + { + "bbox": [ + 441, + 239, + 531, + 300 + ], + "lines": [ + { + "bbox": [ + 441, + 239, + 531, + 300 + ], + "spans": [ + { + "bbox": [ + 441, + 239, + 531, + 300 + ], + "type": "image", + "image_path": "09c1891a830a23fbc620a22ac5e9c58856b78f5efaea2cf8f1da9005971890f1.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 59, + 251, + 160, + 403 + ], + "blocks": [ + { + "bbox": [ + 98, + 239, + 134, + 250 + ], + "lines": [ + { + "bbox": [ + 98, + 239, + 134, + 250 + ], + "spans": [ + { + "bbox": [ + 98, + 239, + 134, + 250 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 251, + 160, + 403 + ], + "lines": [ + { + "bbox": [ + 59, + 251, + 160, + 403 + ], + "spans": [ + { + "bbox": [ + 59, + 251, + 160, + 403 + ], + "type": "image", + "image_path": "7def0160a5af8dfbc2da02f54623763411d8065e9bb925784ad0a2b1dd18007b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 408, + 545, + 444 + ], + "lines": [ + { + "bbox": [ + 46, + 408, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 545, + 444 + ], + "type": "text", + "content": "Figure 5: SAFF object segmentations show balanced quality while recovering a volumetric scene representation (e). Basic DINO-ViT produces low-quality segmentations and misses objects. A state-of-the-art 2D video learning method [19] sometimes has edge detail (Umbrella, legs) but othertimes misses detail and objects (Balloon NBoard)." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 162, + 251, + 252, + 403 + ], + "blocks": [ + { + "bbox": [ + 168, + 239, + 247, + 250 + ], + "lines": [ + { + "bbox": [ + 168, + 239, + 247, + 250 + ], + "spans": [ + { + "bbox": [ + 168, + 239, + 247, + 250 + ], + "type": "text", + "content": "(b) DINO-ViT (2D)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 162, + 251, + 252, + 403 + ], + "lines": [ + { + "bbox": [ + 162, + 251, + 252, + 403 + ], + "spans": [ + { + "bbox": [ + 162, + 251, + 252, + 403 + ], + "type": "image", + "image_path": "4b13ac5c9c77ea60444552e5524c3e891235297addb0747db1403a807f583b6e.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 255, + 251, + 345, + 403 + ], + "blocks": [ + { + "bbox": [ + 255, + 239, + 347, + 250 + ], + "lines": [ + { + "bbox": [ + 255, + 239, + 347, + 250 + ], + "spans": [ + { + "bbox": [ + 255, + 239, + 347, + 250 + ], + "type": "text", + "content": "(c) ProposeReduce [19]" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 255, + 251, + 345, + 403 + ], + "lines": [ + { + "bbox": [ + 255, + 251, + 345, + 403 + ], + "spans": [ + { + "bbox": [ + 255, + 251, + 345, + 403 + ], + "type": "image", + "image_path": "6e89c0f7d9b9ee8d30bfbc5bec0ae618274ab0497d01755188af532b00ba8da6.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 348, + 251, + 438, + 403 + ], + "blocks": [ + { + "bbox": [ + 349, + 239, + 439, + 250 + ], + "lines": [ + { + "bbox": [ + 349, + 239, + 439, + 250 + ], + "spans": [ + { + "bbox": [ + 349, + 239, + 439, + 250 + ], + "type": "text", + "content": "(d) SAFF (Ours; input)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 348, + 251, + 438, + 403 + ], + "lines": [ + { + "bbox": [ + 348, + 251, + 438, + 403 + ], + "spans": [ + { + "bbox": [ + 348, + 251, + 438, + 403 + ], + "type": "image", + "image_path": "a9ed21c21fb0dcc4c27f0d7b78c66d1cf4381b857645386272644700ab7fdba2.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 441, + 304, + 530, + 403 + ], + "blocks": [ + { + "bbox": [ + 441, + 304, + 530, + 403 + ], + "lines": [ + { + "bbox": [ + 441, + 304, + 530, + 403 + ], + "spans": [ + { + "bbox": [ + 441, + 304, + 530, + 403 + ], + "type": "image", + "image_path": "8810bee4ce7e1f93f0f11ff2df60cc4844cb666696cdb3af51a551bec96b38da.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 465, + 287, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 465, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 46, + 465, + 287, + 573 + ], + "type": "text", + "content": "Data: DyCheck Dataset (Masked) For single-camera casual monocular data, we select a subset of Gao et al.'s DyCheck dataset [10]. We remove scenes without an obvious subject, e.g., where telling foreground from background is hard even for a human. We select haru-sit, space-out, spin, paper-windmill, and mochi-high-five. We uniformly sample 48 frames and manually annotate object masks. We take even frames as the input set and odd frames as the hold out set. We use the same hyperparameters between both datasets." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 46, + 579, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 651 + ], + "type": "text", + "content": "Metrics To assess clustering performance, we use the Adjusted Rand Index (ARI; " + }, + { + "bbox": [ + 46, + 579, + 287, + 651 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 46, + 579, + 287, + 651 + ], + "type": "text", + "content": "). This compares the similarity of two assignments without label matching, where random assignment would score " + }, + { + "bbox": [ + 46, + 579, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\approx 0" + }, + { + "bbox": [ + 46, + 579, + 287, + 651 + ], + "type": "text", + "content": ". For foreground segmentation, we compute IoU (Jaccard), and for RGB quality we use PSNR, SSIM, and LPIPS." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 47, + 659, + 224, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 224, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 224, + 672 + ], + "type": "text", + "content": "4.1. Comparisons including ablations" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "SAFF (ours) We optimize upon the input split of each scene, and perform clustering to obtain object segmentations. To produce a foreground, we merge all salient objects." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 305, + 465, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 465, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 305, + 465, + 545, + 490 + ], + "type": "text", + "content": "SAFF (3D) The same as above, but processed in 3D rather than on 2D volume projections." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 495, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 40, + "blocks": [ + { + "bbox": [ + 305, + 495, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 495, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 305, + 495, + 545, + 520 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 305, + 495, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\mathbf{w} / \\mathbf{pyr}\\lambda_{\\hat{\\mathbf{a}}} = \\{1,0,0\\}" + }, + { + "bbox": [ + 305, + 495, + 545, + 520 + ], + "type": "text", + "content": " Pyramid with only coarse saliency (Sec. 3.3) and balanced semantic weight across levels." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 305, + 526, + 545, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 526, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 305, + 526, + 545, + 551 + ], + "type": "text", + "content": "—w/o pyr No pyramid (Sec. 3.3); we optimize with features and saliency extracted from the input image only." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 305, + 556, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 556, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 305, + 556, + 545, + 581 + ], + "type": "text", + "content": "— w/o merge With pyramid, but we remove cluster merging inside the saliency-aware clustering algorithm." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 586, + 547, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 547, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 547, + 647 + ], + "type": "text", + "content": "- w/blend " + }, + { + "bbox": [ + 304, + 586, + 547, + 647 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 586, + 547, + 647 + ], + "type": "text", + "content": " To compare generic dynamic segmentation to saliency segmentation, we use the static/dynamic weight instead of volume saliency to segment foreground objects. We set every pixel below the " + }, + { + "bbox": [ + 304, + 586, + 547, + 647 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 304, + 586, + 547, + 647 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 586, + 547, + 647 + ], + "type": "text", + "content": " quantile in each image to be background, or otherwise foreground." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "— w/post process We add a step after the saliency-aware clustering to refine edges using a conditional random field (please see supplemental material for details). This gains significantly from the depth estimated via volume reconstruction, producing sharp and detailed edges." + } + ] + } + ], + "index": 39 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "21802" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 159, + 176 + ], + "blocks": [ + { + "bbox": [ + 96, + 60, + 132, + 71 + ], + "lines": [ + { + "bbox": [ + 96, + 60, + 132, + 71 + ], + "spans": [ + { + "bbox": [ + 96, + 60, + 132, + 71 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 72, + 159, + 176 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 159, + 176 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 159, + 176 + ], + "type": "image", + "image_path": "36a63135f22edbeb9eb8b71e22f73a3ba4d8b599837e8409cfe50dd8f4c40fc9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 179, + 547, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 179, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 547, + 216 + ], + "type": "text", + "content": "Figure 6: Saliency improves foreground segmentation. Static/dynamic separations are not foreground segmentations, leading to limited use of dynamic NeRF models for downstream tasks. Minor improvements to dynamic blending weight " + }, + { + "bbox": [ + 46, + 179, + 547, + 216 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 179, + 547, + 216 + ], + "type": "text", + "content": " are seen in some sequences (Jumping) by adding the saliency head to the shared backbone." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 160, + 72, + 252, + 176 + ], + "blocks": [ + { + "bbox": [ + 161, + 60, + 351, + 71 + ], + "lines": [ + { + "bbox": [ + 161, + 60, + 351, + 71 + ], + "spans": [ + { + "bbox": [ + 161, + 60, + 351, + 71 + ], + "type": "text", + "content": "(b) NSFF [18] blend " + }, + { + "bbox": [ + 161, + 60, + 351, + 71 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 161, + 60, + 351, + 71 + ], + "type": "text", + "content": " (c) " + }, + { + "bbox": [ + 161, + 60, + 351, + 71 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 161, + 60, + 351, + 71 + ], + "type": "text", + "content": " [40] blend " + }, + { + "bbox": [ + 161, + 60, + 351, + 71 + ], + "type": "inline_equation", + "content": "v" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 160, + 72, + 252, + 176 + ], + "lines": [ + { + "bbox": [ + 160, + 72, + 252, + 176 + ], + "spans": [ + { + "bbox": [ + 160, + 72, + 252, + 176 + ], + "type": "image", + "image_path": "2cc18da6cca93a8b5c3df31bac7ab4722ea3c3d49eb0f8b46f3dc1a1f05dfb45.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 255, + 60, + 351, + 71 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 60, + 351, + 71 + ], + "spans": [ + { + "bbox": [ + 255, + 60, + 351, + 71 + ], + "type": "text", + "content": "c) " + }, + { + "bbox": [ + 255, + 60, + 351, + 71 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 255, + 60, + 351, + 71 + ], + "type": "text", + "content": " [40] blend " + }, + { + "bbox": [ + 255, + 60, + 351, + 71 + ], + "type": "inline_equation", + "content": "v" + } + ] + } + ], + "index": 4, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 255, + 72, + 347, + 120 + ], + "blocks": [ + { + "bbox": [ + 255, + 72, + 347, + 120 + ], + "lines": [ + { + "bbox": [ + 255, + 72, + 347, + 120 + ], + "spans": [ + { + "bbox": [ + 255, + 72, + 347, + 120 + ], + "type": "image", + "image_path": "21cc1dd0431ac7f807f9d4cb3470d324905107b22fdef2f00b6fcc51388a164d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 351, + 72, + 441, + 120 + ], + "blocks": [ + { + "bbox": [ + 362, + 60, + 433, + 70 + ], + "lines": [ + { + "bbox": [ + 362, + 60, + 433, + 70 + ], + "spans": [ + { + "bbox": [ + 362, + 60, + 433, + 70 + ], + "type": "text", + "content": "(d) SAFF blend " + }, + { + "bbox": [ + 362, + 60, + 433, + 70 + ], + "type": "inline_equation", + "content": "v" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 351, + 72, + 441, + 120 + ], + "lines": [ + { + "bbox": [ + 351, + 72, + 441, + 120 + ], + "spans": [ + { + "bbox": [ + 351, + 72, + 441, + 120 + ], + "type": "image", + "image_path": "9e957d9085e676a91bfd7c9c567126cb875382a76823f9a1b0396fca531b832e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 443, + 72, + 533, + 120 + ], + "blocks": [ + { + "bbox": [ + 444, + 60, + 533, + 71 + ], + "lines": [ + { + "bbox": [ + 444, + 60, + 533, + 71 + ], + "spans": [ + { + "bbox": [ + 444, + 60, + 533, + 71 + ], + "type": "text", + "content": "(e) SAFF foregrounds" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 443, + 72, + 533, + 120 + ], + "lines": [ + { + "bbox": [ + 443, + 72, + 533, + 120 + ], + "spans": [ + { + "bbox": [ + 443, + 72, + 533, + 120 + ], + "type": "image", + "image_path": "2fa19228ccb41766b7bf0778ab8bac6cf30332c73aef241eaefc80e232489399.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 227, + 288, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 227, + 288, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 288, + 299 + ], + "type": "text", + "content": "NSFF [18] This method cannot produce semantic clusterings. While saliency and blend weight " + }, + { + "bbox": [ + 46, + 227, + 288, + 299 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 227, + 288, + 299 + ], + "type": "text", + "content": " have different meanings, if we compare our " + }, + { + "bbox": [ + 46, + 227, + 288, + 299 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 227, + 288, + 299 + ], + "type": "text", + "content": " to NSFF's, then we can see any impact of a shared backbone with attention heads upon the static/dynamic separation. We disable the supervised motion mask initialization in NSFF as our method does not use such information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 302, + 288, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 288, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^{2}\\mathbf{N}\\mathbf{e}\\mathbf{R}\\mathbf{F}" + }, + { + "bbox": [ + 46, + 302, + 288, + 363 + ], + "type": "text", + "content": " [40] This method also cannot produce semantic clusterings. Over HyperNeRF [26], it adds a shadow field network and further losses to try to isolate objects into the dynamic NeRF over the separate static one. The paper also compares to NSFF without motion mask initialization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 369, + 288, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 288, + 441 + ], + "type": "text", + "content": "DINO-ViT (2D) [1] We ignore the volume and pass 2D semantic and attention features into the clustering algorithm. This cannot apply to novel viewpoints. Instead, we evaluate the approach upon all multi-view color images—input and hold-out—whereas other methods must render hold-out views. With our added pyramid processing (Sec. 3.3)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 446, + 263, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 446, + 263, + 459 + ], + "spans": [ + { + "bbox": [ + 47, + 446, + 263, + 459 + ], + "type": "text", + "content": "- w/o pyr No pyramid; upsample to input RGB size." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 464, + 288, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 288, + 584 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 288, + 584 + ], + "type": "text", + "content": "ProposeReduce (2D) [19] As a general comparison point, we apply this state-of-the-art 2D video segmentation method. For object segmentation, we use a ProposeReduce network that was pretrained with supervision on YouTube-VIS 2019 [43] for instance segmentation. For foreground segmentation, we use weights pretrained on UVOS [5] intended specifically for unsupervised foreground segmentation. As ProposeReduce is only a 2D method, we provide it with hold-out images for splits with novel views rather than our approach that must render novel views at hold-out poses." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 591, + 111, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 111, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 111, + 604 + ], + "type": "text", + "content": "4.2. Findings" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Dynamic scene decomposition We separate the background and each foreground object individually (Tab. 1). The baseline 2D DINO-ViT method is improved by our pyramid approach. But, being only 2D, this fails to produce a consistent decomposition across novel spacetime views even when given ground truth hold-out images. This shows the value of a volume integration. Next, supervised ProposeReduce produces good results (Fig. 5), but sometimes misses salient objects or fails to join them behind occluding objects, and only" + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 309, + 285, + 544, + 432 + ], + "blocks": [ + { + "bbox": [ + 305, + 225, + 547, + 273 + ], + "lines": [ + { + "bbox": [ + 305, + 225, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 305, + 225, + 547, + 273 + ], + "type": "text", + "content": "Table 1: Spacetime volume integration improves dynamic scene decomposition. Pyramid construction and cluster merging help quantitatively, and ours is comparable to SOTA supervised 2D video segmentation network ProposeReduce." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 274, + 526, + 285 + ], + "lines": [ + { + "bbox": [ + 307, + 274, + 526, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 526, + 285 + ], + "type": "text", + "content": "Metric: Adjusted Rand Index " + }, + { + "bbox": [ + 307, + 274, + 526, + 285 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 307, + 274, + 526, + 285 + ], + "type": "text", + "content": ", higher is better." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 285, + 544, + 432 + ], + "lines": [ + { + "bbox": [ + 309, + 285, + 544, + 432 + ], + "spans": [ + { + "bbox": [ + 309, + 285, + 544, + 432 + ], + "type": "table", + "html": "
InputFix Cam 0Fix Time 0
ProposeReduce (2D)0.7250.7360.742
DINO-ViT (2D)0.5010.4950.321
w/o pyr帽子,hat0.4700.4640.346
SAFF (3D)0.5940.5780.566
SAFF (ours)0.6530.6340.625
w/ pyr λa = {1,0,0}0.6200.5980.592
w/o pyr帽子,hat0.5450.5320.521
w/o merge cluster0.5930.5740.563
w/post process0.7590.7330.735
w/oracle0.8340.8060.800
w/oracle + post process0.9220.8900.880
", + "image_path": "350ded1779f0cc54162f52260ead563617104562086209c2929a161f8df6657e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 441, + 546, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 546, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 546, + 478 + ], + "type": "text", + "content": "sometimes produces better edges than our method without post-processing as it can oversmooth edges. ProposeReduce also receives ground truth images in hold-out sets." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 479, + 547, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 547, + 587 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 547, + 587 + ], + "type": "text", + "content": "Instead, our approach must render hold-out views via the volume reconstruction. This produces more consistent segmentations through spacetime manipulations—this is the added value of volume integration through learned saliency and attention heads. Ablated components show the value of our pyramid step, its coarse-saliency-only variant, and the cluster merge and image post processing steps. Qualitatively, we see good detail (Fig. 5); post-processing additionally improves edge quality and removes small unwanted regions." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Oracle saliency We also include an oracle experiment where saliency voting is replaced by cluster selection based on ground truth masks. This experiment tells us what part of the performance gap lies with saliency itself, and what remains due to volume integration and cluster boundary issues. With oracle clusters, our decomposition performance is 0.8 ARI (Tab. 1) even in hold-out views. This shows that our existing cluster boundaries are accurate, and that accurate saliency for object selection is the larger remaining problem." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "21803" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 134, + 275, + 236 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 131 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 131 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 131 + ], + "type": "text", + "content": "Table 2: Saliency improves foreground segmentation. Adding saliency also slightly aids how much static/dynamic blend weight " + }, + { + "bbox": [ + 47, + 70, + 288, + 131 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 47, + 70, + 288, + 131 + ], + "type": "text", + "content": " represents the foreground (cf. NSFF blend " + }, + { + "bbox": [ + 47, + 70, + 288, + 131 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 47, + 70, + 288, + 131 + ], + "type": "text", + "content": " to SAFF's). Here, ProposeReduce uses unsupervised training and data [5]. Metric: IoU/Jaccard ([0, 1], higher is better)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 134, + 275, + 236 + ], + "lines": [ + { + "bbox": [ + 58, + 134, + 275, + 236 + ], + "spans": [ + { + "bbox": [ + 58, + 134, + 275, + 236 + ], + "type": "table", + "html": "
InputFix Cam 0Fix Time 0
ProposeReduce (2D)0.6090.4640.591
DINO-ViT (2D)0.3810.3820.357
NSFF — blend v0.3220.3090.268
D2NeRF— blend v0.4700.3340.269
SAFF — saliency0.6090.5890.572
— blend v0.3880.3800.329
— post process v0.7200.6940.679
", + "image_path": "cc037c42e5797e3aadb14db85e9ca407d4c278629fbb159978018f9ae1fcf99d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 323, + 285, + 392 + ], + "blocks": [ + { + "bbox": [ + 46, + 244, + 288, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 244, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 288, + 316 + ], + "type": "text", + "content": "Table 3: SAFF semantics generalize to DyCheck Dataset (Masked) Oracle saliency. ProposeReduce is given ground truth test frames, while SAFF must render them. SAFF is comparable to SOTA supervised 2D video segmentation network ProposeReduce. Metric: Adjusted Rand Index " + }, + { + "bbox": [ + 46, + 244, + 288, + 316 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 46, + 244, + 288, + 316 + ], + "type": "text", + "content": ", higher is better); IoU/Jaccard index " + }, + { + "bbox": [ + 46, + 244, + 288, + 316 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 46, + 244, + 288, + 316 + ], + "type": "text", + "content": ", higher is better)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 323, + 285, + 392 + ], + "lines": [ + { + "bbox": [ + 50, + 323, + 285, + 392 + ], + "spans": [ + { + "bbox": [ + 50, + 323, + 285, + 392 + ], + "type": "table", + "html": "
ARIIoU
InputTestInputTest
ProposeReduce (2D)0.7610.7620.7610.762
DINO-ViT (2D)0.8010.8000.7970.797
SAFF (3D)0.9020.8200.9100.812
", + "image_path": "38078e65d1eed18ee574ddb69b341f59d29e0a4b8545ee092f7d422d3faf9930.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 405, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 287, + 525 + ], + "type": "text", + "content": "Foreground segmentation We simplify the problem and consider all objects as simply 'foreground' to compare to methods that do not produce per object masks. Here, the same trend continues (Tab. 2). We note more subtle improvements to static/dynamic blending weights when adding our additional feature heads to the backbone MLP, and overall show that adding top-down information helps produce more useful object masks. Qualitative results show whole objects in the foreground rather than broad regions of possible dynamics (NSFF) or broken objects " + }, + { + "bbox": [ + 46, + 405, + 287, + 525 + ], + "type": "inline_equation", + "content": "(\\mathrm{D}^2\\mathrm{NeRF};" + }, + { + "bbox": [ + 46, + 405, + 287, + 525 + ], + "type": "text", + "content": " Fig. 6)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 532, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 288, + 628 + ], + "type": "text", + "content": "DyCheck evaluation DyCheck often has close-up objects and, even with our selected sequences, saliency struggles to find foreground objects. Thus, we use oracle saliency. DINOViT and ProposeReduce are given test views while SAFF must render them. Quantitative (Tab. 3) and qualitative (Fig. 7) experiments show similar trends as before: ProposeReduce is good but may miss objects and fine details; SAFF may produce finer details and is more geometrically consistent." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 639, + 200, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 200, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 200, + 651 + ], + "type": "text", + "content": "5. Discussion and Limitations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "DINO-ViT features are not instance-aware (Fig. 8). This is in contrast to object-centric learning approaches that aim to identify individual objects. To represent these different approaches, we compare to a result from slot-attention based " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SAVi}++" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " [7]. This method trains on thousands of supervised" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 79, + 366, + 155 + ], + "blocks": [ + { + "bbox": [ + 329, + 70, + 547, + 78 + ], + "lines": [ + { + "bbox": [ + 329, + 70, + 547, + 78 + ], + "spans": [ + { + "bbox": [ + 329, + 70, + 547, + 78 + ], + "type": "text", + "content": "Input ProposeReduce [19] SAFF (Ours; input) SAFF (Ours; novel)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 79, + 366, + 155 + ], + "lines": [ + { + "bbox": [ + 309, + 79, + 366, + 155 + ], + "spans": [ + { + "bbox": [ + 309, + 79, + 366, + 155 + ], + "type": "image", + "image_path": "cdd7b3c6f9d1e1e1425caa90ae6f392fa17fe43167b54091676d627962928cfa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 156, + 367, + 232 + ], + "blocks": [ + { + "bbox": [ + 309, + 156, + 367, + 232 + ], + "lines": [ + { + "bbox": [ + 309, + 156, + 367, + 232 + ], + "spans": [ + { + "bbox": [ + 309, + 156, + 367, + 232 + ], + "type": "image", + "image_path": "273f11278c90df905bb04230318be2162cc45f2dd6ab17510c1db4b067fbc468.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 234, + 547, + 294 + ], + "lines": [ + { + "bbox": [ + 305, + 234, + 547, + 294 + ], + "spans": [ + { + "bbox": [ + 305, + 234, + 547, + 294 + ], + "type": "text", + "content": "Figure 7: SAFF can apply to DyCheck Dataset (Masked). SAFF is comparable to supervised large-scale 2D video learning [19]. Some fine details are improved (bottom, fingers) and novel views maintain their geometry (bottom, gap between legs occluded by head at input timestep)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 369, + 79, + 427, + 155 + ], + "blocks": [ + { + "bbox": [ + 369, + 79, + 427, + 155 + ], + "lines": [ + { + "bbox": [ + 369, + 79, + 427, + 155 + ], + "spans": [ + { + "bbox": [ + 369, + 79, + 427, + 155 + ], + "type": "image", + "image_path": "e822e330344030109a8f27f3b32ab14a8a9a04c53f39ea6bf92763540d178a21.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 369, + 156, + 427, + 232 + ], + "blocks": [ + { + "bbox": [ + 369, + 156, + 427, + 232 + ], + "lines": [ + { + "bbox": [ + 369, + 156, + 427, + 232 + ], + "spans": [ + { + "bbox": [ + 369, + 156, + 427, + 232 + ], + "type": "image", + "image_path": "f118d418e44feb85ea722ffb224829fa95d77cb850e703874a4d2dcaedeb8466.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 429, + 79, + 487, + 155 + ], + "blocks": [ + { + "bbox": [ + 429, + 79, + 487, + 155 + ], + "lines": [ + { + "bbox": [ + 429, + 79, + 487, + 155 + ], + "spans": [ + { + "bbox": [ + 429, + 79, + 487, + 155 + ], + "type": "image", + "image_path": "41e9bf5dedef7a9aa4433f49ac7d51e411c80fd64d4ff68c6fd4a6c02c4dc620.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 429, + 156, + 487, + 232 + ], + "blocks": [ + { + "bbox": [ + 429, + 156, + 487, + 232 + ], + "lines": [ + { + "bbox": [ + 429, + 156, + 487, + 232 + ], + "spans": [ + { + "bbox": [ + 429, + 156, + 487, + 232 + ], + "type": "image", + "image_path": "d63c51f4f565fe03ce474325545808eb92b7f3beafc4605e703a8166445f1dad.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 489, + 79, + 548, + 155 + ], + "blocks": [ + { + "bbox": [ + 489, + 79, + 548, + 155 + ], + "lines": [ + { + "bbox": [ + 489, + 79, + 548, + 155 + ], + "spans": [ + { + "bbox": [ + 489, + 79, + 548, + 155 + ], + "type": "image", + "image_path": "662699dfb3e9528689d2f33db4a4c8da2c57f93c1e930c0ca44483f2db93a7b7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 489, + 156, + 548, + 232 + ], + "blocks": [ + { + "bbox": [ + 489, + 156, + 548, + 232 + ], + "lines": [ + { + "bbox": [ + 489, + 156, + 548, + 232 + ], + "spans": [ + { + "bbox": [ + 489, + 156, + 548, + 232 + ], + "type": "image", + "image_path": "c894697ca2b290957c8e39c04bb2ec7be403cd2d62bcad781bbbe2c4acfe88a0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 309, + 312, + 384, + 367 + ], + "blocks": [ + { + "bbox": [ + 329, + 300, + 365, + 312 + ], + "lines": [ + { + "bbox": [ + 329, + 300, + 365, + 312 + ], + "spans": [ + { + "bbox": [ + 329, + 300, + 365, + 312 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 312, + 384, + 367 + ], + "lines": [ + { + "bbox": [ + 309, + 312, + 384, + 367 + ], + "spans": [ + { + "bbox": [ + 309, + 312, + 384, + 367 + ], + "type": "image", + "image_path": "6729bdccffeb5d82405d77c40eecf02828d678f74eb3d84859773e7e5f4a06ba.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 387, + 312, + 463, + 367 + ], + "blocks": [ + { + "bbox": [ + 396, + 300, + 456, + 312 + ], + "lines": [ + { + "bbox": [ + 396, + 300, + 456, + 312 + ], + "spans": [ + { + "bbox": [ + 396, + 300, + 456, + 312 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 396, + 300, + 456, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{SAV_i + + [7]}" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 387, + 312, + 463, + 367 + ], + "lines": [ + { + "bbox": [ + 387, + 312, + 463, + 367 + ], + "spans": [ + { + "bbox": [ + 387, + 312, + 463, + 367 + ], + "type": "image", + "image_path": "988c7a0f5473de6d3d76cb7af8e50e1eb2966cae1592553828303c6e8169f2ae.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 466, + 312, + 542, + 366 + ], + "blocks": [ + { + "bbox": [ + 486, + 300, + 525, + 310 + ], + "lines": [ + { + "bbox": [ + 486, + 300, + 525, + 310 + ], + "spans": [ + { + "bbox": [ + 486, + 300, + 525, + 310 + ], + "type": "text", + "content": "(c) SAFF" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 466, + 312, + 542, + 366 + ], + "lines": [ + { + "bbox": [ + 466, + 312, + 542, + 366 + ], + "spans": [ + { + "bbox": [ + 466, + 312, + 542, + 366 + ], + "type": "image", + "image_path": "e8844aa33ab7268e2befb4420a8bdff4f44183a0e86e2b9956937aa9eb001cd7.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 309, + 381, + 384, + 422 + ], + "blocks": [ + { + "bbox": [ + 329, + 370, + 365, + 381 + ], + "lines": [ + { + "bbox": [ + 329, + 370, + 365, + 381 + ], + "spans": [ + { + "bbox": [ + 329, + 370, + 365, + 381 + ], + "type": "text", + "content": "(d) Input" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 381, + 384, + 422 + ], + "lines": [ + { + "bbox": [ + 309, + 381, + 384, + 422 + ], + "spans": [ + { + "bbox": [ + 309, + 381, + 384, + 422 + ], + "type": "image", + "image_path": "cc9c34111c72b5cd2873e5abb927357374f13c87078284374297bf3b4b5a6620.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 427, + 547, + 464 + ], + "lines": [ + { + "bbox": [ + 305, + 427, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 427, + 547, + 464 + ], + "type": "text", + "content": "Figure 8: Limitations. DINO-ViT is not instance-aware, causing unwanted grouping (c). Unwanted static objects may be salient; assuming salient objects are dynamic fixes this (f)." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 385, + 381, + 461, + 422 + ], + "blocks": [ + { + "bbox": [ + 391, + 370, + 456, + 381 + ], + "lines": [ + { + "bbox": [ + 391, + 370, + 456, + 381 + ], + "spans": [ + { + "bbox": [ + 391, + 370, + 456, + 381 + ], + "type": "text", + "content": "(e) SAFF salient" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 385, + 381, + 461, + 422 + ], + "lines": [ + { + "bbox": [ + 385, + 381, + 461, + 422 + ], + "spans": [ + { + "bbox": [ + 385, + 381, + 461, + 422 + ], + "type": "image", + "image_path": "d0b34e16db9794f6f8c60fa715c6dfee06a9337e01c9fc2539e48867a77d9b04.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 464, + 381, + 541, + 422 + ], + "blocks": [ + { + "bbox": [ + 463, + 370, + 542, + 381 + ], + "lines": [ + { + "bbox": [ + 463, + 370, + 542, + 381 + ], + "spans": [ + { + "bbox": [ + 463, + 370, + 542, + 381 + ], + "type": "text", + "content": "(f) Salient+dynamic" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 464, + 381, + 541, + 422 + ], + "lines": [ + { + "bbox": [ + 464, + 381, + 541, + 422 + ], + "spans": [ + { + "bbox": [ + 464, + 381, + 541, + 422 + ], + "type": "image", + "image_path": "bfdbd6632afcaeb6b3475a898124129ca7c741b11ecfe7f74ba05018b82b489f.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 471, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 471, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 305, + 471, + 545, + 519 + ], + "type": "text", + "content": "MOVi [11] sequences with per-object masks, whereas we use generic pre-trained features and gain better edges from volume integration. Combining these two approaches could give accurate instance-level scene objects." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 519, + 546, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 546, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 546, + 663 + ], + "type": "text", + "content": "DINO-ViT saliency may attend to unwanted regions. In Figure 8d-e, the static pillars could be isolated using scene flow. But, often our desired subjects do not move (cf. people in Umbrella or Balloon NBoard). For tasks or data that can assume that salient objects are dynamic, we use SAFF's 4D scene reconstruction to reject static-but-salient objects by merging clusters via scene flow: First, we project " + }, + { + "bbox": [ + 304, + 519, + 546, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 304, + 519, + 546, + 663 + ], + "type": "text", + "content": " over each timestep into each input camera pose—this simulates optical flow with a static camera. Clusters are marked as salient per image if mean flow magnitude per cluster " + }, + { + "bbox": [ + 304, + 519, + 546, + 663 + ], + "type": "inline_equation", + "content": "|\\bar{\\mathbf{p}}| > 0.07" + }, + { + "bbox": [ + 304, + 519, + 546, + 663 + ], + "type": "text", + "content": " and mean attention " + }, + { + "bbox": [ + 304, + 519, + 546, + 663 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{a}}_c > 0.07" + }, + { + "bbox": [ + 304, + 519, + 546, + 663 + ], + "type": "text", + "content": ". Finally, as before, a cluster is globally salient if " + }, + { + "bbox": [ + 304, + 519, + 546, + 663 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 304, + 519, + 546, + 663 + ], + "type": "text", + "content": " of images agree (Fig. 8f)." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgements The CV community in New England for feedback, and funding from NSF CNS-2038897 and an Amazon Research Award. Eliot thanks a Randy F. Pausch '82 Undergraduate Summer Research Award at Brown CS." + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "21804" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2021. 2, 4, 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 223 + ], + "type": "text", + "content": "[3] Kenneth Blomqvist, Lionel Ott, Jen Jen Chung, and Roland Y. Siegwart. Baking in the feature: Accelerating volumetric segmentation by rendering feature maps. ArXiv, abs/2209.12744, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 269 + ], + "type": "text", + "content": "[4] Christopher P. Burgess, Loic Matthews, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matthew M. Botvinick, and Alexander Lerchner. Monet: Unsupervised scene decomposition and representation. ArXiv, abs/1901.11390, 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 270, + 288, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 288, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 288, + 313 + ], + "type": "text", + "content": "[5] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv:1905.00737, 2019. 2, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 315, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 315, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 315, + 288, + 369 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 1, 2, 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 370, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 370, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 370, + 287, + 424 + ], + "type": "text", + "content": "[7] Gamaeldin F. Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael C. Mozer, and Thomas Kipf. " + }, + { + "bbox": [ + 53, + 370, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\mathrm{SAVi}++" + }, + { + "bbox": [ + 53, + 370, + 287, + 424 + ], + "type": "text", + "content": ": Towards end-to-end object-centric learning from real-world videos. In Advances in Neural Information Processing Systems, 2022. 1, 2, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 426, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 288, + 468 + ], + "type": "text", + "content": "[8] Zhiwen Fan, Peihao Wang, Yifan Jiang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation on complex scenes. ArXiv, abs/2209.08776, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 470, + 288, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 470, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 470, + 288, + 524 + ], + "type": "text", + "content": "[9] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In International Conference on 3D Vision (3DV), 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "type": "text", + "content": "[10] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. In NeurIPS, 2022. 1, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 690 + ], + "type": "text", + "content": "[11] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam H. Laradji, Hsueh-Ti Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: A scalable dataset generator. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3739-3751, 2022. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "type": "text", + "content": "[12] Klaus Greff, Raphael Lopez Kaufman, Rishabh Kabra, Nicholas Watters, Christopher P. Burgess, Daniel Zoran, Loic" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "type": "text", + "content": "Matthew, Matthew M. Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative variational inference. In ICML, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 107, + 547, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 547, + 152 + ], + "type": "text", + "content": "[13] Basile Van Hoorick, Purva Tendulka, Dídac Surís, Dennis Park, Simon Stent, and Carl Vondrick. Revealing occlusions with 4d neural fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3001-3011, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 153, + 547, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 153, + 547, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 153, + 547, + 207 + ], + "type": "text", + "content": "[14] Rishabh Kabra, Daniel Zoran, Goker Erdogan, Loic Matthey, Antonia Creswell, Matthew M. Botvinick, Alexander Lerchner, and Christopher P. Burgess. Simone: View-invariant, temporally-abstracted object representations via unsupervised video decomposition. ArXiv, abs/2106.03849, 2021. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 209, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 547, + 262 + ], + "type": "text", + "content": "[15] Thomas Kipf, Gamaleldin F. Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional Object-Centric Learning from Video. In International Conference on Learning Representations (ICLR), 2022. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 264, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 547, + 308 + ], + "type": "text", + "content": "[16] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. In Advances in Neural Information Processing Systems, volume 35, 2022. 1, 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 309, + 547, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 547, + 376 + ], + "type": "text", + "content": "[17] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J. Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12871-12881, June 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 376, + 547, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 547, + 431 + ], + "type": "text", + "content": "[18] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 433, + 547, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 547, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 547, + 477 + ], + "type": "text", + "content": "[19] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1719-1728, 2021. 1, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 478, + 547, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 547, + 532 + ], + "type": "text", + "content": "[20] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. arXiv preprint arXiv:2006.15055, 2020.1,2,5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 534, + 547, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 547, + 567 + ], + "type": "text", + "content": "[21] Kirill Mazur, Edgar Sucar, and Andrew J. Davison. Feature-realistic neural fusion for real-time, open set scene understanding. ArXiv, abs/2210.03043, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 568, + 547, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 547, + 612 + ], + "type": "text", + "content": "[22] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In CVPR, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 613, + 547, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 547, + 657 + ], + "type": "text", + "content": "[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 658, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 547, + 713 + ], + "type": "text", + "content": "[24] Tom Monnier, Elliot Vincent, Jean Ponce, and Mathieu Aubry. Unsupervised layered image decomposition into object prototypes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8640-8650, October 2021. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "21805" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[25] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 288, + 172 + ], + "type": "text", + "content": "[26] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), dec 2021. 1, 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 227 + ], + "type": "text", + "content": "[27] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 259 + ], + "type": "text", + "content": "[28] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 316 + ], + "type": "text", + "content": "[29] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, and Francesco Locatello. Bridging the gap to real-world object-centric learning. ArXiv, abs/2209.14860, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 288, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 347 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 347 + ], + "type": "text", + "content": "[30] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In CVPRW, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 288, + 392 + ], + "type": "text", + "content": "[31] Cameron Smith, Hong-Xing Yu, Sergey Zakharov, Frédo Durand, Joshua B. Tenenbaum, Jiajun Wu, and Vincent Sitzmann. Unsupervised discovery and composition of object light fields. ArXiv, abs/2205.03923, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 394, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 288, + 426 + ], + "type": "text", + "content": "[32] Karl Stelzner, Kristian Kersting, and Adam R. Kosiorek. Decomposing 3d scenes into objects via unsupervised volume segmentation. ArXiv, abs/2104.01148, 2021. 1, 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 525 + ], + "type": "text", + "content": "[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 526, + 288, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 288, + 559 + ], + "type": "text", + "content": "[34] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 624 + ], + "type": "text", + "content": "[35] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12939-12950, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "text", + "content": "[36] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural Feature Fusion Fields: 3D distillation of self-supervised 2D image representations. In Proceedings of the International Conference on 3D Vision (3DV), 2022. 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[37] Vadim Tschernezki, Diane Larlus, and Andrea Vedaldi. NeuralDiff: Segmenting 3D objects that move in egocentric videos. In Proceedings of the International Conference on 3D Vision (3DV), 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 575 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 160 + ], + "type": "text", + "content": "[38] Rishi Veerapaneni, John D. Co-Reyes, Michael Chang, Michael Janner, Chelsea Finn, Jiajun Wu, Joshua Tenenbaum, and Sergey Levine. Entity abstraction in visual model-based reinforcement learning. In Leslie Pack Kaelbling, Danica Kragic, and Komei Sugiura, editors, Proceedings of the Conference on Robot Learning, volume 100 of Proceedings of Machine Learning Research, pages 1439-1456. PMLR, 30 Oct-01 Nov 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 162, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 547, + 217 + ], + "type": "text", + "content": "[39] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L. Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 217, + 546, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 546, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 546, + 261 + ], + "type": "text", + "content": "[40] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Öz Tireli. D2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. ArXiv, abs/2205.15838, 2022. 1, 2, 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 262, + 547, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 262, + 547, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 547, + 306 + ], + "type": "text", + "content": "[41] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 308, + 546, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 546, + 361 + ], + "type": "text", + "content": "[42] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for editable scene rendering. In International Conference on Computer Vision (ICCV), October 2021. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 363, + 546, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 363, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 363, + 546, + 396 + ], + "type": "text", + "content": "[43] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Oct 2019. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 397, + 546, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 397, + 546, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 397, + 546, + 441 + ], + "type": "text", + "content": "[44] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snavely. Deformable sprites for unsupervised video decomposition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 441, + 546, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 441, + 546, + 496 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 546, + 496 + ], + "type": "text", + "content": "[45] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 498, + 546, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 546, + 530 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 546, + 530 + ], + "type": "text", + "content": "[46] Hong-Xing Yu, Leonidas J. Guibas, and Jiajun Wu. Unsupervised discovery of object radiance fields. In International Conference on Learning Representations, 2022. 1, 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 532, + 547, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 547, + 575 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 547, + 575 + ], + "type": "text", + "content": "[47] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "21806" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Semantic Information in Contrastive Learning/c3144656-62d8-4975-97d6-ef4630f410df_content_list.json b/2023/Semantic Information in Contrastive Learning/c3144656-62d8-4975-97d6-ef4630f410df_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a036b952980aadd8170d2edbc4cec7118ab6bdfd --- /dev/null +++ b/2023/Semantic Information in Contrastive Learning/c3144656-62d8-4975-97d6-ef4630f410df_content_list.json @@ -0,0 +1,1699 @@ +[ + { + "type": "text", + "text": "Semantic Information in Contrastive Learning", + "text_level": 1, + "bbox": [ + 248, + 130, + 722, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shengjiang Quan $^{1*}$ , Masahiro Hirano $^{2}$ and Yuji Yamakawa $^{3}$ \\\n $^{1}$ Graduate School of Engineering, The University of Tokyo, Japan\\\n $^{2}$ Institute of Industrial Science, The University of Tokyo, Japan\\\n $^{3}$ Interfaculty Initiative in Information Studies, The University of Tokyo, Japan\\\nquan@g.ecc.u-tokyo.ac.jp, {mhirano, y-ymkw}@iis.u-tokyo.ac.jp", + "bbox": [ + 176, + 179, + 792, + 271 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This work investigates the functionality of Semantic information in Contrastive Learning (SemCL). An advanced pretext task is designed: a contrast is performed between each object and its environment, taken from a scene. This allows the SemCL pretrained model to extract objects from their environment in an image, significantly improving the spatial understanding of the pretrained models. Downstream tasks of semantic/instance segmentation, object detection and depth estimation are implemented on PASCAL VOC, Cityscapes, COCO, KITTI, etc. SemCL pretrained models substantially outperform ImageNet pretrained counterparts and are competitive with well-known works on downstream tasks. The results suggest that a dedicated pretext task leveraging semantic information can be powerful in benchmarks related to spatial understanding. The code is available at https://github.com/sjiang95/semcl.", + "bbox": [ + 75, + 335, + 473, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 607, + 209, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Within the field of visual representation learning, contrastive learning attracts special attention for its inspiring performance in transfer learning [57, 49, 31, 22, 8, 9]. The concept of contrastive learning can literally be explained as discovering the difference between positive and negative samples. The definition of positive-negative samples is one of the main subjects of contrastive learning, which directly determines the pretext task and the corresponding loss function.", + "bbox": [ + 75, + 633, + 470, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We humans can recognize an object from its even complicated environment because we have learned and bound the typical geometric feature and the corresponding semantic concept of the object class. For example, we can recognize a cat in a scene thanks to our knowledge of its appearance (typical geometric feature) and the concept of cat (semantic information). Similarly, the target of prevalent", + "bbox": [ + 75, + 770, + 470, + 877 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bab89fd9041f4a80e562caeb78e23e0262a1f6e2656bcaff36d7a6fdd19d7503.jpg", + "image_caption": [ + "Figure 1: Perceptual cognition of SemCL. Leveraging semantic information, SemCL separates an object from its environment. Each sample is divided into objects and surroundings and considered as a contrastive pair. The task is to maximize the distance between the object and its environment in the embedding space." + ], + "image_footnote": [], + "bbox": [ + 504, + 306, + 893, + 452 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "visual tasks, such as semantic/instance segmentation, object detection, etc., can be summarized as distinguishing and localizing subjects from surroundings. Inspired by and to extend the spatial recognition mechanism to CV, we propose our method SemCL which directly teaches models to extract a subject from its environment. Utilizing publicly off-the-shelf datasets providing semantic labels, SemCL can be considered as a supervised contrastive learning approach whose samples consist of subjects and their corresponding surroundings (semantically not the subject). The pretext task of SemCL is to tell the difference between one subject and its surroundings. From a representation learning perspective, the pretext task is to maximize the distance $(L^2$ norm) between a subject and its environment in the embedding space (see Figure 1). The pretext task mimics the spatial recognition mechanism to discriminate objects and surroundings, which significantly improves the spatial information understanding of pretrained models.", + "bbox": [ + 496, + 583, + 893, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In SemCL, contrast is performed pairwise between positive-negative pairs from one image/scene. Inter-scene contrast is considered inappropriate: only the contrasts", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5686", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "among subject-surrounding pairs from one scene are considered valid. Therefore, paired InfoNCE loss is adopted. With the mechanism of contrasting only paired samples, SemCL is characterized by decoupling the number of negative samples from the batch size. Compared to MoCo [22], whose performance is positively correlated with batch size and thus increases hardware cost (e.g. batchsize $= 4096$ on 128 GPUs for ViT-B in MoCo v3 [9]) for satisfactory results, SemCL can achieve consistent yet competitive results on downstream tasks with relatively small batch size (e.g., batchsize $= 64 \\ @ 224 \\times 224$ ). We adopt the MoCo v3 [9] as the pretraining framework.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The primary motivation of representation learning is to pretrain general representations that can be fine-tuned and transferred to downstream tasks. In this work, the SemCL pretrained models are benchmarked on semantic segmentation, object detection and instance segmentation, and depth estimation tasks. SemCL models substantially outperform their ImageNet pretrained counterparts due to the improved ability of spacial information understanding. In a system-level comparison, SemCL models also show gains over previous well-known works (e.g. Deeplabv3+ [7] in semantic segmentation, Mask2Former [10] in instance segmentation and Binsformer [35] in depth estimation).", + "bbox": [ + 75, + 273, + 472, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 470, + 218, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Contrastive Learning", + "text_level": 1, + "bbox": [ + 76, + 497, + 277, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The seminal idea of contrastive learning dates back at least to the 1990s [2, 3, 31], but it has become prominent in recent years thanks to the large pretrained models in the fields of NLP and CV [31]. As the name suggests, contrastive learning mines abstract representations by comparing between \"similar\" and \"dissimilar\" samples: minimizing the representation of \"similar\" samples and maximizing that of \"dissimilar\" samples in the embedding space. Formally, the \"similar\" inputs are called positive samples, and the \"dissimilar\" inputs are called negative samples.", + "bbox": [ + 75, + 521, + 472, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Defining \"similar\" and \"dissimilar\" samples is an important topic in contrastive learning. In instance discrimination [57], which is a ubiquitous approach to define positive-negative samples in CV, each sample in a minibatch $x \\in \\{x_1, \\ldots, x_n\\}$ is considered mutually exclusive to the rest. The encoded output $\\theta(\\cdot)$ of each sample is pushed away from the others in the embedding space, denoted as $\\theta(x_1) \\Leftrightarrow \\{\\theta(x_2), \\ldots, \\theta(x_n)\\} \\Leftrightarrow \\text{push}$ . Contrastive Predictive Coding (CPC) [49] defines positive-negative samples in a generative way: Given a sequential input $\\{x_{t-3}, x_{t-2}, x_{t-1}, x_t\\}$ , a context latent representation $C_t$ is computed to predict the encoded outputs of future inputs $\\{\\theta(\\mathbb{X}_{t+1}), \\theta(\\mathbb{X}_{t+2}), \\theta(\\mathbb{X}_{t+3}), \\ldots\\}$ . Those predicted future embeddings are considered as positive samples to real future embeddings, and exclusive to some other", + "bbox": [ + 75, + 674, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "irrelevant inputs.", + "bbox": [ + 500, + 92, + 614, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Momentum Contrast", + "text_level": 1, + "bbox": [ + 500, + 114, + 697, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "He et al. [22] proposed a momentum contrast framework, MoCo, for contrastive representation learning. MoCo summed unsupervised visual representation learning [57, 49, 26, 69, 25, 48, 1] up as dictionary look-up: the input data are represented and sampled by an encoder network $\\theta_{k}$ as \"keys\" $k = \\theta_{k}(x^{k})$ . The goal of contrastive learning is to train encoders $\\theta_{q}$ by which a \"query\" $q = \\theta_{q}(x^{q})$ should be similar to its matching key and dissimilar to others. Since dictionary consistency is considered crucial for unsupervised learning with contrastive loss [22], MoCo introduces momentum updating. In the MoCo framework, the key coder $\\theta_{k}$ is progressively updated by the weighted average of $\\theta_{k}$ and the query coder $\\theta_{q}$ .", + "bbox": [ + 496, + 138, + 893, + 335 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {k} \\leftarrow m \\theta_ {k} + (1 - m) \\theta_ {q}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 344, + 893, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where the momentum coefficient $m \\in [0,1)$ is set quite high ( $m = 0.999$ in [22]) for slower key encoder update. The MoCo series has proven the power of the Momentum Contrast framework on feature transferring and is therefore employed as the basis for this work.", + "bbox": [ + 496, + 369, + 893, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Labeled Data In Unsupervised Framework", + "text_level": 1, + "bbox": [ + 500, + 453, + 864, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large networks are quite thirsty for training data, while labelling the large amount of collected data is a human resource consuming challenge. Apart from turning to unsupervised learning, which inherently avoids using labelled data, using unlabeled data in supervised learning is also a solution. Pseudo-Label [44, 38, 56] is a trick that uses (usually supervised) pretrained networks to generate features that can play the role of labels. And the generated pseudo-labels participate in the iterations of the target network just like annotations.", + "bbox": [ + 496, + 476, + 893, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, Khosla et al. [28] also investigated the function of labelled data in unsupervised learning. As a contrastive learning research, Khosla et al. [28] intuitively uses the classification label to define samples within the same class are mutually positive, and those from other categories are negative. To exploit the label information, clusters of points belonging to the same class are pulled together in the embedding space, while clusters of samples from different classes are pushed apart [28]: $\\left\\{\\theta\\left(x_1^{c0}\\right) \\rightleftharpoons \\theta\\left(x_2^{c0}\\right) \\rightleftharpoons \\theta\\left(x_{n}^{c0}\\right) \\rightleftharpoons \\theta\\left(x_n^0\\right)$ $\\Leftrightarrow$ $\\left\\{\\theta\\left(x_1^{c1}\\right) \\rightleftharpoons \\theta\\left(x_2^{c1}\\right) \\rightleftharpoons \\theta\\left(x_{m}^{c1}\\right)$ ( $\\rightleftharpoons$ for pull). The experimental result shows that supervised contrastive learning can improve both the accuracy and robustness of classifiers - [28] achieves a top-1 accuracy of $81.4\\%$ on the ImageNet (IN) dataset, which is $0.8\\%$ above the best result reported by ResNet-200 architecture [24]. Likewise, our work investigates the effectiveness of semantic information in unsupervised framework contrastive learning.", + "bbox": [ + 496, + 627, + 895, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5687", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.4. Semantic Information in Contrastive Learning", + "text_level": 1, + "bbox": [ + 76, + 90, + 470, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Other than [28], there are works that use small amounts of labelled data in unsupervised representation learning. Wang et al. [54] proposed a pixel-wise contrastive algorithm for semantic segmentation: for a pixel $i$ , it is pulled towards pixels belonging to the same category and pushed away from pixels in other categories in the embedding space. This contrast is performed across images from the same dataset. In the Cityscapes test benchmark, [54] R101 achieves 79.2 mIoU, which is 1.1 points better than Deeplabv3 [6].", + "bbox": [ + 75, + 113, + 468, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The attempt and results of works using labels in unsupervised representation learning, such as [28, 54], indicate that label information can bring gains to a fully unsupervised representation. It is promising to explore advanced pretext tasks to improve the performance of pretrained models in downstream tasks.", + "bbox": [ + 75, + 250, + 470, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 352, + 168, + 368 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. SemCL Dataset", + "text_level": 1, + "bbox": [ + 76, + 376, + 233, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The core of this work is to define semantically contrastive pairs by utilizing off-the-shelf semantic labels, which are mappings between categories of objects and their coordinates on the raw image at the pixel level. As decided by the authors of a dataset, semantic labels can be color or grayscale images. Each predefined color or grayscale denotes a category, and the distribution of label pixels reflects the location of pixels belonging to a particular category in the raw image.", + "bbox": [ + 75, + 400, + 468, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The SemCL dataset is generated by the mechanism shown in Figure 2, where the raw image and its semantic label are selected from the PASCAL VOC2012 dataset [14]. From the semantic label annotating two objects - a motorcycle and a rider - two binary masks BM0 and BM1 can be extracted to separately represent two annotated objects. Their inverted counterparts $\\neg$ BM0 and $\\neg$ BM1 represent pixels other than the target object. These binary masks are then bitwise applied to the raw image to separate an object (Anchor) from its environment ( $\\neg$ Anchor), forming a semantically contrastive pair. Each element in a raw Anchor- $\\neg$ Anchor pair is considered to be semantically opposite to the other.", + "bbox": [ + 75, + 521, + 468, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In practice, the SemCL-1M dataset is produced from training sets of augmented PASCAL VOC2012 [14] (also known as SBD [20]), Cityscapes [11], ADE20K [67, 68] and COCO [36] (stuff+thing 2017 [4]) datasets. The components of the SemCL dataset are illustrated in Table 1. When generating the dataset, a threshold $t = 0.01$ is set to filter out too small (less than $[\\sqrt{t} \\times \\text{height}] \\times [\\sqrt{t} \\times \\text{width}]$ ) objects that are considered semantically unsaturated.", + "bbox": [ + 75, + 702, + 470, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Paired InfoNCE Loss", + "text_level": 1, + "bbox": [ + 76, + 830, + 277, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since contrastive learning can be considered as a dictionary look-up task [22], it is assumed that within a set of keys $\\{k_0, k_1, k_2, \\ldots\\}$ of a dictionary, a query $q$ matches only one", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/4f0e152aa9bdfd04cd415a97c9454b917d14019e511435d7455f721b3c307df8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VOC2012CityscapesADE20KCOCOTotal
# training samples10,5822,97525,574118,287157,418
# SemCL pairs14,20320,280270,218712,2121,016,913
", + "bbox": [ + 501, + 88, + 893, + 161 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1: Components of the SemCL dataset. The SemCL datasets spawned from VOC2012(aug), Cityscapes, ADE20k and COCO 2017 are named SemCL- $\\{\\mathrm{VOC}$ , City, ADE, COCO\\} respectively.", + "bbox": [ + 496, + 170, + 893, + 232 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "key $k_{+}$ of the dictionary [22]. The value of the objective function should be low if $q$ is similar to its positive key $k_{+}$ and dissimilar to all negative keys $\\{k_{0}, k_{1}, k_{2}, \\ldots\\} \\setminus \\{k_{+}\\}$ . In MoCo [22], the form of a contrastive loss function InfoNCE [49] is adopted", + "bbox": [ + 496, + 258, + 893, + 335 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {q} = - \\log \\frac {\\exp (q \\cdot k _ {+} / \\tau)}{\\sum_ {i = 0} ^ {K} \\exp (q \\cdot k _ {i} / \\tau)} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 345, + 892, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau$ is a scalar temperature hyperparameter controlling the strength of penalties on hard negative samples [50], e.g. in the instance discrimination task [49, 22], there may be potential positive samples in the randomly chosen negative samples. Summing over one positive and $K$ negative samples, Equation (2) is the log loss of a $(\\mathbf{K} + 1)$ -way softmax-based classifier trying to classify $q$ as $k_{+}$ [22]. While in our case an anchor instance $q$ compares only with its own $n$ augmentations (positive) $K^{+} := \\{k_{1}^{+}, k_{2}^{+}, k_{3}^{+}, \\ldots, k_{n}^{+}\\}$ and paired $n + 1$ negatives $K^{-} := \\{k_{0}^{-}, k_{1}^{-}, k_{2}^{-}, k_{3}^{-}, \\ldots, k_{n}^{-}\\}$ using paired InfoNCE", + "bbox": [ + 496, + 392, + 893, + 559 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {q} = - \\sum_ {j = t + 1} ^ {n} \\log \\frac {\\exp \\left(q _ {t} \\cdot k _ {j} ^ {+} / \\tau\\right)}{\\sum_ {i = 0} ^ {K ^ {-}} \\exp \\left(q _ {t} \\cdot k _ {i} ^ {-} / \\tau\\right) + \\exp \\left(q _ {t} \\cdot k _ {j} ^ {+} / \\tau\\right)}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 571, + 901, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For one anchor as query $q_{0}(t = 0)$ , the paired InfoNCE loss tries to classify it as $k_{j}^{+}, j \\in [1, n]$ . The next step is to go forward: classify the first positive $q_{1}(t = 1)$ of the anchor as $k_{j}^{+}, j \\in [2, n]$ , and so on, as shown in the algorithm 1, Appendix A.1. To contrast with the paired InfoNCE, we refer to the original InfoNCE loss as unpaired InfoNCE hereafter. The behavior difference between unpaired and paired InfoNCE losses is compared in Figure 3. The paired InfoNCE only maximizes the distance between anchors and -anchors in the embedding space, neglecting irrelevant samples. The implementation of the paired InfoNCE loss is introduced in detail in Appendix A.2.", + "bbox": [ + 496, + 626, + 893, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Pretext Task", + "text_level": 1, + "bbox": [ + 500, + 816, + 633, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the aim of this work is to enable networks to strip targets (anchors) from their surroundings ( $\\neg$ anchors), we use a restricted instance discrimination task [57]: a query and a key are positive pairs if they are different views (by", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5688", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/94bcab4c34e62920d5c29550c8ba28bbd49c686ee525a78e3d55635c96dac62c.jpg", + "image_caption": [ + "Figure 2: SemCL dataset. The semantic label of a sample is utilized to generate binary masks. Each binary mask depicts the pixel-level information of one object (anchor) in the raw image. Those binary masks, together with their inverts, are applied on the raw image to extract objects (anchors) and corresponding surroundings ( $\\neg$ anchors)." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 890, + 335 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f00a92c7ece71a2863a0c55f5a16a3c78bc0764ebe9e51e37957ba554ae0cf66.jpg", + "image_caption": [ + "Figure 3: Comparison between unpaired (left) and paired InfoNCE losses (right). Left: a sample (anchor) is positive only to its own augmentations, while negative to others in the same mini-batch. Right: A sample (anchor) is positive only to its own augmentations, while negative only to its paired counterparts." + ], + "image_footnote": [], + "bbox": [ + 76, + 405, + 519, + 676 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/562f206f886369d73ce424b9fecb105a9fbdc964e9552b00571342f39e1ab72a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 405, + 893, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "data augmentation) of the same anchor, and are negative pairs if they originate from an anchor and its corresponding surroundings ( $\\neg$ anchors). A PyTorch-style pseudocode is given in the Appendix A.2. The inputs are encoded by base encoder $f_{b}$ and momentum encoder $f_{m}$ , producing queries and keys. Theoretically, the backbone of $f_{b}$ and $f_{m}$ can be any architecture. In practice, we use ResNet [23] and Swin Transformer [37] because they are the most representative architectures in recent years. All keys are stacked along", + "bbox": [ + 75, + 758, + 472, + 896 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dimension 1 for paired InfoNCE loss.", + "bbox": [ + 500, + 758, + 750, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 796, + 633, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering the enormous scale difference between ImageNet and SemCL dataset, all SemCL models are initialized by corresponding IN pretrained weights, which are compared as counterparts in downstream tasks. The total number of pretraining iterations is set to $30k$ to minimize the significant", + "bbox": [ + 496, + 825, + 895, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5689", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "variation in training time caused by scale differences among SemCL sub-datasets. We adopt AdamW [40] with $wd = 0.1$ , and the OneCycle [46] learning rate schedule consisting of warmup [17] for the initial 3, 750 (warmup ratio 0.125) steps and cosine annealing [9, 39]. Following [9], the base $lr$ is set to $1.5 \\times 10^{-4}$ and scaled linearly by $lr \\times batchsize / 256$ and temperature $\\tau = 0.2$ . The initial MoCo momentum is 0.99 and is gradually increased to 1 with a half cycle cosine schedule. The data augmentation strategy includes $224 \\times 224$ random crops after random rescaling the original image with an area ratio in the range [0.08, 1.0], followed by random color jitter, random grayscale, random Gaussian blur, and random horizontal flip. All batch norm (BN) layers are synchronized across GPUs (SyncBN [41]). Limited by the two dual-GPU nodes (RTX3090 $\\times$ 2+Titan RTX $\\times$ 2) we use, a native batch size of 64 can be applied to all ResNets and up to Swin-B for Swin Transformers with crop size $224 \\times 224$ . For Swin-L in particular, we compromise by using a batch size of 32 with gradient accumulation over 2 iterations to achieve an equivalent batch size of 64.", + "bbox": [ + 76, + 90, + 472, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SemCL aims to improve the understanding of spatial information of models for which downstream tasks of semantic segmentation, object detection and depth estimation are conducted. Unless otherwise stated, all SemCL backbones are pretrained on the corresponding SemCL sub-datasets. The implementation details of downstream tasks are listed in Appendix A.3.", + "bbox": [ + 75, + 393, + 472, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Semantic Segmentation", + "text_level": 1, + "bbox": [ + 76, + 508, + 294, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Backbone comparison. In the semantic segmentation benchmark on validation sets of VOC2012, Cityscapes, ADE20K and COCO 2017, SemCL models pretrained on the corresponding SemCL sub dataset are compared with ImageNet pretrained counterparts in Table 2. For VOC2012 and COCO, SemCL outperforms its IN counterparts by a maximum of 0.83 and 0.66 points respectively. SemCL can comprehensively outperform its IN counterparts on all backbones in the Cityscapes and ADE20k benchmarks.", + "bbox": [ + 75, + 532, + 472, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "System-level comparison. SemCL pretrained backbones are compared with other methods on semantic segmentation tasks in Table 3. On VOC2012, SemCL R50/R101 outperform Deeplabv3+ [7] by 2.37/3.12 points. The Deeplabv3+ [7] X-71 is 0.55 points behind SemCL R101 on Cityscapes. For $\\mathrm{CP}^2$ [51] ViT-S/16, SemCL Swin-S significantly outperforms it on all benchmarks by 4.97, 4.53 and 6.71 points. Compared to MAE [21] ViT-B, SemCL Swin-B achieves a gain of 0.58 points. A qualitative comparison between MAE [21] ViT-B and SemCL Swin-B is shown in Figure 4.", + "bbox": [ + 75, + 670, + 472, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Object Detection and Instance Segmentation", + "text_level": 1, + "bbox": [ + 76, + 830, + 455, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Backbone comparison. In the benchmark for object detection and instance segmentation on the test set of VOC07 and validation sets of Cityscapes and COCO 2017, SemCL", + "bbox": [ + 76, + 854, + 472, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4464714652d0d386d11cba4348eea09574b3c6db395b35b868060c353a60eee7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 101, + 733, + 189 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3f932164755daac84f3361f1fbec4912e01ed15afe9034f869ae861a85472cbb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 101, + 888, + 188 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1eb08784a2efb5c367b9822502a620995202afc14e03ad7395adfa34a6525eb2.jpg", + "image_caption": [ + "Figure 4: Qualitative comparison of semantic segmentation between MAE [21] ViT-B (top) and SemCL Swin-B (bottom) on ADE20K semantic segmentation validation set. Both decoders are UPerNet[58]." + ], + "image_footnote": [], + "bbox": [ + 503, + 191, + 733, + 281 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f2f480a4cf239efbceb69cbd05d3d16e03ca52ff8605d6ace2cba415ac391135.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 191, + 888, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "models pretrained on the corresponding SemCL subset are compared with IN pretrained counterparts in Table 4. SemCL comprehensively outperforms its IN counterparts on the Cityscapes and COCO benchmarks. In particular, on the instance segmentation task, SemCL improves by a maximum of 0.5 and 0.3 points on Cityscapes and COCO respectively. However, SemCL performs worse than its IN counterparts in VOC detection, which is considered to be a result of insufficient training samples in the SemCL-VOC dataset and is further investigated in the ablation study (see Appendix A.4).", + "bbox": [ + 496, + 378, + 893, + 545 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/881e02686d5b8419b9404f054eedae8382a4634bf805d80c8a92204695daeeaa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVOC2007 \nAP50Cityscapes†COCO
APboxAPmaskAPboxAPmask
MoCo v3[9] IN-1k81.8842.737.542.637.9
SemCL81.57(-0.31)42.9(+0.2)38.0(+0.5)42.8(+0.2)38.0(+0.1)
", + "bbox": [ + 504, + 556, + 895, + 614 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b66851922523b6f50ca535f239e5dc7ddc52139143ad1dda8bb753c92e07334c.jpg", + "table_caption": [ + "(a) R50" + ], + "table_footnote": [], + "table_body": "
MethodVOC2007 \nAP50Cityscapes†COCO
APboxAPmaskAPboxAPmask
Swin[37] IN-22k86.5845.939.849.243.0
SemCL86.53(-0.05)46.0(+0.1)40.2(+0.4)49.0(-0.2)43.3(+0.3)
", + "bbox": [ + 504, + 643, + 895, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(b) Swin-T", + "bbox": [ + 669, + 705, + 730, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 4: Results of object detection and instance segmentation on various backbones w/ Cascade Mask R-CNN. All results are from our implementation. $\\dagger$ : Sync BN is disabled due to gradient accumulation.", + "bbox": [ + 496, + 729, + 893, + 791 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "System-level comparison. Table 5 compares the object detection and instance segmentation results of SemCL backbones with other models. On Cityscapes and COCO, SemCL backbones comprehensively outperform their twinborn MoCo Specifically, SemCL ResNet50 backbones achieve $+1.7$ over MoCo IG-1B [22] in COCO detection and", + "bbox": [ + 496, + 809, + 895, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5690", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/082d3671feab936b337929b5aef38900265e6cd95b0e1b788214a2974c3c3f21.jpg", + "table_caption": [], + "table_footnote": [ + "$\\dagger$ Although MoCo v3 [9] is originally based on ViT, R50 pretrained models are provided in their repo https://github. com/facebookresearch/moco-v3." + ], + "table_body": "
BackboneMethodVOC2012CityscapesADE20KCOCO
R50MoCo v3[9] IN-1k†79.6579.0142.7140.59
SemCL79.57(-0.08)79.13(+0.12)42.96(+0.25)40.80(+0.21)
R101timm[55] IN-1k81.8279.8744.4341.43
SemCL81.97(+0.15)80.10(+0.23)44.66(+0.23)41.58(+0.15)
Swin-TSwin[37] IN-22k80.9078.6744.6042.39
SemCL81.73(+0.83)78.91(+0.26)44.72(+0.12)42.53(+0.14)
Swin-SSwin[37] IN-22k83.7181.2547.6244.75
SemCL84.47(+0.76)81.53(+0.28)47.91(+0.29)44.57(-0.18)
Swin-BSwin[37] IN-22k84.6481.6148.2144.81
SemCL84.67(+0.03)81.94(+0.33)48.68(+0.47)45.47(+0.66)
", + "bbox": [ + 148, + 88, + 823, + 296 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a88f6f47b2805d265dad4e1af595c06f4e49492eed648c230cdfb35669b82287.jpg", + "table_caption": [ + "Table 2: SemCL vs. ImageNet pretrained backbones on VOC2012, Cityscapes, ADE20K and COCO semantic segmentation (mIoU). All results are from our implementation." + ], + "table_footnote": [ + "(a) VOC2012" + ], + "table_body": "
MethodBackbonemIoU
Deeplabv3+[7]R5077.2
Supervised-IN[22]R5074.4
MoCo IG-1B[22]R5075.5
BYOL[18]R5076.3
SemCLR5079.57
Deeplabv3+[7]R10178.85
Res2Net[15]Res2Net-10180.2
DFN[61]R10180.60
SemCLR10181.97
CP2[51]ViT-S/1679.5
SemCLSwin-S84.47
Leopart[70]ViT-B/876.3
SemCLSwin-B84.67
", + "bbox": [ + 148, + 388, + 370, + 570 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7ea98b2dbfb2b38ce71183e887d7da52cb8af53865505b61f735ea0865ed291f.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Cityscapes" + ], + "table_body": "
MethodBackbonemIoU
MoCo IN-1M[22]R5072.5
MoCo IG-1B[22]R5073.6
Supervised-IN[22]R5074.6
CP2[51]R5076.5
SemCLR5079.13
Deeplabv3+[7]X-7179.55
SemCLR10180.10
Trans4Trans-T[62]PVTv2-B178.23
SemCLSwin-T78.91
CP2[51]ViT-S/1677.0
Trans4Trans-S[62]PVTv2-B280.02
SemCLSwin-S81.53
", + "bbox": [ + 375, + 388, + 607, + 570 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/109acbebfc1947ec79c70f17f2f219f2b04e15e3050ce38334797cd51d81bc36.jpg", + "table_caption": [], + "table_footnote": [ + "(c) ADE20K" + ], + "table_body": "
MethodBackbonemIoU
CP2[51]R5039.2
PSPNet[66]R5042.78
SemCLR5042.96
PSPNet[66]R10143.29
SemCLR10144.66
CP2[51]ViT-S/1641.2
Seg-S/16[47]ViT-S45.37
SemCLSwin-S47.91
Seg-B/16[47]Vi-B/1648.06
MAE[21]ViT-B48.1
SemCLSwin-B48.68
", + "bbox": [ + 611, + 388, + 823, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3: Comparison with previous best results on validation sets of VOC2012, Cityscapes, and ADE20K semantic segmentation (mIoU). Best results are bolded.", + "bbox": [ + 73, + 597, + 895, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "+5.1/+0.6 in Cityscapes/COCO instance segmentation tasks. For SemCL Swin-T backbones, SemCL achieves +1.2/+0.1 over Focal [60] on COCO detection/instance segmentation. Figure 5 demonstrates a qualitative comparison between Mask2Former [10] and SemCL Swin-T on Cityscapes validation set. With the improved spatial information understanding, SemCL model clearly discriminates (even overlapped) instances in Figures 5a and 5b and successfully classify part of an instance to a defined category as in Figure 5c.", + "bbox": [ + 73, + 654, + 470, + 791 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Depth Estimation", + "text_level": 1, + "bbox": [ + 75, + 801, + 250, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Backbone comparison. SemCL pretrained backbones are compared with corresponding IN pretrained ones in Table 6. For cityscapes, SemCL comprehensively outperforms its IN pretrained counterparts. In particular, the ResNet50 backbones improve the RMSE by 0.138 points. For KITTI", + "bbox": [ + 75, + 824, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and NTUv2, all SemCL backbones show improvements over the IN pretrained ones. The SemCL pretrained SwinL model achieves improvements of 0.001/0.026 points on KITTI AbsRel/RMSE, and the SemCL pretrained ResNet50 model achieves improvements of 0.003/0.005 points on NYUv2.", + "bbox": [ + 496, + 654, + 893, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "System-level comparison. We report the comparison of SemCL pretrained backbones with the previous state-of-the-art method Binsformer[35] and other works in Table 7. As for the Cityscapes results shown in Table 7a, SemCL pretrained ResNet50 significantly outperforms SDC [53] by $-0.089$ and $-2.218$ points on AbsRel and RMSE respectively. And SD-SSMDE [43] lags behind SemCL by $0.005 / 3.742$ points. Meanwhile, the test on KITTI (Eigen split [13]) is shown in Table 7b, SemCL pretrained ResNet50 and Swin-T comprehensively outperform the previous SOTA", + "bbox": [ + 496, + 750, + 895, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5691", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f0aea2d67442d3453f0fbb5112f6eacee395552c46628977e1c37de8d71aa79e.jpg", + "table_caption": [], + "table_footnote": [ + "(a) Object det. on VOC2007 test." + ], + "table_body": "
MethodBackboneAP50
Supervised-IN[22]R5081.3
MoCo IN-1M[22]R5081.5
MoCo IG-1B[22]R5082.2
MoCo v2 800ep[8]R5082.5
SemCLR5081.57
UP-DETR/300[12]*80.1
OW-DETR[19]DDETR82.1
SemCLSwin-T86.53
", + "bbox": [ + 106, + 99, + 331, + 224 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/205c13fa2f3d33f22861b55e37ffc24e67b41f95f70bbaef87e043681cae870d.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Instance seg. on Cityscapes val." + ], + "table_body": "
MethodBackboneAPmask
MoCo IN-1M[22]R5032.3
MoCo IG-1B[22]R5032.9
Supervised-IN[22]R5032.9
Mask2Former[10]R5037.4
SemCLR5038.0
BoundaryFormer[30]*38.3
Mask2Former[10]Swin-T39.7
SemCLSwin-T40.2
", + "bbox": [ + 334, + 99, + 578, + 224 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b0115554c9ab6aba34934dba7a61a1063d74f807831ed54f6be7bf3d84103743.jpg", + "table_caption": [], + "table_footnote": [ + "(c) Object det. and instance seg. on COCO val2017." + ], + "table_body": "
MethodBackboneAPboxAPmask
Supervised-IN[22]R5040.636.8
MoCo IN-1M[22]R5040.836.9
MoCo IG-1B[22]R5041.137.4
SemCLR5042.838.0
MST[34]Swin-T42.738.8
VSA[63]Swin-T46.942.1
Focal[60]Focal-Base47.843.2
SemCLSwin-T49.043.3
", + "bbox": [ + 583, + 99, + 864, + 224 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8bf50fb1d870c899a12d71afc831c2a9dc3a804607e544327d9a9c0fa27a4ca5.jpg", + "table_caption": [ + "Table 5: System-level comparison on object detection and instance segmentation. The best results are **bolded**. * claims their backbones are CNN-based, but are considered and compared as transformer-based networks, since transformer blocks are massively used." + ], + "table_footnote": [], + "table_body": "
BackboneMethodCityscapesKITTINYUv2
AbsRelRMSEAbsRelRMSEAbsRelRMSE
R50MoCo v3[9] IN-1k0.1384.8370.0592.3370.1200.405
SemCL0.1384.699(-0.138)0.058(-0.001)2.3370.117(-0.003)0.400(-0.005)
Swin-TSwin[37] IN-22k0.1344.6430.0572.2470.1140.386
SemCL0.133(-0.001)4.577(-0.066)0.0572.237(-0.010)0.112(-0.002)0.381(-0.005)
Swin-LSwin[37] IN-22k0.1304.5360.0532.1300.0950.329
SemCL0.128(-0.002)4.426(-0.110)0.052(-0.001)2.104(-0.026)0.094(-0.001)0.329
", + "bbox": [ + 78, + 311, + 895, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6: SemCL vs. ImageNet pretraining on Cityscapes, KITTI and NYUv2 depth estimation (lower is better). All results are from our implementation.", + "bbox": [ + 75, + 453, + 892, + 484 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4f828fa5596d72bf9b809e81a5743ab171f28b17213a00bec4ae5cf72cb1236d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 508, + 467, + 582 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9eebcdaf5d930380d32a1ce1228b380344b5f5c2f25b0450afc3010ebbb72469.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 583, + 467, + 654 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d14dc188fc07519c4c2d193ac715f644034b4cf0ec9484d690735ac98a5ff5cd.jpg", + "image_caption": [ + "Figure 5: Qualitative comparison of instance segmentation between Mask2Former [10] (left) and SemCL (right) Swin-T on Cityscapes validation set." + ], + "image_footnote": [], + "bbox": [ + 80, + 655, + 467, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Binsformer. In particular, SemCL ResNet50 achieves a gain of $-0.003 / - 0.089$ over Binsformer, and SemCL Swin-L is on par with Binsformer Swin-L. On NYUv2 (see Table 7c), the gains of SemCL pretrained ResNet50 are also high with $-0.010$ AbsRel and $-0.173$ RMSE. The SemCL", + "bbox": [ + 75, + 824, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb131c095b7ec1c7b169c6f45ea819e017f9722d5d9360e6ab2443f96f6cf025.jpg", + "image_caption": [ + "Figure 6: Qualitative comparison of depth estimation between Binsformer[35] and SemCL Swin-T on the NYUv2 test set." + ], + "image_footnote": [], + "bbox": [ + 500, + 521, + 890, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Swin-L RMSE is 0.001 points lower than the SOTA Binsformer. Qualitative comparison between Binsformer [35] and SemCL Swin-T is given in Figure 6.", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5692", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4a2a00e6430d21ed3b2c4f48b4c6933169b22d91c026c758bf0265e2e6acd86b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneAbsRelRMSE
UpProj[29]R500.257[53]7.273[53]
Pad-net[59]R500.246[53]7.117[53]
TRL[64]R500.234[53]7.104[53]
Pilzer, et al.[45]-0.4405.443
SDC-Depth[53]R500.2276.917
DTS-Depth[27]MobileNetv20.1677.785
struct2depth[5]-0.1517.024
SD-SSMDE[43]R500.1438.441
SemCLR500.1384.699
", + "bbox": [ + 99, + 99, + 397, + 224 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8351eebc395c62ca76c3574320eaf60e2e6803d09c41d05afa5ff76716febdd8.jpg", + "table_caption": [ + "(a) Cityscapes" + ], + "table_footnote": [], + "table_body": "
MethodBackboneAbsRelRMSE
EPCDepth [42]R500.0914.207
Binsformer[35]R500.0612.426
SemCLR500.0582.337
MonoViT[65]-0.0984.333
Binsformer [35]Swin-T0.0582.286
SemCLSwin-T0.0572.237
DepthFormer [33]Swin-L +R50CI0.0522.143
Binsformer[35]Swin-L0.0522.098
SemCLSwin-L0.0522.104
", + "bbox": [ + 400, + 99, + 629, + 224 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/71e30f17f65e81b2c4e177c6ae6216589702ba758817f108caa56853fec9b7b7.jpg", + "table_caption": [ + "(b) KITTI" + ], + "table_footnote": [], + "table_body": "
MethodBackboneAbsRelRMSE
Pad-net [59]R500.2140.792
TRL [64]R500.1440.501
SDC-Depth [53]R500.1280.497
UpProj [29]R500.1270.573
SemCLR500.1170.400
Binsformer [35]Swin-T0.1130.379
SemCLSwin-T0.1120.381
Binsformer [35]Swin-L0.0940.330
SemCLSwin-L0.0940.329
", + "bbox": [ + 632, + 99, + 872, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(c) NYUv2", + "bbox": [ + 720, + 227, + 781, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7: System-level comparison on depth estimation. Best results (lower is better) are bolded.", + "bbox": [ + 166, + 252, + 797, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Attention Map Visualization", + "text_level": 1, + "bbox": [ + 76, + 294, + 330, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To improve the qualitative understanding of the SemCL representation, we visualize attention maps of Swin-T pretrained on SemCL-VOC and ImageNet using PyTorch library for CAM methods [16] with the Score-CAM method [52]. These samples in Figure 7 selected from VOC2012 val and have not been seen by the pretrained models are characterized by simple centered subjects (Figure 7a), complex scene with occlusion (Figure 7b) and multiple subjects (Figures 7c and 7d).", + "bbox": [ + 75, + 316, + 470, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figure 7a, the IN pretrained model failed to focus on the subject, while the SemCL model accurately depicted the area of the main subject. Meanwhile, SemCL performs better in complex scenes containing foreground-background relationships, as shown in Figure 7b. The IN pretrained model pays no attention to an occluded object, but the SemCL model recognizes that the occluded part also belongs to the horse, as shown in Figure 7b. In Figure 7c, the IN model directly ignores the distant person, but the SemCL model not only annotates two people but also focuses more on the nearby one, suggesting that the SemCL model can understand the spatial distribution of subjects in an image. Also, in Figure 7d, the SemCL model successfully recognizes one plant and its pot as one object.", + "bbox": [ + 75, + 453, + 468, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Summary. SemCL pretrained backbones can substantially outperform ImageNet pretrained counterparts on semantic segmentation, object detection, instance segmentation and depth estimation tasks. Using semantically contrastive pairs generated from off-the-shelf datasets, SemCL significantly improves the spatial information understanding of pretrained models. Such an improvement is also promising to benefit more challenging tasks including generative ones [32]. Ablation study concerning paired/unpaired InfoNCE loss, pretraining batch size, training length and dataset scale are given in Appendix A.4.", + "bbox": [ + 75, + 665, + 470, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 844, + 194, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We investigate the effectiveness of semantic information in the contrastive learning pretext task on the spatial un", + "bbox": [ + 76, + 869, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5226e012f04effbc40caf2ee71320be0dd930e0f6780ee6e035ff270d3c00d18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 292, + 638, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bc20e95d32622b1d719688fb7165e8bcba5717039984cbeea7d857567e6845de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 383, + 638, + 474 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dde65f9160e7393015f2b97881d5058707c118ec74bc9b84b93e11c0565da30e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 474, + 638, + 564 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7919bb40da1c18c0f38cacca0e4346a87da9f4a3fa4da570a4ea3730d441e554.jpg", + "image_caption": [ + "Figure 7: Swin-T attention map visualization. Left: raw image. Middle: attention maps of the IN pretrained model. Right: attention maps of the SemCL pretrained model." + ], + "image_footnote": [], + "bbox": [ + 501, + 565, + 638, + 655 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1fc2ba11d96a5b524ff739c0e9eb982a41fd4562c0c91572ad37b10b6e55bb2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 292, + 764, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/693d5bde387383fc97b44205cca0feb880c9879a5708401aa62593fb1774048e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 383, + 764, + 473 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/648068eebffd0c743bfea2d1146cabbd5569dc9fa89dc62364f507530818824e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 473, + 764, + 564 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/231350e5e0a7098c25bc15d484f0c3ff1aca3dd2c956250a23babc7618d29816.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 564, + 764, + 655 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/537e899522c70cbcab38245ca5a0863705d05756a4dcac98dba417070839438f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 292, + 890, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a6caed9c7280655bf1a27b87c9a3c67be42a401e0db67034708997b7285fbc54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 383, + 890, + 474 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/689106e8ce1df92ee8ea95243925d7e07effab326b0a18106a216c5342d7e681.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 474, + 890, + 564 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4f2b7f5cdecb8ffc041152283ccb5171172d6cb5b0e25a425fada50073325794.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 564, + 890, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "derstanding ability of models. Tests are conducted on semantic/instance segmentation, object detection and depth estimation tasks with both ResNets and Swin Transformers, and SemCL pretrained backbones substantially outperform their ImageNet pretrained counterparts. By supporting small batch sizes and fast pretraining, SemCL is a lightweight yet effective approach. In contrastive representation learning, if dedicated pretext task is designed properly, four ounces can move a thousand pounds. We hope that SemCL will inspire more advanced pretext tasks for contrastive learning.", + "bbox": [ + 496, + 739, + 893, + 890 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5693", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019. 2", + "[2] Suzanne Becker and Geoffrey E Hinton. Self-organizing neural network that discovers surfaces in random-dot stereograms. Nature, 355(6356):161-163, 1992. 2", + "[3] Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah. Signature verification using a “siamese” time delay neural network. Advances in neural information processing systems, 6, 1993. 2", + "[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Computer vision and pattern recognition (CVPR), 2018 IEEE conference on. IEEE, 2018. 3", + "[5] Vincent Casser, Soeren Pirk, Reza Mahjourian, and Anelia Angelova. Unsupervised monocular depth and ego-motion learning with structure and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 8", + "[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 3", + "[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 2, 5, 6", + "[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 7", + "[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 1, 2, 5, 6, 7", + "[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 2, 6, 7", + "[11] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3", + "[12] Zhigang Dai, Bolun Cai, Yugeng Lin, and Junying Chen. Updetr: Unsupervised pre-training for object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1601-1610, 2021. 7", + "[13] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 6" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html. 3", + "[15] Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multiscale backbone architecture. IEEE transactions on pattern analysis and machine intelligence, 43(2):652-662, 2019. 6", + "[16] Jacob Gildenblat and contributors. Pytorch library for cam methods. https://github.com/jacobgil/pytorch-grad-cam, 2021.8", + "[17] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: TrainingImagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 5", + "[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems, 33:21271-21284, 2020. 6", + "[19] Akshita Gupta, Sanath Narayan, KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Ow-detr: Open-world detection transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9235-9244, 2022. 7", + "[20] Bharath Hariharan, Pablo Arbelaez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In International Conference on Computer Vision (ICCV), 2011. 3", + "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 5, 6", + "[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 2, 3, 5, 6, 7", + "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4", + "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pages 630-645. Springer, 2016. 2", + "[25] Olivier Henaff. Data-efficient image recognition with contrastive predictive coding. In International Conference on Machine Learning, pages 4182-4192. PMLR, 2020. 2", + "[26] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In International Conference on Learning Representations, 2019. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5694", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Hatem Ibrahim, Ahmed Salem, and Hyun-Soo Kang. Dts-depth: Real-time single-image depth estimation using depth-to-space image construction. Sensors, 22(5):1914, 2022. 8", + "[28] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 2, 3", + "[29] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In 2016 Fourth international conference on 3D vision (3DV), pages 239-248. IEEE, 2016. 8", + "[30] Justin Lazarow, Weijian Xu, and Zhuowen Tu. Instance segmentation with mask-supervised polygonal boundary transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4382-4391, 2022. 7", + "[31] Phuc H. Le-Khac, Graham Healy, and Alan F. Smeaton. Contrastive representation learning: A framework and review. IEEE Access, 8:193907-193934, 2020. 1, 2", + "[32] Yuheng Li, Yijun Li, Jingwan Lu, Eli Shechtman, Yong Jae Lee, and Krishna Kumar Singh. Contrastive learning for diverse disentangled foreground generation. In European Conference on Computer Vision, pages 334-351. Springer, 2022. 8", + "[33] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. arXiv preprint arXiv:2203.14211, 2022. 8", + "[34] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, et al. Mst: Masked self-supervised transformer for visual representation. Advances in Neural Information Processing Systems, 34:13165-13176, 2021. 7", + "[35] Zhenyu Li, Xuyang Wang, Xianming Liu, and Junjun Jiang. Binsformer: Revisiting adaptive bins for monocular depth estimation. arXiv preprint arXiv:2204.00987, 2022. 2, 6, 7, 8", + "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 3", + "[37] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 4, 5, 6, 7", + "[38] Vishnu Suresh Lokhande, Songwong Tasneeyapant, Abhay Venkatesh, Sathya N Ravi, and Vikas Singh. Generating accurate pseudo-labels in semi-supervised learning and avoiding overconfident predictions via hermite polynomial activations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11435-11443, 2020. 2", + "[39] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, April" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017.5", + "[40] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5", + "[41] Chao Peng, Tete Xiao, Zeming Li, Yuning Jiang, Xiangyu Zhang, Kai Jia, Gang Yu, and Jian Sun. Megdet: A large minibatch object detector. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6181-6189, 2018. 5", + "[42] Rui Peng, Ronggang Wang, Yawen Lai, Luyang Tang, and Yangang Cai. Excavating the potential capacity of self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15560-15569, 2021. 8", + "[43] Andra Petrovai and Sergiu Nedevschi. Exploiting pseudo labels in a self-supervised learning framework for improved monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1578-1588, 2022. 6, 8", + "[44] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V Le. Meta pseudo labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11557-11568, 2021. 2", + "[45] Andrea Pilzer, Dan Xu, Mihai Puscas, Elisa Ricci, and Nicu Sebe. Unsupervised adversarial depth estimation using cycled generative networks. In 2018 international conference on 3D vision (3DV), pages 587-595. IEEE, 2018. 8", + "[46] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, volume 11006, pages 369-386. SPIE, 2019. 5", + "[47] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7262-7272, 2021. 6", + "[48] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In European conference on computer vision, pages 776-794. Springer, 2020. 2", + "[49] Aaron Van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv e-prints, pages arXiv-1807, 2018. 1, 2, 3", + "[50] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021. 3", + "[51] Feng Wang, Huiyu Wang, Chen Wei, Alan Yuille, and Wei Shen. Cp 2: Copy-paste contrastive pretraining for semantic segmentation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXX, pages 499-515. Springer, 2022. 5, 6", + "[52] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 8" + ], + "bbox": [ + 503, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5695", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Lijun Wang, Jianming Zhang, Oliver Wang, Zhe Lin, and Huchuan Lu. Sdc-depth: Semantic divide-and-conquer network for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 541-550, 2020. 6, 8", + "[54] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7303-7313, 2021. 3", + "[55] Ross Wightman, Hugo Touvron, and Hervé Jégou. Resnet strikes back: An improved training procedure in timm. arXiv preprint arXiv:2110.00476, 2021. 6", + "[56] Hao Wu and Saurabh Prasad. Semi-supervised deep learning using pseudo labels for hyperspectral image classification. IEEE Transactions on Image Processing, 27(3):1259-1270, 2017. 2", + "[57] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 1, 2, 3", + "[58] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 5", + "[59] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Padnet: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 675-684, 2018. 8", + "[60] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal attention for long-range interactions in vision transformers. Advances in Neural Information Processing Systems, 34:30008-30022, 2021. 6, 7", + "[61] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Learning a discriminative feature network for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1857-1866, 2018. 6", + "[62] Jiaming Zhang, Kailun Yang, Angela Constantinescu, Kunyu Peng, Karin Müller, and Rainer Stiefelhagen. Trans4trans: Efficient transformer for transparent object and semantic scene segmentation in real-world navigation assistance. IEEE Transactions on Intelligent Transportation Systems, 23(10):19173-19186, 2022. 6", + "[63] Qiming Zhang, Yufei Xu, Jing Zhang, and Dacheng Tao. Vsa: learning varied-size window attention in vision transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXV, pages 466-483. Springer, 2022. 7", + "[64] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Zequn Jie, Xiang Li, and Jian Yang. Joint task-recursive learning for semantic segmentation and depth estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 235–251, 2018. 8", + "[65] Chaoqiang Zhao, Youmin Zhang, Matteo Poggi, Fabio Tosi, Xianda Guo, Zheng Zhu, Guan Huang, Yang Tang, and" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Stefano Mattoccia. Monovit: Self-supervised monocular depth estimation with a vision transformer. arXiv preprint arXiv:2208.03543, 2022. 8", + "[66] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 6", + "[67] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 3", + "[68] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 3", + "[69] Chengxu Zhuang, Alex Lin Zhai, and Daniel Yamins. Local aggregation for unsupervised learning of visual embeddings. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6002-6012, 2019. 2", + "[70] Adrian Ziegler and Yuki M Asano. Self-supervised learning of object parts for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14502-14511, 2022. 6" + ], + "bbox": [ + 503, + 92, + 890, + 416 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5696", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Semantic Information in Contrastive Learning/c3144656-62d8-4975-97d6-ef4630f410df_model.json b/2023/Semantic Information in Contrastive Learning/c3144656-62d8-4975-97d6-ef4630f410df_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7eb4c31dfcdbd44bf1f445d749458baeb1b0c81d --- /dev/null +++ b/2023/Semantic Information in Contrastive Learning/c3144656-62d8-4975-97d6-ef4630f410df_model.json @@ -0,0 +1,2488 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.131, + 0.723, + 0.154 + ], + "angle": 0, + "content": "Semantic Information in Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.18, + 0.793, + 0.272 + ], + "angle": 0, + "content": "Shengjiang Quan\\(^{1*}\\), Masahiro Hirano\\(^{2}\\) and Yuji Yamakawa\\(^{3}\\)\\\n\\(^{1}\\)Graduate School of Engineering, The University of Tokyo, Japan\\\n\\(^{2}\\)Institute of Industrial Science, The University of Tokyo, Japan\\\n\\(^{3}\\)Interfaculty Initiative in Information Studies, The University of Tokyo, Japan\\\nquan@g.ecc.u-tokyo.ac.jp, {mhirano, y-ymkw}@iis.u-tokyo.ac.jp" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.474, + 0.58 + ], + "angle": 0, + "content": "This work investigates the functionality of Semantic information in Contrastive Learning (SemCL). An advanced pretext task is designed: a contrast is performed between each object and its environment, taken from a scene. This allows the SemCL pretrained model to extract objects from their environment in an image, significantly improving the spatial understanding of the pretrained models. Downstream tasks of semantic/instance segmentation, object detection and depth estimation are implemented on PASCAL VOC, Cityscapes, COCO, KITTI, etc. SemCL pretrained models substantially outperform ImageNet pretrained counterparts and are competitive with well-known works on downstream tasks. The results suggest that a dedicated pretext task leveraging semantic information can be powerful in benchmarks related to spatial understanding. The code is available at https://github.com/sjiang95/semcl." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.608, + 0.21, + 0.625 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.634, + 0.471, + 0.769 + ], + "angle": 0, + "content": "Within the field of visual representation learning, contrastive learning attracts special attention for its inspiring performance in transfer learning [57, 49, 31, 22, 8, 9]. The concept of contrastive learning can literally be explained as discovering the difference between positive and negative samples. The definition of positive-negative samples is one of the main subjects of contrastive learning, which directly determines the pretext task and the corresponding loss function." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.771, + 0.472, + 0.878 + ], + "angle": 0, + "content": "We humans can recognize an object from its even complicated environment because we have learned and bound the typical geometric feature and the corresponding semantic concept of the object class. For example, we can recognize a cat in a scene thanks to our knowledge of its appearance (typical geometric feature) and the concept of cat (semantic information). Similarly, the target of prevalent" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.308, + 0.895, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.465, + 0.896, + 0.556 + ], + "angle": 0, + "content": "Figure 1: Perceptual cognition of SemCL. Leveraging semantic information, SemCL separates an object from its environment. Each sample is divided into objects and surroundings and considered as a contrastive pair. The task is to maximize the distance between the object and its environment in the embedding space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.854 + ], + "angle": 0, + "content": "visual tasks, such as semantic/instance segmentation, object detection, etc., can be summarized as distinguishing and localizing subjects from surroundings. Inspired by and to extend the spatial recognition mechanism to CV, we propose our method SemCL which directly teaches models to extract a subject from its environment. Utilizing publicly off-the-shelf datasets providing semantic labels, SemCL can be considered as a supervised contrastive learning approach whose samples consist of subjects and their corresponding surroundings (semantically not the subject). The pretext task of SemCL is to tell the difference between one subject and its surroundings. From a representation learning perspective, the pretext task is to maximize the distance \\((L^2\\) norm) between a subject and its environment in the embedding space (see Figure 1). The pretext task mimics the spatial recognition mechanism to discriminate objects and surroundings, which significantly improves the spatial information understanding of pretrained models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In SemCL, contrast is performed pairwise between positive-negative pairs from one image/scene. Inter-scene contrast is considered inappropriate: only the contrasts" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5686" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "among subject-surrounding pairs from one scene are considered valid. Therefore, paired InfoNCE loss is adopted. With the mechanism of contrasting only paired samples, SemCL is characterized by decoupling the number of negative samples from the batch size. Compared to MoCo [22], whose performance is positively correlated with batch size and thus increases hardware cost (e.g. batchsize \\(= 4096\\) on 128 GPUs for ViT-B in MoCo v3 [9]) for satisfactory results, SemCL can achieve consistent yet competitive results on downstream tasks with relatively small batch size (e.g., batchsize \\(= 64 \\ @ 224 \\times 224\\)). We adopt the MoCo v3 [9] as the pretraining framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.473, + 0.456 + ], + "angle": 0, + "content": "The primary motivation of representation learning is to pretrain general representations that can be fine-tuned and transferred to downstream tasks. In this work, the SemCL pretrained models are benchmarked on semantic segmentation, object detection and instance segmentation, and depth estimation tasks. SemCL models substantially outperform their ImageNet pretrained counterparts due to the improved ability of spacial information understanding. In a system-level comparison, SemCL models also show gains over previous well-known works (e.g. Deeplabv3+ [7] in semantic segmentation, Mask2Former [10] in instance segmentation and Binsformer [35] in depth estimation)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.472, + 0.22, + 0.487 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.498, + 0.279, + 0.515 + ], + "angle": 0, + "content": "2.1. Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.473, + 0.674 + ], + "angle": 0, + "content": "The seminal idea of contrastive learning dates back at least to the 1990s [2, 3, 31], but it has become prominent in recent years thanks to the large pretrained models in the fields of NLP and CV [31]. As the name suggests, contrastive learning mines abstract representations by comparing between \"similar\" and \"dissimilar\" samples: minimizing the representation of \"similar\" samples and maximizing that of \"dissimilar\" samples in the embedding space. Formally, the \"similar\" inputs are called positive samples, and the \"dissimilar\" inputs are called negative samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Defining \"similar\" and \"dissimilar\" samples is an important topic in contrastive learning. In instance discrimination [57], which is a ubiquitous approach to define positive-negative samples in CV, each sample in a minibatch \\( x \\in \\{x_1, \\ldots, x_n\\} \\) is considered mutually exclusive to the rest. The encoded output \\( \\theta(\\cdot) \\) of each sample is pushed away from the others in the embedding space, denoted as \\( \\theta(x_1) \\Leftrightarrow \\{\\theta(x_2), \\ldots, \\theta(x_n)\\} \\Leftrightarrow \\text{push} \\). Contrastive Predictive Coding (CPC) [49] defines positive-negative samples in a generative way: Given a sequential input \\( \\{x_{t-3}, x_{t-2}, x_{t-1}, x_t\\} \\), a context latent representation \\( C_t \\) is computed to predict the encoded outputs of future inputs \\( \\{\\theta(\\mathbb{X}_{t+1}), \\theta(\\mathbb{X}_{t+2}), \\theta(\\mathbb{X}_{t+3}), \\ldots\\} \\). Those predicted future embeddings are considered as positive samples to real future embeddings, and exclusive to some other" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.093, + 0.615, + 0.107 + ], + "angle": 0, + "content": "irrelevant inputs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.699, + 0.13 + ], + "angle": 0, + "content": "2.2. Momentum Contrast" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.139, + 0.895, + 0.336 + ], + "angle": 0, + "content": "He et al. [22] proposed a momentum contrast framework, MoCo, for contrastive representation learning. MoCo summed unsupervised visual representation learning [57, 49, 26, 69, 25, 48, 1] up as dictionary look-up: the input data are represented and sampled by an encoder network \\(\\theta_{k}\\) as \"keys\" \\(k = \\theta_{k}(x^{k})\\). The goal of contrastive learning is to train encoders \\(\\theta_{q}\\) by which a \"query\" \\(q = \\theta_{q}(x^{q})\\) should be similar to its matching key and dissimilar to others. Since dictionary consistency is considered crucial for unsupervised learning with contrastive loss [22], MoCo introduces momentum updating. In the MoCo framework, the key coder \\(\\theta_{k}\\) is progressively updated by the weighted average of \\(\\theta_{k}\\) and the query coder \\(\\theta_{q}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.345, + 0.894, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\theta_ {k} \\leftarrow m \\theta_ {k} + (1 - m) \\theta_ {q}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.894, + 0.446 + ], + "angle": 0, + "content": "where the momentum coefficient \\( m \\in [0,1) \\) is set quite high (\\( m = 0.999 \\) in [22]) for slower key encoder update. The MoCo series has proven the power of the Momentum Contrast framework on feature transferring and is therefore employed as the basis for this work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.454, + 0.866, + 0.47 + ], + "angle": 0, + "content": "2.3. Labeled Data In Unsupervised Framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.895, + 0.627 + ], + "angle": 0, + "content": "Large networks are quite thirsty for training data, while labelling the large amount of collected data is a human resource consuming challenge. Apart from turning to unsupervised learning, which inherently avoids using labelled data, using unlabeled data in supervised learning is also a solution. Pseudo-Label [44, 38, 56] is a trick that uses (usually supervised) pretrained networks to generate features that can play the role of labels. And the generated pseudo-labels participate in the iterations of the target network just like annotations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.897, + 0.902 + ], + "angle": 0, + "content": "In contrast, Khosla et al. [28] also investigated the function of labelled data in unsupervised learning. As a contrastive learning research, Khosla et al. [28] intuitively uses the classification label to define samples within the same class are mutually positive, and those from other categories are negative. To exploit the label information, clusters of points belonging to the same class are pulled together in the embedding space, while clusters of samples from different classes are pushed apart [28]: \\(\\left\\{\\theta\\left(x_1^{c0}\\right) \\rightleftharpoons \\theta\\left(x_2^{c0}\\right) \\rightleftharpoons \\theta\\left(x_{n}^{c0}\\right) \\rightleftharpoons \\theta\\left(x_n^0\\right)\\) \\(\\Leftrightarrow\\) \\(\\left\\{\\theta\\left(x_1^{c1}\\right) \\rightleftharpoons \\theta\\left(x_2^{c1}\\right) \\rightleftharpoons \\theta\\left(x_{m}^{c1}\\right)\\) (\\(\\rightleftharpoons\\) for pull). The experimental result shows that supervised contrastive learning can improve both the accuracy and robustness of classifiers - [28] achieves a top-1 accuracy of \\(81.4\\%\\) on the ImageNet (IN) dataset, which is \\(0.8\\%\\) above the best result reported by ResNet-200 architecture [24]. Likewise, our work investigates the effectiveness of semantic information in unsupervised framework contrastive learning." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5687" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.471, + 0.108 + ], + "angle": 0, + "content": "2.4. Semantic Information in Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.25 + ], + "angle": 0, + "content": "Other than [28], there are works that use small amounts of labelled data in unsupervised representation learning. Wang et al. [54] proposed a pixel-wise contrastive algorithm for semantic segmentation: for a pixel \\(i\\), it is pulled towards pixels belonging to the same category and pushed away from pixels in other categories in the embedding space. This contrast is performed across images from the same dataset. In the Cityscapes test benchmark, [54] R101 achieves 79.2 mIoU, which is 1.1 points better than Deeplabv3 [6]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.25, + 0.471, + 0.34 + ], + "angle": 0, + "content": "The attempt and results of works using labels in unsupervised representation learning, such as [28, 54], indicate that label information can bring gains to a fully unsupervised representation. It is promising to explore advanced pretext tasks to improve the performance of pretrained models in downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.353, + 0.169, + 0.369 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.377, + 0.235, + 0.393 + ], + "angle": 0, + "content": "3.1. SemCL Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.47, + 0.522 + ], + "angle": 0, + "content": "The core of this work is to define semantically contrastive pairs by utilizing off-the-shelf semantic labels, which are mappings between categories of objects and their coordinates on the raw image at the pixel level. As decided by the authors of a dataset, semantic labels can be color or grayscale images. Each predefined color or grayscale denotes a category, and the distribution of label pixels reflects the location of pixels belonging to a particular category in the raw image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.704 + ], + "angle": 0, + "content": "The SemCL dataset is generated by the mechanism shown in Figure 2, where the raw image and its semantic label are selected from the PASCAL VOC2012 dataset [14]. From the semantic label annotating two objects - a motorcycle and a rider - two binary masks BM0 and BM1 can be extracted to separately represent two annotated objects. Their inverted counterparts \\(\\neg\\)BM0 and \\(\\neg\\)BM1 represent pixels other than the target object. These binary masks are then bitwise applied to the raw image to separate an object (Anchor) from its environment (\\(\\neg\\)Anchor), forming a semantically contrastive pair. Each element in a raw Anchor-\\(\\neg\\)Anchor pair is considered to be semantically opposite to the other." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.471, + 0.825 + ], + "angle": 0, + "content": "In practice, the SemCL-1M dataset is produced from training sets of augmented PASCAL VOC2012 [14] (also known as SBD [20]), Cityscapes [11], ADE20K [67, 68] and COCO [36] (stuff+thing 2017 [4]) datasets. The components of the SemCL dataset are illustrated in Table 1. When generating the dataset, a threshold \\( t = 0.01 \\) is set to filter out too small (less than \\( [\\sqrt{t} \\times \\text{height}] \\times [\\sqrt{t} \\times \\text{width}] \\)) objects that are considered semantically unsaturated." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.279, + 0.848 + ], + "angle": 0, + "content": "3.2. Paired InfoNCE Loss" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Since contrastive learning can be considered as a dictionary look-up task [22], it is assumed that within a set of keys \\(\\{k_0, k_1, k_2, \\ldots\\}\\) of a dictionary, a query \\(q\\) matches only one" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.089, + 0.894, + 0.162 + ], + "angle": 0, + "content": "
VOC2012CityscapesADE20KCOCOTotal
# training samples10,5822,97525,574118,287157,418
# SemCL pairs14,20320,280270,218712,2121,016,913
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.171, + 0.894, + 0.233 + ], + "angle": 0, + "content": "Table 1: Components of the SemCL dataset. The SemCL datasets spawned from VOC2012(aug), Cityscapes, ADE20k and COCO 2017 are named SemCL-\\(\\{\\mathrm{VOC}\\), City, ADE, COCO\\} respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.894, + 0.336 + ], + "angle": 0, + "content": "key \\( k_{+} \\) of the dictionary [22]. The value of the objective function should be low if \\( q \\) is similar to its positive key \\( k_{+} \\) and dissimilar to all negative keys \\( \\{k_{0}, k_{1}, k_{2}, \\ldots\\} \\setminus \\{k_{+}\\} \\). In MoCo [22], the form of a contrastive loss function InfoNCE [49] is adopted" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.347, + 0.893, + 0.384 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {q} = - \\log \\frac {\\exp (q \\cdot k _ {+} / \\tau)}{\\sum_ {i = 0} ^ {K} \\exp (q \\cdot k _ {i} / \\tau)} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.895, + 0.56 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a scalar temperature hyperparameter controlling the strength of penalties on hard negative samples [50], e.g. in the instance discrimination task [49, 22], there may be potential positive samples in the randomly chosen negative samples. Summing over one positive and \\(K\\) negative samples, Equation (2) is the log loss of a \\((\\mathbf{K} + 1)\\)-way softmax-based classifier trying to classify \\(q\\) as \\(k_{+}\\) [22]. While in our case an anchor instance \\(q\\) compares only with its own \\(n\\) augmentations (positive) \\(K^{+} := \\{k_{1}^{+}, k_{2}^{+}, k_{3}^{+}, \\ldots, k_{n}^{+}\\}\\) and paired \\(n + 1\\) negatives \\(K^{-} := \\{k_{0}^{-}, k_{1}^{-}, k_{2}^{-}, k_{3}^{-}, \\ldots, k_{n}^{-}\\}\\) using paired InfoNCE" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.572, + 0.903, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {q} = - \\sum_ {j = t + 1} ^ {n} \\log \\frac {\\exp \\left(q _ {t} \\cdot k _ {j} ^ {+} / \\tau\\right)}{\\sum_ {i = 0} ^ {K ^ {-}} \\exp \\left(q _ {t} \\cdot k _ {i} ^ {-} / \\tau\\right) + \\exp \\left(q _ {t} \\cdot k _ {j} ^ {+} / \\tau\\right)}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.895, + 0.807 + ], + "angle": 0, + "content": "For one anchor as query \\( q_{0}(t = 0) \\), the paired InfoNCE loss tries to classify it as \\( k_{j}^{+}, j \\in [1, n] \\). The next step is to go forward: classify the first positive \\( q_{1}(t = 1) \\) of the anchor as \\( k_{j}^{+}, j \\in [2, n] \\), and so on, as shown in the algorithm 1, Appendix A.1. To contrast with the paired InfoNCE, we refer to the original InfoNCE loss as unpaired InfoNCE hereafter. The behavior difference between unpaired and paired InfoNCE losses is compared in Figure 3. The paired InfoNCE only maximizes the distance between anchors and -anchors in the embedding space, neglecting irrelevant samples. The implementation of the paired InfoNCE loss is introduced in detail in Appendix A.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.634, + 0.832 + ], + "angle": 0, + "content": "3.3. Pretext Task" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Since the aim of this work is to enable networks to strip targets (anchors) from their surroundings (\\(\\neg\\)anchors), we use a restricted instance discrimination task [57]: a query and a key are positive pairs if they are different views (by" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5688" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.891, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.344, + 0.895, + 0.391 + ], + "angle": 0, + "content": "Figure 2: SemCL dataset. The semantic label of a sample is utilized to generate binary masks. Each binary mask depicts the pixel-level information of one object (anchor) in the raw image. Those binary masks, together with their inverts, are applied on the raw image to extract objects (anchors) and corresponding surroundings (\\(\\neg\\)anchors)." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.406, + 0.52, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.406, + 0.895, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.687, + 0.895, + 0.734 + ], + "angle": 0, + "content": "Figure 3: Comparison between unpaired (left) and paired InfoNCE losses (right). Left: a sample (anchor) is positive only to its own augmentations, while negative to others in the same mini-batch. Right: A sample (anchor) is positive only to its own augmentations, while negative only to its paired counterparts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.76, + 0.473, + 0.897 + ], + "angle": 0, + "content": "data augmentation) of the same anchor, and are negative pairs if they originate from an anchor and its corresponding surroundings (\\(\\neg\\)anchors). A PyTorch-style pseudocode is given in the Appendix A.2. The inputs are encoded by base encoder \\(f_{b}\\) and momentum encoder \\(f_{m}\\), producing queries and keys. Theoretically, the backbone of \\(f_{b}\\) and \\(f_{m}\\) can be any architecture. In practice, we use ResNet [23] and Swin Transformer [37] because they are the most representative architectures in recent years. All keys are stacked along" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.76, + 0.751, + 0.775 + ], + "angle": 0, + "content": "dimension 1 for paired InfoNCE loss." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.797, + 0.634, + 0.814 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Considering the enormous scale difference between ImageNet and SemCL dataset, all SemCL models are initialized by corresponding IN pretrained weights, which are compared as counterparts in downstream tasks. The total number of pretraining iterations is set to \\(30k\\) to minimize the significant" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5689" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.392 + ], + "angle": 0, + "content": "variation in training time caused by scale differences among SemCL sub-datasets. We adopt AdamW [40] with \\(wd = 0.1\\), and the OneCycle [46] learning rate schedule consisting of warmup [17] for the initial 3, 750 (warmup ratio 0.125) steps and cosine annealing [9, 39]. Following [9], the base \\(lr\\) is set to \\(1.5 \\times 10^{-4}\\) and scaled linearly by \\(lr \\times batchsize / 256\\) and temperature \\(\\tau = 0.2\\). The initial MoCo momentum is 0.99 and is gradually increased to 1 with a half cycle cosine schedule. The data augmentation strategy includes \\(224 \\times 224\\) random crops after random rescaling the original image with an area ratio in the range [0.08, 1.0], followed by random color jitter, random grayscale, random Gaussian blur, and random horizontal flip. All batch norm (BN) layers are synchronized across GPUs (SyncBN [41]). Limited by the two dual-GPU nodes (RTX3090 \\(\\times\\) 2+Titan RTX \\(\\times\\) 2) we use, a native batch size of 64 can be applied to all ResNets and up to Swin-B for Swin Transformers with crop size \\(224 \\times 224\\). For Swin-L in particular, we compromise by using a batch size of 32 with gradient accumulation over 2 iterations to achieve an equivalent batch size of 64." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.473, + 0.501 + ], + "angle": 0, + "content": "SemCL aims to improve the understanding of spatial information of models for which downstream tasks of semantic segmentation, object detection and depth estimation are conducted. Unless otherwise stated, all SemCL backbones are pretrained on the corresponding SemCL sub-datasets. The implementation details of downstream tasks are listed in Appendix A.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.51, + 0.295, + 0.527 + ], + "angle": 0, + "content": "4.1. Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.473, + 0.67 + ], + "angle": 0, + "content": "Backbone comparison. In the semantic segmentation benchmark on validation sets of VOC2012, Cityscapes, ADE20K and COCO 2017, SemCL models pretrained on the corresponding SemCL sub dataset are compared with ImageNet pretrained counterparts in Table 2. For VOC2012 and COCO, SemCL outperforms its IN counterparts by a maximum of 0.83 and 0.66 points respectively. SemCL can comprehensively outperform its IN counterparts on all backbones in the Cityscapes and ADE20k benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.473, + 0.822 + ], + "angle": 0, + "content": "System-level comparison. SemCL pretrained backbones are compared with other methods on semantic segmentation tasks in Table 3. On VOC2012, SemCL R50/R101 outperform Deeplabv3+ [7] by 2.37/3.12 points. The Deeplabv3+ [7] X-71 is 0.55 points behind SemCL R101 on Cityscapes. For \\(\\mathrm{CP}^2\\) [51] ViT-S/16, SemCL Swin-S significantly outperforms it on all benchmarks by 4.97, 4.53 and 6.71 points. Compared to MAE [21] ViT-B, SemCL Swin-B achieves a gain of 0.58 points. A qualitative comparison between MAE [21] ViT-B and SemCL Swin-B is shown in Figure 4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.456, + 0.849 + ], + "angle": 0, + "content": "4.2. Object Detection and Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Backbone comparison. In the benchmark for object detection and instance segmentation on the test set of VOC07 and validation sets of Cityscapes and COCO 2017, SemCL" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.102, + 0.734, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.102, + 0.89, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.192, + 0.735, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.192, + 0.889, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.294, + 0.895, + 0.355 + ], + "angle": 0, + "content": "Figure 4: Qualitative comparison of semantic segmentation between MAE [21] ViT-B (top) and SemCL Swin-B (bottom) on ADE20K semantic segmentation validation set. Both decoders are UPerNet[58]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.895, + 0.546 + ], + "angle": 0, + "content": "models pretrained on the corresponding SemCL subset are compared with IN pretrained counterparts in Table 4. SemCL comprehensively outperforms its IN counterparts on the Cityscapes and COCO benchmarks. In particular, on the instance segmentation task, SemCL improves by a maximum of 0.5 and 0.3 points on Cityscapes and COCO respectively. However, SemCL performs worse than its IN counterparts in VOC detection, which is considered to be a result of insufficient training samples in the SemCL-VOC dataset and is further investigated in the ablation study (see Appendix A.4)." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.558, + 0.897, + 0.616 + ], + "angle": 0, + "content": "
MethodVOC2007 \nAP50Cityscapes†COCO
APboxAPmaskAPboxAPmask
MoCo v3[9] IN-1k81.8842.737.542.637.9
SemCL81.57(-0.31)42.9(+0.2)38.0(+0.5)42.8(+0.2)38.0(+0.1)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.679, + 0.619, + 0.722, + 0.631 + ], + "angle": 0, + "content": "(a) R50" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.644, + 0.897, + 0.703 + ], + "angle": 0, + "content": "
MethodVOC2007 \nAP50Cityscapes†COCO
APboxAPmaskAPboxAPmask
Swin[37] IN-22k86.5845.939.849.243.0
SemCL86.53(-0.05)46.0(+0.1)40.2(+0.4)49.0(-0.2)43.3(+0.3)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.67, + 0.706, + 0.732, + 0.719 + ], + "angle": 0, + "content": "(b) Swin-T" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.731, + 0.895, + 0.792 + ], + "angle": 0, + "content": "Table 4: Results of object detection and instance segmentation on various backbones w/ Cascade Mask R-CNN. All results are from our implementation. \\(\\dagger\\): Sync BN is disabled due to gradient accumulation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.897, + 0.901 + ], + "angle": 0, + "content": "System-level comparison. Table 5 compares the object detection and instance segmentation results of SemCL backbones with other models. On Cityscapes and COCO, SemCL backbones comprehensively outperform their twinborn MoCo Specifically, SemCL ResNet50 backbones achieve \\(+1.7\\) over MoCo IG-1B [22] in COCO detection and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5690" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.149, + 0.089, + 0.825, + 0.297 + ], + "angle": 0, + "content": "
BackboneMethodVOC2012CityscapesADE20KCOCO
R50MoCo v3[9] IN-1k†79.6579.0142.7140.59
SemCL79.57(-0.08)79.13(+0.12)42.96(+0.25)40.80(+0.21)
R101timm[55] IN-1k81.8279.8744.4341.43
SemCL81.97(+0.15)80.10(+0.23)44.66(+0.23)41.58(+0.15)
Swin-TSwin[37] IN-22k80.9078.6744.6042.39
SemCL81.73(+0.83)78.91(+0.26)44.72(+0.12)42.53(+0.14)
Swin-SSwin[37] IN-22k83.7181.2547.6244.75
SemCL84.47(+0.76)81.53(+0.28)47.91(+0.29)44.57(-0.18)
Swin-BSwin[37] IN-22k84.6481.6148.2144.81
SemCL84.67(+0.03)81.94(+0.33)48.68(+0.47)45.47(+0.66)
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.16, + 0.298, + 0.825, + 0.323 + ], + "angle": 0, + "content": "\\(\\dagger\\) Although MoCo v3 [9] is originally based on ViT, R50 pretrained models are provided in their repo https://github. com/facebookresearch/moco-v3." + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.332, + 0.895, + 0.364 + ], + "angle": 0, + "content": "Table 2: SemCL vs. ImageNet pretrained backbones on VOC2012, Cityscapes, ADE20K and COCO semantic segmentation (mIoU). All results are from our implementation." + }, + { + "type": "table", + "bbox": [ + 0.149, + 0.39, + 0.372, + 0.571 + ], + "angle": 0, + "content": "
MethodBackbonemIoU
Deeplabv3+[7]R5077.2
Supervised-IN[22]R5074.4
MoCo IG-1B[22]R5075.5
BYOL[18]R5076.3
SemCLR5079.57
Deeplabv3+[7]R10178.85
Res2Net[15]Res2Net-10180.2
DFN[61]R10180.60
SemCLR10181.97
CP2[51]ViT-S/1679.5
SemCLSwin-S84.47
Leopart[70]ViT-B/876.3
SemCLSwin-B84.67
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.223, + 0.574, + 0.298, + 0.585 + ], + "angle": 0, + "content": "(a) VOC2012" + }, + { + "type": "table", + "bbox": [ + 0.377, + 0.389, + 0.609, + 0.571 + ], + "angle": 0, + "content": "
MethodBackbonemIoU
MoCo IN-1M[22]R5072.5
MoCo IG-1B[22]R5073.6
Supervised-IN[22]R5074.6
CP2[51]R5076.5
SemCLR5079.13
Deeplabv3+[7]X-7179.55
SemCLR10180.10
Trans4Trans-T[62]PVTv2-B178.23
SemCLSwin-T78.91
CP2[51]ViT-S/1677.0
Trans4Trans-S[62]PVTv2-B280.02
SemCLSwin-S81.53
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.453, + 0.574, + 0.531, + 0.586 + ], + "angle": 0, + "content": "(b) Cityscapes" + }, + { + "type": "table", + "bbox": [ + 0.612, + 0.389, + 0.825, + 0.571 + ], + "angle": 0, + "content": "
MethodBackbonemIoU
CP2[51]R5039.2
PSPNet[66]R5042.78
SemCLR5042.96
PSPNet[66]R10143.29
SemCLR10144.66
CP2[51]ViT-S/1641.2
Seg-S/16[47]ViT-S45.37
SemCLSwin-S47.91
Seg-B/16[47]Vi-B/1648.06
MAE[21]ViT-B48.1
SemCLSwin-B48.68
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.683, + 0.574, + 0.752, + 0.585 + ], + "angle": 0, + "content": "(c) ADE20K" + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.598, + 0.897, + 0.629 + ], + "angle": 0, + "content": "Table 3: Comparison with previous best results on validation sets of VOC2012, Cityscapes, and ADE20K semantic segmentation (mIoU). Best results are bolded." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.655, + 0.472, + 0.792 + ], + "angle": 0, + "content": "+5.1/+0.6 in Cityscapes/COCO instance segmentation tasks. For SemCL Swin-T backbones, SemCL achieves +1.2/+0.1 over Focal [60] on COCO detection/instance segmentation. Figure 5 demonstrates a qualitative comparison between Mask2Former [10] and SemCL Swin-T on Cityscapes validation set. With the improved spatial information understanding, SemCL model clearly discriminates (even overlapped) instances in Figures 5a and 5b and successfully classify part of an instance to a defined category as in Figure 5c." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.802, + 0.25, + 0.818 + ], + "angle": 0, + "content": "4.3. Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Backbone comparison. SemCL pretrained backbones are compared with corresponding IN pretrained ones in Table 6. For cityscapes, SemCL comprehensively outperforms its IN pretrained counterparts. In particular, the ResNet50 backbones improve the RMSE by 0.138 points. For KITTI" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.895, + 0.745 + ], + "angle": 0, + "content": "and NTUv2, all SemCL backbones show improvements over the IN pretrained ones. The SemCL pretrained SwinL model achieves improvements of 0.001/0.026 points on KITTI AbsRel/RMSE, and the SemCL pretrained ResNet50 model achieves improvements of 0.003/0.005 points on NYUv2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.897, + 0.903 + ], + "angle": 0, + "content": "System-level comparison. We report the comparison of SemCL pretrained backbones with the previous state-of-the-art method Binsformer[35] and other works in Table 7. As for the Cityscapes results shown in Table 7a, SemCL pretrained ResNet50 significantly outperforms SDC [53] by \\(-0.089\\) and \\(-2.218\\) points on AbsRel and RMSE respectively. And SD-SSMDE [43] lags behind SemCL by \\(0.005 / 3.742\\) points. Meanwhile, the test on KITTI (Eigen split [13]) is shown in Table 7b, SemCL pretrained ResNet50 and Swin-T comprehensively outperform the previous SOTA" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "5691" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.101, + 0.332, + 0.226 + ], + "angle": 0, + "content": "
MethodBackboneAP50
Supervised-IN[22]R5081.3
MoCo IN-1M[22]R5081.5
MoCo IG-1B[22]R5082.2
MoCo v2 800ep[8]R5082.5
SemCLR5081.57
UP-DETR/300[12]*80.1
OW-DETR[19]DDETR82.1
SemCLSwin-T86.53
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.131, + 0.228, + 0.308, + 0.24 + ], + "angle": 0, + "content": "(a) Object det. on VOC2007 test." + }, + { + "type": "table", + "bbox": [ + 0.336, + 0.101, + 0.58, + 0.226 + ], + "angle": 0, + "content": "
MethodBackboneAPmask
MoCo IN-1M[22]R5032.3
MoCo IG-1B[22]R5032.9
Supervised-IN[22]R5032.9
Mask2Former[10]R5037.4
SemCLR5038.0
BoundaryFormer[30]*38.3
Mask2Former[10]Swin-T39.7
SemCLSwin-T40.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.362, + 0.228, + 0.551, + 0.241 + ], + "angle": 0, + "content": "(b) Instance seg. on Cityscapes val." + }, + { + "type": "table", + "bbox": [ + 0.584, + 0.101, + 0.865, + 0.226 + ], + "angle": 0, + "content": "
MethodBackboneAPboxAPmask
Supervised-IN[22]R5040.636.8
MoCo IN-1M[22]R5040.836.9
MoCo IG-1B[22]R5041.137.4
SemCLR5042.838.0
MST[34]Swin-T42.738.8
VSA[63]Swin-T46.942.1
Focal[60]Focal-Base47.843.2
SemCLSwin-T49.043.3
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.585, + 0.228, + 0.86, + 0.24 + ], + "angle": 0, + "content": "(c) Object det. and instance seg. on COCO val2017." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.253, + 0.895, + 0.299 + ], + "angle": 0, + "content": "Table 5: System-level comparison on object detection and instance segmentation. The best results are **bolded**. * claims their backbones are CNN-based, but are considered and compared as transformer-based networks, since transformer blocks are massively used." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.312, + 0.896, + 0.446 + ], + "angle": 0, + "content": "
BackboneMethodCityscapesKITTINYUv2
AbsRelRMSEAbsRelRMSEAbsRelRMSE
R50MoCo v3[9] IN-1k0.1384.8370.0592.3370.1200.405
SemCL0.1384.699(-0.138)0.058(-0.001)2.3370.117(-0.003)0.400(-0.005)
Swin-TSwin[37] IN-22k0.1344.6430.0572.2470.1140.386
SemCL0.133(-0.001)4.577(-0.066)0.0572.237(-0.010)0.112(-0.002)0.381(-0.005)
Swin-LSwin[37] IN-22k0.1304.5360.0532.1300.0950.329
SemCL0.128(-0.002)4.426(-0.110)0.052(-0.001)2.104(-0.026)0.094(-0.001)0.329
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.454, + 0.893, + 0.486 + ], + "angle": 0, + "content": "Table 6: SemCL vs. ImageNet pretraining on Cityscapes, KITTI and NYUv2 depth estimation (lower is better). All results are from our implementation." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.51, + 0.468, + 0.583 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.584, + 0.468, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.656, + 0.468, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.741, + 0.47, + 0.786 + ], + "angle": 0, + "content": "Figure 5: Qualitative comparison of instance segmentation between Mask2Former [10] (left) and SemCL (right) Swin-T on Cityscapes validation set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Binsformer. In particular, SemCL ResNet50 achieves a gain of \\(-0.003 / - 0.089\\) over Binsformer, and SemCL Swin-L is on par with Binsformer Swin-L. On NYUv2 (see Table 7c), the gains of SemCL pretrained ResNet50 are also high with \\(-0.010\\) AbsRel and \\(-0.173\\) RMSE. The SemCL" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.522, + 0.892, + 0.769 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.78, + 0.895, + 0.824 + ], + "angle": 0, + "content": "Figure 6: Qualitative comparison of depth estimation between Binsformer[35] and SemCL Swin-T on the NYUv2 test set." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Swin-L RMSE is 0.001 points lower than the SOTA Binsformer. Qualitative comparison between Binsformer [35] and SemCL Swin-T is given in Figure 6." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5692" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.101, + 0.398, + 0.226 + ], + "angle": 0, + "content": "
MethodBackboneAbsRelRMSE
UpProj[29]R500.257[53]7.273[53]
Pad-net[59]R500.246[53]7.117[53]
TRL[64]R500.234[53]7.104[53]
Pilzer, et al.[45]-0.4405.443
SDC-Depth[53]R500.2276.917
DTS-Depth[27]MobileNetv20.1677.785
struct2depth[5]-0.1517.024
SD-SSMDE[43]R500.1438.441
SemCLR500.1384.699
" + }, + { + "type": "table_caption", + "bbox": [ + 0.21, + 0.228, + 0.287, + 0.241 + ], + "angle": 0, + "content": "(a) Cityscapes" + }, + { + "type": "table", + "bbox": [ + 0.401, + 0.101, + 0.63, + 0.226 + ], + "angle": 0, + "content": "
MethodBackboneAbsRelRMSE
EPCDepth [42]R500.0914.207
Binsformer[35]R500.0612.426
SemCLR500.0582.337
MonoViT[65]-0.0984.333
Binsformer [35]Swin-T0.0582.286
SemCLSwin-T0.0572.237
DepthFormer [33]Swin-L +R50CI0.0522.143
Binsformer[35]Swin-L0.0522.098
SemCLSwin-L0.0522.104
" + }, + { + "type": "table_caption", + "bbox": [ + 0.487, + 0.228, + 0.542, + 0.24 + ], + "angle": 0, + "content": "(b) KITTI" + }, + { + "type": "table", + "bbox": [ + 0.633, + 0.101, + 0.873, + 0.226 + ], + "angle": 0, + "content": "
MethodBackboneAbsRelRMSE
Pad-net [59]R500.2140.792
TRL [64]R500.1440.501
SDC-Depth [53]R500.1280.497
UpProj [29]R500.1270.573
SemCLR500.1170.400
Binsformer [35]Swin-T0.1130.379
SemCLSwin-T0.1120.381
Binsformer [35]Swin-L0.0940.330
SemCLSwin-L0.0940.329
" + }, + { + "type": "table_caption", + "bbox": [ + 0.721, + 0.228, + 0.782, + 0.24 + ], + "angle": 0, + "content": "(c) NYUv2" + }, + { + "type": "table_caption", + "bbox": [ + 0.168, + 0.253, + 0.799, + 0.268 + ], + "angle": 0, + "content": "Table 7: System-level comparison on depth estimation. Best results (lower is better) are bolded." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.295, + 0.331, + 0.311 + ], + "angle": 0, + "content": "4.4. Attention Map Visualization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.471, + 0.453 + ], + "angle": 0, + "content": "To improve the qualitative understanding of the SemCL representation, we visualize attention maps of Swin-T pretrained on SemCL-VOC and ImageNet using PyTorch library for CAM methods [16] with the Score-CAM method [52]. These samples in Figure 7 selected from VOC2012 val and have not been seen by the pretrained models are characterized by simple centered subjects (Figure 7a), complex scene with occlusion (Figure 7b) and multiple subjects (Figures 7c and 7d)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.454, + 0.47, + 0.665 + ], + "angle": 0, + "content": "In Figure 7a, the IN pretrained model failed to focus on the subject, while the SemCL model accurately depicted the area of the main subject. Meanwhile, SemCL performs better in complex scenes containing foreground-background relationships, as shown in Figure 7b. The IN pretrained model pays no attention to an occluded object, but the SemCL model recognizes that the occluded part also belongs to the horse, as shown in Figure 7b. In Figure 7c, the IN model directly ignores the distant person, but the SemCL model not only annotates two people but also focuses more on the nearby one, suggesting that the SemCL model can understand the spatial distribution of subjects in an image. Also, in Figure 7d, the SemCL model successfully recognizes one plant and its pot as one object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.666, + 0.471, + 0.833 + ], + "angle": 0, + "content": "Summary. SemCL pretrained backbones can substantially outperform ImageNet pretrained counterparts on semantic segmentation, object detection, instance segmentation and depth estimation tasks. Using semantically contrastive pairs generated from off-the-shelf datasets, SemCL significantly improves the spatial information understanding of pretrained models. Such an improvement is also promising to benefit more challenging tasks including generative ones [32]. Ablation study concerning paired/unpaired InfoNCE loss, pretraining batch size, training length and dataset scale are given in Appendix A.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.195, + 0.86 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We investigate the effectiveness of semantic information in the contrastive learning pretext task on the spatial un" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.293, + 0.64, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.385, + 0.64, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.475, + 0.639, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.566, + 0.64, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.293, + 0.765, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.384, + 0.765, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.474, + 0.765, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.565, + 0.766, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.293, + 0.892, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.385, + 0.891, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.475, + 0.891, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.565, + 0.891, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.668, + 0.894, + 0.713 + ], + "angle": 0, + "content": "Figure 7: Swin-T attention map visualization. Left: raw image. Middle: attention maps of the IN pretrained model. Right: attention maps of the SemCL pretrained model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.74, + 0.894, + 0.891 + ], + "angle": 0, + "content": "derstanding ability of models. Tests are conducted on semantic/instance segmentation, object detection and depth estimation tasks with both ResNets and Swin Transformers, and SemCL pretrained backbones substantially outperform their ImageNet pretrained counterparts. By supporting small batch sizes and fast pretraining, SemCL is a lightweight yet effective approach. In contrastive representation learning, if dedicated pretext task is designed properly, four ounces can move a thousand pounds. We hope that SemCL will inspire more advanced pretext tasks for contrastive learning." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "5693" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.212 + ], + "angle": 0, + "content": "[2] Suzanne Becker and Geoffrey E Hinton. Self-organizing neural network that discovers surfaces in random-dot stereograms. Nature, 355(6356):161-163, 1992. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.269 + ], + "angle": 0, + "content": "[3] Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah. Signature verification using a “siamese” time delay neural network. Advances in neural information processing systems, 6, 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.471, + 0.324 + ], + "angle": 0, + "content": "[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Computer vision and pattern recognition (CVPR), 2018 IEEE conference on. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.396 + ], + "angle": 0, + "content": "[5] Vincent Casser, Soeren Pirk, Reza Mahjourian, and Anelia Angelova. Unsupervised monocular depth and ego-motion learning with structure and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.471, + 0.45 + ], + "angle": 0, + "content": "[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.47, + 0.521 + ], + "angle": 0, + "content": "[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.62 + ], + "angle": 0, + "content": "[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.621, + 0.471, + 0.69 + ], + "angle": 0, + "content": "[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.691, + 0.471, + 0.773 + ], + "angle": 0, + "content": "[11] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.775, + 0.471, + 0.843 + ], + "angle": 0, + "content": "[12] Zhigang Dai, Bolun Cai, Yugeng Lin, and Junying Chen. Updetr: Unsupervised pre-training for object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1601-1610, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.845, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[13] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 6" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.16 + ], + "angle": 0, + "content": "[14] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.219 + ], + "angle": 0, + "content": "[15] Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multiscale backbone architecture. IEEE transactions on pattern analysis and machine intelligence, 43(2):652-662, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.894, + 0.262 + ], + "angle": 0, + "content": "[16] Jacob Gildenblat and contributors. Pytorch library for cam methods. https://github.com/jacobgil/pytorch-grad-cam, 2021.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.894, + 0.319 + ], + "angle": 0, + "content": "[17] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: TrainingImagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.894, + 0.403 + ], + "angle": 0, + "content": "[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems, 33:21271-21284, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.894, + 0.473 + ], + "angle": 0, + "content": "[19] Akshita Gupta, Sanath Narayan, KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Ow-detr: Open-world detection transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9235-9244, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.475, + 0.894, + 0.531 + ], + "angle": 0, + "content": "[20] Bharath Hariharan, Pablo Arbelaez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In International Conference on Computer Vision (ICCV), 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.894, + 0.601 + ], + "angle": 0, + "content": "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.603, + 0.894, + 0.672 + ], + "angle": 0, + "content": "[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.894, + 0.731 + ], + "angle": 0, + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.894, + 0.785 + ], + "angle": 0, + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pages 630-645. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.894, + 0.83 + ], + "angle": 0, + "content": "[25] Olivier Henaff. Data-efficient image recognition with contrastive predictive coding. In International Conference on Machine Learning, pages 4182-4192. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[26] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In International Conference on Learning Representations, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5694" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.135 + ], + "angle": 0, + "content": "[27] Hatem Ibrahim, Ahmed Salem, and Hyun-Soo Kang. Dts-depth: Real-time single-image depth estimation using depth-to-space image construction. Sensors, 22(5):1914, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[28] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.472, + 0.275 + ], + "angle": 0, + "content": "[29] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In 2016 Fourth international conference on 3D vision (3DV), pages 239-248. IEEE, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.472, + 0.346 + ], + "angle": 0, + "content": "[30] Justin Lazarow, Weijian Xu, and Zhuowen Tu. Instance segmentation with mask-supervised polygonal boundary transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4382-4391, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.472, + 0.389 + ], + "angle": 0, + "content": "[31] Phuc H. Le-Khac, Graham Healy, and Alan F. Smeaton. Contrastive representation learning: A framework and review. IEEE Access, 8:193907-193934, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.472, + 0.459 + ], + "angle": 0, + "content": "[32] Yuheng Li, Yijun Li, Jingwan Lu, Eli Shechtman, Yong Jae Lee, and Krishna Kumar Singh. Contrastive learning for diverse disentangled foreground generation. In European Conference on Computer Vision, pages 334-351. Springer, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.472, + 0.518 + ], + "angle": 0, + "content": "[33] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. arXiv preprint arXiv:2203.14211, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.52, + 0.472, + 0.588 + ], + "angle": 0, + "content": "[34] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, et al. Mst: Masked self-supervised transformer for visual representation. Advances in Neural Information Processing Systems, 34:13165-13176, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.472, + 0.631 + ], + "angle": 0, + "content": "[35] Zhenyu Li, Xuyang Wang, Xianming Liu, and Junjun Jiang. Binsformer: Revisiting adaptive bins for monocular depth estimation. arXiv preprint arXiv:2204.00987, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.472, + 0.701 + ], + "angle": 0, + "content": "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[37] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[38] Vishnu Suresh Lokhande, Songwong Tasneeyapant, Abhay Venkatesh, Sathya N Ravi, and Vikas Singh. Generating accurate pseudo-labels in semi-supervised learning and avoiding overconfident predictions via hermite polynomial activations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11435-11443, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[39] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, April" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.532, + 0.093, + 0.893, + 0.119 + ], + "angle": 0, + "content": "24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[40] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.164, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[41] Chao Peng, Tete Xiao, Zeming Li, Yuning Jiang, Xiangyu Zhang, Kai Jia, Gang Yu, and Jian Sun. Megdet: A large minibatch object detector. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6181-6189, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.894, + 0.302 + ], + "angle": 0, + "content": "[42] Rui Peng, Ronggang Wang, Yawen Lai, Luyang Tang, and Yangang Cai. Excavating the potential capacity of self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15560-15569, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.304, + 0.893, + 0.371 + ], + "angle": 0, + "content": "[43] Andra Petrovai and Sergiu Nedevschi. Exploiting pseudo labels in a self-supervised learning framework for improved monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1578-1588, 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.373, + 0.893, + 0.426 + ], + "angle": 0, + "content": "[44] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V Le. Meta pseudo labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11557-11568, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.893, + 0.482 + ], + "angle": 0, + "content": "[45] Andrea Pilzer, Dan Xu, Mihai Puscas, Elisa Ricci, and Nicu Sebe. Unsupervised adversarial depth estimation using cycled generative networks. In 2018 international conference on 3D vision (3DV), pages 587-595. IEEE, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.485, + 0.893, + 0.551 + ], + "angle": 0, + "content": "[46] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, volume 11006, pages 369-386. SPIE, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.893, + 0.608 + ], + "angle": 0, + "content": "[47] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7262-7272, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.61, + 0.893, + 0.65 + ], + "angle": 0, + "content": "[48] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In European conference on computer vision, pages 776-794. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.652, + 0.893, + 0.692 + ], + "angle": 0, + "content": "[49] Aaron Van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv e-prints, pages arXiv-1807, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.694, + 0.893, + 0.747 + ], + "angle": 0, + "content": "[50] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.893, + 0.817 + ], + "angle": 0, + "content": "[51] Feng Wang, Huiyu Wang, Chen Wei, Alan Yuille, and Wei Shen. Cp 2: Copy-paste contrastive pretraining for semantic segmentation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXX, pages 499-515. Springer, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.819, + 0.893, + 0.9 + ], + "angle": 0, + "content": "[52] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 8" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5695" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[53] Lijun Wang, Jianming Zhang, Oliver Wang, Zhe Lin, and Huchuan Lu. Sdc-depth: Semantic divide-and-conquer network for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 541-550, 2020. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[54] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7303-7313, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.472, + 0.273 + ], + "angle": 0, + "content": "[55] Ross Wightman, Hugo Touvron, and Hervé Jégou. Resnet strikes back: An improved training procedure in timm. arXiv preprint arXiv:2110.00476, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.472, + 0.327 + ], + "angle": 0, + "content": "[56] Hao Wu and Saurabh Prasad. Semi-supervised deep learning using pseudo labels for hyperspectral image classification. IEEE Transactions on Image Processing, 27(3):1259-1270, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.472, + 0.398 + ], + "angle": 0, + "content": "[57] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.472, + 0.455 + ], + "angle": 0, + "content": "[58] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.472, + 0.524 + ], + "angle": 0, + "content": "[59] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Padnet: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 675-684, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.472, + 0.58 + ], + "angle": 0, + "content": "[60] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal attention for long-range interactions in vision transformers. Advances in Neural Information Processing Systems, 34:30008-30022, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.472, + 0.649 + ], + "angle": 0, + "content": "[61] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Learning a discriminative feature network for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1857-1866, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.472, + 0.733 + ], + "angle": 0, + "content": "[62] Jiaming Zhang, Kailun Yang, Angela Constantinescu, Kunyu Peng, Karin Müller, and Rainer Stiefelhagen. Trans4trans: Efficient transformer for transparent object and semantic scene segmentation in real-world navigation assistance. IEEE Transactions on Intelligent Transportation Systems, 23(10):19173-19186, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.472, + 0.803 + ], + "angle": 0, + "content": "[63] Qiming Zhang, Yufei Xu, Jing Zhang, and Dacheng Tao. Vsa: learning varied-size window attention in vision transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXV, pages 466-483. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[64] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Zequn Jie, Xiang Li, and Jian Yang. Joint task-recursive learning for semantic segmentation and depth estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 235–251, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[65] Chaoqiang Zhao, Youmin Zhang, Matteo Poggi, Fabio Tosi, Xianda Guo, Zheng Zhu, Guan Huang, Yang Tang, and" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.532, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Stefano Mattoccia. Monovit: Self-supervised monocular depth estimation with a vision transformer. arXiv preprint arXiv:2208.03543, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[66] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[67] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[68] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[69] Chengxu Zhuang, Alex Lin Zhai, and Daniel Yamins. Local aggregation for unsupervised learning of visual embeddings. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6002-6012, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[70] Adrian Ziegler and Yuki M Asano. Self-supervised learning of object parts for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14502-14511, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5696" + } + ] +] \ No newline at end of file diff --git a/2023/Semantic Information in Contrastive Learning/c3144656-62d8-4975-97d6-ef4630f410df_origin.pdf b/2023/Semantic Information in Contrastive Learning/c3144656-62d8-4975-97d6-ef4630f410df_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2fd983386092644e872607273806714f0f110199 --- /dev/null +++ b/2023/Semantic Information in Contrastive Learning/c3144656-62d8-4975-97d6-ef4630f410df_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e218b4fccb8a5db52bc216fe320d7f3b8cd1d5776971d71e082d902d83667a89 +size 8409101 diff --git a/2023/Semantic Information in Contrastive Learning/full.md b/2023/Semantic Information in Contrastive Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3e12e60ac8bf6f884ed3f6244604a5a58e435a2d --- /dev/null +++ b/2023/Semantic Information in Contrastive Learning/full.md @@ -0,0 +1,346 @@ +# Semantic Information in Contrastive Learning + +Shengjiang Quan $^{1*}$ , Masahiro Hirano $^{2}$ and Yuji Yamakawa $^{3}$ \ + $^{1}$ Graduate School of Engineering, The University of Tokyo, Japan\ + $^{2}$ Institute of Industrial Science, The University of Tokyo, Japan\ + $^{3}$ Interfaculty Initiative in Information Studies, The University of Tokyo, Japan\ +quan@g.ecc.u-tokyo.ac.jp, {mhirano, y-ymkw}@iis.u-tokyo.ac.jp + +# Abstract + +This work investigates the functionality of Semantic information in Contrastive Learning (SemCL). An advanced pretext task is designed: a contrast is performed between each object and its environment, taken from a scene. This allows the SemCL pretrained model to extract objects from their environment in an image, significantly improving the spatial understanding of the pretrained models. Downstream tasks of semantic/instance segmentation, object detection and depth estimation are implemented on PASCAL VOC, Cityscapes, COCO, KITTI, etc. SemCL pretrained models substantially outperform ImageNet pretrained counterparts and are competitive with well-known works on downstream tasks. The results suggest that a dedicated pretext task leveraging semantic information can be powerful in benchmarks related to spatial understanding. The code is available at https://github.com/sjiang95/semcl. + +# 1. Introduction + +Within the field of visual representation learning, contrastive learning attracts special attention for its inspiring performance in transfer learning [57, 49, 31, 22, 8, 9]. The concept of contrastive learning can literally be explained as discovering the difference between positive and negative samples. The definition of positive-negative samples is one of the main subjects of contrastive learning, which directly determines the pretext task and the corresponding loss function. + +We humans can recognize an object from its even complicated environment because we have learned and bound the typical geometric feature and the corresponding semantic concept of the object class. For example, we can recognize a cat in a scene thanks to our knowledge of its appearance (typical geometric feature) and the concept of cat (semantic information). Similarly, the target of prevalent + +![](images/bab89fd9041f4a80e562caeb78e23e0262a1f6e2656bcaff36d7a6fdd19d7503.jpg) +Figure 1: Perceptual cognition of SemCL. Leveraging semantic information, SemCL separates an object from its environment. Each sample is divided into objects and surroundings and considered as a contrastive pair. The task is to maximize the distance between the object and its environment in the embedding space. + +visual tasks, such as semantic/instance segmentation, object detection, etc., can be summarized as distinguishing and localizing subjects from surroundings. Inspired by and to extend the spatial recognition mechanism to CV, we propose our method SemCL which directly teaches models to extract a subject from its environment. Utilizing publicly off-the-shelf datasets providing semantic labels, SemCL can be considered as a supervised contrastive learning approach whose samples consist of subjects and their corresponding surroundings (semantically not the subject). The pretext task of SemCL is to tell the difference between one subject and its surroundings. From a representation learning perspective, the pretext task is to maximize the distance $(L^2$ norm) between a subject and its environment in the embedding space (see Figure 1). The pretext task mimics the spatial recognition mechanism to discriminate objects and surroundings, which significantly improves the spatial information understanding of pretrained models. + +In SemCL, contrast is performed pairwise between positive-negative pairs from one image/scene. Inter-scene contrast is considered inappropriate: only the contrasts + +among subject-surrounding pairs from one scene are considered valid. Therefore, paired InfoNCE loss is adopted. With the mechanism of contrasting only paired samples, SemCL is characterized by decoupling the number of negative samples from the batch size. Compared to MoCo [22], whose performance is positively correlated with batch size and thus increases hardware cost (e.g. batchsize $= 4096$ on 128 GPUs for ViT-B in MoCo v3 [9]) for satisfactory results, SemCL can achieve consistent yet competitive results on downstream tasks with relatively small batch size (e.g., batchsize $= 64 \ @ 224 \times 224$ ). We adopt the MoCo v3 [9] as the pretraining framework. + +The primary motivation of representation learning is to pretrain general representations that can be fine-tuned and transferred to downstream tasks. In this work, the SemCL pretrained models are benchmarked on semantic segmentation, object detection and instance segmentation, and depth estimation tasks. SemCL models substantially outperform their ImageNet pretrained counterparts due to the improved ability of spacial information understanding. In a system-level comparison, SemCL models also show gains over previous well-known works (e.g. Deeplabv3+ [7] in semantic segmentation, Mask2Former [10] in instance segmentation and Binsformer [35] in depth estimation). + +# 2. Related Work + +# 2.1. Contrastive Learning + +The seminal idea of contrastive learning dates back at least to the 1990s [2, 3, 31], but it has become prominent in recent years thanks to the large pretrained models in the fields of NLP and CV [31]. As the name suggests, contrastive learning mines abstract representations by comparing between "similar" and "dissimilar" samples: minimizing the representation of "similar" samples and maximizing that of "dissimilar" samples in the embedding space. Formally, the "similar" inputs are called positive samples, and the "dissimilar" inputs are called negative samples. + +Defining "similar" and "dissimilar" samples is an important topic in contrastive learning. In instance discrimination [57], which is a ubiquitous approach to define positive-negative samples in CV, each sample in a minibatch $x \in \{x_1, \ldots, x_n\}$ is considered mutually exclusive to the rest. The encoded output $\theta(\cdot)$ of each sample is pushed away from the others in the embedding space, denoted as $\theta(x_1) \Leftrightarrow \{\theta(x_2), \ldots, \theta(x_n)\} \Leftrightarrow \text{push}$ . Contrastive Predictive Coding (CPC) [49] defines positive-negative samples in a generative way: Given a sequential input $\{x_{t-3}, x_{t-2}, x_{t-1}, x_t\}$ , a context latent representation $C_t$ is computed to predict the encoded outputs of future inputs $\{\theta(\mathbb{X}_{t+1}), \theta(\mathbb{X}_{t+2}), \theta(\mathbb{X}_{t+3}), \ldots\}$ . Those predicted future embeddings are considered as positive samples to real future embeddings, and exclusive to some other + +irrelevant inputs. + +# 2.2. Momentum Contrast + +He et al. [22] proposed a momentum contrast framework, MoCo, for contrastive representation learning. MoCo summed unsupervised visual representation learning [57, 49, 26, 69, 25, 48, 1] up as dictionary look-up: the input data are represented and sampled by an encoder network $\theta_{k}$ as "keys" $k = \theta_{k}(x^{k})$ . The goal of contrastive learning is to train encoders $\theta_{q}$ by which a "query" $q = \theta_{q}(x^{q})$ should be similar to its matching key and dissimilar to others. Since dictionary consistency is considered crucial for unsupervised learning with contrastive loss [22], MoCo introduces momentum updating. In the MoCo framework, the key coder $\theta_{k}$ is progressively updated by the weighted average of $\theta_{k}$ and the query coder $\theta_{q}$ . + +$$ +\theta_ {k} \leftarrow m \theta_ {k} + (1 - m) \theta_ {q}, \tag {1} +$$ + +where the momentum coefficient $m \in [0,1)$ is set quite high ( $m = 0.999$ in [22]) for slower key encoder update. The MoCo series has proven the power of the Momentum Contrast framework on feature transferring and is therefore employed as the basis for this work. + +# 2.3. Labeled Data In Unsupervised Framework + +Large networks are quite thirsty for training data, while labelling the large amount of collected data is a human resource consuming challenge. Apart from turning to unsupervised learning, which inherently avoids using labelled data, using unlabeled data in supervised learning is also a solution. Pseudo-Label [44, 38, 56] is a trick that uses (usually supervised) pretrained networks to generate features that can play the role of labels. And the generated pseudo-labels participate in the iterations of the target network just like annotations. + +In contrast, Khosla et al. [28] also investigated the function of labelled data in unsupervised learning. As a contrastive learning research, Khosla et al. [28] intuitively uses the classification label to define samples within the same class are mutually positive, and those from other categories are negative. To exploit the label information, clusters of points belonging to the same class are pulled together in the embedding space, while clusters of samples from different classes are pushed apart [28]: $\left\{\theta\left(x_1^{c0}\right) \rightleftharpoons \theta\left(x_2^{c0}\right) \rightleftharpoons \theta\left(x_{n}^{c0}\right) \rightleftharpoons \theta\left(x_n^0\right)$ $\Leftrightarrow$ $\left\{\theta\left(x_1^{c1}\right) \rightleftharpoons \theta\left(x_2^{c1}\right) \rightleftharpoons \theta\left(x_{m}^{c1}\right)$ ( $\rightleftharpoons$ for pull). The experimental result shows that supervised contrastive learning can improve both the accuracy and robustness of classifiers - [28] achieves a top-1 accuracy of $81.4\%$ on the ImageNet (IN) dataset, which is $0.8\%$ above the best result reported by ResNet-200 architecture [24]. Likewise, our work investigates the effectiveness of semantic information in unsupervised framework contrastive learning. + +# 2.4. Semantic Information in Contrastive Learning + +Other than [28], there are works that use small amounts of labelled data in unsupervised representation learning. Wang et al. [54] proposed a pixel-wise contrastive algorithm for semantic segmentation: for a pixel $i$ , it is pulled towards pixels belonging to the same category and pushed away from pixels in other categories in the embedding space. This contrast is performed across images from the same dataset. In the Cityscapes test benchmark, [54] R101 achieves 79.2 mIoU, which is 1.1 points better than Deeplabv3 [6]. + +The attempt and results of works using labels in unsupervised representation learning, such as [28, 54], indicate that label information can bring gains to a fully unsupervised representation. It is promising to explore advanced pretext tasks to improve the performance of pretrained models in downstream tasks. + +# 3. Method + +# 3.1. SemCL Dataset + +The core of this work is to define semantically contrastive pairs by utilizing off-the-shelf semantic labels, which are mappings between categories of objects and their coordinates on the raw image at the pixel level. As decided by the authors of a dataset, semantic labels can be color or grayscale images. Each predefined color or grayscale denotes a category, and the distribution of label pixels reflects the location of pixels belonging to a particular category in the raw image. + +The SemCL dataset is generated by the mechanism shown in Figure 2, where the raw image and its semantic label are selected from the PASCAL VOC2012 dataset [14]. From the semantic label annotating two objects - a motorcycle and a rider - two binary masks BM0 and BM1 can be extracted to separately represent two annotated objects. Their inverted counterparts $\neg$ BM0 and $\neg$ BM1 represent pixels other than the target object. These binary masks are then bitwise applied to the raw image to separate an object (Anchor) from its environment ( $\neg$ Anchor), forming a semantically contrastive pair. Each element in a raw Anchor- $\neg$ Anchor pair is considered to be semantically opposite to the other. + +In practice, the SemCL-1M dataset is produced from training sets of augmented PASCAL VOC2012 [14] (also known as SBD [20]), Cityscapes [11], ADE20K [67, 68] and COCO [36] (stuff+thing 2017 [4]) datasets. The components of the SemCL dataset are illustrated in Table 1. When generating the dataset, a threshold $t = 0.01$ is set to filter out too small (less than $[\sqrt{t} \times \text{height}] \times [\sqrt{t} \times \text{width}]$ ) objects that are considered semantically unsaturated. + +# 3.2. Paired InfoNCE Loss + +Since contrastive learning can be considered as a dictionary look-up task [22], it is assumed that within a set of keys $\{k_0, k_1, k_2, \ldots\}$ of a dictionary, a query $q$ matches only one + +
VOC2012CityscapesADE20KCOCOTotal
# training samples10,5822,97525,574118,287157,418
# SemCL pairs14,20320,280270,218712,2121,016,913
+ +Table 1: Components of the SemCL dataset. The SemCL datasets spawned from VOC2012(aug), Cityscapes, ADE20k and COCO 2017 are named SemCL- $\{\mathrm{VOC}$ , City, ADE, COCO\} respectively. + +key $k_{+}$ of the dictionary [22]. The value of the objective function should be low if $q$ is similar to its positive key $k_{+}$ and dissimilar to all negative keys $\{k_{0}, k_{1}, k_{2}, \ldots\} \setminus \{k_{+}\}$ . In MoCo [22], the form of a contrastive loss function InfoNCE [49] is adopted + +$$ +\mathcal {L} _ {q} = - \log \frac {\exp (q \cdot k _ {+} / \tau)}{\sum_ {i = 0} ^ {K} \exp (q \cdot k _ {i} / \tau)} \tag {2} +$$ + +where $\tau$ is a scalar temperature hyperparameter controlling the strength of penalties on hard negative samples [50], e.g. in the instance discrimination task [49, 22], there may be potential positive samples in the randomly chosen negative samples. Summing over one positive and $K$ negative samples, Equation (2) is the log loss of a $(\mathbf{K} + 1)$ -way softmax-based classifier trying to classify $q$ as $k_{+}$ [22]. While in our case an anchor instance $q$ compares only with its own $n$ augmentations (positive) $K^{+} := \{k_{1}^{+}, k_{2}^{+}, k_{3}^{+}, \ldots, k_{n}^{+}\}$ and paired $n + 1$ negatives $K^{-} := \{k_{0}^{-}, k_{1}^{-}, k_{2}^{-}, k_{3}^{-}, \ldots, k_{n}^{-}\}$ using paired InfoNCE + +$$ +\mathcal {L} _ {q} = - \sum_ {j = t + 1} ^ {n} \log \frac {\exp \left(q _ {t} \cdot k _ {j} ^ {+} / \tau\right)}{\sum_ {i = 0} ^ {K ^ {-}} \exp \left(q _ {t} \cdot k _ {i} ^ {-} / \tau\right) + \exp \left(q _ {t} \cdot k _ {j} ^ {+} / \tau\right)}. \tag {3} +$$ + +For one anchor as query $q_{0}(t = 0)$ , the paired InfoNCE loss tries to classify it as $k_{j}^{+}, j \in [1, n]$ . The next step is to go forward: classify the first positive $q_{1}(t = 1)$ of the anchor as $k_{j}^{+}, j \in [2, n]$ , and so on, as shown in the algorithm 1, Appendix A.1. To contrast with the paired InfoNCE, we refer to the original InfoNCE loss as unpaired InfoNCE hereafter. The behavior difference between unpaired and paired InfoNCE losses is compared in Figure 3. The paired InfoNCE only maximizes the distance between anchors and -anchors in the embedding space, neglecting irrelevant samples. The implementation of the paired InfoNCE loss is introduced in detail in Appendix A.2. + +# 3.3. Pretext Task + +Since the aim of this work is to enable networks to strip targets (anchors) from their surroundings ( $\neg$ anchors), we use a restricted instance discrimination task [57]: a query and a key are positive pairs if they are different views (by + +![](images/94bcab4c34e62920d5c29550c8ba28bbd49c686ee525a78e3d55635c96dac62c.jpg) +Figure 2: SemCL dataset. The semantic label of a sample is utilized to generate binary masks. Each binary mask depicts the pixel-level information of one object (anchor) in the raw image. Those binary masks, together with their inverts, are applied on the raw image to extract objects (anchors) and corresponding surroundings ( $\neg$ anchors). + +![](images/f00a92c7ece71a2863a0c55f5a16a3c78bc0764ebe9e51e37957ba554ae0cf66.jpg) +Figure 3: Comparison between unpaired (left) and paired InfoNCE losses (right). Left: a sample (anchor) is positive only to its own augmentations, while negative to others in the same mini-batch. Right: A sample (anchor) is positive only to its own augmentations, while negative only to its paired counterparts. + +![](images/562f206f886369d73ce424b9fecb105a9fbdc964e9552b00571342f39e1ab72a.jpg) + +data augmentation) of the same anchor, and are negative pairs if they originate from an anchor and its corresponding surroundings ( $\neg$ anchors). A PyTorch-style pseudocode is given in the Appendix A.2. The inputs are encoded by base encoder $f_{b}$ and momentum encoder $f_{m}$ , producing queries and keys. Theoretically, the backbone of $f_{b}$ and $f_{m}$ can be any architecture. In practice, we use ResNet [23] and Swin Transformer [37] because they are the most representative architectures in recent years. All keys are stacked along + +dimension 1 for paired InfoNCE loss. + +# 4. Experiments + +Considering the enormous scale difference between ImageNet and SemCL dataset, all SemCL models are initialized by corresponding IN pretrained weights, which are compared as counterparts in downstream tasks. The total number of pretraining iterations is set to $30k$ to minimize the significant + +variation in training time caused by scale differences among SemCL sub-datasets. We adopt AdamW [40] with $wd = 0.1$ , and the OneCycle [46] learning rate schedule consisting of warmup [17] for the initial 3, 750 (warmup ratio 0.125) steps and cosine annealing [9, 39]. Following [9], the base $lr$ is set to $1.5 \times 10^{-4}$ and scaled linearly by $lr \times batchsize / 256$ and temperature $\tau = 0.2$ . The initial MoCo momentum is 0.99 and is gradually increased to 1 with a half cycle cosine schedule. The data augmentation strategy includes $224 \times 224$ random crops after random rescaling the original image with an area ratio in the range [0.08, 1.0], followed by random color jitter, random grayscale, random Gaussian blur, and random horizontal flip. All batch norm (BN) layers are synchronized across GPUs (SyncBN [41]). Limited by the two dual-GPU nodes (RTX3090 $\times$ 2+Titan RTX $\times$ 2) we use, a native batch size of 64 can be applied to all ResNets and up to Swin-B for Swin Transformers with crop size $224 \times 224$ . For Swin-L in particular, we compromise by using a batch size of 32 with gradient accumulation over 2 iterations to achieve an equivalent batch size of 64. + +SemCL aims to improve the understanding of spatial information of models for which downstream tasks of semantic segmentation, object detection and depth estimation are conducted. Unless otherwise stated, all SemCL backbones are pretrained on the corresponding SemCL sub-datasets. The implementation details of downstream tasks are listed in Appendix A.3. + +# 4.1. Semantic Segmentation + +Backbone comparison. In the semantic segmentation benchmark on validation sets of VOC2012, Cityscapes, ADE20K and COCO 2017, SemCL models pretrained on the corresponding SemCL sub dataset are compared with ImageNet pretrained counterparts in Table 2. For VOC2012 and COCO, SemCL outperforms its IN counterparts by a maximum of 0.83 and 0.66 points respectively. SemCL can comprehensively outperform its IN counterparts on all backbones in the Cityscapes and ADE20k benchmarks. + +System-level comparison. SemCL pretrained backbones are compared with other methods on semantic segmentation tasks in Table 3. On VOC2012, SemCL R50/R101 outperform Deeplabv3+ [7] by 2.37/3.12 points. The Deeplabv3+ [7] X-71 is 0.55 points behind SemCL R101 on Cityscapes. For $\mathrm{CP}^2$ [51] ViT-S/16, SemCL Swin-S significantly outperforms it on all benchmarks by 4.97, 4.53 and 6.71 points. Compared to MAE [21] ViT-B, SemCL Swin-B achieves a gain of 0.58 points. A qualitative comparison between MAE [21] ViT-B and SemCL Swin-B is shown in Figure 4. + +# 4.2. Object Detection and Instance Segmentation + +Backbone comparison. In the benchmark for object detection and instance segmentation on the test set of VOC07 and validation sets of Cityscapes and COCO 2017, SemCL + +![](images/4464714652d0d386d11cba4348eea09574b3c6db395b35b868060c353a60eee7.jpg) + +![](images/3f932164755daac84f3361f1fbec4912e01ed15afe9034f869ae861a85472cbb.jpg) + +![](images/1eb08784a2efb5c367b9822502a620995202afc14e03ad7395adfa34a6525eb2.jpg) +Figure 4: Qualitative comparison of semantic segmentation between MAE [21] ViT-B (top) and SemCL Swin-B (bottom) on ADE20K semantic segmentation validation set. Both decoders are UPerNet[58]. + +![](images/f2f480a4cf239efbceb69cbd05d3d16e03ca52ff8605d6ace2cba415ac391135.jpg) + +models pretrained on the corresponding SemCL subset are compared with IN pretrained counterparts in Table 4. SemCL comprehensively outperforms its IN counterparts on the Cityscapes and COCO benchmarks. In particular, on the instance segmentation task, SemCL improves by a maximum of 0.5 and 0.3 points on Cityscapes and COCO respectively. However, SemCL performs worse than its IN counterparts in VOC detection, which is considered to be a result of insufficient training samples in the SemCL-VOC dataset and is further investigated in the ablation study (see Appendix A.4). + +
MethodVOC2007 +AP50Cityscapes†COCO
APboxAPmaskAPboxAPmask
MoCo v3[9] IN-1k81.8842.737.542.637.9
SemCL81.57(-0.31)42.9(+0.2)38.0(+0.5)42.8(+0.2)38.0(+0.1)
+ +(a) R50 + +
MethodVOC2007 +AP50Cityscapes†COCO
APboxAPmaskAPboxAPmask
Swin[37] IN-22k86.5845.939.849.243.0
SemCL86.53(-0.05)46.0(+0.1)40.2(+0.4)49.0(-0.2)43.3(+0.3)
+ +(b) Swin-T + +Table 4: Results of object detection and instance segmentation on various backbones w/ Cascade Mask R-CNN. All results are from our implementation. $\dagger$ : Sync BN is disabled due to gradient accumulation. + +System-level comparison. Table 5 compares the object detection and instance segmentation results of SemCL backbones with other models. On Cityscapes and COCO, SemCL backbones comprehensively outperform their twinborn MoCo Specifically, SemCL ResNet50 backbones achieve $+1.7$ over MoCo IG-1B [22] in COCO detection and + +
BackboneMethodVOC2012CityscapesADE20KCOCO
R50MoCo v3[9] IN-1k†79.6579.0142.7140.59
SemCL79.57(-0.08)79.13(+0.12)42.96(+0.25)40.80(+0.21)
R101timm[55] IN-1k81.8279.8744.4341.43
SemCL81.97(+0.15)80.10(+0.23)44.66(+0.23)41.58(+0.15)
Swin-TSwin[37] IN-22k80.9078.6744.6042.39
SemCL81.73(+0.83)78.91(+0.26)44.72(+0.12)42.53(+0.14)
Swin-SSwin[37] IN-22k83.7181.2547.6244.75
SemCL84.47(+0.76)81.53(+0.28)47.91(+0.29)44.57(-0.18)
Swin-BSwin[37] IN-22k84.6481.6148.2144.81
SemCL84.67(+0.03)81.94(+0.33)48.68(+0.47)45.47(+0.66)
+ +$\dagger$ Although MoCo v3 [9] is originally based on ViT, R50 pretrained models are provided in their repo https://github. com/facebookresearch/moco-v3. + +Table 2: SemCL vs. ImageNet pretrained backbones on VOC2012, Cityscapes, ADE20K and COCO semantic segmentation (mIoU). All results are from our implementation. + +
MethodBackbonemIoU
Deeplabv3+[7]R5077.2
Supervised-IN[22]R5074.4
MoCo IG-1B[22]R5075.5
BYOL[18]R5076.3
SemCLR5079.57
Deeplabv3+[7]R10178.85
Res2Net[15]Res2Net-10180.2
DFN[61]R10180.60
SemCLR10181.97
CP2[51]ViT-S/1679.5
SemCLSwin-S84.47
Leopart[70]ViT-B/876.3
SemCLSwin-B84.67
+ +(a) VOC2012 + +
MethodBackbonemIoU
MoCo IN-1M[22]R5072.5
MoCo IG-1B[22]R5073.6
Supervised-IN[22]R5074.6
CP2[51]R5076.5
SemCLR5079.13
Deeplabv3+[7]X-7179.55
SemCLR10180.10
Trans4Trans-T[62]PVTv2-B178.23
SemCLSwin-T78.91
CP2[51]ViT-S/1677.0
Trans4Trans-S[62]PVTv2-B280.02
SemCLSwin-S81.53
+ +(b) Cityscapes + +
MethodBackbonemIoU
CP2[51]R5039.2
PSPNet[66]R5042.78
SemCLR5042.96
PSPNet[66]R10143.29
SemCLR10144.66
CP2[51]ViT-S/1641.2
Seg-S/16[47]ViT-S45.37
SemCLSwin-S47.91
Seg-B/16[47]Vi-B/1648.06
MAE[21]ViT-B48.1
SemCLSwin-B48.68
+ +(c) ADE20K + +Table 3: Comparison with previous best results on validation sets of VOC2012, Cityscapes, and ADE20K semantic segmentation (mIoU). Best results are bolded. + ++5.1/+0.6 in Cityscapes/COCO instance segmentation tasks. For SemCL Swin-T backbones, SemCL achieves +1.2/+0.1 over Focal [60] on COCO detection/instance segmentation. Figure 5 demonstrates a qualitative comparison between Mask2Former [10] and SemCL Swin-T on Cityscapes validation set. With the improved spatial information understanding, SemCL model clearly discriminates (even overlapped) instances in Figures 5a and 5b and successfully classify part of an instance to a defined category as in Figure 5c. + +# 4.3. Depth Estimation + +Backbone comparison. SemCL pretrained backbones are compared with corresponding IN pretrained ones in Table 6. For cityscapes, SemCL comprehensively outperforms its IN pretrained counterparts. In particular, the ResNet50 backbones improve the RMSE by 0.138 points. For KITTI + +and NTUv2, all SemCL backbones show improvements over the IN pretrained ones. The SemCL pretrained SwinL model achieves improvements of 0.001/0.026 points on KITTI AbsRel/RMSE, and the SemCL pretrained ResNet50 model achieves improvements of 0.003/0.005 points on NYUv2. + +System-level comparison. We report the comparison of SemCL pretrained backbones with the previous state-of-the-art method Binsformer[35] and other works in Table 7. As for the Cityscapes results shown in Table 7a, SemCL pretrained ResNet50 significantly outperforms SDC [53] by $-0.089$ and $-2.218$ points on AbsRel and RMSE respectively. And SD-SSMDE [43] lags behind SemCL by $0.005 / 3.742$ points. Meanwhile, the test on KITTI (Eigen split [13]) is shown in Table 7b, SemCL pretrained ResNet50 and Swin-T comprehensively outperform the previous SOTA + +
MethodBackboneAP50
Supervised-IN[22]R5081.3
MoCo IN-1M[22]R5081.5
MoCo IG-1B[22]R5082.2
MoCo v2 800ep[8]R5082.5
SemCLR5081.57
UP-DETR/300[12]*80.1
OW-DETR[19]DDETR82.1
SemCLSwin-T86.53
+ +(a) Object det. on VOC2007 test. + +
MethodBackboneAPmask
MoCo IN-1M[22]R5032.3
MoCo IG-1B[22]R5032.9
Supervised-IN[22]R5032.9
Mask2Former[10]R5037.4
SemCLR5038.0
BoundaryFormer[30]*38.3
Mask2Former[10]Swin-T39.7
SemCLSwin-T40.2
+ +(b) Instance seg. on Cityscapes val. + +
MethodBackboneAPboxAPmask
Supervised-IN[22]R5040.636.8
MoCo IN-1M[22]R5040.836.9
MoCo IG-1B[22]R5041.137.4
SemCLR5042.838.0
MST[34]Swin-T42.738.8
VSA[63]Swin-T46.942.1
Focal[60]Focal-Base47.843.2
SemCLSwin-T49.043.3
+ +(c) Object det. and instance seg. on COCO val2017. + +Table 5: System-level comparison on object detection and instance segmentation. The best results are **bolded**. * claims their backbones are CNN-based, but are considered and compared as transformer-based networks, since transformer blocks are massively used. + +
BackboneMethodCityscapesKITTINYUv2
AbsRelRMSEAbsRelRMSEAbsRelRMSE
R50MoCo v3[9] IN-1k0.1384.8370.0592.3370.1200.405
SemCL0.1384.699(-0.138)0.058(-0.001)2.3370.117(-0.003)0.400(-0.005)
Swin-TSwin[37] IN-22k0.1344.6430.0572.2470.1140.386
SemCL0.133(-0.001)4.577(-0.066)0.0572.237(-0.010)0.112(-0.002)0.381(-0.005)
Swin-LSwin[37] IN-22k0.1304.5360.0532.1300.0950.329
SemCL0.128(-0.002)4.426(-0.110)0.052(-0.001)2.104(-0.026)0.094(-0.001)0.329
+ +Table 6: SemCL vs. ImageNet pretraining on Cityscapes, KITTI and NYUv2 depth estimation (lower is better). All results are from our implementation. + +![](images/4f828fa5596d72bf9b809e81a5743ab171f28b17213a00bec4ae5cf72cb1236d.jpg) + +![](images/9eebcdaf5d930380d32a1ce1228b380344b5f5c2f25b0450afc3010ebbb72469.jpg) + +![](images/d14dc188fc07519c4c2d193ac715f644034b4cf0ec9484d690735ac98a5ff5cd.jpg) +Figure 5: Qualitative comparison of instance segmentation between Mask2Former [10] (left) and SemCL (right) Swin-T on Cityscapes validation set. + +Binsformer. In particular, SemCL ResNet50 achieves a gain of $-0.003 / - 0.089$ over Binsformer, and SemCL Swin-L is on par with Binsformer Swin-L. On NYUv2 (see Table 7c), the gains of SemCL pretrained ResNet50 are also high with $-0.010$ AbsRel and $-0.173$ RMSE. The SemCL + +![](images/eb131c095b7ec1c7b169c6f45ea819e017f9722d5d9360e6ab2443f96f6cf025.jpg) +Figure 6: Qualitative comparison of depth estimation between Binsformer[35] and SemCL Swin-T on the NYUv2 test set. + +Swin-L RMSE is 0.001 points lower than the SOTA Binsformer. Qualitative comparison between Binsformer [35] and SemCL Swin-T is given in Figure 6. + +
MethodBackboneAbsRelRMSE
UpProj[29]R500.257[53]7.273[53]
Pad-net[59]R500.246[53]7.117[53]
TRL[64]R500.234[53]7.104[53]
Pilzer, et al.[45]-0.4405.443
SDC-Depth[53]R500.2276.917
DTS-Depth[27]MobileNetv20.1677.785
struct2depth[5]-0.1517.024
SD-SSMDE[43]R500.1438.441
SemCLR500.1384.699
+ +(a) Cityscapes + +
MethodBackboneAbsRelRMSE
EPCDepth [42]R500.0914.207
Binsformer[35]R500.0612.426
SemCLR500.0582.337
MonoViT[65]-0.0984.333
Binsformer [35]Swin-T0.0582.286
SemCLSwin-T0.0572.237
DepthFormer [33]Swin-L +R50CI0.0522.143
Binsformer[35]Swin-L0.0522.098
SemCLSwin-L0.0522.104
+ +(b) KITTI + +
MethodBackboneAbsRelRMSE
Pad-net [59]R500.2140.792
TRL [64]R500.1440.501
SDC-Depth [53]R500.1280.497
UpProj [29]R500.1270.573
SemCLR500.1170.400
Binsformer [35]Swin-T0.1130.379
SemCLSwin-T0.1120.381
Binsformer [35]Swin-L0.0940.330
SemCLSwin-L0.0940.329
+ +(c) NYUv2 + +Table 7: System-level comparison on depth estimation. Best results (lower is better) are bolded. + +# 4.4. Attention Map Visualization + +To improve the qualitative understanding of the SemCL representation, we visualize attention maps of Swin-T pretrained on SemCL-VOC and ImageNet using PyTorch library for CAM methods [16] with the Score-CAM method [52]. These samples in Figure 7 selected from VOC2012 val and have not been seen by the pretrained models are characterized by simple centered subjects (Figure 7a), complex scene with occlusion (Figure 7b) and multiple subjects (Figures 7c and 7d). + +In Figure 7a, the IN pretrained model failed to focus on the subject, while the SemCL model accurately depicted the area of the main subject. Meanwhile, SemCL performs better in complex scenes containing foreground-background relationships, as shown in Figure 7b. The IN pretrained model pays no attention to an occluded object, but the SemCL model recognizes that the occluded part also belongs to the horse, as shown in Figure 7b. In Figure 7c, the IN model directly ignores the distant person, but the SemCL model not only annotates two people but also focuses more on the nearby one, suggesting that the SemCL model can understand the spatial distribution of subjects in an image. Also, in Figure 7d, the SemCL model successfully recognizes one plant and its pot as one object. + +Summary. SemCL pretrained backbones can substantially outperform ImageNet pretrained counterparts on semantic segmentation, object detection, instance segmentation and depth estimation tasks. Using semantically contrastive pairs generated from off-the-shelf datasets, SemCL significantly improves the spatial information understanding of pretrained models. Such an improvement is also promising to benefit more challenging tasks including generative ones [32]. Ablation study concerning paired/unpaired InfoNCE loss, pretraining batch size, training length and dataset scale are given in Appendix A.4. + +# 5. Conclusion + +We investigate the effectiveness of semantic information in the contrastive learning pretext task on the spatial un + +![](images/5226e012f04effbc40caf2ee71320be0dd930e0f6780ee6e035ff270d3c00d18.jpg) + +![](images/bc20e95d32622b1d719688fb7165e8bcba5717039984cbeea7d857567e6845de.jpg) + +![](images/dde65f9160e7393015f2b97881d5058707c118ec74bc9b84b93e11c0565da30e.jpg) + +![](images/7919bb40da1c18c0f38cacca0e4346a87da9f4a3fa4da570a4ea3730d441e554.jpg) +Figure 7: Swin-T attention map visualization. Left: raw image. Middle: attention maps of the IN pretrained model. Right: attention maps of the SemCL pretrained model. + +![](images/1fc2ba11d96a5b524ff739c0e9eb982a41fd4562c0c91572ad37b10b6e55bb2b.jpg) + +![](images/693d5bde387383fc97b44205cca0feb880c9879a5708401aa62593fb1774048e.jpg) + +![](images/648068eebffd0c743bfea2d1146cabbd5569dc9fa89dc62364f507530818824e.jpg) + +![](images/231350e5e0a7098c25bc15d484f0c3ff1aca3dd2c956250a23babc7618d29816.jpg) + +![](images/537e899522c70cbcab38245ca5a0863705d05756a4dcac98dba417070839438f.jpg) + +![](images/a6caed9c7280655bf1a27b87c9a3c67be42a401e0db67034708997b7285fbc54.jpg) + +![](images/689106e8ce1df92ee8ea95243925d7e07effab326b0a18106a216c5342d7e681.jpg) + +![](images/4f2b7f5cdecb8ffc041152283ccb5171172d6cb5b0e25a425fada50073325794.jpg) + +derstanding ability of models. Tests are conducted on semantic/instance segmentation, object detection and depth estimation tasks with both ResNets and Swin Transformers, and SemCL pretrained backbones substantially outperform their ImageNet pretrained counterparts. By supporting small batch sizes and fast pretraining, SemCL is a lightweight yet effective approach. In contrastive representation learning, if dedicated pretext task is designed properly, four ounces can move a thousand pounds. We hope that SemCL will inspire more advanced pretext tasks for contrastive learning. + +# References + +[1] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019. 2 +[2] Suzanne Becker and Geoffrey E Hinton. Self-organizing neural network that discovers surfaces in random-dot stereograms. Nature, 355(6356):161-163, 1992. 2 +[3] Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah. Signature verification using a “siamese” time delay neural network. Advances in neural information processing systems, 6, 1993. 2 +[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Computer vision and pattern recognition (CVPR), 2018 IEEE conference on. IEEE, 2018. 3 +[5] Vincent Casser, Soeren Pirk, Reza Mahjourian, and Anelia Angelova. Unsupervised monocular depth and ego-motion learning with structure and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 8 +[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 3 +[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 2, 5, 6 +[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 7 +[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 1, 2, 5, 6, 7 +[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 2, 6, 7 +[11] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3 +[12] Zhigang Dai, Bolun Cai, Yugeng Lin, and Junying Chen. Updetr: Unsupervised pre-training for object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1601-1610, 2021. 7 +[13] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 6 + +[14] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html. 3 +[15] Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multiscale backbone architecture. IEEE transactions on pattern analysis and machine intelligence, 43(2):652-662, 2019. 6 +[16] Jacob Gildenblat and contributors. Pytorch library for cam methods. https://github.com/jacobgil/pytorch-grad-cam, 2021.8 +[17] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: TrainingImagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 5 +[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems, 33:21271-21284, 2020. 6 +[19] Akshita Gupta, Sanath Narayan, KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Ow-detr: Open-world detection transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9235-9244, 2022. 7 +[20] Bharath Hariharan, Pablo Arbelaez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In International Conference on Computer Vision (ICCV), 2011. 3 +[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 5, 6 +[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 2, 3, 5, 6, 7 +[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 +[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pages 630-645. Springer, 2016. 2 +[25] Olivier Henaff. Data-efficient image recognition with contrastive predictive coding. In International Conference on Machine Learning, pages 4182-4192. PMLR, 2020. 2 +[26] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In International Conference on Learning Representations, 2019. 2 + +[27] Hatem Ibrahim, Ahmed Salem, and Hyun-Soo Kang. Dts-depth: Real-time single-image depth estimation using depth-to-space image construction. Sensors, 22(5):1914, 2022. 8 +[28] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 2, 3 +[29] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In 2016 Fourth international conference on 3D vision (3DV), pages 239-248. IEEE, 2016. 8 +[30] Justin Lazarow, Weijian Xu, and Zhuowen Tu. Instance segmentation with mask-supervised polygonal boundary transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4382-4391, 2022. 7 +[31] Phuc H. Le-Khac, Graham Healy, and Alan F. Smeaton. Contrastive representation learning: A framework and review. IEEE Access, 8:193907-193934, 2020. 1, 2 +[32] Yuheng Li, Yijun Li, Jingwan Lu, Eli Shechtman, Yong Jae Lee, and Krishna Kumar Singh. Contrastive learning for diverse disentangled foreground generation. In European Conference on Computer Vision, pages 334-351. Springer, 2022. 8 +[33] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. arXiv preprint arXiv:2203.14211, 2022. 8 +[34] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, et al. Mst: Masked self-supervised transformer for visual representation. Advances in Neural Information Processing Systems, 34:13165-13176, 2021. 7 +[35] Zhenyu Li, Xuyang Wang, Xianming Liu, and Junjun Jiang. Binsformer: Revisiting adaptive bins for monocular depth estimation. arXiv preprint arXiv:2204.00987, 2022. 2, 6, 7, 8 +[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 3 +[37] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 4, 5, 6, 7 +[38] Vishnu Suresh Lokhande, Songwong Tasneeyapant, Abhay Venkatesh, Sathya N Ravi, and Vikas Singh. Generating accurate pseudo-labels in semi-supervised learning and avoiding overconfident predictions via hermite polynomial activations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11435-11443, 2020. 2 +[39] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, April + +24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017.5 +[40] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5 +[41] Chao Peng, Tete Xiao, Zeming Li, Yuning Jiang, Xiangyu Zhang, Kai Jia, Gang Yu, and Jian Sun. Megdet: A large minibatch object detector. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6181-6189, 2018. 5 +[42] Rui Peng, Ronggang Wang, Yawen Lai, Luyang Tang, and Yangang Cai. Excavating the potential capacity of self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15560-15569, 2021. 8 +[43] Andra Petrovai and Sergiu Nedevschi. Exploiting pseudo labels in a self-supervised learning framework for improved monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1578-1588, 2022. 6, 8 +[44] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V Le. Meta pseudo labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11557-11568, 2021. 2 +[45] Andrea Pilzer, Dan Xu, Mihai Puscas, Elisa Ricci, and Nicu Sebe. Unsupervised adversarial depth estimation using cycled generative networks. In 2018 international conference on 3D vision (3DV), pages 587-595. IEEE, 2018. 8 +[46] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, volume 11006, pages 369-386. SPIE, 2019. 5 +[47] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7262-7272, 2021. 6 +[48] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In European conference on computer vision, pages 776-794. Springer, 2020. 2 +[49] Aaron Van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv e-prints, pages arXiv-1807, 2018. 1, 2, 3 +[50] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021. 3 +[51] Feng Wang, Huiyu Wang, Chen Wei, Alan Yuille, and Wei Shen. Cp 2: Copy-paste contrastive pretraining for semantic segmentation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXX, pages 499-515. Springer, 2022. 5, 6 +[52] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 8 + +[53] Lijun Wang, Jianming Zhang, Oliver Wang, Zhe Lin, and Huchuan Lu. Sdc-depth: Semantic divide-and-conquer network for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 541-550, 2020. 6, 8 +[54] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7303-7313, 2021. 3 +[55] Ross Wightman, Hugo Touvron, and Hervé Jégou. Resnet strikes back: An improved training procedure in timm. arXiv preprint arXiv:2110.00476, 2021. 6 +[56] Hao Wu and Saurabh Prasad. Semi-supervised deep learning using pseudo labels for hyperspectral image classification. IEEE Transactions on Image Processing, 27(3):1259-1270, 2017. 2 +[57] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 1, 2, 3 +[58] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 5 +[59] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Padnet: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 675-684, 2018. 8 +[60] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal attention for long-range interactions in vision transformers. Advances in Neural Information Processing Systems, 34:30008-30022, 2021. 6, 7 +[61] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Learning a discriminative feature network for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1857-1866, 2018. 6 +[62] Jiaming Zhang, Kailun Yang, Angela Constantinescu, Kunyu Peng, Karin Müller, and Rainer Stiefelhagen. Trans4trans: Efficient transformer for transparent object and semantic scene segmentation in real-world navigation assistance. IEEE Transactions on Intelligent Transportation Systems, 23(10):19173-19186, 2022. 6 +[63] Qiming Zhang, Yufei Xu, Jing Zhang, and Dacheng Tao. Vsa: learning varied-size window attention in vision transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXV, pages 466-483. Springer, 2022. 7 +[64] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Zequn Jie, Xiang Li, and Jian Yang. Joint task-recursive learning for semantic segmentation and depth estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 235–251, 2018. 8 +[65] Chaoqiang Zhao, Youmin Zhang, Matteo Poggi, Fabio Tosi, Xianda Guo, Zheng Zhu, Guan Huang, Yang Tang, and + +Stefano Mattoccia. Monovit: Self-supervised monocular depth estimation with a vision transformer. arXiv preprint arXiv:2208.03543, 2022. 8 +[66] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 6 +[67] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 3 +[68] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 3 +[69] Chengxu Zhuang, Alex Lin Zhai, and Daniel Yamins. Local aggregation for unsupervised learning of visual embeddings. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6002-6012, 2019. 2 +[70] Adrian Ziegler and Yuki M Asano. Self-supervised learning of object parts for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14502-14511, 2022. 6 \ No newline at end of file diff --git a/2023/Semantic Information in Contrastive Learning/images.zip b/2023/Semantic Information in Contrastive Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9950f06c9f6a8dd5fd9409c6bb971e014f33485e --- /dev/null +++ b/2023/Semantic Information in Contrastive Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:483d0ebdbf1bba339d203bdc5f5cc8bae61ba0a9230ee118850c8a9e38fdb290 +size 863772 diff --git a/2023/Semantic Information in Contrastive Learning/layout.json b/2023/Semantic Information in Contrastive Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..68956559d7daa305f8d1c9df2167aba8b02e38a2 --- /dev/null +++ b/2023/Semantic Information in Contrastive Learning/layout.json @@ -0,0 +1,9548 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 152, + 103, + 442, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 103, + 442, + 121 + ], + "spans": [ + { + "bbox": [ + 152, + 103, + 442, + 121 + ], + "type": "text", + "content": "Semantic Information in Contrastive Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 142, + 485, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 142, + 485, + 215 + ], + "spans": [ + { + "bbox": [ + 108, + 142, + 485, + 215 + ], + "type": "text", + "content": "Shengjiang Quan" + }, + { + "bbox": [ + 108, + 142, + 485, + 215 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 108, + 142, + 485, + 215 + ], + "type": "text", + "content": ", Masahiro Hirano" + }, + { + "bbox": [ + 108, + 142, + 485, + 215 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 142, + 485, + 215 + ], + "type": "text", + "content": " and Yuji Yamakawa" + }, + { + "bbox": [ + 108, + 142, + 485, + 215 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 108, + 142, + 485, + 215 + ], + "type": "text", + "content": "\\\n" + }, + { + "bbox": [ + 108, + 142, + 485, + 215 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 142, + 485, + 215 + ], + "type": "text", + "content": "Graduate School of Engineering, The University of Tokyo, Japan\\\n" + }, + { + "bbox": [ + 108, + 142, + 485, + 215 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 142, + 485, + 215 + ], + "type": "text", + "content": "Institute of Industrial Science, The University of Tokyo, Japan\\\n" + }, + { + "bbox": [ + 108, + 142, + 485, + 215 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 108, + 142, + 485, + 215 + ], + "type": "text", + "content": "Interfaculty Initiative in Information Studies, The University of Tokyo, Japan\\\nquan@g.ecc.u-tokyo.ac.jp, {mhirano, y-ymkw}@iis.u-tokyo.ac.jp" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 266, + 290, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 459 + ], + "type": "text", + "content": "This work investigates the functionality of Semantic information in Contrastive Learning (SemCL). An advanced pretext task is designed: a contrast is performed between each object and its environment, taken from a scene. This allows the SemCL pretrained model to extract objects from their environment in an image, significantly improving the spatial understanding of the pretrained models. Downstream tasks of semantic/instance segmentation, object detection and depth estimation are implemented on PASCAL VOC, Cityscapes, COCO, KITTI, etc. SemCL pretrained models substantially outperform ImageNet pretrained counterparts and are competitive with well-known works on downstream tasks. The results suggest that a dedicated pretext task leveraging semantic information can be powerful in benchmarks related to spatial understanding. The code is available at https://github.com/sjiang95/semcl." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 481, + 128, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 481, + 128, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 481, + 128, + 495 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 502, + 288, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 288, + 609 + ], + "type": "text", + "content": "Within the field of visual representation learning, contrastive learning attracts special attention for its inspiring performance in transfer learning [57, 49, 31, 22, 8, 9]. The concept of contrastive learning can literally be explained as discovering the difference between positive and negative samples. The definition of positive-negative samples is one of the main subjects of contrastive learning, which directly determines the pretext task and the corresponding loss function." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 610, + 288, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 610, + 288, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 610, + 288, + 695 + ], + "type": "text", + "content": "We humans can recognize an object from its even complicated environment because we have learned and bound the typical geometric feature and the corresponding semantic concept of the object class. For example, we can recognize a cat in a scene thanks to our knowledge of its appearance (typical geometric feature) and the concept of cat (semantic information). Similarly, the target of prevalent" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 243, + 547, + 358 + ], + "blocks": [ + { + "bbox": [ + 309, + 243, + 547, + 358 + ], + "lines": [ + { + "bbox": [ + 309, + 243, + 547, + 358 + ], + "spans": [ + { + "bbox": [ + 309, + 243, + 547, + 358 + ], + "type": "image", + "image_path": "bab89fd9041f4a80e562caeb78e23e0262a1f6e2656bcaff36d7a6fdd19d7503.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 368, + 548, + 440 + ], + "lines": [ + { + "bbox": [ + 305, + 368, + 548, + 440 + ], + "spans": [ + { + "bbox": [ + 305, + 368, + 548, + 440 + ], + "type": "text", + "content": "Figure 1: Perceptual cognition of SemCL. Leveraging semantic information, SemCL separates an object from its environment. Each sample is divided into objects and surroundings and considered as a contrastive pair. The task is to maximize the distance between the object and its environment in the embedding space." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 462, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 676 + ], + "type": "text", + "content": "visual tasks, such as semantic/instance segmentation, object detection, etc., can be summarized as distinguishing and localizing subjects from surroundings. Inspired by and to extend the spatial recognition mechanism to CV, we propose our method SemCL which directly teaches models to extract a subject from its environment. Utilizing publicly off-the-shelf datasets providing semantic labels, SemCL can be considered as a supervised contrastive learning approach whose samples consist of subjects and their corresponding surroundings (semantically not the subject). The pretext task of SemCL is to tell the difference between one subject and its surroundings. From a representation learning perspective, the pretext task is to maximize the distance " + }, + { + "bbox": [ + 304, + 462, + 547, + 676 + ], + "type": "inline_equation", + "content": "(L^2" + }, + { + "bbox": [ + 304, + 462, + 547, + 676 + ], + "type": "text", + "content": " norm) between a subject and its environment in the embedding space (see Figure 1). The pretext task mimics the spatial recognition mechanism to discriminate objects and surroundings, which significantly improves the spatial information understanding of pretrained models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "In SemCL, contrast is performed pairwise between positive-negative pairs from one image/scene. Inter-scene contrast is considered inappropriate: only the contrasts" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5686" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "among subject-surrounding pairs from one scene are considered valid. Therefore, paired InfoNCE loss is adopted. With the mechanism of contrasting only paired samples, SemCL is characterized by decoupling the number of negative samples from the batch size. Compared to MoCo [22], whose performance is positively correlated with batch size and thus increases hardware cost (e.g. batchsize " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "= 4096" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " on 128 GPUs for ViT-B in MoCo v3 [9]) for satisfactory results, SemCL can achieve consistent yet competitive results on downstream tasks with relatively small batch size (e.g., batchsize " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "= 64 \\ @ 224 \\times 224" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "). We adopt the MoCo v3 [9] as the pretraining framework." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 289, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 361 + ], + "type": "text", + "content": "The primary motivation of representation learning is to pretrain general representations that can be fine-tuned and transferred to downstream tasks. In this work, the SemCL pretrained models are benchmarked on semantic segmentation, object detection and instance segmentation, and depth estimation tasks. SemCL models substantially outperform their ImageNet pretrained counterparts due to the improved ability of spacial information understanding. In a system-level comparison, SemCL models also show gains over previous well-known works (e.g. Deeplabv3+ [7] in semantic segmentation, Mask2Former [10] in instance segmentation and Binsformer [35] in depth estimation)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 373, + 134, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 134, + 385 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 134, + 385 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 394, + 170, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 170, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 170, + 407 + ], + "type": "text", + "content": "2.1. Contrastive Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 413, + 289, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 289, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 289, + 533 + ], + "type": "text", + "content": "The seminal idea of contrastive learning dates back at least to the 1990s [2, 3, 31], but it has become prominent in recent years thanks to the large pretrained models in the fields of NLP and CV [31]. As the name suggests, contrastive learning mines abstract representations by comparing between \"similar\" and \"dissimilar\" samples: minimizing the representation of \"similar\" samples and maximizing that of \"dissimilar\" samples in the embedding space. Formally, the \"similar\" inputs are called positive samples, and the \"dissimilar\" inputs are called negative samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "content": "Defining \"similar\" and \"dissimilar\" samples is an important topic in contrastive learning. In instance discrimination [57], which is a ubiquitous approach to define positive-negative samples in CV, each sample in a minibatch " + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "inline_equation", + "content": "x \\in \\{x_1, \\ldots, x_n\\}" + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "content": " is considered mutually exclusive to the rest. The encoded output " + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\theta(\\cdot)" + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "content": " of each sample is pushed away from the others in the embedding space, denoted as " + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\theta(x_1) \\Leftrightarrow \\{\\theta(x_2), \\ldots, \\theta(x_n)\\} \\Leftrightarrow \\text{push}" + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "content": ". Contrastive Predictive Coding (CPC) [49] defines positive-negative samples in a generative way: Given a sequential input " + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\{x_{t-3}, x_{t-2}, x_{t-1}, x_t\\}" + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "content": ", a context latent representation " + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "content": " is computed to predict the encoded outputs of future inputs " + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\{\\theta(\\mathbb{X}_{t+1}), \\theta(\\mathbb{X}_{t+2}), \\theta(\\mathbb{X}_{t+3}), \\ldots\\}" + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "content": ". Those predicted future embeddings are considered as positive samples to real future embeddings, and exclusive to some other" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 73, + 376, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 73, + 376, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 73, + 376, + 84 + ], + "type": "text", + "content": "irrelevant inputs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 91, + 427, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 427, + 102 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 427, + 102 + ], + "type": "text", + "content": "2.2. Momentum Contrast" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 110, + 547, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 547, + 266 + ], + "type": "text", + "content": "He et al. [22] proposed a momentum contrast framework, MoCo, for contrastive representation learning. MoCo summed unsupervised visual representation learning [57, 49, 26, 69, 25, 48, 1] up as dictionary look-up: the input data are represented and sampled by an encoder network " + }, + { + "bbox": [ + 304, + 110, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\theta_{k}" + }, + { + "bbox": [ + 304, + 110, + 547, + 266 + ], + "type": "text", + "content": " as \"keys\" " + }, + { + "bbox": [ + 304, + 110, + 547, + 266 + ], + "type": "inline_equation", + "content": "k = \\theta_{k}(x^{k})" + }, + { + "bbox": [ + 304, + 110, + 547, + 266 + ], + "type": "text", + "content": ". The goal of contrastive learning is to train encoders " + }, + { + "bbox": [ + 304, + 110, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\theta_{q}" + }, + { + "bbox": [ + 304, + 110, + 547, + 266 + ], + "type": "text", + "content": " by which a \"query\" " + }, + { + "bbox": [ + 304, + 110, + 547, + 266 + ], + "type": "inline_equation", + "content": "q = \\theta_{q}(x^{q})" + }, + { + "bbox": [ + 304, + 110, + 547, + 266 + ], + "type": "text", + "content": " should be similar to its matching key and dissimilar to others. Since dictionary consistency is considered crucial for unsupervised learning with contrastive loss [22], MoCo introduces momentum updating. In the MoCo framework, the key coder " + }, + { + "bbox": [ + 304, + 110, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\theta_{k}" + }, + { + "bbox": [ + 304, + 110, + 547, + 266 + ], + "type": "text", + "content": " is progressively updated by the weighted average of " + }, + { + "bbox": [ + 304, + 110, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\theta_{k}" + }, + { + "bbox": [ + 304, + 110, + 547, + 266 + ], + "type": "text", + "content": " and the query coder " + }, + { + "bbox": [ + 304, + 110, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\theta_{q}" + }, + { + "bbox": [ + 304, + 110, + 547, + 266 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 372, + 273, + 547, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 273, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 372, + 273, + 547, + 287 + ], + "type": "interline_equation", + "content": "\\theta_ {k} \\leftarrow m \\theta_ {k} + (1 - m) \\theta_ {q}, \\tag {1}", + "image_path": "511a5c45e82ca356ade6db28bc336a1dd29e72dd77c14510c4dfe0522b86a59c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 293, + 547, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 547, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 547, + 353 + ], + "type": "text", + "content": "where the momentum coefficient " + }, + { + "bbox": [ + 304, + 293, + 547, + 353 + ], + "type": "inline_equation", + "content": "m \\in [0,1)" + }, + { + "bbox": [ + 304, + 293, + 547, + 353 + ], + "type": "text", + "content": " is set quite high (" + }, + { + "bbox": [ + 304, + 293, + 547, + 353 + ], + "type": "inline_equation", + "content": "m = 0.999" + }, + { + "bbox": [ + 304, + 293, + 547, + 353 + ], + "type": "text", + "content": " in [22]) for slower key encoder update. The MoCo series has proven the power of the Momentum Contrast framework on feature transferring and is therefore employed as the basis for this work." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 359, + 529, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 359, + 529, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 359, + 529, + 372 + ], + "type": "text", + "content": "2.3. Labeled Data In Unsupervised Framework" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 377, + 547, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 547, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 547, + 496 + ], + "type": "text", + "content": "Large networks are quite thirsty for training data, while labelling the large amount of collected data is a human resource consuming challenge. Apart from turning to unsupervised learning, which inherently avoids using labelled data, using unlabeled data in supervised learning is also a solution. Pseudo-Label [44, 38, 56] is a trick that uses (usually supervised) pretrained networks to generate features that can play the role of labels. And the generated pseudo-labels participate in the iterations of the target network just like annotations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 497, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 548, + 714 + ], + "type": "text", + "content": "In contrast, Khosla et al. [28] also investigated the function of labelled data in unsupervised learning. As a contrastive learning research, Khosla et al. [28] intuitively uses the classification label to define samples within the same class are mutually positive, and those from other categories are negative. To exploit the label information, clusters of points belonging to the same class are pulled together in the embedding space, while clusters of samples from different classes are pushed apart [28]: " + }, + { + "bbox": [ + 304, + 497, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\left\\{\\theta\\left(x_1^{c0}\\right) \\rightleftharpoons \\theta\\left(x_2^{c0}\\right) \\rightleftharpoons \\theta\\left(x_{n}^{c0}\\right) \\rightleftharpoons \\theta\\left(x_n^0\\right)" + }, + { + "bbox": [ + 304, + 497, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\Leftrightarrow" + }, + { + "bbox": [ + 304, + 497, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\left\\{\\theta\\left(x_1^{c1}\\right) \\rightleftharpoons \\theta\\left(x_2^{c1}\\right) \\rightleftharpoons \\theta\\left(x_{m}^{c1}\\right)" + }, + { + "bbox": [ + 304, + 497, + 548, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 497, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\rightleftharpoons" + }, + { + "bbox": [ + 304, + 497, + 548, + 714 + ], + "type": "text", + "content": " for pull). The experimental result shows that supervised contrastive learning can improve both the accuracy and robustness of classifiers - [28] achieves a top-1 accuracy of " + }, + { + "bbox": [ + 304, + 497, + 548, + 714 + ], + "type": "inline_equation", + "content": "81.4\\%" + }, + { + "bbox": [ + 304, + 497, + 548, + 714 + ], + "type": "text", + "content": " on the ImageNet (IN) dataset, which is " + }, + { + "bbox": [ + 304, + 497, + 548, + 714 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 304, + 497, + 548, + 714 + ], + "type": "text", + "content": " above the best result reported by ResNet-200 architecture [24]. Likewise, our work investigates the effectiveness of semantic information in unsupervised framework contrastive learning." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5687" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 85 + ], + "type": "text", + "content": "2.4. Semantic Information in Contrastive Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 198 + ], + "type": "text", + "content": "Other than [28], there are works that use small amounts of labelled data in unsupervised representation learning. Wang et al. [54] proposed a pixel-wise contrastive algorithm for semantic segmentation: for a pixel " + }, + { + "bbox": [ + 46, + 90, + 287, + 198 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 90, + 287, + 198 + ], + "type": "text", + "content": ", it is pulled towards pixels belonging to the same category and pushed away from pixels in other categories in the embedding space. This contrast is performed across images from the same dataset. In the Cityscapes test benchmark, [54] R101 achieves 79.2 mIoU, which is 1.1 points better than Deeplabv3 [6]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 198, + 288, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 288, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 288, + 269 + ], + "type": "text", + "content": "The attempt and results of works using labels in unsupervised representation learning, such as [28, 54], indicate that label information can bring gains to a fully unsupervised representation. It is promising to explore advanced pretext tasks to improve the performance of pretrained models in downstream tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 279, + 103, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 279, + 103, + 292 + ], + "spans": [ + { + "bbox": [ + 47, + 279, + 103, + 292 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 298, + 143, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 298, + 143, + 311 + ], + "spans": [ + { + "bbox": [ + 47, + 298, + 143, + 311 + ], + "type": "text", + "content": "3.1. SemCL Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 317, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 287, + 413 + ], + "type": "text", + "content": "The core of this work is to define semantically contrastive pairs by utilizing off-the-shelf semantic labels, which are mappings between categories of objects and their coordinates on the raw image at the pixel level. As decided by the authors of a dataset, semantic labels can be color or grayscale images. Each predefined color or grayscale denotes a category, and the distribution of label pixels reflects the location of pixels belonging to a particular category in the raw image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 413, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 557 + ], + "type": "text", + "content": "The SemCL dataset is generated by the mechanism shown in Figure 2, where the raw image and its semantic label are selected from the PASCAL VOC2012 dataset [14]. From the semantic label annotating two objects - a motorcycle and a rider - two binary masks BM0 and BM1 can be extracted to separately represent two annotated objects. Their inverted counterparts " + }, + { + "bbox": [ + 46, + 413, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\neg" + }, + { + "bbox": [ + 46, + 413, + 287, + 557 + ], + "type": "text", + "content": "BM0 and " + }, + { + "bbox": [ + 46, + 413, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\neg" + }, + { + "bbox": [ + 46, + 413, + 287, + 557 + ], + "type": "text", + "content": "BM1 represent pixels other than the target object. These binary masks are then bitwise applied to the raw image to separate an object (Anchor) from its environment (" + }, + { + "bbox": [ + 46, + 413, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\neg" + }, + { + "bbox": [ + 46, + 413, + 287, + 557 + ], + "type": "text", + "content": "Anchor), forming a semantically contrastive pair. Each element in a raw Anchor-" + }, + { + "bbox": [ + 46, + 413, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\neg" + }, + { + "bbox": [ + 46, + 413, + 287, + 557 + ], + "type": "text", + "content": "Anchor pair is considered to be semantically opposite to the other." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 556, + 288, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 288, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 288, + 653 + ], + "type": "text", + "content": "In practice, the SemCL-1M dataset is produced from training sets of augmented PASCAL VOC2012 [14] (also known as SBD [20]), Cityscapes [11], ADE20K [67, 68] and COCO [36] (stuff+thing 2017 [4]) datasets. The components of the SemCL dataset are illustrated in Table 1. When generating the dataset, a threshold " + }, + { + "bbox": [ + 46, + 556, + 288, + 653 + ], + "type": "inline_equation", + "content": "t = 0.01" + }, + { + "bbox": [ + 46, + 556, + 288, + 653 + ], + "type": "text", + "content": " is set to filter out too small (less than " + }, + { + "bbox": [ + 46, + 556, + 288, + 653 + ], + "type": "inline_equation", + "content": "[\\sqrt{t} \\times \\text{height}] \\times [\\sqrt{t} \\times \\text{width}]" + }, + { + "bbox": [ + 46, + 556, + 288, + 653 + ], + "type": "text", + "content": ") objects that are considered semantically unsaturated." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 658, + 170, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 170, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 170, + 671 + ], + "type": "text", + "content": "3.2. Paired InfoNCE Loss" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "Since contrastive learning can be considered as a dictionary look-up task [22], it is assumed that within a set of keys " + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\{k_0, k_1, k_2, \\ldots\\}" + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": " of a dictionary, a query " + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": " matches only one" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 547, + 128 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 547, + 128 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 547, + 128 + ], + "type": "table", + "html": "
VOC2012CityscapesADE20KCOCOTotal
# training samples10,5822,97525,574118,287157,418
# SemCL pairs14,20320,280270,218712,2121,016,913
", + "image_path": "4f0e152aa9bdfd04cd415a97c9454b917d14019e511435d7455f721b3c307df8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 135, + 547, + 184 + ], + "lines": [ + { + "bbox": [ + 304, + 135, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 304, + 135, + 547, + 184 + ], + "type": "text", + "content": "Table 1: Components of the SemCL dataset. The SemCL datasets spawned from VOC2012(aug), Cityscapes, ADE20k and COCO 2017 are named SemCL-" + }, + { + "bbox": [ + 304, + 135, + 547, + 184 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{VOC}" + }, + { + "bbox": [ + 304, + 135, + 547, + 184 + ], + "type": "text", + "content": ", City, ADE, COCO\\} respectively." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 205, + 547, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 547, + 266 + ], + "type": "text", + "content": "key " + }, + { + "bbox": [ + 304, + 205, + 547, + 266 + ], + "type": "inline_equation", + "content": "k_{+}" + }, + { + "bbox": [ + 304, + 205, + 547, + 266 + ], + "type": "text", + "content": " of the dictionary [22]. The value of the objective function should be low if " + }, + { + "bbox": [ + 304, + 205, + 547, + 266 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 205, + 547, + 266 + ], + "type": "text", + "content": " is similar to its positive key " + }, + { + "bbox": [ + 304, + 205, + 547, + 266 + ], + "type": "inline_equation", + "content": "k_{+}" + }, + { + "bbox": [ + 304, + 205, + 547, + 266 + ], + "type": "text", + "content": " and dissimilar to all negative keys " + }, + { + "bbox": [ + 304, + 205, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\{k_{0}, k_{1}, k_{2}, \\ldots\\} \\setminus \\{k_{+}\\}" + }, + { + "bbox": [ + 304, + 205, + 547, + 266 + ], + "type": "text", + "content": ". In MoCo [22], the form of a contrastive loss function InfoNCE [49] is adopted" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 358, + 274, + 546, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 274, + 546, + 304 + ], + "spans": [ + { + "bbox": [ + 358, + 274, + 546, + 304 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {q} = - \\log \\frac {\\exp (q \\cdot k _ {+} / \\tau)}{\\sum_ {i = 0} ^ {K} \\exp (q \\cdot k _ {i} / \\tau)} \\tag {2}", + "image_path": "eeac554b469d1ab0dfe2a0a499173d80c38fafa3e0ddbca78e58c8d5b473031d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "text", + "content": " is a scalar temperature hyperparameter controlling the strength of penalties on hard negative samples [50], e.g. in the instance discrimination task [49, 22], there may be potential positive samples in the randomly chosen negative samples. Summing over one positive and " + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "text", + "content": " negative samples, Equation (2) is the log loss of a " + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "inline_equation", + "content": "(\\mathbf{K} + 1)" + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "text", + "content": "-way softmax-based classifier trying to classify " + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "inline_equation", + "content": "k_{+}" + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "text", + "content": " [22]. While in our case an anchor instance " + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "text", + "content": " compares only with its own " + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "text", + "content": " augmentations (positive) " + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "inline_equation", + "content": "K^{+} := \\{k_{1}^{+}, k_{2}^{+}, k_{3}^{+}, \\ldots, k_{n}^{+}\\}" + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "text", + "content": " and paired " + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "inline_equation", + "content": "n + 1" + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "text", + "content": " negatives " + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "inline_equation", + "content": "K^{-} := \\{k_{0}^{-}, k_{1}^{-}, k_{2}^{-}, k_{3}^{-}, \\ldots, k_{n}^{-}\\}" + }, + { + "bbox": [ + 304, + 311, + 547, + 443 + ], + "type": "text", + "content": " using paired InfoNCE" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 453, + 552, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 552, + 495 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 552, + 495 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {q} = - \\sum_ {j = t + 1} ^ {n} \\log \\frac {\\exp \\left(q _ {t} \\cdot k _ {j} ^ {+} / \\tau\\right)}{\\sum_ {i = 0} ^ {K ^ {-}} \\exp \\left(q _ {t} \\cdot k _ {i} ^ {-} / \\tau\\right) + \\exp \\left(q _ {t} \\cdot k _ {j} ^ {+} / \\tau\\right)}. \\tag {3}", + "image_path": "3e7dc0b76d854ccf398496d121fcd1eafbdee0118817d78d0cc684fefcf18ce2.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 496, + 547, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 547, + 639 + ], + "type": "text", + "content": "For one anchor as query " + }, + { + "bbox": [ + 304, + 496, + 547, + 639 + ], + "type": "inline_equation", + "content": "q_{0}(t = 0)" + }, + { + "bbox": [ + 304, + 496, + 547, + 639 + ], + "type": "text", + "content": ", the paired InfoNCE loss tries to classify it as " + }, + { + "bbox": [ + 304, + 496, + 547, + 639 + ], + "type": "inline_equation", + "content": "k_{j}^{+}, j \\in [1, n]" + }, + { + "bbox": [ + 304, + 496, + 547, + 639 + ], + "type": "text", + "content": ". The next step is to go forward: classify the first positive " + }, + { + "bbox": [ + 304, + 496, + 547, + 639 + ], + "type": "inline_equation", + "content": "q_{1}(t = 1)" + }, + { + "bbox": [ + 304, + 496, + 547, + 639 + ], + "type": "text", + "content": " of the anchor as " + }, + { + "bbox": [ + 304, + 496, + 547, + 639 + ], + "type": "inline_equation", + "content": "k_{j}^{+}, j \\in [2, n]" + }, + { + "bbox": [ + 304, + 496, + 547, + 639 + ], + "type": "text", + "content": ", and so on, as shown in the algorithm 1, Appendix A.1. To contrast with the paired InfoNCE, we refer to the original InfoNCE loss as unpaired InfoNCE hereafter. The behavior difference between unpaired and paired InfoNCE losses is compared in Figure 3. The paired InfoNCE only maximizes the distance between anchors and -anchors in the embedding space, neglecting irrelevant samples. The implementation of the paired InfoNCE loss is introduced in detail in Appendix A.2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 647, + 388, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 388, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 388, + 658 + ], + "type": "text", + "content": "3.3. Pretext Task" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Since the aim of this work is to enable networks to strip targets (anchors) from their surroundings (" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\neg" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "anchors), we use a restricted instance discrimination task [57]: a query and a key are positive pairs if they are different views (by" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5688" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 545, + 266 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 266 + ], + "type": "image", + "image_path": "94bcab4c34e62920d5c29550c8ba28bbd49c686ee525a78e3d55635c96dac62c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 272, + 547, + 309 + ], + "lines": [ + { + "bbox": [ + 46, + 272, + 547, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 547, + 309 + ], + "type": "text", + "content": "Figure 2: SemCL dataset. The semantic label of a sample is utilized to generate binary masks. Each binary mask depicts the pixel-level information of one object (anchor) in the raw image. Those binary masks, together with their inverts, are applied on the raw image to extract objects (anchors) and corresponding surroundings (" + }, + { + "bbox": [ + 46, + 272, + 547, + 309 + ], + "type": "inline_equation", + "content": "\\neg" + }, + { + "bbox": [ + 46, + 272, + 547, + 309 + ], + "type": "text", + "content": "anchors)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 321, + 318, + 536 + ], + "blocks": [ + { + "bbox": [ + 47, + 321, + 318, + 536 + ], + "lines": [ + { + "bbox": [ + 47, + 321, + 318, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 321, + 318, + 536 + ], + "type": "image", + "image_path": "f00a92c7ece71a2863a0c55f5a16a3c78bc0764ebe9e51e37957ba554ae0cf66.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 544, + 547, + 581 + ], + "lines": [ + { + "bbox": [ + 46, + 544, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 547, + 581 + ], + "type": "text", + "content": "Figure 3: Comparison between unpaired (left) and paired InfoNCE losses (right). Left: a sample (anchor) is positive only to its own augmentations, while negative to others in the same mini-batch. Right: A sample (anchor) is positive only to its own augmentations, while negative only to its paired counterparts." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 320, + 321, + 547, + 536 + ], + "blocks": [ + { + "bbox": [ + 320, + 321, + 547, + 536 + ], + "lines": [ + { + "bbox": [ + 320, + 321, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 320, + 321, + 547, + 536 + ], + "type": "image", + "image_path": "562f206f886369d73ce424b9fecb105a9fbdc964e9552b00571342f39e1ab72a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 601, + 289, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 289, + 710 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 289, + 710 + ], + "type": "text", + "content": "data augmentation) of the same anchor, and are negative pairs if they originate from an anchor and its corresponding surroundings (" + }, + { + "bbox": [ + 46, + 601, + 289, + 710 + ], + "type": "inline_equation", + "content": "\\neg" + }, + { + "bbox": [ + 46, + 601, + 289, + 710 + ], + "type": "text", + "content": "anchors). A PyTorch-style pseudocode is given in the Appendix A.2. The inputs are encoded by base encoder " + }, + { + "bbox": [ + 46, + 601, + 289, + 710 + ], + "type": "inline_equation", + "content": "f_{b}" + }, + { + "bbox": [ + 46, + 601, + 289, + 710 + ], + "type": "text", + "content": " and momentum encoder " + }, + { + "bbox": [ + 46, + 601, + 289, + 710 + ], + "type": "inline_equation", + "content": "f_{m}" + }, + { + "bbox": [ + 46, + 601, + 289, + 710 + ], + "type": "text", + "content": ", producing queries and keys. Theoretically, the backbone of " + }, + { + "bbox": [ + 46, + 601, + 289, + 710 + ], + "type": "inline_equation", + "content": "f_{b}" + }, + { + "bbox": [ + 46, + 601, + 289, + 710 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 601, + 289, + 710 + ], + "type": "inline_equation", + "content": "f_{m}" + }, + { + "bbox": [ + 46, + 601, + 289, + 710 + ], + "type": "text", + "content": " can be any architecture. In practice, we use ResNet [23] and Swin Transformer [37] because they are the most representative architectures in recent years. All keys are stacked along" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 601, + 459, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 601, + 459, + 613 + ], + "spans": [ + { + "bbox": [ + 306, + 601, + 459, + 613 + ], + "type": "text", + "content": "dimension 1 for paired InfoNCE loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 631, + 388, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 631, + 388, + 644 + ], + "spans": [ + { + "bbox": [ + 306, + 631, + 388, + 644 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "content": "Considering the enormous scale difference between ImageNet and SemCL dataset, all SemCL models are initialized by corresponding IN pretrained weights, which are compared as counterparts in downstream tasks. The total number of pretraining iterations is set to " + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "inline_equation", + "content": "30k" + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "content": " to minimize the significant" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5689" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "type": "text", + "content": "variation in training time caused by scale differences among SemCL sub-datasets. We adopt AdamW [40] with " + }, + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "type": "inline_equation", + "content": "wd = 0.1" + }, + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "type": "text", + "content": ", and the OneCycle [46] learning rate schedule consisting of warmup [17] for the initial 3, 750 (warmup ratio 0.125) steps and cosine annealing [9, 39]. Following [9], the base " + }, + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "type": "inline_equation", + "content": "lr" + }, + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "type": "inline_equation", + "content": "1.5 \\times 10^{-4}" + }, + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "type": "text", + "content": " and scaled linearly by " + }, + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "type": "inline_equation", + "content": "lr \\times batchsize / 256" + }, + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "type": "text", + "content": " and temperature " + }, + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "type": "inline_equation", + "content": "\\tau = 0.2" + }, + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "type": "text", + "content": ". The initial MoCo momentum is 0.99 and is gradually increased to 1 with a half cycle cosine schedule. The data augmentation strategy includes " + }, + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "type": "text", + "content": " random crops after random rescaling the original image with an area ratio in the range [0.08, 1.0], followed by random color jitter, random grayscale, random Gaussian blur, and random horizontal flip. All batch norm (BN) layers are synchronized across GPUs (SyncBN [41]). Limited by the two dual-GPU nodes (RTX3090 " + }, + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "type": "text", + "content": " 2+Titan RTX " + }, + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "type": "text", + "content": " 2) we use, a native batch size of 64 can be applied to all ResNets and up to Swin-B for Swin Transformers with crop size " + }, + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 47, + 72, + 289, + 310 + ], + "type": "text", + "content": ". For Swin-L in particular, we compromise by using a batch size of 32 with gradient accumulation over 2 iterations to achieve an equivalent batch size of 64." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 312, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 289, + 396 + ], + "type": "text", + "content": "SemCL aims to improve the understanding of spatial information of models for which downstream tasks of semantic segmentation, object detection and depth estimation are conducted. Unless otherwise stated, all SemCL backbones are pretrained on the corresponding SemCL sub-datasets. The implementation details of downstream tasks are listed in Appendix A.3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 403, + 180, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 180, + 417 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 180, + 417 + ], + "type": "text", + "content": "4.1. Semantic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 422, + 289, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 289, + 530 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 289, + 530 + ], + "type": "text", + "content": "Backbone comparison. In the semantic segmentation benchmark on validation sets of VOC2012, Cityscapes, ADE20K and COCO 2017, SemCL models pretrained on the corresponding SemCL sub dataset are compared with ImageNet pretrained counterparts in Table 2. For VOC2012 and COCO, SemCL outperforms its IN counterparts by a maximum of 0.83 and 0.66 points respectively. SemCL can comprehensively outperform its IN counterparts on all backbones in the Cityscapes and ADE20k benchmarks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 531, + 289, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 289, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 289, + 651 + ], + "type": "text", + "content": "System-level comparison. SemCL pretrained backbones are compared with other methods on semantic segmentation tasks in Table 3. On VOC2012, SemCL R50/R101 outperform Deeplabv3+ [7] by 2.37/3.12 points. The Deeplabv3+ [7] X-71 is 0.55 points behind SemCL R101 on Cityscapes. For " + }, + { + "bbox": [ + 46, + 531, + 289, + 651 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}^2" + }, + { + "bbox": [ + 46, + 531, + 289, + 651 + ], + "type": "text", + "content": " [51] ViT-S/16, SemCL Swin-S significantly outperforms it on all benchmarks by 4.97, 4.53 and 6.71 points. Compared to MAE [21] ViT-B, SemCL Swin-B achieves a gain of 0.58 points. A qualitative comparison between MAE [21] ViT-B and SemCL Swin-B is shown in Figure 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 658, + 279, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 279, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 279, + 672 + ], + "type": "text", + "content": "4.2. Object Detection and Instance Segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 289, + 713 + ], + "type": "text", + "content": "Backbone comparison. In the benchmark for object detection and instance segmentation on the test set of VOC07 and validation sets of Cityscapes and COCO 2017, SemCL" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 80, + 449, + 150 + ], + "blocks": [ + { + "bbox": [ + 309, + 80, + 449, + 150 + ], + "lines": [ + { + "bbox": [ + 309, + 80, + 449, + 150 + ], + "spans": [ + { + "bbox": [ + 309, + 80, + 449, + 150 + ], + "type": "image", + "image_path": "4464714652d0d386d11cba4348eea09574b3c6db395b35b868060c353a60eee7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 449, + 80, + 544, + 149 + ], + "blocks": [ + { + "bbox": [ + 449, + 80, + 544, + 149 + ], + "lines": [ + { + "bbox": [ + 449, + 80, + 544, + 149 + ], + "spans": [ + { + "bbox": [ + 449, + 80, + 544, + 149 + ], + "type": "image", + "image_path": "3f932164755daac84f3361f1fbec4912e01ed15afe9034f869ae861a85472cbb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 152, + 449, + 223 + ], + "blocks": [ + { + "bbox": [ + 308, + 152, + 449, + 223 + ], + "lines": [ + { + "bbox": [ + 308, + 152, + 449, + 223 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 449, + 223 + ], + "type": "image", + "image_path": "1eb08784a2efb5c367b9822502a620995202afc14e03ad7395adfa34a6525eb2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 232, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 304, + 232, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 547, + 281 + ], + "type": "text", + "content": "Figure 4: Qualitative comparison of semantic segmentation between MAE [21] ViT-B (top) and SemCL Swin-B (bottom) on ADE20K semantic segmentation validation set. Both decoders are UPerNet[58]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 449, + 152, + 544, + 223 + ], + "blocks": [ + { + "bbox": [ + 449, + 152, + 544, + 223 + ], + "lines": [ + { + "bbox": [ + 449, + 152, + 544, + 223 + ], + "spans": [ + { + "bbox": [ + 449, + 152, + 544, + 223 + ], + "type": "image", + "image_path": "f2f480a4cf239efbceb69cbd05d3d16e03ca52ff8605d6ace2cba415ac391135.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 300, + 547, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 547, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 547, + 432 + ], + "type": "text", + "content": "models pretrained on the corresponding SemCL subset are compared with IN pretrained counterparts in Table 4. SemCL comprehensively outperforms its IN counterparts on the Cityscapes and COCO benchmarks. In particular, on the instance segmentation task, SemCL improves by a maximum of 0.5 and 0.3 points on Cityscapes and COCO respectively. However, SemCL performs worse than its IN counterparts in VOC detection, which is considered to be a result of insufficient training samples in the SemCL-VOC dataset and is further investigated in the ablation study (see Appendix A.4)." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 309, + 441, + 548, + 487 + ], + "blocks": [ + { + "bbox": [ + 309, + 441, + 548, + 487 + ], + "lines": [ + { + "bbox": [ + 309, + 441, + 548, + 487 + ], + "spans": [ + { + "bbox": [ + 309, + 441, + 548, + 487 + ], + "type": "table", + "html": "
MethodVOC2007 \nAP50Cityscapes†COCO
APboxAPmaskAPboxAPmask
MoCo v3[9] IN-1k81.8842.737.542.637.9
SemCL81.57(-0.31)42.9(+0.2)38.0(+0.5)42.8(+0.2)38.0(+0.1)
", + "image_path": "881e02686d5b8419b9404f054eedae8382a4634bf805d80c8a92204695daeeaa.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 309, + 510, + 548, + 556 + ], + "blocks": [ + { + "bbox": [ + 415, + 490, + 441, + 499 + ], + "lines": [ + { + "bbox": [ + 415, + 490, + 441, + 499 + ], + "spans": [ + { + "bbox": [ + 415, + 490, + 441, + 499 + ], + "type": "text", + "content": "(a) R50" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 510, + 548, + 556 + ], + "lines": [ + { + "bbox": [ + 309, + 510, + 548, + 556 + ], + "spans": [ + { + "bbox": [ + 309, + 510, + 548, + 556 + ], + "type": "table", + "html": "
MethodVOC2007 \nAP50Cityscapes†COCO
APboxAPmaskAPboxAPmask
Swin[37] IN-22k86.5845.939.849.243.0
SemCL86.53(-0.05)46.0(+0.1)40.2(+0.4)49.0(-0.2)43.3(+0.3)
", + "image_path": "b66851922523b6f50ca535f239e5dc7ddc52139143ad1dda8bb753c92e07334c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 410, + 559, + 447, + 569 + ], + "lines": [ + { + "bbox": [ + 410, + 559, + 447, + 569 + ], + "spans": [ + { + "bbox": [ + 410, + 559, + 447, + 569 + ], + "type": "text", + "content": "(b) Swin-T" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 578, + 547, + 627 + ], + "lines": [ + { + "bbox": [ + 304, + 578, + 547, + 627 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 547, + 627 + ], + "type": "text", + "content": "Table 4: Results of object detection and instance segmentation on various backbones w/ Cascade Mask R-CNN. All results are from our implementation. " + }, + { + "bbox": [ + 304, + 578, + 547, + 627 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 304, + 578, + 547, + 627 + ], + "type": "text", + "content": ": Sync BN is disabled due to gradient accumulation." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 641, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 548, + 713 + ], + "type": "text", + "content": "System-level comparison. Table 5 compares the object detection and instance segmentation results of SemCL backbones with other models. On Cityscapes and COCO, SemCL backbones comprehensively outperform their twinborn MoCo Specifically, SemCL ResNet50 backbones achieve " + }, + { + "bbox": [ + 304, + 641, + 548, + 713 + ], + "type": "inline_equation", + "content": "+1.7" + }, + { + "bbox": [ + 304, + 641, + 548, + 713 + ], + "type": "text", + "content": " over MoCo IG-1B [22] in COCO detection and" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5690" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 91, + 70, + 504, + 235 + ], + "blocks": [ + { + "bbox": [ + 91, + 70, + 504, + 235 + ], + "lines": [ + { + "bbox": [ + 91, + 70, + 504, + 235 + ], + "spans": [ + { + "bbox": [ + 91, + 70, + 504, + 235 + ], + "type": "table", + "html": "
BackboneMethodVOC2012CityscapesADE20KCOCO
R50MoCo v3[9] IN-1k†79.6579.0142.7140.59
SemCL79.57(-0.08)79.13(+0.12)42.96(+0.25)40.80(+0.21)
R101timm[55] IN-1k81.8279.8744.4341.43
SemCL81.97(+0.15)80.10(+0.23)44.66(+0.23)41.58(+0.15)
Swin-TSwin[37] IN-22k80.9078.6744.6042.39
SemCL81.73(+0.83)78.91(+0.26)44.72(+0.12)42.53(+0.14)
Swin-SSwin[37] IN-22k83.7181.2547.6244.75
SemCL84.47(+0.76)81.53(+0.28)47.91(+0.29)44.57(-0.18)
Swin-BSwin[37] IN-22k84.6481.6148.2144.81
SemCL84.67(+0.03)81.94(+0.33)48.68(+0.47)45.47(+0.66)
", + "image_path": "082d3671feab936b337929b5aef38900265e6cd95b0e1b788214a2974c3c3f21.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 97, + 236, + 504, + 255 + ], + "lines": [ + { + "bbox": [ + 97, + 236, + 504, + 255 + ], + "spans": [ + { + "bbox": [ + 97, + 236, + 504, + 255 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 97, + 236, + 504, + 255 + ], + "type": "text", + "content": " Although MoCo v3 [9] is originally based on ViT, R50 pretrained models are provided in their repo https://github. com/facebookresearch/moco-v3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 91, + 308, + 227, + 452 + ], + "blocks": [ + { + "bbox": [ + 45, + 262, + 547, + 288 + ], + "lines": [ + { + "bbox": [ + 45, + 262, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 45, + 262, + 547, + 288 + ], + "type": "text", + "content": "Table 2: SemCL vs. ImageNet pretrained backbones on VOC2012, Cityscapes, ADE20K and COCO semantic segmentation (mIoU). All results are from our implementation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 91, + 308, + 227, + 452 + ], + "lines": [ + { + "bbox": [ + 91, + 308, + 227, + 452 + ], + "spans": [ + { + "bbox": [ + 91, + 308, + 227, + 452 + ], + "type": "table", + "html": "
MethodBackbonemIoU
Deeplabv3+[7]R5077.2
Supervised-IN[22]R5074.4
MoCo IG-1B[22]R5075.5
BYOL[18]R5076.3
SemCLR5079.57
Deeplabv3+[7]R10178.85
Res2Net[15]Res2Net-10180.2
DFN[61]R10180.60
SemCLR10181.97
CP2[51]ViT-S/1679.5
SemCLSwin-S84.47
Leopart[70]ViT-B/876.3
SemCLSwin-B84.67
", + "image_path": "a88f6f47b2805d265dad4e1af595c06f4e49492eed648c230cdfb35669b82287.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 136, + 454, + 182, + 463 + ], + "lines": [ + { + "bbox": [ + 136, + 454, + 182, + 463 + ], + "spans": [ + { + "bbox": [ + 136, + 454, + 182, + 463 + ], + "type": "text", + "content": "(a) VOC2012" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 230, + 308, + 372, + 452 + ], + "blocks": [ + { + "bbox": [ + 230, + 308, + 372, + 452 + ], + "lines": [ + { + "bbox": [ + 230, + 308, + 372, + 452 + ], + "spans": [ + { + "bbox": [ + 230, + 308, + 372, + 452 + ], + "type": "table", + "html": "
MethodBackbonemIoU
MoCo IN-1M[22]R5072.5
MoCo IG-1B[22]R5073.6
Supervised-IN[22]R5074.6
CP2[51]R5076.5
SemCLR5079.13
Deeplabv3+[7]X-7179.55
SemCLR10180.10
Trans4Trans-T[62]PVTv2-B178.23
SemCLSwin-T78.91
CP2[51]ViT-S/1677.0
Trans4Trans-S[62]PVTv2-B280.02
SemCLSwin-S81.53
", + "image_path": "7ea98b2dbfb2b38ce71183e887d7da52cb8af53865505b61f735ea0865ed291f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 277, + 454, + 324, + 464 + ], + "lines": [ + { + "bbox": [ + 277, + 454, + 324, + 464 + ], + "spans": [ + { + "bbox": [ + 277, + 454, + 324, + 464 + ], + "type": "text", + "content": "(b) Cityscapes" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 374, + 308, + 504, + 452 + ], + "blocks": [ + { + "bbox": [ + 374, + 308, + 504, + 452 + ], + "lines": [ + { + "bbox": [ + 374, + 308, + 504, + 452 + ], + "spans": [ + { + "bbox": [ + 374, + 308, + 504, + 452 + ], + "type": "table", + "html": "
MethodBackbonemIoU
CP2[51]R5039.2
PSPNet[66]R5042.78
SemCLR5042.96
PSPNet[66]R10143.29
SemCLR10144.66
CP2[51]ViT-S/1641.2
Seg-S/16[47]ViT-S45.37
SemCLSwin-S47.91
Seg-B/16[47]Vi-B/1648.06
MAE[21]ViT-B48.1
SemCLSwin-B48.68
", + "image_path": "109acbebfc1947ec79c70f17f2f219f2b04e15e3050ce38334797cd51d81bc36.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 417, + 454, + 460, + 463 + ], + "lines": [ + { + "bbox": [ + 417, + 454, + 460, + 463 + ], + "spans": [ + { + "bbox": [ + 417, + 454, + 460, + 463 + ], + "type": "text", + "content": "(c) ADE20K" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 473, + 548, + 498 + ], + "lines": [ + { + "bbox": [ + 45, + 473, + 548, + 498 + ], + "spans": [ + { + "bbox": [ + 45, + 473, + 548, + 498 + ], + "type": "text", + "content": "Table 3: Comparison with previous best results on validation sets of VOC2012, Cityscapes, and ADE20K semantic segmentation (mIoU). Best results are bolded." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 45, + 518, + 288, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 518, + 288, + 627 + ], + "spans": [ + { + "bbox": [ + 45, + 518, + 288, + 627 + ], + "type": "text", + "content": "+5.1/+0.6 in Cityscapes/COCO instance segmentation tasks. For SemCL Swin-T backbones, SemCL achieves +1.2/+0.1 over Focal [60] on COCO detection/instance segmentation. Figure 5 demonstrates a qualitative comparison between Mask2Former [10] and SemCL Swin-T on Cityscapes validation set. With the improved spatial information understanding, SemCL model clearly discriminates (even overlapped) instances in Figures 5a and 5b and successfully classify part of an instance to a defined category as in Figure 5c." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 635, + 153, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 153, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 153, + 647 + ], + "type": "text", + "content": "4.3. Depth Estimation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": "Backbone comparison. SemCL pretrained backbones are compared with corresponding IN pretrained ones in Table 6. For cityscapes, SemCL comprehensively outperforms its IN pretrained counterparts. In particular, the ResNet50 backbones improve the RMSE by 0.138 points. For KITTI" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 518, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 547, + 590 + ], + "type": "text", + "content": "and NTUv2, all SemCL backbones show improvements over the IN pretrained ones. The SemCL pretrained SwinL model achieves improvements of 0.001/0.026 points on KITTI AbsRel/RMSE, and the SemCL pretrained ResNet50 model achieves improvements of 0.003/0.005 points on NYUv2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "content": "System-level comparison. We report the comparison of SemCL pretrained backbones with the previous state-of-the-art method Binsformer[35] and other works in Table 7. As for the Cityscapes results shown in Table 7a, SemCL pretrained ResNet50 significantly outperforms SDC [53] by " + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "inline_equation", + "content": "-0.089" + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "inline_equation", + "content": "-2.218" + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "content": " points on AbsRel and RMSE respectively. And SD-SSMDE [43] lags behind SemCL by " + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "inline_equation", + "content": "0.005 / 3.742" + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "content": " points. Meanwhile, the test on KITTI (Eigen split [13]) is shown in Table 7b, SemCL pretrained ResNet50 and Swin-T comprehensively outperform the previous SOTA" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "5691" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 79, + 203, + 178 + ], + "blocks": [ + { + "bbox": [ + 65, + 79, + 203, + 178 + ], + "lines": [ + { + "bbox": [ + 65, + 79, + 203, + 178 + ], + "spans": [ + { + "bbox": [ + 65, + 79, + 203, + 178 + ], + "type": "table", + "html": "
MethodBackboneAP50
Supervised-IN[22]R5081.3
MoCo IN-1M[22]R5081.5
MoCo IG-1B[22]R5082.2
MoCo v2 800ep[8]R5082.5
SemCLR5081.57
UP-DETR/300[12]*80.1
OW-DETR[19]DDETR82.1
SemCLSwin-T86.53
", + "image_path": "f0aea2d67442d3453f0fbb5112f6eacee395552c46628977e1c37de8d71aa79e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 80, + 180, + 188, + 190 + ], + "lines": [ + { + "bbox": [ + 80, + 180, + 188, + 190 + ], + "spans": [ + { + "bbox": [ + 80, + 180, + 188, + 190 + ], + "type": "text", + "content": "(a) Object det. on VOC2007 test." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 205, + 79, + 354, + 178 + ], + "blocks": [ + { + "bbox": [ + 205, + 79, + 354, + 178 + ], + "lines": [ + { + "bbox": [ + 205, + 79, + 354, + 178 + ], + "spans": [ + { + "bbox": [ + 205, + 79, + 354, + 178 + ], + "type": "table", + "html": "
MethodBackboneAPmask
MoCo IN-1M[22]R5032.3
MoCo IG-1B[22]R5032.9
Supervised-IN[22]R5032.9
Mask2Former[10]R5037.4
SemCLR5038.0
BoundaryFormer[30]*38.3
Mask2Former[10]Swin-T39.7
SemCLSwin-T40.2
", + "image_path": "205c13fa2f3d33f22861b55e37ffc24e67b41f95f70bbaef87e043681cae870d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 221, + 180, + 337, + 190 + ], + "lines": [ + { + "bbox": [ + 221, + 180, + 337, + 190 + ], + "spans": [ + { + "bbox": [ + 221, + 180, + 337, + 190 + ], + "type": "text", + "content": "(b) Instance seg. on Cityscapes val." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 357, + 79, + 529, + 178 + ], + "blocks": [ + { + "bbox": [ + 357, + 79, + 529, + 178 + ], + "lines": [ + { + "bbox": [ + 357, + 79, + 529, + 178 + ], + "spans": [ + { + "bbox": [ + 357, + 79, + 529, + 178 + ], + "type": "table", + "html": "
MethodBackboneAPboxAPmask
Supervised-IN[22]R5040.636.8
MoCo IN-1M[22]R5040.836.9
MoCo IG-1B[22]R5041.137.4
SemCLR5042.838.0
MST[34]Swin-T42.738.8
VSA[63]Swin-T46.942.1
Focal[60]Focal-Base47.843.2
SemCLSwin-T49.043.3
", + "image_path": "b0115554c9ab6aba34934dba7a61a1063d74f807831ed54f6be7bf3d84103743.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 358, + 180, + 526, + 190 + ], + "lines": [ + { + "bbox": [ + 358, + 180, + 526, + 190 + ], + "spans": [ + { + "bbox": [ + 358, + 180, + 526, + 190 + ], + "type": "text", + "content": "(c) Object det. and instance seg. on COCO val2017." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 48, + 247, + 548, + 353 + ], + "blocks": [ + { + "bbox": [ + 46, + 200, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 200, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 200, + 547, + 236 + ], + "type": "text", + "content": "Table 5: System-level comparison on object detection and instance segmentation. The best results are **bolded**. * claims their backbones are CNN-based, but are considered and compared as transformer-based networks, since transformer blocks are massively used." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 247, + 548, + 353 + ], + "lines": [ + { + "bbox": [ + 48, + 247, + 548, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 247, + 548, + 353 + ], + "type": "table", + "html": "
BackboneMethodCityscapesKITTINYUv2
AbsRelRMSEAbsRelRMSEAbsRelRMSE
R50MoCo v3[9] IN-1k0.1384.8370.0592.3370.1200.405
SemCL0.1384.699(-0.138)0.058(-0.001)2.3370.117(-0.003)0.400(-0.005)
Swin-TSwin[37] IN-22k0.1344.6430.0572.2470.1140.386
SemCL0.133(-0.001)4.577(-0.066)0.0572.237(-0.010)0.112(-0.002)0.381(-0.005)
Swin-LSwin[37] IN-22k0.1304.5360.0532.1300.0950.329
SemCL0.128(-0.002)4.426(-0.110)0.052(-0.001)2.104(-0.026)0.094(-0.001)0.329
", + "image_path": "8bf50fb1d870c899a12d71afc831c2a9dc3a804607e544327d9a9c0fa27a4ca5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 359, + 546, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 359, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 546, + 384 + ], + "type": "text", + "content": "Table 6: SemCL vs. ImageNet pretraining on Cityscapes, KITTI and NYUv2 depth estimation (lower is better). All results are from our implementation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 49, + 403, + 286, + 461 + ], + "blocks": [ + { + "bbox": [ + 49, + 403, + 286, + 461 + ], + "lines": [ + { + "bbox": [ + 49, + 403, + 286, + 461 + ], + "spans": [ + { + "bbox": [ + 49, + 403, + 286, + 461 + ], + "type": "image", + "image_path": "4f828fa5596d72bf9b809e81a5743ab171f28b17213a00bec4ae5cf72cb1236d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 49, + 462, + 286, + 518 + ], + "blocks": [ + { + "bbox": [ + 49, + 462, + 286, + 518 + ], + "lines": [ + { + "bbox": [ + 49, + 462, + 286, + 518 + ], + "spans": [ + { + "bbox": [ + 49, + 462, + 286, + 518 + ], + "type": "image", + "image_path": "9eebcdaf5d930380d32a1ce1228b380344b5f5c2f25b0450afc3010ebbb72469.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 49, + 519, + 286, + 578 + ], + "blocks": [ + { + "bbox": [ + 49, + 519, + 286, + 578 + ], + "lines": [ + { + "bbox": [ + 49, + 519, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 49, + 519, + 286, + 578 + ], + "type": "image", + "image_path": "d14dc188fc07519c4c2d193ac715f644034b4cf0ec9484d690735ac98a5ff5cd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 586, + 287, + 622 + ], + "lines": [ + { + "bbox": [ + 46, + 586, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 287, + 622 + ], + "type": "text", + "content": "Figure 5: Qualitative comparison of instance segmentation between Mask2Former [10] (left) and SemCL (right) Swin-T on Cityscapes validation set." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": "Binsformer. In particular, SemCL ResNet50 achieves a gain of " + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "inline_equation", + "content": "-0.003 / - 0.089" + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": " over Binsformer, and SemCL Swin-L is on par with Binsformer Swin-L. On NYUv2 (see Table 7c), the gains of SemCL pretrained ResNet50 are also high with " + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "inline_equation", + "content": "-0.010" + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": " AbsRel and " + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "inline_equation", + "content": "-0.173" + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": " RMSE. The SemCL" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 306, + 413, + 545, + 609 + ], + "blocks": [ + { + "bbox": [ + 306, + 413, + 545, + 609 + ], + "lines": [ + { + "bbox": [ + 306, + 413, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 413, + 545, + 609 + ], + "type": "image", + "image_path": "eb131c095b7ec1c7b169c6f45ea819e017f9722d5d9360e6ab2443f96f6cf025.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 617, + 547, + 652 + ], + "lines": [ + { + "bbox": [ + 305, + 617, + 547, + 652 + ], + "spans": [ + { + "bbox": [ + 305, + 617, + 547, + 652 + ], + "type": "text", + "content": "Figure 6: Qualitative comparison of depth estimation between Binsformer[35] and SemCL Swin-T on the NYUv2 test set." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "Swin-L RMSE is 0.001 points lower than the SOTA Binsformer. Qualitative comparison between Binsformer [35] and SemCL Swin-T is given in Figure 6." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5692" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 79, + 243, + 178 + ], + "blocks": [ + { + "bbox": [ + 61, + 79, + 243, + 178 + ], + "lines": [ + { + "bbox": [ + 61, + 79, + 243, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 79, + 243, + 178 + ], + "type": "table", + "html": "
MethodBackboneAbsRelRMSE
UpProj[29]R500.257[53]7.273[53]
Pad-net[59]R500.246[53]7.117[53]
TRL[64]R500.234[53]7.104[53]
Pilzer, et al.[45]-0.4405.443
SDC-Depth[53]R500.2276.917
DTS-Depth[27]MobileNetv20.1677.785
struct2depth[5]-0.1517.024
SD-SSMDE[43]R500.1438.441
SemCLR500.1384.699
", + "image_path": "4a2a00e6430d21ed3b2c4f48b4c6933169b22d91c026c758bf0265e2e6acd86b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 245, + 79, + 385, + 178 + ], + "blocks": [ + { + "bbox": [ + 128, + 180, + 175, + 190 + ], + "lines": [ + { + "bbox": [ + 128, + 180, + 175, + 190 + ], + "spans": [ + { + "bbox": [ + 128, + 180, + 175, + 190 + ], + "type": "text", + "content": "(a) Cityscapes" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 245, + 79, + 385, + 178 + ], + "lines": [ + { + "bbox": [ + 245, + 79, + 385, + 178 + ], + "spans": [ + { + "bbox": [ + 245, + 79, + 385, + 178 + ], + "type": "table", + "html": "
MethodBackboneAbsRelRMSE
EPCDepth [42]R500.0914.207
Binsformer[35]R500.0612.426
SemCLR500.0582.337
MonoViT[65]-0.0984.333
Binsformer [35]Swin-T0.0582.286
SemCLSwin-T0.0572.237
DepthFormer [33]Swin-L +R50CI0.0522.143
Binsformer[35]Swin-L0.0522.098
SemCLSwin-L0.0522.104
", + "image_path": "8351eebc395c62ca76c3574320eaf60e2e6803d09c41d05afa5ff76716febdd8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 387, + 79, + 534, + 178 + ], + "blocks": [ + { + "bbox": [ + 298, + 180, + 331, + 190 + ], + "lines": [ + { + "bbox": [ + 298, + 180, + 331, + 190 + ], + "spans": [ + { + "bbox": [ + 298, + 180, + 331, + 190 + ], + "type": "text", + "content": "(b) KITTI" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 387, + 79, + 534, + 178 + ], + "lines": [ + { + "bbox": [ + 387, + 79, + 534, + 178 + ], + "spans": [ + { + "bbox": [ + 387, + 79, + 534, + 178 + ], + "type": "table", + "html": "
MethodBackboneAbsRelRMSE
Pad-net [59]R500.2140.792
TRL [64]R500.1440.501
SDC-Depth [53]R500.1280.497
UpProj [29]R500.1270.573
SemCLR500.1170.400
Binsformer [35]Swin-T0.1130.379
SemCLSwin-T0.1120.381
Binsformer [35]Swin-L0.0940.330
SemCLSwin-L0.0940.329
", + "image_path": "71e30f17f65e81b2c4e177c6ae6216589702ba758817f108caa56853fec9b7b7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 441, + 180, + 478, + 190 + ], + "lines": [ + { + "bbox": [ + 441, + 180, + 478, + 190 + ], + "spans": [ + { + "bbox": [ + 441, + 180, + 478, + 190 + ], + "type": "text", + "content": "(c) NYUv2" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 102, + 200, + 488, + 212 + ], + "lines": [ + { + "bbox": [ + 102, + 200, + 488, + 212 + ], + "spans": [ + { + "bbox": [ + 102, + 200, + 488, + 212 + ], + "type": "text", + "content": "Table 7: System-level comparison on depth estimation. Best results (lower is better) are bolded." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 233, + 202, + 246 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 233, + 202, + 246 + ], + "spans": [ + { + "bbox": [ + 47, + 233, + 202, + 246 + ], + "type": "text", + "content": "4.4. Attention Map Visualization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "text", + "content": "To improve the qualitative understanding of the SemCL representation, we visualize attention maps of Swin-T pretrained on SemCL-VOC and ImageNet using PyTorch library for CAM methods [16] with the Score-CAM method [52]. These samples in Figure 7 selected from VOC2012 val and have not been seen by the pretrained models are characterized by simple centered subjects (Figure 7a), complex scene with occlusion (Figure 7b) and multiple subjects (Figures 7c and 7d)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 359, + 287, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 359, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 287, + 526 + ], + "type": "text", + "content": "In Figure 7a, the IN pretrained model failed to focus on the subject, while the SemCL model accurately depicted the area of the main subject. Meanwhile, SemCL performs better in complex scenes containing foreground-background relationships, as shown in Figure 7b. The IN pretrained model pays no attention to an occluded object, but the SemCL model recognizes that the occluded part also belongs to the horse, as shown in Figure 7b. In Figure 7c, the IN model directly ignores the distant person, but the SemCL model not only annotates two people but also focuses more on the nearby one, suggesting that the SemCL model can understand the spatial distribution of subjects in an image. Also, in Figure 7d, the SemCL model successfully recognizes one plant and its pot as one object." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 527, + 288, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 527, + 288, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 527, + 288, + 659 + ], + "type": "text", + "content": "Summary. SemCL pretrained backbones can substantially outperform ImageNet pretrained counterparts on semantic segmentation, object detection, instance segmentation and depth estimation tasks. Using semantically contrastive pairs generated from off-the-shelf datasets, SemCL significantly improves the spatial information understanding of pretrained models. Such an improvement is also promising to benefit more challenging tasks including generative ones [32]. Ablation study concerning paired/unpaired InfoNCE loss, pretraining batch size, training length and dataset scale are given in Appendix A.4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "content": "We investigate the effectiveness of semantic information in the contrastive learning pretext task on the spatial un" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 232, + 391, + 304 + ], + "blocks": [ + { + "bbox": [ + 307, + 232, + 391, + 304 + ], + "lines": [ + { + "bbox": [ + 307, + 232, + 391, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 391, + 304 + ], + "type": "image", + "image_path": "5226e012f04effbc40caf2ee71320be0dd930e0f6780ee6e035ff270d3c00d18.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 304, + 391, + 376 + ], + "blocks": [ + { + "bbox": [ + 307, + 304, + 391, + 376 + ], + "lines": [ + { + "bbox": [ + 307, + 304, + 391, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 391, + 376 + ], + "type": "image", + "image_path": "bc20e95d32622b1d719688fb7165e8bcba5717039984cbeea7d857567e6845de.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 376, + 391, + 447 + ], + "blocks": [ + { + "bbox": [ + 307, + 376, + 391, + 447 + ], + "lines": [ + { + "bbox": [ + 307, + 376, + 391, + 447 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 391, + 447 + ], + "type": "image", + "image_path": "dde65f9160e7393015f2b97881d5058707c118ec74bc9b84b93e11c0565da30e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 307, + 448, + 391, + 519 + ], + "blocks": [ + { + "bbox": [ + 307, + 448, + 391, + 519 + ], + "lines": [ + { + "bbox": [ + 307, + 448, + 391, + 519 + ], + "spans": [ + { + "bbox": [ + 307, + 448, + 391, + 519 + ], + "type": "image", + "image_path": "7919bb40da1c18c0f38cacca0e4346a87da9f4a3fa4da570a4ea3730d441e554.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 529, + 547, + 564 + ], + "lines": [ + { + "bbox": [ + 305, + 529, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 305, + 529, + 547, + 564 + ], + "type": "text", + "content": "Figure 7: Swin-T attention map visualization. Left: raw image. Middle: attention maps of the IN pretrained model. Right: attention maps of the SemCL pretrained model." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 395, + 232, + 468, + 304 + ], + "blocks": [ + { + "bbox": [ + 395, + 232, + 468, + 304 + ], + "lines": [ + { + "bbox": [ + 395, + 232, + 468, + 304 + ], + "spans": [ + { + "bbox": [ + 395, + 232, + 468, + 304 + ], + "type": "image", + "image_path": "1fc2ba11d96a5b524ff739c0e9eb982a41fd4562c0c91572ad37b10b6e55bb2b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 395, + 304, + 468, + 375 + ], + "blocks": [ + { + "bbox": [ + 395, + 304, + 468, + 375 + ], + "lines": [ + { + "bbox": [ + 395, + 304, + 468, + 375 + ], + "spans": [ + { + "bbox": [ + 395, + 304, + 468, + 375 + ], + "type": "image", + "image_path": "693d5bde387383fc97b44205cca0feb880c9879a5708401aa62593fb1774048e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 395, + 375, + 468, + 447 + ], + "blocks": [ + { + "bbox": [ + 395, + 375, + 468, + 447 + ], + "lines": [ + { + "bbox": [ + 395, + 375, + 468, + 447 + ], + "spans": [ + { + "bbox": [ + 395, + 375, + 468, + 447 + ], + "type": "image", + "image_path": "648068eebffd0c743bfea2d1146cabbd5569dc9fa89dc62364f507530818824e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 395, + 447, + 468, + 519 + ], + "blocks": [ + { + "bbox": [ + 395, + 447, + 468, + 519 + ], + "lines": [ + { + "bbox": [ + 395, + 447, + 468, + 519 + ], + "spans": [ + { + "bbox": [ + 395, + 447, + 468, + 519 + ], + "type": "image", + "image_path": "231350e5e0a7098c25bc15d484f0c3ff1aca3dd2c956250a23babc7618d29816.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 473, + 232, + 545, + 304 + ], + "blocks": [ + { + "bbox": [ + 473, + 232, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 473, + 232, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 473, + 232, + 545, + 304 + ], + "type": "image", + "image_path": "537e899522c70cbcab38245ca5a0863705d05756a4dcac98dba417070839438f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 473, + 304, + 545, + 376 + ], + "blocks": [ + { + "bbox": [ + 473, + 304, + 545, + 376 + ], + "lines": [ + { + "bbox": [ + 473, + 304, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 473, + 304, + 545, + 376 + ], + "type": "image", + "image_path": "a6caed9c7280655bf1a27b87c9a3c67be42a401e0db67034708997b7285fbc54.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 473, + 376, + 545, + 447 + ], + "blocks": [ + { + "bbox": [ + 473, + 376, + 545, + 447 + ], + "lines": [ + { + "bbox": [ + 473, + 376, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 473, + 376, + 545, + 447 + ], + "type": "image", + "image_path": "689106e8ce1df92ee8ea95243925d7e07effab326b0a18106a216c5342d7e681.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 473, + 447, + 545, + 519 + ], + "blocks": [ + { + "bbox": [ + 473, + 447, + 545, + 519 + ], + "lines": [ + { + "bbox": [ + 473, + 447, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 473, + 447, + 545, + 519 + ], + "type": "image", + "image_path": "4f2b7f5cdecb8ffc041152283ccb5171172d6cb5b0e25a425fada50073325794.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 586, + 547, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 547, + 705 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 547, + 705 + ], + "type": "text", + "content": "derstanding ability of models. Tests are conducted on semantic/instance segmentation, object detection and depth estimation tasks with both ResNets and Swin Transformers, and SemCL pretrained backbones substantially outperform their ImageNet pretrained counterparts. By supporting small batch sizes and fast pretraining, SemCL is a lightweight yet effective approach. In contrastive representation learning, if dedicated pretext task is designed properly, four ounces can move a thousand pounds. We hope that SemCL will inspire more advanced pretext tasks for contrastive learning." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5693" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 167 + ], + "type": "text", + "content": "[2] Suzanne Becker and Geoffrey E Hinton. Self-organizing neural network that discovers surfaces in random-dot stereograms. Nature, 355(6356):161-163, 1992. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "type": "text", + "content": "[3] Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah. Signature verification using a “siamese” time delay neural network. Advances in neural information processing systems, 6, 1993. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 288, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 256 + ], + "type": "text", + "content": "[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Computer vision and pattern recognition (CVPR), 2018 IEEE conference on. IEEE, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 288, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 313 + ], + "type": "text", + "content": "[5] Vincent Casser, Soeren Pirk, Reza Mahjourian, and Anelia Angelova. Unsupervised monocular depth and ego-motion learning with structure and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 314, + 288, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 288, + 356 + ], + "type": "text", + "content": "[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 357, + 287, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 412 + ], + "type": "text", + "content": "[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 413, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 288, + 446 + ], + "type": "text", + "content": "[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "text", + "content": "[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 288, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 288, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 288, + 546 + ], + "type": "text", + "content": "[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 547, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 288, + 612 + ], + "type": "text", + "content": "[11] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "type": "text", + "content": "[12] Zhigang Dai, Bolun Cai, Yugeng Lin, and Junying Chen. Updetr: Unsupervised pre-training for object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1601-1610, 2021. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "type": "text", + "content": "[13] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 126 + ], + "type": "text", + "content": "[14] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 173 + ], + "type": "text", + "content": "[15] Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multiscale backbone architecture. IEEE transactions on pattern analysis and machine intelligence, 43(2):652-662, 2019. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 547, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 207 + ], + "type": "text", + "content": "[16] Jacob Gildenblat and contributors. Pytorch library for cam methods. https://github.com/jacobgil/pytorch-grad-cam, 2021.8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 208, + 547, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 547, + 252 + ], + "type": "text", + "content": "[17] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: TrainingImagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 253, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 547, + 319 + ], + "type": "text", + "content": "[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems, 33:21271-21284, 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 320, + 547, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 547, + 374 + ], + "type": "text", + "content": "[19] Akshita Gupta, Sanath Narayan, KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Ow-detr: Open-world detection transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9235-9244, 2022. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 376, + 547, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 547, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 547, + 420 + ], + "type": "text", + "content": "[20] Bharath Hariharan, Pablo Arbelaez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In International Conference on Computer Vision (ICCV), 2011. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 421, + 547, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 547, + 475 + ], + "type": "text", + "content": "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 477, + 547, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 477, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 477, + 547, + 532 + ], + "type": "text", + "content": "[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 533, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 547, + 578 + ], + "type": "text", + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 578, + 547, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 547, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 547, + 621 + ], + "type": "text", + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pages 630-645. Springer, 2016. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 624, + 547, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 547, + 657 + ], + "type": "text", + "content": "[25] Olivier Henaff. Data-efficient image recognition with contrastive predictive coding. In International Conference on Machine Learning, pages 4182-4192. PMLR, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "type": "text", + "content": "[26] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In International Conference on Learning Representations, 2019. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5694" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "type": "text", + "content": "[27] Hatem Ibrahim, Ahmed Salem, and Hyun-Soo Kang. Dts-depth: Real-time single-image depth estimation using depth-to-space image construction. Sensors, 22(5):1914, 2022. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "type": "text", + "content": "[28] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "text", + "content": "[29] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In 2016 Fourth international conference on 3D vision (3DV), pages 239-248. IEEE, 2016. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 274 + ], + "type": "text", + "content": "[30] Justin Lazarow, Weijian Xu, and Zhuowen Tu. Instance segmentation with mask-supervised polygonal boundary transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4382-4391, 2022. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 288, + 308 + ], + "type": "text", + "content": "[31] Phuc H. Le-Khac, Graham Healy, and Alan F. Smeaton. Contrastive representation learning: A framework and review. IEEE Access, 8:193907-193934, 2020. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 310, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 288, + 363 + ], + "type": "text", + "content": "[32] Yuheng Li, Yijun Li, Jingwan Lu, Eli Shechtman, Yong Jae Lee, and Krishna Kumar Singh. Contrastive learning for diverse disentangled foreground generation. In European Conference on Computer Vision, pages 334-351. Springer, 2022. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 288, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 288, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 288, + 410 + ], + "type": "text", + "content": "[33] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. arXiv preprint arXiv:2203.14211, 2022. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 288, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 288, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 288, + 465 + ], + "type": "text", + "content": "[34] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, et al. Mst: Masked self-supervised transformer for visual representation. Advances in Neural Information Processing Systems, 34:13165-13176, 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 468, + 288, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 288, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 288, + 499 + ], + "type": "text", + "content": "[35] Zhenyu Li, Xuyang Wang, Xianming Liu, and Junjun Jiang. Binsformer: Revisiting adaptive bins for monocular depth estimation. arXiv preprint arXiv:2204.00987, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "type": "text", + "content": "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "type": "text", + "content": "[37] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 4, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 678 + ], + "type": "text", + "content": "[38] Vishnu Suresh Lokhande, Songwong Tasneeyapant, Abhay Venkatesh, Sathya N Ravi, and Vikas Singh. Generating accurate pseudo-labels in semi-supervised learning and avoiding overconfident predictions via hermite polynomial activations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11435-11443, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "text", + "content": "[39] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, April" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 325, + 73, + 546, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 73, + 546, + 94 + ], + "spans": [ + { + "bbox": [ + 325, + 73, + 546, + 94 + ], + "type": "text", + "content": "24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017.5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 547, + 128 + ], + "type": "text", + "content": "[40] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 129, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 547, + 183 + ], + "type": "text", + "content": "[41] Chao Peng, Tete Xiao, Zeming Li, Yuning Jiang, Xiangyu Zhang, Kai Jia, Gang Yu, and Jian Sun. Megdet: A large minibatch object detector. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6181-6189, 2018. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 185, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 547, + 239 + ], + "type": "text", + "content": "[42] Rui Peng, Ronggang Wang, Yawen Lai, Luyang Tang, and Yangang Cai. Excavating the potential capacity of self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15560-15569, 2021. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 240, + 546, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 240, + 546, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 240, + 546, + 293 + ], + "type": "text", + "content": "[43] Andra Petrovai and Sergiu Nedevschi. Exploiting pseudo labels in a self-supervised learning framework for improved monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1578-1588, 2022. 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 295, + 546, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 295, + 546, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 295, + 546, + 337 + ], + "type": "text", + "content": "[44] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V Le. Meta pseudo labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11557-11568, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 338, + 546, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 546, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 546, + 381 + ], + "type": "text", + "content": "[45] Andrea Pilzer, Dan Xu, Mihai Puscas, Elisa Ricci, and Nicu Sebe. Unsupervised adversarial depth estimation using cycled generative networks. In 2018 international conference on 3D vision (3DV), pages 587-595. IEEE, 2018. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 384, + 546, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 384, + 546, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 546, + 436 + ], + "type": "text", + "content": "[46] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, volume 11006, pages 369-386. SPIE, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 438, + 546, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 546, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 546, + 481 + ], + "type": "text", + "content": "[47] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7262-7272, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 483, + 546, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 483, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 483, + 546, + 514 + ], + "type": "text", + "content": "[48] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In European conference on computer vision, pages 776-794. Springer, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 516, + 546, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 516, + 546, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 516, + 546, + 548 + ], + "type": "text", + "content": "[49] Aaron Van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv e-prints, pages arXiv-1807, 2018. 1, 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 549, + 546, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 549, + 546, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 549, + 546, + 591 + ], + "type": "text", + "content": "[50] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 593, + 546, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 546, + 647 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 546, + 647 + ], + "type": "text", + "content": "[51] Feng Wang, Huiyu Wang, Chen Wei, Alan Yuille, and Wei Shen. Cp 2: Copy-paste contrastive pretraining for semantic segmentation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXX, pages 499-515. Springer, 2022. 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 648, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 648, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 648, + 546, + 712 + ], + "type": "text", + "content": "[52] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5695" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[53] Lijun Wang, Jianming Zhang, Oliver Wang, Zhe Lin, and Huchuan Lu. Sdc-depth: Semantic divide-and-conquer network for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 541-550, 2020. 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "text", + "content": "[54] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7303-7313, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 216 + ], + "type": "text", + "content": "[55] Ross Wightman, Hugo Touvron, and Hervé Jégou. Resnet strikes back: An improved training procedure in timm. arXiv preprint arXiv:2110.00476, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 258 + ], + "type": "text", + "content": "[56] Hao Wu and Saurabh Prasad. Semi-supervised deep learning using pseudo labels for hyperspectral image classification. IEEE Transactions on Image Processing, 27(3):1259-1270, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 288, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 315 + ], + "type": "text", + "content": "[57] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "type": "text", + "content": "[58] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 361, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 288, + 415 + ], + "type": "text", + "content": "[59] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Padnet: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 675-684, 2018. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 459 + ], + "type": "text", + "content": "[60] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal attention for long-range interactions in vision transformers. Advances in Neural Information Processing Systems, 34:30008-30022, 2021. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 288, + 514 + ], + "type": "text", + "content": "[61] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Learning a discriminative feature network for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1857-1866, 2018. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 288, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 580 + ], + "type": "text", + "content": "[62] Jiaming Zhang, Kailun Yang, Angela Constantinescu, Kunyu Peng, Karin Müller, and Rainer Stiefelhagen. Trans4trans: Efficient transformer for transparent object and semantic scene segmentation in real-world navigation assistance. IEEE Transactions on Intelligent Transportation Systems, 23(10):19173-19186, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 581, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 635 + ], + "type": "text", + "content": "[63] Qiming Zhang, Yufei Xu, Jing Zhang, and Dacheng Tao. Vsa: learning varied-size window attention in vision transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXV, pages 466-483. Springer, 2022. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "type": "text", + "content": "[64] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Zequn Jie, Xiang Li, and Jian Yang. Joint task-recursive learning for semantic segmentation and depth estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 235–251, 2018. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[65] Chaoqiang Zhao, Youmin Zhang, Matteo Poggi, Fabio Tosi, Xianda Guo, Zheng Zhu, Guan Huang, Yang Tang, and" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 330 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 325, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 325, + 73, + 545, + 106 + ], + "type": "text", + "content": "Stefano Mattoccia. Monovit: Self-supervised monocular depth estimation with a vision transformer. arXiv preprint arXiv:2208.03543, 2022. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "text", + "content": "[66] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 152, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 196 + ], + "type": "text", + "content": "[67] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 197, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 240 + ], + "type": "text", + "content": "[68] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 545, + 285 + ], + "type": "text", + "content": "[69] Chengxu Zhuang, Alex Lin Zhai, and Daniel Yamins. Local aggregation for unsupervised learning of visual embeddings. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6002-6012, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "type": "text", + "content": "[70] Adrian Ziegler and Yuki M Asano. Self-supervised learning of object parts for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14502-14511, 2022. 6" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5696" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/706f36ef-c7c5-4ea7-8481-e3a3bb79a94c_content_list.json b/2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/706f36ef-c7c5-4ea7-8481-e3a3bb79a94c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2d246f9de7ce897c4a4413c26a4217d5c54e6266 --- /dev/null +++ b/2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/706f36ef-c7c5-4ea7-8481-e3a3bb79a94c_content_list.json @@ -0,0 +1,2069 @@ +[ + { + "type": "text", + "text": "Semantic-Aware Dynamic Parameter for Video Inpainting Transformer", + "text_level": 1, + "bbox": [ + 122, + 130, + 844, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Eunhye Lee $^{1*}$ , Jinsu Yoo $^{2*}$ , Yunjeong Yang $^{2}$ , Sungyong Baik $^{2,3}$ , Tae Hyun Kim $^{1\\dagger}$ \n{dldms1345, jinsuyoo, yunjoeongyang, dsybaik, taehyunkim}@hanyang.ac.kr \n $^{1}$ Dept. of Computer Science, $^{2}$ Dept. of Artificial Intelligence, $^{3}$ Dept. of Data Science, Hanyang University", + "bbox": [ + 163, + 179, + 803, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent learning-based video inpainting approaches have achieved considerable progress. However, they still cannot fully utilize semantic information within the video frames and predict improper scene layout, failing to restore clear object boundaries for mixed scenes. To mitigate this problem, we introduce a new transformer-based video inpainting technique that can exploit semantic information within the input and considerably improve reconstruction quality. In this study, we use the mixture-of-experts scheme and train multiple experts to handle mixed scenes, including various semantics. We leverage these multiple experts and produce locally (token-wise) different network parameters to achieve semantic-aware inpainting results. Extensive experiments on YouTube-VOS and DAVIS benchmark datasets demonstrate that, compared with existing conventional video inpainting approaches, the proposed method has superior performance in synthesizing visually pleasing videos with much clearer semantic structures and textures.", + "bbox": [ + 75, + 300, + 472, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 588, + 209, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A constant increase in demand for video content in our daily lives (e.g., YouTube or TikTok) has led to the development of video inpainting methods, which aim to complete a missing region or erase unwanted areas such as watermarks and captions from a given input video. In particular, recent convolutional neural networks (CNNs) for video inpainting tasks have shown promising results [34, 15, 2, 43, 37, 9]. In addition, transformer-based inpainting networks [39, 26, 23, 29] significantly escalate the quality of inpainting results through their large network capacity and superior local/global connectivity based on attention mechanisms.", + "bbox": [ + 75, + 607, + 468, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, conventional approaches are unable to fully utilize the semantic information within the input video or distinguish the class-specific characteristics of objects with different semantics. As a result, they frequently fail to recover proper object structure, texture, and scene layout. Several studies [32, 24, 19, 25] have suggested that utiliz", + "bbox": [ + 75, + 773, + 470, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing semantic information can lead to better results and produce visually more plausible images when filling missing regions. However, research on video inpainting has yet to explore the potential of incorporating semantic maps.", + "bbox": [ + 500, + 270, + 890, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video inpainting is a challenging task that requires to render temporally consistent video frames. To address this issue, optical flow has been extensively explored, although it is computationally intensive [15, 37, 9, 23, 41, 40, 13]. In this study, we demonstrate that semantic information can also be used to enforce temporal consistency, and we introduce a new technique that further facilitates the use of semantic cues in a video by dynamically mixing multiple class-specific experts to handle objects with distinct semantics adaptively.", + "bbox": [ + 496, + 334, + 893, + 487 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although previous semantic-guided inpainting networks used the predicted segmentation map as additional input [32] or developed a dedicated semantic-aware inference module [24] to learn semantic-aware parameters, the restoration performance is limited, given the use of shared network parameters for the different semantics. This problem has been addressed in a recent study [25] by using a semantic-aware attention module that enables the features to attend solely to regions with identical semantic labels. However, the computational cost increases proportionally to the number of classes because the attention mechanism is separately carried out for each category.", + "bbox": [ + 496, + 488, + 893, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To mitigate these problems, we propose a novel semantic-aware dynamic parameter selection approach that effectively utilizes the semantic information within input video frames while retaining the number of operations during the inference phase. Specifically, inspired by Cond-Conv [38], we introduce conditionally parameterized linear operations to learn semantic-aware experts, and produce locally varying (dynamic) parameters by leveraging the given semantic cues based on the notion of mixture-of-experts [31, 30, 38]. Our linear operations with dynamically determined parameters replace the standard feed-forward operation within a conventional transformer block. Notably, we can keep the number of parameters for inference by mixing the expert parameters before performing linear operations rather than aggregating the output features cal", + "bbox": [ + 496, + 674, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 101, + 875, + 228, + 886 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 101, + 887, + 228, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12949", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ulated by each expert. Moreover, in contrast to the original CondConv [38] that produces conditional, but locally uniform parameters, ours can generate conditional and locally (token-wise) varying parameters, thus improving each token's representation power.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our extensive experiments witness the outstanding performance of the proposed Semantic-Aware Video Inpainting Transformer (SAVIT) in improving video inpainting results. Specifically, SAVIT elevates quantitative performance on conventional video inpainting benchmark datasets (YouTube-VOS [36] and DAVIS [28]) and produces visually more superior results against state-of-the-art methods. We summarize our contribution as follows:", + "bbox": [ + 75, + 167, + 468, + 287 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We tackle leveraging semantic information within the given input video frames for the video inpainting task.", + "- We introduce a novel semantic-aware video inpainting transformer based on a mixture-of-experts scheme.", + "- We propose a semantic-aware dynamic linear operation to exploit local semantic cues effectively.", + "- Extensive experiments demonstrate the superiority and efficacy of our method, especially in recovering semantic structures and textures." + ], + "bbox": [ + 94, + 292, + 467, + 433 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 450, + 217, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video inpainting. An active research topic in video inpainting is generating a temporally consistent video. Various approaches, including 3D-CNN-based, flow-based, and transformer-based methods, have been proposed to fill in a missing region inside a video plausibly. Despite their ability to extract spatio-temporal information, 3D-CNN methods [34, 2, 3] fail to fully utilize global correspondence due to limited temporal receptive field. Flow-based methods [15, 37, 9, 23, 41, 40, 13] first estimate motion flow in the missing region and use the flow to predict the missing pixels. However, these methods highly rely on pre-trained inpainting networks and often require manual operations [37, 9]. Recently, transformer-based methods [39, 26, 23, 29] have attracted research attention due to their superiority in capturing information across long-ranged frames. Built upon the vision transformer architecture [8], FuseFormer [26] proposed soft tokenization with spatial overlapping to interact information among neighboring tokens. Followup studies have improved the performance with dedicated modules such as focal attention [23] or discrete latent mapping [29]. In our study, we develop our method on top of transformers to capture long-range dependencies and advocate the effectiveness of leveraging semantic information.", + "bbox": [ + 75, + 470, + 468, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic-guided inpainting. Several works have been proposed to utilize the semantic map for inpainting [32, 24, 25], and its related fields [14, 4, 22, 12]. Among them, SPGNet [32] first fills the missing segmentation map, then", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "synthesizes the inpainted image based on the estimated semantic results. Moreover, SGENet [24] introduced iterative feature refinement using semantic map-based normalization [27] to handle more complex holes in mixed scenes, including objects with different semantics. Recent work [25] proposed to calculate semantic-wise attention to learn the corresponding semantic information. However, most of the previous works have developed upon image inpainting, and leveraging semantic cues within a video has yet to be explored due to the lack of annotated data. Our study mitigates the data acquisition problem by gathering pseudo-annotated videos with pre-trained segmentation network [17], and demonstrates that semantic information can elevate video inpainting performance.", + "bbox": [ + 496, + 90, + 890, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic filter. As an alternative to fixed filter, scene-adaptive dynamic filter networks [11, 7, 42] adapt their parameters based on a given input image and/or features. These methods allow flexibility in the network in treating various inputs with different characteristics. Notably, CondInst [33] dynamically generates instance-aware parameters for instance segmentation. In this study, we adapt the transformer parameters in reasoning each token conditioned to its features and semantic information.", + "bbox": [ + 496, + 306, + 890, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mixture-of-experts. By combining the results from multiple expert models, the network can further improve its representation power [30, 10, 31]. Instead of increasing the computational cost during the inference proportional to the number of experts, recent works [38, 5] parameterized the mixed experts to combine the knowledge efficiently. Our work is motivated by CondConv [38]; we mix the experts' parameters before the feature calculation. Specifically, we divide the experts' roles by semantic categories, then combine semantic-aware experts' knowledge at each token.", + "bbox": [ + 496, + 446, + 890, + 597 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 500, + 604, + 671, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we first introduce the annotation procedure to acquire the semantic maps for videos, followed by a detailed description of the proposed framework SAVIT, which leverages semantic information for video inpainting.", + "bbox": [ + 496, + 623, + 890, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Semantic data acquisition", + "text_level": 1, + "bbox": [ + 500, + 686, + 733, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike previous semantic-aware approaches [32, 25, 24], which simultaneously estimate the segmentation maps for inpainting, we assume that semantic segmentation maps are predictable from the given clean video frames, which include unmasked regions to remove, using off-the-shelf segmentation methods. However, currently available video segmentation datasets, such as DAVIS [28] or CityScape [6], are limited to specific categories and/or particular objects, leading acquiring ground-truth segmentation maps for general video frames difficult. Alternatively, we utilize a fully pre-trained panoptic segmentation network [17] to predict the foreground and background semantics in more general scenes.", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12950", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/19747dbd49d052ab8d77d77bccd7236ed356a66ac7a4be1bae95457e866bafa9.jpg", + "image_caption": [ + "(a) Initial segmentation map" + ], + "image_footnote": [], + "bbox": [ + 125, + 97, + 354, + 299 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/586640cd863417072ec09c23f2d6109ec25d0599f18f8ed8b3b60c0160b7e41b.jpg", + "image_caption": [ + "(b) Super-categorize", + "Figure 1: Overview of the semantic data preparation: (a) Temporally inconsistent rough prediction using a pre-trained panoptic segmentation model; (b) Grouping raw predictions into corresponding super-categories; (c) Super-categorized segmentation map produces temporally consistent labels." + ], + "image_footnote": [], + "bbox": [ + 393, + 95, + 653, + 294 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/51008e07925610910e6c24fb9b9ab66fcc56ead2769c1b726f357d65046c93b0.jpg", + "image_caption": [ + "temporally consistent", + "(c) Super-categorized segmentation map" + ], + "image_footnote": [], + "bbox": [ + 725, + 137, + 828, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 1 illustrates the overall semantic data acquisition process used for training and inference. For training, we first apply the panoptic segmentation network [17] to collect estimated segmentation maps for video frames in conventional video inpainting datasets (e.g., YouTube-VOS [36]). The initial segmentation results consist of 53 background and 80 foreground classes [1]. However, these initial segmentation maps with a large number of classes are erroneous and temporally inconsistent, as shown in Fig. 1 (a). Thus, we enforce the temporal consistency of semantics by grouping the predicted labels to the corresponding supercategory of the class. For example, dog, horse, and sheep regions are grouped into a super-category of animal (Fig. 1 (b)). Although we can lose semantic details to some extent by the super-categorization, such constraint helps the training of our semantic-aware inpainting network through the enforced temporal semantic consistency (Fig. 1 (c)).", + "bbox": [ + 75, + 388, + 468, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our work, we define eight different super-categories based on the texture similarities and the frequencies of categories, which are measured on the entire training dataset. Specifically, we have five super-categories for foreground objects: 'animal,' 'person,' 'vehicle,' 'plants,' and 'sports.' Moreover, we have two super-categories for background according to their texture; 'plain' and 'patterned.' Finally, we group the rest of the remaining sub-categories into 'etc.' During the test phase, we also obtain the semantic maps in the same fashion.", + "bbox": [ + 75, + 647, + 468, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Semantic-aware dynamic transformer", + "text_level": 1, + "bbox": [ + 76, + 806, + 405, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To handle diverse scenes with various objects across different classes in videos, we propose a semantic-aware dynamic transformer. In particular, our transformer employs a mixture-of-experts scheme, in which each expert is a single fully connected layer with its parameters reserved for each", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "corresponding semantic class (i.e., super-category) obtained from Section 3.1. Conditioned on the semantic features of each token, we create a token-wise fully connected layer, where its parameters are determined by weighted summing the class-specific parameters, and the weights for the mixture are produced by our semantic router as in the following.", + "bbox": [ + 496, + 388, + 890, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning semantic router. To produce locally (i.e., token-wise) varying parameters conditioned on the semantic information, the network requires a routing function that predicts blending weights to mix the class-specific expert parameters. Our router computes routing weights (i.e., mixture coefficients) for each token by:", + "bbox": [ + 496, + 482, + 890, + 573 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nr _ {i} = M L P \\left(f _ {i}\\right) + h a r d m a x \\left(s _ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 584, + 890, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $r_i \\in \\mathbb{R}^C$ is a produced mixture weight vector; $f_i$ is the feature of an input token; $s_i \\in \\mathbb{R}^C$ is a histogram of class distribution on the segmentation map for the given $i$ -th token; $C$ is the number of semantic experts ( $C = 8$ in this work unless stated otherwise); and MLP is a standard multi-layer perceptron (MLP) composed of two fully-connected layers with Leaky ReLU activation and sigmoid output. Our router uses the most dominant semantic label from the input histogram $s_i$ by using a hard selection strategy (hardmax operation) rather than a softmax result. In addition, we model our router to be lightweight (i.e., two layers) to minimize the overhead, keeping our framework to have similar computational costs to the standard transformers during inference.", + "bbox": [ + 496, + 609, + 890, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semantic-aware dynamic parameterization. By using the routing weights $r_i$ generated by the router for the $i$ -th token with features $f_i$ , we produce semantic-aware dynamic parameters $W_i$ for linear (fully-connected) layers of a transformer block. The formulation of semantic-aware dynamic", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12951", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4b3569247df550c85411f6c17597e8cda4bb180c0b91273b3fecaaf60d7fad56.jpg", + "image_caption": [ + "Figure 2: Illustration of SAFFN. With the histogram of semantic categories (e.g., hardmax) and additional linear layers, the semantic router outputs a routing weight for a specific token. Using the routing weight, SAL first acquires token-specific parameters by the weighted sum of experts, then operates the linear transformation of the token with the parameters. Here, we illustrate the operational flow of a single token. $\\oplus$ indicates element-wise summation." + ], + "image_footnote": [], + "bbox": [ + 112, + 93, + 416, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "layer (SAL) is given by:", + "bbox": [ + 76, + 585, + 240, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S A L} \\left(f _ {i}\\right) = f _ {i} \\cdot W _ {i}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 609, + 468, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the mixed expert parameter $W_{i}$ for token $f_{i}$ yields:", + "bbox": [ + 76, + 635, + 455, + 650 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW _ {i} = \\sum_ {c = 1} ^ {C} r _ {i} (c) \\cdot w ^ {c}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 657, + 468, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $w^{c}$ indicates the parameters of each class-specific expert, while $r_{i}(c)$ is the $c$ -th element of routing weight vector $r_{i}$ and means the mixture coefficient for the corresponding expert $w^{c}$ . Consequently, our semantic-aware dynamic parameterization produces locally varying (token-wise) parameters. For example, a token corresponding to a goat in the input frame can be handled by parameters specialized for animal super-category, and another token on the rock can be dealt with specific parameters for the plain background. Notably, our framework can also handle a token of complex scenes where semantically different objects are coexisting. This is because our router is designed to produce locally varying weights and give high routing weights", + "bbox": [ + 75, + 703, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for experts corresponding to coexisting objects, rather than choosing a single expert.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we devise a semantic-aware dynamic feedforward network (SAFFN) as depicted in Fig. 2. SAFFN includes two SAL using the dynamically generated semantic-aware parameters as:", + "bbox": [ + 498, + 121, + 890, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {S A F F N} \\left(f _ {i}\\right) = \\operatorname {S A L} \\left(\\operatorname {R e L U} \\left(\\operatorname {S A L} \\left(f _ {i}; W _ {i} ^ {l}\\right)\\right); W _ {i} ^ {l + 1}\\right) \\tag {4} \\\\ = \\max \\left(\\mathbf {0}, f _ {i} \\cdot W _ {i} ^ {l}\\right) \\cdot W _ {i} ^ {l + 1}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 186, + 890, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our SAFFN, each SAL is parameterized by our dynamic parameters $W_{i}^{l}$ , where $l$ denotes the layer index of the linear transformation, and ReLU activation is included between two layers. Bias parameters are omitted for brevity.", + "bbox": [ + 498, + 233, + 890, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Overall architecture", + "text_level": 1, + "bbox": [ + 500, + 296, + 691, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 3 illustrates the overall pipeline of the proposed video inpainting network SAVIT. The network is composed of encoder, decoder, and multiple transformer blocks between the encoder and the decoder. First, the encoder produces output features by taking $N$ consecutive video frames with masks and the corresponding semantic segmentation maps as input. We employ the encoder in FuseFormer [26] and additionally concatenate corresponding semantic maps to input. Next, we carry out the tokenization procedure [26], then forward to several transformer blocks to extract powerful features. Finally, the tokenized features are rearranged into the image-like shape, which is then fed into the decoder that predicts $N$ inpainted video frames as previous works [26, 23].", + "bbox": [ + 498, + 311, + 890, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that our semantic-aware mixture-of-experts schemes can be plugged into any conventional transformer-based approach (e.g., FuseFormer [26], E2FGVI [23]), where a conventional transformer block is just replaced with our semantic-aware transformer block. In our experiments, we embed our dynamic transformer within FuseFormer [26]. Specifically, SAVIT contains eight FuseFormer blocks, each of which consists of a multi-head self-attention (MHSA) operation and fusion feed-forward network (F3N). We replace one of the eight FuseFormer blocks with our dynamic transformer block. Precisely, within the F3N in the fourth FuseFormer block, we modify the two standard linear layers with our dynamic ones and add a semantic router.", + "bbox": [ + 496, + 523, + 890, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic-aware dynamic discriminator. To facilitate each expert's learning of the corresponding category's representation and texture details, we employ a semantic-aware discriminator (SAD) that predicts the real/fake score for each token. Fig. 4 illustrates the architecture of SAD, which has similar structures to the encoder of SAVIT, except for the last layer, where a single SAL is added. Notably, we utilize the same mixture coefficients for dynamic parameters for both SAVIT (generator) and SAD (discriminator), encouraging each expert to learn detailed and discriminative category-specific representation.", + "bbox": [ + 496, + 738, + 890, + 904 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12952", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c92922e1b79655c66e6835ca0662a9d7faa306d7f5c9e69c77aa9099a5715e58.jpg", + "image_caption": [ + "Figure 3: Overall flow of SAVIT. Our method utilizes additional semantic cues with a mixture-of-experts scheme. The transformer parameters are dynamically adapted to the local semantic information for each token region. With these token-wise dynamic parameters, our network produces more precise inpainting results, generating an object's boundary or synthesizing detailed textures." + ], + "image_footnote": [], + "bbox": [ + 106, + 93, + 872, + 296 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/44b6beda0f9f12c1ea8d3d53d535b9efeb1a69caa4c0efb15182ecd802aba7a1.jpg", + "image_caption": [ + "Figure 4: Schematic illustration of SAD. With the specific mixture-of-experts parameters produced by using the same routing weight from the generator, SAD discriminates whether the inpainted results are realistic or not." + ], + "image_footnote": [], + "bbox": [ + 86, + 373, + 455, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Loss function", + "text_level": 1, + "bbox": [ + 76, + 579, + 217, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Auxiliary token classification loss. In the mixture-of-experts scheme, it is crucial that each expert is assigned with distinct role [18, 30]. If our semantic experts have been assigned properly to each corresponding category, the output tokens from the semantic-aware dynamic block should be helpful for classification. We drive collaborative role division across our semantic experts. Specifically, we regularize the learning of experts during training by performing auxiliary classification tasks on the output tokens from the semantic-aware dynamic block. Such token-wise classification is performed by an auxiliary single linear layer classifier. The classifier predicts the semantic labels of the output tokens, as illustrated with the green box in Fig. 3. As a result, each expert can effectively learn the feature representation corresponding to a specific semantic label.", + "bbox": [ + 75, + 597, + 468, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall objective function. Our overall objective function $\\mathcal{L}_{total}$ is the weighted sum of the following four different loss terms:", + "bbox": [ + 75, + 825, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p i x}} + \\lambda_ {\\text {a d v}} \\mathcal {L} _ {\\text {a d v}} + \\lambda_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}} + \\lambda_ {\\text {s a d}} \\mathcal {L} _ {\\text {s a d}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 882, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{pix}$ is a pixel-wise reconstruction loss, $\\mathcal{L}_{adv}$ is an adversarial loss, $\\mathcal{L}_{cls}$ is token classification loss, and $\\mathcal{L}_{sad}$ is semantic adversarial loss, while $\\{\\lambda_{adv},\\lambda_{cls},\\lambda_{sad}\\}$ control the weight of associated loss term. Given the predicted video frame $\\hat{Y}$ and its ground truth counterpart $Y$ , we employ the L1 pixel-wise reconstruction loss as:", + "bbox": [ + 498, + 369, + 890, + 460 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p i x} = \\left| \\left| \\hat {Y} - Y \\right| \\right| _ {1}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 468, + 890, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As for the adversarial loss, we train our inpainting network to fool a discriminator $D$ as follows:", + "bbox": [ + 498, + 497, + 890, + 526 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a d v} = - E _ {\\hat {Y}} [ D (\\hat {Y}) ]. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 536, + 890, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on a temporal patch discriminator [26, 23], the discriminator $D$ is trained to optimize", + "bbox": [ + 498, + 564, + 890, + 594 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D} = E _ {Y} \\left[ \\operatorname {R e L U} (1 - D (Y)) \\right] + E _ {\\hat {Y}} \\left[ \\operatorname {R e L U} (1 + D (\\hat {Y})) \\right]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 603, + 890, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The token classification loss term is given as", + "bbox": [ + 500, + 636, + 795, + 650 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} = - \\sum_ {i} \\sum_ {c = 1} ^ {C} y _ {i} ^ {c} \\log \\left(\\text {s o f t m a x} \\left(\\hat {y} _ {i}\\right)\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 660, + 890, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{y}_i\\in \\mathbb{R}^C$ is the predicted score from the single-layer classifier conditioned on the output tokens from the semantic-aware dynamic block, while $y_{i}^{c}$ is the target probability for class $c$ at the $i$ -th token. For the ground truth target $y_{i}^{c}$ , we use the input semantic segmentation map (i.e., $y_{i}^{c} = s_{i}(c)$ ). The semantic-aware adversarial loss $\\mathcal{L}_{sad}$ is similar to $\\mathcal{L}_{adv}$ in that (8) and (7) are used, but with the semantic-aware discriminator (SAD) instead of $D$ .", + "bbox": [ + 498, + 710, + 890, + 833 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 840, + 632, + 856 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we quantitatively and qualitatively demonstrate the effectiveness of SAVIT. Please refer to the supplementary material for more experimental results.", + "bbox": [ + 498, + 858, + 890, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12953", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4870b981a1c5e6971b8ce26899f6f126eff11ef3b2c1467013882087aeb7be56.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
YouTube-VOS [36]DAVIS [28]
MethodPSNR↑SSIM↑VFID↓E_warp(%)↓PSNR↑SSIM↑VFID↓E_warp(%)↓
VINet [15]29.200.94340.0720.149028.960.94110.1990.1785
LGTSM [3]29.740.95040.0700.185928.570.94090.1700.1640
CAP [21]31.580.96070.0710.147030.280.95210.1820.1533
STTN [39]32.340.96550.0530.090730.670.95600.1490.1449
E2FGVI [23]33.710.97000.0460.086433.010.97210.1160.1315
ISVI [41]31.190.95690.0530.117332.280.96690.1290.1589
FuseFormer [26]33.160.96730.0510.090032.540.97000.1380.1362
SAVIT (Ours)33.970.97270.0430.043633.140.97480.1070.0673
", + "bbox": [ + 135, + 88, + 831, + 237 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/47b4c628aef37a12370a5b7acd28a8bcd2ca46da086d404859a1073a7a4fa63b.jpg", + "table_caption": [ + "Table 1: Quantitative comparison of SAVIT against state-of-the-art methods. SAVIT outperforms the existing approaches in terms of various evaluation metrics. Notably, SAVIT significantly elevates the inpainting performance of FuseFormer [26], which is our baseline architecture, witnessing the superiority of leveraging semantic cues within the input video." + ], + "table_footnote": [ + "Table 2: Ablation experiments on various network configurations under DAVIS [28] dataset. Starting from the performance of the baseline architecture [26], the inpainting performance gradually increases as the configurations are added since the network can benefit from the semantic cues within a video." + ], + "table_body": "
ConfigurationPSNR↑VFID↓
AFuseFormersmall [26]30.510.182
B+ super-categorized segmentation map30.900.161
C+ SAL30.920.164
D+ token classification loss (λcls)31.190.165
E+ SAD31.010.157
", + "bbox": [ + 94, + 295, + 447, + 381 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 486, + 290, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and evaluation metrics. We use two conventional video inpainting datasets, YouTube-VOS [36] and DAVIS [28], for our experiments. YouTube-VOS dataset consists of 3471 training, 474 validation, and 508 test videos from mixed scenes. Meanwhile, DAVIS dataset comprises 150 video clips, 60 of which are used for training and the remaining 90 clips are used for evaluation. Following previous works [26, 23], we use only the training videos in the YouTube-VOS for training, while the evaluation is conducted on YouTube-VOS testset and 50 selected clips of DAVIS testset. During training, we create random mask shapes for input masked frames as in [39, 26, 23] and use the method discussed in Section 3.1 to obtain a segmentation map for the entire dataset.", + "bbox": [ + 75, + 503, + 468, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The inpainting performance is evaluated under two standard scenarios: fixed region inpainting and foreground object removal. For the fixed region inpainting, we use free-form fixed masks for each video provided by FuseFormer [26]. As for the foreground object removal, we remove the main objects using frame-by-frame annotations provided by the stakeholders. Notably, we employ a simple k-nearest neighbors algorithm to complete the missing object region in the segmentation maps for the object removal task (please refer to supplementary material for detail).", + "bbox": [ + 75, + 717, + 468, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the results with respect to four evaluation metrics: PSNR, SSIM, temporal warping error $E_{\\text{warp}}$ [20],", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and Video-based Fréchet Inception Distance (VFID) [35]. PSNR and SSIM are traditional metrics used to measure the average quality of produced frames. VFID and $E_{\\text{warp}}$ measure the video's perceptual quality and temporal stability (consistency), respectively, and low values indicate better quality.", + "bbox": [ + 496, + 297, + 890, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training settings. We follow the training settings of our baseline network FuseFormer [26]. Specifically, we randomly sample five frames from a video, resize the frames to $240 \\times 432$ , and utilize random horizontal flipping augmentation. The network is trained for $500k$ iterations with Adam optimizer [16] and a batch size of eight. The initial learning rate is 1e-4 and decreases to 1e-5 after $400k$ iterations. The coefficients for our loss functions, $\\lambda_{adv}$ , $\\lambda_{cls}$ , and $\\lambda_{sad}$ are 0.01, 0.01, and 0.001, respectively. We implement our network on the PyTorch framework and use eight NVIDIA RTX A6000 GPUs for training.", + "bbox": [ + 496, + 393, + 892, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Video inpainting results", + "text_level": 1, + "bbox": [ + 498, + 561, + 718, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results. Table 1 quantitatively compares our video inpainting results on fixed region inpainting setting against seven state-of-the-art video inpainting networks: VINet [15], LGTSM [3], CAP [21], STTN [39], E2FVGI [23], ISVI [41], and FuseFormer [26]. The result shows that our method remarkably elevates the performance of FuseFormer [26], on which we base our architecture, and achieves superior performances on both YouTube-VOS and DAVIS datasets compared to the state-of-the-art methods. Notably, the warping error has also significantly improved, indicating that semantic information benefits in generating temporally consistent video. These results show the effectiveness of employing mixture-of-experts to dynamically leverage the semantic cues within given frames for video inpainting.", + "bbox": [ + 496, + 579, + 892, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative results. Fig. 5 shows that SAVIT has visually pleasing inpainting results, especially in synthesizing semantic structure, boundary, and texture of objects. Notably, unlike the baseline methods that rely solely on inherent video clues and cannot accurately restore the shape of the human body, SAVIT effectively completes the region", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12954", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8f9ef754c0e2f525e324a8415e1ff9fad5cd58bd9f064040d0f75fe35a052a96.jpg", + "image_caption": [ + "Masked Frames" + ], + "image_footnote": [], + "bbox": [ + 138, + 74, + 274, + 133 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e12573d17e4ed4bbdef05d514d4cfcb748be6c60ec3ea4406cc4f14a89dcc42.jpg", + "image_caption": [ + "STTN" + ], + "image_footnote": [], + "bbox": [ + 276, + 74, + 418, + 133 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83c9e68628e4dbb172d29abcd1c981678252f40549c178078f8137a4c8933100.jpg", + "image_caption": [ + "FuseFormer" + ], + "image_footnote": [], + "bbox": [ + 418, + 74, + 555, + 136 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db5af5a1e67cd1cd0acaf227123157af4be6685c972685b42ce5c94e51797cd3.jpg", + "image_caption": [ + "E2FGVI" + ], + "image_footnote": [], + "bbox": [ + 555, + 74, + 689, + 136 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/61ddfec66c64ef64f1e9d49cf480898400a41ac45fb5c6bfe0be325c09473e8b.jpg", + "image_caption": [ + "SAVIT (Ours)" + ], + "image_footnote": [], + "bbox": [ + 689, + 74, + 828, + 136 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c4147a3778a7f13a666cc6763133e5c893c7817b42f7aa996fffa7b08e93034.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 138, + 135, + 274, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4317c7295e4383c139139d008c49d540460ab19648ef61ed0cf72cbaef892e5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 135, + 418, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ecaeab738521227e51db4e6f4d7f990f33ba22d70d047337a3000c2d4e66acca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 135, + 553, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/596c988000b2e45b2463901bdaf47f6f65d7aa85adf85a2b3a5baf6a4af84b05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 135, + 689, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d38d9b0ac28599f1352c93ab68bfc66c176a0638a94010869bb1a98aa065e1e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 135, + 828, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a1a24f9ceac5f5e6855c058b28efb731e8dea5b271b002eeb283b9b5c66c6125.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 138, + 194, + 274, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ffc4d195e1f7d7db840bb62f8f3a22634d9d0814194e2b3aaa1e8ca4a05858b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 194, + 418, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/37a9502c5f70e0f52997fd5faf314adf0d73e5c8aa12dffbaa726d0e6a2ffca3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 194, + 553, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/650781ecd42b506da60b43ac943a79287cd01d68a99b4e566ebc8a09b2b0a2d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 194, + 689, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7df933a9d73d7f4f7bd3ad244b9e0a1deeef2701e8cd1a7628d7a6f83b3c36db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 194, + 828, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5e557ee2e7efeac777aa2741c8494cb4f211df93eb81c60d5d6a36303eff8b3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 138, + 253, + 274, + 313 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dcf1dd3bbcea01e073c1197ba4cbebcd62a0a184871f1a5ca58fdcabbe11c58f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 253, + 418, + 313 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/42fd4d9da2aab0f05f9d2a04a54016784b64d742ba53a31b6a55936660320420.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 253, + 553, + 313 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d93db6a3e4bbe625e5f75f2b0655a892517bbc3be5b6146d13c362c99a0414bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 253, + 689, + 313 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/553275ed3a56e3b5667a7bdab89edf9dd46a4fcc0d412a38112262cb2cfdd065.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 253, + 828, + 314 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c828aff8235dcc0408eb3b55ccdd947e47ef8ef765be8108fdff4cf5d7ec5bed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 138, + 314, + 274, + 373 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c0f95f58d30ef09e84edbffc1a1b192cdbe5f452721b1cf40217a51f2e623af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 314, + 418, + 373 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e3c27885d3d128e97fba23afbbce14e479568f22a10023d3c1bfe49bf2c8b4a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 314, + 553, + 373 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/17c81c33b45fe4596f76f1f2309547b48631b4081b48ed8a112cb15b1a55cd37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 314, + 689, + 373 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/24999c0908660f75dffaefda4aacc2d52fc0ebbcdfd15e9cf96a9c276f4ac988.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 314, + 828, + 373 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f0f4cb76128310bec266cfe7d02c9a6186e03a5ac2bcba8fdb10ef43e18e283c.jpg", + "image_caption": [ + "Figure 5: Qualitative comparison of SAVIT with baseline video inpainting networks in fixed region inpainting and object removal. SAVIT more accurately recovers the object's boundaries and appropriate textures." + ], + "image_footnote": [], + "bbox": [ + 138, + 375, + 274, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bca9e385492a285e57d66712941e82a4338c0d4129693e942c32479790841a37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 375, + 418, + 434 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9f42f31a2daef94f9b2b42752ca633f7b518305740b6117c0885179112dcfbe2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 375, + 553, + 434 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb457ca45898752bc9c94b4e07c1966d9b8b20b1cb5d2f9cbd80f4781d8335d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 375, + 689, + 434 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/13229ab8c7ddaeef1b4fb360903aaa689533aff721ecdfc4730ce24fae886665.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 375, + 828, + 434 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e1640d5d2232cd6f429e96210ddf9ca508d1ce50b726c4733789844f3c55c9aa.jpg", + "image_caption": [ + "Figure 6: User study result." + ], + "image_footnote": [], + "bbox": [ + 78, + 484, + 174, + 556 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c307646ea73a9204afc13553f891a750ba243675718221088d8c5b4d33b39756.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ConfigurationPSNR†
hardmax (SAVIT)31.01
softmax30.91
random routing weight19.86
", + "bbox": [ + 232, + 488, + 436, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3: Effect of the routing weight on the DAVIS [28] dataset.", + "bbox": [ + 207, + 558, + 464, + 587 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "by leveraging additional semantic information (third and fourth rows). Furthermore, SAVIT demonstrates the capability of generating more semantically plausible textures, as observed from the detailed textures of synthesized grass and wall (second and fifth rows).", + "bbox": [ + 75, + 603, + 468, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User study. We conduct a user study on the foreground object removal task, where ground truth target video frames for evaluation are not available, to investigate the real-world applicability of our method using DAVIS [28] testset. We ask 24 raters to choose a better video between two randomly ordered videos, including the inpainted results by FuseFormer [26] and ours, fifteen times per rater. The result presented in Fig. 6 shows that raters prefer ours to the baseline, indicating the impact of exploiting semantic cues in completing missing object regions.", + "bbox": [ + 75, + 686, + 468, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Ablation studies", + "text_level": 1, + "bbox": [ + 76, + 847, + 236, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform ablation experiments based on a smaller FuseFormer network (FuseFormersmall) by halving the dimension of the hidden layer.", + "bbox": [ + 75, + 867, + 468, + 912 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d79bd840aa5d68ad9ebdacbe7ef6254649b43d37d35a6105892e0055d143ebc.jpg", + "image_caption": [ + "SAVIT w/o SAD", + "Figure 7: Effect of semantic-aware discriminator (SAD). The texture of the black swan's body is blurry in the case without SAD." + ], + "image_footnote": [], + "bbox": [ + 501, + 489, + 692, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7b4fe8511bf1a6cea21b03c4df82bc973f361a247cdf9aa04169684d0c5aae11.jpg", + "image_caption": [ + "SAVIT w/ SAD" + ], + "image_footnote": [], + "bbox": [ + 697, + 488, + 890, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Network configuration. Table 2 compares the inpainting performance for various configurations of SAVIT. First, we observe that the baseline network's performance improves with additional segmentation map inputs (rows A and B). However, a subtle additional performance improvement (0.02dB PSNR gain, rows B and C) is achieved when we naively add our proposed dynamic linear layer to the network. Rather, applying the proposed token-wise classification loss to the objective boosts the performance (0.2dB PSNR gain, row D). Overall performance improvement by applying mixture-of-experts (rows C and D) highlights the importance and effectiveness of employing mixture-of-experts to handle semantics dynamically. Moreover, using a semantic-aware dynamic discriminator (row E) degrades PSNR while improving VFID. However, as shown in Fig. 7, utilizing the semantic-aware discriminator produces more visually plausible results, which solidifies the advantage of", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12955", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e6a537c1d41f40c7c2e54af2f8543267c4121a3c6aa7101e40a7c65fc67c471.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 85, + 181, + 138 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/683f84253a3d4df672db5955b8591fbc2ca835d10e89c5ec21dfbeccd40c1a70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 85, + 284, + 138 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/11ccf9fa0c90e978f58ca19e20a7cac7e10d8d28e023726e7caa73f05c919c7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 85, + 383, + 138 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a8d45a3c5ca0fdbd138b0ba804b9beae8318c856b29c33689fde41ac4da28eca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 85, + 486, + 138 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fb3a4e1974f878c4a2d0956f15bd8ac4a1cef61e231d6fe7729a0c353fa4e9b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 85, + 586, + 138 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/77a8f6095b22a6baff8c284663afe9ad4981f940925da0c6a9cecb1f2b7d1b2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 85, + 686, + 138 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b022bd318f8cd0d92db791a753f0df51156f333fe89c87dbfe32607c873c0515.jpg", + "image_caption": [ + "Figure 8: Visualization of the semantic router results. (1) to (8) are the visualization of routing weights of the corresponding super-category. The routing weights are rearranged to the image-like features. Brighter pixel values indicate higher routing weights." + ], + "image_footnote": [], + "bbox": [ + 81, + 145, + 181, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/caf30627110a1a0f1af152c9b326d99547050e2cc717445342f9dec54715d7e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 145, + 284, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e7f834c23e09fd6ebd51fa147a7a1c3a815dd1dd386c76e6d98730eb5e259917.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 145, + 383, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/86c02db76eb995e3c627518ee35bf74f1185d927e568fc96194ffdf13e483d9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 145, + 486, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/36285cb7505347e38e657601e9874d0e418bb7a51053a0d0ba0493839ba9fe31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 145, + 586, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ef1a5e631bc5ee88923b3a65cc4799f2042c7af9ec19d28047201a5ca510acca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 145, + 686, + 196 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/eb94e0b8cacb880f280f963286a163d8a1646f2f6bcde629380f53f81d7c2ec5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
# ExpertsPSNR↑VFID↓
430.830.165
630.900.168
831.010.157
", + "bbox": [ + 697, + 88, + 885, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a34b416150bc72a0c10c74a5618c4197e585ae0fd65437606b1f410293f19073.jpg", + "image_caption": [ + "Figure 9: Qualitative comparison on different routing weight strategies." + ], + "image_footnote": [], + "bbox": [ + 78, + 253, + 274, + 299 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/563d625ad537ac3e52b57ea182197f70af9719ba06347e067b5f16e05b581f56.jpg", + "table_caption": [ + "Table 4: Inpainting results by changing the numbers of experts under the DAVIS [28] dataset. More experts render better inpainting results." + ], + "table_footnote": [], + "table_body": "
# blocks# params.PSNR↑
126.0M30.40
233.1M30.45
", + "bbox": [ + 285, + 255, + 464, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the discriminator in improving perceptual quality.", + "bbox": [ + 76, + 369, + 405, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of routing weights. Fig. 8 visualizes the routing weights of each expert from the trained semantic router. We rearrange the tokens to image-like features for visualization, and brighter values indicate higher routing weights in the corresponding region. The bright regions among different experts indicate that the experts collaborate during the inpainting procedure rather than simply utilizing a specific expert. We see that routing weights tend to change locally smoothly while the input segmentation map changes abruptly.", + "bbox": [ + 75, + 388, + 468, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Designing choice on semantic router. Table 3 compares different design choices of our semantic router. The performance drop by replacing the trained routing weights (i.e., mixture coefficients) with random weights indicates the discriminative role of trained experts. Moreover, the results of softmax routing strategy show a noticeable performance degradation against hardmax strategy quantitatively. Fig. 9 also shows that softmax results tend to have blurry texture, especially on the edge area. Therefore, Hardmax was chosen for our final model to facilitate role division between experts better, as it enables the learning of distinct parameters by dividing the input for each expert before the router blends the token based on its features.", + "bbox": [ + 75, + 544, + 468, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of the number of experts. Table 4 compares the inpainting performance by changing the number of experts. Notably, as the number of experts (i.e., semantic categories) increases, the network can benefit from more diverse semantic clues in terms of PSNR and VFID values, and we set the number of experts for our final model to eight.", + "bbox": [ + 75, + 744, + 468, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of the number of semantic blocks. Table 5 compares the inpainting performance by applying our semantic module on a different number of blocks. The result indicates that using more semantic blocks produces a slightly", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/65f8596b3ea6a40f58f013b1318484f4198424a7be21fd7b00fc04542280202c.jpg", + "table_caption": [ + "Table 5: Ablation on different number of semantic blocks." + ], + "table_footnote": [], + "table_body": "
MethodMask ratioPSNR†VFID↓
FuseFormersmall30%22.570.554
60%18.101.059
FuseFormersmall + Ours30%23.570.412
60%19.210.750
", + "bbox": [ + 537, + 255, + 849, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6: Comparison on varying mask ratio.", + "bbox": [ + 549, + 333, + 841, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "better inpainting performance. Although we apply our semantic module one time due to hardware limits for our final model, it is expected that the results will be naturally improved by increasing the number of semantic blocks.", + "bbox": [ + 498, + 364, + 890, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Input video with large corruption. Table 6 compares our method against FuseFormersmall under different corrupted region ratio settings. It is shown that, despite larger masks, our method still outperforms the baseline by a large margin by exploiting the semantic information.", + "bbox": [ + 498, + 439, + 890, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Future Works", + "text_level": 1, + "bbox": [ + 500, + 529, + 776, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we have explored leveraging semantic information within a given video for transformer-based video inpainting. Accordingly, we propose an effective semantic-aware dynamic transformer with the notion of mixture-of-experts. Our dynamic transformer contains class-specific expert parameters where experts are efficiently combined to perform token-wise dynamic linear operations based on the token feature and its semantic information. The experiments demonstrate the effectiveness of our method, especially in recovering semantic structures. Although we have utilized pseudo labels for our study, we believe more accurate and denser annotated datasets can improve the inpainting performance further since our method relies on the quality of the segmentation label.", + "bbox": [ + 496, + 550, + 890, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 782, + 666, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2023-00222776) and Samsung Electronics Co., Ltd, and Samsung Research Funding Center of Samsung Electronics under Project Number SRFCIT1901-06.", + "bbox": [ + 498, + 810, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12956", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3", + "[2] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 1, 2", + "[3] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Learnable gated temporal shift module for deep video inpainting. In Proceedings of the British Machine Vision Conference (BMVC), 2019. 2, 6", + "[4] Cheng Chen, Jiayin Cai, Yao Hu, Xu Tang, Xinggang Wang, Chun Yuan, Xiang Bai, and Song Bai. Deep interactive video inpainting: An invisibility cloak for harry potter. In Proceedings of the ACM International Conference on Multimedia (ACM MM), 2021. 2", + "[5] Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolution: Attention over convolution kernels. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2", + "[7] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[9] Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2", + "[10] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 1991. 2", + "[11] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. In Advances in Neural Information Processing Systems (NeurIPS), 2016. 2", + "[12] Jun-Gyu Jin, Jaehyun Bae, Han-gyul Baek, and Sang-hyo Park. Object-ratio-preserving video retargeting framework based on segmentation and inpainting. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 2", + "[13] Jaeyeon Kang, Seoung Wug Oh, and Seon Joo Kim. Error compensation framework for flow-guided video inpainting." + ], + "bbox": [ + 78, + 114, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 2", + "[14] Bholeshwar Khurana, Soumya Ranjan Dash, Abhishek Bhatia, Aniruddha Mahapatra, Hrituraj Singh, and Kuldeep Kulkarni. Semie: Semantically-aware image extrapolation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 2", + "[15] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Deep video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 6", + "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[17] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3", + "[18] Xiangtao Kong, Hengyuan Zhao, Yu Qiao, and Chao Dong. Classsr: A general framework to accelerate super-resolution networks by data characteristic. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5", + "[19] Avisek Lahiri, Arnav Kumar Jain, Sanskar Agrawal, Pabitra Mitra, and Prabir Kumar Biswas. Prior guided gan based semantic inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1", + "[20] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 6", + "[21] Sungho Lee, Seoul Wug Oh, DaeYeun Won, and Seon Joo Kim. Copy-and-paste networks for deep video inpainting. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 6", + "[22] Jiacheng Li, Chang Chen, and Zhiwei Xiong. Contextual outpainting with object-level contrastive learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[23] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 4, 5, 6", + "[24] Liang Liao, Jing Xiao, Zheng Wang, Chia-Wen Lin, and Shin'ichi Satoh. Guidance and evaluation: Semantic-aware image inpainting for mixed scenes. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2", + "[25] Liang Liao, Jing Xiao, Zheng Wang, Chia-Wen Lin, and Shin'ichi Satoh. Image inpainting guided by coherence priors of semantics and textures. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2", + "[26] Rui Liu, Hanming Deng, Yangyi Huang, Xiaoyu Shi, Lewei Lu, Wenxiu Sun, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fuseformer: Fusing fine-grained information in" + ], + "bbox": [ + 503, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12957", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "transformers for video inpainting. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1, 2, 4, 5, 6, 7", + "[27] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[28] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6, 7, 8", + "[29] Jingjing Ren, Qingqing Zheng, Yuanyuan Zhao, Xuemiao Xu, and Chen Li. Dlformer: Discrete latent transformer for video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2", + "[30] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 1, 2, 5", + "[31] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In Proceedings of the International Conference on Learning Representations (ICLR), 2017. 1, 2", + "[32] Yuhang Song, Chao Yang, Yeji Shen, Peng Wang, Qin Huang, and C-C Jay Kuo. Spg-net: Segmentation prediction and guidance network for image inpainting. In Proceedings of the British Machine Vision Conference (BMVC), 2018. 1, 2", + "[33] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2", + "[34] Chuan Wang, Haibin Huang, Xiaoguang Han, and Jue Wang. Video inpainting by jointly learning temporal structure and spatial details. In Association for the Advancement of Artificial Intelligence (AAAI), 2019. 1, 2", + "[35] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2018. 6", + "[36] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 2, 3, 6", + "[37] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep flow-guided video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2", + "[38] Brandon Yang, Gabriel Bender, Quoc V Le, and Jiquan Ngiam. Condconv: Conditionally parameterized convolutions for efficient inference. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 1, 2", + "[39] Yanhong Zeng, Jianlong Fu, and Hongyang Chao. Learning joint spatial-temporal transformations for video inpainting." + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 6", + "[40] Kaidong Zhang, Jingjing Fu, and Dong Liu. Flow-guided transformer for video inpainting. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 2", + "[41] Kaidong Zhang, Jingjing Fu, and Dong Liu. Inertia-guided flow completion and style fusion for video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6", + "[42] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[43] Xueyan Zou, Linjie Yang, Ding Liu, and Yong Jae Lee. Progressive temporal feature alignment network for video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1" + ], + "bbox": [ + 501, + 92, + 893, + 345 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12958", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/706f36ef-c7c5-4ea7-8481-e3a3bb79a94c_model.json b/2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/706f36ef-c7c5-4ea7-8481-e3a3bb79a94c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..87166154095b8d40fe1143037d23a9f507e699a8 --- /dev/null +++ b/2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/706f36ef-c7c5-4ea7-8481-e3a3bb79a94c_model.json @@ -0,0 +1,2662 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.124, + 0.131, + 0.846, + 0.154 + ], + "angle": 0, + "content": "Semantic-Aware Dynamic Parameter for Video Inpainting Transformer" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.18, + 0.804, + 0.235 + ], + "angle": 0, + "content": "Eunhye Lee\\(^{1*}\\), Jinsu Yoo\\(^{2*}\\), Yunjeong Yang\\(^{2}\\), Sungyong Baik\\(^{2,3}\\), Tae Hyun Kim\\(^{1\\dagger}\\) \n{dldms1345, jinsuyoo, yunjoeongyang, dsybaik, taehyunkim}@hanyang.ac.kr \n\\(^{1}\\)Dept. of Computer Science, \\(^{2}\\)Dept. of Artificial Intelligence, \\(^{3}\\)Dept. of Data Science, Hanyang University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.473, + 0.573 + ], + "angle": 0, + "content": "Recent learning-based video inpainting approaches have achieved considerable progress. However, they still cannot fully utilize semantic information within the video frames and predict improper scene layout, failing to restore clear object boundaries for mixed scenes. To mitigate this problem, we introduce a new transformer-based video inpainting technique that can exploit semantic information within the input and considerably improve reconstruction quality. In this study, we use the mixture-of-experts scheme and train multiple experts to handle mixed scenes, including various semantics. We leverage these multiple experts and produce locally (token-wise) different network parameters to achieve semantic-aware inpainting results. Extensive experiments on YouTube-VOS and DAVIS benchmark datasets demonstrate that, compared with existing conventional video inpainting approaches, the proposed method has superior performance in synthesizing visually pleasing videos with much clearer semantic structures and textures." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.589, + 0.21, + 0.605 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.775 + ], + "angle": 0, + "content": "A constant increase in demand for video content in our daily lives (e.g., YouTube or TikTok) has led to the development of video inpainting methods, which aim to complete a missing region or erase unwanted areas such as watermarks and captions from a given input video. In particular, recent convolutional neural networks (CNNs) for video inpainting tasks have shown promising results [34, 15, 2, 43, 37, 9]. In addition, transformer-based inpainting networks [39, 26, 23, 29] significantly escalate the quality of inpainting results through their large network capacity and superior local/global connectivity based on attention mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.775, + 0.471, + 0.866 + ], + "angle": 0, + "content": "However, conventional approaches are unable to fully utilize the semantic information within the input video or distinguish the class-specific characteristics of objects with different semantics. As a result, they frequently fail to recover proper object structure, texture, and scene layout. Several studies [32, 24, 19, 25] have suggested that utiliz" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.271, + 0.892, + 0.332 + ], + "angle": 0, + "content": "ing semantic information can lead to better results and produce visually more plausible images when filling missing regions. However, research on video inpainting has yet to explore the potential of incorporating semantic maps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.895, + 0.488 + ], + "angle": 0, + "content": "Video inpainting is a challenging task that requires to render temporally consistent video frames. To address this issue, optical flow has been extensively explored, although it is computationally intensive [15, 37, 9, 23, 41, 40, 13]. In this study, we demonstrate that semantic information can also be used to enforce temporal consistency, and we introduce a new technique that further facilitates the use of semantic cues in a video by dynamically mixing multiple class-specific experts to handle objects with distinct semantics adaptively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.895, + 0.672 + ], + "angle": 0, + "content": "Although previous semantic-guided inpainting networks used the predicted segmentation map as additional input [32] or developed a dedicated semantic-aware inference module [24] to learn semantic-aware parameters, the restoration performance is limited, given the use of shared network parameters for the different semantics. This problem has been addressed in a recent study [25] by using a semantic-aware attention module that enables the features to attend solely to regions with identical semantic labels. However, the computational cost increases proportionally to the number of classes because the attention mechanism is separately carried out for each category." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To mitigate these problems, we propose a novel semantic-aware dynamic parameter selection approach that effectively utilizes the semantic information within input video frames while retaining the number of operations during the inference phase. Specifically, inspired by Cond-Conv [38], we introduce conditionally parameterized linear operations to learn semantic-aware experts, and produce locally varying (dynamic) parameters by leveraging the given semantic cues based on the notion of mixture-of-experts [31, 30, 38]. Our linear operations with dynamically determined parameters replace the standard feed-forward operation within a conventional transformer block. Notably, we can keep the number of parameters for inference by mixing the expert parameters before performing linear operations rather than aggregating the output features cal" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.875, + 0.23, + 0.887 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.229, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.875, + 0.23, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12949" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "ulated by each expert. Moreover, in contrast to the original CondConv [38] that produces conditional, but locally uniform parameters, ours can generate conditional and locally (token-wise) varying parameters, thus improving each token's representation power." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.47, + 0.288 + ], + "angle": 0, + "content": "Our extensive experiments witness the outstanding performance of the proposed Semantic-Aware Video Inpainting Transformer (SAVIT) in improving video inpainting results. Specifically, SAVIT elevates quantitative performance on conventional video inpainting benchmark datasets (YouTube-VOS [36] and DAVIS [28]) and produces visually more superior results against state-of-the-art methods. We summarize our contribution as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.294, + 0.468, + 0.323 + ], + "angle": 0, + "content": "- We tackle leveraging semantic information within the given input video frames for the video inpainting task." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.326, + 0.468, + 0.355 + ], + "angle": 0, + "content": "- We introduce a novel semantic-aware video inpainting transformer based on a mixture-of-experts scheme." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.36, + 0.468, + 0.388 + ], + "angle": 0, + "content": "- We propose a semantic-aware dynamic linear operation to exploit local semantic cues effectively." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.392, + 0.468, + 0.434 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate the superiority and efficacy of our method, especially in recovering semantic structures and textures." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.294, + 0.468, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.452, + 0.218, + 0.467 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.47, + 0.835 + ], + "angle": 0, + "content": "Video inpainting. An active research topic in video inpainting is generating a temporally consistent video. Various approaches, including 3D-CNN-based, flow-based, and transformer-based methods, have been proposed to fill in a missing region inside a video plausibly. Despite their ability to extract spatio-temporal information, 3D-CNN methods [34, 2, 3] fail to fully utilize global correspondence due to limited temporal receptive field. Flow-based methods [15, 37, 9, 23, 41, 40, 13] first estimate motion flow in the missing region and use the flow to predict the missing pixels. However, these methods highly rely on pre-trained inpainting networks and often require manual operations [37, 9]. Recently, transformer-based methods [39, 26, 23, 29] have attracted research attention due to their superiority in capturing information across long-ranged frames. Built upon the vision transformer architecture [8], FuseFormer [26] proposed soft tokenization with spatial overlapping to interact information among neighboring tokens. Followup studies have improved the performance with dedicated modules such as focal attention [23] or discrete latent mapping [29]. In our study, we develop our method on top of transformers to capture long-range dependencies and advocate the effectiveness of leveraging semantic information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Semantic-guided inpainting. Several works have been proposed to utilize the semantic map for inpainting [32, 24, 25], and its related fields [14, 4, 22, 12]. Among them, SPGNet [32] first fills the missing segmentation map, then" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.303 + ], + "angle": 0, + "content": "synthesizes the inpainted image based on the estimated semantic results. Moreover, SGENet [24] introduced iterative feature refinement using semantic map-based normalization [27] to handle more complex holes in mixed scenes, including objects with different semantics. Recent work [25] proposed to calculate semantic-wise attention to learn the corresponding semantic information. However, most of the previous works have developed upon image inpainting, and leveraging semantic cues within a video has yet to be explored due to the lack of annotated data. Our study mitigates the data acquisition problem by gathering pseudo-annotated videos with pre-trained segmentation network [17], and demonstrates that semantic information can elevate video inpainting performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Dynamic filter. As an alternative to fixed filter, scene-adaptive dynamic filter networks [11, 7, 42] adapt their parameters based on a given input image and/or features. These methods allow flexibility in the network in treating various inputs with different characteristics. Notably, CondInst [33] dynamically generates instance-aware parameters for instance segmentation. In this study, we adapt the transformer parameters in reasoning each token conditioned to its features and semantic information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.598 + ], + "angle": 0, + "content": "Mixture-of-experts. By combining the results from multiple expert models, the network can further improve its representation power [30, 10, 31]. Instead of increasing the computational cost during the inference proportional to the number of experts, recent works [38, 5] parameterized the mixed experts to combine the knowledge efficiently. Our work is motivated by CondConv [38]; we mix the experts' parameters before the feature calculation. Specifically, we divide the experts' roles by semantic categories, then combine semantic-aware experts' knowledge at each token." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.673, + 0.621 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.892, + 0.685 + ], + "angle": 0, + "content": "In this section, we first introduce the annotation procedure to acquire the semantic maps for videos, followed by a detailed description of the proposed framework SAVIT, which leverages semantic information for video inpainting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.688, + 0.734, + 0.703 + ], + "angle": 0, + "content": "3.1. Semantic data acquisition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Unlike previous semantic-aware approaches [32, 25, 24], which simultaneously estimate the segmentation maps for inpainting, we assume that semantic segmentation maps are predictable from the given clean video frames, which include unmasked regions to remove, using off-the-shelf segmentation methods. However, currently available video segmentation datasets, such as DAVIS [28] or CityScape [6], are limited to specific categories and/or particular objects, leading acquiring ground-truth segmentation maps for general video frames difficult. Alternatively, we utilize a fully pre-trained panoptic segmentation network [17] to predict the foreground and background semantics in more general scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12950" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.098, + 0.355, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.304, + 0.318, + 0.318 + ], + "angle": 0, + "content": "(a) Initial segmentation map" + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.096, + 0.655, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.304, + 0.589, + 0.318 + ], + "angle": 0, + "content": "(b) Super-categorize" + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.121, + 0.836, + 0.133 + ], + "angle": 0, + "content": "temporally consistent" + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.138, + 0.829, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.305, + 0.838, + 0.331 + ], + "angle": 0, + "content": "(c) Super-categorized segmentation map" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.331, + 0.892, + 0.377 + ], + "angle": 0, + "content": "Figure 1: Overview of the semantic data preparation: (a) Temporally inconsistent rough prediction using a pre-trained panoptic segmentation model; (b) Grouping raw predictions into corresponding super-categories; (c) Super-categorized segmentation map produces temporally consistent labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.47, + 0.646 + ], + "angle": 0, + "content": "Fig. 1 illustrates the overall semantic data acquisition process used for training and inference. For training, we first apply the panoptic segmentation network [17] to collect estimated segmentation maps for video frames in conventional video inpainting datasets (e.g., YouTube-VOS [36]). The initial segmentation results consist of 53 background and 80 foreground classes [1]. However, these initial segmentation maps with a large number of classes are erroneous and temporally inconsistent, as shown in Fig. 1 (a). Thus, we enforce the temporal consistency of semantics by grouping the predicted labels to the corresponding supercategory of the class. For example, dog, horse, and sheep regions are grouped into a super-category of animal (Fig. 1 (b)). Although we can lose semantic details to some extent by the super-categorization, such constraint helps the training of our semantic-aware inpainting network through the enforced temporal semantic consistency (Fig. 1 (c))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.648, + 0.47, + 0.799 + ], + "angle": 0, + "content": "In our work, we define eight different super-categories based on the texture similarities and the frequencies of categories, which are measured on the entire training dataset. Specifically, we have five super-categories for foreground objects: 'animal,' 'person,' 'vehicle,' 'plants,' and 'sports.' Moreover, we have two super-categories for background according to their texture; 'plain' and 'patterned.' Finally, we group the rest of the remaining sub-categories into 'etc.' During the test phase, we also obtain the semantic maps in the same fashion." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.807, + 0.406, + 0.823 + ], + "angle": 0, + "content": "3.2. Semantic-aware dynamic transformer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To handle diverse scenes with various objects across different classes in videos, we propose a semantic-aware dynamic transformer. In particular, our transformer employs a mixture-of-experts scheme, in which each expert is a single fully connected layer with its parameters reserved for each" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.892, + 0.48 + ], + "angle": 0, + "content": "corresponding semantic class (i.e., super-category) obtained from Section 3.1. Conditioned on the semantic features of each token, we create a token-wise fully connected layer, where its parameters are determined by weighted summing the class-specific parameters, and the weights for the mixture are produced by our semantic router as in the following." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.483, + 0.892, + 0.574 + ], + "angle": 0, + "content": "Learning semantic router. To produce locally (i.e., token-wise) varying parameters conditioned on the semantic information, the network requires a routing function that predicts blending weights to mix the class-specific expert parameters. Our router computes routing weights (i.e., mixture coefficients) for each token by:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.585, + 0.891, + 0.601 + ], + "angle": 0, + "content": "\\[\nr _ {i} = M L P \\left(f _ {i}\\right) + h a r d m a x \\left(s _ {i}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.892, + 0.822 + ], + "angle": 0, + "content": "where \\( r_i \\in \\mathbb{R}^C \\) is a produced mixture weight vector; \\( f_i \\) is the feature of an input token; \\( s_i \\in \\mathbb{R}^C \\) is a histogram of class distribution on the segmentation map for the given \\( i \\)-th token; \\( C \\) is the number of semantic experts (\\( C = 8 \\) in this work unless stated otherwise); and MLP is a standard multi-layer perceptron (MLP) composed of two fully-connected layers with Leaky ReLU activation and sigmoid output. Our router uses the most dominant semantic label from the input histogram \\( s_i \\) by using a hard selection strategy (hardmax operation) rather than a softmax result. In addition, we model our router to be lightweight (i.e., two layers) to minimize the overhead, keeping our framework to have similar computational costs to the standard transformers during inference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Semantic-aware dynamic parameterization. By using the routing weights \\( r_i \\) generated by the router for the \\( i \\)-th token with features \\( f_i \\), we produce semantic-aware dynamic parameters \\( W_i \\) for linear (fully-connected) layers of a transformer block. The formulation of semantic-aware dynamic" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12951" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.094, + 0.418, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.461, + 0.47, + 0.576 + ], + "angle": 0, + "content": "Figure 2: Illustration of SAFFN. With the histogram of semantic categories (e.g., hardmax) and additional linear layers, the semantic router outputs a routing weight for a specific token. Using the routing weight, SAL first acquires token-specific parameters by the weighted sum of experts, then operates the linear transformation of the token with the parameters. Here, we illustrate the operational flow of a single token. \\(\\oplus\\) indicates element-wise summation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.586, + 0.241, + 0.601 + ], + "angle": 0, + "content": "layer (SAL) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.61, + 0.469, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S A L} \\left(f _ {i}\\right) = f _ {i} \\cdot W _ {i}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.636, + 0.457, + 0.651 + ], + "angle": 0, + "content": "where the mixed expert parameter \\( W_{i} \\) for token \\( f_{i} \\) yields:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.658, + 0.469, + 0.699 + ], + "angle": 0, + "content": "\\[\nW _ {i} = \\sum_ {c = 1} ^ {C} r _ {i} (c) \\cdot w ^ {c}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Here, \\( w^{c} \\) indicates the parameters of each class-specific expert, while \\( r_{i}(c) \\) is the \\( c \\)-th element of routing weight vector \\( r_{i} \\) and means the mixture coefficient for the corresponding expert \\( w^{c} \\). Consequently, our semantic-aware dynamic parameterization produces locally varying (token-wise) parameters. For example, a token corresponding to a goat in the input frame can be handled by parameters specialized for animal super-category, and another token on the rock can be dealt with specific parameters for the plain background. Notably, our framework can also handle a token of complex scenes where semantically different objects are coexisting. This is because our router is designed to produce locally varying weights and give high routing weights" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.891, + 0.121 + ], + "angle": 0, + "content": "for experts corresponding to coexisting objects, rather than choosing a single expert." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.122, + 0.892, + 0.182 + ], + "angle": 0, + "content": "Finally, we devise a semantic-aware dynamic feedforward network (SAFFN) as depicted in Fig. 2. SAFFN includes two SAL using the dynamically generated semantic-aware parameters as:" + }, + { + "type": "equation", + "bbox": [ + 0.519, + 0.188, + 0.892, + 0.228 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {S A F F N} \\left(f _ {i}\\right) = \\operatorname {S A L} \\left(\\operatorname {R e L U} \\left(\\operatorname {S A L} \\left(f _ {i}; W _ {i} ^ {l}\\right)\\right); W _ {i} ^ {l + 1}\\right) \\tag {4} \\\\ = \\max \\left(\\mathbf {0}, f _ {i} \\cdot W _ {i} ^ {l}\\right) \\cdot W _ {i} ^ {l + 1}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.234, + 0.892, + 0.295 + ], + "angle": 0, + "content": "In our SAFFN, each SAL is parameterized by our dynamic parameters \\( W_{i}^{l} \\), where \\( l \\) denotes the layer index of the linear transformation, and ReLU activation is included between two layers. Bias parameters are omitted for brevity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.297, + 0.692, + 0.31 + ], + "angle": 0, + "content": "3.3. Overall architecture" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.313, + 0.892, + 0.525 + ], + "angle": 0, + "content": "Fig. 3 illustrates the overall pipeline of the proposed video inpainting network SAVIT. The network is composed of encoder, decoder, and multiple transformer blocks between the encoder and the decoder. First, the encoder produces output features by taking \\( N \\) consecutive video frames with masks and the corresponding semantic segmentation maps as input. We employ the encoder in FuseFormer [26] and additionally concatenate corresponding semantic maps to input. Next, we carry out the tokenization procedure [26], then forward to several transformer blocks to extract powerful features. Finally, the tokenized features are rearranged into the image-like shape, which is then fed into the decoder that predicts \\( N \\) inpainted video frames as previous works [26, 23]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.525, + 0.892, + 0.736 + ], + "angle": 0, + "content": "Note that our semantic-aware mixture-of-experts schemes can be plugged into any conventional transformer-based approach (e.g., FuseFormer [26], E2FGVI [23]), where a conventional transformer block is just replaced with our semantic-aware transformer block. In our experiments, we embed our dynamic transformer within FuseFormer [26]. Specifically, SAVIT contains eight FuseFormer blocks, each of which consists of a multi-head self-attention (MHSA) operation and fusion feed-forward network (F3N). We replace one of the eight FuseFormer blocks with our dynamic transformer block. Precisely, within the F3N in the fourth FuseFormer block, we modify the two standard linear layers with our dynamic ones and add a semantic router." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.739, + 0.892, + 0.905 + ], + "angle": 0, + "content": "Semantic-aware dynamic discriminator. To facilitate each expert's learning of the corresponding category's representation and texture details, we employ a semantic-aware discriminator (SAD) that predicts the real/fake score for each token. Fig. 4 illustrates the architecture of SAD, which has similar structures to the encoder of SAVIT, except for the last layer, where a single SAL is added. Notably, we utilize the same mixture coefficients for dynamic parameters for both SAVIT (generator) and SAD (discriminator), encouraging each expert to learn detailed and discriminative category-specific representation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12952" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.094, + 0.873, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.304, + 0.895, + 0.361 + ], + "angle": 0, + "content": "Figure 3: Overall flow of SAVIT. Our method utilizes additional semantic cues with a mixture-of-experts scheme. The transformer parameters are dynamically adapted to the local semantic information for each token region. With these token-wise dynamic parameters, our network produces more precise inpainting results, generating an object's boundary or synthesizing detailed textures." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.374, + 0.457, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.511, + 0.471, + 0.569 + ], + "angle": 0, + "content": "Figure 4: Schematic illustration of SAD. With the specific mixture-of-experts parameters produced by using the same routing weight from the generator, SAD discriminates whether the inpainted results are realistic or not." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.58, + 0.218, + 0.594 + ], + "angle": 0, + "content": "3.4. Loss function" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.824 + ], + "angle": 0, + "content": "Auxiliary token classification loss. In the mixture-of-experts scheme, it is crucial that each expert is assigned with distinct role [18, 30]. If our semantic experts have been assigned properly to each corresponding category, the output tokens from the semantic-aware dynamic block should be helpful for classification. We drive collaborative role division across our semantic experts. Specifically, we regularize the learning of experts during training by performing auxiliary classification tasks on the output tokens from the semantic-aware dynamic block. Such token-wise classification is performed by an auxiliary single linear layer classifier. The classifier predicts the semantic labels of the output tokens, as illustrated with the green box in Fig. 3. As a result, each expert can effectively learn the feature representation corresponding to a specific semantic label." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.827, + 0.469, + 0.871 + ], + "angle": 0, + "content": "Overall objective function. Our overall objective function \\(\\mathcal{L}_{total}\\) is the weighted sum of the following four different loss terms:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.883, + 0.47, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p i x}} + \\lambda_ {\\text {a d v}} \\mathcal {L} _ {\\text {a d v}} + \\lambda_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}} + \\lambda_ {\\text {s a d}} \\mathcal {L} _ {\\text {s a d}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.37, + 0.892, + 0.461 + ], + "angle": 0, + "content": "where \\(L_{pix}\\) is a pixel-wise reconstruction loss, \\(\\mathcal{L}_{adv}\\) is an adversarial loss, \\(\\mathcal{L}_{cls}\\) is token classification loss, and \\(\\mathcal{L}_{sad}\\) is semantic adversarial loss, while \\(\\{\\lambda_{adv},\\lambda_{cls},\\lambda_{sad}\\}\\) control the weight of associated loss term. Given the predicted video frame \\(\\hat{Y}\\) and its ground truth counterpart \\(Y\\), we employ the L1 pixel-wise reconstruction loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.469, + 0.892, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p i x} = \\left| \\left| \\hat {Y} - Y \\right| \\right| _ {1}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.498, + 0.892, + 0.527 + ], + "angle": 0, + "content": "As for the adversarial loss, we train our inpainting network to fool a discriminator \\( D \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.537, + 0.892, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a d v} = - E _ {\\hat {Y}} [ D (\\hat {Y}) ]. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.565, + 0.891, + 0.595 + ], + "angle": 0, + "content": "Based on a temporal patch discriminator [26, 23], the discriminator \\(D\\) is trained to optimize" + }, + { + "type": "equation", + "bbox": [ + 0.506, + 0.604, + 0.891, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D} = E _ {Y} \\left[ \\operatorname {R e L U} (1 - D (Y)) \\right] + E _ {\\hat {Y}} \\left[ \\operatorname {R e L U} (1 + D (\\hat {Y})) \\right]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.637, + 0.796, + 0.651 + ], + "angle": 0, + "content": "The token classification loss term is given as" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.661, + 0.892, + 0.702 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} = - \\sum_ {i} \\sum_ {c = 1} ^ {C} y _ {i} ^ {c} \\log \\left(\\text {s o f t m a x} \\left(\\hat {y} _ {i}\\right)\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.712, + 0.892, + 0.834 + ], + "angle": 0, + "content": "where \\(\\hat{y}_i\\in \\mathbb{R}^C\\) is the predicted score from the single-layer classifier conditioned on the output tokens from the semantic-aware dynamic block, while \\(y_{i}^{c}\\) is the target probability for class \\(c\\) at the \\(i\\)-th token. For the ground truth target \\(y_{i}^{c}\\), we use the input semantic segmentation map (i.e., \\(y_{i}^{c} = s_{i}(c)\\)). The semantic-aware adversarial loss \\(\\mathcal{L}_{sad}\\) is similar to \\(\\mathcal{L}_{adv}\\) in that (8) and (7) are used, but with the semantic-aware discriminator (SAD) instead of \\(D\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.841, + 0.633, + 0.857 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.859, + 0.892, + 0.905 + ], + "angle": 0, + "content": "In this section, we quantitatively and qualitatively demonstrate the effectiveness of SAVIT. Please refer to the supplementary material for more experimental results." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12953" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.137, + 0.089, + 0.833, + 0.238 + ], + "angle": 0, + "content": "
YouTube-VOS [36]DAVIS [28]
MethodPSNR↑SSIM↑VFID↓E_warp(%)↓PSNR↑SSIM↑VFID↓E_warp(%)↓
VINet [15]29.200.94340.0720.149028.960.94110.1990.1785
LGTSM [3]29.740.95040.0700.185928.570.94090.1700.1640
CAP [21]31.580.96070.0710.147030.280.95210.1820.1533
STTN [39]32.340.96550.0530.090730.670.95600.1490.1449
E2FGVI [23]33.710.97000.0460.086433.010.97210.1160.1315
ISVI [41]31.190.95690.0530.117332.280.96690.1290.1589
FuseFormer [26]33.160.96730.0510.090032.540.97000.1380.1362
SAVIT (Ours)33.970.97270.0430.043633.140.97480.1070.0673
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.241, + 0.895, + 0.288 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison of SAVIT against state-of-the-art methods. SAVIT outperforms the existing approaches in terms of various evaluation metrics. Notably, SAVIT significantly elevates the inpainting performance of FuseFormer [26], which is our baseline architecture, witnessing the superiority of leveraging semantic cues within the input video." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.296, + 0.449, + 0.382 + ], + "angle": 0, + "content": "
ConfigurationPSNR↑VFID↓
AFuseFormersmall [26]30.510.182
B+ super-categorized segmentation map30.900.161
C+ SAL30.920.164
D+ token classification loss (λcls)31.190.165
E+ SAD31.010.157
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.385, + 0.472, + 0.472 + ], + "angle": 0, + "content": "Table 2: Ablation experiments on various network configurations under DAVIS [28] dataset. Starting from the performance of the baseline architecture [26], the inpainting performance gradually increases as the configurations are added since the network can benefit from the semantic cues within a video." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.487, + 0.292, + 0.503 + ], + "angle": 0, + "content": "4.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.47, + 0.717 + ], + "angle": 0, + "content": "Datasets and evaluation metrics. We use two conventional video inpainting datasets, YouTube-VOS [36] and DAVIS [28], for our experiments. YouTube-VOS dataset consists of 3471 training, 474 validation, and 508 test videos from mixed scenes. Meanwhile, DAVIS dataset comprises 150 video clips, 60 of which are used for training and the remaining 90 clips are used for evaluation. Following previous works [26, 23], we use only the training videos in the YouTube-VOS for training, while the evaluation is conducted on YouTube-VOS testset and 50 selected clips of DAVIS testset. During training, we create random mask shapes for input masked frames as in [39, 26, 23] and use the method discussed in Section 3.1 to obtain a segmentation map for the entire dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.869 + ], + "angle": 0, + "content": "The inpainting performance is evaluated under two standard scenarios: fixed region inpainting and foreground object removal. For the fixed region inpainting, we use free-form fixed masks for each video provided by FuseFormer [26]. As for the foreground object removal, we remove the main objects using frame-by-frame annotations provided by the stakeholders. Notably, we employ a simple k-nearest neighbors algorithm to complete the missing object region in the segmentation maps for the object removal task (please refer to supplementary material for detail)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We compare the results with respect to four evaluation metrics: PSNR, SSIM, temporal warping error \\( E_{\\text{warp}} \\) [20]," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.299, + 0.892, + 0.39 + ], + "angle": 0, + "content": "and Video-based Fréchet Inception Distance (VFID) [35]. PSNR and SSIM are traditional metrics used to measure the average quality of produced frames. VFID and \\( E_{\\text{warp}} \\) measure the video's perceptual quality and temporal stability (consistency), respectively, and low values indicate better quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.394, + 0.893, + 0.56 + ], + "angle": 0, + "content": "Training settings. We follow the training settings of our baseline network FuseFormer [26]. Specifically, we randomly sample five frames from a video, resize the frames to \\(240 \\times 432\\), and utilize random horizontal flipping augmentation. The network is trained for \\(500k\\) iterations with Adam optimizer [16] and a batch size of eight. The initial learning rate is 1e-4 and decreases to 1e-5 after \\(400k\\) iterations. The coefficients for our loss functions, \\(\\lambda_{adv}\\), \\(\\lambda_{cls}\\), and \\(\\lambda_{sad}\\) are 0.01, 0.01, and 0.001, respectively. We implement our network on the PyTorch framework and use eight NVIDIA RTX A6000 GPUs for training." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.563, + 0.72, + 0.579 + ], + "angle": 0, + "content": "4.2. Video inpainting results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.893, + 0.807 + ], + "angle": 0, + "content": "Quantitative results. Table 1 quantitatively compares our video inpainting results on fixed region inpainting setting against seven state-of-the-art video inpainting networks: VINet [15], LGTSM [3], CAP [21], STTN [39], E2FVGI [23], ISVI [41], and FuseFormer [26]. The result shows that our method remarkably elevates the performance of FuseFormer [26], on which we base our architecture, and achieves superior performances on both YouTube-VOS and DAVIS datasets compared to the state-of-the-art methods. Notably, the warping error has also significantly improved, indicating that semantic information benefits in generating temporally consistent video. These results show the effectiveness of employing mixture-of-experts to dynamically leverage the semantic cues within given frames for video inpainting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Qualitative results. Fig. 5 shows that SAVIT has visually pleasing inpainting results, especially in synthesizing semantic structure, boundary, and texture of objects. Notably, unlike the baseline methods that rely solely on inherent video clues and cannot accurately restore the shape of the human body, SAVIT effectively completes the region" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12954" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.062, + 0.257, + 0.074 + ], + "angle": 0, + "content": "Masked Frames" + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.062, + 0.365, + 0.073 + ], + "angle": 0, + "content": "STTN" + }, + { + "type": "image_caption", + "bbox": [ + 0.446, + 0.062, + 0.518, + 0.074 + ], + "angle": 0, + "content": "FuseFormer" + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.062, + 0.643, + 0.074 + ], + "angle": 0, + "content": "E2FGVI" + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.062, + 0.796, + 0.075 + ], + "angle": 0, + "content": "SAVIT (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.075, + 0.276, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.075, + 0.419, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.075, + 0.556, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.075, + 0.691, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.075, + 0.829, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.136, + 0.276, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.136, + 0.419, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.136, + 0.555, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.136, + 0.691, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.136, + 0.829, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.195, + 0.276, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.195, + 0.419, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.195, + 0.555, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.195, + 0.691, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.195, + 0.829, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.255, + 0.276, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.255, + 0.419, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.255, + 0.555, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.255, + 0.691, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.255, + 0.829, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.315, + 0.276, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.315, + 0.419, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.315, + 0.555, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.315, + 0.691, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.315, + 0.829, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.375, + 0.276, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.375, + 0.419, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.375, + 0.555, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.375, + 0.691, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.375, + 0.829, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.448, + 0.892, + 0.478 + ], + "angle": 0, + "content": "Figure 5: Qualitative comparison of SAVIT with baseline video inpainting networks in fixed region inpainting and object removal. SAVIT more accurately recovers the object's boundaries and appropriate textures." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.485, + 0.176, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.559, + 0.187, + 0.588 + ], + "angle": 0, + "content": "Figure 6: User study result." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.489, + 0.437, + 0.555 + ], + "angle": 0, + "content": "
ConfigurationPSNR†
hardmax (SAVIT)31.01
softmax30.91
random routing weight19.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.208, + 0.559, + 0.465, + 0.588 + ], + "angle": 0, + "content": "Table 3: Effect of the routing weight on the DAVIS [28] dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.47, + 0.68 + ], + "angle": 0, + "content": "by leveraging additional semantic information (third and fourth rows). Furthermore, SAVIT demonstrates the capability of generating more semantically plausible textures, as observed from the detailed textures of synthesized grass and wall (second and fifth rows)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.839 + ], + "angle": 0, + "content": "User study. We conduct a user study on the foreground object removal task, where ground truth target video frames for evaluation are not available, to investigate the real-world applicability of our method using DAVIS [28] testset. We ask 24 raters to choose a better video between two randomly ordered videos, including the inpainted results by FuseFormer [26] and ours, fifteen times per rater. The result presented in Fig. 6 shows that raters prefer ours to the baseline, indicating the impact of exploiting semantic cues in completing missing object regions." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.238, + 0.864 + ], + "angle": 0, + "content": "5. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.868, + 0.469, + 0.914 + ], + "angle": 0, + "content": "We perform ablation experiments based on a smaller FuseFormer network (FuseFormersmall) by halving the dimension of the hidden layer." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.49, + 0.694, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.574, + 0.638, + 0.584 + ], + "angle": 0, + "content": "SAVIT w/o SAD" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.489, + 0.892, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.759, + 0.575, + 0.833, + 0.584 + ], + "angle": 0, + "content": "SAVIT w/ SAD" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.587, + 0.892, + 0.63 + ], + "angle": 0, + "content": "Figure 7: Effect of semantic-aware discriminator (SAD). The texture of the black swan's body is blurry in the case without SAD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Network configuration. Table 2 compares the inpainting performance for various configurations of SAVIT. First, we observe that the baseline network's performance improves with additional segmentation map inputs (rows A and B). However, a subtle additional performance improvement (0.02dB PSNR gain, rows B and C) is achieved when we naively add our proposed dynamic linear layer to the network. Rather, applying the proposed token-wise classification loss to the objective boosts the performance (0.2dB PSNR gain, row D). Overall performance improvement by applying mixture-of-experts (rows C and D) highlights the importance and effectiveness of employing mixture-of-experts to handle semantics dynamically. Moreover, using a semantic-aware dynamic discriminator (row E) degrades PSNR while improving VFID. However, as shown in Fig. 7, utilizing the semantic-aware discriminator produces more visually plausible results, which solidifies the advantage of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12955" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.086, + 0.182, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.087, + 0.285, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.087, + 0.385, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.087, + 0.487, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.087, + 0.587, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.087, + 0.687, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.146, + 0.182, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.146, + 0.285, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.146, + 0.385, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.146, + 0.487, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.146, + 0.587, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.146, + 0.687, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.2, + 0.688, + 0.245 + ], + "angle": 0, + "content": "Figure 8: Visualization of the semantic router results. (1) to (8) are the visualization of routing weights of the corresponding super-category. The routing weights are rearranged to the image-like features. Brighter pixel values indicate higher routing weights." + }, + { + "type": "table", + "bbox": [ + 0.698, + 0.089, + 0.887, + 0.154 + ], + "angle": 0, + "content": "
# ExpertsPSNR↑VFID↓
430.830.165
630.900.168
831.010.157
" + }, + { + "type": "table_caption", + "bbox": [ + 0.696, + 0.157, + 0.888, + 0.244 + ], + "angle": 0, + "content": "Table 4: Inpainting results by changing the numbers of experts under the DAVIS [28] dataset. More experts render better inpainting results." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.255, + 0.276, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.307, + 0.276, + 0.351 + ], + "angle": 0, + "content": "Figure 9: Qualitative comparison on different routing weight strategies." + }, + { + "type": "table", + "bbox": [ + 0.286, + 0.256, + 0.465, + 0.304 + ], + "angle": 0, + "content": "
# blocks# params.PSNR↑
126.0M30.40
233.1M30.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.285, + 0.308, + 0.465, + 0.351 + ], + "angle": 0, + "content": "Table 5: Ablation on different number of semantic blocks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.371, + 0.406, + 0.386 + ], + "angle": 0, + "content": "the discriminator in improving perceptual quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.469, + 0.542 + ], + "angle": 0, + "content": "Visualization of routing weights. Fig. 8 visualizes the routing weights of each expert from the trained semantic router. We rearrange the tokens to image-like features for visualization, and brighter values indicate higher routing weights in the corresponding region. The bright regions among different experts indicate that the experts collaborate during the inpainting procedure rather than simply utilizing a specific expert. We see that routing weights tend to change locally smoothly while the input segmentation map changes abruptly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.469, + 0.741 + ], + "angle": 0, + "content": "Designing choice on semantic router. Table 3 compares different design choices of our semantic router. The performance drop by replacing the trained routing weights (i.e., mixture coefficients) with random weights indicates the discriminative role of trained experts. Moreover, the results of softmax routing strategy show a noticeable performance degradation against hardmax strategy quantitatively. Fig. 9 also shows that softmax results tend to have blurry texture, especially on the edge area. Therefore, Hardmax was chosen for our final model to facilitate role division between experts better, as it enables the learning of distinct parameters by dividing the input for each expert before the router blends the token based on its features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.746, + 0.469, + 0.837 + ], + "angle": 0, + "content": "Impact of the number of experts. Table 4 compares the inpainting performance by changing the number of experts. Notably, as the number of experts (i.e., semantic categories) increases, the network can benefit from more diverse semantic clues in terms of PSNR and VFID values, and we set the number of experts for our final model to eight." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Impact of the number of semantic blocks. Table 5 compares the inpainting performance by applying our semantic module on a different number of blocks. The result indicates that using more semantic blocks produces a slightly" + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.256, + 0.851, + 0.331 + ], + "angle": 0, + "content": "
MethodMask ratioPSNR†VFID↓
FuseFormersmall30%22.570.554
60%18.101.059
FuseFormersmall + Ours30%23.570.412
60%19.210.750
" + }, + { + "type": "table_caption", + "bbox": [ + 0.55, + 0.334, + 0.842, + 0.348 + ], + "angle": 0, + "content": "Table 6: Comparison on varying mask ratio." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.365, + 0.892, + 0.426 + ], + "angle": 0, + "content": "better inpainting performance. Although we apply our semantic module one time due to hardware limits for our final model, it is expected that the results will be naturally improved by increasing the number of semantic blocks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.44, + 0.892, + 0.516 + ], + "angle": 0, + "content": "Input video with large corruption. Table 6 compares our method against FuseFormersmall under different corrupted region ratio settings. It is shown that, despite larger masks, our method still outperforms the baseline by a large margin by exploiting the semantic information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.53, + 0.777, + 0.546 + ], + "angle": 0, + "content": "6. Conclusion and Future Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.763 + ], + "angle": 0, + "content": "In this study, we have explored leveraging semantic information within a given video for transformer-based video inpainting. Accordingly, we propose an effective semantic-aware dynamic transformer with the notion of mixture-of-experts. Our dynamic transformer contains class-specific expert parameters where experts are efficiently combined to perform token-wise dynamic linear operations based on the token feature and its semantic information. The experiments demonstrate the effectiveness of our method, especially in recovering semantic structures. Although we have utilized pseudo labels for our study, we believe more accurate and denser annotated datasets can improve the inpainting performance further since our method relies on the quality of the segmentation label." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.783, + 0.667, + 0.8 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.811, + 0.892, + 0.899 + ], + "angle": 0, + "content": "This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2023-00222776) and Samsung Electronics Co., Ltd, and Samsung Research Funding Center of Samsung Electronics under Project Number SRFCIT1901-06." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12956" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.239 + ], + "angle": 0, + "content": "[2] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.241, + 0.47, + 0.297 + ], + "angle": 0, + "content": "[3] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Learnable gated temporal shift module for deep video inpainting. In Proceedings of the British Machine Vision Conference (BMVC), 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.47, + 0.367 + ], + "angle": 0, + "content": "[4] Cheng Chen, Jiayin Cai, Yao Hu, Xu Tang, Xinggang Wang, Chun Yuan, Xiang Bai, and Song Bai. Deep interactive video inpainting: An invisibility cloak for harry potter. In Proceedings of the ACM International Conference on Multimedia (ACM MM), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.47, + 0.436 + ], + "angle": 0, + "content": "[5] Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolution: Attention over convolution kernels. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.521 + ], + "angle": 0, + "content": "[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.47, + 0.577 + ], + "angle": 0, + "content": "[7] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.578, + 0.47, + 0.661 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.662, + 0.47, + 0.717 + ], + "angle": 0, + "content": "[9] Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.718, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[10] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 1991. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.761, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[11] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. In Advances in Neural Information Processing Systems (NeurIPS), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.803, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[12] Jun-Gyu Jin, Jaehyun Bae, Han-gyul Baek, and Sang-hyo Park. Object-ratio-preserving video retargeting framework based on segmentation and inpainting. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.872, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[13] Jaeyeon Kang, Seoung Wug Oh, and Seon Joo Kim. Error compensation framework for flow-guided video inpainting." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[14] Bholeshwar Khurana, Soumya Ranjan Dash, Abhishek Bhatia, Aniruddha Mahapatra, Hrituraj Singh, and Kuldeep Kulkarni. Semie: Semantically-aware image extrapolation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[15] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Deep video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.246, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[17] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[18] Xiangtao Kong, Hengyuan Zhao, Yu Qiao, and Chao Dong. Classsr: A general framework to accelerate super-resolution networks by data characteristic. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[19] Avisek Lahiri, Arnav Kumar Jain, Sanskar Agrawal, Pabitra Mitra, and Prabir Kumar Biswas. Prior guided gan based semantic inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.483, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[20] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[21] Sungho Lee, Seoul Wug Oh, DaeYeun Won, and Seon Joo Kim. Copy-and-paste networks for deep video inpainting. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[22] Jiacheng Li, Chang Chen, and Zhiwei Xiong. Contextual outpainting with object-level contrastive learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[23] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.719, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[24] Liang Liao, Jing Xiao, Zheng Wang, Chia-Wen Lin, and Shin'ichi Satoh. Guidance and evaluation: Semantic-aware image inpainting for mixed scenes. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[25] Liang Liao, Jing Xiao, Zheng Wang, Chia-Wen Lin, and Shin'ichi Satoh. Image inpainting guided by coherence priors of semantics and textures. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[26] Rui Liu, Hanming Deng, Yangyi Huang, Xiaoyu Shi, Lewei Lu, Wenxiu Sun, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fuseformer: Fusing fine-grained information in" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12957" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "transformers for video inpainting. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[27] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[28] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[29] Jingjing Ren, Qingqing Zheng, Yuanyuan Zhao, Xuemiao Xu, and Chen Li. Dlformer: Discrete latent transformer for video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.342, + 0.471, + 0.412 + ], + "angle": 0, + "content": "[30] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[31] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In Proceedings of the International Conference on Learning Representations (ICLR), 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.471, + 0.55 + ], + "angle": 0, + "content": "[32] Yuhang Song, Chao Yang, Yeji Shen, Peng Wang, Qin Huang, and C-C Jay Kuo. Spg-net: Segmentation prediction and guidance network for image inpainting. In Proceedings of the British Machine Vision Conference (BMVC), 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.471, + 0.594 + ], + "angle": 0, + "content": "[33] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.471, + 0.649 + ], + "angle": 0, + "content": "[34] Chuan Wang, Haibin Huang, Xiaoguang Han, and Jue Wang. Video inpainting by jointly learning temporal structure and spatial details. In Association for the Advancement of Artificial Intelligence (AAAI), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.471, + 0.705 + ], + "angle": 0, + "content": "[35] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[36] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.471, + 0.816 + ], + "angle": 0, + "content": "[37] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep flow-guided video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[38] Brandon Yang, Gabriel Bender, Quoc V Le, and Jiquan Ngiam. Condconv: Conditionally parameterized convolutions for efficient inference. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[39] Yanhong Zeng, Jianlong Fu, and Hongyang Chao. Learning joint spatial-temporal transformations for video inpainting." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[40] Kaidong Zhang, Jingjing Fu, and Dong Liu. Flow-guided transformer for video inpainting. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[41] Kaidong Zhang, Jingjing Fu, and Dong Liu. Inertia-guided flow completion and style fusion for video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.894, + 0.29 + ], + "angle": 0, + "content": "[42] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.894, + 0.347 + ], + "angle": 0, + "content": "[43] Xueyan Zou, Linjie Yang, Ding Liu, and Yong Jae Lee. Progressive temporal feature alignment network for video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12958" + } + ] +] \ No newline at end of file diff --git a/2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/706f36ef-c7c5-4ea7-8481-e3a3bb79a94c_origin.pdf b/2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/706f36ef-c7c5-4ea7-8481-e3a3bb79a94c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f155490c66e6939023002bb2e59d12cc2a2c5e42 --- /dev/null +++ b/2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/706f36ef-c7c5-4ea7-8481-e3a3bb79a94c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f98da5dd607f7198d8257b265b39db0b86beaad76cf43d71e327fff047225a35 +size 6988144 diff --git a/2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/full.md b/2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e38b75e80e3514f6083f00e79c673a87e06ff4eb --- /dev/null +++ b/2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/full.md @@ -0,0 +1,394 @@ +# Semantic-Aware Dynamic Parameter for Video Inpainting Transformer + +Eunhye Lee $^{1*}$ , Jinsu Yoo $^{2*}$ , Yunjeong Yang $^{2}$ , Sungyong Baik $^{2,3}$ , Tae Hyun Kim $^{1\dagger}$ +{dldms1345, jinsuyoo, yunjoeongyang, dsybaik, taehyunkim}@hanyang.ac.kr + $^{1}$ Dept. of Computer Science, $^{2}$ Dept. of Artificial Intelligence, $^{3}$ Dept. of Data Science, Hanyang University + +# Abstract + +Recent learning-based video inpainting approaches have achieved considerable progress. However, they still cannot fully utilize semantic information within the video frames and predict improper scene layout, failing to restore clear object boundaries for mixed scenes. To mitigate this problem, we introduce a new transformer-based video inpainting technique that can exploit semantic information within the input and considerably improve reconstruction quality. In this study, we use the mixture-of-experts scheme and train multiple experts to handle mixed scenes, including various semantics. We leverage these multiple experts and produce locally (token-wise) different network parameters to achieve semantic-aware inpainting results. Extensive experiments on YouTube-VOS and DAVIS benchmark datasets demonstrate that, compared with existing conventional video inpainting approaches, the proposed method has superior performance in synthesizing visually pleasing videos with much clearer semantic structures and textures. + +# 1. Introduction + +A constant increase in demand for video content in our daily lives (e.g., YouTube or TikTok) has led to the development of video inpainting methods, which aim to complete a missing region or erase unwanted areas such as watermarks and captions from a given input video. In particular, recent convolutional neural networks (CNNs) for video inpainting tasks have shown promising results [34, 15, 2, 43, 37, 9]. In addition, transformer-based inpainting networks [39, 26, 23, 29] significantly escalate the quality of inpainting results through their large network capacity and superior local/global connectivity based on attention mechanisms. + +However, conventional approaches are unable to fully utilize the semantic information within the input video or distinguish the class-specific characteristics of objects with different semantics. As a result, they frequently fail to recover proper object structure, texture, and scene layout. Several studies [32, 24, 19, 25] have suggested that utiliz + +ing semantic information can lead to better results and produce visually more plausible images when filling missing regions. However, research on video inpainting has yet to explore the potential of incorporating semantic maps. + +Video inpainting is a challenging task that requires to render temporally consistent video frames. To address this issue, optical flow has been extensively explored, although it is computationally intensive [15, 37, 9, 23, 41, 40, 13]. In this study, we demonstrate that semantic information can also be used to enforce temporal consistency, and we introduce a new technique that further facilitates the use of semantic cues in a video by dynamically mixing multiple class-specific experts to handle objects with distinct semantics adaptively. + +Although previous semantic-guided inpainting networks used the predicted segmentation map as additional input [32] or developed a dedicated semantic-aware inference module [24] to learn semantic-aware parameters, the restoration performance is limited, given the use of shared network parameters for the different semantics. This problem has been addressed in a recent study [25] by using a semantic-aware attention module that enables the features to attend solely to regions with identical semantic labels. However, the computational cost increases proportionally to the number of classes because the attention mechanism is separately carried out for each category. + +To mitigate these problems, we propose a novel semantic-aware dynamic parameter selection approach that effectively utilizes the semantic information within input video frames while retaining the number of operations during the inference phase. Specifically, inspired by Cond-Conv [38], we introduce conditionally parameterized linear operations to learn semantic-aware experts, and produce locally varying (dynamic) parameters by leveraging the given semantic cues based on the notion of mixture-of-experts [31, 30, 38]. Our linear operations with dynamically determined parameters replace the standard feed-forward operation within a conventional transformer block. Notably, we can keep the number of parameters for inference by mixing the expert parameters before performing linear operations rather than aggregating the output features cal + +ulated by each expert. Moreover, in contrast to the original CondConv [38] that produces conditional, but locally uniform parameters, ours can generate conditional and locally (token-wise) varying parameters, thus improving each token's representation power. + +Our extensive experiments witness the outstanding performance of the proposed Semantic-Aware Video Inpainting Transformer (SAVIT) in improving video inpainting results. Specifically, SAVIT elevates quantitative performance on conventional video inpainting benchmark datasets (YouTube-VOS [36] and DAVIS [28]) and produces visually more superior results against state-of-the-art methods. We summarize our contribution as follows: + +- We tackle leveraging semantic information within the given input video frames for the video inpainting task. +- We introduce a novel semantic-aware video inpainting transformer based on a mixture-of-experts scheme. +- We propose a semantic-aware dynamic linear operation to exploit local semantic cues effectively. +- Extensive experiments demonstrate the superiority and efficacy of our method, especially in recovering semantic structures and textures. + +# 2. Related Work + +Video inpainting. An active research topic in video inpainting is generating a temporally consistent video. Various approaches, including 3D-CNN-based, flow-based, and transformer-based methods, have been proposed to fill in a missing region inside a video plausibly. Despite their ability to extract spatio-temporal information, 3D-CNN methods [34, 2, 3] fail to fully utilize global correspondence due to limited temporal receptive field. Flow-based methods [15, 37, 9, 23, 41, 40, 13] first estimate motion flow in the missing region and use the flow to predict the missing pixels. However, these methods highly rely on pre-trained inpainting networks and often require manual operations [37, 9]. Recently, transformer-based methods [39, 26, 23, 29] have attracted research attention due to their superiority in capturing information across long-ranged frames. Built upon the vision transformer architecture [8], FuseFormer [26] proposed soft tokenization with spatial overlapping to interact information among neighboring tokens. Followup studies have improved the performance with dedicated modules such as focal attention [23] or discrete latent mapping [29]. In our study, we develop our method on top of transformers to capture long-range dependencies and advocate the effectiveness of leveraging semantic information. + +Semantic-guided inpainting. Several works have been proposed to utilize the semantic map for inpainting [32, 24, 25], and its related fields [14, 4, 22, 12]. Among them, SPGNet [32] first fills the missing segmentation map, then + +synthesizes the inpainted image based on the estimated semantic results. Moreover, SGENet [24] introduced iterative feature refinement using semantic map-based normalization [27] to handle more complex holes in mixed scenes, including objects with different semantics. Recent work [25] proposed to calculate semantic-wise attention to learn the corresponding semantic information. However, most of the previous works have developed upon image inpainting, and leveraging semantic cues within a video has yet to be explored due to the lack of annotated data. Our study mitigates the data acquisition problem by gathering pseudo-annotated videos with pre-trained segmentation network [17], and demonstrates that semantic information can elevate video inpainting performance. + +Dynamic filter. As an alternative to fixed filter, scene-adaptive dynamic filter networks [11, 7, 42] adapt their parameters based on a given input image and/or features. These methods allow flexibility in the network in treating various inputs with different characteristics. Notably, CondInst [33] dynamically generates instance-aware parameters for instance segmentation. In this study, we adapt the transformer parameters in reasoning each token conditioned to its features and semantic information. + +Mixture-of-experts. By combining the results from multiple expert models, the network can further improve its representation power [30, 10, 31]. Instead of increasing the computational cost during the inference proportional to the number of experts, recent works [38, 5] parameterized the mixed experts to combine the knowledge efficiently. Our work is motivated by CondConv [38]; we mix the experts' parameters before the feature calculation. Specifically, we divide the experts' roles by semantic categories, then combine semantic-aware experts' knowledge at each token. + +# 3. Proposed Method + +In this section, we first introduce the annotation procedure to acquire the semantic maps for videos, followed by a detailed description of the proposed framework SAVIT, which leverages semantic information for video inpainting. + +# 3.1. Semantic data acquisition + +Unlike previous semantic-aware approaches [32, 25, 24], which simultaneously estimate the segmentation maps for inpainting, we assume that semantic segmentation maps are predictable from the given clean video frames, which include unmasked regions to remove, using off-the-shelf segmentation methods. However, currently available video segmentation datasets, such as DAVIS [28] or CityScape [6], are limited to specific categories and/or particular objects, leading acquiring ground-truth segmentation maps for general video frames difficult. Alternatively, we utilize a fully pre-trained panoptic segmentation network [17] to predict the foreground and background semantics in more general scenes. + +![](images/19747dbd49d052ab8d77d77bccd7236ed356a66ac7a4be1bae95457e866bafa9.jpg) +(a) Initial segmentation map + +![](images/586640cd863417072ec09c23f2d6109ec25d0599f18f8ed8b3b60c0160b7e41b.jpg) +(b) Super-categorize +Figure 1: Overview of the semantic data preparation: (a) Temporally inconsistent rough prediction using a pre-trained panoptic segmentation model; (b) Grouping raw predictions into corresponding super-categories; (c) Super-categorized segmentation map produces temporally consistent labels. + +![](images/51008e07925610910e6c24fb9b9ab66fcc56ead2769c1b726f357d65046c93b0.jpg) +temporally consistent +(c) Super-categorized segmentation map + +Fig. 1 illustrates the overall semantic data acquisition process used for training and inference. For training, we first apply the panoptic segmentation network [17] to collect estimated segmentation maps for video frames in conventional video inpainting datasets (e.g., YouTube-VOS [36]). The initial segmentation results consist of 53 background and 80 foreground classes [1]. However, these initial segmentation maps with a large number of classes are erroneous and temporally inconsistent, as shown in Fig. 1 (a). Thus, we enforce the temporal consistency of semantics by grouping the predicted labels to the corresponding supercategory of the class. For example, dog, horse, and sheep regions are grouped into a super-category of animal (Fig. 1 (b)). Although we can lose semantic details to some extent by the super-categorization, such constraint helps the training of our semantic-aware inpainting network through the enforced temporal semantic consistency (Fig. 1 (c)). + +In our work, we define eight different super-categories based on the texture similarities and the frequencies of categories, which are measured on the entire training dataset. Specifically, we have five super-categories for foreground objects: 'animal,' 'person,' 'vehicle,' 'plants,' and 'sports.' Moreover, we have two super-categories for background according to their texture; 'plain' and 'patterned.' Finally, we group the rest of the remaining sub-categories into 'etc.' During the test phase, we also obtain the semantic maps in the same fashion. + +# 3.2. Semantic-aware dynamic transformer + +To handle diverse scenes with various objects across different classes in videos, we propose a semantic-aware dynamic transformer. In particular, our transformer employs a mixture-of-experts scheme, in which each expert is a single fully connected layer with its parameters reserved for each + +corresponding semantic class (i.e., super-category) obtained from Section 3.1. Conditioned on the semantic features of each token, we create a token-wise fully connected layer, where its parameters are determined by weighted summing the class-specific parameters, and the weights for the mixture are produced by our semantic router as in the following. + +Learning semantic router. To produce locally (i.e., token-wise) varying parameters conditioned on the semantic information, the network requires a routing function that predicts blending weights to mix the class-specific expert parameters. Our router computes routing weights (i.e., mixture coefficients) for each token by: + +$$ +r _ {i} = M L P \left(f _ {i}\right) + h a r d m a x \left(s _ {i}\right), \tag {1} +$$ + +where $r_i \in \mathbb{R}^C$ is a produced mixture weight vector; $f_i$ is the feature of an input token; $s_i \in \mathbb{R}^C$ is a histogram of class distribution on the segmentation map for the given $i$ -th token; $C$ is the number of semantic experts ( $C = 8$ in this work unless stated otherwise); and MLP is a standard multi-layer perceptron (MLP) composed of two fully-connected layers with Leaky ReLU activation and sigmoid output. Our router uses the most dominant semantic label from the input histogram $s_i$ by using a hard selection strategy (hardmax operation) rather than a softmax result. In addition, we model our router to be lightweight (i.e., two layers) to minimize the overhead, keeping our framework to have similar computational costs to the standard transformers during inference. + +Semantic-aware dynamic parameterization. By using the routing weights $r_i$ generated by the router for the $i$ -th token with features $f_i$ , we produce semantic-aware dynamic parameters $W_i$ for linear (fully-connected) layers of a transformer block. The formulation of semantic-aware dynamic + +![](images/4b3569247df550c85411f6c17597e8cda4bb180c0b91273b3fecaaf60d7fad56.jpg) +Figure 2: Illustration of SAFFN. With the histogram of semantic categories (e.g., hardmax) and additional linear layers, the semantic router outputs a routing weight for a specific token. Using the routing weight, SAL first acquires token-specific parameters by the weighted sum of experts, then operates the linear transformation of the token with the parameters. Here, we illustrate the operational flow of a single token. $\oplus$ indicates element-wise summation. + +layer (SAL) is given by: + +$$ +\operatorname {S A L} \left(f _ {i}\right) = f _ {i} \cdot W _ {i}, \tag {2} +$$ + +where the mixed expert parameter $W_{i}$ for token $f_{i}$ yields: + +$$ +W _ {i} = \sum_ {c = 1} ^ {C} r _ {i} (c) \cdot w ^ {c}. \tag {3} +$$ + +Here, $w^{c}$ indicates the parameters of each class-specific expert, while $r_{i}(c)$ is the $c$ -th element of routing weight vector $r_{i}$ and means the mixture coefficient for the corresponding expert $w^{c}$ . Consequently, our semantic-aware dynamic parameterization produces locally varying (token-wise) parameters. For example, a token corresponding to a goat in the input frame can be handled by parameters specialized for animal super-category, and another token on the rock can be dealt with specific parameters for the plain background. Notably, our framework can also handle a token of complex scenes where semantically different objects are coexisting. This is because our router is designed to produce locally varying weights and give high routing weights + +for experts corresponding to coexisting objects, rather than choosing a single expert. + +Finally, we devise a semantic-aware dynamic feedforward network (SAFFN) as depicted in Fig. 2. SAFFN includes two SAL using the dynamically generated semantic-aware parameters as: + +$$ +\begin{array}{l} \operatorname {S A F F N} \left(f _ {i}\right) = \operatorname {S A L} \left(\operatorname {R e L U} \left(\operatorname {S A L} \left(f _ {i}; W _ {i} ^ {l}\right)\right); W _ {i} ^ {l + 1}\right) \tag {4} \\ = \max \left(\mathbf {0}, f _ {i} \cdot W _ {i} ^ {l}\right) \cdot W _ {i} ^ {l + 1}. \\ \end{array} +$$ + +In our SAFFN, each SAL is parameterized by our dynamic parameters $W_{i}^{l}$ , where $l$ denotes the layer index of the linear transformation, and ReLU activation is included between two layers. Bias parameters are omitted for brevity. + +# 3.3. Overall architecture + +Fig. 3 illustrates the overall pipeline of the proposed video inpainting network SAVIT. The network is composed of encoder, decoder, and multiple transformer blocks between the encoder and the decoder. First, the encoder produces output features by taking $N$ consecutive video frames with masks and the corresponding semantic segmentation maps as input. We employ the encoder in FuseFormer [26] and additionally concatenate corresponding semantic maps to input. Next, we carry out the tokenization procedure [26], then forward to several transformer blocks to extract powerful features. Finally, the tokenized features are rearranged into the image-like shape, which is then fed into the decoder that predicts $N$ inpainted video frames as previous works [26, 23]. + +Note that our semantic-aware mixture-of-experts schemes can be plugged into any conventional transformer-based approach (e.g., FuseFormer [26], E2FGVI [23]), where a conventional transformer block is just replaced with our semantic-aware transformer block. In our experiments, we embed our dynamic transformer within FuseFormer [26]. Specifically, SAVIT contains eight FuseFormer blocks, each of which consists of a multi-head self-attention (MHSA) operation and fusion feed-forward network (F3N). We replace one of the eight FuseFormer blocks with our dynamic transformer block. Precisely, within the F3N in the fourth FuseFormer block, we modify the two standard linear layers with our dynamic ones and add a semantic router. + +Semantic-aware dynamic discriminator. To facilitate each expert's learning of the corresponding category's representation and texture details, we employ a semantic-aware discriminator (SAD) that predicts the real/fake score for each token. Fig. 4 illustrates the architecture of SAD, which has similar structures to the encoder of SAVIT, except for the last layer, where a single SAL is added. Notably, we utilize the same mixture coefficients for dynamic parameters for both SAVIT (generator) and SAD (discriminator), encouraging each expert to learn detailed and discriminative category-specific representation. + +![](images/c92922e1b79655c66e6835ca0662a9d7faa306d7f5c9e69c77aa9099a5715e58.jpg) +Figure 3: Overall flow of SAVIT. Our method utilizes additional semantic cues with a mixture-of-experts scheme. The transformer parameters are dynamically adapted to the local semantic information for each token region. With these token-wise dynamic parameters, our network produces more precise inpainting results, generating an object's boundary or synthesizing detailed textures. + +![](images/44b6beda0f9f12c1ea8d3d53d535b9efeb1a69caa4c0efb15182ecd802aba7a1.jpg) +Figure 4: Schematic illustration of SAD. With the specific mixture-of-experts parameters produced by using the same routing weight from the generator, SAD discriminates whether the inpainted results are realistic or not. + +# 3.4. Loss function + +Auxiliary token classification loss. In the mixture-of-experts scheme, it is crucial that each expert is assigned with distinct role [18, 30]. If our semantic experts have been assigned properly to each corresponding category, the output tokens from the semantic-aware dynamic block should be helpful for classification. We drive collaborative role division across our semantic experts. Specifically, we regularize the learning of experts during training by performing auxiliary classification tasks on the output tokens from the semantic-aware dynamic block. Such token-wise classification is performed by an auxiliary single linear layer classifier. The classifier predicts the semantic labels of the output tokens, as illustrated with the green box in Fig. 3. As a result, each expert can effectively learn the feature representation corresponding to a specific semantic label. + +Overall objective function. Our overall objective function $\mathcal{L}_{total}$ is the weighted sum of the following four different loss terms: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {p i x}} + \lambda_ {\text {a d v}} \mathcal {L} _ {\text {a d v}} + \lambda_ {\text {c l s}} \mathcal {L} _ {\text {c l s}} + \lambda_ {\text {s a d}} \mathcal {L} _ {\text {s a d}}, \tag {5} +$$ + +where $L_{pix}$ is a pixel-wise reconstruction loss, $\mathcal{L}_{adv}$ is an adversarial loss, $\mathcal{L}_{cls}$ is token classification loss, and $\mathcal{L}_{sad}$ is semantic adversarial loss, while $\{\lambda_{adv},\lambda_{cls},\lambda_{sad}\}$ control the weight of associated loss term. Given the predicted video frame $\hat{Y}$ and its ground truth counterpart $Y$ , we employ the L1 pixel-wise reconstruction loss as: + +$$ +\mathcal {L} _ {p i x} = \left| \left| \hat {Y} - Y \right| \right| _ {1}. \tag {6} +$$ + +As for the adversarial loss, we train our inpainting network to fool a discriminator $D$ as follows: + +$$ +\mathcal {L} _ {a d v} = - E _ {\hat {Y}} [ D (\hat {Y}) ]. \tag {7} +$$ + +Based on a temporal patch discriminator [26, 23], the discriminator $D$ is trained to optimize + +$$ +\mathcal {L} _ {D} = E _ {Y} \left[ \operatorname {R e L U} (1 - D (Y)) \right] + E _ {\hat {Y}} \left[ \operatorname {R e L U} (1 + D (\hat {Y})) \right]. \tag {8} +$$ + +The token classification loss term is given as + +$$ +\mathcal {L} _ {c l s} = - \sum_ {i} \sum_ {c = 1} ^ {C} y _ {i} ^ {c} \log \left(\text {s o f t m a x} \left(\hat {y} _ {i}\right)\right), \tag {9} +$$ + +where $\hat{y}_i\in \mathbb{R}^C$ is the predicted score from the single-layer classifier conditioned on the output tokens from the semantic-aware dynamic block, while $y_{i}^{c}$ is the target probability for class $c$ at the $i$ -th token. For the ground truth target $y_{i}^{c}$ , we use the input semantic segmentation map (i.e., $y_{i}^{c} = s_{i}(c)$ ). The semantic-aware adversarial loss $\mathcal{L}_{sad}$ is similar to $\mathcal{L}_{adv}$ in that (8) and (7) are used, but with the semantic-aware discriminator (SAD) instead of $D$ . + +# 4. Experiments + +In this section, we quantitatively and qualitatively demonstrate the effectiveness of SAVIT. Please refer to the supplementary material for more experimental results. + +
YouTube-VOS [36]DAVIS [28]
MethodPSNR↑SSIM↑VFID↓E_warp(%)↓PSNR↑SSIM↑VFID↓E_warp(%)↓
VINet [15]29.200.94340.0720.149028.960.94110.1990.1785
LGTSM [3]29.740.95040.0700.185928.570.94090.1700.1640
CAP [21]31.580.96070.0710.147030.280.95210.1820.1533
STTN [39]32.340.96550.0530.090730.670.95600.1490.1449
E2FGVI [23]33.710.97000.0460.086433.010.97210.1160.1315
ISVI [41]31.190.95690.0530.117332.280.96690.1290.1589
FuseFormer [26]33.160.96730.0510.090032.540.97000.1380.1362
SAVIT (Ours)33.970.97270.0430.043633.140.97480.1070.0673
+ +Table 1: Quantitative comparison of SAVIT against state-of-the-art methods. SAVIT outperforms the existing approaches in terms of various evaluation metrics. Notably, SAVIT significantly elevates the inpainting performance of FuseFormer [26], which is our baseline architecture, witnessing the superiority of leveraging semantic cues within the input video. + +
ConfigurationPSNR↑VFID↓
AFuseFormersmall [26]30.510.182
B+ super-categorized segmentation map30.900.161
C+ SAL30.920.164
D+ token classification loss (λcls)31.190.165
E+ SAD31.010.157
+ +Table 2: Ablation experiments on various network configurations under DAVIS [28] dataset. Starting from the performance of the baseline architecture [26], the inpainting performance gradually increases as the configurations are added since the network can benefit from the semantic cues within a video. + +# 4.1. Implementation details + +Datasets and evaluation metrics. We use two conventional video inpainting datasets, YouTube-VOS [36] and DAVIS [28], for our experiments. YouTube-VOS dataset consists of 3471 training, 474 validation, and 508 test videos from mixed scenes. Meanwhile, DAVIS dataset comprises 150 video clips, 60 of which are used for training and the remaining 90 clips are used for evaluation. Following previous works [26, 23], we use only the training videos in the YouTube-VOS for training, while the evaluation is conducted on YouTube-VOS testset and 50 selected clips of DAVIS testset. During training, we create random mask shapes for input masked frames as in [39, 26, 23] and use the method discussed in Section 3.1 to obtain a segmentation map for the entire dataset. + +The inpainting performance is evaluated under two standard scenarios: fixed region inpainting and foreground object removal. For the fixed region inpainting, we use free-form fixed masks for each video provided by FuseFormer [26]. As for the foreground object removal, we remove the main objects using frame-by-frame annotations provided by the stakeholders. Notably, we employ a simple k-nearest neighbors algorithm to complete the missing object region in the segmentation maps for the object removal task (please refer to supplementary material for detail). + +We compare the results with respect to four evaluation metrics: PSNR, SSIM, temporal warping error $E_{\text{warp}}$ [20], + +and Video-based Fréchet Inception Distance (VFID) [35]. PSNR and SSIM are traditional metrics used to measure the average quality of produced frames. VFID and $E_{\text{warp}}$ measure the video's perceptual quality and temporal stability (consistency), respectively, and low values indicate better quality. + +Training settings. We follow the training settings of our baseline network FuseFormer [26]. Specifically, we randomly sample five frames from a video, resize the frames to $240 \times 432$ , and utilize random horizontal flipping augmentation. The network is trained for $500k$ iterations with Adam optimizer [16] and a batch size of eight. The initial learning rate is 1e-4 and decreases to 1e-5 after $400k$ iterations. The coefficients for our loss functions, $\lambda_{adv}$ , $\lambda_{cls}$ , and $\lambda_{sad}$ are 0.01, 0.01, and 0.001, respectively. We implement our network on the PyTorch framework and use eight NVIDIA RTX A6000 GPUs for training. + +# 4.2. Video inpainting results + +Quantitative results. Table 1 quantitatively compares our video inpainting results on fixed region inpainting setting against seven state-of-the-art video inpainting networks: VINet [15], LGTSM [3], CAP [21], STTN [39], E2FVGI [23], ISVI [41], and FuseFormer [26]. The result shows that our method remarkably elevates the performance of FuseFormer [26], on which we base our architecture, and achieves superior performances on both YouTube-VOS and DAVIS datasets compared to the state-of-the-art methods. Notably, the warping error has also significantly improved, indicating that semantic information benefits in generating temporally consistent video. These results show the effectiveness of employing mixture-of-experts to dynamically leverage the semantic cues within given frames for video inpainting. + +Qualitative results. Fig. 5 shows that SAVIT has visually pleasing inpainting results, especially in synthesizing semantic structure, boundary, and texture of objects. Notably, unlike the baseline methods that rely solely on inherent video clues and cannot accurately restore the shape of the human body, SAVIT effectively completes the region + +![](images/8f9ef754c0e2f525e324a8415e1ff9fad5cd58bd9f064040d0f75fe35a052a96.jpg) +Masked Frames + +![](images/2e12573d17e4ed4bbdef05d514d4cfcb748be6c60ec3ea4406cc4f14a89dcc42.jpg) +STTN + +![](images/83c9e68628e4dbb172d29abcd1c981678252f40549c178078f8137a4c8933100.jpg) +FuseFormer + +![](images/db5af5a1e67cd1cd0acaf227123157af4be6685c972685b42ce5c94e51797cd3.jpg) +E2FGVI + +![](images/61ddfec66c64ef64f1e9d49cf480898400a41ac45fb5c6bfe0be325c09473e8b.jpg) +SAVIT (Ours) + +![](images/7c4147a3778a7f13a666cc6763133e5c893c7817b42f7aa996fffa7b08e93034.jpg) + +![](images/4317c7295e4383c139139d008c49d540460ab19648ef61ed0cf72cbaef892e5e.jpg) + +![](images/ecaeab738521227e51db4e6f4d7f990f33ba22d70d047337a3000c2d4e66acca.jpg) + +![](images/596c988000b2e45b2463901bdaf47f6f65d7aa85adf85a2b3a5baf6a4af84b05.jpg) + +![](images/d38d9b0ac28599f1352c93ab68bfc66c176a0638a94010869bb1a98aa065e1e3.jpg) + +![](images/a1a24f9ceac5f5e6855c058b28efb731e8dea5b271b002eeb283b9b5c66c6125.jpg) + +![](images/7ffc4d195e1f7d7db840bb62f8f3a22634d9d0814194e2b3aaa1e8ca4a05858b.jpg) + +![](images/37a9502c5f70e0f52997fd5faf314adf0d73e5c8aa12dffbaa726d0e6a2ffca3.jpg) + +![](images/650781ecd42b506da60b43ac943a79287cd01d68a99b4e566ebc8a09b2b0a2d4.jpg) + +![](images/7df933a9d73d7f4f7bd3ad244b9e0a1deeef2701e8cd1a7628d7a6f83b3c36db.jpg) + +![](images/5e557ee2e7efeac777aa2741c8494cb4f211df93eb81c60d5d6a36303eff8b3f.jpg) + +![](images/dcf1dd3bbcea01e073c1197ba4cbebcd62a0a184871f1a5ca58fdcabbe11c58f.jpg) + +![](images/42fd4d9da2aab0f05f9d2a04a54016784b64d742ba53a31b6a55936660320420.jpg) + +![](images/d93db6a3e4bbe625e5f75f2b0655a892517bbc3be5b6146d13c362c99a0414bf.jpg) + +![](images/553275ed3a56e3b5667a7bdab89edf9dd46a4fcc0d412a38112262cb2cfdd065.jpg) + +![](images/c828aff8235dcc0408eb3b55ccdd947e47ef8ef765be8108fdff4cf5d7ec5bed.jpg) + +![](images/4c0f95f58d30ef09e84edbffc1a1b192cdbe5f452721b1cf40217a51f2e623af.jpg) + +![](images/e3c27885d3d128e97fba23afbbce14e479568f22a10023d3c1bfe49bf2c8b4a6.jpg) + +![](images/17c81c33b45fe4596f76f1f2309547b48631b4081b48ed8a112cb15b1a55cd37.jpg) + +![](images/24999c0908660f75dffaefda4aacc2d52fc0ebbcdfd15e9cf96a9c276f4ac988.jpg) + +![](images/f0f4cb76128310bec266cfe7d02c9a6186e03a5ac2bcba8fdb10ef43e18e283c.jpg) +Figure 5: Qualitative comparison of SAVIT with baseline video inpainting networks in fixed region inpainting and object removal. SAVIT more accurately recovers the object's boundaries and appropriate textures. + +![](images/bca9e385492a285e57d66712941e82a4338c0d4129693e942c32479790841a37.jpg) + +![](images/9f42f31a2daef94f9b2b42752ca633f7b518305740b6117c0885179112dcfbe2.jpg) + +![](images/cb457ca45898752bc9c94b4e07c1966d9b8b20b1cb5d2f9cbd80f4781d8335d0.jpg) + +![](images/13229ab8c7ddaeef1b4fb360903aaa689533aff721ecdfc4730ce24fae886665.jpg) + +![](images/e1640d5d2232cd6f429e96210ddf9ca508d1ce50b726c4733789844f3c55c9aa.jpg) +Figure 6: User study result. + +
ConfigurationPSNR†
hardmax (SAVIT)31.01
softmax30.91
random routing weight19.86
+ +Table 3: Effect of the routing weight on the DAVIS [28] dataset. + +by leveraging additional semantic information (third and fourth rows). Furthermore, SAVIT demonstrates the capability of generating more semantically plausible textures, as observed from the detailed textures of synthesized grass and wall (second and fifth rows). + +User study. We conduct a user study on the foreground object removal task, where ground truth target video frames for evaluation are not available, to investigate the real-world applicability of our method using DAVIS [28] testset. We ask 24 raters to choose a better video between two randomly ordered videos, including the inpainted results by FuseFormer [26] and ours, fifteen times per rater. The result presented in Fig. 6 shows that raters prefer ours to the baseline, indicating the impact of exploiting semantic cues in completing missing object regions. + +# 5. Ablation studies + +We perform ablation experiments based on a smaller FuseFormer network (FuseFormersmall) by halving the dimension of the hidden layer. + +![](images/7d79bd840aa5d68ad9ebdacbe7ef6254649b43d37d35a6105892e0055d143ebc.jpg) +SAVIT w/o SAD +Figure 7: Effect of semantic-aware discriminator (SAD). The texture of the black swan's body is blurry in the case without SAD. + +![](images/7b4fe8511bf1a6cea21b03c4df82bc973f361a247cdf9aa04169684d0c5aae11.jpg) +SAVIT w/ SAD + +Network configuration. Table 2 compares the inpainting performance for various configurations of SAVIT. First, we observe that the baseline network's performance improves with additional segmentation map inputs (rows A and B). However, a subtle additional performance improvement (0.02dB PSNR gain, rows B and C) is achieved when we naively add our proposed dynamic linear layer to the network. Rather, applying the proposed token-wise classification loss to the objective boosts the performance (0.2dB PSNR gain, row D). Overall performance improvement by applying mixture-of-experts (rows C and D) highlights the importance and effectiveness of employing mixture-of-experts to handle semantics dynamically. Moreover, using a semantic-aware dynamic discriminator (row E) degrades PSNR while improving VFID. However, as shown in Fig. 7, utilizing the semantic-aware discriminator produces more visually plausible results, which solidifies the advantage of + +![](images/7e6a537c1d41f40c7c2e54af2f8543267c4121a3c6aa7101e40a7c65fc67c471.jpg) + +![](images/683f84253a3d4df672db5955b8591fbc2ca835d10e89c5ec21dfbeccd40c1a70.jpg) + +![](images/11ccf9fa0c90e978f58ca19e20a7cac7e10d8d28e023726e7caa73f05c919c7a.jpg) + +![](images/a8d45a3c5ca0fdbd138b0ba804b9beae8318c856b29c33689fde41ac4da28eca.jpg) + +![](images/fb3a4e1974f878c4a2d0956f15bd8ac4a1cef61e231d6fe7729a0c353fa4e9b7.jpg) + +![](images/77a8f6095b22a6baff8c284663afe9ad4981f940925da0c6a9cecb1f2b7d1b2d.jpg) + +![](images/b022bd318f8cd0d92db791a753f0df51156f333fe89c87dbfe32607c873c0515.jpg) +Figure 8: Visualization of the semantic router results. (1) to (8) are the visualization of routing weights of the corresponding super-category. The routing weights are rearranged to the image-like features. Brighter pixel values indicate higher routing weights. + +![](images/caf30627110a1a0f1af152c9b326d99547050e2cc717445342f9dec54715d7e5.jpg) + +![](images/e7f834c23e09fd6ebd51fa147a7a1c3a815dd1dd386c76e6d98730eb5e259917.jpg) + +![](images/86c02db76eb995e3c627518ee35bf74f1185d927e568fc96194ffdf13e483d9c.jpg) + +![](images/36285cb7505347e38e657601e9874d0e418bb7a51053a0d0ba0493839ba9fe31.jpg) + +![](images/ef1a5e631bc5ee88923b3a65cc4799f2042c7af9ec19d28047201a5ca510acca.jpg) + +
# ExpertsPSNR↑VFID↓
430.830.165
630.900.168
831.010.157
+ +![](images/a34b416150bc72a0c10c74a5618c4197e585ae0fd65437606b1f410293f19073.jpg) +Figure 9: Qualitative comparison on different routing weight strategies. + +Table 4: Inpainting results by changing the numbers of experts under the DAVIS [28] dataset. More experts render better inpainting results. + +
# blocks# params.PSNR↑
126.0M30.40
233.1M30.45
+ +the discriminator in improving perceptual quality. + +Visualization of routing weights. Fig. 8 visualizes the routing weights of each expert from the trained semantic router. We rearrange the tokens to image-like features for visualization, and brighter values indicate higher routing weights in the corresponding region. The bright regions among different experts indicate that the experts collaborate during the inpainting procedure rather than simply utilizing a specific expert. We see that routing weights tend to change locally smoothly while the input segmentation map changes abruptly. + +Designing choice on semantic router. Table 3 compares different design choices of our semantic router. The performance drop by replacing the trained routing weights (i.e., mixture coefficients) with random weights indicates the discriminative role of trained experts. Moreover, the results of softmax routing strategy show a noticeable performance degradation against hardmax strategy quantitatively. Fig. 9 also shows that softmax results tend to have blurry texture, especially on the edge area. Therefore, Hardmax was chosen for our final model to facilitate role division between experts better, as it enables the learning of distinct parameters by dividing the input for each expert before the router blends the token based on its features. + +Impact of the number of experts. Table 4 compares the inpainting performance by changing the number of experts. Notably, as the number of experts (i.e., semantic categories) increases, the network can benefit from more diverse semantic clues in terms of PSNR and VFID values, and we set the number of experts for our final model to eight. + +Impact of the number of semantic blocks. Table 5 compares the inpainting performance by applying our semantic module on a different number of blocks. The result indicates that using more semantic blocks produces a slightly + +Table 5: Ablation on different number of semantic blocks. + +
MethodMask ratioPSNR†VFID↓
FuseFormersmall30%22.570.554
60%18.101.059
FuseFormersmall + Ours30%23.570.412
60%19.210.750
+ +Table 6: Comparison on varying mask ratio. + +better inpainting performance. Although we apply our semantic module one time due to hardware limits for our final model, it is expected that the results will be naturally improved by increasing the number of semantic blocks. + +Input video with large corruption. Table 6 compares our method against FuseFormersmall under different corrupted region ratio settings. It is shown that, despite larger masks, our method still outperforms the baseline by a large margin by exploiting the semantic information. + +# 6. Conclusion and Future Works + +In this study, we have explored leveraging semantic information within a given video for transformer-based video inpainting. Accordingly, we propose an effective semantic-aware dynamic transformer with the notion of mixture-of-experts. Our dynamic transformer contains class-specific expert parameters where experts are efficiently combined to perform token-wise dynamic linear operations based on the token feature and its semantic information. The experiments demonstrate the effectiveness of our method, especially in recovering semantic structures. Although we have utilized pseudo labels for our study, we believe more accurate and denser annotated datasets can improve the inpainting performance further since our method relies on the quality of the segmentation label. + +# Acknowledgements + +This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2023-00222776) and Samsung Electronics Co., Ltd, and Samsung Research Funding Center of Samsung Electronics under Project Number SRFCIT1901-06. + +# References + +[1] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3 +[2] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 1, 2 +[3] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Learnable gated temporal shift module for deep video inpainting. In Proceedings of the British Machine Vision Conference (BMVC), 2019. 2, 6 +[4] Cheng Chen, Jiayin Cai, Yao Hu, Xu Tang, Xinggang Wang, Chun Yuan, Xiang Bai, and Song Bai. Deep interactive video inpainting: An invisibility cloak for harry potter. In Proceedings of the ACM International Conference on Multimedia (ACM MM), 2021. 2 +[5] Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolution: Attention over convolution kernels. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2 +[7] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[9] Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2 +[10] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 1991. 2 +[11] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. In Advances in Neural Information Processing Systems (NeurIPS), 2016. 2 +[12] Jun-Gyu Jin, Jaehyun Bae, Han-gyul Baek, and Sang-hyo Park. Object-ratio-preserving video retargeting framework based on segmentation and inpainting. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 2 +[13] Jaeyeon Kang, Seoung Wug Oh, and Seon Joo Kim. Error compensation framework for flow-guided video inpainting. + +In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 2 +[14] Bholeshwar Khurana, Soumya Ranjan Dash, Abhishek Bhatia, Aniruddha Mahapatra, Hrituraj Singh, and Kuldeep Kulkarni. Semie: Semantically-aware image extrapolation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 2 +[15] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Deep video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 6 +[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[17] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3 +[18] Xiangtao Kong, Hengyuan Zhao, Yu Qiao, and Chao Dong. Classsr: A general framework to accelerate super-resolution networks by data characteristic. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5 +[19] Avisek Lahiri, Arnav Kumar Jain, Sanskar Agrawal, Pabitra Mitra, and Prabir Kumar Biswas. Prior guided gan based semantic inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1 +[20] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 6 +[21] Sungho Lee, Seoul Wug Oh, DaeYeun Won, and Seon Joo Kim. Copy-and-paste networks for deep video inpainting. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 6 +[22] Jiacheng Li, Chang Chen, and Zhiwei Xiong. Contextual outpainting with object-level contrastive learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[23] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 4, 5, 6 +[24] Liang Liao, Jing Xiao, Zheng Wang, Chia-Wen Lin, and Shin'ichi Satoh. Guidance and evaluation: Semantic-aware image inpainting for mixed scenes. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2 +[25] Liang Liao, Jing Xiao, Zheng Wang, Chia-Wen Lin, and Shin'ichi Satoh. Image inpainting guided by coherence priors of semantics and textures. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2 +[26] Rui Liu, Hanming Deng, Yangyi Huang, Xiaoyu Shi, Lewei Lu, Wenxiu Sun, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fuseformer: Fusing fine-grained information in + +transformers for video inpainting. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1, 2, 4, 5, 6, 7 +[27] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[28] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6, 7, 8 +[29] Jingjing Ren, Qingqing Zheng, Yuanyuan Zhao, Xuemiao Xu, and Chen Li. Dlformer: Discrete latent transformer for video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2 +[30] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 1, 2, 5 +[31] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In Proceedings of the International Conference on Learning Representations (ICLR), 2017. 1, 2 +[32] Yuhang Song, Chao Yang, Yeji Shen, Peng Wang, Qin Huang, and C-C Jay Kuo. Spg-net: Segmentation prediction and guidance network for image inpainting. In Proceedings of the British Machine Vision Conference (BMVC), 2018. 1, 2 +[33] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2 +[34] Chuan Wang, Haibin Huang, Xiaoguang Han, and Jue Wang. Video inpainting by jointly learning temporal structure and spatial details. In Association for the Advancement of Artificial Intelligence (AAAI), 2019. 1, 2 +[35] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2018. 6 +[36] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 2, 3, 6 +[37] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep flow-guided video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2 +[38] Brandon Yang, Gabriel Bender, Quoc V Le, and Jiquan Ngiam. Condconv: Conditionally parameterized convolutions for efficient inference. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 1, 2 +[39] Yanhong Zeng, Jianlong Fu, and Hongyang Chao. Learning joint spatial-temporal transformations for video inpainting. + +In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 6 +[40] Kaidong Zhang, Jingjing Fu, and Dong Liu. Flow-guided transformer for video inpainting. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 2 +[41] Kaidong Zhang, Jingjing Fu, and Dong Liu. Inertia-guided flow completion and style fusion for video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6 +[42] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[43] Xueyan Zou, Linjie Yang, Ding Liu, and Yong Jae Lee. Progressive temporal feature alignment network for video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1 \ No newline at end of file diff --git a/2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/images.zip b/2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ced9adacf8a4a23c886ba6452d28d51fcfae49ea --- /dev/null +++ b/2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da700231c69591bc0d4d0041aa45734362a3306cf6e95a0576d5362558a1a057 +size 702409 diff --git a/2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/layout.json b/2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..09611e4932b2c8d06153adafa98a3e0e1c850c47 --- /dev/null +++ b/2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/layout.json @@ -0,0 +1,9779 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 103, + 517, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 103, + 517, + 121 + ], + "spans": [ + { + "bbox": [ + 75, + 103, + 517, + 121 + ], + "type": "text", + "content": "Semantic-Aware Dynamic Parameter for Video Inpainting Transformer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 142, + 492, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 142, + 492, + 186 + ], + "spans": [ + { + "bbox": [ + 100, + 142, + 492, + 186 + ], + "type": "text", + "content": "Eunhye Lee" + }, + { + "bbox": [ + 100, + 142, + 492, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 100, + 142, + 492, + 186 + ], + "type": "text", + "content": ", Jinsu Yoo" + }, + { + "bbox": [ + 100, + 142, + 492, + 186 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 100, + 142, + 492, + 186 + ], + "type": "text", + "content": ", Yunjeong Yang" + }, + { + "bbox": [ + 100, + 142, + 492, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 142, + 492, + 186 + ], + "type": "text", + "content": ", Sungyong Baik" + }, + { + "bbox": [ + 100, + 142, + 492, + 186 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 100, + 142, + 492, + 186 + ], + "type": "text", + "content": ", Tae Hyun Kim" + }, + { + "bbox": [ + 100, + 142, + 492, + 186 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 100, + 142, + 492, + 186 + ], + "type": "text", + "content": " \n{dldms1345, jinsuyoo, yunjoeongyang, dsybaik, taehyunkim}@hanyang.ac.kr \n" + }, + { + "bbox": [ + 100, + 142, + 492, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 142, + 492, + 186 + ], + "type": "text", + "content": "Dept. of Computer Science, " + }, + { + "bbox": [ + 100, + 142, + 492, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 142, + 492, + 186 + ], + "type": "text", + "content": "Dept. of Artificial Intelligence, " + }, + { + "bbox": [ + 100, + 142, + 492, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 100, + 142, + 492, + 186 + ], + "type": "text", + "content": "Dept. of Data Science, Hanyang University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 238, + 289, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 289, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 289, + 453 + ], + "type": "text", + "content": "Recent learning-based video inpainting approaches have achieved considerable progress. However, they still cannot fully utilize semantic information within the video frames and predict improper scene layout, failing to restore clear object boundaries for mixed scenes. To mitigate this problem, we introduce a new transformer-based video inpainting technique that can exploit semantic information within the input and considerably improve reconstruction quality. In this study, we use the mixture-of-experts scheme and train multiple experts to handle mixed scenes, including various semantics. We leverage these multiple experts and produce locally (token-wise) different network parameters to achieve semantic-aware inpainting results. Extensive experiments on YouTube-VOS and DAVIS benchmark datasets demonstrate that, compared with existing conventional video inpainting approaches, the proposed method has superior performance in synthesizing visually pleasing videos with much clearer semantic structures and textures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 466, + 128, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 128, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 128, + 479 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 481, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 613 + ], + "type": "text", + "content": "A constant increase in demand for video content in our daily lives (e.g., YouTube or TikTok) has led to the development of video inpainting methods, which aim to complete a missing region or erase unwanted areas such as watermarks and captions from a given input video. In particular, recent convolutional neural networks (CNNs) for video inpainting tasks have shown promising results [34, 15, 2, 43, 37, 9]. In addition, transformer-based inpainting networks [39, 26, 23, 29] significantly escalate the quality of inpainting results through their large network capacity and superior local/global connectivity based on attention mechanisms." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 613, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 288, + 685 + ], + "type": "text", + "content": "However, conventional approaches are unable to fully utilize the semantic information within the input video or distinguish the class-specific characteristics of objects with different semantics. As a result, they frequently fail to recover proper object structure, texture, and scene layout. Several studies [32, 24, 19, 25] have suggested that utiliz" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 214, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 214, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 214, + 545, + 262 + ], + "type": "text", + "content": "ing semantic information can lead to better results and produce visually more plausible images when filling missing regions. However, research on video inpainting has yet to explore the potential of incorporating semantic maps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 265, + 547, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 547, + 386 + ], + "type": "text", + "content": "Video inpainting is a challenging task that requires to render temporally consistent video frames. To address this issue, optical flow has been extensively explored, although it is computationally intensive [15, 37, 9, 23, 41, 40, 13]. In this study, we demonstrate that semantic information can also be used to enforce temporal consistency, and we introduce a new technique that further facilitates the use of semantic cues in a video by dynamically mixing multiple class-specific experts to handle objects with distinct semantics adaptively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 387, + 547, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 547, + 532 + ], + "type": "text", + "content": "Although previous semantic-guided inpainting networks used the predicted segmentation map as additional input [32] or developed a dedicated semantic-aware inference module [24] to learn semantic-aware parameters, the restoration performance is limited, given the use of shared network parameters for the different semantics. This problem has been addressed in a recent study [25] by using a semantic-aware attention module that enables the features to attend solely to regions with identical semantic labels. However, the computational cost increases proportionally to the number of classes because the attention mechanism is separately carried out for each category." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": "To mitigate these problems, we propose a novel semantic-aware dynamic parameter selection approach that effectively utilizes the semantic information within input video frames while retaining the number of operations during the inference phase. Specifically, inspired by Cond-Conv [38], we introduce conditionally parameterized linear operations to learn semantic-aware experts, and produce locally varying (dynamic) parameters by leveraging the given semantic cues based on the notion of mixture-of-experts [31, 30, 38]. Our linear operations with dynamically determined parameters replace the standard feed-forward operation within a conventional transformer block. Notably, we can keep the number of parameters for inference by mixing the expert parameters before performing linear operations rather than aggregating the output features cal" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 693, + 140, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 693, + 140, + 702 + ], + "spans": [ + { + "bbox": [ + 62, + 693, + 140, + 702 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 703, + 140, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 140, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 140, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12949" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "ulated by each expert. Moreover, in contrast to the original CondConv [38] that produces conditional, but locally uniform parameters, ours can generate conditional and locally (token-wise) varying parameters, thus improving each token's representation power." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 287, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 287, + 228 + ], + "type": "text", + "content": "Our extensive experiments witness the outstanding performance of the proposed Semantic-Aware Video Inpainting Transformer (SAVIT) in improving video inpainting results. Specifically, SAVIT elevates quantitative performance on conventional video inpainting benchmark datasets (YouTube-VOS [36] and DAVIS [28]) and produces visually more superior results against state-of-the-art methods. We summarize our contribution as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 232, + 286, + 343 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 58, + 232, + 286, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 232, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 58, + 232, + 286, + 255 + ], + "type": "text", + "content": "- We tackle leveraging semantic information within the given input video frames for the video inpainting task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 258, + 286, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 258, + 286, + 281 + ], + "spans": [ + { + "bbox": [ + 58, + 258, + 286, + 281 + ], + "type": "text", + "content": "- We introduce a novel semantic-aware video inpainting transformer based on a mixture-of-experts scheme." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 285, + 286, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 285, + 286, + 307 + ], + "spans": [ + { + "bbox": [ + 58, + 285, + 286, + 307 + ], + "type": "text", + "content": "- We propose a semantic-aware dynamic linear operation to exploit local semantic cues effectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 310, + 286, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 310, + 286, + 343 + ], + "spans": [ + { + "bbox": [ + 58, + 310, + 286, + 343 + ], + "type": "text", + "content": "- Extensive experiments demonstrate the superiority and efficacy of our method, especially in recovering semantic structures and textures." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 357, + 133, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 133, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 133, + 369 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 373, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 661 + ], + "type": "text", + "content": "Video inpainting. An active research topic in video inpainting is generating a temporally consistent video. Various approaches, including 3D-CNN-based, flow-based, and transformer-based methods, have been proposed to fill in a missing region inside a video plausibly. Despite their ability to extract spatio-temporal information, 3D-CNN methods [34, 2, 3] fail to fully utilize global correspondence due to limited temporal receptive field. Flow-based methods [15, 37, 9, 23, 41, 40, 13] first estimate motion flow in the missing region and use the flow to predict the missing pixels. However, these methods highly rely on pre-trained inpainting networks and often require manual operations [37, 9]. Recently, transformer-based methods [39, 26, 23, 29] have attracted research attention due to their superiority in capturing information across long-ranged frames. Built upon the vision transformer architecture [8], FuseFormer [26] proposed soft tokenization with spatial overlapping to interact information among neighboring tokens. Followup studies have improved the performance with dedicated modules such as focal attention [23] or discrete latent mapping [29]. In our study, we develop our method on top of transformers to capture long-range dependencies and advocate the effectiveness of leveraging semantic information." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Semantic-guided inpainting. Several works have been proposed to utilize the semantic map for inpainting [32, 24, 25], and its related fields [14, 4, 22, 12]. Among them, SPGNet [32] first fills the missing segmentation map, then" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "content": "synthesizes the inpainted image based on the estimated semantic results. Moreover, SGENet [24] introduced iterative feature refinement using semantic map-based normalization [27] to handle more complex holes in mixed scenes, including objects with different semantics. Recent work [25] proposed to calculate semantic-wise attention to learn the corresponding semantic information. However, most of the previous works have developed upon image inpainting, and leveraging semantic cues within a video has yet to be explored due to the lack of annotated data. Our study mitigates the data acquisition problem by gathering pseudo-annotated videos with pre-trained segmentation network [17], and demonstrates that semantic information can elevate video inpainting performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "type": "text", + "content": "Dynamic filter. As an alternative to fixed filter, scene-adaptive dynamic filter networks [11, 7, 42] adapt their parameters based on a given input image and/or features. These methods allow flexibility in the network in treating various inputs with different characteristics. Notably, CondInst [33] dynamically generates instance-aware parameters for instance segmentation. In this study, we adapt the transformer parameters in reasoning each token conditioned to its features and semantic information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 354, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 473 + ], + "type": "text", + "content": "Mixture-of-experts. By combining the results from multiple expert models, the network can further improve its representation power [30, 10, 31]. Instead of increasing the computational cost during the inference proportional to the number of experts, recent works [38, 5] parameterized the mixed experts to combine the knowledge efficiently. Our work is motivated by CondConv [38]; we mix the experts' parameters before the feature calculation. Specifically, we divide the experts' roles by semantic categories, then combine semantic-aware experts' knowledge at each token." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 479, + 411, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 411, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 411, + 491 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 494, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 545, + 542 + ], + "type": "text", + "content": "In this section, we first introduce the annotation procedure to acquire the semantic maps for videos, followed by a detailed description of the proposed framework SAVIT, which leverages semantic information for video inpainting." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 544, + 449, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 449, + 556 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 449, + 556 + ], + "type": "text", + "content": "3.1. Semantic data acquisition" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Unlike previous semantic-aware approaches [32, 25, 24], which simultaneously estimate the segmentation maps for inpainting, we assume that semantic segmentation maps are predictable from the given clean video frames, which include unmasked regions to remove, using off-the-shelf segmentation methods. However, currently available video segmentation datasets, such as DAVIS [28] or CityScape [6], are limited to specific categories and/or particular objects, leading acquiring ground-truth segmentation maps for general video frames difficult. Alternatively, we utilize a fully pre-trained panoptic segmentation network [17] to predict the foreground and background semantics in more general scenes." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12950" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 77, + 217, + 237 + ], + "blocks": [ + { + "bbox": [ + 77, + 77, + 217, + 237 + ], + "lines": [ + { + "bbox": [ + 77, + 77, + 217, + 237 + ], + "spans": [ + { + "bbox": [ + 77, + 77, + 217, + 237 + ], + "type": "image", + "image_path": "19747dbd49d052ab8d77d77bccd7236ed356a66ac7a4be1bae95457e866bafa9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 240, + 194, + 251 + ], + "lines": [ + { + "bbox": [ + 93, + 240, + 194, + 251 + ], + "spans": [ + { + "bbox": [ + 93, + 240, + 194, + 251 + ], + "type": "text", + "content": "(a) Initial segmentation map" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 241, + 76, + 400, + 233 + ], + "blocks": [ + { + "bbox": [ + 241, + 76, + 400, + 233 + ], + "lines": [ + { + "bbox": [ + 241, + 76, + 400, + 233 + ], + "spans": [ + { + "bbox": [ + 241, + 76, + 400, + 233 + ], + "type": "image", + "image_path": "586640cd863417072ec09c23f2d6109ec25d0599f18f8ed8b3b60c0160b7e41b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 240, + 360, + 251 + ], + "lines": [ + { + "bbox": [ + 284, + 240, + 360, + 251 + ], + "spans": [ + { + "bbox": [ + 284, + 240, + 360, + 251 + ], + "type": "text", + "content": "(b) Super-categorize" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 262, + 545, + 298 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 545, + 298 + ], + "type": "text", + "content": "Figure 1: Overview of the semantic data preparation: (a) Temporally inconsistent rough prediction using a pre-trained panoptic segmentation model; (b) Grouping raw predictions into corresponding super-categories; (c) Super-categorized segmentation map produces temporally consistent labels." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 444, + 109, + 507, + 217 + ], + "blocks": [ + { + "bbox": [ + 437, + 95, + 511, + 105 + ], + "lines": [ + { + "bbox": [ + 437, + 95, + 511, + 105 + ], + "spans": [ + { + "bbox": [ + 437, + 95, + 511, + 105 + ], + "type": "text", + "content": "temporally consistent" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 444, + 109, + 507, + 217 + ], + "lines": [ + { + "bbox": [ + 444, + 109, + 507, + 217 + ], + "spans": [ + { + "bbox": [ + 444, + 109, + 507, + 217 + ], + "type": "image", + "image_path": "51008e07925610910e6c24fb9b9ab66fcc56ead2769c1b726f357d65046c93b0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 241, + 512, + 262 + ], + "lines": [ + { + "bbox": [ + 432, + 241, + 512, + 262 + ], + "spans": [ + { + "bbox": [ + 432, + 241, + 512, + 262 + ], + "type": "text", + "content": "(c) Super-categorized segmentation map" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 308, + 287, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 511 + ], + "type": "text", + "content": "Fig. 1 illustrates the overall semantic data acquisition process used for training and inference. For training, we first apply the panoptic segmentation network [17] to collect estimated segmentation maps for video frames in conventional video inpainting datasets (e.g., YouTube-VOS [36]). The initial segmentation results consist of 53 background and 80 foreground classes [1]. However, these initial segmentation maps with a large number of classes are erroneous and temporally inconsistent, as shown in Fig. 1 (a). Thus, we enforce the temporal consistency of semantics by grouping the predicted labels to the corresponding supercategory of the class. For example, dog, horse, and sheep regions are grouped into a super-category of animal (Fig. 1 (b)). Although we can lose semantic details to some extent by the super-categorization, such constraint helps the training of our semantic-aware inpainting network through the enforced temporal semantic consistency (Fig. 1 (c))." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 513, + 287, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 513, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 513, + 287, + 632 + ], + "type": "text", + "content": "In our work, we define eight different super-categories based on the texture similarities and the frequencies of categories, which are measured on the entire training dataset. Specifically, we have five super-categories for foreground objects: 'animal,' 'person,' 'vehicle,' 'plants,' and 'sports.' Moreover, we have two super-categories for background according to their texture; 'plain' and 'patterned.' Finally, we group the rest of the remaining sub-categories into 'etc.' During the test phase, we also obtain the semantic maps in the same fashion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 639, + 248, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 248, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 248, + 651 + ], + "type": "text", + "content": "3.2. Semantic-aware dynamic transformer" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "To handle diverse scenes with various objects across different classes in videos, we propose a semantic-aware dynamic transformer. In particular, our transformer employs a mixture-of-experts scheme, in which each expert is a single fully connected layer with its parameters reserved for each" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 308, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 380 + ], + "type": "text", + "content": "corresponding semantic class (i.e., super-category) obtained from Section 3.1. Conditioned on the semantic features of each token, we create a token-wise fully connected layer, where its parameters are determined by weighted summing the class-specific parameters, and the weights for the mixture are produced by our semantic router as in the following." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 382, + 545, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 382, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 545, + 454 + ], + "type": "text", + "content": "Learning semantic router. To produce locally (i.e., token-wise) varying parameters conditioned on the semantic information, the network requires a routing function that predicts blending weights to mix the class-specific expert parameters. Our router computes routing weights (i.e., mixture coefficients) for each token by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 357, + 463, + 545, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 463, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 357, + 463, + 545, + 475 + ], + "type": "interline_equation", + "content": "r _ {i} = M L P \\left(f _ {i}\\right) + h a r d m a x \\left(s _ {i}\\right), \\tag {1}", + "image_path": "46e248dc9878e880ced12e2949a8c076fc8495dd7d841457542921b8a884f531.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "inline_equation", + "content": "r_i \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "text", + "content": " is a produced mixture weight vector; " + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "text", + "content": " is the feature of an input token; " + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "inline_equation", + "content": "s_i \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "text", + "content": " is a histogram of class distribution on the segmentation map for the given " + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "text", + "content": "-th token; " + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "text", + "content": " is the number of semantic experts (" + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "inline_equation", + "content": "C = 8" + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "text", + "content": " in this work unless stated otherwise); and MLP is a standard multi-layer perceptron (MLP) composed of two fully-connected layers with Leaky ReLU activation and sigmoid output. Our router uses the most dominant semantic label from the input histogram " + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "text", + "content": " by using a hard selection strategy (hardmax operation) rather than a softmax result. In addition, we model our router to be lightweight (i.e., two layers) to minimize the overhead, keeping our framework to have similar computational costs to the standard transformers during inference." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Semantic-aware dynamic parameterization. By using the routing weights " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " generated by the router for the " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "-th token with features " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", we produce semantic-aware dynamic parameters " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "W_i" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " for linear (fully-connected) layers of a transformer block. The formulation of semantic-aware dynamic" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12951" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 74, + 255, + 361 + ], + "blocks": [ + { + "bbox": [ + 69, + 74, + 255, + 361 + ], + "lines": [ + { + "bbox": [ + 69, + 74, + 255, + 361 + ], + "spans": [ + { + "bbox": [ + 69, + 74, + 255, + 361 + ], + "type": "image", + "image_path": "4b3569247df550c85411f6c17597e8cda4bb180c0b91273b3fecaaf60d7fad56.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 365, + 287, + 456 + ], + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 456 + ], + "type": "text", + "content": "Figure 2: Illustration of SAFFN. With the histogram of semantic categories (e.g., hardmax) and additional linear layers, the semantic router outputs a routing weight for a specific token. Using the routing weight, SAL first acquires token-specific parameters by the weighted sum of experts, then operates the linear transformation of the token with the parameters. Here, we illustrate the operational flow of a single token. " + }, + { + "bbox": [ + 46, + 365, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 46, + 365, + 287, + 456 + ], + "type": "text", + "content": " indicates element-wise summation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 464, + 147, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 147, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 147, + 475 + ], + "type": "text", + "content": "layer (SAL) is given by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 125, + 483, + 287, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 483, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 125, + 483, + 287, + 496 + ], + "type": "interline_equation", + "content": "\\operatorname {S A L} \\left(f _ {i}\\right) = f _ {i} \\cdot W _ {i}, \\tag {2}", + "image_path": "db63d7b1476657afcfc16daf92da3bbd064376732b24540d22621f08d3a2eba5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 503, + 279, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 279, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 279, + 515 + ], + "type": "text", + "content": "where the mixed expert parameter " + }, + { + "bbox": [ + 47, + 503, + 279, + 515 + ], + "type": "inline_equation", + "content": "W_{i}" + }, + { + "bbox": [ + 47, + 503, + 279, + 515 + ], + "type": "text", + "content": " for token " + }, + { + "bbox": [ + 47, + 503, + 279, + 515 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 47, + 503, + 279, + 515 + ], + "type": "text", + "content": " yields:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 123, + 521, + 287, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 521, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 123, + 521, + 287, + 553 + ], + "type": "interline_equation", + "content": "W _ {i} = \\sum_ {c = 1} ^ {C} r _ {i} (c) \\cdot w ^ {c}. \\tag {3}", + "image_path": "f15e1a0ae575e1b513b2bbf50ec62cf9e96c92ebed435c21d86b218715a33ebf.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 557, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 557, + 287, + 714 + ], + "type": "inline_equation", + "content": "w^{c}" + }, + { + "bbox": [ + 46, + 557, + 287, + 714 + ], + "type": "text", + "content": " indicates the parameters of each class-specific expert, while " + }, + { + "bbox": [ + 46, + 557, + 287, + 714 + ], + "type": "inline_equation", + "content": "r_{i}(c)" + }, + { + "bbox": [ + 46, + 557, + 287, + 714 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 557, + 287, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 557, + 287, + 714 + ], + "type": "text", + "content": "-th element of routing weight vector " + }, + { + "bbox": [ + 46, + 557, + 287, + 714 + ], + "type": "inline_equation", + "content": "r_{i}" + }, + { + "bbox": [ + 46, + 557, + 287, + 714 + ], + "type": "text", + "content": " and means the mixture coefficient for the corresponding expert " + }, + { + "bbox": [ + 46, + 557, + 287, + 714 + ], + "type": "inline_equation", + "content": "w^{c}" + }, + { + "bbox": [ + 46, + 557, + 287, + 714 + ], + "type": "text", + "content": ". Consequently, our semantic-aware dynamic parameterization produces locally varying (token-wise) parameters. For example, a token corresponding to a goat in the input frame can be handled by parameters specialized for animal super-category, and another token on the rock can be dealt with specific parameters for the plain background. Notably, our framework can also handle a token of complex scenes where semantically different objects are coexisting. This is because our router is designed to produce locally varying weights and give high routing weights" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "for experts corresponding to coexisting objects, rather than choosing a single expert." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 96, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 96, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 96, + 545, + 144 + ], + "type": "text", + "content": "Finally, we devise a semantic-aware dynamic feedforward network (SAFFN) as depicted in Fig. 2. SAFFN includes two SAL using the dynamically generated semantic-aware parameters as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 148, + 545, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 148, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 317, + 148, + 545, + 180 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {S A F F N} \\left(f _ {i}\\right) = \\operatorname {S A L} \\left(\\operatorname {R e L U} \\left(\\operatorname {S A L} \\left(f _ {i}; W _ {i} ^ {l}\\right)\\right); W _ {i} ^ {l + 1}\\right) \\tag {4} \\\\ = \\max \\left(\\mathbf {0}, f _ {i} \\cdot W _ {i} ^ {l}\\right) \\cdot W _ {i} ^ {l + 1}. \\\\ \\end{array}", + "image_path": "b1187ae4951a7023a3fcc93e6760370ca52d692028ba2a1f651694b1cdaa7f60.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 185, + 545, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 185, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 305, + 185, + 545, + 233 + ], + "type": "text", + "content": "In our SAFFN, each SAL is parameterized by our dynamic parameters " + }, + { + "bbox": [ + 305, + 185, + 545, + 233 + ], + "type": "inline_equation", + "content": "W_{i}^{l}" + }, + { + "bbox": [ + 305, + 185, + 545, + 233 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 185, + 545, + 233 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 185, + 545, + 233 + ], + "type": "text", + "content": " denotes the layer index of the linear transformation, and ReLU activation is included between two layers. Bias parameters are omitted for brevity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 235, + 423, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 235, + 423, + 245 + ], + "spans": [ + { + "bbox": [ + 306, + 235, + 423, + 245 + ], + "type": "text", + "content": "3.3. Overall architecture" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 247, + 545, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 247, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 545, + 415 + ], + "type": "text", + "content": "Fig. 3 illustrates the overall pipeline of the proposed video inpainting network SAVIT. The network is composed of encoder, decoder, and multiple transformer blocks between the encoder and the decoder. First, the encoder produces output features by taking " + }, + { + "bbox": [ + 305, + 247, + 545, + 415 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 247, + 545, + 415 + ], + "type": "text", + "content": " consecutive video frames with masks and the corresponding semantic segmentation maps as input. We employ the encoder in FuseFormer [26] and additionally concatenate corresponding semantic maps to input. Next, we carry out the tokenization procedure [26], then forward to several transformer blocks to extract powerful features. Finally, the tokenized features are rearranged into the image-like shape, which is then fed into the decoder that predicts " + }, + { + "bbox": [ + 305, + 247, + 545, + 415 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 247, + 545, + 415 + ], + "type": "text", + "content": " inpainted video frames as previous works [26, 23]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 415, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 582 + ], + "type": "text", + "content": "Note that our semantic-aware mixture-of-experts schemes can be plugged into any conventional transformer-based approach (e.g., FuseFormer [26], E2FGVI [23]), where a conventional transformer block is just replaced with our semantic-aware transformer block. In our experiments, we embed our dynamic transformer within FuseFormer [26]. Specifically, SAVIT contains eight FuseFormer blocks, each of which consists of a multi-head self-attention (MHSA) operation and fusion feed-forward network (F3N). We replace one of the eight FuseFormer blocks with our dynamic transformer block. Precisely, within the F3N in the fourth FuseFormer block, we modify the two standard linear layers with our dynamic ones and add a semantic router." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 585, + 545, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 585, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 304, + 585, + 545, + 716 + ], + "type": "text", + "content": "Semantic-aware dynamic discriminator. To facilitate each expert's learning of the corresponding category's representation and texture details, we employ a semantic-aware discriminator (SAD) that predicts the real/fake score for each token. Fig. 4 illustrates the architecture of SAD, which has similar structures to the encoder of SAVIT, except for the last layer, where a single SAL is added. Notably, we utilize the same mixture coefficients for dynamic parameters for both SAVIT (generator) and SAD (discriminator), encouraging each expert to learn detailed and discriminative category-specific representation." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12952" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 74, + 534, + 235 + ], + "blocks": [ + { + "bbox": [ + 65, + 74, + 534, + 235 + ], + "lines": [ + { + "bbox": [ + 65, + 74, + 534, + 235 + ], + "spans": [ + { + "bbox": [ + 65, + 74, + 534, + 235 + ], + "type": "image", + "image_path": "c92922e1b79655c66e6835ca0662a9d7faa306d7f5c9e69c77aa9099a5715e58.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 240, + 547, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 240, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 547, + 285 + ], + "type": "text", + "content": "Figure 3: Overall flow of SAVIT. Our method utilizes additional semantic cues with a mixture-of-experts scheme. The transformer parameters are dynamically adapted to the local semantic information for each token region. With these token-wise dynamic parameters, our network produces more precise inpainting results, generating an object's boundary or synthesizing detailed textures." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 296, + 279, + 400 + ], + "blocks": [ + { + "bbox": [ + 53, + 296, + 279, + 400 + ], + "lines": [ + { + "bbox": [ + 53, + 296, + 279, + 400 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 279, + 400 + ], + "type": "image", + "image_path": "44b6beda0f9f12c1ea8d3d53d535b9efeb1a69caa4c0efb15182ecd802aba7a1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 404, + 288, + 450 + ], + "lines": [ + { + "bbox": [ + 46, + 404, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 288, + 450 + ], + "type": "text", + "content": "Figure 4: Schematic illustration of SAD. With the specific mixture-of-experts parameters produced by using the same routing weight from the generator, SAD discriminates whether the inpainted results are realistic or not." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 459, + 133, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 459, + 133, + 470 + ], + "spans": [ + { + "bbox": [ + 47, + 459, + 133, + 470 + ], + "type": "text", + "content": "3.4. Loss function" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 473, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 652 + ], + "type": "text", + "content": "Auxiliary token classification loss. In the mixture-of-experts scheme, it is crucial that each expert is assigned with distinct role [18, 30]. If our semantic experts have been assigned properly to each corresponding category, the output tokens from the semantic-aware dynamic block should be helpful for classification. We drive collaborative role division across our semantic experts. Specifically, we regularize the learning of experts during training by performing auxiliary classification tasks on the output tokens from the semantic-aware dynamic block. Such token-wise classification is performed by an auxiliary single linear layer classifier. The classifier predicts the semantic labels of the output tokens, as illustrated with the green box in Fig. 3. As a result, each expert can effectively learn the feature representation corresponding to a specific semantic label." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "text", + "content": "Overall objective function. Our overall objective function " + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{total}" + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "text", + "content": " is the weighted sum of the following four different loss terms:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 699, + 287, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 699, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 699, + 287, + 713 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p i x}} + \\lambda_ {\\text {a d v}} \\mathcal {L} _ {\\text {a d v}} + \\lambda_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}} + \\lambda_ {\\text {s a d}} \\mathcal {L} _ {\\text {s a d}}, \\tag {5}", + "image_path": "917e24705fb561bab313683ef8edb2462a88b4db1f935c3d04ed1ffcf0f27665.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 293, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 545, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 293, + 545, + 365 + ], + "type": "inline_equation", + "content": "L_{pix}" + }, + { + "bbox": [ + 305, + 293, + 545, + 365 + ], + "type": "text", + "content": " is a pixel-wise reconstruction loss, " + }, + { + "bbox": [ + 305, + 293, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv}" + }, + { + "bbox": [ + 305, + 293, + 545, + 365 + ], + "type": "text", + "content": " is an adversarial loss, " + }, + { + "bbox": [ + 305, + 293, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 305, + 293, + 545, + 365 + ], + "type": "text", + "content": " is token classification loss, and " + }, + { + "bbox": [ + 305, + 293, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sad}" + }, + { + "bbox": [ + 305, + 293, + 545, + 365 + ], + "type": "text", + "content": " is semantic adversarial loss, while " + }, + { + "bbox": [ + 305, + 293, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\{\\lambda_{adv},\\lambda_{cls},\\lambda_{sad}\\}" + }, + { + "bbox": [ + 305, + 293, + 545, + 365 + ], + "type": "text", + "content": " control the weight of associated loss term. Given the predicted video frame " + }, + { + "bbox": [ + 305, + 293, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 305, + 293, + 545, + 365 + ], + "type": "text", + "content": " and its ground truth counterpart " + }, + { + "bbox": [ + 305, + 293, + 545, + 365 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 305, + 293, + 545, + 365 + ], + "type": "text", + "content": ", we employ the L1 pixel-wise reconstruction loss as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 384, + 371, + 545, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 371, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 384, + 371, + 545, + 386 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p i x} = \\left| \\left| \\hat {Y} - Y \\right| \\right| _ {1}. \\tag {6}", + "image_path": "706171454292e377fb1bdb714cc03e0e6a1e1c8cb254970132b0de9066dc82ca.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 394, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 394, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 305, + 394, + 545, + 417 + ], + "type": "text", + "content": "As for the adversarial loss, we train our inpainting network to fool a discriminator " + }, + { + "bbox": [ + 305, + 394, + 545, + 417 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 394, + 545, + 417 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 380, + 425, + 545, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 425, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 380, + 425, + 545, + 440 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a d v} = - E _ {\\hat {Y}} [ D (\\hat {Y}) ]. \\tag {7}", + "image_path": "3950f3fd85d912631ea834b8fb81b2e8c1a487c2fa0cbb04f3b1d7b33e486d3c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 447, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 447, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 305, + 447, + 545, + 471 + ], + "type": "text", + "content": "Based on a temporal patch discriminator [26, 23], the discriminator " + }, + { + "bbox": [ + 305, + 447, + 545, + 471 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 447, + 545, + 471 + ], + "type": "text", + "content": " is trained to optimize" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 309, + 478, + 545, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 478, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 309, + 478, + 545, + 503 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D} = E _ {Y} \\left[ \\operatorname {R e L U} (1 - D (Y)) \\right] + E _ {\\hat {Y}} \\left[ \\operatorname {R e L U} (1 + D (\\hat {Y})) \\right]. \\tag {8}", + "image_path": "26a691c7f6e388de4e1f9827e12c252f98b1fa4238d7cd6e7ae9a8d75b3841b4.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 504, + 487, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 504, + 487, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 504, + 487, + 515 + ], + "type": "text", + "content": "The token classification loss term is given as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 343, + 523, + 545, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 343, + 523, + 545, + 555 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} = - \\sum_ {i} \\sum_ {c = 1} ^ {C} y _ {i} ^ {c} \\log \\left(\\text {s o f t m a x} \\left(\\hat {y} _ {i}\\right)\\right), \\tag {9}", + "image_path": "71a1cf1fdf110f1f1a8ec6e6314684bfddf1a3aca00bbe0dff6c39087d3ca015.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "type": "inline_equation", + "content": "\\hat{y}_i\\in \\mathbb{R}^C" + }, + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "type": "text", + "content": " is the predicted score from the single-layer classifier conditioned on the output tokens from the semantic-aware dynamic block, while " + }, + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "type": "inline_equation", + "content": "y_{i}^{c}" + }, + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "type": "text", + "content": " is the target probability for class " + }, + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "type": "text", + "content": "-th token. For the ground truth target " + }, + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "type": "inline_equation", + "content": "y_{i}^{c}" + }, + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "type": "text", + "content": ", we use the input semantic segmentation map (i.e., " + }, + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "type": "inline_equation", + "content": "y_{i}^{c} = s_{i}(c)" + }, + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "type": "text", + "content": "). The semantic-aware adversarial loss " + }, + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sad}" + }, + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "type": "text", + "content": " is similar to " + }, + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv}" + }, + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "type": "text", + "content": " in that (8) and (7) are used, but with the semantic-aware discriminator (SAD) instead of " + }, + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 563, + 545, + 660 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 666, + 387, + 678 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 666, + 387, + 678 + ], + "spans": [ + { + "bbox": [ + 306, + 666, + 387, + 678 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 680, + 545, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 680, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 305, + 680, + 545, + 716 + ], + "type": "text", + "content": "In this section, we quantitatively and qualitatively demonstrate the effectiveness of SAVIT. Please refer to the supplementary material for more experimental results." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12953" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 509, + 188 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 509, + 188 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 509, + 188 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 509, + 188 + ], + "type": "table", + "html": "
YouTube-VOS [36]DAVIS [28]
MethodPSNR↑SSIM↑VFID↓E_warp(%)↓PSNR↑SSIM↑VFID↓E_warp(%)↓
VINet [15]29.200.94340.0720.149028.960.94110.1990.1785
LGTSM [3]29.740.95040.0700.185928.570.94090.1700.1640
CAP [21]31.580.96070.0710.147030.280.95210.1820.1533
STTN [39]32.340.96550.0530.090730.670.95600.1490.1449
E2FGVI [23]33.710.97000.0460.086433.010.97210.1160.1315
ISVI [41]31.190.95690.0530.117332.280.96690.1290.1589
FuseFormer [26]33.160.96730.0510.090032.540.97000.1380.1362
SAVIT (Ours)33.970.97270.0430.043633.140.97480.1070.0673
", + "image_path": "4870b981a1c5e6971b8ce26899f6f126eff11ef3b2c1467013882087aeb7be56.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 234, + 274, + 302 + ], + "blocks": [ + { + "bbox": [ + 46, + 190, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 228 + ], + "type": "text", + "content": "Table 1: Quantitative comparison of SAVIT against state-of-the-art methods. SAVIT outperforms the existing approaches in terms of various evaluation metrics. Notably, SAVIT significantly elevates the inpainting performance of FuseFormer [26], which is our baseline architecture, witnessing the superiority of leveraging semantic cues within the input video." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 234, + 274, + 302 + ], + "lines": [ + { + "bbox": [ + 58, + 234, + 274, + 302 + ], + "spans": [ + { + "bbox": [ + 58, + 234, + 274, + 302 + ], + "type": "table", + "html": "
ConfigurationPSNR↑VFID↓
AFuseFormersmall [26]30.510.182
B+ super-categorized segmentation map30.900.161
C+ SAL30.920.164
D+ token classification loss (λcls)31.190.165
E+ SAD31.010.157
", + "image_path": "47b4c628aef37a12370a5b7acd28a8bcd2ca46da086d404859a1073a7a4fa63b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 304, + 288, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 304, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 288, + 373 + ], + "type": "text", + "content": "Table 2: Ablation experiments on various network configurations under DAVIS [28] dataset. Starting from the performance of the baseline architecture [26], the inpainting performance gradually increases as the configurations are added since the network can benefit from the semantic cues within a video." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 385, + 178, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 385, + 178, + 398 + ], + "spans": [ + { + "bbox": [ + 47, + 385, + 178, + 398 + ], + "type": "text", + "content": "4.1. Implementation details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 567 + ], + "type": "text", + "content": "Datasets and evaluation metrics. We use two conventional video inpainting datasets, YouTube-VOS [36] and DAVIS [28], for our experiments. YouTube-VOS dataset consists of 3471 training, 474 validation, and 508 test videos from mixed scenes. Meanwhile, DAVIS dataset comprises 150 video clips, 60 of which are used for training and the remaining 90 clips are used for evaluation. Following previous works [26, 23], we use only the training videos in the YouTube-VOS for training, while the evaluation is conducted on YouTube-VOS testset and 50 selected clips of DAVIS testset. During training, we create random mask shapes for input masked frames as in [39, 26, 23] and use the method discussed in Section 3.1 to obtain a segmentation map for the entire dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 568, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 688 + ], + "type": "text", + "content": "The inpainting performance is evaluated under two standard scenarios: fixed region inpainting and foreground object removal. For the fixed region inpainting, we use free-form fixed masks for each video provided by FuseFormer [26]. As for the foreground object removal, we remove the main objects using frame-by-frame annotations provided by the stakeholders. Notably, we employ a simple k-nearest neighbors algorithm to complete the missing object region in the segmentation maps for the object removal task (please refer to supplementary material for detail)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "We compare the results with respect to four evaluation metrics: PSNR, SSIM, temporal warping error " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": " [20]," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 236, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 545, + 308 + ], + "type": "text", + "content": "and Video-based Fréchet Inception Distance (VFID) [35]. PSNR and SSIM are traditional metrics used to measure the average quality of produced frames. VFID and " + }, + { + "bbox": [ + 304, + 236, + 545, + 308 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}" + }, + { + "bbox": [ + 304, + 236, + 545, + 308 + ], + "type": "text", + "content": " measure the video's perceptual quality and temporal stability (consistency), respectively, and low values indicate better quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 312, + 546, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 546, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 546, + 443 + ], + "type": "text", + "content": "Training settings. We follow the training settings of our baseline network FuseFormer [26]. Specifically, we randomly sample five frames from a video, resize the frames to " + }, + { + "bbox": [ + 304, + 312, + 546, + 443 + ], + "type": "inline_equation", + "content": "240 \\times 432" + }, + { + "bbox": [ + 304, + 312, + 546, + 443 + ], + "type": "text", + "content": ", and utilize random horizontal flipping augmentation. The network is trained for " + }, + { + "bbox": [ + 304, + 312, + 546, + 443 + ], + "type": "inline_equation", + "content": "500k" + }, + { + "bbox": [ + 304, + 312, + 546, + 443 + ], + "type": "text", + "content": " iterations with Adam optimizer [16] and a batch size of eight. The initial learning rate is 1e-4 and decreases to 1e-5 after " + }, + { + "bbox": [ + 304, + 312, + 546, + 443 + ], + "type": "inline_equation", + "content": "400k" + }, + { + "bbox": [ + 304, + 312, + 546, + 443 + ], + "type": "text", + "content": " iterations. The coefficients for our loss functions, " + }, + { + "bbox": [ + 304, + 312, + 546, + 443 + ], + "type": "inline_equation", + "content": "\\lambda_{adv}" + }, + { + "bbox": [ + 304, + 312, + 546, + 443 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 312, + 546, + 443 + ], + "type": "inline_equation", + "content": "\\lambda_{cls}" + }, + { + "bbox": [ + 304, + 312, + 546, + 443 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 312, + 546, + 443 + ], + "type": "inline_equation", + "content": "\\lambda_{sad}" + }, + { + "bbox": [ + 304, + 312, + 546, + 443 + ], + "type": "text", + "content": " are 0.01, 0.01, and 0.001, respectively. We implement our network on the PyTorch framework and use eight NVIDIA RTX A6000 GPUs for training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 445, + 440, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 445, + 440, + 458 + ], + "spans": [ + { + "bbox": [ + 305, + 445, + 440, + 458 + ], + "type": "text", + "content": "4.2. Video inpainting results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 459, + 546, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 546, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 546, + 639 + ], + "type": "text", + "content": "Quantitative results. Table 1 quantitatively compares our video inpainting results on fixed region inpainting setting against seven state-of-the-art video inpainting networks: VINet [15], LGTSM [3], CAP [21], STTN [39], E2FVGI [23], ISVI [41], and FuseFormer [26]. The result shows that our method remarkably elevates the performance of FuseFormer [26], on which we base our architecture, and achieves superior performances on both YouTube-VOS and DAVIS datasets compared to the state-of-the-art methods. Notably, the warping error has also significantly improved, indicating that semantic information benefits in generating temporally consistent video. These results show the effectiveness of employing mixture-of-experts to dynamically leverage the semantic cues within given frames for video inpainting." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Qualitative results. Fig. 5 shows that SAVIT has visually pleasing inpainting results, especially in synthesizing semantic structure, boundary, and texture of objects. Notably, unlike the baseline methods that rely solely on inherent video clues and cannot accurately restore the shape of the human body, SAVIT effectively completes the region" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12954" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 59, + 168, + 106 + ], + "blocks": [ + { + "bbox": [ + 85, + 49, + 157, + 58 + ], + "lines": [ + { + "bbox": [ + 85, + 49, + 157, + 58 + ], + "spans": [ + { + "bbox": [ + 85, + 49, + 157, + 58 + ], + "type": "text", + "content": "Masked Frames" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 59, + 168, + 106 + ], + "lines": [ + { + "bbox": [ + 85, + 59, + 168, + 106 + ], + "spans": [ + { + "bbox": [ + 85, + 59, + 168, + 106 + ], + "type": "image", + "image_path": "8f9ef754c0e2f525e324a8415e1ff9fad5cd58bd9f064040d0f75fe35a052a96.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 169, + 59, + 256, + 106 + ], + "blocks": [ + { + "bbox": [ + 200, + 49, + 223, + 57 + ], + "lines": [ + { + "bbox": [ + 200, + 49, + 223, + 57 + ], + "spans": [ + { + "bbox": [ + 200, + 49, + 223, + 57 + ], + "type": "text", + "content": "STTN" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 169, + 59, + 256, + 106 + ], + "lines": [ + { + "bbox": [ + 169, + 59, + 256, + 106 + ], + "spans": [ + { + "bbox": [ + 169, + 59, + 256, + 106 + ], + "type": "image", + "image_path": "2e12573d17e4ed4bbdef05d514d4cfcb748be6c60ec3ea4406cc4f14a89dcc42.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 256, + 59, + 340, + 108 + ], + "blocks": [ + { + "bbox": [ + 272, + 49, + 317, + 58 + ], + "lines": [ + { + "bbox": [ + 272, + 49, + 317, + 58 + ], + "spans": [ + { + "bbox": [ + 272, + 49, + 317, + 58 + ], + "type": "text", + "content": "FuseFormer" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 256, + 59, + 340, + 108 + ], + "lines": [ + { + "bbox": [ + 256, + 59, + 340, + 108 + ], + "spans": [ + { + "bbox": [ + 256, + 59, + 340, + 108 + ], + "type": "image", + "image_path": "83c9e68628e4dbb172d29abcd1c981678252f40549c178078f8137a4c8933100.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 340, + 59, + 422, + 108 + ], + "blocks": [ + { + "bbox": [ + 362, + 49, + 393, + 58 + ], + "lines": [ + { + "bbox": [ + 362, + 49, + 393, + 58 + ], + "spans": [ + { + "bbox": [ + 362, + 49, + 393, + 58 + ], + "type": "text", + "content": "E2FGVI" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 59, + 422, + 108 + ], + "lines": [ + { + "bbox": [ + 340, + 59, + 422, + 108 + ], + "spans": [ + { + "bbox": [ + 340, + 59, + 422, + 108 + ], + "type": "image", + "image_path": "db5af5a1e67cd1cd0acaf227123157af4be6685c972685b42ce5c94e51797cd3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 422, + 59, + 507, + 108 + ], + "blocks": [ + { + "bbox": [ + 436, + 49, + 487, + 59 + ], + "lines": [ + { + "bbox": [ + 436, + 49, + 487, + 59 + ], + "spans": [ + { + "bbox": [ + 436, + 49, + 487, + 59 + ], + "type": "text", + "content": "SAVIT (Ours)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 422, + 59, + 507, + 108 + ], + "lines": [ + { + "bbox": [ + 422, + 59, + 507, + 108 + ], + "spans": [ + { + "bbox": [ + 422, + 59, + 507, + 108 + ], + "type": "image", + "image_path": "61ddfec66c64ef64f1e9d49cf480898400a41ac45fb5c6bfe0be325c09473e8b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 85, + 107, + 168, + 152 + ], + "blocks": [ + { + "bbox": [ + 85, + 107, + 168, + 152 + ], + "lines": [ + { + "bbox": [ + 85, + 107, + 168, + 152 + ], + "spans": [ + { + "bbox": [ + 85, + 107, + 168, + 152 + ], + "type": "image", + "image_path": "7c4147a3778a7f13a666cc6763133e5c893c7817b42f7aa996fffa7b08e93034.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 169, + 107, + 256, + 154 + ], + "blocks": [ + { + "bbox": [ + 169, + 107, + 256, + 154 + ], + "lines": [ + { + "bbox": [ + 169, + 107, + 256, + 154 + ], + "spans": [ + { + "bbox": [ + 169, + 107, + 256, + 154 + ], + "type": "image", + "image_path": "4317c7295e4383c139139d008c49d540460ab19648ef61ed0cf72cbaef892e5e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 256, + 107, + 339, + 154 + ], + "blocks": [ + { + "bbox": [ + 256, + 107, + 339, + 154 + ], + "lines": [ + { + "bbox": [ + 256, + 107, + 339, + 154 + ], + "spans": [ + { + "bbox": [ + 256, + 107, + 339, + 154 + ], + "type": "image", + "image_path": "ecaeab738521227e51db4e6f4d7f990f33ba22d70d047337a3000c2d4e66acca.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 340, + 107, + 422, + 154 + ], + "blocks": [ + { + "bbox": [ + 340, + 107, + 422, + 154 + ], + "lines": [ + { + "bbox": [ + 340, + 107, + 422, + 154 + ], + "spans": [ + { + "bbox": [ + 340, + 107, + 422, + 154 + ], + "type": "image", + "image_path": "596c988000b2e45b2463901bdaf47f6f65d7aa85adf85a2b3a5baf6a4af84b05.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 422, + 107, + 507, + 155 + ], + "blocks": [ + { + "bbox": [ + 422, + 107, + 507, + 155 + ], + "lines": [ + { + "bbox": [ + 422, + 107, + 507, + 155 + ], + "spans": [ + { + "bbox": [ + 422, + 107, + 507, + 155 + ], + "type": "image", + "image_path": "d38d9b0ac28599f1352c93ab68bfc66c176a0638a94010869bb1a98aa065e1e3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 85, + 154, + 168, + 200 + ], + "blocks": [ + { + "bbox": [ + 85, + 154, + 168, + 200 + ], + "lines": [ + { + "bbox": [ + 85, + 154, + 168, + 200 + ], + "spans": [ + { + "bbox": [ + 85, + 154, + 168, + 200 + ], + "type": "image", + "image_path": "a1a24f9ceac5f5e6855c058b28efb731e8dea5b271b002eeb283b9b5c66c6125.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 169, + 154, + 256, + 201 + ], + "blocks": [ + { + "bbox": [ + 169, + 154, + 256, + 201 + ], + "lines": [ + { + "bbox": [ + 169, + 154, + 256, + 201 + ], + "spans": [ + { + "bbox": [ + 169, + 154, + 256, + 201 + ], + "type": "image", + "image_path": "7ffc4d195e1f7d7db840bb62f8f3a22634d9d0814194e2b3aaa1e8ca4a05858b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 256, + 154, + 339, + 201 + ], + "blocks": [ + { + "bbox": [ + 256, + 154, + 339, + 201 + ], + "lines": [ + { + "bbox": [ + 256, + 154, + 339, + 201 + ], + "spans": [ + { + "bbox": [ + 256, + 154, + 339, + 201 + ], + "type": "image", + "image_path": "37a9502c5f70e0f52997fd5faf314adf0d73e5c8aa12dffbaa726d0e6a2ffca3.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 340, + 154, + 422, + 201 + ], + "blocks": [ + { + "bbox": [ + 340, + 154, + 422, + 201 + ], + "lines": [ + { + "bbox": [ + 340, + 154, + 422, + 201 + ], + "spans": [ + { + "bbox": [ + 340, + 154, + 422, + 201 + ], + "type": "image", + "image_path": "650781ecd42b506da60b43ac943a79287cd01d68a99b4e566ebc8a09b2b0a2d4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 422, + 154, + 507, + 201 + ], + "blocks": [ + { + "bbox": [ + 422, + 154, + 507, + 201 + ], + "lines": [ + { + "bbox": [ + 422, + 154, + 507, + 201 + ], + "spans": [ + { + "bbox": [ + 422, + 154, + 507, + 201 + ], + "type": "image", + "image_path": "7df933a9d73d7f4f7bd3ad244b9e0a1deeef2701e8cd1a7628d7a6f83b3c36db.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 85, + 201, + 168, + 248 + ], + "blocks": [ + { + "bbox": [ + 85, + 201, + 168, + 248 + ], + "lines": [ + { + "bbox": [ + 85, + 201, + 168, + 248 + ], + "spans": [ + { + "bbox": [ + 85, + 201, + 168, + 248 + ], + "type": "image", + "image_path": "5e557ee2e7efeac777aa2741c8494cb4f211df93eb81c60d5d6a36303eff8b3f.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 169, + 201, + 256, + 248 + ], + "blocks": [ + { + "bbox": [ + 169, + 201, + 256, + 248 + ], + "lines": [ + { + "bbox": [ + 169, + 201, + 256, + 248 + ], + "spans": [ + { + "bbox": [ + 169, + 201, + 256, + 248 + ], + "type": "image", + "image_path": "dcf1dd3bbcea01e073c1197ba4cbebcd62a0a184871f1a5ca58fdcabbe11c58f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 256, + 201, + 339, + 248 + ], + "blocks": [ + { + "bbox": [ + 256, + 201, + 339, + 248 + ], + "lines": [ + { + "bbox": [ + 256, + 201, + 339, + 248 + ], + "spans": [ + { + "bbox": [ + 256, + 201, + 339, + 248 + ], + "type": "image", + "image_path": "42fd4d9da2aab0f05f9d2a04a54016784b64d742ba53a31b6a55936660320420.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 340, + 201, + 422, + 248 + ], + "blocks": [ + { + "bbox": [ + 340, + 201, + 422, + 248 + ], + "lines": [ + { + "bbox": [ + 340, + 201, + 422, + 248 + ], + "spans": [ + { + "bbox": [ + 340, + 201, + 422, + 248 + ], + "type": "image", + "image_path": "d93db6a3e4bbe625e5f75f2b0655a892517bbc3be5b6146d13c362c99a0414bf.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 422, + 201, + 507, + 249 + ], + "blocks": [ + { + "bbox": [ + 422, + 201, + 507, + 249 + ], + "lines": [ + { + "bbox": [ + 422, + 201, + 507, + 249 + ], + "spans": [ + { + "bbox": [ + 422, + 201, + 507, + 249 + ], + "type": "image", + "image_path": "553275ed3a56e3b5667a7bdab89edf9dd46a4fcc0d412a38112262cb2cfdd065.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 85, + 249, + 168, + 296 + ], + "blocks": [ + { + "bbox": [ + 85, + 249, + 168, + 296 + ], + "lines": [ + { + "bbox": [ + 85, + 249, + 168, + 296 + ], + "spans": [ + { + "bbox": [ + 85, + 249, + 168, + 296 + ], + "type": "image", + "image_path": "c828aff8235dcc0408eb3b55ccdd947e47ef8ef765be8108fdff4cf5d7ec5bed.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 169, + 249, + 256, + 296 + ], + "blocks": [ + { + "bbox": [ + 169, + 249, + 256, + 296 + ], + "lines": [ + { + "bbox": [ + 169, + 249, + 256, + 296 + ], + "spans": [ + { + "bbox": [ + 169, + 249, + 256, + 296 + ], + "type": "image", + "image_path": "4c0f95f58d30ef09e84edbffc1a1b192cdbe5f452721b1cf40217a51f2e623af.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 256, + 249, + 339, + 296 + ], + "blocks": [ + { + "bbox": [ + 256, + 249, + 339, + 296 + ], + "lines": [ + { + "bbox": [ + 256, + 249, + 339, + 296 + ], + "spans": [ + { + "bbox": [ + 256, + 249, + 339, + 296 + ], + "type": "image", + "image_path": "e3c27885d3d128e97fba23afbbce14e479568f22a10023d3c1bfe49bf2c8b4a6.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 340, + 249, + 422, + 296 + ], + "blocks": [ + { + "bbox": [ + 340, + 249, + 422, + 296 + ], + "lines": [ + { + "bbox": [ + 340, + 249, + 422, + 296 + ], + "spans": [ + { + "bbox": [ + 340, + 249, + 422, + 296 + ], + "type": "image", + "image_path": "17c81c33b45fe4596f76f1f2309547b48631b4081b48ed8a112cb15b1a55cd37.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 422, + 249, + 507, + 296 + ], + "blocks": [ + { + "bbox": [ + 422, + 249, + 507, + 296 + ], + "lines": [ + { + "bbox": [ + 422, + 249, + 507, + 296 + ], + "spans": [ + { + "bbox": [ + 422, + 249, + 507, + 296 + ], + "type": "image", + "image_path": "24999c0908660f75dffaefda4aacc2d52fc0ebbcdfd15e9cf96a9c276f4ac988.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 85, + 297, + 168, + 343 + ], + "blocks": [ + { + "bbox": [ + 85, + 297, + 168, + 343 + ], + "lines": [ + { + "bbox": [ + 85, + 297, + 168, + 343 + ], + "spans": [ + { + "bbox": [ + 85, + 297, + 168, + 343 + ], + "type": "image", + "image_path": "f0f4cb76128310bec266cfe7d02c9a6186e03a5ac2bcba8fdb10ef43e18e283c.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 354, + 545, + 378 + ], + "lines": [ + { + "bbox": [ + 46, + 354, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 545, + 378 + ], + "type": "text", + "content": "Figure 5: Qualitative comparison of SAVIT with baseline video inpainting networks in fixed region inpainting and object removal. SAVIT more accurately recovers the object's boundaries and appropriate textures." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 169, + 297, + 256, + 344 + ], + "blocks": [ + { + "bbox": [ + 169, + 297, + 256, + 344 + ], + "lines": [ + { + "bbox": [ + 169, + 297, + 256, + 344 + ], + "spans": [ + { + "bbox": [ + 169, + 297, + 256, + 344 + ], + "type": "image", + "image_path": "bca9e385492a285e57d66712941e82a4338c0d4129693e942c32479790841a37.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 256, + 297, + 339, + 344 + ], + "blocks": [ + { + "bbox": [ + 256, + 297, + 339, + 344 + ], + "lines": [ + { + "bbox": [ + 256, + 297, + 339, + 344 + ], + "spans": [ + { + "bbox": [ + 256, + 297, + 339, + 344 + ], + "type": "image", + "image_path": "9f42f31a2daef94f9b2b42752ca633f7b518305740b6117c0885179112dcfbe2.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 340, + 297, + 422, + 344 + ], + "blocks": [ + { + "bbox": [ + 340, + 297, + 422, + 344 + ], + "lines": [ + { + "bbox": [ + 340, + 297, + 422, + 344 + ], + "spans": [ + { + "bbox": [ + 340, + 297, + 422, + 344 + ], + "type": "image", + "image_path": "cb457ca45898752bc9c94b4e07c1966d9b8b20b1cb5d2f9cbd80f4781d8335d0.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 422, + 297, + 507, + 344 + ], + "blocks": [ + { + "bbox": [ + 422, + 297, + 507, + 344 + ], + "lines": [ + { + "bbox": [ + 422, + 297, + 507, + 344 + ], + "spans": [ + { + "bbox": [ + 422, + 297, + 507, + 344 + ], + "type": "image", + "image_path": "13229ab8c7ddaeef1b4fb360903aaa689533aff721ecdfc4730ce24fae886665.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 48, + 384, + 107, + 441 + ], + "blocks": [ + { + "bbox": [ + 48, + 384, + 107, + 441 + ], + "lines": [ + { + "bbox": [ + 48, + 384, + 107, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 384, + 107, + 441 + ], + "type": "image", + "image_path": "e1640d5d2232cd6f429e96210ddf9ca508d1ce50b726c4733789844f3c55c9aa.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 442, + 114, + 465 + ], + "lines": [ + { + "bbox": [ + 47, + 442, + 114, + 465 + ], + "spans": [ + { + "bbox": [ + 47, + 442, + 114, + 465 + ], + "type": "text", + "content": "Figure 6: User study result." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "table", + "bbox": [ + 142, + 387, + 267, + 439 + ], + "blocks": [ + { + "bbox": [ + 142, + 387, + 267, + 439 + ], + "lines": [ + { + "bbox": [ + 142, + 387, + 267, + 439 + ], + "spans": [ + { + "bbox": [ + 142, + 387, + 267, + 439 + ], + "type": "table", + "html": "
ConfigurationPSNR†
hardmax (SAVIT)31.01
softmax30.91
random routing weight19.86
", + "image_path": "c307646ea73a9204afc13553f891a750ba243675718221088d8c5b4d33b39756.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "table_body" + } + ], + "index": 38 + }, + { + "bbox": [ + 127, + 442, + 284, + 465 + ], + "lines": [ + { + "bbox": [ + 127, + 442, + 284, + 465 + ], + "spans": [ + { + "bbox": [ + 127, + 442, + 284, + 465 + ], + "type": "text", + "content": "Table 3: Effect of the routing weight on the DAVIS [28] dataset." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 478, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 478, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 287, + 538 + ], + "type": "text", + "content": "by leveraging additional semantic information (third and fourth rows). Furthermore, SAVIT demonstrates the capability of generating more semantically plausible textures, as observed from the detailed textures of synthesized grass and wall (second and fifth rows)." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 46, + 544, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 664 + ], + "type": "text", + "content": "User study. We conduct a user study on the foreground object removal task, where ground truth target video frames for evaluation are not available, to investigate the real-world applicability of our method using DAVIS [28] testset. We ask 24 raters to choose a better video between two randomly ordered videos, including the inpainted results by FuseFormer [26] and ours, fifteen times per rater. The result presented in Fig. 6 shows that raters prefer ours to the baseline, indicating the impact of exploiting semantic cues in completing missing object regions." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 47, + 671, + 145, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 145, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 145, + 684 + ], + "type": "text", + "content": "5. Ablation studies" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 46, + 687, + 287, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 687, + 287, + 723 + ], + "spans": [ + { + "bbox": [ + 46, + 687, + 287, + 723 + ], + "type": "text", + "content": "We perform ablation experiments based on a smaller FuseFormer network (FuseFormersmall) by halving the dimension of the hidden layer." + } + ] + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 307, + 388, + 424, + 453 + ], + "blocks": [ + { + "bbox": [ + 307, + 388, + 424, + 453 + ], + "lines": [ + { + "bbox": [ + 307, + 388, + 424, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 424, + 453 + ], + "type": "image", + "image_path": "7d79bd840aa5d68ad9ebdacbe7ef6254649b43d37d35a6105892e0055d143ebc.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 454, + 390, + 462 + ], + "lines": [ + { + "bbox": [ + 341, + 454, + 390, + 462 + ], + "spans": [ + { + "bbox": [ + 341, + 454, + 390, + 462 + ], + "type": "text", + "content": "SAVIT w/o SAD" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 464, + 545, + 498 + ], + "lines": [ + { + "bbox": [ + 306, + 464, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 545, + 498 + ], + "type": "text", + "content": "Figure 7: Effect of semantic-aware discriminator (SAD). The texture of the black swan's body is blurry in the case without SAD." + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 427, + 387, + 545, + 453 + ], + "blocks": [ + { + "bbox": [ + 427, + 387, + 545, + 453 + ], + "lines": [ + { + "bbox": [ + 427, + 387, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 427, + 387, + 545, + 453 + ], + "type": "image", + "image_path": "7b4fe8511bf1a6cea21b03c4df82bc973f361a247cdf9aa04169684d0c5aae11.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 464, + 455, + 509, + 462 + ], + "lines": [ + { + "bbox": [ + 464, + 455, + 509, + 462 + ], + "spans": [ + { + "bbox": [ + 464, + 455, + 509, + 462 + ], + "type": "text", + "content": "SAVIT w/ SAD" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Network configuration. Table 2 compares the inpainting performance for various configurations of SAVIT. First, we observe that the baseline network's performance improves with additional segmentation map inputs (rows A and B). However, a subtle additional performance improvement (0.02dB PSNR gain, rows B and C) is achieved when we naively add our proposed dynamic linear layer to the network. Rather, applying the proposed token-wise classification loss to the objective boosts the performance (0.2dB PSNR gain, row D). Overall performance improvement by applying mixture-of-experts (rows C and D) highlights the importance and effectiveness of employing mixture-of-experts to handle semantics dynamically. Moreover, using a semantic-aware dynamic discriminator (row E) degrades PSNR while improving VFID. However, as shown in Fig. 7, utilizing the semantic-aware discriminator produces more visually plausible results, which solidifies the advantage of" + } + ] + } + ], + "index": 49 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12955" + } + ] + } + ], + "index": 50 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 68, + 111, + 110 + ], + "blocks": [ + { + "bbox": [ + 50, + 68, + 111, + 110 + ], + "lines": [ + { + "bbox": [ + 50, + 68, + 111, + 110 + ], + "spans": [ + { + "bbox": [ + 50, + 68, + 111, + 110 + ], + "type": "image", + "image_path": "7e6a537c1d41f40c7c2e54af2f8543267c4121a3c6aa7101e40a7c65fc67c471.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 112, + 68, + 174, + 110 + ], + "blocks": [ + { + "bbox": [ + 112, + 68, + 174, + 110 + ], + "lines": [ + { + "bbox": [ + 112, + 68, + 174, + 110 + ], + "spans": [ + { + "bbox": [ + 112, + 68, + 174, + 110 + ], + "type": "image", + "image_path": "683f84253a3d4df672db5955b8591fbc2ca835d10e89c5ec21dfbeccd40c1a70.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 175, + 68, + 235, + 110 + ], + "blocks": [ + { + "bbox": [ + 175, + 68, + 235, + 110 + ], + "lines": [ + { + "bbox": [ + 175, + 68, + 235, + 110 + ], + "spans": [ + { + "bbox": [ + 175, + 68, + 235, + 110 + ], + "type": "image", + "image_path": "11ccf9fa0c90e978f58ca19e20a7cac7e10d8d28e023726e7caa73f05c919c7a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 236, + 68, + 298, + 110 + ], + "blocks": [ + { + "bbox": [ + 236, + 68, + 298, + 110 + ], + "lines": [ + { + "bbox": [ + 236, + 68, + 298, + 110 + ], + "spans": [ + { + "bbox": [ + 236, + 68, + 298, + 110 + ], + "type": "image", + "image_path": "a8d45a3c5ca0fdbd138b0ba804b9beae8318c856b29c33689fde41ac4da28eca.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 298, + 68, + 359, + 110 + ], + "blocks": [ + { + "bbox": [ + 298, + 68, + 359, + 110 + ], + "lines": [ + { + "bbox": [ + 298, + 68, + 359, + 110 + ], + "spans": [ + { + "bbox": [ + 298, + 68, + 359, + 110 + ], + "type": "image", + "image_path": "fb3a4e1974f878c4a2d0956f15bd8ac4a1cef61e231d6fe7729a0c353fa4e9b7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 359, + 68, + 420, + 110 + ], + "blocks": [ + { + "bbox": [ + 359, + 68, + 420, + 110 + ], + "lines": [ + { + "bbox": [ + 359, + 68, + 420, + 110 + ], + "spans": [ + { + "bbox": [ + 359, + 68, + 420, + 110 + ], + "type": "image", + "image_path": "77a8f6095b22a6baff8c284663afe9ad4981f940925da0c6a9cecb1f2b7d1b2d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 115, + 111, + 156 + ], + "blocks": [ + { + "bbox": [ + 50, + 115, + 111, + 156 + ], + "lines": [ + { + "bbox": [ + 50, + 115, + 111, + 156 + ], + "spans": [ + { + "bbox": [ + 50, + 115, + 111, + 156 + ], + "type": "image", + "image_path": "b022bd318f8cd0d92db791a753f0df51156f333fe89c87dbfe32607c873c0515.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 158, + 421, + 194 + ], + "lines": [ + { + "bbox": [ + 47, + 158, + 421, + 194 + ], + "spans": [ + { + "bbox": [ + 47, + 158, + 421, + 194 + ], + "type": "text", + "content": "Figure 8: Visualization of the semantic router results. (1) to (8) are the visualization of routing weights of the corresponding super-category. The routing weights are rearranged to the image-like features. Brighter pixel values indicate higher routing weights." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 112, + 115, + 174, + 156 + ], + "blocks": [ + { + "bbox": [ + 112, + 115, + 174, + 156 + ], + "lines": [ + { + "bbox": [ + 112, + 115, + 174, + 156 + ], + "spans": [ + { + "bbox": [ + 112, + 115, + 174, + 156 + ], + "type": "image", + "image_path": "caf30627110a1a0f1af152c9b326d99547050e2cc717445342f9dec54715d7e5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 175, + 115, + 235, + 156 + ], + "blocks": [ + { + "bbox": [ + 175, + 115, + 235, + 156 + ], + "lines": [ + { + "bbox": [ + 175, + 115, + 235, + 156 + ], + "spans": [ + { + "bbox": [ + 175, + 115, + 235, + 156 + ], + "type": "image", + "image_path": "e7f834c23e09fd6ebd51fa147a7a1c3a815dd1dd386c76e6d98730eb5e259917.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 236, + 115, + 298, + 156 + ], + "blocks": [ + { + "bbox": [ + 236, + 115, + 298, + 156 + ], + "lines": [ + { + "bbox": [ + 236, + 115, + 298, + 156 + ], + "spans": [ + { + "bbox": [ + 236, + 115, + 298, + 156 + ], + "type": "image", + "image_path": "86c02db76eb995e3c627518ee35bf74f1185d927e568fc96194ffdf13e483d9c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 298, + 115, + 359, + 156 + ], + "blocks": [ + { + "bbox": [ + 298, + 115, + 359, + 156 + ], + "lines": [ + { + "bbox": [ + 298, + 115, + 359, + 156 + ], + "spans": [ + { + "bbox": [ + 298, + 115, + 359, + 156 + ], + "type": "image", + "image_path": "36285cb7505347e38e657601e9874d0e418bb7a51053a0d0ba0493839ba9fe31.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 359, + 115, + 420, + 156 + ], + "blocks": [ + { + "bbox": [ + 359, + 115, + 420, + 156 + ], + "lines": [ + { + "bbox": [ + 359, + 115, + 420, + 156 + ], + "spans": [ + { + "bbox": [ + 359, + 115, + 420, + 156 + ], + "type": "image", + "image_path": "ef1a5e631bc5ee88923b3a65cc4799f2042c7af9ec19d28047201a5ca510acca.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 427, + 70, + 542, + 121 + ], + "blocks": [ + { + "bbox": [ + 427, + 70, + 542, + 121 + ], + "lines": [ + { + "bbox": [ + 427, + 70, + 542, + 121 + ], + "spans": [ + { + "bbox": [ + 427, + 70, + 542, + 121 + ], + "type": "table", + "html": "
# ExpertsPSNR↑VFID↓
430.830.165
630.900.168
831.010.157
", + "image_path": "eb94e0b8cacb880f280f963286a163d8a1646f2f6bcde629380f53f81d7c2ec5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 48, + 201, + 168, + 237 + ], + "blocks": [ + { + "bbox": [ + 48, + 201, + 168, + 237 + ], + "lines": [ + { + "bbox": [ + 48, + 201, + 168, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 201, + 168, + 237 + ], + "type": "image", + "image_path": "a34b416150bc72a0c10c74a5618c4197e585ae0fd65437606b1f410293f19073.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 243, + 168, + 277 + ], + "lines": [ + { + "bbox": [ + 47, + 243, + 168, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 243, + 168, + 277 + ], + "type": "text", + "content": "Figure 9: Qualitative comparison on different routing weight strategies." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 175, + 202, + 284, + 240 + ], + "blocks": [ + { + "bbox": [ + 425, + 124, + 543, + 193 + ], + "lines": [ + { + "bbox": [ + 425, + 124, + 543, + 193 + ], + "spans": [ + { + "bbox": [ + 425, + 124, + 543, + 193 + ], + "type": "text", + "content": "Table 4: Inpainting results by changing the numbers of experts under the DAVIS [28] dataset. More experts render better inpainting results." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 175, + 202, + 284, + 240 + ], + "lines": [ + { + "bbox": [ + 175, + 202, + 284, + 240 + ], + "spans": [ + { + "bbox": [ + 175, + 202, + 284, + 240 + ], + "type": "table", + "html": "
# blocks# params.PSNR↑
126.0M30.40
233.1M30.45
", + "image_path": "563d625ad537ac3e52b57ea182197f70af9719ba06347e067b5f16e05b581f56.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 293, + 248, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 248, + 305 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 248, + 305 + ], + "type": "text", + "content": "the discriminator in improving perceptual quality." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 308, + 287, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 429 + ], + "type": "text", + "content": "Visualization of routing weights. Fig. 8 visualizes the routing weights of each expert from the trained semantic router. We rearrange the tokens to image-like features for visualization, and brighter values indicate higher routing weights in the corresponding region. The bright regions among different experts indicate that the experts collaborate during the inpainting procedure rather than simply utilizing a specific expert. We see that routing weights tend to change locally smoothly while the input segmentation map changes abruptly." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 431, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 287, + 586 + ], + "type": "text", + "content": "Designing choice on semantic router. Table 3 compares different design choices of our semantic router. The performance drop by replacing the trained routing weights (i.e., mixture coefficients) with random weights indicates the discriminative role of trained experts. Moreover, the results of softmax routing strategy show a noticeable performance degradation against hardmax strategy quantitatively. Fig. 9 also shows that softmax results tend to have blurry texture, especially on the edge area. Therefore, Hardmax was chosen for our final model to facilitate role division between experts better, as it enables the learning of distinct parameters by dividing the input for each expert before the router blends the token based on its features." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 590, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 662 + ], + "type": "text", + "content": "Impact of the number of experts. Table 4 compares the inpainting performance by changing the number of experts. Notably, as the number of experts (i.e., semantic categories) increases, the network can benefit from more diverse semantic clues in terms of PSNR and VFID values, and we set the number of experts for our final model to eight." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Impact of the number of semantic blocks. Table 5 compares the inpainting performance by applying our semantic module on a different number of blocks. The result indicates that using more semantic blocks produces a slightly" + } + ] + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 329, + 202, + 520, + 262 + ], + "blocks": [ + { + "bbox": [ + 174, + 243, + 284, + 277 + ], + "lines": [ + { + "bbox": [ + 174, + 243, + 284, + 277 + ], + "spans": [ + { + "bbox": [ + 174, + 243, + 284, + 277 + ], + "type": "text", + "content": "Table 5: Ablation on different number of semantic blocks." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 202, + 520, + 262 + ], + "lines": [ + { + "bbox": [ + 329, + 202, + 520, + 262 + ], + "spans": [ + { + "bbox": [ + 329, + 202, + 520, + 262 + ], + "type": "table", + "html": "
MethodMask ratioPSNR†VFID↓
FuseFormersmall30%22.570.554
60%18.101.059
FuseFormersmall + Ours30%23.570.412
60%19.210.750
", + "image_path": "65f8596b3ea6a40f58f013b1318484f4198424a7be21fd7b00fc04542280202c.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 336, + 264, + 515, + 275 + ], + "lines": [ + { + "bbox": [ + 336, + 264, + 515, + 275 + ], + "spans": [ + { + "bbox": [ + 336, + 264, + 515, + 275 + ], + "type": "text", + "content": "Table 6: Comparison on varying mask ratio." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 289, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 545, + 337 + ], + "type": "text", + "content": "better inpainting performance. Although we apply our semantic module one time due to hardware limits for our final model, it is expected that the results will be naturally improved by increasing the number of semantic blocks." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 348, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 348, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 305, + 348, + 545, + 408 + ], + "type": "text", + "content": "Input video with large corruption. Table 6 compares our method against FuseFormersmall under different corrupted region ratio settings. It is shown that, despite larger masks, our method still outperforms the baseline by a large margin by exploiting the semantic information." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 419, + 475, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 419, + 475, + 432 + ], + "spans": [ + { + "bbox": [ + 306, + 419, + 475, + 432 + ], + "type": "text", + "content": "6. Conclusion and Future Works" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 436, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 604 + ], + "type": "text", + "content": "In this study, we have explored leveraging semantic information within a given video for transformer-based video inpainting. Accordingly, we propose an effective semantic-aware dynamic transformer with the notion of mixture-of-experts. Our dynamic transformer contains class-specific expert parameters where experts are efficiently combined to perform token-wise dynamic linear operations based on the token feature and its semantic information. The experiments demonstrate the effectiveness of our method, especially in recovering semantic structures. Although we have utilized pseudo labels for our study, we believe more accurate and denser annotated datasets can improve the inpainting performance further since our method relies on the quality of the segmentation label." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 620, + 408, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 620, + 408, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 620, + 408, + 633 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 642, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 642, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 642, + 545, + 712 + ], + "type": "text", + "content": "This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2023-00222776) and Samsung Electronics Co., Ltd, and Samsung Research Funding Center of Samsung Electronics under Project Number SRFCIT1901-06." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12956" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 189 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 189 + ], + "type": "text", + "content": "[2] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 190, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 190, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 190, + 287, + 235 + ], + "type": "text", + "content": "[3] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Learnable gated temporal shift module for deep video inpainting. In Proceedings of the British Machine Vision Conference (BMVC), 2019. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 290 + ], + "type": "text", + "content": "[4] Cheng Chen, Jiayin Cai, Yao Hu, Xu Tang, Xinggang Wang, Chun Yuan, Xiang Bai, and Song Bai. Deep interactive video inpainting: An invisibility cloak for harry potter. In Proceedings of the ACM International Conference on Multimedia (ACM MM), 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 345 + ], + "type": "text", + "content": "[5] Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolution: Attention over convolution kernels. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 287, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 412 + ], + "type": "text", + "content": "[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 413, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 287, + 456 + ], + "type": "text", + "content": "[7] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 457, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 457, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 457, + 287, + 523 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 524, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 524, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 53, + 524, + 287, + 567 + ], + "type": "text", + "content": "[9] Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 602 + ], + "type": "text", + "content": "[10] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 1991. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "text", + "content": "[11] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. In Advances in Neural Information Processing Systems (NeurIPS), 2016. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[12] Jun-Gyu Jin, Jaehyun Bae, Han-gyul Baek, and Sang-hyo Park. Object-ratio-preserving video retargeting framework based on segmentation and inpainting. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 690, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 715 + ], + "type": "text", + "content": "[13] Jaeyeon Kang, Seoung Wug Oh, and Seon Joo Kim. Error compensation framework for flow-guided video inpainting." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 95, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 149 + ], + "type": "text", + "content": "[14] Bholeshwar Khurana, Soumya Ranjan Dash, Abhishek Bhatia, Aniruddha Mahapatra, Hrituraj Singh, and Kuldeep Kulkarni. Semie: Semantically-aware image extrapolation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 150, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 194 + ], + "type": "text", + "content": "[15] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Deep video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 194, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 194, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 194, + 545, + 226 + ], + "type": "text", + "content": "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "text", + "content": "[17] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "type": "text", + "content": "[18] Xiangtao Kong, Hengyuan Zhao, Yu Qiao, and Chao Dong. Classsr: A general framework to accelerate super-resolution networks by data characteristic. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 327, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 380 + ], + "type": "text", + "content": "[19] Avisek Lahiri, Arnav Kumar Jain, Sanskar Agrawal, Pabitra Mitra, and Prabir Kumar Biswas. Prior guided gan based semantic inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 382, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 545, + 426 + ], + "type": "text", + "content": "[20] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "type": "text", + "content": "[21] Sungho Lee, Seoul Wug Oh, DaeYeun Won, and Seon Joo Kim. Copy-and-paste networks for deep video inpainting. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "text", + "content": "[22] Jiacheng Li, Chang Chen, and Zhiwei Xiong. Contextual outpainting with object-level contrastive learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 514, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 568 + ], + "type": "text", + "content": "[23] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 569, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 569, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 569, + 545, + 624 + ], + "type": "text", + "content": "[24] Liang Liao, Jing Xiao, Zheng Wang, Chia-Wen Lin, and Shin'ichi Satoh. Guidance and evaluation: Semantic-aware image inpainting for mixed scenes. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 625, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 679 + ], + "type": "text", + "content": "[25] Liang Liao, Jing Xiao, Zheng Wang, Chia-Wen Lin, and Shin'ichi Satoh. Image inpainting guided by coherence priors of semantics and textures. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "type": "text", + "content": "[26] Rui Liu, Hanming Deng, Yangyi Huang, Xiaoyu Shi, Lewei Lu, Wenxiu Sun, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fuseformer: Fusing fine-grained information in" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12957" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "transformers for video inpainting. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "text", + "content": "[27] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 288, + 215 + ], + "type": "text", + "content": "[28] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 270 + ], + "type": "text", + "content": "[29] Jingjing Ren, Qingqing Zheng, Yuanyuan Zhao, Xuemiao Xu, and Chen Li. Dlformer: Discrete latent transformer for video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 270, + 288, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 270, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 270, + 288, + 326 + ], + "type": "text", + "content": "[30] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 1, 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 381 + ], + "type": "text", + "content": "[31] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In Proceedings of the International Conference on Learning Representations (ICLR), 2017. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 382, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 288, + 435 + ], + "type": "text", + "content": "[32] Yuhang Song, Chao Yang, Yeji Shen, Peng Wang, Qin Huang, and C-C Jay Kuo. Spg-net: Segmentation prediction and guidance network for image inpainting. In Proceedings of the British Machine Vision Conference (BMVC), 2018. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 437, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 288, + 470 + ], + "type": "text", + "content": "[33] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "type": "text", + "content": "[34] Chuan Wang, Haibin Huang, Xiaoguang Han, and Jue Wang. Video inpainting by jointly learning temporal structure and spatial details. In Association for the Advancement of Artificial Intelligence (AAAI), 2019. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 288, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 558 + ], + "type": "text", + "content": "[35] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2018. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 602 + ], + "type": "text", + "content": "[36] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "type": "text", + "content": "[37] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep flow-guided video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 690 + ], + "type": "text", + "content": "[38] Brandon Yang, Gabriel Bender, Quoc V Le, and Jiquan Ngiam. Condconv: Conditionally parameterized convolutions for efficient inference. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[39] Yanhong Zeng, Jianlong Fu, and Hongyang Chao. Learning joint spatial-temporal transformations for video inpainting." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 274 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 547, + 139 + ], + "type": "text", + "content": "[40] Kaidong Zhang, Jingjing Fu, and Dong Liu. Flow-guided transformer for video inpainting. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 140, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 547, + 184 + ], + "type": "text", + "content": "[41] Kaidong Zhang, Jingjing Fu, and Dong Liu. Inertia-guided flow completion and style fusion for video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 185, + 547, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 547, + 229 + ], + "type": "text", + "content": "[42] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 230, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 547, + 274 + ], + "type": "text", + "content": "[43] Xueyan Zou, Linjie Yang, Ding Liu, and Yong Jae Lee. Progressive temporal feature alignment network for video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12958" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/840dd92d-8817-48f0-ac2c-586a3d27151d_content_list.json b/2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/840dd92d-8817-48f0-ac2c-586a3d27151d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d85b6708060cecb143a2fe62df3bbb85de9c6df4 --- /dev/null +++ b/2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/840dd92d-8817-48f0-ac2c-586a3d27151d_content_list.json @@ -0,0 +1,1773 @@ +[ + { + "type": "text", + "text": "Semantic-Aware Implicit Template Learning via Part Deformation Consistency", + "text_level": 1, + "bbox": [ + 84, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sihyeon Kim Minseok Joo Jaewon Lee \nJuyeon Ko Juhan Cha Hyunwoo J. Kim*", + "bbox": [ + 287, + 180, + 676, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Computer Science and Engineering, Korea University", + "bbox": [ + 210, + 227, + 756, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sh_bs15, wlgkcjf87, 2jlejyu, juyon98, hanchaa, hyunwoojkim}@korea.ac.kr", + "bbox": [ + 168, + 247, + 800, + 262 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 297, + 313, + 313 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning implicit templates as neural fields has recently shown impressive performance in unsupervised shape correspondence. Despite the success, we observe current approaches, which solely rely on geometric information, often learn suboptimal deformation across generic object shapes, which have high structural variability. In this paper, we highlight the importance of part deformation consistency and propose a semantic-aware implicit template learning framework to enable semantically plausible deformation. By leveraging semantic prior from a self-supervised feature extractor, we suggest local conditioning with novel semantic-aware deformation code and deformation consistency regularizations regarding part deformation, global deformation, and global scaling. Our extensive experiments demonstrate the superiority of the proposed method over baselines in various tasks: keypoint transfer, part label transfer, and texture transfer. More interestingly, our framework shows a larger performance gain under more challenging settings. We also provide qualitative analyses to validate the effectiveness of semantic-aware deformation. The code is available at https://github.com/mlvlab/PDC.", + "bbox": [ + 73, + 328, + 473, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 656, + 209, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Template learning is essential for shape analysis since the template, which is a compact representation of arbitrary shapes, provides a good prior for shape fitting so that we can establish dense correspondence and transfer key attributes such as texture or keypoints between shapes. Traditionally, handcrafted templates such as meshes, part elements with primitive shapes (e.g., sphere, cylinder, cone), or parametric models have been widely used for human/animal body [3, 4, 5, 6, 7, 8, 9], human face [10, 11], human hand [12, 13, 14], and generic objects [15, 16, 17]. Defining templates for less complex shapes with consistent topologies and similar structures is relatively straightforward. However, for shapes like generic object shapes with high structural variability or", + "bbox": [ + 75, + 681, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b8c7719d3d700afb1d874be71a8a7714a1fb5c826244ecbfeb71ba0467281d7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 297, + 893, + 512 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/068252e044fa8ddddc802a3c0d2b98ea36759befa0f5a17a5bcffdd3449bdcb3.jpg", + "image_caption": [ + "Figure 1. Importance of part deformation consistency. Given input shapes with high structural variability, only considering low-level geometry during template learning [1, 2] can lead to suboptimal deformation (e.g., cannot distinguish between seat and arm where two parts are geometrically close while semantically distant: low seat-low arm, high seat-high arm). Ours succeeds in transferring shape attributes by encouraging part deformation consistency with semantic information." + ], + "image_footnote": [], + "bbox": [ + 503, + 513, + 893, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "non-rigid shapes with large-scale deformations, manually defining suitable templates becomes challenging.", + "bbox": [ + 496, + 750, + 890, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this challenge, recent works [1, 2, 18, 19] have focused on learning implicit templates based on neural fields [20]. Neural fields, which are known for strong representation power, efficiency, and flexibility, have been successful in multiple 3D vision tasks, e.g., shape/scene reconstruction [21, 22, 23, 24], neural rendering [25, 26, 27], and human digitization [28, 29, 30]. Likewise, neural fields-based template learning have shown superior perfor", + "bbox": [ + 496, + 779, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "593", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mance on dense correspondence between complex shapes. Both [1] and [2] suggest decomposing the implicit representation of shapes into a deformation field and a template field. Here, the template is learned as a continuous function, which is expected to capture shared structures among the given shapes within a category (e.g., airplane, chair), while the deformation depends solely on the conditioned geometry of each shape (i.e., SDF value and surface normal). We observe current approaches often learn suboptimal deformation especially when a shape category has high structural variability like diverse local deformation scales and unseen/missing part structures; see Figure 1. That is, incorporating extra knowledge of part semantics is necessary to handle such cases.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel framework to learn semantically plausible deformation by distilling the knowledge of self-supervised feature extractor [31] and employing it as a semantic prior to satisfy part deformation consistency during the template learning process. Part deformation consistency is our inductive bias that the deformation result within the same part should be consistent. By leveraging semantic prior to understand part arrangements of shapes, we suggest local conditioning with semantic-aware deformation code and carefully designed regularizations to encourage semantically corresponding deformation. Semantic-aware deformation code is a point-wise soft assignment of part deformation priors, where part deformation priors are multiple latent codes to encode information on corresponding part deformation. We calculate assigned weights considering the part semantics of each point. In addition, we propose input space regularization and latent space regularization to encourage part-level distances to be close, providing an explicit way to control the deformation field so that the model can learn flexible deformation and successfully learn common structures under diverse shapes. Lastly, we suggest global scale consistency regularization to preserve the volume of the template against large-scale deformations. To validate the effectiveness of our framework, we conduct extensive surrogate experiments reflecting different levels of correspondence such as keypoint transfer, part label transfer, and texture transfer, and demonstrate competitive performance over baselines with qualitative analyses.", + "bbox": [ + 75, + 306, + 472, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows: ① We propose a new framework that learns global implicit template fields founded on part deformation consistency while understanding part semantics to enable semantically plausible deformation, ② We impose hybrid conditioning by combining global conditioning of geometry and local conditioning of part semantics for flexible deformation, ③ We design novel regularizations to successfully manipulate the template fields, ④ We show the effectiveness of our framework with both qualitative and quantitative experiments.", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 89, + 650, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Template learning for shape correspondence. Defining a template for a set of arbitrary shapes has been studied for a long time. A high-quality template provides a good initialization for shape alignment and various applications in vision and graphics. For shapes with consistent topologies and structures such as faces or bodies of humans and animals, handcrafted templates have been popular [3, 5, 7, 8, 9] and still widely used today. On the other hand, it was more challenging for generic object shapes with inconsistent topology and diverse part structures. Prior works suggest to predefine or learn local templates from data as in primitive parts [15] or part elements [32, 16, 33, 17, 34] to handle the inconsistent structure. Recent works have tried to learn a canonical shape of generic objects through implicit template field [1, 2, 35, 36], which are proven to be effective for representing shapes. However, existing approaches mostly lack explicit consideration of semantic knowledge, which enables semantically plausible deformation so that the learned template captures a common structure even with the various structured shapes.", + "bbox": [ + 496, + 130, + 893, + 434 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural fields for 3D representations. Recent studies have shown that representing 3D shapes or scenes as implicit representations with a neural network has several advantages: powerful representation, flexibility, and memory efficiency. These are called neural fields [20]. For instance, Occupancy networks [22] proposed a method to represent 3D geometry as a continuous occupancy function. DeepSDF [23] learns 3D complex surfaces with a neural network of auto-decoder framework for 3D modeling. Further, implicit representation has been introduced to various applications [37, 38]. To leverage the advantages, many applications based on neural fields have emerged such as neural rendering [25, 26, 27], human digitization [28, 29, 30], and generative modeling [39, 40, 41, 42]. They have also been applied to learn the template of complex shapes as mentioned above.", + "bbox": [ + 496, + 446, + 893, + 676 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Utilizing self-supervised segmentation models. Recently, there has been a growing body of research on analyzing complex shapes in a self-supervised manner due to the high cost of annotation for 3D data. Unsupervised semantic segmentation/co-segmentation models like BAE-Net [31], RIM-Net [43], AdaCoSeg [44], and PartSLIP [45] have demonstrated the ability to comprehend part semantics without ground truth labels. For 2D and multi-modal tasks, using large-scale self-supervised segmentation models like DINO [46] as a feature extractor has been widely adopted, including pixelNeRF [47], Feature Field [48], and ClipN-erF [49]. Inspired by these recent works, we utilize pretrained unsupervised part segmentation models to harness part semantic information for template learning.", + "bbox": [ + 496, + 688, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "594", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2c8c62fcf1b166e69f5abf8d598351ffd928a7e37ec0a0d5fc4bebd571ae0dd2.jpg", + "image_caption": [ + "Figure 2. Framework of semantic-aware implicit template learning. To learn semantic-aware implicit template field, (a) we extract part semantics from pretrained feature extractor $E$ and incorporate knowledge into semantic-aware deformation code $(\\alpha)$ and (b) part deformation consistency regularizations. With optimizing with the reconstruction loss $\\mathcal{L}_{\\mathrm{rec}}$ and several regularizers, our framework succeeds to learn semantically plausible deformation." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 279, + 242 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ee7a919c08e58d123d46598646c37fef17ae02e62c65416039063fcee5aef637.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 282, + 88, + 555, + 242 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f523e7f3c92f502460306e545568713d7bd79a9809d81f4cd9455b4a4c8abcb5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 85, + 684, + 241 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/801ade73fef454d3add5f9bef7712229db9601be40b7f48707a00ebd1f041aa5.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 696, + 85, + 888, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 337, + 215, + 353 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Implicit templates have been studied with neural fields. Given a 3D shape $P$ , a neural field, which is a neural network $\\mathcal{F}_{\\theta}$ parameterized by $\\theta$ , approximates continuous functions such as an occupancy function [21, 22] or a Signed Distance Function (SDF) [23] as", + "bbox": [ + 75, + 362, + 468, + 438 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\theta} (\\boldsymbol {x}; \\boldsymbol {z} _ {P}) = s. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 450, + 468, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Herein, $\\pmb{x} \\in \\mathbb{R}^3$ is a query point and $s$ is an occupancy/SDF function value. To represent multiple 3D shapes, for each shape, a shape latent code $\\pmb{z}_P \\in \\mathbb{R}^d$ is learned. In general, a surface is defined as the zero-level set of a SDF, i.e., $\\{x | \\mathcal{F}_{\\theta}(x; z_P) = 0\\}$ as in DeepSDF [23].", + "bbox": [ + 76, + 476, + 468, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The neural field can be further decomposed into an implicit template field $\\mathcal{T}$ and a deformation field $\\mathcal{D}[1,2]$ :", + "bbox": [ + 75, + 551, + 468, + 582 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\theta} (\\boldsymbol {x}; \\boldsymbol {z} _ {P}) = \\mathcal {T} _ {\\theta_ {2}} \\left(\\mathcal {D} _ {\\theta_ {1}} (\\boldsymbol {x}; \\boldsymbol {z} _ {P})\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 593, + 468, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The implicit template field $\\mathcal{T}_{\\theta_2}$ captures a common structure and outputs an occupancy/SDF value $s$ in the template space. $\\mathcal{D}_{\\theta_1}$ maps a point $x$ from a shape-specific space to the common template space by adding a deformation vector $\\Delta \\pmb{x} \\in \\mathbb{R}^3$ , i.e., $\\pmb{x} + \\Delta \\pmb{x}$ . In DIF-Net [2], $\\mathcal{D}_{\\theta_1}$ additionally predicts a correction value $\\Delta s \\in \\mathbb{R}$ to capture instance-specific local structures. In this case, the neural field $\\mathcal{F}_{\\theta}$ is defined as", + "bbox": [ + 75, + 619, + 468, + 739 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\theta} (\\boldsymbol {x}; \\boldsymbol {z} _ {P}) = \\mathcal {T} _ {\\theta_ {2}} (\\boldsymbol {x} + \\Delta \\boldsymbol {x}) + \\Delta s \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 750, + 468, + 773 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta s, \\Delta \\boldsymbol {x} = \\mathcal {D} _ {\\theta_ {1}} (\\boldsymbol {x}; \\boldsymbol {z} _ {P}).\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 768, + 336, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The existing approaches [1, 2] learn the implicit template field by minimizing the distance between the template and the deformed individual shapes. The distance is measured by SDF values and surface normals. Also, several regularizers are imposed such as deformation smoothness (e.g., penalizing the spatial gradient of $\\mathcal{D}_{\\theta_1}$ and restricting the deformation scale between points). However, the loss functions", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and regularizers are computed by low-level geometric information. They are not effective enough to handle generic objects with large structural variability. Hence, in this paper, we take part semantics into account for template learning to achieve 'semantically plausible' deformation and dense correspondence.", + "bbox": [ + 496, + 338, + 890, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 441, + 589, + 457 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present a semantic-aware implicit template learning framework based on part deformation consistency. In this section, we first delineate the overall architecture of our method equipped with part deformation priors and semantic-aware deformation code. Then, we introduce our deformation consistency regularizers including part deformation consistency regularizations. Lastly, we discuss the training procedure with several additional loss functions.", + "bbox": [ + 496, + 468, + 890, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Semantic-aware implicit template field", + "text_level": 1, + "bbox": [ + 498, + 597, + 831, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We here present a novel semantic-aware implicit template learning framework. The overall pipeline of the proposed method is shown in Figure 2(a).", + "bbox": [ + 496, + 621, + 890, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Template field and deformation field. We extend the implicit template learning by incorporating semantic information into the deformation field. Specifically, given a query point $\\pmb{x}$ of a shape $P$ , the deformation field $\\mathcal{D}_{\\theta_1}$ predicts a deformation vector $\\Delta x$ and a correction value $\\Delta s$ conditioned on a point-wise semantic-aware deformation code (SDC) $\\alpha(x; P)$ in addition to a shape latent code $z_P$ . Then, our framework is defined as", + "bbox": [ + 496, + 666, + 890, + 785 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\theta} (\\boldsymbol {x}; P, \\mathcal {E}) := \\mathcal {T} _ {\\theta_ {2}} (\\boldsymbol {x} + \\Delta \\boldsymbol {x}) + \\Delta s, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 794, + 890, + 818 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\boldsymbol {x}, \\Delta s = \\mathcal {D} _ {\\theta_ {1}} (\\boldsymbol {x}; \\boldsymbol {z} _ {P}, \\boldsymbol {\\alpha} (\\boldsymbol {x}; P)),\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 814, + 818, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{E}$ is a pretrained feature extractor. We assign $\\alpha (\\pmb {x};P)$ based on a point-wise semantic feature, which is extracted from $\\mathcal{E}$ . For simplicity, we denote the neural field $\\mathcal{F}_{\\theta}(\\pmb {x};P,\\mathcal{E})$ by $\\mathcal{F}_{\\theta}(\\pmb {x})$ and SDC $\\alpha (\\pmb {x};P)$ by $\\alpha (\\pmb {x})$ .", + "bbox": [ + 500, + 839, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "595", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature extractor. In this paper, for feature extraction, we mainly use BAE-Net [31] as an encoder $\\mathcal{E}$ , which is proven effective in capturing part semantics. Note that we pretrained BAE-Net on selected categories of ShapeNetV2 [50] by self-supervision without any extra labels or data for a fair comparison. The features from BAE-Net are $k$ dimensional vectors, i.e., $\\boldsymbol{o}^P(\\boldsymbol{x}) = \\mathcal{E}(\\boldsymbol{x}, P) \\in \\mathbb{R}^k$ . $k$ is a hyperparameter, which is closely related to the number of semantic parts. Since the number of semantic parts is unknown in self-supervised learning, we simply chose sufficiently large $k$ .", + "bbox": [ + 75, + 90, + 472, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic-aware deformation code. Our framework assumes that different semantic parts necessitate different deformations. To achieve this, the proposed method learns a deformation prior for each semantic part. The deformation priors are encoded as latent codes $e_1, \\dots, e_k \\in \\mathbb{R}^{d'}$ . A point $x$ is softly assigned to $k$ semantic parts using the softmax values of semantic features $o^P(x)$ . Combining the soft assignment with deformation priors, semantic-aware deformation code (SDC) denoted as $\\alpha(x) \\in \\mathbb{R}^{d'}$ is defined as", + "bbox": [ + 75, + 257, + 472, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\alpha} (\\boldsymbol {x}) = \\sum_ {i = 1} ^ {k} \\frac {\\exp \\left(\\boldsymbol {o} _ {i} ^ {P} (\\boldsymbol {x})\\right) \\boldsymbol {e} _ {i}}{\\sum_ {j = 1} ^ {k} \\exp \\left(\\boldsymbol {o} _ {j} ^ {P} (\\boldsymbol {x})\\right)}, \\forall i \\in \\{1, \\dots , k \\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 416, + 468, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{o}_i^P(\\pmb{x})$ denotes the $i$ -th dimension value of the semantic features. The SDC at $\\pmb{x}$ is computed by a linear combination of deformation priors. We learn the point's probability of belonging to each of the $k$ parts, which is served as part semantics.", + "bbox": [ + 75, + 468, + 468, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remarks. For comparison, we also applied hard assignment of part deformation priors for SDC, i.e., $\\alpha(\\pmb{x}) = e_{i^*}$ , where $i^* = \\operatorname{argmax}_i (\\pmb{o}_i^P(\\pmb{x}))$ . However, as points are completely separated with local conditioning of the corresponding part deformation prior $e_{i^*}$ , the final deformation result tends to break the shape geometry (see Figure 3), leading to performance degradation as shown in Table 1. Thereby, we choose the soft assignment to enable flexible deformation while preserving the shape geometry.", + "bbox": [ + 75, + 545, + 468, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We train the deformation field $\\mathcal{D}_{\\theta_1}$ and $\\mathcal{T}_{\\theta_2}$ in an end-to-end fashion. In our experiments, $\\mathcal{D}_{\\theta_1}$ is a 6-layer MLP conditioned on both global geometry $z_P$ via hypernetwork $\\Psi$ [51, 52] and local part semantics $\\alpha(x)$ via concatenation with input coordinates. The shape latent code $z_P$ and part deformation priors $\\{\\pmb{e}_i\\}$ are randomly initialized and jointly optimized with neural fields. $\\mathcal{T}_{\\theta_2}$ is a 5-layer MLP.", + "bbox": [ + 75, + 681, + 468, + 787 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/ed6d6b55ebcc8e221a658f50958afd732547feaaea2845337de00adced3979fd.jpg", + "table_caption": [ + "Table 1. Ablations of different assignment strategies for SDC. We conduct a quantitative comparison between hard and soft assignments with 5-shot part label transfer on ShapeNet chair." + ], + "table_footnote": [], + "table_body": "
Assignment strategiesHardSoft
chair (5-shot PLT)77.084.2
", + "bbox": [ + 143, + 840, + 403, + 883 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3ffcd3cae66c1d892adb41f0722cb7d9eaec6987c5d6eee85e1dd68309f33035.jpg", + "image_caption": [ + "Figure 3. Effects of different assignment strategies for SDC. Hard assignment breaks the shape geometry, while soft assignment preserves the geometry while learning flexible deformations." + ], + "image_footnote": [], + "bbox": [ + 550, + 102, + 846, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Part deformation consistency regularizations", + "text_level": 1, + "bbox": [ + 498, + 257, + 879, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to conditioning neural fields, leveraging semantic information, we propose novel part deformation consistency regularizations to encourage semantically plausible deformation. Semantically similar points should be close after deformation in both an input space and a latent space. That is, the deformation of points from the same part should be consistent. Figure 2(b) illustrates the concept of our regularizations. Given a pair of deformed shapes $P', Q'$ , part deformation consistency regularizations are defined as", + "bbox": [ + 496, + 280, + 892, + 417 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {p d c - g e o}} = \\sum_ {i = 1} ^ {k} \\left(\\frac {1}{| \\mathcal {P} _ {i} ^ {\\prime} |} \\sum_ {\\boldsymbol {x} \\in \\mathcal {P} _ {i} ^ {\\prime}} \\| \\boldsymbol {x} - C (\\boldsymbol {x}, \\mathcal {Q} _ {i} ^ {\\prime}) \\| _ {2} ^ {2} \\right. \\\\ \\left. + \\frac {1}{\\left| \\mathcal {Q} _ {i} ^ {\\prime} \\right|} \\sum_ {\\boldsymbol {y} \\in \\mathcal {Q} _ {i} ^ {\\prime}} \\| \\boldsymbol {y} - C (\\boldsymbol {y}, \\mathcal {P} _ {i} ^ {\\prime}) \\| _ {2} ^ {2}\\right), \\text {a n d} \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 440, + 890, + 540 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {p d c} \\text {- s e m}} = \\sum_ {i = 1} ^ {k} \\left(\\frac {1}{| \\mathcal {P} _ {i} ^ {\\prime} |} \\sum_ {\\boldsymbol {x} \\in \\mathcal {P} _ {i} ^ {\\prime}} \\| \\boldsymbol {o} ^ {P ^ {\\prime}} (\\boldsymbol {x}) - \\boldsymbol {o} ^ {Q ^ {\\prime}} \\left(C (\\boldsymbol {x}, \\mathcal {Q} _ {i} ^ {\\prime})\\right) \\| _ {2} ^ {2} \\right. \\\\ \\left. + \\frac {1}{| \\mathcal {Q} _ {i} ^ {\\prime} |} \\sum_ {\\boldsymbol {y} \\in \\mathcal {Q} _ {i} ^ {\\prime}} \\| \\boldsymbol {o} ^ {Q ^ {\\prime}} (\\boldsymbol {y}) - \\boldsymbol {o} ^ {P ^ {\\prime}} \\left(C (\\boldsymbol {y}, \\mathcal {P} _ {i} ^ {\\prime})\\right) \\| _ {2} ^ {2}\\right), \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 566, + 890, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $C$ is a correspondence matching function for a point $\\pmb{x}$ and its nearest point in a set $\\mathcal{Q}$ , i.e., $C(\\pmb{x},\\mathcal{Q}) = \\operatorname*{argmin}_{\\pmb{y}\\in \\mathcal{Q}}\\| \\pmb{x} - \\pmb{y}\\|_2^2$ . $\\mathcal{P}_i'$ and $\\mathcal{Q}_i'$ denote sampled point sets of", + "bbox": [ + 496, + 691, + 893, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$i^{\\mathrm{th}}$ deformed part shapes of $P^{\\prime}$ and $Q^{\\prime}$ . They are defined as $\\mathcal{P}_i^\\prime = \\{\\pmb {x} + \\mathcal{D}_{\\theta_1}(\\pmb {x};\\pmb {z}_P,\\pmb {\\alpha}(\\pmb {x};P))\\mid i = \\operatorname *{argmax}_{k}(\\pmb {o}_k^P (\\pmb {x})),\\pmb {x}\\in$", + "bbox": [ + 496, + 748, + 890, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\Omega_P\\}$ and $\\mathcal{Q}_i^\\prime = \\{\\pmb {x} + \\mathcal{D}_{\\theta_1}(\\pmb {x};\\pmb {z}_Q,\\pmb {\\alpha}(\\pmb {x};Q))|i =$ argmax $(\\pmb{o}_k^Q (\\pmb {x})),\\pmb {x}\\in \\Omega_Q\\}$ , where $\\Omega_P$ and $\\Omega_{Q}$ are sam", + "bbox": [ + 496, + 785, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pled query point sets for shape $P$ and $Q$ . Note that each deformed part shape is extracted based on part semantics from original shapes $P$ and $Q$ . Here, we explicitly guide the part-level deformation among various shapes to be consistent. Moreover, since we understand part arrangement and", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "596", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c11deb1e2f34ad9c566244693f1e65d187963d0795697e23acc5908163a4b361.jpg", + "image_caption": [ + "Figure 4. Keypoint transfer result on ShapeNet. We transfer keypoint labels from source shapes (colored in ivory) to target shapes and compare ours with DIT [1] and DIF [2]. For additional visualizations including results on ScanObjectNN [53], see the supplement." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 888, + 234 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cafeb9d58e02ba8740aa2249583b33b699f5429902f293ffbc09e7ad539f23d0.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 96, + 314, + 171, + 359 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4b36f80cd31eee68e6a5b448f1efa22671b904462eb9211f55d754c17fb51573.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 192, + 316, + 302, + 353 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f5876dceecdf1c4231267053c9f0d994cbe0c8858bedf5f786589aa659e3d8c1.jpg", + "image_caption": [ + "(c)", + "Figure 5. Efficacy of global scaling consistency $\\mathcal{L}_{\\mathrm{scale}}$" + ], + "image_footnote": [], + "bbox": [ + 321, + 306, + 455, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "only penalize matched parts, our method does not forcibly deform \"unseen/missing match\" cases.", + "bbox": [ + 75, + 402, + 467, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We partly observed spatial distortion when part deformation consistency was too strongly imposed. This is because using external semantics for deformation could break the continuity of geometric fields. In addition, we propose global scale consistency and global deformation consistency to address the concern. For the former one, we disentangle a scaling operation from the entire non-rigid deformation $\\mathcal{D}_{\\theta_1}$ to calculate the global scale of the learned implicit template field. That is, we want to approximate $\\mathcal{D}$ with the global scaling of the scaling factor of $r$ , i.e., $\\mathcal{D} \\approx rI$ . Then the problem becomes finding $r$ from", + "bbox": [ + 75, + 433, + 468, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {r} \\sum_ {s} \\sum_ {i} \\| \\boldsymbol {x} _ {i} ^ {s} + \\Delta \\boldsymbol {x} _ {i} ^ {s} - r \\boldsymbol {x} _ {i} ^ {s} \\| _ {2} ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 606, + 468, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s$ and $i$ are an index for shape and point, respectively. A closed-form solution of (8) is $r = \\sum_{s}\\sum_{i}\\frac{\\pmb{x}_i^{s\\top}(\\pmb{x}_i^s + \\Delta\\pmb{x}_i^s)}{\\pmb{x}_i^{s\\top}\\pmb{x}_i^s}$ . See the supplement for proof. Since we want the learned $\\mathcal{D}$ to be insensitive to global scale operation, i.e., the implicit template to keep its scale by the average scale of input shapes, we penalize the expected $\\mathbb{E}[r]$ over input shapes $P_{1}\\dots P_{N}$ in a single batch as:", + "bbox": [ + 75, + 643, + 468, + 762 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s c a l e}} = | \\mathbb {E} [ r ] - 1 | \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 770, + 468, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 5 illuminates the effect of $\\mathcal{L}_{\\mathrm{scale}}$ . For the same input shapes, Figure 5(a) is a car template learned without $\\mathcal{L}_{\\mathrm{scale}}$ , Figure 5(b) is a car template learned with $\\mathcal{L}_{\\mathrm{scale}}$ , and Figure 5(c) is the result of placing two shapes in the same space. It shows $\\mathcal{L}_{\\mathrm{scale}}$ helps the model to learn a more desirable template by penalizing the change in the global scale of the learned template fields.", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the latter one, we add global deformation consistency for a pair of deformed shapes $P', Q'$ :", + "bbox": [ + 498, + 305, + 890, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {g e o}} = \\mathbf {C D} \\left(P ^ {\\prime}, Q ^ {\\prime}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 340, + 890, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{CD}$ denotes a conventional Chamfer loss. Thus, the deformation consistency regularization can be written as:", + "bbox": [ + 498, + 364, + 890, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d c}} = \\gamma_ {1} \\mathcal {L} _ {\\mathrm {p d c - g e o}} + \\gamma_ {2} \\mathcal {L} _ {\\mathrm {p d c - s e m}} + \\gamma_ {3} \\mathcal {L} _ {\\text {s c a l e}} + \\gamma_ {4} \\mathcal {L} _ {\\text {g e o}}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 400, + 890, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Optimization", + "text_level": 1, + "bbox": [ + 500, + 445, + 638, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train our deformation and template fields with a reconstruction loss and several regularizers including our deformation consistency regularization. Following [2], the reconstruction loss $\\mathcal{L}_{\\mathrm{rec}}$ for shape $P$ is", + "bbox": [ + 496, + 468, + 890, + 529 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {r e c}} = \\sum_ {\\boldsymbol {x} \\in \\Omega_ {P}} | \\mathcal {F} _ {\\theta} (\\boldsymbol {x}) - \\bar {s} | + \\sum_ {\\boldsymbol {x} \\in P} (1 - \\langle \\nabla \\mathcal {F} _ {\\theta} (\\boldsymbol {x}), \\bar {n} \\rangle) \\\\ + \\sum_ {\\boldsymbol {x} \\in \\Omega_ {P}} | \\| \\nabla \\mathcal {F} _ {\\theta} (\\boldsymbol {x}) \\| _ {2} - 1 | + \\sum_ {\\boldsymbol {x} \\in \\Omega_ {P} \\backslash P} \\exp (- \\delta \\cdot | \\mathcal {F} _ {\\theta} (\\boldsymbol {x}) |), \\tag {12} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 534, + 890, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\bar{s}$ denotes the ground-truth SDF value of each query point, $\\bar{n}$ is the ground truth surface normal vector of each surface point and $\\nabla \\mathcal{F}_{\\theta}$ the spatial derivative of a 3D field. The second term of (12) is derived from the property of the SDF function that the gradient of an SDF is identical to the surface normal for surface points and the third is implicit geometric regularization [54] derived from the Eikonal equation. The last term pertains to penalizing off-surface points to be away from the surface points, which is controlled by $\\delta \\gg 1$ .", + "bbox": [ + 496, + 608, + 890, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since we do not know ground truth mapping between the template and input shapes, we manipulate the deformation field through multiple regularizers. First, we employ our proposed deformation consistency regularization (11). For further regularizers, we adopt [2] such as deformation smoothness $\\mathcal{L}_{\\mathrm{smooth}}$ , normal consistency $\\mathcal{L}_{\\mathrm{normal}}$ and minimal correction $\\mathcal{L}_{\\mathrm{c}}$ , which are defined as", + "bbox": [ + 496, + 758, + 890, + 864 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s m o o t h}} = \\sum_ {\\boldsymbol {x} \\in \\Omega_ {P}} \\sum_ {d \\in \\{X, Y, Z \\}} \\| \\nabla \\mathcal {D} _ {\\theta_ {1}} ^ {v} (\\cdot , \\boldsymbol {z} _ {P}) | _ {d} (\\boldsymbol {x}) \\| _ {2} ^ {2}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 869, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "597", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/da82899a2c9ce4b1dc7fc4cccec1d8fceb2404da3d183bb1d68551ee615a29cb.jpg", + "image_caption": [ + "Figure 6. Part label transfer result on ShapeNet. We transfer part segment labels from source shapes (colored in ivory) to target shapes and compare ours with DIT [1] and DIF [2]. For additional visualizations, please refer to the supplement." + ], + "image_footnote": [], + "bbox": [ + 91, + 92, + 883, + 232 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {n o r m a l}} = \\sum_ {i} \\sum_ {\\boldsymbol {x} \\in \\mathcal {S} _ {i}} \\left(1 - \\left\\langle \\nabla \\mathcal {T} _ {\\theta_ {2}} \\left(\\boldsymbol {x} + D _ {\\theta_ {1}} ^ {v} (\\boldsymbol {x})\\right), \\bar {n} \\right\\rangle\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 310, + 468, + 357 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c}} = \\sum_ {i} \\sum_ {\\boldsymbol {x} \\in \\Omega_ {P}} | \\mathcal {D} _ {\\theta_ {1}} ^ {\\Delta s} (\\boldsymbol {x}) |. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 368, + 468, + 401 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also apply $\\ell_2$ regularization on shape latent codes and part deformation priors", + "bbox": [ + 76, + 407, + 468, + 438 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {e m b}} = \\sum_ {i} ^ {N} \\| z _ {i} \\| _ {2} ^ {2} + \\sum_ {j} ^ {k} \\| e _ {j} \\| _ {2} ^ {2}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 448, + 468, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The regularizers are collectively written as", + "bbox": [ + 76, + 501, + 359, + 517 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e g}} = \\gamma_ {5} \\mathcal {L} _ {\\text {s m o o t h}} + \\gamma_ {6} \\mathcal {L} _ {\\text {n o r m a l}} + \\gamma_ {7} \\mathcal {L} _ {\\mathrm {c}} + \\gamma_ {8} \\mathcal {L} _ {\\text {e m b}}. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 527, + 468, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hence, the final loss becomes", + "bbox": [ + 76, + 556, + 274, + 570 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {\\mathrm {d c}} + \\mathcal {L} _ {\\text {r e g}}. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 583, + 468, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluating dense correspondence. Given two pairs of seen or unseen shapes $P, Q$ , we can evaluate the dense correspondence between $P$ and $Q$ by deforming each shape to the common template space. Moreover, the uncertainty of dense correspondence performance between two points $p \\in P$ and $q \\in Q$ can be measured as below:", + "bbox": [ + 75, + 611, + 468, + 702 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nu (p, q) = 1 - \\exp (- \\gamma \\| (p + \\Delta p) - (q + \\Delta q) \\| _ {2} ^ {2}). \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 712, + 468, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\gamma$ is a hyperparameter for modulation and $\\Delta p, \\Delta q$ are output from $\\mathcal{D}_{\\theta_1}$ . Note that since each part deformation prior $e$ is shared based on part semantics across entire training shapes, we do not need a test-time optimization on $e$ for an unseen shape at inference time.", + "bbox": [ + 75, + 741, + 468, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 829, + 209, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of our proposed framework and provide both quantitative and qualitative results. In Section 5.1, we conduct multiple label transfer", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/14e884a9b2c350111152cd9b325302420efe9ad71d8c1301a84b84ea6fe4b7b3.jpg", + "table_caption": [ + "Table 2. Keypoint correspondence accuracy in KeypointNet [55]. PCK scores with threshold of $0.05 / 0.1$ are reported." + ], + "table_footnote": [], + "table_body": "
PCKairplanecarchairtable
AtlasNetV2 [34]23.9 / 45.333.5 / 64.025.6 / 50.225.3 / 48.0
DIT [2]13.6 / 34.258.9 / 87.733.3 / 66.726.2 / 55.5
DIF [1]16.5 / 40.567.8 / 90.434.3 / 57.637.9 / 62.9
Ours24.2 / 52.369.3 / 91.861.2 / 83.250.8/77.9
", + "bbox": [ + 501, + 311, + 893, + 388 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2d58bb777779ba8ee363b11e57efc839f42847be5442e9977970070eaba1d71c.jpg", + "table_caption": [ + "Table 3. 5-shot part label transfer in four categories of ShapeNet. The reported numbers are overall mIoU." + ], + "table_footnote": [], + "table_body": "
modelairplanecarchairtable
AtlasNetV2 [34]57.859.366.970.6
DIT [1]64.266.180.486.0
DIF [2]73.969.682.186.6
Ours75.170.484.286.9
", + "bbox": [ + 517, + 428, + 875, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tasks for dense correspondence evaluation. In Section 5.2, we measure shape reconstruction accuracy to evaluate the representational power for unknown shapes. In Section 5.3, we present advanced dense correspondence analysis under more challenging settings. Lastly, in Section 5.4, we perform ablation studies to analyze the proposed regularization and the sensitivity over feature extractor [31].", + "bbox": [ + 496, + 549, + 890, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Baselines. We train and validate our model and baselines on four categories (e.g., chair, table, airplane and car) of ShapeNetV2 [50]. We follow the training split of DIF [2] for each category. For label transfer tasks, we use the labels from ShapeNet-Part [56] and KeypointNet [55]. DIT [1] and DIF [2] are the most relevant baseline models as they both globally learn implicit template fields, and we select AtlasNetV2 [34] as an additional baseline. Dense correspondence is measured with local patch composition in [34]. For more implementation details, analyses, and visualizations, see the supplement.", + "bbox": [ + 496, + 657, + 890, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Dense correspondence", + "text_level": 1, + "bbox": [ + 500, + 832, + 707, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the quality of learned dense correspondence, we conduct various surrogate tasks including keypoint transfer, part label transfer, and texture transfer due to", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "598", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d992a9a2429a7e8c487f80fb2c12c05698fa643a016b4a70d35d2eca80ee5f3e.jpg", + "image_caption": [ + "Figure 7. Texture transfer result on ShapeNet. We transfer part segment labels from source shapes (colored in ivory) to target shapes and compare ours with DIT [1] and DIF [2]. For additional visualizations, please refer to the supplement." + ], + "image_footnote": [], + "bbox": [ + 89, + 92, + 877, + 258 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/90e6e2b99689eff2b04221393f9d7235b1b3c75beda2f97bc90fbbe1218e35c2.jpg", + "table_caption": [ + "Table 4. Reconstruction accuracy for unseen shapes in four categories of ShapeNet. The reported numbers are chamfer distance (CD) between ground truth and reconstructed meshes at a resolution of $256^{3}$ ." + ], + "table_footnote": [], + "table_body": "
CD (×103)↓airplanecarchairtable
DIT [1]0.22610.40820.30630.6187
DIF [2]0.20930.31570.42290.2640
Ours0.18590.31500.22650.2796
", + "bbox": [ + 84, + 356, + 460, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the lack of labeled datasets for 3D dense correspondence. For the following tasks, we transfer respective labels from source shapes to target shapes in a common template space.", + "bbox": [ + 75, + 470, + 468, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Keypoint label transfer. Table 2 presents the quantitative result in a 5-shot keypoint transfer task. We measure the performance by a percentage of correct keypoints (PCK) [57]. We compute the distance between the transferred keypoints and the ground truth points to record the PCK score with the threshold of $0.05 / 0.1$ . Our method achieves the best performance in all four categories with a significant improvement. Especially for categories with high structural variability (e.g., chair and table), a large performance gain is observed. For instance, even compared to the best baseline, on the chair category with the threshold of 0.1, our method (83.2) outperforms DIT [1] (66.7) by $16.5\\%$ , and on the table category with the threshold of 0.1, our method (77.9) outperforms DIF [2] (62.9) by $15\\%$ . We additionally provide the visualization of keypoint transfer in Figure 4, which also support the importance of understanding part semantics for shape correspondence. In particular, two keypoints of pea-green and orange from the source chair are geometrically close but semantically different since one is for the arm part and the other is for the seat part. Among several methods, only our framework succeeds in disentangling two keypoints in the target chair. We provide more qualitative results regarding keypoint transfer results in real-world data [53] to validate the robustness of our method in the supplement.", + "bbox": [ + 75, + 523, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Part label transfer. For few-shot part label transfer tasks, we transfer part segment labels from source shapes to target shapes. In detail, for each (target) point of the deformed target shape, we find the $n$ nearest points of the deformed source shapes and perform weighted voting using their labels and distances to the target point. In our experiments, we used $n = 10$ nearest points. For a 5-shot part label transfer task in Table 3, five source shapes were used just the same as [2] for every category. Table 3 shows our framework achieves the best performance over all four categories. Moreover, Figure 6 illustrates the qualitative results of one-shot part label transfer on ShapeNet [50]. Similar to keypoint transfer, Figure 6 also proves the effectiveness of semantic knowledge in dense correspondence.", + "bbox": [ + 496, + 306, + 893, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Texture transfer. Figure 7 shows texture transfer results on ShapeNetV2 [50] to compare established dense correspondence between baselines and ours with higher-frequency label transfer. We observe notable texture transfer results. For instance, our framework successfully transfers the source car's texture of the windshield (gray) to the corresponding part of the target cars, whereas baselines transfer the same textures to the roof part of the target cars.", + "bbox": [ + 496, + 518, + 893, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Shape reconstruction", + "text_level": 1, + "bbox": [ + 500, + 650, + 700, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report the reconstruction quality of unseen shapes with Chamfer Distance (CD) to validate the representational power of our framework in Table 4. A smaller CD means more accurate reconstruction. As in previous work, we report the $\\mathrm{CD} \\times 10^{3}$ for readability. Also, we consistently discuss the CD and performance gain below at this scale. We optimize the corresponding shape latent code for a given unseen shape. We measure CD as in DIF [2], where reconstructed meshes are extracted at a resolution of $256^{3}$ , and CD is calculated with $10K$ sampled points. Our framework on average achieved the best reconstruction results surpassing DIF by a margin of 0.0512. On average, our method and DIF exhibit 0.2518 and 0.3030 reconstruction errors, respectively. In three out of four categories, our method achieved the best reconstruction accuracy.", + "bbox": [ + 496, + 674, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "599", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. More challenging setups", + "text_level": 1, + "bbox": [ + 76, + 90, + 302, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further validate the effectiveness of our method, we perform additional experiments and compare the results with DIF [2]. Overall, the results below imply current methods that lack part semantics and do not consider part deformation consistency may fail to learn semantically plausible deformation in challenging cases.", + "bbox": [ + 75, + 113, + 468, + 204 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "2-shot part label transfer within different subcategories. Our objective is to assess dense correspondence performance under difficult settings. We conduct 2-shot part label transfer within two different subcategories belonging to the same category, where shapes from different subcategories tend to have different geometry for same semantics. Since ShapeNet [50] includes incorrect subcategory labels [58], we manually cleaned subcategory labels for four categories (e.g., airplane: airliner/jet, car: sedan/jeep, chair: straight chair/sofa, table: pedestal table/short table). Figure 8 displays examples of shapes from each subcategory. For robust evaluations, we randomly select the source shape 5 times for each subcategory. Table 5 shows our framework achieves a significant performance gain in all cases by over $7.2\\%$ on average compared to DIF [2].", + "bbox": [ + 75, + 205, + 470, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unsupervised dense correspondence in DFAUST [59]. Furthermore, we evaluate unsupervised dense correspondence with non-rigid shapes [59]. For more challenging settings, we train both DIF [2] and our model for two different subjects with different sizes and heights (e.g., subject ID of 50002, 50026) and four distinct actions (e.g., shake arms, chicken wings, running on spot, and jumping jacks). Colormap in Figure 9 demonstrates the correspondence quality, where we transfer color over the up-axis. In the source shape, the arms are labeled in red. After transferring these color labels to the target shapes, our method preserves the red/orange labeling of the arms across different motions, whereas the baseline fails to capture the same labeling (i.e., yellow hands) since it lacks semantic knowledge during training. Additionally, we conduct keypoint transfer tasks and measure the PCK score as identical as in Section 5.1. Our model achieves 74.9/85.2 with a threshold of 0.05/0.1, whereas DIF [2] achieves 67.0/80.7 with the same threshold, showing our superior performance in the challenging settings (see the supplement for the full keypoint transfer result). Note that we can also learn implicit templates for non-rigid shapes with supervision as in [19]. We conduct experiments in unsupervised settings since our focus is to validate the effectiveness of utilizing additional semantic information for template learning.", + "bbox": [ + 75, + 431, + 470, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablations", + "text_level": 1, + "bbox": [ + 76, + 816, + 187, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on regularizations. We conduct ablation studies regarding the proposed regularizations including part deformation consistency regularizations, global scale consistency regularization, and global deformation consistency", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6cd10adf61d83a66ca80750a8786e2ac526e6ea83495786caa65bca3c3b03ef5.jpg", + "image_caption": [ + "Figure 8. Types of source shapes for 2-shot part label transfer." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 905, + 152 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dc1a2082addac86b30bcfc88ccc5cdc6b77036672b43248d1c28ef43db747ff9.jpg", + "table_caption": [ + "Table 5. 2-shot part label transfer in subcategories of ShapeNet. The reported numbers are overall mIoU over 5 runs." + ], + "table_footnote": [], + "table_body": "
subcategoriesDIF [2]Ours
airplane(airliner/jet)64.5±1.6172.8±0.99 (+8.3↑)
car(sedan/jeep)65.1±2.3469.1±1.46 (+4.0↑)
chair(straight/sofa)66.3±6.0677.5±5.81 (+11.2↑)
table(pedestal/short)74.4±10.179.7±6.26 (+5.3↑)
", + "bbox": [ + 509, + 210, + 885, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/87a877925235bd64d7f0802809ba8f9a944ff6a557b07ab22f5af790ef6063d1.jpg", + "image_caption": [ + "Figure 9. Color transfer on DFAUST [59]. We transfer different color labels over the up-axis from source shapes (colored in ivory) to target shapes and compare ours with DIF [2]." + ], + "image_footnote": [], + "bbox": [ + 506, + 316, + 893, + 436 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4a4aa9ec1574d04903a1e52c1ee7901ed039391446da6f32798f612695cd34be.jpg", + "table_caption": [ + "Table 6. Ablations for every proposed regularization. We report mIoU for 5-shot part label transfer on ShapeNet chair. Note that $\\mathcal{L}_{\\mathrm{pdc}}$ denotes $\\mathcal{L}_{\\mathrm{pdc\\_geo}} + \\mathcal{L}_{\\mathrm{pdc\\_sem}}$ ." + ], + "table_footnote": [], + "table_body": "
LpdcLgeoLscalechair
(a)✓83.4
(b)✓✓83.9
(c)✓✓✓84.2
", + "bbox": [ + 573, + 537, + 820, + 616 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/16617a34d40e30f85270118a5840eaf7ff9adf29a7f500f6fcf9c99f2b8a6a47.jpg", + "table_caption": [ + "Table 7. BAE-Net $k$ ablation. We report mIoU for 5-shot part label transfer on ShapeNet chair over 3 runs." + ], + "table_footnote": [], + "table_body": "
k=2k=4k=6k=8k=10
chair83.8±0.5183.5±1.4882.8±1.2783.6±0.8583.7±1.33
", + "bbox": [ + 506, + 656, + 879, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "regularization. We conduct 5-shot part label transfer within a chair category, as same as Section 5.1. Table 6 shows that each component of our proposed regularizers contributes to performance improvement for dense correspondence. Note that by only using $\\mathcal{L}_{\\mathrm{pdc}}$ , i.e., using $\\mathcal{L}_{\\mathrm{pdc\\_geo}}$ and $\\mathcal{L}_{\\mathrm{pdc\\_sem}}$ , it still outperforms than baselines (80.4 [1] and 82.1 [2]). The best performance was achieved by incorporating all proposed regularizations.", + "bbox": [ + 496, + 703, + 890, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "BAE-Net [31] with different $k$ . We evaluate the effect of $k$ for BAE-Net [31] with mean/standard deviation over 3 runs for ShapeNet [50] chair in We train BAE-Net [31] with a range of $k = (2,4,6,8,10)$ , where $k = 4$ is used in the main paper, and report mIoU for 5-shot part label", + "bbox": [ + 496, + 824, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "600", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "transfer task. Table 7 shows the efficacy of semantic priors, where our framework achieves the average mIoU of 83.5, which shows higher performance than baselines. More importantly, we observed that the performance is fairly robust over $k$ .", + "bbox": [ + 75, + 90, + 472, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 179, + 196, + 194 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We have proposed a new framework that learns a semantic-aware implicit template field. Our framework extracts semantic knowledge from a self-supervised feature extractor given unlabelled point clouds. By leveraging the semantic information, we propose novel conditioning with semantic-aware deformation code and part deformation consistency regularizations to encourage semantically plausible deformation. Furthermore, we suggest global scaling consistency and global deformation consistency regularizations to enhance the learning process. Through quantitative and qualitative evaluation, we demonstrate the superiority of our proposed method and verified the effect of all proposed components in our framework.", + "bbox": [ + 75, + 203, + 472, + 400 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgments. This research was supported in part by the MSIT (Ministry of Science and ICT), Korea, under the ICT Creative Consilience program (IITP-2023-2020-0-01819) supervised by the IITP (Institute for Information & communications Technology Planning & Evaluation); the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF-2023R1A2C2005373); and Google Cloud Research Credits program with the award (GCLF-67WL-WU6D-MF14, V9P3-99XD-64CL-EH5H).", + "bbox": [ + 75, + 417, + 472, + 568 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 579, + 174, + 594 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Zerong Zheng, Tao Yu, Qionghai Dai, and Yebin Liu. Deep implicit templates for 3d shape representation. In CVPR, 2021. 1, 2, 3, 5, 6, 7, 8", + "[2] Yu Deng, Jiaolong Yang, and Xin Tong. Deformed implicit field: Modeling 3d shapes with learned dense correspondence. In CVPR, 2021. 1, 2, 3, 5, 6, 7, 8", + "[3] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: a skinned multiperson linear model. TOG, 2015. 1, 2", + "[4] Brett Allen, Brian Curless, and Zoran Popovic. Articulated body deformation from range scan data. TOG, 2002. 1", + "[5] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan C. Russell, and Mathieu Aubry. 3d-coded: 3d correspondences by deep deformation. In ECCV, 2018. 1, 2", + "[6] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In CVPR, 2018. 1", + "[7] Wang Zeng, Wanli Ouyang, Ping Luo, Wentao Liu, and Xiaogang Wang. 3d human mesh regression with dense correspondence. In CVPR, 2020. 1, 2" + ], + "bbox": [ + 86, + 604, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[8] Sing Chun Lee and Misha Kazhdan. Dense point-to-point correspondences between genus-zero shapes. In Computer Graphics Forum, 2019. 1, 2", + "[9] Silvia Zuffi, Angjoo Kanazawa, David W. Jacobs, and Michael J. Black. 3d menagerie: Modeling the 3d shape and pose of animals. In CVPR, 2017. 1, 2", + "[10] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In AVSS, 2009. 1", + "[11] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. TOG, 2017. 1", + "[12] Luca Ballan, Aparna Taneja, Jürgen Gall, Luc Van Gool, and Marc Pollefeys. Motion capture of hands in action using discriminative salient points. In ECCV, 2012. 1", + "[13] Srinath Sridhar, Antti Oulasvirta, and Christian Theobalt. Interactive markerless articulated hand motion tracking using rgb and depth data. In ICCV, 2013. 1", + "[14] Iason Oikonomidis, Manolis IA Lourakis, and Antonis A Argyros. Evolutionary quasi-random search for hand articulations tracking. In CVPR, 2014. 1", + "[15] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In CVPR, 2017. 1, 2", + "[16] Vignesh Ganapathi-Subramanian, Olga Diamanti, Soeren Pirk, Chengcheng Tang, Matthias Niessner, and Leonidas Guibas. Parsing geometry using structure-aware shape templates. In 3DV, 2018. 1, 2", + "[17] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan Russell, and Mathieu Aubry. AtlasNet: A Papier-Mâché Approach to Learning 3D Surface Generation. In CVPR, 2018. 1, 2", + "[18] Pablo Palafox, Nikolaos Sarafianos, Tony Tung, and Angela Dai. Spams: Structured implicit parametric models. In CVPR, 2022. 1", + "[19] Ramana Sundararaman, Gautam Pai, and Maks Ovsjanikov. Implicit field supervision for robust non-rigid shape matching. In ECCV, 2022. 1, 8", + "[20] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, 2022. 1, 2", + "[21] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. 2019. 1, 3", + "[22] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 1, 2, 3", + "[23] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 1, 2, 3" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "601", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In ECCV, 2020. 1", + "[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2", + "[26] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 1, 2", + "[27] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In CVPR, 2021. 1, 2", + "[28] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, 2019. 1, 2", + "[29] Tarun Yenamandra, Ayush Tewari, Florian Bernard, Hans-Peter Seidel, Mohamed Elgharib, Daniel Cremers, and Christian Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In CVPR, 2021. 1, 2", + "[30] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR, 2020. 1, 2", + "[31] Zhiqin Chen, Kangxue Yin, Matthew Fisher, Siddhartha Chaudhuri, and Hao Zhang. Bae-net: Branched autoencoder for shape co-segmentation. In ICCV, 2019. 2, 4, 6, 8", + "[32] Vladimir G Kim, Wilmot Li, Niloy J Mitra, Siddhartha Chaudhuri, Stephen DiVerdi, and Thomas Funkhouser. Learning part-based templates from large collections of 3d shapes. TOG, 2013. 2", + "[33] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T. Freeman, and Thomas A. Funkhouser. Learning shape templates with structured implicit functions. In ICCV, 2019. 2", + "[34] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir Kim, Bryan Russell, and Mathieu Aubry. Learning elementary structures for 3d shape generation and matching. In NeurIPS, 2019. 2, 6", + "[35] Jiahui Lei and Kostas Daniilidis. Cadex: Learning canonical deformation coordinate space for dynamic surface representation via neural homeomorphism. In CVPR, 2022. 2", + "[36] Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural template: Topology-aware reconstruction and disentangled generation of 3d meshes. In CVPR, 2022. 2", + "[37] Zekun Hao, Hadar Averbuch-Elor, Noah Snavely, and Serge Belongie. Dualsdf: Semantic shape manipulation using a two-level representation. In CVPR, 2020. 2", + "[38] Yueqi Duan, Haidong Zhu, He Wang, Li Yi, Ram Nevatia, and Leonidas J. Guibas. Curriculum deepsdf. In ECCV, 2020. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In ICLR, 2022. 2", + "[40] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2", + "[41] Michael Niemeyer and Andreas Geiger. Campari: Camera-aware decomposed generative neural radiance fields. In 3DV, 2021. 2", + "[42] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, et al. Gaudi: A neural architect for immersive 3d scene generation. In NeurIPS, 2022. 2", + "[43] Chengjie Niu, Manyi Li, Kai Xu, and Hao Zhang. Rimnet: Recursive implicit fields for unsupervised learning of hierarchical shape structures. In CVPR, 2022. 2", + "[44] Chenyang Zhu, Kai Xu, Siddhartha Chaudhuri, Li Yi, Leonidas J Guibas, and Hao Zhang. Adacoseg: Adaptive shape co-segmentation with group consistency loss. In CVPR, 2020. 2", + "[45] Minghua Liu, Yinhao Zhu, Hong Cai, Shizhong Han, Zhan Ling, Fatih Porikli, and Hao Su. Part slip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In CVPR, 2023. 2", + "[46] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2", + "[47] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 2", + "[48] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. In NeurIPS, 2022. 2", + "[49] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In CVPR, 2022. 2", + "[50] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv:1512.03012, 2015. 4, 6, 7, 8", + "[51] David Ha, Andrew M. Dai, and Quoc V. Le. Hypernetworks. In ICLR, 2017. 4", + "[52] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In NeurIPS, 2020. 4", + "[53] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 5, 7" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "602", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 5", + "[55] Yang You, Yujing Lou, Chengkun Li, Zhoujun Cheng, Liangwei Li, Lizhuang Ma, Cewu Lu, and Weiming Wang. Keypointnet: A large-scale 3d keypoint dataset aggregated from numerous human annotations. In CVPR, 2020. 6", + "[56] Li Yi, Vladimir G. Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas J. Guibas. A scalable active framework for region annotation in 3d shape collections. TOG, 2016. 6", + "[57] Li Yi, Hao Su, Xingwen Guo, and Leonidas Guibas. Syncspeccnn: Synchronized spectral cnn for 3d shape segmentation. In CVPR, 2017. 7", + "[58] Xinhai Liu, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Fine-grained 3d shape classification with hierarchical part-view attention. IEEE Transactions on Image Processing, 2021. 8", + "[59] Federica Bogo, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Dynamic FAUST: Registering human bodies in motion. In CVPR, 2017. 8" + ], + "bbox": [ + 78, + 90, + 468, + 407 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "603", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/840dd92d-8817-48f0-ac2c-586a3d27151d_model.json b/2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/840dd92d-8817-48f0-ac2c-586a3d27151d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e790fcedc2b7c3aa3c1b32f0e87fdb8150ef5f1d --- /dev/null +++ b/2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/840dd92d-8817-48f0-ac2c-586a3d27151d_model.json @@ -0,0 +1,2477 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.086, + 0.131, + 0.884, + 0.154 + ], + "angle": 0, + "content": "Semantic-Aware Implicit Template Learning via Part Deformation Consistency" + }, + { + "type": "text", + "bbox": [ + 0.288, + 0.181, + 0.677, + 0.218 + ], + "angle": 0, + "content": "Sihyeon Kim Minseok Joo Jaewon Lee \nJuyeon Ko Juhan Cha Hyunwoo J. Kim*" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.228, + 0.758, + 0.246 + ], + "angle": 0, + "content": "Department of Computer Science and Engineering, Korea University" + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.248, + 0.802, + 0.263 + ], + "angle": 0, + "content": "{sh_bs15, wlgkcjf87, 2jlejyu, juyon98, hanchaa, hyunwoojkim}@korea.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.298, + 0.314, + 0.314 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.329, + 0.474, + 0.649 + ], + "angle": 0, + "content": "Learning implicit templates as neural fields has recently shown impressive performance in unsupervised shape correspondence. Despite the success, we observe current approaches, which solely rely on geometric information, often learn suboptimal deformation across generic object shapes, which have high structural variability. In this paper, we highlight the importance of part deformation consistency and propose a semantic-aware implicit template learning framework to enable semantically plausible deformation. By leveraging semantic prior from a self-supervised feature extractor, we suggest local conditioning with novel semantic-aware deformation code and deformation consistency regularizations regarding part deformation, global deformation, and global scaling. Our extensive experiments demonstrate the superiority of the proposed method over baselines in various tasks: keypoint transfer, part label transfer, and texture transfer. More interestingly, our framework shows a larger performance gain under more challenging settings. We also provide qualitative analyses to validate the effectiveness of semantic-aware deformation. The code is available at https://github.com/mlvlab/PDC." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.657, + 0.21, + 0.672 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.471, + 0.879 + ], + "angle": 0, + "content": "Template learning is essential for shape analysis since the template, which is a compact representation of arbitrary shapes, provides a good prior for shape fitting so that we can establish dense correspondence and transfer key attributes such as texture or keypoints between shapes. Traditionally, handcrafted templates such as meshes, part elements with primitive shapes (e.g., sphere, cylinder, cone), or parametric models have been widely used for human/animal body [3, 4, 5, 6, 7, 8, 9], human face [10, 11], human hand [12, 13, 14], and generic objects [15, 16, 17]. Defining templates for less complex shapes with consistent topologies and similar structures is relatively straightforward. However, for shapes like generic object shapes with high structural variability or" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.299, + 0.895, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.515, + 0.895, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.633, + 0.895, + 0.744 + ], + "angle": 0, + "content": "Figure 1. Importance of part deformation consistency. Given input shapes with high structural variability, only considering low-level geometry during template learning [1, 2] can lead to suboptimal deformation (e.g., cannot distinguish between seat and arm where two parts are geometrically close while semantically distant: low seat-low arm, high seat-high arm). Ours succeeds in transferring shape attributes by encouraging part deformation consistency with semantic information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.779 + ], + "angle": 0, + "content": "non-rigid shapes with large-scale deformations, manually defining suitable templates becomes challenging." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.902 + ], + "angle": 0, + "content": "To address this challenge, recent works [1, 2, 18, 19] have focused on learning implicit templates based on neural fields [20]. Neural fields, which are known for strong representation power, efficiency, and flexibility, have been successful in multiple 3D vision tasks, e.g., shape/scene reconstruction [21, 22, 23, 24], neural rendering [25, 26, 27], and human digitization [28, 29, 30]. Likewise, neural fields-based template learning have shown superior perfor" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "593" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "mance on dense correspondence between complex shapes. Both [1] and [2] suggest decomposing the implicit representation of shapes into a deformation field and a template field. Here, the template is learned as a continuous function, which is expected to capture shared structures among the given shapes within a category (e.g., airplane, chair), while the deformation depends solely on the conditioned geometry of each shape (i.e., SDF value and surface normal). We observe current approaches often learn suboptimal deformation especially when a shape category has high structural variability like diverse local deformation scales and unseen/missing part structures; see Figure 1. That is, incorporating extra knowledge of part semantics is necessary to handle such cases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.473, + 0.746 + ], + "angle": 0, + "content": "In this paper, we propose a novel framework to learn semantically plausible deformation by distilling the knowledge of self-supervised feature extractor [31] and employing it as a semantic prior to satisfy part deformation consistency during the template learning process. Part deformation consistency is our inductive bias that the deformation result within the same part should be consistent. By leveraging semantic prior to understand part arrangements of shapes, we suggest local conditioning with semantic-aware deformation code and carefully designed regularizations to encourage semantically corresponding deformation. Semantic-aware deformation code is a point-wise soft assignment of part deformation priors, where part deformation priors are multiple latent codes to encode information on corresponding part deformation. We calculate assigned weights considering the part semantics of each point. In addition, we propose input space regularization and latent space regularization to encourage part-level distances to be close, providing an explicit way to control the deformation field so that the model can learn flexible deformation and successfully learn common structures under diverse shapes. Lastly, we suggest global scale consistency regularization to preserve the volume of the template against large-scale deformations. To validate the effectiveness of our framework, we conduct extensive surrogate experiments reflecting different levels of correspondence such as keypoint transfer, part label transfer, and texture transfer, and demonstrate competitive performance over baselines with qualitative analyses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Our contributions are summarized as follows: ① We propose a new framework that learns global implicit template fields founded on part deformation consistency while understanding part semantics to enable semantically plausible deformation, ② We impose hybrid conditioning by combining global conditioning of geometry and local conditioning of part semantics for flexible deformation, ③ We design novel regularizations to successfully manipulate the template fields, ④ We show the effectiveness of our framework with both qualitative and quantitative experiments." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.651, + 0.107 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.131, + 0.895, + 0.435 + ], + "angle": 0, + "content": "Template learning for shape correspondence. Defining a template for a set of arbitrary shapes has been studied for a long time. A high-quality template provides a good initialization for shape alignment and various applications in vision and graphics. For shapes with consistent topologies and structures such as faces or bodies of humans and animals, handcrafted templates have been popular [3, 5, 7, 8, 9] and still widely used today. On the other hand, it was more challenging for generic object shapes with inconsistent topology and diverse part structures. Prior works suggest to predefine or learn local templates from data as in primitive parts [15] or part elements [32, 16, 33, 17, 34] to handle the inconsistent structure. Recent works have tried to learn a canonical shape of generic objects through implicit template field [1, 2, 35, 36], which are proven to be effective for representing shapes. However, existing approaches mostly lack explicit consideration of semantic knowledge, which enables semantically plausible deformation so that the learned template captures a common structure even with the various structured shapes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.895, + 0.677 + ], + "angle": 0, + "content": "Neural fields for 3D representations. Recent studies have shown that representing 3D shapes or scenes as implicit representations with a neural network has several advantages: powerful representation, flexibility, and memory efficiency. These are called neural fields [20]. For instance, Occupancy networks [22] proposed a method to represent 3D geometry as a continuous occupancy function. DeepSDF [23] learns 3D complex surfaces with a neural network of auto-decoder framework for 3D modeling. Further, implicit representation has been introduced to various applications [37, 38]. To leverage the advantages, many applications based on neural fields have emerged such as neural rendering [25, 26, 27], human digitization [28, 29, 30], and generative modeling [39, 40, 41, 42]. They have also been applied to learn the template of complex shapes as mentioned above." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Utilizing self-supervised segmentation models. Recently, there has been a growing body of research on analyzing complex shapes in a self-supervised manner due to the high cost of annotation for 3D data. Unsupervised semantic segmentation/co-segmentation models like BAE-Net [31], RIM-Net [43], AdaCoSeg [44], and PartSLIP [45] have demonstrated the ability to comprehend part semantics without ground truth labels. For 2D and multi-modal tasks, using large-scale self-supervised segmentation models like DINO [46] as a feature extractor has been widely adopted, including pixelNeRF [47], Feature Field [48], and ClipN-erF [49]. Inspired by these recent works, we utilize pretrained unsupervised part segmentation models to harness part semantic information for template learning." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "594" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.28, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.089, + 0.556, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.087, + 0.686, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.087, + 0.889, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.249, + 0.39, + 0.262 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.784, + 0.249, + 0.803, + 0.262 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.271, + 0.894, + 0.327 + ], + "angle": 0, + "content": "Figure 2. Framework of semantic-aware implicit template learning. To learn semantic-aware implicit template field, (a) we extract part semantics from pretrained feature extractor \\(E\\) and incorporate knowledge into semantic-aware deformation code \\((\\alpha)\\) and (b) part deformation consistency regularizations. With optimizing with the reconstruction loss \\(\\mathcal{L}_{\\mathrm{rec}}\\) and several regularizers, our framework succeeds to learn semantically plausible deformation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.338, + 0.216, + 0.354 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.47, + 0.439 + ], + "angle": 0, + "content": "Implicit templates have been studied with neural fields. Given a 3D shape \\(P\\), a neural field, which is a neural network \\(\\mathcal{F}_{\\theta}\\) parameterized by \\(\\theta\\), approximates continuous functions such as an occupancy function [21, 22] or a Signed Distance Function (SDF) [23] as" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.451, + 0.469, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\theta} (\\boldsymbol {x}; \\boldsymbol {z} _ {P}) = s. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.477, + 0.469, + 0.553 + ], + "angle": 0, + "content": "Herein, \\( \\pmb{x} \\in \\mathbb{R}^3 \\) is a query point and \\( s \\) is an occupancy/SDF function value. To represent multiple 3D shapes, for each shape, a shape latent code \\( \\pmb{z}_P \\in \\mathbb{R}^d \\) is learned. In general, a surface is defined as the zero-level set of a SDF, i.e., \\( \\{x | \\mathcal{F}_{\\theta}(x; z_P) = 0\\} \\) as in DeepSDF [23]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.469, + 0.583 + ], + "angle": 0, + "content": "The neural field can be further decomposed into an implicit template field \\(\\mathcal{T}\\) and a deformation field \\(\\mathcal{D}[1,2]\\):" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.594, + 0.469, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\theta} (\\boldsymbol {x}; \\boldsymbol {z} _ {P}) = \\mathcal {T} _ {\\theta_ {2}} \\left(\\mathcal {D} _ {\\theta_ {1}} (\\boldsymbol {x}; \\boldsymbol {z} _ {P})\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.469, + 0.741 + ], + "angle": 0, + "content": "The implicit template field \\(\\mathcal{T}_{\\theta_2}\\) captures a common structure and outputs an occupancy/SDF value \\(s\\) in the template space. \\(\\mathcal{D}_{\\theta_1}\\) maps a point \\(x\\) from a shape-specific space to the common template space by adding a deformation vector \\(\\Delta \\pmb{x} \\in \\mathbb{R}^3\\), i.e., \\(\\pmb{x} + \\Delta \\pmb{x}\\). In DIF-Net [2], \\(\\mathcal{D}_{\\theta_1}\\) additionally predicts a correction value \\(\\Delta s \\in \\mathbb{R}\\) to capture instance-specific local structures. In this case, the neural field \\(\\mathcal{F}_{\\theta}\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.75, + 0.469, + 0.774 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\theta} (\\boldsymbol {x}; \\boldsymbol {z} _ {P}) = \\mathcal {T} _ {\\theta_ {2}} (\\boldsymbol {x} + \\Delta \\boldsymbol {x}) + \\Delta s \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.77, + 0.337, + 0.785 + ], + "angle": 0, + "content": "\\[\n\\Delta s, \\Delta \\boldsymbol {x} = \\mathcal {D} _ {\\theta_ {1}} (\\boldsymbol {x}; \\boldsymbol {z} _ {P}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The existing approaches [1, 2] learn the implicit template field by minimizing the distance between the template and the deformed individual shapes. The distance is measured by SDF values and surface normals. Also, several regularizers are imposed such as deformation smoothness (e.g., penalizing the spatial gradient of \\(\\mathcal{D}_{\\theta_1}\\) and restricting the deformation scale between points). However, the loss functions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.43 + ], + "angle": 0, + "content": "and regularizers are computed by low-level geometric information. They are not effective enough to handle generic objects with large structural variability. Hence, in this paper, we take part semantics into account for template learning to achieve 'semantically plausible' deformation and dense correspondence." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.443, + 0.591, + 0.458 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.469, + 0.892, + 0.59 + ], + "angle": 0, + "content": "We present a semantic-aware implicit template learning framework based on part deformation consistency. In this section, we first delineate the overall architecture of our method equipped with part deformation priors and semantic-aware deformation code. Then, we introduce our deformation consistency regularizers including part deformation consistency regularizations. Lastly, we discuss the training procedure with several additional loss functions." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.598, + 0.833, + 0.615 + ], + "angle": 0, + "content": "4.1. Semantic-aware implicit template field" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.891, + 0.667 + ], + "angle": 0, + "content": "We here present a novel semantic-aware implicit template learning framework. The overall pipeline of the proposed method is shown in Figure 2(a)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.667, + 0.892, + 0.786 + ], + "angle": 0, + "content": "Template field and deformation field. We extend the implicit template learning by incorporating semantic information into the deformation field. Specifically, given a query point \\( \\pmb{x} \\) of a shape \\( P \\), the deformation field \\( \\mathcal{D}_{\\theta_1} \\) predicts a deformation vector \\( \\Delta x \\) and a correction value \\( \\Delta s \\) conditioned on a point-wise semantic-aware deformation code (SDC) \\( \\alpha(x; P) \\) in addition to a shape latent code \\( z_P \\). Then, our framework is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.795, + 0.892, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\theta} (\\boldsymbol {x}; P, \\mathcal {E}) := \\mathcal {T} _ {\\theta_ {2}} (\\boldsymbol {x} + \\Delta \\boldsymbol {x}) + \\Delta s, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.815, + 0.82, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\Delta \\boldsymbol {x}, \\Delta s = \\mathcal {D} _ {\\theta_ {1}} (\\boldsymbol {x}; \\boldsymbol {z} _ {P}, \\boldsymbol {\\alpha} (\\boldsymbol {x}; P)),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}\\) is a pretrained feature extractor. We assign \\(\\alpha (\\pmb {x};P)\\) based on a point-wise semantic feature, which is extracted from \\(\\mathcal{E}\\). For simplicity, we denote the neural field \\(\\mathcal{F}_{\\theta}(\\pmb {x};P,\\mathcal{E})\\) by \\(\\mathcal{F}_{\\theta}(\\pmb {x})\\) and SDC \\(\\alpha (\\pmb {x};P)\\) by \\(\\alpha (\\pmb {x})\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "595" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.257 + ], + "angle": 0, + "content": "Feature extractor. In this paper, for feature extraction, we mainly use BAE-Net [31] as an encoder \\(\\mathcal{E}\\), which is proven effective in capturing part semantics. Note that we pretrained BAE-Net on selected categories of ShapeNetV2 [50] by self-supervision without any extra labels or data for a fair comparison. The features from BAE-Net are \\(k\\) dimensional vectors, i.e., \\(\\boldsymbol{o}^P(\\boldsymbol{x}) = \\mathcal{E}(\\boldsymbol{x}, P) \\in \\mathbb{R}^k\\). \\(k\\) is a hyperparameter, which is closely related to the number of semantic parts. Since the number of semantic parts is unknown in self-supervised learning, we simply chose sufficiently large \\(k\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.473, + 0.395 + ], + "angle": 0, + "content": "Semantic-aware deformation code. Our framework assumes that different semantic parts necessitate different deformations. To achieve this, the proposed method learns a deformation prior for each semantic part. The deformation priors are encoded as latent codes \\( e_1, \\dots, e_k \\in \\mathbb{R}^{d'} \\). A point \\( x \\) is softly assigned to \\( k \\) semantic parts using the softmax values of semantic features \\( o^P(x) \\). Combining the soft assignment with deformation priors, semantic-aware deformation code (SDC) denoted as \\( \\alpha(x) \\in \\mathbb{R}^{d'} \\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.417, + 0.47, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\alpha} (\\boldsymbol {x}) = \\sum_ {i = 1} ^ {k} \\frac {\\exp \\left(\\boldsymbol {o} _ {i} ^ {P} (\\boldsymbol {x})\\right) \\boldsymbol {e} _ {i}}{\\sum_ {j = 1} ^ {k} \\exp \\left(\\boldsymbol {o} _ {j} ^ {P} (\\boldsymbol {x})\\right)}, \\forall i \\in \\{1, \\dots , k \\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.546 + ], + "angle": 0, + "content": "where \\( \\pmb{o}_i^P(\\pmb{x}) \\) denotes the \\( i \\)-th dimension value of the semantic features. The SDC at \\( \\pmb{x} \\) is computed by a linear combination of deformation priors. We learn the point's probability of belonging to each of the \\( k \\) parts, which is served as part semantics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.47, + 0.682 + ], + "angle": 0, + "content": "Remarks. For comparison, we also applied hard assignment of part deformation priors for SDC, i.e., \\(\\alpha(\\pmb{x}) = e_{i^*}\\), where \\(i^* = \\operatorname{argmax}_i (\\pmb{o}_i^P(\\pmb{x}))\\). However, as points are completely separated with local conditioning of the corresponding part deformation prior \\(e_{i^*}\\), the final deformation result tends to break the shape geometry (see Figure 3), leading to performance degradation as shown in Table 1. Thereby, we choose the soft assignment to enable flexible deformation while preserving the shape geometry." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.47, + 0.789 + ], + "angle": 0, + "content": "We train the deformation field \\(\\mathcal{D}_{\\theta_1}\\) and \\(\\mathcal{T}_{\\theta_2}\\) in an end-to-end fashion. In our experiments, \\(\\mathcal{D}_{\\theta_1}\\) is a 6-layer MLP conditioned on both global geometry \\(z_P\\) via hypernetwork \\(\\Psi\\) [51, 52] and local part semantics \\(\\alpha(x)\\) via concatenation with input coordinates. The shape latent code \\(z_P\\) and part deformation priors \\(\\{\\pmb{e}_i\\}\\) are randomly initialized and jointly optimized with neural fields. \\(\\mathcal{T}_{\\theta_2}\\) is a 5-layer MLP." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.799, + 0.469, + 0.841 + ], + "angle": 0, + "content": "Table 1. Ablations of different assignment strategies for SDC. We conduct a quantitative comparison between hard and soft assignments with 5-shot part label transfer on ShapeNet chair." + }, + { + "type": "table", + "bbox": [ + 0.145, + 0.841, + 0.404, + 0.884 + ], + "angle": 0, + "content": "
Assignment strategiesHardSoft
chair (5-shot PLT)77.084.2
" + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.103, + 0.847, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.202, + 0.894, + 0.245 + ], + "angle": 0, + "content": "Figure 3. Effects of different assignment strategies for SDC. Hard assignment breaks the shape geometry, while soft assignment preserves the geometry while learning flexible deformations." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.258, + 0.88, + 0.274 + ], + "angle": 0, + "content": "4.2. Part deformation consistency regularizations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.281, + 0.893, + 0.418 + ], + "angle": 0, + "content": "In addition to conditioning neural fields, leveraging semantic information, we propose novel part deformation consistency regularizations to encourage semantically plausible deformation. Semantically similar points should be close after deformation in both an input space and a latent space. That is, the deformation of points from the same part should be consistent. Figure 2(b) illustrates the concept of our regularizations. Given a pair of deformed shapes \\( P', Q' \\), part deformation consistency regularizations are defined as" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.441, + 0.892, + 0.541 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {p d c - g e o}} = \\sum_ {i = 1} ^ {k} \\left(\\frac {1}{| \\mathcal {P} _ {i} ^ {\\prime} |} \\sum_ {\\boldsymbol {x} \\in \\mathcal {P} _ {i} ^ {\\prime}} \\| \\boldsymbol {x} - C (\\boldsymbol {x}, \\mathcal {Q} _ {i} ^ {\\prime}) \\| _ {2} ^ {2} \\right. \\\\ \\left. + \\frac {1}{\\left| \\mathcal {Q} _ {i} ^ {\\prime} \\right|} \\sum_ {\\boldsymbol {y} \\in \\mathcal {Q} _ {i} ^ {\\prime}} \\| \\boldsymbol {y} - C (\\boldsymbol {y}, \\mathcal {P} _ {i} ^ {\\prime}) \\| _ {2} ^ {2}\\right), \\text {a n d} \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.567, + 0.892, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {p d c} \\text {- s e m}} = \\sum_ {i = 1} ^ {k} \\left(\\frac {1}{| \\mathcal {P} _ {i} ^ {\\prime} |} \\sum_ {\\boldsymbol {x} \\in \\mathcal {P} _ {i} ^ {\\prime}} \\| \\boldsymbol {o} ^ {P ^ {\\prime}} (\\boldsymbol {x}) - \\boldsymbol {o} ^ {Q ^ {\\prime}} \\left(C (\\boldsymbol {x}, \\mathcal {Q} _ {i} ^ {\\prime})\\right) \\| _ {2} ^ {2} \\right. \\\\ \\left. + \\frac {1}{| \\mathcal {Q} _ {i} ^ {\\prime} |} \\sum_ {\\boldsymbol {y} \\in \\mathcal {Q} _ {i} ^ {\\prime}} \\| \\boldsymbol {o} ^ {Q ^ {\\prime}} (\\boldsymbol {y}) - \\boldsymbol {o} ^ {P ^ {\\prime}} \\left(C (\\boldsymbol {y}, \\mathcal {P} _ {i} ^ {\\prime})\\right) \\| _ {2} ^ {2}\\right), \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.693, + 0.894, + 0.749 + ], + "angle": 0, + "content": "where \\(C\\) is a correspondence matching function for a point \\(\\pmb{x}\\) and its nearest point in a set \\(\\mathcal{Q}\\), i.e., \\(C(\\pmb{x},\\mathcal{Q}) = \\operatorname*{argmin}_{\\pmb{y}\\in \\mathcal{Q}}\\| \\pmb{x} - \\pmb{y}\\|_2^2\\). \\(\\mathcal{P}_i'\\) and \\(\\mathcal{Q}_i'\\) denote sampled point sets of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.786 + ], + "angle": 0, + "content": "\\(i^{\\mathrm{th}}\\) deformed part shapes of \\(P^{\\prime}\\) and \\(Q^{\\prime}\\). They are defined as \\(\\mathcal{P}_i^\\prime = \\{\\pmb {x} + \\mathcal{D}_{\\theta_1}(\\pmb {x};\\pmb {z}_P,\\pmb {\\alpha}(\\pmb {x};P))\\mid i = \\operatorname *{argmax}_{k}(\\pmb {o}_k^P (\\pmb {x})),\\pmb {x}\\in\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.786, + 0.892, + 0.825 + ], + "angle": 0, + "content": "\\(\\Omega_P\\}\\) and \\(\\mathcal{Q}_i^\\prime = \\{\\pmb {x} + \\mathcal{D}_{\\theta_1}(\\pmb {x};\\pmb {z}_Q,\\pmb {\\alpha}(\\pmb {x};Q))|i =\\) argmax \\((\\pmb{o}_k^Q (\\pmb {x})),\\pmb {x}\\in \\Omega_Q\\}\\) , where \\(\\Omega_P\\) and \\(\\Omega_{Q}\\) are sam" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "pled query point sets for shape \\(P\\) and \\(Q\\). Note that each deformed part shape is extracted based on part semantics from original shapes \\(P\\) and \\(Q\\). Here, we explicitly guide the part-level deformation among various shapes to be consistent. Moreover, since we understand part arrangement and" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "596" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.889, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.244, + 0.893, + 0.274 + ], + "angle": 0, + "content": "Figure 4. Keypoint transfer result on ShapeNet. We transfer keypoint labels from source shapes (colored in ivory) to target shapes and compare ours with DIT [1] and DIF [2]. For additional visualizations including results on ScanObjectNN [53], see the supplement." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.315, + 0.172, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.364, + 0.144, + 0.376 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.193, + 0.318, + 0.303, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.363, + 0.258, + 0.375 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.322, + 0.308, + 0.456, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.384, + 0.363, + 0.401, + 0.376 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.381, + 0.438, + 0.395 + ], + "angle": 0, + "content": "Figure 5. Efficacy of global scaling consistency \\(\\mathcal{L}_{\\mathrm{scale}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.468, + 0.433 + ], + "angle": 0, + "content": "only penalize matched parts, our method does not forcibly deform \"unseen/missing match\" cases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.434, + 0.469, + 0.601 + ], + "angle": 0, + "content": "We partly observed spatial distortion when part deformation consistency was too strongly imposed. This is because using external semantics for deformation could break the continuity of geometric fields. In addition, we propose global scale consistency and global deformation consistency to address the concern. For the former one, we disentangle a scaling operation from the entire non-rigid deformation \\(\\mathcal{D}_{\\theta_1}\\) to calculate the global scale of the learned implicit template field. That is, we want to approximate \\(\\mathcal{D}\\) with the global scaling of the scaling factor of \\(r\\), i.e., \\(\\mathcal{D} \\approx rI\\). Then the problem becomes finding \\(r\\) from" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.607, + 0.469, + 0.638 + ], + "angle": 0, + "content": "\\[\n\\min _ {r} \\sum_ {s} \\sum_ {i} \\| \\boldsymbol {x} _ {i} ^ {s} + \\Delta \\boldsymbol {x} _ {i} ^ {s} - r \\boldsymbol {x} _ {i} ^ {s} \\| _ {2} ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.469, + 0.763 + ], + "angle": 0, + "content": "where \\(s\\) and \\(i\\) are an index for shape and point, respectively. A closed-form solution of (8) is \\(r = \\sum_{s}\\sum_{i}\\frac{\\pmb{x}_i^{s\\top}(\\pmb{x}_i^s + \\Delta\\pmb{x}_i^s)}{\\pmb{x}_i^{s\\top}\\pmb{x}_i^s}\\). See the supplement for proof. Since we want the learned \\(\\mathcal{D}\\) to be insensitive to global scale operation, i.e., the implicit template to keep its scale by the average scale of input shapes, we penalize the expected \\(\\mathbb{E}[r]\\) over input shapes \\(P_{1}\\dots P_{N}\\) in a single batch as:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.771, + 0.469, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s c a l e}} = | \\mathbb {E} [ r ] - 1 | \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Figure 5 illuminates the effect of \\(\\mathcal{L}_{\\mathrm{scale}}\\). For the same input shapes, Figure 5(a) is a car template learned without \\(\\mathcal{L}_{\\mathrm{scale}}\\), Figure 5(b) is a car template learned with \\(\\mathcal{L}_{\\mathrm{scale}}\\), and Figure 5(c) is the result of placing two shapes in the same space. It shows \\(\\mathcal{L}_{\\mathrm{scale}}\\) helps the model to learn a more desirable template by penalizing the change in the global scale of the learned template fields." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.306, + 0.892, + 0.336 + ], + "angle": 0, + "content": "For the latter one, we add global deformation consistency for a pair of deformed shapes \\(P', Q'\\):" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.342, + 0.892, + 0.359 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {g e o}} = \\mathbf {C D} \\left(P ^ {\\prime}, Q ^ {\\prime}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.365, + 0.892, + 0.395 + ], + "angle": 0, + "content": "where \\(\\mathbf{CD}\\) denotes a conventional Chamfer loss. Thus, the deformation consistency regularization can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.401, + 0.892, + 0.419 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d c}} = \\gamma_ {1} \\mathcal {L} _ {\\mathrm {p d c - g e o}} + \\gamma_ {2} \\mathcal {L} _ {\\mathrm {p d c - s e m}} + \\gamma_ {3} \\mathcal {L} _ {\\text {s c a l e}} + \\gamma_ {4} \\mathcal {L} _ {\\text {g e o}}. \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.446, + 0.64, + 0.463 + ], + "angle": 0, + "content": "4.3. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.469, + 0.892, + 0.53 + ], + "angle": 0, + "content": "We train our deformation and template fields with a reconstruction loss and several regularizers including our deformation consistency regularization. Following [2], the reconstruction loss \\(\\mathcal{L}_{\\mathrm{rec}}\\) for shape \\(P\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.535, + 0.891, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {r e c}} = \\sum_ {\\boldsymbol {x} \\in \\Omega_ {P}} | \\mathcal {F} _ {\\theta} (\\boldsymbol {x}) - \\bar {s} | + \\sum_ {\\boldsymbol {x} \\in P} (1 - \\langle \\nabla \\mathcal {F} _ {\\theta} (\\boldsymbol {x}), \\bar {n} \\rangle) \\\\ + \\sum_ {\\boldsymbol {x} \\in \\Omega_ {P}} | \\| \\nabla \\mathcal {F} _ {\\theta} (\\boldsymbol {x}) \\| _ {2} - 1 | + \\sum_ {\\boldsymbol {x} \\in \\Omega_ {P} \\backslash P} \\exp (- \\delta \\cdot | \\mathcal {F} _ {\\theta} (\\boldsymbol {x}) |), \\tag {12} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.892, + 0.758 + ], + "angle": 0, + "content": "where \\(\\bar{s}\\) denotes the ground-truth SDF value of each query point, \\(\\bar{n}\\) is the ground truth surface normal vector of each surface point and \\(\\nabla \\mathcal{F}_{\\theta}\\) the spatial derivative of a 3D field. The second term of (12) is derived from the property of the SDF function that the gradient of an SDF is identical to the surface normal for surface points and the third is implicit geometric regularization [54] derived from the Eikonal equation. The last term pertains to penalizing off-surface points to be away from the surface points, which is controlled by \\(\\delta \\gg 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.759, + 0.892, + 0.865 + ], + "angle": 0, + "content": "Since we do not know ground truth mapping between the template and input shapes, we manipulate the deformation field through multiple regularizers. First, we employ our proposed deformation consistency regularization (11). For further regularizers, we adopt [2] such as deformation smoothness \\(\\mathcal{L}_{\\mathrm{smooth}}\\), normal consistency \\(\\mathcal{L}_{\\mathrm{normal}}\\) and minimal correction \\(\\mathcal{L}_{\\mathrm{c}}\\), which are defined as" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.87, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s m o o t h}} = \\sum_ {\\boldsymbol {x} \\in \\Omega_ {P}} \\sum_ {d \\in \\{X, Y, Z \\}} \\| \\nabla \\mathcal {D} _ {\\theta_ {1}} ^ {v} (\\cdot , \\boldsymbol {z} _ {P}) | _ {d} (\\boldsymbol {x}) \\| _ {2} ^ {2}, \\tag {13}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "597" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.093, + 0.885, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.241, + 0.895, + 0.27 + ], + "angle": 0, + "content": "Figure 6. Part label transfer result on ShapeNet. We transfer part segment labels from source shapes (colored in ivory) to target shapes and compare ours with DIT [1] and DIF [2]. For additional visualizations, please refer to the supplement." + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.311, + 0.47, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {n o r m a l}} = \\sum_ {i} \\sum_ {\\boldsymbol {x} \\in \\mathcal {S} _ {i}} \\left(1 - \\left\\langle \\nabla \\mathcal {T} _ {\\theta_ {2}} \\left(\\boldsymbol {x} + D _ {\\theta_ {1}} ^ {v} (\\boldsymbol {x})\\right), \\bar {n} \\right\\rangle\\right), \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.369, + 0.47, + 0.402 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c}} = \\sum_ {i} \\sum_ {\\boldsymbol {x} \\in \\Omega_ {P}} | \\mathcal {D} _ {\\theta_ {1}} ^ {\\Delta s} (\\boldsymbol {x}) |. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.408, + 0.47, + 0.439 + ], + "angle": 0, + "content": "We also apply \\(\\ell_2\\) regularization on shape latent codes and part deformation priors" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.449, + 0.47, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {e m b}} = \\sum_ {i} ^ {N} \\| z _ {i} \\| _ {2} ^ {2} + \\sum_ {j} ^ {k} \\| e _ {j} \\| _ {2} ^ {2}. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.502, + 0.361, + 0.518 + ], + "angle": 0, + "content": "The regularizers are collectively written as" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.529, + 0.47, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e g}} = \\gamma_ {5} \\mathcal {L} _ {\\text {s m o o t h}} + \\gamma_ {6} \\mathcal {L} _ {\\text {n o r m a l}} + \\gamma_ {7} \\mathcal {L} _ {\\mathrm {c}} + \\gamma_ {8} \\mathcal {L} _ {\\text {e m b}}. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.557, + 0.276, + 0.571 + ], + "angle": 0, + "content": "Hence, the final loss becomes" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.584, + 0.47, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {\\mathrm {d c}} + \\mathcal {L} _ {\\text {r e g}}. \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.703 + ], + "angle": 0, + "content": "Evaluating dense correspondence. Given two pairs of seen or unseen shapes \\( P, Q \\), we can evaluate the dense correspondence between \\( P \\) and \\( Q \\) by deforming each shape to the common template space. Moreover, the uncertainty of dense correspondence performance between two points \\( p \\in P \\) and \\( q \\in Q \\) can be measured as below:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.713, + 0.47, + 0.731 + ], + "angle": 0, + "content": "\\[\nu (p, q) = 1 - \\exp (- \\gamma \\| (p + \\Delta p) - (q + \\Delta q) \\| _ {2} ^ {2}). \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.469, + 0.817 + ], + "angle": 0, + "content": "\\(\\gamma\\) is a hyperparameter for modulation and \\(\\Delta p, \\Delta q\\) are output from \\(\\mathcal{D}_{\\theta_1}\\). Note that since each part deformation prior \\(e\\) is shared based on part semantics across entire training shapes, we do not need a test-time optimization on \\(e\\) for an unseen shape at inference time." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.83, + 0.21, + 0.848 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We demonstrate the effectiveness of our proposed framework and provide both quantitative and qualitative results. In Section 5.1, we conduct multiple label transfer" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.284, + 0.892, + 0.311 + ], + "angle": 0, + "content": "Table 2. Keypoint correspondence accuracy in KeypointNet [55]. PCK scores with threshold of \\(0.05 / 0.1\\) are reported." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.312, + 0.895, + 0.39 + ], + "angle": 0, + "content": "
PCKairplanecarchairtable
AtlasNetV2 [34]23.9 / 45.333.5 / 64.025.6 / 50.225.3 / 48.0
DIT [2]13.6 / 34.258.9 / 87.733.3 / 66.726.2 / 55.5
DIF [1]16.5 / 40.567.8 / 90.434.3 / 57.637.9 / 62.9
Ours24.2 / 52.369.3 / 91.861.2 / 83.250.8/77.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.4, + 0.894, + 0.428 + ], + "angle": 0, + "content": "Table 3. 5-shot part label transfer in four categories of ShapeNet. The reported numbers are overall mIoU." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.429, + 0.876, + 0.526 + ], + "angle": 0, + "content": "
modelairplanecarchairtable
AtlasNetV2 [34]57.859.366.970.6
DIT [1]64.266.180.486.0
DIF [2]73.969.682.186.6
Ours75.170.484.286.9
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.656 + ], + "angle": 0, + "content": "tasks for dense correspondence evaluation. In Section 5.2, we measure shape reconstruction accuracy to evaluate the representational power for unknown shapes. In Section 5.3, we present advanced dense correspondence analysis under more challenging settings. Lastly, in Section 5.4, we perform ablation studies to analyze the proposed regularization and the sensitivity over feature extractor [31]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Datasets and Baselines. We train and validate our model and baselines on four categories (e.g., chair, table, airplane and car) of ShapeNetV2 [50]. We follow the training split of DIF [2] for each category. For label transfer tasks, we use the labels from ShapeNet-Part [56] and KeypointNet [55]. DIT [1] and DIF [2] are the most relevant baseline models as they both globally learn implicit template fields, and we select AtlasNetV2 [34] as an additional baseline. Dense correspondence is measured with local patch composition in [34]. For more implementation details, analyses, and visualizations, see the supplement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.708, + 0.849 + ], + "angle": 0, + "content": "5.1. Dense correspondence" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To evaluate the quality of learned dense correspondence, we conduct various surrogate tasks including keypoint transfer, part label transfer, and texture transfer due to" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "598" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.093, + 0.878, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.267, + 0.895, + 0.297 + ], + "angle": 0, + "content": "Figure 7. Texture transfer result on ShapeNet. We transfer part segment labels from source shapes (colored in ivory) to target shapes and compare ours with DIT [1] and DIF [2]. For additional visualizations, please refer to the supplement." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.304, + 0.471, + 0.357 + ], + "angle": 0, + "content": "Table 4. Reconstruction accuracy for unseen shapes in four categories of ShapeNet. The reported numbers are chamfer distance (CD) between ground truth and reconstructed meshes at a resolution of \\(256^{3}\\)." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.357, + 0.462, + 0.44 + ], + "angle": 0, + "content": "
CD (×103)↓airplanecarchairtable
DIT [1]0.22610.40820.30630.6187
DIF [2]0.20930.31570.42290.2640
Ours0.18590.31500.22650.2796
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.471, + 0.47, + 0.518 + ], + "angle": 0, + "content": "the lack of labeled datasets for 3D dense correspondence. For the following tasks, we transfer respective labels from source shapes to target shapes in a common template space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Keypoint label transfer. Table 2 presents the quantitative result in a 5-shot keypoint transfer task. We measure the performance by a percentage of correct keypoints (PCK) [57]. We compute the distance between the transferred keypoints and the ground truth points to record the PCK score with the threshold of \\(0.05 / 0.1\\). Our method achieves the best performance in all four categories with a significant improvement. Especially for categories with high structural variability (e.g., chair and table), a large performance gain is observed. For instance, even compared to the best baseline, on the chair category with the threshold of 0.1, our method (83.2) outperforms DIT [1] (66.7) by \\(16.5\\%\\), and on the table category with the threshold of 0.1, our method (77.9) outperforms DIF [2] (62.9) by \\(15\\%\\). We additionally provide the visualization of keypoint transfer in Figure 4, which also support the importance of understanding part semantics for shape correspondence. In particular, two keypoints of pea-green and orange from the source chair are geometrically close but semantically different since one is for the arm part and the other is for the seat part. Among several methods, only our framework succeeds in disentangling two keypoints in the target chair. We provide more qualitative results regarding keypoint transfer results in real-world data [53] to validate the robustness of our method in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.895, + 0.519 + ], + "angle": 0, + "content": "Part label transfer. For few-shot part label transfer tasks, we transfer part segment labels from source shapes to target shapes. In detail, for each (target) point of the deformed target shape, we find the \\( n \\) nearest points of the deformed source shapes and perform weighted voting using their labels and distances to the target point. In our experiments, we used \\( n = 10 \\) nearest points. For a 5-shot part label transfer task in Table 3, five source shapes were used just the same as [2] for every category. Table 3 shows our framework achieves the best performance over all four categories. Moreover, Figure 6 illustrates the qualitative results of one-shot part label transfer on ShapeNet [50]. Similar to keypoint transfer, Figure 6 also proves the effectiveness of semantic knowledge in dense correspondence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.895, + 0.641 + ], + "angle": 0, + "content": "Texture transfer. Figure 7 shows texture transfer results on ShapeNetV2 [50] to compare established dense correspondence between baselines and ours with higher-frequency label transfer. We observe notable texture transfer results. For instance, our framework successfully transfers the source car's texture of the windshield (gray) to the corresponding part of the target cars, whereas baselines transfer the same textures to the roof part of the target cars." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.702, + 0.667 + ], + "angle": 0, + "content": "5.2. Shape reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We report the reconstruction quality of unseen shapes with Chamfer Distance (CD) to validate the representational power of our framework in Table 4. A smaller CD means more accurate reconstruction. As in previous work, we report the \\(\\mathrm{CD} \\times 10^{3}\\) for readability. Also, we consistently discuss the CD and performance gain below at this scale. We optimize the corresponding shape latent code for a given unseen shape. We measure CD as in DIF [2], where reconstructed meshes are extracted at a resolution of \\(256^{3}\\), and CD is calculated with \\(10K\\) sampled points. Our framework on average achieved the best reconstruction results surpassing DIF by a margin of 0.0512. On average, our method and DIF exhibit 0.2518 and 0.3030 reconstruction errors, respectively. In three out of four categories, our method achieved the best reconstruction accuracy." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "599" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.303, + 0.108 + ], + "angle": 0, + "content": "5.3. More challenging setups" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.205 + ], + "angle": 0, + "content": "To further validate the effectiveness of our method, we perform additional experiments and compare the results with DIF [2]. Overall, the results below imply current methods that lack part semantics and do not consider part deformation consistency may fail to learn semantically plausible deformation in challenging cases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.206, + 0.471, + 0.431 + ], + "angle": 0, + "content": "2-shot part label transfer within different subcategories. Our objective is to assess dense correspondence performance under difficult settings. We conduct 2-shot part label transfer within two different subcategories belonging to the same category, where shapes from different subcategories tend to have different geometry for same semantics. Since ShapeNet [50] includes incorrect subcategory labels [58], we manually cleaned subcategory labels for four categories (e.g., airplane: airliner/jet, car: sedan/jeep, chair: straight chair/sofa, table: pedestal table/short table). Figure 8 displays examples of shapes from each subcategory. For robust evaluations, we randomly select the source shape 5 times for each subcategory. Table 5 shows our framework achieves a significant performance gain in all cases by over \\(7.2\\%\\) on average compared to DIF [2]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.471, + 0.81 + ], + "angle": 0, + "content": "Unsupervised dense correspondence in DFAUST [59]. Furthermore, we evaluate unsupervised dense correspondence with non-rigid shapes [59]. For more challenging settings, we train both DIF [2] and our model for two different subjects with different sizes and heights (e.g., subject ID of 50002, 50026) and four distinct actions (e.g., shake arms, chicken wings, running on spot, and jumping jacks). Colormap in Figure 9 demonstrates the correspondence quality, where we transfer color over the up-axis. In the source shape, the arms are labeled in red. After transferring these color labels to the target shapes, our method preserves the red/orange labeling of the arms across different motions, whereas the baseline fails to capture the same labeling (i.e., yellow hands) since it lacks semantic knowledge during training. Additionally, we conduct keypoint transfer tasks and measure the PCK score as identical as in Section 5.1. Our model achieves 74.9/85.2 with a threshold of 0.05/0.1, whereas DIF [2] achieves 67.0/80.7 with the same threshold, showing our superior performance in the challenging settings (see the supplement for the full keypoint transfer result). Note that we can also learn implicit templates for non-rigid shapes with supervision as in [19]. We conduct experiments in unsupervised settings since our focus is to validate the effectiveness of utilizing additional semantic information for template learning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.189, + 0.832 + ], + "angle": 0, + "content": "5.4. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Ablation on regularizations. We conduct ablation studies regarding the proposed regularizations including part deformation consistency regularizations, global scale consistency regularization, and global deformation consistency" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.091, + 0.906, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.156, + 0.889, + 0.171 + ], + "angle": 0, + "content": "Figure 8. Types of source shapes for 2-shot part label transfer." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.183, + 0.894, + 0.211 + ], + "angle": 0, + "content": "Table 5. 2-shot part label transfer in subcategories of ShapeNet. The reported numbers are overall mIoU over 5 runs." + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.212, + 0.887, + 0.302 + ], + "angle": 0, + "content": "
subcategoriesDIF [2]Ours
airplane(airliner/jet)64.5±1.6172.8±0.99 (+8.3↑)
car(sedan/jeep)65.1±2.3469.1±1.46 (+4.0↑)
chair(straight/sofa)66.3±6.0677.5±5.81 (+11.2↑)
table(pedestal/short)74.4±10.179.7±6.26 (+5.3↑)
" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.317, + 0.895, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.442, + 0.894, + 0.485 + ], + "angle": 0, + "content": "Figure 9. Color transfer on DFAUST [59]. We transfer different color labels over the up-axis from source shapes (colored in ivory) to target shapes and compare ours with DIF [2]." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.496, + 0.893, + 0.538 + ], + "angle": 0, + "content": "Table 6. Ablations for every proposed regularization. We report mIoU for 5-shot part label transfer on ShapeNet chair. Note that \\(\\mathcal{L}_{\\mathrm{pdc}}\\) denotes \\(\\mathcal{L}_{\\mathrm{pdc\\_geo}} + \\mathcal{L}_{\\mathrm{pdc\\_sem}}\\)." + }, + { + "type": "table", + "bbox": [ + 0.574, + 0.538, + 0.821, + 0.617 + ], + "angle": 0, + "content": "
LpdcLgeoLscalechair
(a)✓83.4
(b)✓✓83.9
(c)✓✓✓84.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.63, + 0.892, + 0.657 + ], + "angle": 0, + "content": "Table 7. BAE-Net \\( k \\) ablation. We report mIoU for 5-shot part label transfer on ShapeNet chair over 3 runs." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.657, + 0.88, + 0.69 + ], + "angle": 0, + "content": "
k=2k=4k=6k=8k=10
chair83.8±0.5183.5±1.4882.8±1.2783.6±0.8583.7±1.33
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.825 + ], + "angle": 0, + "content": "regularization. We conduct 5-shot part label transfer within a chair category, as same as Section 5.1. Table 6 shows that each component of our proposed regularizers contributes to performance improvement for dense correspondence. Note that by only using \\(\\mathcal{L}_{\\mathrm{pdc}}\\), i.e., using \\(\\mathcal{L}_{\\mathrm{pdc\\_geo}}\\) and \\(\\mathcal{L}_{\\mathrm{pdc\\_sem}}\\), it still outperforms than baselines (80.4 [1] and 82.1 [2]). The best performance was achieved by incorporating all proposed regularizations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.902 + ], + "angle": 0, + "content": "BAE-Net [31] with different \\( k \\). We evaluate the effect of \\( k \\) for BAE-Net [31] with mean/standard deviation over 3 runs for ShapeNet [50] chair in We train BAE-Net [31] with a range of \\( k = (2,4,6,8,10) \\), where \\( k = 4 \\) is used in the main paper, and report mIoU for 5-shot part label" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "600" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.168 + ], + "angle": 0, + "content": "transfer task. Table 7 shows the efficacy of semantic priors, where our framework achieves the average mIoU of 83.5, which shows higher performance than baselines. More importantly, we observed that the performance is fairly robust over \\( k \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.18, + 0.197, + 0.195 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.204, + 0.473, + 0.401 + ], + "angle": 0, + "content": "We have proposed a new framework that learns a semantic-aware implicit template field. Our framework extracts semantic knowledge from a self-supervised feature extractor given unlabelled point clouds. By leveraging the semantic information, we propose novel conditioning with semantic-aware deformation code and part deformation consistency regularizations to encourage semantically plausible deformation. Furthermore, we suggest global scaling consistency and global deformation consistency regularizations to enhance the learning process. Through quantitative and qualitative evaluation, we demonstrate the superiority of our proposed method and verified the effect of all proposed components in our framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.473, + 0.569 + ], + "angle": 0, + "content": "Acknowledgments. This research was supported in part by the MSIT (Ministry of Science and ICT), Korea, under the ICT Creative Consilience program (IITP-2023-2020-0-01819) supervised by the IITP (Institute for Information & communications Technology Planning & Evaluation); the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF-2023R1A2C2005373); and Google Cloud Research Credits program with the award (GCLF-67WL-WU6D-MF14, V9P3-99XD-64CL-EH5H)." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.58, + 0.175, + 0.595 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.605, + 0.47, + 0.647 + ], + "angle": 0, + "content": "[1] Zerong Zheng, Tao Yu, Qionghai Dai, and Yebin Liu. Deep implicit templates for 3d shape representation. In CVPR, 2021. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.65, + 0.47, + 0.692 + ], + "angle": 0, + "content": "[2] Yu Deng, Jiaolong Yang, and Xin Tong. Deformed implicit field: Modeling 3d shapes with learned dense correspondence. In CVPR, 2021. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.694, + 0.47, + 0.736 + ], + "angle": 0, + "content": "[3] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: a skinned multiperson linear model. TOG, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.739, + 0.469, + 0.768 + ], + "angle": 0, + "content": "[4] Brett Allen, Brian Curless, and Zoran Popovic. Articulated body deformation from range scan data. TOG, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.77, + 0.469, + 0.812 + ], + "angle": 0, + "content": "[5] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan C. Russell, and Mathieu Aubry. 3d-coded: 3d correspondences by deep deformation. In ECCV, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.814, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[6] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[7] Wang Zeng, Wanli Ouyang, Ping Luo, Wentao Liu, and Xiaogang Wang. 3d human mesh regression with dense correspondence. In CVPR, 2020. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.605, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[8] Sing Chun Lee and Misha Kazhdan. Dense point-to-point correspondences between genus-zero shapes. In Computer Graphics Forum, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.138, + 0.892, + 0.18 + ], + "angle": 0, + "content": "[9] Silvia Zuffi, Angjoo Kanazawa, David W. Jacobs, and Michael J. Black. 3d menagerie: Modeling the 3d shape and pose of animals. In CVPR, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.184, + 0.892, + 0.238 + ], + "angle": 0, + "content": "[10] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In AVSS, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.244, + 0.892, + 0.284 + ], + "angle": 0, + "content": "[11] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. TOG, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[12] Luca Ballan, Aparna Taneja, Jürgen Gall, Luc Van Gool, and Marc Pollefeys. Motion capture of hands in action using discriminative salient points. In ECCV, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[13] Srinath Sridhar, Antti Oulasvirta, and Christian Theobalt. Interactive markerless articulated hand motion tracking using rgb and depth data. In ICCV, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.38, + 0.892, + 0.422 + ], + "angle": 0, + "content": "[14] Iason Oikonomidis, Manolis IA Lourakis, and Antonis A Argyros. Evolutionary quasi-random search for hand articulations tracking. In CVPR, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.426, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[15] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In CVPR, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.471, + 0.892, + 0.527 + ], + "angle": 0, + "content": "[16] Vignesh Ganapathi-Subramanian, Olga Diamanti, Soeren Pirk, Chengcheng Tang, Matthias Niessner, and Leonidas Guibas. Parsing geometry using structure-aware shape templates. In 3DV, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.53, + 0.892, + 0.585 + ], + "angle": 0, + "content": "[17] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan Russell, and Mathieu Aubry. AtlasNet: A Papier-Mâché Approach to Learning 3D Surface Generation. In CVPR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[18] Pablo Palafox, Nikolaos Sarafianos, Tony Tung, and Angela Dai. Spams: Structured implicit parametric models. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.635, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[19] Ramana Sundararaman, Gautam Pai, and Maks Ovsjanikov. Implicit field supervision for robust non-rigid shape matching. In ECCV, 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.681, + 0.892, + 0.75 + ], + "angle": 0, + "content": "[20] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.754, + 0.892, + 0.783 + ], + "angle": 0, + "content": "[21] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.786, + 0.892, + 0.841 + ], + "angle": 0, + "content": "[22] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[23] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "601" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[24] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.139, + 0.47, + 0.194 + ], + "angle": 0, + "content": "[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.2, + 0.469, + 0.256 + ], + "angle": 0, + "content": "[26] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.469, + 0.316 + ], + "angle": 0, + "content": "[27] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.322, + 0.469, + 0.377 + ], + "angle": 0, + "content": "[28] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.382, + 0.469, + 0.437 + ], + "angle": 0, + "content": "[29] Tarun Yenamandra, Ayush Tewari, Florian Bernard, Hans-Peter Seidel, Mohamed Elgharib, Daniel Cremers, and Christian Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.443, + 0.469, + 0.485 + ], + "angle": 0, + "content": "[30] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.532 + ], + "angle": 0, + "content": "[31] Zhiqin Chen, Kangxue Yin, Matthew Fisher, Siddhartha Chaudhuri, and Hao Zhang. Bae-net: Branched autoencoder for shape co-segmentation. In ICCV, 2019. 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.537, + 0.469, + 0.592 + ], + "angle": 0, + "content": "[32] Vladimir G Kim, Wilmot Li, Niloy J Mitra, Siddhartha Chaudhuri, Stephen DiVerdi, and Thomas Funkhouser. Learning part-based templates from large collections of 3d shapes. TOG, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.598, + 0.469, + 0.652 + ], + "angle": 0, + "content": "[33] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T. Freeman, and Thomas A. Funkhouser. Learning shape templates with structured implicit functions. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.658, + 0.469, + 0.713 + ], + "angle": 0, + "content": "[34] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir Kim, Bryan Russell, and Mathieu Aubry. Learning elementary structures for 3d shape generation and matching. In NeurIPS, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[35] Jiahui Lei and Kostas Daniilidis. Cadex: Learning canonical deformation coordinate space for dynamic surface representation via neural homeomorphism. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.766, + 0.469, + 0.807 + ], + "angle": 0, + "content": "[36] Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural template: Topology-aware reconstruction and disentangled generation of 3d meshes. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.813, + 0.469, + 0.854 + ], + "angle": 0, + "content": "[37] Zekun Hao, Hadar Averbuch-Elor, Noah Snavely, and Serge Belongie. Dualsdf: Semantic shape manipulation using a two-level representation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[38] Yueqi Duan, Haidong Zhu, He Wang, Li Yi, Ram Nevatia, and Leonidas J. Guibas. Curriculum deepsdf. In ECCV, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[39] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.14, + 0.892, + 0.18 + ], + "angle": 0, + "content": "[40] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.186, + 0.892, + 0.227 + ], + "angle": 0, + "content": "[41] Michael Niemeyer and Andreas Geiger. Campari: Camera-aware decomposed generative neural radiance fields. In 3DV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[42] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, et al. Gaudi: A neural architect for immersive 3d scene generation. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.308, + 0.892, + 0.35 + ], + "angle": 0, + "content": "[43] Chengjie Niu, Manyi Li, Kai Xu, and Hao Zhang. Rimnet: Recursive implicit fields for unsupervised learning of hierarchical shape structures. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.355, + 0.892, + 0.409 + ], + "angle": 0, + "content": "[44] Chenyang Zhu, Kai Xu, Siddhartha Chaudhuri, Li Yi, Leonidas J Guibas, and Hao Zhang. Adacoseg: Adaptive shape co-segmentation with group consistency loss. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.415, + 0.892, + 0.47 + ], + "angle": 0, + "content": "[45] Minghua Liu, Yinhao Zhu, Hong Cai, Shizhong Han, Zhan Ling, Fatih Porikli, and Hao Su. Part slip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[46] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.536, + 0.892, + 0.578 + ], + "angle": 0, + "content": "[47] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.583, + 0.892, + 0.625 + ], + "angle": 0, + "content": "[48] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[49] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.677, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[50] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv:1512.03012, 2015. 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.752, + 0.892, + 0.778 + ], + "angle": 0, + "content": "[51] David Ha, Andrew M. Dai, and Quoc V. Le. Hypernetworks. In ICLR, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.784, + 0.892, + 0.839 + ], + "angle": 0, + "content": "[52] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In NeurIPS, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[53] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "602" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[54] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.139, + 0.47, + 0.194 + ], + "angle": 0, + "content": "[55] Yang You, Yujing Lou, Chengkun Li, Zhoujun Cheng, Liangwei Li, Lizhuang Ma, Cewu Lu, and Weiming Wang. Keypointnet: A large-scale 3d keypoint dataset aggregated from numerous human annotations. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.199, + 0.47, + 0.255 + ], + "angle": 0, + "content": "[56] Li Yi, Vladimir G. Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas J. Guibas. A scalable active framework for region annotation in 3d shape collections. TOG, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.47, + 0.301 + ], + "angle": 0, + "content": "[57] Li Yi, Hao Su, Xingwen Guo, and Leonidas Guibas. Syncspeccnn: Synchronized spectral cnn for 3d shape segmentation. In CVPR, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.306, + 0.47, + 0.361 + ], + "angle": 0, + "content": "[58] Xinhai Liu, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Fine-grained 3d shape classification with hierarchical part-view attention. IEEE Transactions on Image Processing, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.367, + 0.47, + 0.408 + ], + "angle": 0, + "content": "[59] Federica Bogo, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Dynamic FAUST: Registering human bodies in motion. In CVPR, 2017. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "603" + } + ] +] \ No newline at end of file diff --git a/2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/840dd92d-8817-48f0-ac2c-586a3d27151d_origin.pdf b/2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/840dd92d-8817-48f0-ac2c-586a3d27151d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b311e321c5a7ca1e8c71909416710009c1537e12 --- /dev/null +++ b/2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/840dd92d-8817-48f0-ac2c-586a3d27151d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2588ef90fd0697b8f30d88d98ed503c417918fb95bc4f5c5ad7a745dd7094b2f +size 3691891 diff --git a/2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/full.md b/2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2c98ba40e8b225896170fa5c9b5746cc5a4adfa4 --- /dev/null +++ b/2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/full.md @@ -0,0 +1,382 @@ +# Semantic-Aware Implicit Template Learning via Part Deformation Consistency + +Sihyeon Kim Minseok Joo Jaewon Lee +Juyeon Ko Juhan Cha Hyunwoo J. Kim* + +Department of Computer Science and Engineering, Korea University + +{sh_bs15, wlgkcjf87, 2jlejyu, juyon98, hanchaa, hyunwoojkim}@korea.ac.kr + +# Abstract + +Learning implicit templates as neural fields has recently shown impressive performance in unsupervised shape correspondence. Despite the success, we observe current approaches, which solely rely on geometric information, often learn suboptimal deformation across generic object shapes, which have high structural variability. In this paper, we highlight the importance of part deformation consistency and propose a semantic-aware implicit template learning framework to enable semantically plausible deformation. By leveraging semantic prior from a self-supervised feature extractor, we suggest local conditioning with novel semantic-aware deformation code and deformation consistency regularizations regarding part deformation, global deformation, and global scaling. Our extensive experiments demonstrate the superiority of the proposed method over baselines in various tasks: keypoint transfer, part label transfer, and texture transfer. More interestingly, our framework shows a larger performance gain under more challenging settings. We also provide qualitative analyses to validate the effectiveness of semantic-aware deformation. The code is available at https://github.com/mlvlab/PDC. + +# 1. Introduction + +Template learning is essential for shape analysis since the template, which is a compact representation of arbitrary shapes, provides a good prior for shape fitting so that we can establish dense correspondence and transfer key attributes such as texture or keypoints between shapes. Traditionally, handcrafted templates such as meshes, part elements with primitive shapes (e.g., sphere, cylinder, cone), or parametric models have been widely used for human/animal body [3, 4, 5, 6, 7, 8, 9], human face [10, 11], human hand [12, 13, 14], and generic objects [15, 16, 17]. Defining templates for less complex shapes with consistent topologies and similar structures is relatively straightforward. However, for shapes like generic object shapes with high structural variability or + +![](images/b8c7719d3d700afb1d874be71a8a7714a1fb5c826244ecbfeb71ba0467281d7f.jpg) + +![](images/068252e044fa8ddddc802a3c0d2b98ea36759befa0f5a17a5bcffdd3449bdcb3.jpg) +Figure 1. Importance of part deformation consistency. Given input shapes with high structural variability, only considering low-level geometry during template learning [1, 2] can lead to suboptimal deformation (e.g., cannot distinguish between seat and arm where two parts are geometrically close while semantically distant: low seat-low arm, high seat-high arm). Ours succeeds in transferring shape attributes by encouraging part deformation consistency with semantic information. + +non-rigid shapes with large-scale deformations, manually defining suitable templates becomes challenging. + +To address this challenge, recent works [1, 2, 18, 19] have focused on learning implicit templates based on neural fields [20]. Neural fields, which are known for strong representation power, efficiency, and flexibility, have been successful in multiple 3D vision tasks, e.g., shape/scene reconstruction [21, 22, 23, 24], neural rendering [25, 26, 27], and human digitization [28, 29, 30]. Likewise, neural fields-based template learning have shown superior perfor + +mance on dense correspondence between complex shapes. Both [1] and [2] suggest decomposing the implicit representation of shapes into a deformation field and a template field. Here, the template is learned as a continuous function, which is expected to capture shared structures among the given shapes within a category (e.g., airplane, chair), while the deformation depends solely on the conditioned geometry of each shape (i.e., SDF value and surface normal). We observe current approaches often learn suboptimal deformation especially when a shape category has high structural variability like diverse local deformation scales and unseen/missing part structures; see Figure 1. That is, incorporating extra knowledge of part semantics is necessary to handle such cases. + +In this paper, we propose a novel framework to learn semantically plausible deformation by distilling the knowledge of self-supervised feature extractor [31] and employing it as a semantic prior to satisfy part deformation consistency during the template learning process. Part deformation consistency is our inductive bias that the deformation result within the same part should be consistent. By leveraging semantic prior to understand part arrangements of shapes, we suggest local conditioning with semantic-aware deformation code and carefully designed regularizations to encourage semantically corresponding deformation. Semantic-aware deformation code is a point-wise soft assignment of part deformation priors, where part deformation priors are multiple latent codes to encode information on corresponding part deformation. We calculate assigned weights considering the part semantics of each point. In addition, we propose input space regularization and latent space regularization to encourage part-level distances to be close, providing an explicit way to control the deformation field so that the model can learn flexible deformation and successfully learn common structures under diverse shapes. Lastly, we suggest global scale consistency regularization to preserve the volume of the template against large-scale deformations. To validate the effectiveness of our framework, we conduct extensive surrogate experiments reflecting different levels of correspondence such as keypoint transfer, part label transfer, and texture transfer, and demonstrate competitive performance over baselines with qualitative analyses. + +Our contributions are summarized as follows: ① We propose a new framework that learns global implicit template fields founded on part deformation consistency while understanding part semantics to enable semantically plausible deformation, ② We impose hybrid conditioning by combining global conditioning of geometry and local conditioning of part semantics for flexible deformation, ③ We design novel regularizations to successfully manipulate the template fields, ④ We show the effectiveness of our framework with both qualitative and quantitative experiments. + +# 2. Related Works + +Template learning for shape correspondence. Defining a template for a set of arbitrary shapes has been studied for a long time. A high-quality template provides a good initialization for shape alignment and various applications in vision and graphics. For shapes with consistent topologies and structures such as faces or bodies of humans and animals, handcrafted templates have been popular [3, 5, 7, 8, 9] and still widely used today. On the other hand, it was more challenging for generic object shapes with inconsistent topology and diverse part structures. Prior works suggest to predefine or learn local templates from data as in primitive parts [15] or part elements [32, 16, 33, 17, 34] to handle the inconsistent structure. Recent works have tried to learn a canonical shape of generic objects through implicit template field [1, 2, 35, 36], which are proven to be effective for representing shapes. However, existing approaches mostly lack explicit consideration of semantic knowledge, which enables semantically plausible deformation so that the learned template captures a common structure even with the various structured shapes. + +Neural fields for 3D representations. Recent studies have shown that representing 3D shapes or scenes as implicit representations with a neural network has several advantages: powerful representation, flexibility, and memory efficiency. These are called neural fields [20]. For instance, Occupancy networks [22] proposed a method to represent 3D geometry as a continuous occupancy function. DeepSDF [23] learns 3D complex surfaces with a neural network of auto-decoder framework for 3D modeling. Further, implicit representation has been introduced to various applications [37, 38]. To leverage the advantages, many applications based on neural fields have emerged such as neural rendering [25, 26, 27], human digitization [28, 29, 30], and generative modeling [39, 40, 41, 42]. They have also been applied to learn the template of complex shapes as mentioned above. + +Utilizing self-supervised segmentation models. Recently, there has been a growing body of research on analyzing complex shapes in a self-supervised manner due to the high cost of annotation for 3D data. Unsupervised semantic segmentation/co-segmentation models like BAE-Net [31], RIM-Net [43], AdaCoSeg [44], and PartSLIP [45] have demonstrated the ability to comprehend part semantics without ground truth labels. For 2D and multi-modal tasks, using large-scale self-supervised segmentation models like DINO [46] as a feature extractor has been widely adopted, including pixelNeRF [47], Feature Field [48], and ClipN-erF [49]. Inspired by these recent works, we utilize pretrained unsupervised part segmentation models to harness part semantic information for template learning. + +![](images/2c8c62fcf1b166e69f5abf8d598351ffd928a7e37ec0a0d5fc4bebd571ae0dd2.jpg) +Figure 2. Framework of semantic-aware implicit template learning. To learn semantic-aware implicit template field, (a) we extract part semantics from pretrained feature extractor $E$ and incorporate knowledge into semantic-aware deformation code $(\alpha)$ and (b) part deformation consistency regularizations. With optimizing with the reconstruction loss $\mathcal{L}_{\mathrm{rec}}$ and several regularizers, our framework succeeds to learn semantically plausible deformation. + +![](images/ee7a919c08e58d123d46598646c37fef17ae02e62c65416039063fcee5aef637.jpg) +(a) + +![](images/f523e7f3c92f502460306e545568713d7bd79a9809d81f4cd9455b4a4c8abcb5.jpg) + +![](images/801ade73fef454d3add5f9bef7712229db9601be40b7f48707a00ebd1f041aa5.jpg) +(b) + +# 3. Preliminaries + +Implicit templates have been studied with neural fields. Given a 3D shape $P$ , a neural field, which is a neural network $\mathcal{F}_{\theta}$ parameterized by $\theta$ , approximates continuous functions such as an occupancy function [21, 22] or a Signed Distance Function (SDF) [23] as + +$$ +\mathcal {F} _ {\theta} (\boldsymbol {x}; \boldsymbol {z} _ {P}) = s. \tag {1} +$$ + +Herein, $\pmb{x} \in \mathbb{R}^3$ is a query point and $s$ is an occupancy/SDF function value. To represent multiple 3D shapes, for each shape, a shape latent code $\pmb{z}_P \in \mathbb{R}^d$ is learned. In general, a surface is defined as the zero-level set of a SDF, i.e., $\{x | \mathcal{F}_{\theta}(x; z_P) = 0\}$ as in DeepSDF [23]. + +The neural field can be further decomposed into an implicit template field $\mathcal{T}$ and a deformation field $\mathcal{D}[1,2]$ : + +$$ +\mathcal {F} _ {\theta} (\boldsymbol {x}; \boldsymbol {z} _ {P}) = \mathcal {T} _ {\theta_ {2}} \left(\mathcal {D} _ {\theta_ {1}} (\boldsymbol {x}; \boldsymbol {z} _ {P})\right). \tag {2} +$$ + +The implicit template field $\mathcal{T}_{\theta_2}$ captures a common structure and outputs an occupancy/SDF value $s$ in the template space. $\mathcal{D}_{\theta_1}$ maps a point $x$ from a shape-specific space to the common template space by adding a deformation vector $\Delta \pmb{x} \in \mathbb{R}^3$ , i.e., $\pmb{x} + \Delta \pmb{x}$ . In DIF-Net [2], $\mathcal{D}_{\theta_1}$ additionally predicts a correction value $\Delta s \in \mathbb{R}$ to capture instance-specific local structures. In this case, the neural field $\mathcal{F}_{\theta}$ is defined as + +$$ +\mathcal {F} _ {\theta} (\boldsymbol {x}; \boldsymbol {z} _ {P}) = \mathcal {T} _ {\theta_ {2}} (\boldsymbol {x} + \Delta \boldsymbol {x}) + \Delta s \tag {3} +$$ + +$$ +\Delta s, \Delta \boldsymbol {x} = \mathcal {D} _ {\theta_ {1}} (\boldsymbol {x}; \boldsymbol {z} _ {P}). +$$ + +The existing approaches [1, 2] learn the implicit template field by minimizing the distance between the template and the deformed individual shapes. The distance is measured by SDF values and surface normals. Also, several regularizers are imposed such as deformation smoothness (e.g., penalizing the spatial gradient of $\mathcal{D}_{\theta_1}$ and restricting the deformation scale between points). However, the loss functions + +and regularizers are computed by low-level geometric information. They are not effective enough to handle generic objects with large structural variability. Hence, in this paper, we take part semantics into account for template learning to achieve 'semantically plausible' deformation and dense correspondence. + +# 4. Method + +We present a semantic-aware implicit template learning framework based on part deformation consistency. In this section, we first delineate the overall architecture of our method equipped with part deformation priors and semantic-aware deformation code. Then, we introduce our deformation consistency regularizers including part deformation consistency regularizations. Lastly, we discuss the training procedure with several additional loss functions. + +# 4.1. Semantic-aware implicit template field + +We here present a novel semantic-aware implicit template learning framework. The overall pipeline of the proposed method is shown in Figure 2(a). + +Template field and deformation field. We extend the implicit template learning by incorporating semantic information into the deformation field. Specifically, given a query point $\pmb{x}$ of a shape $P$ , the deformation field $\mathcal{D}_{\theta_1}$ predicts a deformation vector $\Delta x$ and a correction value $\Delta s$ conditioned on a point-wise semantic-aware deformation code (SDC) $\alpha(x; P)$ in addition to a shape latent code $z_P$ . Then, our framework is defined as + +$$ +\mathcal {F} _ {\theta} (\boldsymbol {x}; P, \mathcal {E}) := \mathcal {T} _ {\theta_ {2}} (\boldsymbol {x} + \Delta \boldsymbol {x}) + \Delta s, \tag {4} +$$ + +$$ +\Delta \boldsymbol {x}, \Delta s = \mathcal {D} _ {\theta_ {1}} (\boldsymbol {x}; \boldsymbol {z} _ {P}, \boldsymbol {\alpha} (\boldsymbol {x}; P)), +$$ + +where $\mathcal{E}$ is a pretrained feature extractor. We assign $\alpha (\pmb {x};P)$ based on a point-wise semantic feature, which is extracted from $\mathcal{E}$ . For simplicity, we denote the neural field $\mathcal{F}_{\theta}(\pmb {x};P,\mathcal{E})$ by $\mathcal{F}_{\theta}(\pmb {x})$ and SDC $\alpha (\pmb {x};P)$ by $\alpha (\pmb {x})$ . + +Feature extractor. In this paper, for feature extraction, we mainly use BAE-Net [31] as an encoder $\mathcal{E}$ , which is proven effective in capturing part semantics. Note that we pretrained BAE-Net on selected categories of ShapeNetV2 [50] by self-supervision without any extra labels or data for a fair comparison. The features from BAE-Net are $k$ dimensional vectors, i.e., $\boldsymbol{o}^P(\boldsymbol{x}) = \mathcal{E}(\boldsymbol{x}, P) \in \mathbb{R}^k$ . $k$ is a hyperparameter, which is closely related to the number of semantic parts. Since the number of semantic parts is unknown in self-supervised learning, we simply chose sufficiently large $k$ . + +Semantic-aware deformation code. Our framework assumes that different semantic parts necessitate different deformations. To achieve this, the proposed method learns a deformation prior for each semantic part. The deformation priors are encoded as latent codes $e_1, \dots, e_k \in \mathbb{R}^{d'}$ . A point $x$ is softly assigned to $k$ semantic parts using the softmax values of semantic features $o^P(x)$ . Combining the soft assignment with deformation priors, semantic-aware deformation code (SDC) denoted as $\alpha(x) \in \mathbb{R}^{d'}$ is defined as + +$$ +\boldsymbol {\alpha} (\boldsymbol {x}) = \sum_ {i = 1} ^ {k} \frac {\exp \left(\boldsymbol {o} _ {i} ^ {P} (\boldsymbol {x})\right) \boldsymbol {e} _ {i}}{\sum_ {j = 1} ^ {k} \exp \left(\boldsymbol {o} _ {j} ^ {P} (\boldsymbol {x})\right)}, \forall i \in \{1, \dots , k \}, \tag {5} +$$ + +where $\pmb{o}_i^P(\pmb{x})$ denotes the $i$ -th dimension value of the semantic features. The SDC at $\pmb{x}$ is computed by a linear combination of deformation priors. We learn the point's probability of belonging to each of the $k$ parts, which is served as part semantics. + +Remarks. For comparison, we also applied hard assignment of part deformation priors for SDC, i.e., $\alpha(\pmb{x}) = e_{i^*}$ , where $i^* = \operatorname{argmax}_i (\pmb{o}_i^P(\pmb{x}))$ . However, as points are completely separated with local conditioning of the corresponding part deformation prior $e_{i^*}$ , the final deformation result tends to break the shape geometry (see Figure 3), leading to performance degradation as shown in Table 1. Thereby, we choose the soft assignment to enable flexible deformation while preserving the shape geometry. + +We train the deformation field $\mathcal{D}_{\theta_1}$ and $\mathcal{T}_{\theta_2}$ in an end-to-end fashion. In our experiments, $\mathcal{D}_{\theta_1}$ is a 6-layer MLP conditioned on both global geometry $z_P$ via hypernetwork $\Psi$ [51, 52] and local part semantics $\alpha(x)$ via concatenation with input coordinates. The shape latent code $z_P$ and part deformation priors $\{\pmb{e}_i\}$ are randomly initialized and jointly optimized with neural fields. $\mathcal{T}_{\theta_2}$ is a 5-layer MLP. + +Table 1. Ablations of different assignment strategies for SDC. We conduct a quantitative comparison between hard and soft assignments with 5-shot part label transfer on ShapeNet chair. + +
Assignment strategiesHardSoft
chair (5-shot PLT)77.084.2
+ +![](images/3ffcd3cae66c1d892adb41f0722cb7d9eaec6987c5d6eee85e1dd68309f33035.jpg) +Figure 3. Effects of different assignment strategies for SDC. Hard assignment breaks the shape geometry, while soft assignment preserves the geometry while learning flexible deformations. + +# 4.2. Part deformation consistency regularizations + +In addition to conditioning neural fields, leveraging semantic information, we propose novel part deformation consistency regularizations to encourage semantically plausible deformation. Semantically similar points should be close after deformation in both an input space and a latent space. That is, the deformation of points from the same part should be consistent. Figure 2(b) illustrates the concept of our regularizations. Given a pair of deformed shapes $P', Q'$ , part deformation consistency regularizations are defined as + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {p d c - g e o}} = \sum_ {i = 1} ^ {k} \left(\frac {1}{| \mathcal {P} _ {i} ^ {\prime} |} \sum_ {\boldsymbol {x} \in \mathcal {P} _ {i} ^ {\prime}} \| \boldsymbol {x} - C (\boldsymbol {x}, \mathcal {Q} _ {i} ^ {\prime}) \| _ {2} ^ {2} \right. \\ \left. + \frac {1}{\left| \mathcal {Q} _ {i} ^ {\prime} \right|} \sum_ {\boldsymbol {y} \in \mathcal {Q} _ {i} ^ {\prime}} \| \boldsymbol {y} - C (\boldsymbol {y}, \mathcal {P} _ {i} ^ {\prime}) \| _ {2} ^ {2}\right), \text {a n d} \tag {6} \\ \end{array} +$$ + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {p d c} \text {- s e m}} = \sum_ {i = 1} ^ {k} \left(\frac {1}{| \mathcal {P} _ {i} ^ {\prime} |} \sum_ {\boldsymbol {x} \in \mathcal {P} _ {i} ^ {\prime}} \| \boldsymbol {o} ^ {P ^ {\prime}} (\boldsymbol {x}) - \boldsymbol {o} ^ {Q ^ {\prime}} \left(C (\boldsymbol {x}, \mathcal {Q} _ {i} ^ {\prime})\right) \| _ {2} ^ {2} \right. \\ \left. + \frac {1}{| \mathcal {Q} _ {i} ^ {\prime} |} \sum_ {\boldsymbol {y} \in \mathcal {Q} _ {i} ^ {\prime}} \| \boldsymbol {o} ^ {Q ^ {\prime}} (\boldsymbol {y}) - \boldsymbol {o} ^ {P ^ {\prime}} \left(C (\boldsymbol {y}, \mathcal {P} _ {i} ^ {\prime})\right) \| _ {2} ^ {2}\right), \tag {7} \\ \end{array} +$$ + +where $C$ is a correspondence matching function for a point $\pmb{x}$ and its nearest point in a set $\mathcal{Q}$ , i.e., $C(\pmb{x},\mathcal{Q}) = \operatorname*{argmin}_{\pmb{y}\in \mathcal{Q}}\| \pmb{x} - \pmb{y}\|_2^2$ . $\mathcal{P}_i'$ and $\mathcal{Q}_i'$ denote sampled point sets of + +$i^{\mathrm{th}}$ deformed part shapes of $P^{\prime}$ and $Q^{\prime}$ . They are defined as $\mathcal{P}_i^\prime = \{\pmb {x} + \mathcal{D}_{\theta_1}(\pmb {x};\pmb {z}_P,\pmb {\alpha}(\pmb {x};P))\mid i = \operatorname *{argmax}_{k}(\pmb {o}_k^P (\pmb {x})),\pmb {x}\in$ + +$\Omega_P\}$ and $\mathcal{Q}_i^\prime = \{\pmb {x} + \mathcal{D}_{\theta_1}(\pmb {x};\pmb {z}_Q,\pmb {\alpha}(\pmb {x};Q))|i =$ argmax $(\pmb{o}_k^Q (\pmb {x})),\pmb {x}\in \Omega_Q\}$ , where $\Omega_P$ and $\Omega_{Q}$ are sam + +pled query point sets for shape $P$ and $Q$ . Note that each deformed part shape is extracted based on part semantics from original shapes $P$ and $Q$ . Here, we explicitly guide the part-level deformation among various shapes to be consistent. Moreover, since we understand part arrangement and + +![](images/c11deb1e2f34ad9c566244693f1e65d187963d0795697e23acc5908163a4b361.jpg) +Figure 4. Keypoint transfer result on ShapeNet. We transfer keypoint labels from source shapes (colored in ivory) to target shapes and compare ours with DIT [1] and DIF [2]. For additional visualizations including results on ScanObjectNN [53], see the supplement. + +![](images/cafeb9d58e02ba8740aa2249583b33b699f5429902f293ffbc09e7ad539f23d0.jpg) +(a) + +![](images/4b36f80cd31eee68e6a5b448f1efa22671b904462eb9211f55d754c17fb51573.jpg) +(b) + +![](images/f5876dceecdf1c4231267053c9f0d994cbe0c8858bedf5f786589aa659e3d8c1.jpg) +(c) +Figure 5. Efficacy of global scaling consistency $\mathcal{L}_{\mathrm{scale}}$ + +only penalize matched parts, our method does not forcibly deform "unseen/missing match" cases. + +We partly observed spatial distortion when part deformation consistency was too strongly imposed. This is because using external semantics for deformation could break the continuity of geometric fields. In addition, we propose global scale consistency and global deformation consistency to address the concern. For the former one, we disentangle a scaling operation from the entire non-rigid deformation $\mathcal{D}_{\theta_1}$ to calculate the global scale of the learned implicit template field. That is, we want to approximate $\mathcal{D}$ with the global scaling of the scaling factor of $r$ , i.e., $\mathcal{D} \approx rI$ . Then the problem becomes finding $r$ from + +$$ +\min _ {r} \sum_ {s} \sum_ {i} \| \boldsymbol {x} _ {i} ^ {s} + \Delta \boldsymbol {x} _ {i} ^ {s} - r \boldsymbol {x} _ {i} ^ {s} \| _ {2} ^ {2}, \tag {8} +$$ + +where $s$ and $i$ are an index for shape and point, respectively. A closed-form solution of (8) is $r = \sum_{s}\sum_{i}\frac{\pmb{x}_i^{s\top}(\pmb{x}_i^s + \Delta\pmb{x}_i^s)}{\pmb{x}_i^{s\top}\pmb{x}_i^s}$ . See the supplement for proof. Since we want the learned $\mathcal{D}$ to be insensitive to global scale operation, i.e., the implicit template to keep its scale by the average scale of input shapes, we penalize the expected $\mathbb{E}[r]$ over input shapes $P_{1}\dots P_{N}$ in a single batch as: + +$$ +\mathcal {L} _ {\text {s c a l e}} = | \mathbb {E} [ r ] - 1 | \tag {9} +$$ + +Figure 5 illuminates the effect of $\mathcal{L}_{\mathrm{scale}}$ . For the same input shapes, Figure 5(a) is a car template learned without $\mathcal{L}_{\mathrm{scale}}$ , Figure 5(b) is a car template learned with $\mathcal{L}_{\mathrm{scale}}$ , and Figure 5(c) is the result of placing two shapes in the same space. It shows $\mathcal{L}_{\mathrm{scale}}$ helps the model to learn a more desirable template by penalizing the change in the global scale of the learned template fields. + +For the latter one, we add global deformation consistency for a pair of deformed shapes $P', Q'$ : + +$$ +\mathcal {L} _ {\mathrm {g e o}} = \mathbf {C D} \left(P ^ {\prime}, Q ^ {\prime}\right), \tag {10} +$$ + +where $\mathbf{CD}$ denotes a conventional Chamfer loss. Thus, the deformation consistency regularization can be written as: + +$$ +\mathcal {L} _ {\mathrm {d c}} = \gamma_ {1} \mathcal {L} _ {\mathrm {p d c - g e o}} + \gamma_ {2} \mathcal {L} _ {\mathrm {p d c - s e m}} + \gamma_ {3} \mathcal {L} _ {\text {s c a l e}} + \gamma_ {4} \mathcal {L} _ {\text {g e o}}. \tag {11} +$$ + +# 4.3. Optimization + +We train our deformation and template fields with a reconstruction loss and several regularizers including our deformation consistency regularization. Following [2], the reconstruction loss $\mathcal{L}_{\mathrm{rec}}$ for shape $P$ is + +$$ +\begin{array}{l} \mathcal {L} _ {\text {r e c}} = \sum_ {\boldsymbol {x} \in \Omega_ {P}} | \mathcal {F} _ {\theta} (\boldsymbol {x}) - \bar {s} | + \sum_ {\boldsymbol {x} \in P} (1 - \langle \nabla \mathcal {F} _ {\theta} (\boldsymbol {x}), \bar {n} \rangle) \\ + \sum_ {\boldsymbol {x} \in \Omega_ {P}} | \| \nabla \mathcal {F} _ {\theta} (\boldsymbol {x}) \| _ {2} - 1 | + \sum_ {\boldsymbol {x} \in \Omega_ {P} \backslash P} \exp (- \delta \cdot | \mathcal {F} _ {\theta} (\boldsymbol {x}) |), \tag {12} \\ \end{array} +$$ + +where $\bar{s}$ denotes the ground-truth SDF value of each query point, $\bar{n}$ is the ground truth surface normal vector of each surface point and $\nabla \mathcal{F}_{\theta}$ the spatial derivative of a 3D field. The second term of (12) is derived from the property of the SDF function that the gradient of an SDF is identical to the surface normal for surface points and the third is implicit geometric regularization [54] derived from the Eikonal equation. The last term pertains to penalizing off-surface points to be away from the surface points, which is controlled by $\delta \gg 1$ . + +Since we do not know ground truth mapping between the template and input shapes, we manipulate the deformation field through multiple regularizers. First, we employ our proposed deformation consistency regularization (11). For further regularizers, we adopt [2] such as deformation smoothness $\mathcal{L}_{\mathrm{smooth}}$ , normal consistency $\mathcal{L}_{\mathrm{normal}}$ and minimal correction $\mathcal{L}_{\mathrm{c}}$ , which are defined as + +$$ +\mathcal {L} _ {\text {s m o o t h}} = \sum_ {\boldsymbol {x} \in \Omega_ {P}} \sum_ {d \in \{X, Y, Z \}} \| \nabla \mathcal {D} _ {\theta_ {1}} ^ {v} (\cdot , \boldsymbol {z} _ {P}) | _ {d} (\boldsymbol {x}) \| _ {2} ^ {2}, \tag {13} +$$ + +![](images/da82899a2c9ce4b1dc7fc4cccec1d8fceb2404da3d183bb1d68551ee615a29cb.jpg) +Figure 6. Part label transfer result on ShapeNet. We transfer part segment labels from source shapes (colored in ivory) to target shapes and compare ours with DIT [1] and DIF [2]. For additional visualizations, please refer to the supplement. + +$$ +\mathcal {L} _ {\text {n o r m a l}} = \sum_ {i} \sum_ {\boldsymbol {x} \in \mathcal {S} _ {i}} \left(1 - \left\langle \nabla \mathcal {T} _ {\theta_ {2}} \left(\boldsymbol {x} + D _ {\theta_ {1}} ^ {v} (\boldsymbol {x})\right), \bar {n} \right\rangle\right), \tag {14} +$$ + +$$ +\mathcal {L} _ {\mathrm {c}} = \sum_ {i} \sum_ {\boldsymbol {x} \in \Omega_ {P}} | \mathcal {D} _ {\theta_ {1}} ^ {\Delta s} (\boldsymbol {x}) |. \tag {15} +$$ + +We also apply $\ell_2$ regularization on shape latent codes and part deformation priors + +$$ +\mathcal {L} _ {\mathrm {e m b}} = \sum_ {i} ^ {N} \| z _ {i} \| _ {2} ^ {2} + \sum_ {j} ^ {k} \| e _ {j} \| _ {2} ^ {2}. \tag {16} +$$ + +The regularizers are collectively written as + +$$ +\mathcal {L} _ {\text {r e g}} = \gamma_ {5} \mathcal {L} _ {\text {s m o o t h}} + \gamma_ {6} \mathcal {L} _ {\text {n o r m a l}} + \gamma_ {7} \mathcal {L} _ {\mathrm {c}} + \gamma_ {8} \mathcal {L} _ {\text {e m b}}. \tag {17} +$$ + +Hence, the final loss becomes + +$$ +\mathcal {L} = \mathcal {L} _ {\text {r e c}} + \mathcal {L} _ {\mathrm {d c}} + \mathcal {L} _ {\text {r e g}}. \tag {18} +$$ + +Evaluating dense correspondence. Given two pairs of seen or unseen shapes $P, Q$ , we can evaluate the dense correspondence between $P$ and $Q$ by deforming each shape to the common template space. Moreover, the uncertainty of dense correspondence performance between two points $p \in P$ and $q \in Q$ can be measured as below: + +$$ +u (p, q) = 1 - \exp (- \gamma \| (p + \Delta p) - (q + \Delta q) \| _ {2} ^ {2}). \tag {19} +$$ + +$\gamma$ is a hyperparameter for modulation and $\Delta p, \Delta q$ are output from $\mathcal{D}_{\theta_1}$ . Note that since each part deformation prior $e$ is shared based on part semantics across entire training shapes, we do not need a test-time optimization on $e$ for an unseen shape at inference time. + +# 5. Experiments + +We demonstrate the effectiveness of our proposed framework and provide both quantitative and qualitative results. In Section 5.1, we conduct multiple label transfer + +Table 2. Keypoint correspondence accuracy in KeypointNet [55]. PCK scores with threshold of $0.05 / 0.1$ are reported. + +
PCKairplanecarchairtable
AtlasNetV2 [34]23.9 / 45.333.5 / 64.025.6 / 50.225.3 / 48.0
DIT [2]13.6 / 34.258.9 / 87.733.3 / 66.726.2 / 55.5
DIF [1]16.5 / 40.567.8 / 90.434.3 / 57.637.9 / 62.9
Ours24.2 / 52.369.3 / 91.861.2 / 83.250.8/77.9
+ +Table 3. 5-shot part label transfer in four categories of ShapeNet. The reported numbers are overall mIoU. + +
modelairplanecarchairtable
AtlasNetV2 [34]57.859.366.970.6
DIT [1]64.266.180.486.0
DIF [2]73.969.682.186.6
Ours75.170.484.286.9
+ +tasks for dense correspondence evaluation. In Section 5.2, we measure shape reconstruction accuracy to evaluate the representational power for unknown shapes. In Section 5.3, we present advanced dense correspondence analysis under more challenging settings. Lastly, in Section 5.4, we perform ablation studies to analyze the proposed regularization and the sensitivity over feature extractor [31]. + +Datasets and Baselines. We train and validate our model and baselines on four categories (e.g., chair, table, airplane and car) of ShapeNetV2 [50]. We follow the training split of DIF [2] for each category. For label transfer tasks, we use the labels from ShapeNet-Part [56] and KeypointNet [55]. DIT [1] and DIF [2] are the most relevant baseline models as they both globally learn implicit template fields, and we select AtlasNetV2 [34] as an additional baseline. Dense correspondence is measured with local patch composition in [34]. For more implementation details, analyses, and visualizations, see the supplement. + +# 5.1. Dense correspondence + +To evaluate the quality of learned dense correspondence, we conduct various surrogate tasks including keypoint transfer, part label transfer, and texture transfer due to + +![](images/d992a9a2429a7e8c487f80fb2c12c05698fa643a016b4a70d35d2eca80ee5f3e.jpg) +Figure 7. Texture transfer result on ShapeNet. We transfer part segment labels from source shapes (colored in ivory) to target shapes and compare ours with DIT [1] and DIF [2]. For additional visualizations, please refer to the supplement. + +Table 4. Reconstruction accuracy for unseen shapes in four categories of ShapeNet. The reported numbers are chamfer distance (CD) between ground truth and reconstructed meshes at a resolution of $256^{3}$ . + +
CD (×103)↓airplanecarchairtable
DIT [1]0.22610.40820.30630.6187
DIF [2]0.20930.31570.42290.2640
Ours0.18590.31500.22650.2796
+ +the lack of labeled datasets for 3D dense correspondence. For the following tasks, we transfer respective labels from source shapes to target shapes in a common template space. + +Keypoint label transfer. Table 2 presents the quantitative result in a 5-shot keypoint transfer task. We measure the performance by a percentage of correct keypoints (PCK) [57]. We compute the distance between the transferred keypoints and the ground truth points to record the PCK score with the threshold of $0.05 / 0.1$ . Our method achieves the best performance in all four categories with a significant improvement. Especially for categories with high structural variability (e.g., chair and table), a large performance gain is observed. For instance, even compared to the best baseline, on the chair category with the threshold of 0.1, our method (83.2) outperforms DIT [1] (66.7) by $16.5\%$ , and on the table category with the threshold of 0.1, our method (77.9) outperforms DIF [2] (62.9) by $15\%$ . We additionally provide the visualization of keypoint transfer in Figure 4, which also support the importance of understanding part semantics for shape correspondence. In particular, two keypoints of pea-green and orange from the source chair are geometrically close but semantically different since one is for the arm part and the other is for the seat part. Among several methods, only our framework succeeds in disentangling two keypoints in the target chair. We provide more qualitative results regarding keypoint transfer results in real-world data [53] to validate the robustness of our method in the supplement. + +Part label transfer. For few-shot part label transfer tasks, we transfer part segment labels from source shapes to target shapes. In detail, for each (target) point of the deformed target shape, we find the $n$ nearest points of the deformed source shapes and perform weighted voting using their labels and distances to the target point. In our experiments, we used $n = 10$ nearest points. For a 5-shot part label transfer task in Table 3, five source shapes were used just the same as [2] for every category. Table 3 shows our framework achieves the best performance over all four categories. Moreover, Figure 6 illustrates the qualitative results of one-shot part label transfer on ShapeNet [50]. Similar to keypoint transfer, Figure 6 also proves the effectiveness of semantic knowledge in dense correspondence. + +Texture transfer. Figure 7 shows texture transfer results on ShapeNetV2 [50] to compare established dense correspondence between baselines and ours with higher-frequency label transfer. We observe notable texture transfer results. For instance, our framework successfully transfers the source car's texture of the windshield (gray) to the corresponding part of the target cars, whereas baselines transfer the same textures to the roof part of the target cars. + +# 5.2. Shape reconstruction + +We report the reconstruction quality of unseen shapes with Chamfer Distance (CD) to validate the representational power of our framework in Table 4. A smaller CD means more accurate reconstruction. As in previous work, we report the $\mathrm{CD} \times 10^{3}$ for readability. Also, we consistently discuss the CD and performance gain below at this scale. We optimize the corresponding shape latent code for a given unseen shape. We measure CD as in DIF [2], where reconstructed meshes are extracted at a resolution of $256^{3}$ , and CD is calculated with $10K$ sampled points. Our framework on average achieved the best reconstruction results surpassing DIF by a margin of 0.0512. On average, our method and DIF exhibit 0.2518 and 0.3030 reconstruction errors, respectively. In three out of four categories, our method achieved the best reconstruction accuracy. + +# 5.3. More challenging setups + +To further validate the effectiveness of our method, we perform additional experiments and compare the results with DIF [2]. Overall, the results below imply current methods that lack part semantics and do not consider part deformation consistency may fail to learn semantically plausible deformation in challenging cases. + +2-shot part label transfer within different subcategories. Our objective is to assess dense correspondence performance under difficult settings. We conduct 2-shot part label transfer within two different subcategories belonging to the same category, where shapes from different subcategories tend to have different geometry for same semantics. Since ShapeNet [50] includes incorrect subcategory labels [58], we manually cleaned subcategory labels for four categories (e.g., airplane: airliner/jet, car: sedan/jeep, chair: straight chair/sofa, table: pedestal table/short table). Figure 8 displays examples of shapes from each subcategory. For robust evaluations, we randomly select the source shape 5 times for each subcategory. Table 5 shows our framework achieves a significant performance gain in all cases by over $7.2\%$ on average compared to DIF [2]. + +Unsupervised dense correspondence in DFAUST [59]. Furthermore, we evaluate unsupervised dense correspondence with non-rigid shapes [59]. For more challenging settings, we train both DIF [2] and our model for two different subjects with different sizes and heights (e.g., subject ID of 50002, 50026) and four distinct actions (e.g., shake arms, chicken wings, running on spot, and jumping jacks). Colormap in Figure 9 demonstrates the correspondence quality, where we transfer color over the up-axis. In the source shape, the arms are labeled in red. After transferring these color labels to the target shapes, our method preserves the red/orange labeling of the arms across different motions, whereas the baseline fails to capture the same labeling (i.e., yellow hands) since it lacks semantic knowledge during training. Additionally, we conduct keypoint transfer tasks and measure the PCK score as identical as in Section 5.1. Our model achieves 74.9/85.2 with a threshold of 0.05/0.1, whereas DIF [2] achieves 67.0/80.7 with the same threshold, showing our superior performance in the challenging settings (see the supplement for the full keypoint transfer result). Note that we can also learn implicit templates for non-rigid shapes with supervision as in [19]. We conduct experiments in unsupervised settings since our focus is to validate the effectiveness of utilizing additional semantic information for template learning. + +# 5.4. Ablations + +Ablation on regularizations. We conduct ablation studies regarding the proposed regularizations including part deformation consistency regularizations, global scale consistency regularization, and global deformation consistency + +![](images/6cd10adf61d83a66ca80750a8786e2ac526e6ea83495786caa65bca3c3b03ef5.jpg) +Figure 8. Types of source shapes for 2-shot part label transfer. + +Table 5. 2-shot part label transfer in subcategories of ShapeNet. The reported numbers are overall mIoU over 5 runs. + +
subcategoriesDIF [2]Ours
airplane(airliner/jet)64.5±1.6172.8±0.99 (+8.3↑)
car(sedan/jeep)65.1±2.3469.1±1.46 (+4.0↑)
chair(straight/sofa)66.3±6.0677.5±5.81 (+11.2↑)
table(pedestal/short)74.4±10.179.7±6.26 (+5.3↑)
+ +![](images/87a877925235bd64d7f0802809ba8f9a944ff6a557b07ab22f5af790ef6063d1.jpg) +Figure 9. Color transfer on DFAUST [59]. We transfer different color labels over the up-axis from source shapes (colored in ivory) to target shapes and compare ours with DIF [2]. + +Table 6. Ablations for every proposed regularization. We report mIoU for 5-shot part label transfer on ShapeNet chair. Note that $\mathcal{L}_{\mathrm{pdc}}$ denotes $\mathcal{L}_{\mathrm{pdc\_geo}} + \mathcal{L}_{\mathrm{pdc\_sem}}$ . + +
LpdcLgeoLscalechair
(a)✓83.4
(b)✓✓83.9
(c)✓✓✓84.2
+ +Table 7. BAE-Net $k$ ablation. We report mIoU for 5-shot part label transfer on ShapeNet chair over 3 runs. + +
k=2k=4k=6k=8k=10
chair83.8±0.5183.5±1.4882.8±1.2783.6±0.8583.7±1.33
+ +regularization. We conduct 5-shot part label transfer within a chair category, as same as Section 5.1. Table 6 shows that each component of our proposed regularizers contributes to performance improvement for dense correspondence. Note that by only using $\mathcal{L}_{\mathrm{pdc}}$ , i.e., using $\mathcal{L}_{\mathrm{pdc\_geo}}$ and $\mathcal{L}_{\mathrm{pdc\_sem}}$ , it still outperforms than baselines (80.4 [1] and 82.1 [2]). The best performance was achieved by incorporating all proposed regularizations. + +BAE-Net [31] with different $k$ . We evaluate the effect of $k$ for BAE-Net [31] with mean/standard deviation over 3 runs for ShapeNet [50] chair in We train BAE-Net [31] with a range of $k = (2,4,6,8,10)$ , where $k = 4$ is used in the main paper, and report mIoU for 5-shot part label + +transfer task. Table 7 shows the efficacy of semantic priors, where our framework achieves the average mIoU of 83.5, which shows higher performance than baselines. More importantly, we observed that the performance is fairly robust over $k$ . + +# 6. Conclusion + +We have proposed a new framework that learns a semantic-aware implicit template field. Our framework extracts semantic knowledge from a self-supervised feature extractor given unlabelled point clouds. By leveraging the semantic information, we propose novel conditioning with semantic-aware deformation code and part deformation consistency regularizations to encourage semantically plausible deformation. Furthermore, we suggest global scaling consistency and global deformation consistency regularizations to enhance the learning process. Through quantitative and qualitative evaluation, we demonstrate the superiority of our proposed method and verified the effect of all proposed components in our framework. + +Acknowledgments. This research was supported in part by the MSIT (Ministry of Science and ICT), Korea, under the ICT Creative Consilience program (IITP-2023-2020-0-01819) supervised by the IITP (Institute for Information & communications Technology Planning & Evaluation); the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF-2023R1A2C2005373); and Google Cloud Research Credits program with the award (GCLF-67WL-WU6D-MF14, V9P3-99XD-64CL-EH5H). + +# References + +[1] Zerong Zheng, Tao Yu, Qionghai Dai, and Yebin Liu. Deep implicit templates for 3d shape representation. In CVPR, 2021. 1, 2, 3, 5, 6, 7, 8 +[2] Yu Deng, Jiaolong Yang, and Xin Tong. Deformed implicit field: Modeling 3d shapes with learned dense correspondence. In CVPR, 2021. 1, 2, 3, 5, 6, 7, 8 +[3] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: a skinned multiperson linear model. TOG, 2015. 1, 2 +[4] Brett Allen, Brian Curless, and Zoran Popovic. Articulated body deformation from range scan data. TOG, 2002. 1 +[5] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan C. Russell, and Mathieu Aubry. 3d-coded: 3d correspondences by deep deformation. In ECCV, 2018. 1, 2 +[6] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In CVPR, 2018. 1 +[7] Wang Zeng, Wanli Ouyang, Ping Luo, Wentao Liu, and Xiaogang Wang. 3d human mesh regression with dense correspondence. In CVPR, 2020. 1, 2 + +[8] Sing Chun Lee and Misha Kazhdan. Dense point-to-point correspondences between genus-zero shapes. In Computer Graphics Forum, 2019. 1, 2 +[9] Silvia Zuffi, Angjoo Kanazawa, David W. Jacobs, and Michael J. Black. 3d menagerie: Modeling the 3d shape and pose of animals. In CVPR, 2017. 1, 2 +[10] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In AVSS, 2009. 1 +[11] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. TOG, 2017. 1 +[12] Luca Ballan, Aparna Taneja, Jürgen Gall, Luc Van Gool, and Marc Pollefeys. Motion capture of hands in action using discriminative salient points. In ECCV, 2012. 1 +[13] Srinath Sridhar, Antti Oulasvirta, and Christian Theobalt. Interactive markerless articulated hand motion tracking using rgb and depth data. In ICCV, 2013. 1 +[14] Iason Oikonomidis, Manolis IA Lourakis, and Antonis A Argyros. Evolutionary quasi-random search for hand articulations tracking. In CVPR, 2014. 1 +[15] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In CVPR, 2017. 1, 2 +[16] Vignesh Ganapathi-Subramanian, Olga Diamanti, Soeren Pirk, Chengcheng Tang, Matthias Niessner, and Leonidas Guibas. Parsing geometry using structure-aware shape templates. In 3DV, 2018. 1, 2 +[17] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan Russell, and Mathieu Aubry. AtlasNet: A Papier-Mâché Approach to Learning 3D Surface Generation. In CVPR, 2018. 1, 2 +[18] Pablo Palafox, Nikolaos Sarafianos, Tony Tung, and Angela Dai. Spams: Structured implicit parametric models. In CVPR, 2022. 1 +[19] Ramana Sundararaman, Gautam Pai, and Maks Ovsjanikov. Implicit field supervision for robust non-rigid shape matching. In ECCV, 2022. 1, 8 +[20] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, 2022. 1, 2 +[21] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. 2019. 1, 3 +[22] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 1, 2, 3 +[23] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 1, 2, 3 + +[24] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In ECCV, 2020. 1 +[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2 +[26] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 1, 2 +[27] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In CVPR, 2021. 1, 2 +[28] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, 2019. 1, 2 +[29] Tarun Yenamandra, Ayush Tewari, Florian Bernard, Hans-Peter Seidel, Mohamed Elgharib, Daniel Cremers, and Christian Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In CVPR, 2021. 1, 2 +[30] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR, 2020. 1, 2 +[31] Zhiqin Chen, Kangxue Yin, Matthew Fisher, Siddhartha Chaudhuri, and Hao Zhang. Bae-net: Branched autoencoder for shape co-segmentation. In ICCV, 2019. 2, 4, 6, 8 +[32] Vladimir G Kim, Wilmot Li, Niloy J Mitra, Siddhartha Chaudhuri, Stephen DiVerdi, and Thomas Funkhouser. Learning part-based templates from large collections of 3d shapes. TOG, 2013. 2 +[33] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T. Freeman, and Thomas A. Funkhouser. Learning shape templates with structured implicit functions. In ICCV, 2019. 2 +[34] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir Kim, Bryan Russell, and Mathieu Aubry. Learning elementary structures for 3d shape generation and matching. In NeurIPS, 2019. 2, 6 +[35] Jiahui Lei and Kostas Daniilidis. Cadex: Learning canonical deformation coordinate space for dynamic surface representation via neural homeomorphism. In CVPR, 2022. 2 +[36] Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural template: Topology-aware reconstruction and disentangled generation of 3d meshes. In CVPR, 2022. 2 +[37] Zekun Hao, Hadar Averbuch-Elor, Noah Snavely, and Serge Belongie. Dualsdf: Semantic shape manipulation using a two-level representation. In CVPR, 2020. 2 +[38] Yueqi Duan, Haidong Zhu, He Wang, Li Yi, Ram Nevatia, and Leonidas J. Guibas. Curriculum deepsdf. In ECCV, 2020. 2 + +[39] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In ICLR, 2022. 2 +[40] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2 +[41] Michael Niemeyer and Andreas Geiger. Campari: Camera-aware decomposed generative neural radiance fields. In 3DV, 2021. 2 +[42] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, et al. Gaudi: A neural architect for immersive 3d scene generation. In NeurIPS, 2022. 2 +[43] Chengjie Niu, Manyi Li, Kai Xu, and Hao Zhang. Rimnet: Recursive implicit fields for unsupervised learning of hierarchical shape structures. In CVPR, 2022. 2 +[44] Chenyang Zhu, Kai Xu, Siddhartha Chaudhuri, Li Yi, Leonidas J Guibas, and Hao Zhang. Adacoseg: Adaptive shape co-segmentation with group consistency loss. In CVPR, 2020. 2 +[45] Minghua Liu, Yinhao Zhu, Hong Cai, Shizhong Han, Zhan Ling, Fatih Porikli, and Hao Su. Part slip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In CVPR, 2023. 2 +[46] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2 +[47] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 2 +[48] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. In NeurIPS, 2022. 2 +[49] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In CVPR, 2022. 2 +[50] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv:1512.03012, 2015. 4, 6, 7, 8 +[51] David Ha, Andrew M. Dai, and Quoc V. Le. Hypernetworks. In ICLR, 2017. 4 +[52] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In NeurIPS, 2020. 4 +[53] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 5, 7 + +[54] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 5 +[55] Yang You, Yujing Lou, Chengkun Li, Zhoujun Cheng, Liangwei Li, Lizhuang Ma, Cewu Lu, and Weiming Wang. Keypointnet: A large-scale 3d keypoint dataset aggregated from numerous human annotations. In CVPR, 2020. 6 +[56] Li Yi, Vladimir G. Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas J. Guibas. A scalable active framework for region annotation in 3d shape collections. TOG, 2016. 6 +[57] Li Yi, Hao Su, Xingwen Guo, and Leonidas Guibas. Syncspeccnn: Synchronized spectral cnn for 3d shape segmentation. In CVPR, 2017. 7 +[58] Xinhai Liu, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Fine-grained 3d shape classification with hierarchical part-view attention. IEEE Transactions on Image Processing, 2021. 8 +[59] Federica Bogo, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Dynamic FAUST: Registering human bodies in motion. In CVPR, 2017. 8 \ No newline at end of file diff --git a/2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/images.zip b/2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..857b7eb6a9dc519cb109eea7b4cfdebc3b602bdb --- /dev/null +++ b/2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d19d7439e6b01d2d6515880d933d28b0044d57274205d157f166d02f3018aa0e +size 603754 diff --git a/2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/layout.json b/2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f45e6f8b08f6563f370dd03be384662325ef2a9f --- /dev/null +++ b/2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/layout.json @@ -0,0 +1,10473 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "type": "text", + "content": "Semantic-Aware Implicit Template Learning via Part Deformation Consistency" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 176, + 143, + 414, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 143, + 414, + 172 + ], + "spans": [ + { + "bbox": [ + 176, + 143, + 414, + 172 + ], + "type": "text", + "content": "Sihyeon Kim Minseok Joo Jaewon Lee \nJuyeon Ko Juhan Cha Hyunwoo J. Kim*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 180, + 463, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 180, + 463, + 194 + ], + "spans": [ + { + "bbox": [ + 129, + 180, + 463, + 194 + ], + "type": "text", + "content": "Department of Computer Science and Engineering, Korea University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 103, + 196, + 490, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 196, + 490, + 208 + ], + "spans": [ + { + "bbox": [ + 103, + 196, + 490, + 208 + ], + "type": "text", + "content": "{sh_bs15, wlgkcjf87, 2jlejyu, juyon98, hanchaa, hyunwoojkim}@korea.ac.kr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 236, + 192, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 236, + 192, + 248 + ], + "spans": [ + { + "bbox": [ + 143, + 236, + 192, + 248 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 260, + 290, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 260, + 290, + 514 + ], + "spans": [ + { + "bbox": [ + 45, + 260, + 290, + 514 + ], + "type": "text", + "content": "Learning implicit templates as neural fields has recently shown impressive performance in unsupervised shape correspondence. Despite the success, we observe current approaches, which solely rely on geometric information, often learn suboptimal deformation across generic object shapes, which have high structural variability. In this paper, we highlight the importance of part deformation consistency and propose a semantic-aware implicit template learning framework to enable semantically plausible deformation. By leveraging semantic prior from a self-supervised feature extractor, we suggest local conditioning with novel semantic-aware deformation code and deformation consistency regularizations regarding part deformation, global deformation, and global scaling. Our extensive experiments demonstrate the superiority of the proposed method over baselines in various tasks: keypoint transfer, part label transfer, and texture transfer. More interestingly, our framework shows a larger performance gain under more challenging settings. We also provide qualitative analyses to validate the effectiveness of semantic-aware deformation. The code is available at https://github.com/mlvlab/PDC." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 520, + 128, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 128, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 128, + 532 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 540, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 288, + 696 + ], + "type": "text", + "content": "Template learning is essential for shape analysis since the template, which is a compact representation of arbitrary shapes, provides a good prior for shape fitting so that we can establish dense correspondence and transfer key attributes such as texture or keypoints between shapes. Traditionally, handcrafted templates such as meshes, part elements with primitive shapes (e.g., sphere, cylinder, cone), or parametric models have been widely used for human/animal body [3, 4, 5, 6, 7, 8, 9], human face [10, 11], human hand [12, 13, 14], and generic objects [15, 16, 17]. Defining templates for less complex shapes with consistent topologies and similar structures is relatively straightforward. However, for shapes like generic object shapes with high structural variability or" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 236, + 547, + 406 + ], + "blocks": [ + { + "bbox": [ + 309, + 236, + 547, + 406 + ], + "lines": [ + { + "bbox": [ + 309, + 236, + 547, + 406 + ], + "spans": [ + { + "bbox": [ + 309, + 236, + 547, + 406 + ], + "type": "image", + "image_path": "b8c7719d3d700afb1d874be71a8a7714a1fb5c826244ecbfeb71ba0467281d7f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 308, + 407, + 547, + 495 + ], + "blocks": [ + { + "bbox": [ + 308, + 407, + 547, + 495 + ], + "lines": [ + { + "bbox": [ + 308, + 407, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 308, + 407, + 547, + 495 + ], + "type": "image", + "image_path": "068252e044fa8ddddc802a3c0d2b98ea36759befa0f5a17a5bcffdd3449bdcb3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 501, + 547, + 589 + ], + "lines": [ + { + "bbox": [ + 304, + 501, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 547, + 589 + ], + "type": "text", + "content": "Figure 1. Importance of part deformation consistency. Given input shapes with high structural variability, only considering low-level geometry during template learning [1, 2] can lead to suboptimal deformation (e.g., cannot distinguish between seat and arm where two parts are geometrically close while semantically distant: low seat-low arm, high seat-high arm). Ours succeeds in transferring shape attributes by encouraging part deformation consistency with semantic information." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 616 + ], + "type": "text", + "content": "non-rigid shapes with large-scale deformations, manually defining suitable templates becomes challenging." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": "To address this challenge, recent works [1, 2, 18, 19] have focused on learning implicit templates based on neural fields [20]. Neural fields, which are known for strong representation power, efficiency, and flexibility, have been successful in multiple 3D vision tasks, e.g., shape/scene reconstruction [21, 22, 23, 24], neural rendering [25, 26, 27], and human digitization [28, 29, 30]. Likewise, neural fields-based template learning have shown superior perfor" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "593" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "mance on dense correspondence between complex shapes. Both [1] and [2] suggest decomposing the implicit representation of shapes into a deformation field and a template field. Here, the template is learned as a continuous function, which is expected to capture shared structures among the given shapes within a category (e.g., airplane, chair), while the deformation depends solely on the conditioned geometry of each shape (i.e., SDF value and surface normal). We observe current approaches often learn suboptimal deformation especially when a shape category has high structural variability like diverse local deformation scales and unseen/missing part structures; see Figure 1. That is, incorporating extra knowledge of part semantics is necessary to handle such cases." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 243, + 289, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 289, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 289, + 590 + ], + "type": "text", + "content": "In this paper, we propose a novel framework to learn semantically plausible deformation by distilling the knowledge of self-supervised feature extractor [31] and employing it as a semantic prior to satisfy part deformation consistency during the template learning process. Part deformation consistency is our inductive bias that the deformation result within the same part should be consistent. By leveraging semantic prior to understand part arrangements of shapes, we suggest local conditioning with semantic-aware deformation code and carefully designed regularizations to encourage semantically corresponding deformation. Semantic-aware deformation code is a point-wise soft assignment of part deformation priors, where part deformation priors are multiple latent codes to encode information on corresponding part deformation. We calculate assigned weights considering the part semantics of each point. In addition, we propose input space regularization and latent space regularization to encourage part-level distances to be close, providing an explicit way to control the deformation field so that the model can learn flexible deformation and successfully learn common structures under diverse shapes. Lastly, we suggest global scale consistency regularization to preserve the volume of the template against large-scale deformations. To validate the effectiveness of our framework, we conduct extensive surrogate experiments reflecting different levels of correspondence such as keypoint transfer, part label transfer, and texture transfer, and demonstrate competitive performance over baselines with qualitative analyses." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Our contributions are summarized as follows: ① We propose a new framework that learns global implicit template fields founded on part deformation consistency while understanding part semantics to enable semantically plausible deformation, ② We impose hybrid conditioning by combining global conditioning of geometry and local conditioning of part semantics for flexible deformation, ③ We design novel regularizations to successfully manipulate the template fields, ④ We show the effectiveness of our framework with both qualitative and quantitative experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 71, + 398, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 398, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 398, + 84 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 103, + 547, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 103, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 103, + 547, + 344 + ], + "type": "text", + "content": "Template learning for shape correspondence. Defining a template for a set of arbitrary shapes has been studied for a long time. A high-quality template provides a good initialization for shape alignment and various applications in vision and graphics. For shapes with consistent topologies and structures such as faces or bodies of humans and animals, handcrafted templates have been popular [3, 5, 7, 8, 9] and still widely used today. On the other hand, it was more challenging for generic object shapes with inconsistent topology and diverse part structures. Prior works suggest to predefine or learn local templates from data as in primitive parts [15] or part elements [32, 16, 33, 17, 34] to handle the inconsistent structure. Recent works have tried to learn a canonical shape of generic objects through implicit template field [1, 2, 35, 36], which are proven to be effective for representing shapes. However, existing approaches mostly lack explicit consideration of semantic knowledge, which enables semantically plausible deformation so that the learned template captures a common structure even with the various structured shapes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 354, + 547, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 536 + ], + "type": "text", + "content": "Neural fields for 3D representations. Recent studies have shown that representing 3D shapes or scenes as implicit representations with a neural network has several advantages: powerful representation, flexibility, and memory efficiency. These are called neural fields [20]. For instance, Occupancy networks [22] proposed a method to represent 3D geometry as a continuous occupancy function. DeepSDF [23] learns 3D complex surfaces with a neural network of auto-decoder framework for 3D modeling. Further, implicit representation has been introduced to various applications [37, 38]. To leverage the advantages, many applications based on neural fields have emerged such as neural rendering [25, 26, 27], human digitization [28, 29, 30], and generative modeling [39, 40, 41, 42]. They have also been applied to learn the template of complex shapes as mentioned above." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "Utilizing self-supervised segmentation models. Recently, there has been a growing body of research on analyzing complex shapes in a self-supervised manner due to the high cost of annotation for 3D data. Unsupervised semantic segmentation/co-segmentation models like BAE-Net [31], RIM-Net [43], AdaCoSeg [44], and PartSLIP [45] have demonstrated the ability to comprehend part semantics without ground truth labels. For 2D and multi-modal tasks, using large-scale self-supervised segmentation models like DINO [46] as a feature extractor has been widely adopted, including pixelNeRF [47], Feature Field [48], and ClipN-erF [49]. Inspired by these recent works, we utilize pretrained unsupervised part segmentation models to harness part semantic information for template learning." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "594" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 171, + 192 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 171, + 192 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 171, + 192 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 171, + 192 + ], + "type": "image", + "image_path": "2c8c62fcf1b166e69f5abf8d598351ffd928a7e37ec0a0d5fc4bebd571ae0dd2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 214, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 214, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 214, + 547, + 258 + ], + "type": "text", + "content": "Figure 2. Framework of semantic-aware implicit template learning. To learn semantic-aware implicit template field, (a) we extract part semantics from pretrained feature extractor " + }, + { + "bbox": [ + 46, + 214, + 547, + 258 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 214, + 547, + 258 + ], + "type": "text", + "content": " and incorporate knowledge into semantic-aware deformation code " + }, + { + "bbox": [ + 46, + 214, + 547, + 258 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 46, + 214, + 547, + 258 + ], + "type": "text", + "content": " and (b) part deformation consistency regularizations. With optimizing with the reconstruction loss " + }, + { + "bbox": [ + 46, + 214, + 547, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 46, + 214, + 547, + 258 + ], + "type": "text", + "content": " and several regularizers, our framework succeeds to learn semantically plausible deformation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 340, + 192 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 340, + 192 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 340, + 192 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 340, + 192 + ], + "type": "image", + "image_path": "ee7a919c08e58d123d46598646c37fef17ae02e62c65416039063fcee5aef637.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 197, + 238, + 207 + ], + "lines": [ + { + "bbox": [ + 227, + 197, + 238, + 207 + ], + "spans": [ + { + "bbox": [ + 227, + 197, + 238, + 207 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 342, + 68, + 419, + 191 + ], + "blocks": [ + { + "bbox": [ + 342, + 68, + 419, + 191 + ], + "lines": [ + { + "bbox": [ + 342, + 68, + 419, + 191 + ], + "spans": [ + { + "bbox": [ + 342, + 68, + 419, + 191 + ], + "type": "image", + "image_path": "f523e7f3c92f502460306e545568713d7bd79a9809d81f4cd9455b4a4c8abcb5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 426, + 68, + 544, + 192 + ], + "blocks": [ + { + "bbox": [ + 426, + 68, + 544, + 192 + ], + "lines": [ + { + "bbox": [ + 426, + 68, + 544, + 192 + ], + "spans": [ + { + "bbox": [ + 426, + 68, + 544, + 192 + ], + "type": "image", + "image_path": "801ade73fef454d3add5f9bef7712229db9601be40b7f48707a00ebd1f041aa5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 197, + 491, + 207 + ], + "lines": [ + { + "bbox": [ + 479, + 197, + 491, + 207 + ], + "spans": [ + { + "bbox": [ + 479, + 197, + 491, + 207 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 267, + 132, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 267, + 132, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 267, + 132, + 280 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 287, + 287, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 287, + 347 + ], + "type": "text", + "content": "Implicit templates have been studied with neural fields. Given a 3D shape " + }, + { + "bbox": [ + 46, + 287, + 287, + 347 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 287, + 287, + 347 + ], + "type": "text", + "content": ", a neural field, which is a neural network " + }, + { + "bbox": [ + 46, + 287, + 287, + 347 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 46, + 287, + 287, + 347 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 46, + 287, + 287, + 347 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 287, + 287, + 347 + ], + "type": "text", + "content": ", approximates continuous functions such as an occupancy function [21, 22] or a Signed Distance Function (SDF) [23] as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 357, + 287, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 357, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 133, + 357, + 287, + 369 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\theta} (\\boldsymbol {x}; \\boldsymbol {z} _ {P}) = s. \\tag {1}", + "image_path": "effc6ecf9acba9455bbf11240d3c5bb32892ad3680834144cca9608227def40e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 377, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 377, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 377, + 287, + 437 + ], + "type": "text", + "content": "Herein, " + }, + { + "bbox": [ + 47, + 377, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 47, + 377, + 287, + 437 + ], + "type": "text", + "content": " is a query point and " + }, + { + "bbox": [ + 47, + 377, + 287, + 437 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 377, + 287, + 437 + ], + "type": "text", + "content": " is an occupancy/SDF function value. To represent multiple 3D shapes, for each shape, a shape latent code " + }, + { + "bbox": [ + 47, + 377, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\pmb{z}_P \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 47, + 377, + 287, + 437 + ], + "type": "text", + "content": " is learned. In general, a surface is defined as the zero-level set of a SDF, i.e., " + }, + { + "bbox": [ + 47, + 377, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\{x | \\mathcal{F}_{\\theta}(x; z_P) = 0\\}" + }, + { + "bbox": [ + 47, + 377, + 287, + 437 + ], + "type": "text", + "content": " as in DeepSDF [23]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 437, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 461 + ], + "type": "text", + "content": "The neural field can be further decomposed into an implicit template field " + }, + { + "bbox": [ + 46, + 437, + 287, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 437, + 287, + 461 + ], + "type": "text", + "content": " and a deformation field " + }, + { + "bbox": [ + 46, + 437, + 287, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{D}[1,2]" + }, + { + "bbox": [ + 46, + 437, + 287, + 461 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 101, + 470, + 287, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 470, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 101, + 470, + 287, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\theta} (\\boldsymbol {x}; \\boldsymbol {z} _ {P}) = \\mathcal {T} _ {\\theta_ {2}} \\left(\\mathcal {D} _ {\\theta_ {1}} (\\boldsymbol {x}; \\boldsymbol {z} _ {P})\\right). \\tag {2}", + "image_path": "4e24e84f95bdc4be4c6ca83898a5fea05e1d79f3d66af82cc8bc6aacb8452b91.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "type": "text", + "content": "The implicit template field " + }, + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\theta_2}" + }, + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "type": "text", + "content": " captures a common structure and outputs an occupancy/SDF value " + }, + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "type": "text", + "content": " in the template space. " + }, + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\theta_1}" + }, + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "type": "text", + "content": " maps a point " + }, + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "type": "text", + "content": " from a shape-specific space to the common template space by adding a deformation vector " + }, + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\Delta \\pmb{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\pmb{x} + \\Delta \\pmb{x}" + }, + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "type": "text", + "content": ". In DIF-Net [2], " + }, + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\theta_1}" + }, + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "type": "text", + "content": " additionally predicts a correction value " + }, + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\Delta s \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "type": "text", + "content": " to capture instance-specific local structures. In this case, the neural field " + }, + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 46, + 491, + 287, + 586 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 96, + 594, + 287, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 594, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 96, + 594, + 287, + 613 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\theta} (\\boldsymbol {x}; \\boldsymbol {z} _ {P}) = \\mathcal {T} _ {\\theta_ {2}} (\\boldsymbol {x} + \\Delta \\boldsymbol {x}) + \\Delta s \\tag {3}", + "image_path": "221700e030793692edf24e58ecc3bbcdc6d0725cf22a08bf51d9c2cb07e4e672.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 109, + 609, + 206, + 621 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 609, + 206, + 621 + ], + "spans": [ + { + "bbox": [ + 109, + 609, + 206, + 621 + ], + "type": "interline_equation", + "content": "\\Delta s, \\Delta \\boldsymbol {x} = \\mathcal {D} _ {\\theta_ {1}} (\\boldsymbol {x}; \\boldsymbol {z} _ {P}).", + "image_path": "b7297572b4ee1b32ad4aa5c162473828bce2458c65703db9dc3473eac50505e6.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "The existing approaches [1, 2] learn the implicit template field by minimizing the distance between the template and the deformed individual shapes. The distance is measured by SDF values and surface normals. Also, several regularizers are imposed such as deformation smoothness (e.g., penalizing the spatial gradient of " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\theta_1}" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": " and restricting the deformation scale between points). However, the loss functions" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 268, + 545, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 340 + ], + "type": "text", + "content": "and regularizers are computed by low-level geometric information. They are not effective enough to handle generic objects with large structural variability. Hence, in this paper, we take part semantics into account for template learning to achieve 'semantically plausible' deformation and dense correspondence." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 350, + 361, + 362 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 350, + 361, + 362 + ], + "spans": [ + { + "bbox": [ + 306, + 350, + 361, + 362 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 371, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 545, + 467 + ], + "type": "text", + "content": "We present a semantic-aware implicit template learning framework based on part deformation consistency. In this section, we first delineate the overall architecture of our method equipped with part deformation priors and semantic-aware deformation code. Then, we introduce our deformation consistency regularizers including part deformation consistency regularizations. Lastly, we discuss the training procedure with several additional loss functions." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 473, + 509, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 473, + 509, + 487 + ], + "spans": [ + { + "bbox": [ + 305, + 473, + 509, + 487 + ], + "type": "text", + "content": "4.1. Semantic-aware implicit template field" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 492, + 545, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 545, + 528 + ], + "type": "text", + "content": "We here present a novel semantic-aware implicit template learning framework. The overall pipeline of the proposed method is shown in Figure 2(a)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 528, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 528, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 528, + 545, + 622 + ], + "type": "text", + "content": "Template field and deformation field. We extend the implicit template learning by incorporating semantic information into the deformation field. Specifically, given a query point " + }, + { + "bbox": [ + 304, + 528, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 528, + 545, + 622 + ], + "type": "text", + "content": " of a shape " + }, + { + "bbox": [ + 304, + 528, + 545, + 622 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 528, + 545, + 622 + ], + "type": "text", + "content": ", the deformation field " + }, + { + "bbox": [ + 304, + 528, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\theta_1}" + }, + { + "bbox": [ + 304, + 528, + 545, + 622 + ], + "type": "text", + "content": " predicts a deformation vector " + }, + { + "bbox": [ + 304, + 528, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\Delta x" + }, + { + "bbox": [ + 304, + 528, + 545, + 622 + ], + "type": "text", + "content": " and a correction value " + }, + { + "bbox": [ + 304, + 528, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\Delta s" + }, + { + "bbox": [ + 304, + 528, + 545, + 622 + ], + "type": "text", + "content": " conditioned on a point-wise semantic-aware deformation code (SDC) " + }, + { + "bbox": [ + 304, + 528, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\alpha(x; P)" + }, + { + "bbox": [ + 304, + 528, + 545, + 622 + ], + "type": "text", + "content": " in addition to a shape latent code " + }, + { + "bbox": [ + 304, + 528, + 545, + 622 + ], + "type": "inline_equation", + "content": "z_P" + }, + { + "bbox": [ + 304, + 528, + 545, + 622 + ], + "type": "text", + "content": ". Then, our framework is defined as" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 348, + 629, + 545, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 629, + 545, + 648 + ], + "spans": [ + { + "bbox": [ + 348, + 629, + 545, + 648 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\theta} (\\boldsymbol {x}; P, \\mathcal {E}) := \\mathcal {T} _ {\\theta_ {2}} (\\boldsymbol {x} + \\Delta \\boldsymbol {x}) + \\Delta s, \\tag {4}", + "image_path": "450894c897357d9d26849202c963bdf236598ee4ee502a4fccbff67575c68152.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 366, + 645, + 501, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 645, + 501, + 658 + ], + "spans": [ + { + "bbox": [ + 366, + 645, + 501, + 658 + ], + "type": "interline_equation", + "content": "\\Delta \\boldsymbol {x}, \\Delta s = \\mathcal {D} _ {\\theta_ {1}} (\\boldsymbol {x}; \\boldsymbol {z} _ {P}, \\boldsymbol {\\alpha} (\\boldsymbol {x}; P)),", + "image_path": "ab61a04e596b4416e919893b784f964b46128fdbb389182dfe676e31f735b216.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 665, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 306, + 665, + 545, + 714 + ], + "type": "text", + "content": " is a pretrained feature extractor. We assign " + }, + { + "bbox": [ + 306, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\alpha (\\pmb {x};P)" + }, + { + "bbox": [ + 306, + 665, + 545, + 714 + ], + "type": "text", + "content": " based on a point-wise semantic feature, which is extracted from " + }, + { + "bbox": [ + 306, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 306, + 665, + 545, + 714 + ], + "type": "text", + "content": ". For simplicity, we denote the neural field " + }, + { + "bbox": [ + 306, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}(\\pmb {x};P,\\mathcal{E})" + }, + { + "bbox": [ + 306, + 665, + 545, + 714 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 306, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}(\\pmb {x})" + }, + { + "bbox": [ + 306, + 665, + 545, + 714 + ], + "type": "text", + "content": " and SDC " + }, + { + "bbox": [ + 306, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\alpha (\\pmb {x};P)" + }, + { + "bbox": [ + 306, + 665, + 545, + 714 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 306, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\alpha (\\pmb {x})" + }, + { + "bbox": [ + 306, + 665, + 545, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "595" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": "Feature extractor. In this paper, for feature extraction, we mainly use BAE-Net [31] as an encoder " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": ", which is proven effective in capturing part semantics. Note that we pretrained BAE-Net on selected categories of ShapeNetV2 [50] by self-supervision without any extra labels or data for a fair comparison. The features from BAE-Net are " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": " dimensional vectors, i.e., " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\boldsymbol{o}^P(\\boldsymbol{x}) = \\mathcal{E}(\\boldsymbol{x}, P) \\in \\mathbb{R}^k" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": " is a hyperparameter, which is closely related to the number of semantic parts. Since the number of semantic parts is unknown in self-supervised learning, we simply chose sufficiently large " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "text", + "content": "Semantic-aware deformation code. Our framework assumes that different semantic parts necessitate different deformations. To achieve this, the proposed method learns a deformation prior for each semantic part. The deformation priors are encoded as latent codes " + }, + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "inline_equation", + "content": "e_1, \\dots, e_k \\in \\mathbb{R}^{d'}" + }, + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "text", + "content": ". A point " + }, + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "text", + "content": " is softly assigned to " + }, + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "text", + "content": " semantic parts using the softmax values of semantic features " + }, + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "inline_equation", + "content": "o^P(x)" + }, + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "text", + "content": ". Combining the soft assignment with deformation priors, semantic-aware deformation code (SDC) denoted as " + }, + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\alpha(x) \\in \\mathbb{R}^{d'}" + }, + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 330, + 287, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 330, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 59, + 330, + 287, + 364 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\alpha} (\\boldsymbol {x}) = \\sum_ {i = 1} ^ {k} \\frac {\\exp \\left(\\boldsymbol {o} _ {i} ^ {P} (\\boldsymbol {x})\\right) \\boldsymbol {e} _ {i}}{\\sum_ {j = 1} ^ {k} \\exp \\left(\\boldsymbol {o} _ {j} ^ {P} (\\boldsymbol {x})\\right)}, \\forall i \\in \\{1, \\dots , k \\}, \\tag {5}", + "image_path": "920fdbc8633901dbc53dbda85c3715a491fb47a4941e482ac1898349f03da9ec.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 371, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 432 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 371, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\pmb{o}_i^P(\\pmb{x})" + }, + { + "bbox": [ + 46, + 371, + 287, + 432 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 371, + 287, + 432 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 371, + 287, + 432 + ], + "type": "text", + "content": "-th dimension value of the semantic features. The SDC at " + }, + { + "bbox": [ + 46, + 371, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 371, + 287, + 432 + ], + "type": "text", + "content": " is computed by a linear combination of deformation priors. We learn the point's probability of belonging to each of the " + }, + { + "bbox": [ + 46, + 371, + 287, + 432 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 371, + 287, + 432 + ], + "type": "text", + "content": " parts, which is served as part semantics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 432, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 287, + 540 + ], + "type": "text", + "content": "Remarks. For comparison, we also applied hard assignment of part deformation priors for SDC, i.e., " + }, + { + "bbox": [ + 46, + 432, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\alpha(\\pmb{x}) = e_{i^*}" + }, + { + "bbox": [ + 46, + 432, + 287, + 540 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 432, + 287, + 540 + ], + "type": "inline_equation", + "content": "i^* = \\operatorname{argmax}_i (\\pmb{o}_i^P(\\pmb{x}))" + }, + { + "bbox": [ + 46, + 432, + 287, + 540 + ], + "type": "text", + "content": ". However, as points are completely separated with local conditioning of the corresponding part deformation prior " + }, + { + "bbox": [ + 46, + 432, + 287, + 540 + ], + "type": "inline_equation", + "content": "e_{i^*}" + }, + { + "bbox": [ + 46, + 432, + 287, + 540 + ], + "type": "text", + "content": ", the final deformation result tends to break the shape geometry (see Figure 3), leading to performance degradation as shown in Table 1. Thereby, we choose the soft assignment to enable flexible deformation while preserving the shape geometry." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "text", + "content": "We train the deformation field " + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\theta_1}" + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\theta_2}" + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "text", + "content": " in an end-to-end fashion. In our experiments, " + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\theta_1}" + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "text", + "content": " is a 6-layer MLP conditioned on both global geometry " + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "inline_equation", + "content": "z_P" + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "text", + "content": " via hypernetwork " + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "text", + "content": " [51, 52] and local part semantics " + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\alpha(x)" + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "text", + "content": " via concatenation with input coordinates. The shape latent code " + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "inline_equation", + "content": "z_P" + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "text", + "content": " and part deformation priors " + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\{\\pmb{e}_i\\}" + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "text", + "content": " are randomly initialized and jointly optimized with neural fields. " + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\theta_2}" + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "text", + "content": " is a 5-layer MLP." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 88, + 666, + 247, + 700 + ], + "blocks": [ + { + "bbox": [ + 47, + 632, + 287, + 666 + ], + "lines": [ + { + "bbox": [ + 47, + 632, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 47, + 632, + 287, + 666 + ], + "type": "text", + "content": "Table 1. Ablations of different assignment strategies for SDC. We conduct a quantitative comparison between hard and soft assignments with 5-shot part label transfer on ShapeNet chair." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 88, + 666, + 247, + 700 + ], + "lines": [ + { + "bbox": [ + 88, + 666, + 247, + 700 + ], + "spans": [ + { + "bbox": [ + 88, + 666, + 247, + 700 + ], + "type": "table", + "html": "
Assignment strategiesHardSoft
chair (5-shot PLT)77.084.2
", + "image_path": "ed6d6b55ebcc8e221a658f50958afd732547feaaea2845337de00adced3979fd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 337, + 81, + 518, + 156 + ], + "blocks": [ + { + "bbox": [ + 337, + 81, + 518, + 156 + ], + "lines": [ + { + "bbox": [ + 337, + 81, + 518, + 156 + ], + "spans": [ + { + "bbox": [ + 337, + 81, + 518, + 156 + ], + "type": "image", + "image_path": "3ffcd3cae66c1d892adb41f0722cb7d9eaec6987c5d6eee85e1dd68309f33035.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 159, + 547, + 194 + ], + "lines": [ + { + "bbox": [ + 304, + 159, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 547, + 194 + ], + "type": "text", + "content": "Figure 3. Effects of different assignment strategies for SDC. Hard assignment breaks the shape geometry, while soft assignment preserves the geometry while learning flexible deformations." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 204, + 538, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 204, + 538, + 217 + ], + "spans": [ + { + "bbox": [ + 305, + 204, + 538, + 217 + ], + "type": "text", + "content": "4.2. Part deformation consistency regularizations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 222, + 546, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 222, + 546, + 331 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 546, + 331 + ], + "type": "text", + "content": "In addition to conditioning neural fields, leveraging semantic information, we propose novel part deformation consistency regularizations to encourage semantically plausible deformation. Semantically similar points should be close after deformation in both an input space and a latent space. That is, the deformation of points from the same part should be consistent. Figure 2(b) illustrates the concept of our regularizations. Given a pair of deformed shapes " + }, + { + "bbox": [ + 304, + 222, + 546, + 331 + ], + "type": "inline_equation", + "content": "P', Q'" + }, + { + "bbox": [ + 304, + 222, + 546, + 331 + ], + "type": "text", + "content": ", part deformation consistency regularizations are defined as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 322, + 349, + 545, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 349, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 322, + 349, + 545, + 428 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {p d c - g e o}} = \\sum_ {i = 1} ^ {k} \\left(\\frac {1}{| \\mathcal {P} _ {i} ^ {\\prime} |} \\sum_ {\\boldsymbol {x} \\in \\mathcal {P} _ {i} ^ {\\prime}} \\| \\boldsymbol {x} - C (\\boldsymbol {x}, \\mathcal {Q} _ {i} ^ {\\prime}) \\| _ {2} ^ {2} \\right. \\\\ \\left. + \\frac {1}{\\left| \\mathcal {Q} _ {i} ^ {\\prime} \\right|} \\sum_ {\\boldsymbol {y} \\in \\mathcal {Q} _ {i} ^ {\\prime}} \\| \\boldsymbol {y} - C (\\boldsymbol {y}, \\mathcal {P} _ {i} ^ {\\prime}) \\| _ {2} ^ {2}\\right), \\text {a n d} \\tag {6} \\\\ \\end{array}", + "image_path": "eb804e4f1a6c095ea5e176e4911704270fc343529aec014f4e3106315609a9e8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 310, + 449, + 545, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 449, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 310, + 449, + 545, + 539 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {p d c} \\text {- s e m}} = \\sum_ {i = 1} ^ {k} \\left(\\frac {1}{| \\mathcal {P} _ {i} ^ {\\prime} |} \\sum_ {\\boldsymbol {x} \\in \\mathcal {P} _ {i} ^ {\\prime}} \\| \\boldsymbol {o} ^ {P ^ {\\prime}} (\\boldsymbol {x}) - \\boldsymbol {o} ^ {Q ^ {\\prime}} \\left(C (\\boldsymbol {x}, \\mathcal {Q} _ {i} ^ {\\prime})\\right) \\| _ {2} ^ {2} \\right. \\\\ \\left. + \\frac {1}{| \\mathcal {Q} _ {i} ^ {\\prime} |} \\sum_ {\\boldsymbol {y} \\in \\mathcal {Q} _ {i} ^ {\\prime}} \\| \\boldsymbol {o} ^ {Q ^ {\\prime}} (\\boldsymbol {y}) - \\boldsymbol {o} ^ {P ^ {\\prime}} \\left(C (\\boldsymbol {y}, \\mathcal {P} _ {i} ^ {\\prime})\\right) \\| _ {2} ^ {2}\\right), \\tag {7} \\\\ \\end{array}", + "image_path": "6b5a906522e995bc237a0af2ca402c28caa48256de980495f092b1440705b598.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 548, + 547, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 547, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 547, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 548, + 547, + 593 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 548, + 547, + 593 + ], + "type": "text", + "content": " is a correspondence matching function for a point " + }, + { + "bbox": [ + 304, + 548, + 547, + 593 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 548, + 547, + 593 + ], + "type": "text", + "content": " and its nearest point in a set " + }, + { + "bbox": [ + 304, + 548, + 547, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 304, + 548, + 547, + 593 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 548, + 547, + 593 + ], + "type": "inline_equation", + "content": "C(\\pmb{x},\\mathcal{Q}) = \\operatorname*{argmin}_{\\pmb{y}\\in \\mathcal{Q}}\\| \\pmb{x} - \\pmb{y}\\|_2^2" + }, + { + "bbox": [ + 304, + 548, + 547, + 593 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 548, + 547, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i'" + }, + { + "bbox": [ + 304, + 548, + 547, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 548, + 547, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_i'" + }, + { + "bbox": [ + 304, + 548, + 547, + 593 + ], + "type": "text", + "content": " denote sampled point sets of" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 593, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 622 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 593, + 545, + 622 + ], + "type": "text", + "content": " deformed part shapes of " + }, + { + "bbox": [ + 304, + 593, + 545, + 622 + ], + "type": "inline_equation", + "content": "P^{\\prime}" + }, + { + "bbox": [ + 304, + 593, + 545, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 593, + 545, + 622 + ], + "type": "inline_equation", + "content": "Q^{\\prime}" + }, + { + "bbox": [ + 304, + 593, + 545, + 622 + ], + "type": "text", + "content": ". They are defined as " + }, + { + "bbox": [ + 304, + 593, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i^\\prime = \\{\\pmb {x} + \\mathcal{D}_{\\theta_1}(\\pmb {x};\\pmb {z}_P,\\pmb {\\alpha}(\\pmb {x};P))\\mid i = \\operatorname *{argmax}_{k}(\\pmb {o}_k^P (\\pmb {x})),\\pmb {x}\\in" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 622, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 622, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 622, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\Omega_P\\}" + }, + { + "bbox": [ + 304, + 622, + 545, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 622, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_i^\\prime = \\{\\pmb {x} + \\mathcal{D}_{\\theta_1}(\\pmb {x};\\pmb {z}_Q,\\pmb {\\alpha}(\\pmb {x};Q))|i =" + }, + { + "bbox": [ + 304, + 622, + 545, + 653 + ], + "type": "text", + "content": " argmax " + }, + { + "bbox": [ + 304, + 622, + 545, + 653 + ], + "type": "inline_equation", + "content": "(\\pmb{o}_k^Q (\\pmb {x})),\\pmb {x}\\in \\Omega_Q\\}" + }, + { + "bbox": [ + 304, + 622, + 545, + 653 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 304, + 622, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\Omega_P" + }, + { + "bbox": [ + 304, + 622, + 545, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 622, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\Omega_{Q}" + }, + { + "bbox": [ + 304, + 622, + 545, + 653 + ], + "type": "text", + "content": " are sam" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "pled query point sets for shape " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": ". Note that each deformed part shape is extracted based on part semantics from original shapes " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": ". Here, we explicitly guide the part-level deformation among various shapes to be consistent. Moreover, since we understand part arrangement and" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "596" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 544, + 186 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 544, + 186 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 544, + 186 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 544, + 186 + ], + "type": "image", + "image_path": "c11deb1e2f34ad9c566244693f1e65d187963d0795697e23acc5908163a4b361.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 193, + 546, + 217 + ], + "lines": [ + { + "bbox": [ + 46, + 193, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 546, + 217 + ], + "type": "text", + "content": "Figure 4. Keypoint transfer result on ShapeNet. We transfer keypoint labels from source shapes (colored in ivory) to target shapes and compare ours with DIT [1] and DIF [2]. For additional visualizations including results on ScanObjectNN [53], see the supplement." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 249, + 105, + 285 + ], + "blocks": [ + { + "bbox": [ + 59, + 249, + 105, + 285 + ], + "lines": [ + { + "bbox": [ + 59, + 249, + 105, + 285 + ], + "spans": [ + { + "bbox": [ + 59, + 249, + 105, + 285 + ], + "type": "image", + "image_path": "cafeb9d58e02ba8740aa2249583b33b699f5429902f293ffbc09e7ad539f23d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 288, + 88, + 297 + ], + "lines": [ + { + "bbox": [ + 78, + 288, + 88, + 297 + ], + "spans": [ + { + "bbox": [ + 78, + 288, + 88, + 297 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 118, + 251, + 185, + 280 + ], + "blocks": [ + { + "bbox": [ + 118, + 251, + 185, + 280 + ], + "lines": [ + { + "bbox": [ + 118, + 251, + 185, + 280 + ], + "spans": [ + { + "bbox": [ + 118, + 251, + 185, + 280 + ], + "type": "image", + "image_path": "4b36f80cd31eee68e6a5b448f1efa22671b904462eb9211f55d754c17fb51573.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 287, + 157, + 297 + ], + "lines": [ + { + "bbox": [ + 146, + 287, + 157, + 297 + ], + "spans": [ + { + "bbox": [ + 146, + 287, + 157, + 297 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 197, + 243, + 279, + 284 + ], + "blocks": [ + { + "bbox": [ + 197, + 243, + 279, + 284 + ], + "lines": [ + { + "bbox": [ + 197, + 243, + 279, + 284 + ], + "spans": [ + { + "bbox": [ + 197, + 243, + 279, + 284 + ], + "type": "image", + "image_path": "f5876dceecdf1c4231267053c9f0d994cbe0c8858bedf5f786589aa659e3d8c1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 287, + 245, + 297 + ], + "lines": [ + { + "bbox": [ + 235, + 287, + 245, + 297 + ], + "spans": [ + { + "bbox": [ + 235, + 287, + 245, + 297 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 301, + 268, + 312 + ], + "lines": [ + { + "bbox": [ + 66, + 301, + 268, + 312 + ], + "spans": [ + { + "bbox": [ + 66, + 301, + 268, + 312 + ], + "type": "text", + "content": "Figure 5. Efficacy of global scaling consistency " + }, + { + "bbox": [ + 66, + 301, + 268, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{scale}}" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 319, + 286, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 286, + 342 + ], + "type": "text", + "content": "only penalize matched parts, our method does not forcibly deform \"unseen/missing match\" cases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 343, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 343, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 343, + 287, + 475 + ], + "type": "text", + "content": "We partly observed spatial distortion when part deformation consistency was too strongly imposed. This is because using external semantics for deformation could break the continuity of geometric fields. In addition, we propose global scale consistency and global deformation consistency to address the concern. For the former one, we disentangle a scaling operation from the entire non-rigid deformation " + }, + { + "bbox": [ + 46, + 343, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\theta_1}" + }, + { + "bbox": [ + 46, + 343, + 287, + 475 + ], + "type": "text", + "content": " to calculate the global scale of the learned implicit template field. That is, we want to approximate " + }, + { + "bbox": [ + 46, + 343, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 343, + 287, + 475 + ], + "type": "text", + "content": " with the global scaling of the scaling factor of " + }, + { + "bbox": [ + 46, + 343, + 287, + 475 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 343, + 287, + 475 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 343, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{D} \\approx rI" + }, + { + "bbox": [ + 46, + 343, + 287, + 475 + ], + "type": "text", + "content": ". Then the problem becomes finding " + }, + { + "bbox": [ + 46, + 343, + 287, + 475 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 343, + 287, + 475 + ], + "type": "text", + "content": " from" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 96, + 480, + 287, + 505 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 480, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 96, + 480, + 287, + 505 + ], + "type": "interline_equation", + "content": "\\min _ {r} \\sum_ {s} \\sum_ {i} \\| \\boldsymbol {x} _ {i} ^ {s} + \\Delta \\boldsymbol {x} _ {i} ^ {s} - r \\boldsymbol {x} _ {i} ^ {s} \\| _ {2} ^ {2}, \\tag {8}", + "image_path": "8a58656bc44896d4b75ddecc165494e539296039f53ae0c0d6b8bfc6fc2ea77f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 510, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 510, + 287, + 604 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 510, + 287, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 287, + 604 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 510, + 287, + 604 + ], + "type": "text", + "content": " are an index for shape and point, respectively. A closed-form solution of (8) is " + }, + { + "bbox": [ + 46, + 510, + 287, + 604 + ], + "type": "inline_equation", + "content": "r = \\sum_{s}\\sum_{i}\\frac{\\pmb{x}_i^{s\\top}(\\pmb{x}_i^s + \\Delta\\pmb{x}_i^s)}{\\pmb{x}_i^{s\\top}\\pmb{x}_i^s}" + }, + { + "bbox": [ + 46, + 510, + 287, + 604 + ], + "type": "text", + "content": ". See the supplement for proof. Since we want the learned " + }, + { + "bbox": [ + 46, + 510, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 510, + 287, + 604 + ], + "type": "text", + "content": " to be insensitive to global scale operation, i.e., the implicit template to keep its scale by the average scale of input shapes, we penalize the expected " + }, + { + "bbox": [ + 46, + 510, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[r]" + }, + { + "bbox": [ + 46, + 510, + 287, + 604 + ], + "type": "text", + "content": " over input shapes " + }, + { + "bbox": [ + 46, + 510, + 287, + 604 + ], + "type": "inline_equation", + "content": "P_{1}\\dots P_{N}" + }, + { + "bbox": [ + 46, + 510, + 287, + 604 + ], + "type": "text", + "content": " in a single batch as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 127, + 610, + 287, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 610, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 127, + 610, + 287, + 624 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s c a l e}} = | \\mathbb {E} [ r ] - 1 | \\tag {9}", + "image_path": "9bd8f3633a2ce08aab296269a01905240157cfeba9e2e2b420754aceb85eaae1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Figure 5 illuminates the effect of " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{scale}}" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": ". For the same input shapes, Figure 5(a) is a car template learned without " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{scale}}" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": ", Figure 5(b) is a car template learned with " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{scale}}" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": ", and Figure 5(c) is the result of placing two shapes in the same space. It shows " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{scale}}" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": " helps the model to learn a more desirable template by penalizing the change in the global scale of the learned template fields." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 242, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 242, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 305, + 242, + 545, + 266 + ], + "type": "text", + "content": "For the latter one, we add global deformation consistency for a pair of deformed shapes " + }, + { + "bbox": [ + 305, + 242, + 545, + 266 + ], + "type": "inline_equation", + "content": "P', Q'" + }, + { + "bbox": [ + 305, + 242, + 545, + 266 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 383, + 270, + 545, + 284 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 270, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 383, + 270, + 545, + 284 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {g e o}} = \\mathbf {C D} \\left(P ^ {\\prime}, Q ^ {\\prime}\\right), \\tag {10}", + "image_path": "0afea46b797c9ff7c739bed6a68b69be8dffbfaf4b4bfbeee61d9783a452b925.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 289, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 545, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 289, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{CD}" + }, + { + "bbox": [ + 305, + 289, + 545, + 312 + ], + "type": "text", + "content": " denotes a conventional Chamfer loss. Thus, the deformation consistency regularization can be written as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 317, + 545, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 545, + 331 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d c}} = \\gamma_ {1} \\mathcal {L} _ {\\mathrm {p d c - g e o}} + \\gamma_ {2} \\mathcal {L} _ {\\mathrm {p d c - s e m}} + \\gamma_ {3} \\mathcal {L} _ {\\text {s c a l e}} + \\gamma_ {4} \\mathcal {L} _ {\\text {g e o}}. \\tag {11}", + "image_path": "b62615906480c252e678cee37d02921f7c8f05ad2b01182df3fee75018b7bae1.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 353, + 391, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 353, + 391, + 366 + ], + "spans": [ + { + "bbox": [ + 306, + 353, + 391, + 366 + ], + "type": "text", + "content": "4.3. Optimization" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 371, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 545, + 419 + ], + "type": "text", + "content": "We train our deformation and template fields with a reconstruction loss and several regularizers including our deformation consistency regularization. Following [2], the reconstruction loss " + }, + { + "bbox": [ + 304, + 371, + 545, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 304, + 371, + 545, + 419 + ], + "type": "text", + "content": " for shape " + }, + { + "bbox": [ + 304, + 371, + 545, + 419 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 371, + 545, + 419 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 318, + 423, + 545, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 423, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 318, + 423, + 545, + 480 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {r e c}} = \\sum_ {\\boldsymbol {x} \\in \\Omega_ {P}} | \\mathcal {F} _ {\\theta} (\\boldsymbol {x}) - \\bar {s} | + \\sum_ {\\boldsymbol {x} \\in P} (1 - \\langle \\nabla \\mathcal {F} _ {\\theta} (\\boldsymbol {x}), \\bar {n} \\rangle) \\\\ + \\sum_ {\\boldsymbol {x} \\in \\Omega_ {P}} | \\| \\nabla \\mathcal {F} _ {\\theta} (\\boldsymbol {x}) \\| _ {2} - 1 | + \\sum_ {\\boldsymbol {x} \\in \\Omega_ {P} \\backslash P} \\exp (- \\delta \\cdot | \\mathcal {F} _ {\\theta} (\\boldsymbol {x}) |), \\tag {12} \\\\ \\end{array}", + "image_path": "d4ffa865edbad67fffe2560ca260ce259689b20ef63a6c50d9e981b7798f0a9b.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 482, + 545, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 545, + 600 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 482, + 545, + 600 + ], + "type": "inline_equation", + "content": "\\bar{s}" + }, + { + "bbox": [ + 304, + 482, + 545, + 600 + ], + "type": "text", + "content": " denotes the ground-truth SDF value of each query point, " + }, + { + "bbox": [ + 304, + 482, + 545, + 600 + ], + "type": "inline_equation", + "content": "\\bar{n}" + }, + { + "bbox": [ + 304, + 482, + 545, + 600 + ], + "type": "text", + "content": " is the ground truth surface normal vector of each surface point and " + }, + { + "bbox": [ + 304, + 482, + 545, + 600 + ], + "type": "inline_equation", + "content": "\\nabla \\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 304, + 482, + 545, + 600 + ], + "type": "text", + "content": " the spatial derivative of a 3D field. The second term of (12) is derived from the property of the SDF function that the gradient of an SDF is identical to the surface normal for surface points and the third is implicit geometric regularization [54] derived from the Eikonal equation. The last term pertains to penalizing off-surface points to be away from the surface points, which is controlled by " + }, + { + "bbox": [ + 304, + 482, + 545, + 600 + ], + "type": "inline_equation", + "content": "\\delta \\gg 1" + }, + { + "bbox": [ + 304, + 482, + 545, + 600 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "text", + "content": "Since we do not know ground truth mapping between the template and input shapes, we manipulate the deformation field through multiple regularizers. First, we employ our proposed deformation consistency regularization (11). For further regularizers, we adopt [2] such as deformation smoothness " + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{smooth}}" + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "text", + "content": ", normal consistency " + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{normal}}" + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "text", + "content": " and minimal correction " + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{c}}" + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "text", + "content": ", which are defined as" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 689, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 315, + 689, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s m o o t h}} = \\sum_ {\\boldsymbol {x} \\in \\Omega_ {P}} \\sum_ {d \\in \\{X, Y, Z \\}} \\| \\nabla \\mathcal {D} _ {\\theta_ {1}} ^ {v} (\\cdot , \\boldsymbol {z} _ {P}) | _ {d} (\\boldsymbol {x}) \\| _ {2} ^ {2}, \\tag {13}", + "image_path": "e49c4c557d313561cdc4724452104ef4ca31d8d871f4307ab4edfbe23c6c49a8.jpg" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "597" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 73, + 541, + 184 + ], + "blocks": [ + { + "bbox": [ + 56, + 73, + 541, + 184 + ], + "lines": [ + { + "bbox": [ + 56, + 73, + 541, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 541, + 184 + ], + "type": "image", + "image_path": "da82899a2c9ce4b1dc7fc4cccec1d8fceb2404da3d183bb1d68551ee615a29cb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 190, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 213 + ], + "type": "text", + "content": "Figure 6. Part label transfer result on ShapeNet. We transfer part segment labels from source shapes (colored in ivory) to target shapes and compare ours with DIT [1] and DIF [2]. For additional visualizations, please refer to the supplement." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 246, + 287, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 246, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 59, + 246, + 287, + 283 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {n o r m a l}} = \\sum_ {i} \\sum_ {\\boldsymbol {x} \\in \\mathcal {S} _ {i}} \\left(1 - \\left\\langle \\nabla \\mathcal {T} _ {\\theta_ {2}} \\left(\\boldsymbol {x} + D _ {\\theta_ {1}} ^ {v} (\\boldsymbol {x})\\right), \\bar {n} \\right\\rangle\\right), \\tag {14}", + "image_path": "472be19e3b175f2902fcdb97acf7c8c26cda341d8f513168ed08bf6d4f98c159.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 292, + 287, + 318 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 292, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 113, + 292, + 287, + 318 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c}} = \\sum_ {i} \\sum_ {\\boldsymbol {x} \\in \\Omega_ {P}} | \\mathcal {D} _ {\\theta_ {1}} ^ {\\Delta s} (\\boldsymbol {x}) |. \\tag {15}", + "image_path": "604bd491150524f3f77d1a86436119195115c29c97e686da0992bacb0b5df465.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 323, + 287, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 287, + 347 + ], + "type": "text", + "content": "We also apply " + }, + { + "bbox": [ + 47, + 323, + 287, + 347 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 47, + 323, + 287, + 347 + ], + "type": "text", + "content": " regularization on shape latent codes and part deformation priors" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 101, + 355, + 287, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 355, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 101, + 355, + 287, + 389 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {e m b}} = \\sum_ {i} ^ {N} \\| z _ {i} \\| _ {2} ^ {2} + \\sum_ {j} ^ {k} \\| e _ {j} \\| _ {2} ^ {2}. \\tag {16}", + "image_path": "7bcaf8f957b695d56e324bec8a38d879a443cc89e81669847b1b8cff26e4caf4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 397, + 220, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 397, + 220, + 410 + ], + "spans": [ + { + "bbox": [ + 47, + 397, + 220, + 410 + ], + "type": "text", + "content": "The regularizers are collectively written as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 418, + 287, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 418, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 61, + 418, + 287, + 432 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e g}} = \\gamma_ {5} \\mathcal {L} _ {\\text {s m o o t h}} + \\gamma_ {6} \\mathcal {L} _ {\\text {n o r m a l}} + \\gamma_ {7} \\mathcal {L} _ {\\mathrm {c}} + \\gamma_ {8} \\mathcal {L} _ {\\text {e m b}}. \\tag {17}", + "image_path": "aff6a9195c790ad53a0397655c84cd03e21db7889b2c303c516c0a5ee1a331ca.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 441, + 168, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 168, + 452 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 168, + 452 + ], + "type": "text", + "content": "Hence, the final loss becomes" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 118, + 462, + 287, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 462, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 118, + 462, + 287, + 475 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {\\mathrm {d c}} + \\mathcal {L} _ {\\text {r e g}}. \\tag {18}", + "image_path": "c37e65141d64e1477be0700b9bd7492fb3d6d3ac20a42fcb4f28679399756a5d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 484, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 556 + ], + "type": "text", + "content": "Evaluating dense correspondence. Given two pairs of seen or unseen shapes " + }, + { + "bbox": [ + 46, + 484, + 287, + 556 + ], + "type": "inline_equation", + "content": "P, Q" + }, + { + "bbox": [ + 46, + 484, + 287, + 556 + ], + "type": "text", + "content": ", we can evaluate the dense correspondence between " + }, + { + "bbox": [ + 46, + 484, + 287, + 556 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 484, + 287, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 484, + 287, + 556 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 484, + 287, + 556 + ], + "type": "text", + "content": " by deforming each shape to the common template space. Moreover, the uncertainty of dense correspondence performance between two points " + }, + { + "bbox": [ + 46, + 484, + 287, + 556 + ], + "type": "inline_equation", + "content": "p \\in P" + }, + { + "bbox": [ + 46, + 484, + 287, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 484, + 287, + 556 + ], + "type": "inline_equation", + "content": "q \\in Q" + }, + { + "bbox": [ + 46, + 484, + 287, + 556 + ], + "type": "text", + "content": " can be measured as below:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 564, + 287, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 564, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 58, + 564, + 287, + 578 + ], + "type": "interline_equation", + "content": "u (p, q) = 1 - \\exp (- \\gamma \\| (p + \\Delta p) - (q + \\Delta q) \\| _ {2} ^ {2}). \\tag {19}", + "image_path": "3fe86cd8e902b29bc84e6e7dade001cd7c462b05b3aa5ec8395367960acb12af.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 587, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 587, + 287, + 647 + ], + "type": "text", + "content": " is a hyperparameter for modulation and " + }, + { + "bbox": [ + 46, + 587, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\Delta p, \\Delta q" + }, + { + "bbox": [ + 46, + 587, + 287, + 647 + ], + "type": "text", + "content": " are output from " + }, + { + "bbox": [ + 46, + 587, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\theta_1}" + }, + { + "bbox": [ + 46, + 587, + 287, + 647 + ], + "type": "text", + "content": ". Note that since each part deformation prior " + }, + { + "bbox": [ + 46, + 587, + 287, + 647 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 587, + 287, + 647 + ], + "type": "text", + "content": " is shared based on part semantics across entire training shapes, we do not need a test-time optimization on " + }, + { + "bbox": [ + 46, + 587, + 287, + 647 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 587, + 287, + 647 + ], + "type": "text", + "content": " for an unseen shape at inference time." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 657, + 128, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 128, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 128, + 671 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "We demonstrate the effectiveness of our proposed framework and provide both quantitative and qualitative results. In Section 5.1, we conduct multiple label transfer" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 307, + 247, + 547, + 308 + ], + "blocks": [ + { + "bbox": [ + 305, + 224, + 545, + 246 + ], + "lines": [ + { + "bbox": [ + 305, + 224, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 545, + 246 + ], + "type": "text", + "content": "Table 2. Keypoint correspondence accuracy in KeypointNet [55]. PCK scores with threshold of " + }, + { + "bbox": [ + 305, + 224, + 545, + 246 + ], + "type": "inline_equation", + "content": "0.05 / 0.1" + }, + { + "bbox": [ + 305, + 224, + 545, + 246 + ], + "type": "text", + "content": " are reported." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 247, + 547, + 308 + ], + "lines": [ + { + "bbox": [ + 307, + 247, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 247, + 547, + 308 + ], + "type": "table", + "html": "
PCKairplanecarchairtable
AtlasNetV2 [34]23.9 / 45.333.5 / 64.025.6 / 50.225.3 / 48.0
DIT [2]13.6 / 34.258.9 / 87.733.3 / 66.726.2 / 55.5
DIF [1]16.5 / 40.567.8 / 90.434.3 / 57.637.9 / 62.9
Ours24.2 / 52.369.3 / 91.861.2 / 83.250.8/77.9
", + "image_path": "14e884a9b2c350111152cd9b325302420efe9ad71d8c1301a84b84ea6fe4b7b3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 317, + 339, + 536, + 416 + ], + "blocks": [ + { + "bbox": [ + 305, + 316, + 547, + 338 + ], + "lines": [ + { + "bbox": [ + 305, + 316, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 547, + 338 + ], + "type": "text", + "content": "Table 3. 5-shot part label transfer in four categories of ShapeNet. The reported numbers are overall mIoU." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 339, + 536, + 416 + ], + "lines": [ + { + "bbox": [ + 317, + 339, + 536, + 416 + ], + "spans": [ + { + "bbox": [ + 317, + 339, + 536, + 416 + ], + "type": "table", + "html": "
modelairplanecarchairtable
AtlasNetV2 [34]57.859.366.970.6
DIT [1]64.266.180.486.0
DIF [2]73.969.682.186.6
Ours75.170.484.286.9
", + "image_path": "2d58bb777779ba8ee363b11e57efc839f42847be5442e9977970070eaba1d71c.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 435, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 519 + ], + "type": "text", + "content": "tasks for dense correspondence evaluation. In Section 5.2, we measure shape reconstruction accuracy to evaluate the representational power for unknown shapes. In Section 5.3, we present advanced dense correspondence analysis under more challenging settings. Lastly, in Section 5.4, we perform ablation studies to analyze the proposed regularization and the sensitivity over feature extractor [31]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 521, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 652 + ], + "type": "text", + "content": "Datasets and Baselines. We train and validate our model and baselines on four categories (e.g., chair, table, airplane and car) of ShapeNetV2 [50]. We follow the training split of DIF [2] for each category. For label transfer tasks, we use the labels from ShapeNet-Part [56] and KeypointNet [55]. DIT [1] and DIF [2] are the most relevant baseline models as they both globally learn implicit template fields, and we select AtlasNetV2 [34] as an additional baseline. Dense correspondence is measured with local patch composition in [34]. For more implementation details, analyses, and visualizations, see the supplement." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 659, + 433, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 433, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 433, + 672 + ], + "type": "text", + "content": "5.1. Dense correspondence" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "To evaluate the quality of learned dense correspondence, we conduct various surrogate tasks including keypoint transfer, part label transfer, and texture transfer due to" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "598" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 73, + 537, + 205 + ], + "blocks": [ + { + "bbox": [ + 55, + 73, + 537, + 205 + ], + "lines": [ + { + "bbox": [ + 55, + 73, + 537, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 73, + 537, + 205 + ], + "type": "image", + "image_path": "d992a9a2429a7e8c487f80fb2c12c05698fa643a016b4a70d35d2eca80ee5f3e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 211, + 547, + 235 + ], + "lines": [ + { + "bbox": [ + 46, + 211, + 547, + 235 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 547, + 235 + ], + "type": "text", + "content": "Figure 7. Texture transfer result on ShapeNet. We transfer part segment labels from source shapes (colored in ivory) to target shapes and compare ours with DIT [1] and DIF [2]. For additional visualizations, please refer to the supplement." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 282, + 282, + 348 + ], + "blocks": [ + { + "bbox": [ + 47, + 240, + 288, + 282 + ], + "lines": [ + { + "bbox": [ + 47, + 240, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 288, + 282 + ], + "type": "text", + "content": "Table 4. Reconstruction accuracy for unseen shapes in four categories of ShapeNet. The reported numbers are chamfer distance (CD) between ground truth and reconstructed meshes at a resolution of " + }, + { + "bbox": [ + 47, + 240, + 288, + 282 + ], + "type": "inline_equation", + "content": "256^{3}" + }, + { + "bbox": [ + 47, + 240, + 288, + 282 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 282, + 282, + 348 + ], + "lines": [ + { + "bbox": [ + 52, + 282, + 282, + 348 + ], + "spans": [ + { + "bbox": [ + 52, + 282, + 282, + 348 + ], + "type": "table", + "html": "
CD (×103)↓airplanecarchairtable
DIT [1]0.22610.40820.30630.6187
DIF [2]0.20930.31570.42290.2640
Ours0.18590.31500.22650.2796
", + "image_path": "90e6e2b99689eff2b04221393f9d7235b1b3c75beda2f97bc90fbbe1218e35c2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 373, + 287, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 410 + ], + "type": "text", + "content": "the lack of labeled datasets for 3D dense correspondence. For the following tasks, we transfer respective labels from source shapes to target shapes in a common template space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 415, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 289, + 715 + ], + "type": "text", + "content": "Keypoint label transfer. Table 2 presents the quantitative result in a 5-shot keypoint transfer task. We measure the performance by a percentage of correct keypoints (PCK) [57]. We compute the distance between the transferred keypoints and the ground truth points to record the PCK score with the threshold of " + }, + { + "bbox": [ + 46, + 415, + 289, + 715 + ], + "type": "inline_equation", + "content": "0.05 / 0.1" + }, + { + "bbox": [ + 46, + 415, + 289, + 715 + ], + "type": "text", + "content": ". Our method achieves the best performance in all four categories with a significant improvement. Especially for categories with high structural variability (e.g., chair and table), a large performance gain is observed. For instance, even compared to the best baseline, on the chair category with the threshold of 0.1, our method (83.2) outperforms DIT [1] (66.7) by " + }, + { + "bbox": [ + 46, + 415, + 289, + 715 + ], + "type": "inline_equation", + "content": "16.5\\%" + }, + { + "bbox": [ + 46, + 415, + 289, + 715 + ], + "type": "text", + "content": ", and on the table category with the threshold of 0.1, our method (77.9) outperforms DIF [2] (62.9) by " + }, + { + "bbox": [ + 46, + 415, + 289, + 715 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 46, + 415, + 289, + 715 + ], + "type": "text", + "content": ". We additionally provide the visualization of keypoint transfer in Figure 4, which also support the importance of understanding part semantics for shape correspondence. In particular, two keypoints of pea-green and orange from the source chair are geometrically close but semantically different since one is for the arm part and the other is for the seat part. Among several methods, only our framework succeeds in disentangling two keypoints in the target chair. We provide more qualitative results regarding keypoint transfer results in real-world data [53] to validate the robustness of our method in the supplement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 243, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 547, + 411 + ], + "type": "text", + "content": "Part label transfer. For few-shot part label transfer tasks, we transfer part segment labels from source shapes to target shapes. In detail, for each (target) point of the deformed target shape, we find the " + }, + { + "bbox": [ + 304, + 243, + 547, + 411 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 243, + 547, + 411 + ], + "type": "text", + "content": " nearest points of the deformed source shapes and perform weighted voting using their labels and distances to the target point. In our experiments, we used " + }, + { + "bbox": [ + 304, + 243, + 547, + 411 + ], + "type": "inline_equation", + "content": "n = 10" + }, + { + "bbox": [ + 304, + 243, + 547, + 411 + ], + "type": "text", + "content": " nearest points. For a 5-shot part label transfer task in Table 3, five source shapes were used just the same as [2] for every category. Table 3 shows our framework achieves the best performance over all four categories. Moreover, Figure 6 illustrates the qualitative results of one-shot part label transfer on ShapeNet [50]. Similar to keypoint transfer, Figure 6 also proves the effectiveness of semantic knowledge in dense correspondence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 411, + 547, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 507 + ], + "type": "text", + "content": "Texture transfer. Figure 7 shows texture transfer results on ShapeNetV2 [50] to compare established dense correspondence between baselines and ours with higher-frequency label transfer. We observe notable texture transfer results. For instance, our framework successfully transfers the source car's texture of the windshield (gray) to the corresponding part of the target cars, whereas baselines transfer the same textures to the roof part of the target cars." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 515, + 429, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 429, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 429, + 528 + ], + "type": "text", + "content": "5.2. Shape reconstruction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": "We report the reconstruction quality of unseen shapes with Chamfer Distance (CD) to validate the representational power of our framework in Table 4. A smaller CD means more accurate reconstruction. As in previous work, we report the " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CD} \\times 10^{3}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " for readability. Also, we consistently discuss the CD and performance gain below at this scale. We optimize the corresponding shape latent code for a given unseen shape. We measure CD as in DIF [2], where reconstructed meshes are extracted at a resolution of " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "256^{3}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": ", and CD is calculated with " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "10K" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " sampled points. Our framework on average achieved the best reconstruction results surpassing DIF by a margin of 0.0512. On average, our method and DIF exhibit 0.2518 and 0.3030 reconstruction errors, respectively. In three out of four categories, our method achieved the best reconstruction accuracy." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "599" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 185, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 185, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 185, + 85 + ], + "type": "text", + "content": "5.3. More challenging setups" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 162 + ], + "type": "text", + "content": "To further validate the effectiveness of our method, we perform additional experiments and compare the results with DIF [2]. Overall, the results below imply current methods that lack part semantics and do not consider part deformation consistency may fail to learn semantically plausible deformation in challenging cases." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 163, + 288, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 163, + 288, + 341 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 288, + 341 + ], + "type": "text", + "content": "2-shot part label transfer within different subcategories. Our objective is to assess dense correspondence performance under difficult settings. We conduct 2-shot part label transfer within two different subcategories belonging to the same category, where shapes from different subcategories tend to have different geometry for same semantics. Since ShapeNet [50] includes incorrect subcategory labels [58], we manually cleaned subcategory labels for four categories (e.g., airplane: airliner/jet, car: sedan/jeep, chair: straight chair/sofa, table: pedestal table/short table). Figure 8 displays examples of shapes from each subcategory. For robust evaluations, we randomly select the source shape 5 times for each subcategory. Table 5 shows our framework achieves a significant performance gain in all cases by over " + }, + { + "bbox": [ + 46, + 163, + 288, + 341 + ], + "type": "inline_equation", + "content": "7.2\\%" + }, + { + "bbox": [ + 46, + 163, + 288, + 341 + ], + "type": "text", + "content": " on average compared to DIF [2]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 342, + 288, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 288, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 288, + 641 + ], + "type": "text", + "content": "Unsupervised dense correspondence in DFAUST [59]. Furthermore, we evaluate unsupervised dense correspondence with non-rigid shapes [59]. For more challenging settings, we train both DIF [2] and our model for two different subjects with different sizes and heights (e.g., subject ID of 50002, 50026) and four distinct actions (e.g., shake arms, chicken wings, running on spot, and jumping jacks). Colormap in Figure 9 demonstrates the correspondence quality, where we transfer color over the up-axis. In the source shape, the arms are labeled in red. After transferring these color labels to the target shapes, our method preserves the red/orange labeling of the arms across different motions, whereas the baseline fails to capture the same labeling (i.e., yellow hands) since it lacks semantic knowledge during training. Additionally, we conduct keypoint transfer tasks and measure the PCK score as identical as in Section 5.1. Our model achieves 74.9/85.2 with a threshold of 0.05/0.1, whereas DIF [2] achieves 67.0/80.7 with the same threshold, showing our superior performance in the challenging settings (see the supplement for the full keypoint transfer result). Note that we can also learn implicit templates for non-rigid shapes with supervision as in [19]. We conduct experiments in unsupervised settings since our focus is to validate the effectiveness of utilizing additional semantic information for template learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 647, + 115, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 115, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 115, + 658 + ], + "type": "text", + "content": "5.4. Ablations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "Ablation on regularizations. We conduct ablation studies regarding the proposed regularizations including part deformation consistency regularizations, global scale consistency regularization, and global deformation consistency" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 554, + 121 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 554, + 121 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 554, + 121 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 554, + 121 + ], + "type": "image", + "image_path": "6cd10adf61d83a66ca80750a8786e2ac526e6ea83495786caa65bca3c3b03ef5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 123, + 544, + 135 + ], + "lines": [ + { + "bbox": [ + 306, + 123, + 544, + 135 + ], + "spans": [ + { + "bbox": [ + 306, + 123, + 544, + 135 + ], + "type": "text", + "content": "Figure 8. Types of source shapes for 2-shot part label transfer." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 312, + 167, + 542, + 239 + ], + "blocks": [ + { + "bbox": [ + 305, + 144, + 547, + 167 + ], + "lines": [ + { + "bbox": [ + 305, + 144, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 547, + 167 + ], + "type": "text", + "content": "Table 5. 2-shot part label transfer in subcategories of ShapeNet. The reported numbers are overall mIoU over 5 runs." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 167, + 542, + 239 + ], + "lines": [ + { + "bbox": [ + 312, + 167, + 542, + 239 + ], + "spans": [ + { + "bbox": [ + 312, + 167, + 542, + 239 + ], + "type": "table", + "html": "
subcategoriesDIF [2]Ours
airplane(airliner/jet)64.5±1.6172.8±0.99 (+8.3↑)
car(sedan/jeep)65.1±2.3469.1±1.46 (+4.0↑)
chair(straight/sofa)66.3±6.0677.5±5.81 (+11.2↑)
table(pedestal/short)74.4±10.179.7±6.26 (+5.3↑)
", + "image_path": "dc1a2082addac86b30bcfc88ccc5cdc6b77036672b43248d1c28ef43db747ff9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 251, + 547, + 346 + ], + "blocks": [ + { + "bbox": [ + 310, + 251, + 547, + 346 + ], + "lines": [ + { + "bbox": [ + 310, + 251, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 310, + 251, + 547, + 346 + ], + "type": "image", + "image_path": "87a877925235bd64d7f0802809ba8f9a944ff6a557b07ab22f5af790ef6063d1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 350, + 547, + 384 + ], + "lines": [ + { + "bbox": [ + 304, + 350, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 547, + 384 + ], + "type": "text", + "content": "Figure 9. Color transfer on DFAUST [59]. We transfer different color labels over the up-axis from source shapes (colored in ivory) to target shapes and compare ours with DIF [2]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 351, + 426, + 502, + 488 + ], + "blocks": [ + { + "bbox": [ + 305, + 392, + 546, + 426 + ], + "lines": [ + { + "bbox": [ + 305, + 392, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 392, + 546, + 426 + ], + "type": "text", + "content": "Table 6. Ablations for every proposed regularization. We report mIoU for 5-shot part label transfer on ShapeNet chair. Note that " + }, + { + "bbox": [ + 305, + 392, + 546, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pdc}}" + }, + { + "bbox": [ + 305, + 392, + 546, + 426 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 305, + 392, + 546, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pdc\\_geo}} + \\mathcal{L}_{\\mathrm{pdc\\_sem}}" + }, + { + "bbox": [ + 305, + 392, + 546, + 426 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 351, + 426, + 502, + 488 + ], + "lines": [ + { + "bbox": [ + 351, + 426, + 502, + 488 + ], + "spans": [ + { + "bbox": [ + 351, + 426, + 502, + 488 + ], + "type": "table", + "html": "
LpdcLgeoLscalechair
(a)✓83.4
(b)✓✓83.9
(c)✓✓✓84.2
", + "image_path": "4a4aa9ec1574d04903a1e52c1ee7901ed039391446da6f32798f612695cd34be.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 310, + 520, + 538, + 546 + ], + "blocks": [ + { + "bbox": [ + 306, + 498, + 545, + 520 + ], + "lines": [ + { + "bbox": [ + 306, + 498, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 545, + 520 + ], + "type": "text", + "content": "Table 7. BAE-Net " + }, + { + "bbox": [ + 306, + 498, + 545, + 520 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 498, + 545, + 520 + ], + "type": "text", + "content": " ablation. We report mIoU for 5-shot part label transfer on ShapeNet chair over 3 runs." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 520, + 538, + 546 + ], + "lines": [ + { + "bbox": [ + 310, + 520, + 538, + 546 + ], + "spans": [ + { + "bbox": [ + 310, + 520, + 538, + 546 + ], + "type": "table", + "html": "
k=2k=4k=6k=8k=10
chair83.8±0.5183.5±1.4882.8±1.2783.6±0.8583.7±1.33
", + "image_path": "16617a34d40e30f85270118a5840eaf7ff9adf29a7f500f6fcf9c99f2b8a6a47.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "content": "regularization. We conduct 5-shot part label transfer within a chair category, as same as Section 5.1. Table 6 shows that each component of our proposed regularizers contributes to performance improvement for dense correspondence. Note that by only using " + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pdc}}" + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "content": ", i.e., using " + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pdc\\_geo}}" + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pdc\\_sem}}" + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "content": ", it still outperforms than baselines (80.4 [1] and 82.1 [2]). The best performance was achieved by incorporating all proposed regularizations." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": "BAE-Net [31] with different " + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": ". We evaluate the effect of " + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": " for BAE-Net [31] with mean/standard deviation over 3 runs for ShapeNet [50] chair in We train BAE-Net [31] with a range of " + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "inline_equation", + "content": "k = (2,4,6,8,10)" + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "inline_equation", + "content": "k = 4" + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": " is used in the main paper, and report mIoU for 5-shot part label" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "600" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "content": "transfer task. Table 7 shows the efficacy of semantic priors, where our framework achieves the average mIoU of 83.5, which shows higher performance than baselines. More importantly, we observed that the performance is fairly robust over " + }, + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 142, + 120, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 142, + 120, + 154 + ], + "spans": [ + { + "bbox": [ + 47, + 142, + 120, + 154 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 161, + 289, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 161, + 289, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 161, + 289, + 317 + ], + "type": "text", + "content": "We have proposed a new framework that learns a semantic-aware implicit template field. Our framework extracts semantic knowledge from a self-supervised feature extractor given unlabelled point clouds. By leveraging the semantic information, we propose novel conditioning with semantic-aware deformation code and part deformation consistency regularizations to encourage semantically plausible deformation. Furthermore, we suggest global scaling consistency and global deformation consistency regularizations to enhance the learning process. Through quantitative and qualitative evaluation, we demonstrate the superiority of our proposed method and verified the effect of all proposed components in our framework." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 331, + 289, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 289, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 289, + 450 + ], + "type": "text", + "content": "Acknowledgments. This research was supported in part by the MSIT (Ministry of Science and ICT), Korea, under the ICT Creative Consilience program (IITP-2023-2020-0-01819) supervised by the IITP (Institute for Information & communications Technology Planning & Evaluation); the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF-2023R1A2C2005373); and Google Cloud Research Credits program with the award (GCLF-67WL-WU6D-MF14, V9P3-99XD-64CL-EH5H)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 459, + 107, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 107, + 471 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 107, + 471 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 479, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 479, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 479, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 479, + 287, + 512 + ], + "type": "text", + "content": "[1] Zerong Zheng, Tao Yu, Qionghai Dai, and Yebin Liu. Deep implicit templates for 3d shape representation. In CVPR, 2021. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 514, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 287, + 548 + ], + "type": "text", + "content": "[2] Yu Deng, Jiaolong Yang, and Xin Tong. Deformed implicit field: Modeling 3d shapes with learned dense correspondence. In CVPR, 2021. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 549, + 287, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 549, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 53, + 549, + 287, + 582 + ], + "type": "text", + "content": "[3] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: a skinned multiperson linear model. TOG, 2015. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 585, + 287, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 585, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 53, + 585, + 287, + 608 + ], + "type": "text", + "content": "[4] Brett Allen, Brian Curless, and Zoran Popovic. Articulated body deformation from range scan data. TOG, 2002. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 609, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 609, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 53, + 609, + 287, + 643 + ], + "type": "text", + "content": "[5] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan C. Russell, and Mathieu Aubry. 3d-coded: 3d correspondences by deep deformation. In ECCV, 2018. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 644, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 644, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 53, + 644, + 287, + 677 + ], + "type": "text", + "content": "[6] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In CVPR, 2018. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 680, + 287, + 713 + ], + "type": "text", + "content": "[7] Wang Zeng, Wanli Ouyang, Ping Luo, Wentao Liu, and Xiaogang Wang. 3d human mesh regression with dense correspondence. In CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "text", + "content": "[8] Sing Chun Lee and Misha Kazhdan. Dense point-to-point correspondences between genus-zero shapes. In Computer Graphics Forum, 2019. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 109, + 545, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 109, + 545, + 142 + ], + "spans": [ + { + "bbox": [ + 312, + 109, + 545, + 142 + ], + "type": "text", + "content": "[9] Silvia Zuffi, Angjoo Kanazawa, David W. Jacobs, and Michael J. Black. 3d menagerie: Modeling the 3d shape and pose of animals. In CVPR, 2017. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 145, + 545, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 145, + 545, + 188 + ], + "spans": [ + { + "bbox": [ + 308, + 145, + 545, + 188 + ], + "type": "text", + "content": "[10] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In AVSS, 2009. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 193, + 545, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 193, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 308, + 193, + 545, + 224 + ], + "type": "text", + "content": "[11] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. TOG, 2017. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 228, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 261 + ], + "type": "text", + "content": "[12] Luca Ballan, Aparna Taneja, Jürgen Gall, Luc Van Gool, and Marc Pollefeys. Motion capture of hands in action using discriminative salient points. In ECCV, 2012. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 297 + ], + "type": "text", + "content": "[13] Srinath Sridhar, Antti Oulasvirta, and Christian Theobalt. Interactive markerless articulated hand motion tracking using rgb and depth data. In ICCV, 2013. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 300, + 545, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 300, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 308, + 300, + 545, + 334 + ], + "type": "text", + "content": "[14] Iason Oikonomidis, Manolis IA Lourakis, and Antonis A Argyros. Evolutionary quasi-random search for hand articulations tracking. In CVPR, 2014. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 337, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 337, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 337, + 545, + 370 + ], + "type": "text", + "content": "[15] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In CVPR, 2017. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 373, + 545, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 373, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 308, + 373, + 545, + 417 + ], + "type": "text", + "content": "[16] Vignesh Ganapathi-Subramanian, Olga Diamanti, Soeren Pirk, Chengcheng Tang, Matthias Niessner, and Leonidas Guibas. Parsing geometry using structure-aware shape templates. In 3DV, 2018. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 419, + 545, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 419, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 308, + 419, + 545, + 463 + ], + "type": "text", + "content": "[17] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan Russell, and Mathieu Aubry. AtlasNet: A Papier-Mâché Approach to Learning 3D Surface Generation. In CVPR, 2018. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 467, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 499 + ], + "type": "text", + "content": "[18] Pablo Palafox, Nikolaos Sarafianos, Tony Tung, and Angela Dai. Spams: Structured implicit parametric models. In CVPR, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 502, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 502, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 502, + 545, + 536 + ], + "type": "text", + "content": "[19] Ramana Sundararaman, Gautam Pai, and Maks Ovsjanikov. Implicit field supervision for robust non-rigid shape matching. In ECCV, 2022. 1, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 539, + 545, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 539, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 308, + 539, + 545, + 594 + ], + "type": "text", + "content": "[20] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, 2022. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 597, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 597, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 308, + 597, + 545, + 620 + ], + "type": "text", + "content": "[21] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. 2019. 1, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 622, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 622, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 308, + 622, + 545, + 666 + ], + "type": "text", + "content": "[22] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 1, 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "text", + "content": "[23] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 1, 2, 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "601" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[24] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In ECCV, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 110, + 287, + 153 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 110, + 287, + 153 + ], + "spans": [ + { + "bbox": [ + 48, + 110, + 287, + 153 + ], + "type": "text", + "content": "[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 158, + 287, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 158, + 287, + 202 + ], + "spans": [ + { + "bbox": [ + 48, + 158, + 287, + 202 + ], + "type": "text", + "content": "[26] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 206, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 287, + 250 + ], + "type": "text", + "content": "[27] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 255, + 287, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 287, + 298 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 287, + 298 + ], + "type": "text", + "content": "[28] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 302, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 302, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 302, + 287, + 346 + ], + "type": "text", + "content": "[29] Tarun Yenamandra, Ayush Tewari, Florian Bernard, Hans-Peter Seidel, Mohamed Elgharib, Daniel Cremers, and Christian Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 384 + ], + "type": "text", + "content": "[30] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "type": "text", + "content": "[31] Zhiqin Chen, Kangxue Yin, Matthew Fisher, Siddhartha Chaudhuri, and Hao Zhang. Bae-net: Branched autoencoder for shape co-segmentation. In ICCV, 2019. 2, 4, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 425, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 425, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 425, + 287, + 468 + ], + "type": "text", + "content": "[32] Vladimir G Kim, Wilmot Li, Niloy J Mitra, Siddhartha Chaudhuri, Stephen DiVerdi, and Thomas Funkhouser. Learning part-based templates from large collections of 3d shapes. TOG, 2013. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 473, + 287, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 473, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 48, + 473, + 287, + 516 + ], + "type": "text", + "content": "[33] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T. Freeman, and Thomas A. Funkhouser. Learning shape templates with structured implicit functions. In ICCV, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 521, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 564 + ], + "type": "text", + "content": "[34] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir Kim, Bryan Russell, and Mathieu Aubry. Learning elementary structures for 3d shape generation and matching. In NeurIPS, 2019. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "type": "text", + "content": "[35] Jiahui Lei and Kostas Daniilidis. Cadex: Learning canonical deformation coordinate space for dynamic surface representation via neural homeomorphism. In CVPR, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 606, + 287, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 606, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 48, + 606, + 287, + 639 + ], + "type": "text", + "content": "[36] Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural template: Topology-aware reconstruction and disentangled generation of 3d meshes. In CVPR, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 643, + 287, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 643, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 48, + 643, + 287, + 676 + ], + "type": "text", + "content": "[37] Zekun Hao, Hadar Averbuch-Elor, Noah Snavely, and Serge Belongie. Dualsdf: Semantic shape manipulation using a two-level representation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[38] Yueqi Duan, Haidong Zhu, He Wang, Li Yi, Ram Nevatia, and Leonidas J. Guibas. Curriculum deepsdf. In ECCV, 2020. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[39] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In ICLR, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 110, + 545, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 110, + 545, + 142 + ], + "spans": [ + { + "bbox": [ + 307, + 110, + 545, + 142 + ], + "type": "text", + "content": "[40] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 147, + 545, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 147, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 307, + 147, + 545, + 179 + ], + "type": "text", + "content": "[41] Michael Niemeyer and Andreas Geiger. Campari: Camera-aware decomposed generative neural radiance fields. In 3DV, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 184, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 545, + 239 + ], + "type": "text", + "content": "[42] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, et al. Gaudi: A neural architect for immersive 3d scene generation. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 243, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 277 + ], + "type": "text", + "content": "[43] Chengjie Niu, Manyi Li, Kai Xu, and Hao Zhang. Rimnet: Recursive implicit fields for unsupervised learning of hierarchical shape structures. In CVPR, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 281, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 281, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 281, + 545, + 323 + ], + "type": "text", + "content": "[44] Chenyang Zhu, Kai Xu, Siddhartha Chaudhuri, Li Yi, Leonidas J Guibas, and Hao Zhang. Adacoseg: Adaptive shape co-segmentation with group consistency loss. In CVPR, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 328, + 545, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 328, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 307, + 328, + 545, + 372 + ], + "type": "text", + "content": "[45] Minghua Liu, Yinhao Zhu, Hong Cai, Shizhong Han, Zhan Ling, Fatih Porikli, and Hao Su. Part slip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In CVPR, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "text", + "content": "[46] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 424, + 545, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 424, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 307, + 424, + 545, + 457 + ], + "type": "text", + "content": "[47] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 461, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 461, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 461, + 545, + 495 + ], + "type": "text", + "content": "[48] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 498, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 531 + ], + "type": "text", + "content": "[49] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In CVPR, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 536, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 536, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 536, + 545, + 590 + ], + "type": "text", + "content": "[50] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv:1512.03012, 2015. 4, 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 595, + 545, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 595, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 307, + 595, + 545, + 616 + ], + "type": "text", + "content": "[51] David Ha, Andrew M. Dai, and Quoc V. Le. Hypernetworks. In ICLR, 2017. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 620, + 545, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 620, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 307, + 620, + 545, + 664 + ], + "type": "text", + "content": "[52] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In NeurIPS, 2020. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[53] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 5, 7" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "602" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 323 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[54] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 110, + 287, + 153 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 110, + 287, + 153 + ], + "spans": [ + { + "bbox": [ + 48, + 110, + 287, + 153 + ], + "type": "text", + "content": "[55] Yang You, Yujing Lou, Chengkun Li, Zhoujun Cheng, Liangwei Li, Lizhuang Ma, Cewu Lu, and Weiming Wang. Keypointnet: A large-scale 3d keypoint dataset aggregated from numerous human annotations. In CVPR, 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 157, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 157, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 48, + 157, + 287, + 201 + ], + "type": "text", + "content": "[56] Li Yi, Vladimir G. Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas J. Guibas. A scalable active framework for region annotation in 3d shape collections. TOG, 2016. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 238 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 238 + ], + "type": "text", + "content": "[57] Li Yi, Hao Su, Xingwen Guo, and Leonidas Guibas. Syncspeccnn: Synchronized spectral cnn for 3d shape segmentation. In CVPR, 2017. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 242, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 242, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 242, + 287, + 285 + ], + "type": "text", + "content": "[58] Xinhai Liu, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Fine-grained 3d shape classification with hierarchical part-view attention. IEEE Transactions on Image Processing, 2021. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 290, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 290, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 48, + 290, + 287, + 323 + ], + "type": "text", + "content": "[59] Federica Bogo, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Dynamic FAUST: Registering human bodies in motion. In CVPR, 2017. 8" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "603" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/e3cccf23-d23a-4b30-90e1-8c04059b69dd_content_list.json b/2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/e3cccf23-d23a-4b30-90e1-8c04059b69dd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6d7df41ac0e7788b87d2f8162ed13686d1ad57a2 --- /dev/null +++ b/2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/e3cccf23-d23a-4b30-90e1-8c04059b69dd_content_list.json @@ -0,0 +1,2001 @@ +[ + { + "type": "text", + "text": "Semantically Structured Image Compression via Irregular Group-Based Decoupling", + "text_level": 1, + "bbox": [ + 238, + 130, + 733, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruoyu Feng $^{1*}$ Yixin Gao $^{1*}$ Xin Jin $^{2}$ Runsen Feng $^{1}$ Zhibo Chen $^{1,\\dagger}$", + "bbox": [ + 194, + 202, + 774, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Science and Technology of China 2Eastern Institute of Technology, Ningbo", + "bbox": [ + 210, + 223, + 759, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 273, + 313, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image compression techniques typically focus on compressing rectangular images for human consumption, however, resulting in transmitting redundant content for downstream applications. To overcome this limitation, some previous works propose to semantically structure the bitstream, which can meet specific application requirements by selective transmission and reconstruction. Nevertheless, they divide the input image into multiple rectangular regions according to semantics and ignore avoiding information interaction among them, causing waste of bitrate and distorted reconstruction of region boundaries. In this paper, we propose to decouple an image into multiple groups with irregular shapes based on a customized group mask and compress them independently. Our group mask describes the image at a finer granularity, enabling significant bitrate saving by reducing the transmission of redundant content. Moreover, to ensure the fidelity of selective reconstruction, this paper proposes the concept of group-independent transform that maintain the independence among distinct groups. And we instantiate it by the proposed Group-Independent Swin-Block (GI Swin-Block). Experimental results demonstrate that our framework structures the bitstream with negligible cost, and exhibits superior performance on both visual quality and intelligent task supporting.", + "bbox": [ + 76, + 306, + 473, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 715, + 209, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image signal serves as a critical information carrier for various applications in modern society. Image compression techniques aim at converting images into compact representations (i.e. bitstreams) to save transmission and storage resources. Lossy image compression is one of the most practical techniques, as it allows for the restoration of important content while discarding a small amount of inessential information. In the past decades, traditional image compres", + "bbox": [ + 75, + 741, + 470, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/82944d9ba34b0a08d6b1cec5006054c78948fdbab5baaf9006f2c024e98a9be7.jpg", + "image_caption": [ + "Figure 1. The input image is decoupled into groups according to distinct semantics. Then the semantically structured bitstream (SSB) is generated by compressing the image based on the partitioned groups. The SSB facilitates downstream applications by selective bitstream transmission and partial reconstruction, depending on the specific task requirements." + ], + "image_footnote": [], + "bbox": [ + 506, + 270, + 890, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sion standards [10,54,61,65,67] have been extensively studied and utilized. With the fast development of deep learning, neural image codes [7,8,16,27,34,40,41,50-52,75] rapidly evolved and achieved promising results. Meanwhile, more and more media contents tend to be handled by machine vision algorithms, such as recognition [17,26,29,32,47], detection [43,44,55-57], and segmentation [5,9,11,12,28,46,48,69,73]. However, most compression methods are mainly developed for compressing regular rectangular images for human consumption, without considering the efficiency and functionality for downstream tasks or human-machine interaction scenarios.", + "bbox": [ + 496, + 623, + 893, + 805 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, the field of image coding for machines (ICM) has emerged to develop a joint efficient and analytical framework for supporting intelligent analytics. End-to-end optimization of the trade-off between specific task loss and compression rate is a promising way [4, 13, 14, 18, 19, 38, 39, 49, 60], but it lacks generalization for massively diverse", + "bbox": [ + 496, + 809, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* First two authors contributed equally.", + "bbox": [ + 76, + 875, + 287, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 80, + 888, + 209, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17237", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "applications. To overcome this limitation, Feng et al. [21] propose to compress general and compact features learned by self-supervised learning under entropy constraints for supporting downstream tasks. Nevertheless, it requires retraining the task model with the proposed features as inputs, which makes the overall performance heavily depend on the effectiveness of feature extraction. The methods mentioned above are designed specifically for scenarios of compression for machine vision without considering situations in which human involvement is required. Semantically structured image compression (SSIC) [62] proposes to generate a semantically structured bitstream (SSB) by separately compressing rectangular regions of detected objects using a pre-prepared object detection toolbox. Although SSB is efficient in supporting intelligent tasks and human-machine interaction through partial transmission and reconstruction of the bitstream, its division approach based on rectangular regions may encounter problems with overlapped objects. SSIC addresses this issue by replacing the overlapping objects with a larger rectangular region, which can result in a waste of bitrate. In addition, SSIC generates the bitstream of each object by compressing the corresponding latent variables in the latent domain directly, without considering the interaction and dependence of features during the transform process. This can lead to blurry and distorted group boundaries in the partial reconstruction scenario, which in turn affects the reconstruction quality.", + "bbox": [ + 76, + 90, + 472, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Going beyond the rectangular-based division, this paper proposes to decouple the image into multiple groups with irregular shapes based on a customized group mask. Then the SSB is generated by independently compressing these groups and can support various requirements via selective transmission and reconstruction. Notably, the generation of the group mask offers high flexibility in terms of shapes of groups, means of pre-analysis, and partition criteria, enabling customization to suit diverse application scenarios and requirements. Moreover, to avoid potential quality degradation in the partial reconstruction scenario, we propose the concept of the group-independent transform, which ensures the independence among groups in the latent representations and therefore the quality of the selective reconstruction will not be affected by the absence of other groups. More specifically, we instantiate it by carefully designing the Group-Independent Swin-Block (GI Swin-Block), which is an extension of Swin Transformer [47] tailored to our situation and requirements. GI SwinBlocks make use of the hierarchical modeling capability of the Swin-Transformer [47, 75], achieving high coding efficiency under the premise of group independence. By combining the group mask based decoupling and the Group-Independent Swin-Block, our proposed method can efficiently support various downstream applications including human-machine interaction and machine vision tasks with", + "bbox": [ + 76, + 508, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "only one bitstream generated.", + "bbox": [ + 500, + 90, + 697, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of our approach are summarized as follows:", + "bbox": [ + 500, + 106, + 890, + 135 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose to decouple an image into multiple groups with irregular shapes for structuring the bitstream. Our group mask can describe spatial division at a finer granularity manner than a typical rectangle, saving bitrate by reducing redundant content transmission.", + "- We propose the group-independent transform and instantiate it by carefully designing the Group-Independent Swin-Block (GI Swin-Block), which maintains powerful transformation capability and ensures the independence among groups in the latent representations.", + "- Experimental results demonstrate that our proposed model achieves state-of-the-art compression ability and superior downstream tasks performance, which is a codec with both high compression efficiency and functionality." + ], + "bbox": [ + 516, + 143, + 890, + 401 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 409, + 650, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Compression", + "text_level": 1, + "bbox": [ + 500, + 434, + 689, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional Image Compression. Traditional image compression standards, such as JPEG [65], JPEG2000 [54], HEVC [61], and VVC Intra [10], have been extensively used in practice after several decades of development. These standards rely on transform coding [23], which decomposes the lossy image compression task into three parts: transform, quantization, and entropy coding. Each module of these standards is manually designed with multiple modes, and rate-distortion optimization is performed to determine the optimal mode. However, the completely hand-crafted structure of traditional CODECs limits their flexibility and scalability to support various objectives, such as MS-SSIM and classification accuracy, as they cannot be optimized in an end-to-end manner.", + "bbox": [ + 496, + 460, + 890, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learned Image Compression. In recent years, learned image compression methods based on nonlinear transform coding [6] have achieved rapid progress. Early works in this area concentrated on enabling end-to-end training by developing differential quantization and rate estimation techniques [1, 7, 63]. Subsequently, a considerable amount of work design powerful neural network modules to enhance the transform [16, 75], quantization [25, 70], and entropy model [8, 27, 52, 53]. As a result, some neural image CODECs [16, 24, 27, 52, 53] can achieve comparable or even superior performance to traditional coding standards like HEVC [61] and VVC [10]. Moreover, some works [2, 51, 58] introduce perceptual loss to effectively improve reconstruction quality. However, existing image compression methods focus on compressing the entire image", + "bbox": [ + 496, + 674, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17238", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "without considering selective transmission and reconstruction of the bitstream for arbitrary reconstruction, leading to significant bandwidth waste when addressing downstream applications with different requirements.", + "bbox": [ + 76, + 90, + 468, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Image Coding for Machines", + "text_level": 1, + "bbox": [ + 76, + 162, + 326, + 178 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image coding for machines aims at compress source images to serve downstream tasks, such as recognition [17, 26, 29, 32, 47], detection [43, 44, 55-57], and segmentation [5, 9, 11, 12, 28, 46, 48, 69, 73]. Joint optimization of task loss and bitrate [3, 31, 38, 42, 66] in an end-to-end manner is a natural and promising way. Another direction is compress the features corresponding to downstream tasks [4, 13, 14, 18, 19, 49, 60]. However, these methods tend to be biased towards specific tasks and may lack generalizability to a wide range of intelligent applications. To address the limitation of generalization, a novel representation learning based method is proposed [21] to learn general and compact features by combining contrastive learning with entropy constraints. The learned features are used to replace the original images as the new source for compression and transmission, resulting in a significant improvement in coding efficiency across various intelligent tasks. Nevertheless, this approach necessitates re-training the task model using the proposed features as input, and its overall performance is heavily reliant on the effectiveness of feature extraction. In this paper, we propose a unified compression framework for machine vision support without task-oriented joint training or additional feature learning. Moreover, it can also conduct full or selective reconstruction for human perception according to requirements.", + "bbox": [ + 76, + 186, + 470, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Object-oriented Image Compression", + "text_level": 1, + "bbox": [ + 76, + 574, + 393, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object-oriented image compression is first proposed in MPEG-4 Visual [20, 35, 59] by compressing visual object planes (VOP) with arbitrary shapes, targeting content-based interactivity and scalability for the image compression technique. However, due to the difficulty in annotating the complex and intricate shapes of the target and the lack of actual corresponding downstream applications, MPEG-4 Visual has not been widely used. Recently, with the fast development of deep learning [29, 64], the practical scenarios of image compression for downstream applications have increased rapidly. Sun et al. [33, 62] propose the semantically structured image compression (SSIC) of decoupling the image into objects and compressing the corresponding latent variables independently of the neural image codec to generate the semantically structured bitstream. Nevertheless, the rectangular-based division adopted by SSIC lacks generalization to irregular objects and flexibility for customization. Additionally, selecting the representation of objects directly in the latent space ignores the interaction among elements during the encoding transform process, leading to", + "bbox": [ + 76, + 599, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "significant edge distortion of objects in partial reconstruction. This paper proposes to conduct semantically structured image compression by decoupling the image into irregular groups based on the group mask, which is flexible and customizable. Furthermore, we propose the group-independent transform conducted during both the encoding and decoding processes, which enables efficient partial or complete reconstruction of images under various requirements.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 241, + 591, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 267, + 609, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed method is an efficient and flexible version of semantically structured image compression [62]. It can satisfy multiple application requirements via transmitting and reconstructing partial spatial regions guided by a block-wise group mask. Our model takes the group mask as guidance to ensure that interactions of transform only occur within the same group, thereby achieving group independence during the redundancy removal process. The group mask, which is generated based on the pre-analysis, such as object detection, semantic segmentation, and saliency detection, provides high flexibility and customization for structuring the bitstream and is considered as side information. Then the entropy coding process is conducted on the variables of each group distinctly, resulting in a bitstream structured by semantics. The bitstream can be partially or fully transmitted according to the requirements of the decoder side, and then the receiver conducts entropy decoding on the bitstream and reorganizes the spatial-wise arrangement of latent variables according to the group mask and the group indexes.", + "bbox": [ + 496, + 291, + 890, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall network structure is illustrated in Fig. 2, which incorporates the group mask and GI Swin-Block for guiding and instantiating the group-independent transform, respectively. Section 3.2 and Section 3.3 introduce the group mask and group independent, respectively. The detailed implementation of our network architecture is presented in Section 3.4.", + "bbox": [ + 496, + 593, + 890, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Group Mask", + "text_level": 1, + "bbox": [ + 500, + 710, + 635, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Before compression, the group mask is generated according to the results of pre-analysis techniques such as object detection, instance segmentation, and saliency detection. Figure 3 provides an example of this process. More specifically, for the input image $x$ with a height of $H$ and a width of $W$ , the spatial resolution of the group mask is the same as the input image, and it consists of $\\frac{H}{B} \\times \\frac{W}{B}$ blocks, where $B$ is the side length of each block. And the value of it indicates the index of the corresponding group.", + "bbox": [ + 496, + 734, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed group mask can clearly divide the elements in the latent variable space after the one-pass down", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17239", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d76eb38af4aa73a7eb406dfec3fa55e911ae3f0cb944f421209f8772f5801d39.jpg", + "image_caption": [ + "Figure 2. The network architecture of our proposed model with the channel-wise auto-regressive model (ChARM). ConvT denotes transposed convolution. AE and AD are respectively arithmetic encoding and arithmetic decoding. In Ours-Hyper model, we remove the ChARM component and instead output $\\mu$ and $\\sigma$ directly from the hyperdecoder $h_s$ ." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 345 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d018b1156d0090bc1bf3f1b2206388b15125c66e51427de92ecb2ae225e90d97.jpg", + "image_caption": [ + "Figure 3. An example of group mask generation. The input image is pre-analyzed by object detection, then the group mask is generated based on the results of the pre-analysis." + ], + "image_footnote": [], + "bbox": [ + 91, + 425, + 467, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sampling transform. Besides, downsampling the group mask by a factor of $B$ before conducting compression can greatly reduce the overhead bitrate. The block size and the overhead bitrate have an inverse relationship. Furthermore, as shown in Figure 4, the generation of the group mask can be flexibly customized, in terms of the way of pre-analysis and the criteria of block allocation, etc.", + "bbox": [ + 75, + 606, + 468, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Group-Independent transform", + "text_level": 1, + "bbox": [ + 76, + 724, + 349, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lossy image compression based on transform coding can be divided into three modularized components: transform, quantization, and entropy coding. While quantization and entropy coding do not affect the generation and usage of the semantically structured bitstream (SSB), traditional transform applied to the entire image to remove spatial redundancy inevitably creates inter-group dependencies during compression. In situations of selective transmission and reconstruction, the incomplete representation that results from the reorganization of a subset of all groups can", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lead to inaccurate reconstruction due to the lack of intergroup dependencies. Therefore, we propose the concept of the group-independent transform for semantically structured bitstream generation. The basic idea is to constrain the transform to be conducted only within each group.", + "bbox": [ + 496, + 428, + 890, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A straight and natural way is by using transformer [64] with the customized attention map corresponding to the group mask. However, image compression differs significantly from high-level understanding tasks, which the transformer excels at. The presence of numerous complex long-range dependencies can impede convergence and adversely affect the final performance. In order to achieve high coding efficiency while maintaining the group-independent property, and inspired by the strong hierarchical representation modeling capability of Swin-Transformer [47, 75], we extend the Swin-Block of Swin-Transformer to our proposed group-independent Swin-Block (GI Swin-Block). The GI Swin-Block serves as the core component of the transform. To be more specific, as shown in Figure 5a, by merging the window partition and the group partition into the group-independent window partition, allowing the self-attention to be computed within the partitioned local regions, enabling us to achieve high coding efficiency while maintaining the group-independent property. Additionally, the cross-window connections are introduced by merging the shifted window partition and group partition similarly, as depicted in Figure 5b.", + "bbox": [ + 496, + 503, + 892, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Network architecture and pipeline", + "text_level": 1, + "bbox": [ + 500, + 845, + 802, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 2 illustrate the proposed architecture. The main transform is performed in a staggered manner through", + "bbox": [ + 500, + 869, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17240", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/64e2a7e5d133e8584b985bb8fc4289ce2b243157912d5b5644d5c8bb2a027ef0.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 83, + 97, + 151, + 161 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/847ce4cb612360f0315619348f88a9c1ae9f73c5efdf51227898cba8b0507cf1.jpg", + "image_caption": [ + "Pre-Analysis" + ], + "image_footnote": [], + "bbox": [ + 151, + 98, + 218, + 161 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7da7c82cd4385373807b1a9982a8b99f7f115051fb63e0034447d20768c16861.jpg", + "image_caption": [ + "(a)", + "Possible Group Masks" + ], + "image_footnote": [], + "bbox": [ + 218, + 97, + 308, + 161 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/57afce836be7bf66ddd76947463897e79365a5ec9ab2725cade8d1ad37afb6c0.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 316, + 97, + 383, + 161 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7a899820c4768ae05823d7a91afd1b4c4e267e1306e2d69a2fa71390d363631d.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 390, + 97, + 459, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the incorporation of upsample/downsample and group-independent Swin-Block (GI Swin-Block). Upsample contains a pixel shuffle operation and a convolution layer with $1 \\times 1$ kernel size, whereas downsample contains a pixel unshuffle operation and a convolution layer with $1 \\times 1$ kernel size. For entropy coding, we adopt the Mean & Scale (M&S) Hyperprior model [52] with the strong channel-wise auto-regressive model (ChARM) [53] to predict the probability distribution of latent variables.", + "bbox": [ + 75, + 309, + 468, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, given a source image $\\pmb{x}$ , the encoder $g_{a}$ converts it to latent representations $\\pmb{y}$ conditioned on the given group mask $\\pmb{m} = \\sum_{i=1}^{N} \\pmb{m}_{i}$ , where $i \\in \\{1,2,3,\\dots,N\\}$ , indicating the index of the corresponding group in all $N$ groups. In each $\\pmb{m}_{i}$ , the values of positions corresponding to the $i_{th}$ group are set as $i$ , and all values of other positions are set as 0. Similarly, we can define $\\pmb{x} = \\sum_{i=1}^{N} \\pmb{x}_{i}$ and $\\pmb{y} = \\sum_{i=1}^{N} \\pmb{y}_{i}$ from the perspective of group, and transform only occurs within each group, which is given by", + "bbox": [ + 76, + 445, + 468, + 583 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {i} = g _ {a} \\left(\\boldsymbol {x} _ {i} \\mid \\boldsymbol {m} _ {i}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 590, + 468, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thanks to the group-independent transform, we can easily implement the transform of all groups within one forward-pass, thus, the process can also be written as $\\pmb{y} = g_{a}(\\pmb{x}|\\pmb{m})$ .", + "bbox": [ + 76, + 617, + 468, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then the latents are quantized to discrete representations $\\hat{\\pmb{y}}$ and selectively transmitted to the decoder side with predicted probability distribution through the entropy model. Concretely, the side information $\\pmb{z}$ is extracted by the hyper-encoder $\\pmb{z} = h_{a}(\\pmb{y})$ . The quantized hyper-latent $\\hat{\\pmb{z}} = Q(\\pmb{z})$ is modeled and entropy-coded with a learned factorized prior. In ChARM, latent $\\hat{\\pmb{y}}$ is split into $S$ (we choose $S = 10$ ) slices along with channel dimension, and each slice $\\hat{\\pmb{y}}_s$ is entropy-coded based on the previous slices $\\hat{\\pmb{y}}_{< s}$ . $\\hat{\\pmb{y}}$ is modeled by a conditional Gaussian distribution convolving with a unit uniform noise.", + "bbox": [ + 76, + 674, + 468, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After selectively entropy decoding the quantized latents, we have $\\hat{\\pmb{y}}_p = \\sum_{j\\in \\mathbf{P}}\\hat{\\pmb{y}}_j$ , where $\\mathbf{P}\\subseteq \\{1,2,\\dots ,N\\}$ is the gather of required groups. The reconstruction image $\\hat{x}_p = \\sum_{j\\in \\mathbf{P}}\\hat{x}_j$ will be calculated as follows", + "bbox": [ + 76, + 839, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/260198688b7f50c623cc5d37cb6e3268728877286a4ba52d8e5c7cfbd0f3b894.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 84, + 890, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/29995e04e180ef31d391f4849e88039ea00f178dd36534d0563f50a847c3de6a.jpg", + "image_caption": [ + "Figure 4. Illustration of flexibility to customize the group mask: (a). Taking overlapping objects as one group with the guidance of bounding boxes. (b). Taking overlapping objects as one group with the guidance of instance masks. (c) Taking overlapping objects as distinct groups with the guidance of instance masks.", + "(a) GI window partition under regular window partition.", + "(b) GI window partition under shifted window partition.", + "Figure 5. Group-independent window partition of GI SwinBlock under regular and shifted window partition. Self-attention is conducted inside each local region." + ], + "image_footnote": [], + "bbox": [ + 506, + 202, + 893, + 290 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {x}} _ {j} = g _ {s} \\left(\\hat {\\boldsymbol {y}} _ {j} \\mid \\boldsymbol {m} _ {j}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 386, + 890, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to Equation (1), Equation (2) can also be written as $\\hat{\\pmb{x}} = g_s(\\hat{\\pmb{y}} |\\pmb {m})$ . Note that the rounding operation is non-differentiable, thus we apply uniform noise for learning the entropy model and use a rounded one as the input of the decoder $g_{s}$ as in [53].", + "bbox": [ + 498, + 412, + 890, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 500, + 632, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 500, + 525, + 674, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Codec. We train our high bitrate models $(>= 0.6\\mathrm{bpp})$ on the Flickr 2K dataset, which has the same setting as [74] and train our low bitrate models $(< 0.6\\mathrm{bpp})$ on the COCO 2017 training set [45]. Ours-Hyper are trained with 3.2M iterations, while Ours-ChARM models are trained by initializing their weights from higher bitrate hyperprior models for another 1.2M iterations. Each batch contains 8 random $256 \\times 256$ crops from the training dataset. The learning rate is set as $5e - 5$ and is decayed by a factor of 10 at $2.8M$ iterations. Training loss $L = R + \\beta D$ is the weighted combination of the rate-distortion trade-off, balanced by the Lagrange multiplier $\\beta$ . We adopt the mean squared error (MSE) in the RGB color space as the distortion metric $D$ . The bitrate range of our image codec is achieved by applying different $\\beta$ that $\\beta \\in \\{512, 1024, 2048, 4096\\}$ . During the training stage, we randomly generate the group masks for the training of the group-independent transform, the detailed operations are shown in the supplementary material.", + "bbox": [ + 496, + 550, + 890, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Datasets and Protocol. For the entire image reconstruction quality, we use the widely-used Kodak dataset [37] to evaluate the coding efficiency of our models. We evaluate the selective reconstruction quality on objects", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17241", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9b12c7d4a0e73874705e51a40266a9d7bb47f380cc053c4c9c90bc580acf9213.jpg", + "image_caption": [ + "(a) Entire image reconstruction on Kodak." + ], + "image_footnote": [], + "bbox": [ + 151, + 80, + 486, + 337 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/454acd5af833d1e28fe6c84492629d937cfdd848dcee5678633f682360df2fbb.jpg", + "image_caption": [ + "(b) All categories objects reconstruction on COCO." + ], + "image_footnote": [], + "bbox": [ + 517, + 80, + 839, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/05afc7ad8aab7a1a4c1e804b9e235000091a7895845f4c471278342b1924d8c5.jpg", + "image_caption": [ + "(c) Human category objects reconstruction on COCO." + ], + "image_footnote": [], + "bbox": [ + 519, + 218, + 836, + 337 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/86f7d85cbcfd47b5265e68850ba59a0ba7be47d47417f88613bd42ec786aa812.jpg", + "image_caption": [ + "Figure 6. Rate distortion comparison of compression efficiency on both entire and partial reconstruction scenarios. The PSNR of partial reconstructions is only calculated on the pre-detected bounding boxes." + ], + "image_footnote": [], + "bbox": [ + 86, + 400, + 272, + 511 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/353c8bda89cf67445284c9a35188dad4bc74df5b7b498bec08ef0bae8c6d4a9e.jpg", + "image_caption": [ + "Figure 7. Performance comparison on instance segmentation (left) and pose estimation (right) on COCO." + ], + "image_footnote": [], + "bbox": [ + 277, + 400, + 460, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of all categories and human objects using 40 and 20 images, which are randomly selected from the COCO 2017 validation set. The PSNR of the region of interest serves as the metric for measuring the objective quality. Additionally, the bits per pixel (bpp) are calculated by dividing the total bits of the transmitted bitstream by the number of pixels in the region of interest. To verify the effectiveness of our proposed method in supporting downstream tasks, we conduct experiments on instance segmentation and pose estimation using the COCO 2017 validation set. [45]. This dataset is widely used for dense prediction tasks and comprises 118K training, 5K validation, and 20K test-dev images. We evaluate the performances of instance segmentation and human keypoint detection using Mask R-CNN (X101-FPN backbone) and Keypoint R-CNN (X101-FPN backbone), respectively, implemented in the Detectron2 toolbox [68].", + "bbox": [ + 75, + 555, + 472, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 76, + 800, + 272, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For comparisons, we choose recent neural image codes [16,27,52,75], powerful classical image codes [10, 61], and the functional codec that supports semantically structured bitstream [62]. Among them, ELIC [27] is reproduced and performs closely as their report. It is crit-", + "bbox": [ + 75, + 825, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ical to emphasize that the encoder does not possess prior knowledge regarding the regions of interest required by the decoder. And we utilize object detection as the pre-analysis by default, which is one of the most fundamental visual analysis tools. The semantically structured bitstreams are generated based on the bounding boxes detected by the detector of Mask R-CNN (X101-FPN backbone), and are selectively decoded and reconstructed into images for evaluation according to the requirements of relative semantics. In all experiments of quantitative results, the block size $B$ of group masks is set to 32 to balance the trade-off between performance and overhead bitrate, and we take the union of overlapping groups as one group. More experiments compared with other methods of image coding for machines are presented in the supplementary.", + "bbox": [ + 496, + 400, + 893, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reconstruction Quality. For entire image reconstruction, Figure. 6a demonstrates that our proposed model achieves state-of-the-art rate-distortion performance, surpassing VTM 18.2 in terms of PSNR at all bitrates. Notably, structuring the bitstream according to semantics introduces negligible effects. Therefore, our model has a strong potential to achieve both high coding efficiency and functionality simultaneously. For partial image reconstruction in scenarios of specific interest, as shown in Figure. 6b and Figure. 6c, our model has achieved a significant improvement compared to other CODECs. Specifically, the semantically structured bitstream enabled our model and SSIC [62] to avoid transmitting and decoding bitstreams corresponding to the entire image, which is required of ELIC [27]. Additionally, our model's flexible block-level group mask and group-independent transform spatially decouple latents from the semantic level, resulting in sharper and more re", + "bbox": [ + 496, + 643, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17242", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9a7d6896638011c7da58bfb5d5a4d44d52557be07a841e5bd38414a2e789d8df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 79, + 222, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d26c13f5ecbf86b15f0e1465c92066dad84819ca749f521a992c3f2938c59cc1.jpg", + "image_caption": [ + "0.516/29.36" + ], + "image_footnote": [], + "bbox": [ + 227, + 80, + 357, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b82832be0cdaefe0fb89afa2c13c282dce936f6ee9ef022d4683c3edd1066af1.jpg", + "image_caption": [ + "0.251/29.40" + ], + "image_footnote": [], + "bbox": [ + 364, + 80, + 493, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/91153e25e2b2538a795999db486136efc42bc8df64aa3cf632dee8600fcbb228.jpg", + "image_caption": [ + "0.195 / 30.14" + ], + "image_footnote": [], + "bbox": [ + 501, + 80, + 630, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2321acaaa637c289a5c0a1f8e3f959ffd0cb018b4c55e85bea0cd98459d59d91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 80, + 710, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e1c2e28dc631c2fdcd56c02a5221d6e48d989aa79fe96945ba03b26ca87ddfb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 80, + 766, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f357cb43ff7ddff1b85d23af659ca4e98589f68dcb4b8cd5c7ee436d77d2a67d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 80, + 823, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/60b72b7b5778c6a5edad1e354da5f0f2a811a9de8c487dbf196f0444f14857a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 80, + 880, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6596592032ef0de60f2972318ae404dc93639bf21f1ce1008c80b23f6288bb76.jpg", + "image_caption": [ + "Original Image" + ], + "image_footnote": [], + "bbox": [ + 91, + 161, + 220, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5815dc1ed2654d37d11e6a5ea48fc4b30b7d7b13e331035ac5fda3288bfc736b.jpg", + "image_caption": [ + "0.318/28.61", + "ELIC" + ], + "image_footnote": [], + "bbox": [ + 227, + 161, + 357, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c9a827a04c52c7686e159b55bdd6a0952adfae61c751cdbf42badde15c87a27c.jpg", + "image_caption": [ + "0.295/28.53", + "SSIC" + ], + "image_footnote": [], + "bbox": [ + 364, + 161, + 493, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c8c1e82b2c0da3fcbbfaaee39d9ebac3dd039bccb0aceaf89619198a4d7d54f8.jpg", + "image_caption": [ + "0.316 / 29.25", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 501, + 161, + 630, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f2953d350535e7801e274afa9f8d1063982131c77f25427a099cd4d84f3be2e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 161, + 707, + 193 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/adece8258b93916b6efd26abe44ab7f10348c55c69fafce5b5d6b4d2f710cfce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 194, + 707, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9911aecd9cc119cf8b57d5b021fd2ab6a46f042fb417a95b1bd287df5dc9bbdd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 161, + 766, + 193 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ea4415c651ebfca4edbf6c5af7750d68a43b3fac712ef5bb3eccb1f4c7dfe15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 194, + 766, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/608398d18aaf1880d92b461c09528722f51fa2b0b3bbda7116c4551a529a0d70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 161, + 823, + 193 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ade7db95446bd023e7632e8331c570c6f4ac7f7e4808aec24940599d2122fba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 194, + 823, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c2fc8807996df6a36b79859eeeca1e3317e3151ff8419059fb1432c567730fc3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 161, + 880, + 193 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8415f5019a1d8711b9a137210601a8f997b8e81047da306068fcf626e1d7b406.jpg", + "image_caption": [ + "Figure 8. Visualization of reconstructed images for human consumption from the COCO 2017 dataset, in the scenario where foreground objects are the regions of interest. The numbers below the images correspond to bits per pixel and PSNR." + ], + "image_footnote": [], + "bbox": [ + 828, + 194, + 880, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a13a0777271ce508fd1754ea8677adf56d6b83d632d1b3afe9937524c4a52099.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 300, + 236, + 390 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa1c98fa10d31e1c2136f8e7d61b4021f37b7bae999e9f18f5f1ff57e58abf21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 300, + 398, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/861d3d4a4d8256d62b1215bbe6b9b627198ca1d1994661459f2a6fda843512dc.jpg", + "image_caption": [ + "1.850bpp" + ], + "image_footnote": [], + "bbox": [ + 406, + 300, + 560, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/330b8d1349b7574c9d6833e3b9f55fc3ebf4f4452a1c3e7aa65971fb0e685025.jpg", + "image_caption": [ + "1.592bpp" + ], + "image_footnote": [], + "bbox": [ + 568, + 300, + 723, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/49fa984a24ca38df917f62282ed49a2bea31b9f456fb45145a7541e5c816f5ff.jpg", + "image_caption": [ + "1.211bpp" + ], + "image_footnote": [], + "bbox": [ + 732, + 300, + 883, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/acd3064c9b691bceecd9ea5b338e637633cc619e06e0d49549e35befe6077a76.jpg", + "image_caption": [ + "Original Image", + "Figure 9. Visualization of reconstructed images for downstream tasks from the COCO 2017 dataset. The first line corresponds to instance segmentation, and the second one corresponds to pose estimation." + ], + "image_footnote": [], + "bbox": [ + 81, + 402, + 235, + 474 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1506b4316dbd3ba8017e120cfb41ee1f8d2d8f01e5922c859f08579b8a1222cf.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 243, + 402, + 398, + 474 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f7f548d9eee40c0acab91e4583932cf93071d80c036a465f43df8a5714770586.jpg", + "image_caption": [ + "2.229bpp", + "ELIC" + ], + "image_footnote": [], + "bbox": [ + 406, + 402, + 558, + 474 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d811158090ed31fc98c9a7d750aff7fa179417fb1a258abe2e11091460623ccb.jpg", + "image_caption": [ + "1.335bpp", + "SSIC" + ], + "image_footnote": [], + "bbox": [ + 566, + 402, + 720, + 473 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8f696fbfa4d9872515a66aa0bee837dbf4d4afb5e2e216a55c05aae7b68a03eb.jpg", + "image_caption": [ + "1.048bpp", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 730, + 402, + 883, + 473 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "alistic boundaries compared to SSIC [62]. Moreover, in the case of overlapping regions, our model can significantly save bitrate by grouping them into an irregular group instead of replacing them with a larger bounding box, which would introduce more distortion.", + "bbox": [ + 75, + 541, + 468, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Downstream Task Supporting. Figure. 7 demonstrates our excellent performance on instance segmentation and pose estimation. On instance segmentation, our model achieves significantly improved performance at low bitrates (<0.6bpp) compared with other methods, which can be attributed to the faithful reconstruction of RoI boundaries in our approach. It is worth noting that the superior performance of our method on pose estimation is due to the ability of our model to preserve fine details and boundaries, which are critical for accurately localizing human keypoints. In addition, human objects are a minority in the image and are often sparsely distributed, and compared with SSIC, our method can significantly reduce bitrate by avoiding transmitting the latents corresponding to the rectangular area containing all overlapped objects.", + "bbox": [ + 75, + 619, + 470, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 76, + 849, + 261, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reconstruction for Human Consumption. When the reconstructed images are mainly intended for human con", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sumption with specific semantic interests, as shown in Figure 8, bitstreams generated by CODEs designed for compressing regular rectangular images are required to be fully transmitted and serve full reconstruction, regardless of the requirements and the content of images. Although SSIC [62] partially addresses this issue through generating SSB through bounding box based division, it lacks flexibility and may introduce additional irrelevant disturbance when merging overlapping objects into larger bounding boxes. Moreover, SSIC generates the bitstream by directly dividing the corresponding positions in the latent variable space after the transform based on ConvNets, which can lead to lost dependencies during selective reconstruction, resulting in blurred and distorted region boundaries. With group mask based partitioning and group-independent transform, our method decouples images to generate the SSB more efficiently while ensuring that no distortion or blurring is introduced after selective reconstruction, leading to both significant bitrate saving and a better visual experience.", + "bbox": [ + 496, + 541, + 890, + 829 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reconstruction for Intelligent Tasks. When using reconstructed images for downstream intelligent tasks, selectively transmitting and decoding the semantically structured bitstream based on the prior knowledge of relative seman", + "bbox": [ + 496, + 839, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17243", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/799f42b5c5c2dcb10d91b8b35101445a0edbdc4ea3afb0a026a1b970294ae9bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 80, + 171, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e7fa75907c6ea135bb0420e8e2dd0678a0a05ebfef73509e511ed710500ac633.jpg", + "image_caption": [ + "Original Image" + ], + "image_footnote": [], + "bbox": [ + 81, + 130, + 171, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/37fc29c7ed7d0877f883a4ea8c36c41c4fa9ce64956e6d5946ec7108b3b187eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 80, + 267, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d4db6a3afb325cfeeef3f20e5835f794e0dca570673340f8d777d1a6a7c915ac.jpg", + "image_caption": [ + "Pre-Analysis" + ], + "image_footnote": [], + "bbox": [ + 179, + 130, + 267, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a2799c7805e83a4a04e9f5a0784d05410ae31451129c4fd211fa829470a7017b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 80, + 366, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/43a71bd43fd4e03274e00f62bdb929c436d733b3f0c4338dd5c132687033a308.jpg", + "image_caption": [ + "Based on" + ], + "image_footnote": [], + "bbox": [ + 274, + 130, + 366, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8769c172d99d89cd87ba721a8bb8fcb780119bb35cd6cf6d0b62192c2879d15d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 80, + 464, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f6557947609d733be5376ff9746a43819fdc09b4b7e1962eeb2d422f0efc629.jpg", + "image_caption": [ + "Bounding Boxes", + "Based on", + "Instance Maps" + ], + "image_footnote": [], + "bbox": [ + 372, + 130, + 464, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6fa09b9e4356de1d3cfc882bdc6df10b92f7f6b79b3643749b9d5c70e87784fe.jpg", + "image_caption": [ + "Figure 10. Selective reconstructions from the semantically structured bitstream. The bitstream can be generated based on different pre-analysis methods.", + "Original Image", + "Figure 11. Illustration of semantically-aware encryption based on our proposed method." + ], + "image_footnote": [], + "bbox": [ + 84, + 253, + 171, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b261705210627fc10e0391bf0a3ff55d4f1a540e4577571c0703c93c0f99b029.jpg", + "image_caption": [ + "Pre-Analysis" + ], + "image_footnote": [], + "bbox": [ + 181, + 253, + 267, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0c2f30714811b2af2c34764895014333a29fcf7d3308a8cf5052e087eede41ef.jpg", + "image_caption": [ + "Encrypt People" + ], + "image_footnote": [], + "bbox": [ + 276, + 253, + 362, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b773c9f4052d737dee2ed6239d88df1083153744309efc5282895de23c2fcdcc.jpg", + "image_caption": [ + "Encrypt People", + "and Vehicles" + ], + "image_footnote": [], + "bbox": [ + 372, + 253, + 460, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tics can significantly save bitrate. As illustrated in Figure 9, conventional codecs designed for compressing rectangular images require the entire bitstream to be transmitted and decoded. Downstream models then perform intelligent analytics on the fully reconstructed image, resulting in a waste of bitrate. Similar to image reconstruction for human consumption, SSIC can perform selective transmission to save bitrate. However, its bounding box based partitioning may not be optimal for bitrate savings, and the resulting blurred and distorted region boundaries and irrelevant content can further impede downstream intelligent analytics. (misidentifying the car as a person shown in Figure 9). Our method can more efficiently generate the SSB and support various downstream tasks with specific requirements with both accuracy and coding efficiency.", + "bbox": [ + 75, + 364, + 467, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Customizability and Flexibility. It's flexible to customize the semantically structured bitstream based on different semantical partition criteria. For instance, as shown in Figure 10, the generation and usage of semantically structured bitstreams can be based on either object detection or instance segmentation, depending on the specific requirements. It should be noted that the methods of pre-analysis are not limited to object detection and instance segmentation as shown in Figure 10, but can include other techniques such as saliency detection [22, 30, 72], semantic segmentation [5, 48, 69, 71, 73], panoptic segmentation [15, 36], etc., and even human annotation.", + "bbox": [ + 75, + 595, + 467, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Semantically-Aware Encryption. As presented in Figure 11, our proposed method can be applied to bitstream encryption, enabling selective and even layered encryption with semantic priors based on the user's security level. In some cases, the selectively encrypted SSB can allow for secure transmission and storage of sensitive information while minimizing the impact on visual quality and downstream analytics. Please refer to the supplementary material for", + "bbox": [ + 75, + 780, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/456c61c5f8400fe58ecc3baef621b45b8a82c4c31dc7286b67f525c6784f8833.jpg", + "image_caption": [ + "(a) RoI reconstruction.", + "Figure 12. Ablation study." + ], + "image_footnote": [], + "bbox": [ + 509, + 80, + 687, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c3bcf2a27219e7eb2317217dfe73097f5e1251e12e9589ea6cdc245eef733f9d.jpg", + "image_caption": [ + "(b) Intelligent task supporting." + ], + "image_footnote": [], + "bbox": [ + 702, + 80, + 883, + 190 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "details of implementation.", + "bbox": [ + 500, + 234, + 673, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 266, + 650, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In order to show the effectiveness of our proposed group mask and group-independent transform (GIT) for both RoI-aware selective reconstruction and downstream intelligent analytics, we conducted the ablation study at different bitrates. Specifically, we use people as the region of interest for RoI reconstruction, and the results for taking all foreground objects as RoI are presented in the supplementary material. Additionally, we conducted an ablation study on pose estimation of COCO 2017 for intelligent task support, and the results on instance segmentation are also included in the supplementary material.", + "bbox": [ + 496, + 292, + 890, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Group Mask. Figure 12 shows that group mask based partitioning saves significant bitrate compared with naive bounding box based partitioning. Meanwhile, both the visual quality of RoI reconstruction and downstream task performances would not be impeded.", + "bbox": [ + 496, + 460, + 890, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Group-Independent transform. As shown in Figure 12a, group-independent transform is crucial for the visual quality of RoI reconstruction. And the downstream task performance would be impeded by the blurred and distorted region boundaries without group-independent transform, which is shown in Figure 12b.", + "bbox": [ + 496, + 540, + 890, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 647, + 617, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose to generate semantically structured bitstream with strong functionality based on group masks, which are highly flexible, customizable, while maintaining lightweight overhead. Moreover, we first propose the concept of group-independent transform and instantiate it by designing the Group-Independent Swin-Block (GI-Swin Block) to ensure independence among distinct groups, resulting in pleasing reconstructions of RoI regions. Particularly, our proposed method outperforms VTM-18.2 and achieves comparable coding efficiency with the SOTA neural image CODECs, while offering strong functionality. The experimental results have demonstrated the effectiveness of our proposed method across several applications, including human consumption of regions of interest, downstream task support, and semantically-aware encryption.", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17244", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported in part by NSFC under Grant U1908209, 62021001. This work was also supported in part by ZJNSFC under Grant LQ23F010008.", + "bbox": [ + 76, + 90, + 472, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 165, + 174, + 181 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eirikur Agustsson, Fabian Mentzer, Michael Tschannen, Lukas Cavigelli, Radu Timofte, Luca Benini, and Luc V Gool. Soft-to-hard vector quantization for end-to-end learning compressible representations. *NeurIPS*, 30, 2017. 2", + "[2] Eirikur Agustsson, Michael Tschannen, Fabian Mentzer, Radu Timofte, and Luc Van Gool. Generative adversarial networks for extreme learned image compression. In ICCV, pages 221-231, 2019. 2", + "[3] Mohammad Akbari, Jie Liang, and Jingning Han. Dsslic: Deep semantic segmentation-based layered image compression. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2042-2046. IEEE, 2019. 3", + "[4] Artem Babenko, Anton Slesarev, Alexandr Chigorin, and Victor Lempitsky. Neural codes for image retrieval. In ECCV, pages 584-599. Springer, 2014. 1, 3", + "[5] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. TPAMI, 39(12):2481-2495, 2017. 1, 3, 8", + "[6] Johannes Ballé, Philip A Chou, David Minnen, Saurabh Singh, Nick Johnston, Eirikur Agustsson, Sung Jin Hwang, and George Toderici. Nonlinear transform coding. IEEE Journal of Selected Topics in Signal Processing, 15(2):339-353, 2020. 2", + "[7] Johannes Ballé, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. In *ICLR*, 2017. 1, 2", + "[8] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In ICLR, 2018. 1, 2", + "[9] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In ICCV, pages 9157-9166, 2019. 1, 3", + "[10] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. TCSVT, 2021. 1, 2, 6", + "[11] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 40(4):834-848, 2017. 1, 3", + "[12] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, pages 801-818, 2018. 1, 3", + "[13] Zhuo Chen, Kui Fan, Shiqi Wang, Lingyu Duan, Weisi Lin, and Alex Chichung Kot. Toward intelligent sensing: Intermediate deep feature compression. TIP, 29:2230-2243, 2019. 1, 3" + ], + "bbox": [ + 78, + 189, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Zhuo Chen, Kui Fan, Shiqi Wang, Ling-Yu Duan, Weisi Lin, and Alex Kot. Lossy intermediate deep learning feature compression and evaluation. In ACM MM, pages 2414-2422, 2019. 1, 3", + "[15] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, pages 12475-12485, 2020. 8", + "[16] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In CVPR, pages 7939-7948, 2020. 1, 2, 6", + "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 1, 3", + "[18] Ling-Yu Duan, Vijay Chandrasekhar, Jie Chen, Jie Lin, Zhe Wang, Tiejun Huang, Bernd Girod, and Wen Gao. Overview of thempeg-cdvs standard. TIP, 25(1):179-194, 2015. 1, 3", + "[19] Ling-Yu Duan, Yihang Lou, Yan Bai, Tiejun Huang, Wen Gao, Vijay Chandrasekhar, Jie Lin, Shiqi Wang, and Alex Chichung Kot. Compact descriptors for video analysis: The emerging mpg standard. IEEE MultiMedia, 26(2):44-54, 2018. 1, 3", + "[20] Touradj Ebrahimi and Caspar Horne. Mpeg-4 natural video coding—an overview. Signal Processing: Image Communication, 15(4-5):365-385, 2000. 3", + "[21] Ruoyu Feng, Xin Jin, Zongyu Guo, Runsen Feng, Yixin Gao, Tianyu He, Zhizheng Zhang, Simeng Sun, and Zhibo Chen. Image coding for machines with omnipotent feature learning. In ECCV, pages 510-528. Springer, 2022. 2, 3", + "[22] Stas Goferman, Lihi Zelnik-Manor, and Ayellet Tal. Context-aware saliency detection. TPAMI, 34(10):1915-1926, 2011. 8", + "[23] Vivek K Goyal. Theoretical foundations of transform coding. IEEE Signal Processing Magazine, 18(5):9-21, 2001. 2", + "[24] Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Causal contextual prediction for learned image compression. TCSVT, 32(4):2329-2341, 2021. 2", + "[25] Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Soft then hard: Rethinking the quantization in neural image compression. In ICML, pages 3920-3929. PMLR, 2021. 2", + "[26] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. NeurIPS, 34:15908-15919, 2021. 1, 3", + "[27] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In CVPR, pages 5718-5727, 2022. 1, 2, 6", + "[28] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 1, 3", + "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 3" + ], + "bbox": [ + 501, + 90, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17245", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Xiaodi Hou and Liqing Zhang. Saliency detection: A spectral residual approach. In CVPR, pages 1-8. IEEE, 2007. 8", + "[31] Yueyu Hu, Shuai Yang, Wenhan Yang, Ling-Yu Duan, and Jiaying Liu. Towards coding for human and machine vision: A scalable image coding approach. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 3", + "[32] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. ECCV, 2022. 1, 3", + "[33] Xin Jin, Ruoyu Feng, Simeng Sun, Runsen Feng, Tianyu He, and Zhibo Chen. Semantically video coding: Instill static-dynamic clues into structured bitstream for ai tasks. arXiv preprint arXiv:2201.10162, 2022. 3", + "[34] Nick Johnston, Damien Vincent, David Minnen, Michele Covell, Saurabh Singh, Troy Chinen, Sung Jin Hwang, Joel Shor, and George Toderici. Improved lossy image compression with priming and spatially adaptive bit rates for recurrent networks. In CVPR, pages 4385-4393, 2018. 1", + "[35] Aggelos K Katsaggelos, Lisimachos P Kondi, Fabian W Meier, Jorn Ostermann, and Guido M Schuster. Mpeg-4 and rate-distortion-based shape-coding techniques. Proceedings of the IEEE, 86(6):1126-1154, 1998. 3", + "[36] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, pages 9404-9413, 2019. 8", + "[37] Eastman Kodak. Kodak lossless true color image suite (photocd pcd0992). http://r0k.us/graphics/kodak, 1993.5", + "[38] Nam Le, Honglei Zhang, Francesco Cricri, Ramin GhaznaviYouvalari, and Esa Rahtu. Image coding for machines: An end-to-end learned approach. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1590-1594. IEEE, 2021. 1, 3", + "[39] Nam Le, Honglei Zhang, Francesco Cricri, Ramin Ghaznavi Youvalari, Hamed Rezazadegan Tavakoli, and Esa Rahtu. Learned image coding for machines: A content-adaptive approach. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 1", + "[40] Mu Li, Wangmeng Zuo, Shuhang Gu, Jane You, and David Zhang. Learning content-weighted deep image compression. TPAMI, 2020. 1", + "[41] Mu Li, Wangmeng Zuo, Shuhang Gu, Debin Zhao, and David Zhang. Learning convolutional networks for content-weighted image compression. In CVPR, pages 3214-3223, 2018. 1", + "[42] Xin Li, Jun Shi, and Zhibo Chen. Task-driven semantic coding via reinforcement learning. TIP, 2021. 3", + "[43] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. ECCV, 2022. 1, 3", + "[44] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 1, 3", + "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence" + ], + "bbox": [ + 78, + 92, + 467, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 5, 6", + "[46] Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In CVPR, pages 8759-8768, 2018. 1, 3", + "[47] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 1, 2, 3, 4", + "[48] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 1, 3, 8", + "[49] Siwei Ma, Xiang Zhang, Shiqi Wang, Xinfeng Zhang, Chuanmin Jia, and Shanshe Wang. Joint feature and texture coding: Toward smart video representation via front-end intelligence. TCSVT, 29(10):3095-3105, 2018. 1, 3", + "[50] Fabian Mentzer, Eirikur Agustsson, Michael Tschannen, Radu Timofte, and Luc Van Gool. Conditional probability models for deep image compression. In CVPR, pages 4394-4402, 2018. 1", + "[51] Fabian Mentzer, George Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. arXiv preprint arXiv:2006.09965, 2020. 1, 2", + "[52] David Minnen, Johannes Balle, and George Toderici. Joint autoregressive and hierarchical priors for learned image compression. In NeurIPS, 2018. 1, 2, 5, 6", + "[53] David Minnen and Saurabh Singh. Channel-wise autoregressive entropy models for learned image compression. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3339-3343. IEEE, 2020. 2, 5", + "[54] Majid Rabbani and Rajan Joshi. An overview of the JPEG 2000 still image compression standard. Signal processing: Image communication, 17(1):3-48, 2002. 1, 2", + "[55] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 1, 3", + "[56] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In CVPR, pages 7263-7271, 2017. 1, 3", + "[57] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28:91-99, 2015. 1, 3", + "[58] Oren Rippel and Lubomir Bourdev. Real-time adaptive image compression. In ICML, pages 2922-2930. PMLR, 2017. 2", + "[59] Thomas Sikora. Thempeg-4 video standard verification model. TCSVT, 7(1):19-31, 1997. 3", + "[60] Saurabh Singh, Sami Abu-El-Haija, Nick Johnston, Johannes Balle, Abhinav Shrivastava, and George Toderici. End-to-end learning of compressible features. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3349–3353. IEEE, 2020. 1, 3", + "[61] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. TCSVT, 22(12):1649-1668, 2012. 1, 2, 6", + "[62] Simeng Sun, Tianyu He, and Zhibo Chen. Semantic structured image coding framework for multiple intelligent applications. TCSVT, 2020. 2, 3, 6, 7" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17246", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Lucas Theis, Wenzhe Shi, Andrew Cunningham, and Ferenc Huszár. Lossy image compression with compressive autoencoders. arXiv preprint arXiv:1703.00395, 2017. 2", + "[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 3, 4", + "[65] Gregory K Wallace. The JPEG still picture compression standard. IEEE transactions on consumer electronics, 38(1):xviii-xxxiv, 1992. 1, 2", + "[66] Shurun Wang, Shiqi Wang, Wenhan Yang, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. Towards analysis-friendly face representation with scalable feature and texture compression. TMM, 2021. 3", + "[67] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. TCSVT, 13(7):560-576, 2003. 1", + "[68] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6", + "[69] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. NeurIPS, 34:12077-12090, 2021. 1, 3, 8", + "[70] Yibo Yang, Robert Bamler, and Stephan Mandt. Improving inference for neural image compression. NeurIPS, 33:573-584, 2020. 2", + "[71] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, pages 2881-2890, 2017. 8", + "[72] Rui Zhao, Wanli Ouyang, Hongsheng Li, and Xiaogang Wang. Saliency detection by multi-context deep learning. In CVPR, pages 1265-1274, 2015. 8", + "[73] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. 1, 3, 8", + "[74] Lei Zhou, Zhenhong Sun, Xiangji Wu, and Junmin Wu. End-to-end optimized image compression with attention mechanism. In CVPR workshops, page 0, 2019. 5", + "[75] Yinhao Zhu, Yang Yang, and Taco Cohen. Transformer-based transform coding. In ICLR, 2022. 1, 2, 4, 6" + ], + "bbox": [ + 78, + 90, + 468, + 702 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17247", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/e3cccf23-d23a-4b30-90e1-8c04059b69dd_model.json b/2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/e3cccf23-d23a-4b30-90e1-8c04059b69dd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..00d37d594ff5b647f9a754293625a8dc2f27a38f --- /dev/null +++ b/2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/e3cccf23-d23a-4b30-90e1-8c04059b69dd_model.json @@ -0,0 +1,3203 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.24, + 0.131, + 0.734, + 0.178 + ], + "angle": 0, + "content": "Semantically Structured Image Compression via Irregular Group-Based Decoupling" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.203, + 0.775, + 0.223 + ], + "angle": 0, + "content": "Ruoyu Feng\\(^{1*}\\) Yixin Gao\\(^{1*}\\) Xin Jin\\(^{2}\\) Runsen Feng\\(^{1}\\) Zhibo Chen\\(^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.224, + 0.761, + 0.24 + ], + "angle": 0, + "content": "1University of Science and Technology of China 2Eastern Institute of Technology, Ningbo" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.274, + 0.314, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.307, + 0.474, + 0.685 + ], + "angle": 0, + "content": "Image compression techniques typically focus on compressing rectangular images for human consumption, however, resulting in transmitting redundant content for downstream applications. To overcome this limitation, some previous works propose to semantically structure the bitstream, which can meet specific application requirements by selective transmission and reconstruction. Nevertheless, they divide the input image into multiple rectangular regions according to semantics and ignore avoiding information interaction among them, causing waste of bitrate and distorted reconstruction of region boundaries. In this paper, we propose to decouple an image into multiple groups with irregular shapes based on a customized group mask and compress them independently. Our group mask describes the image at a finer granularity, enabling significant bitrate saving by reducing the transmission of redundant content. Moreover, to ensure the fidelity of selective reconstruction, this paper proposes the concept of group-independent transform that maintain the independence among distinct groups. And we instantiate it by the proposed Group-Independent Swin-Block (GI Swin-Block). Experimental results demonstrate that our framework structures the bitstream with negligible cost, and exhibits superior performance on both visual quality and intelligent task supporting." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.716, + 0.21, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.471, + 0.864 + ], + "angle": 0, + "content": "Image signal serves as a critical information carrier for various applications in modern society. Image compression techniques aim at converting images into compact representations (i.e. bitstreams) to save transmission and storage resources. Lossy image compression is one of the most practical techniques, as it allows for the restoration of important content while discarding a small amount of inessential information. In the past decades, traditional image compres" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.271, + 0.891, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.481, + 0.895, + 0.588 + ], + "angle": 0, + "content": "Figure 1. The input image is decoupled into groups according to distinct semantics. Then the semantically structured bitstream (SSB) is generated by compressing the image based on the partitioned groups. The SSB facilitates downstream applications by selective bitstream transmission and partial reconstruction, depending on the specific task requirements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.895, + 0.806 + ], + "angle": 0, + "content": "sion standards [10,54,61,65,67] have been extensively studied and utilized. With the fast development of deep learning, neural image codes [7,8,16,27,34,40,41,50-52,75] rapidly evolved and achieved promising results. Meanwhile, more and more media contents tend to be handled by machine vision algorithms, such as recognition [17,26,29,32,47], detection [43,44,55-57], and segmentation [5,9,11,12,28,46,48,69,73]. However, most compression methods are mainly developed for compressing regular rectangular images for human consumption, without considering the efficiency and functionality for downstream tasks or human-machine interaction scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Recently, the field of image coding for machines (ICM) has emerged to develop a joint efficient and analytical framework for supporting intelligent analytics. End-to-end optimization of the trade-off between specific task loss and compression rate is a promising way [4, 13, 14, 18, 19, 38, 39, 49, 60], but it lacks generalization for massively diverse" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.288, + 0.888 + ], + "angle": 0, + "content": "* First two authors contributed equally." + }, + { + "type": "page_footnote", + "bbox": [ + 0.081, + 0.889, + 0.21, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.875, + 0.288, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17237" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.501 + ], + "angle": 0, + "content": "applications. To overcome this limitation, Feng et al. [21] propose to compress general and compact features learned by self-supervised learning under entropy constraints for supporting downstream tasks. Nevertheless, it requires retraining the task model with the proposed features as inputs, which makes the overall performance heavily depend on the effectiveness of feature extraction. The methods mentioned above are designed specifically for scenarios of compression for machine vision without considering situations in which human involvement is required. Semantically structured image compression (SSIC) [62] proposes to generate a semantically structured bitstream (SSB) by separately compressing rectangular regions of detected objects using a pre-prepared object detection toolbox. Although SSB is efficient in supporting intelligent tasks and human-machine interaction through partial transmission and reconstruction of the bitstream, its division approach based on rectangular regions may encounter problems with overlapped objects. SSIC addresses this issue by replacing the overlapping objects with a larger rectangular region, which can result in a waste of bitrate. In addition, SSIC generates the bitstream of each object by compressing the corresponding latent variables in the latent domain directly, without considering the interaction and dependence of features during the transform process. This can lead to blurry and distorted group boundaries in the partial reconstruction scenario, which in turn affects the reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.509, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Going beyond the rectangular-based division, this paper proposes to decouple the image into multiple groups with irregular shapes based on a customized group mask. Then the SSB is generated by independently compressing these groups and can support various requirements via selective transmission and reconstruction. Notably, the generation of the group mask offers high flexibility in terms of shapes of groups, means of pre-analysis, and partition criteria, enabling customization to suit diverse application scenarios and requirements. Moreover, to avoid potential quality degradation in the partial reconstruction scenario, we propose the concept of the group-independent transform, which ensures the independence among groups in the latent representations and therefore the quality of the selective reconstruction will not be affected by the absence of other groups. More specifically, we instantiate it by carefully designing the Group-Independent Swin-Block (GI Swin-Block), which is an extension of Swin Transformer [47] tailored to our situation and requirements. GI SwinBlocks make use of the hierarchical modeling capability of the Swin-Transformer [47, 75], achieving high coding efficiency under the premise of group independence. By combining the group mask based decoupling and the Group-Independent Swin-Block, our proposed method can efficiently support various downstream applications including human-machine interaction and machine vision tasks with" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.698, + 0.107 + ], + "angle": 0, + "content": "only one bitstream generated." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.107, + 0.892, + 0.136 + ], + "angle": 0, + "content": "The main contributions of our approach are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.145, + 0.892, + 0.22 + ], + "angle": 0, + "content": "- We propose to decouple an image into multiple groups with irregular shapes for structuring the bitstream. Our group mask can describe spatial division at a finer granularity manner than a typical rectangle, saving bitrate by reducing redundant content transmission." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.228, + 0.892, + 0.318 + ], + "angle": 0, + "content": "- We propose the group-independent transform and instantiate it by carefully designing the Group-Independent Swin-Block (GI Swin-Block), which maintains powerful transformation capability and ensures the independence among groups in the latent representations." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.327, + 0.892, + 0.402 + ], + "angle": 0, + "content": "- Experimental results demonstrate that our proposed model achieves state-of-the-art compression ability and superior downstream tasks performance, which is a codec with both high compression efficiency and functionality." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.145, + 0.892, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.41, + 0.651, + 0.426 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.435, + 0.69, + 0.451 + ], + "angle": 0, + "content": "2.1. Image Compression" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.672 + ], + "angle": 0, + "content": "Traditional Image Compression. Traditional image compression standards, such as JPEG [65], JPEG2000 [54], HEVC [61], and VVC Intra [10], have been extensively used in practice after several decades of development. These standards rely on transform coding [23], which decomposes the lossy image compression task into three parts: transform, quantization, and entropy coding. Each module of these standards is manually designed with multiple modes, and rate-distortion optimization is performed to determine the optimal mode. However, the completely hand-crafted structure of traditional CODECs limits their flexibility and scalability to support various objectives, such as MS-SSIM and classification accuracy, as they cannot be optimized in an end-to-end manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Learned Image Compression. In recent years, learned image compression methods based on nonlinear transform coding [6] have achieved rapid progress. Early works in this area concentrated on enabling end-to-end training by developing differential quantization and rate estimation techniques [1, 7, 63]. Subsequently, a considerable amount of work design powerful neural network modules to enhance the transform [16, 75], quantization [25, 70], and entropy model [8, 27, 52, 53]. As a result, some neural image CODECs [16, 24, 27, 52, 53] can achieve comparable or even superior performance to traditional coding standards like HEVC [61] and VVC [10]. Moreover, some works [2, 51, 58] introduce perceptual loss to effectively improve reconstruction quality. However, existing image compression methods focus on compressing the entire image" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17238" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "without considering selective transmission and reconstruction of the bitstream for arbitrary reconstruction, leading to significant bandwidth waste when addressing downstream applications with different requirements." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.163, + 0.328, + 0.179 + ], + "angle": 0, + "content": "2.2. Image Coding for Machines" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.187, + 0.471, + 0.565 + ], + "angle": 0, + "content": "Image coding for machines aims at compress source images to serve downstream tasks, such as recognition [17, 26, 29, 32, 47], detection [43, 44, 55-57], and segmentation [5, 9, 11, 12, 28, 46, 48, 69, 73]. Joint optimization of task loss and bitrate [3, 31, 38, 42, 66] in an end-to-end manner is a natural and promising way. Another direction is compress the features corresponding to downstream tasks [4, 13, 14, 18, 19, 49, 60]. However, these methods tend to be biased towards specific tasks and may lack generalizability to a wide range of intelligent applications. To address the limitation of generalization, a novel representation learning based method is proposed [21] to learn general and compact features by combining contrastive learning with entropy constraints. The learned features are used to replace the original images as the new source for compression and transmission, resulting in a significant improvement in coding efficiency across various intelligent tasks. Nevertheless, this approach necessitates re-training the task model using the proposed features as input, and its overall performance is heavily reliant on the effectiveness of feature extraction. In this paper, we propose a unified compression framework for machine vision support without task-oriented joint training or additional feature learning. Moreover, it can also conduct full or selective reconstruction for human perception according to requirements." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.575, + 0.394, + 0.592 + ], + "angle": 0, + "content": "2.3. Object-oriented Image Compression" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.6, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Object-oriented image compression is first proposed in MPEG-4 Visual [20, 35, 59] by compressing visual object planes (VOP) with arbitrary shapes, targeting content-based interactivity and scalability for the image compression technique. However, due to the difficulty in annotating the complex and intricate shapes of the target and the lack of actual corresponding downstream applications, MPEG-4 Visual has not been widely used. Recently, with the fast development of deep learning [29, 64], the practical scenarios of image compression for downstream applications have increased rapidly. Sun et al. [33, 62] propose the semantically structured image compression (SSIC) of decoupling the image into objects and compressing the corresponding latent variables independently of the neural image codec to generate the semantically structured bitstream. Nevertheless, the rectangular-based division adopted by SSIC lacks generalization to irregular objects and flexibility for customization. Additionally, selecting the representation of objects directly in the latent space ignores the interaction among elements during the encoding transform process, leading to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "significant edge distortion of objects in partial reconstruction. This paper proposes to conduct semantically structured image compression by decoupling the image into irregular groups based on the group mask, which is flexible and customizable. Furthermore, we propose the group-independent transform conducted during both the encoding and decoding processes, which enables efficient partial or complete reconstruction of images under various requirements." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.242, + 0.593, + 0.258 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.268, + 0.611, + 0.283 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.593 + ], + "angle": 0, + "content": "The proposed method is an efficient and flexible version of semantically structured image compression [62]. It can satisfy multiple application requirements via transmitting and reconstructing partial spatial regions guided by a block-wise group mask. Our model takes the group mask as guidance to ensure that interactions of transform only occur within the same group, thereby achieving group independence during the redundancy removal process. The group mask, which is generated based on the pre-analysis, such as object detection, semantic segmentation, and saliency detection, provides high flexibility and customization for structuring the bitstream and is considered as side information. Then the entropy coding process is conducted on the variables of each group distinctly, resulting in a bitstream structured by semantics. The bitstream can be partially or fully transmitted according to the requirements of the decoder side, and then the receiver conducts entropy decoding on the bitstream and reorganizes the spatial-wise arrangement of latent variables according to the group mask and the group indexes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.892, + 0.7 + ], + "angle": 0, + "content": "The overall network structure is illustrated in Fig. 2, which incorporates the group mask and GI Swin-Block for guiding and instantiating the group-independent transform, respectively. Section 3.2 and Section 3.3 introduce the group mask and group independent, respectively. The detailed implementation of our network architecture is presented in Section 3.4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.637, + 0.727 + ], + "angle": 0, + "content": "3.2. Group Mask" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Before compression, the group mask is generated according to the results of pre-analysis techniques such as object detection, instance segmentation, and saliency detection. Figure 3 provides an example of this process. More specifically, for the input image \\( x \\) with a height of \\( H \\) and a width of \\( W \\), the spatial resolution of the group mask is the same as the input image, and it consists of \\( \\frac{H}{B} \\times \\frac{W}{B} \\) blocks, where \\( B \\) is the side length of each block. And the value of it indicates the index of the corresponding group." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our proposed group mask can clearly divide the elements in the latent variable space after the one-pass down" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17239" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.895, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.356, + 0.895, + 0.404 + ], + "angle": 0, + "content": "Figure 2. The network architecture of our proposed model with the channel-wise auto-regressive model (ChARM). ConvT denotes transposed convolution. AE and AD are respectively arithmetic encoding and arithmetic decoding. In Ours-Hyper model, we remove the ChARM component and instead output \\(\\mu\\) and \\(\\sigma\\) directly from the hyperdecoder \\(h_s\\)." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.426, + 0.468, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.532, + 0.472, + 0.58 + ], + "angle": 0, + "content": "Figure 3. An example of group mask generation. The input image is pre-analyzed by object detection, then the group mask is generated based on the results of the pre-analysis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.607, + 0.47, + 0.713 + ], + "angle": 0, + "content": "sampling transform. Besides, downsampling the group mask by a factor of \\( B \\) before conducting compression can greatly reduce the overhead bitrate. The block size and the overhead bitrate have an inverse relationship. Furthermore, as shown in Figure 4, the generation of the group mask can be flexibly customized, in terms of the way of pre-analysis and the criteria of block allocation, etc." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.726, + 0.35, + 0.743 + ], + "angle": 0, + "content": "3.3. Group-Independent transform" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Lossy image compression based on transform coding can be divided into three modularized components: transform, quantization, and entropy coding. While quantization and entropy coding do not affect the generation and usage of the semantically structured bitstream (SSB), traditional transform applied to the entire image to remove spatial redundancy inevitably creates inter-group dependencies during compression. In situations of selective transmission and reconstruction, the incomplete representation that results from the reorganization of a subset of all groups can" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.892, + 0.504 + ], + "angle": 0, + "content": "lead to inaccurate reconstruction due to the lack of intergroup dependencies. Therefore, we propose the concept of the group-independent transform for semantically structured bitstream generation. The basic idea is to constrain the transform to be conducted only within each group." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.893, + 0.838 + ], + "angle": 0, + "content": "A straight and natural way is by using transformer [64] with the customized attention map corresponding to the group mask. However, image compression differs significantly from high-level understanding tasks, which the transformer excels at. The presence of numerous complex long-range dependencies can impede convergence and adversely affect the final performance. In order to achieve high coding efficiency while maintaining the group-independent property, and inspired by the strong hierarchical representation modeling capability of Swin-Transformer [47, 75], we extend the Swin-Block of Swin-Transformer to our proposed group-independent Swin-Block (GI Swin-Block). The GI Swin-Block serves as the core component of the transform. To be more specific, as shown in Figure 5a, by merging the window partition and the group partition into the group-independent window partition, allowing the self-attention to be computed within the partitioned local regions, enabling us to achieve high coding efficiency while maintaining the group-independent property. Additionally, the cross-window connections are introduced by merging the shifted window partition and group partition similarly, as depicted in Figure 5b." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.803, + 0.865 + ], + "angle": 0, + "content": "3.4. Network architecture and pipeline" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Figure 2 illustrate the proposed architecture. The main transform is performed in a staggered manner through" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17240" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.098, + 0.152, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.166, + 0.147, + 0.176 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.153, + 0.099, + 0.22, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.166, + 0.215, + 0.176 + ], + "angle": 0, + "content": "Pre-Analysis" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.098, + 0.309, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.164, + 0.282, + 0.171 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.098, + 0.385, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.164, + 0.358, + 0.171 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.392, + 0.098, + 0.46, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.164, + 0.431, + 0.171 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.3, + 0.172, + 0.402, + 0.182 + ], + "angle": 0, + "content": "Possible Group Masks" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.193, + 0.47, + 0.284 + ], + "angle": 0, + "content": "Figure 4. Illustration of flexibility to customize the group mask: (a). Taking overlapping objects as one group with the guidance of bounding boxes. (b). Taking overlapping objects as one group with the guidance of instance masks. (c) Taking overlapping objects as distinct groups with the guidance of instance masks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.469, + 0.445 + ], + "angle": 0, + "content": "the incorporation of upsample/downsample and group-independent Swin-Block (GI Swin-Block). Upsample contains a pixel shuffle operation and a convolution layer with \\(1 \\times 1\\) kernel size, whereas downsample contains a pixel unshuffle operation and a convolution layer with \\(1 \\times 1\\) kernel size. For entropy coding, we adopt the Mean & Scale (M&S) Hyperprior model [52] with the strong channel-wise auto-regressive model (ChARM) [53] to predict the probability distribution of latent variables." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.446, + 0.47, + 0.584 + ], + "angle": 0, + "content": "Specifically, given a source image \\(\\pmb{x}\\), the encoder \\(g_{a}\\) converts it to latent representations \\(\\pmb{y}\\) conditioned on the given group mask \\(\\pmb{m} = \\sum_{i=1}^{N} \\pmb{m}_{i}\\), where \\(i \\in \\{1,2,3,\\dots,N\\}\\), indicating the index of the corresponding group in all \\(N\\) groups. In each \\(\\pmb{m}_{i}\\), the values of positions corresponding to the \\(i_{th}\\) group are set as \\(i\\), and all values of other positions are set as 0. Similarly, we can define \\(\\pmb{x} = \\sum_{i=1}^{N} \\pmb{x}_{i}\\) and \\(\\pmb{y} = \\sum_{i=1}^{N} \\pmb{y}_{i}\\) from the perspective of group, and transform only occurs within each group, which is given by" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.592, + 0.469, + 0.608 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {i} = g _ {a} \\left(\\boldsymbol {x} _ {i} \\mid \\boldsymbol {m} _ {i}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.618, + 0.469, + 0.675 + ], + "angle": 0, + "content": "Thanks to the group-independent transform, we can easily implement the transform of all groups within one forward-pass, thus, the process can also be written as \\( \\pmb{y} = g_{a}(\\pmb{x}|\\pmb{m}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.675, + 0.469, + 0.84 + ], + "angle": 0, + "content": "Then the latents are quantized to discrete representations \\(\\hat{\\pmb{y}}\\) and selectively transmitted to the decoder side with predicted probability distribution through the entropy model. Concretely, the side information \\(\\pmb{z}\\) is extracted by the hyper-encoder \\(\\pmb{z} = h_{a}(\\pmb{y})\\). The quantized hyper-latent \\(\\hat{\\pmb{z}} = Q(\\pmb{z})\\) is modeled and entropy-coded with a learned factorized prior. In ChARM, latent \\(\\hat{\\pmb{y}}\\) is split into \\(S\\) (we choose \\(S = 10\\)) slices along with channel dimension, and each slice \\(\\hat{\\pmb{y}}_s\\) is entropy-coded based on the previous slices \\(\\hat{\\pmb{y}}_{< s}\\). \\(\\hat{\\pmb{y}}\\) is modeled by a conditional Gaussian distribution convolving with a unit uniform noise." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.469, + 0.904 + ], + "angle": 0, + "content": "After selectively entropy decoding the quantized latents, we have \\(\\hat{\\pmb{y}}_p = \\sum_{j\\in \\mathbf{P}}\\hat{\\pmb{y}}_j\\) , where \\(\\mathbf{P}\\subseteq \\{1,2,\\dots ,N\\}\\) is the gather of required groups. The reconstruction image \\(\\hat{x}_p = \\sum_{j\\in \\mathbf{P}}\\hat{x}_j\\) will be calculated as follows" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.085, + 0.892, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.183, + 0.862, + 0.197 + ], + "angle": 0, + "content": "(a) GI window partition under regular window partition." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.203, + 0.895, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.299, + 0.862, + 0.313 + ], + "angle": 0, + "content": "(b) GI window partition under shifted window partition." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.325, + 0.892, + 0.371 + ], + "angle": 0, + "content": "Figure 5. Group-independent window partition of GI SwinBlock under regular and shifted window partition. Self-attention is conducted inside each local region." + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.387, + 0.891, + 0.406 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {x}} _ {j} = g _ {s} \\left(\\hat {\\boldsymbol {y}} _ {j} \\mid \\boldsymbol {m} _ {j}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.413, + 0.892, + 0.489 + ], + "angle": 0, + "content": "Similar to Equation (1), Equation (2) can also be written as \\(\\hat{\\pmb{x}} = g_s(\\hat{\\pmb{y}} |\\pmb {m})\\). Note that the rounding operation is non-differentiable, thus we apply uniform noise for learning the entropy model and use a rounded one as the input of the decoder \\(g_{s}\\) as in [53]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.501, + 0.633, + 0.518 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.526, + 0.676, + 0.542 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Image Codec. We train our high bitrate models \\((>= 0.6\\mathrm{bpp})\\) on the Flickr 2K dataset, which has the same setting as [74] and train our low bitrate models \\((< 0.6\\mathrm{bpp})\\) on the COCO 2017 training set [45]. Ours-Hyper are trained with 3.2M iterations, while Ours-ChARM models are trained by initializing their weights from higher bitrate hyperprior models for another 1.2M iterations. Each batch contains 8 random \\(256 \\times 256\\) crops from the training dataset. The learning rate is set as \\(5e - 5\\) and is decayed by a factor of 10 at \\(2.8M\\) iterations. Training loss \\(L = R + \\beta D\\) is the weighted combination of the rate-distortion trade-off, balanced by the Lagrange multiplier \\(\\beta\\). We adopt the mean squared error (MSE) in the RGB color space as the distortion metric \\(D\\). The bitrate range of our image codec is achieved by applying different \\(\\beta\\) that \\(\\beta \\in \\{512, 1024, 2048, 4096\\}\\). During the training stage, we randomly generate the group masks for the training of the group-independent transform, the detailed operations are shown in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation Datasets and Protocol. For the entire image reconstruction quality, we use the widely-used Kodak dataset [37] to evaluate the coding efficiency of our models. We evaluate the selective reconstruction quality on objects" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17241" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.152, + 0.082, + 0.488, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.34, + 0.443, + 0.353 + ], + "angle": 0, + "content": "(a) Entire image reconstruction on Kodak." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.081, + 0.841, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.203, + 0.832, + 0.215 + ], + "angle": 0, + "content": "(b) All categories objects reconstruction on COCO." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.219, + 0.837, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.34, + 0.838, + 0.353 + ], + "angle": 0, + "content": "(c) Human category objects reconstruction on COCO." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.358, + 0.892, + 0.39 + ], + "angle": 0, + "content": "Figure 6. Rate distortion comparison of compression efficiency on both entire and partial reconstruction scenarios. The PSNR of partial reconstructions is only calculated on the pre-detected bounding boxes." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.401, + 0.273, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.401, + 0.461, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.515, + 0.47, + 0.545 + ], + "angle": 0, + "content": "Figure 7. Performance comparison on instance segmentation (left) and pose estimation (right) on COCO." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.556, + 0.473, + 0.799 + ], + "angle": 0, + "content": "of all categories and human objects using 40 and 20 images, which are randomly selected from the COCO 2017 validation set. The PSNR of the region of interest serves as the metric for measuring the objective quality. Additionally, the bits per pixel (bpp) are calculated by dividing the total bits of the transmitted bitstream by the number of pixels in the region of interest. To verify the effectiveness of our proposed method in supporting downstream tasks, we conduct experiments on instance segmentation and pose estimation using the COCO 2017 validation set. [45]. This dataset is widely used for dense prediction tasks and comprises 118K training, 5K validation, and 20K test-dev images. We evaluate the performances of instance segmentation and human keypoint detection using Mask R-CNN (X101-FPN backbone) and Keypoint R-CNN (X101-FPN backbone), respectively, implemented in the Detectron2 toolbox [68]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.273, + 0.817 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.903 + ], + "angle": 0, + "content": "For comparisons, we choose recent neural image codes [16,27,52,75], powerful classical image codes [10, 61], and the functional codec that supports semantically structured bitstream [62]. Among them, ELIC [27] is reproduced and performs closely as their report. It is crit-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.895, + 0.629 + ], + "angle": 0, + "content": "ical to emphasize that the encoder does not possess prior knowledge regarding the regions of interest required by the decoder. And we utilize object detection as the pre-analysis by default, which is one of the most fundamental visual analysis tools. The semantically structured bitstreams are generated based on the bounding boxes detected by the detector of Mask R-CNN (X101-FPN backbone), and are selectively decoded and reconstructed into images for evaluation according to the requirements of relative semantics. In all experiments of quantitative results, the block size \\( B \\) of group masks is set to 32 to balance the trade-off between performance and overhead bitrate, and we take the union of overlapping groups as one group. More experiments compared with other methods of image coding for machines are presented in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Reconstruction Quality. For entire image reconstruction, Figure. 6a demonstrates that our proposed model achieves state-of-the-art rate-distortion performance, surpassing VTM 18.2 in terms of PSNR at all bitrates. Notably, structuring the bitstream according to semantics introduces negligible effects. Therefore, our model has a strong potential to achieve both high coding efficiency and functionality simultaneously. For partial image reconstruction in scenarios of specific interest, as shown in Figure. 6b and Figure. 6c, our model has achieved a significant improvement compared to other CODECs. Specifically, the semantically structured bitstream enabled our model and SSIC [62] to avoid transmitting and decoding bitstreams corresponding to the entire image, which is required of ELIC [27]. Additionally, our model's flexible block-level group mask and group-independent transform spatially decouple latents from the semantic level, resulting in sharper and more re" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17242" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.08, + 0.223, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.081, + 0.358, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.15, + 0.329, + 0.16 + ], + "angle": 0, + "content": "0.516/29.36" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.081, + 0.495, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.397, + 0.15, + 0.467, + 0.16 + ], + "angle": 0, + "content": "0.251/29.40" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.081, + 0.631, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.15, + 0.603, + 0.16 + ], + "angle": 0, + "content": "0.195 / 30.14" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.081, + 0.711, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.081, + 0.767, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.081, + 0.825, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.081, + 0.882, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.162, + 0.222, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.24, + 0.19, + 0.252 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.162, + 0.358, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.228, + 0.328, + 0.239 + ], + "angle": 0, + "content": "0.318/28.61" + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.24, + 0.304, + 0.251 + ], + "angle": 0, + "content": "ELIC" + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.162, + 0.495, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.228, + 0.464, + 0.239 + ], + "angle": 0, + "content": "0.295/28.53" + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.24, + 0.444, + 0.25 + ], + "angle": 0, + "content": "SSIC" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.162, + 0.631, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.228, + 0.601, + 0.239 + ], + "angle": 0, + "content": "0.316 / 29.25" + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.24, + 0.582, + 0.25 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.162, + 0.709, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.195, + 0.709, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.162, + 0.767, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.195, + 0.767, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.162, + 0.824, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.195, + 0.824, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.162, + 0.882, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.195, + 0.881, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.254, + 0.892, + 0.285 + ], + "angle": 0, + "content": "Figure 8. Visualization of reconstructed images for human consumption from the COCO 2017 dataset, in the scenario where foreground objects are the regions of interest. The numbers below the images correspond to bits per pixel and PSNR." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.301, + 0.237, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.301, + 0.4, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.301, + 0.561, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.458, + 0.391, + 0.508, + 0.402 + ], + "angle": 0, + "content": "1.850bpp" + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.301, + 0.724, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.391, + 0.671, + 0.402 + ], + "angle": 0, + "content": "1.592bpp" + }, + { + "type": "image", + "bbox": [ + 0.733, + 0.301, + 0.885, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.391, + 0.832, + 0.402 + ], + "angle": 0, + "content": "1.211bpp" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.403, + 0.236, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.489, + 0.198, + 0.502 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.403, + 0.399, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.489, + 0.353, + 0.501 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.403, + 0.56, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.458, + 0.476, + 0.508, + 0.487 + ], + "angle": 0, + "content": "2.229bpp" + }, + { + "type": "image_caption", + "bbox": [ + 0.472, + 0.489, + 0.496, + 0.499 + ], + "angle": 0, + "content": "ELIC" + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.403, + 0.721, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.476, + 0.671, + 0.487 + ], + "angle": 0, + "content": "1.335bpp" + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.489, + 0.658, + 0.499 + ], + "angle": 0, + "content": "SSIC" + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.403, + 0.885, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.476, + 0.833, + 0.487 + ], + "angle": 0, + "content": "1.048bpp" + }, + { + "type": "image_caption", + "bbox": [ + 0.794, + 0.489, + 0.822, + 0.499 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.505, + 0.892, + 0.535 + ], + "angle": 0, + "content": "Figure 9. Visualization of reconstructed images for downstream tasks from the COCO 2017 dataset. The first line corresponds to instance segmentation, and the second one corresponds to pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.47, + 0.617 + ], + "angle": 0, + "content": "alistic boundaries compared to SSIC [62]. Moreover, in the case of overlapping regions, our model can significantly save bitrate by grouping them into an irregular group instead of replacing them with a larger bounding box, which would introduce more distortion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.471, + 0.847 + ], + "angle": 0, + "content": "Downstream Task Supporting. Figure. 7 demonstrates our excellent performance on instance segmentation and pose estimation. On instance segmentation, our model achieves significantly improved performance at low bitrates (<0.6bpp) compared with other methods, which can be attributed to the faithful reconstruction of RoI boundaries in our approach. It is worth noting that the superior performance of our method on pose estimation is due to the ability of our model to preserve fine details and boundaries, which are critical for accurately localizing human keypoints. In addition, human objects are a minority in the image and are often sparsely distributed, and compared with SSIC, our method can significantly reduce bitrate by avoiding transmitting the latents corresponding to the rectangular area containing all overlapped objects." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.85, + 0.262, + 0.865 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Reconstruction for Human Consumption. When the reconstructed images are mainly intended for human con" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.892, + 0.83 + ], + "angle": 0, + "content": "sumption with specific semantic interests, as shown in Figure 8, bitstreams generated by CODEs designed for compressing regular rectangular images are required to be fully transmitted and serve full reconstruction, regardless of the requirements and the content of images. Although SSIC [62] partially addresses this issue through generating SSB through bounding box based division, it lacks flexibility and may introduce additional irrelevant disturbance when merging overlapping objects into larger bounding boxes. Moreover, SSIC generates the bitstream by directly dividing the corresponding positions in the latent variable space after the transform based on ConvNets, which can lead to lost dependencies during selective reconstruction, resulting in blurred and distorted region boundaries. With group mask based partitioning and group-independent transform, our method decouples images to generate the SSB more efficiently while ensuring that no distortion or blurring is introduced after selective reconstruction, leading to both significant bitrate saving and a better visual experience." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Reconstruction for Intelligent Tasks. When using reconstructed images for downstream intelligent tasks, selectively transmitting and decoding the semantically structured bitstream based on the prior knowledge of relative seman" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17243" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.081, + 0.172, + 0.128 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.131, + 0.172, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.184, + 0.157, + 0.193 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.081, + 0.269, + 0.128 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.131, + 0.269, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.198, + 0.184, + 0.25, + 0.193 + ], + "angle": 0, + "content": "Pre-Analysis" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.081, + 0.367, + 0.128 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.131, + 0.367, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.18, + 0.341, + 0.188 + ], + "angle": 0, + "content": "Based on" + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.189, + 0.355, + 0.197 + ], + "angle": 0, + "content": "Bounding Boxes" + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.081, + 0.465, + 0.128 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.131, + 0.465, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.18, + 0.44, + 0.188 + ], + "angle": 0, + "content": "Based on" + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.189, + 0.45, + 0.197 + ], + "angle": 0, + "content": "Instance Maps" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.2, + 0.468, + 0.246 + ], + "angle": 0, + "content": "Figure 10. Selective reconstructions from the semantically structured bitstream. The bitstream can be generated based on different pre-analysis methods." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.255, + 0.172, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.306, + 0.159, + 0.314 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.255, + 0.268, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.2, + 0.306, + 0.25, + 0.314 + ], + "angle": 0, + "content": "Pre-Analysis" + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.255, + 0.364, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.306, + 0.352, + 0.314 + ], + "angle": 0, + "content": "Encrypt People" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.255, + 0.461, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.303, + 0.448, + 0.311 + ], + "angle": 0, + "content": "Encrypt People" + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.311, + 0.443, + 0.318 + ], + "angle": 0, + "content": "and Vehicles" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.32, + 0.468, + 0.35 + ], + "angle": 0, + "content": "Figure 11. Illustration of semantically-aware encryption based on our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.468, + 0.591 + ], + "angle": 0, + "content": "tics can significantly save bitrate. As illustrated in Figure 9, conventional codecs designed for compressing rectangular images require the entire bitstream to be transmitted and decoded. Downstream models then perform intelligent analytics on the fully reconstructed image, resulting in a waste of bitrate. Similar to image reconstruction for human consumption, SSIC can perform selective transmission to save bitrate. However, its bounding box based partitioning may not be optimal for bitrate savings, and the resulting blurred and distorted region boundaries and irrelevant content can further impede downstream intelligent analytics. (misidentifying the car as a person shown in Figure 9). Our method can more efficiently generate the SSB and support various downstream tasks with specific requirements with both accuracy and coding efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.468, + 0.776 + ], + "angle": 0, + "content": "Customizability and Flexibility. It's flexible to customize the semantically structured bitstream based on different semantical partition criteria. For instance, as shown in Figure 10, the generation and usage of semantically structured bitstreams can be based on either object detection or instance segmentation, depending on the specific requirements. It should be noted that the methods of pre-analysis are not limited to object detection and instance segmentation as shown in Figure 10, but can include other techniques such as saliency detection [22, 30, 72], semantic segmentation [5, 48, 69, 71, 73], panoptic segmentation [15, 36], etc., and even human annotation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Semantically-Aware Encryption. As presented in Figure 11, our proposed method can be applied to bitstream encryption, enabling selective and even layered encryption with semantic priors based on the user's security level. In some cases, the selectively encrypted SSB can allow for secure transmission and storage of sensitive information while minimizing the impact on visual quality and downstream analytics. Please refer to the supplementary material for" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.082, + 0.689, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.194, + 0.667, + 0.206 + ], + "angle": 0, + "content": "(a) RoI reconstruction." + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.082, + 0.885, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.194, + 0.883, + 0.207 + ], + "angle": 0, + "content": "(b) Intelligent task supporting." + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.209, + 0.783, + 0.224 + ], + "angle": 0, + "content": "Figure 12. Ablation study." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.235, + 0.674, + 0.25 + ], + "angle": 0, + "content": "details of implementation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.267, + 0.651, + 0.284 + ], + "angle": 0, + "content": "5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.294, + 0.892, + 0.459 + ], + "angle": 0, + "content": "In order to show the effectiveness of our proposed group mask and group-independent transform (GIT) for both RoI-aware selective reconstruction and downstream intelligent analytics, we conducted the ablation study at different bitrates. Specifically, we use people as the region of interest for RoI reconstruction, and the results for taking all foreground objects as RoI are presented in the supplementary material. Additionally, we conducted an ablation study on pose estimation of COCO 2017 for intelligent task support, and the results on instance segmentation are also included in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.538 + ], + "angle": 0, + "content": "Group Mask. Figure 12 shows that group mask based partitioning saves significant bitrate compared with naive bounding box based partitioning. Meanwhile, both the visual quality of RoI reconstruction and downstream task performances would not be impeded." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.541, + 0.892, + 0.632 + ], + "angle": 0, + "content": "Group-Independent transform. As shown in Figure 12a, group-independent transform is crucial for the visual quality of RoI reconstruction. And the downstream task performance would be impeded by the blurred and distorted region boundaries without group-independent transform, which is shown in Figure 12b." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.648, + 0.619, + 0.664 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose to generate semantically structured bitstream with strong functionality based on group masks, which are highly flexible, customizable, while maintaining lightweight overhead. Moreover, we first propose the concept of group-independent transform and instantiate it by designing the Group-Independent Swin-Block (GI-Swin Block) to ensure independence among distinct groups, resulting in pleasing reconstructions of RoI regions. Particularly, our proposed method outperforms VTM-18.2 and achieves comparable coding efficiency with the SOTA neural image CODECs, while offering strong functionality. The experimental results have demonstrated the effectiveness of our proposed method across several applications, including human consumption of regions of interest, downstream task support, and semantically-aware encryption." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17244" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.152 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported in part by NSFC under Grant U1908209, 62021001. This work was also supported in part by ZJNSFC under Grant LQ23F010008." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.166, + 0.175, + 0.183 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.19, + 0.47, + 0.247 + ], + "angle": 0, + "content": "[1] Eirikur Agustsson, Fabian Mentzer, Michael Tschannen, Lukas Cavigelli, Radu Timofte, Luca Benini, and Luc V Gool. Soft-to-hard vector quantization for end-to-end learning compressible representations. *NeurIPS*, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.249, + 0.472, + 0.303 + ], + "angle": 0, + "content": "[2] Eirikur Agustsson, Michael Tschannen, Fabian Mentzer, Radu Timofte, and Luc Van Gool. Generative adversarial networks for extreme learned image compression. In ICCV, pages 221-231, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.305, + 0.47, + 0.374 + ], + "angle": 0, + "content": "[3] Mohammad Akbari, Jie Liang, and Jingning Han. Dsslic: Deep semantic segmentation-based layered image compression. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2042-2046. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.375, + 0.47, + 0.417 + ], + "angle": 0, + "content": "[4] Artem Babenko, Anton Slesarev, Alexandr Chigorin, and Victor Lempitsky. Neural codes for image retrieval. In ECCV, pages 584-599. Springer, 2014. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.419, + 0.47, + 0.473 + ], + "angle": 0, + "content": "[5] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. TPAMI, 39(12):2481-2495, 2017. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.475, + 0.47, + 0.544 + ], + "angle": 0, + "content": "[6] Johannes Ballé, Philip A Chou, David Minnen, Saurabh Singh, Nick Johnston, Eirikur Agustsson, Sung Jin Hwang, and George Toderici. Nonlinear transform coding. IEEE Journal of Selected Topics in Signal Processing, 15(2):339-353, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.546, + 0.47, + 0.574 + ], + "angle": 0, + "content": "[7] Johannes Ballé, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. In *ICLR*, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.576, + 0.47, + 0.617 + ], + "angle": 0, + "content": "[8] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In ICLR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.619, + 0.47, + 0.659 + ], + "angle": 0, + "content": "[9] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In ICCV, pages 9157-9166, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.661, + 0.47, + 0.716 + ], + "angle": 0, + "content": "[10] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. TCSVT, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.718, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[11] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 40(4):834-848, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.789, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[12] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, pages 801-818, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[13] Zhuo Chen, Kui Fan, Shiqi Wang, Lingyu Duan, Weisi Lin, and Alex Chichung Kot. Toward intelligent sensing: Intermediate deep feature compression. TIP, 29:2230-2243, 2019. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.19, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Zhuo Chen, Kui Fan, Shiqi Wang, Ling-Yu Duan, Weisi Lin, and Alex Kot. Lossy intermediate deep learning feature compression and evaluation. In ACM MM, pages 2414-2422, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.217 + ], + "angle": 0, + "content": "[15] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, pages 12475-12485, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[16] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In CVPR, pages 7939-7948, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[18] Ling-Yu Duan, Vijay Chandrasekhar, Jie Chen, Jie Lin, Zhe Wang, Tiejun Huang, Bernd Girod, and Wen Gao. Overview of thempeg-cdvs standard. TIP, 25(1):179-194, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[19] Ling-Yu Duan, Yihang Lou, Yan Bai, Tiejun Huang, Wen Gao, Vijay Chandrasekhar, Jie Lin, Shiqi Wang, and Alex Chichung Kot. Compact descriptors for video analysis: The emerging mpg standard. IEEE MultiMedia, 26(2):44-54, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[20] Touradj Ebrahimi and Caspar Horne. Mpeg-4 natural video coding—an overview. Signal Processing: Image Communication, 15(4-5):365-385, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[21] Ruoyu Feng, Xin Jin, Zongyu Guo, Runsen Feng, Yixin Gao, Tianyu He, Zhizheng Zhang, Simeng Sun, and Zhibo Chen. Image coding for machines with omnipotent feature learning. In ECCV, pages 510-528. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[22] Stas Goferman, Lihi Zelnik-Manor, and Ayellet Tal. Context-aware saliency detection. TPAMI, 34(10):1915-1926, 2011. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[23] Vivek K Goyal. Theoretical foundations of transform coding. IEEE Signal Processing Magazine, 18(5):9-21, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[24] Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Causal contextual prediction for learned image compression. TCSVT, 32(4):2329-2341, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.664, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[25] Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Soft then hard: Rethinking the quantization in neural image compression. In ICML, pages 3920-3929. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.719, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[26] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. NeurIPS, 34:15908-15919, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[27] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In CVPR, pages 5718-5727, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[28] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17245" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "[30] Xiaodi Hou and Liqing Zhang. Saliency detection: A spectral residual approach. In CVPR, pages 1-8. IEEE, 2007. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.189 + ], + "angle": 0, + "content": "[31] Yueyu Hu, Shuai Yang, Wenhan Yang, Ling-Yu Duan, and Jiaying Liu. Towards coding for human and machine vision: A scalable image coding approach. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.468, + 0.234 + ], + "angle": 0, + "content": "[32] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. ECCV, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.468, + 0.29 + ], + "angle": 0, + "content": "[33] Xin Jin, Ruoyu Feng, Simeng Sun, Runsen Feng, Tianyu He, and Zhibo Chen. Semantically video coding: Instill static-dynamic clues into structured bitstream for ai tasks. arXiv preprint arXiv:2201.10162, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.468, + 0.36 + ], + "angle": 0, + "content": "[34] Nick Johnston, Damien Vincent, David Minnen, Michele Covell, Saurabh Singh, Troy Chinen, Sung Jin Hwang, Joel Shor, and George Toderici. Improved lossy image compression with priming and spatially adaptive bit rates for recurrent networks. In CVPR, pages 4385-4393, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.468, + 0.417 + ], + "angle": 0, + "content": "[35] Aggelos K Katsaggelos, Lisimachos P Kondi, Fabian W Meier, Jorn Ostermann, and Guido M Schuster. Mpeg-4 and rate-distortion-based shape-coding techniques. Proceedings of the IEEE, 86(6):1126-1154, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.468, + 0.46 + ], + "angle": 0, + "content": "[36] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, pages 9404-9413, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.468, + 0.502 + ], + "angle": 0, + "content": "[37] Eastman Kodak. Kodak lossless true color image suite (photocd pcd0992). http://r0k.us/graphics/kodak, 1993.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.468, + 0.573 + ], + "angle": 0, + "content": "[38] Nam Le, Honglei Zhang, Francesco Cricri, Ramin GhaznaviYouvalari, and Esa Rahtu. Image coding for machines: An end-to-end learned approach. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1590-1594. IEEE, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.468, + 0.644 + ], + "angle": 0, + "content": "[39] Nam Le, Honglei Zhang, Francesco Cricri, Ramin Ghaznavi Youvalari, Hamed Rezazadegan Tavakoli, and Esa Rahtu. Learned image coding for machines: A content-adaptive approach. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.468, + 0.685 + ], + "angle": 0, + "content": "[40] Mu Li, Wangmeng Zuo, Shuhang Gu, Jane You, and David Zhang. Learning content-weighted deep image compression. TPAMI, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.468, + 0.742 + ], + "angle": 0, + "content": "[41] Mu Li, Wangmeng Zuo, Shuhang Gu, Debin Zhao, and David Zhang. Learning convolutional networks for content-weighted image compression. In CVPR, pages 3214-3223, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.468, + 0.772 + ], + "angle": 0, + "content": "[42] Xin Li, Jun Shi, and Zhibo Chen. Task-driven semantic coding via reinforcement learning. TIP, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.468, + 0.814 + ], + "angle": 0, + "content": "[43] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. ECCV, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.468, + 0.87 + ], + "angle": 0, + "content": "[44] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[46] Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In CVPR, pages 8759-8768, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.164, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[47] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.22, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[48] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[49] Siwei Ma, Xiang Zhang, Shiqi Wang, Xinfeng Zhang, Chuanmin Jia, and Shanshe Wang. Joint feature and texture coding: Toward smart video representation via front-end intelligence. TCSVT, 29(10):3095-3105, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[50] Fabian Mentzer, Eirikur Agustsson, Michael Tschannen, Radu Timofte, and Luc Van Gool. Conditional probability models for deep image compression. In CVPR, pages 4394-4402, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[51] Fabian Mentzer, George Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. arXiv preprint arXiv:2006.09965, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[52] David Minnen, Johannes Balle, and George Toderici. Joint autoregressive and hierarchical priors for learned image compression. In NeurIPS, 2018. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.457, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[53] David Minnen and Saurabh Singh. Channel-wise autoregressive entropy models for learned image compression. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3339-3343. IEEE, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[54] Majid Rabbani and Rajan Joshi. An overview of the JPEG 2000 still image compression standard. Signal processing: Image communication, 17(1):3-48, 2002. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[55] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.596, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[56] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In CVPR, pages 7263-7271, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.624, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[57] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28:91-99, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[58] Oren Rippel and Lubomir Bourdev. Real-time adaptive image compression. In ICML, pages 2922-2930. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[59] Thomas Sikora. Thempeg-4 video standard verification model. TCSVT, 7(1):19-31, 1997. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.736, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[60] Saurabh Singh, Sami Abu-El-Haija, Nick Johnston, Johannes Balle, Abhinav Shrivastava, and George Toderici. End-to-end learning of compressible features. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3349–3353. IEEE, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[61] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. TCSVT, 22(12):1649-1668, 2012. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[62] Simeng Sun, Tianyu He, and Zhibo Chen. Semantic structured image coding framework for multiple intelligent applications. TCSVT, 2020. 2, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17246" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[63] Lucas Theis, Wenzhe Shi, Andrew Cunningham, and Ferenc Huszár. Lossy image compression with compressive autoencoders. arXiv preprint arXiv:1703.00395, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[65] Gregory K Wallace. The JPEG still picture compression standard. IEEE transactions on consumer electronics, 38(1):xviii-xxxiv, 1992. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.469, + 0.29 + ], + "angle": 0, + "content": "[66] Shurun Wang, Shiqi Wang, Wenhan Yang, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. Towards analysis-friendly face representation with scalable feature and texture compression. TMM, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[67] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. TCSVT, 13(7):560-576, 2003. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.335, + 0.468, + 0.375 + ], + "angle": 0, + "content": "[68] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[69] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. NeurIPS, 34:12077-12090, 2021. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.474 + ], + "angle": 0, + "content": "[70] Yibo Yang, Robert Bamler, and Stephan Mandt. Improving inference for neural image compression. NeurIPS, 33:573-584, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.469, + 0.518 + ], + "angle": 0, + "content": "[71] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, pages 2881-2890, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.52, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[72] Rui Zhao, Wanli Ouyang, Hongsheng Li, and Xiaogang Wang. Saliency detection by multi-context deep learning. In CVPR, pages 1265-1274, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[73] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.674 + ], + "angle": 0, + "content": "[74] Lei Zhou, Zhenhong Sun, Xiangji Wu, and Junmin Wu. End-to-end optimized image compression with attention mechanism. In CVPR workshops, page 0, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[75] Yinhao Zhu, Yang Yang, and Taco Cohen. Transformer-based transform coding. In ICLR, 2022. 1, 2, 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.703 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17247" + } + ] +] \ No newline at end of file diff --git a/2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/e3cccf23-d23a-4b30-90e1-8c04059b69dd_origin.pdf b/2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/e3cccf23-d23a-4b30-90e1-8c04059b69dd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d2d0dad3d5afa1043388b2a8a7a7e7568bf92c09 --- /dev/null +++ b/2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/e3cccf23-d23a-4b30-90e1-8c04059b69dd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4251bfefca76eacf3a242926a64e9843bea1312e77831c613554e6e02de7a1de +size 6743904 diff --git a/2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/full.md b/2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..78176c73f90cfdfdc84fab91776f3f662b3d6a33 --- /dev/null +++ b/2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/full.md @@ -0,0 +1,407 @@ +# Semantically Structured Image Compression via Irregular Group-Based Decoupling + +Ruoyu Feng $^{1*}$ Yixin Gao $^{1*}$ Xin Jin $^{2}$ Runsen Feng $^{1}$ Zhibo Chen $^{1,\dagger}$ + +1University of Science and Technology of China 2Eastern Institute of Technology, Ningbo + +# Abstract + +Image compression techniques typically focus on compressing rectangular images for human consumption, however, resulting in transmitting redundant content for downstream applications. To overcome this limitation, some previous works propose to semantically structure the bitstream, which can meet specific application requirements by selective transmission and reconstruction. Nevertheless, they divide the input image into multiple rectangular regions according to semantics and ignore avoiding information interaction among them, causing waste of bitrate and distorted reconstruction of region boundaries. In this paper, we propose to decouple an image into multiple groups with irregular shapes based on a customized group mask and compress them independently. Our group mask describes the image at a finer granularity, enabling significant bitrate saving by reducing the transmission of redundant content. Moreover, to ensure the fidelity of selective reconstruction, this paper proposes the concept of group-independent transform that maintain the independence among distinct groups. And we instantiate it by the proposed Group-Independent Swin-Block (GI Swin-Block). Experimental results demonstrate that our framework structures the bitstream with negligible cost, and exhibits superior performance on both visual quality and intelligent task supporting. + +# 1. Introduction + +Image signal serves as a critical information carrier for various applications in modern society. Image compression techniques aim at converting images into compact representations (i.e. bitstreams) to save transmission and storage resources. Lossy image compression is one of the most practical techniques, as it allows for the restoration of important content while discarding a small amount of inessential information. In the past decades, traditional image compres + +![](images/82944d9ba34b0a08d6b1cec5006054c78948fdbab5baaf9006f2c024e98a9be7.jpg) +Figure 1. The input image is decoupled into groups according to distinct semantics. Then the semantically structured bitstream (SSB) is generated by compressing the image based on the partitioned groups. The SSB facilitates downstream applications by selective bitstream transmission and partial reconstruction, depending on the specific task requirements. + +sion standards [10,54,61,65,67] have been extensively studied and utilized. With the fast development of deep learning, neural image codes [7,8,16,27,34,40,41,50-52,75] rapidly evolved and achieved promising results. Meanwhile, more and more media contents tend to be handled by machine vision algorithms, such as recognition [17,26,29,32,47], detection [43,44,55-57], and segmentation [5,9,11,12,28,46,48,69,73]. However, most compression methods are mainly developed for compressing regular rectangular images for human consumption, without considering the efficiency and functionality for downstream tasks or human-machine interaction scenarios. + +Recently, the field of image coding for machines (ICM) has emerged to develop a joint efficient and analytical framework for supporting intelligent analytics. End-to-end optimization of the trade-off between specific task loss and compression rate is a promising way [4, 13, 14, 18, 19, 38, 39, 49, 60], but it lacks generalization for massively diverse + +applications. To overcome this limitation, Feng et al. [21] propose to compress general and compact features learned by self-supervised learning under entropy constraints for supporting downstream tasks. Nevertheless, it requires retraining the task model with the proposed features as inputs, which makes the overall performance heavily depend on the effectiveness of feature extraction. The methods mentioned above are designed specifically for scenarios of compression for machine vision without considering situations in which human involvement is required. Semantically structured image compression (SSIC) [62] proposes to generate a semantically structured bitstream (SSB) by separately compressing rectangular regions of detected objects using a pre-prepared object detection toolbox. Although SSB is efficient in supporting intelligent tasks and human-machine interaction through partial transmission and reconstruction of the bitstream, its division approach based on rectangular regions may encounter problems with overlapped objects. SSIC addresses this issue by replacing the overlapping objects with a larger rectangular region, which can result in a waste of bitrate. In addition, SSIC generates the bitstream of each object by compressing the corresponding latent variables in the latent domain directly, without considering the interaction and dependence of features during the transform process. This can lead to blurry and distorted group boundaries in the partial reconstruction scenario, which in turn affects the reconstruction quality. + +Going beyond the rectangular-based division, this paper proposes to decouple the image into multiple groups with irregular shapes based on a customized group mask. Then the SSB is generated by independently compressing these groups and can support various requirements via selective transmission and reconstruction. Notably, the generation of the group mask offers high flexibility in terms of shapes of groups, means of pre-analysis, and partition criteria, enabling customization to suit diverse application scenarios and requirements. Moreover, to avoid potential quality degradation in the partial reconstruction scenario, we propose the concept of the group-independent transform, which ensures the independence among groups in the latent representations and therefore the quality of the selective reconstruction will not be affected by the absence of other groups. More specifically, we instantiate it by carefully designing the Group-Independent Swin-Block (GI Swin-Block), which is an extension of Swin Transformer [47] tailored to our situation and requirements. GI SwinBlocks make use of the hierarchical modeling capability of the Swin-Transformer [47, 75], achieving high coding efficiency under the premise of group independence. By combining the group mask based decoupling and the Group-Independent Swin-Block, our proposed method can efficiently support various downstream applications including human-machine interaction and machine vision tasks with + +only one bitstream generated. + +The main contributions of our approach are summarized as follows: + +- We propose to decouple an image into multiple groups with irregular shapes for structuring the bitstream. Our group mask can describe spatial division at a finer granularity manner than a typical rectangle, saving bitrate by reducing redundant content transmission. +- We propose the group-independent transform and instantiate it by carefully designing the Group-Independent Swin-Block (GI Swin-Block), which maintains powerful transformation capability and ensures the independence among groups in the latent representations. +- Experimental results demonstrate that our proposed model achieves state-of-the-art compression ability and superior downstream tasks performance, which is a codec with both high compression efficiency and functionality. + +# 2. Related Works + +# 2.1. Image Compression + +Traditional Image Compression. Traditional image compression standards, such as JPEG [65], JPEG2000 [54], HEVC [61], and VVC Intra [10], have been extensively used in practice after several decades of development. These standards rely on transform coding [23], which decomposes the lossy image compression task into three parts: transform, quantization, and entropy coding. Each module of these standards is manually designed with multiple modes, and rate-distortion optimization is performed to determine the optimal mode. However, the completely hand-crafted structure of traditional CODECs limits their flexibility and scalability to support various objectives, such as MS-SSIM and classification accuracy, as they cannot be optimized in an end-to-end manner. + +Learned Image Compression. In recent years, learned image compression methods based on nonlinear transform coding [6] have achieved rapid progress. Early works in this area concentrated on enabling end-to-end training by developing differential quantization and rate estimation techniques [1, 7, 63]. Subsequently, a considerable amount of work design powerful neural network modules to enhance the transform [16, 75], quantization [25, 70], and entropy model [8, 27, 52, 53]. As a result, some neural image CODECs [16, 24, 27, 52, 53] can achieve comparable or even superior performance to traditional coding standards like HEVC [61] and VVC [10]. Moreover, some works [2, 51, 58] introduce perceptual loss to effectively improve reconstruction quality. However, existing image compression methods focus on compressing the entire image + +without considering selective transmission and reconstruction of the bitstream for arbitrary reconstruction, leading to significant bandwidth waste when addressing downstream applications with different requirements. + +# 2.2. Image Coding for Machines + +Image coding for machines aims at compress source images to serve downstream tasks, such as recognition [17, 26, 29, 32, 47], detection [43, 44, 55-57], and segmentation [5, 9, 11, 12, 28, 46, 48, 69, 73]. Joint optimization of task loss and bitrate [3, 31, 38, 42, 66] in an end-to-end manner is a natural and promising way. Another direction is compress the features corresponding to downstream tasks [4, 13, 14, 18, 19, 49, 60]. However, these methods tend to be biased towards specific tasks and may lack generalizability to a wide range of intelligent applications. To address the limitation of generalization, a novel representation learning based method is proposed [21] to learn general and compact features by combining contrastive learning with entropy constraints. The learned features are used to replace the original images as the new source for compression and transmission, resulting in a significant improvement in coding efficiency across various intelligent tasks. Nevertheless, this approach necessitates re-training the task model using the proposed features as input, and its overall performance is heavily reliant on the effectiveness of feature extraction. In this paper, we propose a unified compression framework for machine vision support without task-oriented joint training or additional feature learning. Moreover, it can also conduct full or selective reconstruction for human perception according to requirements. + +# 2.3. Object-oriented Image Compression + +Object-oriented image compression is first proposed in MPEG-4 Visual [20, 35, 59] by compressing visual object planes (VOP) with arbitrary shapes, targeting content-based interactivity and scalability for the image compression technique. However, due to the difficulty in annotating the complex and intricate shapes of the target and the lack of actual corresponding downstream applications, MPEG-4 Visual has not been widely used. Recently, with the fast development of deep learning [29, 64], the practical scenarios of image compression for downstream applications have increased rapidly. Sun et al. [33, 62] propose the semantically structured image compression (SSIC) of decoupling the image into objects and compressing the corresponding latent variables independently of the neural image codec to generate the semantically structured bitstream. Nevertheless, the rectangular-based division adopted by SSIC lacks generalization to irregular objects and flexibility for customization. Additionally, selecting the representation of objects directly in the latent space ignores the interaction among elements during the encoding transform process, leading to + +significant edge distortion of objects in partial reconstruction. This paper proposes to conduct semantically structured image compression by decoupling the image into irregular groups based on the group mask, which is flexible and customizable. Furthermore, we propose the group-independent transform conducted during both the encoding and decoding processes, which enables efficient partial or complete reconstruction of images under various requirements. + +# 3. Method + +# 3.1. Overview + +The proposed method is an efficient and flexible version of semantically structured image compression [62]. It can satisfy multiple application requirements via transmitting and reconstructing partial spatial regions guided by a block-wise group mask. Our model takes the group mask as guidance to ensure that interactions of transform only occur within the same group, thereby achieving group independence during the redundancy removal process. The group mask, which is generated based on the pre-analysis, such as object detection, semantic segmentation, and saliency detection, provides high flexibility and customization for structuring the bitstream and is considered as side information. Then the entropy coding process is conducted on the variables of each group distinctly, resulting in a bitstream structured by semantics. The bitstream can be partially or fully transmitted according to the requirements of the decoder side, and then the receiver conducts entropy decoding on the bitstream and reorganizes the spatial-wise arrangement of latent variables according to the group mask and the group indexes. + +The overall network structure is illustrated in Fig. 2, which incorporates the group mask and GI Swin-Block for guiding and instantiating the group-independent transform, respectively. Section 3.2 and Section 3.3 introduce the group mask and group independent, respectively. The detailed implementation of our network architecture is presented in Section 3.4. + +# 3.2. Group Mask + +Before compression, the group mask is generated according to the results of pre-analysis techniques such as object detection, instance segmentation, and saliency detection. Figure 3 provides an example of this process. More specifically, for the input image $x$ with a height of $H$ and a width of $W$ , the spatial resolution of the group mask is the same as the input image, and it consists of $\frac{H}{B} \times \frac{W}{B}$ blocks, where $B$ is the side length of each block. And the value of it indicates the index of the corresponding group. + +Our proposed group mask can clearly divide the elements in the latent variable space after the one-pass down + +![](images/d76eb38af4aa73a7eb406dfec3fa55e911ae3f0cb944f421209f8772f5801d39.jpg) +Figure 2. The network architecture of our proposed model with the channel-wise auto-regressive model (ChARM). ConvT denotes transposed convolution. AE and AD are respectively arithmetic encoding and arithmetic decoding. In Ours-Hyper model, we remove the ChARM component and instead output $\mu$ and $\sigma$ directly from the hyperdecoder $h_s$ . + +![](images/d018b1156d0090bc1bf3f1b2206388b15125c66e51427de92ecb2ae225e90d97.jpg) +Figure 3. An example of group mask generation. The input image is pre-analyzed by object detection, then the group mask is generated based on the results of the pre-analysis. + +sampling transform. Besides, downsampling the group mask by a factor of $B$ before conducting compression can greatly reduce the overhead bitrate. The block size and the overhead bitrate have an inverse relationship. Furthermore, as shown in Figure 4, the generation of the group mask can be flexibly customized, in terms of the way of pre-analysis and the criteria of block allocation, etc. + +# 3.3. Group-Independent transform + +Lossy image compression based on transform coding can be divided into three modularized components: transform, quantization, and entropy coding. While quantization and entropy coding do not affect the generation and usage of the semantically structured bitstream (SSB), traditional transform applied to the entire image to remove spatial redundancy inevitably creates inter-group dependencies during compression. In situations of selective transmission and reconstruction, the incomplete representation that results from the reorganization of a subset of all groups can + +lead to inaccurate reconstruction due to the lack of intergroup dependencies. Therefore, we propose the concept of the group-independent transform for semantically structured bitstream generation. The basic idea is to constrain the transform to be conducted only within each group. + +A straight and natural way is by using transformer [64] with the customized attention map corresponding to the group mask. However, image compression differs significantly from high-level understanding tasks, which the transformer excels at. The presence of numerous complex long-range dependencies can impede convergence and adversely affect the final performance. In order to achieve high coding efficiency while maintaining the group-independent property, and inspired by the strong hierarchical representation modeling capability of Swin-Transformer [47, 75], we extend the Swin-Block of Swin-Transformer to our proposed group-independent Swin-Block (GI Swin-Block). The GI Swin-Block serves as the core component of the transform. To be more specific, as shown in Figure 5a, by merging the window partition and the group partition into the group-independent window partition, allowing the self-attention to be computed within the partitioned local regions, enabling us to achieve high coding efficiency while maintaining the group-independent property. Additionally, the cross-window connections are introduced by merging the shifted window partition and group partition similarly, as depicted in Figure 5b. + +# 3.4. Network architecture and pipeline + +Figure 2 illustrate the proposed architecture. The main transform is performed in a staggered manner through + +![](images/64e2a7e5d133e8584b985bb8fc4289ce2b243157912d5b5644d5c8bb2a027ef0.jpg) +Input Image + +![](images/847ce4cb612360f0315619348f88a9c1ae9f73c5efdf51227898cba8b0507cf1.jpg) +Pre-Analysis + +![](images/7da7c82cd4385373807b1a9982a8b99f7f115051fb63e0034447d20768c16861.jpg) +(a) +Possible Group Masks + +![](images/57afce836be7bf66ddd76947463897e79365a5ec9ab2725cade8d1ad37afb6c0.jpg) +(b) + +![](images/7a899820c4768ae05823d7a91afd1b4c4e267e1306e2d69a2fa71390d363631d.jpg) +(c) + +the incorporation of upsample/downsample and group-independent Swin-Block (GI Swin-Block). Upsample contains a pixel shuffle operation and a convolution layer with $1 \times 1$ kernel size, whereas downsample contains a pixel unshuffle operation and a convolution layer with $1 \times 1$ kernel size. For entropy coding, we adopt the Mean & Scale (M&S) Hyperprior model [52] with the strong channel-wise auto-regressive model (ChARM) [53] to predict the probability distribution of latent variables. + +Specifically, given a source image $\pmb{x}$ , the encoder $g_{a}$ converts it to latent representations $\pmb{y}$ conditioned on the given group mask $\pmb{m} = \sum_{i=1}^{N} \pmb{m}_{i}$ , where $i \in \{1,2,3,\dots,N\}$ , indicating the index of the corresponding group in all $N$ groups. In each $\pmb{m}_{i}$ , the values of positions corresponding to the $i_{th}$ group are set as $i$ , and all values of other positions are set as 0. Similarly, we can define $\pmb{x} = \sum_{i=1}^{N} \pmb{x}_{i}$ and $\pmb{y} = \sum_{i=1}^{N} \pmb{y}_{i}$ from the perspective of group, and transform only occurs within each group, which is given by + +$$ +\boldsymbol {y} _ {i} = g _ {a} \left(\boldsymbol {x} _ {i} \mid \boldsymbol {m} _ {i}\right). \tag {1} +$$ + +Thanks to the group-independent transform, we can easily implement the transform of all groups within one forward-pass, thus, the process can also be written as $\pmb{y} = g_{a}(\pmb{x}|\pmb{m})$ . + +Then the latents are quantized to discrete representations $\hat{\pmb{y}}$ and selectively transmitted to the decoder side with predicted probability distribution through the entropy model. Concretely, the side information $\pmb{z}$ is extracted by the hyper-encoder $\pmb{z} = h_{a}(\pmb{y})$ . The quantized hyper-latent $\hat{\pmb{z}} = Q(\pmb{z})$ is modeled and entropy-coded with a learned factorized prior. In ChARM, latent $\hat{\pmb{y}}$ is split into $S$ (we choose $S = 10$ ) slices along with channel dimension, and each slice $\hat{\pmb{y}}_s$ is entropy-coded based on the previous slices $\hat{\pmb{y}}_{< s}$ . $\hat{\pmb{y}}$ is modeled by a conditional Gaussian distribution convolving with a unit uniform noise. + +After selectively entropy decoding the quantized latents, we have $\hat{\pmb{y}}_p = \sum_{j\in \mathbf{P}}\hat{\pmb{y}}_j$ , where $\mathbf{P}\subseteq \{1,2,\dots ,N\}$ is the gather of required groups. The reconstruction image $\hat{x}_p = \sum_{j\in \mathbf{P}}\hat{x}_j$ will be calculated as follows + +![](images/260198688b7f50c623cc5d37cb6e3268728877286a4ba52d8e5c7cfbd0f3b894.jpg) + +![](images/29995e04e180ef31d391f4849e88039ea00f178dd36534d0563f50a847c3de6a.jpg) +Figure 4. Illustration of flexibility to customize the group mask: (a). Taking overlapping objects as one group with the guidance of bounding boxes. (b). Taking overlapping objects as one group with the guidance of instance masks. (c) Taking overlapping objects as distinct groups with the guidance of instance masks. +(a) GI window partition under regular window partition. +(b) GI window partition under shifted window partition. +Figure 5. Group-independent window partition of GI SwinBlock under regular and shifted window partition. Self-attention is conducted inside each local region. + +$$ +\hat {\boldsymbol {x}} _ {j} = g _ {s} \left(\hat {\boldsymbol {y}} _ {j} \mid \boldsymbol {m} _ {j}\right). \tag {2} +$$ + +Similar to Equation (1), Equation (2) can also be written as $\hat{\pmb{x}} = g_s(\hat{\pmb{y}} |\pmb {m})$ . Note that the rounding operation is non-differentiable, thus we apply uniform noise for learning the entropy model and use a rounded one as the input of the decoder $g_{s}$ as in [53]. + +# 4. Experiments + +# 4.1. Experiment Setup + +Image Codec. We train our high bitrate models $(>= 0.6\mathrm{bpp})$ on the Flickr 2K dataset, which has the same setting as [74] and train our low bitrate models $(< 0.6\mathrm{bpp})$ on the COCO 2017 training set [45]. Ours-Hyper are trained with 3.2M iterations, while Ours-ChARM models are trained by initializing their weights from higher bitrate hyperprior models for another 1.2M iterations. Each batch contains 8 random $256 \times 256$ crops from the training dataset. The learning rate is set as $5e - 5$ and is decayed by a factor of 10 at $2.8M$ iterations. Training loss $L = R + \beta D$ is the weighted combination of the rate-distortion trade-off, balanced by the Lagrange multiplier $\beta$ . We adopt the mean squared error (MSE) in the RGB color space as the distortion metric $D$ . The bitrate range of our image codec is achieved by applying different $\beta$ that $\beta \in \{512, 1024, 2048, 4096\}$ . During the training stage, we randomly generate the group masks for the training of the group-independent transform, the detailed operations are shown in the supplementary material. + +Evaluation Datasets and Protocol. For the entire image reconstruction quality, we use the widely-used Kodak dataset [37] to evaluate the coding efficiency of our models. We evaluate the selective reconstruction quality on objects + +![](images/9b12c7d4a0e73874705e51a40266a9d7bb47f380cc053c4c9c90bc580acf9213.jpg) +(a) Entire image reconstruction on Kodak. + +![](images/454acd5af833d1e28fe6c84492629d937cfdd848dcee5678633f682360df2fbb.jpg) +(b) All categories objects reconstruction on COCO. + +![](images/05afc7ad8aab7a1a4c1e804b9e235000091a7895845f4c471278342b1924d8c5.jpg) +(c) Human category objects reconstruction on COCO. + +![](images/86f7d85cbcfd47b5265e68850ba59a0ba7be47d47417f88613bd42ec786aa812.jpg) +Figure 6. Rate distortion comparison of compression efficiency on both entire and partial reconstruction scenarios. The PSNR of partial reconstructions is only calculated on the pre-detected bounding boxes. + +![](images/353c8bda89cf67445284c9a35188dad4bc74df5b7b498bec08ef0bae8c6d4a9e.jpg) +Figure 7. Performance comparison on instance segmentation (left) and pose estimation (right) on COCO. + +of all categories and human objects using 40 and 20 images, which are randomly selected from the COCO 2017 validation set. The PSNR of the region of interest serves as the metric for measuring the objective quality. Additionally, the bits per pixel (bpp) are calculated by dividing the total bits of the transmitted bitstream by the number of pixels in the region of interest. To verify the effectiveness of our proposed method in supporting downstream tasks, we conduct experiments on instance segmentation and pose estimation using the COCO 2017 validation set. [45]. This dataset is widely used for dense prediction tasks and comprises 118K training, 5K validation, and 20K test-dev images. We evaluate the performances of instance segmentation and human keypoint detection using Mask R-CNN (X101-FPN backbone) and Keypoint R-CNN (X101-FPN backbone), respectively, implemented in the Detectron2 toolbox [68]. + +# 4.2. Quantitative Results + +For comparisons, we choose recent neural image codes [16,27,52,75], powerful classical image codes [10, 61], and the functional codec that supports semantically structured bitstream [62]. Among them, ELIC [27] is reproduced and performs closely as their report. It is crit- + +ical to emphasize that the encoder does not possess prior knowledge regarding the regions of interest required by the decoder. And we utilize object detection as the pre-analysis by default, which is one of the most fundamental visual analysis tools. The semantically structured bitstreams are generated based on the bounding boxes detected by the detector of Mask R-CNN (X101-FPN backbone), and are selectively decoded and reconstructed into images for evaluation according to the requirements of relative semantics. In all experiments of quantitative results, the block size $B$ of group masks is set to 32 to balance the trade-off between performance and overhead bitrate, and we take the union of overlapping groups as one group. More experiments compared with other methods of image coding for machines are presented in the supplementary. + +Reconstruction Quality. For entire image reconstruction, Figure. 6a demonstrates that our proposed model achieves state-of-the-art rate-distortion performance, surpassing VTM 18.2 in terms of PSNR at all bitrates. Notably, structuring the bitstream according to semantics introduces negligible effects. Therefore, our model has a strong potential to achieve both high coding efficiency and functionality simultaneously. For partial image reconstruction in scenarios of specific interest, as shown in Figure. 6b and Figure. 6c, our model has achieved a significant improvement compared to other CODECs. Specifically, the semantically structured bitstream enabled our model and SSIC [62] to avoid transmitting and decoding bitstreams corresponding to the entire image, which is required of ELIC [27]. Additionally, our model's flexible block-level group mask and group-independent transform spatially decouple latents from the semantic level, resulting in sharper and more re + +![](images/9a7d6896638011c7da58bfb5d5a4d44d52557be07a841e5bd38414a2e789d8df.jpg) + +![](images/d26c13f5ecbf86b15f0e1465c92066dad84819ca749f521a992c3f2938c59cc1.jpg) +0.516/29.36 + +![](images/b82832be0cdaefe0fb89afa2c13c282dce936f6ee9ef022d4683c3edd1066af1.jpg) +0.251/29.40 + +![](images/91153e25e2b2538a795999db486136efc42bc8df64aa3cf632dee8600fcbb228.jpg) +0.195 / 30.14 + +![](images/2321acaaa637c289a5c0a1f8e3f959ffd0cb018b4c55e85bea0cd98459d59d91.jpg) + +![](images/2e1c2e28dc631c2fdcd56c02a5221d6e48d989aa79fe96945ba03b26ca87ddfb.jpg) + +![](images/f357cb43ff7ddff1b85d23af659ca4e98589f68dcb4b8cd5c7ee436d77d2a67d.jpg) + +![](images/60b72b7b5778c6a5edad1e354da5f0f2a811a9de8c487dbf196f0444f14857a6.jpg) + +![](images/6596592032ef0de60f2972318ae404dc93639bf21f1ce1008c80b23f6288bb76.jpg) +Original Image + +![](images/5815dc1ed2654d37d11e6a5ea48fc4b30b7d7b13e331035ac5fda3288bfc736b.jpg) +0.318/28.61 +ELIC + +![](images/c9a827a04c52c7686e159b55bdd6a0952adfae61c751cdbf42badde15c87a27c.jpg) +0.295/28.53 +SSIC + +![](images/c8c1e82b2c0da3fcbbfaaee39d9ebac3dd039bccb0aceaf89619198a4d7d54f8.jpg) +0.316 / 29.25 +Ours + +![](images/5f2953d350535e7801e274afa9f8d1063982131c77f25427a099cd4d84f3be2e.jpg) + +![](images/adece8258b93916b6efd26abe44ab7f10348c55c69fafce5b5d6b4d2f710cfce.jpg) + +![](images/9911aecd9cc119cf8b57d5b021fd2ab6a46f042fb417a95b1bd287df5dc9bbdd.jpg) + +![](images/2ea4415c651ebfca4edbf6c5af7750d68a43b3fac712ef5bb3eccb1f4c7dfe15.jpg) + +![](images/608398d18aaf1880d92b461c09528722f51fa2b0b3bbda7116c4551a529a0d70.jpg) + +![](images/2ade7db95446bd023e7632e8331c570c6f4ac7f7e4808aec24940599d2122fba.jpg) + +![](images/c2fc8807996df6a36b79859eeeca1e3317e3151ff8419059fb1432c567730fc3.jpg) + +![](images/8415f5019a1d8711b9a137210601a8f997b8e81047da306068fcf626e1d7b406.jpg) +Figure 8. Visualization of reconstructed images for human consumption from the COCO 2017 dataset, in the scenario where foreground objects are the regions of interest. The numbers below the images correspond to bits per pixel and PSNR. + +![](images/a13a0777271ce508fd1754ea8677adf56d6b83d632d1b3afe9937524c4a52099.jpg) + +![](images/aa1c98fa10d31e1c2136f8e7d61b4021f37b7bae999e9f18f5f1ff57e58abf21.jpg) + +![](images/861d3d4a4d8256d62b1215bbe6b9b627198ca1d1994661459f2a6fda843512dc.jpg) +1.850bpp + +![](images/330b8d1349b7574c9d6833e3b9f55fc3ebf4f4452a1c3e7aa65971fb0e685025.jpg) +1.592bpp + +![](images/49fa984a24ca38df917f62282ed49a2bea31b9f456fb45145a7541e5c816f5ff.jpg) +1.211bpp + +![](images/acd3064c9b691bceecd9ea5b338e637633cc619e06e0d49549e35befe6077a76.jpg) +Original Image +Figure 9. Visualization of reconstructed images for downstream tasks from the COCO 2017 dataset. The first line corresponds to instance segmentation, and the second one corresponds to pose estimation. + +![](images/1506b4316dbd3ba8017e120cfb41ee1f8d2d8f01e5922c859f08579b8a1222cf.jpg) +Ground Truth + +![](images/f7f548d9eee40c0acab91e4583932cf93071d80c036a465f43df8a5714770586.jpg) +2.229bpp +ELIC + +![](images/d811158090ed31fc98c9a7d750aff7fa179417fb1a258abe2e11091460623ccb.jpg) +1.335bpp +SSIC + +![](images/8f696fbfa4d9872515a66aa0bee837dbf4d4afb5e2e216a55c05aae7b68a03eb.jpg) +1.048bpp +Ours + +alistic boundaries compared to SSIC [62]. Moreover, in the case of overlapping regions, our model can significantly save bitrate by grouping them into an irregular group instead of replacing them with a larger bounding box, which would introduce more distortion. + +Downstream Task Supporting. Figure. 7 demonstrates our excellent performance on instance segmentation and pose estimation. On instance segmentation, our model achieves significantly improved performance at low bitrates (<0.6bpp) compared with other methods, which can be attributed to the faithful reconstruction of RoI boundaries in our approach. It is worth noting that the superior performance of our method on pose estimation is due to the ability of our model to preserve fine details and boundaries, which are critical for accurately localizing human keypoints. In addition, human objects are a minority in the image and are often sparsely distributed, and compared with SSIC, our method can significantly reduce bitrate by avoiding transmitting the latents corresponding to the rectangular area containing all overlapped objects. + +# 4.3. Qualitative Results + +Reconstruction for Human Consumption. When the reconstructed images are mainly intended for human con + +sumption with specific semantic interests, as shown in Figure 8, bitstreams generated by CODEs designed for compressing regular rectangular images are required to be fully transmitted and serve full reconstruction, regardless of the requirements and the content of images. Although SSIC [62] partially addresses this issue through generating SSB through bounding box based division, it lacks flexibility and may introduce additional irrelevant disturbance when merging overlapping objects into larger bounding boxes. Moreover, SSIC generates the bitstream by directly dividing the corresponding positions in the latent variable space after the transform based on ConvNets, which can lead to lost dependencies during selective reconstruction, resulting in blurred and distorted region boundaries. With group mask based partitioning and group-independent transform, our method decouples images to generate the SSB more efficiently while ensuring that no distortion or blurring is introduced after selective reconstruction, leading to both significant bitrate saving and a better visual experience. + +Reconstruction for Intelligent Tasks. When using reconstructed images for downstream intelligent tasks, selectively transmitting and decoding the semantically structured bitstream based on the prior knowledge of relative seman + +![](images/799f42b5c5c2dcb10d91b8b35101445a0edbdc4ea3afb0a026a1b970294ae9bf.jpg) + +![](images/e7fa75907c6ea135bb0420e8e2dd0678a0a05ebfef73509e511ed710500ac633.jpg) +Original Image + +![](images/37fc29c7ed7d0877f883a4ea8c36c41c4fa9ce64956e6d5946ec7108b3b187eb.jpg) + +![](images/d4db6a3afb325cfeeef3f20e5835f794e0dca570673340f8d777d1a6a7c915ac.jpg) +Pre-Analysis + +![](images/a2799c7805e83a4a04e9f5a0784d05410ae31451129c4fd211fa829470a7017b.jpg) + +![](images/43a71bd43fd4e03274e00f62bdb929c436d733b3f0c4338dd5c132687033a308.jpg) +Based on + +![](images/8769c172d99d89cd87ba721a8bb8fcb780119bb35cd6cf6d0b62192c2879d15d.jpg) + +![](images/0f6557947609d733be5376ff9746a43819fdc09b4b7e1962eeb2d422f0efc629.jpg) +Bounding Boxes +Based on +Instance Maps + +![](images/6fa09b9e4356de1d3cfc882bdc6df10b92f7f6b79b3643749b9d5c70e87784fe.jpg) +Figure 10. Selective reconstructions from the semantically structured bitstream. The bitstream can be generated based on different pre-analysis methods. +Original Image +Figure 11. Illustration of semantically-aware encryption based on our proposed method. + +![](images/b261705210627fc10e0391bf0a3ff55d4f1a540e4577571c0703c93c0f99b029.jpg) +Pre-Analysis + +![](images/0c2f30714811b2af2c34764895014333a29fcf7d3308a8cf5052e087eede41ef.jpg) +Encrypt People + +![](images/b773c9f4052d737dee2ed6239d88df1083153744309efc5282895de23c2fcdcc.jpg) +Encrypt People +and Vehicles + +tics can significantly save bitrate. As illustrated in Figure 9, conventional codecs designed for compressing rectangular images require the entire bitstream to be transmitted and decoded. Downstream models then perform intelligent analytics on the fully reconstructed image, resulting in a waste of bitrate. Similar to image reconstruction for human consumption, SSIC can perform selective transmission to save bitrate. However, its bounding box based partitioning may not be optimal for bitrate savings, and the resulting blurred and distorted region boundaries and irrelevant content can further impede downstream intelligent analytics. (misidentifying the car as a person shown in Figure 9). Our method can more efficiently generate the SSB and support various downstream tasks with specific requirements with both accuracy and coding efficiency. + +Customizability and Flexibility. It's flexible to customize the semantically structured bitstream based on different semantical partition criteria. For instance, as shown in Figure 10, the generation and usage of semantically structured bitstreams can be based on either object detection or instance segmentation, depending on the specific requirements. It should be noted that the methods of pre-analysis are not limited to object detection and instance segmentation as shown in Figure 10, but can include other techniques such as saliency detection [22, 30, 72], semantic segmentation [5, 48, 69, 71, 73], panoptic segmentation [15, 36], etc., and even human annotation. + +Semantically-Aware Encryption. As presented in Figure 11, our proposed method can be applied to bitstream encryption, enabling selective and even layered encryption with semantic priors based on the user's security level. In some cases, the selectively encrypted SSB can allow for secure transmission and storage of sensitive information while minimizing the impact on visual quality and downstream analytics. Please refer to the supplementary material for + +![](images/456c61c5f8400fe58ecc3baef621b45b8a82c4c31dc7286b67f525c6784f8833.jpg) +(a) RoI reconstruction. +Figure 12. Ablation study. + +![](images/c3bcf2a27219e7eb2317217dfe73097f5e1251e12e9589ea6cdc245eef733f9d.jpg) +(b) Intelligent task supporting. + +details of implementation. + +# 5. Ablation Study + +In order to show the effectiveness of our proposed group mask and group-independent transform (GIT) for both RoI-aware selective reconstruction and downstream intelligent analytics, we conducted the ablation study at different bitrates. Specifically, we use people as the region of interest for RoI reconstruction, and the results for taking all foreground objects as RoI are presented in the supplementary material. Additionally, we conducted an ablation study on pose estimation of COCO 2017 for intelligent task support, and the results on instance segmentation are also included in the supplementary material. + +Group Mask. Figure 12 shows that group mask based partitioning saves significant bitrate compared with naive bounding box based partitioning. Meanwhile, both the visual quality of RoI reconstruction and downstream task performances would not be impeded. + +Group-Independent transform. As shown in Figure 12a, group-independent transform is crucial for the visual quality of RoI reconstruction. And the downstream task performance would be impeded by the blurred and distorted region boundaries without group-independent transform, which is shown in Figure 12b. + +# 6. Conclusion + +In this work, we propose to generate semantically structured bitstream with strong functionality based on group masks, which are highly flexible, customizable, while maintaining lightweight overhead. Moreover, we first propose the concept of group-independent transform and instantiate it by designing the Group-Independent Swin-Block (GI-Swin Block) to ensure independence among distinct groups, resulting in pleasing reconstructions of RoI regions. Particularly, our proposed method outperforms VTM-18.2 and achieves comparable coding efficiency with the SOTA neural image CODECs, while offering strong functionality. The experimental results have demonstrated the effectiveness of our proposed method across several applications, including human consumption of regions of interest, downstream task support, and semantically-aware encryption. + +Acknowledgement. This work was supported in part by NSFC under Grant U1908209, 62021001. This work was also supported in part by ZJNSFC under Grant LQ23F010008. + +# References + +[1] Eirikur Agustsson, Fabian Mentzer, Michael Tschannen, Lukas Cavigelli, Radu Timofte, Luca Benini, and Luc V Gool. Soft-to-hard vector quantization for end-to-end learning compressible representations. *NeurIPS*, 30, 2017. 2 +[2] Eirikur Agustsson, Michael Tschannen, Fabian Mentzer, Radu Timofte, and Luc Van Gool. Generative adversarial networks for extreme learned image compression. In ICCV, pages 221-231, 2019. 2 +[3] Mohammad Akbari, Jie Liang, and Jingning Han. Dsslic: Deep semantic segmentation-based layered image compression. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2042-2046. IEEE, 2019. 3 +[4] Artem Babenko, Anton Slesarev, Alexandr Chigorin, and Victor Lempitsky. Neural codes for image retrieval. In ECCV, pages 584-599. Springer, 2014. 1, 3 +[5] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. TPAMI, 39(12):2481-2495, 2017. 1, 3, 8 +[6] Johannes Ballé, Philip A Chou, David Minnen, Saurabh Singh, Nick Johnston, Eirikur Agustsson, Sung Jin Hwang, and George Toderici. Nonlinear transform coding. IEEE Journal of Selected Topics in Signal Processing, 15(2):339-353, 2020. 2 +[7] Johannes Ballé, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. In *ICLR*, 2017. 1, 2 +[8] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In ICLR, 2018. 1, 2 +[9] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In ICCV, pages 9157-9166, 2019. 1, 3 +[10] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. TCSVT, 2021. 1, 2, 6 +[11] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 40(4):834-848, 2017. 1, 3 +[12] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, pages 801-818, 2018. 1, 3 +[13] Zhuo Chen, Kui Fan, Shiqi Wang, Lingyu Duan, Weisi Lin, and Alex Chichung Kot. Toward intelligent sensing: Intermediate deep feature compression. TIP, 29:2230-2243, 2019. 1, 3 + +[14] Zhuo Chen, Kui Fan, Shiqi Wang, Ling-Yu Duan, Weisi Lin, and Alex Kot. Lossy intermediate deep learning feature compression and evaluation. In ACM MM, pages 2414-2422, 2019. 1, 3 +[15] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, pages 12475-12485, 2020. 8 +[16] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In CVPR, pages 7939-7948, 2020. 1, 2, 6 +[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 1, 3 +[18] Ling-Yu Duan, Vijay Chandrasekhar, Jie Chen, Jie Lin, Zhe Wang, Tiejun Huang, Bernd Girod, and Wen Gao. Overview of thempeg-cdvs standard. TIP, 25(1):179-194, 2015. 1, 3 +[19] Ling-Yu Duan, Yihang Lou, Yan Bai, Tiejun Huang, Wen Gao, Vijay Chandrasekhar, Jie Lin, Shiqi Wang, and Alex Chichung Kot. Compact descriptors for video analysis: The emerging mpg standard. IEEE MultiMedia, 26(2):44-54, 2018. 1, 3 +[20] Touradj Ebrahimi and Caspar Horne. Mpeg-4 natural video coding—an overview. Signal Processing: Image Communication, 15(4-5):365-385, 2000. 3 +[21] Ruoyu Feng, Xin Jin, Zongyu Guo, Runsen Feng, Yixin Gao, Tianyu He, Zhizheng Zhang, Simeng Sun, and Zhibo Chen. Image coding for machines with omnipotent feature learning. In ECCV, pages 510-528. Springer, 2022. 2, 3 +[22] Stas Goferman, Lihi Zelnik-Manor, and Ayellet Tal. Context-aware saliency detection. TPAMI, 34(10):1915-1926, 2011. 8 +[23] Vivek K Goyal. Theoretical foundations of transform coding. IEEE Signal Processing Magazine, 18(5):9-21, 2001. 2 +[24] Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Causal contextual prediction for learned image compression. TCSVT, 32(4):2329-2341, 2021. 2 +[25] Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Soft then hard: Rethinking the quantization in neural image compression. In ICML, pages 3920-3929. PMLR, 2021. 2 +[26] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. NeurIPS, 34:15908-15919, 2021. 1, 3 +[27] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In CVPR, pages 5718-5727, 2022. 1, 2, 6 +[28] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 1, 3 +[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 3 + +[30] Xiaodi Hou and Liqing Zhang. Saliency detection: A spectral residual approach. In CVPR, pages 1-8. IEEE, 2007. 8 +[31] Yueyu Hu, Shuai Yang, Wenhan Yang, Ling-Yu Duan, and Jiaying Liu. Towards coding for human and machine vision: A scalable image coding approach. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 3 +[32] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. ECCV, 2022. 1, 3 +[33] Xin Jin, Ruoyu Feng, Simeng Sun, Runsen Feng, Tianyu He, and Zhibo Chen. Semantically video coding: Instill static-dynamic clues into structured bitstream for ai tasks. arXiv preprint arXiv:2201.10162, 2022. 3 +[34] Nick Johnston, Damien Vincent, David Minnen, Michele Covell, Saurabh Singh, Troy Chinen, Sung Jin Hwang, Joel Shor, and George Toderici. Improved lossy image compression with priming and spatially adaptive bit rates for recurrent networks. In CVPR, pages 4385-4393, 2018. 1 +[35] Aggelos K Katsaggelos, Lisimachos P Kondi, Fabian W Meier, Jorn Ostermann, and Guido M Schuster. Mpeg-4 and rate-distortion-based shape-coding techniques. Proceedings of the IEEE, 86(6):1126-1154, 1998. 3 +[36] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, pages 9404-9413, 2019. 8 +[37] Eastman Kodak. Kodak lossless true color image suite (photocd pcd0992). http://r0k.us/graphics/kodak, 1993.5 +[38] Nam Le, Honglei Zhang, Francesco Cricri, Ramin GhaznaviYouvalari, and Esa Rahtu. Image coding for machines: An end-to-end learned approach. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1590-1594. IEEE, 2021. 1, 3 +[39] Nam Le, Honglei Zhang, Francesco Cricri, Ramin Ghaznavi Youvalari, Hamed Rezazadegan Tavakoli, and Esa Rahtu. Learned image coding for machines: A content-adaptive approach. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 1 +[40] Mu Li, Wangmeng Zuo, Shuhang Gu, Jane You, and David Zhang. Learning content-weighted deep image compression. TPAMI, 2020. 1 +[41] Mu Li, Wangmeng Zuo, Shuhang Gu, Debin Zhao, and David Zhang. Learning convolutional networks for content-weighted image compression. In CVPR, pages 3214-3223, 2018. 1 +[42] Xin Li, Jun Shi, and Zhibo Chen. Task-driven semantic coding via reinforcement learning. TIP, 2021. 3 +[43] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. ECCV, 2022. 1, 3 +[44] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 1, 3 +[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence + +Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 5, 6 +[46] Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In CVPR, pages 8759-8768, 2018. 1, 3 +[47] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 1, 2, 3, 4 +[48] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 1, 3, 8 +[49] Siwei Ma, Xiang Zhang, Shiqi Wang, Xinfeng Zhang, Chuanmin Jia, and Shanshe Wang. Joint feature and texture coding: Toward smart video representation via front-end intelligence. TCSVT, 29(10):3095-3105, 2018. 1, 3 +[50] Fabian Mentzer, Eirikur Agustsson, Michael Tschannen, Radu Timofte, and Luc Van Gool. Conditional probability models for deep image compression. In CVPR, pages 4394-4402, 2018. 1 +[51] Fabian Mentzer, George Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. arXiv preprint arXiv:2006.09965, 2020. 1, 2 +[52] David Minnen, Johannes Balle, and George Toderici. Joint autoregressive and hierarchical priors for learned image compression. In NeurIPS, 2018. 1, 2, 5, 6 +[53] David Minnen and Saurabh Singh. Channel-wise autoregressive entropy models for learned image compression. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3339-3343. IEEE, 2020. 2, 5 +[54] Majid Rabbani and Rajan Joshi. An overview of the JPEG 2000 still image compression standard. Signal processing: Image communication, 17(1):3-48, 2002. 1, 2 +[55] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 1, 3 +[56] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In CVPR, pages 7263-7271, 2017. 1, 3 +[57] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28:91-99, 2015. 1, 3 +[58] Oren Rippel and Lubomir Bourdev. Real-time adaptive image compression. In ICML, pages 2922-2930. PMLR, 2017. 2 +[59] Thomas Sikora. Thempeg-4 video standard verification model. TCSVT, 7(1):19-31, 1997. 3 +[60] Saurabh Singh, Sami Abu-El-Haija, Nick Johnston, Johannes Balle, Abhinav Shrivastava, and George Toderici. End-to-end learning of compressible features. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3349–3353. IEEE, 2020. 1, 3 +[61] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. TCSVT, 22(12):1649-1668, 2012. 1, 2, 6 +[62] Simeng Sun, Tianyu He, and Zhibo Chen. Semantic structured image coding framework for multiple intelligent applications. TCSVT, 2020. 2, 3, 6, 7 + +[63] Lucas Theis, Wenzhe Shi, Andrew Cunningham, and Ferenc Huszár. Lossy image compression with compressive autoencoders. arXiv preprint arXiv:1703.00395, 2017. 2 +[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 3, 4 +[65] Gregory K Wallace. The JPEG still picture compression standard. IEEE transactions on consumer electronics, 38(1):xviii-xxxiv, 1992. 1, 2 +[66] Shurun Wang, Shiqi Wang, Wenhan Yang, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. Towards analysis-friendly face representation with scalable feature and texture compression. TMM, 2021. 3 +[67] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. TCSVT, 13(7):560-576, 2003. 1 +[68] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6 +[69] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. NeurIPS, 34:12077-12090, 2021. 1, 3, 8 +[70] Yibo Yang, Robert Bamler, and Stephan Mandt. Improving inference for neural image compression. NeurIPS, 33:573-584, 2020. 2 +[71] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, pages 2881-2890, 2017. 8 +[72] Rui Zhao, Wanli Ouyang, Hongsheng Li, and Xiaogang Wang. Saliency detection by multi-context deep learning. In CVPR, pages 1265-1274, 2015. 8 +[73] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. 1, 3, 8 +[74] Lei Zhou, Zhenhong Sun, Xiangji Wu, and Junmin Wu. End-to-end optimized image compression with attention mechanism. In CVPR workshops, page 0, 2019. 5 +[75] Yinhao Zhu, Yang Yang, and Taco Cohen. Transformer-based transform coding. In ICLR, 2022. 1, 2, 4, 6 \ No newline at end of file diff --git a/2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/images.zip b/2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8872413a6f5d5350c65c4c6dd2a528829239e71e --- /dev/null +++ b/2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d2b0b40ab3c1a7892d32c706c19675285b8b8532f66d0e623351a278c32fda9 +size 661304 diff --git a/2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/layout.json b/2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..83ddabca121516ee1d203b67a0e79ab0309a204a --- /dev/null +++ b/2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/layout.json @@ -0,0 +1,11297 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 103, + 449, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 103, + 449, + 140 + ], + "spans": [ + { + "bbox": [ + 146, + 103, + 449, + 140 + ], + "type": "text", + "content": "Semantically Structured Image Compression via Irregular Group-Based Decoupling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 160, + 474, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 160, + 474, + 176 + ], + "spans": [ + { + "bbox": [ + 119, + 160, + 474, + 176 + ], + "type": "text", + "content": "Ruoyu Feng" + }, + { + "bbox": [ + 119, + 160, + 474, + 176 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 119, + 160, + 474, + 176 + ], + "type": "text", + "content": " Yixin Gao" + }, + { + "bbox": [ + 119, + 160, + 474, + 176 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 119, + 160, + 474, + 176 + ], + "type": "text", + "content": " Xin Jin" + }, + { + "bbox": [ + 119, + 160, + 474, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 160, + 474, + 176 + ], + "type": "text", + "content": " Runsen Feng" + }, + { + "bbox": [ + 119, + 160, + 474, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 160, + 474, + 176 + ], + "type": "text", + "content": " Zhibo Chen" + }, + { + "bbox": [ + 119, + 160, + 474, + 176 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 177, + 465, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 177, + 465, + 190 + ], + "spans": [ + { + "bbox": [ + 129, + 177, + 465, + 190 + ], + "type": "text", + "content": "1University of Science and Technology of China 2Eastern Institute of Technology, Ningbo" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "spans": [ + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 243, + 290, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 243, + 290, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 243, + 290, + 542 + ], + "type": "text", + "content": "Image compression techniques typically focus on compressing rectangular images for human consumption, however, resulting in transmitting redundant content for downstream applications. To overcome this limitation, some previous works propose to semantically structure the bitstream, which can meet specific application requirements by selective transmission and reconstruction. Nevertheless, they divide the input image into multiple rectangular regions according to semantics and ignore avoiding information interaction among them, causing waste of bitrate and distorted reconstruction of region boundaries. In this paper, we propose to decouple an image into multiple groups with irregular shapes based on a customized group mask and compress them independently. Our group mask describes the image at a finer granularity, enabling significant bitrate saving by reducing the transmission of redundant content. Moreover, to ensure the fidelity of selective reconstruction, this paper proposes the concept of group-independent transform that maintain the independence among distinct groups. And we instantiate it by the proposed Group-Independent Swin-Block (GI Swin-Block). Experimental results demonstrate that our framework structures the bitstream with negligible cost, and exhibits superior performance on both visual quality and intelligent task supporting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 587, + 288, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 288, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 288, + 684 + ], + "type": "text", + "content": "Image signal serves as a critical information carrier for various applications in modern society. Image compression techniques aim at converting images into compact representations (i.e. bitstreams) to save transmission and storage resources. Lossy image compression is one of the most practical techniques, as it allows for the restoration of important content while discarding a small amount of inessential information. In the past decades, traditional image compres" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 214, + 545, + 373 + ], + "blocks": [ + { + "bbox": [ + 310, + 214, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 310, + 214, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 310, + 214, + 545, + 373 + ], + "type": "image", + "image_path": "82944d9ba34b0a08d6b1cec5006054c78948fdbab5baaf9006f2c024e98a9be7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 380, + 547, + 465 + ], + "lines": [ + { + "bbox": [ + 304, + 380, + 547, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 547, + 465 + ], + "type": "text", + "content": "Figure 1. The input image is decoupled into groups according to distinct semantics. Then the semantically structured bitstream (SSB) is generated by compressing the image based on the partitioned groups. The SSB facilitates downstream applications by selective bitstream transmission and partial reconstruction, depending on the specific task requirements." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 494, + 547, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 547, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 547, + 638 + ], + "type": "text", + "content": "sion standards [10,54,61,65,67] have been extensively studied and utilized. With the fast development of deep learning, neural image codes [7,8,16,27,34,40,41,50-52,75] rapidly evolved and achieved promising results. Meanwhile, more and more media contents tend to be handled by machine vision algorithms, such as recognition [17,26,29,32,47], detection [43,44,55-57], and segmentation [5,9,11,12,28,46,48,69,73]. However, most compression methods are mainly developed for compressing regular rectangular images for human consumption, without considering the efficiency and functionality for downstream tasks or human-machine interaction scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": "Recently, the field of image coding for machines (ICM) has emerged to develop a joint efficient and analytical framework for supporting intelligent analytics. End-to-end optimization of the trade-off between specific task loss and compression rate is a promising way [4, 13, 14, 18, 19, 38, 39, 49, 60], but it lacks generalization for massively diverse" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 176, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 176, + 703 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 176, + 703 + ], + "type": "text", + "content": "* First two authors contributed equally." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 704, + 128, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 704, + 128, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 704, + 128, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17237" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "text", + "content": "applications. To overcome this limitation, Feng et al. [21] propose to compress general and compact features learned by self-supervised learning under entropy constraints for supporting downstream tasks. Nevertheless, it requires retraining the task model with the proposed features as inputs, which makes the overall performance heavily depend on the effectiveness of feature extraction. The methods mentioned above are designed specifically for scenarios of compression for machine vision without considering situations in which human involvement is required. Semantically structured image compression (SSIC) [62] proposes to generate a semantically structured bitstream (SSB) by separately compressing rectangular regions of detected objects using a pre-prepared object detection toolbox. Although SSB is efficient in supporting intelligent tasks and human-machine interaction through partial transmission and reconstruction of the bitstream, its division approach based on rectangular regions may encounter problems with overlapped objects. SSIC addresses this issue by replacing the overlapping objects with a larger rectangular region, which can result in a waste of bitrate. In addition, SSIC generates the bitstream of each object by compressing the corresponding latent variables in the latent domain directly, without considering the interaction and dependence of features during the transform process. This can lead to blurry and distorted group boundaries in the partial reconstruction scenario, which in turn affects the reconstruction quality." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 403, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 289, + 714 + ], + "type": "text", + "content": "Going beyond the rectangular-based division, this paper proposes to decouple the image into multiple groups with irregular shapes based on a customized group mask. Then the SSB is generated by independently compressing these groups and can support various requirements via selective transmission and reconstruction. Notably, the generation of the group mask offers high flexibility in terms of shapes of groups, means of pre-analysis, and partition criteria, enabling customization to suit diverse application scenarios and requirements. Moreover, to avoid potential quality degradation in the partial reconstruction scenario, we propose the concept of the group-independent transform, which ensures the independence among groups in the latent representations and therefore the quality of the selective reconstruction will not be affected by the absence of other groups. More specifically, we instantiate it by carefully designing the Group-Independent Swin-Block (GI Swin-Block), which is an extension of Swin Transformer [47] tailored to our situation and requirements. GI SwinBlocks make use of the hierarchical modeling capability of the Swin-Transformer [47, 75], achieving high coding efficiency under the premise of group independence. By combining the group mask based decoupling and the Group-Independent Swin-Block, our proposed method can efficiently support various downstream applications including human-machine interaction and machine vision tasks with" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 306, + 72, + 427, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 427, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 427, + 84 + ], + "type": "text", + "content": "only one bitstream generated." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 84, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 84, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 306, + 84, + 545, + 107 + ], + "type": "text", + "content": "The main contributions of our approach are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 316, + 114, + 545, + 318 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 316, + 114, + 545, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 114, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 114, + 545, + 174 + ], + "type": "text", + "content": "- We propose to decouple an image into multiple groups with irregular shapes for structuring the bitstream. Our group mask can describe spatial division at a finer granularity manner than a typical rectangle, saving bitrate by reducing redundant content transmission." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 316, + 180, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 180, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 180, + 545, + 251 + ], + "type": "text", + "content": "- We propose the group-independent transform and instantiate it by carefully designing the Group-Independent Swin-Block (GI Swin-Block), which maintains powerful transformation capability and ensures the independence among groups in the latent representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 258, + 545, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 258, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 545, + 318 + ], + "type": "text", + "content": "- Experimental results demonstrate that our proposed model achieves state-of-the-art compression ability and superior downstream tasks performance, which is a codec with both high compression efficiency and functionality." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 324, + 398, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 324, + 398, + 337 + ], + "spans": [ + { + "bbox": [ + 306, + 324, + 398, + 337 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 344, + 422, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 344, + 422, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 344, + 422, + 357 + ], + "type": "text", + "content": "2.1. Image Compression" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 365, + 545, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 532 + ], + "type": "text", + "content": "Traditional Image Compression. Traditional image compression standards, such as JPEG [65], JPEG2000 [54], HEVC [61], and VVC Intra [10], have been extensively used in practice after several decades of development. These standards rely on transform coding [23], which decomposes the lossy image compression task into three parts: transform, quantization, and entropy coding. Each module of these standards is manually designed with multiple modes, and rate-distortion optimization is performed to determine the optimal mode. However, the completely hand-crafted structure of traditional CODECs limits their flexibility and scalability to support various objectives, such as MS-SSIM and classification accuracy, as they cannot be optimized in an end-to-end manner." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 534, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 715 + ], + "type": "text", + "content": "Learned Image Compression. In recent years, learned image compression methods based on nonlinear transform coding [6] have achieved rapid progress. Early works in this area concentrated on enabling end-to-end training by developing differential quantization and rate estimation techniques [1, 7, 63]. Subsequently, a considerable amount of work design powerful neural network modules to enhance the transform [16, 75], quantization [25, 70], and entropy model [8, 27, 52, 53]. As a result, some neural image CODECs [16, 24, 27, 52, 53] can achieve comparable or even superior performance to traditional coding standards like HEVC [61] and VVC [10]. Moreover, some works [2, 51, 58] introduce perceptual loss to effectively improve reconstruction quality. However, existing image compression methods focus on compressing the entire image" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17238" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": "without considering selective transmission and reconstruction of the bitstream for arbitrary reconstruction, leading to significant bandwidth waste when addressing downstream applications with different requirements." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 129, + 200, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 129, + 200, + 141 + ], + "spans": [ + { + "bbox": [ + 47, + 129, + 200, + 141 + ], + "type": "text", + "content": "2.2. Image Coding for Machines" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 148, + 288, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 148, + 288, + 447 + ], + "spans": [ + { + "bbox": [ + 47, + 148, + 288, + 447 + ], + "type": "text", + "content": "Image coding for machines aims at compress source images to serve downstream tasks, such as recognition [17, 26, 29, 32, 47], detection [43, 44, 55-57], and segmentation [5, 9, 11, 12, 28, 46, 48, 69, 73]. Joint optimization of task loss and bitrate [3, 31, 38, 42, 66] in an end-to-end manner is a natural and promising way. Another direction is compress the features corresponding to downstream tasks [4, 13, 14, 18, 19, 49, 60]. However, these methods tend to be biased towards specific tasks and may lack generalizability to a wide range of intelligent applications. To address the limitation of generalization, a novel representation learning based method is proposed [21] to learn general and compact features by combining contrastive learning with entropy constraints. The learned features are used to replace the original images as the new source for compression and transmission, resulting in a significant improvement in coding efficiency across various intelligent tasks. Nevertheless, this approach necessitates re-training the task model using the proposed features as input, and its overall performance is heavily reliant on the effectiveness of feature extraction. In this paper, we propose a unified compression framework for machine vision support without task-oriented joint training or additional feature learning. Moreover, it can also conduct full or selective reconstruction for human perception according to requirements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 455, + 241, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 455, + 241, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 455, + 241, + 468 + ], + "type": "text", + "content": "2.3. Object-oriented Image Compression" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 475, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 287, + 713 + ], + "type": "text", + "content": "Object-oriented image compression is first proposed in MPEG-4 Visual [20, 35, 59] by compressing visual object planes (VOP) with arbitrary shapes, targeting content-based interactivity and scalability for the image compression technique. However, due to the difficulty in annotating the complex and intricate shapes of the target and the lack of actual corresponding downstream applications, MPEG-4 Visual has not been widely used. Recently, with the fast development of deep learning [29, 64], the practical scenarios of image compression for downstream applications have increased rapidly. Sun et al. [33, 62] propose the semantically structured image compression (SSIC) of decoupling the image into objects and compressing the corresponding latent variables independently of the neural image codec to generate the semantically structured bitstream. Nevertheless, the rectangular-based division adopted by SSIC lacks generalization to irregular objects and flexibility for customization. Additionally, selecting the representation of objects directly in the latent space ignores the interaction among elements during the encoding transform process, leading to" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "significant edge distortion of objects in partial reconstruction. This paper proposes to conduct semantically structured image compression by decoupling the image into irregular groups based on the group mask, which is flexible and customizable. Furthermore, we propose the group-independent transform conducted during both the encoding and decoding processes, which enables efficient partial or complete reconstruction of images under various requirements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 191, + 362, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 191, + 362, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 191, + 362, + 204 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 212, + 373, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 212, + 373, + 224 + ], + "spans": [ + { + "bbox": [ + 306, + 212, + 373, + 224 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 231, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 469 + ], + "type": "text", + "content": "The proposed method is an efficient and flexible version of semantically structured image compression [62]. It can satisfy multiple application requirements via transmitting and reconstructing partial spatial regions guided by a block-wise group mask. Our model takes the group mask as guidance to ensure that interactions of transform only occur within the same group, thereby achieving group independence during the redundancy removal process. The group mask, which is generated based on the pre-analysis, such as object detection, semantic segmentation, and saliency detection, provides high flexibility and customization for structuring the bitstream and is considered as side information. Then the entropy coding process is conducted on the variables of each group distinctly, resulting in a bitstream structured by semantics. The bitstream can be partially or fully transmitted according to the requirements of the decoder side, and then the receiver conducts entropy decoding on the bitstream and reorganizes the spatial-wise arrangement of latent variables according to the group mask and the group indexes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 470, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 545, + 554 + ], + "type": "text", + "content": "The overall network structure is illustrated in Fig. 2, which incorporates the group mask and GI Swin-Block for guiding and instantiating the group-independent transform, respectively. Section 3.2 and Section 3.3 introduce the group mask and group independent, respectively. The detailed implementation of our network architecture is presented in Section 3.4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 563, + 389, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 389, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 389, + 575 + ], + "type": "text", + "content": "3.2. Group Mask" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "content": "Before compression, the group mask is generated according to the results of pre-analysis techniques such as object detection, instance segmentation, and saliency detection. Figure 3 provides an example of this process. More specifically, for the input image " + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "content": " with a height of " + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "content": " and a width of " + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "content": ", the spatial resolution of the group mask is the same as the input image, and it consists of " + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\frac{H}{B} \\times \\frac{W}{B}" + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "content": " blocks, where " + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "content": " is the side length of each block. And the value of it indicates the index of the corresponding group." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Our proposed group mask can clearly divide the elements in the latent variable space after the one-pass down" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17239" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 274 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 274 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 274 + ], + "type": "image", + "image_path": "d76eb38af4aa73a7eb406dfec3fa55e911ae3f0cb944f421209f8772f5801d39.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 281, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 281, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 547, + 319 + ], + "type": "text", + "content": "Figure 2. The network architecture of our proposed model with the channel-wise auto-regressive model (ChARM). ConvT denotes transposed convolution. AE and AD are respectively arithmetic encoding and arithmetic decoding. In Ours-Hyper model, we remove the ChARM component and instead output " + }, + { + "bbox": [ + 46, + 281, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 281, + 547, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 281, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 281, + 547, + 319 + ], + "type": "text", + "content": " directly from the hyperdecoder " + }, + { + "bbox": [ + 46, + 281, + 547, + 319 + ], + "type": "inline_equation", + "content": "h_s" + }, + { + "bbox": [ + 46, + 281, + 547, + 319 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 337, + 286, + 412 + ], + "blocks": [ + { + "bbox": [ + 56, + 337, + 286, + 412 + ], + "lines": [ + { + "bbox": [ + 56, + 337, + 286, + 412 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 286, + 412 + ], + "type": "image", + "image_path": "d018b1156d0090bc1bf3f1b2206388b15125c66e51427de92ecb2ae225e90d97.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 421, + 288, + 459 + ], + "lines": [ + { + "bbox": [ + 46, + 421, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 288, + 459 + ], + "type": "text", + "content": "Figure 3. An example of group mask generation. The input image is pre-analyzed by object detection, then the group mask is generated based on the results of the pre-analysis." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 480, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 480, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 287, + 564 + ], + "type": "text", + "content": "sampling transform. Besides, downsampling the group mask by a factor of " + }, + { + "bbox": [ + 46, + 480, + 287, + 564 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 480, + 287, + 564 + ], + "type": "text", + "content": " before conducting compression can greatly reduce the overhead bitrate. The block size and the overhead bitrate have an inverse relationship. Furthermore, as shown in Figure 4, the generation of the group mask can be flexibly customized, in terms of the way of pre-analysis and the criteria of block allocation, etc." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 574, + 214, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 214, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 214, + 588 + ], + "type": "text", + "content": "3.3. Group-Independent transform" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "Lossy image compression based on transform coding can be divided into three modularized components: transform, quantization, and entropy coding. While quantization and entropy coding do not affect the generation and usage of the semantically structured bitstream (SSB), traditional transform applied to the entire image to remove spatial redundancy inevitably creates inter-group dependencies during compression. In situations of selective transmission and reconstruction, the incomplete representation that results from the reorganization of a subset of all groups can" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 339, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 545, + 399 + ], + "type": "text", + "content": "lead to inaccurate reconstruction due to the lack of intergroup dependencies. Therefore, we propose the concept of the group-independent transform for semantically structured bitstream generation. The basic idea is to constrain the transform to be conducted only within each group." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 399, + 546, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 546, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 546, + 663 + ], + "type": "text", + "content": "A straight and natural way is by using transformer [64] with the customized attention map corresponding to the group mask. However, image compression differs significantly from high-level understanding tasks, which the transformer excels at. The presence of numerous complex long-range dependencies can impede convergence and adversely affect the final performance. In order to achieve high coding efficiency while maintaining the group-independent property, and inspired by the strong hierarchical representation modeling capability of Swin-Transformer [47, 75], we extend the Swin-Block of Swin-Transformer to our proposed group-independent Swin-Block (GI Swin-Block). The GI Swin-Block serves as the core component of the transform. To be more specific, as shown in Figure 5a, by merging the window partition and the group partition into the group-independent window partition, allowing the self-attention to be computed within the partitioned local regions, enabling us to achieve high coding efficiency while maintaining the group-independent property. Additionally, the cross-window connections are introduced by merging the shifted window partition and group partition similarly, as depicted in Figure 5b." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 670, + 491, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 491, + 685 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 491, + 685 + ], + "type": "text", + "content": "3.4. Network architecture and pipeline" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "content": "Figure 2 illustrate the proposed architecture. The main transform is performed in a staggered manner through" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17240" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 77, + 93, + 128 + ], + "blocks": [ + { + "bbox": [ + 51, + 77, + 93, + 128 + ], + "lines": [ + { + "bbox": [ + 51, + 77, + 93, + 128 + ], + "spans": [ + { + "bbox": [ + 51, + 77, + 93, + 128 + ], + "type": "image", + "image_path": "64e2a7e5d133e8584b985bb8fc4289ce2b243157912d5b5644d5c8bb2a027ef0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 131, + 89, + 139 + ], + "lines": [ + { + "bbox": [ + 54, + 131, + 89, + 139 + ], + "spans": [ + { + "bbox": [ + 54, + 131, + 89, + 139 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 93, + 78, + 134, + 128 + ], + "blocks": [ + { + "bbox": [ + 93, + 78, + 134, + 128 + ], + "lines": [ + { + "bbox": [ + 93, + 78, + 134, + 128 + ], + "spans": [ + { + "bbox": [ + 93, + 78, + 134, + 128 + ], + "type": "image", + "image_path": "847ce4cb612360f0315619348f88a9c1ae9f73c5efdf51227898cba8b0507cf1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 95, + 131, + 131, + 139 + ], + "lines": [ + { + "bbox": [ + 95, + 131, + 131, + 139 + ], + "spans": [ + { + "bbox": [ + 95, + 131, + 131, + 139 + ], + "type": "text", + "content": "Pre-Analysis" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 77, + 189, + 128 + ], + "blocks": [ + { + "bbox": [ + 134, + 77, + 189, + 128 + ], + "lines": [ + { + "bbox": [ + 134, + 77, + 189, + 128 + ], + "spans": [ + { + "bbox": [ + 134, + 77, + 189, + 128 + ], + "type": "image", + "image_path": "7da7c82cd4385373807b1a9982a8b99f7f115051fb63e0034447d20768c16861.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 129, + 172, + 135 + ], + "lines": [ + { + "bbox": [ + 165, + 129, + 172, + 135 + ], + "spans": [ + { + "bbox": [ + 165, + 129, + 172, + 135 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 183, + 136, + 246, + 144 + ], + "lines": [ + { + "bbox": [ + 183, + 136, + 246, + 144 + ], + "spans": [ + { + "bbox": [ + 183, + 136, + 246, + 144 + ], + "type": "text", + "content": "Possible Group Masks" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 194, + 77, + 235, + 128 + ], + "blocks": [ + { + "bbox": [ + 194, + 77, + 235, + 128 + ], + "lines": [ + { + "bbox": [ + 194, + 77, + 235, + 128 + ], + "spans": [ + { + "bbox": [ + 194, + 77, + 235, + 128 + ], + "type": "image", + "image_path": "57afce836be7bf66ddd76947463897e79365a5ec9ab2725cade8d1ad37afb6c0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 129, + 219, + 135 + ], + "lines": [ + { + "bbox": [ + 211, + 129, + 219, + 135 + ], + "spans": [ + { + "bbox": [ + 211, + 129, + 219, + 135 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 239, + 77, + 281, + 128 + ], + "blocks": [ + { + "bbox": [ + 239, + 77, + 281, + 128 + ], + "lines": [ + { + "bbox": [ + 239, + 77, + 281, + 128 + ], + "spans": [ + { + "bbox": [ + 239, + 77, + 281, + 128 + ], + "type": "image", + "image_path": "7a899820c4768ae05823d7a91afd1b4c4e267e1306e2d69a2fa71390d363631d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 256, + 129, + 263, + 135 + ], + "lines": [ + { + "bbox": [ + 256, + 129, + 263, + 135 + ], + "spans": [ + { + "bbox": [ + 256, + 129, + 263, + 135 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 245, + 287, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 245, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 287, + 352 + ], + "type": "text", + "content": "the incorporation of upsample/downsample and group-independent Swin-Block (GI Swin-Block). Upsample contains a pixel shuffle operation and a convolution layer with " + }, + { + "bbox": [ + 46, + 245, + 287, + 352 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 245, + 287, + 352 + ], + "type": "text", + "content": " kernel size, whereas downsample contains a pixel unshuffle operation and a convolution layer with " + }, + { + "bbox": [ + 46, + 245, + 287, + 352 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 245, + 287, + 352 + ], + "type": "text", + "content": " kernel size. For entropy coding, we adopt the Mean & Scale (M&S) Hyperprior model [52] with the strong channel-wise auto-regressive model (ChARM) [53] to predict the probability distribution of latent variables." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "text", + "content": "Specifically, given a source image " + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "text", + "content": ", the encoder " + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "text", + "content": " converts it to latent representations " + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "text", + "content": " conditioned on the given group mask " + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\pmb{m} = \\sum_{i=1}^{N} \\pmb{m}_{i}" + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "inline_equation", + "content": "i \\in \\{1,2,3,\\dots,N\\}" + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "text", + "content": ", indicating the index of the corresponding group in all " + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "text", + "content": " groups. In each " + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\pmb{m}_{i}" + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "text", + "content": ", the values of positions corresponding to the " + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "text", + "content": " group are set as " + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "text", + "content": ", and all values of other positions are set as 0. Similarly, we can define " + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\pmb{x} = \\sum_{i=1}^{N} \\pmb{x}_{i}" + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\pmb{y} = \\sum_{i=1}^{N} \\pmb{y}_{i}" + }, + { + "bbox": [ + 47, + 353, + 287, + 462 + ], + "type": "text", + "content": " from the perspective of group, and transform only occurs within each group, which is given by" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 129, + 468, + 287, + 481 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 468, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 129, + 468, + 287, + 481 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {i} = g _ {a} \\left(\\boldsymbol {x} _ {i} \\mid \\boldsymbol {m} _ {i}\\right). \\tag {1}", + "image_path": "606751a87b8f5865e374c7183c55f89c00b55fa99ced9980da82bf57c98b1ecd.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 489, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 287, + 534 + ], + "type": "text", + "content": "Thanks to the group-independent transform, we can easily implement the transform of all groups within one forward-pass, thus, the process can also be written as " + }, + { + "bbox": [ + 47, + 489, + 287, + 534 + ], + "type": "inline_equation", + "content": "\\pmb{y} = g_{a}(\\pmb{x}|\\pmb{m})" + }, + { + "bbox": [ + 47, + 489, + 287, + 534 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "text", + "content": "Then the latents are quantized to discrete representations " + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "text", + "content": " and selectively transmitted to the decoder side with predicted probability distribution through the entropy model. Concretely, the side information " + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{z}" + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "text", + "content": " is extracted by the hyper-encoder " + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{z} = h_{a}(\\pmb{y})" + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "text", + "content": ". The quantized hyper-latent " + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{z}} = Q(\\pmb{z})" + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "text", + "content": " is modeled and entropy-coded with a learned factorized prior. In ChARM, latent " + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "text", + "content": " is split into " + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "text", + "content": " (we choose " + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "inline_equation", + "content": "S = 10" + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "text", + "content": ") slices along with channel dimension, and each slice " + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_s" + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "text", + "content": " is entropy-coded based on the previous slices " + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_{< s}" + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 47, + 534, + 287, + 665 + ], + "type": "text", + "content": " is modeled by a conditional Gaussian distribution convolving with a unit uniform noise." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "type": "text", + "content": "After selectively entropy decoding the quantized latents, we have " + }, + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_p = \\sum_{j\\in \\mathbf{P}}\\hat{\\pmb{y}}_j" + }, + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{P}\\subseteq \\{1,2,\\dots ,N\\}" + }, + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "type": "text", + "content": " is the gather of required groups. The reconstruction image " + }, + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\hat{x}_p = \\sum_{j\\in \\mathbf{P}}\\hat{x}_j" + }, + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "type": "text", + "content": " will be calculated as follows" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 310, + 67, + 545, + 137 + ], + "blocks": [ + { + "bbox": [ + 310, + 67, + 545, + 137 + ], + "lines": [ + { + "bbox": [ + 310, + 67, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 310, + 67, + 545, + 137 + ], + "type": "image", + "image_path": "260198688b7f50c623cc5d37cb6e3268728877286a4ba52d8e5c7cfbd0f3b894.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 310, + 160, + 547, + 230 + ], + "blocks": [ + { + "bbox": [ + 46, + 152, + 287, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 152, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 287, + 224 + ], + "type": "text", + "content": "Figure 4. Illustration of flexibility to customize the group mask: (a). Taking overlapping objects as one group with the guidance of bounding boxes. (b). Taking overlapping objects as one group with the guidance of instance masks. (c) Taking overlapping objects as distinct groups with the guidance of instance masks." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 324, + 144, + 527, + 156 + ], + "lines": [ + { + "bbox": [ + 324, + 144, + 527, + 156 + ], + "spans": [ + { + "bbox": [ + 324, + 144, + 527, + 156 + ], + "type": "text", + "content": "(a) GI window partition under regular window partition." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 160, + 547, + 230 + ], + "lines": [ + { + "bbox": [ + 310, + 160, + 547, + 230 + ], + "spans": [ + { + "bbox": [ + 310, + 160, + 547, + 230 + ], + "type": "image", + "image_path": "29995e04e180ef31d391f4849e88039ea00f178dd36534d0563f50a847c3de6a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 236, + 527, + 247 + ], + "lines": [ + { + "bbox": [ + 324, + 236, + 527, + 247 + ], + "spans": [ + { + "bbox": [ + 324, + 236, + 527, + 247 + ], + "type": "text", + "content": "(b) GI window partition under shifted window partition." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 257, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 305, + 257, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 257, + 545, + 293 + ], + "type": "text", + "content": "Figure 5. Group-independent window partition of GI SwinBlock under regular and shifted window partition. Self-attention is conducted inside each local region." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 385, + 306, + 545, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 306, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 385, + 306, + 545, + 321 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {x}} _ {j} = g _ {s} \\left(\\hat {\\boldsymbol {y}} _ {j} \\mid \\boldsymbol {m} _ {j}\\right). \\tag {2}", + "image_path": "39f9ffcf044414dafd5af4d0984d9d4a01a2ed14a6c3f165785dd4cdb8a4a35a.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 327, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 327, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 545, + 387 + ], + "type": "text", + "content": "Similar to Equation (1), Equation (2) can also be written as " + }, + { + "bbox": [ + 305, + 327, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}} = g_s(\\hat{\\pmb{y}} |\\pmb {m})" + }, + { + "bbox": [ + 305, + 327, + 545, + 387 + ], + "type": "text", + "content": ". Note that the rounding operation is non-differentiable, thus we apply uniform noise for learning the entropy model and use a rounded one as the input of the decoder " + }, + { + "bbox": [ + 305, + 327, + 545, + 387 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 305, + 327, + 545, + 387 + ], + "type": "text", + "content": " as in [53]." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 396, + 387, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 396, + 387, + 410 + ], + "spans": [ + { + "bbox": [ + 306, + 396, + 387, + 410 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 416, + 413, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 416, + 413, + 429 + ], + "spans": [ + { + "bbox": [ + 306, + 416, + 413, + 429 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "text", + "content": "Image Codec. We train our high bitrate models " + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "inline_equation", + "content": "(>= 0.6\\mathrm{bpp})" + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "text", + "content": " on the Flickr 2K dataset, which has the same setting as [74] and train our low bitrate models " + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "inline_equation", + "content": "(< 0.6\\mathrm{bpp})" + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "text", + "content": " on the COCO 2017 training set [45]. Ours-Hyper are trained with 3.2M iterations, while Ours-ChARM models are trained by initializing their weights from higher bitrate hyperprior models for another 1.2M iterations. Each batch contains 8 random " + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "text", + "content": " crops from the training dataset. The learning rate is set as " + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "inline_equation", + "content": "5e - 5" + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "text", + "content": " and is decayed by a factor of 10 at " + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "inline_equation", + "content": "2.8M" + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "text", + "content": " iterations. Training loss " + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "inline_equation", + "content": "L = R + \\beta D" + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "text", + "content": " is the weighted combination of the rate-distortion trade-off, balanced by the Lagrange multiplier " + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "text", + "content": ". We adopt the mean squared error (MSE) in the RGB color space as the distortion metric " + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "text", + "content": ". The bitrate range of our image codec is achieved by applying different " + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "text", + "content": " that " + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\beta \\in \\{512, 1024, 2048, 4096\\}" + }, + { + "bbox": [ + 304, + 436, + 545, + 664 + ], + "type": "text", + "content": ". During the training stage, we randomly generate the group masks for the training of the group-independent transform, the detailed operations are shown in the supplementary material." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "Evaluation Datasets and Protocol. For the entire image reconstruction quality, we use the widely-used Kodak dataset [37] to evaluate the coding efficiency of our models. We evaluate the selective reconstruction quality on objects" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17241" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 64, + 298, + 267 + ], + "blocks": [ + { + "bbox": [ + 93, + 64, + 298, + 267 + ], + "lines": [ + { + "bbox": [ + 93, + 64, + 298, + 267 + ], + "spans": [ + { + "bbox": [ + 93, + 64, + 298, + 267 + ], + "type": "image", + "image_path": "9b12c7d4a0e73874705e51a40266a9d7bb47f380cc053c4c9c90bc580acf9213.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 269, + 271, + 279 + ], + "lines": [ + { + "bbox": [ + 119, + 269, + 271, + 279 + ], + "spans": [ + { + "bbox": [ + 119, + 269, + 271, + 279 + ], + "type": "text", + "content": "(a) Entire image reconstruction on Kodak." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 317, + 64, + 514, + 159 + ], + "blocks": [ + { + "bbox": [ + 317, + 64, + 514, + 159 + ], + "lines": [ + { + "bbox": [ + 317, + 64, + 514, + 159 + ], + "spans": [ + { + "bbox": [ + 317, + 64, + 514, + 159 + ], + "type": "image", + "image_path": "454acd5af833d1e28fe6c84492629d937cfdd848dcee5678633f682360df2fbb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 160, + 509, + 170 + ], + "lines": [ + { + "bbox": [ + 322, + 160, + 509, + 170 + ], + "spans": [ + { + "bbox": [ + 322, + 160, + 509, + 170 + ], + "type": "text", + "content": "(b) All categories objects reconstruction on COCO." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 318, + 173, + 512, + 267 + ], + "blocks": [ + { + "bbox": [ + 318, + 173, + 512, + 267 + ], + "lines": [ + { + "bbox": [ + 318, + 173, + 512, + 267 + ], + "spans": [ + { + "bbox": [ + 318, + 173, + 512, + 267 + ], + "type": "image", + "image_path": "05afc7ad8aab7a1a4c1e804b9e235000091a7895845f4c471278342b1924d8c5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 269, + 512, + 279 + ], + "lines": [ + { + "bbox": [ + 318, + 269, + 512, + 279 + ], + "spans": [ + { + "bbox": [ + 318, + 269, + 512, + 279 + ], + "type": "text", + "content": "(c) Human category objects reconstruction on COCO." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 53, + 317, + 167, + 405 + ], + "blocks": [ + { + "bbox": [ + 46, + 283, + 545, + 308 + ], + "lines": [ + { + "bbox": [ + 46, + 283, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 545, + 308 + ], + "type": "text", + "content": "Figure 6. Rate distortion comparison of compression efficiency on both entire and partial reconstruction scenarios. The PSNR of partial reconstructions is only calculated on the pre-detected bounding boxes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 317, + 167, + 405 + ], + "lines": [ + { + "bbox": [ + 53, + 317, + 167, + 405 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 167, + 405 + ], + "type": "image", + "image_path": "86f7d85cbcfd47b5265e68850ba59a0ba7be47d47417f88613bd42ec786aa812.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 170, + 317, + 282, + 405 + ], + "blocks": [ + { + "bbox": [ + 170, + 317, + 282, + 405 + ], + "lines": [ + { + "bbox": [ + 170, + 317, + 282, + 405 + ], + "spans": [ + { + "bbox": [ + 170, + 317, + 282, + 405 + ], + "type": "image", + "image_path": "353c8bda89cf67445284c9a35188dad4bc74df5b7b498bec08ef0bae8c6d4a9e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 407, + 287, + 431 + ], + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 431 + ], + "type": "text", + "content": "Figure 7. Performance comparison on instance segmentation (left) and pose estimation (right) on COCO." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 440, + 289, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 440, + 289, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 289, + 632 + ], + "type": "text", + "content": "of all categories and human objects using 40 and 20 images, which are randomly selected from the COCO 2017 validation set. The PSNR of the region of interest serves as the metric for measuring the objective quality. Additionally, the bits per pixel (bpp) are calculated by dividing the total bits of the transmitted bitstream by the number of pixels in the region of interest. To verify the effectiveness of our proposed method in supporting downstream tasks, we conduct experiments on instance segmentation and pose estimation using the COCO 2017 validation set. [45]. This dataset is widely used for dense prediction tasks and comprises 118K training, 5K validation, and 20K test-dev images. We evaluate the performances of instance segmentation and human keypoint detection using Mask R-CNN (X101-FPN backbone) and Keypoint R-CNN (X101-FPN backbone), respectively, implemented in the Detectron2 toolbox [68]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 634, + 167, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 167, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 167, + 647 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "content": "For comparisons, we choose recent neural image codes [16,27,52,75], powerful classical image codes [10, 61], and the functional codec that supports semantically structured bitstream [62]. Among them, ELIC [27] is reproduced and performs closely as their report. It is crit-" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 317, + 547, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 547, + 498 + ], + "type": "text", + "content": "ical to emphasize that the encoder does not possess prior knowledge regarding the regions of interest required by the decoder. And we utilize object detection as the pre-analysis by default, which is one of the most fundamental visual analysis tools. The semantically structured bitstreams are generated based on the bounding boxes detected by the detector of Mask R-CNN (X101-FPN backbone), and are selectively decoded and reconstructed into images for evaluation according to the requirements of relative semantics. In all experiments of quantitative results, the block size " + }, + { + "bbox": [ + 304, + 317, + 547, + 498 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 317, + 547, + 498 + ], + "type": "text", + "content": " of group masks is set to 32 to balance the trade-off between performance and overhead bitrate, and we take the union of overlapping groups as one group. More experiments compared with other methods of image coding for machines are presented in the supplementary." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": "Reconstruction Quality. For entire image reconstruction, Figure. 6a demonstrates that our proposed model achieves state-of-the-art rate-distortion performance, surpassing VTM 18.2 in terms of PSNR at all bitrates. Notably, structuring the bitstream according to semantics introduces negligible effects. Therefore, our model has a strong potential to achieve both high coding efficiency and functionality simultaneously. For partial image reconstruction in scenarios of specific interest, as shown in Figure. 6b and Figure. 6c, our model has achieved a significant improvement compared to other CODECs. Specifically, the semantically structured bitstream enabled our model and SSIC [62] to avoid transmitting and decoding bitstreams corresponding to the entire image, which is required of ELIC [27]. Additionally, our model's flexible block-level group mask and group-independent transform spatially decouple latents from the semantic level, resulting in sharper and more re" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "17242" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 63, + 136, + 118 + ], + "blocks": [ + { + "bbox": [ + 56, + 63, + 136, + 118 + ], + "lines": [ + { + "bbox": [ + 56, + 63, + 136, + 118 + ], + "spans": [ + { + "bbox": [ + 56, + 63, + 136, + 118 + ], + "type": "image", + "image_path": "9a7d6896638011c7da58bfb5d5a4d44d52557be07a841e5bd38414a2e789d8df.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 139, + 64, + 219, + 118 + ], + "blocks": [ + { + "bbox": [ + 139, + 64, + 219, + 118 + ], + "lines": [ + { + "bbox": [ + 139, + 64, + 219, + 118 + ], + "spans": [ + { + "bbox": [ + 139, + 64, + 219, + 118 + ], + "type": "image", + "image_path": "d26c13f5ecbf86b15f0e1465c92066dad84819ca749f521a992c3f2938c59cc1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 118, + 201, + 126 + ], + "lines": [ + { + "bbox": [ + 157, + 118, + 201, + 126 + ], + "spans": [ + { + "bbox": [ + 157, + 118, + 201, + 126 + ], + "type": "text", + "content": "0.516/29.36" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 223, + 64, + 302, + 118 + ], + "blocks": [ + { + "bbox": [ + 223, + 64, + 302, + 118 + ], + "lines": [ + { + "bbox": [ + 223, + 64, + 302, + 118 + ], + "spans": [ + { + "bbox": [ + 223, + 64, + 302, + 118 + ], + "type": "image", + "image_path": "b82832be0cdaefe0fb89afa2c13c282dce936f6ee9ef022d4683c3edd1066af1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 118, + 285, + 126 + ], + "lines": [ + { + "bbox": [ + 242, + 118, + 285, + 126 + ], + "spans": [ + { + "bbox": [ + 242, + 118, + 285, + 126 + ], + "type": "text", + "content": "0.251/29.40" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 307, + 64, + 386, + 118 + ], + "blocks": [ + { + "bbox": [ + 307, + 64, + 386, + 118 + ], + "lines": [ + { + "bbox": [ + 307, + 64, + 386, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 64, + 386, + 118 + ], + "type": "image", + "image_path": "91153e25e2b2538a795999db486136efc42bc8df64aa3cf632dee8600fcbb228.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 118, + 369, + 126 + ], + "lines": [ + { + "bbox": [ + 326, + 118, + 369, + 126 + ], + "spans": [ + { + "bbox": [ + 326, + 118, + 369, + 126 + ], + "type": "text", + "content": "0.195 / 30.14" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 402, + 64, + 435, + 118 + ], + "blocks": [ + { + "bbox": [ + 402, + 64, + 435, + 118 + ], + "lines": [ + { + "bbox": [ + 402, + 64, + 435, + 118 + ], + "spans": [ + { + "bbox": [ + 402, + 64, + 435, + 118 + ], + "type": "image", + "image_path": "2321acaaa637c289a5c0a1f8e3f959ffd0cb018b4c55e85bea0cd98459d59d91.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 438, + 64, + 469, + 118 + ], + "blocks": [ + { + "bbox": [ + 438, + 64, + 469, + 118 + ], + "lines": [ + { + "bbox": [ + 438, + 64, + 469, + 118 + ], + "spans": [ + { + "bbox": [ + 438, + 64, + 469, + 118 + ], + "type": "image", + "image_path": "2e1c2e28dc631c2fdcd56c02a5221d6e48d989aa79fe96945ba03b26ca87ddfb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 473, + 64, + 504, + 118 + ], + "blocks": [ + { + "bbox": [ + 473, + 64, + 504, + 118 + ], + "lines": [ + { + "bbox": [ + 473, + 64, + 504, + 118 + ], + "spans": [ + { + "bbox": [ + 473, + 64, + 504, + 118 + ], + "type": "image", + "image_path": "f357cb43ff7ddff1b85d23af659ca4e98589f68dcb4b8cd5c7ee436d77d2a67d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 507, + 64, + 539, + 118 + ], + "blocks": [ + { + "bbox": [ + 507, + 64, + 539, + 118 + ], + "lines": [ + { + "bbox": [ + 507, + 64, + 539, + 118 + ], + "spans": [ + { + "bbox": [ + 507, + 64, + 539, + 118 + ], + "type": "image", + "image_path": "60b72b7b5778c6a5edad1e354da5f0f2a811a9de8c487dbf196f0444f14857a6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 56, + 128, + 135, + 179 + ], + "blocks": [ + { + "bbox": [ + 56, + 128, + 135, + 179 + ], + "lines": [ + { + "bbox": [ + 56, + 128, + 135, + 179 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 135, + 179 + ], + "type": "image", + "image_path": "6596592032ef0de60f2972318ae404dc93639bf21f1ce1008c80b23f6288bb76.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 190, + 116, + 199 + ], + "lines": [ + { + "bbox": [ + 67, + 190, + 116, + 199 + ], + "spans": [ + { + "bbox": [ + 67, + 190, + 116, + 199 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 139, + 128, + 219, + 179 + ], + "blocks": [ + { + "bbox": [ + 139, + 128, + 219, + 179 + ], + "lines": [ + { + "bbox": [ + 139, + 128, + 219, + 179 + ], + "spans": [ + { + "bbox": [ + 139, + 128, + 219, + 179 + ], + "type": "image", + "image_path": "5815dc1ed2654d37d11e6a5ea48fc4b30b7d7b13e331035ac5fda3288bfc736b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 180, + 200, + 189 + ], + "lines": [ + { + "bbox": [ + 157, + 180, + 200, + 189 + ], + "spans": [ + { + "bbox": [ + 157, + 180, + 200, + 189 + ], + "type": "text", + "content": "0.318/28.61" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 170, + 190, + 186, + 198 + ], + "lines": [ + { + "bbox": [ + 170, + 190, + 186, + 198 + ], + "spans": [ + { + "bbox": [ + 170, + 190, + 186, + 198 + ], + "type": "text", + "content": "ELIC" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 223, + 128, + 302, + 179 + ], + "blocks": [ + { + "bbox": [ + 223, + 128, + 302, + 179 + ], + "lines": [ + { + "bbox": [ + 223, + 128, + 302, + 179 + ], + "spans": [ + { + "bbox": [ + 223, + 128, + 302, + 179 + ], + "type": "image", + "image_path": "c9a827a04c52c7686e159b55bdd6a0952adfae61c751cdbf42badde15c87a27c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 180, + 283, + 189 + ], + "lines": [ + { + "bbox": [ + 241, + 180, + 283, + 189 + ], + "spans": [ + { + "bbox": [ + 241, + 180, + 283, + 189 + ], + "type": "text", + "content": "0.295/28.53" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 256, + 190, + 271, + 198 + ], + "lines": [ + { + "bbox": [ + 256, + 190, + 271, + 198 + ], + "spans": [ + { + "bbox": [ + 256, + 190, + 271, + 198 + ], + "type": "text", + "content": "SSIC" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 307, + 128, + 386, + 179 + ], + "blocks": [ + { + "bbox": [ + 307, + 128, + 386, + 179 + ], + "lines": [ + { + "bbox": [ + 307, + 128, + 386, + 179 + ], + "spans": [ + { + "bbox": [ + 307, + 128, + 386, + 179 + ], + "type": "image", + "image_path": "c8c1e82b2c0da3fcbbfaaee39d9ebac3dd039bccb0aceaf89619198a4d7d54f8.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 180, + 367, + 189 + ], + "lines": [ + { + "bbox": [ + 324, + 180, + 367, + 189 + ], + "spans": [ + { + "bbox": [ + 324, + 180, + 367, + 189 + ], + "type": "text", + "content": "0.316 / 29.25" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 339, + 190, + 356, + 198 + ], + "lines": [ + { + "bbox": [ + 339, + 190, + 356, + 198 + ], + "spans": [ + { + "bbox": [ + 339, + 190, + 356, + 198 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 402, + 128, + 433, + 153 + ], + "blocks": [ + { + "bbox": [ + 402, + 128, + 433, + 153 + ], + "lines": [ + { + "bbox": [ + 402, + 128, + 433, + 153 + ], + "spans": [ + { + "bbox": [ + 402, + 128, + 433, + 153 + ], + "type": "image", + "image_path": "5f2953d350535e7801e274afa9f8d1063982131c77f25427a099cd4d84f3be2e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 402, + 154, + 433, + 179 + ], + "blocks": [ + { + "bbox": [ + 402, + 154, + 433, + 179 + ], + "lines": [ + { + "bbox": [ + 402, + 154, + 433, + 179 + ], + "spans": [ + { + "bbox": [ + 402, + 154, + 433, + 179 + ], + "type": "image", + "image_path": "adece8258b93916b6efd26abe44ab7f10348c55c69fafce5b5d6b4d2f710cfce.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 437, + 128, + 469, + 153 + ], + "blocks": [ + { + "bbox": [ + 437, + 128, + 469, + 153 + ], + "lines": [ + { + "bbox": [ + 437, + 128, + 469, + 153 + ], + "spans": [ + { + "bbox": [ + 437, + 128, + 469, + 153 + ], + "type": "image", + "image_path": "9911aecd9cc119cf8b57d5b021fd2ab6a46f042fb417a95b1bd287df5dc9bbdd.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 438, + 154, + 469, + 179 + ], + "blocks": [ + { + "bbox": [ + 438, + 154, + 469, + 179 + ], + "lines": [ + { + "bbox": [ + 438, + 154, + 469, + 179 + ], + "spans": [ + { + "bbox": [ + 438, + 154, + 469, + 179 + ], + "type": "image", + "image_path": "2ea4415c651ebfca4edbf6c5af7750d68a43b3fac712ef5bb3eccb1f4c7dfe15.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 473, + 128, + 504, + 153 + ], + "blocks": [ + { + "bbox": [ + 473, + 128, + 504, + 153 + ], + "lines": [ + { + "bbox": [ + 473, + 128, + 504, + 153 + ], + "spans": [ + { + "bbox": [ + 473, + 128, + 504, + 153 + ], + "type": "image", + "image_path": "608398d18aaf1880d92b461c09528722f51fa2b0b3bbda7116c4551a529a0d70.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 473, + 154, + 504, + 179 + ], + "blocks": [ + { + "bbox": [ + 473, + 154, + 504, + 179 + ], + "lines": [ + { + "bbox": [ + 473, + 154, + 504, + 179 + ], + "spans": [ + { + "bbox": [ + 473, + 154, + 504, + 179 + ], + "type": "image", + "image_path": "2ade7db95446bd023e7632e8331c570c6f4ac7f7e4808aec24940599d2122fba.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 507, + 128, + 539, + 153 + ], + "blocks": [ + { + "bbox": [ + 507, + 128, + 539, + 153 + ], + "lines": [ + { + "bbox": [ + 507, + 128, + 539, + 153 + ], + "spans": [ + { + "bbox": [ + 507, + 128, + 539, + 153 + ], + "type": "image", + "image_path": "c2fc8807996df6a36b79859eeeca1e3317e3151ff8419059fb1432c567730fc3.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 507, + 154, + 539, + 179 + ], + "blocks": [ + { + "bbox": [ + 507, + 154, + 539, + 179 + ], + "lines": [ + { + "bbox": [ + 507, + 154, + 539, + 179 + ], + "spans": [ + { + "bbox": [ + 507, + 154, + 539, + 179 + ], + "type": "image", + "image_path": "8415f5019a1d8711b9a137210601a8f997b8e81047da306068fcf626e1d7b406.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 545, + 225 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 545, + 225 + ], + "type": "text", + "content": "Figure 8. Visualization of reconstructed images for human consumption from the COCO 2017 dataset, in the scenario where foreground objects are the regions of interest. The numbers below the images correspond to bits per pixel and PSNR." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 50, + 238, + 145, + 309 + ], + "blocks": [ + { + "bbox": [ + 50, + 238, + 145, + 309 + ], + "lines": [ + { + "bbox": [ + 50, + 238, + 145, + 309 + ], + "spans": [ + { + "bbox": [ + 50, + 238, + 145, + 309 + ], + "type": "image", + "image_path": "a13a0777271ce508fd1754ea8677adf56d6b83d632d1b3afe9937524c4a52099.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 149, + 238, + 244, + 308 + ], + "blocks": [ + { + "bbox": [ + 149, + 238, + 244, + 308 + ], + "lines": [ + { + "bbox": [ + 149, + 238, + 244, + 308 + ], + "spans": [ + { + "bbox": [ + 149, + 238, + 244, + 308 + ], + "type": "image", + "image_path": "aa1c98fa10d31e1c2136f8e7d61b4021f37b7bae999e9f18f5f1ff57e58abf21.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 249, + 238, + 343, + 308 + ], + "blocks": [ + { + "bbox": [ + 249, + 238, + 343, + 308 + ], + "lines": [ + { + "bbox": [ + 249, + 238, + 343, + 308 + ], + "spans": [ + { + "bbox": [ + 249, + 238, + 343, + 308 + ], + "type": "image", + "image_path": "861d3d4a4d8256d62b1215bbe6b9b627198ca1d1994661459f2a6fda843512dc.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 309, + 310, + 318 + ], + "lines": [ + { + "bbox": [ + 280, + 309, + 310, + 318 + ], + "spans": [ + { + "bbox": [ + 280, + 309, + 310, + 318 + ], + "type": "text", + "content": "1.850bpp" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 348, + 238, + 443, + 308 + ], + "blocks": [ + { + "bbox": [ + 348, + 238, + 443, + 308 + ], + "lines": [ + { + "bbox": [ + 348, + 238, + 443, + 308 + ], + "spans": [ + { + "bbox": [ + 348, + 238, + 443, + 308 + ], + "type": "image", + "image_path": "330b8d1349b7574c9d6833e3b9f55fc3ebf4f4452a1c3e7aa65971fb0e685025.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 309, + 410, + 318 + ], + "lines": [ + { + "bbox": [ + 380, + 309, + 410, + 318 + ], + "spans": [ + { + "bbox": [ + 380, + 309, + 410, + 318 + ], + "type": "text", + "content": "1.592bpp" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 448, + 238, + 541, + 308 + ], + "blocks": [ + { + "bbox": [ + 448, + 238, + 541, + 308 + ], + "lines": [ + { + "bbox": [ + 448, + 238, + 541, + 308 + ], + "spans": [ + { + "bbox": [ + 448, + 238, + 541, + 308 + ], + "type": "image", + "image_path": "49fa984a24ca38df917f62282ed49a2bea31b9f456fb45145a7541e5c816f5ff.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 309, + 509, + 318 + ], + "lines": [ + { + "bbox": [ + 479, + 309, + 509, + 318 + ], + "spans": [ + { + "bbox": [ + 479, + 309, + 509, + 318 + ], + "type": "text", + "content": "1.211bpp" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 50, + 319, + 144, + 376 + ], + "blocks": [ + { + "bbox": [ + 50, + 319, + 144, + 376 + ], + "lines": [ + { + "bbox": [ + 50, + 319, + 144, + 376 + ], + "spans": [ + { + "bbox": [ + 50, + 319, + 144, + 376 + ], + "type": "image", + "image_path": "acd3064c9b691bceecd9ea5b338e637633cc619e06e0d49549e35befe6077a76.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 387, + 121, + 397 + ], + "lines": [ + { + "bbox": [ + 73, + 387, + 121, + 397 + ], + "spans": [ + { + "bbox": [ + 73, + 387, + 121, + 397 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 399, + 545, + 423 + ], + "lines": [ + { + "bbox": [ + 46, + 399, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 545, + 423 + ], + "type": "text", + "content": "Figure 9. Visualization of reconstructed images for downstream tasks from the COCO 2017 dataset. The first line corresponds to instance segmentation, and the second one corresponds to pose estimation." + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 149, + 319, + 244, + 376 + ], + "blocks": [ + { + "bbox": [ + 149, + 319, + 244, + 376 + ], + "lines": [ + { + "bbox": [ + 149, + 319, + 244, + 376 + ], + "spans": [ + { + "bbox": [ + 149, + 319, + 244, + 376 + ], + "type": "image", + "image_path": "1506b4316dbd3ba8017e120cfb41ee1f8d2d8f01e5922c859f08579b8a1222cf.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 387, + 216, + 396 + ], + "lines": [ + { + "bbox": [ + 171, + 387, + 216, + 396 + ], + "spans": [ + { + "bbox": [ + 171, + 387, + 216, + 396 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 249, + 319, + 342, + 376 + ], + "blocks": [ + { + "bbox": [ + 249, + 319, + 342, + 376 + ], + "lines": [ + { + "bbox": [ + 249, + 319, + 342, + 376 + ], + "spans": [ + { + "bbox": [ + 249, + 319, + 342, + 376 + ], + "type": "image", + "image_path": "f7f548d9eee40c0acab91e4583932cf93071d80c036a465f43df8a5714770586.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 376, + 310, + 385 + ], + "lines": [ + { + "bbox": [ + 280, + 376, + 310, + 385 + ], + "spans": [ + { + "bbox": [ + 280, + 376, + 310, + 385 + ], + "type": "text", + "content": "2.229bpp" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 288, + 387, + 303, + 395 + ], + "lines": [ + { + "bbox": [ + 288, + 387, + 303, + 395 + ], + "spans": [ + { + "bbox": [ + 288, + 387, + 303, + 395 + ], + "type": "text", + "content": "ELIC" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 347, + 319, + 441, + 375 + ], + "blocks": [ + { + "bbox": [ + 347, + 319, + 441, + 375 + ], + "lines": [ + { + "bbox": [ + 347, + 319, + 441, + 375 + ], + "spans": [ + { + "bbox": [ + 347, + 319, + 441, + 375 + ], + "type": "image", + "image_path": "d811158090ed31fc98c9a7d750aff7fa179417fb1a258abe2e11091460623ccb.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 379, + 376, + 410, + 385 + ], + "lines": [ + { + "bbox": [ + 379, + 376, + 410, + 385 + ], + "spans": [ + { + "bbox": [ + 379, + 376, + 410, + 385 + ], + "type": "text", + "content": "1.335bpp" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 387, + 387, + 402, + 395 + ], + "lines": [ + { + "bbox": [ + 387, + 387, + 402, + 395 + ], + "spans": [ + { + "bbox": [ + 387, + 387, + 402, + 395 + ], + "type": "text", + "content": "SSIC" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 447, + 319, + 541, + 375 + ], + "blocks": [ + { + "bbox": [ + 447, + 319, + 541, + 375 + ], + "lines": [ + { + "bbox": [ + 447, + 319, + 541, + 375 + ], + "spans": [ + { + "bbox": [ + 447, + 319, + 541, + 375 + ], + "type": "image", + "image_path": "8f696fbfa4d9872515a66aa0bee837dbf4d4afb5e2e216a55c05aae7b68a03eb.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 478, + 376, + 509, + 385 + ], + "lines": [ + { + "bbox": [ + 478, + 376, + 509, + 385 + ], + "spans": [ + { + "bbox": [ + 478, + 376, + 509, + 385 + ], + "type": "text", + "content": "1.048bpp" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 485, + 387, + 503, + 395 + ], + "lines": [ + { + "bbox": [ + 485, + 387, + 503, + 395 + ], + "spans": [ + { + "bbox": [ + 485, + 387, + 503, + 395 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 49 + }, + { + "bbox": [ + 46, + 429, + 287, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 488 + ], + "type": "text", + "content": "alistic boundaries compared to SSIC [62]. Moreover, in the case of overlapping regions, our model can significantly save bitrate by grouping them into an irregular group instead of replacing them with a larger bounding box, which would introduce more distortion." + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 46, + 491, + 288, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 288, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 288, + 670 + ], + "type": "text", + "content": "Downstream Task Supporting. Figure. 7 demonstrates our excellent performance on instance segmentation and pose estimation. On instance segmentation, our model achieves significantly improved performance at low bitrates (<0.6bpp) compared with other methods, which can be attributed to the faithful reconstruction of RoI boundaries in our approach. It is worth noting that the superior performance of our method on pose estimation is due to the ability of our model to preserve fine details and boundaries, which are critical for accurately localizing human keypoints. In addition, human objects are a minority in the image and are often sparsely distributed, and compared with SSIC, our method can significantly reduce bitrate by avoiding transmitting the latents corresponding to the rectangular area containing all overlapped objects." + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 47, + 673, + 160, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 160, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 160, + 685 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Reconstruction for Human Consumption. When the reconstructed images are mainly intended for human con" + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 304, + 429, + 545, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 545, + 657 + ], + "type": "text", + "content": "sumption with specific semantic interests, as shown in Figure 8, bitstreams generated by CODEs designed for compressing regular rectangular images are required to be fully transmitted and serve full reconstruction, regardless of the requirements and the content of images. Although SSIC [62] partially addresses this issue through generating SSB through bounding box based division, it lacks flexibility and may introduce additional irrelevant disturbance when merging overlapping objects into larger bounding boxes. Moreover, SSIC generates the bitstream by directly dividing the corresponding positions in the latent variable space after the transform based on ConvNets, which can lead to lost dependencies during selective reconstruction, resulting in blurred and distorted region boundaries. With group mask based partitioning and group-independent transform, our method decouples images to generate the SSB more efficiently while ensuring that no distortion or blurring is introduced after selective reconstruction, leading to both significant bitrate saving and a better visual experience." + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": "Reconstruction for Intelligent Tasks. When using reconstructed images for downstream intelligent tasks, selectively transmitting and decoding the semantically structured bitstream based on the prior knowledge of relative seman" + } + ] + } + ], + "index": 58 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17243" + } + ] + } + ], + "index": 59 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 64, + 105, + 101 + ], + "blocks": [ + { + "bbox": [ + 50, + 64, + 105, + 101 + ], + "lines": [ + { + "bbox": [ + 50, + 64, + 105, + 101 + ], + "spans": [ + { + "bbox": [ + 50, + 64, + 105, + 101 + ], + "type": "image", + "image_path": "799f42b5c5c2dcb10d91b8b35101445a0edbdc4ea3afb0a026a1b970294ae9bf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 103, + 105, + 140 + ], + "blocks": [ + { + "bbox": [ + 50, + 103, + 105, + 140 + ], + "lines": [ + { + "bbox": [ + 50, + 103, + 105, + 140 + ], + "spans": [ + { + "bbox": [ + 50, + 103, + 105, + 140 + ], + "type": "image", + "image_path": "e7fa75907c6ea135bb0420e8e2dd0678a0a05ebfef73509e511ed710500ac633.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 145, + 96, + 152 + ], + "lines": [ + { + "bbox": [ + 58, + 145, + 96, + 152 + ], + "spans": [ + { + "bbox": [ + 58, + 145, + 96, + 152 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 110, + 64, + 164, + 101 + ], + "blocks": [ + { + "bbox": [ + 110, + 64, + 164, + 101 + ], + "lines": [ + { + "bbox": [ + 110, + 64, + 164, + 101 + ], + "spans": [ + { + "bbox": [ + 110, + 64, + 164, + 101 + ], + "type": "image", + "image_path": "37fc29c7ed7d0877f883a4ea8c36c41c4fa9ce64956e6d5946ec7108b3b187eb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 110, + 103, + 164, + 140 + ], + "blocks": [ + { + "bbox": [ + 110, + 103, + 164, + 140 + ], + "lines": [ + { + "bbox": [ + 110, + 103, + 164, + 140 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 164, + 140 + ], + "type": "image", + "image_path": "d4db6a3afb325cfeeef3f20e5835f794e0dca570673340f8d777d1a6a7c915ac.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 145, + 153, + 152 + ], + "lines": [ + { + "bbox": [ + 121, + 145, + 153, + 152 + ], + "spans": [ + { + "bbox": [ + 121, + 145, + 153, + 152 + ], + "type": "text", + "content": "Pre-Analysis" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 168, + 64, + 224, + 101 + ], + "blocks": [ + { + "bbox": [ + 168, + 64, + 224, + 101 + ], + "lines": [ + { + "bbox": [ + 168, + 64, + 224, + 101 + ], + "spans": [ + { + "bbox": [ + 168, + 64, + 224, + 101 + ], + "type": "image", + "image_path": "a2799c7805e83a4a04e9f5a0784d05410ae31451129c4fd211fa829470a7017b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 168, + 103, + 224, + 140 + ], + "blocks": [ + { + "bbox": [ + 168, + 103, + 224, + 140 + ], + "lines": [ + { + "bbox": [ + 168, + 103, + 224, + 140 + ], + "spans": [ + { + "bbox": [ + 168, + 103, + 224, + 140 + ], + "type": "image", + "image_path": "43a71bd43fd4e03274e00f62bdb929c436d733b3f0c4338dd5c132687033a308.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 142, + 208, + 148 + ], + "lines": [ + { + "bbox": [ + 184, + 142, + 208, + 148 + ], + "spans": [ + { + "bbox": [ + 184, + 142, + 208, + 148 + ], + "type": "text", + "content": "Based on" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 228, + 64, + 284, + 101 + ], + "blocks": [ + { + "bbox": [ + 228, + 64, + 284, + 101 + ], + "lines": [ + { + "bbox": [ + 228, + 64, + 284, + 101 + ], + "spans": [ + { + "bbox": [ + 228, + 64, + 284, + 101 + ], + "type": "image", + "image_path": "8769c172d99d89cd87ba721a8bb8fcb780119bb35cd6cf6d0b62192c2879d15d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 228, + 103, + 284, + 140 + ], + "blocks": [ + { + "bbox": [ + 176, + 149, + 217, + 156 + ], + "lines": [ + { + "bbox": [ + 176, + 149, + 217, + 156 + ], + "spans": [ + { + "bbox": [ + 176, + 149, + 217, + 156 + ], + "type": "text", + "content": "Bounding Boxes" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 103, + 284, + 140 + ], + "lines": [ + { + "bbox": [ + 228, + 103, + 284, + 140 + ], + "spans": [ + { + "bbox": [ + 228, + 103, + 284, + 140 + ], + "type": "image", + "image_path": "0f6557947609d733be5376ff9746a43819fdc09b4b7e1962eeb2d422f0efc629.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 142, + 269, + 148 + ], + "lines": [ + { + "bbox": [ + 244, + 142, + 269, + 148 + ], + "spans": [ + { + "bbox": [ + 244, + 142, + 269, + 148 + ], + "type": "text", + "content": "Based on" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 237, + 149, + 275, + 156 + ], + "lines": [ + { + "bbox": [ + 237, + 149, + 275, + 156 + ], + "spans": [ + { + "bbox": [ + 237, + 149, + 275, + 156 + ], + "type": "text", + "content": "Instance Maps" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 52, + 201, + 105, + 239 + ], + "blocks": [ + { + "bbox": [ + 46, + 158, + 286, + 194 + ], + "lines": [ + { + "bbox": [ + 46, + 158, + 286, + 194 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 286, + 194 + ], + "type": "text", + "content": "Figure 10. Selective reconstructions from the semantically structured bitstream. The bitstream can be generated based on different pre-analysis methods." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 201, + 105, + 239 + ], + "lines": [ + { + "bbox": [ + 52, + 201, + 105, + 239 + ], + "spans": [ + { + "bbox": [ + 52, + 201, + 105, + 239 + ], + "type": "image", + "image_path": "6fa09b9e4356de1d3cfc882bdc6df10b92f7f6b79b3643749b9d5c70e87784fe.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 242, + 97, + 248 + ], + "lines": [ + { + "bbox": [ + 60, + 242, + 97, + 248 + ], + "spans": [ + { + "bbox": [ + 60, + 242, + 97, + 248 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 253, + 286, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 253, + 286, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 286, + 277 + ], + "type": "text", + "content": "Figure 11. Illustration of semantically-aware encryption based on our proposed method." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 111, + 201, + 164, + 239 + ], + "blocks": [ + { + "bbox": [ + 111, + 201, + 164, + 239 + ], + "lines": [ + { + "bbox": [ + 111, + 201, + 164, + 239 + ], + "spans": [ + { + "bbox": [ + 111, + 201, + 164, + 239 + ], + "type": "image", + "image_path": "b261705210627fc10e0391bf0a3ff55d4f1a540e4577571c0703c93c0f99b029.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 122, + 242, + 153, + 248 + ], + "lines": [ + { + "bbox": [ + 122, + 242, + 153, + 248 + ], + "spans": [ + { + "bbox": [ + 122, + 242, + 153, + 248 + ], + "type": "text", + "content": "Pre-Analysis" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 169, + 201, + 222, + 239 + ], + "blocks": [ + { + "bbox": [ + 169, + 201, + 222, + 239 + ], + "lines": [ + { + "bbox": [ + 169, + 201, + 222, + 239 + ], + "spans": [ + { + "bbox": [ + 169, + 201, + 222, + 239 + ], + "type": "image", + "image_path": "0c2f30714811b2af2c34764895014333a29fcf7d3308a8cf5052e087eede41ef.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 242, + 215, + 248 + ], + "lines": [ + { + "bbox": [ + 178, + 242, + 215, + 248 + ], + "spans": [ + { + "bbox": [ + 178, + 242, + 215, + 248 + ], + "type": "text", + "content": "Encrypt People" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 228, + 201, + 282, + 239 + ], + "blocks": [ + { + "bbox": [ + 228, + 201, + 282, + 239 + ], + "lines": [ + { + "bbox": [ + 228, + 201, + 282, + 239 + ], + "spans": [ + { + "bbox": [ + 228, + 201, + 282, + 239 + ], + "type": "image", + "image_path": "b773c9f4052d737dee2ed6239d88df1083153744309efc5282895de23c2fcdcc.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 237, + 239, + 274, + 246 + ], + "lines": [ + { + "bbox": [ + 237, + 239, + 274, + 246 + ], + "spans": [ + { + "bbox": [ + 237, + 239, + 274, + 246 + ], + "type": "text", + "content": "Encrypt People" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 239, + 246, + 271, + 251 + ], + "lines": [ + { + "bbox": [ + 239, + 246, + 271, + 251 + ], + "spans": [ + { + "bbox": [ + 239, + 246, + 271, + 251 + ], + "type": "text", + "content": "and Vehicles" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 289, + 286, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 286, + 468 + ], + "type": "text", + "content": "tics can significantly save bitrate. As illustrated in Figure 9, conventional codecs designed for compressing rectangular images require the entire bitstream to be transmitted and decoded. Downstream models then perform intelligent analytics on the fully reconstructed image, resulting in a waste of bitrate. Similar to image reconstruction for human consumption, SSIC can perform selective transmission to save bitrate. However, its bounding box based partitioning may not be optimal for bitrate savings, and the resulting blurred and distorted region boundaries and irrelevant content can further impede downstream intelligent analytics. (misidentifying the car as a person shown in Figure 9). Our method can more efficiently generate the SSB and support various downstream tasks with specific requirements with both accuracy and coding efficiency." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 472, + 286, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 286, + 614 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 286, + 614 + ], + "type": "text", + "content": "Customizability and Flexibility. It's flexible to customize the semantically structured bitstream based on different semantical partition criteria. For instance, as shown in Figure 10, the generation and usage of semantically structured bitstreams can be based on either object detection or instance segmentation, depending on the specific requirements. It should be noted that the methods of pre-analysis are not limited to object detection and instance segmentation as shown in Figure 10, but can include other techniques such as saliency detection [22, 30, 72], semantic segmentation [5, 48, 69, 71, 73], panoptic segmentation [15, 36], etc., and even human annotation." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "content": "Semantically-Aware Encryption. As presented in Figure 11, our proposed method can be applied to bitstream encryption, enabling selective and even layered encryption with semantic priors based on the user's security level. In some cases, the selectively encrypted SSB can allow for secure transmission and storage of sensitive information while minimizing the impact on visual quality and downstream analytics. Please refer to the supplementary material for" + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 312, + 64, + 421, + 151 + ], + "blocks": [ + { + "bbox": [ + 312, + 64, + 421, + 151 + ], + "lines": [ + { + "bbox": [ + 312, + 64, + 421, + 151 + ], + "spans": [ + { + "bbox": [ + 312, + 64, + 421, + 151 + ], + "type": "image", + "image_path": "456c61c5f8400fe58ecc3baef621b45b8a82c4c31dc7286b67f525c6784f8833.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 153, + 408, + 163 + ], + "lines": [ + { + "bbox": [ + 325, + 153, + 408, + 163 + ], + "spans": [ + { + "bbox": [ + 325, + 153, + 408, + 163 + ], + "type": "text", + "content": "(a) RoI reconstruction." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 372, + 165, + 479, + 177 + ], + "lines": [ + { + "bbox": [ + 372, + 165, + 479, + 177 + ], + "spans": [ + { + "bbox": [ + 372, + 165, + 479, + 177 + ], + "type": "text", + "content": "Figure 12. Ablation study." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 430, + 64, + 541, + 151 + ], + "blocks": [ + { + "bbox": [ + 430, + 64, + 541, + 151 + ], + "lines": [ + { + "bbox": [ + 430, + 64, + 541, + 151 + ], + "spans": [ + { + "bbox": [ + 430, + 64, + 541, + 151 + ], + "type": "image", + "image_path": "c3bcf2a27219e7eb2317217dfe73097f5e1251e12e9589ea6cdc245eef733f9d.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 153, + 540, + 163 + ], + "lines": [ + { + "bbox": [ + 430, + 153, + 540, + 163 + ], + "spans": [ + { + "bbox": [ + 430, + 153, + 540, + 163 + ], + "type": "text", + "content": "(b) Intelligent task supporting." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 186, + 412, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 186, + 412, + 198 + ], + "spans": [ + { + "bbox": [ + 306, + 186, + 412, + 198 + ], + "type": "text", + "content": "details of implementation." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 306, + 211, + 398, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 211, + 398, + 224 + ], + "spans": [ + { + "bbox": [ + 306, + 211, + 398, + 224 + ], + "type": "text", + "content": "5. Ablation Study" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 232, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 363 + ], + "type": "text", + "content": "In order to show the effectiveness of our proposed group mask and group-independent transform (GIT) for both RoI-aware selective reconstruction and downstream intelligent analytics, we conducted the ablation study at different bitrates. Specifically, we use people as the region of interest for RoI reconstruction, and the results for taking all foreground objects as RoI are presented in the supplementary material. Additionally, we conducted an ablation study on pose estimation of COCO 2017 for intelligent task support, and the results on instance segmentation are also included in the supplementary material." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 365, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 426 + ], + "type": "text", + "content": "Group Mask. Figure 12 shows that group mask based partitioning saves significant bitrate compared with naive bounding box based partitioning. Meanwhile, both the visual quality of RoI reconstruction and downstream task performances would not be impeded." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 428, + 545, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 428, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 428, + 545, + 500 + ], + "type": "text", + "content": "Group-Independent transform. As shown in Figure 12a, group-independent transform is crucial for the visual quality of RoI reconstruction. And the downstream task performance would be impeded by the blurred and distorted region boundaries without group-independent transform, which is shown in Figure 12b." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 306, + 513, + 378, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 513, + 378, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 513, + 378, + 525 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "In this work, we propose to generate semantically structured bitstream with strong functionality based on group masks, which are highly flexible, customizable, while maintaining lightweight overhead. Moreover, we first propose the concept of group-independent transform and instantiate it by designing the Group-Independent Swin-Block (GI-Swin Block) to ensure independence among distinct groups, resulting in pleasing reconstructions of RoI regions. Particularly, our proposed method outperforms VTM-18.2 and achieves comparable coding efficiency with the SOTA neural image CODECs, while offering strong functionality. The experimental results have demonstrated the effectiveness of our proposed method across several applications, including human consumption of regions of interest, downstream task support, and semantically-aware encryption." + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17244" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 120 + ], + "type": "text", + "content": "Acknowledgement. This work was supported in part by NSFC under Grant U1908209, 62021001. This work was also supported in part by ZJNSFC under Grant LQ23F010008." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 131, + 107, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 131, + 107, + 144 + ], + "spans": [ + { + "bbox": [ + 48, + 131, + 107, + 144 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 150, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 150, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 287, + 195 + ], + "type": "text", + "content": "[1] Eirikur Agustsson, Fabian Mentzer, Michael Tschannen, Lukas Cavigelli, Radu Timofte, Luca Benini, and Luc V Gool. Soft-to-hard vector quantization for end-to-end learning compressible representations. *NeurIPS*, 30, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 197, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 197, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 197, + 288, + 239 + ], + "type": "text", + "content": "[2] Eirikur Agustsson, Michael Tschannen, Fabian Mentzer, Radu Timofte, and Luc Van Gool. Generative adversarial networks for extreme learned image compression. In ICCV, pages 221-231, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 241, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 241, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 53, + 241, + 287, + 296 + ], + "type": "text", + "content": "[3] Mohammad Akbari, Jie Liang, and Jingning Han. Dsslic: Deep semantic segmentation-based layered image compression. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2042-2046. IEEE, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 297, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 287, + 330 + ], + "type": "text", + "content": "[4] Artem Babenko, Anton Slesarev, Alexandr Chigorin, and Victor Lempitsky. Neural codes for image retrieval. In ECCV, pages 584-599. Springer, 2014. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 331, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 287, + 374 + ], + "type": "text", + "content": "[5] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. TPAMI, 39(12):2481-2495, 2017. 1, 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 376, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 287, + 430 + ], + "type": "text", + "content": "[6] Johannes Ballé, Philip A Chou, David Minnen, Saurabh Singh, Nick Johnston, Eirikur Agustsson, Sung Jin Hwang, and George Toderici. Nonlinear transform coding. IEEE Journal of Selected Topics in Signal Processing, 15(2):339-353, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 432, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 432, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 53, + 432, + 287, + 454 + ], + "type": "text", + "content": "[7] Johannes Ballé, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. In *ICLR*, 2017. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 456, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 456, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 53, + 456, + 287, + 488 + ], + "type": "text", + "content": "[8] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In ICLR, 2018. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 490, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 490, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 53, + 490, + 287, + 521 + ], + "type": "text", + "content": "[9] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In ICCV, pages 9157-9166, 2019. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "text", + "content": "[10] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. TCSVT, 2021. 1, 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "type": "text", + "content": "[11] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 40(4):834-848, 2017. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "text", + "content": "[12] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, pages 801-818, 2018. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[13] Zhuo Chen, Kui Fan, Shiqi Wang, Lingyu Duan, Weisi Lin, and Alex Chichung Kot. Toward intelligent sensing: Intermediate deep feature compression. TIP, 29:2230-2243, 2019. 1, 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "text", + "content": "[14] Zhuo Chen, Kui Fan, Shiqi Wang, Ling-Yu Duan, Weisi Lin, and Alex Kot. Lossy intermediate deep learning feature compression and evaluation. In ACM MM, pages 2414-2422, 2019. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "type": "text", + "content": "[15] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, pages 12475-12485, 2020. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 172, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 215 + ], + "type": "text", + "content": "[16] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In CVPR, pages 7939-7948, 2020. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 217, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 271 + ], + "type": "text", + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "type": "text", + "content": "[18] Ling-Yu Duan, Vijay Chandrasekhar, Jie Chen, Jie Lin, Zhe Wang, Tiejun Huang, Bernd Girod, and Wen Gao. Overview of thempeg-cdvs standard. TIP, 25(1):179-194, 2015. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 304, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 358 + ], + "type": "text", + "content": "[19] Ling-Yu Duan, Yihang Lou, Yan Bai, Tiejun Huang, Wen Gao, Vijay Chandrasekhar, Jie Lin, Shiqi Wang, and Alex Chichung Kot. Compact descriptors for video analysis: The emerging mpg standard. IEEE MultiMedia, 26(2):44-54, 2018. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 360, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 392 + ], + "type": "text", + "content": "[20] Touradj Ebrahimi and Caspar Horne. Mpeg-4 natural video coding—an overview. Signal Processing: Image Communication, 15(4-5):365-385, 2000. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 393, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 437 + ], + "type": "text", + "content": "[21] Ruoyu Feng, Xin Jin, Zongyu Guo, Runsen Feng, Yixin Gao, Tianyu He, Zhizheng Zhang, Simeng Sun, and Zhibo Chen. Image coding for machines with omnipotent feature learning. In ECCV, pages 510-528. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 437, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 469 + ], + "type": "text", + "content": "[22] Stas Goferman, Lihi Zelnik-Manor, and Ayellet Tal. Context-aware saliency detection. TPAMI, 34(10):1915-1926, 2011. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 471, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 491 + ], + "type": "text", + "content": "[23] Vivek K Goyal. Theoretical foundations of transform coding. IEEE Signal Processing Magazine, 18(5):9-21, 2001. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 492, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 545, + 525 + ], + "type": "text", + "content": "[24] Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Causal contextual prediction for learned image compression. TCSVT, 32(4):2329-2341, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 525, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 545, + 568 + ], + "type": "text", + "content": "[25] Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Soft then hard: Rethinking the quantization in neural image compression. In ICML, pages 3920-3929. PMLR, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 569, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 569, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 569, + 545, + 602 + ], + "type": "text", + "content": "[26] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. NeurIPS, 34:15908-15919, 2021. 1, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 602, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 647 + ], + "type": "text", + "content": "[27] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In CVPR, pages 5718-5727, 2022. 1, 2, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "text", + "content": "[28] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 1, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17245" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 286, + 95 + ], + "type": "text", + "content": "[30] Xiaodi Hou and Liqing Zhang. Saliency detection: A spectral residual approach. In CVPR, pages 1-8. IEEE, 2007. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 149 + ], + "type": "text", + "content": "[31] Yueyu Hu, Shuai Yang, Wenhan Yang, Ling-Yu Duan, and Jiaying Liu. Towards coding for human and machine vision: A scalable image coding approach. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 286, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 286, + 185 + ], + "type": "text", + "content": "[32] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. ECCV, 2022. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 286, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 286, + 229 + ], + "type": "text", + "content": "[33] Xin Jin, Ruoyu Feng, Simeng Sun, Runsen Feng, Tianyu He, and Zhibo Chen. Semantically video coding: Instill static-dynamic clues into structured bitstream for ai tasks. arXiv preprint arXiv:2201.10162, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 286, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 286, + 285 + ], + "type": "text", + "content": "[34] Nick Johnston, Damien Vincent, David Minnen, Michele Covell, Saurabh Singh, Troy Chinen, Sung Jin Hwang, Joel Shor, and George Toderici. Improved lossy image compression with priming and spatially adaptive bit rates for recurrent networks. In CVPR, pages 4385-4393, 2018. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 286, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 286, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 286, + 330 + ], + "type": "text", + "content": "[35] Aggelos K Katsaggelos, Lisimachos P Kondi, Fabian W Meier, Jorn Ostermann, and Guido M Schuster. Mpeg-4 and rate-distortion-based shape-coding techniques. Proceedings of the IEEE, 86(6):1126-1154, 1998. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 286, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 286, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 286, + 364 + ], + "type": "text", + "content": "[36] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, pages 9404-9413, 2019. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 365, + 286, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 286, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 286, + 397 + ], + "type": "text", + "content": "[37] Eastman Kodak. Kodak lossless true color image suite (photocd pcd0992). http://r0k.us/graphics/kodak, 1993.5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 286, + 453 + ], + "type": "text", + "content": "[38] Nam Le, Honglei Zhang, Francesco Cricri, Ramin GhaznaviYouvalari, and Esa Rahtu. Image coding for machines: An end-to-end learned approach. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1590-1594. IEEE, 2021. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 286, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 286, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 286, + 510 + ], + "type": "text", + "content": "[39] Nam Le, Honglei Zhang, Francesco Cricri, Ramin Ghaznavi Youvalari, Hamed Rezazadegan Tavakoli, and Esa Rahtu. Learned image coding for machines: A content-adaptive approach. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 511, + 286, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 286, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 286, + 542 + ], + "type": "text", + "content": "[40] Mu Li, Wangmeng Zuo, Shuhang Gu, Jane You, and David Zhang. Learning content-weighted deep image compression. TPAMI, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 286, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 286, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 286, + 587 + ], + "type": "text", + "content": "[41] Mu Li, Wangmeng Zuo, Shuhang Gu, Debin Zhao, and David Zhang. Learning convolutional networks for content-weighted image compression. In CVPR, pages 3214-3223, 2018. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 286, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 286, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 286, + 611 + ], + "type": "text", + "content": "[42] Xin Li, Jun Shi, and Zhibo Chen. Task-driven semantic coding via reinforcement learning. TIP, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 286, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 644 + ], + "type": "text", + "content": "[43] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. ECCV, 2022. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "text", + "content": "[44] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 712 + ], + "type": "text", + "content": "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "type": "text", + "content": "[46] Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In CVPR, pages 8759-8768, 2018. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "type": "text", + "content": "[47] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 174, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 174, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 174, + 545, + 205 + ], + "type": "text", + "content": "[48] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 1, 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "text", + "content": "[49] Siwei Ma, Xiang Zhang, Shiqi Wang, Xinfeng Zhang, Chuanmin Jia, and Shanshe Wang. Joint feature and texture coding: Toward smart video representation via front-end intelligence. TCSVT, 29(10):3095-3105, 2018. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "type": "text", + "content": "[50] Fabian Mentzer, Eirikur Agustsson, Michael Tschannen, Radu Timofte, and Luc Van Gool. Conditional probability models for deep image compression. In CVPR, pages 4394-4402, 2018. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "type": "text", + "content": "[51] Fabian Mentzer, George Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. arXiv preprint arXiv:2006.09965, 2020. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "text", + "content": "[52] David Minnen, Johannes Balle, and George Toderici. Joint autoregressive and hierarchical priors for learned image compression. In NeurIPS, 2018. 1, 2, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 361, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 403 + ], + "type": "text", + "content": "[53] David Minnen and Saurabh Singh. Channel-wise autoregressive entropy models for learned image compression. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3339-3343. IEEE, 2020. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 405, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 545, + 436 + ], + "type": "text", + "content": "[54] Majid Rabbani and Rajan Joshi. An overview of the JPEG 2000 still image compression standard. Signal processing: Image communication, 17(1):3-48, 2002. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "type": "text", + "content": "[55] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 472, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 492 + ], + "type": "text", + "content": "[56] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In CVPR, pages 7263-7271, 2017. 1, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 494, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 494, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 494, + 545, + 525 + ], + "type": "text", + "content": "[57] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28:91-99, 2015. 1, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 526, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 557 + ], + "type": "text", + "content": "[58] Oren Rippel and Lubomir Bourdev. Real-time adaptive image compression. In ICML, pages 2922-2930. PMLR, 2017. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 559, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 581 + ], + "type": "text", + "content": "[59] Thomas Sikora. Thempeg-4 video standard verification model. TCSVT, 7(1):19-31, 1997. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "text", + "content": "[60] Saurabh Singh, Sami Abu-El-Haija, Nick Johnston, Johannes Balle, Abhinav Shrivastava, and George Toderici. End-to-end learning of compressible features. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3349–3353. IEEE, 2020. 1, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 636, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 678 + ], + "type": "text", + "content": "[61] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. TCSVT, 22(12):1649-1668, 2012. 1, 2, 6" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[62] Simeng Sun, Tianyu He, and Zhibo Chen. Semantic structured image coding framework for multiple intelligent applications. TCSVT, 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "17246" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 556 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[63] Lucas Theis, Wenzhe Shi, Andrew Cunningham, and Ferenc Huszár. Lossy image compression with compressive autoencoders. arXiv preprint arXiv:1703.00395, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "text", + "content": "[65] Gregory K Wallace. The JPEG still picture compression standard. IEEE transactions on consumer electronics, 38(1):xviii-xxxiv, 1992. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "type": "text", + "content": "[66] Shurun Wang, Shiqi Wang, Wenhan Yang, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. Towards analysis-friendly face representation with scalable feature and texture compression. TMM, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 263 + ], + "type": "text", + "content": "[67] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. TCSVT, 13(7):560-576, 2003. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 265, + 286, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 265, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 265, + 286, + 297 + ], + "type": "text", + "content": "[68] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "type": "text", + "content": "[69] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. NeurIPS, 34:12077-12090, 2021. 1, 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 343, + 287, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 375 + ], + "type": "text", + "content": "[70] Yibo Yang, Robert Bamler, and Stephan Mandt. Improving inference for neural image compression. NeurIPS, 33:573-584, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 376, + 287, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 410 + ], + "type": "text", + "content": "[71] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, pages 2881-2890, 2017. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "type": "text", + "content": "[72] Rui Zhao, Wanli Ouyang, Hongsheng Li, and Xiaogang Wang. Saliency detection by multi-context deep learning. In CVPR, pages 1265-1274, 2015. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 445, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 499 + ], + "type": "text", + "content": "[73] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. 1, 3, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "type": "text", + "content": "[74] Lei Zhou, Zhenhong Sun, Xiangji Wu, and Junmin Wu. End-to-end optimized image compression with attention mechanism. In CVPR workshops, page 0, 2019. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 535, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 556 + ], + "type": "text", + "content": "[75] Yinhao Zhu, Yang Yang, and Taco Cohen. Transformer-based transform coding. In ICLR, 2022. 1, 2, 4, 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17247" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/0da0877f-f78d-4f97-9a49-55c50b2292de_content_list.json b/2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/0da0877f-f78d-4f97-9a49-55c50b2292de_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0961ee60362cee5079e13d60de019a2bb4742379 --- /dev/null +++ b/2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/0da0877f-f78d-4f97-9a49-55c50b2292de_content_list.json @@ -0,0 +1,1757 @@ +[ + { + "type": "text", + "text": "Semantics Meets Temporal Correspondence: Self-supervised Object-centric Learning in Videos", + "text_level": 1, + "bbox": [ + 179, + 130, + 789, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rui Qian1 Shuangrui Ding1 Xian Liu1 Dahua Lin1,2*", + "bbox": [ + 250, + 202, + 723, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Chinese University of Hong Kong 2Shanghai Artificial Intelligence Laboratory", + "bbox": [ + 143, + 220, + 825, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{qr021,ds023,lx021,dhlin}@ie.cuhk.edu.hk", + "bbox": [ + 307, + 241, + 661, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 312, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervised methods have shown remarkable progress in learning high-level semantics and low-level temporal correspondence. Building on these results, we take one step further and explore the possibility of integrating these two features to enhance object-centric representations. Our preliminary experiments indicate that query slot attention can extract different semantic components from the RGB feature map, while random sampling based slot attention can exploit temporal correspondence cues between frames to assist instance identification. Motivated by this, we propose a novel semantic-aware masked slot attention on top of the fused semantic features and correspondence maps. It comprises two slot attention stages with a set of shared learnable Gaussian distributions. In the first stage, we use the mean vectors as slot initialization to decompose potential semantics and generate semantic segmentation masks through iterative attention. In the second stage, for each semantics, we randomly sample slots from the corresponding Gaussian distribution and perform masked feature aggregation within the semantic area to exploit temporal correspondence patterns for instance identification. We adopt semantic- and instance-level temporal consistency as self-supervision to encourage temporally coherent object-centric representations. Our model effectively identifies multiple object instances with semantic structure, reaching promising results on unsupervised video object discovery. Furthermore, we achieve state-of-the-art performance on dense label propagation tasks, demonstrating the potential for object-centric analysis. The code is released at https://github.com/shvdiwnkozbw/SMTC.", + "bbox": [ + 76, + 323, + 473, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 820, + 209, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As one of the fundamental cognitive capabilities, human beings easily distinguish different objects, establish visual", + "bbox": [ + 76, + 845, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/06fe90c013fc77671267ef4b2e01a9a4c5c5ec86500384249bf7b59b4041c790.jpg", + "image_caption": [ + "(a) Query slot attention on semantic feature." + ], + "image_footnote": [], + "bbox": [ + 519, + 290, + 696, + 368 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/95713c39e96e743b2ba94bccd45808cf70187045eaa76c8e70f3830f7ba15b8e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 290, + 872, + 368 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f54196344cb58166b76125737f0d095a1fff030f384f1f1f8b67aba06fea8ca8.jpg", + "image_caption": [ + "(b) Random sampling slot attention on correspondence map." + ], + "image_footnote": [], + "bbox": [ + 519, + 396, + 696, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a9fc3ad0a33a1ce8c341f7a5026395f04b9f055999083bf9257420247028d229.jpg", + "image_caption": [ + "Figure 1. Fig. 1(a) presents the results of query slot attention on top of the RGB feature map. It successfully decomposes different semantics, e.g., camels and fence. Fig. 1(b) visualizes the correspondence map after PCA dimensionality reduction, showing that different instances have different correspondence patterns. And the slot attention with random sampling coarsely distinguishes two camels with some redundant borders. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 697, + 397, + 872, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "correspondence and perform object-centric analysis from temporally continuous observations. This ability can be attributed to two indispensable visual mechanisms: high-level semantic discrimination as well as low-level temporal correspondence, which enable humans to effectively understand and interact with the world.", + "bbox": [ + 496, + 625, + 893, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motivated by this, computer vision researchers equip machines with these capacities to enhance object-centric perception [79, 60, 34]. To achieve this goal, early works rely on human annotations or weak supervision to perceive object semantics [60, 63, 13], identify geometric positions [39, 71, 31, 91], and establish temporal correspondence [41, 79, 75, 77], but their generalization ability is limited. Recently, there emerge a host of fully unsupervised methods to learn robust representations for semantic discrimination [14, 38, 33, 70, 29, 37, 12] or spatiotemporal correspondence [44, 81, 86, 55, 52, 43], which achieve promising performance. Given this encouraging re", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author. Email: dhlin@ie.cuhk.edu.hk.", + "bbox": [ + 94, + 886, + 382, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16675", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sult, we naturally come up with a question: Is it possible to jointly leverage the semantics and correspondence to discover object instances and distill object-centric representations without human annotations?", + "bbox": [ + 80, + 90, + 465, + 148 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Regarding this problem, our intuition is that the high-level semantics delineates meaningful foreground areas in a top-down manner, while when looking into more frames, the low-level correspondence temporally associates coherent objects and separates individual instances in a bottom-up fashion. For instance, in a football scene, the semantic cue differentiates the foreground that includes several players, while the temporal correspondence links distinct players through dynamic movements and geometric relationships. These two aspects collectively contribute to object-centric representations. Unfortunately, most of the existing works only concentrate on one of these features. [14, 38, 33, 70, 82] succeed in developing high-level semantics, but this abstract semantics alone is insufficient to distinguish instances. Whereas, [44, 81, 78, 55] excel in detailed correspondence, but lack semantic structure and result in redundancy and ambiguity.", + "bbox": [ + 80, + 154, + 465, + 410 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a new architecture, Semantics Meets Temporal Correspondence (SMTC), to jointly leverage semantics and temporal correspondence to distill object-centric representations from RGB sequences. Specifically, we first extract frame-wise visual features as the semantic representation. Then we calculate dense feature correlations between adjacent frames as the correspondence map which encodes temporal relationships. To mine the object-centric knowledge, we take inspiration from [59, 88, 50, 25], and investigate using different formulations of slot attention on them. The preliminary experiments show that the original slot attention with random sampling on RGB feature map suffers from complex scene components in real-world videos [59, 72], but the revised query slot attention [88, 46] can decompose different semantic components as shown in Fig. 1(a). As for the correspondence map, different objects present diverse temporal correspondence patterns. Comparing to semantic features, these patterns reveal low-level geometric relationships, which are comparatively simple but vary with specific scenes. Hence, query slot attention fails but the random sampling based formulation performs surprisingly well as shown in Fig. 1(b), coarsely separating different object areas with some redundant borders.", + "bbox": [ + 80, + 415, + 465, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by this, we develop semantic-aware masked slot attention, which comprises a set of Gaussian distributions with learnable mean and standard deviation vectors, on top of the fused semantic and correspondence representations. The intuition is that the mean vectors could represent potential semantic centers, which act similarly to the query slot attention to separate semantic components. While the deviation vectors introduce perturbations around", + "bbox": [ + 80, + 780, + 465, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the semantic centers to capture distinct temporal correspondence patterns of different instances. Technically, we formulate two slot attention stages to achieve this goal. Firstly, we use the mean vectors as slot initialization to generate semantic segmentation masks. Secondly, for each semantics, we randomly sample slot vectors from the Gaussian distribution, then perform iterative attention and masked aggregation within the corresponding semantic mask area to distinguish instances. We enforce temporal consistency on the semantic masks as well as object instance slots to enhance temporal coherency and refine object-centric representations. Comparing with existing works on object-centric learning in videos [25, 50, 88, 84], our model is free of pre-computed motion or depth prior, and explicitly identifies multiple objects with semantic structure.", + "bbox": [ + 503, + 90, + 888, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are: (1) We propose a novel self-supervised architecture that unifies semantic discrimination and temporal correspondence to distill object-centric representations in videos. (2) We demonstrate that simple feature correlation can effectively represent temporal correspondence cues when used in conjunction with semantic features. Building on this observation, we develop semantic-aware masked slot attention, which operates on fused visual features and correspondence maps, to distinguish multiple object instances with semantic structure without relying on motion or depth priors. (3) We achieve promising results on unsupervised object discovery in both single and multiple object scenarios, and reach state-of-the-art performance on label propagation tasks, demonstrating that we learn discriminative and temporally consistent object-centric representations.", + "bbox": [ + 503, + 318, + 888, + 558 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 503, + 571, + 638, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised Object Discovery is an essential process to formulate object-centric representations, which aims to identify objects without human annotations. There exist a series of works focusing on this problem [59, 32, 26, 27, 28, 7]. Typically, [59] develops slot attention to iteratively update latent object representations from random initializations but it has difficulty scaling to complex real-world scenes. To tackle this challenge, [72] employs feature reconstruction as objective to reduce redundancy, [46] develops query slot initialization to encode visual concepts. Further, a line of works extend object discovery to video domain [50, 48, 17, 5, 88]. Most of them build on slot attention architecture, and adopt extra pre-computed priors, e.g., optical flow [88, 84, 16, 21], depth [25], geometric positions [50], as input or supervision to assist object discovery. In our work, we find that the simple feature correlation provides informative temporal cues. And we develop semantic-aware masked slot attention to identify object instances with semantic structure without resorting to extra priors.", + "bbox": [ + 503, + 599, + 888, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised Representation Learning aims to learn", + "bbox": [ + 503, + 886, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16676", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fe93decb9a5c6db629d51dcb51e2632af2e45a5db995b6fabaad9a7b39399dd8.jpg", + "image_caption": [ + "Figure 2. An overview of our framework. We first extract frame-wise features and calculate dense feature correlation, then fuse them to pass through semantic-aware masked slot attention, which comprises two slot attention stages with $N$ shared learnable Gaussian distributions. In the first semantic slot attention stage, the mean vectors serve as slot initialization to generate a set of segmentation masks for semantic decomposition. In the second masked slot attention stage, which runs on $N$ semantics in parallel, we randomly sample slots from the Gaussian distribution of each semantics and perform masked feature aggregation within the semantic area to identify distinct instances. We enforce semantic and instance temporal consistency to train the architecture in a teacher-student manner, with the teacher marked in gray." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 890, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "robust representation without human annotations. Early works design various pretext tasks to generate pseudo labels as self-supervision [49, 30, 47, 10, 1, 62]. Later, contrastive learning with instance discrimination revolutionized the field [38, 14, 70, 29, 35, 83, 65, 42, 11, 20, 3, 68, 35, 36, 69, 19, 22]. These works generally contribute to general semantics but lack the ability to distill object-centric representations. To delve into this problem, [82, 85] preserve spatial sensitivity and perform contrastive learning on dense features. [86, 12] point out that semantic pre-training implicitly encodes object correspondence. While in our work, we learn to explicitly identify different objects with instance-level consistency to progressively refine object-centric representations.", + "bbox": [ + 75, + 465, + 472, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-supervised Correspondence is a fundamental problem in computer vision [55, 78, 81, 80]. Typical works rely on low-level statistics like color to generate the dense correspondence flow [51, 52, 78], or establish temporal cycles to track image regions [81, 44, 6]. Further, [43, 86, 12] incorporate correspondence learning with semantic discrimination, but they either regard correspondence as a side effect of semantic discrimination [86, 12] or independently learn two pathway features with late fusion [43]. Among these works, high-level semantics is not well utilized, and there exists high redundancy due to large background areas. In contrast, our method explicitly separates semantic components to alleviate redundancy and formulate instance-level correspondence from an object-centric perspective.", + "bbox": [ + 75, + 688, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 464, + 589, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our framework is shown in Fig. 2, which adopts a teacher-student structure similar to recent self-supervised methods [38, 12, 33]. Specifically, we first extract framewise features from an RGB sequence, and calculate dense feature correlation as temporal correspondence representations. After that, we fuse them to pass through semantic-aware masked slot attention to separate multiple object instances with semantic structure, and enforce temporal consistency to refine the object-centric representations.", + "bbox": [ + 496, + 491, + 890, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Feature Encoding", + "text_level": 1, + "bbox": [ + 500, + 643, + 671, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a RGB sequence $v = \\{I_1, I_2, \\dots, I_T\\}$ , we employ a shared visual encoder $f$ to extract frame-wise features:", + "bbox": [ + 498, + 669, + 890, + 700 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {t} = f \\left(I _ {t}\\right) \\in \\mathbb {R} ^ {H W \\times D}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 729, + 890, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $H, W, D$ respectively denotes height, width and channel dimension, $t$ indicates frame index. For simplicity, we omit the subscript indicating student or teacher pathway, and use $\\hat{F}$ to denote teacher outputs. Intuitively, $F_{t}$ contains appearance and semantic cues in the input sequence, and the next step is to formulate an effective representation for the temporal correspondence cues. To do this, inspired by the cost volume in motion estimation [24, 73, 74], we calculate dense feature correlation to generate the temporal", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16677", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/27bc5aa89f1d02eb10a5f10b7873e1d35a06f0564f41870597c03af712daf58b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FeatureQueryRandom
RGB58.138.4
Correlation21.556.3
", + "bbox": [ + 156, + 88, + 390, + 148 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Preliminary results on RGB feature map and feature correlation. 'Query' ('Random') denotes slot attention with query (random sampling) initialization. We report IoU on DAVIS-2016.", + "bbox": [ + 76, + 148, + 468, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "correspondence map for each timestamp:", + "bbox": [ + 76, + 214, + 349, + 229 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {t j} = F _ {t} F _ {j} ^ {T} \\in \\mathbb {R} ^ {H W \\times H W}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 239, + 468, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we randomly sample one frame index $j \\neq t$ to compute feature correlation. The channel activations of $C_{tj}$ encode low-level geometric correlations between frame $t$ and $j$ , which indicates potential partitions of different objects as illustrated in Fig. 1(b).", + "bbox": [ + 76, + 268, + 468, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To jointly leverage the semantic and correspondence cues, we employ two linear transformation heads, $h_f$ and $h_c$ , to respectively project $F_t$ and $C_{tj}$ to a shared $D$ -dimensional embedding space, which are then fused with element-wise summation:", + "bbox": [ + 76, + 345, + 468, + 419 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {t} = h _ {f} \\left(F _ {t}\\right) + h _ {c} \\left(C _ {t j}\\right) \\in \\mathbb {R} ^ {H W \\times D}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 430, + 468, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this way, $R_{t}$ simultaneously encodes semantic and temporal correspondence information, serving as an intermediate feature for object-centric analysis.", + "bbox": [ + 76, + 460, + 468, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Preliminary Experiments with Slot Attention", + "text_level": 1, + "bbox": [ + 76, + 513, + 457, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given these representations, the next step is to formulate effective ways to identify individual objects. To do this, we take inspiration from slot attention [59, 50, 88, 46], which employs a set of slot vectors to iteratively attend to specific objects. Generally, there are two alternatives for slot initialization, random sampling from a Gaussian distribution with learnable parameters $\\mu, \\sigma \\in \\mathbb{R}^D$ [59, 50] or directly inheriting from a set of learnable queries [46, 88].", + "bbox": [ + 75, + 537, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To investigate the effectiveness of these two formulations, we conduct preliminary experiments on the semantic feature $F_{t}$ and correspondence map $C_{tj}$ respectively. We follow the settings in [72], using a frozen DINO pre-trained ViT to extract visual feature and generate correspondence map, on top of which we perform two kinds of slot attention. We train and evaluate them on DAVIS-2016 dataset [66]. As indicated by Fig. 1 and Table. 1, we find that query slot attention decomposes distinct semantic components on $F_{t}$ , while the original slot attention with random sampling performs much worse and struggles in complex real-world scenes. Conversely, query slot attention does not work on the correspondence map, and random sampling initialization enables the model to effectively utilize the temporal correspondence patterns in $C_{tj}$ to distinguish instances. An intuitive explanation into these observations is the RGB", + "bbox": [ + 75, + 659, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "features contain rich semantics and there exist consistent patterns for similar objects. Slot attention with learnable queries can capture inter-sample semantic patterns to identify objects. While random initialization fails to memorize semantic patterns, thus resulting in worse performance. In contrast, correspondence features reveal inter-frame geometric correlations, shown in Fig. 1(b), which vary with specific scenes. There are no consistent patterns for similar objects, so the learnable queries fail but random initialization works well to capture correspondence cues. The complementarity between these two conditions motivates two-stage slot attention design to combine the attributes of query and random sampling initialization, thus better exploiting the semantic and correspondence cues in $R_{t}$ .", + "bbox": [ + 496, + 90, + 890, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Semantic-aware Masked Slot Attention", + "text_level": 1, + "bbox": [ + 500, + 310, + 834, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose semantic-aware masked slot attention, which comprises $N$ Gaussian distributions with learnable means $\\pmb{\\mu} = \\{\\mu_1,\\mu_2,\\dots,\\mu_N\\} \\in \\mathbb{R}^{N\\times D}$ and standard deviations $\\sigma = \\{\\sigma_{1},\\sigma_{2},\\dots,\\sigma_{N}\\} \\in \\mathbb{R}^{N\\times D}$ . With this design, the mean vectors act similarly to the learnable queries, each representing a potential semantic center. And the deviations introduce perturbations around semantic centers to model temporal correspondence patterns. To achieve this goal, we employ two stages of iterative attention with shared learnable parameters. First, we use the mean vectors as slot initialization to decompose different semantics in the form of segmentation. Second, we randomly sample slots from each Gaussian distribution $(\\mu_n,\\mathrm{diag}(\\sigma_n))$ , and perform masked aggregation within the corresponding semantic area to identify instances. The process is detailed as follows.", + "bbox": [ + 496, + 333, + 890, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic Decomposition. In the first stage, we directly use $\\pmb{\\mu}$ without randomness as slot initialization $S = \\pmb{\\mu} \\in \\mathbb{R}^{N \\times D}$ . Then following the standard slot attention procedure [59], we employ three linear transformation heads to map $S$ and $R_{t}$ into $Q \\in \\mathbb{R}^{N \\times D}, K, V \\in \\mathbb{R}^{HW \\times D}$ , and iteratively calculate attention score and update slot representations. Mathematically, we formulate each iteration as:", + "bbox": [ + 496, + 560, + 890, + 666 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM := \\operatorname {s o f t m a x} \\left(\\frac {1}{\\sqrt {D}} Q K ^ {T}, 0\\right) \\in \\mathbb {R} ^ {N \\times H W}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 685, + 890, + 717 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA [ n, i ] := \\frac {M [ n , i ]}{\\sum_ {l} M [ n , l ] + \\epsilon}, A \\in \\mathbb {R} ^ {N \\times H W}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 718, + 890, + 753 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS := \\operatorname {G R U} (\\text {i n p u t} = A V, \\text {s t a t e s} = S). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 755, + 890, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that the slot attention weight $M$ is normalized along the query axis, and the weighted mean coefficient $A$ is employed to aggregate the value to update the slots. This mechanism induces competition among slots and enforce different slots to take over different semantic features. We iterate this process for three times, and take the slot attention weight (slot vectors) of the final iteration as the segmentation mask (representation) for different semantic compo", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16678", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "nents. To avoid ambiguity in presentation, we denote them as $\\mathcal{M} \\in \\mathbb{R}^{N \\times HW}$ and $S \\in \\mathbb{R}^{N \\times D}$ , where $S[n]$ presents $n$ -th semantic center representation, $\\mathcal{M}[n]$ indicates the probability that each pixel belongs to $n$ -th semantics.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instance Identification. Within each semantics, there could exist multiple object instances. The Gaussian distributions we introduce help to distinguish them with temporal correspondence cues. Specifically, in the second stage, for each semantics, we randomly sample slots from the Gaussian distribution, using the perturbations around the semantic center to capture distinct temporal correspondence patterns. For simplicity, we take the $n$ -th semantics as an example. First, we sample $P$ vectors, $S \\sim \\mathcal{N}(\\mu_n, \\mathrm{diag}(\\sigma_n)) \\in \\mathbb{R}^{P \\times D}$ , as slot initialization to represent $P$ potential instances. Next, we follow the first stage formulation to respectively project $S$ and $R_t$ into $Q \\in \\mathbb{R}^{P \\times D}$ and $K, V \\in \\mathbb{R}^{HW \\times D}$ , and obtain the attention score $M \\in \\mathbb{R}^{P \\times HW}$ . To preserve the semantic attributes and discriminate correspondence cues, we perform masked feature aggregation within the semantic mask areas. The only mathematical difference in this process is the weighted mean coefficient computation:", + "bbox": [ + 75, + 151, + 472, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} A [ p, i ] := \\frac {M [ p , i ] \\widetilde {\\mathcal {M}} [ n , i ]}{\\sum_ {l} M [ p , l ] \\widetilde {\\mathcal {M}} [ n , l ] + \\epsilon}, A \\in \\mathbb {R} ^ {P \\times H W}, \\tag {7} \\\\ \\widetilde {\\mathcal {M}} [ n, i ] = \\operatorname {b i n a r i z e} (\\mathcal {M} [ n, i ], \\tau), \\mathcal {M} \\in \\mathbb {R} ^ {N \\times H W}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 433, + 468, + 493 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we use the $n$ -th semantic mask $\\widetilde{\\mathcal{M}}[n] \\in \\{0,1\\}^{HW}$ binarized with threshold $\\tau$ to constrain the model to only aggregate the visual features related to $n$ -th semantics. And we empirically find a comparatively large threshold $\\tau = 0.5$ helps to preserve object regions with clear semantics and filter out background. The masked slot attention for all semantics follows the same rule and is calculated in parallel. Similarly, we iterate the process for three times, use the final slot attention weight as the instance segmentation masks, and take the final slot vectors as the object-centric representations $\\mathcal{O} \\in \\mathbb{R}^{N \\times P \\times D}$ , where $\\mathcal{O}[n,p] \\in \\mathbb{R}^D$ denotes the $p$ -th potential object instance of $n$ -th semantics. In this way, our semantic-aware masked slot attention explicitly segments multiple instances with semantic structure.", + "bbox": [ + 75, + 503, + 472, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training", + "text_level": 1, + "bbox": [ + 76, + 726, + 181, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use temporal consistency as self-supervision to optimize semantic masks $\\mathcal{M}$ and instance representations $\\mathcal{O}$ .", + "bbox": [ + 75, + 750, + 468, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dense Semantic Alignment. Given the semantic mask of each timestamp $\\mathcal{M}_t\\in \\mathbb{R}^{N\\times HW}$ , as well as $\\hat{\\mathcal{M}}_t$ from teacher pathway, we aim to align the spatially dense semantic distributions across time. To achieve this, the first step is to determine the corresponding patches between different frames. Utilizing the feature correlation $C_{tj}\\in \\mathbb{R}^{HW\\times HW}$ which indicates dense feature similarity between timestamps $t,j$ , we can infer patch-level correspondence from", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "this cue with optimal transport [11, 2, 58, 18]. Formally, we take $-\\hat{C}_{tj}$ from teacher pathway as cost matrix, and solve the optimal transport strategy $\\pi_{tj}^{*} \\in \\mathbb{R}^{HW \\times HW}$ between two uniform distributions to indicate patch correspondence:", + "bbox": [ + 496, + 90, + 892, + 152 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\pi_ {t j}} \\sum_ {u = 1} ^ {H W} \\sum_ {v = 1} ^ {H W} - \\hat {C} _ {t j} [ u, v ] \\pi_ {t j} [ u, v ] \\\\ \\text {s . t .} \\sum_ {v = 1} ^ {H W} \\pi_ {t j} [ \\cdot , v ] = \\frac {1}{H W} \\mathbf {1} ^ {H W}, \\tag {8} \\\\ \\sum_ {u = 1} ^ {H W} \\pi_ {t j} [ u, \\cdot ] = \\frac {1}{H W} \\mathbf {1} ^ {H W}, \\\\ \\pi_ {t j} [ u, v ] \\geq 0 \\quad u, v \\in \\{1, 2, \\dots , H W \\}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 172, + 890, + 323 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $u, v$ denotes spatial index, $\\mathbf{1}^{HW}$ is a $HW$ -dimensional vector of all ones. Note that it is feasible to adopt other marginal distribution formulations with prior knowledge, e.g., class-agnostic activation map [4], to facilitate training, with details discussed in Supplementary Material. We employ Sinkhorn-Knopp algorithm [18] to obtain the optimal transport matrix $\\pi_{t,j}^*$ , and formulate the dense semantic alignment objective in the form of cross entropy:", + "bbox": [ + 496, + 332, + 892, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s e m} = - \\sum_ {\\substack {t, j = 1 \\\\ t \\neq j}} ^ {T} \\sum_ {u, v = 1} ^ {H W} \\sum_ {n = 1} ^ {N} \\pi_ {t j} ^ {*} [ u, v ] \\hat {\\mathcal {M}} _ {j} [ n, v ] \\log \\mathcal {M} _ {t} [ n, u ]. \\tag{9}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 476, + 890, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By minimizing the weighted sum of cross entropy, we achieve temporally consistent semantic distributions among corresponding spatial areas.", + "bbox": [ + 496, + 550, + 890, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic Mask Regularization. To encourage the semantic centers to emphasize different visual contents and cover diverse semantics, we apply a simple regularization to $\\mathcal{M}_t$ :", + "bbox": [ + 496, + 595, + 890, + 642 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e g} = \\sum_ {t = 1} ^ {T} \\sum_ {\\substack {n, j = 1 \\\\ j \\neq n}} ^ {N} \\frac {\\mathcal {M} _ {t} [ n ] ^ {T} \\mathcal {M} _ {t} [ j ]}{\\| \\mathcal {M} _ {t} [ n ] \\| _ {2} \\| \\mathcal {M} _ {t} [ j ] \\| _ {2}}. \\tag{10}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 662, + 890, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This simple regularization suppresses the cosine similarity between different semantic masks to avoid collapse.", + "bbox": [ + 496, + 719, + 890, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instance Representation Consistency. Apart from the semantic distributions, it is necessary to ensure that the representations of each object instance are temporally consistent. Given the instance representations $\\mathcal{O}_t\\in \\mathbb{R}^{N\\times P\\times HW}$ , as well as $\\hat{\\mathcal{O}}_t$ from teacher pathway, we need to first match corresponding instances between different timestamps. For illustration, considering $n$ -th semantics of time $t,j$ , we adopt bipartite matching [9, 15] based on the cosine similarity between $\\mathcal{O}_t[n]$ and $\\hat{\\mathcal{O}}_j[n]$ to generate one-to-one instance correspondence, with the matching function denoted as $\\varepsilon (\\cdot)$ .", + "bbox": [ + 496, + 750, + 892, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16679", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, $\\mathcal{O}_t[n,p]$ and $\\hat{\\mathcal{O}}_j[n,\\varepsilon(p)]$ are considered as a positive pair to be aligned. However, for each video, there exists absent semantics and object occlusion such that the number of visible object instances varies. Thus, there could be redundant slots not attending to objects, and it is crucial to filter out these invalid instance representations to reduce distractions. Mathematically, for timestamp $t$ , we introduce $\\mathcal{I}_t \\in \\{0,1\\}^{N \\times P}$ as the valid instance indicator. For the $p$ -th instance of $n$ -th semantics to be valid, two criteria must be met: (1) The ratio of the $n$ -th semantic area is above threshold $\\tau_1 = 0.2$ to filter out non-existing semantics; (2) The instance representation is close to the semantic center representation with cosine similarity larger than $\\tau_2 = 0.5$ to exclude redundant slots. The conditions are formulated as:", + "bbox": [ + 75, + 89, + 472, + 303 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {t} [ n, p ] = 1 \\Leftrightarrow \\left\\{ \\begin{array}{l} \\frac {1}{H W} \\sum_ {u = 1} ^ {H W} \\widetilde {\\mathcal {M}} _ {t} [ n, u ] \\geq \\tau_ {1}, \\\\ \\cos \\left(\\mathcal {S} _ {t} [ n ], \\mathcal {O} _ {t} [ n, p ]\\right) \\geq \\tau_ {2}. \\end{array} \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 320, + 468, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "And $\\hat{\\mathcal{L}}_j\\in \\{0,1\\}^{N\\times P}$ is defined in the same manner. We use a margin loss to encourage object detail consistency over valid instance representations:", + "bbox": [ + 76, + 390, + 470, + 436 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal{L}_{obj} = \\sum_{\\substack{t,j = 1\\\\ t\\neq j}}^{T}\\sum_{n = 1}^{N}\\sum_{p = 1}^{P}\\mathcal{I}_{t}[n,p]\\big\\{ \\\\ \\hat {\\mathcal {I}} _ {j} [ n, \\varepsilon (p) ] \\| \\mathcal {O} _ {t} [ n, p ] - \\hat {\\mathcal {O}} _ {j} [ n, \\varepsilon (p) ] \\| _ {2} \\tag {12} \\\\ + \\sum_{\\substack{q = 1\\\\ q\\neq \\varepsilon (p)}}^{P}\\mathrm{relu}(\\lambda -\\| \\mathcal{O}_{t}[n,p] - \\hat{\\mathcal{O}}_{j}[n,q]\\|_{2})\\} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 443, + 468, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ is a margin hyper-parameter. Note that with the $\\mathcal{I}_t$ and $\\hat{\\mathcal{I}}_j$ constraints, our formulation can handle object occlusions which lead to varying number of visible instances in different frames. By minimizing the margin loss, we encourage the model to distill discriminative and temporally consistent object-centric representations.", + "bbox": [ + 75, + 569, + 468, + 659 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall, we take the summation of three objectives for training:", + "bbox": [ + 75, + 660, + 468, + 690 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {s e m}} + \\mathcal {L} _ {\\text {o b j}} + \\mathcal {L} _ {\\text {r e g}}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 696, + 468, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The student pathway $\\theta$ is updated with gradient descent, and the teacher parameters $\\hat{\\theta}$ are updated with momentum as:", + "bbox": [ + 76, + 719, + 468, + 750 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\theta} \\leftarrow m \\hat {\\theta} + (1 - m) \\theta , \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 768, + 468, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $m$ is momentum coefficient set to 0.999 in default. This momentum update mechanism results in a slowly evolving teacher network, which progressively distills object knowledge, provides reliable self-supervision signals and enables us to train semantic-aware masked slot attention without relying on widely adopted reconstruction objective in existing works [59, 50, 88, 84, 72, 46].", + "bbox": [ + 75, + 794, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 89, + 633, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 500, + 114, + 594, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train our model on YouTube-VOS [87], a challenging video dataset that contains multiple object instances of distinct semantics in each video. We evaluate our method on two lines of tasks: (1) Unsupervised object segmentation on DAVIS-2016 [66], SegTrack-v2 [54], FMBS-59 [64] and challenging multiple object segmentation on DAVIS-2017-Unsupervised [8]. We respectively calculate the mean per frame intersection over union (IoU) and $\\mathcal{J}\\&\\mathcal{F}$ on single and multiple object discovery benchmarks. (2) Lable propagation tasks including semi-supervised video object segmentation on DAVIS-2017 [67], human pose tracking on JHMDB [45], human part tracking on VIP [90]. We adopt the same settings as [44, 43] and report standard $\\mathcal{J}$ and $\\mathcal{F}$ score on DAVIS, probability of a correct pose (PCK) on JHMDB and mean intersection over union (mIoU) on VIP.", + "bbox": [ + 496, + 137, + 893, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 375, + 718, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We sample $T = 4$ frames with frame rate FPS $= 4$ as the input RGB sequence, where each frame is augmented with random crop, horizontal flip and color jitter, and finally resized to $256 \\times 256$ . We adopt ViT-Small/16 [23] and ResNet-50 [40] as the visual encoder, specified in each experiment, to extract frame-wise features. Then for semantic-aware masked slot attention, we set the number of learnable Gaussian distributions to $N = 16$ , the number of instances of each semantics to $P = 4$ in default and use 3 iterations to update slot representations and attention maps.", + "bbox": [ + 496, + 398, + 892, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In training, the visual encoder is initialized with self-supervised pre-trained weights from DINO ViT-Small/16 [12] or MoCo ResNet-50 [38]. We use AdamW optimizer [61] with learning rate $2 \\times 10^{-4}$ for batch size 128 to update student parameters, and the teacher pathway is updated in momentum. In inference, we employ the same evaluation protocol as [44] to evaluate the temporal object correspondence performance on label propagation tasks. And for object discovery, we maintain the valid instances filtered with Eq. 11 as candidate objects. For single object evaluation, we merge all candidate objects as foreground areas to calculate IoU. For multiple object benchmark, we follow [8] to match ground-truth and our predictions and report $\\mathcal{J} \\& \\mathcal{F}$ score.", + "bbox": [ + 496, + 550, + 890, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-art", + "text_level": 1, + "bbox": [ + 498, + 771, + 794, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Single Object Discovery. To measure our model's ability to decompose different objects, we first present the quantitative results on single object discovery without post processing (e.g., CRF, spectral clustering) in Table 2. Many existing works resort to optical flow [89, 53, 88, 84] or synthetic data [84] as weak supervision to learn temporal dynamics and generate semantic-agnostic segmentation masks for", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16680", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/74915c80b8afafea9848b2410f8c537e12d5a07caee08cd07f2c42f0f5b7a3e5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelRGBFlowDAVISST-v2FBMS
CIS [89]✓✓71.562.563.5
AMD [57]✓X57.857.047.5
DINO [12]✓X52.346.550.3
SIMO [53]X✓67.862.0-
MG [88]X✓68.358.653.1
OCLR [84]X✓72.167.665.4
GWM [16]✓✓71.269.066.9
SMTC✓X71.869.368.4
\\( SMTC^† \\)✓X70.868.466.5
", + "bbox": [ + 78, + 88, + 470, + 258 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/243e8f7e53a13f3659749e48c6f79e8126816d114352e82806decb9861caa6fb.jpg", + "table_caption": [ + "Table 2. Quantitative results on single object discovery. We compare per frame mean IoU on DAVIS-2016, SegTrack-v2 and FBMS-59 without any post-processing. $\\mathsf{SMTC}^{\\dagger}$ denotes only with first slot attention stage for semantic decomposition in inference." + ], + "table_footnote": [], + "table_body": "
ModelBackboneJ&FJF
DINOSAUR [72]ViT-S/1621.419.223.7
SMTCViT-S/1640.536.444.6
SMTCResNet-5039.035.542.6
RVOS [76]ResNet-10141.236.845.7
ProReduce [56]ResNet-10168.365.071.6
", + "bbox": [ + 78, + 327, + 470, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Quantitative results on multiple object discovery on DAVIS-2017-Unsupervised. Gray denotes supervised training.", + "bbox": [ + 76, + 439, + 468, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "moving objects. DINO baseline [12] uses the self-attention map from the last layer as an object segmentation mask. While our method only relies on RGB frames and explicitly discriminates object semantics in a fully self-supervised manner. From the comparison, we observe that our SMTC with only semantic decomposition largely exceeds RGB-only methods [12, 57] and comparable to optical flow-based methods, and our formulation with instance identification further improves the performance. This phenomenon indicates that the learned semantic decomposition with temporal correspondence cues have the potential to substitute precomputed optical flow to guide single object segmentation. Multiple Object Discovery. Most of the existing video segmentation methods with fully unsupervised training cannot explicitly distinguish multiple object instances. OCLR [84] adopts layered flow representations to identify multiple objects on re-annotated DAVIS-2017-motion dataset, but it cannot distinguish instances with common motion. On the contrary, our method jointly utilizes semantics and temporal correspondence to separate arbitrary object instances on DAVIS-2017-Unsupervised benchmark. For fair comparison, we re-run DINOSAUR [72] on DAVIS-2017 with 5 slots as an unsupervised baseline. As presented in Table 3, our method significantly outperforms DINOSAUR. This is because our method explicitly decomposes different semantics, exploits temporal correspondence among multiple frames to distinguish instances and encourages instance", + "bbox": [ + 73, + 492, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "level temporal consistency. While DINOSAUR is only supervised with DINO feature reconstruction without considering temporal relationships and semantic discrimination. Besides, our method is comparable with supervised baseline RVOS [76], and the gap to supervised state-of-the-art [56] is majorly due to human annotated masks as supervision leading to much more precise segmentation boundaries. This conjecture can be demonstrated by Fig. 3, which also reveals that our model is able to learn discriminative object-centric representations to separate different instances.", + "bbox": [ + 496, + 90, + 890, + 241 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Label Propagation. Finally, we compare the performance on label propagation tasks in Table 4 to validate the temporal consistency of our learned features. Though not specifically designed for these tasks, our model achieves state-of-the-art results on three benchmarks. The most related work to our SMTC is SFC [43], which also utilizes both high-level semantic and low-level correspondence. The difference is that SFC only employs late fusion to incorporate these two cues in inference, while our model jointly exploits semantics and temporal correspondence in training to identify object instances, perform instance-level alignment and refine temporally consistent object-centric representations.", + "bbox": [ + 496, + 243, + 890, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 436, + 653, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report IoU on DAVIS-2016, and $\\mathcal{J}\\&\\mathcal{F}$ on DAVIS-2017-Unsupervised for single and multiple object discovery. Refer to Supplementary Material for more ablations.", + "bbox": [ + 496, + 460, + 890, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Slot Attention Formulation. We compare the different formulations of semantic-aware masked slot attention in Table 5. With only the first slot attention stage, i.e., the 'Semantic' part, the model cannot distinguish object instances of the same semantics, thus the performance on multiple object discovery drops significantly as indicated by Ours-B. While with only the second stage, i.e., the 'Instance' part, the performance on both single and multiple object segmentation drops as indicated by Ours-C. This reveals that the pre-computed semantic masks play an important role in alleviating interference in instance slot update and filtering valid sample for instance-level alignment.", + "bbox": [ + 496, + 506, + 890, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of Learnable Gaussian Distributions. Comparing Ours-A, Ours-E and Ours-F in Table 5, we observe that the performance improves when $N$ increases since larger $N$ guides our model to distill more fine-grained semantics. And we present an extreme setting with $N = 1$ in Ours-D, where our formulation degenerates into the original slot attention with random sampling from one Gaussian distribution [59]. Note that although Ours-C also has no access to semantic masks, it still maintains the potential to discriminate different semantics due to multiple learnable means. Its superiority to Ours-D demonstrates the effectiveness of semantic discrimination in object discovery.", + "bbox": [ + 496, + 688, + 890, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Feature Usage. We explore using different feature input to our semantic-aware masked slot attention in Table 6. We", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16681", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2e991ca17115f8e6d9d1a5316ff1385b34d19b4ff7406228e0854913bf4017c1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelBackboneDAVISJHMDBVIP
J&FJFPCK@0.1PCK@0.2mIoU
Supervised [40]ResNet-5066.063.768.459.278.339.5
MoCo [38]ResNet-5065.463.267.660.479.336.1
VFS [86]ResNet-5068.966.571.360.980.743.2
DINO [12]ViT-S/1661.860.263.445.475.237.9
TimeCycle [81]ResNet-5040.741.939.457.778.528.9
UVC [55]ResNet-5056.354.558.156.576.634.2
MAST [51]ResNet-1865.563.367.6---
CRW [44]ResNet-1867.664.870.258.880.337.6
SFC [43]ResNet-18+ResNet-5071.268.374.061.983.038.4
SMTCViT-S/1667.664.171.253.279.639.2
SMTCResNet-5073.069.476.662.584.138.8
", + "bbox": [ + 143, + 88, + 826, + 316 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/22c366b2462887e8ff4e437ad3aa5886c4e1a7fb3cd60b8dd7c41fccaa81a877.jpg", + "table_caption": [ + "Table 4. Quantitative results on label propagation tasks: semi-supervised video object segmentation on DAVIS-2017, pose tracking on JHMDB, human part tracking on VIP. We present the backbone for comprehensive comparison, and report comparing results from [86, 43]." + ], + "table_footnote": [ + "Table 5. Ablation studies on the number of learnable Gaussian distributions, and the formulation of semantic-aware masked slot attention. 'Semantic' ('Instance') denotes the first (second) attention stage for semantic decomposition (instance identification)." + ], + "table_body": "
ModelNSemanticMaskedIoUJ&F
Ours-A16✓✓71.840.5
Ours-B16✓X66.424.5
Ours-C16X✓52.130.3
Ours-D1-✓47.422.1
Ours-E8✓✓68.437.9
Ours-F32✓✓71.940.7
", + "bbox": [ + 89, + 367, + 455, + 500 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/638ed2c7d3ac7867504703ab9de585d5f0d4d604b119a2c18184f98b1906a9fb.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6. Ablation studies on the feature usage. We compare different slot attention formulations on different feature inputs." + ], + "table_body": "
FeatureSemanticInstanceIoUJ&F
Fused✓✓71.840.5
RGB-only✓✓67.720.1
✓×67.518.8
×✓41.515.3
Correlation-only✓✓38.814.2
✓×21.110.7
×✓59.427.5
", + "bbox": [ + 89, + 573, + 455, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "observe that with RGB-only input, the 'Semantic' part guarantees satisfactory performance on single object discovery, but performs poor on multiple object benchmarks even with the 'Instance' part. This coincides with our intuition that it is difficult to distinguish object instances with solely semantic cues, instead we need to look into more frames and resort to temporal correspondence. While for the correlation-only input, the 'Instance' part coarsely segments objects and", + "bbox": [ + 75, + 779, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/58193a7cbbe98de680f926172e4fb7141545481b1e678e4b592cbe2510928ba5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LsemLregLobjIoUJ&F
✓XX61.321.6
✓✓X66.424.5
XX✓50.630.4
✓✓✓71.740.5
", + "bbox": [ + 562, + 367, + 828, + 445 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/01df299f42619f1a2af5bada16f016a24e292ae388de20451732167ae9c373b3.jpg", + "table_caption": [ + "Table 7. Ablation studies on the learning objectives. We report the results on DAVIS-2016 and DAVIS-2017-Unsupervised." + ], + "table_footnote": [ + "Table 8. Ablation studies on the hyper-parameters of frame sampling. We compare different number of frames per clip and different FPS." + ], + "table_body": "
TFPSIoUJ&F
2465.436.9
4266.337.7
4471.840.5
8471.940.8
", + "bbox": [ + 602, + 484, + 790, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the 'Semantic' part exerts negative impact. This is because the correspondence map contains little semantic information, the semantic decomposition results in severe ambiguity, thus impairing the performance.", + "bbox": [ + 496, + 627, + 890, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learning Objectives. We also present a comprehensive ablation on the learning objectives in Table 7, where we report single object discovery on DAVIS-2016 and multiple object discovery results on DAVIS-2017-Unsupervised. From the comparison, we observe that $\\mathcal{L}_{sem}$ is fundamental to decompose object semantics, which demonstrates the necessity of semantic decomposition in the first stage. $\\mathcal{L}_{reg}$ also leads to improvement by encouraging the learnable queries to cover more diverse semantics. As for $\\mathcal{L}_{obj}$ , this objective introduces instance-level alignment and significantly enhances multiple object discovery performance on DAVIS-2017-Unsupervised.", + "bbox": [ + 496, + 689, + 890, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Frame Sampling. Another important thing in training is the frame sampling procedure. The number of frames per", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16682", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/321920c59f87079ae3354195f90e07dba2a40c7ad71b4348a4a4c2136517e82f.jpg", + "image_caption": [ + "Figure 3. Visualization of semantic and instance segmentation map. The red boxes outline the ambiguous areas." + ], + "image_footnote": [], + "bbox": [ + 84, + 90, + 468, + 170 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5c852b66a77f8f4dba7488c70f5cf5273325a7f16a637a4829c96daa9cffb1c0.jpg", + "image_caption": [ + "Figure 4. Visualization of instance alignment. The arrows point out the matched instances across time." + ], + "image_footnote": [], + "bbox": [ + 78, + 213, + 468, + 271 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "clip and sampling frame rate determines the temporal reception field of our model. We compare different sampling hyper-parameters in Table 8. The results indicate that it is necessary to use large FPS to provide rich temporal dynamics. And by setting comparatively large FPS, our model can have access to more temporal information without introducing more frames, reaching a trade-off between performance and efficiency.", + "bbox": [ + 75, + 324, + 467, + 445 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.5. Further Discussion", + "text_level": 1, + "bbox": [ + 76, + 454, + 261, + 469 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Visualization of Object Discovery. We visualize the our semantic decomposition map and instance identification map in Fig. 3, where the same color denotes the segmentation mask of the same semantics. The visualization reveals that our method is able to distinguish multiple object instances with semantic structure. For example, different people belong to the same semantic center, and quadrupeds such as dogs and pigs belong to another semantics. And our model also separates different instances of the same semantics, such as multiple persons and different pigs, with minor ambiguity on small object part, e.g., pig legs, mobile phones. Besides, we also visualize our instance-level alignment during training in Fig. 4. Our bipartite matching well aligns the corresponding instances in different frames, and the valid sample constraint of Eq. 11 effectively handles object occlusion in videos. For example, when encountering an occluded person instance, our method filters out the redundant slots not attending to objects, only correlating valid slots for robust instance-level alignment.", + "bbox": [ + 75, + 477, + 467, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitation. Due to the lack of the human annotated segmentation masks, our model faces challenges in generating precise boundaries for each instance, particularly for small objects. One potential solution is to incorporate multi-scale feature pyramid to further improve dense perception. We leave it in the future work. Despite this limitation, our work effectively demonstrates the benefits of leveraging both semantics and temporal correspondence to discover object instances with semantic structure and distill discriminative", + "bbox": [ + 75, + 765, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "and temporally consistent object-centric representations.", + "bbox": [ + 500, + 90, + 870, + 106 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 119, + 617, + 135 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we propose a novel self-supervised framework jointly exploiting high-level semantics and low-level temporal correspondence to enhance object-centric perception. Specifically, we represent semantic and temporal correspondence cues using the RGB feature map and dense feature correlation, respectively. These cues are fused and fed into semantic-aware masked slot attention which comprises a set of learnable Gaussian distributions. This design allows us to leverage the mean vectors as potential semantic centers for semantic decomposition, and use the perturbations introduced by the standard deviation vectors around the semantic centers to make use of the temporal correspondence cues for instance identification. To distill discriminative and temporally consistent object-centric representations, we devise semantic- and instance-level alignment that is robust to object occlusion as self-supervision. We demonstrate the effectiveness of our model for object-centric analysis through the state-of-the-art performance on label propagation tasks, as well as the promising results on unsupervised object discovery in both single and multiple object scenarios.", + "bbox": [ + 496, + 145, + 890, + 448 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 462, + 660, + 479 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank Tianfan Xue for constructive feedback. This project is funded in part by Shanghai AI Laboratory (P23KS00020, 2022ZD0160201), CUHK Interdisciplinary AI Research Institute, and the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK.", + "bbox": [ + 496, + 487, + 890, + 579 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 592, + 596, + 607 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. Advances in Neural Information Processing Systems, 33:9758-9770, 2020.", + "[2] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. arXiv preprint arXiv:1911.05371, 2019.", + "[3] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019.", + "[4] Kyungjune Baek, Minhyun Lee, and Hyunjung Shim. Psynet: Self-supervised approach to object localization using point symmetric transformation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10451-10459, 2020.", + "[5] Beril Besbinar and Pascal Frossard. Self-supervision by prediction for object discovery in videos. In 2021 IEEE Interna" + ], + "bbox": [ + 506, + 617, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16683", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tional Conference on Image Processing (ICIP), pages 1509-1513. IEEE, 2021.", + "[6] Zhangxing Bian, Allan Jabri, Alexei A Efros, and Andrew Owens. Learning pixel trajectories with multiscale contrastive random walks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6508-6519, 2022.", + "[7] Christopher P Burgess, Loic Matthey, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matt Botvinick, and Alexander Lerchner. Monet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390, 2019.", + "[8] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv preprint arXiv:1905.00737, 2019.", + "[9] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020.", + "[10] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018.", + "[11] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020.", + "[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021.", + "[13] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017.", + "[14] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020.", + "[15] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021.", + "[16] Subhabrata Choudhury, Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Guess what moves: unsupervised video and image segmentation by anticipating motion. arXiv preprint arXiv:2205.07844, 2022.", + "[17] Eric Crawford and Joelle Pineau. Exploiting spatial invariance for scalable unsupervised object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 3684-3692, 2020.", + "[18] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013." + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9716-9726, 2022.", + "[20] Shuangrui Ding, Rui Qian, and Hongkai Xiong. Dual contrastive learning for spatio-temporal representation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5649-5658, 2022.", + "[21] Shuangrui Ding, Weidi Xie, Yabo Chen, Rui Qian, Xiaopeng Zhang, Hongkai Xiong, and Qi Tian. Motion-inductive self-supervised object discovery in videos. arXiv preprint arXiv:2210.00221, 2022.", + "[22] Shuangrui Ding, Peisen Zhao, Xiaopeng Zhang, Rui Qian, Hongkai Xiong, and Qi Tian. Prune spatio-temporal tokens by semantic-aware temporal accumulation. arXiv preprint arXiv:2308.04549, 2023.", + "[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.", + "[24] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015.", + "[25] Gamaleldin F Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael C Mozer, and Thomas Kipf. Savi++: Towards end-to-end object-centric learning from real-world videos. arXiv preprint arXiv:2206.07764, 2022.", + "[26] Patrick Emami, Pan He, Sanjay Ranka, and Anand Rangarajan. Efficient iterative amortized inference for learning symmetric and disentangled multi-object representations. In International Conference on Machine Learning, pages 2970-2981. PMLR, 2021.", + "[27] Martin Engelcke, Adam R Kosiorek, Oiwi Parker Jones, and Ingmar Posner. Genesis: Generative scene inference and sampling with object-centric latent representations. arXiv preprint arXiv:1907.13052, 2019.", + "[28] Martin Engelcke, Oiwi Parker Jones, and Ingmar Posner. Genesis-v2: Inferring unordered object representations without iterative refinement. Advances in Neural Information Processing Systems, 34:8085-8094, 2021.", + "[29] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3299-3309, 2021.", + "[30] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16684", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015.", + "[32] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthey, Matthew Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative variational inference. In International Conference on Machine Learning, pages 2424-2433. PMLR, 2019.", + "[33] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020.", + "[34] Kai Han, Rafael S Rezende, Bumsub Ham, Kwan-Yee K Wong, Minsu Cho, Cordelia Schmid, and Jean Ponce. Sc net: Learning semantic correspondence. In Proceedings of the IEEE international conference on computer vision, pages 1831-1840, 2017.", + "[35] Tengda Han, Weidi Xie, and Andrew Zisserman. Video representation learning by dense predictive coding. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019.", + "[36] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. Advances in Neural Information Processing Systems, 33:5679-5690, 2020.", + "[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022.", + "[38] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020.", + "[39] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017.", + "[40] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.", + "[41] David Held, Sebastian Thrun, and Silvio Savarese. Learning to track at 100 fps with deep regression networks. In European conference on computer vision, pages 749-765. Springer, 2016.", + "[42] Olivier Henaff. Data-efficient image recognition with contrastive predictive coding. In International conference on machine learning, pages 4182-4192. PMLR, 2020.", + "[43] Yingdong Hu, Renhao Wang, Kaifeng Zhang, and Yang Gao. Semantic-aware fine-grained correspondence. arXiv preprint arXiv:2207.10456, 2022.", + "[44] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "in neural information processing systems, 33:19545-19560, 2020.", + "[45] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013.", + "[46] Baoxiong Jia, Yu Liu, and Siyuan Huang. Unsupervised object-centric learning with bi-level optimized query slot attention. arXiv preprint arXiv:2210.08990, 2022.", + "[47] Longlong Jing, Xiaodong Yang, Jingen Liu, and Yingli Tian. Self-supervised spatiotemporal feature learning via video rotation prediction. arXiv preprint arXiv:1811.11387, 2018.", + "[48] Rishabh Kabra, Daniel Zoran, Goker Erdogan, Loic Matthew, Antonia Creswell, Matt Botvinick, Alexander Lerchner, and Chris Burgess. Simone: View-invariant, temporally-abstracted object representations via unsupervised video decomposition. Advances in Neural Information Processing Systems, 34:20146-20159, 2021.", + "[49] Dahun Kim, Donghyeon Cho, Donggeun Yoo, and In So Kweon. Learning image representations by completing damaged jigsaw puzzles. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 793-802. IEEE, 2018.", + "[50] Thomas Kipf, Gamaleldin F Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional object-centric learning from video. arXiv preprint arXiv:2111.12594, 2021.", + "[51] Zihang Lai, Erika Lu, and Weidi Xie. Mast: A memory-augmented self-supervised tracker. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6479-6488, 2020.", + "[52] Zihang Lai and Weidi Xie. Self-supervised learning for video correspondence flow. arXiv preprint arXiv:1905.00875, 2019.", + "[53] Hala Lamdouar, Weidi Xie, and Andrew Zisserman. Segmenting invisible moving objects. 2021.", + "[54] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In Proceedings of the IEEE international conference on computer vision, pages 2192-2199, 2013.", + "[55] Xueting Li, Sifei Liu, Shalini De Mello, Xiaolong Wang, Jan Kautz, and Ming-Hsuan Yang. Joint-task self-supervised learning for temporal correspondence. Advances in Neural Information Processing Systems, 32, 2019.", + "[56] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1739-1748, 2021.", + "[57] Runtao Liu, Zhirong Wu, Stella Yu, and Stephen Lin. The emergence of objectness: Learning zero-shot segmentation from videos. Advances in Neural Information Processing Systems, 34:13137-13152, 2021.", + "[58] Songtao Liu, Zeming Li, and Jian Sun. Self-emd: Self-supervised object detection withoutImagenet. arXiv preprint arXiv:2011.13677, 2020." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16685", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in Neural Information Processing Systems, 33:11525-11538, 2020.", + "[60] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015.", + "[61] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018.", + "[62] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European conference on computer vision, pages 527-544. Springer, 2016.", + "[63] Joe Yue-Hei Ng, Jonghyun Choi, Jan Neumann, and Larry S Davis. Actionflownet: Learning motion representation for action recognition. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1616–1624. IEEE, 2018.", + "[64] Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. IEEE transactions on pattern analysis and machine intelligence, 36(6):1187-1200, 2013.", + "[65] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018.", + "[66] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016.", + "[67] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017.", + "[68] Rui Qian, Shuangrui Ding, Xian Liu, and Dahua Lin. Static and dynamic concepts for self-supervised video representation learning. In European Conference on Computer Vision, pages 145-164. Springer, 2022.", + "[69] Rui Qian, Yuxi Li, Huabin Liu, John See, Shuangrui Ding, Xian Liu, Dian Li, and Weiyao Lin. Enhancing self-supervised video representation learning via multi-level feature optimization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7990-8001, 2021.", + "[70] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6964-6974, 2021.", + "[71] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[72] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, et al. Bridging the gap to real-world object-centric learning. arXiv preprint arXiv:2209.14860, 2022.", + "[73] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8934-8943, 2018.", + "[74] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020.", + "[75] Jack Valmadre, Luca Bertinetto, Joao Henriques, Andrea Vedaldi, and Philip HS Torr. End-to-end representation learning for correlation filter based tracking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2805-2813, 2017.", + "[76] Carles Ventura, Miriam Bellver, Andreu Girbau, Amaia Salvador, Ferran Marques, and Xavier Giro-i Nieto. Rvos: End-to-end recurrent network for video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5277-5286, 2019.", + "[77] Paul Voigtlaender and Bastian Leibe. Online adaptation of convolutional neural networks for video object segmentation. arXiv preprint arXiv:1706.09364, 2017.", + "[78] Carl Vondrick, Abhinav Shrivastava, Alireza Fathi, Sergio Guadarrama, and Kevin Murphy. Tracking emerges by colorizing videos. In Proceedings of the European conference on computer vision (ECCV), pages 391-408, 2018.", + "[79] Naiyan Wang and Dit-Yan Yeung. Learning a deep compact image representation for visual tracking. Advances in neural information processing systems, 26, 2013.", + "[80] Xiaolong Wang, Kaiming He, and Abhinav Gupta. Transitive invariance for self-supervised visual representation learning. In Proceedings of the IEEE international conference on computer vision, pages 1329-1338, 2017.", + "[81] Xiaolong Wang, Allan Jabri, and Alexei A Efros. Learning correspondence from the cycle-consistency of time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2566-2576, 2019.", + "[82] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021.", + "[83] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018.", + "[84] Junyu Xie, Weidi Xie, and Andrew Zisserman. Segmenting moving objects via an object-centric layered representation. arXiv preprint arXiv:2207.02206, 2022.", + "[85] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level" + ], + "bbox": [ + 501, + 90, + 892, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "16686", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021.", + "[86] Jiarui Xu and Xiaolong Wang. Rethinking self-supervised correspondence learning: A video frame-level similarity perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10075-10085, 2021.", + "[87] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen, and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 585-601, 2018.", + "[88] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7177-7188, 2021.", + "[89] Yanchao Yang, Antonio Loquercio, Davide Scaramuzza, and Stefano Soatto. Unsupervised moving object detection via contextual information separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 879-888, 2019.", + "[90] Qixian Zhou, Xiaodan Liang, Ke Gong, and Liang Lin. Adaptive temporal encoding network for video instance-level human parsing. In Proceedings of the 26th ACM international conference on Multimedia, pages 1527-1535, 2018.", + "[91] Xizhou Zhu, Yujie Wang, Jifeng Dai, Lu Yuan, and Yichen Wei. Flow-guided feature aggregation for video object detection. In Proceedings of the IEEE international conference on computer vision, pages 408-417, 2017." + ], + "bbox": [ + 78, + 92, + 468, + 527 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "16687", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/0da0877f-f78d-4f97-9a49-55c50b2292de_model.json b/2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/0da0877f-f78d-4f97-9a49-55c50b2292de_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cdff9ee07b2b835d067415becd9ec5c9685ac381 --- /dev/null +++ b/2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/0da0877f-f78d-4f97-9a49-55c50b2292de_model.json @@ -0,0 +1,2745 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.18, + 0.131, + 0.79, + 0.176 + ], + "angle": 0, + "content": "Semantics Meets Temporal Correspondence: Self-supervised Object-centric Learning in Videos" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.203, + 0.724, + 0.221 + ], + "angle": 0, + "content": "Rui Qian1 Shuangrui Ding1 Xian Liu1 Dahua Lin1,2*" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.222, + 0.826, + 0.24 + ], + "angle": 0, + "content": "1The Chinese University of Hong Kong 2Shanghai Artificial Intelligence Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.242, + 0.662, + 0.256 + ], + "angle": 0, + "content": "{qr021,ds023,lx021,dhlin}@ie.cuhk.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.313, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.324, + 0.474, + 0.794 + ], + "angle": 0, + "content": "Self-supervised methods have shown remarkable progress in learning high-level semantics and low-level temporal correspondence. Building on these results, we take one step further and explore the possibility of integrating these two features to enhance object-centric representations. Our preliminary experiments indicate that query slot attention can extract different semantic components from the RGB feature map, while random sampling based slot attention can exploit temporal correspondence cues between frames to assist instance identification. Motivated by this, we propose a novel semantic-aware masked slot attention on top of the fused semantic features and correspondence maps. It comprises two slot attention stages with a set of shared learnable Gaussian distributions. In the first stage, we use the mean vectors as slot initialization to decompose potential semantics and generate semantic segmentation masks through iterative attention. In the second stage, for each semantics, we randomly sample slots from the corresponding Gaussian distribution and perform masked feature aggregation within the semantic area to exploit temporal correspondence patterns for instance identification. We adopt semantic- and instance-level temporal consistency as self-supervision to encourage temporally coherent object-centric representations. Our model effectively identifies multiple object instances with semantic structure, reaching promising results on unsupervised video object discovery. Furthermore, we achieve state-of-the-art performance on dense label propagation tasks, demonstrating the potential for object-centric analysis. The code is released at https://github.com/shvdiwnkozbw/SMTC." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.821, + 0.21, + 0.837 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.847, + 0.47, + 0.879 + ], + "angle": 0, + "content": "As one of the fundamental cognitive capabilities, human beings easily distinguish different objects, establish visual" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.291, + 0.697, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.374, + 0.815, + 0.386 + ], + "angle": 0, + "content": "(a) Query slot attention on semantic feature." + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.291, + 0.874, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.397, + 0.697, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.48, + 0.856, + 0.493 + ], + "angle": 0, + "content": "(b) Random sampling slot attention on correspondence map." + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.398, + 0.874, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.496, + 0.895, + 0.594 + ], + "angle": 0, + "content": "Figure 1. Fig. 1(a) presents the results of query slot attention on top of the RGB feature map. It successfully decomposes different semantics, e.g., camels and fence. Fig. 1(b) visualizes the correspondence map after PCA dimensionality reduction, showing that different instances have different correspondence patterns. And the slot attention with random sampling coarsely distinguishes two camels with some redundant borders. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.894, + 0.716 + ], + "angle": 0, + "content": "correspondence and perform object-centric analysis from temporally continuous observations. This ability can be attributed to two indispensable visual mechanisms: high-level semantic discrimination as well as low-level temporal correspondence, which enable humans to effectively understand and interact with the world." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Motivated by this, computer vision researchers equip machines with these capacities to enhance object-centric perception [79, 60, 34]. To achieve this goal, early works rely on human annotations or weak supervision to perceive object semantics [60, 63, 13], identify geometric positions [39, 71, 31, 91], and establish temporal correspondence [41, 79, 75, 77], but their generalization ability is limited. Recently, there emerge a host of fully unsupervised methods to learn robust representations for semantic discrimination [14, 38, 33, 70, 29, 37, 12] or spatiotemporal correspondence [44, 81, 86, 55, 52, 43], which achieve promising performance. Given this encouraging re" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.383, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author. Email: dhlin@ie.cuhk.edu.hk." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16675" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.467, + 0.15 + ], + "angle": 0, + "content": "sult, we naturally come up with a question: Is it possible to jointly leverage the semantics and correspondence to discover object instances and distill object-centric representations without human annotations?" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.155, + 0.467, + 0.411 + ], + "angle": 0, + "content": "Regarding this problem, our intuition is that the high-level semantics delineates meaningful foreground areas in a top-down manner, while when looking into more frames, the low-level correspondence temporally associates coherent objects and separates individual instances in a bottom-up fashion. For instance, in a football scene, the semantic cue differentiates the foreground that includes several players, while the temporal correspondence links distinct players through dynamic movements and geometric relationships. These two aspects collectively contribute to object-centric representations. Unfortunately, most of the existing works only concentrate on one of these features. [14, 38, 33, 70, 82] succeed in developing high-level semantics, but this abstract semantics alone is insufficient to distinguish instances. Whereas, [44, 81, 78, 55] excel in detailed correspondence, but lack semantic structure and result in redundancy and ambiguity." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.416, + 0.467, + 0.775 + ], + "angle": 0, + "content": "In this paper, we propose a new architecture, Semantics Meets Temporal Correspondence (SMTC), to jointly leverage semantics and temporal correspondence to distill object-centric representations from RGB sequences. Specifically, we first extract frame-wise visual features as the semantic representation. Then we calculate dense feature correlations between adjacent frames as the correspondence map which encodes temporal relationships. To mine the object-centric knowledge, we take inspiration from [59, 88, 50, 25], and investigate using different formulations of slot attention on them. The preliminary experiments show that the original slot attention with random sampling on RGB feature map suffers from complex scene components in real-world videos [59, 72], but the revised query slot attention [88, 46] can decompose different semantic components as shown in Fig. 1(a). As for the correspondence map, different objects present diverse temporal correspondence patterns. Comparing to semantic features, these patterns reveal low-level geometric relationships, which are comparatively simple but vary with specific scenes. Hence, query slot attention fails but the random sampling based formulation performs surprisingly well as shown in Fig. 1(b), coarsely separating different object areas with some redundant borders." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.781, + 0.467, + 0.9 + ], + "angle": 0, + "content": "Motivated by this, we develop semantic-aware masked slot attention, which comprises a set of Gaussian distributions with learnable mean and standard deviation vectors, on top of the fused semantic and correspondence representations. The intuition is that the mean vectors could represent potential semantic centers, which act similarly to the query slot attention to separate semantic components. While the deviation vectors introduce perturbations around" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.89, + 0.317 + ], + "angle": 0, + "content": "the semantic centers to capture distinct temporal correspondence patterns of different instances. Technically, we formulate two slot attention stages to achieve this goal. Firstly, we use the mean vectors as slot initialization to generate semantic segmentation masks. Secondly, for each semantics, we randomly sample slot vectors from the Gaussian distribution, then perform iterative attention and masked aggregation within the corresponding semantic mask area to distinguish instances. We enforce temporal consistency on the semantic masks as well as object instance slots to enhance temporal coherency and refine object-centric representations. Comparing with existing works on object-centric learning in videos [25, 50, 88, 84], our model is free of pre-computed motion or depth prior, and explicitly identifies multiple objects with semantic structure." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.319, + 0.89, + 0.559 + ], + "angle": 0, + "content": "In summary, our contributions are: (1) We propose a novel self-supervised architecture that unifies semantic discrimination and temporal correspondence to distill object-centric representations in videos. (2) We demonstrate that simple feature correlation can effectively represent temporal correspondence cues when used in conjunction with semantic features. Building on this observation, we develop semantic-aware masked slot attention, which operates on fused visual features and correspondence maps, to distinguish multiple object instances with semantic structure without relying on motion or depth priors. (3) We achieve promising results on unsupervised object discovery in both single and multiple object scenarios, and reach state-of-the-art performance on label propagation tasks, demonstrating that we learn discriminative and temporally consistent object-centric representations." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.573, + 0.64, + 0.589 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.6, + 0.89, + 0.884 + ], + "angle": 0, + "content": "Unsupervised Object Discovery is an essential process to formulate object-centric representations, which aims to identify objects without human annotations. There exist a series of works focusing on this problem [59, 32, 26, 27, 28, 7]. Typically, [59] develops slot attention to iteratively update latent object representations from random initializations but it has difficulty scaling to complex real-world scenes. To tackle this challenge, [72] employs feature reconstruction as objective to reduce redundancy, [46] develops query slot initialization to encode visual concepts. Further, a line of works extend object discovery to video domain [50, 48, 17, 5, 88]. Most of them build on slot attention architecture, and adopt extra pre-computed priors, e.g., optical flow [88, 84, 16, 21], depth [25], geometric positions [50], as input or supervision to assist object discovery. In our work, we find that the simple feature correlation provides informative temporal cues. And we develop semantic-aware masked slot attention to identify object instances with semantic structure without resorting to extra priors." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.887, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Self-supervised Representation Learning aims to learn" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16676" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.891, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.358, + 0.893, + 0.442 + ], + "angle": 0, + "content": "Figure 2. An overview of our framework. We first extract frame-wise features and calculate dense feature correlation, then fuse them to pass through semantic-aware masked slot attention, which comprises two slot attention stages with \\( N \\) shared learnable Gaussian distributions. In the first semantic slot attention stage, the mean vectors serve as slot initialization to generate a set of segmentation masks for semantic decomposition. In the second masked slot attention stage, which runs on \\( N \\) semantics in parallel, we randomly sample slots from the Gaussian distribution of each semantics and perform masked feature aggregation within the semantic area to identify distinct instances. We enforce semantic and instance temporal consistency to train the architecture in a teacher-student manner, with the teacher marked in gray." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.473, + 0.678 + ], + "angle": 0, + "content": "robust representation without human annotations. Early works design various pretext tasks to generate pseudo labels as self-supervision [49, 30, 47, 10, 1, 62]. Later, contrastive learning with instance discrimination revolutionized the field [38, 14, 70, 29, 35, 83, 65, 42, 11, 20, 3, 68, 35, 36, 69, 19, 22]. These works generally contribute to general semantics but lack the ability to distill object-centric representations. To delve into this problem, [82, 85] preserve spatial sensitivity and perform contrastive learning on dense features. [86, 12] point out that semantic pre-training implicitly encodes object correspondence. While in our work, we learn to explicitly identify different objects with instance-level consistency to progressively refine object-centric representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Self-supervised Correspondence is a fundamental problem in computer vision [55, 78, 81, 80]. Typical works rely on low-level statistics like color to generate the dense correspondence flow [51, 52, 78], or establish temporal cycles to track image regions [81, 44, 6]. Further, [43, 86, 12] incorporate correspondence learning with semantic discrimination, but they either regard correspondence as a side effect of semantic discrimination [86, 12] or independently learn two pathway features with late fusion [43]. Among these works, high-level semantics is not well utilized, and there exists high redundancy due to large background areas. In contrast, our method explicitly separates semantic components to alleviate redundancy and formulate instance-level correspondence from an object-centric perspective." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.465, + 0.591, + 0.48 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Our framework is shown in Fig. 2, which adopts a teacher-student structure similar to recent self-supervised methods [38, 12, 33]. Specifically, we first extract framewise features from an RGB sequence, and calculate dense feature correlation as temporal correspondence representations. After that, we fuse them to pass through semantic-aware masked slot attention to separate multiple object instances with semantic structure, and enforce temporal consistency to refine the object-centric representations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.645, + 0.673, + 0.661 + ], + "angle": 0, + "content": "3.1. Feature Encoding" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.67, + 0.892, + 0.701 + ], + "angle": 0, + "content": "Given a RGB sequence \\( v = \\{I_1, I_2, \\dots, I_T\\} \\), we employ a shared visual encoder \\( f \\) to extract frame-wise features:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.73, + 0.891, + 0.749 + ], + "angle": 0, + "content": "\\[\nF _ {t} = f \\left(I _ {t}\\right) \\in \\mathbb {R} ^ {H W \\times D}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "where \\( H, W, D \\) respectively denotes height, width and channel dimension, \\( t \\) indicates frame index. For simplicity, we omit the subscript indicating student or teacher pathway, and use \\( \\hat{F} \\) to denote teacher outputs. Intuitively, \\( F_{t} \\) contains appearance and semantic cues in the input sequence, and the next step is to formulate an effective representation for the temporal correspondence cues. To do this, inspired by the cost volume in motion estimation [24, 73, 74], we calculate dense feature correlation to generate the temporal" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16677" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.158, + 0.089, + 0.391, + 0.149 + ], + "angle": 0, + "content": "
FeatureQueryRandom
RGB58.138.4
Correlation21.556.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "Table 1. Preliminary results on RGB feature map and feature correlation. 'Query' ('Random') denotes slot attention with query (random sampling) initialization. We report IoU on DAVIS-2016." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.215, + 0.351, + 0.23 + ], + "angle": 0, + "content": "correspondence map for each timestamp:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.24, + 0.469, + 0.26 + ], + "angle": 0, + "content": "\\[\nC _ {t j} = F _ {t} F _ {j} ^ {T} \\in \\mathbb {R} ^ {H W \\times H W}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.27, + 0.469, + 0.346 + ], + "angle": 0, + "content": "where we randomly sample one frame index \\( j \\neq t \\) to compute feature correlation. The channel activations of \\( C_{tj} \\) encode low-level geometric correlations between frame \\( t \\) and \\( j \\), which indicates potential partitions of different objects as illustrated in Fig. 1(b)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.346, + 0.469, + 0.42 + ], + "angle": 0, + "content": "To jointly leverage the semantic and correspondence cues, we employ two linear transformation heads, \\( h_f \\) and \\( h_c \\), to respectively project \\( F_t \\) and \\( C_{tj} \\) to a shared \\( D \\)-dimensional embedding space, which are then fused with element-wise summation:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.431, + 0.469, + 0.45 + ], + "angle": 0, + "content": "\\[\nR _ {t} = h _ {f} \\left(F _ {t}\\right) + h _ {c} \\left(C _ {t j}\\right) \\in \\mathbb {R} ^ {H W \\times D}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.461, + 0.469, + 0.506 + ], + "angle": 0, + "content": "In this way, \\( R_{t} \\) simultaneously encodes semantic and temporal correspondence information, serving as an intermediate feature for object-centric analysis." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.458, + 0.531 + ], + "angle": 0, + "content": "3.2. Preliminary Experiments with Slot Attention" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.469, + 0.659 + ], + "angle": 0, + "content": "Given these representations, the next step is to formulate effective ways to identify individual objects. To do this, we take inspiration from slot attention [59, 50, 88, 46], which employs a set of slot vectors to iteratively attend to specific objects. Generally, there are two alternatives for slot initialization, random sampling from a Gaussian distribution with learnable parameters \\(\\mu, \\sigma \\in \\mathbb{R}^D\\) [59, 50] or directly inheriting from a set of learnable queries [46, 88]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To investigate the effectiveness of these two formulations, we conduct preliminary experiments on the semantic feature \\( F_{t} \\) and correspondence map \\( C_{tj} \\) respectively. We follow the settings in [72], using a frozen DINO pre-trained ViT to extract visual feature and generate correspondence map, on top of which we perform two kinds of slot attention. We train and evaluate them on DAVIS-2016 dataset [66]. As indicated by Fig. 1 and Table. 1, we find that query slot attention decomposes distinct semantic components on \\( F_{t} \\), while the original slot attention with random sampling performs much worse and struggles in complex real-world scenes. Conversely, query slot attention does not work on the correspondence map, and random sampling initialization enables the model to effectively utilize the temporal correspondence patterns in \\( C_{tj} \\) to distinguish instances. An intuitive explanation into these observations is the RGB" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.304 + ], + "angle": 0, + "content": "features contain rich semantics and there exist consistent patterns for similar objects. Slot attention with learnable queries can capture inter-sample semantic patterns to identify objects. While random initialization fails to memorize semantic patterns, thus resulting in worse performance. In contrast, correspondence features reveal inter-frame geometric correlations, shown in Fig. 1(b), which vary with specific scenes. There are no consistent patterns for similar objects, so the learnable queries fail but random initialization works well to capture correspondence cues. The complementarity between these two conditions motivates two-stage slot attention design to combine the attributes of query and random sampling initialization, thus better exploiting the semantic and correspondence cues in \\( R_{t} \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.311, + 0.836, + 0.326 + ], + "angle": 0, + "content": "3.3. Semantic-aware Masked Slot Attention" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.56 + ], + "angle": 0, + "content": "We propose semantic-aware masked slot attention, which comprises \\(N\\) Gaussian distributions with learnable means \\(\\pmb{\\mu} = \\{\\mu_1,\\mu_2,\\dots,\\mu_N\\} \\in \\mathbb{R}^{N\\times D}\\) and standard deviations \\(\\sigma = \\{\\sigma_{1},\\sigma_{2},\\dots,\\sigma_{N}\\} \\in \\mathbb{R}^{N\\times D}\\). With this design, the mean vectors act similarly to the learnable queries, each representing a potential semantic center. And the deviations introduce perturbations around semantic centers to model temporal correspondence patterns. To achieve this goal, we employ two stages of iterative attention with shared learnable parameters. First, we use the mean vectors as slot initialization to decompose different semantics in the form of segmentation. Second, we randomly sample slots from each Gaussian distribution \\((\\mu_n,\\mathrm{diag}(\\sigma_n))\\), and perform masked aggregation within the corresponding semantic area to identify instances. The process is detailed as follows." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.892, + 0.667 + ], + "angle": 0, + "content": "Semantic Decomposition. In the first stage, we directly use \\(\\pmb{\\mu}\\) without randomness as slot initialization \\(S = \\pmb{\\mu} \\in \\mathbb{R}^{N \\times D}\\). Then following the standard slot attention procedure [59], we employ three linear transformation heads to map \\(S\\) and \\(R_{t}\\) into \\(Q \\in \\mathbb{R}^{N \\times D}, K, V \\in \\mathbb{R}^{HW \\times D}\\), and iteratively calculate attention score and update slot representations. Mathematically, we formulate each iteration as:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.686, + 0.891, + 0.718 + ], + "angle": 0, + "content": "\\[\nM := \\operatorname {s o f t m a x} \\left(\\frac {1}{\\sqrt {D}} Q K ^ {T}, 0\\right) \\in \\mathbb {R} ^ {N \\times H W}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.719, + 0.891, + 0.754 + ], + "angle": 0, + "content": "\\[\nA [ n, i ] := \\frac {M [ n , i ]}{\\sum_ {l} M [ n , l ] + \\epsilon}, A \\in \\mathbb {R} ^ {N \\times H W}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.756, + 0.891, + 0.772 + ], + "angle": 0, + "content": "\\[\nS := \\operatorname {G R U} (\\text {i n p u t} = A V, \\text {s t a t e s} = S). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Note that the slot attention weight \\( M \\) is normalized along the query axis, and the weighted mean coefficient \\( A \\) is employed to aggregate the value to update the slots. This mechanism induces competition among slots and enforce different slots to take over different semantic features. We iterate this process for three times, and take the slot attention weight (slot vectors) of the final iteration as the segmentation mask (representation) for different semantic compo" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16678" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "nents. To avoid ambiguity in presentation, we denote them as \\(\\mathcal{M} \\in \\mathbb{R}^{N \\times HW}\\) and \\(S \\in \\mathbb{R}^{N \\times D}\\), where \\(S[n]\\) presents \\(n\\)-th semantic center representation, \\(\\mathcal{M}[n]\\) indicates the probability that each pixel belongs to \\(n\\)-th semantics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.473, + 0.424 + ], + "angle": 0, + "content": "Instance Identification. Within each semantics, there could exist multiple object instances. The Gaussian distributions we introduce help to distinguish them with temporal correspondence cues. Specifically, in the second stage, for each semantics, we randomly sample slots from the Gaussian distribution, using the perturbations around the semantic center to capture distinct temporal correspondence patterns. For simplicity, we take the \\(n\\)-th semantics as an example. First, we sample \\(P\\) vectors, \\(S \\sim \\mathcal{N}(\\mu_n, \\mathrm{diag}(\\sigma_n)) \\in \\mathbb{R}^{P \\times D}\\), as slot initialization to represent \\(P\\) potential instances. Next, we follow the first stage formulation to respectively project \\(S\\) and \\(R_t\\) into \\(Q \\in \\mathbb{R}^{P \\times D}\\) and \\(K, V \\in \\mathbb{R}^{HW \\times D}\\), and obtain the attention score \\(M \\in \\mathbb{R}^{P \\times HW}\\). To preserve the semantic attributes and discriminate correspondence cues, we perform masked feature aggregation within the semantic mask areas. The only mathematical difference in this process is the weighted mean coefficient computation:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.434, + 0.47, + 0.494 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} A [ p, i ] := \\frac {M [ p , i ] \\widetilde {\\mathcal {M}} [ n , i ]}{\\sum_ {l} M [ p , l ] \\widetilde {\\mathcal {M}} [ n , l ] + \\epsilon}, A \\in \\mathbb {R} ^ {P \\times H W}, \\tag {7} \\\\ \\widetilde {\\mathcal {M}} [ n, i ] = \\operatorname {b i n a r i z e} (\\mathcal {M} [ n, i ], \\tau), \\mathcal {M} \\in \\mathbb {R} ^ {N \\times H W}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.473, + 0.719 + ], + "angle": 0, + "content": "where we use the \\(n\\)-th semantic mask \\(\\widetilde{\\mathcal{M}}[n] \\in \\{0,1\\}^{HW}\\) binarized with threshold \\(\\tau\\) to constrain the model to only aggregate the visual features related to \\(n\\)-th semantics. And we empirically find a comparatively large threshold \\(\\tau = 0.5\\) helps to preserve object regions with clear semantics and filter out background. The masked slot attention for all semantics follows the same rule and is calculated in parallel. Similarly, we iterate the process for three times, use the final slot attention weight as the instance segmentation masks, and take the final slot vectors as the object-centric representations \\(\\mathcal{O} \\in \\mathbb{R}^{N \\times P \\times D}\\), where \\(\\mathcal{O}[n,p] \\in \\mathbb{R}^D\\) denotes the \\(p\\)-th potential object instance of \\(n\\)-th semantics. In this way, our semantic-aware masked slot attention explicitly segments multiple instances with semantic structure." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.182, + 0.743 + ], + "angle": 0, + "content": "3.4. Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.78 + ], + "angle": 0, + "content": "We use temporal consistency as self-supervision to optimize semantic masks \\(\\mathcal{M}\\) and instance representations \\(\\mathcal{O}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Dense Semantic Alignment. Given the semantic mask of each timestamp \\(\\mathcal{M}_t\\in \\mathbb{R}^{N\\times HW}\\), as well as \\(\\hat{\\mathcal{M}}_t\\) from teacher pathway, we aim to align the spatially dense semantic distributions across time. To achieve this, the first step is to determine the corresponding patches between different frames. Utilizing the feature correlation \\(C_{tj}\\in \\mathbb{R}^{HW\\times HW}\\) which indicates dense feature similarity between timestamps \\(t,j\\), we can infer patch-level correspondence from" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.893, + 0.153 + ], + "angle": 0, + "content": "this cue with optimal transport [11, 2, 58, 18]. Formally, we take \\(-\\hat{C}_{tj}\\) from teacher pathway as cost matrix, and solve the optimal transport strategy \\(\\pi_{tj}^{*} \\in \\mathbb{R}^{HW \\times HW}\\) between two uniform distributions to indicate patch correspondence:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.174, + 0.892, + 0.324 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\pi_ {t j}} \\sum_ {u = 1} ^ {H W} \\sum_ {v = 1} ^ {H W} - \\hat {C} _ {t j} [ u, v ] \\pi_ {t j} [ u, v ] \\\\ \\text {s . t .} \\sum_ {v = 1} ^ {H W} \\pi_ {t j} [ \\cdot , v ] = \\frac {1}{H W} \\mathbf {1} ^ {H W}, \\tag {8} \\\\ \\sum_ {u = 1} ^ {H W} \\pi_ {t j} [ u, \\cdot ] = \\frac {1}{H W} \\mathbf {1} ^ {H W}, \\\\ \\pi_ {t j} [ u, v ] \\geq 0 \\quad u, v \\in \\{1, 2, \\dots , H W \\}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.893, + 0.457 + ], + "angle": 0, + "content": "where \\( u, v \\) denotes spatial index, \\( \\mathbf{1}^{HW} \\) is a \\( HW \\)-dimensional vector of all ones. Note that it is feasible to adopt other marginal distribution formulations with prior knowledge, e.g., class-agnostic activation map [4], to facilitate training, with details discussed in Supplementary Material. We employ Sinkhorn-Knopp algorithm [18] to obtain the optimal transport matrix \\( \\pi_{t,j}^* \\), and formulate the dense semantic alignment objective in the form of cross entropy:" + }, + { + "type": "equation", + "bbox": [ + 0.503, + 0.477, + 0.892, + 0.541 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s e m} = - \\sum_ {\\substack {t, j = 1 \\\\ t \\neq j}} ^ {T} \\sum_ {u, v = 1} ^ {H W} \\sum_ {n = 1} ^ {N} \\pi_ {t j} ^ {*} [ u, v ] \\hat {\\mathcal {M}} _ {j} [ n, v ] \\log \\mathcal {M} _ {t} [ n, u ]. \\tag{9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.595 + ], + "angle": 0, + "content": "By minimizing the weighted sum of cross entropy, we achieve temporally consistent semantic distributions among corresponding spatial areas." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.892, + 0.643 + ], + "angle": 0, + "content": "Semantic Mask Regularization. To encourage the semantic centers to emphasize different visual contents and cover diverse semantics, we apply a simple regularization to \\(\\mathcal{M}_t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.663, + 0.892, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e g} = \\sum_ {t = 1} ^ {T} \\sum_ {\\substack {n, j = 1 \\\\ j \\neq n}} ^ {N} \\frac {\\mathcal {M} _ {t} [ n ] ^ {T} \\mathcal {M} _ {t} [ j ]}{\\| \\mathcal {M} _ {t} [ n ] \\| _ {2} \\| \\mathcal {M} _ {t} [ j ] \\| _ {2}}. \\tag{10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.75 + ], + "angle": 0, + "content": "This simple regularization suppresses the cosine similarity between different semantic masks to avoid collapse." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Instance Representation Consistency. Apart from the semantic distributions, it is necessary to ensure that the representations of each object instance are temporally consistent. Given the instance representations \\(\\mathcal{O}_t\\in \\mathbb{R}^{N\\times P\\times HW}\\), as well as \\(\\hat{\\mathcal{O}}_t\\) from teacher pathway, we need to first match corresponding instances between different timestamps. For illustration, considering \\(n\\)-th semantics of time \\(t,j\\), we adopt bipartite matching [9, 15] based on the cosine similarity between \\(\\mathcal{O}_t[n]\\) and \\(\\hat{\\mathcal{O}}_j[n]\\) to generate one-to-one instance correspondence, with the matching function denoted as \\(\\varepsilon (\\cdot)\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16679" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.09, + 0.473, + 0.304 + ], + "angle": 0, + "content": "In this way, \\(\\mathcal{O}_t[n,p]\\) and \\(\\hat{\\mathcal{O}}_j[n,\\varepsilon(p)]\\) are considered as a positive pair to be aligned. However, for each video, there exists absent semantics and object occlusion such that the number of visible object instances varies. Thus, there could be redundant slots not attending to objects, and it is crucial to filter out these invalid instance representations to reduce distractions. Mathematically, for timestamp \\(t\\), we introduce \\(\\mathcal{I}_t \\in \\{0,1\\}^{N \\times P}\\) as the valid instance indicator. For the \\(p\\)-th instance of \\(n\\)-th semantics to be valid, two criteria must be met: (1) The ratio of the \\(n\\)-th semantic area is above threshold \\(\\tau_1 = 0.2\\) to filter out non-existing semantics; (2) The instance representation is close to the semantic center representation with cosine similarity larger than \\(\\tau_2 = 0.5\\) to exclude redundant slots. The conditions are formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.321, + 0.47, + 0.386 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {t} [ n, p ] = 1 \\Leftrightarrow \\left\\{ \\begin{array}{l} \\frac {1}{H W} \\sum_ {u = 1} ^ {H W} \\widetilde {\\mathcal {M}} _ {t} [ n, u ] \\geq \\tau_ {1}, \\\\ \\cos \\left(\\mathcal {S} _ {t} [ n ], \\mathcal {O} _ {t} [ n, p ]\\right) \\geq \\tau_ {2}. \\end{array} \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.391, + 0.472, + 0.438 + ], + "angle": 0, + "content": "And \\(\\hat{\\mathcal{L}}_j\\in \\{0,1\\}^{N\\times P}\\) is defined in the same manner. We use a margin loss to encourage object detail consistency over valid instance representations:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.444, + 0.47, + 0.564 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal{L}_{obj} = \\sum_{\\substack{t,j = 1\\\\ t\\neq j}}^{T}\\sum_{n = 1}^{N}\\sum_{p = 1}^{P}\\mathcal{I}_{t}[n,p]\\big\\{ \\\\ \\hat {\\mathcal {I}} _ {j} [ n, \\varepsilon (p) ] \\| \\mathcal {O} _ {t} [ n, p ] - \\hat {\\mathcal {O}} _ {j} [ n, \\varepsilon (p) ] \\| _ {2} \\tag {12} \\\\ + \\sum_{\\substack{q = 1\\\\ q\\neq \\varepsilon (p)}}^{P}\\mathrm{relu}(\\lambda -\\| \\mathcal{O}_{t}[n,p] - \\hat{\\mathcal{O}}_{j}[n,q]\\|_{2})\\} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.47, + 0.66 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a margin hyper-parameter. Note that with the \\(\\mathcal{I}_t\\) and \\(\\hat{\\mathcal{I}}_j\\) constraints, our formulation can handle object occlusions which lead to varying number of visible instances in different frames. By minimizing the margin loss, we encourage the model to distill discriminative and temporally consistent object-centric representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.661, + 0.47, + 0.691 + ], + "angle": 0, + "content": "Overall, we take the summation of three objectives for training:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.698, + 0.469, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {s e m}} + \\mathcal {L} _ {\\text {o b j}} + \\mathcal {L} _ {\\text {r e g}}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.47, + 0.751 + ], + "angle": 0, + "content": "The student pathway \\(\\theta\\) is updated with gradient descent, and the teacher parameters \\(\\hat{\\theta}\\) are updated with momentum as:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.77, + 0.469, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\hat {\\theta} \\leftarrow m \\hat {\\theta} + (1 - m) \\theta , \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\( m \\) is momentum coefficient set to 0.999 in default. This momentum update mechanism results in a slowly evolving teacher network, which progressively distills object knowledge, provides reliable self-supervision signals and enables us to train semantic-aware masked slot attention without relying on widely adopted reconstruction objective in existing works [59, 50, 88, 84, 72, 46]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.634, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.596, + 0.13 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.894, + 0.366 + ], + "angle": 0, + "content": "We train our model on YouTube-VOS [87], a challenging video dataset that contains multiple object instances of distinct semantics in each video. We evaluate our method on two lines of tasks: (1) Unsupervised object segmentation on DAVIS-2016 [66], SegTrack-v2 [54], FMBS-59 [64] and challenging multiple object segmentation on DAVIS-2017-Unsupervised [8]. We respectively calculate the mean per frame intersection over union (IoU) and \\(\\mathcal{J}\\&\\mathcal{F}\\) on single and multiple object discovery benchmarks. (2) Lable propagation tasks including semi-supervised video object segmentation on DAVIS-2017 [67], human pose tracking on JHMDB [45], human part tracking on VIP [90]. We adopt the same settings as [44, 43] and report standard \\(\\mathcal{J}\\) and \\(\\mathcal{F}\\) score on DAVIS, probability of a correct pose (PCK) on JHMDB and mean intersection over union (mIoU) on VIP." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.375, + 0.719, + 0.391 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.893, + 0.55 + ], + "angle": 0, + "content": "We sample \\( T = 4 \\) frames with frame rate FPS \\( = 4 \\) as the input RGB sequence, where each frame is augmented with random crop, horizontal flip and color jitter, and finally resized to \\( 256 \\times 256 \\). We adopt ViT-Small/16 [23] and ResNet-50 [40] as the visual encoder, specified in each experiment, to extract frame-wise features. Then for semantic-aware masked slot attention, we set the number of learnable Gaussian distributions to \\( N = 16 \\), the number of instances of each semantics to \\( P = 4 \\) in default and use 3 iterations to update slot representations and attention maps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.763 + ], + "angle": 0, + "content": "In training, the visual encoder is initialized with self-supervised pre-trained weights from DINO ViT-Small/16 [12] or MoCo ResNet-50 [38]. We use AdamW optimizer [61] with learning rate \\( 2 \\times 10^{-4} \\) for batch size 128 to update student parameters, and the teacher pathway is updated in momentum. In inference, we employ the same evaluation protocol as [44] to evaluate the temporal object correspondence performance on label propagation tasks. And for object discovery, we maintain the valid instances filtered with Eq. 11 as candidate objects. For single object evaluation, we merge all candidate objects as foreground areas to calculate IoU. For multiple object benchmark, we follow [8] to match ground-truth and our predictions and report \\( \\mathcal{J} \\& \\mathcal{F} \\) score." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.772, + 0.795, + 0.788 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Single Object Discovery. To measure our model's ability to decompose different objects, we first present the quantitative results on single object discovery without post processing (e.g., CRF, spectral clustering) in Table 2. Many existing works resort to optical flow [89, 53, 88, 84] or synthetic data [84] as weak supervision to learn temporal dynamics and generate semantic-agnostic segmentation masks for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16680" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.472, + 0.26 + ], + "angle": 0, + "content": "
ModelRGBFlowDAVISST-v2FBMS
CIS [89]✓✓71.562.563.5
AMD [57]✓X57.857.047.5
DINO [12]✓X52.346.550.3
SIMO [53]X✓67.862.0-
MG [88]X✓68.358.653.1
OCLR [84]X✓72.167.665.4
GWM [16]✓✓71.269.066.9
SMTC✓X71.869.368.4
\\( SMTC^† \\)✓X70.868.466.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.261, + 0.47, + 0.318 + ], + "angle": 0, + "content": "Table 2. Quantitative results on single object discovery. We compare per frame mean IoU on DAVIS-2016, SegTrack-v2 and FBMS-59 without any post-processing. \\(\\mathsf{SMTC}^{\\dagger}\\) denotes only with first slot attention stage for semantic decomposition in inference." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.328, + 0.471, + 0.44 + ], + "angle": 0, + "content": "
ModelBackboneJ&FJF
DINOSAUR [72]ViT-S/1621.419.223.7
SMTCViT-S/1640.536.444.6
SMTCResNet-5039.035.542.6
RVOS [76]ResNet-10141.236.845.7
ProReduce [56]ResNet-10168.365.071.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.44, + 0.47, + 0.469 + ], + "angle": 0, + "content": "Table 3. Quantitative results on multiple object discovery on DAVIS-2017-Unsupervised. Gray denotes supervised training." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.493, + 0.47, + 0.902 + ], + "angle": 0, + "content": "moving objects. DINO baseline [12] uses the self-attention map from the last layer as an object segmentation mask. While our method only relies on RGB frames and explicitly discriminates object semantics in a fully self-supervised manner. From the comparison, we observe that our SMTC with only semantic decomposition largely exceeds RGB-only methods [12, 57] and comparable to optical flow-based methods, and our formulation with instance identification further improves the performance. This phenomenon indicates that the learned semantic decomposition with temporal correspondence cues have the potential to substitute precomputed optical flow to guide single object segmentation. Multiple Object Discovery. Most of the existing video segmentation methods with fully unsupervised training cannot explicitly distinguish multiple object instances. OCLR [84] adopts layered flow representations to identify multiple objects on re-annotated DAVIS-2017-motion dataset, but it cannot distinguish instances with common motion. On the contrary, our method jointly utilizes semantics and temporal correspondence to separate arbitrary object instances on DAVIS-2017-Unsupervised benchmark. For fair comparison, we re-run DINOSAUR [72] on DAVIS-2017 with 5 slots as an unsupervised baseline. As presented in Table 3, our method significantly outperforms DINOSAUR. This is because our method explicitly decomposes different semantics, exploits temporal correspondence among multiple frames to distinguish instances and encourages instance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.242 + ], + "angle": 0, + "content": "level temporal consistency. While DINOSAUR is only supervised with DINO feature reconstruction without considering temporal relationships and semantic discrimination. Besides, our method is comparable with supervised baseline RVOS [76], and the gap to supervised state-of-the-art [56] is majorly due to human annotated masks as supervision leading to much more precise segmentation boundaries. This conjecture can be demonstrated by Fig. 3, which also reveals that our model is able to learn discriminative object-centric representations to separate different instances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.892, + 0.426 + ], + "angle": 0, + "content": "Label Propagation. Finally, we compare the performance on label propagation tasks in Table 4 to validate the temporal consistency of our learned features. Though not specifically designed for these tasks, our model achieves state-of-the-art results on three benchmarks. The most related work to our SMTC is SFC [43], which also utilizes both high-level semantic and low-level correspondence. The difference is that SFC only employs late fusion to incorporate these two cues in inference, while our model jointly exploits semantics and temporal correspondence in training to identify object instances, perform instance-level alignment and refine temporally consistent object-centric representations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.437, + 0.655, + 0.453 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.506 + ], + "angle": 0, + "content": "We report IoU on DAVIS-2016, and \\(\\mathcal{J}\\&\\mathcal{F}\\) on DAVIS-2017-Unsupervised for single and multiple object discovery. Refer to Supplementary Material for more ablations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.688 + ], + "angle": 0, + "content": "Slot Attention Formulation. We compare the different formulations of semantic-aware masked slot attention in Table 5. With only the first slot attention stage, i.e., the 'Semantic' part, the model cannot distinguish object instances of the same semantics, thus the performance on multiple object discovery drops significantly as indicated by Ours-B. While with only the second stage, i.e., the 'Instance' part, the performance on both single and multiple object segmentation drops as indicated by Ours-C. This reveals that the pre-computed semantic masks play an important role in alleviating interference in instance slot update and filtering valid sample for instance-level alignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Number of Learnable Gaussian Distributions. Comparing Ours-A, Ours-E and Ours-F in Table 5, we observe that the performance improves when \\( N \\) increases since larger \\( N \\) guides our model to distill more fine-grained semantics. And we present an extreme setting with \\( N = 1 \\) in Ours-D, where our formulation degenerates into the original slot attention with random sampling from one Gaussian distribution [59]. Note that although Ours-C also has no access to semantic masks, it still maintains the potential to discriminate different semantics due to multiple learnable means. Its superiority to Ours-D demonstrates the effectiveness of semantic discrimination in object discovery." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Feature Usage. We explore using different feature input to our semantic-aware masked slot attention in Table 6. We" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16681" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.145, + 0.089, + 0.827, + 0.318 + ], + "angle": 0, + "content": "
ModelBackboneDAVISJHMDBVIP
J&FJFPCK@0.1PCK@0.2mIoU
Supervised [40]ResNet-5066.063.768.459.278.339.5
MoCo [38]ResNet-5065.463.267.660.479.336.1
VFS [86]ResNet-5068.966.571.360.980.743.2
DINO [12]ViT-S/1661.860.263.445.475.237.9
TimeCycle [81]ResNet-5040.741.939.457.778.528.9
UVC [55]ResNet-5056.354.558.156.576.634.2
MAST [51]ResNet-1865.563.367.6---
CRW [44]ResNet-1867.664.870.258.880.337.6
SFC [43]ResNet-18+ResNet-5071.268.374.061.983.038.4
SMTCViT-S/1667.664.171.253.279.639.2
SMTCResNet-5073.069.476.662.584.138.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.318, + 0.892, + 0.348 + ], + "angle": 0, + "content": "Table 4. Quantitative results on label propagation tasks: semi-supervised video object segmentation on DAVIS-2017, pose tracking on JHMDB, human part tracking on VIP. We present the backbone for comprehensive comparison, and report comparing results from [86, 43]." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.368, + 0.457, + 0.5 + ], + "angle": 0, + "content": "
ModelNSemanticMaskedIoUJ&F
Ours-A16✓✓71.840.5
Ours-B16✓X66.424.5
Ours-C16X✓52.130.3
Ours-D1-✓47.422.1
Ours-E8✓✓68.437.9
Ours-F32✓✓71.940.7
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.501, + 0.468, + 0.556 + ], + "angle": 0, + "content": "Table 5. Ablation studies on the number of learnable Gaussian distributions, and the formulation of semantic-aware masked slot attention. 'Semantic' ('Instance') denotes the first (second) attention stage for semantic decomposition (instance identification)." + }, + { + "type": "table", + "bbox": [ + 0.091, + 0.574, + 0.457, + 0.721 + ], + "angle": 0, + "content": "
FeatureSemanticInstanceIoUJ&F
Fused✓✓71.840.5
RGB-only✓✓67.720.1
✓×67.518.8
×✓41.515.3
Correlation-only✓✓38.814.2
✓×21.110.7
×✓59.427.5
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.722, + 0.468, + 0.75 + ], + "angle": 0, + "content": "Table 6. Ablation studies on the feature usage. We compare different slot attention formulations on different feature inputs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.903 + ], + "angle": 0, + "content": "observe that with RGB-only input, the 'Semantic' part guarantees satisfactory performance on single object discovery, but performs poor on multiple object benchmarks even with the 'Instance' part. This coincides with our intuition that it is difficult to distinguish object instances with solely semantic cues, instead we need to look into more frames and resort to temporal correspondence. While for the correlation-only input, the 'Instance' part coarsely segments objects and" + }, + { + "type": "table", + "bbox": [ + 0.563, + 0.368, + 0.83, + 0.446 + ], + "angle": 0, + "content": "
LsemLregLobjIoUJ&F
✓XX61.321.6
✓✓X66.424.5
XX✓50.630.4
✓✓✓71.740.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.446, + 0.892, + 0.473 + ], + "angle": 0, + "content": "Table 7. Ablation studies on the learning objectives. We report the results on DAVIS-2016 and DAVIS-2017-Unsupervised." + }, + { + "type": "table", + "bbox": [ + 0.603, + 0.485, + 0.791, + 0.564 + ], + "angle": 0, + "content": "
TFPSIoUJ&F
2465.436.9
4266.337.7
4471.840.5
8471.940.8
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.564, + 0.891, + 0.604 + ], + "angle": 0, + "content": "Table 8. Ablation studies on the hyper-parameters of frame sampling. We compare different number of frames per clip and different FPS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.891, + 0.689 + ], + "angle": 0, + "content": "the 'Semantic' part exerts negative impact. This is because the correspondence map contains little semantic information, the semantic decomposition results in severe ambiguity, thus impairing the performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Learning Objectives. We also present a comprehensive ablation on the learning objectives in Table 7, where we report single object discovery on DAVIS-2016 and multiple object discovery results on DAVIS-2017-Unsupervised. From the comparison, we observe that \\(\\mathcal{L}_{sem}\\) is fundamental to decompose object semantics, which demonstrates the necessity of semantic decomposition in the first stage. \\(\\mathcal{L}_{reg}\\) also leads to improvement by encouraging the learnable queries to cover more diverse semantics. As for \\(\\mathcal{L}_{obj}\\), this objective introduces instance-level alignment and significantly enhances multiple object discovery performance on DAVIS-2017-Unsupervised." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Frame Sampling. Another important thing in training is the frame sampling procedure. The number of frames per" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16682" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.085, + 0.092, + 0.47, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.175, + 0.468, + 0.203 + ], + "angle": 0, + "content": "Figure 3. Visualization of semantic and instance segmentation map. The red boxes outline the ambiguous areas." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.214, + 0.47, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.276, + 0.468, + 0.303 + ], + "angle": 0, + "content": "Figure 4. Visualization of instance alignment. The arrows point out the matched instances across time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.468, + 0.446 + ], + "angle": 0, + "content": "clip and sampling frame rate determines the temporal reception field of our model. We compare different sampling hyper-parameters in Table 8. The results indicate that it is necessary to use large FPS to provide rich temporal dynamics. And by setting comparatively large FPS, our model can have access to more temporal information without introducing more frames, reaching a trade-off between performance and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.455, + 0.262, + 0.47 + ], + "angle": 0, + "content": "4.5. Further Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.468, + 0.764 + ], + "angle": 0, + "content": "Visualization of Object Discovery. We visualize the our semantic decomposition map and instance identification map in Fig. 3, where the same color denotes the segmentation mask of the same semantics. The visualization reveals that our method is able to distinguish multiple object instances with semantic structure. For example, different people belong to the same semantic center, and quadrupeds such as dogs and pigs belong to another semantics. And our model also separates different instances of the same semantics, such as multiple persons and different pigs, with minor ambiguity on small object part, e.g., pig legs, mobile phones. Besides, we also visualize our instance-level alignment during training in Fig. 4. Our bipartite matching well aligns the corresponding instances in different frames, and the valid sample constraint of Eq. 11 effectively handles object occlusion in videos. For example, when encountering an occluded person instance, our method filters out the redundant slots not attending to objects, only correlating valid slots for robust instance-level alignment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Limitation. Due to the lack of the human annotated segmentation masks, our model faces challenges in generating precise boundaries for each instance, particularly for small objects. One potential solution is to incorporate multi-scale feature pyramid to further improve dense perception. We leave it in the future work. Despite this limitation, our work effectively demonstrates the benefits of leveraging both semantics and temporal correspondence to discover object instances with semantic structure and distill discriminative" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.872, + 0.107 + ], + "angle": 0, + "content": "and temporally consistent object-centric representations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.121, + 0.619, + 0.136 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.146, + 0.892, + 0.449 + ], + "angle": 0, + "content": "In this work, we propose a novel self-supervised framework jointly exploiting high-level semantics and low-level temporal correspondence to enhance object-centric perception. Specifically, we represent semantic and temporal correspondence cues using the RGB feature map and dense feature correlation, respectively. These cues are fused and fed into semantic-aware masked slot attention which comprises a set of learnable Gaussian distributions. This design allows us to leverage the mean vectors as potential semantic centers for semantic decomposition, and use the perturbations introduced by the standard deviation vectors around the semantic centers to make use of the temporal correspondence cues for instance identification. To distill discriminative and temporally consistent object-centric representations, we devise semantic- and instance-level alignment that is robust to object occlusion as self-supervision. We demonstrate the effectiveness of our model for object-centric analysis through the state-of-the-art performance on label propagation tasks, as well as the promising results on unsupervised object discovery in both single and multiple object scenarios." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.463, + 0.661, + 0.48 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.892, + 0.58 + ], + "angle": 0, + "content": "We thank Tianfan Xue for constructive feedback. This project is funded in part by Shanghai AI Laboratory (P23KS00020, 2022ZD0160201), CUHK Interdisciplinary AI Research Institute, and the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.593, + 0.597, + 0.608 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.618, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[1] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. Advances in Neural Information Processing Systems, 33:9758-9770, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.689, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[2] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. arXiv preprint arXiv:1911.05371, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.746, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[3] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.803, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[4] Kyungjune Baek, Minhyun Lee, and Hyunjung Shim. Psynet: Self-supervised approach to object localization using point symmetric transformation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10451-10459, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[5] Beril Besbinar and Pascal Frossard. Self-supervision by prediction for object discovery in videos. In 2021 IEEE Interna" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.618, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16683" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.093, + 0.469, + 0.119 + ], + "angle": 0, + "content": "tional Conference on Image Processing (ICIP), pages 1509-1513. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.121, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[6] Zhangxing Bian, Allan Jabri, Alexei A Efros, and Andrew Owens. Learning pixel trajectories with multiscale contrastive random walks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6508-6519, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.19, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[7] Christopher P Burgess, Loic Matthey, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matt Botvinick, and Alexander Lerchner. Monet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.246, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[8] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv preprint arXiv:1905.00737, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.302, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[9] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.441 + ], + "angle": 0, + "content": "[10] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[11] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.469, + 0.636 + ], + "angle": 0, + "content": "[13] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.469, + 0.692 + ], + "angle": 0, + "content": "[14] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[15] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[16] Subhabrata Choudhury, Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Guess what moves: unsupervised video and image segmentation by anticipating motion. arXiv preprint arXiv:2205.07844, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[17] Eric Crawford and Joelle Pineau. Exploiting spatial invariance for scalable unsupervised object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 3684-3692, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[18] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[19] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9716-9726, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[20] Shuangrui Ding, Rui Qian, and Hongkai Xiong. Dual contrastive learning for spatio-temporal representation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5649-5658, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.292 + ], + "angle": 0, + "content": "[21] Shuangrui Ding, Weidi Xie, Yabo Chen, Rui Qian, Xiaopeng Zhang, Hongkai Xiong, and Qi Tian. Motion-inductive self-supervised object discovery in videos. arXiv preprint arXiv:2210.00221, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.295, + 0.892, + 0.35 + ], + "angle": 0, + "content": "[22] Shuangrui Ding, Peisen Zhao, Xiaopeng Zhang, Rui Qian, Hongkai Xiong, and Qi Tian. Prune spatio-temporal tokens by semantic-aware temporal accumulation. arXiv preprint arXiv:2308.04549, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.354, + 0.892, + 0.436 + ], + "angle": 0, + "content": "[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.439, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[24] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[25] Gamaleldin F Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael C Mozer, and Thomas Kipf. Savi++: Towards end-to-end object-centric learning from real-world videos. arXiv preprint arXiv:2206.07764, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.598, + 0.892, + 0.667 + ], + "angle": 0, + "content": "[26] Patrick Emami, Pan He, Sanjay Ranka, and Anand Rangarajan. Efficient iterative amortized inference for learning symmetric and disentangled multi-object representations. In International Conference on Machine Learning, pages 2970-2981. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.67, + 0.892, + 0.726 + ], + "angle": 0, + "content": "[27] Martin Engelcke, Adam R Kosiorek, Oiwi Parker Jones, and Ingmar Posner. Genesis: Generative scene inference and sampling with object-centric latent representations. arXiv preprint arXiv:1907.13052, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.728, + 0.892, + 0.784 + ], + "angle": 0, + "content": "[28] Martin Engelcke, Oiwi Parker Jones, and Ingmar Posner. Genesis-v2: Inferring unordered object representations without iterative refinement. Advances in Neural Information Processing Systems, 34:8085-8094, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.786, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[29] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3299-3309, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16684" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "[31] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.218 + ], + "angle": 0, + "content": "[32] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthey, Matthew Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative variational inference. In International Conference on Machine Learning, pages 2424-2433. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[33] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.47, + 0.373 + ], + "angle": 0, + "content": "[34] Kai Han, Rafael S Rezende, Bumsub Ham, Kwan-Yee K Wong, Minsu Cho, Cordelia Schmid, and Jean Ponce. Sc net: Learning semantic correspondence. In Proceedings of the IEEE international conference on computer vision, pages 1831-1840, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[35] Tengda Han, Weidi Xie, and Andrew Zisserman. Video representation learning by dense predictive coding. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.469, + 0.487 + ], + "angle": 0, + "content": "[36] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. Advances in Neural Information Processing Systems, 33:5679-5690, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.557 + ], + "angle": 0, + "content": "[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.56, + 0.469, + 0.628 + ], + "angle": 0, + "content": "[38] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.469, + 0.672 + ], + "angle": 0, + "content": "[39] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.674, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[40] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[41] David Held, Sebastian Thrun, and Silvio Savarese. Learning to track at 100 fps with deep regression networks. In European conference on computer vision, pages 749-765. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[42] Olivier Henaff. Data-efficient image recognition with contrastive predictive coding. In International conference on machine learning, pages 4182-4192. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[43] Yingdong Hu, Renhao Wang, Kaifeng Zhang, and Yang Gao. Semantic-aware fine-grained correspondence. arXiv preprint arXiv:2207.10456, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[44] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "in neural information processing systems, 33:19545-19560, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[45] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.892, + 0.222 + ], + "angle": 0, + "content": "[46] Baoxiong Jia, Yu Liu, and Siyuan Huang. Unsupervised object-centric learning with bi-level optimized query slot attention. arXiv preprint arXiv:2210.08990, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.225, + 0.892, + 0.266 + ], + "angle": 0, + "content": "[47] Longlong Jing, Xiaodong Yang, Jingen Liu, and Yingli Tian. Self-supervised spatiotemporal feature learning via video rotation prediction. arXiv preprint arXiv:1811.11387, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.269, + 0.892, + 0.35 + ], + "angle": 0, + "content": "[48] Rishabh Kabra, Daniel Zoran, Goker Erdogan, Loic Matthew, Antonia Creswell, Matt Botvinick, Alexander Lerchner, and Chris Burgess. Simone: View-invariant, temporally-abstracted object representations via unsupervised video decomposition. Advances in Neural Information Processing Systems, 34:20146-20159, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.354, + 0.892, + 0.421 + ], + "angle": 0, + "content": "[49] Dahun Kim, Donghyeon Cho, Donggeun Yoo, and In So Kweon. Learning image representations by completing damaged jigsaw puzzles. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 793-802. IEEE, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.425, + 0.892, + 0.493 + ], + "angle": 0, + "content": "[50] Thomas Kipf, Gamaleldin F Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional object-centric learning from video. arXiv preprint arXiv:2111.12594, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[51] Zihang Lai, Erika Lu, and Weidi Xie. Mast: A memory-augmented self-supervised tracker. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6479-6488, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.555, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[52] Zihang Lai and Weidi Xie. Self-supervised learning for video correspondence flow. arXiv preprint arXiv:1905.00875, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.599, + 0.892, + 0.626 + ], + "angle": 0, + "content": "[53] Hala Lamdouar, Weidi Xie, and Andrew Zisserman. Segmenting invisible moving objects. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.629, + 0.892, + 0.684 + ], + "angle": 0, + "content": "[54] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In Proceedings of the IEEE international conference on computer vision, pages 2192-2199, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.687, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[55] Xueting Li, Sifei Liu, Shalini De Mello, Xiaolong Wang, Jan Kautz, and Ming-Hsuan Yang. Joint-task self-supervised learning for temporal correspondence. Advances in Neural Information Processing Systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[56] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1739-1748, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[57] Runtao Liu, Zhirong Wu, Stella Yu, and Stephen Lin. The emergence of objectness: Learning zero-shot segmentation from videos. Advances in Neural Information Processing Systems, 34:13137-13152, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[58] Songtao Liu, Zeming Li, and Jian Sun. Self-emd: Self-supervised object detection withoutImagenet. arXiv preprint arXiv:2011.13677, 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16685" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[59] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in Neural Information Processing Systems, 33:11525-11538, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.22 + ], + "angle": 0, + "content": "[60] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[61] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.47, + 0.32 + ], + "angle": 0, + "content": "[62] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European conference on computer vision, pages 527-544. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[63] Joe Yue-Hei Ng, Jonghyun Choi, Jan Neumann, and Larry S Davis. Actionflownet: Learning motion representation for action recognition. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1616–1624. IEEE, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.47, + 0.446 + ], + "angle": 0, + "content": "[64] Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. IEEE transactions on pattern analysis and machine intelligence, 36(6):1187-1200, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.47, + 0.49 + ], + "angle": 0, + "content": "[65] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.47, + 0.573 + ], + "angle": 0, + "content": "[66] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.47, + 0.631 + ], + "angle": 0, + "content": "[67] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.47, + 0.689 + ], + "angle": 0, + "content": "[68] Rui Qian, Shuangrui Ding, Xian Liu, and Dahua Lin. Static and dynamic concepts for self-supervised video representation learning. In European Conference on Computer Vision, pages 145-164. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.47, + 0.772 + ], + "angle": 0, + "content": "[69] Rui Qian, Yuxi Li, Huabin Liu, John See, Shuangrui Ding, Xian Liu, Dian Li, and Weiyao Lin. Enhancing self-supervised video representation learning via multi-level feature optimization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7990-8001, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[70] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6964-6974, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[71] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[72] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, et al. Bridging the gap to real-world object-centric learning. arXiv preprint arXiv:2209.14860, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.23 + ], + "angle": 0, + "content": "[73] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8934-8943, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.232, + 0.893, + 0.3 + ], + "angle": 0, + "content": "[74] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.301, + 0.893, + 0.37 + ], + "angle": 0, + "content": "[75] Jack Valmadre, Luca Bertinetto, Joao Henriques, Andrea Vedaldi, and Philip HS Torr. End-to-end representation learning for correlation filter based tracking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2805-2813, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.371, + 0.893, + 0.44 + ], + "angle": 0, + "content": "[76] Carles Ventura, Miriam Bellver, Andreu Girbau, Amaia Salvador, Ferran Marques, and Xavier Giro-i Nieto. Rvos: End-to-end recurrent network for video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5277-5286, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.893, + 0.482 + ], + "angle": 0, + "content": "[77] Paul Voigtlaender and Bastian Leibe. Online adaptation of convolutional neural networks for video object segmentation. arXiv preprint arXiv:1706.09364, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.893, + 0.538 + ], + "angle": 0, + "content": "[78] Carl Vondrick, Abhinav Shrivastava, Alireza Fathi, Sergio Guadarrama, and Kevin Murphy. Tracking emerges by colorizing videos. In Proceedings of the European conference on computer vision (ECCV), pages 391-408, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.893, + 0.58 + ], + "angle": 0, + "content": "[79] Naiyan Wang and Dit-Yan Yeung. Learning a deep compact image representation for visual tracking. Advances in neural information processing systems, 26, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.893, + 0.636 + ], + "angle": 0, + "content": "[80] Xiaolong Wang, Kaiming He, and Abhinav Gupta. Transitive invariance for self-supervised visual representation learning. In Proceedings of the IEEE international conference on computer vision, pages 1329-1338, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.893, + 0.692 + ], + "angle": 0, + "content": "[81] Xiaolong Wang, Allan Jabri, and Alexei A Efros. Learning correspondence from the cycle-consistency of time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2566-2576, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.893, + 0.761 + ], + "angle": 0, + "content": "[82] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.893, + 0.829 + ], + "angle": 0, + "content": "[83] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.893, + 0.872 + ], + "angle": 0, + "content": "[84] Junyu Xie, Weidi Xie, and Andrew Zisserman. Segmenting moving objects via an object-centric layered representation. arXiv preprint arXiv:2207.02206, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[85] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16686" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.469, + 0.134 + ], + "angle": 0, + "content": "consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[86] Jiarui Xu and Xiaolong Wang. Rethinking self-supervised correspondence learning: A video frame-level similarity perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10075-10085, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[87] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen, and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 585-601, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.469, + 0.344 + ], + "angle": 0, + "content": "[88] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7177-7188, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.469, + 0.415 + ], + "angle": 0, + "content": "[89] Yanchao Yang, Antonio Loquercio, Davide Scaramuzza, and Stefano Soatto. Unsupervised moving object detection via contextual information separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 879-888, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.469, + 0.472 + ], + "angle": 0, + "content": "[90] Qixian Zhou, Xiaodan Liang, Ke Gong, and Liang Lin. Adaptive temporal encoding network for video instance-level human parsing. In Proceedings of the 26th ACM international conference on Multimedia, pages 1527-1535, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.474, + 0.469, + 0.529 + ], + "angle": 0, + "content": "[91] Xizhou Zhu, Yujie Wang, Jifeng Dai, Lu Yuan, and Yichen Wei. Flow-guided feature aggregation for video object detection. In Proceedings of the IEEE international conference on computer vision, pages 408-417, 2017." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "16687" + } + ] +] \ No newline at end of file diff --git a/2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/0da0877f-f78d-4f97-9a49-55c50b2292de_origin.pdf b/2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/0da0877f-f78d-4f97-9a49-55c50b2292de_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..064b2b9d723f39e0700678f9f303712bf9e38ad5 --- /dev/null +++ b/2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/0da0877f-f78d-4f97-9a49-55c50b2292de_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d88715185fb924acb4113735c2a10b31bef8875837bc7837f221e9e9d6a23f8 +size 3914233 diff --git a/2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/full.md b/2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..302f7e33f3ab98e5d29899de264807fe2ec3ee64 --- /dev/null +++ b/2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/full.md @@ -0,0 +1,381 @@ +# Semantics Meets Temporal Correspondence: Self-supervised Object-centric Learning in Videos + +Rui Qian1 Shuangrui Ding1 Xian Liu1 Dahua Lin1,2* + +1The Chinese University of Hong Kong 2Shanghai Artificial Intelligence Laboratory + +{qr021,ds023,lx021,dhlin}@ie.cuhk.edu.hk + +# Abstract + +Self-supervised methods have shown remarkable progress in learning high-level semantics and low-level temporal correspondence. Building on these results, we take one step further and explore the possibility of integrating these two features to enhance object-centric representations. Our preliminary experiments indicate that query slot attention can extract different semantic components from the RGB feature map, while random sampling based slot attention can exploit temporal correspondence cues between frames to assist instance identification. Motivated by this, we propose a novel semantic-aware masked slot attention on top of the fused semantic features and correspondence maps. It comprises two slot attention stages with a set of shared learnable Gaussian distributions. In the first stage, we use the mean vectors as slot initialization to decompose potential semantics and generate semantic segmentation masks through iterative attention. In the second stage, for each semantics, we randomly sample slots from the corresponding Gaussian distribution and perform masked feature aggregation within the semantic area to exploit temporal correspondence patterns for instance identification. We adopt semantic- and instance-level temporal consistency as self-supervision to encourage temporally coherent object-centric representations. Our model effectively identifies multiple object instances with semantic structure, reaching promising results on unsupervised video object discovery. Furthermore, we achieve state-of-the-art performance on dense label propagation tasks, demonstrating the potential for object-centric analysis. The code is released at https://github.com/shvdiwnkozbw/SMTC. + +# 1. Introduction + +As one of the fundamental cognitive capabilities, human beings easily distinguish different objects, establish visual + +![](images/06fe90c013fc77671267ef4b2e01a9a4c5c5ec86500384249bf7b59b4041c790.jpg) +(a) Query slot attention on semantic feature. + +![](images/95713c39e96e743b2ba94bccd45808cf70187045eaa76c8e70f3830f7ba15b8e.jpg) + +![](images/f54196344cb58166b76125737f0d095a1fff030f384f1f1f8b67aba06fea8ca8.jpg) +(b) Random sampling slot attention on correspondence map. + +![](images/a9fc3ad0a33a1ce8c341f7a5026395f04b9f055999083bf9257420247028d229.jpg) +Figure 1. Fig. 1(a) presents the results of query slot attention on top of the RGB feature map. It successfully decomposes different semantics, e.g., camels and fence. Fig. 1(b) visualizes the correspondence map after PCA dimensionality reduction, showing that different instances have different correspondence patterns. And the slot attention with random sampling coarsely distinguishes two camels with some redundant borders. Best viewed in color. + +correspondence and perform object-centric analysis from temporally continuous observations. This ability can be attributed to two indispensable visual mechanisms: high-level semantic discrimination as well as low-level temporal correspondence, which enable humans to effectively understand and interact with the world. + +Motivated by this, computer vision researchers equip machines with these capacities to enhance object-centric perception [79, 60, 34]. To achieve this goal, early works rely on human annotations or weak supervision to perceive object semantics [60, 63, 13], identify geometric positions [39, 71, 31, 91], and establish temporal correspondence [41, 79, 75, 77], but their generalization ability is limited. Recently, there emerge a host of fully unsupervised methods to learn robust representations for semantic discrimination [14, 38, 33, 70, 29, 37, 12] or spatiotemporal correspondence [44, 81, 86, 55, 52, 43], which achieve promising performance. Given this encouraging re + +sult, we naturally come up with a question: Is it possible to jointly leverage the semantics and correspondence to discover object instances and distill object-centric representations without human annotations? + +Regarding this problem, our intuition is that the high-level semantics delineates meaningful foreground areas in a top-down manner, while when looking into more frames, the low-level correspondence temporally associates coherent objects and separates individual instances in a bottom-up fashion. For instance, in a football scene, the semantic cue differentiates the foreground that includes several players, while the temporal correspondence links distinct players through dynamic movements and geometric relationships. These two aspects collectively contribute to object-centric representations. Unfortunately, most of the existing works only concentrate on one of these features. [14, 38, 33, 70, 82] succeed in developing high-level semantics, but this abstract semantics alone is insufficient to distinguish instances. Whereas, [44, 81, 78, 55] excel in detailed correspondence, but lack semantic structure and result in redundancy and ambiguity. + +In this paper, we propose a new architecture, Semantics Meets Temporal Correspondence (SMTC), to jointly leverage semantics and temporal correspondence to distill object-centric representations from RGB sequences. Specifically, we first extract frame-wise visual features as the semantic representation. Then we calculate dense feature correlations between adjacent frames as the correspondence map which encodes temporal relationships. To mine the object-centric knowledge, we take inspiration from [59, 88, 50, 25], and investigate using different formulations of slot attention on them. The preliminary experiments show that the original slot attention with random sampling on RGB feature map suffers from complex scene components in real-world videos [59, 72], but the revised query slot attention [88, 46] can decompose different semantic components as shown in Fig. 1(a). As for the correspondence map, different objects present diverse temporal correspondence patterns. Comparing to semantic features, these patterns reveal low-level geometric relationships, which are comparatively simple but vary with specific scenes. Hence, query slot attention fails but the random sampling based formulation performs surprisingly well as shown in Fig. 1(b), coarsely separating different object areas with some redundant borders. + +Motivated by this, we develop semantic-aware masked slot attention, which comprises a set of Gaussian distributions with learnable mean and standard deviation vectors, on top of the fused semantic and correspondence representations. The intuition is that the mean vectors could represent potential semantic centers, which act similarly to the query slot attention to separate semantic components. While the deviation vectors introduce perturbations around + +the semantic centers to capture distinct temporal correspondence patterns of different instances. Technically, we formulate two slot attention stages to achieve this goal. Firstly, we use the mean vectors as slot initialization to generate semantic segmentation masks. Secondly, for each semantics, we randomly sample slot vectors from the Gaussian distribution, then perform iterative attention and masked aggregation within the corresponding semantic mask area to distinguish instances. We enforce temporal consistency on the semantic masks as well as object instance slots to enhance temporal coherency and refine object-centric representations. Comparing with existing works on object-centric learning in videos [25, 50, 88, 84], our model is free of pre-computed motion or depth prior, and explicitly identifies multiple objects with semantic structure. + +In summary, our contributions are: (1) We propose a novel self-supervised architecture that unifies semantic discrimination and temporal correspondence to distill object-centric representations in videos. (2) We demonstrate that simple feature correlation can effectively represent temporal correspondence cues when used in conjunction with semantic features. Building on this observation, we develop semantic-aware masked slot attention, which operates on fused visual features and correspondence maps, to distinguish multiple object instances with semantic structure without relying on motion or depth priors. (3) We achieve promising results on unsupervised object discovery in both single and multiple object scenarios, and reach state-of-the-art performance on label propagation tasks, demonstrating that we learn discriminative and temporally consistent object-centric representations. + +# 2. Related Work + +Unsupervised Object Discovery is an essential process to formulate object-centric representations, which aims to identify objects without human annotations. There exist a series of works focusing on this problem [59, 32, 26, 27, 28, 7]. Typically, [59] develops slot attention to iteratively update latent object representations from random initializations but it has difficulty scaling to complex real-world scenes. To tackle this challenge, [72] employs feature reconstruction as objective to reduce redundancy, [46] develops query slot initialization to encode visual concepts. Further, a line of works extend object discovery to video domain [50, 48, 17, 5, 88]. Most of them build on slot attention architecture, and adopt extra pre-computed priors, e.g., optical flow [88, 84, 16, 21], depth [25], geometric positions [50], as input or supervision to assist object discovery. In our work, we find that the simple feature correlation provides informative temporal cues. And we develop semantic-aware masked slot attention to identify object instances with semantic structure without resorting to extra priors. + +Self-supervised Representation Learning aims to learn + +![](images/fe93decb9a5c6db629d51dcb51e2632af2e45a5db995b6fabaad9a7b39399dd8.jpg) +Figure 2. An overview of our framework. We first extract frame-wise features and calculate dense feature correlation, then fuse them to pass through semantic-aware masked slot attention, which comprises two slot attention stages with $N$ shared learnable Gaussian distributions. In the first semantic slot attention stage, the mean vectors serve as slot initialization to generate a set of segmentation masks for semantic decomposition. In the second masked slot attention stage, which runs on $N$ semantics in parallel, we randomly sample slots from the Gaussian distribution of each semantics and perform masked feature aggregation within the semantic area to identify distinct instances. We enforce semantic and instance temporal consistency to train the architecture in a teacher-student manner, with the teacher marked in gray. + +robust representation without human annotations. Early works design various pretext tasks to generate pseudo labels as self-supervision [49, 30, 47, 10, 1, 62]. Later, contrastive learning with instance discrimination revolutionized the field [38, 14, 70, 29, 35, 83, 65, 42, 11, 20, 3, 68, 35, 36, 69, 19, 22]. These works generally contribute to general semantics but lack the ability to distill object-centric representations. To delve into this problem, [82, 85] preserve spatial sensitivity and perform contrastive learning on dense features. [86, 12] point out that semantic pre-training implicitly encodes object correspondence. While in our work, we learn to explicitly identify different objects with instance-level consistency to progressively refine object-centric representations. + +Self-supervised Correspondence is a fundamental problem in computer vision [55, 78, 81, 80]. Typical works rely on low-level statistics like color to generate the dense correspondence flow [51, 52, 78], or establish temporal cycles to track image regions [81, 44, 6]. Further, [43, 86, 12] incorporate correspondence learning with semantic discrimination, but they either regard correspondence as a side effect of semantic discrimination [86, 12] or independently learn two pathway features with late fusion [43]. Among these works, high-level semantics is not well utilized, and there exists high redundancy due to large background areas. In contrast, our method explicitly separates semantic components to alleviate redundancy and formulate instance-level correspondence from an object-centric perspective. + +# 3. Method + +Our framework is shown in Fig. 2, which adopts a teacher-student structure similar to recent self-supervised methods [38, 12, 33]. Specifically, we first extract framewise features from an RGB sequence, and calculate dense feature correlation as temporal correspondence representations. After that, we fuse them to pass through semantic-aware masked slot attention to separate multiple object instances with semantic structure, and enforce temporal consistency to refine the object-centric representations. + +# 3.1. Feature Encoding + +Given a RGB sequence $v = \{I_1, I_2, \dots, I_T\}$ , we employ a shared visual encoder $f$ to extract frame-wise features: + +$$ +F _ {t} = f \left(I _ {t}\right) \in \mathbb {R} ^ {H W \times D}, \tag {1} +$$ + +where $H, W, D$ respectively denotes height, width and channel dimension, $t$ indicates frame index. For simplicity, we omit the subscript indicating student or teacher pathway, and use $\hat{F}$ to denote teacher outputs. Intuitively, $F_{t}$ contains appearance and semantic cues in the input sequence, and the next step is to formulate an effective representation for the temporal correspondence cues. To do this, inspired by the cost volume in motion estimation [24, 73, 74], we calculate dense feature correlation to generate the temporal + +
FeatureQueryRandom
RGB58.138.4
Correlation21.556.3
+ +Table 1. Preliminary results on RGB feature map and feature correlation. 'Query' ('Random') denotes slot attention with query (random sampling) initialization. We report IoU on DAVIS-2016. + +correspondence map for each timestamp: + +$$ +C _ {t j} = F _ {t} F _ {j} ^ {T} \in \mathbb {R} ^ {H W \times H W}, \tag {2} +$$ + +where we randomly sample one frame index $j \neq t$ to compute feature correlation. The channel activations of $C_{tj}$ encode low-level geometric correlations between frame $t$ and $j$ , which indicates potential partitions of different objects as illustrated in Fig. 1(b). + +To jointly leverage the semantic and correspondence cues, we employ two linear transformation heads, $h_f$ and $h_c$ , to respectively project $F_t$ and $C_{tj}$ to a shared $D$ -dimensional embedding space, which are then fused with element-wise summation: + +$$ +R _ {t} = h _ {f} \left(F _ {t}\right) + h _ {c} \left(C _ {t j}\right) \in \mathbb {R} ^ {H W \times D}. \tag {3} +$$ + +In this way, $R_{t}$ simultaneously encodes semantic and temporal correspondence information, serving as an intermediate feature for object-centric analysis. + +# 3.2. Preliminary Experiments with Slot Attention + +Given these representations, the next step is to formulate effective ways to identify individual objects. To do this, we take inspiration from slot attention [59, 50, 88, 46], which employs a set of slot vectors to iteratively attend to specific objects. Generally, there are two alternatives for slot initialization, random sampling from a Gaussian distribution with learnable parameters $\mu, \sigma \in \mathbb{R}^D$ [59, 50] or directly inheriting from a set of learnable queries [46, 88]. + +To investigate the effectiveness of these two formulations, we conduct preliminary experiments on the semantic feature $F_{t}$ and correspondence map $C_{tj}$ respectively. We follow the settings in [72], using a frozen DINO pre-trained ViT to extract visual feature and generate correspondence map, on top of which we perform two kinds of slot attention. We train and evaluate them on DAVIS-2016 dataset [66]. As indicated by Fig. 1 and Table. 1, we find that query slot attention decomposes distinct semantic components on $F_{t}$ , while the original slot attention with random sampling performs much worse and struggles in complex real-world scenes. Conversely, query slot attention does not work on the correspondence map, and random sampling initialization enables the model to effectively utilize the temporal correspondence patterns in $C_{tj}$ to distinguish instances. An intuitive explanation into these observations is the RGB + +features contain rich semantics and there exist consistent patterns for similar objects. Slot attention with learnable queries can capture inter-sample semantic patterns to identify objects. While random initialization fails to memorize semantic patterns, thus resulting in worse performance. In contrast, correspondence features reveal inter-frame geometric correlations, shown in Fig. 1(b), which vary with specific scenes. There are no consistent patterns for similar objects, so the learnable queries fail but random initialization works well to capture correspondence cues. The complementarity between these two conditions motivates two-stage slot attention design to combine the attributes of query and random sampling initialization, thus better exploiting the semantic and correspondence cues in $R_{t}$ . + +# 3.3. Semantic-aware Masked Slot Attention + +We propose semantic-aware masked slot attention, which comprises $N$ Gaussian distributions with learnable means $\pmb{\mu} = \{\mu_1,\mu_2,\dots,\mu_N\} \in \mathbb{R}^{N\times D}$ and standard deviations $\sigma = \{\sigma_{1},\sigma_{2},\dots,\sigma_{N}\} \in \mathbb{R}^{N\times D}$ . With this design, the mean vectors act similarly to the learnable queries, each representing a potential semantic center. And the deviations introduce perturbations around semantic centers to model temporal correspondence patterns. To achieve this goal, we employ two stages of iterative attention with shared learnable parameters. First, we use the mean vectors as slot initialization to decompose different semantics in the form of segmentation. Second, we randomly sample slots from each Gaussian distribution $(\mu_n,\mathrm{diag}(\sigma_n))$ , and perform masked aggregation within the corresponding semantic area to identify instances. The process is detailed as follows. + +Semantic Decomposition. In the first stage, we directly use $\pmb{\mu}$ without randomness as slot initialization $S = \pmb{\mu} \in \mathbb{R}^{N \times D}$ . Then following the standard slot attention procedure [59], we employ three linear transformation heads to map $S$ and $R_{t}$ into $Q \in \mathbb{R}^{N \times D}, K, V \in \mathbb{R}^{HW \times D}$ , and iteratively calculate attention score and update slot representations. Mathematically, we formulate each iteration as: + +$$ +M := \operatorname {s o f t m a x} \left(\frac {1}{\sqrt {D}} Q K ^ {T}, 0\right) \in \mathbb {R} ^ {N \times H W}, \tag {4} +$$ + +$$ +A [ n, i ] := \frac {M [ n , i ]}{\sum_ {l} M [ n , l ] + \epsilon}, A \in \mathbb {R} ^ {N \times H W}, \tag {5} +$$ + +$$ +S := \operatorname {G R U} (\text {i n p u t} = A V, \text {s t a t e s} = S). \tag {6} +$$ + +Note that the slot attention weight $M$ is normalized along the query axis, and the weighted mean coefficient $A$ is employed to aggregate the value to update the slots. This mechanism induces competition among slots and enforce different slots to take over different semantic features. We iterate this process for three times, and take the slot attention weight (slot vectors) of the final iteration as the segmentation mask (representation) for different semantic compo + +nents. To avoid ambiguity in presentation, we denote them as $\mathcal{M} \in \mathbb{R}^{N \times HW}$ and $S \in \mathbb{R}^{N \times D}$ , where $S[n]$ presents $n$ -th semantic center representation, $\mathcal{M}[n]$ indicates the probability that each pixel belongs to $n$ -th semantics. + +Instance Identification. Within each semantics, there could exist multiple object instances. The Gaussian distributions we introduce help to distinguish them with temporal correspondence cues. Specifically, in the second stage, for each semantics, we randomly sample slots from the Gaussian distribution, using the perturbations around the semantic center to capture distinct temporal correspondence patterns. For simplicity, we take the $n$ -th semantics as an example. First, we sample $P$ vectors, $S \sim \mathcal{N}(\mu_n, \mathrm{diag}(\sigma_n)) \in \mathbb{R}^{P \times D}$ , as slot initialization to represent $P$ potential instances. Next, we follow the first stage formulation to respectively project $S$ and $R_t$ into $Q \in \mathbb{R}^{P \times D}$ and $K, V \in \mathbb{R}^{HW \times D}$ , and obtain the attention score $M \in \mathbb{R}^{P \times HW}$ . To preserve the semantic attributes and discriminate correspondence cues, we perform masked feature aggregation within the semantic mask areas. The only mathematical difference in this process is the weighted mean coefficient computation: + +$$ +\begin{array}{l} A [ p, i ] := \frac {M [ p , i ] \widetilde {\mathcal {M}} [ n , i ]}{\sum_ {l} M [ p , l ] \widetilde {\mathcal {M}} [ n , l ] + \epsilon}, A \in \mathbb {R} ^ {P \times H W}, \tag {7} \\ \widetilde {\mathcal {M}} [ n, i ] = \operatorname {b i n a r i z e} (\mathcal {M} [ n, i ], \tau), \mathcal {M} \in \mathbb {R} ^ {N \times H W}, \\ \end{array} +$$ + +where we use the $n$ -th semantic mask $\widetilde{\mathcal{M}}[n] \in \{0,1\}^{HW}$ binarized with threshold $\tau$ to constrain the model to only aggregate the visual features related to $n$ -th semantics. And we empirically find a comparatively large threshold $\tau = 0.5$ helps to preserve object regions with clear semantics and filter out background. The masked slot attention for all semantics follows the same rule and is calculated in parallel. Similarly, we iterate the process for three times, use the final slot attention weight as the instance segmentation masks, and take the final slot vectors as the object-centric representations $\mathcal{O} \in \mathbb{R}^{N \times P \times D}$ , where $\mathcal{O}[n,p] \in \mathbb{R}^D$ denotes the $p$ -th potential object instance of $n$ -th semantics. In this way, our semantic-aware masked slot attention explicitly segments multiple instances with semantic structure. + +# 3.4. Training + +We use temporal consistency as self-supervision to optimize semantic masks $\mathcal{M}$ and instance representations $\mathcal{O}$ . + +Dense Semantic Alignment. Given the semantic mask of each timestamp $\mathcal{M}_t\in \mathbb{R}^{N\times HW}$ , as well as $\hat{\mathcal{M}}_t$ from teacher pathway, we aim to align the spatially dense semantic distributions across time. To achieve this, the first step is to determine the corresponding patches between different frames. Utilizing the feature correlation $C_{tj}\in \mathbb{R}^{HW\times HW}$ which indicates dense feature similarity between timestamps $t,j$ , we can infer patch-level correspondence from + +this cue with optimal transport [11, 2, 58, 18]. Formally, we take $-\hat{C}_{tj}$ from teacher pathway as cost matrix, and solve the optimal transport strategy $\pi_{tj}^{*} \in \mathbb{R}^{HW \times HW}$ between two uniform distributions to indicate patch correspondence: + +$$ +\begin{array}{l} \min _ {\pi_ {t j}} \sum_ {u = 1} ^ {H W} \sum_ {v = 1} ^ {H W} - \hat {C} _ {t j} [ u, v ] \pi_ {t j} [ u, v ] \\ \text {s . t .} \sum_ {v = 1} ^ {H W} \pi_ {t j} [ \cdot , v ] = \frac {1}{H W} \mathbf {1} ^ {H W}, \tag {8} \\ \sum_ {u = 1} ^ {H W} \pi_ {t j} [ u, \cdot ] = \frac {1}{H W} \mathbf {1} ^ {H W}, \\ \pi_ {t j} [ u, v ] \geq 0 \quad u, v \in \{1, 2, \dots , H W \}, \\ \end{array} +$$ + +where $u, v$ denotes spatial index, $\mathbf{1}^{HW}$ is a $HW$ -dimensional vector of all ones. Note that it is feasible to adopt other marginal distribution formulations with prior knowledge, e.g., class-agnostic activation map [4], to facilitate training, with details discussed in Supplementary Material. We employ Sinkhorn-Knopp algorithm [18] to obtain the optimal transport matrix $\pi_{t,j}^*$ , and formulate the dense semantic alignment objective in the form of cross entropy: + +$$ +\mathcal {L} _ {s e m} = - \sum_ {\substack {t, j = 1 \\ t \neq j}} ^ {T} \sum_ {u, v = 1} ^ {H W} \sum_ {n = 1} ^ {N} \pi_ {t j} ^ {*} [ u, v ] \hat {\mathcal {M}} _ {j} [ n, v ] \log \mathcal {M} _ {t} [ n, u ]. \tag{9} +$$ + +By minimizing the weighted sum of cross entropy, we achieve temporally consistent semantic distributions among corresponding spatial areas. + +Semantic Mask Regularization. To encourage the semantic centers to emphasize different visual contents and cover diverse semantics, we apply a simple regularization to $\mathcal{M}_t$ : + +$$ +\mathcal {L} _ {r e g} = \sum_ {t = 1} ^ {T} \sum_ {\substack {n, j = 1 \\ j \neq n}} ^ {N} \frac {\mathcal {M} _ {t} [ n ] ^ {T} \mathcal {M} _ {t} [ j ]}{\| \mathcal {M} _ {t} [ n ] \| _ {2} \| \mathcal {M} _ {t} [ j ] \| _ {2}}. \tag{10} +$$ + +This simple regularization suppresses the cosine similarity between different semantic masks to avoid collapse. + +Instance Representation Consistency. Apart from the semantic distributions, it is necessary to ensure that the representations of each object instance are temporally consistent. Given the instance representations $\mathcal{O}_t\in \mathbb{R}^{N\times P\times HW}$ , as well as $\hat{\mathcal{O}}_t$ from teacher pathway, we need to first match corresponding instances between different timestamps. For illustration, considering $n$ -th semantics of time $t,j$ , we adopt bipartite matching [9, 15] based on the cosine similarity between $\mathcal{O}_t[n]$ and $\hat{\mathcal{O}}_j[n]$ to generate one-to-one instance correspondence, with the matching function denoted as $\varepsilon (\cdot)$ . + +In this way, $\mathcal{O}_t[n,p]$ and $\hat{\mathcal{O}}_j[n,\varepsilon(p)]$ are considered as a positive pair to be aligned. However, for each video, there exists absent semantics and object occlusion such that the number of visible object instances varies. Thus, there could be redundant slots not attending to objects, and it is crucial to filter out these invalid instance representations to reduce distractions. Mathematically, for timestamp $t$ , we introduce $\mathcal{I}_t \in \{0,1\}^{N \times P}$ as the valid instance indicator. For the $p$ -th instance of $n$ -th semantics to be valid, two criteria must be met: (1) The ratio of the $n$ -th semantic area is above threshold $\tau_1 = 0.2$ to filter out non-existing semantics; (2) The instance representation is close to the semantic center representation with cosine similarity larger than $\tau_2 = 0.5$ to exclude redundant slots. The conditions are formulated as: + +$$ +\mathcal {I} _ {t} [ n, p ] = 1 \Leftrightarrow \left\{ \begin{array}{l} \frac {1}{H W} \sum_ {u = 1} ^ {H W} \widetilde {\mathcal {M}} _ {t} [ n, u ] \geq \tau_ {1}, \\ \cos \left(\mathcal {S} _ {t} [ n ], \mathcal {O} _ {t} [ n, p ]\right) \geq \tau_ {2}. \end{array} \right. \tag {11} +$$ + +And $\hat{\mathcal{L}}_j\in \{0,1\}^{N\times P}$ is defined in the same manner. We use a margin loss to encourage object detail consistency over valid instance representations: + +$$ +\begin{array}{l} \mathcal{L}_{obj} = \sum_{\substack{t,j = 1\\ t\neq j}}^{T}\sum_{n = 1}^{N}\sum_{p = 1}^{P}\mathcal{I}_{t}[n,p]\big\{ \\ \hat {\mathcal {I}} _ {j} [ n, \varepsilon (p) ] \| \mathcal {O} _ {t} [ n, p ] - \hat {\mathcal {O}} _ {j} [ n, \varepsilon (p) ] \| _ {2} \tag {12} \\ + \sum_{\substack{q = 1\\ q\neq \varepsilon (p)}}^{P}\mathrm{relu}(\lambda -\| \mathcal{O}_{t}[n,p] - \hat{\mathcal{O}}_{j}[n,q]\|_{2})\} \\ \end{array} +$$ + +where $\lambda$ is a margin hyper-parameter. Note that with the $\mathcal{I}_t$ and $\hat{\mathcal{I}}_j$ constraints, our formulation can handle object occlusions which lead to varying number of visible instances in different frames. By minimizing the margin loss, we encourage the model to distill discriminative and temporally consistent object-centric representations. + +Overall, we take the summation of three objectives for training: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {s e m}} + \mathcal {L} _ {\text {o b j}} + \mathcal {L} _ {\text {r e g}}. \tag {13} +$$ + +The student pathway $\theta$ is updated with gradient descent, and the teacher parameters $\hat{\theta}$ are updated with momentum as: + +$$ +\hat {\theta} \leftarrow m \hat {\theta} + (1 - m) \theta , \tag {14} +$$ + +where $m$ is momentum coefficient set to 0.999 in default. This momentum update mechanism results in a slowly evolving teacher network, which progressively distills object knowledge, provides reliable self-supervision signals and enables us to train semantic-aware masked slot attention without relying on widely adopted reconstruction objective in existing works [59, 50, 88, 84, 72, 46]. + +# 4. Experiments + +# 4.1. Dataset + +We train our model on YouTube-VOS [87], a challenging video dataset that contains multiple object instances of distinct semantics in each video. We evaluate our method on two lines of tasks: (1) Unsupervised object segmentation on DAVIS-2016 [66], SegTrack-v2 [54], FMBS-59 [64] and challenging multiple object segmentation on DAVIS-2017-Unsupervised [8]. We respectively calculate the mean per frame intersection over union (IoU) and $\mathcal{J}\&\mathcal{F}$ on single and multiple object discovery benchmarks. (2) Lable propagation tasks including semi-supervised video object segmentation on DAVIS-2017 [67], human pose tracking on JHMDB [45], human part tracking on VIP [90]. We adopt the same settings as [44, 43] and report standard $\mathcal{J}$ and $\mathcal{F}$ score on DAVIS, probability of a correct pose (PCK) on JHMDB and mean intersection over union (mIoU) on VIP. + +# 4.2. Implementation Details + +We sample $T = 4$ frames with frame rate FPS $= 4$ as the input RGB sequence, where each frame is augmented with random crop, horizontal flip and color jitter, and finally resized to $256 \times 256$ . We adopt ViT-Small/16 [23] and ResNet-50 [40] as the visual encoder, specified in each experiment, to extract frame-wise features. Then for semantic-aware masked slot attention, we set the number of learnable Gaussian distributions to $N = 16$ , the number of instances of each semantics to $P = 4$ in default and use 3 iterations to update slot representations and attention maps. + +In training, the visual encoder is initialized with self-supervised pre-trained weights from DINO ViT-Small/16 [12] or MoCo ResNet-50 [38]. We use AdamW optimizer [61] with learning rate $2 \times 10^{-4}$ for batch size 128 to update student parameters, and the teacher pathway is updated in momentum. In inference, we employ the same evaluation protocol as [44] to evaluate the temporal object correspondence performance on label propagation tasks. And for object discovery, we maintain the valid instances filtered with Eq. 11 as candidate objects. For single object evaluation, we merge all candidate objects as foreground areas to calculate IoU. For multiple object benchmark, we follow [8] to match ground-truth and our predictions and report $\mathcal{J} \& \mathcal{F}$ score. + +# 4.3. Comparison with State-of-the-art + +Single Object Discovery. To measure our model's ability to decompose different objects, we first present the quantitative results on single object discovery without post processing (e.g., CRF, spectral clustering) in Table 2. Many existing works resort to optical flow [89, 53, 88, 84] or synthetic data [84] as weak supervision to learn temporal dynamics and generate semantic-agnostic segmentation masks for + +
ModelRGBFlowDAVISST-v2FBMS
CIS [89]✓✓71.562.563.5
AMD [57]✓X57.857.047.5
DINO [12]✓X52.346.550.3
SIMO [53]X✓67.862.0-
MG [88]X✓68.358.653.1
OCLR [84]X✓72.167.665.4
GWM [16]✓✓71.269.066.9
SMTC✓X71.869.368.4
\( SMTC^† \)✓X70.868.466.5
+ +Table 2. Quantitative results on single object discovery. We compare per frame mean IoU on DAVIS-2016, SegTrack-v2 and FBMS-59 without any post-processing. $\mathsf{SMTC}^{\dagger}$ denotes only with first slot attention stage for semantic decomposition in inference. + +
ModelBackboneJ&FJF
DINOSAUR [72]ViT-S/1621.419.223.7
SMTCViT-S/1640.536.444.6
SMTCResNet-5039.035.542.6
RVOS [76]ResNet-10141.236.845.7
ProReduce [56]ResNet-10168.365.071.6
+ +Table 3. Quantitative results on multiple object discovery on DAVIS-2017-Unsupervised. Gray denotes supervised training. + +moving objects. DINO baseline [12] uses the self-attention map from the last layer as an object segmentation mask. While our method only relies on RGB frames and explicitly discriminates object semantics in a fully self-supervised manner. From the comparison, we observe that our SMTC with only semantic decomposition largely exceeds RGB-only methods [12, 57] and comparable to optical flow-based methods, and our formulation with instance identification further improves the performance. This phenomenon indicates that the learned semantic decomposition with temporal correspondence cues have the potential to substitute precomputed optical flow to guide single object segmentation. Multiple Object Discovery. Most of the existing video segmentation methods with fully unsupervised training cannot explicitly distinguish multiple object instances. OCLR [84] adopts layered flow representations to identify multiple objects on re-annotated DAVIS-2017-motion dataset, but it cannot distinguish instances with common motion. On the contrary, our method jointly utilizes semantics and temporal correspondence to separate arbitrary object instances on DAVIS-2017-Unsupervised benchmark. For fair comparison, we re-run DINOSAUR [72] on DAVIS-2017 with 5 slots as an unsupervised baseline. As presented in Table 3, our method significantly outperforms DINOSAUR. This is because our method explicitly decomposes different semantics, exploits temporal correspondence among multiple frames to distinguish instances and encourages instance + +level temporal consistency. While DINOSAUR is only supervised with DINO feature reconstruction without considering temporal relationships and semantic discrimination. Besides, our method is comparable with supervised baseline RVOS [76], and the gap to supervised state-of-the-art [56] is majorly due to human annotated masks as supervision leading to much more precise segmentation boundaries. This conjecture can be demonstrated by Fig. 3, which also reveals that our model is able to learn discriminative object-centric representations to separate different instances. + +Label Propagation. Finally, we compare the performance on label propagation tasks in Table 4 to validate the temporal consistency of our learned features. Though not specifically designed for these tasks, our model achieves state-of-the-art results on three benchmarks. The most related work to our SMTC is SFC [43], which also utilizes both high-level semantic and low-level correspondence. The difference is that SFC only employs late fusion to incorporate these two cues in inference, while our model jointly exploits semantics and temporal correspondence in training to identify object instances, perform instance-level alignment and refine temporally consistent object-centric representations. + +# 4.4. Ablation Study + +We report IoU on DAVIS-2016, and $\mathcal{J}\&\mathcal{F}$ on DAVIS-2017-Unsupervised for single and multiple object discovery. Refer to Supplementary Material for more ablations. + +Slot Attention Formulation. We compare the different formulations of semantic-aware masked slot attention in Table 5. With only the first slot attention stage, i.e., the 'Semantic' part, the model cannot distinguish object instances of the same semantics, thus the performance on multiple object discovery drops significantly as indicated by Ours-B. While with only the second stage, i.e., the 'Instance' part, the performance on both single and multiple object segmentation drops as indicated by Ours-C. This reveals that the pre-computed semantic masks play an important role in alleviating interference in instance slot update and filtering valid sample for instance-level alignment. + +Number of Learnable Gaussian Distributions. Comparing Ours-A, Ours-E and Ours-F in Table 5, we observe that the performance improves when $N$ increases since larger $N$ guides our model to distill more fine-grained semantics. And we present an extreme setting with $N = 1$ in Ours-D, where our formulation degenerates into the original slot attention with random sampling from one Gaussian distribution [59]. Note that although Ours-C also has no access to semantic masks, it still maintains the potential to discriminate different semantics due to multiple learnable means. Its superiority to Ours-D demonstrates the effectiveness of semantic discrimination in object discovery. + +Feature Usage. We explore using different feature input to our semantic-aware masked slot attention in Table 6. We + +
ModelBackboneDAVISJHMDBVIP
J&FJFPCK@0.1PCK@0.2mIoU
Supervised [40]ResNet-5066.063.768.459.278.339.5
MoCo [38]ResNet-5065.463.267.660.479.336.1
VFS [86]ResNet-5068.966.571.360.980.743.2
DINO [12]ViT-S/1661.860.263.445.475.237.9
TimeCycle [81]ResNet-5040.741.939.457.778.528.9
UVC [55]ResNet-5056.354.558.156.576.634.2
MAST [51]ResNet-1865.563.367.6---
CRW [44]ResNet-1867.664.870.258.880.337.6
SFC [43]ResNet-18+ResNet-5071.268.374.061.983.038.4
SMTCViT-S/1667.664.171.253.279.639.2
SMTCResNet-5073.069.476.662.584.138.8
+ +Table 4. Quantitative results on label propagation tasks: semi-supervised video object segmentation on DAVIS-2017, pose tracking on JHMDB, human part tracking on VIP. We present the backbone for comprehensive comparison, and report comparing results from [86, 43]. + +
ModelNSemanticMaskedIoUJ&F
Ours-A16✓✓71.840.5
Ours-B16✓X66.424.5
Ours-C16X✓52.130.3
Ours-D1-✓47.422.1
Ours-E8✓✓68.437.9
Ours-F32✓✓71.940.7
+ +Table 5. Ablation studies on the number of learnable Gaussian distributions, and the formulation of semantic-aware masked slot attention. 'Semantic' ('Instance') denotes the first (second) attention stage for semantic decomposition (instance identification). + +
FeatureSemanticInstanceIoUJ&F
Fused✓✓71.840.5
RGB-only✓✓67.720.1
✓×67.518.8
×✓41.515.3
Correlation-only✓✓38.814.2
✓×21.110.7
×✓59.427.5
+ +Table 6. Ablation studies on the feature usage. We compare different slot attention formulations on different feature inputs. + +observe that with RGB-only input, the 'Semantic' part guarantees satisfactory performance on single object discovery, but performs poor on multiple object benchmarks even with the 'Instance' part. This coincides with our intuition that it is difficult to distinguish object instances with solely semantic cues, instead we need to look into more frames and resort to temporal correspondence. While for the correlation-only input, the 'Instance' part coarsely segments objects and + +
LsemLregLobjIoUJ&F
✓XX61.321.6
✓✓X66.424.5
XX✓50.630.4
✓✓✓71.740.5
+ +Table 7. Ablation studies on the learning objectives. We report the results on DAVIS-2016 and DAVIS-2017-Unsupervised. + +
TFPSIoUJ&F
2465.436.9
4266.337.7
4471.840.5
8471.940.8
+ +Table 8. Ablation studies on the hyper-parameters of frame sampling. We compare different number of frames per clip and different FPS. + +the 'Semantic' part exerts negative impact. This is because the correspondence map contains little semantic information, the semantic decomposition results in severe ambiguity, thus impairing the performance. + +Learning Objectives. We also present a comprehensive ablation on the learning objectives in Table 7, where we report single object discovery on DAVIS-2016 and multiple object discovery results on DAVIS-2017-Unsupervised. From the comparison, we observe that $\mathcal{L}_{sem}$ is fundamental to decompose object semantics, which demonstrates the necessity of semantic decomposition in the first stage. $\mathcal{L}_{reg}$ also leads to improvement by encouraging the learnable queries to cover more diverse semantics. As for $\mathcal{L}_{obj}$ , this objective introduces instance-level alignment and significantly enhances multiple object discovery performance on DAVIS-2017-Unsupervised. + +Frame Sampling. Another important thing in training is the frame sampling procedure. The number of frames per + +![](images/321920c59f87079ae3354195f90e07dba2a40c7ad71b4348a4a4c2136517e82f.jpg) +Figure 3. Visualization of semantic and instance segmentation map. The red boxes outline the ambiguous areas. + +![](images/5c852b66a77f8f4dba7488c70f5cf5273325a7f16a637a4829c96daa9cffb1c0.jpg) +Figure 4. Visualization of instance alignment. The arrows point out the matched instances across time. + +clip and sampling frame rate determines the temporal reception field of our model. We compare different sampling hyper-parameters in Table 8. The results indicate that it is necessary to use large FPS to provide rich temporal dynamics. And by setting comparatively large FPS, our model can have access to more temporal information without introducing more frames, reaching a trade-off between performance and efficiency. + +# 4.5. Further Discussion + +Visualization of Object Discovery. We visualize the our semantic decomposition map and instance identification map in Fig. 3, where the same color denotes the segmentation mask of the same semantics. The visualization reveals that our method is able to distinguish multiple object instances with semantic structure. For example, different people belong to the same semantic center, and quadrupeds such as dogs and pigs belong to another semantics. And our model also separates different instances of the same semantics, such as multiple persons and different pigs, with minor ambiguity on small object part, e.g., pig legs, mobile phones. Besides, we also visualize our instance-level alignment during training in Fig. 4. Our bipartite matching well aligns the corresponding instances in different frames, and the valid sample constraint of Eq. 11 effectively handles object occlusion in videos. For example, when encountering an occluded person instance, our method filters out the redundant slots not attending to objects, only correlating valid slots for robust instance-level alignment. + +Limitation. Due to the lack of the human annotated segmentation masks, our model faces challenges in generating precise boundaries for each instance, particularly for small objects. One potential solution is to incorporate multi-scale feature pyramid to further improve dense perception. We leave it in the future work. Despite this limitation, our work effectively demonstrates the benefits of leveraging both semantics and temporal correspondence to discover object instances with semantic structure and distill discriminative + +and temporally consistent object-centric representations. + +# 5. Conclusion + +In this work, we propose a novel self-supervised framework jointly exploiting high-level semantics and low-level temporal correspondence to enhance object-centric perception. Specifically, we represent semantic and temporal correspondence cues using the RGB feature map and dense feature correlation, respectively. These cues are fused and fed into semantic-aware masked slot attention which comprises a set of learnable Gaussian distributions. This design allows us to leverage the mean vectors as potential semantic centers for semantic decomposition, and use the perturbations introduced by the standard deviation vectors around the semantic centers to make use of the temporal correspondence cues for instance identification. To distill discriminative and temporally consistent object-centric representations, we devise semantic- and instance-level alignment that is robust to object occlusion as self-supervision. We demonstrate the effectiveness of our model for object-centric analysis through the state-of-the-art performance on label propagation tasks, as well as the promising results on unsupervised object discovery in both single and multiple object scenarios. + +# Acknowledgement + +We thank Tianfan Xue for constructive feedback. This project is funded in part by Shanghai AI Laboratory (P23KS00020, 2022ZD0160201), CUHK Interdisciplinary AI Research Institute, and the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK. + +# References + +[1] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. Advances in Neural Information Processing Systems, 33:9758-9770, 2020. +[2] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. arXiv preprint arXiv:1911.05371, 2019. +[3] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019. +[4] Kyungjune Baek, Minhyun Lee, and Hyunjung Shim. Psynet: Self-supervised approach to object localization using point symmetric transformation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10451-10459, 2020. +[5] Beril Besbinar and Pascal Frossard. Self-supervision by prediction for object discovery in videos. In 2021 IEEE Interna + +tional Conference on Image Processing (ICIP), pages 1509-1513. IEEE, 2021. +[6] Zhangxing Bian, Allan Jabri, Alexei A Efros, and Andrew Owens. Learning pixel trajectories with multiscale contrastive random walks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6508-6519, 2022. +[7] Christopher P Burgess, Loic Matthey, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matt Botvinick, and Alexander Lerchner. Monet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390, 2019. +[8] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv preprint arXiv:1905.00737, 2019. +[9] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. +[10] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. +[11] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. +[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. +[13] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. +[14] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. +[15] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. +[16] Subhabrata Choudhury, Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Guess what moves: unsupervised video and image segmentation by anticipating motion. arXiv preprint arXiv:2205.07844, 2022. +[17] Eric Crawford and Joelle Pineau. Exploiting spatial invariance for scalable unsupervised object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 3684-3692, 2020. +[18] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. + +[19] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9716-9726, 2022. +[20] Shuangrui Ding, Rui Qian, and Hongkai Xiong. Dual contrastive learning for spatio-temporal representation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5649-5658, 2022. +[21] Shuangrui Ding, Weidi Xie, Yabo Chen, Rui Qian, Xiaopeng Zhang, Hongkai Xiong, and Qi Tian. Motion-inductive self-supervised object discovery in videos. arXiv preprint arXiv:2210.00221, 2022. +[22] Shuangrui Ding, Peisen Zhao, Xiaopeng Zhang, Rui Qian, Hongkai Xiong, and Qi Tian. Prune spatio-temporal tokens by semantic-aware temporal accumulation. arXiv preprint arXiv:2308.04549, 2023. +[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. +[24] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015. +[25] Gamaleldin F Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael C Mozer, and Thomas Kipf. Savi++: Towards end-to-end object-centric learning from real-world videos. arXiv preprint arXiv:2206.07764, 2022. +[26] Patrick Emami, Pan He, Sanjay Ranka, and Anand Rangarajan. Efficient iterative amortized inference for learning symmetric and disentangled multi-object representations. In International Conference on Machine Learning, pages 2970-2981. PMLR, 2021. +[27] Martin Engelcke, Adam R Kosiorek, Oiwi Parker Jones, and Ingmar Posner. Genesis: Generative scene inference and sampling with object-centric latent representations. arXiv preprint arXiv:1907.13052, 2019. +[28] Martin Engelcke, Oiwi Parker Jones, and Ingmar Posner. Genesis-v2: Inferring unordered object representations without iterative refinement. Advances in Neural Information Processing Systems, 34:8085-8094, 2021. +[29] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3299-3309, 2021. +[30] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. + +[31] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015. +[32] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthey, Matthew Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative variational inference. In International Conference on Machine Learning, pages 2424-2433. PMLR, 2019. +[33] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. +[34] Kai Han, Rafael S Rezende, Bumsub Ham, Kwan-Yee K Wong, Minsu Cho, Cordelia Schmid, and Jean Ponce. Sc net: Learning semantic correspondence. In Proceedings of the IEEE international conference on computer vision, pages 1831-1840, 2017. +[35] Tengda Han, Weidi Xie, and Andrew Zisserman. Video representation learning by dense predictive coding. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. +[36] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. Advances in Neural Information Processing Systems, 33:5679-5690, 2020. +[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. +[38] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. +[39] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. +[40] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[41] David Held, Sebastian Thrun, and Silvio Savarese. Learning to track at 100 fps with deep regression networks. In European conference on computer vision, pages 749-765. Springer, 2016. +[42] Olivier Henaff. Data-efficient image recognition with contrastive predictive coding. In International conference on machine learning, pages 4182-4192. PMLR, 2020. +[43] Yingdong Hu, Renhao Wang, Kaifeng Zhang, and Yang Gao. Semantic-aware fine-grained correspondence. arXiv preprint arXiv:2207.10456, 2022. +[44] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances + +in neural information processing systems, 33:19545-19560, 2020. +[45] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013. +[46] Baoxiong Jia, Yu Liu, and Siyuan Huang. Unsupervised object-centric learning with bi-level optimized query slot attention. arXiv preprint arXiv:2210.08990, 2022. +[47] Longlong Jing, Xiaodong Yang, Jingen Liu, and Yingli Tian. Self-supervised spatiotemporal feature learning via video rotation prediction. arXiv preprint arXiv:1811.11387, 2018. +[48] Rishabh Kabra, Daniel Zoran, Goker Erdogan, Loic Matthew, Antonia Creswell, Matt Botvinick, Alexander Lerchner, and Chris Burgess. Simone: View-invariant, temporally-abstracted object representations via unsupervised video decomposition. Advances in Neural Information Processing Systems, 34:20146-20159, 2021. +[49] Dahun Kim, Donghyeon Cho, Donggeun Yoo, and In So Kweon. Learning image representations by completing damaged jigsaw puzzles. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 793-802. IEEE, 2018. +[50] Thomas Kipf, Gamaleldin F Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional object-centric learning from video. arXiv preprint arXiv:2111.12594, 2021. +[51] Zihang Lai, Erika Lu, and Weidi Xie. Mast: A memory-augmented self-supervised tracker. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6479-6488, 2020. +[52] Zihang Lai and Weidi Xie. Self-supervised learning for video correspondence flow. arXiv preprint arXiv:1905.00875, 2019. +[53] Hala Lamdouar, Weidi Xie, and Andrew Zisserman. Segmenting invisible moving objects. 2021. +[54] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In Proceedings of the IEEE international conference on computer vision, pages 2192-2199, 2013. +[55] Xueting Li, Sifei Liu, Shalini De Mello, Xiaolong Wang, Jan Kautz, and Ming-Hsuan Yang. Joint-task self-supervised learning for temporal correspondence. Advances in Neural Information Processing Systems, 32, 2019. +[56] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1739-1748, 2021. +[57] Runtao Liu, Zhirong Wu, Stella Yu, and Stephen Lin. The emergence of objectness: Learning zero-shot segmentation from videos. Advances in Neural Information Processing Systems, 34:13137-13152, 2021. +[58] Songtao Liu, Zeming Li, and Jian Sun. Self-emd: Self-supervised object detection withoutImagenet. arXiv preprint arXiv:2011.13677, 2020. + +[59] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in Neural Information Processing Systems, 33:11525-11538, 2020. +[60] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. +[61] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. +[62] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European conference on computer vision, pages 527-544. Springer, 2016. +[63] Joe Yue-Hei Ng, Jonghyun Choi, Jan Neumann, and Larry S Davis. Actionflownet: Learning motion representation for action recognition. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1616–1624. IEEE, 2018. +[64] Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. IEEE transactions on pattern analysis and machine intelligence, 36(6):1187-1200, 2013. +[65] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. +[66] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. +[67] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. +[68] Rui Qian, Shuangrui Ding, Xian Liu, and Dahua Lin. Static and dynamic concepts for self-supervised video representation learning. In European Conference on Computer Vision, pages 145-164. Springer, 2022. +[69] Rui Qian, Yuxi Li, Huabin Liu, John See, Shuangrui Ding, Xian Liu, Dian Li, and Weiyao Lin. Enhancing self-supervised video representation learning via multi-level feature optimization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7990-8001, 2021. +[70] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6964-6974, 2021. +[71] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. + +[72] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, et al. Bridging the gap to real-world object-centric learning. arXiv preprint arXiv:2209.14860, 2022. +[73] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8934-8943, 2018. +[74] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. +[75] Jack Valmadre, Luca Bertinetto, Joao Henriques, Andrea Vedaldi, and Philip HS Torr. End-to-end representation learning for correlation filter based tracking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2805-2813, 2017. +[76] Carles Ventura, Miriam Bellver, Andreu Girbau, Amaia Salvador, Ferran Marques, and Xavier Giro-i Nieto. Rvos: End-to-end recurrent network for video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5277-5286, 2019. +[77] Paul Voigtlaender and Bastian Leibe. Online adaptation of convolutional neural networks for video object segmentation. arXiv preprint arXiv:1706.09364, 2017. +[78] Carl Vondrick, Abhinav Shrivastava, Alireza Fathi, Sergio Guadarrama, and Kevin Murphy. Tracking emerges by colorizing videos. In Proceedings of the European conference on computer vision (ECCV), pages 391-408, 2018. +[79] Naiyan Wang and Dit-Yan Yeung. Learning a deep compact image representation for visual tracking. Advances in neural information processing systems, 26, 2013. +[80] Xiaolong Wang, Kaiming He, and Abhinav Gupta. Transitive invariance for self-supervised visual representation learning. In Proceedings of the IEEE international conference on computer vision, pages 1329-1338, 2017. +[81] Xiaolong Wang, Allan Jabri, and Alexei A Efros. Learning correspondence from the cycle-consistency of time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2566-2576, 2019. +[82] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. +[83] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. +[84] Junyu Xie, Weidi Xie, and Andrew Zisserman. Segmenting moving objects via an object-centric layered representation. arXiv preprint arXiv:2207.02206, 2022. +[85] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level + +consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021. +[86] Jiarui Xu and Xiaolong Wang. Rethinking self-supervised correspondence learning: A video frame-level similarity perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10075-10085, 2021. +[87] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen, and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 585-601, 2018. +[88] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7177-7188, 2021. +[89] Yanchao Yang, Antonio Loquercio, Davide Scaramuzza, and Stefano Soatto. Unsupervised moving object detection via contextual information separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 879-888, 2019. +[90] Qixian Zhou, Xiaodan Liang, Ke Gong, and Liang Lin. Adaptive temporal encoding network for video instance-level human parsing. In Proceedings of the 26th ACM international conference on Multimedia, pages 1527-1535, 2018. +[91] Xizhou Zhu, Yujie Wang, Jifeng Dai, Lu Yuan, and Yichen Wei. Flow-guided feature aggregation for video object detection. In Proceedings of the IEEE international conference on computer vision, pages 408-417, 2017. \ No newline at end of file diff --git a/2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/images.zip b/2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..070ca9582f2a5b1e26476046d3546c9139b0b475 --- /dev/null +++ b/2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af4d1bc599c9881bd894a8fa9b37216f3935288dde9f46a753a9e14148b794d9 +size 499855 diff --git a/2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/layout.json b/2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d98b8e4f362082cfc5b89355eb3a2bdb0da602af --- /dev/null +++ b/2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/layout.json @@ -0,0 +1,10610 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 103, + 483, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 103, + 483, + 139 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 483, + 139 + ], + "type": "text", + "content": "Semantics Meets Temporal Correspondence: Self-supervised Object-centric Learning in Videos" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 160, + 443, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 160, + 443, + 175 + ], + "spans": [ + { + "bbox": [ + 153, + 160, + 443, + 175 + ], + "type": "text", + "content": "Rui Qian1 Shuangrui Ding1 Xian Liu1 Dahua Lin1,2*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 175, + 505, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 175, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 88, + 175, + 505, + 190 + ], + "type": "text", + "content": "1The Chinese University of Hong Kong 2Shanghai Artificial Intelligence Laboratory" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 188, + 191, + 405, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 191, + 405, + 202 + ], + "spans": [ + { + "bbox": [ + 188, + 191, + 405, + 202 + ], + "type": "text", + "content": "{qr021,ds023,lx021,dhlin}@ie.cuhk.edu.hk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 256, + 290, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 256, + 290, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 256, + 290, + 628 + ], + "type": "text", + "content": "Self-supervised methods have shown remarkable progress in learning high-level semantics and low-level temporal correspondence. Building on these results, we take one step further and explore the possibility of integrating these two features to enhance object-centric representations. Our preliminary experiments indicate that query slot attention can extract different semantic components from the RGB feature map, while random sampling based slot attention can exploit temporal correspondence cues between frames to assist instance identification. Motivated by this, we propose a novel semantic-aware masked slot attention on top of the fused semantic features and correspondence maps. It comprises two slot attention stages with a set of shared learnable Gaussian distributions. In the first stage, we use the mean vectors as slot initialization to decompose potential semantics and generate semantic segmentation masks through iterative attention. In the second stage, for each semantics, we randomly sample slots from the corresponding Gaussian distribution and perform masked feature aggregation within the semantic area to exploit temporal correspondence patterns for instance identification. We adopt semantic- and instance-level temporal consistency as self-supervision to encourage temporally coherent object-centric representations. Our model effectively identifies multiple object instances with semantic structure, reaching promising results on unsupervised video object discovery. Furthermore, we achieve state-of-the-art performance on dense label propagation tasks, demonstrating the potential for object-centric analysis. The code is released at https://github.com/shvdiwnkozbw/SMTC." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 650, + 128, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 650, + 128, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 650, + 128, + 662 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 670, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 287, + 696 + ], + "type": "text", + "content": "As one of the fundamental cognitive capabilities, human beings easily distinguish different objects, establish visual" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 230, + 426, + 292 + ], + "blocks": [ + { + "bbox": [ + 318, + 230, + 426, + 292 + ], + "lines": [ + { + "bbox": [ + 318, + 230, + 426, + 292 + ], + "spans": [ + { + "bbox": [ + 318, + 230, + 426, + 292 + ], + "type": "image", + "image_path": "06fe90c013fc77671267ef4b2e01a9a4c5c5ec86500384249bf7b59b4041c790.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 296, + 498, + 305 + ], + "lines": [ + { + "bbox": [ + 354, + 296, + 498, + 305 + ], + "spans": [ + { + "bbox": [ + 354, + 296, + 498, + 305 + ], + "type": "text", + "content": "(a) Query slot attention on semantic feature." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 427, + 230, + 534, + 292 + ], + "blocks": [ + { + "bbox": [ + 427, + 230, + 534, + 292 + ], + "lines": [ + { + "bbox": [ + 427, + 230, + 534, + 292 + ], + "spans": [ + { + "bbox": [ + 427, + 230, + 534, + 292 + ], + "type": "image", + "image_path": "95713c39e96e743b2ba94bccd45808cf70187045eaa76c8e70f3830f7ba15b8e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 318, + 314, + 426, + 376 + ], + "blocks": [ + { + "bbox": [ + 318, + 314, + 426, + 376 + ], + "lines": [ + { + "bbox": [ + 318, + 314, + 426, + 376 + ], + "spans": [ + { + "bbox": [ + 318, + 314, + 426, + 376 + ], + "type": "image", + "image_path": "f54196344cb58166b76125737f0d095a1fff030f384f1f1f8b67aba06fea8ca8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 380, + 523, + 390 + ], + "lines": [ + { + "bbox": [ + 328, + 380, + 523, + 390 + ], + "spans": [ + { + "bbox": [ + 328, + 380, + 523, + 390 + ], + "type": "text", + "content": "(b) Random sampling slot attention on correspondence map." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 427, + 315, + 534, + 376 + ], + "blocks": [ + { + "bbox": [ + 427, + 315, + 534, + 376 + ], + "lines": [ + { + "bbox": [ + 427, + 315, + 534, + 376 + ], + "spans": [ + { + "bbox": [ + 427, + 315, + 534, + 376 + ], + "type": "image", + "image_path": "a9fc3ad0a33a1ce8c341f7a5026395f04b9f055999083bf9257420247028d229.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 392, + 547, + 470 + ], + "lines": [ + { + "bbox": [ + 305, + 392, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 305, + 392, + 547, + 470 + ], + "type": "text", + "content": "Figure 1. Fig. 1(a) presents the results of query slot attention on top of the RGB feature map. It successfully decomposes different semantics, e.g., camels and fence. Fig. 1(b) visualizes the correspondence map after PCA dimensionality reduction, showing that different instances have different correspondence patterns. And the slot attention with random sampling coarsely distinguishes two camels with some redundant borders. Best viewed in color." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 495, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 567 + ], + "type": "text", + "content": "correspondence and perform object-centric analysis from temporally continuous observations. This ability can be attributed to two indispensable visual mechanisms: high-level semantic discrimination as well as low-level temporal correspondence, which enable humans to effectively understand and interact with the world." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "Motivated by this, computer vision researchers equip machines with these capacities to enhance object-centric perception [79, 60, 34]. To achieve this goal, early works rely on human annotations or weak supervision to perceive object semantics [60, 63, 13], identify geometric positions [39, 71, 31, 91], and establish temporal correspondence [41, 79, 75, 77], but their generalization ability is limited. Recently, there emerge a host of fully unsupervised methods to learn robust representations for semantic discrimination [14, 38, 33, 70, 29, 37, 12] or spatiotemporal correspondence [44, 81, 86, 55, 52, 43], which achieve promising performance. Given this encouraging re" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 234, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 234, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 234, + 712 + ], + "type": "text", + "content": "*Corresponding author. Email: dhlin@ie.cuhk.edu.hk." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16675" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 285, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 285, + 118 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 285, + 118 + ], + "type": "text", + "content": "sult, we naturally come up with a question: Is it possible to jointly leverage the semantics and correspondence to discover object instances and distill object-centric representations without human annotations?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 122, + 285, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 122, + 285, + 325 + ], + "spans": [ + { + "bbox": [ + 49, + 122, + 285, + 325 + ], + "type": "text", + "content": "Regarding this problem, our intuition is that the high-level semantics delineates meaningful foreground areas in a top-down manner, while when looking into more frames, the low-level correspondence temporally associates coherent objects and separates individual instances in a bottom-up fashion. For instance, in a football scene, the semantic cue differentiates the foreground that includes several players, while the temporal correspondence links distinct players through dynamic movements and geometric relationships. These two aspects collectively contribute to object-centric representations. Unfortunately, most of the existing works only concentrate on one of these features. [14, 38, 33, 70, 82] succeed in developing high-level semantics, but this abstract semantics alone is insufficient to distinguish instances. Whereas, [44, 81, 78, 55] excel in detailed correspondence, but lack semantic structure and result in redundancy and ambiguity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 329, + 285, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 329, + 285, + 613 + ], + "spans": [ + { + "bbox": [ + 49, + 329, + 285, + 613 + ], + "type": "text", + "content": "In this paper, we propose a new architecture, Semantics Meets Temporal Correspondence (SMTC), to jointly leverage semantics and temporal correspondence to distill object-centric representations from RGB sequences. Specifically, we first extract frame-wise visual features as the semantic representation. Then we calculate dense feature correlations between adjacent frames as the correspondence map which encodes temporal relationships. To mine the object-centric knowledge, we take inspiration from [59, 88, 50, 25], and investigate using different formulations of slot attention on them. The preliminary experiments show that the original slot attention with random sampling on RGB feature map suffers from complex scene components in real-world videos [59, 72], but the revised query slot attention [88, 46] can decompose different semantic components as shown in Fig. 1(a). As for the correspondence map, different objects present diverse temporal correspondence patterns. Comparing to semantic features, these patterns reveal low-level geometric relationships, which are comparatively simple but vary with specific scenes. Hence, query slot attention fails but the random sampling based formulation performs surprisingly well as shown in Fig. 1(b), coarsely separating different object areas with some redundant borders." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 618, + 285, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 618, + 285, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 618, + 285, + 712 + ], + "type": "text", + "content": "Motivated by this, we develop semantic-aware masked slot attention, which comprises a set of Gaussian distributions with learnable mean and standard deviation vectors, on top of the fused semantic and correspondence representations. The intuition is that the mean vectors could represent potential semantic centers, which act similarly to the query slot attention to separate semantic components. While the deviation vectors introduce perturbations around" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 308, + 72, + 544, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 251 + ], + "type": "text", + "content": "the semantic centers to capture distinct temporal correspondence patterns of different instances. Technically, we formulate two slot attention stages to achieve this goal. Firstly, we use the mean vectors as slot initialization to generate semantic segmentation masks. Secondly, for each semantics, we randomly sample slot vectors from the Gaussian distribution, then perform iterative attention and masked aggregation within the corresponding semantic mask area to distinguish instances. We enforce temporal consistency on the semantic masks as well as object instance slots to enhance temporal coherency and refine object-centric representations. Comparing with existing works on object-centric learning in videos [25, 50, 88, 84], our model is free of pre-computed motion or depth prior, and explicitly identifies multiple objects with semantic structure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 252, + 544, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 252, + 544, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 252, + 544, + 442 + ], + "type": "text", + "content": "In summary, our contributions are: (1) We propose a novel self-supervised architecture that unifies semantic discrimination and temporal correspondence to distill object-centric representations in videos. (2) We demonstrate that simple feature correlation can effectively represent temporal correspondence cues when used in conjunction with semantic features. Building on this observation, we develop semantic-aware masked slot attention, which operates on fused visual features and correspondence maps, to distinguish multiple object instances with semantic structure without relying on motion or depth priors. (3) We achieve promising results on unsupervised object discovery in both single and multiple object scenarios, and reach state-of-the-art performance on label propagation tasks, demonstrating that we learn discriminative and temporally consistent object-centric representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 453, + 391, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 453, + 391, + 466 + ], + "spans": [ + { + "bbox": [ + 308, + 453, + 391, + 466 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 475, + 544, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 544, + 700 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 544, + 700 + ], + "type": "text", + "content": "Unsupervised Object Discovery is an essential process to formulate object-centric representations, which aims to identify objects without human annotations. There exist a series of works focusing on this problem [59, 32, 26, 27, 28, 7]. Typically, [59] develops slot attention to iteratively update latent object representations from random initializations but it has difficulty scaling to complex real-world scenes. To tackle this challenge, [72] employs feature reconstruction as objective to reduce redundancy, [46] develops query slot initialization to encode visual concepts. Further, a line of works extend object discovery to video domain [50, 48, 17, 5, 88]. Most of them build on slot attention architecture, and adopt extra pre-computed priors, e.g., optical flow [88, 84, 16, 21], depth [25], geometric positions [50], as input or supervision to assist object discovery. In our work, we find that the simple feature correlation provides informative temporal cues. And we develop semantic-aware masked slot attention to identify object instances with semantic structure without resorting to extra priors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 702, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 702, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 702, + 544, + 712 + ], + "type": "text", + "content": "Self-supervised Representation Learning aims to learn" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "16676" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 545, + 281 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 545, + 281 + ], + "type": "image", + "image_path": "fe93decb9a5c6db629d51dcb51e2632af2e45a5db995b6fabaad9a7b39399dd8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 283, + 546, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 283, + 546, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 546, + 350 + ], + "type": "text", + "content": "Figure 2. An overview of our framework. We first extract frame-wise features and calculate dense feature correlation, then fuse them to pass through semantic-aware masked slot attention, which comprises two slot attention stages with " + }, + { + "bbox": [ + 46, + 283, + 546, + 350 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 283, + 546, + 350 + ], + "type": "text", + "content": " shared learnable Gaussian distributions. In the first semantic slot attention stage, the mean vectors serve as slot initialization to generate a set of segmentation masks for semantic decomposition. In the second masked slot attention stage, which runs on " + }, + { + "bbox": [ + 46, + 283, + 546, + 350 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 283, + 546, + 350 + ], + "type": "text", + "content": " semantics in parallel, we randomly sample slots from the Gaussian distribution of each semantics and perform masked feature aggregation within the semantic area to identify distinct instances. We enforce semantic and instance temporal consistency to train the architecture in a teacher-student manner, with the teacher marked in gray." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 369, + 289, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 289, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 289, + 536 + ], + "type": "text", + "content": "robust representation without human annotations. Early works design various pretext tasks to generate pseudo labels as self-supervision [49, 30, 47, 10, 1, 62]. Later, contrastive learning with instance discrimination revolutionized the field [38, 14, 70, 29, 35, 83, 65, 42, 11, 20, 3, 68, 35, 36, 69, 19, 22]. These works generally contribute to general semantics but lack the ability to distill object-centric representations. To delve into this problem, [82, 85] preserve spatial sensitivity and perform contrastive learning on dense features. [86, 12] point out that semantic pre-training implicitly encodes object correspondence. While in our work, we learn to explicitly identify different objects with instance-level consistency to progressively refine object-centric representations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": "Self-supervised Correspondence is a fundamental problem in computer vision [55, 78, 81, 80]. Typical works rely on low-level statistics like color to generate the dense correspondence flow [51, 52, 78], or establish temporal cycles to track image regions [81, 44, 6]. Further, [43, 86, 12] incorporate correspondence learning with semantic discrimination, but they either regard correspondence as a side effect of semantic discrimination [86, 12] or independently learn two pathway features with late fusion [43]. Among these works, high-level semantics is not well utilized, and there exists high redundancy due to large background areas. In contrast, our method explicitly separates semantic components to alleviate redundancy and formulate instance-level correspondence from an object-centric perspective." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 368, + 361, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 368, + 361, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 368, + 361, + 380 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 389, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 498 + ], + "type": "text", + "content": "Our framework is shown in Fig. 2, which adopts a teacher-student structure similar to recent self-supervised methods [38, 12, 33]. Specifically, we first extract framewise features from an RGB sequence, and calculate dense feature correlation as temporal correspondence representations. After that, we fuse them to pass through semantic-aware masked slot attention to separate multiple object instances with semantic structure, and enforce temporal consistency to refine the object-centric representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 510, + 411, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 510, + 411, + 523 + ], + "spans": [ + { + "bbox": [ + 306, + 510, + 411, + 523 + ], + "type": "text", + "content": "3.1. Feature Encoding" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 530, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 530, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 530, + 545, + 555 + ], + "type": "text", + "content": "Given a RGB sequence " + }, + { + "bbox": [ + 305, + 530, + 545, + 555 + ], + "type": "inline_equation", + "content": "v = \\{I_1, I_2, \\dots, I_T\\}" + }, + { + "bbox": [ + 305, + 530, + 545, + 555 + ], + "type": "text", + "content": ", we employ a shared visual encoder " + }, + { + "bbox": [ + 305, + 530, + 545, + 555 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 530, + 545, + 555 + ], + "type": "text", + "content": " to extract frame-wise features:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 376, + 578, + 545, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 578, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 376, + 578, + 545, + 593 + ], + "type": "interline_equation", + "content": "F _ {t} = f \\left(I _ {t}\\right) \\in \\mathbb {R} ^ {H W \\times D}, \\tag {1}", + "image_path": "32d78cf18782debab2ccc66dfc7e493aed06a0e301b91e6b9ac5f8fdda53000a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "H, W, D" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " respectively denotes height, width and channel dimension, " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " indicates frame index. For simplicity, we omit the subscript indicating student or teacher pathway, and use " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\hat{F}" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " to denote teacher outputs. Intuitively, " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " contains appearance and semantic cues in the input sequence, and the next step is to formulate an effective representation for the temporal correspondence cues. To do this, inspired by the cost volume in motion estimation [24, 73, 74], we calculate dense feature correlation to generate the temporal" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16677" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 96, + 70, + 239, + 118 + ], + "blocks": [ + { + "bbox": [ + 96, + 70, + 239, + 118 + ], + "lines": [ + { + "bbox": [ + 96, + 70, + 239, + 118 + ], + "spans": [ + { + "bbox": [ + 96, + 70, + 239, + 118 + ], + "type": "table", + "html": "
FeatureQueryRandom
RGB58.138.4
Correlation21.556.3
", + "image_path": "27bc5aa89f1d02eb10a5f10b7873e1d35a06f0564f41870597c03af712daf58b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 118, + 287, + 150 + ], + "lines": [ + { + "bbox": [ + 47, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 47, + 118, + 287, + 150 + ], + "type": "text", + "content": "Table 1. Preliminary results on RGB feature map and feature correlation. 'Query' ('Random') denotes slot attention with query (random sampling) initialization. We report IoU on DAVIS-2016." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 170, + 214, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 170, + 214, + 182 + ], + "spans": [ + { + "bbox": [ + 47, + 170, + 214, + 182 + ], + "type": "text", + "content": "correspondence map for each timestamp:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 190, + 287, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 190, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 109, + 190, + 287, + 205 + ], + "type": "interline_equation", + "content": "C _ {t j} = F _ {t} F _ {j} ^ {T} \\in \\mathbb {R} ^ {H W \\times H W}, \\tag {2}", + "image_path": "f7095cd8fd90947c66bec5c2a190bea89e8774569c521c4c8ac832a535b2be85.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "type": "text", + "content": "where we randomly sample one frame index " + }, + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "type": "inline_equation", + "content": "j \\neq t" + }, + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "type": "text", + "content": " to compute feature correlation. The channel activations of " + }, + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "type": "inline_equation", + "content": "C_{tj}" + }, + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "type": "text", + "content": " encode low-level geometric correlations between frame " + }, + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "type": "text", + "content": ", which indicates potential partitions of different objects as illustrated in Fig. 1(b)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 274, + 287, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 274, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 47, + 274, + 287, + 332 + ], + "type": "text", + "content": "To jointly leverage the semantic and correspondence cues, we employ two linear transformation heads, " + }, + { + "bbox": [ + 47, + 274, + 287, + 332 + ], + "type": "inline_equation", + "content": "h_f" + }, + { + "bbox": [ + 47, + 274, + 287, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 274, + 287, + 332 + ], + "type": "inline_equation", + "content": "h_c" + }, + { + "bbox": [ + 47, + 274, + 287, + 332 + ], + "type": "text", + "content": ", to respectively project " + }, + { + "bbox": [ + 47, + 274, + 287, + 332 + ], + "type": "inline_equation", + "content": "F_t" + }, + { + "bbox": [ + 47, + 274, + 287, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 274, + 287, + 332 + ], + "type": "inline_equation", + "content": "C_{tj}" + }, + { + "bbox": [ + 47, + 274, + 287, + 332 + ], + "type": "text", + "content": " to a shared " + }, + { + "bbox": [ + 47, + 274, + 287, + 332 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 274, + 287, + 332 + ], + "type": "text", + "content": "-dimensional embedding space, which are then fused with element-wise summation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 341, + 287, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 341, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 91, + 341, + 287, + 356 + ], + "type": "interline_equation", + "content": "R _ {t} = h _ {f} \\left(F _ {t}\\right) + h _ {c} \\left(C _ {t j}\\right) \\in \\mathbb {R} ^ {H W \\times D}. \\tag {3}", + "image_path": "4ba167616977aa00b1465fecf1425d6baf098d2e47b64e54598af5ae720eb194.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 365, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 287, + 400 + ], + "type": "text", + "content": "In this way, " + }, + { + "bbox": [ + 47, + 365, + 287, + 400 + ], + "type": "inline_equation", + "content": "R_{t}" + }, + { + "bbox": [ + 47, + 365, + 287, + 400 + ], + "type": "text", + "content": " simultaneously encodes semantic and temporal correspondence information, serving as an intermediate feature for object-centric analysis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 407, + 280, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 280, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 280, + 420 + ], + "type": "text", + "content": "3.2. Preliminary Experiments with Slot Attention" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 426, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 521 + ], + "type": "text", + "content": "Given these representations, the next step is to formulate effective ways to identify individual objects. To do this, we take inspiration from slot attention [59, 50, 88, 46], which employs a set of slot vectors to iteratively attend to specific objects. Generally, there are two alternatives for slot initialization, random sampling from a Gaussian distribution with learnable parameters " + }, + { + "bbox": [ + 46, + 426, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mu, \\sigma \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 426, + 287, + 521 + ], + "type": "text", + "content": " [59, 50] or directly inheriting from a set of learnable queries [46, 88]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": "To investigate the effectiveness of these two formulations, we conduct preliminary experiments on the semantic feature " + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": " and correspondence map " + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "inline_equation", + "content": "C_{tj}" + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": " respectively. We follow the settings in [72], using a frozen DINO pre-trained ViT to extract visual feature and generate correspondence map, on top of which we perform two kinds of slot attention. We train and evaluate them on DAVIS-2016 dataset [66]. As indicated by Fig. 1 and Table. 1, we find that query slot attention decomposes distinct semantic components on " + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": ", while the original slot attention with random sampling performs much worse and struggles in complex real-world scenes. Conversely, query slot attention does not work on the correspondence map, and random sampling initialization enables the model to effectively utilize the temporal correspondence patterns in " + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "inline_equation", + "content": "C_{tj}" + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": " to distinguish instances. An intuitive explanation into these observations is the RGB" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "content": "features contain rich semantics and there exist consistent patterns for similar objects. Slot attention with learnable queries can capture inter-sample semantic patterns to identify objects. While random initialization fails to memorize semantic patterns, thus resulting in worse performance. In contrast, correspondence features reveal inter-frame geometric correlations, shown in Fig. 1(b), which vary with specific scenes. There are no consistent patterns for similar objects, so the learnable queries fail but random initialization works well to capture correspondence cues. The complementarity between these two conditions motivates two-stage slot attention design to combine the attributes of query and random sampling initialization, thus better exploiting the semantic and correspondence cues in " + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "inline_equation", + "content": "R_{t}" + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 246, + 511, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 246, + 511, + 258 + ], + "spans": [ + { + "bbox": [ + 306, + 246, + 511, + 258 + ], + "type": "text", + "content": "3.3. Semantic-aware Masked Slot Attention" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 264, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 443 + ], + "type": "text", + "content": "We propose semantic-aware masked slot attention, which comprises " + }, + { + "bbox": [ + 304, + 264, + 545, + 443 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 264, + 545, + 443 + ], + "type": "text", + "content": " Gaussian distributions with learnable means " + }, + { + "bbox": [ + 304, + 264, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} = \\{\\mu_1,\\mu_2,\\dots,\\mu_N\\} \\in \\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 304, + 264, + 545, + 443 + ], + "type": "text", + "content": " and standard deviations " + }, + { + "bbox": [ + 304, + 264, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\sigma = \\{\\sigma_{1},\\sigma_{2},\\dots,\\sigma_{N}\\} \\in \\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 304, + 264, + 545, + 443 + ], + "type": "text", + "content": ". With this design, the mean vectors act similarly to the learnable queries, each representing a potential semantic center. And the deviations introduce perturbations around semantic centers to model temporal correspondence patterns. To achieve this goal, we employ two stages of iterative attention with shared learnable parameters. First, we use the mean vectors as slot initialization to decompose different semantics in the form of segmentation. Second, we randomly sample slots from each Gaussian distribution " + }, + { + "bbox": [ + 304, + 264, + 545, + 443 + ], + "type": "inline_equation", + "content": "(\\mu_n,\\mathrm{diag}(\\sigma_n))" + }, + { + "bbox": [ + 304, + 264, + 545, + 443 + ], + "type": "text", + "content": ", and perform masked aggregation within the corresponding semantic area to identify instances. The process is detailed as follows." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 444, + 545, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 545, + 528 + ], + "type": "text", + "content": "Semantic Decomposition. In the first stage, we directly use " + }, + { + "bbox": [ + 304, + 444, + 545, + 528 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 304, + 444, + 545, + 528 + ], + "type": "text", + "content": " without randomness as slot initialization " + }, + { + "bbox": [ + 304, + 444, + 545, + 528 + ], + "type": "inline_equation", + "content": "S = \\pmb{\\mu} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 304, + 444, + 545, + 528 + ], + "type": "text", + "content": ". Then following the standard slot attention procedure [59], we employ three linear transformation heads to map " + }, + { + "bbox": [ + 304, + 444, + 545, + 528 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 444, + 545, + 528 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 444, + 545, + 528 + ], + "type": "inline_equation", + "content": "R_{t}" + }, + { + "bbox": [ + 304, + 444, + 545, + 528 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 444, + 545, + 528 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{N \\times D}, K, V \\in \\mathbb{R}^{HW \\times D}" + }, + { + "bbox": [ + 304, + 444, + 545, + 528 + ], + "type": "text", + "content": ", and iteratively calculate attention score and update slot representations. Mathematically, we formulate each iteration as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 337, + 543, + 545, + 568 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 543, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 337, + 543, + 545, + 568 + ], + "type": "interline_equation", + "content": "M := \\operatorname {s o f t m a x} \\left(\\frac {1}{\\sqrt {D}} Q K ^ {T}, 0\\right) \\in \\mathbb {R} ^ {N \\times H W}, \\tag {4}", + "image_path": "904624d2de479b8c946316675fa94397dff3f557c68f7370e02545f609eb196a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 337, + 569, + 545, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 569, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 337, + 569, + 545, + 597 + ], + "type": "interline_equation", + "content": "A [ n, i ] := \\frac {M [ n , i ]}{\\sum_ {l} M [ n , l ] + \\epsilon}, A \\in \\mathbb {R} ^ {N \\times H W}, \\tag {5}", + "image_path": "96bb1b10edd76ef50b430bb43adc0178593b2edce47c0dbe78e79758862f9e16.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 337, + 598, + 545, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 598, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 337, + 598, + 545, + 611 + ], + "type": "interline_equation", + "content": "S := \\operatorname {G R U} (\\text {i n p u t} = A V, \\text {s t a t e s} = S). \\tag {6}", + "image_path": "002a5a25b68d8362ec5f6af026d895442fb8fe562537dfa9894760f06ba05e5c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Note that the slot attention weight " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " is normalized along the query axis, and the weighted mean coefficient " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " is employed to aggregate the value to update the slots. This mechanism induces competition among slots and enforce different slots to take over different semantic features. We iterate this process for three times, and take the slot attention weight (slot vectors) of the final iteration as the segmentation mask (representation) for different semantic compo" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16678" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "nents. To avoid ambiguity in presentation, we denote them as " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\in \\mathbb{R}^{N \\times HW}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "S[n]" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " presents " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "-th semantic center representation, " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{M}[n]" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " indicates the probability that each pixel belongs to " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "-th semantics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "type": "text", + "content": "Instance Identification. Within each semantics, there could exist multiple object instances. The Gaussian distributions we introduce help to distinguish them with temporal correspondence cues. Specifically, in the second stage, for each semantics, we randomly sample slots from the Gaussian distribution, using the perturbations around the semantic center to capture distinct temporal correspondence patterns. For simplicity, we take the " + }, + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "type": "text", + "content": "-th semantics as an example. First, we sample " + }, + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "type": "text", + "content": " vectors, " + }, + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "type": "inline_equation", + "content": "S \\sim \\mathcal{N}(\\mu_n, \\mathrm{diag}(\\sigma_n)) \\in \\mathbb{R}^{P \\times D}" + }, + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "type": "text", + "content": ", as slot initialization to represent " + }, + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "type": "text", + "content": " potential instances. Next, we follow the first stage formulation to respectively project " + }, + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "type": "inline_equation", + "content": "R_t" + }, + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{P \\times D}" + }, + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "type": "inline_equation", + "content": "K, V \\in \\mathbb{R}^{HW \\times D}" + }, + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "type": "text", + "content": ", and obtain the attention score " + }, + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{P \\times HW}" + }, + { + "bbox": [ + 46, + 120, + 289, + 335 + ], + "type": "text", + "content": ". To preserve the semantic attributes and discriminate correspondence cues, we perform masked feature aggregation within the semantic mask areas. The only mathematical difference in this process is the weighted mean coefficient computation:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 343, + 287, + 391 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 343, + 287, + 391 + ], + "spans": [ + { + "bbox": [ + 59, + 343, + 287, + 391 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} A [ p, i ] := \\frac {M [ p , i ] \\widetilde {\\mathcal {M}} [ n , i ]}{\\sum_ {l} M [ p , l ] \\widetilde {\\mathcal {M}} [ n , l ] + \\epsilon}, A \\in \\mathbb {R} ^ {P \\times H W}, \\tag {7} \\\\ \\widetilde {\\mathcal {M}} [ n, i ] = \\operatorname {b i n a r i z e} (\\mathcal {M} [ n, i ], \\tau), \\mathcal {M} \\in \\mathbb {R} ^ {N \\times H W}, \\\\ \\end{array}", + "image_path": "283d4421b86205aba5b7ac1f462d1f28a68e462a0a6b03ba2bd475be58fa3e0b.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "type": "text", + "content": "where we use the " + }, + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "type": "text", + "content": "-th semantic mask " + }, + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{M}}[n] \\in \\{0,1\\}^{HW}" + }, + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "type": "text", + "content": " binarized with threshold " + }, + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "type": "text", + "content": " to constrain the model to only aggregate the visual features related to " + }, + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "type": "text", + "content": "-th semantics. And we empirically find a comparatively large threshold " + }, + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "type": "inline_equation", + "content": "\\tau = 0.5" + }, + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "type": "text", + "content": " helps to preserve object regions with clear semantics and filter out background. The masked slot attention for all semantics follows the same rule and is calculated in parallel. Similarly, we iterate the process for three times, use the final slot attention weight as the instance segmentation masks, and take the final slot vectors as the object-centric representations " + }, + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{O} \\in \\mathbb{R}^{N \\times P \\times D}" + }, + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{O}[n,p] \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "type": "text", + "content": "-th potential object instance of " + }, + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 399, + 289, + 569 + ], + "type": "text", + "content": "-th semantics. In this way, our semantic-aware masked slot attention explicitly segments multiple instances with semantic structure." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 575, + 111, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 111, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 111, + 588 + ], + "type": "text", + "content": "3.4. Training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "type": "text", + "content": "We use temporal consistency as self-supervision to optimize semantic masks " + }, + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "type": "text", + "content": " and instance representations " + }, + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "Dense Semantic Alignment. Given the semantic mask of each timestamp " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t\\in \\mathbb{R}^{N\\times HW}" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": ", as well as " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}_t" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " from teacher pathway, we aim to align the spatially dense semantic distributions across time. To achieve this, the first step is to determine the corresponding patches between different frames. Utilizing the feature correlation " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "C_{tj}\\in \\mathbb{R}^{HW\\times HW}" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " which indicates dense feature similarity between timestamps " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "t,j" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": ", we can infer patch-level correspondence from" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 546, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 546, + 121 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 546, + 121 + ], + "type": "text", + "content": "this cue with optimal transport [11, 2, 58, 18]. Formally, we take " + }, + { + "bbox": [ + 304, + 72, + 546, + 121 + ], + "type": "inline_equation", + "content": "-\\hat{C}_{tj}" + }, + { + "bbox": [ + 304, + 72, + 546, + 121 + ], + "type": "text", + "content": " from teacher pathway as cost matrix, and solve the optimal transport strategy " + }, + { + "bbox": [ + 304, + 72, + 546, + 121 + ], + "type": "inline_equation", + "content": "\\pi_{tj}^{*} \\in \\mathbb{R}^{HW \\times HW}" + }, + { + "bbox": [ + 304, + 72, + 546, + 121 + ], + "type": "text", + "content": " between two uniform distributions to indicate patch correspondence:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 338, + 137, + 545, + 256 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 137, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 338, + 137, + 545, + 256 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\pi_ {t j}} \\sum_ {u = 1} ^ {H W} \\sum_ {v = 1} ^ {H W} - \\hat {C} _ {t j} [ u, v ] \\pi_ {t j} [ u, v ] \\\\ \\text {s . t .} \\sum_ {v = 1} ^ {H W} \\pi_ {t j} [ \\cdot , v ] = \\frac {1}{H W} \\mathbf {1} ^ {H W}, \\tag {8} \\\\ \\sum_ {u = 1} ^ {H W} \\pi_ {t j} [ u, \\cdot ] = \\frac {1}{H W} \\mathbf {1} ^ {H W}, \\\\ \\pi_ {t j} [ u, v ] \\geq 0 \\quad u, v \\in \\{1, 2, \\dots , H W \\}, \\\\ \\end{array}", + "image_path": "d3b330579e671363c42a5527777c42aa965afafa5a687bbb799d96cc9dcd48f3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 263, + 546, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 546, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 263, + 546, + 361 + ], + "type": "inline_equation", + "content": "u, v" + }, + { + "bbox": [ + 304, + 263, + 546, + 361 + ], + "type": "text", + "content": " denotes spatial index, " + }, + { + "bbox": [ + 304, + 263, + 546, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{1}^{HW}" + }, + { + "bbox": [ + 304, + 263, + 546, + 361 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 304, + 263, + 546, + 361 + ], + "type": "inline_equation", + "content": "HW" + }, + { + "bbox": [ + 304, + 263, + 546, + 361 + ], + "type": "text", + "content": "-dimensional vector of all ones. Note that it is feasible to adopt other marginal distribution formulations with prior knowledge, e.g., class-agnostic activation map [4], to facilitate training, with details discussed in Supplementary Material. We employ Sinkhorn-Knopp algorithm [18] to obtain the optimal transport matrix " + }, + { + "bbox": [ + 304, + 263, + 546, + 361 + ], + "type": "inline_equation", + "content": "\\pi_{t,j}^*" + }, + { + "bbox": [ + 304, + 263, + 546, + 361 + ], + "type": "text", + "content": ", and formulate the dense semantic alignment objective in the form of cross entropy:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 377, + 545, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 545, + 428 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s e m} = - \\sum_ {\\substack {t, j = 1 \\\\ t \\neq j}} ^ {T} \\sum_ {u, v = 1} ^ {H W} \\sum_ {n = 1} ^ {N} \\pi_ {t j} ^ {*} [ u, v ] \\hat {\\mathcal {M}} _ {j} [ n, v ] \\log \\mathcal {M} _ {t} [ n, u ]. \\tag{9}", + "image_path": "7efd2b0af1c14b3f3796e678739b1bdcb3972b33cb6c40ed442db90617b7a237.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 436, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 471 + ], + "type": "text", + "content": "By minimizing the weighted sum of cross entropy, we achieve temporally consistent semantic distributions among corresponding spatial areas." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 472, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 509 + ], + "type": "text", + "content": "Semantic Mask Regularization. To encourage the semantic centers to emphasize different visual contents and cover diverse semantics, we apply a simple regularization to " + }, + { + "bbox": [ + 304, + 472, + 545, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t" + }, + { + "bbox": [ + 304, + 472, + 545, + 509 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 345, + 525, + 545, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 525, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 345, + 525, + 545, + 563 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e g} = \\sum_ {t = 1} ^ {T} \\sum_ {\\substack {n, j = 1 \\\\ j \\neq n}} ^ {N} \\frac {\\mathcal {M} _ {t} [ n ] ^ {T} \\mathcal {M} _ {t} [ j ]}{\\| \\mathcal {M} _ {t} [ n ] \\| _ {2} \\| \\mathcal {M} _ {t} [ j ] \\| _ {2}}. \\tag{10}", + "image_path": "ea7dedca0e825e8b0739ccad8343d9d729d1edad69075e2b92dcb8b128855eaa.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 594 + ], + "type": "text", + "content": "This simple regularization suppresses the cosine similarity between different semantic masks to avoid collapse." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "Instance Representation Consistency. Apart from the semantic distributions, it is necessary to ensure that the representations of each object instance are temporally consistent. Given the instance representations " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_t\\in \\mathbb{R}^{N\\times P\\times HW}" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": ", as well as " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{O}}_t" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": " from teacher pathway, we need to first match corresponding instances between different timestamps. For illustration, considering " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "-th semantics of time " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "t,j" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": ", we adopt bipartite matching [9, 15] based on the cosine similarity between " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_t[n]" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{O}}_j[n]" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": " to generate one-to-one instance correspondence, with the matching function denoted as " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\varepsilon (\\cdot)" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16679" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "type": "text", + "content": "In this way, " + }, + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_t[n,p]" + }, + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{O}}_j[n,\\varepsilon(p)]" + }, + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "type": "text", + "content": " are considered as a positive pair to be aligned. However, for each video, there exists absent semantics and object occlusion such that the number of visible object instances varies. Thus, there could be redundant slots not attending to objects, and it is crucial to filter out these invalid instance representations to reduce distractions. Mathematically, for timestamp " + }, + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "type": "text", + "content": ", we introduce " + }, + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_t \\in \\{0,1\\}^{N \\times P}" + }, + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "type": "text", + "content": " as the valid instance indicator. For the " + }, + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "type": "text", + "content": "-th instance of " + }, + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "type": "text", + "content": "-th semantics to be valid, two criteria must be met: (1) The ratio of the " + }, + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "type": "text", + "content": "-th semantic area is above threshold " + }, + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\tau_1 = 0.2" + }, + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "type": "text", + "content": " to filter out non-existing semantics; (2) The instance representation is close to the semantic center representation with cosine similarity larger than " + }, + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\tau_2 = 0.5" + }, + { + "bbox": [ + 46, + 71, + 289, + 240 + ], + "type": "text", + "content": " to exclude redundant slots. The conditions are formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 254, + 287, + 305 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 254, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 67, + 254, + 287, + 305 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {t} [ n, p ] = 1 \\Leftrightarrow \\left\\{ \\begin{array}{l} \\frac {1}{H W} \\sum_ {u = 1} ^ {H W} \\widetilde {\\mathcal {M}} _ {t} [ n, u ] \\geq \\tau_ {1}, \\\\ \\cos \\left(\\mathcal {S} _ {t} [ n ], \\mathcal {O} _ {t} [ n, p ]\\right) \\geq \\tau_ {2}. \\end{array} \\right. \\tag {11}", + "image_path": "2d3f0e1f311ec58529f3bb1de50653dca6f4427e52267c977d90782850933596.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 309, + 288, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 309, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 309, + 288, + 346 + ], + "type": "text", + "content": "And " + }, + { + "bbox": [ + 47, + 309, + 288, + 346 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{L}}_j\\in \\{0,1\\}^{N\\times P}" + }, + { + "bbox": [ + 47, + 309, + 288, + 346 + ], + "type": "text", + "content": " is defined in the same manner. We use a margin loss to encourage object detail consistency over valid instance representations:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 351, + 287, + 446 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 287, + 446 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal{L}_{obj} = \\sum_{\\substack{t,j = 1\\\\ t\\neq j}}^{T}\\sum_{n = 1}^{N}\\sum_{p = 1}^{P}\\mathcal{I}_{t}[n,p]\\big\\{ \\\\ \\hat {\\mathcal {I}} _ {j} [ n, \\varepsilon (p) ] \\| \\mathcal {O} _ {t} [ n, p ] - \\hat {\\mathcal {O}} _ {j} [ n, \\varepsilon (p) ] \\| _ {2} \\tag {12} \\\\ + \\sum_{\\substack{q = 1\\\\ q\\neq \\varepsilon (p)}}^{P}\\mathrm{relu}(\\lambda -\\| \\mathcal{O}_{t}[n,p] - \\hat{\\mathcal{O}}_{j}[n,q]\\|_{2})\\} \\\\ \\end{array}", + "image_path": "30c8d8509f9c12a29e8ca48b27f71027e4ccd4d8d7a21e64472c5363c7488851.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 451, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 522 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 451, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 451, + 287, + 522 + ], + "type": "text", + "content": " is a margin hyper-parameter. Note that with the " + }, + { + "bbox": [ + 46, + 451, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_t" + }, + { + "bbox": [ + 46, + 451, + 287, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 451, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}_j" + }, + { + "bbox": [ + 46, + 451, + 287, + 522 + ], + "type": "text", + "content": " constraints, our formulation can handle object occlusions which lead to varying number of visible instances in different frames. By minimizing the margin loss, we encourage the model to distill discriminative and temporally consistent object-centric representations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 523, + 287, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 523, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 523, + 287, + 547 + ], + "type": "text", + "content": "Overall, we take the summation of three objectives for training:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 552, + 287, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 552, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 111, + 552, + 287, + 566 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {s e m}} + \\mathcal {L} _ {\\text {o b j}} + \\mathcal {L} _ {\\text {r e g}}. \\tag {13}", + "image_path": "71205a47d20b7cdcf8b7e36143092ee517c2bb134c0fe06ab543bc74e4057a38.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 570, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 287, + 594 + ], + "type": "text", + "content": "The student pathway " + }, + { + "bbox": [ + 47, + 570, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 570, + 287, + 594 + ], + "type": "text", + "content": " is updated with gradient descent, and the teacher parameters " + }, + { + "bbox": [ + 47, + 570, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 47, + 570, + 287, + 594 + ], + "type": "text", + "content": " are updated with momentum as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 121, + 609, + 287, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 609, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 121, + 609, + 287, + 624 + ], + "type": "interline_equation", + "content": "\\hat {\\theta} \\leftarrow m \\hat {\\theta} + (1 - m) \\theta , \\tag {14}", + "image_path": "4f8582e572d90d513d254ec617a13fd87318f20f01f243f64cd0d095f2e11094.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": " is momentum coefficient set to 0.999 in default. This momentum update mechanism results in a slowly evolving teacher network, which progressively distills object knowledge, provides reliable self-supervision signals and enables us to train semantic-aware masked slot attention without relying on widely adopted reconstruction objective in existing works [59, 50, 88, 84, 72, 46]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 71, + 388, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 388, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 388, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 91, + 364, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 364, + 102 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 364, + 102 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 109, + 547, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 547, + 289 + ], + "type": "text", + "content": "We train our model on YouTube-VOS [87], a challenging video dataset that contains multiple object instances of distinct semantics in each video. We evaluate our method on two lines of tasks: (1) Unsupervised object segmentation on DAVIS-2016 [66], SegTrack-v2 [54], FMBS-59 [64] and challenging multiple object segmentation on DAVIS-2017-Unsupervised [8]. We respectively calculate the mean per frame intersection over union (IoU) and " + }, + { + "bbox": [ + 304, + 109, + 547, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\&\\mathcal{F}" + }, + { + "bbox": [ + 304, + 109, + 547, + 289 + ], + "type": "text", + "content": " on single and multiple object discovery benchmarks. (2) Lable propagation tasks including semi-supervised video object segmentation on DAVIS-2017 [67], human pose tracking on JHMDB [45], human part tracking on VIP [90]. We adopt the same settings as [44, 43] and report standard " + }, + { + "bbox": [ + 304, + 109, + 547, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{J}" + }, + { + "bbox": [ + 304, + 109, + 547, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 109, + 547, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 109, + 547, + 289 + ], + "type": "text", + "content": " score on DAVIS, probability of a correct pose (PCK) on JHMDB and mean intersection over union (mIoU) on VIP." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 297, + 440, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 297, + 440, + 309 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 440, + 309 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 316, + 546, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 546, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 546, + 435 + ], + "type": "text", + "content": "We sample " + }, + { + "bbox": [ + 304, + 316, + 546, + 435 + ], + "type": "inline_equation", + "content": "T = 4" + }, + { + "bbox": [ + 304, + 316, + 546, + 435 + ], + "type": "text", + "content": " frames with frame rate FPS " + }, + { + "bbox": [ + 304, + 316, + 546, + 435 + ], + "type": "inline_equation", + "content": "= 4" + }, + { + "bbox": [ + 304, + 316, + 546, + 435 + ], + "type": "text", + "content": " as the input RGB sequence, where each frame is augmented with random crop, horizontal flip and color jitter, and finally resized to " + }, + { + "bbox": [ + 304, + 316, + 546, + 435 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 316, + 546, + 435 + ], + "type": "text", + "content": ". We adopt ViT-Small/16 [23] and ResNet-50 [40] as the visual encoder, specified in each experiment, to extract frame-wise features. Then for semantic-aware masked slot attention, we set the number of learnable Gaussian distributions to " + }, + { + "bbox": [ + 304, + 316, + 546, + 435 + ], + "type": "inline_equation", + "content": "N = 16" + }, + { + "bbox": [ + 304, + 316, + 546, + 435 + ], + "type": "text", + "content": ", the number of instances of each semantics to " + }, + { + "bbox": [ + 304, + 316, + 546, + 435 + ], + "type": "inline_equation", + "content": "P = 4" + }, + { + "bbox": [ + 304, + 316, + 546, + 435 + ], + "type": "text", + "content": " in default and use 3 iterations to update slot representations and attention maps." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 436, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 604 + ], + "type": "text", + "content": "In training, the visual encoder is initialized with self-supervised pre-trained weights from DINO ViT-Small/16 [12] or MoCo ResNet-50 [38]. We use AdamW optimizer [61] with learning rate " + }, + { + "bbox": [ + 304, + 436, + 545, + 604 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 436, + 545, + 604 + ], + "type": "text", + "content": " for batch size 128 to update student parameters, and the teacher pathway is updated in momentum. In inference, we employ the same evaluation protocol as [44] to evaluate the temporal object correspondence performance on label propagation tasks. And for object discovery, we maintain the valid instances filtered with Eq. 11 as candidate objects. For single object evaluation, we merge all candidate objects as foreground areas to calculate IoU. For multiple object benchmark, we follow [8] to match ground-truth and our predictions and report " + }, + { + "bbox": [ + 304, + 436, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\& \\mathcal{F}" + }, + { + "bbox": [ + 304, + 436, + 545, + 604 + ], + "type": "text", + "content": " score." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 611, + 486, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 611, + 486, + 624 + ], + "spans": [ + { + "bbox": [ + 305, + 611, + 486, + 624 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-art" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Single Object Discovery. To measure our model's ability to decompose different objects, we first present the quantitative results on single object discovery without post processing (e.g., CRF, spectral clustering) in Table 2. Many existing works resort to optical flow [89, 53, 88, 84] or synthetic data [84] as weak supervision to learn temporal dynamics and generate semantic-agnostic segmentation masks for" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16680" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 288, + 205 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 205 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 205 + ], + "type": "table", + "html": "
ModelRGBFlowDAVISST-v2FBMS
CIS [89]✓✓71.562.563.5
AMD [57]✓X57.857.047.5
DINO [12]✓X52.346.550.3
SIMO [53]X✓67.862.0-
MG [88]X✓68.358.653.1
OCLR [84]X✓72.167.665.4
GWM [16]✓✓71.269.066.9
SMTC✓X71.869.368.4
\\( SMTC^† \\)✓X70.868.466.5
", + "image_path": "74915c80b8afafea9848b2410f8c537e12d5a07caee08cd07f2c42f0f5b7a3e5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 259, + 288, + 348 + ], + "blocks": [ + { + "bbox": [ + 47, + 206, + 287, + 251 + ], + "lines": [ + { + "bbox": [ + 47, + 206, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 287, + 251 + ], + "type": "text", + "content": "Table 2. Quantitative results on single object discovery. We compare per frame mean IoU on DAVIS-2016, SegTrack-v2 and FBMS-59 without any post-processing. " + }, + { + "bbox": [ + 47, + 206, + 287, + 251 + ], + "type": "inline_equation", + "content": "\\mathsf{SMTC}^{\\dagger}" + }, + { + "bbox": [ + 47, + 206, + 287, + 251 + ], + "type": "text", + "content": " denotes only with first slot attention stage for semantic decomposition in inference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 259, + 288, + 348 + ], + "lines": [ + { + "bbox": [ + 48, + 259, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 259, + 288, + 348 + ], + "type": "table", + "html": "
ModelBackboneJ&FJF
DINOSAUR [72]ViT-S/1621.419.223.7
SMTCViT-S/1640.536.444.6
SMTCResNet-5039.035.542.6
RVOS [76]ResNet-10141.236.845.7
ProReduce [56]ResNet-10168.365.071.6
", + "image_path": "243e8f7e53a13f3659749e48c6f79e8126816d114352e82806decb9861caa6fb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 348, + 287, + 371 + ], + "lines": [ + { + "bbox": [ + 47, + 348, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 287, + 371 + ], + "type": "text", + "content": "Table 3. Quantitative results on multiple object discovery on DAVIS-2017-Unsupervised. Gray denotes supervised training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 45, + 390, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 390, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 45, + 390, + 287, + 714 + ], + "type": "text", + "content": "moving objects. DINO baseline [12] uses the self-attention map from the last layer as an object segmentation mask. While our method only relies on RGB frames and explicitly discriminates object semantics in a fully self-supervised manner. From the comparison, we observe that our SMTC with only semantic decomposition largely exceeds RGB-only methods [12, 57] and comparable to optical flow-based methods, and our formulation with instance identification further improves the performance. This phenomenon indicates that the learned semantic decomposition with temporal correspondence cues have the potential to substitute precomputed optical flow to guide single object segmentation. Multiple Object Discovery. Most of the existing video segmentation methods with fully unsupervised training cannot explicitly distinguish multiple object instances. OCLR [84] adopts layered flow representations to identify multiple objects on re-annotated DAVIS-2017-motion dataset, but it cannot distinguish instances with common motion. On the contrary, our method jointly utilizes semantics and temporal correspondence to separate arbitrary object instances on DAVIS-2017-Unsupervised benchmark. For fair comparison, we re-run DINOSAUR [72] on DAVIS-2017 with 5 slots as an unsupervised baseline. As presented in Table 3, our method significantly outperforms DINOSAUR. This is because our method explicitly decomposes different semantics, exploits temporal correspondence among multiple frames to distinguish instances and encourages instance" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": "level temporal consistency. While DINOSAUR is only supervised with DINO feature reconstruction without considering temporal relationships and semantic discrimination. Besides, our method is comparable with supervised baseline RVOS [76], and the gap to supervised state-of-the-art [56] is majorly due to human annotated masks as supervision leading to much more precise segmentation boundaries. This conjecture can be demonstrated by Fig. 3, which also reveals that our model is able to learn discriminative object-centric representations to separate different instances." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 193, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 545, + 337 + ], + "type": "text", + "content": "Label Propagation. Finally, we compare the performance on label propagation tasks in Table 4 to validate the temporal consistency of our learned features. Though not specifically designed for these tasks, our model achieves state-of-the-art results on three benchmarks. The most related work to our SMTC is SFC [43], which also utilizes both high-level semantic and low-level correspondence. The difference is that SFC only employs late fusion to incorporate these two cues in inference, while our model jointly exploits semantics and temporal correspondence in training to identify object instances, perform instance-level alignment and refine temporally consistent object-centric representations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 346, + 400, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 400, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 400, + 358 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 365, + 545, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 400 + ], + "type": "text", + "content": "We report IoU on DAVIS-2016, and " + }, + { + "bbox": [ + 304, + 365, + 545, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\&\\mathcal{F}" + }, + { + "bbox": [ + 304, + 365, + 545, + 400 + ], + "type": "text", + "content": " on DAVIS-2017-Unsupervised for single and multiple object discovery. Refer to Supplementary Material for more ablations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 401, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 544 + ], + "type": "text", + "content": "Slot Attention Formulation. We compare the different formulations of semantic-aware masked slot attention in Table 5. With only the first slot attention stage, i.e., the 'Semantic' part, the model cannot distinguish object instances of the same semantics, thus the performance on multiple object discovery drops significantly as indicated by Ours-B. While with only the second stage, i.e., the 'Instance' part, the performance on both single and multiple object segmentation drops as indicated by Ours-C. This reveals that the pre-computed semantic masks play an important role in alleviating interference in instance slot update and filtering valid sample for instance-level alignment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 545, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 689 + ], + "type": "text", + "content": "Number of Learnable Gaussian Distributions. Comparing Ours-A, Ours-E and Ours-F in Table 5, we observe that the performance improves when " + }, + { + "bbox": [ + 304, + 545, + 545, + 689 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 545, + 545, + 689 + ], + "type": "text", + "content": " increases since larger " + }, + { + "bbox": [ + 304, + 545, + 545, + 689 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 545, + 545, + 689 + ], + "type": "text", + "content": " guides our model to distill more fine-grained semantics. And we present an extreme setting with " + }, + { + "bbox": [ + 304, + 545, + 545, + 689 + ], + "type": "inline_equation", + "content": "N = 1" + }, + { + "bbox": [ + 304, + 545, + 545, + 689 + ], + "type": "text", + "content": " in Ours-D, where our formulation degenerates into the original slot attention with random sampling from one Gaussian distribution [59]. Note that although Ours-C also has no access to semantic masks, it still maintains the potential to discriminate different semantics due to multiple learnable means. Its superiority to Ours-D demonstrates the effectiveness of semantic discrimination in object discovery." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Feature Usage. We explore using different feature input to our semantic-aware masked slot attention in Table 6. We" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16681" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 88, + 70, + 506, + 251 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 506, + 251 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 506, + 251 + ], + "type": "table", + "html": "
ModelBackboneDAVISJHMDBVIP
J&FJFPCK@0.1PCK@0.2mIoU
Supervised [40]ResNet-5066.063.768.459.278.339.5
MoCo [38]ResNet-5065.463.267.660.479.336.1
VFS [86]ResNet-5068.966.571.360.980.743.2
DINO [12]ViT-S/1661.860.263.445.475.237.9
TimeCycle [81]ResNet-5040.741.939.457.778.528.9
UVC [55]ResNet-5056.354.558.156.576.634.2
MAST [51]ResNet-1865.563.367.6---
CRW [44]ResNet-1867.664.870.258.880.337.6
SFC [43]ResNet-18+ResNet-5071.268.374.061.983.038.4
SMTCViT-S/1667.664.171.253.279.639.2
SMTCResNet-5073.069.476.662.584.138.8
", + "image_path": "2e991ca17115f8e6d9d1a5316ff1385b34d19b4ff7406228e0854913bf4017c1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 55, + 291, + 279, + 396 + ], + "blocks": [ + { + "bbox": [ + 47, + 251, + 545, + 275 + ], + "lines": [ + { + "bbox": [ + 47, + 251, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 545, + 275 + ], + "type": "text", + "content": "Table 4. Quantitative results on label propagation tasks: semi-supervised video object segmentation on DAVIS-2017, pose tracking on JHMDB, human part tracking on VIP. We present the backbone for comprehensive comparison, and report comparing results from [86, 43]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 291, + 279, + 396 + ], + "lines": [ + { + "bbox": [ + 55, + 291, + 279, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 279, + 396 + ], + "type": "table", + "html": "
ModelNSemanticMaskedIoUJ&F
Ours-A16✓✓71.840.5
Ours-B16✓X66.424.5
Ours-C16X✓52.130.3
Ours-D1-✓47.422.1
Ours-E8✓✓68.437.9
Ours-F32✓✓71.940.7
", + "image_path": "22c366b2462887e8ff4e437ad3aa5886c4e1a7fb3cd60b8dd7c41fccaa81a877.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 396, + 286, + 440 + ], + "lines": [ + { + "bbox": [ + 47, + 396, + 286, + 440 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 286, + 440 + ], + "type": "text", + "content": "Table 5. Ablation studies on the number of learnable Gaussian distributions, and the formulation of semantic-aware masked slot attention. 'Semantic' ('Instance') denotes the first (second) attention stage for semantic decomposition (instance identification)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 55, + 454, + 279, + 571 + ], + "blocks": [ + { + "bbox": [ + 55, + 454, + 279, + 571 + ], + "lines": [ + { + "bbox": [ + 55, + 454, + 279, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 279, + 571 + ], + "type": "table", + "html": "
FeatureSemanticInstanceIoUJ&F
Fused✓✓71.840.5
RGB-only✓✓67.720.1
✓×67.518.8
×✓41.515.3
Correlation-only✓✓38.814.2
✓×21.110.7
×✓59.427.5
", + "image_path": "638ed2c7d3ac7867504703ab9de585d5f0d4d604b119a2c18184f98b1906a9fb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 571, + 286, + 594 + ], + "lines": [ + { + "bbox": [ + 47, + 571, + 286, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 571, + 286, + 594 + ], + "type": "text", + "content": "Table 6. Ablation studies on the feature usage. We compare different slot attention formulations on different feature inputs." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": "observe that with RGB-only input, the 'Semantic' part guarantees satisfactory performance on single object discovery, but performs poor on multiple object benchmarks even with the 'Instance' part. This coincides with our intuition that it is difficult to distinguish object instances with solely semantic cues, instead we need to look into more frames and resort to temporal correspondence. While for the correlation-only input, the 'Instance' part coarsely segments objects and" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 344, + 291, + 507, + 353 + ], + "blocks": [ + { + "bbox": [ + 344, + 291, + 507, + 353 + ], + "lines": [ + { + "bbox": [ + 344, + 291, + 507, + 353 + ], + "spans": [ + { + "bbox": [ + 344, + 291, + 507, + 353 + ], + "type": "table", + "html": "
LsemLregLobjIoUJ&F
✓XX61.321.6
✓✓X66.424.5
XX✓50.630.4
✓✓✓71.740.5
", + "image_path": "58193a7cbbe98de680f926172e4fb7141545481b1e678e4b592cbe2510928ba5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 369, + 384, + 484, + 446 + ], + "blocks": [ + { + "bbox": [ + 306, + 353, + 545, + 374 + ], + "lines": [ + { + "bbox": [ + 306, + 353, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 306, + 353, + 545, + 374 + ], + "type": "text", + "content": "Table 7. Ablation studies on the learning objectives. We report the results on DAVIS-2016 and DAVIS-2017-Unsupervised." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 369, + 384, + 484, + 446 + ], + "lines": [ + { + "bbox": [ + 369, + 384, + 484, + 446 + ], + "spans": [ + { + "bbox": [ + 369, + 384, + 484, + 446 + ], + "type": "table", + "html": "
TFPSIoUJ&F
2465.436.9
4266.337.7
4471.840.5
8471.940.8
", + "image_path": "01df299f42619f1a2af5bada16f016a24e292ae388de20451732167ae9c373b3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 446, + 545, + 478 + ], + "lines": [ + { + "bbox": [ + 306, + 446, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 446, + 545, + 478 + ], + "type": "text", + "content": "Table 8. Ablation studies on the hyper-parameters of frame sampling. We compare different number of frames per clip and different FPS." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 497, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 545 + ], + "type": "text", + "content": "the 'Semantic' part exerts negative impact. This is because the correspondence map contains little semantic information, the semantic decomposition results in severe ambiguity, thus impairing the performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 546, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 689 + ], + "type": "text", + "content": "Learning Objectives. We also present a comprehensive ablation on the learning objectives in Table 7, where we report single object discovery on DAVIS-2016 and multiple object discovery results on DAVIS-2017-Unsupervised. From the comparison, we observe that " + }, + { + "bbox": [ + 304, + 546, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sem}" + }, + { + "bbox": [ + 304, + 546, + 545, + 689 + ], + "type": "text", + "content": " is fundamental to decompose object semantics, which demonstrates the necessity of semantic decomposition in the first stage. " + }, + { + "bbox": [ + 304, + 546, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 546, + 545, + 689 + ], + "type": "text", + "content": " also leads to improvement by encouraging the learnable queries to cover more diverse semantics. As for " + }, + { + "bbox": [ + 304, + 546, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{obj}" + }, + { + "bbox": [ + 304, + 546, + 545, + 689 + ], + "type": "text", + "content": ", this objective introduces instance-level alignment and significantly enhances multiple object discovery performance on DAVIS-2017-Unsupervised." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "Frame Sampling. Another important thing in training is the frame sampling procedure. The number of frames per" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16682" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 72, + 287, + 135 + ], + "blocks": [ + { + "bbox": [ + 52, + 72, + 287, + 135 + ], + "lines": [ + { + "bbox": [ + 52, + 72, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 52, + 72, + 287, + 135 + ], + "type": "image", + "image_path": "321920c59f87079ae3354195f90e07dba2a40c7ad71b4348a4a4c2136517e82f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 138, + 286, + 160 + ], + "lines": [ + { + "bbox": [ + 47, + 138, + 286, + 160 + ], + "spans": [ + { + "bbox": [ + 47, + 138, + 286, + 160 + ], + "type": "text", + "content": "Figure 3. Visualization of semantic and instance segmentation map. The red boxes outline the ambiguous areas." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 169, + 287, + 215 + ], + "blocks": [ + { + "bbox": [ + 48, + 169, + 287, + 215 + ], + "lines": [ + { + "bbox": [ + 48, + 169, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 169, + 287, + 215 + ], + "type": "image", + "image_path": "5c852b66a77f8f4dba7488c70f5cf5273325a7f16a637a4829c96daa9cffb1c0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 218, + 286, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 218, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 286, + 239 + ], + "type": "text", + "content": "Figure 4. Visualization of instance alignment. The arrows point out the matched instances across time." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 257, + 286, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 286, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 286, + 353 + ], + "type": "text", + "content": "clip and sampling frame rate determines the temporal reception field of our model. We compare different sampling hyper-parameters in Table 8. The results indicate that it is necessary to use large FPS to provide rich temporal dynamics. And by setting comparatively large FPS, our model can have access to more temporal information without introducing more frames, reaching a trade-off between performance and efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 360, + 160, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 160, + 372 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 160, + 372 + ], + "type": "text", + "content": "4.5. Further Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 378, + 286, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 286, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 286, + 605 + ], + "type": "text", + "content": "Visualization of Object Discovery. We visualize the our semantic decomposition map and instance identification map in Fig. 3, where the same color denotes the segmentation mask of the same semantics. The visualization reveals that our method is able to distinguish multiple object instances with semantic structure. For example, different people belong to the same semantic center, and quadrupeds such as dogs and pigs belong to another semantics. And our model also separates different instances of the same semantics, such as multiple persons and different pigs, with minor ambiguity on small object part, e.g., pig legs, mobile phones. Besides, we also visualize our instance-level alignment during training in Fig. 4. Our bipartite matching well aligns the corresponding instances in different frames, and the valid sample constraint of Eq. 11 effectively handles object occlusion in videos. For example, when encountering an occluded person instance, our method filters out the redundant slots not attending to objects, only correlating valid slots for robust instance-level alignment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 606, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 286, + 712 + ], + "type": "text", + "content": "Limitation. Due to the lack of the human annotated segmentation masks, our model faces challenges in generating precise boundaries for each instance, particularly for small objects. One potential solution is to incorporate multi-scale feature pyramid to further improve dense perception. We leave it in the future work. Despite this limitation, our work effectively demonstrates the benefits of leveraging both semantics and temporal correspondence to discover object instances with semantic structure and distill discriminative" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 533, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 533, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 533, + 84 + ], + "type": "text", + "content": "and temporally consistent object-centric representations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 95, + 378, + 107 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 95, + 378, + 107 + ], + "spans": [ + { + "bbox": [ + 306, + 95, + 378, + 107 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 115, + 545, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 115, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 304, + 115, + 545, + 355 + ], + "type": "text", + "content": "In this work, we propose a novel self-supervised framework jointly exploiting high-level semantics and low-level temporal correspondence to enhance object-centric perception. Specifically, we represent semantic and temporal correspondence cues using the RGB feature map and dense feature correlation, respectively. These cues are fused and fed into semantic-aware masked slot attention which comprises a set of learnable Gaussian distributions. This design allows us to leverage the mean vectors as potential semantic centers for semantic decomposition, and use the perturbations introduced by the standard deviation vectors around the semantic centers to make use of the temporal correspondence cues for instance identification. To distill discriminative and temporally consistent object-centric representations, we devise semantic- and instance-level alignment that is robust to object occlusion as self-supervision. We demonstrate the effectiveness of our model for object-centric analysis through the state-of-the-art performance on label propagation tasks, as well as the promising results on unsupervised object discovery in both single and multiple object scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 366, + 404, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 366, + 404, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 366, + 404, + 380 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 386, + 545, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 545, + 459 + ], + "type": "text", + "content": "We thank Tianfan Xue for constructive feedback. This project is funded in part by Shanghai AI Laboratory (P23KS00020, 2022ZD0160201), CUHK Interdisciplinary AI Research Institute, and the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 469, + 365, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 469, + 365, + 481 + ], + "spans": [ + { + "bbox": [ + 306, + 469, + 365, + 481 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 310, + 489, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 310, + 489, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 489, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 310, + 489, + 545, + 543 + ], + "type": "text", + "content": "[1] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. Advances in Neural Information Processing Systems, 33:9758-9770, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 545, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 545, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 311, + 545, + 545, + 588 + ], + "type": "text", + "content": "[2] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. arXiv preprint arXiv:1911.05371, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 590, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 590, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 311, + 590, + 545, + 633 + ], + "type": "text", + "content": "[3] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 311, + 635, + 545, + 689 + ], + "type": "text", + "content": "[4] Kyungjune Baek, Minhyun Lee, and Hyunjung Shim. Psynet: Self-supervised approach to object localization using point symmetric transformation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10451-10459, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 311, + 691, + 545, + 713 + ], + "type": "text", + "content": "[5] Beril Besbinar and Pascal Frossard. Self-supervision by prediction for object discovery in videos. In 2021 IEEE Interna" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16683" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 94 + ], + "type": "text", + "content": "tional Conference on Image Processing (ICIP), pages 1509-1513. IEEE, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 95, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 95, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 53, + 95, + 287, + 149 + ], + "type": "text", + "content": "[6] Zhangxing Bian, Allan Jabri, Alexei A Efros, and Andrew Owens. Learning pixel trajectories with multiscale contrastive random walks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6508-6519, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 150, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 150, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 287, + 194 + ], + "type": "text", + "content": "[7] Christopher P Burgess, Loic Matthey, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matt Botvinick, and Alexander Lerchner. Monet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 239 + ], + "type": "text", + "content": "[8] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv preprint arXiv:1905.00737, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 287, + 304 + ], + "type": "text", + "content": "[9] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 287, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 349 + ], + "type": "text", + "content": "[10] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 349, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 403 + ], + "type": "text", + "content": "[11] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "text", + "content": "[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 459, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 503 + ], + "type": "text", + "content": "[13] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 548 + ], + "type": "text", + "content": "[14] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 590 + ], + "type": "text", + "content": "[15] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "text", + "content": "[16] Subhabrata Choudhury, Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Guess what moves: unsupervised video and image segmentation by anticipating motion. arXiv preprint arXiv:2205.07844, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 680 + ], + "type": "text", + "content": "[17] Eric Crawford and Joelle Pineau. Exploiting spatial invariance for scalable unsupervised object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 3684-3692, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[18] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "type": "text", + "content": "[19] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9716-9726, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "text", + "content": "[20] Shuangrui Ding, Rui Qian, and Hongkai Xiong. Dual contrastive learning for spatio-temporal representation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5649-5658, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 186, + 545, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 231 + ], + "type": "text", + "content": "[21] Shuangrui Ding, Weidi Xie, Yabo Chen, Rui Qian, Xiaopeng Zhang, Hongkai Xiong, and Qi Tian. Motion-inductive self-supervised object discovery in videos. arXiv preprint arXiv:2210.00221, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 233, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 233, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 233, + 545, + 277 + ], + "type": "text", + "content": "[22] Shuangrui Ding, Peisen Zhao, Xiaopeng Zhang, Rui Qian, Hongkai Xiong, and Qi Tian. Prune spatio-temporal tokens by semantic-aware temporal accumulation. arXiv preprint arXiv:2308.04549, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 280, + 545, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 280, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 307, + 280, + 545, + 345 + ], + "type": "text", + "content": "[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 347, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 347, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 347, + 545, + 414 + ], + "type": "text", + "content": "[24] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 416, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 545, + 470 + ], + "type": "text", + "content": "[25] Gamaleldin F Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael C Mozer, and Thomas Kipf. Savi++: Towards end-to-end object-centric learning from real-world videos. arXiv preprint arXiv:2206.07764, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 473, + 545, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 473, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 307, + 473, + 545, + 528 + ], + "type": "text", + "content": "[26] Patrick Emami, Pan He, Sanjay Ranka, and Anand Rangarajan. Efficient iterative amortized inference for learning symmetric and disentangled multi-object representations. In International Conference on Machine Learning, pages 2970-2981. PMLR, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 530, + 545, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 530, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 307, + 530, + 545, + 574 + ], + "type": "text", + "content": "[27] Martin Engelcke, Adam R Kosiorek, Oiwi Parker Jones, and Ingmar Posner. Genesis: Generative scene inference and sampling with object-centric latent representations. arXiv preprint arXiv:1907.13052, 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 576, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 576, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 576, + 545, + 620 + ], + "type": "text", + "content": "[28] Martin Engelcke, Oiwi Parker Jones, and Ingmar Posner. Genesis-v2: Inferring unordered object representations without iterative refinement. Advances in Neural Information Processing Systems, 34:8085-8094, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 622, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 622, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 622, + 545, + 678 + ], + "type": "text", + "content": "[29] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3299-3309, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[30] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16684" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[31] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 172 + ], + "type": "text", + "content": "[32] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthey, Matthew Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative variational inference. In International Conference on Machine Learning, pages 2424-2433. PMLR, 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "type": "text", + "content": "[33] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 287, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 295 + ], + "type": "text", + "content": "[34] Kai Han, Rafael S Rezende, Bumsub Ham, Kwan-Yee K Wong, Minsu Cho, Cordelia Schmid, and Jean Ponce. Sc net: Learning semantic correspondence. In Proceedings of the IEEE international conference on computer vision, pages 1831-1840, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "type": "text", + "content": "[35] Tengda Han, Weidi Xie, and Andrew Zisserman. Video representation learning by dense predictive coding. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 385 + ], + "type": "text", + "content": "[36] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. Advances in Neural Information Processing Systems, 33:5679-5690, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 441 + ], + "type": "text", + "content": "[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 443, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 443, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 443, + 287, + 497 + ], + "type": "text", + "content": "[38] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 499, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 532 + ], + "type": "text", + "content": "[39] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 533, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 577 + ], + "type": "text", + "content": "[40] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "text", + "content": "[41] David Held, Sebastian Thrun, and Silvio Savarese. Learning to track at 100 fps with deep regression networks. In European conference on computer vision, pages 749-765. Springer, 2016." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 655 + ], + "type": "text", + "content": "[42] Olivier Henaff. Data-efficient image recognition with contrastive predictive coding. In International conference on machine learning, pages 4182-4192. PMLR, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "text", + "content": "[43] Yingdong Hu, Renhao Wang, Kaifeng Zhang, and Yang Gao. Semantic-aware fine-grained correspondence. arXiv preprint arXiv:2207.10456, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 692, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 712 + ], + "type": "text", + "content": "[44] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "in neural information processing systems, 33:19545-19560, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 97, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 545, + 140 + ], + "type": "text", + "content": "[45] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 143, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 545, + 175 + ], + "type": "text", + "content": "[46] Baoxiong Jia, Yu Liu, and Siyuan Huang. Unsupervised object-centric learning with bi-level optimized query slot attention. arXiv preprint arXiv:2210.08990, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 178, + 545, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 178, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 307, + 178, + 545, + 210 + ], + "type": "text", + "content": "[47] Longlong Jing, Xiaodong Yang, Jingen Liu, and Yingli Tian. Self-supervised spatiotemporal feature learning via video rotation prediction. arXiv preprint arXiv:1811.11387, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 213, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 213, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 213, + 545, + 277 + ], + "type": "text", + "content": "[48] Rishabh Kabra, Daniel Zoran, Goker Erdogan, Loic Matthew, Antonia Creswell, Matt Botvinick, Alexander Lerchner, and Chris Burgess. Simone: View-invariant, temporally-abstracted object representations via unsupervised video decomposition. Advances in Neural Information Processing Systems, 34:20146-20159, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 280, + 545, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 280, + 545, + 333 + ], + "spans": [ + { + "bbox": [ + 307, + 280, + 545, + 333 + ], + "type": "text", + "content": "[49] Dahun Kim, Donghyeon Cho, Donggeun Yoo, and In So Kweon. Learning image representations by completing damaged jigsaw puzzles. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 793-802. IEEE, 2018." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 336, + 545, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 336, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 307, + 336, + 545, + 390 + ], + "type": "text", + "content": "[50] Thomas Kipf, Gamaleldin F Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional object-centric learning from video. arXiv preprint arXiv:2111.12594, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 393, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 437 + ], + "type": "text", + "content": "[51] Zihang Lai, Erika Lu, and Weidi Xie. Mast: A memory-augmented self-supervised tracker. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6479-6488, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 439, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 439, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 439, + 545, + 470 + ], + "type": "text", + "content": "[52] Zihang Lai and Weidi Xie. Self-supervised learning for video correspondence flow. arXiv preprint arXiv:1905.00875, 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 474, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 474, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 474, + 545, + 495 + ], + "type": "text", + "content": "[53] Hala Lamdouar, Weidi Xie, and Andrew Zisserman. Segmenting invisible moving objects. 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 498, + 545, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 541 + ], + "type": "text", + "content": "[54] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In Proceedings of the IEEE international conference on computer vision, pages 2192-2199, 2013." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 544, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 586 + ], + "type": "text", + "content": "[55] Xueting Li, Sifei Liu, Shalini De Mello, Xiaolong Wang, Jan Kautz, and Ming-Hsuan Yang. Joint-task self-supervised learning for temporal correspondence. Advances in Neural Information Processing Systems, 32, 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "type": "text", + "content": "[56] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1739-1748, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 635, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 677 + ], + "type": "text", + "content": "[57] Runtao Liu, Zhirong Wu, Stella Yu, and Stephen Lin. The emergence of objectness: Learning zero-shot segmentation from videos. Advances in Neural Information Processing Systems, 34:13137-13152, 2021." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[58] Songtao Liu, Zeming Li, and Jian Sun. Self-emd: Self-supervised object detection withoutImagenet. arXiv preprint arXiv:2011.13677, 2020." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16685" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[59] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in Neural Information Processing Systems, 33:11525-11538, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 174 + ], + "type": "text", + "content": "[60] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "text", + "content": "[61] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 253 + ], + "type": "text", + "content": "[62] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European conference on computer vision, pages 527-544. Springer, 2016." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "type": "text", + "content": "[63] Joe Yue-Hei Ng, Jonghyun Choi, Jan Neumann, and Larry S Davis. Actionflownet: Learning motion representation for action recognition. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1616–1624. IEEE, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "type": "text", + "content": "[64] Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. IEEE transactions on pattern analysis and machine intelligence, 36(6):1187-1200, 2013." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 388 + ], + "type": "text", + "content": "[65] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 453 + ], + "type": "text", + "content": "[66] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "type": "text", + "content": "[67] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 501, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 545 + ], + "type": "text", + "content": "[68] Rui Qian, Shuangrui Ding, Xian Liu, and Dahua Lin. Static and dynamic concepts for self-supervised video representation learning. In European Conference on Computer Vision, pages 145-164. Springer, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 546, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 611 + ], + "type": "text", + "content": "[69] Rui Qian, Yuxi Li, Huabin Liu, John See, Shuangrui Ding, Xian Liu, Dian Li, and Weiyao Lin. Enhancing self-supervised video representation learning via multi-level feature optimization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7990-8001, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "text", + "content": "[70] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6964-6974, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[71] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 127 + ], + "type": "text", + "content": "[72] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, et al. Bridging the gap to real-world object-centric learning. arXiv preprint arXiv:2209.14860, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 546, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 182 + ], + "type": "text", + "content": "[73] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8934-8943, 2018." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 183, + 546, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 183, + 546, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 183, + 546, + 237 + ], + "type": "text", + "content": "[74] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 238, + 546, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 238, + 546, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 238, + 546, + 293 + ], + "type": "text", + "content": "[75] Jack Valmadre, Luca Bertinetto, Joao Henriques, Andrea Vedaldi, and Philip HS Torr. End-to-end representation learning for correlation filter based tracking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2805-2813, 2017." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 293, + 546, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 293, + 546, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 293, + 546, + 348 + ], + "type": "text", + "content": "[76] Carles Ventura, Miriam Bellver, Andreu Girbau, Amaia Salvador, Ferran Marques, and Xavier Giro-i Nieto. Rvos: End-to-end recurrent network for video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5277-5286, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 349, + 546, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 546, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 546, + 381 + ], + "type": "text", + "content": "[77] Paul Voigtlaender and Bastian Leibe. Online adaptation of convolutional neural networks for video object segmentation. arXiv preprint arXiv:1706.09364, 2017." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 383, + 546, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 546, + 426 + ], + "type": "text", + "content": "[78] Carl Vondrick, Abhinav Shrivastava, Alireza Fathi, Sergio Guadarrama, and Kevin Murphy. Tracking emerges by colorizing videos. In Proceedings of the European conference on computer vision (ECCV), pages 391-408, 2018." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 426, + 546, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 546, + 459 + ], + "type": "text", + "content": "[79] Naiyan Wang and Dit-Yan Yeung. Learning a deep compact image representation for visual tracking. Advances in neural information processing systems, 26, 2013." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 460, + 546, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 546, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 546, + 503 + ], + "type": "text", + "content": "[80] Xiaolong Wang, Kaiming He, and Abhinav Gupta. Transitive invariance for self-supervised visual representation learning. In Proceedings of the IEEE international conference on computer vision, pages 1329-1338, 2017." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 504, + 546, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 546, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 546, + 548 + ], + "type": "text", + "content": "[81] Xiaolong Wang, Allan Jabri, and Alexei A Efros. Learning correspondence from the cycle-consistency of time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2566-2576, 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 548, + 546, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 546, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 546, + 602 + ], + "type": "text", + "content": "[82] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 603, + 546, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 546, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 546, + 656 + ], + "type": "text", + "content": "[83] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 546, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 546, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 546, + 690 + ], + "type": "text", + "content": "[84] Junyu Xie, Weidi Xie, and Andrew Zisserman. Segmenting moving objects via an object-centric layered representation. arXiv preprint arXiv:2207.02206, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 691, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 546, + 713 + ], + "type": "text", + "content": "[85] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16686" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 418 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 66, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 287, + 106 + ], + "type": "text", + "content": "consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[86] Jiarui Xu and Xiaolong Wang. Rethinking self-supervised correspondence learning: A video frame-level similarity perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10075-10085, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 217 + ], + "type": "text", + "content": "[87] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen, and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 585-601, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 272 + ], + "type": "text", + "content": "[88] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7177-7188, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 287, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 328 + ], + "type": "text", + "content": "[89] Yanchao Yang, Antonio Loquercio, Davide Scaramuzza, and Stefano Soatto. Unsupervised moving object detection via contextual information separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 879-888, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 373 + ], + "type": "text", + "content": "[90] Qixian Zhou, Xiaodan Liang, Ke Gong, and Liang Lin. Adaptive temporal encoding network for video instance-level human parsing. In Proceedings of the 26th ACM international conference on Multimedia, pages 1527-1535, 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 375, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 375, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 375, + 287, + 418 + ], + "type": "text", + "content": "[91] Xizhou Zhu, Yujie Wang, Jifeng Dai, Lu Yuan, and Yichen Wei. Flow-guided feature aggregation for video object detection. In Proceedings of the IEEE international conference on computer vision, pages 408-417, 2017." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16687" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning/fd85c784-c77b-49cb-91b7-f3a6f94e50c3_content_list.json b/2023/Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning/fd85c784-c77b-49cb-91b7-f3a6f94e50c3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0be9b565af9db265b97c37bfbd053b1cb31123fc --- /dev/null +++ b/2023/Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning/fd85c784-c77b-49cb-91b7-f3a6f94e50c3_content_list.json @@ -0,0 +1,1566 @@ +[ + { + "type": "text", + "text": "Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning", + "text_level": 1, + "bbox": [ + 241, + 130, + 728, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kaiyou Song Shan Zhang Zimeng Luo Tong Wang Jin Xie Megvii Technology", + "bbox": [ + 217, + 203, + 746, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{songkaiyou, zhangshan, luozimeng, wangtong, xiejin}@megvii.com", + "bbox": [ + 205, + 250, + 764, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 297, + 313, + 314 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrastive self-supervised learning, the common way to learn discriminative representation is to pull different augmented \"views\" of the same image closer while pushing all other images further apart, which has been proven to be effective. However, it is unavoidable to construct undesirable views containing different semantic concepts during the augmentation procedure. It would damage the semantic consistency of representation to pull these augmentations closer in the feature space indiscriminately. In this study, we introduce feature-level augmentation and propose a novel semantics-consistent feature search (SCFS) method to mitigate this negative effect. The main idea of SCFS is to adaptively search semantics-consistent features to enhance the contrast between semantics-consistent regions in different augmentations. Thus, the trained model can learn to focus on meaningful object regions, improving the semantic representation ability. Extensive experiments conducted on different datasets and tasks demonstrate that SCFS effectively improves the performance of self-supervised learning and achieves state-of-the-art performance on different downstream tasks.", + "bbox": [ + 76, + 332, + 473, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 680, + 209, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to the tremendous potential in learning discriminative feature representation without using data annotations, self-supervised learning, including contrastive learning [16, 5] and masked image modeling (MIM) [1, 15, 38] has received much attention in the representation learning field. Contrastive learning, as a type of discriminative self-supervised learning method, is heavily studied and has shown remarkable progress in the computer vision field in recent years. It aims at pulling different augmented \"views\" of the same image (positive pairs) closer while pushing diverse images (negative pairs) far from each other. To this end, a contrastive", + "bbox": [ + 76, + 707, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4d603971121f051b93cdb76af88d0e95709712c0cdf30fedf59da5215805e844.jpg", + "image_caption": [ + "Figure 1. Semantic inconsistency of over-augmentation. (a) shows three augmentations of two images, in which the third augmentation is over-augmented and contains only background. (b) shows category probability distributions of the corresponding images in (a), which are obtained from a supervised pre-trained ResNet50 [18] model. (c)(d)(e) show three different samples of an image (the data-augmented image, the original image, and the semantics-consistent feature-augmented sample generated by Eq. (7) in this study) and their corresponding probability distributions, which point out that the over-augmented image generates different category with the original image, while the feature-augmented sample gets a balanced category probability." + ], + "image_footnote": [], + "bbox": [ + 511, + 297, + 880, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "loss between the features of different views extracted from an encoder network is employed to train the encoder network end-to-end. According to whether the negative pairs are used, current contrastive learning can be generally divided into two categories.", + "bbox": [ + 496, + 731, + 893, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The first category [16, 5] utilizes both positive pairs and negative pairs for contrast. MoCo [16, 7] uses a momentum update mechanism to maintain a memory bank of negative examples. SimCLR [5, 6] directly trains a single encoder network with a large batch size to ensure sufficient positive and negative samples for learning. Based on MoCo and Sim", + "bbox": [ + 496, + 809, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Code: https://github.com/skyoux/scfs", + "bbox": [ + 94, + 886, + 377, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16099", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CLR, some methods [24, 20, 36, 10, 49, 39, 19, 9, 48, 14] are proposed to improve the performance. For example, MSF [20], ISD [36] and NNCLR [10] aim to search semantics-consistent samples for contrast, solving the false negative problem. While some studies, such as Momentum2Teacher [24] and DCL [49], aim to solve the limitation that large batch size is necessary for satisfactory performance.", + "bbox": [ + 75, + 90, + 470, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The second category of contrastive learning methods [13, 8, 3, 50, 2, 12, 4, 31, 35] only constructs positive pairs for contrast. Based on MoCo [16] and SimCLR [5], respectively, BYOL [13] and SimSiam [8] abandon the negative samples and use an asymmetric architecture to avoid model collapse. SwAV [3] uses online clustering to cluster samples and forces the consistency among cluster assignments of different augmentations. After that, some studies [4, 39, 10, 19] point out that enriching the augmented samples can improve the performance of contrastive learning. In addition, the study in [31] shows that improving the quality of positive augmented samples is important for self-supervised learning.", + "bbox": [ + 75, + 199, + 470, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, it is unavoidable to construct data augmentations containing different semantic concepts. Fig. 1(a) shows three augmentations of two images, in which the third augmentation is over-augmented and contains only the background. Fig. 1(b) shows category probability distributions of the corresponding images in (a), which are obtained from a supervised pre-trained ResNet50 [18]. We observed that the probability distribution of over-augmented images changes greatly compared with the first two augmentations, which indicates that the semantic information of the over-augmented images deviates from the normally-augmented images. Similar observation can be found in Fig. 1(c)(d). The original image in (d) shows a max probability for \"ambulance\", while the over-augmented image in (c) represents the different category \"telescope\". Due to such semantic inconsistency, conducting contrastive learning on these overaugmentations is harmful to representation learning. In this study, we found that semantics-consistent feature augmentation (Fig. 1(e), generated by Eq. (7)) can balance the original semantics \"ambulance\" and the over-augmented semantics \"telescope\", which can alleviate the influence of semantic inconsistency.", + "bbox": [ + 75, + 383, + 470, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by this observation, we propose a novel semantics-consistent feature search (SCFS) method to alleviate the negative influence of semantic inconsistency in contrastive learning. SCFS utilizes the global feature of a view to adaptively search the semantics-consistent features of another view for contrast according to their similarity. It constructs informative feature augmentations and conducts contrast learning between feature augmentations and data augmentations. Thus, the pre-trained model can learn to focus on meaningful object regions to alleviate the negative influence of unmatched semantic alignment in current contrastive learning for better representation learning. In", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "addition, the feature search is conducted on multiple layers of the backbone network, further enhancing the semantic alignment at different scales of features. Extensive experiments conducted on different datasets and tasks demonstrate that SCFS effectively improves the performance of self-supervised learning and achieves state-of-the-art performance on different downstream tasks. For example, it achieves state-of-the-art $75.7\\%$ ImageNet top-1 accuracy under the pre-training setting of 1024 batch size and 800 epochs for ResNet50.", + "bbox": [ + 496, + 90, + 893, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this study are threefold:", + "bbox": [ + 517, + 242, + 851, + 257 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel contrastive learning method, i.e., SCFS, is proposed, and it can enhance semantic alignment in contrastive learning. To our knowledge, this is the first work that defines a feature search task in contrastive learning.", + "- We expand contrastive learning from a data-to-data manner to a feature-to-data manner, which enriches the diversity of augmentations.", + "- The proposed SCFS achieves state-of-the-art performance on different downstream tasks." + ], + "bbox": [ + 517, + 267, + 890, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 452, + 648, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, some studies [33, 44, 42, 45, 46, 25, 47, 40, 23, 41] pointed out that the problem of semantic inconsistency is more serious for downstream dense prediction tasks, such as object detection and instance segmentation. Therefore, these methods utilize region-level and pixel-level features for contrast. In this study, the proposed SCFS construct feature-level augmentations using dense feature maps. Therefore, this section introduces related studies that conduct contrastive learning using region-level and pixel-level features.", + "bbox": [ + 496, + 477, + 893, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Region-level contrastive learning. SCRL [33] minimizes the distance between two local features, which are cropped from two corresponding feature maps of two views. ReSim [44] aligns regional representations by sliding a fixed-sized window across the overlapping area between two views to improve the performance for localization-based tasks. SoCo [42], ORL [46], and UniVIP [25] extract object region proposals and use them to construct region-level features for contrastive learning. They achieve good performance for downstream dense prediction tasks.", + "bbox": [ + 496, + 614, + 893, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pixel-level contrastive learning. To obtain a more fine-grained representation, several studies [47, 40, 41] design pixel-level contrastive learning task, which assumes that features extracted from the same pixel of different views should be treated as positive pairs while pixels from others must be distinguished. PixPro [47] utilizes a pixel propagation module to select similar pixel features for contrast and encourages consistency between positive pixel pairs. DenseCL [40] proposes a dense projection head to generate dense", + "bbox": [ + 496, + 765, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16100", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "feature vectors for pixel-level contrastive learning. SetSim [41] is designed to realize pixel-wise similarity learning by filtering out noisy backgrounds.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As summarized above, data augmentations bring rich information while increasing uncertainty in contrastive learning. While methods that utilize region-level features expand the granularity of feature representation by alleviating the influence of noises. Recently, to solve the problem of semantic inconsistency in video contrastive learning, VITO [30] uses an MLP to learn attention masks for each view independently and fuses the attention pooling features from multi-layers for contrastive learning. Unlike VITO [30], SCFS uses a cross-view search to calculate attention and conducts contrastive learning at multi-layers independently. In addition, our method bridges the correlation between data and feature augmentations and extends the contrastive learning task to a semantics-consistent feature search task.", + "bbox": [ + 75, + 137, + 470, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 76, + 362, + 174, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall architecture of SCFS is shown in Fig. 2. It consists of an encoder and a momentum encoder. The momentum encoder is an exponential-moving-average version of the encoder. SCFS consists of two contrastive learning tasks: the contrast between data augmentations $(\\mathcal{L}_d)$ , which is designed based on the baseline DINO [4], and the contrast between data augmentations and feature augmentations $(\\mathcal{L}_{fs})$ . In this section, we first introduce the contrast between data augmentations in Sec. 3.1. Then, the key feature search module of SCFS is introduced in detail in Sec. 3.2. Next, we will introduce the overall loss of SCFS in Sec. 3.3. Finally, the implementation details are presented in Sec. 3.4.", + "bbox": [ + 75, + 387, + 470, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Contrast Between Data Augmentations", + "text_level": 1, + "bbox": [ + 76, + 579, + 413, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a pair of global augmentations $(I_1$ and $I_{2}$ ) of an input image, the feature representations of the two augmentations are used to calculate the global contrastive loss. Specifically, $\\pmb {f}_1 = E_\\pmb {\\theta}(\\pmb {I}_1)$ and $\\pmb{f^{\\prime}}_2 = E_{\\pmb{\\theta}^{\\prime}}(\\pmb {I}_2)$ , where $\\pmb{\\theta}$ and $\\pmb{\\theta}^{\\prime}$ are parameters of the encoder and the momentum encoder, respectively. $\\pmb {f}_1,\\pmb{f}^{\\prime}_2\\in R^K$ $K$ is the output dimension. $\\pmb{f}_{1}$ is normalized with a softmax function:", + "bbox": [ + 75, + 603, + 468, + 708 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP _ {1} ^ {i} = \\frac {\\exp \\left(f _ {1} ^ {i} / \\tau\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(f _ {1} ^ {k} / \\tau\\right)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 717, + 468, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau > 0$ is a temperature parameter that controls the sharpness of the output distribution. Note that $P_{2}^{\\prime}$ is obtained by normalizing $f_{2}^{\\prime}$ with a similar softmax function with temperature $\\tau^{\\prime}$ . $I_{1}$ and $I_{2}$ are fed to the momentum encoder and encoder symmetrically, and $P_{1}^{\\prime}$ and $P_{2}$ are obtained respectively. Following DINO [4], the cross-entropy loss is employed as the contrastive loss between two global views:", + "bbox": [ + 75, + 767, + 468, + 872 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g} = - \\left(P _ {2} ^ {\\prime} \\log \\left(P _ {1}\\right) + P _ {1} ^ {\\prime} \\log \\left(P _ {2}\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 883, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enrich augmentations, the multi-crop strategy [3] is employed. Multiple local augmentations $\\mathbf{I}_l$ is also constructed and fed to the encoder: $\\mathbf{f}_l = E_{\\boldsymbol{\\theta}}(\\mathbf{I}_l)$ . $P_l$ is obtained by normalizing $\\mathbf{f}_l$ with the softmax function with temperature $\\tau$ . The contrast between local views and global views can be calculated:", + "bbox": [ + 496, + 90, + 893, + 181 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l} = \\sum_ {n = 1} ^ {N} - \\left(P _ {1} ^ {\\prime} \\log \\left(P _ {l} ^ {n}\\right) + P _ {2} ^ {\\prime} \\log \\left(P _ {l} ^ {n}\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 189, + 892, + 231 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N$ denotes the number of local views. Thus, the overall loss is the sum of global loss and local loss:", + "bbox": [ + 496, + 239, + 890, + 268 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = \\mathcal {L} _ {g} + \\mathcal {L} _ {l} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 281, + 892, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Semantics-Consistent Feature Search", + "text_level": 1, + "bbox": [ + 500, + 306, + 821, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As introduced in Sec. 1, it is unavoidable to construct augmentations that contain different semantic concepts during the augmentation procedure. It's harmful to pull these augmentations close indiscriminately in the feature space. Therefore, we propose the SCFS method to enhance the contrast between semantics-consistent regions in different augmentations.", + "bbox": [ + 496, + 329, + 893, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The architecture of SCFS is shown in Fig. 2. By feeding the local augmentations $I_{l}$ to the encoder, feature maps from different stages of the backbone ResNet50 [18] are extracted. Specifically, the output features from different stages, i.e., $Res2$ , $Res3$ and $Res4$ , are utilized to conduct SCFS, ensuring that each stage of the backbone produces discriminative features: $\\{^{2}F_{l},^{3}F_{l},^{4}F_{l}\\} = E_{\\theta}(I_{l})$ , where $^i F_l\\in R^{W_l^i\\times H_l^i\\times C^i}$ , $W_{l}^{i},H_{l}^{i},C^{i}$ denote the width, height and channel dimension, respectively. Next, the global average pooling operation is conducted on each $^i F_{l}$ in the spatial dimensions:", + "bbox": [ + 496, + 435, + 893, + 603 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n{ } ^ { i } \\boldsymbol { f } _ { l } = \\frac { 1 } { W _ { l } ^ { i } \\times H _ { l } ^ { i } } \\sum _ { x = 1 } ^ { W _ { l } ^ { i } } \\sum _ { y = 1 } ^ { H _ { l } ^ { i } } { } ^ { i } \\boldsymbol { F } _ { l } ( x , y , z ) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 613, + 892, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where ${}^i f_l \\in R^{C^i}$ . Meanwhile, the global augmentations $I_g$ ( $g = 1,2$ ) are fed to the momentum encoder to extract feature maps from different stages: $\\{^2 F_g',^3 F_g',^4 F_g'\\} = E_{\\pmb{\\theta}'}(\\pmb {I}_g),{}^i F_g' \\in R^{W_g^i\\times H_g^i\\times C^i}, W_g^i,H_g^i,C^i$ denote width, height and channel dimension, respectively.", + "bbox": [ + 496, + 670, + 890, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, based on ${}^i\\pmb {f}_l$ and ${}^i\\pmb{F}_{g}^{\\prime}$ , SCFS aims to adaptively search the most semantics-consistent features in ${}^i\\pmb{F}_{g}^{\\prime}$ for contrast, while suppressing irrelevant features. In SCFS, each feature ${}^i\\pmb {f}_l$ of the local data augmentations is treated as query, and the features ${}^i\\pmb{F}_{g}^{\\prime}$ of the global augmentations are treated as keys. The similarity between ${}^i\\pmb {f}_l$ and ${}^i\\pmb{F}_{g}^{\\prime}$ is calculated:", + "bbox": [ + 496, + 750, + 893, + 857 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {A} (x, y) = \\frac {{} ^ {i} \\boldsymbol {f} _ {l} \\cdot {} ^ {i} \\boldsymbol {F} _ {g} ^ {\\prime} (x , y)}{\\| ^ {i} \\boldsymbol {f} _ {l} \\| _ {2} \\| ^ {i} \\boldsymbol {F} _ {g} ^ {\\prime} (x , y) \\| _ {2}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 866, + 892, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16101", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2f47ebc92a19132eefcdbaca58e53be431c38d99e09fc3c383ded5eae72b6d7d.jpg", + "image_caption": [ + "Figure 2. Overall architecture of the proposed semantics-consistent feature search (SCFS). It consists of an encoder and a momentum encoder. There are two contrastive learning tasks: the contrast between data augmentations $(\\mathcal{L}_d)$ in the final feature space and the feature search task conducted on multiple layers $(\\mathcal{L}_{fs})$ . The details of the feature search procedure is shown on the right." + ], + "image_footnote": [], + "bbox": [ + 158, + 88, + 558, + 268 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e11192e29bee3c44aa0413a27f29c9325f5cabf6877fd1fcae5ca6defae4acc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 89, + 810, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{A} \\in R^{W_g^i \\times H_g^i}$ is the attention map, and $x = 1, \\ldots, W_g^i$ , $y = 1, \\ldots, H_g^i$ , $\\| \\cdot \\|_2$ is the L2 norm. The attention map $\\mathbf{A}$ activates the semantics-consistent regions of the local augmentation on the global augmentation. Thus, the higher portion of local regions can be searched. To select semantic features and suppress irrelevant local features, we directly multiply the attention map $\\mathbf{A}$ with ${}^i\\mathbf{F}_g'$ to obtain the semantics-consistent feature augmentations:", + "bbox": [ + 75, + 337, + 473, + 460 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n{ } ^ { i } \\boldsymbol { F } _ { l g } ^ { \\prime } = \\boldsymbol { A } \\cdot { } ^ { i } \\boldsymbol { F } _ { g } ^ { \\prime } ( 7 )\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 465, + 468, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This operation can be regarded as attention-weighted average pooling. Through feature search, $N$ local data augmentations $\\pmb{I}_l$ can search $N$ corresponding semantics-consistent features ${}^i\\pmb{F}_{lg}'$ from a global data augmentation $\\pmb{I}_g$ . That is, in terms of the global data augmentation, $N$ different features are constructed in the feature space through the feature search procedure. Therefore, we term ${}^i\\pmb{F}_{lg}'$ as feature-level augmentations.", + "bbox": [ + 75, + 491, + 468, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, ${}^i\\mathbf{F}_l$ and ${}^i\\mathbf{F}_{lg}'$ are fed to corresponding projection heads to obtain their final representations for contrast:", + "bbox": [ + 75, + 611, + 468, + 643 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} ^ {i} \\boldsymbol {f} _ {l} = \\mathrm {H} _ {i} \\left(^ {i} \\boldsymbol {F} _ {l}\\right) \\\\ ^ {i} \\boldsymbol {f} _ {l g} ^ {\\prime} = H _ {i} ^ {\\prime} \\left(^ {i} \\boldsymbol {F} _ {l g} ^ {\\prime}\\right) \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 648, + 468, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{H}_i$ and $\\mathrm{H}'_i$ denote the projection heads on the $i$ -th layer of the encoder and the momentum encoder, respectively. ${}^i\\pmb {f}_l$ and ${}^i\\pmb {f}_{lg}'$ are normalized with softmax function with temperature $\\tau$ and $\\tau'$ , respectively, as the same formulation in Eq. (1). The corresponding output probability ${}^iP_l$ and ${}^iP_{lg}'$ are employed to calculate the contrast loss between local data augmentations and feature augmentations:", + "bbox": [ + 75, + 688, + 472, + 794 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n{ } ^ { i } \\mathcal { L } _ { f s } = \\sum _ { g = 1 } ^ { 2 } \\sum _ { n = 1 } ^ { N } - \\left( ^ { i } P _ { l g } ^ { \\prime } \\log \\left( ^ { i } P _ { l } ^ { n } \\right) \\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 800, + 468, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall feature search loss is the sum of all layers:", + "bbox": [ + 96, + 849, + 457, + 864 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f s} = \\sum_ {i \\in V _ {L}} ^ {i} \\mathcal {L} _ {f s} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 869, + 468, + 904 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $V_{L}$ denotes the set of layers to conduct SCFS.", + "bbox": [ + 498, + 339, + 846, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Overall Loss", + "text_level": 1, + "bbox": [ + 500, + 373, + 633, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall loss is the sum of the contrastive loss between data augmentations and the feature search loss:", + "bbox": [ + 498, + 398, + 890, + 430 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {d} + \\mathcal {L} _ {f s} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 449, + 890, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Through SCFS, the contrast between feature augmentations and data augmentations is bridged. The model can adaptively search the semantics-consistent features for contrast. Therefore, it can enhance the importance of semantics-consistent regions in different augmentations, alleviating the uncertainty in contrastive learning introduced by data augmentations that contain different semantic concepts.", + "bbox": [ + 498, + 486, + 893, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 611, + 718, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SCFS is based on DINO [4] and we follow the most hyper-parameter settings of DINO. For a fair comparison, the standard ResNet50 [18] is employed as the default backbone network.", + "bbox": [ + 498, + 637, + 890, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For data augmentation, the global augmentations consist of random cropping, resizing to $224 \\times 224$ , random horizontal flip, gaussian blur, and color jittering. And the local augmentations consist of random cropping, resizing to $96 \\times 96$ , random horizontal flip, gaussian blur, and color jittering. For feature augmentations in SCFS, the Res2, Res3, and Res4 layers are used. Two global views with $N = 8$ local views are the default setting of augmentation.", + "bbox": [ + 496, + 700, + 893, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The projection head for the contrast between data augmentations consists of a four-layer multi-layer-perceptron (MLP) with the same architecture as DINO [4]. The projection head for feature search consists of three convolutional layers and two FC layers.", + "bbox": [ + 496, + 825, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16102", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/a261d5d96710eba8221daecb261b08400865a1aabfe57128aaf815ebbacea0bf.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Linear probing and $k$ -NN accuracy (\\%) on ImageNet. The result with “*” is reproduced for fair comparison. LP denotes linear probing. Bold font and underline indicate the best results under the setting of 256 batch size and 200 epochs and the setting of 1024 batch size and 800 epochs, respectively." + ], + "table_body": "
MethodBatch SizeEpochsLPk-NN
Supervised25610076.274.8
SimCLR [5]4096100069.3-
BYOL [13]4096100074.366.9
BYOL [13]409620070.6-
SwAV [3]409680075.3-
SwAV [3]25620072.7-
MoCo-v2 [16]25620067.554.3
SimSiam [8]25620070.0-
ISD [36]25620069.862.0
MSF [20]25620071.464.0
NNCLR [10]409620070.7-
Barlow Twins [50]2048100073.2-
VICReg [2]2048100073.2-
OBoW [12]25620073.8-
DCL [49]25620066.9-
CLSA [39]25620073.3-
AdCo [19]25620073.2-
DetCo [45]25620068.6-
UniVIP [25]409620073.1-
HCSC [14]25620073.3-
MoCo-v3 [9]409630072.8-
MoCo-v3 [9]4096100074.6-
DINO* [4]25620073.064.0
DINO [4]408080075.367.5
SCFS25620073.965.5
SCFS102480075.768.5
", + "bbox": [ + 84, + 88, + 460, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 625, + 209, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, comprehensive experiments are conducted to demonstrate the effectiveness of SCFS. We evaluate the performance on different downstream tasks, including ImageNet classification, object detection, instance segmentation, and other classification task on small datasets. In addition, we conduct ablation experiments to analyze the influence of each component in SCFS.", + "bbox": [ + 75, + 651, + 470, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Comparing with SSL methods on ImageNet", + "text_level": 1, + "bbox": [ + 76, + 768, + 447, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$k$ -NN and Linear Probing Accuracy on ImageNet. After pre-training on the ImageNet ILSVRC-2012 [34] training set, the pre-trained models are evaluated on the ImageNet ILSVRC-2012 validation set. For $k$ -NN, it is evaluated as in study [43]. For linear probing, we train a linear classifier from scratch based on the feature extracted by a fixed backbone with 100 epochs [16]. The top-1 accuracy is adopted", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/557e7ef00384ae78338e786dfa9c39cb02142d5ddddc4717c6cc15d21f3d672e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBatch Size EpochsTop-1Top-5
1%10%1%10%
Supervised [51]2569025.456.448.480.4
SimCLR [5]4096100048.365.675.587.8
BYOL [13]4096100053.268.878.489.0
SwAV [3]409680053.970.278.589.9
DINO [4]408080050.269.374.089.1
SCFS102480054.370.578.690.2
", + "bbox": [ + 509, + 88, + 883, + 232 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2588c583ac30b4edffc275433e0569e3afdec46d580cedca47138db6e2a5ed9d.jpg", + "table_caption": [ + "Table 2. Evaluation on small labeled ImageNet. Bold font indicates the best result." + ], + "table_footnote": [], + "table_body": "
MethodEpochsAPbAPb50APb75
Scratch-33.860.233.1
Supervised9053.581.358.8
SimCLR [5]100056.381.962.5
BYOL [13]30051.981.056.5
SwAV [3]40045.177.446.5
DINO [4]80055.982.162.3
SCFS80057.483.063.6
", + "bbox": [ + 511, + 279, + 880, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3. Results for PASCAL VOC object detection using Faster R-CNN [32] with ResNet50-C4. Bold font indicates the best result.", + "bbox": [ + 500, + 422, + 893, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as the evaluation metric.", + "bbox": [ + 500, + 474, + 663, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The results are reported in Tab. 1. With the standard ResNet50 [18] architecture and pre-trained with 256 batch size for 200 epoch, the proposed SCFS achieves the best $k$ -NN top-1 accuracy $65.5\\%$ and the best linear probing top-1 accuracy $73.9\\%$ , outperforming its baseline DINO [4] by $1.5\\%$ and $0.9\\%$ , respectively. In addition, with 1024 batch size and 800 epoch, SCFS achieves the best $k$ -NN accuracy $(68.5\\%)$ and linear probing accuracy $(75.7\\%)$ , outperforming the accuracy of DINO [4] trained with 4080 batch size for 800 epoch. This result demonstrates that SCFS can improve the representation learning performance by searching semantics-consistent features for contrast.", + "bbox": [ + 496, + 489, + 893, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semi-Supervised Learning on ImageNet. In this part, we evaluate the performance of SCFS under the semi-supervised setting. Specifically, we use $1\\%$ and $10\\%$ of the labeled training data from ImageNet [34] for finetuning, which follows the semi-supervised protocol in SimCLR [5]. The same splits of $1\\%$ and $10\\%$ of ImageNet labeled training data in SimCLRv2 [6] are used. The results are reported in Tab. 2. After finetuning using $1\\%$ and $10\\%$ training data, SCFS outperforms all the compared methods. The results demonstrate that SCFS achieves the best feature representation quality.", + "bbox": [ + 496, + 671, + 895, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Transfer Learning on Downstream Tasks", + "text_level": 1, + "bbox": [ + 500, + 832, + 851, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Object Detection and Instance Segmentation. In this part, we evaluate the representations of SCFS on dense prediction tasks, i.e., object detection and instance segmentation,", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16103", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9805798297ae8665ee55cf52316dd8d4159f48c59879cfe4b813a0bac8a47fab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodEpochs1×schedule2×schedule
\\(AP^b\\)\\(AP^b_{50}\\)\\(AP^b_{75}\\)\\(AP^s\\)\\(AP^s_{50}\\)\\(AP^s_{75}\\)\\(AP^b\\)\\(AP^b_{50}\\)\\(AP^b_{75}\\)\\(AP^s\\)\\(AP^s_{50}\\)\\(AP^s_{75}\\)
Scratch-31.049.533.228.546.830.438.457.542.034.754.837.2
Supervised9038.959.642.735.456.538.141.361.345.037.358.340.3
MoCo [16]20038.558.942.035.155.937.740.861.644.736.958.439.7
MoCo v2 [7]20040.460.244.236.457.238.941.761.645.637.658.740.5
BYOL [13]30040.461.644.137.258.839.842.362.646.238.359.641.1
SwAV [3]400------42.362.846.338.260.041.0
\\(ReSim-FPN^T\\) [44]20039.860.243.536.057.138.641.461.945.437.559.140.3
SetSim [41]20040.260.743.936.457.739.041.662.445.937.759.440.6
DenseCL [40]20040.359.944.336.457.039.241.261.945.137.358.940.1
DSC [23]20039.458.943.235.756.138.3------
HSA [48]80040.260.943.936.557.939.142.263.046.138.159.940.9
DetCo [45]80040.161.043.936.458.038.9------
ORL* [46]80040.360.244.436.357.338.9------
DINO [4]80040.061.643.436.558.639.141.962.646.037.859.740.6
SCFS80040.561.844.036.758.839.242.163.446.138.160.241.0
", + "bbox": [ + 127, + 88, + 841, + 363 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3ecca5e5e90135872ecaeddb8722b14257db43a983d5059353969ddff9008647.jpg", + "table_caption": [ + "Table 4. Object detection and instance segmentation on COCO using Mask R-CNN [17] with ResNet50-FPN. Bold font indicates the best result. * indicates pre-training on COCO [26]." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR-10CIFAR-100CUB-BirdStanford-CarsAircraftOxford-Pets
Supervised97.586.481.392.186.092.1
SimCLR[5]97.785.9-91.388.189.2
BYOL[13]97.886.1-91.688.191.7
DINO[4]*97.786.681.091.187.491.5
SCFS97.886.782.791.688.591.9
", + "bbox": [ + 187, + 411, + 784, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5. Transfer learning results from ImageNet with the standard ResNet50 [18]. * denotes the results are reproduced in this study. Bold font indicates the best result.", + "bbox": [ + 75, + 523, + 893, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "on mainstream datasets PASCAL VOC [11] and MS COCO [26] datasets. On the PASCAL VOC dataset [11], the trainval07+12 set is used as the training set, and the test2007 set is used as the test set. Following [42], Faster R-CNN detector [32] with the ResNet50-C4 backbone initialized by the self-supervised pre-trained model is trained end-to-end. On the COCO dataset, the train2017 set is used for training and the val2017 set is used for evaluation. The Mask R-CNN [17] with R50-FPN is used. The $\\mathrm{AP^{b}}$ , $\\mathrm{AP_{50}^{b}}$ and $\\mathrm{AP_{75}^{b}}$ metrics are used for object detection. While the $\\mathrm{AP^{s}}$ , $\\mathrm{AP_{50}^{s}}$ and $\\mathrm{AP_{75}^{s}}$ metrics are used for instance segmentation.", + "bbox": [ + 75, + 573, + 472, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The experimental results are shown in Tab. 3 and Tab. 4. SCFS achieves best performance on the two datasets. For example, on VOC, SCFS achieves $57.4\\% \\mathrm{AP}^{\\mathrm{b}}$ , $83.0\\% \\mathrm{AP}_{50}^{\\mathrm{b}}$ and $63.6\\% \\mathrm{AP}_{75}^{\\mathrm{b}}$ . The $\\mathrm{AP}^{\\mathrm{b}}$ of SCFS outperforms its baseline DINO by $1.5\\%$ . These results show that SCFS also has good transfer ability on dense prediction tasks.", + "bbox": [ + 75, + 744, + 470, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Other Classification Tasks. In this part, we focus on the performance of self-supervised models when they are finetuned on small datasets, including CIFAR [22] and fine grained datasets [21, 37, 28, 29]. The results are shown in Tab. 5.", + "bbox": [ + 75, + 839, + 472, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed SCFS shows the best performance on all the small datasets, which demonstrates that SCFS has good generalization ability.", + "bbox": [ + 496, + 574, + 893, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Pre-training on Uncurated Dataset", + "text_level": 1, + "bbox": [ + 498, + 633, + 803, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed SCFS can solve the problem of semantic inconsistency during pre-training, which is important when pre-training on uncurated datasets since this problem is more serious. To verify this, we pre-train SCFS and DINO on COCO [26], which is much more Uncurated than ImageNet. The same hyper-parameters used on ImageNet are applied to train the models with 512 batch size for 500 epochs. After pre-training, we fine-tune the pre-trained models on COCO for object detection and instance segmentation. The Mask R-CNN [17] with R50-FPN is used. As shown in Tab. 6, SCFS improves the performance significantly compared to its baseline DINO. In addition, compared to other dense pixel-level and region-level methods, such as DenseCL [40] and ORL [46], SCFS also achieves the best performance. This experiment verifies that SCFS can effectively solve the problem of semantic inconsistency during pre-training.", + "bbox": [ + 496, + 657, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16104", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1478d593ae9846acc2ccbcbeeebf2af6822945c5c83eacdd8dacbe41198dc827.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6. Pre-training and than Fine-tuning on COCO [26] using Mask R-CNN [17] with ResNet50-FPN and $1 \\times$ schedule. All models pre-trained on COCO are pre-trained with 512 batch size for 800 epochs. Bold font indicates the best result. * indicates pre-training on ImageNet [34]" + ], + "table_body": "
Methodk-NNAPbAPb50APb75APsAPs50APs75
Scratch-31.049.533.228.546.830.4
Supervised*-38.959.642.735.456.538.1
SimCLR [5]35.337.056.840.333.753.836.1
MoCov2 [7]39.438.558.142.134.855.337.3
BYOL [13]42.639.559.343.235.656.538.2
DenseCL [40]32.539.659.343.335.756.538.4
ORL [46]43.640.360.244.436.357.338.9
UniVIP [25]-40.8--36.8--
DINO [4]42.639.059.642.935.656.838.0
SCFS44.640.961.644.436.958.439.5
", + "bbox": [ + 81, + 88, + 460, + 266 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 371, + 240, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We analyze the influence of each component in SCFS. To speed up the training time, the ImageNet100 dataset, which contains 100 randomly selected categories from ImageNet [34], is adopted. All the models are pre-trained on the ImageNet100 training set with 256 batch size for 200 epoch, and tested on the validation set. The $k$ -NN and linear probing top-1 accuracy are used as the evaluation metrics.", + "bbox": [ + 75, + 396, + 468, + 501 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Influence of Different Contrast Modes. The contrast mode can be divided into three types: contrast between two global data augmentations used in all contrastive learning methods $(\\mathrm{G_d}2\\mathrm{G_d})$ ; contrast between local data augmentations and global data augmentations used in multi-crop strategy $(\\mathrm{L_d}2\\mathrm{G_d})$ ; and contrast between local data augmentations and local feature augmentations used in SCFS $(\\mathrm{L_d}2\\mathrm{L_f})$ .", + "bbox": [ + 75, + 503, + 468, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results are shown in Tab. 7. With multi-crop, DINO [4] (81.1%) improves accuracy by 3.0% compared to DINO without multi-crop baseline. SCFS (84.8%) further improves accuracy by 3.7% by introducing a contrast between local data augmentation and local feature augmentation. Some attention maps of SCFS and DINO are shown in Fig. 3. SCFS can more accurately focus on semantics-consistent regions between global view and local views, while DINO is easily influenced by background.", + "bbox": [ + 75, + 609, + 470, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also add multi-layer feature contrastive learning on DINO. The result in Tab. 7 (the \"DINO w ML\" row) verifies the improvements of SCFS are not totally owed to multi-layer contrast.", + "bbox": [ + 75, + 748, + 470, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, we directly crop the corresponding region of local augmentation on the feature map of global augmentation for contrastive learning. As shown in Tab. 7, this variant (ROIAlign) of SCFS also outperforms the DINO baseline, which shows that the directly cropped features are also beneficial for contrastive learning. And the ROIAlign variant", + "bbox": [ + 75, + 810, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/287e3074a6911c2fc81ac5c04e8a4c21eaf394efbd1df87d966a1160e6257505.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Contrast ModeGd2GdLd2GdLd2LfGd2Gfk-NNLP
DINO w/o MC✓78.183.7
DINO✓✓81.187.0
DINO w ML✓✓82.287.4
SCFS✓✓✓84.889.2
ROI Crop✓✓✓83.988.1
SCFS w/o MC✓79.786.3
", + "bbox": [ + 504, + 89, + 888, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Influence of different contrast modes. MC, ML, and LP denote multi-crop, multi-layer, and linear probing, respectively.", + "bbox": [ + 498, + 214, + 890, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f8aa03f2395644f4897a8537d84e58c10244e784f755028b6dc16c7b3747956.jpg", + "image_caption": [ + "Figure 3. Attention maps of SCFS (the third row) compared with DINO [4] (the second row). In each example, (a) shows a global image, and its four local images in (b) are constructed by $2 \\times 2$ jigsaw. (d) and (f) show the attention maps that highlight the semantics-consistent regions between the local images in (b) and the global image in (a). They are obtained by multiplying the globally average pooled feature maps from the encoder (Res4) of the local images in (b) with the feature map (Res4) of the global image in (a). And the encoder is the trained DINO ResNet50 model and SCFS ResNet50 model in (d) and (f), respectively. (c) and (e) show the mean attention maps of DINO and SCFS respectively, which are obtained by multiplying the mean globally average pooled feature map of the four local images in (b) with feature map of the global image in (a)." + ], + "image_footnote": [], + "bbox": [ + 511, + 261, + 875, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of SCFS achieves lower accuracy than SCFS, demonstrating that the soft feature search in SCFS is better than the hard ROIAlign since ROIAlign may damage the continuous semantic context of the feature map.", + "bbox": [ + 496, + 640, + 893, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Further, we also test the performance of SCFS under the setting without multi-crop. That is, the feature search is conducted between two global data augmentations. We term this contrast mode as $\\mathrm{G_d}2\\mathrm{G_f}$ . As shown in the \"SCFS w/o MC\" row, SCFS also improves the performance compared to its baseline (the \"DINO w/o MC\" row), which proves that SCFS is also helpful in solving the semantic inconsistency caused by other augmentations, not only the multi-crop augmentation strategy.", + "bbox": [ + 496, + 702, + 893, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Influence of Multi-Layer Contrast. The influence of the feature layer that is used for feature search is analyzed. The Res2, Res3 and Res4 in the ResNet50 [18] backbone are evaluated. As shown in Tab. 8, the performance improves", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16105", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2e8294c96dda461c7703e22846aa5179dd74bab00a7f7c92563d475e60abd0eb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Res2Res3Res4k-NNLinear Probing
✓82.086.1
✓✓84.387.5
✓✓✓84.889.2
", + "bbox": [ + 106, + 88, + 442, + 160 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/32cc7bcf55c389eba83a68856410513d9704b827fa327950f9ffe264b5128b8b.jpg", + "image_caption": [ + "Figure 4. The $k$ -NN accuracy of features from different layers." + ], + "image_footnote": [], + "bbox": [ + 114, + 196, + 428, + 345 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b5fcca0d8ad18777935c3d077f1178bacd21da419d0ae5694771f3ec4b1e8ecc.jpg", + "image_caption": [ + "Figure 5. Influence of local augmentation number." + ], + "image_footnote": [], + "bbox": [ + 114, + 377, + 428, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with the increase of feature layer numbers, which demonstrates that conducting feature search on more layers is helpful for representation learning.", + "bbox": [ + 75, + 566, + 470, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Further, we evaluate the $k$ -NN accuracy using feature maps from different layers to observe the influence of feature search on the representation of middle layers. We also choose the features extracted by the Res2, Res3 and Res4 layer of ResNet50. The results are shown in Fig. 4. Compared with DINO [4], SCFS achieves better performance with features from all middle layers on ImageNet100, which verifies that enhancing the semantic consistency can improve the semantic representation of shallow layers. Compared with supervised learning, SCFS has higher performance on res2 and res3 layer, which shows that SCFS is more advantageous in the shallow layer feature representation.", + "bbox": [ + 75, + 613, + 468, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Influence of Local Augmentation Number. In this part, we analyze the performance difference with the change of local augmentation numbers. The results are shown in Fig. 5. The performance of DINO and SCFS is steadily improved when adding more local augmentations for contrast. In addition, SCFS improves the performance under different local augmentation numbers, which demonstrates that semantics-", + "bbox": [ + 75, + 795, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e9c309b5021e66bfb4a1297ecb79190f00e75c004ec45b7e6280ea2c0d9d1573.jpg", + "table_caption": [ + "Table 8. Influence of different feature augmentation layer." + ], + "table_footnote": [], + "table_body": "
Cropping Scale0.05-0.140.14-0.250.25-0.40.4-0.6
DINO [4]87.087.285.682.6
SCFS89.288.787.083.1
", + "bbox": [ + 504, + 88, + 890, + 137 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c9da0c2713e13b0ec73f1c0914867db92bcfadc816e901347292fb8bf7915f3f.jpg", + "table_caption": [ + "Table 9. Influence of random cropping scale. We report the linear probing top-1 accuracy." + ], + "table_footnote": [], + "table_body": "
MethodBackboneBatch SizeEpochsk-NNLP
DINO [4]R10125620081.086.3
SCFSR10125620085.188.3
DINO [4]ViT-S-1625620075.080.4
SCFSViT-S-1625620076.381.0
DINO [4]ViT-B-1625620076.280.7
SCFSViT-B-1625620077.282.3
", + "bbox": [ + 504, + 189, + 883, + 308 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 10. Experiments on other backbones. LP denotes linear probing.", + "bbox": [ + 498, + 316, + 890, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "consistent feature search is helpful to alleviate the influence of semantics inconsistent data augmentations.", + "bbox": [ + 498, + 371, + 890, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Influence of Random Cropping Scale. The hyperparameter of the random cropping scale influences the degree of semantic inconsistency. In this experiment, we analyze the performance of SCFS and DINO [4] under different random cropping scales. As shown in Tab. 9, owing to the feature search, SCFS is more adaptive to different cropping scales than the baseline DINO [4].", + "bbox": [ + 498, + 402, + 892, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experiments on Other Backbones. In this part, we conduct experiments on other backbones to further evaluate the effectiveness of SCFS. Apart for the default Reset50 used in other experiments, ResNet101 and Vision Transformer (ViT-S and ViT-B) are tested. The results are shown in Tab. 10. SCFS achieves significant improvement on different backbones compared to its baseline DINO, which demonstrates that SCFS is applicable to different backbones.", + "bbox": [ + 498, + 510, + 893, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 647, + 625, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we aim to alleviate the problem of unmatched semantic alignment in current contrastive learning by expanding the augmentations from data space to feature space. The proposed semantics-consistent feature search (SCFS) adaptively searches semantics-consistent local features between different views for contrast, while suppressing irrelevant local features during pre-training. It conducts contrast learning between feature augmentation and data augmentation. The experimental results demonstrate that SCFS can learn to focus on meaningful object regions and effectively improve the performance of self-supervised learning. The feature search procedure in SCFS is learnable parameter-free. We will utilize the self-attention mechanism in Transformer to perform the feature search procedure to further boost its performance in future work.", + "bbox": [ + 496, + 674, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16106", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022. 1", + "[2] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021. 2, 5", + "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. NeurIPS, 33:9912-9924, 2020. 2, 3, 5, 6, 12", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 2, 3, 4, 5, 6, 7, 8, 11, 12", + "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. PMLR, 2020. 1, 2, 5, 6, 7, 12", + "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. NeurIPS, 33:22243-22255, 2020. 1, 5", + "[7] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 6, 7", + "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021. 2, 5", + "[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021. 2, 5, 12", + "[10] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In ICCV, pages 9588–9597, 2021. 2, 5", + "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 88(2):303–338, 2010. 6", + "[12] Spyros Gidaris, Andrei Bursuc, Gilles Puy, Nikos Komodakis, Matthieu Cord, and Patrick Perez. Obow: Online bag-of-visual-words generation for self-supervised learning. In CVPR, pages 6830–6840, 2021. 2, 5", + "[13] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 33:21271-21284, 2020. 2, 5, 6, 7, 12", + "[14] Yuanfan Guo, Minghao Xu, Jiawen Li, Bingbing Ni, Xuanyu Zhu, Zhenbang Sun, and Yi Xu. Hcsc: Hierarchical contrastive selective coding. In CVPR, pages 9706-9715, 2022. 2, 5", + "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 1, 2, 5, 6", + "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 6, 7", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 2, 3, 4, 5, 6, 7", + "[19] Qianjiang Hu, Xiao Wang, Wei Hu, and Guo-Jun Qi. Adco: Adversarial contrast for efficient learning of unsupervised representations from self-trained negative adversaries. In CVPR, pages 1074-1083, 2021. 2, 5", + "[20] Soroush Abbasi Koohpayegani, Ajinkya Tejankar, and Hamed Piri siavash. Mean shift for self-supervised learning. In ICCV, pages 10326-10335, 2021. 2, 5", + "[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 6", + "[22] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[23] Xiaoni Li, Yu Zhou, Yifei Zhang, Aoting Zhang, Wei Wang, Ning Jiang, Haiying Wu, and Weiping Wang. Dense semantic contrast for self-supervised visual representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1368-1376, 2021. 2, 6", + "[24] Zeming Li, Songtao Liu, and Jian Sun. Momentum^2 teacher: Momentum teacher with momentum statistics for self-supervised learning. arXiv preprint arXiv:2101.07525, 2021. 2", + "[25] Zhaowen Li, Yousong Zhu, Fan Yang, Wei Li, Chaoyang Zhao, Yingying Chen, Zhiyang Chen, Jiahao Xie, Liwei Wu, Rui Zhao, et al. Univip: A unified framework for self-supervised visual pre-training. In CVPR, pages 14627-14636, 2022. 2, 5, 7", + "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 6, 7", + "[27] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 11", + "[28] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6", + "[29] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In $CVPR$ , pages 3498-3505. IEEE, 2012. 6", + "[30] Nikhil Parthasarathy, SM Eslami, João Carreira, and Olivier J Henaff. Self-supervised video pretraining yields strong image representations. arXiv preprint arXiv:2210.06433, 2022. 3, 12", + "[31] Xiangyu Peng, Kai Wang, Zheng Zhu, Mang Wang, and Yang You. Crafting better contrastive views for siamese representation learning. In CVPR, pages 16031-16040, 2022. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16107", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 5, 6", + "[33] Byungseok Roh, Wuhyun Shin, Ildoo Kim, and Sungwooong Kim. Spatially consistent representation learning. In CVPR, pages 1144–1153, 2021. 2", + "[34] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015. 5, 7", + "[35] Kaiyou Song, Jin Xie, Shan Zhang, and Zimeng Luo. Multi-mode online knowledge distillation for self-supervised visual representation learning. In CVPR, 2023. 2", + "[36] Ajinkya Tejankar, Soroush Abbasi Koohpayegani, Vipin Pillai, Paolo Favaro, and Hamed Pirsiavash. Isd: Self-supervised learning by iterative similarity distillation. In ICCV, pages 9609-9618, 2021. 2, 5", + "[37] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6", + "[38] Haochen Wang, Kaiyou Song, Junsong Fan, Yuxi Wang, Jin Xie, and Zhaoxiang Zhang. Hard patches mining for masked image modeling. In CVPR, 2023. 1", + "[39] Xiao Wang and Guo-Jun Qi. Contrastive learning with stronger augmentations. IEEE TPAMI, 2022. 2, 5", + "[40] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In CVPR, pages 3024-3033, 2021. 2, 6, 7", + "[41] Zhaoqing Wang, Qiang Li, Guoxin Zhang, Pengfei Wan, Wen Zheng, Nannan Wang, Mingming Gong, and Tongliang Liu. Exploring set similarity for dense self-supervised representation learning. In CVPR, pages 16590-16599, 2022. 2, 3, 6", + "[42] Fangyun Wei, Yue Gao, Zhirong Wu, Han Hu, and Stephen Lin. Aligning pretraining for detection via object-level contrastive learning. NeurIPS, 34, 2021. 2, 6", + "[43] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, pages 3733-3742, 2018. 5", + "[44] Tete Xiao, Colorado J Reed, Xiaolong Wang, Kurt Keutzer, and Trevor Darrell. Region similarity representation learning. In ICCV, pages 10539–10548, 2021. 2, 6", + "[45] Enze Xie, Jian Ding, Wenhai Wang, Xiaohang Zhan, Hang Xu, Peize Sun, Zhenguo Li, and Ping Luo. Detco: Unsupervised contrastive learning for object detection. In ICCV, pages 8392-8401, 2021. 2, 5, 6", + "[46] Jiahao Xie, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Unsupervised object-level representation learning from scene images. NeurIPS, 34:28864-28876, 2021. 2, 6, 7", + "[47] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In CVPR, pages 16684-16693, 2021. 2", + "[48] Haohang Xu, Xiaopeng Zhang, Hao Li, Lingxi Xie, Wenrui Dai, Hongkai Xiong, and Qi Tian. Seed the views: Hierarchi" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "cal semantic alignment for contrastive representation learning. IEEE TPAMI, 2022. 2, 6", + "[49] Chun-Hsiao Yeh, Cheng-Yao Hong, Yen-Chi Hsu, Tyng-Luh Liu, Yubei Chen, and Yann LeCun. Decoupled contrastive learning. In ECCV, pages 668-684. Springer, 2022. 2, 5", + "[50] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, pages 12310-12320. PMLR, 2021. 2, 5", + "[51] Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. In ICCV, pages 1476-1485, 2019. 5" + ], + "bbox": [ + 503, + 92, + 892, + 247 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16108", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning/fd85c784-c77b-49cb-91b7-f3a6f94e50c3_model.json b/2023/Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning/fd85c784-c77b-49cb-91b7-f3a6f94e50c3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..61c37409aba760c1b6d353a5955c1fb9cc2da19d --- /dev/null +++ b/2023/Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning/fd85c784-c77b-49cb-91b7-f3a6f94e50c3_model.json @@ -0,0 +1,2156 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.243, + 0.131, + 0.73, + 0.177 + ], + "angle": 0, + "content": "Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.204, + 0.748, + 0.245 + ], + "angle": 0, + "content": "Kaiyou Song Shan Zhang Zimeng Luo Tong Wang Jin Xie Megvii Technology" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.25, + 0.766, + 0.265 + ], + "angle": 0, + "content": "{songkaiyou, zhangshan, luozimeng, wangtong, xiejin}@megvii.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.299, + 0.314, + 0.315 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.333, + 0.474, + 0.65 + ], + "angle": 0, + "content": "In contrastive self-supervised learning, the common way to learn discriminative representation is to pull different augmented \"views\" of the same image closer while pushing all other images further apart, which has been proven to be effective. However, it is unavoidable to construct undesirable views containing different semantic concepts during the augmentation procedure. It would damage the semantic consistency of representation to pull these augmentations closer in the feature space indiscriminately. In this study, we introduce feature-level augmentation and propose a novel semantics-consistent feature search (SCFS) method to mitigate this negative effect. The main idea of SCFS is to adaptively search semantics-consistent features to enhance the contrast between semantics-consistent regions in different augmentations. Thus, the trained model can learn to focus on meaningful object regions, improving the semantic representation ability. Extensive experiments conducted on different datasets and tasks demonstrate that SCFS effectively improves the performance of self-supervised learning and achieves state-of-the-art performance on different downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.681, + 0.21, + 0.697 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.708, + 0.472, + 0.875 + ], + "angle": 0, + "content": "Due to the tremendous potential in learning discriminative feature representation without using data annotations, self-supervised learning, including contrastive learning [16, 5] and masked image modeling (MIM) [1, 15, 38] has received much attention in the representation learning field. Contrastive learning, as a type of discriminative self-supervised learning method, is heavily studied and has shown remarkable progress in the computer vision field in recent years. It aims at pulling different augmented \"views\" of the same image (positive pairs) closer while pushing diverse images (negative pairs) far from each other. To this end, a contrastive" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.299, + 0.882, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.537, + 0.896, + 0.704 + ], + "angle": 0, + "content": "Figure 1. Semantic inconsistency of over-augmentation. (a) shows three augmentations of two images, in which the third augmentation is over-augmented and contains only background. (b) shows category probability distributions of the corresponding images in (a), which are obtained from a supervised pre-trained ResNet50 [18] model. (c)(d)(e) show three different samples of an image (the data-augmented image, the original image, and the semantics-consistent feature-augmented sample generated by Eq. (7) in this study) and their corresponding probability distributions, which point out that the over-augmented image generates different category with the original image, while the feature-augmented sample gets a balanced category probability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.895, + 0.809 + ], + "angle": 0, + "content": "loss between the features of different views extracted from an encoder network is employed to train the encoder network end-to-end. According to whether the negative pairs are used, current contrastive learning can be generally divided into two categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.896, + 0.902 + ], + "angle": 0, + "content": "The first category [16, 5] utilizes both positive pairs and negative pairs for contrast. MoCo [16, 7] uses a momentum update mechanism to maintain a memory bank of negative examples. SimCLR [5, 6] directly trains a single encoder network with a large batch size to ensure sufficient positive and negative samples for learning. Based on MoCo and Sim" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.378, + 0.901 + ], + "angle": 0, + "content": "1Code: https://github.com/skyoux/scfs" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16099" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.198 + ], + "angle": 0, + "content": "CLR, some methods [24, 20, 36, 10, 49, 39, 19, 9, 48, 14] are proposed to improve the performance. For example, MSF [20], ISD [36] and NNCLR [10] aim to search semantics-consistent samples for contrast, solving the false negative problem. While some studies, such as Momentum2Teacher [24] and DCL [49], aim to solve the limitation that large batch size is necessary for satisfactory performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.472, + 0.382 + ], + "angle": 0, + "content": "The second category of contrastive learning methods [13, 8, 3, 50, 2, 12, 4, 31, 35] only constructs positive pairs for contrast. Based on MoCo [16] and SimCLR [5], respectively, BYOL [13] and SimSiam [8] abandon the negative samples and use an asymmetric architecture to avoid model collapse. SwAV [3] uses online clustering to cluster samples and forces the consistency among cluster assignments of different augmentations. After that, some studies [4, 39, 10, 19] point out that enriching the augmented samples can improve the performance of contrastive learning. In addition, the study in [31] shows that improving the quality of positive augmented samples is important for self-supervised learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.472, + 0.717 + ], + "angle": 0, + "content": "However, it is unavoidable to construct data augmentations containing different semantic concepts. Fig. 1(a) shows three augmentations of two images, in which the third augmentation is over-augmented and contains only the background. Fig. 1(b) shows category probability distributions of the corresponding images in (a), which are obtained from a supervised pre-trained ResNet50 [18]. We observed that the probability distribution of over-augmented images changes greatly compared with the first two augmentations, which indicates that the semantic information of the over-augmented images deviates from the normally-augmented images. Similar observation can be found in Fig. 1(c)(d). The original image in (d) shows a max probability for \"ambulance\", while the over-augmented image in (c) represents the different category \"telescope\". Due to such semantic inconsistency, conducting contrastive learning on these overaugmentations is harmful to representation learning. In this study, we found that semantics-consistent feature augmentation (Fig. 1(e), generated by Eq. (7)) can balance the original semantics \"ambulance\" and the over-augmented semantics \"telescope\", which can alleviate the influence of semantic inconsistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Motivated by this observation, we propose a novel semantics-consistent feature search (SCFS) method to alleviate the negative influence of semantic inconsistency in contrastive learning. SCFS utilizes the global feature of a view to adaptively search the semantics-consistent features of another view for contrast according to their similarity. It constructs informative feature augmentations and conducts contrast learning between feature augmentations and data augmentations. Thus, the pre-trained model can learn to focus on meaningful object regions to alleviate the negative influence of unmatched semantic alignment in current contrastive learning for better representation learning. In" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.242 + ], + "angle": 0, + "content": "addition, the feature search is conducted on multiple layers of the backbone network, further enhancing the semantic alignment at different scales of features. Extensive experiments conducted on different datasets and tasks demonstrate that SCFS effectively improves the performance of self-supervised learning and achieves state-of-the-art performance on different downstream tasks. For example, it achieves state-of-the-art \\(75.7\\%\\) ImageNet top-1 accuracy under the pre-training setting of 1024 batch size and 800 epochs for ResNet50." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.243, + 0.852, + 0.258 + ], + "angle": 0, + "content": "The main contributions of this study are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.268, + 0.892, + 0.342 + ], + "angle": 0, + "content": "- A novel contrastive learning method, i.e., SCFS, is proposed, and it can enhance semantic alignment in contrastive learning. To our knowledge, this is the first work that defines a feature search task in contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.354, + 0.892, + 0.398 + ], + "angle": 0, + "content": "- We expand contrastive learning from a data-to-data manner to a feature-to-data manner, which enriches the diversity of augmentations." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.41, + 0.892, + 0.438 + ], + "angle": 0, + "content": "- The proposed SCFS achieves state-of-the-art performance on different downstream tasks." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.268, + 0.892, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.453, + 0.65, + 0.469 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.895, + 0.614 + ], + "angle": 0, + "content": "Recently, some studies [33, 44, 42, 45, 46, 25, 47, 40, 23, 41] pointed out that the problem of semantic inconsistency is more serious for downstream dense prediction tasks, such as object detection and instance segmentation. Therefore, these methods utilize region-level and pixel-level features for contrast. In this study, the proposed SCFS construct feature-level augmentations using dense feature maps. Therefore, this section introduces related studies that conduct contrastive learning using region-level and pixel-level features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.895, + 0.765 + ], + "angle": 0, + "content": "Region-level contrastive learning. SCRL [33] minimizes the distance between two local features, which are cropped from two corresponding feature maps of two views. ReSim [44] aligns regional representations by sliding a fixed-sized window across the overlapping area between two views to improve the performance for localization-based tasks. SoCo [42], ORL [46], and UniVIP [25] extract object region proposals and use them to construct region-level features for contrastive learning. They achieve good performance for downstream dense prediction tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Pixel-level contrastive learning. To obtain a more fine-grained representation, several studies [47, 40, 41] design pixel-level contrastive learning task, which assumes that features extracted from the same pixel of different views should be treated as positive pairs while pixels from others must be distinguished. PixPro [47] utilizes a pixel propagation module to select similar pixel features for contrast and encourages consistency between positive pixel pairs. DenseCL [40] proposes a dense projection head to generate dense" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16100" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "feature vectors for pixel-level contrastive learning. SetSim [41] is designed to realize pixel-wise similarity learning by filtering out noisy backgrounds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.471, + 0.348 + ], + "angle": 0, + "content": "As summarized above, data augmentations bring rich information while increasing uncertainty in contrastive learning. While methods that utilize region-level features expand the granularity of feature representation by alleviating the influence of noises. Recently, to solve the problem of semantic inconsistency in video contrastive learning, VITO [30] uses an MLP to learn attention masks for each view independently and fuses the attention pooling features from multi-layers for contrastive learning. Unlike VITO [30], SCFS uses a cross-view search to calculate attention and conducts contrastive learning at multi-layers independently. In addition, our method bridges the correlation between data and feature augmentations and extends the contrastive learning task to a semantics-consistent feature search task." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.363, + 0.176, + 0.379 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.471, + 0.57 + ], + "angle": 0, + "content": "The overall architecture of SCFS is shown in Fig. 2. It consists of an encoder and a momentum encoder. The momentum encoder is an exponential-moving-average version of the encoder. SCFS consists of two contrastive learning tasks: the contrast between data augmentations \\((\\mathcal{L}_d)\\), which is designed based on the baseline DINO [4], and the contrast between data augmentations and feature augmentations \\((\\mathcal{L}_{fs})\\). In this section, we first introduce the contrast between data augmentations in Sec. 3.1. Then, the key feature search module of SCFS is introduced in detail in Sec. 3.2. Next, we will introduce the overall loss of SCFS in Sec. 3.3. Finally, the implementation details are presented in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.58, + 0.414, + 0.596 + ], + "angle": 0, + "content": "3.1. Contrast Between Data Augmentations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.47, + 0.709 + ], + "angle": 0, + "content": "Given a pair of global augmentations \\((I_1\\) and \\(I_{2}\\) ) of an input image, the feature representations of the two augmentations are used to calculate the global contrastive loss. Specifically, \\(\\pmb {f}_1 = E_\\pmb {\\theta}(\\pmb {I}_1)\\) and \\(\\pmb{f^{\\prime}}_2 = E_{\\pmb{\\theta}^{\\prime}}(\\pmb {I}_2)\\) , where \\(\\pmb{\\theta}\\) and \\(\\pmb{\\theta}^{\\prime}\\) are parameters of the encoder and the momentum encoder, respectively. \\(\\pmb {f}_1,\\pmb{f}^{\\prime}_2\\in R^K\\) \\(K\\) is the output dimension. \\(\\pmb{f}_{1}\\) is normalized with a softmax function:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.718, + 0.47, + 0.757 + ], + "angle": 0, + "content": "\\[\nP _ {1} ^ {i} = \\frac {\\exp \\left(f _ {1} ^ {i} / \\tau\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(f _ {1} ^ {k} / \\tau\\right)} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.768, + 0.47, + 0.873 + ], + "angle": 0, + "content": "where \\(\\tau > 0\\) is a temperature parameter that controls the sharpness of the output distribution. Note that \\(P_{2}^{\\prime}\\) is obtained by normalizing \\(f_{2}^{\\prime}\\) with a similar softmax function with temperature \\(\\tau^{\\prime}\\). \\(I_{1}\\) and \\(I_{2}\\) are fed to the momentum encoder and encoder symmetrically, and \\(P_{1}^{\\prime}\\) and \\(P_{2}\\) are obtained respectively. Following DINO [4], the cross-entropy loss is employed as the contrastive loss between two global views:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.885, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g} = - \\left(P _ {2} ^ {\\prime} \\log \\left(P _ {1}\\right) + P _ {1} ^ {\\prime} \\log \\left(P _ {2}\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.182 + ], + "angle": 0, + "content": "To enrich augmentations, the multi-crop strategy [3] is employed. Multiple local augmentations \\( \\mathbf{I}_l \\) is also constructed and fed to the encoder: \\( \\mathbf{f}_l = E_{\\boldsymbol{\\theta}}(\\mathbf{I}_l) \\). \\( P_l \\) is obtained by normalizing \\( \\mathbf{f}_l \\) with the softmax function with temperature \\( \\tau \\). The contrast between local views and global views can be calculated:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.19, + 0.893, + 0.232 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l} = \\sum_ {n = 1} ^ {N} - \\left(P _ {1} ^ {\\prime} \\log \\left(P _ {l} ^ {n}\\right) + P _ {2} ^ {\\prime} \\log \\left(P _ {l} ^ {n}\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.24, + 0.892, + 0.27 + ], + "angle": 0, + "content": "where \\(N\\) denotes the number of local views. Thus, the overall loss is the sum of global loss and local loss:" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.282, + 0.893, + 0.298 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = \\mathcal {L} _ {g} + \\mathcal {L} _ {l} \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.307, + 0.822, + 0.322 + ], + "angle": 0, + "content": "3.2. Semantics-Consistent Feature Search" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.33, + 0.894, + 0.435 + ], + "angle": 0, + "content": "As introduced in Sec. 1, it is unavoidable to construct augmentations that contain different semantic concepts during the augmentation procedure. It's harmful to pull these augmentations close indiscriminately in the feature space. Therefore, we propose the SCFS method to enhance the contrast between semantics-consistent regions in different augmentations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.436, + 0.894, + 0.604 + ], + "angle": 0, + "content": "The architecture of SCFS is shown in Fig. 2. By feeding the local augmentations \\(I_{l}\\) to the encoder, feature maps from different stages of the backbone ResNet50 [18] are extracted. Specifically, the output features from different stages, i.e., \\(Res2\\), \\(Res3\\) and \\(Res4\\), are utilized to conduct SCFS, ensuring that each stage of the backbone produces discriminative features: \\(\\{^{2}F_{l},^{3}F_{l},^{4}F_{l}\\} = E_{\\theta}(I_{l})\\), where \\(^i F_l\\in R^{W_l^i\\times H_l^i\\times C^i}\\), \\(W_{l}^{i},H_{l}^{i},C^{i}\\) denote the width, height and channel dimension, respectively. Next, the global average pooling operation is conducted on each \\(^i F_{l}\\) in the spatial dimensions:" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.614, + 0.893, + 0.658 + ], + "angle": 0, + "content": "\\[\n{ } ^ { i } \\boldsymbol { f } _ { l } = \\frac { 1 } { W _ { l } ^ { i } \\times H _ { l } ^ { i } } \\sum _ { x = 1 } ^ { W _ { l } ^ { i } } \\sum _ { y = 1 } ^ { H _ { l } ^ { i } } { } ^ { i } \\boldsymbol { F } _ { l } ( x , y , z ) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.751 + ], + "angle": 0, + "content": "where \\( {}^i f_l \\in R^{C^i} \\). Meanwhile, the global augmentations \\( I_g \\) (\\( g = 1,2 \\)) are fed to the momentum encoder to extract feature maps from different stages: \\( \\{^2 F_g',^3 F_g',^4 F_g'\\} = E_{\\pmb{\\theta}'}(\\pmb {I}_g),{}^i F_g' \\in R^{W_g^i\\times H_g^i\\times C^i}, W_g^i,H_g^i,C^i \\) denote width, height and channel dimension, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.751, + 0.894, + 0.858 + ], + "angle": 0, + "content": "Then, based on \\( {}^i\\pmb {f}_l \\) and \\( {}^i\\pmb{F}_{g}^{\\prime} \\), SCFS aims to adaptively search the most semantics-consistent features in \\( {}^i\\pmb{F}_{g}^{\\prime} \\) for contrast, while suppressing irrelevant features. In SCFS, each feature \\( {}^i\\pmb {f}_l \\) of the local data augmentations is treated as query, and the features \\( {}^i\\pmb{F}_{g}^{\\prime} \\) of the global augmentations are treated as keys. The similarity between \\( {}^i\\pmb {f}_l \\) and \\( {}^i\\pmb{F}_{g}^{\\prime} \\) is calculated:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.867, + 0.893, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {A} (x, y) = \\frac {{} ^ {i} \\boldsymbol {f} _ {l} \\cdot {} ^ {i} \\boldsymbol {F} _ {g} ^ {\\prime} (x , y)}{\\| ^ {i} \\boldsymbol {f} _ {l} \\| _ {2} \\| ^ {i} \\boldsymbol {F} _ {g} ^ {\\prime} (x , y) \\| _ {2}} \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16101" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.16, + 0.089, + 0.56, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.09, + 0.812, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.273, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Figure 2. Overall architecture of the proposed semantics-consistent feature search (SCFS). It consists of an encoder and a momentum encoder. There are two contrastive learning tasks: the contrast between data augmentations \\((\\mathcal{L}_d)\\) in the final feature space and the feature search task conducted on multiple layers \\((\\mathcal{L}_{fs})\\). The details of the feature search procedure is shown on the right." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.475, + 0.462 + ], + "angle": 0, + "content": "where \\( \\mathbf{A} \\in R^{W_g^i \\times H_g^i} \\) is the attention map, and \\( x = 1, \\ldots, W_g^i \\), \\( y = 1, \\ldots, H_g^i \\), \\( \\| \\cdot \\|_2 \\) is the L2 norm. The attention map \\( \\mathbf{A} \\) activates the semantics-consistent regions of the local augmentation on the global augmentation. Thus, the higher portion of local regions can be searched. To select semantic features and suppress irrelevant local features, we directly multiply the attention map \\( \\mathbf{A} \\) with \\( {}^i\\mathbf{F}_g' \\) to obtain the semantics-consistent feature augmentations:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.467, + 0.47, + 0.488 + ], + "angle": 0, + "content": "\\[\n{ } ^ { i } \\boldsymbol { F } _ { l g } ^ { \\prime } = \\boldsymbol { A } \\cdot { } ^ { i } \\boldsymbol { F } _ { g } ^ { \\prime } ( 7 )\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.47, + 0.612 + ], + "angle": 0, + "content": "This operation can be regarded as attention-weighted average pooling. Through feature search, \\(N\\) local data augmentations \\(\\pmb{I}_l\\) can search \\(N\\) corresponding semantics-consistent features \\({}^i\\pmb{F}_{lg}'\\) from a global data augmentation \\(\\pmb{I}_g\\). That is, in terms of the global data augmentation, \\(N\\) different features are constructed in the feature space through the feature search procedure. Therefore, we term \\({}^i\\pmb{F}_{lg}'\\) as feature-level augmentations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.644 + ], + "angle": 0, + "content": "Next, \\( {}^i\\mathbf{F}_l \\) and \\( {}^i\\mathbf{F}_{lg}' \\) are fed to corresponding projection heads to obtain their final representations for contrast:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.649, + 0.47, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} ^ {i} \\boldsymbol {f} _ {l} = \\mathrm {H} _ {i} \\left(^ {i} \\boldsymbol {F} _ {l}\\right) \\\\ ^ {i} \\boldsymbol {f} _ {l g} ^ {\\prime} = H _ {i} ^ {\\prime} \\left(^ {i} \\boldsymbol {F} _ {l g} ^ {\\prime}\\right) \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.795 + ], + "angle": 0, + "content": "where \\(\\mathrm{H}_i\\) and \\(\\mathrm{H}'_i\\) denote the projection heads on the \\(i\\)-th layer of the encoder and the momentum encoder, respectively. \\({}^i\\pmb {f}_l\\) and \\({}^i\\pmb {f}_{lg}'\\) are normalized with softmax function with temperature \\(\\tau\\) and \\(\\tau'\\), respectively, as the same formulation in Eq. (1). The corresponding output probability \\({}^iP_l\\) and \\({}^iP_{lg}'\\) are employed to calculate the contrast loss between local data augmentations and feature augmentations:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.801, + 0.47, + 0.844 + ], + "angle": 0, + "content": "\\[\n{ } ^ { i } \\mathcal { L } _ { f s } = \\sum _ { g = 1 } ^ { 2 } \\sum _ { n = 1 } ^ { N } - \\left( ^ { i } P _ { l g } ^ { \\prime } \\log \\left( ^ { i } P _ { l } ^ { n } \\right) \\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.85, + 0.458, + 0.865 + ], + "angle": 0, + "content": "The overall feature search loss is the sum of all layers:" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.871, + 0.47, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f s} = \\sum_ {i \\in V _ {L}} ^ {i} \\mathcal {L} _ {f s} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.34, + 0.847, + 0.355 + ], + "angle": 0, + "content": "where \\(V_{L}\\) denotes the set of layers to conduct SCFS." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.374, + 0.635, + 0.389 + ], + "angle": 0, + "content": "3.3. Overall Loss" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.4, + 0.892, + 0.431 + ], + "angle": 0, + "content": "The overall loss is the sum of the contrastive loss between data augmentations and the feature search loss:" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.45, + 0.892, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {d} + \\mathcal {L} _ {f s} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.487, + 0.895, + 0.593 + ], + "angle": 0, + "content": "Through SCFS, the contrast between feature augmentations and data augmentations is bridged. The model can adaptively search the semantics-consistent features for contrast. Therefore, it can enhance the importance of semantics-consistent regions in different augmentations, alleviating the uncertainty in contrastive learning introduced by data augmentations that contain different semantic concepts." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.612, + 0.719, + 0.629 + ], + "angle": 0, + "content": "3.4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.638, + 0.892, + 0.698 + ], + "angle": 0, + "content": "SCFS is based on DINO [4] and we follow the most hyper-parameter settings of DINO. For a fair comparison, the standard ResNet50 [18] is employed as the default backbone network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.895, + 0.822 + ], + "angle": 0, + "content": "For data augmentation, the global augmentations consist of random cropping, resizing to \\(224 \\times 224\\), random horizontal flip, gaussian blur, and color jittering. And the local augmentations consist of random cropping, resizing to \\(96 \\times 96\\), random horizontal flip, gaussian blur, and color jittering. For feature augmentations in SCFS, the Res2, Res3, and Res4 layers are used. Two global views with \\(N = 8\\) local views are the default setting of augmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.902 + ], + "angle": 0, + "content": "The projection head for the contrast between data augmentations consists of a four-layer multi-layer-perceptron (MLP) with the same architecture as DINO [4]. The projection head for feature search consists of three convolutional layers and two FC layers." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16102" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.089, + 0.462, + 0.519 + ], + "angle": 0, + "content": "
MethodBatch SizeEpochsLPk-NN
Supervised25610076.274.8
SimCLR [5]4096100069.3-
BYOL [13]4096100074.366.9
BYOL [13]409620070.6-
SwAV [3]409680075.3-
SwAV [3]25620072.7-
MoCo-v2 [16]25620067.554.3
SimSiam [8]25620070.0-
ISD [36]25620069.862.0
MSF [20]25620071.464.0
NNCLR [10]409620070.7-
Barlow Twins [50]2048100073.2-
VICReg [2]2048100073.2-
OBoW [12]25620073.8-
DCL [49]25620066.9-
CLSA [39]25620073.3-
AdCo [19]25620073.2-
DetCo [45]25620068.6-
UniVIP [25]409620073.1-
HCSC [14]25620073.3-
MoCo-v3 [9]409630072.8-
MoCo-v3 [9]4096100074.6-
DINO* [4]25620073.064.0
DINO [4]408080075.367.5
SCFS25620073.965.5
SCFS102480075.768.5
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.529, + 0.472, + 0.6 + ], + "angle": 0, + "content": "Table 1. Linear probing and \\( k \\)-NN accuracy (\\%) on ImageNet. The result with “*” is reproduced for fair comparison. LP denotes linear probing. Bold font and underline indicate the best results under the setting of 256 batch size and 200 epochs and the setting of 1024 batch size and 800 epochs, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.625, + 0.21, + 0.643 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.652, + 0.472, + 0.757 + ], + "angle": 0, + "content": "In this section, comprehensive experiments are conducted to demonstrate the effectiveness of SCFS. We evaluate the performance on different downstream tasks, including ImageNet classification, object detection, instance segmentation, and other classification task on small datasets. In addition, we conduct ablation experiments to analyze the influence of each component in SCFS." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.449, + 0.788 + ], + "angle": 0, + "content": "4.1. Comparing with SSL methods on ImageNet" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.472, + 0.903 + ], + "angle": 0, + "content": "\\(k\\)-NN and Linear Probing Accuracy on ImageNet. After pre-training on the ImageNet ILSVRC-2012 [34] training set, the pre-trained models are evaluated on the ImageNet ILSVRC-2012 validation set. For \\(k\\)-NN, it is evaluated as in study [43]. For linear probing, we train a linear classifier from scratch based on the feature extracted by a fixed backbone with 100 epochs [16]. The top-1 accuracy is adopted" + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.089, + 0.885, + 0.233 + ], + "angle": 0, + "content": "
MethodBatch Size EpochsTop-1Top-5
1%10%1%10%
Supervised [51]2569025.456.448.480.4
SimCLR [5]4096100048.365.675.587.8
BYOL [13]4096100053.268.878.489.0
SwAV [3]409680053.970.278.589.9
DINO [4]408080050.269.374.089.1
SCFS102480054.370.578.690.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.242, + 0.892, + 0.269 + ], + "angle": 0, + "content": "Table 2. Evaluation on small labeled ImageNet. Bold font indicates the best result." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.28, + 0.882, + 0.413 + ], + "angle": 0, + "content": "
MethodEpochsAPbAPb50APb75
Scratch-33.860.233.1
Supervised9053.581.358.8
SimCLR [5]100056.381.962.5
BYOL [13]30051.981.056.5
SwAV [3]40045.177.446.5
DINO [4]80055.982.162.3
SCFS80057.483.063.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.423, + 0.895, + 0.451 + ], + "angle": 0, + "content": "Table 3. Results for PASCAL VOC object detection using Faster R-CNN [32] with ResNet50-C4. Bold font indicates the best result." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.476, + 0.664, + 0.489 + ], + "angle": 0, + "content": "as the evaluation metric." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.895, + 0.672 + ], + "angle": 0, + "content": "The results are reported in Tab. 1. With the standard ResNet50 [18] architecture and pre-trained with 256 batch size for 200 epoch, the proposed SCFS achieves the best \\( k \\)-NN top-1 accuracy \\( 65.5\\% \\) and the best linear probing top-1 accuracy \\( 73.9\\% \\), outperforming its baseline DINO [4] by \\( 1.5\\% \\) and \\( 0.9\\% \\), respectively. In addition, with 1024 batch size and 800 epoch, SCFS achieves the best \\( k \\)-NN accuracy \\( (68.5\\%) \\) and linear probing accuracy \\( (75.7\\%) \\), outperforming the accuracy of DINO [4] trained with 4080 batch size for 800 epoch. This result demonstrates that SCFS can improve the representation learning performance by searching semantics-consistent features for contrast." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.896, + 0.824 + ], + "angle": 0, + "content": "Semi-Supervised Learning on ImageNet. In this part, we evaluate the performance of SCFS under the semi-supervised setting. Specifically, we use \\(1\\%\\) and \\(10\\%\\) of the labeled training data from ImageNet [34] for finetuning, which follows the semi-supervised protocol in SimCLR [5]. The same splits of \\(1\\%\\) and \\(10\\%\\) of ImageNet labeled training data in SimCLRv2 [6] are used. The results are reported in Tab. 2. After finetuning using \\(1\\%\\) and \\(10\\%\\) training data, SCFS outperforms all the compared methods. The results demonstrate that SCFS achieves the best feature representation quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.852, + 0.849 + ], + "angle": 0, + "content": "4.2. Transfer Learning on Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Object Detection and Instance Segmentation. In this part, we evaluate the representations of SCFS on dense prediction tasks, i.e., object detection and instance segmentation," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16103" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.128, + 0.089, + 0.842, + 0.364 + ], + "angle": 0, + "content": "
MethodEpochs1×schedule2×schedule
\\(AP^b\\)\\(AP^b_{50}\\)\\(AP^b_{75}\\)\\(AP^s\\)\\(AP^s_{50}\\)\\(AP^s_{75}\\)\\(AP^b\\)\\(AP^b_{50}\\)\\(AP^b_{75}\\)\\(AP^s\\)\\(AP^s_{50}\\)\\(AP^s_{75}\\)
Scratch-31.049.533.228.546.830.438.457.542.034.754.837.2
Supervised9038.959.642.735.456.538.141.361.345.037.358.340.3
MoCo [16]20038.558.942.035.155.937.740.861.644.736.958.439.7
MoCo v2 [7]20040.460.244.236.457.238.941.761.645.637.658.740.5
BYOL [13]30040.461.644.137.258.839.842.362.646.238.359.641.1
SwAV [3]400------42.362.846.338.260.041.0
\\(ReSim-FPN^T\\) [44]20039.860.243.536.057.138.641.461.945.437.559.140.3
SetSim [41]20040.260.743.936.457.739.041.662.445.937.759.440.6
DenseCL [40]20040.359.944.336.457.039.241.261.945.137.358.940.1
DSC [23]20039.458.943.235.756.138.3------
HSA [48]80040.260.943.936.557.939.142.263.046.138.159.940.9
DetCo [45]80040.161.043.936.458.038.9------
ORL* [46]80040.360.244.436.357.338.9------
DINO [4]80040.061.643.436.558.639.141.962.646.037.859.740.6
SCFS80040.561.844.036.758.839.242.163.446.138.160.241.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.372, + 0.895, + 0.403 + ], + "angle": 0, + "content": "Table 4. Object detection and instance segmentation on COCO using Mask R-CNN [17] with ResNet50-FPN. Bold font indicates the best result. * indicates pre-training on COCO [26]." + }, + { + "type": "table", + "bbox": [ + 0.189, + 0.412, + 0.785, + 0.515 + ], + "angle": 0, + "content": "
MethodCIFAR-10CIFAR-100CUB-BirdStanford-CarsAircraftOxford-Pets
Supervised97.586.481.392.186.092.1
SimCLR[5]97.785.9-91.388.189.2
BYOL[13]97.886.1-91.688.191.7
DINO[4]*97.786.681.091.187.491.5
SCFS97.886.782.791.688.591.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.524, + 0.895, + 0.553 + ], + "angle": 0, + "content": "Table 5. Transfer learning results from ImageNet with the standard ResNet50 [18]. * denotes the results are reproduced in this study. Bold font indicates the best result." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.574, + 0.473, + 0.743 + ], + "angle": 0, + "content": "on mainstream datasets PASCAL VOC [11] and MS COCO [26] datasets. On the PASCAL VOC dataset [11], the trainval07+12 set is used as the training set, and the test2007 set is used as the test set. Following [42], Faster R-CNN detector [32] with the ResNet50-C4 backbone initialized by the self-supervised pre-trained model is trained end-to-end. On the COCO dataset, the train2017 set is used for training and the val2017 set is used for evaluation. The Mask R-CNN [17] with R50-FPN is used. The \\(\\mathrm{AP^{b}}\\), \\(\\mathrm{AP_{50}^{b}}\\) and \\(\\mathrm{AP_{75}^{b}}\\) metrics are used for object detection. While the \\(\\mathrm{AP^{s}}\\), \\(\\mathrm{AP_{50}^{s}}\\) and \\(\\mathrm{AP_{75}^{s}}\\) metrics are used for instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.745, + 0.472, + 0.837 + ], + "angle": 0, + "content": "The experimental results are shown in Tab. 3 and Tab. 4. SCFS achieves best performance on the two datasets. For example, on VOC, SCFS achieves \\(57.4\\% \\mathrm{AP}^{\\mathrm{b}}\\), \\(83.0\\% \\mathrm{AP}_{50}^{\\mathrm{b}}\\) and \\(63.6\\% \\mathrm{AP}_{75}^{\\mathrm{b}}\\). The \\(\\mathrm{AP}^{\\mathrm{b}}\\) of SCFS outperforms its baseline DINO by \\(1.5\\%\\). These results show that SCFS also has good transfer ability on dense prediction tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Other Classification Tasks. In this part, we focus on the performance of self-supervised models when they are finetuned on small datasets, including CIFAR [22] and fine grained datasets [21, 37, 28, 29]. The results are shown in Tab. 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.895, + 0.622 + ], + "angle": 0, + "content": "The proposed SCFS shows the best performance on all the small datasets, which demonstrates that SCFS has good generalization ability." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.635, + 0.805, + 0.651 + ], + "angle": 0, + "content": "4.3. Pre-training on Uncurated Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The proposed SCFS can solve the problem of semantic inconsistency during pre-training, which is important when pre-training on uncurated datasets since this problem is more serious. To verify this, we pre-train SCFS and DINO on COCO [26], which is much more Uncurated than ImageNet. The same hyper-parameters used on ImageNet are applied to train the models with 512 batch size for 500 epochs. After pre-training, we fine-tune the pre-trained models on COCO for object detection and instance segmentation. The Mask R-CNN [17] with R50-FPN is used. As shown in Tab. 6, SCFS improves the performance significantly compared to its baseline DINO. In addition, compared to other dense pixel-level and region-level methods, such as DenseCL [40] and ORL [46], SCFS also achieves the best performance. This experiment verifies that SCFS can effectively solve the problem of semantic inconsistency during pre-training." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16104" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.462, + 0.267 + ], + "angle": 0, + "content": "
Methodk-NNAPbAPb50APb75APsAPs50APs75
Scratch-31.049.533.228.546.830.4
Supervised*-38.959.642.735.456.538.1
SimCLR [5]35.337.056.840.333.753.836.1
MoCov2 [7]39.438.558.142.134.855.337.3
BYOL [13]42.639.559.343.235.656.538.2
DenseCL [40]32.539.659.343.335.756.538.4
ORL [46]43.640.360.244.436.357.338.9
UniVIP [25]-40.8--36.8--
DINO [4]42.639.059.642.935.656.838.0
SCFS44.640.961.644.436.958.439.5
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.275, + 0.47, + 0.345 + ], + "angle": 0, + "content": "Table 6. Pre-training and than Fine-tuning on COCO [26] using Mask R-CNN [17] with ResNet50-FPN and \\(1 \\times\\) schedule. All models pre-trained on COCO are pre-trained with 512 batch size for 800 epochs. Bold font indicates the best result. * indicates pre-training on ImageNet [34]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.372, + 0.241, + 0.387 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.47, + 0.502 + ], + "angle": 0, + "content": "We analyze the influence of each component in SCFS. To speed up the training time, the ImageNet100 dataset, which contains 100 randomly selected categories from ImageNet [34], is adopted. All the models are pre-trained on the ImageNet100 training set with 256 batch size for 200 epoch, and tested on the validation set. The \\(k\\)-NN and linear probing top-1 accuracy are used as the evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.61 + ], + "angle": 0, + "content": "Influence of Different Contrast Modes. The contrast mode can be divided into three types: contrast between two global data augmentations used in all contrastive learning methods \\((\\mathrm{G_d}2\\mathrm{G_d})\\); contrast between local data augmentations and global data augmentations used in multi-crop strategy \\((\\mathrm{L_d}2\\mathrm{G_d})\\); and contrast between local data augmentations and local feature augmentations used in SCFS \\((\\mathrm{L_d}2\\mathrm{L_f})\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.471, + 0.748 + ], + "angle": 0, + "content": "The results are shown in Tab. 7. With multi-crop, DINO [4] (81.1%) improves accuracy by 3.0% compared to DINO without multi-crop baseline. SCFS (84.8%) further improves accuracy by 3.7% by introducing a contrast between local data augmentation and local feature augmentation. Some attention maps of SCFS and DINO are shown in Fig. 3. SCFS can more accurately focus on semantics-consistent regions between global view and local views, while DINO is easily influenced by background." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.471, + 0.809 + ], + "angle": 0, + "content": "We also add multi-layer feature contrastive learning on DINO. The result in Tab. 7 (the \"DINO w ML\" row) verifies the improvements of SCFS are not totally owed to multi-layer contrast." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In addition, we directly crop the corresponding region of local augmentation on the feature map of global augmentation for contrastive learning. As shown in Tab. 7, this variant (ROIAlign) of SCFS also outperforms the DINO baseline, which shows that the directly cropped features are also beneficial for contrastive learning. And the ROIAlign variant" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.09, + 0.89, + 0.206 + ], + "angle": 0, + "content": "
Contrast ModeGd2GdLd2GdLd2LfGd2Gfk-NNLP
DINO w/o MC✓78.183.7
DINO✓✓81.187.0
DINO w ML✓✓82.287.4
SCFS✓✓✓84.889.2
ROI Crop✓✓✓83.988.1
SCFS w/o MC✓79.786.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.215, + 0.892, + 0.244 + ], + "angle": 0, + "content": "Table 7. Influence of different contrast modes. MC, ML, and LP denote multi-crop, multi-layer, and linear probing, respectively." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.262, + 0.877, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.433, + 0.895, + 0.614 + ], + "angle": 0, + "content": "Figure 3. Attention maps of SCFS (the third row) compared with DINO [4] (the second row). In each example, (a) shows a global image, and its four local images in (b) are constructed by \\(2 \\times 2\\) jigsaw. (d) and (f) show the attention maps that highlight the semantics-consistent regions between the local images in (b) and the global image in (a). They are obtained by multiplying the globally average pooled feature maps from the encoder (Res4) of the local images in (b) with the feature map (Res4) of the global image in (a). And the encoder is the trained DINO ResNet50 model and SCFS ResNet50 model in (d) and (f), respectively. (c) and (e) show the mean attention maps of DINO and SCFS respectively, which are obtained by multiplying the mean globally average pooled feature map of the four local images in (b) with feature map of the global image in (a)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.894, + 0.702 + ], + "angle": 0, + "content": "of SCFS achieves lower accuracy than SCFS, demonstrating that the soft feature search in SCFS is better than the hard ROIAlign since ROIAlign may damage the continuous semantic context of the feature map." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.894, + 0.84 + ], + "angle": 0, + "content": "Further, we also test the performance of SCFS under the setting without multi-crop. That is, the feature search is conducted between two global data augmentations. We term this contrast mode as \\( \\mathrm{G_d}2\\mathrm{G_f} \\). As shown in the \"SCFS w/o MC\" row, SCFS also improves the performance compared to its baseline (the \"DINO w/o MC\" row), which proves that SCFS is also helpful in solving the semantic inconsistency caused by other augmentations, not only the multi-crop augmentation strategy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Influence of Multi-Layer Contrast. The influence of the feature layer that is used for feature search is analyzed. The Res2, Res3 and Res4 in the ResNet50 [18] backbone are evaluated. As shown in Tab. 8, the performance improves" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16105" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.444, + 0.161 + ], + "angle": 0, + "content": "
Res2Res3Res4k-NNLinear Probing
✓82.086.1
✓✓84.387.5
✓✓✓84.889.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.17, + 0.444, + 0.185 + ], + "angle": 0, + "content": "Table 8. Influence of different feature augmentation layer." + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.197, + 0.429, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.351, + 0.449, + 0.365 + ], + "angle": 0, + "content": "Figure 4. The \\(k\\)-NN accuracy of features from different layers." + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.378, + 0.429, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.527, + 0.422, + 0.541 + ], + "angle": 0, + "content": "Figure 5. Influence of local augmentation number." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.471, + 0.613 + ], + "angle": 0, + "content": "with the increase of feature layer numbers, which demonstrates that conducting feature search on more layers is helpful for representation learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.794 + ], + "angle": 0, + "content": "Further, we evaluate the \\(k\\)-NN accuracy using feature maps from different layers to observe the influence of feature search on the representation of middle layers. We also choose the features extracted by the Res2, Res3 and Res4 layer of ResNet50. The results are shown in Fig. 4. Compared with DINO [4], SCFS achieves better performance with features from all middle layers on ImageNet100, which verifies that enhancing the semantic consistency can improve the semantic representation of shallow layers. Compared with supervised learning, SCFS has higher performance on res2 and res3 layer, which shows that SCFS is more advantageous in the shallow layer feature representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Influence of Local Augmentation Number. In this part, we analyze the performance difference with the change of local augmentation numbers. The results are shown in Fig. 5. The performance of DINO and SCFS is steadily improved when adding more local augmentations for contrast. In addition, SCFS improves the performance under different local augmentation numbers, which demonstrates that semantics-" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.892, + 0.138 + ], + "angle": 0, + "content": "
Cropping Scale0.05-0.140.14-0.250.25-0.40.4-0.6
DINO [4]87.087.285.682.6
SCFS89.288.787.083.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.148, + 0.892, + 0.176 + ], + "angle": 0, + "content": "Table 9. Influence of random cropping scale. We report the linear probing top-1 accuracy." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.19, + 0.885, + 0.309 + ], + "angle": 0, + "content": "
MethodBackboneBatch SizeEpochsk-NNLP
DINO [4]R10125620081.086.3
SCFSR10125620085.188.3
DINO [4]ViT-S-1625620075.080.4
SCFSViT-S-1625620076.381.0
DINO [4]ViT-B-1625620076.280.7
SCFSViT-B-1625620077.282.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.317, + 0.892, + 0.345 + ], + "angle": 0, + "content": "Table 10. Experiments on other backbones. LP denotes linear probing." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.372, + 0.891, + 0.402 + ], + "angle": 0, + "content": "consistent feature search is helpful to alleviate the influence of semantics inconsistent data augmentations." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.404, + 0.893, + 0.509 + ], + "angle": 0, + "content": "Influence of Random Cropping Scale. The hyperparameter of the random cropping scale influences the degree of semantic inconsistency. In this experiment, we analyze the performance of SCFS and DINO [4] under different random cropping scales. As shown in Tab. 9, owing to the feature search, SCFS is more adaptive to different cropping scales than the baseline DINO [4]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.511, + 0.894, + 0.631 + ], + "angle": 0, + "content": "Experiments on Other Backbones. In this part, we conduct experiments on other backbones to further evaluate the effectiveness of SCFS. Apart for the default Reset50 used in other experiments, ResNet101 and Vision Transformer (ViT-S and ViT-B) are tested. The results are shown in Tab. 10. SCFS achieves significant improvement on different backbones compared to its baseline DINO, which demonstrates that SCFS is applicable to different backbones." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.648, + 0.627, + 0.664 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this study, we aim to alleviate the problem of unmatched semantic alignment in current contrastive learning by expanding the augmentations from data space to feature space. The proposed semantics-consistent feature search (SCFS) adaptively searches semantics-consistent local features between different views for contrast, while suppressing irrelevant local features during pre-training. It conducts contrast learning between feature augmentation and data augmentation. The experimental results demonstrate that SCFS can learn to focus on meaningful object regions and effectively improve the performance of self-supervised learning. The feature search procedure in SCFS is learnable parameter-free. We will utilize the self-attention mechanism in Transformer to perform the feature search procedure to further boost its performance in future work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16106" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.472, + 0.2 + ], + "angle": 0, + "content": "[2] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.203, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. NeurIPS, 33:9912-9924, 2020. 2, 3, 5, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.472, + 0.315 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 2, 3, 4, 5, 6, 7, 8, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.472, + 0.372 + ], + "angle": 0, + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. PMLR, 2020. 1, 2, 5, 6, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.472, + 0.428 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. NeurIPS, 33:22243-22255, 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.472, + 0.472 + ], + "angle": 0, + "content": "[7] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.472, + 0.514 + ], + "angle": 0, + "content": "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.472, + 0.559 + ], + "angle": 0, + "content": "[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021. 2, 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.47, + 0.616 + ], + "angle": 0, + "content": "[10] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In ICCV, pages 9588–9597, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.47, + 0.657 + ], + "angle": 0, + "content": "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 88(2):303–338, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.47, + 0.716 + ], + "angle": 0, + "content": "[12] Spyros Gidaris, Andrei Bursuc, Gilles Puy, Nikos Komodakis, Matthieu Cord, and Patrick Perez. Obow: Online bag-of-visual-words generation for self-supervised learning. In CVPR, pages 6830–6840, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[13] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 33:21271-21284, 2020. 2, 5, 6, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[14] Yuanfan Guo, Minghao Xu, Jiawen Li, Bingbing Ni, Xuanyu Zhu, Zhenbang Sun, and Yi Xu. Hcsc: Hierarchical contrastive selective coding. In CVPR, pages 9706-9715, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.146 + ], + "angle": 0, + "content": "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.188 + ], + "angle": 0, + "content": "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.894, + 0.231 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.894, + 0.287 + ], + "angle": 0, + "content": "[19] Qianjiang Hu, Xiao Wang, Wei Hu, and Guo-Jun Qi. Adco: Adversarial contrast for efficient learning of unsupervised representations from self-trained negative adversaries. In CVPR, pages 1074-1083, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.894, + 0.329 + ], + "angle": 0, + "content": "[20] Soroush Abbasi Koohpayegani, Ajinkya Tejankar, and Hamed Piri siavash. Mean shift for self-supervised learning. In ICCV, pages 10326-10335, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.894, + 0.385 + ], + "angle": 0, + "content": "[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[22] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[23] Xiaoni Li, Yu Zhou, Yifei Zhang, Aoting Zhang, Wei Wang, Ning Jiang, Haiying Wu, and Weiping Wang. Dense semantic contrast for self-supervised visual representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1368-1376, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[24] Zeming Li, Songtao Liu, and Jian Sun. Momentum^2 teacher: Momentum teacher with momentum statistics for self-supervised learning. arXiv preprint arXiv:2101.07525, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.894, + 0.607 + ], + "angle": 0, + "content": "[25] Zhaowen Li, Yousong Zhu, Fan Yang, Wei Li, Chaoyang Zhao, Yingying Chen, Zhiyang Chen, Jiahao Xie, Liwei Wu, Rui Zhao, et al. Univip: A unified framework for self-supervised visual pre-training. In CVPR, pages 14627-14636, 2022. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[27] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.706, + 0.894, + 0.759 + ], + "angle": 0, + "content": "[28] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.894, + 0.801 + ], + "angle": 0, + "content": "[29] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In \\(CVPR\\), pages 3498-3505. IEEE, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.894, + 0.857 + ], + "angle": 0, + "content": "[30] Nikhil Parthasarathy, SM Eslami, João Carreira, and Olivier J Henaff. Self-supervised video pretraining yields strong image representations. arXiv preprint arXiv:2210.06433, 2022. 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[31] Xiangyu Peng, Kai Wang, Zheng Zhu, Mang Wang, and Yang You. Crafting better contrastive views for siamese representation learning. In CVPR, pages 16031-16040, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16107" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[32] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.471, + 0.178 + ], + "angle": 0, + "content": "[33] Byungseok Roh, Wuhyun Shin, Ildoo Kim, and Sungwooong Kim. Spatially consistent representation learning. In CVPR, pages 1144–1153, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.472, + 0.245 + ], + "angle": 0, + "content": "[34] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[35] Kaiyou Song, Jin Xie, Shan Zhang, and Zimeng Luo. Multi-mode online knowledge distillation for self-supervised visual representation learning. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.292, + 0.471, + 0.346 + ], + "angle": 0, + "content": "[36] Ajinkya Tejankar, Soroush Abbasi Koohpayegani, Vipin Pillai, Paolo Favaro, and Hamed Pirsiavash. Isd: Self-supervised learning by iterative similarity distillation. In ICCV, pages 9609-9618, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.471, + 0.388 + ], + "angle": 0, + "content": "[37] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.391, + 0.471, + 0.432 + ], + "angle": 0, + "content": "[38] Haochen Wang, Kaiyou Song, Junsong Fan, Yuxi Wang, Jin Xie, and Zhaoxiang Zhang. Hard patches mining for masked image modeling. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.434, + 0.471, + 0.46 + ], + "angle": 0, + "content": "[39] Xiao Wang and Guo-Jun Qi. Contrastive learning with stronger augmentations. IEEE TPAMI, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.463, + 0.471, + 0.504 + ], + "angle": 0, + "content": "[40] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In CVPR, pages 3024-3033, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.506, + 0.471, + 0.572 + ], + "angle": 0, + "content": "[41] Zhaoqing Wang, Qiang Li, Guoxin Zhang, Pengfei Wan, Wen Zheng, Nannan Wang, Mingming Gong, and Tongliang Liu. Exploring set similarity for dense self-supervised representation learning. In CVPR, pages 16590-16599, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.576, + 0.471, + 0.616 + ], + "angle": 0, + "content": "[42] Fangyun Wei, Yue Gao, Zhirong Wu, Han Hu, and Stephen Lin. Aligning pretraining for detection via object-level contrastive learning. NeurIPS, 34, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.619, + 0.471, + 0.659 + ], + "angle": 0, + "content": "[43] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, pages 3733-3742, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.661, + 0.471, + 0.702 + ], + "angle": 0, + "content": "[44] Tete Xiao, Colorado J Reed, Xiaolong Wang, Kurt Keutzer, and Trevor Darrell. Region similarity representation learning. In ICCV, pages 10539–10548, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.704, + 0.471, + 0.759 + ], + "angle": 0, + "content": "[45] Enze Xie, Jian Ding, Wenhai Wang, Xiaohang Zhan, Hang Xu, Peize Sun, Zhenguo Li, and Ping Luo. Detco: Unsupervised contrastive learning for object detection. In ICCV, pages 8392-8401, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.761, + 0.471, + 0.814 + ], + "angle": 0, + "content": "[46] Jiahao Xie, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Unsupervised object-level representation learning from scene images. NeurIPS, 34:28864-28876, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.817, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[47] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In CVPR, pages 16684-16693, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.874, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[48] Haohang Xu, Xiaopeng Zhang, Hao Li, Lingxi Xie, Wenrui Dai, Hongkai Xiong, and Qi Tian. Seed the views: Hierarchi" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.893, + 0.12 + ], + "angle": 0, + "content": "cal semantic alignment for contrastive representation learning. IEEE TPAMI, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.893, + 0.163 + ], + "angle": 0, + "content": "[49] Chun-Hsiao Yeh, Cheng-Yao Hong, Yen-Chi Hsu, Tyng-Luh Liu, Yubei Chen, and Yann LeCun. Decoupled contrastive learning. In ECCV, pages 668-684. Springer, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.893, + 0.206 + ], + "angle": 0, + "content": "[50] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, pages 12310-12320. PMLR, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.208, + 0.893, + 0.248 + ], + "angle": 0, + "content": "[51] Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. In ICCV, pages 1476-1485, 2019. 5" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.893, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16108" + } + ] +] \ No newline at end of file diff --git a/2023/Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning/fd85c784-c77b-49cb-91b7-f3a6f94e50c3_origin.pdf b/2023/Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning/fd85c784-c77b-49cb-91b7-f3a6f94e50c3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6989212fa0857e0f8159d14677d55fcd5d959b3d --- /dev/null +++ b/2023/Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning/fd85c784-c77b-49cb-91b7-f3a6f94e50c3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2db5b22173fbaeb4a40b7286cf0ccc2b595cbc19bd2bb7e5d889597ce4e3326b +size 1335799 diff --git a/2023/Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning/full.md b/2023/Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..92f4f2d5a58b9485e6a4a28a7f6dbdab18af10e4 --- /dev/null +++ b/2023/Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning/full.md @@ -0,0 +1,318 @@ +# Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning + +Kaiyou Song Shan Zhang Zimeng Luo Tong Wang Jin Xie Megvii Technology + +{songkaiyou, zhangshan, luozimeng, wangtong, xiejin}@megvii.com + +# Abstract + +In contrastive self-supervised learning, the common way to learn discriminative representation is to pull different augmented "views" of the same image closer while pushing all other images further apart, which has been proven to be effective. However, it is unavoidable to construct undesirable views containing different semantic concepts during the augmentation procedure. It would damage the semantic consistency of representation to pull these augmentations closer in the feature space indiscriminately. In this study, we introduce feature-level augmentation and propose a novel semantics-consistent feature search (SCFS) method to mitigate this negative effect. The main idea of SCFS is to adaptively search semantics-consistent features to enhance the contrast between semantics-consistent regions in different augmentations. Thus, the trained model can learn to focus on meaningful object regions, improving the semantic representation ability. Extensive experiments conducted on different datasets and tasks demonstrate that SCFS effectively improves the performance of self-supervised learning and achieves state-of-the-art performance on different downstream tasks. + +# 1. Introduction + +Due to the tremendous potential in learning discriminative feature representation without using data annotations, self-supervised learning, including contrastive learning [16, 5] and masked image modeling (MIM) [1, 15, 38] has received much attention in the representation learning field. Contrastive learning, as a type of discriminative self-supervised learning method, is heavily studied and has shown remarkable progress in the computer vision field in recent years. It aims at pulling different augmented "views" of the same image (positive pairs) closer while pushing diverse images (negative pairs) far from each other. To this end, a contrastive + +![](images/4d603971121f051b93cdb76af88d0e95709712c0cdf30fedf59da5215805e844.jpg) +Figure 1. Semantic inconsistency of over-augmentation. (a) shows three augmentations of two images, in which the third augmentation is over-augmented and contains only background. (b) shows category probability distributions of the corresponding images in (a), which are obtained from a supervised pre-trained ResNet50 [18] model. (c)(d)(e) show three different samples of an image (the data-augmented image, the original image, and the semantics-consistent feature-augmented sample generated by Eq. (7) in this study) and their corresponding probability distributions, which point out that the over-augmented image generates different category with the original image, while the feature-augmented sample gets a balanced category probability. + +loss between the features of different views extracted from an encoder network is employed to train the encoder network end-to-end. According to whether the negative pairs are used, current contrastive learning can be generally divided into two categories. + +The first category [16, 5] utilizes both positive pairs and negative pairs for contrast. MoCo [16, 7] uses a momentum update mechanism to maintain a memory bank of negative examples. SimCLR [5, 6] directly trains a single encoder network with a large batch size to ensure sufficient positive and negative samples for learning. Based on MoCo and Sim + +CLR, some methods [24, 20, 36, 10, 49, 39, 19, 9, 48, 14] are proposed to improve the performance. For example, MSF [20], ISD [36] and NNCLR [10] aim to search semantics-consistent samples for contrast, solving the false negative problem. While some studies, such as Momentum2Teacher [24] and DCL [49], aim to solve the limitation that large batch size is necessary for satisfactory performance. + +The second category of contrastive learning methods [13, 8, 3, 50, 2, 12, 4, 31, 35] only constructs positive pairs for contrast. Based on MoCo [16] and SimCLR [5], respectively, BYOL [13] and SimSiam [8] abandon the negative samples and use an asymmetric architecture to avoid model collapse. SwAV [3] uses online clustering to cluster samples and forces the consistency among cluster assignments of different augmentations. After that, some studies [4, 39, 10, 19] point out that enriching the augmented samples can improve the performance of contrastive learning. In addition, the study in [31] shows that improving the quality of positive augmented samples is important for self-supervised learning. + +However, it is unavoidable to construct data augmentations containing different semantic concepts. Fig. 1(a) shows three augmentations of two images, in which the third augmentation is over-augmented and contains only the background. Fig. 1(b) shows category probability distributions of the corresponding images in (a), which are obtained from a supervised pre-trained ResNet50 [18]. We observed that the probability distribution of over-augmented images changes greatly compared with the first two augmentations, which indicates that the semantic information of the over-augmented images deviates from the normally-augmented images. Similar observation can be found in Fig. 1(c)(d). The original image in (d) shows a max probability for "ambulance", while the over-augmented image in (c) represents the different category "telescope". Due to such semantic inconsistency, conducting contrastive learning on these overaugmentations is harmful to representation learning. In this study, we found that semantics-consistent feature augmentation (Fig. 1(e), generated by Eq. (7)) can balance the original semantics "ambulance" and the over-augmented semantics "telescope", which can alleviate the influence of semantic inconsistency. + +Motivated by this observation, we propose a novel semantics-consistent feature search (SCFS) method to alleviate the negative influence of semantic inconsistency in contrastive learning. SCFS utilizes the global feature of a view to adaptively search the semantics-consistent features of another view for contrast according to their similarity. It constructs informative feature augmentations and conducts contrast learning between feature augmentations and data augmentations. Thus, the pre-trained model can learn to focus on meaningful object regions to alleviate the negative influence of unmatched semantic alignment in current contrastive learning for better representation learning. In + +addition, the feature search is conducted on multiple layers of the backbone network, further enhancing the semantic alignment at different scales of features. Extensive experiments conducted on different datasets and tasks demonstrate that SCFS effectively improves the performance of self-supervised learning and achieves state-of-the-art performance on different downstream tasks. For example, it achieves state-of-the-art $75.7\%$ ImageNet top-1 accuracy under the pre-training setting of 1024 batch size and 800 epochs for ResNet50. + +The main contributions of this study are threefold: + +- A novel contrastive learning method, i.e., SCFS, is proposed, and it can enhance semantic alignment in contrastive learning. To our knowledge, this is the first work that defines a feature search task in contrastive learning. +- We expand contrastive learning from a data-to-data manner to a feature-to-data manner, which enriches the diversity of augmentations. +- The proposed SCFS achieves state-of-the-art performance on different downstream tasks. + +# 2. Related Works + +Recently, some studies [33, 44, 42, 45, 46, 25, 47, 40, 23, 41] pointed out that the problem of semantic inconsistency is more serious for downstream dense prediction tasks, such as object detection and instance segmentation. Therefore, these methods utilize region-level and pixel-level features for contrast. In this study, the proposed SCFS construct feature-level augmentations using dense feature maps. Therefore, this section introduces related studies that conduct contrastive learning using region-level and pixel-level features. + +Region-level contrastive learning. SCRL [33] minimizes the distance between two local features, which are cropped from two corresponding feature maps of two views. ReSim [44] aligns regional representations by sliding a fixed-sized window across the overlapping area between two views to improve the performance for localization-based tasks. SoCo [42], ORL [46], and UniVIP [25] extract object region proposals and use them to construct region-level features for contrastive learning. They achieve good performance for downstream dense prediction tasks. + +Pixel-level contrastive learning. To obtain a more fine-grained representation, several studies [47, 40, 41] design pixel-level contrastive learning task, which assumes that features extracted from the same pixel of different views should be treated as positive pairs while pixels from others must be distinguished. PixPro [47] utilizes a pixel propagation module to select similar pixel features for contrast and encourages consistency between positive pixel pairs. DenseCL [40] proposes a dense projection head to generate dense + +feature vectors for pixel-level contrastive learning. SetSim [41] is designed to realize pixel-wise similarity learning by filtering out noisy backgrounds. + +As summarized above, data augmentations bring rich information while increasing uncertainty in contrastive learning. While methods that utilize region-level features expand the granularity of feature representation by alleviating the influence of noises. Recently, to solve the problem of semantic inconsistency in video contrastive learning, VITO [30] uses an MLP to learn attention masks for each view independently and fuses the attention pooling features from multi-layers for contrastive learning. Unlike VITO [30], SCFS uses a cross-view search to calculate attention and conducts contrastive learning at multi-layers independently. In addition, our method bridges the correlation between data and feature augmentations and extends the contrastive learning task to a semantics-consistent feature search task. + +# 3. Methods + +The overall architecture of SCFS is shown in Fig. 2. It consists of an encoder and a momentum encoder. The momentum encoder is an exponential-moving-average version of the encoder. SCFS consists of two contrastive learning tasks: the contrast between data augmentations $(\mathcal{L}_d)$ , which is designed based on the baseline DINO [4], and the contrast between data augmentations and feature augmentations $(\mathcal{L}_{fs})$ . In this section, we first introduce the contrast between data augmentations in Sec. 3.1. Then, the key feature search module of SCFS is introduced in detail in Sec. 3.2. Next, we will introduce the overall loss of SCFS in Sec. 3.3. Finally, the implementation details are presented in Sec. 3.4. + +# 3.1. Contrast Between Data Augmentations + +Given a pair of global augmentations $(I_1$ and $I_{2}$ ) of an input image, the feature representations of the two augmentations are used to calculate the global contrastive loss. Specifically, $\pmb {f}_1 = E_\pmb {\theta}(\pmb {I}_1)$ and $\pmb{f^{\prime}}_2 = E_{\pmb{\theta}^{\prime}}(\pmb {I}_2)$ , where $\pmb{\theta}$ and $\pmb{\theta}^{\prime}$ are parameters of the encoder and the momentum encoder, respectively. $\pmb {f}_1,\pmb{f}^{\prime}_2\in R^K$ $K$ is the output dimension. $\pmb{f}_{1}$ is normalized with a softmax function: + +$$ +P _ {1} ^ {i} = \frac {\exp \left(f _ {1} ^ {i} / \tau\right)}{\sum_ {k = 1} ^ {K} \exp \left(f _ {1} ^ {k} / \tau\right)} \tag {1} +$$ + +where $\tau > 0$ is a temperature parameter that controls the sharpness of the output distribution. Note that $P_{2}^{\prime}$ is obtained by normalizing $f_{2}^{\prime}$ with a similar softmax function with temperature $\tau^{\prime}$ . $I_{1}$ and $I_{2}$ are fed to the momentum encoder and encoder symmetrically, and $P_{1}^{\prime}$ and $P_{2}$ are obtained respectively. Following DINO [4], the cross-entropy loss is employed as the contrastive loss between two global views: + +$$ +\mathcal {L} _ {g} = - \left(P _ {2} ^ {\prime} \log \left(P _ {1}\right) + P _ {1} ^ {\prime} \log \left(P _ {2}\right)\right) \tag {2} +$$ + +To enrich augmentations, the multi-crop strategy [3] is employed. Multiple local augmentations $\mathbf{I}_l$ is also constructed and fed to the encoder: $\mathbf{f}_l = E_{\boldsymbol{\theta}}(\mathbf{I}_l)$ . $P_l$ is obtained by normalizing $\mathbf{f}_l$ with the softmax function with temperature $\tau$ . The contrast between local views and global views can be calculated: + +$$ +\mathcal {L} _ {l} = \sum_ {n = 1} ^ {N} - \left(P _ {1} ^ {\prime} \log \left(P _ {l} ^ {n}\right) + P _ {2} ^ {\prime} \log \left(P _ {l} ^ {n}\right)\right) \tag {3} +$$ + +where $N$ denotes the number of local views. Thus, the overall loss is the sum of global loss and local loss: + +$$ +\mathcal {L} _ {d} = \mathcal {L} _ {g} + \mathcal {L} _ {l} \tag {4} +$$ + +# 3.2. Semantics-Consistent Feature Search + +As introduced in Sec. 1, it is unavoidable to construct augmentations that contain different semantic concepts during the augmentation procedure. It's harmful to pull these augmentations close indiscriminately in the feature space. Therefore, we propose the SCFS method to enhance the contrast between semantics-consistent regions in different augmentations. + +The architecture of SCFS is shown in Fig. 2. By feeding the local augmentations $I_{l}$ to the encoder, feature maps from different stages of the backbone ResNet50 [18] are extracted. Specifically, the output features from different stages, i.e., $Res2$ , $Res3$ and $Res4$ , are utilized to conduct SCFS, ensuring that each stage of the backbone produces discriminative features: $\{^{2}F_{l},^{3}F_{l},^{4}F_{l}\} = E_{\theta}(I_{l})$ , where $^i F_l\in R^{W_l^i\times H_l^i\times C^i}$ , $W_{l}^{i},H_{l}^{i},C^{i}$ denote the width, height and channel dimension, respectively. Next, the global average pooling operation is conducted on each $^i F_{l}$ in the spatial dimensions: + +$$ +{ } ^ { i } \boldsymbol { f } _ { l } = \frac { 1 } { W _ { l } ^ { i } \times H _ { l } ^ { i } } \sum _ { x = 1 } ^ { W _ { l } ^ { i } } \sum _ { y = 1 } ^ { H _ { l } ^ { i } } { } ^ { i } \boldsymbol { F } _ { l } ( x , y , z ) \tag {5} +$$ + +where ${}^i f_l \in R^{C^i}$ . Meanwhile, the global augmentations $I_g$ ( $g = 1,2$ ) are fed to the momentum encoder to extract feature maps from different stages: $\{^2 F_g',^3 F_g',^4 F_g'\} = E_{\pmb{\theta}'}(\pmb {I}_g),{}^i F_g' \in R^{W_g^i\times H_g^i\times C^i}, W_g^i,H_g^i,C^i$ denote width, height and channel dimension, respectively. + +Then, based on ${}^i\pmb {f}_l$ and ${}^i\pmb{F}_{g}^{\prime}$ , SCFS aims to adaptively search the most semantics-consistent features in ${}^i\pmb{F}_{g}^{\prime}$ for contrast, while suppressing irrelevant features. In SCFS, each feature ${}^i\pmb {f}_l$ of the local data augmentations is treated as query, and the features ${}^i\pmb{F}_{g}^{\prime}$ of the global augmentations are treated as keys. The similarity between ${}^i\pmb {f}_l$ and ${}^i\pmb{F}_{g}^{\prime}$ is calculated: + +$$ +\boldsymbol {A} (x, y) = \frac {{} ^ {i} \boldsymbol {f} _ {l} \cdot {} ^ {i} \boldsymbol {F} _ {g} ^ {\prime} (x , y)}{\| ^ {i} \boldsymbol {f} _ {l} \| _ {2} \| ^ {i} \boldsymbol {F} _ {g} ^ {\prime} (x , y) \| _ {2}} \tag {6} +$$ + +![](images/2f47ebc92a19132eefcdbaca58e53be431c38d99e09fc3c383ded5eae72b6d7d.jpg) +Figure 2. Overall architecture of the proposed semantics-consistent feature search (SCFS). It consists of an encoder and a momentum encoder. There are two contrastive learning tasks: the contrast between data augmentations $(\mathcal{L}_d)$ in the final feature space and the feature search task conducted on multiple layers $(\mathcal{L}_{fs})$ . The details of the feature search procedure is shown on the right. + +![](images/e11192e29bee3c44aa0413a27f29c9325f5cabf6877fd1fcae5ca6defae4acc1.jpg) + +where $\mathbf{A} \in R^{W_g^i \times H_g^i}$ is the attention map, and $x = 1, \ldots, W_g^i$ , $y = 1, \ldots, H_g^i$ , $\| \cdot \|_2$ is the L2 norm. The attention map $\mathbf{A}$ activates the semantics-consistent regions of the local augmentation on the global augmentation. Thus, the higher portion of local regions can be searched. To select semantic features and suppress irrelevant local features, we directly multiply the attention map $\mathbf{A}$ with ${}^i\mathbf{F}_g'$ to obtain the semantics-consistent feature augmentations: + +$$ +{ } ^ { i } \boldsymbol { F } _ { l g } ^ { \prime } = \boldsymbol { A } \cdot { } ^ { i } \boldsymbol { F } _ { g } ^ { \prime } ( 7 ) +$$ + +This operation can be regarded as attention-weighted average pooling. Through feature search, $N$ local data augmentations $\pmb{I}_l$ can search $N$ corresponding semantics-consistent features ${}^i\pmb{F}_{lg}'$ from a global data augmentation $\pmb{I}_g$ . That is, in terms of the global data augmentation, $N$ different features are constructed in the feature space through the feature search procedure. Therefore, we term ${}^i\pmb{F}_{lg}'$ as feature-level augmentations. + +Next, ${}^i\mathbf{F}_l$ and ${}^i\mathbf{F}_{lg}'$ are fed to corresponding projection heads to obtain their final representations for contrast: + +$$ +\left\{ \begin{array}{l} ^ {i} \boldsymbol {f} _ {l} = \mathrm {H} _ {i} \left(^ {i} \boldsymbol {F} _ {l}\right) \\ ^ {i} \boldsymbol {f} _ {l g} ^ {\prime} = H _ {i} ^ {\prime} \left(^ {i} \boldsymbol {F} _ {l g} ^ {\prime}\right) \end{array} \right. \tag {8} +$$ + +where $\mathrm{H}_i$ and $\mathrm{H}'_i$ denote the projection heads on the $i$ -th layer of the encoder and the momentum encoder, respectively. ${}^i\pmb {f}_l$ and ${}^i\pmb {f}_{lg}'$ are normalized with softmax function with temperature $\tau$ and $\tau'$ , respectively, as the same formulation in Eq. (1). The corresponding output probability ${}^iP_l$ and ${}^iP_{lg}'$ are employed to calculate the contrast loss between local data augmentations and feature augmentations: + +$$ +{ } ^ { i } \mathcal { L } _ { f s } = \sum _ { g = 1 } ^ { 2 } \sum _ { n = 1 } ^ { N } - \left( ^ { i } P _ { l g } ^ { \prime } \log \left( ^ { i } P _ { l } ^ { n } \right) \right) \tag {9} +$$ + +The overall feature search loss is the sum of all layers: + +$$ +\mathcal {L} _ {f s} = \sum_ {i \in V _ {L}} ^ {i} \mathcal {L} _ {f s} \tag {10} +$$ + +where $V_{L}$ denotes the set of layers to conduct SCFS. + +# 3.3. Overall Loss + +The overall loss is the sum of the contrastive loss between data augmentations and the feature search loss: + +$$ +\mathcal {L} = \mathcal {L} _ {d} + \mathcal {L} _ {f s} \tag {11} +$$ + +Through SCFS, the contrast between feature augmentations and data augmentations is bridged. The model can adaptively search the semantics-consistent features for contrast. Therefore, it can enhance the importance of semantics-consistent regions in different augmentations, alleviating the uncertainty in contrastive learning introduced by data augmentations that contain different semantic concepts. + +# 3.4. Implementation Details + +SCFS is based on DINO [4] and we follow the most hyper-parameter settings of DINO. For a fair comparison, the standard ResNet50 [18] is employed as the default backbone network. + +For data augmentation, the global augmentations consist of random cropping, resizing to $224 \times 224$ , random horizontal flip, gaussian blur, and color jittering. And the local augmentations consist of random cropping, resizing to $96 \times 96$ , random horizontal flip, gaussian blur, and color jittering. For feature augmentations in SCFS, the Res2, Res3, and Res4 layers are used. Two global views with $N = 8$ local views are the default setting of augmentation. + +The projection head for the contrast between data augmentations consists of a four-layer multi-layer-perceptron (MLP) with the same architecture as DINO [4]. The projection head for feature search consists of three convolutional layers and two FC layers. + +
MethodBatch SizeEpochsLPk-NN
Supervised25610076.274.8
SimCLR [5]4096100069.3-
BYOL [13]4096100074.366.9
BYOL [13]409620070.6-
SwAV [3]409680075.3-
SwAV [3]25620072.7-
MoCo-v2 [16]25620067.554.3
SimSiam [8]25620070.0-
ISD [36]25620069.862.0
MSF [20]25620071.464.0
NNCLR [10]409620070.7-
Barlow Twins [50]2048100073.2-
VICReg [2]2048100073.2-
OBoW [12]25620073.8-
DCL [49]25620066.9-
CLSA [39]25620073.3-
AdCo [19]25620073.2-
DetCo [45]25620068.6-
UniVIP [25]409620073.1-
HCSC [14]25620073.3-
MoCo-v3 [9]409630072.8-
MoCo-v3 [9]4096100074.6-
DINO* [4]25620073.064.0
DINO [4]408080075.367.5
SCFS25620073.965.5
SCFS102480075.768.5
+ +Table 1. Linear probing and $k$ -NN accuracy (\%) on ImageNet. The result with “*” is reproduced for fair comparison. LP denotes linear probing. Bold font and underline indicate the best results under the setting of 256 batch size and 200 epochs and the setting of 1024 batch size and 800 epochs, respectively. + +# 4. Experiments + +In this section, comprehensive experiments are conducted to demonstrate the effectiveness of SCFS. We evaluate the performance on different downstream tasks, including ImageNet classification, object detection, instance segmentation, and other classification task on small datasets. In addition, we conduct ablation experiments to analyze the influence of each component in SCFS. + +# 4.1. Comparing with SSL methods on ImageNet + +$k$ -NN and Linear Probing Accuracy on ImageNet. After pre-training on the ImageNet ILSVRC-2012 [34] training set, the pre-trained models are evaluated on the ImageNet ILSVRC-2012 validation set. For $k$ -NN, it is evaluated as in study [43]. For linear probing, we train a linear classifier from scratch based on the feature extracted by a fixed backbone with 100 epochs [16]. The top-1 accuracy is adopted + +
MethodBatch Size EpochsTop-1Top-5
1%10%1%10%
Supervised [51]2569025.456.448.480.4
SimCLR [5]4096100048.365.675.587.8
BYOL [13]4096100053.268.878.489.0
SwAV [3]409680053.970.278.589.9
DINO [4]408080050.269.374.089.1
SCFS102480054.370.578.690.2
+ +Table 2. Evaluation on small labeled ImageNet. Bold font indicates the best result. + +
MethodEpochsAPbAPb50APb75
Scratch-33.860.233.1
Supervised9053.581.358.8
SimCLR [5]100056.381.962.5
BYOL [13]30051.981.056.5
SwAV [3]40045.177.446.5
DINO [4]80055.982.162.3
SCFS80057.483.063.6
+ +Table 3. Results for PASCAL VOC object detection using Faster R-CNN [32] with ResNet50-C4. Bold font indicates the best result. + +as the evaluation metric. + +The results are reported in Tab. 1. With the standard ResNet50 [18] architecture and pre-trained with 256 batch size for 200 epoch, the proposed SCFS achieves the best $k$ -NN top-1 accuracy $65.5\%$ and the best linear probing top-1 accuracy $73.9\%$ , outperforming its baseline DINO [4] by $1.5\%$ and $0.9\%$ , respectively. In addition, with 1024 batch size and 800 epoch, SCFS achieves the best $k$ -NN accuracy $(68.5\%)$ and linear probing accuracy $(75.7\%)$ , outperforming the accuracy of DINO [4] trained with 4080 batch size for 800 epoch. This result demonstrates that SCFS can improve the representation learning performance by searching semantics-consistent features for contrast. + +Semi-Supervised Learning on ImageNet. In this part, we evaluate the performance of SCFS under the semi-supervised setting. Specifically, we use $1\%$ and $10\%$ of the labeled training data from ImageNet [34] for finetuning, which follows the semi-supervised protocol in SimCLR [5]. The same splits of $1\%$ and $10\%$ of ImageNet labeled training data in SimCLRv2 [6] are used. The results are reported in Tab. 2. After finetuning using $1\%$ and $10\%$ training data, SCFS outperforms all the compared methods. The results demonstrate that SCFS achieves the best feature representation quality. + +# 4.2. Transfer Learning on Downstream Tasks + +Object Detection and Instance Segmentation. In this part, we evaluate the representations of SCFS on dense prediction tasks, i.e., object detection and instance segmentation, + +
MethodEpochs1×schedule2×schedule
\(AP^b\)\(AP^b_{50}\)\(AP^b_{75}\)\(AP^s\)\(AP^s_{50}\)\(AP^s_{75}\)\(AP^b\)\(AP^b_{50}\)\(AP^b_{75}\)\(AP^s\)\(AP^s_{50}\)\(AP^s_{75}\)
Scratch-31.049.533.228.546.830.438.457.542.034.754.837.2
Supervised9038.959.642.735.456.538.141.361.345.037.358.340.3
MoCo [16]20038.558.942.035.155.937.740.861.644.736.958.439.7
MoCo v2 [7]20040.460.244.236.457.238.941.761.645.637.658.740.5
BYOL [13]30040.461.644.137.258.839.842.362.646.238.359.641.1
SwAV [3]400------42.362.846.338.260.041.0
\(ReSim-FPN^T\) [44]20039.860.243.536.057.138.641.461.945.437.559.140.3
SetSim [41]20040.260.743.936.457.739.041.662.445.937.759.440.6
DenseCL [40]20040.359.944.336.457.039.241.261.945.137.358.940.1
DSC [23]20039.458.943.235.756.138.3------
HSA [48]80040.260.943.936.557.939.142.263.046.138.159.940.9
DetCo [45]80040.161.043.936.458.038.9------
ORL* [46]80040.360.244.436.357.338.9------
DINO [4]80040.061.643.436.558.639.141.962.646.037.859.740.6
SCFS80040.561.844.036.758.839.242.163.446.138.160.241.0
+ +Table 4. Object detection and instance segmentation on COCO using Mask R-CNN [17] with ResNet50-FPN. Bold font indicates the best result. * indicates pre-training on COCO [26]. + +
MethodCIFAR-10CIFAR-100CUB-BirdStanford-CarsAircraftOxford-Pets
Supervised97.586.481.392.186.092.1
SimCLR[5]97.785.9-91.388.189.2
BYOL[13]97.886.1-91.688.191.7
DINO[4]*97.786.681.091.187.491.5
SCFS97.886.782.791.688.591.9
+ +Table 5. Transfer learning results from ImageNet with the standard ResNet50 [18]. * denotes the results are reproduced in this study. Bold font indicates the best result. + +on mainstream datasets PASCAL VOC [11] and MS COCO [26] datasets. On the PASCAL VOC dataset [11], the trainval07+12 set is used as the training set, and the test2007 set is used as the test set. Following [42], Faster R-CNN detector [32] with the ResNet50-C4 backbone initialized by the self-supervised pre-trained model is trained end-to-end. On the COCO dataset, the train2017 set is used for training and the val2017 set is used for evaluation. The Mask R-CNN [17] with R50-FPN is used. The $\mathrm{AP^{b}}$ , $\mathrm{AP_{50}^{b}}$ and $\mathrm{AP_{75}^{b}}$ metrics are used for object detection. While the $\mathrm{AP^{s}}$ , $\mathrm{AP_{50}^{s}}$ and $\mathrm{AP_{75}^{s}}$ metrics are used for instance segmentation. + +The experimental results are shown in Tab. 3 and Tab. 4. SCFS achieves best performance on the two datasets. For example, on VOC, SCFS achieves $57.4\% \mathrm{AP}^{\mathrm{b}}$ , $83.0\% \mathrm{AP}_{50}^{\mathrm{b}}$ and $63.6\% \mathrm{AP}_{75}^{\mathrm{b}}$ . The $\mathrm{AP}^{\mathrm{b}}$ of SCFS outperforms its baseline DINO by $1.5\%$ . These results show that SCFS also has good transfer ability on dense prediction tasks. + +Other Classification Tasks. In this part, we focus on the performance of self-supervised models when they are finetuned on small datasets, including CIFAR [22] and fine grained datasets [21, 37, 28, 29]. The results are shown in Tab. 5. + +The proposed SCFS shows the best performance on all the small datasets, which demonstrates that SCFS has good generalization ability. + +# 4.3. Pre-training on Uncurated Dataset + +The proposed SCFS can solve the problem of semantic inconsistency during pre-training, which is important when pre-training on uncurated datasets since this problem is more serious. To verify this, we pre-train SCFS and DINO on COCO [26], which is much more Uncurated than ImageNet. The same hyper-parameters used on ImageNet are applied to train the models with 512 batch size for 500 epochs. After pre-training, we fine-tune the pre-trained models on COCO for object detection and instance segmentation. The Mask R-CNN [17] with R50-FPN is used. As shown in Tab. 6, SCFS improves the performance significantly compared to its baseline DINO. In addition, compared to other dense pixel-level and region-level methods, such as DenseCL [40] and ORL [46], SCFS also achieves the best performance. This experiment verifies that SCFS can effectively solve the problem of semantic inconsistency during pre-training. + +
Methodk-NNAPbAPb50APb75APsAPs50APs75
Scratch-31.049.533.228.546.830.4
Supervised*-38.959.642.735.456.538.1
SimCLR [5]35.337.056.840.333.753.836.1
MoCov2 [7]39.438.558.142.134.855.337.3
BYOL [13]42.639.559.343.235.656.538.2
DenseCL [40]32.539.659.343.335.756.538.4
ORL [46]43.640.360.244.436.357.338.9
UniVIP [25]-40.8--36.8--
DINO [4]42.639.059.642.935.656.838.0
SCFS44.640.961.644.436.958.439.5
+ +Table 6. Pre-training and than Fine-tuning on COCO [26] using Mask R-CNN [17] with ResNet50-FPN and $1 \times$ schedule. All models pre-trained on COCO are pre-trained with 512 batch size for 800 epochs. Bold font indicates the best result. * indicates pre-training on ImageNet [34] + +# 4.4. Ablation Studies + +We analyze the influence of each component in SCFS. To speed up the training time, the ImageNet100 dataset, which contains 100 randomly selected categories from ImageNet [34], is adopted. All the models are pre-trained on the ImageNet100 training set with 256 batch size for 200 epoch, and tested on the validation set. The $k$ -NN and linear probing top-1 accuracy are used as the evaluation metrics. + +Influence of Different Contrast Modes. The contrast mode can be divided into three types: contrast between two global data augmentations used in all contrastive learning methods $(\mathrm{G_d}2\mathrm{G_d})$ ; contrast between local data augmentations and global data augmentations used in multi-crop strategy $(\mathrm{L_d}2\mathrm{G_d})$ ; and contrast between local data augmentations and local feature augmentations used in SCFS $(\mathrm{L_d}2\mathrm{L_f})$ . + +The results are shown in Tab. 7. With multi-crop, DINO [4] (81.1%) improves accuracy by 3.0% compared to DINO without multi-crop baseline. SCFS (84.8%) further improves accuracy by 3.7% by introducing a contrast between local data augmentation and local feature augmentation. Some attention maps of SCFS and DINO are shown in Fig. 3. SCFS can more accurately focus on semantics-consistent regions between global view and local views, while DINO is easily influenced by background. + +We also add multi-layer feature contrastive learning on DINO. The result in Tab. 7 (the "DINO w ML" row) verifies the improvements of SCFS are not totally owed to multi-layer contrast. + +In addition, we directly crop the corresponding region of local augmentation on the feature map of global augmentation for contrastive learning. As shown in Tab. 7, this variant (ROIAlign) of SCFS also outperforms the DINO baseline, which shows that the directly cropped features are also beneficial for contrastive learning. And the ROIAlign variant + +
Contrast ModeGd2GdLd2GdLd2LfGd2Gfk-NNLP
DINO w/o MC✓78.183.7
DINO✓✓81.187.0
DINO w ML✓✓82.287.4
SCFS✓✓✓84.889.2
ROI Crop✓✓✓83.988.1
SCFS w/o MC✓79.786.3
+ +Table 7. Influence of different contrast modes. MC, ML, and LP denote multi-crop, multi-layer, and linear probing, respectively. + +![](images/7f8aa03f2395644f4897a8537d84e58c10244e784f755028b6dc16c7b3747956.jpg) +Figure 3. Attention maps of SCFS (the third row) compared with DINO [4] (the second row). In each example, (a) shows a global image, and its four local images in (b) are constructed by $2 \times 2$ jigsaw. (d) and (f) show the attention maps that highlight the semantics-consistent regions between the local images in (b) and the global image in (a). They are obtained by multiplying the globally average pooled feature maps from the encoder (Res4) of the local images in (b) with the feature map (Res4) of the global image in (a). And the encoder is the trained DINO ResNet50 model and SCFS ResNet50 model in (d) and (f), respectively. (c) and (e) show the mean attention maps of DINO and SCFS respectively, which are obtained by multiplying the mean globally average pooled feature map of the four local images in (b) with feature map of the global image in (a). + +of SCFS achieves lower accuracy than SCFS, demonstrating that the soft feature search in SCFS is better than the hard ROIAlign since ROIAlign may damage the continuous semantic context of the feature map. + +Further, we also test the performance of SCFS under the setting without multi-crop. That is, the feature search is conducted between two global data augmentations. We term this contrast mode as $\mathrm{G_d}2\mathrm{G_f}$ . As shown in the "SCFS w/o MC" row, SCFS also improves the performance compared to its baseline (the "DINO w/o MC" row), which proves that SCFS is also helpful in solving the semantic inconsistency caused by other augmentations, not only the multi-crop augmentation strategy. + +Influence of Multi-Layer Contrast. The influence of the feature layer that is used for feature search is analyzed. The Res2, Res3 and Res4 in the ResNet50 [18] backbone are evaluated. As shown in Tab. 8, the performance improves + +
Res2Res3Res4k-NNLinear Probing
✓82.086.1
✓✓84.387.5
✓✓✓84.889.2
+ +![](images/32cc7bcf55c389eba83a68856410513d9704b827fa327950f9ffe264b5128b8b.jpg) +Figure 4. The $k$ -NN accuracy of features from different layers. + +![](images/b5fcca0d8ad18777935c3d077f1178bacd21da419d0ae5694771f3ec4b1e8ecc.jpg) +Figure 5. Influence of local augmentation number. + +with the increase of feature layer numbers, which demonstrates that conducting feature search on more layers is helpful for representation learning. + +Further, we evaluate the $k$ -NN accuracy using feature maps from different layers to observe the influence of feature search on the representation of middle layers. We also choose the features extracted by the Res2, Res3 and Res4 layer of ResNet50. The results are shown in Fig. 4. Compared with DINO [4], SCFS achieves better performance with features from all middle layers on ImageNet100, which verifies that enhancing the semantic consistency can improve the semantic representation of shallow layers. Compared with supervised learning, SCFS has higher performance on res2 and res3 layer, which shows that SCFS is more advantageous in the shallow layer feature representation. + +Influence of Local Augmentation Number. In this part, we analyze the performance difference with the change of local augmentation numbers. The results are shown in Fig. 5. The performance of DINO and SCFS is steadily improved when adding more local augmentations for contrast. In addition, SCFS improves the performance under different local augmentation numbers, which demonstrates that semantics- + +Table 8. Influence of different feature augmentation layer. + +
Cropping Scale0.05-0.140.14-0.250.25-0.40.4-0.6
DINO [4]87.087.285.682.6
SCFS89.288.787.083.1
+ +Table 9. Influence of random cropping scale. We report the linear probing top-1 accuracy. + +
MethodBackboneBatch SizeEpochsk-NNLP
DINO [4]R10125620081.086.3
SCFSR10125620085.188.3
DINO [4]ViT-S-1625620075.080.4
SCFSViT-S-1625620076.381.0
DINO [4]ViT-B-1625620076.280.7
SCFSViT-B-1625620077.282.3
+ +Table 10. Experiments on other backbones. LP denotes linear probing. + +consistent feature search is helpful to alleviate the influence of semantics inconsistent data augmentations. + +Influence of Random Cropping Scale. The hyperparameter of the random cropping scale influences the degree of semantic inconsistency. In this experiment, we analyze the performance of SCFS and DINO [4] under different random cropping scales. As shown in Tab. 9, owing to the feature search, SCFS is more adaptive to different cropping scales than the baseline DINO [4]. + +Experiments on Other Backbones. In this part, we conduct experiments on other backbones to further evaluate the effectiveness of SCFS. Apart for the default Reset50 used in other experiments, ResNet101 and Vision Transformer (ViT-S and ViT-B) are tested. The results are shown in Tab. 10. SCFS achieves significant improvement on different backbones compared to its baseline DINO, which demonstrates that SCFS is applicable to different backbones. + +# 5. Conclusions + +In this study, we aim to alleviate the problem of unmatched semantic alignment in current contrastive learning by expanding the augmentations from data space to feature space. The proposed semantics-consistent feature search (SCFS) adaptively searches semantics-consistent local features between different views for contrast, while suppressing irrelevant local features during pre-training. It conducts contrast learning between feature augmentation and data augmentation. The experimental results demonstrate that SCFS can learn to focus on meaningful object regions and effectively improve the performance of self-supervised learning. The feature search procedure in SCFS is learnable parameter-free. We will utilize the self-attention mechanism in Transformer to perform the feature search procedure to further boost its performance in future work. + +# References + +[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022. 1 +[2] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021. 2, 5 +[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. NeurIPS, 33:9912-9924, 2020. 2, 3, 5, 6, 12 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 2, 3, 4, 5, 6, 7, 8, 11, 12 +[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. PMLR, 2020. 1, 2, 5, 6, 7, 12 +[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. NeurIPS, 33:22243-22255, 2020. 1, 5 +[7] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 6, 7 +[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021. 2, 5 +[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021. 2, 5, 12 +[10] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In ICCV, pages 9588–9597, 2021. 2, 5 +[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 88(2):303–338, 2010. 6 +[12] Spyros Gidaris, Andrei Bursuc, Gilles Puy, Nikos Komodakis, Matthieu Cord, and Patrick Perez. Obow: Online bag-of-visual-words generation for self-supervised learning. In CVPR, pages 6830–6840, 2021. 2, 5 +[13] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 33:21271-21284, 2020. 2, 5, 6, 7, 12 +[14] Yuanfan Guo, Minghao Xu, Jiawen Li, Bingbing Ni, Xuanyu Zhu, Zhenbang Sun, and Yi Xu. Hcsc: Hierarchical contrastive selective coding. In CVPR, pages 9706-9715, 2022. 2, 5 +[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1 + +[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 1, 2, 5, 6 +[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 6, 7 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 2, 3, 4, 5, 6, 7 +[19] Qianjiang Hu, Xiao Wang, Wei Hu, and Guo-Jun Qi. Adco: Adversarial contrast for efficient learning of unsupervised representations from self-trained negative adversaries. In CVPR, pages 1074-1083, 2021. 2, 5 +[20] Soroush Abbasi Koohpayegani, Ajinkya Tejankar, and Hamed Piri siavash. Mean shift for self-supervised learning. In ICCV, pages 10326-10335, 2021. 2, 5 +[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 6 +[22] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[23] Xiaoni Li, Yu Zhou, Yifei Zhang, Aoting Zhang, Wei Wang, Ning Jiang, Haiying Wu, and Weiping Wang. Dense semantic contrast for self-supervised visual representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1368-1376, 2021. 2, 6 +[24] Zeming Li, Songtao Liu, and Jian Sun. Momentum^2 teacher: Momentum teacher with momentum statistics for self-supervised learning. arXiv preprint arXiv:2101.07525, 2021. 2 +[25] Zhaowen Li, Yousong Zhu, Fan Yang, Wei Li, Chaoyang Zhao, Yingying Chen, Zhiyang Chen, Jiahao Xie, Liwei Wu, Rui Zhao, et al. Univip: A unified framework for self-supervised visual pre-training. In CVPR, pages 14627-14636, 2022. 2, 5, 7 +[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 6, 7 +[27] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 11 +[28] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6 +[29] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In $CVPR$ , pages 3498-3505. IEEE, 2012. 6 +[30] Nikhil Parthasarathy, SM Eslami, João Carreira, and Olivier J Henaff. Self-supervised video pretraining yields strong image representations. arXiv preprint arXiv:2210.06433, 2022. 3, 12 +[31] Xiangyu Peng, Kai Wang, Zheng Zhu, Mang Wang, and Yang You. Crafting better contrastive views for siamese representation learning. In CVPR, pages 16031-16040, 2022. 2 + +[32] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 5, 6 +[33] Byungseok Roh, Wuhyun Shin, Ildoo Kim, and Sungwooong Kim. Spatially consistent representation learning. In CVPR, pages 1144–1153, 2021. 2 +[34] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015. 5, 7 +[35] Kaiyou Song, Jin Xie, Shan Zhang, and Zimeng Luo. Multi-mode online knowledge distillation for self-supervised visual representation learning. In CVPR, 2023. 2 +[36] Ajinkya Tejankar, Soroush Abbasi Koohpayegani, Vipin Pillai, Paolo Favaro, and Hamed Pirsiavash. Isd: Self-supervised learning by iterative similarity distillation. In ICCV, pages 9609-9618, 2021. 2, 5 +[37] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6 +[38] Haochen Wang, Kaiyou Song, Junsong Fan, Yuxi Wang, Jin Xie, and Zhaoxiang Zhang. Hard patches mining for masked image modeling. In CVPR, 2023. 1 +[39] Xiao Wang and Guo-Jun Qi. Contrastive learning with stronger augmentations. IEEE TPAMI, 2022. 2, 5 +[40] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In CVPR, pages 3024-3033, 2021. 2, 6, 7 +[41] Zhaoqing Wang, Qiang Li, Guoxin Zhang, Pengfei Wan, Wen Zheng, Nannan Wang, Mingming Gong, and Tongliang Liu. Exploring set similarity for dense self-supervised representation learning. In CVPR, pages 16590-16599, 2022. 2, 3, 6 +[42] Fangyun Wei, Yue Gao, Zhirong Wu, Han Hu, and Stephen Lin. Aligning pretraining for detection via object-level contrastive learning. NeurIPS, 34, 2021. 2, 6 +[43] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, pages 3733-3742, 2018. 5 +[44] Tete Xiao, Colorado J Reed, Xiaolong Wang, Kurt Keutzer, and Trevor Darrell. Region similarity representation learning. In ICCV, pages 10539–10548, 2021. 2, 6 +[45] Enze Xie, Jian Ding, Wenhai Wang, Xiaohang Zhan, Hang Xu, Peize Sun, Zhenguo Li, and Ping Luo. Detco: Unsupervised contrastive learning for object detection. In ICCV, pages 8392-8401, 2021. 2, 5, 6 +[46] Jiahao Xie, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Unsupervised object-level representation learning from scene images. NeurIPS, 34:28864-28876, 2021. 2, 6, 7 +[47] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In CVPR, pages 16684-16693, 2021. 2 +[48] Haohang Xu, Xiaopeng Zhang, Hao Li, Lingxi Xie, Wenrui Dai, Hongkai Xiong, and Qi Tian. Seed the views: Hierarchi + +cal semantic alignment for contrastive representation learning. IEEE TPAMI, 2022. 2, 6 +[49] Chun-Hsiao Yeh, Cheng-Yao Hong, Yen-Chi Hsu, Tyng-Luh Liu, Yubei Chen, and Yann LeCun. Decoupled contrastive learning. In ECCV, pages 668-684. Springer, 2022. 2, 5 +[50] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, pages 12310-12320. PMLR, 2021. 2, 5 +[51] Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. In ICCV, pages 1476-1485, 2019. 5 \ No newline at end of file diff --git a/2023/Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning/images.zip b/2023/Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..394a584ec0a954b8cdaaff12a41bec3c0e9921dc --- /dev/null +++ b/2023/Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4445cd39809a26096d156658466fa04edd1a304d83f238f8df57bcf29caa7668 +size 727689 diff --git a/2023/Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning/layout.json b/2023/Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b1fc03d8f714251beb173edcea14a2880ef01ccd --- /dev/null +++ b/2023/Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning/layout.json @@ -0,0 +1,8896 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 103, + 446, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 103, + 446, + 140 + ], + "spans": [ + { + "bbox": [ + 148, + 103, + 446, + 140 + ], + "type": "text", + "content": "Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 161, + 457, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 457, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 457, + 194 + ], + "type": "text", + "content": "Kaiyou Song Shan Zhang Zimeng Luo Tong Wang Jin Xie Megvii Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 198, + 468, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 198, + 468, + 209 + ], + "spans": [ + { + "bbox": [ + 126, + 198, + 468, + 209 + ], + "type": "text", + "content": "{songkaiyou, zhangshan, luozimeng, wangtong, xiejin}@megvii.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 236, + 192, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 236, + 192, + 249 + ], + "spans": [ + { + "bbox": [ + 143, + 236, + 192, + 249 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 263, + 290, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 263, + 290, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 263, + 290, + 514 + ], + "type": "text", + "content": "In contrastive self-supervised learning, the common way to learn discriminative representation is to pull different augmented \"views\" of the same image closer while pushing all other images further apart, which has been proven to be effective. However, it is unavoidable to construct undesirable views containing different semantic concepts during the augmentation procedure. It would damage the semantic consistency of representation to pull these augmentations closer in the feature space indiscriminately. In this study, we introduce feature-level augmentation and propose a novel semantics-consistent feature search (SCFS) method to mitigate this negative effect. The main idea of SCFS is to adaptively search semantics-consistent features to enhance the contrast between semantics-consistent regions in different augmentations. Thus, the trained model can learn to focus on meaningful object regions, improving the semantic representation ability. Extensive experiments conducted on different datasets and tasks demonstrate that SCFS effectively improves the performance of self-supervised learning and achieves state-of-the-art performance on different downstream tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 560, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 560, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 560, + 288, + 693 + ], + "type": "text", + "content": "Due to the tremendous potential in learning discriminative feature representation without using data annotations, self-supervised learning, including contrastive learning [16, 5] and masked image modeling (MIM) [1, 15, 38] has received much attention in the representation learning field. Contrastive learning, as a type of discriminative self-supervised learning method, is heavily studied and has shown remarkable progress in the computer vision field in recent years. It aims at pulling different augmented \"views\" of the same image (positive pairs) closer while pushing diverse images (negative pairs) far from each other. To this end, a contrastive" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 236, + 539, + 422 + ], + "blocks": [ + { + "bbox": [ + 313, + 236, + 539, + 422 + ], + "lines": [ + { + "bbox": [ + 313, + 236, + 539, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 539, + 422 + ], + "type": "image", + "image_path": "4d603971121f051b93cdb76af88d0e95709712c0cdf30fedf59da5215805e844.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 425, + 548, + 557 + ], + "lines": [ + { + "bbox": [ + 304, + 425, + 548, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 548, + 557 + ], + "type": "text", + "content": "Figure 1. Semantic inconsistency of over-augmentation. (a) shows three augmentations of two images, in which the third augmentation is over-augmented and contains only background. (b) shows category probability distributions of the corresponding images in (a), which are obtained from a supervised pre-trained ResNet50 [18] model. (c)(d)(e) show three different samples of an image (the data-augmented image, the original image, and the semantics-consistent feature-augmented sample generated by Eq. (7) in this study) and their corresponding probability distributions, which point out that the over-augmented image generates different category with the original image, while the feature-augmented sample gets a balanced category probability." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 579, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 547, + 640 + ], + "type": "text", + "content": "loss between the features of different views extracted from an encoder network is employed to train the encoder network end-to-end. According to whether the negative pairs are used, current contrastive learning can be generally divided into two categories." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 548, + 714 + ], + "type": "text", + "content": "The first category [16, 5] utilizes both positive pairs and negative pairs for contrast. MoCo [16, 7] uses a momentum update mechanism to maintain a memory bank of negative examples. SimCLR [5, 6] directly trains a single encoder network with a large batch size to ensure sufficient positive and negative samples for learning. Based on MoCo and Sim" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 231, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 231, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 231, + 713 + ], + "type": "text", + "content": "1Code: https://github.com/skyoux/scfs" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16099" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": "CLR, some methods [24, 20, 36, 10, 49, 39, 19, 9, 48, 14] are proposed to improve the performance. For example, MSF [20], ISD [36] and NNCLR [10] aim to search semantics-consistent samples for contrast, solving the false negative problem. While some studies, such as Momentum2Teacher [24] and DCL [49], aim to solve the limitation that large batch size is necessary for satisfactory performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 158, + 288, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 288, + 302 + ], + "type": "text", + "content": "The second category of contrastive learning methods [13, 8, 3, 50, 2, 12, 4, 31, 35] only constructs positive pairs for contrast. Based on MoCo [16] and SimCLR [5], respectively, BYOL [13] and SimSiam [8] abandon the negative samples and use an asymmetric architecture to avoid model collapse. SwAV [3] uses online clustering to cluster samples and forces the consistency among cluster assignments of different augmentations. After that, some studies [4, 39, 10, 19] point out that enriching the augmented samples can improve the performance of contrastive learning. In addition, the study in [31] shows that improving the quality of positive augmented samples is important for self-supervised learning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 304, + 288, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 288, + 567 + ], + "type": "text", + "content": "However, it is unavoidable to construct data augmentations containing different semantic concepts. Fig. 1(a) shows three augmentations of two images, in which the third augmentation is over-augmented and contains only the background. Fig. 1(b) shows category probability distributions of the corresponding images in (a), which are obtained from a supervised pre-trained ResNet50 [18]. We observed that the probability distribution of over-augmented images changes greatly compared with the first two augmentations, which indicates that the semantic information of the over-augmented images deviates from the normally-augmented images. Similar observation can be found in Fig. 1(c)(d). The original image in (d) shows a max probability for \"ambulance\", while the over-augmented image in (c) represents the different category \"telescope\". Due to such semantic inconsistency, conducting contrastive learning on these overaugmentations is harmful to representation learning. In this study, we found that semantics-consistent feature augmentation (Fig. 1(e), generated by Eq. (7)) can balance the original semantics \"ambulance\" and the over-augmented semantics \"telescope\", which can alleviate the influence of semantic inconsistency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Motivated by this observation, we propose a novel semantics-consistent feature search (SCFS) method to alleviate the negative influence of semantic inconsistency in contrastive learning. SCFS utilizes the global feature of a view to adaptively search the semantics-consistent features of another view for contrast according to their similarity. It constructs informative feature augmentations and conducts contrast learning between feature augmentations and data augmentations. Thus, the pre-trained model can learn to focus on meaningful object regions to alleviate the negative influence of unmatched semantic alignment in current contrastive learning for better representation learning. In" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 191 + ], + "type": "text", + "content": "addition, the feature search is conducted on multiple layers of the backbone network, further enhancing the semantic alignment at different scales of features. Extensive experiments conducted on different datasets and tasks demonstrate that SCFS effectively improves the performance of self-supervised learning and achieves state-of-the-art performance on different downstream tasks. For example, it achieves state-of-the-art " + }, + { + "bbox": [ + 304, + 72, + 547, + 191 + ], + "type": "inline_equation", + "content": "75.7\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 191 + ], + "type": "text", + "content": " ImageNet top-1 accuracy under the pre-training setting of 1024 batch size and 800 epochs for ResNet50." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 192, + 521, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 192, + 521, + 204 + ], + "spans": [ + { + "bbox": [ + 317, + 192, + 521, + 204 + ], + "type": "text", + "content": "The main contributions of this study are threefold:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 212, + 545, + 346 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 317, + 212, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 212, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 317, + 212, + 545, + 270 + ], + "type": "text", + "content": "- A novel contrastive learning method, i.e., SCFS, is proposed, and it can enhance semantic alignment in contrastive learning. To our knowledge, this is the first work that defines a feature search task in contrastive learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 280, + 545, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 280, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 317, + 280, + 545, + 315 + ], + "type": "text", + "content": "- We expand contrastive learning from a data-to-data manner to a feature-to-data manner, which enriches the diversity of augmentations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 324, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 324, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 317, + 324, + 545, + 346 + ], + "type": "text", + "content": "- The proposed SCFS achieves state-of-the-art performance on different downstream tasks." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 358, + 397, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 358, + 397, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 358, + 397, + 371 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 378, + 547, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 547, + 486 + ], + "type": "text", + "content": "Recently, some studies [33, 44, 42, 45, 46, 25, 47, 40, 23, 41] pointed out that the problem of semantic inconsistency is more serious for downstream dense prediction tasks, such as object detection and instance segmentation. Therefore, these methods utilize region-level and pixel-level features for contrast. In this study, the proposed SCFS construct feature-level augmentations using dense feature maps. Therefore, this section introduces related studies that conduct contrastive learning using region-level and pixel-level features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 487, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 547, + 605 + ], + "type": "text", + "content": "Region-level contrastive learning. SCRL [33] minimizes the distance between two local features, which are cropped from two corresponding feature maps of two views. ReSim [44] aligns regional representations by sliding a fixed-sized window across the overlapping area between two views to improve the performance for localization-based tasks. SoCo [42], ORL [46], and UniVIP [25] extract object region proposals and use them to construct region-level features for contrastive learning. They achieve good performance for downstream dense prediction tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "text", + "content": "Pixel-level contrastive learning. To obtain a more fine-grained representation, several studies [47, 40, 41] design pixel-level contrastive learning task, which assumes that features extracted from the same pixel of different views should be treated as positive pairs while pixels from others must be distinguished. PixPro [47] utilizes a pixel propagation module to select similar pixel features for contrast and encourages consistency between positive pixel pairs. DenseCL [40] proposes a dense projection head to generate dense" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16100" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "feature vectors for pixel-level contrastive learning. SetSim [41] is designed to realize pixel-wise similarity learning by filtering out noisy backgrounds." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 288, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 288, + 275 + ], + "type": "text", + "content": "As summarized above, data augmentations bring rich information while increasing uncertainty in contrastive learning. While methods that utilize region-level features expand the granularity of feature representation by alleviating the influence of noises. Recently, to solve the problem of semantic inconsistency in video contrastive learning, VITO [30] uses an MLP to learn attention masks for each view independently and fuses the attention pooling features from multi-layers for contrastive learning. Unlike VITO [30], SCFS uses a cross-view search to calculate attention and conducts contrastive learning at multi-layers independently. In addition, our method bridges the correlation between data and feature augmentations and extends the contrastive learning task to a semantics-consistent feature search task." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 287, + 107, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 287, + 107, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 287, + 107, + 300 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 307, + 288, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 288, + 451 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 288, + 451 + ], + "type": "text", + "content": "The overall architecture of SCFS is shown in Fig. 2. It consists of an encoder and a momentum encoder. The momentum encoder is an exponential-moving-average version of the encoder. SCFS consists of two contrastive learning tasks: the contrast between data augmentations " + }, + { + "bbox": [ + 46, + 307, + 288, + 451 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_d)" + }, + { + "bbox": [ + 46, + 307, + 288, + 451 + ], + "type": "text", + "content": ", which is designed based on the baseline DINO [4], and the contrast between data augmentations and feature augmentations " + }, + { + "bbox": [ + 46, + 307, + 288, + 451 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{fs})" + }, + { + "bbox": [ + 46, + 307, + 288, + 451 + ], + "type": "text", + "content": ". In this section, we first introduce the contrast between data augmentations in Sec. 3.1. Then, the key feature search module of SCFS is introduced in detail in Sec. 3.2. Next, we will introduce the overall loss of SCFS in Sec. 3.3. Finally, the implementation details are presented in Sec. 3.4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 459, + 253, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 459, + 253, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 459, + 253, + 472 + ], + "type": "text", + "content": "3.1. Contrast Between Data Augmentations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 478, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 478, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 287, + 561 + ], + "type": "text", + "content": "Given a pair of global augmentations " + }, + { + "bbox": [ + 46, + 478, + 287, + 561 + ], + "type": "inline_equation", + "content": "(I_1" + }, + { + "bbox": [ + 46, + 478, + 287, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 478, + 287, + 561 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 46, + 478, + 287, + 561 + ], + "type": "text", + "content": " ) of an input image, the feature representations of the two augmentations are used to calculate the global contrastive loss. Specifically, " + }, + { + "bbox": [ + 46, + 478, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\pmb {f}_1 = E_\\pmb {\\theta}(\\pmb {I}_1)" + }, + { + "bbox": [ + 46, + 478, + 287, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 478, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\pmb{f^{\\prime}}_2 = E_{\\pmb{\\theta}^{\\prime}}(\\pmb {I}_2)" + }, + { + "bbox": [ + 46, + 478, + 287, + 561 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 46, + 478, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 46, + 478, + 287, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 478, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}^{\\prime}" + }, + { + "bbox": [ + 46, + 478, + 287, + 561 + ], + "type": "text", + "content": " are parameters of the encoder and the momentum encoder, respectively. " + }, + { + "bbox": [ + 46, + 478, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\pmb {f}_1,\\pmb{f}^{\\prime}_2\\in R^K" + }, + { + "bbox": [ + 46, + 478, + 287, + 561 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 478, + 287, + 561 + ], + "type": "text", + "content": " is the output dimension. " + }, + { + "bbox": [ + 46, + 478, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{1}" + }, + { + "bbox": [ + 46, + 478, + 287, + 561 + ], + "type": "text", + "content": " is normalized with a softmax function:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 568, + 287, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 568, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 114, + 568, + 287, + 599 + ], + "type": "interline_equation", + "content": "P _ {1} ^ {i} = \\frac {\\exp \\left(f _ {1} ^ {i} / \\tau\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(f _ {1} ^ {k} / \\tau\\right)} \\tag {1}", + "image_path": "643b81ea57477a633cf89a9817603f5d3b87449e656cca163712f1d2facd2801.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 608, + 287, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 608, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 46, + 608, + 287, + 691 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 608, + 287, + 691 + ], + "type": "inline_equation", + "content": "\\tau > 0" + }, + { + "bbox": [ + 46, + 608, + 287, + 691 + ], + "type": "text", + "content": " is a temperature parameter that controls the sharpness of the output distribution. Note that " + }, + { + "bbox": [ + 46, + 608, + 287, + 691 + ], + "type": "inline_equation", + "content": "P_{2}^{\\prime}" + }, + { + "bbox": [ + 46, + 608, + 287, + 691 + ], + "type": "text", + "content": " is obtained by normalizing " + }, + { + "bbox": [ + 46, + 608, + 287, + 691 + ], + "type": "inline_equation", + "content": "f_{2}^{\\prime}" + }, + { + "bbox": [ + 46, + 608, + 287, + 691 + ], + "type": "text", + "content": " with a similar softmax function with temperature " + }, + { + "bbox": [ + 46, + 608, + 287, + 691 + ], + "type": "inline_equation", + "content": "\\tau^{\\prime}" + }, + { + "bbox": [ + 46, + 608, + 287, + 691 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 608, + 287, + 691 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 46, + 608, + 287, + 691 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 608, + 287, + 691 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 46, + 608, + 287, + 691 + ], + "type": "text", + "content": " are fed to the momentum encoder and encoder symmetrically, and " + }, + { + "bbox": [ + 46, + 608, + 287, + 691 + ], + "type": "inline_equation", + "content": "P_{1}^{\\prime}" + }, + { + "bbox": [ + 46, + 608, + 287, + 691 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 608, + 287, + 691 + ], + "type": "inline_equation", + "content": "P_{2}" + }, + { + "bbox": [ + 46, + 608, + 287, + 691 + ], + "type": "text", + "content": " are obtained respectively. Following DINO [4], the cross-entropy loss is employed as the contrastive loss between two global views:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 92, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 92, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g} = - \\left(P _ {2} ^ {\\prime} \\log \\left(P _ {1}\\right) + P _ {1} ^ {\\prime} \\log \\left(P _ {2}\\right)\\right) \\tag {2}", + "image_path": "ae48d58da58b92be76cdf7dbd593ccbfae028651bf55f16a5938fde906f7f1e6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": "To enrich augmentations, the multi-crop strategy [3] is employed. Multiple local augmentations " + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_l" + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": " is also constructed and fed to the encoder: " + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_l = E_{\\boldsymbol{\\theta}}(\\mathbf{I}_l)" + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "inline_equation", + "content": "P_l" + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": " is obtained by normalizing " + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_l" + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": " with the softmax function with temperature " + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": ". The contrast between local views and global views can be calculated:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 337, + 150, + 546, + 183 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 150, + 546, + 183 + ], + "spans": [ + { + "bbox": [ + 337, + 150, + 546, + 183 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l} = \\sum_ {n = 1} ^ {N} - \\left(P _ {1} ^ {\\prime} \\log \\left(P _ {l} ^ {n}\\right) + P _ {2} ^ {\\prime} \\log \\left(P _ {l} ^ {n}\\right)\\right) \\tag {3}", + "image_path": "39b18704c0b57bd38690c5a9eda879ce8fa981f711ab764b7f65c69df76fe577.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 190, + 545, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 545, + 213 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 190, + 545, + 213 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 190, + 545, + 213 + ], + "type": "text", + "content": " denotes the number of local views. Thus, the overall loss is the sum of global loss and local loss:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 394, + 223, + 546, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 223, + 546, + 236 + ], + "spans": [ + { + "bbox": [ + 394, + 223, + 546, + 236 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = \\mathcal {L} _ {g} + \\mathcal {L} _ {l} \\tag {4}", + "image_path": "9ed74a2415fd1f0d15d0becb8435fdb9938a2bed4f9f408c048dc643726b6e9b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 243, + 503, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 243, + 503, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 503, + 255 + ], + "type": "text", + "content": "3.2. Semantics-Consistent Feature Search" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 261, + 547, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 261, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 547, + 344 + ], + "type": "text", + "content": "As introduced in Sec. 1, it is unavoidable to construct augmentations that contain different semantic concepts during the augmentation procedure. It's harmful to pull these augmentations close indiscriminately in the feature space. Therefore, we propose the SCFS method to enhance the contrast between semantics-consistent regions in different augmentations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 345, + 547, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 345, + 547, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 345, + 547, + 478 + ], + "type": "text", + "content": "The architecture of SCFS is shown in Fig. 2. By feeding the local augmentations " + }, + { + "bbox": [ + 304, + 345, + 547, + 478 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 304, + 345, + 547, + 478 + ], + "type": "text", + "content": " to the encoder, feature maps from different stages of the backbone ResNet50 [18] are extracted. Specifically, the output features from different stages, i.e., " + }, + { + "bbox": [ + 304, + 345, + 547, + 478 + ], + "type": "inline_equation", + "content": "Res2" + }, + { + "bbox": [ + 304, + 345, + 547, + 478 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 345, + 547, + 478 + ], + "type": "inline_equation", + "content": "Res3" + }, + { + "bbox": [ + 304, + 345, + 547, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 345, + 547, + 478 + ], + "type": "inline_equation", + "content": "Res4" + }, + { + "bbox": [ + 304, + 345, + 547, + 478 + ], + "type": "text", + "content": ", are utilized to conduct SCFS, ensuring that each stage of the backbone produces discriminative features: " + }, + { + "bbox": [ + 304, + 345, + 547, + 478 + ], + "type": "inline_equation", + "content": "\\{^{2}F_{l},^{3}F_{l},^{4}F_{l}\\} = E_{\\theta}(I_{l})" + }, + { + "bbox": [ + 304, + 345, + 547, + 478 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 345, + 547, + 478 + ], + "type": "inline_equation", + "content": "^i F_l\\in R^{W_l^i\\times H_l^i\\times C^i}" + }, + { + "bbox": [ + 304, + 345, + 547, + 478 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 345, + 547, + 478 + ], + "type": "inline_equation", + "content": "W_{l}^{i},H_{l}^{i},C^{i}" + }, + { + "bbox": [ + 304, + 345, + 547, + 478 + ], + "type": "text", + "content": " denote the width, height and channel dimension, respectively. Next, the global average pooling operation is conducted on each " + }, + { + "bbox": [ + 304, + 345, + 547, + 478 + ], + "type": "inline_equation", + "content": "^i F_{l}" + }, + { + "bbox": [ + 304, + 345, + 547, + 478 + ], + "type": "text", + "content": " in the spatial dimensions:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 348, + 486, + 546, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 486, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 348, + 486, + 546, + 521 + ], + "type": "interline_equation", + "content": "{ } ^ { i } \\boldsymbol { f } _ { l } = \\frac { 1 } { W _ { l } ^ { i } \\times H _ { l } ^ { i } } \\sum _ { x = 1 } ^ { W _ { l } ^ { i } } \\sum _ { y = 1 } ^ { H _ { l } ^ { i } } { } ^ { i } \\boldsymbol { F } _ { l } ( x , y , z ) \\tag {5}", + "image_path": "a0f0c01c595b592d4a26d07867491bc40e85ec43c465cbd715c27133d5bc5dbf.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 531, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 531, + 545, + 594 + ], + "type": "inline_equation", + "content": "{}^i f_l \\in R^{C^i}" + }, + { + "bbox": [ + 304, + 531, + 545, + 594 + ], + "type": "text", + "content": ". Meanwhile, the global augmentations " + }, + { + "bbox": [ + 304, + 531, + 545, + 594 + ], + "type": "inline_equation", + "content": "I_g" + }, + { + "bbox": [ + 304, + 531, + 545, + 594 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 531, + 545, + 594 + ], + "type": "inline_equation", + "content": "g = 1,2" + }, + { + "bbox": [ + 304, + 531, + 545, + 594 + ], + "type": "text", + "content": ") are fed to the momentum encoder to extract feature maps from different stages: " + }, + { + "bbox": [ + 304, + 531, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\{^2 F_g',^3 F_g',^4 F_g'\\} = E_{\\pmb{\\theta}'}(\\pmb {I}_g),{}^i F_g' \\in R^{W_g^i\\times H_g^i\\times C^i}, W_g^i,H_g^i,C^i" + }, + { + "bbox": [ + 304, + 531, + 545, + 594 + ], + "type": "text", + "content": " denote width, height and channel dimension, respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 594, + 547, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 679 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 679 + ], + "type": "text", + "content": "Then, based on " + }, + { + "bbox": [ + 304, + 594, + 547, + 679 + ], + "type": "inline_equation", + "content": "{}^i\\pmb {f}_l" + }, + { + "bbox": [ + 304, + 594, + 547, + 679 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 547, + 679 + ], + "type": "inline_equation", + "content": "{}^i\\pmb{F}_{g}^{\\prime}" + }, + { + "bbox": [ + 304, + 594, + 547, + 679 + ], + "type": "text", + "content": ", SCFS aims to adaptively search the most semantics-consistent features in " + }, + { + "bbox": [ + 304, + 594, + 547, + 679 + ], + "type": "inline_equation", + "content": "{}^i\\pmb{F}_{g}^{\\prime}" + }, + { + "bbox": [ + 304, + 594, + 547, + 679 + ], + "type": "text", + "content": " for contrast, while suppressing irrelevant features. In SCFS, each feature " + }, + { + "bbox": [ + 304, + 594, + 547, + 679 + ], + "type": "inline_equation", + "content": "{}^i\\pmb {f}_l" + }, + { + "bbox": [ + 304, + 594, + 547, + 679 + ], + "type": "text", + "content": " of the local data augmentations is treated as query, and the features " + }, + { + "bbox": [ + 304, + 594, + 547, + 679 + ], + "type": "inline_equation", + "content": "{}^i\\pmb{F}_{g}^{\\prime}" + }, + { + "bbox": [ + 304, + 594, + 547, + 679 + ], + "type": "text", + "content": " of the global augmentations are treated as keys. The similarity between " + }, + { + "bbox": [ + 304, + 594, + 547, + 679 + ], + "type": "inline_equation", + "content": "{}^i\\pmb {f}_l" + }, + { + "bbox": [ + 304, + 594, + 547, + 679 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 547, + 679 + ], + "type": "inline_equation", + "content": "{}^i\\pmb{F}_{g}^{\\prime}" + }, + { + "bbox": [ + 304, + 594, + 547, + 679 + ], + "type": "text", + "content": " is calculated:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 357, + 686, + 546, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 686, + 546, + 717 + ], + "spans": [ + { + "bbox": [ + 357, + 686, + 546, + 717 + ], + "type": "interline_equation", + "content": "\\boldsymbol {A} (x, y) = \\frac {{} ^ {i} \\boldsymbol {f} _ {l} \\cdot {} ^ {i} \\boldsymbol {F} _ {g} ^ {\\prime} (x , y)}{\\| ^ {i} \\boldsymbol {f} _ {l} \\| _ {2} \\| ^ {i} \\boldsymbol {F} _ {g} ^ {\\prime} (x , y) \\| _ {2}} \\tag {6}", + "image_path": "92b39bcc968bef48ec7c2469d6685692abbf42dd6e9c467bcdca0b85c97e37fc.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16101" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 70, + 342, + 213 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 342, + 213 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 342, + 213 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 342, + 213 + ], + "type": "image", + "image_path": "2f47ebc92a19132eefcdbaca58e53be431c38d99e09fc3c383ded5eae72b6d7d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "type": "text", + "content": "Figure 2. Overall architecture of the proposed semantics-consistent feature search (SCFS). It consists of an encoder and a momentum encoder. There are two contrastive learning tasks: the contrast between data augmentations " + }, + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_d)" + }, + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "type": "text", + "content": " in the final feature space and the feature search task conducted on multiple layers " + }, + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{fs})" + }, + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "type": "text", + "content": ". The details of the feature search procedure is shown on the right." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 345, + 71, + 496, + 213 + ], + "blocks": [ + { + "bbox": [ + 345, + 71, + 496, + 213 + ], + "lines": [ + { + "bbox": [ + 345, + 71, + 496, + 213 + ], + "spans": [ + { + "bbox": [ + 345, + 71, + 496, + 213 + ], + "type": "image", + "image_path": "e11192e29bee3c44aa0413a27f29c9325f5cabf6877fd1fcae5ca6defae4acc1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 267, + 290, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 290, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 290, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 267, + 290, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in R^{W_g^i \\times H_g^i}" + }, + { + "bbox": [ + 46, + 267, + 290, + 365 + ], + "type": "text", + "content": " is the attention map, and " + }, + { + "bbox": [ + 46, + 267, + 290, + 365 + ], + "type": "inline_equation", + "content": "x = 1, \\ldots, W_g^i" + }, + { + "bbox": [ + 46, + 267, + 290, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 267, + 290, + 365 + ], + "type": "inline_equation", + "content": "y = 1, \\ldots, H_g^i" + }, + { + "bbox": [ + 46, + 267, + 290, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 267, + 290, + 365 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_2" + }, + { + "bbox": [ + 46, + 267, + 290, + 365 + ], + "type": "text", + "content": " is the L2 norm. The attention map " + }, + { + "bbox": [ + 46, + 267, + 290, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 267, + 290, + 365 + ], + "type": "text", + "content": " activates the semantics-consistent regions of the local augmentation on the global augmentation. Thus, the higher portion of local regions can be searched. To select semantic features and suppress irrelevant local features, we directly multiply the attention map " + }, + { + "bbox": [ + 46, + 267, + 290, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 267, + 290, + 365 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 267, + 290, + 365 + ], + "type": "inline_equation", + "content": "{}^i\\mathbf{F}_g'" + }, + { + "bbox": [ + 46, + 267, + 290, + 365 + ], + "type": "text", + "content": " to obtain the semantics-consistent feature augmentations:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 369, + 287, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 369, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 133, + 369, + 287, + 386 + ], + "type": "interline_equation", + "content": "{ } ^ { i } \\boldsymbol { F } _ { l g } ^ { \\prime } = \\boldsymbol { A } \\cdot { } ^ { i } \\boldsymbol { F } _ { g } ^ { \\prime } ( 7 )", + "image_path": "3102805253634cb08118e448ab1b4d21410c9d417123ef6b7cef0f6173b130bb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 389, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 484 + ], + "type": "text", + "content": "This operation can be regarded as attention-weighted average pooling. Through feature search, " + }, + { + "bbox": [ + 46, + 389, + 287, + 484 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 389, + 287, + 484 + ], + "type": "text", + "content": " local data augmentations " + }, + { + "bbox": [ + 46, + 389, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\pmb{I}_l" + }, + { + "bbox": [ + 46, + 389, + 287, + 484 + ], + "type": "text", + "content": " can search " + }, + { + "bbox": [ + 46, + 389, + 287, + 484 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 389, + 287, + 484 + ], + "type": "text", + "content": " corresponding semantics-consistent features " + }, + { + "bbox": [ + 46, + 389, + 287, + 484 + ], + "type": "inline_equation", + "content": "{}^i\\pmb{F}_{lg}'" + }, + { + "bbox": [ + 46, + 389, + 287, + 484 + ], + "type": "text", + "content": " from a global data augmentation " + }, + { + "bbox": [ + 46, + 389, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\pmb{I}_g" + }, + { + "bbox": [ + 46, + 389, + 287, + 484 + ], + "type": "text", + "content": ". That is, in terms of the global data augmentation, " + }, + { + "bbox": [ + 46, + 389, + 287, + 484 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 389, + 287, + 484 + ], + "type": "text", + "content": " different features are constructed in the feature space through the feature search procedure. Therefore, we term " + }, + { + "bbox": [ + 46, + 389, + 287, + 484 + ], + "type": "inline_equation", + "content": "{}^i\\pmb{F}_{lg}'" + }, + { + "bbox": [ + 46, + 389, + 287, + 484 + ], + "type": "text", + "content": " as feature-level augmentations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 484, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 510 + ], + "type": "text", + "content": "Next, " + }, + { + "bbox": [ + 46, + 484, + 287, + 510 + ], + "type": "inline_equation", + "content": "{}^i\\mathbf{F}_l" + }, + { + "bbox": [ + 46, + 484, + 287, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 484, + 287, + 510 + ], + "type": "inline_equation", + "content": "{}^i\\mathbf{F}_{lg}'" + }, + { + "bbox": [ + 46, + 484, + 287, + 510 + ], + "type": "text", + "content": " are fed to corresponding projection heads to obtain their final representations for contrast:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 118, + 514, + 287, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 514, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 118, + 514, + 287, + 541 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} ^ {i} \\boldsymbol {f} _ {l} = \\mathrm {H} _ {i} \\left(^ {i} \\boldsymbol {F} _ {l}\\right) \\\\ ^ {i} \\boldsymbol {f} _ {l g} ^ {\\prime} = H _ {i} ^ {\\prime} \\left(^ {i} \\boldsymbol {F} _ {l g} ^ {\\prime}\\right) \\end{array} \\right. \\tag {8}", + "image_path": "a1ed75ea059a1f00e4951cb91083a8589a00663948339b0a0ff62ba4202d9cb2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{H}_i" + }, + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{H}'_i" + }, + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "type": "text", + "content": " denote the projection heads on the " + }, + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "type": "text", + "content": "-th layer of the encoder and the momentum encoder, respectively. " + }, + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "type": "inline_equation", + "content": "{}^i\\pmb {f}_l" + }, + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "type": "inline_equation", + "content": "{}^i\\pmb {f}_{lg}'" + }, + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "type": "text", + "content": " are normalized with softmax function with temperature " + }, + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "type": "inline_equation", + "content": "\\tau'" + }, + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "type": "text", + "content": ", respectively, as the same formulation in Eq. (1). The corresponding output probability " + }, + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "type": "inline_equation", + "content": "{}^iP_l" + }, + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "type": "inline_equation", + "content": "{}^iP_{lg}'" + }, + { + "bbox": [ + 46, + 545, + 289, + 629 + ], + "type": "text", + "content": " are employed to calculate the contrast loss between local data augmentations and feature augmentations:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 91, + 634, + 287, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 634, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 91, + 634, + 287, + 668 + ], + "type": "interline_equation", + "content": "{ } ^ { i } \\mathcal { L } _ { f s } = \\sum _ { g = 1 } ^ { 2 } \\sum _ { n = 1 } ^ { N } - \\left( ^ { i } P _ { l g } ^ { \\prime } \\log \\left( ^ { i } P _ { l } ^ { n } \\right) \\right) \\tag {9}", + "image_path": "6a2794ea78b6e4c53946d3386afaea353ed2d1eee53388fd4e996323f6d276bc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 673, + 280, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 673, + 280, + 685 + ], + "spans": [ + { + "bbox": [ + 59, + 673, + 280, + 685 + ], + "type": "text", + "content": "The overall feature search loss is the sum of all layers:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 689, + 287, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 689, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 131, + 689, + 287, + 716 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f s} = \\sum_ {i \\in V _ {L}} ^ {i} \\mathcal {L} _ {f s} \\tag {10}", + "image_path": "074bc569acc0e0630a9f102440d9a1f7bcaeaedb62c3bb510a6b52f0c1031513.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 269, + 518, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 269, + 518, + 281 + ], + "spans": [ + { + "bbox": [ + 305, + 269, + 518, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 269, + 518, + 281 + ], + "type": "inline_equation", + "content": "V_{L}" + }, + { + "bbox": [ + 305, + 269, + 518, + 281 + ], + "type": "text", + "content": " denotes the set of layers to conduct SCFS." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 296, + 388, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 296, + 388, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 296, + 388, + 308 + ], + "type": "text", + "content": "3.3. Overall Loss" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 316, + 545, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 316, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 545, + 341 + ], + "type": "text", + "content": "The overall loss is the sum of the contrastive loss between data augmentations and the feature search loss:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 394, + 356, + 545, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 356, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 394, + 356, + 545, + 369 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {d} + \\mathcal {L} _ {f s} \\tag {11}", + "image_path": "1b894187f62bee4813ddca812b96b6773bb426454363fe54b35ed549d7fc832f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 385, + 547, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 385, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 305, + 385, + 547, + 469 + ], + "type": "text", + "content": "Through SCFS, the contrast between feature augmentations and data augmentations is bridged. The model can adaptively search the semantics-consistent features for contrast. Therefore, it can enhance the importance of semantics-consistent regions in different augmentations, alleviating the uncertainty in contrastive learning introduced by data augmentations that contain different semantic concepts." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 484, + 440, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 484, + 440, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 484, + 440, + 498 + ], + "type": "text", + "content": "3.4. Implementation Details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 505, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 505, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 305, + 505, + 545, + 552 + ], + "type": "text", + "content": "SCFS is based on DINO [4] and we follow the most hyper-parameter settings of DINO. For a fair comparison, the standard ResNet50 [18] is employed as the default backbone network." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "type": "text", + "content": "For data augmentation, the global augmentations consist of random cropping, resizing to " + }, + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "type": "text", + "content": ", random horizontal flip, gaussian blur, and color jittering. And the local augmentations consist of random cropping, resizing to " + }, + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "type": "inline_equation", + "content": "96 \\times 96" + }, + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "type": "text", + "content": ", random horizontal flip, gaussian blur, and color jittering. For feature augmentations in SCFS, the Res2, Res3, and Res4 layers are used. Two global views with " + }, + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "type": "inline_equation", + "content": "N = 8" + }, + { + "bbox": [ + 304, + 555, + 547, + 651 + ], + "type": "text", + "content": " local views are the default setting of augmentation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "text", + "content": "The projection head for the contrast between data augmentations consists of a four-layer multi-layer-perceptron (MLP) with the same architecture as DINO [4]. The projection head for feature search consists of three convolutional layers and two FC layers." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16102" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 282, + 411 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 282, + 411 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 282, + 411 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 282, + 411 + ], + "type": "table", + "html": "
MethodBatch SizeEpochsLPk-NN
Supervised25610076.274.8
SimCLR [5]4096100069.3-
BYOL [13]4096100074.366.9
BYOL [13]409620070.6-
SwAV [3]409680075.3-
SwAV [3]25620072.7-
MoCo-v2 [16]25620067.554.3
SimSiam [8]25620070.0-
ISD [36]25620069.862.0
MSF [20]25620071.464.0
NNCLR [10]409620070.7-
Barlow Twins [50]2048100073.2-
VICReg [2]2048100073.2-
OBoW [12]25620073.8-
DCL [49]25620066.9-
CLSA [39]25620073.3-
AdCo [19]25620073.2-
DetCo [45]25620068.6-
UniVIP [25]409620073.1-
HCSC [14]25620073.3-
MoCo-v3 [9]409630072.8-
MoCo-v3 [9]4096100074.6-
DINO* [4]25620073.064.0
DINO [4]408080075.367.5
SCFS25620073.965.5
SCFS102480075.768.5
", + "image_path": "a261d5d96710eba8221daecb261b08400865a1aabfe57128aaf815ebbacea0bf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 418, + 288, + 475 + ], + "lines": [ + { + "bbox": [ + 46, + 418, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 288, + 475 + ], + "type": "text", + "content": "Table 1. Linear probing and " + }, + { + "bbox": [ + 46, + 418, + 288, + 475 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 418, + 288, + 475 + ], + "type": "text", + "content": "-NN accuracy (\\%) on ImageNet. The result with “*” is reproduced for fair comparison. LP denotes linear probing. Bold font and underline indicate the best results under the setting of 256 batch size and 200 epochs and the setting of 1024 batch size and 800 epochs, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 495, + 128, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 495, + 128, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 495, + 128, + 509 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 516, + 288, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 516, + 288, + 599 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 288, + 599 + ], + "type": "text", + "content": "In this section, comprehensive experiments are conducted to demonstrate the effectiveness of SCFS. We evaluate the performance on different downstream tasks, including ImageNet classification, object detection, instance segmentation, and other classification task on small datasets. In addition, we conduct ablation experiments to analyze the influence of each component in SCFS." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 609, + 274, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 274, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 274, + 624 + ], + "type": "text", + "content": "4.1. Comparing with SSL methods on ImageNet" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "-NN and Linear Probing Accuracy on ImageNet. After pre-training on the ImageNet ILSVRC-2012 [34] training set, the pre-trained models are evaluated on the ImageNet ILSVRC-2012 validation set. For " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "-NN, it is evaluated as in study [43]. For linear probing, we train a linear classifier from scratch based on the feature extracted by a fixed backbone with 100 epochs [16]. The top-1 accuracy is adopted" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 312, + 70, + 541, + 184 + ], + "blocks": [ + { + "bbox": [ + 312, + 70, + 541, + 184 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 541, + 184 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 541, + 184 + ], + "type": "table", + "html": "
MethodBatch Size EpochsTop-1Top-5
1%10%1%10%
Supervised [51]2569025.456.448.480.4
SimCLR [5]4096100048.365.675.587.8
BYOL [13]4096100053.268.878.489.0
SwAV [3]409680053.970.278.589.9
DINO [4]408080050.269.374.089.1
SCFS102480054.370.578.690.2
", + "image_path": "557e7ef00384ae78338e786dfa9c39cb02142d5ddddc4717c6cc15d21f3d672e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 313, + 221, + 539, + 327 + ], + "blocks": [ + { + "bbox": [ + 306, + 191, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 306, + 191, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 306, + 191, + 545, + 213 + ], + "type": "text", + "content": "Table 2. Evaluation on small labeled ImageNet. Bold font indicates the best result." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 221, + 539, + 327 + ], + "lines": [ + { + "bbox": [ + 313, + 221, + 539, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 539, + 327 + ], + "type": "table", + "html": "
MethodEpochsAPbAPb50APb75
Scratch-33.860.233.1
Supervised9053.581.358.8
SimCLR [5]100056.381.962.5
BYOL [13]30051.981.056.5
SwAV [3]40045.177.446.5
DINO [4]80055.982.162.3
SCFS80057.483.063.6
", + "image_path": "2588c583ac30b4edffc275433e0569e3afdec46d580cedca47138db6e2a5ed9d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 335, + 547, + 357 + ], + "lines": [ + { + "bbox": [ + 306, + 335, + 547, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 547, + 357 + ], + "type": "text", + "content": "Table 3. Results for PASCAL VOC object detection using Faster R-CNN [32] with ResNet50-C4. Bold font indicates the best result." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 376, + 406, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 406, + 387 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 406, + 387 + ], + "type": "text", + "content": "as the evaluation metric." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 388, + 547, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 547, + 532 + ], + "type": "text", + "content": "The results are reported in Tab. 1. With the standard ResNet50 [18] architecture and pre-trained with 256 batch size for 200 epoch, the proposed SCFS achieves the best " + }, + { + "bbox": [ + 304, + 388, + 547, + 532 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 388, + 547, + 532 + ], + "type": "text", + "content": "-NN top-1 accuracy " + }, + { + "bbox": [ + 304, + 388, + 547, + 532 + ], + "type": "inline_equation", + "content": "65.5\\%" + }, + { + "bbox": [ + 304, + 388, + 547, + 532 + ], + "type": "text", + "content": " and the best linear probing top-1 accuracy " + }, + { + "bbox": [ + 304, + 388, + 547, + 532 + ], + "type": "inline_equation", + "content": "73.9\\%" + }, + { + "bbox": [ + 304, + 388, + 547, + 532 + ], + "type": "text", + "content": ", outperforming its baseline DINO [4] by " + }, + { + "bbox": [ + 304, + 388, + 547, + 532 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 304, + 388, + 547, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 388, + 547, + 532 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 304, + 388, + 547, + 532 + ], + "type": "text", + "content": ", respectively. In addition, with 1024 batch size and 800 epoch, SCFS achieves the best " + }, + { + "bbox": [ + 304, + 388, + 547, + 532 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 388, + 547, + 532 + ], + "type": "text", + "content": "-NN accuracy " + }, + { + "bbox": [ + 304, + 388, + 547, + 532 + ], + "type": "inline_equation", + "content": "(68.5\\%)" + }, + { + "bbox": [ + 304, + 388, + 547, + 532 + ], + "type": "text", + "content": " and linear probing accuracy " + }, + { + "bbox": [ + 304, + 388, + 547, + 532 + ], + "type": "inline_equation", + "content": "(75.7\\%)" + }, + { + "bbox": [ + 304, + 388, + 547, + 532 + ], + "type": "text", + "content": ", outperforming the accuracy of DINO [4] trained with 4080 batch size for 800 epoch. This result demonstrates that SCFS can improve the representation learning performance by searching semantics-consistent features for contrast." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 532, + 548, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 548, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 548, + 652 + ], + "type": "text", + "content": "Semi-Supervised Learning on ImageNet. In this part, we evaluate the performance of SCFS under the semi-supervised setting. Specifically, we use " + }, + { + "bbox": [ + 304, + 532, + 548, + 652 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 532, + 548, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 532, + 548, + 652 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 532, + 548, + 652 + ], + "type": "text", + "content": " of the labeled training data from ImageNet [34] for finetuning, which follows the semi-supervised protocol in SimCLR [5]. The same splits of " + }, + { + "bbox": [ + 304, + 532, + 548, + 652 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 532, + 548, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 532, + 548, + 652 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 532, + 548, + 652 + ], + "type": "text", + "content": " of ImageNet labeled training data in SimCLRv2 [6] are used. The results are reported in Tab. 2. After finetuning using " + }, + { + "bbox": [ + 304, + 532, + 548, + 652 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 532, + 548, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 532, + 548, + 652 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 532, + 548, + 652 + ], + "type": "text", + "content": " training data, SCFS outperforms all the compared methods. The results demonstrate that SCFS achieves the best feature representation quality." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 659, + 521, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 521, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 521, + 672 + ], + "type": "text", + "content": "4.2. Transfer Learning on Downstream Tasks" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Object Detection and Instance Segmentation. In this part, we evaluate the representations of SCFS on dense prediction tasks, i.e., object detection and instance segmentation," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16103" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 78, + 70, + 515, + 288 + ], + "blocks": [ + { + "bbox": [ + 78, + 70, + 515, + 288 + ], + "lines": [ + { + "bbox": [ + 78, + 70, + 515, + 288 + ], + "spans": [ + { + "bbox": [ + 78, + 70, + 515, + 288 + ], + "type": "table", + "html": "
MethodEpochs1×schedule2×schedule
\\(AP^b\\)\\(AP^b_{50}\\)\\(AP^b_{75}\\)\\(AP^s\\)\\(AP^s_{50}\\)\\(AP^s_{75}\\)\\(AP^b\\)\\(AP^b_{50}\\)\\(AP^b_{75}\\)\\(AP^s\\)\\(AP^s_{50}\\)\\(AP^s_{75}\\)
Scratch-31.049.533.228.546.830.438.457.542.034.754.837.2
Supervised9038.959.642.735.456.538.141.361.345.037.358.340.3
MoCo [16]20038.558.942.035.155.937.740.861.644.736.958.439.7
MoCo v2 [7]20040.460.244.236.457.238.941.761.645.637.658.740.5
BYOL [13]30040.461.644.137.258.839.842.362.646.238.359.641.1
SwAV [3]400------42.362.846.338.260.041.0
\\(ReSim-FPN^T\\) [44]20039.860.243.536.057.138.641.461.945.437.559.140.3
SetSim [41]20040.260.743.936.457.739.041.662.445.937.759.440.6
DenseCL [40]20040.359.944.336.457.039.241.261.945.137.358.940.1
DSC [23]20039.458.943.235.756.138.3------
HSA [48]80040.260.943.936.557.939.142.263.046.138.159.940.9
DetCo [45]80040.161.043.936.458.038.9------
ORL* [46]80040.360.244.436.357.338.9------
DINO [4]80040.061.643.436.558.639.141.962.646.037.859.740.6
SCFS80040.561.844.036.758.839.242.163.446.138.160.241.0
", + "image_path": "9805798297ae8665ee55cf52316dd8d4159f48c59879cfe4b813a0bac8a47fab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 115, + 326, + 480, + 407 + ], + "blocks": [ + { + "bbox": [ + 46, + 294, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 294, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 547, + 319 + ], + "type": "text", + "content": "Table 4. Object detection and instance segmentation on COCO using Mask R-CNN [17] with ResNet50-FPN. Bold font indicates the best result. * indicates pre-training on COCO [26]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 115, + 326, + 480, + 407 + ], + "lines": [ + { + "bbox": [ + 115, + 326, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 115, + 326, + 480, + 407 + ], + "type": "table", + "html": "
MethodCIFAR-10CIFAR-100CUB-BirdStanford-CarsAircraftOxford-Pets
Supervised97.586.481.392.186.092.1
SimCLR[5]97.785.9-91.388.189.2
BYOL[13]97.886.1-91.688.191.7
DINO[4]*97.786.681.091.187.491.5
SCFS97.886.782.791.688.591.9
", + "image_path": "3ecca5e5e90135872ecaeddb8722b14257db43a983d5059353969ddff9008647.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 415, + 547, + 437 + ], + "lines": [ + { + "bbox": [ + 46, + 415, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 547, + 437 + ], + "type": "text", + "content": "Table 5. Transfer learning results from ImageNet with the standard ResNet50 [18]. * denotes the results are reproduced in this study. Bold font indicates the best result." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 454, + 289, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 454, + 289, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 454, + 289, + 588 + ], + "type": "text", + "content": "on mainstream datasets PASCAL VOC [11] and MS COCO [26] datasets. On the PASCAL VOC dataset [11], the trainval07+12 set is used as the training set, and the test2007 set is used as the test set. Following [42], Faster R-CNN detector [32] with the ResNet50-C4 backbone initialized by the self-supervised pre-trained model is trained end-to-end. On the COCO dataset, the train2017 set is used for training and the val2017 set is used for evaluation. The Mask R-CNN [17] with R50-FPN is used. The " + }, + { + "bbox": [ + 46, + 454, + 289, + 588 + ], + "type": "inline_equation", + "content": "\\mathrm{AP^{b}}" + }, + { + "bbox": [ + 46, + 454, + 289, + 588 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 454, + 289, + 588 + ], + "type": "inline_equation", + "content": "\\mathrm{AP_{50}^{b}}" + }, + { + "bbox": [ + 46, + 454, + 289, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 454, + 289, + 588 + ], + "type": "inline_equation", + "content": "\\mathrm{AP_{75}^{b}}" + }, + { + "bbox": [ + 46, + 454, + 289, + 588 + ], + "type": "text", + "content": " metrics are used for object detection. While the " + }, + { + "bbox": [ + 46, + 454, + 289, + 588 + ], + "type": "inline_equation", + "content": "\\mathrm{AP^{s}}" + }, + { + "bbox": [ + 46, + 454, + 289, + 588 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 454, + 289, + 588 + ], + "type": "inline_equation", + "content": "\\mathrm{AP_{50}^{s}}" + }, + { + "bbox": [ + 46, + 454, + 289, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 454, + 289, + 588 + ], + "type": "inline_equation", + "content": "\\mathrm{AP_{75}^{s}}" + }, + { + "bbox": [ + 46, + 454, + 289, + 588 + ], + "type": "text", + "content": " metrics are used for instance segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 590, + 288, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 288, + 662 + ], + "type": "text", + "content": "The experimental results are shown in Tab. 3 and Tab. 4. SCFS achieves best performance on the two datasets. For example, on VOC, SCFS achieves " + }, + { + "bbox": [ + 46, + 590, + 288, + 662 + ], + "type": "inline_equation", + "content": "57.4\\% \\mathrm{AP}^{\\mathrm{b}}" + }, + { + "bbox": [ + 46, + 590, + 288, + 662 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 590, + 288, + 662 + ], + "type": "inline_equation", + "content": "83.0\\% \\mathrm{AP}_{50}^{\\mathrm{b}}" + }, + { + "bbox": [ + 46, + 590, + 288, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 590, + 288, + 662 + ], + "type": "inline_equation", + "content": "63.6\\% \\mathrm{AP}_{75}^{\\mathrm{b}}" + }, + { + "bbox": [ + 46, + 590, + 288, + 662 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 590, + 288, + 662 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^{\\mathrm{b}}" + }, + { + "bbox": [ + 46, + 590, + 288, + 662 + ], + "type": "text", + "content": " of SCFS outperforms its baseline DINO by " + }, + { + "bbox": [ + 46, + 590, + 288, + 662 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 46, + 590, + 288, + 662 + ], + "type": "text", + "content": ". These results show that SCFS also has good transfer ability on dense prediction tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "content": "Other Classification Tasks. In this part, we focus on the performance of self-supervised models when they are finetuned on small datasets, including CIFAR [22] and fine grained datasets [21, 37, 28, 29]. The results are shown in Tab. 5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 455, + 547, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 547, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 547, + 492 + ], + "type": "text", + "content": "The proposed SCFS shows the best performance on all the small datasets, which demonstrates that SCFS has good generalization ability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 502, + 492, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 502, + 492, + 515 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 492, + 515 + ], + "type": "text", + "content": "4.3. Pre-training on Uncurated Dataset" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": "The proposed SCFS can solve the problem of semantic inconsistency during pre-training, which is important when pre-training on uncurated datasets since this problem is more serious. To verify this, we pre-train SCFS and DINO on COCO [26], which is much more Uncurated than ImageNet. The same hyper-parameters used on ImageNet are applied to train the models with 512 batch size for 500 epochs. After pre-training, we fine-tune the pre-trained models on COCO for object detection and instance segmentation. The Mask R-CNN [17] with R50-FPN is used. As shown in Tab. 6, SCFS improves the performance significantly compared to its baseline DINO. In addition, compared to other dense pixel-level and region-level methods, such as DenseCL [40] and ORL [46], SCFS also achieves the best performance. This experiment verifies that SCFS can effectively solve the problem of semantic inconsistency during pre-training." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16104" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 282, + 211 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 282, + 211 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 282, + 211 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 282, + 211 + ], + "type": "table", + "html": "
Methodk-NNAPbAPb50APb75APsAPs50APs75
Scratch-31.049.533.228.546.830.4
Supervised*-38.959.642.735.456.538.1
SimCLR [5]35.337.056.840.333.753.836.1
MoCov2 [7]39.438.558.142.134.855.337.3
BYOL [13]42.639.559.343.235.656.538.2
DenseCL [40]32.539.659.343.335.756.538.4
ORL [46]43.640.360.244.436.357.338.9
UniVIP [25]-40.8--36.8--
DINO [4]42.639.059.642.935.656.838.0
SCFS44.640.961.644.436.958.439.5
", + "image_path": "1478d593ae9846acc2ccbcbeeebf2af6822945c5c83eacdd8dacbe41198dc827.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 217, + 287, + 273 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 287, + 273 + ], + "type": "text", + "content": "Table 6. Pre-training and than Fine-tuning on COCO [26] using Mask R-CNN [17] with ResNet50-FPN and " + }, + { + "bbox": [ + 46, + 217, + 287, + 273 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 46, + 217, + 287, + 273 + ], + "type": "text", + "content": " schedule. All models pre-trained on COCO are pre-trained with 512 batch size for 800 epochs. Bold font indicates the best result. * indicates pre-training on ImageNet [34]" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 294, + 147, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 294, + 147, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 294, + 147, + 306 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 314, + 287, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 287, + 397 + ], + "type": "text", + "content": "We analyze the influence of each component in SCFS. To speed up the training time, the ImageNet100 dataset, which contains 100 randomly selected categories from ImageNet [34], is adopted. All the models are pre-trained on the ImageNet100 training set with 256 batch size for 200 epoch, and tested on the validation set. The " + }, + { + "bbox": [ + 46, + 314, + 287, + 397 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 314, + 287, + 397 + ], + "type": "text", + "content": "-NN and linear probing top-1 accuracy are used as the evaluation metrics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "text", + "content": "Influence of Different Contrast Modes. The contrast mode can be divided into three types: contrast between two global data augmentations used in all contrastive learning methods " + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "inline_equation", + "content": "(\\mathrm{G_d}2\\mathrm{G_d})" + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "text", + "content": "; contrast between local data augmentations and global data augmentations used in multi-crop strategy " + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "inline_equation", + "content": "(\\mathrm{L_d}2\\mathrm{G_d})" + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "text", + "content": "; and contrast between local data augmentations and local feature augmentations used in SCFS " + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "inline_equation", + "content": "(\\mathrm{L_d}2\\mathrm{L_f})" + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 483, + 288, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 288, + 592 + ], + "type": "text", + "content": "The results are shown in Tab. 7. With multi-crop, DINO [4] (81.1%) improves accuracy by 3.0% compared to DINO without multi-crop baseline. SCFS (84.8%) further improves accuracy by 3.7% by introducing a contrast between local data augmentation and local feature augmentation. Some attention maps of SCFS and DINO are shown in Fig. 3. SCFS can more accurately focus on semantics-consistent regions between global view and local views, while DINO is easily influenced by background." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 593, + 288, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 288, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 288, + 640 + ], + "type": "text", + "content": "We also add multi-layer feature contrastive learning on DINO. The result in Tab. 7 (the \"DINO w ML\" row) verifies the improvements of SCFS are not totally owed to multi-layer contrast." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 713 + ], + "type": "text", + "content": "In addition, we directly crop the corresponding region of local augmentation on the feature map of global augmentation for contrastive learning. As shown in Tab. 7, this variant (ROIAlign) of SCFS also outperforms the DINO baseline, which shows that the directly cropped features are also beneficial for contrastive learning. And the ROIAlign variant" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 544, + 163 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 544, + 163 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 544, + 163 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 544, + 163 + ], + "type": "table", + "html": "
Contrast ModeGd2GdLd2GdLd2LfGd2Gfk-NNLP
DINO w/o MC✓78.183.7
DINO✓✓81.187.0
DINO w ML✓✓82.287.4
SCFS✓✓✓84.889.2
ROI Crop✓✓✓83.988.1
SCFS w/o MC✓79.786.3
", + "image_path": "287e3074a6911c2fc81ac5c04e8a4c21eaf394efbd1df87d966a1160e6257505.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 170, + 545, + 193 + ], + "lines": [ + { + "bbox": [ + 305, + 170, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 305, + 170, + 545, + 193 + ], + "type": "text", + "content": "Table 7. Influence of different contrast modes. MC, ML, and LP denote multi-crop, multi-layer, and linear probing, respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 313, + 207, + 536, + 341 + ], + "blocks": [ + { + "bbox": [ + 313, + 207, + 536, + 341 + ], + "lines": [ + { + "bbox": [ + 313, + 207, + 536, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 536, + 341 + ], + "type": "image", + "image_path": "7f8aa03f2395644f4897a8537d84e58c10244e784f755028b6dc16c7b3747956.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 342, + 547, + 486 + ], + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 486 + ], + "type": "text", + "content": "Figure 3. Attention maps of SCFS (the third row) compared with DINO [4] (the second row). In each example, (a) shows a global image, and its four local images in (b) are constructed by " + }, + { + "bbox": [ + 304, + 342, + 547, + 486 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 304, + 342, + 547, + 486 + ], + "type": "text", + "content": " jigsaw. (d) and (f) show the attention maps that highlight the semantics-consistent regions between the local images in (b) and the global image in (a). They are obtained by multiplying the globally average pooled feature maps from the encoder (Res4) of the local images in (b) with the feature map (Res4) of the global image in (a). And the encoder is the trained DINO ResNet50 model and SCFS ResNet50 model in (d) and (f), respectively. (c) and (e) show the mean attention maps of DINO and SCFS respectively, which are obtained by multiplying the mean globally average pooled feature map of the four local images in (b) with feature map of the global image in (a)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 507, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 547, + 555 + ], + "type": "text", + "content": "of SCFS achieves lower accuracy than SCFS, demonstrating that the soft feature search in SCFS is better than the hard ROIAlign since ROIAlign may damage the continuous semantic context of the feature map." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 556, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 547, + 665 + ], + "type": "text", + "content": "Further, we also test the performance of SCFS under the setting without multi-crop. That is, the feature search is conducted between two global data augmentations. We term this contrast mode as " + }, + { + "bbox": [ + 304, + 556, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{G_d}2\\mathrm{G_f}" + }, + { + "bbox": [ + 304, + 556, + 547, + 665 + ], + "type": "text", + "content": ". As shown in the \"SCFS w/o MC\" row, SCFS also improves the performance compared to its baseline (the \"DINO w/o MC\" row), which proves that SCFS is also helpful in solving the semantic inconsistency caused by other augmentations, not only the multi-crop augmentation strategy." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Influence of Multi-Layer Contrast. The influence of the feature layer that is used for feature search is analyzed. The Res2, Res3 and Res4 in the ResNet50 [18] backbone are evaluated. As shown in Tab. 8, the performance improves" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16105" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 271, + 127 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 271, + 127 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 271, + 127 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 271, + 127 + ], + "type": "table", + "html": "
Res2Res3Res4k-NNLinear Probing
✓82.086.1
✓✓84.387.5
✓✓✓84.889.2
", + "image_path": "2e8294c96dda461c7703e22846aa5179dd74bab00a7f7c92563d475e60abd0eb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 70, + 156, + 262, + 274 + ], + "blocks": [ + { + "bbox": [ + 70, + 156, + 262, + 274 + ], + "lines": [ + { + "bbox": [ + 70, + 156, + 262, + 274 + ], + "spans": [ + { + "bbox": [ + 70, + 156, + 262, + 274 + ], + "type": "image", + "image_path": "32cc7bcf55c389eba83a68856410513d9704b827fa327950f9ffe264b5128b8b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 277, + 274, + 289 + ], + "lines": [ + { + "bbox": [ + 47, + 277, + 274, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 274, + 289 + ], + "type": "text", + "content": "Figure 4. The " + }, + { + "bbox": [ + 47, + 277, + 274, + 289 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 277, + 274, + 289 + ], + "type": "text", + "content": "-NN accuracy of features from different layers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 70, + 299, + 262, + 415 + ], + "blocks": [ + { + "bbox": [ + 70, + 299, + 262, + 415 + ], + "lines": [ + { + "bbox": [ + 70, + 299, + 262, + 415 + ], + "spans": [ + { + "bbox": [ + 70, + 299, + 262, + 415 + ], + "type": "image", + "image_path": "b5fcca0d8ad18777935c3d077f1178bacd21da419d0ae5694771f3ec4b1e8ecc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 417, + 258, + 428 + ], + "lines": [ + { + "bbox": [ + 75, + 417, + 258, + 428 + ], + "spans": [ + { + "bbox": [ + 75, + 417, + 258, + 428 + ], + "type": "text", + "content": "Figure 5. Influence of local augmentation number." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 288, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 288, + 485 + ], + "type": "text", + "content": "with the increase of feature layer numbers, which demonstrates that conducting feature search on more layers is helpful for representation learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 628 + ], + "type": "text", + "content": "Further, we evaluate the " + }, + { + "bbox": [ + 46, + 486, + 287, + 628 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 486, + 287, + 628 + ], + "type": "text", + "content": "-NN accuracy using feature maps from different layers to observe the influence of feature search on the representation of middle layers. We also choose the features extracted by the Res2, Res3 and Res4 layer of ResNet50. The results are shown in Fig. 4. Compared with DINO [4], SCFS achieves better performance with features from all middle layers on ImageNet100, which verifies that enhancing the semantic consistency can improve the semantic representation of shallow layers. Compared with supervised learning, SCFS has higher performance on res2 and res3 layer, which shows that SCFS is more advantageous in the shallow layer feature representation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": "Influence of Local Augmentation Number. In this part, we analyze the performance difference with the change of local augmentation numbers. The results are shown in Fig. 5. The performance of DINO and SCFS is steadily improved when adding more local augmentations for contrast. In addition, SCFS improves the performance under different local augmentation numbers, which demonstrates that semantics-" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 545, + 109 + ], + "blocks": [ + { + "bbox": [ + 62, + 134, + 271, + 146 + ], + "lines": [ + { + "bbox": [ + 62, + 134, + 271, + 146 + ], + "spans": [ + { + "bbox": [ + 62, + 134, + 271, + 146 + ], + "type": "text", + "content": "Table 8. Influence of different feature augmentation layer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 545, + 109 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 109 + ], + "type": "table", + "html": "
Cropping Scale0.05-0.140.14-0.250.25-0.40.4-0.6
DINO [4]87.087.285.682.6
SCFS89.288.787.083.1
", + "image_path": "e9c309b5021e66bfb4a1297ecb79190f00e75c004ec45b7e6280ea2c0d9d1573.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 309, + 150, + 541, + 244 + ], + "blocks": [ + { + "bbox": [ + 306, + 117, + 545, + 139 + ], + "lines": [ + { + "bbox": [ + 306, + 117, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 306, + 117, + 545, + 139 + ], + "type": "text", + "content": "Table 9. Influence of random cropping scale. We report the linear probing top-1 accuracy." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 150, + 541, + 244 + ], + "lines": [ + { + "bbox": [ + 309, + 150, + 541, + 244 + ], + "spans": [ + { + "bbox": [ + 309, + 150, + 541, + 244 + ], + "type": "table", + "html": "
MethodBackboneBatch SizeEpochsk-NNLP
DINO [4]R10125620081.086.3
SCFSR10125620085.188.3
DINO [4]ViT-S-1625620075.080.4
SCFSViT-S-1625620076.381.0
DINO [4]ViT-B-1625620076.280.7
SCFSViT-B-1625620077.282.3
", + "image_path": "c9da0c2713e13b0ec73f1c0914867db92bcfadc816e901347292fb8bf7915f3f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 251, + 545, + 273 + ], + "lines": [ + { + "bbox": [ + 305, + 251, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 305, + 251, + 545, + 273 + ], + "type": "text", + "content": "Table 10. Experiments on other backbones. LP denotes linear probing." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 294, + 545, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 294, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 305, + 294, + 545, + 318 + ], + "type": "text", + "content": "consistent feature search is helpful to alleviate the influence of semantics inconsistent data augmentations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 319, + 546, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 546, + 403 + ], + "type": "text", + "content": "Influence of Random Cropping Scale. The hyperparameter of the random cropping scale influences the degree of semantic inconsistency. In this experiment, we analyze the performance of SCFS and DINO [4] under different random cropping scales. As shown in Tab. 9, owing to the feature search, SCFS is more adaptive to different cropping scales than the baseline DINO [4]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 404, + 547, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 404, + 547, + 499 + ], + "spans": [ + { + "bbox": [ + 305, + 404, + 547, + 499 + ], + "type": "text", + "content": "Experiments on Other Backbones. In this part, we conduct experiments on other backbones to further evaluate the effectiveness of SCFS. Apart for the default Reset50 used in other experiments, ResNet101 and Vision Transformer (ViT-S and ViT-B) are tested. The results are shown in Tab. 10. SCFS achieves significant improvement on different backbones compared to its baseline DINO, which demonstrates that SCFS is applicable to different backbones." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 513, + 383, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 513, + 383, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 513, + 383, + 525 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": "In this study, we aim to alleviate the problem of unmatched semantic alignment in current contrastive learning by expanding the augmentations from data space to feature space. The proposed semantics-consistent feature search (SCFS) adaptively searches semantics-consistent local features between different views for contrast, while suppressing irrelevant local features during pre-training. It conducts contrast learning between feature augmentation and data augmentation. The experimental results demonstrate that SCFS can learn to focus on meaningful object regions and effectively improve the performance of self-supervised learning. The feature search procedure in SCFS is learnable parameter-free. We will utilize the self-attention mechanism in Transformer to perform the feature search procedure to further boost its performance in future work." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16106" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 288, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 288, + 158 + ], + "type": "text", + "content": "[2] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 160, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 288, + 203 + ], + "type": "text", + "content": "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. NeurIPS, 33:9912-9924, 2020. 2, 3, 5, 6, 12" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 249 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 2, 3, 4, 5, 6, 7, 8, 11, 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 294 + ], + "type": "text", + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. PMLR, 2020. 1, 2, 5, 6, 7, 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 296, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 288, + 338 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. NeurIPS, 33:22243-22255, 2020. 1, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 341, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 288, + 373 + ], + "type": "text", + "content": "[7] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 375, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 288, + 407 + ], + "type": "text", + "content": "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 409, + 288, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 288, + 442 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 288, + 442 + ], + "type": "text", + "content": "[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021. 2, 5, 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "type": "text", + "content": "[10] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In ICCV, pages 9588–9597, 2021. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 489, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 520 + ], + "type": "text", + "content": "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 88(2):303–338, 2010. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "text", + "content": "[12] Spyros Gidaris, Andrei Bursuc, Gilles Puy, Nikos Komodakis, Matthieu Cord, and Patrick Perez. Obow: Online bag-of-visual-words generation for self-supervised learning. In CVPR, pages 6830–6840, 2021. 2, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 568, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 633 + ], + "type": "text", + "content": "[13] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 33:21271-21284, 2020. 2, 5, 6, 7, 12" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[14] Yuanfan Guo, Minghao Xu, Jiawen Li, Bingbing Ni, Xuanyu Zhu, Zhenbang Sun, and Yi Xu. Hcsc: Hierarchical contrastive selective coding. In CVPR, pages 9706-9715, 2022. 2, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 115 + ], + "type": "text", + "content": "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 547, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 148 + ], + "type": "text", + "content": "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 150, + 547, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 547, + 182 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 184, + 547, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 547, + 227 + ], + "type": "text", + "content": "[19] Qianjiang Hu, Xiao Wang, Wei Hu, and Guo-Jun Qi. Adco: Adversarial contrast for efficient learning of unsupervised representations from self-trained negative adversaries. In CVPR, pages 1074-1083, 2021. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 228, + 547, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 547, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 547, + 260 + ], + "type": "text", + "content": "[20] Soroush Abbasi Koohpayegani, Ajinkya Tejankar, and Hamed Piri siavash. Mean shift for self-supervised learning. In ICCV, pages 10326-10335, 2021. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 261, + 547, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 547, + 304 + ], + "type": "text", + "content": "[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 305, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 545, + 327 + ], + "type": "text", + "content": "[22] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 327, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 381 + ], + "type": "text", + "content": "[23] Xiaoni Li, Yu Zhou, Yifei Zhang, Aoting Zhang, Wei Wang, Ning Jiang, Haiying Wu, and Weiping Wang. Dense semantic contrast for self-supervised visual representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1368-1376, 2021. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 383, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 425 + ], + "type": "text", + "content": "[24] Zeming Li, Songtao Liu, and Jian Sun. Momentum^2 teacher: Momentum teacher with momentum statistics for self-supervised learning. arXiv preprint arXiv:2101.07525, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 426, + 547, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 547, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 547, + 480 + ], + "type": "text", + "content": "[25] Zhaowen Li, Yousong Zhu, Fan Yang, Wei Li, Chaoyang Zhao, Yingying Chen, Zhiyang Chen, Jiahao Xie, Liwei Wu, Rui Zhao, et al. Univip: A unified framework for self-supervised visual pre-training. In CVPR, pages 14627-14636, 2022. 2, 5, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 526, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 545, + 557 + ], + "type": "text", + "content": "[27] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 11" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 559, + 547, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 547, + 601 + ], + "type": "text", + "content": "[28] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 603, + 547, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 547, + 634 + ], + "type": "text", + "content": "[29] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In " + }, + { + "bbox": [ + 307, + 603, + 547, + 634 + ], + "type": "inline_equation", + "content": "CVPR" + }, + { + "bbox": [ + 307, + 603, + 547, + 634 + ], + "type": "text", + "content": ", pages 3498-3505. IEEE, 2012. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 636, + 547, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 547, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 547, + 678 + ], + "type": "text", + "content": "[30] Nikhil Parthasarathy, SM Eslami, João Carreira, and Olivier J Henaff. Self-supervised video pretraining yields strong image representations. arXiv preprint arXiv:2210.06433, 2022. 3, 12" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 680, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 547, + 713 + ], + "type": "text", + "content": "[31] Xiangyu Peng, Kai Wang, Zheng Zhu, Mang Wang, and Yang You. Crafting better contrastive views for siamese representation learning. In CVPR, pages 16031-16040, 2022. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16107" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[32] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 288, + 140 + ], + "type": "text", + "content": "[33] Byungseok Roh, Wuhyun Shin, Ildoo Kim, and Sungwooong Kim. Spatially consistent representation learning. In CVPR, pages 1144–1153, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 288, + 194 + ], + "type": "text", + "content": "[34] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015. 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 288, + 229 + ], + "type": "text", + "content": "[35] Kaiyou Song, Jin Xie, Shan Zhang, and Zimeng Luo. Multi-mode online knowledge distillation for self-supervised visual representation learning. In CVPR, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 231, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 231, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 231, + 288, + 274 + ], + "type": "text", + "content": "[36] Ajinkya Tejankar, Soroush Abbasi Koohpayegani, Vipin Pillai, Paolo Favaro, and Hamed Pirsiavash. Isd: Self-supervised learning by iterative similarity distillation. In ICCV, pages 9609-9618, 2021. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 276, + 288, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 288, + 307 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 288, + 307 + ], + "type": "text", + "content": "[37] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 309, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 309, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 309, + 288, + 342 + ], + "type": "text", + "content": "[38] Haochen Wang, Kaiyou Song, Junsong Fan, Yuxi Wang, Jin Xie, and Zhaoxiang Zhang. Hard patches mining for masked image modeling. In CVPR, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 343, + 288, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 343, + 288, + 364 + ], + "spans": [ + { + "bbox": [ + 49, + 343, + 288, + 364 + ], + "type": "text", + "content": "[39] Xiao Wang and Guo-Jun Qi. Contrastive learning with stronger augmentations. IEEE TPAMI, 2022. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 366, + 288, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 366, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 49, + 366, + 288, + 399 + ], + "type": "text", + "content": "[40] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In CVPR, pages 3024-3033, 2021. 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 400, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 400, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 400, + 288, + 453 + ], + "type": "text", + "content": "[41] Zhaoqing Wang, Qiang Li, Guoxin Zhang, Pengfei Wan, Wen Zheng, Nannan Wang, Mingming Gong, and Tongliang Liu. Exploring set similarity for dense self-supervised representation learning. In CVPR, pages 16590-16599, 2022. 2, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 456, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 456, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 49, + 456, + 288, + 487 + ], + "type": "text", + "content": "[42] Fangyun Wei, Yue Gao, Zhirong Wu, Han Hu, and Stephen Lin. Aligning pretraining for detection via object-level contrastive learning. NeurIPS, 34, 2021. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 490, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 490, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 49, + 490, + 288, + 521 + ], + "type": "text", + "content": "[43] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, pages 3733-3742, 2018. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 523, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 523, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 523, + 288, + 555 + ], + "type": "text", + "content": "[44] Tete Xiao, Colorado J Reed, Xiaolong Wang, Kurt Keutzer, and Trevor Darrell. Region similarity representation learning. In ICCV, pages 10539–10548, 2021. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 557, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 557, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 49, + 557, + 288, + 601 + ], + "type": "text", + "content": "[45] Enze Xie, Jian Ding, Wenhai Wang, Xiaohang Zhan, Hang Xu, Peize Sun, Zhenguo Li, and Ping Luo. Detco: Unsupervised contrastive learning for object detection. In ICCV, pages 8392-8401, 2021. 2, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 602, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 602, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 49, + 602, + 288, + 644 + ], + "type": "text", + "content": "[46] Jiahao Xie, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Unsupervised object-level representation learning from scene images. NeurIPS, 34:28864-28876, 2021. 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 647, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 288, + 689 + ], + "type": "text", + "content": "[47] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In CVPR, pages 16684-16693, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 692, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 692, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 692, + 288, + 713 + ], + "type": "text", + "content": "[48] Haohang Xu, Xiaopeng Zhang, Hao Li, Lingxi Xie, Wenrui Dai, Hongkai Xiong, and Qi Tian. Seed the views: Hierarchi" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 546, + 196 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 327, + 73, + 546, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 546, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 546, + 95 + ], + "type": "text", + "content": "cal semantic alignment for contrastive representation learning. IEEE TPAMI, 2022. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 96, + 546, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 546, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 546, + 129 + ], + "type": "text", + "content": "[49] Chun-Hsiao Yeh, Cheng-Yao Hong, Yen-Chi Hsu, Tyng-Luh Liu, Yubei Chen, and Yann LeCun. Decoupled contrastive learning. In ECCV, pages 668-684. Springer, 2022. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 130, + 546, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 546, + 163 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 546, + 163 + ], + "type": "text", + "content": "[50] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, pages 12310-12320. PMLR, 2021. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 164, + 546, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 164, + 546, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 546, + 196 + ], + "type": "text", + "content": "[51] Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. In ICCV, pages 1476-1485, 2019. 5" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16108" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Semantify_ Simplifying the Control of 3D Morphable Models Using CLIP/ff2d3341-5fbc-47bb-9ce8-0e16c61108bc_content_list.json b/2023/Semantify_ Simplifying the Control of 3D Morphable Models Using CLIP/ff2d3341-5fbc-47bb-9ce8-0e16c61108bc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..40a8ddc0246a5dfefa5927677c1c0cf9ee69e7ee --- /dev/null +++ b/2023/Semantify_ Simplifying the Control of 3D Morphable Models Using CLIP/ff2d3341-5fbc-47bb-9ce8-0e16c61108bc_content_list.json @@ -0,0 +1,1359 @@ +[ + { + "type": "text", + "text": "Semantify: Simplifying the Control of 3D Morphable Models using CLIP", + "text_level": 1, + "bbox": [ + 114, + 130, + 854, + 154 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/473800955cb810138a380f9346d36cad6a30af74e50d3e7fbaf71ad28b51722a.jpg", + "image_caption": [ + "Figure 1: Semantify offers a method to create and edit a 3D parametric model using semantically meaningful descriptors. Semantify is based on a self-supervised method that utilizes the semantic power of CLIP language-vision model to build a mapping between semantic descriptors to 3DMM model coefficients. This can be used in an interactive application defining a slider for each descriptor (a), or to fit a model to an image in a zero shot manner by feeding the image into CLIP and obtaining a vector of semantic scores that can be mapped to shape parameters (b)." + ], + "image_footnote": [], + "bbox": [ + 158, + 178, + 767, + 443 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 549, + 313, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present Semantify: a self-supervised method that utilizes the semantic power of CLIP language-vision foundation model [32] to simplify the control of 3D morphable models. Given a parametric model, training data is created by randomly sampling the model's parameters, creating various shapes and rendering them. The similarity between the output images and a set of word descriptors is calculated in CLIP's latent space. Our key idea is first to choose a small set of semantically meaningful and disentangled descriptors that characterize the 3DMM, and then learn a non-linear mapping from scores across this set to the parametric coefficients of the given 3DMM. The nonlinear mapping is defined by training a neural network without a human-in-the-loop. We present results on numerous 3DMMs: body shape models, face shape and expression models, as well as animal shapes. We demonstrate how our method defines a simple slider interface for intuitive modeling, and show how the mapping can be used to instantly fit a 3D parametric body shape to in-the-wild images. See our project page at https://omergral.github.io/Semantify/", + "bbox": [ + 75, + 580, + 473, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 549, + 632, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D modeling techniques have evolved tremendously over the last few years. Such techniques are used in countless industries including the burgeoning AR/VR industry, fashion design, game development, film, and many others. However, designing a high-quality 3D model is not a simple task for most people and might require a well-trained 3D artist. Even when using more recent parametric morphable models (3DMM) [24, 31, 45, 3, 42] it is still hard for humans to understand how to choose the correct set of parameters, e.g. to achieve a specific human body shape or facial expression. It is also difficult to find what are the limits of the given parametric model in terms of coverage and expressiveness. The reason is that in most cases, the provided set of parameters is not interpretable, as they are commonly calculated using automatic optimization mechanisms followed by dimensionality reduction using PCA. Thus, they carry no clear semantic meaning.", + "bbox": [ + 496, + 578, + 893, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To simplify the use of 3DMMs and allow for natural interactive human modeling, a key research question is - how to insert semantics into 3DMM control? Previous approaches [34] relied mostly on human intelligence and la", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14554", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "beling, which is often time consuming and expensive. In addition, previous methods create a large number of semantic control descriptors that are often correlated and entangled. This means it is difficult to anticipate the effect of each descriptor as changing a value in one may modify others, resulting in difficulties to control the model.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, instead of using human labeling, we rely on the remarkable abilities of huge foundation models that combine natural language and visual understanding. We present a self-supervised method that utilizes the semantic power of CLIP [32] to define a method to control 3DMMs that carries two main advantages. First, it is more natural for humans to use as it allows modeling using a small number of semantically meaningful descriptors, that cover the space of deformations but are disentangled. Second, it covers even extreme examples in the shape/pose space of parametric models as it utilizes CLIP's pre-training on a large number of images. The main idea is first to use CLIP to select a small subset of semantically meaningful and disentangled descriptors, and then learn a non-linear mapping of this set to the coefficients of the given 3DMM.", + "bbox": [ + 75, + 184, + 470, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a parametric 3DMM, we sample its parameter space to create a dataset containing a variety of 3D mesh shapes. We then render each mesh from different camera views to create a diverse set of images corresponding to the parameter samples (see Figure 2). Next, we gather a set of semantically descriptive textual terms related to the parametric 3D model, which we call descriptors. We encode both the images (using CLIP's image encoder) and the descriptors (using CLIP's text encoder) into CLIP's latent space and compare them. This defines a vector of similarity scores between the input vector of 3DMM coefficients (of each image sample) and each corresponding semantic descriptor. Next, we define a selection scheme to choose a small number of descriptors that are de-correlated to control the model. Lastly, we train a neural network to learn a mapping between the vector of similarity scores to the vector of 3DMM parameters.", + "bbox": [ + 75, + 412, + 470, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate how the learned mapping can be used to define an interface to control a 3DMM in a way that is simple and effective using a small set of semantically meaningful sliders (see Figure 1 and 4). Such sliders are easy to employ for designing high-quality 3D models and cover the shape space well. We demonstrate this for four parametric models: human face's shape and expression (FLAME [24]), human body shapes (SMPL [25] and SMPL-X [31]), and even animals (SMAL [45]). We also show how the mapping can be used to instantly fit a 3D parametric body shape to an input image that works well \"in the wild\" even in extreme poses and body shapes.", + "bbox": [ + 75, + 671, + 470, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contribution is a novel, self-supervised method for defining a small set of semantic descriptors to control a parametric model more naturally, by learning a mapping", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "from a semantic space to a parametric representation without human-in-the-loop. We demonstrate the effectiveness of our approach on several parametric models and utilize it to define a simple interface for modeling and to instantly fit a 3D model to images in the wild. We will release our code for future research.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 196, + 640, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work encompasses a wide range of disciplines, including 3D parametric modeling, language-vision models, and zero-shot reconstruction.", + "bbox": [ + 496, + 222, + 890, + 267 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Morphable Models. Much research has gone into improving 3D modeling of humans bodies, faces and other object classes, such as animals [11, 4, 25, 31]. 3D Morphable Models (3DMM) are a powerful tool to parameterize the variation in geometry of objects belonging to a certain class. Dating back to 1999, Blantz et al. [4] proposed 3DMMs to capture the variation of human faces and applied dimensionality reduction using Principal Component Analysis (PCA).", + "bbox": [ + 496, + 287, + 890, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast to a 3D artist freely controlling the vertices of a mesh, using each of these PCA axes to deform the vertices induces a strong statistical prior on the resulting geometry. The resulting deformations are global in nature, strongly constraining the geometry to the manifold of statistically plausible meshes, as the per-axis coefficients correlate with their likelihood. While initially proposed for human faces, statistical 3DMMs have been developed for human body models [3, 25, 40, 31, 42, 29], facial models [24, 37, 41, 43], and even 3D animal models [45, 44].", + "bbox": [ + 496, + 410, + 890, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic Models. PCA-based 3DMMs are efficiently and easily constructed, and admit an orthogonal and importance-ordered basis for the deformation. One disadvantage of these models is their lack of interpretability for humans. The axes of maximal variation in geometry, which form the deformation basis for each vertex, do not necessarily correspond to any semantically meaningful geometric change (e.g., raising an eyebrow), but instead, each axis induces a rather global change that may introduce correlations or effects that are unwanted to a human modeler, making it hard to obtain the desired shape or expression.", + "bbox": [ + 496, + 582, + 890, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some 3DMMs [17, 14] alleviate this issue by using deformation bases that are hand-crafted by artists, often referred to as blendshapes. While this alleviates the interpretability issue, it loses the data-driven and explicit statistical prior, requires manual work, and does not guarantee the expressiveness of the resulting model.", + "bbox": [ + 496, + 750, + 890, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First to address the semantic issue of PCA-based 3DMMs in the human body domain were Seo et al. [33] which used metrics such as hip-to-waist ratio, body fat percentage, and height to regress 3DMM shape parame", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14555", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/66f4710b2aea4be682757d90d68a3adcd51df36fb3a70871158233b62a01ae87.jpg", + "image_caption": [ + "Figure 2: Learning a mapping from Semantic to Parametric space. (a) Given a coefficient vector $\\vec{\\xi}$ we create the 3DMM mesh. The mesh is rendered from several views. Each Rendered image $I'$ is passed into CLIP along with a set of semantic descriptors $\\vec{\\ell}$ . The difference between each descriptor and the image in CLIP latent space is calculated and stored in the corresponding entry of the similarity vector $\\vec{\\Omega}$ . (b) Using a large set of such random pairs of $(\\vec{\\Omega},\\vec{\\xi})$ , we train a network to learn the mapping from semantic space to parametric space." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 890, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ters. Hill et al.[15] suggested examining the relationship between body shapes and description, by linking two similarity spaces - one created from body descriptions and the other from full-body laser scans. BodyTalk [34] followed suit and approached this problem using the \"wisdom of the crowd\". In their work, they captured 256 male and female body shapes and represented each shape with 30 body descriptor words (e.g. curvy, fit, heavyet, round-Apple, etc.). They hired labelers to discretely rate the bodies with respect to these descriptors and designed a regression model to learn the relationship between the discrete ratings and the parameters of the corresponding body shape. In a more recent publication, SHAPY [7] bridges this gap by curating datasets of images with corresponding body measurements (e.g., from model agencies) followed by human ratings respective to a set of linguistic shape attributes. They train models to predict SMPLX shape parameters from attributes and/or body measurements, and vice-versa.", + "bbox": [ + 75, + 351, + 472, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Language-Guided 3D Modeling. Recent works have shown great success in learning latent representations that are capable of coupling visual signals with language. One such example is CLIP [32] model. CLIP uses a contrastive training scheme [6] on over 400 million pairs of images and their captions to build a shared latent representation for visual-textual content. In this space, similarity between images and text can be computed. It has been shown that CLIP is an effective tool for image generation tasks. For example, providing natural-language guided manipulation of human face imagery in StyleGAN space [30, 26, 1]. Similarly, CLIP can be employed in 3D: using differentiable rendering, images of a 3D scene are rendered in a forward pass and are scored against a text prompt with CLIP. Taking the similarity score as an optimization objective, gradients are back-propagated through the CLIP network and the rendering process, back to the underlying representation of the 3D scene. Text2Mesh [27] uses CLIP to optimize for col", + "bbox": [ + 75, + 628, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ors and positions of mesh vertices to match a text query. CLIP-Mesh [20] followed suit and proposed to use CLIP to guide normal and texture maps. This basic recipe has been applied to a variety of scene representations, such as point clouds [28] and Neural Fields [38, 18, 39, 13]. In the context of human faces and bodies, CLIP is used for generation of animation sequences [21, 36, 16, 35, 10, 2].", + "bbox": [ + 496, + 351, + 893, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Zero-Shot 3DMM Shape Reconstruction. Recovering accurate, explicit meshes from 2D signals such as images or video is an under-constrained and over-parameterized objective. The low-dimensional underlying representation of 3DMMs proves useful not only for modeling, but for 3D reconstruction of geometry from images of humans. The parameters can serve as a data-driven regularization term for shape and pose, and the low dimensional representation convexifies the optimization problem of fitting such a model to an image, rather than a freely deforming mesh. Recent 3DMM-from-image methods either use iterative optimization schemes to fit the parameters of the 3DMM to the image [23] or directly regress the shape and pose parameters from an Image [12, 8, 22, 19]. Our method allows to use CLIP image and text mapping and simply feed-forward the scores through an MLP network to get a body shape.", + "bbox": [ + 496, + 464, + 893, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 722, + 591, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overview of our method can be seen in Figure 2. In essence, our method defines a non-linear mapping from semantic space to parametric space by training a neural network to predict 3DMM coefficients from a vector of semantic attribute scores. We first create a dataset of rendered images of randomly sampled shapes (3.1), gather their corresponding semantic scores (3.2). Then, we employ a scheme to reduce the semantic descriptors to a subset of least-correlated descriptors (3.3) and then train the network (3.4).", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14556", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/0434f26d053aff10504472a4a25123f0221de16f6d76b69550466fb7eace7c32.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Coverage / Overlap
MethodAlgorithm Choice2 descriptors5 descriptors10 descriptors15 descriptors
SMPLX-male(6) / 92.1% / 64.5%38.6% / 30.1%51.5% / 71.4%49.7% / 83.5%98.3% / 80.5%
SMPLX-female(4) / 49.7% / 56.7%27.1% / 49.5%50.0% / 67.1%47.4% / 83.0%82.4% / 83.4%
SMPLX-neutral(8) / 52.0% / 83.0%40.2% / 51.2%50.0% / 67.1%51.9% / 86.9%52.1% / 90.3%
SMPL-male(4) / 97.5% / 60.2%40.1% / 46.7%97.7% / 66.5%98.3% / 83.9%98.2% / 89.3%
SMPL-female(5) / 95.7% / 68.6%36.3% / 47.3%95.7% / 68.6%97.8% / 85.2%99.1% / 89.9%
SMPL-neutral(6) / 96.1% / 76.0%80.4% / 48.7%96.2% / 71.5%99.6% / 85.2%99.8% / 90.8%
FLAME-expression(3) / 13.4% / 56.8%10.4% / 47.4%19.3% / 67.2%27.0% / 77.3%34.4% / 85.0%
", + "bbox": [ + 101, + 88, + 870, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1: Choosing different number of descriptors for a model. We evaluate the coverage and overlap of vertices for a mapper that was trained by using different number of descriptors (see Section 5.1, the threshold used to measure coverage is 0.3). The number chosen by our algorithm is shown in parentheses at the beginning of each row. As can be seen, the more descriptors are chosen the larger the cover is, but the larger the overlap between the descriptors is as well.", + "bbox": [ + 75, + 257, + 893, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Dataset Creation", + "text_level": 1, + "bbox": [ + 76, + 344, + 243, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The process of creating the data for training our model is similar across all 3DMMs (we demonstrate it on four different ones in this paper). For each model, we only use the first 10 principal components, that cover above $95\\%$ of the variance. We draw $N_{\\text{samples}}$ 10-dimensional random vectors of coefficients. For body models, we draw a random shape coefficients vector $\\vec{\\beta}$ , which yields a random shape. From the expression basis of the FLAME model, we randomly sample an expression coefficient vector $\\vec{\\psi}$ which yields a random expression. Since random values may also lead to noisy and unrealistic outputs, we limit the values of the sampled coefficients to a range of $k$ standard deviations of the model's coefficients ( $k = 2$ for body shapes and $k = 4$ for face expressions), yielding more realistic results. The other parameters of the 3DMM $(\\theta, \\gamma)$ remain neutral.", + "bbox": [ + 75, + 368, + 470, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Defining the semantic representation begins by gathering an over-complete set $\\ell$ of $N_{\\ell}$ word descriptors that correspond to each 3DMM. Our method supports any set of words that relate to the model: we couple the body model with body descriptors, and choose face descriptors for the face model. For the body models SMPL and SMPL-X, we adopt the set of descriptors used in [34], for FLAME face expression we use an expanded set based on [9], and for FLAME face shape we used our own set of descriptors. Lastly, since the SMAL model was trained on 5 animal families (Felidae-big cats, Canidae-dogs, Equidae-horses, Bovidae-cows, Hippopotamidae-hippos) we used a set of descriptors that contains animals from these given families. The full list of descriptors can be found in the supplemental file. We note that our method can be used with any desired set of relevant descriptors.", + "bbox": [ + 75, + 595, + 470, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. CLIP Ratings", + "text_level": 1, + "bbox": [ + 76, + 845, + 220, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use CLIP [32] to encode each text descriptor $\\ell_{i}$ using their text encoder $e^{\\ell_i} = CLIP_{Text}(\\ell_i)$ . Each mesh created", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from a random coefficient vector $F(\\vec{\\beta}, \\theta, \\gamma)$ is rendered to create a set of images $I_{j}$ , and these images $I_{j}$ are encoded by CLIP image encoder $e^{I_j} = CLIP_{Image}(I_j)$ to the same latent space of the encoded text (see Figure 2). In this space, we can compute the compatibility between the encoded image $e^{I_j}$ and each encoded label $e^{\\ell_i}$ using cosine similarity to get a score $\\Omega(I_j, \\ell_i) = \\cos(e^{I_j}, e^{\\ell_i})$ . This way, each random coefficients vector $\\vec{\\beta}$ or $\\vec{\\psi}$ is paired with a vector of the descriptors' scores:", + "bbox": [ + 496, + 344, + 893, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega_ {i = 1} ^ {N _ {\\ell}} = [ \\Omega (I _ {j}, \\ell_ {1}), \\dots , \\Omega (I _ {j}, \\ell_ {N _ {\\ell}}) ] = [ \\Omega_ {1}, \\dots , \\Omega_ {N _ {\\ell}} ]\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 492, + 867, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "containing $N_{\\ell}$ scores.", + "bbox": [ + 500, + 521, + 645, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Descriptors Selection", + "text_level": 1, + "bbox": [ + 500, + 545, + 697, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There are many words that could describe a body shape, face shape or facial expression. As the score of a given image and word descriptor in CLIP's embedding space is calculated by their semantic similarity, there might be many possible word descriptors for a single image, and many possible images for a single word descriptor. For example, both the words \"happy\" and \"smile\" would have a high correlation with a smiling face, and therefore, their effect on a facial expression is entangled. Similarly, \"tall\" and \"short\" are entangled in a body shape model. Our key observation is that a large number of descriptors are difficult to handle for interactive modeling, and the larger the number, the more entangled they are in terms of their effect on the shape - making it even harder to achieve the desired shape results (see Table 1).", + "bbox": [ + 496, + 568, + 890, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contrary to BodyTalk [34], our goal is to use a minimal set of semantic word descriptors, while covering as much of the PCA shape space of each 3DMM as possible. We present an algorithm for selecting a subset of semantic descriptors with two main objectives: they should cover a large part of the shape space, and they should be disentangled as much as possible.", + "bbox": [ + 496, + 795, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14557", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our idea is to first segment the shape space into regions, and assign descriptors to each region. This segmentation assures that we cover the shape space as we do not neglect any region. Next, we select for each region a set of descriptors by checking their variance (i.e., promoting coverage), and removing correlated descriptors, leaving the remaining ones as disentangled as possible.", + "bbox": [ + 76, + 90, + 468, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Clustering. To segment the shape space, we use our $N_{\\text{samples}}$ dataset as a proxy to represent a sampling of the space and cluster their images (created as described in Section 3.1) encoded to CLIP's latent space. We use K-means algorithm with Silhouette score to find $K$ .", + "bbox": [ + 76, + 198, + 468, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we compare all encoded images $e^{I}$ in each cluster $C_{i}$ to all $N_{\\ell}$ encoded word descriptors $e^{\\ell_i}$ for $i = 1,\\dots ,N_{\\ell}$ . The top 5 most similar descriptors in a cluster are given a vote for this cluster. We normalize the value of votes by the size of each cluster. Then, each word descriptor is assigned exclusively to the cluster with the largest sum of votes for this descriptor.", + "bbox": [ + 76, + 275, + 468, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Choosing descriptors. To anticipate the effect of a given descriptor on the 3DMM, we use the variance of its score $\\Omega(I', \\ell_i)$ on all images $I'$ in the dataset. The larger the variance, the more descriptive this descriptor will be in terms of shape variation. To anticipate the level of entanglement between descriptors, we use their correlation. For each descriptor $\\ell_i$ we build a vector of $N_{\\text{samples}}$ scores $\\Omega(I', \\ell_i)$ for all images $I'$ in the dataset. Correlation is measured between these vectors. The larger the correlation between two descriptors, the more similar the effect of this descriptor is on the shape, and the more entangled they are (we show some correlation plot examples in the supplemental file). Lastly, we apply antonyms and synonyms detection to verify there are no such pairs in our final set of descriptors.", + "bbox": [ + 76, + 383, + 468, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The process of choosing the descriptors is first performed on each cluster separately, and then merging the lists. We sort the cluster descriptors according to their variance in descending order. Thus, the first descriptor that is chosen from each cluster is the descriptor with the highest variance. Once a descriptor is chosen, we iterate over the other descriptors that are left for this cluster and filter out the ones that have a high correlation with the chosen descriptor (we use the median correlation value of all pairs as the threshold). In addition, before adding the next descriptor to the set, we check if its synonyms or antonyms are already chosen, and if so, we skip it. This process continues until the list of descriptors is exhausted.", + "bbox": [ + 76, + 597, + 468, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lastly, we merge the lists created from all clusters and sort again according to the descriptor variance. In a similar manner as we did for each cluster, we filter from the merged list all correlated descriptors, synonyms and antonyms. Finally, we arrive at a set of $d$ chosen semantic descriptors for this model. $d$ can be different for each model, but our algorithm can also support any preset number of descrip", + "bbox": [ + 76, + 795, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2275e4d54f8c2c0a0b7a7c9673ad5499299f332252d0bf8304e84c3b08c3fc34.jpg", + "image_caption": [ + "Figure 3: Models created using an interactive slider application for SMAL[45] model (the blue model on the left is the neutral model). These animals were created by interpolating between four semantic descriptors - lion, cow, cat and dog. As can be seen, other animals could also be designed as well as intermixed creatures." + ], + "image_footnote": [], + "bbox": [ + 506, + 87, + 901, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tors by removing descriptors from the bottom of the list or adding back descriptors that were filtered out according to their variance order (see Table 1).", + "bbox": [ + 496, + 431, + 890, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our algorithm can also support a preset list of descriptors if the user seeks to fit a 3D model with respect to some specific descriptors. In this case, the process described above for choosing the final set of descriptors is simply initialized with the user pre-defined descriptors, so that they are always contained in the final set.", + "bbox": [ + 496, + 481, + 890, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training", + "text_level": 1, + "bbox": [ + 500, + 594, + 602, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For simplicity, we will denote the coefficients vector by $\\vec{\\xi}$ for both shape and expression parameters. Our goal is to define the mapping from the semantic representation of the $d$ descriptors to the parametric representation represented by $\\vec{\\xi}$ . From our dataset creation (Section 3.1), each coefficients vector $\\vec{\\xi} = [\\xi_1,\\dots,\\xi_{10}]$ is paired with a scores vector $\\vec{\\Omega} = [\\Omega_1,\\dots,\\Omega_{N_\\ell}]$ . Hence, for a given scores vector $\\vec{\\Omega}$ , the goal is to predict the corresponding coefficients vector $\\vec{\\xi}$ . This mapping resembles the one presented in [34], only we rely on CLIP score to label our data, providing a single score rather than multiple scores obtained by crowd sourcing, and we do not assume a linear relationship between the word descriptors and the 3DMM's coefficients vector. Instead, we train a multi-layer-perceptron with ReLU activations as the mapping function using our paired data and a simple $L_{2}$ loss $\\mathcal{L} = ||\\hat{\\vec{\\xi}} -\\vec{\\xi} ||_2$ . The network consists of hidden layers (500 and 800 neurons), and is trained for 50 epochs.", + "bbox": [ + 496, + 622, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14558", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Applications", + "text_level": 1, + "bbox": [ + 76, + 89, + 207, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Interactive Sliders", + "text_level": 1, + "bbox": [ + 76, + 114, + 253, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using the set of semantic descriptors $\\ell$ chosen by our algorithm, we can build a simple and intuitive interface to control the 3DMM. This is because the descriptors are least entangled and cover the PCA space well. First, we train a mapper $M$ with the chosen set of descriptors $\\ell$ as described above (Section 3.4). Next, we create a slider for each descriptor whose value represents the desired CLIP score for that specific descriptor (see Figure 4). The user can now change each slider to trigger a re-prediction of the 3DMM coefficients. The computation is interactive as the interface takes all sliders' values as input $\\vec{\\Omega}$ to $M$ , which maps them in a forward pass to the 3DMM coefficients vector $\\vec{\\xi}$ and renders the corresponding 3D mesh. Thus, the modeler can interact and understand the relationships between the descriptors and their effect on the mesh. Figure 3 demonstrates examples of 3D meshes of different animals that were created using an interface created for the SMAL model.", + "bbox": [ + 75, + 137, + 472, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Zero-Shot Image to Shape Reconstruction", + "text_level": 1, + "bbox": [ + 76, + 419, + 434, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method allows to leverage CLIP's semantic understanding to define a zero-shot image to shape reconstruction. Given any image of a person, we embed it into CLIP's latent space and use the similarity scores of the image against the set of descriptors $\\ell^{M}$ chosen for a mapper $M$ to obtain the vector $\\vec{\\Omega}$ . Next, we simply feed $\\vec{\\Omega}$ through the network to get the 3DMM coefficients vector $\\vec{\\xi}$ that best fits the image and create the shape from $\\vec{\\xi}$ . Our method differs from state-of-the-art methods in that it enables the user to refine the zero-shot prediction very simply using semantic sliders starting from a relatively good initial guess.", + "bbox": [ + 75, + 441, + 468, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 622, + 209, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Coverage Evaluation", + "text_level": 1, + "bbox": [ + 76, + 647, + 272, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate a mapper $M$ that was trained with a given set of descriptors $\\ell$ , we would like to measure its effect on the 3DMM. For each descriptor $\\ell_i \\in \\ell$ we do a forward pass through the mapper twice: the first pass by setting a low score to $\\Omega_i$ and the second pass by setting a high score to $\\Omega_i$ . The other coefficients are set to a default value and remain constant. Each such pass creates a parameter vector $\\vec{\\xi} = [\\xi_1, \\dots, \\xi_{10}]$ , which represents the resulting 3DMM shape or expression. We denote these vectors as $\\vec{\\xi}(i)_{low}$ and $\\vec{\\xi}(i)_{high}$ .", + "bbox": [ + 75, + 671, + 468, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We measure the geometric effect of a descriptor $\\ell_{i}$ on a 3DMM by examining the deformation of the vertices of the mesh. This is evaluated by comparing the position of each vertex in the two extreme cases of $\\vec{\\xi} (i)_{low}$ and $\\vec{\\xi} (i)_{high}$ . Hence, for each vertex $v$ the size of deformation is de", + "bbox": [ + 75, + 825, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ba2e8233a3463231635a82b4dd5861c7c0bad839b00885c2da23b673b0f54061.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 87, + 624, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b446a98b47353fb1fb61a099791f9d52189f37b7ae78b972d87dc3e14dc131f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 174, + 620, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0aa37f85af56025902920a1d234eb5600e573a056737d354461c241c7dafc3e6.jpg", + "image_caption": [ + "Figure 4: A simple and intuitive interactive application can be defined using sliders for each semantic descriptor. With Semantiy, the sliders have strong semantic meaning and are disentangled so that changing one will have very limited affect on others." + ], + "image_footnote": [], + "bbox": [ + 500, + 237, + 620, + 316 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b810a44ef572a968a50973da2d71ea70decc818848d9e8d0a61f80ea36d12521.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 117, + 893, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "fined as $\\delta'(v) = ||v_{low} - v_{high}||_2$ . We can normalize these values by $\\delta_{\\max} = \\max_{v \\in Mesh} \\delta'(v)$ and get a value $\\delta(v) = \\delta'(v) / \\delta_{\\max}$ between 0 and 1 (see Figure 5).", + "bbox": [ + 496, + 441, + 890, + 488 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2b25c53b186ea56238c5157ab4294f6b64e2ccfbab03d23e9831276848c7345a.jpg", + "image_caption": [ + "Figure 5: The geometric effect of a given descriptor. The darker the color of the vertex is (more red), the more it is affected by the descriptor. To define coverage we collect all vertices with a value larger than a given threshold $\\theta$ ." + ], + "image_footnote": [], + "bbox": [ + 509, + 505, + 774, + 686 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b86ca843a083907af627e79beb5fed8dc8b91d41a4d60b9eb2b0ad6a0f9137e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 551, + 895, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using these differences, we can evaluate the coverage of a specific descriptor by measuring the number and positions of vertices it affects in the mesh. For example, we can look at all vertices $v$ whose value is above a threshold $\\delta(v) > \\theta$ for any given descriptor. We can also compare the overlap between two descriptors by measuring the intersection-over-union of their covered vertices. We use these measure", + "bbox": [ + 496, + 794, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14559", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in our ablation studies as described next in Section 5.2. Note that this measure is only an approximation as it measures the size of the change of vertices and not their direction.", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Ablations", + "text_level": 1, + "bbox": [ + 76, + 147, + 187, + 161 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further evaluate our algorithm's performance, we conducted experiments that include training the mapper with different numbers of descriptors and training on different sizes of training data. As we do not have valid ground truth to estimate the accuracy of our mapper we created a test set that contains several body shapes that were generated by the first 10 blendshapes $\\vec{\\xi}$ of the SMPL-X model. In each experiment we optimize the descriptors' values $\\vec{\\Omega}$ , to minimize the error $||\\hat{\\vec{\\xi}} - \\vec{\\xi}||_2$ . The goal of these experiments is to test the expressive power of the given configuration to match a given shape. The smaller the error the more expressive the method is. We show the results of these ablations in the supplemental file. We used this evaluation to choose the hyper-parameters we use in our algorithm: training with 3K sample images with texture for all models.", + "bbox": [ + 75, + 170, + 472, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Interactive Sliders User Study", + "text_level": 1, + "bbox": [ + 76, + 422, + 344, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate Semantify's performance we conducted a user study that compares our approach against BodyTalk [34], as well as the original basis of the 3DMM. The user is presented with a target 3D model (randomly created using the 3DMM) and is tasked to fit a 3DMM mesh model to the target shape using a set of sliders, using the two compared interfaces (see Figure 4) at random order. Users were given up to 5 minutes to complete the task. BodyTalk was implemented on the SMPL body model, therefore we compare it to our SMPL male and female models. In another study, we compare our SMPLX male and female models to the original PCA-based axes. In both studies, we seek to evaluate the accuracy of the user-fitted 3DMMs, the time that it took to fit the model, as well as the overall subjective experience of users. Since we cannot obtain the outputs of BodyTalk in the first study, we asked humans to rate the results, and determine which of them fits more accurately to the input shape. We had 10 users perform both studies (5 males, 5 females aged between 23 to 54), 7 of which are novices and 3 of which were professional 3D modelers. We present the quantitative results of these studies in Table 2. As can be seen, Semantify achieves better accuracy in less time. Note that theoretically only the baseline can reach zero error, but because it is very difficult to handle, the error obtained was higher than Semantify. In addition, we asked the users for feedback on the experience of using the different applications (A was ours, B was the alternative). Here are some examples of such reviews (more examples could be found in the supplemental file):", + "bbox": [ + 75, + 446, + 472, + 885 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "\"Using application B, occasionally when changing one", + "bbox": [ + 96, + 885, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "slider it affects the other, which causes the user to start all over again\".", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "\"Using application B every minor change in a certain slider generated major changes in the rest of the sliders\"", + "bbox": [ + 498, + 121, + 890, + 152 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "\"Application A was far friendlier and I sensed as though I maintained much more control over the different body features\"", + "bbox": [ + 498, + 152, + 890, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "\"The abundance of sliders on application B only made it harder to control, not the other way around\".", + "bbox": [ + 498, + 198, + 890, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Zero Shot Reconstruction", + "text_level": 1, + "bbox": [ + 498, + 239, + 733, + 255 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To measure our zero-shot 3D-shape reconstruction method of SMPLX model, we use HBW (Human Bodies in-the-wild) dataset presented in SHAPY [7], containing ground-truth 3D body scans with in-the-wild images. Table 3 shows the performance of our method compared to SHAPY and PixIE [12] on all of HBW validation set. Note that SHAPY also takes keypoints (usually obtained from openPose [5]) as input. Some examples are shown in Figure 6. As can be seen, our zero-shot performance are on-par with state of the art methods, but we have the advantage of using semantic descriptors. This allows the user to fine tune the results from a very good initial guess, simply by using the semantic sliders. Some examples results with around a minute of fine-tuning can be found in the supplemental file.", + "bbox": [ + 496, + 263, + 890, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 491, + 617, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We have presented a method to edit a 3D parametric model using semantically meaningful descriptors by first choosing a subset of descriptors judiciously and then learning a non-linear mapping from this set to the 3DMM coefficients. All this is done in a self-supervised manner harnessing the abilities of CLIP visual-language model. Our method can support a more intuitive user interface that allows even novices to model 3D shapes. It also supports other applications such as zero-shot reconstruction and fine-tuning of shapes from images.", + "bbox": [ + 496, + 516, + 890, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Limitations and future directions. There are several limitations to our method. First, our mapper's performance relies heavily on the data that is created (Section 3.1). For example, a dataset must contain extreme samples to achieve a more expressive mapper. Second, our goal with Semantify was to build a semantic representation of a given 3DMM without a human-in-the-loop. However, in some cases manually tuned modeling for a specific 3DMM would probably improve the 3DMM mapper's performance (we further elaborate on this in the supplemental file). Lastly, as shown in many other works, CLIP has rich semantic understanding, but there are cases in which its performance degrades. For example, performance is better with textured meshes, so using various textures for each 3DMM may improve the", + "bbox": [ + 496, + 689, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14560", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/496f0ccadf6c7582fd65d2e6abfef15bd4e73c4e6671a91bfe46eda410bcd999.jpg", + "image_caption": [ + "Figure 6: Results of zero-shot image to shape reconstruction. We compare Semantify against SHAPY and PixIE on SMPLX 3DMM. Colors indicate error compared to ground truth (darker means larger error)." + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 869, + 448 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5f18fb9bfaacce1607bc68dc382d5cf2003119dc20f9a9f482a5543ae927e176.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Time (min)Score
MalesFemalesTotal
BodyTalk3.223.163.190.04%
ours2.121.451.590.96%
", + "bbox": [ + 86, + 510, + 426, + 595 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d57469d8d99552b6e4c8880640eac051b2f6a5f0f4bb79cfc9315fa6f3278414.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Mean Error (cm)Time (min)
MalesFemalesTotalMalesFemalesTotal
Baseline0.1220.1110.1163.182.092.43
ours0.1210.0990.1102.222.172.19
", + "bbox": [ + 452, + 503, + 908, + 588 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1ff0efca4da9766220533610ecfe6e2da76c1410c7c8964b37173ae2b9c98683.jpg", + "table_caption": [ + "Table 2: User-study results for the SMPL 3DMM compared against BodyTalk [34] (left, Score means the percent that this model was selected as better fitting the input), and SMPL-X parametric model compared against the raw blendshapes (right)." + ], + "table_footnote": [], + "table_body": "
MalesFemalesTotal
SHAPY0.01969340.0144220.017085
PIXIE0.0247790.0152650.020324
Semantify (ours)0.0226320.0165640.019666
", + "bbox": [ + 81, + 667, + 467, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3: Comparing our zero-shot shape reconstruction from image method against SHAPY[7] and PixIE [12]. The values demonstrate the average MSE over HBW's in-the-wild validation set.", + "bbox": [ + 75, + 768, + 468, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mapper's performance. Furthermore, more subjective descriptors such as \"cute\", \"scary\", \"pretty\" etc., are harder to", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rate using CLIP, and would probably not work well in our model.", + "bbox": [ + 498, + 669, + 890, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Future work can also improve the zero-shot 3D shape reconstruction by using prior information about its context (for example, the gender of a person, body keypoints as used in [7] etc.). Adding such priors along with CLIP's semantic understanding may surpass the existing state-of-the-art solutions.", + "bbox": [ + 496, + 704, + 893, + 796 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 824, + 679, + 842 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was supported by the Israel Science Foundation (grant No. 1390/19). We would also like to thank Eyal Gomel for valuable discussions.", + "bbox": [ + 498, + 854, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14561", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdal, Peihao Zhu, John Femiani, Niloy J. Mitra, and Peter Wonka. Clip2stylegan: Unsupervised extraction of stylegan edit directions. CoRR, abs/2112.05219, 2021.", + "[2] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Clipface: Text-guided editing of textured 3d morphable models. arXiv preprint arXiv:2212.01406, 2022.", + "[3] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. Scape: Shape completion and animation of people. ACM Transactions on Graphics (TOG), 24:408-416, 07 2005.", + "[4] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999.", + "[5] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43:172–186, 2018.", + "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations, 2020.", + "[7] Vasileios Choutas, Lea Müller, Chun-Hao P. Huang, Siyu Tang, Dimitrios Tzionas, and Michael J. Black. Accurate 3d body shape regression using metric and semantic attributes. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[8] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Monocular expressive body regression through body-driven attention. In European Conference on Computer Vision (ECCV), 2020.", + "[9] Radek Daněček, Michael J. Black, and Timo Bolkart. Emotion driven monocular face capture and animation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20311-20322, June 2022.", + "[10] Ginger Delmas, Philippe Weinzaepfel, Thomas Lucas, Francesc Moreno-Noguer, and Grégory Rogez. Postscript: 3d human poses from natural language. In European Conference on Computer Vision, pages 346-362. Springer, 2022.", + "[11] Bernhard Egger, William AP Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski," + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Sami Romdhani, et al. 3d morphable face models—past, present, and future. ACM Transactions on Graphics (TOG), 39(5):1-38, 2020.", + "[12] Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael Black. Collaborative regression of expressive bodies using moderation. In International Conference on 3D Vision (3DV), pages 792-804, Dec. 2021.", + "[13] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8649-8658, June 2021.", + "[14] Thomas Gereg, Andreas Morel-Forster, Clemens Blumer, Bernhard Egger, Marcel Luthi, Sandro Schoenborn, and Thomas Vetter. Morphable face models - an open framework. 2018 13th IEEE International Conference on Automatic Face amp; Gesture Recognition (FG 2018), May 2018.", + "[15] Matthew Hill, Stephan Streuber, Carina Hahn, Michael Black, and Alice O'Toole. Exploring the relationship between body shapes and descriptions by linking similarity spaces. Journal of Vision, 15(12):931-931, 2015.", + "[16] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhonggang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022.", + "[17] IEEE. A 3D Face Model for Pose and Illumination Invariant Face Recognition, Genova, Italy, 2009.", + "[18] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, P. Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 857-866, 2021.", + "[19] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), 2018.", + "[20] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. SIGGRAPH Asia 2022 Conference Papers, 2022.", + "[21] Youwang Kim, Ji-Yeon Kim, and Tae-Hyun Oh. Clip-actor: Text-driven recommendation and stylization for animating human meshes. In European Conference on Computer Vision, 2022.", + "[22] Muhammed Kocabas, Chun-Hao P. Huang, Joachim Tesch, Lea Müller, Otmar Hilliges, and Michael J." + ], + "bbox": [ + 501, + 90, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14562", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Black. SPEC: Seeing people in the wild with an estimated camera. In Proc. International Conference on Computer Vision (ICCV), pages 11035-11045, Oct. 2021.", + "[23] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, 2019.", + "[24] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017.", + "[25] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. *Smpl: A skinned multi-person linear model*. ACM Trans. Graph., 34(6), nov 2015.", + "[26] Xudong Lou, Yiguang Liu, and Xuwei Li. Tecm-clip: Text-based controllable multi-attribute face image manipulation. In Proceedings of the Asian Conference on Computer Vision, pages 1942–1958, 2022.", + "[27] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022.", + "[28] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. ArXiv, abs/2212.08751, 2022.", + "[29] Ahmed A. A. Osman, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Supr: A sparse unified part-based human representation. In Shai Avidan, Gabriel Brostow, Moustapha Cissé, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision – ECCV 2022, pages 568–585, Cham, 2022. Springer Nature Switzerland.", + "[30] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021.", + "[31] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019.", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark," + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021.", + "[33] Hyewon Seo, Frederic Cordier, and Nadia Thalmann. Synthesizing animatable body models with parameterized shape modifications. 07 2003.", + "[34] Stephan Streuber, M. Alejandra Quiros-Ramirez, Matthew Q. Hill, Carina A. Hahn, Silvia Zuffi, Alice O'Toole, and Michael J. Black. Body Talk: Crowd-shaping realistic 3D avatars with words. ACM Trans. Graph. (Proc. SIGGRAPH), 35(4), July 2016.", + "[35] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022.", + "[36] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022.", + "[37] Ayush Tewari, Michael Zollhöfer, Pablo Garrido, Florian Bernard, Hyeongwoo Kim, Patrick Pérez, and Christian Theobalt. Self-supervised multi-level face model learning for monocular reconstruction at over 250 hz. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018.", + "[38] Can Wang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2022.", + "[39] Can Wang, Ruixia Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. ArXiv, abs/2212.08070, 2022.", + "[40] Jinbao Wang, Shujie Tan, Xiantong Zhen, Shuo Xu, Feng Zheng, Zhenyu He, and Ling Shao. Deep 3d human pose estimation: A review. Computer Vision and Image Understanding, 210:103225, 2021.", + "[41] Lizhen Wang, Zhiyua Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR2022), June 2022.", + "[42] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Ghum & ghuml: Generative 3d human shape and articulated pose models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6184-6193, 2020." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14563", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. *Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction.* In *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition* (CVPR), 2020.", + "[44] Silvia Zuffi, Angjoo Kanazawa, and Michael J. Black. Lions and tigers and bears: Capturing non-rigid, 3d, articulated shape from images. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3955-3963, 2018.", + "[45] Silvia Zuffi, Angjoo Kanazawa, David Jacobs, and Michael J. Black. 3D menagerie: Modeling the 3D shape and pose of animals. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), July 2017." + ], + "bbox": [ + 78, + 90, + 468, + 339 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14564", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Semantify_ Simplifying the Control of 3D Morphable Models Using CLIP/ff2d3341-5fbc-47bb-9ce8-0e16c61108bc_model.json b/2023/Semantify_ Simplifying the Control of 3D Morphable Models Using CLIP/ff2d3341-5fbc-47bb-9ce8-0e16c61108bc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..285bfe2f055ee89a1ba35ee79222585f05f194cf --- /dev/null +++ b/2023/Semantify_ Simplifying the Control of 3D Morphable Models Using CLIP/ff2d3341-5fbc-47bb-9ce8-0e16c61108bc_model.json @@ -0,0 +1,1861 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.116, + 0.131, + 0.856, + 0.155 + ], + "angle": 0, + "content": "Semantify: Simplifying the Control of 3D Morphable Models using CLIP" + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.179, + 0.768, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.461, + 0.896, + 0.537 + ], + "angle": 0, + "content": "Figure 1: Semantify offers a method to create and edit a 3D parametric model using semantically meaningful descriptors. Semantify is based on a self-supervised method that utilizes the semantic power of CLIP language-vision model to build a mapping between semantic descriptors to 3DMM model coefficients. This can be used in an interactive application defining a slider for each descriptor (a), or to fit a model to an image in a zero shot manner by feeding the image into CLIP and obtaining a vector of semantic scores that can be mapped to shape parameters (b)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.55, + 0.314, + 0.565 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.475, + 0.885 + ], + "angle": 0, + "content": "We present Semantify: a self-supervised method that utilizes the semantic power of CLIP language-vision foundation model [32] to simplify the control of 3D morphable models. Given a parametric model, training data is created by randomly sampling the model's parameters, creating various shapes and rendering them. The similarity between the output images and a set of word descriptors is calculated in CLIP's latent space. Our key idea is first to choose a small set of semantically meaningful and disentangled descriptors that characterize the 3DMM, and then learn a non-linear mapping from scores across this set to the parametric coefficients of the given 3DMM. The nonlinear mapping is defined by training a neural network without a human-in-the-loop. We present results on numerous 3DMMs: body shape models, face shape and expression models, as well as animal shapes. We demonstrate how our method defines a simple slider interface for intuitive modeling, and show how the mapping can be used to instantly fit a 3D parametric body shape to in-the-wild images. See our project page at https://omergral.github.io/Semantify/" + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.55, + 0.633, + 0.565 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.895, + 0.837 + ], + "angle": 0, + "content": "3D modeling techniques have evolved tremendously over the last few years. Such techniques are used in countless industries including the burgeoning AR/VR industry, fashion design, game development, film, and many others. However, designing a high-quality 3D model is not a simple task for most people and might require a well-trained 3D artist. Even when using more recent parametric morphable models (3DMM) [24, 31, 45, 3, 42] it is still hard for humans to understand how to choose the correct set of parameters, e.g. to achieve a specific human body shape or facial expression. It is also difficult to find what are the limits of the given parametric model in terms of coverage and expressiveness. The reason is that in most cases, the provided set of parameters is not interpretable, as they are commonly calculated using automatic optimization mechanisms followed by dimensionality reduction using PCA. Thus, they carry no clear semantic meaning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To simplify the use of 3DMMs and allow for natural interactive human modeling, a key research question is - how to insert semantics into 3DMM control? Previous approaches [34] relied mostly on human intelligence and la" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14554" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "beling, which is often time consuming and expensive. In addition, previous methods create a large number of semantic control descriptors that are often correlated and entangled. This means it is difficult to anticipate the effect of each descriptor as changing a value in one may modify others, resulting in difficulties to control the model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.185, + 0.471, + 0.412 + ], + "angle": 0, + "content": "In this paper, instead of using human labeling, we rely on the remarkable abilities of huge foundation models that combine natural language and visual understanding. We present a self-supervised method that utilizes the semantic power of CLIP [32] to define a method to control 3DMMs that carries two main advantages. First, it is more natural for humans to use as it allows modeling using a small number of semantically meaningful descriptors, that cover the space of deformations but are disentangled. Second, it covers even extreme examples in the shape/pose space of parametric models as it utilizes CLIP's pre-training on a large number of images. The main idea is first to use CLIP to select a small subset of semantically meaningful and disentangled descriptors, and then learn a non-linear mapping of this set to the coefficients of the given 3DMM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.471, + 0.67 + ], + "angle": 0, + "content": "Given a parametric 3DMM, we sample its parameter space to create a dataset containing a variety of 3D mesh shapes. We then render each mesh from different camera views to create a diverse set of images corresponding to the parameter samples (see Figure 2). Next, we gather a set of semantically descriptive textual terms related to the parametric 3D model, which we call descriptors. We encode both the images (using CLIP's image encoder) and the descriptors (using CLIP's text encoder) into CLIP's latent space and compare them. This defines a vector of similarity scores between the input vector of 3DMM coefficients (of each image sample) and each corresponding semantic descriptor. Next, we define a selection scheme to choose a small number of descriptors that are de-correlated to control the model. Lastly, we train a neural network to learn a mapping between the vector of similarity scores to the vector of 3DMM parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.471, + 0.853 + ], + "angle": 0, + "content": "We demonstrate how the learned mapping can be used to define an interface to control a 3DMM in a way that is simple and effective using a small set of semantically meaningful sliders (see Figure 1 and 4). Such sliders are easy to employ for designing high-quality 3D models and cover the shape space well. We demonstrate this for four parametric models: human face's shape and expression (FLAME [24]), human body shapes (SMPL [25] and SMPL-X [31]), and even animals (SMAL [45]). We also show how the mapping can be used to instantly fit a 3D parametric body shape to an input image that works well \"in the wild\" even in extreme poses and body shapes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Our main contribution is a novel, self-supervised method for defining a small set of semantic descriptors to control a parametric model more naturally, by learning a mapping" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "from a semantic space to a parametric representation without human-in-the-loop. We demonstrate the effectiveness of our approach on several parametric models and utilize it to define a simple interface for modeling and to instantly fit a 3D model to images in the wild. We will release our code for future research." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.197, + 0.642, + 0.213 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.223, + 0.892, + 0.268 + ], + "angle": 0, + "content": "Our work encompasses a wide range of disciplines, including 3D parametric modeling, language-vision models, and zero-shot reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.892, + 0.41 + ], + "angle": 0, + "content": "3D Morphable Models. Much research has gone into improving 3D modeling of humans bodies, faces and other object classes, such as animals [11, 4, 25, 31]. 3D Morphable Models (3DMM) are a powerful tool to parameterize the variation in geometry of objects belonging to a certain class. Dating back to 1999, Blantz et al. [4] proposed 3DMMs to capture the variation of human faces and applied dimensionality reduction using Principal Component Analysis (PCA)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.892, + 0.562 + ], + "angle": 0, + "content": "In contrast to a 3D artist freely controlling the vertices of a mesh, using each of these PCA axes to deform the vertices induces a strong statistical prior on the resulting geometry. The resulting deformations are global in nature, strongly constraining the geometry to the manifold of statistically plausible meshes, as the per-axis coefficients correlate with their likelihood. While initially proposed for human faces, statistical 3DMMs have been developed for human body models [3, 25, 40, 31, 42, 29], facial models [24, 37, 41, 43], and even 3D animal models [45, 44]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Semantic Models. PCA-based 3DMMs are efficiently and easily constructed, and admit an orthogonal and importance-ordered basis for the deformation. One disadvantage of these models is their lack of interpretability for humans. The axes of maximal variation in geometry, which form the deformation basis for each vertex, do not necessarily correspond to any semantically meaningful geometric change (e.g., raising an eyebrow), but instead, each axis induces a rather global change that may introduce correlations or effects that are unwanted to a human modeler, making it hard to obtain the desired shape or expression." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Some 3DMMs [17, 14] alleviate this issue by using deformation bases that are hand-crafted by artists, often referred to as blendshapes. While this alleviates the interpretability issue, it loses the data-driven and explicit statistical prior, requires manual work, and does not guarantee the expressiveness of the resulting model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "First to address the semantic issue of PCA-based 3DMMs in the human body domain were Seo et al. [33] which used metrics such as hip-to-waist ratio, body fat percentage, and height to regress 3DMM shape parame" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14555" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.891, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.249, + 0.896, + 0.327 + ], + "angle": 0, + "content": "Figure 2: Learning a mapping from Semantic to Parametric space. (a) Given a coefficient vector \\(\\vec{\\xi}\\) we create the 3DMM mesh. The mesh is rendered from several views. Each Rendered image \\(I'\\) is passed into CLIP along with a set of semantic descriptors \\(\\vec{\\ell}\\). The difference between each descriptor and the image in CLIP latent space is calculated and stored in the corresponding entry of the similarity vector \\(\\vec{\\Omega}\\). (b) Using a large set of such random pairs of \\((\\vec{\\Omega},\\vec{\\xi})\\), we train a network to learn the mapping from semantic space to parametric space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.473, + 0.625 + ], + "angle": 0, + "content": "ters. Hill et al.[15] suggested examining the relationship between body shapes and description, by linking two similarity spaces - one created from body descriptions and the other from full-body laser scans. BodyTalk [34] followed suit and approached this problem using the \"wisdom of the crowd\". In their work, they captured 256 male and female body shapes and represented each shape with 30 body descriptor words (e.g. curvy, fit, heavyet, round-Apple, etc.). They hired labelers to discretely rate the bodies with respect to these descriptors and designed a regression model to learn the relationship between the discrete ratings and the parameters of the corresponding body shape. In a more recent publication, SHAPY [7] bridges this gap by curating datasets of images with corresponding body measurements (e.g., from model agencies) followed by human ratings respective to a set of linguistic shape attributes. They train models to predict SMPLX shape parameters from attributes and/or body measurements, and vice-versa." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Language-Guided 3D Modeling. Recent works have shown great success in learning latent representations that are capable of coupling visual signals with language. One such example is CLIP [32] model. CLIP uses a contrastive training scheme [6] on over 400 million pairs of images and their captions to build a shared latent representation for visual-textual content. In this space, similarity between images and text can be computed. It has been shown that CLIP is an effective tool for image generation tasks. For example, providing natural-language guided manipulation of human face imagery in StyleGAN space [30, 26, 1]. Similarly, CLIP can be employed in 3D: using differentiable rendering, images of a 3D scene are rendered in a forward pass and are scored against a text prompt with CLIP. Taking the similarity score as an optimization objective, gradients are back-propagated through the CLIP network and the rendering process, back to the underlying representation of the 3D scene. Text2Mesh [27] uses CLIP to optimize for col" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.895, + 0.459 + ], + "angle": 0, + "content": "ors and positions of mesh vertices to match a text query. CLIP-Mesh [20] followed suit and proposed to use CLIP to guide normal and texture maps. This basic recipe has been applied to a variety of scene representations, such as point clouds [28] and Neural Fields [38, 18, 39, 13]. In the context of human faces and bodies, CLIP is used for generation of animation sequences [21, 36, 16, 35, 10, 2]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.465, + 0.895, + 0.709 + ], + "angle": 0, + "content": "Zero-Shot 3DMM Shape Reconstruction. Recovering accurate, explicit meshes from 2D signals such as images or video is an under-constrained and over-parameterized objective. The low-dimensional underlying representation of 3DMMs proves useful not only for modeling, but for 3D reconstruction of geometry from images of humans. The parameters can serve as a data-driven regularization term for shape and pose, and the low dimensional representation convexifies the optimization problem of fitting such a model to an image, rather than a freely deforming mesh. Recent 3DMM-from-image methods either use iterative optimization schemes to fit the parameters of the 3DMM to the image [23] or directly regress the shape and pose parameters from an Image [12, 8, 22, 19]. Our method allows to use CLIP image and text mapping and simply feed-forward the scores through an MLP network to get a body shape." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.723, + 0.593, + 0.74 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "An overview of our method can be seen in Figure 2. In essence, our method defines a non-linear mapping from semantic space to parametric space by training a neural network to predict 3DMM coefficients from a vector of semantic attribute scores. We first create a dataset of rendered images of randomly sampled shapes (3.1), gather their corresponding semantic scores (3.2). Then, we employ a scheme to reduce the semantic descriptors to a subset of least-correlated descriptors (3.3) and then train the network (3.4)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14556" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.872, + 0.248 + ], + "angle": 0, + "content": "
Coverage / Overlap
MethodAlgorithm Choice2 descriptors5 descriptors10 descriptors15 descriptors
SMPLX-male(6) / 92.1% / 64.5%38.6% / 30.1%51.5% / 71.4%49.7% / 83.5%98.3% / 80.5%
SMPLX-female(4) / 49.7% / 56.7%27.1% / 49.5%50.0% / 67.1%47.4% / 83.0%82.4% / 83.4%
SMPLX-neutral(8) / 52.0% / 83.0%40.2% / 51.2%50.0% / 67.1%51.9% / 86.9%52.1% / 90.3%
SMPL-male(4) / 97.5% / 60.2%40.1% / 46.7%97.7% / 66.5%98.3% / 83.9%98.2% / 89.3%
SMPL-female(5) / 95.7% / 68.6%36.3% / 47.3%95.7% / 68.6%97.8% / 85.2%99.1% / 89.9%
SMPL-neutral(6) / 96.1% / 76.0%80.4% / 48.7%96.2% / 71.5%99.6% / 85.2%99.8% / 90.8%
FLAME-expression(3) / 13.4% / 56.8%10.4% / 47.4%19.3% / 67.2%27.0% / 77.3%34.4% / 85.0%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.258, + 0.895, + 0.32 + ], + "angle": 0, + "content": "Table 1: Choosing different number of descriptors for a model. We evaluate the coverage and overlap of vertices for a mapper that was trained by using different number of descriptors (see Section 5.1, the threshold used to measure coverage is 0.3). The number chosen by our algorithm is shown in parentheses at the beginning of each row. As can be seen, the more descriptors are chosen the larger the cover is, but the larger the overlap between the descriptors is as well." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.345, + 0.245, + 0.36 + ], + "angle": 0, + "content": "3.1. Dataset Creation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.471, + 0.595 + ], + "angle": 0, + "content": "The process of creating the data for training our model is similar across all 3DMMs (we demonstrate it on four different ones in this paper). For each model, we only use the first 10 principal components, that cover above \\(95\\%\\) of the variance. We draw \\(N_{\\text{samples}}\\) 10-dimensional random vectors of coefficients. For body models, we draw a random shape coefficients vector \\(\\vec{\\beta}\\), which yields a random shape. From the expression basis of the FLAME model, we randomly sample an expression coefficient vector \\(\\vec{\\psi}\\) which yields a random expression. Since random values may also lead to noisy and unrealistic outputs, we limit the values of the sampled coefficients to a range of \\(k\\) standard deviations of the model's coefficients (\\(k = 2\\) for body shapes and \\(k = 4\\) for face expressions), yielding more realistic results. The other parameters of the 3DMM \\((\\theta, \\gamma)\\) remain neutral." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.471, + 0.839 + ], + "angle": 0, + "content": "Defining the semantic representation begins by gathering an over-complete set \\(\\ell\\) of \\(N_{\\ell}\\) word descriptors that correspond to each 3DMM. Our method supports any set of words that relate to the model: we couple the body model with body descriptors, and choose face descriptors for the face model. For the body models SMPL and SMPL-X, we adopt the set of descriptors used in [34], for FLAME face expression we use an expanded set based on [9], and for FLAME face shape we used our own set of descriptors. Lastly, since the SMAL model was trained on 5 animal families (Felidae-big cats, Canidae-dogs, Equidae-horses, Bovidae-cows, Hippopotamidae-hippos) we used a set of descriptors that contains animals from these given families. The full list of descriptors can be found in the supplemental file. We note that our method can be used with any desired set of relevant descriptors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.222, + 0.864 + ], + "angle": 0, + "content": "3.2. CLIP Ratings" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We use CLIP [32] to encode each text descriptor \\(\\ell_{i}\\) using their text encoder \\(e^{\\ell_i} = CLIP_{Text}(\\ell_i)\\). Each mesh created" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.345, + 0.895, + 0.485 + ], + "angle": 0, + "content": "from a random coefficient vector \\( F(\\vec{\\beta}, \\theta, \\gamma) \\) is rendered to create a set of images \\( I_{j} \\), and these images \\( I_{j} \\) are encoded by CLIP image encoder \\( e^{I_j} = CLIP_{Image}(I_j) \\) to the same latent space of the encoded text (see Figure 2). In this space, we can compute the compatibility between the encoded image \\( e^{I_j} \\) and each encoded label \\( e^{\\ell_i} \\) using cosine similarity to get a score \\( \\Omega(I_j, \\ell_i) = \\cos(e^{I_j}, e^{\\ell_i}) \\). This way, each random coefficients vector \\( \\vec{\\beta} \\) or \\( \\vec{\\psi} \\) is paired with a vector of the descriptors' scores:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.493, + 0.868, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\Omega_ {i = 1} ^ {N _ {\\ell}} = [ \\Omega (I _ {j}, \\ell_ {1}), \\dots , \\Omega (I _ {j}, \\ell_ {N _ {\\ell}}) ] = [ \\Omega_ {1}, \\dots , \\Omega_ {N _ {\\ell}} ]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.522, + 0.646, + 0.537 + ], + "angle": 0, + "content": "containing \\(N_{\\ell}\\) scores." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.546, + 0.699, + 0.563 + ], + "angle": 0, + "content": "3.3. Descriptors Selection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.795 + ], + "angle": 0, + "content": "There are many words that could describe a body shape, face shape or facial expression. As the score of a given image and word descriptor in CLIP's embedding space is calculated by their semantic similarity, there might be many possible word descriptors for a single image, and many possible images for a single word descriptor. For example, both the words \"happy\" and \"smile\" would have a high correlation with a smiling face, and therefore, their effect on a facial expression is entangled. Similarly, \"tall\" and \"short\" are entangled in a body shape model. Our key observation is that a large number of descriptors are difficult to handle for interactive modeling, and the larger the number, the more entangled they are in terms of their effect on the shape - making it even harder to achieve the desired shape results (see Table 1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Contrary to BodyTalk [34], our goal is to use a minimal set of semantic word descriptors, while covering as much of the PCA shape space of each 3DMM as possible. We present an algorithm for selecting a subset of semantic descriptors with two main objectives: they should cover a large part of the shape space, and they should be disentangled as much as possible." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14557" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "Our idea is to first segment the shape space into regions, and assign descriptors to each region. This segmentation assures that we cover the shape space as we do not neglect any region. Next, we select for each region a set of descriptors by checking their variance (i.e., promoting coverage), and removing correlated descriptors, leaving the remaining ones as disentangled as possible." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.199, + 0.47, + 0.274 + ], + "angle": 0, + "content": "Clustering. To segment the shape space, we use our \\( N_{\\text{samples}} \\) dataset as a proxy to represent a sampling of the space and cluster their images (created as described in Section 3.1) encoded to CLIP's latent space. We use K-means algorithm with Silhouette score to find \\( K \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.276, + 0.469, + 0.382 + ], + "angle": 0, + "content": "Next, we compare all encoded images \\( e^{I} \\) in each cluster \\( C_{i} \\) to all \\( N_{\\ell} \\) encoded word descriptors \\( e^{\\ell_i} \\) for \\( i = 1,\\dots ,N_{\\ell} \\). The top 5 most similar descriptors in a cluster are given a vote for this cluster. We normalize the value of votes by the size of each cluster. Then, each word descriptor is assigned exclusively to the cluster with the largest sum of votes for this descriptor." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.384, + 0.469, + 0.595 + ], + "angle": 0, + "content": "Choosing descriptors. To anticipate the effect of a given descriptor on the 3DMM, we use the variance of its score \\(\\Omega(I', \\ell_i)\\) on all images \\(I'\\) in the dataset. The larger the variance, the more descriptive this descriptor will be in terms of shape variation. To anticipate the level of entanglement between descriptors, we use their correlation. For each descriptor \\(\\ell_i\\) we build a vector of \\(N_{\\text{samples}}\\) scores \\(\\Omega(I', \\ell_i)\\) for all images \\(I'\\) in the dataset. Correlation is measured between these vectors. The larger the correlation between two descriptors, the more similar the effect of this descriptor is on the shape, and the more entangled they are (we show some correlation plot examples in the supplemental file). Lastly, we apply antonyms and synonyms detection to verify there are no such pairs in our final set of descriptors." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.598, + 0.469, + 0.793 + ], + "angle": 0, + "content": "The process of choosing the descriptors is first performed on each cluster separately, and then merging the lists. We sort the cluster descriptors according to their variance in descending order. Thus, the first descriptor that is chosen from each cluster is the descriptor with the highest variance. Once a descriptor is chosen, we iterate over the other descriptors that are left for this cluster and filter out the ones that have a high correlation with the chosen descriptor (we use the median correlation value of all pairs as the threshold). In addition, before adding the next descriptor to the set, we check if its synonyms or antonyms are already chosen, and if so, we skip it. This process continues until the list of descriptors is exhausted." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Lastly, we merge the lists created from all clusters and sort again according to the descriptor variance. In a similar manner as we did for each cluster, we filter from the merged list all correlated descriptors, synonyms and antonyms. Finally, we arrive at a set of \\( d \\) chosen semantic descriptors for this model. \\( d \\) can be different for each model, but our algorithm can also support any preset number of descrip" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.088, + 0.902, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.293, + 0.892, + 0.382 + ], + "angle": 0, + "content": "Figure 3: Models created using an interactive slider application for SMAL[45] model (the blue model on the left is the neutral model). These animals were created by interpolating between four semantic descriptors - lion, cow, cat and dog. As can be seen, other animals could also be designed as well as intermixed creatures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.892, + 0.477 + ], + "angle": 0, + "content": "tors by removing descriptors from the bottom of the list or adding back descriptors that were filtered out according to their variance order (see Table 1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.892, + 0.571 + ], + "angle": 0, + "content": "Our algorithm can also support a preset list of descriptors if the user seeks to fit a 3D model with respect to some specific descriptors. In this case, the process described above for choosing the final set of descriptors is simply initialized with the user pre-defined descriptors, so that they are always contained in the final set." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.595, + 0.604, + 0.611 + ], + "angle": 0, + "content": "3.4. Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For simplicity, we will denote the coefficients vector by \\(\\vec{\\xi}\\) for both shape and expression parameters. Our goal is to define the mapping from the semantic representation of the \\(d\\) descriptors to the parametric representation represented by \\(\\vec{\\xi}\\). From our dataset creation (Section 3.1), each coefficients vector \\(\\vec{\\xi} = [\\xi_1,\\dots,\\xi_{10}]\\) is paired with a scores vector \\(\\vec{\\Omega} = [\\Omega_1,\\dots,\\Omega_{N_\\ell}]\\). Hence, for a given scores vector \\(\\vec{\\Omega}\\), the goal is to predict the corresponding coefficients vector \\(\\vec{\\xi}\\). This mapping resembles the one presented in [34], only we rely on CLIP score to label our data, providing a single score rather than multiple scores obtained by crowd sourcing, and we do not assume a linear relationship between the word descriptors and the 3DMM's coefficients vector. Instead, we train a multi-layer-perceptron with ReLU activations as the mapping function using our paired data and a simple \\(L_{2}\\) loss \\(\\mathcal{L} = ||\\hat{\\vec{\\xi}} -\\vec{\\xi} ||_2\\). The network consists of hidden layers (500 and 800 neurons), and is trained for 50 epochs." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14558" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.09, + 0.208, + 0.108 + ], + "angle": 0, + "content": "4. Applications" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.115, + 0.254, + 0.129 + ], + "angle": 0, + "content": "4.1. Interactive Sliders" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.473, + 0.41 + ], + "angle": 0, + "content": "Using the set of semantic descriptors \\(\\ell\\) chosen by our algorithm, we can build a simple and intuitive interface to control the 3DMM. This is because the descriptors are least entangled and cover the PCA space well. First, we train a mapper \\(M\\) with the chosen set of descriptors \\(\\ell\\) as described above (Section 3.4). Next, we create a slider for each descriptor whose value represents the desired CLIP score for that specific descriptor (see Figure 4). The user can now change each slider to trigger a re-prediction of the 3DMM coefficients. The computation is interactive as the interface takes all sliders' values as input \\(\\vec{\\Omega}\\) to \\(M\\), which maps them in a forward pass to the 3DMM coefficients vector \\(\\vec{\\xi}\\) and renders the corresponding 3D mesh. Thus, the modeler can interact and understand the relationships between the descriptors and their effect on the mesh. Figure 3 demonstrates examples of 3D meshes of different animals that were created using an interface created for the SMAL model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.42, + 0.436, + 0.436 + ], + "angle": 0, + "content": "4.2. Zero-Shot Image to Shape Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.443, + 0.47, + 0.61 + ], + "angle": 0, + "content": "Our method allows to leverage CLIP's semantic understanding to define a zero-shot image to shape reconstruction. Given any image of a person, we embed it into CLIP's latent space and use the similarity scores of the image against the set of descriptors \\(\\ell^{M}\\) chosen for a mapper \\(M\\) to obtain the vector \\(\\vec{\\Omega}\\). Next, we simply feed \\(\\vec{\\Omega}\\) through the network to get the 3DMM coefficients vector \\(\\vec{\\xi}\\) that best fits the image and create the shape from \\(\\vec{\\xi}\\). Our method differs from state-of-the-art methods in that it enables the user to refine the zero-shot prediction very simply using semantic sliders starting from a relatively good initial guess." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.623, + 0.21, + 0.64 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.648, + 0.274, + 0.664 + ], + "angle": 0, + "content": "5.1. Coverage Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.47, + 0.826 + ], + "angle": 0, + "content": "To evaluate a mapper \\(M\\) that was trained with a given set of descriptors \\(\\ell\\), we would like to measure its effect on the 3DMM. For each descriptor \\(\\ell_i \\in \\ell\\) we do a forward pass through the mapper twice: the first pass by setting a low score to \\(\\Omega_i\\) and the second pass by setting a high score to \\(\\Omega_i\\). The other coefficients are set to a default value and remain constant. Each such pass creates a parameter vector \\(\\vec{\\xi} = [\\xi_1, \\dots, \\xi_{10}]\\), which represents the resulting 3DMM shape or expression. We denote these vectors as \\(\\vec{\\xi}(i)_{low}\\) and \\(\\vec{\\xi}(i)_{high}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.9 + ], + "angle": 0, + "content": "We measure the geometric effect of a descriptor \\(\\ell_{i}\\) on a 3DMM by examining the deformation of the vertices of the mesh. This is evaluated by comparing the position of each vertex in the two extreme cases of \\(\\vec{\\xi} (i)_{low}\\) and \\(\\vec{\\xi} (i)_{high}\\). Hence, for each vertex \\(v\\) the size of deformation is de" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.088, + 0.625, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.175, + 0.622, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.238, + 0.621, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.118, + 0.895, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.329, + 0.895, + 0.404 + ], + "angle": 0, + "content": "Figure 4: A simple and intuitive interactive application can be defined using sliders for each semantic descriptor. With Semantiy, the sliders have strong semantic meaning and are disentangled so that changing one will have very limited affect on others." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.892, + 0.489 + ], + "angle": 0, + "content": "fined as \\(\\delta'(v) = ||v_{low} - v_{high}||_2\\). We can normalize these values by \\(\\delta_{\\max} = \\max_{v \\in Mesh} \\delta'(v)\\) and get a value \\(\\delta(v) = \\delta'(v) / \\delta_{\\max}\\) between 0 and 1 (see Figure 5)." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.506, + 0.775, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.802, + 0.553, + 0.897, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.761 + ], + "angle": 0, + "content": "Figure 5: The geometric effect of a given descriptor. The darker the color of the vertex is (more red), the more it is affected by the descriptor. To define coverage we collect all vertices with a value larger than a given threshold \\(\\theta\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Using these differences, we can evaluate the coverage of a specific descriptor by measuring the number and positions of vertices it affects in the mesh. For example, we can look at all vertices \\( v \\) whose value is above a threshold \\( \\delta(v) > \\theta \\) for any given descriptor. We can also compare the overlap between two descriptors by measuring the intersection-over-union of their covered vertices. We use these measure" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14559" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "in our ablation studies as described next in Section 5.2. Note that this measure is only an approximation as it measures the size of the change of vertices and not their direction." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.148, + 0.188, + 0.162 + ], + "angle": 0, + "content": "5.2. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.171, + 0.473, + 0.403 + ], + "angle": 0, + "content": "To further evaluate our algorithm's performance, we conducted experiments that include training the mapper with different numbers of descriptors and training on different sizes of training data. As we do not have valid ground truth to estimate the accuracy of our mapper we created a test set that contains several body shapes that were generated by the first 10 blendshapes \\(\\vec{\\xi}\\) of the SMPL-X model. In each experiment we optimize the descriptors' values \\(\\vec{\\Omega}\\), to minimize the error \\(||\\hat{\\vec{\\xi}} - \\vec{\\xi}||_2\\). The goal of these experiments is to test the expressive power of the given configuration to match a given shape. The smaller the error the more expressive the method is. We show the results of these ablations in the supplemental file. We used this evaluation to choose the hyper-parameters we use in our algorithm: training with 3K sample images with texture for all models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.424, + 0.345, + 0.44 + ], + "angle": 0, + "content": "5.3. Interactive Sliders User Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.473, + 0.886 + ], + "angle": 0, + "content": "To evaluate Semantify's performance we conducted a user study that compares our approach against BodyTalk [34], as well as the original basis of the 3DMM. The user is presented with a target 3D model (randomly created using the 3DMM) and is tasked to fit a 3DMM mesh model to the target shape using a set of sliders, using the two compared interfaces (see Figure 4) at random order. Users were given up to 5 minutes to complete the task. BodyTalk was implemented on the SMPL body model, therefore we compare it to our SMPL male and female models. In another study, we compare our SMPLX male and female models to the original PCA-based axes. In both studies, we seek to evaluate the accuracy of the user-fitted 3DMMs, the time that it took to fit the model, as well as the overall subjective experience of users. Since we cannot obtain the outputs of BodyTalk in the first study, we asked humans to rate the results, and determine which of them fits more accurately to the input shape. We had 10 users perform both studies (5 males, 5 females aged between 23 to 54), 7 of which are novices and 3 of which were professional 3D modelers. We present the quantitative results of these studies in Table 2. As can be seen, Semantify achieves better accuracy in less time. Note that theoretically only the baseline can reach zero error, but because it is very difficult to handle, the error obtained was higher than Semantify. In addition, we asked the users for feedback on the experience of using the different applications (A was ours, B was the alternative). Here are some examples of such reviews (more examples could be found in the supplemental file):" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "\"Using application B, occasionally when changing one" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "slider it affects the other, which causes the user to start all over again\"." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.122, + 0.892, + 0.153 + ], + "angle": 0, + "content": "\"Using application B every minor change in a certain slider generated major changes in the rest of the sliders\"" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.153, + 0.892, + 0.197 + ], + "angle": 0, + "content": "\"Application A was far friendlier and I sensed as though I maintained much more control over the different body features\"" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.199, + 0.892, + 0.23 + ], + "angle": 0, + "content": "\"The abundance of sliders on application B only made it harder to control, not the other way around\"." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.241, + 0.734, + 0.256 + ], + "angle": 0, + "content": "5.4. Zero Shot Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.265, + 0.892, + 0.476 + ], + "angle": 0, + "content": "To measure our zero-shot 3D-shape reconstruction method of SMPLX model, we use HBW (Human Bodies in-the-wild) dataset presented in SHAPY [7], containing ground-truth 3D body scans with in-the-wild images. Table 3 shows the performance of our method compared to SHAPY and PixIE [12] on all of HBW validation set. Note that SHAPY also takes keypoints (usually obtained from openPose [5]) as input. Some examples are shown in Figure 6. As can be seen, our zero-shot performance are on-par with state of the art methods, but we have the advantage of using semantic descriptors. This allows the user to fine tune the results from a very good initial guess, simply by using the semantic sliders. Some examples results with around a minute of fine-tuning can be found in the supplemental file." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.492, + 0.619, + 0.507 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.892, + 0.669 + ], + "angle": 0, + "content": "We have presented a method to edit a 3D parametric model using semantically meaningful descriptors by first choosing a subset of descriptors judiciously and then learning a non-linear mapping from this set to the 3DMM coefficients. All this is done in a self-supervised manner harnessing the abilities of CLIP visual-language model. Our method can support a more intuitive user interface that allows even novices to model 3D shapes. It also supports other applications such as zero-shot reconstruction and fine-tuning of shapes from images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Limitations and future directions. There are several limitations to our method. First, our mapper's performance relies heavily on the data that is created (Section 3.1). For example, a dataset must contain extreme samples to achieve a more expressive mapper. Second, our goal with Semantify was to build a semantic representation of a given 3DMM without a human-in-the-loop. However, in some cases manually tuned modeling for a specific 3DMM would probably improve the 3DMM mapper's performance (we further elaborate on this in the supplemental file). Lastly, as shown in many other works, CLIP has rich semantic understanding, but there are cases in which its performance degrades. For example, performance is better with textured meshes, so using various textures for each 3DMM may improve the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14560" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.87, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.459, + 0.895, + 0.491 + ], + "angle": 0, + "content": "Figure 6: Results of zero-shot image to shape reconstruction. We compare Semantify against SHAPY and PixIE on SMPLX 3DMM. Colors indicate error compared to ground truth (darker means larger error)." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.511, + 0.428, + 0.596 + ], + "angle": 0, + "content": "
Time (min)Score
MalesFemalesTotal
BodyTalk3.223.163.190.04%
ours2.121.451.590.96%
" + }, + { + "type": "table", + "bbox": [ + 0.453, + 0.504, + 0.91, + 0.589 + ], + "angle": 0, + "content": "
Mean Error (cm)Time (min)
MalesFemalesTotalMalesFemalesTotal
Baseline0.1220.1110.1163.182.092.43
ours0.1210.0990.1102.222.172.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.613, + 0.895, + 0.644 + ], + "angle": 0, + "content": "Table 2: User-study results for the SMPL 3DMM compared against BodyTalk [34] (left, Score means the percent that this model was selected as better fitting the input), and SMPL-X parametric model compared against the raw blendshapes (right)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.668, + 0.468, + 0.745 + ], + "angle": 0, + "content": "
MalesFemalesTotal
SHAPY0.01969340.0144220.017085
PIXIE0.0247790.0152650.020324
Semantify (ours)0.0226320.0165640.019666
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.77, + 0.47, + 0.83 + ], + "angle": 0, + "content": "Table 3: Comparing our zero-shot shape reconstruction from image method against SHAPY[7] and PixIE [12]. The values demonstrate the average MSE over HBW's in-the-wild validation set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "mapper's performance. Furthermore, more subjective descriptors such as \"cute\", \"scary\", \"pretty\" etc., are harder to" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.67, + 0.892, + 0.699 + ], + "angle": 0, + "content": "rate using CLIP, and would probably not work well in our model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.894, + 0.797 + ], + "angle": 0, + "content": "Future work can also improve the zero-shot 3D shape reconstruction by using prior information about its context (for example, the gender of a person, body keypoints as used in [7] etc.). Adding such priors along with CLIP's semantic understanding may surpass the existing state-of-the-art solutions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.825, + 0.681, + 0.843 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "This research was supported by the Israel Science Foundation (grant No. 1390/19). We would also like to thank Eyal Gomel for valuable discussions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14561" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.116, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[1] Rameen Abdal, Peihao Zhu, John Femiani, Niloy J. Mitra, and Peter Wonka. Clip2stylegan: Unsupervised extraction of stylegan edit directions. CoRR, abs/2112.05219, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.181, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Clipface: Text-guided editing of textured 3d morphable models. arXiv preprint arXiv:2212.01406, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.246, + 0.471, + 0.322 + ], + "angle": 0, + "content": "[3] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. Scape: Shape completion and animation of people. ACM Transactions on Graphics (TOG), 24:408-416, 07 2005." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.326, + 0.471, + 0.387 + ], + "angle": 0, + "content": "[4] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.391, + 0.472, + 0.466 + ], + "angle": 0, + "content": "[5] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43:172–186, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.471, + 0.47, + 0.517 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.521, + 0.471, + 0.61 + ], + "angle": 0, + "content": "[7] Vasileios Choutas, Lea Müller, Chun-Hao P. Huang, Siyu Tang, Dimitrios Tzionas, and Michael J. Black. Accurate 3d body shape regression using metric and semantic attributes. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.616, + 0.471, + 0.691 + ], + "angle": 0, + "content": "[8] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Monocular expressive body regression through body-driven attention. In European Conference on Computer Vision (ECCV), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.696, + 0.471, + 0.77 + ], + "angle": 0, + "content": "[9] Radek Daněček, Michael J. Black, and Timo Bolkart. Emotion driven monocular face capture and animation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20311-20322, June 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.851 + ], + "angle": 0, + "content": "[10] Ginger Delmas, Philippe Weinzaepfel, Thomas Lucas, Francesc Moreno-Noguer, and Grégory Rogez. Postscript: 3d human poses from natural language. In European Conference on Computer Vision, pages 346-362. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[11] Bernhard Egger, William AP Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.536, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "Sami Romdhani, et al. 3d morphable face models—past, present, and future. ACM Transactions on Graphics (TOG), 39(5):1-38, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.141, + 0.894, + 0.215 + ], + "angle": 0, + "content": "[12] Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael Black. Collaborative regression of expressive bodies using moderation. In International Conference on 3D Vision (3DV), pages 792-804, Dec. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.894, + 0.309 + ], + "angle": 0, + "content": "[13] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8649-8658, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.314, + 0.894, + 0.405 + ], + "angle": 0, + "content": "[14] Thomas Gereg, Andreas Morel-Forster, Clemens Blumer, Bernhard Egger, Marcel Luthi, Sandro Schoenborn, and Thomas Vetter. Morphable face models - an open framework. 2018 13th IEEE International Conference on Automatic Face amp; Gesture Recognition (FG 2018), May 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.408, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[15] Matthew Hill, Stephan Streuber, Carina Hahn, Michael Black, and Alice O'Toole. Exploring the relationship between body shapes and descriptions by linking similarity spaces. Journal of Vision, 15(12):931-931, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.487, + 0.892, + 0.548 + ], + "angle": 0, + "content": "[16] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhonggang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.551, + 0.892, + 0.582 + ], + "angle": 0, + "content": "[17] IEEE. A 3D Face Model for Pose and Illumination Invariant Face Recognition, Genova, Italy, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.585, + 0.892, + 0.66 + ], + "angle": 0, + "content": "[18] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, P. Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 857-866, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.664, + 0.892, + 0.724 + ], + "angle": 0, + "content": "[19] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.728, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[20] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. SIGGRAPH Asia 2022 Conference Papers, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.807, + 0.892, + 0.868 + ], + "angle": 0, + "content": "[21] Youwang Kim, Ji-Yeon Kim, and Tae-Hyun Oh. Clip-actor: Text-driven recommendation and stylization for animating human meshes. In European Conference on Computer Vision, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[22] Muhammed Kocabas, Chun-Hao P. Huang, Joachim Tesch, Lea Müller, Otmar Hilliges, and Michael J." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14562" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.114, + 0.092, + 0.469, + 0.151 + ], + "angle": 0, + "content": "Black. SPEC: Seeing people in the wild with an estimated camera. In Proc. International Conference on Computer Vision (ICCV), pages 11035-11045, Oct. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.156, + 0.47, + 0.215 + ], + "angle": 0, + "content": "[23] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.469, + 0.295 + ], + "angle": 0, + "content": "[24] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.3, + 0.469, + 0.359 + ], + "angle": 0, + "content": "[25] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. *Smpl: A skinned multi-person linear model*. ACM Trans. Graph., 34(6), nov 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.365, + 0.469, + 0.424 + ], + "angle": 0, + "content": "[26] Xudong Lou, Yiguang Liu, and Xuwei Li. Tecm-clip: Text-based controllable multi-attribute face image manipulation. In Proceedings of the Asian Conference on Computer Vision, pages 1942–1958, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.429, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[27] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.508, + 0.469, + 0.567 + ], + "angle": 0, + "content": "[28] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. ArXiv, abs/2212.08751, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.573, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[29] Ahmed A. A. Osman, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Supr: A sparse unified part-based human representation. In Shai Avidan, Gabriel Brostow, Moustapha Cissé, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision – ECCV 2022, pages 568–585, Cham, 2022. Springer Nature Switzerland." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.682, + 0.47, + 0.757 + ], + "angle": 0, + "content": "[30] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.852 + ], + "angle": 0, + "content": "[31] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.536, + 0.093, + 0.892, + 0.137 + ], + "angle": 0, + "content": "et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.142, + 0.891, + 0.185 + ], + "angle": 0, + "content": "[33] Hyewon Seo, Frederic Cordier, and Nadia Thalmann. Synthesizing animatable body models with parameterized shape modifications. 07 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.265 + ], + "angle": 0, + "content": "[34] Stephan Streuber, M. Alejandra Quiros-Ramirez, Matthew Q. Hill, Carina A. Hahn, Silvia Zuffi, Alice O'Toole, and Michael J. Black. Body Talk: Crowd-shaping realistic 3D avatars with words. ACM Trans. Graph. (Proc. SIGGRAPH), 35(4), July 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.27, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[35] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.892, + 0.393 + ], + "angle": 0, + "content": "[36] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.399, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[37] Ayush Tewari, Michael Zollhöfer, Pablo Garrido, Florian Bernard, Hyeongwoo Kim, Patrick Pérez, and Christian Theobalt. Self-supervised multi-level face model learning for monocular reconstruction at over 250 hz. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.508, + 0.892, + 0.582 + ], + "angle": 0, + "content": "[38] Can Wang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.588, + 0.892, + 0.647 + ], + "angle": 0, + "content": "[39] Can Wang, Ruixia Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. ArXiv, abs/2212.08070, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.652, + 0.892, + 0.712 + ], + "angle": 0, + "content": "[40] Jinbao Wang, Shujie Tan, Xiantong Zhen, Shuo Xu, Feng Zheng, Zhenyu He, and Ling Shao. Deep 3d human pose estimation: A review. Computer Vision and Image Understanding, 210:103225, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.892, + 0.805 + ], + "angle": 0, + "content": "[41] Lizhen Wang, Zhiyua Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR2022), June 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[42] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Ghum & ghuml: Generative 3d human shape and articulated pose models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6184-6193, 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14563" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.091, + 0.47, + 0.183 + ], + "angle": 0, + "content": "[43] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. *Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction.* In *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition* (CVPR), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.186, + 0.47, + 0.263 + ], + "angle": 0, + "content": "[44] Silvia Zuffi, Angjoo Kanazawa, and Michael J. Black. Lions and tigers and bears: Capturing non-rigid, 3d, articulated shape from images. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3955-3963, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.266, + 0.47, + 0.34 + ], + "angle": 0, + "content": "[45] Silvia Zuffi, Angjoo Kanazawa, David Jacobs, and Michael J. Black. 3D menagerie: Modeling the 3D shape and pose of animals. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), July 2017." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.091, + 0.47, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14564" + } + ] +] \ No newline at end of file diff --git a/2023/Semantify_ Simplifying the Control of 3D Morphable Models Using CLIP/ff2d3341-5fbc-47bb-9ce8-0e16c61108bc_origin.pdf b/2023/Semantify_ Simplifying the Control of 3D Morphable Models Using CLIP/ff2d3341-5fbc-47bb-9ce8-0e16c61108bc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cd79fc9963920c2a78576f4e1e5a2dc245d5e65e --- /dev/null +++ b/2023/Semantify_ Simplifying the Control of 3D Morphable Models Using CLIP/ff2d3341-5fbc-47bb-9ce8-0e16c61108bc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4aba61f98d7e07370a9a32cb7563f41e4c07cd434f72794c26c9f8c2c4a49e25 +size 2673490 diff --git a/2023/Semantify_ Simplifying the Control of 3D Morphable Models Using CLIP/full.md b/2023/Semantify_ Simplifying the Control of 3D Morphable Models Using CLIP/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0ac8a0bb44099c482b5ca67a784c7b3be58d7f3d --- /dev/null +++ b/2023/Semantify_ Simplifying the Control of 3D Morphable Models Using CLIP/full.md @@ -0,0 +1,250 @@ +# Semantify: Simplifying the Control of 3D Morphable Models using CLIP + +![](images/473800955cb810138a380f9346d36cad6a30af74e50d3e7fbaf71ad28b51722a.jpg) +Figure 1: Semantify offers a method to create and edit a 3D parametric model using semantically meaningful descriptors. Semantify is based on a self-supervised method that utilizes the semantic power of CLIP language-vision model to build a mapping between semantic descriptors to 3DMM model coefficients. This can be used in an interactive application defining a slider for each descriptor (a), or to fit a model to an image in a zero shot manner by feeding the image into CLIP and obtaining a vector of semantic scores that can be mapped to shape parameters (b). + +# Abstract + +We present Semantify: a self-supervised method that utilizes the semantic power of CLIP language-vision foundation model [32] to simplify the control of 3D morphable models. Given a parametric model, training data is created by randomly sampling the model's parameters, creating various shapes and rendering them. The similarity between the output images and a set of word descriptors is calculated in CLIP's latent space. Our key idea is first to choose a small set of semantically meaningful and disentangled descriptors that characterize the 3DMM, and then learn a non-linear mapping from scores across this set to the parametric coefficients of the given 3DMM. The nonlinear mapping is defined by training a neural network without a human-in-the-loop. We present results on numerous 3DMMs: body shape models, face shape and expression models, as well as animal shapes. We demonstrate how our method defines a simple slider interface for intuitive modeling, and show how the mapping can be used to instantly fit a 3D parametric body shape to in-the-wild images. See our project page at https://omergral.github.io/Semantify/ + +# 1. Introduction + +3D modeling techniques have evolved tremendously over the last few years. Such techniques are used in countless industries including the burgeoning AR/VR industry, fashion design, game development, film, and many others. However, designing a high-quality 3D model is not a simple task for most people and might require a well-trained 3D artist. Even when using more recent parametric morphable models (3DMM) [24, 31, 45, 3, 42] it is still hard for humans to understand how to choose the correct set of parameters, e.g. to achieve a specific human body shape or facial expression. It is also difficult to find what are the limits of the given parametric model in terms of coverage and expressiveness. The reason is that in most cases, the provided set of parameters is not interpretable, as they are commonly calculated using automatic optimization mechanisms followed by dimensionality reduction using PCA. Thus, they carry no clear semantic meaning. + +To simplify the use of 3DMMs and allow for natural interactive human modeling, a key research question is - how to insert semantics into 3DMM control? Previous approaches [34] relied mostly on human intelligence and la + +beling, which is often time consuming and expensive. In addition, previous methods create a large number of semantic control descriptors that are often correlated and entangled. This means it is difficult to anticipate the effect of each descriptor as changing a value in one may modify others, resulting in difficulties to control the model. + +In this paper, instead of using human labeling, we rely on the remarkable abilities of huge foundation models that combine natural language and visual understanding. We present a self-supervised method that utilizes the semantic power of CLIP [32] to define a method to control 3DMMs that carries two main advantages. First, it is more natural for humans to use as it allows modeling using a small number of semantically meaningful descriptors, that cover the space of deformations but are disentangled. Second, it covers even extreme examples in the shape/pose space of parametric models as it utilizes CLIP's pre-training on a large number of images. The main idea is first to use CLIP to select a small subset of semantically meaningful and disentangled descriptors, and then learn a non-linear mapping of this set to the coefficients of the given 3DMM. + +Given a parametric 3DMM, we sample its parameter space to create a dataset containing a variety of 3D mesh shapes. We then render each mesh from different camera views to create a diverse set of images corresponding to the parameter samples (see Figure 2). Next, we gather a set of semantically descriptive textual terms related to the parametric 3D model, which we call descriptors. We encode both the images (using CLIP's image encoder) and the descriptors (using CLIP's text encoder) into CLIP's latent space and compare them. This defines a vector of similarity scores between the input vector of 3DMM coefficients (of each image sample) and each corresponding semantic descriptor. Next, we define a selection scheme to choose a small number of descriptors that are de-correlated to control the model. Lastly, we train a neural network to learn a mapping between the vector of similarity scores to the vector of 3DMM parameters. + +We demonstrate how the learned mapping can be used to define an interface to control a 3DMM in a way that is simple and effective using a small set of semantically meaningful sliders (see Figure 1 and 4). Such sliders are easy to employ for designing high-quality 3D models and cover the shape space well. We demonstrate this for four parametric models: human face's shape and expression (FLAME [24]), human body shapes (SMPL [25] and SMPL-X [31]), and even animals (SMAL [45]). We also show how the mapping can be used to instantly fit a 3D parametric body shape to an input image that works well "in the wild" even in extreme poses and body shapes. + +Our main contribution is a novel, self-supervised method for defining a small set of semantic descriptors to control a parametric model more naturally, by learning a mapping + +from a semantic space to a parametric representation without human-in-the-loop. We demonstrate the effectiveness of our approach on several parametric models and utilize it to define a simple interface for modeling and to instantly fit a 3D model to images in the wild. We will release our code for future research. + +# 2. Related Work + +Our work encompasses a wide range of disciplines, including 3D parametric modeling, language-vision models, and zero-shot reconstruction. + +3D Morphable Models. Much research has gone into improving 3D modeling of humans bodies, faces and other object classes, such as animals [11, 4, 25, 31]. 3D Morphable Models (3DMM) are a powerful tool to parameterize the variation in geometry of objects belonging to a certain class. Dating back to 1999, Blantz et al. [4] proposed 3DMMs to capture the variation of human faces and applied dimensionality reduction using Principal Component Analysis (PCA). + +In contrast to a 3D artist freely controlling the vertices of a mesh, using each of these PCA axes to deform the vertices induces a strong statistical prior on the resulting geometry. The resulting deformations are global in nature, strongly constraining the geometry to the manifold of statistically plausible meshes, as the per-axis coefficients correlate with their likelihood. While initially proposed for human faces, statistical 3DMMs have been developed for human body models [3, 25, 40, 31, 42, 29], facial models [24, 37, 41, 43], and even 3D animal models [45, 44]. + +Semantic Models. PCA-based 3DMMs are efficiently and easily constructed, and admit an orthogonal and importance-ordered basis for the deformation. One disadvantage of these models is their lack of interpretability for humans. The axes of maximal variation in geometry, which form the deformation basis for each vertex, do not necessarily correspond to any semantically meaningful geometric change (e.g., raising an eyebrow), but instead, each axis induces a rather global change that may introduce correlations or effects that are unwanted to a human modeler, making it hard to obtain the desired shape or expression. + +Some 3DMMs [17, 14] alleviate this issue by using deformation bases that are hand-crafted by artists, often referred to as blendshapes. While this alleviates the interpretability issue, it loses the data-driven and explicit statistical prior, requires manual work, and does not guarantee the expressiveness of the resulting model. + +First to address the semantic issue of PCA-based 3DMMs in the human body domain were Seo et al. [33] which used metrics such as hip-to-waist ratio, body fat percentage, and height to regress 3DMM shape parame + +![](images/66f4710b2aea4be682757d90d68a3adcd51df36fb3a70871158233b62a01ae87.jpg) +Figure 2: Learning a mapping from Semantic to Parametric space. (a) Given a coefficient vector $\vec{\xi}$ we create the 3DMM mesh. The mesh is rendered from several views. Each Rendered image $I'$ is passed into CLIP along with a set of semantic descriptors $\vec{\ell}$ . The difference between each descriptor and the image in CLIP latent space is calculated and stored in the corresponding entry of the similarity vector $\vec{\Omega}$ . (b) Using a large set of such random pairs of $(\vec{\Omega},\vec{\xi})$ , we train a network to learn the mapping from semantic space to parametric space. + +ters. Hill et al.[15] suggested examining the relationship between body shapes and description, by linking two similarity spaces - one created from body descriptions and the other from full-body laser scans. BodyTalk [34] followed suit and approached this problem using the "wisdom of the crowd". In their work, they captured 256 male and female body shapes and represented each shape with 30 body descriptor words (e.g. curvy, fit, heavyet, round-Apple, etc.). They hired labelers to discretely rate the bodies with respect to these descriptors and designed a regression model to learn the relationship between the discrete ratings and the parameters of the corresponding body shape. In a more recent publication, SHAPY [7] bridges this gap by curating datasets of images with corresponding body measurements (e.g., from model agencies) followed by human ratings respective to a set of linguistic shape attributes. They train models to predict SMPLX shape parameters from attributes and/or body measurements, and vice-versa. + +Language-Guided 3D Modeling. Recent works have shown great success in learning latent representations that are capable of coupling visual signals with language. One such example is CLIP [32] model. CLIP uses a contrastive training scheme [6] on over 400 million pairs of images and their captions to build a shared latent representation for visual-textual content. In this space, similarity between images and text can be computed. It has been shown that CLIP is an effective tool for image generation tasks. For example, providing natural-language guided manipulation of human face imagery in StyleGAN space [30, 26, 1]. Similarly, CLIP can be employed in 3D: using differentiable rendering, images of a 3D scene are rendered in a forward pass and are scored against a text prompt with CLIP. Taking the similarity score as an optimization objective, gradients are back-propagated through the CLIP network and the rendering process, back to the underlying representation of the 3D scene. Text2Mesh [27] uses CLIP to optimize for col + +ors and positions of mesh vertices to match a text query. CLIP-Mesh [20] followed suit and proposed to use CLIP to guide normal and texture maps. This basic recipe has been applied to a variety of scene representations, such as point clouds [28] and Neural Fields [38, 18, 39, 13]. In the context of human faces and bodies, CLIP is used for generation of animation sequences [21, 36, 16, 35, 10, 2]. + +Zero-Shot 3DMM Shape Reconstruction. Recovering accurate, explicit meshes from 2D signals such as images or video is an under-constrained and over-parameterized objective. The low-dimensional underlying representation of 3DMMs proves useful not only for modeling, but for 3D reconstruction of geometry from images of humans. The parameters can serve as a data-driven regularization term for shape and pose, and the low dimensional representation convexifies the optimization problem of fitting such a model to an image, rather than a freely deforming mesh. Recent 3DMM-from-image methods either use iterative optimization schemes to fit the parameters of the 3DMM to the image [23] or directly regress the shape and pose parameters from an Image [12, 8, 22, 19]. Our method allows to use CLIP image and text mapping and simply feed-forward the scores through an MLP network to get a body shape. + +# 3. Method + +An overview of our method can be seen in Figure 2. In essence, our method defines a non-linear mapping from semantic space to parametric space by training a neural network to predict 3DMM coefficients from a vector of semantic attribute scores. We first create a dataset of rendered images of randomly sampled shapes (3.1), gather their corresponding semantic scores (3.2). Then, we employ a scheme to reduce the semantic descriptors to a subset of least-correlated descriptors (3.3) and then train the network (3.4). + +
Coverage / Overlap
MethodAlgorithm Choice2 descriptors5 descriptors10 descriptors15 descriptors
SMPLX-male(6) / 92.1% / 64.5%38.6% / 30.1%51.5% / 71.4%49.7% / 83.5%98.3% / 80.5%
SMPLX-female(4) / 49.7% / 56.7%27.1% / 49.5%50.0% / 67.1%47.4% / 83.0%82.4% / 83.4%
SMPLX-neutral(8) / 52.0% / 83.0%40.2% / 51.2%50.0% / 67.1%51.9% / 86.9%52.1% / 90.3%
SMPL-male(4) / 97.5% / 60.2%40.1% / 46.7%97.7% / 66.5%98.3% / 83.9%98.2% / 89.3%
SMPL-female(5) / 95.7% / 68.6%36.3% / 47.3%95.7% / 68.6%97.8% / 85.2%99.1% / 89.9%
SMPL-neutral(6) / 96.1% / 76.0%80.4% / 48.7%96.2% / 71.5%99.6% / 85.2%99.8% / 90.8%
FLAME-expression(3) / 13.4% / 56.8%10.4% / 47.4%19.3% / 67.2%27.0% / 77.3%34.4% / 85.0%
+ +Table 1: Choosing different number of descriptors for a model. We evaluate the coverage and overlap of vertices for a mapper that was trained by using different number of descriptors (see Section 5.1, the threshold used to measure coverage is 0.3). The number chosen by our algorithm is shown in parentheses at the beginning of each row. As can be seen, the more descriptors are chosen the larger the cover is, but the larger the overlap between the descriptors is as well. + +# 3.1. Dataset Creation + +The process of creating the data for training our model is similar across all 3DMMs (we demonstrate it on four different ones in this paper). For each model, we only use the first 10 principal components, that cover above $95\%$ of the variance. We draw $N_{\text{samples}}$ 10-dimensional random vectors of coefficients. For body models, we draw a random shape coefficients vector $\vec{\beta}$ , which yields a random shape. From the expression basis of the FLAME model, we randomly sample an expression coefficient vector $\vec{\psi}$ which yields a random expression. Since random values may also lead to noisy and unrealistic outputs, we limit the values of the sampled coefficients to a range of $k$ standard deviations of the model's coefficients ( $k = 2$ for body shapes and $k = 4$ for face expressions), yielding more realistic results. The other parameters of the 3DMM $(\theta, \gamma)$ remain neutral. + +Defining the semantic representation begins by gathering an over-complete set $\ell$ of $N_{\ell}$ word descriptors that correspond to each 3DMM. Our method supports any set of words that relate to the model: we couple the body model with body descriptors, and choose face descriptors for the face model. For the body models SMPL and SMPL-X, we adopt the set of descriptors used in [34], for FLAME face expression we use an expanded set based on [9], and for FLAME face shape we used our own set of descriptors. Lastly, since the SMAL model was trained on 5 animal families (Felidae-big cats, Canidae-dogs, Equidae-horses, Bovidae-cows, Hippopotamidae-hippos) we used a set of descriptors that contains animals from these given families. The full list of descriptors can be found in the supplemental file. We note that our method can be used with any desired set of relevant descriptors. + +# 3.2. CLIP Ratings + +We use CLIP [32] to encode each text descriptor $\ell_{i}$ using their text encoder $e^{\ell_i} = CLIP_{Text}(\ell_i)$ . Each mesh created + +from a random coefficient vector $F(\vec{\beta}, \theta, \gamma)$ is rendered to create a set of images $I_{j}$ , and these images $I_{j}$ are encoded by CLIP image encoder $e^{I_j} = CLIP_{Image}(I_j)$ to the same latent space of the encoded text (see Figure 2). In this space, we can compute the compatibility between the encoded image $e^{I_j}$ and each encoded label $e^{\ell_i}$ using cosine similarity to get a score $\Omega(I_j, \ell_i) = \cos(e^{I_j}, e^{\ell_i})$ . This way, each random coefficients vector $\vec{\beta}$ or $\vec{\psi}$ is paired with a vector of the descriptors' scores: + +$$ +\Omega_ {i = 1} ^ {N _ {\ell}} = [ \Omega (I _ {j}, \ell_ {1}), \dots , \Omega (I _ {j}, \ell_ {N _ {\ell}}) ] = [ \Omega_ {1}, \dots , \Omega_ {N _ {\ell}} ] +$$ + +containing $N_{\ell}$ scores. + +# 3.3. Descriptors Selection + +There are many words that could describe a body shape, face shape or facial expression. As the score of a given image and word descriptor in CLIP's embedding space is calculated by their semantic similarity, there might be many possible word descriptors for a single image, and many possible images for a single word descriptor. For example, both the words "happy" and "smile" would have a high correlation with a smiling face, and therefore, their effect on a facial expression is entangled. Similarly, "tall" and "short" are entangled in a body shape model. Our key observation is that a large number of descriptors are difficult to handle for interactive modeling, and the larger the number, the more entangled they are in terms of their effect on the shape - making it even harder to achieve the desired shape results (see Table 1). + +Contrary to BodyTalk [34], our goal is to use a minimal set of semantic word descriptors, while covering as much of the PCA shape space of each 3DMM as possible. We present an algorithm for selecting a subset of semantic descriptors with two main objectives: they should cover a large part of the shape space, and they should be disentangled as much as possible. + +Our idea is to first segment the shape space into regions, and assign descriptors to each region. This segmentation assures that we cover the shape space as we do not neglect any region. Next, we select for each region a set of descriptors by checking their variance (i.e., promoting coverage), and removing correlated descriptors, leaving the remaining ones as disentangled as possible. + +Clustering. To segment the shape space, we use our $N_{\text{samples}}$ dataset as a proxy to represent a sampling of the space and cluster their images (created as described in Section 3.1) encoded to CLIP's latent space. We use K-means algorithm with Silhouette score to find $K$ . + +Next, we compare all encoded images $e^{I}$ in each cluster $C_{i}$ to all $N_{\ell}$ encoded word descriptors $e^{\ell_i}$ for $i = 1,\dots ,N_{\ell}$ . The top 5 most similar descriptors in a cluster are given a vote for this cluster. We normalize the value of votes by the size of each cluster. Then, each word descriptor is assigned exclusively to the cluster with the largest sum of votes for this descriptor. + +Choosing descriptors. To anticipate the effect of a given descriptor on the 3DMM, we use the variance of its score $\Omega(I', \ell_i)$ on all images $I'$ in the dataset. The larger the variance, the more descriptive this descriptor will be in terms of shape variation. To anticipate the level of entanglement between descriptors, we use their correlation. For each descriptor $\ell_i$ we build a vector of $N_{\text{samples}}$ scores $\Omega(I', \ell_i)$ for all images $I'$ in the dataset. Correlation is measured between these vectors. The larger the correlation between two descriptors, the more similar the effect of this descriptor is on the shape, and the more entangled they are (we show some correlation plot examples in the supplemental file). Lastly, we apply antonyms and synonyms detection to verify there are no such pairs in our final set of descriptors. + +The process of choosing the descriptors is first performed on each cluster separately, and then merging the lists. We sort the cluster descriptors according to their variance in descending order. Thus, the first descriptor that is chosen from each cluster is the descriptor with the highest variance. Once a descriptor is chosen, we iterate over the other descriptors that are left for this cluster and filter out the ones that have a high correlation with the chosen descriptor (we use the median correlation value of all pairs as the threshold). In addition, before adding the next descriptor to the set, we check if its synonyms or antonyms are already chosen, and if so, we skip it. This process continues until the list of descriptors is exhausted. + +Lastly, we merge the lists created from all clusters and sort again according to the descriptor variance. In a similar manner as we did for each cluster, we filter from the merged list all correlated descriptors, synonyms and antonyms. Finally, we arrive at a set of $d$ chosen semantic descriptors for this model. $d$ can be different for each model, but our algorithm can also support any preset number of descrip + +![](images/2275e4d54f8c2c0a0b7a7c9673ad5499299f332252d0bf8304e84c3b08c3fc34.jpg) +Figure 3: Models created using an interactive slider application for SMAL[45] model (the blue model on the left is the neutral model). These animals were created by interpolating between four semantic descriptors - lion, cow, cat and dog. As can be seen, other animals could also be designed as well as intermixed creatures. + +tors by removing descriptors from the bottom of the list or adding back descriptors that were filtered out according to their variance order (see Table 1). + +Our algorithm can also support a preset list of descriptors if the user seeks to fit a 3D model with respect to some specific descriptors. In this case, the process described above for choosing the final set of descriptors is simply initialized with the user pre-defined descriptors, so that they are always contained in the final set. + +# 3.4. Training + +For simplicity, we will denote the coefficients vector by $\vec{\xi}$ for both shape and expression parameters. Our goal is to define the mapping from the semantic representation of the $d$ descriptors to the parametric representation represented by $\vec{\xi}$ . From our dataset creation (Section 3.1), each coefficients vector $\vec{\xi} = [\xi_1,\dots,\xi_{10}]$ is paired with a scores vector $\vec{\Omega} = [\Omega_1,\dots,\Omega_{N_\ell}]$ . Hence, for a given scores vector $\vec{\Omega}$ , the goal is to predict the corresponding coefficients vector $\vec{\xi}$ . This mapping resembles the one presented in [34], only we rely on CLIP score to label our data, providing a single score rather than multiple scores obtained by crowd sourcing, and we do not assume a linear relationship between the word descriptors and the 3DMM's coefficients vector. Instead, we train a multi-layer-perceptron with ReLU activations as the mapping function using our paired data and a simple $L_{2}$ loss $\mathcal{L} = ||\hat{\vec{\xi}} -\vec{\xi} ||_2$ . The network consists of hidden layers (500 and 800 neurons), and is trained for 50 epochs. + +# 4. Applications + +# 4.1. Interactive Sliders + +Using the set of semantic descriptors $\ell$ chosen by our algorithm, we can build a simple and intuitive interface to control the 3DMM. This is because the descriptors are least entangled and cover the PCA space well. First, we train a mapper $M$ with the chosen set of descriptors $\ell$ as described above (Section 3.4). Next, we create a slider for each descriptor whose value represents the desired CLIP score for that specific descriptor (see Figure 4). The user can now change each slider to trigger a re-prediction of the 3DMM coefficients. The computation is interactive as the interface takes all sliders' values as input $\vec{\Omega}$ to $M$ , which maps them in a forward pass to the 3DMM coefficients vector $\vec{\xi}$ and renders the corresponding 3D mesh. Thus, the modeler can interact and understand the relationships between the descriptors and their effect on the mesh. Figure 3 demonstrates examples of 3D meshes of different animals that were created using an interface created for the SMAL model. + +# 4.2. Zero-Shot Image to Shape Reconstruction + +Our method allows to leverage CLIP's semantic understanding to define a zero-shot image to shape reconstruction. Given any image of a person, we embed it into CLIP's latent space and use the similarity scores of the image against the set of descriptors $\ell^{M}$ chosen for a mapper $M$ to obtain the vector $\vec{\Omega}$ . Next, we simply feed $\vec{\Omega}$ through the network to get the 3DMM coefficients vector $\vec{\xi}$ that best fits the image and create the shape from $\vec{\xi}$ . Our method differs from state-of-the-art methods in that it enables the user to refine the zero-shot prediction very simply using semantic sliders starting from a relatively good initial guess. + +# 5. Experiments + +# 5.1. Coverage Evaluation + +To evaluate a mapper $M$ that was trained with a given set of descriptors $\ell$ , we would like to measure its effect on the 3DMM. For each descriptor $\ell_i \in \ell$ we do a forward pass through the mapper twice: the first pass by setting a low score to $\Omega_i$ and the second pass by setting a high score to $\Omega_i$ . The other coefficients are set to a default value and remain constant. Each such pass creates a parameter vector $\vec{\xi} = [\xi_1, \dots, \xi_{10}]$ , which represents the resulting 3DMM shape or expression. We denote these vectors as $\vec{\xi}(i)_{low}$ and $\vec{\xi}(i)_{high}$ . + +We measure the geometric effect of a descriptor $\ell_{i}$ on a 3DMM by examining the deformation of the vertices of the mesh. This is evaluated by comparing the position of each vertex in the two extreme cases of $\vec{\xi} (i)_{low}$ and $\vec{\xi} (i)_{high}$ . Hence, for each vertex $v$ the size of deformation is de + +![](images/ba2e8233a3463231635a82b4dd5861c7c0bad839b00885c2da23b673b0f54061.jpg) + +![](images/b446a98b47353fb1fb61a099791f9d52189f37b7ae78b972d87dc3e14dc131f5.jpg) + +![](images/0aa37f85af56025902920a1d234eb5600e573a056737d354461c241c7dafc3e6.jpg) +Figure 4: A simple and intuitive interactive application can be defined using sliders for each semantic descriptor. With Semantiy, the sliders have strong semantic meaning and are disentangled so that changing one will have very limited affect on others. + +![](images/b810a44ef572a968a50973da2d71ea70decc818848d9e8d0a61f80ea36d12521.jpg) + +fined as $\delta'(v) = ||v_{low} - v_{high}||_2$ . We can normalize these values by $\delta_{\max} = \max_{v \in Mesh} \delta'(v)$ and get a value $\delta(v) = \delta'(v) / \delta_{\max}$ between 0 and 1 (see Figure 5). + +![](images/2b25c53b186ea56238c5157ab4294f6b64e2ccfbab03d23e9831276848c7345a.jpg) +Figure 5: The geometric effect of a given descriptor. The darker the color of the vertex is (more red), the more it is affected by the descriptor. To define coverage we collect all vertices with a value larger than a given threshold $\theta$ . + +![](images/b86ca843a083907af627e79beb5fed8dc8b91d41a4d60b9eb2b0ad6a0f9137e9.jpg) + +Using these differences, we can evaluate the coverage of a specific descriptor by measuring the number and positions of vertices it affects in the mesh. For example, we can look at all vertices $v$ whose value is above a threshold $\delta(v) > \theta$ for any given descriptor. We can also compare the overlap between two descriptors by measuring the intersection-over-union of their covered vertices. We use these measure + +in our ablation studies as described next in Section 5.2. Note that this measure is only an approximation as it measures the size of the change of vertices and not their direction. + +# 5.2. Ablations + +To further evaluate our algorithm's performance, we conducted experiments that include training the mapper with different numbers of descriptors and training on different sizes of training data. As we do not have valid ground truth to estimate the accuracy of our mapper we created a test set that contains several body shapes that were generated by the first 10 blendshapes $\vec{\xi}$ of the SMPL-X model. In each experiment we optimize the descriptors' values $\vec{\Omega}$ , to minimize the error $||\hat{\vec{\xi}} - \vec{\xi}||_2$ . The goal of these experiments is to test the expressive power of the given configuration to match a given shape. The smaller the error the more expressive the method is. We show the results of these ablations in the supplemental file. We used this evaluation to choose the hyper-parameters we use in our algorithm: training with 3K sample images with texture for all models. + +# 5.3. Interactive Sliders User Study + +To evaluate Semantify's performance we conducted a user study that compares our approach against BodyTalk [34], as well as the original basis of the 3DMM. The user is presented with a target 3D model (randomly created using the 3DMM) and is tasked to fit a 3DMM mesh model to the target shape using a set of sliders, using the two compared interfaces (see Figure 4) at random order. Users were given up to 5 minutes to complete the task. BodyTalk was implemented on the SMPL body model, therefore we compare it to our SMPL male and female models. In another study, we compare our SMPLX male and female models to the original PCA-based axes. In both studies, we seek to evaluate the accuracy of the user-fitted 3DMMs, the time that it took to fit the model, as well as the overall subjective experience of users. Since we cannot obtain the outputs of BodyTalk in the first study, we asked humans to rate the results, and determine which of them fits more accurately to the input shape. We had 10 users perform both studies (5 males, 5 females aged between 23 to 54), 7 of which are novices and 3 of which were professional 3D modelers. We present the quantitative results of these studies in Table 2. As can be seen, Semantify achieves better accuracy in less time. Note that theoretically only the baseline can reach zero error, but because it is very difficult to handle, the error obtained was higher than Semantify. In addition, we asked the users for feedback on the experience of using the different applications (A was ours, B was the alternative). Here are some examples of such reviews (more examples could be found in the supplemental file): + +"Using application B, occasionally when changing one + +slider it affects the other, which causes the user to start all over again". + +"Using application B every minor change in a certain slider generated major changes in the rest of the sliders" + +"Application A was far friendlier and I sensed as though I maintained much more control over the different body features" + +"The abundance of sliders on application B only made it harder to control, not the other way around". + +# 5.4. Zero Shot Reconstruction + +To measure our zero-shot 3D-shape reconstruction method of SMPLX model, we use HBW (Human Bodies in-the-wild) dataset presented in SHAPY [7], containing ground-truth 3D body scans with in-the-wild images. Table 3 shows the performance of our method compared to SHAPY and PixIE [12] on all of HBW validation set. Note that SHAPY also takes keypoints (usually obtained from openPose [5]) as input. Some examples are shown in Figure 6. As can be seen, our zero-shot performance are on-par with state of the art methods, but we have the advantage of using semantic descriptors. This allows the user to fine tune the results from a very good initial guess, simply by using the semantic sliders. Some examples results with around a minute of fine-tuning can be found in the supplemental file. + +# 6. Conclusion + +We have presented a method to edit a 3D parametric model using semantically meaningful descriptors by first choosing a subset of descriptors judiciously and then learning a non-linear mapping from this set to the 3DMM coefficients. All this is done in a self-supervised manner harnessing the abilities of CLIP visual-language model. Our method can support a more intuitive user interface that allows even novices to model 3D shapes. It also supports other applications such as zero-shot reconstruction and fine-tuning of shapes from images. + +Limitations and future directions. There are several limitations to our method. First, our mapper's performance relies heavily on the data that is created (Section 3.1). For example, a dataset must contain extreme samples to achieve a more expressive mapper. Second, our goal with Semantify was to build a semantic representation of a given 3DMM without a human-in-the-loop. However, in some cases manually tuned modeling for a specific 3DMM would probably improve the 3DMM mapper's performance (we further elaborate on this in the supplemental file). Lastly, as shown in many other works, CLIP has rich semantic understanding, but there are cases in which its performance degrades. For example, performance is better with textured meshes, so using various textures for each 3DMM may improve the + +![](images/496f0ccadf6c7582fd65d2e6abfef15bd4e73c4e6671a91bfe46eda410bcd999.jpg) +Figure 6: Results of zero-shot image to shape reconstruction. We compare Semantify against SHAPY and PixIE on SMPLX 3DMM. Colors indicate error compared to ground truth (darker means larger error). + +
Time (min)Score
MalesFemalesTotal
BodyTalk3.223.163.190.04%
ours2.121.451.590.96%
+ +
Mean Error (cm)Time (min)
MalesFemalesTotalMalesFemalesTotal
Baseline0.1220.1110.1163.182.092.43
ours0.1210.0990.1102.222.172.19
+ +Table 2: User-study results for the SMPL 3DMM compared against BodyTalk [34] (left, Score means the percent that this model was selected as better fitting the input), and SMPL-X parametric model compared against the raw blendshapes (right). + +
MalesFemalesTotal
SHAPY0.01969340.0144220.017085
PIXIE0.0247790.0152650.020324
Semantify (ours)0.0226320.0165640.019666
+ +Table 3: Comparing our zero-shot shape reconstruction from image method against SHAPY[7] and PixIE [12]. The values demonstrate the average MSE over HBW's in-the-wild validation set. + +mapper's performance. Furthermore, more subjective descriptors such as "cute", "scary", "pretty" etc., are harder to + +rate using CLIP, and would probably not work well in our model. + +Future work can also improve the zero-shot 3D shape reconstruction by using prior information about its context (for example, the gender of a person, body keypoints as used in [7] etc.). Adding such priors along with CLIP's semantic understanding may surpass the existing state-of-the-art solutions. + +# 7. Acknowledgments + +This research was supported by the Israel Science Foundation (grant No. 1390/19). We would also like to thank Eyal Gomel for valuable discussions. + +# References + +[1] Rameen Abdal, Peihao Zhu, John Femiani, Niloy J. Mitra, and Peter Wonka. Clip2stylegan: Unsupervised extraction of stylegan edit directions. CoRR, abs/2112.05219, 2021. +[2] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Clipface: Text-guided editing of textured 3d morphable models. arXiv preprint arXiv:2212.01406, 2022. +[3] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. Scape: Shape completion and animation of people. ACM Transactions on Graphics (TOG), 24:408-416, 07 2005. +[4] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999. +[5] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43:172–186, 2018. +[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations, 2020. +[7] Vasileios Choutas, Lea Müller, Chun-Hao P. Huang, Siyu Tang, Dimitrios Tzionas, and Michael J. Black. Accurate 3d body shape regression using metric and semantic attributes. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. +[8] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Monocular expressive body regression through body-driven attention. In European Conference on Computer Vision (ECCV), 2020. +[9] Radek Daněček, Michael J. Black, and Timo Bolkart. Emotion driven monocular face capture and animation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20311-20322, June 2022. +[10] Ginger Delmas, Philippe Weinzaepfel, Thomas Lucas, Francesc Moreno-Noguer, and Grégory Rogez. Postscript: 3d human poses from natural language. In European Conference on Computer Vision, pages 346-362. Springer, 2022. +[11] Bernhard Egger, William AP Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski, + +Sami Romdhani, et al. 3d morphable face models—past, present, and future. ACM Transactions on Graphics (TOG), 39(5):1-38, 2020. +[12] Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael Black. Collaborative regression of expressive bodies using moderation. In International Conference on 3D Vision (3DV), pages 792-804, Dec. 2021. +[13] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8649-8658, June 2021. +[14] Thomas Gereg, Andreas Morel-Forster, Clemens Blumer, Bernhard Egger, Marcel Luthi, Sandro Schoenborn, and Thomas Vetter. Morphable face models - an open framework. 2018 13th IEEE International Conference on Automatic Face amp; Gesture Recognition (FG 2018), May 2018. +[15] Matthew Hill, Stephan Streuber, Carina Hahn, Michael Black, and Alice O'Toole. Exploring the relationship between body shapes and descriptions by linking similarity spaces. Journal of Vision, 15(12):931-931, 2015. +[16] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhonggang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. +[17] IEEE. A 3D Face Model for Pose and Illumination Invariant Face Recognition, Genova, Italy, 2009. +[18] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, P. Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 857-866, 2021. +[19] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), 2018. +[20] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. SIGGRAPH Asia 2022 Conference Papers, 2022. +[21] Youwang Kim, Ji-Yeon Kim, and Tae-Hyun Oh. Clip-actor: Text-driven recommendation and stylization for animating human meshes. In European Conference on Computer Vision, 2022. +[22] Muhammed Kocabas, Chun-Hao P. Huang, Joachim Tesch, Lea Müller, Otmar Hilliges, and Michael J. + +Black. SPEC: Seeing people in the wild with an estimated camera. In Proc. International Conference on Computer Vision (ICCV), pages 11035-11045, Oct. 2021. +[23] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, 2019. +[24] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. +[25] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. *Smpl: A skinned multi-person linear model*. ACM Trans. Graph., 34(6), nov 2015. +[26] Xudong Lou, Yiguang Liu, and Xuwei Li. Tecm-clip: Text-based controllable multi-attribute face image manipulation. In Proceedings of the Asian Conference on Computer Vision, pages 1942–1958, 2022. +[27] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. +[28] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. ArXiv, abs/2212.08751, 2022. +[29] Ahmed A. A. Osman, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Supr: A sparse unified part-based human representation. In Shai Avidan, Gabriel Brostow, Moustapha Cissé, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision – ECCV 2022, pages 568–585, Cham, 2022. Springer Nature Switzerland. +[30] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. +[31] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, + +et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. +[33] Hyewon Seo, Frederic Cordier, and Nadia Thalmann. Synthesizing animatable body models with parameterized shape modifications. 07 2003. +[34] Stephan Streuber, M. Alejandra Quiros-Ramirez, Matthew Q. Hill, Carina A. Hahn, Silvia Zuffi, Alice O'Toole, and Michael J. Black. Body Talk: Crowd-shaping realistic 3D avatars with words. ACM Trans. Graph. (Proc. SIGGRAPH), 35(4), July 2016. +[35] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022. +[36] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. +[37] Ayush Tewari, Michael Zollhöfer, Pablo Garrido, Florian Bernard, Hyeongwoo Kim, Patrick Pérez, and Christian Theobalt. Self-supervised multi-level face model learning for monocular reconstruction at over 250 hz. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. +[38] Can Wang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2022. +[39] Can Wang, Ruixia Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. ArXiv, abs/2212.08070, 2022. +[40] Jinbao Wang, Shujie Tan, Xiantong Zhen, Shuo Xu, Feng Zheng, Zhenyu He, and Ling Shao. Deep 3d human pose estimation: A review. Computer Vision and Image Understanding, 210:103225, 2021. +[41] Lizhen Wang, Zhiyua Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR2022), June 2022. +[42] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Ghum & ghuml: Generative 3d human shape and articulated pose models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6184-6193, 2020. + +[43] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. *Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction.* In *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition* (CVPR), 2020. +[44] Silvia Zuffi, Angjoo Kanazawa, and Michael J. Black. Lions and tigers and bears: Capturing non-rigid, 3d, articulated shape from images. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3955-3963, 2018. +[45] Silvia Zuffi, Angjoo Kanazawa, David Jacobs, and Michael J. Black. 3D menagerie: Modeling the 3D shape and pose of animals. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), July 2017. \ No newline at end of file diff --git a/2023/Semantify_ Simplifying the Control of 3D Morphable Models Using CLIP/images.zip b/2023/Semantify_ Simplifying the Control of 3D Morphable Models Using CLIP/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1e0701b84848ea6267af5df5524d9599df153f71 --- /dev/null +++ b/2023/Semantify_ Simplifying the Control of 3D Morphable Models Using CLIP/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26cb9334b2bfa96bdc0a4b82a50b917be81b4724a63fc266f16175faf79029fc +size 354646 diff --git a/2023/Semantify_ Simplifying the Control of 3D Morphable Models Using CLIP/layout.json b/2023/Semantify_ Simplifying the Control of 3D Morphable Models Using CLIP/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..750271ebc7bdb7c60f5d9c7e90eac592d5f37409 --- /dev/null +++ b/2023/Semantify_ Simplifying the Control of 3D Morphable Models Using CLIP/layout.json @@ -0,0 +1,7396 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 103, + 523, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 103, + 523, + 122 + ], + "spans": [ + { + "bbox": [ + 70, + 103, + 523, + 122 + ], + "type": "text", + "content": "Semantify: Simplifying the Control of 3D Morphable Models using CLIP" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 97, + 141, + 470, + 351 + ], + "blocks": [ + { + "bbox": [ + 97, + 141, + 470, + 351 + ], + "lines": [ + { + "bbox": [ + 97, + 141, + 470, + 351 + ], + "spans": [ + { + "bbox": [ + 97, + 141, + 470, + 351 + ], + "type": "image", + "image_path": "473800955cb810138a380f9346d36cad6a30af74e50d3e7fbaf71ad28b51722a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 365, + 548, + 425 + ], + "lines": [ + { + "bbox": [ + 45, + 365, + 548, + 425 + ], + "spans": [ + { + "bbox": [ + 45, + 365, + 548, + 425 + ], + "type": "text", + "content": "Figure 1: Semantify offers a method to create and edit a 3D parametric model using semantically meaningful descriptors. Semantify is based on a self-supervised method that utilizes the semantic power of CLIP language-vision model to build a mapping between semantic descriptors to 3DMM model coefficients. This can be used in an interactive application defining a slider for each descriptor (a), or to fit a model to an image in a zero shot manner by feeding the image into CLIP and obtaining a vector of semantic scores that can be mapped to shape parameters (b)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 435, + 192, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 435, + 192, + 447 + ], + "spans": [ + { + "bbox": [ + 143, + 435, + 192, + 447 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 460, + 290, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 290, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 290, + 700 + ], + "type": "text", + "content": "We present Semantify: a self-supervised method that utilizes the semantic power of CLIP language-vision foundation model [32] to simplify the control of 3D morphable models. Given a parametric model, training data is created by randomly sampling the model's parameters, creating various shapes and rendering them. The similarity between the output images and a set of word descriptors is calculated in CLIP's latent space. Our key idea is first to choose a small set of semantically meaningful and disentangled descriptors that characterize the 3DMM, and then learn a non-linear mapping from scores across this set to the parametric coefficients of the given 3DMM. The nonlinear mapping is defined by training a neural network without a human-in-the-loop. We present results on numerous 3DMMs: body shape models, face shape and expression models, as well as animal shapes. We demonstrate how our method defines a simple slider interface for intuitive modeling, and show how the mapping can be used to instantly fit a 3D parametric body shape to in-the-wild images. See our project page at https://omergral.github.io/Semantify/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 435, + 387, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 435, + 387, + 447 + ], + "spans": [ + { + "bbox": [ + 307, + 435, + 387, + 447 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 458, + 547, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 547, + 662 + ], + "type": "text", + "content": "3D modeling techniques have evolved tremendously over the last few years. Such techniques are used in countless industries including the burgeoning AR/VR industry, fashion design, game development, film, and many others. However, designing a high-quality 3D model is not a simple task for most people and might require a well-trained 3D artist. Even when using more recent parametric morphable models (3DMM) [24, 31, 45, 3, 42] it is still hard for humans to understand how to choose the correct set of parameters, e.g. to achieve a specific human body shape or facial expression. It is also difficult to find what are the limits of the given parametric model in terms of coverage and expressiveness. The reason is that in most cases, the provided set of parameters is not interpretable, as they are commonly calculated using automatic optimization mechanisms followed by dimensionality reduction using PCA. Thus, they carry no clear semantic meaning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "To simplify the use of 3DMMs and allow for natural interactive human modeling, a key research question is - how to insert semantics into 3DMM control? Previous approaches [34] relied mostly on human intelligence and la" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14554" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "beling, which is often time consuming and expensive. In addition, previous methods create a large number of semantic control descriptors that are often correlated and entangled. This means it is difficult to anticipate the effect of each descriptor as changing a value in one may modify others, resulting in difficulties to control the model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 146, + 288, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 146, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 288, + 326 + ], + "type": "text", + "content": "In this paper, instead of using human labeling, we rely on the remarkable abilities of huge foundation models that combine natural language and visual understanding. We present a self-supervised method that utilizes the semantic power of CLIP [32] to define a method to control 3DMMs that carries two main advantages. First, it is more natural for humans to use as it allows modeling using a small number of semantically meaningful descriptors, that cover the space of deformations but are disentangled. Second, it covers even extreme examples in the shape/pose space of parametric models as it utilizes CLIP's pre-training on a large number of images. The main idea is first to use CLIP to select a small subset of semantically meaningful and disentangled descriptors, and then learn a non-linear mapping of this set to the coefficients of the given 3DMM." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 327, + 288, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 288, + 530 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 288, + 530 + ], + "type": "text", + "content": "Given a parametric 3DMM, we sample its parameter space to create a dataset containing a variety of 3D mesh shapes. We then render each mesh from different camera views to create a diverse set of images corresponding to the parameter samples (see Figure 2). Next, we gather a set of semantically descriptive textual terms related to the parametric 3D model, which we call descriptors. We encode both the images (using CLIP's image encoder) and the descriptors (using CLIP's text encoder) into CLIP's latent space and compare them. This defines a vector of similarity scores between the input vector of 3DMM coefficients (of each image sample) and each corresponding semantic descriptor. Next, we define a selection scheme to choose a small number of descriptors that are de-correlated to control the model. Lastly, we train a neural network to learn a mapping between the vector of similarity scores to the vector of 3DMM parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 532, + 288, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 288, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 288, + 675 + ], + "type": "text", + "content": "We demonstrate how the learned mapping can be used to define an interface to control a 3DMM in a way that is simple and effective using a small set of semantically meaningful sliders (see Figure 1 and 4). Such sliders are easy to employ for designing high-quality 3D models and cover the shape space well. We demonstrate this for four parametric models: human face's shape and expression (FLAME [24]), human body shapes (SMPL [25] and SMPL-X [31]), and even animals (SMAL [45]). We also show how the mapping can be used to instantly fit a 3D parametric body shape to an input image that works well \"in the wild\" even in extreme poses and body shapes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "Our main contribution is a novel, self-supervised method for defining a small set of semantic descriptors to control a parametric model more naturally, by learning a mapping" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "from a semantic space to a parametric representation without human-in-the-loop. We demonstrate the effectiveness of our approach on several parametric models and utilize it to define a simple interface for modeling and to instantly fit a 3D model to images in the wild. We will release our code for future research." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 156, + 392, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 156, + 392, + 168 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 392, + 168 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 176, + 545, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 176, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 304, + 176, + 545, + 212 + ], + "type": "text", + "content": "Our work encompasses a wide range of disciplines, including 3D parametric modeling, language-vision models, and zero-shot reconstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 228, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 324 + ], + "type": "text", + "content": "3D Morphable Models. Much research has gone into improving 3D modeling of humans bodies, faces and other object classes, such as animals [11, 4, 25, 31]. 3D Morphable Models (3DMM) are a powerful tool to parameterize the variation in geometry of objects belonging to a certain class. Dating back to 1999, Blantz et al. [4] proposed 3DMMs to capture the variation of human faces and applied dimensionality reduction using Principal Component Analysis (PCA)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 325, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 545, + 445 + ], + "type": "text", + "content": "In contrast to a 3D artist freely controlling the vertices of a mesh, using each of these PCA axes to deform the vertices induces a strong statistical prior on the resulting geometry. The resulting deformations are global in nature, strongly constraining the geometry to the manifold of statistically plausible meshes, as the per-axis coefficients correlate with their likelihood. While initially proposed for human faces, statistical 3DMMs have been developed for human body models [3, 25, 40, 31, 42, 29], facial models [24, 37, 41, 43], and even 3D animal models [45, 44]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 461, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 593 + ], + "type": "text", + "content": "Semantic Models. PCA-based 3DMMs are efficiently and easily constructed, and admit an orthogonal and importance-ordered basis for the deformation. One disadvantage of these models is their lack of interpretability for humans. The axes of maximal variation in geometry, which form the deformation basis for each vertex, do not necessarily correspond to any semantically meaningful geometric change (e.g., raising an eyebrow), but instead, each axis induces a rather global change that may introduce correlations or effects that are unwanted to a human modeler, making it hard to obtain the desired shape or expression." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "content": "Some 3DMMs [17, 14] alleviate this issue by using deformation bases that are hand-crafted by artists, often referred to as blendshapes. While this alleviates the interpretability issue, it loses the data-driven and explicit statistical prior, requires manual work, and does not guarantee the expressiveness of the resulting model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "First to address the semantic issue of PCA-based 3DMMs in the human body domain were Seo et al. [33] which used metrics such as hip-to-waist ratio, body fat percentage, and height to regress 3DMM shape parame" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14555" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 545, + 186 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 186 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 186 + ], + "type": "image", + "image_path": "66f4710b2aea4be682757d90d68a3adcd51df36fb3a70871158233b62a01ae87.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 197, + 548, + 258 + ], + "lines": [ + { + "bbox": [ + 45, + 197, + 548, + 258 + ], + "spans": [ + { + "bbox": [ + 45, + 197, + 548, + 258 + ], + "type": "text", + "content": "Figure 2: Learning a mapping from Semantic to Parametric space. (a) Given a coefficient vector " + }, + { + "bbox": [ + 45, + 197, + 548, + 258 + ], + "type": "inline_equation", + "content": "\\vec{\\xi}" + }, + { + "bbox": [ + 45, + 197, + 548, + 258 + ], + "type": "text", + "content": " we create the 3DMM mesh. The mesh is rendered from several views. Each Rendered image " + }, + { + "bbox": [ + 45, + 197, + 548, + 258 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 45, + 197, + 548, + 258 + ], + "type": "text", + "content": " is passed into CLIP along with a set of semantic descriptors " + }, + { + "bbox": [ + 45, + 197, + 548, + 258 + ], + "type": "inline_equation", + "content": "\\vec{\\ell}" + }, + { + "bbox": [ + 45, + 197, + 548, + 258 + ], + "type": "text", + "content": ". The difference between each descriptor and the image in CLIP latent space is calculated and stored in the corresponding entry of the similarity vector " + }, + { + "bbox": [ + 45, + 197, + 548, + 258 + ], + "type": "inline_equation", + "content": "\\vec{\\Omega}" + }, + { + "bbox": [ + 45, + 197, + 548, + 258 + ], + "type": "text", + "content": ". (b) Using a large set of such random pairs of " + }, + { + "bbox": [ + 45, + 197, + 548, + 258 + ], + "type": "inline_equation", + "content": "(\\vec{\\Omega},\\vec{\\xi})" + }, + { + "bbox": [ + 45, + 197, + 548, + 258 + ], + "type": "text", + "content": ", we train a network to learn the mapping from semantic space to parametric space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 278, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 289, + 495 + ], + "type": "text", + "content": "ters. Hill et al.[15] suggested examining the relationship between body shapes and description, by linking two similarity spaces - one created from body descriptions and the other from full-body laser scans. BodyTalk [34] followed suit and approached this problem using the \"wisdom of the crowd\". In their work, they captured 256 male and female body shapes and represented each shape with 30 body descriptor words (e.g. curvy, fit, heavyet, round-Apple, etc.). They hired labelers to discretely rate the bodies with respect to these descriptors and designed a regression model to learn the relationship between the discrete ratings and the parameters of the corresponding body shape. In a more recent publication, SHAPY [7] bridges this gap by curating datasets of images with corresponding body measurements (e.g., from model agencies) followed by human ratings respective to a set of linguistic shape attributes. They train models to predict SMPLX shape parameters from attributes and/or body measurements, and vice-versa." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": "Language-Guided 3D Modeling. Recent works have shown great success in learning latent representations that are capable of coupling visual signals with language. One such example is CLIP [32] model. CLIP uses a contrastive training scheme [6] on over 400 million pairs of images and their captions to build a shared latent representation for visual-textual content. In this space, similarity between images and text can be computed. It has been shown that CLIP is an effective tool for image generation tasks. For example, providing natural-language guided manipulation of human face imagery in StyleGAN space [30, 26, 1]. Similarly, CLIP can be employed in 3D: using differentiable rendering, images of a 3D scene are rendered in a forward pass and are scored against a text prompt with CLIP. Taking the similarity score as an optimization objective, gradients are back-propagated through the CLIP network and the rendering process, back to the underlying representation of the 3D scene. Text2Mesh [27] uses CLIP to optimize for col" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 278, + 547, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 547, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 547, + 363 + ], + "type": "text", + "content": "ors and positions of mesh vertices to match a text query. CLIP-Mesh [20] followed suit and proposed to use CLIP to guide normal and texture maps. This basic recipe has been applied to a variety of scene representations, such as point clouds [28] and Neural Fields [38, 18, 39, 13]. In the context of human faces and bodies, CLIP is used for generation of animation sequences [21, 36, 16, 35, 10, 2]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 368, + 547, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 368, + 547, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 547, + 561 + ], + "type": "text", + "content": "Zero-Shot 3DMM Shape Reconstruction. Recovering accurate, explicit meshes from 2D signals such as images or video is an under-constrained and over-parameterized objective. The low-dimensional underlying representation of 3DMMs proves useful not only for modeling, but for 3D reconstruction of geometry from images of humans. The parameters can serve as a data-driven regularization term for shape and pose, and the low dimensional representation convexifies the optimization problem of fitting such a model to an image, rather than a freely deforming mesh. Recent 3DMM-from-image methods either use iterative optimization schemes to fit the parameters of the 3DMM to the image [23] or directly regress the shape and pose parameters from an Image [12, 8, 22, 19]. Our method allows to use CLIP image and text mapping and simply feed-forward the scores through an MLP network to get a body shape." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 572, + 362, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 572, + 362, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 572, + 362, + 586 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "An overview of our method can be seen in Figure 2. In essence, our method defines a non-linear mapping from semantic space to parametric space by training a neural network to predict 3DMM coefficients from a vector of semantic attribute scores. We first create a dataset of rendered images of randomly sampled shapes (3.1), gather their corresponding semantic scores (3.2). Then, we employ a scheme to reduce the semantic descriptors to a subset of least-correlated descriptors (3.3) and then train the network (3.4)." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14556" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 533, + 196 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 533, + 196 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 533, + 196 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 533, + 196 + ], + "type": "table", + "html": "
Coverage / Overlap
MethodAlgorithm Choice2 descriptors5 descriptors10 descriptors15 descriptors
SMPLX-male(6) / 92.1% / 64.5%38.6% / 30.1%51.5% / 71.4%49.7% / 83.5%98.3% / 80.5%
SMPLX-female(4) / 49.7% / 56.7%27.1% / 49.5%50.0% / 67.1%47.4% / 83.0%82.4% / 83.4%
SMPLX-neutral(8) / 52.0% / 83.0%40.2% / 51.2%50.0% / 67.1%51.9% / 86.9%52.1% / 90.3%
SMPL-male(4) / 97.5% / 60.2%40.1% / 46.7%97.7% / 66.5%98.3% / 83.9%98.2% / 89.3%
SMPL-female(5) / 95.7% / 68.6%36.3% / 47.3%95.7% / 68.6%97.8% / 85.2%99.1% / 89.9%
SMPL-neutral(6) / 96.1% / 76.0%80.4% / 48.7%96.2% / 71.5%99.6% / 85.2%99.8% / 90.8%
FLAME-expression(3) / 13.4% / 56.8%10.4% / 47.4%19.3% / 67.2%27.0% / 77.3%34.4% / 85.0%
", + "image_path": "0434f26d053aff10504472a4a25123f0221de16f6d76b69550466fb7eace7c32.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 547, + 253 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 547, + 253 + ], + "type": "text", + "content": "Table 1: Choosing different number of descriptors for a model. We evaluate the coverage and overlap of vertices for a mapper that was trained by using different number of descriptors (see Section 5.1, the threshold used to measure coverage is 0.3). The number chosen by our algorithm is shown in parentheses at the beginning of each row. As can be seen, the more descriptors are chosen the larger the cover is, but the larger the overlap between the descriptors is as well." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 273, + 149, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 273, + 149, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 273, + 149, + 285 + ], + "type": "text", + "content": "3.1. Dataset Creation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 292, + 288, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 288, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 288, + 471 + ], + "type": "text", + "content": "The process of creating the data for training our model is similar across all 3DMMs (we demonstrate it on four different ones in this paper). For each model, we only use the first 10 principal components, that cover above " + }, + { + "bbox": [ + 46, + 292, + 288, + 471 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 292, + 288, + 471 + ], + "type": "text", + "content": " of the variance. We draw " + }, + { + "bbox": [ + 46, + 292, + 288, + 471 + ], + "type": "inline_equation", + "content": "N_{\\text{samples}}" + }, + { + "bbox": [ + 46, + 292, + 288, + 471 + ], + "type": "text", + "content": " 10-dimensional random vectors of coefficients. For body models, we draw a random shape coefficients vector " + }, + { + "bbox": [ + 46, + 292, + 288, + 471 + ], + "type": "inline_equation", + "content": "\\vec{\\beta}" + }, + { + "bbox": [ + 46, + 292, + 288, + 471 + ], + "type": "text", + "content": ", which yields a random shape. From the expression basis of the FLAME model, we randomly sample an expression coefficient vector " + }, + { + "bbox": [ + 46, + 292, + 288, + 471 + ], + "type": "inline_equation", + "content": "\\vec{\\psi}" + }, + { + "bbox": [ + 46, + 292, + 288, + 471 + ], + "type": "text", + "content": " which yields a random expression. Since random values may also lead to noisy and unrealistic outputs, we limit the values of the sampled coefficients to a range of " + }, + { + "bbox": [ + 46, + 292, + 288, + 471 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 292, + 288, + 471 + ], + "type": "text", + "content": " standard deviations of the model's coefficients (" + }, + { + "bbox": [ + 46, + 292, + 288, + 471 + ], + "type": "inline_equation", + "content": "k = 2" + }, + { + "bbox": [ + 46, + 292, + 288, + 471 + ], + "type": "text", + "content": " for body shapes and " + }, + { + "bbox": [ + 46, + 292, + 288, + 471 + ], + "type": "inline_equation", + "content": "k = 4" + }, + { + "bbox": [ + 46, + 292, + 288, + 471 + ], + "type": "text", + "content": " for face expressions), yielding more realistic results. The other parameters of the 3DMM " + }, + { + "bbox": [ + 46, + 292, + 288, + 471 + ], + "type": "inline_equation", + "content": "(\\theta, \\gamma)" + }, + { + "bbox": [ + 46, + 292, + 288, + 471 + ], + "type": "text", + "content": " remain neutral." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 472, + 288, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 288, + 664 + ], + "type": "text", + "content": "Defining the semantic representation begins by gathering an over-complete set " + }, + { + "bbox": [ + 46, + 472, + 288, + 664 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 46, + 472, + 288, + 664 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 472, + 288, + 664 + ], + "type": "inline_equation", + "content": "N_{\\ell}" + }, + { + "bbox": [ + 46, + 472, + 288, + 664 + ], + "type": "text", + "content": " word descriptors that correspond to each 3DMM. Our method supports any set of words that relate to the model: we couple the body model with body descriptors, and choose face descriptors for the face model. For the body models SMPL and SMPL-X, we adopt the set of descriptors used in [34], for FLAME face expression we use an expanded set based on [9], and for FLAME face shape we used our own set of descriptors. Lastly, since the SMAL model was trained on 5 animal families (Felidae-big cats, Canidae-dogs, Equidae-horses, Bovidae-cows, Hippopotamidae-hippos) we used a set of descriptors that contains animals from these given families. The full list of descriptors can be found in the supplemental file. We note that our method can be used with any desired set of relevant descriptors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 670, + 135, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 135, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 135, + 684 + ], + "type": "text", + "content": "3.2. CLIP Ratings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "We use CLIP [32] to encode each text descriptor " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\ell_{i}" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": " using their text encoder " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "e^{\\ell_i} = CLIP_{Text}(\\ell_i)" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": ". Each mesh created" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "type": "text", + "content": "from a random coefficient vector " + }, + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "type": "inline_equation", + "content": "F(\\vec{\\beta}, \\theta, \\gamma)" + }, + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "type": "text", + "content": " is rendered to create a set of images " + }, + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "type": "text", + "content": ", and these images " + }, + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "type": "text", + "content": " are encoded by CLIP image encoder " + }, + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "type": "inline_equation", + "content": "e^{I_j} = CLIP_{Image}(I_j)" + }, + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "type": "text", + "content": " to the same latent space of the encoded text (see Figure 2). In this space, we can compute the compatibility between the encoded image " + }, + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "type": "inline_equation", + "content": "e^{I_j}" + }, + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "type": "text", + "content": " and each encoded label " + }, + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "type": "inline_equation", + "content": "e^{\\ell_i}" + }, + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "type": "text", + "content": " using cosine similarity to get a score " + }, + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "type": "inline_equation", + "content": "\\Omega(I_j, \\ell_i) = \\cos(e^{I_j}, e^{\\ell_i})" + }, + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "type": "text", + "content": ". This way, each random coefficients vector " + }, + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "type": "inline_equation", + "content": "\\vec{\\beta}" + }, + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "type": "inline_equation", + "content": "\\vec{\\psi}" + }, + { + "bbox": [ + 304, + 273, + 547, + 384 + ], + "type": "text", + "content": " is paired with a vector of the descriptors' scores:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 320, + 390, + 531, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 390, + 531, + 406 + ], + "spans": [ + { + "bbox": [ + 320, + 390, + 531, + 406 + ], + "type": "interline_equation", + "content": "\\Omega_ {i = 1} ^ {N _ {\\ell}} = [ \\Omega (I _ {j}, \\ell_ {1}), \\dots , \\Omega (I _ {j}, \\ell_ {N _ {\\ell}}) ] = [ \\Omega_ {1}, \\dots , \\Omega_ {N _ {\\ell}} ]", + "image_path": "d1edbba4a606a678c1bd01447e13fd65b3ebe418e62d25dc820003c8d60d1317.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 413, + 395, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 413, + 395, + 425 + ], + "spans": [ + { + "bbox": [ + 306, + 413, + 395, + 425 + ], + "type": "text", + "content": "containing " + }, + { + "bbox": [ + 306, + 413, + 395, + 425 + ], + "type": "inline_equation", + "content": "N_{\\ell}" + }, + { + "bbox": [ + 306, + 413, + 395, + 425 + ], + "type": "text", + "content": " scores." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 432, + 427, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 432, + 427, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 432, + 427, + 445 + ], + "type": "text", + "content": "3.3. Descriptors Selection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "type": "text", + "content": "There are many words that could describe a body shape, face shape or facial expression. As the score of a given image and word descriptor in CLIP's embedding space is calculated by their semantic similarity, there might be many possible word descriptors for a single image, and many possible images for a single word descriptor. For example, both the words \"happy\" and \"smile\" would have a high correlation with a smiling face, and therefore, their effect on a facial expression is entangled. Similarly, \"tall\" and \"short\" are entangled in a body shape model. Our key observation is that a large number of descriptors are difficult to handle for interactive modeling, and the larger the number, the more entangled they are in terms of their effect on the shape - making it even harder to achieve the desired shape results (see Table 1)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "content": "Contrary to BodyTalk [34], our goal is to use a minimal set of semantic word descriptors, while covering as much of the PCA shape space of each 3DMM as possible. We present an algorithm for selecting a subset of semantic descriptors with two main objectives: they should cover a large part of the shape space, and they should be disentangled as much as possible." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14557" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 156 + ], + "type": "text", + "content": "Our idea is to first segment the shape space into regions, and assign descriptors to each region. This segmentation assures that we cover the shape space as we do not neglect any region. Next, we select for each region a set of descriptors by checking their variance (i.e., promoting coverage), and removing correlated descriptors, leaving the remaining ones as disentangled as possible." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 157, + 287, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 157, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 157, + 287, + 217 + ], + "type": "text", + "content": "Clustering. To segment the shape space, we use our " + }, + { + "bbox": [ + 47, + 157, + 287, + 217 + ], + "type": "inline_equation", + "content": "N_{\\text{samples}}" + }, + { + "bbox": [ + 47, + 157, + 287, + 217 + ], + "type": "text", + "content": " dataset as a proxy to represent a sampling of the space and cluster their images (created as described in Section 3.1) encoded to CLIP's latent space. We use K-means algorithm with Silhouette score to find " + }, + { + "bbox": [ + 47, + 157, + 287, + 217 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 157, + 287, + 217 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 218, + 287, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 218, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 47, + 218, + 287, + 302 + ], + "type": "text", + "content": "Next, we compare all encoded images " + }, + { + "bbox": [ + 47, + 218, + 287, + 302 + ], + "type": "inline_equation", + "content": "e^{I}" + }, + { + "bbox": [ + 47, + 218, + 287, + 302 + ], + "type": "text", + "content": " in each cluster " + }, + { + "bbox": [ + 47, + 218, + 287, + 302 + ], + "type": "inline_equation", + "content": "C_{i}" + }, + { + "bbox": [ + 47, + 218, + 287, + 302 + ], + "type": "text", + "content": " to all " + }, + { + "bbox": [ + 47, + 218, + 287, + 302 + ], + "type": "inline_equation", + "content": "N_{\\ell}" + }, + { + "bbox": [ + 47, + 218, + 287, + 302 + ], + "type": "text", + "content": " encoded word descriptors " + }, + { + "bbox": [ + 47, + 218, + 287, + 302 + ], + "type": "inline_equation", + "content": "e^{\\ell_i}" + }, + { + "bbox": [ + 47, + 218, + 287, + 302 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 218, + 287, + 302 + ], + "type": "inline_equation", + "content": "i = 1,\\dots ,N_{\\ell}" + }, + { + "bbox": [ + 47, + 218, + 287, + 302 + ], + "type": "text", + "content": ". The top 5 most similar descriptors in a cluster are given a vote for this cluster. We normalize the value of votes by the size of each cluster. Then, each word descriptor is assigned exclusively to the cluster with the largest sum of votes for this descriptor." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 304, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 304, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 287, + 471 + ], + "type": "text", + "content": "Choosing descriptors. To anticipate the effect of a given descriptor on the 3DMM, we use the variance of its score " + }, + { + "bbox": [ + 47, + 304, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\Omega(I', \\ell_i)" + }, + { + "bbox": [ + 47, + 304, + 287, + 471 + ], + "type": "text", + "content": " on all images " + }, + { + "bbox": [ + 47, + 304, + 287, + 471 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 47, + 304, + 287, + 471 + ], + "type": "text", + "content": " in the dataset. The larger the variance, the more descriptive this descriptor will be in terms of shape variation. To anticipate the level of entanglement between descriptors, we use their correlation. For each descriptor " + }, + { + "bbox": [ + 47, + 304, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\ell_i" + }, + { + "bbox": [ + 47, + 304, + 287, + 471 + ], + "type": "text", + "content": " we build a vector of " + }, + { + "bbox": [ + 47, + 304, + 287, + 471 + ], + "type": "inline_equation", + "content": "N_{\\text{samples}}" + }, + { + "bbox": [ + 47, + 304, + 287, + 471 + ], + "type": "text", + "content": " scores " + }, + { + "bbox": [ + 47, + 304, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\Omega(I', \\ell_i)" + }, + { + "bbox": [ + 47, + 304, + 287, + 471 + ], + "type": "text", + "content": " for all images " + }, + { + "bbox": [ + 47, + 304, + 287, + 471 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 47, + 304, + 287, + 471 + ], + "type": "text", + "content": " in the dataset. Correlation is measured between these vectors. The larger the correlation between two descriptors, the more similar the effect of this descriptor is on the shape, and the more entangled they are (we show some correlation plot examples in the supplemental file). Lastly, we apply antonyms and synonyms detection to verify there are no such pairs in our final set of descriptors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 473, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 473, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 473, + 287, + 628 + ], + "type": "text", + "content": "The process of choosing the descriptors is first performed on each cluster separately, and then merging the lists. We sort the cluster descriptors according to their variance in descending order. Thus, the first descriptor that is chosen from each cluster is the descriptor with the highest variance. Once a descriptor is chosen, we iterate over the other descriptors that are left for this cluster and filter out the ones that have a high correlation with the chosen descriptor (we use the median correlation value of all pairs as the threshold). In addition, before adding the next descriptor to the set, we check if its synonyms or antonyms are already chosen, and if so, we skip it. This process continues until the list of descriptors is exhausted." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "text", + "content": "Lastly, we merge the lists created from all clusters and sort again according to the descriptor variance. In a similar manner as we did for each cluster, we filter from the merged list all correlated descriptors, synonyms and antonyms. Finally, we arrive at a set of " + }, + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "text", + "content": " chosen semantic descriptors for this model. " + }, + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "text", + "content": " can be different for each model, but our algorithm can also support any preset number of descrip" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 69, + 552, + 222 + ], + "blocks": [ + { + "bbox": [ + 310, + 69, + 552, + 222 + ], + "lines": [ + { + "bbox": [ + 310, + 69, + 552, + 222 + ], + "spans": [ + { + "bbox": [ + 310, + 69, + 552, + 222 + ], + "type": "image", + "image_path": "2275e4d54f8c2c0a0b7a7c9673ad5499299f332252d0bf8304e84c3b08c3fc34.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 232, + 545, + 302 + ], + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 302 + ], + "type": "text", + "content": "Figure 3: Models created using an interactive slider application for SMAL[45] model (the blue model on the left is the neutral model). These animals were created by interpolating between four semantic descriptors - lion, cow, cat and dog. As can be seen, other animals could also be designed as well as intermixed creatures." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 342, + 545, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 377 + ], + "type": "text", + "content": "tors by removing descriptors from the bottom of the list or adding back descriptors that were filtered out according to their variance order (see Table 1)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 381, + 545, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 545, + 452 + ], + "type": "text", + "content": "Our algorithm can also support a preset list of descriptors if the user seeks to fit a 3D model with respect to some specific descriptors. In this case, the process described above for choosing the final set of descriptors is simply initialized with the user pre-defined descriptors, so that they are always contained in the final set." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 471, + 369, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 471, + 369, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 471, + 369, + 483 + ], + "type": "text", + "content": "3.4. Training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "type": "text", + "content": "For simplicity, we will denote the coefficients vector by " + }, + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\vec{\\xi}" + }, + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "type": "text", + "content": " for both shape and expression parameters. Our goal is to define the mapping from the semantic representation of the " + }, + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "type": "text", + "content": " descriptors to the parametric representation represented by " + }, + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\vec{\\xi}" + }, + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "type": "text", + "content": ". From our dataset creation (Section 3.1), each coefficients vector " + }, + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\vec{\\xi} = [\\xi_1,\\dots,\\xi_{10}]" + }, + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "type": "text", + "content": " is paired with a scores vector " + }, + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\vec{\\Omega} = [\\Omega_1,\\dots,\\Omega_{N_\\ell}]" + }, + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "type": "text", + "content": ". Hence, for a given scores vector " + }, + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\vec{\\Omega}" + }, + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "type": "text", + "content": ", the goal is to predict the corresponding coefficients vector " + }, + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\vec{\\xi}" + }, + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "type": "text", + "content": ". This mapping resembles the one presented in [34], only we rely on CLIP score to label our data, providing a single score rather than multiple scores obtained by crowd sourcing, and we do not assume a linear relationship between the word descriptors and the 3DMM's coefficients vector. Instead, we train a multi-layer-perceptron with ReLU activations as the mapping function using our paired data and a simple " + }, + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "type": "text", + "content": " loss " + }, + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = ||\\hat{\\vec{\\xi}} -\\vec{\\xi} ||_2" + }, + { + "bbox": [ + 304, + 493, + 545, + 713 + ], + "type": "text", + "content": ". The network consists of hidden layers (500 and 800 neurons), and is trained for 50 epochs." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14558" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 127, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 127, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 127, + 85 + ], + "type": "text", + "content": "4. Applications" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 155, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 155, + 102 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 155, + 102 + ], + "type": "text", + "content": "4.1. Interactive Sliders" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 109, + 289, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 289, + 324 + ], + "type": "text", + "content": "Using the set of semantic descriptors " + }, + { + "bbox": [ + 46, + 109, + 289, + 324 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 46, + 109, + 289, + 324 + ], + "type": "text", + "content": " chosen by our algorithm, we can build a simple and intuitive interface to control the 3DMM. This is because the descriptors are least entangled and cover the PCA space well. First, we train a mapper " + }, + { + "bbox": [ + 46, + 109, + 289, + 324 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 109, + 289, + 324 + ], + "type": "text", + "content": " with the chosen set of descriptors " + }, + { + "bbox": [ + 46, + 109, + 289, + 324 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 46, + 109, + 289, + 324 + ], + "type": "text", + "content": " as described above (Section 3.4). Next, we create a slider for each descriptor whose value represents the desired CLIP score for that specific descriptor (see Figure 4). The user can now change each slider to trigger a re-prediction of the 3DMM coefficients. The computation is interactive as the interface takes all sliders' values as input " + }, + { + "bbox": [ + 46, + 109, + 289, + 324 + ], + "type": "inline_equation", + "content": "\\vec{\\Omega}" + }, + { + "bbox": [ + 46, + 109, + 289, + 324 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 109, + 289, + 324 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 109, + 289, + 324 + ], + "type": "text", + "content": ", which maps them in a forward pass to the 3DMM coefficients vector " + }, + { + "bbox": [ + 46, + 109, + 289, + 324 + ], + "type": "inline_equation", + "content": "\\vec{\\xi}" + }, + { + "bbox": [ + 46, + 109, + 289, + 324 + ], + "type": "text", + "content": " and renders the corresponding 3D mesh. Thus, the modeler can interact and understand the relationships between the descriptors and their effect on the mesh. Figure 3 demonstrates examples of 3D meshes of different animals that were created using an interface created for the SMAL model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 332, + 266, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 332, + 266, + 345 + ], + "spans": [ + { + "bbox": [ + 47, + 332, + 266, + 345 + ], + "type": "text", + "content": "4.2. Zero-Shot Image to Shape Reconstruction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 350, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 483 + ], + "type": "text", + "content": "Our method allows to leverage CLIP's semantic understanding to define a zero-shot image to shape reconstruction. Given any image of a person, we embed it into CLIP's latent space and use the similarity scores of the image against the set of descriptors " + }, + { + "bbox": [ + 46, + 350, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\ell^{M}" + }, + { + "bbox": [ + 46, + 350, + 287, + 483 + ], + "type": "text", + "content": " chosen for a mapper " + }, + { + "bbox": [ + 46, + 350, + 287, + 483 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 350, + 287, + 483 + ], + "type": "text", + "content": " to obtain the vector " + }, + { + "bbox": [ + 46, + 350, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\vec{\\Omega}" + }, + { + "bbox": [ + 46, + 350, + 287, + 483 + ], + "type": "text", + "content": ". Next, we simply feed " + }, + { + "bbox": [ + 46, + 350, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\vec{\\Omega}" + }, + { + "bbox": [ + 46, + 350, + 287, + 483 + ], + "type": "text", + "content": " through the network to get the 3DMM coefficients vector " + }, + { + "bbox": [ + 46, + 350, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\vec{\\xi}" + }, + { + "bbox": [ + 46, + 350, + 287, + 483 + ], + "type": "text", + "content": " that best fits the image and create the shape from " + }, + { + "bbox": [ + 46, + 350, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\vec{\\xi}" + }, + { + "bbox": [ + 46, + 350, + 287, + 483 + ], + "type": "text", + "content": ". Our method differs from state-of-the-art methods in that it enables the user to refine the zero-shot prediction very simply using semantic sliders starting from a relatively good initial guess." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 493, + 128, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 493, + 128, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 493, + 128, + 506 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 513, + 167, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 167, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 167, + 525 + ], + "type": "text", + "content": "5.1. Coverage Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 532, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 654 + ], + "type": "text", + "content": "To evaluate a mapper " + }, + { + "bbox": [ + 46, + 532, + 287, + 654 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 532, + 287, + 654 + ], + "type": "text", + "content": " that was trained with a given set of descriptors " + }, + { + "bbox": [ + 46, + 532, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 46, + 532, + 287, + 654 + ], + "type": "text", + "content": ", we would like to measure its effect on the 3DMM. For each descriptor " + }, + { + "bbox": [ + 46, + 532, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\ell_i \\in \\ell" + }, + { + "bbox": [ + 46, + 532, + 287, + 654 + ], + "type": "text", + "content": " we do a forward pass through the mapper twice: the first pass by setting a low score to " + }, + { + "bbox": [ + 46, + 532, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\Omega_i" + }, + { + "bbox": [ + 46, + 532, + 287, + 654 + ], + "type": "text", + "content": " and the second pass by setting a high score to " + }, + { + "bbox": [ + 46, + 532, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\Omega_i" + }, + { + "bbox": [ + 46, + 532, + 287, + 654 + ], + "type": "text", + "content": ". The other coefficients are set to a default value and remain constant. Each such pass creates a parameter vector " + }, + { + "bbox": [ + 46, + 532, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\vec{\\xi} = [\\xi_1, \\dots, \\xi_{10}]" + }, + { + "bbox": [ + 46, + 532, + 287, + 654 + ], + "type": "text", + "content": ", which represents the resulting 3DMM shape or expression. We denote these vectors as " + }, + { + "bbox": [ + 46, + 532, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\vec{\\xi}(i)_{low}" + }, + { + "bbox": [ + 46, + 532, + 287, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 532, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\vec{\\xi}(i)_{high}" + }, + { + "bbox": [ + 46, + 532, + 287, + 654 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "text", + "content": "We measure the geometric effect of a descriptor " + }, + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\ell_{i}" + }, + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "text", + "content": " on a 3DMM by examining the deformation of the vertices of the mesh. This is evaluated by comparing the position of each vertex in the two extreme cases of " + }, + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\vec{\\xi} (i)_{low}" + }, + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\vec{\\xi} (i)_{high}" + }, + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "text", + "content": ". Hence, for each vertex " + }, + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "text", + "content": " the size of deformation is de" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 304, + 69, + 382, + 126 + ], + "blocks": [ + { + "bbox": [ + 304, + 69, + 382, + 126 + ], + "lines": [ + { + "bbox": [ + 304, + 69, + 382, + 126 + ], + "spans": [ + { + "bbox": [ + 304, + 69, + 382, + 126 + ], + "type": "image", + "image_path": "ba2e8233a3463231635a82b4dd5861c7c0bad839b00885c2da23b673b0f54061.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 304, + 138, + 380, + 173 + ], + "blocks": [ + { + "bbox": [ + 304, + 138, + 380, + 173 + ], + "lines": [ + { + "bbox": [ + 304, + 138, + 380, + 173 + ], + "spans": [ + { + "bbox": [ + 304, + 138, + 380, + 173 + ], + "type": "image", + "image_path": "b446a98b47353fb1fb61a099791f9d52189f37b7ae78b972d87dc3e14dc131f5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 188, + 380, + 251 + ], + "blocks": [ + { + "bbox": [ + 306, + 188, + 380, + 251 + ], + "lines": [ + { + "bbox": [ + 306, + 188, + 380, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 188, + 380, + 251 + ], + "type": "image", + "image_path": "0aa37f85af56025902920a1d234eb5600e573a056737d354461c241c7dafc3e6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 260, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 304, + 260, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 260, + 547, + 319 + ], + "type": "text", + "content": "Figure 4: A simple and intuitive interactive application can be defined using sliders for each semantic descriptor. With Semantiy, the sliders have strong semantic meaning and are disentangled so that changing one will have very limited affect on others." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 402, + 93, + 547, + 241 + ], + "blocks": [ + { + "bbox": [ + 402, + 93, + 547, + 241 + ], + "lines": [ + { + "bbox": [ + 402, + 93, + 547, + 241 + ], + "spans": [ + { + "bbox": [ + 402, + 93, + 547, + 241 + ], + "type": "image", + "image_path": "b810a44ef572a968a50973da2d71ea70decc818848d9e8d0a61f80ea36d12521.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "text", + "content": "fined as " + }, + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\delta'(v) = ||v_{low} - v_{high}||_2" + }, + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "text", + "content": ". We can normalize these values by " + }, + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\delta_{\\max} = \\max_{v \\in Mesh} \\delta'(v)" + }, + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "text", + "content": " and get a value " + }, + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\delta(v) = \\delta'(v) / \\delta_{\\max}" + }, + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "text", + "content": " between 0 and 1 (see Figure 5)." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 312, + 400, + 474, + 544 + ], + "blocks": [ + { + "bbox": [ + 312, + 400, + 474, + 544 + ], + "lines": [ + { + "bbox": [ + 312, + 400, + 474, + 544 + ], + "spans": [ + { + "bbox": [ + 312, + 400, + 474, + 544 + ], + "type": "image", + "image_path": "2b25c53b186ea56238c5157ab4294f6b64e2ccfbab03d23e9831276848c7345a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 554, + 545, + 602 + ], + "lines": [ + { + "bbox": [ + 304, + 554, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 545, + 602 + ], + "type": "text", + "content": "Figure 5: The geometric effect of a given descriptor. The darker the color of the vertex is (more red), the more it is affected by the descriptor. To define coverage we collect all vertices with a value larger than a given threshold " + }, + { + "bbox": [ + 304, + 554, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 554, + 545, + 602 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 490, + 437, + 548, + 543 + ], + "blocks": [ + { + "bbox": [ + 490, + 437, + 548, + 543 + ], + "lines": [ + { + "bbox": [ + 490, + 437, + 548, + 543 + ], + "spans": [ + { + "bbox": [ + 490, + 437, + 548, + 543 + ], + "type": "image", + "image_path": "b86ca843a083907af627e79beb5fed8dc8b91d41a4d60b9eb2b0ad6a0f9137e9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 629, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 712 + ], + "type": "text", + "content": "Using these differences, we can evaluate the coverage of a specific descriptor by measuring the number and positions of vertices it affects in the mesh. For example, we can look at all vertices " + }, + { + "bbox": [ + 304, + 629, + 545, + 712 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 629, + 545, + 712 + ], + "type": "text", + "content": " whose value is above a threshold " + }, + { + "bbox": [ + 304, + 629, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\delta(v) > \\theta" + }, + { + "bbox": [ + 304, + 629, + 545, + 712 + ], + "type": "text", + "content": " for any given descriptor. We can also compare the overlap between two descriptors by measuring the intersection-over-union of their covered vertices. We use these measure" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14559" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "in our ablation studies as described next in Section 5.2. Note that this measure is only an approximation as it measures the size of the change of vertices and not their direction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 117, + 115, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 117, + 115, + 128 + ], + "spans": [ + { + "bbox": [ + 47, + 117, + 115, + 128 + ], + "type": "text", + "content": "5.2. Ablations" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 135, + 289, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 135, + 289, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 135, + 289, + 319 + ], + "type": "text", + "content": "To further evaluate our algorithm's performance, we conducted experiments that include training the mapper with different numbers of descriptors and training on different sizes of training data. As we do not have valid ground truth to estimate the accuracy of our mapper we created a test set that contains several body shapes that were generated by the first 10 blendshapes " + }, + { + "bbox": [ + 46, + 135, + 289, + 319 + ], + "type": "inline_equation", + "content": "\\vec{\\xi}" + }, + { + "bbox": [ + 46, + 135, + 289, + 319 + ], + "type": "text", + "content": " of the SMPL-X model. In each experiment we optimize the descriptors' values " + }, + { + "bbox": [ + 46, + 135, + 289, + 319 + ], + "type": "inline_equation", + "content": "\\vec{\\Omega}" + }, + { + "bbox": [ + 46, + 135, + 289, + 319 + ], + "type": "text", + "content": ", to minimize the error " + }, + { + "bbox": [ + 46, + 135, + 289, + 319 + ], + "type": "inline_equation", + "content": "||\\hat{\\vec{\\xi}} - \\vec{\\xi}||_2" + }, + { + "bbox": [ + 46, + 135, + 289, + 319 + ], + "type": "text", + "content": ". The goal of these experiments is to test the expressive power of the given configuration to match a given shape. The smaller the error the more expressive the method is. We show the results of these ablations in the supplemental file. We used this evaluation to choose the hyper-parameters we use in our algorithm: training with 3K sample images with texture for all models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 335, + 211, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 211, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 211, + 348 + ], + "type": "text", + "content": "5.3. Interactive Sliders User Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 354, + 289, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 289, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 289, + 701 + ], + "type": "text", + "content": "To evaluate Semantify's performance we conducted a user study that compares our approach against BodyTalk [34], as well as the original basis of the 3DMM. The user is presented with a target 3D model (randomly created using the 3DMM) and is tasked to fit a 3DMM mesh model to the target shape using a set of sliders, using the two compared interfaces (see Figure 4) at random order. Users were given up to 5 minutes to complete the task. BodyTalk was implemented on the SMPL body model, therefore we compare it to our SMPL male and female models. In another study, we compare our SMPLX male and female models to the original PCA-based axes. In both studies, we seek to evaluate the accuracy of the user-fitted 3DMMs, the time that it took to fit the model, as well as the overall subjective experience of users. Since we cannot obtain the outputs of BodyTalk in the first study, we asked humans to rate the results, and determine which of them fits more accurately to the input shape. We had 10 users perform both studies (5 males, 5 females aged between 23 to 54), 7 of which are novices and 3 of which were professional 3D modelers. We present the quantitative results of these studies in Table 2. As can be seen, Semantify achieves better accuracy in less time. Note that theoretically only the baseline can reach zero error, but because it is very difficult to handle, the error obtained was higher than Semantify. In addition, we asked the users for feedback on the experience of using the different applications (A was ours, B was the alternative). Here are some examples of such reviews (more examples could be found in the supplemental file):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": "\"Using application B, occasionally when changing one" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "slider it affects the other, which causes the user to start all over again\"." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 96, + 545, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 96, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 305, + 96, + 545, + 121 + ], + "type": "text", + "content": "\"Using application B every minor change in a certain slider generated major changes in the rest of the sliders\"" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 121, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 121, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 305, + 121, + 545, + 156 + ], + "type": "text", + "content": "\"Application A was far friendlier and I sensed as though I maintained much more control over the different body features\"" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 157, + 545, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 157, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 305, + 157, + 545, + 182 + ], + "type": "text", + "content": "\"The abundance of sliders on application B only made it harder to control, not the other way around\"." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 190, + 449, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 190, + 449, + 202 + ], + "spans": [ + { + "bbox": [ + 305, + 190, + 449, + 202 + ], + "type": "text", + "content": "5.4. Zero Shot Reconstruction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 209, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 376 + ], + "type": "text", + "content": "To measure our zero-shot 3D-shape reconstruction method of SMPLX model, we use HBW (Human Bodies in-the-wild) dataset presented in SHAPY [7], containing ground-truth 3D body scans with in-the-wild images. Table 3 shows the performance of our method compared to SHAPY and PixIE [12] on all of HBW validation set. Note that SHAPY also takes keypoints (usually obtained from openPose [5]) as input. Some examples are shown in Figure 6. As can be seen, our zero-shot performance are on-par with state of the art methods, but we have the advantage of using semantic descriptors. This allows the user to fine tune the results from a very good initial guess, simply by using the semantic sliders. Some examples results with around a minute of fine-tuning can be found in the supplemental file." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 389, + 378, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 389, + 378, + 401 + ], + "spans": [ + { + "bbox": [ + 305, + 389, + 378, + 401 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 545, + 529 + ], + "type": "text", + "content": "We have presented a method to edit a 3D parametric model using semantically meaningful descriptors by first choosing a subset of descriptors judiciously and then learning a non-linear mapping from this set to the 3DMM coefficients. All this is done in a self-supervised manner harnessing the abilities of CLIP visual-language model. Our method can support a more intuitive user interface that allows even novices to model 3D shapes. It also supports other applications such as zero-shot reconstruction and fine-tuning of shapes from images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": "Limitations and future directions. There are several limitations to our method. First, our mapper's performance relies heavily on the data that is created (Section 3.1). For example, a dataset must contain extreme samples to achieve a more expressive mapper. Second, our goal with Semantify was to build a semantic representation of a given 3DMM without a human-in-the-loop. However, in some cases manually tuned modeling for a specific 3DMM would probably improve the 3DMM mapper's performance (we further elaborate on this in the supplemental file). Lastly, as shown in many other works, CLIP has rich semantic understanding, but there are cases in which its performance degrades. For example, performance is better with textured meshes, so using various textures for each 3DMM may improve the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14560" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 532, + 355 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 532, + 355 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 532, + 355 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 532, + 355 + ], + "type": "image", + "image_path": "496f0ccadf6c7582fd65d2e6abfef15bd4e73c4e6671a91bfe46eda410bcd999.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 363, + 547, + 388 + ], + "lines": [ + { + "bbox": [ + 46, + 363, + 547, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 547, + 388 + ], + "type": "text", + "content": "Figure 6: Results of zero-shot image to shape reconstruction. We compare Semantify against SHAPY and PixIE on SMPLX 3DMM. Colors indicate error compared to ground truth (darker means larger error)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 404, + 261, + 472 + ], + "blocks": [ + { + "bbox": [ + 53, + 404, + 261, + 472 + ], + "lines": [ + { + "bbox": [ + 53, + 404, + 261, + 472 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 261, + 472 + ], + "type": "table", + "html": "
Time (min)Score
MalesFemalesTotal
BodyTalk3.223.163.190.04%
ours2.121.451.590.96%
", + "image_path": "5f18fb9bfaacce1607bc68dc382d5cf2003119dc20f9a9f482a5543ae927e176.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 277, + 399, + 556, + 466 + ], + "blocks": [ + { + "bbox": [ + 277, + 399, + 556, + 466 + ], + "lines": [ + { + "bbox": [ + 277, + 399, + 556, + 466 + ], + "spans": [ + { + "bbox": [ + 277, + 399, + 556, + 466 + ], + "type": "table", + "html": "
Mean Error (cm)Time (min)
MalesFemalesTotalMalesFemalesTotal
Baseline0.1220.1110.1163.182.092.43
ours0.1210.0990.1102.222.172.19
", + "image_path": "d57469d8d99552b6e4c8880640eac051b2f6a5f0f4bb79cfc9315fa6f3278414.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 50, + 529, + 286, + 590 + ], + "blocks": [ + { + "bbox": [ + 46, + 485, + 547, + 510 + ], + "lines": [ + { + "bbox": [ + 46, + 485, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 547, + 510 + ], + "type": "text", + "content": "Table 2: User-study results for the SMPL 3DMM compared against BodyTalk [34] (left, Score means the percent that this model was selected as better fitting the input), and SMPL-X parametric model compared against the raw blendshapes (right)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 529, + 286, + 590 + ], + "lines": [ + { + "bbox": [ + 50, + 529, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 50, + 529, + 286, + 590 + ], + "type": "table", + "html": "
MalesFemalesTotal
SHAPY0.01969340.0144220.017085
PIXIE0.0247790.0152650.020324
Semantify (ours)0.0226320.0165640.019666
", + "image_path": "1ff0efca4da9766220533610ecfe6e2da76c1410c7c8964b37173ae2b9c98683.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "lines": [ + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "text", + "content": "Table 3: Comparing our zero-shot shape reconstruction from image method against SHAPY[7] and PixIE [12]. The values demonstrate the average MSE over HBW's in-the-wild validation set." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "mapper's performance. Furthermore, more subjective descriptors such as \"cute\", \"scary\", \"pretty\" etc., are harder to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 530, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 530, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 305, + 530, + 545, + 553 + ], + "type": "text", + "content": "rate using CLIP, and would probably not work well in our model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 558, + 547, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 631 + ], + "type": "text", + "content": "Future work can also improve the zero-shot 3D shape reconstruction by using prior information about its context (for example, the gender of a person, body keypoints as used in [7] etc.). Adding such priors along with CLIP's semantic understanding may surpass the existing state-of-the-art solutions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 653, + 416, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 653, + 416, + 667 + ], + "spans": [ + { + "bbox": [ + 306, + 653, + 416, + 667 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "content": "This research was supported by the Israel Science Foundation (grant No. 1390/19). We would also like to thank Eyal Gomel for valuable discussions." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14561" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 139 + ], + "type": "text", + "content": "[1] Rameen Abdal, Peihao Zhu, John Femiani, Niloy J. Mitra, and Peter Wonka. Clip2stylegan: Unsupervised extraction of stylegan edit directions. CoRR, abs/2112.05219, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 143, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 143, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 143, + 288, + 192 + ], + "type": "text", + "content": "[2] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Clipface: Text-guided editing of textured 3d morphable models. arXiv preprint arXiv:2212.01406, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 194, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 194, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 54, + 194, + 288, + 255 + ], + "type": "text", + "content": "[3] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. Scape: Shape completion and animation of people. ACM Transactions on Graphics (TOG), 24:408-416, 07 2005." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 258, + 288, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 288, + 306 + ], + "type": "text", + "content": "[4] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 309, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 309, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 54, + 309, + 288, + 369 + ], + "type": "text", + "content": "[5] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43:172–186, 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 373, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 287, + 409 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 412, + 288, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 412, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 412, + 288, + 483 + ], + "type": "text", + "content": "[7] Vasileios Choutas, Lea Müller, Chun-Hao P. Huang, Siyu Tang, Dimitrios Tzionas, and Michael J. Black. Accurate 3d body shape regression using metric and semantic attributes. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 487, + 288, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 487, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 54, + 487, + 288, + 547 + ], + "type": "text", + "content": "[8] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Monocular expressive body regression through body-driven attention. In European Conference on Computer Vision (ECCV), 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 551, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 551, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 54, + 551, + 288, + 609 + ], + "type": "text", + "content": "[9] Radek Daněček, Michael J. Black, and Timo Bolkart. Emotion driven monocular face capture and animation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20311-20322, June 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 614, + 287, + 673 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 673 + ], + "type": "text", + "content": "[10] Ginger Delmas, Philippe Weinzaepfel, Thomas Lucas, Francesc Moreno-Noguer, and Grégory Rogez. Postscript: 3d human poses from natural language. In European Conference on Computer Vision, pages 346-362. Springer, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 677, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 677, + 287, + 714 + ], + "type": "text", + "content": "[11] Bernhard Egger, William AP Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 328, + 72, + 545, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 545, + 108 + ], + "type": "text", + "content": "Sami Romdhani, et al. 3d morphable face models—past, present, and future. ACM Transactions on Graphics (TOG), 39(5):1-38, 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 111, + 547, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 111, + 547, + 170 + ], + "spans": [ + { + "bbox": [ + 307, + 111, + 547, + 170 + ], + "type": "text", + "content": "[12] Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael Black. Collaborative regression of expressive bodies using moderation. In International Conference on 3D Vision (3DV), pages 792-804, Dec. 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 174, + 547, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 547, + 244 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 547, + 244 + ], + "type": "text", + "content": "[13] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8649-8658, June 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 248, + 547, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 248, + 547, + 320 + ], + "spans": [ + { + "bbox": [ + 307, + 248, + 547, + 320 + ], + "type": "text", + "content": "[14] Thomas Gereg, Andreas Morel-Forster, Clemens Blumer, Bernhard Egger, Marcel Luthi, Sandro Schoenborn, and Thomas Vetter. Morphable face models - an open framework. 2018 13th IEEE International Conference on Automatic Face amp; Gesture Recognition (FG 2018), May 2018." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 323, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 323, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 307, + 323, + 545, + 382 + ], + "type": "text", + "content": "[15] Matthew Hill, Stephan Streuber, Carina Hahn, Michael Black, and Alice O'Toole. Exploring the relationship between body shapes and descriptions by linking similarity spaces. Journal of Vision, 15(12):931-931, 2015." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 385, + 545, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 385, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 307, + 385, + 545, + 434 + ], + "type": "text", + "content": "[16] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhonggang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 436, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 436, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 436, + 545, + 460 + ], + "type": "text", + "content": "[17] IEEE. A 3D Face Model for Pose and Illumination Invariant Face Recognition, Genova, Italy, 2009." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 463, + 545, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 463, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 307, + 463, + 545, + 522 + ], + "type": "text", + "content": "[18] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, P. Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 857-866, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 525, + 545, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 545, + 573 + ], + "type": "text", + "content": "[19] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 576, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 576, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 576, + 545, + 635 + ], + "type": "text", + "content": "[20] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. SIGGRAPH Asia 2022 Conference Papers, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 639, + 545, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 639, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 307, + 639, + 545, + 687 + ], + "type": "text", + "content": "[21] Youwang Kim, Ji-Yeon Kim, and Tae-Hyun Oh. Clip-actor: Text-driven recommendation and stylization for animating human meshes. In European Conference on Computer Vision, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 689, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 689, + 545, + 714 + ], + "type": "text", + "content": "[22] Muhammed Kocabas, Chun-Hao P. Huang, Joachim Tesch, Lea Müller, Otmar Hilliges, and Michael J." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14562" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 69, + 72, + 287, + 119 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 287, + 119 + ], + "type": "text", + "content": "Black. SPEC: Seeing people in the wild with an estimated camera. In Proc. International Conference on Computer Vision (ICCV), pages 11035-11045, Oct. 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 123, + 287, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 123, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 48, + 123, + 287, + 170 + ], + "type": "text", + "content": "[23] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 233 + ], + "type": "text", + "content": "[24] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 237, + 287, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 237, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 237, + 287, + 284 + ], + "type": "text", + "content": "[25] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. *Smpl: A skinned multi-person linear model*. ACM Trans. Graph., 34(6), nov 2015." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 289, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 289, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 289, + 287, + 335 + ], + "type": "text", + "content": "[26] Xudong Lou, Yiguang Liu, and Xuwei Li. Tecm-clip: Text-based controllable multi-attribute face image manipulation. In Proceedings of the Asian Conference on Computer Vision, pages 1942–1958, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 339, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 339, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 339, + 287, + 398 + ], + "type": "text", + "content": "[27] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 402, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 402, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 402, + 287, + 449 + ], + "type": "text", + "content": "[28] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. ArXiv, abs/2212.08751, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 453, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 287, + 535 + ], + "type": "text", + "content": "[29] Ahmed A. A. Osman, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Supr: A sparse unified part-based human representation. In Shai Avidan, Gabriel Brostow, Moustapha Cissé, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision – ECCV 2022, pages 568–585, Cham, 2022. Springer Nature Switzerland." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 540, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 540, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 540, + 287, + 599 + ], + "type": "text", + "content": "[30] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 603, + 287, + 674 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 674 + ], + "type": "text", + "content": "[31] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 677, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 677, + 287, + 713 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark," + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 328, + 73, + 545, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 73, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 328, + 73, + 545, + 108 + ], + "type": "text", + "content": "et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 112, + 545, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 112, + 545, + 146 + ], + "spans": [ + { + "bbox": [ + 307, + 112, + 545, + 146 + ], + "type": "text", + "content": "[33] Hyewon Seo, Frederic Cordier, and Nadia Thalmann. Synthesizing animatable body models with parameterized shape modifications. 07 2003." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 150, + 545, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 209 + ], + "type": "text", + "content": "[34] Stephan Streuber, M. Alejandra Quiros-Ramirez, Matthew Q. Hill, Carina A. Hahn, Silvia Zuffi, Alice O'Toole, and Michael J. Black. Body Talk: Crowd-shaping realistic 3D avatars with words. ACM Trans. Graph. (Proc. SIGGRAPH), 35(4), July 2016." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 213, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 213, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 213, + 545, + 261 + ], + "type": "text", + "content": "[35] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 265, + 545, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 545, + 311 + ], + "type": "text", + "content": "[36] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 316, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 397 + ], + "type": "text", + "content": "[37] Ayush Tewari, Michael Zollhöfer, Pablo Garrido, Florian Bernard, Hyeongwoo Kim, Patrick Pérez, and Christian Theobalt. Self-supervised multi-level face model learning for monocular reconstruction at over 250 hz. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 402, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 402, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 402, + 545, + 460 + ], + "type": "text", + "content": "[38] Can Wang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 465, + 545, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 465, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 307, + 465, + 545, + 512 + ], + "type": "text", + "content": "[39] Can Wang, Ruixia Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. ArXiv, abs/2212.08070, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 516, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 516, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 516, + 545, + 563 + ], + "type": "text", + "content": "[40] Jinbao Wang, Shujie Tan, Xiantong Zhen, Shuo Xu, Feng Zheng, Zhenyu He, and Ling Shao. Deep 3d human pose estimation: A review. Computer Vision and Image Understanding, 210:103225, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 567, + 545, + 637 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 637 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 637 + ], + "type": "text", + "content": "[41] Lizhen Wang, Zhiyua Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR2022), June 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 642, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 642, + 545, + 713 + ], + "type": "text", + "content": "[42] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Ghum & ghuml: Generative 3d human shape and articulated pose models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6184-6193, 2020." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14563" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 269 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 144 + ], + "type": "text", + "content": "[43] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. *Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction.* In *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition* (CVPR), 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 147, + 287, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 147, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 147, + 287, + 208 + ], + "type": "text", + "content": "[44] Silvia Zuffi, Angjoo Kanazawa, and Michael J. Black. Lions and tigers and bears: Capturing non-rigid, 3d, articulated shape from images. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3955-3963, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 210, + 287, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 210, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 287, + 269 + ], + "type": "text", + "content": "[45] Silvia Zuffi, Angjoo Kanazawa, David Jacobs, and Michael J. Black. 3D menagerie: Modeling the 3D shape and pose of animals. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), July 2017." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14564" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Semi-Supervised Learning via Weight-Aware Distillation under Class Distribution Mismatch/ab302b84-1623-4a04-b35b-84d3ca9ad536_content_list.json b/2023/Semi-Supervised Learning via Weight-Aware Distillation under Class Distribution Mismatch/ab302b84-1623-4a04-b35b-84d3ca9ad536_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d6d278892c597a6151778879324580f87ef2e756 --- /dev/null +++ b/2023/Semi-Supervised Learning via Weight-Aware Distillation under Class Distribution Mismatch/ab302b84-1623-4a04-b35b-84d3ca9ad536_content_list.json @@ -0,0 +1,1769 @@ +[ + { + "type": "text", + "text": "Semi-Supervised Learning via Weight-aware Distillation under Class Distribution Mismatch", + "text_level": 1, + "bbox": [ + 137, + 128, + 831, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pan Du $^{1,2}$ , Suyun Zhao $^{1,2,*}$ , Zisen Sheng $^{1,2}$ , Cuiping Li $^{1,2}$ , Hong Chen $^{1,2}$ \nKey Lab of Data Engineering and Knowledge Engineering of MOE Renmin University of China $^{1}$ \nRenmin University of China, Beijing, China $^{2}$", + "bbox": [ + 101, + 202, + 867, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{duPan,shengzisen}@163.com,{zhaosuyun,zisen,licuiping,chong}@ruc.edu.cn", + "bbox": [ + 147, + 258, + 812, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 309, + 313, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semi-Supervised Learning (SSL) under class distribution mismatch aims to tackle a challenging problem wherein unlabeled data contain lots of unknown categories unseen in the labeled ones. In such mismatch scenarios, traditional SSL suffers severe performance damage due to the harmful invasion of the instances with unknown categories into the target classifier. In this study, by strict mathematical reasoning, we reveal that the SSL error under class distribution mismatch is composed of pseudo-labeling error and invasion error, both of which jointly bound the SSL population risk. To alleviate the SSL error, we propose a robust SSL framework called Weight-Aware Distillation (WAD) that, by weights, selectively transfers knowledge beneficial to the target task from unsupervised contrastive representation to the target classifier. Specifically, WAD captures adaptive weights and high-quality pseudolabels to target instances by exploring point mutual information (PMI) in representation space to maximize the role of unlabeled data and filter unknown categories. Theoretically, we prove that WAD has a tight upper bound of population risk under class distribution mismatch. Experimentally, extensive results demonstrate that WAD outperforms five state-of-the-art SSL approaches and one standard baseline on two benchmark datasets, CIFAR10 and CIFAR100, and an artificial cross-dataset. The code is available at https://github.com/RUC-DWBI-ML/research/tree/main/WAD-master.", + "bbox": [ + 75, + 340, + 473, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 777, + 209, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) have achieved remarkable success in fully-supervised learning tasks. However, sufficient labeled data are usually unavailable in real applications due to the expensive annotation cost or even domain-specific knowledge required [8, 11, 12, 13]. Semi-", + "bbox": [ + 75, + 801, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17e59410467b98a65d776d258b4c60844e60bdac60c7d897a455cc036f59a9d3.jpg", + "image_caption": [ + "Figure 1. Example of class distribution mismatch. The unlabeled data contains categories that are unseen in labeled ones." + ], + "image_footnote": [], + "bbox": [ + 503, + 306, + 888, + 483 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "supervised learning (SSL), as a powerful weakly-supervised technique, provides an effective way to improve DNNs by exploiting massive unlabeled data, and then it weakens the demand for human annotation [9, 14, 24, 34]. Generally, traditional SSL approaches assume that the labeled and unlabeled instances share the same class distribution, i.e., they come from identical categories. However, in real scenarios, this assumption hardly holds as unlabeled data inevitably contains lots of categories unseen in labeled ones. For instance, if unlabeled data are collected from the internet using keywords \"cat\" and \"dog\" (target categories), they may contain instances unrelated to these categories, such as \"deer,\" \"horse,\" or \"airplane\"(unknown categories), as shown in Figure 1. Similar scenarios occur in medical diagnoses [11, 15] and house annotations of remote-sensing images [12, 13]. SSL in such mismatch scenarios is called SSL under class distribution mismatch [12, 15].", + "bbox": [ + 496, + 534, + 892, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Under class distribution mismatch, some SSL approaches [8, 11, 15, 19, 37] have been proposed. Usually, most of them exploit pseudo-labeling or consistency regularization to expand the labeled pool, as well as filter instances with unknown categories by weights, just as shown in Figure 2. UASD [11] and T2T [19] filter out the instances with unknown categories by leveraging a hard weight, i.e., a", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 101, + 887, + 230, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16410", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6a3dfd6d1f31232dc90e781df99df420becd0c1abf22d49a313a685da74a29c6.jpg", + "image_caption": [ + "Figure 2. The paradigm of SSL under class distribution mismatch." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 467, + 172 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "threshold, on the accumulated network's output or the out-of-distribution score. Although these two approaches reduce the invasion of unknown categories, it is inevitable to keep off amounts of unlabeled instances with target categories. Instead of hard weights, Guo et al. [15] assign a soft weight to the unlabeled instances according to the consistent empirical risk loss. In such case, many instances with unknown categories tend to have consistent outputs and get high weights, just as shown in Appendix 4.3, and then they may invade the target classifier and impair its performance.", + "bbox": [ + 75, + 220, + 467, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, the existing SSL approaches with consistency regularization and pseudo-labeling heavily rely on the performance of the target classifier. Both [15] and [19] annotate pseudo labels by leveraging the prediction of the target classifier in training. Once the target classifier trained on limited labeled instances is biased by some instances with unknown categories, the subsequently updated target classifier may allow more unknown instances to invade. Accordingly, it is promising to propose a novel SSL approach that captures pseudo labels from representations produced by all available data rather than an immature classifier.", + "bbox": [ + 75, + 371, + 467, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this study, by strict theoretical analyses, we decouple the SSL error under class distribution mismatch into pseudo-labeling error and invasion error (seen in Subsection 3.2). According to this discovery, a robust SSL framework called weight-aware distillation (WAD) is then proposed to distill pseudo labels and weights from the representation space to the target classifier. Unlike the conventional distillation approaches [7, 17, 28] that simply train the student model using the prediction probability of the teacher model, WAD is a weight-aware distillation framework that adapts to mismatch problems. Specifically, we learn the representations from labeled and unlabeled data by unsupervised contrastive coding, as the teacher model. Then WAD captures adaptive weights as well as high-quality pseudo labels from the teacher model by leveraging point mutual information(PMI), and thus, the target classifier could selectively utilize the instances from target categories while filtering the ones with unknown categories.", + "bbox": [ + 75, + 537, + 467, + 810 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are listed as follows.", + "bbox": [ + 96, + 811, + 388, + 825 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "i) We theoretically analyze the population risk in an SSL manner and reveal that the SSL error under class distribution mismatch is jointly controlled by pseudolabeling error and invasion error.", + "bbox": [ + 91, + 839, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "ii) We propose a distillation-based SSL framework, WAD, that captures weights as well as pseudo labels from robust representations to the target classifier to filter unknown categories and make full use of targeted unlabeled instances as well.", + "iii) Theoretically, we verify that the population risk of WAD is tightly bounded. Experimentally, WAD outperforms five state-of-the-art SSL approaches and one standard baseline on several datasets." + ], + "bbox": [ + 504, + 90, + 890, + 238 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 253, + 638, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This section reviews the SSL approaches under class distribution match and mismatch. For contrastive learning, please refer to Appendix 1.", + "bbox": [ + 498, + 280, + 890, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semi-Supervised Learning. The traditional SSL strategies include entropy minimization, consistency regularization, and pseudo-label. Entropy minimization [14] incorporates unlabeled data in supervised learning by minimizing the entropy of the unlabeled instance's prediction. The consistency regularization [24, 31, 34] techniques mainly make the prediction on two views of one instance consistent. II-Model [31] focuses on reducing the distance of prediction between one instance and its stochastic perturbation. Unlike the II-Model, temporal ensembling [24] adopts the ensemble of predictions as the target to achieve more stable performance, while Virtual Adversarial Training (VAT) [27] explores adversarial disturbances of the unlabeled instances on the prediction of the target classifier. Pseudo-Labeled based approaches [3, 4, 26, 33] annotate some unlabeled instances with pseudo labels to expand the labeled data. By leveraging the class probability of the unlabeled data, a pseudo-labeling method is proposed [26]. Furthermore, FixMatch [33] uses the weakly augmented unlabeled instances to create a pseudo label and enforce consistent prediction against its strong augmented version.", + "bbox": [ + 498, + 325, + 890, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These traditional SSL approaches perform well when the class distribution is matched, but they suffer severe performance degradation under class distribution mismatch.", + "bbox": [ + 498, + 643, + 890, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semi-Supervised Learning under Class Distribution Mismatch. To tackle class distribution mismatch, several studies [11, 15, 19, 37] adopt the traditional SSL strategies with the assistance of soft or hard weights. UASD [11] leverages a threshold to the accumulated network's output to eliminate the instances with unknown categories, followed by pseudo-labeling highly confident ones. Similarly, T2T [19] adopts a hard weight on the out-of-distribution score to conduct filtering and leverages consistency constraints to expand the labeled pool. Furtherly, CCSSL [37] filters out unknown instances by taking both hard and soft weights into consideration. These approaches with hard weights may eliminate too many instances from target categories. Instead of hard weights, $\\mathrm{DS}^{3}\\mathrm{L}$ [15] assigns soft", + "bbox": [ + 498, + 688, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16411", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e32d869aec5aa08a0452a34edb56cd1157118e18e71a3d72877894b0a4aa87f9.jpg", + "image_caption": [ + "Figure 3. The selective distillation in WAD. Traditional distillation methods force the output of the student model $(s_i)$ to align with that of the teacher model $(t_i)$ , resulting in a wrong decision boundary. However, WAD selectively distills benefit knowledge to student and filter the negatives, such as $s_2$ and $s_3$ , by weights to rectify the decision boundary and solve the problem of class distribution mismatch." + ], + "image_footnote": [], + "bbox": [ + 89, + 90, + 379, + 243 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8110fc305963ff6ee202a9d658d763a695f120893a11b718707935137afb7953.jpg", + "image_caption": [], + "image_footnote": [ + "- : Target category 1\n- : Target category 2\n- : Unknown category\n- : Decision boundary\n- >: Knowledge distillation" + ], + "bbox": [ + 401, + 90, + 715, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "weights to unlabeled instances according to the consistent empirical risk loss. However, SSL with pseudo labeling or consistency regularization heavily rely on the performance of the target classifier, and thus they are susceptible to being invaded by instances with unknown categories.", + "bbox": [ + 75, + 316, + 470, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, a model-level approach [40] is proposed by modifying batch normalization to counter the unknown categories. Also, ORCA [8], a novelty detection approach, leverages uncertainty-based adaptive margins to circumvent the bias caused by the mismatched distribution.", + "bbox": [ + 75, + 395, + 470, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Knowledge Distillation. Knowledge distillation aims to transfer knowledge from a big model (teacher model) to a smaller one (student model) [35]. It is widely applied to two distinct fields: model compression and knowledge transfer. Model compression is training a small student model to mimic the big teacher model or the ensemble of models. Bucilua et al. [7] compress the ensembles of the neural networks into a single one. While the approaches based on transfer knowledge concentrate more on effectively transferring and are mainly divided into logits-based and representation-based distillation [39]. The logits-based distillation approaches usually train the student model by leveraging the output of the teacher model as the soft label [35]. Ba et al. [1] propose to push the logits, i.e., the output before the softmax function, of the shallow neural network to mimic the ones from a deep neural network. Furtherly, Hinton et al. [17] suggest training a student model to match the combination of the softmax distribution of the teacher model and ground truth. Representation-based approaches enable the student model to learn information from the intermediate layers [35]. Kim et al. [21] propose transferring the attention map from the teacher to the student. Park et al. [30] introduce a novel approach that transfers the mutual relationship of the instances learned from the teacher to the student, similar to our intention.", + "bbox": [ + 75, + 474, + 470, + 851 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, these approaches mentioned above aim to transfer as much information as possible to the student model and ignore the unknown instances under class dis", + "bbox": [ + 76, + 854, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ribution mismatch, which may severely hurt the training of the student model. Unlike the conventional approaches, WAD is a weight-aware distillation framework that selectively transfers the knowledge to the student model, as shown in Figure 3, to fully use the beneficial knowledge and filter the unknown ones by weights. Specifically, WAD distills high-quality pseudo labels to the instances with target categories and filters the instances with unknown categories by assigning them tiny weights.", + "bbox": [ + 496, + 316, + 893, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 464, + 589, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we propose WAD, an SSL framework under class distribution mismatch. Concretely, Subsection 3.1 introduces the problem statement, followed by analyses of the SSL error in Subsection 3.2. Subsection 3.3 subsequently presents WAD. Finally, theoretical studies of WAD are conducted in Subsection 3.4.", + "bbox": [ + 496, + 489, + 893, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Statement", + "text_level": 1, + "bbox": [ + 500, + 589, + 684, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this study, we investigate the $K$ classification problem in an SSL manner wherein limited labeled data $\\mathcal{D}_l = \\{(x_{i,l},y_{i,l})\\}_{i = 1}^m$ and massive unlabeled instances $\\mathcal{D}_u = \\{x_{i,u}\\}_{i = 1}^n$ are accessible, $x_{i,l}\\in \\mathcal{X}$ $y_{i,l}\\in \\mathcal{Y}$ $\\mathcal{V} = \\{1,\\dots,K\\}$ and $m\\ll n$ . Under class distribution mismatch, the unlabeled instances are not guaranteed to belong to the $K$ target categories in $\\mathcal{V}$ .", + "bbox": [ + 498, + 612, + 893, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Population Risk Analysis", + "text_level": 1, + "bbox": [ + 500, + 726, + 728, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To make full use of the unlabeled data, we assign a pseudo label to each unlabeled instance, denoted as $\\hat{y}$ , and then build the target classifier, $h_{\\hat{T}}: \\mathcal{X} \\to \\mathcal{Y}$ , to map the given instance to one of the known categories in $\\mathcal{Y}$ , where $\\hat{T} = \\{x_{i,l}, y_{i,l}\\}_{i=1}^{m} \\cup \\{x_{i,u}, \\hat{y}_{i,u}\\}_{i=1}^{n}$ , $\\hat{y}_{i,u} \\in \\mathcal{Y}$ . Here, $\\hat{T}$ indicates the instances in hand, that is, labeled instances and unlabeled instances assigned with pseudo labels. Then, the population risk [32] of the target classifier learned from both labeled and unlabeled data with the pseudo label $(\\hat{T})$ is controlled by the generalization gap, training error, and SSL", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16412", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "error, as shown in Eq.1. The generalization gap is the gap between the population risk and the average prediction loss across all instances with target categories $(T)$ . Note that $T$ contains all the accessible instances with target categories, including labeled and unlabeled. And every instance in $T$ is assumed with ground truth labels in ideal. The training error is the average empirical loss across $\\hat{T}$ . The SSL error is the gap between the average empirical loss across the instances with target categories $(T)$ and the average empirical loss across both labeled data and unlabeled ones with pseudo labels $(\\hat{T})$ . We depict the relations among these sets in Figure 4.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbb {E} _ {(\\boldsymbol {x}, \\boldsymbol {y}) \\sim D} [ l (\\boldsymbol {x}, \\boldsymbol {y}; h _ {\\hat {T}}) ] \\\\ \\leq \\underbrace {\\left| \\mathbb {E} _ {(\\boldsymbol {x} , y) \\sim D} [ l (\\boldsymbol {x} , y ; h _ {\\hat {T}}) ] - \\frac {1}{| T |} \\sum_ {(\\boldsymbol {x} , y) \\in T} l (\\boldsymbol {x} , y ; h _ {\\hat {T}}) \\right|} _ {\\text {g e n e r a l i z a t i o n g a p}} \\\\ + \\underbrace {\\left| \\frac {1}{| \\hat {T} |} \\sum_ {(\\boldsymbol {x} , y) \\in \\hat {T}} l (\\boldsymbol {x} , y ; h _ {\\hat {T}}) \\right|} _ {\\text {t r a i n i n g e r r o r}} \\tag {1} \\\\ + \\underbrace {\\left| \\frac {1}{| T |} \\sum_ {(\\pmb {x} , y) \\in T} l (\\pmb {x} , y ; h _ {\\hat {T}}) - \\frac {1}{| \\hat {T} |} \\sum_ {(\\pmb {x} , y) \\in \\hat {T}} l (\\pmb {x} , y ; h _ {\\hat {T}}) \\right|} _ {\\mathrm {S S L e r r o r}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 287, + 468, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{D}$ is the data distribution of the instances that belong to target categories in the realistic world, i.e., $\\mathcal{D} = \\mathcal{X} \\times \\mathcal{Y}$ . $l(\\cdot, \\cdot; h_{\\hat{T}}): \\mathcal{X} \\times \\mathcal{Y} \\to \\mathcal{R}$ denotes the loss function of the classifier $h_{\\hat{T}}$ learned from $\\hat{T}$ .", + "bbox": [ + 76, + 481, + 468, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theoretical analyses [36] have confirmed that the generalization gap of DNNs can be bounded, and empirical evidence suggests that the training error of DNNs can be reduced almost to zero [32]. Thus, the essential component concerning population risk is the SSL error. Under class distribution mismatch, in addition to the wrongly annotated instances with target categories, the ones with unknown categories also contribute to the SSL error as they invade the training of the target classifier as outliers. Accordingly, we decouple the SSL error into pseudo-labeling and invasion error, as shown in Eq.2. For a detailed derivation process, please refer to Appendix 5.2.", + "bbox": [ + 75, + 545, + 468, + 724 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left| \\frac {1}{| T |} \\sum_ {(\\boldsymbol {x}, y) \\in T} l (\\boldsymbol {x}, y; h _ {\\hat {T}}) - \\frac {1}{| \\hat {T} |} \\sum_ {(\\boldsymbol {x}, y) \\in \\hat {T}} l (\\boldsymbol {x}, y; h _ {\\hat {T}}) \\right| \\\\ \\leq \\underbrace {\\left| \\frac {1}{| T |} \\sum_ {(\\boldsymbol {x} , y) \\in T} l (\\boldsymbol {x} , y ; h _ {\\hat {T}}) - \\frac {1}{| \\hat {T} |} \\sum_ {(\\boldsymbol {x} , y) \\in \\hat {T} \\setminus U} l (\\boldsymbol {x} , y ; h _ {\\hat {T}}) \\right|} _ {\\textbf {P s e u d o - l a b e l i n g e r r o r}} \\\\ + \\underbrace {\\left| \\frac {1}{| \\hat {T} |} \\sum_ {(\\boldsymbol {x} , y) \\in U} l (\\boldsymbol {x} , y ; h _ {\\hat {T}}) \\right|} _ {\\text {I n v a s i o n e r r o r}} \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 729, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bb87031d6d3253f1852ffe8e7123ca4b44767cb6b1624f83879b14e51a1b8d72.jpg", + "image_caption": [ + "Figure 4. The relations among data sets $\\mathcal{D}_l$ , $\\mathcal{D}_u$ , $T$ , $U$ , $\\hat{T}$ . Note that $\\hat{T} \\neq T \\cup U$ due to the instances with target categories in $\\mathcal{D}_u$ are assigned with pseudo labels while not the ground truth ones in $T$ ." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 887, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $U$ indicates the unlabeled instances with unknown categories and $\\hat{T} = \\hat{T}\\backslash U\\cup U$", + "bbox": [ + 498, + 242, + 890, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Eq.2, the pseudo-labeling error is contributed by the wrongly annotated instances with target categories, as it is the gap of the average empirical loss caused by the inconsistency of the ground truth and pseudo labels. Thus, the quality of pseudo-labels assigned to unlabeled instances within the target distribution determines this error, and accurate pseudo-labeling may alleviate it. By contrast, the invasion error is the average empirical loss across the instances with unknown categories that is caused by the negative effect of those untargeted instances. By Eq.1& Eq.2, we find that the population risk of the target model is jointly controlled by pseudo-labeling error and invasion error. Accordingly, to mitigate the SSL error, we need to filter those instances with unknown categories and accurately annotate the unlabeled instances with target categories as well.", + "bbox": [ + 498, + 273, + 892, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Weight-aware Distillation Framework", + "text_level": 1, + "bbox": [ + 500, + 511, + 826, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the aim of mitigating the pseudo-labeling and invasion errors, we design an SSL framework named WAD, which delivers the knowledge of pseudo labels and weights from robust representations to the target classifier.", + "bbox": [ + 498, + 535, + 890, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Pseudo Label Learning", + "text_level": 1, + "bbox": [ + 500, + 618, + 718, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Most existing SSL approaches produce pseudo labels by leveraging an immature target classifier, which cause catastrophic error once invaded by some instances with unknown categories, just as discussed in Section 1. To solve this problem, we distill the pseudo labels from a representation space (Teacher model) which is learned from all labeled and unlabeled instances by contrastive learning in an unsupervised manner and then transfer it to the target classifier (Student model). The teacher model could produce closely aligned representations for instances from the same categories and maximize the mutual information among them [2, 18, 20, 25].", + "bbox": [ + 496, + 642, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Denoted the labeled and unlabeled representations learned by the teacher model, $\\phi$ , as $\\mathcal{Z}_l = \\{z_{j,l,k}\\}_{j=1}^m$ and $\\mathcal{Z}_u = \\{z_{i,u}\\}_{i=1}^n$ , respectively, where $k \\in \\mathcal{V}$ . Inspired by the characteristic of contrastive learning, one effective approach for building the pseudo-label of an unlabeled in-", + "bbox": [ + 498, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16413", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b7cf7136a5b561ffcc6a96843bf126f9d3e4e021c4974f9e1d3dcac29573989d.jpg", + "image_caption": [ + "Figure 5. Illustration of WAD. The pseudo labels and weights determined by point mutual information (PMI) in robust representations of the teacher model are used to participate in training the target classifier. Then, some reliable instances selected according to $\\mathcal{L}_{\\mathrm{WAD}}$ are regarded as labeled ones to update the knowledge gradually. “ $\\bullet$ ” and “ $\\bullet$ ” indicates the labeled and unlabeled instance respectively. “//” means stop gradient." + ], + "image_footnote": [], + "bbox": [ + 88, + 85, + 883, + 323 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "stance is to identify the labeled instance with the highest PMI and then assign the label of it to unlabeled ones. The PMI between the unlabeled and labeled representation is formulated as Eq.3.", + "bbox": [ + 75, + 406, + 470, + 467 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {P M I} \\left(z _ {i, u}, z _ {j, l, k}\\right) = \\log \\left[ \\frac {p \\left(z _ {j , l , k} \\mid z _ {i , u}\\right)}{p \\left(z _ {j , l , k}\\right)} \\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 476, + 468, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although the conditional and marginals distributions, i.e., $p(z_{j,l,k}|z_{i,u})$ and $p(z_{j,l,k})$ , cannot be directly evaluated, we prove that PMI is proportional to the inner product in Appendix 2, as described in Eq.4.", + "bbox": [ + 76, + 518, + 468, + 580 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf \\left(z _ {i, u}, z _ {j, l, k}\\right) \\propto \\operatorname {P M I} \\left(z _ {i, u}, z _ {j, l, k}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 590, + 468, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f = \\cos(\\cdot, \\cdot)$ , $\\| z_i \\| = 1$ , and $\\propto$ stands for \"proportional to\".", + "bbox": [ + 75, + 617, + 468, + 646 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, the pseudo label is formulated as Eq.5.", + "bbox": [ + 96, + 648, + 426, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {i, u} = \\underset {k} {\\arg \\max } f \\left(z _ {i, u}, z _ {j, l, k}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 674, + 468, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consequently, the class label of the labeled instance with the maximum PMI is assigned to the unlabeled one. The Eq.5 can precisely capture the PMI from the representation space to produce high-quality pseudo labels and then mitigate pseudo-labeling error.", + "bbox": [ + 75, + 707, + 468, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Unknown Categories Filtering", + "text_level": 1, + "bbox": [ + 76, + 801, + 341, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To mitigate the invasion error, the instances with unknown categories should be filtered out. Following the Subsection 3.3.1, a higher PMI between the labeled and unlabeled instance suggests a stronger association or similarity between the two instances, further indicating a higher", + "bbox": [ + 75, + 825, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "likelihood of the unlabeled instance belonging to the same class distribution as the labeled one. However, some hard instances that have similar PMI between two target categories, i.e., laid on the decision boundary of two target categories, may introduce incorrect pseudo labels and hurt the performance of the target classifier. Hence, we also propose a ratio among the first and second maximum PMI to evaluate the confidence of the pseudo labels. Then, the weight is defined as Eq.6 to avoid the negative effect caused by the wrong labels and unknown categories.", + "bbox": [ + 496, + 406, + 890, + 556 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {w} _ {i, u} = g _ {1} \\left(\\widetilde {p} _ {i, u}\\right) \\times g _ {2} \\left(1 - \\frac {\\widetilde {q} _ {i , u}}{\\widetilde {p} _ {i , u}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 565, + 890, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "wherein,", + "bbox": [ + 500, + 606, + 560, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {p} _ {i, u} = \\max _ {j} f (z _ {i, u}, z _ {j, l, k})\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 619, + 771, + 642 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {q} _ {i, u} = \\max _ {v, k \\neq \\hat {y} _ {i, u}} f (z _ {i, u}, z _ {v, l, k})\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 645, + 790, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Eq.6, $g_{1}(\\cdot)$ and $g_{2}(\\cdot)$ can be interpreted as any monotonically increasing functions. The former in Eq.6 aims to estimate the likelihood of the unlabeled instance belonging to target categories. The higher this item, the more chances of the instance in the target class distribution are. The latter, $g_{2}(\\cdot)$ , penalizes instances whose labels are ambiguous between the nearest and second-nearest target categories. The lower this item, the larger probability of incorrect pseudo labels is. As shown in Figure 6, the weight could filter instances with unknown categories and those incorrectly annotated ones with target categories, while the ones from target categories with high-quality pseudo labels are encouraged. Thus, by weight, WAD selectively distills the knowledge beneficial to the target classifier from the teacher model, and the invasion error is then mitigated.", + "bbox": [ + 500, + 674, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16414", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3 Weight-aware Knowledge Distillation", + "text_level": 1, + "bbox": [ + 76, + 90, + 392, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Weight-aware knowledge distillation loss. The knowledge of pseudo labels and weights captured from robust representations is applied in the distillation process. In each feed-forward process, pseudo labels and weights are aggregated to the target classifier, as shown in Figure 5. Then, we propose the weight-aware knowledge distillation loss, including the traditional supervised loss $\\mathcal{L}_l$ in labeled data and weight-aware supervised loss $\\mathcal{L}_u$ in unlabeled data as Eq.7.", + "bbox": [ + 75, + 114, + 468, + 250 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {W A D}} = \\mathcal {L} _ {l} + \\mathcal {L} _ {u}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 250, + 468, + 266 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "wherein,", + "bbox": [ + 76, + 272, + 137, + 285 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l} = \\frac {1}{| \\mathcal {D} _ {l} |} \\sum_ {(x _ {i, l}, y _ {i, l}) \\in \\mathcal {D} _ {l}} \\ell (h (x _ {i, l}; \\theta), y _ {i, l})\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 296, + 395, + 335 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {u} = \\frac {1}{| \\mathcal {D} _ {u} |} \\sum_ {x _ {i, u} \\in \\mathcal {D} _ {u}} \\boldsymbol {w} _ {i, u} \\ell (h (x _ {i, u}; \\theta), \\hat {y} _ {i, u})\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 339, + 413, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The traditional supervised loss $\\mathcal{L}_l$ aims to minimize the distance between the predicted probability and the ground truth label. While the weight-aware supervised loss $\\mathcal{L}_u$ mainly focuses on selectively transferring the beneficial knowledge from the teacher model to the student model by weights to mitigate the negative effect from unknown categories and improve the target classifier as well. Moreover, $\\mathcal{L}_{\\mathrm{WAD}}$ is the loss function that is adopted to train the target classifier $h_{\\hat{T}}$ mentioned in Eq.1. Consequently, WAD leverages the pseudo labels and weights to mitigate the pseudolabeling and invasion errors, following alleviating the SSL error, which has been proved in Subsection 3.4.", + "bbox": [ + 75, + 381, + 468, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Knowledge-update in Training. The knowledge of pseudo labels and weights may be biased as the labeled data is limited. Accordingly, after several forward iterations, we progressively add some reliable instances to labeled data. Because the feedback from the target classifier, i.e., loss, is highly related to the weights and reflects the training error, we consider the reliability according to it. Then, the criterion for updating is formulated as Eq.8.", + "bbox": [ + 75, + 561, + 468, + 683 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nc _ {i, u} = \\ell \\left(h \\left(x _ {i, u}; \\theta_ {t}\\right), \\hat {y} _ {i, u}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 693, + 468, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\ell (\\cdot ,\\cdot)$ is the cross-entropy function, and $\\theta_{t}$ is the parameters of the target classifier in the current iteration.", + "bbox": [ + 75, + 719, + 468, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The reliability of $x_{i,u}$ is enhanced when $c_{i,u}$ takes a lower value. Then, WAD leverages Eq.8 to identify the top $\\alpha \\%$ reliable instances from the unlabeled data and puts them in the labeled data while removing them from the unlabeled data. Moreover, we adopt the polynomial decay [5] to dynamically adjust $\\alpha$ to prevent the gradually increased negative effect from unknown categories with the iteration. The details are shown in Appendix 3. A visualization of the number of selected reliable instances with target categories is also provided in Appendix 4.4. Consequently, the pseudo", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "labels and weights are updated in the subsequent distillation steps, as shown in Figure 5, with the aim of optimizing the target classifier. Finally, the schematic diagram and algorithm process is presented in Figure 6 and Algorithm 1, respectively.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Theoretical Studies", + "text_level": 1, + "bbox": [ + 498, + 176, + 684, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This subsection provides the theoretical studies about the WAD's SSL error, as shown in Theorem 1. Detailed proof of Theorem 1 is given in Appendix 5.", + "bbox": [ + 496, + 200, + 890, + 246 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Theorem 1 Given $|T|$ instances that i.i.d. sampled from $\\mathcal{D}$ as $\\{(x_i, y_i)\\}_{i=1}^{|T|}$ , $|U|$ instances that is not i.i.d with $\\mathcal{D}$ , and $\\hat{T} = \\{(x, \\hat{y}) | (x, y) \\in T \\cup U, \\hat{y} \\in \\mathcal{Y}\\}$ where $\\mathcal{D} = \\mathcal{X} \\times \\mathcal{Y}$ . Assume the loss function $l(\\cdot, y; h_{\\hat{T}})$ is $\\lambda^l$ -Lipschitz continuous for all $y, h_{\\hat{T}}$ and bounded by $H$ , the regression function is $\\lambda^\\mu$ -Lipschitz continuous, training error $l(\\pmb{x}, y; h_{\\hat{T}}) = 0$ , $\\forall (\\pmb{x}, y) \\in \\hat{T}$ . $\\overline{\\pmb{w}}$ indicates the average of weights, and $\\xi$ is the maximum PMI which determines the pseudo label, with the probability of at least $1 - \\gamma$ ,", + "bbox": [ + 496, + 255, + 892, + 391 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left| \\frac {1}{| T |} \\sum_ {(\\boldsymbol {x}, y) \\in T} l (\\boldsymbol {x}, y; h _ {\\hat {T}}) - \\frac {1}{| \\hat {T} |} \\sum_ {(\\boldsymbol {x}, y) \\in \\hat {T} \\backslash U} l (\\boldsymbol {x}, y; h _ {\\hat {T}}) \\right| \\\\ + \\left| \\frac {1}{| \\hat {T} |} \\sum_ {(\\boldsymbol {x}, y) \\in U} l (\\boldsymbol {x}, y; h _ {\\hat {T}}) \\right| \\tag {9} \\\\ \\leq \\sqrt {4 - 4 \\xi} (\\lambda^ {l} + \\lambda^ {\\mu} H K) + \\frac {\\overline {{\\boldsymbol {w}}} | U | H}{| \\hat {T} |} + \\sqrt {\\frac {2 H ^ {2} l o g (1 / \\gamma)}{| T |}}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 400, + 890, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "From Theorem 1, we find that the smaller $\\overline{\\boldsymbol{w}}$ and the larger $\\xi$ , the tighter the bound in the SSL error is. Specifically, just as verified in Appendix 5, the pseudo-labeling error bounded by $\\sqrt{4 - 4\\xi} (\\lambda^l +\\lambda^\\mu HK) + \\sqrt{\\frac{2H^2\\log(1 / \\gamma)}{|T|}}$ and the invasion error bounded by $\\frac{\\overline{\\boldsymbol{w}}|U|H}{|\\hat{T}|}$ can be reduced by minimizing the weights $w$ of unlabeled instances with unknown categories and maximizing the confidence of pseudo labels, just as WAD does. Thus, WAD's SSL error has a tight upper bound.", + "bbox": [ + 496, + 525, + 890, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 690, + 632, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Subsection 4.2 presents the comparison results between WAD and five state-of-the-art SSL approaches, as well as one standard baseline. Furthermore, an ablation experiment is conducted in Subsection 4.3, while sensitivity analyses and visualization are carried out in Subsection 4.4 and Subsection 4.5, respectively. For more experiments, please refer to Appendix 4.2 & 4.5.", + "bbox": [ + 496, + 715, + 890, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 500, + 830, + 696, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Our experiments are conducted on two benchmark datasets, CIFAR10 [23] and CIFAR100 [23], as well as an artificial cross-dataset that comprises subsamples", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16415", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/260de664b251c6a1fd12038f4d5209101deb3a37159b56ed23c28d8b25f262c7.jpg", + "image_caption": [ + "(1) Representations" + ], + "image_footnote": [], + "bbox": [ + 86, + 98, + 251, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d24f695f2723ee244b45bbce8e33c06e4f1455f025f3dc1b437a843f811a242.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 145, + 302, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/15f88588df540a25dbcfbfa75841d3003cc5301b7ab0c20c529cc74e79808704.jpg", + "image_caption": [ + "(2) Representations with pseudo labels", + "Figure 6. The schematic diagram of how WAD works. First, each unlabeled instance is assigned a pseudo label according to the max PMI with labeled data in (1), as shown in (2). Then, the instances with wrong labels and unknown categories are assigned a tiny weight to avoid the negative effect, as shown in (3)." + ], + "image_footnote": [], + "bbox": [ + 310, + 101, + 473, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/426970f158d050b18b5f61b483d740728f483fe28c04b40889846aa264359976.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 485, + 146, + 524, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/28e5a25df893d65ff4a28262441006760e553f35c270b3ace622efb8ced6aeff.jpg", + "image_caption": [ + "(3) Representations with pseudo labels and weights" + ], + "image_footnote": [], + "bbox": [ + 535, + 88, + 699, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/38932131b7c55b034e14a03fd0fd5fafafb45aa2136cf5302b0bb206a24032dc.jpg", + "image_caption": [ + "Unlabeled data", + "Unknown categories", + "Cat Dog" + ], + "image_footnote": [], + "bbox": [ + 758, + 90, + 864, + 202 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Weight-aware distillation framework" + ], + "code_body": "Input: Labeled data: $\\mathcal{D}_l$ , Unlabeled data: $\\mathcal{D}_u$ , Max iterations: $N$ , Initial value of $\\alpha$ : $\\alpha_0$ , Set of update steps: $G$ . Output: target classifier $h_{\\hat{T}}$ . \n1 Embedding: $\\mathcal{Z}_l \\gets \\phi(\\mathcal{D}_l)$ , $\\mathcal{Z}_u \\gets \\phi(\\mathcal{D}_u)$ . \n2 Initialize model parameter; \n3 for $t = 0$ to $N - 1$ do \n4 $\\forall x_{i,u} \\in \\mathcal{D}_u$ , distill the knowledge about pseudo label, $\\hat{y}_{i,u}$ , using Eq.5; \n5 $\\forall x_{i,u} \\in \\mathcal{D}_u$ , distill the knowledge about weight, $w_{i,u}$ , using Eq.6; \n6 $\\alpha =$ polynomial Decay $(\\alpha_0)$ ; \n7 Using Eq.7 training the target classifier; \n8 if $t \\in G$ then \n9 $C = \\emptyset$ ; \n10 Calculate the reliability of each unlabeled instance using Eq.8. \n11 Ascending the instances according to reliability and add the top $\\alpha\\%$ instances with pseudo labels to $C$ . \n12 $\\mathcal{D}_l = \\mathcal{D}_l \\cup C$ , $\\mathcal{D}_u = \\mathcal{D}_u \\backslash \\{x | (x,y) \\in C\\}$ . \n13 end \n14 Return target classifier $h_{\\hat{T}}$ .", + "bbox": [ + 81, + 323, + 442, + 675 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "from CIFAR10, CIFAR100, Flowers [29], Food-101 [6], and Places-365 [41]. The CIFAR10 and CIFAR100 datasets consist of 50,000 training and 10,000 testing images of 10 and 100 categories, respectively. The cross-dataset contains 138,000 unlabeled instances from 674 categories. All images from the datasets are resized to $32 \\times 32$ . For further details, please refer to Appendix 4.1.", + "bbox": [ + 75, + 714, + 468, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Settings. i) The proportion of the instances with unknown categories in unlabeled data, named as mismatch proportion, are set as $20\\%$ , $40\\%$ , $60\\%$ , and $80\\%$ in this work. For instance, the unlabeled data has a $60\\%$ mismatch proportion with 4,000 instances with target categories and", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6,000 instances with unknown categories. ii) Randomly sampled $8\\%$ instances from the training dataset that belong to target categories are regarded as labeled data. The remaining $92\\%$ of instances with target categories and some instances with unknown categories are composed of unlabeled data according to the mismatch proportion.", + "bbox": [ + 496, + 303, + 890, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Details. The teacher model is with a Resnet-18 [16] backbone and is trained using SimCLR [10]. And we maintain consistency with SimCLR in all implementation details. The target classifier is a WideResnet-28-2 network [38] with input size $32 \\times 32$ , following Huang et al. [19]. Both the encoder and target classifier are trained from scratch. We train the target classifier using the Adam optimizer [22] with a learning rate of $5 \\times 10^{-4}$ . Furthermore, the epochs and batch size are set as 100 and 32, respectively. The augmentations include random horizontal flipping, random translation by up to 2 pixels, and Gaussian input noise with a standard deviation of 0.15 is used in the training of the target classifier as same as Guo et al. [15]. Moreover, we apply global contrast normalization and ZCA normalized, which is widely used in the pretreatment [15, 11], on CIFAR10. For simplicity, the functions $g_{1}(\\cdot)$ and $g_{2}(\\cdot)$ act as identical mapping with no additional constraints. The initial value of $\\alpha$ is set as 0.1 and decayed five times until it reached 0. It remains the same across all experiments unless otherwise specified. Finally, the approaches on each dataset run three times, and the mean accuracy and standard deviation are reported; the best one is highlighted in bold.", + "bbox": [ + 496, + 395, + 892, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. WAD is compared to five state-of-the-art approaches, including $\\mathrm{DS}^{3}\\mathrm{L}$ [15], T2T [19], CCSSL [37], UASD [11] and ORCA [8], as well as one baseline model that only trains labeled data. Moreover, T2T and ORCA are performed without pretraining tasks for fairness, indicated by \"T2T w\\o pre.\" and \"ORCA w\\o pre\".", + "bbox": [ + 496, + 743, + 890, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 845, + 702, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This subsection presents the experimental results of the classification tasks performed on CIFAR10, CIFAR100,", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16416", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1031d01beef778613eef940e1f4e03759a4e07e822da9af0aa23edc25b5d3f76.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCIFAR10CIFAR100
20%40%60%80%20%40%60%80%
Baseline94.33±0.4594.33±0.4594.33±0.4594.33±0.4536.98±1.7936.98±1.7936.98±1.7936.98±1.79
DS3L91.82±1.8991.38±1.7392.47±0.2590.82±1.5023.92±2.7824.92±4.4126.20±4.2924.55±3.67
UASD95.02±0.7795.03±0.7793.87±0.1393.37±0.3539.85±0.3537.55±2.2436.03±0.7329.87±2.07
CCSSL86.08±0.1284.00±0.1783.13±0.1981.15±0.2541.72 ±0.8541.20±0.5840.60 ±0.2239.67±0.31
T2T----43.70±0.5042.82±0.4540.12±0.7137.35±1.10
T2Tw\\o pre.----39.40±0.3636.78±0.1636.65 ±1.0934.62±1.68
ORCA95.40±0.7494.13±1.1694.35±0.6793.82±0.9329.50±0.2531.12±0.7131.18±0.4031.65±1.86
ORCA w\\o pre.93.32±0.9992.55±2.0292.37±0.9089.65±6.9522.13±1.3323.98±0.7923.37±1.1422.98±0.53
WAD98.43±0.1497.88±0.3397.90±0.2097.77±0.3351.65±2.8650.00±1.4346.88±0.2045.45±1.73
", + "bbox": [ + 84, + 88, + 890, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and a cross-dataset. For CIFAR10, we designated two categories as the target and eight as unknown, while twenty classes are considered as target categories and eighty categories as unknown in CIFAR100. Moreover, we constructed a cross-dataset integrated with five datasets to evaluate WAD in the case that the unlabeled data contains massive unknown categories. Specifically, six classes from CIFAR10 were assigned as target categories, and 668 categories from four external datasets are unknown. The experimental results conducted on CIFAR10, CIFAR100, and the cross-dataset are presented in Table 1 and Table 2.", + "bbox": [ + 75, + 286, + 468, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on CIFAR10 and CIFAR100. From Table 1, we have four findings as follows. i) WAD outperforms all compared methods on CIFAR10 and CIFAR100 with different mismatch proportions, demonstrating its remarkable performance. ii) WAD retains stable performance improvement under different mismatch proportions, exhibiting further improvement of $4.1\\%$ , $3.55\\%$ , $2.91\\%$ , and $3.44\\%$ for mismatch proportions of $20\\%$ , $40\\%$ , $60\\%$ , and $80\\%$ on CIFAR10. This highlights that WAD can achieve robust performance even under a high mismatch proportion. iii) The accuracies of $\\mathrm{DS}^{3}\\mathrm{L}$ on CIFAR10 and CIFAR100 are lower than baseline, as ORCA does. This is because it weights the instances according to consistent empirical risk loss, resulting in the invasion of many unknown categories in training, as shown in Appendix 4.3. iv) In CIFAR100, WAD surpasses baseline $8.47\\%$ for $80\\%$ mismatch proportion. This demonstrates that WAD is still effective when the unlabeled data contains large unknown categories. Therefore, WAD achieves outstanding performance on datasets with different mismatch proportions and exhibits excellent robustness to the scale of unknown categories. Notably, T2T can not apply to the binary classification task, and the accuracy is not reported here.", + "bbox": [ + 75, + 457, + 468, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on cross-dataset. Further, we investigate the limits of WAD's tolerance for unknown categories and then perform the experiments on an artificial cross-dataset containing 668 unknown categories from four datasets. From Table 2, we observe that WAD still maintains an improvement compared to the baseline. Obviously, the other compared", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9f48715d6668b055ca273e79c3614d6913398e25e04a6ba0baef322495d84853.jpg", + "table_caption": [ + "Table 1. Experimental results on CIFAR10 and CIFAR100 under different mismatch proportions." + ], + "table_footnote": [], + "table_body": "
Cross-dataset
Method20%40%60%80%
Baseline66.83±1.3766.83±1.3766.83±1.3766.83±1.37
DS3L50.02±6.6950.69±5.2649.03±5.9351.46±6.99
UASD61.18±0.2957.02±0.5854.70±2.2545.67±1.72
CCSSL64.83±0.2765.15±0.5664.16±0.5864.16±0.45
T2T66.56±2.8065.08±0.7663.76±0.5362.83±0.77
T2T w\\o pre.64.44±0.1562.47±0.7962.23±0.7561.42 ±0.65
ORCA65.53±0.8565.51±1.2566.44±0.8066.46±1.28
ORCA w\\o pre.65.37±0.7863.63±0.6464.42±0.5366.34±1.05
WAD67.13±0.5967.20±1.6567.80±0.0767.88±0.37
", + "bbox": [ + 501, + 284, + 890, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Experimental results on cross-dataset under different mismatch proportions.", + "bbox": [ + 500, + 434, + 890, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "methods were lower than the baseline. This indicates that WAD could boost the performance even on a dataset that contains massive instances with unknown categories.", + "bbox": [ + 498, + 487, + 890, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 542, + 663, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted ablation studies on the CIFAR10 dataset using different models: \"+\"Pse.\" (trained with labeled data and unlabeled instances with pseudo labels), \"+\"Pse.&W.\" (trained with pseudo labels and fix weights), and the WAD model (trained with all components). We also examined the impact of the weight function and explored alternative choices for $g(\\cdot)$ through identical mappings, $g_{i}(\\cdot)$ , and the transformation $\\tilde{g}_{i}(\\cdot) = \\exp (\\cdot)$ . Results are presented in Table 3, and w\\o $g_{i}(\\cdot)$ means removing $g_{i}(\\cdot)$ from Eq.6.", + "bbox": [ + 496, + 566, + 890, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of pseudo labels. From Table 3, we observe that compared with the baseline, $2.72\\%$ and $1.52\\%$ accuracy improvement can be obtained by leveraging the unlabeled instances with pseudo labels, under $20\\%$ and $80\\%$ mismatch proportion, respectively. This indicates that the pseudo labels are beneficial to improving performance.", + "bbox": [ + 496, + 704, + 890, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of weights. According to Table 3, we observe two findings about weights. i) Training by leveraging both pseudo labels and weights exhibits the comparable performance to that without weights under $20\\%$ mismatch proportion. This is because there are fewer instances with unknown categories under $20\\%$ . Then, the model training with fixed weights will result in a sub-optimal model", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16417", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "compared to explicit labels. ii) The weights improve the accuracy by $0.99\\%$ over without it, under $40\\%$ mismatch proportion, while the gap decreases with the mismatch proportion increasing. This demonstrates that the weights are effective in filtering the instances with unknown categories. And the performance degradation is because the absence of the knowledge-update makes the algorithm fail to prevent the increased negative effect from unknown categories.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Effects of knowledge-update. We have two findings according to Table 3. i) The accuracy with knowledge-update surpasses the one only leveraging pseudo labels and weights, and the gap between them reaches $1.77\\%$ under $80\\%$ mismatch proportion. This indicates that knowledge-update plays important roles in WAD. ii) WAD, training with all components, shows its outstanding performance compared to the ones removing other parts. This demonstrates that the aggregation of all the proposed parts could achieve significant improvement.", + "bbox": [ + 75, + 212, + 472, + 362 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Effects of each part of Eq.6. From the Table 3, we have the following two findings. i) both “ $\\mathrm{w} \\backslash \\mathrm{o} g_{1}(\\cdot)$ ” and “ $\\mathrm{w} \\backslash \\mathrm{o} g_{2}(\\cdot)$ ” are worse than WAD, illustrating their equal importance for WAD. ii) Assigning the same mappings to $g_{1}(\\cdot)$ and $g_{2}(\\cdot)$ yields better performance, as the same mappings share the same scales.", + "bbox": [ + 75, + 362, + 470, + 454 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/d61de22678afc4d81c96fe2aa5c72b6b653d5cb731fd6816dfb10db20605c5c0.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Ablation Studies under different mismatch proportions." + ], + "table_body": "
Setting20%40%60%80%
Baseline94.33±0.4594.33±0.4094.33±0.494.33±0.45
+Pse.97.05±0.4895.98±0.7596.65±0.3595.85±0.88
+Pse.&W96.62±0.4796.97±0.7897.22±0.3896.00±0.48
w\\o g1(·)97.85±0.5796.98±0.1194.38±1.5294.38±0.74
w\\o g2(·)97.98±0.5396.85±0.1494.58±0.4695.85±0.35
g1(·) × g2(·)97.50±0.0797.15±0.7194.95±0.1494.48±0.46
g1(·) × g2(·)96.60±0.5796.93±0.0495.65±0.2195.65±0.07
WAD98.43±0.1497.88±0.3397.90±0.2097.77±0.33
", + "bbox": [ + 81, + 462, + 467, + 595 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4. Sensitivity Analysis", + "text_level": 1, + "bbox": [ + 76, + 633, + 264, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This subsection investigates the influence of parameter $\\alpha$ , which controls how many instances with high reliability will be added to labeled data. Hence, we vary the initial value of $\\alpha$ and evaluate WAD's performance on CIFAR10. The results are reported in Table 4. We find that WAD depicts the comparable performance with different values of $\\alpha$ , although lower values of $\\alpha$ achieve slightly better performance with $20\\%$ and $40\\%$ mismatch proportions. This indicates that WAD is not sensitive to $\\alpha$ because the selected instances may have higher similarities. Thus, WAD can achieve a robust performance for a wide range of $\\alpha$ but not too large, preventing the invasion of unknown categories.", + "bbox": [ + 75, + 657, + 472, + 839 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.5. Visualization", + "text_level": 1, + "bbox": [ + 76, + 847, + 215, + 861 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To comprehend how WAD works, we visualize pseudo labels assigned to unlabeled instances with target categories", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/adda62f75bbbd378271a29f55c30d8d164efb5429558dd57caaf8b9665221aad.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Sensitivity analysis under varied mismatch proportions." + ], + "table_body": "
Setting20%40%60%80%
α = 0.0598.62±0.2898.12±0.3997.63±0.0397.63±0.19
α = 0.1098.43±0.1497.88±0.3397.90±0.2097.77±0.33
α = 0.1598.35±0.2597.73±0.1597.93±0.0897.42±0.63
", + "bbox": [ + 506, + 89, + 890, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "alongside the ground truths of labeled ones in different colors, as depicted in the left part of Figure 7. We observe that instances with target categories are separated into two clusters and follow the same distribution as labeled instances. Additionally, the weight distribution, shown on the right side of Figure 7, depicts that WAD assigns smaller weights to unknown categories and larger ones to target ones, making it feasible to filter out harmful unknown categories and to distill useful information from target ones.", + "bbox": [ + 496, + 196, + 893, + 333 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5a4b3ae2308cf15aed7c1a6ef281f7c379a6091b62e8477b7d6baf2acd87748b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 357, + 679, + 445 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fa0514afd8bc2583d6db4a38b94d0a0ebc63b26ac2421cbaa0e7b826bd92c352.jpg", + "image_caption": [ + "Figure 7. Visualization of pseudo labels and weights on CIFAR10 under $60\\%$ class distribution mismatch." + ], + "image_footnote": [], + "bbox": [ + 700, + 356, + 867, + 452 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 503, + 627, + 520 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To tackle class distribution mismatch in an SSL manner, we theoretically reveal that the SSL error is composed of pseudo-labeling error and invasion error under mismatch scenarios. Then, a distillation-based SSL framework, WAD, is proposed to transfer knowledge, such as pseudo labels and weights, from the representations to the target model. Theoretical analyses verify that the population risk of WAD is tightly bounded. Extensive experiments on two benchmark datasets and a cross-dataset demonstrate the superiority of WAD.", + "bbox": [ + 496, + 530, + 890, + 680 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the near future, we would like to investigate whether some instances from unknown categories are beneficial to target task and how to utilize them if so.", + "bbox": [ + 496, + 681, + 893, + 726 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 739, + 679, + 757 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Key Research & Develop Plan(2018YFB1004401), National Natural Science Foundation of China(62276270,62072460, 62172424), Beijing Natural Science Foundation(4212022), Fundamental Research Funds for Central University, and the Research Funds of Renmin University of China. It is also partially supported by the Opening Fund of Hebei Key Laboratory of Machine Learning and Computational Intelligence.", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16418", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? Advances in neural information processing systems, 27, 2014. 3", + "[2] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. arXiv preprint arXiv:1906.00910, 2019. 4", + "[3] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix-match: Semi-supervised learning with distribution matching and augmentation anchoring. In International Conference on Learning Representations, 2019. 2", + "[4] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 2", + "[5] Alexander Borichev and Yuri Tomilov. Optimal polynomial decay of functions and operator semigroups. Mathematische Annalen, 347(2):455-478, 2010. 6", + "[6] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European conference on computer vision, pages 446-461. Springer, 2014. 7", + "[7] Cristian Bucilua, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. In Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 535-541, 2006. 2, 3", + "[8] Kaidi Cao, Maria Brbic, and Jure Leskovec. Open-world semi-supervised learning. In International Conference on Learning Representations, 2021. 1, 3, 7", + "[9] Olivier Chapelle, Bernhard Scholkopf, and Alexander Zien. Semi-supervised learning (chapelle, o. et al., eds.; 2006)[book reviews]. IEEE Transactions on Neural Networks, 20(3):542-542, 2009. 1", + "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 7", + "[11] Yanbei Chen, Xiatian Zhu, Wei Li, and Shaogang Gong. Semi-supervised learning under class distribution mismatch. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 3569-3576, 2020. 1, 2, 7", + "[12] Pan Du, Hui Chen, Suyun Zhao, Shuwen Chai, Hong Chen, and Cuiping Li. Contrastive active learning under class distribution mismatch. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-13, 2022. 1", + "[13] Pan Du, Suyun Zhao, Hui Chen, Shuwen Chai, Hong Chen, and Cuiping Li. Contrastive coding for active learning under class distribution mismatch. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8927-8936, October 2021. 1", + "[14] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004. 1, 2", + "[15] Lan-Zhe Guo, Zhen-Yu Zhang, Yuan Jiang, Yu-Feng Li, and Zhi-Hua Zhou. Safe deep semi-supervised learning for" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "unseen-class unlabeled data. In International Conference on Machine Learning, pages 3897-3906. PMLR, 2020. 1, 2, 7", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7", + "[17] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. In NIPS Deep Learning and Representation Learning Workshop, 2015. 2, 3", + "[18] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In International Conference on Learning Representations, 2018. 4", + "[19] Junkai Huang, Chaowei Fang, Weikai Chen, Zhenhua Chai, Xiaolin Wei, Pengxu Wei, Liang Lin, and Guanbin Li. Trash to treasure: Harvesting ood data with cross-modal matching for open-set semi-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8310-8319, 2021. 1, 2, 7", + "[20] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. A survey on contrastive self-supervised learning. Technologies, 9(1):2, 2021. 4", + "[21] Wonsik Kim, Bhavya Goyal, Kunal Chawla, Jungmin Lee, and Keunjoo Kwon. Attention-based ensemble for deep metric learning. In Proceedings of the European conference on computer vision (ECCV), pages 736-751, 2018. 3", + "[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 7", + "[23] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[24] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 1, 2", + "[25] Phuc H Le-Khac, Graham Healy, and Alan F Smeaton. Contrastive representation learning: A framework and review. IEEE Access, 2020. 4", + "[26] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 2", + "[27] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979–1993, 2018. 2", + "[28] Gaurav Kumar Nayak, Konda Reddy Mopuri, Vaisakh Shaj, Venkatesh Babu Radhakrishnan, and Anirban Chakraborty. Zero-shot knowledge distillation in deep networks. In International Conference on Machine Learning, pages 4743-4751. PMLR, 2019. 2", + "[29] M-E Nilsback and Andrew Zisserman. A visual vocabulary for flower classification. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), volume 2, pages 1447-1454. IEEE, 2006. 7" + ], + "bbox": [ + 503, + 93, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16419", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3967-3976, 2019. 3", + "[31] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems, 29, 2016. 2", + "[32] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 3, 4", + "[33] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 2", + "[34] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 1, 2", + "[35] Lin Wang and Kuk-Jin Yoon. Knowledge distillation and student-teacher learning for visual intelligence: A review and new outlooks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 3", + "[36] Huan Xu and Shie Mannor. Robustness and generalization. Machine learning, 86(3):391-423, 2012. 4", + "[37] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14421-14430, 2022. 1, 2, 7", + "[38] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In *British Machine Vision Conference* 2016. British Machine Vision Association, 2016. 7", + "[39] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 11953-11962, 2022. 3", + "[40] Xujiang Zhao, Killamsetty Krishnateja, Rishabh Iyer, and Feng Chen. Robust semi-supervised learning with out of distribution data. arXiv preprint arXiv:2010.03658, 2020. 3", + "[41] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 7" + ], + "bbox": [ + 78, + 90, + 470, + 742 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16420", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Semi-Supervised Learning via Weight-Aware Distillation under Class Distribution Mismatch/ab302b84-1623-4a04-b35b-84d3ca9ad536_model.json b/2023/Semi-Supervised Learning via Weight-Aware Distillation under Class Distribution Mismatch/ab302b84-1623-4a04-b35b-84d3ca9ad536_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3ff9cfbbdbee52e5751123438d2680ac4b933045 --- /dev/null +++ b/2023/Semi-Supervised Learning via Weight-Aware Distillation under Class Distribution Mismatch/ab302b84-1623-4a04-b35b-84d3ca9ad536_model.json @@ -0,0 +1,2312 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.138, + 0.13, + 0.833, + 0.175 + ], + "angle": 0, + "content": "Semi-Supervised Learning via Weight-aware Distillation under Class Distribution Mismatch" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.203, + 0.868, + 0.257 + ], + "angle": 0, + "content": "Pan Du\\(^{1,2}\\), Suyun Zhao\\(^{1,2,*}\\), Zisen Sheng\\(^{1,2}\\), Cuiping Li\\(^{1,2}\\), Hong Chen\\(^{1,2}\\) \nKey Lab of Data Engineering and Knowledge Engineering of MOE Renmin University of China\\(^{1}\\) \nRenmin University of China, Beijing, China\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.26, + 0.813, + 0.275 + ], + "angle": 0, + "content": "{duPan,shengzisen}@163.com,{zhaosuyun,zisen,licuiping,chong}@ruc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.31, + 0.314, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.474, + 0.751 + ], + "angle": 0, + "content": "Semi-Supervised Learning (SSL) under class distribution mismatch aims to tackle a challenging problem wherein unlabeled data contain lots of unknown categories unseen in the labeled ones. In such mismatch scenarios, traditional SSL suffers severe performance damage due to the harmful invasion of the instances with unknown categories into the target classifier. In this study, by strict mathematical reasoning, we reveal that the SSL error under class distribution mismatch is composed of pseudo-labeling error and invasion error, both of which jointly bound the SSL population risk. To alleviate the SSL error, we propose a robust SSL framework called Weight-Aware Distillation (WAD) that, by weights, selectively transfers knowledge beneficial to the target task from unsupervised contrastive representation to the target classifier. Specifically, WAD captures adaptive weights and high-quality pseudolabels to target instances by exploring point mutual information (PMI) in representation space to maximize the role of unlabeled data and filter unknown categories. Theoretically, we prove that WAD has a tight upper bound of population risk under class distribution mismatch. Experimentally, extensive results demonstrate that WAD outperforms five state-of-the-art SSL approaches and one standard baseline on two benchmark datasets, CIFAR10 and CIFAR100, and an artificial cross-dataset. The code is available at https://github.com/RUC-DWBI-ML/research/tree/main/WAD-master." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.778, + 0.21, + 0.793 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.803, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) have achieved remarkable success in fully-supervised learning tasks. However, sufficient labeled data are usually unavailable in real applications due to the expensive annotation cost or even domain-specific knowledge required [8, 11, 12, 13]. Semi-" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.307, + 0.889, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.488, + 0.892, + 0.517 + ], + "angle": 0, + "content": "Figure 1. Example of class distribution mismatch. The unlabeled data contains categories that are unseen in labeled ones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.893, + 0.791 + ], + "angle": 0, + "content": "supervised learning (SSL), as a powerful weakly-supervised technique, provides an effective way to improve DNNs by exploiting massive unlabeled data, and then it weakens the demand for human annotation [9, 14, 24, 34]. Generally, traditional SSL approaches assume that the labeled and unlabeled instances share the same class distribution, i.e., they come from identical categories. However, in real scenarios, this assumption hardly holds as unlabeled data inevitably contains lots of categories unseen in labeled ones. For instance, if unlabeled data are collected from the internet using keywords \"cat\" and \"dog\" (target categories), they may contain instances unrelated to these categories, such as \"deer,\" \"horse,\" or \"airplane\"(unknown categories), as shown in Figure 1. Similar scenarios occur in medical diagnoses [11, 15] and house annotations of remote-sensing images [12, 13]. SSL in such mismatch scenarios is called SSL under class distribution mismatch [12, 15]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Under class distribution mismatch, some SSL approaches [8, 11, 15, 19, 37] have been proposed. Usually, most of them exploit pseudo-labeling or consistency regularization to expand the labeled pool, as well as filter instances with unknown categories by weights, just as shown in Figure 2. UASD [11] and T2T [19] filter out the instances with unknown categories by leveraging a hard weight, i.e., a" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.232, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16410" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.091, + 0.468, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.179, + 0.468, + 0.194 + ], + "angle": 0, + "content": "Figure 2. The paradigm of SSL under class distribution mismatch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.221, + 0.468, + 0.371 + ], + "angle": 0, + "content": "threshold, on the accumulated network's output or the out-of-distribution score. Although these two approaches reduce the invasion of unknown categories, it is inevitable to keep off amounts of unlabeled instances with target categories. Instead of hard weights, Guo et al. [15] assign a soft weight to the unlabeled instances according to the consistent empirical risk loss. In such case, many instances with unknown categories tend to have consistent outputs and get high weights, just as shown in Appendix 4.3, and then they may invade the target classifier and impair its performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.468, + 0.537 + ], + "angle": 0, + "content": "Moreover, the existing SSL approaches with consistency regularization and pseudo-labeling heavily rely on the performance of the target classifier. Both [15] and [19] annotate pseudo labels by leveraging the prediction of the target classifier in training. Once the target classifier trained on limited labeled instances is biased by some instances with unknown categories, the subsequently updated target classifier may allow more unknown instances to invade. Accordingly, it is promising to propose a novel SSL approach that captures pseudo labels from representations produced by all available data rather than an immature classifier." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.468, + 0.811 + ], + "angle": 0, + "content": "In this study, by strict theoretical analyses, we decouple the SSL error under class distribution mismatch into pseudo-labeling error and invasion error (seen in Subsection 3.2). According to this discovery, a robust SSL framework called weight-aware distillation (WAD) is then proposed to distill pseudo labels and weights from the representation space to the target classifier. Unlike the conventional distillation approaches [7, 17, 28] that simply train the student model using the prediction probability of the teacher model, WAD is a weight-aware distillation framework that adapts to mismatch problems. Specifically, we learn the representations from labeled and unlabeled data by unsupervised contrastive coding, as the teacher model. Then WAD captures adaptive weights as well as high-quality pseudo labels from the teacher model by leveraging point mutual information(PMI), and thus, the target classifier could selectively utilize the instances from target categories while filtering the ones with unknown categories." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.813, + 0.39, + 0.827 + ], + "angle": 0, + "content": "Our main contributions are listed as follows." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.84, + 0.468, + 0.9 + ], + "angle": 0, + "content": "i) We theoretically analyze the population risk in an SSL manner and reveal that the SSL error under class distribution mismatch is jointly controlled by pseudolabeling error and invasion error." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "ii) We propose a distillation-based SSL framework, WAD, that captures weights as well as pseudo labels from robust representations to the target classifier to filter unknown categories and make full use of targeted unlabeled instances as well." + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.179, + 0.892, + 0.239 + ], + "angle": 0, + "content": "iii) Theoretically, we verify that the population risk of WAD is tightly bounded. Experimentally, WAD outperforms five state-of-the-art SSL approaches and one standard baseline on several datasets." + }, + { + "type": "list", + "bbox": [ + 0.506, + 0.092, + 0.892, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.255, + 0.64, + 0.27 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.281, + 0.891, + 0.326 + ], + "angle": 0, + "content": "This section reviews the SSL approaches under class distribution match and mismatch. For contrastive learning, please refer to Appendix 1." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.327, + 0.892, + 0.643 + ], + "angle": 0, + "content": "Semi-Supervised Learning. The traditional SSL strategies include entropy minimization, consistency regularization, and pseudo-label. Entropy minimization [14] incorporates unlabeled data in supervised learning by minimizing the entropy of the unlabeled instance's prediction. The consistency regularization [24, 31, 34] techniques mainly make the prediction on two views of one instance consistent. II-Model [31] focuses on reducing the distance of prediction between one instance and its stochastic perturbation. Unlike the II-Model, temporal ensembling [24] adopts the ensemble of predictions as the target to achieve more stable performance, while Virtual Adversarial Training (VAT) [27] explores adversarial disturbances of the unlabeled instances on the prediction of the target classifier. Pseudo-Labeled based approaches [3, 4, 26, 33] annotate some unlabeled instances with pseudo labels to expand the labeled data. By leveraging the class probability of the unlabeled data, a pseudo-labeling method is proposed [26]. Furthermore, FixMatch [33] uses the weakly augmented unlabeled instances to create a pseudo label and enforce consistent prediction against its strong augmented version." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.644, + 0.891, + 0.688 + ], + "angle": 0, + "content": "These traditional SSL approaches perform well when the class distribution is matched, but they suffer severe performance degradation under class distribution mismatch." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Semi-Supervised Learning under Class Distribution Mismatch. To tackle class distribution mismatch, several studies [11, 15, 19, 37] adopt the traditional SSL strategies with the assistance of soft or hard weights. UASD [11] leverages a threshold to the accumulated network's output to eliminate the instances with unknown categories, followed by pseudo-labeling highly confident ones. Similarly, T2T [19] adopts a hard weight on the out-of-distribution score to conduct filtering and leverages consistency constraints to expand the labeled pool. Furtherly, CCSSL [37] filters out unknown instances by taking both hard and soft weights into consideration. These approaches with hard weights may eliminate too many instances from target categories. Instead of hard weights, \\(\\mathrm{DS}^{3}\\mathrm{L}\\) [15] assigns soft" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16411" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.092, + 0.38, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.091, + 0.716, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.724, + 0.128, + 0.887, + 0.226 + ], + "angle": 0, + "content": "- : Target category 1\n- : Target category 2\n- : Unknown category\n- : Decision boundary\n- >: Knowledge distillation" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.25, + 0.895, + 0.293 + ], + "angle": 0, + "content": "Figure 3. The selective distillation in WAD. Traditional distillation methods force the output of the student model \\((s_i)\\) to align with that of the teacher model \\((t_i)\\), resulting in a wrong decision boundary. However, WAD selectively distills benefit knowledge to student and filter the negatives, such as \\(s_2\\) and \\(s_3\\), by weights to rectify the decision boundary and solve the problem of class distribution mismatch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.317, + 0.471, + 0.392 + ], + "angle": 0, + "content": "weights to unlabeled instances according to the consistent empirical risk loss. However, SSL with pseudo labeling or consistency regularization heavily rely on the performance of the target classifier, and thus they are susceptible to being invaded by instances with unknown categories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.471, + 0.472 + ], + "angle": 0, + "content": "Additionally, a model-level approach [40] is proposed by modifying batch normalization to counter the unknown categories. Also, ORCA [8], a novelty detection approach, leverages uncertainty-based adaptive margins to circumvent the bias caused by the mismatched distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.471, + 0.852 + ], + "angle": 0, + "content": "Knowledge Distillation. Knowledge distillation aims to transfer knowledge from a big model (teacher model) to a smaller one (student model) [35]. It is widely applied to two distinct fields: model compression and knowledge transfer. Model compression is training a small student model to mimic the big teacher model or the ensemble of models. Bucilua et al. [7] compress the ensembles of the neural networks into a single one. While the approaches based on transfer knowledge concentrate more on effectively transferring and are mainly divided into logits-based and representation-based distillation [39]. The logits-based distillation approaches usually train the student model by leveraging the output of the teacher model as the soft label [35]. Ba et al. [1] propose to push the logits, i.e., the output before the softmax function, of the shallow neural network to mimic the ones from a deep neural network. Furtherly, Hinton et al. [17] suggest training a student model to match the combination of the softmax distribution of the teacher model and ground truth. Representation-based approaches enable the student model to learn information from the intermediate layers [35]. Kim et al. [21] propose transferring the attention map from the teacher to the student. Park et al. [30] introduce a novel approach that transfers the mutual relationship of the instances learned from the teacher to the student, similar to our intention." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "However, these approaches mentioned above aim to transfer as much information as possible to the student model and ignore the unknown instances under class dis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.895, + 0.453 + ], + "angle": 0, + "content": "ribution mismatch, which may severely hurt the training of the student model. Unlike the conventional approaches, WAD is a weight-aware distillation framework that selectively transfers the knowledge to the student model, as shown in Figure 3, to fully use the beneficial knowledge and filter the unknown ones by weights. Specifically, WAD distills high-quality pseudo labels to the instances with target categories and filters the instances with unknown categories by assigning them tiny weights." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.465, + 0.591, + 0.48 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.894, + 0.581 + ], + "angle": 0, + "content": "In this section, we propose WAD, an SSL framework under class distribution mismatch. Concretely, Subsection 3.1 introduces the problem statement, followed by analyses of the SSL error in Subsection 3.2. Subsection 3.3 subsequently presents WAD. Finally, theoretical studies of WAD are conducted in Subsection 3.4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.59, + 0.685, + 0.604 + ], + "angle": 0, + "content": "3.1. Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.613, + 0.894, + 0.719 + ], + "angle": 0, + "content": "In this study, we investigate the \\(K\\) classification problem in an SSL manner wherein limited labeled data \\(\\mathcal{D}_l = \\{(x_{i,l},y_{i,l})\\}_{i = 1}^m\\) and massive unlabeled instances \\(\\mathcal{D}_u = \\{x_{i,u}\\}_{i = 1}^n\\) are accessible, \\(x_{i,l}\\in \\mathcal{X}\\) \\(y_{i,l}\\in \\mathcal{Y}\\) \\(\\mathcal{V} = \\{1,\\dots,K\\}\\) and \\(m\\ll n\\). Under class distribution mismatch, the unlabeled instances are not guaranteed to belong to the \\(K\\) target categories in \\(\\mathcal{V}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.729, + 0.743 + ], + "angle": 0, + "content": "3.2. Population Risk Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To make full use of the unlabeled data, we assign a pseudo label to each unlabeled instance, denoted as \\(\\hat{y}\\), and then build the target classifier, \\(h_{\\hat{T}}: \\mathcal{X} \\to \\mathcal{Y}\\), to map the given instance to one of the known categories in \\(\\mathcal{Y}\\), where \\(\\hat{T} = \\{x_{i,l}, y_{i,l}\\}_{i=1}^{m} \\cup \\{x_{i,u}, \\hat{y}_{i,u}\\}_{i=1}^{n}\\), \\(\\hat{y}_{i,u} \\in \\mathcal{Y}\\). Here, \\(\\hat{T}\\) indicates the instances in hand, that is, labeled instances and unlabeled instances assigned with pseudo labels. Then, the population risk [32] of the target classifier learned from both labeled and unlabeled data with the pseudo label \\((\\hat{T})\\) is controlled by the generalization gap, training error, and SSL" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16412" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "error, as shown in Eq.1. The generalization gap is the gap between the population risk and the average prediction loss across all instances with target categories \\((T)\\). Note that \\(T\\) contains all the accessible instances with target categories, including labeled and unlabeled. And every instance in \\(T\\) is assumed with ground truth labels in ideal. The training error is the average empirical loss across \\(\\hat{T}\\). The SSL error is the gap between the average empirical loss across the instances with target categories \\((T)\\) and the average empirical loss across both labeled data and unlabeled ones with pseudo labels \\((\\hat{T})\\). We depict the relations among these sets in Figure 4." + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.289, + 0.47, + 0.478 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbb {E} _ {(\\boldsymbol {x}, \\boldsymbol {y}) \\sim D} [ l (\\boldsymbol {x}, \\boldsymbol {y}; h _ {\\hat {T}}) ] \\\\ \\leq \\underbrace {\\left| \\mathbb {E} _ {(\\boldsymbol {x} , y) \\sim D} [ l (\\boldsymbol {x} , y ; h _ {\\hat {T}}) ] - \\frac {1}{| T |} \\sum_ {(\\boldsymbol {x} , y) \\in T} l (\\boldsymbol {x} , y ; h _ {\\hat {T}}) \\right|} _ {\\text {g e n e r a l i z a t i o n g a p}} \\\\ + \\underbrace {\\left| \\frac {1}{| \\hat {T} |} \\sum_ {(\\boldsymbol {x} , y) \\in \\hat {T}} l (\\boldsymbol {x} , y ; h _ {\\hat {T}}) \\right|} _ {\\text {t r a i n i n g e r r o r}} \\tag {1} \\\\ + \\underbrace {\\left| \\frac {1}{| T |} \\sum_ {(\\pmb {x} , y) \\in T} l (\\pmb {x} , y ; h _ {\\hat {T}}) - \\frac {1}{| \\hat {T} |} \\sum_ {(\\pmb {x} , y) \\in \\hat {T}} l (\\pmb {x} , y ; h _ {\\hat {T}}) \\right|} _ {\\mathrm {S S L e r r o r}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.482, + 0.47, + 0.545 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}\\) is the data distribution of the instances that belong to target categories in the realistic world, i.e., \\(\\mathcal{D} = \\mathcal{X} \\times \\mathcal{Y}\\). \\(l(\\cdot, \\cdot; h_{\\hat{T}}): \\mathcal{X} \\times \\mathcal{Y} \\to \\mathcal{R}\\) denotes the loss function of the classifier \\(h_{\\hat{T}}\\) learned from \\(\\hat{T}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.47, + 0.726 + ], + "angle": 0, + "content": "Theoretical analyses [36] have confirmed that the generalization gap of DNNs can be bounded, and empirical evidence suggests that the training error of DNNs can be reduced almost to zero [32]. Thus, the essential component concerning population risk is the SSL error. Under class distribution mismatch, in addition to the wrongly annotated instances with target categories, the ones with unknown categories also contribute to the SSL error as they invade the training of the target classifier as outliers. Accordingly, we decouple the SSL error into pseudo-labeling and invasion error, as shown in Eq.2. For a detailed derivation process, please refer to Appendix 5.2." + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.731, + 0.469, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left| \\frac {1}{| T |} \\sum_ {(\\boldsymbol {x}, y) \\in T} l (\\boldsymbol {x}, y; h _ {\\hat {T}}) - \\frac {1}{| \\hat {T} |} \\sum_ {(\\boldsymbol {x}, y) \\in \\hat {T}} l (\\boldsymbol {x}, y; h _ {\\hat {T}}) \\right| \\\\ \\leq \\underbrace {\\left| \\frac {1}{| T |} \\sum_ {(\\boldsymbol {x} , y) \\in T} l (\\boldsymbol {x} , y ; h _ {\\hat {T}}) - \\frac {1}{| \\hat {T} |} \\sum_ {(\\boldsymbol {x} , y) \\in \\hat {T} \\setminus U} l (\\boldsymbol {x} , y ; h _ {\\hat {T}}) \\right|} _ {\\textbf {P s e u d o - l a b e l i n g e r r o r}} \\\\ + \\underbrace {\\left| \\frac {1}{| \\hat {T} |} \\sum_ {(\\boldsymbol {x} , y) \\in U} l (\\boldsymbol {x} , y ; h _ {\\hat {T}}) \\right|} _ {\\text {I n v a s i o n e r r o r}} \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.888, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.163, + 0.892, + 0.217 + ], + "angle": 0, + "content": "Figure 4. The relations among data sets \\(\\mathcal{D}_l\\), \\(\\mathcal{D}_u\\), \\(T\\), \\(U\\), \\(\\hat{T}\\). Note that \\(\\hat{T} \\neq T \\cup U\\) due to the instances with target categories in \\(\\mathcal{D}_u\\) are assigned with pseudo labels while not the ground truth ones in \\(T\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.243, + 0.892, + 0.273 + ], + "angle": 0, + "content": "where \\(U\\) indicates the unlabeled instances with unknown categories and \\(\\hat{T} = \\hat{T}\\backslash U\\cup U\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.274, + 0.893, + 0.501 + ], + "angle": 0, + "content": "In Eq.2, the pseudo-labeling error is contributed by the wrongly annotated instances with target categories, as it is the gap of the average empirical loss caused by the inconsistency of the ground truth and pseudo labels. Thus, the quality of pseudo-labels assigned to unlabeled instances within the target distribution determines this error, and accurate pseudo-labeling may alleviate it. By contrast, the invasion error is the average empirical loss across the instances with unknown categories that is caused by the negative effect of those untargeted instances. By Eq.1& Eq.2, we find that the population risk of the target model is jointly controlled by pseudo-labeling error and invasion error. Accordingly, to mitigate the SSL error, we need to filter those instances with unknown categories and accurately annotate the unlabeled instances with target categories as well." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.512, + 0.828, + 0.528 + ], + "angle": 0, + "content": "3.3. Weight-aware Distillation Framework" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.536, + 0.892, + 0.597 + ], + "angle": 0, + "content": "With the aim of mitigating the pseudo-labeling and invasion errors, we design an SSL framework named WAD, which delivers the knowledge of pseudo labels and weights from robust representations to the target classifier." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.619, + 0.719, + 0.635 + ], + "angle": 0, + "content": "3.3.1 Pseudo Label Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Most existing SSL approaches produce pseudo labels by leveraging an immature target classifier, which cause catastrophic error once invaded by some instances with unknown categories, just as discussed in Section 1. To solve this problem, we distill the pseudo labels from a representation space (Teacher model) which is learned from all labeled and unlabeled instances by contrastive learning in an unsupervised manner and then transfer it to the target classifier (Student model). The teacher model could produce closely aligned representations for instances from the same categories and maximize the mutual information among them [2, 18, 20, 25]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Denoted the labeled and unlabeled representations learned by the teacher model, \\(\\phi\\), as \\(\\mathcal{Z}_l = \\{z_{j,l,k}\\}_{j=1}^m\\) and \\(\\mathcal{Z}_u = \\{z_{i,u}\\}_{i=1}^n\\), respectively, where \\(k \\in \\mathcal{V}\\). Inspired by the characteristic of contrastive learning, one effective approach for building the pseudo-label of an unlabeled in-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16413" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.089, + 0.087, + 0.885, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.326, + 0.895, + 0.385 + ], + "angle": 0, + "content": "Figure 5. Illustration of WAD. The pseudo labels and weights determined by point mutual information (PMI) in robust representations of the teacher model are used to participate in training the target classifier. Then, some reliable instances selected according to \\(\\mathcal{L}_{\\mathrm{WAD}}\\) are regarded as labeled ones to update the knowledge gradually. “\\(\\bullet\\)” and “\\(\\bullet\\)” indicates the labeled and unlabeled instance respectively. “//” means stop gradient." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.407, + 0.471, + 0.468 + ], + "angle": 0, + "content": "stance is to identify the labeled instance with the highest PMI and then assign the label of it to unlabeled ones. The PMI between the unlabeled and labeled representation is formulated as Eq.3." + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.477, + 0.47, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\operatorname {P M I} \\left(z _ {i, u}, z _ {j, l, k}\\right) = \\log \\left[ \\frac {p \\left(z _ {j , l , k} \\mid z _ {i , u}\\right)}{p \\left(z _ {j , l , k}\\right)} \\right] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.52, + 0.47, + 0.581 + ], + "angle": 0, + "content": "Although the conditional and marginals distributions, i.e., \\( p(z_{j,l,k}|z_{i,u}) \\) and \\( p(z_{j,l,k}) \\), cannot be directly evaluated, we prove that PMI is proportional to the inner product in Appendix 2, as described in Eq.4." + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.591, + 0.47, + 0.61 + ], + "angle": 0, + "content": "\\[\nf \\left(z _ {i, u}, z _ {j, l, k}\\right) \\propto \\operatorname {P M I} \\left(z _ {i, u}, z _ {j, l, k}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.618, + 0.47, + 0.647 + ], + "angle": 0, + "content": "where \\( f = \\cos(\\cdot, \\cdot) \\), \\( \\| z_i \\| = 1 \\), and \\( \\propto \\) stands for \"proportional to\"." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.649, + 0.427, + 0.664 + ], + "angle": 0, + "content": "Therefore, the pseudo label is formulated as Eq.5." + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.675, + 0.469, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {i, u} = \\underset {k} {\\arg \\max } f \\left(z _ {i, u}, z _ {j, l, k}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.708, + 0.469, + 0.785 + ], + "angle": 0, + "content": "Consequently, the class label of the labeled instance with the maximum PMI is assigned to the unlabeled one. The Eq.5 can precisely capture the PMI from the representation space to produce high-quality pseudo labels and then mitigate pseudo-labeling error." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.343, + 0.818 + ], + "angle": 0, + "content": "3.3.2 Unknown Categories Filtering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To mitigate the invasion error, the instances with unknown categories should be filtered out. Following the Subsection 3.3.1, a higher PMI between the labeled and unlabeled instance suggests a stronger association or similarity between the two instances, further indicating a higher" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.892, + 0.558 + ], + "angle": 0, + "content": "likelihood of the unlabeled instance belonging to the same class distribution as the labeled one. However, some hard instances that have similar PMI between two target categories, i.e., laid on the decision boundary of two target categories, may introduce incorrect pseudo labels and hurt the performance of the target classifier. Hence, we also propose a ratio among the first and second maximum PMI to evaluate the confidence of the pseudo labels. Then, the weight is defined as Eq.6 to avoid the negative effect caused by the wrong labels and unknown categories." + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.566, + 0.892, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {w} _ {i, u} = g _ {1} \\left(\\widetilde {p} _ {i, u}\\right) \\times g _ {2} \\left(1 - \\frac {\\widetilde {q} _ {i , u}}{\\widetilde {p} _ {i , u}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.607, + 0.562, + 0.62 + ], + "angle": 0, + "content": "wherein," + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.62, + 0.772, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\widetilde {p} _ {i, u} = \\max _ {j} f (z _ {i, u}, z _ {j, l, k})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.646, + 0.792, + 0.67 + ], + "angle": 0, + "content": "\\[\n\\widetilde {q} _ {i, u} = \\max _ {v, k \\neq \\hat {y} _ {i, u}} f (z _ {i, u}, z _ {v, l, k})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In Eq.6, \\( g_{1}(\\cdot) \\) and \\( g_{2}(\\cdot) \\) can be interpreted as any monotonically increasing functions. The former in Eq.6 aims to estimate the likelihood of the unlabeled instance belonging to target categories. The higher this item, the more chances of the instance in the target class distribution are. The latter, \\( g_{2}(\\cdot) \\), penalizes instances whose labels are ambiguous between the nearest and second-nearest target categories. The lower this item, the larger probability of incorrect pseudo labels is. As shown in Figure 6, the weight could filter instances with unknown categories and those incorrectly annotated ones with target categories, while the ones from target categories with high-quality pseudo labels are encouraged. Thus, by weight, WAD selectively distills the knowledge beneficial to the target classifier from the teacher model, and the invasion error is then mitigated." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16414" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.092, + 0.393, + 0.108 + ], + "angle": 0, + "content": "3.3.3 Weight-aware Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.25 + ], + "angle": 0, + "content": "Weight-aware knowledge distillation loss. The knowledge of pseudo labels and weights captured from robust representations is applied in the distillation process. In each feed-forward process, pseudo labels and weights are aggregated to the target classifier, as shown in Figure 5. Then, we propose the weight-aware knowledge distillation loss, including the traditional supervised loss \\(\\mathcal{L}_l\\) in labeled data and weight-aware supervised loss \\(\\mathcal{L}_u\\) in unlabeled data as Eq.7." + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.251, + 0.469, + 0.267 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {W A D}} = \\mathcal {L} _ {l} + \\mathcal {L} _ {u}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.273, + 0.138, + 0.286 + ], + "angle": 0, + "content": "wherein," + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.297, + 0.397, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l} = \\frac {1}{| \\mathcal {D} _ {l} |} \\sum_ {(x _ {i, l}, y _ {i, l}) \\in \\mathcal {D} _ {l}} \\ell (h (x _ {i, l}; \\theta), y _ {i, l})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.34, + 0.414, + 0.376 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {u} = \\frac {1}{| \\mathcal {D} _ {u} |} \\sum_ {x _ {i, u} \\in \\mathcal {D} _ {u}} \\boldsymbol {w} _ {i, u} \\ell (h (x _ {i, u}; \\theta), \\hat {y} _ {i, u})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.469, + 0.563 + ], + "angle": 0, + "content": "The traditional supervised loss \\(\\mathcal{L}_l\\) aims to minimize the distance between the predicted probability and the ground truth label. While the weight-aware supervised loss \\(\\mathcal{L}_u\\) mainly focuses on selectively transferring the beneficial knowledge from the teacher model to the student model by weights to mitigate the negative effect from unknown categories and improve the target classifier as well. Moreover, \\(\\mathcal{L}_{\\mathrm{WAD}}\\) is the loss function that is adopted to train the target classifier \\(h_{\\hat{T}}\\) mentioned in Eq.1. Consequently, WAD leverages the pseudo labels and weights to mitigate the pseudolabeling and invasion errors, following alleviating the SSL error, which has been proved in Subsection 3.4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.469, + 0.684 + ], + "angle": 0, + "content": "Knowledge-update in Training. The knowledge of pseudo labels and weights may be biased as the labeled data is limited. Accordingly, after several forward iterations, we progressively add some reliable instances to labeled data. Because the feedback from the target classifier, i.e., loss, is highly related to the weights and reflects the training error, we consider the reliability according to it. Then, the criterion for updating is formulated as Eq.8." + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.694, + 0.469, + 0.71 + ], + "angle": 0, + "content": "\\[\nc _ {i, u} = \\ell \\left(h \\left(x _ {i, u}; \\theta_ {t}\\right), \\hat {y} _ {i, u}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.749 + ], + "angle": 0, + "content": "where \\(\\ell (\\cdot ,\\cdot)\\) is the cross-entropy function, and \\(\\theta_{t}\\) is the parameters of the target classifier in the current iteration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The reliability of \\( x_{i,u} \\) is enhanced when \\( c_{i,u} \\) takes a lower value. Then, WAD leverages Eq.8 to identify the top \\( \\alpha \\% \\) reliable instances from the unlabeled data and puts them in the labeled data while removing them from the unlabeled data. Moreover, we adopt the polynomial decay [5] to dynamically adjust \\( \\alpha \\) to prevent the gradually increased negative effect from unknown categories with the iteration. The details are shown in Appendix 3. A visualization of the number of selected reliable instances with target categories is also provided in Appendix 4.4. Consequently, the pseudo" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "labels and weights are updated in the subsequent distillation steps, as shown in Figure 5, with the aim of optimizing the target classifier. Finally, the schematic diagram and algorithm process is presented in Figure 6 and Algorithm 1, respectively." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.177, + 0.686, + 0.192 + ], + "angle": 0, + "content": "3.4. Theoretical Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.201, + 0.892, + 0.247 + ], + "angle": 0, + "content": "This subsection provides the theoretical studies about the WAD's SSL error, as shown in Theorem 1. Detailed proof of Theorem 1 is given in Appendix 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.893, + 0.392 + ], + "angle": 0, + "content": "Theorem 1 Given \\(|T|\\) instances that i.i.d. sampled from \\(\\mathcal{D}\\) as \\(\\{(x_i, y_i)\\}_{i=1}^{|T|}\\), \\(|U|\\) instances that is not i.i.d with \\(\\mathcal{D}\\), and \\(\\hat{T} = \\{(x, \\hat{y}) | (x, y) \\in T \\cup U, \\hat{y} \\in \\mathcal{Y}\\}\\) where \\(\\mathcal{D} = \\mathcal{X} \\times \\mathcal{Y}\\). Assume the loss function \\(l(\\cdot, y; h_{\\hat{T}})\\) is \\(\\lambda^l\\)-Lipschitz continuous for all \\(y, h_{\\hat{T}}\\) and bounded by \\(H\\), the regression function is \\(\\lambda^\\mu\\)-Lipschitz continuous, training error \\(l(\\pmb{x}, y; h_{\\hat{T}}) = 0\\), \\(\\forall (\\pmb{x}, y) \\in \\hat{T}\\). \\(\\overline{\\pmb{w}}\\) indicates the average of weights, and \\(\\xi\\) is the maximum PMI which determines the pseudo label, with the probability of at least \\(1 - \\gamma\\)," + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.401, + 0.891, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left| \\frac {1}{| T |} \\sum_ {(\\boldsymbol {x}, y) \\in T} l (\\boldsymbol {x}, y; h _ {\\hat {T}}) - \\frac {1}{| \\hat {T} |} \\sum_ {(\\boldsymbol {x}, y) \\in \\hat {T} \\backslash U} l (\\boldsymbol {x}, y; h _ {\\hat {T}}) \\right| \\\\ + \\left| \\frac {1}{| \\hat {T} |} \\sum_ {(\\boldsymbol {x}, y) \\in U} l (\\boldsymbol {x}, y; h _ {\\hat {T}}) \\right| \\tag {9} \\\\ \\leq \\sqrt {4 - 4 \\xi} (\\lambda^ {l} + \\lambda^ {\\mu} H K) + \\frac {\\overline {{\\boldsymbol {w}}} | U | H}{| \\hat {T} |} + \\sqrt {\\frac {2 H ^ {2} l o g (1 / \\gamma)}{| T |}}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.526, + 0.892, + 0.676 + ], + "angle": 0, + "content": "From Theorem 1, we find that the smaller \\(\\overline{\\boldsymbol{w}}\\) and the larger \\(\\xi\\), the tighter the bound in the SSL error is. Specifically, just as verified in Appendix 5, the pseudo-labeling error bounded by \\(\\sqrt{4 - 4\\xi} (\\lambda^l +\\lambda^\\mu HK) + \\sqrt{\\frac{2H^2\\log(1 / \\gamma)}{|T|}}\\) and the invasion error bounded by \\(\\frac{\\overline{\\boldsymbol{w}}|U|H}{|\\hat{T}|}\\) can be reduced by minimizing the weights \\(w\\) of unlabeled instances with unknown categories and maximizing the confidence of pseudo labels, just as WAD does. Thus, WAD's SSL error has a tight upper bound." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.691, + 0.633, + 0.708 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.716, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Subsection 4.2 presents the comparison results between WAD and five state-of-the-art SSL approaches, as well as one standard baseline. Furthermore, an ablation experiment is conducted in Subsection 4.3, while sensitivity analyses and visualization are carried out in Subsection 4.4 and Subsection 4.5, respectively. For more experiments, please refer to Appendix 4.2 & 4.5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.697, + 0.849 + ], + "angle": 0, + "content": "4.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Datasets. Our experiments are conducted on two benchmark datasets, CIFAR10 [23] and CIFAR100 [23], as well as an artificial cross-dataset that comprises subsamples" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16415" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.099, + 0.252, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.221, + 0.207, + 0.232 + ], + "angle": 0, + "content": "(1) Representations" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.146, + 0.303, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.102, + 0.474, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.221, + 0.475, + 0.232 + ], + "angle": 0, + "content": "(2) Representations with pseudo labels" + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.147, + 0.526, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.089, + 0.7, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.221, + 0.726, + 0.232 + ], + "angle": 0, + "content": "(3) Representations with pseudo labels and weights" + }, + { + "type": "image", + "bbox": [ + 0.759, + 0.091, + 0.866, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.76, + 0.203, + 0.866, + 0.211 + ], + "angle": 0, + "content": "Unlabeled data" + }, + { + "type": "image_caption", + "bbox": [ + 0.76, + 0.211, + 0.866, + 0.219 + ], + "angle": 0, + "content": "Unknown categories" + }, + { + "type": "image_caption", + "bbox": [ + 0.76, + 0.22, + 0.847, + 0.231 + ], + "angle": 0, + "content": "Cat Dog" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.237, + 0.895, + 0.28 + ], + "angle": 0, + "content": "Figure 6. The schematic diagram of how WAD works. First, each unlabeled instance is assigned a pseudo label according to the max PMI with labeled data in (1), as shown in (2). Then, the instances with wrong labels and unknown categories are assigned a tiny weight to avoid the negative effect, as shown in (3)." + }, + { + "type": "code_caption", + "bbox": [ + 0.088, + 0.308, + 0.425, + 0.322 + ], + "angle": 0, + "content": "Algorithm 1: Weight-aware distillation framework" + }, + { + "type": "algorithm", + "bbox": [ + 0.082, + 0.324, + 0.443, + 0.676 + ], + "angle": 0, + "content": "Input: Labeled data: \\(\\mathcal{D}_l\\), Unlabeled data: \\(\\mathcal{D}_u\\), Max iterations: \\(N\\), Initial value of \\(\\alpha\\): \\(\\alpha_0\\), Set of update steps: \\(G\\). Output: target classifier \\(h_{\\hat{T}}\\). \n1 Embedding: \\(\\mathcal{Z}_l \\gets \\phi(\\mathcal{D}_l)\\), \\(\\mathcal{Z}_u \\gets \\phi(\\mathcal{D}_u)\\). \n2 Initialize model parameter; \n3 for \\(t = 0\\) to \\(N - 1\\) do \n4 \\(\\forall x_{i,u} \\in \\mathcal{D}_u\\), distill the knowledge about pseudo label, \\(\\hat{y}_{i,u}\\), using Eq.5; \n5 \\(\\forall x_{i,u} \\in \\mathcal{D}_u\\), distill the knowledge about weight, \\(w_{i,u}\\), using Eq.6; \n6 \\(\\alpha =\\) polynomial Decay \\((\\alpha_0)\\); \n7 Using Eq.7 training the target classifier; \n8 if \\(t \\in G\\) then \n9 \\(C = \\emptyset\\); \n10 Calculate the reliability of each unlabeled instance using Eq.8. \n11 Ascending the instances according to reliability and add the top \\(\\alpha\\%\\) instances with pseudo labels to \\(C\\). \n12 \\(\\mathcal{D}_l = \\mathcal{D}_l \\cup C\\), \\(\\mathcal{D}_u = \\mathcal{D}_u \\backslash \\{x | (x,y) \\in C\\}\\). \n13 end \n14 Return target classifier \\(h_{\\hat{T}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.715, + 0.47, + 0.822 + ], + "angle": 0, + "content": "from CIFAR10, CIFAR100, Flowers [29], Food-101 [6], and Places-365 [41]. The CIFAR10 and CIFAR100 datasets consist of 50,000 training and 10,000 testing images of 10 and 100 categories, respectively. The cross-dataset contains 138,000 unlabeled instances from 674 categories. All images from the datasets are resized to \\(32 \\times 32\\). For further details, please refer to Appendix 4.1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Settings. i) The proportion of the instances with unknown categories in unlabeled data, named as mismatch proportion, are set as \\(20\\%\\), \\(40\\%\\), \\(60\\%\\), and \\(80\\%\\) in this work. For instance, the unlabeled data has a \\(60\\%\\) mismatch proportion with 4,000 instances with target categories and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.395 + ], + "angle": 0, + "content": "6,000 instances with unknown categories. ii) Randomly sampled \\(8\\%\\) instances from the training dataset that belong to target categories are regarded as labeled data. The remaining \\(92\\%\\) of instances with target categories and some instances with unknown categories are composed of unlabeled data according to the mismatch proportion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.396, + 0.893, + 0.743 + ], + "angle": 0, + "content": "Details. The teacher model is with a Resnet-18 [16] backbone and is trained using SimCLR [10]. And we maintain consistency with SimCLR in all implementation details. The target classifier is a WideResnet-28-2 network [38] with input size \\( 32 \\times 32 \\), following Huang et al. [19]. Both the encoder and target classifier are trained from scratch. We train the target classifier using the Adam optimizer [22] with a learning rate of \\( 5 \\times 10^{-4} \\). Furthermore, the epochs and batch size are set as 100 and 32, respectively. The augmentations include random horizontal flipping, random translation by up to 2 pixels, and Gaussian input noise with a standard deviation of 0.15 is used in the training of the target classifier as same as Guo et al. [15]. Moreover, we apply global contrast normalization and ZCA normalized, which is widely used in the pretreatment [15, 11], on CIFAR10. For simplicity, the functions \\( g_{1}(\\cdot) \\) and \\( g_{2}(\\cdot) \\) act as identical mapping with no additional constraints. The initial value of \\( \\alpha \\) is set as 0.1 and decayed five times until it reached 0. It remains the same across all experiments unless otherwise specified. Finally, the approaches on each dataset run three times, and the mean accuracy and standard deviation are reported; the best one is highlighted in bold." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.744, + 0.892, + 0.835 + ], + "angle": 0, + "content": "Baselines. WAD is compared to five state-of-the-art approaches, including \\(\\mathrm{DS}^{3}\\mathrm{L}\\) [15], T2T [19], CCSSL [37], UASD [11] and ORCA [8], as well as one baseline model that only trains labeled data. Moreover, T2T and ORCA are performed without pretraining tasks for fairness, indicated by \"T2T w\\o pre.\" and \"ORCA w\\o pre\"." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.704, + 0.863 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This subsection presents the experimental results of the classification tasks performed on CIFAR10, CIFAR100," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16416" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.089, + 0.891, + 0.25 + ], + "angle": 0, + "content": "
MethodCIFAR10CIFAR100
20%40%60%80%20%40%60%80%
Baseline94.33±0.4594.33±0.4594.33±0.4594.33±0.4536.98±1.7936.98±1.7936.98±1.7936.98±1.79
DS3L91.82±1.8991.38±1.7392.47±0.2590.82±1.5023.92±2.7824.92±4.4126.20±4.2924.55±3.67
UASD95.02±0.7795.03±0.7793.87±0.1393.37±0.3539.85±0.3537.55±2.2436.03±0.7329.87±2.07
CCSSL86.08±0.1284.00±0.1783.13±0.1981.15±0.2541.72 ±0.8541.20±0.5840.60 ±0.2239.67±0.31
T2T----43.70±0.5042.82±0.4540.12±0.7137.35±1.10
T2Tw\\o pre.----39.40±0.3636.78±0.1636.65 ±1.0934.62±1.68
ORCA95.40±0.7494.13±1.1694.35±0.6793.82±0.9329.50±0.2531.12±0.7131.18±0.4031.65±1.86
ORCA w\\o pre.93.32±0.9992.55±2.0292.37±0.9089.65±6.9522.13±1.3323.98±0.7923.37±1.1422.98±0.53
WAD98.43±0.1497.88±0.3397.90±0.2097.77±0.3351.65±2.8650.00±1.4346.88±0.2045.45±1.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.198, + 0.25, + 0.772, + 0.262 + ], + "angle": 0, + "content": "Table 1. Experimental results on CIFAR10 and CIFAR100 under different mismatch proportions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.287, + 0.47, + 0.452 + ], + "angle": 0, + "content": "and a cross-dataset. For CIFAR10, we designated two categories as the target and eight as unknown, while twenty classes are considered as target categories and eighty categories as unknown in CIFAR100. Moreover, we constructed a cross-dataset integrated with five datasets to evaluate WAD in the case that the unlabeled data contains massive unknown categories. Specifically, six classes from CIFAR10 were assigned as target categories, and 668 categories from four external datasets are unknown. The experimental results conducted on CIFAR10, CIFAR100, and the cross-dataset are presented in Table 1 and Table 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.47, + 0.805 + ], + "angle": 0, + "content": "Results on CIFAR10 and CIFAR100. From Table 1, we have four findings as follows. i) WAD outperforms all compared methods on CIFAR10 and CIFAR100 with different mismatch proportions, demonstrating its remarkable performance. ii) WAD retains stable performance improvement under different mismatch proportions, exhibiting further improvement of \\(4.1\\%\\), \\(3.55\\%\\), \\(2.91\\%\\), and \\(3.44\\%\\) for mismatch proportions of \\(20\\%\\), \\(40\\%\\), \\(60\\%\\), and \\(80\\%\\) on CIFAR10. This highlights that WAD can achieve robust performance even under a high mismatch proportion. iii) The accuracies of \\(\\mathrm{DS}^{3}\\mathrm{L}\\) on CIFAR10 and CIFAR100 are lower than baseline, as ORCA does. This is because it weights the instances according to consistent empirical risk loss, resulting in the invasion of many unknown categories in training, as shown in Appendix 4.3. iv) In CIFAR100, WAD surpasses baseline \\(8.47\\%\\) for \\(80\\%\\) mismatch proportion. This demonstrates that WAD is still effective when the unlabeled data contains large unknown categories. Therefore, WAD achieves outstanding performance on datasets with different mismatch proportions and exhibits excellent robustness to the scale of unknown categories. Notably, T2T can not apply to the binary classification task, and the accuracy is not reported here." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Results on cross-dataset. Further, we investigate the limits of WAD's tolerance for unknown categories and then perform the experiments on an artificial cross-dataset containing 668 unknown categories from four datasets. From Table 2, we observe that WAD still maintains an improvement compared to the baseline. Obviously, the other compared" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.285, + 0.891, + 0.435 + ], + "angle": 0, + "content": "
Cross-dataset
Method20%40%60%80%
Baseline66.83±1.3766.83±1.3766.83±1.3766.83±1.37
DS3L50.02±6.6950.69±5.2649.03±5.9351.46±6.99
UASD61.18±0.2957.02±0.5854.70±2.2545.67±1.72
CCSSL64.83±0.2765.15±0.5664.16±0.5864.16±0.45
T2T66.56±2.8065.08±0.7663.76±0.5362.83±0.77
T2T w\\o pre.64.44±0.1562.47±0.7962.23±0.7561.42 ±0.65
ORCA65.53±0.8565.51±1.2566.44±0.8066.46±1.28
ORCA w\\o pre.65.37±0.7863.63±0.6464.42±0.5366.34±1.05
WAD67.13±0.5967.20±1.6567.80±0.0767.88±0.37
" + }, + { + "type": "table_caption", + "bbox": [ + 0.501, + 0.435, + 0.891, + 0.462 + ], + "angle": 0, + "content": "Table 2. Experimental results on cross-dataset under different mismatch proportions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.488, + 0.892, + 0.534 + ], + "angle": 0, + "content": "methods were lower than the baseline. This indicates that WAD could boost the performance even on a dataset that contains massive instances with unknown categories." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.544, + 0.665, + 0.559 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.704 + ], + "angle": 0, + "content": "We conducted ablation studies on the CIFAR10 dataset using different models: \"+\"Pse.\" (trained with labeled data and unlabeled instances with pseudo labels), \"+\"Pse.&W.\" (trained with pseudo labels and fix weights), and the WAD model (trained with all components). We also examined the impact of the weight function and explored alternative choices for \\( g(\\cdot) \\) through identical mappings, \\( g_{i}(\\cdot) \\), and the transformation \\( \\tilde{g}_{i}(\\cdot) = \\exp (\\cdot) \\). Results are presented in Table 3, and w\\o \\( g_{i}(\\cdot) \\) means removing \\( g_{i}(\\cdot) \\) from Eq.6." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Effects of pseudo labels. From Table 3, we observe that compared with the baseline, \\(2.72\\%\\) and \\(1.52\\%\\) accuracy improvement can be obtained by leveraging the unlabeled instances with pseudo labels, under \\(20\\%\\) and \\(80\\%\\) mismatch proportion, respectively. This indicates that the pseudo labels are beneficial to improving performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effects of weights. According to Table 3, we observe two findings about weights. i) Training by leveraging both pseudo labels and weights exhibits the comparable performance to that without weights under \\(20\\%\\) mismatch proportion. This is because there are fewer instances with unknown categories under \\(20\\%\\). Then, the model training with fixed weights will result in a sub-optimal model" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16417" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "compared to explicit labels. ii) The weights improve the accuracy by \\(0.99\\%\\) over without it, under \\(40\\%\\) mismatch proportion, while the gap decreases with the mismatch proportion increasing. This demonstrates that the weights are effective in filtering the instances with unknown categories. And the performance degradation is because the absence of the knowledge-update makes the algorithm fail to prevent the increased negative effect from unknown categories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.473, + 0.363 + ], + "angle": 0, + "content": "Effects of knowledge-update. We have two findings according to Table 3. i) The accuracy with knowledge-update surpasses the one only leveraging pseudo labels and weights, and the gap between them reaches \\(1.77\\%\\) under \\(80\\%\\) mismatch proportion. This indicates that knowledge-update plays important roles in WAD. ii) WAD, training with all components, shows its outstanding performance compared to the ones removing other parts. This demonstrates that the aggregation of all the proposed parts could achieve significant improvement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.472, + 0.455 + ], + "angle": 0, + "content": "Effects of each part of Eq.6. From the Table 3, we have the following two findings. i) both “\\(\\mathrm{w} \\backslash \\mathrm{o} g_{1}(\\cdot)\\)” and “\\(\\mathrm{w} \\backslash \\mathrm{o} g_{2}(\\cdot)\\)” are worse than WAD, illustrating their equal importance for WAD. ii) Assigning the same mappings to \\(g_{1}(\\cdot)\\) and \\(g_{2}(\\cdot)\\) yields better performance, as the same mappings share the same scales." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.463, + 0.468, + 0.596 + ], + "angle": 0, + "content": "
Setting20%40%60%80%
Baseline94.33±0.4594.33±0.4094.33±0.494.33±0.45
+Pse.97.05±0.4895.98±0.7596.65±0.3595.85±0.88
+Pse.&W96.62±0.4796.97±0.7897.22±0.3896.00±0.48
w\\o g1(·)97.85±0.5796.98±0.1194.38±1.5294.38±0.74
w\\o g2(·)97.98±0.5396.85±0.1494.58±0.4695.85±0.35
g1(·) × g2(·)97.50±0.0797.15±0.7194.95±0.1494.48±0.46
g1(·) × g2(·)96.60±0.5796.93±0.0495.65±0.2195.65±0.07
WAD98.43±0.1497.88±0.3397.90±0.2097.77±0.33
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.078, + 0.595, + 0.458, + 0.609 + ], + "angle": 0, + "content": "Table 3. Ablation Studies under different mismatch proportions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.634, + 0.265, + 0.65 + ], + "angle": 0, + "content": "4.4. Sensitivity Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.473, + 0.84 + ], + "angle": 0, + "content": "This subsection investigates the influence of parameter \\(\\alpha\\), which controls how many instances with high reliability will be added to labeled data. Hence, we vary the initial value of \\(\\alpha\\) and evaluate WAD's performance on CIFAR10. The results are reported in Table 4. We find that WAD depicts the comparable performance with different values of \\(\\alpha\\), although lower values of \\(\\alpha\\) achieve slightly better performance with \\(20\\%\\) and \\(40\\%\\) mismatch proportions. This indicates that WAD is not sensitive to \\(\\alpha\\) because the selected instances may have higher similarities. Thus, WAD can achieve a robust performance for a wide range of \\(\\alpha\\) but not too large, preventing the invasion of unknown categories." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.216, + 0.862 + ], + "angle": 0, + "content": "4.5. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "To comprehend how WAD works, we visualize pseudo labels assigned to unlabeled instances with target categories" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.09, + 0.892, + 0.161 + ], + "angle": 0, + "content": "
Setting20%40%60%80%
α = 0.0598.62±0.2898.12±0.3997.63±0.0397.63±0.19
α = 0.1098.43±0.1497.88±0.3397.90±0.2097.77±0.33
α = 0.1598.35±0.2597.73±0.1597.93±0.0897.42±0.63
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.501, + 0.162, + 0.882, + 0.175 + ], + "angle": 0, + "content": "Table 4. Sensitivity analysis under varied mismatch proportions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.895, + 0.334 + ], + "angle": 0, + "content": "alongside the ground truths of labeled ones in different colors, as depicted in the left part of Figure 7. We observe that instances with target categories are separated into two clusters and follow the same distribution as labeled instances. Additionally, the weight distribution, shown on the right side of Figure 7, depicts that WAD assigns smaller weights to unknown categories and larger ones to target ones, making it feasible to filter out harmful unknown categories and to distill useful information from target ones." + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.358, + 0.68, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.357, + 0.868, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.453, + 0.894, + 0.481 + ], + "angle": 0, + "content": "Figure 7. Visualization of pseudo labels and weights on CIFAR10 under \\(60\\%\\) class distribution mismatch." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.505, + 0.628, + 0.521 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.892, + 0.681 + ], + "angle": 0, + "content": "To tackle class distribution mismatch in an SSL manner, we theoretically reveal that the SSL error is composed of pseudo-labeling error and invasion error under mismatch scenarios. Then, a distillation-based SSL framework, WAD, is proposed to transfer knowledge, such as pseudo labels and weights, from the representations to the target model. Theoretical analyses verify that the population risk of WAD is tightly bounded. Extensive experiments on two benchmark datasets and a cross-dataset demonstrate the superiority of WAD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.894, + 0.727 + ], + "angle": 0, + "content": "In the near future, we would like to investigate whether some instances from unknown categories are beneficial to target task and how to utilize them if so." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.681, + 0.758 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "This work is supported by the National Key Research & Develop Plan(2018YFB1004401), National Natural Science Foundation of China(62276270,62072460, 62172424), Beijing Natural Science Foundation(4212022), Fundamental Research Funds for Central University, and the Research Funds of Renmin University of China. It is also partially supported by the Opening Fund of Hebei Key Laboratory of Machine Learning and Computational Intelligence." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16418" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.469, + 0.156 + ], + "angle": 0, + "content": "[1] Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? Advances in neural information processing systems, 27, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.469, + 0.2 + ], + "angle": 0, + "content": "[2] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. arXiv preprint arXiv:1906.00910, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.202, + 0.469, + 0.27 + ], + "angle": 0, + "content": "[3] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix-match: Semi-supervised learning with distribution matching and augmentation anchoring. In International Conference on Learning Representations, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[4] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.469, + 0.371 + ], + "angle": 0, + "content": "[5] Alexander Borichev and Yuri Tomilov. Optimal polynomial decay of functions and operator semigroups. Mathematische Annalen, 347(2):455-478, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.469, + 0.429 + ], + "angle": 0, + "content": "[6] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European conference on computer vision, pages 446-461. Springer, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.469, + 0.486 + ], + "angle": 0, + "content": "[7] Cristian Bucilua, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. In Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 535-541, 2006. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.469, + 0.528 + ], + "angle": 0, + "content": "[8] Kaidi Cao, Maria Brbic, and Jure Leskovec. Open-world semi-supervised learning. In International Conference on Learning Representations, 2021. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.469, + 0.585 + ], + "angle": 0, + "content": "[9] Olivier Chapelle, Bernhard Scholkopf, and Alexander Zien. Semi-supervised learning (chapelle, o. et al., eds.; 2006)[book reviews]. IEEE Transactions on Neural Networks, 20(3):542-542, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[11] Yanbei Chen, Xiatian Zhu, Wei Li, and Shaogang Gong. Semi-supervised learning under class distribution mismatch. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 3569-3576, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[12] Pan Du, Hui Chen, Suyun Zhao, Shuwen Chai, Hong Chen, and Cuiping Li. Contrastive active learning under class distribution mismatch. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-13, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[13] Pan Du, Suyun Zhao, Hui Chen, Shuwen Chai, Hong Chen, and Cuiping Li. Contrastive coding for active learning under class distribution mismatch. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8927-8936, October 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[14] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Lan-Zhe Guo, Zhen-Yu Zhang, Yuan Jiang, Yu-Feng Li, and Zhi-Hua Zhou. Safe deep semi-supervised learning for" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "unseen-class unlabeled data. In International Conference on Machine Learning, pages 3897-3906. PMLR, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[17] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. In NIPS Deep Learning and Representation Learning Workshop, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.222, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[18] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In International Conference on Learning Representations, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[19] Junkai Huang, Chaowei Fang, Weikai Chen, Zhenhua Chai, Xiaolin Wei, Pengxu Wei, Liang Lin, and Guanbin Li. Trash to treasure: Harvesting ood data with cross-modal matching for open-set semi-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8310-8319, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[20] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. A survey on contrastive self-supervised learning. Technologies, 9(1):2, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[21] Wonsik Kim, Bhavya Goyal, Kunal Chawla, Jungmin Lee, and Keunjoo Kwon. Attention-based ensemble for deep metric learning. In Proceedings of the European conference on computer vision (ECCV), pages 736-751, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[23] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[24] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[25] Phuc H Le-Khac, Graham Healy, and Alan F Smeaton. Contrastive representation learning: A framework and review. IEEE Access, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.703 + ], + "angle": 0, + "content": "[26] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.705, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[27] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979–1993, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[28] Gaurav Kumar Nayak, Konda Reddy Mopuri, Vaisakh Shaj, Venkatesh Babu Radhakrishnan, and Anirban Chakraborty. Zero-shot knowledge distillation in deep networks. In International Conference on Machine Learning, pages 4743-4751. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] M-E Nilsback and Andrew Zisserman. A visual vocabulary for flower classification. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), volume 2, pages 1447-1454. IEEE, 2006. 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16419" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[30] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3967-3976, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[31] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[32] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.25, + 0.47, + 0.331 + ], + "angle": 0, + "content": "[33] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.333, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[34] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.391, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[35] Lin Wang and Kuk-Jin Yoon. Knowledge distillation and student-teacher learning for visual intelligence: A review and new outlooks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.447, + 0.468, + 0.473 + ], + "angle": 0, + "content": "[36] Huan Xu and Shie Mannor. Robustness and generalization. Machine learning, 86(3):391-423, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.476, + 0.469, + 0.545 + ], + "angle": 0, + "content": "[37] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14421-14430, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.547, + 0.469, + 0.586 + ], + "angle": 0, + "content": "[38] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In *British Machine Vision Conference* 2016. British Machine Vision Association, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.589, + 0.47, + 0.643 + ], + "angle": 0, + "content": "[39] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 11953-11962, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.646, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[40] Xujiang Zhao, Killamsetty Krishnateja, Rishabh Iyer, and Feng Chen. Robust semi-supervised learning with out of distribution data. arXiv preprint arXiv:2010.03658, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.688, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[41] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16420" + } + ] +] \ No newline at end of file diff --git a/2023/Semi-Supervised Learning via Weight-Aware Distillation under Class Distribution Mismatch/ab302b84-1623-4a04-b35b-84d3ca9ad536_origin.pdf b/2023/Semi-Supervised Learning via Weight-Aware Distillation under Class Distribution Mismatch/ab302b84-1623-4a04-b35b-84d3ca9ad536_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f760831dd63c685878b9930559fbb4e154764018 --- /dev/null +++ b/2023/Semi-Supervised Learning via Weight-Aware Distillation under Class Distribution Mismatch/ab302b84-1623-4a04-b35b-84d3ca9ad536_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc3802c269c76ac4c3a93753f5c9bb3bc3ae04727f9d2cda311c7baad9dd1ae1 +size 1571954 diff --git a/2023/Semi-Supervised Learning via Weight-Aware Distillation under Class Distribution Mismatch/full.md b/2023/Semi-Supervised Learning via Weight-Aware Distillation under Class Distribution Mismatch/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a767ee3b85f8c8b7a54fbf4d901dbea69ec49d87 --- /dev/null +++ b/2023/Semi-Supervised Learning via Weight-Aware Distillation under Class Distribution Mismatch/full.md @@ -0,0 +1,370 @@ +# Semi-Supervised Learning via Weight-aware Distillation under Class Distribution Mismatch + +Pan Du $^{1,2}$ , Suyun Zhao $^{1,2,*}$ , Zisen Sheng $^{1,2}$ , Cuiping Li $^{1,2}$ , Hong Chen $^{1,2}$ +Key Lab of Data Engineering and Knowledge Engineering of MOE Renmin University of China $^{1}$ +Renmin University of China, Beijing, China $^{2}$ + +{duPan,shengzisen}@163.com,{zhaosuyun,zisen,licuiping,chong}@ruc.edu.cn + +# Abstract + +Semi-Supervised Learning (SSL) under class distribution mismatch aims to tackle a challenging problem wherein unlabeled data contain lots of unknown categories unseen in the labeled ones. In such mismatch scenarios, traditional SSL suffers severe performance damage due to the harmful invasion of the instances with unknown categories into the target classifier. In this study, by strict mathematical reasoning, we reveal that the SSL error under class distribution mismatch is composed of pseudo-labeling error and invasion error, both of which jointly bound the SSL population risk. To alleviate the SSL error, we propose a robust SSL framework called Weight-Aware Distillation (WAD) that, by weights, selectively transfers knowledge beneficial to the target task from unsupervised contrastive representation to the target classifier. Specifically, WAD captures adaptive weights and high-quality pseudolabels to target instances by exploring point mutual information (PMI) in representation space to maximize the role of unlabeled data and filter unknown categories. Theoretically, we prove that WAD has a tight upper bound of population risk under class distribution mismatch. Experimentally, extensive results demonstrate that WAD outperforms five state-of-the-art SSL approaches and one standard baseline on two benchmark datasets, CIFAR10 and CIFAR100, and an artificial cross-dataset. The code is available at https://github.com/RUC-DWBI-ML/research/tree/main/WAD-master. + +# 1. Introduction + +Deep neural networks (DNNs) have achieved remarkable success in fully-supervised learning tasks. However, sufficient labeled data are usually unavailable in real applications due to the expensive annotation cost or even domain-specific knowledge required [8, 11, 12, 13]. Semi- + +![](images/17e59410467b98a65d776d258b4c60844e60bdac60c7d897a455cc036f59a9d3.jpg) +Figure 1. Example of class distribution mismatch. The unlabeled data contains categories that are unseen in labeled ones. + +supervised learning (SSL), as a powerful weakly-supervised technique, provides an effective way to improve DNNs by exploiting massive unlabeled data, and then it weakens the demand for human annotation [9, 14, 24, 34]. Generally, traditional SSL approaches assume that the labeled and unlabeled instances share the same class distribution, i.e., they come from identical categories. However, in real scenarios, this assumption hardly holds as unlabeled data inevitably contains lots of categories unseen in labeled ones. For instance, if unlabeled data are collected from the internet using keywords "cat" and "dog" (target categories), they may contain instances unrelated to these categories, such as "deer," "horse," or "airplane"(unknown categories), as shown in Figure 1. Similar scenarios occur in medical diagnoses [11, 15] and house annotations of remote-sensing images [12, 13]. SSL in such mismatch scenarios is called SSL under class distribution mismatch [12, 15]. + +Under class distribution mismatch, some SSL approaches [8, 11, 15, 19, 37] have been proposed. Usually, most of them exploit pseudo-labeling or consistency regularization to expand the labeled pool, as well as filter instances with unknown categories by weights, just as shown in Figure 2. UASD [11] and T2T [19] filter out the instances with unknown categories by leveraging a hard weight, i.e., a + +![](images/6a3dfd6d1f31232dc90e781df99df420becd0c1abf22d49a313a685da74a29c6.jpg) +Figure 2. The paradigm of SSL under class distribution mismatch. + +threshold, on the accumulated network's output or the out-of-distribution score. Although these two approaches reduce the invasion of unknown categories, it is inevitable to keep off amounts of unlabeled instances with target categories. Instead of hard weights, Guo et al. [15] assign a soft weight to the unlabeled instances according to the consistent empirical risk loss. In such case, many instances with unknown categories tend to have consistent outputs and get high weights, just as shown in Appendix 4.3, and then they may invade the target classifier and impair its performance. + +Moreover, the existing SSL approaches with consistency regularization and pseudo-labeling heavily rely on the performance of the target classifier. Both [15] and [19] annotate pseudo labels by leveraging the prediction of the target classifier in training. Once the target classifier trained on limited labeled instances is biased by some instances with unknown categories, the subsequently updated target classifier may allow more unknown instances to invade. Accordingly, it is promising to propose a novel SSL approach that captures pseudo labels from representations produced by all available data rather than an immature classifier. + +In this study, by strict theoretical analyses, we decouple the SSL error under class distribution mismatch into pseudo-labeling error and invasion error (seen in Subsection 3.2). According to this discovery, a robust SSL framework called weight-aware distillation (WAD) is then proposed to distill pseudo labels and weights from the representation space to the target classifier. Unlike the conventional distillation approaches [7, 17, 28] that simply train the student model using the prediction probability of the teacher model, WAD is a weight-aware distillation framework that adapts to mismatch problems. Specifically, we learn the representations from labeled and unlabeled data by unsupervised contrastive coding, as the teacher model. Then WAD captures adaptive weights as well as high-quality pseudo labels from the teacher model by leveraging point mutual information(PMI), and thus, the target classifier could selectively utilize the instances from target categories while filtering the ones with unknown categories. + +Our main contributions are listed as follows. + +i) We theoretically analyze the population risk in an SSL manner and reveal that the SSL error under class distribution mismatch is jointly controlled by pseudolabeling error and invasion error. + +ii) We propose a distillation-based SSL framework, WAD, that captures weights as well as pseudo labels from robust representations to the target classifier to filter unknown categories and make full use of targeted unlabeled instances as well. +iii) Theoretically, we verify that the population risk of WAD is tightly bounded. Experimentally, WAD outperforms five state-of-the-art SSL approaches and one standard baseline on several datasets. + +# 2. Related Work + +This section reviews the SSL approaches under class distribution match and mismatch. For contrastive learning, please refer to Appendix 1. + +Semi-Supervised Learning. The traditional SSL strategies include entropy minimization, consistency regularization, and pseudo-label. Entropy minimization [14] incorporates unlabeled data in supervised learning by minimizing the entropy of the unlabeled instance's prediction. The consistency regularization [24, 31, 34] techniques mainly make the prediction on two views of one instance consistent. II-Model [31] focuses on reducing the distance of prediction between one instance and its stochastic perturbation. Unlike the II-Model, temporal ensembling [24] adopts the ensemble of predictions as the target to achieve more stable performance, while Virtual Adversarial Training (VAT) [27] explores adversarial disturbances of the unlabeled instances on the prediction of the target classifier. Pseudo-Labeled based approaches [3, 4, 26, 33] annotate some unlabeled instances with pseudo labels to expand the labeled data. By leveraging the class probability of the unlabeled data, a pseudo-labeling method is proposed [26]. Furthermore, FixMatch [33] uses the weakly augmented unlabeled instances to create a pseudo label and enforce consistent prediction against its strong augmented version. + +These traditional SSL approaches perform well when the class distribution is matched, but they suffer severe performance degradation under class distribution mismatch. + +Semi-Supervised Learning under Class Distribution Mismatch. To tackle class distribution mismatch, several studies [11, 15, 19, 37] adopt the traditional SSL strategies with the assistance of soft or hard weights. UASD [11] leverages a threshold to the accumulated network's output to eliminate the instances with unknown categories, followed by pseudo-labeling highly confident ones. Similarly, T2T [19] adopts a hard weight on the out-of-distribution score to conduct filtering and leverages consistency constraints to expand the labeled pool. Furtherly, CCSSL [37] filters out unknown instances by taking both hard and soft weights into consideration. These approaches with hard weights may eliminate too many instances from target categories. Instead of hard weights, $\mathrm{DS}^{3}\mathrm{L}$ [15] assigns soft + +![](images/e32d869aec5aa08a0452a34edb56cd1157118e18e71a3d72877894b0a4aa87f9.jpg) +Figure 3. The selective distillation in WAD. Traditional distillation methods force the output of the student model $(s_i)$ to align with that of the teacher model $(t_i)$ , resulting in a wrong decision boundary. However, WAD selectively distills benefit knowledge to student and filter the negatives, such as $s_2$ and $s_3$ , by weights to rectify the decision boundary and solve the problem of class distribution mismatch. + +![](images/8110fc305963ff6ee202a9d658d763a695f120893a11b718707935137afb7953.jpg) +- : Target category 1 +- : Target category 2 +- : Unknown category +- : Decision boundary +- >: Knowledge distillation + +weights to unlabeled instances according to the consistent empirical risk loss. However, SSL with pseudo labeling or consistency regularization heavily rely on the performance of the target classifier, and thus they are susceptible to being invaded by instances with unknown categories. + +Additionally, a model-level approach [40] is proposed by modifying batch normalization to counter the unknown categories. Also, ORCA [8], a novelty detection approach, leverages uncertainty-based adaptive margins to circumvent the bias caused by the mismatched distribution. + +Knowledge Distillation. Knowledge distillation aims to transfer knowledge from a big model (teacher model) to a smaller one (student model) [35]. It is widely applied to two distinct fields: model compression and knowledge transfer. Model compression is training a small student model to mimic the big teacher model or the ensemble of models. Bucilua et al. [7] compress the ensembles of the neural networks into a single one. While the approaches based on transfer knowledge concentrate more on effectively transferring and are mainly divided into logits-based and representation-based distillation [39]. The logits-based distillation approaches usually train the student model by leveraging the output of the teacher model as the soft label [35]. Ba et al. [1] propose to push the logits, i.e., the output before the softmax function, of the shallow neural network to mimic the ones from a deep neural network. Furtherly, Hinton et al. [17] suggest training a student model to match the combination of the softmax distribution of the teacher model and ground truth. Representation-based approaches enable the student model to learn information from the intermediate layers [35]. Kim et al. [21] propose transferring the attention map from the teacher to the student. Park et al. [30] introduce a novel approach that transfers the mutual relationship of the instances learned from the teacher to the student, similar to our intention. + +However, these approaches mentioned above aim to transfer as much information as possible to the student model and ignore the unknown instances under class dis + +ribution mismatch, which may severely hurt the training of the student model. Unlike the conventional approaches, WAD is a weight-aware distillation framework that selectively transfers the knowledge to the student model, as shown in Figure 3, to fully use the beneficial knowledge and filter the unknown ones by weights. Specifically, WAD distills high-quality pseudo labels to the instances with target categories and filters the instances with unknown categories by assigning them tiny weights. + +# 3. Method + +In this section, we propose WAD, an SSL framework under class distribution mismatch. Concretely, Subsection 3.1 introduces the problem statement, followed by analyses of the SSL error in Subsection 3.2. Subsection 3.3 subsequently presents WAD. Finally, theoretical studies of WAD are conducted in Subsection 3.4. + +# 3.1. Problem Statement + +In this study, we investigate the $K$ classification problem in an SSL manner wherein limited labeled data $\mathcal{D}_l = \{(x_{i,l},y_{i,l})\}_{i = 1}^m$ and massive unlabeled instances $\mathcal{D}_u = \{x_{i,u}\}_{i = 1}^n$ are accessible, $x_{i,l}\in \mathcal{X}$ $y_{i,l}\in \mathcal{Y}$ $\mathcal{V} = \{1,\dots,K\}$ and $m\ll n$ . Under class distribution mismatch, the unlabeled instances are not guaranteed to belong to the $K$ target categories in $\mathcal{V}$ . + +# 3.2. Population Risk Analysis + +To make full use of the unlabeled data, we assign a pseudo label to each unlabeled instance, denoted as $\hat{y}$ , and then build the target classifier, $h_{\hat{T}}: \mathcal{X} \to \mathcal{Y}$ , to map the given instance to one of the known categories in $\mathcal{Y}$ , where $\hat{T} = \{x_{i,l}, y_{i,l}\}_{i=1}^{m} \cup \{x_{i,u}, \hat{y}_{i,u}\}_{i=1}^{n}$ , $\hat{y}_{i,u} \in \mathcal{Y}$ . Here, $\hat{T}$ indicates the instances in hand, that is, labeled instances and unlabeled instances assigned with pseudo labels. Then, the population risk [32] of the target classifier learned from both labeled and unlabeled data with the pseudo label $(\hat{T})$ is controlled by the generalization gap, training error, and SSL + +error, as shown in Eq.1. The generalization gap is the gap between the population risk and the average prediction loss across all instances with target categories $(T)$ . Note that $T$ contains all the accessible instances with target categories, including labeled and unlabeled. And every instance in $T$ is assumed with ground truth labels in ideal. The training error is the average empirical loss across $\hat{T}$ . The SSL error is the gap between the average empirical loss across the instances with target categories $(T)$ and the average empirical loss across both labeled data and unlabeled ones with pseudo labels $(\hat{T})$ . We depict the relations among these sets in Figure 4. + +$$ +\begin{array}{l} \mathbb {E} _ {(\boldsymbol {x}, \boldsymbol {y}) \sim D} [ l (\boldsymbol {x}, \boldsymbol {y}; h _ {\hat {T}}) ] \\ \leq \underbrace {\left| \mathbb {E} _ {(\boldsymbol {x} , y) \sim D} [ l (\boldsymbol {x} , y ; h _ {\hat {T}}) ] - \frac {1}{| T |} \sum_ {(\boldsymbol {x} , y) \in T} l (\boldsymbol {x} , y ; h _ {\hat {T}}) \right|} _ {\text {g e n e r a l i z a t i o n g a p}} \\ + \underbrace {\left| \frac {1}{| \hat {T} |} \sum_ {(\boldsymbol {x} , y) \in \hat {T}} l (\boldsymbol {x} , y ; h _ {\hat {T}}) \right|} _ {\text {t r a i n i n g e r r o r}} \tag {1} \\ + \underbrace {\left| \frac {1}{| T |} \sum_ {(\pmb {x} , y) \in T} l (\pmb {x} , y ; h _ {\hat {T}}) - \frac {1}{| \hat {T} |} \sum_ {(\pmb {x} , y) \in \hat {T}} l (\pmb {x} , y ; h _ {\hat {T}}) \right|} _ {\mathrm {S S L e r r o r}}, \\ \end{array} +$$ + +where $\mathcal{D}$ is the data distribution of the instances that belong to target categories in the realistic world, i.e., $\mathcal{D} = \mathcal{X} \times \mathcal{Y}$ . $l(\cdot, \cdot; h_{\hat{T}}): \mathcal{X} \times \mathcal{Y} \to \mathcal{R}$ denotes the loss function of the classifier $h_{\hat{T}}$ learned from $\hat{T}$ . + +Theoretical analyses [36] have confirmed that the generalization gap of DNNs can be bounded, and empirical evidence suggests that the training error of DNNs can be reduced almost to zero [32]. Thus, the essential component concerning population risk is the SSL error. Under class distribution mismatch, in addition to the wrongly annotated instances with target categories, the ones with unknown categories also contribute to the SSL error as they invade the training of the target classifier as outliers. Accordingly, we decouple the SSL error into pseudo-labeling and invasion error, as shown in Eq.2. For a detailed derivation process, please refer to Appendix 5.2. + +$$ +\begin{array}{l} \left| \frac {1}{| T |} \sum_ {(\boldsymbol {x}, y) \in T} l (\boldsymbol {x}, y; h _ {\hat {T}}) - \frac {1}{| \hat {T} |} \sum_ {(\boldsymbol {x}, y) \in \hat {T}} l (\boldsymbol {x}, y; h _ {\hat {T}}) \right| \\ \leq \underbrace {\left| \frac {1}{| T |} \sum_ {(\boldsymbol {x} , y) \in T} l (\boldsymbol {x} , y ; h _ {\hat {T}}) - \frac {1}{| \hat {T} |} \sum_ {(\boldsymbol {x} , y) \in \hat {T} \setminus U} l (\boldsymbol {x} , y ; h _ {\hat {T}}) \right|} _ {\textbf {P s e u d o - l a b e l i n g e r r o r}} \\ + \underbrace {\left| \frac {1}{| \hat {T} |} \sum_ {(\boldsymbol {x} , y) \in U} l (\boldsymbol {x} , y ; h _ {\hat {T}}) \right|} _ {\text {I n v a s i o n e r r o r}} \tag {2} \\ \end{array} +$$ + +![](images/bb87031d6d3253f1852ffe8e7123ca4b44767cb6b1624f83879b14e51a1b8d72.jpg) +Figure 4. The relations among data sets $\mathcal{D}_l$ , $\mathcal{D}_u$ , $T$ , $U$ , $\hat{T}$ . Note that $\hat{T} \neq T \cup U$ due to the instances with target categories in $\mathcal{D}_u$ are assigned with pseudo labels while not the ground truth ones in $T$ . + +where $U$ indicates the unlabeled instances with unknown categories and $\hat{T} = \hat{T}\backslash U\cup U$ + +In Eq.2, the pseudo-labeling error is contributed by the wrongly annotated instances with target categories, as it is the gap of the average empirical loss caused by the inconsistency of the ground truth and pseudo labels. Thus, the quality of pseudo-labels assigned to unlabeled instances within the target distribution determines this error, and accurate pseudo-labeling may alleviate it. By contrast, the invasion error is the average empirical loss across the instances with unknown categories that is caused by the negative effect of those untargeted instances. By Eq.1& Eq.2, we find that the population risk of the target model is jointly controlled by pseudo-labeling error and invasion error. Accordingly, to mitigate the SSL error, we need to filter those instances with unknown categories and accurately annotate the unlabeled instances with target categories as well. + +# 3.3. Weight-aware Distillation Framework + +With the aim of mitigating the pseudo-labeling and invasion errors, we design an SSL framework named WAD, which delivers the knowledge of pseudo labels and weights from robust representations to the target classifier. + +# 3.3.1 Pseudo Label Learning + +Most existing SSL approaches produce pseudo labels by leveraging an immature target classifier, which cause catastrophic error once invaded by some instances with unknown categories, just as discussed in Section 1. To solve this problem, we distill the pseudo labels from a representation space (Teacher model) which is learned from all labeled and unlabeled instances by contrastive learning in an unsupervised manner and then transfer it to the target classifier (Student model). The teacher model could produce closely aligned representations for instances from the same categories and maximize the mutual information among them [2, 18, 20, 25]. + +Denoted the labeled and unlabeled representations learned by the teacher model, $\phi$ , as $\mathcal{Z}_l = \{z_{j,l,k}\}_{j=1}^m$ and $\mathcal{Z}_u = \{z_{i,u}\}_{i=1}^n$ , respectively, where $k \in \mathcal{V}$ . Inspired by the characteristic of contrastive learning, one effective approach for building the pseudo-label of an unlabeled in- + +![](images/b7cf7136a5b561ffcc6a96843bf126f9d3e4e021c4974f9e1d3dcac29573989d.jpg) +Figure 5. Illustration of WAD. The pseudo labels and weights determined by point mutual information (PMI) in robust representations of the teacher model are used to participate in training the target classifier. Then, some reliable instances selected according to $\mathcal{L}_{\mathrm{WAD}}$ are regarded as labeled ones to update the knowledge gradually. “ $\bullet$ ” and “ $\bullet$ ” indicates the labeled and unlabeled instance respectively. “//” means stop gradient. + +stance is to identify the labeled instance with the highest PMI and then assign the label of it to unlabeled ones. The PMI between the unlabeled and labeled representation is formulated as Eq.3. + +$$ +\operatorname {P M I} \left(z _ {i, u}, z _ {j, l, k}\right) = \log \left[ \frac {p \left(z _ {j , l , k} \mid z _ {i , u}\right)}{p \left(z _ {j , l , k}\right)} \right] \tag {3} +$$ + +Although the conditional and marginals distributions, i.e., $p(z_{j,l,k}|z_{i,u})$ and $p(z_{j,l,k})$ , cannot be directly evaluated, we prove that PMI is proportional to the inner product in Appendix 2, as described in Eq.4. + +$$ +f \left(z _ {i, u}, z _ {j, l, k}\right) \propto \operatorname {P M I} \left(z _ {i, u}, z _ {j, l, k}\right). \tag {4} +$$ + +where $f = \cos(\cdot, \cdot)$ , $\| z_i \| = 1$ , and $\propto$ stands for "proportional to". + +Therefore, the pseudo label is formulated as Eq.5. + +$$ +\hat {y} _ {i, u} = \underset {k} {\arg \max } f \left(z _ {i, u}, z _ {j, l, k}\right) \tag {5} +$$ + +Consequently, the class label of the labeled instance with the maximum PMI is assigned to the unlabeled one. The Eq.5 can precisely capture the PMI from the representation space to produce high-quality pseudo labels and then mitigate pseudo-labeling error. + +# 3.3.2 Unknown Categories Filtering + +To mitigate the invasion error, the instances with unknown categories should be filtered out. Following the Subsection 3.3.1, a higher PMI between the labeled and unlabeled instance suggests a stronger association or similarity between the two instances, further indicating a higher + +likelihood of the unlabeled instance belonging to the same class distribution as the labeled one. However, some hard instances that have similar PMI between two target categories, i.e., laid on the decision boundary of two target categories, may introduce incorrect pseudo labels and hurt the performance of the target classifier. Hence, we also propose a ratio among the first and second maximum PMI to evaluate the confidence of the pseudo labels. Then, the weight is defined as Eq.6 to avoid the negative effect caused by the wrong labels and unknown categories. + +$$ +\boldsymbol {w} _ {i, u} = g _ {1} \left(\widetilde {p} _ {i, u}\right) \times g _ {2} \left(1 - \frac {\widetilde {q} _ {i , u}}{\widetilde {p} _ {i , u}}\right), \tag {6} +$$ + +wherein, + +$$ +\widetilde {p} _ {i, u} = \max _ {j} f (z _ {i, u}, z _ {j, l, k}) +$$ + +$$ +\widetilde {q} _ {i, u} = \max _ {v, k \neq \hat {y} _ {i, u}} f (z _ {i, u}, z _ {v, l, k}) +$$ + +In Eq.6, $g_{1}(\cdot)$ and $g_{2}(\cdot)$ can be interpreted as any monotonically increasing functions. The former in Eq.6 aims to estimate the likelihood of the unlabeled instance belonging to target categories. The higher this item, the more chances of the instance in the target class distribution are. The latter, $g_{2}(\cdot)$ , penalizes instances whose labels are ambiguous between the nearest and second-nearest target categories. The lower this item, the larger probability of incorrect pseudo labels is. As shown in Figure 6, the weight could filter instances with unknown categories and those incorrectly annotated ones with target categories, while the ones from target categories with high-quality pseudo labels are encouraged. Thus, by weight, WAD selectively distills the knowledge beneficial to the target classifier from the teacher model, and the invasion error is then mitigated. + +# 3.3.3 Weight-aware Knowledge Distillation + +Weight-aware knowledge distillation loss. The knowledge of pseudo labels and weights captured from robust representations is applied in the distillation process. In each feed-forward process, pseudo labels and weights are aggregated to the target classifier, as shown in Figure 5. Then, we propose the weight-aware knowledge distillation loss, including the traditional supervised loss $\mathcal{L}_l$ in labeled data and weight-aware supervised loss $\mathcal{L}_u$ in unlabeled data as Eq.7. + +$$ +\mathcal {L} _ {\mathrm {W A D}} = \mathcal {L} _ {l} + \mathcal {L} _ {u}, \tag {7} +$$ + +wherein, + +$$ +\mathcal {L} _ {l} = \frac {1}{| \mathcal {D} _ {l} |} \sum_ {(x _ {i, l}, y _ {i, l}) \in \mathcal {D} _ {l}} \ell (h (x _ {i, l}; \theta), y _ {i, l}) +$$ + +$$ +\mathcal {L} _ {u} = \frac {1}{| \mathcal {D} _ {u} |} \sum_ {x _ {i, u} \in \mathcal {D} _ {u}} \boldsymbol {w} _ {i, u} \ell (h (x _ {i, u}; \theta), \hat {y} _ {i, u}) +$$ + +The traditional supervised loss $\mathcal{L}_l$ aims to minimize the distance between the predicted probability and the ground truth label. While the weight-aware supervised loss $\mathcal{L}_u$ mainly focuses on selectively transferring the beneficial knowledge from the teacher model to the student model by weights to mitigate the negative effect from unknown categories and improve the target classifier as well. Moreover, $\mathcal{L}_{\mathrm{WAD}}$ is the loss function that is adopted to train the target classifier $h_{\hat{T}}$ mentioned in Eq.1. Consequently, WAD leverages the pseudo labels and weights to mitigate the pseudolabeling and invasion errors, following alleviating the SSL error, which has been proved in Subsection 3.4. + +Knowledge-update in Training. The knowledge of pseudo labels and weights may be biased as the labeled data is limited. Accordingly, after several forward iterations, we progressively add some reliable instances to labeled data. Because the feedback from the target classifier, i.e., loss, is highly related to the weights and reflects the training error, we consider the reliability according to it. Then, the criterion for updating is formulated as Eq.8. + +$$ +c _ {i, u} = \ell \left(h \left(x _ {i, u}; \theta_ {t}\right), \hat {y} _ {i, u}\right) \tag {8} +$$ + +where $\ell (\cdot ,\cdot)$ is the cross-entropy function, and $\theta_{t}$ is the parameters of the target classifier in the current iteration. + +The reliability of $x_{i,u}$ is enhanced when $c_{i,u}$ takes a lower value. Then, WAD leverages Eq.8 to identify the top $\alpha \%$ reliable instances from the unlabeled data and puts them in the labeled data while removing them from the unlabeled data. Moreover, we adopt the polynomial decay [5] to dynamically adjust $\alpha$ to prevent the gradually increased negative effect from unknown categories with the iteration. The details are shown in Appendix 3. A visualization of the number of selected reliable instances with target categories is also provided in Appendix 4.4. Consequently, the pseudo + +labels and weights are updated in the subsequent distillation steps, as shown in Figure 5, with the aim of optimizing the target classifier. Finally, the schematic diagram and algorithm process is presented in Figure 6 and Algorithm 1, respectively. + +# 3.4. Theoretical Studies + +This subsection provides the theoretical studies about the WAD's SSL error, as shown in Theorem 1. Detailed proof of Theorem 1 is given in Appendix 5. + +Theorem 1 Given $|T|$ instances that i.i.d. sampled from $\mathcal{D}$ as $\{(x_i, y_i)\}_{i=1}^{|T|}$ , $|U|$ instances that is not i.i.d with $\mathcal{D}$ , and $\hat{T} = \{(x, \hat{y}) | (x, y) \in T \cup U, \hat{y} \in \mathcal{Y}\}$ where $\mathcal{D} = \mathcal{X} \times \mathcal{Y}$ . Assume the loss function $l(\cdot, y; h_{\hat{T}})$ is $\lambda^l$ -Lipschitz continuous for all $y, h_{\hat{T}}$ and bounded by $H$ , the regression function is $\lambda^\mu$ -Lipschitz continuous, training error $l(\pmb{x}, y; h_{\hat{T}}) = 0$ , $\forall (\pmb{x}, y) \in \hat{T}$ . $\overline{\pmb{w}}$ indicates the average of weights, and $\xi$ is the maximum PMI which determines the pseudo label, with the probability of at least $1 - \gamma$ , + +$$ +\begin{array}{l} \left| \frac {1}{| T |} \sum_ {(\boldsymbol {x}, y) \in T} l (\boldsymbol {x}, y; h _ {\hat {T}}) - \frac {1}{| \hat {T} |} \sum_ {(\boldsymbol {x}, y) \in \hat {T} \backslash U} l (\boldsymbol {x}, y; h _ {\hat {T}}) \right| \\ + \left| \frac {1}{| \hat {T} |} \sum_ {(\boldsymbol {x}, y) \in U} l (\boldsymbol {x}, y; h _ {\hat {T}}) \right| \tag {9} \\ \leq \sqrt {4 - 4 \xi} (\lambda^ {l} + \lambda^ {\mu} H K) + \frac {\overline {{\boldsymbol {w}}} | U | H}{| \hat {T} |} + \sqrt {\frac {2 H ^ {2} l o g (1 / \gamma)}{| T |}}. \\ \end{array} +$$ + +From Theorem 1, we find that the smaller $\overline{\boldsymbol{w}}$ and the larger $\xi$ , the tighter the bound in the SSL error is. Specifically, just as verified in Appendix 5, the pseudo-labeling error bounded by $\sqrt{4 - 4\xi} (\lambda^l +\lambda^\mu HK) + \sqrt{\frac{2H^2\log(1 / \gamma)}{|T|}}$ and the invasion error bounded by $\frac{\overline{\boldsymbol{w}}|U|H}{|\hat{T}|}$ can be reduced by minimizing the weights $w$ of unlabeled instances with unknown categories and maximizing the confidence of pseudo labels, just as WAD does. Thus, WAD's SSL error has a tight upper bound. + +# 4. Experiments + +Subsection 4.2 presents the comparison results between WAD and five state-of-the-art SSL approaches, as well as one standard baseline. Furthermore, an ablation experiment is conducted in Subsection 4.3, while sensitivity analyses and visualization are carried out in Subsection 4.4 and Subsection 4.5, respectively. For more experiments, please refer to Appendix 4.2 & 4.5. + +# 4.1. Experimental Setups + +Datasets. Our experiments are conducted on two benchmark datasets, CIFAR10 [23] and CIFAR100 [23], as well as an artificial cross-dataset that comprises subsamples + +![](images/260de664b251c6a1fd12038f4d5209101deb3a37159b56ed23c28d8b25f262c7.jpg) +(1) Representations + +![](images/1d24f695f2723ee244b45bbce8e33c06e4f1455f025f3dc1b437a843f811a242.jpg) + +![](images/15f88588df540a25dbcfbfa75841d3003cc5301b7ab0c20c529cc74e79808704.jpg) +(2) Representations with pseudo labels +Figure 6. The schematic diagram of how WAD works. First, each unlabeled instance is assigned a pseudo label according to the max PMI with labeled data in (1), as shown in (2). Then, the instances with wrong labels and unknown categories are assigned a tiny weight to avoid the negative effect, as shown in (3). + +![](images/426970f158d050b18b5f61b483d740728f483fe28c04b40889846aa264359976.jpg) + +![](images/28e5a25df893d65ff4a28262441006760e553f35c270b3ace622efb8ced6aeff.jpg) +(3) Representations with pseudo labels and weights + +![](images/38932131b7c55b034e14a03fd0fd5fafafb45aa2136cf5302b0bb206a24032dc.jpg) +Unlabeled data +Unknown categories +Cat Dog + +Algorithm 1: Weight-aware distillation framework +Input: Labeled data: $\mathcal{D}_l$ , Unlabeled data: $\mathcal{D}_u$ , Max iterations: $N$ , Initial value of $\alpha$ : $\alpha_0$ , Set of update steps: $G$ . Output: target classifier $h_{\hat{T}}$ . +1 Embedding: $\mathcal{Z}_l \gets \phi(\mathcal{D}_l)$ , $\mathcal{Z}_u \gets \phi(\mathcal{D}_u)$ . +2 Initialize model parameter; +3 for $t = 0$ to $N - 1$ do +4 $\forall x_{i,u} \in \mathcal{D}_u$ , distill the knowledge about pseudo label, $\hat{y}_{i,u}$ , using Eq.5; +5 $\forall x_{i,u} \in \mathcal{D}_u$ , distill the knowledge about weight, $w_{i,u}$ , using Eq.6; +6 $\alpha =$ polynomial Decay $(\alpha_0)$ ; +7 Using Eq.7 training the target classifier; +8 if $t \in G$ then +9 $C = \emptyset$ ; +10 Calculate the reliability of each unlabeled instance using Eq.8. +11 Ascending the instances according to reliability and add the top $\alpha\%$ instances with pseudo labels to $C$ . +12 $\mathcal{D}_l = \mathcal{D}_l \cup C$ , $\mathcal{D}_u = \mathcal{D}_u \backslash \{x | (x,y) \in C\}$ . +13 end +14 Return target classifier $h_{\hat{T}}$ . + +from CIFAR10, CIFAR100, Flowers [29], Food-101 [6], and Places-365 [41]. The CIFAR10 and CIFAR100 datasets consist of 50,000 training and 10,000 testing images of 10 and 100 categories, respectively. The cross-dataset contains 138,000 unlabeled instances from 674 categories. All images from the datasets are resized to $32 \times 32$ . For further details, please refer to Appendix 4.1. + +Settings. i) The proportion of the instances with unknown categories in unlabeled data, named as mismatch proportion, are set as $20\%$ , $40\%$ , $60\%$ , and $80\%$ in this work. For instance, the unlabeled data has a $60\%$ mismatch proportion with 4,000 instances with target categories and + +6,000 instances with unknown categories. ii) Randomly sampled $8\%$ instances from the training dataset that belong to target categories are regarded as labeled data. The remaining $92\%$ of instances with target categories and some instances with unknown categories are composed of unlabeled data according to the mismatch proportion. + +Details. The teacher model is with a Resnet-18 [16] backbone and is trained using SimCLR [10]. And we maintain consistency with SimCLR in all implementation details. The target classifier is a WideResnet-28-2 network [38] with input size $32 \times 32$ , following Huang et al. [19]. Both the encoder and target classifier are trained from scratch. We train the target classifier using the Adam optimizer [22] with a learning rate of $5 \times 10^{-4}$ . Furthermore, the epochs and batch size are set as 100 and 32, respectively. The augmentations include random horizontal flipping, random translation by up to 2 pixels, and Gaussian input noise with a standard deviation of 0.15 is used in the training of the target classifier as same as Guo et al. [15]. Moreover, we apply global contrast normalization and ZCA normalized, which is widely used in the pretreatment [15, 11], on CIFAR10. For simplicity, the functions $g_{1}(\cdot)$ and $g_{2}(\cdot)$ act as identical mapping with no additional constraints. The initial value of $\alpha$ is set as 0.1 and decayed five times until it reached 0. It remains the same across all experiments unless otherwise specified. Finally, the approaches on each dataset run three times, and the mean accuracy and standard deviation are reported; the best one is highlighted in bold. + +Baselines. WAD is compared to five state-of-the-art approaches, including $\mathrm{DS}^{3}\mathrm{L}$ [15], T2T [19], CCSSL [37], UASD [11] and ORCA [8], as well as one baseline model that only trains labeled data. Moreover, T2T and ORCA are performed without pretraining tasks for fairness, indicated by "T2T w\o pre." and "ORCA w\o pre". + +# 4.2. Experimental Results + +This subsection presents the experimental results of the classification tasks performed on CIFAR10, CIFAR100, + +
MethodCIFAR10CIFAR100
20%40%60%80%20%40%60%80%
Baseline94.33±0.4594.33±0.4594.33±0.4594.33±0.4536.98±1.7936.98±1.7936.98±1.7936.98±1.79
DS3L91.82±1.8991.38±1.7392.47±0.2590.82±1.5023.92±2.7824.92±4.4126.20±4.2924.55±3.67
UASD95.02±0.7795.03±0.7793.87±0.1393.37±0.3539.85±0.3537.55±2.2436.03±0.7329.87±2.07
CCSSL86.08±0.1284.00±0.1783.13±0.1981.15±0.2541.72 ±0.8541.20±0.5840.60 ±0.2239.67±0.31
T2T----43.70±0.5042.82±0.4540.12±0.7137.35±1.10
T2Tw\o pre.----39.40±0.3636.78±0.1636.65 ±1.0934.62±1.68
ORCA95.40±0.7494.13±1.1694.35±0.6793.82±0.9329.50±0.2531.12±0.7131.18±0.4031.65±1.86
ORCA w\o pre.93.32±0.9992.55±2.0292.37±0.9089.65±6.9522.13±1.3323.98±0.7923.37±1.1422.98±0.53
WAD98.43±0.1497.88±0.3397.90±0.2097.77±0.3351.65±2.8650.00±1.4346.88±0.2045.45±1.73
+ +and a cross-dataset. For CIFAR10, we designated two categories as the target and eight as unknown, while twenty classes are considered as target categories and eighty categories as unknown in CIFAR100. Moreover, we constructed a cross-dataset integrated with five datasets to evaluate WAD in the case that the unlabeled data contains massive unknown categories. Specifically, six classes from CIFAR10 were assigned as target categories, and 668 categories from four external datasets are unknown. The experimental results conducted on CIFAR10, CIFAR100, and the cross-dataset are presented in Table 1 and Table 2. + +Results on CIFAR10 and CIFAR100. From Table 1, we have four findings as follows. i) WAD outperforms all compared methods on CIFAR10 and CIFAR100 with different mismatch proportions, demonstrating its remarkable performance. ii) WAD retains stable performance improvement under different mismatch proportions, exhibiting further improvement of $4.1\%$ , $3.55\%$ , $2.91\%$ , and $3.44\%$ for mismatch proportions of $20\%$ , $40\%$ , $60\%$ , and $80\%$ on CIFAR10. This highlights that WAD can achieve robust performance even under a high mismatch proportion. iii) The accuracies of $\mathrm{DS}^{3}\mathrm{L}$ on CIFAR10 and CIFAR100 are lower than baseline, as ORCA does. This is because it weights the instances according to consistent empirical risk loss, resulting in the invasion of many unknown categories in training, as shown in Appendix 4.3. iv) In CIFAR100, WAD surpasses baseline $8.47\%$ for $80\%$ mismatch proportion. This demonstrates that WAD is still effective when the unlabeled data contains large unknown categories. Therefore, WAD achieves outstanding performance on datasets with different mismatch proportions and exhibits excellent robustness to the scale of unknown categories. Notably, T2T can not apply to the binary classification task, and the accuracy is not reported here. + +Results on cross-dataset. Further, we investigate the limits of WAD's tolerance for unknown categories and then perform the experiments on an artificial cross-dataset containing 668 unknown categories from four datasets. From Table 2, we observe that WAD still maintains an improvement compared to the baseline. Obviously, the other compared + +Table 1. Experimental results on CIFAR10 and CIFAR100 under different mismatch proportions. + +
Cross-dataset
Method20%40%60%80%
Baseline66.83±1.3766.83±1.3766.83±1.3766.83±1.37
DS3L50.02±6.6950.69±5.2649.03±5.9351.46±6.99
UASD61.18±0.2957.02±0.5854.70±2.2545.67±1.72
CCSSL64.83±0.2765.15±0.5664.16±0.5864.16±0.45
T2T66.56±2.8065.08±0.7663.76±0.5362.83±0.77
T2T w\o pre.64.44±0.1562.47±0.7962.23±0.7561.42 ±0.65
ORCA65.53±0.8565.51±1.2566.44±0.8066.46±1.28
ORCA w\o pre.65.37±0.7863.63±0.6464.42±0.5366.34±1.05
WAD67.13±0.5967.20±1.6567.80±0.0767.88±0.37
+ +Table 2. Experimental results on cross-dataset under different mismatch proportions. + +methods were lower than the baseline. This indicates that WAD could boost the performance even on a dataset that contains massive instances with unknown categories. + +# 4.3. Ablation Studies + +We conducted ablation studies on the CIFAR10 dataset using different models: "+"Pse." (trained with labeled data and unlabeled instances with pseudo labels), "+"Pse.&W." (trained with pseudo labels and fix weights), and the WAD model (trained with all components). We also examined the impact of the weight function and explored alternative choices for $g(\cdot)$ through identical mappings, $g_{i}(\cdot)$ , and the transformation $\tilde{g}_{i}(\cdot) = \exp (\cdot)$ . Results are presented in Table 3, and w\o $g_{i}(\cdot)$ means removing $g_{i}(\cdot)$ from Eq.6. + +Effects of pseudo labels. From Table 3, we observe that compared with the baseline, $2.72\%$ and $1.52\%$ accuracy improvement can be obtained by leveraging the unlabeled instances with pseudo labels, under $20\%$ and $80\%$ mismatch proportion, respectively. This indicates that the pseudo labels are beneficial to improving performance. + +Effects of weights. According to Table 3, we observe two findings about weights. i) Training by leveraging both pseudo labels and weights exhibits the comparable performance to that without weights under $20\%$ mismatch proportion. This is because there are fewer instances with unknown categories under $20\%$ . Then, the model training with fixed weights will result in a sub-optimal model + +compared to explicit labels. ii) The weights improve the accuracy by $0.99\%$ over without it, under $40\%$ mismatch proportion, while the gap decreases with the mismatch proportion increasing. This demonstrates that the weights are effective in filtering the instances with unknown categories. And the performance degradation is because the absence of the knowledge-update makes the algorithm fail to prevent the increased negative effect from unknown categories. + +Effects of knowledge-update. We have two findings according to Table 3. i) The accuracy with knowledge-update surpasses the one only leveraging pseudo labels and weights, and the gap between them reaches $1.77\%$ under $80\%$ mismatch proportion. This indicates that knowledge-update plays important roles in WAD. ii) WAD, training with all components, shows its outstanding performance compared to the ones removing other parts. This demonstrates that the aggregation of all the proposed parts could achieve significant improvement. + +Effects of each part of Eq.6. From the Table 3, we have the following two findings. i) both “ $\mathrm{w} \backslash \mathrm{o} g_{1}(\cdot)$ ” and “ $\mathrm{w} \backslash \mathrm{o} g_{2}(\cdot)$ ” are worse than WAD, illustrating their equal importance for WAD. ii) Assigning the same mappings to $g_{1}(\cdot)$ and $g_{2}(\cdot)$ yields better performance, as the same mappings share the same scales. + +
Setting20%40%60%80%
Baseline94.33±0.4594.33±0.4094.33±0.494.33±0.45
+Pse.97.05±0.4895.98±0.7596.65±0.3595.85±0.88
+Pse.&W96.62±0.4796.97±0.7897.22±0.3896.00±0.48
w\o g1(·)97.85±0.5796.98±0.1194.38±1.5294.38±0.74
w\o g2(·)97.98±0.5396.85±0.1494.58±0.4695.85±0.35
g1(·) × g2(·)97.50±0.0797.15±0.7194.95±0.1494.48±0.46
g1(·) × g2(·)96.60±0.5796.93±0.0495.65±0.2195.65±0.07
WAD98.43±0.1497.88±0.3397.90±0.2097.77±0.33
+ +Table 3. Ablation Studies under different mismatch proportions. + +# 4.4. Sensitivity Analysis + +This subsection investigates the influence of parameter $\alpha$ , which controls how many instances with high reliability will be added to labeled data. Hence, we vary the initial value of $\alpha$ and evaluate WAD's performance on CIFAR10. The results are reported in Table 4. We find that WAD depicts the comparable performance with different values of $\alpha$ , although lower values of $\alpha$ achieve slightly better performance with $20\%$ and $40\%$ mismatch proportions. This indicates that WAD is not sensitive to $\alpha$ because the selected instances may have higher similarities. Thus, WAD can achieve a robust performance for a wide range of $\alpha$ but not too large, preventing the invasion of unknown categories. + +# 4.5. Visualization + +To comprehend how WAD works, we visualize pseudo labels assigned to unlabeled instances with target categories + +
Setting20%40%60%80%
α = 0.0598.62±0.2898.12±0.3997.63±0.0397.63±0.19
α = 0.1098.43±0.1497.88±0.3397.90±0.2097.77±0.33
α = 0.1598.35±0.2597.73±0.1597.93±0.0897.42±0.63
+ +Table 4. Sensitivity analysis under varied mismatch proportions. + +alongside the ground truths of labeled ones in different colors, as depicted in the left part of Figure 7. We observe that instances with target categories are separated into two clusters and follow the same distribution as labeled instances. Additionally, the weight distribution, shown on the right side of Figure 7, depicts that WAD assigns smaller weights to unknown categories and larger ones to target ones, making it feasible to filter out harmful unknown categories and to distill useful information from target ones. + +![](images/5a4b3ae2308cf15aed7c1a6ef281f7c379a6091b62e8477b7d6baf2acd87748b.jpg) + +![](images/fa0514afd8bc2583d6db4a38b94d0a0ebc63b26ac2421cbaa0e7b826bd92c352.jpg) +Figure 7. Visualization of pseudo labels and weights on CIFAR10 under $60\%$ class distribution mismatch. + +# 5. Conclusions + +To tackle class distribution mismatch in an SSL manner, we theoretically reveal that the SSL error is composed of pseudo-labeling error and invasion error under mismatch scenarios. Then, a distillation-based SSL framework, WAD, is proposed to transfer knowledge, such as pseudo labels and weights, from the representations to the target model. Theoretical analyses verify that the population risk of WAD is tightly bounded. Extensive experiments on two benchmark datasets and a cross-dataset demonstrate the superiority of WAD. + +In the near future, we would like to investigate whether some instances from unknown categories are beneficial to target task and how to utilize them if so. + +# 6. Acknowledgement + +This work is supported by the National Key Research & Develop Plan(2018YFB1004401), National Natural Science Foundation of China(62276270,62072460, 62172424), Beijing Natural Science Foundation(4212022), Fundamental Research Funds for Central University, and the Research Funds of Renmin University of China. It is also partially supported by the Opening Fund of Hebei Key Laboratory of Machine Learning and Computational Intelligence. + +# References + +[1] Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? Advances in neural information processing systems, 27, 2014. 3 +[2] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. arXiv preprint arXiv:1906.00910, 2019. 4 +[3] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix-match: Semi-supervised learning with distribution matching and augmentation anchoring. In International Conference on Learning Representations, 2019. 2 +[4] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 2 +[5] Alexander Borichev and Yuri Tomilov. Optimal polynomial decay of functions and operator semigroups. Mathematische Annalen, 347(2):455-478, 2010. 6 +[6] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European conference on computer vision, pages 446-461. Springer, 2014. 7 +[7] Cristian Bucilua, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. In Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 535-541, 2006. 2, 3 +[8] Kaidi Cao, Maria Brbic, and Jure Leskovec. Open-world semi-supervised learning. In International Conference on Learning Representations, 2021. 1, 3, 7 +[9] Olivier Chapelle, Bernhard Scholkopf, and Alexander Zien. Semi-supervised learning (chapelle, o. et al., eds.; 2006)[book reviews]. IEEE Transactions on Neural Networks, 20(3):542-542, 2009. 1 +[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 7 +[11] Yanbei Chen, Xiatian Zhu, Wei Li, and Shaogang Gong. Semi-supervised learning under class distribution mismatch. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 3569-3576, 2020. 1, 2, 7 +[12] Pan Du, Hui Chen, Suyun Zhao, Shuwen Chai, Hong Chen, and Cuiping Li. Contrastive active learning under class distribution mismatch. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-13, 2022. 1 +[13] Pan Du, Suyun Zhao, Hui Chen, Shuwen Chai, Hong Chen, and Cuiping Li. Contrastive coding for active learning under class distribution mismatch. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8927-8936, October 2021. 1 +[14] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004. 1, 2 +[15] Lan-Zhe Guo, Zhen-Yu Zhang, Yuan Jiang, Yu-Feng Li, and Zhi-Hua Zhou. Safe deep semi-supervised learning for + +unseen-class unlabeled data. In International Conference on Machine Learning, pages 3897-3906. PMLR, 2020. 1, 2, 7 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7 +[17] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. In NIPS Deep Learning and Representation Learning Workshop, 2015. 2, 3 +[18] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In International Conference on Learning Representations, 2018. 4 +[19] Junkai Huang, Chaowei Fang, Weikai Chen, Zhenhua Chai, Xiaolin Wei, Pengxu Wei, Liang Lin, and Guanbin Li. Trash to treasure: Harvesting ood data with cross-modal matching for open-set semi-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8310-8319, 2021. 1, 2, 7 +[20] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. A survey on contrastive self-supervised learning. Technologies, 9(1):2, 2021. 4 +[21] Wonsik Kim, Bhavya Goyal, Kunal Chawla, Jungmin Lee, and Keunjoo Kwon. Attention-based ensemble for deep metric learning. In Proceedings of the European conference on computer vision (ECCV), pages 736-751, 2018. 3 +[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 7 +[23] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[24] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 1, 2 +[25] Phuc H Le-Khac, Graham Healy, and Alan F Smeaton. Contrastive representation learning: A framework and review. IEEE Access, 2020. 4 +[26] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 2 +[27] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979–1993, 2018. 2 +[28] Gaurav Kumar Nayak, Konda Reddy Mopuri, Vaisakh Shaj, Venkatesh Babu Radhakrishnan, and Anirban Chakraborty. Zero-shot knowledge distillation in deep networks. In International Conference on Machine Learning, pages 4743-4751. PMLR, 2019. 2 +[29] M-E Nilsback and Andrew Zisserman. A visual vocabulary for flower classification. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), volume 2, pages 1447-1454. IEEE, 2006. 7 + +[30] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3967-3976, 2019. 3 +[31] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems, 29, 2016. 2 +[32] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 3, 4 +[33] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 2 +[34] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 1, 2 +[35] Lin Wang and Kuk-Jin Yoon. Knowledge distillation and student-teacher learning for visual intelligence: A review and new outlooks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 3 +[36] Huan Xu and Shie Mannor. Robustness and generalization. Machine learning, 86(3):391-423, 2012. 4 +[37] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14421-14430, 2022. 1, 2, 7 +[38] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In *British Machine Vision Conference* 2016. British Machine Vision Association, 2016. 7 +[39] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 11953-11962, 2022. 3 +[40] Xujiang Zhao, Killamsetty Krishnateja, Rishabh Iyer, and Feng Chen. Robust semi-supervised learning with out of distribution data. arXiv preprint arXiv:2010.03658, 2020. 3 +[41] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 7 \ No newline at end of file diff --git a/2023/Semi-Supervised Learning via Weight-Aware Distillation under Class Distribution Mismatch/images.zip b/2023/Semi-Supervised Learning via Weight-Aware Distillation under Class Distribution Mismatch/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c5f1302d2068c1aee1777885663686f7f3b111f9 --- /dev/null +++ b/2023/Semi-Supervised Learning via Weight-Aware Distillation under Class Distribution Mismatch/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7635c2a4d77bdc4ef1b265f72d8e11b9dc6e6a8026fd58ce8a13a87a93e1f009 +size 575725 diff --git a/2023/Semi-Supervised Learning via Weight-Aware Distillation under Class Distribution Mismatch/layout.json b/2023/Semi-Supervised Learning via Weight-Aware Distillation under Class Distribution Mismatch/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dac7c10d2aa9f574ad161adf8c91c27358db5301 --- /dev/null +++ b/2023/Semi-Supervised Learning via Weight-Aware Distillation under Class Distribution Mismatch/layout.json @@ -0,0 +1,10621 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 84, + 102, + 509, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 102, + 509, + 138 + ], + "spans": [ + { + "bbox": [ + 84, + 102, + 509, + 138 + ], + "type": "text", + "content": "Semi-Supervised Learning via Weight-aware Distillation under Class Distribution Mismatch" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 160, + 531, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 160, + 531, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 160, + 531, + 203 + ], + "type": "text", + "content": "Pan Du" + }, + { + "bbox": [ + 62, + 160, + 531, + 203 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 62, + 160, + 531, + 203 + ], + "type": "text", + "content": ", Suyun Zhao" + }, + { + "bbox": [ + 62, + 160, + 531, + 203 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 62, + 160, + 531, + 203 + ], + "type": "text", + "content": ", Zisen Sheng" + }, + { + "bbox": [ + 62, + 160, + 531, + 203 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 62, + 160, + 531, + 203 + ], + "type": "text", + "content": ", Cuiping Li" + }, + { + "bbox": [ + 62, + 160, + 531, + 203 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 62, + 160, + 531, + 203 + ], + "type": "text", + "content": ", Hong Chen" + }, + { + "bbox": [ + 62, + 160, + 531, + 203 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 62, + 160, + 531, + 203 + ], + "type": "text", + "content": " \nKey Lab of Data Engineering and Knowledge Engineering of MOE Renmin University of China" + }, + { + "bbox": [ + 62, + 160, + 531, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 531, + 203 + ], + "type": "text", + "content": " \nRenmin University of China, Beijing, China" + }, + { + "bbox": [ + 62, + 160, + 531, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 90, + 205, + 497, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 205, + 497, + 217 + ], + "spans": [ + { + "bbox": [ + 90, + 205, + 497, + 217 + ], + "type": "text", + "content": "{duPan,shengzisen}@163.com,{zhaosuyun,zisen,licuiping,chong}@ruc.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 245, + 192, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 245, + 192, + 257 + ], + "spans": [ + { + "bbox": [ + 143, + 245, + 192, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 270, + 290, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 290, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 290, + 594 + ], + "type": "text", + "content": "Semi-Supervised Learning (SSL) under class distribution mismatch aims to tackle a challenging problem wherein unlabeled data contain lots of unknown categories unseen in the labeled ones. In such mismatch scenarios, traditional SSL suffers severe performance damage due to the harmful invasion of the instances with unknown categories into the target classifier. In this study, by strict mathematical reasoning, we reveal that the SSL error under class distribution mismatch is composed of pseudo-labeling error and invasion error, both of which jointly bound the SSL population risk. To alleviate the SSL error, we propose a robust SSL framework called Weight-Aware Distillation (WAD) that, by weights, selectively transfers knowledge beneficial to the target task from unsupervised contrastive representation to the target classifier. Specifically, WAD captures adaptive weights and high-quality pseudolabels to target instances by exploring point mutual information (PMI) in representation space to maximize the role of unlabeled data and filter unknown categories. Theoretically, we prove that WAD has a tight upper bound of population risk under class distribution mismatch. Experimentally, extensive results demonstrate that WAD outperforms five state-of-the-art SSL approaches and one standard baseline on two benchmark datasets, CIFAR10 and CIFAR100, and an artificial cross-dataset. The code is available at https://github.com/RUC-DWBI-ML/research/tree/main/WAD-master." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "content": "Deep neural networks (DNNs) have achieved remarkable success in fully-supervised learning tasks. However, sufficient labeled data are usually unavailable in real applications due to the expensive annotation cost or even domain-specific knowledge required [8, 11, 12, 13]. Semi-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 243, + 544, + 383 + ], + "blocks": [ + { + "bbox": [ + 308, + 243, + 544, + 383 + ], + "lines": [ + { + "bbox": [ + 308, + 243, + 544, + 383 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 544, + 383 + ], + "type": "image", + "image_path": "17e59410467b98a65d776d258b4c60844e60bdac60c7d897a455cc036f59a9d3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 386, + 545, + 409 + ], + "lines": [ + { + "bbox": [ + 305, + 386, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 305, + 386, + 545, + 409 + ], + "type": "text", + "content": "Figure 1. Example of class distribution mismatch. The unlabeled data contains categories that are unseen in labeled ones." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 423, + 546, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 546, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 546, + 626 + ], + "type": "text", + "content": "supervised learning (SSL), as a powerful weakly-supervised technique, provides an effective way to improve DNNs by exploiting massive unlabeled data, and then it weakens the demand for human annotation [9, 14, 24, 34]. Generally, traditional SSL approaches assume that the labeled and unlabeled instances share the same class distribution, i.e., they come from identical categories. However, in real scenarios, this assumption hardly holds as unlabeled data inevitably contains lots of categories unseen in labeled ones. For instance, if unlabeled data are collected from the internet using keywords \"cat\" and \"dog\" (target categories), they may contain instances unrelated to these categories, such as \"deer,\" \"horse,\" or \"airplane\"(unknown categories), as shown in Figure 1. Similar scenarios occur in medical diagnoses [11, 15] and house annotations of remote-sensing images [12, 13]. SSL in such mismatch scenarios is called SSL under class distribution mismatch [12, 15]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Under class distribution mismatch, some SSL approaches [8, 11, 15, 19, 37] have been proposed. Usually, most of them exploit pseudo-labeling or consistency regularization to expand the labeled pool, as well as filter instances with unknown categories by weights, just as shown in Figure 2. UASD [11] and T2T [19] filter out the instances with unknown categories by leveraging a hard weight, i.e., a" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 141, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 141, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 141, + 712 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16410" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 286, + 137 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 286, + 137 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 286, + 137 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 286, + 137 + ], + "type": "image", + "image_path": "6a3dfd6d1f31232dc90e781df99df420becd0c1abf22d49a313a685da74a29c6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 141, + 286, + 153 + ], + "lines": [ + { + "bbox": [ + 47, + 141, + 286, + 153 + ], + "spans": [ + { + "bbox": [ + 47, + 141, + 286, + 153 + ], + "type": "text", + "content": "Figure 2. The paradigm of SSL under class distribution mismatch." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 175, + 286, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 286, + 293 + ], + "type": "text", + "content": "threshold, on the accumulated network's output or the out-of-distribution score. Although these two approaches reduce the invasion of unknown categories, it is inevitable to keep off amounts of unlabeled instances with target categories. Instead of hard weights, Guo et al. [15] assign a soft weight to the unlabeled instances according to the consistent empirical risk loss. In such case, many instances with unknown categories tend to have consistent outputs and get high weights, just as shown in Appendix 4.3, and then they may invade the target classifier and impair its performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 294, + 286, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 286, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 286, + 425 + ], + "type": "text", + "content": "Moreover, the existing SSL approaches with consistency regularization and pseudo-labeling heavily rely on the performance of the target classifier. Both [15] and [19] annotate pseudo labels by leveraging the prediction of the target classifier in training. Once the target classifier trained on limited labeled instances is biased by some instances with unknown categories, the subsequently updated target classifier may allow more unknown instances to invade. Accordingly, it is promising to propose a novel SSL approach that captures pseudo labels from representations produced by all available data rather than an immature classifier." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 286, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 286, + 642 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 286, + 642 + ], + "type": "text", + "content": "In this study, by strict theoretical analyses, we decouple the SSL error under class distribution mismatch into pseudo-labeling error and invasion error (seen in Subsection 3.2). According to this discovery, a robust SSL framework called weight-aware distillation (WAD) is then proposed to distill pseudo labels and weights from the representation space to the target classifier. Unlike the conventional distillation approaches [7, 17, 28] that simply train the student model using the prediction probability of the teacher model, WAD is a weight-aware distillation framework that adapts to mismatch problems. Specifically, we learn the representations from labeled and unlabeled data by unsupervised contrastive coding, as the teacher model. Then WAD captures adaptive weights as well as high-quality pseudo labels from the teacher model by leveraging point mutual information(PMI), and thus, the target classifier could selectively utilize the instances from target categories while filtering the ones with unknown categories." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 643, + 238, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 643, + 238, + 654 + ], + "spans": [ + { + "bbox": [ + 59, + 643, + 238, + 654 + ], + "type": "text", + "content": "Our main contributions are listed as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 665, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 665, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 665, + 286, + 712 + ], + "type": "text", + "content": "i) We theoretically analyze the population risk in an SSL manner and reveal that the SSL error under class distribution mismatch is jointly controlled by pseudolabeling error and invasion error." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 309, + 72, + 545, + 189 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 312, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 545, + 131 + ], + "type": "text", + "content": "ii) We propose a distillation-based SSL framework, WAD, that captures weights as well as pseudo labels from robust representations to the target classifier to filter unknown categories and make full use of targeted unlabeled instances as well." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 309, + 141, + 545, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 141, + 545, + 189 + ], + "spans": [ + { + "bbox": [ + 309, + 141, + 545, + 189 + ], + "type": "text", + "content": "iii) Theoretically, we verify that the population risk of WAD is tightly bounded. Experimentally, WAD outperforms five state-of-the-art SSL approaches and one standard baseline on several datasets." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 201, + 391, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 201, + 391, + 213 + ], + "spans": [ + { + "bbox": [ + 306, + 201, + 391, + 213 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 222, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 222, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 305, + 222, + 545, + 258 + ], + "type": "text", + "content": "This section reviews the SSL approaches under class distribution match and mismatch. For contrastive learning, please refer to Appendix 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 258, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 258, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 545, + 509 + ], + "type": "text", + "content": "Semi-Supervised Learning. The traditional SSL strategies include entropy minimization, consistency regularization, and pseudo-label. Entropy minimization [14] incorporates unlabeled data in supervised learning by minimizing the entropy of the unlabeled instance's prediction. The consistency regularization [24, 31, 34] techniques mainly make the prediction on two views of one instance consistent. II-Model [31] focuses on reducing the distance of prediction between one instance and its stochastic perturbation. Unlike the II-Model, temporal ensembling [24] adopts the ensemble of predictions as the target to achieve more stable performance, while Virtual Adversarial Training (VAT) [27] explores adversarial disturbances of the unlabeled instances on the prediction of the target classifier. Pseudo-Labeled based approaches [3, 4, 26, 33] annotate some unlabeled instances with pseudo labels to expand the labeled data. By leveraging the class probability of the unlabeled data, a pseudo-labeling method is proposed [26]. Furthermore, FixMatch [33] uses the weakly augmented unlabeled instances to create a pseudo label and enforce consistent prediction against its strong augmented version." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 510, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 510, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 510, + 545, + 544 + ], + "type": "text", + "content": "These traditional SSL approaches perform well when the class distribution is matched, but they suffer severe performance degradation under class distribution mismatch." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 545, + 545, + 713 + ], + "type": "text", + "content": "Semi-Supervised Learning under Class Distribution Mismatch. To tackle class distribution mismatch, several studies [11, 15, 19, 37] adopt the traditional SSL strategies with the assistance of soft or hard weights. UASD [11] leverages a threshold to the accumulated network's output to eliminate the instances with unknown categories, followed by pseudo-labeling highly confident ones. Similarly, T2T [19] adopts a hard weight on the out-of-distribution score to conduct filtering and leverages consistency constraints to expand the labeled pool. Furtherly, CCSSL [37] filters out unknown instances by taking both hard and soft weights into consideration. These approaches with hard weights may eliminate too many instances from target categories. Instead of hard weights, " + }, + { + "bbox": [ + 305, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{DS}^{3}\\mathrm{L}" + }, + { + "bbox": [ + 305, + 545, + 545, + 713 + ], + "type": "text", + "content": " [15] assigns soft" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16411" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 72, + 232, + 193 + ], + "blocks": [ + { + "bbox": [ + 55, + 72, + 232, + 193 + ], + "lines": [ + { + "bbox": [ + 55, + 72, + 232, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 232, + 193 + ], + "type": "image", + "image_path": "e32d869aec5aa08a0452a34edb56cd1157118e18e71a3d72877894b0a4aa87f9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "text", + "content": "Figure 3. The selective distillation in WAD. Traditional distillation methods force the output of the student model " + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "inline_equation", + "content": "(s_i)" + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "text", + "content": " to align with that of the teacher model " + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "inline_equation", + "content": "(t_i)" + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "text", + "content": ", resulting in a wrong decision boundary. However, WAD selectively distills benefit knowledge to student and filter the negatives, such as " + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "inline_equation", + "content": "s_2" + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "inline_equation", + "content": "s_3" + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "text", + "content": ", by weights to rectify the decision boundary and solve the problem of class distribution mismatch." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 246, + 72, + 438, + 191 + ], + "blocks": [ + { + "bbox": [ + 246, + 72, + 438, + 191 + ], + "lines": [ + { + "bbox": [ + 246, + 72, + 438, + 191 + ], + "spans": [ + { + "bbox": [ + 246, + 72, + 438, + 191 + ], + "type": "image", + "image_path": "8110fc305963ff6ee202a9d658d763a695f120893a11b718707935137afb7953.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 101, + 542, + 178 + ], + "lines": [ + { + "bbox": [ + 443, + 101, + 542, + 178 + ], + "spans": [ + { + "bbox": [ + 443, + 101, + 542, + 178 + ], + "type": "text", + "content": "- : Target category 1\n- : Target category 2\n- : Unknown category\n- : Decision boundary\n- >: Knowledge distillation" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 251, + 288, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 288, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 288, + 310 + ], + "type": "text", + "content": "weights to unlabeled instances according to the consistent empirical risk loss. However, SSL with pseudo labeling or consistency regularization heavily rely on the performance of the target classifier, and thus they are susceptible to being invaded by instances with unknown categories." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 313, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 288, + 373 + ], + "type": "text", + "content": "Additionally, a model-level approach [40] is proposed by modifying batch normalization to counter the unknown categories. Also, ORCA [8], a novelty detection approach, leverages uncertainty-based adaptive margins to circumvent the bias caused by the mismatched distribution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 376, + 288, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 288, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 288, + 674 + ], + "type": "text", + "content": "Knowledge Distillation. Knowledge distillation aims to transfer knowledge from a big model (teacher model) to a smaller one (student model) [35]. It is widely applied to two distinct fields: model compression and knowledge transfer. Model compression is training a small student model to mimic the big teacher model or the ensemble of models. Bucilua et al. [7] compress the ensembles of the neural networks into a single one. While the approaches based on transfer knowledge concentrate more on effectively transferring and are mainly divided into logits-based and representation-based distillation [39]. The logits-based distillation approaches usually train the student model by leveraging the output of the teacher model as the soft label [35]. Ba et al. [1] propose to push the logits, i.e., the output before the softmax function, of the shallow neural network to mimic the ones from a deep neural network. Furtherly, Hinton et al. [17] suggest training a student model to match the combination of the softmax distribution of the teacher model and ground truth. Representation-based approaches enable the student model to learn information from the intermediate layers [35]. Kim et al. [21] propose transferring the attention map from the teacher to the student. Park et al. [30] introduce a novel approach that transfers the mutual relationship of the instances learned from the teacher to the student, similar to our intention." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": "However, these approaches mentioned above aim to transfer as much information as possible to the student model and ignore the unknown instances under class dis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 251, + 547, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 547, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 547, + 358 + ], + "type": "text", + "content": "ribution mismatch, which may severely hurt the training of the student model. Unlike the conventional approaches, WAD is a weight-aware distillation framework that selectively transfers the knowledge to the student model, as shown in Figure 3, to fully use the beneficial knowledge and filter the unknown ones by weights. Specifically, WAD distills high-quality pseudo labels to the instances with target categories and filters the instances with unknown categories by assigning them tiny weights." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 368, + 361, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 368, + 361, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 368, + 361, + 380 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 388, + 547, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 547, + 460 + ], + "type": "text", + "content": "In this section, we propose WAD, an SSL framework under class distribution mismatch. Concretely, Subsection 3.1 introduces the problem statement, followed by analyses of the SSL error in Subsection 3.2. Subsection 3.3 subsequently presents WAD. Finally, theoretical studies of WAD are conducted in Subsection 3.4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 467, + 419, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 467, + 419, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 467, + 419, + 478 + ], + "type": "text", + "content": "3.1. Problem Statement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 485, + 547, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 485, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 305, + 485, + 547, + 569 + ], + "type": "text", + "content": "In this study, we investigate the " + }, + { + "bbox": [ + 305, + 485, + 547, + 569 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 485, + 547, + 569 + ], + "type": "text", + "content": " classification problem in an SSL manner wherein limited labeled data " + }, + { + "bbox": [ + 305, + 485, + 547, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_l = \\{(x_{i,l},y_{i,l})\\}_{i = 1}^m" + }, + { + "bbox": [ + 305, + 485, + 547, + 569 + ], + "type": "text", + "content": " and massive unlabeled instances " + }, + { + "bbox": [ + 305, + 485, + 547, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u = \\{x_{i,u}\\}_{i = 1}^n" + }, + { + "bbox": [ + 305, + 485, + 547, + 569 + ], + "type": "text", + "content": " are accessible, " + }, + { + "bbox": [ + 305, + 485, + 547, + 569 + ], + "type": "inline_equation", + "content": "x_{i,l}\\in \\mathcal{X}" + }, + { + "bbox": [ + 305, + 485, + 547, + 569 + ], + "type": "inline_equation", + "content": "y_{i,l}\\in \\mathcal{Y}" + }, + { + "bbox": [ + 305, + 485, + 547, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{1,\\dots,K\\}" + }, + { + "bbox": [ + 305, + 485, + 547, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 485, + 547, + 569 + ], + "type": "inline_equation", + "content": "m\\ll n" + }, + { + "bbox": [ + 305, + 485, + 547, + 569 + ], + "type": "text", + "content": ". Under class distribution mismatch, the unlabeled instances are not guaranteed to belong to the " + }, + { + "bbox": [ + 305, + 485, + 547, + 569 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 485, + 547, + 569 + ], + "type": "text", + "content": " target categories in " + }, + { + "bbox": [ + 305, + 485, + 547, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 305, + 485, + 547, + 569 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 575, + 446, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 446, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 446, + 588 + ], + "type": "text", + "content": "3.2. Population Risk Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "To make full use of the unlabeled data, we assign a pseudo label to each unlabeled instance, denoted as " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ", and then build the target classifier, " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "h_{\\hat{T}}: \\mathcal{X} \\to \\mathcal{Y}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ", to map the given instance to one of the known categories in " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{T} = \\{x_{i,l}, y_{i,l}\\}_{i=1}^{m} \\cup \\{x_{i,u}, \\hat{y}_{i,u}\\}_{i=1}^{n}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{y}_{i,u} \\in \\mathcal{Y}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{T}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " indicates the instances in hand, that is, labeled instances and unlabeled instances assigned with pseudo labels. Then, the population risk [32] of the target classifier learned from both labeled and unlabeled data with the pseudo label " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "(\\hat{T})" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " is controlled by the generalization gap, training error, and SSL" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16412" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "error, as shown in Eq.1. The generalization gap is the gap between the population risk and the average prediction loss across all instances with target categories " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "(T)" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " contains all the accessible instances with target categories, including labeled and unlabeled. And every instance in " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " is assumed with ground truth labels in ideal. The training error is the average empirical loss across " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\hat{T}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": ". The SSL error is the gap between the average empirical loss across the instances with target categories " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "(T)" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " and the average empirical loss across both labeled data and unlabeled ones with pseudo labels " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "(\\hat{T})" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": ". We depict the relations among these sets in Figure 4." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 228, + 287, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 228, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 59, + 228, + 287, + 378 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbb {E} _ {(\\boldsymbol {x}, \\boldsymbol {y}) \\sim D} [ l (\\boldsymbol {x}, \\boldsymbol {y}; h _ {\\hat {T}}) ] \\\\ \\leq \\underbrace {\\left| \\mathbb {E} _ {(\\boldsymbol {x} , y) \\sim D} [ l (\\boldsymbol {x} , y ; h _ {\\hat {T}}) ] - \\frac {1}{| T |} \\sum_ {(\\boldsymbol {x} , y) \\in T} l (\\boldsymbol {x} , y ; h _ {\\hat {T}}) \\right|} _ {\\text {g e n e r a l i z a t i o n g a p}} \\\\ + \\underbrace {\\left| \\frac {1}{| \\hat {T} |} \\sum_ {(\\boldsymbol {x} , y) \\in \\hat {T}} l (\\boldsymbol {x} , y ; h _ {\\hat {T}}) \\right|} _ {\\text {t r a i n i n g e r r o r}} \\tag {1} \\\\ + \\underbrace {\\left| \\frac {1}{| T |} \\sum_ {(\\pmb {x} , y) \\in T} l (\\pmb {x} , y ; h _ {\\hat {T}}) - \\frac {1}{| \\hat {T} |} \\sum_ {(\\pmb {x} , y) \\in \\hat {T}} l (\\pmb {x} , y ; h _ {\\hat {T}}) \\right|} _ {\\mathrm {S S L e r r o r}}, \\\\ \\end{array}", + "image_path": "d3b8c3536ea761f994f153680cbdb7474e24b3cea33f0fa3c66cf067ea266539.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 381, + 287, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 287, + 431 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 381, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 47, + 381, + 287, + 431 + ], + "type": "text", + "content": " is the data distribution of the instances that belong to target categories in the realistic world, i.e., " + }, + { + "bbox": [ + 47, + 381, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\mathcal{X} \\times \\mathcal{Y}" + }, + { + "bbox": [ + 47, + 381, + 287, + 431 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 381, + 287, + 431 + ], + "type": "inline_equation", + "content": "l(\\cdot, \\cdot; h_{\\hat{T}}): \\mathcal{X} \\times \\mathcal{Y} \\to \\mathcal{R}" + }, + { + "bbox": [ + 47, + 381, + 287, + 431 + ], + "type": "text", + "content": " denotes the loss function of the classifier " + }, + { + "bbox": [ + 47, + 381, + 287, + 431 + ], + "type": "inline_equation", + "content": "h_{\\hat{T}}" + }, + { + "bbox": [ + 47, + 381, + 287, + 431 + ], + "type": "text", + "content": " learned from " + }, + { + "bbox": [ + 47, + 381, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\hat{T}" + }, + { + "bbox": [ + 47, + 381, + 287, + 431 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 432, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 287, + 574 + ], + "type": "text", + "content": "Theoretical analyses [36] have confirmed that the generalization gap of DNNs can be bounded, and empirical evidence suggests that the training error of DNNs can be reduced almost to zero [32]. Thus, the essential component concerning population risk is the SSL error. Under class distribution mismatch, in addition to the wrongly annotated instances with target categories, the ones with unknown categories also contribute to the SSL error as they invade the training of the target classifier as outliers. Accordingly, we decouple the SSL error into pseudo-labeling and invasion error, as shown in Eq.2. For a detailed derivation process, please refer to Appendix 5.2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 578, + 287, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 578, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 578, + 287, + 712 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left| \\frac {1}{| T |} \\sum_ {(\\boldsymbol {x}, y) \\in T} l (\\boldsymbol {x}, y; h _ {\\hat {T}}) - \\frac {1}{| \\hat {T} |} \\sum_ {(\\boldsymbol {x}, y) \\in \\hat {T}} l (\\boldsymbol {x}, y; h _ {\\hat {T}}) \\right| \\\\ \\leq \\underbrace {\\left| \\frac {1}{| T |} \\sum_ {(\\boldsymbol {x} , y) \\in T} l (\\boldsymbol {x} , y ; h _ {\\hat {T}}) - \\frac {1}{| \\hat {T} |} \\sum_ {(\\boldsymbol {x} , y) \\in \\hat {T} \\setminus U} l (\\boldsymbol {x} , y ; h _ {\\hat {T}}) \\right|} _ {\\textbf {P s e u d o - l a b e l i n g e r r o r}} \\\\ + \\underbrace {\\left| \\frac {1}{| \\hat {T} |} \\sum_ {(\\boldsymbol {x} , y) \\in U} l (\\boldsymbol {x} , y ; h _ {\\hat {T}}) \\right|} _ {\\text {I n v a s i o n e r r o r}} \\tag {2} \\\\ \\end{array}", + "image_path": "6f2f4856a3e22efd4b1137cdfd9436c86b7ece90f8ccbc0be34e4043dae432df.jpg" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 543, + 127 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 543, + 127 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 543, + 127 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 543, + 127 + ], + "type": "image", + "image_path": "bb87031d6d3253f1852ffe8e7123ca4b44767cb6b1624f83879b14e51a1b8d72.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 129, + 545, + 171 + ], + "lines": [ + { + "bbox": [ + 305, + 129, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 305, + 129, + 545, + 171 + ], + "type": "text", + "content": "Figure 4. The relations among data sets " + }, + { + "bbox": [ + 305, + 129, + 545, + 171 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_l" + }, + { + "bbox": [ + 305, + 129, + 545, + 171 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 129, + 545, + 171 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u" + }, + { + "bbox": [ + 305, + 129, + 545, + 171 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 129, + 545, + 171 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 129, + 545, + 171 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 129, + 545, + 171 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 305, + 129, + 545, + 171 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 129, + 545, + 171 + ], + "type": "inline_equation", + "content": "\\hat{T}" + }, + { + "bbox": [ + 305, + 129, + 545, + 171 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 305, + 129, + 545, + 171 + ], + "type": "inline_equation", + "content": "\\hat{T} \\neq T \\cup U" + }, + { + "bbox": [ + 305, + 129, + 545, + 171 + ], + "type": "text", + "content": " due to the instances with target categories in " + }, + { + "bbox": [ + 305, + 129, + 545, + 171 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u" + }, + { + "bbox": [ + 305, + 129, + 545, + 171 + ], + "type": "text", + "content": " are assigned with pseudo labels while not the ground truth ones in " + }, + { + "bbox": [ + 305, + 129, + 545, + 171 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 129, + 545, + 171 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 192, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 192, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 305, + 192, + 545, + 216 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 192, + 545, + 216 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 305, + 192, + 545, + 216 + ], + "type": "text", + "content": " indicates the unlabeled instances with unknown categories and " + }, + { + "bbox": [ + 305, + 192, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\hat{T} = \\hat{T}\\backslash U\\cup U" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 217, + 546, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 217, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 546, + 396 + ], + "type": "text", + "content": "In Eq.2, the pseudo-labeling error is contributed by the wrongly annotated instances with target categories, as it is the gap of the average empirical loss caused by the inconsistency of the ground truth and pseudo labels. Thus, the quality of pseudo-labels assigned to unlabeled instances within the target distribution determines this error, and accurate pseudo-labeling may alleviate it. By contrast, the invasion error is the average empirical loss across the instances with unknown categories that is caused by the negative effect of those untargeted instances. By Eq.1& Eq.2, we find that the population risk of the target model is jointly controlled by pseudo-labeling error and invasion error. Accordingly, to mitigate the SSL error, we need to filter those instances with unknown categories and accurately annotate the unlabeled instances with target categories as well." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 405, + 506, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 405, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 405, + 506, + 418 + ], + "type": "text", + "content": "3.3. Weight-aware Distillation Framework" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 424, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 424, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 424, + 545, + 472 + ], + "type": "text", + "content": "With the aim of mitigating the pseudo-labeling and invasion errors, we design an SSL framework named WAD, which delivers the knowledge of pseudo labels and weights from robust representations to the target classifier." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 490, + 440, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 440, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 440, + 502 + ], + "type": "text", + "content": "3.3.1 Pseudo Label Learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 509, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 653 + ], + "type": "text", + "content": "Most existing SSL approaches produce pseudo labels by leveraging an immature target classifier, which cause catastrophic error once invaded by some instances with unknown categories, just as discussed in Section 1. To solve this problem, we distill the pseudo labels from a representation space (Teacher model) which is learned from all labeled and unlabeled instances by contrastive learning in an unsupervised manner and then transfer it to the target classifier (Student model). The teacher model could produce closely aligned representations for instances from the same categories and maximize the mutual information among them [2, 18, 20, 25]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": "Denoted the labeled and unlabeled representations learned by the teacher model, " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": ", as " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_l = \\{z_{j,l,k}\\}_{j=1}^m" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_u = \\{z_{i,u}\\}_{i=1}^n" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "k \\in \\mathcal{V}" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": ". Inspired by the characteristic of contrastive learning, one effective approach for building the pseudo-label of an unlabeled in-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16413" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 54, + 68, + 541, + 256 + ], + "blocks": [ + { + "bbox": [ + 54, + 68, + 541, + 256 + ], + "lines": [ + { + "bbox": [ + 54, + 68, + 541, + 256 + ], + "spans": [ + { + "bbox": [ + 54, + 68, + 541, + 256 + ], + "type": "image", + "image_path": "b7cf7136a5b561ffcc6a96843bf126f9d3e4e021c4974f9e1d3dcac29573989d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 258, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 258, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 547, + 304 + ], + "type": "text", + "content": "Figure 5. Illustration of WAD. The pseudo labels and weights determined by point mutual information (PMI) in robust representations of the teacher model are used to participate in training the target classifier. Then, some reliable instances selected according to " + }, + { + "bbox": [ + 46, + 258, + 547, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{WAD}}" + }, + { + "bbox": [ + 46, + 258, + 547, + 304 + ], + "type": "text", + "content": " are regarded as labeled ones to update the knowledge gradually. “" + }, + { + "bbox": [ + 46, + 258, + 547, + 304 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 46, + 258, + 547, + 304 + ], + "type": "text", + "content": "” and “" + }, + { + "bbox": [ + 46, + 258, + 547, + 304 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 46, + 258, + 547, + 304 + ], + "type": "text", + "content": "” indicates the labeled and unlabeled instance respectively. “//” means stop gradient." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 322, + 288, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 322, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 288, + 370 + ], + "type": "text", + "content": "stance is to identify the labeled instance with the highest PMI and then assign the label of it to unlabeled ones. The PMI between the unlabeled and labeled representation is formulated as Eq.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 377, + 287, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 377, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 86, + 377, + 287, + 406 + ], + "type": "interline_equation", + "content": "\\operatorname {P M I} \\left(z _ {i, u}, z _ {j, l, k}\\right) = \\log \\left[ \\frac {p \\left(z _ {j , l , k} \\mid z _ {i , u}\\right)}{p \\left(z _ {j , l , k}\\right)} \\right] \\tag {3}", + "image_path": "6399f70e7b89ecb66ba7671b278319981a0a29513fe04a071d9f5ab750501e17.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 411, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 287, + 460 + ], + "type": "text", + "content": "Although the conditional and marginals distributions, i.e., " + }, + { + "bbox": [ + 47, + 411, + 287, + 460 + ], + "type": "inline_equation", + "content": "p(z_{j,l,k}|z_{i,u})" + }, + { + "bbox": [ + 47, + 411, + 287, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 411, + 287, + 460 + ], + "type": "inline_equation", + "content": "p(z_{j,l,k})" + }, + { + "bbox": [ + 47, + 411, + 287, + 460 + ], + "type": "text", + "content": ", cannot be directly evaluated, we prove that PMI is proportional to the inner product in Appendix 2, as described in Eq.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 468, + 287, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 468, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 97, + 468, + 287, + 483 + ], + "type": "interline_equation", + "content": "f \\left(z _ {i, u}, z _ {j, l, k}\\right) \\propto \\operatorname {P M I} \\left(z _ {i, u}, z _ {j, l, k}\\right). \\tag {4}", + "image_path": "c302dac40bf750c2b3da6f58ce2a7fac215ef0b13f9a2bb7bf681f0c55ec0fd9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 489, + 287, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 489, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 287, + 512 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 489, + 287, + 512 + ], + "type": "inline_equation", + "content": "f = \\cos(\\cdot, \\cdot)" + }, + { + "bbox": [ + 46, + 489, + 287, + 512 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 489, + 287, + 512 + ], + "type": "inline_equation", + "content": "\\| z_i \\| = 1" + }, + { + "bbox": [ + 46, + 489, + 287, + 512 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 489, + 287, + 512 + ], + "type": "inline_equation", + "content": "\\propto" + }, + { + "bbox": [ + 46, + 489, + 287, + 512 + ], + "type": "text", + "content": " stands for \"proportional to\"." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 514, + 261, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 514, + 261, + 525 + ], + "spans": [ + { + "bbox": [ + 59, + 514, + 261, + 525 + ], + "type": "text", + "content": "Therefore, the pseudo label is formulated as Eq.5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 534, + 287, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 534, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 287, + 554 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {i, u} = \\underset {k} {\\arg \\max } f \\left(z _ {i, u}, z _ {j, l, k}\\right) \\tag {5}", + "image_path": "f400ee351a226c7fa8399855fa8566b21cc4dcb926d861c7ac490420230932bb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 560, + 287, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 560, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 46, + 560, + 287, + 621 + ], + "type": "text", + "content": "Consequently, the class label of the labeled instance with the maximum PMI is assigned to the unlabeled one. The Eq.5 can precisely capture the PMI from the representation space to produce high-quality pseudo labels and then mitigate pseudo-labeling error." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 635, + 209, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 209, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 209, + 647 + ], + "type": "text", + "content": "3.3.2 Unknown Categories Filtering" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "content": "To mitigate the invasion error, the instances with unknown categories should be filtered out. Following the Subsection 3.3.1, a higher PMI between the labeled and unlabeled instance suggests a stronger association or similarity between the two instances, further indicating a higher" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 322, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 545, + 441 + ], + "type": "text", + "content": "likelihood of the unlabeled instance belonging to the same class distribution as the labeled one. However, some hard instances that have similar PMI between two target categories, i.e., laid on the decision boundary of two target categories, may introduce incorrect pseudo labels and hurt the performance of the target classifier. Hence, we also propose a ratio among the first and second maximum PMI to evaluate the confidence of the pseudo labels. Then, the weight is defined as Eq.6 to avoid the negative effect caused by the wrong labels and unknown categories." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 352, + 448, + 545, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 448, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 352, + 448, + 545, + 475 + ], + "type": "interline_equation", + "content": "\\boldsymbol {w} _ {i, u} = g _ {1} \\left(\\widetilde {p} _ {i, u}\\right) \\times g _ {2} \\left(1 - \\frac {\\widetilde {q} _ {i , u}}{\\widetilde {p} _ {i , u}}\\right), \\tag {6}", + "image_path": "07e7ef6330b9cab941b0567be0ab8c7faac50e77f37542bd0500a3ef85659481.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 480, + 343, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 480, + 343, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 480, + 343, + 491 + ], + "type": "text", + "content": "wherein," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 365, + 491, + 472, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 491, + 472, + 509 + ], + "spans": [ + { + "bbox": [ + 365, + 491, + 472, + 509 + ], + "type": "interline_equation", + "content": "\\widetilde {p} _ {i, u} = \\max _ {j} f (z _ {i, u}, z _ {j, l, k})", + "image_path": "97542babb2d5757746bd1d3c7bc22a406ce84bfbfe5d405bf56eb484c18f0a35.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 366, + 511, + 484, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 511, + 484, + 530 + ], + "spans": [ + { + "bbox": [ + 366, + 511, + 484, + 530 + ], + "type": "interline_equation", + "content": "\\widetilde {q} _ {i, u} = \\max _ {v, k \\neq \\hat {y} _ {i, u}} f (z _ {i, u}, z _ {v, l, k})", + "image_path": "5fa5a749c862bd4f3a18593e2b0a08a790bf307450574d623e95771c8c32d7f3.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "text", + "content": "In Eq.6, " + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "g_{1}(\\cdot)" + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "g_{2}(\\cdot)" + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "text", + "content": " can be interpreted as any monotonically increasing functions. The former in Eq.6 aims to estimate the likelihood of the unlabeled instance belonging to target categories. The higher this item, the more chances of the instance in the target class distribution are. The latter, " + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "g_{2}(\\cdot)" + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "text", + "content": ", penalizes instances whose labels are ambiguous between the nearest and second-nearest target categories. The lower this item, the larger probability of incorrect pseudo labels is. As shown in Figure 6, the weight could filter instances with unknown categories and those incorrectly annotated ones with target categories, while the ones from target categories with high-quality pseudo labels are encouraged. Thus, by weight, WAD selectively distills the knowledge beneficial to the target classifier from the teacher model, and the invasion error is then mitigated." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16414" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 240, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 240, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 240, + 85 + ], + "type": "text", + "content": "3.3.3 Weight-aware Knowledge Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 198 + ], + "type": "text", + "content": "Weight-aware knowledge distillation loss. The knowledge of pseudo labels and weights captured from robust representations is applied in the distillation process. In each feed-forward process, pseudo labels and weights are aggregated to the target classifier, as shown in Figure 5. Then, we propose the weight-aware knowledge distillation loss, including the traditional supervised loss " + }, + { + "bbox": [ + 46, + 91, + 287, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_l" + }, + { + "bbox": [ + 46, + 91, + 287, + 198 + ], + "type": "text", + "content": " in labeled data and weight-aware supervised loss " + }, + { + "bbox": [ + 46, + 91, + 287, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 46, + 91, + 287, + 198 + ], + "type": "text", + "content": " in unlabeled data as Eq.7." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 124, + 198, + 287, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 198, + 287, + 211 + ], + "spans": [ + { + "bbox": [ + 124, + 198, + 287, + 211 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {W A D}} = \\mathcal {L} _ {l} + \\mathcal {L} _ {u}, \\tag {7}", + "image_path": "8c2a123b64d19da567d8c239324094f650391794f359928bb464f5400b2ab8c3.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 216, + 84, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 216, + 84, + 226 + ], + "spans": [ + { + "bbox": [ + 47, + 216, + 84, + 226 + ], + "type": "text", + "content": "wherein," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 235, + 242, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 235, + 242, + 266 + ], + "spans": [ + { + "bbox": [ + 80, + 235, + 242, + 266 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l} = \\frac {1}{| \\mathcal {D} _ {l} |} \\sum_ {(x _ {i, l}, y _ {i, l}) \\in \\mathcal {D} _ {l}} \\ell (h (x _ {i, l}; \\theta), y _ {i, l})", + "image_path": "756466fee6832a20bc8cee9ae9e5c95fce38ebf680886d4b62b91963ec9c4c3a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 269, + 253, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 269, + 253, + 297 + ], + "spans": [ + { + "bbox": [ + 82, + 269, + 253, + 297 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {u} = \\frac {1}{| \\mathcal {D} _ {u} |} \\sum_ {x _ {i, u} \\in \\mathcal {D} _ {u}} \\boldsymbol {w} _ {i, u} \\ell (h (x _ {i, u}; \\theta), \\hat {y} _ {i, u})", + "image_path": "3f0052e1ff7542fb7d099ce8f481eb568d442c3b2eb9ac3fb3d56465ecb347ce.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 302, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 287, + 445 + ], + "type": "text", + "content": "The traditional supervised loss " + }, + { + "bbox": [ + 46, + 302, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_l" + }, + { + "bbox": [ + 46, + 302, + 287, + 445 + ], + "type": "text", + "content": " aims to minimize the distance between the predicted probability and the ground truth label. While the weight-aware supervised loss " + }, + { + "bbox": [ + 46, + 302, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 46, + 302, + 287, + 445 + ], + "type": "text", + "content": " mainly focuses on selectively transferring the beneficial knowledge from the teacher model to the student model by weights to mitigate the negative effect from unknown categories and improve the target classifier as well. Moreover, " + }, + { + "bbox": [ + 46, + 302, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{WAD}}" + }, + { + "bbox": [ + 46, + 302, + 287, + 445 + ], + "type": "text", + "content": " is the loss function that is adopted to train the target classifier " + }, + { + "bbox": [ + 46, + 302, + 287, + 445 + ], + "type": "inline_equation", + "content": "h_{\\hat{T}}" + }, + { + "bbox": [ + 46, + 302, + 287, + 445 + ], + "type": "text", + "content": " mentioned in Eq.1. Consequently, WAD leverages the pseudo labels and weights to mitigate the pseudolabeling and invasion errors, following alleviating the SSL error, which has been proved in Subsection 3.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 445, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 541 + ], + "type": "text", + "content": "Knowledge-update in Training. The knowledge of pseudo labels and weights may be biased as the labeled data is limited. Accordingly, after several forward iterations, we progressively add some reliable instances to labeled data. Because the feedback from the target classifier, i.e., loss, is highly related to the weights and reflects the training error, we consider the reliability according to it. Then, the criterion for updating is formulated as Eq.8." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 114, + 549, + 287, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 549, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 114, + 549, + 287, + 562 + ], + "type": "interline_equation", + "content": "c _ {i, u} = \\ell \\left(h \\left(x _ {i, u}; \\theta_ {t}\\right), \\hat {y} _ {i, u}\\right) \\tag {8}", + "image_path": "190a173d436fdea620c81bc0ac6ed052a95423e0dd7e3735eeffe41deb7ab6d3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 570, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 570, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\ell (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 46, + 570, + 287, + 593 + ], + "type": "text", + "content": " is the cross-entropy function, and " + }, + { + "bbox": [ + 46, + 570, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\theta_{t}" + }, + { + "bbox": [ + 46, + 570, + 287, + 593 + ], + "type": "text", + "content": " is the parameters of the target classifier in the current iteration." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "The reliability of " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "x_{i,u}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " is enhanced when " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "c_{i,u}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " takes a lower value. Then, WAD leverages Eq.8 to identify the top " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\alpha \\%" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " reliable instances from the unlabeled data and puts them in the labeled data while removing them from the unlabeled data. Moreover, we adopt the polynomial decay [5] to dynamically adjust " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " to prevent the gradually increased negative effect from unknown categories with the iteration. The details are shown in Appendix 3. A visualization of the number of selected reliable instances with target categories is also provided in Appendix 4.4. Consequently, the pseudo" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "labels and weights are updated in the subsequent distillation steps, as shown in Figure 5, with the aim of optimizing the target classifier. Finally, the schematic diagram and algorithm process is presented in Figure 6 and Algorithm 1, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 140, + 419, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 140, + 419, + 152 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 419, + 152 + ], + "type": "text", + "content": "3.4. Theoretical Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 159, + 545, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 545, + 195 + ], + "type": "text", + "content": "This subsection provides the theoretical studies about the WAD's SSL error, as shown in Theorem 1. Detailed proof of Theorem 1 is given in Appendix 5." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "text", + "content": "Theorem 1 Given " + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "inline_equation", + "content": "|T|" + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "text", + "content": " instances that i.i.d. sampled from " + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "inline_equation", + "content": "\\{(x_i, y_i)\\}_{i=1}^{|T|}" + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "inline_equation", + "content": "|U|" + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "text", + "content": " instances that is not i.i.d with " + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "inline_equation", + "content": "\\hat{T} = \\{(x, \\hat{y}) | (x, y) \\in T \\cup U, \\hat{y} \\in \\mathcal{Y}\\}" + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\mathcal{X} \\times \\mathcal{Y}" + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "text", + "content": ". Assume the loss function " + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "inline_equation", + "content": "l(\\cdot, y; h_{\\hat{T}})" + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "inline_equation", + "content": "\\lambda^l" + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "text", + "content": "-Lipschitz continuous for all " + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "inline_equation", + "content": "y, h_{\\hat{T}}" + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "text", + "content": " and bounded by " + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "text", + "content": ", the regression function is " + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "inline_equation", + "content": "\\lambda^\\mu" + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "text", + "content": "-Lipschitz continuous, training error " + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "inline_equation", + "content": "l(\\pmb{x}, y; h_{\\hat{T}}) = 0" + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "inline_equation", + "content": "\\forall (\\pmb{x}, y) \\in \\hat{T}" + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "inline_equation", + "content": "\\overline{\\pmb{w}}" + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "text", + "content": " indicates the average of weights, and " + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "text", + "content": " is the maximum PMI which determines the pseudo label, with the probability of at least " + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "inline_equation", + "content": "1 - \\gamma" + }, + { + "bbox": [ + 304, + 202, + 546, + 310 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 317, + 545, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 317, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 315, + 317, + 545, + 410 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left| \\frac {1}{| T |} \\sum_ {(\\boldsymbol {x}, y) \\in T} l (\\boldsymbol {x}, y; h _ {\\hat {T}}) - \\frac {1}{| \\hat {T} |} \\sum_ {(\\boldsymbol {x}, y) \\in \\hat {T} \\backslash U} l (\\boldsymbol {x}, y; h _ {\\hat {T}}) \\right| \\\\ + \\left| \\frac {1}{| \\hat {T} |} \\sum_ {(\\boldsymbol {x}, y) \\in U} l (\\boldsymbol {x}, y; h _ {\\hat {T}}) \\right| \\tag {9} \\\\ \\leq \\sqrt {4 - 4 \\xi} (\\lambda^ {l} + \\lambda^ {\\mu} H K) + \\frac {\\overline {{\\boldsymbol {w}}} | U | H}{| \\hat {T} |} + \\sqrt {\\frac {2 H ^ {2} l o g (1 / \\gamma)}{| T |}}. \\\\ \\end{array}", + "image_path": "c35fa12803fa83e155c1163fc3718b5f49565242c71094503fc10a160373ba53.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 416, + 545, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 416, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 416, + 545, + 535 + ], + "type": "text", + "content": "From Theorem 1, we find that the smaller " + }, + { + "bbox": [ + 304, + 416, + 545, + 535 + ], + "type": "inline_equation", + "content": "\\overline{\\boldsymbol{w}}" + }, + { + "bbox": [ + 304, + 416, + 545, + 535 + ], + "type": "text", + "content": " and the larger " + }, + { + "bbox": [ + 304, + 416, + 545, + 535 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 304, + 416, + 545, + 535 + ], + "type": "text", + "content": ", the tighter the bound in the SSL error is. Specifically, just as verified in Appendix 5, the pseudo-labeling error bounded by " + }, + { + "bbox": [ + 304, + 416, + 545, + 535 + ], + "type": "inline_equation", + "content": "\\sqrt{4 - 4\\xi} (\\lambda^l +\\lambda^\\mu HK) + \\sqrt{\\frac{2H^2\\log(1 / \\gamma)}{|T|}}" + }, + { + "bbox": [ + 304, + 416, + 545, + 535 + ], + "type": "text", + "content": " and the invasion error bounded by " + }, + { + "bbox": [ + 304, + 416, + 545, + 535 + ], + "type": "inline_equation", + "content": "\\frac{\\overline{\\boldsymbol{w}}|U|H}{|\\hat{T}|}" + }, + { + "bbox": [ + 304, + 416, + 545, + 535 + ], + "type": "text", + "content": " can be reduced by minimizing the weights " + }, + { + "bbox": [ + 304, + 416, + 545, + 535 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 416, + 545, + 535 + ], + "type": "text", + "content": " of unlabeled instances with unknown categories and maximizing the confidence of pseudo labels, just as WAD does. Thus, WAD's SSL error has a tight upper bound." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 547, + 387, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 547, + 387, + 560 + ], + "spans": [ + { + "bbox": [ + 306, + 547, + 387, + 560 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "text", + "content": "Subsection 4.2 presents the comparison results between WAD and five state-of-the-art SSL approaches, as well as one standard baseline. Furthermore, an ablation experiment is conducted in Subsection 4.3, while sensitivity analyses and visualization are carried out in Subsection 4.4 and Subsection 4.5, respectively. For more experiments, please refer to Appendix 4.2 & 4.5." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 658, + 426, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 426, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 426, + 672 + ], + "type": "text", + "content": "4.1. Experimental Setups" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "Datasets. Our experiments are conducted on two benchmark datasets, CIFAR10 [23] and CIFAR100 [23], as well as an artificial cross-dataset that comprises subsamples" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16415" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 78, + 154, + 167 + ], + "blocks": [ + { + "bbox": [ + 53, + 78, + 154, + 167 + ], + "lines": [ + { + "bbox": [ + 53, + 78, + 154, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 78, + 154, + 167 + ], + "type": "image", + "image_path": "260de664b251c6a1fd12038f4d5209101deb3a37159b56ed23c28d8b25f262c7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 175, + 126, + 183 + ], + "lines": [ + { + "bbox": [ + 72, + 175, + 126, + 183 + ], + "spans": [ + { + "bbox": [ + 72, + 175, + 126, + 183 + ], + "type": "text", + "content": "(1) Representations" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 159, + 115, + 185, + 132 + ], + "blocks": [ + { + "bbox": [ + 159, + 115, + 185, + 132 + ], + "lines": [ + { + "bbox": [ + 159, + 115, + 185, + 132 + ], + "spans": [ + { + "bbox": [ + 159, + 115, + 185, + 132 + ], + "type": "image", + "image_path": "1d24f695f2723ee244b45bbce8e33c06e4f1455f025f3dc1b437a843f811a242.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 190, + 80, + 290, + 167 + ], + "blocks": [ + { + "bbox": [ + 190, + 80, + 290, + 167 + ], + "lines": [ + { + "bbox": [ + 190, + 80, + 290, + 167 + ], + "spans": [ + { + "bbox": [ + 190, + 80, + 290, + 167 + ], + "type": "image", + "image_path": "15f88588df540a25dbcfbfa75841d3003cc5301b7ab0c20c529cc74e79808704.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 175, + 290, + 183 + ], + "lines": [ + { + "bbox": [ + 187, + 175, + 290, + 183 + ], + "spans": [ + { + "bbox": [ + 187, + 175, + 290, + 183 + ], + "type": "text", + "content": "(2) Representations with pseudo labels" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 187, + 547, + 221 + ], + "lines": [ + { + "bbox": [ + 46, + 187, + 547, + 221 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 547, + 221 + ], + "type": "text", + "content": "Figure 6. The schematic diagram of how WAD works. First, each unlabeled instance is assigned a pseudo label according to the max PMI with labeled data in (1), as shown in (2). Then, the instances with wrong labels and unknown categories are assigned a tiny weight to avoid the negative effect, as shown in (3)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 297, + 116, + 321, + 133 + ], + "blocks": [ + { + "bbox": [ + 297, + 116, + 321, + 133 + ], + "lines": [ + { + "bbox": [ + 297, + 116, + 321, + 133 + ], + "spans": [ + { + "bbox": [ + 297, + 116, + 321, + 133 + ], + "type": "image", + "image_path": "426970f158d050b18b5f61b483d740728f483fe28c04b40889846aa264359976.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 328, + 70, + 428, + 167 + ], + "blocks": [ + { + "bbox": [ + 328, + 70, + 428, + 167 + ], + "lines": [ + { + "bbox": [ + 328, + 70, + 428, + 167 + ], + "spans": [ + { + "bbox": [ + 328, + 70, + 428, + 167 + ], + "type": "image", + "image_path": "28e5a25df893d65ff4a28262441006760e553f35c270b3ace622efb8ced6aeff.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 175, + 444, + 183 + ], + "lines": [ + { + "bbox": [ + 318, + 175, + 444, + 183 + ], + "spans": [ + { + "bbox": [ + 318, + 175, + 444, + 183 + ], + "type": "text", + "content": "(3) Representations with pseudo labels and weights" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 464, + 72, + 529, + 160 + ], + "blocks": [ + { + "bbox": [ + 464, + 72, + 529, + 160 + ], + "lines": [ + { + "bbox": [ + 464, + 72, + 529, + 160 + ], + "spans": [ + { + "bbox": [ + 464, + 72, + 529, + 160 + ], + "type": "image", + "image_path": "38932131b7c55b034e14a03fd0fd5fafafb45aa2136cf5302b0bb206a24032dc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 160, + 529, + 167 + ], + "lines": [ + { + "bbox": [ + 465, + 160, + 529, + 167 + ], + "spans": [ + { + "bbox": [ + 465, + 160, + 529, + 167 + ], + "type": "text", + "content": "Unlabeled data" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 465, + 167, + 529, + 173 + ], + "lines": [ + { + "bbox": [ + 465, + 167, + 529, + 173 + ], + "spans": [ + { + "bbox": [ + 465, + 167, + 529, + 173 + ], + "type": "text", + "content": "Unknown categories" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 465, + 174, + 518, + 182 + ], + "lines": [ + { + "bbox": [ + 465, + 174, + 518, + 182 + ], + "spans": [ + { + "bbox": [ + 465, + 174, + 518, + 182 + ], + "type": "text", + "content": "Cat Dog" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "code", + "bbox": [ + 50, + 256, + 271, + 535 + ], + "blocks": [ + { + "bbox": [ + 53, + 243, + 260, + 255 + ], + "lines": [ + { + "bbox": [ + 53, + 243, + 260, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 243, + 260, + 255 + ], + "type": "text", + "content": "Algorithm 1: Weight-aware distillation framework" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "lines": [ + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "spans": [ + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": "Input: Labeled data: " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_l" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": ", Unlabeled data: " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": ", Max iterations: " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": ", Initial value of " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": ", Set of update steps: " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": ". Output: target classifier " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "h_{\\hat{T}}" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": ". \n1 Embedding: " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_l \\gets \\phi(\\mathcal{D}_l)" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_u \\gets \\phi(\\mathcal{D}_u)" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": ". \n2 Initialize model parameter; \n3 for " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": " do \n4 " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "\\forall x_{i,u} \\in \\mathcal{D}_u" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": ", distill the knowledge about pseudo label, " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "\\hat{y}_{i,u}" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": ", using Eq.5; \n5 " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "\\forall x_{i,u} \\in \\mathcal{D}_u" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": ", distill the knowledge about weight, " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "w_{i,u}" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": ", using Eq.6; \n6 " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "\\alpha =" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": " polynomial Decay " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "(\\alpha_0)" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": "; \n7 Using Eq.7 training the target classifier; \n8 if " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "t \\in G" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": " then \n9 " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "C = \\emptyset" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": "; \n10 Calculate the reliability of each unlabeled instance using Eq.8. \n11 Ascending the instances according to reliability and add the top " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "\\alpha\\%" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": " instances with pseudo labels to " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": ". \n12 " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_l = \\mathcal{D}_l \\cup C" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u = \\mathcal{D}_u \\backslash \\{x | (x,y) \\in C\\}" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": ". \n13 end \n14 Return target classifier " + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "inline_equation", + "content": "h_{\\hat{T}}" + }, + { + "bbox": [ + 50, + 256, + 271, + 535 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "code_body" + } + ], + "index": 14, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 566, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 566, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 287, + 651 + ], + "type": "text", + "content": "from CIFAR10, CIFAR100, Flowers [29], Food-101 [6], and Places-365 [41]. The CIFAR10 and CIFAR100 datasets consist of 50,000 training and 10,000 testing images of 10 and 100 categories, respectively. The cross-dataset contains 138,000 unlabeled instances from 674 categories. All images from the datasets are resized to " + }, + { + "bbox": [ + 46, + 566, + 287, + 651 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 46, + 566, + 287, + 651 + ], + "type": "text", + "content": ". For further details, please refer to Appendix 4.1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Settings. i) The proportion of the instances with unknown categories in unlabeled data, named as mismatch proportion, are set as " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " in this work. For instance, the unlabeled data has a " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " mismatch proportion with 4,000 instances with target categories and" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 240, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 312 + ], + "type": "text", + "content": "6,000 instances with unknown categories. ii) Randomly sampled " + }, + { + "bbox": [ + 304, + 240, + 545, + 312 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 304, + 240, + 545, + 312 + ], + "type": "text", + "content": " instances from the training dataset that belong to target categories are regarded as labeled data. The remaining " + }, + { + "bbox": [ + 304, + 240, + 545, + 312 + ], + "type": "inline_equation", + "content": "92\\%" + }, + { + "bbox": [ + 304, + 240, + 545, + 312 + ], + "type": "text", + "content": " of instances with target categories and some instances with unknown categories are composed of unlabeled data according to the mismatch proportion." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 313, + 546, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 313, + 546, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 546, + 588 + ], + "type": "text", + "content": "Details. The teacher model is with a Resnet-18 [16] backbone and is trained using SimCLR [10]. And we maintain consistency with SimCLR in all implementation details. The target classifier is a WideResnet-28-2 network [38] with input size " + }, + { + "bbox": [ + 304, + 313, + 546, + 588 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 313, + 546, + 588 + ], + "type": "text", + "content": ", following Huang et al. [19]. Both the encoder and target classifier are trained from scratch. We train the target classifier using the Adam optimizer [22] with a learning rate of " + }, + { + "bbox": [ + 304, + 313, + 546, + 588 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 313, + 546, + 588 + ], + "type": "text", + "content": ". Furthermore, the epochs and batch size are set as 100 and 32, respectively. The augmentations include random horizontal flipping, random translation by up to 2 pixels, and Gaussian input noise with a standard deviation of 0.15 is used in the training of the target classifier as same as Guo et al. [15]. Moreover, we apply global contrast normalization and ZCA normalized, which is widely used in the pretreatment [15, 11], on CIFAR10. For simplicity, the functions " + }, + { + "bbox": [ + 304, + 313, + 546, + 588 + ], + "type": "inline_equation", + "content": "g_{1}(\\cdot)" + }, + { + "bbox": [ + 304, + 313, + 546, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 313, + 546, + 588 + ], + "type": "inline_equation", + "content": "g_{2}(\\cdot)" + }, + { + "bbox": [ + 304, + 313, + 546, + 588 + ], + "type": "text", + "content": " act as identical mapping with no additional constraints. The initial value of " + }, + { + "bbox": [ + 304, + 313, + 546, + 588 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 313, + 546, + 588 + ], + "type": "text", + "content": " is set as 0.1 and decayed five times until it reached 0. It remains the same across all experiments unless otherwise specified. Finally, the approaches on each dataset run three times, and the mean accuracy and standard deviation are reported; the best one is highlighted in bold." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 589, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 589, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 589, + 545, + 661 + ], + "type": "text", + "content": "Baselines. WAD is compared to five state-of-the-art approaches, including " + }, + { + "bbox": [ + 304, + 589, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\mathrm{DS}^{3}\\mathrm{L}" + }, + { + "bbox": [ + 304, + 589, + 545, + 661 + ], + "type": "text", + "content": " [15], T2T [19], CCSSL [37], UASD [11] and ORCA [8], as well as one baseline model that only trains labeled data. Moreover, T2T and ORCA are performed without pretraining tasks for fairness, indicated by \"T2T w\\o pre.\" and \"ORCA w\\o pre\"." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 670, + 430, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 430, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 430, + 683 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "This subsection presents the experimental results of the classification tasks performed on CIFAR10, CIFAR100," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16416" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 545, + 198 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 545, + 198 + ], + "type": "table", + "html": "
MethodCIFAR10CIFAR100
20%40%60%80%20%40%60%80%
Baseline94.33±0.4594.33±0.4594.33±0.4594.33±0.4536.98±1.7936.98±1.7936.98±1.7936.98±1.79
DS3L91.82±1.8991.38±1.7392.47±0.2590.82±1.5023.92±2.7824.92±4.4126.20±4.2924.55±3.67
UASD95.02±0.7795.03±0.7793.87±0.1393.37±0.3539.85±0.3537.55±2.2436.03±0.7329.87±2.07
CCSSL86.08±0.1284.00±0.1783.13±0.1981.15±0.2541.72 ±0.8541.20±0.5840.60 ±0.2239.67±0.31
T2T----43.70±0.5042.82±0.4540.12±0.7137.35±1.10
T2Tw\\o pre.----39.40±0.3636.78±0.1636.65 ±1.0934.62±1.68
ORCA95.40±0.7494.13±1.1694.35±0.6793.82±0.9329.50±0.2531.12±0.7131.18±0.4031.65±1.86
ORCA w\\o pre.93.32±0.9992.55±2.0292.37±0.9089.65±6.9522.13±1.3323.98±0.7923.37±1.1422.98±0.53
WAD98.43±0.1497.88±0.3397.90±0.2097.77±0.3351.65±2.8650.00±1.4346.88±0.2045.45±1.73
", + "image_path": "1031d01beef778613eef940e1f4e03759a4e07e822da9af0aa23edc25b5d3f76.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 227, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 227, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 287, + 357 + ], + "type": "text", + "content": "and a cross-dataset. For CIFAR10, we designated two categories as the target and eight as unknown, while twenty classes are considered as target categories and eighty categories as unknown in CIFAR100. Moreover, we constructed a cross-dataset integrated with five datasets to evaluate WAD in the case that the unlabeled data contains massive unknown categories. Specifically, six classes from CIFAR10 were assigned as target categories, and 668 categories from four external datasets are unknown. The experimental results conducted on CIFAR10, CIFAR100, and the cross-dataset are presented in Table 1 and Table 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "text", + "content": "Results on CIFAR10 and CIFAR100. From Table 1, we have four findings as follows. i) WAD outperforms all compared methods on CIFAR10 and CIFAR100 with different mismatch proportions, demonstrating its remarkable performance. ii) WAD retains stable performance improvement under different mismatch proportions, exhibiting further improvement of " + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "inline_equation", + "content": "3.55\\%" + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "inline_equation", + "content": "2.91\\%" + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "inline_equation", + "content": "3.44\\%" + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "text", + "content": " for mismatch proportions of " + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "text", + "content": " on CIFAR10. This highlights that WAD can achieve robust performance even under a high mismatch proportion. iii) The accuracies of " + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "inline_equation", + "content": "\\mathrm{DS}^{3}\\mathrm{L}" + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "text", + "content": " on CIFAR10 and CIFAR100 are lower than baseline, as ORCA does. This is because it weights the instances according to consistent empirical risk loss, resulting in the invasion of many unknown categories in training, as shown in Appendix 4.3. iv) In CIFAR100, WAD surpasses baseline " + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "inline_equation", + "content": "8.47\\%" + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 362, + 287, + 637 + ], + "type": "text", + "content": " mismatch proportion. This demonstrates that WAD is still effective when the unlabeled data contains large unknown categories. Therefore, WAD achieves outstanding performance on datasets with different mismatch proportions and exhibits excellent robustness to the scale of unknown categories. Notably, T2T can not apply to the binary classification task, and the accuracy is not reported here." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "Results on cross-dataset. Further, we investigate the limits of WAD's tolerance for unknown categories and then perform the experiments on an artificial cross-dataset containing 668 unknown categories from four datasets. From Table 2, we observe that WAD still maintains an improvement compared to the baseline. Obviously, the other compared" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 225, + 545, + 344 + ], + "blocks": [ + { + "bbox": [ + 121, + 198, + 472, + 207 + ], + "lines": [ + { + "bbox": [ + 121, + 198, + 472, + 207 + ], + "spans": [ + { + "bbox": [ + 121, + 198, + 472, + 207 + ], + "type": "text", + "content": "Table 1. Experimental results on CIFAR10 and CIFAR100 under different mismatch proportions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 225, + 545, + 344 + ], + "lines": [ + { + "bbox": [ + 307, + 225, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 307, + 225, + 545, + 344 + ], + "type": "table", + "html": "
Cross-dataset
Method20%40%60%80%
Baseline66.83±1.3766.83±1.3766.83±1.3766.83±1.37
DS3L50.02±6.6950.69±5.2649.03±5.9351.46±6.99
UASD61.18±0.2957.02±0.5854.70±2.2545.67±1.72
CCSSL64.83±0.2765.15±0.5664.16±0.5864.16±0.45
T2T66.56±2.8065.08±0.7663.76±0.5362.83±0.77
T2T w\\o pre.64.44±0.1562.47±0.7962.23±0.7561.42 ±0.65
ORCA65.53±0.8565.51±1.2566.44±0.8066.46±1.28
ORCA w\\o pre.65.37±0.7863.63±0.6464.42±0.5366.34±1.05
WAD67.13±0.5967.20±1.6567.80±0.0767.88±0.37
", + "image_path": "9f48715d6668b055ca273e79c3614d6913398e25e04a6ba0baef322495d84853.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 344, + 545, + 365 + ], + "lines": [ + { + "bbox": [ + 306, + 344, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 306, + 344, + 545, + 365 + ], + "type": "text", + "content": "Table 2. Experimental results on cross-dataset under different mismatch proportions." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 386, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 386, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 386, + 545, + 422 + ], + "type": "text", + "content": "methods were lower than the baseline. This indicates that WAD could boost the performance even on a dataset that contains massive instances with unknown categories." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 430, + 406, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 406, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 406, + 442 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 449, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 557 + ], + "type": "text", + "content": "We conducted ablation studies on the CIFAR10 dataset using different models: \"+\"Pse.\" (trained with labeled data and unlabeled instances with pseudo labels), \"+\"Pse.&W.\" (trained with pseudo labels and fix weights), and the WAD model (trained with all components). We also examined the impact of the weight function and explored alternative choices for " + }, + { + "bbox": [ + 304, + 449, + 545, + 557 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 304, + 449, + 545, + 557 + ], + "type": "text", + "content": " through identical mappings, " + }, + { + "bbox": [ + 304, + 449, + 545, + 557 + ], + "type": "inline_equation", + "content": "g_{i}(\\cdot)" + }, + { + "bbox": [ + 304, + 449, + 545, + 557 + ], + "type": "text", + "content": ", and the transformation " + }, + { + "bbox": [ + 304, + 449, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\tilde{g}_{i}(\\cdot) = \\exp (\\cdot)" + }, + { + "bbox": [ + 304, + 449, + 545, + 557 + ], + "type": "text", + "content": ". Results are presented in Table 3, and w\\o " + }, + { + "bbox": [ + 304, + 449, + 545, + 557 + ], + "type": "inline_equation", + "content": "g_{i}(\\cdot)" + }, + { + "bbox": [ + 304, + 449, + 545, + 557 + ], + "type": "text", + "content": " means removing " + }, + { + "bbox": [ + 304, + 449, + 545, + 557 + ], + "type": "inline_equation", + "content": "g_{i}(\\cdot)" + }, + { + "bbox": [ + 304, + 449, + 545, + 557 + ], + "type": "text", + "content": " from Eq.6." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "content": "Effects of pseudo labels. From Table 3, we observe that compared with the baseline, " + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "inline_equation", + "content": "2.72\\%" + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "inline_equation", + "content": "1.52\\%" + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "content": " accuracy improvement can be obtained by leveraging the unlabeled instances with pseudo labels, under " + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "content": " mismatch proportion, respectively. This indicates that the pseudo labels are beneficial to improving performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Effects of weights. According to Table 3, we observe two findings about weights. i) Training by leveraging both pseudo labels and weights exhibits the comparable performance to that without weights under " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " mismatch proportion. This is because there are fewer instances with unknown categories under " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". Then, the model training with fixed weights will result in a sub-optimal model" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16417" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "compared to explicit labels. ii) The weights improve the accuracy by " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "0.99\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": " over without it, under " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": " mismatch proportion, while the gap decreases with the mismatch proportion increasing. This demonstrates that the weights are effective in filtering the instances with unknown categories. And the performance degradation is because the absence of the knowledge-update makes the algorithm fail to prevent the increased negative effect from unknown categories." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 289, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 289, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 289, + 287 + ], + "type": "text", + "content": "Effects of knowledge-update. We have two findings according to Table 3. i) The accuracy with knowledge-update surpasses the one only leveraging pseudo labels and weights, and the gap between them reaches " + }, + { + "bbox": [ + 46, + 168, + 289, + 287 + ], + "type": "inline_equation", + "content": "1.77\\%" + }, + { + "bbox": [ + 46, + 168, + 289, + 287 + ], + "type": "text", + "content": " under " + }, + { + "bbox": [ + 46, + 168, + 289, + 287 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 168, + 289, + 287 + ], + "type": "text", + "content": " mismatch proportion. This indicates that knowledge-update plays important roles in WAD. ii) WAD, training with all components, shows its outstanding performance compared to the ones removing other parts. This demonstrates that the aggregation of all the proposed parts could achieve significant improvement." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 287, + 288, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 288, + 360 + ], + "type": "text", + "content": "Effects of each part of Eq.6. From the Table 3, we have the following two findings. i) both “" + }, + { + "bbox": [ + 46, + 287, + 288, + 360 + ], + "type": "inline_equation", + "content": "\\mathrm{w} \\backslash \\mathrm{o} g_{1}(\\cdot)" + }, + { + "bbox": [ + 46, + 287, + 288, + 360 + ], + "type": "text", + "content": "” and “" + }, + { + "bbox": [ + 46, + 287, + 288, + 360 + ], + "type": "inline_equation", + "content": "\\mathrm{w} \\backslash \\mathrm{o} g_{2}(\\cdot)" + }, + { + "bbox": [ + 46, + 287, + 288, + 360 + ], + "type": "text", + "content": "” are worse than WAD, illustrating their equal importance for WAD. ii) Assigning the same mappings to " + }, + { + "bbox": [ + 46, + 287, + 288, + 360 + ], + "type": "inline_equation", + "content": "g_{1}(\\cdot)" + }, + { + "bbox": [ + 46, + 287, + 288, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 287, + 288, + 360 + ], + "type": "inline_equation", + "content": "g_{2}(\\cdot)" + }, + { + "bbox": [ + 46, + 287, + 288, + 360 + ], + "type": "text", + "content": " yields better performance, as the same mappings share the same scales." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 366, + 286, + 472 + ], + "blocks": [ + { + "bbox": [ + 50, + 366, + 286, + 472 + ], + "lines": [ + { + "bbox": [ + 50, + 366, + 286, + 472 + ], + "spans": [ + { + "bbox": [ + 50, + 366, + 286, + 472 + ], + "type": "table", + "html": "
Setting20%40%60%80%
Baseline94.33±0.4594.33±0.4094.33±0.494.33±0.45
+Pse.97.05±0.4895.98±0.7596.65±0.3595.85±0.88
+Pse.&W96.62±0.4796.97±0.7897.22±0.3896.00±0.48
w\\o g1(·)97.85±0.5796.98±0.1194.38±1.5294.38±0.74
w\\o g2(·)97.98±0.5396.85±0.1494.58±0.4695.85±0.35
g1(·) × g2(·)97.50±0.0797.15±0.7194.95±0.1494.48±0.46
g1(·) × g2(·)96.60±0.5796.93±0.0495.65±0.2195.65±0.07
WAD98.43±0.1497.88±0.3397.90±0.2097.77±0.33
", + "image_path": "d61de22678afc4d81c96fe2aa5c72b6b653d5cb731fd6816dfb10db20605c5c0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 471, + 280, + 482 + ], + "lines": [ + { + "bbox": [ + 47, + 471, + 280, + 482 + ], + "spans": [ + { + "bbox": [ + 47, + 471, + 280, + 482 + ], + "type": "text", + "content": "Table 3. Ablation Studies under different mismatch proportions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 502, + 162, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 162, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 162, + 514 + ], + "type": "text", + "content": "4.4. Sensitivity Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 521, + 289, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 665 + ], + "type": "text", + "content": "This subsection investigates the influence of parameter " + }, + { + "bbox": [ + 46, + 521, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 521, + 289, + 665 + ], + "type": "text", + "content": ", which controls how many instances with high reliability will be added to labeled data. Hence, we vary the initial value of " + }, + { + "bbox": [ + 46, + 521, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 521, + 289, + 665 + ], + "type": "text", + "content": " and evaluate WAD's performance on CIFAR10. The results are reported in Table 4. We find that WAD depicts the comparable performance with different values of " + }, + { + "bbox": [ + 46, + 521, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 521, + 289, + 665 + ], + "type": "text", + "content": ", although lower values of " + }, + { + "bbox": [ + 46, + 521, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 521, + 289, + 665 + ], + "type": "text", + "content": " achieve slightly better performance with " + }, + { + "bbox": [ + 46, + 521, + 289, + 665 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 521, + 289, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 289, + 665 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 521, + 289, + 665 + ], + "type": "text", + "content": " mismatch proportions. This indicates that WAD is not sensitive to " + }, + { + "bbox": [ + 46, + 521, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 521, + 289, + 665 + ], + "type": "text", + "content": " because the selected instances may have higher similarities. Thus, WAD can achieve a robust performance for a wide range of " + }, + { + "bbox": [ + 46, + 521, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 521, + 289, + 665 + ], + "type": "text", + "content": " but not too large, preventing the invasion of unknown categories." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 671, + 132, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 132, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 132, + 682 + ], + "type": "text", + "content": "4.5. Visualization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "To comprehend how WAD works, we visualize pseudo labels assigned to unlabeled instances with target categories" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 310, + 71, + 545, + 127 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 545, + 127 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 545, + 127 + ], + "type": "table", + "html": "
Setting20%40%60%80%
α = 0.0598.62±0.2898.12±0.3997.63±0.0397.63±0.19
α = 0.1098.43±0.1497.88±0.3397.90±0.2097.77±0.33
α = 0.1598.35±0.2597.73±0.1597.93±0.0897.42±0.63
", + "image_path": "adda62f75bbbd378271a29f55c30d8d164efb5429558dd57caaf8b9665221aad.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 128, + 539, + 138 + ], + "lines": [ + { + "bbox": [ + 306, + 128, + 539, + 138 + ], + "spans": [ + { + "bbox": [ + 306, + 128, + 539, + 138 + ], + "type": "text", + "content": "Table 4. Sensitivity analysis under varied mismatch proportions." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 156, + 547, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 547, + 264 + ], + "type": "text", + "content": "alongside the ground truths of labeled ones in different colors, as depicted in the left part of Figure 7. We observe that instances with target categories are separated into two clusters and follow the same distribution as labeled instances. Additionally, the weight distribution, shown on the right side of Figure 7, depicts that WAD assigns smaller weights to unknown categories and larger ones to target ones, making it feasible to filter out harmful unknown categories and to distill useful information from target ones." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 324, + 283, + 416, + 353 + ], + "blocks": [ + { + "bbox": [ + 324, + 283, + 416, + 353 + ], + "lines": [ + { + "bbox": [ + 324, + 283, + 416, + 353 + ], + "spans": [ + { + "bbox": [ + 324, + 283, + 416, + 353 + ], + "type": "image", + "image_path": "5a4b3ae2308cf15aed7c1a6ef281f7c379a6091b62e8477b7d6baf2acd87748b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 429, + 282, + 531, + 358 + ], + "blocks": [ + { + "bbox": [ + 429, + 282, + 531, + 358 + ], + "lines": [ + { + "bbox": [ + 429, + 282, + 531, + 358 + ], + "spans": [ + { + "bbox": [ + 429, + 282, + 531, + 358 + ], + "type": "image", + "image_path": "fa0514afd8bc2583d6db4a38b94d0a0ebc63b26ac2421cbaa0e7b826bd92c352.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 358, + 547, + 380 + ], + "lines": [ + { + "bbox": [ + 305, + 358, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 305, + 358, + 547, + 380 + ], + "type": "text", + "content": "Figure 7. Visualization of pseudo labels and weights on CIFAR10 under " + }, + { + "bbox": [ + 305, + 358, + 547, + 380 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 305, + 358, + 547, + 380 + ], + "type": "text", + "content": " class distribution mismatch." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 399, + 384, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 384, + 412 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 384, + 412 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 420, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 545, + 539 + ], + "type": "text", + "content": "To tackle class distribution mismatch in an SSL manner, we theoretically reveal that the SSL error is composed of pseudo-labeling error and invasion error under mismatch scenarios. Then, a distillation-based SSL framework, WAD, is proposed to transfer knowledge, such as pseudo labels and weights, from the representations to the target model. Theoretical analyses verify that the population risk of WAD is tightly bounded. Extensive experiments on two benchmark datasets and a cross-dataset demonstrate the superiority of WAD." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 540, + 547, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 547, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 547, + 575 + ], + "type": "text", + "content": "In the near future, we would like to investigate whether some instances from unknown categories are beneficial to target task and how to utilize them if so." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 586, + 416, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 416, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 416, + 600 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "This work is supported by the National Key Research & Develop Plan(2018YFB1004401), National Natural Science Foundation of China(62276270,62072460, 62172424), Beijing Natural Science Foundation(4212022), Fundamental Research Funds for Central University, and the Research Funds of Renmin University of China. It is also partially supported by the Opening Fund of Hebei Key Laboratory of Machine Learning and Computational Intelligence." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16418" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? Advances in neural information processing systems, 27, 2014. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 287, + 158 + ], + "type": "text", + "content": "[2] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. arXiv preprint arXiv:1906.00910, 2019. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 287, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 287, + 213 + ], + "type": "text", + "content": "[3] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix-match: Semi-supervised learning with distribution matching and augmentation anchoring. In International Conference on Learning Representations, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 287, + 259 + ], + "type": "text", + "content": "[4] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 287, + 293 + ], + "type": "text", + "content": "[5] Alexander Borichev and Yuri Tomilov. Optimal polynomial decay of functions and operator semigroups. Mathematische Annalen, 347(2):455-478, 2010. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 295, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 287, + 339 + ], + "type": "text", + "content": "[6] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European conference on computer vision, pages 446-461. Springer, 2014. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 341, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 287, + 384 + ], + "type": "text", + "content": "[7] Cristian Bucilua, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. In Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 535-541, 2006. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 386, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 287, + 418 + ], + "type": "text", + "content": "[8] Kaidi Cao, Maria Brbic, and Jure Leskovec. Open-world semi-supervised learning. In International Conference on Learning Representations, 2021. 1, 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 420, + 287, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 287, + 463 + ], + "type": "text", + "content": "[9] Olivier Chapelle, Bernhard Scholkopf, and Alexander Zien. Semi-supervised learning (chapelle, o. et al., eds.; 2006)[book reviews]. IEEE Transactions on Neural Networks, 20(3):542-542, 2009. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 465, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 287, + 509 + ], + "type": "text", + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "type": "text", + "content": "[11] Yanbei Chen, Xiatian Zhu, Wei Li, and Shaogang Gong. Semi-supervised learning under class distribution mismatch. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 3569-3576, 2020. 1, 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "text", + "content": "[12] Pan Du, Hui Chen, Suyun Zhao, Shuwen Chai, Hong Chen, and Cuiping Li. Contrastive active learning under class distribution mismatch. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-13, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "type": "text", + "content": "[13] Pan Du, Suyun Zhao, Hui Chen, Shuwen Chai, Hong Chen, and Cuiping Li. Contrastive coding for active learning under class distribution mismatch. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8927-8936, October 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "type": "text", + "content": "[14] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[15] Lan-Zhe Guo, Zhen-Yu Zhang, Yuan Jiang, Yu-Feng Li, and Zhi-Hua Zhou. Safe deep semi-supervised learning for" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "text", + "content": "unseen-class unlabeled data. In International Conference on Machine Learning, pages 3897-3906. PMLR, 2020. 1, 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 141, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 174 + ], + "type": "text", + "content": "[17] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. In NIPS Deep Learning and Representation Learning Workshop, 2015. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 175, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 230 + ], + "type": "text", + "content": "[18] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In International Conference on Learning Representations, 2018. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 232, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 296 + ], + "type": "text", + "content": "[19] Junkai Huang, Chaowei Fang, Weikai Chen, Zhenhua Chai, Xiaolin Wei, Pengxu Wei, Liang Lin, and Guanbin Li. Trash to treasure: Harvesting ood data with cross-modal matching for open-set semi-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8310-8319, 2021. 1, 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 298, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 341 + ], + "type": "text", + "content": "[20] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. A survey on contrastive self-supervised learning. Technologies, 9(1):2, 2021. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 342, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 387 + ], + "type": "text", + "content": "[21] Wonsik Kim, Bhavya Goyal, Kunal Chawla, Jungmin Lee, and Keunjoo Kwon. Attention-based ensemble for deep metric learning. In Proceedings of the European conference on computer vision (ECCV), pages 736-751, 2018. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 388, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 420 + ], + "type": "text", + "content": "[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 422, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 443 + ], + "type": "text", + "content": "[23] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 445, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 476 + ], + "type": "text", + "content": "[24] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "type": "text", + "content": "[25] Phuc H Le-Khac, Graham Healy, and Alan F Smeaton. Contrastive representation learning: A framework and review. IEEE Access, 2020. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 512, + 545, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 556 + ], + "type": "text", + "content": "[26] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 558, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 558, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 558, + 545, + 611 + ], + "type": "text", + "content": "[27] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979–1993, 2018. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "type": "text", + "content": "[28] Gaurav Kumar Nayak, Konda Reddy Mopuri, Vaisakh Shaj, Venkatesh Babu Radhakrishnan, and Anirban Chakraborty. Zero-shot knowledge distillation in deep networks. In International Conference on Machine Learning, pages 4743-4751. PMLR, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[29] M-E Nilsback and Andrew Zisserman. A visual vocabulary for flower classification. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), volume 2, pages 1447-1454. IEEE, 2006. 7" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16419" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 588 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[30] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3967-3976, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "text", + "content": "[31] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 288, + 196 + ], + "type": "text", + "content": "[32] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 198, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 287, + 262 + ], + "type": "text", + "content": "[33] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 263, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 263, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 263, + 287, + 308 + ], + "type": "text", + "content": "[34] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 309, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 309, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 49, + 309, + 287, + 352 + ], + "type": "text", + "content": "[35] Lin Wang and Kuk-Jin Yoon. Knowledge distillation and student-teacher learning for visual intelligence: A review and new outlooks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 354, + 286, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 354, + 286, + 374 + ], + "spans": [ + { + "bbox": [ + 49, + 354, + 286, + 374 + ], + "type": "text", + "content": "[36] Huan Xu and Shie Mannor. Robustness and generalization. Machine learning, 86(3):391-423, 2012. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 376, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 376, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 49, + 376, + 287, + 431 + ], + "type": "text", + "content": "[37] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14421-14430, 2022. 1, 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 433, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 433, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 49, + 433, + 287, + 464 + ], + "type": "text", + "content": "[38] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In *British Machine Vision Conference* 2016. British Machine Vision Association, 2016. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 466, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 466, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 49, + 466, + 287, + 509 + ], + "type": "text", + "content": "[39] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 11953-11962, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 511, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 511, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 49, + 511, + 287, + 543 + ], + "type": "text", + "content": "[40] Xujiang Zhao, Killamsetty Krishnateja, Rishabh Iyer, and Feng Chen. Robust semi-supervised learning with out of distribution data. arXiv preprint arXiv:2010.03658, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 544, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 544, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 49, + 544, + 287, + 588 + ], + "type": "text", + "content": "[41] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16420" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups/8e5c15a5-f270-46c8-829c-81448a8fde6f_content_list.json b/2023/Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups/8e5c15a5-f270-46c8-829c-81448a8fde6f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b02a9d76c81b17379ac170a62f8d2d1cf5250428 --- /dev/null +++ b/2023/Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups/8e5c15a5-f270-46c8-829c-81448a8fde6f_content_list.json @@ -0,0 +1,1599 @@ +[ + { + "type": "text", + "text": "Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups", + "text_level": 1, + "bbox": [ + 181, + 130, + 789, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Peixia Li $^{1,2}$ , Pulak Purkait $^{2}$ , Thalaiyasingam Ajanthan $^{2}$ , Majid Abdolshah $^{2}$ , Ravi Garg $^{2,4}$ , Hisham Husain $^{2}$ , Chenchen Xu $^{2,3}$ , Stephen Gould $^{3}$ , Wanli Ouyang $^{1}$ , Anton van den Hengel $^{2,4}$ , The University of Sydney, $^{2}$ Amazon, Australian National University, $^{4}$ The University of Adelaide", + "bbox": [ + 109, + 202, + 857, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semi-supervised semantic segmentation methods use a small amount of clean pixel-level annotations to guide the interpretation of a larger quantity of unlabelled image data. The challenges of providing pixel-accurate annotations at scale mean that the labels are typically noisy, and this contaminates the final results. In this work, we propose an approach that is robust to label noise in the annotated data. The method uses two diverse learning groups with different network architectures to effectively handle both label noise and unlabelled images. Each learning group consists of a teacher network, a student network and a novel filter module. The filter module of each learning group utilizes pixel-level features from the teacher network to detect incorrectly labelled pixels. To reduce confirmation bias, we employ the labels cleaned by the filter module from one learning group to train the other learning group. Experimental results on two different benchmarks and settings demonstrate the superiority of our method over state-of-the-art approaches.", + "bbox": [ + 75, + 342, + 472, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 632, + 209, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic segmentation is a critical step in many computer vision tasks, and one that, like so many, is now predominantly achieved through deep learning [5, 35]. A key factor in the success of these methods has been access to high-quality annotated datasets. Pixel-accurate annotation of segmentation data is challenging, and becomes impractical as datasets grow. Unlabelled data is abundant and thus semi-supervised semantic segmentation has received increasing attention [15, 20, 8]. The predominant semisupervised approach employs a teacher-student framework to allocate pseudo labels to unlabelled data [11, 23, 36, 40]. These methods rely on the availability of a small but perfectly annotated dataset, however. This is a challenge in many practical situations where annotations are collected via crowd-sourcing (e.g., Sagemaker MTurk), or automated labelling techniques [31, 3], and are thus inevitably noisy.", + "bbox": [ + 75, + 659, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fabdcc8a9daaa5e8c802b4313e12b071775d52eaeab3a87dfc4aec2851407865.jpg", + "image_caption": [ + "Figure 1: Existing semi-supervised semantic segmentation models are vulnerable to the inevitable label noise in the pixel-level annotations. The proposed method performs significantly better than the current state-of-the-art [20] in the presence label noise. We apply three types of noise (labelled PL, RDE and SCP in Sec. 5.2) to achieve a $9\\%$ error rate in the labelled data." + ], + "image_footnote": [], + "bbox": [ + 503, + 309, + 893, + 453 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We consider here the practically important case whereby semi-supervised semantic segmentation is to be applied despite the fact that the provided pixel labels contain noise. We propose an approach that is robust to this noise to the extent that it outperforms the baselines by a significant margin (see Fig. 1), despite being relatively simple to implement and incurring no extra computation cost at test time.", + "bbox": [ + 496, + 568, + 893, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "It has been shown recently [19] that supervised segmentation models trained on noisy pixel-level labels first fit the clean labels during an \"early-learning\" phase before eventually memorizing the noisy labels. The above learning dynamics are exploited in [19] for weakly supervised segmentation by adding a multi-scale consistency loss. However, no noise detection module is used in [19] and the effect of unsupervised consistency loss on the label noise is weak.", + "bbox": [ + 496, + 674, + 893, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently there has been a surge of interest in data-centric methods [38, 39] that utilize the $k$ -NN of the feature representation of the data-points to detect annotation noise in classification datasets. A straightforward extension of this approach to semantic segmentation would lead to computing the $k$ -NN of the feature representation at pixel-level over all images, which would be infeasible. In this work, we", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1229", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "adapt [39] to semi-supervised semantic segmentation by introducing an explicit label noise detection module that predicts label noise at the pixel level without significant additional memory or computation requirements.", + "bbox": [ + 75, + 90, + 467, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pseudo labelling approaches [11, 23, 36, 2] for semi-supervised segmentation often use a single model to generate pseudo labels on the unlabelled data as well as train on them. This leads to confirmation bias as the model can overfit to its prediction errors. The same bias is apparent even when a single model is employed to filter label noise and train on those cleaned labels [19]. To mitigate this effect, we introduce two diverse learning groups with different network architectures that encourage mutual knowledge distillation and effective noise filtering. These learning groups also exploit complementary learning paths that boost the competency of each group. Our contributions are thus summarized as follows:", + "bbox": [ + 75, + 152, + 467, + 347 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an approach to semi-supervised semantic segmentation that is robust to label noise.", + "- We present a general architecture that maintains two diverse learning groups to overcome confirmation bias in label assignment problems.", + "- We introduce a filter module that extends [39] for semantic segmentation while demonstrating a good trade-off between accuracy and computation cost.", + "- We conduct extensive experiments showing that our method significantly outperforms semi/weakly-supervised semantic segmentation baselines regardless of the presence or the type of label noise." + ], + "bbox": [ + 94, + 353, + 467, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 558, + 217, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semi-supervised Semantic Segmentation. Pixel-level annotation for semantic segmentation is time-consuming, costly, and error prone. Semi-supervised learning addresses the first two elements of this problem using unlabeled data. These approaches can be broadly categorized into self-training (a.k.a self-learning) based methods [11, 23, 10, 25, 29] and consistency regularization-based methods [40, 26, 6, 14, 20]. Self-training-based methods use models trained on labelled examples to predict pseudo-segmentation masks over the unlabeled examples, thus iteratively expanding the training set. The effectiveness of these methods depends on accurately identifying high-confidence pseudo labels via crafted auxiliary networks [11, 23, 36] or GANs [10, 25, 29]. Consistency regularization methods enforce consistency among predictions under varying perturbations, where the perturbations can be applied to the input image (French [8], PseudoSeg [40], CAC [16]), the intermediate feature representations (CCT [26]) or the network itself (CPS [6], GCT [14]). Most effort in this vein aims to identify perturbations that are challenging enough to support generalization without", + "bbox": [ + 75, + 583, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "changing the underlying true labels. Particularly, recent work [20] on Mean Teacher [30] demonstrates superior performance and flexibility in the choice of perturbations. Our work builds upon [20] and introduces a general framework with two learning groups and a noise filter module to effectively learn from noisy labelled data and unlabelled data.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic Segmentation with Label Noise. Segmentation is a more challenging task than classification, and the labelled data is more prone to label noise. To achieve robustness some methods explicitly model the human annotation errors [12], or add a regularization term [28, 32]. Interestingly, [22] shows how high-level structure information (about microscopy images) learned by the model can be used as additional signal to remedy wrong annotations. In [24], a two stream mutual-attention network is devised to smooth out the influence of noisy labels during backpropagation. ADELE [19] exploits the learning dynamics demonstrated by noisy examples and proposes an early-learning strategy accordingly. These supervised segmentation methods assume all of the data is labellebd and that the majority of the labels are correct. They thus focus on alleviating the influence a small set of incorrect labels.", + "bbox": [ + 496, + 193, + 890, + 434 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The approach most similar to that proposed here is that of DivideMix [18] whereby two divergent networks are applied to separate data partitions. The method there is specifically designed for classification, however, as it is based on an image-level Gaussian Mixture Model. We propose here instead a pixel-level filter module appropriate for semantic segmentation. Additionally, in contrast to [18], we employ multiple learning groups with varying architectures to encourage diversity.", + "bbox": [ + 496, + 435, + 890, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 578, + 635, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Problem Setup", + "text_level": 1, + "bbox": [ + 500, + 603, + 648, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We consider semi-supervised segmentation in the case where we have a small set of potentially noisy labelled examples and a large set of unlabelled examples. Let $\\mathcal{D}_L = \\{(x_i, y_i)\\}_{i=1}^{N_L}$ be the labelled set where $x_i \\in \\mathcal{X} := [0,1]^{h \\times w \\times 3}$ and $y_i \\in \\mathcal{Y} := \\mathcal{C}^{h \\times w}$ denote an image and the corresponding pixel-wise segmentation mask respectively. Here, $\\mathcal{C} = \\{1,2,\\dots,\\ell\\}$ denotes the set of labels. Similarly, let $\\mathcal{D}_U = \\{x_i\\}_{i=1}^{N_U}$ denote the unlabelled set of images.", + "bbox": [ + 496, + 626, + 890, + 761 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast to the standard semi-supervised segmentation problem, the segmentation masks $\\{\\pmb{y}_i\\}$ may contain erroneous labels.", + "bbox": [ + 496, + 762, + 890, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.2. Noisy Label Detection", + "text_level": 1, + "bbox": [ + 500, + 816, + 705, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here we briefly review the noisy label detection approach SimiFeat [38] developed for robust image classification. Our work extends this framework to the segmentation problem by pixel-level reasoning over noisy labels in", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1230", + "bbox": [ + 483, + 924, + 514, + 935 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ce750cb3b1cc39a49e1b4b65d976a0d803c597153fbe2c6441f37dbfa1307083.jpg", + "image_caption": [ + "Figure 2: The pipelines of a) the baseline model [20] and b) the proposed approach. In contrast to the baseline that uses two teachers and a single student network, we propose a framework with two diverse Learning Groups (LGs) $(t^1, s^1)$ and $(t^2, s^2)$ where each group contains a teacher and a student network. The two learning groups propagate complementary information and reduce the confirmation bias. Here, $t$ denotes a teacher model, $s$ is a student model, solid lines show the forward flow of labelled data, and dotted lines show the flow of unlabelled data. $L_{ce}$ denotes the cross-entropy loss and $L_{con}$ is the confidence-weighted cross-entropy loss." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 867, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sec. 4.2. SimiFeat estimates the label noise proportion for each class and then compute a per-class \"noisiness\" score for all the images in order to rank them and eliminate top $R_{c}$ percentile of images, where $R_{c}$ is the noise percentage for class $c$ .", + "bbox": [ + 75, + 410, + 468, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\tilde{\\mathcal{D}} = \\left\\{(\\pmb{x}_i,\\tilde{y}_i)\\right\\}$ be a dataset of images $\\pmb{x}_i\\in \\mathcal{X}$ with noisy labels $\\tilde{y}_i\\in \\mathcal{C}$ , and $g(\\cdot):\\mathcal{X}\\to \\mathbb{R}^d$ be a feature extractor. The noise proportion is computed from the prior probability of clean labels and the noise transition probability, which are estimated via High Order Consensus (HOC) (refer to [38] for more details) as follows:", + "bbox": [ + 75, + 484, + 468, + 575 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {P} (Y), \\mathbb {P} (\\tilde {Y} | Y) \\leftarrow \\operatorname {H O C} \\left(\\left\\{\\left(g (\\boldsymbol {x}), \\tilde {y}\\right) \\mid (\\boldsymbol {x}, \\tilde {y}) \\in \\tilde {\\mathcal {D}} \\right\\}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 585, + 468, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Y$ and $\\tilde{Y}$ denote the random variables corresponding to clean and noisy labels respectively. The noise proportion is then calculated for a class $c \\in \\mathcal{C}$ as:", + "bbox": [ + 75, + 614, + 468, + 659 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR _ {c} = 1 - \\mathbb {P} (Y = c | \\tilde {Y} = c). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 669, + 468, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note, $\\mathbb{P}(Y|\\hat{Y})$ is obtained via Bayes' rule:", + "bbox": [ + 76, + 696, + 356, + 713 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {P} (Y | \\tilde {Y}) = \\mathbb {P} (\\tilde {Y} | Y) \\mathbb {P} (Y) / \\mathbb {P} (\\tilde {Y}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 720, + 468, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbb{P}(\\tilde{Y})$ is computed from the observed labels and $\\mathbb{P}(Y),\\mathbb{P}(\\tilde{Y}|Y)$ are computed using HOC (Eq. (1)).", + "bbox": [ + 75, + 748, + 468, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To compute the noisiness score, a $k$ -NN based soft label is employed in the feature space via nearest-neighbour voting. Then, the score for an image is determined as the negative cosine similarity between the soft label of that image and the one-hot encoding of the class label [38]. Finally, the images are ranked based on the score and the noise proportion from Eq. (2) is utilized to determine the percentile of images to be declared as noisy and consequently removed.", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Our Approach", + "text_level": 1, + "bbox": [ + 500, + 409, + 645, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce a general pipeline with two Learning Groups (LGs), as shown in Fig. 2. Each LG consists of a teacher-student framework and a noisy label detection module (Filter Modules). The teacher network is used to produce pseudo labels for the unlabelled data similar to the semi-supervised learning pipeline in [20]. To stabilize training, we first train both LGs independently in a fully-supervised manner with the labelled dataset $\\mathcal{D}_L$ for a few epochs. After this warm-up, at each training epoch, we activate the filter module that detects the pixels with noisy labels in the labelled set $\\mathcal{D}_L$ and move them to the unlabeled set $\\mathcal{D}_U$ . Then, we train the pipeline in the semi-supervised manner. The noise detection and model training are repeated until the end of the training.", + "bbox": [ + 496, + 463, + 890, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our filter module uses the feature representation of the teacher model from one LG to partition the labelled data into clean and noisy sets. The filter module then ignores the labels of the noisy set and moves the corresponding elements into the unlabelled set to obtain a revised labelled and unlabelled dataset pair. The updated dataset pair generated by one LG is consumed by the other. As training progresses each LG receives a different set of dataset pair due to the decoupled noise filtering, which enables complementary information to propagate in the two learning paths - reducing confirmation bias. This, along with the different network architectures for the two LGs, enhances the performance of our framework in the presence of noise.", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1231", + "bbox": [ + 483, + 924, + 511, + 935 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1a9d8edcb8e60da39647ab5c66cba070c516d5462d14d88629062acd8d9464e6.jpg", + "image_caption": [ + "Figure 3: The noise detection results from the two LGs. The images from left to right are: input, clean ground-truth followed by noise detection results from the two LGs when we use the same architecture (two R50s) and different architectures (R50 and PVTv2-B2 [33]). We get more diverse noise detection results when using different architectures." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 472, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Learning Group", + "text_level": 1, + "bbox": [ + 76, + 335, + 240, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The architecture of the LG used in our pipeline is inspired by the recent teacher-student framework [20] as depicted in Fig. 2 (a). We refer the reader to [20] for more details.", + "bbox": [ + 75, + 361, + 468, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At a training epoch $k$ , the first LG receives the revised labelled set $\\mathcal{D}_{L,k}^{2}$ and unlabelled set $\\mathcal{D}_{U,k}^{2}$ after filtering the noisy labels predicted by the second LG (details of $\\mathcal{D}_{L,k}^{2}$ and $\\mathcal{D}_{U,k}^{2}$ can be found in Sec. 4.2). Then the student network of the first LG minimizes the following loss:", + "bbox": [ + 75, + 422, + 472, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(\\mathcal {D} _ {L, k} ^ {2}, \\mathcal {D} _ {U, k} ^ {2}, \\boldsymbol {\\theta} _ {s ^ {1}}\\right) := L _ {c e} \\left(\\mathcal {D} _ {L, k} ^ {2}, \\boldsymbol {\\theta} _ {s ^ {1}}\\right) + \\alpha L _ {c o n} \\left(\\mathcal {D} _ {U, k} ^ {2}, \\boldsymbol {\\theta} _ {s ^ {1}}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 515, + 475, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $L_{ce}$ denotes the cross-entropy loss on the labelled data, $L_{con}$ is a confidence-weighted cross-entropy loss that enforces consistency between the teacher and student model predictions on the unlabelled data, $\\theta_{s^1}$ denotes the parameters of the student network in the first LG, and $\\alpha > 0$ provides a relative weight between the two losses. Analogously, the parameters of the second student network $\\theta_{s^2}$ are optimized utilizing the revised labelled set $\\mathcal{D}_{L,k}^{1}$ and the unlabelled dataset $\\mathcal{D}_{U,k}^{1}$ predicted by the first LG. The teacher network is updated by the moving average of the student network parameters of the same LG.", + "bbox": [ + 75, + 547, + 468, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that, at each training epoch, the set of labelled data varies depending on the filter module output. Specifically, the detected pixels with noisy labels will be removed from the label set, which leads to modifying the segmentation mask of the labelled examples with an \"ignore-pixel\" label that enables the loss Eq. (4) to be defined in terms of $L_{con}$ instead of $L_{ce}$ at those pixels.", + "bbox": [ + 75, + 717, + 468, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To encourage diversity of the LGs, we utilize R50 [9] for the first LG and PVTv2-B2 [33] for the second. In Fig. 3, we show that the noise detection results from two R50 (the third and fourth columns) have less diversity than those from different network architectures (the last two", + "bbox": [ + 75, + 824, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a4d0dd96f18de45b65674825f033bd7cb9e374e1dc309cea7f41aba3eb0d0003.jpg", + "image_caption": [ + "Figure 4: The architecture of our filter module. First, we generate a representative dataset $\\bar{\\mathcal{D}}$ and calculate the noise proportion based on HOC [39]. Then, we get the adaptive threshold based on the noise proportion and example score. According to the adaptive threshold and example score, the original dataset $\\mathcal{D}_L$ is divided into a clean set $\\mathcal{D}_L^1$ and a noisy set. The noisy set is added to the unlabelled dataset $\\mathcal{D}_U$ ." + ], + "image_footnote": [], + "bbox": [ + 503, + 88, + 893, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "columns).1 We also show that two different architectures bring better performance in our experiments (Fig. 7).", + "bbox": [ + 496, + 388, + 890, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.Filter Module", + "text_level": 1, + "bbox": [ + 498, + 431, + 643, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The task of the filter module is to detect incorrectly labelled pixels efficiently. To this end, rather than relying on label correction approaches [19], we design a filter module that accurately detects pixel-level noise scalably. The filter module is shown in Fig. 4.", + "bbox": [ + 496, + 455, + 890, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that, according to Eq. (1), the HOC approach requires a dataset of feature encodings and the corresponding labels. Since semantic segmentation is a classification problem at the pixel-level, the natural extension is to treat each pixel as a datapoint and directly apply HOC. Unfortunately, this becomes intractable as the computational complexity is quadratic in number of pixels, labels, and number of images in the dataset.", + "bbox": [ + 496, + 532, + 892, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To mitigate this, we use a small subset of pixels per class to generate a representative set $\\bar{\\mathcal{D}} = \\{(\\pmb{x}_i,j,\\tilde{y}_{i,j})\\}$ , which comprises of a fixed number of pixels per image subsampled from $\\mathcal{D}_L$ . Here, $\\pmb{x}_i\\in \\mathcal{X}$ is an image, $j\\in \\mathcal{I}(\\pmb{x}_i)$ is the index of a pixel in image $\\pmb{x}_i$ , and $\\tilde{y}_{i,j}\\in \\mathcal{C}$ is the label of the $j$ -th pixel. In order to generate such a set, we found that random pixel sampling is more effective than elaborate methods such as $k$ -means clustering, and is more computationally efficient also.", + "bbox": [ + 496, + 652, + 893, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $g_{j}^{1}(\\cdot):\\mathcal{X}\\to \\mathbb{R}^{d}$ denotes the backbone feature extractor, i.e., the penultimate layer output for the first LG at pixel $j$ (where $j\\in \\mathcal{I}(\\pmb {x}_i)\\coloneqq \\{1,\\dots ,w\\times h\\}$ if images are resized to a fixed size). We construct the embedded repre", + "bbox": [ + 496, + 789, + 893, + 849 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1It should be noted that the two networks with R50 architecture are trained with different initialization independently on the labelled data for a warm start for a fair comparison.", + "bbox": [ + 498, + 862, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1232", + "bbox": [ + 483, + 924, + 514, + 936 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sentative set $\\bar{\\mathcal{D}}_G^1 \\coloneqq \\left\\{(g_j^1(\\boldsymbol{x}_i), \\tilde{y}_{i,j}) \\mid (\\boldsymbol{x}_i, j, \\tilde{y}_{i,j}) \\in \\bar{\\mathcal{D}}\\right\\}$ and invoke HOC (Eq. (1)) to estimate the noise proportion $\\{R_c^1\\}_{c \\in \\mathcal{C}}$ at pixel-level (Eq. (2)). Analogously, the noise proportion $\\{R_c^2\\}_{c \\in \\mathcal{C}}$ for the other LG can be estimated by constructing $\\bar{\\mathcal{D}}_G^2$ .", + "bbox": [ + 75, + 89, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $h_j^1(\\cdot) : \\mathcal{X} \\to [0,1]^{\\ell}$ and $h_j^2(\\cdot) : \\mathcal{X} \\to [0,1]^{\\ell}$ be the teacher model mappings to softmax scores in the first and second LG respectively. We compute a noisiness score based on the teacher model predictions, as opposed to using $k$ -NN in SimiFeat [38]. This eliminates the computation cost of running $k$ -NN for every pixel over all images in $\\mathcal{D}_L$ . We define the score for the first LG at pixel $j$ in image $x_i$ as:", + "bbox": [ + 75, + 166, + 470, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS ^ {1} \\left(\\boldsymbol {x} _ {i}, j\\right) = \\operatorname {C E} \\left(h _ {j} ^ {1} \\left(\\boldsymbol {x} _ {i}\\right), \\boldsymbol {e} _ {i, j}\\right) + \\lambda \\operatorname {K L} \\left(h _ {j} ^ {1} \\left(\\boldsymbol {x} _ {i}\\right) \\| h _ {j} ^ {2} \\left(\\boldsymbol {x} _ {i}\\right)\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 295, + 468, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where CE and KL denote the cross-entropy and Kullback-Leibler (KL) divergence respectively, $e_{i,j} \\in \\{0,1\\}^{\\ell}$ is the one-hot encoded noisy label at pixel $j$ in image $x_i$ , and $\\lambda > 0$ . CE measures the alignment between label noise and model predictions, and KL measures the inconsistency between the two teacher model predictions. Large CE at a pixel implies a high misalignment between the predicted class label and annotated label and hence high model confidence of the label to be noisy unless the current model overfits at that pixel. On the other hand, if one of the models starts to overfit the noisy annotations, then the KL measure will be high. If any of them is high, such pixel corresponds to a high score, and it is marked as noisy. Experimentally, we verify that the above choice of score works better than the other alternatives (shown in Fig. 8(b)).", + "bbox": [ + 75, + 327, + 468, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once the noisiness score and the noise proportion have been calculated, we get the adaptive noise threshold as follows:", + "bbox": [ + 75, + 555, + 468, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\beta^ {1} = \\frac {\\sigma \\ell S _ {a v g 1} ^ {1}}{\\sum_ {c \\in \\mathcal {C}} R _ {c} ^ {1}}, \\text {w h e r e} S _ {a v g 1} ^ {1} = n ^ {- 1} \\sum_ {\\substack {\\left(\\boldsymbol {x} _ {i},\\cdot\\right) \\in \\mathcal {D} _ {L}, \\\\ j \\in \\mathcal {I} \\left(\\boldsymbol {x} _ {i}\\right)}} S ^ {1} \\left(\\boldsymbol {x} _ {i}, j\\right), \\tag{6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 604, + 468, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma$ is a hyperparameter and $n$ denotes the number of all the pixels in $\\mathcal{D}_L$ . Experimentally we found that the noise proportion $R_{c}$ predicted by HOC is inaccurate for some of the classes and a single threshold $\\beta^1$ , computed from the average as in Eq. (6), works better than class-wise thresholds. A higher $\\beta$ means filtering out less noisy annotations. We utilize the adaptive threshold $\\beta^1$ to determine the pixels to be removed and marked in the segmentation masks to reflect the pixels with noisy labels. Let us denote the set of pixels that are detected as label noise by the first LG as:", + "bbox": [ + 75, + 665, + 470, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {k} ^ {1} := \\left\\{\\left(\\boldsymbol {x} _ {i}, j\\right) \\mid S ^ {1} \\left(\\boldsymbol {x} _ {i}, j\\right) > \\beta^ {1}, \\left(\\boldsymbol {x} _ {i}, \\cdot\\right) \\in \\mathcal {D} _ {L}, j \\in \\mathcal {I} \\left(\\boldsymbol {x} _ {i}\\right) \\right\\} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 821, + 480, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The updated datasets at epoch $k$ are created by removing the pixels with noise labels from the labelled data and placing the images with noisy pixels in the unlabelled set, i.e., the", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "Algorithm 1 Our training pipeline \nInput: $\\mathcal{D}_L, \\mathcal{D}_U$ , total epochs $E$ , warm-up epochs $N$ \nOutput: Optimized $\\{\\theta_{s^1}, \\theta_{t^1}, \\theta_{s^2}, \\theta_{t^2}\\}$ \n1: Initialize parameters $\\{\\theta_{s^1}, \\theta_{t^1}, \\theta_{s^2}, \\theta_{t^2}\\}$ , \n2: for $k \\in \\{0, 1, \\dots, E\\}$ do \n3: if $k < N$ then \n4: $\\mathcal{D}_{L,k}^{1}, \\mathcal{D}_{L,k}^{2} \\coloneqq \\mathcal{D}_{L}$ and $\\mathcal{D}_{U,k}^{1}, \\mathcal{D}_{U,k}^{2} \\coloneqq \\emptyset$ \n5: else \n6: $\\{(x_i, j, \\tilde{y}_{i,j})\\} \\gets$ Sample from $\\mathcal{D}_L$ \n7: for $l \\in \\{1, 2\\}$ do \n8: $\\{R_c^l\\} \\gets$ HOC $\\left(\\left(g_j^l(x_i), \\tilde{y}_{i,j}\\right)\\right)$ \n9: Scores $\\{S^l\\} \\gets$ Using Eq. (5) \n10: $\\beta^l \\gets$ Adaptive $\\left(\\{S^l\\}\\right)$ [Eq. (6)] \n11: $\\mathcal{M}_k^l \\gets$ Using Eq. (7) \n12: $\\mathcal{D}_{L,k}^l \\gets \\mathcal{D}_L \\backslash \\mathcal{M}_k^l$ \n13: $\\mathcal{D}_{U,k}^l \\gets \\mathcal{D}_U \\bigcup \\{x_i | (x_i, \\cdot) \\in \\mathcal{M}_k^l\\}$ \n14: repeat \n15: $\\theta_{s^1} \\gets$ Backprop. $\\mathcal{L}(\\mathcal{D}_{L,k}^2, \\mathcal{D}_{U,k}^2, \\theta_{s^1})$ [Eq. (4)] \n16: $\\theta_{s^2} \\gets$ Backprop. $\\mathcal{L}(\\mathcal{D}_{L,k}^1, \\mathcal{D}_{U,k}^1, \\theta_{s^2})$ [Eq. (4)] \n17: until $M$ times \n18: Update $\\theta_{t^1}, \\theta_{t^2} \\gets$ Moving Average of $\\theta_{s^1}, \\theta_{s^2}$ .", + "bbox": [ + 500, + 90, + 890, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "clean labelled set $\\mathcal{D}_{L,k}^{1} \\coloneqq \\mathcal{D}_{L} \\setminus \\mathcal{M}_{k}^{12}$ and the unlabelled dataset $\\mathcal{D}_{U,k}^{1} \\coloneqq \\mathcal{D}_{U} \\bigcup \\{\\pmb{x}_{i} \\mid (\\pmb{x}_{i},\\cdot) \\in \\mathcal{M}_{k}^{1}\\}$ . The student network of the second LG is then trained on these datasets according to Eq. (4). Analogously, the score for the second LG can be computed by switching $h_{j}^{1}$ and $h_{j}^{2}$ in Eq. (5), and the datasets $\\mathcal{D}_{L,k}^{2}, \\mathcal{D}_{U,k}^{2}$ are created in the above manner on which the student network of the first LG is trained. Our final algorithm is summarized in Algorithm 1.", + "bbox": [ + 496, + 450, + 890, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 582, + 632, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation has been carried out following [19, 20] to allow direct comparison. As in [20], the labelled data is augmented with synthetic label noise. Performance is measured with, and without an unlabelled dataset to illuminate performance differences between the weakly and semi supervised settings.", + "bbox": [ + 496, + 607, + 890, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 707, + 718, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize R50 [9] and PVTv2-B2 [33] as the backbones in the two LGs to encourage diversity. Both the teacher and student models are based on DeeplabV3+ [4]. For inference, we use the average of the predictions from the two teacher networks as in [20]. The $\\lambda$ in Eq. (5) is set to 0.1 to balance the contribution from the CE and KL terms in our loss. We randomly sample 100 pixels per class in each image to get the representative dataset $\\bar{D}$ . The hyperparameter $\\sigma$ in Eq. (6) is set to be 10 in our experiments. Other", + "bbox": [ + 496, + 729, + 890, + 867 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2With a slight abuse of notation, this operation indicates marking all pixels in $\\mathcal{M}_k^1$ with \"ignore-pixel\" label in $\\mathcal{D}_L$ .", + "bbox": [ + 500, + 875, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1233", + "bbox": [ + 483, + 924, + 514, + 936 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5f58705d0981877f2403b48916a66bbf44e65d2cd4a9f7f68fcb4a55fe2c0d25.jpg", + "table_caption": [ + "Table 1: Comparison against SotA approaches on Pascal VOC 2012 in a semi-supervised semantic segmentation setting. All baselines are based on the DeeplabV3+ architecture. The * indicates results reported by [6]. When we replace $50\\%$ of the ground truth of images with 'PL', the proportion of noisy pixels is $9\\%$ . We keep the same pixel level noise proportion while introducing noise using 'RDE' and 'SCP' for parity." + ], + "table_footnote": [], + "table_body": "
MethodsYearNoisy pixelNoise typeLabelled Data Ratio
1/16 (662)1/8(1323)1/4 (2646)1/2 (5291)
MT* [30]20170%None66.7070.7873.2275.41
French [8]20190%None68.9070.7072.4674.49
CCT* [26]20200%None65.2270.8773.4374.75
GCT* [14]20200%None64.0570.4773.4575.20
ECS* [23]20200%None-67.3870.7072.89
CPS* [6]20210%None71.9873.6774.9076.15
CAC* [16]20210%None70.1072.4074.00-
PS-MT [20]20220%None72.8375.7076.4377.88
Ours-0%None77.75 (+4.92)79.31 (+3.61)79.14 (+2.71)79.54 (+1.66)
PS-MT [20]-9%PL61.9065.1465.5366.78
Ours-9%PL71.60 (+9.70)73.95 (+8.81)74.53 (+9.00)74.44 (+7.66)
PS-MT [20]-9%RDE58.8762.1863.8163.46
Ours-9%RDE66.82 (+7.95)69.95 (+7.77)73.82 (+10.01)74.47 (+11.01)
PS-MT [20]-9%SCP52.1457.1956.8153.69
Ours-9%SCP67.80 (+15.66)70.83 (+13.64)69.38 (+12.57)75.95 (+22.26)
", + "bbox": [ + 76, + 162, + 898, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "training and inference settings are the same as those of the baseline method [20]. In all of our experiments, the last epoch checkpoint model is used unless otherwise specified. Further details of the implementation can be found in the supplementary materials.", + "bbox": [ + 75, + 479, + 470, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Semi-supervised Semantic Segmentation", + "text_level": 1, + "bbox": [ + 76, + 566, + 423, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pascal VOC 2012 [7] is employed for evaluating the performance of our model in a semi-supervised semantic segmentation setting. This dataset contains more than $13\\mathrm{k}$ images with 21 different classes. To introduce realistic noise in $\\mathcal{D}_L$ , the following three approaches are considered:", + "bbox": [ + 75, + 590, + 468, + 666 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Pseudo Labels (PL). Following [19], a portion of clean labels are replaced by pseudo-labels from a weakly-supervised segmentation method [34].", + "- Random Dilation and Erosion (RDE). Similar to [19, 21], we apply random degrees of dilation and erosion to the ground-truth masks.", + "- Similar Class Perturbation (SCP). We introduce another type of noise by randomly replacing ground-truth label of a segmentation mask with a different class that resembles the texture of original clean class. This label perturbation simulates a more natural scenario pertaining to annotators confusing semantics." + ], + "bbox": [ + 94, + 676, + 470, + 864 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/258e736694f9fbb32fe84882b00d6438ecedc057edcbc0aa017b6d9439ac6b81.jpg", + "image_caption": [ + "Figure 5: A performance comparison under different noise proportions achieved by replacing (from $0\\%$ to $90\\%$ ) clean segmentation masks with the PL from [34] at image-level. The green font shows the corresponding noisy pixel proportions. All experiments are conducted with $1/8$ labelled data." + ], + "image_footnote": [], + "bbox": [ + 519, + 478, + 870, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Details about the implementation and visualization of noise generation can be found in the supplementary material.", + "bbox": [ + 496, + 729, + 890, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the context of semi-supervised semantic segmentation, we have compared our approach with the following studies: MT [30], French [8], CCT [26], GCT [14], ECS [23], CPS [6], CAC [16], and PS-MT [20]. Following the previous studies [19, 20], we report the mean Intersection-over-Union (mIoU) on the validation set. Table 1 shows that our approach ('Ours') outperforms all baselines trained with clean labelled data. In the presence of noise, we compare the best performing semi-supervised segmentation method", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3Note that, SCP does not change the shape of the clean segmentation masks, but only labels may be switched.", + "bbox": [ + 76, + 875, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1234", + "bbox": [ + 483, + 924, + 514, + 936 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2983e5d01764c9b7d9a2f4b7dd06e285ccb1fa7b63366e193750e5d925a6466a.jpg", + "table_caption": [ + "Table 2: The mIoU of the weakly-supervised semantic segmentation baselines and ours on the test set of SegTHOR. Best Val and Last Epoch correspond to the checkpoint that performs best on the validation set and the last checkpoint." + ], + "table_footnote": [], + "table_body": "
ModelBest ValLast Epoch
Base-ADELE [19]62.659.1
ADELE [19]71.170.8
Ours73.273.5
", + "bbox": [ + 98, + 161, + 446, + 233 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with no noise, i.e., PS-MT [20], against our method with varying noise types (see Table 1) and noise proportions (see Fig. 5). The results demonstrate significant improvements in mIoU in all cases, e.g., $+22.26\\%$ with half of the labelled data on the more practical SCP noise. Overall, the superiority of our method is consistent across all labelled data to unlabelled data ratios, regardless of the presence of noise or the type of applied noise.", + "bbox": [ + 75, + 250, + 468, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Weakly-supervised Semantic Segmentation", + "text_level": 1, + "bbox": [ + 76, + 380, + 442, + 397 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our method in a weakly-supervised semantic segmentation setting on the following datasets: (i) SegTHOR [17], a medical dataset, and (ii) pseudo labelled Pascal VOC 2012.", + "bbox": [ + 75, + 402, + 468, + 464 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Medical image segmentation on SegTHOR. Following [19], we evaluate our approach on the SegTHOR dataset, which consists of 3D CT scans. Each pixel of these images is assigned to one of these five classes: 'esophagus', 'heart', 'trachea', 'aorta', or 'background'. The slices from the images are randomly split into a training set of 3638 slices, a validation set of 570 slices, and a test set of 580 slices [19, 17]. To introduce noise, we apply 'RDE' to the ground-truth segmentation labels, similar to [19]. Note that all annotations in the training set are corrupted by this noise, whereas the validation and test sets remain uncorrupted.", + "bbox": [ + 75, + 470, + 468, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our approach with ADELE [19], which corrects noisy annotations with model predictions in the early training stage, and its variant, Base-ADELE [19], which is based on a UNet trained with multi-scale inputs. Detailed implementations of these approaches are in the supplementary material. Following the setting in ADELE [19], we evaluate two checkpoints: specifically, the one with the highest mIoU on the validation set and the last checkpoint. In Table 2, our approach outperforms both baselines.", + "bbox": [ + 75, + 637, + 468, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Weakly-supervised segmentation on Pascal VOC 2012. We demonstrate the effectiveness of our approach in comparison to the baselines in the weakly-supervised setting on the PASCAL VOC 2012 [7]. Similar to ADELE [19], we use only noisy pseudo labels generated by [34] to train the models and evaluate on the original test set. Table 3 shows that our approach outperforms recent baselines, such as ADELE combined with SEAM [19]. In addition, PS-MT", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "increases the mIoU marginally with PVTv2 compared to R50 with 66.7 vs. 64.0. However, after using our framework, the mIoU is further improved from 66.7 to 71.7, clearly demonstrating the effectiveness of our method.", + "bbox": [ + 496, + 90, + 890, + 152 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/515410c5c936a50758f0de1d6c0e88e2079c05fc331648458f796ca8ebb56334.jpg", + "table_caption": [ + "Table 3: Comparison with state-of-the-art weakly supervised semantic segmentation methods on the Pascal VOC 2012 dataset using mIoU (\\%) of the validation set." + ], + "table_footnote": [], + "table_body": "
ModelYearmIoU (%)
AffinityNet [1]201861.7
SSDD [27]201964.9
SEAM [34]202064.5
CONTA [37]202066.1
SPML [13]202169.5
ADELE [19] + SEAM202269.3
PS-MT [20] (R50)202264.0
PS-MT [20] (PVTv2-B2)202266.7
Ours-71.7
", + "bbox": [ + 514, + 222, + 875, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 411, + 651, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All experiments in this section are conducted with $1/8$ labelled data ratio and $9\\%$ PL noisy annotations on Pascal VOC. We verify the effectiveness of the two main components in our approach, the diverse LGs and filter module.", + "bbox": [ + 496, + 434, + 890, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Diversity in the LGs with different architectures. In contrast to previous works [20, 18], we found that using different network architectures in the two LGs provides more complementary information. In Fig. 3, we show the qualitative results of noise detection when we use the same and different architectures in the two LGs. The two LGs provide more diverse noise detection results (lower overlap between predicted noisy labels) when we employ different architectures. In Fig. 6, the predictions from the two LGs with two different architectures $(\\mathrm{e})\\& (\\mathrm{f})$ are more complementary than those from the two LGs with the same architecture $(\\mathrm{(c)})\\& (\\mathrm{(d)})$ . Fig. 7 shows the mIoU of our pipeline when we use different network combinations. The combination of R50 and PVTv2-B2 [33] achieves the best performance.", + "bbox": [ + 496, + 496, + 892, + 707 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Noise detection with different LGs. To avoid confirmation bias, we utilize the teacher network from one LG to detect noise for the other LG. If we utilize the teacher network from the same LG for noise detection (i.e., replacing $\\theta_{s^1}$ with $\\theta_{s^2}$ in Eq. (4)), the mIoU drops from $74.0\\%$ to $72.4\\%$ on Pascal VOC, demonstrating the importance of mutual noise detection with a different LG in our pipeline.", + "bbox": [ + 496, + 709, + 890, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Significance of the filter module. In Table 4, in addition to the baseline PS-MT and ours, we evaluate the performance of the following models: 1) 'PS-MT + FM' by adding the filter module to PS-MT (using R50)4. 2) 'Ours - FM' by", + "bbox": [ + 496, + 816, + 890, + 878 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "4Note, PS-MT can be regarded as single LG (Fig. 2 (a)) and the mean", + "bbox": [ + 514, + 886, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1235", + "bbox": [ + 483, + 924, + 514, + 936 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a3db4253c085eb04cdbfe4fe6c83696905a8191a60ac4f6d3fb283aa39bd780.jpg", + "image_caption": [ + "(a) Input Image" + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 223, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c3b5a11eede915f6419485aea4029dc925796f7fc31657df27010a2bab5e27dd.jpg", + "image_caption": [ + "(b) Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 223, + 90, + 352, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8c003e5fd52937d8a5ffab945a6e83b638ae187b59ec986f7dc6b5cf9f29ce0a.jpg", + "image_caption": [ + "(c) R50LG1 Prediction" + ], + "image_footnote": [], + "bbox": [ + 354, + 90, + 483, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6b9b8830d77bc507268576a0cb02b8bcea6acd9c924c76e1472c5f479bfbdead.jpg", + "image_caption": [ + "(d) R50LG2 Prediction" + ], + "image_footnote": [], + "bbox": [ + 486, + 90, + 614, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d9c76a0db3e7ec9963e96ec63cf1d93271eec4d13f121f3acb6e588de4cc440.jpg", + "image_caption": [ + "(e) R50LG1 Prediction", + "(f) PVT $^{\\mathrm{LG2}}$ Prediction" + ], + "image_footnote": [], + "bbox": [ + 616, + 90, + 746, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8460ca2812ce4034caad854523fae174e22b50c336a51cfb945117f94ae98073.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 90, + 879, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/735f67257470922de3132b40b16bd722951ddc80d3ed1afc5b1ab0b186d69b27.jpg", + "table_caption": [ + "Table 4: Ablation study of FM and LG on Pascal VOC." + ], + "table_footnote": [], + "table_body": "
Method2 LGsFilter Module (FM)mIoU (%)
PS-MT [20]65.1
PS-MT + FM✓67.6 (+2.5)
Ours - FM✓69.9 (+4.8)
Ours✓✓74.0 (+8.9)
", + "bbox": [ + 94, + 294, + 452, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bc98f6d422e6c3b0ebd1df1df6d9fe07ddcdf6443f7957bc070b2247c618456e.jpg", + "image_caption": [ + "(a) Different Architectures", + "Figure 7: (a) shows the mIoU score of different network combinations on clean data. (b) shows the mIoU score of different combinations on different settings. R, S and P denote R50 [9], Segformer [35] and PVTv2-B2 [33] respectively. -FM_C denotes training our model (without the filter module) with clean data. -FM_N denotes training model with noisy data without using filter module. +FM_N is our complete model trained with noisy data." + ], + "image_footnote": [], + "bbox": [ + 81, + 402, + 263, + 513 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e94072a5805a4bebca868e757affaedee2b0acacfdbca7dcaccfb246204a5c09.jpg", + "image_caption": [ + "(b) Clean and Noisy Data", + "Figure 8: (a) shows detected noise proportion of GMM [18] and ours, where $x$ -axis shows the noisy image ratio and $y$ -axis shows the detected noisy pixel ratio. The detected noise proportion of ours is closer to the GT than GMM. (b) shows the F1 score of noise detection with different criterion. CE is cross entropy loss; KL is KL divergence between the predictions from the two LGs. Our $\\mathrm{CE} + \\mathrm{KL}$ performs the best." + ], + "image_footnote": [], + "bbox": [ + 266, + 402, + 467, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "removing the filter module from our framework. It can be seen that the filter module helps to improve the final performance for both models. The filter module contributes more significantly (+4.1) to our framework due to the diverse learning groups (i.e., using 2 LGs with different architectures). This superiority of our framework is further evidenced by the better performance of 'Ours - FM' compared to PS-MT variants. In Fig. 7 (b), our Filter Module helps to improve the performance under noisy annotations ('+FM_N' vs. '-FM_N') with different architectures.", + "bbox": [ + 75, + 672, + 467, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison of the filter module with GMM. Previous work [18] utilizes Gaussian Mixture Model (GMM) to detect noise for image classification. We compare our filter", + "bbox": [ + 75, + 827, + 467, + 872 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e09b7a2ab53bb65aa0c01d31ed3520279e4011d22e393a5e93d2b1ab14654628.jpg", + "image_caption": [ + "Figure 6: Visualizations from the two LGs when we use the same architecture ((c) and (d)) and different architectures ((e) and (f)). The $1st$ and $2nd$ columns show the inputs and the ground truth. R50 and PVT denote R50 and PVTv2-B2 [33].", + "(a) Detected Noise Ratio", + "(b) F1 Score of Noise Detection" + ], + "image_footnote": [], + "bbox": [ + 514, + 268, + 707, + 398 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eda195d2494cfea237ab1c65eec41cd4ee24dfda863400c1cf98d8a3569b300a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 268, + 844, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "module with GMM (including two components) on the representative dataset. Fig. 8 (a) shows the detected noise proportions. We can see that the detected noise proportion of our filter module is closer to the groundtruth than GMM.", + "bbox": [ + 496, + 544, + 890, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Criterion to compute score. We utilize cross entropy loss and KL divergence $(\\mathrm{CE} + \\mathrm{KL})$ to compute score in Eq. (5). Here, we compare this criterion with others, including only cross entropy loss (CE), and only KL divergence (KL). Fig. 8 (b) shows that ours achieves the best F1 score.", + "bbox": [ + 496, + 607, + 890, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Memory and computation costs. PVTv2-B2 has similar GFLOPs (45.8) with R50 (45.6). Ours only have one more student network than PS-MT, which is not used during inference. Thus, our model needs similar memory and computation costs compared with PS-MT during inference.", + "bbox": [ + 496, + 684, + 890, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 784, + 617, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a learning framework for semi-supervised semantic segmentation involving two diverse learning groups that can handle various labelling errors. Our framework employs a carefully designed filter module to divide the labelled dataset into pixels with and without label noise in a scalable manner. The potentially misla", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "teacher output is used for filter module.", + "bbox": [ + 76, + 887, + 285, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1236", + "bbox": [ + 483, + 925, + 514, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "belled pixels filtered by the first learning group are utilised for training the second learning group and vice versa. We show that our proposed framework provides state-of-the-art results for semi-supervised learning, by addressing confirmation bias while training, that can handle various amounts of annotation noise to accommodate practical inaccuracies in annotations.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 208, + 174, + 224 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4981-4990, 2018.", + "[2] Eric Arazo, Diego Ortego, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2020.", + "[3] Liyi Chen, Weiwei Wu, Chenchen Fu, Xiao Han, and Yuntao Zhang. Weakly supervised semantic segmentation with boundary exploration. In European Conference on Computer Vision, pages 347-362. Springer, 2020.", + "[4] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017.", + "[5] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017.", + "[6] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021.", + "[7] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010.", + "[8] Geoff French, Timo Aila, Samuli Laine, Michal Mackiewicz, and Graham Finlayson. Semi-supervised semantic segmentation needs strong, high-dimensional perturbations. 2019.", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.", + "[10] Wei-Chih Hung, Yi-Hsuan Tsai, Yan-Ting Liou, Yen-Yu Lin, and Ming-Hsuan Yang. Adversarial learning for semi-supervised semantic segmentation. arXiv preprint arXiv:1802.07934, 2018.", + "[11] Mostafa S Ibrahim, Arash Vahdat, Mani Ranjbar, and William G Macready. Semi-supervised semantic image segmentation with self-correcting networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12715-12725, 2020." + ], + "bbox": [ + 78, + 233, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Joseph Jacob, Olga Ciccarelli, Frederik Barkhof, and Daniel C Alexander. Disentangling human error from the ground truth in segmentation of medical images.", + "[13] Tsung-Wei Ke, Jyh-Jing Hwang, and Stella X Yu. Universal weakly supervised segmentation by pixel-to-segment contrastive learning. arXiv preprint arXiv:2105.00957, 2021.", + "[14] Zhanghan Ke, Di Qiu, Kaican Li, Qiong Yan, and Rynson WH Lau. Guided collaborative training for pixel-wise semi-supervised learning. In European conference on computer vision, pages 429-445. Springer, 2020.", + "[15] Donghyeon Kwon and Suha Kwak. Semi-supervised semantic segmentation with error localization network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9957-9967, 2022.", + "[16] Xin Lai, Zhuotao Tian, Li Jiang, Shu Liu, Hengshuang Zhao, Liwei Wang, and Jiaya Jia. Semi-supervised semantic segmentation with directional context-aware consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1205-1214, 2021.", + "[17] Zoé Lambert, Caroline Petitjean, Bernard Dubray, and Su Kuan. Segthor: segmentation of thoracic organs at risk in ct images. In 2020 Tenth International Conference on Image Processing Theory, Tools and Applications (IPTA), pages 1-6. IEEE, 2020.", + "[18] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394, 2020.", + "[19] Sheng Liu, Kangning Liu, Weicheng Zhu, Yiqiu Shen, and Carlos Fernandez-Granda. Adaptive early-learning correction for segmentation from noisy annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2606–2616, 2022.", + "[20] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4258-4267, 2022.", + "[21] Yaoru Luo, Guole Liu, Yuanhao Guo, and Ge Yang. Elucidating meta-structures of noisy labels in semantic segmentation by deep neural networks. arXiv preprint arXiv:2205.00160, 2022.", + "[22] Yaoru Luo, Guole Liu, Wenjing Li, Yuanhao Guo, and Ge Yang. Deep neural networks learn meta-structures to segment fluorescence microscopy images. arXiv preprint arXiv:2103.11594, 2021.", + "[23] Robert Mendel, Luis Antonio de Souza, David Rauber, Joao Paulo Papa, and Christoph Palm. Semi-supervised segmentation based on error-correcting supervision. In European Conference on Computer Vision, pages 141-157. Springer, 2020.", + "[24] Shaobo Min, Xuejin Chen, Zheng-Jun Zha, Feng Wu, and Yongdong Zhang. A two-stream mutual attention network for semi-supervised biomedical segmentation with noisy labels. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 4578-4585, 2019." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1237", + "bbox": [ + 483, + 924, + 514, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Sudhanshu Mittal, Maxim Tatarchenko, and Thomas Brox. Semi-supervised semantic segmentation with high-and low-level consistency. IEEE transactions on pattern analysis and machine intelligence, 43(4):1369-1379, 2019.", + "[26] Yassine Ouali, Céline Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross-consistency training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12674-12684, 2020.", + "[27] Wataru Shimoda and Keiji Yanai. Self-supervised difference detection for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5208-5217, 2019.", + "[28] Yucheng Shu, Xiao Wu, and Weisheng Li. Lvc-net: Medical image segmentation with noisy label based on local visual cues. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 558-566. Springer, 2019.", + "[29] Nasim Souly, Concetto Spampinato, and Mubarak Shah. Semi supervised semantic segmentation using generative adversarial network. In Proceedings of the IEEE international conference on computer vision, pages 5688-5696, 2017.", + "[30] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017.", + "[31] Luis Garcia Ugarriza, Eli Saber, Sreenath Rao Vantaram, Vincent Amuso, Mark Shaw, and Ranjit Bhaskar. Automatic image segmentation by dynamic region growth and multiresolution merging. IEEE transactions on image processing, 18(10):2275-2288, 2009.", + "[32] Guotai Wang, Xinglong Liu, Chaoping Li, Zhiyong Xu, Jiugen Ruan, Haifeng Zhu, Tao Meng, Kang Li, Ning Huang, and Shaoting Zhang. A noise-robust framework for automatic segmentation of Covid-19 pneumonia lesions from ct images. IEEE Transactions on Medical Imaging, 39(8):2653-2663, 2020.", + "[33] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022.", + "[34] Yude Wang, Jie Zhang, Meina Kan, Shiguang Shan, and Xilin Chen. Self-supervised equivariant attention mechanism for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12275-12284, 2020.", + "[35] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021.", + "[36] Rumeng Yi, Yaping Huang, Qingji Guan, Mengyang Pu, and Runsheng Zhang. Learning from pixel-level label noise: A new perspective for semi-supervised semantic segmentation. IEEE Transactions on Image Processing, 31:623-635, 2021.", + "[37] Dong Zhang, Hanwang Zhang, Jinhui Tang, Xian-Sheng Hua, and Qianru Sun. Causal intervention for weakly-" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "supervised semantic segmentation. Advances in Neural Information Processing Systems, 33:655-666, 2020.", + "[38] Zhaowei Zhu, Zihao Dong, and Yang Liu. Detecting corrupted labels without training a model to predict. In International Conference on Machine Learning, pages 27412-27427. PMLR, 2022.", + "[39] Zhaowei Zhu, Yiwen Song, and Yang Liu. Clusterability as an alternative to anchor points when learning with noisy labels. In International Conference on Machine Learning, pages 12912-12923. PMLR, 2021.", + "[40] Yuliang Zou, Zizhao Zhang, Han Zhang, Chun-Liang Li, Xiao Bian, Jia-Bin Huang, and Tomas Pfister. Pseudoseg: Designing pseudo labels for semantic segmentation. arXiv preprint arXiv:2010.09713, 2020." + ], + "bbox": [ + 501, + 92, + 893, + 290 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1238", + "bbox": [ + 483, + 924, + 514, + 936 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups/8e5c15a5-f270-46c8-829c-81448a8fde6f_model.json b/2023/Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups/8e5c15a5-f270-46c8-829c-81448a8fde6f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4009a46befd25ec8336ac629d09498287c275f72 --- /dev/null +++ b/2023/Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups/8e5c15a5-f270-46c8-829c-81448a8fde6f_model.json @@ -0,0 +1,2222 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.183, + 0.131, + 0.79, + 0.177 + ], + "angle": 0, + "content": "Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.203, + 0.859, + 0.276 + ], + "angle": 0, + "content": "Peixia Li\\(^{1,2}\\), Pulak Purkait\\(^{2}\\), Thalaiyasingam Ajanthan\\(^{2}\\), Majid Abdolshah\\(^{2}\\), Ravi Garg\\(^{2,4}\\), Hisham Husain\\(^{2}\\), Chenchen Xu\\(^{2,3}\\), Stephen Gould\\(^{3}\\), Wanli Ouyang\\(^{1}\\), Anton van den Hengel\\(^{2,4}\\), The University of Sydney,\\(^{2}\\) Amazon, Australian National University,\\(^{4}\\) The University of Adelaide" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.473, + 0.616 + ], + "angle": 0, + "content": "Semi-supervised semantic segmentation methods use a small amount of clean pixel-level annotations to guide the interpretation of a larger quantity of unlabelled image data. The challenges of providing pixel-accurate annotations at scale mean that the labels are typically noisy, and this contaminates the final results. In this work, we propose an approach that is robust to label noise in the annotated data. The method uses two diverse learning groups with different network architectures to effectively handle both label noise and unlabelled images. Each learning group consists of a teacher network, a student network and a novel filter module. The filter module of each learning group utilizes pixel-level features from the teacher network to detect incorrectly labelled pixels. To reduce confirmation bias, we employ the labels cleaned by the filter module from one learning group to train the other learning group. Experimental results on two different benchmarks and settings demonstrate the superiority of our method over state-of-the-art approaches." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.633, + 0.21, + 0.648 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Semantic segmentation is a critical step in many computer vision tasks, and one that, like so many, is now predominantly achieved through deep learning [5, 35]. A key factor in the success of these methods has been access to high-quality annotated datasets. Pixel-accurate annotation of segmentation data is challenging, and becomes impractical as datasets grow. Unlabelled data is abundant and thus semi-supervised semantic segmentation has received increasing attention [15, 20, 8]. The predominant semisupervised approach employs a teacher-student framework to allocate pseudo labels to unlabelled data [11, 23, 36, 40]. These methods rely on the availability of a small but perfectly annotated dataset, however. This is a challenge in many practical situations where annotations are collected via crowd-sourcing (e.g., Sagemaker MTurk), or automated labelling techniques [31, 3], and are thus inevitably noisy." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.31, + 0.895, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.457, + 0.895, + 0.563 + ], + "angle": 0, + "content": "Figure 1: Existing semi-supervised semantic segmentation models are vulnerable to the inevitable label noise in the pixel-level annotations. The proposed method performs significantly better than the current state-of-the-art [20] in the presence label noise. We apply three types of noise (labelled PL, RDE and SCP in Sec. 5.2) to achieve a \\(9\\%\\) error rate in the labelled data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.894, + 0.674 + ], + "angle": 0, + "content": "We consider here the practically important case whereby semi-supervised semantic segmentation is to be applied despite the fact that the provided pixel labels contain noise. We propose an approach that is robust to this noise to the extent that it outperforms the baselines by a significant margin (see Fig. 1), despite being relatively simple to implement and incurring no extra computation cost at test time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.795 + ], + "angle": 0, + "content": "It has been shown recently [19] that supervised segmentation models trained on noisy pixel-level labels first fit the clean labels during an \"early-learning\" phase before eventually memorizing the noisy labels. The above learning dynamics are exploited in [19] for weakly supervised segmentation by adding a multi-scale consistency loss. However, no noise detection module is used in [19] and the effect of unsupervised consistency loss on the label noise is weak." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Recently there has been a surge of interest in data-centric methods [38, 39] that utilize the \\( k \\)-NN of the feature representation of the data-points to detect annotation noise in classification datasets. A straightforward extension of this approach to semantic segmentation would lead to computing the \\( k \\)-NN of the feature representation at pixel-level over all images, which would be infeasible. In this work, we" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "1229" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.152 + ], + "angle": 0, + "content": "adapt [39] to semi-supervised semantic segmentation by introducing an explicit label noise detection module that predicts label noise at the pixel level without significant additional memory or computation requirements." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.468, + 0.348 + ], + "angle": 0, + "content": "Pseudo labelling approaches [11, 23, 36, 2] for semi-supervised segmentation often use a single model to generate pseudo labels on the unlabelled data as well as train on them. This leads to confirmation bias as the model can overfit to its prediction errors. The same bias is apparent even when a single model is employed to filter label noise and train on those cleaned labels [19]. To mitigate this effect, we introduce two diverse learning groups with different network architectures that encourage mutual knowledge distillation and effective noise filtering. These learning groups also exploit complementary learning paths that boost the competency of each group. Our contributions are thus summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.354, + 0.468, + 0.384 + ], + "angle": 0, + "content": "- We propose an approach to semi-supervised semantic segmentation that is robust to label noise." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.387, + 0.468, + 0.431 + ], + "angle": 0, + "content": "- We present a general architecture that maintains two diverse learning groups to overcome confirmation bias in label assignment problems." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.434, + 0.468, + 0.479 + ], + "angle": 0, + "content": "- We introduce a filter module that extends [39] for semantic segmentation while demonstrating a good trade-off between accuracy and computation cost." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.483, + 0.468, + 0.542 + ], + "angle": 0, + "content": "- We conduct extensive experiments showing that our method significantly outperforms semi/weakly-supervised semantic segmentation baselines regardless of the presence or the type of label noise." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.354, + 0.468, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.559, + 0.218, + 0.573 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Semi-supervised Semantic Segmentation. Pixel-level annotation for semantic segmentation is time-consuming, costly, and error prone. Semi-supervised learning addresses the first two elements of this problem using unlabeled data. These approaches can be broadly categorized into self-training (a.k.a self-learning) based methods [11, 23, 10, 25, 29] and consistency regularization-based methods [40, 26, 6, 14, 20]. Self-training-based methods use models trained on labelled examples to predict pseudo-segmentation masks over the unlabeled examples, thus iteratively expanding the training set. The effectiveness of these methods depends on accurately identifying high-confidence pseudo labels via crafted auxiliary networks [11, 23, 36] or GANs [10, 25, 29]. Consistency regularization methods enforce consistency among predictions under varying perturbations, where the perturbations can be applied to the input image (French [8], PseudoSeg [40], CAC [16]), the intermediate feature representations (CCT [26]) or the network itself (CPS [6], GCT [14]). Most effort in this vein aims to identify perturbations that are challenging enough to support generalization without" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "changing the underlying true labels. Particularly, recent work [20] on Mean Teacher [30] demonstrates superior performance and flexibility in the choice of perturbations. Our work builds upon [20] and introduces a general framework with two learning groups and a noise filter module to effectively learn from noisy labelled data and unlabelled data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.194, + 0.892, + 0.435 + ], + "angle": 0, + "content": "Semantic Segmentation with Label Noise. Segmentation is a more challenging task than classification, and the labelled data is more prone to label noise. To achieve robustness some methods explicitly model the human annotation errors [12], or add a regularization term [28, 32]. Interestingly, [22] shows how high-level structure information (about microscopy images) learned by the model can be used as additional signal to remedy wrong annotations. In [24], a two stream mutual-attention network is devised to smooth out the influence of noisy labels during backpropagation. ADELE [19] exploits the learning dynamics demonstrated by noisy examples and proposes an early-learning strategy accordingly. These supervised segmentation methods assume all of the data is labellebd and that the majority of the labels are correct. They thus focus on alleviating the influence a small set of incorrect labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.436, + 0.892, + 0.572 + ], + "angle": 0, + "content": "The approach most similar to that proposed here is that of DivideMix [18] whereby two divergent networks are applied to separate data partitions. The method there is specifically designed for classification, however, as it is based on an image-level Gaussian Mixture Model. We propose here instead a pixel-level filter module appropriate for semantic segmentation. Additionally, in contrast to [18], we employ multiple learning groups with varying architectures to encourage diversity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.579, + 0.637, + 0.593 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.65, + 0.62 + ], + "angle": 0, + "content": "3.1. Problem Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.762 + ], + "angle": 0, + "content": "We consider semi-supervised segmentation in the case where we have a small set of potentially noisy labelled examples and a large set of unlabelled examples. Let \\(\\mathcal{D}_L = \\{(x_i, y_i)\\}_{i=1}^{N_L}\\) be the labelled set where \\(x_i \\in \\mathcal{X} := [0,1]^{h \\times w \\times 3}\\) and \\(y_i \\in \\mathcal{Y} := \\mathcal{C}^{h \\times w}\\) denote an image and the corresponding pixel-wise segmentation mask respectively. Here, \\(\\mathcal{C} = \\{1,2,\\dots,\\ell\\}\\) denotes the set of labels. Similarly, let \\(\\mathcal{D}_U = \\{x_i\\}_{i=1}^{N_U}\\) denote the unlabelled set of images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.807 + ], + "angle": 0, + "content": "In contrast to the standard semi-supervised segmentation problem, the segmentation masks \\(\\{\\pmb{y}_i\\}\\) may contain erroneous labels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.706, + 0.832 + ], + "angle": 0, + "content": "3.2. Noisy Label Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Here we briefly review the noisy label detection approach SimiFeat [38] developed for robust image classification. Our work extends this framework to the segmentation problem by pixel-level reasoning over noisy labels in" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.925, + 0.516, + 0.936 + ], + "angle": 0, + "content": "1230" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.09, + 0.868, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.304, + 0.892, + 0.396 + ], + "angle": 0, + "content": "Figure 2: The pipelines of a) the baseline model [20] and b) the proposed approach. In contrast to the baseline that uses two teachers and a single student network, we propose a framework with two diverse Learning Groups (LGs) \\((t^1, s^1)\\) and \\((t^2, s^2)\\) where each group contains a teacher and a student network. The two learning groups propagate complementary information and reduce the confirmation bias. Here, \\(t\\) denotes a teacher model, \\(s\\) is a student model, solid lines show the forward flow of labelled data, and dotted lines show the flow of unlabelled data. \\(L_{ce}\\) denotes the cross-entropy loss and \\(L_{con}\\) is the confidence-weighted cross-entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.469, + 0.485 + ], + "angle": 0, + "content": "Sec. 4.2. SimiFeat estimates the label noise proportion for each class and then compute a per-class \"noisiness\" score for all the images in order to rank them and eliminate top \\( R_{c} \\) percentile of images, where \\( R_{c} \\) is the noise percentage for class \\( c \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.469, + 0.577 + ], + "angle": 0, + "content": "Let \\(\\tilde{\\mathcal{D}} = \\left\\{(\\pmb{x}_i,\\tilde{y}_i)\\right\\}\\) be a dataset of images \\(\\pmb{x}_i\\in \\mathcal{X}\\) with noisy labels \\(\\tilde{y}_i\\in \\mathcal{C}\\), and \\(g(\\cdot):\\mathcal{X}\\to \\mathbb{R}^d\\) be a feature extractor. The noise proportion is computed from the prior probability of clean labels and the noise transition probability, which are estimated via High Order Consensus (HOC) (refer to [38] for more details) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.586, + 0.469, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\mathbb {P} (Y), \\mathbb {P} (\\tilde {Y} | Y) \\leftarrow \\operatorname {H O C} \\left(\\left\\{\\left(g (\\boldsymbol {x}), \\tilde {y}\\right) \\mid (\\boldsymbol {x}, \\tilde {y}) \\in \\tilde {\\mathcal {D}} \\right\\}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.469, + 0.66 + ], + "angle": 0, + "content": "where \\( Y \\) and \\( \\tilde{Y} \\) denote the random variables corresponding to clean and noisy labels respectively. The noise proportion is then calculated for a class \\( c \\in \\mathcal{C} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.67, + 0.469, + 0.688 + ], + "angle": 0, + "content": "\\[\nR _ {c} = 1 - \\mathbb {P} (Y = c | \\tilde {Y} = c). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.697, + 0.357, + 0.714 + ], + "angle": 0, + "content": "Note, \\(\\mathbb{P}(Y|\\hat{Y})\\) is obtained via Bayes' rule:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.722, + 0.469, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\mathbb {P} (Y | \\tilde {Y}) = \\mathbb {P} (\\tilde {Y} | Y) \\mathbb {P} (Y) / \\mathbb {P} (\\tilde {Y}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.469, + 0.78 + ], + "angle": 0, + "content": "where \\(\\mathbb{P}(\\tilde{Y})\\) is computed from the observed labels and \\(\\mathbb{P}(Y),\\mathbb{P}(\\tilde{Y}|Y)\\) are computed using HOC (Eq. (1))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To compute the noisiness score, a \\(k\\)-NN based soft label is employed in the feature space via nearest-neighbour voting. Then, the score for an image is determined as the negative cosine similarity between the soft label of that image and the one-hot encoding of the class label [38]. Finally, the images are ranked based on the score and the noise proportion from Eq. (2) is utilized to determine the percentile of images to be declared as noisy and consequently removed." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.41, + 0.647, + 0.427 + ], + "angle": 0, + "content": "4. Our Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.464, + 0.892, + 0.676 + ], + "angle": 0, + "content": "We introduce a general pipeline with two Learning Groups (LGs), as shown in Fig. 2. Each LG consists of a teacher-student framework and a noisy label detection module (Filter Modules). The teacher network is used to produce pseudo labels for the unlabelled data similar to the semi-supervised learning pipeline in [20]. To stabilize training, we first train both LGs independently in a fully-supervised manner with the labelled dataset \\( \\mathcal{D}_L \\) for a few epochs. After this warm-up, at each training epoch, we activate the filter module that detects the pixels with noisy labels in the labelled set \\( \\mathcal{D}_L \\) and move them to the unlabeled set \\( \\mathcal{D}_U \\). Then, we train the pipeline in the semi-supervised manner. The noise detection and model training are repeated until the end of the training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our filter module uses the feature representation of the teacher model from one LG to partition the labelled data into clean and noisy sets. The filter module then ignores the labels of the noisy set and moves the corresponding elements into the unlabelled set to obtain a revised labelled and unlabelled dataset pair. The updated dataset pair generated by one LG is consumed by the other. As training progresses each LG receives a different set of dataset pair due to the decoupled noise filtering, which enables complementary information to propagate in the two learning paths - reducing confirmation bias. This, along with the different network architectures for the two LGs, enhances the performance of our framework in the presence of noise." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.925, + 0.513, + 0.936 + ], + "angle": 0, + "content": "1231" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.473, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.216, + 0.473, + 0.309 + ], + "angle": 0, + "content": "Figure 3: The noise detection results from the two LGs. The images from left to right are: input, clean ground-truth followed by noise detection results from the two LGs when we use the same architecture (two R50s) and different architectures (R50 and PVTv2-B2 [33]). We get more diverse noise detection results when using different architectures." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.337, + 0.241, + 0.354 + ], + "angle": 0, + "content": "4.1. Learning Group" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.362, + 0.47, + 0.422 + ], + "angle": 0, + "content": "The architecture of the LG used in our pipeline is inspired by the recent teacher-student framework [20] as depicted in Fig. 2 (a). We refer the reader to [20] for more details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.473, + 0.504 + ], + "angle": 0, + "content": "At a training epoch \\(k\\), the first LG receives the revised labelled set \\(\\mathcal{D}_{L,k}^{2}\\) and unlabelled set \\(\\mathcal{D}_{U,k}^{2}\\) after filtering the noisy labels predicted by the second LG (details of \\(\\mathcal{D}_{L,k}^{2}\\) and \\(\\mathcal{D}_{U,k}^{2}\\) can be found in Sec. 4.2). Then the student network of the first LG minimizes the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.516, + 0.476, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(\\mathcal {D} _ {L, k} ^ {2}, \\mathcal {D} _ {U, k} ^ {2}, \\boldsymbol {\\theta} _ {s ^ {1}}\\right) := L _ {c e} \\left(\\mathcal {D} _ {L, k} ^ {2}, \\boldsymbol {\\theta} _ {s ^ {1}}\\right) + \\alpha L _ {c o n} \\left(\\mathcal {D} _ {U, k} ^ {2}, \\boldsymbol {\\theta} _ {s ^ {1}}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.47, + 0.717 + ], + "angle": 0, + "content": "Here, \\( L_{ce} \\) denotes the cross-entropy loss on the labelled data, \\( L_{con} \\) is a confidence-weighted cross-entropy loss that enforces consistency between the teacher and student model predictions on the unlabelled data, \\( \\theta_{s^1} \\) denotes the parameters of the student network in the first LG, and \\( \\alpha > 0 \\) provides a relative weight between the two losses. Analogously, the parameters of the second student network \\( \\theta_{s^2} \\) are optimized utilizing the revised labelled set \\( \\mathcal{D}_{L,k}^{1} \\) and the unlabelled dataset \\( \\mathcal{D}_{U,k}^{1} \\) predicted by the first LG. The teacher network is updated by the moving average of the student network parameters of the same LG." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Note that, at each training epoch, the set of labelled data varies depending on the filter module output. Specifically, the detected pixels with noisy labels will be removed from the label set, which leads to modifying the segmentation mask of the labelled examples with an \"ignore-pixel\" label that enables the loss Eq. (4) to be defined in terms of \\( L_{con} \\) instead of \\( L_{ce} \\) at those pixels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.901 + ], + "angle": 0, + "content": "To encourage diversity of the LGs, we utilize R50 [9] for the first LG and PVTv2-B2 [33] for the second. In Fig. 3, we show that the noise detection results from two R50 (the third and fourth columns) have less diversity than those from different network architectures (the last two" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.089, + 0.895, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.24, + 0.895, + 0.363 + ], + "angle": 0, + "content": "Figure 4: The architecture of our filter module. First, we generate a representative dataset \\(\\bar{\\mathcal{D}}\\) and calculate the noise proportion based on HOC [39]. Then, we get the adaptive threshold based on the noise proportion and example score. According to the adaptive threshold and example score, the original dataset \\(\\mathcal{D}_L\\) is divided into a clean set \\(\\mathcal{D}_L^1\\) and a noisy set. The noisy set is added to the unlabelled dataset \\(\\mathcal{D}_U\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.389, + 0.892, + 0.422 + ], + "angle": 0, + "content": "columns).1 We also show that two different architectures bring better performance in our experiments (Fig. 7)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.432, + 0.644, + 0.447 + ], + "angle": 0, + "content": "4.2.Filter Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.892, + 0.532 + ], + "angle": 0, + "content": "The task of the filter module is to detect incorrectly labelled pixels efficiently. To this end, rather than relying on label correction approaches [19], we design a filter module that accurately detects pixel-level noise scalably. The filter module is shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.893, + 0.652 + ], + "angle": 0, + "content": "Note that, according to Eq. (1), the HOC approach requires a dataset of feature encodings and the corresponding labels. Since semantic segmentation is a classification problem at the pixel-level, the natural extension is to treat each pixel as a datapoint and directly apply HOC. Unfortunately, this becomes intractable as the computational complexity is quadratic in number of pixels, labels, and number of images in the dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.894, + 0.789 + ], + "angle": 0, + "content": "To mitigate this, we use a small subset of pixels per class to generate a representative set \\(\\bar{\\mathcal{D}} = \\{(\\pmb{x}_i,j,\\tilde{y}_{i,j})\\}\\), which comprises of a fixed number of pixels per image subsampled from \\(\\mathcal{D}_L\\). Here, \\(\\pmb{x}_i\\in \\mathcal{X}\\) is an image, \\(j\\in \\mathcal{I}(\\pmb{x}_i)\\) is the index of a pixel in image \\(\\pmb{x}_i\\), and \\(\\tilde{y}_{i,j}\\in \\mathcal{C}\\) is the label of the \\(j\\)-th pixel. In order to generate such a set, we found that random pixel sampling is more effective than elaborate methods such as \\(k\\)-means clustering, and is more computationally efficient also." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.79, + 0.894, + 0.851 + ], + "angle": 0, + "content": "Let \\( g_{j}^{1}(\\cdot):\\mathcal{X}\\to \\mathbb{R}^{d} \\) denotes the backbone feature extractor, i.e., the penultimate layer output for the first LG at pixel \\( j \\) (where \\( j\\in \\mathcal{I}(\\pmb {x}_i)\\coloneqq \\{1,\\dots ,w\\times h\\} \\) if images are resized to a fixed size). We construct the embedded repre" + }, + { + "type": "page_footnote", + "bbox": [ + 0.499, + 0.863, + 0.894, + 0.901 + ], + "angle": 0, + "content": "1It should be noted that the two networks with R50 architecture are trained with different initialization independently on the labelled data for a warm start for a fair comparison." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "1232" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.09, + 0.47, + 0.168 + ], + "angle": 0, + "content": "sentative set \\(\\bar{\\mathcal{D}}_G^1 \\coloneqq \\left\\{(g_j^1(\\boldsymbol{x}_i), \\tilde{y}_{i,j}) \\mid (\\boldsymbol{x}_i, j, \\tilde{y}_{i,j}) \\in \\bar{\\mathcal{D}}\\right\\}\\) and invoke HOC (Eq. (1)) to estimate the noise proportion \\(\\{R_c^1\\}_{c \\in \\mathcal{C}}\\) at pixel-level (Eq. (2)). Analogously, the noise proportion \\(\\{R_c^2\\}_{c \\in \\mathcal{C}}\\) for the other LG can be estimated by constructing \\(\\bar{\\mathcal{D}}_G^2\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.167, + 0.471, + 0.288 + ], + "angle": 0, + "content": "Let \\( h_j^1(\\cdot) : \\mathcal{X} \\to [0,1]^{\\ell} \\) and \\( h_j^2(\\cdot) : \\mathcal{X} \\to [0,1]^{\\ell} \\) be the teacher model mappings to softmax scores in the first and second LG respectively. We compute a noisiness score based on the teacher model predictions, as opposed to using \\( k \\)-NN in SimiFeat [38]. This eliminates the computation cost of running \\( k \\)-NN for every pixel over all images in \\( \\mathcal{D}_L \\). We define the score for the first LG at pixel \\( j \\) in image \\( x_i \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.296, + 0.47, + 0.318 + ], + "angle": 0, + "content": "\\[\nS ^ {1} \\left(\\boldsymbol {x} _ {i}, j\\right) = \\operatorname {C E} \\left(h _ {j} ^ {1} \\left(\\boldsymbol {x} _ {i}\\right), \\boldsymbol {e} _ {i, j}\\right) + \\lambda \\operatorname {K L} \\left(h _ {j} ^ {1} \\left(\\boldsymbol {x} _ {i}\\right) \\| h _ {j} ^ {2} \\left(\\boldsymbol {x} _ {i}\\right)\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.328, + 0.47, + 0.554 + ], + "angle": 0, + "content": "where CE and KL denote the cross-entropy and Kullback-Leibler (KL) divergence respectively, \\( e_{i,j} \\in \\{0,1\\}^{\\ell} \\) is the one-hot encoded noisy label at pixel \\( j \\) in image \\( x_i \\), and \\( \\lambda > 0 \\). CE measures the alignment between label noise and model predictions, and KL measures the inconsistency between the two teacher model predictions. Large CE at a pixel implies a high misalignment between the predicted class label and annotated label and hence high model confidence of the label to be noisy unless the current model overfits at that pixel. On the other hand, if one of the models starts to overfit the noisy annotations, then the KL measure will be high. If any of them is high, such pixel corresponds to a high score, and it is marked as noisy. Experimentally, we verify that the above choice of score works better than the other alternatives (shown in Fig. 8(b))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.556, + 0.47, + 0.601 + ], + "angle": 0, + "content": "Once the noisiness score and the noise proportion have been calculated, we get the adaptive noise threshold as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.606, + 0.47, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\beta^ {1} = \\frac {\\sigma \\ell S _ {a v g 1} ^ {1}}{\\sum_ {c \\in \\mathcal {C}} R _ {c} ^ {1}}, \\text {w h e r e} S _ {a v g 1} ^ {1} = n ^ {- 1} \\sum_ {\\substack {\\left(\\boldsymbol {x} _ {i},\\cdot\\right) \\in \\mathcal {D} _ {L}, \\\\ j \\in \\mathcal {I} \\left(\\boldsymbol {x} _ {i}\\right)}} S ^ {1} \\left(\\boldsymbol {x} _ {i}, j\\right), \\tag{6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.666, + 0.472, + 0.817 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is a hyperparameter and \\(n\\) denotes the number of all the pixels in \\(\\mathcal{D}_L\\). Experimentally we found that the noise proportion \\(R_{c}\\) predicted by HOC is inaccurate for some of the classes and a single threshold \\(\\beta^1\\), computed from the average as in Eq. (6), works better than class-wise thresholds. A higher \\(\\beta\\) means filtering out less noisy annotations. We utilize the adaptive threshold \\(\\beta^1\\) to determine the pixels to be removed and marked in the segmentation masks to reflect the pixels with noisy labels. Let us denote the set of pixels that are detected as label noise by the first LG as:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.823, + 0.482, + 0.855 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {k} ^ {1} := \\left\\{\\left(\\boldsymbol {x} _ {i}, j\\right) \\mid S ^ {1} \\left(\\boldsymbol {x} _ {i}, j\\right) > \\beta^ {1}, \\left(\\boldsymbol {x} _ {i}, \\cdot\\right) \\in \\mathcal {D} _ {L}, j \\in \\mathcal {I} \\left(\\boldsymbol {x} _ {i}\\right) \\right\\} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The updated datasets at epoch \\( k \\) are created by removing the pixels with noise labels from the labelled data and placing the images with noisy pixels in the unlabelled set, i.e., the" + }, + { + "type": "algorithm", + "bbox": [ + 0.501, + 0.091, + 0.892, + 0.429 + ], + "angle": 0, + "content": "Algorithm 1 Our training pipeline \nInput: \\(\\mathcal{D}_L, \\mathcal{D}_U\\), total epochs \\(E\\), warm-up epochs \\(N\\) \nOutput: Optimized \\(\\{\\theta_{s^1}, \\theta_{t^1}, \\theta_{s^2}, \\theta_{t^2}\\}\\) \n1: Initialize parameters \\(\\{\\theta_{s^1}, \\theta_{t^1}, \\theta_{s^2}, \\theta_{t^2}\\}\\), \n2: for \\(k \\in \\{0, 1, \\dots, E\\}\\) do \n3: if \\(k < N\\) then \n4: \\(\\mathcal{D}_{L,k}^{1}, \\mathcal{D}_{L,k}^{2} \\coloneqq \\mathcal{D}_{L}\\) and \\(\\mathcal{D}_{U,k}^{1}, \\mathcal{D}_{U,k}^{2} \\coloneqq \\emptyset\\) \n5: else \n6: \\(\\{(x_i, j, \\tilde{y}_{i,j})\\} \\gets\\) Sample from \\(\\mathcal{D}_L\\) \n7: for \\(l \\in \\{1, 2\\}\\) do \n8: \\(\\{R_c^l\\} \\gets\\) HOC \\(\\left(\\left(g_j^l(x_i), \\tilde{y}_{i,j}\\right)\\right)\\) \n9: Scores \\(\\{S^l\\} \\gets\\) Using Eq. (5) \n10: \\(\\beta^l \\gets\\) Adaptive \\(\\left(\\{S^l\\}\\right)\\) [Eq. (6)] \n11: \\(\\mathcal{M}_k^l \\gets\\) Using Eq. (7) \n12: \\(\\mathcal{D}_{L,k}^l \\gets \\mathcal{D}_L \\backslash \\mathcal{M}_k^l\\) \n13: \\(\\mathcal{D}_{U,k}^l \\gets \\mathcal{D}_U \\bigcup \\{x_i | (x_i, \\cdot) \\in \\mathcal{M}_k^l\\}\\) \n14: repeat \n15: \\(\\theta_{s^1} \\gets\\) Backprop. \\(\\mathcal{L}(\\mathcal{D}_{L,k}^2, \\mathcal{D}_{U,k}^2, \\theta_{s^1})\\) [Eq. (4)] \n16: \\(\\theta_{s^2} \\gets\\) Backprop. \\(\\mathcal{L}(\\mathcal{D}_{L,k}^1, \\mathcal{D}_{U,k}^1, \\theta_{s^2})\\) [Eq. (4)] \n17: until \\(M\\) times \n18: Update \\(\\theta_{t^1}, \\theta_{t^2} \\gets\\) Moving Average of \\(\\theta_{s^1}, \\theta_{s^2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.452, + 0.892, + 0.578 + ], + "angle": 0, + "content": "clean labelled set \\(\\mathcal{D}_{L,k}^{1} \\coloneqq \\mathcal{D}_{L} \\setminus \\mathcal{M}_{k}^{12}\\) and the unlabelled dataset \\(\\mathcal{D}_{U,k}^{1} \\coloneqq \\mathcal{D}_{U} \\bigcup \\{\\pmb{x}_{i} \\mid (\\pmb{x}_{i},\\cdot) \\in \\mathcal{M}_{k}^{1}\\}\\). The student network of the second LG is then trained on these datasets according to Eq. (4). Analogously, the score for the second LG can be computed by switching \\(h_{j}^{1}\\) and \\(h_{j}^{2}\\) in Eq. (5), and the datasets \\(\\mathcal{D}_{L,k}^{2}, \\mathcal{D}_{U,k}^{2}\\) are created in the above manner on which the student network of the first LG is trained. Our final algorithm is summarized in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.583, + 0.633, + 0.601 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.892, + 0.7 + ], + "angle": 0, + "content": "Evaluation has been carried out following [19, 20] to allow direct comparison. As in [20], the labelled data is augmented with synthetic label noise. Performance is measured with, and without an unlabelled dataset to illuminate performance differences between the weakly and semi supervised settings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.708, + 0.719, + 0.724 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.73, + 0.892, + 0.868 + ], + "angle": 0, + "content": "We utilize R50 [9] and PVTv2-B2 [33] as the backbones in the two LGs to encourage diversity. Both the teacher and student models are based on DeeplabV3+ [4]. For inference, we use the average of the predictions from the two teacher networks as in [20]. The \\(\\lambda\\) in Eq. (5) is set to 0.1 to balance the contribution from the CE and KL terms in our loss. We randomly sample 100 pixels per class in each image to get the representative dataset \\(\\bar{D}\\). The hyperparameter \\(\\sigma\\) in Eq. (6) is set to be 10 in our experiments. Other" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.902 + ], + "angle": 0, + "content": "2With a slight abuse of notation, this operation indicates marking all pixels in \\(\\mathcal{M}_k^1\\) with \"ignore-pixel\" label in \\(\\mathcal{D}_L\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "1233" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.152 + ], + "angle": 0, + "content": "Table 1: Comparison against SotA approaches on Pascal VOC 2012 in a semi-supervised semantic segmentation setting. All baselines are based on the DeeplabV3+ architecture. The * indicates results reported by [6]. When we replace \\(50\\%\\) of the ground truth of images with 'PL', the proportion of noisy pixels is \\(9\\%\\). We keep the same pixel level noise proportion while introducing noise using 'RDE' and 'SCP' for parity." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.163, + 0.899, + 0.446 + ], + "angle": 0, + "content": "
MethodsYearNoisy pixelNoise typeLabelled Data Ratio
1/16 (662)1/8(1323)1/4 (2646)1/2 (5291)
MT* [30]20170%None66.7070.7873.2275.41
French [8]20190%None68.9070.7072.4674.49
CCT* [26]20200%None65.2270.8773.4374.75
GCT* [14]20200%None64.0570.4773.4575.20
ECS* [23]20200%None-67.3870.7072.89
CPS* [6]20210%None71.9873.6774.9076.15
CAC* [16]20210%None70.1072.4074.00-
PS-MT [20]20220%None72.8375.7076.4377.88
Ours-0%None77.75 (+4.92)79.31 (+3.61)79.14 (+2.71)79.54 (+1.66)
PS-MT [20]-9%PL61.9065.1465.5366.78
Ours-9%PL71.60 (+9.70)73.95 (+8.81)74.53 (+9.00)74.44 (+7.66)
PS-MT [20]-9%RDE58.8762.1863.8163.46
Ours-9%RDE66.82 (+7.95)69.95 (+7.77)73.82 (+10.01)74.47 (+11.01)
PS-MT [20]-9%SCP52.1457.1956.8153.69
Ours-9%SCP67.80 (+15.66)70.83 (+13.64)69.38 (+12.57)75.95 (+22.26)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.471, + 0.557 + ], + "angle": 0, + "content": "training and inference settings are the same as those of the baseline method [20]. In all of our experiments, the last epoch checkpoint model is used unless otherwise specified. Further details of the implementation can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.567, + 0.424, + 0.584 + ], + "angle": 0, + "content": "5.2. Semi-supervised Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.47, + 0.667 + ], + "angle": 0, + "content": "Pascal VOC 2012 [7] is employed for evaluating the performance of our model in a semi-supervised semantic segmentation setting. This dataset contains more than \\(13\\mathrm{k}\\) images with 21 different classes. To introduce realistic noise in \\(\\mathcal{D}_L\\), the following three approaches are considered:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.677, + 0.47, + 0.723 + ], + "angle": 0, + "content": "- Pseudo Labels (PL). Following [19], a portion of clean labels are replaced by pseudo-labels from a weakly-supervised segmentation method [34]." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.726, + 0.47, + 0.771 + ], + "angle": 0, + "content": "- Random Dilation and Erosion (RDE). Similar to [19, 21], we apply random degrees of dilation and erosion to the ground-truth masks." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.774, + 0.471, + 0.866 + ], + "angle": 0, + "content": "- Similar Class Perturbation (SCP). We introduce another type of noise by randomly replacing ground-truth label of a segmentation mask with a different class that resembles the texture of original clean class. This label perturbation simulates a more natural scenario pertaining to annotators confusing semantics." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.677, + 0.471, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.479, + 0.871, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.633, + 0.892, + 0.71 + ], + "angle": 0, + "content": "Figure 5: A performance comparison under different noise proportions achieved by replacing (from \\(0\\%\\) to \\(90\\%\\)) clean segmentation masks with the PL from [34] at image-level. The green font shows the corresponding noisy pixel proportions. All experiments are conducted with \\(1/8\\) labelled data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Details about the implementation and visualization of noise generation can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In the context of semi-supervised semantic segmentation, we have compared our approach with the following studies: MT [30], French [8], CCT [26], GCT [14], ECS [23], CPS [6], CAC [16], and PS-MT [20]. Following the previous studies [19, 20], we report the mean Intersection-over-Union (mIoU) on the validation set. Table 1 shows that our approach ('Ours') outperforms all baselines trained with clean labelled data. In the presence of noise, we compare the best performing semi-supervised segmentation method" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.471, + 0.901 + ], + "angle": 0, + "content": "3Note that, SCP does not change the shape of the clean segmentation masks, but only labels may be switched." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "1234" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.152 + ], + "angle": 0, + "content": "Table 2: The mIoU of the weakly-supervised semantic segmentation baselines and ours on the test set of SegTHOR. Best Val and Last Epoch correspond to the checkpoint that performs best on the validation set and the last checkpoint." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.162, + 0.447, + 0.234 + ], + "angle": 0, + "content": "
ModelBest ValLast Epoch
Base-ADELE [19]62.659.1
ADELE [19]71.170.8
Ours73.273.5
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.251, + 0.47, + 0.372 + ], + "angle": 0, + "content": "with no noise, i.e., PS-MT [20], against our method with varying noise types (see Table 1) and noise proportions (see Fig. 5). The results demonstrate significant improvements in mIoU in all cases, e.g., \\(+22.26\\%\\) with half of the labelled data on the more practical SCP noise. Overall, the superiority of our method is consistent across all labelled data to unlabelled data ratios, regardless of the presence of noise or the type of applied noise." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.381, + 0.444, + 0.398 + ], + "angle": 0, + "content": "5.3. Weakly-supervised Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.47, + 0.465 + ], + "angle": 0, + "content": "We evaluate our method in a weakly-supervised semantic segmentation setting on the following datasets: (i) SegTHOR [17], a medical dataset, and (ii) pseudo labelled Pascal VOC 2012." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.471, + 0.469, + 0.637 + ], + "angle": 0, + "content": "Medical image segmentation on SegTHOR. Following [19], we evaluate our approach on the SegTHOR dataset, which consists of 3D CT scans. Each pixel of these images is assigned to one of these five classes: 'esophagus', 'heart', 'trachea', 'aorta', or 'background'. The slices from the images are randomly split into a training set of 3638 slices, a validation set of 570 slices, and a test set of 580 slices [19, 17]. To introduce noise, we apply 'RDE' to the ground-truth segmentation labels, similar to [19]. Note that all annotations in the training set are corrupted by this noise, whereas the validation and test sets remain uncorrupted." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.47, + 0.775 + ], + "angle": 0, + "content": "We compare our approach with ADELE [19], which corrects noisy annotations with model predictions in the early training stage, and its variant, Base-ADELE [19], which is based on a UNet trained with multi-scale inputs. Detailed implementations of these approaches are in the supplementary material. Following the setting in ADELE [19], we evaluate two checkpoints: specifically, the one with the highest mIoU on the validation set and the last checkpoint. In Table 2, our approach outperforms both baselines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Weakly-supervised segmentation on Pascal VOC 2012. We demonstrate the effectiveness of our approach in comparison to the baselines in the weakly-supervised setting on the PASCAL VOC 2012 [7]. Similar to ADELE [19], we use only noisy pseudo labels generated by [34] to train the models and evaluate on the original test set. Table 3 shows that our approach outperforms recent baselines, such as ADELE combined with SEAM [19]. In addition, PS-MT" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.153 + ], + "angle": 0, + "content": "increases the mIoU marginally with PVTv2 compared to R50 with 66.7 vs. 64.0. However, after using our framework, the mIoU is further improved from 66.7 to 71.7, clearly demonstrating the effectiveness of our method." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.166, + 0.892, + 0.212 + ], + "angle": 0, + "content": "Table 3: Comparison with state-of-the-art weakly supervised semantic segmentation methods on the Pascal VOC 2012 dataset using mIoU (\\%) of the validation set." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.223, + 0.877, + 0.386 + ], + "angle": 0, + "content": "
ModelYearmIoU (%)
AffinityNet [1]201861.7
SSDD [27]201964.9
SEAM [34]202064.5
CONTA [37]202066.1
SPML [13]202169.5
ADELE [19] + SEAM202269.3
PS-MT [20] (R50)202264.0
PS-MT [20] (PVTv2-B2)202266.7
Ours-71.7
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.412, + 0.653, + 0.428 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.435, + 0.892, + 0.496 + ], + "angle": 0, + "content": "All experiments in this section are conducted with \\(1/8\\) labelled data ratio and \\(9\\%\\) PL noisy annotations on Pascal VOC. We verify the effectiveness of the two main components in our approach, the diverse LGs and filter module." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.893, + 0.708 + ], + "angle": 0, + "content": "Diversity in the LGs with different architectures. In contrast to previous works [20, 18], we found that using different network architectures in the two LGs provides more complementary information. In Fig. 3, we show the qualitative results of noise detection when we use the same and different architectures in the two LGs. The two LGs provide more diverse noise detection results (lower overlap between predicted noisy labels) when we employ different architectures. In Fig. 6, the predictions from the two LGs with two different architectures \\((\\mathrm{e})\\& (\\mathrm{f})\\) are more complementary than those from the two LGs with the same architecture \\((\\mathrm{(c)})\\& (\\mathrm{(d)})\\). Fig. 7 shows the mIoU of our pipeline when we use different network combinations. The combination of R50 and PVTv2-B2 [33] achieves the best performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.71, + 0.892, + 0.816 + ], + "angle": 0, + "content": "Noise detection with different LGs. To avoid confirmation bias, we utilize the teacher network from one LG to detect noise for the other LG. If we utilize the teacher network from the same LG for noise detection (i.e., replacing \\(\\theta_{s^1}\\) with \\(\\theta_{s^2}\\) in Eq. (4)), the mIoU drops from \\(74.0\\%\\) to \\(72.4\\%\\) on Pascal VOC, demonstrating the importance of mutual noise detection with a different LG in our pipeline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.817, + 0.892, + 0.879 + ], + "angle": 0, + "content": "Significance of the filter module. In Table 4, in addition to the baseline PS-MT and ours, we evaluate the performance of the following models: 1) 'PS-MT + FM' by adding the filter module to PS-MT (using R50)4. 2) 'Ours - FM' by" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.892, + 0.902 + ], + "angle": 0, + "content": "4Note, PS-MT can be regarded as single LG (Fig. 2 (a)) and the mean" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "1235" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.09, + 0.225, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.204, + 0.203, + 0.217 + ], + "angle": 0, + "content": "(a) Input Image" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.091, + 0.354, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.204, + 0.338, + 0.216 + ], + "angle": 0, + "content": "(b) Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.091, + 0.485, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.204, + 0.484, + 0.216 + ], + "angle": 0, + "content": "(c) R50LG1 Prediction" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.091, + 0.616, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.204, + 0.615, + 0.216 + ], + "angle": 0, + "content": "(d) R50LG2 Prediction" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.091, + 0.747, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.204, + 0.746, + 0.216 + ], + "angle": 0, + "content": "(e) R50LG1 Prediction" + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.091, + 0.88, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.204, + 0.88, + 0.216 + ], + "angle": 0, + "content": "(f) PVT\\(^{\\mathrm{LG2}}\\) Prediction" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.225, + 0.894, + 0.256 + ], + "angle": 0, + "content": "Figure 6: Visualizations from the two LGs when we use the same architecture ((c) and (d)) and different architectures ((e) and (f)). The \\(1st\\) and \\(2nd\\) columns show the inputs and the ground truth. R50 and PVT denote R50 and PVTv2-B2 [33]." + }, + { + "type": "table_caption", + "bbox": [ + 0.088, + 0.269, + 0.454, + 0.285 + ], + "angle": 0, + "content": "Table 4: Ablation study of FM and LG on Pascal VOC." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.295, + 0.454, + 0.387 + ], + "angle": 0, + "content": "
Method2 LGsFilter Module (FM)mIoU (%)
PS-MT [20]65.1
PS-MT + FM✓67.6 (+2.5)
Ours - FM✓69.9 (+4.8)
Ours✓✓74.0 (+8.9)
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.404, + 0.264, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.515, + 0.24, + 0.525 + ], + "angle": 0, + "content": "(a) Different Architectures" + }, + { + "type": "image", + "bbox": [ + 0.267, + 0.404, + 0.468, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.514, + 0.422, + 0.524 + ], + "angle": 0, + "content": "(b) Clean and Noisy Data" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.655 + ], + "angle": 0, + "content": "Figure 7: (a) shows the mIoU score of different network combinations on clean data. (b) shows the mIoU score of different combinations on different settings. R, S and P denote R50 [9], Segformer [35] and PVTv2-B2 [33] respectively. -FM_C denotes training our model (without the filter module) with clean data. -FM_N denotes training model with noisy data without using filter module. +FM_N is our complete model trained with noisy data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.468, + 0.824 + ], + "angle": 0, + "content": "removing the filter module from our framework. It can be seen that the filter module helps to improve the final performance for both models. The filter module contributes more significantly (+4.1) to our framework due to the diverse learning groups (i.e., using 2 LGs with different architectures). This superiority of our framework is further evidenced by the better performance of 'Ours - FM' compared to PS-MT variants. In Fig. 7 (b), our Filter Module helps to improve the performance under noisy annotations ('+FM_N' vs. '-FM_N') with different architectures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.828, + 0.468, + 0.873 + ], + "angle": 0, + "content": "Comparison of the filter module with GMM. Previous work [18] utilizes Gaussian Mixture Model (GMM) to detect noise for image classification. We compare our filter" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.269, + 0.709, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.4, + 0.686, + 0.41 + ], + "angle": 0, + "content": "(a) Detected Noise Ratio" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.269, + 0.845, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.4, + 0.882, + 0.411 + ], + "angle": 0, + "content": "(b) F1 Score of Noise Detection" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.422, + 0.892, + 0.527 + ], + "angle": 0, + "content": "Figure 8: (a) shows detected noise proportion of GMM [18] and ours, where \\( x \\)-axis shows the noisy image ratio and \\( y \\)-axis shows the detected noisy pixel ratio. The detected noise proportion of ours is closer to the GT than GMM. (b) shows the F1 score of noise detection with different criterion. CE is cross entropy loss; KL is KL divergence between the predictions from the two LGs. Our \\( \\mathrm{CE} + \\mathrm{KL} \\) performs the best." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.892, + 0.605 + ], + "angle": 0, + "content": "module with GMM (including two components) on the representative dataset. Fig. 8 (a) shows the detected noise proportions. We can see that the detected noise proportion of our filter module is closer to the groundtruth than GMM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.892, + 0.684 + ], + "angle": 0, + "content": "Criterion to compute score. We utilize cross entropy loss and KL divergence \\((\\mathrm{CE} + \\mathrm{KL})\\) to compute score in Eq. (5). Here, we compare this criterion with others, including only cross entropy loss (CE), and only KL divergence (KL). Fig. 8 (b) shows that ours achieves the best F1 score." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.892, + 0.76 + ], + "angle": 0, + "content": "Memory and computation costs. PVTv2-B2 has similar GFLOPs (45.8) with R50 (45.6). Ours only have one more student network than PS-MT, which is not used during inference. Thus, our model needs similar memory and computation costs compared with PS-MT during inference." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.619, + 0.8 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose a learning framework for semi-supervised semantic segmentation involving two diverse learning groups that can handle various labelling errors. Our framework employs a carefully designed filter module to divide the labelled dataset into pixels with and without label noise in a scalable manner. The potentially misla" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.888, + 0.287, + 0.9 + ], + "angle": 0, + "content": "teacher output is used for filter module." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.926, + 0.516, + 0.937 + ], + "angle": 0, + "content": "1236" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "belled pixels filtered by the first learning group are utilised for training the second learning group and vice versa. We show that our proposed framework provides state-of-the-art results for semi-supervised learning, by addressing confirmation bias while training, that can handle various amounts of annotation noise to accommodate practical inaccuracies in annotations." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.209, + 0.175, + 0.225 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.234, + 0.47, + 0.302 + ], + "angle": 0, + "content": "[1] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4981-4990, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.304, + 0.472, + 0.371 + ], + "angle": 0, + "content": "[2] Eric Arazo, Diego Ortego, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.373, + 0.47, + 0.429 + ], + "angle": 0, + "content": "[3] Liyi Chen, Weiwei Wu, Chenchen Fu, Xiao Han, and Yuntao Zhang. Weakly supervised semantic segmentation with boundary exploration. In European Conference on Computer Vision, pages 347-362. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.43, + 0.47, + 0.497 + ], + "angle": 0, + "content": "[4] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.498, + 0.47, + 0.552 + ], + "angle": 0, + "content": "[5] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.554, + 0.47, + 0.621 + ], + "angle": 0, + "content": "[6] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.623, + 0.47, + 0.678 + ], + "angle": 0, + "content": "[7] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.679, + 0.47, + 0.72 + ], + "angle": 0, + "content": "[8] Geoff French, Timo Aila, Samuli Laine, Michal Mackiewicz, and Graham Finlayson. Semi-supervised semantic segmentation needs strong, high-dimensional perturbations. 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.721, + 0.47, + 0.776 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.777, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[10] Wei-Chih Hung, Yi-Hsuan Tsai, Yan-Ting Liou, Yen-Yu Lin, and Ming-Hsuan Yang. Adversarial learning for semi-supervised semantic segmentation. arXiv preprint arXiv:1802.07934, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[11] Mostafa S Ibrahim, Arash Vahdat, Mani Ranjbar, and William G Macready. Semi-supervised semantic image segmentation with self-correcting networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12715-12725, 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.234, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[12] Joseph Jacob, Olga Ciccarelli, Frederik Barkhof, and Daniel C Alexander. Disentangling human error from the ground truth in segmentation of medical images." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[13] Tsung-Wei Ke, Jyh-Jing Hwang, and Stella X Yu. Universal weakly supervised segmentation by pixel-to-segment contrastive learning. arXiv preprint arXiv:2105.00957, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.892, + 0.237 + ], + "angle": 0, + "content": "[14] Zhanghan Ke, Di Qiu, Kaican Li, Qiong Yan, and Rynson WH Lau. Guided collaborative training for pixel-wise semi-supervised learning. In European conference on computer vision, pages 429-445. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.239, + 0.892, + 0.295 + ], + "angle": 0, + "content": "[15] Donghyeon Kwon and Suha Kwak. Semi-supervised semantic segmentation with error localization network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9957-9967, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.298, + 0.892, + 0.367 + ], + "angle": 0, + "content": "[16] Xin Lai, Zhuotao Tian, Li Jiang, Shu Liu, Hengshuang Zhao, Liwei Wang, and Jiaya Jia. Semi-supervised semantic segmentation with directional context-aware consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1205-1214, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.369, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[17] Zoé Lambert, Caroline Petitjean, Bernard Dubray, and Su Kuan. Segthor: segmentation of thoracic organs at risk in ct images. In 2020 Tenth International Conference on Image Processing Theory, Tools and Applications (IPTA), pages 1-6. IEEE, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[18] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.486, + 0.892, + 0.555 + ], + "angle": 0, + "content": "[19] Sheng Liu, Kangning Liu, Weicheng Zhu, Yiqiu Shen, and Carlos Fernandez-Granda. Adaptive early-learning correction for segmentation from noisy annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2606–2616, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.557, + 0.892, + 0.64 + ], + "angle": 0, + "content": "[20] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4258-4267, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.643, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[21] Yaoru Luo, Guole Liu, Yuanhao Guo, and Ge Yang. Elucidating meta-structures of noisy labels in semantic segmentation by deep neural networks. arXiv preprint arXiv:2205.00160, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.756 + ], + "angle": 0, + "content": "[22] Yaoru Luo, Guole Liu, Wenjing Li, Yuanhao Guo, and Ge Yang. Deep neural networks learn meta-structures to segment fluorescence microscopy images. arXiv preprint arXiv:2103.11594, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[23] Robert Mendel, Luis Antonio de Souza, David Rauber, Joao Paulo Papa, and Christoph Palm. Semi-supervised segmentation based on error-correcting supervision. In European Conference on Computer Vision, pages 141-157. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Shaobo Min, Xuejin Chen, Zheng-Jun Zha, Feng Wu, and Yongdong Zhang. A two-stream mutual attention network for semi-supervised biomedical segmentation with noisy labels. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 4578-4585, 2019." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "1237" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[25] Sudhanshu Mittal, Maxim Tatarchenko, and Thomas Brox. Semi-supervised semantic segmentation with high-and low-level consistency. IEEE transactions on pattern analysis and machine intelligence, 43(4):1369-1379, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[26] Yassine Ouali, Céline Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross-consistency training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12674-12684, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.276 + ], + "angle": 0, + "content": "[27] Wataru Shimoda and Keiji Yanai. Self-supervised difference detection for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5208-5217, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.47, + 0.348 + ], + "angle": 0, + "content": "[28] Yucheng Shu, Xiao Wu, and Weisheng Li. Lvc-net: Medical image segmentation with noisy label based on local visual cues. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 558-566. Springer, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.47, + 0.404 + ], + "angle": 0, + "content": "[29] Nasim Souly, Concetto Spampinato, and Mubarak Shah. Semi supervised semantic segmentation using generative adversarial network. In Proceedings of the IEEE international conference on computer vision, pages 5688-5696, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.47, + 0.462 + ], + "angle": 0, + "content": "[30] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.463, + 0.47, + 0.531 + ], + "angle": 0, + "content": "[31] Luis Garcia Ugarriza, Eli Saber, Sreenath Rao Vantaram, Vincent Amuso, Mark Shaw, and Ranjit Bhaskar. Automatic image segmentation by dynamic region growth and multiresolution merging. IEEE transactions on image processing, 18(10):2275-2288, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.47, + 0.615 + ], + "angle": 0, + "content": "[32] Guotai Wang, Xinglong Liu, Chaoping Li, Zhiyong Xu, Jiugen Ruan, Haifeng Zhu, Tao Meng, Kang Li, Ning Huang, and Shaoting Zhang. A noise-robust framework for automatic segmentation of Covid-19 pneumonia lesions from ct images. IEEE Transactions on Medical Imaging, 39(8):2653-2663, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.47, + 0.673 + ], + "angle": 0, + "content": "[33] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.47, + 0.744 + ], + "angle": 0, + "content": "[34] Yude Wang, Jie Zhang, Meina Kan, Shiguang Shan, and Xilin Chen. Self-supervised equivariant attention mechanism for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12275-12284, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[35] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[36] Rumeng Yi, Yaping Huang, Qingji Guan, Mengyang Pu, and Runsheng Zhang. Learning from pixel-level label noise: A new perspective for semi-supervised semantic segmentation. IEEE Transactions on Image Processing, 31:623-635, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[37] Dong Zhang, Hanwang Zhang, Jinhui Tang, Xian-Sheng Hua, and Qianru Sun. Causal intervention for weakly-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "supervised semantic segmentation. Advances in Neural Information Processing Systems, 33:655-666, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[38] Zhaowei Zhu, Zihao Dong, and Yang Liu. Detecting corrupted labels without training a model to predict. In International Conference on Machine Learning, pages 27412-27427. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[39] Zhaowei Zhu, Yiwen Song, and Yang Liu. Clusterability as an alternative to anchor points when learning with noisy labels. In International Conference on Machine Learning, pages 12912-12923. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.894, + 0.291 + ], + "angle": 0, + "content": "[40] Yuliang Zou, Zizhao Zhang, Han Zhang, Chun-Liang Li, Xiao Bian, Jia-Bin Huang, and Tomas Pfister. Pseudoseg: Designing pseudo labels for semantic segmentation. arXiv preprint arXiv:2010.09713, 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "1238" + } + ] +] \ No newline at end of file diff --git a/2023/Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups/8e5c15a5-f270-46c8-829c-81448a8fde6f_origin.pdf b/2023/Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups/8e5c15a5-f270-46c8-829c-81448a8fde6f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cb9dcdaf04b0ca800ef192334aba8a8fed23fda3 --- /dev/null +++ b/2023/Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups/8e5c15a5-f270-46c8-829c-81448a8fde6f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b28ebdd9cad0546453465b83650a8f7a44bd0d48258e41868782bb8c0a30e9df +size 1898744 diff --git a/2023/Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups/full.md b/2023/Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups/full.md new file mode 100644 index 0000000000000000000000000000000000000000..feeb6d6420840a5509914cee6d1727f43a3aca36 --- /dev/null +++ b/2023/Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups/full.md @@ -0,0 +1,327 @@ +# Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups + +Peixia Li $^{1,2}$ , Pulak Purkait $^{2}$ , Thalaiyasingam Ajanthan $^{2}$ , Majid Abdolshah $^{2}$ , Ravi Garg $^{2,4}$ , Hisham Husain $^{2}$ , Chenchen Xu $^{2,3}$ , Stephen Gould $^{3}$ , Wanli Ouyang $^{1}$ , Anton van den Hengel $^{2,4}$ , The University of Sydney, $^{2}$ Amazon, Australian National University, $^{4}$ The University of Adelaide + +# Abstract + +Semi-supervised semantic segmentation methods use a small amount of clean pixel-level annotations to guide the interpretation of a larger quantity of unlabelled image data. The challenges of providing pixel-accurate annotations at scale mean that the labels are typically noisy, and this contaminates the final results. In this work, we propose an approach that is robust to label noise in the annotated data. The method uses two diverse learning groups with different network architectures to effectively handle both label noise and unlabelled images. Each learning group consists of a teacher network, a student network and a novel filter module. The filter module of each learning group utilizes pixel-level features from the teacher network to detect incorrectly labelled pixels. To reduce confirmation bias, we employ the labels cleaned by the filter module from one learning group to train the other learning group. Experimental results on two different benchmarks and settings demonstrate the superiority of our method over state-of-the-art approaches. + +# 1. Introduction + +Semantic segmentation is a critical step in many computer vision tasks, and one that, like so many, is now predominantly achieved through deep learning [5, 35]. A key factor in the success of these methods has been access to high-quality annotated datasets. Pixel-accurate annotation of segmentation data is challenging, and becomes impractical as datasets grow. Unlabelled data is abundant and thus semi-supervised semantic segmentation has received increasing attention [15, 20, 8]. The predominant semisupervised approach employs a teacher-student framework to allocate pseudo labels to unlabelled data [11, 23, 36, 40]. These methods rely on the availability of a small but perfectly annotated dataset, however. This is a challenge in many practical situations where annotations are collected via crowd-sourcing (e.g., Sagemaker MTurk), or automated labelling techniques [31, 3], and are thus inevitably noisy. + +![](images/fabdcc8a9daaa5e8c802b4313e12b071775d52eaeab3a87dfc4aec2851407865.jpg) +Figure 1: Existing semi-supervised semantic segmentation models are vulnerable to the inevitable label noise in the pixel-level annotations. The proposed method performs significantly better than the current state-of-the-art [20] in the presence label noise. We apply three types of noise (labelled PL, RDE and SCP in Sec. 5.2) to achieve a $9\%$ error rate in the labelled data. + +We consider here the practically important case whereby semi-supervised semantic segmentation is to be applied despite the fact that the provided pixel labels contain noise. We propose an approach that is robust to this noise to the extent that it outperforms the baselines by a significant margin (see Fig. 1), despite being relatively simple to implement and incurring no extra computation cost at test time. + +It has been shown recently [19] that supervised segmentation models trained on noisy pixel-level labels first fit the clean labels during an "early-learning" phase before eventually memorizing the noisy labels. The above learning dynamics are exploited in [19] for weakly supervised segmentation by adding a multi-scale consistency loss. However, no noise detection module is used in [19] and the effect of unsupervised consistency loss on the label noise is weak. + +Recently there has been a surge of interest in data-centric methods [38, 39] that utilize the $k$ -NN of the feature representation of the data-points to detect annotation noise in classification datasets. A straightforward extension of this approach to semantic segmentation would lead to computing the $k$ -NN of the feature representation at pixel-level over all images, which would be infeasible. In this work, we + +adapt [39] to semi-supervised semantic segmentation by introducing an explicit label noise detection module that predicts label noise at the pixel level without significant additional memory or computation requirements. + +Pseudo labelling approaches [11, 23, 36, 2] for semi-supervised segmentation often use a single model to generate pseudo labels on the unlabelled data as well as train on them. This leads to confirmation bias as the model can overfit to its prediction errors. The same bias is apparent even when a single model is employed to filter label noise and train on those cleaned labels [19]. To mitigate this effect, we introduce two diverse learning groups with different network architectures that encourage mutual knowledge distillation and effective noise filtering. These learning groups also exploit complementary learning paths that boost the competency of each group. Our contributions are thus summarized as follows: + +- We propose an approach to semi-supervised semantic segmentation that is robust to label noise. +- We present a general architecture that maintains two diverse learning groups to overcome confirmation bias in label assignment problems. +- We introduce a filter module that extends [39] for semantic segmentation while demonstrating a good trade-off between accuracy and computation cost. +- We conduct extensive experiments showing that our method significantly outperforms semi/weakly-supervised semantic segmentation baselines regardless of the presence or the type of label noise. + +# 2. Related Work + +Semi-supervised Semantic Segmentation. Pixel-level annotation for semantic segmentation is time-consuming, costly, and error prone. Semi-supervised learning addresses the first two elements of this problem using unlabeled data. These approaches can be broadly categorized into self-training (a.k.a self-learning) based methods [11, 23, 10, 25, 29] and consistency regularization-based methods [40, 26, 6, 14, 20]. Self-training-based methods use models trained on labelled examples to predict pseudo-segmentation masks over the unlabeled examples, thus iteratively expanding the training set. The effectiveness of these methods depends on accurately identifying high-confidence pseudo labels via crafted auxiliary networks [11, 23, 36] or GANs [10, 25, 29]. Consistency regularization methods enforce consistency among predictions under varying perturbations, where the perturbations can be applied to the input image (French [8], PseudoSeg [40], CAC [16]), the intermediate feature representations (CCT [26]) or the network itself (CPS [6], GCT [14]). Most effort in this vein aims to identify perturbations that are challenging enough to support generalization without + +changing the underlying true labels. Particularly, recent work [20] on Mean Teacher [30] demonstrates superior performance and flexibility in the choice of perturbations. Our work builds upon [20] and introduces a general framework with two learning groups and a noise filter module to effectively learn from noisy labelled data and unlabelled data. + +Semantic Segmentation with Label Noise. Segmentation is a more challenging task than classification, and the labelled data is more prone to label noise. To achieve robustness some methods explicitly model the human annotation errors [12], or add a regularization term [28, 32]. Interestingly, [22] shows how high-level structure information (about microscopy images) learned by the model can be used as additional signal to remedy wrong annotations. In [24], a two stream mutual-attention network is devised to smooth out the influence of noisy labels during backpropagation. ADELE [19] exploits the learning dynamics demonstrated by noisy examples and proposes an early-learning strategy accordingly. These supervised segmentation methods assume all of the data is labellebd and that the majority of the labels are correct. They thus focus on alleviating the influence a small set of incorrect labels. + +The approach most similar to that proposed here is that of DivideMix [18] whereby two divergent networks are applied to separate data partitions. The method there is specifically designed for classification, however, as it is based on an image-level Gaussian Mixture Model. We propose here instead a pixel-level filter module appropriate for semantic segmentation. Additionally, in contrast to [18], we employ multiple learning groups with varying architectures to encourage diversity. + +# 3. Preliminaries + +# 3.1. Problem Setup + +We consider semi-supervised segmentation in the case where we have a small set of potentially noisy labelled examples and a large set of unlabelled examples. Let $\mathcal{D}_L = \{(x_i, y_i)\}_{i=1}^{N_L}$ be the labelled set where $x_i \in \mathcal{X} := [0,1]^{h \times w \times 3}$ and $y_i \in \mathcal{Y} := \mathcal{C}^{h \times w}$ denote an image and the corresponding pixel-wise segmentation mask respectively. Here, $\mathcal{C} = \{1,2,\dots,\ell\}$ denotes the set of labels. Similarly, let $\mathcal{D}_U = \{x_i\}_{i=1}^{N_U}$ denote the unlabelled set of images. + +In contrast to the standard semi-supervised segmentation problem, the segmentation masks $\{\pmb{y}_i\}$ may contain erroneous labels. + +# 3.2. Noisy Label Detection + +Here we briefly review the noisy label detection approach SimiFeat [38] developed for robust image classification. Our work extends this framework to the segmentation problem by pixel-level reasoning over noisy labels in + +![](images/ce750cb3b1cc39a49e1b4b65d976a0d803c597153fbe2c6441f37dbfa1307083.jpg) +Figure 2: The pipelines of a) the baseline model [20] and b) the proposed approach. In contrast to the baseline that uses two teachers and a single student network, we propose a framework with two diverse Learning Groups (LGs) $(t^1, s^1)$ and $(t^2, s^2)$ where each group contains a teacher and a student network. The two learning groups propagate complementary information and reduce the confirmation bias. Here, $t$ denotes a teacher model, $s$ is a student model, solid lines show the forward flow of labelled data, and dotted lines show the flow of unlabelled data. $L_{ce}$ denotes the cross-entropy loss and $L_{con}$ is the confidence-weighted cross-entropy loss. + +Sec. 4.2. SimiFeat estimates the label noise proportion for each class and then compute a per-class "noisiness" score for all the images in order to rank them and eliminate top $R_{c}$ percentile of images, where $R_{c}$ is the noise percentage for class $c$ . + +Let $\tilde{\mathcal{D}} = \left\{(\pmb{x}_i,\tilde{y}_i)\right\}$ be a dataset of images $\pmb{x}_i\in \mathcal{X}$ with noisy labels $\tilde{y}_i\in \mathcal{C}$ , and $g(\cdot):\mathcal{X}\to \mathbb{R}^d$ be a feature extractor. The noise proportion is computed from the prior probability of clean labels and the noise transition probability, which are estimated via High Order Consensus (HOC) (refer to [38] for more details) as follows: + +$$ +\mathbb {P} (Y), \mathbb {P} (\tilde {Y} | Y) \leftarrow \operatorname {H O C} \left(\left\{\left(g (\boldsymbol {x}), \tilde {y}\right) \mid (\boldsymbol {x}, \tilde {y}) \in \tilde {\mathcal {D}} \right\}\right), \tag {1} +$$ + +where $Y$ and $\tilde{Y}$ denote the random variables corresponding to clean and noisy labels respectively. The noise proportion is then calculated for a class $c \in \mathcal{C}$ as: + +$$ +R _ {c} = 1 - \mathbb {P} (Y = c | \tilde {Y} = c). \tag {2} +$$ + +Note, $\mathbb{P}(Y|\hat{Y})$ is obtained via Bayes' rule: + +$$ +\mathbb {P} (Y | \tilde {Y}) = \mathbb {P} (\tilde {Y} | Y) \mathbb {P} (Y) / \mathbb {P} (\tilde {Y}), \tag {3} +$$ + +where $\mathbb{P}(\tilde{Y})$ is computed from the observed labels and $\mathbb{P}(Y),\mathbb{P}(\tilde{Y}|Y)$ are computed using HOC (Eq. (1)). + +To compute the noisiness score, a $k$ -NN based soft label is employed in the feature space via nearest-neighbour voting. Then, the score for an image is determined as the negative cosine similarity between the soft label of that image and the one-hot encoding of the class label [38]. Finally, the images are ranked based on the score and the noise proportion from Eq. (2) is utilized to determine the percentile of images to be declared as noisy and consequently removed. + +# 4. Our Approach + +We introduce a general pipeline with two Learning Groups (LGs), as shown in Fig. 2. Each LG consists of a teacher-student framework and a noisy label detection module (Filter Modules). The teacher network is used to produce pseudo labels for the unlabelled data similar to the semi-supervised learning pipeline in [20]. To stabilize training, we first train both LGs independently in a fully-supervised manner with the labelled dataset $\mathcal{D}_L$ for a few epochs. After this warm-up, at each training epoch, we activate the filter module that detects the pixels with noisy labels in the labelled set $\mathcal{D}_L$ and move them to the unlabeled set $\mathcal{D}_U$ . Then, we train the pipeline in the semi-supervised manner. The noise detection and model training are repeated until the end of the training. + +Our filter module uses the feature representation of the teacher model from one LG to partition the labelled data into clean and noisy sets. The filter module then ignores the labels of the noisy set and moves the corresponding elements into the unlabelled set to obtain a revised labelled and unlabelled dataset pair. The updated dataset pair generated by one LG is consumed by the other. As training progresses each LG receives a different set of dataset pair due to the decoupled noise filtering, which enables complementary information to propagate in the two learning paths - reducing confirmation bias. This, along with the different network architectures for the two LGs, enhances the performance of our framework in the presence of noise. + +![](images/1a9d8edcb8e60da39647ab5c66cba070c516d5462d14d88629062acd8d9464e6.jpg) +Figure 3: The noise detection results from the two LGs. The images from left to right are: input, clean ground-truth followed by noise detection results from the two LGs when we use the same architecture (two R50s) and different architectures (R50 and PVTv2-B2 [33]). We get more diverse noise detection results when using different architectures. + +# 4.1. Learning Group + +The architecture of the LG used in our pipeline is inspired by the recent teacher-student framework [20] as depicted in Fig. 2 (a). We refer the reader to [20] for more details. + +At a training epoch $k$ , the first LG receives the revised labelled set $\mathcal{D}_{L,k}^{2}$ and unlabelled set $\mathcal{D}_{U,k}^{2}$ after filtering the noisy labels predicted by the second LG (details of $\mathcal{D}_{L,k}^{2}$ and $\mathcal{D}_{U,k}^{2}$ can be found in Sec. 4.2). Then the student network of the first LG minimizes the following loss: + +$$ +\mathcal {L} \left(\mathcal {D} _ {L, k} ^ {2}, \mathcal {D} _ {U, k} ^ {2}, \boldsymbol {\theta} _ {s ^ {1}}\right) := L _ {c e} \left(\mathcal {D} _ {L, k} ^ {2}, \boldsymbol {\theta} _ {s ^ {1}}\right) + \alpha L _ {c o n} \left(\mathcal {D} _ {U, k} ^ {2}, \boldsymbol {\theta} _ {s ^ {1}}\right). \tag {4} +$$ + +Here, $L_{ce}$ denotes the cross-entropy loss on the labelled data, $L_{con}$ is a confidence-weighted cross-entropy loss that enforces consistency between the teacher and student model predictions on the unlabelled data, $\theta_{s^1}$ denotes the parameters of the student network in the first LG, and $\alpha > 0$ provides a relative weight between the two losses. Analogously, the parameters of the second student network $\theta_{s^2}$ are optimized utilizing the revised labelled set $\mathcal{D}_{L,k}^{1}$ and the unlabelled dataset $\mathcal{D}_{U,k}^{1}$ predicted by the first LG. The teacher network is updated by the moving average of the student network parameters of the same LG. + +Note that, at each training epoch, the set of labelled data varies depending on the filter module output. Specifically, the detected pixels with noisy labels will be removed from the label set, which leads to modifying the segmentation mask of the labelled examples with an "ignore-pixel" label that enables the loss Eq. (4) to be defined in terms of $L_{con}$ instead of $L_{ce}$ at those pixels. + +To encourage diversity of the LGs, we utilize R50 [9] for the first LG and PVTv2-B2 [33] for the second. In Fig. 3, we show that the noise detection results from two R50 (the third and fourth columns) have less diversity than those from different network architectures (the last two + +![](images/a4d0dd96f18de45b65674825f033bd7cb9e374e1dc309cea7f41aba3eb0d0003.jpg) +Figure 4: The architecture of our filter module. First, we generate a representative dataset $\bar{\mathcal{D}}$ and calculate the noise proportion based on HOC [39]. Then, we get the adaptive threshold based on the noise proportion and example score. According to the adaptive threshold and example score, the original dataset $\mathcal{D}_L$ is divided into a clean set $\mathcal{D}_L^1$ and a noisy set. The noisy set is added to the unlabelled dataset $\mathcal{D}_U$ . + +columns).1 We also show that two different architectures bring better performance in our experiments (Fig. 7). + +# 4.2.Filter Module + +The task of the filter module is to detect incorrectly labelled pixels efficiently. To this end, rather than relying on label correction approaches [19], we design a filter module that accurately detects pixel-level noise scalably. The filter module is shown in Fig. 4. + +Note that, according to Eq. (1), the HOC approach requires a dataset of feature encodings and the corresponding labels. Since semantic segmentation is a classification problem at the pixel-level, the natural extension is to treat each pixel as a datapoint and directly apply HOC. Unfortunately, this becomes intractable as the computational complexity is quadratic in number of pixels, labels, and number of images in the dataset. + +To mitigate this, we use a small subset of pixels per class to generate a representative set $\bar{\mathcal{D}} = \{(\pmb{x}_i,j,\tilde{y}_{i,j})\}$ , which comprises of a fixed number of pixels per image subsampled from $\mathcal{D}_L$ . Here, $\pmb{x}_i\in \mathcal{X}$ is an image, $j\in \mathcal{I}(\pmb{x}_i)$ is the index of a pixel in image $\pmb{x}_i$ , and $\tilde{y}_{i,j}\in \mathcal{C}$ is the label of the $j$ -th pixel. In order to generate such a set, we found that random pixel sampling is more effective than elaborate methods such as $k$ -means clustering, and is more computationally efficient also. + +Let $g_{j}^{1}(\cdot):\mathcal{X}\to \mathbb{R}^{d}$ denotes the backbone feature extractor, i.e., the penultimate layer output for the first LG at pixel $j$ (where $j\in \mathcal{I}(\pmb {x}_i)\coloneqq \{1,\dots ,w\times h\}$ if images are resized to a fixed size). We construct the embedded repre + +sentative set $\bar{\mathcal{D}}_G^1 \coloneqq \left\{(g_j^1(\boldsymbol{x}_i), \tilde{y}_{i,j}) \mid (\boldsymbol{x}_i, j, \tilde{y}_{i,j}) \in \bar{\mathcal{D}}\right\}$ and invoke HOC (Eq. (1)) to estimate the noise proportion $\{R_c^1\}_{c \in \mathcal{C}}$ at pixel-level (Eq. (2)). Analogously, the noise proportion $\{R_c^2\}_{c \in \mathcal{C}}$ for the other LG can be estimated by constructing $\bar{\mathcal{D}}_G^2$ . + +Let $h_j^1(\cdot) : \mathcal{X} \to [0,1]^{\ell}$ and $h_j^2(\cdot) : \mathcal{X} \to [0,1]^{\ell}$ be the teacher model mappings to softmax scores in the first and second LG respectively. We compute a noisiness score based on the teacher model predictions, as opposed to using $k$ -NN in SimiFeat [38]. This eliminates the computation cost of running $k$ -NN for every pixel over all images in $\mathcal{D}_L$ . We define the score for the first LG at pixel $j$ in image $x_i$ as: + +$$ +S ^ {1} \left(\boldsymbol {x} _ {i}, j\right) = \operatorname {C E} \left(h _ {j} ^ {1} \left(\boldsymbol {x} _ {i}\right), \boldsymbol {e} _ {i, j}\right) + \lambda \operatorname {K L} \left(h _ {j} ^ {1} \left(\boldsymbol {x} _ {i}\right) \| h _ {j} ^ {2} \left(\boldsymbol {x} _ {i}\right)\right) \tag {5} +$$ + +where CE and KL denote the cross-entropy and Kullback-Leibler (KL) divergence respectively, $e_{i,j} \in \{0,1\}^{\ell}$ is the one-hot encoded noisy label at pixel $j$ in image $x_i$ , and $\lambda > 0$ . CE measures the alignment between label noise and model predictions, and KL measures the inconsistency between the two teacher model predictions. Large CE at a pixel implies a high misalignment between the predicted class label and annotated label and hence high model confidence of the label to be noisy unless the current model overfits at that pixel. On the other hand, if one of the models starts to overfit the noisy annotations, then the KL measure will be high. If any of them is high, such pixel corresponds to a high score, and it is marked as noisy. Experimentally, we verify that the above choice of score works better than the other alternatives (shown in Fig. 8(b)). + +Once the noisiness score and the noise proportion have been calculated, we get the adaptive noise threshold as follows: + +$$ +\beta^ {1} = \frac {\sigma \ell S _ {a v g 1} ^ {1}}{\sum_ {c \in \mathcal {C}} R _ {c} ^ {1}}, \text {w h e r e} S _ {a v g 1} ^ {1} = n ^ {- 1} \sum_ {\substack {\left(\boldsymbol {x} _ {i},\cdot\right) \in \mathcal {D} _ {L}, \\ j \in \mathcal {I} \left(\boldsymbol {x} _ {i}\right)}} S ^ {1} \left(\boldsymbol {x} _ {i}, j\right), \tag{6} +$$ + +where $\sigma$ is a hyperparameter and $n$ denotes the number of all the pixels in $\mathcal{D}_L$ . Experimentally we found that the noise proportion $R_{c}$ predicted by HOC is inaccurate for some of the classes and a single threshold $\beta^1$ , computed from the average as in Eq. (6), works better than class-wise thresholds. A higher $\beta$ means filtering out less noisy annotations. We utilize the adaptive threshold $\beta^1$ to determine the pixels to be removed and marked in the segmentation masks to reflect the pixels with noisy labels. Let us denote the set of pixels that are detected as label noise by the first LG as: + +$$ +\mathcal {M} _ {k} ^ {1} := \left\{\left(\boldsymbol {x} _ {i}, j\right) \mid S ^ {1} \left(\boldsymbol {x} _ {i}, j\right) > \beta^ {1}, \left(\boldsymbol {x} _ {i}, \cdot\right) \in \mathcal {D} _ {L}, j \in \mathcal {I} \left(\boldsymbol {x} _ {i}\right) \right\} \tag {7} +$$ + +The updated datasets at epoch $k$ are created by removing the pixels with noise labels from the labelled data and placing the images with noisy pixels in the unlabelled set, i.e., the + +Algorithm 1 Our training pipeline +Input: $\mathcal{D}_L, \mathcal{D}_U$ , total epochs $E$ , warm-up epochs $N$ +Output: Optimized $\{\theta_{s^1}, \theta_{t^1}, \theta_{s^2}, \theta_{t^2}\}$ +1: Initialize parameters $\{\theta_{s^1}, \theta_{t^1}, \theta_{s^2}, \theta_{t^2}\}$ , +2: for $k \in \{0, 1, \dots, E\}$ do +3: if $k < N$ then +4: $\mathcal{D}_{L,k}^{1}, \mathcal{D}_{L,k}^{2} \coloneqq \mathcal{D}_{L}$ and $\mathcal{D}_{U,k}^{1}, \mathcal{D}_{U,k}^{2} \coloneqq \emptyset$ +5: else +6: $\{(x_i, j, \tilde{y}_{i,j})\} \gets$ Sample from $\mathcal{D}_L$ +7: for $l \in \{1, 2\}$ do +8: $\{R_c^l\} \gets$ HOC $\left(\left(g_j^l(x_i), \tilde{y}_{i,j}\right)\right)$ +9: Scores $\{S^l\} \gets$ Using Eq. (5) +10: $\beta^l \gets$ Adaptive $\left(\{S^l\}\right)$ [Eq. (6)] +11: $\mathcal{M}_k^l \gets$ Using Eq. (7) +12: $\mathcal{D}_{L,k}^l \gets \mathcal{D}_L \backslash \mathcal{M}_k^l$ +13: $\mathcal{D}_{U,k}^l \gets \mathcal{D}_U \bigcup \{x_i | (x_i, \cdot) \in \mathcal{M}_k^l\}$ +14: repeat +15: $\theta_{s^1} \gets$ Backprop. $\mathcal{L}(\mathcal{D}_{L,k}^2, \mathcal{D}_{U,k}^2, \theta_{s^1})$ [Eq. (4)] +16: $\theta_{s^2} \gets$ Backprop. $\mathcal{L}(\mathcal{D}_{L,k}^1, \mathcal{D}_{U,k}^1, \theta_{s^2})$ [Eq. (4)] +17: until $M$ times +18: Update $\theta_{t^1}, \theta_{t^2} \gets$ Moving Average of $\theta_{s^1}, \theta_{s^2}$ . + +clean labelled set $\mathcal{D}_{L,k}^{1} \coloneqq \mathcal{D}_{L} \setminus \mathcal{M}_{k}^{12}$ and the unlabelled dataset $\mathcal{D}_{U,k}^{1} \coloneqq \mathcal{D}_{U} \bigcup \{\pmb{x}_{i} \mid (\pmb{x}_{i},\cdot) \in \mathcal{M}_{k}^{1}\}$ . The student network of the second LG is then trained on these datasets according to Eq. (4). Analogously, the score for the second LG can be computed by switching $h_{j}^{1}$ and $h_{j}^{2}$ in Eq. (5), and the datasets $\mathcal{D}_{L,k}^{2}, \mathcal{D}_{U,k}^{2}$ are created in the above manner on which the student network of the first LG is trained. Our final algorithm is summarized in Algorithm 1. + +# 5. Experiments + +Evaluation has been carried out following [19, 20] to allow direct comparison. As in [20], the labelled data is augmented with synthetic label noise. Performance is measured with, and without an unlabelled dataset to illuminate performance differences between the weakly and semi supervised settings. + +# 5.1. Implementation Details + +We utilize R50 [9] and PVTv2-B2 [33] as the backbones in the two LGs to encourage diversity. Both the teacher and student models are based on DeeplabV3+ [4]. For inference, we use the average of the predictions from the two teacher networks as in [20]. The $\lambda$ in Eq. (5) is set to 0.1 to balance the contribution from the CE and KL terms in our loss. We randomly sample 100 pixels per class in each image to get the representative dataset $\bar{D}$ . The hyperparameter $\sigma$ in Eq. (6) is set to be 10 in our experiments. Other + +Table 1: Comparison against SotA approaches on Pascal VOC 2012 in a semi-supervised semantic segmentation setting. All baselines are based on the DeeplabV3+ architecture. The * indicates results reported by [6]. When we replace $50\%$ of the ground truth of images with 'PL', the proportion of noisy pixels is $9\%$ . We keep the same pixel level noise proportion while introducing noise using 'RDE' and 'SCP' for parity. + +
MethodsYearNoisy pixelNoise typeLabelled Data Ratio
1/16 (662)1/8(1323)1/4 (2646)1/2 (5291)
MT* [30]20170%None66.7070.7873.2275.41
French [8]20190%None68.9070.7072.4674.49
CCT* [26]20200%None65.2270.8773.4374.75
GCT* [14]20200%None64.0570.4773.4575.20
ECS* [23]20200%None-67.3870.7072.89
CPS* [6]20210%None71.9873.6774.9076.15
CAC* [16]20210%None70.1072.4074.00-
PS-MT [20]20220%None72.8375.7076.4377.88
Ours-0%None77.75 (+4.92)79.31 (+3.61)79.14 (+2.71)79.54 (+1.66)
PS-MT [20]-9%PL61.9065.1465.5366.78
Ours-9%PL71.60 (+9.70)73.95 (+8.81)74.53 (+9.00)74.44 (+7.66)
PS-MT [20]-9%RDE58.8762.1863.8163.46
Ours-9%RDE66.82 (+7.95)69.95 (+7.77)73.82 (+10.01)74.47 (+11.01)
PS-MT [20]-9%SCP52.1457.1956.8153.69
Ours-9%SCP67.80 (+15.66)70.83 (+13.64)69.38 (+12.57)75.95 (+22.26)
+ +training and inference settings are the same as those of the baseline method [20]. In all of our experiments, the last epoch checkpoint model is used unless otherwise specified. Further details of the implementation can be found in the supplementary materials. + +# 5.2. Semi-supervised Semantic Segmentation + +Pascal VOC 2012 [7] is employed for evaluating the performance of our model in a semi-supervised semantic segmentation setting. This dataset contains more than $13\mathrm{k}$ images with 21 different classes. To introduce realistic noise in $\mathcal{D}_L$ , the following three approaches are considered: + +- Pseudo Labels (PL). Following [19], a portion of clean labels are replaced by pseudo-labels from a weakly-supervised segmentation method [34]. +- Random Dilation and Erosion (RDE). Similar to [19, 21], we apply random degrees of dilation and erosion to the ground-truth masks. +- Similar Class Perturbation (SCP). We introduce another type of noise by randomly replacing ground-truth label of a segmentation mask with a different class that resembles the texture of original clean class. This label perturbation simulates a more natural scenario pertaining to annotators confusing semantics. + +![](images/258e736694f9fbb32fe84882b00d6438ecedc057edcbc0aa017b6d9439ac6b81.jpg) +Figure 5: A performance comparison under different noise proportions achieved by replacing (from $0\%$ to $90\%$ ) clean segmentation masks with the PL from [34] at image-level. The green font shows the corresponding noisy pixel proportions. All experiments are conducted with $1/8$ labelled data. + +Details about the implementation and visualization of noise generation can be found in the supplementary material. + +In the context of semi-supervised semantic segmentation, we have compared our approach with the following studies: MT [30], French [8], CCT [26], GCT [14], ECS [23], CPS [6], CAC [16], and PS-MT [20]. Following the previous studies [19, 20], we report the mean Intersection-over-Union (mIoU) on the validation set. Table 1 shows that our approach ('Ours') outperforms all baselines trained with clean labelled data. In the presence of noise, we compare the best performing semi-supervised segmentation method + +Table 2: The mIoU of the weakly-supervised semantic segmentation baselines and ours on the test set of SegTHOR. Best Val and Last Epoch correspond to the checkpoint that performs best on the validation set and the last checkpoint. + +
ModelBest ValLast Epoch
Base-ADELE [19]62.659.1
ADELE [19]71.170.8
Ours73.273.5
+ +with no noise, i.e., PS-MT [20], against our method with varying noise types (see Table 1) and noise proportions (see Fig. 5). The results demonstrate significant improvements in mIoU in all cases, e.g., $+22.26\%$ with half of the labelled data on the more practical SCP noise. Overall, the superiority of our method is consistent across all labelled data to unlabelled data ratios, regardless of the presence of noise or the type of applied noise. + +# 5.3. Weakly-supervised Semantic Segmentation + +We evaluate our method in a weakly-supervised semantic segmentation setting on the following datasets: (i) SegTHOR [17], a medical dataset, and (ii) pseudo labelled Pascal VOC 2012. + +Medical image segmentation on SegTHOR. Following [19], we evaluate our approach on the SegTHOR dataset, which consists of 3D CT scans. Each pixel of these images is assigned to one of these five classes: 'esophagus', 'heart', 'trachea', 'aorta', or 'background'. The slices from the images are randomly split into a training set of 3638 slices, a validation set of 570 slices, and a test set of 580 slices [19, 17]. To introduce noise, we apply 'RDE' to the ground-truth segmentation labels, similar to [19]. Note that all annotations in the training set are corrupted by this noise, whereas the validation and test sets remain uncorrupted. + +We compare our approach with ADELE [19], which corrects noisy annotations with model predictions in the early training stage, and its variant, Base-ADELE [19], which is based on a UNet trained with multi-scale inputs. Detailed implementations of these approaches are in the supplementary material. Following the setting in ADELE [19], we evaluate two checkpoints: specifically, the one with the highest mIoU on the validation set and the last checkpoint. In Table 2, our approach outperforms both baselines. + +Weakly-supervised segmentation on Pascal VOC 2012. We demonstrate the effectiveness of our approach in comparison to the baselines in the weakly-supervised setting on the PASCAL VOC 2012 [7]. Similar to ADELE [19], we use only noisy pseudo labels generated by [34] to train the models and evaluate on the original test set. Table 3 shows that our approach outperforms recent baselines, such as ADELE combined with SEAM [19]. In addition, PS-MT + +increases the mIoU marginally with PVTv2 compared to R50 with 66.7 vs. 64.0. However, after using our framework, the mIoU is further improved from 66.7 to 71.7, clearly demonstrating the effectiveness of our method. + +Table 3: Comparison with state-of-the-art weakly supervised semantic segmentation methods on the Pascal VOC 2012 dataset using mIoU (\%) of the validation set. + +
ModelYearmIoU (%)
AffinityNet [1]201861.7
SSDD [27]201964.9
SEAM [34]202064.5
CONTA [37]202066.1
SPML [13]202169.5
ADELE [19] + SEAM202269.3
PS-MT [20] (R50)202264.0
PS-MT [20] (PVTv2-B2)202266.7
Ours-71.7
+ +# 5.4. Ablation Study + +All experiments in this section are conducted with $1/8$ labelled data ratio and $9\%$ PL noisy annotations on Pascal VOC. We verify the effectiveness of the two main components in our approach, the diverse LGs and filter module. + +Diversity in the LGs with different architectures. In contrast to previous works [20, 18], we found that using different network architectures in the two LGs provides more complementary information. In Fig. 3, we show the qualitative results of noise detection when we use the same and different architectures in the two LGs. The two LGs provide more diverse noise detection results (lower overlap between predicted noisy labels) when we employ different architectures. In Fig. 6, the predictions from the two LGs with two different architectures $(\mathrm{e})\& (\mathrm{f})$ are more complementary than those from the two LGs with the same architecture $(\mathrm{(c)})\& (\mathrm{(d)})$ . Fig. 7 shows the mIoU of our pipeline when we use different network combinations. The combination of R50 and PVTv2-B2 [33] achieves the best performance. + +Noise detection with different LGs. To avoid confirmation bias, we utilize the teacher network from one LG to detect noise for the other LG. If we utilize the teacher network from the same LG for noise detection (i.e., replacing $\theta_{s^1}$ with $\theta_{s^2}$ in Eq. (4)), the mIoU drops from $74.0\%$ to $72.4\%$ on Pascal VOC, demonstrating the importance of mutual noise detection with a different LG in our pipeline. + +Significance of the filter module. In Table 4, in addition to the baseline PS-MT and ours, we evaluate the performance of the following models: 1) 'PS-MT + FM' by adding the filter module to PS-MT (using R50)4. 2) 'Ours - FM' by + +![](images/6a3db4253c085eb04cdbfe4fe6c83696905a8191a60ac4f6d3fb283aa39bd780.jpg) +(a) Input Image + +![](images/c3b5a11eede915f6419485aea4029dc925796f7fc31657df27010a2bab5e27dd.jpg) +(b) Ground Truth + +![](images/8c003e5fd52937d8a5ffab945a6e83b638ae187b59ec986f7dc6b5cf9f29ce0a.jpg) +(c) R50LG1 Prediction + +![](images/6b9b8830d77bc507268576a0cb02b8bcea6acd9c924c76e1472c5f479bfbdead.jpg) +(d) R50LG2 Prediction + +![](images/5d9c76a0db3e7ec9963e96ec63cf1d93271eec4d13f121f3acb6e588de4cc440.jpg) +(e) R50LG1 Prediction +(f) PVT $^{\mathrm{LG2}}$ Prediction + +![](images/8460ca2812ce4034caad854523fae174e22b50c336a51cfb945117f94ae98073.jpg) + +Table 4: Ablation study of FM and LG on Pascal VOC. + +
Method2 LGsFilter Module (FM)mIoU (%)
PS-MT [20]65.1
PS-MT + FM✓67.6 (+2.5)
Ours - FM✓69.9 (+4.8)
Ours✓✓74.0 (+8.9)
+ +![](images/bc98f6d422e6c3b0ebd1df1df6d9fe07ddcdf6443f7957bc070b2247c618456e.jpg) +(a) Different Architectures +Figure 7: (a) shows the mIoU score of different network combinations on clean data. (b) shows the mIoU score of different combinations on different settings. R, S and P denote R50 [9], Segformer [35] and PVTv2-B2 [33] respectively. -FM_C denotes training our model (without the filter module) with clean data. -FM_N denotes training model with noisy data without using filter module. +FM_N is our complete model trained with noisy data. + +![](images/e94072a5805a4bebca868e757affaedee2b0acacfdbca7dcaccfb246204a5c09.jpg) +(b) Clean and Noisy Data +Figure 8: (a) shows detected noise proportion of GMM [18] and ours, where $x$ -axis shows the noisy image ratio and $y$ -axis shows the detected noisy pixel ratio. The detected noise proportion of ours is closer to the GT than GMM. (b) shows the F1 score of noise detection with different criterion. CE is cross entropy loss; KL is KL divergence between the predictions from the two LGs. Our $\mathrm{CE} + \mathrm{KL}$ performs the best. + +removing the filter module from our framework. It can be seen that the filter module helps to improve the final performance for both models. The filter module contributes more significantly (+4.1) to our framework due to the diverse learning groups (i.e., using 2 LGs with different architectures). This superiority of our framework is further evidenced by the better performance of 'Ours - FM' compared to PS-MT variants. In Fig. 7 (b), our Filter Module helps to improve the performance under noisy annotations ('+FM_N' vs. '-FM_N') with different architectures. + +Comparison of the filter module with GMM. Previous work [18] utilizes Gaussian Mixture Model (GMM) to detect noise for image classification. We compare our filter + +![](images/e09b7a2ab53bb65aa0c01d31ed3520279e4011d22e393a5e93d2b1ab14654628.jpg) +Figure 6: Visualizations from the two LGs when we use the same architecture ((c) and (d)) and different architectures ((e) and (f)). The $1st$ and $2nd$ columns show the inputs and the ground truth. R50 and PVT denote R50 and PVTv2-B2 [33]. +(a) Detected Noise Ratio +(b) F1 Score of Noise Detection + +![](images/eda195d2494cfea237ab1c65eec41cd4ee24dfda863400c1cf98d8a3569b300a.jpg) + +module with GMM (including two components) on the representative dataset. Fig. 8 (a) shows the detected noise proportions. We can see that the detected noise proportion of our filter module is closer to the groundtruth than GMM. + +Criterion to compute score. We utilize cross entropy loss and KL divergence $(\mathrm{CE} + \mathrm{KL})$ to compute score in Eq. (5). Here, we compare this criterion with others, including only cross entropy loss (CE), and only KL divergence (KL). Fig. 8 (b) shows that ours achieves the best F1 score. + +Memory and computation costs. PVTv2-B2 has similar GFLOPs (45.8) with R50 (45.6). Ours only have one more student network than PS-MT, which is not used during inference. Thus, our model needs similar memory and computation costs compared with PS-MT during inference. + +# 6. Conclusion + +In this work, we propose a learning framework for semi-supervised semantic segmentation involving two diverse learning groups that can handle various labelling errors. Our framework employs a carefully designed filter module to divide the labelled dataset into pixels with and without label noise in a scalable manner. The potentially misla + +belled pixels filtered by the first learning group are utilised for training the second learning group and vice versa. We show that our proposed framework provides state-of-the-art results for semi-supervised learning, by addressing confirmation bias while training, that can handle various amounts of annotation noise to accommodate practical inaccuracies in annotations. + +# References + +[1] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4981-4990, 2018. +[2] Eric Arazo, Diego Ortego, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2020. +[3] Liyi Chen, Weiwei Wu, Chenchen Fu, Xiao Han, and Yuntao Zhang. Weakly supervised semantic segmentation with boundary exploration. In European Conference on Computer Vision, pages 347-362. Springer, 2020. +[4] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. +[5] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. +[6] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021. +[7] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. +[8] Geoff French, Timo Aila, Samuli Laine, Michal Mackiewicz, and Graham Finlayson. Semi-supervised semantic segmentation needs strong, high-dimensional perturbations. 2019. +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[10] Wei-Chih Hung, Yi-Hsuan Tsai, Yan-Ting Liou, Yen-Yu Lin, and Ming-Hsuan Yang. Adversarial learning for semi-supervised semantic segmentation. arXiv preprint arXiv:1802.07934, 2018. +[11] Mostafa S Ibrahim, Arash Vahdat, Mani Ranjbar, and William G Macready. Semi-supervised semantic image segmentation with self-correcting networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12715-12725, 2020. + +[12] Joseph Jacob, Olga Ciccarelli, Frederik Barkhof, and Daniel C Alexander. Disentangling human error from the ground truth in segmentation of medical images. +[13] Tsung-Wei Ke, Jyh-Jing Hwang, and Stella X Yu. Universal weakly supervised segmentation by pixel-to-segment contrastive learning. arXiv preprint arXiv:2105.00957, 2021. +[14] Zhanghan Ke, Di Qiu, Kaican Li, Qiong Yan, and Rynson WH Lau. Guided collaborative training for pixel-wise semi-supervised learning. In European conference on computer vision, pages 429-445. Springer, 2020. +[15] Donghyeon Kwon and Suha Kwak. Semi-supervised semantic segmentation with error localization network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9957-9967, 2022. +[16] Xin Lai, Zhuotao Tian, Li Jiang, Shu Liu, Hengshuang Zhao, Liwei Wang, and Jiaya Jia. Semi-supervised semantic segmentation with directional context-aware consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1205-1214, 2021. +[17] Zoé Lambert, Caroline Petitjean, Bernard Dubray, and Su Kuan. Segthor: segmentation of thoracic organs at risk in ct images. In 2020 Tenth International Conference on Image Processing Theory, Tools and Applications (IPTA), pages 1-6. IEEE, 2020. +[18] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394, 2020. +[19] Sheng Liu, Kangning Liu, Weicheng Zhu, Yiqiu Shen, and Carlos Fernandez-Granda. Adaptive early-learning correction for segmentation from noisy annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2606–2616, 2022. +[20] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4258-4267, 2022. +[21] Yaoru Luo, Guole Liu, Yuanhao Guo, and Ge Yang. Elucidating meta-structures of noisy labels in semantic segmentation by deep neural networks. arXiv preprint arXiv:2205.00160, 2022. +[22] Yaoru Luo, Guole Liu, Wenjing Li, Yuanhao Guo, and Ge Yang. Deep neural networks learn meta-structures to segment fluorescence microscopy images. arXiv preprint arXiv:2103.11594, 2021. +[23] Robert Mendel, Luis Antonio de Souza, David Rauber, Joao Paulo Papa, and Christoph Palm. Semi-supervised segmentation based on error-correcting supervision. In European Conference on Computer Vision, pages 141-157. Springer, 2020. +[24] Shaobo Min, Xuejin Chen, Zheng-Jun Zha, Feng Wu, and Yongdong Zhang. A two-stream mutual attention network for semi-supervised biomedical segmentation with noisy labels. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 4578-4585, 2019. + +[25] Sudhanshu Mittal, Maxim Tatarchenko, and Thomas Brox. Semi-supervised semantic segmentation with high-and low-level consistency. IEEE transactions on pattern analysis and machine intelligence, 43(4):1369-1379, 2019. +[26] Yassine Ouali, Céline Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross-consistency training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12674-12684, 2020. +[27] Wataru Shimoda and Keiji Yanai. Self-supervised difference detection for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5208-5217, 2019. +[28] Yucheng Shu, Xiao Wu, and Weisheng Li. Lvc-net: Medical image segmentation with noisy label based on local visual cues. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 558-566. Springer, 2019. +[29] Nasim Souly, Concetto Spampinato, and Mubarak Shah. Semi supervised semantic segmentation using generative adversarial network. In Proceedings of the IEEE international conference on computer vision, pages 5688-5696, 2017. +[30] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. +[31] Luis Garcia Ugarriza, Eli Saber, Sreenath Rao Vantaram, Vincent Amuso, Mark Shaw, and Ranjit Bhaskar. Automatic image segmentation by dynamic region growth and multiresolution merging. IEEE transactions on image processing, 18(10):2275-2288, 2009. +[32] Guotai Wang, Xinglong Liu, Chaoping Li, Zhiyong Xu, Jiugen Ruan, Haifeng Zhu, Tao Meng, Kang Li, Ning Huang, and Shaoting Zhang. A noise-robust framework for automatic segmentation of Covid-19 pneumonia lesions from ct images. IEEE Transactions on Medical Imaging, 39(8):2653-2663, 2020. +[33] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. +[34] Yude Wang, Jie Zhang, Meina Kan, Shiguang Shan, and Xilin Chen. Self-supervised equivariant attention mechanism for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12275-12284, 2020. +[35] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. +[36] Rumeng Yi, Yaping Huang, Qingji Guan, Mengyang Pu, and Runsheng Zhang. Learning from pixel-level label noise: A new perspective for semi-supervised semantic segmentation. IEEE Transactions on Image Processing, 31:623-635, 2021. +[37] Dong Zhang, Hanwang Zhang, Jinhui Tang, Xian-Sheng Hua, and Qianru Sun. Causal intervention for weakly- + +supervised semantic segmentation. Advances in Neural Information Processing Systems, 33:655-666, 2020. +[38] Zhaowei Zhu, Zihao Dong, and Yang Liu. Detecting corrupted labels without training a model to predict. In International Conference on Machine Learning, pages 27412-27427. PMLR, 2022. +[39] Zhaowei Zhu, Yiwen Song, and Yang Liu. Clusterability as an alternative to anchor points when learning with noisy labels. In International Conference on Machine Learning, pages 12912-12923. PMLR, 2021. +[40] Yuliang Zou, Zizhao Zhang, Han Zhang, Chun-Liang Li, Xiao Bian, Jia-Bin Huang, and Tomas Pfister. Pseudoseg: Designing pseudo labels for semantic segmentation. arXiv preprint arXiv:2010.09713, 2020. \ No newline at end of file diff --git a/2023/Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups/images.zip b/2023/Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3deedc9c1aecaf0e7499cfabc01df0ac613ebca1 --- /dev/null +++ b/2023/Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82b4ba3b18d832395b8ed1fd39459b323236718d240f3cb37a3b79ce59c94dff +size 456132 diff --git a/2023/Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups/layout.json b/2023/Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6193460ddf14aed800b5b736975f0510d44c56be --- /dev/null +++ b/2023/Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups/layout.json @@ -0,0 +1,9974 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 483, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 483, + 140 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 483, + 140 + ], + "type": "text", + "content": "Semi-Supervised Semantic Segmentation under Label Noise via Diverse Learning Groups" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "spans": [ + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "text", + "content": "Peixia Li" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "text", + "content": ", Pulak Purkait" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "text", + "content": ", Thalaiyasingam Ajanthan" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "text", + "content": ", Majid Abdolshah" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "text", + "content": ", Ravi Garg" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "text", + "content": ", Hisham Husain" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "text", + "content": ", Chenchen Xu" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "text", + "content": ", Stephen Gould" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "text", + "content": ", Wanli Ouyang" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "text", + "content": ", Anton van den Hengel" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "text", + "content": ", The University of Sydney," + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "text", + "content": " Amazon, Australian National University," + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 160, + 525, + 218 + ], + "type": "text", + "content": " The University of Adelaide" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 271, + 289, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 289, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 289, + 487 + ], + "type": "text", + "content": "Semi-supervised semantic segmentation methods use a small amount of clean pixel-level annotations to guide the interpretation of a larger quantity of unlabelled image data. The challenges of providing pixel-accurate annotations at scale mean that the labels are typically noisy, and this contaminates the final results. In this work, we propose an approach that is robust to label noise in the annotated data. The method uses two diverse learning groups with different network architectures to effectively handle both label noise and unlabelled images. Each learning group consists of a teacher network, a student network and a novel filter module. The filter module of each learning group utilizes pixel-level features from the teacher network to detect incorrectly labelled pixels. To reduce confirmation bias, we employ the labels cleaned by the filter module from one learning group to train the other learning group. Experimental results on two different benchmarks and settings demonstrate the superiority of our method over state-of-the-art approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 501, + 128, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 128, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 128, + 513 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "text", + "content": "Semantic segmentation is a critical step in many computer vision tasks, and one that, like so many, is now predominantly achieved through deep learning [5, 35]. A key factor in the success of these methods has been access to high-quality annotated datasets. Pixel-accurate annotation of segmentation data is challenging, and becomes impractical as datasets grow. Unlabelled data is abundant and thus semi-supervised semantic segmentation has received increasing attention [15, 20, 8]. The predominant semisupervised approach employs a teacher-student framework to allocate pseudo labels to unlabelled data [11, 23, 36, 40]. These methods rely on the availability of a small but perfectly annotated dataset, however. This is a challenge in many practical situations where annotations are collected via crowd-sourcing (e.g., Sagemaker MTurk), or automated labelling techniques [31, 3], and are thus inevitably noisy." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 245, + 547, + 359 + ], + "blocks": [ + { + "bbox": [ + 308, + 245, + 547, + 359 + ], + "lines": [ + { + "bbox": [ + 308, + 245, + 547, + 359 + ], + "spans": [ + { + "bbox": [ + 308, + 245, + 547, + 359 + ], + "type": "image", + "image_path": "fabdcc8a9daaa5e8c802b4313e12b071775d52eaeab3a87dfc4aec2851407865.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 361, + 547, + 445 + ], + "lines": [ + { + "bbox": [ + 304, + 361, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 547, + 445 + ], + "type": "text", + "content": "Figure 1: Existing semi-supervised semantic segmentation models are vulnerable to the inevitable label noise in the pixel-level annotations. The proposed method performs significantly better than the current state-of-the-art [20] in the presence label noise. We apply three types of noise (labelled PL, RDE and SCP in Sec. 5.2) to achieve a " + }, + { + "bbox": [ + 304, + 361, + 547, + 445 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 304, + 361, + 547, + 445 + ], + "type": "text", + "content": " error rate in the labelled data." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 450, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 533 + ], + "type": "text", + "content": "We consider here the practically important case whereby semi-supervised semantic segmentation is to be applied despite the fact that the provided pixel labels contain noise. We propose an approach that is robust to this noise to the extent that it outperforms the baselines by a significant margin (see Fig. 1), despite being relatively simple to implement and incurring no extra computation cost at test time." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 534, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 629 + ], + "type": "text", + "content": "It has been shown recently [19] that supervised segmentation models trained on noisy pixel-level labels first fit the clean labels during an \"early-learning\" phase before eventually memorizing the noisy labels. The above learning dynamics are exploited in [19] for weakly supervised segmentation by adding a multi-scale consistency loss. However, no noise detection module is used in [19] and the effect of unsupervised consistency loss on the label noise is weak." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Recently there has been a surge of interest in data-centric methods [38, 39] that utilize the " + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "-NN of the feature representation of the data-points to detect annotation noise in classification datasets. A straightforward extension of this approach to semantic segmentation would lead to computing the " + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "-NN of the feature representation at pixel-level over all images, which would be infeasible. In this work, we" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "1229" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "type": "text", + "content": "adapt [39] to semi-supervised semantic segmentation by introducing an explicit label noise detection module that predicts label noise at the pixel level without significant additional memory or computation requirements." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 286, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 286, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 286, + 275 + ], + "type": "text", + "content": "Pseudo labelling approaches [11, 23, 36, 2] for semi-supervised segmentation often use a single model to generate pseudo labels on the unlabelled data as well as train on them. This leads to confirmation bias as the model can overfit to its prediction errors. The same bias is apparent even when a single model is employed to filter label noise and train on those cleaned labels [19]. To mitigate this effect, we introduce two diverse learning groups with different network architectures that encourage mutual knowledge distillation and effective noise filtering. These learning groups also exploit complementary learning paths that boost the competency of each group. Our contributions are thus summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 280, + 286, + 429 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 58, + 280, + 286, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 280, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 58, + 280, + 286, + 304 + ], + "type": "text", + "content": "- We propose an approach to semi-supervised semantic segmentation that is robust to label noise." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 306, + 286, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 306, + 286, + 341 + ], + "spans": [ + { + "bbox": [ + 58, + 306, + 286, + 341 + ], + "type": "text", + "content": "- We present a general architecture that maintains two diverse learning groups to overcome confirmation bias in label assignment problems." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 343, + 286, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 343, + 286, + 379 + ], + "spans": [ + { + "bbox": [ + 58, + 343, + 286, + 379 + ], + "type": "text", + "content": "- We introduce a filter module that extends [39] for semantic segmentation while demonstrating a good trade-off between accuracy and computation cost." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 382, + 286, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 382, + 286, + 429 + ], + "spans": [ + { + "bbox": [ + 58, + 382, + 286, + 429 + ], + "type": "text", + "content": "- We conduct extensive experiments showing that our method significantly outperforms semi/weakly-supervised semantic segmentation baselines regardless of the presence or the type of label noise." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 442, + 133, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 442, + 133, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 442, + 133, + 453 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 462, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 286, + 713 + ], + "type": "text", + "content": "Semi-supervised Semantic Segmentation. Pixel-level annotation for semantic segmentation is time-consuming, costly, and error prone. Semi-supervised learning addresses the first two elements of this problem using unlabeled data. These approaches can be broadly categorized into self-training (a.k.a self-learning) based methods [11, 23, 10, 25, 29] and consistency regularization-based methods [40, 26, 6, 14, 20]. Self-training-based methods use models trained on labelled examples to predict pseudo-segmentation masks over the unlabeled examples, thus iteratively expanding the training set. The effectiveness of these methods depends on accurately identifying high-confidence pseudo labels via crafted auxiliary networks [11, 23, 36] or GANs [10, 25, 29]. Consistency regularization methods enforce consistency among predictions under varying perturbations, where the perturbations can be applied to the input image (French [8], PseudoSeg [40], CAC [16]), the intermediate feature representations (CCT [26]) or the network itself (CPS [6], GCT [14]). Most effort in this vein aims to identify perturbations that are challenging enough to support generalization without" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "changing the underlying true labels. Particularly, recent work [20] on Mean Teacher [30] demonstrates superior performance and flexibility in the choice of perturbations. Our work builds upon [20] and introduces a general framework with two learning groups and a noise filter module to effectively learn from noisy labelled data and unlabelled data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 153, + 545, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 153, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 153, + 545, + 344 + ], + "type": "text", + "content": "Semantic Segmentation with Label Noise. Segmentation is a more challenging task than classification, and the labelled data is more prone to label noise. To achieve robustness some methods explicitly model the human annotation errors [12], or add a regularization term [28, 32]. Interestingly, [22] shows how high-level structure information (about microscopy images) learned by the model can be used as additional signal to remedy wrong annotations. In [24], a two stream mutual-attention network is devised to smooth out the influence of noisy labels during backpropagation. ADELE [19] exploits the learning dynamics demonstrated by noisy examples and proposes an early-learning strategy accordingly. These supervised segmentation methods assume all of the data is labellebd and that the majority of the labels are correct. They thus focus on alleviating the influence a small set of incorrect labels." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 345, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 345, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 345, + 545, + 453 + ], + "type": "text", + "content": "The approach most similar to that proposed here is that of DivideMix [18] whereby two divergent networks are applied to separate data partitions. The method there is specifically designed for classification, however, as it is based on an image-level Gaussian Mixture Model. We propose here instead a pixel-level filter module appropriate for semantic segmentation. Additionally, in contrast to [18], we employ multiple learning groups with varying architectures to encourage diversity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 458, + 389, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 458, + 389, + 469 + ], + "spans": [ + { + "bbox": [ + 306, + 458, + 389, + 469 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 478, + 397, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 397, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 397, + 491 + ], + "type": "text", + "content": "3.1. Problem Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 496, + 545, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 603 + ], + "type": "text", + "content": "We consider semi-supervised segmentation in the case where we have a small set of potentially noisy labelled examples and a large set of unlabelled examples. Let " + }, + { + "bbox": [ + 304, + 496, + 545, + 603 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L = \\{(x_i, y_i)\\}_{i=1}^{N_L}" + }, + { + "bbox": [ + 304, + 496, + 545, + 603 + ], + "type": "text", + "content": " be the labelled set where " + }, + { + "bbox": [ + 304, + 496, + 545, + 603 + ], + "type": "inline_equation", + "content": "x_i \\in \\mathcal{X} := [0,1]^{h \\times w \\times 3}" + }, + { + "bbox": [ + 304, + 496, + 545, + 603 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 496, + 545, + 603 + ], + "type": "inline_equation", + "content": "y_i \\in \\mathcal{Y} := \\mathcal{C}^{h \\times w}" + }, + { + "bbox": [ + 304, + 496, + 545, + 603 + ], + "type": "text", + "content": " denote an image and the corresponding pixel-wise segmentation mask respectively. Here, " + }, + { + "bbox": [ + 304, + 496, + 545, + 603 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{1,2,\\dots,\\ell\\}" + }, + { + "bbox": [ + 304, + 496, + 545, + 603 + ], + "type": "text", + "content": " denotes the set of labels. Similarly, let " + }, + { + "bbox": [ + 304, + 496, + 545, + 603 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_U = \\{x_i\\}_{i=1}^{N_U}" + }, + { + "bbox": [ + 304, + 496, + 545, + 603 + ], + "type": "text", + "content": " denote the unlabelled set of images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 604, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 639 + ], + "type": "text", + "content": "In contrast to the standard semi-supervised segmentation problem, the segmentation masks " + }, + { + "bbox": [ + 304, + 604, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\{\\pmb{y}_i\\}" + }, + { + "bbox": [ + 304, + 604, + 545, + 639 + ], + "type": "text", + "content": " may contain erroneous labels." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 647, + 432, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 432, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 432, + 658 + ], + "type": "text", + "content": "3.2. Noisy Label Detection" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Here we briefly review the noisy label detection approach SimiFeat [38] developed for robust image classification. Our work extends this framework to the segmentation problem by pixel-level reasoning over noisy labels in" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 732, + 315, + 741 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 732, + 315, + 741 + ], + "spans": [ + { + "bbox": [ + 296, + 732, + 315, + 741 + ], + "type": "text", + "content": "1230" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 531, + 232 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 531, + 232 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 531, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 531, + 232 + ], + "type": "image", + "image_path": "ce750cb3b1cc39a49e1b4b65d976a0d803c597153fbe2c6441f37dbfa1307083.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 240, + 545, + 313 + ], + "lines": [ + { + "bbox": [ + 46, + 240, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 545, + 313 + ], + "type": "text", + "content": "Figure 2: The pipelines of a) the baseline model [20] and b) the proposed approach. In contrast to the baseline that uses two teachers and a single student network, we propose a framework with two diverse Learning Groups (LGs) " + }, + { + "bbox": [ + 46, + 240, + 545, + 313 + ], + "type": "inline_equation", + "content": "(t^1, s^1)" + }, + { + "bbox": [ + 46, + 240, + 545, + 313 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 240, + 545, + 313 + ], + "type": "inline_equation", + "content": "(t^2, s^2)" + }, + { + "bbox": [ + 46, + 240, + 545, + 313 + ], + "type": "text", + "content": " where each group contains a teacher and a student network. The two learning groups propagate complementary information and reduce the confirmation bias. Here, " + }, + { + "bbox": [ + 46, + 240, + 545, + 313 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 240, + 545, + 313 + ], + "type": "text", + "content": " denotes a teacher model, " + }, + { + "bbox": [ + 46, + 240, + 545, + 313 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 240, + 545, + 313 + ], + "type": "text", + "content": " is a student model, solid lines show the forward flow of labelled data, and dotted lines show the flow of unlabelled data. " + }, + { + "bbox": [ + 46, + 240, + 545, + 313 + ], + "type": "inline_equation", + "content": "L_{ce}" + }, + { + "bbox": [ + 46, + 240, + 545, + 313 + ], + "type": "text", + "content": " denotes the cross-entropy loss and " + }, + { + "bbox": [ + 46, + 240, + 545, + 313 + ], + "type": "inline_equation", + "content": "L_{con}" + }, + { + "bbox": [ + 46, + 240, + 545, + 313 + ], + "type": "text", + "content": " is the confidence-weighted cross-entropy loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 325, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 287, + 384 + ], + "type": "text", + "content": "Sec. 4.2. SimiFeat estimates the label noise proportion for each class and then compute a per-class \"noisiness\" score for all the images in order to rank them and eliminate top " + }, + { + "bbox": [ + 46, + 325, + 287, + 384 + ], + "type": "inline_equation", + "content": "R_{c}" + }, + { + "bbox": [ + 46, + 325, + 287, + 384 + ], + "type": "text", + "content": " percentile of images, where " + }, + { + "bbox": [ + 46, + 325, + 287, + 384 + ], + "type": "inline_equation", + "content": "R_{c}" + }, + { + "bbox": [ + 46, + 325, + 287, + 384 + ], + "type": "text", + "content": " is the noise percentage for class " + }, + { + "bbox": [ + 46, + 325, + 287, + 384 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 325, + 287, + 384 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}} = \\left\\{(\\pmb{x}_i,\\tilde{y}_i)\\right\\}" + }, + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "text", + "content": " be a dataset of images " + }, + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i\\in \\mathcal{X}" + }, + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "text", + "content": " with noisy labels " + }, + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\tilde{y}_i\\in \\mathcal{C}" + }, + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "inline_equation", + "content": "g(\\cdot):\\mathcal{X}\\to \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "text", + "content": " be a feature extractor. The noise proportion is computed from the prior probability of clean labels and the noise transition probability, which are estimated via High Order Consensus (HOC) (refer to [38] for more details) as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 464, + 287, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 464, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 287, + 479 + ], + "type": "interline_equation", + "content": "\\mathbb {P} (Y), \\mathbb {P} (\\tilde {Y} | Y) \\leftarrow \\operatorname {H O C} \\left(\\left\\{\\left(g (\\boldsymbol {x}), \\tilde {y}\\right) \\mid (\\boldsymbol {x}, \\tilde {y}) \\in \\tilde {\\mathcal {D}} \\right\\}\\right), \\tag {1}", + "image_path": "24812b545881a6e59ba8f37c8102791d3c47a8080502c3760b62d0481a1176c9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 487, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 522 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 487, + 287, + 522 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 487, + 287, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 487, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\tilde{Y}" + }, + { + "bbox": [ + 46, + 487, + 287, + 522 + ], + "type": "text", + "content": " denote the random variables corresponding to clean and noisy labels respectively. The noise proportion is then calculated for a class " + }, + { + "bbox": [ + 46, + 487, + 287, + 522 + ], + "type": "inline_equation", + "content": "c \\in \\mathcal{C}" + }, + { + "bbox": [ + 46, + 487, + 287, + 522 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 108, + 530, + 287, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 530, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 108, + 530, + 287, + 544 + ], + "type": "interline_equation", + "content": "R _ {c} = 1 - \\mathbb {P} (Y = c | \\tilde {Y} = c). \\tag {2}", + "image_path": "3144f5474bb407b4028f3e0da48a200dc995839c28e530edb26ff0c241d50fb2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 552, + 218, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 218, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 218, + 565 + ], + "type": "text", + "content": "Note, " + }, + { + "bbox": [ + 47, + 552, + 218, + 565 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(Y|\\hat{Y})" + }, + { + "bbox": [ + 47, + 552, + 218, + 565 + ], + "type": "text", + "content": " is obtained via Bayes' rule:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 99, + 571, + 287, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 571, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 99, + 571, + 287, + 586 + ], + "type": "interline_equation", + "content": "\\mathbb {P} (Y | \\tilde {Y}) = \\mathbb {P} (\\tilde {Y} | Y) \\mathbb {P} (Y) / \\mathbb {P} (\\tilde {Y}), \\tag {3}", + "image_path": "2a9d785c18d74826e305258d7e640cfb0121d657199fa996230cb9d91b19095a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 593, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 593, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(\\tilde{Y})" + }, + { + "bbox": [ + 46, + 593, + 287, + 617 + ], + "type": "text", + "content": " is computed from the observed labels and " + }, + { + "bbox": [ + 46, + 593, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(Y),\\mathbb{P}(\\tilde{Y}|Y)" + }, + { + "bbox": [ + 46, + 593, + 287, + 617 + ], + "type": "text", + "content": " are computed using HOC (Eq. (1))." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "To compute the noisiness score, a " + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "-NN based soft label is employed in the feature space via nearest-neighbour voting. Then, the score for an image is determined as the negative cosine similarity between the soft label of that image and the one-hot encoding of the class label [38]. Finally, the images are ranked based on the score and the noise proportion from Eq. (2) is utilized to determine the percentile of images to be declared as noisy and consequently removed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 324, + 395, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 324, + 395, + 338 + ], + "spans": [ + { + "bbox": [ + 306, + 324, + 395, + 338 + ], + "type": "text", + "content": "4. Our Approach" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 367, + 545, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 367, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 545, + 535 + ], + "type": "text", + "content": "We introduce a general pipeline with two Learning Groups (LGs), as shown in Fig. 2. Each LG consists of a teacher-student framework and a noisy label detection module (Filter Modules). The teacher network is used to produce pseudo labels for the unlabelled data similar to the semi-supervised learning pipeline in [20]. To stabilize training, we first train both LGs independently in a fully-supervised manner with the labelled dataset " + }, + { + "bbox": [ + 304, + 367, + 545, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L" + }, + { + "bbox": [ + 304, + 367, + 545, + 535 + ], + "type": "text", + "content": " for a few epochs. After this warm-up, at each training epoch, we activate the filter module that detects the pixels with noisy labels in the labelled set " + }, + { + "bbox": [ + 304, + 367, + 545, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L" + }, + { + "bbox": [ + 304, + 367, + 545, + 535 + ], + "type": "text", + "content": " and move them to the unlabeled set " + }, + { + "bbox": [ + 304, + 367, + 545, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_U" + }, + { + "bbox": [ + 304, + 367, + 545, + 535 + ], + "type": "text", + "content": ". Then, we train the pipeline in the semi-supervised manner. The noise detection and model training are repeated until the end of the training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Our filter module uses the feature representation of the teacher model from one LG to partition the labelled data into clean and noisy sets. The filter module then ignores the labels of the noisy set and moves the corresponding elements into the unlabelled set to obtain a revised labelled and unlabelled dataset pair. The updated dataset pair generated by one LG is consumed by the other. As training progresses each LG receives a different set of dataset pair due to the decoupled noise filtering, which enables complementary information to propagate in the two learning paths - reducing confirmation bias. This, along with the different network architectures for the two LGs, enhances the performance of our framework in the presence of noise." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 732, + 313, + 741 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 732, + 313, + 741 + ], + "spans": [ + { + "bbox": [ + 296, + 732, + 313, + 741 + ], + "type": "text", + "content": "1231" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 289, + 163 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 289, + 163 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 289, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 289, + 163 + ], + "type": "image", + "image_path": "1a9d8edcb8e60da39647ab5c66cba070c516d5462d14d88629062acd8d9464e6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 171, + 289, + 244 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 289, + 244 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 289, + 244 + ], + "type": "text", + "content": "Figure 3: The noise detection results from the two LGs. The images from left to right are: input, clean ground-truth followed by noise detection results from the two LGs when we use the same architecture (two R50s) and different architectures (R50 and PVTv2-B2 [33]). We get more diverse noise detection results when using different architectures." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 266, + 147, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 147, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 147, + 280 + ], + "type": "text", + "content": "4.1. Learning Group" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 286, + 287, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 286, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 287, + 334 + ], + "type": "text", + "content": "The architecture of the LG used in our pipeline is inspired by the recent teacher-student framework [20] as depicted in Fig. 2 (a). We refer the reader to [20] for more details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 335, + 289, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 289, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 289, + 399 + ], + "type": "text", + "content": "At a training epoch " + }, + { + "bbox": [ + 46, + 335, + 289, + 399 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 335, + 289, + 399 + ], + "type": "text", + "content": ", the first LG receives the revised labelled set " + }, + { + "bbox": [ + 46, + 335, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{L,k}^{2}" + }, + { + "bbox": [ + 46, + 335, + 289, + 399 + ], + "type": "text", + "content": " and unlabelled set " + }, + { + "bbox": [ + 46, + 335, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{U,k}^{2}" + }, + { + "bbox": [ + 46, + 335, + 289, + 399 + ], + "type": "text", + "content": " after filtering the noisy labels predicted by the second LG (details of " + }, + { + "bbox": [ + 46, + 335, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{L,k}^{2}" + }, + { + "bbox": [ + 46, + 335, + 289, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 335, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{U,k}^{2}" + }, + { + "bbox": [ + 46, + 335, + 289, + 399 + ], + "type": "text", + "content": " can be found in Sec. 4.2). Then the student network of the first LG minimizes the following loss:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 408, + 291, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 408, + 291, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 408, + 291, + 434 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(\\mathcal {D} _ {L, k} ^ {2}, \\mathcal {D} _ {U, k} ^ {2}, \\boldsymbol {\\theta} _ {s ^ {1}}\\right) := L _ {c e} \\left(\\mathcal {D} _ {L, k} ^ {2}, \\boldsymbol {\\theta} _ {s ^ {1}}\\right) + \\alpha L _ {c o n} \\left(\\mathcal {D} _ {U, k} ^ {2}, \\boldsymbol {\\theta} _ {s ^ {1}}\\right). \\tag {4}", + "image_path": "e13648f6a3d4bf4049ab9407ecceb15610ce812c98f369a8def7676399235ad1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 434, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 567 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 434, + 287, + 567 + ], + "type": "inline_equation", + "content": "L_{ce}" + }, + { + "bbox": [ + 46, + 434, + 287, + 567 + ], + "type": "text", + "content": " denotes the cross-entropy loss on the labelled data, " + }, + { + "bbox": [ + 46, + 434, + 287, + 567 + ], + "type": "inline_equation", + "content": "L_{con}" + }, + { + "bbox": [ + 46, + 434, + 287, + 567 + ], + "type": "text", + "content": " is a confidence-weighted cross-entropy loss that enforces consistency between the teacher and student model predictions on the unlabelled data, " + }, + { + "bbox": [ + 46, + 434, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\theta_{s^1}" + }, + { + "bbox": [ + 46, + 434, + 287, + 567 + ], + "type": "text", + "content": " denotes the parameters of the student network in the first LG, and " + }, + { + "bbox": [ + 46, + 434, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\alpha > 0" + }, + { + "bbox": [ + 46, + 434, + 287, + 567 + ], + "type": "text", + "content": " provides a relative weight between the two losses. Analogously, the parameters of the second student network " + }, + { + "bbox": [ + 46, + 434, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\theta_{s^2}" + }, + { + "bbox": [ + 46, + 434, + 287, + 567 + ], + "type": "text", + "content": " are optimized utilizing the revised labelled set " + }, + { + "bbox": [ + 46, + 434, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{L,k}^{1}" + }, + { + "bbox": [ + 46, + 434, + 287, + 567 + ], + "type": "text", + "content": " and the unlabelled dataset " + }, + { + "bbox": [ + 46, + 434, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{U,k}^{1}" + }, + { + "bbox": [ + 46, + 434, + 287, + 567 + ], + "type": "text", + "content": " predicted by the first LG. The teacher network is updated by the moving average of the student network parameters of the same LG." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "text", + "content": "Note that, at each training epoch, the set of labelled data varies depending on the filter module output. Specifically, the detected pixels with noisy labels will be removed from the label set, which leads to modifying the segmentation mask of the labelled examples with an \"ignore-pixel\" label that enables the loss Eq. (4) to be defined in terms of " + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "inline_equation", + "content": "L_{con}" + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "inline_equation", + "content": "L_{ce}" + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "text", + "content": " at those pixels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "content": "To encourage diversity of the LGs, we utilize R50 [9] for the first LG and PVTv2-B2 [33] for the second. In Fig. 3, we show that the noise detection results from two R50 (the third and fourth columns) have less diversity than those from different network architectures (the last two" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 70, + 547, + 182 + ], + "blocks": [ + { + "bbox": [ + 308, + 70, + 547, + 182 + ], + "lines": [ + { + "bbox": [ + 308, + 70, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 70, + 547, + 182 + ], + "type": "image", + "image_path": "a4d0dd96f18de45b65674825f033bd7cb9e374e1dc309cea7f41aba3eb0d0003.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 190, + 547, + 287 + ], + "lines": [ + { + "bbox": [ + 304, + 190, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 547, + 287 + ], + "type": "text", + "content": "Figure 4: The architecture of our filter module. First, we generate a representative dataset " + }, + { + "bbox": [ + 304, + 190, + 547, + 287 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{D}}" + }, + { + "bbox": [ + 304, + 190, + 547, + 287 + ], + "type": "text", + "content": " and calculate the noise proportion based on HOC [39]. Then, we get the adaptive threshold based on the noise proportion and example score. According to the adaptive threshold and example score, the original dataset " + }, + { + "bbox": [ + 304, + 190, + 547, + 287 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L" + }, + { + "bbox": [ + 304, + 190, + 547, + 287 + ], + "type": "text", + "content": " is divided into a clean set " + }, + { + "bbox": [ + 304, + 190, + 547, + 287 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L^1" + }, + { + "bbox": [ + 304, + 190, + 547, + 287 + ], + "type": "text", + "content": " and a noisy set. The noisy set is added to the unlabelled dataset " + }, + { + "bbox": [ + 304, + 190, + 547, + 287 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_U" + }, + { + "bbox": [ + 304, + 190, + 547, + 287 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 308, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 334 + ], + "type": "text", + "content": "columns).1 We also show that two different architectures bring better performance in our experiments (Fig. 7)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 342, + 394, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 342, + 394, + 354 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 394, + 354 + ], + "type": "text", + "content": "4.2.Filter Module" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 361, + 545, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 421 + ], + "type": "text", + "content": "The task of the filter module is to detect incorrectly labelled pixels efficiently. To this end, rather than relying on label correction approaches [19], we design a filter module that accurately detects pixel-level noise scalably. The filter module is shown in Fig. 4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 422, + 546, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 546, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 546, + 516 + ], + "type": "text", + "content": "Note that, according to Eq. (1), the HOC approach requires a dataset of feature encodings and the corresponding labels. Since semantic segmentation is a classification problem at the pixel-level, the natural extension is to treat each pixel as a datapoint and directly apply HOC. Unfortunately, this becomes intractable as the computational complexity is quadratic in number of pixels, labels, and number of images in the dataset." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 517, + 547, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 547, + 624 + ], + "type": "text", + "content": "To mitigate this, we use a small subset of pixels per class to generate a representative set " + }, + { + "bbox": [ + 304, + 517, + 547, + 624 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{D}} = \\{(\\pmb{x}_i,j,\\tilde{y}_{i,j})\\}" + }, + { + "bbox": [ + 304, + 517, + 547, + 624 + ], + "type": "text", + "content": ", which comprises of a fixed number of pixels per image subsampled from " + }, + { + "bbox": [ + 304, + 517, + 547, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L" + }, + { + "bbox": [ + 304, + 517, + 547, + 624 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 517, + 547, + 624 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i\\in \\mathcal{X}" + }, + { + "bbox": [ + 304, + 517, + 547, + 624 + ], + "type": "text", + "content": " is an image, " + }, + { + "bbox": [ + 304, + 517, + 547, + 624 + ], + "type": "inline_equation", + "content": "j\\in \\mathcal{I}(\\pmb{x}_i)" + }, + { + "bbox": [ + 304, + 517, + 547, + 624 + ], + "type": "text", + "content": " is the index of a pixel in image " + }, + { + "bbox": [ + 304, + 517, + 547, + 624 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i" + }, + { + "bbox": [ + 304, + 517, + 547, + 624 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 517, + 547, + 624 + ], + "type": "inline_equation", + "content": "\\tilde{y}_{i,j}\\in \\mathcal{C}" + }, + { + "bbox": [ + 304, + 517, + 547, + 624 + ], + "type": "text", + "content": " is the label of the " + }, + { + "bbox": [ + 304, + 517, + 547, + 624 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 517, + 547, + 624 + ], + "type": "text", + "content": "-th pixel. In order to generate such a set, we found that random pixel sampling is more effective than elaborate methods such as " + }, + { + "bbox": [ + 304, + 517, + 547, + 624 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 517, + 547, + 624 + ], + "type": "text", + "content": "-means clustering, and is more computationally efficient also." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 625, + 547, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 625, + 547, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 625, + 547, + 673 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 625, + 547, + 673 + ], + "type": "inline_equation", + "content": "g_{j}^{1}(\\cdot):\\mathcal{X}\\to \\mathbb{R}^{d}" + }, + { + "bbox": [ + 304, + 625, + 547, + 673 + ], + "type": "text", + "content": " denotes the backbone feature extractor, i.e., the penultimate layer output for the first LG at pixel " + }, + { + "bbox": [ + 304, + 625, + 547, + 673 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 625, + 547, + 673 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 304, + 625, + 547, + 673 + ], + "type": "inline_equation", + "content": "j\\in \\mathcal{I}(\\pmb {x}_i)\\coloneqq \\{1,\\dots ,w\\times h\\}" + }, + { + "bbox": [ + 304, + 625, + 547, + 673 + ], + "type": "text", + "content": " if images are resized to a fixed size). We construct the embedded repre" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 305, + 683, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 683, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 683, + 547, + 713 + ], + "type": "text", + "content": "1It should be noted that the two networks with R50 architecture are trained with different initialization independently on the labelled data for a warm start for a fair comparison." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "text", + "content": "1232" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 71, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 71, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 287, + 133 + ], + "type": "text", + "content": "sentative set " + }, + { + "bbox": [ + 46, + 71, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{D}}_G^1 \\coloneqq \\left\\{(g_j^1(\\boldsymbol{x}_i), \\tilde{y}_{i,j}) \\mid (\\boldsymbol{x}_i, j, \\tilde{y}_{i,j}) \\in \\bar{\\mathcal{D}}\\right\\}" + }, + { + "bbox": [ + 46, + 71, + 287, + 133 + ], + "type": "text", + "content": " and invoke HOC (Eq. (1)) to estimate the noise proportion " + }, + { + "bbox": [ + 46, + 71, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\{R_c^1\\}_{c \\in \\mathcal{C}}" + }, + { + "bbox": [ + 46, + 71, + 287, + 133 + ], + "type": "text", + "content": " at pixel-level (Eq. (2)). Analogously, the noise proportion " + }, + { + "bbox": [ + 46, + 71, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\{R_c^2\\}_{c \\in \\mathcal{C}}" + }, + { + "bbox": [ + 46, + 71, + 287, + 133 + ], + "type": "text", + "content": " for the other LG can be estimated by constructing " + }, + { + "bbox": [ + 46, + 71, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{D}}_G^2" + }, + { + "bbox": [ + 46, + 71, + 287, + 133 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 132, + 288, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 132, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 132, + 288, + 228 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 132, + 288, + 228 + ], + "type": "inline_equation", + "content": "h_j^1(\\cdot) : \\mathcal{X} \\to [0,1]^{\\ell}" + }, + { + "bbox": [ + 46, + 132, + 288, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 132, + 288, + 228 + ], + "type": "inline_equation", + "content": "h_j^2(\\cdot) : \\mathcal{X} \\to [0,1]^{\\ell}" + }, + { + "bbox": [ + 46, + 132, + 288, + 228 + ], + "type": "text", + "content": " be the teacher model mappings to softmax scores in the first and second LG respectively. We compute a noisiness score based on the teacher model predictions, as opposed to using " + }, + { + "bbox": [ + 46, + 132, + 288, + 228 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 132, + 288, + 228 + ], + "type": "text", + "content": "-NN in SimiFeat [38]. This eliminates the computation cost of running " + }, + { + "bbox": [ + 46, + 132, + 288, + 228 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 132, + 288, + 228 + ], + "type": "text", + "content": "-NN for every pixel over all images in " + }, + { + "bbox": [ + 46, + 132, + 288, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L" + }, + { + "bbox": [ + 46, + 132, + 288, + 228 + ], + "type": "text", + "content": ". We define the score for the first LG at pixel " + }, + { + "bbox": [ + 46, + 132, + 288, + 228 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 132, + 288, + 228 + ], + "type": "text", + "content": " in image " + }, + { + "bbox": [ + 46, + 132, + 288, + 228 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 46, + 132, + 288, + 228 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 52, + 234, + 287, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 234, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 52, + 234, + 287, + 251 + ], + "type": "interline_equation", + "content": "S ^ {1} \\left(\\boldsymbol {x} _ {i}, j\\right) = \\operatorname {C E} \\left(h _ {j} ^ {1} \\left(\\boldsymbol {x} _ {i}\\right), \\boldsymbol {e} _ {i, j}\\right) + \\lambda \\operatorname {K L} \\left(h _ {j} ^ {1} \\left(\\boldsymbol {x} _ {i}\\right) \\| h _ {j} ^ {2} \\left(\\boldsymbol {x} _ {i}\\right)\\right) \\tag {5}", + "image_path": "8c42cb96074f73ab957e9c1c3acde96993ca837f7013347f7880af7d83e66d76.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 259, + 287, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 259, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 287, + 438 + ], + "type": "text", + "content": "where CE and KL denote the cross-entropy and Kullback-Leibler (KL) divergence respectively, " + }, + { + "bbox": [ + 46, + 259, + 287, + 438 + ], + "type": "inline_equation", + "content": "e_{i,j} \\in \\{0,1\\}^{\\ell}" + }, + { + "bbox": [ + 46, + 259, + 287, + 438 + ], + "type": "text", + "content": " is the one-hot encoded noisy label at pixel " + }, + { + "bbox": [ + 46, + 259, + 287, + 438 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 259, + 287, + 438 + ], + "type": "text", + "content": " in image " + }, + { + "bbox": [ + 46, + 259, + 287, + 438 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 46, + 259, + 287, + 438 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 259, + 287, + 438 + ], + "type": "inline_equation", + "content": "\\lambda > 0" + }, + { + "bbox": [ + 46, + 259, + 287, + 438 + ], + "type": "text", + "content": ". CE measures the alignment between label noise and model predictions, and KL measures the inconsistency between the two teacher model predictions. Large CE at a pixel implies a high misalignment between the predicted class label and annotated label and hence high model confidence of the label to be noisy unless the current model overfits at that pixel. On the other hand, if one of the models starts to overfit the noisy annotations, then the KL measure will be high. If any of them is high, such pixel corresponds to a high score, and it is marked as noisy. Experimentally, we verify that the above choice of score works better than the other alternatives (shown in Fig. 8(b))." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 440, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 440, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 287, + 475 + ], + "type": "text", + "content": "Once the noisiness score and the noise proportion have been calculated, we get the adaptive noise threshold as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 479, + 287, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 287, + 528 + ], + "type": "interline_equation", + "content": "\\beta^ {1} = \\frac {\\sigma \\ell S _ {a v g 1} ^ {1}}{\\sum_ {c \\in \\mathcal {C}} R _ {c} ^ {1}}, \\text {w h e r e} S _ {a v g 1} ^ {1} = n ^ {- 1} \\sum_ {\\substack {\\left(\\boldsymbol {x} _ {i},\\cdot\\right) \\in \\mathcal {D} _ {L}, \\\\ j \\in \\mathcal {I} \\left(\\boldsymbol {x} _ {i}\\right)}} S ^ {1} \\left(\\boldsymbol {x} _ {i}, j\\right), \\tag{6}", + "image_path": "6cd8816cda34a67d254edaac7d22d8f2f7f8eed4d0f1e96d4d1d6b2184311dc0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 527, + 288, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 527, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 527, + 288, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 527, + 288, + 647 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 527, + 288, + 647 + ], + "type": "text", + "content": " is a hyperparameter and " + }, + { + "bbox": [ + 46, + 527, + 288, + 647 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 527, + 288, + 647 + ], + "type": "text", + "content": " denotes the number of all the pixels in " + }, + { + "bbox": [ + 46, + 527, + 288, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L" + }, + { + "bbox": [ + 46, + 527, + 288, + 647 + ], + "type": "text", + "content": ". Experimentally we found that the noise proportion " + }, + { + "bbox": [ + 46, + 527, + 288, + 647 + ], + "type": "inline_equation", + "content": "R_{c}" + }, + { + "bbox": [ + 46, + 527, + 288, + 647 + ], + "type": "text", + "content": " predicted by HOC is inaccurate for some of the classes and a single threshold " + }, + { + "bbox": [ + 46, + 527, + 288, + 647 + ], + "type": "inline_equation", + "content": "\\beta^1" + }, + { + "bbox": [ + 46, + 527, + 288, + 647 + ], + "type": "text", + "content": ", computed from the average as in Eq. (6), works better than class-wise thresholds. A higher " + }, + { + "bbox": [ + 46, + 527, + 288, + 647 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 527, + 288, + 647 + ], + "type": "text", + "content": " means filtering out less noisy annotations. We utilize the adaptive threshold " + }, + { + "bbox": [ + 46, + 527, + 288, + 647 + ], + "type": "inline_equation", + "content": "\\beta^1" + }, + { + "bbox": [ + 46, + 527, + 288, + 647 + ], + "type": "text", + "content": " to determine the pixels to be removed and marked in the segmentation masks to reflect the pixels with noisy labels. Let us denote the set of pixels that are detected as label noise by the first LG as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 651, + 294, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 294, + 677 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {k} ^ {1} := \\left\\{\\left(\\boldsymbol {x} _ {i}, j\\right) \\mid S ^ {1} \\left(\\boldsymbol {x} _ {i}, j\\right) > \\beta^ {1}, \\left(\\boldsymbol {x} _ {i}, \\cdot\\right) \\in \\mathcal {D} _ {L}, j \\in \\mathcal {I} \\left(\\boldsymbol {x} _ {i}\\right) \\right\\} \\tag {7}", + "image_path": "fd6c8e8938fb60f5455c396eb404404638ddee9ffda26e17d4acdca17af129f6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "The updated datasets at epoch " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " are created by removing the pixels with noise labels from the labelled data and placing the images with noisy pixels in the unlabelled set, i.e., the" + } + ] + } + ], + "index": 8 + }, + { + "type": "code", + "bbox": [ + 306, + 72, + 545, + 339 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": "Algorithm 1 Our training pipeline \nInput: " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L, \\mathcal{D}_U" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": ", total epochs " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": ", warm-up epochs " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " \nOutput: Optimized " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\{\\theta_{s^1}, \\theta_{t^1}, \\theta_{s^2}, \\theta_{t^2}\\}" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " \n1: Initialize parameters " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\{\\theta_{s^1}, \\theta_{t^1}, \\theta_{s^2}, \\theta_{t^2}\\}" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": ", \n2: for " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "k \\in \\{0, 1, \\dots, E\\}" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " do \n3: if " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "k < N" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " then \n4: " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{L,k}^{1}, \\mathcal{D}_{L,k}^{2} \\coloneqq \\mathcal{D}_{L}" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{U,k}^{1}, \\mathcal{D}_{U,k}^{2} \\coloneqq \\emptyset" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " \n5: else \n6: " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\{(x_i, j, \\tilde{y}_{i,j})\\} \\gets" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " Sample from " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " \n7: for " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "l \\in \\{1, 2\\}" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " do \n8: " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\{R_c^l\\} \\gets" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " HOC " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\left(\\left(g_j^l(x_i), \\tilde{y}_{i,j}\\right)\\right)" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " \n9: Scores " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\{S^l\\} \\gets" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " Using Eq. (5) \n10: " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\beta^l \\gets" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " Adaptive " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\left(\\{S^l\\}\\right)" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " [Eq. (6)] \n11: " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_k^l \\gets" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " Using Eq. (7) \n12: " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{L,k}^l \\gets \\mathcal{D}_L \\backslash \\mathcal{M}_k^l" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " \n13: " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{U,k}^l \\gets \\mathcal{D}_U \\bigcup \\{x_i | (x_i, \\cdot) \\in \\mathcal{M}_k^l\\}" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " \n14: repeat \n15: " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\theta_{s^1} \\gets" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " Backprop. " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\mathcal{D}_{L,k}^2, \\mathcal{D}_{U,k}^2, \\theta_{s^1})" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " [Eq. (4)] \n16: " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\theta_{s^2} \\gets" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " Backprop. " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\mathcal{D}_{L,k}^1, \\mathcal{D}_{U,k}^1, \\theta_{s^2})" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " [Eq. (4)] \n17: until " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " times \n18: Update " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\theta_{t^1}, \\theta_{t^2} \\gets" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": " Moving Average of " + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\theta_{s^1}, \\theta_{s^2}" + }, + { + "bbox": [ + 306, + 72, + 545, + 339 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_body" + } + ], + "index": 9, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 357, + 545, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 545, + 457 + ], + "type": "text", + "content": "clean labelled set " + }, + { + "bbox": [ + 304, + 357, + 545, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{L,k}^{1} \\coloneqq \\mathcal{D}_{L} \\setminus \\mathcal{M}_{k}^{12}" + }, + { + "bbox": [ + 304, + 357, + 545, + 457 + ], + "type": "text", + "content": " and the unlabelled dataset " + }, + { + "bbox": [ + 304, + 357, + 545, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{U,k}^{1} \\coloneqq \\mathcal{D}_{U} \\bigcup \\{\\pmb{x}_{i} \\mid (\\pmb{x}_{i},\\cdot) \\in \\mathcal{M}_{k}^{1}\\}" + }, + { + "bbox": [ + 304, + 357, + 545, + 457 + ], + "type": "text", + "content": ". The student network of the second LG is then trained on these datasets according to Eq. (4). Analogously, the score for the second LG can be computed by switching " + }, + { + "bbox": [ + 304, + 357, + 545, + 457 + ], + "type": "inline_equation", + "content": "h_{j}^{1}" + }, + { + "bbox": [ + 304, + 357, + 545, + 457 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 357, + 545, + 457 + ], + "type": "inline_equation", + "content": "h_{j}^{2}" + }, + { + "bbox": [ + 304, + 357, + 545, + 457 + ], + "type": "text", + "content": " in Eq. (5), and the datasets " + }, + { + "bbox": [ + 304, + 357, + 545, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{L,k}^{2}, \\mathcal{D}_{U,k}^{2}" + }, + { + "bbox": [ + 304, + 357, + 545, + 457 + ], + "type": "text", + "content": " are created in the above manner on which the student network of the first LG is trained. Our final algorithm is summarized in Algorithm 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 461, + 387, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 461, + 387, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 461, + 387, + 475 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 481, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 545, + 554 + ], + "type": "text", + "content": "Evaluation has been carried out following [19, 20] to allow direct comparison. As in [20], the labelled data is augmented with synthetic label noise. Performance is measured with, and without an unlabelled dataset to illuminate performance differences between the weakly and semi supervised settings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 560, + 440, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 560, + 440, + 573 + ], + "spans": [ + { + "bbox": [ + 306, + 560, + 440, + 573 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 578, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 687 + ], + "type": "text", + "content": "We utilize R50 [9] and PVTv2-B2 [33] as the backbones in the two LGs to encourage diversity. Both the teacher and student models are based on DeeplabV3+ [4]. For inference, we use the average of the predictions from the two teacher networks as in [20]. The " + }, + { + "bbox": [ + 304, + 578, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 578, + 545, + 687 + ], + "type": "text", + "content": " in Eq. (5) is set to 0.1 to balance the contribution from the CE and KL terms in our loss. We randomly sample 100 pixels per class in each image to get the representative dataset " + }, + { + "bbox": [ + 304, + 578, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\bar{D}" + }, + { + "bbox": [ + 304, + 578, + 545, + 687 + ], + "type": "text", + "content": ". The hyperparameter " + }, + { + "bbox": [ + 304, + 578, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 578, + 545, + 687 + ], + "type": "text", + "content": " in Eq. (6) is set to be 10 in our experiments. Other" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 714 + ], + "type": "text", + "content": "2With a slight abuse of notation, this operation indicates marking all pixels in " + }, + { + "bbox": [ + 306, + 693, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_k^1" + }, + { + "bbox": [ + 306, + 693, + 545, + 714 + ], + "type": "text", + "content": " with \"ignore-pixel\" label in " + }, + { + "bbox": [ + 306, + 693, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L" + }, + { + "bbox": [ + 306, + 693, + 545, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "text", + "content": "1233" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 129, + 550, + 353 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 120 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 120 + ], + "type": "text", + "content": "Table 1: Comparison against SotA approaches on Pascal VOC 2012 in a semi-supervised semantic segmentation setting. All baselines are based on the DeeplabV3+ architecture. The * indicates results reported by [6]. When we replace " + }, + { + "bbox": [ + 46, + 70, + 547, + 120 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 70, + 547, + 120 + ], + "type": "text", + "content": " of the ground truth of images with 'PL', the proportion of noisy pixels is " + }, + { + "bbox": [ + 46, + 70, + 547, + 120 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 46, + 70, + 547, + 120 + ], + "type": "text", + "content": ". We keep the same pixel level noise proportion while introducing noise using 'RDE' and 'SCP' for parity." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 129, + 550, + 353 + ], + "lines": [ + { + "bbox": [ + 47, + 129, + 550, + 353 + ], + "spans": [ + { + "bbox": [ + 47, + 129, + 550, + 353 + ], + "type": "table", + "html": "
MethodsYearNoisy pixelNoise typeLabelled Data Ratio
1/16 (662)1/8(1323)1/4 (2646)1/2 (5291)
MT* [30]20170%None66.7070.7873.2275.41
French [8]20190%None68.9070.7072.4674.49
CCT* [26]20200%None65.2270.8773.4374.75
GCT* [14]20200%None64.0570.4773.4575.20
ECS* [23]20200%None-67.3870.7072.89
CPS* [6]20210%None71.9873.6774.9076.15
CAC* [16]20210%None70.1072.4074.00-
PS-MT [20]20220%None72.8375.7076.4377.88
Ours-0%None77.75 (+4.92)79.31 (+3.61)79.14 (+2.71)79.54 (+1.66)
PS-MT [20]-9%PL61.9065.1465.5366.78
Ours-9%PL71.60 (+9.70)73.95 (+8.81)74.53 (+9.00)74.44 (+7.66)
PS-MT [20]-9%RDE58.8762.1863.8163.46
Ours-9%RDE66.82 (+7.95)69.95 (+7.77)73.82 (+10.01)74.47 (+11.01)
PS-MT [20]-9%SCP52.1457.1956.8153.69
Ours-9%SCP67.80 (+15.66)70.83 (+13.64)69.38 (+12.57)75.95 (+22.26)
", + "image_path": "5f58705d0981877f2403b48916a66bbf44e65d2cd4a9f7f68fcb4a55fe2c0d25.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 380, + 288, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 288, + 441 + ], + "type": "text", + "content": "training and inference settings are the same as those of the baseline method [20]. In all of our experiments, the last epoch checkpoint model is used unless otherwise specified. Further details of the implementation can be found in the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 449, + 259, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 259, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 259, + 462 + ], + "type": "text", + "content": "5.2. Semi-supervised Semantic Segmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "type": "text", + "content": "Pascal VOC 2012 [7] is employed for evaluating the performance of our model in a semi-supervised semantic segmentation setting. This dataset contains more than " + }, + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "type": "inline_equation", + "content": "13\\mathrm{k}" + }, + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "type": "text", + "content": " images with 21 different classes. To introduce realistic noise in " + }, + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L" + }, + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "type": "text", + "content": ", the following three approaches are considered:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 536, + 288, + 685 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 58, + 536, + 287, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 536, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 58, + 536, + 287, + 572 + ], + "type": "text", + "content": "- Pseudo Labels (PL). Following [19], a portion of clean labels are replaced by pseudo-labels from a weakly-supervised segmentation method [34]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 574, + 287, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 574, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 58, + 574, + 287, + 610 + ], + "type": "text", + "content": "- Random Dilation and Erosion (RDE). Similar to [19, 21], we apply random degrees of dilation and erosion to the ground-truth masks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 613, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 613, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 58, + 613, + 288, + 685 + ], + "type": "text", + "content": "- Similar Class Perturbation (SCP). We introduce another type of noise by randomly replacing ground-truth label of a segmentation mask with a different class that resembles the texture of original clean class. This label perturbation simulates a more natural scenario pertaining to annotators confusing semantics." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 318, + 379, + 533, + 498 + ], + "blocks": [ + { + "bbox": [ + 318, + 379, + 533, + 498 + ], + "lines": [ + { + "bbox": [ + 318, + 379, + 533, + 498 + ], + "spans": [ + { + "bbox": [ + 318, + 379, + 533, + 498 + ], + "type": "image", + "image_path": "258e736694f9fbb32fe84882b00d6438ecedc057edcbc0aa017b6d9439ac6b81.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 501, + 545, + 562 + ], + "lines": [ + { + "bbox": [ + 304, + 501, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 545, + 562 + ], + "type": "text", + "content": "Figure 5: A performance comparison under different noise proportions achieved by replacing (from " + }, + { + "bbox": [ + 304, + 501, + 545, + 562 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 304, + 501, + 545, + 562 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 501, + 545, + 562 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 501, + 545, + 562 + ], + "type": "text", + "content": ") clean segmentation masks with the PL from [34] at image-level. The green font shows the corresponding noisy pixel proportions. All experiments are conducted with " + }, + { + "bbox": [ + 304, + 501, + 545, + 562 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 304, + 501, + 545, + 562 + ], + "type": "text", + "content": " labelled data." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 578, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 604 + ], + "type": "text", + "content": "Details about the implementation and visualization of noise generation can be found in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "In the context of semi-supervised semantic segmentation, we have compared our approach with the following studies: MT [30], French [8], CCT [26], GCT [14], ECS [23], CPS [6], CAC [16], and PS-MT [20]. Following the previous studies [19, 20], we report the mean Intersection-over-Union (mIoU) on the validation set. Table 1 shows that our approach ('Ours') outperforms all baselines trained with clean labelled data. In the presence of noise, we compare the best performing semi-supervised segmentation method" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 288, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 288, + 713 + ], + "type": "text", + "content": "3Note that, SCP does not change the shape of the clean segmentation masks, but only labels may be switched." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "text", + "content": "1234" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 128, + 273, + 185 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 120 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 120 + ], + "type": "text", + "content": "Table 2: The mIoU of the weakly-supervised semantic segmentation baselines and ours on the test set of SegTHOR. Best Val and Last Epoch correspond to the checkpoint that performs best on the validation set and the last checkpoint." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 128, + 273, + 185 + ], + "lines": [ + { + "bbox": [ + 60, + 128, + 273, + 185 + ], + "spans": [ + { + "bbox": [ + 60, + 128, + 273, + 185 + ], + "type": "table", + "html": "
ModelBest ValLast Epoch
Base-ADELE [19]62.659.1
ADELE [19]71.170.8
Ours73.273.5
", + "image_path": "2983e5d01764c9b7d9a2f4b7dd06e285ccb1fa7b63366e193750e5d925a6466a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 198, + 287, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 294 + ], + "type": "text", + "content": "with no noise, i.e., PS-MT [20], against our method with varying noise types (see Table 1) and noise proportions (see Fig. 5). The results demonstrate significant improvements in mIoU in all cases, e.g., " + }, + { + "bbox": [ + 46, + 198, + 287, + 294 + ], + "type": "inline_equation", + "content": "+22.26\\%" + }, + { + "bbox": [ + 46, + 198, + 287, + 294 + ], + "type": "text", + "content": " with half of the labelled data on the more practical SCP noise. Overall, the superiority of our method is consistent across all labelled data to unlabelled data ratios, regardless of the presence of noise or the type of applied noise." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 301, + 271, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 301, + 271, + 315 + ], + "spans": [ + { + "bbox": [ + 47, + 301, + 271, + 315 + ], + "type": "text", + "content": "5.3. Weakly-supervised Semantic Segmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 319, + 287, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 368 + ], + "type": "text", + "content": "We evaluate our method in a weakly-supervised semantic segmentation setting on the following datasets: (i) SegTHOR [17], a medical dataset, and (ii) pseudo labelled Pascal VOC 2012." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 373, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 504 + ], + "type": "text", + "content": "Medical image segmentation on SegTHOR. Following [19], we evaluate our approach on the SegTHOR dataset, which consists of 3D CT scans. Each pixel of these images is assigned to one of these five classes: 'esophagus', 'heart', 'trachea', 'aorta', or 'background'. The slices from the images are randomly split into a training set of 3638 slices, a validation set of 570 slices, and a test set of 580 slices [19, 17]. To introduce noise, we apply 'RDE' to the ground-truth segmentation labels, similar to [19]. Note that all annotations in the training set are corrupted by this noise, whereas the validation and test sets remain uncorrupted." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 505, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 287, + 613 + ], + "type": "text", + "content": "We compare our approach with ADELE [19], which corrects noisy annotations with model predictions in the early training stage, and its variant, Base-ADELE [19], which is based on a UNet trained with multi-scale inputs. Detailed implementations of these approaches are in the supplementary material. Following the setting in ADELE [19], we evaluate two checkpoints: specifically, the one with the highest mIoU on the validation set and the last checkpoint. In Table 2, our approach outperforms both baselines." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "Weakly-supervised segmentation on Pascal VOC 2012. We demonstrate the effectiveness of our approach in comparison to the baselines in the weakly-supervised setting on the PASCAL VOC 2012 [7]. Similar to ADELE [19], we use only noisy pseudo labels generated by [34] to train the models and evaluate on the original test set. Table 3 shows that our approach outperforms recent baselines, such as ADELE combined with SEAM [19]. In addition, PS-MT" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "content": "increases the mIoU marginally with PVTv2 compared to R50 with 66.7 vs. 64.0. However, after using our framework, the mIoU is further improved from 66.7 to 71.7, clearly demonstrating the effectiveness of our method." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 176, + 536, + 305 + ], + "blocks": [ + { + "bbox": [ + 304, + 131, + 545, + 167 + ], + "lines": [ + { + "bbox": [ + 304, + 131, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 131, + 545, + 167 + ], + "type": "text", + "content": "Table 3: Comparison with state-of-the-art weakly supervised semantic segmentation methods on the Pascal VOC 2012 dataset using mIoU (\\%) of the validation set." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 176, + 536, + 305 + ], + "lines": [ + { + "bbox": [ + 315, + 176, + 536, + 305 + ], + "spans": [ + { + "bbox": [ + 315, + 176, + 536, + 305 + ], + "type": "table", + "html": "
ModelYearmIoU (%)
AffinityNet [1]201861.7
SSDD [27]201964.9
SEAM [34]202064.5
CONTA [37]202066.1
SPML [13]202169.5
ADELE [19] + SEAM202269.3
PS-MT [20] (R50)202264.0
PS-MT [20] (PVTv2-B2)202266.7
Ours-71.7
", + "image_path": "515410c5c936a50758f0de1d6c0e88e2079c05fc331648458f796ca8ebb56334.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 326, + 399, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 326, + 399, + 338 + ], + "spans": [ + { + "bbox": [ + 306, + 326, + 399, + 338 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 344, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 545, + 392 + ], + "type": "text", + "content": "All experiments in this section are conducted with " + }, + { + "bbox": [ + 304, + 344, + 545, + 392 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 304, + 344, + 545, + 392 + ], + "type": "text", + "content": " labelled data ratio and " + }, + { + "bbox": [ + 304, + 344, + 545, + 392 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 304, + 344, + 545, + 392 + ], + "type": "text", + "content": " PL noisy annotations on Pascal VOC. We verify the effectiveness of the two main components in our approach, the diverse LGs and filter module." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 393, + 546, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 546, + 560 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 546, + 560 + ], + "type": "text", + "content": "Diversity in the LGs with different architectures. In contrast to previous works [20, 18], we found that using different network architectures in the two LGs provides more complementary information. In Fig. 3, we show the qualitative results of noise detection when we use the same and different architectures in the two LGs. The two LGs provide more diverse noise detection results (lower overlap between predicted noisy labels) when we employ different architectures. In Fig. 6, the predictions from the two LGs with two different architectures " + }, + { + "bbox": [ + 304, + 393, + 546, + 560 + ], + "type": "inline_equation", + "content": "(\\mathrm{e})\\& (\\mathrm{f})" + }, + { + "bbox": [ + 304, + 393, + 546, + 560 + ], + "type": "text", + "content": " are more complementary than those from the two LGs with the same architecture " + }, + { + "bbox": [ + 304, + 393, + 546, + 560 + ], + "type": "inline_equation", + "content": "(\\mathrm{(c)})\\& (\\mathrm{(d)})" + }, + { + "bbox": [ + 304, + 393, + 546, + 560 + ], + "type": "text", + "content": ". Fig. 7 shows the mIoU of our pipeline when we use different network combinations. The combination of R50 and PVTv2-B2 [33] achieves the best performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 562, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 562, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 562, + 545, + 646 + ], + "type": "text", + "content": "Noise detection with different LGs. To avoid confirmation bias, we utilize the teacher network from one LG to detect noise for the other LG. If we utilize the teacher network from the same LG for noise detection (i.e., replacing " + }, + { + "bbox": [ + 304, + 562, + 545, + 646 + ], + "type": "inline_equation", + "content": "\\theta_{s^1}" + }, + { + "bbox": [ + 304, + 562, + 545, + 646 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 562, + 545, + 646 + ], + "type": "inline_equation", + "content": "\\theta_{s^2}" + }, + { + "bbox": [ + 304, + 562, + 545, + 646 + ], + "type": "text", + "content": " in Eq. (4)), the mIoU drops from " + }, + { + "bbox": [ + 304, + 562, + 545, + 646 + ], + "type": "inline_equation", + "content": "74.0\\%" + }, + { + "bbox": [ + 304, + 562, + 545, + 646 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 562, + 545, + 646 + ], + "type": "inline_equation", + "content": "72.4\\%" + }, + { + "bbox": [ + 304, + 562, + 545, + 646 + ], + "type": "text", + "content": " on Pascal VOC, demonstrating the importance of mutual noise detection with a different LG in our pipeline." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "text", + "content": "Significance of the filter module. In Table 4, in addition to the baseline PS-MT and ours, we evaluate the performance of the following models: 1) 'PS-MT + FM' by adding the filter module to PS-MT (using R50)4. 2) 'Ours - FM' by" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 545, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 545, + 714 + ], + "type": "text", + "content": "4Note, PS-MT can be regarded as single LG (Fig. 2 (a)) and the mean" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "text", + "content": "1235" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 137, + 159 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 137, + 159 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 137, + 159 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 137, + 159 + ], + "type": "image", + "image_path": "6a3db4253c085eb04cdbfe4fe6c83696905a8191a60ac4f6d3fb283aa39bd780.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 161, + 124, + 171 + ], + "lines": [ + { + "bbox": [ + 63, + 161, + 124, + 171 + ], + "spans": [ + { + "bbox": [ + 63, + 161, + 124, + 171 + ], + "type": "text", + "content": "(a) Input Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 137, + 72, + 216, + 159 + ], + "blocks": [ + { + "bbox": [ + 137, + 72, + 216, + 159 + ], + "lines": [ + { + "bbox": [ + 137, + 72, + 216, + 159 + ], + "spans": [ + { + "bbox": [ + 137, + 72, + 216, + 159 + ], + "type": "image", + "image_path": "c3b5a11eede915f6419485aea4029dc925796f7fc31657df27010a2bab5e27dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 161, + 206, + 171 + ], + "lines": [ + { + "bbox": [ + 140, + 161, + 206, + 171 + ], + "spans": [ + { + "bbox": [ + 140, + 161, + 206, + 171 + ], + "type": "text", + "content": "(b) Ground Truth" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 217, + 72, + 296, + 159 + ], + "blocks": [ + { + "bbox": [ + 217, + 72, + 296, + 159 + ], + "lines": [ + { + "bbox": [ + 217, + 72, + 296, + 159 + ], + "spans": [ + { + "bbox": [ + 217, + 72, + 296, + 159 + ], + "type": "image", + "image_path": "8c003e5fd52937d8a5ffab945a6e83b638ae187b59ec986f7dc6b5cf9f29ce0a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 161, + 296, + 171 + ], + "lines": [ + { + "bbox": [ + 216, + 161, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 216, + 161, + 296, + 171 + ], + "type": "text", + "content": "(c) R50LG1 Prediction" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 298, + 72, + 376, + 159 + ], + "blocks": [ + { + "bbox": [ + 298, + 72, + 376, + 159 + ], + "lines": [ + { + "bbox": [ + 298, + 72, + 376, + 159 + ], + "spans": [ + { + "bbox": [ + 298, + 72, + 376, + 159 + ], + "type": "image", + "image_path": "6b9b8830d77bc507268576a0cb02b8bcea6acd9c924c76e1472c5f479bfbdead.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 297, + 161, + 376, + 171 + ], + "lines": [ + { + "bbox": [ + 297, + 161, + 376, + 171 + ], + "spans": [ + { + "bbox": [ + 297, + 161, + 376, + 171 + ], + "type": "text", + "content": "(d) R50LG2 Prediction" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 377, + 72, + 457, + 159 + ], + "blocks": [ + { + "bbox": [ + 377, + 72, + 457, + 159 + ], + "lines": [ + { + "bbox": [ + 377, + 72, + 457, + 159 + ], + "spans": [ + { + "bbox": [ + 377, + 72, + 457, + 159 + ], + "type": "image", + "image_path": "5d9c76a0db3e7ec9963e96ec63cf1d93271eec4d13f121f3acb6e588de4cc440.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 161, + 456, + 171 + ], + "lines": [ + { + "bbox": [ + 376, + 161, + 456, + 171 + ], + "spans": [ + { + "bbox": [ + 376, + 161, + 456, + 171 + ], + "type": "text", + "content": "(e) R50LG1 Prediction" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 456, + 161, + 538, + 171 + ], + "lines": [ + { + "bbox": [ + 456, + 161, + 538, + 171 + ], + "spans": [ + { + "bbox": [ + 456, + 161, + 538, + 171 + ], + "type": "text", + "content": "(f) PVT" + }, + { + "bbox": [ + 456, + 161, + 538, + 171 + ], + "type": "inline_equation", + "content": "^{\\mathrm{LG2}}" + }, + { + "bbox": [ + 456, + 161, + 538, + 171 + ], + "type": "text", + "content": " Prediction" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 457, + 72, + 538, + 159 + ], + "blocks": [ + { + "bbox": [ + 457, + 72, + 538, + 159 + ], + "lines": [ + { + "bbox": [ + 457, + 72, + 538, + 159 + ], + "spans": [ + { + "bbox": [ + 457, + 72, + 538, + 159 + ], + "type": "image", + "image_path": "8460ca2812ce4034caad854523fae174e22b50c336a51cfb945117f94ae98073.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 58, + 233, + 277, + 306 + ], + "blocks": [ + { + "bbox": [ + 53, + 213, + 277, + 225 + ], + "lines": [ + { + "bbox": [ + 53, + 213, + 277, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 277, + 225 + ], + "type": "text", + "content": "Table 4: Ablation study of FM and LG on Pascal VOC." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 233, + 277, + 306 + ], + "lines": [ + { + "bbox": [ + 58, + 233, + 277, + 306 + ], + "spans": [ + { + "bbox": [ + 58, + 233, + 277, + 306 + ], + "type": "table", + "html": "
Method2 LGsFilter Module (FM)mIoU (%)
PS-MT [20]65.1
PS-MT + FM✓67.6 (+2.5)
Ours - FM✓69.9 (+4.8)
Ours✓✓74.0 (+8.9)
", + "image_path": "735f67257470922de3132b40b16bd722951ddc80d3ed1afc5b1ab0b186d69b27.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 50, + 319, + 161, + 407 + ], + "blocks": [ + { + "bbox": [ + 50, + 319, + 161, + 407 + ], + "lines": [ + { + "bbox": [ + 50, + 319, + 161, + 407 + ], + "spans": [ + { + "bbox": [ + 50, + 319, + 161, + 407 + ], + "type": "image", + "image_path": "bc98f6d422e6c3b0ebd1df1df6d9fe07ddcdf6443f7957bc070b2247c618456e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 407, + 146, + 415 + ], + "lines": [ + { + "bbox": [ + 59, + 407, + 146, + 415 + ], + "spans": [ + { + "bbox": [ + 59, + 407, + 146, + 415 + ], + "type": "text", + "content": "(a) Different Architectures" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 422, + 287, + 518 + ], + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 518 + ], + "type": "text", + "content": "Figure 7: (a) shows the mIoU score of different network combinations on clean data. (b) shows the mIoU score of different combinations on different settings. R, S and P denote R50 [9], Segformer [35] and PVTv2-B2 [33] respectively. -FM_C denotes training our model (without the filter module) with clean data. -FM_N denotes training model with noisy data without using filter module. +FM_N is our complete model trained with noisy data." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 163, + 319, + 286, + 407 + ], + "blocks": [ + { + "bbox": [ + 163, + 319, + 286, + 407 + ], + "lines": [ + { + "bbox": [ + 163, + 319, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 163, + 319, + 286, + 407 + ], + "type": "image", + "image_path": "e94072a5805a4bebca868e757affaedee2b0acacfdbca7dcaccfb246204a5c09.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 407, + 258, + 415 + ], + "lines": [ + { + "bbox": [ + 174, + 407, + 258, + 415 + ], + "spans": [ + { + "bbox": [ + 174, + 407, + 258, + 415 + ], + "type": "text", + "content": "(b) Clean and Noisy Data" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 334, + 545, + 417 + ], + "lines": [ + { + "bbox": [ + 304, + 334, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 545, + 417 + ], + "type": "text", + "content": "Figure 8: (a) shows detected noise proportion of GMM [18] and ours, where " + }, + { + "bbox": [ + 304, + 334, + 545, + 417 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 334, + 545, + 417 + ], + "type": "text", + "content": "-axis shows the noisy image ratio and " + }, + { + "bbox": [ + 304, + 334, + 545, + 417 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 334, + 545, + 417 + ], + "type": "text", + "content": "-axis shows the detected noisy pixel ratio. The detected noise proportion of ours is closer to the GT than GMM. (b) shows the F1 score of noise detection with different criterion. CE is cross entropy loss; KL is KL divergence between the predictions from the two LGs. Our " + }, + { + "bbox": [ + 304, + 334, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\mathrm{CE} + \\mathrm{KL}" + }, + { + "bbox": [ + 304, + 334, + 545, + 417 + ], + "type": "text", + "content": " performs the best." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 533, + 286, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 286, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 286, + 652 + ], + "type": "text", + "content": "removing the filter module from our framework. It can be seen that the filter module helps to improve the final performance for both models. The filter module contributes more significantly (+4.1) to our framework due to the diverse learning groups (i.e., using 2 LGs with different architectures). This superiority of our framework is further evidenced by the better performance of 'Ours - FM' compared to PS-MT variants. In Fig. 7 (b), our Filter Module helps to improve the performance under noisy annotations ('+FM_N' vs. '-FM_N') with different architectures." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 655, + 286, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 655, + 286, + 691 + ], + "spans": [ + { + "bbox": [ + 46, + 655, + 286, + 691 + ], + "type": "text", + "content": "Comparison of the filter module with GMM. Previous work [18] utilizes Gaussian Mixture Model (GMM) to detect noise for image classification. We compare our filter" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 315, + 213, + 433, + 316 + ], + "blocks": [ + { + "bbox": [ + 46, + 178, + 547, + 202 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 547, + 202 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 547, + 202 + ], + "type": "text", + "content": "Figure 6: Visualizations from the two LGs when we use the same architecture ((c) and (d)) and different architectures ((e) and (f)). The " + }, + { + "bbox": [ + 46, + 178, + 547, + 202 + ], + "type": "inline_equation", + "content": "1st" + }, + { + "bbox": [ + 46, + 178, + 547, + 202 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 178, + 547, + 202 + ], + "type": "inline_equation", + "content": "2nd" + }, + { + "bbox": [ + 46, + 178, + 547, + 202 + ], + "type": "text", + "content": " columns show the inputs and the ground truth. R50 and PVT denote R50 and PVTv2-B2 [33]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 213, + 433, + 316 + ], + "lines": [ + { + "bbox": [ + 315, + 213, + 433, + 316 + ], + "spans": [ + { + "bbox": [ + 315, + 213, + 433, + 316 + ], + "type": "image", + "image_path": "e09b7a2ab53bb65aa0c01d31ed3520279e4011d22e393a5e93d2b1ab14654628.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 316, + 419, + 324 + ], + "lines": [ + { + "bbox": [ + 334, + 316, + 419, + 324 + ], + "spans": [ + { + "bbox": [ + 334, + 316, + 419, + 324 + ], + "type": "text", + "content": "(a) Detected Noise Ratio" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 432, + 316, + 539, + 325 + ], + "lines": [ + { + "bbox": [ + 432, + 316, + 539, + 325 + ], + "spans": [ + { + "bbox": [ + 432, + 316, + 539, + 325 + ], + "type": "text", + "content": "(b) F1 Score of Noise Detection" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 436, + 213, + 517, + 316 + ], + "blocks": [ + { + "bbox": [ + 436, + 213, + 517, + 316 + ], + "lines": [ + { + "bbox": [ + 436, + 213, + 517, + 316 + ], + "spans": [ + { + "bbox": [ + 436, + 213, + 517, + 316 + ], + "type": "image", + "image_path": "eda195d2494cfea237ab1c65eec41cd4ee24dfda863400c1cf98d8a3569b300a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 431, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 545, + 479 + ], + "type": "text", + "content": "module with GMM (including two components) on the representative dataset. Fig. 8 (a) shows the detected noise proportions. We can see that the detected noise proportion of our filter module is closer to the groundtruth than GMM." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 481, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 545, + 541 + ], + "type": "text", + "content": "Criterion to compute score. We utilize cross entropy loss and KL divergence " + }, + { + "bbox": [ + 304, + 481, + 545, + 541 + ], + "type": "inline_equation", + "content": "(\\mathrm{CE} + \\mathrm{KL})" + }, + { + "bbox": [ + 304, + 481, + 545, + 541 + ], + "type": "text", + "content": " to compute score in Eq. (5). Here, we compare this criterion with others, including only cross entropy loss (CE), and only KL divergence (KL). Fig. 8 (b) shows that ours achieves the best F1 score." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 542, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 545, + 601 + ], + "type": "text", + "content": "Memory and computation costs. PVTv2-B2 has similar GFLOPs (45.8) with R50 (45.6). Ours only have one more student network than PS-MT, which is not used during inference. Thus, our model needs similar memory and computation costs compared with PS-MT during inference." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 621, + 378, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 378, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 378, + 633 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "In this work, we propose a learning framework for semi-supervised semantic segmentation involving two diverse learning groups that can handle various labelling errors. Our framework employs a carefully designed filter module to divide the labelled dataset into pixels with and without label noise in a scalable manner. The potentially misla" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 703, + 175, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 703, + 175, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 703, + 175, + 712 + ], + "type": "text", + "content": "teacher output is used for filter module." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 296, + 733, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 733, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 296, + 733, + 315, + 742 + ], + "type": "text", + "content": "1236" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "belled pixels filtered by the first learning group are utilised for training the second learning group and vice versa. We show that our proposed framework provides state-of-the-art results for semi-supervised learning, by addressing confirmation bias while training, that can handle various amounts of annotation noise to accommodate practical inaccuracies in annotations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 165, + 107, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 165, + 107, + 178 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 107, + 178 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 185, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 185, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 185, + 287, + 239 + ], + "type": "text", + "content": "[1] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4981-4990, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 240, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 240, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 240, + 288, + 293 + ], + "type": "text", + "content": "[2] Eric Arazo, Diego Ortego, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 295, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 287, + 339 + ], + "type": "text", + "content": "[3] Liyi Chen, Weiwei Wu, Chenchen Fu, Xiao Han, and Yuntao Zhang. Weakly supervised semantic segmentation with boundary exploration. In European Conference on Computer Vision, pages 347-362. Springer, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 340, + 287, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 287, + 393 + ], + "type": "text", + "content": "[4] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 394, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 394, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 53, + 394, + 287, + 437 + ], + "type": "text", + "content": "[5] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 438, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 438, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 438, + 287, + 491 + ], + "type": "text", + "content": "[6] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 493, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 493, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 53, + 493, + 287, + 536 + ], + "type": "text", + "content": "[7] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 537, + 287, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 537, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 53, + 537, + 287, + 570 + ], + "type": "text", + "content": "[8] Geoff French, Timo Aila, Samuli Laine, Michal Mackiewicz, and Graham Finlayson. Semi-supervised semantic segmentation needs strong, high-dimensional perturbations. 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 571, + 287, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 571, + 287, + 614 + ], + "spans": [ + { + "bbox": [ + 53, + 571, + 287, + 614 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 615, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 615, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 615, + 287, + 657 + ], + "type": "text", + "content": "[10] Wei-Chih Hung, Yi-Hsuan Tsai, Yan-Ting Liou, Yen-Yu Lin, and Ming-Hsuan Yang. Adversarial learning for semi-supervised semantic segmentation. arXiv preprint arXiv:1802.07934, 2018." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[11] Mostafa S Ibrahim, Arash Vahdat, Mani Ranjbar, and William G Macready. Semi-supervised semantic image segmentation with self-correcting networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12715-12725, 2020." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[12] Joseph Jacob, Olga Ciccarelli, Frederik Barkhof, and Daniel C Alexander. Disentangling human error from the ground truth in segmentation of medical images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[13] Tsung-Wei Ke, Jyh-Jing Hwang, and Stella X Yu. Universal weakly supervised segmentation by pixel-to-segment contrastive learning. arXiv preprint arXiv:2105.00957, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 143, + 545, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 545, + 187 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 545, + 187 + ], + "type": "text", + "content": "[14] Zhanghan Ke, Di Qiu, Kaican Li, Qiong Yan, and Rynson WH Lau. Guided collaborative training for pixel-wise semi-supervised learning. In European conference on computer vision, pages 429-445. Springer, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 189, + 545, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 189, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 307, + 189, + 545, + 233 + ], + "type": "text", + "content": "[15] Donghyeon Kwon and Suha Kwak. Semi-supervised semantic segmentation with error localization network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9957-9967, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 236, + 545, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 236, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 307, + 236, + 545, + 290 + ], + "type": "text", + "content": "[16] Xin Lai, Zhuotao Tian, Li Jiang, Shu Liu, Hengshuang Zhao, Liwei Wang, and Jiaya Jia. Semi-supervised semantic segmentation with directional context-aware consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1205-1214, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 292, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 292, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 292, + 545, + 346 + ], + "type": "text", + "content": "[17] Zoé Lambert, Caroline Petitjean, Bernard Dubray, and Su Kuan. Segthor: segmentation of thoracic organs at risk in ct images. In 2020 Tenth International Conference on Image Processing Theory, Tools and Applications (IPTA), pages 1-6. IEEE, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 350, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 382 + ], + "type": "text", + "content": "[18] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 384, + 545, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 384, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 307, + 384, + 545, + 439 + ], + "type": "text", + "content": "[19] Sheng Liu, Kangning Liu, Weicheng Zhu, Yiqiu Shen, and Carlos Fernandez-Granda. Adaptive early-learning correction for segmentation from noisy annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2606–2616, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 441, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 441, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 545, + 506 + ], + "type": "text", + "content": "[20] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4258-4267, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 509, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 509, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 509, + 545, + 553 + ], + "type": "text", + "content": "[21] Yaoru Luo, Guole Liu, Yuanhao Guo, and Ge Yang. Elucidating meta-structures of noisy labels in semantic segmentation by deep neural networks. arXiv preprint arXiv:2205.00160, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 555, + 545, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 598 + ], + "type": "text", + "content": "[22] Yaoru Luo, Guole Liu, Wenjing Li, Yuanhao Guo, and Ge Yang. Deep neural networks learn meta-structures to segment fluorescence microscopy images. arXiv preprint arXiv:2103.11594, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "type": "text", + "content": "[23] Robert Mendel, Luis Antonio de Souza, David Rauber, Joao Paulo Papa, and Christoph Palm. Semi-supervised segmentation based on error-correcting supervision. In European Conference on Computer Vision, pages 141-157. Springer, 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[24] Shaobo Min, Xuejin Chen, Zheng-Jun Zha, Feng Wu, and Yongdong Zhang. A two-stream mutual attention network for semi-supervised biomedical segmentation with noisy labels. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 4578-4585, 2019." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "text", + "content": "1237" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[25] Sudhanshu Mittal, Maxim Tatarchenko, and Thomas Brox. Semi-supervised semantic segmentation with high-and low-level consistency. IEEE transactions on pattern analysis and machine intelligence, 43(4):1369-1379, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "type": "text", + "content": "[26] Yassine Ouali, Céline Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross-consistency training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12674-12684, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "text", + "content": "[27] Wataru Shimoda and Keiji Yanai. Self-supervised difference detection for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5208-5217, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 275 + ], + "type": "text", + "content": "[28] Yucheng Shu, Xiao Wu, and Weisheng Li. Lvc-net: Medical image segmentation with noisy label based on local visual cues. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 558-566. Springer, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "type": "text", + "content": "[29] Nasim Souly, Concetto Spampinato, and Mubarak Shah. Semi supervised semantic segmentation using generative adversarial network. In Proceedings of the IEEE international conference on computer vision, pages 5688-5696, 2017." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 365 + ], + "type": "text", + "content": "[30] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 366, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 366, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 287, + 420 + ], + "type": "text", + "content": "[31] Luis Garcia Ugarriza, Eli Saber, Sreenath Rao Vantaram, Vincent Amuso, Mark Shaw, and Ranjit Bhaskar. Automatic image segmentation by dynamic region growth and multiresolution merging. IEEE transactions on image processing, 18(10):2275-2288, 2009." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 487 + ], + "type": "text", + "content": "[32] Guotai Wang, Xinglong Liu, Chaoping Li, Zhiyong Xu, Jiugen Ruan, Haifeng Zhu, Tao Meng, Kang Li, Ning Huang, and Shaoting Zhang. A noise-robust framework for automatic segmentation of Covid-19 pneumonia lesions from ct images. IEEE Transactions on Medical Imaging, 39(8):2653-2663, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 488, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 287, + 533 + ], + "type": "text", + "content": "[33] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 534, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 589 + ], + "type": "text", + "content": "[34] Yude Wang, Jie Zhang, Meina Kan, Shiguang Shan, and Xilin Chen. Self-supervised equivariant attention mechanism for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12275-12284, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "type": "text", + "content": "[35] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 646, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 690 + ], + "type": "text", + "content": "[36] Rumeng Yi, Yaping Huang, Qingji Guan, Mengyang Pu, and Runsheng Zhang. Learning from pixel-level label noise: A new perspective for semi-supervised semantic segmentation. IEEE Transactions on Image Processing, 31:623-635, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[37] Dong Zhang, Hanwang Zhang, Jinhui Tang, Xian-Sheng Hua, and Qianru Sun. Causal intervention for weakly-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 230 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "supervised semantic segmentation. Advances in Neural Information Processing Systems, 33:655-666, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 139 + ], + "type": "text", + "content": "[38] Zhaowei Zhu, Zihao Dong, and Yang Liu. Detecting corrupted labels without training a model to predict. In International Conference on Machine Learning, pages 27412-27427. PMLR, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 184 + ], + "type": "text", + "content": "[39] Zhaowei Zhu, Yiwen Song, and Yang Liu. Clusterability as an alternative to anchor points when learning with noisy labels. In International Conference on Machine Learning, pages 12912-12923. PMLR, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 186, + 547, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 230 + ], + "type": "text", + "content": "[40] Yuliang Zou, Zizhao Zhang, Han Zhang, Chun-Liang Li, Xiao Bian, Jia-Bin Huang, and Tomas Pfister. Pseudoseg: Designing pseudo labels for semantic segmentation. arXiv preprint arXiv:2010.09713, 2020." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "text", + "content": "1238" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction/487dfb22-6094-4f31-976c-310213635421_content_list.json b/2023/Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction/487dfb22-6094-4f31-976c-310213635421_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..17feac72463e3e70688bd80cbc8fb9f123ca4da0 --- /dev/null +++ b/2023/Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction/487dfb22-6094-4f31-976c-310213635421_content_list.json @@ -0,0 +1,1699 @@ +[ + { + "type": "text", + "text": "Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction", + "text_level": 1, + "bbox": [ + 181, + 130, + 785, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruochen Jiao", + "bbox": [ + 290, + 204, + 398, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Northwestern University", + "bbox": [ + 246, + 222, + 444, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ruochen.jiao@u.northwestern.edu", + "bbox": [ + 207, + 242, + 483, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiangguo Liu", + "bbox": [ + 593, + 204, + 707, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Northwestern University", + "bbox": [ + 552, + 222, + 748, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xg.liu@u.northwestern.edu", + "bbox": [ + 539, + 242, + 761, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Takami Sato", + "bbox": [ + 173, + 263, + 276, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of California, Irvine", + "bbox": [ + 102, + 282, + 348, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "takamis@uci.edu", + "bbox": [ + 156, + 303, + 292, + 315 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of California, Irvine", + "bbox": [ + 385, + 282, + 633, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "alfchen@uci.edu", + "bbox": [ + 442, + 303, + 576, + 315 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qi Zhu", + "bbox": [ + 738, + 265, + 799, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Northwestern University", + "bbox": [ + 671, + 282, + 867, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "qzhu@northwestern.edu", + "bbox": [ + 676, + 303, + 862, + 315 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 351, + 312, + 367 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Predicting the trajectories of surrounding objects is a critical task for self-driving vehicles and many other autonomous systems. Recent works demonstrate that adversarial attacks on trajectory prediction, where small crafted perturbations are introduced to history trajectories, may significantly mislead the prediction of future trajectories and induce unsafe planning. However, few works have addressed enhancing the robustness of this important safety-critical task. In this paper, we present a novel adversarial training method for trajectory prediction. Compared with typical adversarial training on image tasks, our work is challenged by more random input with rich context and a lack of class labels. To address these challenges, we propose a method based on a semi-supervised adversarial autoencoder, which models disentangled semantic features with domain knowledge and provides additional latent labels for the adversarial training. Extensive experiments with different types of attacks demonstrate that our Semi-supervised Semantics-guided Adversarial Training (SSAT1) method can effectively mitigate the impact of adversarial attacks by up to $73\\%$ and outperform other popular defense methods. In addition, experiments show that our method can significantly improve the system's robust generalization to unseen patterns of attacks. We believe that such semantics-guided architecture and advancement on robust generalization is an important step for developing robust prediction models and enabling safe decision making.", + "bbox": [ + 76, + 383, + 472, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 820, + 207, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Connected and autonomous vehicles (CAVs) have shown great promise to revolutionize the transportation systems,", + "bbox": [ + 76, + 845, + 468, + 877 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "but also face significant concerns regarding their safety and security [15, 36, 21, 17, 50, 51]. Many of these concerns, which also apply to many other autonomous systems, arise from the increasing adoption of advanced deep neural network (DNN)-based machine learning techniques across system sensing, perception, prediction, planning, control, and general decision-making [52, 53].", + "bbox": [ + 500, + 353, + 890, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In particular, the adversarial robustness of DNNs has drawn significant attention in recent years. State-of-the-art classifiers can be misled to make erroneous predictions by slight but carefully-optimized perturbations [39], and addressing this is critical to the safety of autonomous systems that employ these classifiers [11, 40]. Besides classification, adversarial attacks are also targeted at various other tasks such as object detection, image generation and natural language processing. To defend against adversarial examples, adversarial training is commonly used to enhance the intrinsic robustness of models and has been shown to be very effective among various defense strategies [24, 25, 34]. On the other hand, the works in [30, 43, 37] show that adversarial training may suffer poor robust generalization on unseen attacks and [33, 24, 38] demonstrate a trade-off between robustness and accuracy. In this paper, we address trajectory prediction, a safety-critical task that is common in self-driving vehicles and other autonomous systems, and propose a novel architecture to enhance its adversarial robustness and robust generalization by introducing semantic features and a semi-supervised adversarial autoencoder (AAE) [26] into the adversarial training process.", + "bbox": [ + 496, + 460, + 892, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We focus on the trajectory prediction module in autonomous driving, where large DNNs have enabled breakthroughs in recent years. More specifically, an autonomous driving system typically consists of several modules such as perception, localization, prediction, planning (route, behavior, trajectory, and motion planning), and control. The perception module is to detect agents and obstacles based", + "bbox": [ + 496, + 795, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Our codes: https://github.com/jrcblue/SSAT-for-Motion-Prediction", + "bbox": [ + 94, + 886, + 452, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8207", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "on the sensing inputs (e.g., images and 3D point clouds), and the effectiveness and robustness of the perception module are well-studied in the research community. The prediction module is to predict future trajectories of surrounding vehicles based on the observed history trajectories of those vehicles from the perception module and the map context. The trajectory prediction plays a crucial role in understanding the environment and making safety-critical decisions in the following planning module [18, 20, 19]. Recent works [23, 16, 5, 45, 12] have applied various deep learning techniques such as graph neural networks and transformers to this task and achieved impressive performance in terms of reducing average errors. However, few works studied the robustness of vehicle trajectory prediction under adversarial attacks, which is in fact of vital importance because 1) autonomous driving is a safety-critical task by nature, 2) recent work [48, 10] demonstrates that the prediction module is vulnerable to adversarial attacks if surrounding vehicles drive along a seemingly natural but crafted trajectory, and 3) current prediction models are often overfitted on limited patterns in the datasets but suffer the long-tailed distribution of driving scenarios and behaviors. The threat model of the adversarial trajectory is illustrated in an example in Fig. 1, where a crafted input trajectory can intentionally mislead the target vehicle to the wrong prediction and induce dangerous planning decisions.", + "bbox": [ + 76, + 90, + 472, + 484 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/54cf1db9e7e81c910006b20e1b071f79068f13807b7324360d4b129d527e4b56.jpg", + "image_caption": [ + "Figure 1: An illustrative example of the threat model: In the upper benign scenario, the target vehicle predicts the future trajectory of a surrounding vehicle based on its history trajectory. In the lower adversarial scenario, the surrounding vehicle is malicious and drives along a crafted (history) trajectory, which may mislead the target vehicle to have a wrong (attacked) prediction of the surrounding vehicle's future trajectory. In this particular case, the target vehicle wrongly predicts that the malicious vehicle will cut into its lane, and thus takes an unsafe emergency braking." + ], + "image_footnote": [], + "bbox": [ + 81, + 496, + 468, + 645 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To defend against adversarial attacks on vehicle trajectory prediction, there are important challenges that are different from the cases of images and audios. First, vehicle trajectory prediction is a time-series regression problem with rich contexts while most existing adversarial attacks", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and corresponding defense methods are targeted at classification tasks. The attack patterns are more random and there are no well-defined class labels, which means that the robust model is difficult to train and generalize. Second, vehicle trajectory can convey semantic and behavior information. For instance, people can infer the behavior of a vehicle such as moving forward or changing lanes from its trajectory. Therefore, as shown in our experiments later, some popular defense methods such as TRADES [46] and data augmentation methods [6, 29, 35] either are inapplicable or have poor performance in trajectory prediction task. In this work, we first propose an adversarial training pipeline for the trajectory prediction task, and then further exploit the semantic features to design a semi-supervised AAE architecture that can be added after the feature extractor, to improve the adversarial robustness and its generalization. The methodology of the proposed architecture such as enhancing disentanglement and defining semantic labels may be further applied to adversarial training for general regression and generation problems. Our main contributions are summarized as follows:", + "bbox": [ + 496, + 90, + 893, + 407 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel adversarial training method against adversarial attacks on trajectory predictions.", + "- We develop a semi-supervised architecture with domain knowledge and semantic features to enhance the adversarial robustness and its generalization among different types of attack patterns.", + "- Extensive experiments demonstrate that our proposed method (SSAT) effectively improves the adversarial robustness and outperforms popular defense baselines.", + "- We further explore to balance the robustness-accuracy trade-off in this task by leveraging the MixUp technique [47], and we also test data augmentation methods for trajectory prediction robustness." + ], + "bbox": [ + 517, + 417, + 890, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 656, + 638, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Adversarial Attacks on Trajectory Prediction", + "text_level": 1, + "bbox": [ + 500, + 681, + 883, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent research [48, 10] demonstrates that trajectory prediction in autonomous driving can be fooled by the adversarial behavior of a surrounding vehicle, where the adversarial behavior is optimized with Projected Gradient Descent (PGD) [24]. The threat model is illustrated in Fig. 1. In this work, we consider such attacks and assume the worst setting, i.e., the attacker has full knowledge of the target system and tries to maximize the attack impact. However, malicious agents cannot directly manipulate anything inside the target vehicle and they can only slightly change their own trajectories to mislead the target vehicle indirectly.", + "bbox": [ + 496, + 704, + 890, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Different from image classification, trajectory prediction has no class labels, but it has directional information in the", + "bbox": [ + 498, + 869, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8208", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/daac5439f6a180240c33f5df31b6ceb98f3b8cc786b363d794c8583f936b544d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 108, + 370, + 210 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c2e5e6a3d52e198afd9da362d85a221d25c792c26f6b407ffcca27f02778a530.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 108, + 566, + 210 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2122c7654d92294b38e41a5fe8a72d17a432d8fa8f453474b7c6cea828732a9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 108, + 764, + 210 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a3523bdf036540e85f87e7bd6a8520ba720374a4d2cca6b14741edd372de6b16.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 215, + 234, + 369, + 337 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/74c62a6c7f807384e4fd52f213824e25bb86c7094274909743643f05baecaf73.jpg", + "image_caption": [ + "(b)", + "Figure 2: Adversarial trajectories generated by different types of attacks. The figures in the top row are adversarial (red line) and benign (blue line) input history trajectories for prediction and they look very close to human eyes. The figures in the bottom row show the corresponding attacked future trajectory prediction (red dashed line) and the ground truth trajectory in the benign case (green dashed line). The differences between the two are clearly visible, showing the effectiveness of the attacks. Figure (a) is the ADE attack, which will randomly lead to maximum average deviation. Figure (b) is the lateral attack, which mainly leads the vehicle to deviate to the left or right. Figure (c) is the longitudinal attack, which will mainly lead to longitudinal deviation." + ], + "image_footnote": [], + "bbox": [ + 413, + 234, + 566, + 337 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a8b6ecb4f0eaa094f299713246f3e62c1e7f163d618ea6d2bf04098a0eb65b16.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 612, + 234, + 764, + 337 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "context, such as moving forward, turning or changing lanes to the right. Therefore, the attacker can conduct targeted attacks besides random ones. In Fig. 2, three types of adversarial attacks are presented and we observe that they can cause significant and directional errors. [48] proposes directional error metrics for the optimization of targeted attacks, as shown in Eq. (1):", + "bbox": [ + 75, + 508, + 468, + 614 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD (\\alpha , R) = \\left(p _ {\\alpha} - s _ {\\alpha}\\right) ^ {T} \\cdot R \\left(s _ {\\alpha + 1}, s _ {\\alpha}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 626, + 468, + 646 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha$ denotes the time frame ID. $p$ and $s$ are two-dimensional vectors representing predicted and ground-truth vehicle locations, respectively. $R$ is a function generating the unit vector to a specific direction (lateral or longitudinal). The longitudinal direction is approximated as the vector defined by the adjacent two waypoints of ground truth $s_{\\alpha +1}^{n} - s_{\\alpha}^{n}$", + "bbox": [ + 75, + 657, + 468, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to the directional attack, a random attack can be designed to maximize the Average Displacement Error (ADE), which is the average of the root mean squared error between the predicted waypoints and the ground-truth trajectory waypoints.", + "bbox": [ + 75, + 763, + 468, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hard constraints are also applied to the maximum deviation of waypoints during the optimization so as to make the adversarial history trajectory physically feasible and not perform unrealistic behaviors.", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/0c8a7115b5f11351741f610baf2ff029a72528835eb5ce116673b82499cb4e0a.jpg", + "table_caption": [ + "Table 1: Trajectory prediction errors (in meters) in benign or adversarial scenarios under different types of attacks." + ], + "table_footnote": [], + "table_body": "
ModelADELateralLongitudinal
benign/attackbenign/attackbenign/attack
LaneGCN [16]1.32/5.17-0.01/1.58-0.25/3.33
Trajectory++[31]2.69/6.810.107/2.25-0.526 / 3.79
", + "bbox": [ + 532, + 549, + 859, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1 quantitatively shows that the predicted trajectories can be seriously deviated by any of the three attack types. Generally, in the US, $0.3\\mathrm{m}$ lateral deviation is enough to invade adjacent lanes on local roads [32, 8]. In addition, the final displacement error (the distance between the last points of predicted and ground-truth trajectories) is around two to three times larger than the average error shown here. Therefore, the attacker can apply either random or directional attacks that will greatly challenge the defense methods, especially on the generalization performance.", + "bbox": [ + 496, + 646, + 890, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[48, 2] propose to utilize approaches such as smoothing and adversarial training on the conditional variational encoder (CVAE) architectures to mitigate the impact of such adversarial attacks. However, they do not explicitly address driving semantics and robustness generalization, which are critical for many real road scenarios.", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8209", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Adversarial Training Methods and Robust Generalization", + "text_level": 1, + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adversarial training [24, 25, 34] is shown to be one of the most effective approaches to improve the robustness of DNN models. In practice, the PGD attack is commonly deployed for evaluation because of its strong attack ability in white-box settings and the work in [24] formulates the adversarial training as a min-max problem. [6, 29, 35] show that synthesized data can significantly boost robustness for the image classification task. Thus, we select similar methods as baselines for comparison in our experiments.", + "bbox": [ + 76, + 128, + 468, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent research shows that a specific type of adversarial attack is not sufficient to represent the diverse space of adversarial examples and many adversarially trained models are only robust to specific attacks. This does limit the application of adversarial training in practice, especially for trajectory regression tasks due to its nature of long-tailed distribution. The works in [33, 49, 37] try to explain and improve the robust generalization in various perspectives such as sample complexity and latent feature representation. In this work, we demonstrate that robust generalization for trajectory prediction can be enhanced by explicitly introducing disentangled and semantic features in the latent space.", + "bbox": [ + 75, + 265, + 470, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3. Adversarial Autoencoder Architecture", + "text_level": 1, + "bbox": [ + 76, + 455, + 406, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The adversarial autoencoder (AAE) is a variant of the variational autoencoders (VAE) [14, 13], which provides a principled method for jointly learning deep latent-variable models and corresponding inference models using stochastic gradient descent. AAE imposes various distributions on the latent vector by utilizing adversarial learning instead of KL divergence. Due to the flexibility of adversarial learning, AAE is superior to VAE in terms of imposing complicated distributions over latent space. For adversarial robustness, the works in [27, 41] demonstrate that the disentangled latent representations produce VAEs that are more robust to adversarial attacks. In our work, we design an AAE-based architecture that can be added after the feature extractor of prediction modules. We utilize this architecture to model diverse semantic features and enhance disentanglement in the latent space.", + "bbox": [ + 75, + 479, + 468, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Proposed Adversarial Training Method for Trajectory Prediction", + "text_level": 1, + "bbox": [ + 76, + 734, + 467, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Domain Knowledge-guided Semi-supervised Architecture", + "text_level": 1, + "bbox": [ + 76, + 777, + 467, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.1 Overall Design", + "text_level": 1, + "bbox": [ + 76, + 816, + 233, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike adversarial training for image classification tasks, there are no class labels to guide the training for trajectory prediction, and the attack patterns in the trajectory regression are more random. We thus introduce domain knowl", + "bbox": [ + 76, + 839, + 467, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "edge to facilitate the modeling of the semantic information in both benign and adversarial cases, based on an AAE architecture. The model learns the directional semantic latent vector in a semi-supervised way because the ground truths are only available for limited scenarios but their distributions can be derived from domain knowledge and statistics. Therefore, our latent space modeling contains two levels – unsupervised distribution modeling and semi-supervised learning when labels are available.", + "bbox": [ + 500, + 90, + 890, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The architecture of our proposed model is shown in Fig. 3. The feature extractor [16] utilizes a one-dimensional dilated convolutional neural network to obtain the embedding of the time-series trajectory and uses a graph neural network to model the lane context and interaction between objects. The encoder maps high-dimensional features to the semantics-guided latent space with distribution regularization and semi-supervised training. In particular, the latent space is divided into three parts: longitudinal features $z_{lon}$ , following one-dimensional log-normal distribution, lateral features $z_{lat}$ , following three-dimensional categorical distribution and remaining features, following Gaussian distribution [12]. Finally, the decoder maps semantic vectors along with other disentangled latent vectors to the future trajectories. Note that we develop the AAE instead of traditional VAE architecture to model these different and complex distributions.", + "bbox": [ + 500, + 227, + 890, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the attacked scenarios, the impact of attacks will be decomposed to different latent vectors and the attack patterns will be explicitly modeled by the semantic features. Let us take the lateral directional vector as an example. If the attack is not targeted at the lateral dimension, the encoder will decompose the attack effect into other vectors and the mapping for lateral direction will remain stable. Otherwise, if the attack causes errors in the lateral vector, the feature extractor and encoder will be adversarial trained on the label of the direction and the corresponding mapping from the latent distribution to the final trajectories will also be updated.", + "bbox": [ + 500, + 484, + 890, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared to existing trajectory prediction works that only apply adversarial training on the final trajectory waypoints, our proposed method is designed to capture semantic features in the latent space, and it can benefit the adversarial robustness and its generalization in various aspects:", + "bbox": [ + 500, + 666, + 890, + 739 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The architecture maps the high-dimensional features into disentangled latent space that can decompose the attacks into different orthogonal patterns. Such disentanglement will boost the adversarial robustness [41].", + "- The semantic vectors provide more context information and interpretable labels for the adversarial training and the semi-supervised learning makes the adversarial training process more efficient and easier to generalize to the unseen scenarios." + ], + "bbox": [ + 519, + 753, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8210", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/466c3ad5503ae0455b8dabcff162510d6441254a56aab57a6cbdef8a130aadd4.jpg", + "image_caption": [ + "Figure 3: The overall adversarial training pipeline for the proposed semi-supervised semantics-guided architecture. We leverage the feature extractor in [16]. In the latent space, the D nets represent the discriminator that models the latent distributions by distinguishing the samples of real distribution from the predicted ones. When the semantic labels are available in the scenarios, they are used to adversarially train the semantic vectors (parts in green). Finally, we conduct adversarial training on attacked trajectories." + ], + "image_footnote": [], + "bbox": [ + 158, + 88, + 816, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- The encoder-decoder architecture can filter high-frequency noise by dimension reduction.", + "bbox": [ + 94, + 469, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The detailed design and optimization of the architecture are described in the following.", + "bbox": [ + 76, + 511, + 468, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1.2 Semi-supervised Semantic Features Modeling", + "text_level": 1, + "bbox": [ + 76, + 560, + 447, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To model the high-level semantic information in the trajectory prediction task, it is natural to decompose the trajectories into two dimensions – longitudinal and lateral directions. We want to utilize domain knowledge to guide the modeling by providing representative metrics and prior distributions. In addition to the semantic features, the architecture also maps high-dimensional information into latent vectors in Gaussian distributions to represent other low-level and more random features. We will mainly explain the semantic feature modeling below.", + "bbox": [ + 75, + 583, + 468, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the longitudinal direction, speed and acceleration are often used to model the vehicle's dynamics but their values are always changing and do not contain enough semantic information. In our model, we apply the time headway to effectively extract the longitudinal feature, which measures the time difference between two successive vehicles when they cross a given point. The time headway represents a relatively stable behavior pattern for a certain agent and takes interaction with other vehicles into consideration. Recent works [44, 22] also use time headway as a measurement of aggressiveness in specific scenarios such as lane changing", + "bbox": [ + 75, + 734, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "or merging. In our proposed architecture, the model will represent the time headway as a one-dimensional vector in the latent space. In both benign and adversarial cases, the encoder will be trained by the regularization loss to force the longitudinal feature to a certain distribution that the time headway follows in statistics. Prior works [7] in the transportation domain show that the time headway in urban scenarios can be best described with the Log-normal distribution. We verify this and estimate the parameters in the Argoverse 1 motion forecasting dataset [4]. The longitudinal feature follows the distribution shown in Eq. (2):", + "bbox": [ + 496, + 469, + 890, + 635 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf (x) = \\frac {1}{x \\sigma \\sqrt {2 \\pi}} e ^ {- \\frac {1}{2} \\left(\\frac {\\ln (x) - \\mu}{\\sigma}\\right) ^ {2}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 648, + 890, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu$ is the location parameter and $\\sigma$ is the scale parameter. We can explicitly obtain the true time headway values for semi-supervised training when there is observable interaction between the attacked target and the front vehicle. We consider the semi-supervised longitudinal feature encoding as a regression problem and optimize it by minimizing the mean square error. For the lateral directional features, we represent them by three simple but effective classes: moving forward, turning/changing lanes to the left, and turning/changing lanes to the right. These three intentions are discrete by nature and we model them with the categorical distribution. In the adversarial training process, only vehicles with clear intentions in a long enough time frame will be labeled and we utilize the cross-entropy to optimize this", + "bbox": [ + 496, + 689, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8211", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "classification task.", + "bbox": [ + 76, + 90, + 202, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For all the semantic and Gaussian latent variables, they are regularized to the target distribution by the adversarial generation loss in Eq. (3). The discriminators are trained to maximize the log probability of real latent samples $s$ and the log of the inverse probability for fake latent samples, as in Eq. (4):", + "bbox": [ + 75, + 106, + 468, + 196 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL o s s _ {G} (x) = \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\log \\left(1 - D _ {i} (G (x))\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 205, + 468, + 244 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {L o s s} _ {D _ {i}} (x, s) = \\log D _ {i} \\left(s _ {i}\\right) + \\log \\left(1 - D _ {i} (G (x))\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 248, + 468, + 265 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $x$ is the high-dimensional features and $m$ is the number of different kinds of latent vectors. $G$ and $D$ are encoders and distribution discriminators, respectively.", + "bbox": [ + 75, + 273, + 468, + 320 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. Adversarial Training Process", + "text_level": 1, + "bbox": [ + 76, + 328, + 336, + 344 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2.1 Adversarial Training Algorithm", + "text_level": 1, + "bbox": [ + 76, + 351, + 354, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For each sample, we utilize the PGD attack to generate the adversarial trajectory only for the target vehicle and keep other surrounding vehicles' original trajectories. This constrains the adversarial attack's impact on the whole scenario. If the error between the prediction under attack and the ground truth is greater than a threshold, we consider the attack a successful one and conduct adversarial training on this sample. Since the perturbation is very small, we consider the real future trajectory as the ground truth $y_{i}$ for the adversarial training and optimize the whole pipeline with L1-smooth loss in Eq. (5). This is a general but somewhat naive adversarial training process.", + "bbox": [ + 75, + 375, + 468, + 556 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {L o s s} _ {\\text {t r a j}} \\left(y _ {i}, \\hat {y} _ {i}\\right) = \\left\\{ \\begin{array}{l l} 0. 5 \\left(y _ {i} - \\hat {y} _ {i}\\right) ^ {2} & \\text {i f} \\| y _ {i} - \\hat {y} _ {i} \\| < 1, \\\\ \\| y _ {i} - \\hat {y} _ {i} \\| - 0. 5 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 565, + 468, + 618 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Thus, to further facilitate the adversarial training, we exploit semantic features and their corresponding labels in our proposed architecture. The encoder is optimized to minimize the mean square error of longitudinal features and the cross entropy of lateral features between the ground truths and predictions in the latent space. The semi-supervised loss function is shown as follows:", + "bbox": [ + 75, + 619, + 468, + 723 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL o s s _ {S e m i} (z, g) = - \\sum_ {i = 1} ^ {3} g _ {l a t} \\log z _ {l a t} + \\left(g _ {l o n} - z _ {l o n}\\right) ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 729, + 468, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $z$ represents the predicted semantic vectors when under attack and $g$ represents the ground truth in the benign scenarios.", + "bbox": [ + 75, + 780, + 468, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, we further adapt the adversarial training process with lateral semantic vectors because the lateral directional prediction can be regarded as a classification problem with clear behavior meaning. When the adversarial example leads to the wrong classification of lateral behavior, we", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "will set higher weights of the semi-supervised loss for the adversarial training. In this way, our model will first guarantee the correctness of high-level semantic prediction and then tune the regression error, which could help avoid significant adversarial deviation and enhance the generalization performance.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Our Adversarial Training Pipeline" + ], + "code_body": "1: Initialize: feature extractor $F$ , AAE encoder $G$ , decoder $R$ , discriminator $D_{i}$ , target distribution $p_{i}$ , $i = 1,2,3$ , adversarial example generator $Adv$ . \n2: Input: past trajectories $t$ , future trajectories ground truth $y$ , map context $c$ , prediction model $m$ . \n3: for each sample do \n4: Generate adversarial trajectory for target vehicle $t_{adv} = Adv(t,y,m,c)$ . \n5: Input $t_{adv}$ to the predictor. $\\hat{y}', \\hat{z}' = m(t_{adv},c)$ , where $z = \\{z_{lat},z_{lon},z_{gaussian}\\}$ . \n6: if err $(\\hat{y}', y) > Threshold$ then \n7: if $\\text{argmax}(z_{lat}) \\neq \\text{argmax}(z_{lat}')$ then \n8: Update model $m$ with higher weights of semi-supervised learning loss $Loss_{semi}$ in Eq. (6). \n9: Update model $m$ with $Loss_{D}$ in Eq. (4), $Loss_{G}$ in Eq. (3), and $Loss_{traj}$ in Eq. (5). \n10: else \n11: Update model $m$ with $Loss_{D}$ in Eq. (4), $Loss_{G}$ in Eq. (3), $Loss_{semi}$ in Eq. (6), and $Loss_{traj}$ in Eq. (5). \n12: end if \n13: end if \n14: end for", + "bbox": [ + 508, + 227, + 893, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2.2 Balance Accuracy and Robustness", + "text_level": 1, + "bbox": [ + 500, + 654, + 792, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our preliminary experiments, we notice a trade-off between standard accuracy and adversarial robustness. A similar phenomenon has been observed in classification tasks [46, 28]. Methods such as TRADES [46], robust self-training [3] and MixUp [47, 1] have been proposed to balance such trade-off. However, there are few methods that can be applied to trajectory prediction because such time-series regression problems have no class labels and are more sensitive to errors introduced by augmented data. In this work, we utilize the MixUp [47] technique to mix the adversarial scenarios and benign scenarios in the adversarial training process. The experiments demonstrate that a balance between adversarial robustness and standard accuracy can be achieved in trajectory prediction, as shown later.", + "bbox": [ + 496, + 689, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8212", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/99d3903851becc8cbcd2232ccdd8a7e5bbad83b3c8b25749bdf19e97a37310e7.jpg", + "table_caption": [ + "Table 2: Comparison of different defense methods when under various attacks and in the benign case. For the adversarial training methods, we calculate the mean error of models adversarially trained on different attacks." + ], + "table_footnote": [], + "table_body": "
MethodsADE AttackLateral AttackLongitudinal AttackBenign
ADE/mLateral Error/mLongitudinal Error/mADE/m
DatasetArgo1Argo2ApolloArgo1Argo2ApolloArgo1Argo2ApolloArgo1Argo2Apollo
Original Model5.174.523.931.661.271.543.783.173.421.430.791.78
Train-time Smo. [48]4.674.323.861.451.271.493.233.003.411.500.832.11
Test-time Smo. [48]4.323.513.490.750.891.093.542.423.051.681.232.07
Heuristic Aug. [29]4.623.523.841.091.211.343.244.243.211.540.972.01
Data-driven Aug. [29]4.534.433.530.740.780.542.732.613.022.501.992.24
Standard AT3.793.673.680.670.641.152.441.893.051.671.061.87
SSAT (ours)3.512.672.900.610.530.411.761.441.261.751.201.87
Mixup-SSAT (ours)3.733.333.170.720.510.532.131.631.351.641.071.86
", + "bbox": [ + 99, + 130, + 870, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 306, + 282, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 76, + 330, + 253, + 347 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1.1 Dataset", + "text_level": 1, + "bbox": [ + 76, + 354, + 184, + 368 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We train and evaluate different defense methods using three popular benchmarks - Argoverse 1 [4], Argoverse 2 [42], and ApolloScape datasets [9]. The datasets contain more than $250\\mathrm{k}$ real driving scenarios in different cities, such as Miami and Pittsburgh. For Argoverse 1 and Argoverse 2, each scenario consists of a road graph and multiple agents' trajectories sampled at a frequency of $10\\mathrm{Hz}$ . We choose 20 waypoints as the history trajectory and the models will predict 30 waypoints in the future. Scenarios in ApolloScape are simpler. They have no maps but 6 waypoints for both history and future trajectories.", + "bbox": [ + 75, + 377, + 470, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1.2 Attack Settings", + "text_level": 1, + "bbox": [ + 76, + 561, + 238, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the experiments, we study three different types of attacks [48] to the vehicle trajectory prediction algorithms - lateral directional attack (shift to the right), longitudinal directional attack (shift forward), and ADE attack (deviate randomly). More details of them are in Section 2.1. The trajectory prediction models are adversarially trained on the three types of attacks, respectively. We constrain the maximum deviation between the attacked and the benign input trajectories to be 1 meter.", + "bbox": [ + 75, + 585, + 468, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1.3 Training Settings", + "text_level": 1, + "bbox": [ + 76, + 739, + 250, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since our architecture is an encoder-decoder module that can be combined with different feature extractors, we first fine-tune the model on the benign data. In the experiments, we use the feature extractor from LaneGCN [16], an attention-based graph neural network. We notice that the AAE architecture introduces a slight accuracy drop on the benign data, mainly due to the dimension reduction. For adversarial training, we train prediction models on adversarial samples generated from scenarios in datasets.", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Experimental Results and Analysis", + "text_level": 1, + "bbox": [ + 500, + 306, + 805, + 323 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct experiments with various defense methods under different patterns of attacks, including our semi-supervised semantics-guided (SSAT) method and a Mixup-SSAT method that combines SSAT with the MixUp technique for balancing standard accuracy and adversarial robustness. In the following, we first compare the average robustness improvement among various defense methods, which demonstrates the advantage of SSAT in improving robustness under various types of attacks and the effectiveness of Mixup-SSAT in balancing robustness and accuracy. Then, we show that SSAT can significantly enhance the robust generalization to unseen types of attacks. In addition, we evaluate an unsupervised version of SSAT to explicitly show how the semi-supervised semantic-guided latent space modeling can boost the adversarial robustness, which also serves as an ablation study.", + "bbox": [ + 496, + 330, + 893, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1 Effectiveness of SSAT Methods", + "text_level": 1, + "bbox": [ + 500, + 589, + 769, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our SSAT and Mixup-SSAT methods with the original model and five different defense methods, including train-time smoothing [48], test-time smoothing [48], heuristic data augmentation [29], data-driven augmentation [29], and the standard adversarial training (Standard AT). All the models share the same feature extractor LaneGCN in this setting. Note that we compare to the two data augmentation methods as they are effective for image classification tasks [6, 29, 35]. For the data-driven augmentation, we design an additional decoder to augment input trajectories and it can generate more inputs by adding Gaussian noises to the latent vectors of the real inputs. For the heuristic augmentation, we simply add random perturbation to the benign inputs, with the same constraints of maximum deviation.", + "bbox": [ + 496, + 613, + 893, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 shows the prediction errors of different methods when under various types of attacks and in the benign case (note that when under lateral and longitudinal attacks, we measure lateral and longitudinal errors). We can see that our SSAT method significantly outperforms all other de", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8213", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa1253b2226a86d5e0b88a2cdffb393e62baeed3c97f70f7621773b97277dca5.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 114, + 99, + 321, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e6cee0f856483597832978a138c131ea0afc0ce22d9f48def8bf231c285ee866.jpg", + "image_caption": [ + "(b)", + "Figure 4: Adversarial robustness comparisons of the original model, the standard adversarial training on final trajectories (Standard AT), and the SSAT method proposed by us. Figures (a), (b), (c) show results from adversarial training (AT) that targets ADE attacks, lateral attacks, and longitudinal attacks, respectively." + ], + "image_footnote": [], + "bbox": [ + 375, + 99, + 584, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1bbcab51993c0b022ee18572adfccec7179ed6b2333c7d2b9ac9d51a37701c18.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 640, + 99, + 846, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fense methods in improving the robustness of trajectory prediction. Compared with the original model, SSAT can reduce the prediction error by $32\\% - 73\\%$ when under different types of attacks.", + "bbox": [ + 75, + 321, + 467, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, while our SSAT method improves robustness, we also observe a drop in standard accuracy in the benign case. Mixup-SSAT enables effective trade-off between these two objectives (i.e., better performance than SSAT in benign cases at the expense of worse performance under attacks), by setting the mixup ratio between adversarial and benign examples to different values (the results in Table 2 are based on a mixup ratio of 2).", + "bbox": [ + 75, + 385, + 467, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also notice that both data-driven and heuristic data augmentation methods offer very limited improvement over the original model. This is likely due to the challenge of regression tasks with rich context, which makes augmentation methods perform poorer than they do for image classification tasks.", + "bbox": [ + 75, + 508, + 467, + 598 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4f7872b9bf072a0eacea54bec277099b8aec978fb286742787ed1b60d5411e05.jpg", + "table_caption": [ + "Table 3: Comparison of different methods when adversarial training is conducted for ADE attack, and tested on ADE, lateral, and longitudinal attacks." + ], + "table_footnote": [], + "table_body": "
MethodsPrediction Error/m, w/o AT → with AT
ADE AttackLat AttackLon Attack
Standard AT5.17 → 3.761.58 → 0.833.59 → 2.66
SSAT5.16 → 3.241.66 → 0.733.78 → 1.45
Unsup-SSAT5.16 → 3.481.66 → 0.743.78 → 1.49
", + "bbox": [ + 109, + 672, + 436, + 737 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2f921a19736a0eaa01d3421c9c5d854abb183354a0599d379d2e7e216aff15db.jpg", + "table_caption": [ + "Table 4: Comparison of different methods when adversarial training is conducted for lateral attack, and tested on ADE, lateral, and longitudinal attacks." + ], + "table_footnote": [], + "table_body": "
MethodsPrediction Error/m, w/o AT → with AT
ADE AttackLat AttackLon Attack
Standard AT5.17 → 4.281.58 → 0.593.59 → 3.01
SSAT5.16 → 4.021.66 → 0.593.78 → 2.52
Unsup-SSAT5.16 → 4.111.66 → 0.693.78 → 2.54
", + "bbox": [ + 109, + 834, + 436, + 897 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.2 Effectiveness of SSAT in Robust Generalization on Different Types of Attacks", + "text_level": 1, + "bbox": [ + 500, + 321, + 890, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We observe that there is an adversarial robust generalization gap when the training and test are under different types of attacks. The comparisons in Tables 3, 4, 5 show that our SSAT method is better generalized to unseen types of attacks, compared to the standard adversarial training. In every training scenario, our method is more robust to the various unseen patterns of attacks. For instance, Table 3 shows that when applying SSAT to train under the random ADE attack, its results outperform other models on all seen (i.e., ADE) and unseen (i.e., lateral and longitudinal) types of attacks, which demonstrates that our SSAT method can effectively decompose and learn semantic features from random ADE attacks. Tables 4 and 5 show similar trends, where our SSAT methods are better at defending against unseen attacks and mitigating overfitting on specific patterns of attacks. Fig. 4 further visualizes the results from these three tables for the original model, the standard adversarial training (Standard AT), and SSAT.", + "bbox": [ + 496, + 359, + 890, + 632 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7b2736180437cf2c0f4416e8b3233a059fc090081b6969ef4e0f277b75347fea.jpg", + "table_caption": [ + "Table 5: Comparison of different methods when adversarial training is conducted for longitudinal attack, and tested on ADE, lateral, and longitudinal attacks." + ], + "table_footnote": [], + "table_body": "
MethodsPrediction Error/m, w/o AT → with AT
ADE AttackLat AttackLon Attack
Standard AT5.17 → 3.331.58 → 0.603.59 → 1.64
SSAT5.16 → 3.281.66 → 0.523.78 → 1.30
Unsup-SSAT5.16 → 3.401.66 → 0.613.78 → 1.49
", + "bbox": [ + 532, + 702, + 859, + 766 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.3 Impact from Latent Space Modeling", + "text_level": 1, + "bbox": [ + 500, + 786, + 807, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also conduct adversarial training that only has regularization on the latent distributions but without supervision on latent vectors. We name it Unsup-SSAT. The comparisons between the standard adversarial training (Standard AT) and Unsup-SSAT in Tables 3, 4, and 5 demonstrate that even without labels, the partial disentanglement and", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8214", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "distribution modeling in Unsup-SSAT will benefit the adversarial training on trajectory prediction and outperform the baseline Standard AT in most cases. However, when compared with SSAT, we find that the adversarial robustness will be further improved with the extra labels in the semi-supervised phase (in practice, we often have access to those labels).", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 210, + 194, + 226 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we propose an adversarial training method for trajectory prediction. To tackle the challenge of random inputs with rich context, diverse types of attacks, and lack of class labels, we develop a novel AAE architecture that exploits the disentanglement and semantic features for enhancing model robustness and its generalization. Our proposed SSAT method significantly outperforms a number of baselines from the literature, reducing the prediction errors under attacks by up to $32\\% - 73\\%$ when compared with the original prediction models. Our method is also shown to be effective for defending against unseen attacks.", + "bbox": [ + 75, + 236, + 472, + 402 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 76, + 416, + 228, + 431 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We gratefully acknowledge the support from NSF grants 1834701, 1724341, 2038853, 1929771, 2145493, 1932464, and ONR grant N00014-19-1-2496.", + "bbox": [ + 76, + 441, + 468, + 486 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 500, + 173, + 516 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Guillaume P Archambault, Yongyi Mao, Hongyu Guo, and Richong Zhang. Mixup as directional adversarial training. arXiv preprint arXiv:1906.06875, 2019.", + "[2] Yulong Cao, Danfei Xu, Xinshuo Weng, Zhuoqing Mao, Anima Anandkumar, Chaowei Xiao, and Marco Pavone. Robust trajectory prediction against adversarial attacks. In 6th Annual Conference on Robot Learning.", + "[3] Yair Carmon, Aditi Raghunathan, Ludwig Schmidt, John C Duchi, and Percy S Liang. Unlabeled data improves adversarial robustness. Advances in Neural Information Processing Systems, 32, 2019.", + "[4] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8748-8757, 2019.", + "[5] Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia" + ], + "bbox": [ + 86, + 525, + 472, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11525-11533, 2020.", + "[6] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 34:4218-4233, 2021.", + "[7] Duy-Hung Ha, Maurice Aron, and Simon Cohen. Time headway variable and probabilistic modeling. Transportation Research Part C: Emerging Technologies, 25:181-201, 2012.", + "[8] Michael W Hancock and Bud Wright. A policy on geometric design of highways and streets. American Association of State Highway and Transportation Officials: Washington, DC, USA, 2013.", + "[9] Xinyu Huang, Xinjing Cheng, Qichuan Geng, Binbin Cao, Dingfu Zhou, Peng Wang, Yuanqing Lin, and Ruigang Yang. The apolloscape dataset for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 954-960, 2018.", + "[10] Ruochen Jiao, Juyang Bai, Xiangguo Liu, Takami Sato, Xiaowei Yuan, Qi Alfred Chen, and Qi Zhu. Learning representation for anomaly detection of vehicle trajectories. 2023.", + "[11] Ruochen Jiao, Hengyi Liang, Takami Sato, Junjie Shen, Qi Alfred Chen, and Qi Zhu. End-to-end uncertainty-based mitigation of adversarial attacks to automated lane centering. In 2021 IEEE Intelligent Vehicles Symposium (IV), pages 266-273. IEEE, 2021.", + "[12] Ruochen Jiao, Xiangguo Liu, Bowen Zheng, Dave Liang, and Qi Zhu. Tae: A semi-supervised controllable behavior-aware trajectory generator and predictor. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 12534-12541. IEEE, 2022.", + "[13] Diederik Kingma and Max Welling. An introduction to variational autoencoders. arXiv preprint arXiv:1906.02691, 2019.", + "[14] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013.", + "[15] Dasom Lee and David J Hess. Regulations for on-road testing of connected and automated vehicles: Assessing the potential for global safety harmonization." + ], + "bbox": [ + 501, + 90, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8215", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Transportation Research Part A: Policy and Practice, 136:85-98, 2020.", + "[16] Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph representations for motion forecasting. In European Conference on Computer Vision, pages 541-556. Springer, 2020.", + "[17] C. Lin, Bowen Zheng, Qi Zhu, and Alberto Sangiovanni-Vincentelli. Security-Aware Design Methodology and Optimization for Automotive Systems. ACM Transactions on Design Automation of Electronic Systems (TODAES), 21(1):18:1-18:26, December 2015.", + "[18] Xiangguo Liu, Chao Huang, Yixuan Wang, Bowen Zheng, and Qi Zhu. Physics-aware safety-assured design of hierarchical neural network based planner. In 2022 ACM/IEEE 13th International Conference on Cyber-Physical Systems (ICCPS), pages 137-146. IEEE, 2022.", + "[19] Xiangguo Liu, Ruochen Jiao, Yixuan Wang, Yimin Han, Bowen Zheng, and Qi Zhu. Safety-assured speculative planning with adaptive prediction. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023.", + "[20] Xiangguo Liu, Ruochen Jiao, Bowen Zheng, Dave Liang, and Qi Zhu. Safety-driven interactive planning for neural network-based lane changing. In Proceedings of the 28th Asia and South Pacific Design Automation Conference, pages 39-45, 2023.", + "[21] Xiangguo Liu, Yunpeng Luo, Anthony Goeckner, Trishna Chakraborty, Ruochen Jiao, Ningfei Wang, Yixuan Wang, Takami Sato, Qi Alfred Chen, and Qi Zhu. Invited: Waving the double-edged sword: Building resilient cavs with edge and cloud computing. In Proceedings of the 60th Annual Design Automation Conference, 2023.", + "[22] Xiangguo Liu, Neda Masoud, and Qi Zhu. Impact of sharing driving attitude information: A quantitative study on lane changing. In 2020 IEEE Intelligent Vehicles Symposium (IV), pages 1998-2005. IEEE, 2020.", + "[23] Yicheng Liu, Jinghuai Zhang, Liangji Fang, Qinhong Jiang, and Bolei Zhou. Multimodal motion prediction with stacked transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7577-7586, 2021.", + "[24] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Pratyush Maini, Eric Wong, and Zico Kolter. Adversarial robustness against the union of multiple perturbation models. In International Conference on Machine Learning, pages 6640-6650. PMLR, 2020.", + "[26] Alireza Makhzani, Jonathon Shlens, Navdeep Jaitly, Ian Goodfellow, and Brendan Frey. Adversarial autoencoders. arXiv preprint arXiv:1511.05644, 2015.", + "[27] Emile Mathieu, Tom Rainforth, Nana Siddharth, and Yee Whye Teh. Disentangling disentanglement in variational autoencoders. In International Conference on Machine Learning, pages 4402-4412. PMLR, 2019.", + "[28] Aditi Raghunathan, Sang Michael Xie, Fanny Yang, John Duchi, and Percy Liang. Understanding and mitigating the tradeoff between robustness and accuracy. arXiv preprint arXiv:2002.10716, 2020.", + "[29] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan A Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Fixing data augmentation to improve adversarial robustness. arXiv preprint arXiv:2103.01946, 2021.", + "[30] Leslie Rice, Eric Wong, and Zico Kolter. Overfitting in adversarially robust deep learning. In International Conference on Machine Learning, pages 8093-8104. PMLR, 2020.", + "[31] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory++: Dynamically-feasible trajectory forecasting with heterogeneous data. In European Conference on Computer Vision, pages 683-700. Springer, 2020.", + "[32] Takami Sato, Junjie Shen, Ningfei Wang, Yunhan Jia, Xue Lin, and Qi Alfred Chen. Dirty road can attack: Security of deep learning based automated lane centering under {Physical-World} attack. In 30th USENIX Security Symposium (USENIX Security 21), pages 3309–3326, 2021.", + "[33] Ludwig Schmidt, Shibani Santurkar, Dimitris Tsipras, Kunal Talwar, and Aleksander Madry. Adversarily robust generalization requires more data. Advances in neural information processing systems, 31, 2018.", + "[34] Lukas Schott, Jonas Rauber, Matthias Bethge, and Wieland Brendel. Towards the first adversarially robust neural network model on mnist. arXiv preprint arXiv:1805.09190, 2018.", + "[35] Vikash Sehwag, Saeed Mahloujifar, Tinashe Handina, Sihui Dai, Chong Xiang, Mung Chiang, and Prateek Mittal. Robust learning meets generative models: Can proxy distributions improve adversarial robustness? arXiv preprint arXiv:2104.09425, 2021." + ], + "bbox": [ + 501, + 90, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8216", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Amolika Sinha, Sai Chand, Kasun P Wijayaratna, Navreet Virdi, and Vinayak Dixit. Comprehensive safety assessment in mixed fleets with connected and automated vehicles: A crash severity and rate evaluation of conventional vehicles. Accident Analysis & Prevention, 142:105567, 2020.", + "[37] Chuanbiao Song, Kun He, Jiadong Lin, Liwei Wang, and John E Hopcroft. Robust local features for improving the generalization of adversarial training. arXiv preprint arXiv:1909.10147, 2019.", + "[38] Dong Su, Huan Zhang, Hongge Chen, Jinfeng Yi, Pin-Yu Chen, and Yupeng Gao. Is robustness the cost of accuracy? a comprehensive study on the robustness of 18 deep image classification models. In Proceedings of the European Conference on Computer Vision (ECCV), pages 631-648, 2018.", + "[39] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013.", + "[40] Zhilu Wang, Chao Huang, and Qi Zhu. Efficient global robustness certification of neural networks via interleaving twin-network encoding. In DATE'22: Proceedings of the Conference on Design, Automation and Test in Europe, 2022.", + "[41] Matthew Willetts, Alexander Camuto, Tom Rainforth, Stephen Roberts, and Chris Holmes. Improving vaes' robustness to adversarial attack. arXiv preprint arXiv:1906.00230, 2019.", + "[42] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, et al. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2).", + "[43] Dongxian Wu, Shu-Tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization. Advances in Neural Information Processing Systems, 33:2958-2969, 2020.", + "[44] Hongtao Yu, H Eric Tseng, and Reza Langari. A human-like game theory-based controller for automatic lane changing. Transportation Research Part C: Emerging Technologies, 88:140-158, 2018.", + "[45] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M Kitani. Agentformer: Agent-aware transformers for" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9813-9823, 2021.", + "[46] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019.", + "[47] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017.", + "[48] Qingzhao Zhang, Shengtuo Hu, Jiachen Sun, Qi Alfred Chen, and Z Morley Mao. On adversarial robustness of trajectory prediction for autonomous vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15159-15168, 2022.", + "[49] Shufei Zhang, Zhuang Qian, Kaizhu Huang, Qiufeng Wang, Rui Zhang, and Xinping Yi. Towards better robust generalization with shift consistency regularization. In International Conference on Machine Learning, pages 12524-12534. PMLR, 2021.", + "[50] Bowen Zheng, Hengyi Liang, Qi Zhu, Huafeng Yu, and Chung-Wei Lin. Next generation automotive architecture modeling and exploration for autonomous driving. In 2016 IEEE Computer Society Annual Symposium on VLSI (ISVLSI), pages 53–58, 2016.", + "[51] Bowen Zheng, Chung-Wei Lin, et al. Design and Analysis of Delay-Tolerant Intelligent Intersection Management. ACM T-CPS, 4(1):3:1-3:27, November 2019.", + "[52] Qi Zhu, Chao Huang, Ruochen Jiao, Shuyue Lan, Hengyi Liang, Xiangguo Liu, Yixuan Wang, Zhilu Wang, and Shichao Xu. Safety-assured design and adaptation of learning-enabled autonomous systems. In Proceedings of the 26th Asia and South Pacific Design Automation Conference, 2021.", + "[53] Qi Zhu, Wenchao Li, Hyoseung Kim, Yecheng Xiang, Kacper Wardega, Zhilu Wang, Yixuan Wang, Hengyi Liang, Chao Huang, Jiameng Fan, and Hyunjong Choi. Know the unknowns: Addressing disturbances and uncertainties in autonomous systems. In Proceedings of the 39th International Conference on Computer-Aided Design, ICCAD '20, New York, NY, USA, 2020. Association for Computing Machinery." + ], + "bbox": [ + 501, + 90, + 890, + 866 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8217", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction/487dfb22-6094-4f31-976c-310213635421_model.json b/2023/Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction/487dfb22-6094-4f31-976c-310213635421_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b56db8817fa46ba44c8cf4a2dc34ce3a404335bc --- /dev/null +++ b/2023/Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction/487dfb22-6094-4f31-976c-310213635421_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.183, + 0.131, + 0.787, + 0.176 + ], + "angle": 0, + "content": "Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction" + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.205, + 0.4, + 0.221 + ], + "angle": 0, + "content": "Ruochen Jiao" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.223, + 0.445, + 0.24 + ], + "angle": 0, + "content": "Northwestern University" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.243, + 0.484, + 0.256 + ], + "angle": 0, + "content": "ruochen.jiao@u.northwestern.edu" + }, + { + "type": "text", + "bbox": [ + 0.594, + 0.205, + 0.708, + 0.222 + ], + "angle": 0, + "content": "Xiangguo Liu" + }, + { + "type": "text", + "bbox": [ + 0.553, + 0.223, + 0.749, + 0.24 + ], + "angle": 0, + "content": "Northwestern University" + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.243, + 0.762, + 0.256 + ], + "angle": 0, + "content": "xg.liu@u.northwestern.edu" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.265, + 0.277, + 0.28 + ], + "angle": 0, + "content": "Takami Sato" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.283, + 0.349, + 0.3 + ], + "angle": 0, + "content": "University of California, Irvine" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.304, + 0.293, + 0.316 + ], + "angle": 0, + "content": "takamis@uci.edu" + }, + { + "type": "text", + "bbox": [ + 0.387, + 0.283, + 0.634, + 0.3 + ], + "angle": 0, + "content": "University of California, Irvine" + }, + { + "type": "text", + "bbox": [ + 0.443, + 0.304, + 0.578, + 0.316 + ], + "angle": 0, + "content": "alfchen@uci.edu" + }, + { + "type": "text", + "bbox": [ + 0.74, + 0.266, + 0.8, + 0.282 + ], + "angle": 0, + "content": "Qi Zhu" + }, + { + "type": "text", + "bbox": [ + 0.672, + 0.284, + 0.868, + 0.3 + ], + "angle": 0, + "content": "Northwestern University" + }, + { + "type": "text", + "bbox": [ + 0.677, + 0.304, + 0.864, + 0.316 + ], + "angle": 0, + "content": "qzhu@northwestern.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.352, + 0.313, + 0.368 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.384, + 0.473, + 0.793 + ], + "angle": 0, + "content": "Predicting the trajectories of surrounding objects is a critical task for self-driving vehicles and many other autonomous systems. Recent works demonstrate that adversarial attacks on trajectory prediction, where small crafted perturbations are introduced to history trajectories, may significantly mislead the prediction of future trajectories and induce unsafe planning. However, few works have addressed enhancing the robustness of this important safety-critical task. In this paper, we present a novel adversarial training method for trajectory prediction. Compared with typical adversarial training on image tasks, our work is challenged by more random input with rich context and a lack of class labels. To address these challenges, we propose a method based on a semi-supervised adversarial autoencoder, which models disentangled semantic features with domain knowledge and provides additional latent labels for the adversarial training. Extensive experiments with different types of attacks demonstrate that our Semi-supervised Semantics-guided Adversarial Training (SSAT1) method can effectively mitigate the impact of adversarial attacks by up to \\(73\\%\\) and outperform other popular defense methods. In addition, experiments show that our method can significantly improve the system's robust generalization to unseen patterns of attacks. We believe that such semantics-guided architecture and advancement on robust generalization is an important step for developing robust prediction models and enabling safe decision making." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.821, + 0.208, + 0.837 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.847, + 0.469, + 0.878 + ], + "angle": 0, + "content": "Connected and autonomous vehicles (CAVs) have shown great promise to revolutionize the transportation systems," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.354, + 0.892, + 0.459 + ], + "angle": 0, + "content": "but also face significant concerns regarding their safety and security [15, 36, 21, 17, 50, 51]. Many of these concerns, which also apply to many other autonomous systems, arise from the increasing adoption of advanced deep neural network (DNN)-based machine learning techniques across system sensing, perception, prediction, planning, control, and general decision-making [52, 53]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.893, + 0.793 + ], + "angle": 0, + "content": "In particular, the adversarial robustness of DNNs has drawn significant attention in recent years. State-of-the-art classifiers can be misled to make erroneous predictions by slight but carefully-optimized perturbations [39], and addressing this is critical to the safety of autonomous systems that employ these classifiers [11, 40]. Besides classification, adversarial attacks are also targeted at various other tasks such as object detection, image generation and natural language processing. To defend against adversarial examples, adversarial training is commonly used to enhance the intrinsic robustness of models and has been shown to be very effective among various defense strategies [24, 25, 34]. On the other hand, the works in [30, 43, 37] show that adversarial training may suffer poor robust generalization on unseen attacks and [33, 24, 38] demonstrate a trade-off between robustness and accuracy. In this paper, we address trajectory prediction, a safety-critical task that is common in self-driving vehicles and other autonomous systems, and propose a novel architecture to enhance its adversarial robustness and robust generalization by introducing semantic features and a semi-supervised adversarial autoencoder (AAE) [26] into the adversarial training process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We focus on the trajectory prediction module in autonomous driving, where large DNNs have enabled breakthroughs in recent years. More specifically, an autonomous driving system typically consists of several modules such as perception, localization, prediction, planning (route, behavior, trajectory, and motion planning), and control. The perception module is to detect agents and obstacles based" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.454, + 0.901 + ], + "angle": 0, + "content": "1Our codes: https://github.com/jrcblue/SSAT-for-Motion-Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8207" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.485 + ], + "angle": 0, + "content": "on the sensing inputs (e.g., images and 3D point clouds), and the effectiveness and robustness of the perception module are well-studied in the research community. The prediction module is to predict future trajectories of surrounding vehicles based on the observed history trajectories of those vehicles from the perception module and the map context. The trajectory prediction plays a crucial role in understanding the environment and making safety-critical decisions in the following planning module [18, 20, 19]. Recent works [23, 16, 5, 45, 12] have applied various deep learning techniques such as graph neural networks and transformers to this task and achieved impressive performance in terms of reducing average errors. However, few works studied the robustness of vehicle trajectory prediction under adversarial attacks, which is in fact of vital importance because 1) autonomous driving is a safety-critical task by nature, 2) recent work [48, 10] demonstrates that the prediction module is vulnerable to adversarial attacks if surrounding vehicles drive along a seemingly natural but crafted trajectory, and 3) current prediction models are often overfitted on limited patterns in the datasets but suffer the long-tailed distribution of driving scenarios and behaviors. The threat model of the adversarial trajectory is illustrated in an example in Fig. 1, where a crafted input trajectory can intentionally mislead the target vehicle to the wrong prediction and induce dangerous planning decisions." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.497, + 0.47, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.807 + ], + "angle": 0, + "content": "Figure 1: An illustrative example of the threat model: In the upper benign scenario, the target vehicle predicts the future trajectory of a surrounding vehicle based on its history trajectory. In the lower adversarial scenario, the surrounding vehicle is malicious and drives along a crafted (history) trajectory, which may mislead the target vehicle to have a wrong (attacked) prediction of the surrounding vehicle's future trajectory. In this particular case, the target vehicle wrongly predicts that the malicious vehicle will cut into its lane, and thus takes an unsafe emergency braking." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To defend against adversarial attacks on vehicle trajectory prediction, there are important challenges that are different from the cases of images and audios. First, vehicle trajectory prediction is a time-series regression problem with rich contexts while most existing adversarial attacks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.409 + ], + "angle": 0, + "content": "and corresponding defense methods are targeted at classification tasks. The attack patterns are more random and there are no well-defined class labels, which means that the robust model is difficult to train and generalize. Second, vehicle trajectory can convey semantic and behavior information. For instance, people can infer the behavior of a vehicle such as moving forward or changing lanes from its trajectory. Therefore, as shown in our experiments later, some popular defense methods such as TRADES [46] and data augmentation methods [6, 29, 35] either are inapplicable or have poor performance in trajectory prediction task. In this work, we first propose an adversarial training pipeline for the trajectory prediction task, and then further exploit the semantic features to design a semi-supervised AAE architecture that can be added after the feature extractor, to improve the adversarial robustness and its generalization. The methodology of the proposed architecture such as enhancing disentanglement and defining semantic labels may be further applied to adversarial training for general regression and generation problems. Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.418, + 0.891, + 0.449 + ], + "angle": 0, + "content": "- We propose a novel adversarial training method against adversarial attacks on trajectory predictions." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.458, + 0.892, + 0.519 + ], + "angle": 0, + "content": "- We develop a semi-supervised architecture with domain knowledge and semantic features to enhance the adversarial robustness and its generalization among different types of attack patterns." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.528, + 0.892, + 0.574 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that our proposed method (SSAT) effectively improves the adversarial robustness and outperforms popular defense baselines." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.583, + 0.891, + 0.644 + ], + "angle": 0, + "content": "- We further explore to balance the robustness-accuracy trade-off in this task by leveraging the MixUp technique [47], and we also test data augmentation methods for trajectory prediction robustness." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.418, + 0.892, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.657, + 0.639, + 0.672 + ], + "angle": 0, + "content": "2. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.682, + 0.884, + 0.699 + ], + "angle": 0, + "content": "2.1. Adversarial Attacks on Trajectory Prediction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Recent research [48, 10] demonstrates that trajectory prediction in autonomous driving can be fooled by the adversarial behavior of a surrounding vehicle, where the adversarial behavior is optimized with Projected Gradient Descent (PGD) [24]. The threat model is illustrated in Fig. 1. In this work, we consider such attacks and assume the worst setting, i.e., the attacker has full knowledge of the target system and tries to maximize the attack impact. However, malicious agents cannot directly manipulate anything inside the target vehicle and they can only slightly change their own trajectories to mislead the target vehicle indirectly." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Different from image classification, trajectory prediction has no class labels, but it has directional information in the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8208" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.217, + 0.109, + 0.371, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.109, + 0.567, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.109, + 0.765, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.236, + 0.37, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.349, + 0.294, + 0.36 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.236, + 0.567, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.349, + 0.492, + 0.36 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.236, + 0.765, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.349, + 0.69, + 0.36 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.377, + 0.892, + 0.483 + ], + "angle": 0, + "content": "Figure 2: Adversarial trajectories generated by different types of attacks. The figures in the top row are adversarial (red line) and benign (blue line) input history trajectories for prediction and they look very close to human eyes. The figures in the bottom row show the corresponding attacked future trajectory prediction (red dashed line) and the ground truth trajectory in the benign case (green dashed line). The differences between the two are clearly visible, showing the effectiveness of the attacks. Figure (a) is the ADE attack, which will randomly lead to maximum average deviation. Figure (b) is the lateral attack, which mainly leads the vehicle to deviate to the left or right. Figure (c) is the longitudinal attack, which will mainly lead to longitudinal deviation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.47, + 0.616 + ], + "angle": 0, + "content": "context, such as moving forward, turning or changing lanes to the right. Therefore, the attacker can conduct targeted attacks besides random ones. In Fig. 2, three types of adversarial attacks are presented and we observe that they can cause significant and directional errors. [48] proposes directional error metrics for the optimization of targeted attacks, as shown in Eq. (1):" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.627, + 0.469, + 0.647 + ], + "angle": 0, + "content": "\\[\nD (\\alpha , R) = \\left(p _ {\\alpha} - s _ {\\alpha}\\right) ^ {T} \\cdot R \\left(s _ {\\alpha + 1}, s _ {\\alpha}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.469, + 0.764 + ], + "angle": 0, + "content": "where \\(\\alpha\\) denotes the time frame ID. \\(p\\) and \\(s\\) are two-dimensional vectors representing predicted and ground-truth vehicle locations, respectively. \\(R\\) is a function generating the unit vector to a specific direction (lateral or longitudinal). The longitudinal direction is approximated as the vector defined by the adjacent two waypoints of ground truth \\(s_{\\alpha +1}^{n} - s_{\\alpha}^{n}\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.84 + ], + "angle": 0, + "content": "In addition to the directional attack, a random attack can be designed to maximize the Average Displacement Error (ADE), which is the average of the root mean squared error between the predicted waypoints and the ground-truth trajectory waypoints." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Hard constraints are also applied to the maximum deviation of waypoints during the optimization so as to make the adversarial history trajectory physically feasible and not perform unrealistic behaviors." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.508, + 0.892, + 0.539 + ], + "angle": 0, + "content": "Table 1: Trajectory prediction errors (in meters) in benign or adversarial scenarios under different types of attacks." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.55, + 0.86, + 0.598 + ], + "angle": 0, + "content": "
ModelADELateralLongitudinal
benign/attackbenign/attackbenign/attack
LaneGCN [16]1.32/5.17-0.01/1.58-0.25/3.33
Trajectory++[31]2.69/6.810.107/2.25-0.526 / 3.79
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.647, + 0.892, + 0.799 + ], + "angle": 0, + "content": "Table 1 quantitatively shows that the predicted trajectories can be seriously deviated by any of the three attack types. Generally, in the US, \\(0.3\\mathrm{m}\\) lateral deviation is enough to invade adjacent lanes on local roads [32, 8]. In addition, the final displacement error (the distance between the last points of predicted and ground-truth trajectories) is around two to three times larger than the average error shown here. Therefore, the attacker can apply either random or directional attacks that will greatly challenge the defense methods, especially on the generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[48, 2] propose to utilize approaches such as smoothing and adversarial training on the conditional variational encoder (CVAE) architectures to mitigate the impact of such adversarial attacks. However, they do not explicitly address driving semantics and robustness generalization, which are critical for many real road scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8209" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.47, + 0.121 + ], + "angle": 0, + "content": "2.2. Adversarial Training Methods and Robust Generalization" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.129, + 0.47, + 0.265 + ], + "angle": 0, + "content": "Adversarial training [24, 25, 34] is shown to be one of the most effective approaches to improve the robustness of DNN models. In practice, the PGD attack is commonly deployed for evaluation because of its strong attack ability in white-box settings and the work in [24] formulates the adversarial training as a min-max problem. [6, 29, 35] show that synthesized data can significantly boost robustness for the image classification task. Thus, we select similar methods as baselines for comparison in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.266, + 0.471, + 0.448 + ], + "angle": 0, + "content": "Recent research shows that a specific type of adversarial attack is not sufficient to represent the diverse space of adversarial examples and many adversarially trained models are only robust to specific attacks. This does limit the application of adversarial training in practice, especially for trajectory regression tasks due to its nature of long-tailed distribution. The works in [33, 49, 37] try to explain and improve the robust generalization in various perspectives such as sample complexity and latent feature representation. In this work, we demonstrate that robust generalization for trajectory prediction can be enhanced by explicitly introducing disentangled and semantic features in the latent space." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.457, + 0.408, + 0.471 + ], + "angle": 0, + "content": "2.3. Adversarial Autoencoder Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.48, + 0.47, + 0.722 + ], + "angle": 0, + "content": "The adversarial autoencoder (AAE) is a variant of the variational autoencoders (VAE) [14, 13], which provides a principled method for jointly learning deep latent-variable models and corresponding inference models using stochastic gradient descent. AAE imposes various distributions on the latent vector by utilizing adversarial learning instead of KL divergence. Due to the flexibility of adversarial learning, AAE is superior to VAE in terms of imposing complicated distributions over latent space. For adversarial robustness, the works in [27, 41] demonstrate that the disentangled latent representations produce VAEs that are more robust to adversarial attacks. In our work, we design an AAE-based architecture that can be added after the feature extractor of prediction modules. We utilize this architecture to model diverse semantic features and enhance disentanglement in the latent space." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.736, + 0.468, + 0.771 + ], + "angle": 0, + "content": "3. Proposed Adversarial Training Method for Trajectory Prediction" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.778, + 0.468, + 0.809 + ], + "angle": 0, + "content": "3.1. Domain Knowledge-guided Semi-supervised Architecture" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.817, + 0.235, + 0.832 + ], + "angle": 0, + "content": "3.1.1 Overall Design" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.84, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Unlike adversarial training for image classification tasks, there are no class labels to guide the training for trajectory prediction, and the attack patterns in the trajectory regression are more random. We thus introduce domain knowl" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "edge to facilitate the modeling of the semantic information in both benign and adversarial cases, based on an AAE architecture. The model learns the directional semantic latent vector in a semi-supervised way because the ground truths are only available for limited scenarios but their distributions can be derived from domain knowledge and statistics. Therefore, our latent space modeling contains two levels – unsupervised distribution modeling and semi-supervised learning when labels are available." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.228, + 0.892, + 0.483 + ], + "angle": 0, + "content": "The architecture of our proposed model is shown in Fig. 3. The feature extractor [16] utilizes a one-dimensional dilated convolutional neural network to obtain the embedding of the time-series trajectory and uses a graph neural network to model the lane context and interaction between objects. The encoder maps high-dimensional features to the semantics-guided latent space with distribution regularization and semi-supervised training. In particular, the latent space is divided into three parts: longitudinal features \\( z_{lon} \\), following one-dimensional log-normal distribution, lateral features \\( z_{lat} \\), following three-dimensional categorical distribution and remaining features, following Gaussian distribution [12]. Finally, the decoder maps semantic vectors along with other disentangled latent vectors to the future trajectories. Note that we develop the AAE instead of traditional VAE architecture to model these different and complex distributions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.485, + 0.892, + 0.665 + ], + "angle": 0, + "content": "In the attacked scenarios, the impact of attacks will be decomposed to different latent vectors and the attack patterns will be explicitly modeled by the semantic features. Let us take the lateral directional vector as an example. If the attack is not targeted at the lateral dimension, the encoder will decompose the attack effect into other vectors and the mapping for lateral direction will remain stable. Otherwise, if the attack causes errors in the lateral vector, the feature extractor and encoder will be adversarial trained on the label of the direction and the corresponding mapping from the latent distribution to the final trajectories will also be updated." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.667, + 0.891, + 0.741 + ], + "angle": 0, + "content": "Compared to existing trajectory prediction works that only apply adversarial training on the final trajectory waypoints, our proposed method is designed to capture semantic features in the latent space, and it can benefit the adversarial robustness and its generalization in various aspects:" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.755, + 0.891, + 0.814 + ], + "angle": 0, + "content": "- The architecture maps the high-dimensional features into disentangled latent space that can decompose the attacks into different orthogonal patterns. Such disentanglement will boost the adversarial robustness [41]." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.826, + 0.891, + 0.899 + ], + "angle": 0, + "content": "- The semantic vectors provide more context information and interpretable labels for the adversarial training and the semi-supervised learning makes the adversarial training process more efficient and easier to generalize to the unseen scenarios." + }, + { + "type": "list", + "bbox": [ + 0.52, + 0.755, + 0.891, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8210" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.159, + 0.089, + 0.817, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.367, + 0.892, + 0.443 + ], + "angle": 0, + "content": "Figure 3: The overall adversarial training pipeline for the proposed semi-supervised semantics-guided architecture. We leverage the feature extractor in [16]. In the latent space, the D nets represent the discriminator that models the latent distributions by distinguishing the samples of real distribution from the predicted ones. When the semantic labels are available in the scenarios, they are used to adversarially train the semantic vectors (parts in green). Finally, we conduct adversarial training on attacked trajectories." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.47, + 0.469, + 0.501 + ], + "angle": 0, + "content": "- The encoder-decoder architecture can filter high-frequency noise by dimension reduction." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.512, + 0.469, + 0.543 + ], + "angle": 0, + "content": "The detailed design and optimization of the architecture are described in the following." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.561, + 0.449, + 0.576 + ], + "angle": 0, + "content": "3.1.2 Semi-supervised Semantic Features Modeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.469, + 0.734 + ], + "angle": 0, + "content": "To model the high-level semantic information in the trajectory prediction task, it is natural to decompose the trajectories into two dimensions – longitudinal and lateral directions. We want to utilize domain knowledge to guide the modeling by providing representative metrics and prior distributions. In addition to the semantic features, the architecture also maps high-dimensional information into latent vectors in Gaussian distributions to represent other low-level and more random features. We will mainly explain the semantic feature modeling below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.902 + ], + "angle": 0, + "content": "In the longitudinal direction, speed and acceleration are often used to model the vehicle's dynamics but their values are always changing and do not contain enough semantic information. In our model, we apply the time headway to effectively extract the longitudinal feature, which measures the time difference between two successive vehicles when they cross a given point. The time headway represents a relatively stable behavior pattern for a certain agent and takes interaction with other vehicles into consideration. Recent works [44, 22] also use time headway as a measurement of aggressiveness in specific scenarios such as lane changing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.892, + 0.636 + ], + "angle": 0, + "content": "or merging. In our proposed architecture, the model will represent the time headway as a one-dimensional vector in the latent space. In both benign and adversarial cases, the encoder will be trained by the regularization loss to force the longitudinal feature to a certain distribution that the time headway follows in statistics. Prior works [7] in the transportation domain show that the time headway in urban scenarios can be best described with the Log-normal distribution. We verify this and estimate the parameters in the Argoverse 1 motion forecasting dataset [4]. The longitudinal feature follows the distribution shown in Eq. (2):" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.65, + 0.892, + 0.683 + ], + "angle": 0, + "content": "\\[\nf (x) = \\frac {1}{x \\sigma \\sqrt {2 \\pi}} e ^ {- \\frac {1}{2} \\left(\\frac {\\ln (x) - \\mu}{\\sigma}\\right) ^ {2}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mu\\) is the location parameter and \\(\\sigma\\) is the scale parameter. We can explicitly obtain the true time headway values for semi-supervised training when there is observable interaction between the attacked target and the front vehicle. We consider the semi-supervised longitudinal feature encoding as a regression problem and optimize it by minimizing the mean square error. For the lateral directional features, we represent them by three simple but effective classes: moving forward, turning/changing lanes to the left, and turning/changing lanes to the right. These three intentions are discrete by nature and we model them with the categorical distribution. In the adversarial training process, only vehicles with clear intentions in a long enough time frame will be labeled and we utilize the cross-entropy to optimize this" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8211" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.203, + 0.105 + ], + "angle": 0, + "content": "classification task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.47, + 0.198 + ], + "angle": 0, + "content": "For all the semantic and Gaussian latent variables, they are regularized to the target distribution by the adversarial generation loss in Eq. (3). The discriminators are trained to maximize the log probability of real latent samples \\( s \\) and the log of the inverse probability for fake latent samples, as in Eq. (4):" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.207, + 0.47, + 0.245 + ], + "angle": 0, + "content": "\\[\nL o s s _ {G} (x) = \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\log \\left(1 - D _ {i} (G (x))\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.249, + 0.469, + 0.266 + ], + "angle": 0, + "content": "\\[\n\\operatorname {L o s s} _ {D _ {i}} (x, s) = \\log D _ {i} \\left(s _ {i}\\right) + \\log \\left(1 - D _ {i} (G (x))\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.275, + 0.47, + 0.321 + ], + "angle": 0, + "content": "where \\( x \\) is the high-dimensional features and \\( m \\) is the number of different kinds of latent vectors. \\( G \\) and \\( D \\) are encoders and distribution discriminators, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.329, + 0.338, + 0.345 + ], + "angle": 0, + "content": "3.2. Adversarial Training Process" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.352, + 0.356, + 0.367 + ], + "angle": 0, + "content": "3.2.1 Adversarial Training Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.376, + 0.47, + 0.557 + ], + "angle": 0, + "content": "For each sample, we utilize the PGD attack to generate the adversarial trajectory only for the target vehicle and keep other surrounding vehicles' original trajectories. This constrains the adversarial attack's impact on the whole scenario. If the error between the prediction under attack and the ground truth is greater than a threshold, we consider the attack a successful one and conduct adversarial training on this sample. Since the perturbation is very small, we consider the real future trajectory as the ground truth \\( y_{i} \\) for the adversarial training and optimize the whole pipeline with L1-smooth loss in Eq. (5). This is a general but somewhat naive adversarial training process." + }, + { + "type": "equation", + "bbox": [ + 0.079, + 0.566, + 0.469, + 0.619 + ], + "angle": 0, + "content": "\\[\n\\operatorname {L o s s} _ {\\text {t r a j}} \\left(y _ {i}, \\hat {y} _ {i}\\right) = \\left\\{ \\begin{array}{l l} 0. 5 \\left(y _ {i} - \\hat {y} _ {i}\\right) ^ {2} & \\text {i f} \\| y _ {i} - \\hat {y} _ {i} \\| < 1, \\\\ \\| y _ {i} - \\hat {y} _ {i} \\| - 0. 5 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.47, + 0.724 + ], + "angle": 0, + "content": "Thus, to further facilitate the adversarial training, we exploit semantic features and their corresponding labels in our proposed architecture. The encoder is optimized to minimize the mean square error of longitudinal features and the cross entropy of lateral features between the ground truths and predictions in the latent space. The semi-supervised loss function is shown as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.731, + 0.469, + 0.772 + ], + "angle": 0, + "content": "\\[\nL o s s _ {S e m i} (z, g) = - \\sum_ {i = 1} ^ {3} g _ {l a t} \\log z _ {l a t} + \\left(g _ {l o n} - z _ {l o n}\\right) ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.824 + ], + "angle": 0, + "content": "where \\(z\\) represents the predicted semantic vectors when under attack and \\(g\\) represents the ground truth in the benign scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Moreover, we further adapt the adversarial training process with lateral semantic vectors because the lateral directional prediction can be regarded as a classification problem with clear behavior meaning. When the adversarial example leads to the wrong classification of lateral behavior, we" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "will set higher weights of the semi-supervised loss for the adversarial training. In this way, our model will first guarantee the correctness of high-level semantic prediction and then tune the regression error, which could help avoid significant adversarial deviation and enhance the generalization performance." + }, + { + "type": "code_caption", + "bbox": [ + 0.502, + 0.209, + 0.816, + 0.226 + ], + "angle": 0, + "content": "Algorithm 1 Our Adversarial Training Pipeline" + }, + { + "type": "algorithm", + "bbox": [ + 0.509, + 0.228, + 0.895, + 0.578 + ], + "angle": 0, + "content": "1: Initialize: feature extractor \\(F\\), AAE encoder \\(G\\), decoder \\(R\\), discriminator \\(D_{i}\\), target distribution \\(p_{i}\\), \\(i = 1,2,3\\), adversarial example generator \\(Adv\\). \n2: Input: past trajectories \\(t\\), future trajectories ground truth \\(y\\), map context \\(c\\), prediction model \\(m\\). \n3: for each sample do \n4: Generate adversarial trajectory for target vehicle \\(t_{adv} = Adv(t,y,m,c)\\). \n5: Input \\(t_{adv}\\) to the predictor. \\(\\hat{y}', \\hat{z}' = m(t_{adv},c)\\), where \\(z = \\{z_{lat},z_{lon},z_{gaussian}\\}\\). \n6: if err \\((\\hat{y}', y) > Threshold\\) then \n7: if \\(\\text{argmax}(z_{lat}) \\neq \\text{argmax}(z_{lat}')\\) then \n8: Update model \\(m\\) with higher weights of semi-supervised learning loss \\(Loss_{semi}\\) in Eq. (6). \n9: Update model \\(m\\) with \\(Loss_{D}\\) in Eq. (4), \\(Loss_{G}\\) in Eq. (3), and \\(Loss_{traj}\\) in Eq. (5). \n10: else \n11: Update model \\(m\\) with \\(Loss_{D}\\) in Eq. (4), \\(Loss_{G}\\) in Eq. (3), \\(Loss_{semi}\\) in Eq. (6), and \\(Loss_{traj}\\) in Eq. (5). \n12: end if \n13: end if \n14: end for" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.655, + 0.794, + 0.67 + ], + "angle": 0, + "content": "3.2.2 Balance Accuracy and Robustness" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In our preliminary experiments, we notice a trade-off between standard accuracy and adversarial robustness. A similar phenomenon has been observed in classification tasks [46, 28]. Methods such as TRADES [46], robust self-training [3] and MixUp [47, 1] have been proposed to balance such trade-off. However, there are few methods that can be applied to trajectory prediction because such time-series regression problems have no class labels and are more sensitive to errors introduced by augmented data. In this work, we utilize the MixUp [47] technique to mix the adversarial scenarios and benign scenarios in the adversarial training process. The experiments demonstrate that a balance between adversarial robustness and standard accuracy can be achieved in trajectory prediction, as shown later." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8212" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.121 + ], + "angle": 0, + "content": "Table 2: Comparison of different defense methods when under various attacks and in the benign case. For the adversarial training methods, we calculate the mean error of models adversarially trained on different attacks." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.131, + 0.872, + 0.283 + ], + "angle": 0, + "content": "
MethodsADE AttackLateral AttackLongitudinal AttackBenign
ADE/mLateral Error/mLongitudinal Error/mADE/m
DatasetArgo1Argo2ApolloArgo1Argo2ApolloArgo1Argo2ApolloArgo1Argo2Apollo
Original Model5.174.523.931.661.271.543.783.173.421.430.791.78
Train-time Smo. [48]4.674.323.861.451.271.493.233.003.411.500.832.11
Test-time Smo. [48]4.323.513.490.750.891.093.542.423.051.681.232.07
Heuristic Aug. [29]4.623.523.841.091.211.343.244.243.211.540.972.01
Data-driven Aug. [29]4.534.433.530.740.780.542.732.613.022.501.992.24
Standard AT3.793.673.680.670.641.152.441.893.051.671.061.87
SSAT (ours)3.512.672.900.610.530.411.761.441.261.751.201.87
Mixup-SSAT (ours)3.733.333.170.720.510.532.131.631.351.641.071.86
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.307, + 0.283, + 0.323 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.331, + 0.254, + 0.348 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.355, + 0.185, + 0.369 + ], + "angle": 0, + "content": "4.1.1 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.378, + 0.471, + 0.545 + ], + "angle": 0, + "content": "We train and evaluate different defense methods using three popular benchmarks - Argoverse 1 [4], Argoverse 2 [42], and ApolloScape datasets [9]. The datasets contain more than \\(250\\mathrm{k}\\) real driving scenarios in different cities, such as Miami and Pittsburgh. For Argoverse 1 and Argoverse 2, each scenario consists of a road graph and multiple agents' trajectories sampled at a frequency of \\(10\\mathrm{Hz}\\). We choose 20 waypoints as the history trajectory and the models will predict 30 waypoints in the future. Scenarios in ApolloScape are simpler. They have no maps but 6 waypoints for both history and future trajectories." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.563, + 0.24, + 0.579 + ], + "angle": 0, + "content": "4.1.2 Attack Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.587, + 0.47, + 0.723 + ], + "angle": 0, + "content": "In the experiments, we study three different types of attacks [48] to the vehicle trajectory prediction algorithms - lateral directional attack (shift to the right), longitudinal directional attack (shift forward), and ADE attack (deviate randomly). More details of them are in Section 2.1. The trajectory prediction models are adversarially trained on the three types of attacks, respectively. We constrain the maximum deviation between the attacked and the benign input trajectories to be 1 meter." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.251, + 0.757 + ], + "angle": 0, + "content": "4.1.3 Training Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Since our architecture is an encoder-decoder module that can be combined with different feature extractors, we first fine-tune the model on the benign data. In the experiments, we use the feature extractor from LaneGCN [16], an attention-based graph neural network. We notice that the AAE architecture introduces a slight accuracy drop on the benign data, mainly due to the dimension reduction. For adversarial training, we train prediction models on adversarial samples generated from scenarios in datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.308, + 0.807, + 0.324 + ], + "angle": 0, + "content": "4.2. Experimental Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.331, + 0.895, + 0.572 + ], + "angle": 0, + "content": "In this section, we conduct experiments with various defense methods under different patterns of attacks, including our semi-supervised semantics-guided (SSAT) method and a Mixup-SSAT method that combines SSAT with the MixUp technique for balancing standard accuracy and adversarial robustness. In the following, we first compare the average robustness improvement among various defense methods, which demonstrates the advantage of SSAT in improving robustness under various types of attacks and the effectiveness of Mixup-SSAT in balancing robustness and accuracy. Then, we show that SSAT can significantly enhance the robust generalization to unseen types of attacks. In addition, we evaluate an unsupervised version of SSAT to explicitly show how the semi-supervised semantic-guided latent space modeling can boost the adversarial robustness, which also serves as an ablation study." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.59, + 0.771, + 0.605 + ], + "angle": 0, + "content": "4.2.1 Effectiveness of SSAT Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.894, + 0.825 + ], + "angle": 0, + "content": "We compare our SSAT and Mixup-SSAT methods with the original model and five different defense methods, including train-time smoothing [48], test-time smoothing [48], heuristic data augmentation [29], data-driven augmentation [29], and the standard adversarial training (Standard AT). All the models share the same feature extractor LaneGCN in this setting. Note that we compare to the two data augmentation methods as they are effective for image classification tasks [6, 29, 35]. For the data-driven augmentation, we design an additional decoder to augment input trajectories and it can generate more inputs by adding Gaussian noises to the latent vectors of the real inputs. For the heuristic augmentation, we simply add random perturbation to the benign inputs, with the same constraints of maximum deviation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Table 2 shows the prediction errors of different methods when under various types of attacks and in the benign case (note that when under lateral and longitudinal attacks, we measure lateral and longitudinal errors). We can see that our SSAT method significantly outperforms all other de" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8213" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.101, + 0.322, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.24, + 0.228, + 0.251 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.101, + 0.585, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.24, + 0.492, + 0.251 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.101, + 0.848, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.24, + 0.754, + 0.251 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.268, + 0.895, + 0.314 + ], + "angle": 0, + "content": "Figure 4: Adversarial robustness comparisons of the original model, the standard adversarial training on final trajectories (Standard AT), and the SSAT method proposed by us. Figures (a), (b), (c) show results from adversarial training (AT) that targets ADE attacks, lateral attacks, and longitudinal attacks, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.468, + 0.382 + ], + "angle": 0, + "content": "fense methods in improving the robustness of trajectory prediction. Compared with the original model, SSAT can reduce the prediction error by \\(32\\% - 73\\%\\) when under different types of attacks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.468, + 0.506 + ], + "angle": 0, + "content": "Moreover, while our SSAT method improves robustness, we also observe a drop in standard accuracy in the benign case. Mixup-SSAT enables effective trade-off between these two objectives (i.e., better performance than SSAT in benign cases at the expense of worse performance under attacks), by setting the mixup ratio between adversarial and benign examples to different values (the results in Table 2 are based on a mixup ratio of 2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.468, + 0.599 + ], + "angle": 0, + "content": "We also notice that both data-driven and heuristic data augmentation methods offer very limited improvement over the original model. This is likely due to the challenge of regression tasks with rich context, which makes augmentation methods perform poorer than they do for image classification tasks." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.617, + 0.468, + 0.663 + ], + "angle": 0, + "content": "Table 3: Comparison of different methods when adversarial training is conducted for ADE attack, and tested on ADE, lateral, and longitudinal attacks." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.674, + 0.437, + 0.738 + ], + "angle": 0, + "content": "
MethodsPrediction Error/m, w/o AT → with AT
ADE AttackLat AttackLon Attack
Standard AT5.17 → 3.761.58 → 0.833.59 → 2.66
SSAT5.16 → 3.241.66 → 0.733.78 → 1.45
Unsup-SSAT5.16 → 3.481.66 → 0.743.78 → 1.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.778, + 0.468, + 0.824 + ], + "angle": 0, + "content": "Table 4: Comparison of different methods when adversarial training is conducted for lateral attack, and tested on ADE, lateral, and longitudinal attacks." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.835, + 0.437, + 0.898 + ], + "angle": 0, + "content": "
MethodsPrediction Error/m, w/o AT → with AT
ADE AttackLat AttackLon Attack
Standard AT5.17 → 4.281.58 → 0.593.59 → 3.01
SSAT5.16 → 4.021.66 → 0.593.78 → 2.52
Unsup-SSAT5.16 → 4.111.66 → 0.693.78 → 2.54
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.323, + 0.892, + 0.353 + ], + "angle": 0, + "content": "4.2.2 Effectiveness of SSAT in Robust Generalization on Different Types of Attacks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.361, + 0.892, + 0.633 + ], + "angle": 0, + "content": "We observe that there is an adversarial robust generalization gap when the training and test are under different types of attacks. The comparisons in Tables 3, 4, 5 show that our SSAT method is better generalized to unseen types of attacks, compared to the standard adversarial training. In every training scenario, our method is more robust to the various unseen patterns of attacks. For instance, Table 3 shows that when applying SSAT to train under the random ADE attack, its results outperform other models on all seen (i.e., ADE) and unseen (i.e., lateral and longitudinal) types of attacks, which demonstrates that our SSAT method can effectively decompose and learn semantic features from random ADE attacks. Tables 4 and 5 show similar trends, where our SSAT methods are better at defending against unseen attacks and mitigating overfitting on specific patterns of attacks. Fig. 4 further visualizes the results from these three tables for the original model, the standard adversarial training (Standard AT), and SSAT." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.646, + 0.892, + 0.692 + ], + "angle": 0, + "content": "Table 5: Comparison of different methods when adversarial training is conducted for longitudinal attack, and tested on ADE, lateral, and longitudinal attacks." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.703, + 0.86, + 0.767 + ], + "angle": 0, + "content": "
MethodsPrediction Error/m, w/o AT → with AT
ADE AttackLat AttackLon Attack
Standard AT5.17 → 3.331.58 → 0.603.59 → 1.64
SSAT5.16 → 3.281.66 → 0.523.78 → 1.30
Unsup-SSAT5.16 → 3.401.66 → 0.613.78 → 1.49
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.808, + 0.803 + ], + "angle": 0, + "content": "4.2.3 Impact from Latent Space Modeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We also conduct adversarial training that only has regularization on the latent distributions but without supervision on latent vectors. We name it Unsup-SSAT. The comparisons between the standard adversarial training (Standard AT) and Unsup-SSAT in Tables 3, 4, and 5 demonstrate that even without labels, the partial disentanglement and" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8214" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "distribution modeling in Unsup-SSAT will benefit the adversarial training on trajectory prediction and outperform the baseline Standard AT in most cases. However, when compared with SSAT, we find that the adversarial robustness will be further improved with the extra labels in the semi-supervised phase (in practice, we often have access to those labels)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.211, + 0.196, + 0.227 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.237, + 0.473, + 0.404 + ], + "angle": 0, + "content": "In this work, we propose an adversarial training method for trajectory prediction. To tackle the challenge of random inputs with rich context, diverse types of attacks, and lack of class labels, we develop a novel AAE architecture that exploits the disentanglement and semantic features for enhancing model robustness and its generalization. Our proposed SSAT method significantly outperforms a number of baselines from the literature, reducing the prediction errors under attacks by up to \\(32\\% - 73\\%\\) when compared with the original prediction models. Our method is also shown to be effective for defending against unseen attacks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.417, + 0.229, + 0.433 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.442, + 0.47, + 0.487 + ], + "angle": 0, + "content": "We gratefully acknowledge the support from NSF grants 1834701, 1724341, 2038853, 1929771, 2145493, 1932464, and ONR grant N00014-19-1-2496." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.501, + 0.174, + 0.517 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.526, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[1] Guillaume P Archambault, Yongyi Mao, Hongyu Guo, and Richong Zhang. Mixup as directional adversarial training. arXiv preprint arXiv:1906.06875, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.597, + 0.472, + 0.674 + ], + "angle": 0, + "content": "[2] Yulong Cao, Danfei Xu, Xinshuo Weng, Zhuoqing Mao, Anima Anandkumar, Chaowei Xiao, and Marco Pavone. Robust trajectory prediction against adversarial attacks. In 6th Annual Conference on Robot Learning." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.684, + 0.473, + 0.745 + ], + "angle": 0, + "content": "[3] Yair Carmon, Aditi Raghunathan, Ludwig Schmidt, John C Duchi, and Percy S Liang. Unlabeled data improves adversarial robustness. Advances in Neural Information Processing Systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.754, + 0.472, + 0.86 + ], + "angle": 0, + "content": "[4] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8748-8757, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[5] Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.526, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.092, + 0.894, + 0.152 + ], + "angle": 0, + "content": "Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11525-11533, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.162, + 0.895, + 0.237 + ], + "angle": 0, + "content": "[6] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 34:4218-4233, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.247, + 0.895, + 0.308 + ], + "angle": 0, + "content": "[7] Duy-Hung Ha, Maurice Aron, and Simon Cohen. Time headway variable and probabilistic modeling. Transportation Research Part C: Emerging Technologies, 25:181-201, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.318, + 0.895, + 0.379 + ], + "angle": 0, + "content": "[8] Michael W Hancock and Bud Wright. A policy on geometric design of highways and streets. American Association of State Highway and Transportation Officials: Washington, DC, USA, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.388, + 0.895, + 0.48 + ], + "angle": 0, + "content": "[9] Xinyu Huang, Xinjing Cheng, Qichuan Geng, Binbin Cao, Dingfu Zhou, Peng Wang, Yuanqing Lin, and Ruigang Yang. The apolloscape dataset for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 954-960, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.488, + 0.892, + 0.549 + ], + "angle": 0, + "content": "[10] Ruochen Jiao, Juyang Bai, Xiangguo Liu, Takami Sato, Xiaowei Yuan, Qi Alfred Chen, and Qi Zhu. Learning representation for anomaly detection of vehicle trajectories. 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.559, + 0.895, + 0.636 + ], + "angle": 0, + "content": "[11] Ruochen Jiao, Hengyi Liang, Takami Sato, Junjie Shen, Qi Alfred Chen, and Qi Zhu. End-to-end uncertainty-based mitigation of adversarial attacks to automated lane centering. In 2021 IEEE Intelligent Vehicles Symposium (IV), pages 266-273. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.644, + 0.895, + 0.735 + ], + "angle": 0, + "content": "[12] Ruochen Jiao, Xiangguo Liu, Bowen Zheng, Dave Liang, and Qi Zhu. Tae: A semi-supervised controllable behavior-aware trajectory generator and predictor. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 12534-12541. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.895, + 0.79 + ], + "angle": 0, + "content": "[13] Diederik Kingma and Max Welling. An introduction to variational autoencoders. arXiv preprint arXiv:1906.02691, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.8, + 0.895, + 0.845 + ], + "angle": 0, + "content": "[14] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.856, + 0.895, + 0.902 + ], + "angle": 0, + "content": "[15] Dasom Lee and David J Hess. Regulations for on-road testing of connected and automated vehicles: Assessing the potential for global safety harmonization." + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.092, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8215" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.114, + 0.092, + 0.468, + 0.121 + ], + "angle": 0, + "content": "Transportation Research Part A: Policy and Practice, 136:85-98, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.13, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[16] Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph representations for motion forecasting. In European Conference on Computer Vision, pages 541-556. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.213, + 0.47, + 0.302 + ], + "angle": 0, + "content": "[17] C. Lin, Bowen Zheng, Qi Zhu, and Alberto Sangiovanni-Vincentelli. Security-Aware Design Methodology and Optimization for Automotive Systems. ACM Transactions on Design Automation of Electronic Systems (TODAES), 21(1):18:1-18:26, December 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.311, + 0.469, + 0.401 + ], + "angle": 0, + "content": "[18] Xiangguo Liu, Chao Huang, Yixuan Wang, Bowen Zheng, and Qi Zhu. Physics-aware safety-assured design of hierarchical neural network based planner. In 2022 ACM/IEEE 13th International Conference on Cyber-Physical Systems (ICCPS), pages 137-146. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.41, + 0.469, + 0.485 + ], + "angle": 0, + "content": "[19] Xiangguo Liu, Ruochen Jiao, Yixuan Wang, Yimin Han, Bowen Zheng, and Qi Zhu. Safety-assured speculative planning with adaptive prediction. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.493, + 0.469, + 0.568 + ], + "angle": 0, + "content": "[20] Xiangguo Liu, Ruochen Jiao, Bowen Zheng, Dave Liang, and Qi Zhu. Safety-driven interactive planning for neural network-based lane changing. In Proceedings of the 28th Asia and South Pacific Design Automation Conference, pages 39-45, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.682 + ], + "angle": 0, + "content": "[21] Xiangguo Liu, Yunpeng Luo, Anthony Goeckner, Trishna Chakraborty, Ruochen Jiao, Ningfei Wang, Yixuan Wang, Takami Sato, Qi Alfred Chen, and Qi Zhu. Invited: Waving the double-edged sword: Building resilient cavs with edge and cloud computing. In Proceedings of the 60th Annual Design Automation Conference, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.749 + ], + "angle": 0, + "content": "[22] Xiangguo Liu, Neda Masoud, and Qi Zhu. Impact of sharing driving attitude information: A quantitative study on lane changing. In 2020 IEEE Intelligent Vehicles Symposium (IV), pages 1998-2005. IEEE, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.469, + 0.833 + ], + "angle": 0, + "content": "[23] Yicheng Liu, Jinghuai Zhang, Liangji Fang, Qinhong Jiang, and Bolei Zhou. Multimodal motion prediction with stacked transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7577-7586, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.841, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[24] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "[25] Pratyush Maini, Eric Wong, and Zico Kolter. Adversarial robustness against the union of multiple perturbation models. In International Conference on Machine Learning, pages 6640-6650. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.16, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[26] Alireza Makhzani, Jonathon Shlens, Navdeep Jaitly, Ian Goodfellow, and Brendan Frey. Adversarial autoencoders. arXiv preprint arXiv:1511.05644, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.214, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[27] Emile Mathieu, Tom Rainforth, Nana Siddharth, and Yee Whye Teh. Disentangling disentanglement in variational autoencoders. In International Conference on Machine Learning, pages 4402-4412. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.298, + 0.892, + 0.358 + ], + "angle": 0, + "content": "[28] Aditi Raghunathan, Sang Michael Xie, Fanny Yang, John Duchi, and Percy Liang. Understanding and mitigating the tradeoff between robustness and accuracy. arXiv preprint arXiv:2002.10716, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.367, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[29] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan A Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Fixing data augmentation to improve adversarial robustness. arXiv preprint arXiv:2103.01946, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.436, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[30] Leslie Rice, Eric Wong, and Zico Kolter. Overfitting in adversarially robust deep learning. In International Conference on Machine Learning, pages 8093-8104. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[31] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory++: Dynamically-feasible trajectory forecasting with heterogeneous data. In European Conference on Computer Vision, pages 683-700. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.679 + ], + "angle": 0, + "content": "[32] Takami Sato, Junjie Shen, Ningfei Wang, Yunhan Jia, Xue Lin, and Qi Alfred Chen. Dirty road can attack: Security of deep learning based automated lane centering under {Physical-World} attack. In 30th USENIX Security Symposium (USENIX Security 21), pages 3309–3326, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.749 + ], + "angle": 0, + "content": "[33] Ludwig Schmidt, Shibani Santurkar, Dimitris Tsipras, Kunal Talwar, and Aleksander Madry. Adversarily robust generalization requires more data. Advances in neural information processing systems, 31, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.757, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[34] Lukas Schott, Jonas Rauber, Matthias Bethge, and Wieland Brendel. Towards the first adversarially robust neural network model on mnist. arXiv preprint arXiv:1805.09190, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[35] Vikash Sehwag, Saeed Mahloujifar, Tinashe Handina, Sihui Dai, Chong Xiang, Mung Chiang, and Prateek Mittal. Robust learning meets generative models: Can proxy distributions improve adversarial robustness? arXiv preprint arXiv:2104.09425, 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8216" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "[36] Amolika Sinha, Sai Chand, Kasun P Wijayaratna, Navreet Virdi, and Vinayak Dixit. Comprehensive safety assessment in mixed fleets with connected and automated vehicles: A crash severity and rate evaluation of conventional vehicles. Accident Analysis & Prevention, 142:105567, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.47, + 0.256 + ], + "angle": 0, + "content": "[37] Chuanbiao Song, Kun He, Jiadong Lin, Liwei Wang, and John E Hopcroft. Robust local features for improving the generalization of adversarial training. arXiv preprint arXiv:1909.10147, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.265, + 0.471, + 0.356 + ], + "angle": 0, + "content": "[38] Dong Su, Huan Zhang, Hongge Chen, Jinfeng Yi, Pin-Yu Chen, and Yupeng Gao. Is robustness the cost of accuracy? a comprehensive study on the robustness of 18 deep image classification models. In Proceedings of the European Conference on Computer Vision (ECCV), pages 631-648, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.366, + 0.469, + 0.428 + ], + "angle": 0, + "content": "[39] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.438, + 0.469, + 0.513 + ], + "angle": 0, + "content": "[40] Zhilu Wang, Chao Huang, and Qi Zhu. Efficient global robustness certification of neural networks via interleaving twin-network encoding. In DATE'22: Proceedings of the Conference on Design, Automation and Test in Europe, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.524, + 0.469, + 0.584 + ], + "angle": 0, + "content": "[41] Matthew Willetts, Alexander Camuto, Tom Rainforth, Stephen Roberts, and Chris Holmes. Improving vaes' robustness to adversarial attack. arXiv preprint arXiv:1906.00230, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[42] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, et al. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.727, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[43] Dongxian Wu, Shu-Tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization. Advances in Neural Information Processing Systems, 33:2958-2969, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.799, + 0.469, + 0.86 + ], + "angle": 0, + "content": "[44] Hongtao Yu, H Eric Tseng, and Reza Langari. A human-like game theory-based controller for automatic lane changing. Transportation Research Part C: Emerging Technologies, 88:140-158, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.87, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[45] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M Kitani. Agentformer: Agent-aware transformers for" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9813-9823, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.147, + 0.892, + 0.223 + ], + "angle": 0, + "content": "[46] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[47] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.303, + 0.892, + 0.393 + ], + "angle": 0, + "content": "[48] Qingzhao Zhang, Shengtuo Hu, Jiachen Sun, Qi Alfred Chen, and Z Morley Mao. On adversarial robustness of trajectory prediction for autonomous vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15159-15168, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.403, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[49] Shufei Zhang, Zhuang Qian, Kaizhu Huang, Qiufeng Wang, Rui Zhang, and Xinping Yi. Towards better robust generalization with shift consistency regularization. In International Conference on Machine Learning, pages 12524-12534. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[50] Bowen Zheng, Hengyi Liang, Qi Zhu, Huafeng Yu, and Chung-Wei Lin. Next generation automotive architecture modeling and exploration for autonomous driving. In 2016 IEEE Computer Society Annual Symposium on VLSI (ISVLSI), pages 53–58, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.633 + ], + "angle": 0, + "content": "[51] Bowen Zheng, Chung-Wei Lin, et al. Design and Analysis of Delay-Tolerant Intelligent Intersection Management. ACM T-CPS, 4(1):3:1-3:27, November 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.892, + 0.735 + ], + "angle": 0, + "content": "[52] Qi Zhu, Chao Huang, Ruochen Jiao, Shuyue Lan, Hengyi Liang, Xiangguo Liu, Yixuan Wang, Zhilu Wang, and Shichao Xu. Safety-assured design and adaptation of learning-enabled autonomous systems. In Proceedings of the 26th Asia and South Pacific Design Automation Conference, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.867 + ], + "angle": 0, + "content": "[53] Qi Zhu, Wenchao Li, Hyoseung Kim, Yecheng Xiang, Kacper Wardega, Zhilu Wang, Yixuan Wang, Hengyi Liang, Chao Huang, Jiameng Fan, and Hyunjong Choi. Know the unknowns: Addressing disturbances and uncertainties in autonomous systems. In Proceedings of the 39th International Conference on Computer-Aided Design, ICCAD '20, New York, NY, USA, 2020. Association for Computing Machinery." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "8217" + } + ] +] \ No newline at end of file diff --git a/2023/Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction/487dfb22-6094-4f31-976c-310213635421_origin.pdf b/2023/Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction/487dfb22-6094-4f31-976c-310213635421_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b70e268b1c7a1db1a8b1fcdbe0e25c356b234f60 --- /dev/null +++ b/2023/Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction/487dfb22-6094-4f31-976c-310213635421_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd9f5759dc4312ea34bd303a01c9ff25c0a74e735f52a1b95f961d14e8a9e7cb +size 932071 diff --git a/2023/Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction/full.md b/2023/Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7d2140b5c4913e3352691dcac9f1965540c86bf2 --- /dev/null +++ b/2023/Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction/full.md @@ -0,0 +1,347 @@ +# Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction + +Ruochen Jiao + +Northwestern University + +ruochen.jiao@u.northwestern.edu + +Xiangguo Liu + +Northwestern University + +xg.liu@u.northwestern.edu + +Takami Sato + +University of California, Irvine + +takamis@uci.edu + +University of California, Irvine + +alfchen@uci.edu + +Qi Zhu + +Northwestern University + +qzhu@northwestern.edu + +# Abstract + +Predicting the trajectories of surrounding objects is a critical task for self-driving vehicles and many other autonomous systems. Recent works demonstrate that adversarial attacks on trajectory prediction, where small crafted perturbations are introduced to history trajectories, may significantly mislead the prediction of future trajectories and induce unsafe planning. However, few works have addressed enhancing the robustness of this important safety-critical task. In this paper, we present a novel adversarial training method for trajectory prediction. Compared with typical adversarial training on image tasks, our work is challenged by more random input with rich context and a lack of class labels. To address these challenges, we propose a method based on a semi-supervised adversarial autoencoder, which models disentangled semantic features with domain knowledge and provides additional latent labels for the adversarial training. Extensive experiments with different types of attacks demonstrate that our Semi-supervised Semantics-guided Adversarial Training (SSAT1) method can effectively mitigate the impact of adversarial attacks by up to $73\%$ and outperform other popular defense methods. In addition, experiments show that our method can significantly improve the system's robust generalization to unseen patterns of attacks. We believe that such semantics-guided architecture and advancement on robust generalization is an important step for developing robust prediction models and enabling safe decision making. + +# 1. Introduction + +Connected and autonomous vehicles (CAVs) have shown great promise to revolutionize the transportation systems, + +but also face significant concerns regarding their safety and security [15, 36, 21, 17, 50, 51]. Many of these concerns, which also apply to many other autonomous systems, arise from the increasing adoption of advanced deep neural network (DNN)-based machine learning techniques across system sensing, perception, prediction, planning, control, and general decision-making [52, 53]. + +In particular, the adversarial robustness of DNNs has drawn significant attention in recent years. State-of-the-art classifiers can be misled to make erroneous predictions by slight but carefully-optimized perturbations [39], and addressing this is critical to the safety of autonomous systems that employ these classifiers [11, 40]. Besides classification, adversarial attacks are also targeted at various other tasks such as object detection, image generation and natural language processing. To defend against adversarial examples, adversarial training is commonly used to enhance the intrinsic robustness of models and has been shown to be very effective among various defense strategies [24, 25, 34]. On the other hand, the works in [30, 43, 37] show that adversarial training may suffer poor robust generalization on unseen attacks and [33, 24, 38] demonstrate a trade-off between robustness and accuracy. In this paper, we address trajectory prediction, a safety-critical task that is common in self-driving vehicles and other autonomous systems, and propose a novel architecture to enhance its adversarial robustness and robust generalization by introducing semantic features and a semi-supervised adversarial autoencoder (AAE) [26] into the adversarial training process. + +We focus on the trajectory prediction module in autonomous driving, where large DNNs have enabled breakthroughs in recent years. More specifically, an autonomous driving system typically consists of several modules such as perception, localization, prediction, planning (route, behavior, trajectory, and motion planning), and control. The perception module is to detect agents and obstacles based + +on the sensing inputs (e.g., images and 3D point clouds), and the effectiveness and robustness of the perception module are well-studied in the research community. The prediction module is to predict future trajectories of surrounding vehicles based on the observed history trajectories of those vehicles from the perception module and the map context. The trajectory prediction plays a crucial role in understanding the environment and making safety-critical decisions in the following planning module [18, 20, 19]. Recent works [23, 16, 5, 45, 12] have applied various deep learning techniques such as graph neural networks and transformers to this task and achieved impressive performance in terms of reducing average errors. However, few works studied the robustness of vehicle trajectory prediction under adversarial attacks, which is in fact of vital importance because 1) autonomous driving is a safety-critical task by nature, 2) recent work [48, 10] demonstrates that the prediction module is vulnerable to adversarial attacks if surrounding vehicles drive along a seemingly natural but crafted trajectory, and 3) current prediction models are often overfitted on limited patterns in the datasets but suffer the long-tailed distribution of driving scenarios and behaviors. The threat model of the adversarial trajectory is illustrated in an example in Fig. 1, where a crafted input trajectory can intentionally mislead the target vehicle to the wrong prediction and induce dangerous planning decisions. + +![](images/54cf1db9e7e81c910006b20e1b071f79068f13807b7324360d4b129d527e4b56.jpg) +Figure 1: An illustrative example of the threat model: In the upper benign scenario, the target vehicle predicts the future trajectory of a surrounding vehicle based on its history trajectory. In the lower adversarial scenario, the surrounding vehicle is malicious and drives along a crafted (history) trajectory, which may mislead the target vehicle to have a wrong (attacked) prediction of the surrounding vehicle's future trajectory. In this particular case, the target vehicle wrongly predicts that the malicious vehicle will cut into its lane, and thus takes an unsafe emergency braking. + +To defend against adversarial attacks on vehicle trajectory prediction, there are important challenges that are different from the cases of images and audios. First, vehicle trajectory prediction is a time-series regression problem with rich contexts while most existing adversarial attacks + +and corresponding defense methods are targeted at classification tasks. The attack patterns are more random and there are no well-defined class labels, which means that the robust model is difficult to train and generalize. Second, vehicle trajectory can convey semantic and behavior information. For instance, people can infer the behavior of a vehicle such as moving forward or changing lanes from its trajectory. Therefore, as shown in our experiments later, some popular defense methods such as TRADES [46] and data augmentation methods [6, 29, 35] either are inapplicable or have poor performance in trajectory prediction task. In this work, we first propose an adversarial training pipeline for the trajectory prediction task, and then further exploit the semantic features to design a semi-supervised AAE architecture that can be added after the feature extractor, to improve the adversarial robustness and its generalization. The methodology of the proposed architecture such as enhancing disentanglement and defining semantic labels may be further applied to adversarial training for general regression and generation problems. Our main contributions are summarized as follows: + +- We propose a novel adversarial training method against adversarial attacks on trajectory predictions. +- We develop a semi-supervised architecture with domain knowledge and semantic features to enhance the adversarial robustness and its generalization among different types of attack patterns. +- Extensive experiments demonstrate that our proposed method (SSAT) effectively improves the adversarial robustness and outperforms popular defense baselines. +- We further explore to balance the robustness-accuracy trade-off in this task by leveraging the MixUp technique [47], and we also test data augmentation methods for trajectory prediction robustness. + +# 2. Preliminaries + +# 2.1. Adversarial Attacks on Trajectory Prediction + +Recent research [48, 10] demonstrates that trajectory prediction in autonomous driving can be fooled by the adversarial behavior of a surrounding vehicle, where the adversarial behavior is optimized with Projected Gradient Descent (PGD) [24]. The threat model is illustrated in Fig. 1. In this work, we consider such attacks and assume the worst setting, i.e., the attacker has full knowledge of the target system and tries to maximize the attack impact. However, malicious agents cannot directly manipulate anything inside the target vehicle and they can only slightly change their own trajectories to mislead the target vehicle indirectly. + +Different from image classification, trajectory prediction has no class labels, but it has directional information in the + +![](images/daac5439f6a180240c33f5df31b6ceb98f3b8cc786b363d794c8583f936b544d.jpg) + +![](images/c2e5e6a3d52e198afd9da362d85a221d25c792c26f6b407ffcca27f02778a530.jpg) + +![](images/2122c7654d92294b38e41a5fe8a72d17a432d8fa8f453474b7c6cea828732a9e.jpg) + +![](images/a3523bdf036540e85f87e7bd6a8520ba720374a4d2cca6b14741edd372de6b16.jpg) +(a) + +![](images/74c62a6c7f807384e4fd52f213824e25bb86c7094274909743643f05baecaf73.jpg) +(b) +Figure 2: Adversarial trajectories generated by different types of attacks. The figures in the top row are adversarial (red line) and benign (blue line) input history trajectories for prediction and they look very close to human eyes. The figures in the bottom row show the corresponding attacked future trajectory prediction (red dashed line) and the ground truth trajectory in the benign case (green dashed line). The differences between the two are clearly visible, showing the effectiveness of the attacks. Figure (a) is the ADE attack, which will randomly lead to maximum average deviation. Figure (b) is the lateral attack, which mainly leads the vehicle to deviate to the left or right. Figure (c) is the longitudinal attack, which will mainly lead to longitudinal deviation. + +![](images/a8b6ecb4f0eaa094f299713246f3e62c1e7f163d618ea6d2bf04098a0eb65b16.jpg) +(c) + +context, such as moving forward, turning or changing lanes to the right. Therefore, the attacker can conduct targeted attacks besides random ones. In Fig. 2, three types of adversarial attacks are presented and we observe that they can cause significant and directional errors. [48] proposes directional error metrics for the optimization of targeted attacks, as shown in Eq. (1): + +$$ +D (\alpha , R) = \left(p _ {\alpha} - s _ {\alpha}\right) ^ {T} \cdot R \left(s _ {\alpha + 1}, s _ {\alpha}\right), \tag {1} +$$ + +where $\alpha$ denotes the time frame ID. $p$ and $s$ are two-dimensional vectors representing predicted and ground-truth vehicle locations, respectively. $R$ is a function generating the unit vector to a specific direction (lateral or longitudinal). The longitudinal direction is approximated as the vector defined by the adjacent two waypoints of ground truth $s_{\alpha +1}^{n} - s_{\alpha}^{n}$ + +In addition to the directional attack, a random attack can be designed to maximize the Average Displacement Error (ADE), which is the average of the root mean squared error between the predicted waypoints and the ground-truth trajectory waypoints. + +Hard constraints are also applied to the maximum deviation of waypoints during the optimization so as to make the adversarial history trajectory physically feasible and not perform unrealistic behaviors. + +Table 1: Trajectory prediction errors (in meters) in benign or adversarial scenarios under different types of attacks. + +
ModelADELateralLongitudinal
benign/attackbenign/attackbenign/attack
LaneGCN [16]1.32/5.17-0.01/1.58-0.25/3.33
Trajectory++[31]2.69/6.810.107/2.25-0.526 / 3.79
+ +Table 1 quantitatively shows that the predicted trajectories can be seriously deviated by any of the three attack types. Generally, in the US, $0.3\mathrm{m}$ lateral deviation is enough to invade adjacent lanes on local roads [32, 8]. In addition, the final displacement error (the distance between the last points of predicted and ground-truth trajectories) is around two to three times larger than the average error shown here. Therefore, the attacker can apply either random or directional attacks that will greatly challenge the defense methods, especially on the generalization performance. + +[48, 2] propose to utilize approaches such as smoothing and adversarial training on the conditional variational encoder (CVAE) architectures to mitigate the impact of such adversarial attacks. However, they do not explicitly address driving semantics and robustness generalization, which are critical for many real road scenarios. + +# 2.2. Adversarial Training Methods and Robust Generalization + +Adversarial training [24, 25, 34] is shown to be one of the most effective approaches to improve the robustness of DNN models. In practice, the PGD attack is commonly deployed for evaluation because of its strong attack ability in white-box settings and the work in [24] formulates the adversarial training as a min-max problem. [6, 29, 35] show that synthesized data can significantly boost robustness for the image classification task. Thus, we select similar methods as baselines for comparison in our experiments. + +Recent research shows that a specific type of adversarial attack is not sufficient to represent the diverse space of adversarial examples and many adversarially trained models are only robust to specific attacks. This does limit the application of adversarial training in practice, especially for trajectory regression tasks due to its nature of long-tailed distribution. The works in [33, 49, 37] try to explain and improve the robust generalization in various perspectives such as sample complexity and latent feature representation. In this work, we demonstrate that robust generalization for trajectory prediction can be enhanced by explicitly introducing disentangled and semantic features in the latent space. + +# 2.3. Adversarial Autoencoder Architecture + +The adversarial autoencoder (AAE) is a variant of the variational autoencoders (VAE) [14, 13], which provides a principled method for jointly learning deep latent-variable models and corresponding inference models using stochastic gradient descent. AAE imposes various distributions on the latent vector by utilizing adversarial learning instead of KL divergence. Due to the flexibility of adversarial learning, AAE is superior to VAE in terms of imposing complicated distributions over latent space. For adversarial robustness, the works in [27, 41] demonstrate that the disentangled latent representations produce VAEs that are more robust to adversarial attacks. In our work, we design an AAE-based architecture that can be added after the feature extractor of prediction modules. We utilize this architecture to model diverse semantic features and enhance disentanglement in the latent space. + +# 3. Proposed Adversarial Training Method for Trajectory Prediction + +# 3.1. Domain Knowledge-guided Semi-supervised Architecture + +# 3.1.1 Overall Design + +Unlike adversarial training for image classification tasks, there are no class labels to guide the training for trajectory prediction, and the attack patterns in the trajectory regression are more random. We thus introduce domain knowl + +edge to facilitate the modeling of the semantic information in both benign and adversarial cases, based on an AAE architecture. The model learns the directional semantic latent vector in a semi-supervised way because the ground truths are only available for limited scenarios but their distributions can be derived from domain knowledge and statistics. Therefore, our latent space modeling contains two levels – unsupervised distribution modeling and semi-supervised learning when labels are available. + +The architecture of our proposed model is shown in Fig. 3. The feature extractor [16] utilizes a one-dimensional dilated convolutional neural network to obtain the embedding of the time-series trajectory and uses a graph neural network to model the lane context and interaction between objects. The encoder maps high-dimensional features to the semantics-guided latent space with distribution regularization and semi-supervised training. In particular, the latent space is divided into three parts: longitudinal features $z_{lon}$ , following one-dimensional log-normal distribution, lateral features $z_{lat}$ , following three-dimensional categorical distribution and remaining features, following Gaussian distribution [12]. Finally, the decoder maps semantic vectors along with other disentangled latent vectors to the future trajectories. Note that we develop the AAE instead of traditional VAE architecture to model these different and complex distributions. + +In the attacked scenarios, the impact of attacks will be decomposed to different latent vectors and the attack patterns will be explicitly modeled by the semantic features. Let us take the lateral directional vector as an example. If the attack is not targeted at the lateral dimension, the encoder will decompose the attack effect into other vectors and the mapping for lateral direction will remain stable. Otherwise, if the attack causes errors in the lateral vector, the feature extractor and encoder will be adversarial trained on the label of the direction and the corresponding mapping from the latent distribution to the final trajectories will also be updated. + +Compared to existing trajectory prediction works that only apply adversarial training on the final trajectory waypoints, our proposed method is designed to capture semantic features in the latent space, and it can benefit the adversarial robustness and its generalization in various aspects: + +- The architecture maps the high-dimensional features into disentangled latent space that can decompose the attacks into different orthogonal patterns. Such disentanglement will boost the adversarial robustness [41]. +- The semantic vectors provide more context information and interpretable labels for the adversarial training and the semi-supervised learning makes the adversarial training process more efficient and easier to generalize to the unseen scenarios. + +![](images/466c3ad5503ae0455b8dabcff162510d6441254a56aab57a6cbdef8a130aadd4.jpg) +Figure 3: The overall adversarial training pipeline for the proposed semi-supervised semantics-guided architecture. We leverage the feature extractor in [16]. In the latent space, the D nets represent the discriminator that models the latent distributions by distinguishing the samples of real distribution from the predicted ones. When the semantic labels are available in the scenarios, they are used to adversarially train the semantic vectors (parts in green). Finally, we conduct adversarial training on attacked trajectories. + +- The encoder-decoder architecture can filter high-frequency noise by dimension reduction. + +The detailed design and optimization of the architecture are described in the following. + +# 3.1.2 Semi-supervised Semantic Features Modeling + +To model the high-level semantic information in the trajectory prediction task, it is natural to decompose the trajectories into two dimensions – longitudinal and lateral directions. We want to utilize domain knowledge to guide the modeling by providing representative metrics and prior distributions. In addition to the semantic features, the architecture also maps high-dimensional information into latent vectors in Gaussian distributions to represent other low-level and more random features. We will mainly explain the semantic feature modeling below. + +In the longitudinal direction, speed and acceleration are often used to model the vehicle's dynamics but their values are always changing and do not contain enough semantic information. In our model, we apply the time headway to effectively extract the longitudinal feature, which measures the time difference between two successive vehicles when they cross a given point. The time headway represents a relatively stable behavior pattern for a certain agent and takes interaction with other vehicles into consideration. Recent works [44, 22] also use time headway as a measurement of aggressiveness in specific scenarios such as lane changing + +or merging. In our proposed architecture, the model will represent the time headway as a one-dimensional vector in the latent space. In both benign and adversarial cases, the encoder will be trained by the regularization loss to force the longitudinal feature to a certain distribution that the time headway follows in statistics. Prior works [7] in the transportation domain show that the time headway in urban scenarios can be best described with the Log-normal distribution. We verify this and estimate the parameters in the Argoverse 1 motion forecasting dataset [4]. The longitudinal feature follows the distribution shown in Eq. (2): + +$$ +f (x) = \frac {1}{x \sigma \sqrt {2 \pi}} e ^ {- \frac {1}{2} \left(\frac {\ln (x) - \mu}{\sigma}\right) ^ {2}}, \tag {2} +$$ + +where $\mu$ is the location parameter and $\sigma$ is the scale parameter. We can explicitly obtain the true time headway values for semi-supervised training when there is observable interaction between the attacked target and the front vehicle. We consider the semi-supervised longitudinal feature encoding as a regression problem and optimize it by minimizing the mean square error. For the lateral directional features, we represent them by three simple but effective classes: moving forward, turning/changing lanes to the left, and turning/changing lanes to the right. These three intentions are discrete by nature and we model them with the categorical distribution. In the adversarial training process, only vehicles with clear intentions in a long enough time frame will be labeled and we utilize the cross-entropy to optimize this + +classification task. + +For all the semantic and Gaussian latent variables, they are regularized to the target distribution by the adversarial generation loss in Eq. (3). The discriminators are trained to maximize the log probability of real latent samples $s$ and the log of the inverse probability for fake latent samples, as in Eq. (4): + +$$ +L o s s _ {G} (x) = \frac {1}{m} \sum_ {i = 1} ^ {m} \log \left(1 - D _ {i} (G (x))\right), \tag {3} +$$ + +$$ +\operatorname {L o s s} _ {D _ {i}} (x, s) = \log D _ {i} \left(s _ {i}\right) + \log \left(1 - D _ {i} (G (x))\right), \tag {4} +$$ + +where $x$ is the high-dimensional features and $m$ is the number of different kinds of latent vectors. $G$ and $D$ are encoders and distribution discriminators, respectively. + +# 3.2. Adversarial Training Process + +# 3.2.1 Adversarial Training Algorithm + +For each sample, we utilize the PGD attack to generate the adversarial trajectory only for the target vehicle and keep other surrounding vehicles' original trajectories. This constrains the adversarial attack's impact on the whole scenario. If the error between the prediction under attack and the ground truth is greater than a threshold, we consider the attack a successful one and conduct adversarial training on this sample. Since the perturbation is very small, we consider the real future trajectory as the ground truth $y_{i}$ for the adversarial training and optimize the whole pipeline with L1-smooth loss in Eq. (5). This is a general but somewhat naive adversarial training process. + +$$ +\operatorname {L o s s} _ {\text {t r a j}} \left(y _ {i}, \hat {y} _ {i}\right) = \left\{ \begin{array}{l l} 0. 5 \left(y _ {i} - \hat {y} _ {i}\right) ^ {2} & \text {i f} \| y _ {i} - \hat {y} _ {i} \| < 1, \\ \| y _ {i} - \hat {y} _ {i} \| - 0. 5 & \text {o t h e r w i s e .} \end{array} \right. \tag {5} +$$ + +Thus, to further facilitate the adversarial training, we exploit semantic features and their corresponding labels in our proposed architecture. The encoder is optimized to minimize the mean square error of longitudinal features and the cross entropy of lateral features between the ground truths and predictions in the latent space. The semi-supervised loss function is shown as follows: + +$$ +L o s s _ {S e m i} (z, g) = - \sum_ {i = 1} ^ {3} g _ {l a t} \log z _ {l a t} + \left(g _ {l o n} - z _ {l o n}\right) ^ {2}, \tag {6} +$$ + +where $z$ represents the predicted semantic vectors when under attack and $g$ represents the ground truth in the benign scenarios. + +Moreover, we further adapt the adversarial training process with lateral semantic vectors because the lateral directional prediction can be regarded as a classification problem with clear behavior meaning. When the adversarial example leads to the wrong classification of lateral behavior, we + +will set higher weights of the semi-supervised loss for the adversarial training. In this way, our model will first guarantee the correctness of high-level semantic prediction and then tune the regression error, which could help avoid significant adversarial deviation and enhance the generalization performance. + +Algorithm 1 Our Adversarial Training Pipeline +1: Initialize: feature extractor $F$ , AAE encoder $G$ , decoder $R$ , discriminator $D_{i}$ , target distribution $p_{i}$ , $i = 1,2,3$ , adversarial example generator $Adv$ . +2: Input: past trajectories $t$ , future trajectories ground truth $y$ , map context $c$ , prediction model $m$ . +3: for each sample do +4: Generate adversarial trajectory for target vehicle $t_{adv} = Adv(t,y,m,c)$ . +5: Input $t_{adv}$ to the predictor. $\hat{y}', \hat{z}' = m(t_{adv},c)$ , where $z = \{z_{lat},z_{lon},z_{gaussian}\}$ . +6: if err $(\hat{y}', y) > Threshold$ then +7: if $\text{argmax}(z_{lat}) \neq \text{argmax}(z_{lat}')$ then +8: Update model $m$ with higher weights of semi-supervised learning loss $Loss_{semi}$ in Eq. (6). +9: Update model $m$ with $Loss_{D}$ in Eq. (4), $Loss_{G}$ in Eq. (3), and $Loss_{traj}$ in Eq. (5). +10: else +11: Update model $m$ with $Loss_{D}$ in Eq. (4), $Loss_{G}$ in Eq. (3), $Loss_{semi}$ in Eq. (6), and $Loss_{traj}$ in Eq. (5). +12: end if +13: end if +14: end for + +# 3.2.2 Balance Accuracy and Robustness + +In our preliminary experiments, we notice a trade-off between standard accuracy and adversarial robustness. A similar phenomenon has been observed in classification tasks [46, 28]. Methods such as TRADES [46], robust self-training [3] and MixUp [47, 1] have been proposed to balance such trade-off. However, there are few methods that can be applied to trajectory prediction because such time-series regression problems have no class labels and are more sensitive to errors introduced by augmented data. In this work, we utilize the MixUp [47] technique to mix the adversarial scenarios and benign scenarios in the adversarial training process. The experiments demonstrate that a balance between adversarial robustness and standard accuracy can be achieved in trajectory prediction, as shown later. + +Table 2: Comparison of different defense methods when under various attacks and in the benign case. For the adversarial training methods, we calculate the mean error of models adversarially trained on different attacks. + +
MethodsADE AttackLateral AttackLongitudinal AttackBenign
ADE/mLateral Error/mLongitudinal Error/mADE/m
DatasetArgo1Argo2ApolloArgo1Argo2ApolloArgo1Argo2ApolloArgo1Argo2Apollo
Original Model5.174.523.931.661.271.543.783.173.421.430.791.78
Train-time Smo. [48]4.674.323.861.451.271.493.233.003.411.500.832.11
Test-time Smo. [48]4.323.513.490.750.891.093.542.423.051.681.232.07
Heuristic Aug. [29]4.623.523.841.091.211.343.244.243.211.540.972.01
Data-driven Aug. [29]4.534.433.530.740.780.542.732.613.022.501.992.24
Standard AT3.793.673.680.670.641.152.441.893.051.671.061.87
SSAT (ours)3.512.672.900.610.530.411.761.441.261.751.201.87
Mixup-SSAT (ours)3.733.333.170.720.510.532.131.631.351.641.071.86
+ +# 4. Experimental Results + +# 4.1. Experiment Setup + +# 4.1.1 Dataset + +We train and evaluate different defense methods using three popular benchmarks - Argoverse 1 [4], Argoverse 2 [42], and ApolloScape datasets [9]. The datasets contain more than $250\mathrm{k}$ real driving scenarios in different cities, such as Miami and Pittsburgh. For Argoverse 1 and Argoverse 2, each scenario consists of a road graph and multiple agents' trajectories sampled at a frequency of $10\mathrm{Hz}$ . We choose 20 waypoints as the history trajectory and the models will predict 30 waypoints in the future. Scenarios in ApolloScape are simpler. They have no maps but 6 waypoints for both history and future trajectories. + +# 4.1.2 Attack Settings + +In the experiments, we study three different types of attacks [48] to the vehicle trajectory prediction algorithms - lateral directional attack (shift to the right), longitudinal directional attack (shift forward), and ADE attack (deviate randomly). More details of them are in Section 2.1. The trajectory prediction models are adversarially trained on the three types of attacks, respectively. We constrain the maximum deviation between the attacked and the benign input trajectories to be 1 meter. + +# 4.1.3 Training Settings + +Since our architecture is an encoder-decoder module that can be combined with different feature extractors, we first fine-tune the model on the benign data. In the experiments, we use the feature extractor from LaneGCN [16], an attention-based graph neural network. We notice that the AAE architecture introduces a slight accuracy drop on the benign data, mainly due to the dimension reduction. For adversarial training, we train prediction models on adversarial samples generated from scenarios in datasets. + +# 4.2. Experimental Results and Analysis + +In this section, we conduct experiments with various defense methods under different patterns of attacks, including our semi-supervised semantics-guided (SSAT) method and a Mixup-SSAT method that combines SSAT with the MixUp technique for balancing standard accuracy and adversarial robustness. In the following, we first compare the average robustness improvement among various defense methods, which demonstrates the advantage of SSAT in improving robustness under various types of attacks and the effectiveness of Mixup-SSAT in balancing robustness and accuracy. Then, we show that SSAT can significantly enhance the robust generalization to unseen types of attacks. In addition, we evaluate an unsupervised version of SSAT to explicitly show how the semi-supervised semantic-guided latent space modeling can boost the adversarial robustness, which also serves as an ablation study. + +# 4.2.1 Effectiveness of SSAT Methods + +We compare our SSAT and Mixup-SSAT methods with the original model and five different defense methods, including train-time smoothing [48], test-time smoothing [48], heuristic data augmentation [29], data-driven augmentation [29], and the standard adversarial training (Standard AT). All the models share the same feature extractor LaneGCN in this setting. Note that we compare to the two data augmentation methods as they are effective for image classification tasks [6, 29, 35]. For the data-driven augmentation, we design an additional decoder to augment input trajectories and it can generate more inputs by adding Gaussian noises to the latent vectors of the real inputs. For the heuristic augmentation, we simply add random perturbation to the benign inputs, with the same constraints of maximum deviation. + +Table 2 shows the prediction errors of different methods when under various types of attacks and in the benign case (note that when under lateral and longitudinal attacks, we measure lateral and longitudinal errors). We can see that our SSAT method significantly outperforms all other de + +![](images/aa1253b2226a86d5e0b88a2cdffb393e62baeed3c97f70f7621773b97277dca5.jpg) +(a) + +![](images/e6cee0f856483597832978a138c131ea0afc0ce22d9f48def8bf231c285ee866.jpg) +(b) +Figure 4: Adversarial robustness comparisons of the original model, the standard adversarial training on final trajectories (Standard AT), and the SSAT method proposed by us. Figures (a), (b), (c) show results from adversarial training (AT) that targets ADE attacks, lateral attacks, and longitudinal attacks, respectively. + +![](images/1bbcab51993c0b022ee18572adfccec7179ed6b2333c7d2b9ac9d51a37701c18.jpg) +(c) + +fense methods in improving the robustness of trajectory prediction. Compared with the original model, SSAT can reduce the prediction error by $32\% - 73\%$ when under different types of attacks. + +Moreover, while our SSAT method improves robustness, we also observe a drop in standard accuracy in the benign case. Mixup-SSAT enables effective trade-off between these two objectives (i.e., better performance than SSAT in benign cases at the expense of worse performance under attacks), by setting the mixup ratio between adversarial and benign examples to different values (the results in Table 2 are based on a mixup ratio of 2). + +We also notice that both data-driven and heuristic data augmentation methods offer very limited improvement over the original model. This is likely due to the challenge of regression tasks with rich context, which makes augmentation methods perform poorer than they do for image classification tasks. + +Table 3: Comparison of different methods when adversarial training is conducted for ADE attack, and tested on ADE, lateral, and longitudinal attacks. + +
MethodsPrediction Error/m, w/o AT → with AT
ADE AttackLat AttackLon Attack
Standard AT5.17 → 3.761.58 → 0.833.59 → 2.66
SSAT5.16 → 3.241.66 → 0.733.78 → 1.45
Unsup-SSAT5.16 → 3.481.66 → 0.743.78 → 1.49
+ +Table 4: Comparison of different methods when adversarial training is conducted for lateral attack, and tested on ADE, lateral, and longitudinal attacks. + +
MethodsPrediction Error/m, w/o AT → with AT
ADE AttackLat AttackLon Attack
Standard AT5.17 → 4.281.58 → 0.593.59 → 3.01
SSAT5.16 → 4.021.66 → 0.593.78 → 2.52
Unsup-SSAT5.16 → 4.111.66 → 0.693.78 → 2.54
+ +# 4.2.2 Effectiveness of SSAT in Robust Generalization on Different Types of Attacks + +We observe that there is an adversarial robust generalization gap when the training and test are under different types of attacks. The comparisons in Tables 3, 4, 5 show that our SSAT method is better generalized to unseen types of attacks, compared to the standard adversarial training. In every training scenario, our method is more robust to the various unseen patterns of attacks. For instance, Table 3 shows that when applying SSAT to train under the random ADE attack, its results outperform other models on all seen (i.e., ADE) and unseen (i.e., lateral and longitudinal) types of attacks, which demonstrates that our SSAT method can effectively decompose and learn semantic features from random ADE attacks. Tables 4 and 5 show similar trends, where our SSAT methods are better at defending against unseen attacks and mitigating overfitting on specific patterns of attacks. Fig. 4 further visualizes the results from these three tables for the original model, the standard adversarial training (Standard AT), and SSAT. + +Table 5: Comparison of different methods when adversarial training is conducted for longitudinal attack, and tested on ADE, lateral, and longitudinal attacks. + +
MethodsPrediction Error/m, w/o AT → with AT
ADE AttackLat AttackLon Attack
Standard AT5.17 → 3.331.58 → 0.603.59 → 1.64
SSAT5.16 → 3.281.66 → 0.523.78 → 1.30
Unsup-SSAT5.16 → 3.401.66 → 0.613.78 → 1.49
+ +# 4.2.3 Impact from Latent Space Modeling + +We also conduct adversarial training that only has regularization on the latent distributions but without supervision on latent vectors. We name it Unsup-SSAT. The comparisons between the standard adversarial training (Standard AT) and Unsup-SSAT in Tables 3, 4, and 5 demonstrate that even without labels, the partial disentanglement and + +distribution modeling in Unsup-SSAT will benefit the adversarial training on trajectory prediction and outperform the baseline Standard AT in most cases. However, when compared with SSAT, we find that the adversarial robustness will be further improved with the extra labels in the semi-supervised phase (in practice, we often have access to those labels). + +# 5. Conclusion + +In this work, we propose an adversarial training method for trajectory prediction. To tackle the challenge of random inputs with rich context, diverse types of attacks, and lack of class labels, we develop a novel AAE architecture that exploits the disentanglement and semantic features for enhancing model robustness and its generalization. Our proposed SSAT method significantly outperforms a number of baselines from the literature, reducing the prediction errors under attacks by up to $32\% - 73\%$ when compared with the original prediction models. Our method is also shown to be effective for defending against unseen attacks. + +# Acknowledgment + +We gratefully acknowledge the support from NSF grants 1834701, 1724341, 2038853, 1929771, 2145493, 1932464, and ONR grant N00014-19-1-2496. + +# References + +[1] Guillaume P Archambault, Yongyi Mao, Hongyu Guo, and Richong Zhang. Mixup as directional adversarial training. arXiv preprint arXiv:1906.06875, 2019. +[2] Yulong Cao, Danfei Xu, Xinshuo Weng, Zhuoqing Mao, Anima Anandkumar, Chaowei Xiao, and Marco Pavone. Robust trajectory prediction against adversarial attacks. In 6th Annual Conference on Robot Learning. +[3] Yair Carmon, Aditi Raghunathan, Ludwig Schmidt, John C Duchi, and Percy S Liang. Unlabeled data improves adversarial robustness. Advances in Neural Information Processing Systems, 32, 2019. +[4] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8748-8757, 2019. +[5] Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia + +Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11525-11533, 2020. +[6] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 34:4218-4233, 2021. +[7] Duy-Hung Ha, Maurice Aron, and Simon Cohen. Time headway variable and probabilistic modeling. Transportation Research Part C: Emerging Technologies, 25:181-201, 2012. +[8] Michael W Hancock and Bud Wright. A policy on geometric design of highways and streets. American Association of State Highway and Transportation Officials: Washington, DC, USA, 2013. +[9] Xinyu Huang, Xinjing Cheng, Qichuan Geng, Binbin Cao, Dingfu Zhou, Peng Wang, Yuanqing Lin, and Ruigang Yang. The apolloscape dataset for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 954-960, 2018. +[10] Ruochen Jiao, Juyang Bai, Xiangguo Liu, Takami Sato, Xiaowei Yuan, Qi Alfred Chen, and Qi Zhu. Learning representation for anomaly detection of vehicle trajectories. 2023. +[11] Ruochen Jiao, Hengyi Liang, Takami Sato, Junjie Shen, Qi Alfred Chen, and Qi Zhu. End-to-end uncertainty-based mitigation of adversarial attacks to automated lane centering. In 2021 IEEE Intelligent Vehicles Symposium (IV), pages 266-273. IEEE, 2021. +[12] Ruochen Jiao, Xiangguo Liu, Bowen Zheng, Dave Liang, and Qi Zhu. Tae: A semi-supervised controllable behavior-aware trajectory generator and predictor. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 12534-12541. IEEE, 2022. +[13] Diederik Kingma and Max Welling. An introduction to variational autoencoders. arXiv preprint arXiv:1906.02691, 2019. +[14] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. +[15] Dasom Lee and David J Hess. Regulations for on-road testing of connected and automated vehicles: Assessing the potential for global safety harmonization. + +Transportation Research Part A: Policy and Practice, 136:85-98, 2020. +[16] Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph representations for motion forecasting. In European Conference on Computer Vision, pages 541-556. Springer, 2020. +[17] C. Lin, Bowen Zheng, Qi Zhu, and Alberto Sangiovanni-Vincentelli. Security-Aware Design Methodology and Optimization for Automotive Systems. ACM Transactions on Design Automation of Electronic Systems (TODAES), 21(1):18:1-18:26, December 2015. +[18] Xiangguo Liu, Chao Huang, Yixuan Wang, Bowen Zheng, and Qi Zhu. Physics-aware safety-assured design of hierarchical neural network based planner. In 2022 ACM/IEEE 13th International Conference on Cyber-Physical Systems (ICCPS), pages 137-146. IEEE, 2022. +[19] Xiangguo Liu, Ruochen Jiao, Yixuan Wang, Yimin Han, Bowen Zheng, and Qi Zhu. Safety-assured speculative planning with adaptive prediction. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023. +[20] Xiangguo Liu, Ruochen Jiao, Bowen Zheng, Dave Liang, and Qi Zhu. Safety-driven interactive planning for neural network-based lane changing. In Proceedings of the 28th Asia and South Pacific Design Automation Conference, pages 39-45, 2023. +[21] Xiangguo Liu, Yunpeng Luo, Anthony Goeckner, Trishna Chakraborty, Ruochen Jiao, Ningfei Wang, Yixuan Wang, Takami Sato, Qi Alfred Chen, and Qi Zhu. Invited: Waving the double-edged sword: Building resilient cavs with edge and cloud computing. In Proceedings of the 60th Annual Design Automation Conference, 2023. +[22] Xiangguo Liu, Neda Masoud, and Qi Zhu. Impact of sharing driving attitude information: A quantitative study on lane changing. In 2020 IEEE Intelligent Vehicles Symposium (IV), pages 1998-2005. IEEE, 2020. +[23] Yicheng Liu, Jinghuai Zhang, Liangji Fang, Qinhong Jiang, and Bolei Zhou. Multimodal motion prediction with stacked transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7577-7586, 2021. +[24] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. + +[25] Pratyush Maini, Eric Wong, and Zico Kolter. Adversarial robustness against the union of multiple perturbation models. In International Conference on Machine Learning, pages 6640-6650. PMLR, 2020. +[26] Alireza Makhzani, Jonathon Shlens, Navdeep Jaitly, Ian Goodfellow, and Brendan Frey. Adversarial autoencoders. arXiv preprint arXiv:1511.05644, 2015. +[27] Emile Mathieu, Tom Rainforth, Nana Siddharth, and Yee Whye Teh. Disentangling disentanglement in variational autoencoders. In International Conference on Machine Learning, pages 4402-4412. PMLR, 2019. +[28] Aditi Raghunathan, Sang Michael Xie, Fanny Yang, John Duchi, and Percy Liang. Understanding and mitigating the tradeoff between robustness and accuracy. arXiv preprint arXiv:2002.10716, 2020. +[29] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan A Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Fixing data augmentation to improve adversarial robustness. arXiv preprint arXiv:2103.01946, 2021. +[30] Leslie Rice, Eric Wong, and Zico Kolter. Overfitting in adversarially robust deep learning. In International Conference on Machine Learning, pages 8093-8104. PMLR, 2020. +[31] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory++: Dynamically-feasible trajectory forecasting with heterogeneous data. In European Conference on Computer Vision, pages 683-700. Springer, 2020. +[32] Takami Sato, Junjie Shen, Ningfei Wang, Yunhan Jia, Xue Lin, and Qi Alfred Chen. Dirty road can attack: Security of deep learning based automated lane centering under {Physical-World} attack. In 30th USENIX Security Symposium (USENIX Security 21), pages 3309–3326, 2021. +[33] Ludwig Schmidt, Shibani Santurkar, Dimitris Tsipras, Kunal Talwar, and Aleksander Madry. Adversarily robust generalization requires more data. Advances in neural information processing systems, 31, 2018. +[34] Lukas Schott, Jonas Rauber, Matthias Bethge, and Wieland Brendel. Towards the first adversarially robust neural network model on mnist. arXiv preprint arXiv:1805.09190, 2018. +[35] Vikash Sehwag, Saeed Mahloujifar, Tinashe Handina, Sihui Dai, Chong Xiang, Mung Chiang, and Prateek Mittal. Robust learning meets generative models: Can proxy distributions improve adversarial robustness? arXiv preprint arXiv:2104.09425, 2021. + +[36] Amolika Sinha, Sai Chand, Kasun P Wijayaratna, Navreet Virdi, and Vinayak Dixit. Comprehensive safety assessment in mixed fleets with connected and automated vehicles: A crash severity and rate evaluation of conventional vehicles. Accident Analysis & Prevention, 142:105567, 2020. +[37] Chuanbiao Song, Kun He, Jiadong Lin, Liwei Wang, and John E Hopcroft. Robust local features for improving the generalization of adversarial training. arXiv preprint arXiv:1909.10147, 2019. +[38] Dong Su, Huan Zhang, Hongge Chen, Jinfeng Yi, Pin-Yu Chen, and Yupeng Gao. Is robustness the cost of accuracy? a comprehensive study on the robustness of 18 deep image classification models. In Proceedings of the European Conference on Computer Vision (ECCV), pages 631-648, 2018. +[39] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. +[40] Zhilu Wang, Chao Huang, and Qi Zhu. Efficient global robustness certification of neural networks via interleaving twin-network encoding. In DATE'22: Proceedings of the Conference on Design, Automation and Test in Europe, 2022. +[41] Matthew Willetts, Alexander Camuto, Tom Rainforth, Stephen Roberts, and Chris Holmes. Improving vaes' robustness to adversarial attack. arXiv preprint arXiv:1906.00230, 2019. +[42] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, et al. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2). +[43] Dongxian Wu, Shu-Tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization. Advances in Neural Information Processing Systems, 33:2958-2969, 2020. +[44] Hongtao Yu, H Eric Tseng, and Reza Langari. A human-like game theory-based controller for automatic lane changing. Transportation Research Part C: Emerging Technologies, 88:140-158, 2018. +[45] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M Kitani. Agentformer: Agent-aware transformers for + +socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9813-9823, 2021. +[46] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. +[47] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. +[48] Qingzhao Zhang, Shengtuo Hu, Jiachen Sun, Qi Alfred Chen, and Z Morley Mao. On adversarial robustness of trajectory prediction for autonomous vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15159-15168, 2022. +[49] Shufei Zhang, Zhuang Qian, Kaizhu Huang, Qiufeng Wang, Rui Zhang, and Xinping Yi. Towards better robust generalization with shift consistency regularization. In International Conference on Machine Learning, pages 12524-12534. PMLR, 2021. +[50] Bowen Zheng, Hengyi Liang, Qi Zhu, Huafeng Yu, and Chung-Wei Lin. Next generation automotive architecture modeling and exploration for autonomous driving. In 2016 IEEE Computer Society Annual Symposium on VLSI (ISVLSI), pages 53–58, 2016. +[51] Bowen Zheng, Chung-Wei Lin, et al. Design and Analysis of Delay-Tolerant Intelligent Intersection Management. ACM T-CPS, 4(1):3:1-3:27, November 2019. +[52] Qi Zhu, Chao Huang, Ruochen Jiao, Shuyue Lan, Hengyi Liang, Xiangguo Liu, Yixuan Wang, Zhilu Wang, and Shichao Xu. Safety-assured design and adaptation of learning-enabled autonomous systems. In Proceedings of the 26th Asia and South Pacific Design Automation Conference, 2021. +[53] Qi Zhu, Wenchao Li, Hyoseung Kim, Yecheng Xiang, Kacper Wardega, Zhilu Wang, Yixuan Wang, Hengyi Liang, Chao Huang, Jiameng Fan, and Hyunjong Choi. Know the unknowns: Addressing disturbances and uncertainties in autonomous systems. In Proceedings of the 39th International Conference on Computer-Aided Design, ICCAD '20, New York, NY, USA, 2020. Association for Computing Machinery. \ No newline at end of file diff --git a/2023/Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction/images.zip b/2023/Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..079f2838387033bac339c42200d2b0e2fb7fc456 --- /dev/null +++ b/2023/Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af6ee8759cb1e8a0ee6ecabf7ad017f94849935c58f566ec8ce14623a82dbfca +size 361601 diff --git a/2023/Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction/layout.json b/2023/Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cf2b4568035e1a94f93232d53d7e566feb9cc5b3 --- /dev/null +++ b/2023/Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction/layout.json @@ -0,0 +1,8329 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 481, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 481, + 139 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 481, + 139 + ], + "type": "text", + "content": "Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 178, + 162, + 244, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 162, + 244, + 175 + ], + "spans": [ + { + "bbox": [ + 178, + 162, + 244, + 175 + ], + "type": "text", + "content": "Ruochen Jiao" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 176, + 272, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 176, + 272, + 190 + ], + "spans": [ + { + "bbox": [ + 151, + 176, + 272, + 190 + ], + "type": "text", + "content": "Northwestern University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 192, + 296, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 192, + 296, + 202 + ], + "spans": [ + { + "bbox": [ + 127, + 192, + 296, + 202 + ], + "type": "text", + "content": "ruochen.jiao@u.northwestern.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 363, + 162, + 433, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 162, + 433, + 175 + ], + "spans": [ + { + "bbox": [ + 363, + 162, + 433, + 175 + ], + "type": "text", + "content": "Xiangguo Liu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 338, + 176, + 458, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 176, + 458, + 190 + ], + "spans": [ + { + "bbox": [ + 338, + 176, + 458, + 190 + ], + "type": "text", + "content": "Northwestern University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 330, + 192, + 466, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 192, + 466, + 202 + ], + "spans": [ + { + "bbox": [ + 330, + 192, + 466, + 202 + ], + "type": "text", + "content": "xg.liu@u.northwestern.edu" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 209, + 169, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 209, + 169, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 169, + 221 + ], + "type": "text", + "content": "Takami Sato" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 224, + 213, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 224, + 213, + 237 + ], + "spans": [ + { + "bbox": [ + 63, + 224, + 213, + 237 + ], + "type": "text", + "content": "University of California, Irvine" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 96, + 240, + 179, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 240, + 179, + 250 + ], + "spans": [ + { + "bbox": [ + 96, + 240, + 179, + 250 + ], + "type": "text", + "content": "takamis@uci.edu" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 236, + 224, + 388, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 224, + 388, + 237 + ], + "spans": [ + { + "bbox": [ + 236, + 224, + 388, + 237 + ], + "type": "text", + "content": "University of California, Irvine" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 271, + 240, + 353, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 240, + 353, + 250 + ], + "spans": [ + { + "bbox": [ + 271, + 240, + 353, + 250 + ], + "type": "text", + "content": "alfchen@uci.edu" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 452, + 210, + 489, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 452, + 210, + 489, + 223 + ], + "spans": [ + { + "bbox": [ + 452, + 210, + 489, + 223 + ], + "type": "text", + "content": "Qi Zhu" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 411, + 224, + 531, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 224, + 531, + 237 + ], + "spans": [ + { + "bbox": [ + 411, + 224, + 531, + 237 + ], + "type": "text", + "content": "Northwestern University" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 414, + 240, + 528, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 240, + 528, + 250 + ], + "spans": [ + { + "bbox": [ + 414, + 240, + 528, + 250 + ], + "type": "text", + "content": "qzhu@northwestern.edu" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 143, + 278, + 191, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 278, + 191, + 291 + ], + "spans": [ + { + "bbox": [ + 143, + 278, + 191, + 291 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 304, + 289, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 304, + 289, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 289, + 628 + ], + "type": "text", + "content": "Predicting the trajectories of surrounding objects is a critical task for self-driving vehicles and many other autonomous systems. Recent works demonstrate that adversarial attacks on trajectory prediction, where small crafted perturbations are introduced to history trajectories, may significantly mislead the prediction of future trajectories and induce unsafe planning. However, few works have addressed enhancing the robustness of this important safety-critical task. In this paper, we present a novel adversarial training method for trajectory prediction. Compared with typical adversarial training on image tasks, our work is challenged by more random input with rich context and a lack of class labels. To address these challenges, we propose a method based on a semi-supervised adversarial autoencoder, which models disentangled semantic features with domain knowledge and provides additional latent labels for the adversarial training. Extensive experiments with different types of attacks demonstrate that our Semi-supervised Semantics-guided Adversarial Training (SSAT1) method can effectively mitigate the impact of adversarial attacks by up to " + }, + { + "bbox": [ + 47, + 304, + 289, + 628 + ], + "type": "inline_equation", + "content": "73\\%" + }, + { + "bbox": [ + 47, + 304, + 289, + 628 + ], + "type": "text", + "content": " and outperform other popular defense methods. In addition, experiments show that our method can significantly improve the system's robust generalization to unseen patterns of attacks. We believe that such semantics-guided architecture and advancement on robust generalization is an important step for developing robust prediction models and enabling safe decision making." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 650, + 127, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 650, + 127, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 650, + 127, + 662 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 670, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 287, + 695 + ], + "type": "text", + "content": "Connected and autonomous vehicles (CAVs) have shown great promise to revolutionize the transportation systems," + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 280, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 280, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 306, + 280, + 545, + 363 + ], + "type": "text", + "content": "but also face significant concerns regarding their safety and security [15, 36, 21, 17, 50, 51]. Many of these concerns, which also apply to many other autonomous systems, arise from the increasing adoption of advanced deep neural network (DNN)-based machine learning techniques across system sensing, perception, prediction, planning, control, and general decision-making [52, 53]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 365, + 546, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 546, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 546, + 628 + ], + "type": "text", + "content": "In particular, the adversarial robustness of DNNs has drawn significant attention in recent years. State-of-the-art classifiers can be misled to make erroneous predictions by slight but carefully-optimized perturbations [39], and addressing this is critical to the safety of autonomous systems that employ these classifiers [11, 40]. Besides classification, adversarial attacks are also targeted at various other tasks such as object detection, image generation and natural language processing. To defend against adversarial examples, adversarial training is commonly used to enhance the intrinsic robustness of models and has been shown to be very effective among various defense strategies [24, 25, 34]. On the other hand, the works in [30, 43, 37] show that adversarial training may suffer poor robust generalization on unseen attacks and [33, 24, 38] demonstrate a trade-off between robustness and accuracy. In this paper, we address trajectory prediction, a safety-critical task that is common in self-driving vehicles and other autonomous systems, and propose a novel architecture to enhance its adversarial robustness and robust generalization by introducing semantic features and a semi-supervised adversarial autoencoder (AAE) [26] into the adversarial training process." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": "We focus on the trajectory prediction module in autonomous driving, where large DNNs have enabled breakthroughs in recent years. More specifically, an autonomous driving system typically consists of several modules such as perception, localization, prediction, planning (route, behavior, trajectory, and motion planning), and control. The perception module is to detect agents and obstacles based" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 277, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 277, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 277, + 713 + ], + "type": "text", + "content": "1Our codes: https://github.com/jrcblue/SSAT-for-Motion-Prediction" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8207" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "content": "on the sensing inputs (e.g., images and 3D point clouds), and the effectiveness and robustness of the perception module are well-studied in the research community. The prediction module is to predict future trajectories of surrounding vehicles based on the observed history trajectories of those vehicles from the perception module and the map context. The trajectory prediction plays a crucial role in understanding the environment and making safety-critical decisions in the following planning module [18, 20, 19]. Recent works [23, 16, 5, 45, 12] have applied various deep learning techniques such as graph neural networks and transformers to this task and achieved impressive performance in terms of reducing average errors. However, few works studied the robustness of vehicle trajectory prediction under adversarial attacks, which is in fact of vital importance because 1) autonomous driving is a safety-critical task by nature, 2) recent work [48, 10] demonstrates that the prediction module is vulnerable to adversarial attacks if surrounding vehicles drive along a seemingly natural but crafted trajectory, and 3) current prediction models are often overfitted on limited patterns in the datasets but suffer the long-tailed distribution of driving scenarios and behaviors. The threat model of the adversarial trajectory is illustrated in an example in Fig. 1, where a crafted input trajectory can intentionally mislead the target vehicle to the wrong prediction and induce dangerous planning decisions." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 393, + 287, + 511 + ], + "blocks": [ + { + "bbox": [ + 50, + 393, + 287, + 511 + ], + "lines": [ + { + "bbox": [ + 50, + 393, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 50, + 393, + 287, + 511 + ], + "type": "image", + "image_path": "54cf1db9e7e81c910006b20e1b071f79068f13807b7324360d4b129d527e4b56.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 519, + 287, + 639 + ], + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 639 + ], + "type": "text", + "content": "Figure 1: An illustrative example of the threat model: In the upper benign scenario, the target vehicle predicts the future trajectory of a surrounding vehicle based on its history trajectory. In the lower adversarial scenario, the surrounding vehicle is malicious and drives along a crafted (history) trajectory, which may mislead the target vehicle to have a wrong (attacked) prediction of the surrounding vehicle's future trajectory. In this particular case, the target vehicle wrongly predicts that the malicious vehicle will cut into its lane, and thus takes an unsafe emergency braking." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "To defend against adversarial attacks on vehicle trajectory prediction, there are important challenges that are different from the cases of images and audios. First, vehicle trajectory prediction is a time-series regression problem with rich contexts while most existing adversarial attacks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 323 + ], + "type": "text", + "content": "and corresponding defense methods are targeted at classification tasks. The attack patterns are more random and there are no well-defined class labels, which means that the robust model is difficult to train and generalize. Second, vehicle trajectory can convey semantic and behavior information. For instance, people can infer the behavior of a vehicle such as moving forward or changing lanes from its trajectory. Therefore, as shown in our experiments later, some popular defense methods such as TRADES [46] and data augmentation methods [6, 29, 35] either are inapplicable or have poor performance in trajectory prediction task. In this work, we first propose an adversarial training pipeline for the trajectory prediction task, and then further exploit the semantic features to design a semi-supervised AAE architecture that can be added after the feature extractor, to improve the adversarial robustness and its generalization. The methodology of the proposed architecture such as enhancing disentanglement and defining semantic labels may be further applied to adversarial training for general regression and generation problems. Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 331, + 545, + 510 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 317, + 331, + 545, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 331, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 317, + 331, + 545, + 355 + ], + "type": "text", + "content": "- We propose a novel adversarial training method against adversarial attacks on trajectory predictions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 362, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 362, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 317, + 362, + 545, + 411 + ], + "type": "text", + "content": "- We develop a semi-supervised architecture with domain knowledge and semantic features to enhance the adversarial robustness and its generalization among different types of attack patterns." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 418, + 545, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 418, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 317, + 418, + 545, + 454 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that our proposed method (SSAT) effectively improves the adversarial robustness and outperforms popular defense baselines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 461, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 461, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 461, + 545, + 510 + ], + "type": "text", + "content": "- We further explore to balance the robustness-accuracy trade-off in this task by leveraging the MixUp technique [47], and we also test data augmentation methods for trajectory prediction robustness." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 520, + 391, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 520, + 391, + 532 + ], + "spans": [ + { + "bbox": [ + 306, + 520, + 391, + 532 + ], + "type": "text", + "content": "2. Preliminaries" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 540, + 541, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 541, + 553 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 541, + 553 + ], + "type": "text", + "content": "2.1. Adversarial Attacks on Trajectory Prediction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "content": "Recent research [48, 10] demonstrates that trajectory prediction in autonomous driving can be fooled by the adversarial behavior of a surrounding vehicle, where the adversarial behavior is optimized with Projected Gradient Descent (PGD) [24]. The threat model is illustrated in Fig. 1. In this work, we consider such attacks and assume the worst setting, i.e., the attacker has full knowledge of the target system and tries to maximize the attack impact. However, malicious agents cannot directly manipulate anything inside the target vehicle and they can only slightly change their own trajectories to mislead the target vehicle indirectly." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "type": "text", + "content": "Different from image classification, trajectory prediction has no class labels, but it has directional information in the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8208" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 86, + 227, + 167 + ], + "blocks": [ + { + "bbox": [ + 132, + 86, + 227, + 167 + ], + "lines": [ + { + "bbox": [ + 132, + 86, + 227, + 167 + ], + "spans": [ + { + "bbox": [ + 132, + 86, + 227, + 167 + ], + "type": "image", + "image_path": "daac5439f6a180240c33f5df31b6ceb98f3b8cc786b363d794c8583f936b544d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 253, + 86, + 347, + 167 + ], + "blocks": [ + { + "bbox": [ + 253, + 86, + 347, + 167 + ], + "lines": [ + { + "bbox": [ + 253, + 86, + 347, + 167 + ], + "spans": [ + { + "bbox": [ + 253, + 86, + 347, + 167 + ], + "type": "image", + "image_path": "c2e5e6a3d52e198afd9da362d85a221d25c792c26f6b407ffcca27f02778a530.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 374, + 86, + 468, + 167 + ], + "blocks": [ + { + "bbox": [ + 374, + 86, + 468, + 167 + ], + "lines": [ + { + "bbox": [ + 374, + 86, + 468, + 167 + ], + "spans": [ + { + "bbox": [ + 374, + 86, + 468, + 167 + ], + "type": "image", + "image_path": "2122c7654d92294b38e41a5fe8a72d17a432d8fa8f453474b7c6cea828732a9e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 132, + 186, + 226, + 267 + ], + "blocks": [ + { + "bbox": [ + 132, + 186, + 226, + 267 + ], + "lines": [ + { + "bbox": [ + 132, + 186, + 226, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 186, + 226, + 267 + ], + "type": "image", + "image_path": "a3523bdf036540e85f87e7bd6a8520ba720374a4d2cca6b14741edd372de6b16.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 276, + 179, + 285 + ], + "lines": [ + { + "bbox": [ + 168, + 276, + 179, + 285 + ], + "spans": [ + { + "bbox": [ + 168, + 276, + 179, + 285 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 253, + 186, + 347, + 267 + ], + "blocks": [ + { + "bbox": [ + 253, + 186, + 347, + 267 + ], + "lines": [ + { + "bbox": [ + 253, + 186, + 347, + 267 + ], + "spans": [ + { + "bbox": [ + 253, + 186, + 347, + 267 + ], + "type": "image", + "image_path": "74c62a6c7f807384e4fd52f213824e25bb86c7094274909743643f05baecaf73.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 276, + 301, + 285 + ], + "lines": [ + { + "bbox": [ + 290, + 276, + 301, + 285 + ], + "spans": [ + { + "bbox": [ + 290, + 276, + 301, + 285 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 298, + 545, + 382 + ], + "lines": [ + { + "bbox": [ + 46, + 298, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 545, + 382 + ], + "type": "text", + "content": "Figure 2: Adversarial trajectories generated by different types of attacks. The figures in the top row are adversarial (red line) and benign (blue line) input history trajectories for prediction and they look very close to human eyes. The figures in the bottom row show the corresponding attacked future trajectory prediction (red dashed line) and the ground truth trajectory in the benign case (green dashed line). The differences between the two are clearly visible, showing the effectiveness of the attacks. Figure (a) is the ADE attack, which will randomly lead to maximum average deviation. Figure (b) is the lateral attack, which mainly leads the vehicle to deviate to the left or right. Figure (c) is the longitudinal attack, which will mainly lead to longitudinal deviation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 375, + 186, + 468, + 267 + ], + "blocks": [ + { + "bbox": [ + 375, + 186, + 468, + 267 + ], + "lines": [ + { + "bbox": [ + 375, + 186, + 468, + 267 + ], + "spans": [ + { + "bbox": [ + 375, + 186, + 468, + 267 + ], + "type": "image", + "image_path": "a8b6ecb4f0eaa094f299713246f3e62c1e7f163d618ea6d2bf04098a0eb65b16.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 276, + 422, + 285 + ], + "lines": [ + { + "bbox": [ + 411, + 276, + 422, + 285 + ], + "spans": [ + { + "bbox": [ + 411, + 276, + 422, + 285 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 403, + 287, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 487 + ], + "type": "text", + "content": "context, such as moving forward, turning or changing lanes to the right. Therefore, the attacker can conduct targeted attacks besides random ones. In Fig. 2, three types of adversarial attacks are presented and we observe that they can cause significant and directional errors. [48] proposes directional error metrics for the optimization of targeted attacks, as shown in Eq. (1):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 85, + 496, + 287, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 496, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 85, + 496, + 287, + 512 + ], + "type": "interline_equation", + "content": "D (\\alpha , R) = \\left(p _ {\\alpha} - s _ {\\alpha}\\right) ^ {T} \\cdot R \\left(s _ {\\alpha + 1}, s _ {\\alpha}\\right), \\tag {1}", + "image_path": "1ebf219cd0d3842b483477bc3a848fe5931576f9fb2ef1eebafaa6ee4ec060eb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "content": " denotes the time frame ID. " + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "content": " are two-dimensional vectors representing predicted and ground-truth vehicle locations, respectively. " + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "content": " is a function generating the unit vector to a specific direction (lateral or longitudinal). The longitudinal direction is approximated as the vector defined by the adjacent two waypoints of ground truth " + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "inline_equation", + "content": "s_{\\alpha +1}^{n} - s_{\\alpha}^{n}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "text", + "content": "In addition to the directional attack, a random attack can be designed to maximize the Average Displacement Error (ADE), which is the average of the root mean squared error between the predicted waypoints and the ground-truth trajectory waypoints." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Hard constraints are also applied to the maximum deviation of waypoints during the optimization so as to make the adversarial history trajectory physically feasible and not perform unrealistic behaviors." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 326, + 435, + 526, + 473 + ], + "blocks": [ + { + "bbox": [ + 305, + 402, + 545, + 426 + ], + "lines": [ + { + "bbox": [ + 305, + 402, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 402, + 545, + 426 + ], + "type": "text", + "content": "Table 1: Trajectory prediction errors (in meters) in benign or adversarial scenarios under different types of attacks." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 435, + 526, + 473 + ], + "lines": [ + { + "bbox": [ + 326, + 435, + 526, + 473 + ], + "spans": [ + { + "bbox": [ + 326, + 435, + 526, + 473 + ], + "type": "table", + "html": "
ModelADELateralLongitudinal
benign/attackbenign/attackbenign/attack
LaneGCN [16]1.32/5.17-0.01/1.58-0.25/3.33
Trajectory++[31]2.69/6.810.107/2.25-0.526 / 3.79
", + "image_path": "0c8a7115b5f11351741f610baf2ff029a72528835eb5ce116673b82499cb4e0a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 512, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 512, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 512, + 545, + 632 + ], + "type": "text", + "content": "Table 1 quantitatively shows that the predicted trajectories can be seriously deviated by any of the three attack types. Generally, in the US, " + }, + { + "bbox": [ + 304, + 512, + 545, + 632 + ], + "type": "inline_equation", + "content": "0.3\\mathrm{m}" + }, + { + "bbox": [ + 304, + 512, + 545, + 632 + ], + "type": "text", + "content": " lateral deviation is enough to invade adjacent lanes on local roads [32, 8]. In addition, the final displacement error (the distance between the last points of predicted and ground-truth trajectories) is around two to three times larger than the average error shown here. Therefore, the attacker can apply either random or directional attacks that will greatly challenge the defense methods, especially on the generalization performance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "[48, 2] propose to utilize approaches such as smoothing and adversarial training on the conditional variational encoder (CVAE) architectures to mitigate the impact of such adversarial attacks. However, they do not explicitly address driving semantics and robustness generalization, which are critical for many real road scenarios." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8209" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "2.2. Adversarial Training Methods and Robust Generalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 102, + 287, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 102, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 287, + 209 + ], + "type": "text", + "content": "Adversarial training [24, 25, 34] is shown to be one of the most effective approaches to improve the robustness of DNN models. In practice, the PGD attack is commonly deployed for evaluation because of its strong attack ability in white-box settings and the work in [24] formulates the adversarial training as a min-max problem. [6, 29, 35] show that synthesized data can significantly boost robustness for the image classification task. Thus, we select similar methods as baselines for comparison in our experiments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 210, + 288, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 210, + 288, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 288, + 354 + ], + "type": "text", + "content": "Recent research shows that a specific type of adversarial attack is not sufficient to represent the diverse space of adversarial examples and many adversarially trained models are only robust to specific attacks. This does limit the application of adversarial training in practice, especially for trajectory regression tasks due to its nature of long-tailed distribution. The works in [33, 49, 37] try to explain and improve the robust generalization in various perspectives such as sample complexity and latent feature representation. In this work, we demonstrate that robust generalization for trajectory prediction can be enhanced by explicitly introducing disentangled and semantic features in the latent space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 361, + 249, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 249, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 249, + 373 + ], + "type": "text", + "content": "2.3. Adversarial Autoencoder Architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 380, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 571 + ], + "type": "text", + "content": "The adversarial autoencoder (AAE) is a variant of the variational autoencoders (VAE) [14, 13], which provides a principled method for jointly learning deep latent-variable models and corresponding inference models using stochastic gradient descent. AAE imposes various distributions on the latent vector by utilizing adversarial learning instead of KL divergence. Due to the flexibility of adversarial learning, AAE is superior to VAE in terms of imposing complicated distributions over latent space. For adversarial robustness, the works in [27, 41] demonstrate that the disentangled latent representations produce VAEs that are more robust to adversarial attacks. In our work, we design an AAE-based architecture that can be added after the feature extractor of prediction modules. We utilize this architecture to model diverse semantic features and enhance disentanglement in the latent space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 582, + 286, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 286, + 610 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 286, + 610 + ], + "type": "text", + "content": "3. Proposed Adversarial Training Method for Trajectory Prediction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 616, + 286, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 286, + 640 + ], + "type": "text", + "content": "3.1. Domain Knowledge-guided Semi-supervised Architecture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 647, + 143, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 143, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 143, + 658 + ], + "type": "text", + "content": "3.1.1 Overall Design" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 665, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 286, + 712 + ], + "type": "text", + "content": "Unlike adversarial training for image classification tasks, there are no class labels to guide the training for trajectory prediction, and the attack patterns in the trajectory regression are more random. We thus introduce domain knowl" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "type": "text", + "content": "edge to facilitate the modeling of the semantic information in both benign and adversarial cases, based on an AAE architecture. The model learns the directional semantic latent vector in a semi-supervised way because the ground truths are only available for limited scenarios but their distributions can be derived from domain knowledge and statistics. Therefore, our latent space modeling contains two levels – unsupervised distribution modeling and semi-supervised learning when labels are available." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 180, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 180, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 545, + 382 + ], + "type": "text", + "content": "The architecture of our proposed model is shown in Fig. 3. The feature extractor [16] utilizes a one-dimensional dilated convolutional neural network to obtain the embedding of the time-series trajectory and uses a graph neural network to model the lane context and interaction between objects. The encoder maps high-dimensional features to the semantics-guided latent space with distribution regularization and semi-supervised training. In particular, the latent space is divided into three parts: longitudinal features " + }, + { + "bbox": [ + 306, + 180, + 545, + 382 + ], + "type": "inline_equation", + "content": "z_{lon}" + }, + { + "bbox": [ + 306, + 180, + 545, + 382 + ], + "type": "text", + "content": ", following one-dimensional log-normal distribution, lateral features " + }, + { + "bbox": [ + 306, + 180, + 545, + 382 + ], + "type": "inline_equation", + "content": "z_{lat}" + }, + { + "bbox": [ + 306, + 180, + 545, + 382 + ], + "type": "text", + "content": ", following three-dimensional categorical distribution and remaining features, following Gaussian distribution [12]. Finally, the decoder maps semantic vectors along with other disentangled latent vectors to the future trajectories. Note that we develop the AAE instead of traditional VAE architecture to model these different and complex distributions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 384, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 384, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 384, + 545, + 526 + ], + "type": "text", + "content": "In the attacked scenarios, the impact of attacks will be decomposed to different latent vectors and the attack patterns will be explicitly modeled by the semantic features. Let us take the lateral directional vector as an example. If the attack is not targeted at the lateral dimension, the encoder will decompose the attack effect into other vectors and the mapping for lateral direction will remain stable. Otherwise, if the attack causes errors in the lateral vector, the feature extractor and encoder will be adversarial trained on the label of the direction and the corresponding mapping from the latent distribution to the final trajectories will also be updated." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 528, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 528, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 528, + 545, + 586 + ], + "type": "text", + "content": "Compared to existing trajectory prediction works that only apply adversarial training on the final trajectory waypoints, our proposed method is designed to capture semantic features in the latent space, and it can benefit the adversarial robustness and its generalization in various aspects:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 597, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 318, + 597, + 545, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 597, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 318, + 597, + 545, + 644 + ], + "type": "text", + "content": "- The architecture maps the high-dimensional features into disentangled latent space that can decompose the attacks into different orthogonal patterns. Such disentanglement will boost the adversarial robustness [41]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 318, + 654, + 545, + 712 + ], + "type": "text", + "content": "- The semantic vectors provide more context information and interpretable labels for the adversarial training and the semi-supervised learning makes the adversarial training process more efficient and easier to generalize to the unseen scenarios." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8210" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 70, + 500, + 281 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 500, + 281 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 500, + 281 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 500, + 281 + ], + "type": "image", + "image_path": "466c3ad5503ae0455b8dabcff162510d6441254a56aab57a6cbdef8a130aadd4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 290, + 545, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 290, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 545, + 350 + ], + "type": "text", + "content": "Figure 3: The overall adversarial training pipeline for the proposed semi-supervised semantics-guided architecture. We leverage the feature extractor in [16]. In the latent space, the D nets represent the discriminator that models the latent distributions by distinguishing the samples of real distribution from the predicted ones. When the semantic labels are available in the scenarios, they are used to adversarially train the semantic vectors (parts in green). Finally, we conduct adversarial training on attacked trajectories." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 372, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 372, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 58, + 372, + 287, + 396 + ], + "type": "text", + "content": "- The encoder-decoder architecture can filter high-frequency noise by dimension reduction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 405, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 405, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 405, + 287, + 430 + ], + "type": "text", + "content": "The detailed design and optimization of the architecture are described in the following." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 444, + 274, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 274, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 274, + 456 + ], + "type": "text", + "content": "3.1.2 Semi-supervised Semantic Features Modeling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 462, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 581 + ], + "type": "text", + "content": "To model the high-level semantic information in the trajectory prediction task, it is natural to decompose the trajectories into two dimensions – longitudinal and lateral directions. We want to utilize domain knowledge to guide the modeling by providing representative metrics and prior distributions. In addition to the semantic features, the architecture also maps high-dimensional information into latent vectors in Gaussian distributions to represent other low-level and more random features. We will mainly explain the semantic feature modeling below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": "In the longitudinal direction, speed and acceleration are often used to model the vehicle's dynamics but their values are always changing and do not contain enough semantic information. In our model, we apply the time headway to effectively extract the longitudinal feature, which measures the time difference between two successive vehicles when they cross a given point. The time headway represents a relatively stable behavior pattern for a certain agent and takes interaction with other vehicles into consideration. Recent works [44, 22] also use time headway as a measurement of aggressiveness in specific scenarios such as lane changing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 372, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 545, + 503 + ], + "type": "text", + "content": "or merging. In our proposed architecture, the model will represent the time headway as a one-dimensional vector in the latent space. In both benign and adversarial cases, the encoder will be trained by the regularization loss to force the longitudinal feature to a certain distribution that the time headway follows in statistics. Prior works [7] in the transportation domain show that the time headway in urban scenarios can be best described with the Log-normal distribution. We verify this and estimate the parameters in the Argoverse 1 motion forecasting dataset [4]. The longitudinal feature follows the distribution shown in Eq. (2):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 361, + 514, + 545, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 514, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 361, + 514, + 545, + 540 + ], + "type": "interline_equation", + "content": "f (x) = \\frac {1}{x \\sigma \\sqrt {2 \\pi}} e ^ {- \\frac {1}{2} \\left(\\frac {\\ln (x) - \\mu}{\\sigma}\\right) ^ {2}}, \\tag {2}", + "image_path": "266ea4701bb25b7bb9f6bdd543c12eb648f446f999e79a153d43f0f98425a56c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": " is the location parameter and " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": " is the scale parameter. We can explicitly obtain the true time headway values for semi-supervised training when there is observable interaction between the attacked target and the front vehicle. We consider the semi-supervised longitudinal feature encoding as a regression problem and optimize it by minimizing the mean square error. For the lateral directional features, we represent them by three simple but effective classes: moving forward, turning/changing lanes to the left, and turning/changing lanes to the right. These three intentions are discrete by nature and we model them with the categorical distribution. In the adversarial training process, only vehicles with clear intentions in a long enough time frame will be labeled and we utilize the cross-entropy to optimize this" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "8211" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 124, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 124, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 124, + 83 + ], + "type": "text", + "content": "classification task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "text", + "content": "For all the semantic and Gaussian latent variables, they are regularized to the target distribution by the adversarial generation loss in Eq. (3). The discriminators are trained to maximize the log probability of real latent samples " + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "text", + "content": " and the log of the inverse probability for fake latent samples, as in Eq. (4):" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 163, + 287, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 163, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 67, + 163, + 287, + 194 + ], + "type": "interline_equation", + "content": "L o s s _ {G} (x) = \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\log \\left(1 - D _ {i} (G (x))\\right), \\tag {3}", + "image_path": "00b1a96e545354029bc98e4f883a0957bdbc4fe47fba17b5bf479ac81a1be3a7.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 197, + 287, + 210 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 197, + 287, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 287, + 210 + ], + "type": "interline_equation", + "content": "\\operatorname {L o s s} _ {D _ {i}} (x, s) = \\log D _ {i} \\left(s _ {i}\\right) + \\log \\left(1 - D _ {i} (G (x))\\right), \\tag {4}", + "image_path": "cd59e71b57cae2c4e27fac88886a38ba52b739086974d70d6f80a1f36cf1c7b3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 217, + 287, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 287, + 254 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 287, + 254 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 217, + 287, + 254 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 217, + 287, + 254 + ], + "type": "text", + "content": " is the high-dimensional features and " + }, + { + "bbox": [ + 46, + 217, + 287, + 254 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 217, + 287, + 254 + ], + "type": "text", + "content": " is the number of different kinds of latent vectors. " + }, + { + "bbox": [ + 46, + 217, + 287, + 254 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 217, + 287, + 254 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 217, + 287, + 254 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 217, + 287, + 254 + ], + "type": "text", + "content": " are encoders and distribution discriminators, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 260, + 206, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 260, + 206, + 273 + ], + "spans": [ + { + "bbox": [ + 47, + 260, + 206, + 273 + ], + "type": "text", + "content": "3.2. Adversarial Training Process" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 278, + 217, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 278, + 217, + 290 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 217, + 290 + ], + "type": "text", + "content": "3.2.1 Adversarial Training Algorithm" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 297, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 287, + 441 + ], + "type": "text", + "content": "For each sample, we utilize the PGD attack to generate the adversarial trajectory only for the target vehicle and keep other surrounding vehicles' original trajectories. This constrains the adversarial attack's impact on the whole scenario. If the error between the prediction under attack and the ground truth is greater than a threshold, we consider the attack a successful one and conduct adversarial training on this sample. Since the perturbation is very small, we consider the real future trajectory as the ground truth " + }, + { + "bbox": [ + 46, + 297, + 287, + 441 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 297, + 287, + 441 + ], + "type": "text", + "content": " for the adversarial training and optimize the whole pipeline with L1-smooth loss in Eq. (5). This is a general but somewhat naive adversarial training process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 448, + 287, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 448, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 48, + 448, + 287, + 490 + ], + "type": "interline_equation", + "content": "\\operatorname {L o s s} _ {\\text {t r a j}} \\left(y _ {i}, \\hat {y} _ {i}\\right) = \\left\\{ \\begin{array}{l l} 0. 5 \\left(y _ {i} - \\hat {y} _ {i}\\right) ^ {2} & \\text {i f} \\| y _ {i} - \\hat {y} _ {i} \\| < 1, \\\\ \\| y _ {i} - \\hat {y} _ {i} \\| - 0. 5 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {5}", + "image_path": "4e7e6640489d2ea57e3aea64c54752f37a97edff76e0e1ff6a6e243f1d91b0a0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 491, + 287, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 573 + ], + "type": "text", + "content": "Thus, to further facilitate the adversarial training, we exploit semantic features and their corresponding labels in our proposed architecture. The encoder is optimized to minimize the mean square error of longitudinal features and the cross entropy of lateral features between the ground truths and predictions in the latent space. The semi-supervised loss function is shown as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 52, + 578, + 287, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 578, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 52, + 578, + 287, + 611 + ], + "type": "interline_equation", + "content": "L o s s _ {S e m i} (z, g) = - \\sum_ {i = 1} ^ {3} g _ {l a t} \\log z _ {l a t} + \\left(g _ {l o n} - z _ {l o n}\\right) ^ {2}, \\tag {6}", + "image_path": "4966bb1e60655924a5ad1a6055101d08a0309fe5b3673b0f82bb41f88a9ccc80.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 618, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 618, + 287, + 652 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 618, + 287, + 652 + ], + "type": "text", + "content": " represents the predicted semantic vectors when under attack and " + }, + { + "bbox": [ + 46, + 618, + 287, + 652 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 618, + 287, + 652 + ], + "type": "text", + "content": " represents the ground truth in the benign scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "Moreover, we further adapt the adversarial training process with lateral semantic vectors because the lateral directional prediction can be regarded as a classification problem with clear behavior meaning. When the adversarial example leads to the wrong classification of lateral behavior, we" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "will set higher weights of the semi-supervised loss for the adversarial training. In this way, our model will first guarantee the correctness of high-level semantic prediction and then tune the regression error, which could help avoid significant adversarial deviation and enhance the generalization performance." + } + ] + } + ], + "index": 13 + }, + { + "type": "code", + "bbox": [ + 311, + 180, + 547, + 457 + ], + "blocks": [ + { + "bbox": [ + 307, + 165, + 499, + 178 + ], + "lines": [ + { + "bbox": [ + 307, + 165, + 499, + 178 + ], + "spans": [ + { + "bbox": [ + 307, + 165, + 499, + 178 + ], + "type": "text", + "content": "Algorithm 1 Our Adversarial Training Pipeline" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "lines": [ + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "spans": [ + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": "1: Initialize: feature extractor " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": ", AAE encoder " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": ", decoder " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": ", discriminator " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": ", target distribution " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "p_{i}" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "i = 1,2,3" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": ", adversarial example generator " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "Adv" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": ". \n2: Input: past trajectories " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": ", future trajectories ground truth " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": ", map context " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": ", prediction model " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": ". \n3: for each sample do \n4: Generate adversarial trajectory for target vehicle " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "t_{adv} = Adv(t,y,m,c)" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": ". \n5: Input " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "t_{adv}" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": " to the predictor. " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "\\hat{y}', \\hat{z}' = m(t_{adv},c)" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "z = \\{z_{lat},z_{lon},z_{gaussian}\\}" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": ". \n6: if err " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "(\\hat{y}', y) > Threshold" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": " then \n7: if " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "\\text{argmax}(z_{lat}) \\neq \\text{argmax}(z_{lat}')" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": " then \n8: Update model " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": " with higher weights of semi-supervised learning loss " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "Loss_{semi}" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": " in Eq. (6). \n9: Update model " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "Loss_{D}" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": " in Eq. (4), " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "Loss_{G}" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": " in Eq. (3), and " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "Loss_{traj}" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": " in Eq. (5). \n10: else \n11: Update model " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "Loss_{D}" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": " in Eq. (4), " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "Loss_{G}" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": " in Eq. (3), " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "Loss_{semi}" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": " in Eq. (6), and " + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "inline_equation", + "content": "Loss_{traj}" + }, + { + "bbox": [ + 311, + 180, + 547, + 457 + ], + "type": "text", + "content": " in Eq. (5). \n12: end if \n13: end if \n14: end for" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "code_body" + } + ], + "index": 15, + "sub_type": "algorithm" + }, + { + "bbox": [ + 306, + 518, + 485, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 518, + 485, + 530 + ], + "spans": [ + { + "bbox": [ + 306, + 518, + 485, + 530 + ], + "type": "text", + "content": "3.2.2 Balance Accuracy and Robustness" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": "In our preliminary experiments, we notice a trade-off between standard accuracy and adversarial robustness. A similar phenomenon has been observed in classification tasks [46, 28]. Methods such as TRADES [46], robust self-training [3] and MixUp [47, 1] have been proposed to balance such trade-off. However, there are few methods that can be applied to trajectory prediction because such time-series regression problems have no class labels and are more sensitive to errors introduced by augmented data. In this work, we utilize the MixUp [47] technique to mix the adversarial scenarios and benign scenarios in the adversarial training process. The experiments demonstrate that a balance between adversarial robustness and standard accuracy can be achieved in trajectory prediction, as shown later." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8212" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 103, + 533, + 224 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "text", + "content": "Table 2: Comparison of different defense methods when under various attacks and in the benign case. For the adversarial training methods, we calculate the mean error of models adversarially trained on different attacks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 103, + 533, + 224 + ], + "lines": [ + { + "bbox": [ + 61, + 103, + 533, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 533, + 224 + ], + "type": "table", + "html": "
MethodsADE AttackLateral AttackLongitudinal AttackBenign
ADE/mLateral Error/mLongitudinal Error/mADE/m
DatasetArgo1Argo2ApolloArgo1Argo2ApolloArgo1Argo2ApolloArgo1Argo2Apollo
Original Model5.174.523.931.661.271.543.783.173.421.430.791.78
Train-time Smo. [48]4.674.323.861.451.271.493.233.003.411.500.832.11
Test-time Smo. [48]4.323.513.490.750.891.093.542.423.051.681.232.07
Heuristic Aug. [29]4.623.523.841.091.211.343.244.243.211.540.972.01
Data-driven Aug. [29]4.534.433.530.740.780.542.732.613.022.501.992.24
Standard AT3.793.673.680.670.641.152.441.893.051.671.061.87
SSAT (ours)3.512.672.900.610.530.411.761.441.261.751.201.87
Mixup-SSAT (ours)3.733.333.170.720.510.532.131.631.351.641.071.86
", + "image_path": "99d3903851becc8cbcd2232ccdd8a7e5bbad83b3c8b25749bdf19e97a37310e7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 243, + 173, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 243, + 173, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 243, + 173, + 255 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 262, + 155, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 155, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 155, + 275 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 281, + 113, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 113, + 292 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 113, + 292 + ], + "type": "text", + "content": "4.1.1 Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 299, + 288, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 299, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 288, + 431 + ], + "type": "text", + "content": "We train and evaluate different defense methods using three popular benchmarks - Argoverse 1 [4], Argoverse 2 [42], and ApolloScape datasets [9]. The datasets contain more than " + }, + { + "bbox": [ + 46, + 299, + 288, + 431 + ], + "type": "inline_equation", + "content": "250\\mathrm{k}" + }, + { + "bbox": [ + 46, + 299, + 288, + 431 + ], + "type": "text", + "content": " real driving scenarios in different cities, such as Miami and Pittsburgh. For Argoverse 1 and Argoverse 2, each scenario consists of a road graph and multiple agents' trajectories sampled at a frequency of " + }, + { + "bbox": [ + 46, + 299, + 288, + 431 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 299, + 288, + 431 + ], + "type": "text", + "content": ". We choose 20 waypoints as the history trajectory and the models will predict 30 waypoints in the future. Scenarios in ApolloScape are simpler. They have no maps but 6 waypoints for both history and future trajectories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 445, + 146, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 146, + 458 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 146, + 458 + ], + "type": "text", + "content": "4.1.2 Attack Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 464, + 287, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 572 + ], + "type": "text", + "content": "In the experiments, we study three different types of attacks [48] to the vehicle trajectory prediction algorithms - lateral directional attack (shift to the right), longitudinal directional attack (shift forward), and ADE attack (deviate randomly). More details of them are in Section 2.1. The trajectory prediction models are adversarially trained on the three types of attacks, respectively. We constrain the maximum deviation between the attacked and the benign input trajectories to be 1 meter." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 586, + 153, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 153, + 599 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 153, + 599 + ], + "type": "text", + "content": "4.1.3 Training Settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Since our architecture is an encoder-decoder module that can be combined with different feature extractors, we first fine-tune the model on the benign data. In the experiments, we use the feature extractor from LaneGCN [16], an attention-based graph neural network. We notice that the AAE architecture introduces a slight accuracy drop on the benign data, mainly due to the dimension reduction. For adversarial training, we train prediction models on adversarial samples generated from scenarios in datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 243, + 493, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 243, + 493, + 256 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 493, + 256 + ], + "type": "text", + "content": "4.2. Experimental Results and Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 262, + 547, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 547, + 453 + ], + "type": "text", + "content": "In this section, we conduct experiments with various defense methods under different patterns of attacks, including our semi-supervised semantics-guided (SSAT) method and a Mixup-SSAT method that combines SSAT with the MixUp technique for balancing standard accuracy and adversarial robustness. In the following, we first compare the average robustness improvement among various defense methods, which demonstrates the advantage of SSAT in improving robustness under various types of attacks and the effectiveness of Mixup-SSAT in balancing robustness and accuracy. Then, we show that SSAT can significantly enhance the robust generalization to unseen types of attacks. In addition, we evaluate an unsupervised version of SSAT to explicitly show how the semi-supervised semantic-guided latent space modeling can boost the adversarial robustness, which also serves as an ablation study." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 467, + 471, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 467, + 471, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 467, + 471, + 479 + ], + "type": "text", + "content": "4.2.1 Effectiveness of SSAT Methods" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 486, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 653 + ], + "type": "text", + "content": "We compare our SSAT and Mixup-SSAT methods with the original model and five different defense methods, including train-time smoothing [48], test-time smoothing [48], heuristic data augmentation [29], data-driven augmentation [29], and the standard adversarial training (Standard AT). All the models share the same feature extractor LaneGCN in this setting. Note that we compare to the two data augmentation methods as they are effective for image classification tasks [6, 29, 35]. For the data-driven augmentation, we design an additional decoder to augment input trajectories and it can generate more inputs by adding Gaussian noises to the latent vectors of the real inputs. For the heuristic augmentation, we simply add random perturbation to the benign inputs, with the same constraints of maximum deviation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Table 2 shows the prediction errors of different methods when under various types of attacks and in the benign case (note that when under lateral and longitudinal attacks, we measure lateral and longitudinal errors). We can see that our SSAT method significantly outperforms all other de" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8213" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 79, + 197, + 186 + ], + "blocks": [ + { + "bbox": [ + 70, + 79, + 197, + 186 + ], + "lines": [ + { + "bbox": [ + 70, + 79, + 197, + 186 + ], + "spans": [ + { + "bbox": [ + 70, + 79, + 197, + 186 + ], + "type": "image", + "image_path": "aa1253b2226a86d5e0b88a2cdffb393e62baeed3c97f70f7621773b97277dca5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 190, + 139, + 198 + ], + "lines": [ + { + "bbox": [ + 129, + 190, + 139, + 198 + ], + "spans": [ + { + "bbox": [ + 129, + 190, + 139, + 198 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 230, + 79, + 358, + 186 + ], + "blocks": [ + { + "bbox": [ + 230, + 79, + 358, + 186 + ], + "lines": [ + { + "bbox": [ + 230, + 79, + 358, + 186 + ], + "spans": [ + { + "bbox": [ + 230, + 79, + 358, + 186 + ], + "type": "image", + "image_path": "e6cee0f856483597832978a138c131ea0afc0ce22d9f48def8bf231c285ee866.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 190, + 301, + 198 + ], + "lines": [ + { + "bbox": [ + 290, + 190, + 301, + 198 + ], + "spans": [ + { + "bbox": [ + 290, + 190, + 301, + 198 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 212, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 212, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 547, + 248 + ], + "type": "text", + "content": "Figure 4: Adversarial robustness comparisons of the original model, the standard adversarial training on final trajectories (Standard AT), and the SSAT method proposed by us. Figures (a), (b), (c) show results from adversarial training (AT) that targets ADE attacks, lateral attacks, and longitudinal attacks, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 392, + 79, + 518, + 186 + ], + "blocks": [ + { + "bbox": [ + 392, + 79, + 518, + 186 + ], + "lines": [ + { + "bbox": [ + 392, + 79, + 518, + 186 + ], + "spans": [ + { + "bbox": [ + 392, + 79, + 518, + 186 + ], + "type": "image", + "image_path": "1bbcab51993c0b022ee18572adfccec7179ed6b2333c7d2b9ac9d51a37701c18.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 190, + 461, + 198 + ], + "lines": [ + { + "bbox": [ + 451, + 190, + 461, + 198 + ], + "spans": [ + { + "bbox": [ + 451, + 190, + 461, + 198 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 255, + 286, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 286, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 286, + 302 + ], + "type": "text", + "content": "fense methods in improving the robustness of trajectory prediction. Compared with the original model, SSAT can reduce the prediction error by " + }, + { + "bbox": [ + 46, + 255, + 286, + 302 + ], + "type": "inline_equation", + "content": "32\\% - 73\\%" + }, + { + "bbox": [ + 46, + 255, + 286, + 302 + ], + "type": "text", + "content": " when under different types of attacks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 305, + 286, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 286, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 286, + 400 + ], + "type": "text", + "content": "Moreover, while our SSAT method improves robustness, we also observe a drop in standard accuracy in the benign case. Mixup-SSAT enables effective trade-off between these two objectives (i.e., better performance than SSAT in benign cases at the expense of worse performance under attacks), by setting the mixup ratio between adversarial and benign examples to different values (the results in Table 2 are based on a mixup ratio of 2)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 403, + 286, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 286, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 286, + 474 + ], + "type": "text", + "content": "We also notice that both data-driven and heuristic data augmentation methods offer very limited improvement over the original model. This is likely due to the challenge of regression tasks with rich context, which makes augmentation methods perform poorer than they do for image classification tasks." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 67, + 533, + 267, + 584 + ], + "blocks": [ + { + "bbox": [ + 46, + 488, + 286, + 525 + ], + "lines": [ + { + "bbox": [ + 46, + 488, + 286, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 286, + 525 + ], + "type": "text", + "content": "Table 3: Comparison of different methods when adversarial training is conducted for ADE attack, and tested on ADE, lateral, and longitudinal attacks." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 533, + 267, + 584 + ], + "lines": [ + { + "bbox": [ + 67, + 533, + 267, + 584 + ], + "spans": [ + { + "bbox": [ + 67, + 533, + 267, + 584 + ], + "type": "table", + "html": "
MethodsPrediction Error/m, w/o AT → with AT
ADE AttackLat AttackLon Attack
Standard AT5.17 → 3.761.58 → 0.833.59 → 2.66
SSAT5.16 → 3.241.66 → 0.733.78 → 1.45
Unsup-SSAT5.16 → 3.481.66 → 0.743.78 → 1.49
", + "image_path": "4f7872b9bf072a0eacea54bec277099b8aec978fb286742787ed1b60d5411e05.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 67, + 661, + 267, + 711 + ], + "blocks": [ + { + "bbox": [ + 46, + 616, + 286, + 652 + ], + "lines": [ + { + "bbox": [ + 46, + 616, + 286, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 286, + 652 + ], + "type": "text", + "content": "Table 4: Comparison of different methods when adversarial training is conducted for lateral attack, and tested on ADE, lateral, and longitudinal attacks." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 661, + 267, + 711 + ], + "lines": [ + { + "bbox": [ + 67, + 661, + 267, + 711 + ], + "spans": [ + { + "bbox": [ + 67, + 661, + 267, + 711 + ], + "type": "table", + "html": "
MethodsPrediction Error/m, w/o AT → with AT
ADE AttackLat AttackLon Attack
Standard AT5.17 → 4.281.58 → 0.593.59 → 3.01
SSAT5.16 → 4.021.66 → 0.593.78 → 2.52
Unsup-SSAT5.16 → 4.111.66 → 0.693.78 → 2.54
", + "image_path": "2f921a19736a0eaa01d3421c9c5d854abb183354a0599d379d2e7e216aff15db.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 255, + 545, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 255, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 306, + 255, + 545, + 279 + ], + "type": "text", + "content": "4.2.2 Effectiveness of SSAT in Robust Generalization on Different Types of Attacks" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 285, + 545, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 501 + ], + "type": "text", + "content": "We observe that there is an adversarial robust generalization gap when the training and test are under different types of attacks. The comparisons in Tables 3, 4, 5 show that our SSAT method is better generalized to unseen types of attacks, compared to the standard adversarial training. In every training scenario, our method is more robust to the various unseen patterns of attacks. For instance, Table 3 shows that when applying SSAT to train under the random ADE attack, its results outperform other models on all seen (i.e., ADE) and unseen (i.e., lateral and longitudinal) types of attacks, which demonstrates that our SSAT method can effectively decompose and learn semantic features from random ADE attacks. Tables 4 and 5 show similar trends, where our SSAT methods are better at defending against unseen attacks and mitigating overfitting on specific patterns of attacks. Fig. 4 further visualizes the results from these three tables for the original model, the standard adversarial training (Standard AT), and SSAT." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 326, + 556, + 526, + 607 + ], + "blocks": [ + { + "bbox": [ + 305, + 511, + 545, + 548 + ], + "lines": [ + { + "bbox": [ + 305, + 511, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 305, + 511, + 545, + 548 + ], + "type": "text", + "content": "Table 5: Comparison of different methods when adversarial training is conducted for longitudinal attack, and tested on ADE, lateral, and longitudinal attacks." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 556, + 526, + 607 + ], + "lines": [ + { + "bbox": [ + 326, + 556, + 526, + 607 + ], + "spans": [ + { + "bbox": [ + 326, + 556, + 526, + 607 + ], + "type": "table", + "html": "
MethodsPrediction Error/m, w/o AT → with AT
ADE AttackLat AttackLon Attack
Standard AT5.17 → 3.331.58 → 0.603.59 → 1.64
SSAT5.16 → 3.281.66 → 0.523.78 → 1.30
Unsup-SSAT5.16 → 3.401.66 → 0.613.78 → 1.49
", + "image_path": "7b2736180437cf2c0f4416e8b3233a059fc090081b6969ef4e0f277b75347fea.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 623, + 494, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 494, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 494, + 635 + ], + "type": "text", + "content": "4.2.3 Impact from Latent Space Modeling" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "We also conduct adversarial training that only has regularization on the latent distributions but without supervision on latent vectors. We name it Unsup-SSAT. The comparisons between the standard adversarial training (Standard AT) and Unsup-SSAT in Tables 3, 4, and 5 demonstrate that even without labels, the partial disentanglement and" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8214" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "distribution modeling in Unsup-SSAT will benefit the adversarial training on trajectory prediction and outperform the baseline Standard AT in most cases. However, when compared with SSAT, we find that the adversarial robustness will be further improved with the extra labels in the semi-supervised phase (in practice, we often have access to those labels)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 167, + 119, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 167, + 119, + 179 + ], + "spans": [ + { + "bbox": [ + 47, + 167, + 119, + 179 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 187, + 289, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 187, + 289, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 289, + 319 + ], + "type": "text", + "content": "In this work, we propose an adversarial training method for trajectory prediction. To tackle the challenge of random inputs with rich context, diverse types of attacks, and lack of class labels, we develop a novel AAE architecture that exploits the disentanglement and semantic features for enhancing model robustness and its generalization. Our proposed SSAT method significantly outperforms a number of baselines from the literature, reducing the prediction errors under attacks by up to " + }, + { + "bbox": [ + 46, + 187, + 289, + 319 + ], + "type": "inline_equation", + "content": "32\\% - 73\\%" + }, + { + "bbox": [ + 46, + 187, + 289, + 319 + ], + "type": "text", + "content": " when compared with the original prediction models. Our method is also shown to be effective for defending against unseen attacks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 330, + 140, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 330, + 140, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 330, + 140, + 342 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 350, + 287, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 287, + 385 + ], + "type": "text", + "content": "We gratefully acknowledge the support from NSF grants 1834701, 1724341, 2038853, 1929771, 2145493, 1932464, and ONR grant N00014-19-1-2496." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 396, + 106, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 396, + 106, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 396, + 106, + 409 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 416, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 53, + 416, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 416, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 416, + 287, + 464 + ], + "type": "text", + "content": "[1] Guillaume P Archambault, Yongyi Mao, Hongyu Guo, and Richong Zhang. Mixup as directional adversarial training. arXiv preprint arXiv:1906.06875, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 472, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 472, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 472, + 288, + 533 + ], + "type": "text", + "content": "[2] Yulong Cao, Danfei Xu, Xinshuo Weng, Zhuoqing Mao, Anima Anandkumar, Chaowei Xiao, and Marco Pavone. Robust trajectory prediction against adversarial attacks. In 6th Annual Conference on Robot Learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 541, + 289, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 541, + 289, + 590 + ], + "spans": [ + { + "bbox": [ + 53, + 541, + 289, + 590 + ], + "type": "text", + "content": "[3] Yair Carmon, Aditi Raghunathan, Ludwig Schmidt, John C Duchi, and Percy S Liang. Unlabeled data improves adversarial robustness. Advances in Neural Information Processing Systems, 32, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 597, + 288, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 597, + 288, + 681 + ], + "spans": [ + { + "bbox": [ + 53, + 597, + 288, + 681 + ], + "type": "text", + "content": "[4] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8748-8757, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 689, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 689, + 288, + 714 + ], + "type": "text", + "content": "[5] Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 326, + 72, + 547, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 547, + 120 + ], + "type": "text", + "content": "Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11525-11533, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 128, + 547, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 128, + 547, + 187 + ], + "spans": [ + { + "bbox": [ + 312, + 128, + 547, + 187 + ], + "type": "text", + "content": "[6] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 34:4218-4233, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 195, + 547, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 195, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 312, + 195, + 547, + 243 + ], + "type": "text", + "content": "[7] Duy-Hung Ha, Maurice Aron, and Simon Cohen. Time headway variable and probabilistic modeling. Transportation Research Part C: Emerging Technologies, 25:181-201, 2012." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 251, + 547, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 251, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 312, + 251, + 547, + 300 + ], + "type": "text", + "content": "[8] Michael W Hancock and Bud Wright. A policy on geometric design of highways and streets. American Association of State Highway and Transportation Officials: Washington, DC, USA, 2013." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 307, + 547, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 307, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 312, + 307, + 547, + 380 + ], + "type": "text", + "content": "[9] Xinyu Huang, Xinjing Cheng, Qichuan Geng, Binbin Cao, Dingfu Zhou, Peng Wang, Yuanqing Lin, and Ruigang Yang. The apolloscape dataset for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 954-960, 2018." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 386, + 545, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 386, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 307, + 386, + 545, + 434 + ], + "type": "text", + "content": "[10] Ruochen Jiao, Juyang Bai, Xiangguo Liu, Takami Sato, Xiaowei Yuan, Qi Alfred Chen, and Qi Zhu. Learning representation for anomaly detection of vehicle trajectories. 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 442, + 547, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 442, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 442, + 547, + 503 + ], + "type": "text", + "content": "[11] Ruochen Jiao, Hengyi Liang, Takami Sato, Junjie Shen, Qi Alfred Chen, and Qi Zhu. End-to-end uncertainty-based mitigation of adversarial attacks to automated lane centering. In 2021 IEEE Intelligent Vehicles Symposium (IV), pages 266-273. IEEE, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 510, + 547, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 547, + 582 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 547, + 582 + ], + "type": "text", + "content": "[12] Ruochen Jiao, Xiangguo Liu, Bowen Zheng, Dave Liang, and Qi Zhu. Tae: A semi-supervised controllable behavior-aware trajectory generator and predictor. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 12534-12541. IEEE, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 590, + 547, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 547, + 625 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 547, + 625 + ], + "type": "text", + "content": "[13] Diederik Kingma and Max Welling. An introduction to variational autoencoders. arXiv preprint arXiv:1906.02691, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 633, + 547, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 633, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 633, + 547, + 669 + ], + "type": "text", + "content": "[14] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 677, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 677, + 547, + 714 + ], + "type": "text", + "content": "[15] Dasom Lee and David J Hess. Regulations for on-road testing of connected and automated vehicles: Assessing the potential for global safety harmonization." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8215" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 69, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 286, + 95 + ], + "type": "text", + "content": "Transportation Research Part A: Policy and Practice, 136:85-98, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 102, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 102, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 102, + 287, + 162 + ], + "type": "text", + "content": "[16] Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph representations for motion forecasting. In European Conference on Computer Vision, pages 541-556. Springer, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 168, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 168, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 168, + 287, + 239 + ], + "type": "text", + "content": "[17] C. Lin, Bowen Zheng, Qi Zhu, and Alberto Sangiovanni-Vincentelli. Security-Aware Design Methodology and Optimization for Automotive Systems. ACM Transactions on Design Automation of Electronic Systems (TODAES), 21(1):18:1-18:26, December 2015." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 246, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 246, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 48, + 246, + 287, + 317 + ], + "type": "text", + "content": "[18] Xiangguo Liu, Chao Huang, Yixuan Wang, Bowen Zheng, and Qi Zhu. Physics-aware safety-assured design of hierarchical neural network based planner. In 2022 ACM/IEEE 13th International Conference on Cyber-Physical Systems (ICCPS), pages 137-146. IEEE, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 324, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 324, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 48, + 324, + 287, + 384 + ], + "type": "text", + "content": "[19] Xiangguo Liu, Ruochen Jiao, Yixuan Wang, Yimin Han, Bowen Zheng, and Qi Zhu. Safety-assured speculative planning with adaptive prediction. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 390, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 390, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 390, + 287, + 449 + ], + "type": "text", + "content": "[20] Xiangguo Liu, Ruochen Jiao, Bowen Zheng, Dave Liang, and Qi Zhu. Safety-driven interactive planning for neural network-based lane changing. In Proceedings of the 28th Asia and South Pacific Design Automation Conference, pages 39-45, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 456, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 540 + ], + "type": "text", + "content": "[21] Xiangguo Liu, Yunpeng Luo, Anthony Goeckner, Trishna Chakraborty, Ruochen Jiao, Ningfei Wang, Yixuan Wang, Takami Sato, Qi Alfred Chen, and Qi Zhu. Invited: Waving the double-edged sword: Building resilient cavs with edge and cloud computing. In Proceedings of the 60th Annual Design Automation Conference, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 545, + 287, + 593 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 593 + ], + "type": "text", + "content": "[22] Xiangguo Liu, Neda Masoud, and Qi Zhu. Impact of sharing driving attitude information: A quantitative study on lane changing. In 2020 IEEE Intelligent Vehicles Symposium (IV), pages 1998-2005. IEEE, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 600, + 287, + 659 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 659 + ], + "type": "text", + "content": "[23] Yicheng Liu, Jinghuai Zhang, Liangji Fang, Qinhong Jiang, and Bolei Zhou. Multimodal motion prediction with stacked transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7577-7586, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 666, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 666, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 666, + 287, + 713 + ], + "type": "text", + "content": "[24] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 120 + ], + "type": "text", + "content": "[25] Pratyush Maini, Eric Wong, and Zico Kolter. Adversarial robustness against the union of multiple perturbation models. In International Conference on Machine Learning, pages 6640-6650. PMLR, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 126, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 126, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 126, + 545, + 163 + ], + "type": "text", + "content": "[26] Alireza Makhzani, Jonathon Shlens, Navdeep Jaitly, Ian Goodfellow, and Brendan Frey. Adversarial autoencoders. arXiv preprint arXiv:1511.05644, 2015." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 169, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 169, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 169, + 545, + 228 + ], + "type": "text", + "content": "[27] Emile Mathieu, Tom Rainforth, Nana Siddharth, and Yee Whye Teh. Disentangling disentanglement in variational autoencoders. In International Conference on Machine Learning, pages 4402-4412. PMLR, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 236, + 545, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 236, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 307, + 236, + 545, + 283 + ], + "type": "text", + "content": "[28] Aditi Raghunathan, Sang Michael Xie, Fanny Yang, John Duchi, and Percy Liang. Understanding and mitigating the tradeoff between robustness and accuracy. arXiv preprint arXiv:2002.10716, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 290, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 290, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 290, + 545, + 338 + ], + "type": "text", + "content": "[29] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan A Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Fixing data augmentation to improve adversarial robustness. arXiv preprint arXiv:2103.01946, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 345, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 345, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 345, + 545, + 392 + ], + "type": "text", + "content": "[30] Leslie Rice, Eric Wong, and Zico Kolter. Overfitting in adversarially robust deep learning. In International Conference on Machine Learning, pages 8093-8104. PMLR, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 399, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 459 + ], + "type": "text", + "content": "[31] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory++: Dynamically-feasible trajectory forecasting with heterogeneous data. In European Conference on Computer Vision, pages 683-700. Springer, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 466, + 545, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 537 + ], + "type": "text", + "content": "[32] Takami Sato, Junjie Shen, Ningfei Wang, Yunhan Jia, Xue Lin, and Qi Alfred Chen. Dirty road can attack: Security of deep learning based automated lane centering under {Physical-World} attack. In 30th USENIX Security Symposium (USENIX Security 21), pages 3309–3326, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 544, + 545, + 593 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 593 + ], + "type": "text", + "content": "[33] Ludwig Schmidt, Shibani Santurkar, Dimitris Tsipras, Kunal Talwar, and Aleksander Madry. Adversarily robust generalization requires more data. Advances in neural information processing systems, 31, 2018." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 599, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 599, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 599, + 545, + 646 + ], + "type": "text", + "content": "[34] Lukas Schott, Jonas Rauber, Matthias Bethge, and Wieland Brendel. Towards the first adversarially robust neural network model on mnist. arXiv preprint arXiv:1805.09190, 2018." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 653, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 653, + 545, + 713 + ], + "type": "text", + "content": "[35] Vikash Sehwag, Saeed Mahloujifar, Tinashe Handina, Sihui Dai, Chong Xiang, Mung Chiang, and Prateek Mittal. Robust learning meets generative models: Can proxy distributions improve adversarial robustness? arXiv preprint arXiv:2104.09425, 2021." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8216" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 144 + ], + "type": "text", + "content": "[36] Amolika Sinha, Sai Chand, Kasun P Wijayaratna, Navreet Virdi, and Vinayak Dixit. Comprehensive safety assessment in mixed fleets with connected and automated vehicles: A crash severity and rate evaluation of conventional vehicles. Accident Analysis & Prevention, 142:105567, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 152, + 287, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 202 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 202 + ], + "type": "text", + "content": "[37] Chuanbiao Song, Kun He, Jiadong Lin, Liwei Wang, and John E Hopcroft. Robust local features for improving the generalization of adversarial training. arXiv preprint arXiv:1909.10147, 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 209, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 281 + ], + "type": "text", + "content": "[38] Dong Su, Huan Zhang, Hongge Chen, Jinfeng Yi, Pin-Yu Chen, and Yupeng Gao. Is robustness the cost of accuracy? a comprehensive study on the robustness of 18 deep image classification models. In Proceedings of the European Conference on Computer Vision (ECCV), pages 631-648, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 289, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 289, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 289, + 287, + 338 + ], + "type": "text", + "content": "[39] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 346, + 287, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 346, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 48, + 346, + 287, + 406 + ], + "type": "text", + "content": "[40] Zhilu Wang, Chao Huang, and Qi Zhu. Efficient global robustness certification of neural networks via interleaving twin-network encoding. In DATE'22: Proceedings of the Conference on Design, Automation and Test in Europe, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 415, + 287, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 462 + ], + "type": "text", + "content": "[41] Matthew Willetts, Alexander Camuto, Tom Rainforth, Stephen Roberts, and Chris Holmes. Improving vaes' robustness to adversarial attack. arXiv preprint arXiv:1906.00230, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 471, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 567 + ], + "type": "text", + "content": "[42] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, et al. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 575, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 575, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 575, + 287, + 623 + ], + "type": "text", + "content": "[43] Dongxian Wu, Shu-Tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization. Advances in Neural Information Processing Systems, 33:2958-2969, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 632, + 287, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 632, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 48, + 632, + 287, + 681 + ], + "type": "text", + "content": "[44] Hongtao Yu, H Eric Tseng, and Reza Langari. A human-like game theory-based controller for automatic lane changing. Transportation Research Part C: Emerging Technologies, 88:140-158, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 689, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 689, + 287, + 713 + ], + "type": "text", + "content": "[45] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M Kitani. Agentformer: Agent-aware transformers for" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 686 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 327, + 72, + 545, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 327, + 72, + 545, + 108 + ], + "type": "text", + "content": "socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9813-9823, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 116, + 545, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 116, + 545, + 176 + ], + "spans": [ + { + "bbox": [ + 307, + 116, + 545, + 176 + ], + "type": "text", + "content": "[46] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 184, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 545, + 230 + ], + "type": "text", + "content": "[47] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 239, + 545, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 311 + ], + "type": "text", + "content": "[48] Qingzhao Zhang, Shengtuo Hu, Jiachen Sun, Qi Alfred Chen, and Z Morley Mao. On adversarial robustness of trajectory prediction for autonomous vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15159-15168, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 319, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 380 + ], + "type": "text", + "content": "[49] Shufei Zhang, Zhuang Qian, Kaizhu Huang, Qiufeng Wang, Rui Zhang, and Xinping Yi. Towards better robust generalization with shift consistency regularization. In International Conference on Machine Learning, pages 12524-12534. PMLR, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 387, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 545, + 447 + ], + "type": "text", + "content": "[50] Bowen Zheng, Hengyi Liang, Qi Zhu, Huafeng Yu, and Chung-Wei Lin. Next generation automotive architecture modeling and exploration for autonomous driving. In 2016 IEEE Computer Society Annual Symposium on VLSI (ISVLSI), pages 53–58, 2016." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 455, + 545, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 501 + ], + "type": "text", + "content": "[51] Bowen Zheng, Chung-Wei Lin, et al. Design and Analysis of Delay-Tolerant Intelligent Intersection Management. ACM T-CPS, 4(1):3:1-3:27, November 2019." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 510, + 545, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 582 + ], + "type": "text", + "content": "[52] Qi Zhu, Chao Huang, Ruochen Jiao, Shuyue Lan, Hengyi Liang, Xiangguo Liu, Yixuan Wang, Zhilu Wang, and Shichao Xu. Safety-assured design and adaptation of learning-enabled autonomous systems. In Proceedings of the 26th Asia and South Pacific Design Automation Conference, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 590, + 545, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 686 + ], + "type": "text", + "content": "[53] Qi Zhu, Wenchao Li, Hyoseung Kim, Yecheng Xiang, Kacper Wardega, Zhilu Wang, Yixuan Wang, Hengyi Liang, Chao Huang, Jiameng Fan, and Hyunjong Choi. Know the unknowns: Addressing disturbances and uncertainties in autonomous systems. In Proceedings of the 39th International Conference on Computer-Aided Design, ICCAD '20, New York, NY, USA, 2020. Association for Computing Machinery." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8217" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding/ebc67249-8be2-4ca5-975d-54a569184984_content_list.json b/2023/Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding/ebc67249-8be2-4ca5-975d-54a569184984_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..80a87dda344dcc267883230da1adadd3696cd47d --- /dev/null +++ b/2023/Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding/ebc67249-8be2-4ca5-975d-54a569184984_content_list.json @@ -0,0 +1,1465 @@ +[ + { + "type": "text", + "text": "Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding", + "text_level": 1, + "bbox": [ + 84, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Peiji Yang* Huawei Wei* Yicheng Zhong* Zhisheng Wang \nTencent, Shenzhen, China", + "bbox": [ + 233, + 180, + 733, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{peijiyang, huaweiwei, ajaxhong, plorywang}@tencent.com", + "bbox": [ + 236, + 219, + 732, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing Speech-driven 3D facial animation methods typically follow the supervised paradigm, involving regression from speech to 3D facial animation. This paradigm faces two major challenges: the high cost of supervision acquisition, and the ambiguity in mapping between speech and lip movements. To address these challenges, this study proposes a novel cross-modal semi-supervised framework, comprising a Speech-to-Image Transcoder and a Face-to-Geometry Regressor. The former jointly learns a common representation space from speech and image domains, enabling the transformation of speech into semantically-consistent facial images. The latter is responsible for reconstructing 3D facial meshes from the transformed images. Both modules require minimal effort to acquire the necessary training data, thereby obviating the dependence on costly supervised data. Furthermore, the joint learning scheme enables the fusion of intricate visual features into speech encoding, thereby facilitating the transformation of subtle speech variations into nuanced lip movements, ultimately enhancing the fidelity of 3D face reconstructions. Consequently, the ambiguity of the direct mapping of speech-to-animation is significantly reduced, leading to coherent and high-fidelity generation of lip motion. Extensive experiments demonstrate that our approach produces competitive results compared to supervised methods.", + "bbox": [ + 75, + 301, + 472, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 710, + 209, + 726 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Speech-driven 3D facial animation aims to automatically animate vivid facial expressions of realistic 3D avatars from the speech signal. It has become increasingly popular in many fields, such as games, virtual reality, film production, and online communication, as it enables the generation of lifelike facial expressions with minimal effort.", + "bbox": [ + 75, + 736, + 468, + 827 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently proposed approaches, including VOCA[7], MeshTalk[28], and FaceFormer[10], have shown promising results in recovering 3D facial meshes from speech", + "bbox": [ + 75, + 828, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "signals through the utilization of regressive networks or transformer-based autoregressive models. However, these methods, which fall under the supervised paradigm, encounter two primary challenges. Firstly, obtaining paired supervision of speech and high-fidelity 3D facial animation requires the utilization of a professional and expensive facial motion capture system, leading to considerable costs. Secondly, the mapping of low-dimensional speech signals to high-dimensional 3D facial meshes with significant variability may result in ambiguity issues, leading to suboptimal outcomes.", + "bbox": [ + 500, + 268, + 893, + 435 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This study for the first time proposes a novel cross-modal semi-supervised framework to address the above issues, which consists of a Speech-to-Image Transcoder and a Face-to-Geometry Regressor. The former is designed to transform speech into semantically-consistent facial images, while the latter is responsible for reconstructing 3D facial meshes from the transformed images. The design of the Speech-to-Image Transcoder is inspired by the recent success of unsupervised image-to-image translation [26], which allows for transformation between different input image domains. Our transcoder extends this mechanism to facilitate cross-modal conversion between speech and image domains. To achieve this capability, we learn a common representation space using data from three domains: paired speech and real facial images, as well as synthetic facial images. We train the translation between the real and synthetic image domains in an unsupervised manner, and simultaneously leverage the paired relationship between real images and speech to project speech features into the image representation space. Through this joint training, we construct a common space where the representations of speech and images are tied together. With the learned common representation, we can convert speech into synthetic facial images using our domain-dependent image decoder. Furthermore, we leverage the synthetic facial data, which is paired with rendered facial images and 3D face meshes, to train the Face-to-Geometry Regressor. This allows us to infer the corresponding 3D face meshes from the synthetic faces generated from speech.", + "bbox": [ + 498, + 441, + 893, + 880 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our pipeline is capable of effortlessly acquiring the nec", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 94, + 886, + 351, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21032", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "essary data, which includes synchronized speech and real facial images extracted from readily available video clips, as well as synthetic facial images generated through the deformation of a polygonal face mesh using a rendering engine. Notably, our approach is not bound by any prerequisites for paired speech and 3D face animation, thereby effectively addressing the issue of supervision data scarcity.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, our joint training scheme integrates the speech-to-image regression and image-to-image reconstruction into a cohesive framework, which enables the incorporation of detailed visual features into speech encoding. Leveraging the domain-dependent image decoders, which store intricate facial visual priors, we are able to convert nuanced speech variations into micro-expressions that are rich in detail, thereby facilitating the subsequent generation of high-fidelity 3D face reconstructions. Consequently, the uncertainties associated with the direct mapping of speech-to-animation in the supervised paradigm are considerably reduced, leading to the generation of lip motion with finer granularity.", + "bbox": [ + 75, + 198, + 470, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the main contributions of our research are:", + "bbox": [ + 96, + 393, + 464, + 407 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We make the first attempt to build a semi-supervised framework for speech-driven 3D facial animation, which eliminates the need for paired speech and 3D animation.", + "- Our proposed joint training scheme leverages cross-modal translation to convert subtle speech variations into images that are rich in detail, resulting in more high-fidelity lip motion generation.", + "- Our extensive experiments and user studies demonstrate that our approach produces competitive results compared to supervised methods." + ], + "bbox": [ + 96, + 419, + 468, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 618, + 218, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed method is related to several research fields, such as unsupervised image-to-image translation, speech-driven 2D talking head, and speech-driven 3D face animation. In the following, we review the most relevant approaches.", + "bbox": [ + 75, + 643, + 468, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised image-to-image translation. The unsupervised image-to-image translation aims to convert images from one domain to another domain with certain characteristics of the source images preserved[22, 36]. The translation involves converting from silhouettes to real images [16], from segmentation masks to RGB images[8, 23], from face to face [5, 27], and so on. This technology has wide applications in many fields, such as style transfer [40], face swapping [5, 26, 27], expression transfer [25], pose estimation [20] etc. They typically employ GAN-based networks [40] or encoder-decoder architectures [16] to achieve the goals. Notably, they only focus on image domains,", + "bbox": [ + 75, + 719, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "our method extends this mechanism to achieve cross-modal translation.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Speech-driven 2D talking head. 2D-based talking head generation methods generally utilize different image representations to assist the conversion of speech to 2D faces, such as facial landmarks[4, 31, 37], depth maps[15] and semantic maps[6, 24]. The discriminator loss and other task-specific losses are usually employed for optimization. For example, Chen et al.[3] design an audio-visual derivative correlation loss to ensure cross-modal consistency. In addition, several works use 3D information such as 3D Morphable Model (3DMM)[2, 19, 30, 32, 39] parameters or dynamic neural radiance fields (NeRF)[12, 13, 35] to guide the generation process. Unlike these 2D-based methods, this paper focuses on high-fidelity 3D animation generation from speech.", + "bbox": [ + 496, + 119, + 890, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Speech-driven 3D face animation. Recently, considerable endeavors have been undertaken in this field[7, 10, 11, 18, 28, 34, 38]. VOCA[7] casts the mapping from speech to animation as a regression issue. To train the regressor, it uses the paired data of speech and 3D face animation. FaceFormer[10] employs transformers to model the long-term dependencies of speech, leading to improved performance. Meshtalk[28] puts emphasis on addressing the scalability and realism of the model by decoupling audio-correlated and audio-uncorrelated information. All the above methods belong to the supervised paradigm, which requires costly paired data and suffers from the problem of ambiguous mapping when directly regressing 3D animation from speech. This paper develops a semi-supervised framework to convert speech into images and then reconstruct 3D faces from the transformed images, effectively alleviating the aforementioned two issues.", + "bbox": [ + 496, + 333, + 890, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 602, + 635, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed semi-supervised method is divided into two stages. We first employ the Speech-to-Image Transcoder to learn a common representation space for data from three domains: speech data, real facial images, and synthetic facial images. With this, we transform speech into semantically-consistent synthetic facial images. Subsequently, we infer the 3D meshes of the synthetic images via the Face-to-Geometry Regressor. This section first introduces preliminaries of the necessary data of three domains, then describes the two modules in detail.", + "bbox": [ + 496, + 627, + 890, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Data Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 787, + 681, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike supervised methods that require paired data of speech and 3D facial animation, our framework only necessitates real facial videos synchronized with speech and a collection of rendered facial images.", + "bbox": [ + 496, + 810, + 890, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We record the videos in an environment without background noise to ensure a clean recording of sound. The", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21033", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "camera is positioned to capture the speaker's front view, enabling clear capture of lip movements. In terms of rendered images, we utilize a readily available character for rendering, which is not required to have the same physiognomy as the speaker and can be easily acquired from online resources or produced by artists. Then, varying facial expressions are generated by deforming the character using performances of range-of-motion (ROM) and exercising poses from Facial Action Coding System (FACS)[9]. Consequently, each rendered image has its corresponding 3D facial mesh. More details of the dataset construction are presented in Section 4.1", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Speech-to-Image Transcoder", + "text_level": 1, + "bbox": [ + 76, + 282, + 334, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal{I}_{\\mathcal{R}}^{1:T} = \\left(\\mathcal{I}_{\\mathcal{R}}^{1},\\dots ,\\mathcal{I}_{\\mathcal{R}}^{T}\\right)$ denotes a sequence of real facial frames. Correspondingly, the sequence of speech snippets $\\mathcal{A}^{1:T} = \\left(\\mathbf{a}^1,\\dots ,\\mathbf{a}^T\\right)$ is defined such that each $\\mathbf{a}^t$ is synchronized with the respective $\\mathcal{I}_{\\mathcal{R}}^t$ . $\\{\\mathcal{I}_S\\}$ is a collection of synthetic facial images, and each facial image has its corresponding 3D face mesh $\\nu_{S}$ . These three types of data are utilized by the Speech-to-Image Transcoder to construct a common representation space for speech and images.", + "bbox": [ + 75, + 303, + 468, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Network architecture. Our proposed Speech-to-Image Transcoder consists of two encoders and two decoders. We employ a speech encoder $\\mathcal{E}_{\\mathcal{A}}$ to transform speech into a compact representation. A shared visual encoder $\\mathcal{E}_{\\mathcal{V}}$ is used to encode real and synthetic faces. Additionally, two decoders, $\\mathcal{D}_{\\mathcal{R}}$ and $\\mathcal{D}_{\\mathcal{S}}$ , are used to reconstruct real and synthetic faces from the encodings. The overview of our Speech-to-Image Transcoder is illustrated in Figure 1.", + "bbox": [ + 75, + 426, + 468, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, the speech encoder $\\mathcal{E}_{\\mathcal{A}}$ consists of an audio feature extractor and a multi-layer transformer encoder[33]. The audio feature extractor is initialized with pre-trained wav2vec2.0[1] weights, responsible for converting raw waveform input into audio features. And the transformer encoder converts audio features into contextualized speech representations $\\mathcal{Z}_{\\mathcal{A}}^{1:T} = (\\mathbf{z}_{\\mathcal{A}}^{1},\\dots,\\mathbf{z}_{\\mathcal{A}}^{T})$ . Simultaneously, real faces aligned to the audio are input to the visual encoder $\\mathcal{E}_{\\mathcal{V}}$ to obtain visual features $\\mathcal{Z}_{\\mathcal{R}}^{1:T} = (\\mathbf{z}_{\\mathcal{R}}^{1},\\dots,\\mathbf{z}_{\\mathcal{R}}^{T})$ . This single visual encoder is shared for encoding synthetic facial features $\\mathcal{Z}_{\\mathcal{S}}^{\\{\\mathcal{B}\\}} = (\\mathbf{z}_{\\mathcal{S}}^{\\{\\mathcal{B}\\}}),\\{\\mathcal{B}\\} \\subseteq \\{\\mathcal{I}_{\\mathcal{S}}\\}$ . Then, $\\mathbf{z}_{\\mathcal{R}}$ and $\\mathbf{z}_{\\mathcal{S}}$ are respectively fed into decoder $\\mathcal{D}_{\\mathcal{R}}$ and $\\mathcal{D}_{\\mathcal{S}}$ for reconstructing the corresponding facial images. In addition, due to the paired relationship between real images and speech, $\\mathcal{D}_{\\mathcal{S}}$ can also be utilized to reconstruct synthetic images from $\\mathbf{z}_{\\mathcal{A}}$ . We borrow the decoder structure from [26], which consists of several upsampling blocks with an upsampling scale of 2. This architecture of a single shared visual encoder and multiple decoders for different domains has been verified to have good applicability in face swapping [27] and expression transfer [25]. In our experiments, $\\mathbf{z}_{\\mathcal{A}},\\mathbf{z}_{\\mathcal{R}}$ and $\\mathbf{z}_{\\mathcal{S}}$ are all $m$ -dim vectors, where $m$ is set as 512.", + "bbox": [ + 75, + 547, + 470, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Speech-to-Image Transcoder learns a semantically", + "bbox": [ + 96, + 885, + 467, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "consistent representation of speech, real faces, and synthetic faces. It enables the domain-dependent decoder $\\mathcal{D}_{\\mathcal{R}}$ or $\\mathcal{D}_S$ to reconstruct the facial image with corresponding semantics from any one of the three encodings.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Expression de-neutralization We observe that in cases where the two facial domains exhibit significant differences, such as one comprising real faces and the other consisting of cartoonized synthetic faces, the semantic distributions between the two domains may not be well aligned after encoding. This misalignment can result in semantically mismatched image-to-image translation. For instance, the decoder $\\mathcal{D}_S$ may erroneously decode the encoding of a real face exhibiting a grinning expression as a synthetic face with pouting expression.", + "bbox": [ + 496, + 152, + 890, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this issue, we propose a simple yet effective operation that we refer to as expression de-neutralization (denoted in Equation 1). Rather than reconstructing the facial expression image directly from the encoding $\\mathbf{z}_i, i \\in \\{\\mathcal{A}, \\mathcal{R}, \\mathcal{S}\\}$ we first subtract the neutral expression code $\\overline{z}_i$ corresponding to the domain (for speech domain, we use $\\overline{z}_{\\mathcal{R}}$ ) and then recover the image with the offset expression code $\\Delta z_i$ . We manually select a neutral expression image for each facial domain and obtain the neutral expression code after encoding. Through expression de-neutralization, two semantically misaligned domains can be shifted to become aligned, leading to semantically-consistent translation. Our experiments profoundly verify the effectiveness of this operation.", + "bbox": [ + 496, + 303, + 890, + 513 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta z _ {i} = z _ {i} - \\bar {z} _ {i}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 529, + 890, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training objectives. We employ two commonly used loss terms for image reconstruction, combined in Equation 2. The first term denotes the mean error of all pixels between the input image and the reconstructed one $(\\widetilde{\\mathcal{I}}_{\\mathcal{R}}$ or $\\widetilde{\\mathcal{I}}_S)$ . The second term $\\mathcal{L}_{percep}$ is the perceptual loss[17] to measure the similarity of two different images. It has obvious benefits for enhancing the fidelity of the reconstructed image.", + "bbox": [ + 496, + 551, + 890, + 657 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\left| \\left| \\mathcal {D} _ {i} \\left(\\Delta z _ {i}\\right) - \\mathcal {I} _ {i} \\right| \\right| _ {2} + \\mathcal {L} _ {p e r c e p}, i \\in \\{\\mathcal {R}, \\mathcal {S} \\} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 683, + 890, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, owing to the fact that the speech and real facial images are inherently paired, we are able to enhance temporal synchronization and consistency between speech and visual features by imposing our designed cross-modal constraint (Equation 3). This constraint is comprised of two terms, with $\\mathcal{L}_{cf}$ representing the cross-modal constraint on features, and $\\mathcal{L}_{cs}$ representing that on image spatial space.", + "bbox": [ + 496, + 710, + 890, + 816 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c r o s s} = \\mathcal {L} _ {c f} + \\mathcal {L} _ {c s},\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 842, + 743, + 859 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c f} = | | z _ {\\mathcal {A}} - z _ {\\mathcal {R}} | | _ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 862, + 740, + 878 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c s} = \\left| \\left| \\mathcal {D} _ {\\mathcal {R}} \\left(\\Delta z _ {\\mathcal {A}}\\right) - \\mathcal {I} _ {\\mathcal {R}} \\right| \\right| _ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 880, + 890, + 897 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21034", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c0c2718a626df8f862a86828bd56bf8222986c7ac95d833b0e99f854fd5edb26.jpg", + "image_caption": [ + "Figure 1. The architecture of the Speech-to-Image Transcoder. Firstly, the speech is converted into $\\mathbf{z}_{\\mathcal{A}}$ through $\\mathcal{E}_{\\mathcal{A}}$ . The red arrow shows how we input $\\mathbf{z}_{\\mathcal{A}}$ into $\\mathcal{D}_{\\mathcal{R}}$ to obtain the reconstructed image for calculating $\\mathcal{L}_{cs}$ . The green arrow shows how we obtain $z_{S}^{*}$ for calculating $\\mathcal{L}_{ccl}$ . We feed the Real/Synthetic face images into the $\\mathcal{E}_{\\mathcal{V}}$ to obtain the visual representation $\\mathbf{z}_{\\mathcal{R}}$ and $\\mathbf{z}_{\\mathcal{S}}$ , which are respectively input into $\\mathcal{D}_{\\mathcal{R}}$ and $\\mathcal{D}_{\\mathcal{S}}$ to obtain reconstructed images for calculating the $\\mathcal{L}_{rec}$ ." + ], + "image_footnote": [], + "bbox": [ + 130, + 89, + 843, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further encourage the transcoder to use a common representation across domains of speech and synthetic faces, we present a latent cycle consistency loss (Equation 4). This loss creates a loop constraint between the latent codes $z_{\\mathcal{A}}$ derived from speech input and the latent codes $z_{\\mathcal{S}}^{*}$ derived from the decoded synthetic images.", + "bbox": [ + 75, + 492, + 470, + 583 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c c l} = | | z _ {\\mathcal {S}} ^ {*} - z _ {\\mathcal {A}} | | _ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 607, + 339, + 623 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {\\mathcal {S}} ^ {*} = \\mathcal {E} _ {\\mathcal {V}} (\\mathcal {D} _ {\\mathcal {S}} (\\Delta z _ {\\mathcal {A}})). \\qquad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 625, + 468, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To sum up, our total loss is depicted in Equation 5.", + "bbox": [ + 96, + 651, + 433, + 667 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {\\text {c r o s s}} + \\mathcal {L} _ {\\text {c c l}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 676, + 468, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By utilizing the proposed cross-modal joint training, encodings from three distinct domains are semantically aligned. This allows for any of the three input modalities to be transformed into semantically-consistent facial images through domain-specific decoders. Specifically, we convert speech into synthetic faces. By further leveraging the Faceto-Geometry Regressor in Section 3.3, 3D facial meshes can be reconstructed from the transformed images.", + "bbox": [ + 75, + 702, + 468, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Face-to-Geometry Regressor", + "text_level": 1, + "bbox": [ + 76, + 832, + 334, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The architecture of the Face-to-geometry Regressor $\\mathcal{E}_{\\mathcal{G}}$ is shown in Figure 2. It is trained in a supervised manner. We exclusively rely on information from the synthetic", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/655c7cc0bbeaab507c9c194996db7a3a84f18493a7aeb4b58be78ba909f89270.jpg", + "image_caption": [ + "Figure 2. The architecture of the Face-to-Geometry Regressor. The synthetic face images are fed into the Regressor $\\mathcal{E}_{\\mathcal{G}}$ , which learns to regress the 3D mesh associated with the synthetic face images using RMSE loss." + ], + "image_footnote": [], + "bbox": [ + 527, + 493, + 867, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "image domain, specifically synthetic facial images and the corresponding 3D face meshes. We utilize Resnet34[14] as the backbone and a single fully connected layer to regress the coordinates of the 3D mesh $\\widetilde{\\mathcal{V}}_S$ . A simple RMSE loss is used to optimize the network, denoted as follows:", + "bbox": [ + 496, + 667, + 890, + 743 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g} = \\left| \\left| \\mathcal {E} _ {\\mathcal {G}} \\left(\\mathcal {I} _ {S}\\right) - \\mathcal {Y} _ {S} \\right| \\right| _ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 765, + 890, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the training phase, we employ data augmentation of color jittering and random affine transformation to enhance the robustness of our model. As a consequence, we are able to achieve satisfactory 3D face reconstruction results, even when the synthetic faces generated by the Speech-to-Image Transcoder are outside the training set.", + "bbox": [ + 496, + 810, + 892, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21035", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9c44563976a6ab42325d48d7464639cf45071f8597cd0929642b7103eca9eedf.jpg", + "image_caption": [ + "Figure 3. Inference pipeline. The raw waveform is fed to the speech encoder $\\mathcal{E}_{\\mathcal{A}}$ . The output is subtracted by $\\overline{z}_{\\mathcal{R}}$ and then fed into $\\mathcal{D}_{\\mathcal{S}}$ to recover the synthetic facial images, which are then input into $\\mathcal{E}_{\\mathcal{G}}$ to generate the 3D face animation." + ], + "image_footnote": [], + "bbox": [ + 89, + 92, + 459, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Inference Phase", + "text_level": 1, + "bbox": [ + 76, + 340, + 235, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Figure 3, the Speech-to-Image Transcoder and Face-to-Geometry Regressor are integrated to infer 3D face animation from speech. The raw waveform is fed into the speech encoder $\\mathcal{E}_A$ to generate the speech representation. Then we subtract the speech representation by $\\overline{z}_{\\mathcal{R}}$ , and the offset code is fed into $\\mathcal{D}_S$ to recover the synthetic facial images. Due to the cross-modal constraint $\\mathcal{L}_{cross}$ , the transformed synthetic facial images are endowed with lip movements reflecting the content of the input speech. Finally, the face-to-geometry model is utilized to generate the 3D face animation from the transformed synthetic facial images.", + "bbox": [ + 75, + 366, + 470, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Discussion", + "text_level": 1, + "bbox": [ + 76, + 566, + 194, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One might raise doubts about the rationale behind not utilizing the image-to-image modules to accomplish the transformation from real faces to synthetic faces, subsequently employing the Face-to-Geometry Regressor to acquire pseudo-labels, and ultimately training a speech-topseudo-label network in a supervised manner. (Hereafter, we refer to this pipeline as the three-stage method.) We notice that while this approach can alleviate the challenge of limited supervision data, it still suffers from the problem of ambiguous mapping between speech and animation, the same as other regression-based supervised methods.", + "bbox": [ + 75, + 595, + 468, + 761 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our cross-modal joint training scheme can identify latent patterns and correlations between speech and images. This allows us to translate subtle variations in speech into detailed lip motions, creating highly-fidelity 3D reconstructions. Therefore, our approach has a distinct advantage in mitigating the uncertainties of direct speech-to-animation mapping. We conduct detailed experiments to demonstrate the advantages of our method compared with the three-stage method in Section 4.", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 89, + 632, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 114, + 601, + 130 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our experiments are conducted on two datasets: the self-built dataset SelfCollected and the public dataset VOCASET[7]. In addition, to support the semi-supervised cross-modal training, we generate two synthetic datasets. The details are described below.", + "bbox": [ + 498, + 140, + 890, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SelfCollected dataset. This dataset contains 148 speaking video clips of one speaker. Each video contains one sentence. We employ the 3D facial capture system's workflow to obtain the ground truth (GT). Specifically, a head-mounted facial capture system is used to record the videos so that detailed lip motions of the speaker can be accurately tracked. Then, the facial tracking module captures several key facial landmarks. For each video, the artists manually pair a real person's face with the CG character at several keyframes. Finally, the facial retargeting module is trained on these keyframes and then automatically directs the motions of other video frames onto the CG character. Note that the character has not the same physiognomy as the speaker, so the range of lip movements is different after face retargeting. Through the workflows, we obtain the corresponding 3D face mesh with the topology of the CG character for each video frame. In total, this dataset contains 45000 frames, each frame corresponding a face mesh with 12696 vertices. We take 128 videos for training and the remaining 20 videos for testing.", + "bbox": [ + 496, + 215, + 892, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VOCASET dataset. VOCASET contains 12 speakers, each with 40 utterances. Facial meshes are scanned at 60FPS and then registered into a unified topology using the FLAME model[21], with each mesh containing 5023 vertices. In our experiments, we only use the data of one speaker for training and testing. We use 36 utterances as the training set and the remaining 4 as the test set.", + "bbox": [ + 496, + 520, + 890, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SyntheticSelf dataset. This dataset is built to assist the semi-supervised training on SelfCollected. We use the same CG character in SelfCollected for rendering. During rendering, we keep the camera facing the character, and the rendering lighting remains unchanged. Animations of performances of range-of-motion (ROM) and exercising poses from the Facial Action Coding System (FACS) are utilized for rendering. Specifically, we require artists manually craft the blendshape weights of ROM and FACS by linearly combining the blendshapes of the CG character, enabling us to achieve a wide range of expressions. Finally, SyntheticSelf comprises 20000 rendered facial images, each corresponding to a facial mesh with 12696 vertices.", + "bbox": [ + 496, + 627, + 892, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SyntheticVOCA dataset. Similar to SyntheticSelf, we create SyntheticVOCA to assist the semi-supervised training on VOCASET. We use FLAME model to fit expression coefficients for each rendered image in SyntheticSelf. Then expressions are transferred to the character (i.e. the", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21036", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "speaker's scan) provided by VOCASET. In total, we obtain 20000 rendered facial images, each corresponding to a FLAME-topology facial mesh with 5023 vertices.", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Compared methods", + "text_level": 1, + "bbox": [ + 76, + 146, + 264, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with several SOTA methods: VOCA[7], MeshTalk[28] and FaceFormer[10], as well as the three-stage method discussed in Section 3.5. All methods are compared on SelfCollected and VOCASET. Notably, since MeshTalk cannot be trained on a small single-speaker dataset, we only perform qualitative comparisons with its officially released model. In addition, we find VOCA cannot converge on a single-speaker dataset, so we use all the data of the other 11 speakers combined with the single speaker's 36 sentences for training. The three-stage method employs an image-to-image and face-to-geometry pipeline to obtain pseudo-labels and then experiments with network architectures of VOCA and Faceformer, respectively. Since VOCASET does not have videos of real faces, leading to that we cannot obtain the corresponding pseudolabels, so we do not compare with the three-stage method on VOCASET. We use the symbol * later to represent the corresponding method using the three-stage scheme.", + "bbox": [ + 75, + 170, + 468, + 443 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To sum up, there are 7 comparisons on SelfCollected: ours, VOCA, FaceFormer, GT, VOCA*, FaceFormer*, and pseudo GT. And 4 comparisons on SelfCollected: ours, VOCA, FaceFormer, and GT.", + "bbox": [ + 75, + 443, + 468, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 513, + 299, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We employ the commonly used lip vertex error[28] to evaluate the quantitative performance, which represents the maximal $l_{2}$ error of vertices in the lip region. In addition, we propose the lip landmark error to measure the lip movement difference between the generated 3D animation and real facial sequence. To obtain this error, we first estimate a similarity transformation based on the projected 2D coordinates of the predicted 3D lip vertices and the 2D lip landmarks of the real face, then align them using the estimated transformation, and finally compute the root mean squared error.", + "bbox": [ + 75, + 537, + 468, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results are illustrated in Table 1 and Table 2. Note that GT is obtained by retargeting the lip motions from the real speaker to the CG character, so it reflects the idiosyncrasies of the character instead of the speaker. Therefore, it is not objective to compare us with the supervised methods in terms of the metric of lip vertex error, and similarly, there is no comparability between supervised methods and ours in terms of the metric of lip landmark error. Nonetheless, our semi-supervised method achieves comparable performance to supervised methods in terms of lip vertex error, and even outperforms VOCA on SelfCollected. As for the lip landmark error, our method significantly outperforms the supervised approach and consistently outperforms the three-stage", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3ff901ff3513ec5e76690b06f9f868d81bf9e12d54ada2f88d5b50a4b7f6b3b1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsLip vertex error (in mm)↓Lip landmark error (in pixel)↓
VOCA3.7185.908
FaceFormer2.6905.845
GT05.041
VOCA*5.8383.790
FaceFormer*4.5363.613
Pseudo GT3.7732.941
Ours3.7083.241
", + "bbox": [ + 501, + 90, + 901, + 243 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/82f25b207707c313a3110a40d43527db056a931ff47b3f0b8acfb40f460a0bc6.jpg", + "table_caption": [ + "Table 1. Quantitative results on SelfCollected." + ], + "table_footnote": [], + "table_body": "
MethodsLip vertex error (in mm)↓
VOCA4.690
FaceFormer3.166
Ours5.102
", + "bbox": [ + 578, + 273, + 815, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative results on VOCASET.", + "bbox": [ + 549, + 369, + 841, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "methods, even approaching the pseudo GT, which is the upper bound of the three-stage methods. This indicates that our joint training approach indeed promotes generating accurate and higher fidelity lip animations.", + "bbox": [ + 496, + 398, + 890, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 500, + 468, + 710, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 4 presents qualitative results obtained by comparing three sentences. Sentences A and B come from SelfCollected-Test, and sentence C comes from VOCASET-Test. They belong to two different languages, namely Chinese and English. We evaluate sentence A with the model trained on SelfCollected and evaluate sentences B, C using the model trained on VOCASET. It should be noted that sentence B does not have its FLAME-topology GT meshes, and sentence C does not have the corresponding real faces.", + "bbox": [ + 496, + 491, + 890, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "From the results of sentence A, we can see that our method achieves competitive results compared to the supervised approaches, and even gains better performance in some cases (highlighted by the red box) in terms of mouth closure and lip movement amplitude. From the results of sentence C, our method achieves results that are highly synchronized with the speech even without using speech data from VOCASET for training. The results of sentence A also demonstrate that compared to the three-stage methods, our method presents results that are more consistent with the lip movements of the real face. This indicates our joint training scheme helps to generate more realistic animations.", + "bbox": [ + 496, + 628, + 890, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Perceptual Evaluation", + "text_level": 1, + "bbox": [ + 500, + 816, + 705, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct user studies to evaluate the animation quality of our method, which comprise a total of 10 participants.", + "bbox": [ + 496, + 840, + 890, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "User Study on SelfCollected. For SelfCollected, we obtain the results of 7 comparisons using 10 random test sen", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21037", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8464b4cf1dc13883dc7fe70c9f5cad641a36c7cc27b62374b0071c1c06937ba2.jpg", + "image_caption": [ + "Figure 4. Qualitative results. We utilize the officially released model to produce MeshTalk results as it is not trainable on SelfCollected and VOCASET datasets. The highlighted section by red box indicates that our method has better performance in terms of mouth closure and lip movement amplitude." + ], + "image_footnote": [], + "bbox": [ + 81, + 98, + 480, + 840 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a9e2d19c5e7011fba0ff14cc3cd9dd122d169abb117b043c0e6a03af0e4a3223.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 98, + 880, + 455 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2893f2164fef244d4ed9c18c988e2258bc386eb75f177610676fade5d78a2900.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 490, + 455, + 880, + 834 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21038", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/12b5c37dabfa8c3f804900c211dd595d7de74d040f44696f4671572f8c5cba77.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Ours vs. CompetitorRealism↑Lip Sync↑Sim↑
Ours vs. VOCA51.053.063.0
Ours vs. FaceFormer45.049.057.0
Ours vs. GT43.045.058.0
Ours vs. VOCA*75.064.060.0
Ours vs. FaceFormer*58.061.054.0
Ours vs. pesudo GT52.051.048.0
", + "bbox": [ + 78, + 89, + 475, + 205 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2c1d81c95fd6b5d9a623a2be9d9b6ed4af90911aaeb7020a18a0f991cd48c772.jpg", + "table_caption": [ + "Table 3. User study results on SelfCollected. We use A/B testing and report the percentage (\\%) of answers where A is preferred over B." + ], + "table_footnote": [], + "table_body": "
Ours vs. CompetitorRealism↑
VOCASETSelf.Avg.
Ours vs. GT17.5——
Ours vs. FaceFormer40.057.548.8
Ours vs. VOCA45.095.070.0
Ours vs. CompetitorLip Sync↑
VOCASETSelf.Avg.
Ours vs. GT5.0——
Ours vs. FaceFormer15.057.536.3
Ours vs. VOCA60.080.070.0
", + "bbox": [ + 91, + 289, + 455, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tences of the test set. Therefore, 60 A vs. B pairs(10 videos x 6 comparisons) are created. For each pair, we ask the participants to choose their favorite one in terms of realism and lip sync, following the user study protocol of Faceformer. Additionally, we propose a new metric \"similarity\" to characterize the preference for which method generates more similar lip movements to that of the real faces. For this metric, we provide both A/B pair and the real facial frames for each sentence. And then we ask the participants to choose the one that has closer lip amplitude to the real face. Table 3 shows the results. We observe that our method is comparable to the SOTA supervised methods in terms of realism and lip sync. And our method significantly outperforms the supervised approach in terms of similarity. Furthermore, the results show that our method consistently performs better than the three-stage methods. It indicates that our joint training of speech and image is indeed effective.", + "bbox": [ + 75, + 488, + 468, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Study on VOCASET. We conduct another user study on VOCASET. Since our model has never seen the speech data in VOCASET, for fairness, we additionally randomly pick 4 sentences for comparison from SelfCollected-Test. In total, 20 A vs. B pairs (there is no FLAMEtopology GT in SelfCollected) are created. For each pair, participants are asked to choose their favorite one based on realism and lip sync. As shown in Table 4, we separately report the results of SelfCollected-Test and VOCASET-Test. The results show that our method consistently performs bet", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b5fc6177763c725446627a12668d53c4eeea2d59e04ccaaf11378241dfe253a3.jpg", + "table_caption": [ + "Table 4. User study results $(\\%)$ on VOCASET." + ], + "table_footnote": [], + "table_body": "
Expression de-neutralizationLcclLip lmk error (in pixel)↓
✓—3.407
—✓3.482
✓✓3.241
", + "bbox": [ + 547, + 88, + 846, + 180 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation study on SelfCollected.", + "bbox": [ + 557, + 184, + 833, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e529259c93538fe5922ec47cb4a4cfd475123d2a22b9d84e73949197d926fad5.jpg", + "image_caption": [ + "Figure 5. Ablation study on expression de-neutralization." + ], + "image_footnote": [], + "bbox": [ + 540, + 212, + 870, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ter in terms of realism. However, our model achieves worse lip sync than FaceFormer on VOCASET-Test. We claim that it is reasonable because our model is not trained with the English speech data of VOCASET. Similarly, the Faceformer's result on SelfCollected-Test is worse than ours.", + "bbox": [ + 498, + 391, + 890, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 477, + 651, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Expression De-neutralization. We investigate the effect of expression de-neutralization by removing it from our method. We conduct the comparison on Self-Collected. As depicted in Table 5, the operation of expression de-neutralization helps to reduce lip landmark error by $0.241$ ( $3.482 \\rightarrow 3.241$ ). We also visualize the decoded synthetic faces in Figure 5. As depicted in Figure 5, the incorporation of expression de-neutralization results in more consistent lip motions of synthetic faces with that of real faces. Without this operation, the decoded faces demonstrate obvious semantic mismatches, particularly when the disparity between the two image domains is substantial (i.e. real faces and rendered FLAME-topology faces). This phenomenon proves that expression de-neutralization helps align semantic distributions of different domains, resulting in more semantically consistent lip movement generation.", + "bbox": [ + 496, + 501, + 890, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Cycle Consistency Loss $\\mathcal{L}_{ccl}$ . We present the quantitative results in Table 5. As shown, $\\mathcal{L}_{ccl}$ helps to reduce lip landmark error by 0.166 (3.407 → 3.241). The qualitative comparison is depicted in Figure 6, it is indicated this loss function promotes the generation of lip movements that are more consistent with the speech.", + "bbox": [ + 496, + 743, + 890, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 844, + 617, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper aims to recover 3D animation from speech, it proposes a novel semi-supervised framework to address the", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21039", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/56ad41857693de04c44c9ee4acdbf184dd514d1e00e9373404a66aab5ecc2104.jpg", + "image_caption": [ + "wo hai bu zhi dao ni ren shi fu lan ke", + "Figure 6. Ablation study on $\\mathcal{L}_{ccl}$ ." + ], + "image_footnote": [], + "bbox": [ + 107, + 103, + 447, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "data scarcity problem of the supervised methods. In addition, a joint training scheme on cross-domain data, namely speech and images, is developed. Extensive experiments show the proposed method achieves satisfactory generation of lip motions. However, the major limitation of our method is its inability for multiple identities. To address this issue, we suggest two possible ways for future work. The first is expanding the decoder for each speaker, which succeeds in faceswap but increases model size with more IDs. The second is inspired by FOMM [29], which uses a universal motion module for facial motion extraction and transfer. Overall, our proposed method shows promising results and opens up opportunities for further research in the field of speech-driven 3D facial animation.", + "bbox": [ + 75, + 318, + 472, + 529 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 76, + 545, + 174, + 561 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020.", + "[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999.", + "[3] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European conference on computer vision (ECCV), pages 520-535, 2018.", + "[4] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7832-7841, 2019.", + "[5] Renwang Chen, Xuanhong Chen, Bingbing Ni, and Yanhao Ge. Simswap: An efficient framework for high fidelity face swapping. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2003–2011, 2020.", + "[6] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan" + ], + "bbox": [ + 84, + 571, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022.", + "[7] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J. Black. Capture, learning, and synthesis of 3d speaking styles. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10093-10103, 2019.", + "[8] Ricard Durall Lopez, Jireh Jam, Dominik Strassel, Moi Hoon Yap, and Janis Keuper. Facialgan: Style transfer and attribute manipulation on synthetic faces. In [32nd British Machine Vision Conference], pages 1-14, 2021.", + "[9] Paul Ekman and Wallace V Friesen. Facial action coding system. Environmental Psychology & Nonverbal Behavior, 1978.", + "[10] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18749-18758, 2021.", + "[11] Ohad Fried, Ayush Tewari, Michael Zollhöfer, Adam Finkelstein, Eli Shechtman, Dan B Goldman, Kyle Genova, Zeyu Jin, Christian Theobalt, and Maneesh Agrawala. Text-based editing of talking-head video. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019.", + "[12] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021.", + "[13] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021.", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.", + "[15] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3397-3406, 2022.", + "[16] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017.", + "[17] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016.", + "[18] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21040", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018.", + "[20] Siyuan Li, Semih Gunel, Mirela Ostrek, Pavan Ramdya, Pascal Fua, and Helge Rhodin. Deformation-aware unpaired image translation for pose estimation on laboratory animals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13158-13168, 2020.", + "[21] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017.", + "[22] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. Advances in neural information processing systems, 30, 2017.", + "[23] Wenshuang Liu, Wenting Chen, Zhanjia Yang, and Linlin Shen. Translate the facial regions you like using self-adaptive region translation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2180-2188, 2021.", + "[24] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural audio-driven video portrait generation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 106-125. Springer, 2022.", + "[25] Lucio Moser, Chinyu Chien, Mark Williams, Jose Serra, Darren Hendler, and Doug Roble. Semi-supervised video-driven facial animation transfer for production. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021.", + "[26] Jacek Naruniec, Leonhard Helminger, Christopher Schroers, and Romann M Weber. High-resolution neural face swapping for visual effects. In Computer Graphics Forum, volume 39, pages 173-184. Wiley Online Library, 2020.", + "[27] Ivan Perov, Daiheng Gao, Nikolay Chervoniy, Kunlin Liu, Sugasa Marangonda, Chris Ume, Mr Dpfks, Carl Shift Facenheim, Luis RP, Jian Jiang, et al. Deepfacelab: Integrated, flexible and extensible face-swapping framework. arXiv preprint arXiv:2005.05535, 2020.", + "[28] Alexander Richard, Michael Zollhoefer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1153-1162, 2021.", + "[29] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019.", + "[30] Linsen Song, Wayne Wu, Chen Qian, Ran He, and Chen Change Loy. Everybody's talkin': Let me talk as you want. IEEE Transactions on Information Forensics and Security, 17:585-598, 2022.", + "[31] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 716-731. Springer, 2020.", + "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.", + "[34] Chih-Chun Yang, Wan-Cyuan Fan, Cheng-Fu Yang, and Yu-Chiang Frank Wang. Cross-modal mutual learning for audiovisual speech recognition and manipulation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 3036-3044, 2022.", + "[35] Shunyu Yao, RuiZhe Zhong, Yichao Yan, Guangtao Zhai, and Xiaokang Yang. Dfa-nerf: personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791, 2022.", + "[36] Zili Yi, Hao Zhang, Ping Tan, and Minglun Gong. Dualgan: Unsupervised dual learning for image-to-image translation. In Proceedings of the IEEE international conference on computer vision, pages 2849-2857, 2017.", + "[37] Lingyun Yu, Jun Yu, and Qiang Ling. Mining audio, text and visual information for talking face generation. In 2019 IEEE International Conference on Data Mining (ICDM), pages 787-795. IEEE, 2019.", + "[38] Sibo Zhang, Jiahong Yuan, Miao Liao, and Liangjun Zhang. Text2video: Text-driven talking-head video synthesis with personalized phoneme-pose dictionary. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2659-2663. IEEE, 2022.", + "[39] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021.", + "[40] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017." + ], + "bbox": [ + 501, + 90, + 893, + 696 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21041", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding/ebc67249-8be2-4ca5-975d-54a569184984_model.json b/2023/Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding/ebc67249-8be2-4ca5-975d-54a569184984_model.json new file mode 100644 index 0000000000000000000000000000000000000000..55c23061aaea4a677381c4efc918ae69a3b7e916 --- /dev/null +++ b/2023/Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding/ebc67249-8be2-4ca5-975d-54a569184984_model.json @@ -0,0 +1,1947 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.085, + 0.131, + 0.885, + 0.153 + ], + "angle": 0, + "content": "Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.181, + 0.734, + 0.216 + ], + "angle": 0, + "content": "Peiji Yang* Huawei Wei* Yicheng Zhong* Zhisheng Wang \nTencent, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.22, + 0.733, + 0.234 + ], + "angle": 0, + "content": "{peijiyang, huaweiwei, ajaxhong, plorywang}@tencent.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.473, + 0.681 + ], + "angle": 0, + "content": "Existing Speech-driven 3D facial animation methods typically follow the supervised paradigm, involving regression from speech to 3D facial animation. This paradigm faces two major challenges: the high cost of supervision acquisition, and the ambiguity in mapping between speech and lip movements. To address these challenges, this study proposes a novel cross-modal semi-supervised framework, comprising a Speech-to-Image Transcoder and a Face-to-Geometry Regressor. The former jointly learns a common representation space from speech and image domains, enabling the transformation of speech into semantically-consistent facial images. The latter is responsible for reconstructing 3D facial meshes from the transformed images. Both modules require minimal effort to acquire the necessary training data, thereby obviating the dependence on costly supervised data. Furthermore, the joint learning scheme enables the fusion of intricate visual features into speech encoding, thereby facilitating the transformation of subtle speech variations into nuanced lip movements, ultimately enhancing the fidelity of 3D face reconstructions. Consequently, the ambiguity of the direct mapping of speech-to-animation is significantly reduced, leading to coherent and high-fidelity generation of lip motion. Extensive experiments demonstrate that our approach produces competitive results compared to supervised methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.711, + 0.21, + 0.727 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.737, + 0.47, + 0.828 + ], + "angle": 0, + "content": "Speech-driven 3D facial animation aims to automatically animate vivid facial expressions of realistic 3D avatars from the speech signal. It has become increasingly popular in many fields, such as games, virtual reality, film production, and online communication, as it enables the generation of lifelike facial expressions with minimal effort." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.829, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Recently proposed approaches, including VOCA[7], MeshTalk[28], and FaceFormer[10], have shown promising results in recovering 3D facial meshes from speech" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.27, + 0.895, + 0.436 + ], + "angle": 0, + "content": "signals through the utilization of regressive networks or transformer-based autoregressive models. However, these methods, which fall under the supervised paradigm, encounter two primary challenges. Firstly, obtaining paired supervision of speech and high-fidelity 3D facial animation requires the utilization of a professional and expensive facial motion capture system, leading to considerable costs. Secondly, the mapping of low-dimensional speech signals to high-dimensional 3D facial meshes with significant variability may result in ambiguity issues, leading to suboptimal outcomes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.442, + 0.895, + 0.881 + ], + "angle": 0, + "content": "This study for the first time proposes a novel cross-modal semi-supervised framework to address the above issues, which consists of a Speech-to-Image Transcoder and a Face-to-Geometry Regressor. The former is designed to transform speech into semantically-consistent facial images, while the latter is responsible for reconstructing 3D facial meshes from the transformed images. The design of the Speech-to-Image Transcoder is inspired by the recent success of unsupervised image-to-image translation [26], which allows for transformation between different input image domains. Our transcoder extends this mechanism to facilitate cross-modal conversion between speech and image domains. To achieve this capability, we learn a common representation space using data from three domains: paired speech and real facial images, as well as synthetic facial images. We train the translation between the real and synthetic image domains in an unsupervised manner, and simultaneously leverage the paired relationship between real images and speech to project speech features into the image representation space. Through this joint training, we construct a common space where the representations of speech and images are tied together. With the learned common representation, we can convert speech into synthetic facial images using our domain-dependent image decoder. Furthermore, we leverage the synthetic facial data, which is paired with rendered facial images and 3D face meshes, to train the Face-to-Geometry Regressor. This allows us to infer the corresponding 3D face meshes from the synthetic faces generated from speech." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Our pipeline is capable of effortlessly acquiring the nec" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.352, + 0.901 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21032" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "essary data, which includes synchronized speech and real facial images extracted from readily available video clips, as well as synthetic facial images generated through the deformation of a polygonal face mesh using a rendering engine. Notably, our approach is not bound by any prerequisites for paired speech and 3D face animation, thereby effectively addressing the issue of supervision data scarcity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.471, + 0.394 + ], + "angle": 0, + "content": "Moreover, our joint training scheme integrates the speech-to-image regression and image-to-image reconstruction into a cohesive framework, which enables the incorporation of detailed visual features into speech encoding. Leveraging the domain-dependent image decoders, which store intricate facial visual priors, we are able to convert nuanced speech variations into micro-expressions that are rich in detail, thereby facilitating the subsequent generation of high-fidelity 3D face reconstructions. Consequently, the uncertainties associated with the direct mapping of speech-to-animation in the supervised paradigm are considerably reduced, leading to the generation of lip motion with finer granularity." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.394, + 0.465, + 0.409 + ], + "angle": 0, + "content": "In summary, the main contributions of our research are:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.42, + 0.468, + 0.48 + ], + "angle": 0, + "content": "- We make the first attempt to build a semi-supervised framework for speech-driven 3D facial animation, which eliminates the need for paired speech and 3D animation." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.49, + 0.469, + 0.552 + ], + "angle": 0, + "content": "- Our proposed joint training scheme leverages cross-modal translation to convert subtle speech variations into images that are rich in detail, resulting in more high-fidelity lip motion generation." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.561, + 0.469, + 0.606 + ], + "angle": 0, + "content": "- Our extensive experiments and user studies demonstrate that our approach produces competitive results compared to supervised methods." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.42, + 0.469, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.619, + 0.22, + 0.635 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.469, + 0.719 + ], + "angle": 0, + "content": "The proposed method is related to several research fields, such as unsupervised image-to-image translation, speech-driven 2D talking head, and speech-driven 3D face animation. In the following, we review the most relevant approaches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Unsupervised image-to-image translation. The unsupervised image-to-image translation aims to convert images from one domain to another domain with certain characteristics of the source images preserved[22, 36]. The translation involves converting from silhouettes to real images [16], from segmentation masks to RGB images[8, 23], from face to face [5, 27], and so on. This technology has wide applications in many fields, such as style transfer [40], face swapping [5, 26, 27], expression transfer [25], pose estimation [20] etc. They typically employ GAN-based networks [40] or encoder-decoder architectures [16] to achieve the goals. Notably, they only focus on image domains," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "our method extends this mechanism to achieve cross-modal translation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.121, + 0.892, + 0.333 + ], + "angle": 0, + "content": "Speech-driven 2D talking head. 2D-based talking head generation methods generally utilize different image representations to assist the conversion of speech to 2D faces, such as facial landmarks[4, 31, 37], depth maps[15] and semantic maps[6, 24]. The discriminator loss and other task-specific losses are usually employed for optimization. For example, Chen et al.[3] design an audio-visual derivative correlation loss to ensure cross-modal consistency. In addition, several works use 3D information such as 3D Morphable Model (3DMM)[2, 19, 30, 32, 39] parameters or dynamic neural radiance fields (NeRF)[12, 13, 35] to guide the generation process. Unlike these 2D-based methods, this paper focuses on high-fidelity 3D animation generation from speech." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.59 + ], + "angle": 0, + "content": "Speech-driven 3D face animation. Recently, considerable endeavors have been undertaken in this field[7, 10, 11, 18, 28, 34, 38]. VOCA[7] casts the mapping from speech to animation as a regression issue. To train the regressor, it uses the paired data of speech and 3D face animation. FaceFormer[10] employs transformers to model the long-term dependencies of speech, leading to improved performance. Meshtalk[28] puts emphasis on addressing the scalability and realism of the model by decoupling audio-correlated and audio-uncorrelated information. All the above methods belong to the supervised paradigm, which requires costly paired data and suffers from the problem of ambiguous mapping when directly regressing 3D animation from speech. This paper develops a semi-supervised framework to convert speech into images and then reconstruct 3D faces from the transformed images, effectively alleviating the aforementioned two issues." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.603, + 0.637, + 0.62 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.779 + ], + "angle": 0, + "content": "The proposed semi-supervised method is divided into two stages. We first employ the Speech-to-Image Transcoder to learn a common representation space for data from three domains: speech data, real facial images, and synthetic facial images. With this, we transform speech into semantically-consistent synthetic facial images. Subsequently, we infer the 3D meshes of the synthetic images via the Face-to-Geometry Regressor. This section first introduces preliminaries of the necessary data of three domains, then describes the two modules in detail." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.788, + 0.682, + 0.803 + ], + "angle": 0, + "content": "3.1. Data Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.871 + ], + "angle": 0, + "content": "Unlike supervised methods that require paired data of speech and 3D facial animation, our framework only necessitates real facial videos synchronized with speech and a collection of rendered facial images." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We record the videos in an environment without background noise to ensure a clean recording of sound. The" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21033" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "camera is positioned to capture the speaker's front view, enabling clear capture of lip movements. In terms of rendered images, we utilize a readily available character for rendering, which is not required to have the same physiognomy as the speaker and can be easily acquired from online resources or produced by artists. Then, varying facial expressions are generated by deforming the character using performances of range-of-motion (ROM) and exercising poses from Facial Action Coding System (FACS)[9]. Consequently, each rendered image has its corresponding 3D facial mesh. More details of the dataset construction are presented in Section 4.1" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.283, + 0.336, + 0.299 + ], + "angle": 0, + "content": "3.2. Speech-to-Image Transcoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.47, + 0.426 + ], + "angle": 0, + "content": "Let \\(\\mathcal{I}_{\\mathcal{R}}^{1:T} = \\left(\\mathcal{I}_{\\mathcal{R}}^{1},\\dots ,\\mathcal{I}_{\\mathcal{R}}^{T}\\right)\\) denotes a sequence of real facial frames. Correspondingly, the sequence of speech snippets \\(\\mathcal{A}^{1:T} = \\left(\\mathbf{a}^1,\\dots ,\\mathbf{a}^T\\right)\\) is defined such that each \\(\\mathbf{a}^t\\) is synchronized with the respective \\(\\mathcal{I}_{\\mathcal{R}}^t\\). \\(\\{\\mathcal{I}_S\\}\\) is a collection of synthetic facial images, and each facial image has its corresponding 3D face mesh \\(\\nu_{S}\\). These three types of data are utilized by the Speech-to-Image Transcoder to construct a common representation space for speech and images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.47, + 0.547 + ], + "angle": 0, + "content": "Network architecture. Our proposed Speech-to-Image Transcoder consists of two encoders and two decoders. We employ a speech encoder \\(\\mathcal{E}_{\\mathcal{A}}\\) to transform speech into a compact representation. A shared visual encoder \\(\\mathcal{E}_{\\mathcal{V}}\\) is used to encode real and synthetic faces. Additionally, two decoders, \\(\\mathcal{D}_{\\mathcal{R}}\\) and \\(\\mathcal{D}_{\\mathcal{S}}\\), are used to reconstruct real and synthetic faces from the encodings. The overview of our Speech-to-Image Transcoder is illustrated in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.471, + 0.884 + ], + "angle": 0, + "content": "Specifically, the speech encoder \\(\\mathcal{E}_{\\mathcal{A}}\\) consists of an audio feature extractor and a multi-layer transformer encoder[33]. The audio feature extractor is initialized with pre-trained wav2vec2.0[1] weights, responsible for converting raw waveform input into audio features. And the transformer encoder converts audio features into contextualized speech representations \\(\\mathcal{Z}_{\\mathcal{A}}^{1:T} = (\\mathbf{z}_{\\mathcal{A}}^{1},\\dots,\\mathbf{z}_{\\mathcal{A}}^{T})\\). Simultaneously, real faces aligned to the audio are input to the visual encoder \\(\\mathcal{E}_{\\mathcal{V}}\\) to obtain visual features \\(\\mathcal{Z}_{\\mathcal{R}}^{1:T} = (\\mathbf{z}_{\\mathcal{R}}^{1},\\dots,\\mathbf{z}_{\\mathcal{R}}^{T})\\). This single visual encoder is shared for encoding synthetic facial features \\(\\mathcal{Z}_{\\mathcal{S}}^{\\{\\mathcal{B}\\}} = (\\mathbf{z}_{\\mathcal{S}}^{\\{\\mathcal{B}\\}}),\\{\\mathcal{B}\\} \\subseteq \\{\\mathcal{I}_{\\mathcal{S}}\\}\\). Then, \\(\\mathbf{z}_{\\mathcal{R}}\\) and \\(\\mathbf{z}_{\\mathcal{S}}\\) are respectively fed into decoder \\(\\mathcal{D}_{\\mathcal{R}}\\) and \\(\\mathcal{D}_{\\mathcal{S}}\\) for reconstructing the corresponding facial images. In addition, due to the paired relationship between real images and speech, \\(\\mathcal{D}_{\\mathcal{S}}\\) can also be utilized to reconstruct synthetic images from \\(\\mathbf{z}_{\\mathcal{A}}\\). We borrow the decoder structure from [26], which consists of several upsampling blocks with an upsampling scale of 2. This architecture of a single shared visual encoder and multiple decoders for different domains has been verified to have good applicability in face swapping [27] and expression transfer [25]. In our experiments, \\(\\mathbf{z}_{\\mathcal{A}},\\mathbf{z}_{\\mathcal{R}}\\) and \\(\\mathbf{z}_{\\mathcal{S}}\\) are all \\(m\\)-dim vectors, where \\(m\\) is set as 512." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.902 + ], + "angle": 0, + "content": "The Speech-to-Image Transcoder learns a semantically" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "consistent representation of speech, real faces, and synthetic faces. It enables the domain-dependent decoder \\(\\mathcal{D}_{\\mathcal{R}}\\) or \\(\\mathcal{D}_S\\) to reconstruct the facial image with corresponding semantics from any one of the three encodings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.892, + 0.303 + ], + "angle": 0, + "content": "Expression de-neutralization We observe that in cases where the two facial domains exhibit significant differences, such as one comprising real faces and the other consisting of cartoonized synthetic faces, the semantic distributions between the two domains may not be well aligned after encoding. This misalignment can result in semantically mismatched image-to-image translation. For instance, the decoder \\(\\mathcal{D}_S\\) may erroneously decode the encoding of a real face exhibiting a grinning expression as a synthetic face with pouting expression." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.514 + ], + "angle": 0, + "content": "To address this issue, we propose a simple yet effective operation that we refer to as expression de-neutralization (denoted in Equation 1). Rather than reconstructing the facial expression image directly from the encoding \\(\\mathbf{z}_i, i \\in \\{\\mathcal{A}, \\mathcal{R}, \\mathcal{S}\\}\\) we first subtract the neutral expression code \\(\\overline{z}_i\\) corresponding to the domain (for speech domain, we use \\(\\overline{z}_{\\mathcal{R}}\\)) and then recover the image with the offset expression code \\(\\Delta z_i\\). We manually select a neutral expression image for each facial domain and obtain the neutral expression code after encoding. Through expression de-neutralization, two semantically misaligned domains can be shifted to become aligned, leading to semantically-consistent translation. Our experiments profoundly verify the effectiveness of this operation." + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.53, + 0.891, + 0.545 + ], + "angle": 0, + "content": "\\[\n\\Delta z _ {i} = z _ {i} - \\bar {z} _ {i}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Training objectives. We employ two commonly used loss terms for image reconstruction, combined in Equation 2. The first term denotes the mean error of all pixels between the input image and the reconstructed one \\((\\widetilde{\\mathcal{I}}_{\\mathcal{R}}\\) or \\(\\widetilde{\\mathcal{I}}_S)\\). The second term \\(\\mathcal{L}_{percep}\\) is the perceptual loss[17] to measure the similarity of two different images. It has obvious benefits for enhancing the fidelity of the reconstructed image." + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.684, + 0.891, + 0.701 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\left| \\left| \\mathcal {D} _ {i} \\left(\\Delta z _ {i}\\right) - \\mathcal {I} _ {i} \\right| \\right| _ {2} + \\mathcal {L} _ {p e r c e p}, i \\in \\{\\mathcal {R}, \\mathcal {S} \\} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.711, + 0.892, + 0.817 + ], + "angle": 0, + "content": "Moreover, owing to the fact that the speech and real facial images are inherently paired, we are able to enhance temporal synchronization and consistency between speech and visual features by imposing our designed cross-modal constraint (Equation 3). This constraint is comprised of two terms, with \\(\\mathcal{L}_{cf}\\) representing the cross-modal constraint on features, and \\(\\mathcal{L}_{cs}\\) representing that on image spatial space." + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.843, + 0.744, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c r o s s} = \\mathcal {L} _ {c f} + \\mathcal {L} _ {c s},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.863, + 0.741, + 0.879 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c f} = | | z _ {\\mathcal {A}} - z _ {\\mathcal {R}} | | _ {2},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.881, + 0.891, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c s} = \\left| \\left| \\mathcal {D} _ {\\mathcal {R}} \\left(\\Delta z _ {\\mathcal {A}}\\right) - \\mathcal {I} _ {\\mathcal {R}} \\right| \\right| _ {2}. \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21034" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.09, + 0.844, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.415, + 0.895, + 0.478 + ], + "angle": 0, + "content": "Figure 1. The architecture of the Speech-to-Image Transcoder. Firstly, the speech is converted into \\(\\mathbf{z}_{\\mathcal{A}}\\) through \\(\\mathcal{E}_{\\mathcal{A}}\\). The red arrow shows how we input \\(\\mathbf{z}_{\\mathcal{A}}\\) into \\(\\mathcal{D}_{\\mathcal{R}}\\) to obtain the reconstructed image for calculating \\(\\mathcal{L}_{cs}\\). The green arrow shows how we obtain \\(z_{S}^{*}\\) for calculating \\(\\mathcal{L}_{ccl}\\). We feed the Real/Synthetic face images into the \\(\\mathcal{E}_{\\mathcal{V}}\\) to obtain the visual representation \\(\\mathbf{z}_{\\mathcal{R}}\\) and \\(\\mathbf{z}_{\\mathcal{S}}\\), which are respectively input into \\(\\mathcal{D}_{\\mathcal{R}}\\) and \\(\\mathcal{D}_{\\mathcal{S}}\\) to obtain reconstructed images for calculating the \\(\\mathcal{L}_{rec}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.471, + 0.584 + ], + "angle": 0, + "content": "To further encourage the transcoder to use a common representation across domains of speech and synthetic faces, we present a latent cycle consistency loss (Equation 4). This loss creates a loop constraint between the latent codes \\( z_{\\mathcal{A}} \\) derived from speech input and the latent codes \\( z_{\\mathcal{S}}^{*} \\) derived from the decoded synthetic images." + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.608, + 0.341, + 0.624 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c c l} = | | z _ {\\mathcal {S}} ^ {*} - z _ {\\mathcal {A}} | | _ {2},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.626, + 0.47, + 0.645 + ], + "angle": 0, + "content": "\\[\nz _ {\\mathcal {S}} ^ {*} = \\mathcal {E} _ {\\mathcal {V}} (\\mathcal {D} _ {\\mathcal {S}} (\\Delta z _ {\\mathcal {A}})). \\qquad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.652, + 0.434, + 0.668 + ], + "angle": 0, + "content": "To sum up, our total loss is depicted in Equation 5." + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.678, + 0.47, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {\\text {c r o s s}} + \\mathcal {L} _ {\\text {c c l}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.825 + ], + "angle": 0, + "content": "By utilizing the proposed cross-modal joint training, encodings from three distinct domains are semantically aligned. This allows for any of the three input modalities to be transformed into semantically-consistent facial images through domain-specific decoders. Specifically, we convert speech into synthetic faces. By further leveraging the Faceto-Geometry Regressor in Section 3.3, 3D facial meshes can be reconstructed from the transformed images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.336, + 0.849 + ], + "angle": 0, + "content": "3.3. Face-to-Geometry Regressor" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The architecture of the Face-to-geometry Regressor \\(\\mathcal{E}_{\\mathcal{G}}\\) is shown in Figure 2. It is trained in a supervised manner. We exclusively rely on information from the synthetic" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.494, + 0.868, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.596, + 0.892, + 0.657 + ], + "angle": 0, + "content": "Figure 2. The architecture of the Face-to-Geometry Regressor. The synthetic face images are fed into the Regressor \\(\\mathcal{E}_{\\mathcal{G}}\\), which learns to regress the 3D mesh associated with the synthetic face images using RMSE loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.668, + 0.892, + 0.744 + ], + "angle": 0, + "content": "image domain, specifically synthetic facial images and the corresponding 3D face meshes. We utilize Resnet34[14] as the backbone and a single fully connected layer to regress the coordinates of the 3D mesh \\(\\widetilde{\\mathcal{V}}_S\\). A simple RMSE loss is used to optimize the network, denoted as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.766, + 0.892, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g} = \\left| \\left| \\mathcal {E} _ {\\mathcal {G}} \\left(\\mathcal {I} _ {S}\\right) - \\mathcal {Y} _ {S} \\right| \\right| _ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.893, + 0.903 + ], + "angle": 0, + "content": "During the training phase, we employ data augmentation of color jittering and random affine transformation to enhance the robustness of our model. As a consequence, we are able to achieve satisfactory 3D face reconstruction results, even when the synthetic faces generated by the Speech-to-Image Transcoder are outside the training set." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21035" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.093, + 0.46, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.269, + 0.47, + 0.33 + ], + "angle": 0, + "content": "Figure 3. Inference pipeline. The raw waveform is fed to the speech encoder \\(\\mathcal{E}_{\\mathcal{A}}\\). The output is subtracted by \\(\\overline{z}_{\\mathcal{R}}\\) and then fed into \\(\\mathcal{D}_{\\mathcal{S}}\\) to recover the synthetic facial images, which are then input into \\(\\mathcal{E}_{\\mathcal{G}}\\) to generate the 3D face animation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.341, + 0.236, + 0.355 + ], + "angle": 0, + "content": "3.4. Inference Phase" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.471, + 0.549 + ], + "angle": 0, + "content": "As illustrated in Figure 3, the Speech-to-Image Transcoder and Face-to-Geometry Regressor are integrated to infer 3D face animation from speech. The raw waveform is fed into the speech encoder \\(\\mathcal{E}_A\\) to generate the speech representation. Then we subtract the speech representation by \\(\\overline{z}_{\\mathcal{R}}\\), and the offset code is fed into \\(\\mathcal{D}_S\\) to recover the synthetic facial images. Due to the cross-modal constraint \\(\\mathcal{L}_{cross}\\), the transformed synthetic facial images are endowed with lip movements reflecting the content of the input speech. Finally, the face-to-geometry model is utilized to generate the 3D face animation from the transformed synthetic facial images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.568, + 0.195, + 0.583 + ], + "angle": 0, + "content": "3.5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.762 + ], + "angle": 0, + "content": "One might raise doubts about the rationale behind not utilizing the image-to-image modules to accomplish the transformation from real faces to synthetic faces, subsequently employing the Face-to-Geometry Regressor to acquire pseudo-labels, and ultimately training a speech-topseudo-label network in a supervised manner. (Hereafter, we refer to this pipeline as the three-stage method.) We notice that while this approach can alleviate the challenge of limited supervision data, it still suffers from the problem of ambiguous mapping between speech and animation, the same as other regression-based supervised methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Our cross-modal joint training scheme can identify latent patterns and correlations between speech and images. This allows us to translate subtle variations in speech into detailed lip motions, creating highly-fidelity 3D reconstructions. Therefore, our approach has a distinct advantage in mitigating the uncertainties of direct speech-to-animation mapping. We conduct detailed experiments to demonstrate the advantages of our method compared with the three-stage method in Section 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.633, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.116, + 0.602, + 0.131 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.141, + 0.892, + 0.215 + ], + "angle": 0, + "content": "Our experiments are conducted on two datasets: the self-built dataset SelfCollected and the public dataset VOCASET[7]. In addition, to support the semi-supervised cross-modal training, we generate two synthetic datasets. The details are described below." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.217, + 0.893, + 0.52 + ], + "angle": 0, + "content": "SelfCollected dataset. This dataset contains 148 speaking video clips of one speaker. Each video contains one sentence. We employ the 3D facial capture system's workflow to obtain the ground truth (GT). Specifically, a head-mounted facial capture system is used to record the videos so that detailed lip motions of the speaker can be accurately tracked. Then, the facial tracking module captures several key facial landmarks. For each video, the artists manually pair a real person's face with the CG character at several keyframes. Finally, the facial retargeting module is trained on these keyframes and then automatically directs the motions of other video frames onto the CG character. Note that the character has not the same physiognomy as the speaker, so the range of lip movements is different after face retargeting. Through the workflows, we obtain the corresponding 3D face mesh with the topology of the CG character for each video frame. In total, this dataset contains 45000 frames, each frame corresponding a face mesh with 12696 vertices. We take 128 videos for training and the remaining 20 videos for testing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.627 + ], + "angle": 0, + "content": "VOCASET dataset. VOCASET contains 12 speakers, each with 40 utterances. Facial meshes are scanned at 60FPS and then registered into a unified topology using the FLAME model[21], with each mesh containing 5023 vertices. In our experiments, we only use the data of one speaker for training and testing. We use 36 utterances as the training set and the remaining 4 as the test set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.893, + 0.824 + ], + "angle": 0, + "content": "SyntheticSelf dataset. This dataset is built to assist the semi-supervised training on SelfCollected. We use the same CG character in SelfCollected for rendering. During rendering, we keep the camera facing the character, and the rendering lighting remains unchanged. Animations of performances of range-of-motion (ROM) and exercising poses from the Facial Action Coding System (FACS) are utilized for rendering. Specifically, we require artists manually craft the blendshape weights of ROM and FACS by linearly combining the blendshapes of the CG character, enabling us to achieve a wide range of expressions. Finally, SyntheticSelf comprises 20000 rendered facial images, each corresponding to a facial mesh with 12696 vertices." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "SyntheticVOCA dataset. Similar to SyntheticSelf, we create SyntheticVOCA to assist the semi-supervised training on VOCASET. We use FLAME model to fit expression coefficients for each rendered image in SyntheticSelf. Then expressions are transferred to the character (i.e. the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21036" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.138 + ], + "angle": 0, + "content": "speaker's scan) provided by VOCASET. In total, we obtain 20000 rendered facial images, each corresponding to a FLAME-topology facial mesh with 5023 vertices." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.147, + 0.266, + 0.163 + ], + "angle": 0, + "content": "4.2. Compared methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.171, + 0.47, + 0.444 + ], + "angle": 0, + "content": "We compare our method with several SOTA methods: VOCA[7], MeshTalk[28] and FaceFormer[10], as well as the three-stage method discussed in Section 3.5. All methods are compared on SelfCollected and VOCASET. Notably, since MeshTalk cannot be trained on a small single-speaker dataset, we only perform qualitative comparisons with its officially released model. In addition, we find VOCA cannot converge on a single-speaker dataset, so we use all the data of the other 11 speakers combined with the single speaker's 36 sentences for training. The three-stage method employs an image-to-image and face-to-geometry pipeline to obtain pseudo-labels and then experiments with network architectures of VOCA and Faceformer, respectively. Since VOCASET does not have videos of real faces, leading to that we cannot obtain the corresponding pseudolabels, so we do not compare with the three-stage method on VOCASET. We use the symbol * later to represent the corresponding method using the three-stage scheme." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.469, + 0.504 + ], + "angle": 0, + "content": "To sum up, there are 7 comparisons on SelfCollected: ours, VOCA, FaceFormer, GT, VOCA*, FaceFormer*, and pseudo GT. And 4 comparisons on SelfCollected: ours, VOCA, FaceFormer, and GT." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.3, + 0.531 + ], + "angle": 0, + "content": "4.3.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.703 + ], + "angle": 0, + "content": "We employ the commonly used lip vertex error[28] to evaluate the quantitative performance, which represents the maximal \\( l_{2} \\) error of vertices in the lip region. In addition, we propose the lip landmark error to measure the lip movement difference between the generated 3D animation and real facial sequence. To obtain this error, we first estimate a similarity transformation based on the projected 2D coordinates of the predicted 3D lip vertices and the 2D lip landmarks of the real face, then align them using the estimated transformation, and finally compute the root mean squared error." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Results are illustrated in Table 1 and Table 2. Note that GT is obtained by retargeting the lip motions from the real speaker to the CG character, so it reflects the idiosyncrasies of the character instead of the speaker. Therefore, it is not objective to compare us with the supervised methods in terms of the metric of lip vertex error, and similarly, there is no comparability between supervised methods and ours in terms of the metric of lip landmark error. Nonetheless, our semi-supervised method achieves comparable performance to supervised methods in terms of lip vertex error, and even outperforms VOCA on SelfCollected. As for the lip landmark error, our method significantly outperforms the supervised approach and consistently outperforms the three-stage" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.091, + 0.903, + 0.244 + ], + "angle": 0, + "content": "
MethodsLip vertex error (in mm)↓Lip landmark error (in pixel)↓
VOCA3.7185.908
FaceFormer2.6905.845
GT05.041
VOCA*5.8383.790
FaceFormer*4.5363.613
Pseudo GT3.7732.941
Ours3.7083.241
" + }, + { + "type": "table_caption", + "bbox": [ + 0.543, + 0.248, + 0.848, + 0.263 + ], + "angle": 0, + "content": "Table 1. Quantitative results on SelfCollected." + }, + { + "type": "table", + "bbox": [ + 0.58, + 0.274, + 0.816, + 0.367 + ], + "angle": 0, + "content": "
MethodsLip vertex error (in mm)↓
VOCA4.690
FaceFormer3.166
Ours5.102
" + }, + { + "type": "table_caption", + "bbox": [ + 0.55, + 0.371, + 0.842, + 0.386 + ], + "angle": 0, + "content": "Table 2. Quantitative results on VOCASET." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.892, + 0.462 + ], + "angle": 0, + "content": "methods, even approaching the pseudo GT, which is the upper bound of the three-stage methods. This indicates that our joint training approach indeed promotes generating accurate and higher fidelity lip animations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.469, + 0.712, + 0.485 + ], + "angle": 0, + "content": "4.4. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.628 + ], + "angle": 0, + "content": "Figure 4 presents qualitative results obtained by comparing three sentences. Sentences A and B come from SelfCollected-Test, and sentence C comes from VOCASET-Test. They belong to two different languages, namely Chinese and English. We evaluate sentence A with the model trained on SelfCollected and evaluate sentences B, C using the model trained on VOCASET. It should be noted that sentence B does not have its FLAME-topology GT meshes, and sentence C does not have the corresponding real faces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.809 + ], + "angle": 0, + "content": "From the results of sentence A, we can see that our method achieves competitive results compared to the supervised approaches, and even gains better performance in some cases (highlighted by the red box) in terms of mouth closure and lip movement amplitude. From the results of sentence C, our method achieves results that are highly synchronized with the speech even without using speech data from VOCASET for training. The results of sentence A also demonstrate that compared to the three-stage methods, our method presents results that are more consistent with the lip movements of the real face. This indicates our joint training scheme helps to generate more realistic animations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.707, + 0.833 + ], + "angle": 0, + "content": "4.5. Perceptual Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.891, + 0.871 + ], + "angle": 0, + "content": "We conduct user studies to evaluate the animation quality of our method, which comprise a total of 10 participants." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "User Study on SelfCollected. For SelfCollected, we obtain the results of 7 comparisons using 10 random test sen" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21037" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.099, + 0.482, + 0.841 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.099, + 0.882, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.457, + 0.882, + 0.835 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.851, + 0.895, + 0.898 + ], + "angle": 0, + "content": "Figure 4. Qualitative results. We utilize the officially released model to produce MeshTalk results as it is not trainable on SelfCollected and VOCASET datasets. The highlighted section by red box indicates that our method has better performance in terms of mouth closure and lip movement amplitude." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21038" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.09, + 0.477, + 0.206 + ], + "angle": 0, + "content": "
Ours vs. CompetitorRealism↑Lip Sync↑Sim↑
Ours vs. VOCA51.053.063.0
Ours vs. FaceFormer45.049.057.0
Ours vs. GT43.045.058.0
Ours vs. VOCA*75.064.060.0
Ours vs. FaceFormer*58.061.054.0
Ours vs. pesudo GT52.051.048.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.211, + 0.471, + 0.256 + ], + "angle": 0, + "content": "Table 3. User study results on SelfCollected. We use A/B testing and report the percentage (\\%) of answers where A is preferred over B." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.29, + 0.456, + 0.457 + ], + "angle": 0, + "content": "
Ours vs. CompetitorRealism↑
VOCASETSelf.Avg.
Ours vs. GT17.5——
Ours vs. FaceFormer40.057.548.8
Ours vs. VOCA45.095.070.0
Ours vs. CompetitorLip Sync↑
VOCASETSelf.Avg.
Ours vs. GT5.0——
Ours vs. FaceFormer15.057.536.3
Ours vs. VOCA60.080.070.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.118, + 0.46, + 0.428, + 0.475 + ], + "angle": 0, + "content": "Table 4. User study results \\((\\%)\\) on VOCASET." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.47, + 0.746 + ], + "angle": 0, + "content": "tences of the test set. Therefore, 60 A vs. B pairs(10 videos x 6 comparisons) are created. For each pair, we ask the participants to choose their favorite one in terms of realism and lip sync, following the user study protocol of Faceformer. Additionally, we propose a new metric \"similarity\" to characterize the preference for which method generates more similar lip movements to that of the real faces. For this metric, we provide both A/B pair and the real facial frames for each sentence. And then we ask the participants to choose the one that has closer lip amplitude to the real face. Table 3 shows the results. We observe that our method is comparable to the SOTA supervised methods in terms of realism and lip sync. And our method significantly outperforms the supervised approach in terms of similarity. Furthermore, the results show that our method consistently performs better than the three-stage methods. It indicates that our joint training of speech and image is indeed effective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "User Study on VOCASET. We conduct another user study on VOCASET. Since our model has never seen the speech data in VOCASET, for fairness, we additionally randomly pick 4 sentences for comparison from SelfCollected-Test. In total, 20 A vs. B pairs (there is no FLAMEtopology GT in SelfCollected) are created. For each pair, participants are asked to choose their favorite one based on realism and lip sync. As shown in Table 4, we separately report the results of SelfCollected-Test and VOCASET-Test. The results show that our method consistently performs bet" + }, + { + "type": "table", + "bbox": [ + 0.548, + 0.089, + 0.848, + 0.181 + ], + "angle": 0, + "content": "
Expression de-neutralizationLcclLip lmk error (in pixel)↓
✓—3.407
—✓3.482
✓✓3.241
" + }, + { + "type": "table_caption", + "bbox": [ + 0.558, + 0.185, + 0.834, + 0.201 + ], + "angle": 0, + "content": "Table 5. Ablation study on SelfCollected." + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.213, + 0.872, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.373, + 0.885, + 0.388 + ], + "angle": 0, + "content": "Figure 5. Ablation study on expression de-neutralization." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.392, + 0.892, + 0.468 + ], + "angle": 0, + "content": "ter in terms of realism. However, our model achieves worse lip sync than FaceFormer on VOCASET-Test. We claim that it is reasonable because our model is not trained with the English speech data of VOCASET. Similarly, the Faceformer's result on SelfCollected-Test is worse than ours." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.478, + 0.653, + 0.495 + ], + "angle": 0, + "content": "4.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.892, + 0.744 + ], + "angle": 0, + "content": "Effect of Expression De-neutralization. We investigate the effect of expression de-neutralization by removing it from our method. We conduct the comparison on Self-Collected. As depicted in Table 5, the operation of expression de-neutralization helps to reduce lip landmark error by \\(0.241\\) (\\(3.482 \\rightarrow 3.241\\)). We also visualize the decoded synthetic faces in Figure 5. As depicted in Figure 5, the incorporation of expression de-neutralization results in more consistent lip motions of synthetic faces with that of real faces. Without this operation, the decoded faces demonstrate obvious semantic mismatches, particularly when the disparity between the two image domains is substantial (i.e. real faces and rendered FLAME-topology faces). This phenomenon proves that expression de-neutralization helps align semantic distributions of different domains, resulting in more semantically consistent lip movement generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.744, + 0.892, + 0.835 + ], + "angle": 0, + "content": "Effect of Cycle Consistency Loss \\(\\mathcal{L}_{ccl}\\). We present the quantitative results in Table 5. As shown, \\(\\mathcal{L}_{ccl}\\) helps to reduce lip landmark error by 0.166 (3.407 → 3.241). The qualitative comparison is depicted in Figure 6, it is indicated this loss function promotes the generation of lip movements that are more consistent with the speech." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.619, + 0.861 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This paper aims to recover 3D animation from speech, it proposes a novel semi-supervised framework to address the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21039" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.198, + 0.092, + 0.393, + 0.103 + ], + "angle": 0, + "content": "wo hai bu zhi dao ni ren shi fu lan ke" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.104, + 0.449, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.292, + 0.384, + 0.307 + ], + "angle": 0, + "content": "Figure 6. Ablation study on \\(\\mathcal{L}_{ccl}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.473, + 0.53 + ], + "angle": 0, + "content": "data scarcity problem of the supervised methods. In addition, a joint training scheme on cross-domain data, namely speech and images, is developed. Extensive experiments show the proposed method achieves satisfactory generation of lip motions. However, the major limitation of our method is its inability for multiple identities. To address this issue, we suggest two possible ways for future work. The first is expanding the decoder for each speaker, which succeeds in faceswap but increases model size with more IDs. The second is inspired by FOMM [29], which uses a universal motion module for facial motion extraction and transfer. Overall, our proposed method shows promising results and opens up opportunities for further research in the field of speech-driven 3D facial animation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.546, + 0.175, + 0.563 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.572, + 0.47, + 0.628 + ], + "angle": 0, + "content": "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.63, + 0.47, + 0.685 + ], + "angle": 0, + "content": "[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.687, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[3] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European conference on computer vision (ECCV), pages 520-535, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.745, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[4] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7832-7841, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.816, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[5] Renwang Chen, Xuanhong Chen, Bingbing Ni, and Yanhao Ge. Simswap: An efficient framework for high fidelity face swapping. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2003–2011, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[6] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan" + }, + { + "type": "list", + "bbox": [ + 0.086, + 0.572, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.136, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[7] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J. Black. Capture, learning, and synthesis of 3d speaking styles. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10093-10103, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.205, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[8] Ricard Durall Lopez, Jireh Jam, Dominik Strassel, Moi Hoon Yap, and Janis Keuper. Facialgan: Style transfer and attribute manipulation on synthetic faces. In [32nd British Machine Vision Conference], pages 1-14, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.261, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[9] Paul Ekman and Wallace V Friesen. Facial action coding system. Environmental Psychology & Nonverbal Behavior, 1978." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[10] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18749-18758, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[11] Ohad Fried, Ayush Tewari, Michael Zollhöfer, Adam Finkelstein, Eli Shechtman, Dan B Goldman, Kyle Genova, Zeyu Jin, Christian Theobalt, and Maneesh Agrawala. Text-based editing of talking-head video. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[12] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[13] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.58, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[15] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3397-3406, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[16] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[17] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[18] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21040" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[19] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[20] Siyuan Li, Semih Gunel, Mirela Ostrek, Pavan Ramdya, Pascal Fua, and Helge Rhodin. Deformation-aware unpaired image translation for pose estimation on laboratory animals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13158-13168, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.47, + 0.276 + ], + "angle": 0, + "content": "[21] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.47, + 0.32 + ], + "angle": 0, + "content": "[22] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[23] Wenshuang Liu, Wenting Chen, Zhanjia Yang, and Linlin Shen. Translate the facial regions you like using self-adaptive region translation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2180-2188, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.47, + 0.475 + ], + "angle": 0, + "content": "[24] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural audio-driven video portrait generation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 106-125. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.47, + 0.532 + ], + "angle": 0, + "content": "[25] Lucio Moser, Chinyu Chien, Mark Williams, Jose Serra, Darren Hendler, and Doug Roble. Semi-supervised video-driven facial animation transfer for production. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.47, + 0.589 + ], + "angle": 0, + "content": "[26] Jacek Naruniec, Leonhard Helminger, Christopher Schroers, and Romann M Weber. High-resolution neural face swapping for visual effects. In Computer Graphics Forum, volume 39, pages 173-184. Wiley Online Library, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.47, + 0.659 + ], + "angle": 0, + "content": "[27] Ivan Perov, Daiheng Gao, Nikolay Chervoniy, Kunlin Liu, Sugasa Marangonda, Chris Ume, Mr Dpfks, Carl Shift Facenheim, Luis RP, Jian Jiang, et al. Deepfacelab: Integrated, flexible and extensible face-swapping framework. arXiv preprint arXiv:2005.05535, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.47, + 0.731 + ], + "angle": 0, + "content": "[28] Alexander Richard, Michael Zollhoefer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1153-1162, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[29] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[30] Linsen Song, Wayne Wu, Chen Qian, Ran He, and Chen Change Loy. Everybody's talkin': Let me talk as you want. IEEE Transactions on Information Forensics and Security, 17:585-598, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[31] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.176 + ], + "angle": 0, + "content": "[32] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 716-731. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.893, + 0.233 + ], + "angle": 0, + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.893, + 0.303 + ], + "angle": 0, + "content": "[34] Chih-Chun Yang, Wan-Cyuan Fan, Cheng-Fu Yang, and Yu-Chiang Frank Wang. Cross-modal mutual learning for audiovisual speech recognition and manipulation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 3036-3044, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.304, + 0.893, + 0.359 + ], + "angle": 0, + "content": "[35] Shunyu Yao, RuiZhe Zhong, Yichao Yan, Guangtao Zhai, and Xiaokang Yang. Dfa-nerf: personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.36, + 0.893, + 0.416 + ], + "angle": 0, + "content": "[36] Zili Yi, Hao Zhang, Ping Tan, and Minglun Gong. Dualgan: Unsupervised dual learning for image-to-image translation. In Proceedings of the IEEE international conference on computer vision, pages 2849-2857, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.893, + 0.472 + ], + "angle": 0, + "content": "[37] Lingyun Yu, Jun Yu, and Qiang Ling. Mining audio, text and visual information for talking face generation. In 2019 IEEE International Conference on Data Mining (ICDM), pages 787-795. IEEE, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.473, + 0.893, + 0.556 + ], + "angle": 0, + "content": "[38] Sibo Zhang, Jiahong Yuan, Miao Liao, and Liangjun Zhang. Text2video: Text-driven talking-head video synthesis with personalized phoneme-pose dictionary. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2659-2663. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.558, + 0.894, + 0.627 + ], + "angle": 0, + "content": "[39] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.628, + 0.893, + 0.697 + ], + "angle": 0, + "content": "[40] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21041" + } + ] +] \ No newline at end of file diff --git a/2023/Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding/ebc67249-8be2-4ca5-975d-54a569184984_origin.pdf b/2023/Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding/ebc67249-8be2-4ca5-975d-54a569184984_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c73b0d86dd1b2f0540c4e363d93d1176e2338d93 --- /dev/null +++ b/2023/Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding/ebc67249-8be2-4ca5-975d-54a569184984_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7feaa3f45964434c75818e09c9d43378e61d7219baaa9e5a8d59a46c449b420 +size 2017982 diff --git a/2023/Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding/full.md b/2023/Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d6168597578ecc436645d892e5e7713c40df59b1 --- /dev/null +++ b/2023/Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding/full.md @@ -0,0 +1,286 @@ +# Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding + +Peiji Yang* Huawei Wei* Yicheng Zhong* Zhisheng Wang +Tencent, Shenzhen, China + +{peijiyang, huaweiwei, ajaxhong, plorywang}@tencent.com + +# Abstract + +Existing Speech-driven 3D facial animation methods typically follow the supervised paradigm, involving regression from speech to 3D facial animation. This paradigm faces two major challenges: the high cost of supervision acquisition, and the ambiguity in mapping between speech and lip movements. To address these challenges, this study proposes a novel cross-modal semi-supervised framework, comprising a Speech-to-Image Transcoder and a Face-to-Geometry Regressor. The former jointly learns a common representation space from speech and image domains, enabling the transformation of speech into semantically-consistent facial images. The latter is responsible for reconstructing 3D facial meshes from the transformed images. Both modules require minimal effort to acquire the necessary training data, thereby obviating the dependence on costly supervised data. Furthermore, the joint learning scheme enables the fusion of intricate visual features into speech encoding, thereby facilitating the transformation of subtle speech variations into nuanced lip movements, ultimately enhancing the fidelity of 3D face reconstructions. Consequently, the ambiguity of the direct mapping of speech-to-animation is significantly reduced, leading to coherent and high-fidelity generation of lip motion. Extensive experiments demonstrate that our approach produces competitive results compared to supervised methods. + +# 1. Introduction + +Speech-driven 3D facial animation aims to automatically animate vivid facial expressions of realistic 3D avatars from the speech signal. It has become increasingly popular in many fields, such as games, virtual reality, film production, and online communication, as it enables the generation of lifelike facial expressions with minimal effort. + +Recently proposed approaches, including VOCA[7], MeshTalk[28], and FaceFormer[10], have shown promising results in recovering 3D facial meshes from speech + +signals through the utilization of regressive networks or transformer-based autoregressive models. However, these methods, which fall under the supervised paradigm, encounter two primary challenges. Firstly, obtaining paired supervision of speech and high-fidelity 3D facial animation requires the utilization of a professional and expensive facial motion capture system, leading to considerable costs. Secondly, the mapping of low-dimensional speech signals to high-dimensional 3D facial meshes with significant variability may result in ambiguity issues, leading to suboptimal outcomes. + +This study for the first time proposes a novel cross-modal semi-supervised framework to address the above issues, which consists of a Speech-to-Image Transcoder and a Face-to-Geometry Regressor. The former is designed to transform speech into semantically-consistent facial images, while the latter is responsible for reconstructing 3D facial meshes from the transformed images. The design of the Speech-to-Image Transcoder is inspired by the recent success of unsupervised image-to-image translation [26], which allows for transformation between different input image domains. Our transcoder extends this mechanism to facilitate cross-modal conversion between speech and image domains. To achieve this capability, we learn a common representation space using data from three domains: paired speech and real facial images, as well as synthetic facial images. We train the translation between the real and synthetic image domains in an unsupervised manner, and simultaneously leverage the paired relationship between real images and speech to project speech features into the image representation space. Through this joint training, we construct a common space where the representations of speech and images are tied together. With the learned common representation, we can convert speech into synthetic facial images using our domain-dependent image decoder. Furthermore, we leverage the synthetic facial data, which is paired with rendered facial images and 3D face meshes, to train the Face-to-Geometry Regressor. This allows us to infer the corresponding 3D face meshes from the synthetic faces generated from speech. + +Our pipeline is capable of effortlessly acquiring the nec + +essary data, which includes synchronized speech and real facial images extracted from readily available video clips, as well as synthetic facial images generated through the deformation of a polygonal face mesh using a rendering engine. Notably, our approach is not bound by any prerequisites for paired speech and 3D face animation, thereby effectively addressing the issue of supervision data scarcity. + +Moreover, our joint training scheme integrates the speech-to-image regression and image-to-image reconstruction into a cohesive framework, which enables the incorporation of detailed visual features into speech encoding. Leveraging the domain-dependent image decoders, which store intricate facial visual priors, we are able to convert nuanced speech variations into micro-expressions that are rich in detail, thereby facilitating the subsequent generation of high-fidelity 3D face reconstructions. Consequently, the uncertainties associated with the direct mapping of speech-to-animation in the supervised paradigm are considerably reduced, leading to the generation of lip motion with finer granularity. + +In summary, the main contributions of our research are: + +- We make the first attempt to build a semi-supervised framework for speech-driven 3D facial animation, which eliminates the need for paired speech and 3D animation. +- Our proposed joint training scheme leverages cross-modal translation to convert subtle speech variations into images that are rich in detail, resulting in more high-fidelity lip motion generation. +- Our extensive experiments and user studies demonstrate that our approach produces competitive results compared to supervised methods. + +# 2. Related Work + +The proposed method is related to several research fields, such as unsupervised image-to-image translation, speech-driven 2D talking head, and speech-driven 3D face animation. In the following, we review the most relevant approaches. + +Unsupervised image-to-image translation. The unsupervised image-to-image translation aims to convert images from one domain to another domain with certain characteristics of the source images preserved[22, 36]. The translation involves converting from silhouettes to real images [16], from segmentation masks to RGB images[8, 23], from face to face [5, 27], and so on. This technology has wide applications in many fields, such as style transfer [40], face swapping [5, 26, 27], expression transfer [25], pose estimation [20] etc. They typically employ GAN-based networks [40] or encoder-decoder architectures [16] to achieve the goals. Notably, they only focus on image domains, + +our method extends this mechanism to achieve cross-modal translation. + +Speech-driven 2D talking head. 2D-based talking head generation methods generally utilize different image representations to assist the conversion of speech to 2D faces, such as facial landmarks[4, 31, 37], depth maps[15] and semantic maps[6, 24]. The discriminator loss and other task-specific losses are usually employed for optimization. For example, Chen et al.[3] design an audio-visual derivative correlation loss to ensure cross-modal consistency. In addition, several works use 3D information such as 3D Morphable Model (3DMM)[2, 19, 30, 32, 39] parameters or dynamic neural radiance fields (NeRF)[12, 13, 35] to guide the generation process. Unlike these 2D-based methods, this paper focuses on high-fidelity 3D animation generation from speech. + +Speech-driven 3D face animation. Recently, considerable endeavors have been undertaken in this field[7, 10, 11, 18, 28, 34, 38]. VOCA[7] casts the mapping from speech to animation as a regression issue. To train the regressor, it uses the paired data of speech and 3D face animation. FaceFormer[10] employs transformers to model the long-term dependencies of speech, leading to improved performance. Meshtalk[28] puts emphasis on addressing the scalability and realism of the model by decoupling audio-correlated and audio-uncorrelated information. All the above methods belong to the supervised paradigm, which requires costly paired data and suffers from the problem of ambiguous mapping when directly regressing 3D animation from speech. This paper develops a semi-supervised framework to convert speech into images and then reconstruct 3D faces from the transformed images, effectively alleviating the aforementioned two issues. + +# 3. Methodology + +The proposed semi-supervised method is divided into two stages. We first employ the Speech-to-Image Transcoder to learn a common representation space for data from three domains: speech data, real facial images, and synthetic facial images. With this, we transform speech into semantically-consistent synthetic facial images. Subsequently, we infer the 3D meshes of the synthetic images via the Face-to-Geometry Regressor. This section first introduces preliminaries of the necessary data of three domains, then describes the two modules in detail. + +# 3.1. Data Preliminaries + +Unlike supervised methods that require paired data of speech and 3D facial animation, our framework only necessitates real facial videos synchronized with speech and a collection of rendered facial images. + +We record the videos in an environment without background noise to ensure a clean recording of sound. The + +camera is positioned to capture the speaker's front view, enabling clear capture of lip movements. In terms of rendered images, we utilize a readily available character for rendering, which is not required to have the same physiognomy as the speaker and can be easily acquired from online resources or produced by artists. Then, varying facial expressions are generated by deforming the character using performances of range-of-motion (ROM) and exercising poses from Facial Action Coding System (FACS)[9]. Consequently, each rendered image has its corresponding 3D facial mesh. More details of the dataset construction are presented in Section 4.1 + +# 3.2. Speech-to-Image Transcoder + +Let $\mathcal{I}_{\mathcal{R}}^{1:T} = \left(\mathcal{I}_{\mathcal{R}}^{1},\dots ,\mathcal{I}_{\mathcal{R}}^{T}\right)$ denotes a sequence of real facial frames. Correspondingly, the sequence of speech snippets $\mathcal{A}^{1:T} = \left(\mathbf{a}^1,\dots ,\mathbf{a}^T\right)$ is defined such that each $\mathbf{a}^t$ is synchronized with the respective $\mathcal{I}_{\mathcal{R}}^t$ . $\{\mathcal{I}_S\}$ is a collection of synthetic facial images, and each facial image has its corresponding 3D face mesh $\nu_{S}$ . These three types of data are utilized by the Speech-to-Image Transcoder to construct a common representation space for speech and images. + +Network architecture. Our proposed Speech-to-Image Transcoder consists of two encoders and two decoders. We employ a speech encoder $\mathcal{E}_{\mathcal{A}}$ to transform speech into a compact representation. A shared visual encoder $\mathcal{E}_{\mathcal{V}}$ is used to encode real and synthetic faces. Additionally, two decoders, $\mathcal{D}_{\mathcal{R}}$ and $\mathcal{D}_{\mathcal{S}}$ , are used to reconstruct real and synthetic faces from the encodings. The overview of our Speech-to-Image Transcoder is illustrated in Figure 1. + +Specifically, the speech encoder $\mathcal{E}_{\mathcal{A}}$ consists of an audio feature extractor and a multi-layer transformer encoder[33]. The audio feature extractor is initialized with pre-trained wav2vec2.0[1] weights, responsible for converting raw waveform input into audio features. And the transformer encoder converts audio features into contextualized speech representations $\mathcal{Z}_{\mathcal{A}}^{1:T} = (\mathbf{z}_{\mathcal{A}}^{1},\dots,\mathbf{z}_{\mathcal{A}}^{T})$ . Simultaneously, real faces aligned to the audio are input to the visual encoder $\mathcal{E}_{\mathcal{V}}$ to obtain visual features $\mathcal{Z}_{\mathcal{R}}^{1:T} = (\mathbf{z}_{\mathcal{R}}^{1},\dots,\mathbf{z}_{\mathcal{R}}^{T})$ . This single visual encoder is shared for encoding synthetic facial features $\mathcal{Z}_{\mathcal{S}}^{\{\mathcal{B}\}} = (\mathbf{z}_{\mathcal{S}}^{\{\mathcal{B}\}}),\{\mathcal{B}\} \subseteq \{\mathcal{I}_{\mathcal{S}}\}$ . Then, $\mathbf{z}_{\mathcal{R}}$ and $\mathbf{z}_{\mathcal{S}}$ are respectively fed into decoder $\mathcal{D}_{\mathcal{R}}$ and $\mathcal{D}_{\mathcal{S}}$ for reconstructing the corresponding facial images. In addition, due to the paired relationship between real images and speech, $\mathcal{D}_{\mathcal{S}}$ can also be utilized to reconstruct synthetic images from $\mathbf{z}_{\mathcal{A}}$ . We borrow the decoder structure from [26], which consists of several upsampling blocks with an upsampling scale of 2. This architecture of a single shared visual encoder and multiple decoders for different domains has been verified to have good applicability in face swapping [27] and expression transfer [25]. In our experiments, $\mathbf{z}_{\mathcal{A}},\mathbf{z}_{\mathcal{R}}$ and $\mathbf{z}_{\mathcal{S}}$ are all $m$ -dim vectors, where $m$ is set as 512. + +The Speech-to-Image Transcoder learns a semantically + +consistent representation of speech, real faces, and synthetic faces. It enables the domain-dependent decoder $\mathcal{D}_{\mathcal{R}}$ or $\mathcal{D}_S$ to reconstruct the facial image with corresponding semantics from any one of the three encodings. + +Expression de-neutralization We observe that in cases where the two facial domains exhibit significant differences, such as one comprising real faces and the other consisting of cartoonized synthetic faces, the semantic distributions between the two domains may not be well aligned after encoding. This misalignment can result in semantically mismatched image-to-image translation. For instance, the decoder $\mathcal{D}_S$ may erroneously decode the encoding of a real face exhibiting a grinning expression as a synthetic face with pouting expression. + +To address this issue, we propose a simple yet effective operation that we refer to as expression de-neutralization (denoted in Equation 1). Rather than reconstructing the facial expression image directly from the encoding $\mathbf{z}_i, i \in \{\mathcal{A}, \mathcal{R}, \mathcal{S}\}$ we first subtract the neutral expression code $\overline{z}_i$ corresponding to the domain (for speech domain, we use $\overline{z}_{\mathcal{R}}$ ) and then recover the image with the offset expression code $\Delta z_i$ . We manually select a neutral expression image for each facial domain and obtain the neutral expression code after encoding. Through expression de-neutralization, two semantically misaligned domains can be shifted to become aligned, leading to semantically-consistent translation. Our experiments profoundly verify the effectiveness of this operation. + +$$ +\Delta z _ {i} = z _ {i} - \bar {z} _ {i}. \tag {1} +$$ + +Training objectives. We employ two commonly used loss terms for image reconstruction, combined in Equation 2. The first term denotes the mean error of all pixels between the input image and the reconstructed one $(\widetilde{\mathcal{I}}_{\mathcal{R}}$ or $\widetilde{\mathcal{I}}_S)$ . The second term $\mathcal{L}_{percep}$ is the perceptual loss[17] to measure the similarity of two different images. It has obvious benefits for enhancing the fidelity of the reconstructed image. + +$$ +\mathcal {L} _ {r e c} = \left| \left| \mathcal {D} _ {i} \left(\Delta z _ {i}\right) - \mathcal {I} _ {i} \right| \right| _ {2} + \mathcal {L} _ {p e r c e p}, i \in \{\mathcal {R}, \mathcal {S} \} \tag {2} +$$ + +Moreover, owing to the fact that the speech and real facial images are inherently paired, we are able to enhance temporal synchronization and consistency between speech and visual features by imposing our designed cross-modal constraint (Equation 3). This constraint is comprised of two terms, with $\mathcal{L}_{cf}$ representing the cross-modal constraint on features, and $\mathcal{L}_{cs}$ representing that on image spatial space. + +$$ +\mathcal {L} _ {c r o s s} = \mathcal {L} _ {c f} + \mathcal {L} _ {c s}, +$$ + +$$ +\mathcal {L} _ {c f} = | | z _ {\mathcal {A}} - z _ {\mathcal {R}} | | _ {2}, +$$ + +$$ +\mathcal {L} _ {c s} = \left| \left| \mathcal {D} _ {\mathcal {R}} \left(\Delta z _ {\mathcal {A}}\right) - \mathcal {I} _ {\mathcal {R}} \right| \right| _ {2}. \tag {3} +$$ + +![](images/c0c2718a626df8f862a86828bd56bf8222986c7ac95d833b0e99f854fd5edb26.jpg) +Figure 1. The architecture of the Speech-to-Image Transcoder. Firstly, the speech is converted into $\mathbf{z}_{\mathcal{A}}$ through $\mathcal{E}_{\mathcal{A}}$ . The red arrow shows how we input $\mathbf{z}_{\mathcal{A}}$ into $\mathcal{D}_{\mathcal{R}}$ to obtain the reconstructed image for calculating $\mathcal{L}_{cs}$ . The green arrow shows how we obtain $z_{S}^{*}$ for calculating $\mathcal{L}_{ccl}$ . We feed the Real/Synthetic face images into the $\mathcal{E}_{\mathcal{V}}$ to obtain the visual representation $\mathbf{z}_{\mathcal{R}}$ and $\mathbf{z}_{\mathcal{S}}$ , which are respectively input into $\mathcal{D}_{\mathcal{R}}$ and $\mathcal{D}_{\mathcal{S}}$ to obtain reconstructed images for calculating the $\mathcal{L}_{rec}$ . + +To further encourage the transcoder to use a common representation across domains of speech and synthetic faces, we present a latent cycle consistency loss (Equation 4). This loss creates a loop constraint between the latent codes $z_{\mathcal{A}}$ derived from speech input and the latent codes $z_{\mathcal{S}}^{*}$ derived from the decoded synthetic images. + +$$ +\mathcal {L} _ {c c l} = | | z _ {\mathcal {S}} ^ {*} - z _ {\mathcal {A}} | | _ {2}, +$$ + +$$ +z _ {\mathcal {S}} ^ {*} = \mathcal {E} _ {\mathcal {V}} (\mathcal {D} _ {\mathcal {S}} (\Delta z _ {\mathcal {A}})). \qquad (4) +$$ + +To sum up, our total loss is depicted in Equation 5. + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {r e c}} + \mathcal {L} _ {\text {c r o s s}} + \mathcal {L} _ {\text {c c l}}. \tag {5} +$$ + +By utilizing the proposed cross-modal joint training, encodings from three distinct domains are semantically aligned. This allows for any of the three input modalities to be transformed into semantically-consistent facial images through domain-specific decoders. Specifically, we convert speech into synthetic faces. By further leveraging the Faceto-Geometry Regressor in Section 3.3, 3D facial meshes can be reconstructed from the transformed images. + +# 3.3. Face-to-Geometry Regressor + +The architecture of the Face-to-geometry Regressor $\mathcal{E}_{\mathcal{G}}$ is shown in Figure 2. It is trained in a supervised manner. We exclusively rely on information from the synthetic + +![](images/655c7cc0bbeaab507c9c194996db7a3a84f18493a7aeb4b58be78ba909f89270.jpg) +Figure 2. The architecture of the Face-to-Geometry Regressor. The synthetic face images are fed into the Regressor $\mathcal{E}_{\mathcal{G}}$ , which learns to regress the 3D mesh associated with the synthetic face images using RMSE loss. + +image domain, specifically synthetic facial images and the corresponding 3D face meshes. We utilize Resnet34[14] as the backbone and a single fully connected layer to regress the coordinates of the 3D mesh $\widetilde{\mathcal{V}}_S$ . A simple RMSE loss is used to optimize the network, denoted as follows: + +$$ +\mathcal {L} _ {g} = \left| \left| \mathcal {E} _ {\mathcal {G}} \left(\mathcal {I} _ {S}\right) - \mathcal {Y} _ {S} \right| \right| _ {2}. \tag {6} +$$ + +During the training phase, we employ data augmentation of color jittering and random affine transformation to enhance the robustness of our model. As a consequence, we are able to achieve satisfactory 3D face reconstruction results, even when the synthetic faces generated by the Speech-to-Image Transcoder are outside the training set. + +![](images/9c44563976a6ab42325d48d7464639cf45071f8597cd0929642b7103eca9eedf.jpg) +Figure 3. Inference pipeline. The raw waveform is fed to the speech encoder $\mathcal{E}_{\mathcal{A}}$ . The output is subtracted by $\overline{z}_{\mathcal{R}}$ and then fed into $\mathcal{D}_{\mathcal{S}}$ to recover the synthetic facial images, which are then input into $\mathcal{E}_{\mathcal{G}}$ to generate the 3D face animation. + +# 3.4. Inference Phase + +As illustrated in Figure 3, the Speech-to-Image Transcoder and Face-to-Geometry Regressor are integrated to infer 3D face animation from speech. The raw waveform is fed into the speech encoder $\mathcal{E}_A$ to generate the speech representation. Then we subtract the speech representation by $\overline{z}_{\mathcal{R}}$ , and the offset code is fed into $\mathcal{D}_S$ to recover the synthetic facial images. Due to the cross-modal constraint $\mathcal{L}_{cross}$ , the transformed synthetic facial images are endowed with lip movements reflecting the content of the input speech. Finally, the face-to-geometry model is utilized to generate the 3D face animation from the transformed synthetic facial images. + +# 3.5. Discussion + +One might raise doubts about the rationale behind not utilizing the image-to-image modules to accomplish the transformation from real faces to synthetic faces, subsequently employing the Face-to-Geometry Regressor to acquire pseudo-labels, and ultimately training a speech-topseudo-label network in a supervised manner. (Hereafter, we refer to this pipeline as the three-stage method.) We notice that while this approach can alleviate the challenge of limited supervision data, it still suffers from the problem of ambiguous mapping between speech and animation, the same as other regression-based supervised methods. + +Our cross-modal joint training scheme can identify latent patterns and correlations between speech and images. This allows us to translate subtle variations in speech into detailed lip motions, creating highly-fidelity 3D reconstructions. Therefore, our approach has a distinct advantage in mitigating the uncertainties of direct speech-to-animation mapping. We conduct detailed experiments to demonstrate the advantages of our method compared with the three-stage method in Section 4. + +# 4. Experiments + +# 4.1. Datasets + +Our experiments are conducted on two datasets: the self-built dataset SelfCollected and the public dataset VOCASET[7]. In addition, to support the semi-supervised cross-modal training, we generate two synthetic datasets. The details are described below. + +SelfCollected dataset. This dataset contains 148 speaking video clips of one speaker. Each video contains one sentence. We employ the 3D facial capture system's workflow to obtain the ground truth (GT). Specifically, a head-mounted facial capture system is used to record the videos so that detailed lip motions of the speaker can be accurately tracked. Then, the facial tracking module captures several key facial landmarks. For each video, the artists manually pair a real person's face with the CG character at several keyframes. Finally, the facial retargeting module is trained on these keyframes and then automatically directs the motions of other video frames onto the CG character. Note that the character has not the same physiognomy as the speaker, so the range of lip movements is different after face retargeting. Through the workflows, we obtain the corresponding 3D face mesh with the topology of the CG character for each video frame. In total, this dataset contains 45000 frames, each frame corresponding a face mesh with 12696 vertices. We take 128 videos for training and the remaining 20 videos for testing. + +VOCASET dataset. VOCASET contains 12 speakers, each with 40 utterances. Facial meshes are scanned at 60FPS and then registered into a unified topology using the FLAME model[21], with each mesh containing 5023 vertices. In our experiments, we only use the data of one speaker for training and testing. We use 36 utterances as the training set and the remaining 4 as the test set. + +SyntheticSelf dataset. This dataset is built to assist the semi-supervised training on SelfCollected. We use the same CG character in SelfCollected for rendering. During rendering, we keep the camera facing the character, and the rendering lighting remains unchanged. Animations of performances of range-of-motion (ROM) and exercising poses from the Facial Action Coding System (FACS) are utilized for rendering. Specifically, we require artists manually craft the blendshape weights of ROM and FACS by linearly combining the blendshapes of the CG character, enabling us to achieve a wide range of expressions. Finally, SyntheticSelf comprises 20000 rendered facial images, each corresponding to a facial mesh with 12696 vertices. + +SyntheticVOCA dataset. Similar to SyntheticSelf, we create SyntheticVOCA to assist the semi-supervised training on VOCASET. We use FLAME model to fit expression coefficients for each rendered image in SyntheticSelf. Then expressions are transferred to the character (i.e. the + +speaker's scan) provided by VOCASET. In total, we obtain 20000 rendered facial images, each corresponding to a FLAME-topology facial mesh with 5023 vertices. + +# 4.2. Compared methods + +We compare our method with several SOTA methods: VOCA[7], MeshTalk[28] and FaceFormer[10], as well as the three-stage method discussed in Section 3.5. All methods are compared on SelfCollected and VOCASET. Notably, since MeshTalk cannot be trained on a small single-speaker dataset, we only perform qualitative comparisons with its officially released model. In addition, we find VOCA cannot converge on a single-speaker dataset, so we use all the data of the other 11 speakers combined with the single speaker's 36 sentences for training. The three-stage method employs an image-to-image and face-to-geometry pipeline to obtain pseudo-labels and then experiments with network architectures of VOCA and Faceformer, respectively. Since VOCASET does not have videos of real faces, leading to that we cannot obtain the corresponding pseudolabels, so we do not compare with the three-stage method on VOCASET. We use the symbol * later to represent the corresponding method using the three-stage scheme. + +To sum up, there are 7 comparisons on SelfCollected: ours, VOCA, FaceFormer, GT, VOCA*, FaceFormer*, and pseudo GT. And 4 comparisons on SelfCollected: ours, VOCA, FaceFormer, and GT. + +# 4.3.Quantitative Evaluation + +We employ the commonly used lip vertex error[28] to evaluate the quantitative performance, which represents the maximal $l_{2}$ error of vertices in the lip region. In addition, we propose the lip landmark error to measure the lip movement difference between the generated 3D animation and real facial sequence. To obtain this error, we first estimate a similarity transformation based on the projected 2D coordinates of the predicted 3D lip vertices and the 2D lip landmarks of the real face, then align them using the estimated transformation, and finally compute the root mean squared error. + +Results are illustrated in Table 1 and Table 2. Note that GT is obtained by retargeting the lip motions from the real speaker to the CG character, so it reflects the idiosyncrasies of the character instead of the speaker. Therefore, it is not objective to compare us with the supervised methods in terms of the metric of lip vertex error, and similarly, there is no comparability between supervised methods and ours in terms of the metric of lip landmark error. Nonetheless, our semi-supervised method achieves comparable performance to supervised methods in terms of lip vertex error, and even outperforms VOCA on SelfCollected. As for the lip landmark error, our method significantly outperforms the supervised approach and consistently outperforms the three-stage + +
MethodsLip vertex error (in mm)↓Lip landmark error (in pixel)↓
VOCA3.7185.908
FaceFormer2.6905.845
GT05.041
VOCA*5.8383.790
FaceFormer*4.5363.613
Pseudo GT3.7732.941
Ours3.7083.241
+ +Table 1. Quantitative results on SelfCollected. + +
MethodsLip vertex error (in mm)↓
VOCA4.690
FaceFormer3.166
Ours5.102
+ +Table 2. Quantitative results on VOCASET. + +methods, even approaching the pseudo GT, which is the upper bound of the three-stage methods. This indicates that our joint training approach indeed promotes generating accurate and higher fidelity lip animations. + +# 4.4. Qualitative Evaluation + +Figure 4 presents qualitative results obtained by comparing three sentences. Sentences A and B come from SelfCollected-Test, and sentence C comes from VOCASET-Test. They belong to two different languages, namely Chinese and English. We evaluate sentence A with the model trained on SelfCollected and evaluate sentences B, C using the model trained on VOCASET. It should be noted that sentence B does not have its FLAME-topology GT meshes, and sentence C does not have the corresponding real faces. + +From the results of sentence A, we can see that our method achieves competitive results compared to the supervised approaches, and even gains better performance in some cases (highlighted by the red box) in terms of mouth closure and lip movement amplitude. From the results of sentence C, our method achieves results that are highly synchronized with the speech even without using speech data from VOCASET for training. The results of sentence A also demonstrate that compared to the three-stage methods, our method presents results that are more consistent with the lip movements of the real face. This indicates our joint training scheme helps to generate more realistic animations. + +# 4.5. Perceptual Evaluation + +We conduct user studies to evaluate the animation quality of our method, which comprise a total of 10 participants. + +User Study on SelfCollected. For SelfCollected, we obtain the results of 7 comparisons using 10 random test sen + +![](images/8464b4cf1dc13883dc7fe70c9f5cad641a36c7cc27b62374b0071c1c06937ba2.jpg) +Figure 4. Qualitative results. We utilize the officially released model to produce MeshTalk results as it is not trainable on SelfCollected and VOCASET datasets. The highlighted section by red box indicates that our method has better performance in terms of mouth closure and lip movement amplitude. + +![](images/a9e2d19c5e7011fba0ff14cc3cd9dd122d169abb117b043c0e6a03af0e4a3223.jpg) + +![](images/2893f2164fef244d4ed9c18c988e2258bc386eb75f177610676fade5d78a2900.jpg) + +
Ours vs. CompetitorRealism↑Lip Sync↑Sim↑
Ours vs. VOCA51.053.063.0
Ours vs. FaceFormer45.049.057.0
Ours vs. GT43.045.058.0
Ours vs. VOCA*75.064.060.0
Ours vs. FaceFormer*58.061.054.0
Ours vs. pesudo GT52.051.048.0
+ +Table 3. User study results on SelfCollected. We use A/B testing and report the percentage (\%) of answers where A is preferred over B. + +
Ours vs. CompetitorRealism↑
VOCASETSelf.Avg.
Ours vs. GT17.5——
Ours vs. FaceFormer40.057.548.8
Ours vs. VOCA45.095.070.0
Ours vs. CompetitorLip Sync↑
VOCASETSelf.Avg.
Ours vs. GT5.0——
Ours vs. FaceFormer15.057.536.3
Ours vs. VOCA60.080.070.0
+ +tences of the test set. Therefore, 60 A vs. B pairs(10 videos x 6 comparisons) are created. For each pair, we ask the participants to choose their favorite one in terms of realism and lip sync, following the user study protocol of Faceformer. Additionally, we propose a new metric "similarity" to characterize the preference for which method generates more similar lip movements to that of the real faces. For this metric, we provide both A/B pair and the real facial frames for each sentence. And then we ask the participants to choose the one that has closer lip amplitude to the real face. Table 3 shows the results. We observe that our method is comparable to the SOTA supervised methods in terms of realism and lip sync. And our method significantly outperforms the supervised approach in terms of similarity. Furthermore, the results show that our method consistently performs better than the three-stage methods. It indicates that our joint training of speech and image is indeed effective. + +User Study on VOCASET. We conduct another user study on VOCASET. Since our model has never seen the speech data in VOCASET, for fairness, we additionally randomly pick 4 sentences for comparison from SelfCollected-Test. In total, 20 A vs. B pairs (there is no FLAMEtopology GT in SelfCollected) are created. For each pair, participants are asked to choose their favorite one based on realism and lip sync. As shown in Table 4, we separately report the results of SelfCollected-Test and VOCASET-Test. The results show that our method consistently performs bet + +Table 4. User study results $(\%)$ on VOCASET. + +
Expression de-neutralizationLcclLip lmk error (in pixel)↓
✓—3.407
—✓3.482
✓✓3.241
+ +Table 5. Ablation study on SelfCollected. + +![](images/e529259c93538fe5922ec47cb4a4cfd475123d2a22b9d84e73949197d926fad5.jpg) +Figure 5. Ablation study on expression de-neutralization. + +ter in terms of realism. However, our model achieves worse lip sync than FaceFormer on VOCASET-Test. We claim that it is reasonable because our model is not trained with the English speech data of VOCASET. Similarly, the Faceformer's result on SelfCollected-Test is worse than ours. + +# 4.6. Ablation Study + +Effect of Expression De-neutralization. We investigate the effect of expression de-neutralization by removing it from our method. We conduct the comparison on Self-Collected. As depicted in Table 5, the operation of expression de-neutralization helps to reduce lip landmark error by $0.241$ ( $3.482 \rightarrow 3.241$ ). We also visualize the decoded synthetic faces in Figure 5. As depicted in Figure 5, the incorporation of expression de-neutralization results in more consistent lip motions of synthetic faces with that of real faces. Without this operation, the decoded faces demonstrate obvious semantic mismatches, particularly when the disparity between the two image domains is substantial (i.e. real faces and rendered FLAME-topology faces). This phenomenon proves that expression de-neutralization helps align semantic distributions of different domains, resulting in more semantically consistent lip movement generation. + +Effect of Cycle Consistency Loss $\mathcal{L}_{ccl}$ . We present the quantitative results in Table 5. As shown, $\mathcal{L}_{ccl}$ helps to reduce lip landmark error by 0.166 (3.407 → 3.241). The qualitative comparison is depicted in Figure 6, it is indicated this loss function promotes the generation of lip movements that are more consistent with the speech. + +# 5. Conclusion + +This paper aims to recover 3D animation from speech, it proposes a novel semi-supervised framework to address the + +![](images/56ad41857693de04c44c9ee4acdbf184dd514d1e00e9373404a66aab5ecc2104.jpg) +wo hai bu zhi dao ni ren shi fu lan ke +Figure 6. Ablation study on $\mathcal{L}_{ccl}$ . + +data scarcity problem of the supervised methods. In addition, a joint training scheme on cross-domain data, namely speech and images, is developed. Extensive experiments show the proposed method achieves satisfactory generation of lip motions. However, the major limitation of our method is its inability for multiple identities. To address this issue, we suggest two possible ways for future work. The first is expanding the decoder for each speaker, which succeeds in faceswap but increases model size with more IDs. The second is inspired by FOMM [29], which uses a universal motion module for facial motion extraction and transfer. Overall, our proposed method shows promising results and opens up opportunities for further research in the field of speech-driven 3D facial animation. + +# References + +[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. +[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999. +[3] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European conference on computer vision (ECCV), pages 520-535, 2018. +[4] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7832-7841, 2019. +[5] Renwang Chen, Xuanhong Chen, Bingbing Ni, and Yanhao Ge. Simswap: An efficient framework for high fidelity face swapping. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2003–2011, 2020. +[6] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan + +Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. +[7] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J. Black. Capture, learning, and synthesis of 3d speaking styles. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10093-10103, 2019. +[8] Ricard Durall Lopez, Jireh Jam, Dominik Strassel, Moi Hoon Yap, and Janis Keuper. Facialgan: Style transfer and attribute manipulation on synthetic faces. In [32nd British Machine Vision Conference], pages 1-14, 2021. +[9] Paul Ekman and Wallace V Friesen. Facial action coding system. Environmental Psychology & Nonverbal Behavior, 1978. +[10] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18749-18758, 2021. +[11] Ohad Fried, Ayush Tewari, Michael Zollhöfer, Adam Finkelstein, Eli Shechtman, Dan B Goldman, Kyle Genova, Zeyu Jin, Christian Theobalt, and Maneesh Agrawala. Text-based editing of talking-head video. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. +[12] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. +[13] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021. +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[15] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3397-3406, 2022. +[16] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. +[17] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. +[18] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. + +[19] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. +[20] Siyuan Li, Semih Gunel, Mirela Ostrek, Pavan Ramdya, Pascal Fua, and Helge Rhodin. Deformation-aware unpaired image translation for pose estimation on laboratory animals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13158-13168, 2020. +[21] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. +[22] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. Advances in neural information processing systems, 30, 2017. +[23] Wenshuang Liu, Wenting Chen, Zhanjia Yang, and Linlin Shen. Translate the facial regions you like using self-adaptive region translation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2180-2188, 2021. +[24] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural audio-driven video portrait generation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 106-125. Springer, 2022. +[25] Lucio Moser, Chinyu Chien, Mark Williams, Jose Serra, Darren Hendler, and Doug Roble. Semi-supervised video-driven facial animation transfer for production. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. +[26] Jacek Naruniec, Leonhard Helminger, Christopher Schroers, and Romann M Weber. High-resolution neural face swapping for visual effects. In Computer Graphics Forum, volume 39, pages 173-184. Wiley Online Library, 2020. +[27] Ivan Perov, Daiheng Gao, Nikolay Chervoniy, Kunlin Liu, Sugasa Marangonda, Chris Ume, Mr Dpfks, Carl Shift Facenheim, Luis RP, Jian Jiang, et al. Deepfacelab: Integrated, flexible and extensible face-swapping framework. arXiv preprint arXiv:2005.05535, 2020. +[28] Alexander Richard, Michael Zollhoefer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1153-1162, 2021. +[29] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. +[30] Linsen Song, Wayne Wu, Chen Qian, Ran He, and Chen Change Loy. Everybody's talkin': Let me talk as you want. IEEE Transactions on Information Forensics and Security, 17:585-598, 2022. +[31] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. + +[32] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 716-731. Springer, 2020. +[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. +[34] Chih-Chun Yang, Wan-Cyuan Fan, Cheng-Fu Yang, and Yu-Chiang Frank Wang. Cross-modal mutual learning for audiovisual speech recognition and manipulation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 3036-3044, 2022. +[35] Shunyu Yao, RuiZhe Zhong, Yichao Yan, Guangtao Zhai, and Xiaokang Yang. Dfa-nerf: personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791, 2022. +[36] Zili Yi, Hao Zhang, Ping Tan, and Minglun Gong. Dualgan: Unsupervised dual learning for image-to-image translation. In Proceedings of the IEEE international conference on computer vision, pages 2849-2857, 2017. +[37] Lingyun Yu, Jun Yu, and Qiang Ling. Mining audio, text and visual information for talking face generation. In 2019 IEEE International Conference on Data Mining (ICDM), pages 787-795. IEEE, 2019. +[38] Sibo Zhang, Jiahong Yuan, Miao Liao, and Liangjun Zhang. Text2video: Text-driven talking-head video synthesis with personalized phoneme-pose dictionary. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2659-2663. IEEE, 2022. +[39] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. +[40] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. \ No newline at end of file diff --git a/2023/Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding/images.zip b/2023/Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e1078e6219f8d99371fac616d26ccbfaa2679488 --- /dev/null +++ b/2023/Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c9f3dd74aebdf60d73010a5af6642c1e141a9ceff99b23399baded69f78f85d6 +size 569887 diff --git a/2023/Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding/layout.json b/2023/Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..35935a1b12e6253a68411940f9b1e58260759a67 --- /dev/null +++ b/2023/Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding/layout.json @@ -0,0 +1,7366 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "type": "text", + "content": "Semi-supervised Speech-driven 3D Facial Animation via Cross-modal Encoding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 143, + 449, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 143, + 449, + 171 + ], + "spans": [ + { + "bbox": [ + 143, + 143, + 449, + 171 + ], + "type": "text", + "content": "Peiji Yang* Huawei Wei* Yicheng Zhong* Zhisheng Wang \nTencent, Shenzhen, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 145, + 174, + 448, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 174, + 448, + 185 + ], + "spans": [ + { + "bbox": [ + 145, + 174, + 448, + 185 + ], + "type": "text", + "content": "{peijiyang, huaweiwei, ajaxhong, plorywang}@tencent.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 239, + 289, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 289, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 289, + 539 + ], + "type": "text", + "content": "Existing Speech-driven 3D facial animation methods typically follow the supervised paradigm, involving regression from speech to 3D facial animation. This paradigm faces two major challenges: the high cost of supervision acquisition, and the ambiguity in mapping between speech and lip movements. To address these challenges, this study proposes a novel cross-modal semi-supervised framework, comprising a Speech-to-Image Transcoder and a Face-to-Geometry Regressor. The former jointly learns a common representation space from speech and image domains, enabling the transformation of speech into semantically-consistent facial images. The latter is responsible for reconstructing 3D facial meshes from the transformed images. Both modules require minimal effort to acquire the necessary training data, thereby obviating the dependence on costly supervised data. Furthermore, the joint learning scheme enables the fusion of intricate visual features into speech encoding, thereby facilitating the transformation of subtle speech variations into nuanced lip movements, ultimately enhancing the fidelity of 3D face reconstructions. Consequently, the ambiguity of the direct mapping of speech-to-animation is significantly reduced, leading to coherent and high-fidelity generation of lip motion. Extensive experiments demonstrate that our approach produces competitive results compared to supervised methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 563, + 128, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 128, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 128, + 575 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 583, + 287, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 583, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 46, + 583, + 287, + 655 + ], + "type": "text", + "content": "Speech-driven 3D facial animation aims to automatically animate vivid facial expressions of realistic 3D avatars from the speech signal. It has become increasingly popular in many fields, such as games, virtual reality, film production, and online communication, as it enables the generation of lifelike facial expressions with minimal effort." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 656, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 656, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 656, + 287, + 693 + ], + "type": "text", + "content": "Recently proposed approaches, including VOCA[7], MeshTalk[28], and FaceFormer[10], have shown promising results in recovering 3D facial meshes from speech" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 213, + 547, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 213, + 547, + 345 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 547, + 345 + ], + "type": "text", + "content": "signals through the utilization of regressive networks or transformer-based autoregressive models. However, these methods, which fall under the supervised paradigm, encounter two primary challenges. Firstly, obtaining paired supervision of speech and high-fidelity 3D facial animation requires the utilization of a professional and expensive facial motion capture system, leading to considerable costs. Secondly, the mapping of low-dimensional speech signals to high-dimensional 3D facial meshes with significant variability may result in ambiguity issues, leading to suboptimal outcomes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 350, + 547, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 350, + 547, + 697 + ], + "spans": [ + { + "bbox": [ + 305, + 350, + 547, + 697 + ], + "type": "text", + "content": "This study for the first time proposes a novel cross-modal semi-supervised framework to address the above issues, which consists of a Speech-to-Image Transcoder and a Face-to-Geometry Regressor. The former is designed to transform speech into semantically-consistent facial images, while the latter is responsible for reconstructing 3D facial meshes from the transformed images. The design of the Speech-to-Image Transcoder is inspired by the recent success of unsupervised image-to-image translation [26], which allows for transformation between different input image domains. Our transcoder extends this mechanism to facilitate cross-modal conversion between speech and image domains. To achieve this capability, we learn a common representation space using data from three domains: paired speech and real facial images, as well as synthetic facial images. We train the translation between the real and synthetic image domains in an unsupervised manner, and simultaneously leverage the paired relationship between real images and speech to project speech features into the image representation space. Through this joint training, we construct a common space where the representations of speech and images are tied together. With the learned common representation, we can convert speech into synthetic facial images using our domain-dependent image decoder. Furthermore, we leverage the synthetic facial data, which is paired with rendered facial images and 3D face meshes, to train the Face-to-Geometry Regressor. This allows us to infer the corresponding 3D face meshes from the synthetic faces generated from speech." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "Our pipeline is capable of effortlessly acquiring the nec" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 215, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 215, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 215, + 713 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21032" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "essary data, which includes synchronized speech and real facial images extracted from readily available video clips, as well as synthetic facial images generated through the deformation of a polygonal face mesh using a rendering engine. Notably, our approach is not bound by any prerequisites for paired speech and 3D face animation, thereby effectively addressing the issue of supervision data scarcity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "text", + "content": "Moreover, our joint training scheme integrates the speech-to-image regression and image-to-image reconstruction into a cohesive framework, which enables the incorporation of detailed visual features into speech encoding. Leveraging the domain-dependent image decoders, which store intricate facial visual priors, we are able to convert nuanced speech variations into micro-expressions that are rich in detail, thereby facilitating the subsequent generation of high-fidelity 3D face reconstructions. Consequently, the uncertainties associated with the direct mapping of speech-to-animation in the supervised paradigm are considerably reduced, leading to the generation of lip motion with finer granularity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 312, + 284, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 312, + 284, + 323 + ], + "spans": [ + { + "bbox": [ + 59, + 312, + 284, + 323 + ], + "type": "text", + "content": "In summary, the main contributions of our research are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 332, + 287, + 479 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 332, + 286, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 332, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 59, + 332, + 286, + 380 + ], + "type": "text", + "content": "- We make the first attempt to build a semi-supervised framework for speech-driven 3D facial animation, which eliminates the need for paired speech and 3D animation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 388, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 388, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 59, + 388, + 287, + 437 + ], + "type": "text", + "content": "- Our proposed joint training scheme leverages cross-modal translation to convert subtle speech variations into images that are rich in detail, resulting in more high-fidelity lip motion generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 444, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 444, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 59, + 444, + 287, + 479 + ], + "type": "text", + "content": "- Our extensive experiments and user studies demonstrate that our approach produces competitive results compared to supervised methods." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 490, + 134, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 134, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 134, + 502 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 510, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 569 + ], + "type": "text", + "content": "The proposed method is related to several research fields, such as unsupervised image-to-image translation, speech-driven 2D talking head, and speech-driven 3D face animation. In the following, we review the most relevant approaches." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": "Unsupervised image-to-image translation. The unsupervised image-to-image translation aims to convert images from one domain to another domain with certain characteristics of the source images preserved[22, 36]. The translation involves converting from silhouettes to real images [16], from segmentation masks to RGB images[8, 23], from face to face [5, 27], and so on. This technology has wide applications in many fields, such as style transfer [40], face swapping [5, 26, 27], expression transfer [25], pose estimation [20] etc. They typically employ GAN-based networks [40] or encoder-decoder architectures [16] to achieve the goals. Notably, they only focus on image domains," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "our method extends this mechanism to achieve cross-modal translation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 95, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 95, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 95, + 545, + 263 + ], + "type": "text", + "content": "Speech-driven 2D talking head. 2D-based talking head generation methods generally utilize different image representations to assist the conversion of speech to 2D faces, such as facial landmarks[4, 31, 37], depth maps[15] and semantic maps[6, 24]. The discriminator loss and other task-specific losses are usually employed for optimization. For example, Chen et al.[3] design an audio-visual derivative correlation loss to ensure cross-modal consistency. In addition, several works use 3D information such as 3D Morphable Model (3DMM)[2, 19, 30, 32, 39] parameters or dynamic neural radiance fields (NeRF)[12, 13, 35] to guide the generation process. Unlike these 2D-based methods, this paper focuses on high-fidelity 3D animation generation from speech." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "text", + "content": "Speech-driven 3D face animation. Recently, considerable endeavors have been undertaken in this field[7, 10, 11, 18, 28, 34, 38]. VOCA[7] casts the mapping from speech to animation as a regression issue. To train the regressor, it uses the paired data of speech and 3D face animation. FaceFormer[10] employs transformers to model the long-term dependencies of speech, leading to improved performance. Meshtalk[28] puts emphasis on addressing the scalability and realism of the model by decoupling audio-correlated and audio-uncorrelated information. All the above methods belong to the supervised paradigm, which requires costly paired data and suffers from the problem of ambiguous mapping when directly regressing 3D animation from speech. This paper develops a semi-supervised framework to convert speech into images and then reconstruct 3D faces from the transformed images, effectively alleviating the aforementioned two issues." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 477, + 389, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 477, + 389, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 477, + 389, + 491 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "type": "text", + "content": "The proposed semi-supervised method is divided into two stages. We first employ the Speech-to-Image Transcoder to learn a common representation space for data from three domains: speech data, real facial images, and synthetic facial images. With this, we transform speech into semantically-consistent synthetic facial images. Subsequently, we infer the 3D meshes of the synthetic images via the Face-to-Geometry Regressor. This section first introduces preliminaries of the necessary data of three domains, then describes the two modules in detail." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 624, + 417, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 624, + 417, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 624, + 417, + 635 + ], + "type": "text", + "content": "3.1. Data Preliminaries" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 642, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 689 + ], + "type": "text", + "content": "Unlike supervised methods that require paired data of speech and 3D facial animation, our framework only necessitates real facial videos synchronized with speech and a collection of rendered facial images." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We record the videos in an environment without background noise to ensure a clean recording of sound. The" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21033" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "camera is positioned to capture the speaker's front view, enabling clear capture of lip movements. In terms of rendered images, we utilize a readily available character for rendering, which is not required to have the same physiognomy as the speaker and can be easily acquired from online resources or produced by artists. Then, varying facial expressions are generated by deforming the character using performances of range-of-motion (ROM) and exercising poses from Facial Action Coding System (FACS)[9]. Consequently, each rendered image has its corresponding 3D facial mesh. More details of the dataset construction are presented in Section 4.1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 224, + 205, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 224, + 205, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 205, + 236 + ], + "type": "text", + "content": "3.2. Speech-to-Image Transcoder" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 240, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 287, + 337 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 240, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathcal{R}}^{1:T} = \\left(\\mathcal{I}_{\\mathcal{R}}^{1},\\dots ,\\mathcal{I}_{\\mathcal{R}}^{T}\\right)" + }, + { + "bbox": [ + 46, + 240, + 287, + 337 + ], + "type": "text", + "content": " denotes a sequence of real facial frames. Correspondingly, the sequence of speech snippets " + }, + { + "bbox": [ + 46, + 240, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{1:T} = \\left(\\mathbf{a}^1,\\dots ,\\mathbf{a}^T\\right)" + }, + { + "bbox": [ + 46, + 240, + 287, + 337 + ], + "type": "text", + "content": " is defined such that each " + }, + { + "bbox": [ + 46, + 240, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\mathbf{a}^t" + }, + { + "bbox": [ + 46, + 240, + 287, + 337 + ], + "type": "text", + "content": " is synchronized with the respective " + }, + { + "bbox": [ + 46, + 240, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathcal{R}}^t" + }, + { + "bbox": [ + 46, + 240, + 287, + 337 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 240, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}_S\\}" + }, + { + "bbox": [ + 46, + 240, + 287, + 337 + ], + "type": "text", + "content": " is a collection of synthetic facial images, and each facial image has its corresponding 3D face mesh " + }, + { + "bbox": [ + 46, + 240, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\nu_{S}" + }, + { + "bbox": [ + 46, + 240, + 287, + 337 + ], + "type": "text", + "content": ". These three types of data are utilized by the Speech-to-Image Transcoder to construct a common representation space for speech and images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 338, + 287, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 433 + ], + "type": "text", + "content": "Network architecture. Our proposed Speech-to-Image Transcoder consists of two encoders and two decoders. We employ a speech encoder " + }, + { + "bbox": [ + 46, + 338, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{A}}" + }, + { + "bbox": [ + 46, + 338, + 287, + 433 + ], + "type": "text", + "content": " to transform speech into a compact representation. A shared visual encoder " + }, + { + "bbox": [ + 46, + 338, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{V}}" + }, + { + "bbox": [ + 46, + 338, + 287, + 433 + ], + "type": "text", + "content": " is used to encode real and synthetic faces. Additionally, two decoders, " + }, + { + "bbox": [ + 46, + 338, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{R}}" + }, + { + "bbox": [ + 46, + 338, + 287, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 338, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{S}}" + }, + { + "bbox": [ + 46, + 338, + 287, + 433 + ], + "type": "text", + "content": ", are used to reconstruct real and synthetic faces from the encodings. The overview of our Speech-to-Image Transcoder is illustrated in Figure 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "text", + "content": "Specifically, the speech encoder " + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{A}}" + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "text", + "content": " consists of an audio feature extractor and a multi-layer transformer encoder[33]. The audio feature extractor is initialized with pre-trained wav2vec2.0[1] weights, responsible for converting raw waveform input into audio features. And the transformer encoder converts audio features into contextualized speech representations " + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_{\\mathcal{A}}^{1:T} = (\\mathbf{z}_{\\mathcal{A}}^{1},\\dots,\\mathbf{z}_{\\mathcal{A}}^{T})" + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "text", + "content": ". Simultaneously, real faces aligned to the audio are input to the visual encoder " + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{V}}" + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "text", + "content": " to obtain visual features " + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_{\\mathcal{R}}^{1:T} = (\\mathbf{z}_{\\mathcal{R}}^{1},\\dots,\\mathbf{z}_{\\mathcal{R}}^{T})" + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "text", + "content": ". This single visual encoder is shared for encoding synthetic facial features " + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_{\\mathcal{S}}^{\\{\\mathcal{B}\\}} = (\\mathbf{z}_{\\mathcal{S}}^{\\{\\mathcal{B}\\}}),\\{\\mathcal{B}\\} \\subseteq \\{\\mathcal{I}_{\\mathcal{S}}\\}" + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathcal{R}}" + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathcal{S}}" + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "text", + "content": " are respectively fed into decoder " + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{R}}" + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{S}}" + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "text", + "content": " for reconstructing the corresponding facial images. In addition, due to the paired relationship between real images and speech, " + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{S}}" + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "text", + "content": " can also be utilized to reconstruct synthetic images from " + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathcal{A}}" + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "text", + "content": ". We borrow the decoder structure from [26], which consists of several upsampling blocks with an upsampling scale of 2. This architecture of a single shared visual encoder and multiple decoders for different domains has been verified to have good applicability in face swapping [27] and expression transfer [25]. In our experiments, " + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathcal{A}},\\mathbf{z}_{\\mathcal{R}}" + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathcal{S}}" + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "text", + "content": " are all " + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "text", + "content": "-dim vectors, where " + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 434, + 288, + 700 + ], + "type": "text", + "content": " is set as 512." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 701, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 714 + ], + "type": "text", + "content": "The Speech-to-Image Transcoder learns a semantically" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "consistent representation of speech, real faces, and synthetic faces. It enables the domain-dependent decoder " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{R}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " to reconstruct the facial image with corresponding semantics from any one of the three encodings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 121, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 545, + 239 + ], + "type": "text", + "content": "Expression de-neutralization We observe that in cases where the two facial domains exhibit significant differences, such as one comprising real faces and the other consisting of cartoonized synthetic faces, the semantic distributions between the two domains may not be well aligned after encoding. This misalignment can result in semantically mismatched image-to-image translation. For instance, the decoder " + }, + { + "bbox": [ + 304, + 121, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S" + }, + { + "bbox": [ + 304, + 121, + 545, + 239 + ], + "type": "text", + "content": " may erroneously decode the encoding of a real face exhibiting a grinning expression as a synthetic face with pouting expression." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 240, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 407 + ], + "type": "text", + "content": "To address this issue, we propose a simple yet effective operation that we refer to as expression de-neutralization (denoted in Equation 1). Rather than reconstructing the facial expression image directly from the encoding " + }, + { + "bbox": [ + 304, + 240, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i, i \\in \\{\\mathcal{A}, \\mathcal{R}, \\mathcal{S}\\}" + }, + { + "bbox": [ + 304, + 240, + 545, + 407 + ], + "type": "text", + "content": " we first subtract the neutral expression code " + }, + { + "bbox": [ + 304, + 240, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\overline{z}_i" + }, + { + "bbox": [ + 304, + 240, + 545, + 407 + ], + "type": "text", + "content": " corresponding to the domain (for speech domain, we use " + }, + { + "bbox": [ + 304, + 240, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\overline{z}_{\\mathcal{R}}" + }, + { + "bbox": [ + 304, + 240, + 545, + 407 + ], + "type": "text", + "content": ") and then recover the image with the offset expression code " + }, + { + "bbox": [ + 304, + 240, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\Delta z_i" + }, + { + "bbox": [ + 304, + 240, + 545, + 407 + ], + "type": "text", + "content": ". We manually select a neutral expression image for each facial domain and obtain the neutral expression code after encoding. Through expression de-neutralization, two semantically misaligned domains can be shifted to become aligned, leading to semantically-consistent translation. Our experiments profoundly verify the effectiveness of this operation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 393, + 419, + 545, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 419, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 393, + 419, + 545, + 431 + ], + "type": "interline_equation", + "content": "\\Delta z _ {i} = z _ {i} - \\bar {z} _ {i}. \\tag {1}", + "image_path": "1b31aa11d73b3fc22dc47ff64568315683a5944e17b96be86fe27e0735024a02.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "type": "text", + "content": "Training objectives. We employ two commonly used loss terms for image reconstruction, combined in Equation 2. The first term denotes the mean error of all pixels between the input image and the reconstructed one " + }, + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "type": "inline_equation", + "content": "(\\widetilde{\\mathcal{I}}_{\\mathcal{R}}" + }, + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{I}}_S)" + }, + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "type": "text", + "content": ". The second term " + }, + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{percep}" + }, + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "type": "text", + "content": " is the perceptual loss[17] to measure the similarity of two different images. It has obvious benefits for enhancing the fidelity of the reconstructed image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 541, + 545, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 541, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 320, + 541, + 545, + 555 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\left| \\left| \\mathcal {D} _ {i} \\left(\\Delta z _ {i}\\right) - \\mathcal {I} _ {i} \\right| \\right| _ {2} + \\mathcal {L} _ {p e r c e p}, i \\in \\{\\mathcal {R}, \\mathcal {S} \\} \\tag {2}", + "image_path": "538f4a1df1c0c7426e48e52c2559dee5f14bf6d7cd801f6449b5213a78fb038f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "text", + "content": "Moreover, owing to the fact that the speech and real facial images are inherently paired, we are able to enhance temporal synchronization and consistency between speech and visual features by imposing our designed cross-modal constraint (Equation 3). This constraint is comprised of two terms, with " + }, + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cf}" + }, + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "text", + "content": " representing the cross-modal constraint on features, and " + }, + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cs}" + }, + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "text", + "content": " representing that on image spatial space." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 367, + 667, + 455, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 667, + 455, + 681 + ], + "spans": [ + { + "bbox": [ + 367, + 667, + 455, + 681 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c r o s s} = \\mathcal {L} _ {c f} + \\mathcal {L} _ {c s},", + "image_path": "548cdbb89d7583519d0437f7a1f33db291a16e7a84adfa49f1265992623c5df4.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 368, + 683, + 453, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 683, + 453, + 696 + ], + "spans": [ + { + "bbox": [ + 368, + 683, + 453, + 696 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c f} = | | z _ {\\mathcal {A}} - z _ {\\mathcal {R}} | | _ {2},", + "image_path": "fc91884382f88f22ae4c5791b41c77d19713b344e3ad10e7ec4b18a119e6acc5.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 368, + 697, + 545, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 697, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 368, + 697, + 545, + 711 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c s} = \\left| \\left| \\mathcal {D} _ {\\mathcal {R}} \\left(\\Delta z _ {\\mathcal {A}}\\right) - \\mathcal {I} _ {\\mathcal {R}} \\right| \\right| _ {2}. \\tag {3}", + "image_path": "005f8c6db300aa6b8f2c1cf97fbc9b5b9873a4fac646c44c73971fd43cf0c938.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21034" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 71, + 516, + 323 + ], + "blocks": [ + { + "bbox": [ + 80, + 71, + 516, + 323 + ], + "lines": [ + { + "bbox": [ + 80, + 71, + 516, + 323 + ], + "spans": [ + { + "bbox": [ + 80, + 71, + 516, + 323 + ], + "type": "image", + "image_path": "c0c2718a626df8f862a86828bd56bf8222986c7ac95d833b0e99f854fd5edb26.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "lines": [ + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "text", + "content": "Figure 1. The architecture of the Speech-to-Image Transcoder. Firstly, the speech is converted into " + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathcal{A}}" + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{A}}" + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "text", + "content": ". The red arrow shows how we input " + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathcal{A}}" + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{R}}" + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "text", + "content": " to obtain the reconstructed image for calculating " + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cs}" + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "text", + "content": ". The green arrow shows how we obtain " + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "inline_equation", + "content": "z_{S}^{*}" + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "text", + "content": " for calculating " + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ccl}" + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "text", + "content": ". We feed the Real/Synthetic face images into the " + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{V}}" + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "text", + "content": " to obtain the visual representation " + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathcal{R}}" + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathcal{S}}" + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "text", + "content": ", which are respectively input into " + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{R}}" + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{S}}" + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "text", + "content": " to obtain reconstructed images for calculating the " + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 46, + 328, + 547, + 378 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 390, + 288, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 288, + 462 + ], + "type": "text", + "content": "To further encourage the transcoder to use a common representation across domains of speech and synthetic faces, we present a latent cycle consistency loss (Equation 4). This loss creates a loop constraint between the latent codes " + }, + { + "bbox": [ + 46, + 390, + 288, + 462 + ], + "type": "inline_equation", + "content": "z_{\\mathcal{A}}" + }, + { + "bbox": [ + 46, + 390, + 288, + 462 + ], + "type": "text", + "content": " derived from speech input and the latent codes " + }, + { + "bbox": [ + 46, + 390, + 288, + 462 + ], + "type": "inline_equation", + "content": "z_{\\mathcal{S}}^{*}" + }, + { + "bbox": [ + 46, + 390, + 288, + 462 + ], + "type": "text", + "content": " derived from the decoded synthetic images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 122, + 481, + 208, + 494 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 481, + 208, + 494 + ], + "spans": [ + { + "bbox": [ + 122, + 481, + 208, + 494 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c c l} = | | z _ {\\mathcal {S}} ^ {*} - z _ {\\mathcal {A}} | | _ {2},", + "image_path": "04cd5555b7e8c49a1eaa5279dbd1fc5853c8895666c7d24114aec35b64f1d17f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 495, + 287, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 495, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 123, + 495, + 287, + 510 + ], + "type": "interline_equation", + "content": "z _ {\\mathcal {S}} ^ {*} = \\mathcal {E} _ {\\mathcal {V}} (\\mathcal {D} _ {\\mathcal {S}} (\\Delta z _ {\\mathcal {A}})). \\qquad (4)", + "image_path": "77dcd09e573047c6cd8588295df65c962d1191330cdc14876d791176579247b9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 516, + 265, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 516, + 265, + 529 + ], + "spans": [ + { + "bbox": [ + 59, + 516, + 265, + 529 + ], + "type": "text", + "content": "To sum up, our total loss is depicted in Equation 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 536, + 287, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 536, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 102, + 536, + 287, + 550 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {\\text {c r o s s}} + \\mathcal {L} _ {\\text {c c l}}. \\tag {5}", + "image_path": "0e42692ba140364f73f52e599955246395ff944c728db431b1fd4dabd825e031.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 556, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 653 + ], + "type": "text", + "content": "By utilizing the proposed cross-modal joint training, encodings from three distinct domains are semantically aligned. This allows for any of the three input modalities to be transformed into semantically-consistent facial images through domain-specific decoders. Specifically, we convert speech into synthetic faces. By further leveraging the Faceto-Geometry Regressor in Section 3.3, 3D facial meshes can be reconstructed from the transformed images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 659, + 205, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 205, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 205, + 672 + ], + "type": "text", + "content": "3.3. Face-to-Geometry Regressor" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "The architecture of the Face-to-geometry Regressor " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{G}}" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": " is shown in Figure 2. It is trained in a supervised manner. We exclusively rely on information from the synthetic" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 323, + 391, + 531, + 462 + ], + "blocks": [ + { + "bbox": [ + 323, + 391, + 531, + 462 + ], + "lines": [ + { + "bbox": [ + 323, + 391, + 531, + 462 + ], + "spans": [ + { + "bbox": [ + 323, + 391, + 531, + 462 + ], + "type": "image", + "image_path": "655c7cc0bbeaab507c9c194996db7a3a84f18493a7aeb4b58be78ba909f89270.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "text", + "content": "Figure 2. The architecture of the Face-to-Geometry Regressor. The synthetic face images are fed into the Regressor " + }, + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{G}}" + }, + { + "bbox": [ + 304, + 472, + 545, + 520 + ], + "type": "text", + "content": ", which learns to regress the 3D mesh associated with the synthetic face images using RMSE loss." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 529, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 545, + 589 + ], + "type": "text", + "content": "image domain, specifically synthetic facial images and the corresponding 3D face meshes. We utilize Resnet34[14] as the backbone and a single fully connected layer to regress the coordinates of the 3D mesh " + }, + { + "bbox": [ + 304, + 529, + 545, + 589 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{V}}_S" + }, + { + "bbox": [ + 304, + 529, + 545, + 589 + ], + "type": "text", + "content": ". A simple RMSE loss is used to optimize the network, denoted as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 375, + 606, + 545, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 606, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 375, + 606, + 545, + 620 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g} = \\left| \\left| \\mathcal {E} _ {\\mathcal {G}} \\left(\\mathcal {I} _ {S}\\right) - \\mathcal {Y} _ {S} \\right| \\right| _ {2}. \\tag {6}", + "image_path": "c07f362cc53af1153b21c348dd4e45a0f2e6d65c67deafc2fd5bd8cccbcfeb18.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 546, + 715 + ], + "type": "text", + "content": "During the training phase, we employ data augmentation of color jittering and random affine transformation to enhance the robustness of our model. As a consequence, we are able to achieve satisfactory 3D face reconstruction results, even when the synthetic faces generated by the Speech-to-Image Transcoder are outside the training set." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21035" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 73, + 281, + 201 + ], + "blocks": [ + { + "bbox": [ + 55, + 73, + 281, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 73, + 281, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 73, + 281, + 201 + ], + "type": "image", + "image_path": "9c44563976a6ab42325d48d7464639cf45071f8597cd0929642b7103eca9eedf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 287, + 261 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 287, + 261 + ], + "type": "text", + "content": "Figure 3. Inference pipeline. The raw waveform is fed to the speech encoder " + }, + { + "bbox": [ + 46, + 213, + 287, + 261 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{A}}" + }, + { + "bbox": [ + 46, + 213, + 287, + 261 + ], + "type": "text", + "content": ". The output is subtracted by " + }, + { + "bbox": [ + 46, + 213, + 287, + 261 + ], + "type": "inline_equation", + "content": "\\overline{z}_{\\mathcal{R}}" + }, + { + "bbox": [ + 46, + 213, + 287, + 261 + ], + "type": "text", + "content": " and then fed into " + }, + { + "bbox": [ + 46, + 213, + 287, + 261 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{S}}" + }, + { + "bbox": [ + 46, + 213, + 287, + 261 + ], + "type": "text", + "content": " to recover the synthetic facial images, which are then input into " + }, + { + "bbox": [ + 46, + 213, + 287, + 261 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{G}}" + }, + { + "bbox": [ + 46, + 213, + 287, + 261 + ], + "type": "text", + "content": " to generate the 3D face animation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 270, + 144, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 144, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 144, + 281 + ], + "type": "text", + "content": "3.4. Inference Phase" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 290, + 288, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 288, + 434 + ], + "type": "text", + "content": "As illustrated in Figure 3, the Speech-to-Image Transcoder and Face-to-Geometry Regressor are integrated to infer 3D face animation from speech. The raw waveform is fed into the speech encoder " + }, + { + "bbox": [ + 46, + 290, + 288, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_A" + }, + { + "bbox": [ + 46, + 290, + 288, + 434 + ], + "type": "text", + "content": " to generate the speech representation. Then we subtract the speech representation by " + }, + { + "bbox": [ + 46, + 290, + 288, + 434 + ], + "type": "inline_equation", + "content": "\\overline{z}_{\\mathcal{R}}" + }, + { + "bbox": [ + 46, + 290, + 288, + 434 + ], + "type": "text", + "content": ", and the offset code is fed into " + }, + { + "bbox": [ + 46, + 290, + 288, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S" + }, + { + "bbox": [ + 46, + 290, + 288, + 434 + ], + "type": "text", + "content": " to recover the synthetic facial images. Due to the cross-modal constraint " + }, + { + "bbox": [ + 46, + 290, + 288, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cross}" + }, + { + "bbox": [ + 46, + 290, + 288, + 434 + ], + "type": "text", + "content": ", the transformed synthetic facial images are endowed with lip movements reflecting the content of the input speech. Finally, the face-to-geometry model is utilized to generate the 3D face animation from the transformed synthetic facial images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 449, + 119, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 119, + 461 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 119, + 461 + ], + "type": "text", + "content": "3.5. Discussion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 472, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 603 + ], + "type": "text", + "content": "One might raise doubts about the rationale behind not utilizing the image-to-image modules to accomplish the transformation from real faces to synthetic faces, subsequently employing the Face-to-Geometry Regressor to acquire pseudo-labels, and ultimately training a speech-topseudo-label network in a supervised manner. (Hereafter, we refer to this pipeline as the three-stage method.) We notice that while this approach can alleviate the challenge of limited supervision data, it still suffers from the problem of ambiguous mapping between speech and animation, the same as other regression-based supervised methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Our cross-modal joint training scheme can identify latent patterns and correlations between speech and images. This allows us to translate subtle variations in speech into detailed lip motions, creating highly-fidelity 3D reconstructions. Therefore, our approach has a distinct advantage in mitigating the uncertainties of direct speech-to-animation mapping. We conduct detailed experiments to demonstrate the advantages of our method compared with the three-stage method in Section 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 71, + 387, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 387, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 387, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 91, + 368, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 368, + 103 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 368, + 103 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 111, + 545, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 111, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 305, + 111, + 545, + 170 + ], + "type": "text", + "content": "Our experiments are conducted on two datasets: the self-built dataset SelfCollected and the public dataset VOCASET[7]. In addition, to support the semi-supervised cross-modal training, we generate two synthetic datasets. The details are described below." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 171, + 546, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 546, + 411 + ], + "type": "text", + "content": "SelfCollected dataset. This dataset contains 148 speaking video clips of one speaker. Each video contains one sentence. We employ the 3D facial capture system's workflow to obtain the ground truth (GT). Specifically, a head-mounted facial capture system is used to record the videos so that detailed lip motions of the speaker can be accurately tracked. Then, the facial tracking module captures several key facial landmarks. For each video, the artists manually pair a real person's face with the CG character at several keyframes. Finally, the facial retargeting module is trained on these keyframes and then automatically directs the motions of other video frames onto the CG character. Note that the character has not the same physiognomy as the speaker, so the range of lip movements is different after face retargeting. Through the workflows, we obtain the corresponding 3D face mesh with the topology of the CG character for each video frame. In total, this dataset contains 45000 frames, each frame corresponding a face mesh with 12696 vertices. We take 128 videos for training and the remaining 20 videos for testing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 412, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 496 + ], + "type": "text", + "content": "VOCASET dataset. VOCASET contains 12 speakers, each with 40 utterances. Facial meshes are scanned at 60FPS and then registered into a unified topology using the FLAME model[21], with each mesh containing 5023 vertices. In our experiments, we only use the data of one speaker for training and testing. We use 36 utterances as the training set and the remaining 4 as the test set." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 497, + 546, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 546, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 546, + 652 + ], + "type": "text", + "content": "SyntheticSelf dataset. This dataset is built to assist the semi-supervised training on SelfCollected. We use the same CG character in SelfCollected for rendering. During rendering, we keep the camera facing the character, and the rendering lighting remains unchanged. Animations of performances of range-of-motion (ROM) and exercising poses from the Facial Action Coding System (FACS) are utilized for rendering. Specifically, we require artists manually craft the blendshape weights of ROM and FACS by linearly combining the blendshapes of the CG character, enabling us to achieve a wide range of expressions. Finally, SyntheticSelf comprises 20000 rendered facial images, each corresponding to a facial mesh with 12696 vertices." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "SyntheticVOCA dataset. Similar to SyntheticSelf, we create SyntheticVOCA to assist the semi-supervised training on VOCASET. We use FLAME model to fit expression coefficients for each rendered image in SyntheticSelf. Then expressions are transferred to the character (i.e. the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21036" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "speaker's scan) provided by VOCASET. In total, we obtain 20000 rendered facial images, each corresponding to a FLAME-topology facial mesh with 5023 vertices." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 116, + 162, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 116, + 162, + 129 + ], + "spans": [ + { + "bbox": [ + 47, + 116, + 162, + 129 + ], + "type": "text", + "content": "4.2. Compared methods" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 135, + 287, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 135, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 135, + 287, + 351 + ], + "type": "text", + "content": "We compare our method with several SOTA methods: VOCA[7], MeshTalk[28] and FaceFormer[10], as well as the three-stage method discussed in Section 3.5. All methods are compared on SelfCollected and VOCASET. Notably, since MeshTalk cannot be trained on a small single-speaker dataset, we only perform qualitative comparisons with its officially released model. In addition, we find VOCA cannot converge on a single-speaker dataset, so we use all the data of the other 11 speakers combined with the single speaker's 36 sentences for training. The three-stage method employs an image-to-image and face-to-geometry pipeline to obtain pseudo-labels and then experiments with network architectures of VOCA and Faceformer, respectively. Since VOCASET does not have videos of real faces, leading to that we cannot obtain the corresponding pseudolabels, so we do not compare with the three-stage method on VOCASET. We use the symbol * later to represent the corresponding method using the three-stage scheme." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 351, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 287, + 399 + ], + "type": "text", + "content": "To sum up, there are 7 comparisons on SelfCollected: ours, VOCA, FaceFormer, GT, VOCA*, FaceFormer*, and pseudo GT. And 4 comparisons on SelfCollected: ours, VOCA, FaceFormer, and GT." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 407, + 183, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 183, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 183, + 420 + ], + "type": "text", + "content": "4.3.Quantitative Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 556 + ], + "type": "text", + "content": "We employ the commonly used lip vertex error[28] to evaluate the quantitative performance, which represents the maximal " + }, + { + "bbox": [ + 46, + 426, + 287, + 556 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 46, + 426, + 287, + 556 + ], + "type": "text", + "content": " error of vertices in the lip region. In addition, we propose the lip landmark error to measure the lip movement difference between the generated 3D animation and real facial sequence. To obtain this error, we first estimate a similarity transformation based on the projected 2D coordinates of the predicted 3D lip vertices and the 2D lip landmarks of the real face, then align them using the estimated transformation, and finally compute the root mean squared error." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "Results are illustrated in Table 1 and Table 2. Note that GT is obtained by retargeting the lip motions from the real speaker to the CG character, so it reflects the idiosyncrasies of the character instead of the speaker. Therefore, it is not objective to compare us with the supervised methods in terms of the metric of lip vertex error, and similarly, there is no comparability between supervised methods and ours in terms of the metric of lip landmark error. Nonetheless, our semi-supervised method achieves comparable performance to supervised methods in terms of lip vertex error, and even outperforms VOCA on SelfCollected. As for the lip landmark error, our method significantly outperforms the supervised approach and consistently outperforms the three-stage" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 72, + 552, + 193 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 552, + 193 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 552, + 193 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 552, + 193 + ], + "type": "table", + "html": "
MethodsLip vertex error (in mm)↓Lip landmark error (in pixel)↓
VOCA3.7185.908
FaceFormer2.6905.845
GT05.041
VOCA*5.8383.790
FaceFormer*4.5363.613
Pseudo GT3.7732.941
Ours3.7083.241
", + "image_path": "3ff901ff3513ec5e76690b06f9f868d81bf9e12d54ada2f88d5b50a4b7f6b3b1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 354, + 217, + 499, + 290 + ], + "blocks": [ + { + "bbox": [ + 332, + 196, + 518, + 208 + ], + "lines": [ + { + "bbox": [ + 332, + 196, + 518, + 208 + ], + "spans": [ + { + "bbox": [ + 332, + 196, + 518, + 208 + ], + "type": "text", + "content": "Table 1. Quantitative results on SelfCollected." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 354, + 217, + 499, + 290 + ], + "lines": [ + { + "bbox": [ + 354, + 217, + 499, + 290 + ], + "spans": [ + { + "bbox": [ + 354, + 217, + 499, + 290 + ], + "type": "table", + "html": "
MethodsLip vertex error (in mm)↓
VOCA4.690
FaceFormer3.166
Ours5.102
", + "image_path": "82f25b207707c313a3110a40d43527db056a931ff47b3f0b8acfb40f460a0bc6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 336, + 293, + 515, + 305 + ], + "lines": [ + { + "bbox": [ + 336, + 293, + 515, + 305 + ], + "spans": [ + { + "bbox": [ + 336, + 293, + 515, + 305 + ], + "type": "text", + "content": "Table 2. Quantitative results on VOCASET." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 316, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 365 + ], + "type": "text", + "content": "methods, even approaching the pseudo GT, which is the upper bound of the three-stage methods. This indicates that our joint training approach indeed promotes generating accurate and higher fidelity lip animations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 371, + 435, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 435, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 435, + 384 + ], + "type": "text", + "content": "4.4. Qualitative Evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 389, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 497 + ], + "type": "text", + "content": "Figure 4 presents qualitative results obtained by comparing three sentences. Sentences A and B come from SelfCollected-Test, and sentence C comes from VOCASET-Test. They belong to two different languages, namely Chinese and English. We evaluate sentence A with the model trained on SelfCollected and evaluate sentences B, C using the model trained on VOCASET. It should be noted that sentence B does not have its FLAME-topology GT meshes, and sentence C does not have the corresponding real faces." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 498, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 640 + ], + "type": "text", + "content": "From the results of sentence A, we can see that our method achieves competitive results compared to the supervised approaches, and even gains better performance in some cases (highlighted by the red box) in terms of mouth closure and lip movement amplitude. From the results of sentence C, our method achieves results that are highly synchronized with the speech even without using speech data from VOCASET for training. The results of sentence A also demonstrate that compared to the three-stage methods, our method presents results that are more consistent with the lip movements of the real face. This indicates our joint training scheme helps to generate more realistic animations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 647, + 432, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 432, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 432, + 659 + ], + "type": "text", + "content": "4.5. Perceptual Evaluation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 666, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 689 + ], + "type": "text", + "content": "We conduct user studies to evaluate the animation quality of our method, which comprise a total of 10 participants." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "User Study on SelfCollected. For SelfCollected, we obtain the results of 7 comparisons using 10 random test sen" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21037" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 78, + 294, + 666 + ], + "blocks": [ + { + "bbox": [ + 50, + 78, + 294, + 666 + ], + "lines": [ + { + "bbox": [ + 50, + 78, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 50, + 78, + 294, + 666 + ], + "type": "image", + "image_path": "8464b4cf1dc13883dc7fe70c9f5cad641a36c7cc27b62374b0071c1c06937ba2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 673, + 547, + 711 + ], + "lines": [ + { + "bbox": [ + 46, + 673, + 547, + 711 + ], + "spans": [ + { + "bbox": [ + 46, + 673, + 547, + 711 + ], + "type": "text", + "content": "Figure 4. Qualitative results. We utilize the officially released model to produce MeshTalk results as it is not trainable on SelfCollected and VOCASET datasets. The highlighted section by red box indicates that our method has better performance in terms of mouth closure and lip movement amplitude." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 299, + 78, + 539, + 361 + ], + "blocks": [ + { + "bbox": [ + 299, + 78, + 539, + 361 + ], + "lines": [ + { + "bbox": [ + 299, + 78, + 539, + 361 + ], + "spans": [ + { + "bbox": [ + 299, + 78, + 539, + 361 + ], + "type": "image", + "image_path": "a9e2d19c5e7011fba0ff14cc3cd9dd122d169abb117b043c0e6a03af0e4a3223.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 300, + 361, + 539, + 661 + ], + "blocks": [ + { + "bbox": [ + 300, + 361, + 539, + 661 + ], + "lines": [ + { + "bbox": [ + 300, + 361, + 539, + 661 + ], + "spans": [ + { + "bbox": [ + 300, + 361, + 539, + 661 + ], + "type": "image", + "image_path": "2893f2164fef244d4ed9c18c988e2258bc386eb75f177610676fade5d78a2900.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21038" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 291, + 163 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 291, + 163 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 291, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 291, + 163 + ], + "type": "table", + "html": "
Ours vs. CompetitorRealism↑Lip Sync↑Sim↑
Ours vs. VOCA51.053.063.0
Ours vs. FaceFormer45.049.057.0
Ours vs. GT43.045.058.0
Ours vs. VOCA*75.064.060.0
Ours vs. FaceFormer*58.061.054.0
Ours vs. pesudo GT52.051.048.0
", + "image_path": "12b5c37dabfa8c3f804900c211dd595d7de74d040f44696f4671572f8c5cba77.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 229, + 279, + 361 + ], + "blocks": [ + { + "bbox": [ + 46, + 167, + 288, + 202 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 288, + 202 + ], + "type": "text", + "content": "Table 3. User study results on SelfCollected. We use A/B testing and report the percentage (\\%) of answers where A is preferred over B." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 229, + 279, + 361 + ], + "lines": [ + { + "bbox": [ + 56, + 229, + 279, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 279, + 361 + ], + "type": "table", + "html": "
Ours vs. CompetitorRealism↑
VOCASETSelf.Avg.
Ours vs. GT17.5——
Ours vs. FaceFormer40.057.548.8
Ours vs. VOCA45.095.070.0
Ours vs. CompetitorLip Sync↑
VOCASETSelf.Avg.
Ours vs. GT5.0——
Ours vs. FaceFormer15.057.536.3
Ours vs. VOCA60.080.070.0
", + "image_path": "2c1d81c95fd6b5d9a623a2be9d9b6ed4af90911aaeb7020a18a0f991cd48c772.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 387, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 287, + 590 + ], + "type": "text", + "content": "tences of the test set. Therefore, 60 A vs. B pairs(10 videos x 6 comparisons) are created. For each pair, we ask the participants to choose their favorite one in terms of realism and lip sync, following the user study protocol of Faceformer. Additionally, we propose a new metric \"similarity\" to characterize the preference for which method generates more similar lip movements to that of the real faces. For this metric, we provide both A/B pair and the real facial frames for each sentence. And then we ask the participants to choose the one that has closer lip amplitude to the real face. Table 3 shows the results. We observe that our method is comparable to the SOTA supervised methods in terms of realism and lip sync. And our method significantly outperforms the supervised approach in terms of similarity. Furthermore, the results show that our method consistently performs better than the three-stage methods. It indicates that our joint training of speech and image is indeed effective." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "User Study on VOCASET. We conduct another user study on VOCASET. Since our model has never seen the speech data in VOCASET, for fairness, we additionally randomly pick 4 sentences for comparison from SelfCollected-Test. In total, 20 A vs. B pairs (there is no FLAMEtopology GT in SelfCollected) are created. For each pair, participants are asked to choose their favorite one based on realism and lip sync. As shown in Table 4, we separately report the results of SelfCollected-Test and VOCASET-Test. The results show that our method consistently performs bet" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 335, + 70, + 518, + 143 + ], + "blocks": [ + { + "bbox": [ + 72, + 364, + 261, + 376 + ], + "lines": [ + { + "bbox": [ + 72, + 364, + 261, + 376 + ], + "spans": [ + { + "bbox": [ + 72, + 364, + 261, + 376 + ], + "type": "text", + "content": "Table 4. User study results " + }, + { + "bbox": [ + 72, + 364, + 261, + 376 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 72, + 364, + 261, + 376 + ], + "type": "text", + "content": " on VOCASET." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 70, + 518, + 143 + ], + "lines": [ + { + "bbox": [ + 335, + 70, + 518, + 143 + ], + "spans": [ + { + "bbox": [ + 335, + 70, + 518, + 143 + ], + "type": "table", + "html": "
Expression de-neutralizationLcclLip lmk error (in pixel)↓
✓—3.407
—✓3.482
✓✓3.241
", + "image_path": "b5fc6177763c725446627a12668d53c4eeea2d59e04ccaaf11378241dfe253a3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 341, + 146, + 510, + 159 + ], + "lines": [ + { + "bbox": [ + 341, + 146, + 510, + 159 + ], + "spans": [ + { + "bbox": [ + 341, + 146, + 510, + 159 + ], + "type": "text", + "content": "Table 5. Ablation study on SelfCollected." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 331, + 168, + 533, + 293 + ], + "blocks": [ + { + "bbox": [ + 331, + 168, + 533, + 293 + ], + "lines": [ + { + "bbox": [ + 331, + 168, + 533, + 293 + ], + "spans": [ + { + "bbox": [ + 331, + 168, + 533, + 293 + ], + "type": "image", + "image_path": "e529259c93538fe5922ec47cb4a4cfd475123d2a22b9d84e73949197d926fad5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 295, + 541, + 307 + ], + "lines": [ + { + "bbox": [ + 309, + 295, + 541, + 307 + ], + "spans": [ + { + "bbox": [ + 309, + 295, + 541, + 307 + ], + "type": "text", + "content": "Figure 5. Ablation study on expression de-neutralization." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 310, + 545, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 310, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 305, + 310, + 545, + 370 + ], + "type": "text", + "content": "ter in terms of realism. However, our model achieves worse lip sync than FaceFormer on VOCASET-Test. We claim that it is reasonable because our model is not trained with the English speech data of VOCASET. Similarly, the Faceformer's result on SelfCollected-Test is worse than ours." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 378, + 399, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 378, + 399, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 378, + 399, + 392 + ], + "type": "text", + "content": "4.6. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 397, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 545, + 589 + ], + "type": "text", + "content": "Effect of Expression De-neutralization. We investigate the effect of expression de-neutralization by removing it from our method. We conduct the comparison on Self-Collected. As depicted in Table 5, the operation of expression de-neutralization helps to reduce lip landmark error by " + }, + { + "bbox": [ + 304, + 397, + 545, + 589 + ], + "type": "inline_equation", + "content": "0.241" + }, + { + "bbox": [ + 304, + 397, + 545, + 589 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 397, + 545, + 589 + ], + "type": "inline_equation", + "content": "3.482 \\rightarrow 3.241" + }, + { + "bbox": [ + 304, + 397, + 545, + 589 + ], + "type": "text", + "content": "). We also visualize the decoded synthetic faces in Figure 5. As depicted in Figure 5, the incorporation of expression de-neutralization results in more consistent lip motions of synthetic faces with that of real faces. Without this operation, the decoded faces demonstrate obvious semantic mismatches, particularly when the disparity between the two image domains is substantial (i.e. real faces and rendered FLAME-topology faces). This phenomenon proves that expression de-neutralization helps align semantic distributions of different domains, resulting in more semantically consistent lip movement generation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 589, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 589, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 589, + 545, + 661 + ], + "type": "text", + "content": "Effect of Cycle Consistency Loss " + }, + { + "bbox": [ + 304, + 589, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ccl}" + }, + { + "bbox": [ + 304, + 589, + 545, + 661 + ], + "type": "text", + "content": ". We present the quantitative results in Table 5. As shown, " + }, + { + "bbox": [ + 304, + 589, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ccl}" + }, + { + "bbox": [ + 304, + 589, + 545, + 661 + ], + "type": "text", + "content": " helps to reduce lip landmark error by 0.166 (3.407 → 3.241). The qualitative comparison is depicted in Figure 6, it is indicated this loss function promotes the generation of lip movements that are more consistent with the speech." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 669, + 378, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 378, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 378, + 681 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "This paper aims to recover 3D animation from speech, it proposes a novel semi-supervised framework to address the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21039" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 82, + 274, + 228 + ], + "blocks": [ + { + "bbox": [ + 121, + 72, + 240, + 81 + ], + "lines": [ + { + "bbox": [ + 121, + 72, + 240, + 81 + ], + "spans": [ + { + "bbox": [ + 121, + 72, + 240, + 81 + ], + "type": "text", + "content": "wo hai bu zhi dao ni ren shi fu lan ke" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 82, + 274, + 228 + ], + "lines": [ + { + "bbox": [ + 66, + 82, + 274, + 228 + ], + "spans": [ + { + "bbox": [ + 66, + 82, + 274, + 228 + ], + "type": "image", + "image_path": "56ad41857693de04c44c9ee4acdbf184dd514d1e00e9373404a66aab5ecc2104.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 231, + 235, + 243 + ], + "lines": [ + { + "bbox": [ + 99, + 231, + 235, + 243 + ], + "spans": [ + { + "bbox": [ + 99, + 231, + 235, + 243 + ], + "type": "text", + "content": "Figure 6. Ablation study on " + }, + { + "bbox": [ + 99, + 231, + 235, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ccl}" + }, + { + "bbox": [ + 99, + 231, + 235, + 243 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 252, + 289, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 289, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 289, + 419 + ], + "type": "text", + "content": "data scarcity problem of the supervised methods. In addition, a joint training scheme on cross-domain data, namely speech and images, is developed. Extensive experiments show the proposed method achieves satisfactory generation of lip motions. However, the major limitation of our method is its inability for multiple identities. To address this issue, we suggest two possible ways for future work. The first is expanding the decoder for each speaker, which succeeds in faceswap but increases model size with more IDs. The second is inspired by FOMM [29], which uses a universal motion module for facial motion extraction and transfer. Overall, our proposed method shows promising results and opens up opportunities for further research in the field of speech-driven 3D facial animation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 432, + 107, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 432, + 107, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 432, + 107, + 445 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 453, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 53, + 453, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 287, + 497 + ], + "type": "text", + "content": "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 498, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 287, + 542 + ], + "type": "text", + "content": "[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 544, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 544, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 53, + 544, + 287, + 588 + ], + "type": "text", + "content": "[3] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European conference on computer vision (ECCV), pages 520-535, 2018." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 590, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 590, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 53, + 590, + 287, + 644 + ], + "type": "text", + "content": "[4] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7832-7841, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 52, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 52, + 646, + 288, + 689 + ], + "type": "text", + "content": "[5] Renwang Chen, Xuanhong Chen, Bingbing Ni, and Yanhao Ge. Simswap: An efficient framework for high fidelity face swapping. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2003–2011, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 287, + 714 + ], + "type": "text", + "content": "[6] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 107, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 545, + 160 + ], + "type": "text", + "content": "[7] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J. Black. Capture, learning, and synthesis of 3d speaking styles. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10093-10103, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 162, + 545, + 205 + ], + "type": "text", + "content": "[8] Ricard Durall Lopez, Jireh Jam, Dominik Strassel, Moi Hoon Yap, and Janis Keuper. Facialgan: Style transfer and attribute manipulation on synthetic faces. In [32nd British Machine Vision Conference], pages 1-14, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 206, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 206, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 312, + 206, + 545, + 237 + ], + "type": "text", + "content": "[9] Paul Ekman and Wallace V Friesen. Facial action coding system. Environmental Psychology & Nonverbal Behavior, 1978." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "type": "text", + "content": "[10] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18749-18758, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 294, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 348 + ], + "type": "text", + "content": "[11] Ohad Fried, Ayush Tewari, Michael Zollhöfer, Adam Finkelstein, Eli Shechtman, Dan B Goldman, Kyle Genova, Zeyu Jin, Christian Theobalt, and Maneesh Agrawala. Text-based editing of talking-head video. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 350, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 404 + ], + "type": "text", + "content": "[12] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "type": "text", + "content": "[13] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 459, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 459, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 459, + 545, + 503 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "type": "text", + "content": "[15] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3397-3406, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "text", + "content": "[16] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "type": "text", + "content": "[17] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[18] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21040" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[19] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "type": "text", + "content": "[20] Siyuan Li, Semih Gunel, Mirela Ostrek, Pavan Ramdya, Pascal Fua, and Helge Rhodin. Deformation-aware unpaired image translation for pose estimation on laboratory animals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13158-13168, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 218 + ], + "type": "text", + "content": "[21] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 253 + ], + "type": "text", + "content": "[22] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "type": "text", + "content": "[23] Wenshuang Liu, Wenting Chen, Zhanjia Yang, and Linlin Shen. Translate the facial regions you like using self-adaptive region translation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2180-2188, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 376 + ], + "type": "text", + "content": "[24] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural audio-driven video portrait generation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 106-125. Springer, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "type": "text", + "content": "[25] Lucio Moser, Chinyu Chien, Mark Williams, Jose Serra, Darren Hendler, and Doug Roble. Semi-supervised video-driven facial animation transfer for production. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 466 + ], + "type": "text", + "content": "[26] Jacek Naruniec, Leonhard Helminger, Christopher Schroers, and Romann M Weber. High-resolution neural face swapping for visual effects. In Computer Graphics Forum, volume 39, pages 173-184. Wiley Online Library, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 467, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 521 + ], + "type": "text", + "content": "[27] Ivan Perov, Daiheng Gao, Nikolay Chervoniy, Kunlin Liu, Sugasa Marangonda, Chris Ume, Mr Dpfks, Carl Shift Facenheim, Luis RP, Jian Jiang, et al. Deepfacelab: Integrated, flexible and extensible face-swapping framework. arXiv preprint arXiv:2005.05535, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 522, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 578 + ], + "type": "text", + "content": "[28] Alexander Richard, Michael Zollhoefer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1153-1162, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 578, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 623 + ], + "type": "text", + "content": "[29] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "text", + "content": "[30] Linsen Song, Wayne Wu, Chen Qian, Ran He, and Chen Change Loy. Everybody's talkin': Let me talk as you want. IEEE Transactions on Information Forensics and Security, 17:585-598, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[31] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 552 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 139 + ], + "type": "text", + "content": "[32] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 716-731. Springer, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 140, + 546, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 546, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 546, + 184 + ], + "type": "text", + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 185, + 546, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 546, + 239 + ], + "type": "text", + "content": "[34] Chih-Chun Yang, Wan-Cyuan Fan, Cheng-Fu Yang, and Yu-Chiang Frank Wang. Cross-modal mutual learning for audiovisual speech recognition and manipulation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 3036-3044, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 240, + 546, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 240, + 546, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 240, + 546, + 284 + ], + "type": "text", + "content": "[35] Shunyu Yao, RuiZhe Zhong, Yichao Yan, Guangtao Zhai, and Xiaokang Yang. Dfa-nerf: personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 285, + 546, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 546, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 546, + 329 + ], + "type": "text", + "content": "[36] Zili Yi, Hao Zhang, Ping Tan, and Minglun Gong. Dualgan: Unsupervised dual learning for image-to-image translation. In Proceedings of the IEEE international conference on computer vision, pages 2849-2857, 2017." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 331, + 546, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 546, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 546, + 373 + ], + "type": "text", + "content": "[37] Lingyun Yu, Jun Yu, and Qiang Ling. Mining audio, text and visual information for talking face generation. In 2019 IEEE International Conference on Data Mining (ICDM), pages 787-795. IEEE, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 374, + 546, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 374, + 546, + 440 + ], + "spans": [ + { + "bbox": [ + 308, + 374, + 546, + 440 + ], + "type": "text", + "content": "[38] Sibo Zhang, Jiahong Yuan, Miao Liao, and Liangjun Zhang. Text2video: Text-driven talking-head video synthesis with personalized phoneme-pose dictionary. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2659-2663. IEEE, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 441, + 547, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 441, + 547, + 496 + ], + "spans": [ + { + "bbox": [ + 308, + 441, + 547, + 496 + ], + "type": "text", + "content": "[39] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 497, + 546, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 497, + 546, + 552 + ], + "spans": [ + { + "bbox": [ + 308, + 497, + 546, + 552 + ], + "type": "text", + "content": "[40] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21041" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning/219bbdca-cfe8-433f-a6db-97c4773a1a10_content_list.json b/2023/Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning/219bbdca-cfe8-433f-a6db-97c4773a1a10_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..23e599803ce7275de78cfef69e731df6649c1ace --- /dev/null +++ b/2023/Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning/219bbdca-cfe8-433f-a6db-97c4773a1a10_content_list.json @@ -0,0 +1,1498 @@ +[ + { + "type": "text", + "text": "Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning", + "text_level": 1, + "bbox": [ + 192, + 130, + 777, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haoyu He $^{1}$ Jianfei Cai $^{1}$ Jing Zhang $^{2}$ Dacheng Tao $^{2}$ Bohan Zhuang $^{1\\dagger}$", + "bbox": [ + 187, + 179, + 787, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 ZIP Lab, Monash University 2 The University of Sydney", + "bbox": [ + 243, + 205, + 723, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 263, + 313, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Parameter-Efficient Fine-Tuning (PEFT) has become a powerful alternative for full fine-tuning so as to adapt pre-trained vision models to downstream tasks, which only tunes a small number of parameters while freezing the vast majority ones to ease storage burden and optimization difficulty. However, existing PEFT methods introduce trainable parameters to the same positions across different tasks depending solely on human heuristics and neglect the domain gaps. To this end, we study where to introduce and how to allocate trainable parameters by proposing a novel Sensitivity-aware Visual Parameter-efficient fine-Tuning (SPT) scheme, which adaptively allocates trainable parameters to task-specific important positions given a desired tunable parameter budget. Specifically, our SPT first quickly identifies the sensitive parameters that require tuning for a given task in a data-dependent way. Next, our SPT further boosts the representational capability for the weight matrices whose number of sensitive parameters exceeds a pre-defined threshold by utilizing existing structured tuning methods, e.g., LoRA [21] or Adapter [20], to replace directly tuning the selected sensitive parameters (unstructured tuning) under the budget. Extensive experiments on a wide range of downstream recognition tasks show that our SPT is complementary to the existing PEFT methods and largely boosts their performance, e.g., SPT improves Adapter with supervised pre-trained ViT-B/16 backbone by $4.2\\%$ and $1.4\\%$ mean Top-1 accuracy, reaching SOTA performance on FGVC and VTAB-1k benchmarks, respectively. Source code is at https://github.com/ziplab/SPT.", + "bbox": [ + 76, + 295, + 473, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 776, + 209, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To effectively adapt the pre-trained representations to the downstream tasks, the de-facto choice is full fine-tuning, which initializes the model with the pre-trained weights and tunes all the parameters. However, vanilla full fine-tuning needs to store a separate instance of parameters for each", + "bbox": [ + 76, + 801, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4f12a796781166c998a97d80560c43c3045d884ad14a5de144773c620fb5fb8e.jpg", + "image_caption": [ + "Figure 1: (a) Existing PEFT methods, such as Adapter [20] introduce trainable parameters to the same positions for all downstream tasks. These methods design task-agnostic positions to employ trainable parameters relying on heuristics and neglect consideration of the distinct domain gaps and characteristics for the downstream tasks. (b) Our Sensitivity-aware visual Parameter-efficient fine-Tuning (SPT) introduces trainable parameters to the task-specific important positions and allocates them with both unstructured and structured tuning granularities, simultaneously. For structured tuning, SPT can exploit any existing structured tuning methods, such as LoRA [21] or Adapter [20]. Red lines and blocks represent trainable parameters and modules, while blue lines represent frozen parameters." + ], + "image_footnote": [], + "bbox": [ + 509, + 263, + 874, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "task and each deployment scenario. It can be extremely storage-intensive as the storage cost grows linearly with the number of possible cases, considering there are vast varieties of downstream tasks and dynamic deployment environments, especially when deploying the large vision models [12, 32, 56] to mobile systems. For example, even storing a single large pre-trained ViT-H [12] model on a local disk requires at least 2.3GB, while the Top-10 U.S. apps required only collectively 2.2GB in May 2021. $^{1}$", + "bbox": [ + 496, + 691, + 893, + 829 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Notably, an emerging solution is to replace vanilla fine-tuning with visual Parameter-Efficient Fine-Tuning (PEFT) [22, 9, 63, 23], which only tunes a small num", + "bbox": [ + 496, + 829, + 893, + 875 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author. E-mail: bohan.zhuang@gmail.com", + "bbox": [ + 94, + 886, + 419, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://sensorTower.com/blog/ios-app-size-growth-2021", + "bbox": [ + 517, + 886, + 816, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11825", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ber of trainable parameters while freezing the vast majority ones that are shared by multiple tasks. As PEFT approaches exhibit less than $1\\%$ of trainable parameters, the storage burden is largely alleviated. Another attractive property of PEFT is that tuning fewer parameters eases the optimization difficulty and mitigates the overfitting issue when adapting large pre-trained models on the target dataset, thereby achieving comparable or even better performance than vanilla fine-tuning [22]. Although promising, the existing PEFT approaches introduce trainable parameters to the same positions for all downstream tasks, relying on human heuristics and neglecting task-specific domain gaps and characteristics, which limits their performance. For instance, in a task-agnostic manner, Prompt Tuning-deep [22] and Adapter [20] respectively add trainable parameters to multi-head self-attention and feed-forward network layers for all distinct tasks as depicted in Figure 1 (a).", + "bbox": [ + 75, + 90, + 472, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this fundamental challenge, we explore where to introduce and how to allocate trainable parameters under a desired parameter budget by presenting a novel Sensitivity-aware visual Parameter-efficient fine-Tuning (SPT) scheme that identifies the task-specific important positions to adaptively allocate trainable parameters. Since the pre-trained weights at distinct positions have varying contributions for different downstream tasks [58, 26, 37], we first propose a new criterion to quickly identify the task-specific sensitive parameters that require tuning in a data-dependent way. Inspired by model pruning metrics [44, 35, 2, 3], we propose to measure the parameter sensitivity with the loss reduction when being tuned, which can be approximated by a first-order Taylor expansion derived within a single forward and backward pass ahead of fine-tuning in one-shot. Our sensitivity criterion is simple and effective, which can identify the task-specific important positions to introduce trainable parameters for any backbone quickly. For instance, calculating the sensitivity for ViT-B/16 backbone takes only 5.5 seconds with a single GPU on any of the VTAB-1k datasets.", + "bbox": [ + 75, + 351, + 472, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With our criterion, we empirically observe that the proportions of the sensitivity parameters for each block indeed vary markedly across different tasks in Section 4.4. To allocate the trainable parameters under a desired trainable parameter budge, an intuitive solution is to directly tune the most sensitive weight connections, which we name unstructured tuning. Despite its simplicity and flexibility, unstructured tuning only tunes a few parameters which still lacks representational capability and is challenging to bridge the domain gap. To this end, we propose to further incorporate structured tuning to replace unstructured tuning at the sensitive weight matrices whose numbers of sensitive parameters exceed a pre-defined threshold to improve the representational capability under a similar parameter budget. Structured tuning can be implemented by any parameter-efficient", + "bbox": [ + 75, + 674, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "structured tuning methods [21, 9, 23, 22] that directly adjust the hidden representations, e.g., inserting an adapter module sequentially after the sensitive weight matrices. Therefore, our SPT adaptively combines both unstructured and structured tuning granularity and allocates trainable parameters with high flexibility and representational capability for each distinct downstream task.", + "bbox": [ + 496, + 90, + 893, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper has the following key contributions. 1) We make the pioneering exploration to identify the task-specific important positions under the PEFT setting, which is fast, effective, versatile to be applied to various backbones with different pre-training strategies, and orthogonal to the existing PEFT methods. 2) Based on the sensitivity criterion, we propose a trainable parameter allocation strategy that adaptively combines both unstructured and structured tuning under a desired parameter budget to achieve high flexibility, large capacity, and favorable trade-off between parameter efficiency and accuracy. 3) Extensive experiments on a total of 24 downstream recognition tasks with both plain and hierarchical vision Transformer backbones under supervised and self-supervised pre-trainings show that our SPT is complementary to the existing PEFT methods and boosts their performance by large margins. For instance, SPT improves Adapter [20] by $4.2\\%$ mean Top-1 accuracy, outperforming the SOTA PEFT methods on the FGVC benchmark.", + "bbox": [ + 496, + 199, + 893, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 483, + 640, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Parameter-efficient fine-tuning. Full fine-tuning is the most predominant approach when adapting a large-scale pre-trained model to downstream tasks, where the model is initialized from the pre-trained weights with all parameters trainable. Yet, when a model becomes larger, parameter-efficient fine-tuning [27, 28] is highly desirable, which tunes only a tiny portion of parameters to alleviate the storage burden. The general PEFT approaches can be categorized into addition-based PEFT methods and reparameterization-based PEFT methods.", + "bbox": [ + 496, + 506, + 890, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Addition-based PEFT attaches additional trainable parameters to the backbone and only tunes these parameters. Apart from Prompt tuning [22] and Adapter [20], recent addition-based methods study connecting or combining existing PEFT methods. For instance, He et al. [18] connect Prompt tuning and Adapter and provide a unified view that all PEFT approaches share the same design to adjust the hidden representations. Zhang et al. [63] search for the optimal configurations to combine multiple PEFT approaches following once-for-all scheme [5, 54]. Since the additional parameters require extra computations compared to full fine-tuning, a few recent works [46, 48] design specific architectures to avoid storing the intermediate activations, thereby alleviating the fine-tuning memory cost. However, it is noteworthy that enhancing training efficiency is not the primary objective of our work.", + "bbox": [ + 496, + 659, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11826", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reparameterization-based PEFT aims to avoid extra computational costs by tuning parameters that are inherently in or can be reparameterized into the backbone during inference. Prior works select the parameters that are inherently in the backbone, including the bias terms [59], the last several layers [58, 4], and weight connections [14, 64]. To reparameterize new parameters into the backbone [17, 29], representative work LoRA [21] optimizes two low-rank matrices which can be further merged into the weight matrices. In contrast to the aforementioned works, we argue the importance of tuning parameters at task-specific important positions and quickly identify them with our proposed parameter sensitivity criterion before tuning, which is complementary to and provides valuable guidance for the existing PEFT methods. Moreover, our SPT can also be inference-efficient when implementing structured tuning with any reparameterization-based structured tuning method. Recently, SSF [29] is proposed to introduce trainable scaling and shifting parameters that can be absorbed into the previous linear layers. However, it cannot scale to higher trainable parameter budgets and requires a complex and time-consuming hyper-parameter search for learning rate, weight decay, and drop-path rate on each individual dataset, thus is not directly comparable to our method.", + "bbox": [ + 76, + 90, + 472, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task-specific transfer learning. The effectiveness of transferring pre-trained models to downstream tasks strongly depends on the relationship between the source and target tasks [43, 53, 26, 40]. This has motivated the community to explore the optimal pre-training data [11, 57], model [47, 38], and weights [16, 55] for the target task. To seek suitable task-specific pre-training data, Cui et al. [11] select the source domain data from the top-k most similar classes measured by Earth Mover's Distance; Yoon et al. [57] weight each class in the source domain with reinforcement learning; and Puigcerver et al. [41] first train a diverse set of experts and then select the most relevant expert for each target task. Another line of work selects a suitable pre-trained model for the target task ahead of fine-tuning by measuring the transferability of pre-trained models to the target domain with interclass covariance between the source data and target classes [1] or conditional cross-entropy [47] between the source and target labels. Considering the transferability of the feature representations at distinct layers for the same pre-trained model is different [58, 37], recent works [15, 45] endeavour transfer task-specific weights by freezing some pre-trained weights and fine-tuning the rest. For example, the task-specific fine-tuned weights are selected by learning a policy network with Gumbel-Softmax [16], optimizing a sparse mask with $L_{0}$ norm [14], and learning binary gates for each parameter [64]. Our SPT also adaptively selects task-specific parameters. In contrast to the previous work, we 1) derive task-specific important positions prior to fine-tuning with only a single", + "bbox": [ + 76, + 463, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "forward and backward pass, which is computationally efficient; 2) mask the gradients for insensitive parameters in unstructured tuning with fixed binary masks, thereby having more affordable fine-tuning memory than optimizing learnable binary masks in [14, 64]. Moreover, we are pioneering work to adaptively allocate task-specific trainable parameters with both fine-grained unstructured and coarse-grained structured tuning granularities to achieve both high flexibility and representational capability.", + "bbox": [ + 496, + 90, + 893, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 498, + 239, + 591, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our sensitivity-aware visual parameter-efficient fine-tuning consists of two stages. In the first stage, SPT measures the task-specific sensitivity for the pre-trained parameters (Section 3.1). Based on the parameter sensitivity and a given parameter budget, SPT then adaptively allocates trainable parameters to task-specific important positions (Section 3.2).", + "bbox": [ + 496, + 263, + 893, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Task-specific Parameter Sensitivity", + "text_level": 1, + "bbox": [ + 498, + 377, + 805, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent research has observed that pre-trained backbone parameters exhibit varying feature patterns [42, 36] and criticality [62, 8] at distinct positions. Moreover, when transferred to downstream tasks, their efficacy varies depending on how much pre-trained features are reused and how well they adapt to the specific domain gap [58, 26, 37]. Motivated by these observations, we argue that not all parameters contribute equally to the performance across different tasks in PEFT and propose a new criterion to measure the sensitivity of the parameters in the pre-trained backbone for a given task.", + "bbox": [ + 496, + 398, + 893, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, given the training dataset $\\mathcal{D}_t$ for the $t$ -th task and the pre-trained model weights $\\boldsymbol{w} = \\{w_1, w_2, \\dots, w_N\\} \\in \\mathbb{R}^N$ where $N$ is the total number of parameters, the objective for the task is to minimize the empirical risk: $\\min_{\\boldsymbol{w}} E(\\mathcal{D}_t, \\boldsymbol{w})$ . We denote the parameter sensitivity set as $\\mathcal{S} = \\{s_1, \\dots, s_N\\}$ and the sensitivity $s_n$ for parameter $w_n$ is measured by the empirical risk difference when tuning it:", + "bbox": [ + 496, + 565, + 893, + 686 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns _ {n} = E \\left(\\mathcal {D} _ {t}, \\boldsymbol {w}\\right) - E \\left(\\mathcal {D} _ {t}, \\boldsymbol {w} \\mid w _ {n} = w _ {n} ^ {*}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 700, + 890, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $w_{n}^{*} = \\operatorname*{argmin}_{w_{n}}(E(\\mathcal{D}_{t},\\boldsymbol {w}))$ . We can reparameterize the tuned parameters as $w_{n}^{*} = w_{n} + \\Delta_{w_{n}}$ , where $\\Delta_{w_n}$ denotes the update for $w_{n}$ after tuning. Here we individually measure the sensitivity of each parameter, which is reasonable given that most of the parameters are frozen during fine-tuning in PEFT. However, it is still computationally intensive to compute Eq. (1) for two reasons. Firstly, getting the empirical risk for $N$ parameters requires forwarding the entire network $N$ times, which is time-consuming. Secondly, it is challenging to derive $\\Delta_{w_n}$ , as we have to tune each individual $w_{n}$ until convergence.", + "bbox": [ + 496, + 727, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11827", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e0de3fc7f9016d1bcc965bd66924f89237eb5daa10b3805b7d071e13ff65f082.jpg", + "image_caption": [ + "Figure 2: Overview of our trainable parameter allocation strategy. With the parameter sensitivity set $S$ , we first get the top- $\\tau$ sensitive parameters. Instead of directly tuning these sensitive parameters, we also boost the representational capability by replacing unstructured tuning with structured tuning at sensitive weight matrices that have a large number of sensitive parameters, which can be implemented by an existing structured tuning method, e.g., LoRA [21] and Adapter [20]. Red lines and blocks represent trainable parameters and modules, while blue lines represent frozen parameters." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 883, + 157 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Computing task-specific parameter sensitivities" + ], + "code_body": "Input: Pre-trained model with network parameters $\\pmb{w}$ training set $\\mathcal{D}_t$ for the $t$ -th task, and number of training samples $C$ used to calculate the parameter sensitivities \nOutput: Sensitivity set $\\mathcal{S} = \\{s_1,\\dots,s_N\\}$ \nInitialize $\\mathcal{S} = \\{0\\}^N$ \nfor $i\\in \\{1,\\ldots ,C\\}$ do \n Get the $i$ -th training sample of $\\mathcal{D}_t$ \n Compute loss $E$ \n Compute gradients $\\pmb{g}$ \n for $n\\in \\{1,\\dots,N\\}$ do \n Update sensitivity for the $n$ -th parameter: $s_n = s_n + g_n^2$ \n end for \nend for", + "bbox": [ + 89, + 273, + 472, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To overcome the first barrier, we simplify the empirical loss by approximating $s_n$ in the vicinity of $w$ by its first-order Taylor expansion", + "bbox": [ + 75, + 511, + 468, + 556 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {n} ^ {(1)} = - g _ {n} \\Delta_ {w _ {n}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 561, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the gradients $\\pmb{g} = \\partial E / \\partial \\pmb{w}$ , and $g_{n}$ is the gradient of the $n$ -th element of $\\pmb{g}$ . To address the second barrier, following [31, 7], we take the one-step unrolled weight as the surrogate for $w_{n}^{*}$ and approximate $\\Delta_{w_n}$ in Eq. (2) with a single step of gradient descent. We can accordingly get $s_n^{(1)}\\approx g_n^2\\epsilon$ , where $\\epsilon$ is the learning rate. Since $\\epsilon$ is the same for all parameters, we can eliminate it when comparing the sensitivity with the other parameters and finally get", + "bbox": [ + 75, + 583, + 468, + 707 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {n} ^ {(1)} \\approx g _ {n} ^ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 709, + 468, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, the sensitivity of a parameter can be efficiently measured by its potential to reduce the loss on the target domain. Note that although our criterion draws inspiration from pruning work [35], it is distinct from it. [35] measures the parameter importance by the squared change in loss when removing them, i.e., $(E(\\mathcal{D}_t,\\boldsymbol {w}) - E(\\mathcal{D}_t,\\boldsymbol {w}\\mid w_n = 0))^2$ and finally derives the parameter importance by $(g_{n}w_{n})^{2}$ , which is different from our formulations in Eqs. (1) and (3).", + "bbox": [ + 75, + 732, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, we accumulate $\\mathcal{S}$ from a total number of $C$ training samples ahead of fine-tuning to generate accu-", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "rate sensitivity as shown in Algorithm 1, where $C$ is a predefined hyper-parameter. In Section 4.3, we show that employing only 400 training samples is sufficient for getting reasonable parameter sensitivity, which requires only 5.5 seconds with a single GPU for any VTAB-1k dataset with ViT-B/16 backbone [12].", + "bbox": [ + 496, + 241, + 893, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Adaptive Trainable Parameters Allocation", + "text_level": 1, + "bbox": [ + 498, + 343, + 861, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our next step is to allocate trainable parameters based on the obtained parameter sensitivity set $S$ and a desired parameter budget $\\tau$ . A straightforward solution is to directly tune the top- $\\tau$ most sensitive unstructured connections (parameters) while keeping the rest frozen, which we name unstructured tuning. Specifically, we select the top- $\\tau$ most sensitive weight connections in $S$ to form the sensitive weight connection set $\\mathcal{T}$ . Then, for a weight matrix $\\boldsymbol{W} \\in \\mathbb{R}^{d_{\\mathrm{in}} \\times d_{\\mathrm{out}}}$ , we can get a binary mask $M \\in \\mathbb{R}^{d_{\\mathrm{in}} \\times d_{\\mathrm{out}}}$ computed by", + "bbox": [ + 496, + 366, + 890, + 518 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {M} ^ {j} = \\left\\{ \\begin{array}{l l} 1 & \\boldsymbol {W} ^ {j} \\in \\mathcal {T} \\\\ 0 & \\boldsymbol {W} ^ {j} \\notin \\mathcal {T} \\end{array} , \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 526, + 890, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W^{j}$ and $M^j$ are the $j$ -th element in $W$ and $M$ , respectively. Accordingly, we can train the sensitive parameters by gradient descent and the updated weight matrix can be formulated as $W' \\gets W - \\epsilon g_W \\odot M$ , where $g_W$ is the gradient for $W$ .", + "bbox": [ + 496, + 565, + 890, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, considering PEFT approaches generally limit the proportion of trainable parameters to less than $1\\%$ , tuning only a small number of unstructured weight connections might not have enough representational capability to handle the downstream datasets with large domain gaps from the source pre-training data. Therefore, to improve the representational capability, we propose to replace unstructured tuning with structured tuning at the sensitive weight matrices that have a high number of sensitive parameters. To preserve the parameter budget, we can implement structured tuning with an existing efficient structured tuning PEFT method [21, 9, 20, 23] that learns to directly adjust all hidden dimensions at once. We depict an overview of our trainable parameter allocation strategy in Figure 2. For example, we can employ the low-rank reparameterization trick LoRA [21] to the sensitive weight matrices and the one-step update for $\\mathbf{W}$ can be formulated as", + "bbox": [ + 496, + 643, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11828", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {W} ^ {\\prime} = \\left\\{ \\begin{array}{l l} \\boldsymbol {W} + \\boldsymbol {W} _ {\\mathrm {d o w n}} \\boldsymbol {W} _ {\\mathrm {u p}} & \\text {i f} \\sum_ {j = 0} ^ {d _ {\\mathrm {i n}} \\times d _ {\\mathrm {o u t}}} \\boldsymbol {M} ^ {j} \\geq \\sigma_ {\\mathrm {o p t}} \\\\ \\boldsymbol {W} - \\epsilon \\boldsymbol {g} _ {\\boldsymbol {W}} \\odot \\boldsymbol {M} & \\text {o t h e r w i s e} \\end{array} \\right.,\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 88, + 470, + 118 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{\\mathrm{down}} \\in \\mathbb{R}^{d_{\\mathrm{in}} \\times r}$ and $W_{\\mathrm{up}} \\in \\mathbb{R}^{r \\times d_{\\mathrm{out}}}$ are two learnable low-rank matrices to approximate the update of $W$ and rank $r$ is a hyper-parameter where $r \\ll \\min(d_{\\mathrm{in}}, d_{\\mathrm{out}})$ . In this way, we perform structured tuning on $W$ when its number of sensitive parameters exceeds $\\sigma_{\\mathrm{opt}}$ , whose value depends on the pre-defined type of structured tuning method. For example, since implementing structured tuning with LoRA requires $2 \\times d_{\\mathrm{in}} \\times d_{\\mathrm{out}} \\times r$ trainable parameters for each sensitive weight matrix, we set $\\sigma_{\\mathrm{LoRA}} \\gets 2 \\times d_{\\mathrm{in}} \\times d_{\\mathrm{out}} \\times r$ to ensure that the number of trainable parameters introduced by structured tuning is always equal to or lower than the number of sensitive parameters.", + "bbox": [ + 76, + 119, + 468, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, our SPT adaptively incorporates both structured and unstructured tuning granularities to enable higher flexibility and stronger representational power, simultaneously. In Section 4.3, we show that structured tuning is important for the downstream tasks with larger domain gaps and both unstructured and structured tuning contribute clearly to the superior performance of our SPT.", + "bbox": [ + 76, + 310, + 468, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 430, + 209, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 454, + 266, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and metrics. We evaluate our SPT on total 24 downstream tasks in two groups following [22]. 1) FGVC is a benchmark for fine-grained visual classification, including CUB-200-2011 [52], NABirds [50], Oxford Flowers [39], Stanford Cars [13], and Stanford Dogs [24] datasets. Each FGVC dataset contains between 55 to 200 classes and a few thousand images for train, validation, and test. We follow the validation splits in [22] if the validation set is unavailable. 2) VTAB-1k [60] is a large-scale transfer learning benchmark consisting of a collection of 19 visual classification tasks. VTAB-1k can further be divided into three groups, including Natural tasks with natural images, Specialized tasks with images captured by specialized equipment, e.g., medical images, and Structured tasks with images mostly generated from synthetic environments. Each of the VTAB-1k dataset has only 800 training and 200 validation samples, while the test set sizes vary. We use top-1 accuracy (\\%) averaged within each group as our main metric following [22].", + "bbox": [ + 75, + 474, + 468, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-trained backbones. We conduct experiments on the plain vision Transformer backbone ViT-B/16 [12] that is pre-trained on ImageNet [25] with different pre-training strategies following [22], including supervised pre-training and self-supervised pre-training with MAE [19] and MoCo v3 [10] following [22]. We also conduct experiments on the representative hierarchical vision Transformer backbone Swin-B [32] under supervised pre-training.", + "bbox": [ + 75, + 763, + 468, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Contenders. We categorize the baseline methods into", + "bbox": [ + 76, + 885, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "addition-based and reparameterization-based PEFT methods as introduced in Section 2. Unless specified, all baseline methods keep the backbone frozen. Addition-based methods require extra computations during inference, including MLP- $k$ , PROMPT-SHALLOW [22], PROMPT-DEEP [22], ADAPTER- $k$ [20], ADAPTFORMER [9], and NOAH [63]. Reparameterization-based methods have no additional computational overhead during inference, including LINEAR, PARTIAL- $k$ , BIAS [59], and LORA- $k$ [21]. Here $k$ represents the number of bottleneck dimension in ADAPTER- $k$ and LORA- $k$ . We also compare with full fine-tuning which is denoted by FULL. We introduce the details of these methods in the supplementary material.", + "bbox": [ + 496, + 90, + 890, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also introduce two variants of our SPT: addition-based SPT-ADAPTER and reparameterization-based SPT-LORA. SPT-ADAPTER directly adjusts the hidden representations that are computed by sensitive weight matrices following [20], while SPT-LORA approximates updating the sensitive weight matrices following [21]. For the two variants, we follow the exact weight initializations that are described in [21] and follow [63] to set the bottleneck dimension as 8.", + "bbox": [ + 496, + 287, + 890, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. Following [63], we use the AdamW optimizer [34] with cosine learning rate decay and set the batch size, learning rate, and weight decay as 64, $1 \\times 10^{-3}$ , and $1 \\times 10^{-4}$ , respectively. We also follow [63] for the standard data augmentation pipeline. We set the number of training samples $C$ used to calculate our parameter sensitivities in Algorithm 1 universally to be 800 for the main experiments.", + "bbox": [ + 496, + 422, + 890, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 500, + 547, + 637, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our method with the baseline methods under different backbones, pre-training strategies, and tasks.", + "bbox": [ + 496, + 568, + 890, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visual recognition with ViT backbone. First, our proposed SPT-ADAPTER and SPT-LORA achieve the best performance under different trainable parameter budgets with supervised pre-trained ViT-B/16 backbone, as shown in Table 1 and Figure 3 (a). For instance, SPT-ADAPTER outperforms the SOTA method NOAA by a clear margin of $0.9\\%$ mean top-1 accuracy over the 19 VTAB-1k datasets with fewer trainable parameters. We speculate that our SPT variants allocate trainable parameters at task-specific positions compared to the heuristically selected positions in the baseline methods, which contributes to our superior performance. We also observe that our SPT-ADAPTER and SPT-LORA achieve large performance gains over ADAPTER and LORA variants, respectively. For example, SPT-ADAPTER and SPT-LORA with $0.41\\%$ trainable parameters respectively improve ADAPTER-8 and LORA-8 significantly by $4.0\\%$ and $3.3\\%$ mean accuracy on the FGVC benchmark. This suggests that identifying task-specific important positions and combining both unstructured and structured tuning granularities with SPT are complementary to the exist", + "bbox": [ + 496, + 599, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11829", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/52e9e580d99653c5874b355ff86b5663ee64d49f6beb7d2fd69a3b6fdc870e32.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ViT-B/16 (85.8M)Total paramsFGVCVTAB-1k
Tuned / TotalMean Acc.Tuned / TotalNaturalSpecializedStructuredMean Acc.
FULL24.02×100%88.5100%75.983.447.669.0
", + "bbox": [ + 156, + 88, + 810, + 141 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/027bc3a87db41efdad7e554ff6c9f943d205363a0eef2a4dbc7a498b81269b1e.jpg", + "table_caption": [ + "Addition-based methods" + ], + "table_footnote": [], + "table_body": "
MLP-31.35×1.50%79.81.42%67.872.830.657.1
PROMPT-SHALLOW1.04×0.31%84.60.13%76.879.747.067.8
PROMPT-DEEP1.18×0.98%89.11.14%78.582.455.072.0
ADAPTER-81.06×0.39%85.50.23%79.084.158.573.9
ADAPTER-321.19×0.95%85.60.71%79.684.058.374.0
ADAPTFORMER1.09×0.44%85.10.36%80.685.458.574.8
NOAH---0.52%80.284.961.375.5
SPT-ADAPTER (Ours)1.08×0.41%89.50.30%81.385.360.875.8
SPT-ADAPTER (Ours)1.10×0.47%89.80.44%82.085.861.476.4
", + "bbox": [ + 156, + 157, + 808, + 280 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d881f765a91e273c18d5a93c0f92754432bc6a187c39c349e3fd2c0246f8160e.jpg", + "table_caption": [ + "Reparameterization-based methods" + ], + "table_footnote": [], + "table_body": "
LINEAR1.02×0.12%79.30.04%68.977.226.857.6
PARTIAL-13.00×8.38%82.68.30%69.478.534.260.7
BIAS1.05×0.13%88.40.13%73.378.344.165.2
LORA-81.07×0.55%86.00.23%79.584.660.574.9
LORA-161.18×0.90%84.80.69%79.884.960.275.0
SPT-LORA (Ours)1.08×0.41%89.30.31%81.585.660.775.9
SPT-LORA (Ours)1.15×0.60%90.10.63%81.985.961.376.4
", + "bbox": [ + 158, + 296, + 807, + 393 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/73d02b8ffe5baffa0fe34b679d257b37d6a48a5f316f75779d64204e3bb772e9.jpg", + "table_caption": [ + "Table 1: Comparisons on FGVC and VTAB-1k [60] benchmarks using supervised pre-trained ViT-B/16 backbone pre-trained on ImageNet-21k. \"Total params\" denotes the ratio of the total number of parameters needed for all downstream tasks relative to the one for the pre-trained backbone, and \"Tuned/Total\" denotes the fraction of trainable parameters. Top-1 accuracy (\\%) is reported. The best result is in bold, and the second-best result is underlined." + ], + "table_footnote": [], + "table_body": "
ViT-B/16 (85.8M)Total ParamsVTAB-1kMAEMean Acc.VTAB-1kMoCo v3Mean Acc.
Tuned / TotalNaturalSpecializedStructuredTuned / TotalNaturalSpecializedStructured
FULL38.02×100%59.379.753.864.3100%72.084.742.069.6
", + "bbox": [ + 76, + 468, + 890, + 518 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/018e6081b6366526f43b03ab9255167db2f7df6876dd54c0bb257b71465bfd61.jpg", + "table_caption": [ + "Addition-based methods" + ], + "table_footnote": [], + "table_body": "
ADAPTER-81.08×0.23%57.278.454.763.40.23%27.670.948.449.0
ADAPTER-321.28×0.95%55.378.853.362.50.99%74.282.747.768.2
PROMPT-SHALLOW1.02×0.12%40.069.727.545.70.12%67.382.337.662.4
PROMPT-DEEP1.05×0.23%36.060.626.641.10.07%70.383.042.465.2
SPT-ADAPTER (Ours)1.07×0.26%64.882.460.469.20.08%76.184.960.173.7
SPT-ADAPTER (Ours)1.13×0.41%65.682.760.769.70.30%76.685.061.774.4
", + "bbox": [ + 78, + 535, + 890, + 619 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d301075368587a57e1ed6afb5b4198374554617065f022187e3aa93a37c8cc55.jpg", + "table_caption": [ + "Reparameterization-based methods" + ], + "table_footnote": [], + "table_body": "
LINEAR1.02×0.04%18.952.723.732.10.04%67.581.130.359.6
PARTIAL-14.16×8.30%58.478.347.661.58.30%72.384.647.968.3
BIAS1.06×0.13%54.675.747.759.30.13%72.981.153.469.2
LORA-81.08×0.23%57.577.757.764.30.23%21.266.745.144.3
LORA-161.28×0.69%57.377.159.964.80.69%16.064.048.742.9
SPT-LORA (Ours)1.11×0.29%63.881.660.068.50.30%76.585.463.075.0
SPT-LORA (Ours)1.23×0.69%65.482.461.569.80.50%76.586.063.675.3
", + "bbox": [ + 78, + 636, + 890, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2: Comparisons on VTAB-1k [60] benchmark using self-supervised ViT-B/16 backbone pre-trained by MAE [19] and MoCo v3 [10]. \"Total params\" denotes the ratio of the total number of parameters needed for all downstream tasks relative to the one for the pre-trained backbone, and \"Tuned/Total\" denotes the fraction of trainable parameters. Top-1 accuracy (\\%) is reported. The best result is in bold, and the second-best result is underlined.", + "bbox": [ + 75, + 738, + 893, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing PEFT methods and boost their performance.", + "bbox": [ + 75, + 806, + 393, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Second, SPT variants outperform baseline methods and full fine-tuning by significant margins with the self-supervised pre-trained ViT-B/16 backbones. As shown in Table 2, existing PEFT approaches exhibit inferior results than full fine-tuning with the self-supervised pre-trained", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "backbones MAE and MoCo v3. It is worth noting that previous PEFT methods yield inconsistent results with the backbones of different pre-training strategies. In contrast, SPT variants consistently outperform full fine-tuning. In particular, SPT-ADAPTER achieves remarkable $5.8\\%$ and $5.5\\%$ mean top-1 accuracy gains over the best-performing", + "bbox": [ + 496, + 806, + 893, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11830", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b83b221b6a8712fcf9845dcf61b273fa9ff05a9477f28e352e9221fec82a7972.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTuned / TotalNaturalSpecializedStructuredMean / Acc.
FULL100%79.186.259.775.0
Addition-based methods
MLP-31.60%73.675.235.761.5
PROMPT-SHALLOW0.04%79.982.537.866.7
PROMPT-DEEP0.23%76.884.553.471.6
ADAPTER-81.18%81.787.361.276.7
SPT-ADAPTER (ours)0.33%83.087.362.177.5
Reparameterization-based methods
LINEAR0.04%73.580.833.562.6
PARTIAL-12.15%73.181.735.063.3
LORA-81.18%81.787.260.176.3
SPT-LORA (ours)0.49%83.187.460.477.2
", + "bbox": [ + 78, + 88, + 468, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "baseline method on VTAB-1k benchmark with only $0.26\\%$ and $0.08\\%$ trainable parameters for MAE and MoCo v3 pretrained backbones, respectively. Moreover, our observation in supplementary material suggests that self-supervised pretrained ViT backbones have more diverse sensitivity distributions and a higher variance in sensitivity across different tasks than the supervised pre-trained one. This leads to the conjecture that baseline methods which assign trainable parameters to the same positions for all tasks may fail to mitigate the distinct domain gaps in individual downstream datasets, whereas our SPT allocates trainable parameters to task-specific positions accurately.", + "bbox": [ + 75, + 334, + 468, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual recognition with Swin and ConvNeXt backbones. From Table 3, we observe that our SPT-LORA and SPT-ADAPTER also achieve SOTA performance with Swin-B backbone on all dataset groups. We also follow VPT [22] to apply SPT variants to ResNet-alike architecture ConvNeXt-Base [33] and report the results in Table 4. We observe that SPT variants achieve better trade-offs between accuracy and parameter efficiency than the baseline methods. The results further demonstrate the versatility and effectiveness of our SPT.", + "bbox": [ + 75, + 517, + 468, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic segmentation. We follow VPT [22] to conduct semantic segmentation on ADE20k dataset. Following the settings of [22], we apply our SPT to SETR-PUP [65] framework with ViT-L backbone and only allocate trainable parameters to the backbone with the head fully finetuned. We report the mIoU results in Table 5. Notably, SPT-LORA and SPT-ADAPTER outperform the baseline methods by large margins, indicating our SPT can be generalized to the semantic segmentation task. We report the results for conducting only structured tuning for both SPT-LORA and SPT-ADAPTER as it yields higher mIoU.", + "bbox": [ + 75, + 670, + 470, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 848, + 230, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of the sensitivity criterion. We investigate the effectiveness of our sensitivity criterion on VTAB-1k by em", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c13cb661bdeb2528c91830e3547653868a6458a6eceb6dc87c0a6f32b4447eab.jpg", + "table_caption": [ + "Table 3: Comparisons on VTAB-1k [60] benchmark with supervised pre-trained Swin-B [32]. \"Tuned/Total\" denotes the fraction of trainable parameters. Top-1 accuracy (\\%) is reported. The best result is in bold." + ], + "table_footnote": [], + "table_body": "
MethodTuned/TotalNaturalSpecializedStructuredMean Acc.
FULL100%78.083.760.474.0
LoRA0.79%82.284.764.177.0
Adapter0.47%83.184.964.677.5
SPT-LoRA (ours)0.57%83.486.765.978.7
SPT-Adapter (ours)0.36%83.786.265.378.4
", + "bbox": [ + 501, + 88, + 890, + 161 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a3b9737a6fd48c3da997bede7dad193f87d56b76c6275e54758785bc5ff94ef2.jpg", + "table_caption": [ + "Table 4: Comparisons on VTAB-1k [60] benchmark with supervised pre-trained ConvNeXt-Base [33] backbone. \"Tuned/Total\" denotes the fraction of trainable parameters. Top-1 accuracy $(\\%)$ is reported. The best result is in bold." + ], + "table_footnote": [], + "table_body": "
MethodmIoU-s.s. (%)mIoU-m.s. (%)Trainable Param.
VPT44.045.615.8M
LoRA43.945.914.7M
Adapter44.446.614.6M
SPT-LoRA (ours)45.447.514.6M
SPT-Adapter (ours)45.247.214.6M
", + "bbox": [ + 501, + 247, + 890, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5: Semantic Segmentation: Comparisons on ADE20k [66] val with SETR [65] on ViT-L backbone. We report both single-scale (s.s.) and multi-scale (m.s.) mIoU results. The best result is in bold.", + "bbox": [ + 498, + 342, + 890, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ploying structured tuning methods from [22, 20, 21] to the task-specific sensitive weight matrices. Note that we do not conduct unstructured tuning to ensure fair comparisons. The results are presented in Figure 3 (b). Our criterion brings consistent $1.1\\%$ , $1.6\\%$ , and $0.8\\%$ performance gains for PROMPT-DEEP, ADAPTER-32, and LORA-16, respectively, which demonstrates the effectiveness and versatility of our sensitivity criterion to identify accurate task-specific important positions.", + "bbox": [ + 496, + 431, + 890, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of structured and unstructured tuning. We investigate the effectiveness of unstructured and structured tuning individually on VTAB-1k. The results are presented in Table 6. We start by applying ADAPTER-8 to the sensitive weight matrices identified by our sensitivity criterion (SPT-ADAPTER w/o unstructured). We observe that our sensitivity criterion boosts the performance of all the dataset groups by clear margins, which again demonstrates the importance of our sensitivity criterion. Next, we observe that allocating the trainable parameters to the unstructured sensitive weight connections also brings accuracy improvement to the Natural and Specialized datasets from ADAPTER-8. However, we find that structured tuning is especially important for achieving good performance on Structured datasets. To further investigate this phenomenon, we observe that Structured datasets have larger domain gaps from the pre-training source domain [25] compared to Natural and Specialized datasets as visualized in Figure 3 (c). We hence conjecture that structured tuning has a higher representational capability than unstructured tuning which facilitates mitigating the large domain gaps during fine-tuning (see the supplementary material for visual examples). Finally, we observe", + "bbox": [ + 496, + 568, + 892, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11831", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fecd4559bdd11b1405641730406683e2e4347922e441880fffb05fb5ddd1f62d.jpg", + "image_caption": [ + "Figure 3: (a) Accuracy vs. parameter efficiency with supervised pre-trained ViT-B/16 backbone on VTAB-1k [60]. SPT variants perform favorably against the other PEFT approaches and are more scalable. (b) Applying other PEFT structured tuning methods [22, 20, 21] to the task-specific sensitive weight matrices (denoted by TSM) identified by our criterion with supervised pre-trained ViT-B/16 backbone on VTAB-1k. Our criterion brings consistent performance gain. (c) Domain vs. performance gaps for different dataset groups in VTAB-1k [60]. The blue bars show the domain gaps between the source domain (ImageNet [25]) and target domains, which are measured by Maximum Mean Discrepancy (MMD) distance [49]. The red line represents the performance gaps between SPT-ADAPTER w/o unstructured and w/o structured, using supervised pre-trained ViT-B/16 backbone. The dataset groups are Natural, Specialized, and Structured. Structured tuning is important for achieving good performance on Structured datasets with larger domain gaps." + ], + "image_footnote": [], + "bbox": [ + 80, + 90, + 336, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/91c0b01112e12170c50af13790a69d271301b7040b4ac17299ae9d52f3719831.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 90, + 602, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c46225b41d6523ea95dde04abeda37757627658fc77b5ac4db421db94dfe578c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 95, + 890, + 250 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/37874411b5d15c1141282c71b549b1cdae33c9c9a8bb12c7887a4c66e5c92973.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTuned / TotalNaturalSpecializedStructuredMean / Acc.
ADAPTER-80.23%79.084.158.573.9
SPT-ADAPTER w/o unstructured0.29%81.285.160.375.5
SPT-ADAPTER w/o structured0.34%81.285.059.675.3
SPT-ADAPTER0.30%81.385.360.875.8
", + "bbox": [ + 76, + 401, + 468, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that incorporating both structured and unstructured tuning at task-specific important positions achieves the highest performance on all dataset groups.", + "bbox": [ + 75, + 566, + 468, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of number of training samples $C$ to get parameter sensitivity. We investigate the effect of the number of training images $C$ for calculating our parameter sensitivity (Algorithm 1 of the main paper). We randomly sample training samples and report the mean results over three runs in Table 8. We find that our SPT is robust to the number of training samples $C$ and randomly sampling 400 out of a total of 800 training samples is sufficient to obtain accurate task-specific important positions, e.g., calculating the sensitivity for ViT-B/16 backbone takes only 5.5 seconds with a single GPU on any of the VTAB-1k datasets and this computation is required only once.", + "bbox": [ + 75, + 614, + 468, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computational cost analysis. We investigate the computational cost of SPT-LoRA by comparing with full fine-tuning, addition-based method PROMPT-DEEP, and reparameterization-based method LORA-16. The results are presented in Table 7. We observe that PROMPT-DEEP has higher inference latency and inference GPU memory due to the additional prompts. In contrast, since the up", + "bbox": [ + 75, + 795, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4c158c3dd426cd4e6175ebe86db26ad61e2f63309b2d28e9e2df03a67bfdbd2a.jpg", + "table_caption": [ + "Table 6: Ablation study on structured and unstructured tuning only with supervised pre-trained ViT-B/16 backbone. Top-1 accuracy $(\\%)$ is reported. We set different parameter constraints to align the fractions of trainable parameters for these cases. The best result is in bold" + ], + "table_footnote": [], + "table_body": "
MethodInference Latency (ms/img)Inference Memory (GB)Fine-tuning Memory (GB)
FULL2.81.311.9
PROMPT-DEEP3.81.913.2
LORA-162.81.38.2
SPT-LORA w/o unstructured2.81.38.3
SPT-LORA2.81.39.8
", + "bbox": [ + 500, + 401, + 890, + 472 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/48c2a6d692fd9e18b59e5986b829ab8e252bdf94d9fb54dd7226cede7ae38e48.jpg", + "table_caption": [ + "Table 7: Cost comparisons with ViT-B/16 backbone on the VTAB-1k [60] benchmark with around $0.70\\%$ fractions of the trainable parameters. We report the latency (ms/img) and the peak memory usage (GB) with image resolution $224 \\times 224$ on a single GeForce 3090 GPU. The best result is in bold." + ], + "table_footnote": [], + "table_body": "
C240400560800
Mean Acc.76.376.476.476.4
", + "bbox": [ + 576, + 590, + 813, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8: Effect of the number of training samples used to get the sensitivity for SPT-LoRA with supervised pre-trained ViT-B/16 backbone on VTAB-1k [60]. Top-1 accuracy (%) is reported. The best result is in bold.", + "bbox": [ + 498, + 636, + 890, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dated parameters after fine-tuning can be reparameterized and merged into the pre-trained model, our SPT-LoRA, SPT-LoRA w/o unstructured tuning, and LoRA-16 are more efficient than PROMPT-DEEP during inference. However, we observe that our SPT-LoRA has slightly higher fine-tuning memory than the full fine-tuning and LoRA-16 which is taken up by updating the unstructurally-tuned parameters with sparse gradients in Eq. (5). Additionally, for memory-intensive scenarios, one can employ SPT-LoRA w/o unstructured tuning for improved performance (0.8% Top-1 accuracy on VTAB-1k) and similar fine-tuning memory as LoRA, as shown in Figure 3 (b) and Table 6.", + "bbox": [ + 496, + 720, + 892, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11832", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dcb4479f9e36da7a176aef5420a12aa733401cc06b89e1800dde4aa6e7a7cd66.jpg", + "image_caption": [ + "Supervised pre-trained ViT-B" + ], + "image_footnote": [], + "bbox": [ + 84, + 108, + 336, + 253 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a20dc078f3323546bbe037059cb0ffe8c468d3fe89167440b381dad03f292337.jpg", + "image_caption": [ + "MAE pre-trained ViT-B" + ], + "image_footnote": [], + "bbox": [ + 356, + 108, + 607, + 253 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3c4674f60c4f5db37bf71b1b05fa21a7117f87e0d605b4e8e426893a45353170.jpg", + "image_caption": [ + "MoCo V3 pre-trained ViT-B" + ], + "image_footnote": [], + "bbox": [ + 625, + 108, + 883, + 253 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/897aed02949367588ae110b14be073a14162de7848bdd9c4054d053132b051a3.jpg", + "image_caption": [ + "Figure 4: Parameter sensitivity patterns under 0.4M trainable parameter budget for ViT-B/16 backbone with different pretraining strategies on three sample tasks from VTAB-1k [61]. The proportions exhibit task-specific varying patterns in terms of network depth (upper figures) and task-agnostic similar patterns in terms of operations (lower figures)." + ], + "image_footnote": [], + "bbox": [ + 84, + 262, + 336, + 407 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e3d68d8f8babb9ecc5bc562070062d1af7b5d1c01bb7fd962d3bc3478b7fbadb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 262, + 607, + 407 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9aaa447abff829082dd93d1feb56772225b8dc2d55604cf3d85e4945eeeb6364.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 260, + 883, + 409 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4. Observations on Sensitivity Patterns", + "text_level": 1, + "bbox": [ + 76, + 472, + 390, + 488 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our sensitivity criterion identifies task-specific important positions, which can reveal the contributions of the pretrained weights to different downstream tasks during transfer learning. We visualize the proportions of the sensitive parameters for the supervised pre-trained ViT-B/16 backbone under 0.4M trainable parameter budget in Figure 4. First, we investigate the most sensitive blocks, whose numbers of sensitive parameters are summed and normalized over the 12 ViT-B/16 blocks. We observe that the patterns of the sensitive parameter proportions vary markedly across different tasks, which echoes the observations made in [16]. This suggests that we should not introduce trainable parameters to the same positions for each individual task but allocate trainable parameters at task-specific ones as we proposed. Next, we investigate the most insensitive weight matrices within a block. A ViT block consists of a query $W_{q}$ , a key $W_{k}$ , a value $W_{v}$ , and an output $W_{o}$ weight matrices in the multi-head self-attention layer and two weight matrices $W_{fc1}$ and $W_{fc2}$ in the feed-forward network as elaborated in [51, 12]. We observe that the query $W_{q}$ and key $W_{k}$ weight matrices have the lowest proportions of sensitive parameters for all three sample tasks. Since $W_{q}$ and $W_{k}$ are responsible for learning the attention scores which indicate the pairwise similarity among the patches, we speculate that although domain changes, the patch relationships learned during pre-training can be efficiently reused when transferred to downstream classification tasks.", + "bbox": [ + 75, + 493, + 472, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 470, + 617, + 487 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we have explored identifying and allocating trainable parameters to task-specific important positions for visual parameter-efficient tuning. Specifically, we have proposed a novel criterion to quickly measure the sensitivity of the pre-trained parameters for each specific task before fine-tuning. Based on the parameter sensitivity, we have proposed a trainable parameter allocation strategy that adaptively combines both unstructured and structured tuning under a desired trainable parameter budget, enabling high representational capability and flexibility. Finally, we have conducted extensive experiments on a total of 24 downstream recognition tasks with both plain and hierarchical vision Transformer backbones under different pre-training strategies to demonstrate the versatility and effectiveness of our proposed SPT. Notably, we have shown that our approach is complementary to the existing PEFT methods and improves their performance significantly. In the future, we will explore adapting large vision models to more downstream tasks with SPT, e.g., dense prediction and vision-and-language tasks, and improve the training efficiency of SPT for on-device training [6, 30].", + "bbox": [ + 496, + 493, + 890, + 811 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement. We thank Jing liu and Ziyi Liu for their helpful discussions. This research is partially supported by Monash FIT Start-up Grant. Dr. Jing Zhang is supported by the Australian Research Council project FL-170100117.", + "bbox": [ + 496, + 811, + 890, + 871 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11833", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Y. Bao, Y. Li, S.-L. Huang, L. Zhang, L. Zheng, A. Zamir, and L. Guibas. An information-theoretic approach to transferability in task transfer learning. In ICIP, pages 2309-2313. IEEE, 2019. 3", + "[2] G. Bender, P.-J. Kindermans, B. Zoph, V. Vasudevan, and Q. Le. Understanding and simplifying one-shot architecture search. In ICML, pages 550-559. PMLR, 2018. 2", + "[3] A. Brock, T. Lim, J. M. Ritchie, and N. Weston. Smash: one-shot model architecture search through hypernetworks. In ICLR, 2018. 2", + "[4] S. Caelles, K.-K. Maninis, J. Pont-Tuset, L. Leal-Taixe, D. Cremers, and L. Van Gool. One-shot video object segmentation. In CVPR, pages 221-230, 2017. 3", + "[5] H. Cai, C. Gan, T. Wang, Z. Zhang, and S. Han. Once-for-all: Train one network and specialize it for efficient deployment. In ICLR, 2020. 2", + "[6] H. Cai, C. Gan, L. Zhu, and S. Han. Tinytl: Reduce memory, not parameters for efficient on-device learning. NeurIPS, 33:11285-11297, 2020. 9", + "[7] H. Cai, L. Zhu, and S. Han. ProxylessNAS: Direct neural architecture search on target task and hardware. In ICLR, 2019. 4", + "[8] N. S. Chatterji, B. Neyshabur, and H. Sedghi. The intriguing role of module criticality in the generalization of deep networks. In ICLR, 2020. 3", + "[9] S. Chen, C. Ge, Z. Tong, J. Wang, Y. Song, J. Wang, and P. Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 2022. 1, 2, 4, 5", + "[10] X. Chen, S. Xie, and K. He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021. 5, 6", + "[11] Y. Cui, Y. Song, C. Sun, A. Howard, and S. Belongie. Large scale fine-grained categorization and domain-specific transfer learning. In CVPR, pages 4109-4118, 2018. 3", + "[12] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 4, 5, 9", + "[13] T. Gebru, J. Krause, Y. Wang, D. Chen, J. Deng, and L. Fei-Fei. Fine-grained car detection for visual census estimation. In AAAI, 2017. 5", + "[14] D. Guo, A. Rush, and Y. Kim. Parameter-efficient transfer learning with diff pruning. In ACL-IJCNLP, 2021. 3", + "[15] Y. Guo, Y. Li, L. Wang, and T. Rosing. Adafilter: Adaptive filter fine-tuning for deep transfer learning. In AAAI, pages 4060-4066, 2020. 3", + "[16] Y. Guo, H. Shi, A. Kumar, K. Grauman, T. Rosing, and R. Feris. Spottune: transfer learning through adaptive fin-tuning. In CVPR, pages 4805-4814, 2019. 3, 9" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] T. Hao, H. Chen, Y. Guo, and G. Ding. Consolidator: Mergable adapter with group connections for vision transformer. In ICLR, 2023. 3", + "[18] J. He, C. Zhou, X. Ma, T. Berg-Kirkpatrick, and G. Neubig. Towards a unified view of parameter-efficient transfer learning. In ICLR, 2022. 2", + "[19] K. He, X. Chen, S. Xie, Y. Li, P. Dollár, and R. Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 5, 6", + "[20] N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. De Laroussilhe, A. Gesmundo, M. Attariyan, and S. Gelly. Parameter-efficient transfer learning for nlp. In ICML, pages 2790-2799, 2019. 1, 2, 4, 5, 7, 8", + "[21] E. J. Hu, yelong shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 1, 2, 3, 4, 5, 7, 8", + "[22] M. Jia, L. Tang, B.-C. Chen, C. Cardie, S. Belongie, B. Hariharan, and S.-N. Lim. Visual prompt tuning. In ECCV, 2022. 1, 2, 5, 7, 8", + "[23] S. Jie and Z.-H. Deng. Convolutional bypasses are better vision transformer adapters. arXiv preprint arXiv:2207.07039, 2022. 1, 2, 4", + "[24] A. Khosla, N. Jayadevaprakash, B. Yao, and L. Fei-Fei. Novel dataset for fine-grained image categorization. In CVPRW, 2011. 5", + "[25] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 5, 7, 8", + "[26] A. Kumar, A. Raghunathan, R. M. Jones, T. Ma, and P. Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In ICLR, 2022. 2, 3", + "[27] B. Lester, R. Al-Rfou, and N. Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021. 2", + "[28] X. L. Li and P. Liang. Prefix-tuning: Optimizing continuous prompts for generation. In IJCNLP, pages 4582-4597. Association for Computational Linguistics, 2021. 2", + "[29] D. Lian, D. Zhou, J. Feng, and X. Wang. Scaling & shifting your features: A new baseline for efficient model tuning. NeurIPS, 2022. 3", + "[30] J. Lin, L. Zhu, W.-M. Chen, W.-C. Wang, C. Gan, and S. Han. On-device training under 256kb memory. arXiv preprint arXiv:2206.15472, 2022.9", + "[31] H. Liu, K. Simonyan, and Y. Yang. DARTS: Differentiable architecture search. In ICLR, 2019. 4", + "[32] Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 5, 7", + "[33] Z. Liu, H. Mao, C.-Y. Wu, C. Feichtenhofer, T. Darrell, and S. Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 7", + "[34] I. Loshchilov and F. Hutter. Fixing weight decay regularization in adam, 2018. 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11834", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] P. Molchanov, A. Mallya, S. Tyree, I. Frosio, and J. Kautz. Importance estimation for neural network pruning. In CVPR, pages 11264-11272, 2019. 2, 4", + "[36] M. M. Naseer, K. Ranasinghe, S. H. Khan, M. Hayat, F. Shahbaz Khan, and M.-H. Yang. Intriguing properties of vision transformers. *NeurIPS*, 34:23296-23308, 2021. 3", + "[37] B. Neyshabur, H. Sedghi, and C. Zhang. What is being transferred in transfer learning? NeurIPS, 33:512-523, 2020. 2, 3", + "[38] C. Nguyen, T. Hassner, M. Seeger, and C. Archambeau. Leep: A new measure to evaluate transferability of learned representations. In ICML, pages 7294-7305. PMLR, 2020. 3", + "[39] M.-E. Nilsback and A. Zisserman. Automated flower classification over a large number of classes. In ICVGIP, pages 722-729. IEEE, 2008. 5", + "[40] J. Pleted and T. Gedeon. Deep transfer learning for image classification: a survey. arXiv preprint arXiv:2205.09904, 2022.3", + "[41] J. Puigcerver, C. Riquelme, B. Mustafa, C. Renggli, A. S. Pinto, S. Gelly, D. Keysers, and N. Houlsby. Scalable transfer learning with expert models. arXiv preprint arXiv:2009.13239, 2020. 3", + "[42] M. Raghu, T. Unterthiner, S. Kornblith, C. Zhang, and A. Dosovitskiy. Do vision transformers see like convolutional neural networks? NeurIPS, 34:12116-12128, 2021. 3", + "[43] M. T. Rosenstein, Z. Marx, L. P. Kaelbling, and T. G. Dietterich. To transfer or not to transfer. In NeurlPSW, 2005. 3", + "[44] R. K. Srivastava, K. Greff, and J. Schmidhuber. Training very deep networks. NeurIPS, 28, 2015. 2", + "[45] X. Sun, R. Panda, R. Feris, and K. Saenko. Adashare: Learning what to share for efficient deep multi-task learning. NeurIPS, 33:8728-8740, 2020. 3", + "[46] Y.-L. Sung, J. Cho, and M. Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. arXiv preprint arXiv:2206.06522, 2022. 2", + "[47] A. T. Tran, C. V. Nguyen, and T. Hassner. Transferability and hardness of supervised classification tasks. In ICCV, pages 1395–1405, 2019. 3", + "[48] C.-H. Tu, Z. Mai, and W.-L. Chao. Visual query tuning: Towards effective usage of intermediate representations for parameter and memory efficient transfer learning. arXiv preprint arXiv:2212.03220, 2022. 2", + "[49] E. Tzeng, J. Hoffman, N. Zhang, K. Saenko, and T. Darrell. Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:1412.3474, 2014. 8", + "[50] G. Van Horn, S. Branson, R. Farrell, S. Haber, J. Barry, P. Ipeirotis, P. Perona, and S. Belongie. Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In CVPR, pages 595–604, 2015. 5" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. NeurIPS, pages 5998-6008, 2017. 9", + "[52] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The caltech-ucsd birds-200-2011 dataset. Tech. Rep. CNS-TR-2011-001, California Institute of Technology, 2011. 5", + "[53] Z. Wang, Z. Dai, B. Póczos, and J. Carbonell. Characterizing and avoiding negative transfer. In CVPR, pages 11293-11302, 2019. 3", + "[54] H. Wu, J. Xu, J. Wang, and M. Long. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. NeurIPS, 34:22419-22430, 2021. 2", + "[55] R. Xu, F. Luo, Z. Zhang, C. Tan, B. Chang, S. Huang, and F. Huang. Raise a child in large language model: Towards effective and generalizable fine-tuning. In EMNLP, 2021. 3", + "[56] Y. Xu, Q. Zhang, J. Zhang, and D. Tao. Vitae: Vision transformer advanced by exploring intrinsic inductive bias. In NeurIPS, 2021. 1", + "[57] J. Yoon, S. Arik, and T. Pfister. Data valuation using reinforcement learning. In ICML, pages 10842-10851. PMLR, 2020. 3", + "[58] J. Yosinski, J. Clune, Y. Bengio, and H. Lipson. How transferable are features in deep neural networks? NeurIPS, 27, 2014. 2, 3", + "[59] E. B. Zaken, Y. Goldberg, and S. Ravfogel. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In ACL, pages 1-9, 2022. 3, 5", + "[60] X. Zhai, J. Puigcerver, A. Kolesnikov, P. Ruyssen, C. Riquelme, M. Lucic, J. Djolonga, A. S. Pinto, M. Neumann, A. Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 5, 6, 7, 8", + "[61] X. Zhai, J. Puigcerver, A. Kolesnikov, P. Ruyssen, C. Riquelme, M. Lucic, J. Djolonga, A. S. Pinto, M. Neumann, A. Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 9", + "[62] C. Zhang, S. Bengio, and Y. Singer. Are all layers created equal? arXiv preprint arXiv:1902.01996, 2019. 3", + "[63] Y. Zhang, K. Zhou, and Z. Liu. Neural prompt search. arXiv preprint arXiv:2206.04673, 2022. 1, 2, 5", + "[64] M. Zhao, T. Lin, F. Mi, M. Jaggi, and H. Schütze. Masking as an efficient alternative to finetuning for pretrained language models. In EMNLP, 2020. 3", + "[65] S. Zheng, J. Lu, H. Zhao, X. Zhu, Z. Luo, Y. Wang, Y. Fu, J. Feng, T. Xiang, P. H. Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. 7", + "[66] B. Zhou, H. Zhao, X. Puig, S. Fidler, A. Barriuso, and A. Torralba. Scene parsing through ade20k dataset. In CVPR, pages 633-641, 2017. 7" + ], + "bbox": [ + 501, + 90, + 892, + 873 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11835", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning/219bbdca-cfe8-433f-a6db-97c4773a1a10_model.json b/2023/Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning/219bbdca-cfe8-433f-a6db-97c4773a1a10_model.json new file mode 100644 index 0000000000000000000000000000000000000000..08d6de375289fc0ac11fb25bb97f46ebbe7a717b --- /dev/null +++ b/2023/Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning/219bbdca-cfe8-433f-a6db-97c4773a1a10_model.json @@ -0,0 +1,2246 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.193, + 0.131, + 0.778, + 0.154 + ], + "angle": 0, + "content": "Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning" + }, + { + "type": "text", + "bbox": [ + 0.188, + 0.18, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Haoyu He\\(^{1}\\) Jianfei Cai\\(^{1}\\) Jing Zhang\\(^{2}\\) Dacheng Tao\\(^{2}\\) Bohan Zhuang\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.207, + 0.725, + 0.227 + ], + "angle": 0, + "content": "1 ZIP Lab, Monash University 2 The University of Sydney" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.264, + 0.314, + 0.28 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.296, + 0.474, + 0.751 + ], + "angle": 0, + "content": "Visual Parameter-Efficient Fine-Tuning (PEFT) has become a powerful alternative for full fine-tuning so as to adapt pre-trained vision models to downstream tasks, which only tunes a small number of parameters while freezing the vast majority ones to ease storage burden and optimization difficulty. However, existing PEFT methods introduce trainable parameters to the same positions across different tasks depending solely on human heuristics and neglect the domain gaps. To this end, we study where to introduce and how to allocate trainable parameters by proposing a novel Sensitivity-aware Visual Parameter-efficient fine-Tuning (SPT) scheme, which adaptively allocates trainable parameters to task-specific important positions given a desired tunable parameter budget. Specifically, our SPT first quickly identifies the sensitive parameters that require tuning for a given task in a data-dependent way. Next, our SPT further boosts the representational capability for the weight matrices whose number of sensitive parameters exceeds a pre-defined threshold by utilizing existing structured tuning methods, e.g., LoRA [21] or Adapter [20], to replace directly tuning the selected sensitive parameters (unstructured tuning) under the budget. Extensive experiments on a wide range of downstream recognition tasks show that our SPT is complementary to the existing PEFT methods and largely boosts their performance, e.g., SPT improves Adapter with supervised pre-trained ViT-B/16 backbone by \\(4.2\\%\\) and \\(1.4\\%\\) mean Top-1 accuracy, reaching SOTA performance on FGVC and VTAB-1k benchmarks, respectively. Source code is at https://github.com/ziplab/SPT." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.777, + 0.21, + 0.793 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.803, + 0.47, + 0.879 + ], + "angle": 0, + "content": "To effectively adapt the pre-trained representations to the downstream tasks, the de-facto choice is full fine-tuning, which initializes the model with the pre-trained weights and tunes all the parameters. However, vanilla full fine-tuning needs to store a separate instance of parameters for each" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.264, + 0.875, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.472, + 0.895, + 0.685 + ], + "angle": 0, + "content": "Figure 1: (a) Existing PEFT methods, such as Adapter [20] introduce trainable parameters to the same positions for all downstream tasks. These methods design task-agnostic positions to employ trainable parameters relying on heuristics and neglect consideration of the distinct domain gaps and characteristics for the downstream tasks. (b) Our Sensitivity-aware visual Parameter-efficient fine-Tuning (SPT) introduces trainable parameters to the task-specific important positions and allocates them with both unstructured and structured tuning granularities, simultaneously. For structured tuning, SPT can exploit any existing structured tuning methods, such as LoRA [21] or Adapter [20]. Red lines and blocks represent trainable parameters and modules, while blue lines represent frozen parameters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.693, + 0.895, + 0.83 + ], + "angle": 0, + "content": "task and each deployment scenario. It can be extremely storage-intensive as the storage cost grows linearly with the number of possible cases, considering there are vast varieties of downstream tasks and dynamic deployment environments, especially when deploying the large vision models [12, 32, 56] to mobile systems. For example, even storing a single large pre-trained ViT-H [12] model on a local disk requires at least 2.3GB, while the Top-10 U.S. apps required only collectively 2.2GB in May 2021.\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.83, + 0.895, + 0.876 + ], + "angle": 0, + "content": "Notably, an emerging solution is to replace vanilla fine-tuning with visual Parameter-Efficient Fine-Tuning (PEFT) [22, 9, 63, 23], which only tunes a small num" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.421, + 0.901 + ], + "angle": 0, + "content": "†Corresponding author. E-mail: bohan.zhuang@gmail.com" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.818, + 0.901 + ], + "angle": 0, + "content": "1https://sensorTower.com/blog/ios-app-size-growth-2021" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11825" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.35 + ], + "angle": 0, + "content": "ber of trainable parameters while freezing the vast majority ones that are shared by multiple tasks. As PEFT approaches exhibit less than \\(1\\%\\) of trainable parameters, the storage burden is largely alleviated. Another attractive property of PEFT is that tuning fewer parameters eases the optimization difficulty and mitigates the overfitting issue when adapting large pre-trained models on the target dataset, thereby achieving comparable or even better performance than vanilla fine-tuning [22]. Although promising, the existing PEFT approaches introduce trainable parameters to the same positions for all downstream tasks, relying on human heuristics and neglecting task-specific domain gaps and characteristics, which limits their performance. For instance, in a task-agnostic manner, Prompt Tuning-deep [22] and Adapter [20] respectively add trainable parameters to multi-head self-attention and feed-forward network layers for all distinct tasks as depicted in Figure 1 (a)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.473, + 0.671 + ], + "angle": 0, + "content": "To address this fundamental challenge, we explore where to introduce and how to allocate trainable parameters under a desired parameter budget by presenting a novel Sensitivity-aware visual Parameter-efficient fine-Tuning (SPT) scheme that identifies the task-specific important positions to adaptively allocate trainable parameters. Since the pre-trained weights at distinct positions have varying contributions for different downstream tasks [58, 26, 37], we first propose a new criterion to quickly identify the task-specific sensitive parameters that require tuning in a data-dependent way. Inspired by model pruning metrics [44, 35, 2, 3], we propose to measure the parameter sensitivity with the loss reduction when being tuned, which can be approximated by a first-order Taylor expansion derived within a single forward and backward pass ahead of fine-tuning in one-shot. Our sensitivity criterion is simple and effective, which can identify the task-specific important positions to introduce trainable parameters for any backbone quickly. For instance, calculating the sensitivity for ViT-B/16 backbone takes only 5.5 seconds with a single GPU on any of the VTAB-1k datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "With our criterion, we empirically observe that the proportions of the sensitivity parameters for each block indeed vary markedly across different tasks in Section 4.4. To allocate the trainable parameters under a desired trainable parameter budge, an intuitive solution is to directly tune the most sensitive weight connections, which we name unstructured tuning. Despite its simplicity and flexibility, unstructured tuning only tunes a few parameters which still lacks representational capability and is challenging to bridge the domain gap. To this end, we propose to further incorporate structured tuning to replace unstructured tuning at the sensitive weight matrices whose numbers of sensitive parameters exceed a pre-defined threshold to improve the representational capability under a similar parameter budget. Structured tuning can be implemented by any parameter-efficient" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.198 + ], + "angle": 0, + "content": "structured tuning methods [21, 9, 23, 22] that directly adjust the hidden representations, e.g., inserting an adapter module sequentially after the sensitive weight matrices. Therefore, our SPT adaptively combines both unstructured and structured tuning granularity and allocates trainable parameters with high flexibility and representational capability for each distinct downstream task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.2, + 0.895, + 0.472 + ], + "angle": 0, + "content": "This paper has the following key contributions. 1) We make the pioneering exploration to identify the task-specific important positions under the PEFT setting, which is fast, effective, versatile to be applied to various backbones with different pre-training strategies, and orthogonal to the existing PEFT methods. 2) Based on the sensitivity criterion, we propose a trainable parameter allocation strategy that adaptively combines both unstructured and structured tuning under a desired parameter budget to achieve high flexibility, large capacity, and favorable trade-off between parameter efficiency and accuracy. 3) Extensive experiments on a total of 24 downstream recognition tasks with both plain and hierarchical vision Transformer backbones under supervised and self-supervised pre-trainings show that our SPT is complementary to the existing PEFT methods and boosts their performance by large margins. For instance, SPT improves Adapter [20] by \\(4.2\\%\\) mean Top-1 accuracy, outperforming the SOTA PEFT methods on the FGVC benchmark." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.484, + 0.642, + 0.499 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.657 + ], + "angle": 0, + "content": "Parameter-efficient fine-tuning. Full fine-tuning is the most predominant approach when adapting a large-scale pre-trained model to downstream tasks, where the model is initialized from the pre-trained weights with all parameters trainable. Yet, when a model becomes larger, parameter-efficient fine-tuning [27, 28] is highly desirable, which tunes only a tiny portion of parameters to alleviate the storage burden. The general PEFT approaches can be categorized into addition-based PEFT methods and reparameterization-based PEFT methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Addition-based PEFT attaches additional trainable parameters to the backbone and only tunes these parameters. Apart from Prompt tuning [22] and Adapter [20], recent addition-based methods study connecting or combining existing PEFT methods. For instance, He et al. [18] connect Prompt tuning and Adapter and provide a unified view that all PEFT approaches share the same design to adjust the hidden representations. Zhang et al. [63] search for the optimal configurations to combine multiple PEFT approaches following once-for-all scheme [5, 54]. Since the additional parameters require extra computations compared to full fine-tuning, a few recent works [46, 48] design specific architectures to avoid storing the intermediate activations, thereby alleviating the fine-tuning memory cost. However, it is noteworthy that enhancing training efficiency is not the primary objective of our work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11826" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.455 + ], + "angle": 0, + "content": "Reparameterization-based PEFT aims to avoid extra computational costs by tuning parameters that are inherently in or can be reparameterized into the backbone during inference. Prior works select the parameters that are inherently in the backbone, including the bias terms [59], the last several layers [58, 4], and weight connections [14, 64]. To reparameterize new parameters into the backbone [17, 29], representative work LoRA [21] optimizes two low-rank matrices which can be further merged into the weight matrices. In contrast to the aforementioned works, we argue the importance of tuning parameters at task-specific important positions and quickly identify them with our proposed parameter sensitivity criterion before tuning, which is complementary to and provides valuable guidance for the existing PEFT methods. Moreover, our SPT can also be inference-efficient when implementing structured tuning with any reparameterization-based structured tuning method. Recently, SSF [29] is proposed to introduce trainable scaling and shifting parameters that can be absorbed into the previous linear layers. However, it cannot scale to higher trainable parameter budgets and requires a complex and time-consuming hyper-parameter search for learning rate, weight decay, and drop-path rate on each individual dataset, thus is not directly comparable to our method." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.464, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Task-specific transfer learning. The effectiveness of transferring pre-trained models to downstream tasks strongly depends on the relationship between the source and target tasks [43, 53, 26, 40]. This has motivated the community to explore the optimal pre-training data [11, 57], model [47, 38], and weights [16, 55] for the target task. To seek suitable task-specific pre-training data, Cui et al. [11] select the source domain data from the top-k most similar classes measured by Earth Mover's Distance; Yoon et al. [57] weight each class in the source domain with reinforcement learning; and Puigcerver et al. [41] first train a diverse set of experts and then select the most relevant expert for each target task. Another line of work selects a suitable pre-trained model for the target task ahead of fine-tuning by measuring the transferability of pre-trained models to the target domain with interclass covariance between the source data and target classes [1] or conditional cross-entropy [47] between the source and target labels. Considering the transferability of the feature representations at distinct layers for the same pre-trained model is different [58, 37], recent works [15, 45] endeavour transfer task-specific weights by freezing some pre-trained weights and fine-tuning the rest. For example, the task-specific fine-tuned weights are selected by learning a policy network with Gumbel-Softmax [16], optimizing a sparse mask with \\( L_{0} \\) norm [14], and learning binary gates for each parameter [64]. Our SPT also adaptively selects task-specific parameters. In contrast to the previous work, we 1) derive task-specific important positions prior to fine-tuning with only a single" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.23 + ], + "angle": 0, + "content": "forward and backward pass, which is computationally efficient; 2) mask the gradients for insensitive parameters in unstructured tuning with fixed binary masks, thereby having more affordable fine-tuning memory than optimizing learnable binary masks in [14, 64]. Moreover, we are pioneering work to adaptively allocate task-specific trainable parameters with both fine-grained unstructured and coarse-grained structured tuning granularities to achieve both high flexibility and representational capability." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.241, + 0.593, + 0.258 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.264, + 0.895, + 0.37 + ], + "angle": 0, + "content": "Our sensitivity-aware visual parameter-efficient fine-tuning consists of two stages. In the first stage, SPT measures the task-specific sensitivity for the pre-trained parameters (Section 3.1). Based on the parameter sensitivity and a given parameter budget, SPT then adaptively allocates trainable parameters to task-specific important positions (Section 3.2)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.378, + 0.807, + 0.395 + ], + "angle": 0, + "content": "3.1. Task-specific Parameter Sensitivity" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.895, + 0.564 + ], + "angle": 0, + "content": "Recent research has observed that pre-trained backbone parameters exhibit varying feature patterns [42, 36] and criticality [62, 8] at distinct positions. Moreover, when transferred to downstream tasks, their efficacy varies depending on how much pre-trained features are reused and how well they adapt to the specific domain gap [58, 26, 37]. Motivated by these observations, we argue that not all parameters contribute equally to the performance across different tasks in PEFT and propose a new criterion to measure the sensitivity of the parameters in the pre-trained backbone for a given task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.895, + 0.688 + ], + "angle": 0, + "content": "Specifically, given the training dataset \\(\\mathcal{D}_t\\) for the \\(t\\)-th task and the pre-trained model weights \\(\\boldsymbol{w} = \\{w_1, w_2, \\dots, w_N\\} \\in \\mathbb{R}^N\\) where \\(N\\) is the total number of parameters, the objective for the task is to minimize the empirical risk: \\(\\min_{\\boldsymbol{w}} E(\\mathcal{D}_t, \\boldsymbol{w})\\). We denote the parameter sensitivity set as \\(\\mathcal{S} = \\{s_1, \\dots, s_N\\}\\) and the sensitivity \\(s_n\\) for parameter \\(w_n\\) is measured by the empirical risk difference when tuning it:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.701, + 0.892, + 0.719 + ], + "angle": 0, + "content": "\\[\ns _ {n} = E \\left(\\mathcal {D} _ {t}, \\boldsymbol {w}\\right) - E \\left(\\mathcal {D} _ {t}, \\boldsymbol {w} \\mid w _ {n} = w _ {n} ^ {*}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.728, + 0.895, + 0.903 + ], + "angle": 0, + "content": "where \\( w_{n}^{*} = \\operatorname*{argmin}_{w_{n}}(E(\\mathcal{D}_{t},\\boldsymbol {w})) \\). We can reparameterize the tuned parameters as \\( w_{n}^{*} = w_{n} + \\Delta_{w_{n}} \\), where \\( \\Delta_{w_n} \\) denotes the update for \\( w_{n} \\) after tuning. Here we individually measure the sensitivity of each parameter, which is reasonable given that most of the parameters are frozen during fine-tuning in PEFT. However, it is still computationally intensive to compute Eq. (1) for two reasons. Firstly, getting the empirical risk for \\( N \\) parameters requires forwarding the entire network \\( N \\) times, which is time-consuming. Secondly, it is challenging to derive \\( \\Delta_{w_n} \\), as we have to tune each individual \\( w_{n} \\) until convergence." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "11827" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.885, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.159, + 0.895, + 0.237 + ], + "angle": 0, + "content": "Figure 2: Overview of our trainable parameter allocation strategy. With the parameter sensitivity set \\( S \\), we first get the top-\\( \\tau \\) sensitive parameters. Instead of directly tuning these sensitive parameters, we also boost the representational capability by replacing unstructured tuning with structured tuning at sensitive weight matrices that have a large number of sensitive parameters, which can be implemented by an existing structured tuning method, e.g., LoRA [21] and Adapter [20]. Red lines and blocks represent trainable parameters and modules, while blue lines represent frozen parameters." + }, + { + "type": "code_caption", + "bbox": [ + 0.078, + 0.242, + 0.47, + 0.271 + ], + "angle": 0, + "content": "Algorithm 1 Computing task-specific parameter sensitivities" + }, + { + "type": "algorithm", + "bbox": [ + 0.091, + 0.274, + 0.473, + 0.487 + ], + "angle": 0, + "content": "Input: Pre-trained model with network parameters \\(\\pmb{w}\\) training set \\(\\mathcal{D}_t\\) for the \\(t\\)-th task, and number of training samples \\(C\\) used to calculate the parameter sensitivities \nOutput: Sensitivity set \\(\\mathcal{S} = \\{s_1,\\dots,s_N\\}\\) \nInitialize \\(\\mathcal{S} = \\{0\\}^N\\) \nfor \\(i\\in \\{1,\\ldots ,C\\}\\) do \n Get the \\(i\\)-th training sample of \\(\\mathcal{D}_t\\) \n Compute loss \\(E\\) \n Compute gradients \\(\\pmb{g}\\) \n for \\(n\\in \\{1,\\dots,N\\}\\) do \n Update sensitivity for the \\(n\\)-th parameter: \\(s_n = s_n + g_n^2\\) \n end for \nend for" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.47, + 0.558 + ], + "angle": 0, + "content": "To overcome the first barrier, we simplify the empirical loss by approximating \\( s_n \\) in the vicinity of \\( w \\) by its first-order Taylor expansion" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.563, + 0.47, + 0.582 + ], + "angle": 0, + "content": "\\[\ns _ {n} ^ {(1)} = - g _ {n} \\Delta_ {w _ {n}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.47, + 0.708 + ], + "angle": 0, + "content": "where the gradients \\( \\pmb{g} = \\partial E / \\partial \\pmb{w} \\), and \\( g_{n} \\) is the gradient of the \\( n \\)-th element of \\( \\pmb{g} \\). To address the second barrier, following [31, 7], we take the one-step unrolled weight as the surrogate for \\( w_{n}^{*} \\) and approximate \\( \\Delta_{w_n} \\) in Eq. (2) with a single step of gradient descent. We can accordingly get \\( s_n^{(1)}\\approx g_n^2\\epsilon \\), where \\( \\epsilon \\) is the learning rate. Since \\( \\epsilon \\) is the same for all parameters, we can eliminate it when comparing the sensitivity with the other parameters and finally get" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.71, + 0.469, + 0.73 + ], + "angle": 0, + "content": "\\[\ns _ {n} ^ {(1)} \\approx g _ {n} ^ {2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Therefore, the sensitivity of a parameter can be efficiently measured by its potential to reduce the loss on the target domain. Note that although our criterion draws inspiration from pruning work [35], it is distinct from it. [35] measures the parameter importance by the squared change in loss when removing them, i.e., \\((E(\\mathcal{D}_t,\\boldsymbol {w}) - E(\\mathcal{D}_t,\\boldsymbol {w}\\mid w_n = 0))^2\\) and finally derives the parameter importance by \\((g_{n}w_{n})^{2}\\), which is different from our formulations in Eqs. (1) and (3)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In practice, we accumulate \\( \\mathcal{S} \\) from a total number of \\( C \\) training samples ahead of fine-tuning to generate accu-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.242, + 0.895, + 0.334 + ], + "angle": 0, + "content": "rate sensitivity as shown in Algorithm 1, where \\( C \\) is a predefined hyper-parameter. In Section 4.3, we show that employing only 400 training samples is sufficient for getting reasonable parameter sensitivity, which requires only 5.5 seconds with a single GPU for any VTAB-1k dataset with ViT-B/16 backbone [12]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.344, + 0.862, + 0.361 + ], + "angle": 0, + "content": "3.2. Adaptive Trainable Parameters Allocation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.519 + ], + "angle": 0, + "content": "Our next step is to allocate trainable parameters based on the obtained parameter sensitivity set \\(S\\) and a desired parameter budget \\(\\tau\\). A straightforward solution is to directly tune the top-\\(\\tau\\) most sensitive unstructured connections (parameters) while keeping the rest frozen, which we name unstructured tuning. Specifically, we select the top-\\(\\tau\\) most sensitive weight connections in \\(S\\) to form the sensitive weight connection set \\(\\mathcal{T}\\). Then, for a weight matrix \\(\\boldsymbol{W} \\in \\mathbb{R}^{d_{\\mathrm{in}} \\times d_{\\mathrm{out}}}\\), we can get a binary mask \\(M \\in \\mathbb{R}^{d_{\\mathrm{in}} \\times d_{\\mathrm{out}}}\\) computed by" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.527, + 0.892, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {M} ^ {j} = \\left\\{ \\begin{array}{l l} 1 & \\boldsymbol {W} ^ {j} \\in \\mathcal {T} \\\\ 0 & \\boldsymbol {W} ^ {j} \\notin \\mathcal {T} \\end{array} , \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.892, + 0.643 + ], + "angle": 0, + "content": "where \\(W^{j}\\) and \\(M^j\\) are the \\(j\\)-th element in \\(W\\) and \\(M\\), respectively. Accordingly, we can train the sensitive parameters by gradient descent and the updated weight matrix can be formulated as \\(W' \\gets W - \\epsilon g_W \\odot M\\), where \\(g_W\\) is the gradient for \\(W\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.893, + 0.901 + ], + "angle": 0, + "content": "However, considering PEFT approaches generally limit the proportion of trainable parameters to less than \\(1\\%\\), tuning only a small number of unstructured weight connections might not have enough representational capability to handle the downstream datasets with large domain gaps from the source pre-training data. Therefore, to improve the representational capability, we propose to replace unstructured tuning with structured tuning at the sensitive weight matrices that have a high number of sensitive parameters. To preserve the parameter budget, we can implement structured tuning with an existing efficient structured tuning PEFT method [21, 9, 20, 23] that learns to directly adjust all hidden dimensions at once. We depict an overview of our trainable parameter allocation strategy in Figure 2. For example, we can employ the low-rank reparameterization trick LoRA [21] to the sensitive weight matrices and the one-step update for \\(\\mathbf{W}\\) can be formulated as" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11828" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.085, + 0.089, + 0.472, + 0.119 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {W} ^ {\\prime} = \\left\\{ \\begin{array}{l l} \\boldsymbol {W} + \\boldsymbol {W} _ {\\mathrm {d o w n}} \\boldsymbol {W} _ {\\mathrm {u p}} & \\text {i f} \\sum_ {j = 0} ^ {d _ {\\mathrm {i n}} \\times d _ {\\mathrm {o u t}}} \\boldsymbol {M} ^ {j} \\geq \\sigma_ {\\mathrm {o p t}} \\\\ \\boldsymbol {W} - \\epsilon \\boldsymbol {g} _ {\\boldsymbol {W}} \\odot \\boldsymbol {M} & \\text {o t h e r w i s e} \\end{array} \\right.,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.12, + 0.47, + 0.31 + ], + "angle": 0, + "content": "where \\(W_{\\mathrm{down}} \\in \\mathbb{R}^{d_{\\mathrm{in}} \\times r}\\) and \\(W_{\\mathrm{up}} \\in \\mathbb{R}^{r \\times d_{\\mathrm{out}}}\\) are two learnable low-rank matrices to approximate the update of \\(W\\) and rank \\(r\\) is a hyper-parameter where \\(r \\ll \\min(d_{\\mathrm{in}}, d_{\\mathrm{out}})\\). In this way, we perform structured tuning on \\(W\\) when its number of sensitive parameters exceeds \\(\\sigma_{\\mathrm{opt}}\\), whose value depends on the pre-defined type of structured tuning method. For example, since implementing structured tuning with LoRA requires \\(2 \\times d_{\\mathrm{in}} \\times d_{\\mathrm{out}} \\times r\\) trainable parameters for each sensitive weight matrix, we set \\(\\sigma_{\\mathrm{LoRA}} \\gets 2 \\times d_{\\mathrm{in}} \\times d_{\\mathrm{out}} \\times r\\) to ensure that the number of trainable parameters introduced by structured tuning is always equal to or lower than the number of sensitive parameters." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.311, + 0.469, + 0.416 + ], + "angle": 0, + "content": "In this way, our SPT adaptively incorporates both structured and unstructured tuning granularities to enable higher flexibility and stronger representational power, simultaneously. In Section 4.3, we show that structured tuning is important for the downstream tasks with larger domain gaps and both unstructured and structured tuning contribute clearly to the superior performance of our SPT." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.431, + 0.21, + 0.448 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.455, + 0.267, + 0.47 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.469, + 0.763 + ], + "angle": 0, + "content": "Datasets and metrics. We evaluate our SPT on total 24 downstream tasks in two groups following [22]. 1) FGVC is a benchmark for fine-grained visual classification, including CUB-200-2011 [52], NABirds [50], Oxford Flowers [39], Stanford Cars [13], and Stanford Dogs [24] datasets. Each FGVC dataset contains between 55 to 200 classes and a few thousand images for train, validation, and test. We follow the validation splits in [22] if the validation set is unavailable. 2) VTAB-1k [60] is a large-scale transfer learning benchmark consisting of a collection of 19 visual classification tasks. VTAB-1k can further be divided into three groups, including Natural tasks with natural images, Specialized tasks with images captured by specialized equipment, e.g., medical images, and Structured tasks with images mostly generated from synthetic environments. Each of the VTAB-1k dataset has only 800 training and 200 validation samples, while the test set sizes vary. We use top-1 accuracy (\\%) averaged within each group as our main metric following [22]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.469, + 0.884 + ], + "angle": 0, + "content": "Pre-trained backbones. We conduct experiments on the plain vision Transformer backbone ViT-B/16 [12] that is pre-trained on ImageNet [25] with different pre-training strategies following [22], including supervised pre-training and self-supervised pre-training with MAE [19] and MoCo v3 [10] following [22]. We also conduct experiments on the representative hierarchical vision Transformer backbone Swin-B [32] under supervised pre-training." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Contenders. We categorize the baseline methods into" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.287 + ], + "angle": 0, + "content": "addition-based and reparameterization-based PEFT methods as introduced in Section 2. Unless specified, all baseline methods keep the backbone frozen. Addition-based methods require extra computations during inference, including MLP-\\(k\\), PROMPT-SHALLOW [22], PROMPT-DEEP [22], ADAPTER-\\(k\\) [20], ADAPTFORMER [9], and NOAH [63]. Reparameterization-based methods have no additional computational overhead during inference, including LINEAR, PARTIAL-\\(k\\), BIAS [59], and LORA-\\(k\\) [21]. Here \\(k\\) represents the number of bottleneck dimension in ADAPTER-\\(k\\) and LORA-\\(k\\). We also compare with full fine-tuning which is denoted by FULL. We introduce the details of these methods in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.422 + ], + "angle": 0, + "content": "We also introduce two variants of our SPT: addition-based SPT-ADAPTER and reparameterization-based SPT-LORA. SPT-ADAPTER directly adjusts the hidden representations that are computed by sensitive weight matrices following [20], while SPT-LORA approximates updating the sensitive weight matrices following [21]. For the two variants, we follow the exact weight initializations that are described in [21] and follow [63] to set the bottleneck dimension as 8." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.544 + ], + "angle": 0, + "content": "Implementation details. Following [63], we use the AdamW optimizer [34] with cosine learning rate decay and set the batch size, learning rate, and weight decay as 64, \\(1 \\times 10^{-3}\\), and \\(1 \\times 10^{-4}\\), respectively. We also follow [63] for the standard data augmentation pipeline. We set the number of training samples \\(C\\) used to calculate our parameter sensitivities in Algorithm 1 universally to be 800 for the main experiments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.548, + 0.638, + 0.562 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.599 + ], + "angle": 0, + "content": "We compare our method with the baseline methods under different backbones, pre-training strategies, and tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Visual recognition with ViT backbone. First, our proposed SPT-ADAPTER and SPT-LORA achieve the best performance under different trainable parameter budgets with supervised pre-trained ViT-B/16 backbone, as shown in Table 1 and Figure 3 (a). For instance, SPT-ADAPTER outperforms the SOTA method NOAA by a clear margin of \\(0.9\\%\\) mean top-1 accuracy over the 19 VTAB-1k datasets with fewer trainable parameters. We speculate that our SPT variants allocate trainable parameters at task-specific positions compared to the heuristically selected positions in the baseline methods, which contributes to our superior performance. We also observe that our SPT-ADAPTER and SPT-LORA achieve large performance gains over ADAPTER and LORA variants, respectively. For example, SPT-ADAPTER and SPT-LORA with \\(0.41\\%\\) trainable parameters respectively improve ADAPTER-8 and LORA-8 significantly by \\(4.0\\%\\) and \\(3.3\\%\\) mean accuracy on the FGVC benchmark. This suggests that identifying task-specific important positions and combining both unstructured and structured tuning granularities with SPT are complementary to the exist" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11829" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.157, + 0.089, + 0.812, + 0.142 + ], + "angle": 0, + "content": "
ViT-B/16 (85.8M)Total paramsFGVCVTAB-1k
Tuned / TotalMean Acc.Tuned / TotalNaturalSpecializedStructuredMean Acc.
FULL24.02×100%88.5100%75.983.447.669.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.409, + 0.144, + 0.558, + 0.155 + ], + "angle": 0, + "content": "Addition-based methods" + }, + { + "type": "table", + "bbox": [ + 0.158, + 0.158, + 0.81, + 0.281 + ], + "angle": 0, + "content": "
MLP-31.35×1.50%79.81.42%67.872.830.657.1
PROMPT-SHALLOW1.04×0.31%84.60.13%76.879.747.067.8
PROMPT-DEEP1.18×0.98%89.11.14%78.582.455.072.0
ADAPTER-81.06×0.39%85.50.23%79.084.158.573.9
ADAPTER-321.19×0.95%85.60.71%79.684.058.374.0
ADAPTFORMER1.09×0.44%85.10.36%80.685.458.574.8
NOAH---0.52%80.284.961.375.5
SPT-ADAPTER (Ours)1.08×0.41%89.50.30%81.385.360.875.8
SPT-ADAPTER (Ours)1.10×0.47%89.80.44%82.085.861.476.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.376, + 0.283, + 0.591, + 0.296 + ], + "angle": 0, + "content": "Reparameterization-based methods" + }, + { + "type": "table", + "bbox": [ + 0.159, + 0.297, + 0.808, + 0.395 + ], + "angle": 0, + "content": "
LINEAR1.02×0.12%79.30.04%68.977.226.857.6
PARTIAL-13.00×8.38%82.68.30%69.478.534.260.7
BIAS1.05×0.13%88.40.13%73.378.344.165.2
LORA-81.07×0.55%86.00.23%79.584.660.574.9
LORA-161.18×0.90%84.80.69%79.884.960.275.0
SPT-LORA (Ours)1.08×0.41%89.30.31%81.585.660.775.9
SPT-LORA (Ours)1.15×0.60%90.10.63%81.985.961.376.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.4, + 0.895, + 0.463 + ], + "angle": 0, + "content": "Table 1: Comparisons on FGVC and VTAB-1k [60] benchmarks using supervised pre-trained ViT-B/16 backbone pre-trained on ImageNet-21k. \"Total params\" denotes the ratio of the total number of parameters needed for all downstream tasks relative to the one for the pre-trained backbone, and \"Tuned/Total\" denotes the fraction of trainable parameters. Top-1 accuracy (\\%) is reported. The best result is in bold, and the second-best result is underlined." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.469, + 0.892, + 0.52 + ], + "angle": 0, + "content": "
ViT-B/16 (85.8M)Total ParamsVTAB-1kMAEMean Acc.VTAB-1kMoCo v3Mean Acc.
Tuned / TotalNaturalSpecializedStructuredTuned / TotalNaturalSpecializedStructured
FULL38.02×100%59.379.753.864.3100%72.084.742.069.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.412, + 0.523, + 0.558, + 0.534 + ], + "angle": 0, + "content": "Addition-based methods" + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.536, + 0.891, + 0.62 + ], + "angle": 0, + "content": "
ADAPTER-81.08×0.23%57.278.454.763.40.23%27.670.948.449.0
ADAPTER-321.28×0.95%55.378.853.362.50.99%74.282.747.768.2
PROMPT-SHALLOW1.02×0.12%40.069.727.545.70.12%67.382.337.662.4
PROMPT-DEEP1.05×0.23%36.060.626.641.10.07%70.383.042.465.2
SPT-ADAPTER (Ours)1.07×0.26%64.882.460.469.20.08%76.184.960.173.7
SPT-ADAPTER (Ours)1.13×0.41%65.682.760.769.70.30%76.685.061.774.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.379, + 0.623, + 0.592, + 0.635 + ], + "angle": 0, + "content": "Reparameterization-based methods" + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.637, + 0.891, + 0.734 + ], + "angle": 0, + "content": "
LINEAR1.02×0.04%18.952.723.732.10.04%67.581.130.359.6
PARTIAL-14.16×8.30%58.478.347.661.58.30%72.384.647.968.3
BIAS1.06×0.13%54.675.747.759.30.13%72.981.153.469.2
LORA-81.08×0.23%57.577.757.764.30.23%21.266.745.144.3
LORA-161.28×0.69%57.377.159.964.80.69%16.064.048.742.9
SPT-LORA (Ours)1.11×0.29%63.881.660.068.50.30%76.585.463.075.0
SPT-LORA (Ours)1.23×0.69%65.482.461.569.80.50%76.586.063.675.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.739, + 0.895, + 0.801 + ], + "angle": 0, + "content": "Table 2: Comparisons on VTAB-1k [60] benchmark using self-supervised ViT-B/16 backbone pre-trained by MAE [19] and MoCo v3 [10]. \"Total params\" denotes the ratio of the total number of parameters needed for all downstream tasks relative to the one for the pre-trained backbone, and \"Tuned/Total\" denotes the fraction of trainable parameters. Top-1 accuracy (\\%) is reported. The best result is in bold, and the second-best result is underlined." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.808, + 0.395, + 0.824 + ], + "angle": 0, + "content": "ing PEFT methods and boost their performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Second, SPT variants outperform baseline methods and full fine-tuning by significant margins with the self-supervised pre-trained ViT-B/16 backbones. As shown in Table 2, existing PEFT approaches exhibit inferior results than full fine-tuning with the self-supervised pre-trained" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.808, + 0.895, + 0.9 + ], + "angle": 0, + "content": "backbones MAE and MoCo v3. It is worth noting that previous PEFT methods yield inconsistent results with the backbones of different pre-training strategies. In contrast, SPT variants consistently outperform full fine-tuning. In particular, SPT-ADAPTER achieves remarkable \\(5.8\\%\\) and \\(5.5\\%\\) mean top-1 accuracy gains over the best-performing" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "11830" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.47, + 0.259 + ], + "angle": 0, + "content": "
MethodTuned / TotalNaturalSpecializedStructuredMean / Acc.
FULL100%79.186.259.775.0
Addition-based methods
MLP-31.60%73.675.235.761.5
PROMPT-SHALLOW0.04%79.982.537.866.7
PROMPT-DEEP0.23%76.884.553.471.6
ADAPTER-81.18%81.787.361.276.7
SPT-ADAPTER (ours)0.33%83.087.362.177.5
Reparameterization-based methods
LINEAR0.04%73.580.833.562.6
PARTIAL-12.15%73.181.735.063.3
LORA-81.18%81.787.260.176.3
SPT-LORA (ours)0.49%83.187.460.477.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.263, + 0.471, + 0.325 + ], + "angle": 0, + "content": "Table 3: Comparisons on VTAB-1k [60] benchmark with supervised pre-trained Swin-B [32]. \"Tuned/Total\" denotes the fraction of trainable parameters. Top-1 accuracy (\\%) is reported. The best result is in bold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.47, + 0.517 + ], + "angle": 0, + "content": "baseline method on VTAB-1k benchmark with only \\(0.26\\%\\) and \\(0.08\\%\\) trainable parameters for MAE and MoCo v3 pretrained backbones, respectively. Moreover, our observation in supplementary material suggests that self-supervised pretrained ViT backbones have more diverse sensitivity distributions and a higher variance in sensitivity across different tasks than the supervised pre-trained one. This leads to the conjecture that baseline methods which assign trainable parameters to the same positions for all tasks may fail to mitigate the distinct domain gaps in individual downstream datasets, whereas our SPT allocates trainable parameters to task-specific positions accurately." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.47, + 0.669 + ], + "angle": 0, + "content": "Visual recognition with Swin and ConvNeXt backbones. From Table 3, we observe that our SPT-LORA and SPT-ADAPTER also achieve SOTA performance with Swin-B backbone on all dataset groups. We also follow VPT [22] to apply SPT variants to ResNet-alike architecture ConvNeXt-Base [33] and report the results in Table 4. We observe that SPT variants achieve better trade-offs between accuracy and parameter efficiency than the baseline methods. The results further demonstrate the versatility and effectiveness of our SPT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.471, + 0.839 + ], + "angle": 0, + "content": "Semantic segmentation. We follow VPT [22] to conduct semantic segmentation on ADE20k dataset. Following the settings of [22], we apply our SPT to SETR-PUP [65] framework with ViT-L backbone and only allocate trainable parameters to the backbone with the head fully finetuned. We report the mIoU results in Table 5. Notably, SPT-LORA and SPT-ADAPTER outperform the baseline methods by large margins, indicating our SPT can be generalized to the semantic segmentation task. We report the results for conducting only structured tuning for both SPT-LORA and SPT-ADAPTER as it yields higher mIoU." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.849, + 0.231, + 0.866 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Effect of the sensitivity criterion. We investigate the effectiveness of our sensitivity criterion on VTAB-1k by em" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.089, + 0.892, + 0.162 + ], + "angle": 0, + "content": "
MethodTuned/TotalNaturalSpecializedStructuredMean Acc.
FULL100%78.083.760.474.0
LoRA0.79%82.284.764.177.0
Adapter0.47%83.184.964.677.5
SPT-LoRA (ours)0.57%83.486.765.978.7
SPT-Adapter (ours)0.36%83.786.265.378.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.172, + 0.892, + 0.234 + ], + "angle": 0, + "content": "Table 4: Comparisons on VTAB-1k [60] benchmark with supervised pre-trained ConvNeXt-Base [33] backbone. \"Tuned/Total\" denotes the fraction of trainable parameters. Top-1 accuracy \\((\\%)\\) is reported. The best result is in bold." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.248, + 0.892, + 0.332 + ], + "angle": 0, + "content": "
MethodmIoU-s.s. (%)mIoU-m.s. (%)Trainable Param.
VPT44.045.615.8M
LoRA43.945.914.7M
Adapter44.446.614.6M
SPT-LoRA (ours)45.447.514.6M
SPT-Adapter (ours)45.247.214.6M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.343, + 0.892, + 0.403 + ], + "angle": 0, + "content": "Table 5: Semantic Segmentation: Comparisons on ADE20k [66] val with SETR [65] on ViT-L backbone. We report both single-scale (s.s.) and multi-scale (m.s.) mIoU results. The best result is in bold." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.892, + 0.568 + ], + "angle": 0, + "content": "ploying structured tuning methods from [22, 20, 21] to the task-specific sensitive weight matrices. Note that we do not conduct unstructured tuning to ensure fair comparisons. The results are presented in Figure 3 (b). Our criterion brings consistent \\(1.1\\%\\), \\(1.6\\%\\), and \\(0.8\\%\\) performance gains for PROMPT-DEEP, ADAPTER-32, and LORA-16, respectively, which demonstrates the effectiveness and versatility of our sensitivity criterion to identify accurate task-specific important positions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Effect of structured and unstructured tuning. We investigate the effectiveness of unstructured and structured tuning individually on VTAB-1k. The results are presented in Table 6. We start by applying ADAPTER-8 to the sensitive weight matrices identified by our sensitivity criterion (SPT-ADAPTER w/o unstructured). We observe that our sensitivity criterion boosts the performance of all the dataset groups by clear margins, which again demonstrates the importance of our sensitivity criterion. Next, we observe that allocating the trainable parameters to the unstructured sensitive weight connections also brings accuracy improvement to the Natural and Specialized datasets from ADAPTER-8. However, we find that structured tuning is especially important for achieving good performance on Structured datasets. To further investigate this phenomenon, we observe that Structured datasets have larger domain gaps from the pre-training source domain [25] compared to Natural and Specialized datasets as visualized in Figure 3 (c). We hence conjecture that structured tuning has a higher representational capability than unstructured tuning which facilitates mitigating the large domain gaps during fine-tuning (see the supplementary material for visual examples). Finally, we observe" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11831" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.092, + 0.337, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.343, + 0.092, + 0.603, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.096, + 0.892, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.895, + 0.392 + ], + "angle": 0, + "content": "Figure 3: (a) Accuracy vs. parameter efficiency with supervised pre-trained ViT-B/16 backbone on VTAB-1k [60]. SPT variants perform favorably against the other PEFT approaches and are more scalable. (b) Applying other PEFT structured tuning methods [22, 20, 21] to the task-specific sensitive weight matrices (denoted by TSM) identified by our criterion with supervised pre-trained ViT-B/16 backbone on VTAB-1k. Our criterion brings consistent performance gain. (c) Domain vs. performance gaps for different dataset groups in VTAB-1k [60]. The blue bars show the domain gaps between the source domain (ImageNet [25]) and target domains, which are measured by Maximum Mean Discrepancy (MMD) distance [49]. The red line represents the performance gaps between SPT-ADAPTER w/o unstructured and w/o structured, using supervised pre-trained ViT-B/16 backbone. The dataset groups are Natural, Specialized, and Structured. Structured tuning is important for achieving good performance on Structured datasets with larger domain gaps." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.402, + 0.47, + 0.468 + ], + "angle": 0, + "content": "
MethodTuned / TotalNaturalSpecializedStructuredMean / Acc.
ADAPTER-80.23%79.084.158.573.9
SPT-ADAPTER w/o unstructured0.29%81.285.160.375.5
SPT-ADAPTER w/o structured0.34%81.285.059.675.3
SPT-ADAPTER0.30%81.385.360.875.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.478, + 0.47, + 0.553 + ], + "angle": 0, + "content": "Table 6: Ablation study on structured and unstructured tuning only with supervised pre-trained ViT-B/16 backbone. Top-1 accuracy \\((\\%)\\) is reported. We set different parameter constraints to align the fractions of trainable parameters for these cases. The best result is in bold" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.469, + 0.614 + ], + "angle": 0, + "content": "that incorporating both structured and unstructured tuning at task-specific important positions achieves the highest performance on all dataset groups." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.469, + 0.795 + ], + "angle": 0, + "content": "Effect of number of training samples \\( C \\) to get parameter sensitivity. We investigate the effect of the number of training images \\( C \\) for calculating our parameter sensitivity (Algorithm 1 of the main paper). We randomly sample training samples and report the mean results over three runs in Table 8. We find that our SPT is robust to the number of training samples \\( C \\) and randomly sampling 400 out of a total of 800 training samples is sufficient to obtain accurate task-specific important positions, e.g., calculating the sensitivity for ViT-B/16 backbone takes only 5.5 seconds with a single GPU on any of the VTAB-1k datasets and this computation is required only once." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Computational cost analysis. We investigate the computational cost of SPT-LoRA by comparing with full fine-tuning, addition-based method PROMPT-DEEP, and reparameterization-based method LORA-16. The results are presented in Table 7. We observe that PROMPT-DEEP has higher inference latency and inference GPU memory due to the additional prompts. In contrast, since the up" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.402, + 0.892, + 0.473 + ], + "angle": 0, + "content": "
MethodInference Latency (ms/img)Inference Memory (GB)Fine-tuning Memory (GB)
FULL2.81.311.9
PROMPT-DEEP3.81.913.2
LORA-162.81.38.2
SPT-LORA w/o unstructured2.81.38.3
SPT-LORA2.81.39.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.478, + 0.893, + 0.568 + ], + "angle": 0, + "content": "Table 7: Cost comparisons with ViT-B/16 backbone on the VTAB-1k [60] benchmark with around \\(0.70\\%\\) fractions of the trainable parameters. We report the latency (ms/img) and the peak memory usage (GB) with image resolution \\(224 \\times 224\\) on a single GeForce 3090 GPU. The best result is in bold." + }, + { + "type": "table", + "bbox": [ + 0.577, + 0.591, + 0.814, + 0.632 + ], + "angle": 0, + "content": "
C240400560800
Mean Acc.76.376.476.476.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.637, + 0.892, + 0.698 + ], + "angle": 0, + "content": "Table 8: Effect of the number of training samples used to get the sensitivity for SPT-LoRA with supervised pre-trained ViT-B/16 backbone on VTAB-1k [60]. Top-1 accuracy (%) is reported. The best result is in bold." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.721, + 0.893, + 0.903 + ], + "angle": 0, + "content": "dated parameters after fine-tuning can be reparameterized and merged into the pre-trained model, our SPT-LoRA, SPT-LoRA w/o unstructured tuning, and LoRA-16 are more efficient than PROMPT-DEEP during inference. However, we observe that our SPT-LoRA has slightly higher fine-tuning memory than the full fine-tuning and LoRA-16 which is taken up by updating the unstructurally-tuned parameters with sparse gradients in Eq. (5). Additionally, for memory-intensive scenarios, one can employ SPT-LoRA w/o unstructured tuning for improved performance (0.8% Top-1 accuracy on VTAB-1k) and similar fine-tuning memory as LoRA, as shown in Figure 3 (b) and Table 6." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11832" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.096, + 0.287, + 0.107 + ], + "angle": 0, + "content": "Supervised pre-trained ViT-B" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.109, + 0.337, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.437, + 0.096, + 0.545, + 0.107 + ], + "angle": 0, + "content": "MAE pre-trained ViT-B" + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.109, + 0.609, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.709, + 0.096, + 0.837, + 0.107 + ], + "angle": 0, + "content": "MoCo V3 pre-trained ViT-B" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.109, + 0.884, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.263, + 0.337, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.263, + 0.609, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.261, + 0.884, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.414, + 0.893, + 0.462 + ], + "angle": 0, + "content": "Figure 4: Parameter sensitivity patterns under 0.4M trainable parameter budget for ViT-B/16 backbone with different pretraining strategies on three sample tasks from VTAB-1k [61]. The proportions exhibit task-specific varying patterns in terms of network depth (upper figures) and task-agnostic similar patterns in terms of operations (lower figures)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.473, + 0.392, + 0.489 + ], + "angle": 0, + "content": "4.4. Observations on Sensitivity Patterns" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Our sensitivity criterion identifies task-specific important positions, which can reveal the contributions of the pretrained weights to different downstream tasks during transfer learning. We visualize the proportions of the sensitive parameters for the supervised pre-trained ViT-B/16 backbone under 0.4M trainable parameter budget in Figure 4. First, we investigate the most sensitive blocks, whose numbers of sensitive parameters are summed and normalized over the 12 ViT-B/16 blocks. We observe that the patterns of the sensitive parameter proportions vary markedly across different tasks, which echoes the observations made in [16]. This suggests that we should not introduce trainable parameters to the same positions for each individual task but allocate trainable parameters at task-specific ones as we proposed. Next, we investigate the most insensitive weight matrices within a block. A ViT block consists of a query \\( W_{q} \\), a key \\( W_{k} \\), a value \\( W_{v} \\), and an output \\( W_{o} \\) weight matrices in the multi-head self-attention layer and two weight matrices \\( W_{fc1} \\) and \\( W_{fc2} \\) in the feed-forward network as elaborated in [51, 12]. We observe that the query \\( W_{q} \\) and key \\( W_{k} \\) weight matrices have the lowest proportions of sensitive parameters for all three sample tasks. Since \\( W_{q} \\) and \\( W_{k} \\) are responsible for learning the attention scores which indicate the pairwise similarity among the patches, we speculate that although domain changes, the patch relationships learned during pre-training can be efficiently reused when transferred to downstream classification tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.472, + 0.619, + 0.488 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.812 + ], + "angle": 0, + "content": "In this paper, we have explored identifying and allocating trainable parameters to task-specific important positions for visual parameter-efficient tuning. Specifically, we have proposed a novel criterion to quickly measure the sensitivity of the pre-trained parameters for each specific task before fine-tuning. Based on the parameter sensitivity, we have proposed a trainable parameter allocation strategy that adaptively combines both unstructured and structured tuning under a desired trainable parameter budget, enabling high representational capability and flexibility. Finally, we have conducted extensive experiments on a total of 24 downstream recognition tasks with both plain and hierarchical vision Transformer backbones under different pre-training strategies to demonstrate the versatility and effectiveness of our proposed SPT. Notably, we have shown that our approach is complementary to the existing PEFT methods and improves their performance significantly. In the future, we will explore adapting large vision models to more downstream tasks with SPT, e.g., dense prediction and vision-and-language tasks, and improve the training efficiency of SPT for on-device training [6, 30]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.812, + 0.892, + 0.872 + ], + "angle": 0, + "content": "Acknowledgement. We thank Jing liu and Ziyi Liu for their helpful discussions. This research is partially supported by Monash FIT Start-up Grant. Dr. Jing Zhang is supported by the Australian Research Council project FL-170100117." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11833" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Y. Bao, Y. Li, S.-L. Huang, L. Zhang, L. Zheng, A. Zamir, and L. Guibas. An information-theoretic approach to transferability in task transfer learning. In ICIP, pages 2309-2313. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.178, + 0.471, + 0.22 + ], + "angle": 0, + "content": "[2] G. Bender, P.-J. Kindermans, B. Zoph, V. Vasudevan, and Q. Le. Understanding and simplifying one-shot architecture search. In ICML, pages 550-559. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.225, + 0.469, + 0.266 + ], + "angle": 0, + "content": "[3] A. Brock, T. Lim, J. M. Ritchie, and N. Weston. Smash: one-shot model architecture search through hypernetworks. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[4] S. Caelles, K.-K. Maninis, J. Pont-Tuset, L. Leal-Taixe, D. Cremers, and L. Van Gool. One-shot video object segmentation. In CVPR, pages 221-230, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.321, + 0.468, + 0.361 + ], + "angle": 0, + "content": "[5] H. Cai, C. Gan, T. Wang, Z. Zhang, and S. Han. Once-for-all: Train one network and specialize it for efficient deployment. In ICLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.469, + 0.409 + ], + "angle": 0, + "content": "[6] H. Cai, C. Gan, L. Zhu, and S. Han. Tinytl: Reduce memory, not parameters for efficient on-device learning. NeurIPS, 33:11285-11297, 2020. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.469, + 0.457 + ], + "angle": 0, + "content": "[7] H. Cai, L. Zhu, and S. Han. ProxylessNAS: Direct neural architecture search on target task and hardware. In ICLR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.464, + 0.469, + 0.505 + ], + "angle": 0, + "content": "[8] N. S. Chatterji, B. Neyshabur, and H. Sedghi. The intriguing role of module criticality in the generalization of deep networks. In ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.512, + 0.469, + 0.553 + ], + "angle": 0, + "content": "[9] S. Chen, C. Ge, Z. Tong, J. Wang, Y. Song, J. Wang, and P. Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 2022. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.559, + 0.469, + 0.6 + ], + "angle": 0, + "content": "[10] X. Chen, S. Xie, and K. He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[11] Y. Cui, Y. Song, C. Sun, A. Howard, and S. Belongie. Large scale fine-grained categorization and domain-specific transfer learning. In CVPR, pages 4109-4118, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.655, + 0.469, + 0.723 + ], + "angle": 0, + "content": "[12] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 4, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[13] T. Gebru, J. Krause, Y. Wang, D. Chen, J. Deng, and L. Fei-Fei. Fine-grained car detection for visual census estimation. In AAAI, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.778, + 0.469, + 0.805 + ], + "angle": 0, + "content": "[14] D. Guo, A. Rush, and Y. Kim. Parameter-efficient transfer learning with diff pruning. In ACL-IJCNLP, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.812, + 0.469, + 0.853 + ], + "angle": 0, + "content": "[15] Y. Guo, Y. Li, L. Wang, and T. Rosing. Adafilter: Adaptive filter fine-tuning for deep transfer learning. In AAAI, pages 4060-4066, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[16] Y. Guo, H. Shi, A. Kumar, K. Grauman, T. Rosing, and R. Feris. Spottune: transfer learning through adaptive fin-tuning. In CVPR, pages 4805-4814, 2019. 3, 9" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[17] T. Hao, H. Chen, Y. Guo, and G. Ding. Consolidator: Mergable adapter with group connections for vision transformer. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.139, + 0.892, + 0.18 + ], + "angle": 0, + "content": "[18] J. He, C. Zhou, X. Ma, T. Berg-Kirkpatrick, and G. Neubig. Towards a unified view of parameter-efficient transfer learning. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.185, + 0.892, + 0.226 + ], + "angle": 0, + "content": "[19] K. He, X. Chen, S. Xie, Y. Li, P. Dollár, and R. Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.231, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[20] N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. De Laroussilhe, A. Gesmundo, M. Attariyan, and S. Gelly. Parameter-efficient transfer learning for nlp. In ICML, pages 2790-2799, 2019. 1, 2, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[21] E. J. Hu, yelong shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 1, 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.35, + 0.892, + 0.391 + ], + "angle": 0, + "content": "[22] M. Jia, L. Tang, B.-C. Chen, C. Cardie, S. Belongie, B. Hariharan, and S.-N. Lim. Visual prompt tuning. In ECCV, 2022. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.396, + 0.892, + 0.437 + ], + "angle": 0, + "content": "[23] S. Jie and Z.-H. Deng. Convolutional bypasses are better vision transformer adapters. arXiv preprint arXiv:2207.07039, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[24] A. Khosla, N. Jayadevaprakash, B. Yao, and L. Fei-Fei. Novel dataset for fine-grained image categorization. In CVPRW, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.488, + 0.892, + 0.528 + ], + "angle": 0, + "content": "[25] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[26] A. Kumar, A. Raghunathan, R. M. Jones, T. Ma, and P. Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In ICLR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.579, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[27] B. Lester, R. Al-Rfou, and N. Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.611, + 0.892, + 0.653 + ], + "angle": 0, + "content": "[28] X. L. Li and P. Liang. Prefix-tuning: Optimizing continuous prompts for generation. In IJCNLP, pages 4582-4597. Association for Computational Linguistics, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.698 + ], + "angle": 0, + "content": "[29] D. Lian, D. Zhou, J. Feng, and X. Wang. Scaling & shifting your features: A new baseline for efficient model tuning. NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[30] J. Lin, L. Zhu, W.-M. Chen, W.-C. Wang, C. Gan, and S. Han. On-device training under 256kb memory. arXiv preprint arXiv:2206.15472, 2022.9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.892, + 0.776 + ], + "angle": 0, + "content": "[31] H. Liu, K. Simonyan, and Y. Yang. DARTS: Differentiable architecture search. In ICLR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.781, + 0.892, + 0.823 + ], + "angle": 0, + "content": "[32] Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.827, + 0.892, + 0.868 + ], + "angle": 0, + "content": "[33] Z. Liu, H. Mao, C.-Y. Wu, C. Feichtenhofer, T. Darrell, and S. Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[34] I. Loshchilov and F. Hutter. Fixing weight decay regularization in adam, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11834" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.135 + ], + "angle": 0, + "content": "[35] P. Molchanov, A. Mallya, S. Tyree, I. Frosio, and J. Kautz. Importance estimation for neural network pruning. In CVPR, pages 11264-11272, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.138, + 0.472, + 0.182 + ], + "angle": 0, + "content": "[36] M. M. Naseer, K. Ranasinghe, S. H. Khan, M. Hayat, F. Shahbaz Khan, and M.-H. Yang. Intriguing properties of vision transformers. *NeurIPS*, 34:23296-23308, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.185, + 0.47, + 0.226 + ], + "angle": 0, + "content": "[37] B. Neyshabur, H. Sedghi, and C. Zhang. What is being transferred in transfer learning? NeurIPS, 33:512-523, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.232, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[38] C. Nguyen, T. Hassner, M. Seeger, and C. Archambeau. Leep: A new measure to evaluate transferability of learned representations. In ICML, pages 7294-7305. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.47, + 0.334 + ], + "angle": 0, + "content": "[39] M.-E. Nilsback and A. Zisserman. Automated flower classification over a large number of classes. In ICVGIP, pages 722-729. IEEE, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.339, + 0.47, + 0.38 + ], + "angle": 0, + "content": "[40] J. Pleted and T. Gedeon. Deep transfer learning for image classification: a survey. arXiv preprint arXiv:2205.09904, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.47, + 0.441 + ], + "angle": 0, + "content": "[41] J. Puigcerver, C. Riquelme, B. Mustafa, C. Renggli, A. S. Pinto, S. Gelly, D. Keysers, and N. Houlsby. Scalable transfer learning with expert models. arXiv preprint arXiv:2009.13239, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[42] M. Raghu, T. Unterthiner, S. Kornblith, C. Zhang, and A. Dosovitskiy. Do vision transformers see like convolutional neural networks? NeurIPS, 34:12116-12128, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.506, + 0.47, + 0.547 + ], + "angle": 0, + "content": "[43] M. T. Rosenstein, Z. Marx, L. P. Kaelbling, and T. G. Dietterich. To transfer or not to transfer. In NeurlPSW, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.47, + 0.581 + ], + "angle": 0, + "content": "[44] R. K. Srivastava, K. Greff, and J. Schmidhuber. Training very deep networks. NeurIPS, 28, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.585, + 0.47, + 0.627 + ], + "angle": 0, + "content": "[45] X. Sun, R. Panda, R. Feris, and K. Saenko. Adashare: Learning what to share for efficient deep multi-task learning. NeurIPS, 33:8728-8740, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.47, + 0.674 + ], + "angle": 0, + "content": "[46] Y.-L. Sung, J. Cho, and M. Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. arXiv preprint arXiv:2206.06522, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.47, + 0.719 + ], + "angle": 0, + "content": "[47] A. T. Tran, C. V. Nguyen, and T. Hassner. Transferability and hardness of supervised classification tasks. In ICCV, pages 1395–1405, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.725, + 0.47, + 0.781 + ], + "angle": 0, + "content": "[48] C.-H. Tu, Z. Mai, and W.-L. Chao. Visual query tuning: Towards effective usage of intermediate representations for parameter and memory efficient transfer learning. arXiv preprint arXiv:2212.03220, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.785, + 0.469, + 0.827 + ], + "angle": 0, + "content": "[49] E. Tzeng, J. Hoffman, N. Zhang, K. Saenko, and T. Darrell. Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:1412.3474, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[50] G. Van Horn, S. Branson, R. Farrell, S. Haber, J. Barry, P. Ipeirotis, P. Perona, and S. Belongie. Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In CVPR, pages 595–604, 2015. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.135 + ], + "angle": 0, + "content": "[51] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. NeurIPS, pages 5998-6008, 2017. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.139, + 0.893, + 0.182 + ], + "angle": 0, + "content": "[52] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The caltech-ucsd birds-200-2011 dataset. Tech. Rep. CNS-TR-2011-001, California Institute of Technology, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.186, + 0.893, + 0.226 + ], + "angle": 0, + "content": "[53] Z. Wang, Z. Dai, B. Póczos, and J. Carbonell. Characterizing and avoiding negative transfer. In CVPR, pages 11293-11302, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.232, + 0.893, + 0.274 + ], + "angle": 0, + "content": "[54] H. Wu, J. Xu, J. Wang, and M. Long. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. NeurIPS, 34:22419-22430, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.279, + 0.893, + 0.321 + ], + "angle": 0, + "content": "[55] R. Xu, F. Luo, Z. Zhang, C. Tan, B. Chang, S. Huang, and F. Huang. Raise a child in large language model: Towards effective and generalizable fine-tuning. In EMNLP, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.326, + 0.893, + 0.367 + ], + "angle": 0, + "content": "[56] Y. Xu, Q. Zhang, J. Zhang, and D. Tao. Vitae: Vision transformer advanced by exploring intrinsic inductive bias. In NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.893, + 0.413 + ], + "angle": 0, + "content": "[57] J. Yoon, S. Arik, and T. Pfister. Data valuation using reinforcement learning. In ICML, pages 10842-10851. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.893, + 0.46 + ], + "angle": 0, + "content": "[58] J. Yosinski, J. Clune, Y. Bengio, and H. Lipson. How transferable are features in deep neural networks? NeurIPS, 27, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.465, + 0.893, + 0.507 + ], + "angle": 0, + "content": "[59] E. B. Zaken, Y. Goldberg, and S. Ravfogel. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In ACL, pages 1-9, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.893, + 0.581 + ], + "angle": 0, + "content": "[60] X. Zhai, J. Puigcerver, A. Kolesnikov, P. Ruyssen, C. Riquelme, M. Lucic, J. Djolonga, A. S. Pinto, M. Neumann, A. Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.585, + 0.893, + 0.655 + ], + "angle": 0, + "content": "[61] X. Zhai, J. Puigcerver, A. Kolesnikov, P. Ruyssen, C. Riquelme, M. Lucic, J. Djolonga, A. S. Pinto, M. Neumann, A. Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.893, + 0.688 + ], + "angle": 0, + "content": "[62] C. Zhang, S. Bengio, and Y. Singer. Are all layers created equal? arXiv preprint arXiv:1902.01996, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.693, + 0.893, + 0.721 + ], + "angle": 0, + "content": "[63] Y. Zhang, K. Zhou, and Z. Liu. Neural prompt search. arXiv preprint arXiv:2206.04673, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.725, + 0.893, + 0.767 + ], + "angle": 0, + "content": "[64] M. Zhao, T. Lin, F. Mi, M. Jaggi, and H. Schütze. Masking as an efficient alternative to finetuning for pretrained language models. In EMNLP, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.772, + 0.893, + 0.828 + ], + "angle": 0, + "content": "[65] S. Zheng, J. Lu, H. Zhao, X. Zhu, Z. Luo, Y. Wang, Y. Fu, J. Feng, T. Xiang, P. H. Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.893, + 0.874 + ], + "angle": 0, + "content": "[66] B. Zhou, H. Zhao, X. Puig, S. Fidler, A. Barriuso, and A. Torralba. Scene parsing through ade20k dataset. In CVPR, pages 633-641, 2017. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.874 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11835" + } + ] +] \ No newline at end of file diff --git a/2023/Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning/219bbdca-cfe8-433f-a6db-97c4773a1a10_origin.pdf b/2023/Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning/219bbdca-cfe8-433f-a6db-97c4773a1a10_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..70898d74b6a629770ed74e759edc05f7394aeebb --- /dev/null +++ b/2023/Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning/219bbdca-cfe8-433f-a6db-97c4773a1a10_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb188c8a715c96216cac39074a41faf58c0b657c84c88c87966f5a9c73c47a49 +size 896636 diff --git a/2023/Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning/full.md b/2023/Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..445cf5d3d90b36a2306d830229850cb1546a71f1 --- /dev/null +++ b/2023/Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning/full.md @@ -0,0 +1,315 @@ +# Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning + +Haoyu He $^{1}$ Jianfei Cai $^{1}$ Jing Zhang $^{2}$ Dacheng Tao $^{2}$ Bohan Zhuang $^{1\dagger}$ + +1 ZIP Lab, Monash University 2 The University of Sydney + +# Abstract + +Visual Parameter-Efficient Fine-Tuning (PEFT) has become a powerful alternative for full fine-tuning so as to adapt pre-trained vision models to downstream tasks, which only tunes a small number of parameters while freezing the vast majority ones to ease storage burden and optimization difficulty. However, existing PEFT methods introduce trainable parameters to the same positions across different tasks depending solely on human heuristics and neglect the domain gaps. To this end, we study where to introduce and how to allocate trainable parameters by proposing a novel Sensitivity-aware Visual Parameter-efficient fine-Tuning (SPT) scheme, which adaptively allocates trainable parameters to task-specific important positions given a desired tunable parameter budget. Specifically, our SPT first quickly identifies the sensitive parameters that require tuning for a given task in a data-dependent way. Next, our SPT further boosts the representational capability for the weight matrices whose number of sensitive parameters exceeds a pre-defined threshold by utilizing existing structured tuning methods, e.g., LoRA [21] or Adapter [20], to replace directly tuning the selected sensitive parameters (unstructured tuning) under the budget. Extensive experiments on a wide range of downstream recognition tasks show that our SPT is complementary to the existing PEFT methods and largely boosts their performance, e.g., SPT improves Adapter with supervised pre-trained ViT-B/16 backbone by $4.2\%$ and $1.4\%$ mean Top-1 accuracy, reaching SOTA performance on FGVC and VTAB-1k benchmarks, respectively. Source code is at https://github.com/ziplab/SPT. + +# 1. Introduction + +To effectively adapt the pre-trained representations to the downstream tasks, the de-facto choice is full fine-tuning, which initializes the model with the pre-trained weights and tunes all the parameters. However, vanilla full fine-tuning needs to store a separate instance of parameters for each + +![](images/4f12a796781166c998a97d80560c43c3045d884ad14a5de144773c620fb5fb8e.jpg) +Figure 1: (a) Existing PEFT methods, such as Adapter [20] introduce trainable parameters to the same positions for all downstream tasks. These methods design task-agnostic positions to employ trainable parameters relying on heuristics and neglect consideration of the distinct domain gaps and characteristics for the downstream tasks. (b) Our Sensitivity-aware visual Parameter-efficient fine-Tuning (SPT) introduces trainable parameters to the task-specific important positions and allocates them with both unstructured and structured tuning granularities, simultaneously. For structured tuning, SPT can exploit any existing structured tuning methods, such as LoRA [21] or Adapter [20]. Red lines and blocks represent trainable parameters and modules, while blue lines represent frozen parameters. + +task and each deployment scenario. It can be extremely storage-intensive as the storage cost grows linearly with the number of possible cases, considering there are vast varieties of downstream tasks and dynamic deployment environments, especially when deploying the large vision models [12, 32, 56] to mobile systems. For example, even storing a single large pre-trained ViT-H [12] model on a local disk requires at least 2.3GB, while the Top-10 U.S. apps required only collectively 2.2GB in May 2021. $^{1}$ + +Notably, an emerging solution is to replace vanilla fine-tuning with visual Parameter-Efficient Fine-Tuning (PEFT) [22, 9, 63, 23], which only tunes a small num + +ber of trainable parameters while freezing the vast majority ones that are shared by multiple tasks. As PEFT approaches exhibit less than $1\%$ of trainable parameters, the storage burden is largely alleviated. Another attractive property of PEFT is that tuning fewer parameters eases the optimization difficulty and mitigates the overfitting issue when adapting large pre-trained models on the target dataset, thereby achieving comparable or even better performance than vanilla fine-tuning [22]. Although promising, the existing PEFT approaches introduce trainable parameters to the same positions for all downstream tasks, relying on human heuristics and neglecting task-specific domain gaps and characteristics, which limits their performance. For instance, in a task-agnostic manner, Prompt Tuning-deep [22] and Adapter [20] respectively add trainable parameters to multi-head self-attention and feed-forward network layers for all distinct tasks as depicted in Figure 1 (a). + +To address this fundamental challenge, we explore where to introduce and how to allocate trainable parameters under a desired parameter budget by presenting a novel Sensitivity-aware visual Parameter-efficient fine-Tuning (SPT) scheme that identifies the task-specific important positions to adaptively allocate trainable parameters. Since the pre-trained weights at distinct positions have varying contributions for different downstream tasks [58, 26, 37], we first propose a new criterion to quickly identify the task-specific sensitive parameters that require tuning in a data-dependent way. Inspired by model pruning metrics [44, 35, 2, 3], we propose to measure the parameter sensitivity with the loss reduction when being tuned, which can be approximated by a first-order Taylor expansion derived within a single forward and backward pass ahead of fine-tuning in one-shot. Our sensitivity criterion is simple and effective, which can identify the task-specific important positions to introduce trainable parameters for any backbone quickly. For instance, calculating the sensitivity for ViT-B/16 backbone takes only 5.5 seconds with a single GPU on any of the VTAB-1k datasets. + +With our criterion, we empirically observe that the proportions of the sensitivity parameters for each block indeed vary markedly across different tasks in Section 4.4. To allocate the trainable parameters under a desired trainable parameter budge, an intuitive solution is to directly tune the most sensitive weight connections, which we name unstructured tuning. Despite its simplicity and flexibility, unstructured tuning only tunes a few parameters which still lacks representational capability and is challenging to bridge the domain gap. To this end, we propose to further incorporate structured tuning to replace unstructured tuning at the sensitive weight matrices whose numbers of sensitive parameters exceed a pre-defined threshold to improve the representational capability under a similar parameter budget. Structured tuning can be implemented by any parameter-efficient + +structured tuning methods [21, 9, 23, 22] that directly adjust the hidden representations, e.g., inserting an adapter module sequentially after the sensitive weight matrices. Therefore, our SPT adaptively combines both unstructured and structured tuning granularity and allocates trainable parameters with high flexibility and representational capability for each distinct downstream task. + +This paper has the following key contributions. 1) We make the pioneering exploration to identify the task-specific important positions under the PEFT setting, which is fast, effective, versatile to be applied to various backbones with different pre-training strategies, and orthogonal to the existing PEFT methods. 2) Based on the sensitivity criterion, we propose a trainable parameter allocation strategy that adaptively combines both unstructured and structured tuning under a desired parameter budget to achieve high flexibility, large capacity, and favorable trade-off between parameter efficiency and accuracy. 3) Extensive experiments on a total of 24 downstream recognition tasks with both plain and hierarchical vision Transformer backbones under supervised and self-supervised pre-trainings show that our SPT is complementary to the existing PEFT methods and boosts their performance by large margins. For instance, SPT improves Adapter [20] by $4.2\%$ mean Top-1 accuracy, outperforming the SOTA PEFT methods on the FGVC benchmark. + +# 2. Related Work + +Parameter-efficient fine-tuning. Full fine-tuning is the most predominant approach when adapting a large-scale pre-trained model to downstream tasks, where the model is initialized from the pre-trained weights with all parameters trainable. Yet, when a model becomes larger, parameter-efficient fine-tuning [27, 28] is highly desirable, which tunes only a tiny portion of parameters to alleviate the storage burden. The general PEFT approaches can be categorized into addition-based PEFT methods and reparameterization-based PEFT methods. + +Addition-based PEFT attaches additional trainable parameters to the backbone and only tunes these parameters. Apart from Prompt tuning [22] and Adapter [20], recent addition-based methods study connecting or combining existing PEFT methods. For instance, He et al. [18] connect Prompt tuning and Adapter and provide a unified view that all PEFT approaches share the same design to adjust the hidden representations. Zhang et al. [63] search for the optimal configurations to combine multiple PEFT approaches following once-for-all scheme [5, 54]. Since the additional parameters require extra computations compared to full fine-tuning, a few recent works [46, 48] design specific architectures to avoid storing the intermediate activations, thereby alleviating the fine-tuning memory cost. However, it is noteworthy that enhancing training efficiency is not the primary objective of our work. + +Reparameterization-based PEFT aims to avoid extra computational costs by tuning parameters that are inherently in or can be reparameterized into the backbone during inference. Prior works select the parameters that are inherently in the backbone, including the bias terms [59], the last several layers [58, 4], and weight connections [14, 64]. To reparameterize new parameters into the backbone [17, 29], representative work LoRA [21] optimizes two low-rank matrices which can be further merged into the weight matrices. In contrast to the aforementioned works, we argue the importance of tuning parameters at task-specific important positions and quickly identify them with our proposed parameter sensitivity criterion before tuning, which is complementary to and provides valuable guidance for the existing PEFT methods. Moreover, our SPT can also be inference-efficient when implementing structured tuning with any reparameterization-based structured tuning method. Recently, SSF [29] is proposed to introduce trainable scaling and shifting parameters that can be absorbed into the previous linear layers. However, it cannot scale to higher trainable parameter budgets and requires a complex and time-consuming hyper-parameter search for learning rate, weight decay, and drop-path rate on each individual dataset, thus is not directly comparable to our method. + +Task-specific transfer learning. The effectiveness of transferring pre-trained models to downstream tasks strongly depends on the relationship between the source and target tasks [43, 53, 26, 40]. This has motivated the community to explore the optimal pre-training data [11, 57], model [47, 38], and weights [16, 55] for the target task. To seek suitable task-specific pre-training data, Cui et al. [11] select the source domain data from the top-k most similar classes measured by Earth Mover's Distance; Yoon et al. [57] weight each class in the source domain with reinforcement learning; and Puigcerver et al. [41] first train a diverse set of experts and then select the most relevant expert for each target task. Another line of work selects a suitable pre-trained model for the target task ahead of fine-tuning by measuring the transferability of pre-trained models to the target domain with interclass covariance between the source data and target classes [1] or conditional cross-entropy [47] between the source and target labels. Considering the transferability of the feature representations at distinct layers for the same pre-trained model is different [58, 37], recent works [15, 45] endeavour transfer task-specific weights by freezing some pre-trained weights and fine-tuning the rest. For example, the task-specific fine-tuned weights are selected by learning a policy network with Gumbel-Softmax [16], optimizing a sparse mask with $L_{0}$ norm [14], and learning binary gates for each parameter [64]. Our SPT also adaptively selects task-specific parameters. In contrast to the previous work, we 1) derive task-specific important positions prior to fine-tuning with only a single + +forward and backward pass, which is computationally efficient; 2) mask the gradients for insensitive parameters in unstructured tuning with fixed binary masks, thereby having more affordable fine-tuning memory than optimizing learnable binary masks in [14, 64]. Moreover, we are pioneering work to adaptively allocate task-specific trainable parameters with both fine-grained unstructured and coarse-grained structured tuning granularities to achieve both high flexibility and representational capability. + +# 3. Method + +Our sensitivity-aware visual parameter-efficient fine-tuning consists of two stages. In the first stage, SPT measures the task-specific sensitivity for the pre-trained parameters (Section 3.1). Based on the parameter sensitivity and a given parameter budget, SPT then adaptively allocates trainable parameters to task-specific important positions (Section 3.2). + +# 3.1. Task-specific Parameter Sensitivity + +Recent research has observed that pre-trained backbone parameters exhibit varying feature patterns [42, 36] and criticality [62, 8] at distinct positions. Moreover, when transferred to downstream tasks, their efficacy varies depending on how much pre-trained features are reused and how well they adapt to the specific domain gap [58, 26, 37]. Motivated by these observations, we argue that not all parameters contribute equally to the performance across different tasks in PEFT and propose a new criterion to measure the sensitivity of the parameters in the pre-trained backbone for a given task. + +Specifically, given the training dataset $\mathcal{D}_t$ for the $t$ -th task and the pre-trained model weights $\boldsymbol{w} = \{w_1, w_2, \dots, w_N\} \in \mathbb{R}^N$ where $N$ is the total number of parameters, the objective for the task is to minimize the empirical risk: $\min_{\boldsymbol{w}} E(\mathcal{D}_t, \boldsymbol{w})$ . We denote the parameter sensitivity set as $\mathcal{S} = \{s_1, \dots, s_N\}$ and the sensitivity $s_n$ for parameter $w_n$ is measured by the empirical risk difference when tuning it: + +$$ +s _ {n} = E \left(\mathcal {D} _ {t}, \boldsymbol {w}\right) - E \left(\mathcal {D} _ {t}, \boldsymbol {w} \mid w _ {n} = w _ {n} ^ {*}\right), \tag {1} +$$ + +where $w_{n}^{*} = \operatorname*{argmin}_{w_{n}}(E(\mathcal{D}_{t},\boldsymbol {w}))$ . We can reparameterize the tuned parameters as $w_{n}^{*} = w_{n} + \Delta_{w_{n}}$ , where $\Delta_{w_n}$ denotes the update for $w_{n}$ after tuning. Here we individually measure the sensitivity of each parameter, which is reasonable given that most of the parameters are frozen during fine-tuning in PEFT. However, it is still computationally intensive to compute Eq. (1) for two reasons. Firstly, getting the empirical risk for $N$ parameters requires forwarding the entire network $N$ times, which is time-consuming. Secondly, it is challenging to derive $\Delta_{w_n}$ , as we have to tune each individual $w_{n}$ until convergence. + +![](images/e0de3fc7f9016d1bcc965bd66924f89237eb5daa10b3805b7d071e13ff65f082.jpg) +Figure 2: Overview of our trainable parameter allocation strategy. With the parameter sensitivity set $S$ , we first get the top- $\tau$ sensitive parameters. Instead of directly tuning these sensitive parameters, we also boost the representational capability by replacing unstructured tuning with structured tuning at sensitive weight matrices that have a large number of sensitive parameters, which can be implemented by an existing structured tuning method, e.g., LoRA [21] and Adapter [20]. Red lines and blocks represent trainable parameters and modules, while blue lines represent frozen parameters. + +Algorithm 1 Computing task-specific parameter sensitivities +Input: Pre-trained model with network parameters $\pmb{w}$ training set $\mathcal{D}_t$ for the $t$ -th task, and number of training samples $C$ used to calculate the parameter sensitivities +Output: Sensitivity set $\mathcal{S} = \{s_1,\dots,s_N\}$ +Initialize $\mathcal{S} = \{0\}^N$ +for $i\in \{1,\ldots ,C\}$ do + Get the $i$ -th training sample of $\mathcal{D}_t$ + Compute loss $E$ + Compute gradients $\pmb{g}$ + for $n\in \{1,\dots,N\}$ do + Update sensitivity for the $n$ -th parameter: $s_n = s_n + g_n^2$ + end for +end for + +To overcome the first barrier, we simplify the empirical loss by approximating $s_n$ in the vicinity of $w$ by its first-order Taylor expansion + +$$ +s _ {n} ^ {(1)} = - g _ {n} \Delta_ {w _ {n}}, \tag {2} +$$ + +where the gradients $\pmb{g} = \partial E / \partial \pmb{w}$ , and $g_{n}$ is the gradient of the $n$ -th element of $\pmb{g}$ . To address the second barrier, following [31, 7], we take the one-step unrolled weight as the surrogate for $w_{n}^{*}$ and approximate $\Delta_{w_n}$ in Eq. (2) with a single step of gradient descent. We can accordingly get $s_n^{(1)}\approx g_n^2\epsilon$ , where $\epsilon$ is the learning rate. Since $\epsilon$ is the same for all parameters, we can eliminate it when comparing the sensitivity with the other parameters and finally get + +$$ +s _ {n} ^ {(1)} \approx g _ {n} ^ {2}. \tag {3} +$$ + +Therefore, the sensitivity of a parameter can be efficiently measured by its potential to reduce the loss on the target domain. Note that although our criterion draws inspiration from pruning work [35], it is distinct from it. [35] measures the parameter importance by the squared change in loss when removing them, i.e., $(E(\mathcal{D}_t,\boldsymbol {w}) - E(\mathcal{D}_t,\boldsymbol {w}\mid w_n = 0))^2$ and finally derives the parameter importance by $(g_{n}w_{n})^{2}$ , which is different from our formulations in Eqs. (1) and (3). + +In practice, we accumulate $\mathcal{S}$ from a total number of $C$ training samples ahead of fine-tuning to generate accu- + +rate sensitivity as shown in Algorithm 1, where $C$ is a predefined hyper-parameter. In Section 4.3, we show that employing only 400 training samples is sufficient for getting reasonable parameter sensitivity, which requires only 5.5 seconds with a single GPU for any VTAB-1k dataset with ViT-B/16 backbone [12]. + +# 3.2. Adaptive Trainable Parameters Allocation + +Our next step is to allocate trainable parameters based on the obtained parameter sensitivity set $S$ and a desired parameter budget $\tau$ . A straightforward solution is to directly tune the top- $\tau$ most sensitive unstructured connections (parameters) while keeping the rest frozen, which we name unstructured tuning. Specifically, we select the top- $\tau$ most sensitive weight connections in $S$ to form the sensitive weight connection set $\mathcal{T}$ . Then, for a weight matrix $\boldsymbol{W} \in \mathbb{R}^{d_{\mathrm{in}} \times d_{\mathrm{out}}}$ , we can get a binary mask $M \in \mathbb{R}^{d_{\mathrm{in}} \times d_{\mathrm{out}}}$ computed by + +$$ +\boldsymbol {M} ^ {j} = \left\{ \begin{array}{l l} 1 & \boldsymbol {W} ^ {j} \in \mathcal {T} \\ 0 & \boldsymbol {W} ^ {j} \notin \mathcal {T} \end{array} , \right. \tag {4} +$$ + +where $W^{j}$ and $M^j$ are the $j$ -th element in $W$ and $M$ , respectively. Accordingly, we can train the sensitive parameters by gradient descent and the updated weight matrix can be formulated as $W' \gets W - \epsilon g_W \odot M$ , where $g_W$ is the gradient for $W$ . + +However, considering PEFT approaches generally limit the proportion of trainable parameters to less than $1\%$ , tuning only a small number of unstructured weight connections might not have enough representational capability to handle the downstream datasets with large domain gaps from the source pre-training data. Therefore, to improve the representational capability, we propose to replace unstructured tuning with structured tuning at the sensitive weight matrices that have a high number of sensitive parameters. To preserve the parameter budget, we can implement structured tuning with an existing efficient structured tuning PEFT method [21, 9, 20, 23] that learns to directly adjust all hidden dimensions at once. We depict an overview of our trainable parameter allocation strategy in Figure 2. For example, we can employ the low-rank reparameterization trick LoRA [21] to the sensitive weight matrices and the one-step update for $\mathbf{W}$ can be formulated as + +$$ +\boldsymbol {W} ^ {\prime} = \left\{ \begin{array}{l l} \boldsymbol {W} + \boldsymbol {W} _ {\mathrm {d o w n}} \boldsymbol {W} _ {\mathrm {u p}} & \text {i f} \sum_ {j = 0} ^ {d _ {\mathrm {i n}} \times d _ {\mathrm {o u t}}} \boldsymbol {M} ^ {j} \geq \sigma_ {\mathrm {o p t}} \\ \boldsymbol {W} - \epsilon \boldsymbol {g} _ {\boldsymbol {W}} \odot \boldsymbol {M} & \text {o t h e r w i s e} \end{array} \right., +$$ + +where $W_{\mathrm{down}} \in \mathbb{R}^{d_{\mathrm{in}} \times r}$ and $W_{\mathrm{up}} \in \mathbb{R}^{r \times d_{\mathrm{out}}}$ are two learnable low-rank matrices to approximate the update of $W$ and rank $r$ is a hyper-parameter where $r \ll \min(d_{\mathrm{in}}, d_{\mathrm{out}})$ . In this way, we perform structured tuning on $W$ when its number of sensitive parameters exceeds $\sigma_{\mathrm{opt}}$ , whose value depends on the pre-defined type of structured tuning method. For example, since implementing structured tuning with LoRA requires $2 \times d_{\mathrm{in}} \times d_{\mathrm{out}} \times r$ trainable parameters for each sensitive weight matrix, we set $\sigma_{\mathrm{LoRA}} \gets 2 \times d_{\mathrm{in}} \times d_{\mathrm{out}} \times r$ to ensure that the number of trainable parameters introduced by structured tuning is always equal to or lower than the number of sensitive parameters. + +In this way, our SPT adaptively incorporates both structured and unstructured tuning granularities to enable higher flexibility and stronger representational power, simultaneously. In Section 4.3, we show that structured tuning is important for the downstream tasks with larger domain gaps and both unstructured and structured tuning contribute clearly to the superior performance of our SPT. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets and metrics. We evaluate our SPT on total 24 downstream tasks in two groups following [22]. 1) FGVC is a benchmark for fine-grained visual classification, including CUB-200-2011 [52], NABirds [50], Oxford Flowers [39], Stanford Cars [13], and Stanford Dogs [24] datasets. Each FGVC dataset contains between 55 to 200 classes and a few thousand images for train, validation, and test. We follow the validation splits in [22] if the validation set is unavailable. 2) VTAB-1k [60] is a large-scale transfer learning benchmark consisting of a collection of 19 visual classification tasks. VTAB-1k can further be divided into three groups, including Natural tasks with natural images, Specialized tasks with images captured by specialized equipment, e.g., medical images, and Structured tasks with images mostly generated from synthetic environments. Each of the VTAB-1k dataset has only 800 training and 200 validation samples, while the test set sizes vary. We use top-1 accuracy (\%) averaged within each group as our main metric following [22]. + +Pre-trained backbones. We conduct experiments on the plain vision Transformer backbone ViT-B/16 [12] that is pre-trained on ImageNet [25] with different pre-training strategies following [22], including supervised pre-training and self-supervised pre-training with MAE [19] and MoCo v3 [10] following [22]. We also conduct experiments on the representative hierarchical vision Transformer backbone Swin-B [32] under supervised pre-training. + +Contenders. We categorize the baseline methods into + +addition-based and reparameterization-based PEFT methods as introduced in Section 2. Unless specified, all baseline methods keep the backbone frozen. Addition-based methods require extra computations during inference, including MLP- $k$ , PROMPT-SHALLOW [22], PROMPT-DEEP [22], ADAPTER- $k$ [20], ADAPTFORMER [9], and NOAH [63]. Reparameterization-based methods have no additional computational overhead during inference, including LINEAR, PARTIAL- $k$ , BIAS [59], and LORA- $k$ [21]. Here $k$ represents the number of bottleneck dimension in ADAPTER- $k$ and LORA- $k$ . We also compare with full fine-tuning which is denoted by FULL. We introduce the details of these methods in the supplementary material. + +We also introduce two variants of our SPT: addition-based SPT-ADAPTER and reparameterization-based SPT-LORA. SPT-ADAPTER directly adjusts the hidden representations that are computed by sensitive weight matrices following [20], while SPT-LORA approximates updating the sensitive weight matrices following [21]. For the two variants, we follow the exact weight initializations that are described in [21] and follow [63] to set the bottleneck dimension as 8. + +Implementation details. Following [63], we use the AdamW optimizer [34] with cosine learning rate decay and set the batch size, learning rate, and weight decay as 64, $1 \times 10^{-3}$ , and $1 \times 10^{-4}$ , respectively. We also follow [63] for the standard data augmentation pipeline. We set the number of training samples $C$ used to calculate our parameter sensitivities in Algorithm 1 universally to be 800 for the main experiments. + +# 4.2. Main Results + +We compare our method with the baseline methods under different backbones, pre-training strategies, and tasks. + +Visual recognition with ViT backbone. First, our proposed SPT-ADAPTER and SPT-LORA achieve the best performance under different trainable parameter budgets with supervised pre-trained ViT-B/16 backbone, as shown in Table 1 and Figure 3 (a). For instance, SPT-ADAPTER outperforms the SOTA method NOAA by a clear margin of $0.9\%$ mean top-1 accuracy over the 19 VTAB-1k datasets with fewer trainable parameters. We speculate that our SPT variants allocate trainable parameters at task-specific positions compared to the heuristically selected positions in the baseline methods, which contributes to our superior performance. We also observe that our SPT-ADAPTER and SPT-LORA achieve large performance gains over ADAPTER and LORA variants, respectively. For example, SPT-ADAPTER and SPT-LORA with $0.41\%$ trainable parameters respectively improve ADAPTER-8 and LORA-8 significantly by $4.0\%$ and $3.3\%$ mean accuracy on the FGVC benchmark. This suggests that identifying task-specific important positions and combining both unstructured and structured tuning granularities with SPT are complementary to the exist + +
ViT-B/16 (85.8M)Total paramsFGVCVTAB-1k
Tuned / TotalMean Acc.Tuned / TotalNaturalSpecializedStructuredMean Acc.
FULL24.02×100%88.5100%75.983.447.669.0
+ +Addition-based methods + +
MLP-31.35×1.50%79.81.42%67.872.830.657.1
PROMPT-SHALLOW1.04×0.31%84.60.13%76.879.747.067.8
PROMPT-DEEP1.18×0.98%89.11.14%78.582.455.072.0
ADAPTER-81.06×0.39%85.50.23%79.084.158.573.9
ADAPTER-321.19×0.95%85.60.71%79.684.058.374.0
ADAPTFORMER1.09×0.44%85.10.36%80.685.458.574.8
NOAH---0.52%80.284.961.375.5
SPT-ADAPTER (Ours)1.08×0.41%89.50.30%81.385.360.875.8
SPT-ADAPTER (Ours)1.10×0.47%89.80.44%82.085.861.476.4
+ +Reparameterization-based methods + +
LINEAR1.02×0.12%79.30.04%68.977.226.857.6
PARTIAL-13.00×8.38%82.68.30%69.478.534.260.7
BIAS1.05×0.13%88.40.13%73.378.344.165.2
LORA-81.07×0.55%86.00.23%79.584.660.574.9
LORA-161.18×0.90%84.80.69%79.884.960.275.0
SPT-LORA (Ours)1.08×0.41%89.30.31%81.585.660.775.9
SPT-LORA (Ours)1.15×0.60%90.10.63%81.985.961.376.4
+ +Table 1: Comparisons on FGVC and VTAB-1k [60] benchmarks using supervised pre-trained ViT-B/16 backbone pre-trained on ImageNet-21k. "Total params" denotes the ratio of the total number of parameters needed for all downstream tasks relative to the one for the pre-trained backbone, and "Tuned/Total" denotes the fraction of trainable parameters. Top-1 accuracy (\%) is reported. The best result is in bold, and the second-best result is underlined. + +
ViT-B/16 (85.8M)Total ParamsVTAB-1kMAEMean Acc.VTAB-1kMoCo v3Mean Acc.
Tuned / TotalNaturalSpecializedStructuredTuned / TotalNaturalSpecializedStructured
FULL38.02×100%59.379.753.864.3100%72.084.742.069.6
+ +Addition-based methods + +
ADAPTER-81.08×0.23%57.278.454.763.40.23%27.670.948.449.0
ADAPTER-321.28×0.95%55.378.853.362.50.99%74.282.747.768.2
PROMPT-SHALLOW1.02×0.12%40.069.727.545.70.12%67.382.337.662.4
PROMPT-DEEP1.05×0.23%36.060.626.641.10.07%70.383.042.465.2
SPT-ADAPTER (Ours)1.07×0.26%64.882.460.469.20.08%76.184.960.173.7
SPT-ADAPTER (Ours)1.13×0.41%65.682.760.769.70.30%76.685.061.774.4
+ +Reparameterization-based methods + +
LINEAR1.02×0.04%18.952.723.732.10.04%67.581.130.359.6
PARTIAL-14.16×8.30%58.478.347.661.58.30%72.384.647.968.3
BIAS1.06×0.13%54.675.747.759.30.13%72.981.153.469.2
LORA-81.08×0.23%57.577.757.764.30.23%21.266.745.144.3
LORA-161.28×0.69%57.377.159.964.80.69%16.064.048.742.9
SPT-LORA (Ours)1.11×0.29%63.881.660.068.50.30%76.585.463.075.0
SPT-LORA (Ours)1.23×0.69%65.482.461.569.80.50%76.586.063.675.3
+ +Table 2: Comparisons on VTAB-1k [60] benchmark using self-supervised ViT-B/16 backbone pre-trained by MAE [19] and MoCo v3 [10]. "Total params" denotes the ratio of the total number of parameters needed for all downstream tasks relative to the one for the pre-trained backbone, and "Tuned/Total" denotes the fraction of trainable parameters. Top-1 accuracy (\%) is reported. The best result is in bold, and the second-best result is underlined. + +ing PEFT methods and boost their performance. + +Second, SPT variants outperform baseline methods and full fine-tuning by significant margins with the self-supervised pre-trained ViT-B/16 backbones. As shown in Table 2, existing PEFT approaches exhibit inferior results than full fine-tuning with the self-supervised pre-trained + +backbones MAE and MoCo v3. It is worth noting that previous PEFT methods yield inconsistent results with the backbones of different pre-training strategies. In contrast, SPT variants consistently outperform full fine-tuning. In particular, SPT-ADAPTER achieves remarkable $5.8\%$ and $5.5\%$ mean top-1 accuracy gains over the best-performing + +
MethodTuned / TotalNaturalSpecializedStructuredMean / Acc.
FULL100%79.186.259.775.0
Addition-based methods
MLP-31.60%73.675.235.761.5
PROMPT-SHALLOW0.04%79.982.537.866.7
PROMPT-DEEP0.23%76.884.553.471.6
ADAPTER-81.18%81.787.361.276.7
SPT-ADAPTER (ours)0.33%83.087.362.177.5
Reparameterization-based methods
LINEAR0.04%73.580.833.562.6
PARTIAL-12.15%73.181.735.063.3
LORA-81.18%81.787.260.176.3
SPT-LORA (ours)0.49%83.187.460.477.2
+ +baseline method on VTAB-1k benchmark with only $0.26\%$ and $0.08\%$ trainable parameters for MAE and MoCo v3 pretrained backbones, respectively. Moreover, our observation in supplementary material suggests that self-supervised pretrained ViT backbones have more diverse sensitivity distributions and a higher variance in sensitivity across different tasks than the supervised pre-trained one. This leads to the conjecture that baseline methods which assign trainable parameters to the same positions for all tasks may fail to mitigate the distinct domain gaps in individual downstream datasets, whereas our SPT allocates trainable parameters to task-specific positions accurately. + +Visual recognition with Swin and ConvNeXt backbones. From Table 3, we observe that our SPT-LORA and SPT-ADAPTER also achieve SOTA performance with Swin-B backbone on all dataset groups. We also follow VPT [22] to apply SPT variants to ResNet-alike architecture ConvNeXt-Base [33] and report the results in Table 4. We observe that SPT variants achieve better trade-offs between accuracy and parameter efficiency than the baseline methods. The results further demonstrate the versatility and effectiveness of our SPT. + +Semantic segmentation. We follow VPT [22] to conduct semantic segmentation on ADE20k dataset. Following the settings of [22], we apply our SPT to SETR-PUP [65] framework with ViT-L backbone and only allocate trainable parameters to the backbone with the head fully finetuned. We report the mIoU results in Table 5. Notably, SPT-LORA and SPT-ADAPTER outperform the baseline methods by large margins, indicating our SPT can be generalized to the semantic segmentation task. We report the results for conducting only structured tuning for both SPT-LORA and SPT-ADAPTER as it yields higher mIoU. + +# 4.3. Ablation Study + +Effect of the sensitivity criterion. We investigate the effectiveness of our sensitivity criterion on VTAB-1k by em + +Table 3: Comparisons on VTAB-1k [60] benchmark with supervised pre-trained Swin-B [32]. "Tuned/Total" denotes the fraction of trainable parameters. Top-1 accuracy (\%) is reported. The best result is in bold. + +
MethodTuned/TotalNaturalSpecializedStructuredMean Acc.
FULL100%78.083.760.474.0
LoRA0.79%82.284.764.177.0
Adapter0.47%83.184.964.677.5
SPT-LoRA (ours)0.57%83.486.765.978.7
SPT-Adapter (ours)0.36%83.786.265.378.4
+ +Table 4: Comparisons on VTAB-1k [60] benchmark with supervised pre-trained ConvNeXt-Base [33] backbone. "Tuned/Total" denotes the fraction of trainable parameters. Top-1 accuracy $(\%)$ is reported. The best result is in bold. + +
MethodmIoU-s.s. (%)mIoU-m.s. (%)Trainable Param.
VPT44.045.615.8M
LoRA43.945.914.7M
Adapter44.446.614.6M
SPT-LoRA (ours)45.447.514.6M
SPT-Adapter (ours)45.247.214.6M
+ +Table 5: Semantic Segmentation: Comparisons on ADE20k [66] val with SETR [65] on ViT-L backbone. We report both single-scale (s.s.) and multi-scale (m.s.) mIoU results. The best result is in bold. + +ploying structured tuning methods from [22, 20, 21] to the task-specific sensitive weight matrices. Note that we do not conduct unstructured tuning to ensure fair comparisons. The results are presented in Figure 3 (b). Our criterion brings consistent $1.1\%$ , $1.6\%$ , and $0.8\%$ performance gains for PROMPT-DEEP, ADAPTER-32, and LORA-16, respectively, which demonstrates the effectiveness and versatility of our sensitivity criterion to identify accurate task-specific important positions. + +Effect of structured and unstructured tuning. We investigate the effectiveness of unstructured and structured tuning individually on VTAB-1k. The results are presented in Table 6. We start by applying ADAPTER-8 to the sensitive weight matrices identified by our sensitivity criterion (SPT-ADAPTER w/o unstructured). We observe that our sensitivity criterion boosts the performance of all the dataset groups by clear margins, which again demonstrates the importance of our sensitivity criterion. Next, we observe that allocating the trainable parameters to the unstructured sensitive weight connections also brings accuracy improvement to the Natural and Specialized datasets from ADAPTER-8. However, we find that structured tuning is especially important for achieving good performance on Structured datasets. To further investigate this phenomenon, we observe that Structured datasets have larger domain gaps from the pre-training source domain [25] compared to Natural and Specialized datasets as visualized in Figure 3 (c). We hence conjecture that structured tuning has a higher representational capability than unstructured tuning which facilitates mitigating the large domain gaps during fine-tuning (see the supplementary material for visual examples). Finally, we observe + +![](images/fecd4559bdd11b1405641730406683e2e4347922e441880fffb05fb5ddd1f62d.jpg) +Figure 3: (a) Accuracy vs. parameter efficiency with supervised pre-trained ViT-B/16 backbone on VTAB-1k [60]. SPT variants perform favorably against the other PEFT approaches and are more scalable. (b) Applying other PEFT structured tuning methods [22, 20, 21] to the task-specific sensitive weight matrices (denoted by TSM) identified by our criterion with supervised pre-trained ViT-B/16 backbone on VTAB-1k. Our criterion brings consistent performance gain. (c) Domain vs. performance gaps for different dataset groups in VTAB-1k [60]. The blue bars show the domain gaps between the source domain (ImageNet [25]) and target domains, which are measured by Maximum Mean Discrepancy (MMD) distance [49]. The red line represents the performance gaps between SPT-ADAPTER w/o unstructured and w/o structured, using supervised pre-trained ViT-B/16 backbone. The dataset groups are Natural, Specialized, and Structured. Structured tuning is important for achieving good performance on Structured datasets with larger domain gaps. + +![](images/91c0b01112e12170c50af13790a69d271301b7040b4ac17299ae9d52f3719831.jpg) + +![](images/c46225b41d6523ea95dde04abeda37757627658fc77b5ac4db421db94dfe578c.jpg) + +
MethodTuned / TotalNaturalSpecializedStructuredMean / Acc.
ADAPTER-80.23%79.084.158.573.9
SPT-ADAPTER w/o unstructured0.29%81.285.160.375.5
SPT-ADAPTER w/o structured0.34%81.285.059.675.3
SPT-ADAPTER0.30%81.385.360.875.8
+ +that incorporating both structured and unstructured tuning at task-specific important positions achieves the highest performance on all dataset groups. + +Effect of number of training samples $C$ to get parameter sensitivity. We investigate the effect of the number of training images $C$ for calculating our parameter sensitivity (Algorithm 1 of the main paper). We randomly sample training samples and report the mean results over three runs in Table 8. We find that our SPT is robust to the number of training samples $C$ and randomly sampling 400 out of a total of 800 training samples is sufficient to obtain accurate task-specific important positions, e.g., calculating the sensitivity for ViT-B/16 backbone takes only 5.5 seconds with a single GPU on any of the VTAB-1k datasets and this computation is required only once. + +Computational cost analysis. We investigate the computational cost of SPT-LoRA by comparing with full fine-tuning, addition-based method PROMPT-DEEP, and reparameterization-based method LORA-16. The results are presented in Table 7. We observe that PROMPT-DEEP has higher inference latency and inference GPU memory due to the additional prompts. In contrast, since the up + +Table 6: Ablation study on structured and unstructured tuning only with supervised pre-trained ViT-B/16 backbone. Top-1 accuracy $(\%)$ is reported. We set different parameter constraints to align the fractions of trainable parameters for these cases. The best result is in bold + +
MethodInference Latency (ms/img)Inference Memory (GB)Fine-tuning Memory (GB)
FULL2.81.311.9
PROMPT-DEEP3.81.913.2
LORA-162.81.38.2
SPT-LORA w/o unstructured2.81.38.3
SPT-LORA2.81.39.8
+ +Table 7: Cost comparisons with ViT-B/16 backbone on the VTAB-1k [60] benchmark with around $0.70\%$ fractions of the trainable parameters. We report the latency (ms/img) and the peak memory usage (GB) with image resolution $224 \times 224$ on a single GeForce 3090 GPU. The best result is in bold. + +
C240400560800
Mean Acc.76.376.476.476.4
+ +Table 8: Effect of the number of training samples used to get the sensitivity for SPT-LoRA with supervised pre-trained ViT-B/16 backbone on VTAB-1k [60]. Top-1 accuracy (%) is reported. The best result is in bold. + +dated parameters after fine-tuning can be reparameterized and merged into the pre-trained model, our SPT-LoRA, SPT-LoRA w/o unstructured tuning, and LoRA-16 are more efficient than PROMPT-DEEP during inference. However, we observe that our SPT-LoRA has slightly higher fine-tuning memory than the full fine-tuning and LoRA-16 which is taken up by updating the unstructurally-tuned parameters with sparse gradients in Eq. (5). Additionally, for memory-intensive scenarios, one can employ SPT-LoRA w/o unstructured tuning for improved performance (0.8% Top-1 accuracy on VTAB-1k) and similar fine-tuning memory as LoRA, as shown in Figure 3 (b) and Table 6. + +![](images/dcb4479f9e36da7a176aef5420a12aa733401cc06b89e1800dde4aa6e7a7cd66.jpg) +Supervised pre-trained ViT-B + +![](images/a20dc078f3323546bbe037059cb0ffe8c468d3fe89167440b381dad03f292337.jpg) +MAE pre-trained ViT-B + +![](images/3c4674f60c4f5db37bf71b1b05fa21a7117f87e0d605b4e8e426893a45353170.jpg) +MoCo V3 pre-trained ViT-B + +![](images/897aed02949367588ae110b14be073a14162de7848bdd9c4054d053132b051a3.jpg) +Figure 4: Parameter sensitivity patterns under 0.4M trainable parameter budget for ViT-B/16 backbone with different pretraining strategies on three sample tasks from VTAB-1k [61]. The proportions exhibit task-specific varying patterns in terms of network depth (upper figures) and task-agnostic similar patterns in terms of operations (lower figures). + +![](images/e3d68d8f8babb9ecc5bc562070062d1af7b5d1c01bb7fd962d3bc3478b7fbadb.jpg) + +![](images/9aaa447abff829082dd93d1feb56772225b8dc2d55604cf3d85e4945eeeb6364.jpg) + +# 4.4. Observations on Sensitivity Patterns + +Our sensitivity criterion identifies task-specific important positions, which can reveal the contributions of the pretrained weights to different downstream tasks during transfer learning. We visualize the proportions of the sensitive parameters for the supervised pre-trained ViT-B/16 backbone under 0.4M trainable parameter budget in Figure 4. First, we investigate the most sensitive blocks, whose numbers of sensitive parameters are summed and normalized over the 12 ViT-B/16 blocks. We observe that the patterns of the sensitive parameter proportions vary markedly across different tasks, which echoes the observations made in [16]. This suggests that we should not introduce trainable parameters to the same positions for each individual task but allocate trainable parameters at task-specific ones as we proposed. Next, we investigate the most insensitive weight matrices within a block. A ViT block consists of a query $W_{q}$ , a key $W_{k}$ , a value $W_{v}$ , and an output $W_{o}$ weight matrices in the multi-head self-attention layer and two weight matrices $W_{fc1}$ and $W_{fc2}$ in the feed-forward network as elaborated in [51, 12]. We observe that the query $W_{q}$ and key $W_{k}$ weight matrices have the lowest proportions of sensitive parameters for all three sample tasks. Since $W_{q}$ and $W_{k}$ are responsible for learning the attention scores which indicate the pairwise similarity among the patches, we speculate that although domain changes, the patch relationships learned during pre-training can be efficiently reused when transferred to downstream classification tasks. + +# 5. Conclusion + +In this paper, we have explored identifying and allocating trainable parameters to task-specific important positions for visual parameter-efficient tuning. Specifically, we have proposed a novel criterion to quickly measure the sensitivity of the pre-trained parameters for each specific task before fine-tuning. Based on the parameter sensitivity, we have proposed a trainable parameter allocation strategy that adaptively combines both unstructured and structured tuning under a desired trainable parameter budget, enabling high representational capability and flexibility. Finally, we have conducted extensive experiments on a total of 24 downstream recognition tasks with both plain and hierarchical vision Transformer backbones under different pre-training strategies to demonstrate the versatility and effectiveness of our proposed SPT. Notably, we have shown that our approach is complementary to the existing PEFT methods and improves their performance significantly. In the future, we will explore adapting large vision models to more downstream tasks with SPT, e.g., dense prediction and vision-and-language tasks, and improve the training efficiency of SPT for on-device training [6, 30]. + +Acknowledgement. We thank Jing liu and Ziyi Liu for their helpful discussions. This research is partially supported by Monash FIT Start-up Grant. Dr. Jing Zhang is supported by the Australian Research Council project FL-170100117. + +# References + +[1] Y. Bao, Y. Li, S.-L. Huang, L. Zhang, L. Zheng, A. Zamir, and L. Guibas. An information-theoretic approach to transferability in task transfer learning. In ICIP, pages 2309-2313. IEEE, 2019. 3 +[2] G. Bender, P.-J. Kindermans, B. Zoph, V. Vasudevan, and Q. Le. Understanding and simplifying one-shot architecture search. In ICML, pages 550-559. PMLR, 2018. 2 +[3] A. Brock, T. Lim, J. M. Ritchie, and N. Weston. Smash: one-shot model architecture search through hypernetworks. In ICLR, 2018. 2 +[4] S. Caelles, K.-K. Maninis, J. Pont-Tuset, L. Leal-Taixe, D. Cremers, and L. Van Gool. One-shot video object segmentation. In CVPR, pages 221-230, 2017. 3 +[5] H. Cai, C. Gan, T. Wang, Z. Zhang, and S. Han. Once-for-all: Train one network and specialize it for efficient deployment. In ICLR, 2020. 2 +[6] H. Cai, C. Gan, L. Zhu, and S. Han. Tinytl: Reduce memory, not parameters for efficient on-device learning. NeurIPS, 33:11285-11297, 2020. 9 +[7] H. Cai, L. Zhu, and S. Han. ProxylessNAS: Direct neural architecture search on target task and hardware. In ICLR, 2019. 4 +[8] N. S. Chatterji, B. Neyshabur, and H. Sedghi. The intriguing role of module criticality in the generalization of deep networks. In ICLR, 2020. 3 +[9] S. Chen, C. Ge, Z. Tong, J. Wang, Y. Song, J. Wang, and P. Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 2022. 1, 2, 4, 5 +[10] X. Chen, S. Xie, and K. He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021. 5, 6 +[11] Y. Cui, Y. Song, C. Sun, A. Howard, and S. Belongie. Large scale fine-grained categorization and domain-specific transfer learning. In CVPR, pages 4109-4118, 2018. 3 +[12] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 4, 5, 9 +[13] T. Gebru, J. Krause, Y. Wang, D. Chen, J. Deng, and L. Fei-Fei. Fine-grained car detection for visual census estimation. In AAAI, 2017. 5 +[14] D. Guo, A. Rush, and Y. Kim. Parameter-efficient transfer learning with diff pruning. In ACL-IJCNLP, 2021. 3 +[15] Y. Guo, Y. Li, L. Wang, and T. Rosing. Adafilter: Adaptive filter fine-tuning for deep transfer learning. In AAAI, pages 4060-4066, 2020. 3 +[16] Y. Guo, H. Shi, A. Kumar, K. Grauman, T. Rosing, and R. Feris. Spottune: transfer learning through adaptive fin-tuning. In CVPR, pages 4805-4814, 2019. 3, 9 + +[17] T. Hao, H. Chen, Y. Guo, and G. Ding. Consolidator: Mergable adapter with group connections for vision transformer. In ICLR, 2023. 3 +[18] J. He, C. Zhou, X. Ma, T. Berg-Kirkpatrick, and G. Neubig. Towards a unified view of parameter-efficient transfer learning. In ICLR, 2022. 2 +[19] K. He, X. Chen, S. Xie, Y. Li, P. Dollár, and R. Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 5, 6 +[20] N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. De Laroussilhe, A. Gesmundo, M. Attariyan, and S. Gelly. Parameter-efficient transfer learning for nlp. In ICML, pages 2790-2799, 2019. 1, 2, 4, 5, 7, 8 +[21] E. J. Hu, yelong shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 1, 2, 3, 4, 5, 7, 8 +[22] M. Jia, L. Tang, B.-C. Chen, C. Cardie, S. Belongie, B. Hariharan, and S.-N. Lim. Visual prompt tuning. In ECCV, 2022. 1, 2, 5, 7, 8 +[23] S. Jie and Z.-H. Deng. Convolutional bypasses are better vision transformer adapters. arXiv preprint arXiv:2207.07039, 2022. 1, 2, 4 +[24] A. Khosla, N. Jayadevaprakash, B. Yao, and L. Fei-Fei. Novel dataset for fine-grained image categorization. In CVPRW, 2011. 5 +[25] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 5, 7, 8 +[26] A. Kumar, A. Raghunathan, R. M. Jones, T. Ma, and P. Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In ICLR, 2022. 2, 3 +[27] B. Lester, R. Al-Rfou, and N. Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021. 2 +[28] X. L. Li and P. Liang. Prefix-tuning: Optimizing continuous prompts for generation. In IJCNLP, pages 4582-4597. Association for Computational Linguistics, 2021. 2 +[29] D. Lian, D. Zhou, J. Feng, and X. Wang. Scaling & shifting your features: A new baseline for efficient model tuning. NeurIPS, 2022. 3 +[30] J. Lin, L. Zhu, W.-M. Chen, W.-C. Wang, C. Gan, and S. Han. On-device training under 256kb memory. arXiv preprint arXiv:2206.15472, 2022.9 +[31] H. Liu, K. Simonyan, and Y. Yang. DARTS: Differentiable architecture search. In ICLR, 2019. 4 +[32] Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 5, 7 +[33] Z. Liu, H. Mao, C.-Y. Wu, C. Feichtenhofer, T. Darrell, and S. Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 7 +[34] I. Loshchilov and F. Hutter. Fixing weight decay regularization in adam, 2018. 5 + +[35] P. Molchanov, A. Mallya, S. Tyree, I. Frosio, and J. Kautz. Importance estimation for neural network pruning. In CVPR, pages 11264-11272, 2019. 2, 4 +[36] M. M. Naseer, K. Ranasinghe, S. H. Khan, M. Hayat, F. Shahbaz Khan, and M.-H. Yang. Intriguing properties of vision transformers. *NeurIPS*, 34:23296-23308, 2021. 3 +[37] B. Neyshabur, H. Sedghi, and C. Zhang. What is being transferred in transfer learning? NeurIPS, 33:512-523, 2020. 2, 3 +[38] C. Nguyen, T. Hassner, M. Seeger, and C. Archambeau. Leep: A new measure to evaluate transferability of learned representations. In ICML, pages 7294-7305. PMLR, 2020. 3 +[39] M.-E. Nilsback and A. Zisserman. Automated flower classification over a large number of classes. In ICVGIP, pages 722-729. IEEE, 2008. 5 +[40] J. Pleted and T. Gedeon. Deep transfer learning for image classification: a survey. arXiv preprint arXiv:2205.09904, 2022.3 +[41] J. Puigcerver, C. Riquelme, B. Mustafa, C. Renggli, A. S. Pinto, S. Gelly, D. Keysers, and N. Houlsby. Scalable transfer learning with expert models. arXiv preprint arXiv:2009.13239, 2020. 3 +[42] M. Raghu, T. Unterthiner, S. Kornblith, C. Zhang, and A. Dosovitskiy. Do vision transformers see like convolutional neural networks? NeurIPS, 34:12116-12128, 2021. 3 +[43] M. T. Rosenstein, Z. Marx, L. P. Kaelbling, and T. G. Dietterich. To transfer or not to transfer. In NeurlPSW, 2005. 3 +[44] R. K. Srivastava, K. Greff, and J. Schmidhuber. Training very deep networks. NeurIPS, 28, 2015. 2 +[45] X. Sun, R. Panda, R. Feris, and K. Saenko. Adashare: Learning what to share for efficient deep multi-task learning. NeurIPS, 33:8728-8740, 2020. 3 +[46] Y.-L. Sung, J. Cho, and M. Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. arXiv preprint arXiv:2206.06522, 2022. 2 +[47] A. T. Tran, C. V. Nguyen, and T. Hassner. Transferability and hardness of supervised classification tasks. In ICCV, pages 1395–1405, 2019. 3 +[48] C.-H. Tu, Z. Mai, and W.-L. Chao. Visual query tuning: Towards effective usage of intermediate representations for parameter and memory efficient transfer learning. arXiv preprint arXiv:2212.03220, 2022. 2 +[49] E. Tzeng, J. Hoffman, N. Zhang, K. Saenko, and T. Darrell. Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:1412.3474, 2014. 8 +[50] G. Van Horn, S. Branson, R. Farrell, S. Haber, J. Barry, P. Ipeirotis, P. Perona, and S. Belongie. Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In CVPR, pages 595–604, 2015. 5 + +[51] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. NeurIPS, pages 5998-6008, 2017. 9 +[52] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The caltech-ucsd birds-200-2011 dataset. Tech. Rep. CNS-TR-2011-001, California Institute of Technology, 2011. 5 +[53] Z. Wang, Z. Dai, B. Póczos, and J. Carbonell. Characterizing and avoiding negative transfer. In CVPR, pages 11293-11302, 2019. 3 +[54] H. Wu, J. Xu, J. Wang, and M. Long. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. NeurIPS, 34:22419-22430, 2021. 2 +[55] R. Xu, F. Luo, Z. Zhang, C. Tan, B. Chang, S. Huang, and F. Huang. Raise a child in large language model: Towards effective and generalizable fine-tuning. In EMNLP, 2021. 3 +[56] Y. Xu, Q. Zhang, J. Zhang, and D. Tao. Vitae: Vision transformer advanced by exploring intrinsic inductive bias. In NeurIPS, 2021. 1 +[57] J. Yoon, S. Arik, and T. Pfister. Data valuation using reinforcement learning. In ICML, pages 10842-10851. PMLR, 2020. 3 +[58] J. Yosinski, J. Clune, Y. Bengio, and H. Lipson. How transferable are features in deep neural networks? NeurIPS, 27, 2014. 2, 3 +[59] E. B. Zaken, Y. Goldberg, and S. Ravfogel. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In ACL, pages 1-9, 2022. 3, 5 +[60] X. Zhai, J. Puigcerver, A. Kolesnikov, P. Ruyssen, C. Riquelme, M. Lucic, J. Djolonga, A. S. Pinto, M. Neumann, A. Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 5, 6, 7, 8 +[61] X. Zhai, J. Puigcerver, A. Kolesnikov, P. Ruyssen, C. Riquelme, M. Lucic, J. Djolonga, A. S. Pinto, M. Neumann, A. Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 9 +[62] C. Zhang, S. Bengio, and Y. Singer. Are all layers created equal? arXiv preprint arXiv:1902.01996, 2019. 3 +[63] Y. Zhang, K. Zhou, and Z. Liu. Neural prompt search. arXiv preprint arXiv:2206.04673, 2022. 1, 2, 5 +[64] M. Zhao, T. Lin, F. Mi, M. Jaggi, and H. Schütze. Masking as an efficient alternative to finetuning for pretrained language models. In EMNLP, 2020. 3 +[65] S. Zheng, J. Lu, H. Zhao, X. Zhu, Z. Luo, Y. Wang, Y. Fu, J. Feng, T. Xiang, P. H. Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. 7 +[66] B. Zhou, H. Zhao, X. Puig, S. Fidler, A. Barriuso, and A. Torralba. Scene parsing through ade20k dataset. In CVPR, pages 633-641, 2017. 7 \ No newline at end of file diff --git a/2023/Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning/images.zip b/2023/Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ce97a6195d040f342577db0c550bd53b0efc0c70 --- /dev/null +++ b/2023/Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c48036ecc60734671d4423823676d973238970bed8947dc7f94261ff34d7b51 +size 610402 diff --git a/2023/Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning/layout.json b/2023/Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2ce8118e094bd7049f66e70bf097801523324822 --- /dev/null +++ b/2023/Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning/layout.json @@ -0,0 +1,9483 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 118, + 103, + 476, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 103, + 476, + 121 + ], + "spans": [ + { + "bbox": [ + 118, + 103, + 476, + 121 + ], + "type": "text", + "content": "Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 142, + 482, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 142, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 115, + 142, + 482, + 159 + ], + "type": "text", + "content": "Haoyu He" + }, + { + "bbox": [ + 115, + 142, + 482, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 115, + 142, + 482, + 159 + ], + "type": "text", + "content": " Jianfei Cai" + }, + { + "bbox": [ + 115, + 142, + 482, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 115, + 142, + 482, + 159 + ], + "type": "text", + "content": " Jing Zhang" + }, + { + "bbox": [ + 115, + 142, + 482, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 115, + 142, + 482, + 159 + ], + "type": "text", + "content": " Dacheng Tao" + }, + { + "bbox": [ + 115, + 142, + 482, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 115, + 142, + 482, + 159 + ], + "type": "text", + "content": " Bohan Zhuang" + }, + { + "bbox": [ + 115, + 142, + 482, + 159 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 163, + 443, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 163, + 443, + 179 + ], + "spans": [ + { + "bbox": [ + 149, + 163, + 443, + 179 + ], + "type": "text", + "content": "1 ZIP Lab, Monash University 2 The University of Sydney" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 209, + 192, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 209, + 192, + 221 + ], + "spans": [ + { + "bbox": [ + 143, + 209, + 192, + 221 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 234, + 290, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 234, + 290, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 234, + 290, + 594 + ], + "type": "text", + "content": "Visual Parameter-Efficient Fine-Tuning (PEFT) has become a powerful alternative for full fine-tuning so as to adapt pre-trained vision models to downstream tasks, which only tunes a small number of parameters while freezing the vast majority ones to ease storage burden and optimization difficulty. However, existing PEFT methods introduce trainable parameters to the same positions across different tasks depending solely on human heuristics and neglect the domain gaps. To this end, we study where to introduce and how to allocate trainable parameters by proposing a novel Sensitivity-aware Visual Parameter-efficient fine-Tuning (SPT) scheme, which adaptively allocates trainable parameters to task-specific important positions given a desired tunable parameter budget. Specifically, our SPT first quickly identifies the sensitive parameters that require tuning for a given task in a data-dependent way. Next, our SPT further boosts the representational capability for the weight matrices whose number of sensitive parameters exceeds a pre-defined threshold by utilizing existing structured tuning methods, e.g., LoRA [21] or Adapter [20], to replace directly tuning the selected sensitive parameters (unstructured tuning) under the budget. Extensive experiments on a wide range of downstream recognition tasks show that our SPT is complementary to the existing PEFT methods and largely boosts their performance, e.g., SPT improves Adapter with supervised pre-trained ViT-B/16 backbone by " + }, + { + "bbox": [ + 47, + 234, + 290, + 594 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 47, + 234, + 290, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 234, + 290, + 594 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 47, + 234, + 290, + 594 + ], + "type": "text", + "content": " mean Top-1 accuracy, reaching SOTA performance on FGVC and VTAB-1k benchmarks, respectively. Source code is at https://github.com/ziplab/SPT." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 615, + 128, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 615, + 128, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 615, + 128, + 628 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "type": "text", + "content": "To effectively adapt the pre-trained representations to the downstream tasks, the de-facto choice is full fine-tuning, which initializes the model with the pre-trained weights and tunes all the parameters. However, vanilla full fine-tuning needs to store a separate instance of parameters for each" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 209, + 535, + 369 + ], + "blocks": [ + { + "bbox": [ + 312, + 209, + 535, + 369 + ], + "lines": [ + { + "bbox": [ + 312, + 209, + 535, + 369 + ], + "spans": [ + { + "bbox": [ + 312, + 209, + 535, + 369 + ], + "type": "image", + "image_path": "4f12a796781166c998a97d80560c43c3045d884ad14a5de144773c620fb5fb8e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 373, + 547, + 542 + ], + "lines": [ + { + "bbox": [ + 304, + 373, + 547, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 547, + 542 + ], + "type": "text", + "content": "Figure 1: (a) Existing PEFT methods, such as Adapter [20] introduce trainable parameters to the same positions for all downstream tasks. These methods design task-agnostic positions to employ trainable parameters relying on heuristics and neglect consideration of the distinct domain gaps and characteristics for the downstream tasks. (b) Our Sensitivity-aware visual Parameter-efficient fine-Tuning (SPT) introduces trainable parameters to the task-specific important positions and allocates them with both unstructured and structured tuning granularities, simultaneously. For structured tuning, SPT can exploit any existing structured tuning methods, such as LoRA [21] or Adapter [20]. Red lines and blocks represent trainable parameters and modules, while blue lines represent frozen parameters." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 548, + 547, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 547, + 657 + ], + "type": "text", + "content": "task and each deployment scenario. It can be extremely storage-intensive as the storage cost grows linearly with the number of possible cases, considering there are vast varieties of downstream tasks and dynamic deployment environments, especially when deploying the large vision models [12, 32, 56] to mobile systems. For example, even storing a single large pre-trained ViT-H [12] model on a local disk requires at least 2.3GB, while the Top-10 U.S. apps required only collectively 2.2GB in May 2021." + }, + { + "bbox": [ + 304, + 548, + 547, + 657 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 657, + 547, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 657, + 547, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 657, + 547, + 693 + ], + "type": "text", + "content": "Notably, an emerging solution is to replace vanilla fine-tuning with visual Parameter-Efficient Fine-Tuning (PEFT) [22, 9, 63, 23], which only tunes a small num" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 257, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 257, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 257, + 713 + ], + "type": "text", + "content": "†Corresponding author. E-mail: bohan.zhuang@gmail.com" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 702, + 500, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 500, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 500, + 713 + ], + "type": "text", + "content": "1https://sensorTower.com/blog/ios-app-size-growth-2021" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11825" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "content": "ber of trainable parameters while freezing the vast majority ones that are shared by multiple tasks. As PEFT approaches exhibit less than " + }, + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "content": " of trainable parameters, the storage burden is largely alleviated. Another attractive property of PEFT is that tuning fewer parameters eases the optimization difficulty and mitigates the overfitting issue when adapting large pre-trained models on the target dataset, thereby achieving comparable or even better performance than vanilla fine-tuning [22]. Although promising, the existing PEFT approaches introduce trainable parameters to the same positions for all downstream tasks, relying on human heuristics and neglecting task-specific domain gaps and characteristics, which limits their performance. For instance, in a task-agnostic manner, Prompt Tuning-deep [22] and Adapter [20] respectively add trainable parameters to multi-head self-attention and feed-forward network layers for all distinct tasks as depicted in Figure 1 (a)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 278, + 289, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 289, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 289, + 531 + ], + "type": "text", + "content": "To address this fundamental challenge, we explore where to introduce and how to allocate trainable parameters under a desired parameter budget by presenting a novel Sensitivity-aware visual Parameter-efficient fine-Tuning (SPT) scheme that identifies the task-specific important positions to adaptively allocate trainable parameters. Since the pre-trained weights at distinct positions have varying contributions for different downstream tasks [58, 26, 37], we first propose a new criterion to quickly identify the task-specific sensitive parameters that require tuning in a data-dependent way. Inspired by model pruning metrics [44, 35, 2, 3], we propose to measure the parameter sensitivity with the loss reduction when being tuned, which can be approximated by a first-order Taylor expansion derived within a single forward and backward pass ahead of fine-tuning in one-shot. Our sensitivity criterion is simple and effective, which can identify the task-specific important positions to introduce trainable parameters for any backbone quickly. For instance, calculating the sensitivity for ViT-B/16 backbone takes only 5.5 seconds with a single GPU on any of the VTAB-1k datasets." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": "With our criterion, we empirically observe that the proportions of the sensitivity parameters for each block indeed vary markedly across different tasks in Section 4.4. To allocate the trainable parameters under a desired trainable parameter budge, an intuitive solution is to directly tune the most sensitive weight connections, which we name unstructured tuning. Despite its simplicity and flexibility, unstructured tuning only tunes a few parameters which still lacks representational capability and is challenging to bridge the domain gap. To this end, we propose to further incorporate structured tuning to replace unstructured tuning at the sensitive weight matrices whose numbers of sensitive parameters exceed a pre-defined threshold to improve the representational capability under a similar parameter budget. Structured tuning can be implemented by any parameter-efficient" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": "structured tuning methods [21, 9, 23, 22] that directly adjust the hidden representations, e.g., inserting an adapter module sequentially after the sensitive weight matrices. Therefore, our SPT adaptively combines both unstructured and structured tuning granularity and allocates trainable parameters with high flexibility and representational capability for each distinct downstream task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 158, + 547, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 547, + 373 + ], + "type": "text", + "content": "This paper has the following key contributions. 1) We make the pioneering exploration to identify the task-specific important positions under the PEFT setting, which is fast, effective, versatile to be applied to various backbones with different pre-training strategies, and orthogonal to the existing PEFT methods. 2) Based on the sensitivity criterion, we propose a trainable parameter allocation strategy that adaptively combines both unstructured and structured tuning under a desired parameter budget to achieve high flexibility, large capacity, and favorable trade-off between parameter efficiency and accuracy. 3) Extensive experiments on a total of 24 downstream recognition tasks with both plain and hierarchical vision Transformer backbones under supervised and self-supervised pre-trainings show that our SPT is complementary to the existing PEFT methods and boosts their performance by large margins. For instance, SPT improves Adapter [20] by " + }, + { + "bbox": [ + 304, + 158, + 547, + 373 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 304, + 158, + 547, + 373 + ], + "type": "text", + "content": " mean Top-1 accuracy, outperforming the SOTA PEFT methods on the FGVC benchmark." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 383, + 392, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 383, + 392, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 383, + 392, + 395 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 401, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 520 + ], + "type": "text", + "content": "Parameter-efficient fine-tuning. Full fine-tuning is the most predominant approach when adapting a large-scale pre-trained model to downstream tasks, where the model is initialized from the pre-trained weights with all parameters trainable. Yet, when a model becomes larger, parameter-efficient fine-tuning [27, 28] is highly desirable, which tunes only a tiny portion of parameters to alleviate the storage burden. The general PEFT approaches can be categorized into addition-based PEFT methods and reparameterization-based PEFT methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "content": "Addition-based PEFT attaches additional trainable parameters to the backbone and only tunes these parameters. Apart from Prompt tuning [22] and Adapter [20], recent addition-based methods study connecting or combining existing PEFT methods. For instance, He et al. [18] connect Prompt tuning and Adapter and provide a unified view that all PEFT approaches share the same design to adjust the hidden representations. Zhang et al. [63] search for the optimal configurations to combine multiple PEFT approaches following once-for-all scheme [5, 54]. Since the additional parameters require extra computations compared to full fine-tuning, a few recent works [46, 48] design specific architectures to avoid storing the intermediate activations, thereby alleviating the fine-tuning memory cost. However, it is noteworthy that enhancing training efficiency is not the primary objective of our work." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11826" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "content": "Reparameterization-based PEFT aims to avoid extra computational costs by tuning parameters that are inherently in or can be reparameterized into the backbone during inference. Prior works select the parameters that are inherently in the backbone, including the bias terms [59], the last several layers [58, 4], and weight connections [14, 64]. To reparameterize new parameters into the backbone [17, 29], representative work LoRA [21] optimizes two low-rank matrices which can be further merged into the weight matrices. In contrast to the aforementioned works, we argue the importance of tuning parameters at task-specific important positions and quickly identify them with our proposed parameter sensitivity criterion before tuning, which is complementary to and provides valuable guidance for the existing PEFT methods. Moreover, our SPT can also be inference-efficient when implementing structured tuning with any reparameterization-based structured tuning method. Recently, SSF [29] is proposed to introduce trainable scaling and shifting parameters that can be absorbed into the previous linear layers. However, it cannot scale to higher trainable parameter budgets and requires a complex and time-consuming hyper-parameter search for learning rate, weight decay, and drop-path rate on each individual dataset, thus is not directly comparable to our method." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 367, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 367, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 367, + 289, + 715 + ], + "type": "text", + "content": "Task-specific transfer learning. The effectiveness of transferring pre-trained models to downstream tasks strongly depends on the relationship between the source and target tasks [43, 53, 26, 40]. This has motivated the community to explore the optimal pre-training data [11, 57], model [47, 38], and weights [16, 55] for the target task. To seek suitable task-specific pre-training data, Cui et al. [11] select the source domain data from the top-k most similar classes measured by Earth Mover's Distance; Yoon et al. [57] weight each class in the source domain with reinforcement learning; and Puigcerver et al. [41] first train a diverse set of experts and then select the most relevant expert for each target task. Another line of work selects a suitable pre-trained model for the target task ahead of fine-tuning by measuring the transferability of pre-trained models to the target domain with interclass covariance between the source data and target classes [1] or conditional cross-entropy [47] between the source and target labels. Considering the transferability of the feature representations at distinct layers for the same pre-trained model is different [58, 37], recent works [15, 45] endeavour transfer task-specific weights by freezing some pre-trained weights and fine-tuning the rest. For example, the task-specific fine-tuned weights are selected by learning a policy network with Gumbel-Softmax [16], optimizing a sparse mask with " + }, + { + "bbox": [ + 47, + 367, + 289, + 715 + ], + "type": "inline_equation", + "content": "L_{0}" + }, + { + "bbox": [ + 47, + 367, + 289, + 715 + ], + "type": "text", + "content": " norm [14], and learning binary gates for each parameter [64]. Our SPT also adaptively selects task-specific parameters. In contrast to the previous work, we 1) derive task-specific important positions prior to fine-tuning with only a single" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "content": "forward and backward pass, which is computationally efficient; 2) mask the gradients for insensitive parameters in unstructured tuning with fixed binary masks, thereby having more affordable fine-tuning memory than optimizing learnable binary masks in [14, 64]. Moreover, we are pioneering work to adaptively allocate task-specific trainable parameters with both fine-grained unstructured and coarse-grained structured tuning granularities to achieve both high flexibility and representational capability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 190, + 362, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 190, + 362, + 204 + ], + "spans": [ + { + "bbox": [ + 305, + 190, + 362, + 204 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 209, + 547, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 547, + 293 + ], + "type": "text", + "content": "Our sensitivity-aware visual parameter-efficient fine-tuning consists of two stages. In the first stage, SPT measures the task-specific sensitivity for the pre-trained parameters (Section 3.1). Based on the parameter sensitivity and a given parameter budget, SPT then adaptively allocates trainable parameters to task-specific important positions (Section 3.2)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 299, + 493, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 299, + 493, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 299, + 493, + 312 + ], + "type": "text", + "content": "3.1. Task-specific Parameter Sensitivity" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 316, + 547, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 547, + 446 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 547, + 446 + ], + "type": "text", + "content": "Recent research has observed that pre-trained backbone parameters exhibit varying feature patterns [42, 36] and criticality [62, 8] at distinct positions. Moreover, when transferred to downstream tasks, their efficacy varies depending on how much pre-trained features are reused and how well they adapt to the specific domain gap [58, 26, 37]. Motivated by these observations, we argue that not all parameters contribute equally to the performance across different tasks in PEFT and propose a new criterion to measure the sensitivity of the parameters in the pre-trained backbone for a given task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 448, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 547, + 544 + ], + "type": "text", + "content": "Specifically, given the training dataset " + }, + { + "bbox": [ + 304, + 448, + 547, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 304, + 448, + 547, + 544 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 304, + 448, + 547, + 544 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 448, + 547, + 544 + ], + "type": "text", + "content": "-th task and the pre-trained model weights " + }, + { + "bbox": [ + 304, + 448, + 547, + 544 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w} = \\{w_1, w_2, \\dots, w_N\\} \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 304, + 448, + 547, + 544 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 448, + 547, + 544 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 448, + 547, + 544 + ], + "type": "text", + "content": " is the total number of parameters, the objective for the task is to minimize the empirical risk: " + }, + { + "bbox": [ + 304, + 448, + 547, + 544 + ], + "type": "inline_equation", + "content": "\\min_{\\boldsymbol{w}} E(\\mathcal{D}_t, \\boldsymbol{w})" + }, + { + "bbox": [ + 304, + 448, + 547, + 544 + ], + "type": "text", + "content": ". We denote the parameter sensitivity set as " + }, + { + "bbox": [ + 304, + 448, + 547, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{s_1, \\dots, s_N\\}" + }, + { + "bbox": [ + 304, + 448, + 547, + 544 + ], + "type": "text", + "content": " and the sensitivity " + }, + { + "bbox": [ + 304, + 448, + 547, + 544 + ], + "type": "inline_equation", + "content": "s_n" + }, + { + "bbox": [ + 304, + 448, + 547, + 544 + ], + "type": "text", + "content": " for parameter " + }, + { + "bbox": [ + 304, + 448, + 547, + 544 + ], + "type": "inline_equation", + "content": "w_n" + }, + { + "bbox": [ + 304, + 448, + 547, + 544 + ], + "type": "text", + "content": " is measured by the empirical risk difference when tuning it:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 341, + 555, + 545, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 555, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 341, + 555, + 545, + 569 + ], + "type": "interline_equation", + "content": "s _ {n} = E \\left(\\mathcal {D} _ {t}, \\boldsymbol {w}\\right) - E \\left(\\mathcal {D} _ {t}, \\boldsymbol {w} \\mid w _ {n} = w _ {n} ^ {*}\\right), \\tag {1}", + "image_path": "98384341632da9183f000e17843b43809eca89aca317de5bfa72b9281464dfd9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "inline_equation", + "content": "w_{n}^{*} = \\operatorname*{argmin}_{w_{n}}(E(\\mathcal{D}_{t},\\boldsymbol {w}))" + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "text", + "content": ". We can reparameterize the tuned parameters as " + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "inline_equation", + "content": "w_{n}^{*} = w_{n} + \\Delta_{w_{n}}" + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\Delta_{w_n}" + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "text", + "content": " denotes the update for " + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "inline_equation", + "content": "w_{n}" + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "text", + "content": " after tuning. Here we individually measure the sensitivity of each parameter, which is reasonable given that most of the parameters are frozen during fine-tuning in PEFT. However, it is still computationally intensive to compute Eq. (1) for two reasons. Firstly, getting the empirical risk for " + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "text", + "content": " parameters requires forwarding the entire network " + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "text", + "content": " times, which is time-consuming. Secondly, it is challenging to derive " + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\Delta_{w_n}" + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "text", + "content": ", as we have to tune each individual " + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "inline_equation", + "content": "w_{n}" + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "text", + "content": " until convergence." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11827" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 541, + 125 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 541, + 125 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 541, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 541, + 125 + ], + "type": "image", + "image_path": "e0de3fc7f9016d1bcc965bd66924f89237eb5daa10b3805b7d071e13ff65f082.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 125, + 547, + 187 + ], + "lines": [ + { + "bbox": [ + 46, + 125, + 547, + 187 + ], + "spans": [ + { + "bbox": [ + 46, + 125, + 547, + 187 + ], + "type": "text", + "content": "Figure 2: Overview of our trainable parameter allocation strategy. With the parameter sensitivity set " + }, + { + "bbox": [ + 46, + 125, + 547, + 187 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 125, + 547, + 187 + ], + "type": "text", + "content": ", we first get the top-" + }, + { + "bbox": [ + 46, + 125, + 547, + 187 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 125, + 547, + 187 + ], + "type": "text", + "content": " sensitive parameters. Instead of directly tuning these sensitive parameters, we also boost the representational capability by replacing unstructured tuning with structured tuning at sensitive weight matrices that have a large number of sensitive parameters, which can be implemented by an existing structured tuning method, e.g., LoRA [21] and Adapter [20]. Red lines and blocks represent trainable parameters and modules, while blue lines represent frozen parameters." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "code", + "bbox": [ + 55, + 217, + 289, + 385 + ], + "blocks": [ + { + "bbox": [ + 47, + 191, + 287, + 214 + ], + "lines": [ + { + "bbox": [ + 47, + 191, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 47, + 191, + 287, + 214 + ], + "type": "text", + "content": "Algorithm 1 Computing task-specific parameter sensitivities" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "text", + "content": "Input: Pre-trained model with network parameters " + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "text", + "content": " training set " + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "text", + "content": "-th task, and number of training samples " + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "text", + "content": " used to calculate the parameter sensitivities \nOutput: Sensitivity set " + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{s_1,\\dots,s_N\\}" + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "text", + "content": " \nInitialize " + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{0\\}^N" + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "text", + "content": " \nfor " + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "inline_equation", + "content": "i\\in \\{1,\\ldots ,C\\}" + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "text", + "content": " do \n Get the " + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "text", + "content": "-th training sample of " + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "text", + "content": " \n Compute loss " + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "text", + "content": " \n Compute gradients " + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "inline_equation", + "content": "\\pmb{g}" + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "text", + "content": " \n for " + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "inline_equation", + "content": "n\\in \\{1,\\dots,N\\}" + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "text", + "content": " do \n Update sensitivity for the " + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "text", + "content": "-th parameter: " + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "inline_equation", + "content": "s_n = s_n + g_n^2" + }, + { + "bbox": [ + 55, + 217, + 289, + 385 + ], + "type": "text", + "content": " \n end for \nend for" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 405, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 287, + 441 + ], + "type": "text", + "content": "To overcome the first barrier, we simplify the empirical loss by approximating " + }, + { + "bbox": [ + 46, + 405, + 287, + 441 + ], + "type": "inline_equation", + "content": "s_n" + }, + { + "bbox": [ + 46, + 405, + 287, + 441 + ], + "type": "text", + "content": " in the vicinity of " + }, + { + "bbox": [ + 46, + 405, + 287, + 441 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 405, + 287, + 441 + ], + "type": "text", + "content": " by its first-order Taylor expansion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 445, + 287, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 287, + 460 + ], + "type": "interline_equation", + "content": "s _ {n} ^ {(1)} = - g _ {n} \\Delta_ {w _ {n}}, \\tag {2}", + "image_path": "a54674c34e54f7de8c311269ac21fb4edec1ac95abe4d832e87baae46cbe260f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "type": "text", + "content": "where the gradients " + }, + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "type": "inline_equation", + "content": "\\pmb{g} = \\partial E / \\partial \\pmb{w}" + }, + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "type": "inline_equation", + "content": "g_{n}" + }, + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "type": "text", + "content": " is the gradient of the " + }, + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "type": "text", + "content": "-th element of " + }, + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "type": "inline_equation", + "content": "\\pmb{g}" + }, + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "type": "text", + "content": ". To address the second barrier, following [31, 7], we take the one-step unrolled weight as the surrogate for " + }, + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "type": "inline_equation", + "content": "w_{n}^{*}" + }, + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "type": "text", + "content": " and approximate " + }, + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "type": "inline_equation", + "content": "\\Delta_{w_n}" + }, + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "type": "text", + "content": " in Eq. (2) with a single step of gradient descent. We can accordingly get " + }, + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "type": "inline_equation", + "content": "s_n^{(1)}\\approx g_n^2\\epsilon" + }, + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "type": "text", + "content": " is the learning rate. Since " + }, + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 462, + 287, + 560 + ], + "type": "text", + "content": " is the same for all parameters, we can eliminate it when comparing the sensitivity with the other parameters and finally get" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 144, + 562, + 287, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 562, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 144, + 562, + 287, + 578 + ], + "type": "interline_equation", + "content": "s _ {n} ^ {(1)} \\approx g _ {n} ^ {2}. \\tag {3}", + "image_path": "000bb1e9d187c072d08516bd90c63bc18eb836b77fecea75a9422e0a80dd8e88.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 580, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 689 + ], + "type": "text", + "content": "Therefore, the sensitivity of a parameter can be efficiently measured by its potential to reduce the loss on the target domain. Note that although our criterion draws inspiration from pruning work [35], it is distinct from it. [35] measures the parameter importance by the squared change in loss when removing them, i.e., " + }, + { + "bbox": [ + 46, + 580, + 287, + 689 + ], + "type": "inline_equation", + "content": "(E(\\mathcal{D}_t,\\boldsymbol {w}) - E(\\mathcal{D}_t,\\boldsymbol {w}\\mid w_n = 0))^2" + }, + { + "bbox": [ + 46, + 580, + 287, + 689 + ], + "type": "text", + "content": " and finally derives the parameter importance by " + }, + { + "bbox": [ + 46, + 580, + 287, + 689 + ], + "type": "inline_equation", + "content": "(g_{n}w_{n})^{2}" + }, + { + "bbox": [ + 46, + 580, + 287, + 689 + ], + "type": "text", + "content": ", which is different from our formulations in Eqs. (1) and (3)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "In practice, we accumulate " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": " from a total number of " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": " training samples ahead of fine-tuning to generate accu-" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 191, + 547, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 191, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 191, + 547, + 264 + ], + "type": "text", + "content": "rate sensitivity as shown in Algorithm 1, where " + }, + { + "bbox": [ + 304, + 191, + 547, + 264 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 191, + 547, + 264 + ], + "type": "text", + "content": " is a predefined hyper-parameter. In Section 4.3, we show that employing only 400 training samples is sufficient for getting reasonable parameter sensitivity, which requires only 5.5 seconds with a single GPU for any VTAB-1k dataset with ViT-B/16 backbone [12]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 272, + 527, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 272, + 527, + 285 + ], + "spans": [ + { + "bbox": [ + 305, + 272, + 527, + 285 + ], + "type": "text", + "content": "3.2. Adaptive Trainable Parameters Allocation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "text", + "content": "Our next step is to allocate trainable parameters based on the obtained parameter sensitivity set " + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "text", + "content": " and a desired parameter budget " + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "text", + "content": ". A straightforward solution is to directly tune the top-" + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "text", + "content": " most sensitive unstructured connections (parameters) while keeping the rest frozen, which we name unstructured tuning. Specifically, we select the top-" + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "text", + "content": " most sensitive weight connections in " + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "text", + "content": " to form the sensitive weight connection set " + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "text", + "content": ". Then, for a weight matrix " + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\boldsymbol{W} \\in \\mathbb{R}^{d_{\\mathrm{in}} \\times d_{\\mathrm{out}}}" + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "text", + "content": ", we can get a binary mask " + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{d_{\\mathrm{in}} \\times d_{\\mathrm{out}}}" + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "text", + "content": " computed by" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 375, + 417, + 545, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 417, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 375, + 417, + 545, + 442 + ], + "type": "interline_equation", + "content": "\\boldsymbol {M} ^ {j} = \\left\\{ \\begin{array}{l l} 1 & \\boldsymbol {W} ^ {j} \\in \\mathcal {T} \\\\ 0 & \\boldsymbol {W} ^ {j} \\notin \\mathcal {T} \\end{array} , \\right. \\tag {4}", + "image_path": "de995c220b315f677e46373afeb6b84ba5cb59468c41655defaa57ed8620ed07.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 448, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 545, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 448, + 545, + 509 + ], + "type": "inline_equation", + "content": "W^{j}" + }, + { + "bbox": [ + 304, + 448, + 545, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 448, + 545, + 509 + ], + "type": "inline_equation", + "content": "M^j" + }, + { + "bbox": [ + 304, + 448, + 545, + 509 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 304, + 448, + 545, + 509 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 448, + 545, + 509 + ], + "type": "text", + "content": "-th element in " + }, + { + "bbox": [ + 304, + 448, + 545, + 509 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 448, + 545, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 448, + 545, + 509 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 448, + 545, + 509 + ], + "type": "text", + "content": ", respectively. Accordingly, we can train the sensitive parameters by gradient descent and the updated weight matrix can be formulated as " + }, + { + "bbox": [ + 304, + 448, + 545, + 509 + ], + "type": "inline_equation", + "content": "W' \\gets W - \\epsilon g_W \\odot M" + }, + { + "bbox": [ + 304, + 448, + 545, + 509 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 448, + 545, + 509 + ], + "type": "inline_equation", + "content": "g_W" + }, + { + "bbox": [ + 304, + 448, + 545, + 509 + ], + "type": "text", + "content": " is the gradient for " + }, + { + "bbox": [ + 304, + 448, + 545, + 509 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 448, + 545, + 509 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "content": "However, considering PEFT approaches generally limit the proportion of trainable parameters to less than " + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "content": ", tuning only a small number of unstructured weight connections might not have enough representational capability to handle the downstream datasets with large domain gaps from the source pre-training data. Therefore, to improve the representational capability, we propose to replace unstructured tuning with structured tuning at the sensitive weight matrices that have a high number of sensitive parameters. To preserve the parameter budget, we can implement structured tuning with an existing efficient structured tuning PEFT method [21, 9, 20, 23] that learns to directly adjust all hidden dimensions at once. We depict an overview of our trainable parameter allocation strategy in Figure 2. For example, we can employ the low-rank reparameterization trick LoRA [21] to the sensitive weight matrices and the one-step update for " + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "content": " can be formulated as" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11828" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 52, + 70, + 288, + 94 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 70, + 288, + 94 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 288, + 94 + ], + "type": "interline_equation", + "content": "\\boldsymbol {W} ^ {\\prime} = \\left\\{ \\begin{array}{l l} \\boldsymbol {W} + \\boldsymbol {W} _ {\\mathrm {d o w n}} \\boldsymbol {W} _ {\\mathrm {u p}} & \\text {i f} \\sum_ {j = 0} ^ {d _ {\\mathrm {i n}} \\times d _ {\\mathrm {o u t}}} \\boldsymbol {M} ^ {j} \\geq \\sigma_ {\\mathrm {o p t}} \\\\ \\boldsymbol {W} - \\epsilon \\boldsymbol {g} _ {\\boldsymbol {W}} \\odot \\boldsymbol {M} & \\text {o t h e r w i s e} \\end{array} \\right.,", + "image_path": "893470f885f01555a0557a98fe10045f7de94f2d40cb52b34f30e5390205da0b.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{down}} \\in \\mathbb{R}^{d_{\\mathrm{in}} \\times r}" + }, + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{up}} \\in \\mathbb{R}^{r \\times d_{\\mathrm{out}}}" + }, + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "type": "text", + "content": " are two learnable low-rank matrices to approximate the update of " + }, + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "type": "text", + "content": " and rank " + }, + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "type": "text", + "content": " is a hyper-parameter where " + }, + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "type": "inline_equation", + "content": "r \\ll \\min(d_{\\mathrm{in}}, d_{\\mathrm{out}})" + }, + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "type": "text", + "content": ". In this way, we perform structured tuning on " + }, + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "type": "text", + "content": " when its number of sensitive parameters exceeds " + }, + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{opt}}" + }, + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "type": "text", + "content": ", whose value depends on the pre-defined type of structured tuning method. For example, since implementing structured tuning with LoRA requires " + }, + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "type": "inline_equation", + "content": "2 \\times d_{\\mathrm{in}} \\times d_{\\mathrm{out}} \\times r" + }, + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "type": "text", + "content": " trainable parameters for each sensitive weight matrix, we set " + }, + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{LoRA}} \\gets 2 \\times d_{\\mathrm{in}} \\times d_{\\mathrm{out}} \\times r" + }, + { + "bbox": [ + 47, + 95, + 287, + 245 + ], + "type": "text", + "content": " to ensure that the number of trainable parameters introduced by structured tuning is always equal to or lower than the number of sensitive parameters." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 246, + 287, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 246, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 47, + 246, + 287, + 329 + ], + "type": "text", + "content": "In this way, our SPT adaptively incorporates both structured and unstructured tuning granularities to enable higher flexibility and stronger representational power, simultaneously. In Section 4.3, we show that structured tuning is important for the downstream tasks with larger domain gaps and both unstructured and structured tuning contribute clearly to the superior performance of our SPT." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 341, + 128, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 341, + 128, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 341, + 128, + 354 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 360, + 163, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 163, + 372 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 163, + 372 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "type": "text", + "content": "Datasets and metrics. We evaluate our SPT on total 24 downstream tasks in two groups following [22]. 1) FGVC is a benchmark for fine-grained visual classification, including CUB-200-2011 [52], NABirds [50], Oxford Flowers [39], Stanford Cars [13], and Stanford Dogs [24] datasets. Each FGVC dataset contains between 55 to 200 classes and a few thousand images for train, validation, and test. We follow the validation splits in [22] if the validation set is unavailable. 2) VTAB-1k [60] is a large-scale transfer learning benchmark consisting of a collection of 19 visual classification tasks. VTAB-1k can further be divided into three groups, including Natural tasks with natural images, Specialized tasks with images captured by specialized equipment, e.g., medical images, and Structured tasks with images mostly generated from synthetic environments. Each of the VTAB-1k dataset has only 800 training and 200 validation samples, while the test set sizes vary. We use top-1 accuracy (\\%) averaged within each group as our main metric following [22]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": "Pre-trained backbones. We conduct experiments on the plain vision Transformer backbone ViT-B/16 [12] that is pre-trained on ImageNet [25] with different pre-training strategies following [22], including supervised pre-training and self-supervised pre-training with MAE [19] and MoCo v3 [10] following [22]. We also conduct experiments on the representative hierarchical vision Transformer backbone Swin-B [32] under supervised pre-training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "content": "Contenders. We categorize the baseline methods into" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": "addition-based and reparameterization-based PEFT methods as introduced in Section 2. Unless specified, all baseline methods keep the backbone frozen. Addition-based methods require extra computations during inference, including MLP-" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": ", PROMPT-SHALLOW [22], PROMPT-DEEP [22], ADAPTER-" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": " [20], ADAPTFORMER [9], and NOAH [63]. Reparameterization-based methods have no additional computational overhead during inference, including LINEAR, PARTIAL-" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": ", BIAS [59], and LORA-" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": " [21]. Here " + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": " represents the number of bottleneck dimension in ADAPTER-" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": " and LORA-" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": ". We also compare with full fine-tuning which is denoted by FULL. We introduce the details of these methods in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 334 + ], + "type": "text", + "content": "We also introduce two variants of our SPT: addition-based SPT-ADAPTER and reparameterization-based SPT-LORA. SPT-ADAPTER directly adjusts the hidden representations that are computed by sensitive weight matrices following [20], while SPT-LORA approximates updating the sensitive weight matrices following [21]. For the two variants, we follow the exact weight initializations that are described in [21] and follow [63] to set the bottleneck dimension as 8." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "text", + "content": "Implementation details. Following [63], we use the AdamW optimizer [34] with cosine learning rate decay and set the batch size, learning rate, and weight decay as 64, " + }, + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "text", + "content": ", respectively. We also follow [63] for the standard data augmentation pipeline. We set the number of training samples " + }, + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "text", + "content": " used to calculate our parameter sensitivities in Algorithm 1 universally to be 800 for the main experiments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 434, + 390, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 434, + 390, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 434, + 390, + 445 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 450, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 474 + ], + "type": "text", + "content": "We compare our method with the baseline methods under different backbones, pre-training strategies, and tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "content": "Visual recognition with ViT backbone. First, our proposed SPT-ADAPTER and SPT-LORA achieve the best performance under different trainable parameter budgets with supervised pre-trained ViT-B/16 backbone, as shown in Table 1 and Figure 3 (a). For instance, SPT-ADAPTER outperforms the SOTA method NOAA by a clear margin of " + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "content": " mean top-1 accuracy over the 19 VTAB-1k datasets with fewer trainable parameters. We speculate that our SPT variants allocate trainable parameters at task-specific positions compared to the heuristically selected positions in the baseline methods, which contributes to our superior performance. We also observe that our SPT-ADAPTER and SPT-LORA achieve large performance gains over ADAPTER and LORA variants, respectively. For example, SPT-ADAPTER and SPT-LORA with " + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "inline_equation", + "content": "0.41\\%" + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "content": " trainable parameters respectively improve ADAPTER-8 and LORA-8 significantly by " + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "inline_equation", + "content": "4.0\\%" + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "content": " mean accuracy on the FGVC benchmark. This suggests that identifying task-specific important positions and combining both unstructured and structured tuning granularities with SPT are complementary to the exist" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11829" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 96, + 70, + 496, + 112 + ], + "blocks": [ + { + "bbox": [ + 96, + 70, + 496, + 112 + ], + "lines": [ + { + "bbox": [ + 96, + 70, + 496, + 112 + ], + "spans": [ + { + "bbox": [ + 96, + 70, + 496, + 112 + ], + "type": "table", + "html": "
ViT-B/16 (85.8M)Total paramsFGVCVTAB-1k
Tuned / TotalMean Acc.Tuned / TotalNaturalSpecializedStructuredMean Acc.
FULL24.02×100%88.5100%75.983.447.669.0
", + "image_path": "52e9e580d99653c5874b355ff86b5663ee64d49f6beb7d2fd69a3b6fdc870e32.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 96, + 125, + 495, + 222 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 341, + 122 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 341, + 122 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 341, + 122 + ], + "type": "text", + "content": "Addition-based methods" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 96, + 125, + 495, + 222 + ], + "lines": [ + { + "bbox": [ + 96, + 125, + 495, + 222 + ], + "spans": [ + { + "bbox": [ + 96, + 125, + 495, + 222 + ], + "type": "table", + "html": "
MLP-31.35×1.50%79.81.42%67.872.830.657.1
PROMPT-SHALLOW1.04×0.31%84.60.13%76.879.747.067.8
PROMPT-DEEP1.18×0.98%89.11.14%78.582.455.072.0
ADAPTER-81.06×0.39%85.50.23%79.084.158.573.9
ADAPTER-321.19×0.95%85.60.71%79.684.058.374.0
ADAPTFORMER1.09×0.44%85.10.36%80.685.458.574.8
NOAH---0.52%80.284.961.375.5
SPT-ADAPTER (Ours)1.08×0.41%89.50.30%81.385.360.875.8
SPT-ADAPTER (Ours)1.10×0.47%89.80.44%82.085.861.476.4
", + "image_path": "027bc3a87db41efdad7e554ff6c9f943d205363a0eef2a4dbc7a498b81269b1e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 97, + 235, + 494, + 312 + ], + "blocks": [ + { + "bbox": [ + 230, + 224, + 361, + 234 + ], + "lines": [ + { + "bbox": [ + 230, + 224, + 361, + 234 + ], + "spans": [ + { + "bbox": [ + 230, + 224, + 361, + 234 + ], + "type": "text", + "content": "Reparameterization-based methods" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 97, + 235, + 494, + 312 + ], + "lines": [ + { + "bbox": [ + 97, + 235, + 494, + 312 + ], + "spans": [ + { + "bbox": [ + 97, + 235, + 494, + 312 + ], + "type": "table", + "html": "
LINEAR1.02×0.12%79.30.04%68.977.226.857.6
PARTIAL-13.00×8.38%82.68.30%69.478.534.260.7
BIAS1.05×0.13%88.40.13%73.378.344.165.2
LORA-81.07×0.55%86.00.23%79.584.660.574.9
LORA-161.18×0.90%84.80.69%79.884.960.275.0
SPT-LORA (Ours)1.08×0.41%89.30.31%81.585.660.775.9
SPT-LORA (Ours)1.15×0.60%90.10.63%81.985.961.376.4
", + "image_path": "d881f765a91e273c18d5a93c0f92754432bc6a187c39c349e3fd2c0246f8160e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 47, + 371, + 545, + 411 + ], + "blocks": [ + { + "bbox": [ + 46, + 316, + 547, + 366 + ], + "lines": [ + { + "bbox": [ + 46, + 316, + 547, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 547, + 366 + ], + "type": "text", + "content": "Table 1: Comparisons on FGVC and VTAB-1k [60] benchmarks using supervised pre-trained ViT-B/16 backbone pre-trained on ImageNet-21k. \"Total params\" denotes the ratio of the total number of parameters needed for all downstream tasks relative to the one for the pre-trained backbone, and \"Tuned/Total\" denotes the fraction of trainable parameters. Top-1 accuracy (\\%) is reported. The best result is in bold, and the second-best result is underlined." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 371, + 545, + 411 + ], + "lines": [ + { + "bbox": [ + 47, + 371, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 545, + 411 + ], + "type": "table", + "html": "
ViT-B/16 (85.8M)Total ParamsVTAB-1kMAEMean Acc.VTAB-1kMoCo v3Mean Acc.
Tuned / TotalNaturalSpecializedStructuredTuned / TotalNaturalSpecializedStructured
FULL38.02×100%59.379.753.864.3100%72.084.742.069.6
", + "image_path": "73d02b8ffe5baffa0fe34b679d257b37d6a48a5f316f75779d64204e3bb772e9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 48, + 424, + 545, + 491 + ], + "blocks": [ + { + "bbox": [ + 252, + 414, + 341, + 422 + ], + "lines": [ + { + "bbox": [ + 252, + 414, + 341, + 422 + ], + "spans": [ + { + "bbox": [ + 252, + 414, + 341, + 422 + ], + "type": "text", + "content": "Addition-based methods" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 424, + 545, + 491 + ], + "lines": [ + { + "bbox": [ + 48, + 424, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 424, + 545, + 491 + ], + "type": "table", + "html": "
ADAPTER-81.08×0.23%57.278.454.763.40.23%27.670.948.449.0
ADAPTER-321.28×0.95%55.378.853.362.50.99%74.282.747.768.2
PROMPT-SHALLOW1.02×0.12%40.069.727.545.70.12%67.382.337.662.4
PROMPT-DEEP1.05×0.23%36.060.626.641.10.07%70.383.042.465.2
SPT-ADAPTER (Ours)1.07×0.26%64.882.460.469.20.08%76.184.960.173.7
SPT-ADAPTER (Ours)1.13×0.41%65.682.760.769.70.30%76.685.061.774.4
", + "image_path": "018e6081b6366526f43b03ab9255167db2f7df6876dd54c0bb257b71465bfd61.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 48, + 504, + 545, + 581 + ], + "blocks": [ + { + "bbox": [ + 231, + 493, + 362, + 502 + ], + "lines": [ + { + "bbox": [ + 231, + 493, + 362, + 502 + ], + "spans": [ + { + "bbox": [ + 231, + 493, + 362, + 502 + ], + "type": "text", + "content": "Reparameterization-based methods" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 504, + 545, + 581 + ], + "lines": [ + { + "bbox": [ + 48, + 504, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 545, + 581 + ], + "type": "table", + "html": "
LINEAR1.02×0.04%18.952.723.732.10.04%67.581.130.359.6
PARTIAL-14.16×8.30%58.478.347.661.58.30%72.384.647.968.3
BIAS1.06×0.13%54.675.747.759.30.13%72.981.153.469.2
LORA-81.08×0.23%57.577.757.764.30.23%21.266.745.144.3
LORA-161.28×0.69%57.377.159.964.80.69%16.064.048.742.9
SPT-LORA (Ours)1.11×0.29%63.881.660.068.50.30%76.585.463.075.0
SPT-LORA (Ours)1.23×0.69%65.482.461.569.80.50%76.586.063.675.3
", + "image_path": "d301075368587a57e1ed6afb5b4198374554617065f022187e3aa93a37c8cc55.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 585, + 547, + 634 + ], + "lines": [ + { + "bbox": [ + 46, + 585, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 46, + 585, + 547, + 634 + ], + "type": "text", + "content": "Table 2: Comparisons on VTAB-1k [60] benchmark using self-supervised ViT-B/16 backbone pre-trained by MAE [19] and MoCo v3 [10]. \"Total params\" denotes the ratio of the total number of parameters needed for all downstream tasks relative to the one for the pre-trained backbone, and \"Tuned/Total\" denotes the fraction of trainable parameters. Top-1 accuracy (\\%) is reported. The best result is in bold, and the second-best result is underlined." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 639, + 241, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 639, + 241, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 639, + 241, + 652 + ], + "type": "text", + "content": "ing PEFT methods and boost their performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Second, SPT variants outperform baseline methods and full fine-tuning by significant margins with the self-supervised pre-trained ViT-B/16 backbones. As shown in Table 2, existing PEFT approaches exhibit inferior results than full fine-tuning with the self-supervised pre-trained" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 639, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 547, + 712 + ], + "type": "text", + "content": "backbones MAE and MoCo v3. It is worth noting that previous PEFT methods yield inconsistent results with the backbones of different pre-training strategies. In contrast, SPT variants consistently outperform full fine-tuning. In particular, SPT-ADAPTER achieves remarkable " + }, + { + "bbox": [ + 304, + 639, + 547, + 712 + ], + "type": "inline_equation", + "content": "5.8\\%" + }, + { + "bbox": [ + 304, + 639, + 547, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 639, + 547, + 712 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 304, + 639, + 547, + 712 + ], + "type": "text", + "content": " mean top-1 accuracy gains over the best-performing" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11830" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 287, + 205 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 287, + 205 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 287, + 205 + ], + "type": "table", + "html": "
MethodTuned / TotalNaturalSpecializedStructuredMean / Acc.
FULL100%79.186.259.775.0
Addition-based methods
MLP-31.60%73.675.235.761.5
PROMPT-SHALLOW0.04%79.982.537.866.7
PROMPT-DEEP0.23%76.884.553.471.6
ADAPTER-81.18%81.787.361.276.7
SPT-ADAPTER (ours)0.33%83.087.362.177.5
Reparameterization-based methods
LINEAR0.04%73.580.833.562.6
PARTIAL-12.15%73.181.735.063.3
LORA-81.18%81.787.260.176.3
SPT-LORA (ours)0.49%83.187.460.477.2
", + "image_path": "b83b221b6a8712fcf9845dcf61b273fa9ff05a9477f28e352e9221fec82a7972.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 265, + 287, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 287, + 409 + ], + "type": "text", + "content": "baseline method on VTAB-1k benchmark with only " + }, + { + "bbox": [ + 46, + 265, + 287, + 409 + ], + "type": "inline_equation", + "content": "0.26\\%" + }, + { + "bbox": [ + 46, + 265, + 287, + 409 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 265, + 287, + 409 + ], + "type": "inline_equation", + "content": "0.08\\%" + }, + { + "bbox": [ + 46, + 265, + 287, + 409 + ], + "type": "text", + "content": " trainable parameters for MAE and MoCo v3 pretrained backbones, respectively. Moreover, our observation in supplementary material suggests that self-supervised pretrained ViT backbones have more diverse sensitivity distributions and a higher variance in sensitivity across different tasks than the supervised pre-trained one. This leads to the conjecture that baseline methods which assign trainable parameters to the same positions for all tasks may fail to mitigate the distinct domain gaps in individual downstream datasets, whereas our SPT allocates trainable parameters to task-specific positions accurately." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 410, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 287, + 529 + ], + "type": "text", + "content": "Visual recognition with Swin and ConvNeXt backbones. From Table 3, we observe that our SPT-LORA and SPT-ADAPTER also achieve SOTA performance with Swin-B backbone on all dataset groups. We also follow VPT [22] to apply SPT variants to ResNet-alike architecture ConvNeXt-Base [33] and report the results in Table 4. We observe that SPT variants achieve better trade-offs between accuracy and parameter efficiency than the baseline methods. The results further demonstrate the versatility and effectiveness of our SPT." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 531, + 288, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 288, + 664 + ], + "type": "text", + "content": "Semantic segmentation. We follow VPT [22] to conduct semantic segmentation on ADE20k dataset. Following the settings of [22], we apply our SPT to SETR-PUP [65] framework with ViT-L backbone and only allocate trainable parameters to the backbone with the head fully finetuned. We report the mIoU results in Table 5. Notably, SPT-LORA and SPT-ADAPTER outperform the baseline methods by large margins, indicating our SPT can be generalized to the semantic segmentation task. We report the results for conducting only structured tuning for both SPT-LORA and SPT-ADAPTER as it yields higher mIoU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 672, + 141, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 672, + 141, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 672, + 141, + 685 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Effect of the sensitivity criterion. We investigate the effectiveness of our sensitivity criterion on VTAB-1k by em" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 545, + 128 + ], + "blocks": [ + { + "bbox": [ + 46, + 208, + 288, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 208, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 288, + 257 + ], + "type": "text", + "content": "Table 3: Comparisons on VTAB-1k [60] benchmark with supervised pre-trained Swin-B [32]. \"Tuned/Total\" denotes the fraction of trainable parameters. Top-1 accuracy (\\%) is reported. The best result is in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 70, + 545, + 128 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 128 + ], + "type": "table", + "html": "
MethodTuned/TotalNaturalSpecializedStructuredMean Acc.
FULL100%78.083.760.474.0
LoRA0.79%82.284.764.177.0
Adapter0.47%83.184.964.677.5
SPT-LoRA (ours)0.57%83.486.765.978.7
SPT-Adapter (ours)0.36%83.786.265.378.4
", + "image_path": "c13cb661bdeb2528c91830e3547653868a6458a6eceb6dc87c0a6f32b4447eab.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 196, + 545, + 262 + ], + "blocks": [ + { + "bbox": [ + 305, + 136, + 545, + 185 + ], + "lines": [ + { + "bbox": [ + 305, + 136, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 305, + 136, + 545, + 185 + ], + "type": "text", + "content": "Table 4: Comparisons on VTAB-1k [60] benchmark with supervised pre-trained ConvNeXt-Base [33] backbone. \"Tuned/Total\" denotes the fraction of trainable parameters. Top-1 accuracy " + }, + { + "bbox": [ + 305, + 136, + 545, + 185 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 305, + 136, + 545, + 185 + ], + "type": "text", + "content": " is reported. The best result is in bold." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 196, + 545, + 262 + ], + "lines": [ + { + "bbox": [ + 307, + 196, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 545, + 262 + ], + "type": "table", + "html": "
MethodmIoU-s.s. (%)mIoU-m.s. (%)Trainable Param.
VPT44.045.615.8M
LoRA43.945.914.7M
Adapter44.446.614.6M
SPT-LoRA (ours)45.447.514.6M
SPT-Adapter (ours)45.247.214.6M
", + "image_path": "a3b9737a6fd48c3da997bede7dad193f87d56b76c6275e54758785bc5ff94ef2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 271, + 545, + 319 + ], + "lines": [ + { + "bbox": [ + 305, + 271, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 271, + 545, + 319 + ], + "type": "text", + "content": "Table 5: Semantic Segmentation: Comparisons on ADE20k [66] val with SETR [65] on ViT-L backbone. We report both single-scale (s.s.) and multi-scale (m.s.) mIoU results. The best result is in bold." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 342, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 449 + ], + "type": "text", + "content": "ploying structured tuning methods from [22, 20, 21] to the task-specific sensitive weight matrices. Note that we do not conduct unstructured tuning to ensure fair comparisons. The results are presented in Figure 3 (b). Our criterion brings consistent " + }, + { + "bbox": [ + 304, + 342, + 545, + 449 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 304, + 342, + 545, + 449 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 342, + 545, + 449 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 304, + 342, + 545, + 449 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 342, + 545, + 449 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 304, + 342, + 545, + 449 + ], + "type": "text", + "content": " performance gains for PROMPT-DEEP, ADAPTER-32, and LORA-16, respectively, which demonstrates the effectiveness and versatility of our sensitivity criterion to identify accurate task-specific important positions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 450, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 546, + 714 + ], + "type": "text", + "content": "Effect of structured and unstructured tuning. We investigate the effectiveness of unstructured and structured tuning individually on VTAB-1k. The results are presented in Table 6. We start by applying ADAPTER-8 to the sensitive weight matrices identified by our sensitivity criterion (SPT-ADAPTER w/o unstructured). We observe that our sensitivity criterion boosts the performance of all the dataset groups by clear margins, which again demonstrates the importance of our sensitivity criterion. Next, we observe that allocating the trainable parameters to the unstructured sensitive weight connections also brings accuracy improvement to the Natural and Specialized datasets from ADAPTER-8. However, we find that structured tuning is especially important for achieving good performance on Structured datasets. To further investigate this phenomenon, we observe that Structured datasets have larger domain gaps from the pre-training source domain [25] compared to Natural and Specialized datasets as visualized in Figure 3 (c). We hence conjecture that structured tuning has a higher representational capability than unstructured tuning which facilitates mitigating the large domain gaps during fine-tuning (see the supplementary material for visual examples). Finally, we observe" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11831" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 72, + 206, + 198 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 206, + 198 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 206, + 198 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 206, + 198 + ], + "type": "image", + "image_path": "fecd4559bdd11b1405641730406683e2e4347922e441880fffb05fb5ddd1f62d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 547, + 310 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 310 + ], + "type": "text", + "content": "Figure 3: (a) Accuracy vs. parameter efficiency with supervised pre-trained ViT-B/16 backbone on VTAB-1k [60]. SPT variants perform favorably against the other PEFT approaches and are more scalable. (b) Applying other PEFT structured tuning methods [22, 20, 21] to the task-specific sensitive weight matrices (denoted by TSM) identified by our criterion with supervised pre-trained ViT-B/16 backbone on VTAB-1k. Our criterion brings consistent performance gain. (c) Domain vs. performance gaps for different dataset groups in VTAB-1k [60]. The blue bars show the domain gaps between the source domain (ImageNet [25]) and target domains, which are measured by Maximum Mean Discrepancy (MMD) distance [49]. The red line represents the performance gaps between SPT-ADAPTER w/o unstructured and w/o structured, using supervised pre-trained ViT-B/16 backbone. The dataset groups are Natural, Specialized, and Structured. Structured tuning is important for achieving good performance on Structured datasets with larger domain gaps." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 209, + 72, + 369, + 198 + ], + "blocks": [ + { + "bbox": [ + 209, + 72, + 369, + 198 + ], + "lines": [ + { + "bbox": [ + 209, + 72, + 369, + 198 + ], + "spans": [ + { + "bbox": [ + 209, + 72, + 369, + 198 + ], + "type": "image", + "image_path": "91c0b01112e12170c50af13790a69d271301b7040b4ac17299ae9d52f3719831.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 372, + 76, + 545, + 198 + ], + "blocks": [ + { + "bbox": [ + 372, + 76, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 372, + 76, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 372, + 76, + 545, + 198 + ], + "type": "image", + "image_path": "c46225b41d6523ea95dde04abeda37757627658fc77b5ac4db421db94dfe578c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 47, + 318, + 287, + 370 + ], + "blocks": [ + { + "bbox": [ + 47, + 318, + 287, + 370 + ], + "lines": [ + { + "bbox": [ + 47, + 318, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 287, + 370 + ], + "type": "table", + "html": "
MethodTuned / TotalNaturalSpecializedStructuredMean / Acc.
ADAPTER-80.23%79.084.158.573.9
SPT-ADAPTER w/o unstructured0.29%81.285.160.375.5
SPT-ADAPTER w/o structured0.34%81.285.059.675.3
SPT-ADAPTER0.30%81.385.360.875.8
", + "image_path": "37874411b5d15c1141282c71b549b1cdae33c9c9a8bb12c7887a4c66e5c92973.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 486 + ], + "type": "text", + "content": "that incorporating both structured and unstructured tuning at task-specific important positions achieves the highest performance on all dataset groups." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 487, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 629 + ], + "type": "text", + "content": "Effect of number of training samples " + }, + { + "bbox": [ + 46, + 487, + 287, + 629 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 487, + 287, + 629 + ], + "type": "text", + "content": " to get parameter sensitivity. We investigate the effect of the number of training images " + }, + { + "bbox": [ + 46, + 487, + 287, + 629 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 487, + 287, + 629 + ], + "type": "text", + "content": " for calculating our parameter sensitivity (Algorithm 1 of the main paper). We randomly sample training samples and report the mean results over three runs in Table 8. We find that our SPT is robust to the number of training samples " + }, + { + "bbox": [ + 46, + 487, + 287, + 629 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 487, + 287, + 629 + ], + "type": "text", + "content": " and randomly sampling 400 out of a total of 800 training samples is sufficient to obtain accurate task-specific important positions, e.g., calculating the sensitivity for ViT-B/16 backbone takes only 5.5 seconds with a single GPU on any of the VTAB-1k datasets and this computation is required only once." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "Computational cost analysis. We investigate the computational cost of SPT-LoRA by comparing with full fine-tuning, addition-based method PROMPT-DEEP, and reparameterization-based method LORA-16. The results are presented in Table 7. We observe that PROMPT-DEEP has higher inference latency and inference GPU memory due to the additional prompts. In contrast, since the up" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 306, + 318, + 545, + 374 + ], + "blocks": [ + { + "bbox": [ + 46, + 378, + 287, + 437 + ], + "lines": [ + { + "bbox": [ + 46, + 378, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 287, + 437 + ], + "type": "text", + "content": "Table 6: Ablation study on structured and unstructured tuning only with supervised pre-trained ViT-B/16 backbone. Top-1 accuracy " + }, + { + "bbox": [ + 46, + 378, + 287, + 437 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 378, + 287, + 437 + ], + "type": "text", + "content": " is reported. We set different parameter constraints to align the fractions of trainable parameters for these cases. The best result is in bold" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 318, + 545, + 374 + ], + "lines": [ + { + "bbox": [ + 306, + 318, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 306, + 318, + 545, + 374 + ], + "type": "table", + "html": "
MethodInference Latency (ms/img)Inference Memory (GB)Fine-tuning Memory (GB)
FULL2.81.311.9
PROMPT-DEEP3.81.913.2
LORA-162.81.38.2
SPT-LORA w/o unstructured2.81.38.3
SPT-LORA2.81.39.8
", + "image_path": "4c158c3dd426cd4e6175ebe86db26ad61e2f63309b2d28e9e2df03a67bfdbd2a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 353, + 468, + 498, + 500 + ], + "blocks": [ + { + "bbox": [ + 305, + 378, + 546, + 449 + ], + "lines": [ + { + "bbox": [ + 305, + 378, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 378, + 546, + 449 + ], + "type": "text", + "content": "Table 7: Cost comparisons with ViT-B/16 backbone on the VTAB-1k [60] benchmark with around " + }, + { + "bbox": [ + 305, + 378, + 546, + 449 + ], + "type": "inline_equation", + "content": "0.70\\%" + }, + { + "bbox": [ + 305, + 378, + 546, + 449 + ], + "type": "text", + "content": " fractions of the trainable parameters. We report the latency (ms/img) and the peak memory usage (GB) with image resolution " + }, + { + "bbox": [ + 305, + 378, + 546, + 449 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 305, + 378, + 546, + 449 + ], + "type": "text", + "content": " on a single GeForce 3090 GPU. The best result is in bold." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 353, + 468, + 498, + 500 + ], + "lines": [ + { + "bbox": [ + 353, + 468, + 498, + 500 + ], + "spans": [ + { + "bbox": [ + 353, + 468, + 498, + 500 + ], + "type": "table", + "html": "
C240400560800
Mean Acc.76.376.476.476.4
", + "image_path": "48c2a6d692fd9e18b59e5986b829ab8e252bdf94d9fb54dd7226cede7ae38e48.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 504, + 545, + 552 + ], + "lines": [ + { + "bbox": [ + 305, + 504, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 305, + 504, + 545, + 552 + ], + "type": "text", + "content": "Table 8: Effect of the number of training samples used to get the sensitivity for SPT-LoRA with supervised pre-trained ViT-B/16 backbone on VTAB-1k [60]. Top-1 accuracy (%) is reported. The best result is in bold." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 571, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 571, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 546, + 715 + ], + "type": "text", + "content": "dated parameters after fine-tuning can be reparameterized and merged into the pre-trained model, our SPT-LoRA, SPT-LoRA w/o unstructured tuning, and LoRA-16 are more efficient than PROMPT-DEEP during inference. However, we observe that our SPT-LoRA has slightly higher fine-tuning memory than the full fine-tuning and LoRA-16 which is taken up by updating the unstructurally-tuned parameters with sparse gradients in Eq. (5). Additionally, for memory-intensive scenarios, one can employ SPT-LoRA w/o unstructured tuning for improved performance (0.8% Top-1 accuracy on VTAB-1k) and similar fine-tuning memory as LoRA, as shown in Figure 3 (b) and Table 6." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11832" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 86, + 206, + 201 + ], + "blocks": [ + { + "bbox": [ + 94, + 76, + 175, + 84 + ], + "lines": [ + { + "bbox": [ + 94, + 76, + 175, + 84 + ], + "spans": [ + { + "bbox": [ + 94, + 76, + 175, + 84 + ], + "type": "text", + "content": "Supervised pre-trained ViT-B" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 86, + 206, + 201 + ], + "lines": [ + { + "bbox": [ + 52, + 86, + 206, + 201 + ], + "spans": [ + { + "bbox": [ + 52, + 86, + 206, + 201 + ], + "type": "image", + "image_path": "dcb4479f9e36da7a176aef5420a12aa733401cc06b89e1800dde4aa6e7a7cd66.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 218, + 86, + 372, + 201 + ], + "blocks": [ + { + "bbox": [ + 267, + 76, + 333, + 84 + ], + "lines": [ + { + "bbox": [ + 267, + 76, + 333, + 84 + ], + "spans": [ + { + "bbox": [ + 267, + 76, + 333, + 84 + ], + "type": "text", + "content": "MAE pre-trained ViT-B" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 218, + 86, + 372, + 201 + ], + "lines": [ + { + "bbox": [ + 218, + 86, + 372, + 201 + ], + "spans": [ + { + "bbox": [ + 218, + 86, + 372, + 201 + ], + "type": "image", + "image_path": "a20dc078f3323546bbe037059cb0ffe8c468d3fe89167440b381dad03f292337.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 383, + 86, + 541, + 201 + ], + "blocks": [ + { + "bbox": [ + 433, + 76, + 512, + 84 + ], + "lines": [ + { + "bbox": [ + 433, + 76, + 512, + 84 + ], + "spans": [ + { + "bbox": [ + 433, + 76, + 512, + 84 + ], + "type": "text", + "content": "MoCo V3 pre-trained ViT-B" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 383, + 86, + 541, + 201 + ], + "lines": [ + { + "bbox": [ + 383, + 86, + 541, + 201 + ], + "spans": [ + { + "bbox": [ + 383, + 86, + 541, + 201 + ], + "type": "image", + "image_path": "3c4674f60c4f5db37bf71b1b05fa21a7117f87e0d605b4e8e426893a45353170.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 52, + 208, + 206, + 323 + ], + "blocks": [ + { + "bbox": [ + 52, + 208, + 206, + 323 + ], + "lines": [ + { + "bbox": [ + 52, + 208, + 206, + 323 + ], + "spans": [ + { + "bbox": [ + 52, + 208, + 206, + 323 + ], + "type": "image", + "image_path": "897aed02949367588ae110b14be073a14162de7848bdd9c4054d053132b051a3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 327, + 546, + 365 + ], + "lines": [ + { + "bbox": [ + 46, + 327, + 546, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 546, + 365 + ], + "type": "text", + "content": "Figure 4: Parameter sensitivity patterns under 0.4M trainable parameter budget for ViT-B/16 backbone with different pretraining strategies on three sample tasks from VTAB-1k [61]. The proportions exhibit task-specific varying patterns in terms of network depth (upper figures) and task-agnostic similar patterns in terms of operations (lower figures)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 219, + 208, + 372, + 323 + ], + "blocks": [ + { + "bbox": [ + 219, + 208, + 372, + 323 + ], + "lines": [ + { + "bbox": [ + 219, + 208, + 372, + 323 + ], + "spans": [ + { + "bbox": [ + 219, + 208, + 372, + 323 + ], + "type": "image", + "image_path": "e3d68d8f8babb9ecc5bc562070062d1af7b5d1c01bb7fd962d3bc3478b7fbadb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 383, + 206, + 541, + 324 + ], + "blocks": [ + { + "bbox": [ + 383, + 206, + 541, + 324 + ], + "lines": [ + { + "bbox": [ + 383, + 206, + 541, + 324 + ], + "spans": [ + { + "bbox": [ + 383, + 206, + 541, + 324 + ], + "type": "image", + "image_path": "9aaa447abff829082dd93d1feb56772225b8dc2d55604cf3d85e4945eeeb6364.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 374, + 239, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 374, + 239, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 374, + 239, + 387 + ], + "type": "text", + "content": "4.4. Observations on Sensitivity Patterns" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "text", + "content": "Our sensitivity criterion identifies task-specific important positions, which can reveal the contributions of the pretrained weights to different downstream tasks during transfer learning. We visualize the proportions of the sensitive parameters for the supervised pre-trained ViT-B/16 backbone under 0.4M trainable parameter budget in Figure 4. First, we investigate the most sensitive blocks, whose numbers of sensitive parameters are summed and normalized over the 12 ViT-B/16 blocks. We observe that the patterns of the sensitive parameter proportions vary markedly across different tasks, which echoes the observations made in [16]. This suggests that we should not introduce trainable parameters to the same positions for each individual task but allocate trainable parameters at task-specific ones as we proposed. Next, we investigate the most insensitive weight matrices within a block. A ViT block consists of a query " + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "inline_equation", + "content": "W_{q}" + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "text", + "content": ", a key " + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "text", + "content": ", a value " + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "inline_equation", + "content": "W_{v}" + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "text", + "content": ", and an output " + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "inline_equation", + "content": "W_{o}" + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "text", + "content": " weight matrices in the multi-head self-attention layer and two weight matrices " + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "inline_equation", + "content": "W_{fc1}" + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "inline_equation", + "content": "W_{fc2}" + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "text", + "content": " in the feed-forward network as elaborated in [51, 12]. We observe that the query " + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "inline_equation", + "content": "W_{q}" + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "text", + "content": " and key " + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "text", + "content": " weight matrices have the lowest proportions of sensitive parameters for all three sample tasks. Since " + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "inline_equation", + "content": "W_{q}" + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 46, + 391, + 289, + 713 + ], + "type": "text", + "content": " are responsible for learning the attention scores which indicate the pairwise similarity among the patches, we speculate that although domain changes, the patch relationships learned during pre-training can be efficiently reused when transferred to downstream classification tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 373, + 378, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 378, + 386 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 378, + 386 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 391, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 643 + ], + "type": "text", + "content": "In this paper, we have explored identifying and allocating trainable parameters to task-specific important positions for visual parameter-efficient tuning. Specifically, we have proposed a novel criterion to quickly measure the sensitivity of the pre-trained parameters for each specific task before fine-tuning. Based on the parameter sensitivity, we have proposed a trainable parameter allocation strategy that adaptively combines both unstructured and structured tuning under a desired trainable parameter budget, enabling high representational capability and flexibility. Finally, we have conducted extensive experiments on a total of 24 downstream recognition tasks with both plain and hierarchical vision Transformer backbones under different pre-training strategies to demonstrate the versatility and effectiveness of our proposed SPT. Notably, we have shown that our approach is complementary to the existing PEFT methods and improves their performance significantly. In the future, we will explore adapting large vision models to more downstream tasks with SPT, e.g., dense prediction and vision-and-language tasks, and improve the training efficiency of SPT for on-device training [6, 30]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 643, + 545, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 643, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 304, + 643, + 545, + 690 + ], + "type": "text", + "content": "Acknowledgement. We thank Jing liu and Ziyi Liu for their helpful discussions. This research is partially supported by Monash FIT Start-up Grant. Dr. Jing Zhang is supported by the Australian Research Council project FL-170100117." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11833" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Y. Bao, Y. Li, S.-L. Huang, L. Zhang, L. Zheng, A. Zamir, and L. Guibas. An information-theoretic approach to transferability in task transfer learning. In ICIP, pages 2309-2313. IEEE, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 140, + 288, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 140, + 288, + 174 + ], + "spans": [ + { + "bbox": [ + 53, + 140, + 288, + 174 + ], + "type": "text", + "content": "[2] G. Bender, P.-J. Kindermans, B. Zoph, V. Vasudevan, and Q. Le. Understanding and simplifying one-shot architecture search. In ICML, pages 550-559. PMLR, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 178, + 287, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 178, + 287, + 210 + ], + "spans": [ + { + "bbox": [ + 53, + 178, + 287, + 210 + ], + "type": "text", + "content": "[3] A. Brock, T. Lim, J. M. Ritchie, and N. Weston. Smash: one-shot model architecture search through hypernetworks. In ICLR, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 287, + 249 + ], + "type": "text", + "content": "[4] S. Caelles, K.-K. Maninis, J. Pont-Tuset, L. Leal-Taixe, D. Cremers, and L. Van Gool. One-shot video object segmentation. In CVPR, pages 221-230, 2017. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 254, + 286, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 254, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 53, + 254, + 286, + 285 + ], + "type": "text", + "content": "[5] H. Cai, C. Gan, T. Wang, Z. Zhang, and S. Han. Once-for-all: Train one network and specialize it for efficient deployment. In ICLR, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 292, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 287, + 323 + ], + "type": "text", + "content": "[6] H. Cai, C. Gan, L. Zhu, and S. Han. Tinytl: Reduce memory, not parameters for efficient on-device learning. NeurIPS, 33:11285-11297, 2020. 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 329, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 287, + 361 + ], + "type": "text", + "content": "[7] H. Cai, L. Zhu, and S. Han. ProxylessNAS: Direct neural architecture search on target task and hardware. In ICLR, 2019. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 367, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 367, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 53, + 367, + 287, + 399 + ], + "type": "text", + "content": "[8] N. S. Chatterji, B. Neyshabur, and H. Sedghi. The intriguing role of module criticality in the generalization of deep networks. In ICLR, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 405, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 405, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 53, + 405, + 287, + 437 + ], + "type": "text", + "content": "[9] S. Chen, C. Ge, Z. Tong, J. Wang, Y. Song, J. Wang, and P. Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 2022. 1, 2, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 442, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 442, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 442, + 287, + 475 + ], + "type": "text", + "content": "[10] X. Chen, S. Xie, and K. He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 480, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 287, + 514 + ], + "type": "text", + "content": "[11] Y. Cui, Y. Song, C. Sun, A. Howard, and S. Belongie. Large scale fine-grained categorization and domain-specific transfer learning. In CVPR, pages 4109-4118, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 518, + 287, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 518, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 48, + 518, + 287, + 572 + ], + "type": "text", + "content": "[12] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 4, 5, 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "type": "text", + "content": "[13] T. Gebru, J. Krause, Y. Wang, D. Chen, J. Deng, and L. Fei-Fei. Fine-grained car detection for visual census estimation. In AAAI, 2017. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 616, + 287, + 637 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 616, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 48, + 616, + 287, + 637 + ], + "type": "text", + "content": "[14] D. Guo, A. Rush, and Y. Kim. Parameter-efficient transfer learning with diff pruning. In ACL-IJCNLP, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 643, + 287, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 643, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 48, + 643, + 287, + 675 + ], + "type": "text", + "content": "[15] Y. Guo, Y. Li, L. Wang, and T. Rosing. Adafilter: Adaptive filter fine-tuning for deep transfer learning. In AAAI, pages 4060-4066, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[16] Y. Guo, H. Shi, A. Kumar, K. Grauman, T. Rosing, and R. Feris. Spottune: transfer learning through adaptive fin-tuning. In CVPR, pages 4805-4814, 2019. 3, 9" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[17] T. Hao, H. Chen, Y. Guo, and G. Ding. Consolidator: Mergable adapter with group connections for vision transformer. In ICLR, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 110, + 545, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 110, + 545, + 142 + ], + "spans": [ + { + "bbox": [ + 307, + 110, + 545, + 142 + ], + "type": "text", + "content": "[18] J. He, C. Zhou, X. Ma, T. Berg-Kirkpatrick, and G. Neubig. Towards a unified view of parameter-efficient transfer learning. In ICLR, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 146, + 545, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 146, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 307, + 146, + 545, + 178 + ], + "type": "text", + "content": "[19] K. He, X. Chen, S. Xie, Y. Li, P. Dollár, and R. Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 182, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 182, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 307, + 182, + 545, + 226 + ], + "type": "text", + "content": "[20] N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. De Laroussilhe, A. Gesmundo, M. Attariyan, and S. Gelly. Parameter-efficient transfer learning for nlp. In ICML, pages 2790-2799, 2019. 1, 2, 4, 5, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 230, + 545, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 272 + ], + "type": "text", + "content": "[21] E. J. Hu, yelong shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 1, 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 277, + 545, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 277, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 545, + 309 + ], + "type": "text", + "content": "[22] M. Jia, L. Tang, B.-C. Chen, C. Cardie, S. Belongie, B. Hariharan, and S.-N. Lim. Visual prompt tuning. In ECCV, 2022. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 313, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 313, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 313, + 545, + 346 + ], + "type": "text", + "content": "[23] S. Jie and Z.-H. Deng. Convolutional bypasses are better vision transformer adapters. arXiv preprint arXiv:2207.07039, 2022. 1, 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 350, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 381 + ], + "type": "text", + "content": "[24] A. Khosla, N. Jayadevaprakash, B. Yao, and L. Fei-Fei. Novel dataset for fine-grained image categorization. In CVPRW, 2011. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 386, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 386, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 386, + 545, + 418 + ], + "type": "text", + "content": "[25] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 5, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 422, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 455 + ], + "type": "text", + "content": "[26] A. Kumar, A. Raghunathan, R. M. Jones, T. Ma, and P. Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In ICLR, 2022. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 458, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 458, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 307, + 458, + 545, + 481 + ], + "type": "text", + "content": "[27] B. Lester, R. Al-Rfou, and N. Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 483, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 483, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 307, + 483, + 545, + 517 + ], + "type": "text", + "content": "[28] X. L. Li and P. Liang. Prefix-tuning: Optimizing continuous prompts for generation. In IJCNLP, pages 4582-4597. Association for Computational Linguistics, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 520, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 552 + ], + "type": "text", + "content": "[29] D. Lian, D. Zhou, J. Feng, and X. Wang. Scaling & shifting your features: A new baseline for efficient model tuning. NeurIPS, 2022. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 556, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 590 + ], + "type": "text", + "content": "[30] J. Lin, L. Zhu, W.-M. Chen, W.-C. Wang, C. Gan, and S. Han. On-device training under 256kb memory. arXiv preprint arXiv:2206.15472, 2022.9" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 593, + 545, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 545, + 614 + ], + "type": "text", + "content": "[31] H. Liu, K. Simonyan, and Y. Yang. DARTS: Differentiable architecture search. In ICLR, 2019. 4" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 618, + 545, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 618, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 307, + 618, + 545, + 651 + ], + "type": "text", + "content": "[32] Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 5, 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 654, + 545, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 654, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 307, + 654, + 545, + 687 + ], + "type": "text", + "content": "[33] Z. Liu, H. Mao, C.-Y. Wu, C. Feichtenhofer, T. Darrell, and S. Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 7" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[34] I. Loshchilov and F. Hutter. Fixing weight decay regularization in adam, 2018. 5" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11834" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[35] P. Molchanov, A. Mallya, S. Tyree, I. Frosio, and J. Kautz. Importance estimation for neural network pruning. In CVPR, pages 11264-11272, 2019. 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 109, + 288, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 109, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 48, + 109, + 288, + 144 + ], + "type": "text", + "content": "[36] M. M. Naseer, K. Ranasinghe, S. H. Khan, M. Hayat, F. Shahbaz Khan, and M.-H. Yang. Intriguing properties of vision transformers. *NeurIPS*, 34:23296-23308, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 146, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 146, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 48, + 146, + 287, + 178 + ], + "type": "text", + "content": "[37] B. Neyshabur, H. Sedghi, and C. Zhang. What is being transferred in transfer learning? NeurIPS, 33:512-523, 2020. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 183, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 183, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 183, + 287, + 228 + ], + "type": "text", + "content": "[38] C. Nguyen, T. Hassner, M. Seeger, and C. Archambeau. Leep: A new measure to evaluate transferability of learned representations. In ICML, pages 7294-7305. PMLR, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 287, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 264 + ], + "type": "text", + "content": "[39] M.-E. Nilsback and A. Zisserman. Automated flower classification over a large number of classes. In ICVGIP, pages 722-729. IEEE, 2008. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 268, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 268, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 268, + 287, + 300 + ], + "type": "text", + "content": "[40] J. Pleted and T. Gedeon. Deep transfer learning for image classification: a survey. arXiv preprint arXiv:2205.09904, 2022.3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 304, + 287, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 349 + ], + "type": "text", + "content": "[41] J. Puigcerver, C. Riquelme, B. Mustafa, C. Renggli, A. S. Pinto, S. Gelly, D. Keysers, and N. Houlsby. Scalable transfer learning with expert models. arXiv preprint arXiv:2009.13239, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 353, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 287, + 396 + ], + "type": "text", + "content": "[42] M. Raghu, T. Unterthiner, S. Kornblith, C. Zhang, and A. Dosovitskiy. Do vision transformers see like convolutional neural networks? NeurIPS, 34:12116-12128, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 400, + 287, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 400, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 48, + 400, + 287, + 433 + ], + "type": "text", + "content": "[43] M. T. Rosenstein, Z. Marx, L. P. Kaelbling, and T. G. Dietterich. To transfer or not to transfer. In NeurlPSW, 2005. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 437, + 287, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 460 + ], + "type": "text", + "content": "[44] R. K. Srivastava, K. Greff, and J. Schmidhuber. Training very deep networks. NeurIPS, 28, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 463, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 463, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 48, + 463, + 287, + 496 + ], + "type": "text", + "content": "[45] X. Sun, R. Panda, R. Feris, and K. Saenko. Adashare: Learning what to share for efficient deep multi-task learning. NeurIPS, 33:8728-8740, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "text", + "content": "[46] Y.-L. Sung, J. Cho, and M. Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. arXiv preprint arXiv:2206.06522, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "type": "text", + "content": "[47] A. T. Tran, C. V. Nguyen, and T. Hassner. Transferability and hardness of supervised classification tasks. In ICCV, pages 1395–1405, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 574, + 287, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 574, + 287, + 618 + ], + "spans": [ + { + "bbox": [ + 48, + 574, + 287, + 618 + ], + "type": "text", + "content": "[48] C.-H. Tu, Z. Mai, and W.-L. Chao. Visual query tuning: Towards effective usage of intermediate representations for parameter and memory efficient transfer learning. arXiv preprint arXiv:2212.03220, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 621, + 287, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 621, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 621, + 287, + 654 + ], + "type": "text", + "content": "[49] E. Tzeng, J. Hoffman, N. Zhang, K. Saenko, and T. Darrell. Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:1412.3474, 2014. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 714 + ], + "type": "text", + "content": "[50] G. Van Horn, S. Branson, R. Farrell, S. Haber, J. Barry, P. Ipeirotis, P. Perona, and S. Belongie. Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In CVPR, pages 595–604, 2015. 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 692 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "type": "text", + "content": "[51] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. NeurIPS, pages 5998-6008, 2017. 9" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 110, + 546, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 110, + 546, + 144 + ], + "spans": [ + { + "bbox": [ + 307, + 110, + 546, + 144 + ], + "type": "text", + "content": "[52] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The caltech-ucsd birds-200-2011 dataset. Tech. Rep. CNS-TR-2011-001, California Institute of Technology, 2011. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 147, + 546, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 147, + 546, + 178 + ], + "spans": [ + { + "bbox": [ + 307, + 147, + 546, + 178 + ], + "type": "text", + "content": "[53] Z. Wang, Z. Dai, B. Póczos, and J. Carbonell. Characterizing and avoiding negative transfer. In CVPR, pages 11293-11302, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 183, + 546, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 183, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 183, + 546, + 217 + ], + "type": "text", + "content": "[54] H. Wu, J. Xu, J. Wang, and M. Long. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. NeurIPS, 34:22419-22430, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 220, + 546, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 546, + 254 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 546, + 254 + ], + "type": "text", + "content": "[55] R. Xu, F. Luo, Z. Zhang, C. Tan, B. Chang, S. Huang, and F. Huang. Raise a child in large language model: Towards effective and generalizable fine-tuning. In EMNLP, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 258, + 546, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 258, + 546, + 290 + ], + "spans": [ + { + "bbox": [ + 307, + 258, + 546, + 290 + ], + "type": "text", + "content": "[56] Y. Xu, Q. Zhang, J. Zhang, and D. Tao. Vitae: Vision transformer advanced by exploring intrinsic inductive bias. In NeurIPS, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 294, + 546, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 546, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 546, + 327 + ], + "type": "text", + "content": "[57] J. Yoon, S. Arik, and T. Pfister. Data valuation using reinforcement learning. In ICML, pages 10842-10851. PMLR, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 331, + 546, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 546, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 546, + 364 + ], + "type": "text", + "content": "[58] J. Yosinski, J. Clune, Y. Bengio, and H. Lipson. How transferable are features in deep neural networks? NeurIPS, 27, 2014. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 368, + 546, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 368, + 546, + 401 + ], + "spans": [ + { + "bbox": [ + 307, + 368, + 546, + 401 + ], + "type": "text", + "content": "[59] E. B. Zaken, Y. Goldberg, and S. Ravfogel. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In ACL, pages 1-9, 2022. 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 405, + 546, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 546, + 460 + ], + "type": "text", + "content": "[60] X. Zhai, J. Puigcerver, A. Kolesnikov, P. Ruyssen, C. Riquelme, M. Lucic, J. Djolonga, A. S. Pinto, M. Neumann, A. Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 463, + 546, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 463, + 546, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 463, + 546, + 518 + ], + "type": "text", + "content": "[61] X. Zhai, J. Puigcerver, A. Kolesnikov, P. Ruyssen, C. Riquelme, M. Lucic, J. Djolonga, A. S. Pinto, M. Neumann, A. Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 9" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 522, + 546, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 546, + 544 + ], + "type": "text", + "content": "[62] C. Zhang, S. Bengio, and Y. Singer. Are all layers created equal? arXiv preprint arXiv:1902.01996, 2019. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 548, + 546, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 546, + 571 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 546, + 571 + ], + "type": "text", + "content": "[63] Y. Zhang, K. Zhou, and Z. Liu. Neural prompt search. arXiv preprint arXiv:2206.04673, 2022. 1, 2, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 574, + 546, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 574, + 546, + 607 + ], + "spans": [ + { + "bbox": [ + 307, + 574, + 546, + 607 + ], + "type": "text", + "content": "[64] M. Zhao, T. Lin, F. Mi, M. Jaggi, and H. Schütze. Masking as an efficient alternative to finetuning for pretrained language models. In EMNLP, 2020. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 611, + 546, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 546, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 546, + 655 + ], + "type": "text", + "content": "[65] S. Zheng, J. Lu, H. Zhao, X. Zhu, Z. Luo, Y. Wang, Y. Fu, J. Feng, T. Xiang, P. H. Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 658, + 546, + 692 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 546, + 692 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 546, + 692 + ], + "type": "text", + "content": "[66] B. Zhou, H. Zhao, X. Puig, S. Fidler, A. Barriuso, and A. Torralba. Scene parsing through ade20k dataset. In CVPR, pages 633-641, 2017. 7" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11835" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Sentence Attention Blocks for Answer Grounding/0622d151-144e-4e04-b267-1cccf0dba507_content_list.json b/2023/Sentence Attention Blocks for Answer Grounding/0622d151-144e-4e04-b267-1cccf0dba507_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..49f9e55db361187c799f4aeb0cda62b06de2e5b2 --- /dev/null +++ b/2023/Sentence Attention Blocks for Answer Grounding/0622d151-144e-4e04-b267-1cccf0dba507_content_list.json @@ -0,0 +1,1764 @@ +[ + { + "type": "text", + "text": "Sentence Attention Blocks for Answer Grounding", + "text_level": 1, + "bbox": [ + 233, + 130, + 736, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seyedalireza Khoshsirat Chandra Kambhametty \nVideo/Image Modeling and Synthesis (VIMS) Lab, University of Delaware \n{alireza, chandrak}@udel.edu", + "bbox": [ + 187, + 181, + 781, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Answer grounding is the task of locating relevant visual evidence for the Visual Question Answering task. While a wide variety of attention methods have been introduced for this task, they suffer from the following three problems: designs that do not allow the usage of pre-trained networks and do not benefit from large data pre-training, custom designs that are not based on well-grounded previous designs, therefore limiting the learning power of the network, or complicated designs that make it challenging to re-implement or improve them. In this paper, we propose a novel architectural block, which we term Sentence Attention Block, to solve these problems. The proposed block re-calibrates channel-wise image feature-maps by explicitly modeling inter-dependencies between the image feature-maps and sentence embedding. We visually demonstrate how this block filters out irrelevant feature-maps channels based on sentence embedding. We start our design with a well-known attention method, and by making minor modifications, we improve the results to achieve state-of-the-art accuracy. The flexibility of our method makes it easy to use different pre-trained backbone networks, and its simplicity makes it easy to understand and be re-implemented. We demonstrate the effectiveness of our method on the TextVQA-X, VQS, VQA-X, and VizWiz-VQA-Grounding datasets. We perform multiple ablation studies to show the effectiveness of our design choices.", + "bbox": [ + 75, + 309, + 472, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 758, + 209, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1.1. Visual Question Answering", + "text_level": 1, + "bbox": [ + 76, + 792, + 323, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Question Answering (VQA) systems try to accurately answer natural language questions regarding an input image [1]. This topic aims at developing systems that can communicate effectively about an image in natural language and comprehend the contents of images similar to humans.", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1.2. Answer Grounding", + "text_level": 1, + "bbox": [ + 500, + 268, + 684, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The answer grounding task is defined as detecting the pixels that can provide evidence for the answer to a given question regarding an image [3]. In other words, the task is to return the image regions used to arrive at the answer for a given visual question (question-image pair) with an answer. Although the VQA community has made significant progress, the best-performing systems are complicated black-box models, raising concerns about whether their answer reasoning is based on correct visual evidence. By understanding the reasoning mechanism of the model, we can evaluate the quality of answers, improve model performance, and provide explanations for end-users. To address this problem, answer grounding has been introduced into VQA systems, which requires the model to locate relevant image regions as well as answer visual questions.", + "bbox": [ + 496, + 292, + 893, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "By providing answer groundings in response to visual questions, numerous applications become possible. First, they allow for evaluating whether a VQA model is reasoning correctly based on visual evidence. This is useful as an explanation as well as for assisting developers with model debugging. Second, answer grounding makes it possible to separate important regions from irrelevant background regions. Given that non-professional users can mistakenly have private information in the background of their pictures, answer grounding is a useful tool to obfuscate the background for privacy preserving. Third, suppose a service can magnify the relevant visual evidence. In that case, users will be able to discover the needed information in less time. This is useful in part because VQA answers can be insufficient sometimes.", + "bbox": [ + 496, + 518, + 893, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1.3. Multimodal Deep Learning", + "text_level": 1, + "bbox": [ + 500, + 755, + 746, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "By definition, the VQA and answer grounding tasks are multimodal tasks since a method for these tasks should be able to process and correlate two different modalities.", + "bbox": [ + 496, + 779, + 890, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal Joint-Embedding Models: These models merge and learn representations of multiple modalities in a joint feature space. Joint-embeddings underpin the building of a lot of cross-modal methods as they can bridge the gap between different modalities. In this joint space, the dis", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6080", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tance of different points is equivalent to the semantic distance between their corresponding original inputs.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal Attention-based Models: The main objective of the attention mechanism is to design systems that use local features of image or text, extract features from different regions, and assign priorities to them. The attention portion in an image determines salient regions. Then the language generation component focuses more on those salient regions for additional processing [35, 2, 12, 16, 26].", + "bbox": [ + 75, + 121, + 467, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.4. Attention Mechanism", + "text_level": 1, + "bbox": [ + 76, + 234, + 279, + 250 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In deep learning, attention is a mechanism that mimics cognitive attention. The goal is to enhance the essential features of the input data and vanish out the rest. Attention methods can be classified into two classes based on their inputs: self-attention and cross-attention. Self-attention (also known as intra-attention) is a type of attention that quantifies the interdependence within the elements of a single input. At the same time, cross-attention (also known as inter-attention) finds the interdependence across two or more inputs [40, 18, 17]. Usually, cross-attention methods are used for multimodal inputs [43]. Cross-attention models first process individual modalities using modality-specific encoders, then the encoded features are fed into cross-attention modules.", + "bbox": [ + 75, + 258, + 467, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Squeeze-and-Excitation method [13] is a channelwise self-attention mechanism widely used in classification networks [37, 38]. It consists of a global average pooling of the input, followed by two linear layers with an interleaved non-linearity and a sigmoid function. Concretely, the output of this method is:", + "bbox": [ + 75, + 469, + 467, + 558 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma \\left(F C (R E L U (F C (g _ {-} a v g _ {-} p o o l (\\mathbf {X})))\\right) \\times \\mathbf {X} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 568, + 467, + 584 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\mathbf{X}$ is the feature-maps. This module aims to dynamically focus on more important channels, essentially increasing the importance of specific channels over others. This is accomplished by scaling the more important channels by a higher value. While many feature descriptors exist to reduce the spatial dimensions of the feature maps to a singular value, this module uses average pooling to keep the required computation low. The next part of the module maps the scaling weights using a Multi-Layer Perceptron (MLP) with a bottleneck structure. The result values are scaled to a range of 0-1 by passing them through a sigmoid activation layer. Afterward, using a common broadcasted element-wise multiplication, the output is applied directly to the input. This multiplication scales each input feature-maps channel with its corresponding weight learned from the MLP in the Excitation module.", + "bbox": [ + 75, + 590, + 467, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The following is the summary of our contributions:", + "bbox": [ + 96, + 834, + 434, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We present a novel attention module based on the Squeeze-and-Eccitation method for the answer grounding task.", + "bbox": [ + 94, + 854, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We evaluate our method on common datasets and show that it achieves new state-of-the-art results.", + "- We compare our design with the top-performing networks.", + "- We perform multiple ablation studies to learn more about this network." + ], + "bbox": [ + 517, + 90, + 888, + 200 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 217, + 640, + 232 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Attention Methods for Answer Grounding", + "text_level": 1, + "bbox": [ + 500, + 242, + 859, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Models based on different attention methods have been actively explored for the answer grounding task.", + "bbox": [ + 500, + 266, + 888, + 295 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MAC-Caps [39] is based on MAC [15] which has a recurrent reasoning architecture that performs $T$ reasoning steps to answer the question. At each reasoning step, MAC uses an attention block to read from image features and writing memory. MAC-Caps adds capsule layers on top of the convolutional layers to obtain visual capsules from the feature-maps.", + "bbox": [ + 498, + 296, + 890, + 400 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Att-MCB [32] is designed to use a Knowledge Base. It is composed of nine modules, out of which two modules use an attention mechanism. One module uses an attention mechanism to generate visual explanations, and the other one to consume and point out relevant information from the Knowledge Base.", + "bbox": [ + 498, + 402, + 890, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A multi-grained attention method is introduced in [14]. It consists of two types of object-level groundings to explore fine-grained information, and a more sophisticated language model for better question representation. A Word-Label Matching attention vector that indicates the weight that should be given to each of the $K$ objects in the image is computed in terms of the semantic similarity between the category labels of the objects and the words in the question. A word-object matching module is exploited to evaluate how likely a question word matches a visual object. A Sentence-Object attention module is used to capture the global semantics of the whole sentence to guide the focus on relevant objects.", + "bbox": [ + 498, + 492, + 890, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multiple custom-designed attention modules are used in [46] to generate an attention map for the given image and question pair. Questions are tokenized and passed through an embedding layer followed by an LSTM layer to generate the question features. An Image Attention Supervision Module is used as an auxiliary classification task; that is, the ground-truth visual grounding labels are used to guide the model to focus on significant parts of the image to answer each question.", + "bbox": [ + 498, + 689, + 890, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A so-called \"accumulated attention\" mechanism is introduced in [5]. It consists of three attention modules, one for the input text, one for the objects, and one for the whole image. All three modules have a similar structure of a linear layer followed by a Softmax function.", + "bbox": [ + 498, + 825, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6081", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e00e29fb6741d093fa214e0c26ad6998080f825a199cb239361181c925f00739.jpg", + "image_caption": [ + "Figure 1. Our proposed sentence attention block. $C$ denotes the number of channels in image feature-maps, and $E$ represents the embedding size of the sentence encoder network. Question and answer embeddings are concatenated into one vector." + ], + "image_footnote": [], + "bbox": [ + 98, + 87, + 450, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Multimodal Attention Methods", + "text_level": 1, + "bbox": [ + 76, + 518, + 354, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since its introduction, the attention mechanism has been widely adopted in the computer vision community due to its special capabilities for many multimodal applications.", + "bbox": [ + 76, + 547, + 468, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In [11], a multimodal attention model is proposed based on the encoder-decoder networks and RNNs for video captioning and sentence generation. In particular, the multimodal attention framework combined image, audio, and motion features by selecting each modality's most relevant context vector.", + "bbox": [ + 75, + 597, + 468, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In [44], utilizing stacked attention networks is suggested to look for image regions that correlate with a query answer and pinpoint representative features of a given question more accurately.", + "bbox": [ + 75, + 693, + 468, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More recently, in [9], a normalized variant of the self-attention mechanism, named normalized self-attention (NSA), is introduced. NSA seeks to encode and decode the image and caption features and normalize the distribution of internal activations during training.", + "bbox": [ + 75, + 758, + 468, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the video question answering task, attention and memory mechanisms are used in [6] to efficiently learn visual features and the semantic correlations that precisely answer questions.", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Answer Grounding Methods", + "text_level": 1, + "bbox": [ + 500, + 90, + 756, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The answer grounding task has also been studied under different names, such as grounded VQA and visual explanation for VQA.", + "bbox": [ + 498, + 114, + 890, + 159 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The VinVL [45] method is based on OSCAR [21] which proposes a novel vision-language pre-training that incorporates image anchor points as input, taken from an object detection network. VinVL improves the visual representations by using a bigger visual model pre-trained on larger training corpora that combine multiple annotated object detection datasets.", + "bbox": [ + 498, + 161, + 892, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LXMERT [36] presents a cross-modality framework for learning the connections between vision and language. The authors build a large-scale Transformer model that consists of three encoders: an object relationship encoder, a language encoder, and a cross-modality encoder. This model is then pre-trained on a large-scale dataset of image-and-sentence pairs.", + "bbox": [ + 498, + 268, + 890, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MTXNet [30] is an end-to-end trainable architecture that generates multimodal explanations focusing on the text in the image. It contains a Graph Attention Network, and each object location and OCR token is treated as a node in the graph. It also contains a multimodal transformer that operates on three modalities: question words, visual objects, and OCR tokens.", + "bbox": [ + 498, + 375, + 890, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "U-CAM [29] is a method that uses gradient-based uncertainty estimates to provide visual attention maps. The certainty measure is improved by adding the certainty gradients to the existing standard Cross-Entropy loss gradients for training the model during back-propagation.", + "bbox": [ + 498, + 481, + 890, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 574, + 589, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Current attention methods have one or more of the following three problems:", + "bbox": [ + 500, + 599, + 890, + 631 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The attention mechanism is not based on well-known and well-tested methods [32, 36, 29]. Well-established methods have been evaluated on many different tasks and datasets, especially, they perform well on large-scale datasets. Their design details have been actively studied, and their advantages and disadvantages are visible to the users.", + "- The complicated design makes it challenging to re-implement them on other platforms [32, 39, 30]. Recently, many applications have been made based on deep neural networks that are expected to be used by the end-users of different platforms. For example, answer grounding methods are helpful in assistive technologies for people with vision impairments [3]. In order to use such technologies offline, it is crucial to implement the method for each platform." + ], + "bbox": [ + 517, + 643, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6082", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/48f385075ec7785d20d82c7fcab8e223b9a8fa1e5bc525fff6785427de4c33e7.jpg", + "image_caption": [ + "Figure 2. The complete proposed network. The multi-scale channel-wise image features contain different regions that are filtered by our Sentence Attention Block (SAB) based on the question and answer embeddings." + ], + "image_footnote": [], + "bbox": [ + 117, + 94, + 851, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- The custom design does not allow the usage of pretrained models [39, 45, 28, 27]. Answer grounding datasets are less diverse and have fewer samples than the classification datasets. Using pre-trained models can improve the accuracy and make the model more robust to noise and unseen samples.", + "bbox": [ + 96, + 563, + 468, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This paper proposes an attention block that avoids the mentioned problems. We design an attention block based on well-established and well-studied methods. It has a simple design to make it easy to re-implement and uses pre-trained models to achieve maximum accuracy.", + "bbox": [ + 75, + 666, + 468, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our proposed method has three main parts: region proposal, sentence embedding, and attention fusion.", + "bbox": [ + 75, + 743, + 468, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Region Proposal", + "text_level": 1, + "bbox": [ + 76, + 784, + 240, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This module's primary goal is to process the image and generate all the possible region candidates. Unlike Region Proposal Networks [31] which generate anchor-based bounding boxes, this module generates dense regions (segmentations).", + "bbox": [ + 75, + 809, + 468, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A pre-trained classification network is used as the back", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "bone. However, the final classifier is removed to get the multi-scale feature-maps. During the end-to-end training of the proposed method, the backbone network learns all the candidate regions. Later, these candidate regions are filtered and combined by the proposed Sentence Attention Blocks. In Section 5.3, we show that the backbone network works as expected by visualizing the feature-maps.", + "bbox": [ + 496, + 563, + 890, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Sentence Embedding", + "text_level": 1, + "bbox": [ + 500, + 678, + 697, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the input questions and answers have variable lengths, they are processed by a sentence embedding network. A pre-trained sentence encoder is employed to encode the given question and answer to a vector space separately. The result is two vectors, one for the question and one for the answer. The two vectors are concatenated and passed to the Sentence Attention Blocks as one vector.", + "bbox": [ + 496, + 700, + 890, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Attention Fusion", + "text_level": 1, + "bbox": [ + 500, + 816, + 666, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given multiple potential regions for an image, the answer grounding task reduces to choosing and/or combining these regions based on the given question and/or answer. The multi-scale channel-wise image features from the back", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6083", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/ad0210b1ec08437248686fadf751d380bf1df21b82344d9d684e0bb320428ee5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMean IoU
MTXNet [30]18.9
LXMERT [36]24.9
VinVL [45]25.6
MAC-Caps [39]26.1
Ours29.0 (+2.9)
", + "bbox": [ + 117, + 88, + 431, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "bone network contain candidate regions. The goal is to filter and combine these regions based on the question and answer vector from the sentence embedding network. Towards this goal, we design an attention block based on the Squeeze-and-Excitation (SE) block [13].", + "bbox": [ + 76, + 243, + 467, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We start with an SE block and make four modifications as follows:", + "bbox": [ + 76, + 319, + 467, + 349 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. The first modification is changing it from self-attention to cross-attention. In other words, unlike a self-attention block, the proposed block pays attention to an external source: the encoded question and answer.", + "2. The second modification is adding a normalization layer. Empirically we found that adding a normalization layer improves the accuracy. Similar to [40], we use LayerNorm since it performs better on sequences.", + "3. The third modification is replacing the sigmoid function with softmax. Since the goal is to force the model to choose a few of the channel-wise image features, the sigmoid function is replaced with softmax.", + "4. The last modification is inverting the bottleneck. Instead of shrinking the embedding space, we expand the question and answer embeddings by a factor of two." + ], + "bbox": [ + 88, + 358, + 467, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Section 5.1, the effect of each modification on the network accuracy is shown. Figure 1 depicts the final structure of the proposed block. Precisely, the proposed attention block is as follows:", + "bbox": [ + 76, + 622, + 467, + 680 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S o f t m a x} (F C (\\operatorname {R e L U} (L N (F C (\\mathbf {Q A})))) \\times \\mathbf {R} _ {i} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 691, + 467, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $FC$ is a single-layer perceptron, $LN$ is LayerNorm, QA is the joint question and answer embeddings, and $\\mathbf{R}_i$ is the image regions at scale $i$ . For each scale, a separate instance of this attention block is created so that each block learns special weights for its corresponding feature-map scale.", + "bbox": [ + 76, + 719, + 467, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After computing the attention block for all the scales, all the outputs at different scales are combined into one output. Starting from the lowest resolution output, a $1 \\times 1$ convolution is used to convert the number of channels to the number of channels of the next lowest resolution output. Then, Bilinear interpolation is used for upsampling the", + "bbox": [ + 75, + 810, + 467, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a4e141adbd6587889c3cc516aec2ec8afdc3a315d10b760fed85e273c2fad022.jpg", + "table_caption": [ + "Table 1. Comparison results on the test set of the TextVQA-X dataset [30]." + ], + "table_footnote": [], + "table_body": "
MethodMean IoU
DeconvNet [7]29.8
Mask Aggregation [7]32.6
LXMERT [36]33.3
VinVL [45]33.9
MAC-Caps [39]34.3
Ours36.8 (+2.5)
", + "bbox": [ + 519, + 88, + 870, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Comparison results on the test set of the VQS dataset [7].", + "bbox": [ + 500, + 207, + 888, + 222 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "output feature-map by a factor of 2. Finally, the output is added to the next lowest resolution output. This process is repeated until no other output is left. In the end, a $1 \\times 1$ convolution is used to reduce the number of channels of the final output to two. Therefore, the final output has two channels representing answer grounding and its background. We use Argmax to output a binary image.", + "bbox": [ + 498, + 247, + 890, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The complete design of the proposed method is depicted in Figure 2.", + "bbox": [ + 498, + 353, + 890, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Comparison to Existing Methods", + "text_level": 1, + "bbox": [ + 500, + 393, + 787, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare the design of our proposed method to the top-performing existing methods for the answer grounding task.", + "bbox": [ + 498, + 417, + 890, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MAC-Caps [39] is one of the best-performing networks that currently exist. This network is based on MAC [15] which employs attention blocks in a recurrent reasoning architecture. MAC-Caps adds capsule layers to process the feature-maps further. Although similar to our method, MAC-Caps employs an attention mechanism. However, the recurrent design brings the usual problems of RNNs, such as gradient vanishing/exploding, longer training time, and not performing well on long sequences. Since our proposed method does not have a recurrent design, it avoids such problems.", + "bbox": [ + 498, + 462, + 890, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VinVL [45] is another top-performing network for the answer grounding task. This method is based on OSCAR [21] which incorporates image anchor points as inputs. VinVL improves the visual representations by using a bigger visual model pre-trained on larger training corpora that combine multiple annotated object detection datasets. This method relies on object detection networks to provide the anchor points, which adds an extra task for the network to learn. Also, using large generic corpora may not improve the accuracy for special datasets such as VizWiz-VQA-Grounding [3]. In contrast, our proposed method relies only on the feature-maps and does not define another task.", + "bbox": [ + 498, + 628, + 890, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Att-MCB [32] is the state-of-the-art network on the VQA-X dataset [27]. This method is designed to use a Knowledge Base. It comprises nine modules and four loss functions, making it a tedious design. This complicated design is resource-demanding and makes the training sensitive", + "bbox": [ + 498, + 825, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6084", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2250720e73dcb5c1a70a567c34b8d9b90d8a1959b2a8d124053acdab1d7cf28b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRank Correlation
PJ-X [27]0.342
CCM [28]0.368
U-CAM [29]0.372
VinVL [45]0.373
Att-MFH [32]0.376
MAC-Caps [39]0.389
Att-MCB [32]0.396
Ours0.421 (+0.025)
", + "bbox": [ + 112, + 88, + 433, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to hyper-parameters. Our proposed method comprises three main modules and one loss function. This minimalistic design helps the gradient flow during back-propagation, and the training is less sensitive to hyper-parameters.", + "bbox": [ + 76, + 289, + 467, + 351 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 363, + 209, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train and test our proposed method on four answer grounding datasets. We use the same setup for all of the experiments.", + "bbox": [ + 76, + 388, + 468, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 76, + 443, + 158, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use a pre-trained EfficientNet [37] to extract the multi-scale image features and we use a MiniLMv2 [42] sentence encoder network to encode the questions. In Section 5.2, we perform an ablation study on different backbones. we use BEiT [41] to generate answers for the datasets where the answers are not publicly available. We use RMI [47] with its default settings along with cross-entropy. Therefore, the overall loss function to minimize is:", + "bbox": [ + 75, + 465, + 468, + 602 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {a l l} (y, p) = \\frac {1}{B} \\sum_ {b = 1} ^ {B} \\lambda \\mathcal {L} _ {c e} \\left(y ^ {(b)}, p ^ {(b)}\\right) \\tag {3} \\\\ + (1 - \\lambda) \\mathcal {L} _ {r m i} (y ^ {(b)}, p ^ {(b)}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 609, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda \\in [0,1]$ is a weight factor, $B$ denotes the number of samples in a mini-batch, $\\mathcal{L}_{ce}(y^{(b)},p^{(b)})$ is the standard cross entropy loss between the $b$ -th sample and its corresponding prediction, respectively, and $\\mathcal{L}_{rmi}(y^{(b)},p^{(b)})$ is the RMI loss as follows:", + "bbox": [ + 76, + 681, + 468, + 758 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r m i} (\\mathbf {Y}, \\mathbf {P}) = \\sum_ {c = 1} ^ {C} \\frac {1}{2 d} \\operatorname {T r a c e} (\\log (\\mathbf {M})) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 766, + 468, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and,", + "bbox": [ + 76, + 816, + 109, + 830 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} = \\boldsymbol {\\Sigma} _ {\\mathbf {Y}} - C o v (\\mathbf {Y}, \\mathbf {P}) \\left(\\boldsymbol {\\Sigma} _ {\\mathbf {P}} ^ {- 1}\\right) ^ {T} C o v (\\mathbf {Y}, \\mathbf {P}) ^ {T} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 840, + 468, + 859 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{Y} = y^{(b)}$ , $\\mathbf{P} = p^{(b)}$ , $\\mathbf{M} \\in \\mathbb{R}^{d \\times d}$ , $d$ denotes the number of pixels, $C$ is the number of object classes, here", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e6d2db0b111815b5d28472fbd53120a991a50b0a886100574e90e771cec73cdc.jpg", + "table_caption": [ + "Table 3. Comparison results on the test set of the VQA-X dataset [27]." + ], + "table_footnote": [], + "table_body": "
TeamMean IoU
MindX66.9
MGTV69.7
hsslab_inspur70.1
GroundTruth70.3
Aurora70.6
SAB (Ours)72.4 (+1.8)
", + "bbox": [ + 565, + 87, + 826, + 198 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/412c2bc294ca3bb77ae20c00c6cc08a290d7e5772cabb44f197ae2a7fa9b0a53.jpg", + "table_caption": [ + "Table 4. The top-performing teams on the VizWiz-VQA-Grounding Challenge 2022 leaderboard [3]. At the time of this writing, none of the methods have been published." + ], + "table_footnote": [], + "table_body": "
Method# of paramsinference time
LXMERT [36]153M56ms
VinVL [45]406M69ms
MAC-Caps [39]80M54ms
Ours88M49ms
", + "bbox": [ + 506, + 263, + 885, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5. Comparison of our proposed method to existing methods regarding the number of parameters and inference time.", + "bbox": [ + 498, + 352, + 890, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$C = 2$ for the answer grounding and its background, $\\Sigma_{\\mathbf{Y}}$ is the variance matrix of $\\mathbf{Y}$ , and $Cov(\\mathbf{Y},\\mathbf{P})$ is the covariance matrix of $\\mathbf{Y}$ and $\\mathbf{P}$ . Similar to [47], we set $\\lambda = 0.5$ .", + "bbox": [ + 498, + 406, + 890, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use AdamW optimizer [25] with a weight decay of 0.05 and batch size of 16. We apply the \"polynomial\" learning rate policy with a poly exponent of 0.9 and an initial learning rate of 0.0001. Synchronized batch normalization is used across multiple GPUs. We use RandAugment [4] for data augmentation.", + "bbox": [ + 496, + 452, + 892, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. TextVQA-X", + "text_level": 1, + "bbox": [ + 500, + 554, + 629, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The TextVQA-X dataset [30] is a subset of the TextVQA dataset [33] such that the answer groundings are generated through a manual segmentation annotation process. TextVQA-X consists of 10,379 training images and 3,354 test images. Since only a few models have been evaluated on this dataset, we train three more models for a fair comparison. Similar to [3], we train top-performing methods that their code is publicly available; specifically, LXMERT [36], VinVL [45], and MAC-Caps [39]. We use the same setup as in [3] to train the three models with the difference that we train the models on the TextVQA-X training set. Similar to the existing methods, we use the test set to evaluate our method and report the mean IoU measure.", + "bbox": [ + 496, + 578, + 890, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.VQS", + "text_level": 1, + "bbox": [ + 500, + 785, + 573, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The VQS dataset [7] builds upon the images, instance segmentation masks, and bounding boxes in COCO [22] and the questions and answers in the VQA dataset [1]. VQS contains 26,995 training, 5,000 validation, and 5,873 test images. Similar to Section 4.2, we train top-performing methods that their code is publicly available; specifically,", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6085", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a16446120986a7091ff6d1e8d801ef17e1a6704d47a2eadcaf1f70b4b28af282.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6. The ablation study of the pre-trained backbone networks. Top Section: Using a MiniLMv2 as a fixed sentence embedding network and replacing the image feature extraction networks. Bottom Section: Using a fixed image feature extraction network (EfficientNet-B7) and replacing the sentence embedding networks." + ], + "table_body": "
BackboneMean IoU
ResNet-50x3 [19]71.9
Swin Transformer V2-B [24]72.8
EfficientNetV2-L [38]73.1
EfficientNet-B7 [37]73.5
RoBERTa [23]72.8
MPNet [34]73.1
MiniLMv2 [42]73.5
", + "bbox": [ + 109, + 88, + 436, + 213 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LXMERT [36], VinVL [45], and MAC-Caps [39]. We use the same setup as in [3] to train the three models with the difference that we train the models on the VQS training set. Similar to the existing methods, we use the test set to evaluate our method and report the mean IoU measure.", + "bbox": [ + 75, + 333, + 468, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.VQA-X", + "text_level": 1, + "bbox": [ + 76, + 425, + 171, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The VQA-X dataset [27] does not have visual annotations for its training set. It has visual annotations only for its validation and test sets. Therefore, we train our method on the training set of the VQS dataset [7], which has related images to the VQA-X validation and test sets. The images of both datasets are from the COCO dataset [22]. The other methods are trained on different datasets, including Visual Genome [20] and VQA v2.0 [8]. The VQA-X dataset consists of 24,876 training, 1,431 validation, and 1,921 test images. Similar to the other methods, we use the test set of the VQA-X dataset to evaluate our method and report the Rank Correlation as in [27].", + "bbox": [ + 75, + 450, + 468, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. VizWiz-VQA-Grounding", + "text_level": 1, + "bbox": [ + 76, + 648, + 305, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The VizWiz-VQA-Grounding dataset [3] is based on the VizWiz-VQA dataset [10], and the images and questions come from visually impaired people who shared them to request visual assistance in their day-to-day lives. VizWiz-VQA-Grounding contains a total of 9,998 VQAs divided into 6,494/1,131/2,373 VQAs for training, validation, and testing. The 2022 VizWiz-VQA-Grounding Challenge is designed around the aforementioned VizWiz-VQA-Grounding dataset. This challenge was held recently, and at the time of this writing, current methods in this challenge have not been published yet. Therefore, we evaluate our method using the VizWiz-VQA-Grounding Challenge online leaderboard [3]. The leaderboard evaluates all the methods against the VizWiz-VQA-Grounding test set. The ground truth labels for the test set are not published yet.", + "bbox": [ + 75, + 672, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/24ac30a491de6b54516a3cce41b18b2aa42a290bc7aecfbfaa3c93fd3baec4b2.jpg", + "table_caption": [], + "table_footnote": [ + "Table 7. The results of our ablation study of the design modifications. The modifications are done cumulatively from top to bottom." + ], + "table_body": "
ModificationMean IoU
Switching to cross-attention70.1
Adding normalization layer71.2 (+1.1)
Switching to Softmax72.7 (+1.5)
Expansion of the embeddings73.5 (+0.8)
", + "bbox": [ + 527, + 88, + 864, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Results", + "text_level": 1, + "bbox": [ + 500, + 247, + 591, + 261 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our proposed method on four answer grounding datasets. Tables 1, 2, and 3 compare the results of our proposed method with the existing methods on the TextVQA-X, VQS, and VQA-X datasets respectively. Our method achieves new state-of-the-art accuracy on three datasets. More specifically, improvements of 2.9, 2.5, and 0.025 on the TextVQA-X, VQS, and VQA-X datasets, respectively.", + "bbox": [ + 496, + 273, + 890, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 lists the top-performing teams on the VizWiz-VQA-Grounding Challenge 2022 leaderboard [3]. At the time of this writing, our method holds first place with considerably higher accuracy than the other methods.", + "bbox": [ + 496, + 396, + 890, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, Table 5 compares our proposed method to existing methods regarding the number of parameters and inference time. The efficient design of our method makes it the fastest method.", + "bbox": [ + 496, + 459, + 890, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 542, + 663, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we perform four ablation studies. All of the experiments are done on the VizWiz-VQA-Grounding [3] validation set and using the setup from Section 4.1 unless otherwise stated.", + "bbox": [ + 496, + 569, + 890, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Design Modifications", + "text_level": 1, + "bbox": [ + 500, + 647, + 696, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We started the design of our sentence attention block from the standard SE block and made four modifications as explained in Section 3.3. In this section, we perform an ablation study on these modifications. Table 7 shows the impact of each modification on the final accuracy. These modifications are done cumulatively. The first modification is changing from self-attention to cross-attention, which makes our baseline for this ablation study. The second modification is adding a normalization layer. This modification improves our baseline by 1.1 percent. The third modification is replacing the Sigmoid function with Softmax. This modification has the highest impact on our method, 1.5 percent. And the fourth modification is the expansion of the embeddings. This last modification brings an addition of 0.8 percent.", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6086", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Backbone Networks", + "text_level": 1, + "bbox": [ + 76, + 90, + 269, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The flexibility of our method enables us to use a variety of pre-trained networks. In this ablation study, we compare and analyze different pre-trained networks. To compare the image backbone networks, we keep the sentence embedding network fixed and change the image backbone network. Similarly, to compare the sentence embedding networks, we keep the image backbone network with the highest accuracy and change the sentence embedding network. We use well-known and well-performing networks whose pre-trained weights are publicly available. Specifically, we use EfficientNet [37], EfficientNetV2 [38], Swin Transformer V2 [24], and ResNet-50x3 [19] for image feature extraction and MiniLMv2 [42], MPNet [34], and RoBERTa [23] for sentence embedding. Table 6 shows the results of this ablation study. The highest accuracy is achieved by using EfficientNet-B7 and MiniLMv2. Since the EfficientNetV2-L reduces the input resolution more aggressively than EfficientNet-B7, it performs worse. Transformer networks usually perform well on large datasets. However, since VizWiz-VQA-Grounding is not a large dataset, Swin Transformer V2 does not achieve the highest accuracy. The ResNet-50x3 is pre-trained on ImageNet21k, but it achieves the least accuracy in this experiment.", + "bbox": [ + 75, + 113, + 473, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Region Proposals", + "text_level": 1, + "bbox": [ + 76, + 468, + 246, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feature-map visualization will provide insight into the internal representations of each layer in a network for a specific input image.", + "bbox": [ + 75, + 492, + 468, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This ablation study aims to show that the image backbone network has learned the potential regions for an image, which are then processed by the proposed sentence attention block. To this aim, we visualize the final output feature-maps of the image backbone network, before feeding them to an attention block. We use the image backbone of our fully trained network on the VizWiz dataset. Figure 3 (a) shows a few manually-chosen feature-maps for a sample image. These feature-maps show that after training, the backbone network has been able to learn the potential important regions of images. Having all the candidate regions, the task of the proposed attention block is to filter the irrelevant regions to the question and answer. Figure 3 (b) shows the final output of the complete network for the same image in (a), using different questions and answers (for brevity, the answers are not shown). This figure shows that our proposed block is able to filter the channel-wise feature-maps based on the input questions and answers.", + "bbox": [ + 75, + 537, + 470, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Multi-scale Fusions", + "text_level": 1, + "bbox": [ + 76, + 816, + 264, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "There is a trade-off in the answer grounding task such that some images are best handled at lower inference resolution and others better handled at higher inference resolution. Fine details, such as the edges of objects or thin", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0109e65c52ea08f2638a00f7618f631593e87b414e618aa912a13861ff1e46f3.jpg", + "image_caption": [ + "Figure 3. (a) Eight manually-selected sample feature-maps for an image. These feature-maps are the output of the feature extraction backbone network, before feeding them to the sentence attention block. The network has learned to output potential regions in different channels. (b) The final output of our complete network for the same image in (a), given different questions and answers (for brevity, the answers are not shown). This figure shows that our proposed attention block can filter the channel-wise feature-maps based on the input questions and answers." + ], + "image_footnote": [], + "bbox": [ + 519, + 88, + 869, + 473 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6a6c76095e9032e0825fd1d24b15b7ff12ace196b9ac009e70198670e02ec290.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ScaleMean IoU
1/3271.4
1/32 + 1/1672.5 (+1.1)
1/32 + 1/16 + 1/873.1 (+0.6)
1/32 + 1/16 + 1/8 + 1/473.5 (+0.4)
", + "bbox": [ + 532, + 625, + 859, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. The gradual effect of multi-scale fusion. From top to bottom, the first row uses only the lowest resolution feature-map, and each row adds higher resolution feature-maps. Each scale is the ratio of the feature-map size to the input image.", + "bbox": [ + 496, + 713, + 893, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "structures, are often better predicted with scaled-up images. At the same time, the prediction of large structures, which requires more global context, is often done better at scaled-down images because the network's receptive field can observe more of the necessary context.", + "bbox": [ + 496, + 794, + 890, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This ablation study shows the impact of adding the extra scales on the accuracy. Table 8 shows the results of this", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6087", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "study. From top row to bottom, we start with only one scale and one attention block and gradually add higher resolution feature-maps. The lowest resolution is the final output feature-map of the backbone network.", + "bbox": [ + 76, + 90, + 472, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 76, + 166, + 204, + 181 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the study, a new architectural block termed the Sentence Attention Block was introduced to address certain challenges. This block recalibrates channel-wise image feature-maps by modeling inter-dependencies between image feature-maps and sentence embedding. The design began with a recognized attention method, and with minor adjustments, enhanced results were achieved, reaching top-tier accuracy. We showed the block's ability to filter out irrelevant feature-map channels based on sentence embedding. The method's adaptability allows for various pre-trained backbone networks, and its straightforward nature facilitates comprehension and re-implementation. The method's efficacy was demonstrated on the TextVQA-X, VQS, VQA-X, and VizWiz-VQA-Grounding datasets, and several ablation studies highlighted the validity of the design decisions.", + "bbox": [ + 76, + 191, + 472, + 434 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 448, + 174, + 464 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015.", + "[2] Khaled Bayoudh, Raja Knani, Fayçal Hamdaoui, and Abdel-latif Mtibaa. A survey on deep multimodal learning for computer vision: advances, trends, applications, and datasets. The Visual Computer, 38(8):2939-2970, 2022.", + "[3] Chongyan Chen, Samreen Anjum, and Danna Gurari. Grounding answers for visual questions asked by visually impaired people. arXiv preprint arXiv:2202.01993, 2022.", + "[4] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020.", + "[5] Chaorui Deng, Qi Wu, Qingyao Wu, Fuyuan Hu, Fan Lyu, and Mingkui Tan. Visual grounding via accumulated attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7746-7755, 2018.", + "[6] Chenyou Fan, Xiaofan Zhang, Shu Zhang, Wensheng Wang, Chi Zhang, and Heng Huang. Heterogeneous memory enhanced multimodal attention model for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1999-2007, 2019.", + "[7] Chuang Gan, Yandong Li, Haoxiang Li, Chen Sun, and Boqing Gong. Vqs: Linking segmentations to questions and answers for supervised attention in vqa and question-focused" + ], + "bbox": [ + 86, + 474, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "semantic segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1811-1820, 2017.", + "[8] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017.", + "[9] Longteng Guo, Jing Liu, Xinxin Zhu, Peng Yao, Shichen Lu, and Hanqing Lu. Normalized and geometry-aware self-attention network for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10327–10336, 2020.", + "[10] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3608-3617, 2018.", + "[11] Chiori Hori, Takaaki Hori, Teng-Yok Lee, Ziming Zhang, Bret Harsham, John R Hershey, Tim K Marks, and Kazuhiko Sumi. Attention-based multimodal fusion for video description. In Proceedings of the IEEE international conference on computer vision, pages 4193-4202, 2017.", + "[12] Parisa Hosseini, Seyedalireza Khoshkirat, Mohammad Jalayer, Subasish Das, and Huaguo Zhou. Application of text mining techniques to identify actual wrong-way driving (wwd) crashes in police reports. International Journal of Transportation Science and Technology, 2022.", + "[13] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018.", + "[14] Pingping Huang, Jianhui Huang, Yuqing Guo, Min Qiao, and Yong Zhu. Multi-grained attention with object-level grounding for visual question answering. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3595-3600, 2019.", + "[15] Drew A Hudson and Christopher D Manning. Compositional attention networks for machine reasoning. arXiv preprint arXiv:1803.03067, 2018.", + "[16] Seyedalireza Khoshsirat and Chandra Kambhamtu. Semantic segmentation using neural ordinary differential equations. Advances in Visual Computing: 17th International Symposium, ISVC 2022, San Diego, CA, USA, October 3-5, 2022, Proceedings, Part I, pages 284-295, 2022.", + "[17] Seyedalireza Khoshsurat and Chandra Kambhametu. Embedding attention blocks for the vizwiz answer grounding challenge. VizWiz Grand Challenge Workshop, 2023.", + "[18] Seyedalireza Khoshsirat and Chandra Kambhamtu. Empowering visually impaired individuals: A novel use of apple live photos and android motion photos. 25th Irish Machine Vision and Image Processing Conference, 2023.", + "[19] Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, and Neil Houlsby. Big transfer (bit): General visual representation learning. In European conference on computer vision, pages 491-507. Springer, 2020." + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6088", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[20] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017.", + "[21] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020.", + "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014.", + "[23] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.", + "[24] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. arXiv preprint arXiv:2111.09883, 2021.", + "[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017.", + "[26] Elham Maserat, Reza Safdari, Hamid Asadzadeh Aghdaei, Alireza Khoshsirat, and Mohammad Reza Zali. 43: Designing evidence based risk assessment system for cancer screening as an applicable approach for the estimating of treatment roadmap. BMJ Open, 7(Suppl 1):bmjopen-2016, 2017.", + "[27] Dong Huk Park, Lisa Anne Hendricks, Zeynep Akata, Anna Rohrbach, Bernt Schiele, Trevor Darrell, and Marcus Rohrbach. Multimodal explanations: Justifying decisions and pointing to the evidence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8779-8788, 2018.", + "[28] Badri Patro, Shivansh Patel, and Vinay Namboodiri. Robust explanations for visual question answering. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1577-1586, 2020.", + "[29] Badri N Patro, Mayank Lunayach, Shivansh Patel, and Vinay P Namboodiri. U-cam: Visual explanation using uncertainty based class activation maps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7444-7453, 2019.", + "[30] Varun Nagaraj Rao, Xingjian Zhen, Karen Hovsepian, and Mingwei Shen. A first look: Towards explainable textvqa models via visual and textual explanations. arXiv preprint arXiv:2105.02626, 2021.", + "[31] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015.", + "[32] Felipe Riquelme, Alfredo De Goyeneche, Yundong Zhang, Juan Carlos Niebles, and Alvaro Soto. Explaining vqa pre" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "dictions using visual grounding and a knowledge base. Image and Vision Computing, 101:103968, 2020.", + "[33] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019.", + "[34] Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. Mpnet: Masked and permuted pre-training for language understanding. Advances in Neural Information Processing Systems, 33:16857-16867, 2020.", + "[35] Jabeen Summaira, Xi Li, Amin Muhammad Shoib, Songyuan Li, and Jabbar Abdul. Recent advances and trends in multimodal deep learning: A review. arXiv preprint arXiv:2105.11087, 2021.", + "[36] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019.", + "[37] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114. PMLR, 2019.", + "[38] Mingxing Tan and Quoc V Le. Efficientnetv2: Smaller models and faster training. arXiv preprint arXiv:2104.00298, 2021.", + "[39] Aisha Urooj, Hilde Kuehne, Kevin Duarte, Chuang Gan, Niels Lobo, and Mubarak Shah. Found a reason for me? weakly-supervised grounded visual question answering using capsules. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8465-8474, 2021.", + "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017.", + "[41] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022.", + "[42] Wenhui Wang, Hangbo Bao, Shaohan Huang, Li Dong, and Furu Wei. Minilmv2: Multi-head self-attention relation distillation for compressing pretrained transformers. arXiv preprint arXiv:2012.15828, 2020.", + "[43] Xi Wei, Tianzhu Zhang, Yan Li, Yongdong Zhang, and Feng Wu. Multi-modality cross attention network for image and sentence matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10941-10950, 2020.", + "[44] Zichao Yang, Xiaodong He, Jianfeng Gao, Li Deng, and Alex Smola. Stacked attention networks for image question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 21-29, 2016.", + "[45] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6089", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computer Vision and Pattern Recognition, pages 5579-5588, 2021.", + "[46] Yundong Zhang, Juan Carlos Niebles, and Alvaro Soto. Interpretable visual question answering by visual grounding from attention supervision mining. In 2019 IEEE winter conference on applications of computer vision (wacv), pages 349-357. IEEE, 2019.", + "[47] Shuai Zhao, Yang Wang, Zheng Yang, and Deng Cai. Region mutual information loss for semantic segmentation. Advances in Neural Information Processing Systems, 32, 2019." + ], + "bbox": [ + 78, + 90, + 468, + 233 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6090", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Sentence Attention Blocks for Answer Grounding/0622d151-144e-4e04-b267-1cccf0dba507_model.json b/2023/Sentence Attention Blocks for Answer Grounding/0622d151-144e-4e04-b267-1cccf0dba507_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6e3f0ca325213563882cef48cee409ae63acbe63 --- /dev/null +++ b/2023/Sentence Attention Blocks for Answer Grounding/0622d151-144e-4e04-b267-1cccf0dba507_model.json @@ -0,0 +1,2367 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.131, + 0.738, + 0.154 + ], + "angle": 0, + "content": "Sentence Attention Blocks for Answer Grounding" + }, + { + "type": "text", + "bbox": [ + 0.188, + 0.182, + 0.782, + 0.234 + ], + "angle": 0, + "content": "Seyedalireza Khoshsirat Chandra Kambhametty \nVideo/Image Modeling and Synthesis (VIMS) Lab, University of Delaware \n{alireza, chandrak}@udel.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.473, + 0.705 + ], + "angle": 0, + "content": "Answer grounding is the task of locating relevant visual evidence for the Visual Question Answering task. While a wide variety of attention methods have been introduced for this task, they suffer from the following three problems: designs that do not allow the usage of pre-trained networks and do not benefit from large data pre-training, custom designs that are not based on well-grounded previous designs, therefore limiting the learning power of the network, or complicated designs that make it challenging to re-implement or improve them. In this paper, we propose a novel architectural block, which we term Sentence Attention Block, to solve these problems. The proposed block re-calibrates channel-wise image feature-maps by explicitly modeling inter-dependencies between the image feature-maps and sentence embedding. We visually demonstrate how this block filters out irrelevant feature-maps channels based on sentence embedding. We start our design with a well-known attention method, and by making minor modifications, we improve the results to achieve state-of-the-art accuracy. The flexibility of our method makes it easy to use different pre-trained backbone networks, and its simplicity makes it easy to understand and be re-implemented. We demonstrate the effectiveness of our method on the TextVQA-X, VQS, VQA-X, and VizWiz-VQA-Grounding datasets. We perform multiple ablation studies to show the effectiveness of our design choices." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.759, + 0.21, + 0.775 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.793, + 0.325, + 0.81 + ], + "angle": 0, + "content": "1.1. Visual Question Answering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Visual Question Answering (VQA) systems try to accurately answer natural language questions regarding an input image [1]. This topic aims at developing systems that can communicate effectively about an image in natural language and comprehend the contents of images similar to humans." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.269, + 0.686, + 0.286 + ], + "angle": 0, + "content": "1.2. Answer Grounding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.294, + 0.895, + 0.52 + ], + "angle": 0, + "content": "The answer grounding task is defined as detecting the pixels that can provide evidence for the answer to a given question regarding an image [3]. In other words, the task is to return the image regions used to arrive at the answer for a given visual question (question-image pair) with an answer. Although the VQA community has made significant progress, the best-performing systems are complicated black-box models, raising concerns about whether their answer reasoning is based on correct visual evidence. By understanding the reasoning mechanism of the model, we can evaluate the quality of answers, improve model performance, and provide explanations for end-users. To address this problem, answer grounding has been introduced into VQA systems, which requires the model to locate relevant image regions as well as answer visual questions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.895, + 0.746 + ], + "angle": 0, + "content": "By providing answer groundings in response to visual questions, numerous applications become possible. First, they allow for evaluating whether a VQA model is reasoning correctly based on visual evidence. This is useful as an explanation as well as for assisting developers with model debugging. Second, answer grounding makes it possible to separate important regions from irrelevant background regions. Given that non-professional users can mistakenly have private information in the background of their pictures, answer grounding is a useful tool to obfuscate the background for privacy preserving. Third, suppose a service can magnify the relevant visual evidence. In that case, users will be able to discover the needed information in less time. This is useful in part because VQA answers can be insufficient sometimes." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.756, + 0.747, + 0.773 + ], + "angle": 0, + "content": "1.3. Multimodal Deep Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.825 + ], + "angle": 0, + "content": "By definition, the VQA and answer grounding tasks are multimodal tasks since a method for these tasks should be able to process and correlate two different modalities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Multimodal Joint-Embedding Models: These models merge and learn representations of multiple modalities in a joint feature space. Joint-embeddings underpin the building of a lot of cross-modal methods as they can bridge the gap between different modalities. In this joint space, the dis" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6080" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "tance of different points is equivalent to the semantic distance between their corresponding original inputs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.228 + ], + "angle": 0, + "content": "Multimodal Attention-based Models: The main objective of the attention mechanism is to design systems that use local features of image or text, extract features from different regions, and assign priorities to them. The attention portion in an image determines salient regions. Then the language generation component focuses more on those salient regions for additional processing [35, 2, 12, 16, 26]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.236, + 0.281, + 0.25 + ], + "angle": 0, + "content": "1.4. Attention Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.468, + 0.468 + ], + "angle": 0, + "content": "In deep learning, attention is a mechanism that mimics cognitive attention. The goal is to enhance the essential features of the input data and vanish out the rest. Attention methods can be classified into two classes based on their inputs: self-attention and cross-attention. Self-attention (also known as intra-attention) is a type of attention that quantifies the interdependence within the elements of a single input. At the same time, cross-attention (also known as inter-attention) finds the interdependence across two or more inputs [40, 18, 17]. Usually, cross-attention methods are used for multimodal inputs [43]. Cross-attention models first process individual modalities using modality-specific encoders, then the encoded features are fed into cross-attention modules." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.468, + 0.559 + ], + "angle": 0, + "content": "The Squeeze-and-Excitation method [13] is a channelwise self-attention mechanism widely used in classification networks [37, 38]. It consists of a global average pooling of the input, followed by two linear layers with an interleaved non-linearity and a sigmoid function. Concretely, the output of this method is:" + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.569, + 0.468, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\sigma \\left(F C (R E L U (F C (g _ {-} a v g _ {-} p o o l (\\mathbf {X})))\\right) \\times \\mathbf {X} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.468, + 0.832 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}\\) is the feature-maps. This module aims to dynamically focus on more important channels, essentially increasing the importance of specific channels over others. This is accomplished by scaling the more important channels by a higher value. While many feature descriptors exist to reduce the spatial dimensions of the feature maps to a singular value, this module uses average pooling to keep the required computation low. The next part of the module maps the scaling weights using a Multi-Layer Perceptron (MLP) with a bottleneck structure. The result values are scaled to a range of 0-1 by passing them through a sigmoid activation layer. Afterward, using a common broadcasted element-wise multiplication, the output is applied directly to the input. This multiplication scales each input feature-maps channel with its corresponding weight learned from the MLP in the Excitation module." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.835, + 0.435, + 0.849 + ], + "angle": 0, + "content": "The following is the summary of our contributions:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- We present a novel attention module based on the Squeeze-and-Eccitation method for the answer grounding task." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.889, + 0.12 + ], + "angle": 0, + "content": "- We evaluate our method on common datasets and show that it achieves new state-of-the-art results." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.133, + 0.89, + 0.16 + ], + "angle": 0, + "content": "- We compare our design with the top-performing networks." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.174, + 0.89, + 0.202 + ], + "angle": 0, + "content": "- We perform multiple ablation studies to learn more about this network." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.092, + 0.89, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.218, + 0.642, + 0.233 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.243, + 0.86, + 0.259 + ], + "angle": 0, + "content": "2.1. Attention Methods for Answer Grounding" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.267, + 0.89, + 0.296 + ], + "angle": 0, + "content": "Models based on different attention methods have been actively explored for the answer grounding task." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.297, + 0.892, + 0.401 + ], + "angle": 0, + "content": "MAC-Caps [39] is based on MAC [15] which has a recurrent reasoning architecture that performs \\( T \\) reasoning steps to answer the question. At each reasoning step, MAC uses an attention block to read from image features and writing memory. MAC-Caps adds capsule layers on top of the convolutional layers to obtain visual capsules from the feature-maps." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.403, + 0.892, + 0.492 + ], + "angle": 0, + "content": "Att-MCB [32] is designed to use a Knowledge Base. It is composed of nine modules, out of which two modules use an attention mechanism. One module uses an attention mechanism to generate visual explanations, and the other one to consume and point out relevant information from the Knowledge Base." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.493, + 0.892, + 0.688 + ], + "angle": 0, + "content": "A multi-grained attention method is introduced in [14]. It consists of two types of object-level groundings to explore fine-grained information, and a more sophisticated language model for better question representation. A Word-Label Matching attention vector that indicates the weight that should be given to each of the \\( K \\) objects in the image is computed in terms of the semantic similarity between the category labels of the objects and the words in the question. A word-object matching module is exploited to evaluate how likely a question word matches a visual object. A Sentence-Object attention module is used to capture the global semantics of the whole sentence to guide the focus on relevant objects." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.69, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Multiple custom-designed attention modules are used in [46] to generate an attention map for the given image and question pair. Questions are tokenized and passed through an embedding layer followed by an LSTM layer to generate the question features. An Image Attention Supervision Module is used as an auxiliary classification task; that is, the ground-truth visual grounding labels are used to guide the model to focus on significant parts of the image to answer each question." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "A so-called \"accumulated attention\" mechanism is introduced in [5]. It consists of three attention modules, one for the input text, one for the objects, and one for the whole image. All three modules have a similar structure of a linear layer followed by a Softmax function." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6081" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.088, + 0.452, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.428, + 0.472, + 0.485 + ], + "angle": 0, + "content": "Figure 1. Our proposed sentence attention block. \\( C \\) denotes the number of channels in image feature-maps, and \\( E \\) represents the embedding size of the sentence encoder network. Question and answer embeddings are concatenated into one vector." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.519, + 0.355, + 0.535 + ], + "angle": 0, + "content": "2.2. Multimodal Attention Methods" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.548, + 0.47, + 0.595 + ], + "angle": 0, + "content": "Since its introduction, the attention mechanism has been widely adopted in the computer vision community due to its special capabilities for many multimodal applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.689 + ], + "angle": 0, + "content": "In [11], a multimodal attention model is proposed based on the encoder-decoder networks and RNNs for video captioning and sentence generation. In particular, the multimodal attention framework combined image, audio, and motion features by selecting each modality's most relevant context vector." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.694, + 0.47, + 0.756 + ], + "angle": 0, + "content": "In [44], utilizing stacked attention networks is suggested to look for image regions that correlate with a query answer and pinpoint representative features of a given question more accurately." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.76, + 0.47, + 0.837 + ], + "angle": 0, + "content": "More recently, in [9], a normalized variant of the self-attention mechanism, named normalized self-attention (NSA), is introduced. NSA seeks to encode and decode the image and caption features and normalize the distribution of internal activations during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "For the video question answering task, attention and memory mechanisms are used in [6] to efficiently learn visual features and the semantic correlations that precisely answer questions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.757, + 0.108 + ], + "angle": 0, + "content": "2.3. Answer Grounding Methods" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.116, + 0.892, + 0.16 + ], + "angle": 0, + "content": "The answer grounding task has also been studied under different names, such as grounded VQA and visual explanation for VQA." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.162, + 0.893, + 0.267 + ], + "angle": 0, + "content": "The VinVL [45] method is based on OSCAR [21] which proposes a novel vision-language pre-training that incorporates image anchor points as input, taken from an object detection network. VinVL improves the visual representations by using a bigger visual model pre-trained on larger training corpora that combine multiple annotated object detection datasets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.269, + 0.892, + 0.375 + ], + "angle": 0, + "content": "LXMERT [36] presents a cross-modality framework for learning the connections between vision and language. The authors build a large-scale Transformer model that consists of three encoders: an object relationship encoder, a language encoder, and a cross-modality encoder. This model is then pre-trained on a large-scale dataset of image-and-sentence pairs." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.376, + 0.892, + 0.481 + ], + "angle": 0, + "content": "MTXNet [30] is an end-to-end trainable architecture that generates multimodal explanations focusing on the text in the image. It contains a Graph Attention Network, and each object location and OCR token is treated as a node in the graph. It also contains a multimodal transformer that operates on three modalities: question words, visual objects, and OCR tokens." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.482, + 0.892, + 0.56 + ], + "angle": 0, + "content": "U-CAM [29] is a method that uses gradient-based uncertainty estimates to provide visual attention maps. The certainty measure is improved by adding the certainty gradients to the existing standard Cross-Entropy loss gradients for training the model during back-propagation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.575, + 0.591, + 0.591 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.601, + 0.892, + 0.632 + ], + "angle": 0, + "content": "Current attention methods have one or more of the following three problems:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.645, + 0.892, + 0.75 + ], + "angle": 0, + "content": "- The attention mechanism is not based on well-known and well-tested methods [32, 36, 29]. Well-established methods have been evaluated on many different tasks and datasets, especially, they perform well on large-scale datasets. Their design details have been actively studied, and their advantages and disadvantages are visible to the users." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "- The complicated design makes it challenging to re-implement them on other platforms [32, 39, 30]. Recently, many applications have been made based on deep neural networks that are expected to be used by the end-users of different platforms. For example, answer grounding methods are helpful in assistive technologies for people with vision impairments [3]. In order to use such technologies offline, it is crucial to implement the method for each platform." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.645, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6082" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.095, + 0.852, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.511, + 0.895, + 0.54 + ], + "angle": 0, + "content": "Figure 2. The complete proposed network. The multi-scale channel-wise image features contain different regions that are filtered by our Sentence Attention Block (SAB) based on the question and answer embeddings." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.564, + 0.47, + 0.655 + ], + "angle": 0, + "content": "- The custom design does not allow the usage of pretrained models [39, 45, 28, 27]. Answer grounding datasets are less diverse and have fewer samples than the classification datasets. Using pre-trained models can improve the accuracy and make the model more robust to noise and unseen samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.47, + 0.743 + ], + "angle": 0, + "content": "This paper proposes an attention block that avoids the mentioned problems. We design an attention block based on well-established and well-studied methods. It has a simple design to make it easy to re-implement and uses pre-trained models to achieve maximum accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.744, + 0.47, + 0.774 + ], + "angle": 0, + "content": "Our proposed method has three main parts: region proposal, sentence embedding, and attention fusion." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.241, + 0.802 + ], + "angle": 0, + "content": "3.1. Region Proposal" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.884 + ], + "angle": 0, + "content": "This module's primary goal is to process the image and generate all the possible region candidates. Unlike Region Proposal Networks [31] which generate anchor-based bounding boxes, this module generates dense regions (segmentations)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "A pre-trained classification network is used as the back" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.892, + 0.67 + ], + "angle": 0, + "content": "bone. However, the final classifier is removed to get the multi-scale feature-maps. During the end-to-end training of the proposed method, the backbone network learns all the candidate regions. Later, these candidate regions are filtered and combined by the proposed Sentence Attention Blocks. In Section 5.3, we show that the backbone network works as expected by visualizing the feature-maps." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.698, + 0.695 + ], + "angle": 0, + "content": "3.2. Sentence Embedding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Since the input questions and answers have variable lengths, they are processed by a sentence embedding network. A pre-trained sentence encoder is employed to encode the given question and answer to a vector space separately. The result is two vectors, one for the question and one for the answer. The two vectors are concatenated and passed to the Sentence Attention Blocks as one vector." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.667, + 0.832 + ], + "angle": 0, + "content": "3.3. Attention Fusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Given multiple potential regions for an image, the answer grounding task reduces to choosing and/or combining these regions based on the given question and/or answer. The multi-scale channel-wise image features from the back" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6083" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.118, + 0.089, + 0.432, + 0.184 + ], + "angle": 0, + "content": "
MethodMean IoU
MTXNet [30]18.9
LXMERT [36]24.9
VinVL [45]25.6
MAC-Caps [39]26.1
Ours29.0 (+2.9)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.193, + 0.47, + 0.222 + ], + "angle": 0, + "content": "Table 1. Comparison results on the test set of the TextVQA-X dataset [30]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.244, + 0.468, + 0.319 + ], + "angle": 0, + "content": "bone network contain candidate regions. The goal is to filter and combine these regions based on the question and answer vector from the sentence embedding network. Towards this goal, we design an attention block based on the Squeeze-and-Excitation (SE) block [13]." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.32, + 0.468, + 0.35 + ], + "angle": 0, + "content": "We start with an SE block and make four modifications as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.468, + 0.42 + ], + "angle": 0, + "content": "1. The first modification is changing it from self-attention to cross-attention. In other words, unlike a self-attention block, the proposed block pays attention to an external source: the encoded question and answer." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.429, + 0.468, + 0.49 + ], + "angle": 0, + "content": "2. The second modification is adding a normalization layer. Empirically we found that adding a normalization layer improves the accuracy. Similar to [40], we use LayerNorm since it performs better on sequences." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.499, + 0.468, + 0.559 + ], + "angle": 0, + "content": "3. The third modification is replacing the sigmoid function with softmax. Since the goal is to force the model to choose a few of the channel-wise image features, the sigmoid function is replaced with softmax." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.569, + 0.468, + 0.613 + ], + "angle": 0, + "content": "4. The last modification is inverting the bottleneck. Instead of shrinking the embedding space, we expand the question and answer embeddings by a factor of two." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.359, + 0.468, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.623, + 0.468, + 0.681 + ], + "angle": 0, + "content": "In Section 5.1, the effect of each modification on the network accuracy is shown. Figure 1 depicts the final structure of the proposed block. Precisely, the proposed attention block is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.693, + 0.468, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S o f t m a x} (F C (\\operatorname {R e L U} (L N (F C (\\mathbf {Q A})))) \\times \\mathbf {R} _ {i} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.468, + 0.809 + ], + "angle": 0, + "content": "where \\(FC\\) is a single-layer perceptron, \\(LN\\) is LayerNorm, QA is the joint question and answer embeddings, and \\(\\mathbf{R}_i\\) is the image regions at scale \\(i\\). For each scale, a separate instance of this attention block is created so that each block learns special weights for its corresponding feature-map scale." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.902 + ], + "angle": 0, + "content": "After computing the attention block for all the scales, all the outputs at different scales are combined into one output. Starting from the lowest resolution output, a \\(1 \\times 1\\) convolution is used to convert the number of channels to the number of channels of the next lowest resolution output. Then, Bilinear interpolation is used for upsampling the" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.089, + 0.872, + 0.199 + ], + "angle": 0, + "content": "
MethodMean IoU
DeconvNet [7]29.8
Mask Aggregation [7]32.6
LXMERT [36]33.3
VinVL [45]33.9
MAC-Caps [39]34.3
Ours36.8 (+2.5)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.208, + 0.89, + 0.223 + ], + "angle": 0, + "content": "Table 2. Comparison results on the test set of the VQS dataset [7]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.248, + 0.892, + 0.353 + ], + "angle": 0, + "content": "output feature-map by a factor of 2. Finally, the output is added to the next lowest resolution output. This process is repeated until no other output is left. In the end, a \\(1 \\times 1\\) convolution is used to reduce the number of channels of the final output to two. Therefore, the final output has two channels representing answer grounding and its background. We use Argmax to output a binary image." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.354, + 0.892, + 0.384 + ], + "angle": 0, + "content": "The complete design of the proposed method is depicted in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.394, + 0.788, + 0.41 + ], + "angle": 0, + "content": "3.4. Comparison to Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.418, + 0.892, + 0.46 + ], + "angle": 0, + "content": "We compare the design of our proposed method to the top-performing existing methods for the answer grounding task." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.463, + 0.892, + 0.629 + ], + "angle": 0, + "content": "MAC-Caps [39] is one of the best-performing networks that currently exist. This network is based on MAC [15] which employs attention blocks in a recurrent reasoning architecture. MAC-Caps adds capsule layers to process the feature-maps further. Although similar to our method, MAC-Caps employs an attention mechanism. However, the recurrent design brings the usual problems of RNNs, such as gradient vanishing/exploding, longer training time, and not performing well on long sequences. Since our proposed method does not have a recurrent design, it avoids such problems." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.63, + 0.892, + 0.824 + ], + "angle": 0, + "content": "VinVL [45] is another top-performing network for the answer grounding task. This method is based on OSCAR [21] which incorporates image anchor points as inputs. VinVL improves the visual representations by using a bigger visual model pre-trained on larger training corpora that combine multiple annotated object detection datasets. This method relies on object detection networks to provide the anchor points, which adds an extra task for the network to learn. Also, using large generic corpora may not improve the accuracy for special datasets such as VizWiz-VQA-Grounding [3]. In contrast, our proposed method relies only on the feature-maps and does not define another task." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Att-MCB [32] is the state-of-the-art network on the VQA-X dataset [27]. This method is designed to use a Knowledge Base. It comprises nine modules and four loss functions, making it a tedious design. This complicated design is resource-demanding and makes the training sensitive" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6084" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.114, + 0.089, + 0.434, + 0.228 + ], + "angle": 0, + "content": "
MethodRank Correlation
PJ-X [27]0.342
CCM [28]0.368
U-CAM [29]0.372
VinVL [45]0.373
Att-MFH [32]0.376
MAC-Caps [39]0.389
Att-MCB [32]0.396
Ours0.421 (+0.025)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.239, + 0.47, + 0.267 + ], + "angle": 0, + "content": "Table 3. Comparison results on the test set of the VQA-X dataset [27]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.29, + 0.468, + 0.352 + ], + "angle": 0, + "content": "to hyper-parameters. Our proposed method comprises three main modules and one loss function. This minimalistic design helps the gradient flow during back-propagation, and the training is less sensitive to hyper-parameters." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.364, + 0.21, + 0.381 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.389, + 0.469, + 0.435 + ], + "angle": 0, + "content": "We train and test our proposed method on four answer grounding datasets. We use the same setup for all of the experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.444, + 0.159, + 0.46 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.47, + 0.603 + ], + "angle": 0, + "content": "We use a pre-trained EfficientNet [37] to extract the multi-scale image features and we use a MiniLMv2 [42] sentence encoder network to encode the questions. In Section 5.2, we perform an ablation study on different backbones. we use BEiT [41] to generate answers for the datasets where the answers are not publicly available. We use RMI [47] with its default settings along with cross-entropy. Therefore, the overall loss function to minimize is:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.611, + 0.469, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {a l l} (y, p) = \\frac {1}{B} \\sum_ {b = 1} ^ {B} \\lambda \\mathcal {L} _ {c e} \\left(y ^ {(b)}, p ^ {(b)}\\right) \\tag {3} \\\\ + (1 - \\lambda) \\mathcal {L} _ {r m i} (y ^ {(b)}, p ^ {(b)}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.683, + 0.469, + 0.759 + ], + "angle": 0, + "content": "where \\(\\lambda \\in [0,1]\\) is a weight factor, \\(B\\) denotes the number of samples in a mini-batch, \\(\\mathcal{L}_{ce}(y^{(b)},p^{(b)})\\) is the standard cross entropy loss between the \\(b\\)-th sample and its corresponding prediction, respectively, and \\(\\mathcal{L}_{rmi}(y^{(b)},p^{(b)})\\) is the RMI loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.767, + 0.469, + 0.808 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r m i} (\\mathbf {Y}, \\mathbf {P}) = \\sum_ {c = 1} ^ {C} \\frac {1}{2 d} \\operatorname {T r a c e} (\\log (\\mathbf {M})) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.818, + 0.11, + 0.831 + ], + "angle": 0, + "content": "and," + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.842, + 0.469, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} = \\boldsymbol {\\Sigma} _ {\\mathbf {Y}} - C o v (\\mathbf {Y}, \\mathbf {P}) \\left(\\boldsymbol {\\Sigma} _ {\\mathbf {P}} ^ {- 1}\\right) ^ {T} C o v (\\mathbf {Y}, \\mathbf {P}) ^ {T} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.87, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{Y} = y^{(b)}\\), \\(\\mathbf{P} = p^{(b)}\\), \\(\\mathbf{M} \\in \\mathbb{R}^{d \\times d}\\), \\(d\\) denotes the number of pixels, \\(C\\) is the number of object classes, here" + }, + { + "type": "table", + "bbox": [ + 0.566, + 0.088, + 0.828, + 0.199 + ], + "angle": 0, + "content": "
TeamMean IoU
MindX66.9
MGTV69.7
hsslab_inspur70.1
GroundTruth70.3
Aurora70.6
SAB (Ours)72.4 (+1.8)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.208, + 0.892, + 0.251 + ], + "angle": 0, + "content": "Table 4. The top-performing teams on the VizWiz-VQA-Grounding Challenge 2022 leaderboard [3]. At the time of this writing, none of the methods have been published." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.264, + 0.886, + 0.344 + ], + "angle": 0, + "content": "
Method# of paramsinference time
LXMERT [36]153M56ms
VinVL [45]406M69ms
MAC-Caps [39]80M54ms
Ours88M49ms
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.353, + 0.892, + 0.381 + ], + "angle": 0, + "content": "Table 5. Comparison of our proposed method to existing methods regarding the number of parameters and inference time." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.407, + 0.892, + 0.453 + ], + "angle": 0, + "content": "\\(C = 2\\) for the answer grounding and its background, \\(\\Sigma_{\\mathbf{Y}}\\) is the variance matrix of \\(\\mathbf{Y}\\), and \\(Cov(\\mathbf{Y},\\mathbf{P})\\) is the covariance matrix of \\(\\mathbf{Y}\\) and \\(\\mathbf{P}\\). Similar to [47], we set \\(\\lambda = 0.5\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.893, + 0.544 + ], + "angle": 0, + "content": "We use AdamW optimizer [25] with a weight decay of 0.05 and batch size of 16. We apply the \"polynomial\" learning rate policy with a poly exponent of 0.9 and an initial learning rate of 0.0001. Synchronized batch normalization is used across multiple GPUs. We use RandAugment [4] for data augmentation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.555, + 0.63, + 0.571 + ], + "angle": 0, + "content": "4.2. TextVQA-X" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.776 + ], + "angle": 0, + "content": "The TextVQA-X dataset [30] is a subset of the TextVQA dataset [33] such that the answer groundings are generated through a manual segmentation annotation process. TextVQA-X consists of 10,379 training images and 3,354 test images. Since only a few models have been evaluated on this dataset, we train three more models for a fair comparison. Similar to [3], we train top-performing methods that their code is publicly available; specifically, LXMERT [36], VinVL [45], and MAC-Caps [39]. We use the same setup as in [3] to train the three models with the difference that we train the models on the TextVQA-X training set. Similar to the existing methods, we use the test set to evaluate our method and report the mean IoU measure." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.786, + 0.574, + 0.803 + ], + "angle": 0, + "content": "4.3.VQS" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The VQS dataset [7] builds upon the images, instance segmentation masks, and bounding boxes in COCO [22] and the questions and answers in the VQA dataset [1]. VQS contains 26,995 training, 5,000 validation, and 5,873 test images. Similar to Section 4.2, we train top-performing methods that their code is publicly available; specifically," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6085" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.111, + 0.089, + 0.437, + 0.214 + ], + "angle": 0, + "content": "
BackboneMean IoU
ResNet-50x3 [19]71.9
Swin Transformer V2-B [24]72.8
EfficientNetV2-L [38]73.1
EfficientNet-B7 [37]73.5
RoBERTa [23]72.8
MPNet [34]73.1
MiniLMv2 [42]73.5
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.224, + 0.47, + 0.306 + ], + "angle": 0, + "content": "Table 6. The ablation study of the pre-trained backbone networks. Top Section: Using a MiniLMv2 as a fixed sentence embedding network and replacing the image feature extraction networks. Bottom Section: Using a fixed image feature extraction network (EfficientNet-B7) and replacing the sentence embedding networks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.469, + 0.41 + ], + "angle": 0, + "content": "LXMERT [36], VinVL [45], and MAC-Caps [39]. We use the same setup as in [3] to train the three models with the difference that we train the models on the VQS training set. Similar to the existing methods, we use the test set to evaluate our method and report the mean IoU measure." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.426, + 0.173, + 0.442 + ], + "angle": 0, + "content": "4.4.VQA-X" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.47, + 0.633 + ], + "angle": 0, + "content": "The VQA-X dataset [27] does not have visual annotations for its training set. It has visual annotations only for its validation and test sets. Therefore, we train our method on the training set of the VQS dataset [7], which has related images to the VQA-X validation and test sets. The images of both datasets are from the COCO dataset [22]. The other methods are trained on different datasets, including Visual Genome [20] and VQA v2.0 [8]. The VQA-X dataset consists of 24,876 training, 1,431 validation, and 1,921 test images. Similar to the other methods, we use the test set of the VQA-X dataset to evaluate our method and report the Rank Correlation as in [27]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.649, + 0.307, + 0.666 + ], + "angle": 0, + "content": "4.5. VizWiz-VQA-Grounding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The VizWiz-VQA-Grounding dataset [3] is based on the VizWiz-VQA dataset [10], and the images and questions come from visually impaired people who shared them to request visual assistance in their day-to-day lives. VizWiz-VQA-Grounding contains a total of 9,998 VQAs divided into 6,494/1,131/2,373 VQAs for training, validation, and testing. The 2022 VizWiz-VQA-Grounding Challenge is designed around the aforementioned VizWiz-VQA-Grounding dataset. This challenge was held recently, and at the time of this writing, current methods in this challenge have not been published yet. Therefore, we evaluate our method using the VizWiz-VQA-Grounding Challenge online leaderboard [3]. The leaderboard evaluates all the methods against the VizWiz-VQA-Grounding test set. The ground truth labels for the test set are not published yet." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.089, + 0.865, + 0.169 + ], + "angle": 0, + "content": "
ModificationMean IoU
Switching to cross-attention70.1
Adding normalization layer71.2 (+1.1)
Switching to Softmax72.7 (+1.5)
Expansion of the embeddings73.5 (+0.8)
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.178, + 0.892, + 0.219 + ], + "angle": 0, + "content": "Table 7. The results of our ablation study of the design modifications. The modifications are done cumulatively from top to bottom." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.248, + 0.593, + 0.262 + ], + "angle": 0, + "content": "4.6. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.892, + 0.394 + ], + "angle": 0, + "content": "We evaluate our proposed method on four answer grounding datasets. Tables 1, 2, and 3 compare the results of our proposed method with the existing methods on the TextVQA-X, VQS, and VQA-X datasets respectively. Our method achieves new state-of-the-art accuracy on three datasets. More specifically, improvements of 2.9, 2.5, and 0.025 on the TextVQA-X, VQS, and VQA-X datasets, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Table 4 lists the top-performing teams on the VizWiz-VQA-Grounding Challenge 2022 leaderboard [3]. At the time of this writing, our method holds first place with considerably higher accuracy than the other methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.52 + ], + "angle": 0, + "content": "Furthermore, Table 5 compares our proposed method to existing methods regarding the number of parameters and inference time. The efficient design of our method makes it the fastest method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.543, + 0.664, + 0.558 + ], + "angle": 0, + "content": "5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.892, + 0.63 + ], + "angle": 0, + "content": "In this section, we perform four ablation studies. All of the experiments are done on the VizWiz-VQA-Grounding [3] validation set and using the setup from Section 4.1 unless otherwise stated." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.648, + 0.697, + 0.664 + ], + "angle": 0, + "content": "5.1. Design Modifications" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We started the design of our sentence attention block from the standard SE block and made four modifications as explained in Section 3.3. In this section, we perform an ablation study on these modifications. Table 7 shows the impact of each modification on the final accuracy. These modifications are done cumulatively. The first modification is changing from self-attention to cross-attention, which makes our baseline for this ablation study. The second modification is adding a normalization layer. This modification improves our baseline by 1.1 percent. The third modification is replacing the Sigmoid function with Softmax. This modification has the highest impact on our method, 1.5 percent. And the fourth modification is the expansion of the embeddings. This last modification brings an addition of 0.8 percent." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6086" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.27, + 0.106 + ], + "angle": 0, + "content": "5.2. Backbone Networks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.474, + 0.465 + ], + "angle": 0, + "content": "The flexibility of our method enables us to use a variety of pre-trained networks. In this ablation study, we compare and analyze different pre-trained networks. To compare the image backbone networks, we keep the sentence embedding network fixed and change the image backbone network. Similarly, to compare the sentence embedding networks, we keep the image backbone network with the highest accuracy and change the sentence embedding network. We use well-known and well-performing networks whose pre-trained weights are publicly available. Specifically, we use EfficientNet [37], EfficientNetV2 [38], Swin Transformer V2 [24], and ResNet-50x3 [19] for image feature extraction and MiniLMv2 [42], MPNet [34], and RoBERTa [23] for sentence embedding. Table 6 shows the results of this ablation study. The highest accuracy is achieved by using EfficientNet-B7 and MiniLMv2. Since the EfficientNetV2-L reduces the input resolution more aggressively than EfficientNet-B7, it performs worse. Transformer networks usually perform well on large datasets. However, since VizWiz-VQA-Grounding is not a large dataset, Swin Transformer V2 does not achieve the highest accuracy. The ResNet-50x3 is pre-trained on ImageNet21k, but it achieves the least accuracy in this experiment." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.469, + 0.248, + 0.485 + ], + "angle": 0, + "content": "5.3. Region Proposals" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.47, + 0.538 + ], + "angle": 0, + "content": "Feature-map visualization will provide insight into the internal representations of each layer in a network for a specific input image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.471, + 0.81 + ], + "angle": 0, + "content": "This ablation study aims to show that the image backbone network has learned the potential regions for an image, which are then processed by the proposed sentence attention block. To this aim, we visualize the final output feature-maps of the image backbone network, before feeding them to an attention block. We use the image backbone of our fully trained network on the VizWiz dataset. Figure 3 (a) shows a few manually-chosen feature-maps for a sample image. These feature-maps show that after training, the backbone network has been able to learn the potential important regions of images. Having all the candidate regions, the task of the proposed attention block is to filter the irrelevant regions to the question and answer. Figure 3 (b) shows the final output of the complete network for the same image in (a), using different questions and answers (for brevity, the answers are not shown). This figure shows that our proposed block is able to filter the channel-wise feature-maps based on the input questions and answers." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.265, + 0.833 + ], + "angle": 0, + "content": "5.4. Multi-scale Fusions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "There is a trade-off in the answer grounding task such that some images are best handled at lower inference resolution and others better handled at higher inference resolution. Fine details, such as the edges of objects or thin" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.089, + 0.87, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.488, + 0.895, + 0.613 + ], + "angle": 0, + "content": "Figure 3. (a) Eight manually-selected sample feature-maps for an image. These feature-maps are the output of the feature extraction backbone network, before feeding them to the sentence attention block. The network has learned to output potential regions in different channels. (b) The final output of our complete network for the same image in (a), given different questions and answers (for brevity, the answers are not shown). This figure shows that our proposed attention block can filter the channel-wise feature-maps based on the input questions and answers." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.625, + 0.861, + 0.705 + ], + "angle": 0, + "content": "
ScaleMean IoU
1/3271.4
1/32 + 1/1672.5 (+1.1)
1/32 + 1/16 + 1/873.1 (+0.6)
1/32 + 1/16 + 1/8 + 1/473.5 (+0.4)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.714, + 0.895, + 0.772 + ], + "angle": 0, + "content": "Table 8. The gradual effect of multi-scale fusion. From top to bottom, the first row uses only the lowest resolution feature-map, and each row adds higher resolution feature-maps. Each scale is the ratio of the feature-map size to the input image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.87 + ], + "angle": 0, + "content": "structures, are often better predicted with scaled-up images. At the same time, the prediction of large structures, which requires more global context, is often done better at scaled-down images because the network's receptive field can observe more of the necessary context." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This ablation study shows the impact of adding the extra scales on the accuracy. Table 8 shows the results of this" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6087" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.152 + ], + "angle": 0, + "content": "study. From top row to bottom, we start with only one scale and one attention block and gradually add higher resolution feature-maps. The lowest resolution is the final output feature-map of the backbone network." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.167, + 0.205, + 0.183 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.193, + 0.473, + 0.435 + ], + "angle": 0, + "content": "In the study, a new architectural block termed the Sentence Attention Block was introduced to address certain challenges. This block recalibrates channel-wise image feature-maps by modeling inter-dependencies between image feature-maps and sentence embedding. The design began with a recognized attention method, and with minor adjustments, enhanced results were achieved, reaching top-tier accuracy. We showed the block's ability to filter out irrelevant feature-map channels based on sentence embedding. The method's adaptability allows for various pre-trained backbone networks, and its straightforward nature facilitates comprehension and re-implementation. The method's efficacy was demonstrated on the TextVQA-X, VQS, VQA-X, and VizWiz-VQA-Grounding datasets, and several ablation studies highlighted the validity of the design decisions." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.449, + 0.175, + 0.465 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.475, + 0.47, + 0.543 + ], + "angle": 0, + "content": "[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.546, + 0.47, + 0.602 + ], + "angle": 0, + "content": "[2] Khaled Bayoudh, Raja Knani, Fayçal Hamdaoui, and Abdel-latif Mtibaa. A survey on deep multimodal learning for computer vision: advances, trends, applications, and datasets. The Visual Computer, 38(8):2939-2970, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.604, + 0.468, + 0.645 + ], + "angle": 0, + "content": "[3] Chongyan Chen, Samreen Anjum, and Danna Gurari. Grounding answers for visual questions asked by visually impaired people. arXiv preprint arXiv:2202.01993, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.647, + 0.468, + 0.715 + ], + "angle": 0, + "content": "[4] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.718, + 0.468, + 0.773 + ], + "angle": 0, + "content": "[5] Chaorui Deng, Qi Wu, Qingyao Wu, Fuyuan Hu, Fan Lyu, and Mingkui Tan. Visual grounding via accumulated attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7746-7755, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.775, + 0.468, + 0.856 + ], + "angle": 0, + "content": "[6] Chenyou Fan, Xiaofan Zhang, Shu Zhang, Wensheng Wang, Chi Zhang, and Heng Huang. Heterogeneous memory enhanced multimodal attention model for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1999-2007, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.859, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[7] Chuang Gan, Yandong Li, Haoxiang Li, Chen Sun, and Boqing Gong. Vqs: Linking segmentations to questions and answers for supervised attention in vqa and question-focused" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.475, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "semantic segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1811-1820, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[8] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.207, + 0.894, + 0.275 + ], + "angle": 0, + "content": "[9] Longteng Guo, Jing Liu, Xinxin Zhu, Peng Yao, Shichen Lu, and Hanqing Lu. Normalized and geometry-aware self-attention network for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10327–10336, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[10] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3608-3617, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[11] Chiori Hori, Takaaki Hori, Teng-Yok Lee, Ziming Zhang, Bret Harsham, John R Hershey, Tim K Marks, and Kazuhiko Sumi. Attention-based multimodal fusion for video description. In Proceedings of the IEEE international conference on computer vision, pages 4193-4202, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[12] Parisa Hosseini, Seyedalireza Khoshkirat, Mohammad Jalayer, Subasish Das, and Huaguo Zhou. Application of text mining techniques to identify actual wrong-way driving (wwd) crashes in police reports. International Journal of Transportation Science and Technology, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[13] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[14] Pingping Huang, Jianhui Huang, Yuqing Guo, Min Qiao, and Yong Zhu. Multi-grained attention with object-level grounding for visual question answering. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3595-3600, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[15] Drew A Hudson and Christopher D Manning. Compositional attention networks for machine reasoning. arXiv preprint arXiv:1803.03067, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[16] Seyedalireza Khoshsirat and Chandra Kambhamtu. Semantic segmentation using neural ordinary differential equations. Advances in Visual Computing: 17th International Symposium, ISVC 2022, San Diego, CA, USA, October 3-5, 2022, Proceedings, Part I, pages 284-295, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[17] Seyedalireza Khoshsurat and Chandra Kambhametu. Embedding attention blocks for the vizwiz answer grounding challenge. VizWiz Grand Challenge Workshop, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[18] Seyedalireza Khoshsirat and Chandra Kambhamtu. Empowering visually impaired individuals: A novel use of apple live photos and android motion photos. 25th Irish Machine Vision and Image Processing Conference, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[19] Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, and Neil Houlsby. Big transfer (bit): General visual representation learning. In European conference on computer vision, pages 491-507. Springer, 2020." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6088" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.175 + ], + "angle": 0, + "content": "[20] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.472, + 0.247 + ], + "angle": 0, + "content": "[21] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.47, + 0.317 + ], + "angle": 0, + "content": "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.319, + 0.47, + 0.387 + ], + "angle": 0, + "content": "[23] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[24] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. arXiv preprint arXiv:2111.09883, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.469, + 0.474 + ], + "angle": 0, + "content": "[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.469, + 0.546 + ], + "angle": 0, + "content": "[26] Elham Maserat, Reza Safdari, Hamid Asadzadeh Aghdaei, Alireza Khoshsirat, and Mohammad Reza Zali. 43: Designing evidence based risk assessment system for cancer screening as an applicable approach for the estimating of treatment roadmap. BMJ Open, 7(Suppl 1):bmjopen-2016, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[27] Dong Huk Park, Lisa Anne Hendricks, Zeynep Akata, Anna Rohrbach, Bernt Schiele, Trevor Darrell, and Marcus Rohrbach. Multimodal explanations: Justifying decisions and pointing to the evidence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8779-8788, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[28] Badri Patro, Shivansh Patel, and Vinay Namboodiri. Robust explanations for visual question answering. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1577-1586, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[29] Badri N Patro, Mayank Lunayach, Shivansh Patel, and Vinay P Namboodiri. U-cam: Visual explanation using uncertainty based class activation maps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7444-7453, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[30] Varun Nagaraj Rao, Xingjian Zhen, Karen Hovsepian, and Mingwei Shen. A first look: Towards explainable textvqa models via visual and textual explanations. arXiv preprint arXiv:2105.02626, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[31] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[32] Felipe Riquelme, Alfredo De Goyeneche, Yundong Zhang, Juan Carlos Niebles, and Alvaro Soto. Explaining vqa pre" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "dictions using visual grounding and a knowledge base. Image and Vision Computing, 101:103968, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[33] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[34] Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. Mpnet: Masked and permuted pre-training for language understanding. Advances in Neural Information Processing Systems, 33:16857-16867, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[35] Jabeen Summaira, Xi Li, Amin Muhammad Shoib, Songyuan Li, and Jabbar Abdul. Recent advances and trends in multimodal deep learning: A review. arXiv preprint arXiv:2105.11087, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[36] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.35, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[37] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.407, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[38] Mingxing Tan and Quoc V Le. Efficientnetv2: Smaller models and faster training. arXiv preprint arXiv:2104.00298, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[39] Aisha Urooj, Hilde Kuehne, Kevin Duarte, Chuang Gan, Niels Lobo, and Mubarak Shah. Found a reason for me? weakly-supervised grounded visual question answering using capsules. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8465-8474, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.534, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[41] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.662, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[42] Wenhui Wang, Hangbo Bao, Shaohan Huang, Li Dong, and Furu Wei. Minilmv2: Multi-head self-attention relation distillation for compressing pretrained transformers. arXiv preprint arXiv:2012.15828, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.719, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[43] Xi Wei, Tianzhu Zhang, Yan Li, Yongdong Zhang, and Feng Wu. Multi-modality cross attention network for image and sentence matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10941-10950, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[44] Zichao Yang, Xiaodong He, Jianfeng Gao, Li Deng, and Alex Smola. Stacked attention networks for image question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 21-29, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[45] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6089" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.119 + ], + "angle": 0, + "content": "Computer Vision and Pattern Recognition, pages 5579-5588, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.189 + ], + "angle": 0, + "content": "[46] Yundong Zhang, Juan Carlos Niebles, and Alvaro Soto. Interpretable visual question answering by visual grounding from attention supervision mining. In 2019 IEEE winter conference on applications of computer vision (wacv), pages 349-357. IEEE, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[47] Shuai Zhao, Yang Wang, Zheng Yang, and Deng Cai. Region mutual information loss for semantic segmentation. Advances in Neural Information Processing Systems, 32, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6090" + } + ] +] \ No newline at end of file diff --git a/2023/Sentence Attention Blocks for Answer Grounding/0622d151-144e-4e04-b267-1cccf0dba507_origin.pdf b/2023/Sentence Attention Blocks for Answer Grounding/0622d151-144e-4e04-b267-1cccf0dba507_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d461452d357c7a7d8bb3790f2007ba3521b96958 --- /dev/null +++ b/2023/Sentence Attention Blocks for Answer Grounding/0622d151-144e-4e04-b267-1cccf0dba507_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a6361299250f8eb3c9bfbeec3ff84953bc532390ea96ead8f452eaabba91bb4 +size 625237 diff --git a/2023/Sentence Attention Blocks for Answer Grounding/full.md b/2023/Sentence Attention Blocks for Answer Grounding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..545140de00e6c5734690a7f7233d78d5f60f8bc7 --- /dev/null +++ b/2023/Sentence Attention Blocks for Answer Grounding/full.md @@ -0,0 +1,341 @@ +# Sentence Attention Blocks for Answer Grounding + +Seyedalireza Khoshsirat Chandra Kambhametty +Video/Image Modeling and Synthesis (VIMS) Lab, University of Delaware +{alireza, chandrak}@udel.edu + +# Abstract + +Answer grounding is the task of locating relevant visual evidence for the Visual Question Answering task. While a wide variety of attention methods have been introduced for this task, they suffer from the following three problems: designs that do not allow the usage of pre-trained networks and do not benefit from large data pre-training, custom designs that are not based on well-grounded previous designs, therefore limiting the learning power of the network, or complicated designs that make it challenging to re-implement or improve them. In this paper, we propose a novel architectural block, which we term Sentence Attention Block, to solve these problems. The proposed block re-calibrates channel-wise image feature-maps by explicitly modeling inter-dependencies between the image feature-maps and sentence embedding. We visually demonstrate how this block filters out irrelevant feature-maps channels based on sentence embedding. We start our design with a well-known attention method, and by making minor modifications, we improve the results to achieve state-of-the-art accuracy. The flexibility of our method makes it easy to use different pre-trained backbone networks, and its simplicity makes it easy to understand and be re-implemented. We demonstrate the effectiveness of our method on the TextVQA-X, VQS, VQA-X, and VizWiz-VQA-Grounding datasets. We perform multiple ablation studies to show the effectiveness of our design choices. + +# 1. Introduction + +# 1.1. Visual Question Answering + +Visual Question Answering (VQA) systems try to accurately answer natural language questions regarding an input image [1]. This topic aims at developing systems that can communicate effectively about an image in natural language and comprehend the contents of images similar to humans. + +# 1.2. Answer Grounding + +The answer grounding task is defined as detecting the pixels that can provide evidence for the answer to a given question regarding an image [3]. In other words, the task is to return the image regions used to arrive at the answer for a given visual question (question-image pair) with an answer. Although the VQA community has made significant progress, the best-performing systems are complicated black-box models, raising concerns about whether their answer reasoning is based on correct visual evidence. By understanding the reasoning mechanism of the model, we can evaluate the quality of answers, improve model performance, and provide explanations for end-users. To address this problem, answer grounding has been introduced into VQA systems, which requires the model to locate relevant image regions as well as answer visual questions. + +By providing answer groundings in response to visual questions, numerous applications become possible. First, they allow for evaluating whether a VQA model is reasoning correctly based on visual evidence. This is useful as an explanation as well as for assisting developers with model debugging. Second, answer grounding makes it possible to separate important regions from irrelevant background regions. Given that non-professional users can mistakenly have private information in the background of their pictures, answer grounding is a useful tool to obfuscate the background for privacy preserving. Third, suppose a service can magnify the relevant visual evidence. In that case, users will be able to discover the needed information in less time. This is useful in part because VQA answers can be insufficient sometimes. + +# 1.3. Multimodal Deep Learning + +By definition, the VQA and answer grounding tasks are multimodal tasks since a method for these tasks should be able to process and correlate two different modalities. + +Multimodal Joint-Embedding Models: These models merge and learn representations of multiple modalities in a joint feature space. Joint-embeddings underpin the building of a lot of cross-modal methods as they can bridge the gap between different modalities. In this joint space, the dis + +tance of different points is equivalent to the semantic distance between their corresponding original inputs. + +Multimodal Attention-based Models: The main objective of the attention mechanism is to design systems that use local features of image or text, extract features from different regions, and assign priorities to them. The attention portion in an image determines salient regions. Then the language generation component focuses more on those salient regions for additional processing [35, 2, 12, 16, 26]. + +# 1.4. Attention Mechanism + +In deep learning, attention is a mechanism that mimics cognitive attention. The goal is to enhance the essential features of the input data and vanish out the rest. Attention methods can be classified into two classes based on their inputs: self-attention and cross-attention. Self-attention (also known as intra-attention) is a type of attention that quantifies the interdependence within the elements of a single input. At the same time, cross-attention (also known as inter-attention) finds the interdependence across two or more inputs [40, 18, 17]. Usually, cross-attention methods are used for multimodal inputs [43]. Cross-attention models first process individual modalities using modality-specific encoders, then the encoded features are fed into cross-attention modules. + +The Squeeze-and-Excitation method [13] is a channelwise self-attention mechanism widely used in classification networks [37, 38]. It consists of a global average pooling of the input, followed by two linear layers with an interleaved non-linearity and a sigmoid function. Concretely, the output of this method is: + +$$ +\sigma \left(F C (R E L U (F C (g _ {-} a v g _ {-} p o o l (\mathbf {X})))\right) \times \mathbf {X} \tag {1} +$$ + +where $\mathbf{X}$ is the feature-maps. This module aims to dynamically focus on more important channels, essentially increasing the importance of specific channels over others. This is accomplished by scaling the more important channels by a higher value. While many feature descriptors exist to reduce the spatial dimensions of the feature maps to a singular value, this module uses average pooling to keep the required computation low. The next part of the module maps the scaling weights using a Multi-Layer Perceptron (MLP) with a bottleneck structure. The result values are scaled to a range of 0-1 by passing them through a sigmoid activation layer. Afterward, using a common broadcasted element-wise multiplication, the output is applied directly to the input. This multiplication scales each input feature-maps channel with its corresponding weight learned from the MLP in the Excitation module. + +The following is the summary of our contributions: + +- We present a novel attention module based on the Squeeze-and-Eccitation method for the answer grounding task. + +- We evaluate our method on common datasets and show that it achieves new state-of-the-art results. +- We compare our design with the top-performing networks. +- We perform multiple ablation studies to learn more about this network. + +# 2. Related Work + +# 2.1. Attention Methods for Answer Grounding + +Models based on different attention methods have been actively explored for the answer grounding task. + +MAC-Caps [39] is based on MAC [15] which has a recurrent reasoning architecture that performs $T$ reasoning steps to answer the question. At each reasoning step, MAC uses an attention block to read from image features and writing memory. MAC-Caps adds capsule layers on top of the convolutional layers to obtain visual capsules from the feature-maps. + +Att-MCB [32] is designed to use a Knowledge Base. It is composed of nine modules, out of which two modules use an attention mechanism. One module uses an attention mechanism to generate visual explanations, and the other one to consume and point out relevant information from the Knowledge Base. + +A multi-grained attention method is introduced in [14]. It consists of two types of object-level groundings to explore fine-grained information, and a more sophisticated language model for better question representation. A Word-Label Matching attention vector that indicates the weight that should be given to each of the $K$ objects in the image is computed in terms of the semantic similarity between the category labels of the objects and the words in the question. A word-object matching module is exploited to evaluate how likely a question word matches a visual object. A Sentence-Object attention module is used to capture the global semantics of the whole sentence to guide the focus on relevant objects. + +Multiple custom-designed attention modules are used in [46] to generate an attention map for the given image and question pair. Questions are tokenized and passed through an embedding layer followed by an LSTM layer to generate the question features. An Image Attention Supervision Module is used as an auxiliary classification task; that is, the ground-truth visual grounding labels are used to guide the model to focus on significant parts of the image to answer each question. + +A so-called "accumulated attention" mechanism is introduced in [5]. It consists of three attention modules, one for the input text, one for the objects, and one for the whole image. All three modules have a similar structure of a linear layer followed by a Softmax function. + +![](images/e00e29fb6741d093fa214e0c26ad6998080f825a199cb239361181c925f00739.jpg) +Figure 1. Our proposed sentence attention block. $C$ denotes the number of channels in image feature-maps, and $E$ represents the embedding size of the sentence encoder network. Question and answer embeddings are concatenated into one vector. + +# 2.2. Multimodal Attention Methods + +Since its introduction, the attention mechanism has been widely adopted in the computer vision community due to its special capabilities for many multimodal applications. + +In [11], a multimodal attention model is proposed based on the encoder-decoder networks and RNNs for video captioning and sentence generation. In particular, the multimodal attention framework combined image, audio, and motion features by selecting each modality's most relevant context vector. + +In [44], utilizing stacked attention networks is suggested to look for image regions that correlate with a query answer and pinpoint representative features of a given question more accurately. + +More recently, in [9], a normalized variant of the self-attention mechanism, named normalized self-attention (NSA), is introduced. NSA seeks to encode and decode the image and caption features and normalize the distribution of internal activations during training. + +For the video question answering task, attention and memory mechanisms are used in [6] to efficiently learn visual features and the semantic correlations that precisely answer questions. + +# 2.3. Answer Grounding Methods + +The answer grounding task has also been studied under different names, such as grounded VQA and visual explanation for VQA. + +The VinVL [45] method is based on OSCAR [21] which proposes a novel vision-language pre-training that incorporates image anchor points as input, taken from an object detection network. VinVL improves the visual representations by using a bigger visual model pre-trained on larger training corpora that combine multiple annotated object detection datasets. + +LXMERT [36] presents a cross-modality framework for learning the connections between vision and language. The authors build a large-scale Transformer model that consists of three encoders: an object relationship encoder, a language encoder, and a cross-modality encoder. This model is then pre-trained on a large-scale dataset of image-and-sentence pairs. + +MTXNet [30] is an end-to-end trainable architecture that generates multimodal explanations focusing on the text in the image. It contains a Graph Attention Network, and each object location and OCR token is treated as a node in the graph. It also contains a multimodal transformer that operates on three modalities: question words, visual objects, and OCR tokens. + +U-CAM [29] is a method that uses gradient-based uncertainty estimates to provide visual attention maps. The certainty measure is improved by adding the certainty gradients to the existing standard Cross-Entropy loss gradients for training the model during back-propagation. + +# 3. Method + +Current attention methods have one or more of the following three problems: + +- The attention mechanism is not based on well-known and well-tested methods [32, 36, 29]. Well-established methods have been evaluated on many different tasks and datasets, especially, they perform well on large-scale datasets. Their design details have been actively studied, and their advantages and disadvantages are visible to the users. +- The complicated design makes it challenging to re-implement them on other platforms [32, 39, 30]. Recently, many applications have been made based on deep neural networks that are expected to be used by the end-users of different platforms. For example, answer grounding methods are helpful in assistive technologies for people with vision impairments [3]. In order to use such technologies offline, it is crucial to implement the method for each platform. + +![](images/48f385075ec7785d20d82c7fcab8e223b9a8fa1e5bc525fff6785427de4c33e7.jpg) +Figure 2. The complete proposed network. The multi-scale channel-wise image features contain different regions that are filtered by our Sentence Attention Block (SAB) based on the question and answer embeddings. + +- The custom design does not allow the usage of pretrained models [39, 45, 28, 27]. Answer grounding datasets are less diverse and have fewer samples than the classification datasets. Using pre-trained models can improve the accuracy and make the model more robust to noise and unseen samples. + +This paper proposes an attention block that avoids the mentioned problems. We design an attention block based on well-established and well-studied methods. It has a simple design to make it easy to re-implement and uses pre-trained models to achieve maximum accuracy. + +Our proposed method has three main parts: region proposal, sentence embedding, and attention fusion. + +# 3.1. Region Proposal + +This module's primary goal is to process the image and generate all the possible region candidates. Unlike Region Proposal Networks [31] which generate anchor-based bounding boxes, this module generates dense regions (segmentations). + +A pre-trained classification network is used as the back + +bone. However, the final classifier is removed to get the multi-scale feature-maps. During the end-to-end training of the proposed method, the backbone network learns all the candidate regions. Later, these candidate regions are filtered and combined by the proposed Sentence Attention Blocks. In Section 5.3, we show that the backbone network works as expected by visualizing the feature-maps. + +# 3.2. Sentence Embedding + +Since the input questions and answers have variable lengths, they are processed by a sentence embedding network. A pre-trained sentence encoder is employed to encode the given question and answer to a vector space separately. The result is two vectors, one for the question and one for the answer. The two vectors are concatenated and passed to the Sentence Attention Blocks as one vector. + +# 3.3. Attention Fusion + +Given multiple potential regions for an image, the answer grounding task reduces to choosing and/or combining these regions based on the given question and/or answer. The multi-scale channel-wise image features from the back + +
MethodMean IoU
MTXNet [30]18.9
LXMERT [36]24.9
VinVL [45]25.6
MAC-Caps [39]26.1
Ours29.0 (+2.9)
+ +bone network contain candidate regions. The goal is to filter and combine these regions based on the question and answer vector from the sentence embedding network. Towards this goal, we design an attention block based on the Squeeze-and-Excitation (SE) block [13]. + +We start with an SE block and make four modifications as follows: + +1. The first modification is changing it from self-attention to cross-attention. In other words, unlike a self-attention block, the proposed block pays attention to an external source: the encoded question and answer. +2. The second modification is adding a normalization layer. Empirically we found that adding a normalization layer improves the accuracy. Similar to [40], we use LayerNorm since it performs better on sequences. +3. The third modification is replacing the sigmoid function with softmax. Since the goal is to force the model to choose a few of the channel-wise image features, the sigmoid function is replaced with softmax. +4. The last modification is inverting the bottleneck. Instead of shrinking the embedding space, we expand the question and answer embeddings by a factor of two. + +In Section 5.1, the effect of each modification on the network accuracy is shown. Figure 1 depicts the final structure of the proposed block. Precisely, the proposed attention block is as follows: + +$$ +\operatorname {S o f t m a x} (F C (\operatorname {R e L U} (L N (F C (\mathbf {Q A})))) \times \mathbf {R} _ {i} \tag {2} +$$ + +where $FC$ is a single-layer perceptron, $LN$ is LayerNorm, QA is the joint question and answer embeddings, and $\mathbf{R}_i$ is the image regions at scale $i$ . For each scale, a separate instance of this attention block is created so that each block learns special weights for its corresponding feature-map scale. + +After computing the attention block for all the scales, all the outputs at different scales are combined into one output. Starting from the lowest resolution output, a $1 \times 1$ convolution is used to convert the number of channels to the number of channels of the next lowest resolution output. Then, Bilinear interpolation is used for upsampling the + +Table 1. Comparison results on the test set of the TextVQA-X dataset [30]. + +
MethodMean IoU
DeconvNet [7]29.8
Mask Aggregation [7]32.6
LXMERT [36]33.3
VinVL [45]33.9
MAC-Caps [39]34.3
Ours36.8 (+2.5)
+ +Table 2. Comparison results on the test set of the VQS dataset [7]. + +output feature-map by a factor of 2. Finally, the output is added to the next lowest resolution output. This process is repeated until no other output is left. In the end, a $1 \times 1$ convolution is used to reduce the number of channels of the final output to two. Therefore, the final output has two channels representing answer grounding and its background. We use Argmax to output a binary image. + +The complete design of the proposed method is depicted in Figure 2. + +# 3.4. Comparison to Existing Methods + +We compare the design of our proposed method to the top-performing existing methods for the answer grounding task. + +MAC-Caps [39] is one of the best-performing networks that currently exist. This network is based on MAC [15] which employs attention blocks in a recurrent reasoning architecture. MAC-Caps adds capsule layers to process the feature-maps further. Although similar to our method, MAC-Caps employs an attention mechanism. However, the recurrent design brings the usual problems of RNNs, such as gradient vanishing/exploding, longer training time, and not performing well on long sequences. Since our proposed method does not have a recurrent design, it avoids such problems. + +VinVL [45] is another top-performing network for the answer grounding task. This method is based on OSCAR [21] which incorporates image anchor points as inputs. VinVL improves the visual representations by using a bigger visual model pre-trained on larger training corpora that combine multiple annotated object detection datasets. This method relies on object detection networks to provide the anchor points, which adds an extra task for the network to learn. Also, using large generic corpora may not improve the accuracy for special datasets such as VizWiz-VQA-Grounding [3]. In contrast, our proposed method relies only on the feature-maps and does not define another task. + +Att-MCB [32] is the state-of-the-art network on the VQA-X dataset [27]. This method is designed to use a Knowledge Base. It comprises nine modules and four loss functions, making it a tedious design. This complicated design is resource-demanding and makes the training sensitive + +
MethodRank Correlation
PJ-X [27]0.342
CCM [28]0.368
U-CAM [29]0.372
VinVL [45]0.373
Att-MFH [32]0.376
MAC-Caps [39]0.389
Att-MCB [32]0.396
Ours0.421 (+0.025)
+ +to hyper-parameters. Our proposed method comprises three main modules and one loss function. This minimalistic design helps the gradient flow during back-propagation, and the training is less sensitive to hyper-parameters. + +# 4. Experiments + +We train and test our proposed method on four answer grounding datasets. We use the same setup for all of the experiments. + +# 4.1. Setup + +We use a pre-trained EfficientNet [37] to extract the multi-scale image features and we use a MiniLMv2 [42] sentence encoder network to encode the questions. In Section 5.2, we perform an ablation study on different backbones. we use BEiT [41] to generate answers for the datasets where the answers are not publicly available. We use RMI [47] with its default settings along with cross-entropy. Therefore, the overall loss function to minimize is: + +$$ +\begin{array}{l} \mathcal {L} _ {a l l} (y, p) = \frac {1}{B} \sum_ {b = 1} ^ {B} \lambda \mathcal {L} _ {c e} \left(y ^ {(b)}, p ^ {(b)}\right) \tag {3} \\ + (1 - \lambda) \mathcal {L} _ {r m i} (y ^ {(b)}, p ^ {(b)}) \\ \end{array} +$$ + +where $\lambda \in [0,1]$ is a weight factor, $B$ denotes the number of samples in a mini-batch, $\mathcal{L}_{ce}(y^{(b)},p^{(b)})$ is the standard cross entropy loss between the $b$ -th sample and its corresponding prediction, respectively, and $\mathcal{L}_{rmi}(y^{(b)},p^{(b)})$ is the RMI loss as follows: + +$$ +\mathcal {L} _ {r m i} (\mathbf {Y}, \mathbf {P}) = \sum_ {c = 1} ^ {C} \frac {1}{2 d} \operatorname {T r a c e} (\log (\mathbf {M})) \tag {4} +$$ + +and, + +$$ +\mathbf {M} = \boldsymbol {\Sigma} _ {\mathbf {Y}} - C o v (\mathbf {Y}, \mathbf {P}) \left(\boldsymbol {\Sigma} _ {\mathbf {P}} ^ {- 1}\right) ^ {T} C o v (\mathbf {Y}, \mathbf {P}) ^ {T} \tag {5} +$$ + +where $\mathbf{Y} = y^{(b)}$ , $\mathbf{P} = p^{(b)}$ , $\mathbf{M} \in \mathbb{R}^{d \times d}$ , $d$ denotes the number of pixels, $C$ is the number of object classes, here + +Table 3. Comparison results on the test set of the VQA-X dataset [27]. + +
TeamMean IoU
MindX66.9
MGTV69.7
hsslab_inspur70.1
GroundTruth70.3
Aurora70.6
SAB (Ours)72.4 (+1.8)
+ +Table 4. The top-performing teams on the VizWiz-VQA-Grounding Challenge 2022 leaderboard [3]. At the time of this writing, none of the methods have been published. + +
Method# of paramsinference time
LXMERT [36]153M56ms
VinVL [45]406M69ms
MAC-Caps [39]80M54ms
Ours88M49ms
+ +Table 5. Comparison of our proposed method to existing methods regarding the number of parameters and inference time. + +$C = 2$ for the answer grounding and its background, $\Sigma_{\mathbf{Y}}$ is the variance matrix of $\mathbf{Y}$ , and $Cov(\mathbf{Y},\mathbf{P})$ is the covariance matrix of $\mathbf{Y}$ and $\mathbf{P}$ . Similar to [47], we set $\lambda = 0.5$ . + +We use AdamW optimizer [25] with a weight decay of 0.05 and batch size of 16. We apply the "polynomial" learning rate policy with a poly exponent of 0.9 and an initial learning rate of 0.0001. Synchronized batch normalization is used across multiple GPUs. We use RandAugment [4] for data augmentation. + +# 4.2. TextVQA-X + +The TextVQA-X dataset [30] is a subset of the TextVQA dataset [33] such that the answer groundings are generated through a manual segmentation annotation process. TextVQA-X consists of 10,379 training images and 3,354 test images. Since only a few models have been evaluated on this dataset, we train three more models for a fair comparison. Similar to [3], we train top-performing methods that their code is publicly available; specifically, LXMERT [36], VinVL [45], and MAC-Caps [39]. We use the same setup as in [3] to train the three models with the difference that we train the models on the TextVQA-X training set. Similar to the existing methods, we use the test set to evaluate our method and report the mean IoU measure. + +# 4.3.VQS + +The VQS dataset [7] builds upon the images, instance segmentation masks, and bounding boxes in COCO [22] and the questions and answers in the VQA dataset [1]. VQS contains 26,995 training, 5,000 validation, and 5,873 test images. Similar to Section 4.2, we train top-performing methods that their code is publicly available; specifically, + +
BackboneMean IoU
ResNet-50x3 [19]71.9
Swin Transformer V2-B [24]72.8
EfficientNetV2-L [38]73.1
EfficientNet-B7 [37]73.5
RoBERTa [23]72.8
MPNet [34]73.1
MiniLMv2 [42]73.5
+ +Table 6. The ablation study of the pre-trained backbone networks. Top Section: Using a MiniLMv2 as a fixed sentence embedding network and replacing the image feature extraction networks. Bottom Section: Using a fixed image feature extraction network (EfficientNet-B7) and replacing the sentence embedding networks. + +LXMERT [36], VinVL [45], and MAC-Caps [39]. We use the same setup as in [3] to train the three models with the difference that we train the models on the VQS training set. Similar to the existing methods, we use the test set to evaluate our method and report the mean IoU measure. + +# 4.4.VQA-X + +The VQA-X dataset [27] does not have visual annotations for its training set. It has visual annotations only for its validation and test sets. Therefore, we train our method on the training set of the VQS dataset [7], which has related images to the VQA-X validation and test sets. The images of both datasets are from the COCO dataset [22]. The other methods are trained on different datasets, including Visual Genome [20] and VQA v2.0 [8]. The VQA-X dataset consists of 24,876 training, 1,431 validation, and 1,921 test images. Similar to the other methods, we use the test set of the VQA-X dataset to evaluate our method and report the Rank Correlation as in [27]. + +# 4.5. VizWiz-VQA-Grounding + +The VizWiz-VQA-Grounding dataset [3] is based on the VizWiz-VQA dataset [10], and the images and questions come from visually impaired people who shared them to request visual assistance in their day-to-day lives. VizWiz-VQA-Grounding contains a total of 9,998 VQAs divided into 6,494/1,131/2,373 VQAs for training, validation, and testing. The 2022 VizWiz-VQA-Grounding Challenge is designed around the aforementioned VizWiz-VQA-Grounding dataset. This challenge was held recently, and at the time of this writing, current methods in this challenge have not been published yet. Therefore, we evaluate our method using the VizWiz-VQA-Grounding Challenge online leaderboard [3]. The leaderboard evaluates all the methods against the VizWiz-VQA-Grounding test set. The ground truth labels for the test set are not published yet. + +
ModificationMean IoU
Switching to cross-attention70.1
Adding normalization layer71.2 (+1.1)
Switching to Softmax72.7 (+1.5)
Expansion of the embeddings73.5 (+0.8)
+ +Table 7. The results of our ablation study of the design modifications. The modifications are done cumulatively from top to bottom. + +# 4.6. Results + +We evaluate our proposed method on four answer grounding datasets. Tables 1, 2, and 3 compare the results of our proposed method with the existing methods on the TextVQA-X, VQS, and VQA-X datasets respectively. Our method achieves new state-of-the-art accuracy on three datasets. More specifically, improvements of 2.9, 2.5, and 0.025 on the TextVQA-X, VQS, and VQA-X datasets, respectively. + +Table 4 lists the top-performing teams on the VizWiz-VQA-Grounding Challenge 2022 leaderboard [3]. At the time of this writing, our method holds first place with considerably higher accuracy than the other methods. + +Furthermore, Table 5 compares our proposed method to existing methods regarding the number of parameters and inference time. The efficient design of our method makes it the fastest method. + +# 5. Ablation Studies + +In this section, we perform four ablation studies. All of the experiments are done on the VizWiz-VQA-Grounding [3] validation set and using the setup from Section 4.1 unless otherwise stated. + +# 5.1. Design Modifications + +We started the design of our sentence attention block from the standard SE block and made four modifications as explained in Section 3.3. In this section, we perform an ablation study on these modifications. Table 7 shows the impact of each modification on the final accuracy. These modifications are done cumulatively. The first modification is changing from self-attention to cross-attention, which makes our baseline for this ablation study. The second modification is adding a normalization layer. This modification improves our baseline by 1.1 percent. The third modification is replacing the Sigmoid function with Softmax. This modification has the highest impact on our method, 1.5 percent. And the fourth modification is the expansion of the embeddings. This last modification brings an addition of 0.8 percent. + +# 5.2. Backbone Networks + +The flexibility of our method enables us to use a variety of pre-trained networks. In this ablation study, we compare and analyze different pre-trained networks. To compare the image backbone networks, we keep the sentence embedding network fixed and change the image backbone network. Similarly, to compare the sentence embedding networks, we keep the image backbone network with the highest accuracy and change the sentence embedding network. We use well-known and well-performing networks whose pre-trained weights are publicly available. Specifically, we use EfficientNet [37], EfficientNetV2 [38], Swin Transformer V2 [24], and ResNet-50x3 [19] for image feature extraction and MiniLMv2 [42], MPNet [34], and RoBERTa [23] for sentence embedding. Table 6 shows the results of this ablation study. The highest accuracy is achieved by using EfficientNet-B7 and MiniLMv2. Since the EfficientNetV2-L reduces the input resolution more aggressively than EfficientNet-B7, it performs worse. Transformer networks usually perform well on large datasets. However, since VizWiz-VQA-Grounding is not a large dataset, Swin Transformer V2 does not achieve the highest accuracy. The ResNet-50x3 is pre-trained on ImageNet21k, but it achieves the least accuracy in this experiment. + +# 5.3. Region Proposals + +Feature-map visualization will provide insight into the internal representations of each layer in a network for a specific input image. + +This ablation study aims to show that the image backbone network has learned the potential regions for an image, which are then processed by the proposed sentence attention block. To this aim, we visualize the final output feature-maps of the image backbone network, before feeding them to an attention block. We use the image backbone of our fully trained network on the VizWiz dataset. Figure 3 (a) shows a few manually-chosen feature-maps for a sample image. These feature-maps show that after training, the backbone network has been able to learn the potential important regions of images. Having all the candidate regions, the task of the proposed attention block is to filter the irrelevant regions to the question and answer. Figure 3 (b) shows the final output of the complete network for the same image in (a), using different questions and answers (for brevity, the answers are not shown). This figure shows that our proposed block is able to filter the channel-wise feature-maps based on the input questions and answers. + +# 5.4. Multi-scale Fusions + +There is a trade-off in the answer grounding task such that some images are best handled at lower inference resolution and others better handled at higher inference resolution. Fine details, such as the edges of objects or thin + +![](images/0109e65c52ea08f2638a00f7618f631593e87b414e618aa912a13861ff1e46f3.jpg) +Figure 3. (a) Eight manually-selected sample feature-maps for an image. These feature-maps are the output of the feature extraction backbone network, before feeding them to the sentence attention block. The network has learned to output potential regions in different channels. (b) The final output of our complete network for the same image in (a), given different questions and answers (for brevity, the answers are not shown). This figure shows that our proposed attention block can filter the channel-wise feature-maps based on the input questions and answers. + +
ScaleMean IoU
1/3271.4
1/32 + 1/1672.5 (+1.1)
1/32 + 1/16 + 1/873.1 (+0.6)
1/32 + 1/16 + 1/8 + 1/473.5 (+0.4)
+ +Table 8. The gradual effect of multi-scale fusion. From top to bottom, the first row uses only the lowest resolution feature-map, and each row adds higher resolution feature-maps. Each scale is the ratio of the feature-map size to the input image. + +structures, are often better predicted with scaled-up images. At the same time, the prediction of large structures, which requires more global context, is often done better at scaled-down images because the network's receptive field can observe more of the necessary context. + +This ablation study shows the impact of adding the extra scales on the accuracy. Table 8 shows the results of this + +study. From top row to bottom, we start with only one scale and one attention block and gradually add higher resolution feature-maps. The lowest resolution is the final output feature-map of the backbone network. + +# 6. Conclusions + +In the study, a new architectural block termed the Sentence Attention Block was introduced to address certain challenges. This block recalibrates channel-wise image feature-maps by modeling inter-dependencies between image feature-maps and sentence embedding. The design began with a recognized attention method, and with minor adjustments, enhanced results were achieved, reaching top-tier accuracy. We showed the block's ability to filter out irrelevant feature-map channels based on sentence embedding. The method's adaptability allows for various pre-trained backbone networks, and its straightforward nature facilitates comprehension and re-implementation. The method's efficacy was demonstrated on the TextVQA-X, VQS, VQA-X, and VizWiz-VQA-Grounding datasets, and several ablation studies highlighted the validity of the design decisions. + +# References + +[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. +[2] Khaled Bayoudh, Raja Knani, Fayçal Hamdaoui, and Abdel-latif Mtibaa. A survey on deep multimodal learning for computer vision: advances, trends, applications, and datasets. The Visual Computer, 38(8):2939-2970, 2022. +[3] Chongyan Chen, Samreen Anjum, and Danna Gurari. Grounding answers for visual questions asked by visually impaired people. arXiv preprint arXiv:2202.01993, 2022. +[4] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. +[5] Chaorui Deng, Qi Wu, Qingyao Wu, Fuyuan Hu, Fan Lyu, and Mingkui Tan. Visual grounding via accumulated attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7746-7755, 2018. +[6] Chenyou Fan, Xiaofan Zhang, Shu Zhang, Wensheng Wang, Chi Zhang, and Heng Huang. Heterogeneous memory enhanced multimodal attention model for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1999-2007, 2019. +[7] Chuang Gan, Yandong Li, Haoxiang Li, Chen Sun, and Boqing Gong. Vqs: Linking segmentations to questions and answers for supervised attention in vqa and question-focused + +semantic segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1811-1820, 2017. +[8] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. +[9] Longteng Guo, Jing Liu, Xinxin Zhu, Peng Yao, Shichen Lu, and Hanqing Lu. Normalized and geometry-aware self-attention network for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10327–10336, 2020. +[10] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3608-3617, 2018. +[11] Chiori Hori, Takaaki Hori, Teng-Yok Lee, Ziming Zhang, Bret Harsham, John R Hershey, Tim K Marks, and Kazuhiko Sumi. Attention-based multimodal fusion for video description. In Proceedings of the IEEE international conference on computer vision, pages 4193-4202, 2017. +[12] Parisa Hosseini, Seyedalireza Khoshkirat, Mohammad Jalayer, Subasish Das, and Huaguo Zhou. Application of text mining techniques to identify actual wrong-way driving (wwd) crashes in police reports. International Journal of Transportation Science and Technology, 2022. +[13] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. +[14] Pingping Huang, Jianhui Huang, Yuqing Guo, Min Qiao, and Yong Zhu. Multi-grained attention with object-level grounding for visual question answering. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3595-3600, 2019. +[15] Drew A Hudson and Christopher D Manning. Compositional attention networks for machine reasoning. arXiv preprint arXiv:1803.03067, 2018. +[16] Seyedalireza Khoshsirat and Chandra Kambhamtu. Semantic segmentation using neural ordinary differential equations. Advances in Visual Computing: 17th International Symposium, ISVC 2022, San Diego, CA, USA, October 3-5, 2022, Proceedings, Part I, pages 284-295, 2022. +[17] Seyedalireza Khoshsurat and Chandra Kambhametu. Embedding attention blocks for the vizwiz answer grounding challenge. VizWiz Grand Challenge Workshop, 2023. +[18] Seyedalireza Khoshsirat and Chandra Kambhamtu. Empowering visually impaired individuals: A novel use of apple live photos and android motion photos. 25th Irish Machine Vision and Image Processing Conference, 2023. +[19] Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, and Neil Houlsby. Big transfer (bit): General visual representation learning. In European conference on computer vision, pages 491-507. Springer, 2020. + +[20] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. +[21] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020. +[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. +[23] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. +[24] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. arXiv preprint arXiv:2111.09883, 2021. +[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. +[26] Elham Maserat, Reza Safdari, Hamid Asadzadeh Aghdaei, Alireza Khoshsirat, and Mohammad Reza Zali. 43: Designing evidence based risk assessment system for cancer screening as an applicable approach for the estimating of treatment roadmap. BMJ Open, 7(Suppl 1):bmjopen-2016, 2017. +[27] Dong Huk Park, Lisa Anne Hendricks, Zeynep Akata, Anna Rohrbach, Bernt Schiele, Trevor Darrell, and Marcus Rohrbach. Multimodal explanations: Justifying decisions and pointing to the evidence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8779-8788, 2018. +[28] Badri Patro, Shivansh Patel, and Vinay Namboodiri. Robust explanations for visual question answering. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1577-1586, 2020. +[29] Badri N Patro, Mayank Lunayach, Shivansh Patel, and Vinay P Namboodiri. U-cam: Visual explanation using uncertainty based class activation maps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7444-7453, 2019. +[30] Varun Nagaraj Rao, Xingjian Zhen, Karen Hovsepian, and Mingwei Shen. A first look: Towards explainable textvqa models via visual and textual explanations. arXiv preprint arXiv:2105.02626, 2021. +[31] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. +[32] Felipe Riquelme, Alfredo De Goyeneche, Yundong Zhang, Juan Carlos Niebles, and Alvaro Soto. Explaining vqa pre + +dictions using visual grounding and a knowledge base. Image and Vision Computing, 101:103968, 2020. +[33] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019. +[34] Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. Mpnet: Masked and permuted pre-training for language understanding. Advances in Neural Information Processing Systems, 33:16857-16867, 2020. +[35] Jabeen Summaira, Xi Li, Amin Muhammad Shoib, Songyuan Li, and Jabbar Abdul. Recent advances and trends in multimodal deep learning: A review. arXiv preprint arXiv:2105.11087, 2021. +[36] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. +[37] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. +[38] Mingxing Tan and Quoc V Le. Efficientnetv2: Smaller models and faster training. arXiv preprint arXiv:2104.00298, 2021. +[39] Aisha Urooj, Hilde Kuehne, Kevin Duarte, Chuang Gan, Niels Lobo, and Mubarak Shah. Found a reason for me? weakly-supervised grounded visual question answering using capsules. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8465-8474, 2021. +[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. +[41] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. +[42] Wenhui Wang, Hangbo Bao, Shaohan Huang, Li Dong, and Furu Wei. Minilmv2: Multi-head self-attention relation distillation for compressing pretrained transformers. arXiv preprint arXiv:2012.15828, 2020. +[43] Xi Wei, Tianzhu Zhang, Yan Li, Yongdong Zhang, and Feng Wu. Multi-modality cross attention network for image and sentence matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10941-10950, 2020. +[44] Zichao Yang, Xiaodong He, Jianfeng Gao, Li Deng, and Alex Smola. Stacked attention networks for image question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 21-29, 2016. +[45] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on + +Computer Vision and Pattern Recognition, pages 5579-5588, 2021. +[46] Yundong Zhang, Juan Carlos Niebles, and Alvaro Soto. Interpretable visual question answering by visual grounding from attention supervision mining. In 2019 IEEE winter conference on applications of computer vision (wacv), pages 349-357. IEEE, 2019. +[47] Shuai Zhao, Yang Wang, Zheng Yang, and Deng Cai. Region mutual information loss for semantic segmentation. Advances in Neural Information Processing Systems, 32, 2019. \ No newline at end of file diff --git a/2023/Sentence Attention Blocks for Answer Grounding/images.zip b/2023/Sentence Attention Blocks for Answer Grounding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8e7422a68f58bedd99ea29b672ab592a96855684 --- /dev/null +++ b/2023/Sentence Attention Blocks for Answer Grounding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de05f263cfff9331d3de610eb4c13494ed8bcfef4363c8dbe23140f4fd691393 +size 293452 diff --git a/2023/Sentence Attention Blocks for Answer Grounding/layout.json b/2023/Sentence Attention Blocks for Answer Grounding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6d37e2f886997b861390ab814620b3421a119c6b --- /dev/null +++ b/2023/Sentence Attention Blocks for Answer Grounding/layout.json @@ -0,0 +1,7699 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 103, + 451, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 103, + 451, + 121 + ], + "spans": [ + { + "bbox": [ + 143, + 103, + 451, + 121 + ], + "type": "text", + "content": "Sentence Attention Blocks for Answer Grounding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 144, + 478, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 144, + 478, + 185 + ], + "spans": [ + { + "bbox": [ + 115, + 144, + 478, + 185 + ], + "type": "text", + "content": "Seyedalireza Khoshsirat Chandra Kambhametty \nVideo/Image Modeling and Synthesis (VIMS) Lab, University of Delaware \n{alireza, chandrak}@udel.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 245, + 289, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 245, + 289, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 289, + 558 + ], + "type": "text", + "content": "Answer grounding is the task of locating relevant visual evidence for the Visual Question Answering task. While a wide variety of attention methods have been introduced for this task, they suffer from the following three problems: designs that do not allow the usage of pre-trained networks and do not benefit from large data pre-training, custom designs that are not based on well-grounded previous designs, therefore limiting the learning power of the network, or complicated designs that make it challenging to re-implement or improve them. In this paper, we propose a novel architectural block, which we term Sentence Attention Block, to solve these problems. The proposed block re-calibrates channel-wise image feature-maps by explicitly modeling inter-dependencies between the image feature-maps and sentence embedding. We visually demonstrate how this block filters out irrelevant feature-maps channels based on sentence embedding. We start our design with a well-known attention method, and by making minor modifications, we improve the results to achieve state-of-the-art accuracy. The flexibility of our method makes it easy to use different pre-trained backbone networks, and its simplicity makes it easy to understand and be re-implemented. We demonstrate the effectiveness of our method on the TextVQA-X, VQS, VQA-X, and VizWiz-VQA-Grounding datasets. We perform multiple ablation studies to show the effectiveness of our design choices." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 628, + 198, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 198, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 198, + 641 + ], + "type": "text", + "content": "1.1. Visual Question Answering" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Visual Question Answering (VQA) systems try to accurately answer natural language questions regarding an input image [1]. This topic aims at developing systems that can communicate effectively about an image in natural language and comprehend the contents of images similar to humans." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 213, + 419, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 213, + 419, + 226 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 419, + 226 + ], + "type": "text", + "content": "1.2. Answer Grounding" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 232, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 547, + 411 + ], + "type": "text", + "content": "The answer grounding task is defined as detecting the pixels that can provide evidence for the answer to a given question regarding an image [3]. In other words, the task is to return the image regions used to arrive at the answer for a given visual question (question-image pair) with an answer. Although the VQA community has made significant progress, the best-performing systems are complicated black-box models, raising concerns about whether their answer reasoning is based on correct visual evidence. By understanding the reasoning mechanism of the model, we can evaluate the quality of answers, improve model performance, and provide explanations for end-users. To address this problem, answer grounding has been introduced into VQA systems, which requires the model to locate relevant image regions as well as answer visual questions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 411, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 590 + ], + "type": "text", + "content": "By providing answer groundings in response to visual questions, numerous applications become possible. First, they allow for evaluating whether a VQA model is reasoning correctly based on visual evidence. This is useful as an explanation as well as for assisting developers with model debugging. Second, answer grounding makes it possible to separate important regions from irrelevant background regions. Given that non-professional users can mistakenly have private information in the background of their pictures, answer grounding is a useful tool to obfuscate the background for privacy preserving. Third, suppose a service can magnify the relevant visual evidence. In that case, users will be able to discover the needed information in less time. This is useful in part because VQA answers can be insufficient sometimes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 598, + 457, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 457, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 457, + 612 + ], + "type": "text", + "content": "1.3. Multimodal Deep Learning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 653 + ], + "type": "text", + "content": "By definition, the VQA and answer grounding tasks are multimodal tasks since a method for these tasks should be able to process and correlate two different modalities." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Multimodal Joint-Embedding Models: These models merge and learn representations of multiple modalities in a joint feature space. Joint-embeddings underpin the building of a lot of cross-modal methods as they can bridge the gap between different modalities. In this joint space, the dis" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "6080" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "tance of different points is equivalent to the semantic distance between their corresponding original inputs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 180 + ], + "type": "text", + "content": "Multimodal Attention-based Models: The main objective of the attention mechanism is to design systems that use local features of image or text, extract features from different regions, and assign priorities to them. The attention portion in an image determines salient regions. Then the language generation component focuses more on those salient regions for additional processing [35, 2, 12, 16, 26]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 186, + 171, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 186, + 171, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 171, + 198 + ], + "type": "text", + "content": "1.4. Attention Mechanism" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 205, + 286, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 286, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 286, + 370 + ], + "type": "text", + "content": "In deep learning, attention is a mechanism that mimics cognitive attention. The goal is to enhance the essential features of the input data and vanish out the rest. Attention methods can be classified into two classes based on their inputs: self-attention and cross-attention. Self-attention (also known as intra-attention) is a type of attention that quantifies the interdependence within the elements of a single input. At the same time, cross-attention (also known as inter-attention) finds the interdependence across two or more inputs [40, 18, 17]. Usually, cross-attention methods are used for multimodal inputs [43]. Cross-attention models first process individual modalities using modality-specific encoders, then the encoded features are fed into cross-attention modules." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 372, + 286, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 286, + 442 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 286, + 442 + ], + "type": "text", + "content": "The Squeeze-and-Excitation method [13] is a channelwise self-attention mechanism widely used in classification networks [37, 38]. It consists of a global average pooling of the input, followed by two linear layers with an interleaved non-linearity and a sigmoid function. Concretely, the output of this method is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 450, + 286, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 450, + 286, + 463 + ], + "spans": [ + { + "bbox": [ + 74, + 450, + 286, + 463 + ], + "type": "interline_equation", + "content": "\\sigma \\left(F C (R E L U (F C (g _ {-} a v g _ {-} p o o l (\\mathbf {X})))\\right) \\times \\mathbf {X} \\tag {1}", + "image_path": "7055026f4c484346a1c09ea237711ad011b46d4cab1f41c6464e83f933818196.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 468, + 286, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 286, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 286, + 658 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 468, + 286, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 468, + 286, + 658 + ], + "type": "text", + "content": " is the feature-maps. This module aims to dynamically focus on more important channels, essentially increasing the importance of specific channels over others. This is accomplished by scaling the more important channels by a higher value. While many feature descriptors exist to reduce the spatial dimensions of the feature maps to a singular value, this module uses average pooling to keep the required computation low. The next part of the module maps the scaling weights using a Multi-Layer Perceptron (MLP) with a bottleneck structure. The result values are scaled to a range of 0-1 by passing them through a sigmoid activation layer. Afterward, using a common broadcasted element-wise multiplication, the output is applied directly to the input. This multiplication scales each input feature-maps channel with its corresponding weight learned from the MLP in the Excitation module." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 661, + 266, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 661, + 266, + 672 + ], + "spans": [ + { + "bbox": [ + 59, + 661, + 266, + 672 + ], + "type": "text", + "content": "The following is the summary of our contributions:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 677, + 286, + 713 + ], + "type": "text", + "content": "- We present a novel attention module based on the Squeeze-and-Eccitation method for the answer grounding task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 72, + 544, + 159 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 317, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 544, + 95 + ], + "type": "text", + "content": "- We evaluate our method on common datasets and show that it achieves new state-of-the-art results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 105, + 544, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 105, + 544, + 126 + ], + "spans": [ + { + "bbox": [ + 317, + 105, + 544, + 126 + ], + "type": "text", + "content": "- We compare our design with the top-performing networks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 137, + 544, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 137, + 544, + 159 + ], + "spans": [ + { + "bbox": [ + 317, + 137, + 544, + 159 + ], + "type": "text", + "content": "- We perform multiple ablation studies to learn more about this network." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 172, + 392, + 184 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 172, + 392, + 184 + ], + "spans": [ + { + "bbox": [ + 306, + 172, + 392, + 184 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 192, + 526, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 192, + 526, + 205 + ], + "spans": [ + { + "bbox": [ + 306, + 192, + 526, + 205 + ], + "type": "text", + "content": "2.1. Attention Methods for Answer Grounding" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 211, + 544, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 211, + 544, + 234 + ], + "spans": [ + { + "bbox": [ + 306, + 211, + 544, + 234 + ], + "type": "text", + "content": "Models based on different attention methods have been actively explored for the answer grounding task." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 235, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 235, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 305, + 235, + 545, + 317 + ], + "type": "text", + "content": "MAC-Caps [39] is based on MAC [15] which has a recurrent reasoning architecture that performs " + }, + { + "bbox": [ + 305, + 235, + 545, + 317 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 235, + 545, + 317 + ], + "type": "text", + "content": " reasoning steps to answer the question. At each reasoning step, MAC uses an attention block to read from image features and writing memory. MAC-Caps adds capsule layers on top of the convolutional layers to obtain visual capsules from the feature-maps." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 319, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 545, + 389 + ], + "type": "text", + "content": "Att-MCB [32] is designed to use a Knowledge Base. It is composed of nine modules, out of which two modules use an attention mechanism. One module uses an attention mechanism to generate visual explanations, and the other one to consume and point out relevant information from the Knowledge Base." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 390, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 390, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 390, + 545, + 544 + ], + "type": "text", + "content": "A multi-grained attention method is introduced in [14]. It consists of two types of object-level groundings to explore fine-grained information, and a more sophisticated language model for better question representation. A Word-Label Matching attention vector that indicates the weight that should be given to each of the " + }, + { + "bbox": [ + 305, + 390, + 545, + 544 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 390, + 545, + 544 + ], + "type": "text", + "content": " objects in the image is computed in terms of the semantic similarity between the category labels of the objects and the words in the question. A word-object matching module is exploited to evaluate how likely a question word matches a visual object. A Sentence-Object attention module is used to capture the global semantics of the whole sentence to guide the focus on relevant objects." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 546, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 546, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 305, + 546, + 545, + 653 + ], + "type": "text", + "content": "Multiple custom-designed attention modules are used in [46] to generate an attention map for the given image and question pair. Questions are tokenized and passed through an embedding layer followed by an LSTM layer to generate the question features. An Image Attention Supervision Module is used as an auxiliary classification task; that is, the ground-truth visual grounding labels are used to guide the model to focus on significant parts of the image to answer each question." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 712 + ], + "type": "text", + "content": "A so-called \"accumulated attention\" mechanism is introduced in [5]. It consists of three attention modules, one for the input text, one for the objects, and one for the whole image. All three modules have a similar structure of a linear layer followed by a Softmax function." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "6081" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 69, + 276, + 326 + ], + "blocks": [ + { + "bbox": [ + 60, + 69, + 276, + 326 + ], + "lines": [ + { + "bbox": [ + 60, + 69, + 276, + 326 + ], + "spans": [ + { + "bbox": [ + 60, + 69, + 276, + 326 + ], + "type": "image", + "image_path": "e00e29fb6741d093fa214e0c26ad6998080f825a199cb239361181c925f00739.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 338, + 288, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 338, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 288, + 384 + ], + "type": "text", + "content": "Figure 1. Our proposed sentence attention block. " + }, + { + "bbox": [ + 46, + 338, + 288, + 384 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 338, + 288, + 384 + ], + "type": "text", + "content": " denotes the number of channels in image feature-maps, and " + }, + { + "bbox": [ + 46, + 338, + 288, + 384 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 338, + 288, + 384 + ], + "type": "text", + "content": " represents the embedding size of the sentence encoder network. Question and answer embeddings are concatenated into one vector." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 411, + 217, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 217, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 217, + 423 + ], + "type": "text", + "content": "2.2. Multimodal Attention Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 434, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 287, + 471 + ], + "type": "text", + "content": "Since its introduction, the attention mechanism has been widely adopted in the computer vision community due to its special capabilities for many multimodal applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 473, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 545 + ], + "type": "text", + "content": "In [11], a multimodal attention model is proposed based on the encoder-decoder networks and RNNs for video captioning and sentence generation. In particular, the multimodal attention framework combined image, audio, and motion features by selecting each modality's most relevant context vector." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 549, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 549, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 549, + 287, + 598 + ], + "type": "text", + "content": "In [44], utilizing stacked attention networks is suggested to look for image regions that correlate with a query answer and pinpoint representative features of a given question more accurately." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 601, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 287, + 662 + ], + "type": "text", + "content": "More recently, in [9], a normalized variant of the self-attention mechanism, named normalized self-attention (NSA), is introduced. NSA seeks to encode and decode the image and caption features and normalize the distribution of internal activations during training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "For the video question answering task, attention and memory mechanisms are used in [6] to efficiently learn visual features and the semantic correlations that precisely answer questions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 463, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 463, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 463, + 85 + ], + "type": "text", + "content": "2.3. Answer Grounding Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 91, + 545, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 545, + 126 + ], + "type": "text", + "content": "The answer grounding task has also been studied under different names, such as grounded VQA and visual explanation for VQA." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 128, + 546, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 128, + 546, + 211 + ], + "spans": [ + { + "bbox": [ + 305, + 128, + 546, + 211 + ], + "type": "text", + "content": "The VinVL [45] method is based on OSCAR [21] which proposes a novel vision-language pre-training that incorporates image anchor points as input, taken from an object detection network. VinVL improves the visual representations by using a bigger visual model pre-trained on larger training corpora that combine multiple annotated object detection datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 213, + 545, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 213, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 545, + 297 + ], + "type": "text", + "content": "LXMERT [36] presents a cross-modality framework for learning the connections between vision and language. The authors build a large-scale Transformer model that consists of three encoders: an object relationship encoder, a language encoder, and a cross-modality encoder. This model is then pre-trained on a large-scale dataset of image-and-sentence pairs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 297, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 297, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 545, + 380 + ], + "type": "text", + "content": "MTXNet [30] is an end-to-end trainable architecture that generates multimodal explanations focusing on the text in the image. It contains a Graph Attention Network, and each object location and OCR token is treated as a node in the graph. It also contains a multimodal transformer that operates on three modalities: question words, visual objects, and OCR tokens." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 381, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 381, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 381, + 545, + 443 + ], + "type": "text", + "content": "U-CAM [29] is a method that uses gradient-based uncertainty estimates to provide visual attention maps. The certainty measure is improved by adding the certainty gradients to the existing standard Cross-Entropy loss gradients for training the model during back-propagation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 455, + 361, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 361, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 361, + 468 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 475, + 545, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 475, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 306, + 475, + 545, + 500 + ], + "type": "text", + "content": "Current attention methods have one or more of the following three problems:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 510, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 317, + 510, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 510, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 317, + 510, + 545, + 594 + ], + "type": "text", + "content": "- The attention mechanism is not based on well-known and well-tested methods [32, 36, 29]. Well-established methods have been evaluated on many different tasks and datasets, especially, they perform well on large-scale datasets. Their design details have been actively studied, and their advantages and disadvantages are visible to the users." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 605, + 545, + 714 + ], + "type": "text", + "content": "- The complicated design makes it challenging to re-implement them on other platforms [32, 39, 30]. Recently, many applications have been made based on deep neural networks that are expected to be used by the end-users of different platforms. For example, answer grounding methods are helpful in assistive technologies for people with vision impairments [3]. In order to use such technologies offline, it is crucial to implement the method for each platform." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6082" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 75, + 521, + 390 + ], + "blocks": [ + { + "bbox": [ + 72, + 75, + 521, + 390 + ], + "lines": [ + { + "bbox": [ + 72, + 75, + 521, + 390 + ], + "spans": [ + { + "bbox": [ + 72, + 75, + 521, + 390 + ], + "type": "image", + "image_path": "48f385075ec7785d20d82c7fcab8e223b9a8fa1e5bc525fff6785427de4c33e7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 404, + 547, + 427 + ], + "lines": [ + { + "bbox": [ + 46, + 404, + 547, + 427 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 547, + 427 + ], + "type": "text", + "content": "Figure 2. The complete proposed network. The multi-scale channel-wise image features contain different regions that are filtered by our Sentence Attention Block (SAB) based on the question and answer embeddings." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 446, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 446, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 59, + 446, + 287, + 518 + ], + "type": "text", + "content": "- The custom design does not allow the usage of pretrained models [39, 45, 28, 27]. Answer grounding datasets are less diverse and have fewer samples than the classification datasets. Using pre-trained models can improve the accuracy and make the model more robust to noise and unseen samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 528, + 287, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 287, + 588 + ], + "type": "text", + "content": "This paper proposes an attention block that avoids the mentioned problems. We design an attention block based on well-established and well-studied methods. It has a simple design to make it easy to re-implement and uses pre-trained models to achieve maximum accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 589, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 589, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 589, + 287, + 613 + ], + "type": "text", + "content": "Our proposed method has three main parts: region proposal, sentence embedding, and attention fusion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 621, + 147, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 147, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 147, + 635 + ], + "type": "text", + "content": "3.1. Region Proposal" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 700 + ], + "type": "text", + "content": "This module's primary goal is to process the image and generate all the possible region candidates. Unlike Region Proposal Networks [31] which generate anchor-based bounding boxes, this module generates dense regions (segmentations)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "A pre-trained classification network is used as the back" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 446, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 545, + 530 + ], + "type": "text", + "content": "bone. However, the final classifier is removed to get the multi-scale feature-maps. During the end-to-end training of the proposed method, the backbone network learns all the candidate regions. Later, these candidate regions are filtered and combined by the proposed Sentence Attention Blocks. In Section 5.3, we show that the backbone network works as expected by visualizing the feature-maps." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 537, + 427, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 427, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 427, + 550 + ], + "type": "text", + "content": "3.2. Sentence Embedding" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "text", + "content": "Since the input questions and answers have variable lengths, they are processed by a sentence embedding network. A pre-trained sentence encoder is employed to encode the given question and answer to a vector space separately. The result is two vectors, one for the question and one for the answer. The two vectors are concatenated and passed to the Sentence Attention Blocks as one vector." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 647, + 408, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 408, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 408, + 658 + ], + "type": "text", + "content": "3.3. Attention Fusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "Given multiple potential regions for an image, the answer grounding task reduces to choosing and/or combining these regions based on the given question and/or answer. The multi-scale channel-wise image features from the back" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6083" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 70, + 264, + 145 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 264, + 145 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 264, + 145 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 264, + 145 + ], + "type": "table", + "html": "
MethodMean IoU
MTXNet [30]18.9
LXMERT [36]24.9
VinVL [45]25.6
MAC-Caps [39]26.1
Ours29.0 (+2.9)
", + "image_path": "ad0210b1ec08437248686fadf751d380bf1df21b82344d9d684e0bb320428ee5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 193, + 286, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 193, + 286, + 252 + ], + "spans": [ + { + "bbox": [ + 47, + 193, + 286, + 252 + ], + "type": "text", + "content": "bone network contain candidate regions. The goal is to filter and combine these regions based on the question and answer vector from the sentence embedding network. Towards this goal, we design an attention block based on the Squeeze-and-Excitation (SE) block [13]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 253, + 286, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 286, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 286, + 277 + ], + "type": "text", + "content": "We start with an SE block and make four modifications as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 284, + 286, + 485 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 284, + 286, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 286, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 286, + 332 + ], + "type": "text", + "content": "1. The first modification is changing it from self-attention to cross-attention. In other words, unlike a self-attention block, the proposed block pays attention to an external source: the encoded question and answer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 339, + 286, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 339, + 286, + 388 + ], + "spans": [ + { + "bbox": [ + 54, + 339, + 286, + 388 + ], + "type": "text", + "content": "2. The second modification is adding a normalization layer. Empirically we found that adding a normalization layer improves the accuracy. Similar to [40], we use LayerNorm since it performs better on sequences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 395, + 286, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 395, + 286, + 442 + ], + "spans": [ + { + "bbox": [ + 54, + 395, + 286, + 442 + ], + "type": "text", + "content": "3. The third modification is replacing the sigmoid function with softmax. Since the goal is to force the model to choose a few of the channel-wise image features, the sigmoid function is replaced with softmax." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 450, + 286, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 450, + 286, + 485 + ], + "spans": [ + { + "bbox": [ + 54, + 450, + 286, + 485 + ], + "type": "text", + "content": "4. The last modification is inverting the bottleneck. Instead of shrinking the embedding space, we expand the question and answer embeddings by a factor of two." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 493, + 286, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 493, + 286, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 493, + 286, + 539 + ], + "type": "text", + "content": "In Section 5.1, the effect of each modification on the network accuracy is shown. Figure 1 depicts the final structure of the proposed block. Precisely, the proposed attention block is as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 548, + 286, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 286, + 562 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 286, + 562 + ], + "type": "interline_equation", + "content": "\\operatorname {S o f t m a x} (F C (\\operatorname {R e L U} (L N (F C (\\mathbf {Q A})))) \\times \\mathbf {R} _ {i} \\tag {2}", + "image_path": "57803364bc432da441fcfddb9967b2e44c69ed5e72a939f72c35a22fabc3fbb1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 570, + 286, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 286, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 570, + 286, + 640 + ], + "type": "inline_equation", + "content": "FC" + }, + { + "bbox": [ + 47, + 570, + 286, + 640 + ], + "type": "text", + "content": " is a single-layer perceptron, " + }, + { + "bbox": [ + 47, + 570, + 286, + 640 + ], + "type": "inline_equation", + "content": "LN" + }, + { + "bbox": [ + 47, + 570, + 286, + 640 + ], + "type": "text", + "content": " is LayerNorm, QA is the joint question and answer embeddings, and " + }, + { + "bbox": [ + 47, + 570, + 286, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i" + }, + { + "bbox": [ + 47, + 570, + 286, + 640 + ], + "type": "text", + "content": " is the image regions at scale " + }, + { + "bbox": [ + 47, + 570, + 286, + 640 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 570, + 286, + 640 + ], + "type": "text", + "content": ". For each scale, a separate instance of this attention block is created so that each block learns special weights for its corresponding feature-map scale." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "text", + "content": "After computing the attention block for all the scales, all the outputs at different scales are combined into one output. Starting from the lowest resolution output, a " + }, + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "text", + "content": " convolution is used to convert the number of channels to the number of channels of the next lowest resolution output. Then, Bilinear interpolation is used for upsampling the" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 318, + 70, + 533, + 157 + ], + "blocks": [ + { + "bbox": [ + 47, + 152, + 287, + 175 + ], + "lines": [ + { + "bbox": [ + 47, + 152, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 287, + 175 + ], + "type": "text", + "content": "Table 1. Comparison results on the test set of the TextVQA-X dataset [30]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 70, + 533, + 157 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 533, + 157 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 533, + 157 + ], + "type": "table", + "html": "
MethodMean IoU
DeconvNet [7]29.8
Mask Aggregation [7]32.6
LXMERT [36]33.3
VinVL [45]33.9
MAC-Caps [39]34.3
Ours36.8 (+2.5)
", + "image_path": "a4e141adbd6587889c3cc516aec2ec8afdc3a315d10b760fed85e273c2fad022.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 164, + 544, + 176 + ], + "lines": [ + { + "bbox": [ + 306, + 164, + 544, + 176 + ], + "spans": [ + { + "bbox": [ + 306, + 164, + 544, + 176 + ], + "type": "text", + "content": "Table 2. Comparison results on the test set of the VQS dataset [7]." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 196, + 545, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 196, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 305, + 196, + 545, + 279 + ], + "type": "text", + "content": "output feature-map by a factor of 2. Finally, the output is added to the next lowest resolution output. This process is repeated until no other output is left. In the end, a " + }, + { + "bbox": [ + 305, + 196, + 545, + 279 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 305, + 196, + 545, + 279 + ], + "type": "text", + "content": " convolution is used to reduce the number of channels of the final output to two. Therefore, the final output has two channels representing answer grounding and its background. We use Argmax to output a binary image." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 280, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 280, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 280, + 545, + 304 + ], + "type": "text", + "content": "The complete design of the proposed method is depicted in Figure 2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 312, + 482, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 312, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 482, + 324 + ], + "type": "text", + "content": "3.4. Comparison to Existing Methods" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 331, + 545, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 331, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 545, + 364 + ], + "type": "text", + "content": "We compare the design of our proposed method to the top-performing existing methods for the answer grounding task." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 366, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 366, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 366, + 545, + 498 + ], + "type": "text", + "content": "MAC-Caps [39] is one of the best-performing networks that currently exist. This network is based on MAC [15] which employs attention blocks in a recurrent reasoning architecture. MAC-Caps adds capsule layers to process the feature-maps further. Although similar to our method, MAC-Caps employs an attention mechanism. However, the recurrent design brings the usual problems of RNNs, such as gradient vanishing/exploding, longer training time, and not performing well on long sequences. Since our proposed method does not have a recurrent design, it avoids such problems." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 498, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 498, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 305, + 498, + 545, + 652 + ], + "type": "text", + "content": "VinVL [45] is another top-performing network for the answer grounding task. This method is based on OSCAR [21] which incorporates image anchor points as inputs. VinVL improves the visual representations by using a bigger visual model pre-trained on larger training corpora that combine multiple annotated object detection datasets. This method relies on object detection networks to provide the anchor points, which adds an extra task for the network to learn. Also, using large generic corpora may not improve the accuracy for special datasets such as VizWiz-VQA-Grounding [3]. In contrast, our proposed method relies only on the feature-maps and does not define another task." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": "Att-MCB [32] is the state-of-the-art network on the VQA-X dataset [27]. This method is designed to use a Knowledge Base. It comprises nine modules and four loss functions, making it a tedious design. This complicated design is resource-demanding and makes the training sensitive" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6084" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 70, + 265, + 180 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 265, + 180 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 265, + 180 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 265, + 180 + ], + "type": "table", + "html": "
MethodRank Correlation
PJ-X [27]0.342
CCM [28]0.368
U-CAM [29]0.372
VinVL [45]0.373
Att-MFH [32]0.376
MAC-Caps [39]0.389
Att-MCB [32]0.396
Ours0.421 (+0.025)
", + "image_path": "2250720e73dcb5c1a70a567c34b8d9b90d8a1959b2a8d124053acdab1d7cf28b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 229, + 286, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 229, + 286, + 278 + ], + "spans": [ + { + "bbox": [ + 47, + 229, + 286, + 278 + ], + "type": "text", + "content": "to hyper-parameters. Our proposed method comprises three main modules and one loss function. This minimalistic design helps the gradient flow during back-propagation, and the training is less sensitive to hyper-parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 288, + 128, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 288, + 128, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 288, + 128, + 301 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 308, + 287, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 287, + 344 + ], + "type": "text", + "content": "We train and test our proposed method on four answer grounding datasets. We use the same setup for all of the experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 351, + 97, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 351, + 97, + 364 + ], + "spans": [ + { + "bbox": [ + 47, + 351, + 97, + 364 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "text", + "content": "We use a pre-trained EfficientNet [37] to extract the multi-scale image features and we use a MiniLMv2 [42] sentence encoder network to encode the questions. In Section 5.2, we perform an ablation study on different backbones. we use BEiT [41] to generate answers for the datasets where the answers are not publicly available. We use RMI [47] with its default settings along with cross-entropy. Therefore, the overall loss function to minimize is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 91, + 483, + 287, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 483, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 91, + 483, + 287, + 533 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {a l l} (y, p) = \\frac {1}{B} \\sum_ {b = 1} ^ {B} \\lambda \\mathcal {L} _ {c e} \\left(y ^ {(b)}, p ^ {(b)}\\right) \\tag {3} \\\\ + (1 - \\lambda) \\mathcal {L} _ {r m i} (y ^ {(b)}, p ^ {(b)}) \\\\ \\end{array}", + "image_path": "94b4e1abcb10737f517c368459adda7f1595e7da529d374d5bced4f0c8452fe5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\lambda \\in [0,1]" + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "text", + "content": " is a weight factor, " + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "text", + "content": " denotes the number of samples in a mini-batch, " + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}(y^{(b)},p^{(b)})" + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "text", + "content": " is the standard cross entropy loss between the " + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "text", + "content": "-th sample and its corresponding prediction, respectively, and " + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rmi}(y^{(b)},p^{(b)})" + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "text", + "content": " is the RMI loss as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 86, + 607, + 287, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 607, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 86, + 607, + 287, + 639 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r m i} (\\mathbf {Y}, \\mathbf {P}) = \\sum_ {c = 1} ^ {C} \\frac {1}{2 d} \\operatorname {T r a c e} (\\log (\\mathbf {M})) \\tag {4}", + "image_path": "b658867b18d3f55b53c8aadd824a3f5f2c351edfd9ede8aff6415eae2b499852.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 647, + 67, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 67, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 67, + 658 + ], + "type": "text", + "content": "and," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 72, + 666, + 287, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 666, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 72, + 666, + 287, + 681 + ], + "type": "interline_equation", + "content": "\\mathbf {M} = \\boldsymbol {\\Sigma} _ {\\mathbf {Y}} - C o v (\\mathbf {Y}, \\mathbf {P}) \\left(\\boldsymbol {\\Sigma} _ {\\mathbf {P}} ^ {- 1}\\right) ^ {T} C o v (\\mathbf {Y}, \\mathbf {P}) ^ {T} \\tag {5}", + "image_path": "b8a25ed399ff6a7135984fe60063633888e8069193c5627af3166d43beb59e55.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} = y^{(b)}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = p^{(b)}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " denotes the number of pixels, " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " is the number of object classes, here" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 346, + 69, + 506, + 157 + ], + "blocks": [ + { + "bbox": [ + 47, + 189, + 287, + 211 + ], + "lines": [ + { + "bbox": [ + 47, + 189, + 287, + 211 + ], + "spans": [ + { + "bbox": [ + 47, + 189, + 287, + 211 + ], + "type": "text", + "content": "Table 3. Comparison results on the test set of the VQA-X dataset [27]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 346, + 69, + 506, + 157 + ], + "lines": [ + { + "bbox": [ + 346, + 69, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 346, + 69, + 506, + 157 + ], + "type": "table", + "html": "
TeamMean IoU
MindX66.9
MGTV69.7
hsslab_inspur70.1
GroundTruth70.3
Aurora70.6
SAB (Ours)72.4 (+1.8)
", + "image_path": "e6d2db0b111815b5d28472fbd53120a991a50b0a886100574e90e771cec73cdc.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 310, + 209, + 542, + 272 + ], + "blocks": [ + { + "bbox": [ + 305, + 164, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 305, + 164, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 305, + 164, + 545, + 198 + ], + "type": "text", + "content": "Table 4. The top-performing teams on the VizWiz-VQA-Grounding Challenge 2022 leaderboard [3]. At the time of this writing, none of the methods have been published." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 209, + 542, + 272 + ], + "lines": [ + { + "bbox": [ + 310, + 209, + 542, + 272 + ], + "spans": [ + { + "bbox": [ + 310, + 209, + 542, + 272 + ], + "type": "table", + "html": "
Method# of paramsinference time
LXMERT [36]153M56ms
VinVL [45]406M69ms
MAC-Caps [39]80M54ms
Ours88M49ms
", + "image_path": "412c2bc294ca3bb77ae20c00c6cc08a290d7e5772cabb44f197ae2a7fa9b0a53.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 279, + 545, + 301 + ], + "lines": [ + { + "bbox": [ + 305, + 279, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 305, + 279, + 545, + 301 + ], + "type": "text", + "content": "Table 5. Comparison of our proposed method to existing methods regarding the number of parameters and inference time." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 322, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 322, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 322, + 545, + 358 + ], + "type": "inline_equation", + "content": "C = 2" + }, + { + "bbox": [ + 305, + 322, + 545, + 358 + ], + "type": "text", + "content": " for the answer grounding and its background, " + }, + { + "bbox": [ + 305, + 322, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\mathbf{Y}}" + }, + { + "bbox": [ + 305, + 322, + 545, + 358 + ], + "type": "text", + "content": " is the variance matrix of " + }, + { + "bbox": [ + 305, + 322, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 305, + 322, + 545, + 358 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 322, + 545, + 358 + ], + "type": "inline_equation", + "content": "Cov(\\mathbf{Y},\\mathbf{P})" + }, + { + "bbox": [ + 305, + 322, + 545, + 358 + ], + "type": "text", + "content": " is the covariance matrix of " + }, + { + "bbox": [ + 305, + 322, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 305, + 322, + 545, + 358 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 322, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 305, + 322, + 545, + 358 + ], + "type": "text", + "content": ". Similar to [47], we set " + }, + { + "bbox": [ + 305, + 322, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\lambda = 0.5" + }, + { + "bbox": [ + 305, + 322, + 545, + 358 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 358, + 546, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 546, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 546, + 430 + ], + "type": "text", + "content": "We use AdamW optimizer [25] with a weight decay of 0.05 and batch size of 16. We apply the \"polynomial\" learning rate policy with a poly exponent of 0.9 and an initial learning rate of 0.0001. Synchronized batch normalization is used across multiple GPUs. We use RandAugment [4] for data augmentation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 439, + 385, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 439, + 385, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 385, + 452 + ], + "type": "text", + "content": "4.2. TextVQA-X" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 458, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 614 + ], + "type": "text", + "content": "The TextVQA-X dataset [30] is a subset of the TextVQA dataset [33] such that the answer groundings are generated through a manual segmentation annotation process. TextVQA-X consists of 10,379 training images and 3,354 test images. Since only a few models have been evaluated on this dataset, we train three more models for a fair comparison. Similar to [3], we train top-performing methods that their code is publicly available; specifically, LXMERT [36], VinVL [45], and MAC-Caps [39]. We use the same setup as in [3] to train the three models with the difference that we train the models on the TextVQA-X training set. Similar to the existing methods, we use the test set to evaluate our method and report the mean IoU measure." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 622, + 351, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 622, + 351, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 622, + 351, + 635 + ], + "type": "text", + "content": "4.3.VQS" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "The VQS dataset [7] builds upon the images, instance segmentation masks, and bounding boxes in COCO [22] and the questions and answers in the VQA dataset [1]. VQS contains 26,995 training, 5,000 validation, and 5,873 test images. Similar to Section 4.2, we train top-performing methods that their code is publicly available; specifically," + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6085" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 267, + 169 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 267, + 169 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 267, + 169 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 267, + 169 + ], + "type": "table", + "html": "
BackboneMean IoU
ResNet-50x3 [19]71.9
Swin Transformer V2-B [24]72.8
EfficientNetV2-L [38]73.1
EfficientNet-B7 [37]73.5
RoBERTa [23]72.8
MPNet [34]73.1
MiniLMv2 [42]73.5
", + "image_path": "a16446120986a7091ff6d1e8d801ef17e1a6704d47a2eadcaf1f70b4b28af282.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 177, + 287, + 242 + ], + "lines": [ + { + "bbox": [ + 46, + 177, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 287, + 242 + ], + "type": "text", + "content": "Table 6. The ablation study of the pre-trained backbone networks. Top Section: Using a MiniLMv2 as a fixed sentence embedding network and replacing the image feature extraction networks. Bottom Section: Using a fixed image feature extraction network (EfficientNet-B7) and replacing the sentence embedding networks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 264, + 287, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 287, + 324 + ], + "type": "text", + "content": "LXMERT [36], VinVL [45], and MAC-Caps [39]. We use the same setup as in [3] to train the three models with the difference that we train the models on the VQS training set. Similar to the existing methods, we use the test set to evaluate our method and report the mean IoU measure." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 337, + 105, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 337, + 105, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 337, + 105, + 350 + ], + "type": "text", + "content": "4.4.VQA-X" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 357, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 287, + 501 + ], + "type": "text", + "content": "The VQA-X dataset [27] does not have visual annotations for its training set. It has visual annotations only for its validation and test sets. Therefore, we train our method on the training set of the VQS dataset [7], which has related images to the VQA-X validation and test sets. The images of both datasets are from the COCO dataset [22]. The other methods are trained on different datasets, including Visual Genome [20] and VQA v2.0 [8]. The VQA-X dataset consists of 24,876 training, 1,431 validation, and 1,921 test images. Similar to the other methods, we use the test set of the VQA-X dataset to evaluate our method and report the Rank Correlation as in [27]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 514, + 187, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 187, + 527 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 187, + 527 + ], + "type": "text", + "content": "4.5. VizWiz-VQA-Grounding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 715 + ], + "type": "text", + "content": "The VizWiz-VQA-Grounding dataset [3] is based on the VizWiz-VQA dataset [10], and the images and questions come from visually impaired people who shared them to request visual assistance in their day-to-day lives. VizWiz-VQA-Grounding contains a total of 9,998 VQAs divided into 6,494/1,131/2,373 VQAs for training, validation, and testing. The 2022 VizWiz-VQA-Grounding Challenge is designed around the aforementioned VizWiz-VQA-Grounding dataset. This challenge was held recently, and at the time of this writing, current methods in this challenge have not been published yet. Therefore, we evaluate our method using the VizWiz-VQA-Grounding Challenge online leaderboard [3]. The leaderboard evaluates all the methods against the VizWiz-VQA-Grounding test set. The ground truth labels for the test set are not published yet." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 323, + 70, + 529, + 133 + ], + "blocks": [ + { + "bbox": [ + 323, + 70, + 529, + 133 + ], + "lines": [ + { + "bbox": [ + 323, + 70, + 529, + 133 + ], + "spans": [ + { + "bbox": [ + 323, + 70, + 529, + 133 + ], + "type": "table", + "html": "
ModificationMean IoU
Switching to cross-attention70.1
Adding normalization layer71.2 (+1.1)
Switching to Softmax72.7 (+1.5)
Expansion of the embeddings73.5 (+0.8)
", + "image_path": "24ac30a491de6b54516a3cce41b18b2aa42a290bc7aecfbfaa3c93fd3baec4b2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 140, + 545, + 173 + ], + "lines": [ + { + "bbox": [ + 305, + 140, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 545, + 173 + ], + "type": "text", + "content": "Table 7. The results of our ablation study of the design modifications. The modifications are done cumulatively from top to bottom." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 196, + 362, + 207 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 196, + 362, + 207 + ], + "spans": [ + { + "bbox": [ + 306, + 196, + 362, + 207 + ], + "type": "text", + "content": "4.6. Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 217, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 312 + ], + "type": "text", + "content": "We evaluate our proposed method on four answer grounding datasets. Tables 1, 2, and 3 compare the results of our proposed method with the existing methods on the TextVQA-X, VQS, and VQA-X datasets respectively. Our method achieves new state-of-the-art accuracy on three datasets. More specifically, improvements of 2.9, 2.5, and 0.025 on the TextVQA-X, VQS, and VQA-X datasets, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 314, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 545, + 361 + ], + "type": "text", + "content": "Table 4 lists the top-performing teams on the VizWiz-VQA-Grounding Challenge 2022 leaderboard [3]. At the time of this writing, our method holds first place with considerably higher accuracy than the other methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 364, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 411 + ], + "type": "text", + "content": "Furthermore, Table 5 compares our proposed method to existing methods regarding the number of parameters and inference time. The efficient design of our method makes it the fastest method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 430, + 406, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 406, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 406, + 441 + ], + "type": "text", + "content": "5. Ablation Studies" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 451, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 545, + 498 + ], + "type": "text", + "content": "In this section, we perform four ablation studies. All of the experiments are done on the VizWiz-VQA-Grounding [3] validation set and using the setup from Section 4.1 unless otherwise stated." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 513, + 426, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 513, + 426, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 513, + 426, + 525 + ], + "type": "text", + "content": "5.1. Design Modifications" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "We started the design of our sentence attention block from the standard SE block and made four modifications as explained in Section 3.3. In this section, we perform an ablation study on these modifications. Table 7 shows the impact of each modification on the final accuracy. These modifications are done cumulatively. The first modification is changing from self-attention to cross-attention, which makes our baseline for this ablation study. The second modification is adding a normalization layer. This modification improves our baseline by 1.1 percent. The third modification is replacing the Sigmoid function with Softmax. This modification has the highest impact on our method, 1.5 percent. And the fourth modification is the expansion of the embeddings. This last modification brings an addition of 0.8 percent." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6086" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 165, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 165, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 165, + 83 + ], + "type": "text", + "content": "5.2. Backbone Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 290, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 290, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 290, + 368 + ], + "type": "text", + "content": "The flexibility of our method enables us to use a variety of pre-trained networks. In this ablation study, we compare and analyze different pre-trained networks. To compare the image backbone networks, we keep the sentence embedding network fixed and change the image backbone network. Similarly, to compare the sentence embedding networks, we keep the image backbone network with the highest accuracy and change the sentence embedding network. We use well-known and well-performing networks whose pre-trained weights are publicly available. Specifically, we use EfficientNet [37], EfficientNetV2 [38], Swin Transformer V2 [24], and ResNet-50x3 [19] for image feature extraction and MiniLMv2 [42], MPNet [34], and RoBERTa [23] for sentence embedding. Table 6 shows the results of this ablation study. The highest accuracy is achieved by using EfficientNet-B7 and MiniLMv2. Since the EfficientNetV2-L reduces the input resolution more aggressively than EfficientNet-B7, it performs worse. Transformer networks usually perform well on large datasets. However, since VizWiz-VQA-Grounding is not a large dataset, Swin Transformer V2 does not achieve the highest accuracy. The ResNet-50x3 is pre-trained on ImageNet21k, but it achieves the least accuracy in this experiment." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 371, + 151, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 371, + 151, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 151, + 384 + ], + "type": "text", + "content": "5.3. Region Proposals" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 390, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 426 + ], + "type": "text", + "content": "Feature-map visualization will provide insight into the internal representations of each layer in a network for a specific input image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 288, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 641 + ], + "type": "text", + "content": "This ablation study aims to show that the image backbone network has learned the potential regions for an image, which are then processed by the proposed sentence attention block. To this aim, we visualize the final output feature-maps of the image backbone network, before feeding them to an attention block. We use the image backbone of our fully trained network on the VizWiz dataset. Figure 3 (a) shows a few manually-chosen feature-maps for a sample image. These feature-maps show that after training, the backbone network has been able to learn the potential important regions of images. Having all the candidate regions, the task of the proposed attention block is to filter the irrelevant regions to the question and answer. Figure 3 (b) shows the final output of the complete network for the same image in (a), using different questions and answers (for brevity, the answers are not shown). This figure shows that our proposed block is able to filter the channel-wise feature-maps based on the input questions and answers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 647, + 162, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 162, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 162, + 659 + ], + "type": "text", + "content": "5.4. Multi-scale Fusions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "There is a trade-off in the answer grounding task such that some images are best handled at lower inference resolution and others better handled at higher inference resolution. Fine details, such as the edges of objects or thin" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 70, + 532, + 375 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 532, + 375 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 532, + 375 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 532, + 375 + ], + "type": "image", + "image_path": "0109e65c52ea08f2638a00f7618f631593e87b414e618aa912a13861ff1e46f3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 386, + 547, + 485 + ], + "lines": [ + { + "bbox": [ + 304, + 386, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 547, + 485 + ], + "type": "text", + "content": "Figure 3. (a) Eight manually-selected sample feature-maps for an image. These feature-maps are the output of the feature extraction backbone network, before feeding them to the sentence attention block. The network has learned to output potential regions in different channels. (b) The final output of our complete network for the same image in (a), given different questions and answers (for brevity, the answers are not shown). This figure shows that our proposed attention block can filter the channel-wise feature-maps based on the input questions and answers." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 326, + 495, + 526, + 558 + ], + "blocks": [ + { + "bbox": [ + 326, + 495, + 526, + 558 + ], + "lines": [ + { + "bbox": [ + 326, + 495, + 526, + 558 + ], + "spans": [ + { + "bbox": [ + 326, + 495, + 526, + 558 + ], + "type": "table", + "html": "
ScaleMean IoU
1/3271.4
1/32 + 1/1672.5 (+1.1)
1/32 + 1/16 + 1/873.1 (+0.6)
1/32 + 1/16 + 1/8 + 1/473.5 (+0.4)
", + "image_path": "6a6c76095e9032e0825fd1d24b15b7ff12ace196b9ac009e70198670e02ec290.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 565, + 547, + 611 + ], + "lines": [ + { + "bbox": [ + 304, + 565, + 547, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 547, + 611 + ], + "type": "text", + "content": "Table 8. The gradual effect of multi-scale fusion. From top to bottom, the first row uses only the lowest resolution feature-map, and each row adds higher resolution feature-maps. Each scale is the ratio of the feature-map size to the input image." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "content": "structures, are often better predicted with scaled-up images. At the same time, the prediction of large structures, which requires more global context, is often done better at scaled-down images because the network's receptive field can observe more of the necessary context." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "This ablation study shows the impact of adding the extra scales on the accuracy. Table 8 shows the results of this" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6087" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 120 + ], + "type": "text", + "content": "study. From top row to bottom, we start with only one scale and one attention block and gradually add higher resolution feature-maps. The lowest resolution is the final output feature-map of the backbone network." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 132, + 125, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 132, + 125, + 144 + ], + "spans": [ + { + "bbox": [ + 47, + 132, + 125, + 144 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 152, + 289, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 152, + 289, + 344 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 289, + 344 + ], + "type": "text", + "content": "In the study, a new architectural block termed the Sentence Attention Block was introduced to address certain challenges. This block recalibrates channel-wise image feature-maps by modeling inter-dependencies between image feature-maps and sentence embedding. The design began with a recognized attention method, and with minor adjustments, enhanced results were achieved, reaching top-tier accuracy. We showed the block's ability to filter out irrelevant feature-map channels based on sentence embedding. The method's adaptability allows for various pre-trained backbone networks, and its straightforward nature facilitates comprehension and re-implementation. The method's efficacy was demonstrated on the TextVQA-X, VQS, VQA-X, and VizWiz-VQA-Grounding datasets, and several ablation studies highlighted the validity of the design decisions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 355, + 107, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 107, + 368 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 107, + 368 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 376, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 53, + 376, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 287, + 430 + ], + "type": "text", + "content": "[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 432, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 432, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 53, + 432, + 287, + 476 + ], + "type": "text", + "content": "[2] Khaled Bayoudh, Raja Knani, Fayçal Hamdaoui, and Abdel-latif Mtibaa. A survey on deep multimodal learning for computer vision: advances, trends, applications, and datasets. The Visual Computer, 38(8):2939-2970, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 478, + 286, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 478, + 286, + 510 + ], + "spans": [ + { + "bbox": [ + 53, + 478, + 286, + 510 + ], + "type": "text", + "content": "[3] Chongyan Chen, Samreen Anjum, and Danna Gurari. Grounding answers for visual questions asked by visually impaired people. arXiv preprint arXiv:2202.01993, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 512, + 286, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 512, + 286, + 566 + ], + "spans": [ + { + "bbox": [ + 53, + 512, + 286, + 566 + ], + "type": "text", + "content": "[4] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 568, + 286, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 568, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 53, + 568, + 286, + 612 + ], + "type": "text", + "content": "[5] Chaorui Deng, Qi Wu, Qingyao Wu, Fuyuan Hu, Fan Lyu, and Mingkui Tan. Visual grounding via accumulated attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7746-7755, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 613, + 286, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 613, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 53, + 613, + 286, + 677 + ], + "type": "text", + "content": "[6] Chenyou Fan, Xiaofan Zhang, Shu Zhang, Wensheng Wang, Chi Zhang, and Heng Huang. Heterogeneous memory enhanced multimodal attention model for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1999-2007, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 680, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 680, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 680, + 286, + 713 + ], + "type": "text", + "content": "[7] Chuang Gan, Yandong Li, Haoxiang Li, Chen Sun, and Boqing Gong. Vqs: Linking segmentations to questions and answers for supervised attention in vqa and question-focused" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "text", + "content": "semantic segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1811-1820, 2017." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 545, + 162 + ], + "type": "text", + "content": "[8] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 163, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 163, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 547, + 217 + ], + "type": "text", + "content": "[9] Longteng Guo, Jing Liu, Xinxin Zhu, Peng Yao, Shichen Lu, and Hanqing Lu. Normalized and geometry-aware self-attention network for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10327–10336, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 220, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 284 + ], + "type": "text", + "content": "[10] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3608-3617, 2018." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "type": "text", + "content": "[11] Chiori Hori, Takaaki Hori, Teng-Yok Lee, Ziming Zhang, Bret Harsham, John R Hershey, Tim K Marks, and Kazuhiko Sumi. Attention-based multimodal fusion for video description. In Proceedings of the IEEE international conference on computer vision, pages 4193-4202, 2017." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 342, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 397 + ], + "type": "text", + "content": "[12] Parisa Hosseini, Seyedalireza Khoshkirat, Mohammad Jalayer, Subasish Das, and Huaguo Zhou. Application of text mining techniques to identify actual wrong-way driving (wwd) crashes in police reports. International Journal of Transportation Science and Technology, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 399, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 431 + ], + "type": "text", + "content": "[13] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 433, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 487 + ], + "type": "text", + "content": "[14] Pingping Huang, Jianhui Huang, Yuqing Guo, Min Qiao, and Yong Zhu. Multi-grained attention with object-level grounding for visual question answering. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3595-3600, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 489, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 545, + 520 + ], + "type": "text", + "content": "[15] Drew A Hudson and Christopher D Manning. Compositional attention networks for machine reasoning. arXiv preprint arXiv:1803.03067, 2018." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "type": "text", + "content": "[16] Seyedalireza Khoshsirat and Chandra Kambhamtu. Semantic segmentation using neural ordinary differential equations. Advances in Visual Computing: 17th International Symposium, ISVC 2022, San Diego, CA, USA, October 3-5, 2022, Proceedings, Part I, pages 284-295, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 579, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 612 + ], + "type": "text", + "content": "[17] Seyedalireza Khoshsurat and Chandra Kambhametu. Embedding attention blocks for the vizwiz answer grounding challenge. VizWiz Grand Challenge Workshop, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "type": "text", + "content": "[18] Seyedalireza Khoshsirat and Chandra Kambhamtu. Empowering visually impaired individuals: A novel use of apple live photos and android motion photos. 25th Irish Machine Vision and Image Processing Conference, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[19] Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, and Neil Houlsby. Big transfer (bit): General visual representation learning. In European conference on computer vision, pages 491-507. Springer, 2020." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6088" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "text", + "content": "[20] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "type": "text", + "content": "[21] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "type": "text", + "content": "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 252, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 252, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 252, + 287, + 306 + ], + "type": "text", + "content": "[23] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "type": "text", + "content": "[24] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. arXiv preprint arXiv:2111.09883, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 354, + 287, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 375 + ], + "type": "text", + "content": "[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 432 + ], + "type": "text", + "content": "[26] Elham Maserat, Reza Safdari, Hamid Asadzadeh Aghdaei, Alireza Khoshsirat, and Mohammad Reza Zali. 43: Designing evidence based risk assessment system for cancer screening as an applicable approach for the estimating of treatment roadmap. BMJ Open, 7(Suppl 1):bmjopen-2016, 2017." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 434, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 498 + ], + "type": "text", + "content": "[27] Dong Huk Park, Lisa Anne Hendricks, Zeynep Akata, Anna Rohrbach, Bernt Schiele, Trevor Darrell, and Marcus Rohrbach. Multimodal explanations: Justifying decisions and pointing to the evidence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8779-8788, 2018." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "text", + "content": "[28] Badri Patro, Shivansh Patel, and Vinay Namboodiri. Robust explanations for visual question answering. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1577-1586, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 545, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 600 + ], + "type": "text", + "content": "[29] Badri N Patro, Mayank Lunayach, Shivansh Patel, and Vinay P Namboodiri. U-cam: Visual explanation using uncertainty based class activation maps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7444-7453, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "text", + "content": "[30] Varun Nagaraj Rao, Xingjian Zhen, Karen Hovsepian, and Mingwei Shen. A first look: Towards explainable textvqa models via visual and textual explanations. arXiv preprint arXiv:2105.02626, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[31] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[32] Felipe Riquelme, Alfredo De Goyeneche, Yundong Zhang, Juan Carlos Niebles, and Alvaro Soto. Explaining vqa pre" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "dictions using visual grounding and a knowledge base. Image and Vision Computing, 101:103968, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 97, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 545, + 150 + ], + "type": "text", + "content": "[33] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[34] Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. Mpnet: Masked and permuted pre-training for language understanding. Advances in Neural Information Processing Systems, 33:16857-16867, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 198, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 240 + ], + "type": "text", + "content": "[35] Jabeen Summaira, Xi Li, Amin Muhammad Shoib, Songyuan Li, and Jabbar Abdul. Recent advances and trends in multimodal deep learning: A review. arXiv preprint arXiv:2105.11087, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "type": "text", + "content": "[36] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 277, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 277, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 545, + 319 + ], + "type": "text", + "content": "[37] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114. PMLR, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 322, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 322, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 322, + 545, + 353 + ], + "type": "text", + "content": "[38] Mingxing Tan and Quoc V Le. Efficientnetv2: Smaller models and faster training. arXiv preprint arXiv:2104.00298, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 355, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 420 + ], + "type": "text", + "content": "[39] Aisha Urooj, Hilde Kuehne, Kevin Duarte, Chuang Gan, Niels Lobo, and Mubarak Shah. Found a reason for me? weakly-supervised grounded visual question answering using capsules. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8465-8474, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 422, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 466 + ], + "type": "text", + "content": "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 468, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 521 + ], + "type": "text", + "content": "[41] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 524, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 524, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 524, + 545, + 567 + ], + "type": "text", + "content": "[42] Wenhui Wang, Hangbo Bao, Shaohan Huang, Li Dong, and Furu Wei. Minilmv2: Multi-head self-attention relation distillation for compressing pretrained transformers. arXiv preprint arXiv:2012.15828, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 569, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 569, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 569, + 545, + 622 + ], + "type": "text", + "content": "[43] Xi Wei, Tianzhu Zhang, Yan Li, Yongdong Zhang, and Feng Wu. Multi-modality cross attention network for image and sentence matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10941-10950, 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "text", + "content": "[44] Zichao Yang, Xiaodong He, Jianfeng Gao, Li Deng, and Alex Smola. Stacked attention networks for image question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 21-29, 2016." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[45] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6089" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 185 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "text", + "content": "Computer Vision and Pattern Recognition, pages 5579-5588, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "text", + "content": "[46] Yundong Zhang, Juan Carlos Niebles, and Alvaro Soto. Interpretable visual question answering by visual grounding from attention supervision mining. In 2019 IEEE winter conference on applications of computer vision (wacv), pages 349-357. IEEE, 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "text", + "content": "[47] Shuai Zhao, Yang Wang, Zheng Yang, and Deng Cai. Region mutual information loss for semantic segmentation. Advances in Neural Information Processing Systems, 32, 2019." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "6090" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions/2c5b9c34-9105-42fb-b593-c16dc3db7c25_content_list.json b/2023/Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions/2c5b9c34-9105-42fb-b593-c16dc3db7c25_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4e4f8bf88b7eefe2c73209f1326738bb82bfe2c9 --- /dev/null +++ b/2023/Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions/2c5b9c34-9105-42fb-b593-c16dc3db7c25_content_list.json @@ -0,0 +1,1434 @@ +[ + { + "type": "text", + "text": "Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions", + "text_level": 1, + "bbox": [ + 94, + 130, + 875, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shuai Li $^{1,3}$ , Sisi Zhuang $^{1}$ , Wenfeng Song $^{2*}$ , Xinyu Zhang $^{2}$ , Hejia Chen $^{1}$ , Aimin Hao $^{1}$ $^{1}$ State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, P.R. China \n $^{2}$ Computer School, Beijing Information Science and Technology University, P.R. China \n $^{3}$ Zhongguancun Laboratory, Beijing, P.R. China", + "bbox": [ + 98, + 179, + 872, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{lishuai, sisizhuang}@buaa.edu.cn, songwenfenga@gmail.com zhangxinyul@bistu.edu.cn,{chenhj2000, ham}@buaa.edu.cn", + "bbox": [ + 236, + 253, + 733, + 287 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0c8a23ec46a6e56d3b2f31ba9f9d108bfbf5fb8600de7b6bfc14f9653bdfa87c.jpg", + "image_caption": [ + "Figure 1: Our method synthesizes semantic human motion with natural transition using free-form sequential texts. The skeleton motion can be easily transferred to any character. The character used for demonstration is from Mixamo [1]." + ], + "image_footnote": [], + "bbox": [ + 78, + 297, + 890, + 400 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 446, + 313, + 462 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The intelligent synthesis/generation of daily-life motion sequences is fundamental and urgently needed for many VR/metaverse-related applications. However, existing approaches commonly focus on monotonic motion generation (e.g., walking, jumping, etc.) based on single instruction-like text, which is still not intelligent enough and can't meet practical demands. To this end, we propose a cohesive human motion sequence synthesis framework based on free-form sequential texts while ensuring semantic connection and natural transitions between adjacent motions. At the technical level, we explore the local-to-global semantic features of previous and current texts to extract relevant information. This information is used to guide the framework in understanding the semantics of the current moment. Moreover, we propose learnable tokens to adaptively learn the influence range of the previous motions towards natural transitions. These tokens can be trained to encode the relevant information into well-designed transition loss. To demonstrate the efficacy of our method, we conduct extensive experiments and comprehensive evaluations on the public dataset as well as a new dataset produced by us. All the experiments confirm that our method outperforms the state-of-the-art methods in terms of semantic matching, realism, and transition fluency. Our project is public available. https://druthrie.github.io/sequential-texts-to-motion/", + "bbox": [ + 73, + 476, + 473, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction and Motivation", + "text_level": 1, + "bbox": [ + 500, + 446, + 764, + 462 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human motion synthesis is fundamental for numerous applications, especially for virtual reality, games, and metaverse-related applications [10, 19, 5, 3, 36, 15, 22, 20, 24], of which, it is in high demand to accurately control the digital human motion with natural language. Existing approaches commonly focus on monotonic motion synthesis/generation based on single instruction-like text description. However, practical applications usually require digital humans to respond to multiple rounds of sustainable interactions, wherein they can continuously generate reasonable responses to the sequential texts. Therefore, given a set of free-form sequential texts, we aim to synthesize a cohesive human motion sequence. Namely, the animation clips should be consistent with the ongoing text descriptions, and the whole motion sequence should have smooth semantic connection and natural transitions.", + "bbox": [ + 496, + 474, + 892, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "At present, sequential texts-driven motion synthesis has not been well studied, mainly due to lacking long-term continuous motion datasets with accompanying free-form text descriptions. Recently, TEACH [6] first attempts to address this problem by proposing a dataset (we refer to it as BABEL-TEACH). Each item in the BABEL-TEACH dataset contains two adjacent text-motion pairs. However, TEACH [6] builds on the one-time method TEMOS [29], wherein the limited 5-frame motion is added to encode together. However, it does not consider the potential benefits of previous text information. In practice, previous texts can provide extra and more accurate semantic information for", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 1, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 101, + 887, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9498", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "current motion synthesis. Therefore, global semantic needs to be extracted from the previous text and the local semantics from the current text, and then fuses the semantics to guide the semantic understanding at each moment.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides, there are still two main challenges. Firstly, the semantic relationship between adjacent texts is largely overlooked in previous research. For example, when synthesizing a motion for the current text \"touching the face with left hand,\" the previous text context should be considered. Specifically, if the previous text was \"sitting on a chair,\" the synthesized motion should depict a person sitting and touching face with left hand (shown in Fig. 2); if the previous text was \"a person kicks a ball with right foot,\" the motion should show a standing person touching face with left hand. This requires a more comprehensive understanding about the context and semantic relationships between the texts. Secondly, existing methods tend to abrupt transitions between adjacent motions when multiple motions are synthesized separately and stitched together. We aim to address this issue by seamlessly blending the synthesized motions. It should intelligently capture the temporal and spatial continuity between the motions to ensure a more natural and coherent sequence of motions.", + "bbox": [ + 75, + 152, + 470, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Considering the easy scalability of autoregressive methods, we should extend previous autoregressive single text-driven motion synthesis methods to multiple motions involved in motion sequence synthesis. We observe that a similar human body posture can be achieved by transferring the end features of the previous motion to the next motion, although this is not always entirely consistent. This suggests that previous motion information should be leveraged more effectively and sufficiently rather than being adopted straightforwardly. Therefore, we plan to introduce a transition reasoning module that adaptively learns attention score from the previous motion information and infers accurate motion features at the transition, making the transition between adjacent motions more natural and smooth.", + "bbox": [ + 75, + 439, + 468, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we create a new dataset with much longer-term motions (2-5 times longer) based on a synthetic method, and the corresponding text descriptions are more abundant. We conduct experiments on both BABEL-TEACH [6] dataset and our sequential texts described motion (STDM) dataset. The results demonstrate that our method outperforms existing methods in terms of semantic matching, realism, and transition fluency. Specially, the salient contributions can be summarized as follows.", + "bbox": [ + 75, + 650, + 468, + 785 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a cohesive motions synthesis framework using sequential texts as inputs. The framework can integrate previous text information to obtain semantic features and adaptively select valid previous motion information to guide the current motion synthesis.", + "- We design a local-to-global (L2G) semantic fusion module to extract accurate contextual information. It" + ], + "bbox": [ + 94, + 791, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/657fd4c19d8f689b3c9cc4a36f4b5e6c8b751da6206a706583b27b768b20a515.jpg", + "image_caption": [ + "Figure 2: Top row: Motion synthesis corresponding to sequential texts by our method. Bottom row: Motion synthesis corresponding to single text by T2M [11]." + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 874, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "can well take into account global-context information from the previous text to infer the current moment's text meaning to guide the semantics-consistent motion synthesis.", + "bbox": [ + 529, + 271, + 890, + 330 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a transition reasoning module to adaptively select motion snippet from previous motion information to make the synthesized motion natural and smooth w.r.t the previous, wherein a well-defined transition loss is employed to further constrain the fluency of the transitions.", + "- We create a new sequential texts described motion (STDM) dataset, which involves more extended motion frames and more diverse text descriptions than the existing dataset." + ], + "bbox": [ + 517, + 334, + 890, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 497, + 640, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We briefly introduce the related works of human motion synthesis considering semantic control, including single text driven motion synthesis and sequential texts driven motion synthesis.", + "bbox": [ + 498, + 523, + 890, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single text driven motion synthesis. The motion synthesis from text began with the synthesized motion from action labels, but this approach was limited in covering most human motions [21, 27, 35, 38, 28, 12, 28, 7]. Later, researchers [4, 9, 34] focused on synthesizing motion from complex free-form text descriptions, with early work generating single, deterministic motions from text. Recent works have aimed to improve the diversity and details of the synthesized motion. For example, TEMOS [29] and T2M [11] utilized VAE architecture to synthesize diverse motions from the same texts. The main difference between them is that the former utilized Transformer architecture to synthesize all moment motions at once, while the latter utilized GRU architecture to synthesize every moment motion auto-regressively. To capture more detailed information in the text, TM2T [14] utilized VQVAE-based motion markers to provide a fair environment when considering motion and text signals, and utilized motion-to-text analysis module to strengthen the constraints on text-to-motion synthesis. With the successful application of the diffusion model in image generation, MotionDiffuse [37] utilized the diffusion model", + "bbox": [ + 496, + 584, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9499", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/564f1bf0b59f23c839f8645c4126c2ac2d024c60f5b9481f4fb669600c52f322.jpg", + "image_caption": [ + "Figure 3: Method Overview: Our model takes sequential texts as input. In the local-to-global (L2G) semantic fusion module, the previous text's global features and the current text's local features are fused and extracted. In the transition reasoning module, the motion snippet code at the transition of adjacent motion ( $\\mathbf{c}_{start}^{0}$ and $\\mathbf{c}_{start}^{1}$ ) is deduced from the previous motion. Input the gate snippet code $\\mathbf{g}^t$ and the semantic text feature $F_{sem}^{t}$ to synthesize the motion snippet code $\\hat{\\mathbf{c}}_i^t$ . The transition loss further limits the fluency of the transition with the previous motion." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 883, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to simulate text driven conditional human motion synthesis, which can better respond to fine-grained instructions of body parts. FLAME [18] utilized transformer-based diffusion model architecture to synthesize human motion, and it can allow editing frames and joints without fine-tuning. Its editing ability can be extended to motion prediction or in-between tasks. Although the diffusion model improves the quality of the synthesized motion, its inference speed is relatively slow, so it is unsuitable for real-time applications.", + "bbox": [ + 75, + 415, + 470, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sequential texts driven motion synthesis. Synthesizing motion sequences from sequential texts is challenging due to the limited availability of long-term continuous motions datasets with text descriptions. Previous works Action2video [13] realized the transition between actions by changing the input action label type in an autoregressive synthesizer. Yet, they only conducted experiments on a limited number of action labels in three datasets: NTURGBD [33], CMU [2], and humanACT12 [12]. Mao et al. [26] proposed a weakly-supervised action-driven motion prediction method, but they only used 20 action categories with clear transitions in BABEL [32] to synthesize the following motion and transition. Based on the existing BABEL dataset [32], TEACH [6] proposed a new dataset, in which each data consisted of two adjacent text descriptions and motion sequences. The model of TEACH [6] encoded five previous motion frames using TEMOS [29], every single motion sequence was synthesized simultaneously, and adjacent motion sequences were auto-regressively synthesized. Our work builds on this by proposing a cohesive human motion sequence synthesis framework based on freeform sequential texts, ensuring semantic connection and natural transitions between adjacent motions.", + "bbox": [ + 75, + 553, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Approach", + "text_level": 1, + "bbox": [ + 498, + 414, + 650, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Method Overview", + "text_level": 1, + "bbox": [ + 500, + 438, + 674, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall pipeline of the proposed model is shown in Fig. 3. The temporal VAE and motion autoencoder used in the model are similar to those used in the T2M model, wherein it is employed to encode the motion sequence into a motion snippet code sequence and reconstruct the motion sequence with a decoder. To take sequential texts as input and ensure semantically coherent synthesized motions, a local-to-global (L2G) semantic fusion module is introduced (Sec. 3.2). This module guides learning accurate semantic features from the input texts at each moment, including historical text information. To ensure smooth and natural transitions between adjacent motions, a transition reasoning module is introduced (Sec. 3.3). This module dynamically deduces the motion snippet code during transition based on previous motion information, which can help the model generate realistic and coherent transitions. Additionally, a transition loss is proposed to enhance the smoothness of the transition motion (Sec. 3.4). Overall, the proposed pipeline combines several modules and techniques to generate realistic and semantically coherent motion sequences from sequential texts input. The model can generate smooth transitions between motions, and can adaptively learn the influence range of previous motions to further improve the quality of the synthesized motion sequences.", + "bbox": [ + 496, + 462, + 893, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminaries. Our method aims to take freeform sequential texts $\\mathbf{S} = (\\mathbf{S}_1, \\mathbf{S}_2, \\dots, \\mathbf{S}_n)$ as input, e.g., $\\mathbf{S} =$ (\"salute with the left hand\", \"salute with the right hand\", \"sit on the chair\", \"crawl\"), and outputs corresponding human motion sequences $\\mathbf{M} =$", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9500", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$(\\mathbf{M}_1, \\mathbf{M}_2, \\dots, \\mathbf{M}_n)$ . The transition between adjacent motions should be natural and smooth, and the synthesized motion should reflect the semantic information of previous texts. The data format is consistent with HumanML3D [11] dataset, including the root angular velocity along the Y-axis, the root linear velocity on the XZ-plane, the root height, the local joints positions, velocities, and rotations [40] in root space, while the motions follow the skeleton structure of SMPL [25] with 22 joints. Each moment in our model represents 4 frames.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Local-to-global Semantic Fusion Module", + "text_level": 1, + "bbox": [ + 76, + 250, + 426, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To synthesize motion that accurately matches the semantics of sequential texts, our model relies on accurately extracting text features at each moment to guide the synthesis process. As the motion we synthesize is based on sequential texts, the current motion is determined not only by the current text semantics but also by previous text semantics. To capture the temporal clues, we encode both the previous and current text, and use an attention mechanism to extract text features that integrate the previous text semantics at each moment (shown in Fig. 4). For previous texts, we use a pretrained CLIP [31] text encoder to extract global features of previous texts. In order to make the features more suitable for our task in the process of training, linear projector is used to extract features further. The global features from the previous texts are expressed as:", + "bbox": [ + 75, + 272, + 472, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} _ {p r e} = f \\left(C L I P _ {t e x t} \\left(\\mathbf {S} _ {i - 1}\\right)\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 506, + 468, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is difficult for the CLIP text encoder to extract the local features of the text, motion synthesis of each moment requires local text semantic guidance, so we employ a Bi-GRU-based text encoder, similar to T2M [11], to obtain the local features $\\mathbf{L}_{1:m}$ from the current text. To capture accurate text semantic features at the arbitrary frames, we use an attention mechanism, wherein the key $\\mathbf{K}(L)$ and value $\\mathbf{V}(L)$ of attention are the local features of the current text. Since the semantic information of the previous text is required for guidance at the beginning of the currently synthesizing motion, we use the global features of the previous text as the initial query $\\mathbf{Q}(G)$ . At subsequent moments, the motion snippet code generator in temporal VAE provides the queries $\\mathbf{Q}(L)$ . This can be expressed mathematically as:", + "bbox": [ + 75, + 527, + 472, + 753 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} (G) = \\mathbf {G} _ {p r e} \\mathbf {W} ^ {Q}, \\mathbf {Q} (L) = \\mathbf {h} ^ {t - 1} \\mathbf {W} ^ {Q},\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 757, + 403, + 775 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K} (L) = \\mathbf {L} _ {1: m} \\mathbf {W} ^ {K}, \\mathbf {V} (L) = \\mathbf {L} _ {1: m} \\mathbf {W} ^ {V}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 777, + 468, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {s e m} ^ {t} = f \\left(\\frac {\\mathbf {Q} (G / L) \\mathbf {K} (L) ^ {T}}{\\sqrt {d _ {s e m}}}\\right) \\mathbf {V} (L),\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 800, + 400, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{W}^k, \\mathbf{W}^V \\in \\mathbb{R}^{d_w \\times d_{sem}}$ and $\\mathbf{W}^Q \\in \\mathbb{R}^{d_h \\times d_{sem}}$ are trainable weights; $d_h, d_w$ and $d_{sem}$ are the number of channels in hidden unit $\\mathbf{h}^{t-1}$ , current text feature $\\mathbf{L}^{1:m}$ and semantic fusion attention layer, respectively. $\\mathbf{h}^t$ is the hidden", + "bbox": [ + 75, + 838, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e664c64417496ee69981a4c62ac4e251ef467f247ab8e2a2e6c9469cab138718.jpg", + "image_caption": [ + "Figure 4: Local-to-global semantic fusion module: After extracting the global feature $\\mathbf{G}_{pre}$ from the previous texts and the local features $\\mathbf{L}_{1:m}$ from the current text, feature fusion is performed to obtain the accurate semantic feature $\\mathbf{F}_{sem}^{t}$ of the current text at moment $t$ ." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 885, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "unit generated by the motion snippet code generator at each moment and $\\mathbf{h}^0$ is obtained from $\\mathbf{G}_{pre}$ . $\\mathbf{F}_{sem}^{t}$ is the text semantic feature at moment $t$ .", + "bbox": [ + 498, + 415, + 892, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Transition Reasoning Module", + "text_level": 1, + "bbox": [ + 500, + 470, + 764, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When synthesizing a new motion, the starting pose is influenced by the ending pose of the previous motion. Grafting the end motion snippet code of the previous motion to the start moment of the current motion can produce similar human poses during the transition. However, the transition between the two motions is still unnatural and abrupt. To address this, we introduce a learnable start motion snippet code token $\\mathbf{c}_{\\text {token }}$ and utilize it along with the previous motion snippet code sequence in a transformer encoder. The output of the token position is used as the start motion snippet code $\\mathbf{c}_{\\text {start }}$ for the current motion, allowing the model to learn the range and weight of the previous motions (shown in Fig. 5). Additionally, the length of the current motion affected by the previous motion information needs to be determined. Through experimentation, we obverse that the best performance is achieved by reasoning the first two motion snippet codes $(\\mathbf{c}_{\\text {start }}^{0}$ and $\\mathbf{c}_{\\text {start }}^{1})$ of the current motion from the previous motions. The verification experiment is in Sec. 4.5. The mathematical expression is as:", + "bbox": [ + 496, + 494, + 893, + 781 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {\\text {s t a r t}} ^ {0}, \\mathbf {c} _ {\\text {s t a r t}} ^ {1} = \\text {T r a n s f o r m e r} \\left(\\hat {\\mathbf {c}} _ {i - 1}, \\mathbf {c} _ {\\text {t o k e n}} ^ {0}, \\mathbf {c} _ {\\text {t o k e n}} ^ {1}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 792, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown by the transition gate in Fig. 3, when synthesizing the motion snippet codes of the first two moments, we choose the motion snippet codes $(\\mathbf{c}_{start}^{0}$ and $\\mathbf{c}_{start}^{1})$ inferred from the previous motion as gate snippet code $\\mathbf{g}^t$ . From the third moment, the motion snippet code $\\hat{\\mathbf{c}}^{t - 1}$ from the previ", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9501", + "bbox": [ + 480, + 944, + 513, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/229c73e4a307153e29732d9aa1410ef2ce26d964a6e8e67919883968a8f44dd2.jpg", + "image_caption": [ + "Figure 5: Transition reasoning module: Given the previous texts to get the previous synthesized motion sequence $\\hat{\\mathbf{M}}_{i-1}$ and motion snippet codes $\\hat{\\mathbf{c}}_{i-1}$ , we use two learnable tokens ( $\\mathbf{c}_{token}^{0}$ and $\\mathbf{c}_{token}^{1}$ ) to learn the previous motion information that needs attention, so as to obtain the motion snippet codes at the starting moment of the current motion ( $\\mathbf{c}_{start}^{0}$ and $\\mathbf{c}_{start}^{1}$ )." + ], + "image_footnote": [], + "bbox": [ + 114, + 87, + 433, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ous moment is used, which is synthesized autoregressively. For gate snippet code $\\mathbf{g}^t$ , the mathematical expression is as:", + "bbox": [ + 75, + 512, + 468, + 544 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {g} ^ {t} = \\left\\{ \\begin{array}{l l} \\mathbf {c} _ {\\text {s t a r t}} ^ {0} & t = 1 \\\\ \\mathbf {c} _ {\\text {s t a r t}} ^ {1} & t = 2 \\\\ \\hat {\\mathbf {c}} ^ {t - 1} & t \\geq 3 \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 551, + 468, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When synthesizing the motion based on the first text $\\mathbf{S}_1$ of the sequential texts, we set the previous texts as the default text $\\mathbf{S}_0 = \\text{\"start\"}$ , and the starting snippet code $\\mathbf{c}_{start}^0$ is obtained from the mean pose.", + "bbox": [ + 75, + 604, + 470, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Transition Loss & Total Loss", + "text_level": 1, + "bbox": [ + 76, + 674, + 336, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct a statistical analysis of the dataset and find that the motion gap (L1 distance) of almost all adjacent frames is less than 0.1. Inspired by this, we design a transition loss $L_{Trans}$ to reduce the motion gap between the adjacent frames and avoid abrupt transition. The representation of the $L_{Trans}$ is", + "bbox": [ + 75, + 696, + 468, + 789 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {T r a n s}} = \\left\\| \\hat {\\mathbf {M}} _ {i} ^ {1} - \\hat {\\mathbf {M}} _ {i - 1} ^ {T _ {i} - 1} \\right\\| _ {1}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 795, + 468, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final loss function of our model consists of 4 parts: motion snippet code reconstruction loss $\\mathcal{L}_{rec}^{code}$ , motion reconstruction loss $\\mathcal{L}_{rec}^{mot}$ , KL loss $\\mathcal{L}_{KL}$ of prior distribution and posterior distribution in temporal VAE, and the transition loss $\\mathcal{L}_{Trans}$ . The $\\lambda_{code},\\lambda_{mot},\\lambda_{KL},\\lambda_{Trans}$ are the", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/21d4dbcc3f871464632073cfeb478317d11501b8dc90809bb8d3161441a1e1af.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetQuantityDurationVocab.
STDM528920.02 h2455
BABEL-TEACH[6]1626621.86 h1037
", + "bbox": [ + 511, + 88, + 883, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: Comparison of datasets: STDM dataset has longer-term frames within a single motion, and the text is more abundant.", + "bbox": [ + 498, + 148, + 893, + 194 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "corresponding weights of the losses.", + "bbox": [ + 500, + 218, + 743, + 233 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {\\text {c o d e}} \\mathcal {L} _ {\\text {r e c}} ^ {\\text {c o d e}} + \\lambda_ {\\text {m o t}} \\mathcal {L} _ {\\text {r e c}} ^ {\\text {m o t}} + \\lambda_ {K L} \\mathcal {L} _ {K L} + \\lambda_ {\\text {T r a n s}} \\mathcal {L} _ {\\text {T r a n s}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 243, + 892, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 282, + 633, + 300 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we introduce our STDM dataset (Sec. 4.1), evaluation metrics (Sec. 4.2), and then conduct quantitative (Sec. 4.3) and qualitative (Sec. 4.6) evaluations. We carry out an ablation study (Sec. 4.4) to prove the effectiveness of our proposed modules and analyze the parameters (Sec. 4.5).", + "bbox": [ + 498, + 306, + 892, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. STDM Dataset", + "text_level": 1, + "bbox": [ + 500, + 409, + 651, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each text-motion pair in most existing human motion datasets [30, 17, 23, 39, 12, 8] is independent of the other, and does not contain adjacent text-motion pairs. In a recent study, TEACH [6] used the BABEL [32] dataset to obtain a BABEL-TEACH dataset containing two adjacent text-motion pairs, which is consistent with our research goal.", + "bbox": [ + 498, + 431, + 892, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To obtain adequate motion sequences, we synthesize sequential texts and motion sequences by modifying the existing single text driven motion synthesis method from the viewpoint of data generation instead of using a mocap device. By randomly selecting and annotating texts from a large pool, we create sequential text fed into our improved T2M [11], generating motion sequences. Based on T2M [11], we find that the motion sequences decoded by the same motion snippet code are similar. Our approach diverges from T2M: In the inference stage, instead of using a fixed mean pose, we use snippet codes from the last eight frames of the previous motion, ensuring smoother and more natural transitions. Our experiment results show that the motion sequences synthesized in this way are similar in transition. Since the motion generated by the generation model may not match the semantics of the input text and the transition may not be smooth, we manually remove data with poor transitions and correct erroneous texts to maintain the quality of our STDM dataset. After filtering, we get 5289 pairs containing two adjacent text-motion. Our dataset is collected under the same structure as HumanML3D [11].", + "bbox": [ + 496, + 523, + 893, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared with the BABEL-TEACH [6] dataset, STDM has longer-term motion frames, with a single motion sequence ranging from 2 to 10 seconds, while most BABEL-TEACH [6] data is less than 3 seconds. However, the text", + "bbox": [ + 498, + 840, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9502", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/768a408617a2d8ad37757a91c9ced008343033504d353dd63dcf4d587d894e98.jpg", + "image_caption": [ + "Figure 6: STDM Dataset: Two examples in the STDM dataset, which involves more extended motion frames and more diverse text descriptions than the BABEL-TEACH [6] dataset. Blue traces represent the digital human's trajectory." + ], + "image_footnote": [], + "bbox": [ + 155, + 88, + 816, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "descriptions in BABEL-TEACH [6] usually describe only one action, such as \"walk forward\". The text descriptions in the STDM are more abundant, such as \"a person raises his right hand and walks forward\". A detailed comparison of the two datasets can be found in Tab. 1. Fig. 6 shows two examples of our STDM dataset.", + "bbox": [ + 75, + 377, + 472, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 76, + 488, + 261, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In order to evaluate the quality of the synthesized motion sequence, we use the feature extractor proposed by Guo et al. [11] to extract the features of text and motion. This feature extractor is trained with contrast loss [16], which can make the matched text-motion pairs produce similar feature vectors. Therefore, the feature extractor can measure whether the text and motion sequence match. We combine two adjacent text-motion pairs into one to train the feature extractor. We use the same evaluation metrics as T2M [11]. The metrics are described in the supplementary materials.", + "bbox": [ + 75, + 516, + 468, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To demonstrate the fluency of different methods when transiting between adjacent motions, we add a metric to evaluate the fluency of transition, by calculating the distance between adjacent frames of two adjacent motion sequences. For transition fluency, the mathematical expression is as:", + "bbox": [ + 75, + 671, + 470, + 748 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF = \\sum_ {i = 1} ^ {N} \\| \\mathbf {M} _ {i 1} ^ {T 1} - \\mathbf {M} _ {i 2} ^ {1} \\| _ {1}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 765, + 468, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, we calculate the average transition distance ( $L_{1}$ distance) of the motion sequence synthesized by $N$ adjacent text pairs. For the $i$ -th adjacent text pair, the synthesized adjacent motion sequence pair is $\\mathbf{M}_{i1} = (\\mathbf{M}_{i1}^{1}, \\mathbf{M}_{i1}^{2}, \\dots, \\mathbf{M}_{i1}^{T1})$ and $\\mathbf{M}_{i2} = (\\mathbf{M}_{i2}^{1}, \\mathbf{M}_{i2}^{2}, \\dots, \\mathbf{M}_{i2}^{T2})$ .", + "bbox": [ + 75, + 825, + 470, + 905 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 498, + 376, + 720, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our method with T2M [11], our two variants T2M-Joint and Complusion-Code, and TEACH [6]. For T2M, we input a single text, output a single motion, and align the root coordinates of the last frame of the previous motion with the root coordinates of the first frame of the next motion. For T2M-Joint, we keep the T2M model structure and train the model by combining two adjacent texts with commas as input. For Complusion-Code, based on the T2M [11], the last motion snippet code of the previous motion is used as the start motion snippet code of the next motion during training (Unlike the STDM dataset generation method, the dataset generation only uses the snippet codes of previous motions in the inference phase. Complusion-Code also uses the snippet codes of previous motions in the training phase). For TEACH [6], we compared two versions using spherical linear interpolation (Slerp) and not using Slerp, respectively. Slerp operation is unrelated to the TEACH [6] model and is simply a way of post-processing synthesized data.", + "bbox": [ + 496, + 400, + 890, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 2 and Tab. 3 show our quantitative comparison results with other baselines on BABEL-TEACH [6] and STDM datasets. We divide the training, testing, and validation sets according to the ratio of $0.8:0.15:0.05$ . For a fair comparison, we conduct 20 experiments and show the values with confidence in the $95\\%$ range. From the results, our method outperforms all other methods, especially in the transition fluency metric, which is important for synthesizing natural and smooth transition motion. T2M [11] can only synthesize a single motion per inference, and there is no connection between adjacent motions, so it performs poorly in all indicators. Because our testing data are composed of two adjacent texts, the T2M-Joint method performs similarly to the real data regarding transition fluency", + "bbox": [ + 496, + 689, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9503", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/acbd463e546943742ce9c5d610040bb1ed822ad27aa242b706d78251e35479cd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsR Precision↑FID↓MultiModal Dist↓Transition Fluency↓
Top 1Top 2Top 3
Real motions0.560±0.0020.750±0.0030.834±0.0020.001±0.0003.448±0.001-
T2M [11]0.376±0.0030.543±0.0020.647±0.0036.890±0.0815.218±0.0101.108±0.004
T2M-Joint0.373±0.0040.534±0.0030.640±0.0032.727±0.0324.983±0.011-
Compulsion-Code0.432±0.0030.613±0.0040.717±0.0033.726±0.0684.612±0.0230.348±0.002
TEACH (no Slerp) [6]0.573±0.0020.756±0.0030.842±0.0022.263±0.0553.487±0.0100.577±0.003
TEACH (Slerp) [6]0.563±0.0030.750±0.0030.839±0.0032.240±0.0533.526±0.0110.118±0.001
Ours0.542±0.0030.728±0.0030.818±0.0031.628±0.0313.662±0.0100.177±0.001
Ours (Slerp)0.519±0.0040.705±0.0040.803±0.0031.548±0.0243.797±0.0120.036±0.000
", + "bbox": [ + 142, + 88, + 823, + 233 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/471ab5bdd2eda13748aa515ffb5b827d896a7afdda2a5729d78d6a77e97e757b.jpg", + "table_caption": [ + "Table 2: Quantitative evaluation on testing data of BABEL-TEACH: ± indicates 95% confidence interval, ↑ and ↓ respectively denotes better performance with larger or lower value. Bold indicates the optimal result, the underscore represents the suboptimal result, while wave line refers to the third best. The results show that the motion synthesized by our model outperforms other baselines in terms of semantic matching, transition fluency, and realism." + ], + "table_footnote": [], + "table_body": "
MethodsR Precision↑FID↓MultiModal Dist↓Transition Fluency↓
Top 1Top 2Top 3
Real motions0.322±0.0040.505±0.0050.625±0.0060.014±0.0023.461±0.0050.454±0.000
T2M [11]0.322±0.0060.493±0.0050.614±0.0051.389±0.0493.500±0.0090.639±0.003
T2M-Joint0.304±0.0050.488±0.0060.613±0.0050.744±0.0423.492±0.011-
Compulsion-Code0.285±0.0070.450±0.0060.572±0.0063.146±0.0803.806±0.0170.330±0.003
TEACH (no Slerp) [6]0.318±0.0060.505±0.0050.634±0.0051.414±0.0553.483±0.0120.246±0.002
TEACH (Slerp) [6]0.326±0.0040.504±0.0050.631±0.0061.416±0.0543.487±0.0130.049±0.000
Ours0.328±0.0060.510±0.0060.633±0.0051.085±0.0633.441±0.0170.109±0.001
Ours (Slerp)0.325±0.0070.515±0.0070.637±0.0051.134±0.0593.440±0.0130.022±0.000
", + "bbox": [ + 142, + 311, + 823, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3: Quantitative evaluation on the testing data of STDM: The results show that the motion synthesized by our model outperforms other baselines in terms of semantic matching, transition fluency, and realism.", + "bbox": [ + 73, + 463, + 892, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(less than 0.1). However, when the input text consists of more than two sentences, the T2M-Joint method cannot guarantee transition fluency. Because Compulsion-Code uses the end motion snippet code of the previous motion as the start snippet code of the next motion, the performance of transition fluency is better than T2M [11], but it is still not as good as ours. Although TEACH [6] is slightly better than ours in R Precision and MultiModal Distance on the BABEL-TEACH [6] dataset, as shown in Tab. 2, its transition fluency is not as good as ours, we can obtain transitional and natural motion sequences without Slerp operation, so our method is better than TEACH [6] in a comprehensive evaluation. Although the synthetic STDM dataset is less natural than actual mocap data in transition fluency, we find that our method can compensate for the defects of the training data after training. In terms of transition fluency, the results synthesized by our method outperform the ground truth of the testing data. This further verifies the effectiveness of our method in synthesizing natural and smooth transition motions.", + "bbox": [ + 75, + 513, + 472, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User study. We randomly select 50 texts from the testing data and arrange 3 to 5 texts together to form a set of sequential texts. Some text combinations are not present in the dataset to test the motion generation effect of each model for text combinations not in the dataset. The syn", + "bbox": [ + 75, + 824, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "thesized motions based on these sequential text groups are randomly shuffled and presented using different methods. Users are required to score synthetic motions from the following three aspects: (1) Matching degree: the degree of semantic matching between the motion and the text; (2) Transition fluency: the natural degree of transition between adjacent motions; (3) Realism: how realistic the motion is. The result is shown in Fig. 7.", + "bbox": [ + 496, + 513, + 893, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 498, + 647, + 653, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In order to illustrate the influence of the modules and training strategies we introduced on the overall model, we carried out the ablation experiment. The results are shown in Tab. 4.", + "bbox": [ + 496, + 672, + 893, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The first experiment is to remove the transition reasoning module. We set the skeleton to mean pose to initialize the motion snippet code at the start moment. However, the model does not know the human posture at the beginning, so it greatly reduces transition fluency (value increased by 0.417). The experiment results show that the performance of the model will decline without the transition reasoning module because it does not aware of the information of previous motion when the model synthesizes the motion sequence of the current text, even though the model can synthesize motion sequences that match the semantics of the", + "bbox": [ + 496, + 734, + 895, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9504", + "bbox": [ + 480, + 944, + 517, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/21fcdca1e37f74135b637dcf32a19139aacd6a414cb48778e0ebdebb2efc633e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsR Precision↑FID↓MultiModal Dist↓Transition Fluency↓
Top 1Top 2Top 3
Real motions0.561±0.0030.747±0.0030.836±0.0020.002±0.0003.449±0.002-
w/o Transition reasoning module0.535±0.0040.713±0.0030.805±0.0032.190±0.0383.719±0.0130.594±0.003
w/o L2G semantic fusion module0.440±0.0030.621±0.0030.720±0.0034.592±0.0714.593±0.0130.151±0.001
w/o Transition loss0.525±0.0040.706±0.0040.801±0.0021.766±0.0363.797±0.0110.293±0.001
w/o Segmented training strategy0.500±0.0020.684±0.0040.783±0.0042.164±0.0384.041±0.0130.328±0.002
Ours0.542±0.0030.726±0.0030.817±0.0021.588±0.0303.655±0.0130.177±0.001
", + "bbox": [ + 117, + 88, + 849, + 203 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a1746572055798abcbc12516db68c65adb872abab14796bc08435158ddc44f33.jpg", + "table_caption": [ + "Table 4: Ablation study: The results show that the transition reasoning module has the most significant impact on transition fluency, the L2G semantic fusion module has the most significant impact on R Precision, FID, and MultiModel Distance, which represent semantic matching and realism, and the transition loss and segmentation training strategy have a certain influence on all indicators." + ], + "table_footnote": [], + "table_body": "
Numbers of snippet codeR Precision↑FID↓MultiModal Dist↓Transition Fluency↓
Top 1Top 2Top 3
Real motions0.561±0.0030.750±0.0020.834±0.0020.002±0.0003.449±0.001-
1 snippet code0.517±0.0030.693±0.0040.786±0.0042.027±0.0433.845±0.0140.173±0.001
2 snippet codes0.540±0.0030.725±0.0030.816±0.0021.602±0.0263.662±0.0110.177±0.001
3 snippet codes0.522±0.0040.702±0.0030.796±0.0022.050±0.0383.838±0.0150.175±0.001
5 snippet codes0.504±0.0030.687±0.0040.785±0.0031.781±0.0293.913±0.0130.213±0.001
", + "bbox": [ + 135, + 281, + 831, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5: Parameter analysis of transition reasoning module: We infer the initial motion snippet code of the current motion sequence from the previous motion and determine the optimal number of the inference code through experiments. The result shows that inferring the first two codes of current motion from the previous motion work best.", + "bbox": [ + 75, + 387, + 893, + 434 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f65555da25ba3cdd78d80957654a4b2c233ac97ed3f40765f4575347f5c06d2.jpg", + "image_caption": [ + "Figure 7: User study: We invite 45 users to score the results of our comparative experiments, and the color bars in the figure indicate the percentage of the scores. The results show that our method outperforms other baselines in terms of semantic matching, transition fluency, and realism." + ], + "image_footnote": [], + "bbox": [ + 84, + 444, + 455, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "current text.", + "text_level": 1, + "bbox": [ + 76, + 750, + 158, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The second experiment is to remove the L2G semantic fusion module. We only encode the features of the current text at each moment without considering the previous text. The experiment results show that the model's performances on R Precision, FID, and MultiModal Distance decrease significantly after removing the L2G semantic fusion module, indicating that previous text information is essential in synthesizing motion sequences that match context semantics. Although there is a slight improvement in the metric of", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "transition fluency after the removal, the gap between them can not be observed from the naked eye. With the L2G semantic fusion module, our model shows better competence in evaluating the implicit relationship between current and previous text.", + "bbox": [ + 496, + 445, + 890, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The third experiment is to remove the transition loss. The experimental results show that after removing the transition loss, the model's performances in all indicators have decreased, indicating that adding transition loss will guide the model to learn more semantic matching and transitional natural motion.", + "bbox": [ + 496, + 520, + 890, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The fourth experiment does not use the segmented training strategy (explained in the supplementary file). The specific approach is that all motion sequences are extended to the maximum length (196 frames) with 0 in the training strategy. The experimental results show that when the segmented training strategy is not used, the performance of each indicator decreases, indicating that the segmented training strategy can make the model learn more thoroughly.", + "bbox": [ + 496, + 611, + 890, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Parameter Analysis", + "text_level": 1, + "bbox": [ + 500, + 741, + 687, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the transition reasoning module, we verify through experiments that we need to infer the number of motion snippet codes from the previous motions at the current initial stage. As shown in Tab. 5, the reasoning motion snippet code number of 1, 2, 3, and 5 are compared. The experiment results show that two motion snippet codes at the beginning of reasoning are optimal because the model can not fully capture the influence of previous motion information by reducing motion snippets inferred from previous motions. At", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9505", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6fd40e06859ddfd216f4e4524a461523b3b2e08f254bd51269083df56b687dbc.jpg", + "image_caption": [ + "Figure 8: Qualitative evaluation: Continuously inputting four texts to show the visual results of our method and other baseline methods, except for the two frames between adjacent motions, the rest only show the keyframes. The results show that the motion synthesized by our method is the best in terms of semantic matching, transition fluency, and realism. More results are in the supplementary file." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 870, + 494 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the same time, the model will introduce too much previous motion information into the current motion if increasing the motion snippets inferred from previous motions.", + "bbox": [ + 75, + 578, + 470, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.6. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 628, + 287, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We make a qualitative comparison with the baselines, as shown in Fig. 8, our synthesized motion is the most natural in transition and nicely matches the semantics of the texts. Transition in T2M [11] is most evident in adjacent motion. T2M-Joint cannot fully express the content of the text description. When the input sequential texts contain more than two sentences, noticeable changes in the adjacent motion transition will occur. Complusion-Code produces a slight change in the transition. There is some semantic mismatch in TEACH [6] and abrupt transitions in the no Slerp version.", + "bbox": [ + 75, + 647, + 468, + 797 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusions and Future Works", + "text_level": 1, + "bbox": [ + 76, + 803, + 362, + 819 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We have advocated a novel framework for synthesizing semantic motion sequences driven by sequential texts while ensuring natural transitions between adjacent motions. Our approach leveraged a local-to-global semantic fusion module and a transition reasoning module to over", + "bbox": [ + 75, + 825, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "come the challenges of semantic association and transition fluency. Extensive experiments show that our method outperforms state-of-the-art techniques in terms of semantic matching, transition fluency, and realism.", + "bbox": [ + 496, + 578, + 890, + 638 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "However, our method still has limitations, such as difficulty in synthesizing head motions like \"nodding\" and \"shaking head\". In future work, we plan to improve this by exploring a way to assign different attention weights to different joints automatically. Additionally, since our STDM dataset is synthesized while not captured, we will further improve the data quality and expand the dataset scale.", + "bbox": [ + 496, + 638, + 890, + 744 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 753, + 668, + 771 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This paper is supported by National Natural Science Foundation of China (62272021, 62102036), Beijing Natural Science Foundation (4222024), R&D Program of Beijing Municipal Education Commission (KM202211232003), Open Project Program of State Key Laboratory of Virtual Reality Technology and Systems, Beihang University (No.VRLAB2022A02), Research Unit of Virtual Human and Virtual Surgery (2019RU004).", + "bbox": [ + 496, + 776, + 890, + 897 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9506", + "bbox": [ + 480, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mixamo. https://www MIXamo.com. Accessed February 17, 2023.", + "[2] CMU graphics lab motion capture database. http://mocap.cs.cmu.edu/, 2003.", + "[3] Chaitanya Ahuja, Dong Won Lee, Yukiko I.Nakano, et al. Style transfer for co-speech gesture animation: A multi-speaker conditional-mixture approach. European Conference on Computer Vision, pages 248-265, 2020.", + "[4] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. 2019 International Conference on 3D Vision (3DV), pages 719-728, 2019.", + "[5] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, et al. Style-controllable speech-driven gesture synthesis using normalising flows. Computer Graphics Forum, 39(2):487-496, 2020.", + "[6] Nikos Athanasiou, Mathis Petrovich, Michael J.Black, et al. TEACH: Temporal action composition for 3d humans. 2022 International Conference on 3D Vision (3DV), pages 414-423, 2022.", + "[7] Pablo Cervantes, Yusuke Sekikawa, Ikuro Sato, et al. Implicit neural representations for variable length human motion generation. European Conference on Computer Vision, pages 356-372, 2022.", + "[8] Jihoon Chung, Cheng hsin Wuu, Hsuan ru Yang, et al. HAA500: Human-centric atomic action dataset with curated videos. IEEE/CVF International Conference on Computer Vision, pages 13465-13474, 2021.", + "[9] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, et al. Synthesis of compositional animations from textual descriptions. IEEE/CVF International Conference on Computer Vision, pages 1396-1406, 2021.", + "[10] Shiry Ginosar, Amir Bar, Gefen Kohavi, et al. Learning individual styles of conversational gesture. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019.", + "[11] Chuan Guo, Shihao Zou, Xinxin Zuo, et al. Generating diverse and natural 3d human motions from text. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022.", + "[12] Chuan Guo, Xinxin Zuo, Sen Wang, et al. Action2motion: Conditioned generation of 3d human motions. 28th ACM International Conference on Multimedia, pages 2021-2029, 2020.", + "[13] Chuan Guo, Xinxin Zuo, Sen Wang, et al. Action2video: Generating videos of human 3d actions. International Journal of Computer Vision, 130(2):285-315, 2022.", + "[14] Chuan Guo, Xinxin Zuo, Sen Wang, et al. TM2T: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. European Conference on Computer Vision, pages 580-597, 2022.", + "[15] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, et al. Learning speech-driven 3d conversational gestures from video. 21st ACM International Conference on Intelligent Virtual Agents, pages 101-108, 2021." + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2:1735-1742, 2006.", + "[17] Yanli Ji, Feixiang Xu, Yang Yang, et al. A large-scale RGB-D database for arbitrary-view human action recognition. 26th ACM International Conference on Multimedia, pages 1510-1518, 2018.", + "[18] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. arXiv preprint arXiv:2209.00349, 2022.", + "[19] Taras Kucherenko, Patrik Jonell, Sanne van Waveren, et al. Gesticulator: A framework for semantically-aware speech-driven gesture generation. 2020 International Conference on Multimodal Interaction, pages 242-250, 2020.", + "[20] Yuanzhi Liang, Qianyu Feng, Linchao Zhu, et al. SEEG: Semantic energized co-speech gesture generation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10473-10482, 2022.", + "[21] Xiao Lin and Mohamed R Amer. Human motion modeling using dvgens. arXiv preprint arXiv:1804.10652, 2018.", + "[22] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, et al. BEAT: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. arXiv preprint arXiv:2203.05297, 2022.", + "[23] Jun Liu, Amir Shahroudy, Mauricio Perez, et al. NTU rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2684-2701, 2019.", + "[24] Xian Liu, Qianyi Wu, Hang Zhou, et al. Learning hierarchical cross-modal association for co-speech gesture generation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022.", + "[25] Matthew Loper, Naureen Mahmood, Naureen Mahmood, et al. SMPL: A skinned multi-person linear model. Acm Transactions on Graphics, 34(6cd):248, 2015.", + "[26] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. Weakly-supervised action transition learning for stochastic human motion prediction. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8151-8160, 2022.", + "[27] Dario Pavllo, David Grangier, and Michael Auli. Quaternet: A quaternion-based recurrent model for human motion. arXiv preprint arXiv:1805.06485, 2018.", + "[28] Mathis Petrovich, Michael J.Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021.", + "[29] Mathis Petrovich, Michael J.Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. arXiv preprint arXiv:2204.14109, 2022.", + "[30] Matthias Plappert, Christian Mandery, and Tamim Asfour. The KIT motion-language dataset. Big Data, 4(4):236-252, 2016.", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, et al. Learning transferable visual models from natural language supervision. International Conference on Machine Learning, pages 8748-8763, 2021." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9507", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Abhinanda R.Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, et al. BABEL: Bodies, action and behavior with english labels. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 722-731, 2021.", + "[33] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, et al. Nturogb+ d: A large scale dataset for 3d human activity analysis. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1010-1019, 2016.", + "[34] Guy Tevet, Brian Gordon, Amir Hertz, et al. MotionCLIP: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022.", + "[35] Sijie Yan, Zhizhong Li, Yuanjun Xiong, et al. Convolutional sequence generation for skeleton-based action synthesis. IEEE/CVF International Conference on Computer Vision, pages 4394-4402, 2019.", + "[36] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, et al. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics, 39(6):1-16, 2020.", + "[37] Mingyuan Zhang, Zhongang Cai, Liang Pan, et al. Motion-diffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022.", + "[38] Yan Zhang, Michael J.Black, and Siyu Tang. Perpetual motion: Generating unbounded human motion. arXiv preprint arXiv:2007.13886, 2020.", + "[39] Hang Zhao, Antonio Torralba, Lorenzo Torresani, et al. Hacs: Human action clips and segments dataset for recognition and temporal localization. IEEE/CVF International Conference on Computer Vision, pages 8668-8678, 2019.", + "[40] Yi Zhou, Connelly Barnes, Jingwan Lu, et al. On the continuity of rotation representations in neural networks. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2019." + ], + "bbox": [ + 78, + 90, + 470, + 559 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9508", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions/2c5b9c34-9105-42fb-b593-c16dc3db7c25_model.json b/2023/Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions/2c5b9c34-9105-42fb-b593-c16dc3db7c25_model.json new file mode 100644 index 0000000000000000000000000000000000000000..05220ca47a7729f7bd4dcb98096636c4c379dc4d --- /dev/null +++ b/2023/Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions/2c5b9c34-9105-42fb-b593-c16dc3db7c25_model.json @@ -0,0 +1,1916 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.002, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.131, + 0.877, + 0.155 + ], + "angle": 0, + "content": "Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.18, + 0.874, + 0.253 + ], + "angle": 0, + "content": "Shuai Li\\(^{1,3}\\), Sisi Zhuang\\(^{1}\\), Wenfeng Song\\(^{2*}\\), Xinyu Zhang\\(^{2}\\), Hejia Chen\\(^{1}\\), Aimin Hao\\(^{1}\\) \n\\(^{1}\\)State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, P.R. China \n\\(^{2}\\)Computer School, Beijing Information Science and Technology University, P.R. China \n\\(^{3}\\)Zhongguancun Laboratory, Beijing, P.R. China" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.255, + 0.735, + 0.288 + ], + "angle": 0, + "content": "{lishuai, sisizhuang}@buaa.edu.cn, songwenfenga@gmail.com zhangxinyul@bistu.edu.cn,{chenhj2000, ham}@buaa.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.298, + 0.891, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.404, + 0.893, + 0.436 + ], + "angle": 0, + "content": "Figure 1: Our method synthesizes semantic human motion with natural transition using free-form sequential texts. The skeleton motion can be easily transferred to any character. The character used for demonstration is from Mixamo [1]." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.448, + 0.314, + 0.463 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.477, + 0.474, + 0.871 + ], + "angle": 0, + "content": "The intelligent synthesis/generation of daily-life motion sequences is fundamental and urgently needed for many VR/metaverse-related applications. However, existing approaches commonly focus on monotonic motion generation (e.g., walking, jumping, etc.) based on single instruction-like text, which is still not intelligent enough and can't meet practical demands. To this end, we propose a cohesive human motion sequence synthesis framework based on free-form sequential texts while ensuring semantic connection and natural transitions between adjacent motions. At the technical level, we explore the local-to-global semantic features of previous and current texts to extract relevant information. This information is used to guide the framework in understanding the semantics of the current moment. Moreover, we propose learnable tokens to adaptively learn the influence range of the previous motions towards natural transitions. These tokens can be trained to encode the relevant information into well-designed transition loss. To demonstrate the efficacy of our method, we conduct extensive experiments and comprehensive evaluations on the public dataset as well as a new dataset produced by us. All the experiments confirm that our method outperforms the state-of-the-art methods in terms of semantic matching, realism, and transition fluency. Our project is public available. https://druthrie.github.io/sequential-texts-to-motion/" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.448, + 0.766, + 0.463 + ], + "angle": 0, + "content": "1. Introduction and Motivation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.893, + 0.717 + ], + "angle": 0, + "content": "Human motion synthesis is fundamental for numerous applications, especially for virtual reality, games, and metaverse-related applications [10, 19, 5, 3, 36, 15, 22, 20, 24], of which, it is in high demand to accurately control the digital human motion with natural language. Existing approaches commonly focus on monotonic motion synthesis/generation based on single instruction-like text description. However, practical applications usually require digital humans to respond to multiple rounds of sustainable interactions, wherein they can continuously generate reasonable responses to the sequential texts. Therefore, given a set of free-form sequential texts, we aim to synthesize a cohesive human motion sequence. Namely, the animation clips should be consistent with the ongoing text descriptions, and the whole motion sequence should have smooth semantic connection and natural transitions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "At present, sequential texts-driven motion synthesis has not been well studied, mainly due to lacking long-term continuous motion datasets with accompanying free-form text descriptions. Recently, TEACH [6] first attempts to address this problem by proposing a dataset (we refer to it as BABEL-TEACH). Each item in the BABEL-TEACH dataset contains two adjacent text-motion pairs. However, TEACH [6] builds on the one-time method TEMOS [29], wherein the limited 5-frame motion is added to encode together. However, it does not consider the potential benefits of previous text information. In practice, previous texts can provide extra and more accurate semantic information for" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.227, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9498" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "current motion synthesis. Therefore, global semantic needs to be extracted from the previous text and the local semantics from the current text, and then fuses the semantics to guide the semantic understanding at each moment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.438 + ], + "angle": 0, + "content": "Besides, there are still two main challenges. Firstly, the semantic relationship between adjacent texts is largely overlooked in previous research. For example, when synthesizing a motion for the current text \"touching the face with left hand,\" the previous text context should be considered. Specifically, if the previous text was \"sitting on a chair,\" the synthesized motion should depict a person sitting and touching face with left hand (shown in Fig. 2); if the previous text was \"a person kicks a ball with right foot,\" the motion should show a standing person touching face with left hand. This requires a more comprehensive understanding about the context and semantic relationships between the texts. Secondly, existing methods tend to abrupt transitions between adjacent motions when multiple motions are synthesized separately and stitched together. We aim to address this issue by seamlessly blending the synthesized motions. It should intelligently capture the temporal and spatial continuity between the motions to ensure a more natural and coherent sequence of motions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.47, + 0.649 + ], + "angle": 0, + "content": "Considering the easy scalability of autoregressive methods, we should extend previous autoregressive single text-driven motion synthesis methods to multiple motions involved in motion sequence synthesis. We observe that a similar human body posture can be achieved by transferring the end features of the previous motion to the next motion, although this is not always entirely consistent. This suggests that previous motion information should be leveraged more effectively and sufficiently rather than being adopted straightforwardly. Therefore, we plan to introduce a transition reasoning module that adaptively learns attention score from the previous motion information and infers accurate motion features at the transition, making the transition between adjacent motions more natural and smooth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.47, + 0.786 + ], + "angle": 0, + "content": "To this end, we create a new dataset with much longer-term motions (2-5 times longer) based on a synthetic method, and the corresponding text descriptions are more abundant. We conduct experiments on both BABEL-TEACH [6] dataset and our sequential texts described motion (STDM) dataset. The results demonstrate that our method outperforms existing methods in terms of semantic matching, realism, and transition fluency. Specially, the salient contributions can be summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.792, + 0.468, + 0.867 + ], + "angle": 0, + "content": "- We propose a cohesive motions synthesis framework using sequential texts as inputs. The framework can integrate previous text information to obtain semantic features and adaptively select valid previous motion information to guide the current motion synthesis." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.871, + 0.469, + 0.899 + ], + "angle": 0, + "content": "- We design a local-to-global (L2G) semantic fusion module to extract accurate contextual information. It" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.792, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.875, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.216, + 0.892, + 0.261 + ], + "angle": 0, + "content": "Figure 2: Top row: Motion synthesis corresponding to sequential texts by our method. Bottom row: Motion synthesis corresponding to single text by T2M [11]." + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.272, + 0.892, + 0.332 + ], + "angle": 0, + "content": "can well take into account global-context information from the previous text to infer the current moment's text meaning to guide the semantics-consistent motion synthesis." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.335, + 0.892, + 0.424 + ], + "angle": 0, + "content": "- We introduce a transition reasoning module to adaptively select motion snippet from previous motion information to make the synthesized motion natural and smooth w.r.t the previous, wherein a well-defined transition loss is employed to further constrain the fluency of the transitions." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.429, + 0.892, + 0.488 + ], + "angle": 0, + "content": "- We create a new sequential texts described motion (STDM) dataset, which involves more extended motion frames and more diverse text descriptions than the existing dataset." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.335, + 0.892, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.498, + 0.641, + 0.513 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.524, + 0.892, + 0.583 + ], + "angle": 0, + "content": "We briefly introduce the related works of human motion synthesis considering semantic control, including single text driven motion synthesis and sequential texts driven motion synthesis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Single text driven motion synthesis. The motion synthesis from text began with the synthesized motion from action labels, but this approach was limited in covering most human motions [21, 27, 35, 38, 28, 12, 28, 7]. Later, researchers [4, 9, 34] focused on synthesizing motion from complex free-form text descriptions, with early work generating single, deterministic motions from text. Recent works have aimed to improve the diversity and details of the synthesized motion. For example, TEMOS [29] and T2M [11] utilized VAE architecture to synthesize diverse motions from the same texts. The main difference between them is that the former utilized Transformer architecture to synthesize all moment motions at once, while the latter utilized GRU architecture to synthesize every moment motion auto-regressively. To capture more detailed information in the text, TM2T [14] utilized VQVAE-based motion markers to provide a fair environment when considering motion and text signals, and utilized motion-to-text analysis module to strengthen the constraints on text-to-motion synthesis. With the successful application of the diffusion model in image generation, MotionDiffuse [37] utilized the diffusion model" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9499" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.089, + 0.885, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.328, + 0.896, + 0.404 + ], + "angle": 0, + "content": "Figure 3: Method Overview: Our model takes sequential texts as input. In the local-to-global (L2G) semantic fusion module, the previous text's global features and the current text's local features are fused and extracted. In the transition reasoning module, the motion snippet code at the transition of adjacent motion (\\(\\mathbf{c}_{start}^{0}\\) and \\(\\mathbf{c}_{start}^{1}\\)) is deduced from the previous motion. Input the gate snippet code \\(\\mathbf{g}^t\\) and the semantic text feature \\(F_{sem}^{t}\\) to synthesize the motion snippet code \\(\\hat{\\mathbf{c}}_i^t\\). The transition loss further limits the fluency of the transition with the previous motion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.471, + 0.552 + ], + "angle": 0, + "content": "to simulate text driven conditional human motion synthesis, which can better respond to fine-grained instructions of body parts. FLAME [18] utilized transformer-based diffusion model architecture to synthesize human motion, and it can allow editing frames and joints without fine-tuning. Its editing ability can be extended to motion prediction or in-between tasks. Although the diffusion model improves the quality of the synthesized motion, its inference speed is relatively slow, so it is unsuitable for real-time applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Sequential texts driven motion synthesis. Synthesizing motion sequences from sequential texts is challenging due to the limited availability of long-term continuous motions datasets with text descriptions. Previous works Action2video [13] realized the transition between actions by changing the input action label type in an autoregressive synthesizer. Yet, they only conducted experiments on a limited number of action labels in three datasets: NTURGBD [33], CMU [2], and humanACT12 [12]. Mao et al. [26] proposed a weakly-supervised action-driven motion prediction method, but they only used 20 action categories with clear transitions in BABEL [32] to synthesize the following motion and transition. Based on the existing BABEL dataset [32], TEACH [6] proposed a new dataset, in which each data consisted of two adjacent text descriptions and motion sequences. The model of TEACH [6] encoded five previous motion frames using TEMOS [29], every single motion sequence was synthesized simultaneously, and adjacent motion sequences were auto-regressively synthesized. Our work builds on this by proposing a cohesive human motion sequence synthesis framework based on freeform sequential texts, ensuring semantic connection and natural transitions between adjacent motions." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.415, + 0.651, + 0.433 + ], + "angle": 0, + "content": "3. Our Approach" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.439, + 0.676, + 0.455 + ], + "angle": 0, + "content": "3.1. Method Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.895, + 0.825 + ], + "angle": 0, + "content": "The overall pipeline of the proposed model is shown in Fig. 3. The temporal VAE and motion autoencoder used in the model are similar to those used in the T2M model, wherein it is employed to encode the motion sequence into a motion snippet code sequence and reconstruct the motion sequence with a decoder. To take sequential texts as input and ensure semantically coherent synthesized motions, a local-to-global (L2G) semantic fusion module is introduced (Sec. 3.2). This module guides learning accurate semantic features from the input texts at each moment, including historical text information. To ensure smooth and natural transitions between adjacent motions, a transition reasoning module is introduced (Sec. 3.3). This module dynamically deduces the motion snippet code during transition based on previous motion information, which can help the model generate realistic and coherent transitions. Additionally, a transition loss is proposed to enhance the smoothness of the transition motion (Sec. 3.4). Overall, the proposed pipeline combines several modules and techniques to generate realistic and semantically coherent motion sequences from sequential texts input. The model can generate smooth transitions between motions, and can adaptively learn the influence range of previous motions to further improve the quality of the synthesized motion sequences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Preliminaries. Our method aims to take freeform sequential texts \\(\\mathbf{S} = (\\mathbf{S}_1, \\mathbf{S}_2, \\dots, \\mathbf{S}_n)\\) as input, e.g., \\(\\mathbf{S} =\\) (\"salute with the left hand\", \"salute with the right hand\", \"sit on the chair\", \"crawl\"), and outputs corresponding human motion sequences \\(\\mathbf{M} =\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "9500" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.242 + ], + "angle": 0, + "content": "\\((\\mathbf{M}_1, \\mathbf{M}_2, \\dots, \\mathbf{M}_n)\\). The transition between adjacent motions should be natural and smooth, and the synthesized motion should reflect the semantic information of previous texts. The data format is consistent with HumanML3D [11] dataset, including the root angular velocity along the Y-axis, the root linear velocity on the XZ-plane, the root height, the local joints positions, velocities, and rotations [40] in root space, while the motions follow the skeleton structure of SMPL [25] with 22 joints. Each moment in our model represents 4 frames." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.25, + 0.427, + 0.266 + ], + "angle": 0, + "content": "3.2. Local-to-global Semantic Fusion Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.473, + 0.501 + ], + "angle": 0, + "content": "To synthesize motion that accurately matches the semantics of sequential texts, our model relies on accurately extracting text features at each moment to guide the synthesis process. As the motion we synthesize is based on sequential texts, the current motion is determined not only by the current text semantics but also by previous text semantics. To capture the temporal clues, we encode both the previous and current text, and use an attention mechanism to extract text features that integrate the previous text semantics at each moment (shown in Fig. 4). For previous texts, we use a pretrained CLIP [31] text encoder to extract global features of previous texts. In order to make the features more suitable for our task in the process of training, linear projector is used to extract features further. The global features from the previous texts are expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.507, + 0.47, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} _ {p r e} = f \\left(C L I P _ {t e x t} \\left(\\mathbf {S} _ {i - 1}\\right)\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.529, + 0.473, + 0.754 + ], + "angle": 0, + "content": "It is difficult for the CLIP text encoder to extract the local features of the text, motion synthesis of each moment requires local text semantic guidance, so we employ a Bi-GRU-based text encoder, similar to T2M [11], to obtain the local features \\(\\mathbf{L}_{1:m}\\) from the current text. To capture accurate text semantic features at the arbitrary frames, we use an attention mechanism, wherein the key \\(\\mathbf{K}(L)\\) and value \\(\\mathbf{V}(L)\\) of attention are the local features of the current text. Since the semantic information of the previous text is required for guidance at the beginning of the currently synthesizing motion, we use the global features of the previous text as the initial query \\(\\mathbf{Q}(G)\\). At subsequent moments, the motion snippet code generator in temporal VAE provides the queries \\(\\mathbf{Q}(L)\\). This can be expressed mathematically as:" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.758, + 0.404, + 0.776 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} (G) = \\mathbf {G} _ {p r e} \\mathbf {W} ^ {Q}, \\mathbf {Q} (L) = \\mathbf {h} ^ {t - 1} \\mathbf {W} ^ {Q},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.779, + 0.469, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K} (L) = \\mathbf {L} _ {1: m} \\mathbf {W} ^ {K}, \\mathbf {V} (L) = \\mathbf {L} _ {1: m} \\mathbf {W} ^ {V}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.801, + 0.401, + 0.835 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {s e m} ^ {t} = f \\left(\\frac {\\mathbf {Q} (G / L) \\mathbf {K} (L) ^ {T}}{\\sqrt {d _ {s e m}}}\\right) \\mathbf {V} (L),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.839, + 0.472, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}^k, \\mathbf{W}^V \\in \\mathbb{R}^{d_w \\times d_{sem}}\\) and \\(\\mathbf{W}^Q \\in \\mathbb{R}^{d_h \\times d_{sem}}\\) are trainable weights; \\(d_h, d_w\\) and \\(d_{sem}\\) are the number of channels in hidden unit \\(\\mathbf{h}^{t-1}\\), current text feature \\(\\mathbf{L}^{1:m}\\) and semantic fusion attention layer, respectively. \\(\\mathbf{h}^t\\) is the hidden" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.887, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.316, + 0.895, + 0.395 + ], + "angle": 0, + "content": "Figure 4: Local-to-global semantic fusion module: After extracting the global feature \\(\\mathbf{G}_{pre}\\) from the previous texts and the local features \\(\\mathbf{L}_{1:m}\\) from the current text, feature fusion is performed to obtain the accurate semantic feature \\(\\mathbf{F}_{sem}^{t}\\) of the current text at moment \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.416, + 0.893, + 0.462 + ], + "angle": 0, + "content": "unit generated by the motion snippet code generator at each moment and \\(\\mathbf{h}^0\\) is obtained from \\(\\mathbf{G}_{pre}\\). \\(\\mathbf{F}_{sem}^{t}\\) is the text semantic feature at moment \\(t\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.471, + 0.766, + 0.487 + ], + "angle": 0, + "content": "3.3. Transition Reasoning Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.495, + 0.895, + 0.782 + ], + "angle": 0, + "content": "When synthesizing a new motion, the starting pose is influenced by the ending pose of the previous motion. Grafting the end motion snippet code of the previous motion to the start moment of the current motion can produce similar human poses during the transition. However, the transition between the two motions is still unnatural and abrupt. To address this, we introduce a learnable start motion snippet code token \\(\\mathbf{c}_{\\text {token }}\\) and utilize it along with the previous motion snippet code sequence in a transformer encoder. The output of the token position is used as the start motion snippet code \\(\\mathbf{c}_{\\text {start }}\\) for the current motion, allowing the model to learn the range and weight of the previous motions (shown in Fig. 5). Additionally, the length of the current motion affected by the previous motion information needs to be determined. Through experimentation, we obverse that the best performance is achieved by reasoning the first two motion snippet codes \\((\\mathbf{c}_{\\text {start }}^{0}\\) and \\(\\mathbf{c}_{\\text {start }}^{1})\\) of the current motion from the previous motions. The verification experiment is in Sec. 4.5. The mathematical expression is as:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.793, + 0.891, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {\\text {s t a r t}} ^ {0}, \\mathbf {c} _ {\\text {s t a r t}} ^ {1} = \\text {T r a n s f o r m e r} \\left(\\hat {\\mathbf {c}} _ {i - 1}, \\mathbf {c} _ {\\text {t o k e n}} ^ {0}, \\mathbf {c} _ {\\text {t o k e n}} ^ {1}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As shown by the transition gate in Fig. 3, when synthesizing the motion snippet codes of the first two moments, we choose the motion snippet codes \\((\\mathbf{c}_{start}^{0}\\) and \\(\\mathbf{c}_{start}^{1})\\) inferred from the previous motion as gate snippet code \\(\\mathbf{g}^t\\). From the third moment, the motion snippet code \\(\\hat{\\mathbf{c}}^{t - 1}\\) from the previ" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "9501" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.088, + 0.434, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.385, + 0.473, + 0.493 + ], + "angle": 0, + "content": "Figure 5: Transition reasoning module: Given the previous texts to get the previous synthesized motion sequence \\(\\hat{\\mathbf{M}}_{i-1}\\) and motion snippet codes \\(\\hat{\\mathbf{c}}_{i-1}\\), we use two learnable tokens (\\(\\mathbf{c}_{token}^{0}\\) and \\(\\mathbf{c}_{token}^{1}\\)) to learn the previous motion information that needs attention, so as to obtain the motion snippet codes at the starting moment of the current motion (\\(\\mathbf{c}_{start}^{0}\\) and \\(\\mathbf{c}_{start}^{1}\\))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.47, + 0.545 + ], + "angle": 0, + "content": "ous moment is used, which is synthesized autoregressively. For gate snippet code \\(\\mathbf{g}^t\\) , the mathematical expression is as:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.553, + 0.47, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\mathbf {g} ^ {t} = \\left\\{ \\begin{array}{l l} \\mathbf {c} _ {\\text {s t a r t}} ^ {0} & t = 1 \\\\ \\mathbf {c} _ {\\text {s t a r t}} ^ {1} & t = 2 \\\\ \\hat {\\mathbf {c}} ^ {t - 1} & t \\geq 3 \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.471, + 0.667 + ], + "angle": 0, + "content": "When synthesizing the motion based on the first text \\(\\mathbf{S}_1\\) of the sequential texts, we set the previous texts as the default text \\(\\mathbf{S}_0 = \\text{\"start\"}\\), and the starting snippet code \\(\\mathbf{c}_{start}^0\\) is obtained from the mean pose." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.675, + 0.337, + 0.691 + ], + "angle": 0, + "content": "3.4. Transition Loss & Total Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.698, + 0.47, + 0.79 + ], + "angle": 0, + "content": "We conduct a statistical analysis of the dataset and find that the motion gap (L1 distance) of almost all adjacent frames is less than 0.1. Inspired by this, we design a transition loss \\( L_{Trans} \\) to reduce the motion gap between the adjacent frames and avoid abrupt transition. The representation of the \\( L_{Trans} \\) is" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.796, + 0.47, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {T r a n s}} = \\left\\| \\hat {\\mathbf {M}} _ {i} ^ {1} - \\hat {\\mathbf {M}} _ {i - 1} ^ {T _ {i} - 1} \\right\\| _ {1}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The final loss function of our model consists of 4 parts: motion snippet code reconstruction loss \\(\\mathcal{L}_{rec}^{code}\\), motion reconstruction loss \\(\\mathcal{L}_{rec}^{mot}\\), KL loss \\(\\mathcal{L}_{KL}\\) of prior distribution and posterior distribution in temporal VAE, and the transition loss \\(\\mathcal{L}_{Trans}\\). The \\(\\lambda_{code},\\lambda_{mot},\\lambda_{KL},\\lambda_{Trans}\\) are the" + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.089, + 0.885, + 0.138 + ], + "angle": 0, + "content": "
DatasetQuantityDurationVocab.
STDM528920.02 h2455
BABEL-TEACH[6]1626621.86 h1037
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.149, + 0.894, + 0.195 + ], + "angle": 0, + "content": "Table 1: Comparison of datasets: STDM dataset has longer-term frames within a single motion, and the text is more abundant." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.219, + 0.744, + 0.234 + ], + "angle": 0, + "content": "corresponding weights of the losses." + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.244, + 0.893, + 0.277 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {\\text {c o d e}} \\mathcal {L} _ {\\text {r e c}} ^ {\\text {c o d e}} + \\lambda_ {\\text {m o t}} \\mathcal {L} _ {\\text {r e c}} ^ {\\text {m o t}} + \\lambda_ {K L} \\mathcal {L} _ {K L} + \\lambda_ {\\text {T r a n s}} \\mathcal {L} _ {\\text {T r a n s}}. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.284, + 0.634, + 0.301 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.308, + 0.893, + 0.4 + ], + "angle": 0, + "content": "First, we introduce our STDM dataset (Sec. 4.1), evaluation metrics (Sec. 4.2), and then conduct quantitative (Sec. 4.3) and qualitative (Sec. 4.6) evaluations. We carry out an ablation study (Sec. 4.4) to prove the effectiveness of our proposed modules and analyze the parameters (Sec. 4.5)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.41, + 0.652, + 0.424 + ], + "angle": 0, + "content": "4.1. STDM Dataset" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.433, + 0.893, + 0.524 + ], + "angle": 0, + "content": "Each text-motion pair in most existing human motion datasets [30, 17, 23, 39, 12, 8] is independent of the other, and does not contain adjacent text-motion pairs. In a recent study, TEACH [6] used the BABEL [32] dataset to obtain a BABEL-TEACH dataset containing two adjacent text-motion pairs, which is consistent with our research goal." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.895, + 0.841 + ], + "angle": 0, + "content": "To obtain adequate motion sequences, we synthesize sequential texts and motion sequences by modifying the existing single text driven motion synthesis method from the viewpoint of data generation instead of using a mocap device. By randomly selecting and annotating texts from a large pool, we create sequential text fed into our improved T2M [11], generating motion sequences. Based on T2M [11], we find that the motion sequences decoded by the same motion snippet code are similar. Our approach diverges from T2M: In the inference stage, instead of using a fixed mean pose, we use snippet codes from the last eight frames of the previous motion, ensuring smoother and more natural transitions. Our experiment results show that the motion sequences synthesized in this way are similar in transition. Since the motion generated by the generation model may not match the semantics of the input text and the transition may not be smooth, we manually remove data with poor transitions and correct erroneous texts to maintain the quality of our STDM dataset. After filtering, we get 5289 pairs containing two adjacent text-motion. Our dataset is collected under the same structure as HumanML3D [11]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.841, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Compared with the BABEL-TEACH [6] dataset, STDM has longer-term motion frames, with a single motion sequence ranging from 2 to 10 seconds, while most BABEL-TEACH [6] data is less than 3 seconds. However, the text" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9502" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.156, + 0.089, + 0.818, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.328, + 0.895, + 0.361 + ], + "angle": 0, + "content": "Figure 6: STDM Dataset: Two examples in the STDM dataset, which involves more extended motion frames and more diverse text descriptions than the BABEL-TEACH [6] dataset. Blue traces represent the digital human's trajectory." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.378, + 0.473, + 0.469 + ], + "angle": 0, + "content": "descriptions in BABEL-TEACH [6] usually describe only one action, such as \"walk forward\". The text descriptions in the STDM are more abundant, such as \"a person raises his right hand and walks forward\". A detailed comparison of the two datasets can be found in Tab. 1. Fig. 6 shows two examples of our STDM dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.489, + 0.262, + 0.505 + ], + "angle": 0, + "content": "4.2. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.47, + 0.669 + ], + "angle": 0, + "content": "In order to evaluate the quality of the synthesized motion sequence, we use the feature extractor proposed by Guo et al. [11] to extract the features of text and motion. This feature extractor is trained with contrast loss [16], which can make the matched text-motion pairs produce similar feature vectors. Therefore, the feature extractor can measure whether the text and motion sequence match. We combine two adjacent text-motion pairs into one to train the feature extractor. We use the same evaluation metrics as T2M [11]. The metrics are described in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.471, + 0.749 + ], + "angle": 0, + "content": "To demonstrate the fluency of different methods when transiting between adjacent motions, we add a metric to evaluate the fluency of transition, by calculating the distance between adjacent frames of two adjacent motion sequences. For transition fluency, the mathematical expression is as:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.766, + 0.47, + 0.808 + ], + "angle": 0, + "content": "\\[\nF = \\sum_ {i = 1} ^ {N} \\| \\mathbf {M} _ {i 1} ^ {T 1} - \\mathbf {M} _ {i 2} ^ {1} \\| _ {1}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.906 + ], + "angle": 0, + "content": "Here, we calculate the average transition distance (\\(L_{1}\\) distance) of the motion sequence synthesized by \\(N\\) adjacent text pairs. For the \\(i\\)-th adjacent text pair, the synthesized adjacent motion sequence pair is \\(\\mathbf{M}_{i1} = (\\mathbf{M}_{i1}^{1}, \\mathbf{M}_{i1}^{2}, \\dots, \\mathbf{M}_{i1}^{T1})\\) and \\(\\mathbf{M}_{i2} = (\\mathbf{M}_{i2}^{1}, \\mathbf{M}_{i2}^{2}, \\dots, \\mathbf{M}_{i2}^{T2})\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.377, + 0.722, + 0.393 + ], + "angle": 0, + "content": "4.3.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.689 + ], + "angle": 0, + "content": "Baselines. We compare our method with T2M [11], our two variants T2M-Joint and Complusion-Code, and TEACH [6]. For T2M, we input a single text, output a single motion, and align the root coordinates of the last frame of the previous motion with the root coordinates of the first frame of the next motion. For T2M-Joint, we keep the T2M model structure and train the model by combining two adjacent texts with commas as input. For Complusion-Code, based on the T2M [11], the last motion snippet code of the previous motion is used as the start motion snippet code of the next motion during training (Unlike the STDM dataset generation method, the dataset generation only uses the snippet codes of previous motions in the inference phase. Complusion-Code also uses the snippet codes of previous motions in the training phase). For TEACH [6], we compared two versions using spherical linear interpolation (Slerp) and not using Slerp, respectively. Slerp operation is unrelated to the TEACH [6] model and is simply a way of post-processing synthesized data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Tab. 2 and Tab. 3 show our quantitative comparison results with other baselines on BABEL-TEACH [6] and STDM datasets. We divide the training, testing, and validation sets according to the ratio of \\(0.8:0.15:0.05\\). For a fair comparison, we conduct 20 experiments and show the values with confidence in the \\(95\\%\\) range. From the results, our method outperforms all other methods, especially in the transition fluency metric, which is important for synthesizing natural and smooth transition motion. T2M [11] can only synthesize a single motion per inference, and there is no connection between adjacent motions, so it performs poorly in all indicators. Because our testing data are composed of two adjacent texts, the T2M-Joint method performs similarly to the real data regarding transition fluency" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9503" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.143, + 0.089, + 0.825, + 0.234 + ], + "angle": 0, + "content": "
MethodsR Precision↑FID↓MultiModal Dist↓Transition Fluency↓
Top 1Top 2Top 3
Real motions0.560±0.0020.750±0.0030.834±0.0020.001±0.0003.448±0.001-
T2M [11]0.376±0.0030.543±0.0020.647±0.0036.890±0.0815.218±0.0101.108±0.004
T2M-Joint0.373±0.0040.534±0.0030.640±0.0032.727±0.0324.983±0.011-
Compulsion-Code0.432±0.0030.613±0.0040.717±0.0033.726±0.0684.612±0.0230.348±0.002
TEACH (no Slerp) [6]0.573±0.0020.756±0.0030.842±0.0022.263±0.0553.487±0.0100.577±0.003
TEACH (Slerp) [6]0.563±0.0030.750±0.0030.839±0.0032.240±0.0533.526±0.0110.118±0.001
Ours0.542±0.0030.728±0.0030.818±0.0031.628±0.0313.662±0.0100.177±0.001
Ours (Slerp)0.519±0.0040.705±0.0040.803±0.0031.548±0.0243.797±0.0120.036±0.000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.237, + 0.895, + 0.3 + ], + "angle": 0, + "content": "Table 2: Quantitative evaluation on testing data of BABEL-TEACH: ± indicates 95% confidence interval, ↑ and ↓ respectively denotes better performance with larger or lower value. Bold indicates the optimal result, the underscore represents the suboptimal result, while wave line refers to the third best. The results show that the motion synthesized by our model outperforms other baselines in terms of semantic matching, transition fluency, and realism." + }, + { + "type": "table", + "bbox": [ + 0.143, + 0.312, + 0.825, + 0.461 + ], + "angle": 0, + "content": "
MethodsR Precision↑FID↓MultiModal Dist↓Transition Fluency↓
Top 1Top 2Top 3
Real motions0.322±0.0040.505±0.0050.625±0.0060.014±0.0023.461±0.0050.454±0.000
T2M [11]0.322±0.0060.493±0.0050.614±0.0051.389±0.0493.500±0.0090.639±0.003
T2M-Joint0.304±0.0050.488±0.0060.613±0.0050.744±0.0423.492±0.011-
Compulsion-Code0.285±0.0070.450±0.0060.572±0.0063.146±0.0803.806±0.0170.330±0.003
TEACH (no Slerp) [6]0.318±0.0060.505±0.0050.634±0.0051.414±0.0553.483±0.0120.246±0.002
TEACH (Slerp) [6]0.326±0.0040.504±0.0050.631±0.0061.416±0.0543.487±0.0130.049±0.000
Ours0.328±0.0060.510±0.0060.633±0.0051.085±0.0633.441±0.0170.109±0.001
Ours (Slerp)0.325±0.0070.515±0.0070.637±0.0051.134±0.0593.440±0.0130.022±0.000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.464, + 0.893, + 0.496 + ], + "angle": 0, + "content": "Table 3: Quantitative evaluation on the testing data of STDM: The results show that the motion synthesized by our model outperforms other baselines in terms of semantic matching, transition fluency, and realism." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.514, + 0.473, + 0.816 + ], + "angle": 0, + "content": "(less than 0.1). However, when the input text consists of more than two sentences, the T2M-Joint method cannot guarantee transition fluency. Because Compulsion-Code uses the end motion snippet code of the previous motion as the start snippet code of the next motion, the performance of transition fluency is better than T2M [11], but it is still not as good as ours. Although TEACH [6] is slightly better than ours in R Precision and MultiModal Distance on the BABEL-TEACH [6] dataset, as shown in Tab. 2, its transition fluency is not as good as ours, we can obtain transitional and natural motion sequences without Slerp operation, so our method is better than TEACH [6] in a comprehensive evaluation. Although the synthetic STDM dataset is less natural than actual mocap data in transition fluency, we find that our method can compensate for the defects of the training data after training. In terms of transition fluency, the results synthesized by our method outperform the ground truth of the testing data. This further verifies the effectiveness of our method in synthesizing natural and smooth transition motions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.473, + 0.903 + ], + "angle": 0, + "content": "User study. We randomly select 50 texts from the testing data and arrange 3 to 5 texts together to form a set of sequential texts. Some text combinations are not present in the dataset to test the motion generation effect of each model for text combinations not in the dataset. The syn" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.514, + 0.895, + 0.637 + ], + "angle": 0, + "content": "thesized motions based on these sequential text groups are randomly shuffled and presented using different methods. Users are required to score synthetic motions from the following three aspects: (1) Matching degree: the degree of semantic matching between the motion and the text; (2) Transition fluency: the natural degree of transition between adjacent motions; (3) Realism: how realistic the motion is. The result is shown in Fig. 7." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.648, + 0.655, + 0.665 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.895, + 0.734 + ], + "angle": 0, + "content": "In order to illustrate the influence of the modules and training strategies we introduced on the overall model, we carried out the ablation experiment. The results are shown in Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.896, + 0.903 + ], + "angle": 0, + "content": "The first experiment is to remove the transition reasoning module. We set the skeleton to mean pose to initialize the motion snippet code at the start moment. However, the model does not know the human posture at the beginning, so it greatly reduces transition fluency (value increased by 0.417). The experiment results show that the performance of the model will decline without the transition reasoning module because it does not aware of the information of previous motion when the model synthesizes the motion sequence of the current text, even though the model can synthesize motion sequences that match the semantics of the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "9504" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.119, + 0.089, + 0.85, + 0.204 + ], + "angle": 0, + "content": "
MethodsR Precision↑FID↓MultiModal Dist↓Transition Fluency↓
Top 1Top 2Top 3
Real motions0.561±0.0030.747±0.0030.836±0.0020.002±0.0003.449±0.002-
w/o Transition reasoning module0.535±0.0040.713±0.0030.805±0.0032.190±0.0383.719±0.0130.594±0.003
w/o L2G semantic fusion module0.440±0.0030.621±0.0030.720±0.0034.592±0.0714.593±0.0130.151±0.001
w/o Transition loss0.525±0.0040.706±0.0040.801±0.0021.766±0.0363.797±0.0110.293±0.001
w/o Segmented training strategy0.500±0.0020.684±0.0040.783±0.0042.164±0.0384.041±0.0130.328±0.002
Ours0.542±0.0030.726±0.0030.817±0.0021.588±0.0303.655±0.0130.177±0.001
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.211, + 0.895, + 0.272 + ], + "angle": 0, + "content": "Table 4: Ablation study: The results show that the transition reasoning module has the most significant impact on transition fluency, the L2G semantic fusion module has the most significant impact on R Precision, FID, and MultiModel Distance, which represent semantic matching and realism, and the transition loss and segmentation training strategy have a certain influence on all indicators." + }, + { + "type": "table", + "bbox": [ + 0.137, + 0.282, + 0.832, + 0.381 + ], + "angle": 0, + "content": "
Numbers of snippet codeR Precision↑FID↓MultiModal Dist↓Transition Fluency↓
Top 1Top 2Top 3
Real motions0.561±0.0030.750±0.0020.834±0.0020.002±0.0003.449±0.001-
1 snippet code0.517±0.0030.693±0.0040.786±0.0042.027±0.0433.845±0.0140.173±0.001
2 snippet codes0.540±0.0030.725±0.0030.816±0.0021.602±0.0263.662±0.0110.177±0.001
3 snippet codes0.522±0.0040.702±0.0030.796±0.0022.050±0.0383.838±0.0150.175±0.001
5 snippet codes0.504±0.0030.687±0.0040.785±0.0031.781±0.0293.913±0.0130.213±0.001
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.388, + 0.895, + 0.435 + ], + "angle": 0, + "content": "Table 5: Parameter analysis of transition reasoning module: We infer the initial motion snippet code of the current motion sequence from the previous motion and determine the optimal number of the inference code through experiments. The result shows that inferring the first two codes of current motion from the previous motion work best." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.445, + 0.457, + 0.661 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.665, + 0.47, + 0.741 + ], + "angle": 0, + "content": "Figure 7: User study: We invite 45 users to score the results of our comparative experiments, and the color bars in the figure indicate the percentage of the scores. The results show that our method outperforms other baselines in terms of semantic matching, transition fluency, and realism." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.751, + 0.16, + 0.763 + ], + "angle": 0, + "content": "current text." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The second experiment is to remove the L2G semantic fusion module. We only encode the features of the current text at each moment without considering the previous text. The experiment results show that the model's performances on R Precision, FID, and MultiModal Distance decrease significantly after removing the L2G semantic fusion module, indicating that previous text information is essential in synthesizing motion sequences that match context semantics. Although there is a slight improvement in the metric of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.892, + 0.52 + ], + "angle": 0, + "content": "transition fluency after the removal, the gap between them can not be observed from the naked eye. With the L2G semantic fusion module, our model shows better competence in evaluating the implicit relationship between current and previous text." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.61 + ], + "angle": 0, + "content": "The third experiment is to remove the transition loss. The experimental results show that after removing the transition loss, the model's performances in all indicators have decreased, indicating that adding transition loss will guide the model to learn more semantic matching and transitional natural motion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.734 + ], + "angle": 0, + "content": "The fourth experiment does not use the segmented training strategy (explained in the supplementary file). The specific approach is that all motion sequences are extended to the maximum length (196 frames) with 0 in the training strategy. The experimental results show that when the segmented training strategy is not used, the performance of each indicator decreases, indicating that the segmented training strategy can make the model learn more thoroughly." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.688, + 0.758 + ], + "angle": 0, + "content": "4.5. Parameter Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the transition reasoning module, we verify through experiments that we need to infer the number of motion snippet codes from the previous motions at the current initial stage. As shown in Tab. 5, the reasoning motion snippet code number of 1, 2, 3, and 5 are compared. The experiment results show that two motion snippet codes at the beginning of reasoning are optimal because the model can not fully capture the influence of previous motion information by reducing motion snippets inferred from previous motions. At" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9505" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.872, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.502, + 0.895, + 0.565 + ], + "angle": 0, + "content": "Figure 8: Qualitative evaluation: Continuously inputting four texts to show the visual results of our method and other baseline methods, except for the two frames between adjacent motions, the rest only show the keyframes. The results show that the motion synthesized by our method is the best in terms of semantic matching, transition fluency, and realism. More results are in the supplementary file." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.471, + 0.625 + ], + "angle": 0, + "content": "the same time, the model will introduce too much previous motion information into the current motion if increasing the motion snippets inferred from previous motions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.629, + 0.288, + 0.644 + ], + "angle": 0, + "content": "4.6. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.648, + 0.47, + 0.799 + ], + "angle": 0, + "content": "We make a qualitative comparison with the baselines, as shown in Fig. 8, our synthesized motion is the most natural in transition and nicely matches the semantics of the texts. Transition in T2M [11] is most evident in adjacent motion. T2M-Joint cannot fully express the content of the text description. When the input sequential texts contain more than two sentences, noticeable changes in the adjacent motion transition will occur. Complusion-Code produces a slight change in the transition. There is some semantic mismatch in TEACH [6] and abrupt transitions in the no Slerp version." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.804, + 0.363, + 0.82 + ], + "angle": 0, + "content": "5. Conclusions and Future Works" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We have advocated a novel framework for synthesizing semantic motion sequences driven by sequential texts while ensuring natural transitions between adjacent motions. Our approach leveraged a local-to-global semantic fusion module and a transition reasoning module to over" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.639 + ], + "angle": 0, + "content": "come the challenges of semantic association and transition fluency. Extensive experiments show that our method outperforms state-of-the-art techniques in terms of semantic matching, transition fluency, and realism." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.892, + 0.745 + ], + "angle": 0, + "content": "However, our method still has limitations, such as difficulty in synthesizing head motions like \"nodding\" and \"shaking head\". In future work, we plan to improve this by exploring a way to assign different attention weights to different joints automatically. Additionally, since our STDM dataset is synthesized while not captured, we will further improve the data quality and expand the dataset scale." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.669, + 0.772 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.777, + 0.892, + 0.898 + ], + "angle": 0, + "content": "This paper is supported by National Natural Science Foundation of China (62272021, 62102036), Beijing Natural Science Foundation (4222024), R&D Program of Beijing Municipal Education Commission (KM202211232003), Open Project Program of State Key Laboratory of Virtual Reality Technology and Systems, Beihang University (No.VRLAB2022A02), Research Unit of Virtual Human and Virtual Surgery (2019RU004)." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9506" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.143 + ], + "angle": 0, + "content": "[1] Mixamo. https://www MIXamo.com. Accessed February 17, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.468, + 0.172 + ], + "angle": 0, + "content": "[2] CMU graphics lab motion capture database. http://mocap.cs.cmu.edu/, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.175, + 0.468, + 0.23 + ], + "angle": 0, + "content": "[3] Chaitanya Ahuja, Dong Won Lee, Yukiko I.Nakano, et al. Style transfer for co-speech gesture animation: A multi-speaker conditional-mixture approach. European Conference on Computer Vision, pages 248-265, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.232, + 0.468, + 0.286 + ], + "angle": 0, + "content": "[4] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. 2019 International Conference on 3D Vision (3DV), pages 719-728, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.289, + 0.468, + 0.343 + ], + "angle": 0, + "content": "[5] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, et al. Style-controllable speech-driven gesture synthesis using normalising flows. Computer Graphics Forum, 39(2):487-496, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.346, + 0.468, + 0.4 + ], + "angle": 0, + "content": "[6] Nikos Athanasiou, Mathis Petrovich, Michael J.Black, et al. TEACH: Temporal action composition for 3d humans. 2022 International Conference on 3D Vision (3DV), pages 414-423, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.403, + 0.468, + 0.458 + ], + "angle": 0, + "content": "[7] Pablo Cervantes, Yusuke Sekikawa, Ikuro Sato, et al. Implicit neural representations for variable length human motion generation. European Conference on Computer Vision, pages 356-372, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.46, + 0.468, + 0.515 + ], + "angle": 0, + "content": "[8] Jihoon Chung, Cheng hsin Wuu, Hsuan ru Yang, et al. HAA500: Human-centric atomic action dataset with curated videos. IEEE/CVF International Conference on Computer Vision, pages 13465-13474, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.468, + 0.572 + ], + "angle": 0, + "content": "[9] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, et al. Synthesis of compositional animations from textual descriptions. IEEE/CVF International Conference on Computer Vision, pages 1396-1406, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.468, + 0.629 + ], + "angle": 0, + "content": "[10] Shiry Ginosar, Amir Bar, Gefen Kohavi, et al. Learning individual styles of conversational gesture. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.468, + 0.686 + ], + "angle": 0, + "content": "[11] Chuan Guo, Shihao Zou, Xinxin Zuo, et al. Generating diverse and natural 3d human motions from text. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.468, + 0.742 + ], + "angle": 0, + "content": "[12] Chuan Guo, Xinxin Zuo, Sen Wang, et al. Action2motion: Conditioned generation of 3d human motions. 28th ACM International Conference on Multimedia, pages 2021-2029, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.468, + 0.787 + ], + "angle": 0, + "content": "[13] Chuan Guo, Xinxin Zuo, Sen Wang, et al. Action2video: Generating videos of human 3d actions. International Journal of Computer Vision, 130(2):285-315, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[14] Chuan Guo, Xinxin Zuo, Sen Wang, et al. TM2T: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. European Conference on Computer Vision, pages 580-597, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[15] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, et al. Learning speech-driven 3d conversational gestures from video. 21st ACM International Conference on Intelligent Virtual Agents, pages 101-108, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[16] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2:1735-1742, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[17] Yanli Ji, Feixiang Xu, Yang Yang, et al. A large-scale RGB-D database for arbitrary-view human action recognition. 26th ACM International Conference on Multimedia, pages 1510-1518, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[18] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. arXiv preprint arXiv:2209.00349, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[19] Taras Kucherenko, Patrik Jonell, Sanne van Waveren, et al. Gesticulator: A framework for semantically-aware speech-driven gesture generation. 2020 International Conference on Multimodal Interaction, pages 242-250, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[20] Yuanzhi Liang, Qianyu Feng, Linchao Zhu, et al. SEEG: Semantic energized co-speech gesture generation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10473-10482, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[21] Xiao Lin and Mohamed R Amer. Human motion modeling using dvgens. arXiv preprint arXiv:1804.10652, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[22] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, et al. BEAT: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. arXiv preprint arXiv:2203.05297, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[23] Jun Liu, Amir Shahroudy, Mauricio Perez, et al. NTU rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2684-2701, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[24] Xian Liu, Qianyi Wu, Hang Zhou, et al. Learning hierarchical cross-modal association for co-speech gesture generation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[25] Matthew Loper, Naureen Mahmood, Naureen Mahmood, et al. SMPL: A skinned multi-person linear model. Acm Transactions on Graphics, 34(6cd):248, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[26] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. Weakly-supervised action transition learning for stochastic human motion prediction. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8151-8160, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[27] Dario Pavllo, David Grangier, and Michael Auli. Quaternet: A quaternion-based recurrent model for human motion. arXiv preprint arXiv:1805.06485, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[28] Mathis Petrovich, Michael J.Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[29] Mathis Petrovich, Michael J.Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. arXiv preprint arXiv:2204.14109, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[30] Matthias Plappert, Christian Mandery, and Tamim Asfour. The KIT motion-language dataset. Big Data, 4(4):236-252, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, et al. Learning transferable visual models from natural language supervision. International Conference on Machine Learning, pages 8748-8763, 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9507" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[32] Abhinanda R.Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, et al. BABEL: Bodies, action and behavior with english labels. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 722-731, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[33] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, et al. Nturogb+ d: A large scale dataset for 3d human activity analysis. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1010-1019, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[34] Guy Tevet, Brian Gordon, Amir Hertz, et al. MotionCLIP: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.471, + 0.304 + ], + "angle": 0, + "content": "[35] Sijie Yan, Zhizhong Li, Yuanjun Xiong, et al. Convolutional sequence generation for skeleton-based action synthesis. IEEE/CVF International Conference on Computer Vision, pages 4394-4402, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.471, + 0.36 + ], + "angle": 0, + "content": "[36] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, et al. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics, 39(6):1-16, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.362, + 0.471, + 0.403 + ], + "angle": 0, + "content": "[37] Mingyuan Zhang, Zhongang Cai, Liang Pan, et al. Motion-diffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.405, + 0.471, + 0.446 + ], + "angle": 0, + "content": "[38] Yan Zhang, Michael J.Black, and Siyu Tang. Perpetual motion: Generating unbounded human motion. arXiv preprint arXiv:2007.13886, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.448, + 0.471, + 0.503 + ], + "angle": 0, + "content": "[39] Hang Zhao, Antonio Torralba, Lorenzo Torresani, et al. Hacs: Human action clips and segments dataset for recognition and temporal localization. IEEE/CVF International Conference on Computer Vision, pages 8668-8678, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.505, + 0.471, + 0.56 + ], + "angle": 0, + "content": "[40] Yi Zhou, Connelly Barnes, Jingwan Lu, et al. On the continuity of rotation representations in neural networks. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.956 + ], + "angle": 0, + "content": "9508" + } + ] +] \ No newline at end of file diff --git a/2023/Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions/2c5b9c34-9105-42fb-b593-c16dc3db7c25_origin.pdf b/2023/Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions/2c5b9c34-9105-42fb-b593-c16dc3db7c25_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6f12147401410ad0800ae2ad23ffca8f4180f9f3 --- /dev/null +++ b/2023/Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions/2c5b9c34-9105-42fb-b593-c16dc3db7c25_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9181534072cdac1bb129cb3541486784700a8c87ce880e9660ede19c98e8f30c +size 5439791 diff --git a/2023/Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions/full.md b/2023/Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..41fdef9518cd3b90b2c884c75dd8e4f565fe0d0d --- /dev/null +++ b/2023/Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions/full.md @@ -0,0 +1,278 @@ +# Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions + +Shuai Li $^{1,3}$ , Sisi Zhuang $^{1}$ , Wenfeng Song $^{2*}$ , Xinyu Zhang $^{2}$ , Hejia Chen $^{1}$ , Aimin Hao $^{1}$ $^{1}$ State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, P.R. China + $^{2}$ Computer School, Beijing Information Science and Technology University, P.R. China + $^{3}$ Zhongguancun Laboratory, Beijing, P.R. China + +{lishuai, sisizhuang}@buaa.edu.cn, songwenfenga@gmail.com zhangxinyul@bistu.edu.cn,{chenhj2000, ham}@buaa.edu.cn + +![](images/0c8a23ec46a6e56d3b2f31ba9f9d108bfbf5fb8600de7b6bfc14f9653bdfa87c.jpg) +Figure 1: Our method synthesizes semantic human motion with natural transition using free-form sequential texts. The skeleton motion can be easily transferred to any character. The character used for demonstration is from Mixamo [1]. + +# Abstract + +The intelligent synthesis/generation of daily-life motion sequences is fundamental and urgently needed for many VR/metaverse-related applications. However, existing approaches commonly focus on monotonic motion generation (e.g., walking, jumping, etc.) based on single instruction-like text, which is still not intelligent enough and can't meet practical demands. To this end, we propose a cohesive human motion sequence synthesis framework based on free-form sequential texts while ensuring semantic connection and natural transitions between adjacent motions. At the technical level, we explore the local-to-global semantic features of previous and current texts to extract relevant information. This information is used to guide the framework in understanding the semantics of the current moment. Moreover, we propose learnable tokens to adaptively learn the influence range of the previous motions towards natural transitions. These tokens can be trained to encode the relevant information into well-designed transition loss. To demonstrate the efficacy of our method, we conduct extensive experiments and comprehensive evaluations on the public dataset as well as a new dataset produced by us. All the experiments confirm that our method outperforms the state-of-the-art methods in terms of semantic matching, realism, and transition fluency. Our project is public available. https://druthrie.github.io/sequential-texts-to-motion/ + +# 1. Introduction and Motivation + +Human motion synthesis is fundamental for numerous applications, especially for virtual reality, games, and metaverse-related applications [10, 19, 5, 3, 36, 15, 22, 20, 24], of which, it is in high demand to accurately control the digital human motion with natural language. Existing approaches commonly focus on monotonic motion synthesis/generation based on single instruction-like text description. However, practical applications usually require digital humans to respond to multiple rounds of sustainable interactions, wherein they can continuously generate reasonable responses to the sequential texts. Therefore, given a set of free-form sequential texts, we aim to synthesize a cohesive human motion sequence. Namely, the animation clips should be consistent with the ongoing text descriptions, and the whole motion sequence should have smooth semantic connection and natural transitions. + +At present, sequential texts-driven motion synthesis has not been well studied, mainly due to lacking long-term continuous motion datasets with accompanying free-form text descriptions. Recently, TEACH [6] first attempts to address this problem by proposing a dataset (we refer to it as BABEL-TEACH). Each item in the BABEL-TEACH dataset contains two adjacent text-motion pairs. However, TEACH [6] builds on the one-time method TEMOS [29], wherein the limited 5-frame motion is added to encode together. However, it does not consider the potential benefits of previous text information. In practice, previous texts can provide extra and more accurate semantic information for + +current motion synthesis. Therefore, global semantic needs to be extracted from the previous text and the local semantics from the current text, and then fuses the semantics to guide the semantic understanding at each moment. + +Besides, there are still two main challenges. Firstly, the semantic relationship between adjacent texts is largely overlooked in previous research. For example, when synthesizing a motion for the current text "touching the face with left hand," the previous text context should be considered. Specifically, if the previous text was "sitting on a chair," the synthesized motion should depict a person sitting and touching face with left hand (shown in Fig. 2); if the previous text was "a person kicks a ball with right foot," the motion should show a standing person touching face with left hand. This requires a more comprehensive understanding about the context and semantic relationships between the texts. Secondly, existing methods tend to abrupt transitions between adjacent motions when multiple motions are synthesized separately and stitched together. We aim to address this issue by seamlessly blending the synthesized motions. It should intelligently capture the temporal and spatial continuity between the motions to ensure a more natural and coherent sequence of motions. + +Considering the easy scalability of autoregressive methods, we should extend previous autoregressive single text-driven motion synthesis methods to multiple motions involved in motion sequence synthesis. We observe that a similar human body posture can be achieved by transferring the end features of the previous motion to the next motion, although this is not always entirely consistent. This suggests that previous motion information should be leveraged more effectively and sufficiently rather than being adopted straightforwardly. Therefore, we plan to introduce a transition reasoning module that adaptively learns attention score from the previous motion information and infers accurate motion features at the transition, making the transition between adjacent motions more natural and smooth. + +To this end, we create a new dataset with much longer-term motions (2-5 times longer) based on a synthetic method, and the corresponding text descriptions are more abundant. We conduct experiments on both BABEL-TEACH [6] dataset and our sequential texts described motion (STDM) dataset. The results demonstrate that our method outperforms existing methods in terms of semantic matching, realism, and transition fluency. Specially, the salient contributions can be summarized as follows. + +- We propose a cohesive motions synthesis framework using sequential texts as inputs. The framework can integrate previous text information to obtain semantic features and adaptively select valid previous motion information to guide the current motion synthesis. +- We design a local-to-global (L2G) semantic fusion module to extract accurate contextual information. It + +![](images/657fd4c19d8f689b3c9cc4a36f4b5e6c8b751da6206a706583b27b768b20a515.jpg) +Figure 2: Top row: Motion synthesis corresponding to sequential texts by our method. Bottom row: Motion synthesis corresponding to single text by T2M [11]. + +can well take into account global-context information from the previous text to infer the current moment's text meaning to guide the semantics-consistent motion synthesis. + +- We introduce a transition reasoning module to adaptively select motion snippet from previous motion information to make the synthesized motion natural and smooth w.r.t the previous, wherein a well-defined transition loss is employed to further constrain the fluency of the transitions. +- We create a new sequential texts described motion (STDM) dataset, which involves more extended motion frames and more diverse text descriptions than the existing dataset. + +# 2. Related Work + +We briefly introduce the related works of human motion synthesis considering semantic control, including single text driven motion synthesis and sequential texts driven motion synthesis. + +Single text driven motion synthesis. The motion synthesis from text began with the synthesized motion from action labels, but this approach was limited in covering most human motions [21, 27, 35, 38, 28, 12, 28, 7]. Later, researchers [4, 9, 34] focused on synthesizing motion from complex free-form text descriptions, with early work generating single, deterministic motions from text. Recent works have aimed to improve the diversity and details of the synthesized motion. For example, TEMOS [29] and T2M [11] utilized VAE architecture to synthesize diverse motions from the same texts. The main difference between them is that the former utilized Transformer architecture to synthesize all moment motions at once, while the latter utilized GRU architecture to synthesize every moment motion auto-regressively. To capture more detailed information in the text, TM2T [14] utilized VQVAE-based motion markers to provide a fair environment when considering motion and text signals, and utilized motion-to-text analysis module to strengthen the constraints on text-to-motion synthesis. With the successful application of the diffusion model in image generation, MotionDiffuse [37] utilized the diffusion model + +![](images/564f1bf0b59f23c839f8645c4126c2ac2d024c60f5b9481f4fb669600c52f322.jpg) +Figure 3: Method Overview: Our model takes sequential texts as input. In the local-to-global (L2G) semantic fusion module, the previous text's global features and the current text's local features are fused and extracted. In the transition reasoning module, the motion snippet code at the transition of adjacent motion ( $\mathbf{c}_{start}^{0}$ and $\mathbf{c}_{start}^{1}$ ) is deduced from the previous motion. Input the gate snippet code $\mathbf{g}^t$ and the semantic text feature $F_{sem}^{t}$ to synthesize the motion snippet code $\hat{\mathbf{c}}_i^t$ . The transition loss further limits the fluency of the transition with the previous motion. + +to simulate text driven conditional human motion synthesis, which can better respond to fine-grained instructions of body parts. FLAME [18] utilized transformer-based diffusion model architecture to synthesize human motion, and it can allow editing frames and joints without fine-tuning. Its editing ability can be extended to motion prediction or in-between tasks. Although the diffusion model improves the quality of the synthesized motion, its inference speed is relatively slow, so it is unsuitable for real-time applications. + +Sequential texts driven motion synthesis. Synthesizing motion sequences from sequential texts is challenging due to the limited availability of long-term continuous motions datasets with text descriptions. Previous works Action2video [13] realized the transition between actions by changing the input action label type in an autoregressive synthesizer. Yet, they only conducted experiments on a limited number of action labels in three datasets: NTURGBD [33], CMU [2], and humanACT12 [12]. Mao et al. [26] proposed a weakly-supervised action-driven motion prediction method, but they only used 20 action categories with clear transitions in BABEL [32] to synthesize the following motion and transition. Based on the existing BABEL dataset [32], TEACH [6] proposed a new dataset, in which each data consisted of two adjacent text descriptions and motion sequences. The model of TEACH [6] encoded five previous motion frames using TEMOS [29], every single motion sequence was synthesized simultaneously, and adjacent motion sequences were auto-regressively synthesized. Our work builds on this by proposing a cohesive human motion sequence synthesis framework based on freeform sequential texts, ensuring semantic connection and natural transitions between adjacent motions. + +# 3. Our Approach + +# 3.1. Method Overview + +The overall pipeline of the proposed model is shown in Fig. 3. The temporal VAE and motion autoencoder used in the model are similar to those used in the T2M model, wherein it is employed to encode the motion sequence into a motion snippet code sequence and reconstruct the motion sequence with a decoder. To take sequential texts as input and ensure semantically coherent synthesized motions, a local-to-global (L2G) semantic fusion module is introduced (Sec. 3.2). This module guides learning accurate semantic features from the input texts at each moment, including historical text information. To ensure smooth and natural transitions between adjacent motions, a transition reasoning module is introduced (Sec. 3.3). This module dynamically deduces the motion snippet code during transition based on previous motion information, which can help the model generate realistic and coherent transitions. Additionally, a transition loss is proposed to enhance the smoothness of the transition motion (Sec. 3.4). Overall, the proposed pipeline combines several modules and techniques to generate realistic and semantically coherent motion sequences from sequential texts input. The model can generate smooth transitions between motions, and can adaptively learn the influence range of previous motions to further improve the quality of the synthesized motion sequences. + +Preliminaries. Our method aims to take freeform sequential texts $\mathbf{S} = (\mathbf{S}_1, \mathbf{S}_2, \dots, \mathbf{S}_n)$ as input, e.g., $\mathbf{S} =$ ("salute with the left hand", "salute with the right hand", "sit on the chair", "crawl"), and outputs corresponding human motion sequences $\mathbf{M} =$ + +$(\mathbf{M}_1, \mathbf{M}_2, \dots, \mathbf{M}_n)$ . The transition between adjacent motions should be natural and smooth, and the synthesized motion should reflect the semantic information of previous texts. The data format is consistent with HumanML3D [11] dataset, including the root angular velocity along the Y-axis, the root linear velocity on the XZ-plane, the root height, the local joints positions, velocities, and rotations [40] in root space, while the motions follow the skeleton structure of SMPL [25] with 22 joints. Each moment in our model represents 4 frames. + +# 3.2. Local-to-global Semantic Fusion Module + +To synthesize motion that accurately matches the semantics of sequential texts, our model relies on accurately extracting text features at each moment to guide the synthesis process. As the motion we synthesize is based on sequential texts, the current motion is determined not only by the current text semantics but also by previous text semantics. To capture the temporal clues, we encode both the previous and current text, and use an attention mechanism to extract text features that integrate the previous text semantics at each moment (shown in Fig. 4). For previous texts, we use a pretrained CLIP [31] text encoder to extract global features of previous texts. In order to make the features more suitable for our task in the process of training, linear projector is used to extract features further. The global features from the previous texts are expressed as: + +$$ +\mathbf {G} _ {p r e} = f \left(C L I P _ {t e x t} \left(\mathbf {S} _ {i - 1}\right)\right). \tag {1} +$$ + +It is difficult for the CLIP text encoder to extract the local features of the text, motion synthesis of each moment requires local text semantic guidance, so we employ a Bi-GRU-based text encoder, similar to T2M [11], to obtain the local features $\mathbf{L}_{1:m}$ from the current text. To capture accurate text semantic features at the arbitrary frames, we use an attention mechanism, wherein the key $\mathbf{K}(L)$ and value $\mathbf{V}(L)$ of attention are the local features of the current text. Since the semantic information of the previous text is required for guidance at the beginning of the currently synthesizing motion, we use the global features of the previous text as the initial query $\mathbf{Q}(G)$ . At subsequent moments, the motion snippet code generator in temporal VAE provides the queries $\mathbf{Q}(L)$ . This can be expressed mathematically as: + +$$ +\mathbf {Q} (G) = \mathbf {G} _ {p r e} \mathbf {W} ^ {Q}, \mathbf {Q} (L) = \mathbf {h} ^ {t - 1} \mathbf {W} ^ {Q}, +$$ + +$$ +\mathbf {K} (L) = \mathbf {L} _ {1: m} \mathbf {W} ^ {K}, \mathbf {V} (L) = \mathbf {L} _ {1: m} \mathbf {W} ^ {V}, \tag {2} +$$ + +$$ +\mathbf {F} _ {s e m} ^ {t} = f \left(\frac {\mathbf {Q} (G / L) \mathbf {K} (L) ^ {T}}{\sqrt {d _ {s e m}}}\right) \mathbf {V} (L), +$$ + +where $\mathbf{W}^k, \mathbf{W}^V \in \mathbb{R}^{d_w \times d_{sem}}$ and $\mathbf{W}^Q \in \mathbb{R}^{d_h \times d_{sem}}$ are trainable weights; $d_h, d_w$ and $d_{sem}$ are the number of channels in hidden unit $\mathbf{h}^{t-1}$ , current text feature $\mathbf{L}^{1:m}$ and semantic fusion attention layer, respectively. $\mathbf{h}^t$ is the hidden + +![](images/e664c64417496ee69981a4c62ac4e251ef467f247ab8e2a2e6c9469cab138718.jpg) +Figure 4: Local-to-global semantic fusion module: After extracting the global feature $\mathbf{G}_{pre}$ from the previous texts and the local features $\mathbf{L}_{1:m}$ from the current text, feature fusion is performed to obtain the accurate semantic feature $\mathbf{F}_{sem}^{t}$ of the current text at moment $t$ . + +unit generated by the motion snippet code generator at each moment and $\mathbf{h}^0$ is obtained from $\mathbf{G}_{pre}$ . $\mathbf{F}_{sem}^{t}$ is the text semantic feature at moment $t$ . + +# 3.3. Transition Reasoning Module + +When synthesizing a new motion, the starting pose is influenced by the ending pose of the previous motion. Grafting the end motion snippet code of the previous motion to the start moment of the current motion can produce similar human poses during the transition. However, the transition between the two motions is still unnatural and abrupt. To address this, we introduce a learnable start motion snippet code token $\mathbf{c}_{\text {token }}$ and utilize it along with the previous motion snippet code sequence in a transformer encoder. The output of the token position is used as the start motion snippet code $\mathbf{c}_{\text {start }}$ for the current motion, allowing the model to learn the range and weight of the previous motions (shown in Fig. 5). Additionally, the length of the current motion affected by the previous motion information needs to be determined. Through experimentation, we obverse that the best performance is achieved by reasoning the first two motion snippet codes $(\mathbf{c}_{\text {start }}^{0}$ and $\mathbf{c}_{\text {start }}^{1})$ of the current motion from the previous motions. The verification experiment is in Sec. 4.5. The mathematical expression is as: + +$$ +\mathbf {c} _ {\text {s t a r t}} ^ {0}, \mathbf {c} _ {\text {s t a r t}} ^ {1} = \text {T r a n s f o r m e r} \left(\hat {\mathbf {c}} _ {i - 1}, \mathbf {c} _ {\text {t o k e n}} ^ {0}, \mathbf {c} _ {\text {t o k e n}} ^ {1}\right). \tag {3} +$$ + +As shown by the transition gate in Fig. 3, when synthesizing the motion snippet codes of the first two moments, we choose the motion snippet codes $(\mathbf{c}_{start}^{0}$ and $\mathbf{c}_{start}^{1})$ inferred from the previous motion as gate snippet code $\mathbf{g}^t$ . From the third moment, the motion snippet code $\hat{\mathbf{c}}^{t - 1}$ from the previ + +![](images/229c73e4a307153e29732d9aa1410ef2ce26d964a6e8e67919883968a8f44dd2.jpg) +Figure 5: Transition reasoning module: Given the previous texts to get the previous synthesized motion sequence $\hat{\mathbf{M}}_{i-1}$ and motion snippet codes $\hat{\mathbf{c}}_{i-1}$ , we use two learnable tokens ( $\mathbf{c}_{token}^{0}$ and $\mathbf{c}_{token}^{1}$ ) to learn the previous motion information that needs attention, so as to obtain the motion snippet codes at the starting moment of the current motion ( $\mathbf{c}_{start}^{0}$ and $\mathbf{c}_{start}^{1}$ ). + +ous moment is used, which is synthesized autoregressively. For gate snippet code $\mathbf{g}^t$ , the mathematical expression is as: + +$$ +\mathbf {g} ^ {t} = \left\{ \begin{array}{l l} \mathbf {c} _ {\text {s t a r t}} ^ {0} & t = 1 \\ \mathbf {c} _ {\text {s t a r t}} ^ {1} & t = 2 \\ \hat {\mathbf {c}} ^ {t - 1} & t \geq 3 \end{array} \right. \tag {4} +$$ + +When synthesizing the motion based on the first text $\mathbf{S}_1$ of the sequential texts, we set the previous texts as the default text $\mathbf{S}_0 = \text{"start"}$ , and the starting snippet code $\mathbf{c}_{start}^0$ is obtained from the mean pose. + +# 3.4. Transition Loss & Total Loss + +We conduct a statistical analysis of the dataset and find that the motion gap (L1 distance) of almost all adjacent frames is less than 0.1. Inspired by this, we design a transition loss $L_{Trans}$ to reduce the motion gap between the adjacent frames and avoid abrupt transition. The representation of the $L_{Trans}$ is + +$$ +\mathcal {L} _ {\text {T r a n s}} = \left\| \hat {\mathbf {M}} _ {i} ^ {1} - \hat {\mathbf {M}} _ {i - 1} ^ {T _ {i} - 1} \right\| _ {1}. \tag {5} +$$ + +The final loss function of our model consists of 4 parts: motion snippet code reconstruction loss $\mathcal{L}_{rec}^{code}$ , motion reconstruction loss $\mathcal{L}_{rec}^{mot}$ , KL loss $\mathcal{L}_{KL}$ of prior distribution and posterior distribution in temporal VAE, and the transition loss $\mathcal{L}_{Trans}$ . The $\lambda_{code},\lambda_{mot},\lambda_{KL},\lambda_{Trans}$ are the + +
DatasetQuantityDurationVocab.
STDM528920.02 h2455
BABEL-TEACH[6]1626621.86 h1037
+ +Table 1: Comparison of datasets: STDM dataset has longer-term frames within a single motion, and the text is more abundant. + +corresponding weights of the losses. + +$$ +\mathcal {L} = \lambda_ {\text {c o d e}} \mathcal {L} _ {\text {r e c}} ^ {\text {c o d e}} + \lambda_ {\text {m o t}} \mathcal {L} _ {\text {r e c}} ^ {\text {m o t}} + \lambda_ {K L} \mathcal {L} _ {K L} + \lambda_ {\text {T r a n s}} \mathcal {L} _ {\text {T r a n s}}. \tag {6} +$$ + +# 4. Experiments + +First, we introduce our STDM dataset (Sec. 4.1), evaluation metrics (Sec. 4.2), and then conduct quantitative (Sec. 4.3) and qualitative (Sec. 4.6) evaluations. We carry out an ablation study (Sec. 4.4) to prove the effectiveness of our proposed modules and analyze the parameters (Sec. 4.5). + +# 4.1. STDM Dataset + +Each text-motion pair in most existing human motion datasets [30, 17, 23, 39, 12, 8] is independent of the other, and does not contain adjacent text-motion pairs. In a recent study, TEACH [6] used the BABEL [32] dataset to obtain a BABEL-TEACH dataset containing two adjacent text-motion pairs, which is consistent with our research goal. + +To obtain adequate motion sequences, we synthesize sequential texts and motion sequences by modifying the existing single text driven motion synthesis method from the viewpoint of data generation instead of using a mocap device. By randomly selecting and annotating texts from a large pool, we create sequential text fed into our improved T2M [11], generating motion sequences. Based on T2M [11], we find that the motion sequences decoded by the same motion snippet code are similar. Our approach diverges from T2M: In the inference stage, instead of using a fixed mean pose, we use snippet codes from the last eight frames of the previous motion, ensuring smoother and more natural transitions. Our experiment results show that the motion sequences synthesized in this way are similar in transition. Since the motion generated by the generation model may not match the semantics of the input text and the transition may not be smooth, we manually remove data with poor transitions and correct erroneous texts to maintain the quality of our STDM dataset. After filtering, we get 5289 pairs containing two adjacent text-motion. Our dataset is collected under the same structure as HumanML3D [11]. + +Compared with the BABEL-TEACH [6] dataset, STDM has longer-term motion frames, with a single motion sequence ranging from 2 to 10 seconds, while most BABEL-TEACH [6] data is less than 3 seconds. However, the text + +![](images/768a408617a2d8ad37757a91c9ced008343033504d353dd63dcf4d587d894e98.jpg) +Figure 6: STDM Dataset: Two examples in the STDM dataset, which involves more extended motion frames and more diverse text descriptions than the BABEL-TEACH [6] dataset. Blue traces represent the digital human's trajectory. + +descriptions in BABEL-TEACH [6] usually describe only one action, such as "walk forward". The text descriptions in the STDM are more abundant, such as "a person raises his right hand and walks forward". A detailed comparison of the two datasets can be found in Tab. 1. Fig. 6 shows two examples of our STDM dataset. + +# 4.2. Evaluation Metrics + +In order to evaluate the quality of the synthesized motion sequence, we use the feature extractor proposed by Guo et al. [11] to extract the features of text and motion. This feature extractor is trained with contrast loss [16], which can make the matched text-motion pairs produce similar feature vectors. Therefore, the feature extractor can measure whether the text and motion sequence match. We combine two adjacent text-motion pairs into one to train the feature extractor. We use the same evaluation metrics as T2M [11]. The metrics are described in the supplementary materials. + +To demonstrate the fluency of different methods when transiting between adjacent motions, we add a metric to evaluate the fluency of transition, by calculating the distance between adjacent frames of two adjacent motion sequences. For transition fluency, the mathematical expression is as: + +$$ +F = \sum_ {i = 1} ^ {N} \| \mathbf {M} _ {i 1} ^ {T 1} - \mathbf {M} _ {i 2} ^ {1} \| _ {1}. \tag {7} +$$ + +Here, we calculate the average transition distance ( $L_{1}$ distance) of the motion sequence synthesized by $N$ adjacent text pairs. For the $i$ -th adjacent text pair, the synthesized adjacent motion sequence pair is $\mathbf{M}_{i1} = (\mathbf{M}_{i1}^{1}, \mathbf{M}_{i1}^{2}, \dots, \mathbf{M}_{i1}^{T1})$ and $\mathbf{M}_{i2} = (\mathbf{M}_{i2}^{1}, \mathbf{M}_{i2}^{2}, \dots, \mathbf{M}_{i2}^{T2})$ . + +# 4.3.Quantitative Evaluation + +Baselines. We compare our method with T2M [11], our two variants T2M-Joint and Complusion-Code, and TEACH [6]. For T2M, we input a single text, output a single motion, and align the root coordinates of the last frame of the previous motion with the root coordinates of the first frame of the next motion. For T2M-Joint, we keep the T2M model structure and train the model by combining two adjacent texts with commas as input. For Complusion-Code, based on the T2M [11], the last motion snippet code of the previous motion is used as the start motion snippet code of the next motion during training (Unlike the STDM dataset generation method, the dataset generation only uses the snippet codes of previous motions in the inference phase. Complusion-Code also uses the snippet codes of previous motions in the training phase). For TEACH [6], we compared two versions using spherical linear interpolation (Slerp) and not using Slerp, respectively. Slerp operation is unrelated to the TEACH [6] model and is simply a way of post-processing synthesized data. + +Tab. 2 and Tab. 3 show our quantitative comparison results with other baselines on BABEL-TEACH [6] and STDM datasets. We divide the training, testing, and validation sets according to the ratio of $0.8:0.15:0.05$ . For a fair comparison, we conduct 20 experiments and show the values with confidence in the $95\%$ range. From the results, our method outperforms all other methods, especially in the transition fluency metric, which is important for synthesizing natural and smooth transition motion. T2M [11] can only synthesize a single motion per inference, and there is no connection between adjacent motions, so it performs poorly in all indicators. Because our testing data are composed of two adjacent texts, the T2M-Joint method performs similarly to the real data regarding transition fluency + +
MethodsR Precision↑FID↓MultiModal Dist↓Transition Fluency↓
Top 1Top 2Top 3
Real motions0.560±0.0020.750±0.0030.834±0.0020.001±0.0003.448±0.001-
T2M [11]0.376±0.0030.543±0.0020.647±0.0036.890±0.0815.218±0.0101.108±0.004
T2M-Joint0.373±0.0040.534±0.0030.640±0.0032.727±0.0324.983±0.011-
Compulsion-Code0.432±0.0030.613±0.0040.717±0.0033.726±0.0684.612±0.0230.348±0.002
TEACH (no Slerp) [6]0.573±0.0020.756±0.0030.842±0.0022.263±0.0553.487±0.0100.577±0.003
TEACH (Slerp) [6]0.563±0.0030.750±0.0030.839±0.0032.240±0.0533.526±0.0110.118±0.001
Ours0.542±0.0030.728±0.0030.818±0.0031.628±0.0313.662±0.0100.177±0.001
Ours (Slerp)0.519±0.0040.705±0.0040.803±0.0031.548±0.0243.797±0.0120.036±0.000
+ +Table 2: Quantitative evaluation on testing data of BABEL-TEACH: ± indicates 95% confidence interval, ↑ and ↓ respectively denotes better performance with larger or lower value. Bold indicates the optimal result, the underscore represents the suboptimal result, while wave line refers to the third best. The results show that the motion synthesized by our model outperforms other baselines in terms of semantic matching, transition fluency, and realism. + +
MethodsR Precision↑FID↓MultiModal Dist↓Transition Fluency↓
Top 1Top 2Top 3
Real motions0.322±0.0040.505±0.0050.625±0.0060.014±0.0023.461±0.0050.454±0.000
T2M [11]0.322±0.0060.493±0.0050.614±0.0051.389±0.0493.500±0.0090.639±0.003
T2M-Joint0.304±0.0050.488±0.0060.613±0.0050.744±0.0423.492±0.011-
Compulsion-Code0.285±0.0070.450±0.0060.572±0.0063.146±0.0803.806±0.0170.330±0.003
TEACH (no Slerp) [6]0.318±0.0060.505±0.0050.634±0.0051.414±0.0553.483±0.0120.246±0.002
TEACH (Slerp) [6]0.326±0.0040.504±0.0050.631±0.0061.416±0.0543.487±0.0130.049±0.000
Ours0.328±0.0060.510±0.0060.633±0.0051.085±0.0633.441±0.0170.109±0.001
Ours (Slerp)0.325±0.0070.515±0.0070.637±0.0051.134±0.0593.440±0.0130.022±0.000
+ +Table 3: Quantitative evaluation on the testing data of STDM: The results show that the motion synthesized by our model outperforms other baselines in terms of semantic matching, transition fluency, and realism. + +(less than 0.1). However, when the input text consists of more than two sentences, the T2M-Joint method cannot guarantee transition fluency. Because Compulsion-Code uses the end motion snippet code of the previous motion as the start snippet code of the next motion, the performance of transition fluency is better than T2M [11], but it is still not as good as ours. Although TEACH [6] is slightly better than ours in R Precision and MultiModal Distance on the BABEL-TEACH [6] dataset, as shown in Tab. 2, its transition fluency is not as good as ours, we can obtain transitional and natural motion sequences without Slerp operation, so our method is better than TEACH [6] in a comprehensive evaluation. Although the synthetic STDM dataset is less natural than actual mocap data in transition fluency, we find that our method can compensate for the defects of the training data after training. In terms of transition fluency, the results synthesized by our method outperform the ground truth of the testing data. This further verifies the effectiveness of our method in synthesizing natural and smooth transition motions. + +User study. We randomly select 50 texts from the testing data and arrange 3 to 5 texts together to form a set of sequential texts. Some text combinations are not present in the dataset to test the motion generation effect of each model for text combinations not in the dataset. The syn + +thesized motions based on these sequential text groups are randomly shuffled and presented using different methods. Users are required to score synthetic motions from the following three aspects: (1) Matching degree: the degree of semantic matching between the motion and the text; (2) Transition fluency: the natural degree of transition between adjacent motions; (3) Realism: how realistic the motion is. The result is shown in Fig. 7. + +# 4.4. Ablation Study + +In order to illustrate the influence of the modules and training strategies we introduced on the overall model, we carried out the ablation experiment. The results are shown in Tab. 4. + +The first experiment is to remove the transition reasoning module. We set the skeleton to mean pose to initialize the motion snippet code at the start moment. However, the model does not know the human posture at the beginning, so it greatly reduces transition fluency (value increased by 0.417). The experiment results show that the performance of the model will decline without the transition reasoning module because it does not aware of the information of previous motion when the model synthesizes the motion sequence of the current text, even though the model can synthesize motion sequences that match the semantics of the + +
MethodsR Precision↑FID↓MultiModal Dist↓Transition Fluency↓
Top 1Top 2Top 3
Real motions0.561±0.0030.747±0.0030.836±0.0020.002±0.0003.449±0.002-
w/o Transition reasoning module0.535±0.0040.713±0.0030.805±0.0032.190±0.0383.719±0.0130.594±0.003
w/o L2G semantic fusion module0.440±0.0030.621±0.0030.720±0.0034.592±0.0714.593±0.0130.151±0.001
w/o Transition loss0.525±0.0040.706±0.0040.801±0.0021.766±0.0363.797±0.0110.293±0.001
w/o Segmented training strategy0.500±0.0020.684±0.0040.783±0.0042.164±0.0384.041±0.0130.328±0.002
Ours0.542±0.0030.726±0.0030.817±0.0021.588±0.0303.655±0.0130.177±0.001
+ +Table 4: Ablation study: The results show that the transition reasoning module has the most significant impact on transition fluency, the L2G semantic fusion module has the most significant impact on R Precision, FID, and MultiModel Distance, which represent semantic matching and realism, and the transition loss and segmentation training strategy have a certain influence on all indicators. + +
Numbers of snippet codeR Precision↑FID↓MultiModal Dist↓Transition Fluency↓
Top 1Top 2Top 3
Real motions0.561±0.0030.750±0.0020.834±0.0020.002±0.0003.449±0.001-
1 snippet code0.517±0.0030.693±0.0040.786±0.0042.027±0.0433.845±0.0140.173±0.001
2 snippet codes0.540±0.0030.725±0.0030.816±0.0021.602±0.0263.662±0.0110.177±0.001
3 snippet codes0.522±0.0040.702±0.0030.796±0.0022.050±0.0383.838±0.0150.175±0.001
5 snippet codes0.504±0.0030.687±0.0040.785±0.0031.781±0.0293.913±0.0130.213±0.001
+ +Table 5: Parameter analysis of transition reasoning module: We infer the initial motion snippet code of the current motion sequence from the previous motion and determine the optimal number of the inference code through experiments. The result shows that inferring the first two codes of current motion from the previous motion work best. + +![](images/2f65555da25ba3cdd78d80957654a4b2c233ac97ed3f40765f4575347f5c06d2.jpg) +Figure 7: User study: We invite 45 users to score the results of our comparative experiments, and the color bars in the figure indicate the percentage of the scores. The results show that our method outperforms other baselines in terms of semantic matching, transition fluency, and realism. + +# current text. + +The second experiment is to remove the L2G semantic fusion module. We only encode the features of the current text at each moment without considering the previous text. The experiment results show that the model's performances on R Precision, FID, and MultiModal Distance decrease significantly after removing the L2G semantic fusion module, indicating that previous text information is essential in synthesizing motion sequences that match context semantics. Although there is a slight improvement in the metric of + +transition fluency after the removal, the gap between them can not be observed from the naked eye. With the L2G semantic fusion module, our model shows better competence in evaluating the implicit relationship between current and previous text. + +The third experiment is to remove the transition loss. The experimental results show that after removing the transition loss, the model's performances in all indicators have decreased, indicating that adding transition loss will guide the model to learn more semantic matching and transitional natural motion. + +The fourth experiment does not use the segmented training strategy (explained in the supplementary file). The specific approach is that all motion sequences are extended to the maximum length (196 frames) with 0 in the training strategy. The experimental results show that when the segmented training strategy is not used, the performance of each indicator decreases, indicating that the segmented training strategy can make the model learn more thoroughly. + +# 4.5. Parameter Analysis + +In the transition reasoning module, we verify through experiments that we need to infer the number of motion snippet codes from the previous motions at the current initial stage. As shown in Tab. 5, the reasoning motion snippet code number of 1, 2, 3, and 5 are compared. The experiment results show that two motion snippet codes at the beginning of reasoning are optimal because the model can not fully capture the influence of previous motion information by reducing motion snippets inferred from previous motions. At + +![](images/6fd40e06859ddfd216f4e4524a461523b3b2e08f254bd51269083df56b687dbc.jpg) +Figure 8: Qualitative evaluation: Continuously inputting four texts to show the visual results of our method and other baseline methods, except for the two frames between adjacent motions, the rest only show the keyframes. The results show that the motion synthesized by our method is the best in terms of semantic matching, transition fluency, and realism. More results are in the supplementary file. + +the same time, the model will introduce too much previous motion information into the current motion if increasing the motion snippets inferred from previous motions. + +# 4.6. Qualitative Evaluation + +We make a qualitative comparison with the baselines, as shown in Fig. 8, our synthesized motion is the most natural in transition and nicely matches the semantics of the texts. Transition in T2M [11] is most evident in adjacent motion. T2M-Joint cannot fully express the content of the text description. When the input sequential texts contain more than two sentences, noticeable changes in the adjacent motion transition will occur. Complusion-Code produces a slight change in the transition. There is some semantic mismatch in TEACH [6] and abrupt transitions in the no Slerp version. + +# 5. Conclusions and Future Works + +We have advocated a novel framework for synthesizing semantic motion sequences driven by sequential texts while ensuring natural transitions between adjacent motions. Our approach leveraged a local-to-global semantic fusion module and a transition reasoning module to over + +come the challenges of semantic association and transition fluency. Extensive experiments show that our method outperforms state-of-the-art techniques in terms of semantic matching, transition fluency, and realism. + +However, our method still has limitations, such as difficulty in synthesizing head motions like "nodding" and "shaking head". In future work, we plan to improve this by exploring a way to assign different attention weights to different joints automatically. Additionally, since our STDM dataset is synthesized while not captured, we will further improve the data quality and expand the dataset scale. + +# Acknowledgements + +This paper is supported by National Natural Science Foundation of China (62272021, 62102036), Beijing Natural Science Foundation (4222024), R&D Program of Beijing Municipal Education Commission (KM202211232003), Open Project Program of State Key Laboratory of Virtual Reality Technology and Systems, Beihang University (No.VRLAB2022A02), Research Unit of Virtual Human and Virtual Surgery (2019RU004). + +# References + +[1] Mixamo. https://www MIXamo.com. Accessed February 17, 2023. +[2] CMU graphics lab motion capture database. http://mocap.cs.cmu.edu/, 2003. +[3] Chaitanya Ahuja, Dong Won Lee, Yukiko I.Nakano, et al. Style transfer for co-speech gesture animation: A multi-speaker conditional-mixture approach. European Conference on Computer Vision, pages 248-265, 2020. +[4] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. 2019 International Conference on 3D Vision (3DV), pages 719-728, 2019. +[5] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, et al. Style-controllable speech-driven gesture synthesis using normalising flows. Computer Graphics Forum, 39(2):487-496, 2020. +[6] Nikos Athanasiou, Mathis Petrovich, Michael J.Black, et al. TEACH: Temporal action composition for 3d humans. 2022 International Conference on 3D Vision (3DV), pages 414-423, 2022. +[7] Pablo Cervantes, Yusuke Sekikawa, Ikuro Sato, et al. Implicit neural representations for variable length human motion generation. European Conference on Computer Vision, pages 356-372, 2022. +[8] Jihoon Chung, Cheng hsin Wuu, Hsuan ru Yang, et al. HAA500: Human-centric atomic action dataset with curated videos. IEEE/CVF International Conference on Computer Vision, pages 13465-13474, 2021. +[9] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, et al. Synthesis of compositional animations from textual descriptions. IEEE/CVF International Conference on Computer Vision, pages 1396-1406, 2021. +[10] Shiry Ginosar, Amir Bar, Gefen Kohavi, et al. Learning individual styles of conversational gesture. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019. +[11] Chuan Guo, Shihao Zou, Xinxin Zuo, et al. Generating diverse and natural 3d human motions from text. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. +[12] Chuan Guo, Xinxin Zuo, Sen Wang, et al. Action2motion: Conditioned generation of 3d human motions. 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. +[13] Chuan Guo, Xinxin Zuo, Sen Wang, et al. Action2video: Generating videos of human 3d actions. International Journal of Computer Vision, 130(2):285-315, 2022. +[14] Chuan Guo, Xinxin Zuo, Sen Wang, et al. TM2T: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. European Conference on Computer Vision, pages 580-597, 2022. +[15] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, et al. Learning speech-driven 3d conversational gestures from video. 21st ACM International Conference on Intelligent Virtual Agents, pages 101-108, 2021. + +[16] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2:1735-1742, 2006. +[17] Yanli Ji, Feixiang Xu, Yang Yang, et al. A large-scale RGB-D database for arbitrary-view human action recognition. 26th ACM International Conference on Multimedia, pages 1510-1518, 2018. +[18] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. arXiv preprint arXiv:2209.00349, 2022. +[19] Taras Kucherenko, Patrik Jonell, Sanne van Waveren, et al. Gesticulator: A framework for semantically-aware speech-driven gesture generation. 2020 International Conference on Multimodal Interaction, pages 242-250, 2020. +[20] Yuanzhi Liang, Qianyu Feng, Linchao Zhu, et al. SEEG: Semantic energized co-speech gesture generation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10473-10482, 2022. +[21] Xiao Lin and Mohamed R Amer. Human motion modeling using dvgens. arXiv preprint arXiv:1804.10652, 2018. +[22] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, et al. BEAT: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. arXiv preprint arXiv:2203.05297, 2022. +[23] Jun Liu, Amir Shahroudy, Mauricio Perez, et al. NTU rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2684-2701, 2019. +[24] Xian Liu, Qianyi Wu, Hang Zhou, et al. Learning hierarchical cross-modal association for co-speech gesture generation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. +[25] Matthew Loper, Naureen Mahmood, Naureen Mahmood, et al. SMPL: A skinned multi-person linear model. Acm Transactions on Graphics, 34(6cd):248, 2015. +[26] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. Weakly-supervised action transition learning for stochastic human motion prediction. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8151-8160, 2022. +[27] Dario Pavllo, David Grangier, and Michael Auli. Quaternet: A quaternion-based recurrent model for human motion. arXiv preprint arXiv:1805.06485, 2018. +[28] Mathis Petrovich, Michael J.Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. +[29] Mathis Petrovich, Michael J.Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. arXiv preprint arXiv:2204.14109, 2022. +[30] Matthias Plappert, Christian Mandery, and Tamim Asfour. The KIT motion-language dataset. Big Data, 4(4):236-252, 2016. +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, et al. Learning transferable visual models from natural language supervision. International Conference on Machine Learning, pages 8748-8763, 2021. + +[32] Abhinanda R.Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, et al. BABEL: Bodies, action and behavior with english labels. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 722-731, 2021. +[33] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, et al. Nturogb+ d: A large scale dataset for 3d human activity analysis. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1010-1019, 2016. +[34] Guy Tevet, Brian Gordon, Amir Hertz, et al. MotionCLIP: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022. +[35] Sijie Yan, Zhizhong Li, Yuanjun Xiong, et al. Convolutional sequence generation for skeleton-based action synthesis. IEEE/CVF International Conference on Computer Vision, pages 4394-4402, 2019. +[36] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, et al. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics, 39(6):1-16, 2020. +[37] Mingyuan Zhang, Zhongang Cai, Liang Pan, et al. Motion-diffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. +[38] Yan Zhang, Michael J.Black, and Siyu Tang. Perpetual motion: Generating unbounded human motion. arXiv preprint arXiv:2007.13886, 2020. +[39] Hang Zhao, Antonio Torralba, Lorenzo Torresani, et al. Hacs: Human action clips and segments dataset for recognition and temporal localization. IEEE/CVF International Conference on Computer Vision, pages 8668-8678, 2019. +[40] Yi Zhou, Connelly Barnes, Jingwan Lu, et al. On the continuity of rotation representations in neural networks. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2019. \ No newline at end of file diff --git a/2023/Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions/images.zip b/2023/Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f148eacdfac21f3f2d35b1b7c0e070e5cb81f4d2 --- /dev/null +++ b/2023/Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f50063a532a8ff2b06b13ce161844073fa97fddf7ab94506657feb32002a9c0e +size 863031 diff --git a/2023/Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions/layout.json b/2023/Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..05551a5b65fc837f12a538176d87f7c76b95e729 --- /dev/null +++ b/2023/Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions/layout.json @@ -0,0 +1,7248 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "type": "text", + "content": "Sequential Texts Driven Cohesive Motions Synthesis with Natural Transitions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 60, + 142, + 534, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 142, + 534, + 200 + ], + "spans": [ + { + "bbox": [ + 60, + 142, + 534, + 200 + ], + "type": "text", + "content": "Shuai Li" + }, + { + "bbox": [ + 60, + 142, + 534, + 200 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 60, + 142, + 534, + 200 + ], + "type": "text", + "content": ", Sisi Zhuang" + }, + { + "bbox": [ + 60, + 142, + 534, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 60, + 142, + 534, + 200 + ], + "type": "text", + "content": ", Wenfeng Song" + }, + { + "bbox": [ + 60, + 142, + 534, + 200 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 60, + 142, + 534, + 200 + ], + "type": "text", + "content": ", Xinyu Zhang" + }, + { + "bbox": [ + 60, + 142, + 534, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 60, + 142, + 534, + 200 + ], + "type": "text", + "content": ", Hejia Chen" + }, + { + "bbox": [ + 60, + 142, + 534, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 60, + 142, + 534, + 200 + ], + "type": "text", + "content": ", Aimin Hao" + }, + { + "bbox": [ + 60, + 142, + 534, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 60, + 142, + 534, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 60, + 142, + 534, + 200 + ], + "type": "text", + "content": "State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, P.R. China \n" + }, + { + "bbox": [ + 60, + 142, + 534, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 60, + 142, + 534, + 200 + ], + "type": "text", + "content": "Computer School, Beijing Information Science and Technology University, P.R. China \n" + }, + { + "bbox": [ + 60, + 142, + 534, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 60, + 142, + 534, + 200 + ], + "type": "text", + "content": "Zhongguancun Laboratory, Beijing, P.R. China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 145, + 201, + 449, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 201, + 449, + 228 + ], + "spans": [ + { + "bbox": [ + 145, + 201, + 449, + 228 + ], + "type": "text", + "content": "{lishuai, sisizhuang}@buaa.edu.cn, songwenfenga@gmail.com zhangxinyul@bistu.edu.cn,{chenhj2000, ham}@buaa.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 236, + 545, + 317 + ], + "blocks": [ + { + "bbox": [ + 48, + 236, + 545, + 317 + ], + "lines": [ + { + "bbox": [ + 48, + 236, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 48, + 236, + 545, + 317 + ], + "type": "image", + "image_path": "0c8a23ec46a6e56d3b2f31ba9f9d108bfbf5fb8600de7b6bfc14f9653bdfa87c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 319, + 546, + 345 + ], + "lines": [ + { + "bbox": [ + 46, + 319, + 546, + 345 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 546, + 345 + ], + "type": "text", + "content": "Figure 1: Our method synthesizes semantic human motion with natural transition using free-form sequential texts. The skeleton motion can be easily transferred to any character. The character used for demonstration is from Mixamo [1]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 354, + 192, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 354, + 192, + 366 + ], + "spans": [ + { + "bbox": [ + 143, + 354, + 192, + 366 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 377, + 290, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 377, + 290, + 689 + ], + "spans": [ + { + "bbox": [ + 45, + 377, + 290, + 689 + ], + "type": "text", + "content": "The intelligent synthesis/generation of daily-life motion sequences is fundamental and urgently needed for many VR/metaverse-related applications. However, existing approaches commonly focus on monotonic motion generation (e.g., walking, jumping, etc.) based on single instruction-like text, which is still not intelligent enough and can't meet practical demands. To this end, we propose a cohesive human motion sequence synthesis framework based on free-form sequential texts while ensuring semantic connection and natural transitions between adjacent motions. At the technical level, we explore the local-to-global semantic features of previous and current texts to extract relevant information. This information is used to guide the framework in understanding the semantics of the current moment. Moreover, we propose learnable tokens to adaptively learn the influence range of the previous motions towards natural transitions. These tokens can be trained to encode the relevant information into well-designed transition loss. To demonstrate the efficacy of our method, we conduct extensive experiments and comprehensive evaluations on the public dataset as well as a new dataset produced by us. All the experiments confirm that our method outperforms the state-of-the-art methods in terms of semantic matching, realism, and transition fluency. Our project is public available. https://druthrie.github.io/sequential-texts-to-motion/" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 354, + 468, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 354, + 468, + 366 + ], + "spans": [ + { + "bbox": [ + 306, + 354, + 468, + 366 + ], + "type": "text", + "content": "1. Introduction and Motivation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "text", + "content": "Human motion synthesis is fundamental for numerous applications, especially for virtual reality, games, and metaverse-related applications [10, 19, 5, 3, 36, 15, 22, 20, 24], of which, it is in high demand to accurately control the digital human motion with natural language. Existing approaches commonly focus on monotonic motion synthesis/generation based on single instruction-like text description. However, practical applications usually require digital humans to respond to multiple rounds of sustainable interactions, wherein they can continuously generate reasonable responses to the sequential texts. Therefore, given a set of free-form sequential texts, we aim to synthesize a cohesive human motion sequence. Namely, the animation clips should be consistent with the ongoing text descriptions, and the whole motion sequence should have smooth semantic connection and natural transitions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "At present, sequential texts-driven motion synthesis has not been well studied, mainly due to lacking long-term continuous motion datasets with accompanying free-form text descriptions. Recently, TEACH [6] first attempts to address this problem by proposing a dataset (we refer to it as BABEL-TEACH). Each item in the BABEL-TEACH dataset contains two adjacent text-motion pairs. However, TEACH [6] builds on the one-time method TEMOS [29], wherein the limited 5-frame motion is added to encode together. However, it does not consider the potential benefits of previous text information. In practice, previous texts can provide extra and more accurate semantic information for" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 1, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 1, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 1, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 138, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9498" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "current motion synthesis. Therefore, global semantic needs to be extracted from the previous text and the local semantics from the current text, and then fuses the semantics to guide the semantic understanding at each moment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 346 + ], + "type": "text", + "content": "Besides, there are still two main challenges. Firstly, the semantic relationship between adjacent texts is largely overlooked in previous research. For example, when synthesizing a motion for the current text \"touching the face with left hand,\" the previous text context should be considered. Specifically, if the previous text was \"sitting on a chair,\" the synthesized motion should depict a person sitting and touching face with left hand (shown in Fig. 2); if the previous text was \"a person kicks a ball with right foot,\" the motion should show a standing person touching face with left hand. This requires a more comprehensive understanding about the context and semantic relationships between the texts. Secondly, existing methods tend to abrupt transitions between adjacent motions when multiple motions are synthesized separately and stitched together. We aim to address this issue by seamlessly blending the synthesized motions. It should intelligently capture the temporal and spatial continuity between the motions to ensure a more natural and coherent sequence of motions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 348, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 514 + ], + "type": "text", + "content": "Considering the easy scalability of autoregressive methods, we should extend previous autoregressive single text-driven motion synthesis methods to multiple motions involved in motion sequence synthesis. We observe that a similar human body posture can be achieved by transferring the end features of the previous motion to the next motion, although this is not always entirely consistent. This suggests that previous motion information should be leveraged more effectively and sufficiently rather than being adopted straightforwardly. Therefore, we plan to introduce a transition reasoning module that adaptively learns attention score from the previous motion information and infers accurate motion features at the transition, making the transition between adjacent motions more natural and smooth." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 515, + 287, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 287, + 622 + ], + "type": "text", + "content": "To this end, we create a new dataset with much longer-term motions (2-5 times longer) based on a synthetic method, and the corresponding text descriptions are more abundant. We conduct experiments on both BABEL-TEACH [6] dataset and our sequential texts described motion (STDM) dataset. The results demonstrate that our method outperforms existing methods in terms of semantic matching, realism, and transition fluency. Specially, the salient contributions can be summarized as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 627, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 58, + 627, + 286, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 627, + 286, + 686 + ], + "spans": [ + { + "bbox": [ + 58, + 627, + 286, + 686 + ], + "type": "text", + "content": "- We propose a cohesive motions synthesis framework using sequential texts as inputs. The framework can integrate previous text information to obtain semantic features and adaptively select valid previous motion information to guide the current motion synthesis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 689, + 287, + 712 + ], + "type": "text", + "content": "- We design a local-to-global (L2G) semantic fusion module to extract accurate contextual information. It" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 535, + 167 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 535, + 167 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 535, + 167 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 535, + 167 + ], + "type": "image", + "image_path": "657fd4c19d8f689b3c9cc4a36f4b5e6c8b751da6206a706583b27b768b20a515.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 171, + 545, + 206 + ], + "lines": [ + { + "bbox": [ + 305, + 171, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 305, + 171, + 545, + 206 + ], + "type": "text", + "content": "Figure 2: Top row: Motion synthesis corresponding to sequential texts by our method. Bottom row: Motion synthesis corresponding to single text by T2M [11]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 324, + 215, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 215, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 324, + 215, + 545, + 262 + ], + "type": "text", + "content": "can well take into account global-context information from the previous text to infer the current moment's text meaning to guide the semantics-consistent motion synthesis." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 265, + 545, + 386 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 317, + 265, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 265, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 317, + 265, + 545, + 335 + ], + "type": "text", + "content": "- We introduce a transition reasoning module to adaptively select motion snippet from previous motion information to make the synthesized motion natural and smooth w.r.t the previous, wherein a well-defined transition loss is employed to further constrain the fluency of the transitions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 339, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 339, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 339, + 545, + 386 + ], + "type": "text", + "content": "- We create a new sequential texts described motion (STDM) dataset, which involves more extended motion frames and more diverse text descriptions than the existing dataset." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 394, + 392, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 392, + 406 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 392, + 406 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 415, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 415, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 545, + 461 + ], + "type": "text", + "content": "We briefly introduce the related works of human motion synthesis considering semantic control, including single text driven motion synthesis and sequential texts driven motion synthesis." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 463, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 546, + 713 + ], + "type": "text", + "content": "Single text driven motion synthesis. The motion synthesis from text began with the synthesized motion from action labels, but this approach was limited in covering most human motions [21, 27, 35, 38, 28, 12, 28, 7]. Later, researchers [4, 9, 34] focused on synthesizing motion from complex free-form text descriptions, with early work generating single, deterministic motions from text. Recent works have aimed to improve the diversity and details of the synthesized motion. For example, TEMOS [29] and T2M [11] utilized VAE architecture to synthesize diverse motions from the same texts. The main difference between them is that the former utilized Transformer architecture to synthesize all moment motions at once, while the latter utilized GRU architecture to synthesize every moment motion auto-regressively. To capture more detailed information in the text, TM2T [14] utilized VQVAE-based motion markers to provide a fair environment when considering motion and text signals, and utilized motion-to-text analysis module to strengthen the constraints on text-to-motion synthesis. With the successful application of the diffusion model in image generation, MotionDiffuse [37] utilized the diffusion model" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9499" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 541, + 258 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 541, + 258 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 541, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 541, + 258 + ], + "type": "image", + "image_path": "564f1bf0b59f23c839f8645c4126c2ac2d024c60f5b9481f4fb669600c52f322.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 259, + 548, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 259, + 548, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 548, + 319 + ], + "type": "text", + "content": "Figure 3: Method Overview: Our model takes sequential texts as input. In the local-to-global (L2G) semantic fusion module, the previous text's global features and the current text's local features are fused and extracted. In the transition reasoning module, the motion snippet code at the transition of adjacent motion (" + }, + { + "bbox": [ + 46, + 259, + 548, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{start}^{0}" + }, + { + "bbox": [ + 46, + 259, + 548, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 259, + 548, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{start}^{1}" + }, + { + "bbox": [ + 46, + 259, + 548, + 319 + ], + "type": "text", + "content": ") is deduced from the previous motion. Input the gate snippet code " + }, + { + "bbox": [ + 46, + 259, + 548, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^t" + }, + { + "bbox": [ + 46, + 259, + 548, + 319 + ], + "type": "text", + "content": " and the semantic text feature " + }, + { + "bbox": [ + 46, + 259, + 548, + 319 + ], + "type": "inline_equation", + "content": "F_{sem}^{t}" + }, + { + "bbox": [ + 46, + 259, + 548, + 319 + ], + "type": "text", + "content": " to synthesize the motion snippet code " + }, + { + "bbox": [ + 46, + 259, + 548, + 319 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}}_i^t" + }, + { + "bbox": [ + 46, + 259, + 548, + 319 + ], + "type": "text", + "content": ". The transition loss further limits the fluency of the transition with the previous motion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 329, + 288, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 288, + 437 + ], + "type": "text", + "content": "to simulate text driven conditional human motion synthesis, which can better respond to fine-grained instructions of body parts. FLAME [18] utilized transformer-based diffusion model architecture to synthesize human motion, and it can allow editing frames and joints without fine-tuning. Its editing ability can be extended to motion prediction or in-between tasks. Although the diffusion model improves the quality of the synthesized motion, its inference speed is relatively slow, so it is unsuitable for real-time applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 438, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 288, + 715 + ], + "type": "text", + "content": "Sequential texts driven motion synthesis. Synthesizing motion sequences from sequential texts is challenging due to the limited availability of long-term continuous motions datasets with text descriptions. Previous works Action2video [13] realized the transition between actions by changing the input action label type in an autoregressive synthesizer. Yet, they only conducted experiments on a limited number of action labels in three datasets: NTURGBD [33], CMU [2], and humanACT12 [12]. Mao et al. [26] proposed a weakly-supervised action-driven motion prediction method, but they only used 20 action categories with clear transitions in BABEL [32] to synthesize the following motion and transition. Based on the existing BABEL dataset [32], TEACH [6] proposed a new dataset, in which each data consisted of two adjacent text descriptions and motion sequences. The model of TEACH [6] encoded five previous motion frames using TEMOS [29], every single motion sequence was synthesized simultaneously, and adjacent motion sequences were auto-regressively synthesized. Our work builds on this by proposing a cohesive human motion sequence synthesis framework based on freeform sequential texts, ensuring semantic connection and natural transitions between adjacent motions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 328, + 398, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 328, + 398, + 342 + ], + "spans": [ + { + "bbox": [ + 305, + 328, + 398, + 342 + ], + "type": "text", + "content": "3. Our Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 347, + 413, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 347, + 413, + 360 + ], + "spans": [ + { + "bbox": [ + 306, + 347, + 413, + 360 + ], + "type": "text", + "content": "3.1. Method Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 547, + 653 + ], + "type": "text", + "content": "The overall pipeline of the proposed model is shown in Fig. 3. The temporal VAE and motion autoencoder used in the model are similar to those used in the T2M model, wherein it is employed to encode the motion sequence into a motion snippet code sequence and reconstruct the motion sequence with a decoder. To take sequential texts as input and ensure semantically coherent synthesized motions, a local-to-global (L2G) semantic fusion module is introduced (Sec. 3.2). This module guides learning accurate semantic features from the input texts at each moment, including historical text information. To ensure smooth and natural transitions between adjacent motions, a transition reasoning module is introduced (Sec. 3.3). This module dynamically deduces the motion snippet code during transition based on previous motion information, which can help the model generate realistic and coherent transitions. Additionally, a transition loss is proposed to enhance the smoothness of the transition motion (Sec. 3.4). Overall, the proposed pipeline combines several modules and techniques to generate realistic and semantically coherent motion sequences from sequential texts input. The model can generate smooth transitions between motions, and can adaptively learn the influence range of previous motions to further improve the quality of the synthesized motion sequences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Preliminaries. Our method aims to take freeform sequential texts " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = (\\mathbf{S}_1, \\mathbf{S}_2, \\dots, \\mathbf{S}_n)" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": " as input, e.g., " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S} =" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": " (\"salute with the left hand\", \"salute with the right hand\", \"sit on the chair\", \"crawl\"), and outputs corresponding human motion sequences " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{M} =" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "9500" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "(\\mathbf{M}_1, \\mathbf{M}_2, \\dots, \\mathbf{M}_n)" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": ". The transition between adjacent motions should be natural and smooth, and the synthesized motion should reflect the semantic information of previous texts. The data format is consistent with HumanML3D [11] dataset, including the root angular velocity along the Y-axis, the root linear velocity on the XZ-plane, the root height, the local joints positions, velocities, and rotations [40] in root space, while the motions follow the skeleton structure of SMPL [25] with 22 joints. Each moment in our model represents 4 frames." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 198, + 261, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 198, + 261, + 210 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 261, + 210 + ], + "type": "text", + "content": "3.2. Local-to-global Semantic Fusion Module" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 216, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 289, + 396 + ], + "type": "text", + "content": "To synthesize motion that accurately matches the semantics of sequential texts, our model relies on accurately extracting text features at each moment to guide the synthesis process. As the motion we synthesize is based on sequential texts, the current motion is determined not only by the current text semantics but also by previous text semantics. To capture the temporal clues, we encode both the previous and current text, and use an attention mechanism to extract text features that integrate the previous text semantics at each moment (shown in Fig. 4). For previous texts, we use a pretrained CLIP [31] text encoder to extract global features of previous texts. In order to make the features more suitable for our task in the process of training, linear projector is used to extract features further. The global features from the previous texts are expressed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 401, + 287, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 401, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 287, + 414 + ], + "type": "interline_equation", + "content": "\\mathbf {G} _ {p r e} = f \\left(C L I P _ {t e x t} \\left(\\mathbf {S} _ {i - 1}\\right)\\right). \\tag {1}", + "image_path": "230fb2c4e7165bb557e02fbd0a0a9ac877503d94ad776715e92bc617466c1a3c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 418, + 289, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 289, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 289, + 597 + ], + "type": "text", + "content": "It is difficult for the CLIP text encoder to extract the local features of the text, motion synthesis of each moment requires local text semantic guidance, so we employ a Bi-GRU-based text encoder, similar to T2M [11], to obtain the local features " + }, + { + "bbox": [ + 46, + 418, + 289, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{1:m}" + }, + { + "bbox": [ + 46, + 418, + 289, + 597 + ], + "type": "text", + "content": " from the current text. To capture accurate text semantic features at the arbitrary frames, we use an attention mechanism, wherein the key " + }, + { + "bbox": [ + 46, + 418, + 289, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{K}(L)" + }, + { + "bbox": [ + 46, + 418, + 289, + 597 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 46, + 418, + 289, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{V}(L)" + }, + { + "bbox": [ + 46, + 418, + 289, + 597 + ], + "type": "text", + "content": " of attention are the local features of the current text. Since the semantic information of the previous text is required for guidance at the beginning of the currently synthesizing motion, we use the global features of the previous text as the initial query " + }, + { + "bbox": [ + 46, + 418, + 289, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}(G)" + }, + { + "bbox": [ + 46, + 418, + 289, + 597 + ], + "type": "text", + "content": ". At subsequent moments, the motion snippet code generator in temporal VAE provides the queries " + }, + { + "bbox": [ + 46, + 418, + 289, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}(L)" + }, + { + "bbox": [ + 46, + 418, + 289, + 597 + ], + "type": "text", + "content": ". This can be expressed mathematically as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 600, + 247, + 614 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 600, + 247, + 614 + ], + "spans": [ + { + "bbox": [ + 86, + 600, + 247, + 614 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} (G) = \\mathbf {G} _ {p r e} \\mathbf {W} ^ {Q}, \\mathbf {Q} (L) = \\mathbf {h} ^ {t - 1} \\mathbf {W} ^ {Q},", + "image_path": "fa293a2d8fcf23386e27079f51c98c0b9c58f53772a3781704d0deb25ceecf8b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 87, + 616, + 287, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 616, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 87, + 616, + 287, + 635 + ], + "type": "interline_equation", + "content": "\\mathbf {K} (L) = \\mathbf {L} _ {1: m} \\mathbf {W} ^ {K}, \\mathbf {V} (L) = \\mathbf {L} _ {1: m} \\mathbf {W} ^ {V}, \\tag {2}", + "image_path": "9a28a401c92078c25f1b14b9bf7ccd558378da8ea5f56101ec7b858a76cd8780.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 634, + 245, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 634, + 245, + 661 + ], + "spans": [ + { + "bbox": [ + 102, + 634, + 245, + 661 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {s e m} ^ {t} = f \\left(\\frac {\\mathbf {Q} (G / L) \\mathbf {K} (L) ^ {T}}{\\sqrt {d _ {s e m}}}\\right) \\mathbf {V} (L),", + "image_path": "6a26be1e0fd5c73146990ae41fda836e2a0644594937c36b894c9ac08bfca46a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 664, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 664, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 664, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 664, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^k, \\mathbf{W}^V \\in \\mathbb{R}^{d_w \\times d_{sem}}" + }, + { + "bbox": [ + 46, + 664, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 664, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^Q \\in \\mathbb{R}^{d_h \\times d_{sem}}" + }, + { + "bbox": [ + 46, + 664, + 288, + 715 + ], + "type": "text", + "content": " are trainable weights; " + }, + { + "bbox": [ + 46, + 664, + 288, + 715 + ], + "type": "inline_equation", + "content": "d_h, d_w" + }, + { + "bbox": [ + 46, + 664, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 664, + 288, + 715 + ], + "type": "inline_equation", + "content": "d_{sem}" + }, + { + "bbox": [ + 46, + 664, + 288, + 715 + ], + "type": "text", + "content": " are the number of channels in hidden unit " + }, + { + "bbox": [ + 46, + 664, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^{t-1}" + }, + { + "bbox": [ + 46, + 664, + 288, + 715 + ], + "type": "text", + "content": ", current text feature " + }, + { + "bbox": [ + 46, + 664, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{L}^{1:m}" + }, + { + "bbox": [ + 46, + 664, + 288, + 715 + ], + "type": "text", + "content": " and semantic fusion attention layer, respectively. " + }, + { + "bbox": [ + 46, + 664, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^t" + }, + { + "bbox": [ + 46, + 664, + 288, + 715 + ], + "type": "text", + "content": " is the hidden" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 542, + 245 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 542, + 245 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 542, + 245 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 542, + 245 + ], + "type": "image", + "image_path": "e664c64417496ee69981a4c62ac4e251ef467f247ab8e2a2e6c9469cab138718.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 250, + 547, + 312 + ], + "lines": [ + { + "bbox": [ + 305, + 250, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 250, + 547, + 312 + ], + "type": "text", + "content": "Figure 4: Local-to-global semantic fusion module: After extracting the global feature " + }, + { + "bbox": [ + 305, + 250, + 547, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{pre}" + }, + { + "bbox": [ + 305, + 250, + 547, + 312 + ], + "type": "text", + "content": " from the previous texts and the local features " + }, + { + "bbox": [ + 305, + 250, + 547, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{1:m}" + }, + { + "bbox": [ + 305, + 250, + 547, + 312 + ], + "type": "text", + "content": " from the current text, feature fusion is performed to obtain the accurate semantic feature " + }, + { + "bbox": [ + 305, + 250, + 547, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{sem}^{t}" + }, + { + "bbox": [ + 305, + 250, + 547, + 312 + ], + "type": "text", + "content": " of the current text at moment " + }, + { + "bbox": [ + 305, + 250, + 547, + 312 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 250, + 547, + 312 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 329, + 546, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 329, + 546, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 329, + 546, + 365 + ], + "type": "text", + "content": "unit generated by the motion snippet code generator at each moment and " + }, + { + "bbox": [ + 305, + 329, + 546, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^0" + }, + { + "bbox": [ + 305, + 329, + 546, + 365 + ], + "type": "text", + "content": " is obtained from " + }, + { + "bbox": [ + 305, + 329, + 546, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{pre}" + }, + { + "bbox": [ + 305, + 329, + 546, + 365 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 329, + 546, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{sem}^{t}" + }, + { + "bbox": [ + 305, + 329, + 546, + 365 + ], + "type": "text", + "content": " is the text semantic feature at moment " + }, + { + "bbox": [ + 305, + 329, + 546, + 365 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 329, + 546, + 365 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 373, + 468, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 468, + 385 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 468, + 385 + ], + "type": "text", + "content": "3.3. Transition Reasoning Module" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 392, + 547, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 547, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 547, + 619 + ], + "type": "text", + "content": "When synthesizing a new motion, the starting pose is influenced by the ending pose of the previous motion. Grafting the end motion snippet code of the previous motion to the start moment of the current motion can produce similar human poses during the transition. However, the transition between the two motions is still unnatural and abrupt. To address this, we introduce a learnable start motion snippet code token " + }, + { + "bbox": [ + 304, + 392, + 547, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\text {token }}" + }, + { + "bbox": [ + 304, + 392, + 547, + 619 + ], + "type": "text", + "content": " and utilize it along with the previous motion snippet code sequence in a transformer encoder. The output of the token position is used as the start motion snippet code " + }, + { + "bbox": [ + 304, + 392, + 547, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\text {start }}" + }, + { + "bbox": [ + 304, + 392, + 547, + 619 + ], + "type": "text", + "content": " for the current motion, allowing the model to learn the range and weight of the previous motions (shown in Fig. 5). Additionally, the length of the current motion affected by the previous motion information needs to be determined. Through experimentation, we obverse that the best performance is achieved by reasoning the first two motion snippet codes " + }, + { + "bbox": [ + 304, + 392, + 547, + 619 + ], + "type": "inline_equation", + "content": "(\\mathbf{c}_{\\text {start }}^{0}" + }, + { + "bbox": [ + 304, + 392, + 547, + 619 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 392, + 547, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\text {start }}^{1})" + }, + { + "bbox": [ + 304, + 392, + 547, + 619 + ], + "type": "text", + "content": " of the current motion from the previous motions. The verification experiment is in Sec. 4.5. The mathematical expression is as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 628, + 545, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 628, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 315, + 628, + 545, + 653 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {\\text {s t a r t}} ^ {0}, \\mathbf {c} _ {\\text {s t a r t}} ^ {1} = \\text {T r a n s f o r m e r} \\left(\\hat {\\mathbf {c}} _ {i - 1}, \\mathbf {c} _ {\\text {t o k e n}} ^ {0}, \\mathbf {c} _ {\\text {t o k e n}} ^ {1}\\right). \\tag {3}", + "image_path": "cfaf3954054bead90fe74e21934d8e16f44815095b806833b7db7a67fc0ebb48.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "As shown by the transition gate in Fig. 3, when synthesizing the motion snippet codes of the first two moments, we choose the motion snippet codes " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{c}_{start}^{0}" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{start}^{1})" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": " inferred from the previous motion as gate snippet code " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^t" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": ". From the third moment, the motion snippet code " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}}^{t - 1}" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": " from the previ" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 758 + ], + "type": "text", + "content": "9501" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 69, + 265, + 299 + ], + "blocks": [ + { + "bbox": [ + 70, + 69, + 265, + 299 + ], + "lines": [ + { + "bbox": [ + 70, + 69, + 265, + 299 + ], + "spans": [ + { + "bbox": [ + 70, + 69, + 265, + 299 + ], + "type": "image", + "image_path": "229c73e4a307153e29732d9aa1410ef2ce26d964a6e8e67919883968a8f44dd2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 304, + 289, + 390 + ], + "lines": [ + { + "bbox": [ + 46, + 304, + 289, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 289, + 390 + ], + "type": "text", + "content": "Figure 5: Transition reasoning module: Given the previous texts to get the previous synthesized motion sequence " + }, + { + "bbox": [ + 46, + 304, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_{i-1}" + }, + { + "bbox": [ + 46, + 304, + 289, + 390 + ], + "type": "text", + "content": " and motion snippet codes " + }, + { + "bbox": [ + 46, + 304, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}}_{i-1}" + }, + { + "bbox": [ + 46, + 304, + 289, + 390 + ], + "type": "text", + "content": ", we use two learnable tokens (" + }, + { + "bbox": [ + 46, + 304, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{token}^{0}" + }, + { + "bbox": [ + 46, + 304, + 289, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 304, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{token}^{1}" + }, + { + "bbox": [ + 46, + 304, + 289, + 390 + ], + "type": "text", + "content": ") to learn the previous motion information that needs attention, so as to obtain the motion snippet codes at the starting moment of the current motion (" + }, + { + "bbox": [ + 46, + 304, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{start}^{0}" + }, + { + "bbox": [ + 46, + 304, + 289, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 304, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{start}^{1}" + }, + { + "bbox": [ + 46, + 304, + 289, + 390 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 406, + 287, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 287, + 431 + ], + "type": "text", + "content": "ous moment is used, which is synthesized autoregressively. For gate snippet code " + }, + { + "bbox": [ + 46, + 406, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^t" + }, + { + "bbox": [ + 46, + 406, + 287, + 431 + ], + "type": "text", + "content": " , the mathematical expression is as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 437, + 287, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 437, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 121, + 437, + 287, + 476 + ], + "type": "interline_equation", + "content": "\\mathbf {g} ^ {t} = \\left\\{ \\begin{array}{l l} \\mathbf {c} _ {\\text {s t a r t}} ^ {0} & t = 1 \\\\ \\mathbf {c} _ {\\text {s t a r t}} ^ {1} & t = 2 \\\\ \\hat {\\mathbf {c}} ^ {t - 1} & t \\geq 3 \\end{array} \\right. \\tag {4}", + "image_path": "2e06baf2a34596f3c6419830bc7b4c966d57077e7c49fc34c9513be2c08b0c3c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 479, + 288, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 288, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 288, + 528 + ], + "type": "text", + "content": "When synthesizing the motion based on the first text " + }, + { + "bbox": [ + 46, + 479, + 288, + 528 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_1" + }, + { + "bbox": [ + 46, + 479, + 288, + 528 + ], + "type": "text", + "content": " of the sequential texts, we set the previous texts as the default text " + }, + { + "bbox": [ + 46, + 479, + 288, + 528 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_0 = \\text{\"start\"}" + }, + { + "bbox": [ + 46, + 479, + 288, + 528 + ], + "type": "text", + "content": ", and the starting snippet code " + }, + { + "bbox": [ + 46, + 479, + 288, + 528 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{start}^0" + }, + { + "bbox": [ + 46, + 479, + 288, + 528 + ], + "type": "text", + "content": " is obtained from the mean pose." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 534, + 206, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 206, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 206, + 547 + ], + "type": "text", + "content": "3.4. Transition Loss & Total Loss" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 552, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 625 + ], + "type": "text", + "content": "We conduct a statistical analysis of the dataset and find that the motion gap (L1 distance) of almost all adjacent frames is less than 0.1. Inspired by this, we design a transition loss " + }, + { + "bbox": [ + 46, + 552, + 287, + 625 + ], + "type": "inline_equation", + "content": "L_{Trans}" + }, + { + "bbox": [ + 46, + 552, + 287, + 625 + ], + "type": "text", + "content": " to reduce the motion gap between the adjacent frames and avoid abrupt transition. The representation of the " + }, + { + "bbox": [ + 46, + 552, + 287, + 625 + ], + "type": "inline_equation", + "content": "L_{Trans}" + }, + { + "bbox": [ + 46, + 552, + 287, + 625 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 630, + 287, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 630, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 107, + 630, + 287, + 647 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {T r a n s}} = \\left\\| \\hat {\\mathbf {M}} _ {i} ^ {1} - \\hat {\\mathbf {M}} _ {i - 1} ^ {T _ {i} - 1} \\right\\| _ {1}. \\tag {5}", + "image_path": "fc2041e893884656f97c7c18130269359a6cc0be52dd0d560c2b37061bf9c27b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "The final loss function of our model consists of 4 parts: motion snippet code reconstruction loss " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}^{code}" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": ", motion reconstruction loss " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}^{mot}" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": ", KL loss " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KL}" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " of prior distribution and posterior distribution in temporal VAE, and the transition loss " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Trans}" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{code},\\lambda_{mot},\\lambda_{KL},\\lambda_{Trans}" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " are the" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 313, + 70, + 541, + 109 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 541, + 109 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 541, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 541, + 109 + ], + "type": "table", + "html": "
DatasetQuantityDurationVocab.
STDM528920.02 h2455
BABEL-TEACH[6]1626621.86 h1037
", + "image_path": "21d4dbcc3f871464632073cfeb478317d11501b8dc90809bb8d3161441a1e1af.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 118, + 547, + 154 + ], + "lines": [ + { + "bbox": [ + 305, + 118, + 547, + 154 + ], + "spans": [ + { + "bbox": [ + 305, + 118, + 547, + 154 + ], + "type": "text", + "content": "Table 1: Comparison of datasets: STDM dataset has longer-term frames within a single motion, and the text is more abundant." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 173, + 455, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 173, + 455, + 185 + ], + "spans": [ + { + "bbox": [ + 306, + 173, + 455, + 185 + ], + "type": "text", + "content": "corresponding weights of the losses." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 193, + 546, + 219 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 193, + 546, + 219 + ], + "spans": [ + { + "bbox": [ + 306, + 193, + 546, + 219 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {\\text {c o d e}} \\mathcal {L} _ {\\text {r e c}} ^ {\\text {c o d e}} + \\lambda_ {\\text {m o t}} \\mathcal {L} _ {\\text {r e c}} ^ {\\text {m o t}} + \\lambda_ {K L} \\mathcal {L} _ {K L} + \\lambda_ {\\text {T r a n s}} \\mathcal {L} _ {\\text {T r a n s}}. \\tag {6}", + "image_path": "6c83a2d7ad4e5338b9f449228f7c48cb7ca70ca08a8d479e847d13a2652196ee.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 224, + 388, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 224, + 388, + 238 + ], + "spans": [ + { + "bbox": [ + 306, + 224, + 388, + 238 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 243, + 546, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 243, + 546, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 546, + 316 + ], + "type": "text", + "content": "First, we introduce our STDM dataset (Sec. 4.1), evaluation metrics (Sec. 4.2), and then conduct quantitative (Sec. 4.3) and qualitative (Sec. 4.6) evaluations. We carry out an ablation study (Sec. 4.4) to prove the effectiveness of our proposed modules and analyze the parameters (Sec. 4.5)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 324, + 399, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 324, + 399, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 324, + 399, + 335 + ], + "type": "text", + "content": "4.1. STDM Dataset" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 342, + 546, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 342, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 546, + 415 + ], + "type": "text", + "content": "Each text-motion pair in most existing human motion datasets [30, 17, 23, 39, 12, 8] is independent of the other, and does not contain adjacent text-motion pairs. In a recent study, TEACH [6] used the BABEL [32] dataset to obtain a BABEL-TEACH dataset containing two adjacent text-motion pairs, which is consistent with our research goal." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 415, + 547, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 547, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 547, + 666 + ], + "type": "text", + "content": "To obtain adequate motion sequences, we synthesize sequential texts and motion sequences by modifying the existing single text driven motion synthesis method from the viewpoint of data generation instead of using a mocap device. By randomly selecting and annotating texts from a large pool, we create sequential text fed into our improved T2M [11], generating motion sequences. Based on T2M [11], we find that the motion sequences decoded by the same motion snippet code are similar. Our approach diverges from T2M: In the inference stage, instead of using a fixed mean pose, we use snippet codes from the last eight frames of the previous motion, ensuring smoother and more natural transitions. Our experiment results show that the motion sequences synthesized in this way are similar in transition. Since the motion generated by the generation model may not match the semantics of the input text and the transition may not be smooth, we manually remove data with poor transitions and correct erroneous texts to maintain the quality of our STDM dataset. After filtering, we get 5289 pairs containing two adjacent text-motion. Our dataset is collected under the same structure as HumanML3D [11]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 666, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 666, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 666, + 547, + 713 + ], + "type": "text", + "content": "Compared with the BABEL-TEACH [6] dataset, STDM has longer-term motion frames, with a single motion sequence ranging from 2 to 10 seconds, while most BABEL-TEACH [6] data is less than 3 seconds. However, the text" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9502" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 95, + 70, + 500, + 251 + ], + "blocks": [ + { + "bbox": [ + 95, + 70, + 500, + 251 + ], + "lines": [ + { + "bbox": [ + 95, + 70, + 500, + 251 + ], + "spans": [ + { + "bbox": [ + 95, + 70, + 500, + 251 + ], + "type": "image", + "image_path": "768a408617a2d8ad37757a91c9ced008343033504d353dd63dcf4d587d894e98.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 259, + 547, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 259, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 547, + 285 + ], + "type": "text", + "content": "Figure 6: STDM Dataset: Two examples in the STDM dataset, which involves more extended motion frames and more diverse text descriptions than the BABEL-TEACH [6] dataset. Blue traces represent the digital human's trajectory." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 299, + 289, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 299, + 289, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 289, + 371 + ], + "type": "text", + "content": "descriptions in BABEL-TEACH [6] usually describe only one action, such as \"walk forward\". The text descriptions in the STDM are more abundant, such as \"a person raises his right hand and walks forward\". A detailed comparison of the two datasets can be found in Tab. 1. Fig. 6 shows two examples of our STDM dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 387, + 160, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 387, + 160, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 160, + 399 + ], + "type": "text", + "content": "4.2. Evaluation Metrics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 409, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 287, + 529 + ], + "type": "text", + "content": "In order to evaluate the quality of the synthesized motion sequence, we use the feature extractor proposed by Guo et al. [11] to extract the features of text and motion. This feature extractor is trained with contrast loss [16], which can make the matched text-motion pairs produce similar feature vectors. Therefore, the feature extractor can measure whether the text and motion sequence match. We combine two adjacent text-motion pairs into one to train the feature extractor. We use the same evaluation metrics as T2M [11]. The metrics are described in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 532, + 288, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 288, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 288, + 593 + ], + "type": "text", + "content": "To demonstrate the fluency of different methods when transiting between adjacent motions, we add a metric to evaluate the fluency of transition, by calculating the distance between adjacent frames of two adjacent motion sequences. For transition fluency, the mathematical expression is as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 112, + 606, + 287, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 606, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 112, + 606, + 287, + 639 + ], + "type": "interline_equation", + "content": "F = \\sum_ {i = 1} ^ {N} \\| \\mathbf {M} _ {i 1} ^ {T 1} - \\mathbf {M} _ {i 2} ^ {1} \\| _ {1}. \\tag {7}", + "image_path": "c16c51f3949021c3e280cc96cde97890c29104b67d4f9b4060d260946fe054a2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 654, + 288, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 717 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 717 + ], + "type": "text", + "content": "Here, we calculate the average transition distance (" + }, + { + "bbox": [ + 46, + 654, + 288, + 717 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 654, + 288, + 717 + ], + "type": "text", + "content": " distance) of the motion sequence synthesized by " + }, + { + "bbox": [ + 46, + 654, + 288, + 717 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 654, + 288, + 717 + ], + "type": "text", + "content": " adjacent text pairs. For the " + }, + { + "bbox": [ + 46, + 654, + 288, + 717 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 654, + 288, + 717 + ], + "type": "text", + "content": "-th adjacent text pair, the synthesized adjacent motion sequence pair is " + }, + { + "bbox": [ + 46, + 654, + 288, + 717 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{i1} = (\\mathbf{M}_{i1}^{1}, \\mathbf{M}_{i1}^{2}, \\dots, \\mathbf{M}_{i1}^{T1})" + }, + { + "bbox": [ + 46, + 654, + 288, + 717 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 654, + 288, + 717 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{i2} = (\\mathbf{M}_{i2}^{1}, \\mathbf{M}_{i2}^{2}, \\dots, \\mathbf{M}_{i2}^{T2})" + }, + { + "bbox": [ + 46, + 654, + 288, + 717 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 298, + 441, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 298, + 441, + 311 + ], + "spans": [ + { + "bbox": [ + 305, + 298, + 441, + 311 + ], + "type": "text", + "content": "4.3.Quantitative Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 317, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 545 + ], + "type": "text", + "content": "Baselines. We compare our method with T2M [11], our two variants T2M-Joint and Complusion-Code, and TEACH [6]. For T2M, we input a single text, output a single motion, and align the root coordinates of the last frame of the previous motion with the root coordinates of the first frame of the next motion. For T2M-Joint, we keep the T2M model structure and train the model by combining two adjacent texts with commas as input. For Complusion-Code, based on the T2M [11], the last motion snippet code of the previous motion is used as the start motion snippet code of the next motion during training (Unlike the STDM dataset generation method, the dataset generation only uses the snippet codes of previous motions in the inference phase. Complusion-Code also uses the snippet codes of previous motions in the training phase). For TEACH [6], we compared two versions using spherical linear interpolation (Slerp) and not using Slerp, respectively. Slerp operation is unrelated to the TEACH [6] model and is simply a way of post-processing synthesized data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 546, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 546, + 714 + ], + "type": "text", + "content": "Tab. 2 and Tab. 3 show our quantitative comparison results with other baselines on BABEL-TEACH [6] and STDM datasets. We divide the training, testing, and validation sets according to the ratio of " + }, + { + "bbox": [ + 304, + 546, + 546, + 714 + ], + "type": "inline_equation", + "content": "0.8:0.15:0.05" + }, + { + "bbox": [ + 304, + 546, + 546, + 714 + ], + "type": "text", + "content": ". For a fair comparison, we conduct 20 experiments and show the values with confidence in the " + }, + { + "bbox": [ + 304, + 546, + 546, + 714 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 304, + 546, + 546, + 714 + ], + "type": "text", + "content": " range. From the results, our method outperforms all other methods, especially in the transition fluency metric, which is important for synthesizing natural and smooth transition motion. T2M [11] can only synthesize a single motion per inference, and there is no connection between adjacent motions, so it performs poorly in all indicators. Because our testing data are composed of two adjacent texts, the T2M-Joint method performs similarly to the real data regarding transition fluency" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9503" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 87, + 70, + 504, + 185 + ], + "blocks": [ + { + "bbox": [ + 87, + 70, + 504, + 185 + ], + "lines": [ + { + "bbox": [ + 87, + 70, + 504, + 185 + ], + "spans": [ + { + "bbox": [ + 87, + 70, + 504, + 185 + ], + "type": "table", + "html": "
MethodsR Precision↑FID↓MultiModal Dist↓Transition Fluency↓
Top 1Top 2Top 3
Real motions0.560±0.0020.750±0.0030.834±0.0020.001±0.0003.448±0.001-
T2M [11]0.376±0.0030.543±0.0020.647±0.0036.890±0.0815.218±0.0101.108±0.004
T2M-Joint0.373±0.0040.534±0.0030.640±0.0032.727±0.0324.983±0.011-
Compulsion-Code0.432±0.0030.613±0.0040.717±0.0033.726±0.0684.612±0.0230.348±0.002
TEACH (no Slerp) [6]0.573±0.0020.756±0.0030.842±0.0022.263±0.0553.487±0.0100.577±0.003
TEACH (Slerp) [6]0.563±0.0030.750±0.0030.839±0.0032.240±0.0533.526±0.0110.118±0.001
Ours0.542±0.0030.728±0.0030.818±0.0031.628±0.0313.662±0.0100.177±0.001
Ours (Slerp)0.519±0.0040.705±0.0040.803±0.0031.548±0.0243.797±0.0120.036±0.000
", + "image_path": "acbd463e546943742ce9c5d610040bb1ed822ad27aa242b706d78251e35479cd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 87, + 247, + 504, + 365 + ], + "blocks": [ + { + "bbox": [ + 45, + 187, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 45, + 187, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 45, + 187, + 547, + 237 + ], + "type": "text", + "content": "Table 2: Quantitative evaluation on testing data of BABEL-TEACH: ± indicates 95% confidence interval, ↑ and ↓ respectively denotes better performance with larger or lower value. Bold indicates the optimal result, the underscore represents the suboptimal result, while wave line refers to the third best. The results show that the motion synthesized by our model outperforms other baselines in terms of semantic matching, transition fluency, and realism." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 87, + 247, + 504, + 365 + ], + "lines": [ + { + "bbox": [ + 87, + 247, + 504, + 365 + ], + "spans": [ + { + "bbox": [ + 87, + 247, + 504, + 365 + ], + "type": "table", + "html": "
MethodsR Precision↑FID↓MultiModal Dist↓Transition Fluency↓
Top 1Top 2Top 3
Real motions0.322±0.0040.505±0.0050.625±0.0060.014±0.0023.461±0.0050.454±0.000
T2M [11]0.322±0.0060.493±0.0050.614±0.0051.389±0.0493.500±0.0090.639±0.003
T2M-Joint0.304±0.0050.488±0.0060.613±0.0050.744±0.0423.492±0.011-
Compulsion-Code0.285±0.0070.450±0.0060.572±0.0063.146±0.0803.806±0.0170.330±0.003
TEACH (no Slerp) [6]0.318±0.0060.505±0.0050.634±0.0051.414±0.0553.483±0.0120.246±0.002
TEACH (Slerp) [6]0.326±0.0040.504±0.0050.631±0.0061.416±0.0543.487±0.0130.049±0.000
Ours0.328±0.0060.510±0.0060.633±0.0051.085±0.0633.441±0.0170.109±0.001
Ours (Slerp)0.325±0.0070.515±0.0070.637±0.0051.134±0.0593.440±0.0130.022±0.000
", + "image_path": "471ab5bdd2eda13748aa515ffb5b827d896a7afdda2a5729d78d6a77e97e757b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 367, + 546, + 392 + ], + "lines": [ + { + "bbox": [ + 45, + 367, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 45, + 367, + 546, + 392 + ], + "type": "text", + "content": "Table 3: Quantitative evaluation on the testing data of STDM: The results show that the motion synthesized by our model outperforms other baselines in terms of semantic matching, transition fluency, and realism." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 407, + 289, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 289, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 289, + 646 + ], + "type": "text", + "content": "(less than 0.1). However, when the input text consists of more than two sentences, the T2M-Joint method cannot guarantee transition fluency. Because Compulsion-Code uses the end motion snippet code of the previous motion as the start snippet code of the next motion, the performance of transition fluency is better than T2M [11], but it is still not as good as ours. Although TEACH [6] is slightly better than ours in R Precision and MultiModal Distance on the BABEL-TEACH [6] dataset, as shown in Tab. 2, its transition fluency is not as good as ours, we can obtain transitional and natural motion sequences without Slerp operation, so our method is better than TEACH [6] in a comprehensive evaluation. Although the synthetic STDM dataset is less natural than actual mocap data in transition fluency, we find that our method can compensate for the defects of the training data after training. In terms of transition fluency, the results synthesized by our method outperform the ground truth of the testing data. This further verifies the effectiveness of our method in synthesizing natural and smooth transition motions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": "User study. We randomly select 50 texts from the testing data and arrange 3 to 5 texts together to form a set of sequential texts. Some text combinations are not present in the dataset to test the motion generation effect of each model for text combinations not in the dataset. The syn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 407, + 547, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 547, + 504 + ], + "type": "text", + "content": "thesized motions based on these sequential text groups are randomly shuffled and presented using different methods. Users are required to score synthetic motions from the following three aspects: (1) Matching degree: the degree of semantic matching between the motion and the text; (2) Transition fluency: the natural degree of transition between adjacent motions; (3) Realism: how realistic the motion is. The result is shown in Fig. 7." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 513, + 400, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 513, + 400, + 526 + ], + "spans": [ + { + "bbox": [ + 305, + 513, + 400, + 526 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 533, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 581 + ], + "type": "text", + "content": "In order to illustrate the influence of the modules and training strategies we introduced on the overall model, we carried out the ablation experiment. The results are shown in Tab. 4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": "The first experiment is to remove the transition reasoning module. We set the skeleton to mean pose to initialize the motion snippet code at the start moment. However, the model does not know the human posture at the beginning, so it greatly reduces transition fluency (value increased by 0.417). The experiment results show that the performance of the model will decline without the transition reasoning module because it does not aware of the information of previous motion when the model synthesizes the motion sequence of the current text, even though the model can synthesize motion sequences that match the semantics of the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "text", + "content": "9504" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 70, + 520, + 161 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 520, + 161 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 520, + 161 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 520, + 161 + ], + "type": "table", + "html": "
MethodsR Precision↑FID↓MultiModal Dist↓Transition Fluency↓
Top 1Top 2Top 3
Real motions0.561±0.0030.747±0.0030.836±0.0020.002±0.0003.449±0.002-
w/o Transition reasoning module0.535±0.0040.713±0.0030.805±0.0032.190±0.0383.719±0.0130.594±0.003
w/o L2G semantic fusion module0.440±0.0030.621±0.0030.720±0.0034.592±0.0714.593±0.0130.151±0.001
w/o Transition loss0.525±0.0040.706±0.0040.801±0.0021.766±0.0363.797±0.0110.293±0.001
w/o Segmented training strategy0.500±0.0020.684±0.0040.783±0.0042.164±0.0384.041±0.0130.328±0.002
Ours0.542±0.0030.726±0.0030.817±0.0021.588±0.0303.655±0.0130.177±0.001
", + "image_path": "21fcdca1e37f74135b637dcf32a19139aacd6a414cb48778e0ebdebb2efc633e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 83, + 223, + 509, + 301 + ], + "blocks": [ + { + "bbox": [ + 46, + 167, + 547, + 215 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 547, + 215 + ], + "type": "text", + "content": "Table 4: Ablation study: The results show that the transition reasoning module has the most significant impact on transition fluency, the L2G semantic fusion module has the most significant impact on R Precision, FID, and MultiModel Distance, which represent semantic matching and realism, and the transition loss and segmentation training strategy have a certain influence on all indicators." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 223, + 509, + 301 + ], + "lines": [ + { + "bbox": [ + 83, + 223, + 509, + 301 + ], + "spans": [ + { + "bbox": [ + 83, + 223, + 509, + 301 + ], + "type": "table", + "html": "
Numbers of snippet codeR Precision↑FID↓MultiModal Dist↓Transition Fluency↓
Top 1Top 2Top 3
Real motions0.561±0.0030.750±0.0020.834±0.0020.002±0.0003.449±0.001-
1 snippet code0.517±0.0030.693±0.0040.786±0.0042.027±0.0433.845±0.0140.173±0.001
2 snippet codes0.540±0.0030.725±0.0030.816±0.0021.602±0.0263.662±0.0110.177±0.001
3 snippet codes0.522±0.0040.702±0.0030.796±0.0022.050±0.0383.838±0.0150.175±0.001
5 snippet codes0.504±0.0030.687±0.0040.785±0.0031.781±0.0293.913±0.0130.213±0.001
", + "image_path": "a1746572055798abcbc12516db68c65adb872abab14796bc08435158ddc44f33.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 307, + 547, + 344 + ], + "lines": [ + { + "bbox": [ + 46, + 307, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 547, + 344 + ], + "type": "text", + "content": "Table 5: Parameter analysis of transition reasoning module: We infer the initial motion snippet code of the current motion sequence from the previous motion and determine the optimal number of the inference code through experiments. The result shows that inferring the first two codes of current motion from the previous motion work best." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 52, + 352, + 279, + 523 + ], + "blocks": [ + { + "bbox": [ + 52, + 352, + 279, + 523 + ], + "lines": [ + { + "bbox": [ + 52, + 352, + 279, + 523 + ], + "spans": [ + { + "bbox": [ + 52, + 352, + 279, + 523 + ], + "type": "image", + "image_path": "2f65555da25ba3cdd78d80957654a4b2c233ac97ed3f40765f4575347f5c06d2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 526, + 287, + 586 + ], + "lines": [ + { + "bbox": [ + 46, + 526, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 526, + 287, + 586 + ], + "type": "text", + "content": "Figure 7: User study: We invite 45 users to score the results of our comparative experiments, and the color bars in the figure indicate the percentage of the scores. The results show that our method outperforms other baselines in terms of semantic matching, transition fluency, and realism." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 594, + 97, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 97, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 97, + 604 + ], + "type": "text", + "content": "current text." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "The second experiment is to remove the L2G semantic fusion module. We only encode the features of the current text at each moment without considering the previous text. The experiment results show that the model's performances on R Precision, FID, and MultiModal Distance decrease significantly after removing the L2G semantic fusion module, indicating that previous text information is essential in synthesizing motion sequences that match context semantics. Although there is a slight improvement in the metric of" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 353, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 545, + 411 + ], + "type": "text", + "content": "transition fluency after the removal, the gap between them can not be observed from the naked eye. With the L2G semantic fusion module, our model shows better competence in evaluating the implicit relationship between current and previous text." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 412, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 483 + ], + "type": "text", + "content": "The third experiment is to remove the transition loss. The experimental results show that after removing the transition loss, the model's performances in all indicators have decreased, indicating that adding transition loss will guide the model to learn more semantic matching and transitional natural motion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 484, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 581 + ], + "type": "text", + "content": "The fourth experiment does not use the segmented training strategy (explained in the supplementary file). The specific approach is that all motion sequences are extended to the maximum length (196 frames) with 0 in the training strategy. The experimental results show that when the segmented training strategy is not used, the performance of each indicator decreases, indicating that the segmented training strategy can make the model learn more thoroughly." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 587, + 421, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 421, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 421, + 600 + ], + "type": "text", + "content": "4.5. Parameter Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "In the transition reasoning module, we verify through experiments that we need to infer the number of motion snippet codes from the previous motions at the current initial stage. As shown in Tab. 5, the reasoning motion snippet code number of 1, 2, 3, and 5 are compared. The experiment results show that two motion snippet codes at the beginning of reasoning are optimal because the model can not fully capture the influence of previous motion information by reducing motion snippets inferred from previous motions. At" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9505" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 533, + 392 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 533, + 392 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 533, + 392 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 533, + 392 + ], + "type": "image", + "image_path": "6fd40e06859ddfd216f4e4524a461523b3b2e08f254bd51269083df56b687dbc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 397, + 547, + 447 + ], + "lines": [ + { + "bbox": [ + 46, + 397, + 547, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 547, + 447 + ], + "type": "text", + "content": "Figure 8: Qualitative evaluation: Continuously inputting four texts to show the visual results of our method and other baseline methods, except for the two frames between adjacent motions, the rest only show the keyframes. The results show that the motion synthesized by our method is the best in terms of semantic matching, transition fluency, and realism. More results are in the supplementary file." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 458, + 288, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 288, + 495 + ], + "type": "text", + "content": "the same time, the model will introduce too much previous motion information into the current motion if increasing the motion snippets inferred from previous motions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 498, + 176, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 176, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 176, + 510 + ], + "type": "text", + "content": "4.6. Qualitative Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 513, + 287, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 513, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 513, + 287, + 632 + ], + "type": "text", + "content": "We make a qualitative comparison with the baselines, as shown in Fig. 8, our synthesized motion is the most natural in transition and nicely matches the semantics of the texts. Transition in T2M [11] is most evident in adjacent motion. T2M-Joint cannot fully express the content of the text description. When the input sequential texts contain more than two sentences, noticeable changes in the adjacent motion transition will occur. Complusion-Code produces a slight change in the transition. There is some semantic mismatch in TEACH [6] and abrupt transitions in the no Slerp version." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 636, + 222, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 636, + 222, + 649 + ], + "spans": [ + { + "bbox": [ + 47, + 636, + 222, + 649 + ], + "type": "text", + "content": "5. Conclusions and Future Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 654, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 713 + ], + "type": "text", + "content": "We have advocated a novel framework for synthesizing semantic motion sequences driven by sequential texts while ensuring natural transitions between adjacent motions. Our approach leveraged a local-to-global semantic fusion module and a transition reasoning module to over" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 458, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 506 + ], + "type": "text", + "content": "come the challenges of semantic association and transition fluency. Extensive experiments show that our method outperforms state-of-the-art techniques in terms of semantic matching, transition fluency, and realism." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 506, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 590 + ], + "type": "text", + "content": "However, our method still has limitations, such as difficulty in synthesizing head motions like \"nodding\" and \"shaking head\". In future work, we plan to improve this by exploring a way to assign different attention weights to different joints automatically. Additionally, since our STDM dataset is synthesized while not captured, we will further improve the data quality and expand the dataset scale." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 597, + 409, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 409, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 409, + 611 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 615, + 545, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 615, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 304, + 615, + 545, + 711 + ], + "type": "text", + "content": "This paper is supported by National Natural Science Foundation of China (62272021, 62102036), Beijing Natural Science Foundation (4222024), R&D Program of Beijing Municipal Education Commission (KM202211232003), Open Project Program of State Key Laboratory of Virtual Reality Technology and Systems, Beihang University (No.VRLAB2022A02), Research Unit of Virtual Human and Virtual Surgery (2019RU004)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "text", + "content": "9506" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "type": "text", + "content": "[1] Mixamo. https://www MIXamo.com. Accessed February 17, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 136 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 136 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 136 + ], + "type": "text", + "content": "[2] CMU graphics lab motion capture database. http://mocap.cs.cmu.edu/, 2003." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 138, + 286, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 138, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 138, + 286, + 182 + ], + "type": "text", + "content": "[3] Chaitanya Ahuja, Dong Won Lee, Yukiko I.Nakano, et al. Style transfer for co-speech gesture animation: A multi-speaker conditional-mixture approach. European Conference on Computer Vision, pages 248-265, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 183, + 286, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 183, + 286, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 183, + 286, + 226 + ], + "type": "text", + "content": "[4] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. 2019 International Conference on 3D Vision (3DV), pages 719-728, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 228, + 286, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 286, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 286, + 271 + ], + "type": "text", + "content": "[5] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, et al. Style-controllable speech-driven gesture synthesis using normalising flows. Computer Graphics Forum, 39(2):487-496, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 274, + 286, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 286, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 286, + 316 + ], + "type": "text", + "content": "[6] Nikos Athanasiou, Mathis Petrovich, Michael J.Black, et al. TEACH: Temporal action composition for 3d humans. 2022 International Conference on 3D Vision (3DV), pages 414-423, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 319, + 286, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 286, + 362 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 286, + 362 + ], + "type": "text", + "content": "[7] Pablo Cervantes, Yusuke Sekikawa, Ikuro Sato, et al. Implicit neural representations for variable length human motion generation. European Conference on Computer Vision, pages 356-372, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 364, + 286, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 364, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 364, + 286, + 407 + ], + "type": "text", + "content": "[8] Jihoon Chung, Cheng hsin Wuu, Hsuan ru Yang, et al. HAA500: Human-centric atomic action dataset with curated videos. IEEE/CVF International Conference on Computer Vision, pages 13465-13474, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 409, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 286, + 453 + ], + "type": "text", + "content": "[9] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, et al. Synthesis of compositional animations from textual descriptions. IEEE/CVF International Conference on Computer Vision, pages 1396-1406, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 454, + 286, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 286, + 498 + ], + "type": "text", + "content": "[10] Shiry Ginosar, Amir Bar, Gefen Kohavi, et al. Learning individual styles of conversational gesture. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 499, + 286, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 286, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 286, + 543 + ], + "type": "text", + "content": "[11] Chuan Guo, Shihao Zou, Xinxin Zuo, et al. Generating diverse and natural 3d human motions from text. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 544, + 286, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 286, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 286, + 587 + ], + "type": "text", + "content": "[12] Chuan Guo, Xinxin Zuo, Sen Wang, et al. Action2motion: Conditioned generation of 3d human motions. 28th ACM International Conference on Multimedia, pages 2021-2029, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 590, + 286, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 286, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 286, + 623 + ], + "type": "text", + "content": "[13] Chuan Guo, Xinxin Zuo, Sen Wang, et al. Action2video: Generating videos of human 3d actions. International Journal of Computer Vision, 130(2):285-315, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 286, + 667 + ], + "type": "text", + "content": "[14] Chuan Guo, Xinxin Zuo, Sen Wang, et al. TM2T: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. European Conference on Computer Vision, pages 580-597, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "text", + "content": "[15] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, et al. Learning speech-driven 3d conversational gestures from video. 21st ACM International Conference on Intelligent Virtual Agents, pages 101-108, 2021." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[16] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2:1735-1742, 2006." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[17] Yanli Ji, Feixiang Xu, Yang Yang, et al. A large-scale RGB-D database for arbitrary-view human action recognition. 26th ACM International Conference on Multimedia, pages 1510-1518, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "text", + "content": "[18] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. arXiv preprint arXiv:2209.00349, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "text", + "content": "[19] Taras Kucherenko, Patrik Jonell, Sanne van Waveren, et al. Gesticulator: A framework for semantically-aware speech-driven gesture generation. 2020 International Conference on Multimodal Interaction, pages 242-250, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "text", + "content": "[20] Yuanzhi Liang, Qianyu Feng, Linchao Zhu, et al. SEEG: Semantic energized co-speech gesture generation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10473-10482, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 286, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 308 + ], + "type": "text", + "content": "[21] Xiao Lin and Mohamed R Amer. Human motion modeling using dvgens. arXiv preprint arXiv:1804.10652, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "type": "text", + "content": "[22] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, et al. BEAT: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. arXiv preprint arXiv:2203.05297, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "type": "text", + "content": "[23] Jun Liu, Amir Shahroudy, Mauricio Perez, et al. NTU rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2684-2701, 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "text", + "content": "[24] Xian Liu, Qianyi Wu, Hang Zhou, et al. Learning hierarchical cross-modal association for co-speech gesture generation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 444, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 477 + ], + "type": "text", + "content": "[25] Matthew Loper, Naureen Mahmood, Naureen Mahmood, et al. SMPL: A skinned multi-person linear model. Acm Transactions on Graphics, 34(6cd):248, 2015." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "text", + "content": "[26] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. Weakly-supervised action transition learning for stochastic human motion prediction. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8151-8160, 2022." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "text", + "content": "[27] Dario Pavllo, David Grangier, and Michael Auli. Quaternet: A quaternion-based recurrent model for human motion. arXiv preprint arXiv:1805.06485, 2018." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 556, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 600 + ], + "type": "text", + "content": "[28] Mathis Petrovich, Michael J.Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "text", + "content": "[29] Mathis Petrovich, Michael J.Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. arXiv preprint arXiv:2204.14109, 2022." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "type": "text", + "content": "[30] Matthias Plappert, Christian Mandery, and Tamim Asfour. The KIT motion-language dataset. Big Data, 4(4):236-252, 2016." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, et al. Learning transferable visual models from natural language supervision. International Conference on Machine Learning, pages 8748-8763, 2021." + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9507" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 443 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[32] Abhinanda R.Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, et al. BABEL: Bodies, action and behavior with english labels. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 722-731, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "text", + "content": "[33] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, et al. Nturogb+ d: A large scale dataset for 3d human activity analysis. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1010-1019, 2016." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 288, + 196 + ], + "type": "text", + "content": "[34] Guy Tevet, Brian Gordon, Amir Hertz, et al. MotionCLIP: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 288, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 288, + 240 + ], + "type": "text", + "content": "[35] Sijie Yan, Zhizhong Li, Yuanjun Xiong, et al. Convolutional sequence generation for skeleton-based action synthesis. IEEE/CVF International Conference on Computer Vision, pages 4394-4402, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 241, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 288, + 285 + ], + "type": "text", + "content": "[36] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, et al. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics, 39(6):1-16, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 286, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 286, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 286, + 288, + 319 + ], + "type": "text", + "content": "[37] Mingyuan Zhang, Zhongang Cai, Liang Pan, et al. Motion-diffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 320, + 288, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 320, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 49, + 320, + 288, + 353 + ], + "type": "text", + "content": "[38] Yan Zhang, Michael J.Black, and Siyu Tang. Perpetual motion: Generating unbounded human motion. arXiv preprint arXiv:2007.13886, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 354, + 288, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 354, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 49, + 354, + 288, + 398 + ], + "type": "text", + "content": "[39] Hang Zhao, Antonio Torralba, Lorenzo Torresani, et al. Hacs: Human action clips and segments dataset for recognition and temporal localization. IEEE/CVF International Conference on Computer Vision, pages 8668-8678, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 399, + 288, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 399, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 49, + 399, + 288, + 443 + ], + "type": "text", + "content": "[40] Yi Zhou, Connelly Barnes, Jingwan Lu, et al. On the continuity of rotation representations in neural networks. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2019." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "9508" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Set-level Guidance Attack_ Boosting Adversarial Transferability of Vision-Language Pre-training Models/4d2215bf-24a1-45b3-bcea-24623e459934_content_list.json b/2023/Set-level Guidance Attack_ Boosting Adversarial Transferability of Vision-Language Pre-training Models/4d2215bf-24a1-45b3-bcea-24623e459934_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fb54f7888d2864184123308d5ff741ec4e5fdf34 --- /dev/null +++ b/2023/Set-level Guidance Attack_ Boosting Adversarial Transferability of Vision-Language Pre-training Models/4d2215bf-24a1-45b3-bcea-24623e459934_content_list.json @@ -0,0 +1,1665 @@ +[ + { + "type": "text", + "text": "Set-level Guidance Attack: Boosting Adversarial Transferability of Vision-Language Pre-training Models", + "text_level": 1, + "bbox": [ + 148, + 130, + 823, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dong Lu1, Zhiqiang Wang1, Teng Wang1,2, Weili Guan3, Hongchang Gao4, Feng Zheng1,5†", + "bbox": [ + 125, + 202, + 841, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Southern University of Science and Technology $^{2}$ The University of Hong Kong", + "bbox": [ + 245, + 223, + 723, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Monash University 4Temple University 5Peng Cheng Laboratory", + "bbox": [ + 287, + 239, + 681, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sammylu\\_outlook.com wangzq_2021\\@outlook.com tengwang@connect.hku.hk", + "bbox": [ + 184, + 258, + 781, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "honeyguan@gmail.com hongchang.gao@temple.edu f.zheng@ieee.org", + "bbox": [ + 212, + 277, + 754, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language pre-training (VLP) models have shown vulnerability to adversarial examples in multimodal tasks. Furthermore, malicious adversaries can be deliberately transferred to attack other black-box models. However, existing work has mainly focused on investigating white-box attacks. In this paper, we present the first study to investigate the adversarial transferability of recent VLP models. We observe that existing methods exhibit much lower transferability, compared to the strong attack performance in white-box settings. The transferability degradation is partly caused by the under-utilization of cross-modal interactions. Particularly, unlike unimodal learning, VLP models rely heavily on cross-modal interactions and the multimodal alignments are many-to-many, e.g., an image can be described in various natural languages. To this end, we propose a highly transferable Set-level Guidance Attack (SGA) that thoroughly leverages modality interactions and incorporates alignment-preserving augmentation with cross-modal guidance. Experimental results demonstrate that SGA could generate adversarial examples that can strongly transfer across different VLP models on multiple downstream vision-language tasks. On image-text retrieval, SGA significantly enhances the attack success rate for transfer attacks from ALBEF to TCL by a large margin (at least $9.78\\%$ and up to $30.21\\%$ ), compared to the state-of-the-art. Our code is available at https://github.com/Zoky-2020/SGA.", + "bbox": [ + 75, + 358, + 473, + 751 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 777, + 209, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent work has shown that vision-language pre-training (VLP) models are still vulnerable to adversarial examples [41], even though they have achieved remarkable performance on a wide range of multimodal tasks [30, 15, 16]. Existing work mainly focuses on white-box attacks, where", + "bbox": [ + 75, + 803, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/57abe961e11c257578b9d684d65e3eb931de0acc13b89bc92fa04105a5eba159.jpg", + "image_caption": [ + "(a) ALBEF to ALBEF" + ], + "image_footnote": [], + "bbox": [ + 506, + 327, + 696, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/703a6a22fcd62c73203da94b438e243c35a7028ca3d55fc818e0500d0ab6f195.jpg", + "image_caption": [ + "(b) ALBEF to TCL" + ], + "image_footnote": [], + "bbox": [ + 696, + 325, + 883, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/79b7480f3214c69978d30e31bb2733b3dfefb4eca3b2d5596566f43489447b39.jpg", + "image_caption": [ + "(c) ALBEF to $\\mathrm{CLIP}_{\\mathrm{VIT}}$", + "Figure 1: Comparison of attack success rates (ASR) using five different attacks on image-text retrieval. Adversarial examples are crafted on the source model (ALBEF) to attack the target white-box model or black-box models. The first three columns refer to the image-only PGD attack [24], text-only BERT-Attack [19] (BA), and the combined separate unimodal attack (SA), which all belong to the methods without cross-modal interactions. The fourth column is the state-of-the-art multimodal Co-Attack [41] (CA) that employs single-pair cross-modal interactions. The last column is the proposed Set-level Guidance Attack (SGA), which leverages multiple set-level cross-modal interactions, successfully attacking the white-box model and transferring to attack all black-box models with the highest ASR. More discussions are in Section 3." + ], + "image_footnote": [], + "bbox": [ + 506, + 449, + 694, + 506 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cab211f7941f67f431d4ce6995fa79556eb3bb5696d8b8da33990891b2e141a3.jpg", + "image_caption": [ + "(d) ALBEF to $\\mathrm{CLIP}_{\\mathrm{CNN}}$" + ], + "image_footnote": [], + "bbox": [ + 697, + 450, + 885, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "information about the victim model is accessible. However, the transferability of adversarial examples across VLP models has not been investigated, which is a more practical setting. It is still unknown whether the adversarial data generated on the source model can successfully attack another model, which poses a serious security risk to the deployment of VLP models in real-world applications.", + "bbox": [ + 496, + 777, + 893, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper makes the first step to investigate the trans", + "bbox": [ + 519, + 885, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Corresponding author.", + "bbox": [ + 94, + 886, + 339, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "102", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/092ee924aedd1ff0cfc95c7e445420276c5b21a939008f621bb56c95c838490b.jpg", + "image_caption": [ + "(a) Sep-Attack" + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 132, + 186 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a59f5fb7f155a9f918d1474c0d37ac11b00d845839703e7bc2b47746618723ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 90, + 181, + 186 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3102c8a89a59e8d2a501f6e7ed757080b89fbaefed782e5e8ad9d6a13833fd86.jpg", + "image_caption": [ + "(b) Co-Attack" + ], + "image_footnote": [], + "bbox": [ + 189, + 90, + 299, + 186 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cae2c94049f04114f405fa8b9b227919dd7078f15ef9c244058a2e99349902f6.jpg", + "image_caption": [ + "(c) SGA", + "Figure 2: Comparison of cross-modal interactions. To generate adversarial examples, existing methods either: (a) separately perturb unimodal data without any cross-modal interactions (Sep-Attack) or, (b) perturb multiple modalities but with single image-text pairs to model cross-modal interactions (Co-Attack). However, our method is capable of learning cross-modal interactions among multiple alignments through (c) set-level guidance (SGA). Specifically, $v$ denotes the input image, and $t$ is the paired caption. $v'$ and $t'$ represent the corresponding adversarial examples, respectively. $U(v)$ and $U(t)$ represent the scale-invariant image set and most matching caption set. Arrows indicate the guidance for generating adversarial examples." + ], + "image_footnote": [], + "bbox": [ + 308, + 90, + 464, + 186 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ferability of adversarial samples within VLP models. Without loss of generality, most of our experiments are based on image-text retrieval tasks. We first empirically evaluate this attack performance with respect to different modalities on multimodal tasks across multiple datasets. Our results show that the adversarial transferability of attacking both modalities (image & text) consistently beats attacking unimodal data (image or text). Unfortunately, even though two modalities are allowed to be perturbed simultaneously, the attack success rates of existing methods [24, 19, 41] still significantly drops when transferring from the white-box to black-box settings, as shown in Figure 1.", + "bbox": [ + 75, + 428, + 468, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Different from recent studies focusing on separate attacks on unimodal data [24, 19], multimodal pairs exhibit intrinsic alignment and complementarity to each other. The modeling of inter-modal correspondence turns out to be a critical problem for transferability. Considering that the alignments between image and text are many-to-many, for example, an image could be described to be with various human perspectives and language styles, a reasonable perturbation direction may be determined with diverse guidance from multiple labels in the other modality. However, recent adversarial attack methods for VLP models [41] usually employ a single image-text pair to generate adversarial samples. Although they exhibit strong performance in white-box settings, the poor diversity of guidance makes adversarial samples highly correlated with the alignment pattern of the white-box model, and therefore impedes generalization to black-box settings.", + "bbox": [ + 75, + 611, + 470, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the weak transferability problem, we propose Set-level Guidance Attack (SGA), which leverages diverse cross-modal interactions among multiple image-text pairs", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(Figure 2). Specifically, we introduce alignment-preserving augmentation which enriches image-text pairs while keeping their alignments intact. The image augmentation is based on the scale-invariant property of deep learning models [22], thus we can construct multi-scale images to increase the diversity. For text augmentation, we select the most matching caption pairs from the dataset. More importantly, SGA generates adversarial examples on multimodal augmented input data with carefully designed cross-modal guidance. In detail, SGA iteratively pushes supplemental information away between two modalities with another modality as supervision to disrupt the interactions for better harmonious perturbations. Note that resultant adversarial samples could perceive the gradients originated from multiple guidance.", + "bbox": [ + 496, + 90, + 893, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct experiments on two well-established multimodal datasets, Flickr30K [27] and MSCOCO [23], to evaluate the performance of our proposed SGA across various Vision-and-Language $(\\mathrm{V} + \\mathrm{L})$ downstream tasks. The experimental results demonstrate the high effectiveness of SGA in generating adversarial examples that can be strongly transferred across VLP models, surpassing the current state-of-the-art attack methods in multimodal learning. In particular, SGA achieves notable improvements in image-text retrieval under black-box settings and also exhibits superior performance in white-box attack settings. Moreover, SGA also outperforms the state-of-the-art methods in image captioning and yields higher fooling rates on visual grounding.", + "bbox": [ + 496, + 303, + 893, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contributions as follows. 1) We make the first attempt to explore the transferability of adversarial examples on popular VLP models with a systematical evaluation; 2) We provide SGA, a novel transferable multimodal attack that enhances adversarial transferability through the effective use of set-level alignment-preserving augmentations and well-designed cross-modal guidance; 3) Extensive experiments show that SGA consistently boosts adversarial transferability across different VLP models than the state-of-the-art methods.", + "bbox": [ + 496, + 501, + 893, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 667, + 640, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Vision-Language Pre-training Models", + "text_level": 1, + "bbox": [ + 500, + 693, + 825, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Language Pre-training (VLP) aims to improve the performance of downstream multimodal tasks by pretraining large-scale image-to-text pairs [17]. Most works are developed upon the pre-trained object detectors with region features to learn the vision-language representations [5, 20, 42, 34]. Recently, with the increasing popularity of Vision Transformer (ViT) [8, 31, 40], some other works propose to use ViT as an image encoder and transform the input into patches in an end-to-end manner [18, 38, 17, 9, 33].", + "bbox": [ + 496, + 717, + 893, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "According to the VLP architectures, VLP models can be classified into two typical types: fused VLP models and aligned VLP models [41]. Specifically, fused VLP models", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "103", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(e.g., ALBEF [18], TCL [38]) first utilize separate unimodal encoders to process token embeddings and visual features, and further use a multimodal encoder to process image and text embeddings to output fused multimodal embeddings. Alternatively, aligned VLP models (e.g., CLIP [28]) have only unimodal encoders with independent image and text modality embeddings. In this paper, we focus on popular architectures with fused and aligned VLP models.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Image-Text Retrieval Task", + "text_level": 1, + "bbox": [ + 76, + 223, + 316, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-Text Retrieval (ITR) aims to retrieve the relevant top-ranked instances from a gallery database with one modality, given an input query from another modality [35, 4, 43, 6]. This task can be divided into two subtasks, image-to-text retrieval (TR) and text-to-image retrieval (IR).", + "bbox": [ + 75, + 247, + 470, + 323 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For ALBEF [18] and TCL [38], the semantic similarity score in the unimodal embedding space will be calculated for all image-text pairs to select top- $k$ candidates. Then the multimodal encoder takes the top- $k$ candidates and computes the image-text matching score for ranking. For CLIP [28], without the multimodal encoder, the final rank list can be obtained based on the similarity in the embedding space between image and text modalities.", + "bbox": [ + 75, + 323, + 468, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Adversarial Transferability", + "text_level": 1, + "bbox": [ + 76, + 455, + 323, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing adversarial attacks can be categorized into two settings: white-box attacks and black-box attacks. In a white-box setting, the target model is fully accessible, but not in a black-box setting. In computer vision, many methods employ gradient information for adversarial attacks in white-box settings, such as FGSM [11], PGD [24], C&W [3], and MI [7]. In contrast, in the field of natural language processing (NLP), current attack methods mainly modify or replace some tokens of the input text [19, 29, 10, 13]. In the multimodal vision-language domain, Zhang et al. [41] proposed a white-box multimodal attack method with respect to popular VLP models on downstream tasks.", + "bbox": [ + 75, + 479, + 470, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, white-box attacks are unrealistic due to the inaccessibility of model information in practical applications. In addition, there is no related work that systematically analyzes the adversarial transferability of multimodal attack methods on VLP models. Therefore, in this work, we mainly focus on generating highly transferable adversaries across different VLP models.", + "bbox": [ + 75, + 662, + 470, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Analysis of Adversarial Transferability", + "text_level": 1, + "bbox": [ + 76, + 782, + 426, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we conduct an empirical study on VLP models to evaluate adversarial transferability using existing methods. A common approach to attack multimodal tasks is combining unimodal adversarial attacks [24, 7, 37, 19] of each modalities together. For instance, the separate unimodal attack (Sep-Attack) includes PGD [24] and BERT-Attack", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/31377960b66f9787e47cdfd1586b6c6be547d5ff88b9faa3601d3e36623fef34.jpg", + "image_caption": [ + "Figure 3: Attack success rates $(\\%)$ of different perturbed modalities on image-text retrieval. Adversarial examples with respect to different modalities from the source model to attack the target model using Sep-Attack. Different colors represent different perturbed modalities." + ], + "image_footnote": [], + "bbox": [ + 511, + 90, + 880, + 210 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c71c566bd1ede3d6e42a054d67a7eb090e7241a954e637d505f0a839317aa523.jpg", + "image_caption": [ + "Figure 4: Attack success rates $(\\%)$ on white-box models and black-box models. Both images and text are perturbed by Sep-Attack (Left) and Co-Attack (Right) on image-text retrieval. Colors indicate different target models." + ], + "image_footnote": [], + "bbox": [ + 511, + 311, + 880, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[19] for attacking image modality and text modality, respectively. Other recent multimodal adversarial attacks, such as Co-Attack [41], consider cross-modal interactions by perturbing image modalities and text modalities collectively.", + "bbox": [ + 496, + 522, + 893, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first present the observations regarding the adversarial transferability of VLP models. Then we discuss the limitations of the existing methods. By conducting this study, we aim to provide insights into the robustness of VLP models against adversarial attacks and the effectiveness of different attack strategies.", + "bbox": [ + 496, + 582, + 893, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Observations", + "text_level": 1, + "bbox": [ + 500, + 684, + 637, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To investigate the adversarial transferability of perturbed inputs with respect to different modalities (i.e., image, text, and image & text) and the effect of different VLP models on transferability, we conduct experiments and present the attack success rates of the adversarial examples generated by the source model to attack the target models in Figure 3 and Figure 4. The observations are summarized below:", + "bbox": [ + 496, + 708, + 893, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- The adversarial transferability of attacking both modalities (image & text) is consistently more effective than attacking any unimodal data alone (image or text). As shown in Figure 3, transferring both adversarial image and text from ALBEF to TCL leads to a much higher", + "bbox": [ + 517, + 824, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "104", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "attack success rate than transferring adversarial examples of any single modality. Notably, ALBEF and TCL are both fused VLP models. Similar observations also exist in the following settings: (1) The source and target models are different types of VLP models but have the same basic architectures (e.g., TCL and $\\mathrm{CLIP}_{\\mathrm{ViT}}$ ). (2) The source and target models are the same types of VLP models, but with different basic architectures (e.g., $\\mathrm{CLIP}_{\\mathrm{ViT}}$ and $\\mathrm{CLIP}_{\\mathrm{CNN}}$ ).", + "bbox": [ + 107, + 90, + 472, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Adversarial multimodal data (i.e., adversarial image & text), which have strong attack performance on the source model, can hardly maintain the same capability when transferring to target models. For example, as illustrated in Figure 4, even though ALBEF and TCL have the same model architecture, the attack success rate sees a significant drop when transferring adversarial examples generated on ALBEF to TCL. The phenomenon exists in both Sep-Attack and Co-Attack.", + "bbox": [ + 96, + 239, + 472, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, the attack methods can have stronger transferability in black-box settings if all the modalities are attacked simultaneously. However, even though two modalities are allowed to be perturbed, existing methods still exhibit much lower transferability. This suggests that attacks with higher transferability should be specifically designed, instead of directly utilizing the existing white-box attack methods.", + "bbox": [ + 76, + 388, + 472, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Discussions", + "text_level": 1, + "bbox": [ + 76, + 505, + 202, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We posit that the degradation in transferability of adversarial examples is mainly due to the limitations of the existing attack methods:", + "bbox": [ + 76, + 527, + 470, + 574 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- One major limitation of Sep-Attack is that it does not take into account the interactions between different modalities. As a combined independent attack method for each modality, it fails to model the inter-modal correspondence that is crucial for successful attacks in multimodal learning. This is particularly evident in multimodal tasks such as image-text retrieval, where the ground truth is not discrete labels (e.g., image classification) but another modality data that corresponds to the input modality. The complete lack of cross-modal interactions in Sep-Attack severely limits the generalization of adversarial examples and reduces their transferability among different VLP models.", + "- While Co-Attack is designed to leverage the collaboration between modalities to generate adversarial examples, it still suffers from a key drawback that hinders its transferability to other VLP models. Unlike unimodal learning, multimodal learning involves multiple complementary modalities with many-to-many cross-modal alignments, which pose unique challenges to" + ], + "bbox": [ + 96, + 585, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "achieving sufficient adversarial transferability. However, Co-Attack only uses single image-text pairs to generate adversarial data, limiting the diversity of guidance from multiple labels in other modalities. This lack of diversity in cross-modal guidance makes adversarial samples highly correlated with the alignment pattern of the white-box model. Therefore, the generality of adversarial examples is restricted, and their effectiveness in transferring to other models drops.", + "bbox": [ + 531, + 90, + 893, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In conclusion, the analysis motivates our investigation into the adversarial transferability of VLP models. Moreover, it highlights the pressing need to explore transferable multimodal attacks for generating adversarial examples that can be effectively transferred across different VLP models.", + "bbox": [ + 496, + 237, + 893, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 498, + 325, + 635, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we propose a transferable multimodal adversarial attack, termed Set-level Guidance Attack (SGA). SGA is designed to enhance adversarial transferability across VLP models by leveraging multimodal interactions and incorporating diverse alignment-preserving inter-modal information with carefully designed cross-modal guidance. We provide our motivation first, followed by relevant notations, and finally present SGA in detail.", + "bbox": [ + 496, + 351, + 895, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Motivation", + "text_level": 1, + "bbox": [ + 500, + 482, + 620, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To improve the transferability of multimodal attacks, we first conduct an investigation into the shortcomings of existing methods in black-box settings. Through a systematic analysis of failure cases, we observe that around half of these cases arise due to the presence of multiple matched captions of the image, as shown in Table 1. More specifically, our findings indicate that while the generated adversarial image may be significantly distant from the single supervised caption in the source model, it is prone to approaching other matched captions in the target model, where the alignments can be modeled and ultimately lead to attack failure.", + "bbox": [ + 496, + 505, + 893, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, to maintain the attack ability of adversarial images when transferring to other models, it is crucial to consider multiple paired captions and push the adversarial image away from all the paired captions, thus preserving the attack ability when transferred to other black-box models. Crafting adversarial text for high transferability follows a similar approach, which can also benefit from more paired images. More discussions can be found in Appendix A.", + "bbox": [ + 496, + 671, + 893, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Notations", + "text_level": 1, + "bbox": [ + 500, + 801, + 611, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $(v,t)$ denote an image-text pair sampled from a multimodal dataset $D$ . For VLP models, we denote $f_{I}$ as the image encoder and $f_{T}$ as the text encoder. The multimodal fusion module in fused VLP models is denoted by $f_{M}$ . Specifically, $f_{I}(v)$ represents the image representation $e_v$ encoded", + "bbox": [ + 496, + 825, + 895, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "105", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/8dfbf039b2399d2749c5cc21e35f6729686d410c65b34667f95708c29724208b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AttackCross-modal Interactionp (Event A) ↑p (Event B | Event A) ↓
ALBEFTCLCLIPViTCLIPCNN
Sep-AttackNo74.60%46.78%41.69%41.82%
Co-AttackSingle-pair80.60%50.50%40.82%42.93%
SGASet-level97.20%28.91%34.67%38.58%
", + "bbox": [ + 84, + 88, + 467, + 170 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: Adversarial images with insufficient cross-modal interactions may get closer to other paired captions when transferring. Event A: success cases in white-box settings. Event B: failure cases in black-box settings influenced by other paired captions. Adversarial data are generated on Flickr30K from ALBEF to attack other target models.", + "bbox": [ + 75, + 181, + 470, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "by $f_{I}$ taking an image $v$ as input, $f_{T}(t)$ denotes the text representation $e_t$ encoded by $f_{T}$ taking a text $t$ as input, and $f_{M}(e_v, e_t)$ denotes the multimodal representation encoded by $f_{M}$ taking image and text representations as inputs.", + "bbox": [ + 75, + 286, + 468, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use $B[v, \\epsilon_v]$ and $B[t, \\epsilon_t]$ to represent the legal searching spaces for optimizing adversarial image and text, respectively. Specifically, $\\epsilon_v$ denotes the maximal perturbation bound for the image, and $\\epsilon_t$ denotes the maximal number of changeable words in the caption.", + "bbox": [ + 75, + 347, + 470, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Transferable Set-Level Guidance Attack", + "text_level": 1, + "bbox": [ + 76, + 431, + 421, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Alignment-preserving Augmentation. The analysis presented in Section 3 highlights the key limitation of existing methods: the inter-modal information used to generate adversarial examples lacks diversity. The limitation will make the generated adversarial examples fail to generalize to other black-box models with strong attack performance, resulting in limited adversarial transferability.", + "bbox": [ + 75, + 455, + 468, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To inject more diversity in the generation of generalizable adversarial examples, we propose using set-level alignment-preserving augmentation to expand multimodal input spaces while maintaining cross-modal alignments intact. Unlike previous methods that only consider a single image-text paired example $(v,t)$ to generate adversarial data, we enlarge the input to a set level of images and captions. Specifically, we select the most matching caption pairs from the dataset of each image $v$ to form an augmented caption set $\\pmb{t} = \\{t_1,t_2,\\dots,t_M\\}$ , and resize each image $v$ into different scales $S = \\{s_{1},s_{2},\\dots,s_{N}\\}$ and then add Gaussian noise to obtain a multi-scale image set $\\pmb{v} = \\{v_{1},v_{2},\\dots,v_{N}\\}$ based on the scale-invariant property. The enlarged input set $(\\pmb{v},\\pmb{t})$ is then used to generate the adversarial data $(v^{\\prime},t^{\\prime})$ .", + "bbox": [ + 75, + 561, + 470, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cross-modal Guidance. Cross-modal interactions play a crucial role in multimodal tasks. For example, in imagetext retrieval, the paired information from another modality provides unique annotation supervision for each sample. Similarly, in adversarial attacks, supervisory information is essential in guiding the search for adversarial examples.", + "bbox": [ + 75, + 779, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To fully utilize the enlarged alignment-preserving multimodal input set $(\\pmb{v},\\pmb{t})$ and further improve the transfer", + "bbox": [ + 75, + 869, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ability of the generated adversarial data, we propose cross-modal guidance to utilize interactions from different modalities. Specifically, we use the paired information from another modality as the supervision to guide the direction of optimizing the adversarial data. This guidance iteratively pushed away the multimodal information and disrupt the cross-modal interaction for better harmonious perturbations. Notably, the resultant adversarial examples can perceive the gradients originated from multiple guidance.", + "bbox": [ + 496, + 90, + 893, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we generate corresponding adversarial captions for all captions in the text set $\\pmb{t}$ , forming an adversarial caption set $\\pmb{t}' = \\{t_1', t_2', \\dots, t_M'\\}$ . The process can be formulated as,", + "bbox": [ + 496, + 227, + 893, + 273 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nt _ {i} ^ {\\prime} = \\underset {t _ {i} ^ {\\prime} \\in B [ t _ {i}, \\epsilon_ {t} ]} {\\arg \\max } - \\frac {f _ {T} \\left(t _ {i} ^ {\\prime}\\right) \\cdot f _ {I} (v)}{\\| f _ {T} \\left(t _ {i} ^ {\\prime}\\right) \\| \\| f _ {I} (v) \\|.} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 280, + 892, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The adversarial caption $t_i'$ is constrained to be dissimilar to the original image $v$ in the embedding space. Next, the adversarial image $v'$ is generated by solving", + "bbox": [ + 496, + 324, + 892, + 371 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nv ^ {\\prime} = \\underset {v ^ {\\prime} \\in B [ v, \\epsilon_ {v} ]} {\\arg \\max } - \\sum_ {i = 1} ^ {M} \\frac {f _ {T} \\left(t _ {i} ^ {\\prime}\\right)}{\\| f _ {T} \\left(t _ {i} ^ {\\prime}\\right) \\|} \\sum_ {s _ {i} \\in S} \\frac {f _ {I} \\left(g \\left(v ^ {\\prime} , s _ {i}\\right)\\right)}{\\| f _ {I} \\left(g \\left(v ^ {\\prime} , s _ {i}\\right)\\right) \\|}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 377, + 892, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $g(v', s_i)$ denotes the resizing function that takes the image $v'$ and the scale coefficient $s_i$ as inputs. All the scaled images derived from $v'$ are encouraged to be far away from all the adversarial captions $t_i'$ in the embedding space. Finally, the adversarial caption $t'$ is generated as follows,", + "bbox": [ + 496, + 422, + 893, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nt ^ {\\prime} = \\underset {t ^ {\\prime} \\in B [ t, \\epsilon_ {t} ]} {\\arg \\max } - \\frac {f _ {T} \\left(t ^ {\\prime}\\right) \\cdot f _ {I} \\left(v ^ {\\prime}\\right)}{\\| f _ {T} \\left(t ^ {\\prime}\\right) \\| \\| f _ {I} \\left(v ^ {\\prime}\\right) \\|}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 505, + 892, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in which $t'$ is encouraged to be far away from the adversarial image $v'$ in the embedding space. The detailed algorithm can be found in Appendix C.", + "bbox": [ + 496, + 544, + 892, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 602, + 633, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present experimental evidence for the advantages of our proposed SGA. We conduct experiments on a diverse set of datasets and popular VLP models. First, we describe the experimental settings in Section 5.1. Then, in Section 5.2, we validate the immediate integration of transfer-based unimodal attacks into multimodal learning. Next, we provide the main evaluation results compared to the state-of-the-art method in Section 5.3. In Section 5.4, we analyze the cross-task transferability between different $\\mathrm{V + L}$ tasks. Finally, we present ablation studies in Section 5.5.", + "bbox": [ + 496, + 627, + 893, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 786, + 705, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We consider two widely used multimodal datasets, Flickr30K [27] and MSCOCO [23]. Flickr30K consists of 31,783 images, each with five corresponding captions. Similarly, MSCOCO comprises 123,287 images, and each image is annotated with around five captions. We adopt the Karpathy split [14] for experimental evaluation.", + "bbox": [ + 496, + 809, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "106", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f14f7364892a82e7d2ecbc1051ceea8cbd8a6dfd5e17d8c350f734ae40793086.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AttackALBEF*TCLCLIPViTCLIPCNN
TR R@1*IR R@1*TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1
Sep-Attack65.6973.9517.6032.9531.1745.2332.8245.49
Sep-Attack + MI58.81(↓)65.25(↓)16.02(↓)28.19(↓)23.07(↓)36.98(↓)26.56(↓)39.31(↓)
Sep-Attack + DIM56.41(↓)64.24(↓)16.75(↓)29.55(↓)24.17(↓)37.60(↓)25.54(↓)38.77(↓)
Sep-Attack + PNA.PO40.56(↓)53.95(↓)18.44(↑)30.98(↓)22.33(↓)37.02(↓)26.95(↓)38.63(↓)
Co-Attack77.1683.8615.2129.4923.6036.4825.1238.89
Co-Attack + MI64.86(↓)75.26(↓)25.40(↑)38.69(↑)24.91(↑)37.11(↑)26.31(↑)38.97(↑)
Co-Attack + DIM47.03(↓)62.28(↓)22.23(↑)35.45(↑)25.64(↑)38.50(↑)26.95(↑)40.58(↑)
SGA97.2497.2845.4255.2533.3844.1634.9346.57
", + "bbox": [ + 161, + 88, + 812, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2: Attack success rates (%) of R@1 of integrating transfer-based image attacks on image-text retrieval. The adversarial data are generated on Flickr30K using the source model ALBEF to attack other target models. * indicates white-box attacks. A higher ASR indicates better adversarial transferability.", + "bbox": [ + 75, + 282, + 892, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VLP Models. We evaluate two popular VLP models, the fused VLP and aligned VLP models. For the fused VLP, we consider ALBEF [18] and TCL [38]. ALBEF contains a 12-layer visual transformer ViT-B/16 [8] and two 6-layer transformers for the image encoder and both the text encoder and the multimodal encoder, respectively. TCL uses the same model architecture as ALBEF but with different pretrain objectives. For the aligned VLP model, we choose to evaluate CLIP [28]. CLIP has two different image encoder choices, namely, $\\mathrm{CLIP}_{\\mathrm{ViT}}$ and $\\mathrm{CLIP}_{\\mathrm{CNN}}$ , that use ViT-B/16 and ResNet-101 [12] as the base architectures for the image encoder, respectively.", + "bbox": [ + 75, + 340, + 472, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Adversarial Attack Settings. To craft adversarial images, we employ PGD [24] with perturbation bound $\\epsilon_v = 2/255$ , step size $\\alpha = 0.5/255$ , and iteration steps $T = 10$ . For attacking text modality, we adopt BERT-Attack [19] with perturbation bound $\\epsilon_t = 1$ and length of word list $W = 10$ . Furthermore, we enlarge the image set by resizing the original image into five scales, $\\{0.50, 0.75, 1.00, 1.25, 1.50\\}$ , using bicubic interpolation. Similarly, the caption set is enlarged by augmenting the most matching caption pairs for each image in the dataset, with the size of approximately five.", + "bbox": [ + 75, + 527, + 470, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We employ Attack Success Rate (ASR) as the metric for evaluating the adversarial robustness and transferability in both white-box and black-box settings. Specifically, ASR evaluates the percentage of attacks that only produce successful adversarial examples. A higher ASR indicates better adversarial transferability.", + "bbox": [ + 75, + 686, + 470, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Transferability Analysis", + "text_level": 1, + "bbox": [ + 76, + 786, + 299, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this paper, we present a systematic study of the adversarial transferability of VLP models, which has not been explored. As demonstrated in Section 3, existing methods, including the separate unimodal adversarial attack (SepAttack) and the multimodal adversarial attack (Co-Attack), exhibit limited transferability to other VLP models.", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To improve transferability in multimodal learning, we intuitively investigate the adoption of transfer-based attacks from unimodal learning such as image classification. Specifically, we consider MI [7], DIM [37], and PNA.PO [36]. However, this approach can be problematic if cross-modal interactions and the unique many-to-many alignments in multimodal learning are not taken into account.", + "bbox": [ + 496, + 340, + 893, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 illustrates that multimodal attack methods that incorporate transfer-based image attacks exhibit minimal improvement in transferability while compromising white-box performance. Specifically, when integrated with MI, Co-Attack drops significantly by $12.3\\%$ in white-box settings, while only maintaining $25.40\\%$ ASR in transferability (ALBEF to TCL). However, our SGA shows superior performance in both white-box and black-box settings. Notably, Sep-Attack combined with transfer-based attacks not only reduces the effectiveness of white-box attacks but also fails to improve adversarial transferability in almost all black-box settings. The results provide empirical evidence that directly combining unimodal adversarial attacks in multimodal learning without considering cross-modal interactions and alignments can be problematic, even when using transfer-based unimodal attacks. Additional discussion is provided in Appendix B.2.", + "bbox": [ + 496, + 446, + 895, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 710, + 702, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multimodal Fusion Modules. First, we investigate VLP models with different fusion modules, namely, fused VLP models and aligned VLP models. We generate adversarial examples on both types of models and evaluate their attack performance when transferred to other VLP models while ensuring consistency in the input size of images. For example, adversarial images generated by ALBEF or TCL are resized to $224 \\times 224$ before performing transfer attacks on CLIP, and adversarial examples generated on CLIP are resized to $384 \\times 384$ before being transferred to ALBEF or TCL.", + "bbox": [ + 496, + 734, + 893, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 3, experimental results demonstrate", + "bbox": [ + 519, + 886, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "107", + "bbox": [ + 486, + 944, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e5bc183d7ed555a792d8d3cfa0b84de04a7a750b25960fb27d318d7ef733b30a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SourceAttackALBEFTCLCLIPViTCLIPCNN
TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1
ALBEFPGD52.45*58.65*3.066.798.9613.2110.3414.65
BERT-Attack11.57*27.46*12.6428.0729.3343.1732.6946.11
Sep-Attack65.69*73.95*17.6032.9531.1745.2332.8245.49
Co-Attack77.16*83.86*15.2129.4923.6036.4825.1238.89
SGA97.24±0.22*97.28±0.15*45.42±0.6055.25±0.0633.38±0.3544.16±0.2534.93±0.9946.57±0.13
TCLPGD6.1510.7877.87*79.48*7.4813.7210.3415.33
BERT-Attack11.8926.8214.54*29.17*29.6944.4933.4646.07
Sep-Attack20.1336.4884.72*86.07*31.2944.6533.3345.80
Co-Attack23.1540.0477.94*85.59*27.8541.1930.7444.11
SGA48.91±0.7460.34±0.1098.37±0.08*98.81±0.07*33.87±0.1844.88±0.5437.74±0.2748.30±0.34
CLIPViTPGD2.504.934.858.1770.92*78.61*5.368.44
BERT-Attack9.5922.6411.8025.0728.34*39.08*30.4037.43
Sep-Attack9.5923.2511.3825.6079.75*86.79*30.7839.76
Co-Attack10.5724.3311.9426.6993.25*95.86*32.5241.82
SGA13.40±0.0727.22 ±0.0616.23±0.4530.76±0.0799.08±0.08*98.94±0.00*38.76±0.2747.79±0.58
CLIPCNNPGD2.094.824.007.811.106.6086.46*92.25*
BERT-Attack8.8623.2712.3325.4827.1237.4430.40*40.10*
Sep-Attack8.5523.4112.6426.1228.3439.4391.44*95.44*
Co-Attack8.7923.7413.1026.0728.7940.0394.76*96.89*
SGA11.42±0.0724.80±0.2814.91±0.0828.82±0.1131.24±0.4242.12±0.1199.24±0.18*99.49±0.05*
", + "bbox": [ + 83, + 88, + 888, + 417 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f5772f1a123944da880051baad37e8c72daf4c9fe40d7537e71756152b8df8be.jpg", + "table_caption": [ + "Table 3: Comparison with state-of-the-art methods on image-text retrieval. We report the attack success rate $(\\%)$ of R@1 for both IR and TR. The source column indicates the source models used to generate the adversarial data on Flickr30K. * indicates white-box attacks. A higher ASR indicates better adversarial transferability." + ], + "table_footnote": [], + "table_body": "
AttackB@4METEORROUGE_LCIDErSPICE
Baseline39.731.060.0133.323.8
Co-Attack37.429.858.4125.522.8
SGA34.828.456.3116.021.4
", + "bbox": [ + 109, + 494, + 439, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4: Cross-Task Transferability: ITR $\\rightarrow$ IC. Adversarial data generated from Image-Text Retrieval (ITR) to attack Image Captioning (IC) on MSCOCO. The source model and target model are ALBEF and BLIP, respectively. The Baseline represents the original performance of IC on clean data. Lower values indicate better adversarial transferability.", + "bbox": [ + 75, + 571, + 470, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the superiority of our proposed SGA over existing multimodal attack methods in all black-box settings. Specifically, our SGA achieves significant improvements in adversarial transferability when the source and target models are of the same type. For instance, SGA outperforms Co-Attack by approximately $30\\%$ in terms of attack success rate when transferring adversarial data from ALBEF to TCL. Moreover, in the more challenging scenario where the source and target models are of different types, SGA also surpasses CoAttack with higher attack success rates. More experiments on the MSCOCO dataset are provided in Appendix B.3.", + "bbox": [ + 75, + 676, + 470, + 843 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model Architectures. Then, we explore VLP models with respect to different model architectures. Many VLP models commonly use ViTs as the vision encoder, where images", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are segmented into patches before being processed by the transformer model. However, in the case of CLIP, the image encoder consists of two distinct architectures: conventional Convolutional Neural Networks (CNNs) and ViTs. The transferability between CNNs and ViTs has been well-studied in unimodal learning. Therefore, we also investigate the adversarial transferability of CNN-based and ViT-based CLIP models in multimodal learning.", + "bbox": [ + 496, + 496, + 893, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 3, we observe a similar phenomenon as unimodal learning that compared to CNNs, ViTs show better robustness against adversarial perturbations [25]. Specifically, for all attack methods, the same adversarial multimodal data have a stronger white-box attack effect on $\\mathrm{CLIP}_{\\mathrm{CNN}}$ compared to $\\mathrm{CLIP}_{\\mathrm{ViT}}$ . Moreover, the adversarial examples generated on $\\mathrm{CLIP}_{\\mathrm{ViT}}$ are found to be more transferable to $\\mathrm{CLIP}_{\\mathrm{CNN}}$ than vice versa (38.76% vs. 31.24%).", + "bbox": [ + 496, + 619, + 893, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, our proposed SGA consistently improves transferability on both CNN-based CLIP and ViT-based CLIP compared to other attacks. For instance, SGA increases the adversarial transferability for $\\mathrm{CLIP}_{\\mathrm{ViT}}$ by $5.83\\%$ and $6.24\\%$ compared to Co-Attack under the white-box setting and black-box setting, respectively.", + "bbox": [ + 496, + 741, + 893, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Cross-Task Transferability", + "text_level": 1, + "bbox": [ + 500, + 845, + 741, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-modal interactions and alignments are the core components of multimodal learning regardless of the task.", + "bbox": [ + 500, + 869, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "108", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2d4b85bac06fd3c74c6c318d32d997b8dfd8dc00ca6ac89fe1d132f3d87f680f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AttackValTestATestB
Baseline58.4665.8946.25
Co-Attack54.2661.8043.81
SGA53.5561.1943.71
", + "bbox": [ + 135, + 88, + 415, + 156 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5: Cross-Task Transferability: ITR $\\rightarrow$ VG. Adversarial data generated from Image-Text Retrieval (ITR) to attack Visual Grounding (VG) on RefCOCO+. The source model and target model are both ALBEF. The Baseline represents the original performance of VG on clean data. Lower values indicate better adversarial transferability.", + "bbox": [ + 75, + 167, + 470, + 260 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Therefore, we conduct extensive experiments to explore the effectiveness of our proposed SGA on two additional V+L tasks: Image Captioning (IC) and Visual Grounding (VG).", + "bbox": [ + 75, + 275, + 470, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Image Captioning. Image captioning is a generation-based task, where an input image is encoded into a feature vector and then decoded into a natural language sentence. In our experiments, we craft adversarial images using the source model (ALBEF) with an image-text retrieval objective and then directly attack the target model (BLIP [17]) on image captioning. We employ the MSCOCO dataset, which is suitable for both two tasks, and utilize various evaluation metrics to measure the quality of the generated captions, including BLEU [26], METEOR [2], ROUGE [21], CIDEr [32], and SPICE [1].", + "bbox": [ + 75, + 330, + 468, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present the performance on image captioning of BLIP after being attacked in Table 4. Experimental results demonstrate clear improvements in the adversarial transferability of the proposed SGA compared to Co-Attack. Specifically, our SGA improves the BLEU score by up to $2.6\\%$ and the CIDEr score by up to $9.5\\%$ .", + "bbox": [ + 75, + 497, + 470, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual Grounding. Visual grounding is another $\\mathrm{V} + \\mathrm{L}$ task that aims to localize the region in an image based on the corresponding specific textual description. Similarly, we generate adversarial images using the source model (ALBEF) from image-text to attack the target model (ALBEF) on visual grounding. Table 5 shows the results on RefCOCO+ [39], where our SGA still outperforms Co-Attack.", + "bbox": [ + 75, + 598, + 470, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 715, + 230, + 732 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To systematically investigate the impact of our set-level alignment-preserving augmentations, we conducted ablation experiments on image-text retrieval to evaluate the effect of varying the number of augmented image sets $N$ with multiscale transformation and the number $M$ of augmented caption sets. Specifically, we employed ALBEF as the source model and $\\mathrm{CLIP}_{\\mathrm{ViT}}$ as the target model on Flickr30K. More details are provided in Appendix B.4.", + "bbox": [ + 75, + 739, + 470, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-scale Image Set. We propose the use of multiple scale-invariant images to generate diverse adversarial data in", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e089b05f0c6b102500fe1ce548125d1e3e81f5360af1128d6b7c5e4ee9bc5ff3.jpg", + "image_caption": [ + "Figure 5: Ablation Study: Attack success rates (\\%) on set-level augmentations, Left for image set and Right for caption set. The performance of SGA is consistently improved by using larger augmented image and caption sets." + ], + "image_footnote": [], + "bbox": [ + 498, + 88, + 694, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3d389908dc4be6587473a28e3085c3c66d1c80bffd7ef7c676abf2bc33238d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 88, + 890, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SGA. Results in the left of Figure 5 reveal that the transferability significantly increases as we introduce more diverse images with different scales, peaking when the scale range is set to [0.50, 1.50] with a step of 0.25. We set the scale range $S = \\{0.50, 0.75, 1.00, 1.25, 1.50\\}$ for optimal performance.", + "bbox": [ + 496, + 303, + 893, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-pair Caption Set. We also conduct experiments to investigate the impact of an enlarged caption set on adversarial transferability. The number of additional captions ranged from 1 to $M$ , where $M$ represents the most matching caption pairs from the dataset for each image. Results presented in the right panel of Figure 5 indicate that if $M > 1$ , the black-box performance increases significantly but eventually plateaus. These results demonstrate the effectiveness of using multiple alignment-preserving inter-modal information to enhance adversarial transferability. Furthermore, we observed that the performance is relatively insensitive to the number of extra captions, but adding more captions can improve the overall adversarial transferability.", + "bbox": [ + 496, + 385, + 895, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 595, + 617, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we make the first attempt to investigate the adversarial transferability of typical VLP models. We systematically evaluate the existing attack methods and reveal that they still exhibit lower transferability, despite their impressive performance in white-box settings. Our investigation highlights the need for specially designed transferable attacks in multimodal learning that can model the many-to-many cross-modal alignments and interactions. We propose SGA, a highly transferable multimodal attack, which leverages set-level alignment-preserving augmentations through cross-modal guidance to thoroughly exploit multimodal interactions. We hope that this work could inspire further research to evaluate and enhance the adversarial robustness of VLP models.", + "bbox": [ + 496, + 622, + 893, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported by the National Key R&D Program of China (Grant NO. 2022YFF1202903) and the National Natural Science Foundation of China (Grant NO. 62122035).", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "109", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In ECCV, 2016. 8", + "[2] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In IEEvaluation@ACL, 2005. 8", + "[3] Nicholas Carlini and David A. Wagner. Towards evaluating the robustness of neural networks. 2017 IEEE Symposium on Security and Privacy (SP), pages 39-57, 2017. 3", + "[4] Hui Chen, Guiguang Ding, Xudong Liu, Zijia Lin, Ji Liu, and Jungong Han. Imram: Iterative matching with recurrent attention memory for cross-modal image-text retrieval. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12652-12660, 2020. 3", + "[5] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 2", + "[6] Mengjun Cheng, Yipeng Sun, Long Wang, Xiongwei Zhu, Kun Yao, Jie Chen, Guoli Song, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Vista: Vision and scene text aggregation for cross-modal retrieval. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5174-5183, 2022. 3", + "[7] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. 3, 6", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv, abs/2010.11929, 2021. 2, 6", + "[9] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuo-hang Wang, Lijuan Wang, Chenguang Zhu, Nanyun Peng, Zicheng Liu, and Michael Zeng. An empirical study of training end-to-end vision-and-language transformers. ArXiv, abs/2111.02387, 2021. 2", + "[10] Ji Gao, Jack Lanchantin, Mary Lou Soffa, and Yanjun Qi. Black-box generation of adversarial text sequences to evade deep learning classifiers. 2018 IEEE Security and Privacy Workshops (SPW), pages 50-56, 2018. 3", + "[11] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. CoRR, abs/1412.6572, 2015. 3", + "[12] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6", + "[13] Di Jin, Zhijing Jin, Joey Tianyi Zhou, and Peter Szolovits. Is bert really robust? a strong baseline for natural language attack on text classification and entailment. In AAAI, 2020. 3", + "[14] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. IEEE Transactions" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Pattern Analysis and Machine Intelligence, 39:664-676, 2017.5", + "[15] Zaid Khan and Yun Raymond Fu. Exploiting bert for multimodal target sentiment classification through input space translation. Proceedings of the 29th ACM International Conference on Multimedia, 2021. 1", + "[16] Chenyi Lei, Shixian Luo, Yong Liu, Wanggui He, Jiamang Wang, Guoxin Wang, Haihong Tang, Chunyan Miao, and Houqiang Li. Understanding chinese video and language via contrastive multimodal pre-training. Proceedings of the 29th ACM International Conference on Multimedia, 2021. 1", + "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. ArXiv, abs/2201.12086, 2022. 2, 8", + "[18] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq R. Joty, Caiming Xiong, and Steven C. H. Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 2, 3, 6", + "[19] Linyang Li, Ruotian Ma, Qipeng Guo, X. Xue, and Xipeng Qiu. Bert-attack: Adversarial attack against bert using bert. ArXiv, abs/2004.09984, 2020. 1, 2, 3, 6", + "[20] Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV, 2020. 2", + "[21] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Annual Meeting of the Association for Computational Linguistics, 2004. 8", + "[22] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. arXiv: Learning, 2019. 2", + "[23] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 5", + "[24] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. ArXiv, abs/1706.06083, 2018. 1, 2, 3, 6", + "[25] Muzammal Naseer, Kanchana Ranasinghe, Salman Hameed Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In Neural Information Processing Systems, 2021. 7", + "[26] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 8", + "[27] Bryan A. Plummer, Liwei Wang, Christopher M. Cervantes, Juan C. Caicedo, J. Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. International Journal of Computer Vision, 123:74–93, 2015. 2, 5", + "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen" + ], + "bbox": [ + 503, + 93, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "110", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 6", + "[29] Shuhuai Ren, Yihe Deng, Kun He, and Wanxiang Che. Generating natural language adversarial examples through probability weighted word saliency. In ACL, 2019. 3", + "[30] Lei Shi, Kai Shuang, Shijie Geng, Peng Gao, Zuohui Fu, Gerard de Melo, Yunpeng Chen, and Sen Su. Dense contrastive visual-linguistic pretraining. Proceedings of the 29th ACM International Conference on Multimedia, 2021. 1", + "[31] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 2", + "[32] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4566-4575, 2015. 8", + "[33] Teng Wang, Yixiao Ge, Feng Zheng, Ran Cheng, Ying Shan, Xiaohu Qie, and Ping Luo. Accelerating vision-language pretraining with free language modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23161-23170, 2023. 2", + "[34] Teng Wang, Wenhao Jiang, Zhichao Lu, Feng Zheng, Ran Cheng, Chengguo Yin, and Ping Luo. Vlmixer: Unpaired vision-language pre-training via cross-modal cutmix. In International Conference on Machine Learning, pages 22680-22690. PMLR, 2022. 2", + "[35] Zihao Wang, Xihui Liu, Hongsheng Li, Lu Sheng, Junjie Yan, Xiaogang Wang, and Jing Shao. Camp: Cross-modal adaptive message passing for text-image retrieval. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5763-5772, 2019. 3", + "[36] Zhipeng Wei, Jingjing Chen, Micah Goldblum, Zuxuan Wu, Tom Goldstein, and Yu-Gang Jiang. Towards transferable adversarial attacks on vision transformers. ArXiv, abs/2109.04176, 2021. 6", + "[37] Cihang Xie, Zhishuai Zhang, Jianyu Wang, Yuyin Zhou, Zhou Ren, and Alan Loddon Yuille. Improving transferability of adversarial examples with input diversity. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2725-2734, 2018. 3, 6", + "[38] Jinyu Yang, Jiali Duan, S. Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul M. Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. ArXiv, abs/2202.10401, 2022. 2, 3, 6", + "[39] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C. Berg, and Tamara L. Berg. Modeling context in referring expressions. ArXiv, abs/1608.00272, 2016. 8", + "[40] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Francis E. H. Tay, Jiashi Feng, and Shuicheng Yan. Tokensto-token vit: Training vision transformers from scratch onImagenet. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 538-547, 2021. 2", + "[41] Jiaming Zhang, Qiaomin Yi, and Jitao Sang. Towards adversarial attack on vision-language pre-training models. Proceedings of the 30th ACM International Conference on Multimedia, 2022. 1, 2, 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5575-5584, 2021. 2", + "[43] Qi Zhang, Zhen Lei, Zhaoxiang Zhang, and S. Li. Context-aware attention network for image-text retrieval. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3533-3542, 2020. 3" + ], + "bbox": [ + 501, + 90, + 893, + 219 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "111", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Set-level Guidance Attack_ Boosting Adversarial Transferability of Vision-Language Pre-training Models/4d2215bf-24a1-45b3-bcea-24623e459934_model.json b/2023/Set-level Guidance Attack_ Boosting Adversarial Transferability of Vision-Language Pre-training Models/4d2215bf-24a1-45b3-bcea-24623e459934_model.json new file mode 100644 index 0000000000000000000000000000000000000000..98d06c6bb8523908917081464dbec51b77b7f7a7 --- /dev/null +++ b/2023/Set-level Guidance Attack_ Boosting Adversarial Transferability of Vision-Language Pre-training Models/4d2215bf-24a1-45b3-bcea-24623e459934_model.json @@ -0,0 +1,2200 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.131, + 0.825, + 0.177 + ], + "angle": 0, + "content": "Set-level Guidance Attack: Boosting Adversarial Transferability of Vision-Language Pre-training Models" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.203, + 0.843, + 0.223 + ], + "angle": 0, + "content": "Dong Lu1, Zhiqiang Wang1, Teng Wang1,2, Weili Guan3, Hongchang Gao4, Feng Zheng1,5†" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.224, + 0.725, + 0.239 + ], + "angle": 0, + "content": "\\(^{1}\\)Southern University of Science and Technology \\(^{2}\\)The University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.288, + 0.241, + 0.682, + 0.257 + ], + "angle": 0, + "content": "3Monash University 4Temple University 5Peng Cheng Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.26, + 0.782, + 0.275 + ], + "angle": 0, + "content": "sammylu\\_outlook.com wangzq_2021\\@outlook.com tengwang@connect.hku.hk" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.278, + 0.756, + 0.293 + ], + "angle": 0, + "content": "honeyguan@gmail.com hongchang.gao@temple.edu f.zheng@ieee.org" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.474, + 0.752 + ], + "angle": 0, + "content": "Vision-language pre-training (VLP) models have shown vulnerability to adversarial examples in multimodal tasks. Furthermore, malicious adversaries can be deliberately transferred to attack other black-box models. However, existing work has mainly focused on investigating white-box attacks. In this paper, we present the first study to investigate the adversarial transferability of recent VLP models. We observe that existing methods exhibit much lower transferability, compared to the strong attack performance in white-box settings. The transferability degradation is partly caused by the under-utilization of cross-modal interactions. Particularly, unlike unimodal learning, VLP models rely heavily on cross-modal interactions and the multimodal alignments are many-to-many, e.g., an image can be described in various natural languages. To this end, we propose a highly transferable Set-level Guidance Attack (SGA) that thoroughly leverages modality interactions and incorporates alignment-preserving augmentation with cross-modal guidance. Experimental results demonstrate that SGA could generate adversarial examples that can strongly transfer across different VLP models on multiple downstream vision-language tasks. On image-text retrieval, SGA significantly enhances the attack success rate for transfer attacks from ALBEF to TCL by a large margin (at least \\(9.78\\%\\) and up to \\(30.21\\%\\)), compared to the state-of-the-art. Our code is available at https://github.com/Zoky-2020/SGA." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.779, + 0.21, + 0.794 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.804, + 0.471, + 0.88 + ], + "angle": 0, + "content": "Recent work has shown that vision-language pre-training (VLP) models are still vulnerable to adversarial examples [41], even though they have achieved remarkable performance on a wide range of multimodal tasks [30, 15, 16]. Existing work mainly focuses on white-box attacks, where" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.328, + 0.697, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.435, + 0.671, + 0.447 + ], + "angle": 0, + "content": "(a) ALBEF to ALBEF" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.327, + 0.884, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.759, + 0.435, + 0.852, + 0.447 + ], + "angle": 0, + "content": "(b) ALBEF to TCL" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.45, + 0.696, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.508, + 0.672, + 0.521 + ], + "angle": 0, + "content": "(c) ALBEF to \\(\\mathrm{CLIP}_{\\mathrm{VIT}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.451, + 0.886, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.752, + 0.508, + 0.86, + 0.52 + ], + "angle": 0, + "content": "(d) ALBEF to \\(\\mathrm{CLIP}_{\\mathrm{CNN}}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.531, + 0.896, + 0.759 + ], + "angle": 0, + "content": "Figure 1: Comparison of attack success rates (ASR) using five different attacks on image-text retrieval. Adversarial examples are crafted on the source model (ALBEF) to attack the target white-box model or black-box models. The first three columns refer to the image-only PGD attack [24], text-only BERT-Attack [19] (BA), and the combined separate unimodal attack (SA), which all belong to the methods without cross-modal interactions. The fourth column is the state-of-the-art multimodal Co-Attack [41] (CA) that employs single-pair cross-modal interactions. The last column is the proposed Set-level Guidance Attack (SGA), which leverages multiple set-level cross-modal interactions, successfully attacking the white-box model and transferring to attack all black-box models with the highest ASR. More discussions are in Section 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.778, + 0.895, + 0.884 + ], + "angle": 0, + "content": "information about the victim model is accessible. However, the transferability of adversarial examples across VLP models has not been investigated, which is a more practical setting. It is still unknown whether the adversarial data generated on the source model can successfully attack another model, which poses a serious security risk to the deployment of VLP models in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.895, + 0.902 + ], + "angle": 0, + "content": "This paper makes the first step to investigate the trans" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.341, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "102" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.133, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.19, + 0.179, + 0.203 + ], + "angle": 0, + "content": "(a) Sep-Attack" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.091, + 0.183, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.19, + 0.091, + 0.3, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.204, + 0.19, + 0.287, + 0.202 + ], + "angle": 0, + "content": "(b) Co-Attack" + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.091, + 0.465, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.19, + 0.411, + 0.203 + ], + "angle": 0, + "content": "(c) SGA" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.213, + 0.473, + 0.41 + ], + "angle": 0, + "content": "Figure 2: Comparison of cross-modal interactions. To generate adversarial examples, existing methods either: (a) separately perturb unimodal data without any cross-modal interactions (Sep-Attack) or, (b) perturb multiple modalities but with single image-text pairs to model cross-modal interactions (Co-Attack). However, our method is capable of learning cross-modal interactions among multiple alignments through (c) set-level guidance (SGA). Specifically, \\( v \\) denotes the input image, and \\( t \\) is the paired caption. \\( v' \\) and \\( t' \\) represent the corresponding adversarial examples, respectively. \\( U(v) \\) and \\( U(t) \\) represent the scale-invariant image set and most matching caption set. Arrows indicate the guidance for generating adversarial examples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.61 + ], + "angle": 0, + "content": "ferability of adversarial samples within VLP models. Without loss of generality, most of our experiments are based on image-text retrieval tasks. We first empirically evaluate this attack performance with respect to different modalities on multimodal tasks across multiple datasets. Our results show that the adversarial transferability of attacking both modalities (image & text) consistently beats attacking unimodal data (image or text). Unfortunately, even though two modalities are allowed to be perturbed simultaneously, the attack success rates of existing methods [24, 19, 41] still significantly drops when transferring from the white-box to black-box settings, as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.471, + 0.853 + ], + "angle": 0, + "content": "Different from recent studies focusing on separate attacks on unimodal data [24, 19], multimodal pairs exhibit intrinsic alignment and complementarity to each other. The modeling of inter-modal correspondence turns out to be a critical problem for transferability. Considering that the alignments between image and text are many-to-many, for example, an image could be described to be with various human perspectives and language styles, a reasonable perturbation direction may be determined with diverse guidance from multiple labels in the other modality. However, recent adversarial attack methods for VLP models [41] usually employ a single image-text pair to generate adversarial samples. Although they exhibit strong performance in white-box settings, the poor diversity of guidance makes adversarial samples highly correlated with the alignment pattern of the white-box model, and therefore impedes generalization to black-box settings." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To address the weak transferability problem, we propose Set-level Guidance Attack (SGA), which leverages diverse cross-modal interactions among multiple image-text pairs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.304 + ], + "angle": 0, + "content": "(Figure 2). Specifically, we introduce alignment-preserving augmentation which enriches image-text pairs while keeping their alignments intact. The image augmentation is based on the scale-invariant property of deep learning models [22], thus we can construct multi-scale images to increase the diversity. For text augmentation, we select the most matching caption pairs from the dataset. More importantly, SGA generates adversarial examples on multimodal augmented input data with carefully designed cross-modal guidance. In detail, SGA iteratively pushes supplemental information away between two modalities with another modality as supervision to disrupt the interactions for better harmonious perturbations. Note that resultant adversarial samples could perceive the gradients originated from multiple guidance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.895, + 0.501 + ], + "angle": 0, + "content": "We conduct experiments on two well-established multimodal datasets, Flickr30K [27] and MSCOCO [23], to evaluate the performance of our proposed SGA across various Vision-and-Language \\((\\mathrm{V} + \\mathrm{L})\\) downstream tasks. The experimental results demonstrate the high effectiveness of SGA in generating adversarial examples that can be strongly transferred across VLP models, surpassing the current state-of-the-art attack methods in multimodal learning. In particular, SGA achieves notable improvements in image-text retrieval under black-box settings and also exhibits superior performance in white-box attack settings. Moreover, SGA also outperforms the state-of-the-art methods in image captioning and yields higher fooling rates on visual grounding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.895, + 0.651 + ], + "angle": 0, + "content": "We summarize our contributions as follows. 1) We make the first attempt to explore the transferability of adversarial examples on popular VLP models with a systematical evaluation; 2) We provide SGA, a novel transferable multimodal attack that enhances adversarial transferability through the effective use of set-level alignment-preserving augmentations and well-designed cross-modal guidance; 3) Extensive experiments show that SGA consistently boosts adversarial transferability across different VLP models than the state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.669, + 0.642, + 0.685 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.694, + 0.826, + 0.71 + ], + "angle": 0, + "content": "2.1. Vision-Language Pre-training Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.894, + 0.854 + ], + "angle": 0, + "content": "Vision-Language Pre-training (VLP) aims to improve the performance of downstream multimodal tasks by pretraining large-scale image-to-text pairs [17]. Most works are developed upon the pre-trained object detectors with region features to learn the vision-language representations [5, 20, 42, 34]. Recently, with the increasing popularity of Vision Transformer (ViT) [8, 31, 40], some other works propose to use ViT as an image encoder and transform the input into patches in an end-to-end manner [18, 38, 17, 9, 33]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "According to the VLP architectures, VLP models can be classified into two typical types: fused VLP models and aligned VLP models [41]. Specifically, fused VLP models" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "103" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.213 + ], + "angle": 0, + "content": "(e.g., ALBEF [18], TCL [38]) first utilize separate unimodal encoders to process token embeddings and visual features, and further use a multimodal encoder to process image and text embeddings to output fused multimodal embeddings. Alternatively, aligned VLP models (e.g., CLIP [28]) have only unimodal encoders with independent image and text modality embeddings. In this paper, we focus on popular architectures with fused and aligned VLP models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.224, + 0.317, + 0.24 + ], + "angle": 0, + "content": "2.2. Image-Text Retrieval Task" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.248, + 0.471, + 0.324 + ], + "angle": 0, + "content": "Image-Text Retrieval (ITR) aims to retrieve the relevant top-ranked instances from a gallery database with one modality, given an input query from another modality [35, 4, 43, 6]. This task can be divided into two subtasks, image-to-text retrieval (TR) and text-to-image retrieval (IR)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.47, + 0.446 + ], + "angle": 0, + "content": "For ALBEF [18] and TCL [38], the semantic similarity score in the unimodal embedding space will be calculated for all image-text pairs to select top-\\(k\\) candidates. Then the multimodal encoder takes the top-\\(k\\) candidates and computes the image-text matching score for ranking. For CLIP [28], without the multimodal encoder, the final rank list can be obtained based on the similarity in the embedding space between image and text modalities." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.457, + 0.325, + 0.473 + ], + "angle": 0, + "content": "2.3. Adversarial Transferability" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.471, + 0.661 + ], + "angle": 0, + "content": "Existing adversarial attacks can be categorized into two settings: white-box attacks and black-box attacks. In a white-box setting, the target model is fully accessible, but not in a black-box setting. In computer vision, many methods employ gradient information for adversarial attacks in white-box settings, such as FGSM [11], PGD [24], C&W [3], and MI [7]. In contrast, in the field of natural language processing (NLP), current attack methods mainly modify or replace some tokens of the input text [19, 29, 10, 13]. In the multimodal vision-language domain, Zhang et al. [41] proposed a white-box multimodal attack method with respect to popular VLP models on downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.663, + 0.471, + 0.768 + ], + "angle": 0, + "content": "However, white-box attacks are unrealistic due to the inaccessibility of model information in practical applications. In addition, there is no related work that systematically analyzes the adversarial transferability of multimodal attack methods on VLP models. Therefore, in this work, we mainly focus on generating highly transferable adversaries across different VLP models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.784, + 0.428, + 0.803 + ], + "angle": 0, + "content": "3. Analysis of Adversarial Transferability" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In this section, we conduct an empirical study on VLP models to evaluate adversarial transferability using existing methods. A common approach to attack multimodal tasks is combining unimodal adversarial attacks [24, 7, 37, 19] of each modalities together. For instance, the separate unimodal attack (Sep-Attack) includes PGD [24] and BERT-Attack" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.091, + 0.882, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.221, + 0.894, + 0.298 + ], + "angle": 0, + "content": "Figure 3: Attack success rates \\((\\%)\\) of different perturbed modalities on image-text retrieval. Adversarial examples with respect to different modalities from the source model to attack the target model using Sep-Attack. Different colors represent different perturbed modalities." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.313, + 0.882, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.447, + 0.893, + 0.507 + ], + "angle": 0, + "content": "Figure 4: Attack success rates \\((\\%)\\) on white-box models and black-box models. Both images and text are perturbed by Sep-Attack (Left) and Co-Attack (Right) on image-text retrieval. Colors indicate different target models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.894, + 0.583 + ], + "angle": 0, + "content": "[19] for attacking image modality and text modality, respectively. Other recent multimodal adversarial attacks, such as Co-Attack [41], consider cross-modal interactions by perturbing image modalities and text modalities collectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.894, + 0.675 + ], + "angle": 0, + "content": "We first present the observations regarding the adversarial transferability of VLP models. Then we discuss the limitations of the existing methods. By conducting this study, we aim to provide insights into the robustness of VLP models against adversarial attacks and the effectiveness of different attack strategies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.685, + 0.638, + 0.7 + ], + "angle": 0, + "content": "3.1. Observations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.709, + 0.894, + 0.814 + ], + "angle": 0, + "content": "To investigate the adversarial transferability of perturbed inputs with respect to different modalities (i.e., image, text, and image & text) and the effect of different VLP models on transferability, we conduct experiments and present the attack success rates of the adversarial examples generated by the source model to attack the target models in Figure 3 and Figure 4. The observations are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "- The adversarial transferability of attacking both modalities (image & text) is consistently more effective than attacking any unimodal data alone (image or text). As shown in Figure 3, transferring both adversarial image and text from ALBEF to TCL leads to a much higher" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "104" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.109, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "attack success rate than transferring adversarial examples of any single modality. Notably, ALBEF and TCL are both fused VLP models. Similar observations also exist in the following settings: (1) The source and target models are different types of VLP models but have the same basic architectures (e.g., TCL and \\(\\mathrm{CLIP}_{\\mathrm{ViT}}\\)). (2) The source and target models are the same types of VLP models, but with different basic architectures (e.g., \\(\\mathrm{CLIP}_{\\mathrm{ViT}}\\) and \\(\\mathrm{CLIP}_{\\mathrm{CNN}}\\))." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.24, + 0.473, + 0.378 + ], + "angle": 0, + "content": "- Adversarial multimodal data (i.e., adversarial image & text), which have strong attack performance on the source model, can hardly maintain the same capability when transferring to target models. For example, as illustrated in Figure 4, even though ALBEF and TCL have the same model architecture, the attack success rate sees a significant drop when transferring adversarial examples generated on ALBEF to TCL. The phenomenon exists in both Sep-Attack and Co-Attack." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.389, + 0.473, + 0.496 + ], + "angle": 0, + "content": "In summary, the attack methods can have stronger transferability in black-box settings if all the modalities are attacked simultaneously. However, even though two modalities are allowed to be perturbed, existing methods still exhibit much lower transferability. This suggests that attacks with higher transferability should be specifically designed, instead of directly utilizing the existing white-box attack methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.506, + 0.203, + 0.52 + ], + "angle": 0, + "content": "3.2. Discussions" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.529, + 0.472, + 0.575 + ], + "angle": 0, + "content": "We posit that the degradation in transferability of adversarial examples is mainly due to the limitations of the existing attack methods:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.587, + 0.473, + 0.784 + ], + "angle": 0, + "content": "- One major limitation of Sep-Attack is that it does not take into account the interactions between different modalities. As a combined independent attack method for each modality, it fails to model the inter-modal correspondence that is crucial for successful attacks in multimodal learning. This is particularly evident in multimodal tasks such as image-text retrieval, where the ground truth is not discrete labels (e.g., image classification) but another modality data that corresponds to the input modality. The complete lack of cross-modal interactions in Sep-Attack severely limits the generalization of adversarial examples and reduces their transferability among different VLP models." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.796, + 0.473, + 0.902 + ], + "angle": 0, + "content": "- While Co-Attack is designed to leverage the collaboration between modalities to generate adversarial examples, it still suffers from a key drawback that hinders its transferability to other VLP models. Unlike unimodal learning, multimodal learning involves multiple complementary modalities with many-to-many cross-modal alignments, which pose unique challenges to" + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.587, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "achieving sufficient adversarial transferability. However, Co-Attack only uses single image-text pairs to generate adversarial data, limiting the diversity of guidance from multiple labels in other modalities. This lack of diversity in cross-modal guidance makes adversarial samples highly correlated with the alignment pattern of the white-box model. Therefore, the generality of adversarial examples is restricted, and their effectiveness in transferring to other models drops." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.238, + 0.895, + 0.314 + ], + "angle": 0, + "content": "In conclusion, the analysis motivates our investigation into the adversarial transferability of VLP models. Moreover, it highlights the pressing need to explore transferable multimodal attacks for generating adversarial examples that can be effectively transferred across different VLP models." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.327, + 0.636, + 0.344 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.897, + 0.473 + ], + "angle": 0, + "content": "In this section, we propose a transferable multimodal adversarial attack, termed Set-level Guidance Attack (SGA). SGA is designed to enhance adversarial transferability across VLP models by leveraging multimodal interactions and incorporating diverse alignment-preserving inter-modal information with carefully designed cross-modal guidance. We provide our motivation first, followed by relevant notations, and finally present SGA in detail." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.483, + 0.622, + 0.497 + ], + "angle": 0, + "content": "4.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.895, + 0.672 + ], + "angle": 0, + "content": "To improve the transferability of multimodal attacks, we first conduct an investigation into the shortcomings of existing methods in black-box settings. Through a systematic analysis of failure cases, we observe that around half of these cases arise due to the presence of multiple matched captions of the image, as shown in Table 1. More specifically, our findings indicate that while the generated adversarial image may be significantly distant from the single supervised caption in the source model, it is prone to approaching other matched captions in the target model, where the alignments can be modeled and ultimately lead to attack failure." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.895, + 0.794 + ], + "angle": 0, + "content": "Therefore, to maintain the attack ability of adversarial images when transferring to other models, it is crucial to consider multiple paired captions and push the adversarial image away from all the paired captions, thus preserving the attack ability when transferred to other black-box models. Crafting adversarial text for high transferability follows a similar approach, which can also benefit from more paired images. More discussions can be found in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.612, + 0.817 + ], + "angle": 0, + "content": "4.2. Notations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Let \\((v,t)\\) denote an image-text pair sampled from a multimodal dataset \\(D\\). For VLP models, we denote \\(f_{I}\\) as the image encoder and \\(f_{T}\\) as the text encoder. The multimodal fusion module in fused VLP models is denoted by \\(f_{M}\\). Specifically, \\(f_{I}(v)\\) represents the image representation \\(e_v\\) encoded" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "105" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.085, + 0.089, + 0.468, + 0.171 + ], + "angle": 0, + "content": "
AttackCross-modal Interactionp (Event A) ↑p (Event B | Event A) ↓
ALBEFTCLCLIPViTCLIPCNN
Sep-AttackNo74.60%46.78%41.69%41.82%
Co-AttackSingle-pair80.60%50.50%40.82%42.93%
SGASet-level97.20%28.91%34.67%38.58%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.182, + 0.472, + 0.274 + ], + "angle": 0, + "content": "Table 1: Adversarial images with insufficient cross-modal interactions may get closer to other paired captions when transferring. Event A: success cases in white-box settings. Event B: failure cases in black-box settings influenced by other paired captions. Adversarial data are generated on Flickr30K from ALBEF to attack other target models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.287, + 0.47, + 0.348 + ], + "angle": 0, + "content": "by \\( f_{I} \\) taking an image \\( v \\) as input, \\( f_{T}(t) \\) denotes the text representation \\( e_t \\) encoded by \\( f_{T} \\) taking a text \\( t \\) as input, and \\( f_{M}(e_v, e_t) \\) denotes the multimodal representation encoded by \\( f_{M} \\) taking image and text representations as inputs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.472, + 0.424 + ], + "angle": 0, + "content": "We use \\( B[v, \\epsilon_v] \\) and \\( B[t, \\epsilon_t] \\) to represent the legal searching spaces for optimizing adversarial image and text, respectively. Specifically, \\( \\epsilon_v \\) denotes the maximal perturbation bound for the image, and \\( \\epsilon_t \\) denotes the maximal number of changeable words in the caption." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.433, + 0.423, + 0.449 + ], + "angle": 0, + "content": "4.3. Transferable Set-Level Guidance Attack" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.47, + 0.562 + ], + "angle": 0, + "content": "Alignment-preserving Augmentation. The analysis presented in Section 3 highlights the key limitation of existing methods: the inter-modal information used to generate adversarial examples lacks diversity. The limitation will make the generated adversarial examples fail to generalize to other black-box models with strong attack performance, resulting in limited adversarial transferability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.471, + 0.775 + ], + "angle": 0, + "content": "To inject more diversity in the generation of generalizable adversarial examples, we propose using set-level alignment-preserving augmentation to expand multimodal input spaces while maintaining cross-modal alignments intact. Unlike previous methods that only consider a single image-text paired example \\((v,t)\\) to generate adversarial data, we enlarge the input to a set level of images and captions. Specifically, we select the most matching caption pairs from the dataset of each image \\(v\\) to form an augmented caption set \\(\\pmb{t} = \\{t_1,t_2,\\dots,t_M\\}\\), and resize each image \\(v\\) into different scales \\(S = \\{s_{1},s_{2},\\dots,s_{N}\\}\\) and then add Gaussian noise to obtain a multi-scale image set \\(\\pmb{v} = \\{v_{1},v_{2},\\dots,v_{N}\\}\\) based on the scale-invariant property. The enlarged input set \\((\\pmb{v},\\pmb{t})\\) is then used to generate the adversarial data \\((v^{\\prime},t^{\\prime})\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Cross-modal Guidance. Cross-modal interactions play a crucial role in multimodal tasks. For example, in imagetext retrieval, the paired information from another modality provides unique annotation supervision for each sample. Similarly, in adversarial attacks, supervisory information is essential in guiding the search for adversarial examples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "To fully utilize the enlarged alignment-preserving multimodal input set \\((\\pmb{v},\\pmb{t})\\) and further improve the transfer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.227 + ], + "angle": 0, + "content": "ability of the generated adversarial data, we propose cross-modal guidance to utilize interactions from different modalities. Specifically, we use the paired information from another modality as the supervision to guide the direction of optimizing the adversarial data. This guidance iteratively pushed away the multimodal information and disrupt the cross-modal interaction for better harmonious perturbations. Notably, the resultant adversarial examples can perceive the gradients originated from multiple guidance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.894, + 0.275 + ], + "angle": 0, + "content": "First, we generate corresponding adversarial captions for all captions in the text set \\( \\pmb{t} \\), forming an adversarial caption set \\( \\pmb{t}' = \\{t_1', t_2', \\dots, t_M'\\} \\). The process can be formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.281, + 0.893, + 0.318 + ], + "angle": 0, + "content": "\\[\nt _ {i} ^ {\\prime} = \\underset {t _ {i} ^ {\\prime} \\in B [ t _ {i}, \\epsilon_ {t} ]} {\\arg \\max } - \\frac {f _ {T} \\left(t _ {i} ^ {\\prime}\\right) \\cdot f _ {I} (v)}{\\| f _ {T} \\left(t _ {i} ^ {\\prime}\\right) \\| \\| f _ {I} (v) \\|.} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.893, + 0.372 + ], + "angle": 0, + "content": "The adversarial caption \\( t_i' \\) is constrained to be dissimilar to the original image \\( v \\) in the embedding space. Next, the adversarial image \\( v' \\) is generated by solving" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.378, + 0.893, + 0.416 + ], + "angle": 0, + "content": "\\[\nv ^ {\\prime} = \\underset {v ^ {\\prime} \\in B [ v, \\epsilon_ {v} ]} {\\arg \\max } - \\sum_ {i = 1} ^ {M} \\frac {f _ {T} \\left(t _ {i} ^ {\\prime}\\right)}{\\| f _ {T} \\left(t _ {i} ^ {\\prime}\\right) \\|} \\sum_ {s _ {i} \\in S} \\frac {f _ {I} \\left(g \\left(v ^ {\\prime} , s _ {i}\\right)\\right)}{\\| f _ {I} \\left(g \\left(v ^ {\\prime} , s _ {i}\\right)\\right) \\|}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.894, + 0.5 + ], + "angle": 0, + "content": "where \\( g(v', s_i) \\) denotes the resizing function that takes the image \\( v' \\) and the scale coefficient \\( s_i \\) as inputs. All the scaled images derived from \\( v' \\) are encouraged to be far away from all the adversarial captions \\( t_i' \\) in the embedding space. Finally, the adversarial caption \\( t' \\) is generated as follows," + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.506, + 0.893, + 0.539 + ], + "angle": 0, + "content": "\\[\nt ^ {\\prime} = \\underset {t ^ {\\prime} \\in B [ t, \\epsilon_ {t} ]} {\\arg \\max } - \\frac {f _ {T} \\left(t ^ {\\prime}\\right) \\cdot f _ {I} \\left(v ^ {\\prime}\\right)}{\\| f _ {T} \\left(t ^ {\\prime}\\right) \\| \\| f _ {I} \\left(v ^ {\\prime}\\right) \\|}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.893, + 0.591 + ], + "angle": 0, + "content": "in which \\( t' \\) is encouraged to be far away from the adversarial image \\( v' \\) in the embedding space. The detailed algorithm can be found in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.603, + 0.634, + 0.62 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.895, + 0.78 + ], + "angle": 0, + "content": "In this section, we present experimental evidence for the advantages of our proposed SGA. We conduct experiments on a diverse set of datasets and popular VLP models. First, we describe the experimental settings in Section 5.1. Then, in Section 5.2, we validate the immediate integration of transfer-based unimodal attacks into multimodal learning. Next, we provide the main evaluation results compared to the state-of-the-art method in Section 5.3. In Section 5.4, we analyze the cross-task transferability between different \\(\\mathrm{V + L}\\) tasks. Finally, we present ablation studies in Section 5.5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.707, + 0.804 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Datasets. We consider two widely used multimodal datasets, Flickr30K [27] and MSCOCO [23]. Flickr30K consists of 31,783 images, each with five corresponding captions. Similarly, MSCOCO comprises 123,287 images, and each image is annotated with around five captions. We adopt the Karpathy split [14] for experimental evaluation." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "106" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.162, + 0.089, + 0.813, + 0.272 + ], + "angle": 0, + "content": "
AttackALBEF*TCLCLIPViTCLIPCNN
TR R@1*IR R@1*TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1
Sep-Attack65.6973.9517.6032.9531.1745.2332.8245.49
Sep-Attack + MI58.81(↓)65.25(↓)16.02(↓)28.19(↓)23.07(↓)36.98(↓)26.56(↓)39.31(↓)
Sep-Attack + DIM56.41(↓)64.24(↓)16.75(↓)29.55(↓)24.17(↓)37.60(↓)25.54(↓)38.77(↓)
Sep-Attack + PNA.PO40.56(↓)53.95(↓)18.44(↑)30.98(↓)22.33(↓)37.02(↓)26.95(↓)38.63(↓)
Co-Attack77.1683.8615.2129.4923.6036.4825.1238.89
Co-Attack + MI64.86(↓)75.26(↓)25.40(↑)38.69(↑)24.91(↑)37.11(↑)26.31(↑)38.97(↑)
Co-Attack + DIM47.03(↓)62.28(↓)22.23(↑)35.45(↑)25.64(↑)38.50(↑)26.95(↑)40.58(↑)
SGA97.2497.2845.4255.2533.3844.1634.9346.57
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.284, + 0.893, + 0.331 + ], + "angle": 0, + "content": "Table 2: Attack success rates (%) of R@1 of integrating transfer-based image attacks on image-text retrieval. The adversarial data are generated on Flickr30K using the source model ALBEF to attack other target models. * indicates white-box attacks. A higher ASR indicates better adversarial transferability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.473, + 0.524 + ], + "angle": 0, + "content": "VLP Models. We evaluate two popular VLP models, the fused VLP and aligned VLP models. For the fused VLP, we consider ALBEF [18] and TCL [38]. ALBEF contains a 12-layer visual transformer ViT-B/16 [8] and two 6-layer transformers for the image encoder and both the text encoder and the multimodal encoder, respectively. TCL uses the same model architecture as ALBEF but with different pretrain objectives. For the aligned VLP model, we choose to evaluate CLIP [28]. CLIP has two different image encoder choices, namely, \\(\\mathrm{CLIP}_{\\mathrm{ViT}}\\) and \\(\\mathrm{CLIP}_{\\mathrm{CNN}}\\), that use ViT-B/16 and ResNet-101 [12] as the base architectures for the image encoder, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.529, + 0.472, + 0.682 + ], + "angle": 0, + "content": "Adversarial Attack Settings. To craft adversarial images, we employ PGD [24] with perturbation bound \\(\\epsilon_v = 2/255\\), step size \\(\\alpha = 0.5/255\\), and iteration steps \\(T = 10\\). For attacking text modality, we adopt BERT-Attack [19] with perturbation bound \\(\\epsilon_t = 1\\) and length of word list \\(W = 10\\). Furthermore, we enlarge the image set by resizing the original image into five scales, \\(\\{0.50, 0.75, 1.00, 1.25, 1.50\\}\\), using bicubic interpolation. Similarly, the caption set is enlarged by augmenting the most matching caption pairs for each image in the dataset, with the size of approximately five." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.472, + 0.779 + ], + "angle": 0, + "content": "Metrics. We employ Attack Success Rate (ASR) as the metric for evaluating the adversarial robustness and transferability in both white-box and black-box settings. Specifically, ASR evaluates the percentage of attacks that only produce successful adversarial examples. A higher ASR indicates better adversarial transferability." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.3, + 0.803 + ], + "angle": 0, + "content": "5.2. Transferability Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.472, + 0.902 + ], + "angle": 0, + "content": "In this paper, we present a systematic study of the adversarial transferability of VLP models, which has not been explored. As demonstrated in Section 3, existing methods, including the separate unimodal adversarial attack (SepAttack) and the multimodal adversarial attack (Co-Attack), exhibit limited transferability to other VLP models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.895, + 0.447 + ], + "angle": 0, + "content": "To improve transferability in multimodal learning, we intuitively investigate the adoption of transfer-based attacks from unimodal learning such as image classification. Specifically, we consider MI [7], DIM [37], and PNA.PO [36]. However, this approach can be problematic if cross-modal interactions and the unique many-to-many alignments in multimodal learning are not taken into account." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.897, + 0.705 + ], + "angle": 0, + "content": "Table 2 illustrates that multimodal attack methods that incorporate transfer-based image attacks exhibit minimal improvement in transferability while compromising white-box performance. Specifically, when integrated with MI, Co-Attack drops significantly by \\(12.3\\%\\) in white-box settings, while only maintaining \\(25.40\\%\\) ASR in transferability (ALBEF to TCL). However, our SGA shows superior performance in both white-box and black-box settings. Notably, Sep-Attack combined with transfer-based attacks not only reduces the effectiveness of white-box attacks but also fails to improve adversarial transferability in almost all black-box settings. The results provide empirical evidence that directly combining unimodal adversarial attacks in multimodal learning without considering cross-modal interactions and alignments can be problematic, even when using transfer-based unimodal attacks. Additional discussion is provided in Appendix B.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.704, + 0.728 + ], + "angle": 0, + "content": "5.3. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.885 + ], + "angle": 0, + "content": "Multimodal Fusion Modules. First, we investigate VLP models with different fusion modules, namely, fused VLP models and aligned VLP models. We generate adversarial examples on both types of models and evaluate their attack performance when transferred to other VLP models while ensuring consistency in the input size of images. For example, adversarial images generated by ALBEF or TCL are resized to \\(224 \\times 224\\) before performing transfer attacks on CLIP, and adversarial examples generated on CLIP are resized to \\(384 \\times 384\\) before being transferred to ALBEF or TCL." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.887, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 3, experimental results demonstrate" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.945, + 0.511, + 0.956 + ], + "angle": 0, + "content": "107" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.89, + 0.419 + ], + "angle": 0, + "content": "
SourceAttackALBEFTCLCLIPViTCLIPCNN
TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1
ALBEFPGD52.45*58.65*3.066.798.9613.2110.3414.65
BERT-Attack11.57*27.46*12.6428.0729.3343.1732.6946.11
Sep-Attack65.69*73.95*17.6032.9531.1745.2332.8245.49
Co-Attack77.16*83.86*15.2129.4923.6036.4825.1238.89
SGA97.24±0.22*97.28±0.15*45.42±0.6055.25±0.0633.38±0.3544.16±0.2534.93±0.9946.57±0.13
TCLPGD6.1510.7877.87*79.48*7.4813.7210.3415.33
BERT-Attack11.8926.8214.54*29.17*29.6944.4933.4646.07
Sep-Attack20.1336.4884.72*86.07*31.2944.6533.3345.80
Co-Attack23.1540.0477.94*85.59*27.8541.1930.7444.11
SGA48.91±0.7460.34±0.1098.37±0.08*98.81±0.07*33.87±0.1844.88±0.5437.74±0.2748.30±0.34
CLIPViTPGD2.504.934.858.1770.92*78.61*5.368.44
BERT-Attack9.5922.6411.8025.0728.34*39.08*30.4037.43
Sep-Attack9.5923.2511.3825.6079.75*86.79*30.7839.76
Co-Attack10.5724.3311.9426.6993.25*95.86*32.5241.82
SGA13.40±0.0727.22 ±0.0616.23±0.4530.76±0.0799.08±0.08*98.94±0.00*38.76±0.2747.79±0.58
CLIPCNNPGD2.094.824.007.811.106.6086.46*92.25*
BERT-Attack8.8623.2712.3325.4827.1237.4430.40*40.10*
Sep-Attack8.5523.4112.6426.1228.3439.4391.44*95.44*
Co-Attack8.7923.7413.1026.0728.7940.0394.76*96.89*
SGA11.42±0.0724.80±0.2814.91±0.0828.82±0.1131.24±0.4242.12±0.1199.24±0.18*99.49±0.05*
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.431, + 0.893, + 0.477 + ], + "angle": 0, + "content": "Table 3: Comparison with state-of-the-art methods on image-text retrieval. We report the attack success rate \\((\\%)\\) of R@1 for both IR and TR. The source column indicates the source models used to generate the adversarial data on Flickr30K. * indicates white-box attacks. A higher ASR indicates better adversarial transferability." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.495, + 0.441, + 0.563 + ], + "angle": 0, + "content": "
AttackB@4METEORROUGE_LCIDErSPICE
Baseline39.731.060.0133.323.8
Co-Attack37.429.858.4125.522.8
SGA34.828.456.3116.021.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.572, + 0.472, + 0.665 + ], + "angle": 0, + "content": "Table 4: Cross-Task Transferability: ITR \\(\\rightarrow\\) IC. Adversarial data generated from Image-Text Retrieval (ITR) to attack Image Captioning (IC) on MSCOCO. The source model and target model are ALBEF and BLIP, respectively. The Baseline represents the original performance of IC on clean data. Lower values indicate better adversarial transferability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.678, + 0.472, + 0.844 + ], + "angle": 0, + "content": "the superiority of our proposed SGA over existing multimodal attack methods in all black-box settings. Specifically, our SGA achieves significant improvements in adversarial transferability when the source and target models are of the same type. For instance, SGA outperforms Co-Attack by approximately \\(30\\%\\) in terms of attack success rate when transferring adversarial data from ALBEF to TCL. Moreover, in the more challenging scenario where the source and target models are of different types, SGA also surpasses CoAttack with higher attack success rates. More experiments on the MSCOCO dataset are provided in Appendix B.3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Model Architectures. Then, we explore VLP models with respect to different model architectures. Many VLP models commonly use ViTs as the vision encoder, where images" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.895, + 0.619 + ], + "angle": 0, + "content": "are segmented into patches before being processed by the transformer model. However, in the case of CLIP, the image encoder consists of two distinct architectures: conventional Convolutional Neural Networks (CNNs) and ViTs. The transferability between CNNs and ViTs has been well-studied in unimodal learning. Therefore, we also investigate the adversarial transferability of CNN-based and ViT-based CLIP models in multimodal learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.895, + 0.74 + ], + "angle": 0, + "content": "As shown in Table 3, we observe a similar phenomenon as unimodal learning that compared to CNNs, ViTs show better robustness against adversarial perturbations [25]. Specifically, for all attack methods, the same adversarial multimodal data have a stronger white-box attack effect on \\(\\mathrm{CLIP}_{\\mathrm{CNN}}\\) compared to \\(\\mathrm{CLIP}_{\\mathrm{ViT}}\\). Moreover, the adversarial examples generated on \\(\\mathrm{CLIP}_{\\mathrm{ViT}}\\) are found to be more transferable to \\(\\mathrm{CLIP}_{\\mathrm{CNN}}\\) than vice versa (38.76% vs. 31.24%)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.742, + 0.895, + 0.834 + ], + "angle": 0, + "content": "Furthermore, our proposed SGA consistently improves transferability on both CNN-based CLIP and ViT-based CLIP compared to other attacks. For instance, SGA increases the adversarial transferability for \\(\\mathrm{CLIP}_{\\mathrm{ViT}}\\) by \\(5.83\\%\\) and \\(6.24\\%\\) compared to Co-Attack under the white-box setting and black-box setting, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.846, + 0.742, + 0.864 + ], + "angle": 0, + "content": "5.4. Cross-Task Transferability" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Cross-modal interactions and alignments are the core components of multimodal learning regardless of the task." + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "108" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.137, + 0.089, + 0.416, + 0.157 + ], + "angle": 0, + "content": "
AttackValTestATestB
Baseline58.4665.8946.25
Co-Attack54.2661.8043.81
SGA53.5561.1943.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.169, + 0.472, + 0.261 + ], + "angle": 0, + "content": "Table 5: Cross-Task Transferability: ITR \\(\\rightarrow\\) VG. Adversarial data generated from Image-Text Retrieval (ITR) to attack Visual Grounding (VG) on RefCOCO+. The source model and target model are both ALBEF. The Baseline represents the original performance of VG on clean data. Lower values indicate better adversarial transferability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.471, + 0.323 + ], + "angle": 0, + "content": "Therefore, we conduct extensive experiments to explore the effectiveness of our proposed SGA on two additional V+L tasks: Image Captioning (IC) and Visual Grounding (VG)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.331, + 0.47, + 0.497 + ], + "angle": 0, + "content": "Image Captioning. Image captioning is a generation-based task, where an input image is encoded into a feature vector and then decoded into a natural language sentence. In our experiments, we craft adversarial images using the source model (ALBEF) with an image-text retrieval objective and then directly attack the target model (BLIP [17]) on image captioning. We employ the MSCOCO dataset, which is suitable for both two tasks, and utilize various evaluation metrics to measure the quality of the generated captions, including BLEU [26], METEOR [2], ROUGE [21], CIDEr [32], and SPICE [1]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.498, + 0.471, + 0.589 + ], + "angle": 0, + "content": "We present the performance on image captioning of BLIP after being attacked in Table 4. Experimental results demonstrate clear improvements in the adversarial transferability of the proposed SGA compared to Co-Attack. Specifically, our SGA improves the BLEU score by up to \\(2.6\\%\\) and the CIDEr score by up to \\(9.5\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.472, + 0.706 + ], + "angle": 0, + "content": "Visual Grounding. Visual grounding is another \\(\\mathrm{V} + \\mathrm{L}\\) task that aims to localize the region in an image based on the corresponding specific textual description. Similarly, we generate adversarial images using the source model (ALBEF) from image-text to attack the target model (ALBEF) on visual grounding. Table 5 shows the results on RefCOCO+ [39], where our SGA still outperforms Co-Attack." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.716, + 0.231, + 0.733 + ], + "angle": 0, + "content": "5.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.74, + 0.471, + 0.862 + ], + "angle": 0, + "content": "To systematically investigate the impact of our set-level alignment-preserving augmentations, we conducted ablation experiments on image-text retrieval to evaluate the effect of varying the number of augmented image sets \\(N\\) with multiscale transformation and the number \\(M\\) of augmented caption sets. Specifically, we employed ALBEF as the source model and \\(\\mathrm{CLIP}_{\\mathrm{ViT}}\\) as the target model on Flickr30K. More details are provided in Appendix B.4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Multi-scale Image Set. We propose the use of multiple scale-invariant images to generate diverse adversarial data in" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.089, + 0.696, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.089, + 0.892, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.228, + 0.895, + 0.29 + ], + "angle": 0, + "content": "Figure 5: Ablation Study: Attack success rates (\\%) on set-level augmentations, Left for image set and Right for caption set. The performance of SGA is consistently improved by using larger augmented image and caption sets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.895, + 0.381 + ], + "angle": 0, + "content": "SGA. Results in the left of Figure 5 reveal that the transferability significantly increases as we introduce more diverse images with different scales, peaking when the scale range is set to [0.50, 1.50] with a step of 0.25. We set the scale range \\( S = \\{0.50, 0.75, 1.00, 1.25, 1.50\\} \\) for optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.897, + 0.583 + ], + "angle": 0, + "content": "Multi-pair Caption Set. We also conduct experiments to investigate the impact of an enlarged caption set on adversarial transferability. The number of additional captions ranged from 1 to \\( M \\), where \\( M \\) represents the most matching caption pairs from the dataset for each image. Results presented in the right panel of Figure 5 indicate that if \\( M > 1 \\), the black-box performance increases significantly but eventually plateaus. These results demonstrate the effectiveness of using multiple alignment-preserving inter-modal information to enhance adversarial transferability. Furthermore, we observed that the performance is relatively insensitive to the number of extra captions, but adding more captions can improve the overall adversarial transferability." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.597, + 0.619, + 0.613 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.895, + 0.833 + ], + "angle": 0, + "content": "In this paper, we make the first attempt to investigate the adversarial transferability of typical VLP models. We systematically evaluate the existing attack methods and reveal that they still exhibit lower transferability, despite their impressive performance in white-box settings. Our investigation highlights the need for specially designed transferable attacks in multimodal learning that can model the many-to-many cross-modal alignments and interactions. We propose SGA, a highly transferable multimodal attack, which leverages set-level alignment-preserving augmentations through cross-modal guidance to thoroughly exploit multimodal interactions. We hope that this work could inspire further research to evaluate and enhance the adversarial robustness of VLP models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported by the National Key R&D Program of China (Grant NO. 2022YFF1202903) and the National Natural Science Foundation of China (Grant NO. 62122035)." + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "109" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In ECCV, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.472, + 0.201 + ], + "angle": 0, + "content": "[2] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In IEEvaluation@ACL, 2005. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.203, + 0.47, + 0.245 + ], + "angle": 0, + "content": "[3] Nicholas Carlini and David A. Wagner. Towards evaluating the robustness of neural networks. 2017 IEEE Symposium on Security and Privacy (SP), pages 39-57, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.247, + 0.471, + 0.316 + ], + "angle": 0, + "content": "[4] Hui Chen, Guiguang Ding, Xudong Liu, Zijia Lin, Ji Liu, and Jungong Han. Imram: Iterative matching with recurrent attention memory for cross-modal image-text retrieval. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12652-12660, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.318, + 0.472, + 0.372 + ], + "angle": 0, + "content": "[5] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.375, + 0.471, + 0.457 + ], + "angle": 0, + "content": "[6] Mengjun Cheng, Yipeng Sun, Long Wang, Xiongwei Zhu, Kun Yao, Jie Chen, Guoli Song, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Vista: Vision and scene text aggregation for cross-modal retrieval. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5174-5183, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.46, + 0.47, + 0.515 + ], + "angle": 0, + "content": "[7] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.517, + 0.471, + 0.599 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv, abs/2010.11929, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.602, + 0.471, + 0.67 + ], + "angle": 0, + "content": "[9] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuo-hang Wang, Lijuan Wang, Chenguang Zhu, Nanyun Peng, Zicheng Liu, and Michael Zeng. An empirical study of training end-to-end vision-and-language transformers. ArXiv, abs/2111.02387, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.47, + 0.728 + ], + "angle": 0, + "content": "[10] Ji Gao, Jack Lanchantin, Mary Lou Soffa, and Yanjun Qi. Black-box generation of adversarial text sequences to evade deep learning classifiers. 2018 IEEE Security and Privacy Workshops (SPW), pages 50-56, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.47, + 0.77 + ], + "angle": 0, + "content": "[11] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. CoRR, abs/1412.6572, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.47, + 0.828 + ], + "angle": 0, + "content": "[12] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[13] Di Jin, Zhijing Jin, Joey Tianyi Zhou, and Peter Szolovits. Is bert really robust? a strong baseline for natural language attack on text classification and entailment. In AAAI, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. IEEE Transactions" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.893, + 0.119 + ], + "angle": 0, + "content": "on Pattern Analysis and Machine Intelligence, 39:664-676, 2017.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.894, + 0.177 + ], + "angle": 0, + "content": "[15] Zaid Khan and Yun Raymond Fu. Exploiting bert for multimodal target sentiment classification through input space translation. Proceedings of the 29th ACM International Conference on Multimedia, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.893, + 0.247 + ], + "angle": 0, + "content": "[16] Chenyi Lei, Shixian Luo, Yong Liu, Wanggui He, Jiamang Wang, Guoxin Wang, Haihong Tang, Chunyan Miao, and Houqiang Li. Understanding chinese video and language via contrastive multimodal pre-training. Proceedings of the 29th ACM International Conference on Multimedia, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.894, + 0.303 + ], + "angle": 0, + "content": "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. ArXiv, abs/2201.12086, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.894, + 0.373 + ], + "angle": 0, + "content": "[18] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq R. Joty, Caiming Xiong, and Steven C. H. Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.893, + 0.417 + ], + "angle": 0, + "content": "[19] Linyang Li, Ruotian Ma, Qipeng Guo, X. Xue, and Xipeng Qiu. Bert-attack: Adversarial attack against bert using bert. ArXiv, abs/2004.09984, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.893, + 0.487 + ], + "angle": 0, + "content": "[20] Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.893, + 0.532 + ], + "angle": 0, + "content": "[21] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Annual Meeting of the Association for Computational Linguistics, 2004. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.893, + 0.574 + ], + "angle": 0, + "content": "[22] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. arXiv: Learning, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.577, + 0.893, + 0.63 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.893, + 0.687 + ], + "angle": 0, + "content": "[24] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. ArXiv, abs/1706.06083, 2018. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.893, + 0.745 + ], + "angle": 0, + "content": "[25] Muzammal Naseer, Kanchana Ranasinghe, Salman Hameed Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In Neural Information Processing Systems, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.893, + 0.787 + ], + "angle": 0, + "content": "[26] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.893, + 0.858 + ], + "angle": 0, + "content": "[27] Bryan A. Plummer, Liwei Wang, Christopher M. Cervantes, Juan C. Caicedo, J. Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. International Journal of Computer Vision, 123:74–93, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.861, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "110" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.133 + ], + "angle": 0, + "content": "Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.472, + 0.176 + ], + "angle": 0, + "content": "[29] Shuhuai Ren, Yihe Deng, Kun He, and Wanxiang Che. Generating natural language adversarial examples through probability weighted word saliency. In ACL, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[30] Lei Shi, Kai Shuang, Shijie Geng, Peng Gao, Zuohui Fu, Gerard de Melo, Yunpeng Chen, and Sen Su. Dense contrastive visual-linguistic pretraining. Proceedings of the 29th ACM International Conference on Multimedia, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.472, + 0.287 + ], + "angle": 0, + "content": "[31] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.472, + 0.343 + ], + "angle": 0, + "content": "[32] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4566-4575, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.472, + 0.413 + ], + "angle": 0, + "content": "[33] Teng Wang, Yixiao Ge, Feng Zheng, Ran Cheng, Ying Shan, Xiaohu Qie, and Ping Luo. Accelerating vision-language pretraining with free language modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23161-23170, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.472, + 0.482 + ], + "angle": 0, + "content": "[34] Teng Wang, Wenhao Jiang, Zhichao Lu, Feng Zheng, Ran Cheng, Chengguo Yin, and Ping Luo. Vlmixer: Unpaired vision-language pre-training via cross-modal cutmix. In International Conference on Machine Learning, pages 22680-22690. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.472, + 0.551 + ], + "angle": 0, + "content": "[35] Zihao Wang, Xihui Liu, Hongsheng Li, Lu Sheng, Junjie Yan, Xiaogang Wang, and Jing Shao. Camp: Cross-modal adaptive message passing for text-image retrieval. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5763-5772, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.472, + 0.607 + ], + "angle": 0, + "content": "[36] Zhipeng Wei, Jingjing Chen, Micah Goldblum, Zuxuan Wu, Tom Goldstein, and Yu-Gang Jiang. Towards transferable adversarial attacks on vision transformers. ArXiv, abs/2109.04176, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.472, + 0.678 + ], + "angle": 0, + "content": "[37] Cihang Xie, Zhishuai Zhang, Jianyu Wang, Yuyin Zhou, Zhou Ren, and Alan Loddon Yuille. Improving transferability of adversarial examples with input diversity. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2725-2734, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.472, + 0.734 + ], + "angle": 0, + "content": "[38] Jinyu Yang, Jiali Duan, S. Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul M. Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. ArXiv, abs/2202.10401, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.472, + 0.775 + ], + "angle": 0, + "content": "[39] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C. Berg, and Tamara L. Berg. Modeling context in referring expressions. ArXiv, abs/1608.00272, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.472, + 0.845 + ], + "angle": 0, + "content": "[40] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Francis E. H. Tay, Jiashi Feng, and Shuicheng Yan. Tokensto-token vit: Training vision transformers from scratch onImagenet. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 538-547, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[41] Jiaming Zhang, Qiaomin Yi, and Jitao Sang. Towards adversarial attack on vision-language pre-training models. Proceedings of the 30th ACM International Conference on Multimedia, 2022. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[42] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5575-5584, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.22 + ], + "angle": 0, + "content": "[43] Qi Zhang, Zhen Lei, Zhaoxiang Zhang, and S. Li. Context-aware attention network for image-text retrieval. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3533-3542, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "111" + } + ] +] \ No newline at end of file diff --git a/2023/Set-level Guidance Attack_ Boosting Adversarial Transferability of Vision-Language Pre-training Models/4d2215bf-24a1-45b3-bcea-24623e459934_origin.pdf b/2023/Set-level Guidance Attack_ Boosting Adversarial Transferability of Vision-Language Pre-training Models/4d2215bf-24a1-45b3-bcea-24623e459934_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3b3a6dc2006e1d0f6538e07ec9a2b277a245e12d --- /dev/null +++ b/2023/Set-level Guidance Attack_ Boosting Adversarial Transferability of Vision-Language Pre-training Models/4d2215bf-24a1-45b3-bcea-24623e459934_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c9f1f59cdff7090b7e788401b92b61f3237dc696dda9b3e9b6b7f5666445285 +size 822955 diff --git a/2023/Set-level Guidance Attack_ Boosting Adversarial Transferability of Vision-Language Pre-training Models/full.md b/2023/Set-level Guidance Attack_ Boosting Adversarial Transferability of Vision-Language Pre-training Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..aa1c0a23dae269dea39da7d118de53b8ed1d44fe --- /dev/null +++ b/2023/Set-level Guidance Attack_ Boosting Adversarial Transferability of Vision-Language Pre-training Models/full.md @@ -0,0 +1,311 @@ +# Set-level Guidance Attack: Boosting Adversarial Transferability of Vision-Language Pre-training Models + +Dong Lu1, Zhiqiang Wang1, Teng Wang1,2, Weili Guan3, Hongchang Gao4, Feng Zheng1,5† + +$^{1}$ Southern University of Science and Technology $^{2}$ The University of Hong Kong + +3Monash University 4Temple University 5Peng Cheng Laboratory + +sammylu\_outlook.com wangzq_2021\@outlook.com tengwang@connect.hku.hk + +honeyguan@gmail.com hongchang.gao@temple.edu f.zheng@ieee.org + +# Abstract + +Vision-language pre-training (VLP) models have shown vulnerability to adversarial examples in multimodal tasks. Furthermore, malicious adversaries can be deliberately transferred to attack other black-box models. However, existing work has mainly focused on investigating white-box attacks. In this paper, we present the first study to investigate the adversarial transferability of recent VLP models. We observe that existing methods exhibit much lower transferability, compared to the strong attack performance in white-box settings. The transferability degradation is partly caused by the under-utilization of cross-modal interactions. Particularly, unlike unimodal learning, VLP models rely heavily on cross-modal interactions and the multimodal alignments are many-to-many, e.g., an image can be described in various natural languages. To this end, we propose a highly transferable Set-level Guidance Attack (SGA) that thoroughly leverages modality interactions and incorporates alignment-preserving augmentation with cross-modal guidance. Experimental results demonstrate that SGA could generate adversarial examples that can strongly transfer across different VLP models on multiple downstream vision-language tasks. On image-text retrieval, SGA significantly enhances the attack success rate for transfer attacks from ALBEF to TCL by a large margin (at least $9.78\%$ and up to $30.21\%$ ), compared to the state-of-the-art. Our code is available at https://github.com/Zoky-2020/SGA. + +# 1. Introduction + +Recent work has shown that vision-language pre-training (VLP) models are still vulnerable to adversarial examples [41], even though they have achieved remarkable performance on a wide range of multimodal tasks [30, 15, 16]. Existing work mainly focuses on white-box attacks, where + +![](images/57abe961e11c257578b9d684d65e3eb931de0acc13b89bc92fa04105a5eba159.jpg) +(a) ALBEF to ALBEF + +![](images/703a6a22fcd62c73203da94b438e243c35a7028ca3d55fc818e0500d0ab6f195.jpg) +(b) ALBEF to TCL + +![](images/79b7480f3214c69978d30e31bb2733b3dfefb4eca3b2d5596566f43489447b39.jpg) +(c) ALBEF to $\mathrm{CLIP}_{\mathrm{VIT}}$ +Figure 1: Comparison of attack success rates (ASR) using five different attacks on image-text retrieval. Adversarial examples are crafted on the source model (ALBEF) to attack the target white-box model or black-box models. The first three columns refer to the image-only PGD attack [24], text-only BERT-Attack [19] (BA), and the combined separate unimodal attack (SA), which all belong to the methods without cross-modal interactions. The fourth column is the state-of-the-art multimodal Co-Attack [41] (CA) that employs single-pair cross-modal interactions. The last column is the proposed Set-level Guidance Attack (SGA), which leverages multiple set-level cross-modal interactions, successfully attacking the white-box model and transferring to attack all black-box models with the highest ASR. More discussions are in Section 3. + +![](images/cab211f7941f67f431d4ce6995fa79556eb3bb5696d8b8da33990891b2e141a3.jpg) +(d) ALBEF to $\mathrm{CLIP}_{\mathrm{CNN}}$ + +information about the victim model is accessible. However, the transferability of adversarial examples across VLP models has not been investigated, which is a more practical setting. It is still unknown whether the adversarial data generated on the source model can successfully attack another model, which poses a serious security risk to the deployment of VLP models in real-world applications. + +This paper makes the first step to investigate the trans + +![](images/092ee924aedd1ff0cfc95c7e445420276c5b21a939008f621bb56c95c838490b.jpg) +(a) Sep-Attack + +![](images/a59f5fb7f155a9f918d1474c0d37ac11b00d845839703e7bc2b47746618723ee.jpg) + +![](images/3102c8a89a59e8d2a501f6e7ed757080b89fbaefed782e5e8ad9d6a13833fd86.jpg) +(b) Co-Attack + +![](images/cae2c94049f04114f405fa8b9b227919dd7078f15ef9c244058a2e99349902f6.jpg) +(c) SGA +Figure 2: Comparison of cross-modal interactions. To generate adversarial examples, existing methods either: (a) separately perturb unimodal data without any cross-modal interactions (Sep-Attack) or, (b) perturb multiple modalities but with single image-text pairs to model cross-modal interactions (Co-Attack). However, our method is capable of learning cross-modal interactions among multiple alignments through (c) set-level guidance (SGA). Specifically, $v$ denotes the input image, and $t$ is the paired caption. $v'$ and $t'$ represent the corresponding adversarial examples, respectively. $U(v)$ and $U(t)$ represent the scale-invariant image set and most matching caption set. Arrows indicate the guidance for generating adversarial examples. + +ferability of adversarial samples within VLP models. Without loss of generality, most of our experiments are based on image-text retrieval tasks. We first empirically evaluate this attack performance with respect to different modalities on multimodal tasks across multiple datasets. Our results show that the adversarial transferability of attacking both modalities (image & text) consistently beats attacking unimodal data (image or text). Unfortunately, even though two modalities are allowed to be perturbed simultaneously, the attack success rates of existing methods [24, 19, 41] still significantly drops when transferring from the white-box to black-box settings, as shown in Figure 1. + +Different from recent studies focusing on separate attacks on unimodal data [24, 19], multimodal pairs exhibit intrinsic alignment and complementarity to each other. The modeling of inter-modal correspondence turns out to be a critical problem for transferability. Considering that the alignments between image and text are many-to-many, for example, an image could be described to be with various human perspectives and language styles, a reasonable perturbation direction may be determined with diverse guidance from multiple labels in the other modality. However, recent adversarial attack methods for VLP models [41] usually employ a single image-text pair to generate adversarial samples. Although they exhibit strong performance in white-box settings, the poor diversity of guidance makes adversarial samples highly correlated with the alignment pattern of the white-box model, and therefore impedes generalization to black-box settings. + +To address the weak transferability problem, we propose Set-level Guidance Attack (SGA), which leverages diverse cross-modal interactions among multiple image-text pairs + +(Figure 2). Specifically, we introduce alignment-preserving augmentation which enriches image-text pairs while keeping their alignments intact. The image augmentation is based on the scale-invariant property of deep learning models [22], thus we can construct multi-scale images to increase the diversity. For text augmentation, we select the most matching caption pairs from the dataset. More importantly, SGA generates adversarial examples on multimodal augmented input data with carefully designed cross-modal guidance. In detail, SGA iteratively pushes supplemental information away between two modalities with another modality as supervision to disrupt the interactions for better harmonious perturbations. Note that resultant adversarial samples could perceive the gradients originated from multiple guidance. + +We conduct experiments on two well-established multimodal datasets, Flickr30K [27] and MSCOCO [23], to evaluate the performance of our proposed SGA across various Vision-and-Language $(\mathrm{V} + \mathrm{L})$ downstream tasks. The experimental results demonstrate the high effectiveness of SGA in generating adversarial examples that can be strongly transferred across VLP models, surpassing the current state-of-the-art attack methods in multimodal learning. In particular, SGA achieves notable improvements in image-text retrieval under black-box settings and also exhibits superior performance in white-box attack settings. Moreover, SGA also outperforms the state-of-the-art methods in image captioning and yields higher fooling rates on visual grounding. + +We summarize our contributions as follows. 1) We make the first attempt to explore the transferability of adversarial examples on popular VLP models with a systematical evaluation; 2) We provide SGA, a novel transferable multimodal attack that enhances adversarial transferability through the effective use of set-level alignment-preserving augmentations and well-designed cross-modal guidance; 3) Extensive experiments show that SGA consistently boosts adversarial transferability across different VLP models than the state-of-the-art methods. + +# 2. Related Work + +# 2.1. Vision-Language Pre-training Models + +Vision-Language Pre-training (VLP) aims to improve the performance of downstream multimodal tasks by pretraining large-scale image-to-text pairs [17]. Most works are developed upon the pre-trained object detectors with region features to learn the vision-language representations [5, 20, 42, 34]. Recently, with the increasing popularity of Vision Transformer (ViT) [8, 31, 40], some other works propose to use ViT as an image encoder and transform the input into patches in an end-to-end manner [18, 38, 17, 9, 33]. + +According to the VLP architectures, VLP models can be classified into two typical types: fused VLP models and aligned VLP models [41]. Specifically, fused VLP models + +(e.g., ALBEF [18], TCL [38]) first utilize separate unimodal encoders to process token embeddings and visual features, and further use a multimodal encoder to process image and text embeddings to output fused multimodal embeddings. Alternatively, aligned VLP models (e.g., CLIP [28]) have only unimodal encoders with independent image and text modality embeddings. In this paper, we focus on popular architectures with fused and aligned VLP models. + +# 2.2. Image-Text Retrieval Task + +Image-Text Retrieval (ITR) aims to retrieve the relevant top-ranked instances from a gallery database with one modality, given an input query from another modality [35, 4, 43, 6]. This task can be divided into two subtasks, image-to-text retrieval (TR) and text-to-image retrieval (IR). + +For ALBEF [18] and TCL [38], the semantic similarity score in the unimodal embedding space will be calculated for all image-text pairs to select top- $k$ candidates. Then the multimodal encoder takes the top- $k$ candidates and computes the image-text matching score for ranking. For CLIP [28], without the multimodal encoder, the final rank list can be obtained based on the similarity in the embedding space between image and text modalities. + +# 2.3. Adversarial Transferability + +Existing adversarial attacks can be categorized into two settings: white-box attacks and black-box attacks. In a white-box setting, the target model is fully accessible, but not in a black-box setting. In computer vision, many methods employ gradient information for adversarial attacks in white-box settings, such as FGSM [11], PGD [24], C&W [3], and MI [7]. In contrast, in the field of natural language processing (NLP), current attack methods mainly modify or replace some tokens of the input text [19, 29, 10, 13]. In the multimodal vision-language domain, Zhang et al. [41] proposed a white-box multimodal attack method with respect to popular VLP models on downstream tasks. + +However, white-box attacks are unrealistic due to the inaccessibility of model information in practical applications. In addition, there is no related work that systematically analyzes the adversarial transferability of multimodal attack methods on VLP models. Therefore, in this work, we mainly focus on generating highly transferable adversaries across different VLP models. + +# 3. Analysis of Adversarial Transferability + +In this section, we conduct an empirical study on VLP models to evaluate adversarial transferability using existing methods. A common approach to attack multimodal tasks is combining unimodal adversarial attacks [24, 7, 37, 19] of each modalities together. For instance, the separate unimodal attack (Sep-Attack) includes PGD [24] and BERT-Attack + +![](images/31377960b66f9787e47cdfd1586b6c6be547d5ff88b9faa3601d3e36623fef34.jpg) +Figure 3: Attack success rates $(\%)$ of different perturbed modalities on image-text retrieval. Adversarial examples with respect to different modalities from the source model to attack the target model using Sep-Attack. Different colors represent different perturbed modalities. + +![](images/c71c566bd1ede3d6e42a054d67a7eb090e7241a954e637d505f0a839317aa523.jpg) +Figure 4: Attack success rates $(\%)$ on white-box models and black-box models. Both images and text are perturbed by Sep-Attack (Left) and Co-Attack (Right) on image-text retrieval. Colors indicate different target models. + +[19] for attacking image modality and text modality, respectively. Other recent multimodal adversarial attacks, such as Co-Attack [41], consider cross-modal interactions by perturbing image modalities and text modalities collectively. + +We first present the observations regarding the adversarial transferability of VLP models. Then we discuss the limitations of the existing methods. By conducting this study, we aim to provide insights into the robustness of VLP models against adversarial attacks and the effectiveness of different attack strategies. + +# 3.1. Observations + +To investigate the adversarial transferability of perturbed inputs with respect to different modalities (i.e., image, text, and image & text) and the effect of different VLP models on transferability, we conduct experiments and present the attack success rates of the adversarial examples generated by the source model to attack the target models in Figure 3 and Figure 4. The observations are summarized below: + +- The adversarial transferability of attacking both modalities (image & text) is consistently more effective than attacking any unimodal data alone (image or text). As shown in Figure 3, transferring both adversarial image and text from ALBEF to TCL leads to a much higher + +attack success rate than transferring adversarial examples of any single modality. Notably, ALBEF and TCL are both fused VLP models. Similar observations also exist in the following settings: (1) The source and target models are different types of VLP models but have the same basic architectures (e.g., TCL and $\mathrm{CLIP}_{\mathrm{ViT}}$ ). (2) The source and target models are the same types of VLP models, but with different basic architectures (e.g., $\mathrm{CLIP}_{\mathrm{ViT}}$ and $\mathrm{CLIP}_{\mathrm{CNN}}$ ). + +- Adversarial multimodal data (i.e., adversarial image & text), which have strong attack performance on the source model, can hardly maintain the same capability when transferring to target models. For example, as illustrated in Figure 4, even though ALBEF and TCL have the same model architecture, the attack success rate sees a significant drop when transferring adversarial examples generated on ALBEF to TCL. The phenomenon exists in both Sep-Attack and Co-Attack. + +In summary, the attack methods can have stronger transferability in black-box settings if all the modalities are attacked simultaneously. However, even though two modalities are allowed to be perturbed, existing methods still exhibit much lower transferability. This suggests that attacks with higher transferability should be specifically designed, instead of directly utilizing the existing white-box attack methods. + +# 3.2. Discussions + +We posit that the degradation in transferability of adversarial examples is mainly due to the limitations of the existing attack methods: + +- One major limitation of Sep-Attack is that it does not take into account the interactions between different modalities. As a combined independent attack method for each modality, it fails to model the inter-modal correspondence that is crucial for successful attacks in multimodal learning. This is particularly evident in multimodal tasks such as image-text retrieval, where the ground truth is not discrete labels (e.g., image classification) but another modality data that corresponds to the input modality. The complete lack of cross-modal interactions in Sep-Attack severely limits the generalization of adversarial examples and reduces their transferability among different VLP models. +- While Co-Attack is designed to leverage the collaboration between modalities to generate adversarial examples, it still suffers from a key drawback that hinders its transferability to other VLP models. Unlike unimodal learning, multimodal learning involves multiple complementary modalities with many-to-many cross-modal alignments, which pose unique challenges to + +achieving sufficient adversarial transferability. However, Co-Attack only uses single image-text pairs to generate adversarial data, limiting the diversity of guidance from multiple labels in other modalities. This lack of diversity in cross-modal guidance makes adversarial samples highly correlated with the alignment pattern of the white-box model. Therefore, the generality of adversarial examples is restricted, and their effectiveness in transferring to other models drops. + +In conclusion, the analysis motivates our investigation into the adversarial transferability of VLP models. Moreover, it highlights the pressing need to explore transferable multimodal attacks for generating adversarial examples that can be effectively transferred across different VLP models. + +# 4. Methodology + +In this section, we propose a transferable multimodal adversarial attack, termed Set-level Guidance Attack (SGA). SGA is designed to enhance adversarial transferability across VLP models by leveraging multimodal interactions and incorporating diverse alignment-preserving inter-modal information with carefully designed cross-modal guidance. We provide our motivation first, followed by relevant notations, and finally present SGA in detail. + +# 4.1. Motivation + +To improve the transferability of multimodal attacks, we first conduct an investigation into the shortcomings of existing methods in black-box settings. Through a systematic analysis of failure cases, we observe that around half of these cases arise due to the presence of multiple matched captions of the image, as shown in Table 1. More specifically, our findings indicate that while the generated adversarial image may be significantly distant from the single supervised caption in the source model, it is prone to approaching other matched captions in the target model, where the alignments can be modeled and ultimately lead to attack failure. + +Therefore, to maintain the attack ability of adversarial images when transferring to other models, it is crucial to consider multiple paired captions and push the adversarial image away from all the paired captions, thus preserving the attack ability when transferred to other black-box models. Crafting adversarial text for high transferability follows a similar approach, which can also benefit from more paired images. More discussions can be found in Appendix A. + +# 4.2. Notations + +Let $(v,t)$ denote an image-text pair sampled from a multimodal dataset $D$ . For VLP models, we denote $f_{I}$ as the image encoder and $f_{T}$ as the text encoder. The multimodal fusion module in fused VLP models is denoted by $f_{M}$ . Specifically, $f_{I}(v)$ represents the image representation $e_v$ encoded + +
AttackCross-modal Interactionp (Event A) ↑p (Event B | Event A) ↓
ALBEFTCLCLIPViTCLIPCNN
Sep-AttackNo74.60%46.78%41.69%41.82%
Co-AttackSingle-pair80.60%50.50%40.82%42.93%
SGASet-level97.20%28.91%34.67%38.58%
+ +Table 1: Adversarial images with insufficient cross-modal interactions may get closer to other paired captions when transferring. Event A: success cases in white-box settings. Event B: failure cases in black-box settings influenced by other paired captions. Adversarial data are generated on Flickr30K from ALBEF to attack other target models. + +by $f_{I}$ taking an image $v$ as input, $f_{T}(t)$ denotes the text representation $e_t$ encoded by $f_{T}$ taking a text $t$ as input, and $f_{M}(e_v, e_t)$ denotes the multimodal representation encoded by $f_{M}$ taking image and text representations as inputs. + +We use $B[v, \epsilon_v]$ and $B[t, \epsilon_t]$ to represent the legal searching spaces for optimizing adversarial image and text, respectively. Specifically, $\epsilon_v$ denotes the maximal perturbation bound for the image, and $\epsilon_t$ denotes the maximal number of changeable words in the caption. + +# 4.3. Transferable Set-Level Guidance Attack + +Alignment-preserving Augmentation. The analysis presented in Section 3 highlights the key limitation of existing methods: the inter-modal information used to generate adversarial examples lacks diversity. The limitation will make the generated adversarial examples fail to generalize to other black-box models with strong attack performance, resulting in limited adversarial transferability. + +To inject more diversity in the generation of generalizable adversarial examples, we propose using set-level alignment-preserving augmentation to expand multimodal input spaces while maintaining cross-modal alignments intact. Unlike previous methods that only consider a single image-text paired example $(v,t)$ to generate adversarial data, we enlarge the input to a set level of images and captions. Specifically, we select the most matching caption pairs from the dataset of each image $v$ to form an augmented caption set $\pmb{t} = \{t_1,t_2,\dots,t_M\}$ , and resize each image $v$ into different scales $S = \{s_{1},s_{2},\dots,s_{N}\}$ and then add Gaussian noise to obtain a multi-scale image set $\pmb{v} = \{v_{1},v_{2},\dots,v_{N}\}$ based on the scale-invariant property. The enlarged input set $(\pmb{v},\pmb{t})$ is then used to generate the adversarial data $(v^{\prime},t^{\prime})$ . + +Cross-modal Guidance. Cross-modal interactions play a crucial role in multimodal tasks. For example, in imagetext retrieval, the paired information from another modality provides unique annotation supervision for each sample. Similarly, in adversarial attacks, supervisory information is essential in guiding the search for adversarial examples. + +To fully utilize the enlarged alignment-preserving multimodal input set $(\pmb{v},\pmb{t})$ and further improve the transfer + +ability of the generated adversarial data, we propose cross-modal guidance to utilize interactions from different modalities. Specifically, we use the paired information from another modality as the supervision to guide the direction of optimizing the adversarial data. This guidance iteratively pushed away the multimodal information and disrupt the cross-modal interaction for better harmonious perturbations. Notably, the resultant adversarial examples can perceive the gradients originated from multiple guidance. + +First, we generate corresponding adversarial captions for all captions in the text set $\pmb{t}$ , forming an adversarial caption set $\pmb{t}' = \{t_1', t_2', \dots, t_M'\}$ . The process can be formulated as, + +$$ +t _ {i} ^ {\prime} = \underset {t _ {i} ^ {\prime} \in B [ t _ {i}, \epsilon_ {t} ]} {\arg \max } - \frac {f _ {T} \left(t _ {i} ^ {\prime}\right) \cdot f _ {I} (v)}{\| f _ {T} \left(t _ {i} ^ {\prime}\right) \| \| f _ {I} (v) \|.} \tag {1} +$$ + +The adversarial caption $t_i'$ is constrained to be dissimilar to the original image $v$ in the embedding space. Next, the adversarial image $v'$ is generated by solving + +$$ +v ^ {\prime} = \underset {v ^ {\prime} \in B [ v, \epsilon_ {v} ]} {\arg \max } - \sum_ {i = 1} ^ {M} \frac {f _ {T} \left(t _ {i} ^ {\prime}\right)}{\| f _ {T} \left(t _ {i} ^ {\prime}\right) \|} \sum_ {s _ {i} \in S} \frac {f _ {I} \left(g \left(v ^ {\prime} , s _ {i}\right)\right)}{\| f _ {I} \left(g \left(v ^ {\prime} , s _ {i}\right)\right) \|}, \tag {2} +$$ + +where $g(v', s_i)$ denotes the resizing function that takes the image $v'$ and the scale coefficient $s_i$ as inputs. All the scaled images derived from $v'$ are encouraged to be far away from all the adversarial captions $t_i'$ in the embedding space. Finally, the adversarial caption $t'$ is generated as follows, + +$$ +t ^ {\prime} = \underset {t ^ {\prime} \in B [ t, \epsilon_ {t} ]} {\arg \max } - \frac {f _ {T} \left(t ^ {\prime}\right) \cdot f _ {I} \left(v ^ {\prime}\right)}{\| f _ {T} \left(t ^ {\prime}\right) \| \| f _ {I} \left(v ^ {\prime}\right) \|}, \tag {3} +$$ + +in which $t'$ is encouraged to be far away from the adversarial image $v'$ in the embedding space. The detailed algorithm can be found in Appendix C. + +# 5. Experiments + +In this section, we present experimental evidence for the advantages of our proposed SGA. We conduct experiments on a diverse set of datasets and popular VLP models. First, we describe the experimental settings in Section 5.1. Then, in Section 5.2, we validate the immediate integration of transfer-based unimodal attacks into multimodal learning. Next, we provide the main evaluation results compared to the state-of-the-art method in Section 5.3. In Section 5.4, we analyze the cross-task transferability between different $\mathrm{V + L}$ tasks. Finally, we present ablation studies in Section 5.5. + +# 5.1. Experimental Settings + +Datasets. We consider two widely used multimodal datasets, Flickr30K [27] and MSCOCO [23]. Flickr30K consists of 31,783 images, each with five corresponding captions. Similarly, MSCOCO comprises 123,287 images, and each image is annotated with around five captions. We adopt the Karpathy split [14] for experimental evaluation. + +
AttackALBEF*TCLCLIPViTCLIPCNN
TR R@1*IR R@1*TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1
Sep-Attack65.6973.9517.6032.9531.1745.2332.8245.49
Sep-Attack + MI58.81(↓)65.25(↓)16.02(↓)28.19(↓)23.07(↓)36.98(↓)26.56(↓)39.31(↓)
Sep-Attack + DIM56.41(↓)64.24(↓)16.75(↓)29.55(↓)24.17(↓)37.60(↓)25.54(↓)38.77(↓)
Sep-Attack + PNA.PO40.56(↓)53.95(↓)18.44(↑)30.98(↓)22.33(↓)37.02(↓)26.95(↓)38.63(↓)
Co-Attack77.1683.8615.2129.4923.6036.4825.1238.89
Co-Attack + MI64.86(↓)75.26(↓)25.40(↑)38.69(↑)24.91(↑)37.11(↑)26.31(↑)38.97(↑)
Co-Attack + DIM47.03(↓)62.28(↓)22.23(↑)35.45(↑)25.64(↑)38.50(↑)26.95(↑)40.58(↑)
SGA97.2497.2845.4255.2533.3844.1634.9346.57
+ +Table 2: Attack success rates (%) of R@1 of integrating transfer-based image attacks on image-text retrieval. The adversarial data are generated on Flickr30K using the source model ALBEF to attack other target models. * indicates white-box attacks. A higher ASR indicates better adversarial transferability. + +VLP Models. We evaluate two popular VLP models, the fused VLP and aligned VLP models. For the fused VLP, we consider ALBEF [18] and TCL [38]. ALBEF contains a 12-layer visual transformer ViT-B/16 [8] and two 6-layer transformers for the image encoder and both the text encoder and the multimodal encoder, respectively. TCL uses the same model architecture as ALBEF but with different pretrain objectives. For the aligned VLP model, we choose to evaluate CLIP [28]. CLIP has two different image encoder choices, namely, $\mathrm{CLIP}_{\mathrm{ViT}}$ and $\mathrm{CLIP}_{\mathrm{CNN}}$ , that use ViT-B/16 and ResNet-101 [12] as the base architectures for the image encoder, respectively. + +Adversarial Attack Settings. To craft adversarial images, we employ PGD [24] with perturbation bound $\epsilon_v = 2/255$ , step size $\alpha = 0.5/255$ , and iteration steps $T = 10$ . For attacking text modality, we adopt BERT-Attack [19] with perturbation bound $\epsilon_t = 1$ and length of word list $W = 10$ . Furthermore, we enlarge the image set by resizing the original image into five scales, $\{0.50, 0.75, 1.00, 1.25, 1.50\}$ , using bicubic interpolation. Similarly, the caption set is enlarged by augmenting the most matching caption pairs for each image in the dataset, with the size of approximately five. + +Metrics. We employ Attack Success Rate (ASR) as the metric for evaluating the adversarial robustness and transferability in both white-box and black-box settings. Specifically, ASR evaluates the percentage of attacks that only produce successful adversarial examples. A higher ASR indicates better adversarial transferability. + +# 5.2. Transferability Analysis + +In this paper, we present a systematic study of the adversarial transferability of VLP models, which has not been explored. As demonstrated in Section 3, existing methods, including the separate unimodal adversarial attack (SepAttack) and the multimodal adversarial attack (Co-Attack), exhibit limited transferability to other VLP models. + +To improve transferability in multimodal learning, we intuitively investigate the adoption of transfer-based attacks from unimodal learning such as image classification. Specifically, we consider MI [7], DIM [37], and PNA.PO [36]. However, this approach can be problematic if cross-modal interactions and the unique many-to-many alignments in multimodal learning are not taken into account. + +Table 2 illustrates that multimodal attack methods that incorporate transfer-based image attacks exhibit minimal improvement in transferability while compromising white-box performance. Specifically, when integrated with MI, Co-Attack drops significantly by $12.3\%$ in white-box settings, while only maintaining $25.40\%$ ASR in transferability (ALBEF to TCL). However, our SGA shows superior performance in both white-box and black-box settings. Notably, Sep-Attack combined with transfer-based attacks not only reduces the effectiveness of white-box attacks but also fails to improve adversarial transferability in almost all black-box settings. The results provide empirical evidence that directly combining unimodal adversarial attacks in multimodal learning without considering cross-modal interactions and alignments can be problematic, even when using transfer-based unimodal attacks. Additional discussion is provided in Appendix B.2. + +# 5.3. Experimental Results + +Multimodal Fusion Modules. First, we investigate VLP models with different fusion modules, namely, fused VLP models and aligned VLP models. We generate adversarial examples on both types of models and evaluate their attack performance when transferred to other VLP models while ensuring consistency in the input size of images. For example, adversarial images generated by ALBEF or TCL are resized to $224 \times 224$ before performing transfer attacks on CLIP, and adversarial examples generated on CLIP are resized to $384 \times 384$ before being transferred to ALBEF or TCL. + +As shown in Table 3, experimental results demonstrate + +
SourceAttackALBEFTCLCLIPViTCLIPCNN
TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1
ALBEFPGD52.45*58.65*3.066.798.9613.2110.3414.65
BERT-Attack11.57*27.46*12.6428.0729.3343.1732.6946.11
Sep-Attack65.69*73.95*17.6032.9531.1745.2332.8245.49
Co-Attack77.16*83.86*15.2129.4923.6036.4825.1238.89
SGA97.24±0.22*97.28±0.15*45.42±0.6055.25±0.0633.38±0.3544.16±0.2534.93±0.9946.57±0.13
TCLPGD6.1510.7877.87*79.48*7.4813.7210.3415.33
BERT-Attack11.8926.8214.54*29.17*29.6944.4933.4646.07
Sep-Attack20.1336.4884.72*86.07*31.2944.6533.3345.80
Co-Attack23.1540.0477.94*85.59*27.8541.1930.7444.11
SGA48.91±0.7460.34±0.1098.37±0.08*98.81±0.07*33.87±0.1844.88±0.5437.74±0.2748.30±0.34
CLIPViTPGD2.504.934.858.1770.92*78.61*5.368.44
BERT-Attack9.5922.6411.8025.0728.34*39.08*30.4037.43
Sep-Attack9.5923.2511.3825.6079.75*86.79*30.7839.76
Co-Attack10.5724.3311.9426.6993.25*95.86*32.5241.82
SGA13.40±0.0727.22 ±0.0616.23±0.4530.76±0.0799.08±0.08*98.94±0.00*38.76±0.2747.79±0.58
CLIPCNNPGD2.094.824.007.811.106.6086.46*92.25*
BERT-Attack8.8623.2712.3325.4827.1237.4430.40*40.10*
Sep-Attack8.5523.4112.6426.1228.3439.4391.44*95.44*
Co-Attack8.7923.7413.1026.0728.7940.0394.76*96.89*
SGA11.42±0.0724.80±0.2814.91±0.0828.82±0.1131.24±0.4242.12±0.1199.24±0.18*99.49±0.05*
+ +Table 3: Comparison with state-of-the-art methods on image-text retrieval. We report the attack success rate $(\%)$ of R@1 for both IR and TR. The source column indicates the source models used to generate the adversarial data on Flickr30K. * indicates white-box attacks. A higher ASR indicates better adversarial transferability. + +
AttackB@4METEORROUGE_LCIDErSPICE
Baseline39.731.060.0133.323.8
Co-Attack37.429.858.4125.522.8
SGA34.828.456.3116.021.4
+ +Table 4: Cross-Task Transferability: ITR $\rightarrow$ IC. Adversarial data generated from Image-Text Retrieval (ITR) to attack Image Captioning (IC) on MSCOCO. The source model and target model are ALBEF and BLIP, respectively. The Baseline represents the original performance of IC on clean data. Lower values indicate better adversarial transferability. + +the superiority of our proposed SGA over existing multimodal attack methods in all black-box settings. Specifically, our SGA achieves significant improvements in adversarial transferability when the source and target models are of the same type. For instance, SGA outperforms Co-Attack by approximately $30\%$ in terms of attack success rate when transferring adversarial data from ALBEF to TCL. Moreover, in the more challenging scenario where the source and target models are of different types, SGA also surpasses CoAttack with higher attack success rates. More experiments on the MSCOCO dataset are provided in Appendix B.3. + +Model Architectures. Then, we explore VLP models with respect to different model architectures. Many VLP models commonly use ViTs as the vision encoder, where images + +are segmented into patches before being processed by the transformer model. However, in the case of CLIP, the image encoder consists of two distinct architectures: conventional Convolutional Neural Networks (CNNs) and ViTs. The transferability between CNNs and ViTs has been well-studied in unimodal learning. Therefore, we also investigate the adversarial transferability of CNN-based and ViT-based CLIP models in multimodal learning. + +As shown in Table 3, we observe a similar phenomenon as unimodal learning that compared to CNNs, ViTs show better robustness against adversarial perturbations [25]. Specifically, for all attack methods, the same adversarial multimodal data have a stronger white-box attack effect on $\mathrm{CLIP}_{\mathrm{CNN}}$ compared to $\mathrm{CLIP}_{\mathrm{ViT}}$ . Moreover, the adversarial examples generated on $\mathrm{CLIP}_{\mathrm{ViT}}$ are found to be more transferable to $\mathrm{CLIP}_{\mathrm{CNN}}$ than vice versa (38.76% vs. 31.24%). + +Furthermore, our proposed SGA consistently improves transferability on both CNN-based CLIP and ViT-based CLIP compared to other attacks. For instance, SGA increases the adversarial transferability for $\mathrm{CLIP}_{\mathrm{ViT}}$ by $5.83\%$ and $6.24\%$ compared to Co-Attack under the white-box setting and black-box setting, respectively. + +# 5.4. Cross-Task Transferability + +Cross-modal interactions and alignments are the core components of multimodal learning regardless of the task. + +
AttackValTestATestB
Baseline58.4665.8946.25
Co-Attack54.2661.8043.81
SGA53.5561.1943.71
+ +Table 5: Cross-Task Transferability: ITR $\rightarrow$ VG. Adversarial data generated from Image-Text Retrieval (ITR) to attack Visual Grounding (VG) on RefCOCO+. The source model and target model are both ALBEF. The Baseline represents the original performance of VG on clean data. Lower values indicate better adversarial transferability. + +Therefore, we conduct extensive experiments to explore the effectiveness of our proposed SGA on two additional V+L tasks: Image Captioning (IC) and Visual Grounding (VG). + +Image Captioning. Image captioning is a generation-based task, where an input image is encoded into a feature vector and then decoded into a natural language sentence. In our experiments, we craft adversarial images using the source model (ALBEF) with an image-text retrieval objective and then directly attack the target model (BLIP [17]) on image captioning. We employ the MSCOCO dataset, which is suitable for both two tasks, and utilize various evaluation metrics to measure the quality of the generated captions, including BLEU [26], METEOR [2], ROUGE [21], CIDEr [32], and SPICE [1]. + +We present the performance on image captioning of BLIP after being attacked in Table 4. Experimental results demonstrate clear improvements in the adversarial transferability of the proposed SGA compared to Co-Attack. Specifically, our SGA improves the BLEU score by up to $2.6\%$ and the CIDEr score by up to $9.5\%$ . + +Visual Grounding. Visual grounding is another $\mathrm{V} + \mathrm{L}$ task that aims to localize the region in an image based on the corresponding specific textual description. Similarly, we generate adversarial images using the source model (ALBEF) from image-text to attack the target model (ALBEF) on visual grounding. Table 5 shows the results on RefCOCO+ [39], where our SGA still outperforms Co-Attack. + +# 5.5. Ablation Study + +To systematically investigate the impact of our set-level alignment-preserving augmentations, we conducted ablation experiments on image-text retrieval to evaluate the effect of varying the number of augmented image sets $N$ with multiscale transformation and the number $M$ of augmented caption sets. Specifically, we employed ALBEF as the source model and $\mathrm{CLIP}_{\mathrm{ViT}}$ as the target model on Flickr30K. More details are provided in Appendix B.4. + +Multi-scale Image Set. We propose the use of multiple scale-invariant images to generate diverse adversarial data in + +![](images/e089b05f0c6b102500fe1ce548125d1e3e81f5360af1128d6b7c5e4ee9bc5ff3.jpg) +Figure 5: Ablation Study: Attack success rates (\%) on set-level augmentations, Left for image set and Right for caption set. The performance of SGA is consistently improved by using larger augmented image and caption sets. + +![](images/f3d389908dc4be6587473a28e3085c3c66d1c80bffd7ef7c676abf2bc33238d1.jpg) + +SGA. Results in the left of Figure 5 reveal that the transferability significantly increases as we introduce more diverse images with different scales, peaking when the scale range is set to [0.50, 1.50] with a step of 0.25. We set the scale range $S = \{0.50, 0.75, 1.00, 1.25, 1.50\}$ for optimal performance. + +Multi-pair Caption Set. We also conduct experiments to investigate the impact of an enlarged caption set on adversarial transferability. The number of additional captions ranged from 1 to $M$ , where $M$ represents the most matching caption pairs from the dataset for each image. Results presented in the right panel of Figure 5 indicate that if $M > 1$ , the black-box performance increases significantly but eventually plateaus. These results demonstrate the effectiveness of using multiple alignment-preserving inter-modal information to enhance adversarial transferability. Furthermore, we observed that the performance is relatively insensitive to the number of extra captions, but adding more captions can improve the overall adversarial transferability. + +# 6. Conclusion + +In this paper, we make the first attempt to investigate the adversarial transferability of typical VLP models. We systematically evaluate the existing attack methods and reveal that they still exhibit lower transferability, despite their impressive performance in white-box settings. Our investigation highlights the need for specially designed transferable attacks in multimodal learning that can model the many-to-many cross-modal alignments and interactions. We propose SGA, a highly transferable multimodal attack, which leverages set-level alignment-preserving augmentations through cross-modal guidance to thoroughly exploit multimodal interactions. We hope that this work could inspire further research to evaluate and enhance the adversarial robustness of VLP models. + +Acknowledgments. This work was supported by the National Key R&D Program of China (Grant NO. 2022YFF1202903) and the National Natural Science Foundation of China (Grant NO. 62122035). + +# References + +[1] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In ECCV, 2016. 8 +[2] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In IEEvaluation@ACL, 2005. 8 +[3] Nicholas Carlini and David A. Wagner. Towards evaluating the robustness of neural networks. 2017 IEEE Symposium on Security and Privacy (SP), pages 39-57, 2017. 3 +[4] Hui Chen, Guiguang Ding, Xudong Liu, Zijia Lin, Ji Liu, and Jungong Han. Imram: Iterative matching with recurrent attention memory for cross-modal image-text retrieval. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12652-12660, 2020. 3 +[5] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 2 +[6] Mengjun Cheng, Yipeng Sun, Long Wang, Xiongwei Zhu, Kun Yao, Jie Chen, Guoli Song, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Vista: Vision and scene text aggregation for cross-modal retrieval. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5174-5183, 2022. 3 +[7] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. 3, 6 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv, abs/2010.11929, 2021. 2, 6 +[9] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuo-hang Wang, Lijuan Wang, Chenguang Zhu, Nanyun Peng, Zicheng Liu, and Michael Zeng. An empirical study of training end-to-end vision-and-language transformers. ArXiv, abs/2111.02387, 2021. 2 +[10] Ji Gao, Jack Lanchantin, Mary Lou Soffa, and Yanjun Qi. Black-box generation of adversarial text sequences to evade deep learning classifiers. 2018 IEEE Security and Privacy Workshops (SPW), pages 50-56, 2018. 3 +[11] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. CoRR, abs/1412.6572, 2015. 3 +[12] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6 +[13] Di Jin, Zhijing Jin, Joey Tianyi Zhou, and Peter Szolovits. Is bert really robust? a strong baseline for natural language attack on text classification and entailment. In AAAI, 2020. 3 +[14] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. IEEE Transactions + +on Pattern Analysis and Machine Intelligence, 39:664-676, 2017.5 +[15] Zaid Khan and Yun Raymond Fu. Exploiting bert for multimodal target sentiment classification through input space translation. Proceedings of the 29th ACM International Conference on Multimedia, 2021. 1 +[16] Chenyi Lei, Shixian Luo, Yong Liu, Wanggui He, Jiamang Wang, Guoxin Wang, Haihong Tang, Chunyan Miao, and Houqiang Li. Understanding chinese video and language via contrastive multimodal pre-training. Proceedings of the 29th ACM International Conference on Multimedia, 2021. 1 +[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. ArXiv, abs/2201.12086, 2022. 2, 8 +[18] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq R. Joty, Caiming Xiong, and Steven C. H. Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 2, 3, 6 +[19] Linyang Li, Ruotian Ma, Qipeng Guo, X. Xue, and Xipeng Qiu. Bert-attack: Adversarial attack against bert using bert. ArXiv, abs/2004.09984, 2020. 1, 2, 3, 6 +[20] Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV, 2020. 2 +[21] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Annual Meeting of the Association for Computational Linguistics, 2004. 8 +[22] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. arXiv: Learning, 2019. 2 +[23] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 5 +[24] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. ArXiv, abs/1706.06083, 2018. 1, 2, 3, 6 +[25] Muzammal Naseer, Kanchana Ranasinghe, Salman Hameed Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In Neural Information Processing Systems, 2021. 7 +[26] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 8 +[27] Bryan A. Plummer, Liwei Wang, Christopher M. Cervantes, Juan C. Caicedo, J. Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. International Journal of Computer Vision, 123:74–93, 2015. 2, 5 +[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen + +Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 6 +[29] Shuhuai Ren, Yihe Deng, Kun He, and Wanxiang Che. Generating natural language adversarial examples through probability weighted word saliency. In ACL, 2019. 3 +[30] Lei Shi, Kai Shuang, Shijie Geng, Peng Gao, Zuohui Fu, Gerard de Melo, Yunpeng Chen, and Sen Su. Dense contrastive visual-linguistic pretraining. Proceedings of the 29th ACM International Conference on Multimedia, 2021. 1 +[31] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 2 +[32] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4566-4575, 2015. 8 +[33] Teng Wang, Yixiao Ge, Feng Zheng, Ran Cheng, Ying Shan, Xiaohu Qie, and Ping Luo. Accelerating vision-language pretraining with free language modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23161-23170, 2023. 2 +[34] Teng Wang, Wenhao Jiang, Zhichao Lu, Feng Zheng, Ran Cheng, Chengguo Yin, and Ping Luo. Vlmixer: Unpaired vision-language pre-training via cross-modal cutmix. In International Conference on Machine Learning, pages 22680-22690. PMLR, 2022. 2 +[35] Zihao Wang, Xihui Liu, Hongsheng Li, Lu Sheng, Junjie Yan, Xiaogang Wang, and Jing Shao. Camp: Cross-modal adaptive message passing for text-image retrieval. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5763-5772, 2019. 3 +[36] Zhipeng Wei, Jingjing Chen, Micah Goldblum, Zuxuan Wu, Tom Goldstein, and Yu-Gang Jiang. Towards transferable adversarial attacks on vision transformers. ArXiv, abs/2109.04176, 2021. 6 +[37] Cihang Xie, Zhishuai Zhang, Jianyu Wang, Yuyin Zhou, Zhou Ren, and Alan Loddon Yuille. Improving transferability of adversarial examples with input diversity. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2725-2734, 2018. 3, 6 +[38] Jinyu Yang, Jiali Duan, S. Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul M. Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. ArXiv, abs/2202.10401, 2022. 2, 3, 6 +[39] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C. Berg, and Tamara L. Berg. Modeling context in referring expressions. ArXiv, abs/1608.00272, 2016. 8 +[40] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Francis E. H. Tay, Jiashi Feng, and Shuicheng Yan. Tokensto-token vit: Training vision transformers from scratch onImagenet. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 538-547, 2021. 2 +[41] Jiaming Zhang, Qiaomin Yi, and Jitao Sang. Towards adversarial attack on vision-language pre-training models. Proceedings of the 30th ACM International Conference on Multimedia, 2022. 1, 2, 3 + +[42] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5575-5584, 2021. 2 +[43] Qi Zhang, Zhen Lei, Zhaoxiang Zhang, and S. Li. Context-aware attention network for image-text retrieval. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3533-3542, 2020. 3 \ No newline at end of file diff --git a/2023/Set-level Guidance Attack_ Boosting Adversarial Transferability of Vision-Language Pre-training Models/images.zip b/2023/Set-level Guidance Attack_ Boosting Adversarial Transferability of Vision-Language Pre-training Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d6c402f820e6e92c1f784590107c2b67b9d1d09e --- /dev/null +++ b/2023/Set-level Guidance Attack_ Boosting Adversarial Transferability of Vision-Language Pre-training Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f05a9555a0e53c512df96081f4f709f695721feb4e797e4f6e77331330e2a0e1 +size 424749 diff --git a/2023/Set-level Guidance Attack_ Boosting Adversarial Transferability of Vision-Language Pre-training Models/layout.json b/2023/Set-level Guidance Attack_ Boosting Adversarial Transferability of Vision-Language Pre-training Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ad0de20216d678dfb648c717135e9fd33cddfd94 --- /dev/null +++ b/2023/Set-level Guidance Attack_ Boosting Adversarial Transferability of Vision-Language Pre-training Models/layout.json @@ -0,0 +1,8637 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 103, + 504, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 103, + 504, + 140 + ], + "spans": [ + { + "bbox": [ + 91, + 103, + 504, + 140 + ], + "type": "text", + "content": "Set-level Guidance Attack: Boosting Adversarial Transferability of Vision-Language Pre-training Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 160, + 515, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 160, + 515, + 176 + ], + "spans": [ + { + "bbox": [ + 77, + 160, + 515, + 176 + ], + "type": "text", + "content": "Dong Lu1, Zhiqiang Wang1, Teng Wang1,2, Weili Guan3, Hongchang Gao4, Feng Zheng1,5†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 150, + 177, + 443, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 177, + 443, + 189 + ], + "spans": [ + { + "bbox": [ + 150, + 177, + 443, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 150, + 177, + 443, + 189 + ], + "type": "text", + "content": "Southern University of Science and Technology " + }, + { + "bbox": [ + 150, + 177, + 443, + 189 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 150, + 177, + 443, + 189 + ], + "type": "text", + "content": "The University of Hong Kong" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 176, + 190, + 417, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 190, + 417, + 203 + ], + "spans": [ + { + "bbox": [ + 176, + 190, + 417, + 203 + ], + "type": "text", + "content": "3Monash University 4Temple University 5Peng Cheng Laboratory" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 205, + 478, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 205, + 478, + 217 + ], + "spans": [ + { + "bbox": [ + 113, + 205, + 478, + 217 + ], + "type": "text", + "content": "sammylu\\_outlook.com wangzq_2021\\@outlook.com tengwang@connect.hku.hk" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 220, + 462, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 462, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 462, + 232 + ], + "type": "text", + "content": "honeyguan@gmail.com hongchang.gao@temple.edu f.zheng@ieee.org" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 284, + 290, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 290, + 595 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 290, + 595 + ], + "type": "text", + "content": "Vision-language pre-training (VLP) models have shown vulnerability to adversarial examples in multimodal tasks. Furthermore, malicious adversaries can be deliberately transferred to attack other black-box models. However, existing work has mainly focused on investigating white-box attacks. In this paper, we present the first study to investigate the adversarial transferability of recent VLP models. We observe that existing methods exhibit much lower transferability, compared to the strong attack performance in white-box settings. The transferability degradation is partly caused by the under-utilization of cross-modal interactions. Particularly, unlike unimodal learning, VLP models rely heavily on cross-modal interactions and the multimodal alignments are many-to-many, e.g., an image can be described in various natural languages. To this end, we propose a highly transferable Set-level Guidance Attack (SGA) that thoroughly leverages modality interactions and incorporates alignment-preserving augmentation with cross-modal guidance. Experimental results demonstrate that SGA could generate adversarial examples that can strongly transfer across different VLP models on multiple downstream vision-language tasks. On image-text retrieval, SGA significantly enhances the attack success rate for transfer attacks from ALBEF to TCL by a large margin (at least " + }, + { + "bbox": [ + 46, + 284, + 290, + 595 + ], + "type": "inline_equation", + "content": "9.78\\%" + }, + { + "bbox": [ + 46, + 284, + 290, + 595 + ], + "type": "text", + "content": " and up to " + }, + { + "bbox": [ + 46, + 284, + 290, + 595 + ], + "type": "inline_equation", + "content": "30.21\\%" + }, + { + "bbox": [ + 46, + 284, + 290, + 595 + ], + "type": "text", + "content": "), compared to the state-of-the-art. Our code is available at https://github.com/Zoky-2020/SGA." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 636, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 636, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 636, + 288, + 696 + ], + "type": "text", + "content": "Recent work has shown that vision-language pre-training (VLP) models are still vulnerable to adversarial examples [41], even though they have achieved remarkable performance on a wide range of multimodal tasks [30, 15, 16]. Existing work mainly focuses on white-box attacks, where" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 259, + 426, + 344 + ], + "blocks": [ + { + "bbox": [ + 310, + 259, + 426, + 344 + ], + "lines": [ + { + "bbox": [ + 310, + 259, + 426, + 344 + ], + "spans": [ + { + "bbox": [ + 310, + 259, + 426, + 344 + ], + "type": "image", + "image_path": "57abe961e11c257578b9d684d65e3eb931de0acc13b89bc92fa04105a5eba159.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 344, + 410, + 354 + ], + "lines": [ + { + "bbox": [ + 345, + 344, + 410, + 354 + ], + "spans": [ + { + "bbox": [ + 345, + 344, + 410, + 354 + ], + "type": "text", + "content": "(a) ALBEF to ALBEF" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 426, + 258, + 541, + 344 + ], + "blocks": [ + { + "bbox": [ + 426, + 258, + 541, + 344 + ], + "lines": [ + { + "bbox": [ + 426, + 258, + 541, + 344 + ], + "spans": [ + { + "bbox": [ + 426, + 258, + 541, + 344 + ], + "type": "image", + "image_path": "703a6a22fcd62c73203da94b438e243c35a7028ca3d55fc818e0500d0ab6f195.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 464, + 344, + 521, + 354 + ], + "lines": [ + { + "bbox": [ + 464, + 344, + 521, + 354 + ], + "spans": [ + { + "bbox": [ + 464, + 344, + 521, + 354 + ], + "type": "text", + "content": "(b) ALBEF to TCL" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 310, + 356, + 425, + 401 + ], + "blocks": [ + { + "bbox": [ + 310, + 356, + 425, + 401 + ], + "lines": [ + { + "bbox": [ + 310, + 356, + 425, + 401 + ], + "spans": [ + { + "bbox": [ + 310, + 356, + 425, + 401 + ], + "type": "image", + "image_path": "79b7480f3214c69978d30e31bb2733b3dfefb4eca3b2d5596566f43489447b39.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 402, + 411, + 412 + ], + "lines": [ + { + "bbox": [ + 345, + 402, + 411, + 412 + ], + "spans": [ + { + "bbox": [ + 345, + 402, + 411, + 412 + ], + "type": "text", + "content": "(c) ALBEF to " + }, + { + "bbox": [ + 345, + 402, + 411, + 412 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{VIT}}" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 420, + 548, + 601 + ], + "lines": [ + { + "bbox": [ + 304, + 420, + 548, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 548, + 601 + ], + "type": "text", + "content": "Figure 1: Comparison of attack success rates (ASR) using five different attacks on image-text retrieval. Adversarial examples are crafted on the source model (ALBEF) to attack the target white-box model or black-box models. The first three columns refer to the image-only PGD attack [24], text-only BERT-Attack [19] (BA), and the combined separate unimodal attack (SA), which all belong to the methods without cross-modal interactions. The fourth column is the state-of-the-art multimodal Co-Attack [41] (CA) that employs single-pair cross-modal interactions. The last column is the proposed Set-level Guidance Attack (SGA), which leverages multiple set-level cross-modal interactions, successfully attacking the white-box model and transferring to attack all black-box models with the highest ASR. More discussions are in Section 3." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 427, + 357, + 542, + 401 + ], + "blocks": [ + { + "bbox": [ + 427, + 357, + 542, + 401 + ], + "lines": [ + { + "bbox": [ + 427, + 357, + 542, + 401 + ], + "spans": [ + { + "bbox": [ + 427, + 357, + 542, + 401 + ], + "type": "image", + "image_path": "cab211f7941f67f431d4ce6995fa79556eb3bb5696d8b8da33990891b2e141a3.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 402, + 526, + 411 + ], + "lines": [ + { + "bbox": [ + 460, + 402, + 526, + 411 + ], + "spans": [ + { + "bbox": [ + 460, + 402, + 526, + 411 + ], + "type": "text", + "content": "(d) ALBEF to " + }, + { + "bbox": [ + 460, + 402, + 526, + 411 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{CNN}}" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 616, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 547, + 700 + ], + "type": "text", + "content": "information about the victim model is accessible. However, the transferability of adversarial examples across VLP models has not been investigated, which is a more practical setting. It is still unknown whether the adversarial data generated on the source model can successfully attack another model, which poses a serious security risk to the deployment of VLP models in real-world applications." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 318, + 701, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 547, + 714 + ], + "type": "text", + "content": "This paper makes the first step to investigate the trans" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 208, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 208, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 208, + 713 + ], + "type": "text", + "content": "* Equal contribution. † Corresponding author." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "102" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 81, + 148 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 81, + 148 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 81, + 148 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 81, + 148 + ], + "type": "image", + "image_path": "092ee924aedd1ff0cfc95c7e445420276c5b21a939008f621bb56c95c838490b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 150, + 109, + 160 + ], + "lines": [ + { + "bbox": [ + 53, + 150, + 109, + 160 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 109, + 160 + ], + "type": "text", + "content": "(a) Sep-Attack" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 82, + 72, + 111, + 148 + ], + "blocks": [ + { + "bbox": [ + 82, + 72, + 111, + 148 + ], + "lines": [ + { + "bbox": [ + 82, + 72, + 111, + 148 + ], + "spans": [ + { + "bbox": [ + 82, + 72, + 111, + 148 + ], + "type": "image", + "image_path": "a59f5fb7f155a9f918d1474c0d37ac11b00d845839703e7bc2b47746618723ee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 116, + 72, + 183, + 148 + ], + "blocks": [ + { + "bbox": [ + 116, + 72, + 183, + 148 + ], + "lines": [ + { + "bbox": [ + 116, + 72, + 183, + 148 + ], + "spans": [ + { + "bbox": [ + 116, + 72, + 183, + 148 + ], + "type": "image", + "image_path": "3102c8a89a59e8d2a501f6e7ed757080b89fbaefed782e5e8ad9d6a13833fd86.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 150, + 175, + 159 + ], + "lines": [ + { + "bbox": [ + 124, + 150, + 175, + 159 + ], + "spans": [ + { + "bbox": [ + 124, + 150, + 175, + 159 + ], + "type": "text", + "content": "(b) Co-Attack" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 189, + 72, + 284, + 148 + ], + "blocks": [ + { + "bbox": [ + 189, + 72, + 284, + 148 + ], + "lines": [ + { + "bbox": [ + 189, + 72, + 284, + 148 + ], + "spans": [ + { + "bbox": [ + 189, + 72, + 284, + 148 + ], + "type": "image", + "image_path": "cae2c94049f04114f405fa8b9b227919dd7078f15ef9c244058a2e99349902f6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 150, + 251, + 160 + ], + "lines": [ + { + "bbox": [ + 221, + 150, + 251, + 160 + ], + "spans": [ + { + "bbox": [ + 221, + 150, + 251, + 160 + ], + "type": "text", + "content": "(c) SGA" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 168, + 289, + 324 + ], + "lines": [ + { + "bbox": [ + 46, + 168, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 289, + 324 + ], + "type": "text", + "content": "Figure 2: Comparison of cross-modal interactions. To generate adversarial examples, existing methods either: (a) separately perturb unimodal data without any cross-modal interactions (Sep-Attack) or, (b) perturb multiple modalities but with single image-text pairs to model cross-modal interactions (Co-Attack). However, our method is capable of learning cross-modal interactions among multiple alignments through (c) set-level guidance (SGA). Specifically, " + }, + { + "bbox": [ + 46, + 168, + 289, + 324 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 168, + 289, + 324 + ], + "type": "text", + "content": " denotes the input image, and " + }, + { + "bbox": [ + 46, + 168, + 289, + 324 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 168, + 289, + 324 + ], + "type": "text", + "content": " is the paired caption. " + }, + { + "bbox": [ + 46, + 168, + 289, + 324 + ], + "type": "inline_equation", + "content": "v'" + }, + { + "bbox": [ + 46, + 168, + 289, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 168, + 289, + 324 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 46, + 168, + 289, + 324 + ], + "type": "text", + "content": " represent the corresponding adversarial examples, respectively. " + }, + { + "bbox": [ + 46, + 168, + 289, + 324 + ], + "type": "inline_equation", + "content": "U(v)" + }, + { + "bbox": [ + 46, + 168, + 289, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 168, + 289, + 324 + ], + "type": "inline_equation", + "content": "U(t)" + }, + { + "bbox": [ + 46, + 168, + 289, + 324 + ], + "type": "text", + "content": " represent the scale-invariant image set and most matching caption set. Arrows indicate the guidance for generating adversarial examples." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 339, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 483 + ], + "type": "text", + "content": "ferability of adversarial samples within VLP models. Without loss of generality, most of our experiments are based on image-text retrieval tasks. We first empirically evaluate this attack performance with respect to different modalities on multimodal tasks across multiple datasets. Our results show that the adversarial transferability of attacking both modalities (image & text) consistently beats attacking unimodal data (image or text). Unfortunately, even though two modalities are allowed to be perturbed simultaneously, the attack success rates of existing methods [24, 19, 41] still significantly drops when transferring from the white-box to black-box settings, as shown in Figure 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 484, + 288, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 288, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 288, + 675 + ], + "type": "text", + "content": "Different from recent studies focusing on separate attacks on unimodal data [24, 19], multimodal pairs exhibit intrinsic alignment and complementarity to each other. The modeling of inter-modal correspondence turns out to be a critical problem for transferability. Considering that the alignments between image and text are many-to-many, for example, an image could be described to be with various human perspectives and language styles, a reasonable perturbation direction may be determined with diverse guidance from multiple labels in the other modality. However, recent adversarial attack methods for VLP models [41] usually employ a single image-text pair to generate adversarial samples. Although they exhibit strong performance in white-box settings, the poor diversity of guidance makes adversarial samples highly correlated with the alignment pattern of the white-box model, and therefore impedes generalization to black-box settings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "To address the weak transferability problem, we propose Set-level Guidance Attack (SGA), which leverages diverse cross-modal interactions among multiple image-text pairs" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": "(Figure 2). Specifically, we introduce alignment-preserving augmentation which enriches image-text pairs while keeping their alignments intact. The image augmentation is based on the scale-invariant property of deep learning models [22], thus we can construct multi-scale images to increase the diversity. For text augmentation, we select the most matching caption pairs from the dataset. More importantly, SGA generates adversarial examples on multimodal augmented input data with carefully designed cross-modal guidance. In detail, SGA iteratively pushes supplemental information away between two modalities with another modality as supervision to disrupt the interactions for better harmonious perturbations. Note that resultant adversarial samples could perceive the gradients originated from multiple guidance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 240, + 547, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 547, + 396 + ], + "type": "text", + "content": "We conduct experiments on two well-established multimodal datasets, Flickr30K [27] and MSCOCO [23], to evaluate the performance of our proposed SGA across various Vision-and-Language " + }, + { + "bbox": [ + 304, + 240, + 547, + 396 + ], + "type": "inline_equation", + "content": "(\\mathrm{V} + \\mathrm{L})" + }, + { + "bbox": [ + 304, + 240, + 547, + 396 + ], + "type": "text", + "content": " downstream tasks. The experimental results demonstrate the high effectiveness of SGA in generating adversarial examples that can be strongly transferred across VLP models, surpassing the current state-of-the-art attack methods in multimodal learning. In particular, SGA achieves notable improvements in image-text retrieval under black-box settings and also exhibits superior performance in white-box attack settings. Moreover, SGA also outperforms the state-of-the-art methods in image captioning and yields higher fooling rates on visual grounding." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 397, + 547, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 547, + 515 + ], + "type": "text", + "content": "We summarize our contributions as follows. 1) We make the first attempt to explore the transferability of adversarial examples on popular VLP models with a systematical evaluation; 2) We provide SGA, a novel transferable multimodal attack that enhances adversarial transferability through the effective use of set-level alignment-preserving augmentations and well-designed cross-modal guidance; 3) Extensive experiments show that SGA consistently boosts adversarial transferability across different VLP models than the state-of-the-art methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 529, + 392, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 392, + 542 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 392, + 542 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 549, + 505, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 549, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 549, + 505, + 562 + ], + "type": "text", + "content": "2.1. Vision-Language Pre-training Models" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "text", + "content": "Vision-Language Pre-training (VLP) aims to improve the performance of downstream multimodal tasks by pretraining large-scale image-to-text pairs [17]. Most works are developed upon the pre-trained object detectors with region features to learn the vision-language representations [5, 20, 42, 34]. Recently, with the increasing popularity of Vision Transformer (ViT) [8, 31, 40], some other works propose to use ViT as an image encoder and transform the input into patches in an end-to-end manner [18, 38, 17, 9, 33]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "According to the VLP architectures, VLP models can be classified into two typical types: fused VLP models and aligned VLP models [41]. Specifically, fused VLP models" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "103" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "(e.g., ALBEF [18], TCL [38]) first utilize separate unimodal encoders to process token embeddings and visual features, and further use a multimodal encoder to process image and text embeddings to output fused multimodal embeddings. Alternatively, aligned VLP models (e.g., CLIP [28]) have only unimodal encoders with independent image and text modality embeddings. In this paper, we focus on popular architectures with fused and aligned VLP models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 177, + 194, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 177, + 194, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 177, + 194, + 190 + ], + "type": "text", + "content": "2.2. Image-Text Retrieval Task" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 196, + 288, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 196, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 288, + 256 + ], + "type": "text", + "content": "Image-Text Retrieval (ITR) aims to retrieve the relevant top-ranked instances from a gallery database with one modality, given an input query from another modality [35, 4, 43, 6]. This task can be divided into two subtasks, image-to-text retrieval (TR) and text-to-image retrieval (IR)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 256, + 287, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 287, + 353 + ], + "type": "text", + "content": "For ALBEF [18] and TCL [38], the semantic similarity score in the unimodal embedding space will be calculated for all image-text pairs to select top-" + }, + { + "bbox": [ + 46, + 256, + 287, + 353 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 256, + 287, + 353 + ], + "type": "text", + "content": " candidates. Then the multimodal encoder takes the top-" + }, + { + "bbox": [ + 46, + 256, + 287, + 353 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 256, + 287, + 353 + ], + "type": "text", + "content": " candidates and computes the image-text matching score for ranking. For CLIP [28], without the multimodal encoder, the final rank list can be obtained based on the similarity in the embedding space between image and text modalities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 361, + 198, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 198, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 198, + 374 + ], + "type": "text", + "content": "2.3. Adversarial Transferability" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 288, + 523 + ], + "type": "text", + "content": "Existing adversarial attacks can be categorized into two settings: white-box attacks and black-box attacks. In a white-box setting, the target model is fully accessible, but not in a black-box setting. In computer vision, many methods employ gradient information for adversarial attacks in white-box settings, such as FGSM [11], PGD [24], C&W [3], and MI [7]. In contrast, in the field of natural language processing (NLP), current attack methods mainly modify or replace some tokens of the input text [19, 29, 10, 13]. In the multimodal vision-language domain, Zhang et al. [41] proposed a white-box multimodal attack method with respect to popular VLP models on downstream tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 525, + 288, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 288, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 288, + 608 + ], + "type": "text", + "content": "However, white-box attacks are unrealistic due to the inaccessibility of model information in practical applications. In addition, there is no related work that systematically analyzes the adversarial transferability of multimodal attack methods on VLP models. Therefore, in this work, we mainly focus on generating highly transferable adversaries across different VLP models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 620, + 261, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 261, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 261, + 635 + ], + "type": "text", + "content": "3. Analysis of Adversarial Transferability" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "In this section, we conduct an empirical study on VLP models to evaluate adversarial transferability using existing methods. A common approach to attack multimodal tasks is combining unimodal adversarial attacks [24, 7, 37, 19] of each modalities together. For instance, the separate unimodal attack (Sep-Attack) includes PGD [24] and BERT-Attack" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 72, + 539, + 167 + ], + "blocks": [ + { + "bbox": [ + 313, + 72, + 539, + 167 + ], + "lines": [ + { + "bbox": [ + 313, + 72, + 539, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 539, + 167 + ], + "type": "image", + "image_path": "31377960b66f9787e47cdfd1586b6c6be547d5ff88b9faa3601d3e36623fef34.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 175, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 304, + 175, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 547, + 236 + ], + "type": "text", + "content": "Figure 3: Attack success rates " + }, + { + "bbox": [ + 304, + 175, + 547, + 236 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 304, + 175, + 547, + 236 + ], + "type": "text", + "content": " of different perturbed modalities on image-text retrieval. Adversarial examples with respect to different modalities from the source model to attack the target model using Sep-Attack. Different colors represent different perturbed modalities." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 313, + 247, + 539, + 345 + ], + "blocks": [ + { + "bbox": [ + 313, + 247, + 539, + 345 + ], + "lines": [ + { + "bbox": [ + 313, + 247, + 539, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 539, + 345 + ], + "type": "image", + "image_path": "c71c566bd1ede3d6e42a054d67a7eb090e7241a954e637d505f0a839317aa523.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 354, + 546, + 401 + ], + "lines": [ + { + "bbox": [ + 304, + 354, + 546, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 546, + 401 + ], + "type": "text", + "content": "Figure 4: Attack success rates " + }, + { + "bbox": [ + 304, + 354, + 546, + 401 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 304, + 354, + 546, + 401 + ], + "type": "text", + "content": " on white-box models and black-box models. Both images and text are perturbed by Sep-Attack (Left) and Co-Attack (Right) on image-text retrieval. Colors indicate different target models." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 414, + 547, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 547, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 547, + 461 + ], + "type": "text", + "content": "[19] for attacking image modality and text modality, respectively. Other recent multimodal adversarial attacks, such as Co-Attack [41], consider cross-modal interactions by perturbing image modalities and text modalities collectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 461, + 547, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 547, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 547, + 534 + ], + "type": "text", + "content": "We first present the observations regarding the adversarial transferability of VLP models. Then we discuss the limitations of the existing methods. By conducting this study, we aim to provide insights into the robustness of VLP models against adversarial attacks and the effectiveness of different attack strategies." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 542, + 390, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 542, + 390, + 554 + ], + "spans": [ + { + "bbox": [ + 306, + 542, + 390, + 554 + ], + "type": "text", + "content": "3.1. Observations" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 561, + 547, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 561, + 547, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 561, + 547, + 644 + ], + "type": "text", + "content": "To investigate the adversarial transferability of perturbed inputs with respect to different modalities (i.e., image, text, and image & text) and the effect of different VLP models on transferability, we conduct experiments and present the attack success rates of the adversarial examples generated by the source model to attack the target models in Figure 3 and Figure 4. The observations are summarized below:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 653, + 547, + 713 + ], + "type": "text", + "content": "- The adversarial transferability of attacking both modalities (image & text) is consistently more effective than attacking any unimodal data alone (image or text). As shown in Figure 3, transferring both adversarial image and text from ALBEF to TCL leads to a much higher" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "104" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 66, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 289, + 180 + ], + "type": "text", + "content": "attack success rate than transferring adversarial examples of any single modality. Notably, ALBEF and TCL are both fused VLP models. Similar observations also exist in the following settings: (1) The source and target models are different types of VLP models but have the same basic architectures (e.g., TCL and " + }, + { + "bbox": [ + 66, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{ViT}}" + }, + { + "bbox": [ + 66, + 72, + 289, + 180 + ], + "type": "text", + "content": "). (2) The source and target models are the same types of VLP models, but with different basic architectures (e.g., " + }, + { + "bbox": [ + 66, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{ViT}}" + }, + { + "bbox": [ + 66, + 72, + 289, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 66, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{CNN}}" + }, + { + "bbox": [ + 66, + 72, + 289, + 180 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 190, + 289, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 190, + 289, + 299 + ], + "spans": [ + { + "bbox": [ + 59, + 190, + 289, + 299 + ], + "type": "text", + "content": "- Adversarial multimodal data (i.e., adversarial image & text), which have strong attack performance on the source model, can hardly maintain the same capability when transferring to target models. For example, as illustrated in Figure 4, even though ALBEF and TCL have the same model architecture, the attack success rate sees a significant drop when transferring adversarial examples generated on ALBEF to TCL. The phenomenon exists in both Sep-Attack and Co-Attack." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 308, + 289, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 289, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 289, + 392 + ], + "type": "text", + "content": "In summary, the attack methods can have stronger transferability in black-box settings if all the modalities are attacked simultaneously. However, even though two modalities are allowed to be perturbed, existing methods still exhibit much lower transferability. This suggests that attacks with higher transferability should be specifically designed, instead of directly utilizing the existing white-box attack methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 400, + 124, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 400, + 124, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 400, + 124, + 411 + ], + "type": "text", + "content": "3.2. Discussions" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 418, + 288, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 288, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 288, + 455 + ], + "type": "text", + "content": "We posit that the degradation in transferability of adversarial examples is mainly due to the limitations of the existing attack methods:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 464, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 59, + 464, + 289, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 464, + 289, + 620 + ], + "spans": [ + { + "bbox": [ + 59, + 464, + 289, + 620 + ], + "type": "text", + "content": "- One major limitation of Sep-Attack is that it does not take into account the interactions between different modalities. As a combined independent attack method for each modality, it fails to model the inter-modal correspondence that is crucial for successful attacks in multimodal learning. This is particularly evident in multimodal tasks such as image-text retrieval, where the ground truth is not discrete labels (e.g., image classification) but another modality data that corresponds to the input modality. The complete lack of cross-modal interactions in Sep-Attack severely limits the generalization of adversarial examples and reduces their transferability among different VLP models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 630, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 630, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 630, + 289, + 714 + ], + "type": "text", + "content": "- While Co-Attack is designed to leverage the collaboration between modalities to generate adversarial examples, it still suffers from a key drawback that hinders its transferability to other VLP models. Unlike unimodal learning, multimodal learning involves multiple complementary modalities with many-to-many cross-modal alignments, which pose unique challenges to" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 325, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 547, + 180 + ], + "type": "text", + "content": "achieving sufficient adversarial transferability. However, Co-Attack only uses single image-text pairs to generate adversarial data, limiting the diversity of guidance from multiple labels in other modalities. This lack of diversity in cross-modal guidance makes adversarial samples highly correlated with the alignment pattern of the white-box model. Therefore, the generality of adversarial examples is restricted, and their effectiveness in transferring to other models drops." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 188, + 547, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 188, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 304, + 188, + 547, + 248 + ], + "type": "text", + "content": "In conclusion, the analysis motivates our investigation into the adversarial transferability of VLP models. Moreover, it highlights the pressing need to explore transferable multimodal attacks for generating adversarial examples that can be effectively transferred across different VLP models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 258, + 389, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 258, + 389, + 272 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 389, + 272 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 278, + 548, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 548, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 548, + 374 + ], + "type": "text", + "content": "In this section, we propose a transferable multimodal adversarial attack, termed Set-level Guidance Attack (SGA). SGA is designed to enhance adversarial transferability across VLP models by leveraging multimodal interactions and incorporating diverse alignment-preserving inter-modal information with carefully designed cross-modal guidance. We provide our motivation first, followed by relevant notations, and finally present SGA in detail." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 382, + 380, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 382, + 380, + 393 + ], + "spans": [ + { + "bbox": [ + 306, + 382, + 380, + 393 + ], + "type": "text", + "content": "4.1. Motivation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 400, + 547, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 547, + 532 + ], + "type": "text", + "content": "To improve the transferability of multimodal attacks, we first conduct an investigation into the shortcomings of existing methods in black-box settings. Through a systematic analysis of failure cases, we observe that around half of these cases arise due to the presence of multiple matched captions of the image, as shown in Table 1. More specifically, our findings indicate that while the generated adversarial image may be significantly distant from the single supervised caption in the source model, it is prone to approaching other matched captions in the target model, where the alignments can be modeled and ultimately lead to attack failure." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 532, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 547, + 628 + ], + "type": "text", + "content": "Therefore, to maintain the attack ability of adversarial images when transferring to other models, it is crucial to consider multiple paired captions and push the adversarial image away from all the paired captions, thus preserving the attack ability when transferred to other black-box models. Crafting adversarial text for high transferability follows a similar approach, which can also benefit from more paired images. More discussions can be found in Appendix A." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 635, + 374, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 374, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 374, + 647 + ], + "type": "text", + "content": "4.2. Notations" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "inline_equation", + "content": "(v,t)" + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "content": " denote an image-text pair sampled from a multimodal dataset " + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "content": ". For VLP models, we denote " + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "inline_equation", + "content": "f_{I}" + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "content": " as the image encoder and " + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "content": " as the text encoder. The multimodal fusion module in fused VLP models is denoted by " + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "inline_equation", + "content": "f_{M}" + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "inline_equation", + "content": "f_{I}(v)" + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "content": " represents the image representation " + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "inline_equation", + "content": "e_v" + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "content": " encoded" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "105" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 286, + 135 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 286, + 135 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 286, + 135 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 286, + 135 + ], + "type": "table", + "html": "
AttackCross-modal Interactionp (Event A) ↑p (Event B | Event A) ↓
ALBEFTCLCLIPViTCLIPCNN
Sep-AttackNo74.60%46.78%41.69%41.82%
Co-AttackSingle-pair80.60%50.50%40.82%42.93%
SGASet-level97.20%28.91%34.67%38.58%
", + "image_path": "8dfbf039b2399d2749c5cc21e35f6729686d410c65b34667f95708c29724208b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 288, + 217 + ], + "lines": [ + { + "bbox": [ + 46, + 144, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 288, + 217 + ], + "type": "text", + "content": "Table 1: Adversarial images with insufficient cross-modal interactions may get closer to other paired captions when transferring. Event A: success cases in white-box settings. Event B: failure cases in black-box settings influenced by other paired captions. Adversarial data are generated on Flickr30K from ALBEF to attack other target models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 227, + 287, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 227, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 287, + 275 + ], + "type": "text", + "content": "by " + }, + { + "bbox": [ + 46, + 227, + 287, + 275 + ], + "type": "inline_equation", + "content": "f_{I}" + }, + { + "bbox": [ + 46, + 227, + 287, + 275 + ], + "type": "text", + "content": " taking an image " + }, + { + "bbox": [ + 46, + 227, + 287, + 275 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 227, + 287, + 275 + ], + "type": "text", + "content": " as input, " + }, + { + "bbox": [ + 46, + 227, + 287, + 275 + ], + "type": "inline_equation", + "content": "f_{T}(t)" + }, + { + "bbox": [ + 46, + 227, + 287, + 275 + ], + "type": "text", + "content": " denotes the text representation " + }, + { + "bbox": [ + 46, + 227, + 287, + 275 + ], + "type": "inline_equation", + "content": "e_t" + }, + { + "bbox": [ + 46, + 227, + 287, + 275 + ], + "type": "text", + "content": " encoded by " + }, + { + "bbox": [ + 46, + 227, + 287, + 275 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 46, + 227, + 287, + 275 + ], + "type": "text", + "content": " taking a text " + }, + { + "bbox": [ + 46, + 227, + 287, + 275 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 227, + 287, + 275 + ], + "type": "text", + "content": " as input, and " + }, + { + "bbox": [ + 46, + 227, + 287, + 275 + ], + "type": "inline_equation", + "content": "f_{M}(e_v, e_t)" + }, + { + "bbox": [ + 46, + 227, + 287, + 275 + ], + "type": "text", + "content": " denotes the multimodal representation encoded by " + }, + { + "bbox": [ + 46, + 227, + 287, + 275 + ], + "type": "inline_equation", + "content": "f_{M}" + }, + { + "bbox": [ + 46, + 227, + 287, + 275 + ], + "type": "text", + "content": " taking image and text representations as inputs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 275, + 288, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 288, + 335 + ], + "type": "text", + "content": "We use " + }, + { + "bbox": [ + 46, + 275, + 288, + 335 + ], + "type": "inline_equation", + "content": "B[v, \\epsilon_v]" + }, + { + "bbox": [ + 46, + 275, + 288, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 275, + 288, + 335 + ], + "type": "inline_equation", + "content": "B[t, \\epsilon_t]" + }, + { + "bbox": [ + 46, + 275, + 288, + 335 + ], + "type": "text", + "content": " to represent the legal searching spaces for optimizing adversarial image and text, respectively. Specifically, " + }, + { + "bbox": [ + 46, + 275, + 288, + 335 + ], + "type": "inline_equation", + "content": "\\epsilon_v" + }, + { + "bbox": [ + 46, + 275, + 288, + 335 + ], + "type": "text", + "content": " denotes the maximal perturbation bound for the image, and " + }, + { + "bbox": [ + 46, + 275, + 288, + 335 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 46, + 275, + 288, + 335 + ], + "type": "text", + "content": " denotes the maximal number of changeable words in the caption." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 342, + 258, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 258, + 355 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 258, + 355 + ], + "type": "text", + "content": "4.3. Transferable Set-Level Guidance Attack" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 361, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 445 + ], + "type": "text", + "content": "Alignment-preserving Augmentation. The analysis presented in Section 3 highlights the key limitation of existing methods: the inter-modal information used to generate adversarial examples lacks diversity. The limitation will make the generated adversarial examples fail to generalize to other black-box models with strong attack performance, resulting in limited adversarial transferability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 445, + 288, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 288, + 613 + ], + "type": "text", + "content": "To inject more diversity in the generation of generalizable adversarial examples, we propose using set-level alignment-preserving augmentation to expand multimodal input spaces while maintaining cross-modal alignments intact. Unlike previous methods that only consider a single image-text paired example " + }, + { + "bbox": [ + 46, + 445, + 288, + 613 + ], + "type": "inline_equation", + "content": "(v,t)" + }, + { + "bbox": [ + 46, + 445, + 288, + 613 + ], + "type": "text", + "content": " to generate adversarial data, we enlarge the input to a set level of images and captions. Specifically, we select the most matching caption pairs from the dataset of each image " + }, + { + "bbox": [ + 46, + 445, + 288, + 613 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 445, + 288, + 613 + ], + "type": "text", + "content": " to form an augmented caption set " + }, + { + "bbox": [ + 46, + 445, + 288, + 613 + ], + "type": "inline_equation", + "content": "\\pmb{t} = \\{t_1,t_2,\\dots,t_M\\}" + }, + { + "bbox": [ + 46, + 445, + 288, + 613 + ], + "type": "text", + "content": ", and resize each image " + }, + { + "bbox": [ + 46, + 445, + 288, + 613 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 445, + 288, + 613 + ], + "type": "text", + "content": " into different scales " + }, + { + "bbox": [ + 46, + 445, + 288, + 613 + ], + "type": "inline_equation", + "content": "S = \\{s_{1},s_{2},\\dots,s_{N}\\}" + }, + { + "bbox": [ + 46, + 445, + 288, + 613 + ], + "type": "text", + "content": " and then add Gaussian noise to obtain a multi-scale image set " + }, + { + "bbox": [ + 46, + 445, + 288, + 613 + ], + "type": "inline_equation", + "content": "\\pmb{v} = \\{v_{1},v_{2},\\dots,v_{N}\\}" + }, + { + "bbox": [ + 46, + 445, + 288, + 613 + ], + "type": "text", + "content": " based on the scale-invariant property. The enlarged input set " + }, + { + "bbox": [ + 46, + 445, + 288, + 613 + ], + "type": "inline_equation", + "content": "(\\pmb{v},\\pmb{t})" + }, + { + "bbox": [ + 46, + 445, + 288, + 613 + ], + "type": "text", + "content": " is then used to generate the adversarial data " + }, + { + "bbox": [ + 46, + 445, + 288, + 613 + ], + "type": "inline_equation", + "content": "(v^{\\prime},t^{\\prime})" + }, + { + "bbox": [ + 46, + 445, + 288, + 613 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 689 + ], + "type": "text", + "content": "Cross-modal Guidance. Cross-modal interactions play a crucial role in multimodal tasks. For example, in imagetext retrieval, the paired information from another modality provides unique annotation supervision for each sample. Similarly, in adversarial attacks, supervisory information is essential in guiding the search for adversarial examples." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "content": "To fully utilize the enlarged alignment-preserving multimodal input set " + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "(\\pmb{v},\\pmb{t})" + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "content": " and further improve the transfer" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "content": "ability of the generated adversarial data, we propose cross-modal guidance to utilize interactions from different modalities. Specifically, we use the paired information from another modality as the supervision to guide the direction of optimizing the adversarial data. This guidance iteratively pushed away the multimodal information and disrupt the cross-modal interaction for better harmonious perturbations. Notably, the resultant adversarial examples can perceive the gradients originated from multiple guidance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 180, + 547, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 547, + 217 + ], + "type": "text", + "content": "First, we generate corresponding adversarial captions for all captions in the text set " + }, + { + "bbox": [ + 304, + 180, + 547, + 217 + ], + "type": "inline_equation", + "content": "\\pmb{t}" + }, + { + "bbox": [ + 304, + 180, + 547, + 217 + ], + "type": "text", + "content": ", forming an adversarial caption set " + }, + { + "bbox": [ + 304, + 180, + 547, + 217 + ], + "type": "inline_equation", + "content": "\\pmb{t}' = \\{t_1', t_2', \\dots, t_M'\\}" + }, + { + "bbox": [ + 304, + 180, + 547, + 217 + ], + "type": "text", + "content": ". The process can be formulated as," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 353, + 222, + 546, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 222, + 546, + 251 + ], + "spans": [ + { + "bbox": [ + 353, + 222, + 546, + 251 + ], + "type": "interline_equation", + "content": "t _ {i} ^ {\\prime} = \\underset {t _ {i} ^ {\\prime} \\in B [ t _ {i}, \\epsilon_ {t} ]} {\\arg \\max } - \\frac {f _ {T} \\left(t _ {i} ^ {\\prime}\\right) \\cdot f _ {I} (v)}{\\| f _ {T} \\left(t _ {i} ^ {\\prime}\\right) \\| \\| f _ {I} (v) \\|.} \\tag {1}", + "image_path": "bcf787d2dc4e996e970da246d058661a0f305969429a5887a6603b082030e715.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 257, + 546, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 546, + 294 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 546, + 294 + ], + "type": "text", + "content": "The adversarial caption " + }, + { + "bbox": [ + 304, + 257, + 546, + 294 + ], + "type": "inline_equation", + "content": "t_i'" + }, + { + "bbox": [ + 304, + 257, + 546, + 294 + ], + "type": "text", + "content": " is constrained to be dissimilar to the original image " + }, + { + "bbox": [ + 304, + 257, + 546, + 294 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 257, + 546, + 294 + ], + "type": "text", + "content": " in the embedding space. Next, the adversarial image " + }, + { + "bbox": [ + 304, + 257, + 546, + 294 + ], + "type": "inline_equation", + "content": "v'" + }, + { + "bbox": [ + 304, + 257, + 546, + 294 + ], + "type": "text", + "content": " is generated by solving" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 299, + 546, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 299, + 546, + 329 + ], + "spans": [ + { + "bbox": [ + 325, + 299, + 546, + 329 + ], + "type": "interline_equation", + "content": "v ^ {\\prime} = \\underset {v ^ {\\prime} \\in B [ v, \\epsilon_ {v} ]} {\\arg \\max } - \\sum_ {i = 1} ^ {M} \\frac {f _ {T} \\left(t _ {i} ^ {\\prime}\\right)}{\\| f _ {T} \\left(t _ {i} ^ {\\prime}\\right) \\|} \\sum_ {s _ {i} \\in S} \\frac {f _ {I} \\left(g \\left(v ^ {\\prime} , s _ {i}\\right)\\right)}{\\| f _ {I} \\left(g \\left(v ^ {\\prime} , s _ {i}\\right)\\right) \\|}, \\tag {2}", + "image_path": "3a218850f66d6472d80e1f925249f88f553127aff7e54a03beafcf261b6d7dd2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 335, + 547, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 547, + 396 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 335, + 547, + 396 + ], + "type": "inline_equation", + "content": "g(v', s_i)" + }, + { + "bbox": [ + 304, + 335, + 547, + 396 + ], + "type": "text", + "content": " denotes the resizing function that takes the image " + }, + { + "bbox": [ + 304, + 335, + 547, + 396 + ], + "type": "inline_equation", + "content": "v'" + }, + { + "bbox": [ + 304, + 335, + 547, + 396 + ], + "type": "text", + "content": " and the scale coefficient " + }, + { + "bbox": [ + 304, + 335, + 547, + 396 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 335, + 547, + 396 + ], + "type": "text", + "content": " as inputs. All the scaled images derived from " + }, + { + "bbox": [ + 304, + 335, + 547, + 396 + ], + "type": "inline_equation", + "content": "v'" + }, + { + "bbox": [ + 304, + 335, + 547, + 396 + ], + "type": "text", + "content": " are encouraged to be far away from all the adversarial captions " + }, + { + "bbox": [ + 304, + 335, + 547, + 396 + ], + "type": "inline_equation", + "content": "t_i'" + }, + { + "bbox": [ + 304, + 335, + 547, + 396 + ], + "type": "text", + "content": " in the embedding space. Finally, the adversarial caption " + }, + { + "bbox": [ + 304, + 335, + 547, + 396 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 304, + 335, + 547, + 396 + ], + "type": "text", + "content": " is generated as follows," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 358, + 400, + 546, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 400, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 358, + 400, + 546, + 426 + ], + "type": "interline_equation", + "content": "t ^ {\\prime} = \\underset {t ^ {\\prime} \\in B [ t, \\epsilon_ {t} ]} {\\arg \\max } - \\frac {f _ {T} \\left(t ^ {\\prime}\\right) \\cdot f _ {I} \\left(v ^ {\\prime}\\right)}{\\| f _ {T} \\left(t ^ {\\prime}\\right) \\| \\| f _ {I} \\left(v ^ {\\prime}\\right) \\|}, \\tag {3}", + "image_path": "e6d549912e4f6508937f6584974f2c7fc9afe16d274360d29d7560f6525954df.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 431, + 546, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 546, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 546, + 468 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 304, + 431, + 546, + 468 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 304, + 431, + 546, + 468 + ], + "type": "text", + "content": " is encouraged to be far away from the adversarial image " + }, + { + "bbox": [ + 304, + 431, + 546, + 468 + ], + "type": "inline_equation", + "content": "v'" + }, + { + "bbox": [ + 304, + 431, + 546, + 468 + ], + "type": "text", + "content": " in the embedding space. The detailed algorithm can be found in Appendix C." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 477, + 388, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 477, + 388, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 477, + 388, + 491 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 497, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 547, + 617 + ], + "type": "text", + "content": "In this section, we present experimental evidence for the advantages of our proposed SGA. We conduct experiments on a diverse set of datasets and popular VLP models. First, we describe the experimental settings in Section 5.1. Then, in Section 5.2, we validate the immediate integration of transfer-based unimodal attacks into multimodal learning. Next, we provide the main evaluation results compared to the state-of-the-art method in Section 5.3. In Section 5.4, we analyze the cross-task transferability between different " + }, + { + "bbox": [ + 304, + 497, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{V + L}" + }, + { + "bbox": [ + 304, + 497, + 547, + 617 + ], + "type": "text", + "content": " tasks. Finally, we present ablation studies in Section 5.5." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 623, + 432, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 432, + 636 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 432, + 636 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": "Datasets. We consider two widely used multimodal datasets, Flickr30K [27] and MSCOCO [23]. Flickr30K consists of 31,783 images, each with five corresponding captions. Similarly, MSCOCO comprises 123,287 images, and each image is annotated with around five captions. We adopt the Karpathy split [14] for experimental evaluation." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "106" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 70, + 497, + 215 + ], + "blocks": [ + { + "bbox": [ + 99, + 70, + 497, + 215 + ], + "lines": [ + { + "bbox": [ + 99, + 70, + 497, + 215 + ], + "spans": [ + { + "bbox": [ + 99, + 70, + 497, + 215 + ], + "type": "table", + "html": "
AttackALBEF*TCLCLIPViTCLIPCNN
TR R@1*IR R@1*TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1
Sep-Attack65.6973.9517.6032.9531.1745.2332.8245.49
Sep-Attack + MI58.81(↓)65.25(↓)16.02(↓)28.19(↓)23.07(↓)36.98(↓)26.56(↓)39.31(↓)
Sep-Attack + DIM56.41(↓)64.24(↓)16.75(↓)29.55(↓)24.17(↓)37.60(↓)25.54(↓)38.77(↓)
Sep-Attack + PNA.PO40.56(↓)53.95(↓)18.44(↑)30.98(↓)22.33(↓)37.02(↓)26.95(↓)38.63(↓)
Co-Attack77.1683.8615.2129.4923.6036.4825.1238.89
Co-Attack + MI64.86(↓)75.26(↓)25.40(↑)38.69(↑)24.91(↑)37.11(↑)26.31(↑)38.97(↑)
Co-Attack + DIM47.03(↓)62.28(↓)22.23(↑)35.45(↑)25.64(↑)38.50(↑)26.95(↑)40.58(↑)
SGA97.2497.2845.4255.2533.3844.1634.9346.57
", + "image_path": "f14f7364892a82e7d2ecbc1051ceea8cbd8a6dfd5e17d8c350f734ae40793086.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 224, + 546, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 224, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 546, + 262 + ], + "type": "text", + "content": "Table 2: Attack success rates (%) of R@1 of integrating transfer-based image attacks on image-text retrieval. The adversarial data are generated on Flickr30K using the source model ALBEF to attack other target models. * indicates white-box attacks. A higher ASR indicates better adversarial transferability." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 270, + 289, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 289, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 289, + 415 + ], + "type": "text", + "content": "VLP Models. We evaluate two popular VLP models, the fused VLP and aligned VLP models. For the fused VLP, we consider ALBEF [18] and TCL [38]. ALBEF contains a 12-layer visual transformer ViT-B/16 [8] and two 6-layer transformers for the image encoder and both the text encoder and the multimodal encoder, respectively. TCL uses the same model architecture as ALBEF but with different pretrain objectives. For the aligned VLP model, we choose to evaluate CLIP [28]. CLIP has two different image encoder choices, namely, " + }, + { + "bbox": [ + 46, + 270, + 289, + 415 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{ViT}}" + }, + { + "bbox": [ + 46, + 270, + 289, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 270, + 289, + 415 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{CNN}}" + }, + { + "bbox": [ + 46, + 270, + 289, + 415 + ], + "type": "text", + "content": ", that use ViT-B/16 and ResNet-101 [12] as the base architectures for the image encoder, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 418, + 288, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 288, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 288, + 540 + ], + "type": "text", + "content": "Adversarial Attack Settings. To craft adversarial images, we employ PGD [24] with perturbation bound " + }, + { + "bbox": [ + 46, + 418, + 288, + 540 + ], + "type": "inline_equation", + "content": "\\epsilon_v = 2/255" + }, + { + "bbox": [ + 46, + 418, + 288, + 540 + ], + "type": "text", + "content": ", step size " + }, + { + "bbox": [ + 46, + 418, + 288, + 540 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5/255" + }, + { + "bbox": [ + 46, + 418, + 288, + 540 + ], + "type": "text", + "content": ", and iteration steps " + }, + { + "bbox": [ + 46, + 418, + 288, + 540 + ], + "type": "inline_equation", + "content": "T = 10" + }, + { + "bbox": [ + 46, + 418, + 288, + 540 + ], + "type": "text", + "content": ". For attacking text modality, we adopt BERT-Attack [19] with perturbation bound " + }, + { + "bbox": [ + 46, + 418, + 288, + 540 + ], + "type": "inline_equation", + "content": "\\epsilon_t = 1" + }, + { + "bbox": [ + 46, + 418, + 288, + 540 + ], + "type": "text", + "content": " and length of word list " + }, + { + "bbox": [ + 46, + 418, + 288, + 540 + ], + "type": "inline_equation", + "content": "W = 10" + }, + { + "bbox": [ + 46, + 418, + 288, + 540 + ], + "type": "text", + "content": ". Furthermore, we enlarge the image set by resizing the original image into five scales, " + }, + { + "bbox": [ + 46, + 418, + 288, + 540 + ], + "type": "inline_equation", + "content": "\\{0.50, 0.75, 1.00, 1.25, 1.50\\}" + }, + { + "bbox": [ + 46, + 418, + 288, + 540 + ], + "type": "text", + "content": ", using bicubic interpolation. Similarly, the caption set is enlarged by augmenting the most matching caption pairs for each image in the dataset, with the size of approximately five." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 544, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 288, + 616 + ], + "type": "text", + "content": "Metrics. We employ Attack Success Rate (ASR) as the metric for evaluating the adversarial robustness and transferability in both white-box and black-box settings. Specifically, ASR evaluates the percentage of attacks that only produce successful adversarial examples. A higher ASR indicates better adversarial transferability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 623, + 183, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 183, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 183, + 635 + ], + "type": "text", + "content": "5.2. Transferability Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "In this paper, we present a systematic study of the adversarial transferability of VLP models, which has not been explored. As demonstrated in Section 3, existing methods, including the separate unimodal adversarial attack (SepAttack) and the multimodal adversarial attack (Co-Attack), exhibit limited transferability to other VLP models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 270, + 547, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 547, + 354 + ], + "type": "text", + "content": "To improve transferability in multimodal learning, we intuitively investigate the adoption of transfer-based attacks from unimodal learning such as image classification. Specifically, we consider MI [7], DIM [37], and PNA.PO [36]. However, this approach can be problematic if cross-modal interactions and the unique many-to-many alignments in multimodal learning are not taken into account." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 354, + 548, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 548, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 548, + 558 + ], + "type": "text", + "content": "Table 2 illustrates that multimodal attack methods that incorporate transfer-based image attacks exhibit minimal improvement in transferability while compromising white-box performance. Specifically, when integrated with MI, Co-Attack drops significantly by " + }, + { + "bbox": [ + 304, + 354, + 548, + 558 + ], + "type": "inline_equation", + "content": "12.3\\%" + }, + { + "bbox": [ + 304, + 354, + 548, + 558 + ], + "type": "text", + "content": " in white-box settings, while only maintaining " + }, + { + "bbox": [ + 304, + 354, + 548, + 558 + ], + "type": "inline_equation", + "content": "25.40\\%" + }, + { + "bbox": [ + 304, + 354, + 548, + 558 + ], + "type": "text", + "content": " ASR in transferability (ALBEF to TCL). However, our SGA shows superior performance in both white-box and black-box settings. Notably, Sep-Attack combined with transfer-based attacks not only reduces the effectiveness of white-box attacks but also fails to improve adversarial transferability in almost all black-box settings. The results provide empirical evidence that directly combining unimodal adversarial attacks in multimodal learning without considering cross-modal interactions and alignments can be problematic, even when using transfer-based unimodal attacks. Additional discussion is provided in Appendix B.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 563, + 430, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 430, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 430, + 576 + ], + "type": "text", + "content": "5.3. Experimental Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 582, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 700 + ], + "type": "text", + "content": "Multimodal Fusion Modules. First, we investigate VLP models with different fusion modules, namely, fused VLP models and aligned VLP models. We generate adversarial examples on both types of models and evaluate their attack performance when transferred to other VLP models while ensuring consistency in the input size of images. For example, adversarial images generated by ALBEF or TCL are resized to " + }, + { + "bbox": [ + 304, + 582, + 547, + 700 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 582, + 547, + 700 + ], + "type": "text", + "content": " before performing transfer attacks on CLIP, and adversarial examples generated on CLIP are resized to " + }, + { + "bbox": [ + 304, + 582, + 547, + 700 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 304, + 582, + 547, + 700 + ], + "type": "text", + "content": " before being transferred to ALBEF or TCL." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 702, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 702, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 702, + 545, + 713 + ], + "type": "text", + "content": "As shown in Table 3, experimental results demonstrate" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 748, + 312, + 757 + ], + "type": "text", + "content": "107" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 544, + 331 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 544, + 331 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 544, + 331 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 544, + 331 + ], + "type": "table", + "html": "
SourceAttackALBEFTCLCLIPViTCLIPCNN
TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1
ALBEFPGD52.45*58.65*3.066.798.9613.2110.3414.65
BERT-Attack11.57*27.46*12.6428.0729.3343.1732.6946.11
Sep-Attack65.69*73.95*17.6032.9531.1745.2332.8245.49
Co-Attack77.16*83.86*15.2129.4923.6036.4825.1238.89
SGA97.24±0.22*97.28±0.15*45.42±0.6055.25±0.0633.38±0.3544.16±0.2534.93±0.9946.57±0.13
TCLPGD6.1510.7877.87*79.48*7.4813.7210.3415.33
BERT-Attack11.8926.8214.54*29.17*29.6944.4933.4646.07
Sep-Attack20.1336.4884.72*86.07*31.2944.6533.3345.80
Co-Attack23.1540.0477.94*85.59*27.8541.1930.7444.11
SGA48.91±0.7460.34±0.1098.37±0.08*98.81±0.07*33.87±0.1844.88±0.5437.74±0.2748.30±0.34
CLIPViTPGD2.504.934.858.1770.92*78.61*5.368.44
BERT-Attack9.5922.6411.8025.0728.34*39.08*30.4037.43
Sep-Attack9.5923.2511.3825.6079.75*86.79*30.7839.76
Co-Attack10.5724.3311.9426.6993.25*95.86*32.5241.82
SGA13.40±0.0727.22 ±0.0616.23±0.4530.76±0.0799.08±0.08*98.94±0.00*38.76±0.2747.79±0.58
CLIPCNNPGD2.094.824.007.811.106.6086.46*92.25*
BERT-Attack8.8623.2712.3325.4827.1237.4430.40*40.10*
Sep-Attack8.5523.4112.6426.1228.3439.4391.44*95.44*
Co-Attack8.7923.7413.1026.0728.7940.0394.76*96.89*
SGA11.42±0.0724.80±0.2814.91±0.0828.82±0.1131.24±0.4242.12±0.1199.24±0.18*99.49±0.05*
", + "image_path": "e5bc183d7ed555a792d8d3cfa0b84de04a7a750b25960fb27d318d7ef733b30a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 67, + 392, + 269, + 445 + ], + "blocks": [ + { + "bbox": [ + 46, + 341, + 546, + 377 + ], + "lines": [ + { + "bbox": [ + 46, + 341, + 546, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 546, + 377 + ], + "type": "text", + "content": "Table 3: Comparison with state-of-the-art methods on image-text retrieval. We report the attack success rate " + }, + { + "bbox": [ + 46, + 341, + 546, + 377 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 341, + 546, + 377 + ], + "type": "text", + "content": " of R@1 for both IR and TR. The source column indicates the source models used to generate the adversarial data on Flickr30K. * indicates white-box attacks. A higher ASR indicates better adversarial transferability." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 392, + 269, + 445 + ], + "lines": [ + { + "bbox": [ + 67, + 392, + 269, + 445 + ], + "spans": [ + { + "bbox": [ + 67, + 392, + 269, + 445 + ], + "type": "table", + "html": "
AttackB@4METEORROUGE_LCIDErSPICE
Baseline39.731.060.0133.323.8
Co-Attack37.429.858.4125.522.8
SGA34.828.456.3116.021.4
", + "image_path": "f5772f1a123944da880051baad37e8c72daf4c9fe40d7537e71756152b8df8be.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 453, + 288, + 526 + ], + "lines": [ + { + "bbox": [ + 46, + 453, + 288, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 288, + 526 + ], + "type": "text", + "content": "Table 4: Cross-Task Transferability: ITR " + }, + { + "bbox": [ + 46, + 453, + 288, + 526 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 453, + 288, + 526 + ], + "type": "text", + "content": " IC. Adversarial data generated from Image-Text Retrieval (ITR) to attack Image Captioning (IC) on MSCOCO. The source model and target model are ALBEF and BLIP, respectively. The Baseline represents the original performance of IC on clean data. Lower values indicate better adversarial transferability." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 536, + 288, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 288, + 668 + ], + "type": "text", + "content": "the superiority of our proposed SGA over existing multimodal attack methods in all black-box settings. Specifically, our SGA achieves significant improvements in adversarial transferability when the source and target models are of the same type. For instance, SGA outperforms Co-Attack by approximately " + }, + { + "bbox": [ + 46, + 536, + 288, + 668 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 536, + 288, + 668 + ], + "type": "text", + "content": " in terms of attack success rate when transferring adversarial data from ALBEF to TCL. Moreover, in the more challenging scenario where the source and target models are of different types, SGA also surpasses CoAttack with higher attack success rates. More experiments on the MSCOCO dataset are provided in Appendix B.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Model Architectures. Then, we explore VLP models with respect to different model architectures. Many VLP models commonly use ViTs as the vision encoder, where images" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 393, + 547, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 547, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 547, + 490 + ], + "type": "text", + "content": "are segmented into patches before being processed by the transformer model. However, in the case of CLIP, the image encoder consists of two distinct architectures: conventional Convolutional Neural Networks (CNNs) and ViTs. The transferability between CNNs and ViTs has been well-studied in unimodal learning. Therefore, we also investigate the adversarial transferability of CNN-based and ViT-based CLIP models in multimodal learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 491, + 547, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 547, + 586 + ], + "type": "text", + "content": "As shown in Table 3, we observe a similar phenomenon as unimodal learning that compared to CNNs, ViTs show better robustness against adversarial perturbations [25]. Specifically, for all attack methods, the same adversarial multimodal data have a stronger white-box attack effect on " + }, + { + "bbox": [ + 304, + 491, + 547, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{CNN}}" + }, + { + "bbox": [ + 304, + 491, + 547, + 586 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 304, + 491, + 547, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{ViT}}" + }, + { + "bbox": [ + 304, + 491, + 547, + 586 + ], + "type": "text", + "content": ". Moreover, the adversarial examples generated on " + }, + { + "bbox": [ + 304, + 491, + 547, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{ViT}}" + }, + { + "bbox": [ + 304, + 491, + 547, + 586 + ], + "type": "text", + "content": " are found to be more transferable to " + }, + { + "bbox": [ + 304, + 491, + 547, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{CNN}}" + }, + { + "bbox": [ + 304, + 491, + 547, + 586 + ], + "type": "text", + "content": " than vice versa (38.76% vs. 31.24%)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 587, + 547, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 587, + 547, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 587, + 547, + 660 + ], + "type": "text", + "content": "Furthermore, our proposed SGA consistently improves transferability on both CNN-based CLIP and ViT-based CLIP compared to other attacks. For instance, SGA increases the adversarial transferability for " + }, + { + "bbox": [ + 304, + 587, + 547, + 660 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{ViT}}" + }, + { + "bbox": [ + 304, + 587, + 547, + 660 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 304, + 587, + 547, + 660 + ], + "type": "inline_equation", + "content": "5.83\\%" + }, + { + "bbox": [ + 304, + 587, + 547, + 660 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 587, + 547, + 660 + ], + "type": "inline_equation", + "content": "6.24\\%" + }, + { + "bbox": [ + 304, + 587, + 547, + 660 + ], + "type": "text", + "content": " compared to Co-Attack under the white-box setting and black-box setting, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 670, + 454, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 454, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 454, + 684 + ], + "type": "text", + "content": "5.4. Cross-Task Transferability" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "content": "Cross-modal interactions and alignments are the core components of multimodal learning regardless of the task." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "108" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 254, + 124 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 254, + 124 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 254, + 124 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 254, + 124 + ], + "type": "table", + "html": "
AttackValTestATestB
Baseline58.4665.8946.25
Co-Attack54.2661.8043.81
SGA53.5561.1943.71
", + "image_path": "2d4b85bac06fd3c74c6c318d32d997b8dfd8dc00ca6ac89fe1d132f3d87f680f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 206 + ], + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 206 + ], + "type": "text", + "content": "Table 5: Cross-Task Transferability: ITR " + }, + { + "bbox": [ + 46, + 133, + 288, + 206 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 133, + 288, + 206 + ], + "type": "text", + "content": " VG. Adversarial data generated from Image-Text Retrieval (ITR) to attack Visual Grounding (VG) on RefCOCO+. The source model and target model are both ALBEF. The Baseline represents the original performance of VG on clean data. Lower values indicate better adversarial transferability." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 218, + 288, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 288, + 255 + ], + "type": "text", + "content": "Therefore, we conduct extensive experiments to explore the effectiveness of our proposed SGA on two additional V+L tasks: Image Captioning (IC) and Visual Grounding (VG)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 262, + 287, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 393 + ], + "type": "text", + "content": "Image Captioning. Image captioning is a generation-based task, where an input image is encoded into a feature vector and then decoded into a natural language sentence. In our experiments, we craft adversarial images using the source model (ALBEF) with an image-text retrieval objective and then directly attack the target model (BLIP [17]) on image captioning. We employ the MSCOCO dataset, which is suitable for both two tasks, and utilize various evaluation metrics to measure the quality of the generated captions, including BLEU [26], METEOR [2], ROUGE [21], CIDEr [32], and SPICE [1]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 394, + 288, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 394, + 288, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 288, + 466 + ], + "type": "text", + "content": "We present the performance on image captioning of BLIP after being attacked in Table 4. Experimental results demonstrate clear improvements in the adversarial transferability of the proposed SGA compared to Co-Attack. Specifically, our SGA improves the BLEU score by up to " + }, + { + "bbox": [ + 46, + 394, + 288, + 466 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 46, + 394, + 288, + 466 + ], + "type": "text", + "content": " and the CIDEr score by up to " + }, + { + "bbox": [ + 46, + 394, + 288, + 466 + ], + "type": "inline_equation", + "content": "9.5\\%" + }, + { + "bbox": [ + 46, + 394, + 288, + 466 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 474, + 288, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 288, + 559 + ], + "type": "text", + "content": "Visual Grounding. Visual grounding is another " + }, + { + "bbox": [ + 46, + 474, + 288, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{V} + \\mathrm{L}" + }, + { + "bbox": [ + 46, + 474, + 288, + 559 + ], + "type": "text", + "content": " task that aims to localize the region in an image based on the corresponding specific textual description. Similarly, we generate adversarial images using the source model (ALBEF) from image-text to attack the target model (ALBEF) on visual grounding. Table 5 shows the results on RefCOCO+ [39], where our SGA still outperforms Co-Attack." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 567, + 141, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 141, + 580 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 141, + 580 + ], + "type": "text", + "content": "5.5. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 586, + 288, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 288, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 288, + 682 + ], + "type": "text", + "content": "To systematically investigate the impact of our set-level alignment-preserving augmentations, we conducted ablation experiments on image-text retrieval to evaluate the effect of varying the number of augmented image sets " + }, + { + "bbox": [ + 46, + 586, + 288, + 682 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 586, + 288, + 682 + ], + "type": "text", + "content": " with multiscale transformation and the number " + }, + { + "bbox": [ + 46, + 586, + 288, + 682 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 586, + 288, + 682 + ], + "type": "text", + "content": " of augmented caption sets. Specifically, we employed ALBEF as the source model and " + }, + { + "bbox": [ + 46, + 586, + 288, + 682 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{ViT}}" + }, + { + "bbox": [ + 46, + 586, + 288, + 682 + ], + "type": "text", + "content": " as the target model on Flickr30K. More details are provided in Appendix B.4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Multi-scale Image Set. We propose the use of multiple scale-invariant images to generate diverse adversarial data in" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 305, + 70, + 425, + 173 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 425, + 173 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 425, + 173 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 425, + 173 + ], + "type": "image", + "image_path": "e089b05f0c6b102500fe1ce548125d1e3e81f5360af1128d6b7c5e4ee9bc5ff3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 180, + 547, + 229 + ], + "lines": [ + { + "bbox": [ + 304, + 180, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 547, + 229 + ], + "type": "text", + "content": "Figure 5: Ablation Study: Attack success rates (\\%) on set-level augmentations, Left for image set and Right for caption set. The performance of SGA is consistently improved by using larger augmented image and caption sets." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 427, + 70, + 545, + 171 + ], + "blocks": [ + { + "bbox": [ + 427, + 70, + 545, + 171 + ], + "lines": [ + { + "bbox": [ + 427, + 70, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 427, + 70, + 545, + 171 + ], + "type": "image", + "image_path": "f3d389908dc4be6587473a28e3085c3c66d1c80bffd7ef7c676abf2bc33238d1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 240, + 547, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 547, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 547, + 301 + ], + "type": "text", + "content": "SGA. Results in the left of Figure 5 reveal that the transferability significantly increases as we introduce more diverse images with different scales, peaking when the scale range is set to [0.50, 1.50] with a step of 0.25. We set the scale range " + }, + { + "bbox": [ + 304, + 240, + 547, + 301 + ], + "type": "inline_equation", + "content": "S = \\{0.50, 0.75, 1.00, 1.25, 1.50\\}" + }, + { + "bbox": [ + 304, + 240, + 547, + 301 + ], + "type": "text", + "content": " for optimal performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 305, + 548, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 548, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 548, + 461 + ], + "type": "text", + "content": "Multi-pair Caption Set. We also conduct experiments to investigate the impact of an enlarged caption set on adversarial transferability. The number of additional captions ranged from 1 to " + }, + { + "bbox": [ + 304, + 305, + 548, + 461 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 305, + 548, + 461 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 305, + 548, + 461 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 305, + 548, + 461 + ], + "type": "text", + "content": " represents the most matching caption pairs from the dataset for each image. Results presented in the right panel of Figure 5 indicate that if " + }, + { + "bbox": [ + 304, + 305, + 548, + 461 + ], + "type": "inline_equation", + "content": "M > 1" + }, + { + "bbox": [ + 304, + 305, + 548, + 461 + ], + "type": "text", + "content": ", the black-box performance increases significantly but eventually plateaus. These results demonstrate the effectiveness of using multiple alignment-preserving inter-modal information to enhance adversarial transferability. Furthermore, we observed that the performance is relatively insensitive to the number of extra captions, but adding more captions can improve the overall adversarial transferability." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 472, + 378, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 472, + 378, + 485 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 378, + 485 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 493, + 547, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 547, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 547, + 659 + ], + "type": "text", + "content": "In this paper, we make the first attempt to investigate the adversarial transferability of typical VLP models. We systematically evaluate the existing attack methods and reveal that they still exhibit lower transferability, despite their impressive performance in white-box settings. Our investigation highlights the need for specially designed transferable attacks in multimodal learning that can model the many-to-many cross-modal alignments and interactions. We propose SGA, a highly transferable multimodal attack, which leverages set-level alignment-preserving augmentations through cross-modal guidance to thoroughly exploit multimodal interactions. We hope that this work could inspire further research to evaluate and enhance the adversarial robustness of VLP models." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgments. This work was supported by the National Key R&D Program of China (Grant NO. 2022YFF1202903) and the National Natural Science Foundation of China (Grant NO. 62122035)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "109" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In ECCV, 2016. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 288, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 288, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 288, + 159 + ], + "type": "text", + "content": "[2] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In IEEvaluation@ACL, 2005. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 160, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 160, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 54, + 160, + 287, + 194 + ], + "type": "text", + "content": "[3] Nicholas Carlini and David A. Wagner. Towards evaluating the robustness of neural networks. 2017 IEEE Symposium on Security and Privacy (SP), pages 39-57, 2017. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 195, + 288, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 195, + 288, + 250 + ], + "spans": [ + { + "bbox": [ + 53, + 195, + 288, + 250 + ], + "type": "text", + "content": "[4] Hui Chen, Guiguang Ding, Xudong Liu, Zijia Lin, Ji Liu, and Jungong Han. Imram: Iterative matching with recurrent attention memory for cross-modal image-text retrieval. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12652-12660, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 294 + ], + "type": "text", + "content": "[5] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 297, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 297, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 288, + 361 + ], + "type": "text", + "content": "[6] Mengjun Cheng, Yipeng Sun, Long Wang, Xiongwei Zhu, Kun Yao, Jie Chen, Guoli Song, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Vista: Vision and scene text aggregation for cross-modal retrieval. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5174-5183, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 364, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 364, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 364, + 287, + 407 + ], + "type": "text", + "content": "[7] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 409, + 288, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 288, + 474 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv, abs/2010.11929, 2021. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 476, + 288, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 288, + 530 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 288, + 530 + ], + "type": "text", + "content": "[9] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuo-hang Wang, Lijuan Wang, Chenguang Zhu, Nanyun Peng, Zicheng Liu, and Michael Zeng. An empirical study of training end-to-end vision-and-language transformers. ArXiv, abs/2111.02387, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 576 + ], + "type": "text", + "content": "[10] Ji Gao, Jack Lanchantin, Mary Lou Soffa, and Yanjun Qi. Black-box generation of adversarial text sequences to evade deep learning classifiers. 2018 IEEE Security and Privacy Workshops (SPW), pages 50-56, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 609 + ], + "type": "text", + "content": "[11] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. CoRR, abs/1412.6572, 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 612, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 655 + ], + "type": "text", + "content": "[12] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "text", + "content": "[13] Di Jin, Zhijing Jin, Joey Tianyi Zhou, and Peter Szolovits. Is bert really robust? a strong baseline for natural language attack on text classification and entailment. In AAAI, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[14] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. IEEE Transactions" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 74, + 546, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 546, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 546, + 94 + ], + "type": "text", + "content": "on Pattern Analysis and Machine Intelligence, 39:664-676, 2017.5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 97, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 547, + 140 + ], + "type": "text", + "content": "[15] Zaid Khan and Yun Raymond Fu. Exploiting bert for multimodal target sentiment classification through input space translation. Proceedings of the 29th ACM International Conference on Multimedia, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 141, + 546, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 546, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 546, + 195 + ], + "type": "text", + "content": "[16] Chenyi Lei, Shixian Luo, Yong Liu, Wanggui He, Jiamang Wang, Guoxin Wang, Haihong Tang, Chunyan Miao, and Houqiang Li. Understanding chinese video and language via contrastive multimodal pre-training. Proceedings of the 29th ACM International Conference on Multimedia, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 198, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 547, + 239 + ], + "type": "text", + "content": "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. ArXiv, abs/2201.12086, 2022. 2, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 243, + 547, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 547, + 295 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 547, + 295 + ], + "type": "text", + "content": "[18] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq R. Joty, Caiming Xiong, and Steven C. H. Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 2, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 298, + 546, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 546, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 546, + 330 + ], + "type": "text", + "content": "[19] Linyang Li, Ruotian Ma, Qipeng Guo, X. Xue, and Xipeng Qiu. Bert-attack: Adversarial attack against bert using bert. ArXiv, abs/2004.09984, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 332, + 546, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 546, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 546, + 385 + ], + "type": "text", + "content": "[20] Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 388, + 546, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 546, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 546, + 421 + ], + "type": "text", + "content": "[21] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Annual Meeting of the Association for Computational Linguistics, 2004. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 422, + 546, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 546, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 546, + 454 + ], + "type": "text", + "content": "[22] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. arXiv: Learning, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 456, + 546, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 546, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 546, + 498 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 501, + 546, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 546, + 544 + ], + "type": "text", + "content": "[24] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. ArXiv, abs/1706.06083, 2018. 1, 2, 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 546, + 546, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 546, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 546, + 590 + ], + "type": "text", + "content": "[25] Muzammal Naseer, Kanchana Ranasinghe, Salman Hameed Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In Neural Information Processing Systems, 2021. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 591, + 546, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 546, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 546, + 623 + ], + "type": "text", + "content": "[26] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 625, + 546, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 546, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 546, + 679 + ], + "type": "text", + "content": "[27] Bryan A. Plummer, Liwei Wang, Christopher M. Cervantes, Juan C. Caicedo, J. Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. International Journal of Computer Vision, 123:74–93, 2015. 2, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 681, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 546, + 713 + ], + "type": "text", + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "110" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 105 + ], + "type": "text", + "content": "Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 288, + 139 + ], + "type": "text", + "content": "[29] Shuhuai Ren, Yihe Deng, Kun He, and Wanxiang Che. Generating natural language adversarial examples through probability weighted word saliency. In ACL, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "type": "text", + "content": "[30] Lei Shi, Kai Shuang, Shijie Geng, Peng Gao, Zuohui Fu, Gerard de Melo, Yunpeng Chen, and Sen Su. Dense contrastive visual-linguistic pretraining. Proceedings of the 29th ACM International Conference on Multimedia, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "type": "text", + "content": "[31] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "text", + "content": "[32] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4566-4575, 2015. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 327 + ], + "type": "text", + "content": "[33] Teng Wang, Yixiao Ge, Feng Zheng, Ran Cheng, Ying Shan, Xiaohu Qie, and Ping Luo. Accelerating vision-language pretraining with free language modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23161-23170, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 381 + ], + "type": "text", + "content": "[34] Teng Wang, Wenhao Jiang, Zhichao Lu, Feng Zheng, Ran Cheng, Chengguo Yin, and Ping Luo. Vlmixer: Unpaired vision-language pre-training via cross-modal cutmix. In International Conference on Machine Learning, pages 22680-22690. PMLR, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 288, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 288, + 436 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 288, + 436 + ], + "type": "text", + "content": "[35] Zihao Wang, Xihui Liu, Hongsheng Li, Lu Sheng, Junjie Yan, Xiaogang Wang, and Jing Shao. Camp: Cross-modal adaptive message passing for text-image retrieval. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5763-5772, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 288, + 480 + ], + "type": "text", + "content": "[36] Zhipeng Wei, Jingjing Chen, Micah Goldblum, Zuxuan Wu, Tom Goldstein, and Yu-Gang Jiang. Towards transferable adversarial attacks on vision transformers. ArXiv, abs/2109.04176, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 481, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 288, + 536 + ], + "type": "text", + "content": "[37] Cihang Xie, Zhishuai Zhang, Jianyu Wang, Yuyin Zhou, Zhou Ren, and Alan Loddon Yuille. Improving transferability of adversarial examples with input diversity. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2725-2734, 2018. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 288, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 288, + 581 + ], + "type": "text", + "content": "[38] Jinyu Yang, Jiali Duan, S. Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul M. Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. ArXiv, abs/2202.10401, 2022. 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 582, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 288, + 613 + ], + "type": "text", + "content": "[39] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C. Berg, and Tamara L. Berg. Modeling context in referring expressions. ArXiv, abs/1608.00272, 2016. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 669 + ], + "type": "text", + "content": "[40] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Francis E. H. Tay, Jiashi Feng, and Shuicheng Yan. Tokensto-token vit: Training vision transformers from scratch onImagenet. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 538-547, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[41] Jiaming Zhang, Qiaomin Yi, and Jitao Sang. Towards adversarial attack on vision-language pre-training models. Proceedings of the 30th ACM International Conference on Multimedia, 2022. 1, 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 174 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "type": "text", + "content": "[42] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5575-5584, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 547, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 174 + ], + "type": "text", + "content": "[43] Qi Zhang, Zhen Lei, Zhaoxiang Zhang, and S. Li. Context-aware attention network for image-text retrieval. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3533-3542, 2020. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "111" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Shape Analysis of Euclidean Curves under Frenet-Serret Framework/235fdcf4-691b-437b-8ee8-c7b76395f3ab_content_list.json b/2023/Shape Analysis of Euclidean Curves under Frenet-Serret Framework/235fdcf4-691b-437b-8ee8-c7b76395f3ab_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1c7ebad7d1489ee2e0fe096e41c0e02fdd87f575 --- /dev/null +++ b/2023/Shape Analysis of Euclidean Curves under Frenet-Serret Framework/235fdcf4-691b-437b-8ee8-c7b76395f3ab_content_list.json @@ -0,0 +1,2019 @@ +[ + { + "type": "text", + "text": "Shape Analysis of Euclidean Curves under Frenet-Serret Framework", + "text_level": 1, + "bbox": [ + 135, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Perrine Chassat1 \n1LaMME, University of Paris-Saclay \nperrine.chassat@univ-evry.fr", + "bbox": [ + 99, + 180, + 392, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Juhyun Park $^{1,2}$ $^{2}$ ENSIIE, Evry \njuhyun.park@ensiie.fr", + "bbox": [ + 431, + 180, + 619, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nicolas Brunel1,2,3 \n3Quantmetry, Paris \nnicolas.brunel@ensiie.fr", + "bbox": [ + 655, + 180, + 869, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Geometric frameworks for analyzing curves are common in applications as they focus on invariant features and provide visually satisfying solutions to standard problems such as computing invariant distances, averaging curves, or registering curves. We show that for any smooth curve in $\\mathbb{R}^d$ , $d > 1$ , the generalized curvatures associated with the Frenet-Serret equation can be used to define a Riemannian geometry that takes into account all the geometric features of the shape. This geometry is based on a Square Root Curvature Transform that extends the square root-velocity transform for Euclidean curves (in any dimensions) and provides likely geodesics that avoid artefacts encountered by representations using only first-order geometric information. Our analysis is supported by simulated data and is especially relevant for analyzing human motions. We consider trajectories acquired from sign language, and show the interest of considering curvature and also torsion in their analysis, both being physically meaningful.", + "bbox": [ + 75, + 301, + 473, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 603, + 209, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Identifying and comparing different types of visual objects is a fundamental task in machine learning and computer vision problems [13, 5, 8]. The shape is one of the essential features of objects that allow us to understand and characterize them. Nowadays, it is much easier to obtain data in the form of shapes, typically as dense point clouds or landmarks. The main task in shape analysis is to define a proper framework to compare and quantify the variation of the shapes. However, the shape space is generally nonlinear, and extracting meaningful information or features is complex. One of the successful approaches to shape analysis utilizes a Riemannian framework of differential geometry, where a metric can be defined between the shapes, which is invariant with respect to shape-preserving transformations such as translation and rotation. For instance, this gives rise to geodesic distances that are naturally invariant to smooth and optimal deformations through geodesic paths between the shapes [11]. This approach is very versatile as it can", + "bbox": [ + 75, + 628, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "be adapted to various kinds of manifold-value data and can be designed to emphasize important geometric information to be preserved. As a consequence, several choices of metrics are possible, such as the class of invariant Sobolev metrics, often called elastic, for the analysis of curves [1]. In this work, we are concerned with curves that often arise in the application as trajectories (function of time) or motions (animation, activity recognition) and with the definition of a framework to compare their shapes. The differential geometry of Euclidean curves is among the simplest (with respect to higher dimensional manifolds), and relatively simple Riemannian metrics are available with different mathematical representations of curves [28]. Quite remarkably, the introduction of the Square Root Velocity (SRV) transform [24] that consists of a particular representation of the shape of a curve enables to define a so-called elastic Riemannian distance, which has proven to be useful for the statistical shape analysis of 2D and 3D curves in applications. The SRV possesses interesting properties such as a principled theoretical framework, efficient computation, and generalization to higher dimensions [2]. Nevertheless, a limitation of the SRV transform and the corresponding elastic distance is the restrictive use of the first-order derivative, while the geometry of 3-D (or $d$ -D) curves depends on the derivatives until order $d$ . Indeed, it is well-known that a 3D curve is characterized by its curvature and its torsion: this is particularly critical when we consider trajectories or human movements, where the curvature and torsion can have a physical meaning.", + "bbox": [ + 496, + 268, + 893, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2. Related works and contributions", + "text_level": 1, + "bbox": [ + 500, + 723, + 797, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As we will recall in section 3, the full geometry of a curve can be given either by the Frenet curvatures (standard curvature and torsion in 3D) or by the path of Frenet frames. There have been few attempts to directly deal with the Frenet curvatures: most of the works have been produced in 2D curves as an alternative representation [23]. Nevertheless, the potential for applications has not been investigated. In [17], the elastic shape analysis framework has been considered for 3D curves based on the Frenet frames, but the link to the physical parameters has been overlooked.", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4027", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We can also mention the shape analysis of curves on Lie groups [4] with application in computer animation. Outside the Riemannian framework, an attempt has been made to use a direct curvature-based interpolation of curves [20, 25].", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce two representations of Euclidean curves for their shape analysis that use their complete geometry through Frenet curvatures. We provide the full development of the Riemannian frameworks associated with these two representations. As a consequence and in comparison with existing methods, our approaches also give explicit formulas for geodesics and geodesic distances. The first representation considered is based directly on unparametrized Frenet curvatures. We show through experiments that it defines a shape analysis framework that lacks elasticity. As the main contribution, we propose the definition of a second representation, called the Square-Root Curvature (SRC) Transform, which takes into account reparameterization and defines a metric on the space of shapes through the quotient space with the group of diffeomorphisms. One can imagine that the classical method associated with the SRVF, defining a Sobolev elastic metric [2], already implicitly uses all the geometric information necessary for a relevant curve analysis. We show here with simple examples that this is not the case. Through experiments on synthetic data, we compare the methods, and illustrate the limitations of the SRVF one, due to its lack of use of geometric information. To be able to judge and compare the quality of these metrics, we compare consistent sets of curves characterized by specific features. The SRC method shows a special strength in defining a framework that remains consistent with these sets. The straightforward example of geodesics between helices with different numbers of spins (Figure 1 in 2D and Figure 2 in 3D) shows that, in contrast, this is not the case for the SRVF method. In addition, we highlight the interest of these Frenet curvatures-based representations in the real application case of human motion trajectory analysis.", + "bbox": [ + 75, + 151, + 472, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Riemannian Geometry on Shape Space", + "text_level": 1, + "bbox": [ + 76, + 662, + 429, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce useful notations and we review the main approach for constructing tractable representations of the shape of a curve and deriving a Riemannian geometry.", + "bbox": [ + 75, + 686, + 468, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Shape Analysis of Euclidean Curves", + "text_level": 1, + "bbox": [ + 76, + 741, + 390, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We consider absolutely continuous curves that are smooth, open, and with values in some Euclidean space $\\mathbb{R}^d$ , we denote this set as $AC([0,1],\\mathbb{R}^d)$ . These curves are typically parametrized by a variable $t$ that can usually be interpreted as time. Nevertheless, from a (statistical) shape analysis point of view, we focus on the geometric shape of curves that do not depend on a specific parametrization or standard transformations such as translations, rotations, scaling, or reparametrizations. To distinguish be", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tween parametrized curves that differ only by translation, we consider the set of absolutely continuous curves where $x(0) = 0$ , denoted by $AC_0([0,1],\\mathbb{R}^d)$ . The natural and intrinsic parametrization that uniquely defines the shape of a curve $x$ is the arc-length parametrization, defined with the arc length function $s(t) = \\int_0^t\\| \\dot{x} (u)\\| du$ , for $t\\in [0,1]$ . In order to remove the scaling variability, the total length of the curve $s(1)$ is set to 1. Under this parametrization, the shape $X:[0,1]\\mapsto \\mathbb{R}^d$ of the curve is the image of the function $x$ such that $x(t) = X(s(t))$ . As we want to study shapes independently of their parameterizations, we introduce the reparametrization group $\\mathrm{Diff}_{+}([0,1])$ , of smooth orientation preserving diffeomorphisms of the interval [0, 1] onto itself. This group acts on the space of absolutely continuous curves by right composition, and this action only alters the parametrization of the curve, not the inherent shape $X$ . The space of such shapes (or unparametrized curves) is often mathematically defined as the quotient space", + "bbox": [ + 496, + 90, + 893, + 364 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} ([ 0, 1 ], \\mathbb {R} ^ {d}) = A C _ {0} ([ 0, 1 ], \\mathbb {R} ^ {d}) / \\operatorname {D i f f} _ {+} ([ 0, 1 ]). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 375, + 890, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The purpose of shape analysis of curves is to define a distance function $d_{S}$ on $S$ and a framework to perform a complete statistical analysis on a set of curves in $S$ (e.g. mean, classification, or Principal Component Analysis etc.). One of the main challenges in defining this distance is to choose an appropriate mathematical representation of the curves that can be made invariant to all shape-preserving transformations - translation, rotation, scaling, and reparametrization. Moreover, one of the stakes of such representation is to offer an (infinite-dimensional) Riemannian manifold structure that brings powerful and flexible tools for studying the geometry of shapes or statistical properties notably thanks to the tangent space of the manifold [11, 22]. In [23], a list of the few possible representations is given - coordinate functions, curvatures, angle function, and square-root velocity function (SRVF) - and a framework for curve analysis is derived for the last two ones. While the angle representation is unparameterized, the SRVF representation depends on the parametrization, which is shown to be very useful as a tool for the registration of points across curves. As a consequence, the parameterization group $\\mathrm{Diff}_{+}([0,1])$ must be eliminated by using a quotient space. The classical approach is to define the Riemannian metric on the shape space through a metric on the space of parametrized representations that is invariant to reparametrization: $\\forall h \\in \\mathrm{Diff}_{+}([0,1])$", + "bbox": [ + 496, + 404, + 892, + 797 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nd _ {A C _ {0}} \\left(x _ {0}, x _ {1}\\right) = d _ {A C _ {0}} \\left(x _ {0} \\circ h, x _ {1} \\circ h\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 808, + 890, + 825 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In that case, the distance on $S$ is defined as the infimum over all possible reparametrization. For $X_0, X_1 \\in S$ ,", + "bbox": [ + 496, + 835, + 890, + 867 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\mathcal {S}} \\left(X _ {0}, X _ {1}\\right) := \\inf _ {h \\in \\operatorname {D i f f} ([ 0, 1 ])} d _ {A C _ {0}} \\left(x _ {0}, x _ {1} \\circ h\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 878, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4028", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/15bd225d8a661ba455ccb22c20f9ce4350d3a602c6cf25deb1378fec83b2dabe.jpg", + "image_caption": [ + "Figure 1: Geodesic paths between two 2D scaled spirals with different number of spins: SRVF $(1^{st}$ row) and SRC $(2^{nd}$ row)." + ], + "image_footnote": [], + "bbox": [ + 148, + 89, + 823, + 238 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following, we will denote with a dot the derivation with respect to the time variable, and with a prime the one with respect to the arc-length parameter.", + "bbox": [ + 75, + 301, + 470, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Square Root Velocity Framework", + "text_level": 1, + "bbox": [ + 76, + 356, + 370, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The square-root velocity function framework is the most commonly used representation for curve shape analysis in $\\mathbb{R}^d$ [24, 23]. The square-root velocity function (SRVF) of $x\\in AC_0([0,1],\\mathbb{R}^d)$ , denoted by $\\mathcal{R}_{\\mathrm{SRVF}}(x) = q$ , is defined as", + "bbox": [ + 76, + 378, + 470, + 450 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq (t) = \\frac {\\dot {x} (t)}{\\sqrt {\\| \\dot {x} (t) \\|}} = \\sqrt {\\dot {s} (t)} T (s (t)) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 450, + 468, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $T(s(t)) = \\frac{\\dot{x}(t)}{\\|\\dot{x}(t)\\|} = X'(s(t))$ is the unit tangent vector of the curve. This transformation is a bijection with $AC_0([0,1],\\mathbb{R}^d)$ and its explicit inverse is $x(t) = \\int_0^t q(u)|q(u)|du$ . As we consider length-normalized curves, the SRVFs have a unit $\\mathbb{L}^2$ norm, and their set is the convenient unit Hilbert sphere, a Riemannian submanifold of $\\mathbb{L}^2([0,1],\\mathbb{R}^d)$ (with the $\\mathbb{L}^2$ inner product). Then, the $\\mathbb{L}^2$ metric on SRVF induces a Riemannian metric on $AC_0([0,1],\\mathbb{R}^d)$ where geodesics are given by the shorter arcs on great circles between SRV functions. The action of $\\mathrm{Diff}_{+}([0,1])$ on $AC_0([0,1],\\mathbb{R}^d)$ is reflected on $q$ by the group action denoted by $*$ and defined as", + "bbox": [ + 75, + 491, + 468, + 679 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n(q * h) (t) = \\sqrt {\\dot {h} (t)} q (h (t)) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 689, + 468, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and if the curve is rotated by a matrix $O \\in SO(d)$ , its SRVF gets rotated by the same matrix. The key property of this representation is the invariance of its associated distance under the action of $\\mathrm{Diff}_{+}([0,1])$ and $SO(d)$ :", + "bbox": [ + 75, + 722, + 468, + 784 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| O \\left(q _ {0} * h\\right) - O \\left(q _ {1} * h\\right) \\right\\| _ {\\mathbb {L} ^ {2}} = \\left\\| q _ {0} - q _ {1} \\right\\| _ {\\mathbb {L} ^ {2}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 792, + 468, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The metric can then be used to define a proper distance on the shape space $\\mathcal{S}([0,1],\\mathbb{R}^d)$", + "bbox": [ + 75, + 819, + 468, + 851 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd_{\\mathcal{S}}^{(\\mathrm{SRVF})}(X_{0},X_{1}):= \\inf_{\\substack{O\\in SO(d)\\\\ h\\in \\mathrm{Diff}([0,1])}}\\cos^{-1}\\langle q_{0},O(q_{1}*h)\\rangle \\tag{7}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 859, + 468, + 897 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and the geodesic path on the shape space is taken between $q_0$ and $q_1 * h$ .", + "bbox": [ + 498, + 301, + 890, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The definition of this distance on the shape space under the SRVF representation can be interpreted as the following registration problem", + "bbox": [ + 498, + 332, + 890, + 377 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh ^ {*}, O ^ {*} = \\underset { \\begin{array}{c} O \\in S O (d) \\\\ h \\in \\operatorname {D i f f} ([ 0, 1 ]) \\end{array} } {\\arg \\min } \\int_ {0} ^ {1} \\| q _ {0} (t) - O q _ {1} (h (t)) \\sqrt {\\dot {h} (t)} \\| _ {2} ^ {2} d t. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 503, + 382, + 890, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In [3], this registration problem has been reformulated with the unit tangent vector and the arc length functions. By defining $\\gamma = s_{1} \\circ h \\circ s_{0}^{-1} \\in \\mathrm{Diff}_{+}([0,1])$ the optimization problem 8 amounts to finding the optimal diffeomorphism of $\\mathrm{Diff}_{+}([0,1])$ that acts on the arc-length parameter $s$ and solves the minimization problem:", + "bbox": [ + 498, + 441, + 890, + 532 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma^ {*}, O ^ {*} = \\underset { \\begin{array}{c} O \\in S O (d) \\\\ \\gamma \\in \\operatorname {D i f f} ([ 0, 1 ]) \\end{array} } {\\arg \\min } \\int_ {0} ^ {1} \\| T _ {0} (s) - O T _ {1} (\\gamma (s)) \\sqrt {\\gamma^ {\\prime} (s)} \\| _ {2} ^ {2} d s. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 498, + 539, + 890, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It should be noted, in this reformulation, that the object $OT_{1}(\\gamma(s))\\sqrt{\\gamma'(s)}$ does not represent the same shape as $X_{1}(s)$ in the shape space. Here the element $\\gamma$ of $\\mathrm{Diff}_{+}([0,1])$ is not used as a reparametrization of the curve but to deform the element of $S([0,1],\\mathbb{R}^{d})$ . Under this point of view, the set of unit tangent vectors that can be reached by deforming the vector $T(s)$ , with the group action $T*\\gamma$ defines an equivalence class of shapes associated with that one, as in the setting of deformable templates of Grenander's theory [29, 30].", + "bbox": [ + 496, + 598, + 890, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, the choice of a parametrized curve representation for shape analysis, discussed in [23], can be seen as the problem of choosing a good geometric representative of the shape as a template and defining an associated registration problem. Hence, an appropriate choice may be seen as a matter of modeling and should be done in interaction with the type of data analyzed and the dimension of the space. In the next sections, we use $h$ to refer to functions of $\\mathrm{Diff}_{+}([0,1])$ that act on the time variable $t$ and $\\gamma$ for ones that act on the arc-length variable $s$ .", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4029", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d3bf35c281d578c5961f542676569d9ea11d9beb808bbab0998bce4336337efb.jpg", + "image_caption": [ + "Figure 2: Geodesic paths between two scaled 3D circular helices with different number of spins: SRVF (1 $^{st}$ row) and SRC (2 $^{nd}$ row)." + ], + "image_footnote": [], + "bbox": [ + 166, + 85, + 803, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Exhaustive geometric information with Frenet representation", + "text_level": 1, + "bbox": [ + 76, + 344, + 470, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on these previous observations and with the intention of developing a more suitable framework for the analysis of three-dimensional curves, Brunel and Park [3] proposed a direct extension of the SRVF method, which considers not only the tangent vector as the geometric representation of the curve but the whole Frenet-Serret frame in three dimensions. Their idea is to use an exhaustive description of the geometry of curves by incorporating higher-order information about the geometry in the representation. To exploit this idea, we propose to study suitable representations based on this Frenet-Serret framework. Proofs of theoretical results are given in the supplementary material.", + "bbox": [ + 75, + 387, + 470, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. The Frenet-Serret framework", + "text_level": 1, + "bbox": [ + 76, + 577, + 341, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce the Frenet-Serret framework for curves of any dimension $d$ . Let $F([0,1],\\mathbb{R}^d)$ be the set of curves $x \\in AC_0([0,1],\\mathbb{R}^d)$ $d$ -times continuously differentiable, and with the first $d$ derivatives linearly independent. $F([0,1],\\mathbb{R}^d)$ is called the set of Frenet curves. In the following, we will restrict the shape space to be the set", + "bbox": [ + 76, + 599, + 468, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} ([ 0, 1 ], \\mathbb {R} ^ {d}) = F ([ 0, 1 ], \\mathbb {R} ^ {d}) / \\operatorname {D i f f} _ {+} ([ 0, 1 ]). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 700, + 468, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Frenet frame $e_1, e_2, \\ldots, e_d$ associated with $X \\in S([0,1],\\mathbb{R}^d)$ is uniquely defined by applying the Gram-Schmidt process to the first $d$ derivatives of $X$ . In dimension 3, the three vectors of the Frenet frame are known as the tangent, normal and bi-normal vector. We define the function $Q$ that maps to $s \\in [0,1]$ along the curve the corresponding Frenet frame", + "bbox": [ + 76, + 728, + 468, + 832 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ (s) = \\left[ e _ {1} (s) \\mid e _ {2} (s) \\mid \\dots \\mid e _ {d} (s) \\right]. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 843, + 468, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The function $Q$ is a measurable curve from $[0,1]$ to the Lie group of rotation matrices $SO(d)$ called the Frenet path.", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1 (Frenet-Serret equation [9]). Let $X \\in S([0,1],\\mathbb{R}^d)$ and $Q(s)$ its associated Frenet path. Then there are functions $\\theta_1, \\ldots, \\theta_{d-1}$ defined on that curve with $\\theta_1, \\ldots, \\theta_{d-2} > 0$ , so that every $\\theta_i$ is $(d-1-i)$ -times continuously differentiable and", + "bbox": [ + 498, + 345, + 893, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ ^ {\\prime} (s) = Q (s) A _ {\\boldsymbol {\\theta}} (s) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 434, + 890, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb {\\theta}(s) = (\\theta_{1}(s),\\dots ,\\theta_{d - 1}(s))^{T}$ and", + "bbox": [ + 500, + 462, + 777, + 479 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA _ {\\pmb {\\theta}} (s) = \\left[ \\begin{array}{c c c c c} 0 & - \\theta_ {1} (s) & 0 & \\dots & 0 \\\\ \\theta_ {1} (s) & 0 & - \\theta_ {2} (s) & \\ddots & \\vdots \\\\ 0 & \\theta_ {2} (s) & \\ddots & \\ddots & 0 \\\\ \\vdots & \\ddots & \\ddots & 0 & - \\theta_ {d - 1} (s) \\\\ 0 & \\dots & 0 & \\theta_ {d - 1} (s) & 0 \\end{array} \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 489, + 877, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $\\theta_{i}$ is called the $i$ -th Frenet curvature, and the equation is called the Frenet-Serret equation.", + "bbox": [ + 498, + 595, + 890, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In dimension 3, the two Frenet curvatures are known as $s \\mapsto \\kappa(s)$ the curvature function and $s \\mapsto \\tau(s)$ the torsion function. They have an interpretable physical meaning. The curvature function measures how sharply the curve changes direction at a given point, and the torsion function measures the degree to which the curve twists and turns as it moves along its path. The Frenet-Serret equation with an initial condition $Q(0) = Q_0$ defines an ordinary differential equation on the Lie group $SO(d)$ where the function $s \\mapsto A_\\theta(s)$ has values in the Lie algebra of skew-symmetric matrices. This equation can also be expressed in function of the time variable $t$ as", + "bbox": [ + 496, + 638, + 890, + 818 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {d Q (s (t))}{d t} = \\dot {s} (t) Q (s (t)) A _ {\\boldsymbol {\\theta}} (s (t)). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 828, + 890, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lemma 1. The Frenet curvatures and the Frenet path are invariant under all Euclidean motions.", + "bbox": [ + 498, + 869, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4030", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This means that for $x \\in F([0,1],\\mathbb{R}^d), O \\in SO(d), a \\in \\mathbb{R}^d$ and $h \\in \\mathrm{Diff}_{+}([0,1])$ , the curve defined by $\\tilde{x}(t) = a + Ox(h(t))$ has the same Frenet curvatures as $x$ and $\\tilde{Q}(s) = OQ(s)$ .", + "bbox": [ + 76, + 90, + 468, + 154 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 2 (Fundamental theorem of the local theory of curves [9]). Let $\\theta_1, \\ldots, \\theta_{d-1} \\in C^\\infty([0,1], \\mathbb{R})$ such that $\\theta_1, \\ldots, \\theta_{d-2} > 0$ . For a given $X_0 \\in \\mathbb{R}^d$ and $Q_0 \\in SO(d)$ there is a unique $X \\in S([0,1], \\mathbb{R}^d)$ parametrized by arc length and satisfying the following three conditions:", + "bbox": [ + 76, + 160, + 470, + 236 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$X(0) = X_0$", + "- $Q_{0}$ is the Frenet frame of $X$ at point $s = 0$ ,", + "- $\\theta_{1},\\ldots ,\\theta_{d - 1}$ are the Frenet curvatures of $X$" + ], + "bbox": [ + 94, + 244, + 395, + 308 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorems 1 and 2 state that there is a bijection, up to a translation and a rotation, between the shape space $\\mathcal{S}([0,1],\\mathbb{R}^d)$ , the set of admissible Frenet curvatures $\\mathcal{H}$ and the set of corresponding Frenet paths $\\mathcal{F}_0$ , where", + "bbox": [ + 76, + 325, + 468, + 387 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} = \\left\\{\\boldsymbol {\\theta} \\in C ^ {\\infty} ([ 0, 1 ], \\mathbb {R} ^ {d - 1}) | \\theta_ {1}, \\dots , \\theta_ {d - 2} > 0 \\right\\}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 395, + 468, + 414 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {0} = \\left\\{ \\begin{array}{c} Q \\in \\mathbb {L} ^ {2} ([ 0, 1 ], S O (d)) \\text {s u c h t h a t} \\\\ Q ^ {\\prime} (s) = Q (s) A _ {\\boldsymbol {\\theta}} (s), Q (0) = I _ {d}, \\boldsymbol {\\theta} \\in \\mathcal {H} \\end{array} \\right\\}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 443, + 468, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From the detailed Frenet-Serret framework, one can think of the direct extension of the square-root velocity function 4 that simply consists in replacing the tangent vector with the entire Frenet frame. The representation of a parametrized curve $x \\in F([0,1],\\mathbb{R}^d)$ will be then", + "bbox": [ + 76, + 497, + 468, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {Q} (x) (t) = \\sqrt {\\dot {s} (t)} Q (s (t)). \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 583, + 468, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This representation is used in [3] to define a new alignment method on $S$ . They extend the SRVF registration problem 9 by using the Frobenius distance between the Frenet frames instead of only the $\\mathbb{L}^2$ distance of the unit tangent vectors. They show to obtain more precise results with their method than the SRVF one. From the previous theorems, it is clear that this representation uniquely defines a parametrized curve $x \\in F([0,1],\\mathbb{R}^d)$ . However, as a result of the Frobenius theorem [12], the set of such representations $\\mathcal{R}_Q$ appears not to be a manifold. We leave the demonstration in the supplementary material.", + "bbox": [ + 76, + 611, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Unparametrized Frenet curvatures", + "text_level": 1, + "bbox": [ + 76, + 785, + 382, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A possible representation of a parametrized curve, already suggested in [23, 25, 20], which keeps the idea of encoding more geometric information, is the unparametrized Frenet curvatures and the arc-length function pair", + "bbox": [ + 76, + 809, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {\\boldsymbol {\\theta}} (x) (t) = \\left(\\sqrt {\\dot {s} (t)}, \\boldsymbol {\\theta} (s (t))\\right). \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 878, + 468, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We denote $\\Psi([0,1])$ , the set of square root velocity functions of length-normalized arc-length functions. This set is well-studied in the literature [15, 27]. It is the unit sphere of the Hilbert space $\\mathbb{L}^2([0,1],\\mathbb{R})$ and therefore a Riemannian manifold equipped with the $\\mathbb{L}^2$ metric. Then, the geodesic distance between two elements in $\\Psi([0,1])$ is", + "bbox": [ + 498, + 90, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\Psi} \\left(\\sqrt {\\dot {s} _ {0}}, \\sqrt {\\dot {s} _ {1}}\\right) = \\cos^ {- 1} \\left(\\left\\langle \\sqrt {\\dot {s} _ {0}}, \\sqrt {\\dot {s} _ {1}} \\right\\rangle\\right) \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 191, + 890, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the geodesic path connecting them is given by", + "bbox": [ + 500, + 224, + 831, + 239 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {\\Psi} (\\tau) = \\frac {\\sin ((1 - \\tau) \\vartheta)}{\\sin (\\vartheta)} \\sqrt {\\dot {s} _ {0}} + \\frac {\\sin (\\tau \\vartheta)}{\\sin (\\vartheta)} \\sqrt {\\dot {s} _ {1}} \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 250, + 890, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\vartheta = d_{\\Psi}(\\sqrt{\\dot{s}_0},\\sqrt{\\dot{s}_1})$ Moreover, any element of $\\mathrm{Diff}_{+}([0,1])$ is uniquely represented by an element of $\\Psi ([0,1])$", + "bbox": [ + 500, + 292, + 890, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 1. The set of Frenet curvatures $\\mathcal{H}$ is a Riemannian submanifold of $\\mathbb{L}^2([0,1],\\mathbb{R}^{d-1})$ .", + "bbox": [ + 498, + 347, + 890, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. The set $M = \\{x \\in \\mathbb{R}^{d-1} | x_1, \\ldots, x_{d-2} > 0\\}$ is a open subset of the Riemannian manifold $\\mathbb{R}^{d-1}$ . Then it is itself a differentiable Riemannian manifold with the standard inner product of $\\mathbb{R}^{d-1}$ , and for any point $p \\in M$ the tangent space $T_p(M)$ is $\\mathbb{R}^{d-1}$ . The set of Frenet curvatures $\\mathcal{H}$ is the set of measurable curves from [0, 1] to the Riemannian manifold $M$ and thus also a manifold ([26]). Its tangent space is $\\mathbb{L}^2([0,1], \\mathbb{R}^{d-1})$ and it can be equipped with the $\\mathbb{L}^2$ Riemannian metric.", + "bbox": [ + 498, + 383, + 890, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consequently, the geodesic distance on $\\mathcal{H}$ is simply the $\\mathbb{L}^2$ norm", + "bbox": [ + 500, + 530, + 890, + 560 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\mathcal {H}} \\left(\\boldsymbol {\\theta} _ {0}, \\boldsymbol {\\theta} _ {1}\\right) = \\left\\| \\boldsymbol {\\theta} _ {0} - \\boldsymbol {\\theta} _ {1} \\right\\| _ {\\mathbb {L} ^ {2}} \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 571, + 890, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the geodesic path is the straight line connecting them", + "bbox": [ + 500, + 598, + 877, + 613 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {\\mathcal {H}} (\\tau) = (1 - \\tau) \\boldsymbol {\\theta} _ {0} + \\tau \\boldsymbol {\\theta} _ {1}. \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 625, + 890, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 2. The map $\\mathcal{R}_{\\theta}: F([0,1], \\mathbb{R}^d) \\to \\Psi([0,1]) \\times \\mathcal{H}$ , defined above, is a bijection.", + "bbox": [ + 500, + 648, + 888, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. The element of $\\Psi([0,1])$ uniquely defines the arclength function by $s(t) = \\int_0^t (\\sqrt{\\dot{s}(u)})^2 du$ . As mentioned before, we have a bijection between the unparametrized Frenet curvatures in $\\mathcal{H}$ and the unparametrized curve in the shape space $\\mathcal{S}$ (1, 2). Then, from $X \\in \\mathcal{S}$ , the initial parametrized curve is simply $x(t) = X(s(t))$ .", + "bbox": [ + 498, + 688, + 890, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The set of such $\\mathcal{R}_{\\theta}$ is the Cartesian product of $\\Psi ([0,1])$ and $\\mathcal{H}$ and, therefore, is also a Riemannian manifold equipped with the product metric $d_{\\Psi}\\oplus d_{\\mathcal{H}}$ [16]. The induced metric on $F([0,1],\\mathbb{R}^d)$ under the representation $\\mathcal{R}_{\\theta}$ is", + "bbox": [ + 498, + 790, + 890, + 864 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\boldsymbol {\\theta}} \\left(x _ {0}, x _ {1}\\right) = d _ {\\Psi} \\left(\\sqrt {\\dot {s} _ {0}}, \\sqrt {\\dot {s} _ {1}}\\right) + d _ {\\mathcal {H}} \\left(\\boldsymbol {\\theta} _ {0}, \\boldsymbol {\\theta} _ {1}\\right). \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 872, + 890, + 897 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4031", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to define a distance on the shape space $S$ from that one, we must quotient out the space $\\mathrm{Diff}_{+}([0,1])$ . The action of $\\mathrm{Diff}_{+}([0,1])$ on $\\Psi ([0,1])$ is the same as 5, and the Frenet curvatures are invariant under reparametrization of the corresponding parametrized curve (Lemma 1). Moreover, by simply taking $h^* = s_1^{-1} \\circ s_0 \\in \\mathrm{Diff}_{+}([0,1])$ , we have $\\sqrt{s_1} * h^* = \\sqrt{s_0}$ , and thus the distance on $\\Psi ([0,1]) / \\mathrm{Diff}_{+}([0,1])$ between $s_0$ and $s_1$ is zero. Hence, the induced distance on the shape space, between $X_0, X_1 \\in S$ , under the representation $\\mathcal{R}_{\\theta}$ is defined as", + "bbox": [ + 76, + 90, + 472, + 243 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\mathcal {S}} ^ {(\\boldsymbol {\\theta})} \\left(X _ {0}, X _ {1}\\right) := d _ {\\mathcal {H}} \\left(\\boldsymbol {\\theta} _ {0}, \\boldsymbol {\\theta} _ {1}\\right) = \\left\\| \\boldsymbol {\\theta} _ {0} - \\boldsymbol {\\theta} _ {1} \\right\\| _ {\\mathbb {L} ^ {2}} \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 253, + 468, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and the geodesic path connecting them is", + "bbox": [ + 76, + 287, + 352, + 303 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {\\mathcal {S}} ^ {(\\boldsymbol {\\theta})} (\\tau) = \\left(\\sqrt {\\dot {s} _ {0}}, \\alpha_ {\\mathcal {H}} (\\tau)\\right). \\tag {24}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 314, + 468, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This immediate representation by the Frenet curvatures appears in the experiments not to be sufficiently elastic (Figure 5). It has somewhat the same weakness as the angle representation proposed in [23], the Frenet curvatures being already independent of the parametrization.", + "bbox": [ + 76, + 351, + 468, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Square Root Curvatures Transform", + "text_level": 1, + "bbox": [ + 76, + 438, + 387, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To overcome the \"non-elasticity\" issue of the representation defined above, we propose a second framework for shape analysis based on Frenet curvatures which uses, like the square root velocity function, the parametrization as a tool to register the curves and define a more \"elastic\" method. The latter is inspired by the square-root velocity transform of $SO(d)$ -valued curves.", + "bbox": [ + 76, + 462, + 468, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Definition 1 (SRV Transform for curves on $SO(d)$ ). Let $P \\in C^{\\infty}([0,1], SO(d))$ . The Square Root Velocity transform of $P$ is the map", + "bbox": [ + 75, + 579, + 468, + 626 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nq (P) (t) = \\frac {L _ {P (t) ^ {- 1}} \\dot {P} (t)}{\\sqrt {\\| \\dot {P} (t) \\| _ {F}}} = \\frac {P (t) ^ {T} \\dot {P} (t)}{\\sqrt {\\| \\dot {P} (t) \\| _ {F}}}, \\tag {25}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 637, + 468, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\| .\\| _F$ is the Frobenius norm associated with the scalar product on the Lie Algebra of skew-symmetric matrices $\\langle A,B\\rangle = \\frac{1}{2} tr(A^T B) = -\\frac{1}{2} tr(AB)$ .", + "bbox": [ + 76, + 695, + 468, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Let $x \\in F([0,1],\\mathbb{R}^d)$ and $Q(t) \\in C^\\infty([0,1],SO(d))$ be its associated Frenet path. Using the Frenet-Serret differential equation, the SRV Transform of the Frenet path is", + "bbox": [ + 76, + 752, + 468, + 799 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nq (Q) (t) = \\sqrt {\\dot {s} (t)} \\frac {A _ {\\boldsymbol {\\theta}} (s (t))}{\\sqrt {\\left\\| A _ {\\boldsymbol {\\theta}} (s (t)) \\right\\| _ {F}}}. \\tag {26}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 811, + 468, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proposition 3. Let $\\theta \\in \\mathcal{H}$ , we have", + "bbox": [ + 76, + 856, + 316, + 871 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| A _ {\\boldsymbol {\\theta}} (s (t)) \\right\\| _ {F} = \\left\\| \\boldsymbol {\\theta} (s (t)) \\right\\| _ {2}.\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 883, + 367, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on the SRV Transform of a Frenet path and Proposition 3, we propose a new transformation of a parametrized curve, which we have called the Square-Root Curvatures (SRC) transform.", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Definition 2 (Square-Root Curvatures Transform). Let $x \\in F([0,1],\\mathbb{R}^d)$ . We consider its associated arc-length function $s(t)$ and Frenet curvatures $\\pmb{\\theta}(s(t))$ defined as in Theorem 1. Then we define its square-root curvatures transform to be the map", + "bbox": [ + 498, + 160, + 893, + 236 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nc (t) = \\sqrt {\\dot {s} (t)} \\frac {\\boldsymbol {\\theta} (s (t))}{\\sqrt {\\| \\boldsymbol {\\theta} (s (t)) \\|}}. \\tag {27}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 244, + 890, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The set of such square-root curvatures transforms is", + "bbox": [ + 519, + 289, + 862, + 304 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} = \\left\\{c \\in \\mathbb {L} ([ 0, 1 ], \\mathbb {R} ^ {d - 1}) \\mid c _ {1}, \\dots , c _ {d - 2} > 0 \\right\\}, \\tag {28}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 314, + 890, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which is the same as the set of admissible Frenet curvatures $\\mathcal{H}$ . We have already shown in the previous section that this set is a Riemannian manifold equipped with the $\\mathbb{L}^2$ metric. Therefore, the geodesic distance between $c_{0}, c_{1} \\in \\mathcal{C}$ is the $\\mathbb{L}^2$ distance between them, and the geodesic path is a straight line. We define the following representation of a parametrized curve $x \\in F([0,1], \\mathbb{R}^d)$ , from its Square-Root Curvatures transform, by", + "bbox": [ + 498, + 342, + 890, + 463 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {\\mathrm {S R C}} (x) (t) = \\left(\\sqrt {\\dot {s} (t)}, c (t)\\right). \\tag {29}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 472, + 890, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proposition 4. The map $\\mathcal{R}_{SRC} : F([0,1],\\mathbb{R}^d) \\to \\Psi([0,1]) \\times \\mathcal{C}$ , defined above, is a bijection.", + "bbox": [ + 500, + 508, + 890, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proof. This is again a result of theorems 1 and 2. To get $x$ from $\\mathcal{R}_{\\mathrm{SRC}}(x)$ , it should be noted firstly that $c(t) \\| c(t) \\| = \\dot{s}(t) \\theta(s(t))$ . From that, the skew-symmetric matrix function of the Frenet-Serret ODE can be reconstructed. By solving the corresponding Frenet-Serret ODE one gets the associated time parametrized Frenet path $Q(t)$ . Then, using the first component of $\\mathcal{R}_{\\mathrm{SRC}}(x)$ , we get $x(t) = X(s(t)) = \\int_0^t \\dot{s}(u) T(s(u)) du$ .", + "bbox": [ + 498, + 547, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The set of such square root curvature representations $\\mathcal{R}_{\\mathrm{SRC}}$ is the Cartesian product $\\Psi ([0,1])\\times \\mathcal{C}$ and therefore a Riemannian manifold with the product metric $d_{\\Psi}\\oplus d_{\\mathcal{C}}$ . This representation is, by definition, invariant under the action of $SO(d)$ . Then, the corresponding shape space is the quotient space $\\Psi ([0,1])\\times \\mathcal{C} / \\mathrm{Diff}([0,1])$ . Let's $x\\in F([0,1],\\mathbb{R}^d)$ and $h\\in \\mathrm{Diff}([0,1])$ . The SRC representation of $\\tilde{x} = x\\circ h$ is", + "bbox": [ + 498, + 681, + 890, + 787 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {\\mathrm {S R C}} (\\tilde {x}) = \\left(\\sqrt {\\dot {s}} * h, c * h\\right) = \\mathcal {R} _ {\\mathrm {S R C}} (x) * h \\tag {30}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 796, + 890, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $*$ is the group action defined in 5.", + "bbox": [ + 500, + 830, + 767, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proposition 5. The metric on $F([0,1],\\mathbb{R}^d)$ induced by the Riemannian metric on $\\Psi ([0,1])\\times \\mathcal{C}$ defined by $d_{SRC}\\coloneqq d_{\\Psi}\\oplus d_{\\mathcal{C}}$ is invariant under the action of $\\text{Diff}_{+}([0,1])$ .", + "bbox": [ + 498, + 854, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4032", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The distance on the shape space $\\mathcal{S}$ under the representation $\\mathcal{R}_{\\mathrm{SRC}}$ , between two elements $X_0, X_1 \\in \\mathcal{S}$ , is defined as", + "bbox": [ + 76, + 90, + 468, + 135 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nd _ {S} ^ {\\left(\\mathrm {S R C}\\right)} \\left(X _ {0}, X _ {1}\\right) := \\inf _ {h \\in \\operatorname {D i f f} _ {+} ([ 0, 1 ])} d _ {\\mathrm {S R C}} \\left(x _ {0}, x _ {1} \\circ h\\right). \\tag {31}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 145, + 468, + 172 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From the optimal wrapping function $h^*$ the geodesic path on $S$ between them is", + "bbox": [ + 76, + 185, + 468, + 214 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {\\mathcal {S}} ^ {(\\mathrm {S R C})} (\\tau) = \\left( \\begin{array}{c} \\frac {\\sin ((1 - \\tau) \\vartheta)}{\\sin (\\vartheta)} \\sqrt {\\dot {s} _ {0}} + \\frac {\\sin (\\tau \\vartheta)}{\\sin (\\vartheta)} \\left(\\sqrt {\\dot {s} _ {1}} * h ^ {*}\\right), \\\\ (1 - \\tau) c _ {0} + \\tau \\left(c _ {1} * h ^ {*}\\right)) \\end{array} \\right. \\tag {32}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 223, + 468, + 276 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\vartheta = d_{\\Psi}(\\sqrt{\\dot{s}_0},\\sqrt{\\dot{s}_1}*h^*)$ . The registration problem consider here is to find the minimizer $h^*$ over $\\mathrm{Diff}_{+}([0,1])$ of", + "bbox": [ + 76, + 276, + 468, + 321 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\int_ {0} ^ {1} \\| c _ {0} (t) - \\left(c _ {1} * h\\right) (t) \\| ^ {2} + \\left\\| \\sqrt {\\dot {s} _ {0} (t)} - \\left(\\sqrt {\\dot {s} _ {1}} * h\\right) (t) \\right\\| ^ {2} d t. \\tag {33}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 330, + 475, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Using the reformulation principle of [3], that is $\\gamma = s_1\\circ h\\circ$ $s_0^{-1}\\in \\mathrm{Diff}_+([0,1])$ , this registration problem is shown to be equivalent to finding $\\gamma^{*}\\in \\mathrm{Diff}_{+}([0,1])$ that minimizes", + "bbox": [ + 76, + 392, + 468, + 439 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\int_ {0} ^ {1} \\left\\| \\frac {\\boldsymbol {\\theta} _ {0} (s)}{\\sqrt {\\lVert \\boldsymbol {\\theta} _ {0} (s) \\rVert}} - \\sqrt {\\gamma^ {\\prime} (s)} \\frac {\\boldsymbol {\\theta} _ {1} (\\gamma (s))}{\\sqrt {\\lVert \\boldsymbol {\\theta} _ {1} (\\gamma (s)) \\rVert}} \\right\\| ^ {2} \\\\ + \\| 1 - \\gamma^ {\\prime} (s) \\| ^ {2} d s. \\tag {34} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 465, + 468, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that this reformulation has the form of a penalized registration problem. The second term represents a penalty term on $\\gamma$ and ensures a certain smoothness of the warping function. In this framework, the deformable templates are the square-root normalized curvatures which encode more geometric information than the unit tangent vector.", + "bbox": [ + 76, + 540, + 468, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 645, + 209, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we report the experimental results of the proposed methods, comparing them with the SRVF method. We use both synthetic and real data. Additional results and figures are available in the supplementary material.", + "bbox": [ + 76, + 670, + 468, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Statistical estimation of the Frenet curvatures", + "text_level": 1, + "bbox": [ + 76, + 741, + 462, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The main limitation of shape analysis methods based on the Frenet curvatures is the need of additional estimates of curvatures. Being dependent on higher-order derivatives (up to order $d$ ), they are quite sensitive to the observation noise of the Euclidean curve. We detail here a simple method that can be used for their smooth estimation, and we refer to [21, 18] for a more complex and detailed statistical estimation algorithm. First, it is possible to use a local polynomial smoothing algorithm to estimate the $d$ first", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "derivatives of the Euclidean curve [6]. From these derivatives, the raw estimates of the Frenet curvatures can be computed by using their extrinsic formulas. We propose here a second method based on the Frenet-Serret ODE approximation, to obtain the raw estimates, that appear to be more stable. A simple middle point approximation of the ODE solution gives", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nQ \\left(s _ {j}\\right) \\approx Q \\left(s _ {i}\\right) \\exp \\left(\\left(s _ {i} - s _ {j}\\right) A _ {\\boldsymbol {\\theta}} \\left(\\frac {s _ {i} + s _ {j}}{2}\\right)\\right). \\tag {35}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 207, + 890, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\exp(.)$ is the exponential map of the Lie group $SO(d)$ . Then, using the inverse logarithm map $\\log(.)$ , this gives an approximation of the matrix $A_{\\theta}((s_i + s_j) / 2) \\approx \\frac{1}{s_i - s_j} \\log(Q(s_i)^T Q(s_j))$ and by identification raw estimates of $\\pmb{\\theta}((s_i + s_j) / 2)$ . As we consider here a problem of estimating a functional parameter, we formulate the final $\\pmb{\\theta}$ estimation problem as a penalized weighted functional regression with the obtained raw estimates, that we solve by using a B-spline approximation of $\\pmb{\\theta}$ .", + "bbox": [ + 496, + 252, + 890, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Experiments with synthetic curves", + "text_level": 1, + "bbox": [ + 500, + 401, + 799, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use synthetic data to highlight the differences between the methods discussed above (SRVF, SRC, and Frenet curvatures). The computations related to the SRVF method are made with the package fdasrsf. The SRC and Frenet curvatures methods are implemented with the code provided as supplementary material, including a dynamic programming algorithm for solving the registration problems.", + "bbox": [ + 496, + 424, + 890, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We consider the simple case of a set of 20 curves in $\\mathbb{R}^2$ with a single large peak of curvature. This one is created by generating curvature functions on $[0,1]$ with one peak of maximum value 60.5, width 0.15, and location chosen randomly between 0.1 and 0.9. These curves have the shape of a loop made with a wire, where the loop is more or less close to the right or left wire end, depending on the location of the curvature peak.", + "bbox": [ + 496, + 545, + 890, + 667 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cbc1abda0254f2cc9fe24155af5d9724f1ef96390b3e05fd1c98b5544b7308a7.jpg", + "image_caption": [ + "Figure 3: Matrices of pairwise SRVF distance (left), SRC distance (middle), and unparametrized Frenet curvatures distance (right) by sorted location of the peak curvature." + ], + "image_footnote": [], + "bbox": [ + 501, + 676, + 888, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare the three methods through the pairwise distance matrices in Figure 3, and the geodesic paths computed", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4033", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5fc0e004decb1b717e2aff42397bce2527a270eba077b5f9c81e20677126ddeb.jpg", + "image_caption": [ + "Figure 4: SRVF $(1^{st}$ row), SRC $(2^{nd}$ row), and Frenet curvatures $(3^{rd}$ row) geodesic paths between curves with curvature peaks located at 0.27 (left) and 0.78 (right)." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 465, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a8a764d07a6fe0f604a75abe30c4ddf12a14d9dd63c2068fe0f1915570ec2bc.jpg", + "image_caption": [ + "Figure 5: Curvatures of the Euclidean curves along the geodesic paths plotted in Figure 4: SRVF $(1^{st}$ row), SRC $(2^{nd}$ row), and Frenet curvatures $(3^{rd}$ row)." + ], + "image_footnote": [], + "bbox": [ + 78, + 321, + 467, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "between two of these curves in Figure 4, with peaks located at 0.27 (red curve) and at 0.78 (blue curve). The corresponding deformations through the variations of the curvature along the different geodesic paths on Euclidean curves are shown in Figure 5, which highlights the strengths and weaknesses of each method. First, it emphasizes the \"nonelasticity\" of the unparametrized Frenet curvatures method, as in the middle of the geodesic path, we have two peaks of curvature and, therefore, a completely different shape without any loop. This explains the inconsistency of the heatmap under this method. Conversely, there is an elastic deformation of the curvature with the SRC transform, and shapes along the geodesic are consistent with the set of curves considered, which is well summarised on the corresponding heatmap where all distances are rather close to zero. For the SRVF method, the chosen example with peaks of curvature that are quite far apart shows that artifacts appear along the geodesic; the middle curve has two small loops at the edges. This phenomenon gives unreliable distances, as shown in the heatmap, where the distances are not monotone as a function of the spacing between the curvature peaks.", + "bbox": [ + 75, + 521, + 468, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By considering a set of curves characterized by specific features, we observe a clear difference in consistency of the shapes along the geodesics with respect to the different", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "methods. This phenomenon is quite visible on the geodesic between helices in 2D, Figure 1 or 3D, Figure 2. In that case, within both geodesic paths under SRVF method, the curves lose the characteristic geometry of the helix. A three-dimensional circular helix is characterized by having a constant curvature and torsion, which is not the case along the SRVF geodesic, but preserved with the SRC method. In that case the geodesic under the Frenet curvature representation is very similar to SRC.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Application to sign language motion data", + "text_level": 1, + "bbox": [ + 500, + 236, + 849, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It appears that curvilinear velocity and Frenet curvatures are particularly relevant parameters for the analysis of human motion. Several laws involving these parameters can be found in the literature [10, 14, 19, 7]; among others, the power laws state a special relationship between the curvature, the torsion, and the velocity of a point trajectory representing human motion. Using a method that conserves the shape of these parameters is, therefore, of particular interest in this application. We demonstrate here with the case of wrist trajectories in sign language, acquired with a motion capture system by the company MocapLab (https://www.mocaplab.com/fr/). We compute the geodesic paths, under each of the frameworks, between the arbitrarily chosen red and blue curves within the set of several repetitions of the sign \"Femme\", shown in Figure 6 with the corresponding time-parameterized Frenet curvatures.", + "bbox": [ + 496, + 258, + 890, + 500 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/12031780e212557ad881f4993741d2ee443aaaf43a8f0ee4cb443105d3e87eb4.jpg", + "image_caption": [ + "Figure 6: Trajectories of the right wrist while signing \"Femme\" in sign language: 3D curves (left), time-parametrized curvatures (top right), torsions (bottom right). The blue and red ones are used to compute the geodesic in Figure 7." + ], + "image_footnote": [], + "bbox": [ + 506, + 513, + 890, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 7 emphasizes the advantage of considering a representation depending on the parameterization, allowing a registration before computing the geodesic. However, it also shows that considering only the tangent vector as a representative object (SRVF) is not sufficient to find the optimal reparametrization that correctly aligns the torsions,", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4034", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d253497a348fef4db1a60db1dab9aaa3da2452f6ac63983d86347dfd6a6ea21b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 89, + 349, + 196 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/24d15a2e1a679858767a95bfd2d27b4406775c97261d47fbaa8ae71f4cb7ff39.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 90, + 607, + 198 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/92e3396783fc03519b46608c579f4863a1642dd9b2d192bc286e3092cb2e3094.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 90, + 870, + 198 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/28b92fb8e0014377beb36b6b3827aa6da9697bc4ae4a51c603e6096c8b6f85e5.jpg", + "image_caption": [ + "(a) SRVF" + ], + "image_footnote": [], + "bbox": [ + 101, + 199, + 348, + 305 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d2655f7d998c4b48da89d11ae702c92a33e690d35bc6e453b57ac33e133d43a8.jpg", + "image_caption": [ + "(b) SRC" + ], + "image_footnote": [], + "bbox": [ + 361, + 200, + 607, + 306 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/355efc6332217c06be6fc0cde171206efd88da0d6e977d02b9b8971691864cf5.jpg", + "image_caption": [ + "(c) Frenet curvatures" + ], + "image_footnote": [], + "bbox": [ + 620, + 200, + 867, + 305 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/7940fa0bb2ddd7117f5c3e07cf8db4e39c36103725fcc4d71cfb7725b3fef637.jpg", + "image_caption": [ + "Figure 7: Comparison between time-parametrized curvature and torsion along the geodesic path under SRVF (left), SRC (middle), and Frenet curvatures (right).", + "(a) Warping functions $h$", + "Figure 8: Comparison of estimated warping functions $h$ (left) and $\\gamma$ (right) to compute the geodesic in Figure 7." + ], + "image_footnote": [], + "bbox": [ + 106, + 406, + 263, + 512 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/386b59cb3851d2575d791fcbdca101540e1cd500a4e70bcdd85b65cc802647e0.jpg", + "image_caption": [ + "(b) Warping functions $\\gamma$" + ], + "image_footnote": [], + "bbox": [ + 281, + 406, + 442, + 512 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "and could affect subsequent analyses made by the SRVF method, such as the mean. This results in the appearance of new minimums, maximums and zeros in the torsion functions along the SRVF geodesic. However, such characteristic points of the curvature, torsion and velocity functions are crucial in the observation of the laws of motion. It is therefore preferable to use a method that preserves these characteristics by directly optimizing the optimal alignment from these parameters, such as the proposed SRC method.", + "bbox": [ + 75, + 617, + 470, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 768, + 194, + 784 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The square-root curvature transform of a Euclidean curve in $\\mathbb{R}^d$ is a representation that encodes more geometric information of the curves, and thus the results are easier to interpret than existing methods. The main limitation lies in the estimation of the Frenet curvatures from real and noisy data, nevertheless recent smooth statistical estimators can be used for computing the SRC [21, 18]. We believe our", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "method is particularly interesting for motion trajectory analysis and could be developed further in the future as a tool for generation, segmentation and classification of complex trajectories.", + "bbox": [ + 498, + 409, + 892, + 469 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 486, + 596, + 501 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Martin Bauer, Nicolas Charon, Eric Klassen, and Alice Le Brigant. Intrinsic riemannian metrics on spaces of curves: Theory and computation. Handbook of Mathematical Models and Algorithms in Computer Vision and Imaging, pages 1-35, 2021. 1", + "[2] Martin Bauer, Nicolas Charon, Eric Klassen, Sebastian Kurtek, Tom Needham, and Thomas Pierron. Elastic Metrics on Spaces of Euclidean Curves: Theory and Algorithms, 2022. arXiv:2209.09862 [math]. 1, 2", + "[3] Nicolas J.-B. Brunel and Juhyun Park. The frenet-serret framework for aligning geometric curves. In Frank Nielsen and Frédéric Barbaresco, editors, Geometric Science of Information, pages 608-617, Cham, 2019. Springer International Publishing. 3, 4, 5, 7", + "[4] Elena Celledoni, Markus Eslitzbichler, and Alexander Schmeding. Shape analysis on lie groups with applications in computer animation. Journal of Geometric Mechanics, 8(3):273-304, 2016. 2", + "[5] Marvin Eisenberger and Daniel Cremers. Hamiltonian dynamics for real-world shape interpolation. In European Conference on Computer Vision (ECCV), 2020. 1", + "[6] Jianqing Fan and Irene Gijbels. Local polynomial modelling and its applications. CRC monographs on statistics and applied probability 66. Chapman & Hall, 1996. 7", + "[7] Tamar Flash and Alain Berthoz. *Space-Time Geometries for Motion and Perception in the Brain and the Arts*. Lecture Notes in Morphogenesis. Springer, 2021. 8" + ], + "bbox": [ + 506, + 512, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4035", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[8] Lukas Koestler, Daniel Grittner, Michael Moeller, Daniel Cremers, and Zorah Lähner. Intrinsic neural fields: Learning functions on manifolds. In European Conference on Computer Vision (ECCV), 2022. 1", + "[9] Wolfgang Kühnel. Differential Geometry Curves - Surfaces, volume 77. 4, 5", + "[10] Francesco Lacquaniti, Carlo Terzuolo, and Paolo Viviani. The law relating the kinematic and figural aspects of drawing movements. Acta Psychologica (Amst.), 54:115-130, 1983. 8", + "[11] Serge Lang. Differential and riemannian manifolds. Springer, 1999:1-6, 2006. 1, 2", + "[12] Robert Lyons. Frobenius theorem two ways. Lecture note, 2016. 5", + "[13] Zorah Lähner, Emanuele Rodola, Frank R. Schmidt, Michael M. Bronstein, and Daniel Cremers. Efficient globally optimal 2d-to-3d deformable shape matching. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), May 2016. 1", + "[14] Uri Maoz, Alain Berthoz, and Tamar Flash. Complex unconstrained three-dimensional hand movement and constant equi-affine speed. Journal of Neurophysiology, 101:1002-1015, 2009. 8", + "[15] J. S. Marron, James O. Ramsay, Laura M. Sangalli, and Anuj Srivastava. Functional data analysis of amplitude and phase variation. Statistical Science, 30:468-484, 2015. 5", + "[16] Turk J Math. Submanifolds of Riemannian Product Manifolds. 29:389-401, 2005. 5", + "[17] Tom Needham. Shape Analysis of Framed Space Curves. Journal of Mathematical Imaging and Vision, 61(8):1154-1172, Oct. 2019. 1", + "[18] Juhyun Park, Nicolas Brunel, and Perrine Chassat. Curvature and torsion estimation of 3d functional data: A geometric approach to build the mean shape under the frenet serret framework, 2022. arXiv:2203.02398 [stat]. 7, 9", + "[19] Frank E. Pollick, Uri Maoz, Amir A. Handzel, Peter J. Giblin, Guillermo Sapiro, and Tamar Flash. Three-dimensional arm movements at constant equi-affine speed. Cortex, 45:325-339, 2009. 8", + "[20] Marianna Saba. On the usage of the curvature for the comparison of planar curves. PhD thesis, University of Cagliari, 2012. 2, 5", + "[21] Laura M Sangalli, Piercesare Secchi, Simone Vantini, and Alessandro Veneziani. Efficient estimation of three-dimensional curves and their derivatives by free-knot regression splines, applied to the analysis of inner carotid artery centrelines. Journal of the Royal Statistical Society: Series C (Applied Statistics), 58(3):285-306, 2009. 7, 9", + "[22] Stefan Sommer, Tom Fletcher, and Xavier Pennec. Introduction to differential and Riemannian geometry. 2020. 2", + "[23] Anuj Srivastava and Eric Klassen. Functional and Shape Data Analysis. Springer Series in Statistics. Springer New York, 2016. 1, 2, 3, 5, 6", + "[24] Anuj Srivastava, Eric Klassen, Shantanu H. Joshi, and Ian H. Jermyn. Shape analysis of elastic curves in euclidean spaces. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(7):1415-1428, 2011. 1, 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Tatiana Surazhsky and Gershon Elber. Metamorphosis of planar parametric curves via curvature interpolation. International Journal of Shape Modeling, 8:201-216, 2002. 2, 5", + "[26] Wang Tixiang. Morse theory on banach manifolds. Acta Mathematica Sinica, 5:250-262, 1989. 5", + "[27] J. Derek Tucker, Wei Wu, and Anuj Srivastava. Generative models for functional data using phase and amplitude separation. Computational Statistics and Data Analysis, 61:50-66, 2013. 5", + "[28] Laurent Younes. Computable Elastic Distances between Shapes. SIAM Journal on Applied Mathematics, 58(2):565-586, 1998. Publisher: Society for Industrial and Applied Mathematics. 1", + "[29] Laurent Younes. *Shapes and Diffeomorphisms*. Applied Mathematical Sciences. Springer, 2010. 3", + "[30] Laurent Younes. Elastic distance between curves under the metamorphosis viewpoint, 2018. arXiv:1804.10155. 3" + ], + "bbox": [ + 501, + 90, + 890, + 347 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4036", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Shape Analysis of Euclidean Curves under Frenet-Serret Framework/235fdcf4-691b-437b-8ee8-c7b76395f3ab_model.json b/2023/Shape Analysis of Euclidean Curves under Frenet-Serret Framework/235fdcf4-691b-437b-8ee8-c7b76395f3ab_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d58f5f6b733b2d441476af67df94d001f4ba0ba2 --- /dev/null +++ b/2023/Shape Analysis of Euclidean Curves under Frenet-Serret Framework/235fdcf4-691b-437b-8ee8-c7b76395f3ab_model.json @@ -0,0 +1,2398 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.131, + 0.834, + 0.153 + ], + "angle": 0, + "content": "Shape Analysis of Euclidean Curves under Frenet-Serret Framework" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.181, + 0.393, + 0.234 + ], + "angle": 0, + "content": "Perrine Chassat1 \n1LaMME, University of Paris-Saclay \nperrine.chassat@univ-evry.fr" + }, + { + "type": "text", + "bbox": [ + 0.432, + 0.181, + 0.62, + 0.234 + ], + "angle": 0, + "content": "Juhyun Park\\(^{1,2}\\) \n\\(^{2}\\)ENSIIE, Evry \njuhyun.park@ensiie.fr" + }, + { + "type": "text", + "bbox": [ + 0.656, + 0.181, + 0.87, + 0.234 + ], + "angle": 0, + "content": "Nicolas Brunel1,2,3 \n3Quantmetry, Paris \nnicolas.brunel@ensiie.fr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.575 + ], + "angle": 0, + "content": "Geometric frameworks for analyzing curves are common in applications as they focus on invariant features and provide visually satisfying solutions to standard problems such as computing invariant distances, averaging curves, or registering curves. We show that for any smooth curve in \\(\\mathbb{R}^d\\), \\(d > 1\\), the generalized curvatures associated with the Frenet-Serret equation can be used to define a Riemannian geometry that takes into account all the geometric features of the shape. This geometry is based on a Square Root Curvature Transform that extends the square root-velocity transform for Euclidean curves (in any dimensions) and provides likely geodesics that avoid artefacts encountered by representations using only first-order geometric information. Our analysis is supported by simulated data and is especially relevant for analyzing human motions. We consider trajectories acquired from sign language, and show the interest of considering curvature and also torsion in their analysis, both being physically meaningful." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.604, + 0.21, + 0.619 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Identifying and comparing different types of visual objects is a fundamental task in machine learning and computer vision problems [13, 5, 8]. The shape is one of the essential features of objects that allow us to understand and characterize them. Nowadays, it is much easier to obtain data in the form of shapes, typically as dense point clouds or landmarks. The main task in shape analysis is to define a proper framework to compare and quantify the variation of the shapes. However, the shape space is generally nonlinear, and extracting meaningful information or features is complex. One of the successful approaches to shape analysis utilizes a Riemannian framework of differential geometry, where a metric can be defined between the shapes, which is invariant with respect to shape-preserving transformations such as translation and rotation. For instance, this gives rise to geodesic distances that are naturally invariant to smooth and optimal deformations through geodesic paths between the shapes [11]. This approach is very versatile as it can" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.895, + 0.71 + ], + "angle": 0, + "content": "be adapted to various kinds of manifold-value data and can be designed to emphasize important geometric information to be preserved. As a consequence, several choices of metrics are possible, such as the class of invariant Sobolev metrics, often called elastic, for the analysis of curves [1]. In this work, we are concerned with curves that often arise in the application as trajectories (function of time) or motions (animation, activity recognition) and with the definition of a framework to compare their shapes. The differential geometry of Euclidean curves is among the simplest (with respect to higher dimensional manifolds), and relatively simple Riemannian metrics are available with different mathematical representations of curves [28]. Quite remarkably, the introduction of the Square Root Velocity (SRV) transform [24] that consists of a particular representation of the shape of a curve enables to define a so-called elastic Riemannian distance, which has proven to be useful for the statistical shape analysis of 2D and 3D curves in applications. The SRV possesses interesting properties such as a principled theoretical framework, efficient computation, and generalization to higher dimensions [2]. Nevertheless, a limitation of the SRV transform and the corresponding elastic distance is the restrictive use of the first-order derivative, while the geometry of 3-D (or \\(d\\)-D) curves depends on the derivatives until order \\(d\\). Indeed, it is well-known that a 3D curve is characterized by its curvature and its torsion: this is particularly critical when we consider trajectories or human movements, where the curvature and torsion can have a physical meaning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.724, + 0.798, + 0.74 + ], + "angle": 0, + "content": "2. Related works and contributions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As we will recall in section 3, the full geometry of a curve can be given either by the Frenet curvatures (standard curvature and torsion in 3D) or by the path of Frenet frames. There have been few attempts to directly deal with the Frenet curvatures: most of the works have been produced in 2D curves as an alternative representation [23]. Nevertheless, the potential for applications has not been investigated. In [17], the elastic shape analysis framework has been considered for 3D curves based on the Frenet frames, but the link to the physical parameters has been overlooked." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4027" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "We can also mention the shape analysis of curves on Lie groups [4] with application in computer animation. Outside the Riemannian framework, an attempt has been made to use a direct curvature-based interpolation of curves [20, 25]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.473, + 0.651 + ], + "angle": 0, + "content": "In this work, we introduce two representations of Euclidean curves for their shape analysis that use their complete geometry through Frenet curvatures. We provide the full development of the Riemannian frameworks associated with these two representations. As a consequence and in comparison with existing methods, our approaches also give explicit formulas for geodesics and geodesic distances. The first representation considered is based directly on unparametrized Frenet curvatures. We show through experiments that it defines a shape analysis framework that lacks elasticity. As the main contribution, we propose the definition of a second representation, called the Square-Root Curvature (SRC) Transform, which takes into account reparameterization and defines a metric on the space of shapes through the quotient space with the group of diffeomorphisms. One can imagine that the classical method associated with the SRVF, defining a Sobolev elastic metric [2], already implicitly uses all the geometric information necessary for a relevant curve analysis. We show here with simple examples that this is not the case. Through experiments on synthetic data, we compare the methods, and illustrate the limitations of the SRVF one, due to its lack of use of geometric information. To be able to judge and compare the quality of these metrics, we compare consistent sets of curves characterized by specific features. The SRC method shows a special strength in defining a framework that remains consistent with these sets. The straightforward example of geodesics between helices with different numbers of spins (Figure 1 in 2D and Figure 2 in 3D) shows that, in contrast, this is not the case for the SRVF method. In addition, we highlight the interest of these Frenet curvatures-based representations in the real application case of human motion trajectory analysis." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.663, + 0.43, + 0.681 + ], + "angle": 0, + "content": "3. Riemannian Geometry on Shape Space" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.735 + ], + "angle": 0, + "content": "We introduce useful notations and we review the main approach for constructing tractable representations of the shape of a curve and deriving a Riemannian geometry." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.391, + 0.759 + ], + "angle": 0, + "content": "3.1. Shape Analysis of Euclidean Curves" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We consider absolutely continuous curves that are smooth, open, and with values in some Euclidean space \\(\\mathbb{R}^d\\), we denote this set as \\(AC([0,1],\\mathbb{R}^d)\\). These curves are typically parametrized by a variable \\(t\\) that can usually be interpreted as time. Nevertheless, from a (statistical) shape analysis point of view, we focus on the geometric shape of curves that do not depend on a specific parametrization or standard transformations such as translations, rotations, scaling, or reparametrizations. To distinguish be" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.366 + ], + "angle": 0, + "content": "tween parametrized curves that differ only by translation, we consider the set of absolutely continuous curves where \\( x(0) = 0 \\), denoted by \\( AC_0([0,1],\\mathbb{R}^d) \\). The natural and intrinsic parametrization that uniquely defines the shape of a curve \\( x \\) is the arc-length parametrization, defined with the arc length function \\( s(t) = \\int_0^t\\| \\dot{x} (u)\\| du \\), for \\( t\\in [0,1] \\). In order to remove the scaling variability, the total length of the curve \\( s(1) \\) is set to 1. Under this parametrization, the shape \\( X:[0,1]\\mapsto \\mathbb{R}^d \\) of the curve is the image of the function \\( x \\) such that \\( x(t) = X(s(t)) \\). As we want to study shapes independently of their parameterizations, we introduce the reparametrization group \\( \\mathrm{Diff}_{+}([0,1]) \\), of smooth orientation preserving diffeomorphisms of the interval [0, 1] onto itself. This group acts on the space of absolutely continuous curves by right composition, and this action only alters the parametrization of the curve, not the inherent shape \\( X \\). The space of such shapes (or unparametrized curves) is often mathematically defined as the quotient space" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.375, + 0.892, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} ([ 0, 1 ], \\mathbb {R} ^ {d}) = A C _ {0} ([ 0, 1 ], \\mathbb {R} ^ {d}) / \\operatorname {D i f f} _ {+} ([ 0, 1 ]). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.405, + 0.893, + 0.799 + ], + "angle": 0, + "content": "The purpose of shape analysis of curves is to define a distance function \\( d_{S} \\) on \\( S \\) and a framework to perform a complete statistical analysis on a set of curves in \\( S \\) (e.g. mean, classification, or Principal Component Analysis etc.). One of the main challenges in defining this distance is to choose an appropriate mathematical representation of the curves that can be made invariant to all shape-preserving transformations - translation, rotation, scaling, and reparametrization. Moreover, one of the stakes of such representation is to offer an (infinite-dimensional) Riemannian manifold structure that brings powerful and flexible tools for studying the geometry of shapes or statistical properties notably thanks to the tangent space of the manifold [11, 22]. In [23], a list of the few possible representations is given - coordinate functions, curvatures, angle function, and square-root velocity function (SRVF) - and a framework for curve analysis is derived for the last two ones. While the angle representation is unparameterized, the SRVF representation depends on the parametrization, which is shown to be very useful as a tool for the registration of points across curves. As a consequence, the parameterization group \\( \\mathrm{Diff}_{+}([0,1]) \\) must be eliminated by using a quotient space. The classical approach is to define the Riemannian metric on the shape space through a metric on the space of parametrized representations that is invariant to reparametrization: \\( \\forall h \\in \\mathrm{Diff}_{+}([0,1]) \\)" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.809, + 0.892, + 0.827 + ], + "angle": 0, + "content": "\\[\nd _ {A C _ {0}} \\left(x _ {0}, x _ {1}\\right) = d _ {A C _ {0}} \\left(x _ {0} \\circ h, x _ {1} \\circ h\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.837, + 0.892, + 0.868 + ], + "angle": 0, + "content": "In that case, the distance on \\(S\\) is defined as the infimum over all possible reparametrization. For \\(X_0, X_1 \\in S\\)," + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.879, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\nd _ {\\mathcal {S}} \\left(X _ {0}, X _ {1}\\right) := \\inf _ {h \\in \\operatorname {D i f f} ([ 0, 1 ])} d _ {A C _ {0}} \\left(x _ {0}, x _ {1} \\circ h\\right). \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4028" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.149, + 0.09, + 0.825, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.26, + 0.892, + 0.278 + ], + "angle": 0, + "content": "Figure 1: Geodesic paths between two 2D scaled spirals with different number of spins: SRVF \\((1^{st}\\) row) and SRC \\((2^{nd}\\) row)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.471, + 0.348 + ], + "angle": 0, + "content": "In the following, we will denote with a dot the derivation with respect to the time variable, and with a prime the one with respect to the arc-length parameter." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.357, + 0.371, + 0.373 + ], + "angle": 0, + "content": "3.2. Square Root Velocity Framework" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.38, + 0.471, + 0.452 + ], + "angle": 0, + "content": "The square-root velocity function framework is the most commonly used representation for curve shape analysis in \\(\\mathbb{R}^d\\) [24, 23]. The square-root velocity function (SRVF) of \\(x\\in AC_0([0,1],\\mathbb{R}^d)\\), denoted by \\(\\mathcal{R}_{\\mathrm{SRVF}}(x) = q\\), is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.452, + 0.469, + 0.487 + ], + "angle": 0, + "content": "\\[\nq (t) = \\frac {\\dot {x} (t)}{\\sqrt {\\| \\dot {x} (t) \\|}} = \\sqrt {\\dot {s} (t)} T (s (t)) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.47, + 0.68 + ], + "angle": 0, + "content": "where \\( T(s(t)) = \\frac{\\dot{x}(t)}{\\|\\dot{x}(t)\\|} = X'(s(t)) \\) is the unit tangent vector of the curve. This transformation is a bijection with \\( AC_0([0,1],\\mathbb{R}^d) \\) and its explicit inverse is \\( x(t) = \\int_0^t q(u)|q(u)|du \\). As we consider length-normalized curves, the SRVFs have a unit \\( \\mathbb{L}^2 \\) norm, and their set is the convenient unit Hilbert sphere, a Riemannian submanifold of \\( \\mathbb{L}^2([0,1],\\mathbb{R}^d) \\) (with the \\( \\mathbb{L}^2 \\) inner product). Then, the \\( \\mathbb{L}^2 \\) metric on SRVF induces a Riemannian metric on \\( AC_0([0,1],\\mathbb{R}^d) \\) where geodesics are given by the shorter arcs on great circles between SRV functions. The action of \\( \\mathrm{Diff}_{+}([0,1]) \\) on \\( AC_0([0,1],\\mathbb{R}^d) \\) is reflected on \\( q \\) by the group action denoted by \\( * \\) and defined as" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.69, + 0.47, + 0.715 + ], + "angle": 0, + "content": "\\[\n(q * h) (t) = \\sqrt {\\dot {h} (t)} q (h (t)) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.723, + 0.47, + 0.785 + ], + "angle": 0, + "content": "and if the curve is rotated by a matrix \\(O \\in SO(d)\\), its SRVF gets rotated by the same matrix. The key property of this representation is the invariance of its associated distance under the action of \\(\\mathrm{Diff}_{+}([0,1])\\) and \\(SO(d)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.794, + 0.469, + 0.812 + ], + "angle": 0, + "content": "\\[\n\\left\\| O \\left(q _ {0} * h\\right) - O \\left(q _ {1} * h\\right) \\right\\| _ {\\mathbb {L} ^ {2}} = \\left\\| q _ {0} - q _ {1} \\right\\| _ {\\mathbb {L} ^ {2}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.82, + 0.469, + 0.852 + ], + "angle": 0, + "content": "The metric can then be used to define a proper distance on the shape space \\(\\mathcal{S}([0,1],\\mathbb{R}^d)\\)" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.86, + 0.469, + 0.898 + ], + "angle": 0, + "content": "\\[\nd_{\\mathcal{S}}^{(\\mathrm{SRVF})}(X_{0},X_{1}):= \\inf_{\\substack{O\\in SO(d)\\\\ h\\in \\mathrm{Diff}([0,1])}}\\cos^{-1}\\langle q_{0},O(q_{1}*h)\\rangle \\tag{7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.303, + 0.892, + 0.333 + ], + "angle": 0, + "content": "and the geodesic path on the shape space is taken between \\( q_0 \\) and \\( q_1 * h \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.333, + 0.892, + 0.378 + ], + "angle": 0, + "content": "The definition of this distance on the shape space under the SRVF representation can be interpreted as the following registration problem" + }, + { + "type": "equation", + "bbox": [ + 0.504, + 0.383, + 0.891, + 0.442 + ], + "angle": 0, + "content": "\\[\nh ^ {*}, O ^ {*} = \\underset { \\begin{array}{c} O \\in S O (d) \\\\ h \\in \\operatorname {D i f f} ([ 0, 1 ]) \\end{array} } {\\arg \\min } \\int_ {0} ^ {1} \\| q _ {0} (t) - O q _ {1} (h (t)) \\sqrt {\\dot {h} (t)} \\| _ {2} ^ {2} d t. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.443, + 0.892, + 0.534 + ], + "angle": 0, + "content": "In [3], this registration problem has been reformulated with the unit tangent vector and the arc length functions. By defining \\(\\gamma = s_{1} \\circ h \\circ s_{0}^{-1} \\in \\mathrm{Diff}_{+}([0,1])\\) the optimization problem 8 amounts to finding the optimal diffeomorphism of \\(\\mathrm{Diff}_{+}([0,1])\\) that acts on the arc-length parameter \\(s\\) and solves the minimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.499, + 0.54, + 0.891, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\gamma^ {*}, O ^ {*} = \\underset { \\begin{array}{c} O \\in S O (d) \\\\ \\gamma \\in \\operatorname {D i f f} ([ 0, 1 ]) \\end{array} } {\\arg \\min } \\int_ {0} ^ {1} \\| T _ {0} (s) - O T _ {1} (\\gamma (s)) \\sqrt {\\gamma^ {\\prime} (s)} \\| _ {2} ^ {2} d s. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.75 + ], + "angle": 0, + "content": "It should be noted, in this reformulation, that the object \\( OT_{1}(\\gamma(s))\\sqrt{\\gamma'(s)} \\) does not represent the same shape as \\( X_{1}(s) \\) in the shape space. Here the element \\( \\gamma \\) of \\( \\mathrm{Diff}_{+}([0,1]) \\) is not used as a reparametrization of the curve but to deform the element of \\( S([0,1],\\mathbb{R}^{d}) \\). Under this point of view, the set of unit tangent vectors that can be reached by deforming the vector \\( T(s) \\), with the group action \\( T*\\gamma \\) defines an equivalence class of shapes associated with that one, as in the setting of deformable templates of Grenander's theory [29, 30]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Finally, the choice of a parametrized curve representation for shape analysis, discussed in [23], can be seen as the problem of choosing a good geometric representative of the shape as a template and defining an associated registration problem. Hence, an appropriate choice may be seen as a matter of modeling and should be done in interaction with the type of data analyzed and the dimension of the space. In the next sections, we use \\( h \\) to refer to functions of \\( \\mathrm{Diff}_{+}([0,1]) \\) that act on the time variable \\( t \\) and \\( \\gamma \\) for ones that act on the arc-length variable \\( s \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4029" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.167, + 0.086, + 0.805, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.289, + 0.895, + 0.32 + ], + "angle": 0, + "content": "Figure 2: Geodesic paths between two scaled 3D circular helices with different number of spins: SRVF (1\\(^{st}\\) row) and SRC (2\\(^{nd}\\) row)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.345, + 0.471, + 0.381 + ], + "angle": 0, + "content": "4. Exhaustive geometric information with Frenet representation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.471, + 0.571 + ], + "angle": 0, + "content": "Based on these previous observations and with the intention of developing a more suitable framework for the analysis of three-dimensional curves, Brunel and Park [3] proposed a direct extension of the SRVF method, which considers not only the tangent vector as the geometric representation of the curve but the whole Frenet-Serret frame in three dimensions. Their idea is to use an exhaustive description of the geometry of curves by incorporating higher-order information about the geometry in the representation. To exploit this idea, we propose to study suitable representations based on this Frenet-Serret framework. Proofs of theoretical results are given in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.578, + 0.343, + 0.593 + ], + "angle": 0, + "content": "4.1. The Frenet-Serret framework" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.601, + 0.47, + 0.692 + ], + "angle": 0, + "content": "We introduce the Frenet-Serret framework for curves of any dimension \\(d\\). Let \\(F([0,1],\\mathbb{R}^d)\\) be the set of curves \\(x \\in AC_0([0,1],\\mathbb{R}^d)\\) \\(d\\)-times continuously differentiable, and with the first \\(d\\) derivatives linearly independent. \\(F([0,1],\\mathbb{R}^d)\\) is called the set of Frenet curves. In the following, we will restrict the shape space to be the set" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.701, + 0.469, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} ([ 0, 1 ], \\mathbb {R} ^ {d}) = F ([ 0, 1 ], \\mathbb {R} ^ {d}) / \\operatorname {D i f f} _ {+} ([ 0, 1 ]). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.729, + 0.47, + 0.833 + ], + "angle": 0, + "content": "The Frenet frame \\( e_1, e_2, \\ldots, e_d \\) associated with \\( X \\in S([0,1],\\mathbb{R}^d) \\) is uniquely defined by applying the Gram-Schmidt process to the first \\( d \\) derivatives of \\( X \\). In dimension 3, the three vectors of the Frenet frame are known as the tangent, normal and bi-normal vector. We define the function \\( Q \\) that maps to \\( s \\in [0,1] \\) along the curve the corresponding Frenet frame" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.844, + 0.469, + 0.863 + ], + "angle": 0, + "content": "\\[\nQ (s) = \\left[ e _ {1} (s) \\mid e _ {2} (s) \\mid \\dots \\mid e _ {d} (s) \\right]. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "The function \\( Q \\) is a measurable curve from \\([0,1]\\) to the Lie group of rotation matrices \\( SO(d) \\) called the Frenet path." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.346, + 0.895, + 0.424 + ], + "angle": 0, + "content": "Theorem 1 (Frenet-Serret equation [9]). Let \\( X \\in S([0,1],\\mathbb{R}^d) \\) and \\( Q(s) \\) its associated Frenet path. Then there are functions \\( \\theta_1, \\ldots, \\theta_{d-1} \\) defined on that curve with \\( \\theta_1, \\ldots, \\theta_{d-2} > 0 \\), so that every \\( \\theta_i \\) is \\( (d-1-i) \\)-times continuously differentiable and" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.435, + 0.892, + 0.453 + ], + "angle": 0, + "content": "\\[\nQ ^ {\\prime} (s) = Q (s) A _ {\\boldsymbol {\\theta}} (s) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.463, + 0.779, + 0.481 + ], + "angle": 0, + "content": "where \\(\\pmb {\\theta}(s) = (\\theta_{1}(s),\\dots ,\\theta_{d - 1}(s))^{T}\\) and" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.491, + 0.878, + 0.589 + ], + "angle": 0, + "content": "\\[\nA _ {\\pmb {\\theta}} (s) = \\left[ \\begin{array}{c c c c c} 0 & - \\theta_ {1} (s) & 0 & \\dots & 0 \\\\ \\theta_ {1} (s) & 0 & - \\theta_ {2} (s) & \\ddots & \\vdots \\\\ 0 & \\theta_ {2} (s) & \\ddots & \\ddots & 0 \\\\ \\vdots & \\ddots & \\ddots & 0 & - \\theta_ {d - 1} (s) \\\\ 0 & \\dots & 0 & \\theta_ {d - 1} (s) & 0 \\end{array} \\right]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.597, + 0.892, + 0.629 + ], + "angle": 0, + "content": "and \\(\\theta_{i}\\) is called the \\(i\\)-th Frenet curvature, and the equation is called the Frenet-Serret equation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.819 + ], + "angle": 0, + "content": "In dimension 3, the two Frenet curvatures are known as \\( s \\mapsto \\kappa(s) \\) the curvature function and \\( s \\mapsto \\tau(s) \\) the torsion function. They have an interpretable physical meaning. The curvature function measures how sharply the curve changes direction at a given point, and the torsion function measures the degree to which the curve twists and turns as it moves along its path. The Frenet-Serret equation with an initial condition \\( Q(0) = Q_0 \\) defines an ordinary differential equation on the Lie group \\( SO(d) \\) where the function \\( s \\mapsto A_\\theta(s) \\) has values in the Lie algebra of skew-symmetric matrices. This equation can also be expressed in function of the time variable \\( t \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.829, + 0.892, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\frac {d Q (s (t))}{d t} = \\dot {s} (t) Q (s (t)) A _ {\\boldsymbol {\\theta}} (s (t)). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Lemma 1. The Frenet curvatures and the Frenet path are invariant under all Euclidean motions." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4030" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.47, + 0.155 + ], + "angle": 0, + "content": "This means that for \\( x \\in F([0,1],\\mathbb{R}^d), O \\in SO(d), a \\in \\mathbb{R}^d \\) and \\( h \\in \\mathrm{Diff}_{+}([0,1]) \\), the curve defined by \\( \\tilde{x}(t) = a + Ox(h(t)) \\) has the same Frenet curvatures as \\( x \\) and \\( \\tilde{Q}(s) = OQ(s) \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.161, + 0.471, + 0.237 + ], + "angle": 0, + "content": "Theorem 2 (Fundamental theorem of the local theory of curves [9]). Let \\(\\theta_1, \\ldots, \\theta_{d-1} \\in C^\\infty([0,1], \\mathbb{R})\\) such that \\(\\theta_1, \\ldots, \\theta_{d-2} > 0\\). For a given \\(X_0 \\in \\mathbb{R}^d\\) and \\(Q_0 \\in SO(d)\\) there is a unique \\(X \\in S([0,1], \\mathbb{R}^d)\\) parametrized by arc length and satisfying the following three conditions:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.245, + 0.197, + 0.26 + ], + "angle": 0, + "content": "\\(X(0) = X_0\\)" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.269, + 0.397, + 0.285 + ], + "angle": 0, + "content": "- \\( Q_{0} \\) is the Frenet frame of \\( X \\) at point \\( s = 0 \\)," + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.294, + 0.397, + 0.309 + ], + "angle": 0, + "content": "- \\(\\theta_{1},\\ldots ,\\theta_{d - 1}\\) are the Frenet curvatures of \\(X\\)" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.245, + 0.397, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.326, + 0.469, + 0.388 + ], + "angle": 0, + "content": "Theorems 1 and 2 state that there is a bijection, up to a translation and a rotation, between the shape space \\(\\mathcal{S}([0,1],\\mathbb{R}^d)\\), the set of admissible Frenet curvatures \\(\\mathcal{H}\\) and the set of corresponding Frenet paths \\(\\mathcal{F}_0\\), where" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.396, + 0.47, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} = \\left\\{\\boldsymbol {\\theta} \\in C ^ {\\infty} ([ 0, 1 ], \\mathbb {R} ^ {d - 1}) | \\theta_ {1}, \\dots , \\theta_ {d - 2} > 0 \\right\\}, \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.444, + 0.469, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {0} = \\left\\{ \\begin{array}{c} Q \\in \\mathbb {L} ^ {2} ([ 0, 1 ], S O (d)) \\text {s u c h t h a t} \\\\ Q ^ {\\prime} (s) = Q (s) A _ {\\boldsymbol {\\theta}} (s), Q (0) = I _ {d}, \\boldsymbol {\\theta} \\in \\mathcal {H} \\end{array} \\right\\}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.498, + 0.469, + 0.573 + ], + "angle": 0, + "content": "From the detailed Frenet-Serret framework, one can think of the direct extension of the square-root velocity function 4 that simply consists in replacing the tangent vector with the entire Frenet frame. The representation of a parametrized curve \\( x \\in F([0,1],\\mathbb{R}^d) \\) will be then" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.584, + 0.469, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {Q} (x) (t) = \\sqrt {\\dot {s} (t)} Q (s (t)). \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.612, + 0.469, + 0.779 + ], + "angle": 0, + "content": "This representation is used in [3] to define a new alignment method on \\(S\\). They extend the SRVF registration problem 9 by using the Frobenius distance between the Frenet frames instead of only the \\(\\mathbb{L}^2\\) distance of the unit tangent vectors. They show to obtain more precise results with their method than the SRVF one. From the previous theorems, it is clear that this representation uniquely defines a parametrized curve \\(x \\in F([0,1],\\mathbb{R}^d)\\). However, as a result of the Frobenius theorem [12], the set of such representations \\(\\mathcal{R}_Q\\) appears not to be a manifold. We leave the demonstration in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.383, + 0.803 + ], + "angle": 0, + "content": "4.2. Unparametrized Frenet curvatures" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.469, + 0.87 + ], + "angle": 0, + "content": "A possible representation of a parametrized curve, already suggested in [23, 25, 20], which keeps the idea of encoding more geometric information, is the unparametrized Frenet curvatures and the arc-length function pair" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.88, + 0.469, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {\\boldsymbol {\\theta}} (x) (t) = \\left(\\sqrt {\\dot {s} (t)}, \\boldsymbol {\\theta} (s (t))\\right). \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "We denote \\(\\Psi([0,1])\\), the set of square root velocity functions of length-normalized arc-length functions. This set is well-studied in the literature [15, 27]. It is the unit sphere of the Hilbert space \\(\\mathbb{L}^2([0,1],\\mathbb{R})\\) and therefore a Riemannian manifold equipped with the \\(\\mathbb{L}^2\\) metric. Then, the geodesic distance between two elements in \\(\\Psi([0,1])\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.192, + 0.892, + 0.218 + ], + "angle": 0, + "content": "\\[\nd _ {\\Psi} \\left(\\sqrt {\\dot {s} _ {0}}, \\sqrt {\\dot {s} _ {1}}\\right) = \\cos^ {- 1} \\left(\\left\\langle \\sqrt {\\dot {s} _ {0}}, \\sqrt {\\dot {s} _ {1}} \\right\\rangle\\right) \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.226, + 0.833, + 0.241 + ], + "angle": 0, + "content": "and the geodesic path connecting them is given by" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.25, + 0.892, + 0.283 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {\\Psi} (\\tau) = \\frac {\\sin ((1 - \\tau) \\vartheta)}{\\sin (\\vartheta)} \\sqrt {\\dot {s} _ {0}} + \\frac {\\sin (\\tau \\vartheta)}{\\sin (\\vartheta)} \\sqrt {\\dot {s} _ {1}} \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.293, + 0.892, + 0.339 + ], + "angle": 0, + "content": "where \\(\\vartheta = d_{\\Psi}(\\sqrt{\\dot{s}_0},\\sqrt{\\dot{s}_1})\\) Moreover, any element of \\(\\mathrm{Diff}_{+}([0,1])\\) is uniquely represented by an element of \\(\\Psi ([0,1])\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.348, + 0.891, + 0.379 + ], + "angle": 0, + "content": "Proposition 1. The set of Frenet curvatures \\(\\mathcal{H}\\) is a Riemannian submanifold of \\(\\mathbb{L}^2([0,1],\\mathbb{R}^{d-1})\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.385, + 0.892, + 0.522 + ], + "angle": 0, + "content": "Proof. The set \\( M = \\{x \\in \\mathbb{R}^{d-1} | x_1, \\ldots, x_{d-2} > 0\\} \\) is a open subset of the Riemannian manifold \\( \\mathbb{R}^{d-1} \\). Then it is itself a differentiable Riemannian manifold with the standard inner product of \\( \\mathbb{R}^{d-1} \\), and for any point \\( p \\in M \\) the tangent space \\( T_p(M) \\) is \\( \\mathbb{R}^{d-1} \\). The set of Frenet curvatures \\( \\mathcal{H} \\) is the set of measurable curves from [0, 1] to the Riemannian manifold \\( M \\) and thus also a manifold ([26]). Its tangent space is \\( \\mathbb{L}^2([0,1], \\mathbb{R}^{d-1}) \\) and it can be equipped with the \\( \\mathbb{L}^2 \\) Riemannian metric." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.531, + 0.891, + 0.561 + ], + "angle": 0, + "content": "Consequently, the geodesic distance on \\(\\mathcal{H}\\) is simply the \\(\\mathbb{L}^2\\) norm" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.572, + 0.891, + 0.589 + ], + "angle": 0, + "content": "\\[\nd _ {\\mathcal {H}} \\left(\\boldsymbol {\\theta} _ {0}, \\boldsymbol {\\theta} _ {1}\\right) = \\left\\| \\boldsymbol {\\theta} _ {0} - \\boldsymbol {\\theta} _ {1} \\right\\| _ {\\mathbb {L} ^ {2}} \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.599, + 0.879, + 0.614 + ], + "angle": 0, + "content": "and the geodesic path is the straight line connecting them" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.625, + 0.891, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {\\mathcal {H}} (\\tau) = (1 - \\tau) \\boldsymbol {\\theta} _ {0} + \\tau \\boldsymbol {\\theta} _ {1}. \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.65, + 0.89, + 0.681 + ], + "angle": 0, + "content": "Proposition 2. The map \\(\\mathcal{R}_{\\theta}: F([0,1], \\mathbb{R}^d) \\to \\Psi([0,1]) \\times \\mathcal{H}\\), defined above, is a bijection." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.689, + 0.892, + 0.782 + ], + "angle": 0, + "content": "Proof. The element of \\(\\Psi([0,1])\\) uniquely defines the arclength function by \\(s(t) = \\int_0^t (\\sqrt{\\dot{s}(u)})^2 du\\). As mentioned before, we have a bijection between the unparametrized Frenet curvatures in \\(\\mathcal{H}\\) and the unparametrized curve in the shape space \\(\\mathcal{S}\\) (1, 2). Then, from \\(X \\in \\mathcal{S}\\), the initial parametrized curve is simply \\(x(t) = X(s(t))\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.791, + 0.892, + 0.866 + ], + "angle": 0, + "content": "The set of such \\(\\mathcal{R}_{\\theta}\\) is the Cartesian product of \\(\\Psi ([0,1])\\) and \\(\\mathcal{H}\\) and, therefore, is also a Riemannian manifold equipped with the product metric \\(d_{\\Psi}\\oplus d_{\\mathcal{H}}\\) [16]. The induced metric on \\(F([0,1],\\mathbb{R}^d)\\) under the representation \\(\\mathcal{R}_{\\theta}\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.873, + 0.891, + 0.898 + ], + "angle": 0, + "content": "\\[\nd _ {\\boldsymbol {\\theta}} \\left(x _ {0}, x _ {1}\\right) = d _ {\\Psi} \\left(\\sqrt {\\dot {s} _ {0}}, \\sqrt {\\dot {s} _ {1}}\\right) + d _ {\\mathcal {H}} \\left(\\boldsymbol {\\theta} _ {0}, \\boldsymbol {\\theta} _ {1}\\right). \\tag {22}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4031" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "In order to define a distance on the shape space \\(S\\) from that one, we must quotient out the space \\(\\mathrm{Diff}_{+}([0,1])\\). The action of \\(\\mathrm{Diff}_{+}([0,1])\\) on \\(\\Psi ([0,1])\\) is the same as 5, and the Frenet curvatures are invariant under reparametrization of the corresponding parametrized curve (Lemma 1). Moreover, by simply taking \\(h^* = s_1^{-1} \\circ s_0 \\in \\mathrm{Diff}_{+}([0,1])\\), we have \\(\\sqrt{s_1} * h^* = \\sqrt{s_0}\\), and thus the distance on \\(\\Psi ([0,1]) / \\mathrm{Diff}_{+}([0,1])\\) between \\(s_0\\) and \\(s_1\\) is zero. Hence, the induced distance on the shape space, between \\(X_0, X_1 \\in S\\), under the representation \\(\\mathcal{R}_{\\theta}\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.255, + 0.47, + 0.277 + ], + "angle": 0, + "content": "\\[\nd _ {\\mathcal {S}} ^ {(\\boldsymbol {\\theta})} \\left(X _ {0}, X _ {1}\\right) := d _ {\\mathcal {H}} \\left(\\boldsymbol {\\theta} _ {0}, \\boldsymbol {\\theta} _ {1}\\right) = \\left\\| \\boldsymbol {\\theta} _ {0} - \\boldsymbol {\\theta} _ {1} \\right\\| _ {\\mathbb {L} ^ {2}} \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.288, + 0.354, + 0.304 + ], + "angle": 0, + "content": "and the geodesic path connecting them is" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.315, + 0.47, + 0.342 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {\\mathcal {S}} ^ {(\\boldsymbol {\\theta})} (\\tau) = \\left(\\sqrt {\\dot {s} _ {0}}, \\alpha_ {\\mathcal {H}} (\\tau)\\right). \\tag {24}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.352, + 0.47, + 0.429 + ], + "angle": 0, + "content": "This immediate representation by the Frenet curvatures appears in the experiments not to be sufficiently elastic (Figure 5). It has somewhat the same weakness as the angle representation proposed in [23], the Frenet curvatures being already independent of the parametrization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.439, + 0.388, + 0.455 + ], + "angle": 0, + "content": "4.3. Square Root Curvatures Transform" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.463, + 0.47, + 0.569 + ], + "angle": 0, + "content": "To overcome the \"non-elasticity\" issue of the representation defined above, we propose a second framework for shape analysis based on Frenet curvatures which uses, like the square root velocity function, the parametrization as a tool to register the curves and define a more \"elastic\" method. The latter is inspired by the square-root velocity transform of \\( SO(d) \\)-valued curves." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.47, + 0.627 + ], + "angle": 0, + "content": "Definition 1 (SRV Transform for curves on \\(SO(d)\\)). Let \\(P \\in C^{\\infty}([0,1], SO(d))\\). The Square Root Velocity transform of \\(P\\) is the map" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.638, + 0.469, + 0.684 + ], + "angle": 0, + "content": "\\[\nq (P) (t) = \\frac {L _ {P (t) ^ {- 1}} \\dot {P} (t)}{\\sqrt {\\| \\dot {P} (t) \\| _ {F}}} = \\frac {P (t) ^ {T} \\dot {P} (t)}{\\sqrt {\\| \\dot {P} (t) \\| _ {F}}}, \\tag {25}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.696, + 0.47, + 0.745 + ], + "angle": 0, + "content": "where \\(\\| .\\| _F\\) is the Frobenius norm associated with the scalar product on the Lie Algebra of skew-symmetric matrices \\(\\langle A,B\\rangle = \\frac{1}{2} tr(A^T B) = -\\frac{1}{2} tr(AB)\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.753, + 0.469, + 0.8 + ], + "angle": 0, + "content": "Let \\(x \\in F([0,1],\\mathbb{R}^d)\\) and \\(Q(t) \\in C^\\infty([0,1],SO(d))\\) be its associated Frenet path. Using the Frenet-Serret differential equation, the SRV Transform of the Frenet path is" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.812, + 0.469, + 0.848 + ], + "angle": 0, + "content": "\\[\nq (Q) (t) = \\sqrt {\\dot {s} (t)} \\frac {A _ {\\boldsymbol {\\theta}} (s (t))}{\\sqrt {\\left\\| A _ {\\boldsymbol {\\theta}} (s (t)) \\right\\| _ {F}}}. \\tag {26}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.857, + 0.318, + 0.872 + ], + "angle": 0, + "content": "Proposition 3. Let \\(\\theta \\in \\mathcal{H}\\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.885, + 0.369, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\left\\| A _ {\\boldsymbol {\\theta}} (s (t)) \\right\\| _ {F} = \\left\\| \\boldsymbol {\\theta} (s (t)) \\right\\| _ {2}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "Based on the SRV Transform of a Frenet path and Proposition 3, we propose a new transformation of a parametrized curve, which we have called the Square-Root Curvatures (SRC) transform." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.161, + 0.894, + 0.237 + ], + "angle": 0, + "content": "Definition 2 (Square-Root Curvatures Transform). Let \\(x \\in F([0,1],\\mathbb{R}^d)\\). We consider its associated arc-length function \\(s(t)\\) and Frenet curvatures \\(\\pmb{\\theta}(s(t))\\) defined as in Theorem 1. Then we define its square-root curvatures transform to be the map" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.246, + 0.892, + 0.282 + ], + "angle": 0, + "content": "\\[\nc (t) = \\sqrt {\\dot {s} (t)} \\frac {\\boldsymbol {\\theta} (s (t))}{\\sqrt {\\| \\boldsymbol {\\theta} (s (t)) \\|}}. \\tag {27}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.29, + 0.863, + 0.305 + ], + "angle": 0, + "content": "The set of such square-root curvatures transforms is" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.315, + 0.892, + 0.334 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} = \\left\\{c \\in \\mathbb {L} ([ 0, 1 ], \\mathbb {R} ^ {d - 1}) \\mid c _ {1}, \\dots , c _ {d - 2} > 0 \\right\\}, \\tag {28}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.343, + 0.892, + 0.464 + ], + "angle": 0, + "content": "which is the same as the set of admissible Frenet curvatures \\(\\mathcal{H}\\). We have already shown in the previous section that this set is a Riemannian manifold equipped with the \\(\\mathbb{L}^2\\) metric. Therefore, the geodesic distance between \\(c_{0}, c_{1} \\in \\mathcal{C}\\) is the \\(\\mathbb{L}^2\\) distance between them, and the geodesic path is a straight line. We define the following representation of a parametrized curve \\(x \\in F([0,1], \\mathbb{R}^d)\\), from its Square-Root Curvatures transform, by" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.473, + 0.892, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {\\mathrm {S R C}} (x) (t) = \\left(\\sqrt {\\dot {s} (t)}, c (t)\\right). \\tag {29}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.509, + 0.892, + 0.541 + ], + "angle": 0, + "content": "Proposition 4. The map \\(\\mathcal{R}_{SRC} : F([0,1],\\mathbb{R}^d) \\to \\Psi([0,1]) \\times \\mathcal{C}\\), defined above, is a bijection." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.549, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Proof. This is again a result of theorems 1 and 2. To get \\(x\\) from \\(\\mathcal{R}_{\\mathrm{SRC}}(x)\\), it should be noted firstly that \\(c(t) \\| c(t) \\| = \\dot{s}(t) \\theta(s(t))\\). From that, the skew-symmetric matrix function of the Frenet-Serret ODE can be reconstructed. By solving the corresponding Frenet-Serret ODE one gets the associated time parametrized Frenet path \\(Q(t)\\). Then, using the first component of \\(\\mathcal{R}_{\\mathrm{SRC}}(x)\\), we get \\(x(t) = X(s(t)) = \\int_0^t \\dot{s}(u) T(s(u)) du\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.682, + 0.892, + 0.788 + ], + "angle": 0, + "content": "The set of such square root curvature representations \\(\\mathcal{R}_{\\mathrm{SRC}}\\) is the Cartesian product \\(\\Psi ([0,1])\\times \\mathcal{C}\\) and therefore a Riemannian manifold with the product metric \\(d_{\\Psi}\\oplus d_{\\mathcal{C}}\\). This representation is, by definition, invariant under the action of \\(SO(d)\\). Then, the corresponding shape space is the quotient space \\(\\Psi ([0,1])\\times \\mathcal{C} / \\mathrm{Diff}([0,1])\\). Let's \\(x\\in F([0,1],\\mathbb{R}^d)\\) and \\(h\\in \\mathrm{Diff}([0,1])\\). The SRC representation of \\(\\tilde{x} = x\\circ h\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.797, + 0.892, + 0.823 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {\\mathrm {S R C}} (\\tilde {x}) = \\left(\\sqrt {\\dot {s}} * h, c * h\\right) = \\mathcal {R} _ {\\mathrm {S R C}} (x) * h \\tag {30}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.832, + 0.768, + 0.847 + ], + "angle": 0, + "content": "where \\( * \\) is the group action defined in 5." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.855, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Proposition 5. The metric on \\( F([0,1],\\mathbb{R}^d) \\) induced by the Riemannian metric on \\( \\Psi ([0,1])\\times \\mathcal{C} \\) defined by \\( d_{SRC}\\coloneqq d_{\\Psi}\\oplus d_{\\mathcal{C}} \\) is invariant under the action of \\( \\text{Diff}_{+}([0,1]) \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4032" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "The distance on the shape space \\( \\mathcal{S} \\) under the representation \\( \\mathcal{R}_{\\mathrm{SRC}} \\), between two elements \\( X_0, X_1 \\in \\mathcal{S} \\), is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.146, + 0.47, + 0.174 + ], + "angle": 0, + "content": "\\[\nd _ {S} ^ {\\left(\\mathrm {S R C}\\right)} \\left(X _ {0}, X _ {1}\\right) := \\inf _ {h \\in \\operatorname {D i f f} _ {+} ([ 0, 1 ])} d _ {\\mathrm {S R C}} \\left(x _ {0}, x _ {1} \\circ h\\right). \\tag {31}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.186, + 0.469, + 0.215 + ], + "angle": 0, + "content": "From the optimal wrapping function \\( h^* \\) the geodesic path on \\( S \\) between them is" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.224, + 0.47, + 0.277 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {\\mathcal {S}} ^ {(\\mathrm {S R C})} (\\tau) = \\left( \\begin{array}{c} \\frac {\\sin ((1 - \\tau) \\vartheta)}{\\sin (\\vartheta)} \\sqrt {\\dot {s} _ {0}} + \\frac {\\sin (\\tau \\vartheta)}{\\sin (\\vartheta)} \\left(\\sqrt {\\dot {s} _ {1}} * h ^ {*}\\right), \\\\ (1 - \\tau) c _ {0} + \\tau \\left(c _ {1} * h ^ {*}\\right)) \\end{array} \\right. \\tag {32}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.277, + 0.469, + 0.322 + ], + "angle": 0, + "content": "where \\(\\vartheta = d_{\\Psi}(\\sqrt{\\dot{s}_0},\\sqrt{\\dot{s}_1}*h^*)\\). The registration problem consider here is to find the minimizer \\(h^*\\) over \\(\\mathrm{Diff}_{+}([0,1])\\) of" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.332, + 0.476, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\int_ {0} ^ {1} \\| c _ {0} (t) - \\left(c _ {1} * h\\right) (t) \\| ^ {2} + \\left\\| \\sqrt {\\dot {s} _ {0} (t)} - \\left(\\sqrt {\\dot {s} _ {1}} * h\\right) (t) \\right\\| ^ {2} d t. \\tag {33}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.393, + 0.469, + 0.44 + ], + "angle": 0, + "content": "Using the reformulation principle of [3], that is \\(\\gamma = s_1\\circ h\\circ\\) \\(s_0^{-1}\\in \\mathrm{Diff}_+([0,1])\\), this registration problem is shown to be equivalent to finding \\(\\gamma^{*}\\in \\mathrm{Diff}_{+}([0,1])\\) that minimizes" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.466, + 0.469, + 0.529 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\int_ {0} ^ {1} \\left\\| \\frac {\\boldsymbol {\\theta} _ {0} (s)}{\\sqrt {\\lVert \\boldsymbol {\\theta} _ {0} (s) \\rVert}} - \\sqrt {\\gamma^ {\\prime} (s)} \\frac {\\boldsymbol {\\theta} _ {1} (\\gamma (s))}{\\sqrt {\\lVert \\boldsymbol {\\theta} _ {1} (\\gamma (s)) \\rVert}} \\right\\| ^ {2} \\\\ + \\| 1 - \\gamma^ {\\prime} (s) \\| ^ {2} d s. \\tag {34} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.541, + 0.469, + 0.632 + ], + "angle": 0, + "content": "Note that this reformulation has the form of a penalized registration problem. The second term represents a penalty term on \\(\\gamma\\) and ensures a certain smoothness of the warping function. In this framework, the deformable templates are the square-root normalized curvatures which encode more geometric information than the unit tangent vector." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.646, + 0.21, + 0.663 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.671, + 0.469, + 0.732 + ], + "angle": 0, + "content": "In this section, we report the experimental results of the proposed methods, comparing them with the SRVF method. We use both synthetic and real data. Additional results and figures are available in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.464, + 0.757 + ], + "angle": 0, + "content": "5.1. Statistical estimation of the Frenet curvatures" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The main limitation of shape analysis methods based on the Frenet curvatures is the need of additional estimates of curvatures. Being dependent on higher-order derivatives (up to order \\(d\\)), they are quite sensitive to the observation noise of the Euclidean curve. We detail here a simple method that can be used for their smooth estimation, and we refer to [21, 18] for a more complex and detailed statistical estimation algorithm. First, it is possible to use a local polynomial smoothing algorithm to estimate the \\(d\\) first" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "derivatives of the Euclidean curve [6]. From these derivatives, the raw estimates of the Frenet curvatures can be computed by using their extrinsic formulas. We propose here a second method based on the Frenet-Serret ODE approximation, to obtain the raw estimates, that appear to be more stable. A simple middle point approximation of the ODE solution gives" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.208, + 0.892, + 0.243 + ], + "angle": 0, + "content": "\\[\nQ \\left(s _ {j}\\right) \\approx Q \\left(s _ {i}\\right) \\exp \\left(\\left(s _ {i} - s _ {j}\\right) A _ {\\boldsymbol {\\theta}} \\left(\\frac {s _ {i} + s _ {j}}{2}\\right)\\right). \\tag {35}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.253, + 0.892, + 0.392 + ], + "angle": 0, + "content": "where \\(\\exp(.)\\) is the exponential map of the Lie group \\(SO(d)\\). Then, using the inverse logarithm map \\(\\log(.)\\), this gives an approximation of the matrix \\(A_{\\theta}((s_i + s_j) / 2) \\approx \\frac{1}{s_i - s_j} \\log(Q(s_i)^T Q(s_j))\\) and by identification raw estimates of \\(\\pmb{\\theta}((s_i + s_j) / 2)\\). As we consider here a problem of estimating a functional parameter, we formulate the final \\(\\pmb{\\theta}\\) estimation problem as a penalized weighted functional regression with the obtained raw estimates, that we solve by using a B-spline approximation of \\(\\pmb{\\theta}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.402, + 0.8, + 0.418 + ], + "angle": 0, + "content": "5.2. Experiments with synthetic curves" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.892, + 0.544 + ], + "angle": 0, + "content": "We use synthetic data to highlight the differences between the methods discussed above (SRVF, SRC, and Frenet curvatures). The computations related to the SRVF method are made with the package fdasrsf. The SRC and Frenet curvatures methods are implemented with the code provided as supplementary material, including a dynamic programming algorithm for solving the registration problems." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.668 + ], + "angle": 0, + "content": "We consider the simple case of a set of 20 curves in \\(\\mathbb{R}^2\\) with a single large peak of curvature. This one is created by generating curvature functions on \\([0,1]\\) with one peak of maximum value 60.5, width 0.15, and location chosen randomly between 0.1 and 0.9. These curves have the shape of a loop made with a wire, where the loop is more or less close to the right or left wire end, depending on the location of the curvature peak." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.678, + 0.89, + 0.785 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.808, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Figure 3: Matrices of pairwise SRVF distance (left), SRC distance (middle), and unparametrized Frenet curvatures distance (right) by sorted location of the peak curvature." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We compare the three methods through the pairwise distance matrices in Figure 3, and the geodesic paths computed" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4033" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.466, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.257, + 0.47, + 0.304 + ], + "angle": 0, + "content": "Figure 4: SRVF \\((1^{st}\\) row), SRC \\((2^{nd}\\) row), and Frenet curvatures \\((3^{rd}\\) row) geodesic paths between curves with curvature peaks located at 0.27 (left) and 0.78 (right)." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.322, + 0.468, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.446, + 0.47, + 0.492 + ], + "angle": 0, + "content": "Figure 5: Curvatures of the Euclidean curves along the geodesic paths plotted in Figure 4: SRVF \\((1^{st}\\) row), SRC \\((2^{nd}\\) row), and Frenet curvatures \\((3^{rd}\\) row)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.853 + ], + "angle": 0, + "content": "between two of these curves in Figure 4, with peaks located at 0.27 (red curve) and at 0.78 (blue curve). The corresponding deformations through the variations of the curvature along the different geodesic paths on Euclidean curves are shown in Figure 5, which highlights the strengths and weaknesses of each method. First, it emphasizes the \"nonelasticity\" of the unparametrized Frenet curvatures method, as in the middle of the geodesic path, we have two peaks of curvature and, therefore, a completely different shape without any loop. This explains the inconsistency of the heatmap under this method. Conversely, there is an elastic deformation of the curvature with the SRC transform, and shapes along the geodesic are consistent with the set of curves considered, which is well summarised on the corresponding heatmap where all distances are rather close to zero. For the SRVF method, the chosen example with peaks of curvature that are quite far apart shows that artifacts appear along the geodesic; the middle curve has two small loops at the edges. This phenomenon gives unreliable distances, as shown in the heatmap, where the distances are not monotone as a function of the spacing between the curvature peaks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "By considering a set of curves characterized by specific features, we observe a clear difference in consistency of the shapes along the geodesics with respect to the different" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "methods. This phenomenon is quite visible on the geodesic between helices in 2D, Figure 1 or 3D, Figure 2. In that case, within both geodesic paths under SRVF method, the curves lose the characteristic geometry of the helix. A three-dimensional circular helix is characterized by having a constant curvature and torsion, which is not the case along the SRVF geodesic, but preserved with the SRC method. In that case the geodesic under the Frenet curvature representation is very similar to SRC." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.237, + 0.851, + 0.253 + ], + "angle": 0, + "content": "5.3. Application to sign language motion data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.26, + 0.892, + 0.501 + ], + "angle": 0, + "content": "It appears that curvilinear velocity and Frenet curvatures are particularly relevant parameters for the analysis of human motion. Several laws involving these parameters can be found in the literature [10, 14, 19, 7]; among others, the power laws state a special relationship between the curvature, the torsion, and the velocity of a point trajectory representing human motion. Using a method that conserves the shape of these parameters is, therefore, of particular interest in this application. We demonstrate here with the case of wrist trajectories in sign language, acquired with a motion capture system by the company MocapLab (https://www.mocaplab.com/fr/). We compute the geodesic paths, under each of the frameworks, between the arbitrarily chosen red and blue curves within the set of several repetitions of the sign \"Femme\", shown in Figure 6 with the corresponding time-parameterized Frenet curvatures." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.514, + 0.892, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Figure 6: Trajectories of the right wrist while signing \"Femme\" in sign language: 3D curves (left), time-parametrized curvatures (top right), torsions (bottom right). The blue and red ones are used to compute the geodesic in Figure 7." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Figure 7 emphasizes the advantage of considering a representation depending on the parameterization, allowing a registration before computing the geodesic. However, it also shows that considering only the tangent vector as a representative object (SRVF) is not sufficient to find the optimal reparametrization that correctly aligns the torsions," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4034" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.351, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.091, + 0.608, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.091, + 0.871, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.2, + 0.349, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.323, + 0.255, + 0.337 + ], + "angle": 0, + "content": "(a) SRVF" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.201, + 0.609, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.46, + 0.324, + 0.512, + 0.337 + ], + "angle": 0, + "content": "(b) SRC" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.201, + 0.868, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.324, + 0.808, + 0.337 + ], + "angle": 0, + "content": "(c) Frenet curvatures" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.352, + 0.894, + 0.384 + ], + "angle": 0, + "content": "Figure 7: Comparison between time-parametrized curvature and torsion along the geodesic path under SRVF (left), SRC (middle), and Frenet curvatures (right)." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.407, + 0.264, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.533, + 0.257, + 0.547 + ], + "angle": 0, + "content": "(a) Warping functions \\( h \\)" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.407, + 0.443, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.533, + 0.435, + 0.548 + ], + "angle": 0, + "content": "(b) Warping functions \\(\\gamma\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.56, + 0.47, + 0.592 + ], + "angle": 0, + "content": "Figure 8: Comparison of estimated warping functions \\( h \\) (left) and \\( \\gamma \\) (right) to compute the geodesic in Figure 7." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.618, + 0.471, + 0.755 + ], + "angle": 0, + "content": "and could affect subsequent analyses made by the SRVF method, such as the mean. This results in the appearance of new minimums, maximums and zeros in the torsion functions along the SRVF geodesic. However, such characteristic points of the curvature, torsion and velocity functions are crucial in the observation of the laws of motion. It is therefore preferable to use a method that preserves these characteristics by directly optimizing the optimal alignment from these parameters, such as the proposed SRC method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.196, + 0.785 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The square-root curvature transform of a Euclidean curve in \\(\\mathbb{R}^d\\) is a representation that encodes more geometric information of the curves, and thus the results are easier to interpret than existing methods. The main limitation lies in the estimation of the Frenet curvatures from real and noisy data, nevertheless recent smooth statistical estimators can be used for computing the SRC [21, 18]. We believe our" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.41, + 0.893, + 0.47 + ], + "angle": 0, + "content": "method is particularly interesting for motion trajectory analysis and could be developed further in the future as a tool for generation, segmentation and classification of complex trajectories." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.487, + 0.597, + 0.502 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.513, + 0.893, + 0.581 + ], + "angle": 0, + "content": "[1] Martin Bauer, Nicolas Charon, Eric Klassen, and Alice Le Brigant. Intrinsic riemannian metrics on spaces of curves: Theory and computation. Handbook of Mathematical Models and Algorithms in Computer Vision and Imaging, pages 1-35, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.584, + 0.894, + 0.64 + ], + "angle": 0, + "content": "[2] Martin Bauer, Nicolas Charon, Eric Klassen, Sebastian Kurtek, Tom Needham, and Thomas Pierron. Elastic Metrics on Spaces of Euclidean Curves: Theory and Algorithms, 2022. arXiv:2209.09862 [math]. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.642, + 0.895, + 0.711 + ], + "angle": 0, + "content": "[3] Nicolas J.-B. Brunel and Juhyun Park. The frenet-serret framework for aligning geometric curves. In Frank Nielsen and Frédéric Barbaresco, editors, Geometric Science of Information, pages 608-617, Cham, 2019. Springer International Publishing. 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.714, + 0.894, + 0.768 + ], + "angle": 0, + "content": "[4] Elena Celledoni, Markus Eslitzbichler, and Alexander Schmeding. Shape analysis on lie groups with applications in computer animation. Journal of Geometric Mechanics, 8(3):273-304, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.771, + 0.895, + 0.813 + ], + "angle": 0, + "content": "[5] Marvin Eisenberger and Daniel Cremers. Hamiltonian dynamics for real-world shape interpolation. In European Conference on Computer Vision (ECCV), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.816, + 0.895, + 0.857 + ], + "angle": 0, + "content": "[6] Jianqing Fan and Irene Gijbels. Local polynomial modelling and its applications. CRC monographs on statistics and applied probability 66. Chapman & Hall, 1996. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.859, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[7] Tamar Flash and Alain Berthoz. *Space-Time Geometries for Motion and Perception in the Brain and the Arts*. Lecture Notes in Morphogenesis. Springer, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.513, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4035" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[8] Lukas Koestler, Daniel Grittner, Michael Moeller, Daniel Cremers, and Zorah Lähner. Intrinsic neural fields: Learning functions on manifolds. In European Conference on Computer Vision (ECCV), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.149, + 0.468, + 0.175 + ], + "angle": 0, + "content": "[9] Wolfgang Kühnel. Differential Geometry Curves - Surfaces, volume 77. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[10] Francesco Lacquaniti, Carlo Terzuolo, and Paolo Viviani. The law relating the kinematic and figural aspects of drawing movements. Acta Psychologica (Amst.), 54:115-130, 1983. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.468, + 0.262 + ], + "angle": 0, + "content": "[11] Serge Lang. Differential and riemannian manifolds. Springer, 1999:1-6, 2006. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.468, + 0.289 + ], + "angle": 0, + "content": "[12] Robert Lyons. Frobenius theorem two ways. Lecture note, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.469, + 0.361 + ], + "angle": 0, + "content": "[13] Zorah Lähner, Emanuele Rodola, Frank R. Schmidt, Michael M. Bronstein, and Daniel Cremers. Efficient globally optimal 2d-to-3d deformable shape matching. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), May 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[14] Uri Maoz, Alain Berthoz, and Tamar Flash. Complex unconstrained three-dimensional hand movement and constant equi-affine speed. Journal of Neurophysiology, 101:1002-1015, 2009. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.469, + 0.46 + ], + "angle": 0, + "content": "[15] J. S. Marron, James O. Ramsay, Laura M. Sangalli, and Anuj Srivastava. Functional data analysis of amplitude and phase variation. Statistical Science, 30:468-484, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[16] Turk J Math. Submanifolds of Riemannian Product Manifolds. 29:389-401, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[17] Tom Needham. Shape Analysis of Framed Space Curves. Journal of Mathematical Imaging and Vision, 61(8):1154-1172, Oct. 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[18] Juhyun Park, Nicolas Brunel, and Perrine Chassat. Curvature and torsion estimation of 3d functional data: A geometric approach to build the mean shape under the frenet serret framework, 2022. arXiv:2203.02398 [stat]. 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[19] Frank E. Pollick, Uri Maoz, Amir A. Handzel, Peter J. Giblin, Guillermo Sapiro, and Tamar Flash. Three-dimensional arm movements at constant equi-affine speed. Cortex, 45:325-339, 2009. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[20] Marianna Saba. On the usage of the curvature for the comparison of planar curves. PhD thesis, University of Cagliari, 2012. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[21] Laura M Sangalli, Piercesare Secchi, Simone Vantini, and Alessandro Veneziani. Efficient estimation of three-dimensional curves and their derivatives by free-knot regression splines, applied to the analysis of inner carotid artery centrelines. Journal of the Royal Statistical Society: Series C (Applied Statistics), 58(3):285-306, 2009. 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[22] Stefan Sommer, Tom Fletcher, and Xavier Pennec. Introduction to differential and Riemannian geometry. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[23] Anuj Srivastava and Eric Klassen. Functional and Shape Data Analysis. Springer Series in Statistics. Springer New York, 2016. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[24] Anuj Srivastava, Eric Klassen, Shantanu H. Joshi, and Ian H. Jermyn. Shape analysis of elastic curves in euclidean spaces. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(7):1415-1428, 2011. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[25] Tatiana Surazhsky and Gershon Elber. Metamorphosis of planar parametric curves via curvature interpolation. International Journal of Shape Modeling, 8:201-216, 2002. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[26] Wang Tixiang. Morse theory on banach manifolds. Acta Mathematica Sinica, 5:250-262, 1989. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[27] J. Derek Tucker, Wei Wu, and Anuj Srivastava. Generative models for functional data using phase and amplitude separation. Computational Statistics and Data Analysis, 61:50-66, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[28] Laurent Younes. Computable Elastic Distances between Shapes. SIAM Journal on Applied Mathematics, 58(2):565-586, 1998. Publisher: Society for Industrial and Applied Mathematics. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[29] Laurent Younes. *Shapes and Diffeomorphisms*. Applied Mathematical Sciences. Springer, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[30] Laurent Younes. Elastic distance between curves under the metamorphosis viewpoint, 2018. arXiv:1804.10155. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "4036" + } + ] +] \ No newline at end of file diff --git a/2023/Shape Analysis of Euclidean Curves under Frenet-Serret Framework/235fdcf4-691b-437b-8ee8-c7b76395f3ab_origin.pdf b/2023/Shape Analysis of Euclidean Curves under Frenet-Serret Framework/235fdcf4-691b-437b-8ee8-c7b76395f3ab_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..89869586c0a3169d7544b858e3e305d8243f4a31 --- /dev/null +++ b/2023/Shape Analysis of Euclidean Curves under Frenet-Serret Framework/235fdcf4-691b-437b-8ee8-c7b76395f3ab_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18ce7dd886313dc7162e6f433efb3073965f9baf4a259c86716a2c3f06e5e101 +size 3422227 diff --git a/2023/Shape Analysis of Euclidean Curves under Frenet-Serret Framework/full.md b/2023/Shape Analysis of Euclidean Curves under Frenet-Serret Framework/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ef985666efb4cbe45f950d82a25c2aa0516aa07b --- /dev/null +++ b/2023/Shape Analysis of Euclidean Curves under Frenet-Serret Framework/full.md @@ -0,0 +1,436 @@ +# Shape Analysis of Euclidean Curves under Frenet-Serret Framework + +Perrine Chassat1 +1LaMME, University of Paris-Saclay +perrine.chassat@univ-evry.fr + +Juhyun Park $^{1,2}$ $^{2}$ ENSIIE, Evry +juhyun.park@ensiie.fr + +Nicolas Brunel1,2,3 +3Quantmetry, Paris +nicolas.brunel@ensiie.fr + +# Abstract + +Geometric frameworks for analyzing curves are common in applications as they focus on invariant features and provide visually satisfying solutions to standard problems such as computing invariant distances, averaging curves, or registering curves. We show that for any smooth curve in $\mathbb{R}^d$ , $d > 1$ , the generalized curvatures associated with the Frenet-Serret equation can be used to define a Riemannian geometry that takes into account all the geometric features of the shape. This geometry is based on a Square Root Curvature Transform that extends the square root-velocity transform for Euclidean curves (in any dimensions) and provides likely geodesics that avoid artefacts encountered by representations using only first-order geometric information. Our analysis is supported by simulated data and is especially relevant for analyzing human motions. We consider trajectories acquired from sign language, and show the interest of considering curvature and also torsion in their analysis, both being physically meaningful. + +# 1. Introduction + +Identifying and comparing different types of visual objects is a fundamental task in machine learning and computer vision problems [13, 5, 8]. The shape is one of the essential features of objects that allow us to understand and characterize them. Nowadays, it is much easier to obtain data in the form of shapes, typically as dense point clouds or landmarks. The main task in shape analysis is to define a proper framework to compare and quantify the variation of the shapes. However, the shape space is generally nonlinear, and extracting meaningful information or features is complex. One of the successful approaches to shape analysis utilizes a Riemannian framework of differential geometry, where a metric can be defined between the shapes, which is invariant with respect to shape-preserving transformations such as translation and rotation. For instance, this gives rise to geodesic distances that are naturally invariant to smooth and optimal deformations through geodesic paths between the shapes [11]. This approach is very versatile as it can + +be adapted to various kinds of manifold-value data and can be designed to emphasize important geometric information to be preserved. As a consequence, several choices of metrics are possible, such as the class of invariant Sobolev metrics, often called elastic, for the analysis of curves [1]. In this work, we are concerned with curves that often arise in the application as trajectories (function of time) or motions (animation, activity recognition) and with the definition of a framework to compare their shapes. The differential geometry of Euclidean curves is among the simplest (with respect to higher dimensional manifolds), and relatively simple Riemannian metrics are available with different mathematical representations of curves [28]. Quite remarkably, the introduction of the Square Root Velocity (SRV) transform [24] that consists of a particular representation of the shape of a curve enables to define a so-called elastic Riemannian distance, which has proven to be useful for the statistical shape analysis of 2D and 3D curves in applications. The SRV possesses interesting properties such as a principled theoretical framework, efficient computation, and generalization to higher dimensions [2]. Nevertheless, a limitation of the SRV transform and the corresponding elastic distance is the restrictive use of the first-order derivative, while the geometry of 3-D (or $d$ -D) curves depends on the derivatives until order $d$ . Indeed, it is well-known that a 3D curve is characterized by its curvature and its torsion: this is particularly critical when we consider trajectories or human movements, where the curvature and torsion can have a physical meaning. + +# 2. Related works and contributions + +As we will recall in section 3, the full geometry of a curve can be given either by the Frenet curvatures (standard curvature and torsion in 3D) or by the path of Frenet frames. There have been few attempts to directly deal with the Frenet curvatures: most of the works have been produced in 2D curves as an alternative representation [23]. Nevertheless, the potential for applications has not been investigated. In [17], the elastic shape analysis framework has been considered for 3D curves based on the Frenet frames, but the link to the physical parameters has been overlooked. + +We can also mention the shape analysis of curves on Lie groups [4] with application in computer animation. Outside the Riemannian framework, an attempt has been made to use a direct curvature-based interpolation of curves [20, 25]. + +In this work, we introduce two representations of Euclidean curves for their shape analysis that use their complete geometry through Frenet curvatures. We provide the full development of the Riemannian frameworks associated with these two representations. As a consequence and in comparison with existing methods, our approaches also give explicit formulas for geodesics and geodesic distances. The first representation considered is based directly on unparametrized Frenet curvatures. We show through experiments that it defines a shape analysis framework that lacks elasticity. As the main contribution, we propose the definition of a second representation, called the Square-Root Curvature (SRC) Transform, which takes into account reparameterization and defines a metric on the space of shapes through the quotient space with the group of diffeomorphisms. One can imagine that the classical method associated with the SRVF, defining a Sobolev elastic metric [2], already implicitly uses all the geometric information necessary for a relevant curve analysis. We show here with simple examples that this is not the case. Through experiments on synthetic data, we compare the methods, and illustrate the limitations of the SRVF one, due to its lack of use of geometric information. To be able to judge and compare the quality of these metrics, we compare consistent sets of curves characterized by specific features. The SRC method shows a special strength in defining a framework that remains consistent with these sets. The straightforward example of geodesics between helices with different numbers of spins (Figure 1 in 2D and Figure 2 in 3D) shows that, in contrast, this is not the case for the SRVF method. In addition, we highlight the interest of these Frenet curvatures-based representations in the real application case of human motion trajectory analysis. + +# 3. Riemannian Geometry on Shape Space + +We introduce useful notations and we review the main approach for constructing tractable representations of the shape of a curve and deriving a Riemannian geometry. + +# 3.1. Shape Analysis of Euclidean Curves + +We consider absolutely continuous curves that are smooth, open, and with values in some Euclidean space $\mathbb{R}^d$ , we denote this set as $AC([0,1],\mathbb{R}^d)$ . These curves are typically parametrized by a variable $t$ that can usually be interpreted as time. Nevertheless, from a (statistical) shape analysis point of view, we focus on the geometric shape of curves that do not depend on a specific parametrization or standard transformations such as translations, rotations, scaling, or reparametrizations. To distinguish be + +tween parametrized curves that differ only by translation, we consider the set of absolutely continuous curves where $x(0) = 0$ , denoted by $AC_0([0,1],\mathbb{R}^d)$ . The natural and intrinsic parametrization that uniquely defines the shape of a curve $x$ is the arc-length parametrization, defined with the arc length function $s(t) = \int_0^t\| \dot{x} (u)\| du$ , for $t\in [0,1]$ . In order to remove the scaling variability, the total length of the curve $s(1)$ is set to 1. Under this parametrization, the shape $X:[0,1]\mapsto \mathbb{R}^d$ of the curve is the image of the function $x$ such that $x(t) = X(s(t))$ . As we want to study shapes independently of their parameterizations, we introduce the reparametrization group $\mathrm{Diff}_{+}([0,1])$ , of smooth orientation preserving diffeomorphisms of the interval [0, 1] onto itself. This group acts on the space of absolutely continuous curves by right composition, and this action only alters the parametrization of the curve, not the inherent shape $X$ . The space of such shapes (or unparametrized curves) is often mathematically defined as the quotient space + +$$ +\mathcal {S} ([ 0, 1 ], \mathbb {R} ^ {d}) = A C _ {0} ([ 0, 1 ], \mathbb {R} ^ {d}) / \operatorname {D i f f} _ {+} ([ 0, 1 ]). \tag {1} +$$ + +The purpose of shape analysis of curves is to define a distance function $d_{S}$ on $S$ and a framework to perform a complete statistical analysis on a set of curves in $S$ (e.g. mean, classification, or Principal Component Analysis etc.). One of the main challenges in defining this distance is to choose an appropriate mathematical representation of the curves that can be made invariant to all shape-preserving transformations - translation, rotation, scaling, and reparametrization. Moreover, one of the stakes of such representation is to offer an (infinite-dimensional) Riemannian manifold structure that brings powerful and flexible tools for studying the geometry of shapes or statistical properties notably thanks to the tangent space of the manifold [11, 22]. In [23], a list of the few possible representations is given - coordinate functions, curvatures, angle function, and square-root velocity function (SRVF) - and a framework for curve analysis is derived for the last two ones. While the angle representation is unparameterized, the SRVF representation depends on the parametrization, which is shown to be very useful as a tool for the registration of points across curves. As a consequence, the parameterization group $\mathrm{Diff}_{+}([0,1])$ must be eliminated by using a quotient space. The classical approach is to define the Riemannian metric on the shape space through a metric on the space of parametrized representations that is invariant to reparametrization: $\forall h \in \mathrm{Diff}_{+}([0,1])$ + +$$ +d _ {A C _ {0}} \left(x _ {0}, x _ {1}\right) = d _ {A C _ {0}} \left(x _ {0} \circ h, x _ {1} \circ h\right). \tag {2} +$$ + +In that case, the distance on $S$ is defined as the infimum over all possible reparametrization. For $X_0, X_1 \in S$ , + +$$ +d _ {\mathcal {S}} \left(X _ {0}, X _ {1}\right) := \inf _ {h \in \operatorname {D i f f} ([ 0, 1 ])} d _ {A C _ {0}} \left(x _ {0}, x _ {1} \circ h\right). \tag {3} +$$ + +![](images/15bd225d8a661ba455ccb22c20f9ce4350d3a602c6cf25deb1378fec83b2dabe.jpg) +Figure 1: Geodesic paths between two 2D scaled spirals with different number of spins: SRVF $(1^{st}$ row) and SRC $(2^{nd}$ row). + +In the following, we will denote with a dot the derivation with respect to the time variable, and with a prime the one with respect to the arc-length parameter. + +# 3.2. Square Root Velocity Framework + +The square-root velocity function framework is the most commonly used representation for curve shape analysis in $\mathbb{R}^d$ [24, 23]. The square-root velocity function (SRVF) of $x\in AC_0([0,1],\mathbb{R}^d)$ , denoted by $\mathcal{R}_{\mathrm{SRVF}}(x) = q$ , is defined as + +$$ +q (t) = \frac {\dot {x} (t)}{\sqrt {\| \dot {x} (t) \|}} = \sqrt {\dot {s} (t)} T (s (t)) \tag {4} +$$ + +where $T(s(t)) = \frac{\dot{x}(t)}{\|\dot{x}(t)\|} = X'(s(t))$ is the unit tangent vector of the curve. This transformation is a bijection with $AC_0([0,1],\mathbb{R}^d)$ and its explicit inverse is $x(t) = \int_0^t q(u)|q(u)|du$ . As we consider length-normalized curves, the SRVFs have a unit $\mathbb{L}^2$ norm, and their set is the convenient unit Hilbert sphere, a Riemannian submanifold of $\mathbb{L}^2([0,1],\mathbb{R}^d)$ (with the $\mathbb{L}^2$ inner product). Then, the $\mathbb{L}^2$ metric on SRVF induces a Riemannian metric on $AC_0([0,1],\mathbb{R}^d)$ where geodesics are given by the shorter arcs on great circles between SRV functions. The action of $\mathrm{Diff}_{+}([0,1])$ on $AC_0([0,1],\mathbb{R}^d)$ is reflected on $q$ by the group action denoted by $*$ and defined as + +$$ +(q * h) (t) = \sqrt {\dot {h} (t)} q (h (t)) \tag {5} +$$ + +and if the curve is rotated by a matrix $O \in SO(d)$ , its SRVF gets rotated by the same matrix. The key property of this representation is the invariance of its associated distance under the action of $\mathrm{Diff}_{+}([0,1])$ and $SO(d)$ : + +$$ +\left\| O \left(q _ {0} * h\right) - O \left(q _ {1} * h\right) \right\| _ {\mathbb {L} ^ {2}} = \left\| q _ {0} - q _ {1} \right\| _ {\mathbb {L} ^ {2}}. \tag {6} +$$ + +The metric can then be used to define a proper distance on the shape space $\mathcal{S}([0,1],\mathbb{R}^d)$ + +$$ +d_{\mathcal{S}}^{(\mathrm{SRVF})}(X_{0},X_{1}):= \inf_{\substack{O\in SO(d)\\ h\in \mathrm{Diff}([0,1])}}\cos^{-1}\langle q_{0},O(q_{1}*h)\rangle \tag{7} +$$ + +and the geodesic path on the shape space is taken between $q_0$ and $q_1 * h$ . + +The definition of this distance on the shape space under the SRVF representation can be interpreted as the following registration problem + +$$ +h ^ {*}, O ^ {*} = \underset { \begin{array}{c} O \in S O (d) \\ h \in \operatorname {D i f f} ([ 0, 1 ]) \end{array} } {\arg \min } \int_ {0} ^ {1} \| q _ {0} (t) - O q _ {1} (h (t)) \sqrt {\dot {h} (t)} \| _ {2} ^ {2} d t. \tag {8} +$$ + +In [3], this registration problem has been reformulated with the unit tangent vector and the arc length functions. By defining $\gamma = s_{1} \circ h \circ s_{0}^{-1} \in \mathrm{Diff}_{+}([0,1])$ the optimization problem 8 amounts to finding the optimal diffeomorphism of $\mathrm{Diff}_{+}([0,1])$ that acts on the arc-length parameter $s$ and solves the minimization problem: + +$$ +\gamma^ {*}, O ^ {*} = \underset { \begin{array}{c} O \in S O (d) \\ \gamma \in \operatorname {D i f f} ([ 0, 1 ]) \end{array} } {\arg \min } \int_ {0} ^ {1} \| T _ {0} (s) - O T _ {1} (\gamma (s)) \sqrt {\gamma^ {\prime} (s)} \| _ {2} ^ {2} d s. \tag {9} +$$ + +It should be noted, in this reformulation, that the object $OT_{1}(\gamma(s))\sqrt{\gamma'(s)}$ does not represent the same shape as $X_{1}(s)$ in the shape space. Here the element $\gamma$ of $\mathrm{Diff}_{+}([0,1])$ is not used as a reparametrization of the curve but to deform the element of $S([0,1],\mathbb{R}^{d})$ . Under this point of view, the set of unit tangent vectors that can be reached by deforming the vector $T(s)$ , with the group action $T*\gamma$ defines an equivalence class of shapes associated with that one, as in the setting of deformable templates of Grenander's theory [29, 30]. + +Finally, the choice of a parametrized curve representation for shape analysis, discussed in [23], can be seen as the problem of choosing a good geometric representative of the shape as a template and defining an associated registration problem. Hence, an appropriate choice may be seen as a matter of modeling and should be done in interaction with the type of data analyzed and the dimension of the space. In the next sections, we use $h$ to refer to functions of $\mathrm{Diff}_{+}([0,1])$ that act on the time variable $t$ and $\gamma$ for ones that act on the arc-length variable $s$ . + +![](images/d3bf35c281d578c5961f542676569d9ea11d9beb808bbab0998bce4336337efb.jpg) +Figure 2: Geodesic paths between two scaled 3D circular helices with different number of spins: SRVF (1 $^{st}$ row) and SRC (2 $^{nd}$ row). + +# 4. Exhaustive geometric information with Frenet representation + +Based on these previous observations and with the intention of developing a more suitable framework for the analysis of three-dimensional curves, Brunel and Park [3] proposed a direct extension of the SRVF method, which considers not only the tangent vector as the geometric representation of the curve but the whole Frenet-Serret frame in three dimensions. Their idea is to use an exhaustive description of the geometry of curves by incorporating higher-order information about the geometry in the representation. To exploit this idea, we propose to study suitable representations based on this Frenet-Serret framework. Proofs of theoretical results are given in the supplementary material. + +# 4.1. The Frenet-Serret framework + +We introduce the Frenet-Serret framework for curves of any dimension $d$ . Let $F([0,1],\mathbb{R}^d)$ be the set of curves $x \in AC_0([0,1],\mathbb{R}^d)$ $d$ -times continuously differentiable, and with the first $d$ derivatives linearly independent. $F([0,1],\mathbb{R}^d)$ is called the set of Frenet curves. In the following, we will restrict the shape space to be the set + +$$ +\mathcal {S} ([ 0, 1 ], \mathbb {R} ^ {d}) = F ([ 0, 1 ], \mathbb {R} ^ {d}) / \operatorname {D i f f} _ {+} ([ 0, 1 ]). \tag {10} +$$ + +The Frenet frame $e_1, e_2, \ldots, e_d$ associated with $X \in S([0,1],\mathbb{R}^d)$ is uniquely defined by applying the Gram-Schmidt process to the first $d$ derivatives of $X$ . In dimension 3, the three vectors of the Frenet frame are known as the tangent, normal and bi-normal vector. We define the function $Q$ that maps to $s \in [0,1]$ along the curve the corresponding Frenet frame + +$$ +Q (s) = \left[ e _ {1} (s) \mid e _ {2} (s) \mid \dots \mid e _ {d} (s) \right]. \tag {11} +$$ + +The function $Q$ is a measurable curve from $[0,1]$ to the Lie group of rotation matrices $SO(d)$ called the Frenet path. + +Theorem 1 (Frenet-Serret equation [9]). Let $X \in S([0,1],\mathbb{R}^d)$ and $Q(s)$ its associated Frenet path. Then there are functions $\theta_1, \ldots, \theta_{d-1}$ defined on that curve with $\theta_1, \ldots, \theta_{d-2} > 0$ , so that every $\theta_i$ is $(d-1-i)$ -times continuously differentiable and + +$$ +Q ^ {\prime} (s) = Q (s) A _ {\boldsymbol {\theta}} (s) \tag {12} +$$ + +where $\pmb {\theta}(s) = (\theta_{1}(s),\dots ,\theta_{d - 1}(s))^{T}$ and + +$$ +A _ {\pmb {\theta}} (s) = \left[ \begin{array}{c c c c c} 0 & - \theta_ {1} (s) & 0 & \dots & 0 \\ \theta_ {1} (s) & 0 & - \theta_ {2} (s) & \ddots & \vdots \\ 0 & \theta_ {2} (s) & \ddots & \ddots & 0 \\ \vdots & \ddots & \ddots & 0 & - \theta_ {d - 1} (s) \\ 0 & \dots & 0 & \theta_ {d - 1} (s) & 0 \end{array} \right] +$$ + +and $\theta_{i}$ is called the $i$ -th Frenet curvature, and the equation is called the Frenet-Serret equation. + +In dimension 3, the two Frenet curvatures are known as $s \mapsto \kappa(s)$ the curvature function and $s \mapsto \tau(s)$ the torsion function. They have an interpretable physical meaning. The curvature function measures how sharply the curve changes direction at a given point, and the torsion function measures the degree to which the curve twists and turns as it moves along its path. The Frenet-Serret equation with an initial condition $Q(0) = Q_0$ defines an ordinary differential equation on the Lie group $SO(d)$ where the function $s \mapsto A_\theta(s)$ has values in the Lie algebra of skew-symmetric matrices. This equation can also be expressed in function of the time variable $t$ as + +$$ +\frac {d Q (s (t))}{d t} = \dot {s} (t) Q (s (t)) A _ {\boldsymbol {\theta}} (s (t)). \tag {13} +$$ + +Lemma 1. The Frenet curvatures and the Frenet path are invariant under all Euclidean motions. + +This means that for $x \in F([0,1],\mathbb{R}^d), O \in SO(d), a \in \mathbb{R}^d$ and $h \in \mathrm{Diff}_{+}([0,1])$ , the curve defined by $\tilde{x}(t) = a + Ox(h(t))$ has the same Frenet curvatures as $x$ and $\tilde{Q}(s) = OQ(s)$ . + +Theorem 2 (Fundamental theorem of the local theory of curves [9]). Let $\theta_1, \ldots, \theta_{d-1} \in C^\infty([0,1], \mathbb{R})$ such that $\theta_1, \ldots, \theta_{d-2} > 0$ . For a given $X_0 \in \mathbb{R}^d$ and $Q_0 \in SO(d)$ there is a unique $X \in S([0,1], \mathbb{R}^d)$ parametrized by arc length and satisfying the following three conditions: + +$X(0) = X_0$ +- $Q_{0}$ is the Frenet frame of $X$ at point $s = 0$ , +- $\theta_{1},\ldots ,\theta_{d - 1}$ are the Frenet curvatures of $X$ + +Theorems 1 and 2 state that there is a bijection, up to a translation and a rotation, between the shape space $\mathcal{S}([0,1],\mathbb{R}^d)$ , the set of admissible Frenet curvatures $\mathcal{H}$ and the set of corresponding Frenet paths $\mathcal{F}_0$ , where + +$$ +\mathcal {H} = \left\{\boldsymbol {\theta} \in C ^ {\infty} ([ 0, 1 ], \mathbb {R} ^ {d - 1}) | \theta_ {1}, \dots , \theta_ {d - 2} > 0 \right\}, \tag {14} +$$ + +$$ +\mathcal {F} _ {0} = \left\{ \begin{array}{c} Q \in \mathbb {L} ^ {2} ([ 0, 1 ], S O (d)) \text {s u c h t h a t} \\ Q ^ {\prime} (s) = Q (s) A _ {\boldsymbol {\theta}} (s), Q (0) = I _ {d}, \boldsymbol {\theta} \in \mathcal {H} \end{array} \right\}. \tag {15} +$$ + +From the detailed Frenet-Serret framework, one can think of the direct extension of the square-root velocity function 4 that simply consists in replacing the tangent vector with the entire Frenet frame. The representation of a parametrized curve $x \in F([0,1],\mathbb{R}^d)$ will be then + +$$ +\mathcal {R} _ {Q} (x) (t) = \sqrt {\dot {s} (t)} Q (s (t)). \tag {16} +$$ + +This representation is used in [3] to define a new alignment method on $S$ . They extend the SRVF registration problem 9 by using the Frobenius distance between the Frenet frames instead of only the $\mathbb{L}^2$ distance of the unit tangent vectors. They show to obtain more precise results with their method than the SRVF one. From the previous theorems, it is clear that this representation uniquely defines a parametrized curve $x \in F([0,1],\mathbb{R}^d)$ . However, as a result of the Frobenius theorem [12], the set of such representations $\mathcal{R}_Q$ appears not to be a manifold. We leave the demonstration in the supplementary material. + +# 4.2. Unparametrized Frenet curvatures + +A possible representation of a parametrized curve, already suggested in [23, 25, 20], which keeps the idea of encoding more geometric information, is the unparametrized Frenet curvatures and the arc-length function pair + +$$ +\mathcal {R} _ {\boldsymbol {\theta}} (x) (t) = \left(\sqrt {\dot {s} (t)}, \boldsymbol {\theta} (s (t))\right). \tag {17} +$$ + +We denote $\Psi([0,1])$ , the set of square root velocity functions of length-normalized arc-length functions. This set is well-studied in the literature [15, 27]. It is the unit sphere of the Hilbert space $\mathbb{L}^2([0,1],\mathbb{R})$ and therefore a Riemannian manifold equipped with the $\mathbb{L}^2$ metric. Then, the geodesic distance between two elements in $\Psi([0,1])$ is + +$$ +d _ {\Psi} \left(\sqrt {\dot {s} _ {0}}, \sqrt {\dot {s} _ {1}}\right) = \cos^ {- 1} \left(\left\langle \sqrt {\dot {s} _ {0}}, \sqrt {\dot {s} _ {1}} \right\rangle\right) \tag {18} +$$ + +and the geodesic path connecting them is given by + +$$ +\alpha_ {\Psi} (\tau) = \frac {\sin ((1 - \tau) \vartheta)}{\sin (\vartheta)} \sqrt {\dot {s} _ {0}} + \frac {\sin (\tau \vartheta)}{\sin (\vartheta)} \sqrt {\dot {s} _ {1}} \tag {19} +$$ + +where $\vartheta = d_{\Psi}(\sqrt{\dot{s}_0},\sqrt{\dot{s}_1})$ Moreover, any element of $\mathrm{Diff}_{+}([0,1])$ is uniquely represented by an element of $\Psi ([0,1])$ + +Proposition 1. The set of Frenet curvatures $\mathcal{H}$ is a Riemannian submanifold of $\mathbb{L}^2([0,1],\mathbb{R}^{d-1})$ . + +Proof. The set $M = \{x \in \mathbb{R}^{d-1} | x_1, \ldots, x_{d-2} > 0\}$ is a open subset of the Riemannian manifold $\mathbb{R}^{d-1}$ . Then it is itself a differentiable Riemannian manifold with the standard inner product of $\mathbb{R}^{d-1}$ , and for any point $p \in M$ the tangent space $T_p(M)$ is $\mathbb{R}^{d-1}$ . The set of Frenet curvatures $\mathcal{H}$ is the set of measurable curves from [0, 1] to the Riemannian manifold $M$ and thus also a manifold ([26]). Its tangent space is $\mathbb{L}^2([0,1], \mathbb{R}^{d-1})$ and it can be equipped with the $\mathbb{L}^2$ Riemannian metric. + +Consequently, the geodesic distance on $\mathcal{H}$ is simply the $\mathbb{L}^2$ norm + +$$ +d _ {\mathcal {H}} \left(\boldsymbol {\theta} _ {0}, \boldsymbol {\theta} _ {1}\right) = \left\| \boldsymbol {\theta} _ {0} - \boldsymbol {\theta} _ {1} \right\| _ {\mathbb {L} ^ {2}} \tag {20} +$$ + +and the geodesic path is the straight line connecting them + +$$ +\alpha_ {\mathcal {H}} (\tau) = (1 - \tau) \boldsymbol {\theta} _ {0} + \tau \boldsymbol {\theta} _ {1}. \tag {21} +$$ + +Proposition 2. The map $\mathcal{R}_{\theta}: F([0,1], \mathbb{R}^d) \to \Psi([0,1]) \times \mathcal{H}$ , defined above, is a bijection. + +Proof. The element of $\Psi([0,1])$ uniquely defines the arclength function by $s(t) = \int_0^t (\sqrt{\dot{s}(u)})^2 du$ . As mentioned before, we have a bijection between the unparametrized Frenet curvatures in $\mathcal{H}$ and the unparametrized curve in the shape space $\mathcal{S}$ (1, 2). Then, from $X \in \mathcal{S}$ , the initial parametrized curve is simply $x(t) = X(s(t))$ . + +The set of such $\mathcal{R}_{\theta}$ is the Cartesian product of $\Psi ([0,1])$ and $\mathcal{H}$ and, therefore, is also a Riemannian manifold equipped with the product metric $d_{\Psi}\oplus d_{\mathcal{H}}$ [16]. The induced metric on $F([0,1],\mathbb{R}^d)$ under the representation $\mathcal{R}_{\theta}$ is + +$$ +d _ {\boldsymbol {\theta}} \left(x _ {0}, x _ {1}\right) = d _ {\Psi} \left(\sqrt {\dot {s} _ {0}}, \sqrt {\dot {s} _ {1}}\right) + d _ {\mathcal {H}} \left(\boldsymbol {\theta} _ {0}, \boldsymbol {\theta} _ {1}\right). \tag {22} +$$ + +In order to define a distance on the shape space $S$ from that one, we must quotient out the space $\mathrm{Diff}_{+}([0,1])$ . The action of $\mathrm{Diff}_{+}([0,1])$ on $\Psi ([0,1])$ is the same as 5, and the Frenet curvatures are invariant under reparametrization of the corresponding parametrized curve (Lemma 1). Moreover, by simply taking $h^* = s_1^{-1} \circ s_0 \in \mathrm{Diff}_{+}([0,1])$ , we have $\sqrt{s_1} * h^* = \sqrt{s_0}$ , and thus the distance on $\Psi ([0,1]) / \mathrm{Diff}_{+}([0,1])$ between $s_0$ and $s_1$ is zero. Hence, the induced distance on the shape space, between $X_0, X_1 \in S$ , under the representation $\mathcal{R}_{\theta}$ is defined as + +$$ +d _ {\mathcal {S}} ^ {(\boldsymbol {\theta})} \left(X _ {0}, X _ {1}\right) := d _ {\mathcal {H}} \left(\boldsymbol {\theta} _ {0}, \boldsymbol {\theta} _ {1}\right) = \left\| \boldsymbol {\theta} _ {0} - \boldsymbol {\theta} _ {1} \right\| _ {\mathbb {L} ^ {2}} \tag {23} +$$ + +and the geodesic path connecting them is + +$$ +\alpha_ {\mathcal {S}} ^ {(\boldsymbol {\theta})} (\tau) = \left(\sqrt {\dot {s} _ {0}}, \alpha_ {\mathcal {H}} (\tau)\right). \tag {24} +$$ + +This immediate representation by the Frenet curvatures appears in the experiments not to be sufficiently elastic (Figure 5). It has somewhat the same weakness as the angle representation proposed in [23], the Frenet curvatures being already independent of the parametrization. + +# 4.3. Square Root Curvatures Transform + +To overcome the "non-elasticity" issue of the representation defined above, we propose a second framework for shape analysis based on Frenet curvatures which uses, like the square root velocity function, the parametrization as a tool to register the curves and define a more "elastic" method. The latter is inspired by the square-root velocity transform of $SO(d)$ -valued curves. + +Definition 1 (SRV Transform for curves on $SO(d)$ ). Let $P \in C^{\infty}([0,1], SO(d))$ . The Square Root Velocity transform of $P$ is the map + +$$ +q (P) (t) = \frac {L _ {P (t) ^ {- 1}} \dot {P} (t)}{\sqrt {\| \dot {P} (t) \| _ {F}}} = \frac {P (t) ^ {T} \dot {P} (t)}{\sqrt {\| \dot {P} (t) \| _ {F}}}, \tag {25} +$$ + +where $\| .\| _F$ is the Frobenius norm associated with the scalar product on the Lie Algebra of skew-symmetric matrices $\langle A,B\rangle = \frac{1}{2} tr(A^T B) = -\frac{1}{2} tr(AB)$ . + +Let $x \in F([0,1],\mathbb{R}^d)$ and $Q(t) \in C^\infty([0,1],SO(d))$ be its associated Frenet path. Using the Frenet-Serret differential equation, the SRV Transform of the Frenet path is + +$$ +q (Q) (t) = \sqrt {\dot {s} (t)} \frac {A _ {\boldsymbol {\theta}} (s (t))}{\sqrt {\left\| A _ {\boldsymbol {\theta}} (s (t)) \right\| _ {F}}}. \tag {26} +$$ + +Proposition 3. Let $\theta \in \mathcal{H}$ , we have + +$$ +\left\| A _ {\boldsymbol {\theta}} (s (t)) \right\| _ {F} = \left\| \boldsymbol {\theta} (s (t)) \right\| _ {2}. +$$ + +Based on the SRV Transform of a Frenet path and Proposition 3, we propose a new transformation of a parametrized curve, which we have called the Square-Root Curvatures (SRC) transform. + +Definition 2 (Square-Root Curvatures Transform). Let $x \in F([0,1],\mathbb{R}^d)$ . We consider its associated arc-length function $s(t)$ and Frenet curvatures $\pmb{\theta}(s(t))$ defined as in Theorem 1. Then we define its square-root curvatures transform to be the map + +$$ +c (t) = \sqrt {\dot {s} (t)} \frac {\boldsymbol {\theta} (s (t))}{\sqrt {\| \boldsymbol {\theta} (s (t)) \|}}. \tag {27} +$$ + +The set of such square-root curvatures transforms is + +$$ +\mathcal {C} = \left\{c \in \mathbb {L} ([ 0, 1 ], \mathbb {R} ^ {d - 1}) \mid c _ {1}, \dots , c _ {d - 2} > 0 \right\}, \tag {28} +$$ + +which is the same as the set of admissible Frenet curvatures $\mathcal{H}$ . We have already shown in the previous section that this set is a Riemannian manifold equipped with the $\mathbb{L}^2$ metric. Therefore, the geodesic distance between $c_{0}, c_{1} \in \mathcal{C}$ is the $\mathbb{L}^2$ distance between them, and the geodesic path is a straight line. We define the following representation of a parametrized curve $x \in F([0,1], \mathbb{R}^d)$ , from its Square-Root Curvatures transform, by + +$$ +\mathcal {R} _ {\mathrm {S R C}} (x) (t) = \left(\sqrt {\dot {s} (t)}, c (t)\right). \tag {29} +$$ + +Proposition 4. The map $\mathcal{R}_{SRC} : F([0,1],\mathbb{R}^d) \to \Psi([0,1]) \times \mathcal{C}$ , defined above, is a bijection. + +Proof. This is again a result of theorems 1 and 2. To get $x$ from $\mathcal{R}_{\mathrm{SRC}}(x)$ , it should be noted firstly that $c(t) \| c(t) \| = \dot{s}(t) \theta(s(t))$ . From that, the skew-symmetric matrix function of the Frenet-Serret ODE can be reconstructed. By solving the corresponding Frenet-Serret ODE one gets the associated time parametrized Frenet path $Q(t)$ . Then, using the first component of $\mathcal{R}_{\mathrm{SRC}}(x)$ , we get $x(t) = X(s(t)) = \int_0^t \dot{s}(u) T(s(u)) du$ . + +The set of such square root curvature representations $\mathcal{R}_{\mathrm{SRC}}$ is the Cartesian product $\Psi ([0,1])\times \mathcal{C}$ and therefore a Riemannian manifold with the product metric $d_{\Psi}\oplus d_{\mathcal{C}}$ . This representation is, by definition, invariant under the action of $SO(d)$ . Then, the corresponding shape space is the quotient space $\Psi ([0,1])\times \mathcal{C} / \mathrm{Diff}([0,1])$ . Let's $x\in F([0,1],\mathbb{R}^d)$ and $h\in \mathrm{Diff}([0,1])$ . The SRC representation of $\tilde{x} = x\circ h$ is + +$$ +\mathcal {R} _ {\mathrm {S R C}} (\tilde {x}) = \left(\sqrt {\dot {s}} * h, c * h\right) = \mathcal {R} _ {\mathrm {S R C}} (x) * h \tag {30} +$$ + +where $*$ is the group action defined in 5. + +Proposition 5. The metric on $F([0,1],\mathbb{R}^d)$ induced by the Riemannian metric on $\Psi ([0,1])\times \mathcal{C}$ defined by $d_{SRC}\coloneqq d_{\Psi}\oplus d_{\mathcal{C}}$ is invariant under the action of $\text{Diff}_{+}([0,1])$ . + +The distance on the shape space $\mathcal{S}$ under the representation $\mathcal{R}_{\mathrm{SRC}}$ , between two elements $X_0, X_1 \in \mathcal{S}$ , is defined as + +$$ +d _ {S} ^ {\left(\mathrm {S R C}\right)} \left(X _ {0}, X _ {1}\right) := \inf _ {h \in \operatorname {D i f f} _ {+} ([ 0, 1 ])} d _ {\mathrm {S R C}} \left(x _ {0}, x _ {1} \circ h\right). \tag {31} +$$ + +From the optimal wrapping function $h^*$ the geodesic path on $S$ between them is + +$$ +\alpha_ {\mathcal {S}} ^ {(\mathrm {S R C})} (\tau) = \left( \begin{array}{c} \frac {\sin ((1 - \tau) \vartheta)}{\sin (\vartheta)} \sqrt {\dot {s} _ {0}} + \frac {\sin (\tau \vartheta)}{\sin (\vartheta)} \left(\sqrt {\dot {s} _ {1}} * h ^ {*}\right), \\ (1 - \tau) c _ {0} + \tau \left(c _ {1} * h ^ {*}\right)) \end{array} \right. \tag {32} +$$ + +where $\vartheta = d_{\Psi}(\sqrt{\dot{s}_0},\sqrt{\dot{s}_1}*h^*)$ . The registration problem consider here is to find the minimizer $h^*$ over $\mathrm{Diff}_{+}([0,1])$ of + +$$ +\int_ {0} ^ {1} \| c _ {0} (t) - \left(c _ {1} * h\right) (t) \| ^ {2} + \left\| \sqrt {\dot {s} _ {0} (t)} - \left(\sqrt {\dot {s} _ {1}} * h\right) (t) \right\| ^ {2} d t. \tag {33} +$$ + +Using the reformulation principle of [3], that is $\gamma = s_1\circ h\circ$ $s_0^{-1}\in \mathrm{Diff}_+([0,1])$ , this registration problem is shown to be equivalent to finding $\gamma^{*}\in \mathrm{Diff}_{+}([0,1])$ that minimizes + +$$ +\begin{array}{l} \int_ {0} ^ {1} \left\| \frac {\boldsymbol {\theta} _ {0} (s)}{\sqrt {\lVert \boldsymbol {\theta} _ {0} (s) \rVert}} - \sqrt {\gamma^ {\prime} (s)} \frac {\boldsymbol {\theta} _ {1} (\gamma (s))}{\sqrt {\lVert \boldsymbol {\theta} _ {1} (\gamma (s)) \rVert}} \right\| ^ {2} \\ + \| 1 - \gamma^ {\prime} (s) \| ^ {2} d s. \tag {34} \\ \end{array} +$$ + +Note that this reformulation has the form of a penalized registration problem. The second term represents a penalty term on $\gamma$ and ensures a certain smoothness of the warping function. In this framework, the deformable templates are the square-root normalized curvatures which encode more geometric information than the unit tangent vector. + +# 5. Experiments + +In this section, we report the experimental results of the proposed methods, comparing them with the SRVF method. We use both synthetic and real data. Additional results and figures are available in the supplementary material. + +# 5.1. Statistical estimation of the Frenet curvatures + +The main limitation of shape analysis methods based on the Frenet curvatures is the need of additional estimates of curvatures. Being dependent on higher-order derivatives (up to order $d$ ), they are quite sensitive to the observation noise of the Euclidean curve. We detail here a simple method that can be used for their smooth estimation, and we refer to [21, 18] for a more complex and detailed statistical estimation algorithm. First, it is possible to use a local polynomial smoothing algorithm to estimate the $d$ first + +derivatives of the Euclidean curve [6]. From these derivatives, the raw estimates of the Frenet curvatures can be computed by using their extrinsic formulas. We propose here a second method based on the Frenet-Serret ODE approximation, to obtain the raw estimates, that appear to be more stable. A simple middle point approximation of the ODE solution gives + +$$ +Q \left(s _ {j}\right) \approx Q \left(s _ {i}\right) \exp \left(\left(s _ {i} - s _ {j}\right) A _ {\boldsymbol {\theta}} \left(\frac {s _ {i} + s _ {j}}{2}\right)\right). \tag {35} +$$ + +where $\exp(.)$ is the exponential map of the Lie group $SO(d)$ . Then, using the inverse logarithm map $\log(.)$ , this gives an approximation of the matrix $A_{\theta}((s_i + s_j) / 2) \approx \frac{1}{s_i - s_j} \log(Q(s_i)^T Q(s_j))$ and by identification raw estimates of $\pmb{\theta}((s_i + s_j) / 2)$ . As we consider here a problem of estimating a functional parameter, we formulate the final $\pmb{\theta}$ estimation problem as a penalized weighted functional regression with the obtained raw estimates, that we solve by using a B-spline approximation of $\pmb{\theta}$ . + +# 5.2. Experiments with synthetic curves + +We use synthetic data to highlight the differences between the methods discussed above (SRVF, SRC, and Frenet curvatures). The computations related to the SRVF method are made with the package fdasrsf. The SRC and Frenet curvatures methods are implemented with the code provided as supplementary material, including a dynamic programming algorithm for solving the registration problems. + +We consider the simple case of a set of 20 curves in $\mathbb{R}^2$ with a single large peak of curvature. This one is created by generating curvature functions on $[0,1]$ with one peak of maximum value 60.5, width 0.15, and location chosen randomly between 0.1 and 0.9. These curves have the shape of a loop made with a wire, where the loop is more or less close to the right or left wire end, depending on the location of the curvature peak. + +![](images/cbc1abda0254f2cc9fe24155af5d9724f1ef96390b3e05fd1c98b5544b7308a7.jpg) +Figure 3: Matrices of pairwise SRVF distance (left), SRC distance (middle), and unparametrized Frenet curvatures distance (right) by sorted location of the peak curvature. + +We compare the three methods through the pairwise distance matrices in Figure 3, and the geodesic paths computed + +![](images/5fc0e004decb1b717e2aff42397bce2527a270eba077b5f9c81e20677126ddeb.jpg) +Figure 4: SRVF $(1^{st}$ row), SRC $(2^{nd}$ row), and Frenet curvatures $(3^{rd}$ row) geodesic paths between curves with curvature peaks located at 0.27 (left) and 0.78 (right). + +![](images/9a8a764d07a6fe0f604a75abe30c4ddf12a14d9dd63c2068fe0f1915570ec2bc.jpg) +Figure 5: Curvatures of the Euclidean curves along the geodesic paths plotted in Figure 4: SRVF $(1^{st}$ row), SRC $(2^{nd}$ row), and Frenet curvatures $(3^{rd}$ row). + +between two of these curves in Figure 4, with peaks located at 0.27 (red curve) and at 0.78 (blue curve). The corresponding deformations through the variations of the curvature along the different geodesic paths on Euclidean curves are shown in Figure 5, which highlights the strengths and weaknesses of each method. First, it emphasizes the "nonelasticity" of the unparametrized Frenet curvatures method, as in the middle of the geodesic path, we have two peaks of curvature and, therefore, a completely different shape without any loop. This explains the inconsistency of the heatmap under this method. Conversely, there is an elastic deformation of the curvature with the SRC transform, and shapes along the geodesic are consistent with the set of curves considered, which is well summarised on the corresponding heatmap where all distances are rather close to zero. For the SRVF method, the chosen example with peaks of curvature that are quite far apart shows that artifacts appear along the geodesic; the middle curve has two small loops at the edges. This phenomenon gives unreliable distances, as shown in the heatmap, where the distances are not monotone as a function of the spacing between the curvature peaks. + +By considering a set of curves characterized by specific features, we observe a clear difference in consistency of the shapes along the geodesics with respect to the different + +methods. This phenomenon is quite visible on the geodesic between helices in 2D, Figure 1 or 3D, Figure 2. In that case, within both geodesic paths under SRVF method, the curves lose the characteristic geometry of the helix. A three-dimensional circular helix is characterized by having a constant curvature and torsion, which is not the case along the SRVF geodesic, but preserved with the SRC method. In that case the geodesic under the Frenet curvature representation is very similar to SRC. + +# 5.3. Application to sign language motion data + +It appears that curvilinear velocity and Frenet curvatures are particularly relevant parameters for the analysis of human motion. Several laws involving these parameters can be found in the literature [10, 14, 19, 7]; among others, the power laws state a special relationship between the curvature, the torsion, and the velocity of a point trajectory representing human motion. Using a method that conserves the shape of these parameters is, therefore, of particular interest in this application. We demonstrate here with the case of wrist trajectories in sign language, acquired with a motion capture system by the company MocapLab (https://www.mocaplab.com/fr/). We compute the geodesic paths, under each of the frameworks, between the arbitrarily chosen red and blue curves within the set of several repetitions of the sign "Femme", shown in Figure 6 with the corresponding time-parameterized Frenet curvatures. + +![](images/12031780e212557ad881f4993741d2ee443aaaf43a8f0ee4cb443105d3e87eb4.jpg) +Figure 6: Trajectories of the right wrist while signing "Femme" in sign language: 3D curves (left), time-parametrized curvatures (top right), torsions (bottom right). The blue and red ones are used to compute the geodesic in Figure 7. + +Figure 7 emphasizes the advantage of considering a representation depending on the parameterization, allowing a registration before computing the geodesic. However, it also shows that considering only the tangent vector as a representative object (SRVF) is not sufficient to find the optimal reparametrization that correctly aligns the torsions, + +![](images/d253497a348fef4db1a60db1dab9aaa3da2452f6ac63983d86347dfd6a6ea21b.jpg) + +![](images/24d15a2e1a679858767a95bfd2d27b4406775c97261d47fbaa8ae71f4cb7ff39.jpg) + +![](images/92e3396783fc03519b46608c579f4863a1642dd9b2d192bc286e3092cb2e3094.jpg) + +![](images/28b92fb8e0014377beb36b6b3827aa6da9697bc4ae4a51c603e6096c8b6f85e5.jpg) +(a) SRVF + +![](images/d2655f7d998c4b48da89d11ae702c92a33e690d35bc6e453b57ac33e133d43a8.jpg) +(b) SRC + +![](images/355efc6332217c06be6fc0cde171206efd88da0d6e977d02b9b8971691864cf5.jpg) +(c) Frenet curvatures + +![](images/7940fa0bb2ddd7117f5c3e07cf8db4e39c36103725fcc4d71cfb7725b3fef637.jpg) +Figure 7: Comparison between time-parametrized curvature and torsion along the geodesic path under SRVF (left), SRC (middle), and Frenet curvatures (right). +(a) Warping functions $h$ +Figure 8: Comparison of estimated warping functions $h$ (left) and $\gamma$ (right) to compute the geodesic in Figure 7. + +![](images/386b59cb3851d2575d791fcbdca101540e1cd500a4e70bcdd85b65cc802647e0.jpg) +(b) Warping functions $\gamma$ + +and could affect subsequent analyses made by the SRVF method, such as the mean. This results in the appearance of new minimums, maximums and zeros in the torsion functions along the SRVF geodesic. However, such characteristic points of the curvature, torsion and velocity functions are crucial in the observation of the laws of motion. It is therefore preferable to use a method that preserves these characteristics by directly optimizing the optimal alignment from these parameters, such as the proposed SRC method. + +# 6. Conclusion + +The square-root curvature transform of a Euclidean curve in $\mathbb{R}^d$ is a representation that encodes more geometric information of the curves, and thus the results are easier to interpret than existing methods. The main limitation lies in the estimation of the Frenet curvatures from real and noisy data, nevertheless recent smooth statistical estimators can be used for computing the SRC [21, 18]. We believe our + +method is particularly interesting for motion trajectory analysis and could be developed further in the future as a tool for generation, segmentation and classification of complex trajectories. + +# References + +[1] Martin Bauer, Nicolas Charon, Eric Klassen, and Alice Le Brigant. Intrinsic riemannian metrics on spaces of curves: Theory and computation. Handbook of Mathematical Models and Algorithms in Computer Vision and Imaging, pages 1-35, 2021. 1 +[2] Martin Bauer, Nicolas Charon, Eric Klassen, Sebastian Kurtek, Tom Needham, and Thomas Pierron. Elastic Metrics on Spaces of Euclidean Curves: Theory and Algorithms, 2022. arXiv:2209.09862 [math]. 1, 2 +[3] Nicolas J.-B. Brunel and Juhyun Park. The frenet-serret framework for aligning geometric curves. In Frank Nielsen and Frédéric Barbaresco, editors, Geometric Science of Information, pages 608-617, Cham, 2019. Springer International Publishing. 3, 4, 5, 7 +[4] Elena Celledoni, Markus Eslitzbichler, and Alexander Schmeding. Shape analysis on lie groups with applications in computer animation. Journal of Geometric Mechanics, 8(3):273-304, 2016. 2 +[5] Marvin Eisenberger and Daniel Cremers. Hamiltonian dynamics for real-world shape interpolation. In European Conference on Computer Vision (ECCV), 2020. 1 +[6] Jianqing Fan and Irene Gijbels. Local polynomial modelling and its applications. CRC monographs on statistics and applied probability 66. Chapman & Hall, 1996. 7 +[7] Tamar Flash and Alain Berthoz. *Space-Time Geometries for Motion and Perception in the Brain and the Arts*. Lecture Notes in Morphogenesis. Springer, 2021. 8 + +[8] Lukas Koestler, Daniel Grittner, Michael Moeller, Daniel Cremers, and Zorah Lähner. Intrinsic neural fields: Learning functions on manifolds. In European Conference on Computer Vision (ECCV), 2022. 1 +[9] Wolfgang Kühnel. Differential Geometry Curves - Surfaces, volume 77. 4, 5 +[10] Francesco Lacquaniti, Carlo Terzuolo, and Paolo Viviani. The law relating the kinematic and figural aspects of drawing movements. Acta Psychologica (Amst.), 54:115-130, 1983. 8 +[11] Serge Lang. Differential and riemannian manifolds. Springer, 1999:1-6, 2006. 1, 2 +[12] Robert Lyons. Frobenius theorem two ways. Lecture note, 2016. 5 +[13] Zorah Lähner, Emanuele Rodola, Frank R. Schmidt, Michael M. Bronstein, and Daniel Cremers. Efficient globally optimal 2d-to-3d deformable shape matching. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), May 2016. 1 +[14] Uri Maoz, Alain Berthoz, and Tamar Flash. Complex unconstrained three-dimensional hand movement and constant equi-affine speed. Journal of Neurophysiology, 101:1002-1015, 2009. 8 +[15] J. S. Marron, James O. Ramsay, Laura M. Sangalli, and Anuj Srivastava. Functional data analysis of amplitude and phase variation. Statistical Science, 30:468-484, 2015. 5 +[16] Turk J Math. Submanifolds of Riemannian Product Manifolds. 29:389-401, 2005. 5 +[17] Tom Needham. Shape Analysis of Framed Space Curves. Journal of Mathematical Imaging and Vision, 61(8):1154-1172, Oct. 2019. 1 +[18] Juhyun Park, Nicolas Brunel, and Perrine Chassat. Curvature and torsion estimation of 3d functional data: A geometric approach to build the mean shape under the frenet serret framework, 2022. arXiv:2203.02398 [stat]. 7, 9 +[19] Frank E. Pollick, Uri Maoz, Amir A. Handzel, Peter J. Giblin, Guillermo Sapiro, and Tamar Flash. Three-dimensional arm movements at constant equi-affine speed. Cortex, 45:325-339, 2009. 8 +[20] Marianna Saba. On the usage of the curvature for the comparison of planar curves. PhD thesis, University of Cagliari, 2012. 2, 5 +[21] Laura M Sangalli, Piercesare Secchi, Simone Vantini, and Alessandro Veneziani. Efficient estimation of three-dimensional curves and their derivatives by free-knot regression splines, applied to the analysis of inner carotid artery centrelines. Journal of the Royal Statistical Society: Series C (Applied Statistics), 58(3):285-306, 2009. 7, 9 +[22] Stefan Sommer, Tom Fletcher, and Xavier Pennec. Introduction to differential and Riemannian geometry. 2020. 2 +[23] Anuj Srivastava and Eric Klassen. Functional and Shape Data Analysis. Springer Series in Statistics. Springer New York, 2016. 1, 2, 3, 5, 6 +[24] Anuj Srivastava, Eric Klassen, Shantanu H. Joshi, and Ian H. Jermyn. Shape analysis of elastic curves in euclidean spaces. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(7):1415-1428, 2011. 1, 3 + +[25] Tatiana Surazhsky and Gershon Elber. Metamorphosis of planar parametric curves via curvature interpolation. International Journal of Shape Modeling, 8:201-216, 2002. 2, 5 +[26] Wang Tixiang. Morse theory on banach manifolds. Acta Mathematica Sinica, 5:250-262, 1989. 5 +[27] J. Derek Tucker, Wei Wu, and Anuj Srivastava. Generative models for functional data using phase and amplitude separation. Computational Statistics and Data Analysis, 61:50-66, 2013. 5 +[28] Laurent Younes. Computable Elastic Distances between Shapes. SIAM Journal on Applied Mathematics, 58(2):565-586, 1998. Publisher: Society for Industrial and Applied Mathematics. 1 +[29] Laurent Younes. *Shapes and Diffeomorphisms*. Applied Mathematical Sciences. Springer, 2010. 3 +[30] Laurent Younes. Elastic distance between curves under the metamorphosis viewpoint, 2018. arXiv:1804.10155. 3 \ No newline at end of file diff --git a/2023/Shape Analysis of Euclidean Curves under Frenet-Serret Framework/images.zip b/2023/Shape Analysis of Euclidean Curves under Frenet-Serret Framework/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..54c16070ebf978f91fe96c03bffee0436317013d --- /dev/null +++ b/2023/Shape Analysis of Euclidean Curves under Frenet-Serret Framework/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a792ce28e5fbb1daf57a413f61fff358291335aeb38cb02421d10c6e95cc7a30 +size 478880 diff --git a/2023/Shape Analysis of Euclidean Curves under Frenet-Serret Framework/layout.json b/2023/Shape Analysis of Euclidean Curves under Frenet-Serret Framework/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..560bec054f294cf1800e9bef43fe34e9e0365eb1 --- /dev/null +++ b/2023/Shape Analysis of Euclidean Curves under Frenet-Serret Framework/layout.json @@ -0,0 +1,12003 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "type": "text", + "content": "Shape Analysis of Euclidean Curves under Frenet-Serret Framework" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 143, + 240, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 143, + 240, + 185 + ], + "spans": [ + { + "bbox": [ + 61, + 143, + 240, + 185 + ], + "type": "text", + "content": "Perrine Chassat1 \n1LaMME, University of Paris-Saclay \nperrine.chassat@univ-evry.fr" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 264, + 143, + 379, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 143, + 379, + 185 + ], + "spans": [ + { + "bbox": [ + 264, + 143, + 379, + 185 + ], + "type": "text", + "content": "Juhyun Park" + }, + { + "bbox": [ + 264, + 143, + 379, + 185 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 264, + 143, + 379, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 264, + 143, + 379, + 185 + ], + "type": "text", + "content": "ENSIIE, Evry \njuhyun.park@ensiie.fr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 401, + 143, + 532, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 401, + 143, + 532, + 185 + ], + "spans": [ + { + "bbox": [ + 401, + 143, + 532, + 185 + ], + "type": "text", + "content": "Nicolas Brunel1,2,3 \n3Quantmetry, Paris \nnicolas.brunel@ensiie.fr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 239, + 290, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 455 + ], + "type": "text", + "content": "Geometric frameworks for analyzing curves are common in applications as they focus on invariant features and provide visually satisfying solutions to standard problems such as computing invariant distances, averaging curves, or registering curves. We show that for any smooth curve in " + }, + { + "bbox": [ + 46, + 239, + 290, + 455 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 239, + 290, + 455 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 239, + 290, + 455 + ], + "type": "inline_equation", + "content": "d > 1" + }, + { + "bbox": [ + 46, + 239, + 290, + 455 + ], + "type": "text", + "content": ", the generalized curvatures associated with the Frenet-Serret equation can be used to define a Riemannian geometry that takes into account all the geometric features of the shape. This geometry is based on a Square Root Curvature Transform that extends the square root-velocity transform for Euclidean curves (in any dimensions) and provides likely geodesics that avoid artefacts encountered by representations using only first-order geometric information. Our analysis is supported by simulated data and is especially relevant for analyzing human motions. We consider trajectories acquired from sign language, and show the interest of considering curvature and also torsion in their analysis, both being physically meaningful." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 478, + 128, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 128, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 128, + 490 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 498, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 715 + ], + "type": "text", + "content": "Identifying and comparing different types of visual objects is a fundamental task in machine learning and computer vision problems [13, 5, 8]. The shape is one of the essential features of objects that allow us to understand and characterize them. Nowadays, it is much easier to obtain data in the form of shapes, typically as dense point clouds or landmarks. The main task in shape analysis is to define a proper framework to compare and quantify the variation of the shapes. However, the shape space is generally nonlinear, and extracting meaningful information or features is complex. One of the successful approaches to shape analysis utilizes a Riemannian framework of differential geometry, where a metric can be defined between the shapes, which is invariant with respect to shape-preserving transformations such as translation and rotation. For instance, this gives rise to geodesic distances that are naturally invariant to smooth and optimal deformations through geodesic paths between the shapes [11]. This approach is very versatile as it can" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 213, + 547, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 562 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 562 + ], + "type": "text", + "content": "be adapted to various kinds of manifold-value data and can be designed to emphasize important geometric information to be preserved. As a consequence, several choices of metrics are possible, such as the class of invariant Sobolev metrics, often called elastic, for the analysis of curves [1]. In this work, we are concerned with curves that often arise in the application as trajectories (function of time) or motions (animation, activity recognition) and with the definition of a framework to compare their shapes. The differential geometry of Euclidean curves is among the simplest (with respect to higher dimensional manifolds), and relatively simple Riemannian metrics are available with different mathematical representations of curves [28]. Quite remarkably, the introduction of the Square Root Velocity (SRV) transform [24] that consists of a particular representation of the shape of a curve enables to define a so-called elastic Riemannian distance, which has proven to be useful for the statistical shape analysis of 2D and 3D curves in applications. The SRV possesses interesting properties such as a principled theoretical framework, efficient computation, and generalization to higher dimensions [2]. Nevertheless, a limitation of the SRV transform and the corresponding elastic distance is the restrictive use of the first-order derivative, while the geometry of 3-D (or " + }, + { + "bbox": [ + 304, + 213, + 547, + 562 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 213, + 547, + 562 + ], + "type": "text", + "content": "-D) curves depends on the derivatives until order " + }, + { + "bbox": [ + 304, + 213, + 547, + 562 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 213, + 547, + 562 + ], + "type": "text", + "content": ". Indeed, it is well-known that a 3D curve is characterized by its curvature and its torsion: this is particularly critical when we consider trajectories or human movements, where the curvature and torsion can have a physical meaning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 573, + 488, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 573, + 488, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 573, + 488, + 586 + ], + "type": "text", + "content": "2. Related works and contributions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "As we will recall in section 3, the full geometry of a curve can be given either by the Frenet curvatures (standard curvature and torsion in 3D) or by the path of Frenet frames. There have been few attempts to directly deal with the Frenet curvatures: most of the works have been produced in 2D curves as an alternative representation [23]. Nevertheless, the potential for applications has not been investigated. In [17], the elastic shape analysis framework has been considered for 3D curves based on the Frenet frames, but the link to the physical parameters has been overlooked." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4027" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "We can also mention the shape analysis of curves on Lie groups [4] with application in computer animation. Outside the Riemannian framework, an attempt has been made to use a direct curvature-based interpolation of curves [20, 25]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 120, + 289, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 289, + 515 + ], + "type": "text", + "content": "In this work, we introduce two representations of Euclidean curves for their shape analysis that use their complete geometry through Frenet curvatures. We provide the full development of the Riemannian frameworks associated with these two representations. As a consequence and in comparison with existing methods, our approaches also give explicit formulas for geodesics and geodesic distances. The first representation considered is based directly on unparametrized Frenet curvatures. We show through experiments that it defines a shape analysis framework that lacks elasticity. As the main contribution, we propose the definition of a second representation, called the Square-Root Curvature (SRC) Transform, which takes into account reparameterization and defines a metric on the space of shapes through the quotient space with the group of diffeomorphisms. One can imagine that the classical method associated with the SRVF, defining a Sobolev elastic metric [2], already implicitly uses all the geometric information necessary for a relevant curve analysis. We show here with simple examples that this is not the case. Through experiments on synthetic data, we compare the methods, and illustrate the limitations of the SRVF one, due to its lack of use of geometric information. To be able to judge and compare the quality of these metrics, we compare consistent sets of curves characterized by specific features. The SRC method shows a special strength in defining a framework that remains consistent with these sets. The straightforward example of geodesics between helices with different numbers of spins (Figure 1 in 2D and Figure 2 in 3D) shows that, in contrast, this is not the case for the SRVF method. In addition, we highlight the interest of these Frenet curvatures-based representations in the real application case of human motion trajectory analysis." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 525, + 263, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 263, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 263, + 539 + ], + "type": "text", + "content": "3. Riemannian Geometry on Shape Space" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 544, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 582 + ], + "type": "text", + "content": "We introduce useful notations and we review the main approach for constructing tractable representations of the shape of a curve and deriving a Riemannian geometry." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 587, + 239, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 239, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 239, + 601 + ], + "type": "text", + "content": "3.1. Shape Analysis of Euclidean Curves" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "We consider absolutely continuous curves that are smooth, open, and with values in some Euclidean space " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": ", we denote this set as " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "AC([0,1],\\mathbb{R}^d)" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": ". These curves are typically parametrized by a variable " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " that can usually be interpreted as time. Nevertheless, from a (statistical) shape analysis point of view, we focus on the geometric shape of curves that do not depend on a specific parametrization or standard transformations such as translations, rotations, scaling, or reparametrizations. To distinguish be" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "content": "tween parametrized curves that differ only by translation, we consider the set of absolutely continuous curves where " + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "inline_equation", + "content": "x(0) = 0" + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "content": ", denoted by " + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "inline_equation", + "content": "AC_0([0,1],\\mathbb{R}^d)" + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "content": ". The natural and intrinsic parametrization that uniquely defines the shape of a curve " + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "content": " is the arc-length parametrization, defined with the arc length function " + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "inline_equation", + "content": "s(t) = \\int_0^t\\| \\dot{x} (u)\\| du" + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "inline_equation", + "content": "t\\in [0,1]" + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "content": ". In order to remove the scaling variability, the total length of the curve " + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "inline_equation", + "content": "s(1)" + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "content": " is set to 1. Under this parametrization, the shape " + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "inline_equation", + "content": "X:[0,1]\\mapsto \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "content": " of the curve is the image of the function " + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "inline_equation", + "content": "x(t) = X(s(t))" + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "content": ". As we want to study shapes independently of their parameterizations, we introduce the reparametrization group " + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "inline_equation", + "content": "\\mathrm{Diff}_{+}([0,1])" + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "content": ", of smooth orientation preserving diffeomorphisms of the interval [0, 1] onto itself. This group acts on the space of absolutely continuous curves by right composition, and this action only alters the parametrization of the curve, not the inherent shape " + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "content": ". The space of such shapes (or unparametrized curves) is often mathematically defined as the quotient space" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 323, + 297, + 545, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 297, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 323, + 297, + 545, + 312 + ], + "type": "interline_equation", + "content": "\\mathcal {S} ([ 0, 1 ], \\mathbb {R} ^ {d}) = A C _ {0} ([ 0, 1 ], \\mathbb {R} ^ {d}) / \\operatorname {D i f f} _ {+} ([ 0, 1 ]). \\tag {1}", + "image_path": "9a09a9f95a9708cd669686d197b6e402da4c91cac441fe0475edcb41152e28ea.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 320, + 546, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 320, + 546, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 320, + 546, + 632 + ], + "type": "text", + "content": "The purpose of shape analysis of curves is to define a distance function " + }, + { + "bbox": [ + 304, + 320, + 546, + 632 + ], + "type": "inline_equation", + "content": "d_{S}" + }, + { + "bbox": [ + 304, + 320, + 546, + 632 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 320, + 546, + 632 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 320, + 546, + 632 + ], + "type": "text", + "content": " and a framework to perform a complete statistical analysis on a set of curves in " + }, + { + "bbox": [ + 304, + 320, + 546, + 632 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 320, + 546, + 632 + ], + "type": "text", + "content": " (e.g. mean, classification, or Principal Component Analysis etc.). One of the main challenges in defining this distance is to choose an appropriate mathematical representation of the curves that can be made invariant to all shape-preserving transformations - translation, rotation, scaling, and reparametrization. Moreover, one of the stakes of such representation is to offer an (infinite-dimensional) Riemannian manifold structure that brings powerful and flexible tools for studying the geometry of shapes or statistical properties notably thanks to the tangent space of the manifold [11, 22]. In [23], a list of the few possible representations is given - coordinate functions, curvatures, angle function, and square-root velocity function (SRVF) - and a framework for curve analysis is derived for the last two ones. While the angle representation is unparameterized, the SRVF representation depends on the parametrization, which is shown to be very useful as a tool for the registration of points across curves. As a consequence, the parameterization group " + }, + { + "bbox": [ + 304, + 320, + 546, + 632 + ], + "type": "inline_equation", + "content": "\\mathrm{Diff}_{+}([0,1])" + }, + { + "bbox": [ + 304, + 320, + 546, + 632 + ], + "type": "text", + "content": " must be eliminated by using a quotient space. The classical approach is to define the Riemannian metric on the shape space through a metric on the space of parametrized representations that is invariant to reparametrization: " + }, + { + "bbox": [ + 304, + 320, + 546, + 632 + ], + "type": "inline_equation", + "content": "\\forall h \\in \\mathrm{Diff}_{+}([0,1])" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 347, + 640, + 545, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 640, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 347, + 640, + 545, + 654 + ], + "type": "interline_equation", + "content": "d _ {A C _ {0}} \\left(x _ {0}, x _ {1}\\right) = d _ {A C _ {0}} \\left(x _ {0} \\circ h, x _ {1} \\circ h\\right). \\tag {2}", + "image_path": "c507f8c99501b8c2b2fd249bb2dc360b7d6545176925e7c6bdaf6426e71e8de4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 662, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 662, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 662, + 545, + 687 + ], + "type": "text", + "content": "In that case, the distance on " + }, + { + "bbox": [ + 304, + 662, + 545, + 687 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 662, + 545, + 687 + ], + "type": "text", + "content": " is defined as the infimum over all possible reparametrization. For " + }, + { + "bbox": [ + 304, + 662, + 545, + 687 + ], + "type": "inline_equation", + "content": "X_0, X_1 \\in S" + }, + { + "bbox": [ + 304, + 662, + 545, + 687 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 333, + 696, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 696, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 333, + 696, + 545, + 715 + ], + "type": "interline_equation", + "content": "d _ {\\mathcal {S}} \\left(X _ {0}, X _ {1}\\right) := \\inf _ {h \\in \\operatorname {D i f f} ([ 0, 1 ])} d _ {A C _ {0}} \\left(x _ {0}, x _ {1} \\circ h\\right). \\tag {3}", + "image_path": "64c3d605839f3fc57739db680422bed9fac0fb31ac9dff2a350b05a706297010.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4028" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 71, + 504, + 189 + ], + "blocks": [ + { + "bbox": [ + 91, + 71, + 504, + 189 + ], + "lines": [ + { + "bbox": [ + 91, + 71, + 504, + 189 + ], + "spans": [ + { + "bbox": [ + 91, + 71, + 504, + 189 + ], + "type": "image", + "image_path": "15bd225d8a661ba455ccb22c20f9ce4350d3a602c6cf25deb1378fec83b2dabe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 205, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 545, + 220 + ], + "type": "text", + "content": "Figure 1: Geodesic paths between two 2D scaled spirals with different number of spins: SRVF " + }, + { + "bbox": [ + 46, + 205, + 545, + 220 + ], + "type": "inline_equation", + "content": "(1^{st}" + }, + { + "bbox": [ + 46, + 205, + 545, + 220 + ], + "type": "text", + "content": " row) and SRC " + }, + { + "bbox": [ + 46, + 205, + 545, + 220 + ], + "type": "inline_equation", + "content": "(2^{nd}" + }, + { + "bbox": [ + 46, + 205, + 545, + 220 + ], + "type": "text", + "content": " row)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 239, + 288, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 288, + 275 + ], + "type": "text", + "content": "In the following, we will denote with a dot the derivation with respect to the time variable, and with a prime the one with respect to the arc-length parameter." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 282, + 227, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 282, + 227, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 227, + 295 + ], + "type": "text", + "content": "3.2. Square Root Velocity Framework" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 300, + 288, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 288, + 357 + ], + "type": "text", + "content": "The square-root velocity function framework is the most commonly used representation for curve shape analysis in " + }, + { + "bbox": [ + 47, + 300, + 288, + 357 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d" + }, + { + "bbox": [ + 47, + 300, + 288, + 357 + ], + "type": "text", + "content": " [24, 23]. The square-root velocity function (SRVF) of " + }, + { + "bbox": [ + 47, + 300, + 288, + 357 + ], + "type": "inline_equation", + "content": "x\\in AC_0([0,1],\\mathbb{R}^d)" + }, + { + "bbox": [ + 47, + 300, + 288, + 357 + ], + "type": "text", + "content": ", denoted by " + }, + { + "bbox": [ + 47, + 300, + 288, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{SRVF}}(x) = q" + }, + { + "bbox": [ + 47, + 300, + 288, + 357 + ], + "type": "text", + "content": ", is defined as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 95, + 357, + 287, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 357, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 95, + 357, + 287, + 385 + ], + "type": "interline_equation", + "content": "q (t) = \\frac {\\dot {x} (t)}{\\sqrt {\\| \\dot {x} (t) \\|}} = \\sqrt {\\dot {s} (t)} T (s (t)) \\tag {4}", + "image_path": "630c833de6b7b909bdeba4f756db401b5ab7fcf39c0b315aa293cd80696c9e38.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "inline_equation", + "content": "T(s(t)) = \\frac{\\dot{x}(t)}{\\|\\dot{x}(t)\\|} = X'(s(t))" + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "text", + "content": " is the unit tangent vector of the curve. This transformation is a bijection with " + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "inline_equation", + "content": "AC_0([0,1],\\mathbb{R}^d)" + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "text", + "content": " and its explicit inverse is " + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "inline_equation", + "content": "x(t) = \\int_0^t q(u)|q(u)|du" + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "text", + "content": ". As we consider length-normalized curves, the SRVFs have a unit " + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^2" + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "text", + "content": " norm, and their set is the convenient unit Hilbert sphere, a Riemannian submanifold of " + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^2([0,1],\\mathbb{R}^d)" + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "text", + "content": " (with the " + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^2" + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "text", + "content": " inner product). Then, the " + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^2" + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "text", + "content": " metric on SRVF induces a Riemannian metric on " + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "inline_equation", + "content": "AC_0([0,1],\\mathbb{R}^d)" + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "text", + "content": " where geodesics are given by the shorter arcs on great circles between SRV functions. The action of " + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "inline_equation", + "content": "\\mathrm{Diff}_{+}([0,1])" + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "inline_equation", + "content": "AC_0([0,1],\\mathbb{R}^d)" + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "text", + "content": " is reflected on " + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "text", + "content": " by the group action denoted by " + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 46, + 389, + 287, + 538 + ], + "type": "text", + "content": " and defined as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 111, + 546, + 287, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 546, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 111, + 546, + 287, + 566 + ], + "type": "interline_equation", + "content": "(q * h) (t) = \\sqrt {\\dot {h} (t)} q (h (t)) \\tag {5}", + "image_path": "500ea2802136318f191d98c3b6c2a6e1bd2549b64adc24a7375ebf2717a11a8b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 572, + 287, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 572, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 46, + 572, + 287, + 621 + ], + "type": "text", + "content": "and if the curve is rotated by a matrix " + }, + { + "bbox": [ + 46, + 572, + 287, + 621 + ], + "type": "inline_equation", + "content": "O \\in SO(d)" + }, + { + "bbox": [ + 46, + 572, + 287, + 621 + ], + "type": "text", + "content": ", its SRVF gets rotated by the same matrix. The key property of this representation is the invariance of its associated distance under the action of " + }, + { + "bbox": [ + 46, + 572, + 287, + 621 + ], + "type": "inline_equation", + "content": "\\mathrm{Diff}_{+}([0,1])" + }, + { + "bbox": [ + 46, + 572, + 287, + 621 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 572, + 287, + 621 + ], + "type": "inline_equation", + "content": "SO(d)" + }, + { + "bbox": [ + 46, + 572, + 287, + 621 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 78, + 628, + 287, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 628, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 78, + 628, + 287, + 643 + ], + "type": "interline_equation", + "content": "\\left\\| O \\left(q _ {0} * h\\right) - O \\left(q _ {1} * h\\right) \\right\\| _ {\\mathbb {L} ^ {2}} = \\left\\| q _ {0} - q _ {1} \\right\\| _ {\\mathbb {L} ^ {2}}. \\tag {6}", + "image_path": "07ab6b61301f2180c45bc8e343e056894dfa14be0d8b57104eb53c1eba8ade5e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 649, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 649, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 649, + 287, + 674 + ], + "type": "text", + "content": "The metric can then be used to define a proper distance on the shape space " + }, + { + "bbox": [ + 46, + 649, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{S}([0,1],\\mathbb{R}^d)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 681, + 287, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 681, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 53, + 681, + 287, + 711 + ], + "type": "interline_equation", + "content": "d_{\\mathcal{S}}^{(\\mathrm{SRVF})}(X_{0},X_{1}):= \\inf_{\\substack{O\\in SO(d)\\\\ h\\in \\mathrm{Diff}([0,1])}}\\cos^{-1}\\langle q_{0},O(q_{1}*h)\\rangle \\tag{7}", + "image_path": "4583505f745dcd7bc2a3265b05037c4298659f26d6c70e2dc490b3be0a7768fb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 239, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 239, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 305, + 239, + 545, + 263 + ], + "type": "text", + "content": "and the geodesic path on the shape space is taken between " + }, + { + "bbox": [ + 305, + 239, + 545, + 263 + ], + "type": "inline_equation", + "content": "q_0" + }, + { + "bbox": [ + 305, + 239, + 545, + 263 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 239, + 545, + 263 + ], + "type": "inline_equation", + "content": "q_1 * h" + }, + { + "bbox": [ + 305, + 239, + 545, + 263 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 263, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 263, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 305, + 263, + 545, + 299 + ], + "type": "text", + "content": "The definition of this distance on the shape space under the SRVF representation can be interpreted as the following registration problem" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 303, + 545, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 303, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 308, + 303, + 545, + 350 + ], + "type": "interline_equation", + "content": "h ^ {*}, O ^ {*} = \\underset { \\begin{array}{c} O \\in S O (d) \\\\ h \\in \\operatorname {D i f f} ([ 0, 1 ]) \\end{array} } {\\arg \\min } \\int_ {0} ^ {1} \\| q _ {0} (t) - O q _ {1} (h (t)) \\sqrt {\\dot {h} (t)} \\| _ {2} ^ {2} d t. \\tag {8}", + "image_path": "bc51e7630fb8e122ae5f6cca0e94549d482e3237dbfd96d3503d05a22ffd2fda.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 350, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 350, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 350, + 545, + 422 + ], + "type": "text", + "content": "In [3], this registration problem has been reformulated with the unit tangent vector and the arc length functions. By defining " + }, + { + "bbox": [ + 305, + 350, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\gamma = s_{1} \\circ h \\circ s_{0}^{-1} \\in \\mathrm{Diff}_{+}([0,1])" + }, + { + "bbox": [ + 305, + 350, + 545, + 422 + ], + "type": "text", + "content": " the optimization problem 8 amounts to finding the optimal diffeomorphism of " + }, + { + "bbox": [ + 305, + 350, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\mathrm{Diff}_{+}([0,1])" + }, + { + "bbox": [ + 305, + 350, + 545, + 422 + ], + "type": "text", + "content": " that acts on the arc-length parameter " + }, + { + "bbox": [ + 305, + 350, + 545, + 422 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 305, + 350, + 545, + 422 + ], + "type": "text", + "content": " and solves the minimization problem:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 427, + 545, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 427, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 305, + 427, + 545, + 473 + ], + "type": "interline_equation", + "content": "\\gamma^ {*}, O ^ {*} = \\underset { \\begin{array}{c} O \\in S O (d) \\\\ \\gamma \\in \\operatorname {D i f f} ([ 0, 1 ]) \\end{array} } {\\arg \\min } \\int_ {0} ^ {1} \\| T _ {0} (s) - O T _ {1} (\\gamma (s)) \\sqrt {\\gamma^ {\\prime} (s)} \\| _ {2} ^ {2} d s. \\tag {9}", + "image_path": "256317468ae50abb36946df0814bfa1e0045a9338e58a7fffd9259335f3d160f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": "It should be noted, in this reformulation, that the object " + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "inline_equation", + "content": "OT_{1}(\\gamma(s))\\sqrt{\\gamma'(s)}" + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": " does not represent the same shape as " + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "inline_equation", + "content": "X_{1}(s)" + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": " in the shape space. Here the element " + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{Diff}_{+}([0,1])" + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": " is not used as a reparametrization of the curve but to deform the element of " + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "inline_equation", + "content": "S([0,1],\\mathbb{R}^{d})" + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": ". Under this point of view, the set of unit tangent vectors that can be reached by deforming the vector " + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "inline_equation", + "content": "T(s)" + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": ", with the group action " + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "inline_equation", + "content": "T*\\gamma" + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": " defines an equivalence class of shapes associated with that one, as in the setting of deformable templates of Grenander's theory [29, 30]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Finally, the choice of a parametrized curve representation for shape analysis, discussed in [23], can be seen as the problem of choosing a good geometric representative of the shape as a template and defining an associated registration problem. Hence, an appropriate choice may be seen as a matter of modeling and should be done in interaction with the type of data analyzed and the dimension of the space. In the next sections, we use " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " to refer to functions of " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Diff}_{+}([0,1])" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " that act on the time variable " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " for ones that act on the arc-length variable " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4029" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 68, + 492, + 214 + ], + "blocks": [ + { + "bbox": [ + 102, + 68, + 492, + 214 + ], + "lines": [ + { + "bbox": [ + 102, + 68, + 492, + 214 + ], + "spans": [ + { + "bbox": [ + 102, + 68, + 492, + 214 + ], + "type": "image", + "image_path": "d3bf35c281d578c5961f542676569d9ea11d9beb808bbab0998bce4336337efb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 228, + 547, + 253 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 547, + 253 + ], + "type": "text", + "content": "Figure 2: Geodesic paths between two scaled 3D circular helices with different number of spins: SRVF (1" + }, + { + "bbox": [ + 46, + 228, + 547, + 253 + ], + "type": "inline_equation", + "content": "^{st}" + }, + { + "bbox": [ + 46, + 228, + 547, + 253 + ], + "type": "text", + "content": " row) and SRC (2" + }, + { + "bbox": [ + 46, + 228, + 547, + 253 + ], + "type": "inline_equation", + "content": "^{nd}" + }, + { + "bbox": [ + 46, + 228, + 547, + 253 + ], + "type": "text", + "content": " row)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 273, + 288, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 273, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 273, + 288, + 301 + ], + "type": "text", + "content": "4. Exhaustive geometric information with Frenet representation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 307, + 288, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 288, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 288, + 452 + ], + "type": "text", + "content": "Based on these previous observations and with the intention of developing a more suitable framework for the analysis of three-dimensional curves, Brunel and Park [3] proposed a direct extension of the SRVF method, which considers not only the tangent vector as the geometric representation of the curve but the whole Frenet-Serret frame in three dimensions. Their idea is to use an exhaustive description of the geometry of curves by incorporating higher-order information about the geometry in the representation. To exploit this idea, we propose to study suitable representations based on this Frenet-Serret framework. Proofs of theoretical results are given in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 457, + 209, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 457, + 209, + 469 + ], + "spans": [ + { + "bbox": [ + 47, + 457, + 209, + 469 + ], + "type": "text", + "content": "4.1. The Frenet-Serret framework" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 475, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 287, + 548 + ], + "type": "text", + "content": "We introduce the Frenet-Serret framework for curves of any dimension " + }, + { + "bbox": [ + 47, + 475, + 287, + 548 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 475, + 287, + 548 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 47, + 475, + 287, + 548 + ], + "type": "inline_equation", + "content": "F([0,1],\\mathbb{R}^d)" + }, + { + "bbox": [ + 47, + 475, + 287, + 548 + ], + "type": "text", + "content": " be the set of curves " + }, + { + "bbox": [ + 47, + 475, + 287, + 548 + ], + "type": "inline_equation", + "content": "x \\in AC_0([0,1],\\mathbb{R}^d)" + }, + { + "bbox": [ + 47, + 475, + 287, + 548 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 475, + 287, + 548 + ], + "type": "text", + "content": "-times continuously differentiable, and with the first " + }, + { + "bbox": [ + 47, + 475, + 287, + 548 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 475, + 287, + 548 + ], + "type": "text", + "content": " derivatives linearly independent. " + }, + { + "bbox": [ + 47, + 475, + 287, + 548 + ], + "type": "inline_equation", + "content": "F([0,1],\\mathbb{R}^d)" + }, + { + "bbox": [ + 47, + 475, + 287, + 548 + ], + "type": "text", + "content": " is called the set of Frenet curves. In the following, we will restrict the shape space to be the set" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 555, + 287, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 555, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 69, + 555, + 287, + 570 + ], + "type": "interline_equation", + "content": "\\mathcal {S} ([ 0, 1 ], \\mathbb {R} ^ {d}) = F ([ 0, 1 ], \\mathbb {R} ^ {d}) / \\operatorname {D i f f} _ {+} ([ 0, 1 ]). \\tag {10}", + "image_path": "129d6fc53fa3fef83854b4ffa9078a35b2651eb7570cd859d1e91c656175298c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 577, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 577, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 577, + 287, + 659 + ], + "type": "text", + "content": "The Frenet frame " + }, + { + "bbox": [ + 47, + 577, + 287, + 659 + ], + "type": "inline_equation", + "content": "e_1, e_2, \\ldots, e_d" + }, + { + "bbox": [ + 47, + 577, + 287, + 659 + ], + "type": "text", + "content": " associated with " + }, + { + "bbox": [ + 47, + 577, + 287, + 659 + ], + "type": "inline_equation", + "content": "X \\in S([0,1],\\mathbb{R}^d)" + }, + { + "bbox": [ + 47, + 577, + 287, + 659 + ], + "type": "text", + "content": " is uniquely defined by applying the Gram-Schmidt process to the first " + }, + { + "bbox": [ + 47, + 577, + 287, + 659 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 577, + 287, + 659 + ], + "type": "text", + "content": " derivatives of " + }, + { + "bbox": [ + 47, + 577, + 287, + 659 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 577, + 287, + 659 + ], + "type": "text", + "content": ". In dimension 3, the three vectors of the Frenet frame are known as the tangent, normal and bi-normal vector. We define the function " + }, + { + "bbox": [ + 47, + 577, + 287, + 659 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 47, + 577, + 287, + 659 + ], + "type": "text", + "content": " that maps to " + }, + { + "bbox": [ + 47, + 577, + 287, + 659 + ], + "type": "inline_equation", + "content": "s \\in [0,1]" + }, + { + "bbox": [ + 47, + 577, + 287, + 659 + ], + "type": "text", + "content": " along the curve the corresponding Frenet frame" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 97, + 668, + 287, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 668, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 97, + 668, + 287, + 683 + ], + "type": "interline_equation", + "content": "Q (s) = \\left[ e _ {1} (s) \\mid e _ {2} (s) \\mid \\dots \\mid e _ {d} (s) \\right]. \\tag {11}", + "image_path": "a3195d048014b885796101b92b51f45b585d2a0147824f596b963d8657f71285.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "The function " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " is a measurable curve from " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " to the Lie group of rotation matrices " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "SO(d)" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " called the Frenet path." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 274, + 547, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 274, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 274, + 547, + 335 + ], + "type": "text", + "content": "Theorem 1 (Frenet-Serret equation [9]). Let " + }, + { + "bbox": [ + 305, + 274, + 547, + 335 + ], + "type": "inline_equation", + "content": "X \\in S([0,1],\\mathbb{R}^d)" + }, + { + "bbox": [ + 305, + 274, + 547, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 274, + 547, + 335 + ], + "type": "inline_equation", + "content": "Q(s)" + }, + { + "bbox": [ + 305, + 274, + 547, + 335 + ], + "type": "text", + "content": " its associated Frenet path. Then there are functions " + }, + { + "bbox": [ + 305, + 274, + 547, + 335 + ], + "type": "inline_equation", + "content": "\\theta_1, \\ldots, \\theta_{d-1}" + }, + { + "bbox": [ + 305, + 274, + 547, + 335 + ], + "type": "text", + "content": " defined on that curve with " + }, + { + "bbox": [ + 305, + 274, + 547, + 335 + ], + "type": "inline_equation", + "content": "\\theta_1, \\ldots, \\theta_{d-2} > 0" + }, + { + "bbox": [ + 305, + 274, + 547, + 335 + ], + "type": "text", + "content": ", so that every " + }, + { + "bbox": [ + 305, + 274, + 547, + 335 + ], + "type": "inline_equation", + "content": "\\theta_i" + }, + { + "bbox": [ + 305, + 274, + 547, + 335 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 305, + 274, + 547, + 335 + ], + "type": "inline_equation", + "content": "(d-1-i)" + }, + { + "bbox": [ + 305, + 274, + 547, + 335 + ], + "type": "text", + "content": "-times continuously differentiable and" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 383, + 344, + 545, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 344, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 383, + 344, + 545, + 358 + ], + "type": "interline_equation", + "content": "Q ^ {\\prime} (s) = Q (s) A _ {\\boldsymbol {\\theta}} (s) \\tag {12}", + "image_path": "e879a646403f4c61520098b704ef45aba4be50013a6c440a208c46dbe3ded416.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 366, + 476, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 366, + 476, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 366, + 476, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 366, + 476, + 380 + ], + "type": "inline_equation", + "content": "\\pmb {\\theta}(s) = (\\theta_{1}(s),\\dots ,\\theta_{d - 1}(s))^{T}" + }, + { + "bbox": [ + 306, + 366, + 476, + 380 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 388, + 537, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 388, + 537, + 466 + ], + "spans": [ + { + "bbox": [ + 312, + 388, + 537, + 466 + ], + "type": "interline_equation", + "content": "A _ {\\pmb {\\theta}} (s) = \\left[ \\begin{array}{c c c c c} 0 & - \\theta_ {1} (s) & 0 & \\dots & 0 \\\\ \\theta_ {1} (s) & 0 & - \\theta_ {2} (s) & \\ddots & \\vdots \\\\ 0 & \\theta_ {2} (s) & \\ddots & \\ddots & 0 \\\\ \\vdots & \\ddots & \\ddots & 0 & - \\theta_ {d - 1} (s) \\\\ 0 & \\dots & 0 & \\theta_ {d - 1} (s) & 0 \\end{array} \\right]", + "image_path": "ad54c7fe75494a6603ba1cec1eb4d1f2fb12a91279b7e78ef973108140b01e45.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 472, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 472, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 545, + 498 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 305, + 472, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 305, + 472, + 545, + 498 + ], + "type": "text", + "content": " is called the " + }, + { + "bbox": [ + 305, + 472, + 545, + 498 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 472, + 545, + 498 + ], + "type": "text", + "content": "-th Frenet curvature, and the equation is called the Frenet-Serret equation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 506, + 545, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 648 + ], + "type": "text", + "content": "In dimension 3, the two Frenet curvatures are known as " + }, + { + "bbox": [ + 304, + 506, + 545, + 648 + ], + "type": "inline_equation", + "content": "s \\mapsto \\kappa(s)" + }, + { + "bbox": [ + 304, + 506, + 545, + 648 + ], + "type": "text", + "content": " the curvature function and " + }, + { + "bbox": [ + 304, + 506, + 545, + 648 + ], + "type": "inline_equation", + "content": "s \\mapsto \\tau(s)" + }, + { + "bbox": [ + 304, + 506, + 545, + 648 + ], + "type": "text", + "content": " the torsion function. They have an interpretable physical meaning. The curvature function measures how sharply the curve changes direction at a given point, and the torsion function measures the degree to which the curve twists and turns as it moves along its path. The Frenet-Serret equation with an initial condition " + }, + { + "bbox": [ + 304, + 506, + 545, + 648 + ], + "type": "inline_equation", + "content": "Q(0) = Q_0" + }, + { + "bbox": [ + 304, + 506, + 545, + 648 + ], + "type": "text", + "content": " defines an ordinary differential equation on the Lie group " + }, + { + "bbox": [ + 304, + 506, + 545, + 648 + ], + "type": "inline_equation", + "content": "SO(d)" + }, + { + "bbox": [ + 304, + 506, + 545, + 648 + ], + "type": "text", + "content": " where the function " + }, + { + "bbox": [ + 304, + 506, + 545, + 648 + ], + "type": "inline_equation", + "content": "s \\mapsto A_\\theta(s)" + }, + { + "bbox": [ + 304, + 506, + 545, + 648 + ], + "type": "text", + "content": " has values in the Lie algebra of skew-symmetric matrices. This equation can also be expressed in function of the time variable " + }, + { + "bbox": [ + 304, + 506, + 545, + 648 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 506, + 545, + 648 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 354, + 656, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 656, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 354, + 656, + 545, + 681 + ], + "type": "interline_equation", + "content": "\\frac {d Q (s (t))}{d t} = \\dot {s} (t) Q (s (t)) A _ {\\boldsymbol {\\theta}} (s (t)). \\tag {13}", + "image_path": "099005030f8d9192923d90e921bda7657e7ae9a88376ac6a6e7efcae4be50be3.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "type": "text", + "content": "Lemma 1. The Frenet curvatures and the Frenet path are invariant under all Euclidean motions." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4030" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 122 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 122 + ], + "type": "text", + "content": "This means that for " + }, + { + "bbox": [ + 47, + 72, + 287, + 122 + ], + "type": "inline_equation", + "content": "x \\in F([0,1],\\mathbb{R}^d), O \\in SO(d), a \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 47, + 72, + 287, + 122 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 122 + ], + "type": "inline_equation", + "content": "h \\in \\mathrm{Diff}_{+}([0,1])" + }, + { + "bbox": [ + 47, + 72, + 287, + 122 + ], + "type": "text", + "content": ", the curve defined by " + }, + { + "bbox": [ + 47, + 72, + 287, + 122 + ], + "type": "inline_equation", + "content": "\\tilde{x}(t) = a + Ox(h(t))" + }, + { + "bbox": [ + 47, + 72, + 287, + 122 + ], + "type": "text", + "content": " has the same Frenet curvatures as " + }, + { + "bbox": [ + 47, + 72, + 287, + 122 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 72, + 287, + 122 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 122 + ], + "type": "inline_equation", + "content": "\\tilde{Q}(s) = OQ(s)" + }, + { + "bbox": [ + 47, + 72, + 287, + 122 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 127, + 288, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 127, + 288, + 187 + ], + "spans": [ + { + "bbox": [ + 47, + 127, + 288, + 187 + ], + "type": "text", + "content": "Theorem 2 (Fundamental theorem of the local theory of curves [9]). Let " + }, + { + "bbox": [ + 47, + 127, + 288, + 187 + ], + "type": "inline_equation", + "content": "\\theta_1, \\ldots, \\theta_{d-1} \\in C^\\infty([0,1], \\mathbb{R})" + }, + { + "bbox": [ + 47, + 127, + 288, + 187 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 47, + 127, + 288, + 187 + ], + "type": "inline_equation", + "content": "\\theta_1, \\ldots, \\theta_{d-2} > 0" + }, + { + "bbox": [ + 47, + 127, + 288, + 187 + ], + "type": "text", + "content": ". For a given " + }, + { + "bbox": [ + 47, + 127, + 288, + 187 + ], + "type": "inline_equation", + "content": "X_0 \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 47, + 127, + 288, + 187 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 127, + 288, + 187 + ], + "type": "inline_equation", + "content": "Q_0 \\in SO(d)" + }, + { + "bbox": [ + 47, + 127, + 288, + 187 + ], + "type": "text", + "content": " there is a unique " + }, + { + "bbox": [ + 47, + 127, + 288, + 187 + ], + "type": "inline_equation", + "content": "X \\in S([0,1], \\mathbb{R}^d)" + }, + { + "bbox": [ + 47, + 127, + 288, + 187 + ], + "type": "text", + "content": " parametrized by arc length and satisfying the following three conditions:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 194, + 242, + 244 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 58, + 194, + 120, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 194, + 120, + 205 + ], + "spans": [ + { + "bbox": [ + 58, + 194, + 120, + 205 + ], + "type": "inline_equation", + "content": "X(0) = X_0" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 213, + 242, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 213, + 242, + 225 + ], + "spans": [ + { + "bbox": [ + 58, + 213, + 242, + 225 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 58, + 213, + 242, + 225 + ], + "type": "inline_equation", + "content": "Q_{0}" + }, + { + "bbox": [ + 58, + 213, + 242, + 225 + ], + "type": "text", + "content": " is the Frenet frame of " + }, + { + "bbox": [ + 58, + 213, + 242, + 225 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 58, + 213, + 242, + 225 + ], + "type": "text", + "content": " at point " + }, + { + "bbox": [ + 58, + 213, + 242, + 225 + ], + "type": "inline_equation", + "content": "s = 0" + }, + { + "bbox": [ + 58, + 213, + 242, + 225 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 232, + 242, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 232, + 242, + 244 + ], + "spans": [ + { + "bbox": [ + 58, + 232, + 242, + 244 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 58, + 232, + 242, + 244 + ], + "type": "inline_equation", + "content": "\\theta_{1},\\ldots ,\\theta_{d - 1}" + }, + { + "bbox": [ + 58, + 232, + 242, + 244 + ], + "type": "text", + "content": " are the Frenet curvatures of " + }, + { + "bbox": [ + 58, + 232, + 242, + 244 + ], + "type": "inline_equation", + "content": "X" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 258, + 287, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 287, + 307 + ], + "type": "text", + "content": "Theorems 1 and 2 state that there is a bijection, up to a translation and a rotation, between the shape space " + }, + { + "bbox": [ + 47, + 258, + 287, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{S}([0,1],\\mathbb{R}^d)" + }, + { + "bbox": [ + 47, + 258, + 287, + 307 + ], + "type": "text", + "content": ", the set of admissible Frenet curvatures " + }, + { + "bbox": [ + 47, + 258, + 287, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 47, + 258, + 287, + 307 + ], + "type": "text", + "content": " and the set of corresponding Frenet paths " + }, + { + "bbox": [ + 47, + 258, + 287, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_0" + }, + { + "bbox": [ + 47, + 258, + 287, + 307 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 313, + 287, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 313, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 57, + 313, + 287, + 328 + ], + "type": "interline_equation", + "content": "\\mathcal {H} = \\left\\{\\boldsymbol {\\theta} \\in C ^ {\\infty} ([ 0, 1 ], \\mathbb {R} ^ {d - 1}) | \\theta_ {1}, \\dots , \\theta_ {d - 2} > 0 \\right\\}, \\tag {14}", + "image_path": "a2b41d03b9a3a7ed26b8fe549f3a06c7f892ad427d569ec2608297bbe6ab2c16.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 351, + 287, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 351, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 57, + 351, + 287, + 388 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {0} = \\left\\{ \\begin{array}{c} Q \\in \\mathbb {L} ^ {2} ([ 0, 1 ], S O (d)) \\text {s u c h t h a t} \\\\ Q ^ {\\prime} (s) = Q (s) A _ {\\boldsymbol {\\theta}} (s), Q (0) = I _ {d}, \\boldsymbol {\\theta} \\in \\mathcal {H} \\end{array} \\right\\}. \\tag {15}", + "image_path": "fdaa9c553a5312c3d9adc9d923e499b526ac873b1261fbc4756d72e8f99bdd74.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 394, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 287, + 453 + ], + "type": "text", + "content": "From the detailed Frenet-Serret framework, one can think of the direct extension of the square-root velocity function 4 that simply consists in replacing the tangent vector with the entire Frenet frame. The representation of a parametrized curve " + }, + { + "bbox": [ + 47, + 394, + 287, + 453 + ], + "type": "inline_equation", + "content": "x \\in F([0,1],\\mathbb{R}^d)" + }, + { + "bbox": [ + 47, + 394, + 287, + 453 + ], + "type": "text", + "content": " will be then" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 108, + 462, + 287, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 462, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 108, + 462, + 287, + 476 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {Q} (x) (t) = \\sqrt {\\dot {s} (t)} Q (s (t)). \\tag {16}", + "image_path": "e66e17ea1fdbb645b1453f255f6ed7ddf8fb19a77a918d7379c9f0ed199a967c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 484, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 484, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 287, + 616 + ], + "type": "text", + "content": "This representation is used in [3] to define a new alignment method on " + }, + { + "bbox": [ + 47, + 484, + 287, + 616 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 484, + 287, + 616 + ], + "type": "text", + "content": ". They extend the SRVF registration problem 9 by using the Frobenius distance between the Frenet frames instead of only the " + }, + { + "bbox": [ + 47, + 484, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^2" + }, + { + "bbox": [ + 47, + 484, + 287, + 616 + ], + "type": "text", + "content": " distance of the unit tangent vectors. They show to obtain more precise results with their method than the SRVF one. From the previous theorems, it is clear that this representation uniquely defines a parametrized curve " + }, + { + "bbox": [ + 47, + 484, + 287, + 616 + ], + "type": "inline_equation", + "content": "x \\in F([0,1],\\mathbb{R}^d)" + }, + { + "bbox": [ + 47, + 484, + 287, + 616 + ], + "type": "text", + "content": ". However, as a result of the Frobenius theorem [12], the set of such representations " + }, + { + "bbox": [ + 47, + 484, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_Q" + }, + { + "bbox": [ + 47, + 484, + 287, + 616 + ], + "type": "text", + "content": " appears not to be a manifold. We leave the demonstration in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 622, + 234, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 234, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 234, + 635 + ], + "type": "text", + "content": "4.2. Unparametrized Frenet curvatures" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 641, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 287, + 689 + ], + "type": "text", + "content": "A possible representation of a parametrized curve, already suggested in [23, 25, 20], which keeps the idea of encoding more geometric information, is the unparametrized Frenet curvatures and the arc-length function pair" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 100, + 696, + 287, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 696, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 100, + 696, + 287, + 716 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {\\boldsymbol {\\theta}} (x) (t) = \\left(\\sqrt {\\dot {s} (t)}, \\boldsymbol {\\theta} (s (t))\\right). \\tag {17}", + "image_path": "81a25ef5f02a31e577e95a690aa04aaa533da2a23ad285fae0647875ce242d46.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": "We denote " + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\Psi([0,1])" + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": ", the set of square root velocity functions of length-normalized arc-length functions. This set is well-studied in the literature [15, 27]. It is the unit sphere of the Hilbert space " + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^2([0,1],\\mathbb{R})" + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": " and therefore a Riemannian manifold equipped with the " + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^2" + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": " metric. Then, the geodesic distance between two elements in " + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\Psi([0,1])" + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 328, + 152, + 545, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 152, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 328, + 152, + 545, + 172 + ], + "type": "interline_equation", + "content": "d _ {\\Psi} \\left(\\sqrt {\\dot {s} _ {0}}, \\sqrt {\\dot {s} _ {1}}\\right) = \\cos^ {- 1} \\left(\\left\\langle \\sqrt {\\dot {s} _ {0}}, \\sqrt {\\dot {s} _ {1}} \\right\\rangle\\right) \\tag {18}", + "image_path": "4f25db163561873f482b6aeb587b6b52efa87f9720f3f6216f711e55d5ef4b91.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 178, + 509, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 178, + 509, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 509, + 190 + ], + "type": "text", + "content": "and the geodesic path connecting them is given by" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 324, + 198, + 545, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 198, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 324, + 198, + 545, + 224 + ], + "type": "interline_equation", + "content": "\\alpha_ {\\Psi} (\\tau) = \\frac {\\sin ((1 - \\tau) \\vartheta)}{\\sin (\\vartheta)} \\sqrt {\\dot {s} _ {0}} + \\frac {\\sin (\\tau \\vartheta)}{\\sin (\\vartheta)} \\sqrt {\\dot {s} _ {1}} \\tag {19}", + "image_path": "53e8d740c8c23bfa3d20598be1134d1bad27b9d9533dd34354ebb02a3e41f484.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 232, + 545, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 232, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 545, + 268 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 232, + 545, + 268 + ], + "type": "inline_equation", + "content": "\\vartheta = d_{\\Psi}(\\sqrt{\\dot{s}_0},\\sqrt{\\dot{s}_1})" + }, + { + "bbox": [ + 306, + 232, + 545, + 268 + ], + "type": "text", + "content": " Moreover, any element of " + }, + { + "bbox": [ + 306, + 232, + 545, + 268 + ], + "type": "inline_equation", + "content": "\\mathrm{Diff}_{+}([0,1])" + }, + { + "bbox": [ + 306, + 232, + 545, + 268 + ], + "type": "text", + "content": " is uniquely represented by an element of " + }, + { + "bbox": [ + 306, + 232, + 545, + 268 + ], + "type": "inline_equation", + "content": "\\Psi ([0,1])" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 275, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 275, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 275, + 545, + 300 + ], + "type": "text", + "content": "Proposition 1. The set of Frenet curvatures " + }, + { + "bbox": [ + 305, + 275, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 305, + 275, + 545, + 300 + ], + "type": "text", + "content": " is a Riemannian submanifold of " + }, + { + "bbox": [ + 305, + 275, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^2([0,1],\\mathbb{R}^{d-1})" + }, + { + "bbox": [ + 305, + 275, + 545, + 300 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "text", + "content": "Proof. The set " + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "inline_equation", + "content": "M = \\{x \\in \\mathbb{R}^{d-1} | x_1, \\ldots, x_{d-2} > 0\\}" + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "text", + "content": " is a open subset of the Riemannian manifold " + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{d-1}" + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "text", + "content": ". Then it is itself a differentiable Riemannian manifold with the standard inner product of " + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{d-1}" + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "text", + "content": ", and for any point " + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "inline_equation", + "content": "p \\in M" + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "text", + "content": " the tangent space " + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "inline_equation", + "content": "T_p(M)" + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{d-1}" + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "text", + "content": ". The set of Frenet curvatures " + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "text", + "content": " is the set of measurable curves from [0, 1] to the Riemannian manifold " + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "text", + "content": " and thus also a manifold ([26]). Its tangent space is " + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^2([0,1], \\mathbb{R}^{d-1})" + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "text", + "content": " and it can be equipped with the " + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^2" + }, + { + "bbox": [ + 305, + 304, + 545, + 413 + ], + "type": "text", + "content": " Riemannian metric." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 420, + 545, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 420, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 306, + 420, + 545, + 444 + ], + "type": "text", + "content": "Consequently, the geodesic distance on " + }, + { + "bbox": [ + 306, + 420, + 545, + 444 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 306, + 420, + 545, + 444 + ], + "type": "text", + "content": " is simply the " + }, + { + "bbox": [ + 306, + 420, + 545, + 444 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^2" + }, + { + "bbox": [ + 306, + 420, + 545, + 444 + ], + "type": "text", + "content": " norm" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 368, + 453, + 545, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 453, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 368, + 453, + 545, + 466 + ], + "type": "interline_equation", + "content": "d _ {\\mathcal {H}} \\left(\\boldsymbol {\\theta} _ {0}, \\boldsymbol {\\theta} _ {1}\\right) = \\left\\| \\boldsymbol {\\theta} _ {0} - \\boldsymbol {\\theta} _ {1} \\right\\| _ {\\mathbb {L} ^ {2}} \\tag {20}", + "image_path": "564182fa0cc85cd0c6d58082b73943308efce0a2dbff1abfbc781592ab0a8da5.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 474, + 537, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 474, + 537, + 486 + ], + "spans": [ + { + "bbox": [ + 306, + 474, + 537, + 486 + ], + "type": "text", + "content": "and the geodesic path is the straight line connecting them" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 367, + 495, + 545, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 495, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 367, + 495, + 545, + 507 + ], + "type": "interline_equation", + "content": "\\alpha_ {\\mathcal {H}} (\\tau) = (1 - \\tau) \\boldsymbol {\\theta} _ {0} + \\tau \\boldsymbol {\\theta} _ {1}. \\tag {21}", + "image_path": "b9279a3bbede4799ee8550cadeaeadc467000cb13294ffe863954df005356dba.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 514, + 544, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 544, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 544, + 539 + ], + "type": "text", + "content": "Proposition 2. The map " + }, + { + "bbox": [ + 306, + 514, + 544, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\theta}: F([0,1], \\mathbb{R}^d) \\to \\Psi([0,1]) \\times \\mathcal{H}" + }, + { + "bbox": [ + 306, + 514, + 544, + 539 + ], + "type": "text", + "content": ", defined above, is a bijection." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 545, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 545, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 305, + 545, + 545, + 619 + ], + "type": "text", + "content": "Proof. The element of " + }, + { + "bbox": [ + 305, + 545, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\Psi([0,1])" + }, + { + "bbox": [ + 305, + 545, + 545, + 619 + ], + "type": "text", + "content": " uniquely defines the arclength function by " + }, + { + "bbox": [ + 305, + 545, + 545, + 619 + ], + "type": "inline_equation", + "content": "s(t) = \\int_0^t (\\sqrt{\\dot{s}(u)})^2 du" + }, + { + "bbox": [ + 305, + 545, + 545, + 619 + ], + "type": "text", + "content": ". As mentioned before, we have a bijection between the unparametrized Frenet curvatures in " + }, + { + "bbox": [ + 305, + 545, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 305, + 545, + 545, + 619 + ], + "type": "text", + "content": " and the unparametrized curve in the shape space " + }, + { + "bbox": [ + 305, + 545, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 305, + 545, + 545, + 619 + ], + "type": "text", + "content": " (1, 2). Then, from " + }, + { + "bbox": [ + 305, + 545, + 545, + 619 + ], + "type": "inline_equation", + "content": "X \\in \\mathcal{S}" + }, + { + "bbox": [ + 305, + 545, + 545, + 619 + ], + "type": "text", + "content": ", the initial parametrized curve is simply " + }, + { + "bbox": [ + 305, + 545, + 545, + 619 + ], + "type": "inline_equation", + "content": "x(t) = X(s(t))" + }, + { + "bbox": [ + 305, + 545, + 545, + 619 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 626, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 626, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 305, + 626, + 545, + 685 + ], + "type": "text", + "content": "The set of such " + }, + { + "bbox": [ + 305, + 626, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\theta}" + }, + { + "bbox": [ + 305, + 626, + 545, + 685 + ], + "type": "text", + "content": " is the Cartesian product of " + }, + { + "bbox": [ + 305, + 626, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\Psi ([0,1])" + }, + { + "bbox": [ + 305, + 626, + 545, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 626, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 305, + 626, + 545, + 685 + ], + "type": "text", + "content": " and, therefore, is also a Riemannian manifold equipped with the product metric " + }, + { + "bbox": [ + 305, + 626, + 545, + 685 + ], + "type": "inline_equation", + "content": "d_{\\Psi}\\oplus d_{\\mathcal{H}}" + }, + { + "bbox": [ + 305, + 626, + 545, + 685 + ], + "type": "text", + "content": " [16]. The induced metric on " + }, + { + "bbox": [ + 305, + 626, + 545, + 685 + ], + "type": "inline_equation", + "content": "F([0,1],\\mathbb{R}^d)" + }, + { + "bbox": [ + 305, + 626, + 545, + 685 + ], + "type": "text", + "content": " under the representation " + }, + { + "bbox": [ + 305, + 626, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\theta}" + }, + { + "bbox": [ + 305, + 626, + 545, + 685 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 322, + 691, + 545, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 691, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 322, + 691, + 545, + 711 + ], + "type": "interline_equation", + "content": "d _ {\\boldsymbol {\\theta}} \\left(x _ {0}, x _ {1}\\right) = d _ {\\Psi} \\left(\\sqrt {\\dot {s} _ {0}}, \\sqrt {\\dot {s} _ {1}}\\right) + d _ {\\mathcal {H}} \\left(\\boldsymbol {\\theta} _ {0}, \\boldsymbol {\\theta} _ {1}\\right). \\tag {22}", + "image_path": "23d3e6bbc7bbe484f85f27f2357b2f55e397637c3fcd4cdbc49d17998b208f6a.jpg" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4031" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": "In order to define a distance on the shape space " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " from that one, we must quotient out the space " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\mathrm{Diff}_{+}([0,1])" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": ". The action of " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\mathrm{Diff}_{+}([0,1])" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\Psi ([0,1])" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " is the same as 5, and the Frenet curvatures are invariant under reparametrization of the corresponding parametrized curve (Lemma 1). Moreover, by simply taking " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "h^* = s_1^{-1} \\circ s_0 \\in \\mathrm{Diff}_{+}([0,1])" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\sqrt{s_1} * h^* = \\sqrt{s_0}" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": ", and thus the distance on " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\Psi ([0,1]) / \\mathrm{Diff}_{+}([0,1])" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "s_0" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "s_1" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " is zero. Hence, the induced distance on the shape space, between " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "X_0, X_1 \\in S" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": ", under the representation " + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\theta}" + }, + { + "bbox": [ + 47, + 72, + 289, + 193 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 201, + 287, + 219 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 201, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 66, + 201, + 287, + 219 + ], + "type": "interline_equation", + "content": "d _ {\\mathcal {S}} ^ {(\\boldsymbol {\\theta})} \\left(X _ {0}, X _ {1}\\right) := d _ {\\mathcal {H}} \\left(\\boldsymbol {\\theta} _ {0}, \\boldsymbol {\\theta} _ {1}\\right) = \\left\\| \\boldsymbol {\\theta} _ {0} - \\boldsymbol {\\theta} _ {1} \\right\\| _ {\\mathbb {L} ^ {2}} \\tag {23}", + "image_path": "bc8d80758aae931e994c6b19a9d46f1ec317de458e4f8660929a7a5991751965.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 228, + 216, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 216, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 216, + 240 + ], + "type": "text", + "content": "and the geodesic path connecting them is" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 249, + 287, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 249, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 108, + 249, + 287, + 270 + ], + "type": "interline_equation", + "content": "\\alpha_ {\\mathcal {S}} ^ {(\\boldsymbol {\\theta})} (\\tau) = \\left(\\sqrt {\\dot {s} _ {0}}, \\alpha_ {\\mathcal {H}} (\\tau)\\right). \\tag {24}", + "image_path": "1e4b355888446df3b347ca7369ceced06c49451c2e2ff2eb2e52fa5fc09edbfd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 278, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 278, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 287, + 339 + ], + "type": "text", + "content": "This immediate representation by the Frenet curvatures appears in the experiments not to be sufficiently elastic (Figure 5). It has somewhat the same weakness as the angle representation proposed in [23], the Frenet curvatures being already independent of the parametrization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 347, + 237, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 347, + 237, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 347, + 237, + 360 + ], + "type": "text", + "content": "4.3. Square Root Curvatures Transform" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 366, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 366, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 47, + 366, + 287, + 450 + ], + "type": "text", + "content": "To overcome the \"non-elasticity\" issue of the representation defined above, we propose a second framework for shape analysis based on Frenet curvatures which uses, like the square root velocity function, the parametrization as a tool to register the curves and define a more \"elastic\" method. The latter is inspired by the square-root velocity transform of " + }, + { + "bbox": [ + 47, + 366, + 287, + 450 + ], + "type": "inline_equation", + "content": "SO(d)" + }, + { + "bbox": [ + 47, + 366, + 287, + 450 + ], + "type": "text", + "content": "-valued curves." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 459, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 496 + ], + "type": "text", + "content": "Definition 1 (SRV Transform for curves on " + }, + { + "bbox": [ + 46, + 459, + 287, + 496 + ], + "type": "inline_equation", + "content": "SO(d)" + }, + { + "bbox": [ + 46, + 459, + 287, + 496 + ], + "type": "text", + "content": "). Let " + }, + { + "bbox": [ + 46, + 459, + 287, + 496 + ], + "type": "inline_equation", + "content": "P \\in C^{\\infty}([0,1], SO(d))" + }, + { + "bbox": [ + 46, + 459, + 287, + 496 + ], + "type": "text", + "content": ". The Square Root Velocity transform of " + }, + { + "bbox": [ + 46, + 459, + 287, + 496 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 459, + 287, + 496 + ], + "type": "text", + "content": " is the map" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 82, + 505, + 287, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 505, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 82, + 505, + 287, + 541 + ], + "type": "interline_equation", + "content": "q (P) (t) = \\frac {L _ {P (t) ^ {- 1}} \\dot {P} (t)}{\\sqrt {\\| \\dot {P} (t) \\| _ {F}}} = \\frac {P (t) ^ {T} \\dot {P} (t)}{\\sqrt {\\| \\dot {P} (t) \\| _ {F}}}, \\tag {25}", + "image_path": "0cf086e7979f1e33a04e8d883190333b5bafe4af449acbc2cac3469601acf9eb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 551, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 287, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 551, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\| .\\| _F" + }, + { + "bbox": [ + 47, + 551, + 287, + 590 + ], + "type": "text", + "content": " is the Frobenius norm associated with the scalar product on the Lie Algebra of skew-symmetric matrices " + }, + { + "bbox": [ + 47, + 551, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\langle A,B\\rangle = \\frac{1}{2} tr(A^T B) = -\\frac{1}{2} tr(AB)" + }, + { + "bbox": [ + 47, + 551, + 287, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 596, + 287, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 287, + 633 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 47, + 596, + 287, + 633 + ], + "type": "inline_equation", + "content": "x \\in F([0,1],\\mathbb{R}^d)" + }, + { + "bbox": [ + 47, + 596, + 287, + 633 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 596, + 287, + 633 + ], + "type": "inline_equation", + "content": "Q(t) \\in C^\\infty([0,1],SO(d))" + }, + { + "bbox": [ + 47, + 596, + 287, + 633 + ], + "type": "text", + "content": " be its associated Frenet path. Using the Frenet-Serret differential equation, the SRV Transform of the Frenet path is" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 94, + 643, + 287, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 643, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 94, + 643, + 287, + 671 + ], + "type": "interline_equation", + "content": "q (Q) (t) = \\sqrt {\\dot {s} (t)} \\frac {A _ {\\boldsymbol {\\theta}} (s (t))}{\\sqrt {\\left\\| A _ {\\boldsymbol {\\theta}} (s (t)) \\right\\| _ {F}}}. \\tag {26}", + "image_path": "223af8b3088c9f68b0063f941b5d65bcaa17d0094433ffadf1b3f5ad65851591.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 678, + 194, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 678, + 194, + 690 + ], + "spans": [ + { + "bbox": [ + 47, + 678, + 194, + 690 + ], + "type": "text", + "content": "Proposition 3. Let " + }, + { + "bbox": [ + 47, + 678, + 194, + 690 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathcal{H}" + }, + { + "bbox": [ + 47, + 678, + 194, + 690 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 108, + 700, + 225, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 700, + 225, + 715 + ], + "spans": [ + { + "bbox": [ + 108, + 700, + 225, + 715 + ], + "type": "interline_equation", + "content": "\\left\\| A _ {\\boldsymbol {\\theta}} (s (t)) \\right\\| _ {F} = \\left\\| \\boldsymbol {\\theta} (s (t)) \\right\\| _ {2}.", + "image_path": "e9d2895a5663249e6823cc19d7d44bbda13168068546b8482e468e0222803977.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "Based on the SRV Transform of a Frenet path and Proposition 3, we propose a new transformation of a parametrized curve, which we have called the Square-Root Curvatures (SRC) transform." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "spans": [ + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "text", + "content": "Definition 2 (Square-Root Curvatures Transform). Let " + }, + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "inline_equation", + "content": "x \\in F([0,1],\\mathbb{R}^d)" + }, + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "text", + "content": ". We consider its associated arc-length function " + }, + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "inline_equation", + "content": "s(t)" + }, + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "text", + "content": " and Frenet curvatures " + }, + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}(s(t))" + }, + { + "bbox": [ + 305, + 127, + 547, + 187 + ], + "type": "text", + "content": " defined as in Theorem 1. Then we define its square-root curvatures transform to be the map" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 367, + 194, + 545, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 194, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 367, + 194, + 545, + 223 + ], + "type": "interline_equation", + "content": "c (t) = \\sqrt {\\dot {s} (t)} \\frac {\\boldsymbol {\\theta} (s (t))}{\\sqrt {\\| \\boldsymbol {\\theta} (s (t)) \\|}}. \\tag {27}", + "image_path": "e45dbf1aa4092b9dac5fc3f0590b2e5408ff9ef70c9718c3cfb603dfe3e512f6.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 229, + 528, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 229, + 528, + 241 + ], + "spans": [ + { + "bbox": [ + 318, + 229, + 528, + 241 + ], + "type": "text", + "content": "The set of such square-root curvatures transforms is" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 249, + 545, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 249, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 321, + 249, + 545, + 264 + ], + "type": "interline_equation", + "content": "\\mathcal {C} = \\left\\{c \\in \\mathbb {L} ([ 0, 1 ], \\mathbb {R} ^ {d - 1}) \\mid c _ {1}, \\dots , c _ {d - 2} > 0 \\right\\}, \\tag {28}", + "image_path": "04bf72191fcf705d485dfdd57de8953ce01c73df98cc98d75f85e39e65c3df31.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 271, + 545, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 271, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 305, + 271, + 545, + 367 + ], + "type": "text", + "content": "which is the same as the set of admissible Frenet curvatures " + }, + { + "bbox": [ + 305, + 271, + 545, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 305, + 271, + 545, + 367 + ], + "type": "text", + "content": ". We have already shown in the previous section that this set is a Riemannian manifold equipped with the " + }, + { + "bbox": [ + 305, + 271, + 545, + 367 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^2" + }, + { + "bbox": [ + 305, + 271, + 545, + 367 + ], + "type": "text", + "content": " metric. Therefore, the geodesic distance between " + }, + { + "bbox": [ + 305, + 271, + 545, + 367 + ], + "type": "inline_equation", + "content": "c_{0}, c_{1} \\in \\mathcal{C}" + }, + { + "bbox": [ + 305, + 271, + 545, + 367 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 305, + 271, + 545, + 367 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^2" + }, + { + "bbox": [ + 305, + 271, + 545, + 367 + ], + "type": "text", + "content": " distance between them, and the geodesic path is a straight line. We define the following representation of a parametrized curve " + }, + { + "bbox": [ + 305, + 271, + 545, + 367 + ], + "type": "inline_equation", + "content": "x \\in F([0,1], \\mathbb{R}^d)" + }, + { + "bbox": [ + 305, + 271, + 545, + 367 + ], + "type": "text", + "content": ", from its Square-Root Curvatures transform, by" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 361, + 374, + 545, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 374, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 361, + 374, + 545, + 396 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {\\mathrm {S R C}} (x) (t) = \\left(\\sqrt {\\dot {s} (t)}, c (t)\\right). \\tag {29}", + "image_path": "c6c23d92ab6b134531ca6af1b54dbc1be5f5aba8b0461e7f7dad43b8f72e09ae.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 403, + 545, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 403, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 306, + 403, + 545, + 428 + ], + "type": "text", + "content": "Proposition 4. The map " + }, + { + "bbox": [ + 306, + 403, + 545, + 428 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{SRC} : F([0,1],\\mathbb{R}^d) \\to \\Psi([0,1]) \\times \\mathcal{C}" + }, + { + "bbox": [ + 306, + 403, + 545, + 428 + ], + "type": "text", + "content": ", defined above, is a bijection." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 434, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 434, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 434, + 545, + 533 + ], + "type": "text", + "content": "Proof. This is again a result of theorems 1 and 2. To get " + }, + { + "bbox": [ + 305, + 434, + 545, + 533 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 434, + 545, + 533 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 305, + 434, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{SRC}}(x)" + }, + { + "bbox": [ + 305, + 434, + 545, + 533 + ], + "type": "text", + "content": ", it should be noted firstly that " + }, + { + "bbox": [ + 305, + 434, + 545, + 533 + ], + "type": "inline_equation", + "content": "c(t) \\| c(t) \\| = \\dot{s}(t) \\theta(s(t))" + }, + { + "bbox": [ + 305, + 434, + 545, + 533 + ], + "type": "text", + "content": ". From that, the skew-symmetric matrix function of the Frenet-Serret ODE can be reconstructed. By solving the corresponding Frenet-Serret ODE one gets the associated time parametrized Frenet path " + }, + { + "bbox": [ + 305, + 434, + 545, + 533 + ], + "type": "inline_equation", + "content": "Q(t)" + }, + { + "bbox": [ + 305, + 434, + 545, + 533 + ], + "type": "text", + "content": ". Then, using the first component of " + }, + { + "bbox": [ + 305, + 434, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{SRC}}(x)" + }, + { + "bbox": [ + 305, + 434, + 545, + 533 + ], + "type": "text", + "content": ", we get " + }, + { + "bbox": [ + 305, + 434, + 545, + 533 + ], + "type": "inline_equation", + "content": "x(t) = X(s(t)) = \\int_0^t \\dot{s}(u) T(s(u)) du" + }, + { + "bbox": [ + 305, + 434, + 545, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 540, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 540, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 305, + 540, + 545, + 624 + ], + "type": "text", + "content": "The set of such square root curvature representations " + }, + { + "bbox": [ + 305, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{SRC}}" + }, + { + "bbox": [ + 305, + 540, + 545, + 624 + ], + "type": "text", + "content": " is the Cartesian product " + }, + { + "bbox": [ + 305, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\Psi ([0,1])\\times \\mathcal{C}" + }, + { + "bbox": [ + 305, + 540, + 545, + 624 + ], + "type": "text", + "content": " and therefore a Riemannian manifold with the product metric " + }, + { + "bbox": [ + 305, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "d_{\\Psi}\\oplus d_{\\mathcal{C}}" + }, + { + "bbox": [ + 305, + 540, + 545, + 624 + ], + "type": "text", + "content": ". This representation is, by definition, invariant under the action of " + }, + { + "bbox": [ + 305, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "SO(d)" + }, + { + "bbox": [ + 305, + 540, + 545, + 624 + ], + "type": "text", + "content": ". Then, the corresponding shape space is the quotient space " + }, + { + "bbox": [ + 305, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\Psi ([0,1])\\times \\mathcal{C} / \\mathrm{Diff}([0,1])" + }, + { + "bbox": [ + 305, + 540, + 545, + 624 + ], + "type": "text", + "content": ". Let's " + }, + { + "bbox": [ + 305, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "x\\in F([0,1],\\mathbb{R}^d)" + }, + { + "bbox": [ + 305, + 540, + 545, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "h\\in \\mathrm{Diff}([0,1])" + }, + { + "bbox": [ + 305, + 540, + 545, + 624 + ], + "type": "text", + "content": ". The SRC representation of " + }, + { + "bbox": [ + 305, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\tilde{x} = x\\circ h" + }, + { + "bbox": [ + 305, + 540, + 545, + 624 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 326, + 631, + 545, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 631, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 326, + 631, + 545, + 651 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {\\mathrm {S R C}} (\\tilde {x}) = \\left(\\sqrt {\\dot {s}} * h, c * h\\right) = \\mathcal {R} _ {\\mathrm {S R C}} (x) * h \\tag {30}", + "image_path": "e1c9c63aaa703a9ee12fa01b21817777da7f6da4e41c3e5573577c09f8815cb8.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 658, + 470, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 470, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 470, + 670 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 658, + 470, + 670 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 306, + 658, + 470, + 670 + ], + "type": "text", + "content": " is the group action defined in 5." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "text", + "content": "Proposition 5. The metric on " + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "inline_equation", + "content": "F([0,1],\\mathbb{R}^d)" + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "text", + "content": " induced by the Riemannian metric on " + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\Psi ([0,1])\\times \\mathcal{C}" + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "text", + "content": " defined by " + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "inline_equation", + "content": "d_{SRC}\\coloneqq d_{\\Psi}\\oplus d_{\\mathcal{C}}" + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "text", + "content": " is invariant under the action of " + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\text{Diff}_{+}([0,1])" + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4032" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "content": "The distance on the shape space " + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "content": " under the representation " + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{SRC}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "content": ", between two elements " + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "X_0, X_1 \\in \\mathcal{S}" + }, + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "content": ", is defined as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 115, + 287, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 115, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 58, + 115, + 287, + 137 + ], + "type": "interline_equation", + "content": "d _ {S} ^ {\\left(\\mathrm {S R C}\\right)} \\left(X _ {0}, X _ {1}\\right) := \\inf _ {h \\in \\operatorname {D i f f} _ {+} ([ 0, 1 ])} d _ {\\mathrm {S R C}} \\left(x _ {0}, x _ {1} \\circ h\\right). \\tag {31}", + "image_path": "3a4e3cdd454c2c0641d58f2208e6d720f451e94ef119d5fc564eda8357f7ba90.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 147, + 287, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 147, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 47, + 147, + 287, + 170 + ], + "type": "text", + "content": "From the optimal wrapping function " + }, + { + "bbox": [ + 47, + 147, + 287, + 170 + ], + "type": "inline_equation", + "content": "h^*" + }, + { + "bbox": [ + 47, + 147, + 287, + 170 + ], + "type": "text", + "content": " the geodesic path on " + }, + { + "bbox": [ + 47, + 147, + 287, + 170 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 147, + 287, + 170 + ], + "type": "text", + "content": " between them is" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 177, + 287, + 219 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 177, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 287, + 219 + ], + "type": "interline_equation", + "content": "\\alpha_ {\\mathcal {S}} ^ {(\\mathrm {S R C})} (\\tau) = \\left( \\begin{array}{c} \\frac {\\sin ((1 - \\tau) \\vartheta)}{\\sin (\\vartheta)} \\sqrt {\\dot {s} _ {0}} + \\frac {\\sin (\\tau \\vartheta)}{\\sin (\\vartheta)} \\left(\\sqrt {\\dot {s} _ {1}} * h ^ {*}\\right), \\\\ (1 - \\tau) c _ {0} + \\tau \\left(c _ {1} * h ^ {*}\\right)) \\end{array} \\right. \\tag {32}", + "image_path": "992699129e299817cf0f577ada613ea581ead4900b176ddb3d023b4ec9fcfef8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 219, + 287, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 219, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 219, + 287, + 255 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 219, + 287, + 255 + ], + "type": "inline_equation", + "content": "\\vartheta = d_{\\Psi}(\\sqrt{\\dot{s}_0},\\sqrt{\\dot{s}_1}*h^*)" + }, + { + "bbox": [ + 47, + 219, + 287, + 255 + ], + "type": "text", + "content": ". The registration problem consider here is to find the minimizer " + }, + { + "bbox": [ + 47, + 219, + 287, + 255 + ], + "type": "inline_equation", + "content": "h^*" + }, + { + "bbox": [ + 47, + 219, + 287, + 255 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 47, + 219, + 287, + 255 + ], + "type": "inline_equation", + "content": "\\mathrm{Diff}_{+}([0,1])" + }, + { + "bbox": [ + 47, + 219, + 287, + 255 + ], + "type": "text", + "content": " of" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 262, + 291, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 291, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 291, + 301 + ], + "type": "interline_equation", + "content": "\\int_ {0} ^ {1} \\| c _ {0} (t) - \\left(c _ {1} * h\\right) (t) \\| ^ {2} + \\left\\| \\sqrt {\\dot {s} _ {0} (t)} - \\left(\\sqrt {\\dot {s} _ {1}} * h\\right) (t) \\right\\| ^ {2} d t. \\tag {33}", + "image_path": "3eaeeff08ba833f78f2791a79b113537cef7ac98aedf4917ae50de685a90546a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 311, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 287, + 348 + ], + "type": "text", + "content": "Using the reformulation principle of [3], that is " + }, + { + "bbox": [ + 47, + 311, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\gamma = s_1\\circ h\\circ" + }, + { + "bbox": [ + 47, + 311, + 287, + 348 + ], + "type": "inline_equation", + "content": "s_0^{-1}\\in \\mathrm{Diff}_+([0,1])" + }, + { + "bbox": [ + 47, + 311, + 287, + 348 + ], + "type": "text", + "content": ", this registration problem is shown to be equivalent to finding " + }, + { + "bbox": [ + 47, + 311, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\gamma^{*}\\in \\mathrm{Diff}_{+}([0,1])" + }, + { + "bbox": [ + 47, + 311, + 287, + 348 + ], + "type": "text", + "content": " that minimizes" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 369, + 287, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 369, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 58, + 369, + 287, + 418 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\int_ {0} ^ {1} \\left\\| \\frac {\\boldsymbol {\\theta} _ {0} (s)}{\\sqrt {\\lVert \\boldsymbol {\\theta} _ {0} (s) \\rVert}} - \\sqrt {\\gamma^ {\\prime} (s)} \\frac {\\boldsymbol {\\theta} _ {1} (\\gamma (s))}{\\sqrt {\\lVert \\boldsymbol {\\theta} _ {1} (\\gamma (s)) \\rVert}} \\right\\| ^ {2} \\\\ + \\| 1 - \\gamma^ {\\prime} (s) \\| ^ {2} d s. \\tag {34} \\\\ \\end{array}", + "image_path": "5ca7d5527a76153224fcfeb87ee056767c9fdbbedf81e9386eba9cd4d58b2318.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 428, + 287, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 428, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 47, + 428, + 287, + 500 + ], + "type": "text", + "content": "Note that this reformulation has the form of a penalized registration problem. The second term represents a penalty term on " + }, + { + "bbox": [ + 47, + 428, + 287, + 500 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 428, + 287, + 500 + ], + "type": "text", + "content": " and ensures a certain smoothness of the warping function. In this framework, the deformable templates are the square-root normalized curvatures which encode more geometric information than the unit tangent vector." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 511, + 128, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 511, + 128, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 511, + 128, + 525 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 531, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 287, + 579 + ], + "type": "text", + "content": "In this section, we report the experimental results of the proposed methods, comparing them with the SRVF method. We use both synthetic and real data. Additional results and figures are available in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 587, + 283, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 283, + 599 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 283, + 599 + ], + "type": "text", + "content": "5.1. Statistical estimation of the Frenet curvatures" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "The main limitation of shape analysis methods based on the Frenet curvatures is the need of additional estimates of curvatures. Being dependent on higher-order derivatives (up to order " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "), they are quite sensitive to the observation noise of the Euclidean curve. We detail here a simple method that can be used for their smooth estimation, and we refer to [21, 18] for a more complex and detailed statistical estimation algorithm. First, it is possible to use a local polynomial smoothing algorithm to estimate the " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " first" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "derivatives of the Euclidean curve [6]. From these derivatives, the raw estimates of the Frenet curvatures can be computed by using their extrinsic formulas. We propose here a second method based on the Frenet-Serret ODE approximation, to obtain the raw estimates, that appear to be more stable. A simple middle point approximation of the ODE solution gives" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 164, + 545, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 545, + 192 + ], + "type": "interline_equation", + "content": "Q \\left(s _ {j}\\right) \\approx Q \\left(s _ {i}\\right) \\exp \\left(\\left(s _ {i} - s _ {j}\\right) A _ {\\boldsymbol {\\theta}} \\left(\\frac {s _ {i} + s _ {j}}{2}\\right)\\right). \\tag {35}", + "image_path": "5231bccab4d57a75af368001c86678de8c439a3fa1e5c863b294edbe37e2255c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 200, + 545, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 200, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 304, + 200, + 545, + 310 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 200, + 545, + 310 + ], + "type": "inline_equation", + "content": "\\exp(.)" + }, + { + "bbox": [ + 304, + 200, + 545, + 310 + ], + "type": "text", + "content": " is the exponential map of the Lie group " + }, + { + "bbox": [ + 304, + 200, + 545, + 310 + ], + "type": "inline_equation", + "content": "SO(d)" + }, + { + "bbox": [ + 304, + 200, + 545, + 310 + ], + "type": "text", + "content": ". Then, using the inverse logarithm map " + }, + { + "bbox": [ + 304, + 200, + 545, + 310 + ], + "type": "inline_equation", + "content": "\\log(.)" + }, + { + "bbox": [ + 304, + 200, + 545, + 310 + ], + "type": "text", + "content": ", this gives an approximation of the matrix " + }, + { + "bbox": [ + 304, + 200, + 545, + 310 + ], + "type": "inline_equation", + "content": "A_{\\theta}((s_i + s_j) / 2) \\approx \\frac{1}{s_i - s_j} \\log(Q(s_i)^T Q(s_j))" + }, + { + "bbox": [ + 304, + 200, + 545, + 310 + ], + "type": "text", + "content": " and by identification raw estimates of " + }, + { + "bbox": [ + 304, + 200, + 545, + 310 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}((s_i + s_j) / 2)" + }, + { + "bbox": [ + 304, + 200, + 545, + 310 + ], + "type": "text", + "content": ". As we consider here a problem of estimating a functional parameter, we formulate the final " + }, + { + "bbox": [ + 304, + 200, + 545, + 310 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 304, + 200, + 545, + 310 + ], + "type": "text", + "content": " estimation problem as a penalized weighted functional regression with the obtained raw estimates, that we solve by using a B-spline approximation of " + }, + { + "bbox": [ + 304, + 200, + 545, + 310 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 304, + 200, + 545, + 310 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 318, + 489, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 318, + 489, + 331 + ], + "spans": [ + { + "bbox": [ + 306, + 318, + 489, + 331 + ], + "type": "text", + "content": "5.2. Experiments with synthetic curves" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 336, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 545, + 430 + ], + "type": "text", + "content": "We use synthetic data to highlight the differences between the methods discussed above (SRVF, SRC, and Frenet curvatures). The computations related to the SRVF method are made with the package fdasrsf. The SRC and Frenet curvatures methods are implemented with the code provided as supplementary material, including a dynamic programming algorithm for solving the registration problems." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 432, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 529 + ], + "type": "text", + "content": "We consider the simple case of a set of 20 curves in " + }, + { + "bbox": [ + 304, + 432, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 304, + 432, + 545, + 529 + ], + "type": "text", + "content": " with a single large peak of curvature. This one is created by generating curvature functions on " + }, + { + "bbox": [ + 304, + 432, + 545, + 529 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 304, + 432, + 545, + 529 + ], + "type": "text", + "content": " with one peak of maximum value 60.5, width 0.15, and location chosen randomly between 0.1 and 0.9. These curves have the shape of a loop made with a wire, where the loop is more or less close to the right or left wire end, depending on the location of the curvature peak." + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 307, + 536, + 544, + 621 + ], + "blocks": [ + { + "bbox": [ + 307, + 536, + 544, + 621 + ], + "lines": [ + { + "bbox": [ + 307, + 536, + 544, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 536, + 544, + 621 + ], + "type": "image", + "image_path": "cbc1abda0254f2cc9fe24155af5d9724f1ef96390b3e05fd1c98b5544b7308a7.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "text", + "content": "Figure 3: Matrices of pairwise SRVF distance (left), SRC distance (middle), and unparametrized Frenet curvatures distance (right) by sorted location of the peak curvature." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We compare the three methods through the pairwise distance matrices in Figure 3, and the geodesic paths computed" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4033" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 285, + 184 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 285, + 184 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 285, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 285, + 184 + ], + "type": "image", + "image_path": "5fc0e004decb1b717e2aff42397bce2527a270eba077b5f9c81e20677126ddeb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 203, + 287, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 203, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 287, + 240 + ], + "type": "text", + "content": "Figure 4: SRVF " + }, + { + "bbox": [ + 46, + 203, + 287, + 240 + ], + "type": "inline_equation", + "content": "(1^{st}" + }, + { + "bbox": [ + 46, + 203, + 287, + 240 + ], + "type": "text", + "content": " row), SRC " + }, + { + "bbox": [ + 46, + 203, + 287, + 240 + ], + "type": "inline_equation", + "content": "(2^{nd}" + }, + { + "bbox": [ + 46, + 203, + 287, + 240 + ], + "type": "text", + "content": " row), and Frenet curvatures " + }, + { + "bbox": [ + 46, + 203, + 287, + 240 + ], + "type": "inline_equation", + "content": "(3^{rd}" + }, + { + "bbox": [ + 46, + 203, + 287, + 240 + ], + "type": "text", + "content": " row) geodesic paths between curves with curvature peaks located at 0.27 (left) and 0.78 (right)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 255, + 286, + 335 + ], + "blocks": [ + { + "bbox": [ + 48, + 255, + 286, + 335 + ], + "lines": [ + { + "bbox": [ + 48, + 255, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 286, + 335 + ], + "type": "image", + "image_path": "9a8a764d07a6fe0f604a75abe30c4ddf12a14d9dd63c2068fe0f1915570ec2bc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 353, + 287, + 389 + ], + "lines": [ + { + "bbox": [ + 46, + 353, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 287, + 389 + ], + "type": "text", + "content": "Figure 5: Curvatures of the Euclidean curves along the geodesic paths plotted in Figure 4: SRVF " + }, + { + "bbox": [ + 46, + 353, + 287, + 389 + ], + "type": "inline_equation", + "content": "(1^{st}" + }, + { + "bbox": [ + 46, + 353, + 287, + 389 + ], + "type": "text", + "content": " row), SRC " + }, + { + "bbox": [ + 46, + 353, + 287, + 389 + ], + "type": "inline_equation", + "content": "(2^{nd}" + }, + { + "bbox": [ + 46, + 353, + 287, + 389 + ], + "type": "text", + "content": " row), and Frenet curvatures " + }, + { + "bbox": [ + 46, + 353, + 287, + 389 + ], + "type": "inline_equation", + "content": "(3^{rd}" + }, + { + "bbox": [ + 46, + 353, + 287, + 389 + ], + "type": "text", + "content": " row)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 413, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 675 + ], + "type": "text", + "content": "between two of these curves in Figure 4, with peaks located at 0.27 (red curve) and at 0.78 (blue curve). The corresponding deformations through the variations of the curvature along the different geodesic paths on Euclidean curves are shown in Figure 5, which highlights the strengths and weaknesses of each method. First, it emphasizes the \"nonelasticity\" of the unparametrized Frenet curvatures method, as in the middle of the geodesic path, we have two peaks of curvature and, therefore, a completely different shape without any loop. This explains the inconsistency of the heatmap under this method. Conversely, there is an elastic deformation of the curvature with the SRC transform, and shapes along the geodesic are consistent with the set of curves considered, which is well summarised on the corresponding heatmap where all distances are rather close to zero. For the SRVF method, the chosen example with peaks of curvature that are quite far apart shows that artifacts appear along the geodesic; the middle curve has two small loops at the edges. This phenomenon gives unreliable distances, as shown in the heatmap, where the distances are not monotone as a function of the spacing between the curvature peaks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "By considering a set of curves characterized by specific features, we observe a clear difference in consistency of the shapes along the geodesics with respect to the different" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "methods. This phenomenon is quite visible on the geodesic between helices in 2D, Figure 1 or 3D, Figure 2. In that case, within both geodesic paths under SRVF method, the curves lose the characteristic geometry of the helix. A three-dimensional circular helix is characterized by having a constant curvature and torsion, which is not the case along the SRVF geodesic, but preserved with the SRC method. In that case the geodesic under the Frenet curvature representation is very similar to SRC." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 187, + 520, + 200 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 187, + 520, + 200 + ], + "spans": [ + { + "bbox": [ + 306, + 187, + 520, + 200 + ], + "type": "text", + "content": "5.3. Application to sign language motion data" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 205, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 396 + ], + "type": "text", + "content": "It appears that curvilinear velocity and Frenet curvatures are particularly relevant parameters for the analysis of human motion. Several laws involving these parameters can be found in the literature [10, 14, 19, 7]; among others, the power laws state a special relationship between the curvature, the torsion, and the velocity of a point trajectory representing human motion. Using a method that conserves the shape of these parameters is, therefore, of particular interest in this application. We demonstrate here with the case of wrist trajectories in sign language, acquired with a motion capture system by the company MocapLab (https://www.mocaplab.com/fr/). We compute the geodesic paths, under each of the frameworks, between the arbitrarily chosen red and blue curves within the set of several repetitions of the sign \"Femme\", shown in Figure 6 with the corresponding time-parameterized Frenet curvatures." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 407, + 545, + 550 + ], + "blocks": [ + { + "bbox": [ + 310, + 407, + 545, + 550 + ], + "lines": [ + { + "bbox": [ + 310, + 407, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 310, + 407, + 545, + 550 + ], + "type": "image", + "image_path": "12031780e212557ad881f4993741d2ee443aaaf43a8f0ee4cb443105d3e87eb4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 569, + 545, + 628 + ], + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 628 + ], + "type": "text", + "content": "Figure 6: Trajectories of the right wrist while signing \"Femme\" in sign language: 3D curves (left), time-parametrized curvatures (top right), torsions (bottom right). The blue and red ones are used to compute the geodesic in Figure 7." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Figure 7 emphasizes the advantage of considering a representation depending on the parameterization, allowing a registration before computing the geodesic. However, it also shows that considering only the tangent vector as a representative object (SRVF) is not sufficient to find the optimal reparametrization that correctly aligns the torsions," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4034" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 214, + 156 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 214, + 156 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 214, + 156 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 214, + 156 + ], + "type": "image", + "image_path": "d253497a348fef4db1a60db1dab9aaa3da2452f6ac63983d86347dfd6a6ea21b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 72, + 372, + 157 + ], + "blocks": [ + { + "bbox": [ + 220, + 72, + 372, + 157 + ], + "lines": [ + { + "bbox": [ + 220, + 72, + 372, + 157 + ], + "spans": [ + { + "bbox": [ + 220, + 72, + 372, + 157 + ], + "type": "image", + "image_path": "24d15a2e1a679858767a95bfd2d27b4406775c97261d47fbaa8ae71f4cb7ff39.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 380, + 72, + 533, + 157 + ], + "blocks": [ + { + "bbox": [ + 380, + 72, + 533, + 157 + ], + "lines": [ + { + "bbox": [ + 380, + 72, + 533, + 157 + ], + "spans": [ + { + "bbox": [ + 380, + 72, + 533, + 157 + ], + "type": "image", + "image_path": "92e3396783fc03519b46608c579f4863a1642dd9b2d192bc286e3092cb2e3094.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 62, + 158, + 213, + 242 + ], + "blocks": [ + { + "bbox": [ + 62, + 158, + 213, + 242 + ], + "lines": [ + { + "bbox": [ + 62, + 158, + 213, + 242 + ], + "spans": [ + { + "bbox": [ + 62, + 158, + 213, + 242 + ], + "type": "image", + "image_path": "28b92fb8e0014377beb36b6b3827aa6da9697bc4ae4a51c603e6096c8b6f85e5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 120, + 255, + 156, + 266 + ], + "lines": [ + { + "bbox": [ + 120, + 255, + 156, + 266 + ], + "spans": [ + { + "bbox": [ + 120, + 255, + 156, + 266 + ], + "type": "text", + "content": "(a) SRVF" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 221, + 159, + 372, + 243 + ], + "blocks": [ + { + "bbox": [ + 221, + 159, + 372, + 243 + ], + "lines": [ + { + "bbox": [ + 221, + 159, + 372, + 243 + ], + "spans": [ + { + "bbox": [ + 221, + 159, + 372, + 243 + ], + "type": "image", + "image_path": "d2655f7d998c4b48da89d11ae702c92a33e690d35bc6e453b57ac33e133d43a8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 281, + 256, + 313, + 266 + ], + "lines": [ + { + "bbox": [ + 281, + 256, + 313, + 266 + ], + "spans": [ + { + "bbox": [ + 281, + 256, + 313, + 266 + ], + "type": "text", + "content": "(b) SRC" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 380, + 159, + 531, + 242 + ], + "blocks": [ + { + "bbox": [ + 380, + 159, + 531, + 242 + ], + "lines": [ + { + "bbox": [ + 380, + 159, + 531, + 242 + ], + "spans": [ + { + "bbox": [ + 380, + 159, + 531, + 242 + ], + "type": "image", + "image_path": "355efc6332217c06be6fc0cde171206efd88da0d6e977d02b9b8971691864cf5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 256, + 494, + 266 + ], + "lines": [ + { + "bbox": [ + 417, + 256, + 494, + 266 + ], + "spans": [ + { + "bbox": [ + 417, + 256, + 494, + 266 + ], + "type": "text", + "content": "(c) Frenet curvatures" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 65, + 322, + 161, + 406 + ], + "blocks": [ + { + "bbox": [ + 46, + 278, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 278, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 547, + 304 + ], + "type": "text", + "content": "Figure 7: Comparison between time-parametrized curvature and torsion along the geodesic path under SRVF (left), SRC (middle), and Frenet curvatures (right)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 322, + 161, + 406 + ], + "lines": [ + { + "bbox": [ + 65, + 322, + 161, + 406 + ], + "spans": [ + { + "bbox": [ + 65, + 322, + 161, + 406 + ], + "type": "image", + "image_path": "7940fa0bb2ddd7117f5c3e07cf8db4e39c36103725fcc4d71cfb7725b3fef637.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 422, + 157, + 433 + ], + "lines": [ + { + "bbox": [ + 69, + 422, + 157, + 433 + ], + "spans": [ + { + "bbox": [ + 69, + 422, + 157, + 433 + ], + "type": "text", + "content": "(a) Warping functions " + }, + { + "bbox": [ + 69, + 422, + 157, + 433 + ], + "type": "inline_equation", + "content": "h" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 443, + 287, + 468 + ], + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 468 + ], + "type": "text", + "content": "Figure 8: Comparison of estimated warping functions " + }, + { + "bbox": [ + 46, + 443, + 287, + 468 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 443, + 287, + 468 + ], + "type": "text", + "content": " (left) and " + }, + { + "bbox": [ + 46, + 443, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 443, + 287, + 468 + ], + "type": "text", + "content": " (right) to compute the geodesic in Figure 7." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 172, + 322, + 271, + 406 + ], + "blocks": [ + { + "bbox": [ + 172, + 322, + 271, + 406 + ], + "lines": [ + { + "bbox": [ + 172, + 322, + 271, + 406 + ], + "spans": [ + { + "bbox": [ + 172, + 322, + 271, + 406 + ], + "type": "image", + "image_path": "386b59cb3851d2575d791fcbdca101540e1cd500a4e70bcdd85b65cc802647e0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 422, + 266, + 434 + ], + "lines": [ + { + "bbox": [ + 178, + 422, + 266, + 434 + ], + "spans": [ + { + "bbox": [ + 178, + 422, + 266, + 434 + ], + "type": "text", + "content": "(b) Warping functions " + }, + { + "bbox": [ + 178, + 422, + 266, + 434 + ], + "type": "inline_equation", + "content": "\\gamma" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 489, + 288, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 489, + 288, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 288, + 597 + ], + "type": "text", + "content": "and could affect subsequent analyses made by the SRVF method, such as the mean. This results in the appearance of new minimums, maximums and zeros in the torsion functions along the SRVF geodesic. However, such characteristic points of the curvature, torsion and velocity functions are crucial in the observation of the laws of motion. It is therefore preferable to use a method that preserves these characteristics by directly optimizing the optimal alignment from these parameters, such as the proposed SRC method." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 609, + 119, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 119, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 119, + 621 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "The square-root curvature transform of a Euclidean curve in " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": " is a representation that encodes more geometric information of the curves, and thus the results are easier to interpret than existing methods. The main limitation lies in the estimation of the Frenet curvatures from real and noisy data, nevertheless recent smooth statistical estimators can be used for computing the SRC [21, 18]. We believe our" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 324, + 546, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 546, + 372 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 546, + 372 + ], + "type": "text", + "content": "method is particularly interesting for motion trajectory analysis and could be developed further in the future as a tool for generation, segmentation and classification of complex trajectories." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 385, + 365, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 385, + 365, + 397 + ], + "spans": [ + { + "bbox": [ + 306, + 385, + 365, + 397 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 310, + 406, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 310, + 406, + 546, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 406, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 310, + 406, + 546, + 460 + ], + "type": "text", + "content": "[1] Martin Bauer, Nicolas Charon, Eric Klassen, and Alice Le Brigant. Intrinsic riemannian metrics on spaces of curves: Theory and computation. Handbook of Mathematical Models and Algorithms in Computer Vision and Imaging, pages 1-35, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 311, + 462, + 547, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 462, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 311, + 462, + 547, + 506 + ], + "type": "text", + "content": "[2] Martin Bauer, Nicolas Charon, Eric Klassen, Sebastian Kurtek, Tom Needham, and Thomas Pierron. Elastic Metrics on Spaces of Euclidean Curves: Theory and Algorithms, 2022. arXiv:2209.09862 [math]. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 311, + 508, + 547, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 508, + 547, + 563 + ], + "spans": [ + { + "bbox": [ + 311, + 508, + 547, + 563 + ], + "type": "text", + "content": "[3] Nicolas J.-B. Brunel and Juhyun Park. The frenet-serret framework for aligning geometric curves. In Frank Nielsen and Frédéric Barbaresco, editors, Geometric Science of Information, pages 608-617, Cham, 2019. Springer International Publishing. 3, 4, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 311, + 565, + 547, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 565, + 547, + 608 + ], + "spans": [ + { + "bbox": [ + 311, + 565, + 547, + 608 + ], + "type": "text", + "content": "[4] Elena Celledoni, Markus Eslitzbichler, and Alexander Schmeding. Shape analysis on lie groups with applications in computer animation. Journal of Geometric Mechanics, 8(3):273-304, 2016. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 311, + 610, + 547, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 610, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 311, + 610, + 547, + 643 + ], + "type": "text", + "content": "[5] Marvin Eisenberger and Daniel Cremers. Hamiltonian dynamics for real-world shape interpolation. In European Conference on Computer Vision (ECCV), 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 311, + 646, + 547, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 646, + 547, + 678 + ], + "spans": [ + { + "bbox": [ + 311, + 646, + 547, + 678 + ], + "type": "text", + "content": "[6] Jianqing Fan and Irene Gijbels. Local polynomial modelling and its applications. CRC monographs on statistics and applied probability 66. Chapman & Hall, 1996. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 311, + 680, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 680, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 311, + 680, + 547, + 713 + ], + "type": "text", + "content": "[7] Tamar Flash and Alain Berthoz. *Space-Time Geometries for Motion and Perception in the Brain and the Arts*. Lecture Notes in Morphogenesis. Springer, 2021. 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4035" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 52, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 52, + 72, + 287, + 117 + ], + "type": "text", + "content": "[8] Lukas Koestler, Daniel Grittner, Michael Moeller, Daniel Cremers, and Zorah Lähner. Intrinsic neural fields: Learning functions on manifolds. In European Conference on Computer Vision (ECCV), 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 118, + 286, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 118, + 286, + 138 + ], + "spans": [ + { + "bbox": [ + 52, + 118, + 286, + 138 + ], + "type": "text", + "content": "[9] Wolfgang Kühnel. Differential Geometry Curves - Surfaces, volume 77. 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "text", + "content": "[10] Francesco Lacquaniti, Carlo Terzuolo, and Paolo Viviani. The law relating the kinematic and figural aspects of drawing movements. Acta Psychologica (Amst.), 54:115-130, 1983. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 185, + 286, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 286, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 286, + 207 + ], + "type": "text", + "content": "[11] Serge Lang. Differential and riemannian manifolds. Springer, 1999:1-6, 2006. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 286, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 286, + 228 + ], + "type": "text", + "content": "[12] Robert Lyons. Frobenius theorem two ways. Lecture note, 2016. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "text", + "content": "[13] Zorah Lähner, Emanuele Rodola, Frank R. Schmidt, Michael M. Bronstein, and Daniel Cremers. Efficient globally optimal 2d-to-3d deformable shape matching. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), May 2016. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "text", + "content": "[14] Uri Maoz, Alain Berthoz, and Tamar Flash. Complex unconstrained three-dimensional hand movement and constant equi-affine speed. Journal of Neurophysiology, 101:1002-1015, 2009. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 331, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 364 + ], + "type": "text", + "content": "[15] J. S. Marron, James O. Ramsay, Laura M. Sangalli, and Anuj Srivastava. Functional data analysis of amplitude and phase variation. Statistical Science, 30:468-484, 2015. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 365, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 387 + ], + "type": "text", + "content": "[16] Turk J Math. Submanifolds of Riemannian Product Manifolds. 29:389-401, 2005. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 388, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 420 + ], + "type": "text", + "content": "[17] Tom Needham. Shape Analysis of Framed Space Curves. Journal of Mathematical Imaging and Vision, 61(8):1154-1172, Oct. 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "text", + "content": "[18] Juhyun Park, Nicolas Brunel, and Perrine Chassat. Curvature and torsion estimation of 3d functional data: A geometric approach to build the mean shape under the frenet serret framework, 2022. arXiv:2203.02398 [stat]. 7, 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "text", + "content": "[19] Frank E. Pollick, Uri Maoz, Amir A. Handzel, Peter J. Giblin, Guillermo Sapiro, and Tamar Flash. Three-dimensional arm movements at constant equi-affine speed. Cortex, 45:325-339, 2009. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 512, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 544 + ], + "type": "text", + "content": "[20] Marianna Saba. On the usage of the curvature for the comparison of planar curves. PhD thesis, University of Cagliari, 2012. 2, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 545, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 610 + ], + "type": "text", + "content": "[21] Laura M Sangalli, Piercesare Secchi, Simone Vantini, and Alessandro Veneziani. Efficient estimation of three-dimensional curves and their derivatives by free-knot regression splines, applied to the analysis of inner carotid artery centrelines. Journal of the Royal Statistical Society: Series C (Applied Statistics), 58(3):285-306, 2009. 7, 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 612, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 635 + ], + "type": "text", + "content": "[22] Stefan Sommer, Tom Fletcher, and Xavier Pennec. Introduction to differential and Riemannian geometry. 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "text", + "content": "[23] Anuj Srivastava and Eric Klassen. Functional and Shape Data Analysis. Springer Series in Statistics. Springer New York, 2016. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[24] Anuj Srivastava, Eric Klassen, Shantanu H. Joshi, and Ian H. Jermyn. Shape analysis of elastic curves in euclidean spaces. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(7):1415-1428, 2011. 1, 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 275 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 115 + ], + "type": "text", + "content": "[25] Tatiana Surazhsky and Gershon Elber. Metamorphosis of planar parametric curves via curvature interpolation. International Journal of Shape Modeling, 8:201-216, 2002. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 138 + ], + "type": "text", + "content": "[26] Wang Tixiang. Morse theory on banach manifolds. Acta Mathematica Sinica, 5:250-262, 1989. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 183 + ], + "type": "text", + "content": "[27] J. Derek Tucker, Wei Wu, and Anuj Srivastava. Generative models for functional data using phase and amplitude separation. Computational Statistics and Data Analysis, 61:50-66, 2013. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "type": "text", + "content": "[28] Laurent Younes. Computable Elastic Distances between Shapes. SIAM Journal on Applied Mathematics, 58(2):565-586, 1998. Publisher: Society for Industrial and Applied Mathematics. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 230, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 252 + ], + "type": "text", + "content": "[29] Laurent Younes. *Shapes and Diffeomorphisms*. Applied Mathematical Sciences. Springer, 2010. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 253, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 275 + ], + "type": "text", + "content": "[30] Laurent Younes. Elastic distance between curves under the metamorphosis viewpoint, 2018. arXiv:1804.10155. 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4036" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Shape Anchor Guided Holistic Indoor Scene Understanding/4f878220-d6cc-48e2-9d1e-82c1cf350095_content_list.json b/2023/Shape Anchor Guided Holistic Indoor Scene Understanding/4f878220-d6cc-48e2-9d1e-82c1cf350095_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2941fc70eb82fb37b267be9a2028d56dadc1dd04 --- /dev/null +++ b/2023/Shape Anchor Guided Holistic Indoor Scene Understanding/4f878220-d6cc-48e2-9d1e-82c1cf350095_content_list.json @@ -0,0 +1,1792 @@ +[ + { + "type": "text", + "text": "Shape Anchor Guided Holistic Indoor Scene Understanding", + "text_level": 1, + "bbox": [ + 181, + 130, + 789, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingyue Dong $^{1}$ , Linxi Huan $^{1}$ , Hanjiang Xiong $^{1}$ , Shuhan Shen $^{2}$ , Xianwei Zheng $^{1*}$ $^{1}$ The State Key Lab. LIESMARS, Wuhan University \n $^{2}$ Institute of Automation, Chinese Academy of Sciences \n{dmy25148, whu.hlx, xionghanjiang, zhengxw}@whu.edu.cn \nshshen@nlpr.iac.ac.cn", + "bbox": [ + 151, + 179, + 823, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper proposes a shape anchor guided learning strategy (AncLearn) for robust holistic indoor scene understanding. We observe that the search space constructed by current methods for proposal feature grouping and instance point sampling often introduces massive noise to instance detection and mesh reconstruction. Accordingly, we develop AncLearn to generate anchors that dynamically fit instance surfaces to (i) unmix noise and target-related features for offering reliable proposals at the detection stage, and (ii) reduce outliers in object point sampling for directly providing well-structured geometry priors without segmentation during reconstruction. We embed AncLearn into a reconstruction-from-detection learning system (AncRec) to generate high-quality semantic scene models in a purely instance-oriented manner. Experiments conducted on the challenging ScanNetv2 dataset demonstrate that our shape anchor-based method consistently achieves state-of-the-art performance in terms of 3D object detection, layout estimation, and shape reconstruction. The code will be available at https://github.com/Geo-Tell/AncRec.", + "bbox": [ + 75, + 335, + 473, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 654, + 209, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Holistic indoor scene understanding from partial observations (e.g., single-view images or 3D scans) is a comprehensive task that provides 3D semantic scene models for indoor applications. Early works studied this task with a reconstruction-from-detection framework that recovers the geometries of room structures and objects from the corresponding detection in a separate way. Later, end-to-end learning methods were proposed to simultaneously perform layout estimation, object detection, and shape prediction in one forward pass for semantic scene reconstruction [27, 37, 23]. With the recent success of point-based 3D detection and instance reconstruction, surging interest has been witnessed in detecting and modeling objects directly", + "bbox": [ + 75, + 680, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ca815fdf223d19a4adf61d75adbbb28b2e18d759a3aa9c5b2e9623b2e1eca591.jpg", + "image_caption": [ + "Figure 1. Comparison between different feature grouping strategies. (a) the original scene model. (b)-(e) The different feature grouping operations all suffer from the issue of confusing nontarget noise with useful features. (f) The proposed shape anchor guided grouper directly generates anchors at the instance surface to merge instance-related features, which largely alleviates noise interference." + ], + "image_footnote": [], + "bbox": [ + 501, + 303, + 890, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "from sparse point clouds [28, 33].", + "bbox": [ + 500, + 700, + 723, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Benefiting from the use of rich geometry information, current scan-based deep methods have improved the performance of semantic scene reconstruction. However, they still left two issues that bottleneck high-quality semantic reconstruction: (1) the noisy instance feature learning at the detection phase, and (2) the difficulty in retrieving instances from sparse point clouds for reconstruction.", + "bbox": [ + 496, + 717, + 890, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "At the detection phase, it is required to group features for instance representation learning. The ball query [30, 31] and 3D convolution[39, 20] are two basic operations for point feature grouping, but they often mix massive noise with informative features due to the fixed grouping range,", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21916", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "as shown in Fig. 1 (b) and (c). To compensate for the deficiency caused by the fixed range, VoteNet [14] and its variants [34, 29, 38, 35] adopt a voting strategy to cluster object features by moving surface points towards object centers. Albeit more flexible than the basic operations, the voting-based strategy often generates an unconstrained grouping area that brings quantities of outliers as illustrated in Fig. 1 (d). BRNet [10] hence restricts the grouping space by sampling around representative points given by coarse box proposals. Nevertheless, limited by box-like grouping area, the sampling points can still fall far beyond targets when the objects are irregularly shaped, as depicted in Fig. 1 (e).", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "At the reconstruction stage, the retrieval of outlier-free object points is a prerequisite for object recovery. Due to the noise introduced by feature grouping in the previous detection phase, the points grouped for localizing objects can hardly serve as ideal reconstruction priors as indicated by Fig. 1. Consequently, the current methods are forced to employ an additional foreground classifier [28] or replace the detector with a complex instance segmentation backbone to sample object points from the raw scans [33]. However, the existence of numerous non-target outliers in the search space challenges instance segmentation, resulting in increased risks of gluing different instances and misclassifying background points.", + "bbox": [ + 75, + 273, + 472, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the discussion above, the two noise interference issues during detection and reconstruction are actually highly coupled and can be resolved together as long as outliers are excluded during feature grouping. To this end, we are motivated to propose a shape anchor-guided learning strategy (AncLearn) that generates surface anchors to fit the feature grouping areas to object shape distributions, as displayed in Fig. 1 (f). With the geometry constraint provided by surface anchors, it is able to merge local target-focused features for predicting reliable object proposals and construct a shape-aware search space for robustly sampling instance points without segmentation during reconstruction. The proposed anchor-guided learning strategy can be easily embedded into an end-to-end learning system to accomplish object detection, layout estimation, and instance reconstruction for holistic scene understanding. The main contributions are summarized as follows:", + "bbox": [ + 75, + 468, + 472, + 726 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a shape anchor guided learning strategy to simultaneously address the issues of noisy feature learning in detection and instance point retrieval during reconstruction.", + "- We embed the proposed anchor-guided learning strategy into an end-to-end learning system to accomplish object detection, layout estimation, and instance reconstruction for holistic scene understanding in a purely instance-oriented way.", + "- Extensive experiments demonstrate that our AncRec achieves high-quality semantic scene reconstruction" + ], + "bbox": [ + 94, + 731, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with state-of-the-art performance in instance detection and mesh prediction on the challenging ScanNetv2 dataset [12] (some ground truths provided by Scan2CAD [1] and SceneCAD [3]).", + "bbox": [ + 531, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 169, + 640, + 184 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic scene understanding has been extensively studied over the past years. Many methods focused on acquiring the semantics of 3D scenes [22, 15, 8], while others recovered scene geometries by scene completion [18, 13, 36]. Recently, increasing interest has emerged in semantic scene reconstruction, which recovers both the semantics and geometric shapes of objects. By treating semantic reconstruction as a problem of holistic scene understanding, promising progress has been made based on a reconstruction-from-detection principle [28]. In the following, we review the research from the two core aspects of the reconstruction-from-detection pipeline, i.e., 3D object detection and scene-aligned instance reconstruction.", + "bbox": [ + 496, + 186, + 890, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 3D Object Detection", + "text_level": 1, + "bbox": [ + 500, + 393, + 692, + 410 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Over the past years, deep detectors have gained great success in 2D object detection [40, 25]. The progress of 2D deep detection has inspired the development of deep learning techniques for recognizing objects from scene point clouds. Compared to image data, point clouds provide rich surface geometry clues for locating objects in real scenes. Nevertheless, the sparse, irregular, and orderless characteristics of point clouds make them hard to be handled by the grid-based convolution model.", + "bbox": [ + 496, + 417, + 890, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Works in the early stage leveraged the 2D proposals as 3D detection constraints or projected point clouds into regular 2D/3D grids. Although these 3D detectors are applicable, their performance is still limited by the 2D detectors and the missing geometric details brought by projection. To directly learn the rich geometry features from the raw points, PointNet [30] and PointNet++ [31] used a ball query operation for grouping point features. Later, PointRCNN [32] adopted the point-wise features extracted by PointNet++ for producing instance proposals with point clouds. Hindered by the fixed feature grouping range of grid-based convolutions or ball query operation, these methods often mixed massive noise with informative features, which impairs the reliability of proposals. Considering that the observed object surface points usually lie far away from the object centers, VoteNet [14] introduced a voting mechanism for proposal generation. Based on the voting mechanism, numerous variants further refined proposal features with context relationships [34], hierarchical clues [35], and hybrid geometric primitives [38]. However, the issue of outliers in votes still blocks the learning of representative box features. BRNet [10] derived proposal features based on the generated virtual points given by coarse box proposals,", + "bbox": [ + 496, + 553, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21917", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/43e033d447d7f6b71e0c3b5db2bb4fda8a1cfc8ec027ee4a06e2d4c3df9f5b85.jpg", + "image_caption": [ + "Figure 2. Overview of the AncRec framework. At the detection stage, seed point features for walls and objects are first learned with a modified dual-branch PointNet++. Following each branch, a voting module and an anchor-guided feature grouper respectively generate proposal features $f_{\\mathrm{obj(wall)}}^{\\mathrm{vote}}$ and $f_{\\mathrm{obj(wall)}}^{\\mathrm{anchor}}$ , which are fed into the decoders to predict object bounding boxes and wall quads. The room layout is then constructed by processing the wall quads into connected corners. At the reconstruction stage, objects with high objectness scores are reconstructed under the guidance of $f_{\\mathrm{obj}}^{\\mathrm{vote}}$ , $f_{\\mathrm{obj}}^{\\mathrm{anchor}}$ , and the geometry priors sampled by shape anchors. Finally, object models are arranged in the scene with the predicted layout according to the spatial alignment provided by predicted bounding boxes." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 890, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "whereas the virtual points can fall into non-target areas and bring noise. In this paper, we directly merge features anchored on target surfaces into robust proposal representations for indoor instance learning.", + "bbox": [ + 75, + 450, + 470, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Scene-aligned Instance Reconstruction", + "text_level": 1, + "bbox": [ + 76, + 526, + 410, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scene-aligned instance reconstruction requires not only modeling 3D object shapes but also correctly arranging the shapes in the 3D scene space. The semantic model of an indoor scene was early constructed with retrieval techniques, which search for CAD shapes or geometric primitives from an offline database and align the approximate object models to input scene data [1, 2, 17, 24]. Albeit able to present delicate scene models, the retrieval-based methods lack generalization ability to various scenes due to the inefficient inference and the limited database scale [37].", + "bbox": [ + 75, + 551, + 468, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The promising advances in deep learnable shape representations motivated scene-aligned instance reconstruction without a finite model pooling [16, 11]. Many prior arts in literature generate 3D objects with explicit or implicit representations with learned features derived from 2D recognition results [19, 27, 37, 23]. Unlike previous single-view modeling methods, RevealNet [21] and the following works [5, 4] predicted the occupancy grids of semantic instances with 3D features extracted from voxelized RGB-D scans. For high scene resolution that is hard to acquire by volumetric representation, RfD-Net [28] and DIMR [33] extracted instance points from scans with segmentation techniques for subsequent single object reconstruction with deep implicit", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "functions [26, 9]. Because of the existence of non-target points, these two state-of-the-art methods have to use an extra foreground classifier or a complex instance segmentation backbone to sample object points for reconstruction. In this paper, we utilized the previously generated surface anchors to localize instance points in scans. Hereby, we introduced an easy but effective anchor-guided sampling strategy to offer geometry priors for the following instance reconstruction without an extra segmentation stage.", + "bbox": [ + 496, + 450, + 893, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 592, + 591, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We illustrate the proposed AncRec in Fig. 2. AncRec achieves semantic scene reconstruction with the proposed anchor-guided learning strategy in an instance-oriented way. It first localizes objects and walls via a dual-branch detector that is equipped with shape anchors learned for grouping point features at the detection stage. The object shape anchors are subsequently leveraged to sample instance geometry priors for predicting object meshes. Finally, the complete semantic scene model is reconstructed by arranging object meshes in the post-processed room layout with alignment to the parsed bounding boxes and poses. In the following, we elaborate on how the shape anchors are learned and leveraged to address the noise interference during instance detection and mesh prediction for high-quality semantic scene reconstruction.", + "bbox": [ + 496, + 611, + 893, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Shape Anchor-guided Instance Detection", + "text_level": 1, + "bbox": [ + 500, + 847, + 849, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We modify the VoteNet into a dual-branch instance detector to simultaneously localize objects and walls in one", + "bbox": [ + 500, + 869, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21918", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "forward pass (Fig. 2). Although the voting mechanism of VoteNet is workable for generating proposals, the reliability of voting-based predictions is often impaired by the outliers in votes. We hence design an anchor-guided feature grouper to learn target-focused features for robust instance detection.", + "bbox": [ + 75, + 90, + 472, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.1 Anchor-guided Feature Grouper", + "text_level": 1, + "bbox": [ + 76, + 188, + 359, + 204 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We illustrate the mechanism of the anchor-guided feature grouper in Fig. 3. In the following, by taking the object detection branch as an example, we describe how it works with the voting module in three steps to predict the oriented 3D bounding boxes.", + "bbox": [ + 75, + 205, + 470, + 282 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/85f2b401fbee1043fc1cd65dfd1e1c3b144c54a075f84651277304500c978e31.jpg", + "image_caption": [ + "Figure 3. The anchor-guided feature grouper. The anchor-guided proposal feature $f_{i}^{\\text{anchor}}$ is the average of features at shape anchors. The shape anchors are deformed from initial anchors $\\{a_{j}\\}$ with offsets derived from the concatenation of the voting-based proposal feature $f_{i}^{\\text{vote}}$ and $\\{a_{j}\\}$ . The anchors are then translated to the cluster center $c_{i}$ ." + ], + "image_footnote": [], + "bbox": [ + 76, + 292, + 472, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 1: Anchor Generation. Given the $i^{th}$ voting-based proposal feature $\\pmb{f}_i^{\\mathrm{vote}}\\in \\mathbb{R}^{128}$ , the anchor-guided grouper first generates shape anchors that depict the geometry of the $i^{th}$ object candidate via template deformation. With $N$ initial anchors $A = \\{\\pmb {a}_j\\in \\mathbb{R}^3 |j = 1,\\dots,N\\}$ uniformly selected from a unit ball surface, a deformation layer is applied to translate the initial anchors to the $i^{th}$ candidate object surface by", + "bbox": [ + 75, + 551, + 468, + 674 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {a}} _ {j} = \\boldsymbol {a} _ {j} + \\underbrace {\\operatorname {T a n h} \\left(\\mathrm {M L P} \\left([ \\boldsymbol {a} _ {j} , \\boldsymbol {f} _ {i} ^ {\\text {v o t e}} ]\\right)\\right)} _ {\\text {d e f o r m a t i o n o f t s e t} \\Delta \\boldsymbol {a} _ {j}} + \\boldsymbol {c} _ {i}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 684, + 468, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As indicated by Eq. (1), the deformation layer moves $\\pmb{a}_j$ with offsets $\\Delta a_j$ inferred by a multi-layer perceptron (MLP) and a Tanh activation function and then places the deformed anchors at the candidate center $\\pmb{c}_i$ clustered by the voting module. The training of the deformation layer is supervised by the Chamfer distance loss defined as", + "bbox": [ + 75, + 732, + 470, + 823 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {a n c h o r}} = \\frac {1}{| P _ {\\mathrm {g t}} |} \\sum_ {\\boldsymbol {p} \\in P _ {\\mathrm {g t}}} \\min _ {\\hat {\\boldsymbol {a}} \\in \\hat {A}} \\| \\boldsymbol {p} - \\hat {\\boldsymbol {a}} \\| _ {2} ^ {2} \\tag {2} \\\\ + \\frac {1}{| \\hat {A} |} \\sum_ {\\hat {\\boldsymbol {a}} \\in \\hat {A}} \\min _ {\\boldsymbol {p} \\in P _ {\\mathrm {g t}}} \\| \\hat {\\boldsymbol {a}} - \\boldsymbol {p} \\| _ {2} ^ {2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 824, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{A}$ and $P_{\\mathrm{gt}}$ denote the sets of deformed anchors and surface points sampled from the corresponding scene-aligned 3D meshes. Thereby, the anchors are learned to fit the $i^{th}$ object surface, and the clustering center $c_{i}$ is also refined with shape constraint provided by Eq. (2). Moreover, due to the supervision with complete surface points, the shape anchors can assist in exploiting context for recovering unobserved structures, e.g., the missing object bottom that is supported by the observed floor areas.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 2: Feature Grouping. The anchor-guided grouper propagates the seed point features (extracted by the dual-branch PointNet++ backbone) to each anchor via interpolation followed by an MLP. As the deformed anchors are mainly located on the object surface, the seed point features for propagation can be reliably selected from target-related areas. The $i^{th}$ noise-reduced proposal representation $\\pmb{f}_i^{\\mathrm{anchor}}\\in \\mathbb{R}^{128}$ is obtained by averaging the anchor features.", + "bbox": [ + 496, + 227, + 890, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 3: Prediction Fusion. We employ two decoders to predict the object category and oriented box parameters from $f_{i}^{\\mathrm{vote}}$ and $f_{i}^{\\mathrm{anchor}}$ respectively for the $i^{th}$ candidate. This is based on the consideration that the voting-based proposal feature $f_{i}^{\\mathrm{vote}}$ contains more contextual information while the anchor-guided proposal feature $f_{i}^{\\mathrm{anchor}}$ is more target-focused. The estimated parameters, denoted as $\\Theta_{i}^{\\mathrm{vote}}$ and $\\Theta_{i}^{\\mathrm{anchor}}$ , are averaged with learnable weights to obtain the final object parameters:", + "bbox": [ + 496, + 349, + 890, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Theta} _ {i} = \\boldsymbol {w} _ {1} \\cdot \\boldsymbol {\\Theta} _ {i} ^ {\\text {v o t e}} + \\boldsymbol {w} _ {2} \\cdot \\boldsymbol {\\Theta} _ {i} ^ {\\text {a n c h o r}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 494, + 890, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From the three steps above, the attributes of objects are parsed with robustness to noise. The anchor-guided detection of wall instances works in a similar way. Considering that walls are generally connected to each other, we additionally deploy the attention operation used in [35] to enhance the anchor-guided wall proposal features with the strong relationship between walls for precise layout estimation.", + "bbox": [ + 496, + 525, + 890, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2 The Detection Training Loss", + "text_level": 1, + "bbox": [ + 498, + 654, + 751, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The total loss function of the anchor-guided instance detector is defined as", + "bbox": [ + 496, + 672, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {o b j}} + \\mathcal {L} _ {\\mathrm {w a l l}} + \\mathcal {L} _ {\\mathrm {o b j}} ^ {\\text {a n c h o r}} + \\mathcal {L} _ {\\mathrm {w a l l}} ^ {\\text {a n c h o r}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 713, + 890, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Eq. (4), $\\mathcal{L}_{\\mathrm{obj}}$ is the object detection loss given by [14] while $\\mathcal{L}_{\\mathrm{wall}}$ is the wall quad loss used in [7]. $\\mathcal{L}_{\\mathrm{obj}}^{\\mathrm{anchor}}$ and $\\mathcal{L}_{\\mathrm{wall}}^{\\mathrm{anchor}}$ are the chamfer distance losses that supervise the anchor deformation learning in terms of positive object and wall candidates.", + "bbox": [ + 496, + 743, + 890, + 820 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Anchor-guided Object Reconstruction", + "text_level": 1, + "bbox": [ + 500, + 830, + 828, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instance points from the input scan are desirable geometry priors for reconstruction. Previous works utilized segmentation operations for instance point sampling. Despite", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21919", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "being applicable, accurate instance segmentation is difficult due to the existence of massive non-target noise in the search space. In contrast, we take advantage of the shape anchors to construct a shape-aware search space, in which object points can be directly localized with little noise interference. Fig. 4 illustrates the workflow of our anchor-guided instance point sampling.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/15d9d3a64151420538e75561d84a3a09ed7d5de5c5fe26b87d97412a90545646.jpg", + "image_caption": [ + "Figure 4. The anchor-guided instance point sampling. Object points are iteratively sampled from the input scan under the guidance of shape anchors. The sampled object points serve as the geometry prior for shape reconstruction." + ], + "image_footnote": [], + "bbox": [ + 78, + 213, + 468, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first add the shape anchors of positive proposals to the original scene scan to enhance indoor objects' structural information. From the enhanced scan, we select points that lie within a given radius of each anchor and update the anchor set with these selected points to expand the searching space. With the searching space compactly fitting the object shape, we can further sample more instance points with high coverage. In this way, our anchor-guided sampling efficiently generates object geometry prior without reliance on an extra segmentation module. In the experiments, the anchor-guided sampling process iterates twice with the sampling radius set to the minimum distance between the anchors.", + "bbox": [ + 75, + 388, + 468, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The geometry priors are next concatenated with the corresponding proposal features $f_{\\mathrm{obj}}^{\\mathrm{vote}}$ and $f_{\\mathrm{obj}}^{\\mathrm{anchor}}$ and encoded into shape embeddings $f_{\\mathrm{obj}}^{\\mathrm{shape}}$ by a ResPointNet [28, 30]. Based on $f_{\\mathrm{obj}}^{\\mathrm{shape}}$ , the decoder of BSP-Net [9] is adopted to predict the signed distances of query points at the canonical coordinates. Following the BSP-Net, we apply the Constructive Solid Geometry (CSG) technique to extract the shape surfaces.", + "bbox": [ + 75, + 588, + 468, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Semantic Scene Reconstruction", + "text_level": 1, + "bbox": [ + 76, + 729, + 354, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We now build the semantic scene model with the anchorm-guided results of the instance detector and shape predictor. The room structure is first constructed by transforming the detected wall quads into orderly connected corners with the merging technique [7]. Next, the models of objects with high objectness are chosen to be arranged in the scene with alignment to the predicted 3D bounding boxes [28]. In the end, the semantic scene model is built as a combination of the reconstructed layout and the scene-aligned object shapes.", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 90, + 633, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Settings", + "text_level": 1, + "bbox": [ + 500, + 114, + 692, + 131 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1 Dataset", + "text_level": 1, + "bbox": [ + 500, + 138, + 607, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We test our method on ScanNetv2 [12] with ground truths from Scan2CAD [1] and SceneCAD datasets [3] for holistic indoor scene understanding. (1) ScanNetv2 is a benchmark for indoor scene analysis with 1,513 scanned room point clouds; (2) Scan2CAD is an alignment dataset that matches ShapeNet models to their counterpart object instances in ScanNet with oriented 3D bounding boxes; and (3) SceneCAD provides 3D layout annotations for the scans in ScanNetv2. We follow [7] and [9] to pre-process the layout polygons and object meshes for network training supervision. The train/test split is kept in line with the previous works with eight challenging object categories considered in the experiments.", + "bbox": [ + 496, + 162, + 890, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.2 Evaluation Metrics", + "text_level": 1, + "bbox": [ + 498, + 371, + 687, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The quality of semantic scene reconstruction is evaluated with the performance of object detection, layout estimation, and scene-aligned shape reconstruction. In line with previous works [28, 7], we use mean average precision across all classes (mAP) with 0.5 IoU threshold for object detection and F1-score for layout estimation. To evaluate the reconstruction quality, we use the chamfer distance (CD) based mAP with thresholds 0.1 and 0.047, light field distance (LFD) based mAP with thresholds 5000 and 2500, as well as the mAP at 3D IoU thresholds 0.25 and 0.5.", + "bbox": [ + 496, + 388, + 890, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.3 Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 547, + 718, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training of our method is conducted on a Titan GPU with two stages, and all parameters are updated by the Adam optimizer. In the first stage, we train the dual-branch instance detector of AncRec for simultaneous optimization of object detection and layout estimation modules. We set the batch size to 8, initialize the learning rate to 1e-3, and adopt the ReducerLrOnPlateu learning scheduler in PyTorch. In the second stage, we use the object proposals predicted by the instance detector for mesh reconstruction. We train the shape predictor with the BSPNet decoder pretrained on the 8-category ShapeNet data [6] for training stability and efficiency. The training process lasts 100 epochs with batch size set to 32, learning rate initialized as 1e-4, and the same training schedule. We found the number of shape anchors $N$ is an insensitive hyper-parameter and thus set it to 18 for computational efficiency.", + "bbox": [ + 496, + 565, + 890, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison and Analysis", + "text_level": 1, + "bbox": [ + 500, + 816, + 733, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scene Reconstruction. We compare our method with the state-of-the-art works, i.e., RfD-Net [28] and DIMR [33]. RfD-Net predicts object shapes with the instance points segmented from proposals given by VoteNet, while", + "bbox": [ + 496, + 840, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21920", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9a9d63bc14a6fefc8c9c0ad840274b479ff52a642bad4c6316dffc0bf9b2e65b.jpg", + "image_caption": [ + "5", + "table", + "sofa" + ], + "image_footnote": [], + "bbox": [ + 120, + 112, + 272, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/244ef087c8e1a8e6490f37fa66405bb095af4e3b30a526c1a202348993c51aea.jpg", + "image_caption": [ + "chair" + ], + "image_footnote": [], + "bbox": [ + 292, + 118, + 416, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/65c39733e6f89e93e8cd31ce4d3ef942fc378e298ec229170563a58ee3209d07.jpg", + "image_caption": [ + "cabinet", + "book", + "bookshelf" + ], + "image_footnote": [], + "bbox": [ + 428, + 122, + 573, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c8cd1e88b6ba4c8a19f35abbaa0e1629191853277d3119d4fd8d5361171a3bd4.jpg", + "image_caption": [ + "display", + "bathtub", + "#" + ], + "image_footnote": [], + "bbox": [ + 606, + 112, + 745, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0d7b1a6a46948fd351510aafb9c90f349af3fb31fee47fbde432d335194d74b2.jpg", + "image_caption": [ + "trash bin" + ], + "image_footnote": [], + "bbox": [ + 756, + 122, + 897, + 213 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/051a5192a1400bea33d7b84646fa08c598e9d390af94c409cff904713aeb197b.jpg", + "image_caption": [ + "Rrnnnne" + ], + "image_footnote": [], + "bbox": [ + 122, + 222, + 267, + 318 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cab921369352c5f501b02ff57407e105bfb8277203377d2bdee2d9f6710b535b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 227, + 413, + 325 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/13c26f9580be58219d55f551355e6160ed756e3bfbdf2a474d1d9c6eee17e283.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 428, + 229, + 575, + 328 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c160c17593b7c733d07ffde23fd2a0fddd6c9d2af4c3189b68adff389d3c8d7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 606, + 218, + 725, + 327 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b7e632ab8419327f73b42f7ca7868338136fe781f7386af1eacb7f8c0c575211.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 234, + 892, + 325 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6a5c945cfe03da0e170de32b30cc25c1039aa17128e5d9906a1b7e03b849450e.jpg", + "image_caption": [ + "DIMR" + ], + "image_footnote": [], + "bbox": [ + 125, + 333, + 271, + 431 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b74a0e32bd68fba9ecf58cd5caa23cf5ef7660e4547e8e6085bcfc3594e7e37e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 339, + 413, + 431 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c1ab0bdc432706fc6f59dd743dc5ead262fce94712b53a3ca67f3b8af222d475.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 339, + 575, + 438 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d3d94a0231fbf452e9aa7a9afe05132540cc7a0d88d1f0c9d66e79a3d8df1094.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 606, + 327, + 745, + 439 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b18b1ed6b027c50697fa2a917e85aba09b1e720366036f318aa57e49af3541f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 347, + 890, + 435 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d7f22aa197240be31258b82429526b114a0bbc565ba5b7ea308c66cfca1665e8.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 122, + 441, + 274, + 542 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/250540929aa94f254436d63792fa6ec9a5d146f629b19de4697749eaff2cdae8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 452, + 413, + 547 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f552f9f5d727282ea51f048220f6c41914ba54956c3db0cfdda01aa48d4bd249.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 455, + 575, + 547 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e91509c395083ed1464fa142d7549cc3522413f945b12f9adc62253c300e040c.jpg", + "image_caption": [ + "Figure 5. Visualization comparison of scene semantic reconstruction on ScanNetv2." + ], + "image_footnote": [], + "bbox": [ + 604, + 443, + 743, + 549 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/29d417f9e7d7f65320d207e6d0b070f7cd462c3b23dc413346ff165bc2246873.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 452, + 901, + 547 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dca2775e00598ce4f3d752f47ff8ae1659f53344114f6035704798f257be2b76.jpg", + "image_caption": [ + "Figure 6. Holistic scene understanding by our proposed AnchorRec. (The comparison methods didn't conduct layout estimation.)" + ], + "image_footnote": [], + "bbox": [ + 124, + 589, + 243, + 694 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7c0a3f1417de627a5f2002560ee33e4e860da3bcc7ef19891b4bb776348adafd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 253, + 589, + 410, + 694 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c23932c9257b599fee521a4d850115784c9c9281598c5d979c6708cc9612a9c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 590, + 547, + 690 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3f4516b8a0b7bc1e7d92eb575b3a8d2c00289db1049ed1fd03dfdc3239de4292.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 589, + 676, + 686 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8f1d612f3e8620f3f856b83312e3303ef2419d005dac866fe2f28b43fea9111f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 598, + 846, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DIMR infers object models from the instance points provided by a complex instance segmentation backbone. As shown in Fig. 5, our AncRec can robustly recognize and reconstruct objects from noisy point clouds with accurate object localization and high-fidelity shape details. With the target-focused proposal feature learning offered by AncLearn in terms of object detection, our method generates fewer false positive models than the compared approaches. With more accurate detection results, our AncRec can leverage outlier-reduced instance priors sampled by AncLearn", + "bbox": [ + 76, + 738, + 472, + 890 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to infer more detailed object structures, e.g., the folding chairs are recovered in the second column of Fig. 5. Tab. 1 presents the quantitative comparison of different methods. Our AncRec outperforms other methods under all metrics. As the comparison methods didn't conduct layout estimation, we illustrate our holistic scene understanding results in Fig. 6. Considering that semantic scene reconstruction is a comprehensive problem, we further analyze the performance of our method on the key sub-tasks, i.e., scene parsing and object reconstruction, as follows.", + "bbox": [ + 496, + 738, + 890, + 888 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21921", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/05fbafc4950b2b8ca7e9789af44baddcbce920c6c06d81f2da0009636d2e92ec.jpg", + "image_caption": [ + "Figure 7. Visualization comparison results of object detection on ScanNetv2" + ], + "image_footnote": [], + "bbox": [ + 125, + 87, + 851, + 366 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c99e359ac9cf89a1befe6d2323d4b47c645a1810b1a123a1d3fc7dc7aab4aa3.jpg", + "image_caption": [ + "Figure 8. The influence of different instance point sampling strategies on object reconstruction. Points from the same instance are painted in the same color. Without the reliance on segmentation like RfD-Net and DIMR, our method can correctly separate adjacent objects and predict shape models." + ], + "image_footnote": [], + "bbox": [ + 116, + 404, + 830, + 661 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4ed1351fb026fbc37315b72783e9463f35b12e148a57f43f6a1d4a5ec4cf4983.jpg", + "table_caption": [ + "Table 1. Quantitative results on the ScanNetv2 dataset. We evaluate the reconstruction quality with mAP at different thresholds. (higher is better)" + ], + "table_footnote": [], + "table_body": "
IoU@0.25IoU@0.5CD@0.1CD@0.047LFD@5000LFD@2500
RfD-Net42.516.945.719.128.67.8
DIMR46.312.551.925.729.58.6
AnchorRec52.918.956.829.430.39.9
", + "bbox": [ + 80, + 767, + 467, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scene Parsing. We compare AncRec with the state-of-the-art methods regarding 8-category object detection and layout estimation. The quantitative results in Tab. 2 show that our AnchorRec edges out the comparison methods by", + "bbox": [ + 76, + 839, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "significant performance gains. Especially for the categories of the display, bathtub, and trash bin, AncRec exceeds the second best by $6.6\\%$ , $5.8\\%$ , and $10.7\\%$ in precision measurement. The results indicate that, with adaptive grouping feature areas, AncLearn works effectively for detecting objects shaped irregularly, e.g., diverse-scaled bathtubs, thin displays, and tiny trash bins. In Tab. 2, we also compare our method with PQ-Transofrmer [7], the current state-of-the-art scene parsing approach, in terms of layout estimation. The results demonstrate that our AncLearn is also applicable to localizing walls that are particularly large and thin.", + "bbox": [ + 496, + 729, + 893, + 897 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21922", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b21f33b7f6a969990b959af9e6bf35ad9b71095beb906cfa32be12bb3de1b8f3.jpg", + "table_caption": [ + "Table 2. Quantitative results of object detection and layout estimation. The AP scores are measured with an IoU threshold of 0.5. Higher values mean better results." + ], + "table_footnote": [], + "table_body": "
ObjectLayout
Methodtablesofachaircabinetbookshelfdisplaybathtubtrash binmAPF1-scores
VoteNet [14]44.547.576.328.820.18.026.014.633.2-
MLCVNet [35]50.546.681.825.519.27.825.319.234.5-
H3DNet [38]53.652.476.025.627.29.420.710.034.4-
BRNet [10]56.443.180.935.426.110.428.211.036.4-
PQ-Transformer [7]52.752.581.132.136.46.720.620.837.867.0
AnchorRec (Ours)57.356.184.536.929.817.034.031.543.470.5
", + "bbox": [ + 80, + 112, + 883, + 256 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4e9e36896b4b53dfb64e6d23e17741e9c664ebb514dc8f59af597c7f3cc6a472.jpg", + "table_caption": [ + "Table 3. Numeric results on object reconstruction. The AP scores are measured with CD at a threshold of 0.1. (Higher is better.)" + ], + "table_footnote": [], + "table_body": "
Methodtablechairbookshelfsofatrash bincabinetdisplaybathtubmAP
RfD-Net [28]25.5182.1132.5344.2144.7428.3765.5142.5745.70
DIMR [33]39.4481.0438.2444.0962.6023.5775.1250.9351.88
AnchorRec (Ours)53.6086.2336.8250.9259.5939.5673.6854.3256.84
", + "bbox": [ + 112, + 281, + 854, + 356 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualized object detection result provided in Fig. 7 further manifests that our AnchorRec can reliably delineate objects with compact bounding box prediction.", + "bbox": [ + 75, + 383, + 468, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Object Mesh Reconstruction. In Tab. 3, we also evaluate the performance of different methods with respect to class-wise object reconstruction. The evaluation is based on how matchable the predicted object meshes are with the scene-aligned ground truths. The numeric results in Tab. 3 show that AncRec achieves the best performance on 5 categories and the overall chamfer distance-based mAP score.", + "bbox": [ + 75, + 429, + 468, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 545, + 241, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AncLearn for Scene Parsing. We study different settings of AncLearn for instance detection and report the results in Tab. 4. Compared to the baseline without AncLearn, the proposed AncLearn brings consistent improvement to object detection and layout estimation. Especially, when cooperating with the self-attention layer for layout estimation, AncLearn enables the dual-branch instance detector of AncRec to obtain the best performance in both parsing tasks.", + "bbox": [ + 75, + 569, + 468, + 705 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2754fb5dfbedf6dc43d4bdea7c5377873a3e81ba066cc1aa5beca15a78ec710d.jpg", + "table_caption": [ + "Table 4. Ablation study of the proposed Anglicen. IoU@0.5 and F1 score respectively reflect the performance of object detection and layout estimation. Higher values mean better results." + ], + "table_footnote": [], + "table_body": "
OD anchorlayout anchorlayout SAIoU@0.5F1 score
41.1260.62
✓43.33(+2.21)60.59(-0.03)
✓40.21(-0.91)62.61(+1.99)
✓✓41.24(+0.12)64.00(+3.38)
✓✓41.16(+0.04)65.34(+4.72)
✓✓✓43.38(+2.26)70.45(+9.83)
", + "bbox": [ + 80, + 750, + 468, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AncLearn for Object Reconstruction. In the shape predictor of AncRec, AncLearn works to sample instance points as geometry priors for reconstruction. With the in", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "stance detector of AncRec, we compare AncLearn with the segmentation-based strategy used in RfD-Net and a box cropping sampling method. The quantitative comparison in Tab. 5 shows that AncLearn outperforms the other two approaches by providing instructive geometry priors with shape anchors in terms of all metrics.", + "bbox": [ + 496, + 383, + 890, + 474 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/89fa5a3dce60b83c0793088ee9c206fab0c933201f0e1d78ca0f1d1ca12b9438.jpg", + "table_caption": [ + "Table 5. Comparison between different instance point sampling methods. Higher values mean better results." + ], + "table_footnote": [], + "table_body": "
IoU@0.5IoU@0.25CD@0.1
Segmentation [28]17.451.455.8
Box Cropping17.751.955.5
AncLearn (Ours)18.952.956.8
CD@0.047LFD@5000LFD@2500
Segmentation [28]29.129.17.8
Box Cropping28.829.99.8
AncLearn (Ours)29.430.39.9
", + "bbox": [ + 504, + 513, + 890, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Contributions of the Vote and Anchor Features", + "text_level": 1, + "bbox": [ + 498, + 665, + 890, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As mentioned in Sec. 3.1.1, we perform instance detection by taking both vote and anchor features into account. To investigate the difference in their contributions to the detection parameters, we visualize the two weight vectors in Eq. (3) for perceptual comparison. As shown in Fig. 9, vote features have a larger impact on scoring objectness, indicating that the context encoded in vote features is useful in differentiating between object and non-object regions. Compared to vote features, anchor features contribute significantly more to angle prediction, which demonstrates the superiority of anchor-guided strategy in providing shape clues for learning shape-sensitive parameters. Considering the gap between the bounding box and the shape of objects, the vote features contribute more to predicting the bounding", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21923", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "box center and size residuals.", + "bbox": [ + 76, + 90, + 274, + 104 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9b28b9bf95f77ff1b8a506ee1fd43357a429e41f3cde3ca003f564ca9e1387d7.jpg", + "image_caption": [ + "Figure 9. The learned weights for combining vote and anchor features." + ], + "image_footnote": [], + "bbox": [ + 78, + 118, + 467, + 190 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.5. The Characteristics of the Learned Anchors", + "text_level": 1, + "bbox": [ + 76, + 250, + 450, + 266 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The supervision with surface points derived from complete 3D models enables the anchors to depict object shapes, even with unobserved structures, which assists in object detection and reconstruction when some parts of objects are missing. There also exists a difference between anchor distribution characteristics for objects and non-object areas. As shown in Fig. 10, the anchors and sampled points depict object shapes, while those for non-object areas scatter irregularly. This characteristic may play a role in differentiating objects from non-object regions.", + "bbox": [ + 75, + 275, + 470, + 426 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/27a52f254a715a3765d6315cb42b56f77eae240d18011cdfe71195be02939a5f.jpg", + "image_caption": [ + "Figure 10. The characteristics of the learned anchors. The anchors and sampled points are shown in red and purple respectively." + ], + "image_footnote": [], + "bbox": [ + 93, + 441, + 145, + 484 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1fde219b68682415596ec503b2e7eea94e5f5bdb6749a7a8a0b082c9326b7186.jpg", + "image_caption": [ + "Object Anchors" + ], + "image_footnote": [], + "bbox": [ + 153, + 441, + 197, + 484 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b864c66883678c816552606d3f21575c14a3f2176ffa97961c88fd0be8ecff3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 441, + 263, + 484 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2bc16df606235930bf5b7f008272ddbc103208e16c249b98001392c3ac4466d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 441, + 323, + 484 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c6ced1631ed69f365482a2da19213fd66670d532ab8f76d52990a6c2ac34a059.jpg", + "image_caption": [ + "Non-object Anchors" + ], + "image_footnote": [], + "bbox": [ + 334, + 441, + 450, + 486 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 76, + 559, + 204, + 575 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we introduce a shape anchor guided learning strategy (AncLearn) that is embedded into a reconstruction-from-detection learning system (AncRec) to handle the issue of noise interference in point-based holistic scene understanding. Extensive experiments demonstrate that AncRec achieves high-quality indoor semantic scene reconstruction. The quantitative and qualitative results show that AncRec outperforms current methods in terms of object detection, layout estimation, and shape modeling. The ablation studies convincingly verify that AncLearn can largely exclude noise from search space for reliable feature grouping and robust instance point sampling. In the future, it is promising to study the application of shape anchor guided learning strategy in other point-based 3D vision tasks.", + "bbox": [ + 75, + 585, + 468, + 811 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 76, + 829, + 227, + 845 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research was supported by the Fundamental Research Funds for the Central Universities of China under Grant 2042022dx0001, NSFC projects under Grant", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "42071370, Wuhan University-Huawei Geoinformatics Innovation Laboratory, and the Open fund of Key Laboratory of Urban Land Resources Monitoring and Simulation, Ministry of Natural Resources under Grant KF202106084.", + "bbox": [ + 498, + 90, + 892, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 501, + 166, + 596, + 181 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 2, 3, 5", + "[2] Armen Avetisyan, Angela Dai, and Matthias Nießner. End-to-end cad model retrieval and 9dof alignment in 3d scans. In Proceedings of the IEEE/CVF International Conference on computer vision, pages 2551-2560, 2019. 3", + "[3] Armen Avetisyan, Tatiana Khanova, Christopher Choy, Denver Dash, Angela Dai, and Matthias Nießner. Scenecad: Predicting object alignments and layouts in rgb-d scans. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 596-612. Springer, 2020. 2, 5", + "[4] Aleksei Bokhovkin and Angela Dai. Neural part priors: Learning to optimize part-based object completion in rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9032-9042, 2023. 3", + "[5] Alexey Bokhovkin, Vladislav Ishimtsev, Emil Bogomolov, Denis Zorin, Alexey Artemov, Evgeny Burnaev, and Angela Dai. Towards part-based understanding of rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7484-7494, 2021. 3", + "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5", + "[7] Xiaoxue Chen, Hao Zhao, Guyue Zhou, and Ya-Qin Zhang. Pq-transformer: Jointly parsing 3d objects and layouts from point clouds. IEEE Robotics and Automation Letters, 7(2):2519-2526, 2022. 4, 5, 7, 8", + "[8] Yujin Chen, Matthias Nießner, and Angela Dai. 4dcontrast: Contrastive learning with dynamic correspondences for 3d scene understanding. In European Conference on Computer Vision, pages 543-560. Springer, 2022. 2", + "[9] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 45-54, 2020. 3, 5", + "[10] Bowen Cheng, Lu Sheng, Shaoshuai Shi, Ming Yang, and Dong Xu. Back-tracing representative points for voting-based 3d object detection in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8963-8972, 2021. 2, 8", + "[11] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In Proceedings of the IEEE/CVF con" + ], + "bbox": [ + 501, + 191, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21924", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on computer vision and pattern recognition, pages 6970-6981, 2020. 3", + "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 5", + "[13] Angela Dai, Christian Diller, and Matthias Nießner. Sg-nn: Sparse generative neural networks for self-supervised scene completion of rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 849–858, 2020. 2", + "[14] Zhipeng Ding, Xu Han, and Marc Niethammer. Volenet: A deep learning label fusion method for multi-atlas segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 202-210. Springer, 2019. 2, 4, 8", + "[15] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9031-9040, 2020. 2", + "[16] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4857-4866, 2020. 3", + "[17] Saurabh Gupta, Pablo Arbeláez, Ross Girshick, and Jitendra Malik. Aligning 3d models to rgb-d images of cluttered scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4731-4740, 2015. 3", + "[18] Xiaoguang Han, Zhaoxuan Zhang, Dong Du, Mingdai Yang, Jingming Yu, Pan Pan, Xin Yang, Ligang Liu, Zixiang Xiong, and Shuguang Cui. Deep reinforcement learning of volume-guided progressive view inpainting for 3d point scene completion from a single depth image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 234-243, 2019. 2", + "[19] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. MeshCNN: a network with an edge. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 3", + "[20] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4421-4430, 2019. 1", + "[21] Ji Hou, Angela Dai, and Matthias Nießner. Revealnet: Seeing behind objects in rgb-d scans. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2095-2104, 2019. 3", + "[22] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15587-15597, 2021. 2", + "[23] Linxi Huan, Xianwei Zheng, and Jianya Gong. Georec: Geometry-enhanced semantic 3d reconstruction of rgb-d in" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "door scenes. ISPRS Journal of Photogrammetry and Remote Sensing, 186:301-314, 2022. 1, 3", + "[24] Hamid Izadinia, Qi Shan, and Steven M Seitz. Im2cad. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5134-5143, 2017. 3", + "[25] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International journal of computer vision, 128:261-318, 2020. 2", + "[26] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 3", + "[27] Yinyu Nie, Xiaoguang Han, Shihui Guo, Yujuan Zheng, Jian Chang, and Jian Jun Zhang. Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 55-64, 2020. 1, 3", + "[28] Yinyu Nie, Ji Hou, Xiaoguang Han, and Matthias Nießner. Rfd-net: Point scene understanding by semantic instance reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4608-4618, 2021. 1, 2, 3, 5, 8", + "[29] Charles R Qi, Xinlei Chen, Or Litany, and Leonidas J Guibas. Imvotenet: Boosting 3d object detection in point clouds with image votes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4404-4413, 2020. 2", + "[30] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 2, 5", + "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 2", + "[32] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. PointCNN: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 2", + "[33] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Point scene understanding via disentangled instance mesh reconstruction. arXiv preprint arXiv:2203.16832, 2022. 1, 2, 3, 5, 8", + "[34] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Dening Lu, Mingqiang Wei, and Jun Wang. Venet: Voting enhancement network for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3712-3721, 2021. 2", + "[35] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Yiming Zhang, Kai Xu, and Jun Wang. Mlcvnet: Multi-level context votenet for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10447-10456, 2020. 2, 4, 8" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21925", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Simon Chen, Yifan Liu, and Chunhua Shen. Towards accurate reconstruction of 3d scene shape from a single monocular image. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):6480-6494, 2022. 2", + "[37] Cheng Zhang, Zhaopeng Cui, Yinda Zhang, Bing Zeng, Marc Pollefeys, and Shuaicheng Liu. Holistic 3d scene understanding from a single image with implicit representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8833-8842, 2021. 1, 3", + "[38] Zaiwei Zhang, Bo Sun, Haitao Yang, and Qixing Huang. H3dnet: 3d object detection using hybrid geometric primitives. In European Conference on Computer Vision, pages 311-329. Springer, 2020. 2, 8", + "[39] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 1", + "[40] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 2023. 2" + ], + "bbox": [ + 78, + 90, + 470, + 401 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21926", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Shape Anchor Guided Holistic Indoor Scene Understanding/4f878220-d6cc-48e2-9d1e-82c1cf350095_model.json b/2023/Shape Anchor Guided Holistic Indoor Scene Understanding/4f878220-d6cc-48e2-9d1e-82c1cf350095_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d8ee256a5bf71c3f4b246356744a6d08b031121b --- /dev/null +++ b/2023/Shape Anchor Guided Holistic Indoor Scene Understanding/4f878220-d6cc-48e2-9d1e-82c1cf350095_model.json @@ -0,0 +1,2422 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.182, + 0.131, + 0.79, + 0.154 + ], + "angle": 0, + "content": "Shape Anchor Guided Holistic Indoor Scene Understanding" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.18, + 0.825, + 0.27 + ], + "angle": 0, + "content": "Mingyue Dong\\(^{1}\\), Linxi Huan\\(^{1}\\), Hanjiang Xiong\\(^{1}\\), Shuhan Shen\\(^{2}\\), Xianwei Zheng\\(^{1*}\\) \n\\(^{1}\\)The State Key Lab. LIESMARS, Wuhan University \n\\(^{2}\\)Institute of Automation, Chinese Academy of Sciences \n{dmy25148, whu.hlx, xionghanjiang, zhengxw}@whu.edu.cn \nshshen@nlpr.iac.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.474, + 0.64 + ], + "angle": 0, + "content": "This paper proposes a shape anchor guided learning strategy (AncLearn) for robust holistic indoor scene understanding. We observe that the search space constructed by current methods for proposal feature grouping and instance point sampling often introduces massive noise to instance detection and mesh reconstruction. Accordingly, we develop AncLearn to generate anchors that dynamically fit instance surfaces to (i) unmix noise and target-related features for offering reliable proposals at the detection stage, and (ii) reduce outliers in object point sampling for directly providing well-structured geometry priors without segmentation during reconstruction. We embed AncLearn into a reconstruction-from-detection learning system (AncRec) to generate high-quality semantic scene models in a purely instance-oriented manner. Experiments conducted on the challenging ScanNetv2 dataset demonstrate that our shape anchor-based method consistently achieves state-of-the-art performance in terms of 3D object detection, layout estimation, and shape reconstruction. The code will be available at https://github.com/Geo-Tell/AncRec." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.655, + 0.21, + 0.672 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Holistic indoor scene understanding from partial observations (e.g., single-view images or 3D scans) is a comprehensive task that provides 3D semantic scene models for indoor applications. Early works studied this task with a reconstruction-from-detection framework that recovers the geometries of room structures and objects from the corresponding detection in a separate way. Later, end-to-end learning methods were proposed to simultaneously perform layout estimation, object detection, and shape prediction in one forward pass for semantic scene reconstruction [27, 37, 23]. With the recent success of point-based 3D detection and instance reconstruction, surging interest has been witnessed in detecting and modeling objects directly" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.304, + 0.891, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.579, + 0.895, + 0.676 + ], + "angle": 0, + "content": "Figure 1. Comparison between different feature grouping strategies. (a) the original scene model. (b)-(e) The different feature grouping operations all suffer from the issue of confusing nontarget noise with useful features. (f) The proposed shape anchor guided grouper directly generates anchors at the instance surface to merge instance-related features, which largely alleviates noise interference." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.702, + 0.725, + 0.717 + ], + "angle": 0, + "content": "from sparse point clouds [28, 33]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Benefiting from the use of rich geometry information, current scan-based deep methods have improved the performance of semantic scene reconstruction. However, they still left two issues that bottleneck high-quality semantic reconstruction: (1) the noisy instance feature learning at the detection phase, and (2) the difficulty in retrieving instances from sparse point clouds for reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "At the detection phase, it is required to group features for instance representation learning. The ball query [30, 31] and 3D convolution[39, 20] are two basic operations for point feature grouping, but they often mix massive noise with informative features due to the fixed grouping range," + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21916" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "as shown in Fig. 1 (b) and (c). To compensate for the deficiency caused by the fixed range, VoteNet [14] and its variants [34, 29, 38, 35] adopt a voting strategy to cluster object features by moving surface points towards object centers. Albeit more flexible than the basic operations, the voting-based strategy often generates an unconstrained grouping area that brings quantities of outliers as illustrated in Fig. 1 (d). BRNet [10] hence restricts the grouping space by sampling around representative points given by coarse box proposals. Nevertheless, limited by box-like grouping area, the sampling points can still fall far beyond targets when the objects are irregularly shaped, as depicted in Fig. 1 (e)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.473, + 0.469 + ], + "angle": 0, + "content": "At the reconstruction stage, the retrieval of outlier-free object points is a prerequisite for object recovery. Due to the noise introduced by feature grouping in the previous detection phase, the points grouped for localizing objects can hardly serve as ideal reconstruction priors as indicated by Fig. 1. Consequently, the current methods are forced to employ an additional foreground classifier [28] or replace the detector with a complex instance segmentation backbone to sample object points from the raw scans [33]. However, the existence of numerous non-target outliers in the search space challenges instance segmentation, resulting in increased risks of gluing different instances and misclassifying background points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.473, + 0.727 + ], + "angle": 0, + "content": "Based on the discussion above, the two noise interference issues during detection and reconstruction are actually highly coupled and can be resolved together as long as outliers are excluded during feature grouping. To this end, we are motivated to propose a shape anchor-guided learning strategy (AncLearn) that generates surface anchors to fit the feature grouping areas to object shape distributions, as displayed in Fig. 1 (f). With the geometry constraint provided by surface anchors, it is able to merge local target-focused features for predicting reliable object proposals and construct a shape-aware search space for robustly sampling instance points without segmentation during reconstruction. The proposed anchor-guided learning strategy can be easily embedded into an end-to-end learning system to accomplish object detection, layout estimation, and instance reconstruction for holistic scene understanding. The main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.732, + 0.47, + 0.793 + ], + "angle": 0, + "content": "- We present a shape anchor guided learning strategy to simultaneously address the issues of noisy feature learning in detection and instance point retrieval during reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.794, + 0.47, + 0.87 + ], + "angle": 0, + "content": "- We embed the proposed anchor-guided learning strategy into an end-to-end learning system to accomplish object detection, layout estimation, and instance reconstruction for holistic scene understanding in a purely instance-oriented way." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that our AncRec achieves high-quality semantic scene reconstruction" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.732, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "with state-of-the-art performance in instance detection and mesh prediction on the challenging ScanNetv2 dataset [12] (some ground truths provided by Scan2CAD [1] and SceneCAD [3])." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.17, + 0.642, + 0.185 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.188, + 0.892, + 0.385 + ], + "angle": 0, + "content": "Semantic scene understanding has been extensively studied over the past years. Many methods focused on acquiring the semantics of 3D scenes [22, 15, 8], while others recovered scene geometries by scene completion [18, 13, 36]. Recently, increasing interest has emerged in semantic scene reconstruction, which recovers both the semantics and geometric shapes of objects. By treating semantic reconstruction as a problem of holistic scene understanding, promising progress has been made based on a reconstruction-from-detection principle [28]. In the following, we review the research from the two core aspects of the reconstruction-from-detection pipeline, i.e., 3D object detection and scene-aligned instance reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.395, + 0.693, + 0.411 + ], + "angle": 0, + "content": "2.1. 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Over the past years, deep detectors have gained great success in 2D object detection [40, 25]. The progress of 2D deep detection has inspired the development of deep learning techniques for recognizing objects from scene point clouds. Compared to image data, point clouds provide rich surface geometry clues for locating objects in real scenes. Nevertheless, the sparse, irregular, and orderless characteristics of point clouds make them hard to be handled by the grid-based convolution model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Works in the early stage leveraged the 2D proposals as 3D detection constraints or projected point clouds into regular 2D/3D grids. Although these 3D detectors are applicable, their performance is still limited by the 2D detectors and the missing geometric details brought by projection. To directly learn the rich geometry features from the raw points, PointNet [30] and PointNet++ [31] used a ball query operation for grouping point features. Later, PointRCNN [32] adopted the point-wise features extracted by PointNet++ for producing instance proposals with point clouds. Hindered by the fixed feature grouping range of grid-based convolutions or ball query operation, these methods often mixed massive noise with informative features, which impairs the reliability of proposals. Considering that the observed object surface points usually lie far away from the object centers, VoteNet [14] introduced a voting mechanism for proposal generation. Based on the voting mechanism, numerous variants further refined proposal features with context relationships [34], hierarchical clues [35], and hybrid geometric primitives [38]. However, the issue of outliers in votes still blocks the learning of representative box features. BRNet [10] derived proposal features based on the generated virtual points given by coarse box proposals," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21917" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.891, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.355, + 0.896, + 0.44 + ], + "angle": 0, + "content": "Figure 2. Overview of the AncRec framework. At the detection stage, seed point features for walls and objects are first learned with a modified dual-branch PointNet++. Following each branch, a voting module and an anchor-guided feature grouper respectively generate proposal features \\( f_{\\mathrm{obj(wall)}}^{\\mathrm{vote}} \\) and \\( f_{\\mathrm{obj(wall)}}^{\\mathrm{anchor}} \\), which are fed into the decoders to predict object bounding boxes and wall quads. The room layout is then constructed by processing the wall quads into connected corners. At the reconstruction stage, objects with high objectness scores are reconstructed under the guidance of \\( f_{\\mathrm{obj}}^{\\mathrm{vote}} \\), \\( f_{\\mathrm{obj}}^{\\mathrm{anchor}} \\), and the geometry priors sampled by shape anchors. Finally, object models are arranged in the scene with the predicted layout according to the spatial alignment provided by predicted bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.451, + 0.471, + 0.512 + ], + "angle": 0, + "content": "whereas the virtual points can fall into non-target areas and bring noise. In this paper, we directly merge features anchored on target surfaces into robust proposal representations for indoor instance learning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.527, + 0.411, + 0.543 + ], + "angle": 0, + "content": "2.2. Scene-aligned Instance Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.703 + ], + "angle": 0, + "content": "Scene-aligned instance reconstruction requires not only modeling 3D object shapes but also correctly arranging the shapes in the 3D scene space. The semantic model of an indoor scene was early constructed with retrieval techniques, which search for CAD shapes or geometric primitives from an offline database and align the approximate object models to input scene data [1, 2, 17, 24]. Albeit able to present delicate scene models, the retrieval-based methods lack generalization ability to various scenes due to the inefficient inference and the limited database scale [37]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The promising advances in deep learnable shape representations motivated scene-aligned instance reconstruction without a finite model pooling [16, 11]. Many prior arts in literature generate 3D objects with explicit or implicit representations with learned features derived from 2D recognition results [19, 27, 37, 23]. Unlike previous single-view modeling methods, RevealNet [21] and the following works [5, 4] predicted the occupancy grids of semantic instances with 3D features extracted from voxelized RGB-D scans. For high scene resolution that is hard to acquire by volumetric representation, RfD-Net [28] and DIMR [33] extracted instance points from scans with segmentation techniques for subsequent single object reconstruction with deep implicit" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.451, + 0.895, + 0.587 + ], + "angle": 0, + "content": "functions [26, 9]. Because of the existence of non-target points, these two state-of-the-art methods have to use an extra foreground classifier or a complex instance segmentation backbone to sample object points for reconstruction. In this paper, we utilized the previously generated surface anchors to localize instance points in scans. Hereby, we introduced an easy but effective anchor-guided sampling strategy to offer geometry priors for the following instance reconstruction without an extra segmentation stage." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.593, + 0.593, + 0.608 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.895, + 0.839 + ], + "angle": 0, + "content": "We illustrate the proposed AncRec in Fig. 2. AncRec achieves semantic scene reconstruction with the proposed anchor-guided learning strategy in an instance-oriented way. It first localizes objects and walls via a dual-branch detector that is equipped with shape anchors learned for grouping point features at the detection stage. The object shape anchors are subsequently leveraged to sample instance geometry priors for predicting object meshes. Finally, the complete semantic scene model is reconstructed by arranging object meshes in the post-processed room layout with alignment to the parsed bounding boxes and poses. In the following, we elaborate on how the shape anchors are learned and leveraged to address the noise interference during instance detection and mesh prediction for high-quality semantic scene reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.851, + 0.864 + ], + "angle": 0, + "content": "3.1. Shape Anchor-guided Instance Detection" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We modify the VoteNet into a dual-branch instance detector to simultaneously localize objects and walls in one" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21918" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.182 + ], + "angle": 0, + "content": "forward pass (Fig. 2). Although the voting mechanism of VoteNet is workable for generating proposals, the reliability of voting-based predictions is often impaired by the outliers in votes. We hence design an anchor-guided feature grouper to learn target-focused features for robust instance detection." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.189, + 0.36, + 0.205 + ], + "angle": 0, + "content": "3.1.1 Anchor-guided Feature Grouper" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.207, + 0.471, + 0.283 + ], + "angle": 0, + "content": "We illustrate the mechanism of the anchor-guided feature grouper in Fig. 3. In the following, by taking the object detection branch as an example, we describe how it works with the voting module in three steps to predict the oriented 3D bounding boxes." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.294, + 0.473, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.455, + 0.471, + 0.539 + ], + "angle": 0, + "content": "Figure 3. The anchor-guided feature grouper. The anchor-guided proposal feature \\( f_{i}^{\\text{anchor}} \\) is the average of features at shape anchors. The shape anchors are deformed from initial anchors \\( \\{a_{j}\\} \\) with offsets derived from the concatenation of the voting-based proposal feature \\( f_{i}^{\\text{vote}} \\) and \\( \\{a_{j}\\} \\). The anchors are then translated to the cluster center \\( c_{i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.675 + ], + "angle": 0, + "content": "Step 1: Anchor Generation. Given the \\(i^{th}\\) voting-based proposal feature \\(\\pmb{f}_i^{\\mathrm{vote}}\\in \\mathbb{R}^{128}\\), the anchor-guided grouper first generates shape anchors that depict the geometry of the \\(i^{th}\\) object candidate via template deformation. With \\(N\\) initial anchors \\(A = \\{\\pmb {a}_j\\in \\mathbb{R}^3 |j = 1,\\dots,N\\}\\) uniformly selected from a unit ball surface, a deformation layer is applied to translate the initial anchors to the \\(i^{th}\\) candidate object surface by" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.685, + 0.47, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {a}} _ {j} = \\boldsymbol {a} _ {j} + \\underbrace {\\operatorname {T a n h} \\left(\\mathrm {M L P} \\left([ \\boldsymbol {a} _ {j} , \\boldsymbol {f} _ {i} ^ {\\text {v o t e}} ]\\right)\\right)} _ {\\text {d e f o r m a t i o n o f t s e t} \\Delta \\boldsymbol {a} _ {j}} + \\boldsymbol {c} _ {i}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.471, + 0.824 + ], + "angle": 0, + "content": "As indicated by Eq. (1), the deformation layer moves \\( \\pmb{a}_j \\) with offsets \\( \\Delta a_j \\) inferred by a multi-layer perceptron (MLP) and a Tanh activation function and then places the deformed anchors at the candidate center \\( \\pmb{c}_i \\) clustered by the voting module. The training of the deformation layer is supervised by the Chamfer distance loss defined as" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.825, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {a n c h o r}} = \\frac {1}{| P _ {\\mathrm {g t}} |} \\sum_ {\\boldsymbol {p} \\in P _ {\\mathrm {g t}}} \\min _ {\\hat {\\boldsymbol {a}} \\in \\hat {A}} \\| \\boldsymbol {p} - \\hat {\\boldsymbol {a}} \\| _ {2} ^ {2} \\tag {2} \\\\ + \\frac {1}{| \\hat {A} |} \\sum_ {\\hat {\\boldsymbol {a}} \\in \\hat {A}} \\min _ {\\boldsymbol {p} \\in P _ {\\mathrm {g t}}} \\| \\hat {\\boldsymbol {a}} - \\boldsymbol {p} \\| _ {2} ^ {2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.227 + ], + "angle": 0, + "content": "where \\(\\hat{A}\\) and \\(P_{\\mathrm{gt}}\\) denote the sets of deformed anchors and surface points sampled from the corresponding scene-aligned 3D meshes. Thereby, the anchors are learned to fit the \\(i^{th}\\) object surface, and the clustering center \\(c_{i}\\) is also refined with shape constraint provided by Eq. (2). Moreover, due to the supervision with complete surface points, the shape anchors can assist in exploiting context for recovering unobserved structures, e.g., the missing object bottom that is supported by the observed floor areas." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.349 + ], + "angle": 0, + "content": "Step 2: Feature Grouping. The anchor-guided grouper propagates the seed point features (extracted by the dual-branch PointNet++ backbone) to each anchor via interpolation followed by an MLP. As the deformed anchors are mainly located on the object surface, the seed point features for propagation can be reliably selected from target-related areas. The \\(i^{th}\\) noise-reduced proposal representation \\(\\pmb{f}_i^{\\mathrm{anchor}}\\in \\mathbb{R}^{128}\\) is obtained by averaging the anchor features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.485 + ], + "angle": 0, + "content": "Step 3: Prediction Fusion. We employ two decoders to predict the object category and oriented box parameters from \\( f_{i}^{\\mathrm{vote}} \\) and \\( f_{i}^{\\mathrm{anchor}} \\) respectively for the \\( i^{th} \\) candidate. This is based on the consideration that the voting-based proposal feature \\( f_{i}^{\\mathrm{vote}} \\) contains more contextual information while the anchor-guided proposal feature \\( f_{i}^{\\mathrm{anchor}} \\) is more target-focused. The estimated parameters, denoted as \\( \\Theta_{i}^{\\mathrm{vote}} \\) and \\( \\Theta_{i}^{\\mathrm{anchor}} \\), are averaged with learnable weights to obtain the final object parameters:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.496, + 0.892, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Theta} _ {i} = \\boldsymbol {w} _ {1} \\cdot \\boldsymbol {\\Theta} _ {i} ^ {\\text {v o t e}} + \\boldsymbol {w} _ {2} \\cdot \\boldsymbol {\\Theta} _ {i} ^ {\\text {a n c h o r}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.526, + 0.892, + 0.646 + ], + "angle": 0, + "content": "From the three steps above, the attributes of objects are parsed with robustness to noise. The anchor-guided detection of wall instances works in a similar way. Considering that walls are generally connected to each other, we additionally deploy the attention operation used in [35] to enhance the anchor-guided wall proposal features with the strong relationship between walls for precise layout estimation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.655, + 0.753, + 0.67 + ], + "angle": 0, + "content": "3.1.2 The Detection Training Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.891, + 0.702 + ], + "angle": 0, + "content": "The total loss function of the anchor-guided instance detector is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.714, + 0.892, + 0.733 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {o b j}} + \\mathcal {L} _ {\\mathrm {w a l l}} + \\mathcal {L} _ {\\mathrm {o b j}} ^ {\\text {a n c h o r}} + \\mathcal {L} _ {\\mathrm {w a l l}} ^ {\\text {a n c h o r}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.744, + 0.892, + 0.821 + ], + "angle": 0, + "content": "In Eq. (4), \\(\\mathcal{L}_{\\mathrm{obj}}\\) is the object detection loss given by [14] while \\(\\mathcal{L}_{\\mathrm{wall}}\\) is the wall quad loss used in [7]. \\(\\mathcal{L}_{\\mathrm{obj}}^{\\mathrm{anchor}}\\) and \\(\\mathcal{L}_{\\mathrm{wall}}^{\\mathrm{anchor}}\\) are the chamfer distance losses that supervise the anchor deformation learning in terms of positive object and wall candidates." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.83, + 0.849 + ], + "angle": 0, + "content": "3.2. Anchor-guided Object Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Instance points from the input scan are desirable geometry priors for reconstruction. Previous works utilized segmentation operations for instance point sampling. Despite" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21919" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "being applicable, accurate instance segmentation is difficult due to the existence of massive non-target noise in the search space. In contrast, we take advantage of the shape anchors to construct a shape-aware search space, in which object points can be directly localized with little noise interference. Fig. 4 illustrates the workflow of our anchor-guided instance point sampling." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.214, + 0.47, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.315, + 0.471, + 0.371 + ], + "angle": 0, + "content": "Figure 4. The anchor-guided instance point sampling. Object points are iteratively sampled from the input scan under the guidance of shape anchors. The sampled object points serve as the geometry prior for shape reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.47, + 0.587 + ], + "angle": 0, + "content": "We first add the shape anchors of positive proposals to the original scene scan to enhance indoor objects' structural information. From the enhanced scan, we select points that lie within a given radius of each anchor and update the anchor set with these selected points to expand the searching space. With the searching space compactly fitting the object shape, we can further sample more instance points with high coverage. In this way, our anchor-guided sampling efficiently generates object geometry prior without reliance on an extra segmentation module. In the experiments, the anchor-guided sampling process iterates twice with the sampling radius set to the minimum distance between the anchors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.47, + 0.718 + ], + "angle": 0, + "content": "The geometry priors are next concatenated with the corresponding proposal features \\( f_{\\mathrm{obj}}^{\\mathrm{vote}} \\) and \\( f_{\\mathrm{obj}}^{\\mathrm{anchor}} \\) and encoded into shape embeddings \\( f_{\\mathrm{obj}}^{\\mathrm{shape}} \\) by a ResPointNet [28, 30]. Based on \\( f_{\\mathrm{obj}}^{\\mathrm{shape}} \\), the decoder of BSP-Net [9] is adopted to predict the signed distances of query points at the canonical coordinates. Following the BSP-Net, we apply the Constructive Solid Geometry (CSG) technique to extract the shape surfaces." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.731, + 0.355, + 0.745 + ], + "angle": 0, + "content": "3.3. Semantic Scene Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We now build the semantic scene model with the anchorm-guided results of the instance detector and shape predictor. The room structure is first constructed by transforming the detected wall quads into orderly connected corners with the merging technique [7]. Next, the models of objects with high objectness are chosen to be arranged in the scene with alignment to the predicted 3D bounding boxes [28]. In the end, the semantic scene model is built as a combination of the reconstructed layout and the scene-aligned object shapes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.634, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.694, + 0.132 + ], + "angle": 0, + "content": "4.1. Experiment Settings" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.139, + 0.608, + 0.153 + ], + "angle": 0, + "content": "4.1.1 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.163, + 0.892, + 0.359 + ], + "angle": 0, + "content": "We test our method on ScanNetv2 [12] with ground truths from Scan2CAD [1] and SceneCAD datasets [3] for holistic indoor scene understanding. (1) ScanNetv2 is a benchmark for indoor scene analysis with 1,513 scanned room point clouds; (2) Scan2CAD is an alignment dataset that matches ShapeNet models to their counterpart object instances in ScanNet with oriented 3D bounding boxes; and (3) SceneCAD provides 3D layout annotations for the scans in ScanNetv2. We follow [7] and [9] to pre-process the layout polygons and object meshes for network training supervision. The train/test split is kept in line with the previous works with eight challenging object categories considered in the experiments." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.372, + 0.689, + 0.386 + ], + "angle": 0, + "content": "4.1.2 Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.892, + 0.542 + ], + "angle": 0, + "content": "The quality of semantic scene reconstruction is evaluated with the performance of object detection, layout estimation, and scene-aligned shape reconstruction. In line with previous works [28, 7], we use mean average precision across all classes (mAP) with 0.5 IoU threshold for object detection and F1-score for layout estimation. To evaluate the reconstruction quality, we use the chamfer distance (CD) based mAP with thresholds 0.1 and 0.047, light field distance (LFD) based mAP with thresholds 5000 and 2500, as well as the mAP at 3D IoU thresholds 0.25 and 0.5." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.548, + 0.719, + 0.563 + ], + "angle": 0, + "content": "4.1.3 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.892, + 0.807 + ], + "angle": 0, + "content": "The training of our method is conducted on a Titan GPU with two stages, and all parameters are updated by the Adam optimizer. In the first stage, we train the dual-branch instance detector of AncRec for simultaneous optimization of object detection and layout estimation modules. We set the batch size to 8, initialize the learning rate to 1e-3, and adopt the ReducerLrOnPlateu learning scheduler in PyTorch. In the second stage, we use the object proposals predicted by the instance detector for mesh reconstruction. We train the shape predictor with the BSPNet decoder pretrained on the 8-category ShapeNet data [6] for training stability and efficiency. The training process lasts 100 epochs with batch size set to 32, learning rate initialized as 1e-4, and the same training schedule. We found the number of shape anchors \\(N\\) is an insensitive hyper-parameter and thus set it to 18 for computational efficiency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.734, + 0.833 + ], + "angle": 0, + "content": "4.2. Comparison and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Scene Reconstruction. We compare our method with the state-of-the-art works, i.e., RfD-Net [28] and DIMR [33]. RfD-Net predicts object shapes with the instance points segmented from proposals given by VoteNet, while" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21920" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.163, + 0.105, + 0.184 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.092, + 0.199, + 0.106 + ], + "angle": 0, + "content": "table" + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.092, + 0.284, + 0.106 + ], + "angle": 0, + "content": "sofa" + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.092, + 0.377, + 0.106 + ], + "angle": 0, + "content": "chair" + }, + { + "type": "image_caption", + "bbox": [ + 0.433, + 0.092, + 0.48, + 0.106 + ], + "angle": 0, + "content": "cabinet" + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.092, + 0.549, + 0.106 + ], + "angle": 0, + "content": "book" + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.092, + 0.582, + 0.106 + ], + "angle": 0, + "content": "bookshelf" + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.092, + 0.657, + 0.107 + ], + "angle": 0, + "content": "display" + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.092, + 0.75, + 0.106 + ], + "angle": 0, + "content": "bathtub" + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.092, + 0.787, + 0.106 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.79, + 0.092, + 0.845, + 0.106 + ], + "angle": 0, + "content": "trash bin" + }, + { + "type": "image", + "bbox": [ + 0.121, + 0.113, + 0.273, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.119, + 0.418, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.123, + 0.574, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.113, + 0.746, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.123, + 0.898, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.25, + 0.105, + 0.293 + ], + "angle": 0, + "content": "Rrnnnne" + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.223, + 0.268, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.228, + 0.415, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.231, + 0.576, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.219, + 0.727, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.235, + 0.893, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.369, + 0.105, + 0.404 + ], + "angle": 0, + "content": "DIMR" + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.334, + 0.272, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.34, + 0.414, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.34, + 0.576, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.328, + 0.746, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.348, + 0.892, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.486, + 0.105, + 0.515 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.442, + 0.275, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.453, + 0.414, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.456, + 0.576, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.444, + 0.745, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.453, + 0.902, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.561, + 0.733, + 0.576 + ], + "angle": 0, + "content": "Figure 5. Visualization comparison of scene semantic reconstruction on ScanNetv2." + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.59, + 0.244, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.59, + 0.411, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.591, + 0.548, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.59, + 0.678, + 0.688 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.599, + 0.848, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.699, + 0.869, + 0.714 + ], + "angle": 0, + "content": "Figure 6. Holistic scene understanding by our proposed AnchorRec. (The comparison methods didn't conduct layout estimation.)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.739, + 0.473, + 0.891 + ], + "angle": 0, + "content": "DIMR infers object models from the instance points provided by a complex instance segmentation backbone. As shown in Fig. 5, our AncRec can robustly recognize and reconstruct objects from noisy point clouds with accurate object localization and high-fidelity shape details. With the target-focused proposal feature learning offered by AncLearn in terms of object detection, our method generates fewer false positive models than the compared approaches. With more accurate detection results, our AncRec can leverage outlier-reduced instance priors sampled by AncLearn" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.739, + 0.892, + 0.89 + ], + "angle": 0, + "content": "to infer more detailed object structures, e.g., the folding chairs are recovered in the second column of Fig. 5. Tab. 1 presents the quantitative comparison of different methods. Our AncRec outperforms other methods under all metrics. As the comparison methods didn't conduct layout estimation, we illustrate our holistic scene understanding results in Fig. 6. Considering that semantic scene reconstruction is a comprehensive problem, we further analyze the performance of our method on the key sub-tasks, i.e., scene parsing and object reconstruction, as follows." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21921" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.126, + 0.088, + 0.852, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.373, + 0.713, + 0.387 + ], + "angle": 0, + "content": "Figure 7. Visualization comparison results of object detection on ScanNetv2" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.405, + 0.831, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.664, + 0.895, + 0.707 + ], + "angle": 0, + "content": "Figure 8. The influence of different instance point sampling strategies on object reconstruction. Points from the same instance are painted in the same color. Without the reliance on segmentation like RfD-Net and DIMR, our method can correctly separate adjacent objects and predict shape models." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.727, + 0.47, + 0.768 + ], + "angle": 0, + "content": "Table 1. Quantitative results on the ScanNetv2 dataset. We evaluate the reconstruction quality with mAP at different thresholds. (higher is better)" + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.768, + 0.468, + 0.816 + ], + "angle": 0, + "content": "
IoU@0.25IoU@0.5CD@0.1CD@0.047LFD@5000LFD@2500
RfD-Net42.516.945.719.128.67.8
DIMR46.312.551.925.729.58.6
AnchorRec52.918.956.829.430.39.9
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Scene Parsing. We compare AncRec with the state-of-the-art methods regarding 8-category object detection and layout estimation. The quantitative results in Tab. 2 show that our AnchorRec edges out the comparison methods by" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.895, + 0.898 + ], + "angle": 0, + "content": "significant performance gains. Especially for the categories of the display, bathtub, and trash bin, AncRec exceeds the second best by \\(6.6\\%\\), \\(5.8\\%\\), and \\(10.7\\%\\) in precision measurement. The results indicate that, with adaptive grouping feature areas, AncLearn works effectively for detecting objects shaped irregularly, e.g., diverse-scaled bathtubs, thin displays, and tiny trash bins. In Tab. 2, we also compare our method with PQ-Transofrmer [7], the current state-of-the-art scene parsing approach, in terms of layout estimation. The results demonstrate that our AncLearn is also applicable to localizing walls that are particularly large and thin." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "21922" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.892, + 0.113 + ], + "angle": 0, + "content": "Table 2. Quantitative results of object detection and layout estimation. The AP scores are measured with an IoU threshold of 0.5. Higher values mean better results." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.113, + 0.885, + 0.257 + ], + "angle": 0, + "content": "
ObjectLayout
Methodtablesofachaircabinetbookshelfdisplaybathtubtrash binmAPF1-scores
VoteNet [14]44.547.576.328.820.18.026.014.633.2-
MLCVNet [35]50.546.681.825.519.27.825.319.234.5-
H3DNet [38]53.652.476.025.627.29.420.710.034.4-
BRNet [10]56.443.180.935.426.110.428.211.036.4-
PQ-Transformer [7]52.752.581.132.136.46.720.620.837.867.0
AnchorRec (Ours)57.356.184.536.929.817.034.031.543.470.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.107, + 0.267, + 0.862, + 0.281 + ], + "angle": 0, + "content": "Table 3. Numeric results on object reconstruction. The AP scores are measured with CD at a threshold of 0.1. (Higher is better.)" + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.282, + 0.855, + 0.357 + ], + "angle": 0, + "content": "
Methodtablechairbookshelfsofatrash bincabinetdisplaybathtubmAP
RfD-Net [28]25.5182.1132.5344.2144.7428.3765.5142.5745.70
DIMR [33]39.4481.0438.2444.0962.6023.5775.1250.9351.88
AnchorRec (Ours)53.6086.2336.8250.9259.5939.5673.6854.3256.84
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.47, + 0.43 + ], + "angle": 0, + "content": "Visualized object detection result provided in Fig. 7 further manifests that our AnchorRec can reliably delineate objects with compact bounding box prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.47, + 0.536 + ], + "angle": 0, + "content": "Object Mesh Reconstruction. In Tab. 3, we also evaluate the performance of different methods with respect to class-wise object reconstruction. The evaluation is based on how matchable the predicted object meshes are with the scene-aligned ground truths. The numeric results in Tab. 3 show that AncRec achieves the best performance on 5 categories and the overall chamfer distance-based mAP score." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.546, + 0.242, + 0.561 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.47, + 0.706 + ], + "angle": 0, + "content": "AncLearn for Scene Parsing. We study different settings of AncLearn for instance detection and report the results in Tab. 4. Compared to the baseline without AncLearn, the proposed AncLearn brings consistent improvement to object detection and layout estimation. Especially, when cooperating with the self-attention layer for layout estimation, AncLearn enables the dual-branch instance detector of AncRec to obtain the best performance in both parsing tasks." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.709, + 0.47, + 0.75 + ], + "angle": 0, + "content": "Table 4. Ablation study of the proposed Anglicen. IoU@0.5 and F1 score respectively reflect the performance of object detection and layout estimation. Higher values mean better results." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.75, + 0.47, + 0.849 + ], + "angle": 0, + "content": "
OD anchorlayout anchorlayout SAIoU@0.5F1 score
41.1260.62
✓43.33(+2.21)60.59(-0.03)
✓40.21(-0.91)62.61(+1.99)
✓✓41.24(+0.12)64.00(+3.38)
✓✓41.16(+0.04)65.34(+4.72)
✓✓✓43.38(+2.26)70.45(+9.83)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "AncLearn for Object Reconstruction. In the shape predictor of AncRec, AncLearn works to sample instance points as geometry priors for reconstruction. With the in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.892, + 0.476 + ], + "angle": 0, + "content": "stance detector of AncRec, we compare AncLearn with the segmentation-based strategy used in RfD-Net and a box cropping sampling method. The quantitative comparison in Tab. 5 shows that AncLearn outperforms the other two approaches by providing instructive geometry priors with shape anchors in terms of all metrics." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.487, + 0.892, + 0.514 + ], + "angle": 0, + "content": "Table 5. Comparison between different instance point sampling methods. Higher values mean better results." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.515, + 0.892, + 0.651 + ], + "angle": 0, + "content": "
IoU@0.5IoU@0.25CD@0.1
Segmentation [28]17.451.455.8
Box Cropping17.751.955.5
AncLearn (Ours)18.952.956.8
CD@0.047LFD@5000LFD@2500
Segmentation [28]29.129.17.8
Box Cropping28.829.99.8
AncLearn (Ours)29.430.39.9
" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.666, + 0.892, + 0.681 + ], + "angle": 0, + "content": "4.4. Contributions of the Vote and Anchor Features" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As mentioned in Sec. 3.1.1, we perform instance detection by taking both vote and anchor features into account. To investigate the difference in their contributions to the detection parameters, we visualize the two weight vectors in Eq. (3) for perceptual comparison. As shown in Fig. 9, vote features have a larger impact on scoring objectness, indicating that the context encoded in vote features is useful in differentiating between object and non-object regions. Compared to vote features, anchor features contribute significantly more to angle prediction, which demonstrates the superiority of anchor-guided strategy in providing shape clues for learning shape-sensitive parameters. Considering the gap between the bounding box and the shape of objects, the vote features contribute more to predicting the bounding" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21923" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.275, + 0.106 + ], + "angle": 0, + "content": "box center and size residuals." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.119, + 0.468, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.195, + 0.47, + 0.222 + ], + "angle": 0, + "content": "Figure 9. The learned weights for combining vote and anchor features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.251, + 0.452, + 0.267 + ], + "angle": 0, + "content": "4.5. The Characteristics of the Learned Anchors" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.471, + 0.427 + ], + "angle": 0, + "content": "The supervision with surface points derived from complete 3D models enables the anchors to depict object shapes, even with unobserved structures, which assists in object detection and reconstruction when some parts of objects are missing. There also exists a difference between anchor distribution characteristics for objects and non-object areas. As shown in Fig. 10, the anchors and sampled points depict object shapes, while those for non-object areas scatter irregularly. This characteristic may play a role in differentiating objects from non-object regions." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.443, + 0.147, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.154, + 0.443, + 0.199, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.443, + 0.264, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.443, + 0.324, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.442, + 0.452, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.491, + 0.242, + 0.5 + ], + "angle": 0, + "content": "Object Anchors" + }, + { + "type": "image_caption", + "bbox": [ + 0.355, + 0.491, + 0.441, + 0.5 + ], + "angle": 0, + "content": "Non-object Anchors" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.501, + 0.468, + 0.529 + ], + "angle": 0, + "content": "Figure 10. The characteristics of the learned anchors. The anchors and sampled points are shown in red and purple respectively." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.56, + 0.205, + 0.576 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.587, + 0.47, + 0.813 + ], + "angle": 0, + "content": "In this paper, we introduce a shape anchor guided learning strategy (AncLearn) that is embedded into a reconstruction-from-detection learning system (AncRec) to handle the issue of noise interference in point-based holistic scene understanding. Extensive experiments demonstrate that AncRec achieves high-quality indoor semantic scene reconstruction. The quantitative and qualitative results show that AncRec outperforms current methods in terms of object detection, layout estimation, and shape modeling. The ablation studies convincingly verify that AncLearn can largely exclude noise from search space for reliable feature grouping and robust instance point sampling. In the future, it is promising to study the application of shape anchor guided learning strategy in other point-based 3D vision tasks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.83, + 0.228, + 0.847 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "This research was supported by the Fundamental Research Funds for the Central Universities of China under Grant 2042022dx0001, NSFC projects under Grant" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.893, + 0.152 + ], + "angle": 0, + "content": "42071370, Wuhan University-Huawei Geoinformatics Innovation Laboratory, and the Open fund of Key Laboratory of Urban Land Resources Monitoring and Simulation, Ministry of Natural Resources under Grant KF202106084." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.167, + 0.597, + 0.182 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.192, + 0.893, + 0.262 + ], + "angle": 0, + "content": "[1] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.263, + 0.894, + 0.319 + ], + "angle": 0, + "content": "[2] Armen Avetisyan, Angela Dai, and Matthias Nießner. End-to-end cad model retrieval and 9dof alignment in 3d scans. In Proceedings of the IEEE/CVF International Conference on computer vision, pages 2551-2560, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.321, + 0.895, + 0.403 + ], + "angle": 0, + "content": "[3] Armen Avetisyan, Tatiana Khanova, Christopher Choy, Denver Dash, Angela Dai, and Matthias Nießner. Scenecad: Predicting object alignments and layouts in rgb-d scans. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 596-612. Springer, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.406, + 0.894, + 0.473 + ], + "angle": 0, + "content": "[4] Aleksei Bokhovkin and Angela Dai. Neural part priors: Learning to optimize part-based object completion in rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9032-9042, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.476, + 0.894, + 0.545 + ], + "angle": 0, + "content": "[5] Alexey Bokhovkin, Vladislav Ishimtsev, Emil Bogomolov, Denis Zorin, Alexey Artemov, Evgeny Burnaev, and Angela Dai. Towards part-based understanding of rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7484-7494, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.547, + 0.894, + 0.615 + ], + "angle": 0, + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.618, + 0.893, + 0.672 + ], + "angle": 0, + "content": "[7] Xiaoxue Chen, Hao Zhao, Guyue Zhou, and Ya-Qin Zhang. Pq-transformer: Jointly parsing 3d objects and layouts from point clouds. IEEE Robotics and Automation Letters, 7(2):2519-2526, 2022. 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.675, + 0.894, + 0.731 + ], + "angle": 0, + "content": "[8] Yujin Chen, Matthias Nießner, and Angela Dai. 4dcontrast: Contrastive learning with dynamic correspondences for 3d scene understanding. In European Conference on Computer Vision, pages 543-560. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.732, + 0.893, + 0.787 + ], + "angle": 0, + "content": "[9] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 45-54, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.789, + 0.894, + 0.857 + ], + "angle": 0, + "content": "[10] Bowen Cheng, Lu Sheng, Shaoshuai Shi, Ming Yang, and Dong Xu. Back-tracing representative points for voting-based 3d object detection in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8963-8972, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.859, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[11] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In Proceedings of the IEEE/CVF con" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.192, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21924" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "ference on computer vision and pattern recognition, pages 6970-6981, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[13] Angela Dai, Christian Diller, and Matthias Nießner. Sg-nn: Sparse generative neural networks for self-supervised scene completion of rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 849–858, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[14] Zhipeng Ding, Xu Han, and Marc Niethammer. Volenet: A deep learning label fusion method for multi-atlas segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 202-210. Springer, 2019. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[15] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9031-9040, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[16] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4857-4866, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.469, + 0.532 + ], + "angle": 0, + "content": "[17] Saurabh Gupta, Pablo Arbeláez, Ross Girshick, and Jitendra Malik. Aligning 3d models to rgb-d images of cluttered scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4731-4740, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[18] Xiaoguang Han, Zhaoxuan Zhang, Dong Du, Mingdai Yang, Jingming Yu, Pan Pan, Xin Yang, Ligang Liu, Zixiang Xiong, and Shuguang Cui. Deep reinforcement learning of volume-guided progressive view inpainting for 3d point scene completion from a single depth image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 234-243, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[19] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. MeshCNN: a network with an edge. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[20] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4421-4430, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[21] Ji Hou, Angela Dai, and Matthias Nießner. Revealnet: Seeing behind objects in rgb-d scans. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2095-2104, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[22] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15587-15597, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[23] Linxi Huan, Xianwei Zheng, and Jianya Gong. Georec: Geometry-enhanced semantic 3d reconstruction of rgb-d in" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "door scenes. ISPRS Journal of Photogrammetry and Remote Sensing, 186:301-314, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[24] Hamid Izadinia, Qi Shan, and Steven M Seitz. Im2cad. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5134-5143, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[25] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International journal of computer vision, 128:261-318, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[26] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[27] Yinyu Nie, Xiaoguang Han, Shihui Guo, Yujuan Zheng, Jian Chang, and Jian Jun Zhang. Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 55-64, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.371, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[28] Yinyu Nie, Ji Hou, Xiaoguang Han, and Matthias Nießner. Rfd-net: Point scene understanding by semantic instance reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4608-4618, 2021. 1, 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[29] Charles R Qi, Xinlei Chen, Or Litany, and Leonidas J Guibas. Imvotenet: Boosting 3d object detection in point clouds with image votes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4404-4413, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[30] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.58, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.636, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[32] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. PointCNN: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.706, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[33] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Point scene understanding via disentangled instance mesh reconstruction. arXiv preprint arXiv:2203.16832, 2022. 1, 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[34] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Dening Lu, Mingqiang Wei, and Jun Wang. Venet: Voting enhancement network for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3712-3721, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[35] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Yiming Zhang, Kai Xu, and Jun Wang. Mlcvnet: Multi-level context votenet for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10447-10456, 2020. 2, 4, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21925" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[36] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Simon Chen, Yifan Liu, and Chunhua Shen. Towards accurate reconstruction of 3d scene shape from a single monocular image. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):6480-6494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.245 + ], + "angle": 0, + "content": "[37] Cheng Zhang, Zhaopeng Cui, Yinda Zhang, Bing Zeng, Marc Pollefeys, and Shuaicheng Liu. Holistic 3d scene understanding from a single image with implicit representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8833-8842, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.248, + 0.47, + 0.302 + ], + "angle": 0, + "content": "[38] Zaiwei Zhang, Bo Sun, Haitao Yang, and Qixing Huang. H3dnet: 3d object detection using hybrid geometric primitives. In European Conference on Computer Vision, pages 311-329. Springer, 2020. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.472, + 0.358 + ], + "angle": 0, + "content": "[39] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.361, + 0.468, + 0.402 + ], + "angle": 0, + "content": "[40] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21926" + } + ] +] \ No newline at end of file diff --git a/2023/Shape Anchor Guided Holistic Indoor Scene Understanding/4f878220-d6cc-48e2-9d1e-82c1cf350095_origin.pdf b/2023/Shape Anchor Guided Holistic Indoor Scene Understanding/4f878220-d6cc-48e2-9d1e-82c1cf350095_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d5adbcdee497e8b8f248634d560892c1aed62af4 --- /dev/null +++ b/2023/Shape Anchor Guided Holistic Indoor Scene Understanding/4f878220-d6cc-48e2-9d1e-82c1cf350095_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fe87de7a4adb697ceae3c3271083495e2aa46d49414b738c985bd9d95a32a9b +size 5014541 diff --git a/2023/Shape Anchor Guided Holistic Indoor Scene Understanding/full.md b/2023/Shape Anchor Guided Holistic Indoor Scene Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e86db29d1fb3805eb2cbbef1be3774037c76e110 --- /dev/null +++ b/2023/Shape Anchor Guided Holistic Indoor Scene Understanding/full.md @@ -0,0 +1,342 @@ +# Shape Anchor Guided Holistic Indoor Scene Understanding + +Mingyue Dong $^{1}$ , Linxi Huan $^{1}$ , Hanjiang Xiong $^{1}$ , Shuhan Shen $^{2}$ , Xianwei Zheng $^{1*}$ $^{1}$ The State Key Lab. LIESMARS, Wuhan University + $^{2}$ Institute of Automation, Chinese Academy of Sciences +{dmy25148, whu.hlx, xionghanjiang, zhengxw}@whu.edu.cn +shshen@nlpr.iac.ac.cn + +# Abstract + +This paper proposes a shape anchor guided learning strategy (AncLearn) for robust holistic indoor scene understanding. We observe that the search space constructed by current methods for proposal feature grouping and instance point sampling often introduces massive noise to instance detection and mesh reconstruction. Accordingly, we develop AncLearn to generate anchors that dynamically fit instance surfaces to (i) unmix noise and target-related features for offering reliable proposals at the detection stage, and (ii) reduce outliers in object point sampling for directly providing well-structured geometry priors without segmentation during reconstruction. We embed AncLearn into a reconstruction-from-detection learning system (AncRec) to generate high-quality semantic scene models in a purely instance-oriented manner. Experiments conducted on the challenging ScanNetv2 dataset demonstrate that our shape anchor-based method consistently achieves state-of-the-art performance in terms of 3D object detection, layout estimation, and shape reconstruction. The code will be available at https://github.com/Geo-Tell/AncRec. + +# 1. Introduction + +Holistic indoor scene understanding from partial observations (e.g., single-view images or 3D scans) is a comprehensive task that provides 3D semantic scene models for indoor applications. Early works studied this task with a reconstruction-from-detection framework that recovers the geometries of room structures and objects from the corresponding detection in a separate way. Later, end-to-end learning methods were proposed to simultaneously perform layout estimation, object detection, and shape prediction in one forward pass for semantic scene reconstruction [27, 37, 23]. With the recent success of point-based 3D detection and instance reconstruction, surging interest has been witnessed in detecting and modeling objects directly + +![](images/ca815fdf223d19a4adf61d75adbbb28b2e18d759a3aa9c5b2e9623b2e1eca591.jpg) +Figure 1. Comparison between different feature grouping strategies. (a) the original scene model. (b)-(e) The different feature grouping operations all suffer from the issue of confusing nontarget noise with useful features. (f) The proposed shape anchor guided grouper directly generates anchors at the instance surface to merge instance-related features, which largely alleviates noise interference. + +from sparse point clouds [28, 33]. + +Benefiting from the use of rich geometry information, current scan-based deep methods have improved the performance of semantic scene reconstruction. However, they still left two issues that bottleneck high-quality semantic reconstruction: (1) the noisy instance feature learning at the detection phase, and (2) the difficulty in retrieving instances from sparse point clouds for reconstruction. + +At the detection phase, it is required to group features for instance representation learning. The ball query [30, 31] and 3D convolution[39, 20] are two basic operations for point feature grouping, but they often mix massive noise with informative features due to the fixed grouping range, + +as shown in Fig. 1 (b) and (c). To compensate for the deficiency caused by the fixed range, VoteNet [14] and its variants [34, 29, 38, 35] adopt a voting strategy to cluster object features by moving surface points towards object centers. Albeit more flexible than the basic operations, the voting-based strategy often generates an unconstrained grouping area that brings quantities of outliers as illustrated in Fig. 1 (d). BRNet [10] hence restricts the grouping space by sampling around representative points given by coarse box proposals. Nevertheless, limited by box-like grouping area, the sampling points can still fall far beyond targets when the objects are irregularly shaped, as depicted in Fig. 1 (e). + +At the reconstruction stage, the retrieval of outlier-free object points is a prerequisite for object recovery. Due to the noise introduced by feature grouping in the previous detection phase, the points grouped for localizing objects can hardly serve as ideal reconstruction priors as indicated by Fig. 1. Consequently, the current methods are forced to employ an additional foreground classifier [28] or replace the detector with a complex instance segmentation backbone to sample object points from the raw scans [33]. However, the existence of numerous non-target outliers in the search space challenges instance segmentation, resulting in increased risks of gluing different instances and misclassifying background points. + +Based on the discussion above, the two noise interference issues during detection and reconstruction are actually highly coupled and can be resolved together as long as outliers are excluded during feature grouping. To this end, we are motivated to propose a shape anchor-guided learning strategy (AncLearn) that generates surface anchors to fit the feature grouping areas to object shape distributions, as displayed in Fig. 1 (f). With the geometry constraint provided by surface anchors, it is able to merge local target-focused features for predicting reliable object proposals and construct a shape-aware search space for robustly sampling instance points without segmentation during reconstruction. The proposed anchor-guided learning strategy can be easily embedded into an end-to-end learning system to accomplish object detection, layout estimation, and instance reconstruction for holistic scene understanding. The main contributions are summarized as follows: + +- We present a shape anchor guided learning strategy to simultaneously address the issues of noisy feature learning in detection and instance point retrieval during reconstruction. +- We embed the proposed anchor-guided learning strategy into an end-to-end learning system to accomplish object detection, layout estimation, and instance reconstruction for holistic scene understanding in a purely instance-oriented way. +- Extensive experiments demonstrate that our AncRec achieves high-quality semantic scene reconstruction + +with state-of-the-art performance in instance detection and mesh prediction on the challenging ScanNetv2 dataset [12] (some ground truths provided by Scan2CAD [1] and SceneCAD [3]). + +# 2. Related Work + +Semantic scene understanding has been extensively studied over the past years. Many methods focused on acquiring the semantics of 3D scenes [22, 15, 8], while others recovered scene geometries by scene completion [18, 13, 36]. Recently, increasing interest has emerged in semantic scene reconstruction, which recovers both the semantics and geometric shapes of objects. By treating semantic reconstruction as a problem of holistic scene understanding, promising progress has been made based on a reconstruction-from-detection principle [28]. In the following, we review the research from the two core aspects of the reconstruction-from-detection pipeline, i.e., 3D object detection and scene-aligned instance reconstruction. + +# 2.1. 3D Object Detection + +Over the past years, deep detectors have gained great success in 2D object detection [40, 25]. The progress of 2D deep detection has inspired the development of deep learning techniques for recognizing objects from scene point clouds. Compared to image data, point clouds provide rich surface geometry clues for locating objects in real scenes. Nevertheless, the sparse, irregular, and orderless characteristics of point clouds make them hard to be handled by the grid-based convolution model. + +Works in the early stage leveraged the 2D proposals as 3D detection constraints or projected point clouds into regular 2D/3D grids. Although these 3D detectors are applicable, their performance is still limited by the 2D detectors and the missing geometric details brought by projection. To directly learn the rich geometry features from the raw points, PointNet [30] and PointNet++ [31] used a ball query operation for grouping point features. Later, PointRCNN [32] adopted the point-wise features extracted by PointNet++ for producing instance proposals with point clouds. Hindered by the fixed feature grouping range of grid-based convolutions or ball query operation, these methods often mixed massive noise with informative features, which impairs the reliability of proposals. Considering that the observed object surface points usually lie far away from the object centers, VoteNet [14] introduced a voting mechanism for proposal generation. Based on the voting mechanism, numerous variants further refined proposal features with context relationships [34], hierarchical clues [35], and hybrid geometric primitives [38]. However, the issue of outliers in votes still blocks the learning of representative box features. BRNet [10] derived proposal features based on the generated virtual points given by coarse box proposals, + +![](images/43e033d447d7f6b71e0c3b5db2bb4fda8a1cfc8ec027ee4a06e2d4c3df9f5b85.jpg) +Figure 2. Overview of the AncRec framework. At the detection stage, seed point features for walls and objects are first learned with a modified dual-branch PointNet++. Following each branch, a voting module and an anchor-guided feature grouper respectively generate proposal features $f_{\mathrm{obj(wall)}}^{\mathrm{vote}}$ and $f_{\mathrm{obj(wall)}}^{\mathrm{anchor}}$ , which are fed into the decoders to predict object bounding boxes and wall quads. The room layout is then constructed by processing the wall quads into connected corners. At the reconstruction stage, objects with high objectness scores are reconstructed under the guidance of $f_{\mathrm{obj}}^{\mathrm{vote}}$ , $f_{\mathrm{obj}}^{\mathrm{anchor}}$ , and the geometry priors sampled by shape anchors. Finally, object models are arranged in the scene with the predicted layout according to the spatial alignment provided by predicted bounding boxes. + +whereas the virtual points can fall into non-target areas and bring noise. In this paper, we directly merge features anchored on target surfaces into robust proposal representations for indoor instance learning. + +# 2.2. Scene-aligned Instance Reconstruction + +Scene-aligned instance reconstruction requires not only modeling 3D object shapes but also correctly arranging the shapes in the 3D scene space. The semantic model of an indoor scene was early constructed with retrieval techniques, which search for CAD shapes or geometric primitives from an offline database and align the approximate object models to input scene data [1, 2, 17, 24]. Albeit able to present delicate scene models, the retrieval-based methods lack generalization ability to various scenes due to the inefficient inference and the limited database scale [37]. + +The promising advances in deep learnable shape representations motivated scene-aligned instance reconstruction without a finite model pooling [16, 11]. Many prior arts in literature generate 3D objects with explicit or implicit representations with learned features derived from 2D recognition results [19, 27, 37, 23]. Unlike previous single-view modeling methods, RevealNet [21] and the following works [5, 4] predicted the occupancy grids of semantic instances with 3D features extracted from voxelized RGB-D scans. For high scene resolution that is hard to acquire by volumetric representation, RfD-Net [28] and DIMR [33] extracted instance points from scans with segmentation techniques for subsequent single object reconstruction with deep implicit + +functions [26, 9]. Because of the existence of non-target points, these two state-of-the-art methods have to use an extra foreground classifier or a complex instance segmentation backbone to sample object points for reconstruction. In this paper, we utilized the previously generated surface anchors to localize instance points in scans. Hereby, we introduced an easy but effective anchor-guided sampling strategy to offer geometry priors for the following instance reconstruction without an extra segmentation stage. + +# 3. Method + +We illustrate the proposed AncRec in Fig. 2. AncRec achieves semantic scene reconstruction with the proposed anchor-guided learning strategy in an instance-oriented way. It first localizes objects and walls via a dual-branch detector that is equipped with shape anchors learned for grouping point features at the detection stage. The object shape anchors are subsequently leveraged to sample instance geometry priors for predicting object meshes. Finally, the complete semantic scene model is reconstructed by arranging object meshes in the post-processed room layout with alignment to the parsed bounding boxes and poses. In the following, we elaborate on how the shape anchors are learned and leveraged to address the noise interference during instance detection and mesh prediction for high-quality semantic scene reconstruction. + +# 3.1. Shape Anchor-guided Instance Detection + +We modify the VoteNet into a dual-branch instance detector to simultaneously localize objects and walls in one + +forward pass (Fig. 2). Although the voting mechanism of VoteNet is workable for generating proposals, the reliability of voting-based predictions is often impaired by the outliers in votes. We hence design an anchor-guided feature grouper to learn target-focused features for robust instance detection. + +# 3.1.1 Anchor-guided Feature Grouper + +We illustrate the mechanism of the anchor-guided feature grouper in Fig. 3. In the following, by taking the object detection branch as an example, we describe how it works with the voting module in three steps to predict the oriented 3D bounding boxes. + +![](images/85f2b401fbee1043fc1cd65dfd1e1c3b144c54a075f84651277304500c978e31.jpg) +Figure 3. The anchor-guided feature grouper. The anchor-guided proposal feature $f_{i}^{\text{anchor}}$ is the average of features at shape anchors. The shape anchors are deformed from initial anchors $\{a_{j}\}$ with offsets derived from the concatenation of the voting-based proposal feature $f_{i}^{\text{vote}}$ and $\{a_{j}\}$ . The anchors are then translated to the cluster center $c_{i}$ . + +Step 1: Anchor Generation. Given the $i^{th}$ voting-based proposal feature $\pmb{f}_i^{\mathrm{vote}}\in \mathbb{R}^{128}$ , the anchor-guided grouper first generates shape anchors that depict the geometry of the $i^{th}$ object candidate via template deformation. With $N$ initial anchors $A = \{\pmb {a}_j\in \mathbb{R}^3 |j = 1,\dots,N\}$ uniformly selected from a unit ball surface, a deformation layer is applied to translate the initial anchors to the $i^{th}$ candidate object surface by + +$$ +\hat {\boldsymbol {a}} _ {j} = \boldsymbol {a} _ {j} + \underbrace {\operatorname {T a n h} \left(\mathrm {M L P} \left([ \boldsymbol {a} _ {j} , \boldsymbol {f} _ {i} ^ {\text {v o t e}} ]\right)\right)} _ {\text {d e f o r m a t i o n o f t s e t} \Delta \boldsymbol {a} _ {j}} + \boldsymbol {c} _ {i}. \tag {1} +$$ + +As indicated by Eq. (1), the deformation layer moves $\pmb{a}_j$ with offsets $\Delta a_j$ inferred by a multi-layer perceptron (MLP) and a Tanh activation function and then places the deformed anchors at the candidate center $\pmb{c}_i$ clustered by the voting module. The training of the deformation layer is supervised by the Chamfer distance loss defined as + +$$ +\begin{array}{l} \mathcal {L} _ {\text {a n c h o r}} = \frac {1}{| P _ {\mathrm {g t}} |} \sum_ {\boldsymbol {p} \in P _ {\mathrm {g t}}} \min _ {\hat {\boldsymbol {a}} \in \hat {A}} \| \boldsymbol {p} - \hat {\boldsymbol {a}} \| _ {2} ^ {2} \tag {2} \\ + \frac {1}{| \hat {A} |} \sum_ {\hat {\boldsymbol {a}} \in \hat {A}} \min _ {\boldsymbol {p} \in P _ {\mathrm {g t}}} \| \hat {\boldsymbol {a}} - \boldsymbol {p} \| _ {2} ^ {2}, \\ \end{array} +$$ + +where $\hat{A}$ and $P_{\mathrm{gt}}$ denote the sets of deformed anchors and surface points sampled from the corresponding scene-aligned 3D meshes. Thereby, the anchors are learned to fit the $i^{th}$ object surface, and the clustering center $c_{i}$ is also refined with shape constraint provided by Eq. (2). Moreover, due to the supervision with complete surface points, the shape anchors can assist in exploiting context for recovering unobserved structures, e.g., the missing object bottom that is supported by the observed floor areas. + +Step 2: Feature Grouping. The anchor-guided grouper propagates the seed point features (extracted by the dual-branch PointNet++ backbone) to each anchor via interpolation followed by an MLP. As the deformed anchors are mainly located on the object surface, the seed point features for propagation can be reliably selected from target-related areas. The $i^{th}$ noise-reduced proposal representation $\pmb{f}_i^{\mathrm{anchor}}\in \mathbb{R}^{128}$ is obtained by averaging the anchor features. + +Step 3: Prediction Fusion. We employ two decoders to predict the object category and oriented box parameters from $f_{i}^{\mathrm{vote}}$ and $f_{i}^{\mathrm{anchor}}$ respectively for the $i^{th}$ candidate. This is based on the consideration that the voting-based proposal feature $f_{i}^{\mathrm{vote}}$ contains more contextual information while the anchor-guided proposal feature $f_{i}^{\mathrm{anchor}}$ is more target-focused. The estimated parameters, denoted as $\Theta_{i}^{\mathrm{vote}}$ and $\Theta_{i}^{\mathrm{anchor}}$ , are averaged with learnable weights to obtain the final object parameters: + +$$ +\boldsymbol {\Theta} _ {i} = \boldsymbol {w} _ {1} \cdot \boldsymbol {\Theta} _ {i} ^ {\text {v o t e}} + \boldsymbol {w} _ {2} \cdot \boldsymbol {\Theta} _ {i} ^ {\text {a n c h o r}}. \tag {3} +$$ + +From the three steps above, the attributes of objects are parsed with robustness to noise. The anchor-guided detection of wall instances works in a similar way. Considering that walls are generally connected to each other, we additionally deploy the attention operation used in [35] to enhance the anchor-guided wall proposal features with the strong relationship between walls for precise layout estimation. + +# 3.1.2 The Detection Training Loss + +The total loss function of the anchor-guided instance detector is defined as + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {o b j}} + \mathcal {L} _ {\mathrm {w a l l}} + \mathcal {L} _ {\mathrm {o b j}} ^ {\text {a n c h o r}} + \mathcal {L} _ {\mathrm {w a l l}} ^ {\text {a n c h o r}}. \tag {4} +$$ + +In Eq. (4), $\mathcal{L}_{\mathrm{obj}}$ is the object detection loss given by [14] while $\mathcal{L}_{\mathrm{wall}}$ is the wall quad loss used in [7]. $\mathcal{L}_{\mathrm{obj}}^{\mathrm{anchor}}$ and $\mathcal{L}_{\mathrm{wall}}^{\mathrm{anchor}}$ are the chamfer distance losses that supervise the anchor deformation learning in terms of positive object and wall candidates. + +# 3.2. Anchor-guided Object Reconstruction + +Instance points from the input scan are desirable geometry priors for reconstruction. Previous works utilized segmentation operations for instance point sampling. Despite + +being applicable, accurate instance segmentation is difficult due to the existence of massive non-target noise in the search space. In contrast, we take advantage of the shape anchors to construct a shape-aware search space, in which object points can be directly localized with little noise interference. Fig. 4 illustrates the workflow of our anchor-guided instance point sampling. + +![](images/15d9d3a64151420538e75561d84a3a09ed7d5de5c5fe26b87d97412a90545646.jpg) +Figure 4. The anchor-guided instance point sampling. Object points are iteratively sampled from the input scan under the guidance of shape anchors. The sampled object points serve as the geometry prior for shape reconstruction. + +We first add the shape anchors of positive proposals to the original scene scan to enhance indoor objects' structural information. From the enhanced scan, we select points that lie within a given radius of each anchor and update the anchor set with these selected points to expand the searching space. With the searching space compactly fitting the object shape, we can further sample more instance points with high coverage. In this way, our anchor-guided sampling efficiently generates object geometry prior without reliance on an extra segmentation module. In the experiments, the anchor-guided sampling process iterates twice with the sampling radius set to the minimum distance between the anchors. + +The geometry priors are next concatenated with the corresponding proposal features $f_{\mathrm{obj}}^{\mathrm{vote}}$ and $f_{\mathrm{obj}}^{\mathrm{anchor}}$ and encoded into shape embeddings $f_{\mathrm{obj}}^{\mathrm{shape}}$ by a ResPointNet [28, 30]. Based on $f_{\mathrm{obj}}^{\mathrm{shape}}$ , the decoder of BSP-Net [9] is adopted to predict the signed distances of query points at the canonical coordinates. Following the BSP-Net, we apply the Constructive Solid Geometry (CSG) technique to extract the shape surfaces. + +# 3.3. Semantic Scene Reconstruction + +We now build the semantic scene model with the anchorm-guided results of the instance detector and shape predictor. The room structure is first constructed by transforming the detected wall quads into orderly connected corners with the merging technique [7]. Next, the models of objects with high objectness are chosen to be arranged in the scene with alignment to the predicted 3D bounding boxes [28]. In the end, the semantic scene model is built as a combination of the reconstructed layout and the scene-aligned object shapes. + +# 4. Experiments + +# 4.1. Experiment Settings + +# 4.1.1 Dataset + +We test our method on ScanNetv2 [12] with ground truths from Scan2CAD [1] and SceneCAD datasets [3] for holistic indoor scene understanding. (1) ScanNetv2 is a benchmark for indoor scene analysis with 1,513 scanned room point clouds; (2) Scan2CAD is an alignment dataset that matches ShapeNet models to their counterpart object instances in ScanNet with oriented 3D bounding boxes; and (3) SceneCAD provides 3D layout annotations for the scans in ScanNetv2. We follow [7] and [9] to pre-process the layout polygons and object meshes for network training supervision. The train/test split is kept in line with the previous works with eight challenging object categories considered in the experiments. + +# 4.1.2 Evaluation Metrics + +The quality of semantic scene reconstruction is evaluated with the performance of object detection, layout estimation, and scene-aligned shape reconstruction. In line with previous works [28, 7], we use mean average precision across all classes (mAP) with 0.5 IoU threshold for object detection and F1-score for layout estimation. To evaluate the reconstruction quality, we use the chamfer distance (CD) based mAP with thresholds 0.1 and 0.047, light field distance (LFD) based mAP with thresholds 5000 and 2500, as well as the mAP at 3D IoU thresholds 0.25 and 0.5. + +# 4.1.3 Implementation Details + +The training of our method is conducted on a Titan GPU with two stages, and all parameters are updated by the Adam optimizer. In the first stage, we train the dual-branch instance detector of AncRec for simultaneous optimization of object detection and layout estimation modules. We set the batch size to 8, initialize the learning rate to 1e-3, and adopt the ReducerLrOnPlateu learning scheduler in PyTorch. In the second stage, we use the object proposals predicted by the instance detector for mesh reconstruction. We train the shape predictor with the BSPNet decoder pretrained on the 8-category ShapeNet data [6] for training stability and efficiency. The training process lasts 100 epochs with batch size set to 32, learning rate initialized as 1e-4, and the same training schedule. We found the number of shape anchors $N$ is an insensitive hyper-parameter and thus set it to 18 for computational efficiency. + +# 4.2. Comparison and Analysis + +Scene Reconstruction. We compare our method with the state-of-the-art works, i.e., RfD-Net [28] and DIMR [33]. RfD-Net predicts object shapes with the instance points segmented from proposals given by VoteNet, while + +![](images/9a9d63bc14a6fefc8c9c0ad840274b479ff52a642bad4c6316dffc0bf9b2e65b.jpg) +5 +table +sofa + +![](images/244ef087c8e1a8e6490f37fa66405bb095af4e3b30a526c1a202348993c51aea.jpg) +chair + +![](images/65c39733e6f89e93e8cd31ce4d3ef942fc378e298ec229170563a58ee3209d07.jpg) +cabinet +book +bookshelf + +![](images/c8cd1e88b6ba4c8a19f35abbaa0e1629191853277d3119d4fd8d5361171a3bd4.jpg) +display +bathtub +# + +![](images/0d7b1a6a46948fd351510aafb9c90f349af3fb31fee47fbde432d335194d74b2.jpg) +trash bin + +![](images/051a5192a1400bea33d7b84646fa08c598e9d390af94c409cff904713aeb197b.jpg) +Rrnnnne + +![](images/cab921369352c5f501b02ff57407e105bfb8277203377d2bdee2d9f6710b535b.jpg) + +![](images/13c26f9580be58219d55f551355e6160ed756e3bfbdf2a474d1d9c6eee17e283.jpg) + +![](images/c160c17593b7c733d07ffde23fd2a0fddd6c9d2af4c3189b68adff389d3c8d7b.jpg) + +![](images/b7e632ab8419327f73b42f7ca7868338136fe781f7386af1eacb7f8c0c575211.jpg) + +![](images/6a5c945cfe03da0e170de32b30cc25c1039aa17128e5d9906a1b7e03b849450e.jpg) +DIMR + +![](images/b74a0e32bd68fba9ecf58cd5caa23cf5ef7660e4547e8e6085bcfc3594e7e37e.jpg) + +![](images/c1ab0bdc432706fc6f59dd743dc5ead262fce94712b53a3ca67f3b8af222d475.jpg) + +![](images/d3d94a0231fbf452e9aa7a9afe05132540cc7a0d88d1f0c9d66e79a3d8df1094.jpg) + +![](images/b18b1ed6b027c50697fa2a917e85aba09b1e720366036f318aa57e49af3541f4.jpg) + +![](images/d7f22aa197240be31258b82429526b114a0bbc565ba5b7ea308c66cfca1665e8.jpg) +Ours + +![](images/250540929aa94f254436d63792fa6ec9a5d146f629b19de4697749eaff2cdae8.jpg) + +![](images/f552f9f5d727282ea51f048220f6c41914ba54956c3db0cfdda01aa48d4bd249.jpg) + +![](images/e91509c395083ed1464fa142d7549cc3522413f945b12f9adc62253c300e040c.jpg) +Figure 5. Visualization comparison of scene semantic reconstruction on ScanNetv2. + +![](images/29d417f9e7d7f65320d207e6d0b070f7cd462c3b23dc413346ff165bc2246873.jpg) + +![](images/dca2775e00598ce4f3d752f47ff8ae1659f53344114f6035704798f257be2b76.jpg) +Figure 6. Holistic scene understanding by our proposed AnchorRec. (The comparison methods didn't conduct layout estimation.) + +![](images/7c0a3f1417de627a5f2002560ee33e4e860da3bcc7ef19891b4bb776348adafd.jpg) + +![](images/c23932c9257b599fee521a4d850115784c9c9281598c5d979c6708cc9612a9c5.jpg) + +![](images/3f4516b8a0b7bc1e7d92eb575b3a8d2c00289db1049ed1fd03dfdc3239de4292.jpg) + +![](images/8f1d612f3e8620f3f856b83312e3303ef2419d005dac866fe2f28b43fea9111f.jpg) + +DIMR infers object models from the instance points provided by a complex instance segmentation backbone. As shown in Fig. 5, our AncRec can robustly recognize and reconstruct objects from noisy point clouds with accurate object localization and high-fidelity shape details. With the target-focused proposal feature learning offered by AncLearn in terms of object detection, our method generates fewer false positive models than the compared approaches. With more accurate detection results, our AncRec can leverage outlier-reduced instance priors sampled by AncLearn + +to infer more detailed object structures, e.g., the folding chairs are recovered in the second column of Fig. 5. Tab. 1 presents the quantitative comparison of different methods. Our AncRec outperforms other methods under all metrics. As the comparison methods didn't conduct layout estimation, we illustrate our holistic scene understanding results in Fig. 6. Considering that semantic scene reconstruction is a comprehensive problem, we further analyze the performance of our method on the key sub-tasks, i.e., scene parsing and object reconstruction, as follows. + +![](images/05fbafc4950b2b8ca7e9789af44baddcbce920c6c06d81f2da0009636d2e92ec.jpg) +Figure 7. Visualization comparison results of object detection on ScanNetv2 + +![](images/9c99e359ac9cf89a1befe6d2323d4b47c645a1810b1a123a1d3fc7dc7aab4aa3.jpg) +Figure 8. The influence of different instance point sampling strategies on object reconstruction. Points from the same instance are painted in the same color. Without the reliance on segmentation like RfD-Net and DIMR, our method can correctly separate adjacent objects and predict shape models. + +Table 1. Quantitative results on the ScanNetv2 dataset. We evaluate the reconstruction quality with mAP at different thresholds. (higher is better) + +
IoU@0.25IoU@0.5CD@0.1CD@0.047LFD@5000LFD@2500
RfD-Net42.516.945.719.128.67.8
DIMR46.312.551.925.729.58.6
AnchorRec52.918.956.829.430.39.9
+ +Scene Parsing. We compare AncRec with the state-of-the-art methods regarding 8-category object detection and layout estimation. The quantitative results in Tab. 2 show that our AnchorRec edges out the comparison methods by + +significant performance gains. Especially for the categories of the display, bathtub, and trash bin, AncRec exceeds the second best by $6.6\%$ , $5.8\%$ , and $10.7\%$ in precision measurement. The results indicate that, with adaptive grouping feature areas, AncLearn works effectively for detecting objects shaped irregularly, e.g., diverse-scaled bathtubs, thin displays, and tiny trash bins. In Tab. 2, we also compare our method with PQ-Transofrmer [7], the current state-of-the-art scene parsing approach, in terms of layout estimation. The results demonstrate that our AncLearn is also applicable to localizing walls that are particularly large and thin. + +Table 2. Quantitative results of object detection and layout estimation. The AP scores are measured with an IoU threshold of 0.5. Higher values mean better results. + +
ObjectLayout
Methodtablesofachaircabinetbookshelfdisplaybathtubtrash binmAPF1-scores
VoteNet [14]44.547.576.328.820.18.026.014.633.2-
MLCVNet [35]50.546.681.825.519.27.825.319.234.5-
H3DNet [38]53.652.476.025.627.29.420.710.034.4-
BRNet [10]56.443.180.935.426.110.428.211.036.4-
PQ-Transformer [7]52.752.581.132.136.46.720.620.837.867.0
AnchorRec (Ours)57.356.184.536.929.817.034.031.543.470.5
+ +Table 3. Numeric results on object reconstruction. The AP scores are measured with CD at a threshold of 0.1. (Higher is better.) + +
Methodtablechairbookshelfsofatrash bincabinetdisplaybathtubmAP
RfD-Net [28]25.5182.1132.5344.2144.7428.3765.5142.5745.70
DIMR [33]39.4481.0438.2444.0962.6023.5775.1250.9351.88
AnchorRec (Ours)53.6086.2336.8250.9259.5939.5673.6854.3256.84
+ +Visualized object detection result provided in Fig. 7 further manifests that our AnchorRec can reliably delineate objects with compact bounding box prediction. + +Object Mesh Reconstruction. In Tab. 3, we also evaluate the performance of different methods with respect to class-wise object reconstruction. The evaluation is based on how matchable the predicted object meshes are with the scene-aligned ground truths. The numeric results in Tab. 3 show that AncRec achieves the best performance on 5 categories and the overall chamfer distance-based mAP score. + +# 4.3. Ablation Studies + +AncLearn for Scene Parsing. We study different settings of AncLearn for instance detection and report the results in Tab. 4. Compared to the baseline without AncLearn, the proposed AncLearn brings consistent improvement to object detection and layout estimation. Especially, when cooperating with the self-attention layer for layout estimation, AncLearn enables the dual-branch instance detector of AncRec to obtain the best performance in both parsing tasks. + +Table 4. Ablation study of the proposed Anglicen. IoU@0.5 and F1 score respectively reflect the performance of object detection and layout estimation. Higher values mean better results. + +
OD anchorlayout anchorlayout SAIoU@0.5F1 score
41.1260.62
✓43.33(+2.21)60.59(-0.03)
✓40.21(-0.91)62.61(+1.99)
✓✓41.24(+0.12)64.00(+3.38)
✓✓41.16(+0.04)65.34(+4.72)
✓✓✓43.38(+2.26)70.45(+9.83)
+ +AncLearn for Object Reconstruction. In the shape predictor of AncRec, AncLearn works to sample instance points as geometry priors for reconstruction. With the in + +stance detector of AncRec, we compare AncLearn with the segmentation-based strategy used in RfD-Net and a box cropping sampling method. The quantitative comparison in Tab. 5 shows that AncLearn outperforms the other two approaches by providing instructive geometry priors with shape anchors in terms of all metrics. + +Table 5. Comparison between different instance point sampling methods. Higher values mean better results. + +
IoU@0.5IoU@0.25CD@0.1
Segmentation [28]17.451.455.8
Box Cropping17.751.955.5
AncLearn (Ours)18.952.956.8
CD@0.047LFD@5000LFD@2500
Segmentation [28]29.129.17.8
Box Cropping28.829.99.8
AncLearn (Ours)29.430.39.9
+ +# 4.4. Contributions of the Vote and Anchor Features + +As mentioned in Sec. 3.1.1, we perform instance detection by taking both vote and anchor features into account. To investigate the difference in their contributions to the detection parameters, we visualize the two weight vectors in Eq. (3) for perceptual comparison. As shown in Fig. 9, vote features have a larger impact on scoring objectness, indicating that the context encoded in vote features is useful in differentiating between object and non-object regions. Compared to vote features, anchor features contribute significantly more to angle prediction, which demonstrates the superiority of anchor-guided strategy in providing shape clues for learning shape-sensitive parameters. Considering the gap between the bounding box and the shape of objects, the vote features contribute more to predicting the bounding + +box center and size residuals. + +![](images/9b28b9bf95f77ff1b8a506ee1fd43357a429e41f3cde3ca003f564ca9e1387d7.jpg) +Figure 9. The learned weights for combining vote and anchor features. + +# 4.5. The Characteristics of the Learned Anchors + +The supervision with surface points derived from complete 3D models enables the anchors to depict object shapes, even with unobserved structures, which assists in object detection and reconstruction when some parts of objects are missing. There also exists a difference between anchor distribution characteristics for objects and non-object areas. As shown in Fig. 10, the anchors and sampled points depict object shapes, while those for non-object areas scatter irregularly. This characteristic may play a role in differentiating objects from non-object regions. + +![](images/27a52f254a715a3765d6315cb42b56f77eae240d18011cdfe71195be02939a5f.jpg) +Figure 10. The characteristics of the learned anchors. The anchors and sampled points are shown in red and purple respectively. + +![](images/1fde219b68682415596ec503b2e7eea94e5f5bdb6749a7a8a0b082c9326b7186.jpg) +Object Anchors + +![](images/b864c66883678c816552606d3f21575c14a3f2176ffa97961c88fd0be8ecff3d.jpg) + +![](images/2bc16df606235930bf5b7f008272ddbc103208e16c249b98001392c3ac4466d9.jpg) + +![](images/c6ced1631ed69f365482a2da19213fd66670d532ab8f76d52990a6c2ac34a059.jpg) +Non-object Anchors + +# 5. Conclusions + +In this paper, we introduce a shape anchor guided learning strategy (AncLearn) that is embedded into a reconstruction-from-detection learning system (AncRec) to handle the issue of noise interference in point-based holistic scene understanding. Extensive experiments demonstrate that AncRec achieves high-quality indoor semantic scene reconstruction. The quantitative and qualitative results show that AncRec outperforms current methods in terms of object detection, layout estimation, and shape modeling. The ablation studies convincingly verify that AncLearn can largely exclude noise from search space for reliable feature grouping and robust instance point sampling. In the future, it is promising to study the application of shape anchor guided learning strategy in other point-based 3D vision tasks. + +# Acknowledgment + +This research was supported by the Fundamental Research Funds for the Central Universities of China under Grant 2042022dx0001, NSFC projects under Grant + +42071370, Wuhan University-Huawei Geoinformatics Innovation Laboratory, and the Open fund of Key Laboratory of Urban Land Resources Monitoring and Simulation, Ministry of Natural Resources under Grant KF202106084. + +# References + +[1] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 2, 3, 5 +[2] Armen Avetisyan, Angela Dai, and Matthias Nießner. End-to-end cad model retrieval and 9dof alignment in 3d scans. In Proceedings of the IEEE/CVF International Conference on computer vision, pages 2551-2560, 2019. 3 +[3] Armen Avetisyan, Tatiana Khanova, Christopher Choy, Denver Dash, Angela Dai, and Matthias Nießner. Scenecad: Predicting object alignments and layouts in rgb-d scans. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 596-612. Springer, 2020. 2, 5 +[4] Aleksei Bokhovkin and Angela Dai. Neural part priors: Learning to optimize part-based object completion in rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9032-9042, 2023. 3 +[5] Alexey Bokhovkin, Vladislav Ishimtsev, Emil Bogomolov, Denis Zorin, Alexey Artemov, Evgeny Burnaev, and Angela Dai. Towards part-based understanding of rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7484-7494, 2021. 3 +[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5 +[7] Xiaoxue Chen, Hao Zhao, Guyue Zhou, and Ya-Qin Zhang. Pq-transformer: Jointly parsing 3d objects and layouts from point clouds. IEEE Robotics and Automation Letters, 7(2):2519-2526, 2022. 4, 5, 7, 8 +[8] Yujin Chen, Matthias Nießner, and Angela Dai. 4dcontrast: Contrastive learning with dynamic correspondences for 3d scene understanding. In European Conference on Computer Vision, pages 543-560. Springer, 2022. 2 +[9] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 45-54, 2020. 3, 5 +[10] Bowen Cheng, Lu Sheng, Shaoshuai Shi, Ming Yang, and Dong Xu. Back-tracing representative points for voting-based 3d object detection in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8963-8972, 2021. 2, 8 +[11] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In Proceedings of the IEEE/CVF con + +ference on computer vision and pattern recognition, pages 6970-6981, 2020. 3 +[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 5 +[13] Angela Dai, Christian Diller, and Matthias Nießner. Sg-nn: Sparse generative neural networks for self-supervised scene completion of rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 849–858, 2020. 2 +[14] Zhipeng Ding, Xu Han, and Marc Niethammer. Volenet: A deep learning label fusion method for multi-atlas segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 202-210. Springer, 2019. 2, 4, 8 +[15] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9031-9040, 2020. 2 +[16] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4857-4866, 2020. 3 +[17] Saurabh Gupta, Pablo Arbeláez, Ross Girshick, and Jitendra Malik. Aligning 3d models to rgb-d images of cluttered scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4731-4740, 2015. 3 +[18] Xiaoguang Han, Zhaoxuan Zhang, Dong Du, Mingdai Yang, Jingming Yu, Pan Pan, Xin Yang, Ligang Liu, Zixiang Xiong, and Shuguang Cui. Deep reinforcement learning of volume-guided progressive view inpainting for 3d point scene completion from a single depth image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 234-243, 2019. 2 +[19] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. MeshCNN: a network with an edge. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 3 +[20] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4421-4430, 2019. 1 +[21] Ji Hou, Angela Dai, and Matthias Nießner. Revealnet: Seeing behind objects in rgb-d scans. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2095-2104, 2019. 3 +[22] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15587-15597, 2021. 2 +[23] Linxi Huan, Xianwei Zheng, and Jianya Gong. Georec: Geometry-enhanced semantic 3d reconstruction of rgb-d in + +door scenes. ISPRS Journal of Photogrammetry and Remote Sensing, 186:301-314, 2022. 1, 3 +[24] Hamid Izadinia, Qi Shan, and Steven M Seitz. Im2cad. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5134-5143, 2017. 3 +[25] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International journal of computer vision, 128:261-318, 2020. 2 +[26] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 3 +[27] Yinyu Nie, Xiaoguang Han, Shihui Guo, Yujuan Zheng, Jian Chang, and Jian Jun Zhang. Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 55-64, 2020. 1, 3 +[28] Yinyu Nie, Ji Hou, Xiaoguang Han, and Matthias Nießner. Rfd-net: Point scene understanding by semantic instance reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4608-4618, 2021. 1, 2, 3, 5, 8 +[29] Charles R Qi, Xinlei Chen, Or Litany, and Leonidas J Guibas. Imvotenet: Boosting 3d object detection in point clouds with image votes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4404-4413, 2020. 2 +[30] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 2, 5 +[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 2 +[32] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. PointCNN: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 2 +[33] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Point scene understanding via disentangled instance mesh reconstruction. arXiv preprint arXiv:2203.16832, 2022. 1, 2, 3, 5, 8 +[34] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Dening Lu, Mingqiang Wei, and Jun Wang. Venet: Voting enhancement network for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3712-3721, 2021. 2 +[35] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Yiming Zhang, Kai Xu, and Jun Wang. Mlcvnet: Multi-level context votenet for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10447-10456, 2020. 2, 4, 8 + +[36] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Simon Chen, Yifan Liu, and Chunhua Shen. Towards accurate reconstruction of 3d scene shape from a single monocular image. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):6480-6494, 2022. 2 +[37] Cheng Zhang, Zhaopeng Cui, Yinda Zhang, Bing Zeng, Marc Pollefeys, and Shuaicheng Liu. Holistic 3d scene understanding from a single image with implicit representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8833-8842, 2021. 1, 3 +[38] Zaiwei Zhang, Bo Sun, Haitao Yang, and Qixing Huang. H3dnet: 3d object detection using hybrid geometric primitives. In European Conference on Computer Vision, pages 311-329. Springer, 2020. 2, 8 +[39] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 1 +[40] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 2023. 2 \ No newline at end of file diff --git a/2023/Shape Anchor Guided Holistic Indoor Scene Understanding/images.zip b/2023/Shape Anchor Guided Holistic Indoor Scene Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..52fa176ba6f9372a137d0202896502a0b3d88f81 --- /dev/null +++ b/2023/Shape Anchor Guided Holistic Indoor Scene Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a805dda5ea21219ab2c8a9ead1c573a963337966d1bdd8d58b2f0164045e2b78 +size 789902 diff --git a/2023/Shape Anchor Guided Holistic Indoor Scene Understanding/layout.json b/2023/Shape Anchor Guided Holistic Indoor Scene Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..58d043890e60ee1fd9c1c96e37a254f92d96e640 --- /dev/null +++ b/2023/Shape Anchor Guided Holistic Indoor Scene Understanding/layout.json @@ -0,0 +1,8729 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 483, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 483, + 121 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 483, + 121 + ], + "type": "text", + "content": "Shape Anchor Guided Holistic Indoor Scene Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 142, + 504, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 142, + 504, + 213 + ], + "spans": [ + { + "bbox": [ + 93, + 142, + 504, + 213 + ], + "type": "text", + "content": "Mingyue Dong" + }, + { + "bbox": [ + 93, + 142, + 504, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 142, + 504, + 213 + ], + "type": "text", + "content": ", Linxi Huan" + }, + { + "bbox": [ + 93, + 142, + 504, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 142, + 504, + 213 + ], + "type": "text", + "content": ", Hanjiang Xiong" + }, + { + "bbox": [ + 93, + 142, + 504, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 142, + 504, + 213 + ], + "type": "text", + "content": ", Shuhan Shen" + }, + { + "bbox": [ + 93, + 142, + 504, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 93, + 142, + 504, + 213 + ], + "type": "text", + "content": ", Xianwei Zheng" + }, + { + "bbox": [ + 93, + 142, + 504, + 213 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 93, + 142, + 504, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 142, + 504, + 213 + ], + "type": "text", + "content": "The State Key Lab. LIESMARS, Wuhan University \n" + }, + { + "bbox": [ + 93, + 142, + 504, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 93, + 142, + 504, + 213 + ], + "type": "text", + "content": "Institute of Automation, Chinese Academy of Sciences \n{dmy25148, whu.hlx, xionghanjiang, zhengxw}@whu.edu.cn \nshshen@nlpr.iac.ac.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 266, + 290, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 506 + ], + "type": "text", + "content": "This paper proposes a shape anchor guided learning strategy (AncLearn) for robust holistic indoor scene understanding. We observe that the search space constructed by current methods for proposal feature grouping and instance point sampling often introduces massive noise to instance detection and mesh reconstruction. Accordingly, we develop AncLearn to generate anchors that dynamically fit instance surfaces to (i) unmix noise and target-related features for offering reliable proposals at the detection stage, and (ii) reduce outliers in object point sampling for directly providing well-structured geometry priors without segmentation during reconstruction. We embed AncLearn into a reconstruction-from-detection learning system (AncRec) to generate high-quality semantic scene models in a purely instance-oriented manner. Experiments conducted on the challenging ScanNetv2 dataset demonstrate that our shape anchor-based method consistently achieves state-of-the-art performance in terms of 3D object detection, layout estimation, and shape reconstruction. The code will be available at https://github.com/Geo-Tell/AncRec." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 518, + 128, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 128, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 128, + 532 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 539, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 696 + ], + "type": "text", + "content": "Holistic indoor scene understanding from partial observations (e.g., single-view images or 3D scans) is a comprehensive task that provides 3D semantic scene models for indoor applications. Early works studied this task with a reconstruction-from-detection framework that recovers the geometries of room structures and objects from the corresponding detection in a separate way. Later, end-to-end learning methods were proposed to simultaneously perform layout estimation, object detection, and shape prediction in one forward pass for semantic scene reconstruction [27, 37, 23]. With the recent success of point-based 3D detection and instance reconstruction, surging interest has been witnessed in detecting and modeling objects directly" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 240, + 545, + 453 + ], + "blocks": [ + { + "bbox": [ + 307, + 240, + 545, + 453 + ], + "lines": [ + { + "bbox": [ + 307, + 240, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 545, + 453 + ], + "type": "image", + "image_path": "ca815fdf223d19a4adf61d75adbbb28b2e18d759a3aa9c5b2e9623b2e1eca591.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 458, + 547, + 535 + ], + "lines": [ + { + "bbox": [ + 304, + 458, + 547, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 547, + 535 + ], + "type": "text", + "content": "Figure 1. Comparison between different feature grouping strategies. (a) the original scene model. (b)-(e) The different feature grouping operations all suffer from the issue of confusing nontarget noise with useful features. (f) The proposed shape anchor guided grouper directly generates anchors at the instance surface to merge instance-related features, which largely alleviates noise interference." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 555, + 443, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 555, + 443, + 567 + ], + "spans": [ + { + "bbox": [ + 306, + 555, + 443, + 567 + ], + "type": "text", + "content": "from sparse point clouds [28, 33]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 568, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 652 + ], + "type": "text", + "content": "Benefiting from the use of rich geometry information, current scan-based deep methods have improved the performance of semantic scene reconstruction. However, they still left two issues that bottleneck high-quality semantic reconstruction: (1) the noisy instance feature learning at the detection phase, and (2) the difficulty in retrieving instances from sparse point clouds for reconstruction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "At the detection phase, it is required to group features for instance representation learning. The ball query [30, 31] and 3D convolution[39, 20] are two basic operations for point feature grouping, but they often mix massive noise with informative features due to the fixed grouping range," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21916" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "as shown in Fig. 1 (b) and (c). To compensate for the deficiency caused by the fixed range, VoteNet [14] and its variants [34, 29, 38, 35] adopt a voting strategy to cluster object features by moving surface points towards object centers. Albeit more flexible than the basic operations, the voting-based strategy often generates an unconstrained grouping area that brings quantities of outliers as illustrated in Fig. 1 (d). BRNet [10] hence restricts the grouping space by sampling around representative points given by coarse box proposals. Nevertheless, limited by box-like grouping area, the sampling points can still fall far beyond targets when the objects are irregularly shaped, as depicted in Fig. 1 (e)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 289, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 371 + ], + "type": "text", + "content": "At the reconstruction stage, the retrieval of outlier-free object points is a prerequisite for object recovery. Due to the noise introduced by feature grouping in the previous detection phase, the points grouped for localizing objects can hardly serve as ideal reconstruction priors as indicated by Fig. 1. Consequently, the current methods are forced to employ an additional foreground classifier [28] or replace the detector with a complex instance segmentation backbone to sample object points from the raw scans [33]. However, the existence of numerous non-target outliers in the search space challenges instance segmentation, resulting in increased risks of gluing different instances and misclassifying background points." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 371, + 289, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 289, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 289, + 575 + ], + "type": "text", + "content": "Based on the discussion above, the two noise interference issues during detection and reconstruction are actually highly coupled and can be resolved together as long as outliers are excluded during feature grouping. To this end, we are motivated to propose a shape anchor-guided learning strategy (AncLearn) that generates surface anchors to fit the feature grouping areas to object shape distributions, as displayed in Fig. 1 (f). With the geometry constraint provided by surface anchors, it is able to merge local target-focused features for predicting reliable object proposals and construct a shape-aware search space for robustly sampling instance points without segmentation during reconstruction. The proposed anchor-guided learning strategy can be easily embedded into an end-to-end learning system to accomplish object detection, layout estimation, and instance reconstruction for holistic scene understanding. The main contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 579, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 58, + 579, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 579, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 58, + 579, + 287, + 628 + ], + "type": "text", + "content": "- We present a shape anchor guided learning strategy to simultaneously address the issues of noisy feature learning in detection and instance point retrieval during reconstruction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 628, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 628, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 58, + 628, + 287, + 689 + ], + "type": "text", + "content": "- We embed the proposed anchor-guided learning strategy into an end-to-end learning system to accomplish object detection, layout estimation, and instance reconstruction for holistic scene understanding in a purely instance-oriented way." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 689, + 287, + 714 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that our AncRec achieves high-quality semantic scene reconstruction" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 325, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 545, + 120 + ], + "type": "text", + "content": "with state-of-the-art performance in instance detection and mesh prediction on the challenging ScanNetv2 dataset [12] (some ground truths provided by Scan2CAD [1] and SceneCAD [3])." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 134, + 392, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 134, + 392, + 146 + ], + "spans": [ + { + "bbox": [ + 306, + 134, + 392, + 146 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 148, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 148, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 545, + 304 + ], + "type": "text", + "content": "Semantic scene understanding has been extensively studied over the past years. Many methods focused on acquiring the semantics of 3D scenes [22, 15, 8], while others recovered scene geometries by scene completion [18, 13, 36]. Recently, increasing interest has emerged in semantic scene reconstruction, which recovers both the semantics and geometric shapes of objects. By treating semantic reconstruction as a problem of holistic scene understanding, promising progress has been made based on a reconstruction-from-detection principle [28]. In the following, we review the research from the two core aspects of the reconstruction-from-detection pipeline, i.e., 3D object detection and scene-aligned instance reconstruction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 312, + 424, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 312, + 424, + 325 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 424, + 325 + ], + "type": "text", + "content": "2.1. 3D Object Detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 331, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 437 + ], + "type": "text", + "content": "Over the past years, deep detectors have gained great success in 2D object detection [40, 25]. The progress of 2D deep detection has inspired the development of deep learning techniques for recognizing objects from scene point clouds. Compared to image data, point clouds provide rich surface geometry clues for locating objects in real scenes. Nevertheless, the sparse, irregular, and orderless characteristics of point clouds make them hard to be handled by the grid-based convolution model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 438, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 715 + ], + "type": "text", + "content": "Works in the early stage leveraged the 2D proposals as 3D detection constraints or projected point clouds into regular 2D/3D grids. Although these 3D detectors are applicable, their performance is still limited by the 2D detectors and the missing geometric details brought by projection. To directly learn the rich geometry features from the raw points, PointNet [30] and PointNet++ [31] used a ball query operation for grouping point features. Later, PointRCNN [32] adopted the point-wise features extracted by PointNet++ for producing instance proposals with point clouds. Hindered by the fixed feature grouping range of grid-based convolutions or ball query operation, these methods often mixed massive noise with informative features, which impairs the reliability of proposals. Considering that the observed object surface points usually lie far away from the object centers, VoteNet [14] introduced a voting mechanism for proposal generation. Based on the voting mechanism, numerous variants further refined proposal features with context relationships [34], hierarchical clues [35], and hybrid geometric primitives [38]. However, the issue of outliers in votes still blocks the learning of representative box features. BRNet [10] derived proposal features based on the generated virtual points given by coarse box proposals," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21917" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 545, + 273 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 545, + 273 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 545, + 273 + ], + "type": "image", + "image_path": "43e033d447d7f6b71e0c3b5db2bb4fda8a1cfc8ec027ee4a06e2d4c3df9f5b85.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 281, + 548, + 348 + ], + "lines": [ + { + "bbox": [ + 46, + 281, + 548, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 548, + 348 + ], + "type": "text", + "content": "Figure 2. Overview of the AncRec framework. At the detection stage, seed point features for walls and objects are first learned with a modified dual-branch PointNet++. Following each branch, a voting module and an anchor-guided feature grouper respectively generate proposal features " + }, + { + "bbox": [ + 46, + 281, + 548, + 348 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{obj(wall)}}^{\\mathrm{vote}}" + }, + { + "bbox": [ + 46, + 281, + 548, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 281, + 548, + 348 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{obj(wall)}}^{\\mathrm{anchor}}" + }, + { + "bbox": [ + 46, + 281, + 548, + 348 + ], + "type": "text", + "content": ", which are fed into the decoders to predict object bounding boxes and wall quads. The room layout is then constructed by processing the wall quads into connected corners. At the reconstruction stage, objects with high objectness scores are reconstructed under the guidance of " + }, + { + "bbox": [ + 46, + 281, + 548, + 348 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{obj}}^{\\mathrm{vote}}" + }, + { + "bbox": [ + 46, + 281, + 548, + 348 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 281, + 548, + 348 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{obj}}^{\\mathrm{anchor}}" + }, + { + "bbox": [ + 46, + 281, + 548, + 348 + ], + "type": "text", + "content": ", and the geometry priors sampled by shape anchors. Finally, object models are arranged in the scene with the predicted layout according to the spatial alignment provided by predicted bounding boxes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 357, + 288, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 288, + 405 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 288, + 405 + ], + "type": "text", + "content": "whereas the virtual points can fall into non-target areas and bring noise. In this paper, we directly merge features anchored on target surfaces into robust proposal representations for indoor instance learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 417, + 251, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 417, + 251, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 417, + 251, + 430 + ], + "type": "text", + "content": "2.2. Scene-aligned Instance Reconstruction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 437, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 556 + ], + "type": "text", + "content": "Scene-aligned instance reconstruction requires not only modeling 3D object shapes but also correctly arranging the shapes in the 3D scene space. The semantic model of an indoor scene was early constructed with retrieval techniques, which search for CAD shapes or geometric primitives from an offline database and align the approximate object models to input scene data [1, 2, 17, 24]. Albeit able to present delicate scene models, the retrieval-based methods lack generalization ability to various scenes due to the inefficient inference and the limited database scale [37]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "The promising advances in deep learnable shape representations motivated scene-aligned instance reconstruction without a finite model pooling [16, 11]. Many prior arts in literature generate 3D objects with explicit or implicit representations with learned features derived from 2D recognition results [19, 27, 37, 23]. Unlike previous single-view modeling methods, RevealNet [21] and the following works [5, 4] predicted the occupancy grids of semantic instances with 3D features extracted from voxelized RGB-D scans. For high scene resolution that is hard to acquire by volumetric representation, RfD-Net [28] and DIMR [33] extracted instance points from scans with segmentation techniques for subsequent single object reconstruction with deep implicit" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 357, + 547, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 547, + 464 + ], + "type": "text", + "content": "functions [26, 9]. Because of the existence of non-target points, these two state-of-the-art methods have to use an extra foreground classifier or a complex instance segmentation backbone to sample object points for reconstruction. In this paper, we utilized the previously generated surface anchors to localize instance points in scans. Hereby, we introduced an easy but effective anchor-guided sampling strategy to offer geometry priors for the following instance reconstruction without an extra segmentation stage." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 469, + 362, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 469, + 362, + 481 + ], + "spans": [ + { + "bbox": [ + 306, + 469, + 362, + 481 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "text", + "content": "We illustrate the proposed AncRec in Fig. 2. AncRec achieves semantic scene reconstruction with the proposed anchor-guided learning strategy in an instance-oriented way. It first localizes objects and walls via a dual-branch detector that is equipped with shape anchors learned for grouping point features at the detection stage. The object shape anchors are subsequently leveraged to sample instance geometry priors for predicting object meshes. Finally, the complete semantic scene model is reconstructed by arranging object meshes in the post-processed room layout with alignment to the parsed bounding boxes and poses. In the following, we elaborate on how the shape anchors are learned and leveraged to address the noise interference during instance detection and mesh prediction for high-quality semantic scene reconstruction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 671, + 520, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 520, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 520, + 684 + ], + "type": "text", + "content": "3.1. Shape Anchor-guided Instance Detection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "type": "text", + "content": "We modify the VoteNet into a dual-branch instance detector to simultaneously localize objects and walls in one" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21918" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": "forward pass (Fig. 2). Although the voting mechanism of VoteNet is workable for generating proposals, the reliability of voting-based predictions is often impaired by the outliers in votes. We hence design an anchor-guided feature grouper to learn target-focused features for robust instance detection." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 149, + 220, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 149, + 220, + 162 + ], + "spans": [ + { + "bbox": [ + 47, + 149, + 220, + 162 + ], + "type": "text", + "content": "3.1.1 Anchor-guided Feature Grouper" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 163, + 288, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 163, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 288, + 224 + ], + "type": "text", + "content": "We illustrate the mechanism of the anchor-guided feature grouper in Fig. 3. In the following, by taking the object detection branch as an example, we describe how it works with the voting module in three steps to predict the oriented 3D bounding boxes." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 47, + 232, + 289, + 348 + ], + "blocks": [ + { + "bbox": [ + 47, + 232, + 289, + 348 + ], + "lines": [ + { + "bbox": [ + 47, + 232, + 289, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 232, + 289, + 348 + ], + "type": "image", + "image_path": "85f2b401fbee1043fc1cd65dfd1e1c3b144c54a075f84651277304500c978e31.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 360, + 288, + 426 + ], + "lines": [ + { + "bbox": [ + 46, + 360, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 288, + 426 + ], + "type": "text", + "content": "Figure 3. The anchor-guided feature grouper. The anchor-guided proposal feature " + }, + { + "bbox": [ + 46, + 360, + 288, + 426 + ], + "type": "inline_equation", + "content": "f_{i}^{\\text{anchor}}" + }, + { + "bbox": [ + 46, + 360, + 288, + 426 + ], + "type": "text", + "content": " is the average of features at shape anchors. The shape anchors are deformed from initial anchors " + }, + { + "bbox": [ + 46, + 360, + 288, + 426 + ], + "type": "inline_equation", + "content": "\\{a_{j}\\}" + }, + { + "bbox": [ + 46, + 360, + 288, + 426 + ], + "type": "text", + "content": " with offsets derived from the concatenation of the voting-based proposal feature " + }, + { + "bbox": [ + 46, + 360, + 288, + 426 + ], + "type": "inline_equation", + "content": "f_{i}^{\\text{vote}}" + }, + { + "bbox": [ + 46, + 360, + 288, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 360, + 288, + 426 + ], + "type": "inline_equation", + "content": "\\{a_{j}\\}" + }, + { + "bbox": [ + 46, + 360, + 288, + 426 + ], + "type": "text", + "content": ". The anchors are then translated to the cluster center " + }, + { + "bbox": [ + 46, + 360, + 288, + 426 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 46, + 360, + 288, + 426 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "text", + "content": "Step 1: Anchor Generation. Given the " + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "text", + "content": " voting-based proposal feature " + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^{\\mathrm{vote}}\\in \\mathbb{R}^{128}" + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "text", + "content": ", the anchor-guided grouper first generates shape anchors that depict the geometry of the " + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "text", + "content": " object candidate via template deformation. With " + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "text", + "content": " initial anchors " + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "inline_equation", + "content": "A = \\{\\pmb {a}_j\\in \\mathbb{R}^3 |j = 1,\\dots,N\\}" + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "text", + "content": " uniformly selected from a unit ball surface, a deformation layer is applied to translate the initial anchors to the " + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "text", + "content": " candidate object surface by" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 542, + 287, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 542, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 83, + 542, + 287, + 572 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {a}} _ {j} = \\boldsymbol {a} _ {j} + \\underbrace {\\operatorname {T a n h} \\left(\\mathrm {M L P} \\left([ \\boldsymbol {a} _ {j} , \\boldsymbol {f} _ {i} ^ {\\text {v o t e}} ]\\right)\\right)} _ {\\text {d e f o r m a t i o n o f t s e t} \\Delta \\boldsymbol {a} _ {j}} + \\boldsymbol {c} _ {i}. \\tag {1}", + "image_path": "be982db9a6ea1455ac98d40dc966199f40d1c5c30e4de0d1aa64e71973cadc95.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 580, + 288, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 288, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 288, + 652 + ], + "type": "text", + "content": "As indicated by Eq. (1), the deformation layer moves " + }, + { + "bbox": [ + 46, + 580, + 288, + 652 + ], + "type": "inline_equation", + "content": "\\pmb{a}_j" + }, + { + "bbox": [ + 46, + 580, + 288, + 652 + ], + "type": "text", + "content": " with offsets " + }, + { + "bbox": [ + 46, + 580, + 288, + 652 + ], + "type": "inline_equation", + "content": "\\Delta a_j" + }, + { + "bbox": [ + 46, + 580, + 288, + 652 + ], + "type": "text", + "content": " inferred by a multi-layer perceptron (MLP) and a Tanh activation function and then places the deformed anchors at the candidate center " + }, + { + "bbox": [ + 46, + 580, + 288, + 652 + ], + "type": "inline_equation", + "content": "\\pmb{c}_i" + }, + { + "bbox": [ + 46, + 580, + 288, + 652 + ], + "type": "text", + "content": " clustered by the voting module. The training of the deformation layer is supervised by the Chamfer distance loss defined as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 89, + 653, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 89, + 653, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {a n c h o r}} = \\frac {1}{| P _ {\\mathrm {g t}} |} \\sum_ {\\boldsymbol {p} \\in P _ {\\mathrm {g t}}} \\min _ {\\hat {\\boldsymbol {a}} \\in \\hat {A}} \\| \\boldsymbol {p} - \\hat {\\boldsymbol {a}} \\| _ {2} ^ {2} \\tag {2} \\\\ + \\frac {1}{| \\hat {A} |} \\sum_ {\\hat {\\boldsymbol {a}} \\in \\hat {A}} \\min _ {\\boldsymbol {p} \\in P _ {\\mathrm {g t}}} \\| \\hat {\\boldsymbol {a}} - \\boldsymbol {p} \\| _ {2} ^ {2}, \\\\ \\end{array}", + "image_path": "3b04fe6b7a54873dea2b62eb98586417e3d0b53fe267cd36b281683a08ab4480.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\hat{A}" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{gt}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " denote the sets of deformed anchors and surface points sampled from the corresponding scene-aligned 3D meshes. Thereby, the anchors are learned to fit the " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " object surface, and the clustering center " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " is also refined with shape constraint provided by Eq. (2). Moreover, due to the supervision with complete surface points, the shape anchors can assist in exploiting context for recovering unobserved structures, e.g., the missing object bottom that is supported by the observed floor areas." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "type": "text", + "content": "Step 2: Feature Grouping. The anchor-guided grouper propagates the seed point features (extracted by the dual-branch PointNet++ backbone) to each anchor via interpolation followed by an MLP. As the deformed anchors are mainly located on the object surface, the seed point features for propagation can be reliably selected from target-related areas. The " + }, + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "type": "text", + "content": " noise-reduced proposal representation " + }, + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^{\\mathrm{anchor}}\\in \\mathbb{R}^{128}" + }, + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "type": "text", + "content": " is obtained by averaging the anchor features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 277, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 384 + ], + "type": "text", + "content": "Step 3: Prediction Fusion. We employ two decoders to predict the object category and oriented box parameters from " + }, + { + "bbox": [ + 304, + 277, + 545, + 384 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathrm{vote}}" + }, + { + "bbox": [ + 304, + 277, + 545, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 277, + 545, + 384 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathrm{anchor}}" + }, + { + "bbox": [ + 304, + 277, + 545, + 384 + ], + "type": "text", + "content": " respectively for the " + }, + { + "bbox": [ + 304, + 277, + 545, + 384 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 277, + 545, + 384 + ], + "type": "text", + "content": " candidate. This is based on the consideration that the voting-based proposal feature " + }, + { + "bbox": [ + 304, + 277, + 545, + 384 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathrm{vote}}" + }, + { + "bbox": [ + 304, + 277, + 545, + 384 + ], + "type": "text", + "content": " contains more contextual information while the anchor-guided proposal feature " + }, + { + "bbox": [ + 304, + 277, + 545, + 384 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathrm{anchor}}" + }, + { + "bbox": [ + 304, + 277, + 545, + 384 + ], + "type": "text", + "content": " is more target-focused. The estimated parameters, denoted as " + }, + { + "bbox": [ + 304, + 277, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\Theta_{i}^{\\mathrm{vote}}" + }, + { + "bbox": [ + 304, + 277, + 545, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 277, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\Theta_{i}^{\\mathrm{anchor}}" + }, + { + "bbox": [ + 304, + 277, + 545, + 384 + ], + "type": "text", + "content": ", are averaged with learnable weights to obtain the final object parameters:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 358, + 392, + 545, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 392, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 358, + 392, + 545, + 407 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Theta} _ {i} = \\boldsymbol {w} _ {1} \\cdot \\boldsymbol {\\Theta} _ {i} ^ {\\text {v o t e}} + \\boldsymbol {w} _ {2} \\cdot \\boldsymbol {\\Theta} _ {i} ^ {\\text {a n c h o r}}. \\tag {3}", + "image_path": "98817579537b1db3ce6dd0c99c175f3cfb059a2b00980d240148e6553ea27440.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 416, + 545, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 416, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 416, + 545, + 511 + ], + "type": "text", + "content": "From the three steps above, the attributes of objects are parsed with robustness to noise. The anchor-guided detection of wall instances works in a similar way. Considering that walls are generally connected to each other, we additionally deploy the attention operation used in [35] to enhance the anchor-guided wall proposal features with the strong relationship between walls for precise layout estimation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 518, + 460, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 518, + 460, + 530 + ], + "spans": [ + { + "bbox": [ + 305, + 518, + 460, + 530 + ], + "type": "text", + "content": "3.1.2 The Detection Training Loss" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 533, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 555 + ], + "type": "text", + "content": "The total loss function of the anchor-guided instance detector is defined as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 350, + 565, + 545, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 565, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 350, + 565, + 545, + 580 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {o b j}} + \\mathcal {L} _ {\\mathrm {w a l l}} + \\mathcal {L} _ {\\mathrm {o b j}} ^ {\\text {a n c h o r}} + \\mathcal {L} _ {\\mathrm {w a l l}} ^ {\\text {a n c h o r}}. \\tag {4}", + "image_path": "a7bac8e4d0e377b23474ce7976384e135b6d176bbd921a6e95de934496603386.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 589, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 589, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 589, + 545, + 650 + ], + "type": "text", + "content": "In Eq. (4), " + }, + { + "bbox": [ + 304, + 589, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 304, + 589, + 545, + 650 + ], + "type": "text", + "content": " is the object detection loss given by [14] while " + }, + { + "bbox": [ + 304, + 589, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{wall}}" + }, + { + "bbox": [ + 304, + 589, + 545, + 650 + ], + "type": "text", + "content": " is the wall quad loss used in [7]. " + }, + { + "bbox": [ + 304, + 589, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{obj}}^{\\mathrm{anchor}}" + }, + { + "bbox": [ + 304, + 589, + 545, + 650 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 589, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{wall}}^{\\mathrm{anchor}}" + }, + { + "bbox": [ + 304, + 589, + 545, + 650 + ], + "type": "text", + "content": " are the chamfer distance losses that supervise the anchor deformation learning in terms of positive object and wall candidates." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 658, + 507, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 507, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 507, + 672 + ], + "type": "text", + "content": "3.2. Anchor-guided Object Reconstruction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "Instance points from the input scan are desirable geometry priors for reconstruction. Previous works utilized segmentation operations for instance point sampling. Despite" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21919" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "being applicable, accurate instance segmentation is difficult due to the existence of massive non-target noise in the search space. In contrast, we take advantage of the shape anchors to construct a shape-aware search space, in which object points can be directly localized with little noise interference. Fig. 4 illustrates the workflow of our anchor-guided instance point sampling." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 169, + 287, + 242 + ], + "blocks": [ + { + "bbox": [ + 48, + 169, + 287, + 242 + ], + "lines": [ + { + "bbox": [ + 48, + 169, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 48, + 169, + 287, + 242 + ], + "type": "image", + "image_path": "15d9d3a64151420538e75561d84a3a09ed7d5de5c5fe26b87d97412a90545646.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 249, + 288, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 249, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 288, + 293 + ], + "type": "text", + "content": "Figure 4. The anchor-guided instance point sampling. Object points are iteratively sampled from the input scan under the guidance of shape anchors. The sampled object points serve as the geometry prior for shape reconstruction." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 308, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 464 + ], + "type": "text", + "content": "We first add the shape anchors of positive proposals to the original scene scan to enhance indoor objects' structural information. From the enhanced scan, we select points that lie within a given radius of each anchor and update the anchor set with these selected points to expand the searching space. With the searching space compactly fitting the object shape, we can further sample more instance points with high coverage. In this way, our anchor-guided sampling efficiently generates object geometry prior without reliance on an extra segmentation module. In the experiments, the anchor-guided sampling process iterates twice with the sampling radius set to the minimum distance between the anchors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 466, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 287, + 568 + ], + "type": "text", + "content": "The geometry priors are next concatenated with the corresponding proposal features " + }, + { + "bbox": [ + 46, + 466, + 287, + 568 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{obj}}^{\\mathrm{vote}}" + }, + { + "bbox": [ + 46, + 466, + 287, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 466, + 287, + 568 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{obj}}^{\\mathrm{anchor}}" + }, + { + "bbox": [ + 46, + 466, + 287, + 568 + ], + "type": "text", + "content": " and encoded into shape embeddings " + }, + { + "bbox": [ + 46, + 466, + 287, + 568 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{obj}}^{\\mathrm{shape}}" + }, + { + "bbox": [ + 46, + 466, + 287, + 568 + ], + "type": "text", + "content": " by a ResPointNet [28, 30]. Based on " + }, + { + "bbox": [ + 46, + 466, + 287, + 568 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{obj}}^{\\mathrm{shape}}" + }, + { + "bbox": [ + 46, + 466, + 287, + 568 + ], + "type": "text", + "content": ", the decoder of BSP-Net [9] is adopted to predict the signed distances of query points at the canonical coordinates. Following the BSP-Net, we apply the Constructive Solid Geometry (CSG) technique to extract the shape surfaces." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 578, + 217, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 217, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 217, + 590 + ], + "type": "text", + "content": "3.3. Semantic Scene Reconstruction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "We now build the semantic scene model with the anchorm-guided results of the instance detector and shape predictor. The room structure is first constructed by transforming the detected wall quads into orderly connected corners with the merging technique [7]. Next, the models of objects with high objectness are chosen to be arranged in the scene with alignment to the predicted 3D bounding boxes [28]. In the end, the semantic scene model is built as a combination of the reconstructed layout and the scene-aligned object shapes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 91, + 424, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 424, + 104 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 424, + 104 + ], + "type": "text", + "content": "4.1. Experiment Settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 110, + 372, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 110, + 372, + 121 + ], + "spans": [ + { + "bbox": [ + 306, + 110, + 372, + 121 + ], + "type": "text", + "content": "4.1.1 Dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 129, + 545, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 129, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 304, + 129, + 545, + 284 + ], + "type": "text", + "content": "We test our method on ScanNetv2 [12] with ground truths from Scan2CAD [1] and SceneCAD datasets [3] for holistic indoor scene understanding. (1) ScanNetv2 is a benchmark for indoor scene analysis with 1,513 scanned room point clouds; (2) Scan2CAD is an alignment dataset that matches ShapeNet models to their counterpart object instances in ScanNet with oriented 3D bounding boxes; and (3) SceneCAD provides 3D layout annotations for the scans in ScanNetv2. We follow [7] and [9] to pre-process the layout polygons and object meshes for network training supervision. The train/test split is kept in line with the previous works with eight challenging object categories considered in the experiments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 294, + 421, + 305 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 294, + 421, + 305 + ], + "spans": [ + { + "bbox": [ + 305, + 294, + 421, + 305 + ], + "type": "text", + "content": "4.1.2 Evaluation Metrics" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 429 + ], + "type": "text", + "content": "The quality of semantic scene reconstruction is evaluated with the performance of object detection, layout estimation, and scene-aligned shape reconstruction. In line with previous works [28, 7], we use mean average precision across all classes (mAP) with 0.5 IoU threshold for object detection and F1-score for layout estimation. To evaluate the reconstruction quality, we use the chamfer distance (CD) based mAP with thresholds 0.1 and 0.047, light field distance (LFD) based mAP with thresholds 5000 and 2500, as well as the mAP at 3D IoU thresholds 0.25 and 0.5." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 434, + 440, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 434, + 440, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 434, + 440, + 445 + ], + "type": "text", + "content": "4.1.3 Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 448, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 545, + 639 + ], + "type": "text", + "content": "The training of our method is conducted on a Titan GPU with two stages, and all parameters are updated by the Adam optimizer. In the first stage, we train the dual-branch instance detector of AncRec for simultaneous optimization of object detection and layout estimation modules. We set the batch size to 8, initialize the learning rate to 1e-3, and adopt the ReducerLrOnPlateu learning scheduler in PyTorch. In the second stage, we use the object proposals predicted by the instance detector for mesh reconstruction. We train the shape predictor with the BSPNet decoder pretrained on the 8-category ShapeNet data [6] for training stability and efficiency. The training process lasts 100 epochs with batch size set to 32, learning rate initialized as 1e-4, and the same training schedule. We found the number of shape anchors " + }, + { + "bbox": [ + 304, + 448, + 545, + 639 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 448, + 545, + 639 + ], + "type": "text", + "content": " is an insensitive hyper-parameter and thus set it to 18 for computational efficiency." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 647, + 449, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 449, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 449, + 659 + ], + "type": "text", + "content": "4.2. Comparison and Analysis" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "text", + "content": "Scene Reconstruction. We compare our method with the state-of-the-art works, i.e., RfD-Net [28] and DIMR [33]. RfD-Net predicts object shapes with the instance points segmented from proposals given by VoteNet, while" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21920" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 89, + 167, + 171 + ], + "blocks": [ + { + "bbox": [ + 51, + 129, + 64, + 145 + ], + "lines": [ + { + "bbox": [ + 51, + 129, + 64, + 145 + ], + "spans": [ + { + "bbox": [ + 51, + 129, + 64, + 145 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 72, + 121, + 83 + ], + "lines": [ + { + "bbox": [ + 85, + 72, + 121, + 83 + ], + "spans": [ + { + "bbox": [ + 85, + 72, + 121, + 83 + ], + "type": "text", + "content": "table" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 140, + 72, + 173, + 83 + ], + "lines": [ + { + "bbox": [ + 140, + 72, + 173, + 83 + ], + "spans": [ + { + "bbox": [ + 140, + 72, + 173, + 83 + ], + "type": "text", + "content": "sofa" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 74, + 89, + 167, + 171 + ], + "lines": [ + { + "bbox": [ + 74, + 89, + 167, + 171 + ], + "spans": [ + { + "bbox": [ + 74, + 89, + 167, + 171 + ], + "type": "image", + "image_path": "9a9d63bc14a6fefc8c9c0ad840274b479ff52a642bad4c6316dffc0bf9b2e65b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 179, + 94, + 255, + 171 + ], + "blocks": [ + { + "bbox": [ + 194, + 72, + 230, + 83 + ], + "lines": [ + { + "bbox": [ + 194, + 72, + 230, + 83 + ], + "spans": [ + { + "bbox": [ + 194, + 72, + 230, + 83 + ], + "type": "text", + "content": "chair" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 179, + 94, + 255, + 171 + ], + "lines": [ + { + "bbox": [ + 179, + 94, + 255, + 171 + ], + "spans": [ + { + "bbox": [ + 179, + 94, + 255, + 171 + ], + "type": "image", + "image_path": "244ef087c8e1a8e6490f37fa66405bb095af4e3b30a526c1a202348993c51aea.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 262, + 97, + 351, + 173 + ], + "blocks": [ + { + "bbox": [ + 264, + 72, + 293, + 83 + ], + "lines": [ + { + "bbox": [ + 264, + 72, + 293, + 83 + ], + "spans": [ + { + "bbox": [ + 264, + 72, + 293, + 83 + ], + "type": "text", + "content": "cabinet" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 72, + 335, + 83 + ], + "lines": [ + { + "bbox": [ + 304, + 72, + 335, + 83 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 335, + 83 + ], + "type": "text", + "content": "book" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 72, + 356, + 83 + ], + "lines": [ + { + "bbox": [ + 319, + 72, + 356, + 83 + ], + "spans": [ + { + "bbox": [ + 319, + 72, + 356, + 83 + ], + "type": "text", + "content": "bookshelf" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 262, + 97, + 351, + 173 + ], + "lines": [ + { + "bbox": [ + 262, + 97, + 351, + 173 + ], + "spans": [ + { + "bbox": [ + 262, + 97, + 351, + 173 + ], + "type": "image", + "image_path": "65c39733e6f89e93e8cd31ce4d3ef942fc378e298ec229170563a58ee3209d07.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 371, + 89, + 456, + 173 + ], + "blocks": [ + { + "bbox": [ + 373, + 72, + 402, + 84 + ], + "lines": [ + { + "bbox": [ + 373, + 72, + 402, + 84 + ], + "spans": [ + { + "bbox": [ + 373, + 72, + 402, + 84 + ], + "type": "text", + "content": "display" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 413, + 72, + 459, + 83 + ], + "lines": [ + { + "bbox": [ + 413, + 72, + 459, + 83 + ], + "spans": [ + { + "bbox": [ + 413, + 72, + 459, + 83 + ], + "type": "text", + "content": "bathtub" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 468, + 72, + 481, + 83 + ], + "lines": [ + { + "bbox": [ + 468, + 72, + 481, + 83 + ], + "spans": [ + { + "bbox": [ + 468, + 72, + 481, + 83 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 371, + 89, + 456, + 173 + ], + "lines": [ + { + "bbox": [ + 371, + 89, + 456, + 173 + ], + "spans": [ + { + "bbox": [ + 371, + 89, + 456, + 173 + ], + "type": "image", + "image_path": "c8cd1e88b6ba4c8a19f35abbaa0e1629191853277d3119d4fd8d5361171a3bd4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 463, + 97, + 549, + 169 + ], + "blocks": [ + { + "bbox": [ + 483, + 72, + 517, + 83 + ], + "lines": [ + { + "bbox": [ + 483, + 72, + 517, + 83 + ], + "spans": [ + { + "bbox": [ + 483, + 72, + 517, + 83 + ], + "type": "text", + "content": "trash bin" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 97, + 549, + 169 + ], + "lines": [ + { + "bbox": [ + 463, + 97, + 549, + 169 + ], + "spans": [ + { + "bbox": [ + 463, + 97, + 549, + 169 + ], + "type": "image", + "image_path": "0d7b1a6a46948fd351510aafb9c90f349af3fb31fee47fbde432d335194d74b2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 75, + 176, + 164, + 252 + ], + "blocks": [ + { + "bbox": [ + 53, + 198, + 64, + 232 + ], + "lines": [ + { + "bbox": [ + 53, + 198, + 64, + 232 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 64, + 232 + ], + "type": "text", + "content": "Rrnnnne" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 176, + 164, + 252 + ], + "lines": [ + { + "bbox": [ + 75, + 176, + 164, + 252 + ], + "spans": [ + { + "bbox": [ + 75, + 176, + 164, + 252 + ], + "type": "image", + "image_path": "051a5192a1400bea33d7b84646fa08c598e9d390af94c409cff904713aeb197b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 181, + 180, + 253, + 258 + ], + "blocks": [ + { + "bbox": [ + 181, + 180, + 253, + 258 + ], + "lines": [ + { + "bbox": [ + 181, + 180, + 253, + 258 + ], + "spans": [ + { + "bbox": [ + 181, + 180, + 253, + 258 + ], + "type": "image", + "image_path": "cab921369352c5f501b02ff57407e105bfb8277203377d2bdee2d9f6710b535b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 262, + 182, + 352, + 260 + ], + "blocks": [ + { + "bbox": [ + 262, + 182, + 352, + 260 + ], + "lines": [ + { + "bbox": [ + 262, + 182, + 352, + 260 + ], + "spans": [ + { + "bbox": [ + 262, + 182, + 352, + 260 + ], + "type": "image", + "image_path": "13c26f9580be58219d55f551355e6160ed756e3bfbdf2a474d1d9c6eee17e283.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 371, + 173, + 444, + 259 + ], + "blocks": [ + { + "bbox": [ + 371, + 173, + 444, + 259 + ], + "lines": [ + { + "bbox": [ + 371, + 173, + 444, + 259 + ], + "spans": [ + { + "bbox": [ + 371, + 173, + 444, + 259 + ], + "type": "image", + "image_path": "c160c17593b7c733d07ffde23fd2a0fddd6c9d2af4c3189b68adff389d3c8d7b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 462, + 186, + 546, + 258 + ], + "blocks": [ + { + "bbox": [ + 462, + 186, + 546, + 258 + ], + "lines": [ + { + "bbox": [ + 462, + 186, + 546, + 258 + ], + "spans": [ + { + "bbox": [ + 462, + 186, + 546, + 258 + ], + "type": "image", + "image_path": "b7e632ab8419327f73b42f7ca7868338136fe781f7386af1eacb7f8c0c575211.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 77, + 264, + 166, + 342 + ], + "blocks": [ + { + "bbox": [ + 53, + 292, + 64, + 319 + ], + "lines": [ + { + "bbox": [ + 53, + 292, + 64, + 319 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 64, + 319 + ], + "type": "text", + "content": "DIMR" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 264, + 166, + 342 + ], + "lines": [ + { + "bbox": [ + 77, + 264, + 166, + 342 + ], + "spans": [ + { + "bbox": [ + 77, + 264, + 166, + 342 + ], + "type": "image", + "image_path": "6a5c945cfe03da0e170de32b30cc25c1039aa17128e5d9906a1b7e03b849450e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 181, + 269, + 253, + 342 + ], + "blocks": [ + { + "bbox": [ + 181, + 269, + 253, + 342 + ], + "lines": [ + { + "bbox": [ + 181, + 269, + 253, + 342 + ], + "spans": [ + { + "bbox": [ + 181, + 269, + 253, + 342 + ], + "type": "image", + "image_path": "b74a0e32bd68fba9ecf58cd5caa23cf5ef7660e4547e8e6085bcfc3594e7e37e.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 263, + 269, + 352, + 347 + ], + "blocks": [ + { + "bbox": [ + 263, + 269, + 352, + 347 + ], + "lines": [ + { + "bbox": [ + 263, + 269, + 352, + 347 + ], + "spans": [ + { + "bbox": [ + 263, + 269, + 352, + 347 + ], + "type": "image", + "image_path": "c1ab0bdc432706fc6f59dd743dc5ead262fce94712b53a3ca67f3b8af222d475.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 371, + 259, + 456, + 348 + ], + "blocks": [ + { + "bbox": [ + 371, + 259, + 456, + 348 + ], + "lines": [ + { + "bbox": [ + 371, + 259, + 456, + 348 + ], + "spans": [ + { + "bbox": [ + 371, + 259, + 456, + 348 + ], + "type": "image", + "image_path": "d3d94a0231fbf452e9aa7a9afe05132540cc7a0d88d1f0c9d66e79a3d8df1094.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 462, + 275, + 545, + 345 + ], + "blocks": [ + { + "bbox": [ + 462, + 275, + 545, + 345 + ], + "lines": [ + { + "bbox": [ + 462, + 275, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 462, + 275, + 545, + 345 + ], + "type": "image", + "image_path": "b18b1ed6b027c50697fa2a917e85aba09b1e720366036f318aa57e49af3541f4.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 75, + 350, + 168, + 430 + ], + "blocks": [ + { + "bbox": [ + 53, + 384, + 64, + 407 + ], + "lines": [ + { + "bbox": [ + 53, + 384, + 64, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 384, + 64, + 407 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 350, + 168, + 430 + ], + "lines": [ + { + "bbox": [ + 75, + 350, + 168, + 430 + ], + "spans": [ + { + "bbox": [ + 75, + 350, + 168, + 430 + ], + "type": "image", + "image_path": "d7f22aa197240be31258b82429526b114a0bbc565ba5b7ea308c66cfca1665e8.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 178, + 358, + 253, + 434 + ], + "blocks": [ + { + "bbox": [ + 178, + 358, + 253, + 434 + ], + "lines": [ + { + "bbox": [ + 178, + 358, + 253, + 434 + ], + "spans": [ + { + "bbox": [ + 178, + 358, + 253, + 434 + ], + "type": "image", + "image_path": "250540929aa94f254436d63792fa6ec9a5d146f629b19de4697749eaff2cdae8.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 263, + 361, + 352, + 434 + ], + "blocks": [ + { + "bbox": [ + 263, + 361, + 352, + 434 + ], + "lines": [ + { + "bbox": [ + 263, + 361, + 352, + 434 + ], + "spans": [ + { + "bbox": [ + 263, + 361, + 352, + 434 + ], + "type": "image", + "image_path": "f552f9f5d727282ea51f048220f6c41914ba54956c3db0cfdda01aa48d4bd249.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 370, + 351, + 455, + 435 + ], + "blocks": [ + { + "bbox": [ + 370, + 351, + 455, + 435 + ], + "lines": [ + { + "bbox": [ + 370, + 351, + 455, + 435 + ], + "spans": [ + { + "bbox": [ + 370, + 351, + 455, + 435 + ], + "type": "image", + "image_path": "e91509c395083ed1464fa142d7549cc3522413f945b12f9adc62253c300e040c.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 444, + 448, + 456 + ], + "lines": [ + { + "bbox": [ + 144, + 444, + 448, + 456 + ], + "spans": [ + { + "bbox": [ + 144, + 444, + 448, + 456 + ], + "type": "text", + "content": "Figure 5. Visualization comparison of scene semantic reconstruction on ScanNetv2." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 466, + 358, + 552, + 434 + ], + "blocks": [ + { + "bbox": [ + 466, + 358, + 552, + 434 + ], + "lines": [ + { + "bbox": [ + 466, + 358, + 552, + 434 + ], + "spans": [ + { + "bbox": [ + 466, + 358, + 552, + 434 + ], + "type": "image", + "image_path": "29d417f9e7d7f65320d207e6d0b070f7cd462c3b23dc413346ff165bc2246873.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 76, + 467, + 149, + 550 + ], + "blocks": [ + { + "bbox": [ + 76, + 467, + 149, + 550 + ], + "lines": [ + { + "bbox": [ + 76, + 467, + 149, + 550 + ], + "spans": [ + { + "bbox": [ + 76, + 467, + 149, + 550 + ], + "type": "image", + "image_path": "dca2775e00598ce4f3d752f47ff8ae1659f53344114f6035704798f257be2b76.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 553, + 531, + 565 + ], + "lines": [ + { + "bbox": [ + 61, + 553, + 531, + 565 + ], + "spans": [ + { + "bbox": [ + 61, + 553, + 531, + 565 + ], + "type": "text", + "content": "Figure 6. Holistic scene understanding by our proposed AnchorRec. (The comparison methods didn't conduct layout estimation.)" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 155, + 467, + 251, + 550 + ], + "blocks": [ + { + "bbox": [ + 155, + 467, + 251, + 550 + ], + "lines": [ + { + "bbox": [ + 155, + 467, + 251, + 550 + ], + "spans": [ + { + "bbox": [ + 155, + 467, + 251, + 550 + ], + "type": "image", + "image_path": "7c0a3f1417de627a5f2002560ee33e4e860da3bcc7ef19891b4bb776348adafd.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 257, + 468, + 335, + 547 + ], + "blocks": [ + { + "bbox": [ + 257, + 468, + 335, + 547 + ], + "lines": [ + { + "bbox": [ + 257, + 468, + 335, + 547 + ], + "spans": [ + { + "bbox": [ + 257, + 468, + 335, + 547 + ], + "type": "image", + "image_path": "c23932c9257b599fee521a4d850115784c9c9281598c5d979c6708cc9612a9c5.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 337, + 467, + 414, + 544 + ], + "blocks": [ + { + "bbox": [ + 337, + 467, + 414, + 544 + ], + "lines": [ + { + "bbox": [ + 337, + 467, + 414, + 544 + ], + "spans": [ + { + "bbox": [ + 337, + 467, + 414, + 544 + ], + "type": "image", + "image_path": "3f4516b8a0b7bc1e7d92eb575b3a8d2c00289db1049ed1fd03dfdc3239de4292.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 424, + 474, + 518, + 547 + ], + "blocks": [ + { + "bbox": [ + 424, + 474, + 518, + 547 + ], + "lines": [ + { + "bbox": [ + 424, + 474, + 518, + 547 + ], + "spans": [ + { + "bbox": [ + 424, + 474, + 518, + 547 + ], + "type": "image", + "image_path": "8f1d612f3e8620f3f856b83312e3303ef2419d005dac866fe2f28b43fea9111f.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 47, + 585, + 289, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 289, + 705 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 289, + 705 + ], + "type": "text", + "content": "DIMR infers object models from the instance points provided by a complex instance segmentation backbone. As shown in Fig. 5, our AncRec can robustly recognize and reconstruct objects from noisy point clouds with accurate object localization and high-fidelity shape details. With the target-focused proposal feature learning offered by AncLearn in terms of object detection, our method generates fewer false positive models than the compared approaches. With more accurate detection results, our AncRec can leverage outlier-reduced instance priors sampled by AncLearn" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 304, + 585, + 545, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 585, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 304, + 585, + 545, + 704 + ], + "type": "text", + "content": "to infer more detailed object structures, e.g., the folding chairs are recovered in the second column of Fig. 5. Tab. 1 presents the quantitative comparison of different methods. Our AncRec outperforms other methods under all metrics. As the comparison methods didn't conduct layout estimation, we illustrate our holistic scene understanding results in Fig. 6. Considering that semantic scene reconstruction is a comprehensive problem, we further analyze the performance of our method on the key sub-tasks, i.e., scene parsing and object reconstruction, as follows." + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21921" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 69, + 521, + 290 + ], + "blocks": [ + { + "bbox": [ + 77, + 69, + 521, + 290 + ], + "lines": [ + { + "bbox": [ + 77, + 69, + 521, + 290 + ], + "spans": [ + { + "bbox": [ + 77, + 69, + 521, + 290 + ], + "type": "image", + "image_path": "05fbafc4950b2b8ca7e9789af44baddcbce920c6c06d81f2da0009636d2e92ec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 295, + 436, + 306 + ], + "lines": [ + { + "bbox": [ + 157, + 295, + 436, + 306 + ], + "spans": [ + { + "bbox": [ + 157, + 295, + 436, + 306 + ], + "type": "text", + "content": "Figure 7. Visualization comparison results of object detection on ScanNetv2" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 71, + 320, + 508, + 524 + ], + "blocks": [ + { + "bbox": [ + 71, + 320, + 508, + 524 + ], + "lines": [ + { + "bbox": [ + 71, + 320, + 508, + 524 + ], + "spans": [ + { + "bbox": [ + 71, + 320, + 508, + 524 + ], + "type": "image", + "image_path": "9c99e359ac9cf89a1befe6d2323d4b47c645a1810b1a123a1d3fc7dc7aab4aa3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 525, + 547, + 559 + ], + "lines": [ + { + "bbox": [ + 46, + 525, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 547, + 559 + ], + "type": "text", + "content": "Figure 8. The influence of different instance point sampling strategies on object reconstruction. Points from the same instance are painted in the same color. Without the reliance on segmentation like RfD-Net and DIMR, our method can correctly separate adjacent objects and predict shape models." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 49, + 608, + 286, + 646 + ], + "blocks": [ + { + "bbox": [ + 47, + 575, + 287, + 608 + ], + "lines": [ + { + "bbox": [ + 47, + 575, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 287, + 608 + ], + "type": "text", + "content": "Table 1. Quantitative results on the ScanNetv2 dataset. We evaluate the reconstruction quality with mAP at different thresholds. (higher is better)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 608, + 286, + 646 + ], + "lines": [ + { + "bbox": [ + 49, + 608, + 286, + 646 + ], + "spans": [ + { + "bbox": [ + 49, + 608, + 286, + 646 + ], + "type": "table", + "html": "
IoU@0.25IoU@0.5CD@0.1CD@0.047LFD@5000LFD@2500
RfD-Net42.516.945.719.128.67.8
DIMR46.312.551.925.729.58.6
AnchorRec52.918.956.829.430.39.9
", + "image_path": "4ed1351fb026fbc37315b72783e9463f35b12e148a57f43f6a1d4a5ec4cf4983.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "type": "text", + "content": "Scene Parsing. We compare AncRec with the state-of-the-art methods regarding 8-category object detection and layout estimation. The quantitative results in Tab. 2 show that our AnchorRec edges out the comparison methods by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 578, + 547, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 547, + 711 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 547, + 711 + ], + "type": "text", + "content": "significant performance gains. Especially for the categories of the display, bathtub, and trash bin, AncRec exceeds the second best by " + }, + { + "bbox": [ + 304, + 578, + 547, + 711 + ], + "type": "inline_equation", + "content": "6.6\\%" + }, + { + "bbox": [ + 304, + 578, + 547, + 711 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 578, + 547, + 711 + ], + "type": "inline_equation", + "content": "5.8\\%" + }, + { + "bbox": [ + 304, + 578, + 547, + 711 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 578, + 547, + 711 + ], + "type": "inline_equation", + "content": "10.7\\%" + }, + { + "bbox": [ + 304, + 578, + 547, + 711 + ], + "type": "text", + "content": " in precision measurement. The results indicate that, with adaptive grouping feature areas, AncLearn works effectively for detecting objects shaped irregularly, e.g., diverse-scaled bathtubs, thin displays, and tiny trash bins. In Tab. 2, we also compare our method with PQ-Transofrmer [7], the current state-of-the-art scene parsing approach, in terms of layout estimation. The results demonstrate that our AncLearn is also applicable to localizing walls that are particularly large and thin." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21922" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 89, + 541, + 203 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 545, + 89 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 545, + 89 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 545, + 89 + ], + "type": "text", + "content": "Table 2. Quantitative results of object detection and layout estimation. The AP scores are measured with an IoU threshold of 0.5. Higher values mean better results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 89, + 541, + 203 + ], + "lines": [ + { + "bbox": [ + 49, + 89, + 541, + 203 + ], + "spans": [ + { + "bbox": [ + 49, + 89, + 541, + 203 + ], + "type": "table", + "html": "
ObjectLayout
Methodtablesofachaircabinetbookshelfdisplaybathtubtrash binmAPF1-scores
VoteNet [14]44.547.576.328.820.18.026.014.633.2-
MLCVNet [35]50.546.681.825.519.27.825.319.234.5-
H3DNet [38]53.652.476.025.627.29.420.710.034.4-
BRNet [10]56.443.180.935.426.110.428.211.036.4-
PQ-Transformer [7]52.752.581.132.136.46.720.620.837.867.0
AnchorRec (Ours)57.356.184.536.929.817.034.031.543.470.5
", + "image_path": "b21f33b7f6a969990b959af9e6bf35ad9b71095beb906cfa32be12bb3de1b8f3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 69, + 223, + 523, + 282 + ], + "blocks": [ + { + "bbox": [ + 65, + 211, + 527, + 222 + ], + "lines": [ + { + "bbox": [ + 65, + 211, + 527, + 222 + ], + "spans": [ + { + "bbox": [ + 65, + 211, + 527, + 222 + ], + "type": "text", + "content": "Table 3. Numeric results on object reconstruction. The AP scores are measured with CD at a threshold of 0.1. (Higher is better.)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 223, + 523, + 282 + ], + "lines": [ + { + "bbox": [ + 69, + 223, + 523, + 282 + ], + "spans": [ + { + "bbox": [ + 69, + 223, + 523, + 282 + ], + "type": "table", + "html": "
Methodtablechairbookshelfsofatrash bincabinetdisplaybathtubmAP
RfD-Net [28]25.5182.1132.5344.2144.7428.3765.5142.5745.70
DIMR [33]39.4481.0438.2444.0962.6023.5775.1250.9351.88
AnchorRec (Ours)53.6086.2336.8250.9259.5939.5673.6854.3256.84
", + "image_path": "4e9e36896b4b53dfb64e6d23e17741e9c664ebb514dc8f59af597c7f3cc6a472.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 304, + 287, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 340 + ], + "type": "text", + "content": "Visualized object detection result provided in Fig. 7 further manifests that our AnchorRec can reliably delineate objects with compact bounding box prediction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 340, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 424 + ], + "type": "text", + "content": "Object Mesh Reconstruction. In Tab. 3, we also evaluate the performance of different methods with respect to class-wise object reconstruction. The evaluation is based on how matchable the predicted object meshes are with the scene-aligned ground truths. The numeric results in Tab. 3 show that AncRec achieves the best performance on 5 categories and the overall chamfer distance-based mAP score." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 432, + 148, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 432, + 148, + 444 + ], + "spans": [ + { + "bbox": [ + 47, + 432, + 148, + 444 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 451, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 559 + ], + "type": "text", + "content": "AncLearn for Scene Parsing. We study different settings of AncLearn for instance detection and report the results in Tab. 4. Compared to the baseline without AncLearn, the proposed AncLearn brings consistent improvement to object detection and layout estimation. Especially, when cooperating with the self-attention layer for layout estimation, AncLearn enables the dual-branch instance detector of AncRec to obtain the best performance in both parsing tasks." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 49, + 594, + 287, + 672 + ], + "blocks": [ + { + "bbox": [ + 46, + 561, + 287, + 594 + ], + "lines": [ + { + "bbox": [ + 46, + 561, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 561, + 287, + 594 + ], + "type": "text", + "content": "Table 4. Ablation study of the proposed Anglicen. IoU@0.5 and F1 score respectively reflect the performance of object detection and layout estimation. Higher values mean better results." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 594, + 287, + 672 + ], + "lines": [ + { + "bbox": [ + 49, + 594, + 287, + 672 + ], + "spans": [ + { + "bbox": [ + 49, + 594, + 287, + 672 + ], + "type": "table", + "html": "
OD anchorlayout anchorlayout SAIoU@0.5F1 score
41.1260.62
✓43.33(+2.21)60.59(-0.03)
✓40.21(-0.91)62.61(+1.99)
✓✓41.24(+0.12)64.00(+3.38)
✓✓41.16(+0.04)65.34(+4.72)
✓✓✓43.38(+2.26)70.45(+9.83)
", + "image_path": "2754fb5dfbedf6dc43d4bdea7c5377873a3e81ba066cc1aa5beca15a78ec710d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "AncLearn for Object Reconstruction. In the shape predictor of AncRec, AncLearn works to sample instance points as geometry priors for reconstruction. With the in" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 304, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 376 + ], + "type": "text", + "content": "stance detector of AncRec, we compare AncLearn with the segmentation-based strategy used in RfD-Net and a box cropping sampling method. The quantitative comparison in Tab. 5 shows that AncLearn outperforms the other two approaches by providing instructive geometry priors with shape anchors in terms of all metrics." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 309, + 407, + 545, + 515 + ], + "blocks": [ + { + "bbox": [ + 305, + 385, + 545, + 407 + ], + "lines": [ + { + "bbox": [ + 305, + 385, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 385, + 545, + 407 + ], + "type": "text", + "content": "Table 5. Comparison between different instance point sampling methods. Higher values mean better results." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 407, + 545, + 515 + ], + "lines": [ + { + "bbox": [ + 309, + 407, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 309, + 407, + 545, + 515 + ], + "type": "table", + "html": "
IoU@0.5IoU@0.25CD@0.1
Segmentation [28]17.451.455.8
Box Cropping17.751.955.5
AncLearn (Ours)18.952.956.8
CD@0.047LFD@5000LFD@2500
Segmentation [28]29.129.17.8
Box Cropping28.829.99.8
AncLearn (Ours)29.430.39.9
", + "image_path": "89fa5a3dce60b83c0793088ee9c206fab0c933201f0e1d78ca0f1d1ca12b9438.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 527, + 545, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 527, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 305, + 527, + 545, + 539 + ], + "type": "text", + "content": "4.4. Contributions of the Vote and Anchor Features" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "As mentioned in Sec. 3.1.1, we perform instance detection by taking both vote and anchor features into account. To investigate the difference in their contributions to the detection parameters, we visualize the two weight vectors in Eq. (3) for perceptual comparison. As shown in Fig. 9, vote features have a larger impact on scoring objectness, indicating that the context encoded in vote features is useful in differentiating between object and non-object regions. Compared to vote features, anchor features contribute significantly more to angle prediction, which demonstrates the superiority of anchor-guided strategy in providing shape clues for learning shape-sensitive parameters. Considering the gap between the bounding box and the shape of objects, the vote features contribute more to predicting the bounding" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21923" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 168, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 168, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 168, + 83 + ], + "type": "text", + "content": "box center and size residuals." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 94, + 286, + 151 + ], + "blocks": [ + { + "bbox": [ + 48, + 94, + 286, + 151 + ], + "lines": [ + { + "bbox": [ + 48, + 94, + 286, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 94, + 286, + 151 + ], + "type": "image", + "image_path": "9b28b9bf95f77ff1b8a506ee1fd43357a429e41f3cde3ca003f564ca9e1387d7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 154, + 287, + 175 + ], + "lines": [ + { + "bbox": [ + 46, + 154, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 46, + 154, + 287, + 175 + ], + "type": "text", + "content": "Figure 9. The learned weights for combining vote and anchor features." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 198, + 276, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 198, + 276, + 211 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 276, + 211 + ], + "type": "text", + "content": "4.5. The Characteristics of the Learned Anchors" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 218, + 288, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 288, + 338 + ], + "type": "text", + "content": "The supervision with surface points derived from complete 3D models enables the anchors to depict object shapes, even with unobserved structures, which assists in object detection and reconstruction when some parts of objects are missing. There also exists a difference between anchor distribution characteristics for objects and non-object areas. As shown in Fig. 10, the anchors and sampled points depict object shapes, while those for non-object areas scatter irregularly. This characteristic may play a role in differentiating objects from non-object regions." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 350, + 89, + 384 + ], + "blocks": [ + { + "bbox": [ + 57, + 350, + 89, + 384 + ], + "lines": [ + { + "bbox": [ + 57, + 350, + 89, + 384 + ], + "spans": [ + { + "bbox": [ + 57, + 350, + 89, + 384 + ], + "type": "image", + "image_path": "27a52f254a715a3765d6315cb42b56f77eae240d18011cdfe71195be02939a5f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 396, + 286, + 418 + ], + "lines": [ + { + "bbox": [ + 47, + 396, + 286, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 286, + 418 + ], + "type": "text", + "content": "Figure 10. The characteristics of the learned anchors. The anchors and sampled points are shown in red and purple respectively." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 94, + 350, + 121, + 384 + ], + "blocks": [ + { + "bbox": [ + 94, + 350, + 121, + 384 + ], + "lines": [ + { + "bbox": [ + 94, + 350, + 121, + 384 + ], + "spans": [ + { + "bbox": [ + 94, + 350, + 121, + 384 + ], + "type": "image", + "image_path": "1fde219b68682415596ec503b2e7eea94e5f5bdb6749a7a8a0b082c9326b7186.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 388, + 148, + 396 + ], + "lines": [ + { + "bbox": [ + 107, + 388, + 148, + 396 + ], + "spans": [ + { + "bbox": [ + 107, + 388, + 148, + 396 + ], + "type": "text", + "content": "Object Anchors" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 132, + 350, + 161, + 384 + ], + "blocks": [ + { + "bbox": [ + 132, + 350, + 161, + 384 + ], + "lines": [ + { + "bbox": [ + 132, + 350, + 161, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 161, + 384 + ], + "type": "image", + "image_path": "b864c66883678c816552606d3f21575c14a3f2176ffa97961c88fd0be8ecff3d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 167, + 350, + 198, + 384 + ], + "blocks": [ + { + "bbox": [ + 167, + 350, + 198, + 384 + ], + "lines": [ + { + "bbox": [ + 167, + 350, + 198, + 384 + ], + "spans": [ + { + "bbox": [ + 167, + 350, + 198, + 384 + ], + "type": "image", + "image_path": "2bc16df606235930bf5b7f008272ddbc103208e16c249b98001392c3ac4466d9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 205, + 350, + 276, + 385 + ], + "blocks": [ + { + "bbox": [ + 205, + 350, + 276, + 385 + ], + "lines": [ + { + "bbox": [ + 205, + 350, + 276, + 385 + ], + "spans": [ + { + "bbox": [ + 205, + 350, + 276, + 385 + ], + "type": "image", + "image_path": "c6ced1631ed69f365482a2da19213fd66670d532ab8f76d52990a6c2ac34a059.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 388, + 269, + 396 + ], + "lines": [ + { + "bbox": [ + 217, + 388, + 269, + 396 + ], + "spans": [ + { + "bbox": [ + 217, + 388, + 269, + 396 + ], + "type": "text", + "content": "Non-object Anchors" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 443, + 125, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 125, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 125, + 456 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 464, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 643 + ], + "type": "text", + "content": "In this paper, we introduce a shape anchor guided learning strategy (AncLearn) that is embedded into a reconstruction-from-detection learning system (AncRec) to handle the issue of noise interference in point-based holistic scene understanding. Extensive experiments demonstrate that AncRec achieves high-quality indoor semantic scene reconstruction. The quantitative and qualitative results show that AncRec outperforms current methods in terms of object detection, layout estimation, and shape modeling. The ablation studies convincingly verify that AncLearn can largely exclude noise from search space for reliable feature grouping and robust instance point sampling. In the future, it is promising to study the application of shape anchor guided learning strategy in other point-based 3D vision tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 657, + 139, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 139, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 139, + 670 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "This research was supported by the Fundamental Research Funds for the Central Universities of China under Grant 2042022dx0001, NSFC projects under Grant" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 72, + 546, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 546, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 546, + 120 + ], + "type": "text", + "content": "42071370, Wuhan University-Huawei Geoinformatics Innovation Laboratory, and the Open fund of Key Laboratory of Urban Land Resources Monitoring and Simulation, Ministry of Natural Resources under Grant KF202106084." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 132, + 365, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 132, + 365, + 144 + ], + "spans": [ + { + "bbox": [ + 307, + 132, + 365, + 144 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 311, + 152, + 546, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 152, + 546, + 207 + ], + "spans": [ + { + "bbox": [ + 311, + 152, + 546, + 207 + ], + "type": "text", + "content": "[1] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 2, 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 311, + 208, + 547, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 208, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 311, + 208, + 547, + 252 + ], + "type": "text", + "content": "[2] Armen Avetisyan, Angela Dai, and Matthias Nießner. End-to-end cad model retrieval and 9dof alignment in 3d scans. In Proceedings of the IEEE/CVF International Conference on computer vision, pages 2551-2560, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 311, + 254, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 254, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 311, + 254, + 547, + 319 + ], + "type": "text", + "content": "[3] Armen Avetisyan, Tatiana Khanova, Christopher Choy, Denver Dash, Angela Dai, and Matthias Nießner. Scenecad: Predicting object alignments and layouts in rgb-d scans. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 596-612. Springer, 2020. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 311, + 321, + 547, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 321, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 311, + 321, + 547, + 374 + ], + "type": "text", + "content": "[4] Aleksei Bokhovkin and Angela Dai. Neural part priors: Learning to optimize part-based object completion in rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9032-9042, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 311, + 376, + 547, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 376, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 311, + 376, + 547, + 431 + ], + "type": "text", + "content": "[5] Alexey Bokhovkin, Vladislav Ishimtsev, Emil Bogomolov, Denis Zorin, Alexey Artemov, Evgeny Burnaev, and Angela Dai. Towards part-based understanding of rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7484-7494, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 311, + 433, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 433, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 311, + 433, + 547, + 487 + ], + "type": "text", + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 311, + 489, + 546, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 489, + 546, + 532 + ], + "spans": [ + { + "bbox": [ + 311, + 489, + 546, + 532 + ], + "type": "text", + "content": "[7] Xiaoxue Chen, Hao Zhao, Guyue Zhou, and Ya-Qin Zhang. Pq-transformer: Jointly parsing 3d objects and layouts from point clouds. IEEE Robotics and Automation Letters, 7(2):2519-2526, 2022. 4, 5, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 311, + 534, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 534, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 311, + 534, + 547, + 578 + ], + "type": "text", + "content": "[8] Yujin Chen, Matthias Nießner, and Angela Dai. 4dcontrast: Contrastive learning with dynamic correspondences for 3d scene understanding. In European Conference on Computer Vision, pages 543-560. Springer, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 311, + 579, + 546, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 579, + 546, + 623 + ], + "spans": [ + { + "bbox": [ + 311, + 579, + 546, + 623 + ], + "type": "text", + "content": "[9] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 45-54, 2020. 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 624, + 547, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 547, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 547, + 678 + ], + "type": "text", + "content": "[10] Bowen Cheng, Lu Sheng, Shaoshuai Shi, Ming Yang, and Dong Xu. Back-tracing representative points for voting-based 3d object detection in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8963-8972, 2021. 2, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "type": "text", + "content": "[11] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In Proceedings of the IEEE/CVF con" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21924" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "type": "text", + "content": "ference on computer vision and pattern recognition, pages 6970-6981, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "text", + "content": "[13] Angela Dai, Christian Diller, and Matthias Nießner. Sg-nn: Sparse generative neural networks for self-supervised scene completion of rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 849–858, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "text", + "content": "[14] Zhipeng Ding, Xu Han, and Marc Niethammer. Volenet: A deep learning label fusion method for multi-atlas segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 202-210. Springer, 2019. 2, 4, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "text", + "content": "[15] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9031-9040, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "type": "text", + "content": "[16] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4857-4866, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "type": "text", + "content": "[17] Saurabh Gupta, Pablo Arbeláez, Ross Girshick, and Jitendra Malik. Aligning 3d models to rgb-d images of cluttered scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4731-4740, 2015. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 498 + ], + "type": "text", + "content": "[18] Xiaoguang Han, Zhaoxuan Zhang, Dong Du, Mingdai Yang, Jingming Yu, Pan Pan, Xin Yang, Ligang Liu, Zixiang Xiong, and Shuguang Cui. Deep reinforcement learning of volume-guided progressive view inpainting for 3d point scene completion from a single depth image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 234-243, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 499, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 543 + ], + "type": "text", + "content": "[19] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. MeshCNN: a network with an edge. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 544, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 589 + ], + "type": "text", + "content": "[20] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4421-4430, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "text", + "content": "[21] Ji Hou, Angela Dai, and Matthias Nießner. Revealnet: Seeing behind objects in rgb-d scans. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2095-2104, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[22] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15587-15597, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[23] Linxi Huan, Xianwei Zheng, and Jianya Gong. Georec: Geometry-enhanced semantic 3d reconstruction of rgb-d in" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "door scenes. ISPRS Journal of Photogrammetry and Remote Sensing, 186:301-314, 2022. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 95, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 128 + ], + "type": "text", + "content": "[24] Hamid Izadinia, Qi Shan, and Steven M Seitz. Im2cad. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5134-5143, 2017. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "text", + "content": "[25] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International journal of computer vision, 128:261-318, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 228 + ], + "type": "text", + "content": "[26] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 228, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 293 + ], + "type": "text", + "content": "[27] Yinyu Nie, Xiaoguang Han, Shihui Guo, Yujuan Zheng, Jian Chang, and Jian Jun Zhang. Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 55-64, 2020. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 293, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 293, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 293, + 545, + 348 + ], + "type": "text", + "content": "[28] Yinyu Nie, Ji Hou, Xiaoguang Han, and Matthias Nießner. Rfd-net: Point scene understanding by semantic instance reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4608-4618, 2021. 1, 2, 3, 5, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 349, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 545, + 404 + ], + "type": "text", + "content": "[29] Charles R Qi, Xinlei Chen, Or Litany, and Leonidas J Guibas. Imvotenet: Boosting 3d object detection in point clouds with image votes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4404-4413, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "type": "text", + "content": "[30] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 459, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 459, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 545, + 503 + ], + "type": "text", + "content": "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 503, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 503, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 307, + 503, + 545, + 557 + ], + "type": "text", + "content": "[32] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. PointCNN: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "type": "text", + "content": "[33] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Point scene understanding via disentangled instance mesh reconstruction. arXiv preprint arXiv:2203.16832, 2022. 1, 2, 3, 5, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 602, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 658 + ], + "type": "text", + "content": "[34] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Dening Lu, Mingqiang Wei, and Jun Wang. Venet: Voting enhancement network for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3712-3721, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[35] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Yiming Zhang, Kai Xu, and Jun Wang. Mlcvnet: Multi-level context votenet for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10447-10456, 2020. 2, 4, 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21925" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 318 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[36] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Simon Chen, Yifan Liu, and Chunhua Shen. Towards accurate reconstruction of 3d scene shape from a single monocular image. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):6480-6494, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 194 + ], + "type": "text", + "content": "[37] Cheng Zhang, Zhaopeng Cui, Yinda Zhang, Bing Zeng, Marc Pollefeys, and Shuaicheng Liu. Holistic 3d scene understanding from a single image with implicit representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8833-8842, 2021. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 196, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 196, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 196, + 287, + 239 + ], + "type": "text", + "content": "[38] Zaiwei Zhang, Bo Sun, Haitao Yang, and Qixing Huang. H3dnet: 3d object detection using hybrid geometric primitives. In European Conference on Computer Vision, pages 311-329. Springer, 2020. 2, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 240, + 288, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 288, + 283 + ], + "type": "text", + "content": "[39] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 285, + 286, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 286, + 318 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 286, + 318 + ], + "type": "text", + "content": "[40] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 2023. 2" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21926" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/ShapeScaffolder_ Structure-Aware 3D Shape Generation from Text/3212f743-a520-438c-b4e2-ff41229e0a86_content_list.json b/2023/ShapeScaffolder_ Structure-Aware 3D Shape Generation from Text/3212f743-a520-438c-b4e2-ff41229e0a86_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3ce3b763531e3a2e8dd16d9bf00644c7437f7e42 --- /dev/null +++ b/2023/ShapeScaffolder_ Structure-Aware 3D Shape Generation from Text/3212f743-a520-438c-b4e2-ff41229e0a86_content_list.json @@ -0,0 +1,1300 @@ +[ + { + "type": "text", + "text": "ShapeScaffold: Structure-Aware 3D Shape Generation from Text", + "text_level": 1, + "bbox": [ + 145, + 130, + 823, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xi Tian \nUniversity of Bath \nBath, UK \nxt275@bath.ac.uk", + "bbox": [ + 168, + 181, + 316, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yong-Liang Yang University of Bath Bath, UK y.yang@cs.bath.ac.uk", + "bbox": [ + 364, + 181, + 545, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qi Wu \nUniversity of Adelaide \nAdelaide, Australia \nqi.wu01@adelaide.edu.au", + "bbox": [ + 593, + 181, + 802, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present ShapeScaffold, a structure-based neural network for generating colored 3D shapes based on text input. The approach, similar to providing scaffolds as internal structural supports and adding more details to them, aims to capture finer text-shape connections and improve the quality of generated shapes. Traditional text-to-shape methods often generate 3D shapes as a whole. However, humans tend to understand both shape and text as being structure-based. For example, a table is interpreted as being composed of legs, a seat, and a back; similarly, texts possess inherent linguistic structures that can be analyzed as dependency graphs, depicting the relationships between entities within the text. We believe structure-aware shape generation can bring finer text-shape connections and improve shape generation quality. However, the lack of explicit shape structure and the high freedom of text structure make cross-modality learning challenging. To address these challenges, we first build the structured shape implicit fields in an unsupervised manner. We then propose the part-level attention mechanism between shape parts and textual graph nodes to align the two modalities at the structural level. Finally, we employ a shape refiner to add further detail to the predicted structure, yielding the final results. Extensive experimentation demonstrates that our approaches outperform state-of-the-art methods in terms of both shape fidelity and shape-text matching. Our methods also allow for part-level manipulation and improved part-level completeness.", + "bbox": [ + 75, + 316, + 473, + 727 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 753, + 209, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The advance of 3D representation learning and generative models has sparked increasing interest in 3D shape generation. However, text-guided shape generation remains a challenging task. Many current approaches generate the 3D shape as a whole when no part information is utilized, while others treat text as a simple collection of words in order to provide finer guidance. Both shape and text, however, possess inherent internal structures that can be leveraged to", + "bbox": [ + 75, + 779, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/88baeeeb684e01923109f07d576ce08639792c69feaf7abf5b15f6997ef3dc84.jpg", + "image_caption": [ + "Figure 1. Text to shape generation with structure awareness. Raw description generates the initial coarse shape with structures. Further correspondence is built between shape parts and text graph nodes for fine-grained refinement." + ], + "image_footnote": [], + "bbox": [ + 506, + 287, + 888, + 462 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "improve the alignment between the two modalities.", + "bbox": [ + 498, + 551, + 839, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans are born with a sense of physical understanding of the world [3]. Studies have shown that the human visual system perceives objects as hierarchical arrangements of parts [4], comprising high-level structural properties and low-level details. In addition, humans use language to describe objects, including their appearance, structure, and functions. It has been discovered that language, shaped by the human brain over time, has evolved to have hierarchical linguistic structures in order to convey more complex meanings [10]. This suggests that structure-based reasoning is present in both the visual and language modalities. While this is intuitive and effortless for humans, it is a challenge for intelligent algorithms to understand the structural connection between these two modalities.", + "bbox": [ + 496, + 566, + 892, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this research, we aim to augment shape generation from text with an awareness of structure in both shape and language modalities. This would allow for shape generation to be guided not only by high-level descriptions, such as a sentence but also refined with part-level guidance from semantic text entities. This process is similar to how humans create visual artworks: sketching an idea's gist and then progressively refining the details. However, there are", + "bbox": [ + 496, + 780, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2715", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "several challenges to overcome. One challenge is that the shape structures are unknown unless annotated segmentation labels are provided. Even with annotation, where the process is time-consuming and labor-intensive, the definition of \"parts\" in shapes is subjective and ambiguous at multiple levels. Another challenge is that text descriptions of shapes are diverse and complex, making it difficult to optimize learning. In addition, for colored shapes, the alignment between structure and color representation must also be taken into consideration.", + "bbox": [ + 75, + 90, + 468, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the challenges above, we propose a two-stage approach. First, we leverage an unsupervised hierarchical shape structure given part hierarchy is a prevalent and effective approach for representing shape structures, as demonstrated in previous research [22, 28]. To achieve this, we introduce a decomposition model to progressively decompose the shape into parts at multiple levels of granularity, from coarse to fine, such that each level can reconstruct the entire shape. Additionally, we learn color reconstruction at the part level. Fig. 1 illustrates the process. While the structures and parts identified in the first stage hold significance within the shape domain, they do not have a corresponding relationship with the text. In the second step, we establish a connection between the shape structure and the text structure. We utilize linguistic dependency graphs, which can be used to represent the structural elements present in the text, as attention signals to form the structure-level connection. Through these two steps, we are able to improve the generation results at the part level by building upon the established structure.", + "bbox": [ + 75, + 242, + 470, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our experimentation on the ShapeNet dataset [7], we have demonstrated the efficacy of utilizing a structure-aware representation in improving the correspondence between text and shape. As a result of the part-aware refinement, we observe a significant improvement in the quality of generated shapes, particularly in terms of their constituent parts. Additionally, our approach attains state-of-the-art results regarding several commonly used metrics such as Intersection over Union (IOU), Accuracy, and Inception Score (IS). Furthermore, through further exploration of the latent shape structure space, it is demonstrated that our method has the capability of generating previously unseen shapes.", + "bbox": [ + 75, + 544, + 468, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this work are: (1) An unsupervised hierarchical decoder for building shape structural implicit field. (2) A structure-aware text-to-shape generation method that improves part-level correspondence and refinement. (3) Comprehensive experiments that demonstrate the effectiveness of our method.", + "bbox": [ + 75, + 739, + 468, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 844, + 218, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text Guided 3D Shape Generation. Deep learning has led to a surge in works exploring text-guided 3D shape or", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "scene generation [6, 8, 1, 25, 27, 13]. Recently, to achieve faithful results, some works tend to adopt pre-trained CLIP models [35, 17, 26] or diffusion-based methods [33, 23] for large-scale learning on open datasets. Our work, however, focuses on the classic ShapeNet dataset [7], examining the generation of structure-aware, colored shapes from text and comparing our results to previous baselines.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generating 3D shapes directly from text remains a challenge, particularly when attempting to concurrently generate colors. [8] proposes a method that learns a joint text-shape embedding space and then uses a generative adversarial network (GAN) [14] to generate colored voxels from text. However, this approach has not been successful in terms of shape resolution, color quality, and cross-modality consistency. To address these issues, [25] proposes a method that uses an implicit 3D representation with word-level attention to improve control over generated shape structures and colors. [27] and [13] both adopt a discrete autoencoder to capture patch-based shape priors and subsequently employ a transformer for autoregressive generation. However, it is worth noting that these methods do not take into account color information, which is an essential attribute of object appearance. Also, all these methods have treated the shape as a whole, ignoring the internal structure within the shape. Additionally, text modeling has been limited to language-level features, such as sentence and word features, rather than considering the structure of the text for structure-level matching with the shape domain.", + "bbox": [ + 496, + 198, + 892, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Structure-Aware Representation and Parsing. Although structure-based shape generation with text guidance has not been previously explored, there is a significant amount of research on learning structure-aware 3D shape representation or parsing 3D shapes into parts. For supervised methods, [22] proposes the use of a symmetry hierarchy [39] for hierarchical shape structure representation, while StructureNet [28] represents shapes using a graph neural network that considers both primitives and hierarchies. Similarly, [16] utilizes a binary parsing tree to represent the 3D structure of a cable-stayed bridge. Recently, semantic-based shape decomposition or parsing methods have been proposed in [19, 18], which use learned operations to obtain grammar-level shape parts. Unsupervised methods such as [37, 31] use convolutional neural networks (CNNs) to generate primitive shapes, e.g., cuboids or superquadrics, which abstract the input shape. BAE-NET [9] attempts to co-segment shapes into parts using a branched autoencoder architecture that disentangles common features for segmentation. However, these methods only predict the parts without considering their dependencies in the shape structure. [30] recovers the 3D objects into a hierarchy of parts directly from an RGB image using CNN, while the parts are primitive and lack details. RIM-NET [29] infers", + "bbox": [ + 496, + 537, + 892, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2716", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "hierarchical structures of 3D shapes using recursive implicit fields. It decomposes an input shape into two parts at each level, resulting in a binary tree hierarchy, where each level of the tree corresponds to an assembly of shape parts represented as implicit functions. However, the method does not consider color information and only processes one shape at a time during training. Our work uses a hierarchical tree structure for shape representation, but also includes color information for part-level implicit representation. In addition, our model is able to learn a shared latent space for a set of shapes.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text as Textual Graph Guidance. Text-to-image (T2I) generation has seen significant progress in recent years, with a focus on both generating realistic images and increasing consistency between texts and images. Some approaches use attention-based methods [40] for fine-grained alignment, while others leverage shared structures in images and texts. Scene graphs, which are derived from texts and contain expressive structural relationships between entities, have been widely used to reduce the cross-modality gap between texts and images for tasks such as text-image matching [24], fine-grained image generation [2], and image manipulation [12]. Given that both shapes and texts contain structural elements, we opt to employ textual graphs as additional guidance in conjunction with the generation of structural parts.", + "bbox": [ + 75, + 277, + 472, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 518, + 169, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 544, + 187, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The task is to generate 3D shapes with colors based on the given text description $T$ . The predicted object is denoted as $O \\in \\mathbb{R}^{N \\times 4}$ , comprising the shape occupancy values $S \\in \\mathbb{R}^{N \\times 1}$ and the color RGB values $C \\in \\mathbb{R}^{N \\times 3}$ , where $N$ is the number of sample points.", + "bbox": [ + 75, + 568, + 468, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overview of our method is illustrated in Fig. 2. Our framework consists of a text encoder $E_{T}$ , textual graph encoder $E_{G}$ , hierarchical structure-aware shape decoder $D_{T}$ and shape encoder $E_{S}$ . In contrast to general methods that generate the whole shape at once, our approach, called ShapeScaffold, aims to generate the 3D shape by considering its internal structures, enabling refinement at the part level and improvement of text-shape correspondence. To achieve this, we propose to capture the part-level information from both shape and text. For the shape, we introduce the concept of structural implicit field in the shape decoder to capture the underlying structure and relationships of shape parts and colors. As shown in Fig. 2 (middle right), the structure latent fields capture the semantic shape parts for a chair, including arm, cushion, legs, etc. For the text, in addition to encoding it as a global feature such as a sentence vector, we also propose using a linguistic tex", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tual graph parsed from the text that includes explicit entities with relations. This enables the structure latent fields to interact with textual entities at a finer level, resulting in more direct and explicit correspondence between the text description, such as \"curved armrests\" and the corresponding arm part of the generated shape. However, it is non-trivial to train a hierarchical structure-aware generator conditioned on text. We approach this task by dividing the training process into two stages: (A) for learning the structural implicit field using an autoencoder, and (B) for generating shapes guided by text, as summarized in Fig. 2. Stage B can be further divided into text encoding (B-1) and shape generation (B-2). The following describes each stage in detail.", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Shape-Specific Pre-training", + "text_level": 1, + "bbox": [ + 498, + 297, + 750, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To capture the internal semantic structures of shapes and facilitate shape decoding conditioned on text, we employ a pre-training technique using a shape-specific hierarchical decoder $D_{S}$ and a shape encoder $E_{S}$ , forming an autoencoder (see Fig. 2 (A)). The encoder $E_{S}$ , which is a convolutional neural network (CNN), takes the volume of an object $O$ as input and produces latent shape and color features to be used by the decoder. The decoder then creates an implicit field by partitioning the shape $(\\bar{f}_s)$ and color $(\\bar{f}_c)$ latents at each hierarchical level, reconstructing the latent features into parts and assembling them into the final object. An example of a successfully learned structural implicit field is shown in Fig. 2 (right), where semantic shape parts can be reconstructed from their corresponding latent positions.", + "bbox": [ + 496, + 321, + 893, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hierarchical Shape Decoder The Hierarchical Decoder consists of two modules at each level: the Structure Divider (SD) and the Part Generator (PG). Given a pair of shape/color latent features $(f_s^{l,i} \\in \\mathbb{R}^d, f_c^{l,i} \\in \\mathbb{R}^d)$ at level $l$ and position $i(1 \\leq i \\leq 2^l)$ , the SD first divides the shape latent into two child shape features and then generates the color latent for each child conditioned on the corresponding child shape latent. The SD constructs the latent space for both shape and color, as depicted by the grey and yellow squares in Fig. 2. The PG then generates shape parts one by one from its shape/color latents at a given set of point coordinates $\\mathbf{p} = \\{(x,y,z) \\mid x,y,z \\in \\mathbb{R}\\}$ . For each point $p$ , the input is the concatenation of $p$ with the latent value, and SD outputs one occupancy probability for the shape channel and three values (RGB) for the color channel. The resulting generated shape $O_l$ at level $l$ is a composite of generated parts, where a point is occupied by a part with the highest occupancy probability, provided that this probability exceeds a pre-defined threshold. The color of each point is determined based on the corresponding color values of the occupying part. Both SD and PG are implemented using multi-layer perception (MLPs). The SD is implemented using a two-layer MLP for partitioning the shape latent space", + "bbox": [ + 496, + 553, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2717", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6d54f528674ff29953f5258de83556ceed6e92a624e69882d34b7d76aea344a7.jpg", + "image_caption": [ + "(B-2) T2S: Hierarchical Generation" + ], + "image_footnote": [], + "bbox": [ + 91, + 94, + 424, + 229 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a1b3617f7ad718a23a1aa3393332ec78b391dbc9a9f69a24902cda075583b42b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 109, + 676, + 165 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/33116be9703471b80db4230b10f9ea70463e80f592ba955af3fcc77e720a6eca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 92, + 880, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9555a389380805c697c84df83a9ee72ac5495f30b273f409b10af1ea303c4e97.jpg", + "image_caption": [ + "(B-1) T2S: Text Encoding" + ], + "image_footnote": [], + "bbox": [ + 91, + 257, + 429, + 354 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3b1a71df0864ffd4d10b5db68800b64f4c7fca8b4eca02d5128d532bdf35cdfd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 166, + 880, + 375 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/45d8c22361b1a09b010ff5a3fb38fe2d2a7932e6d424d5344312b70fd1e6afe9.jpg", + "image_caption": [ + "Figure 2. Method overview. The proposed method includes two phases: shape autoencoder pre-training (A) and text-to-shape (T2S) phase including (B-1) and (B-2). Phase (A) trains the shape-specific decoder using shape/color reconstruction loss and hierarchical bond loss. Phase (B-1) utilizes the text encoder $E_{T}$ to convert text into global-level shape $(\\bar{t}_s)$ and color $(\\bar{t}_c)$ representations, as well as local-level features in the form of a textual graph $(g)$ . In Phase (B-2), the shape decoder generates the object hierarchically using the global input from Phase (B-1), with further use of a part-node attention module and refiner $(R)$ to connect part latents and graph node features for more precise guidance." + ], + "image_footnote": [], + "bbox": [ + 93, + 356, + 419, + 458 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9ad6853aa55a15ed8e6e18d0b2c15f2ba8b2e1307a08e61505bec802aa8c3f67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 375, + 880, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and a one-layer MLP for partitioning the color latent space. The PG has three layers for the shape latents and two layers for the color latents. Both SD and PG use LeakyReLU as intermediate activation layers and Sigmoid as the final one. Note SD and PG have shared weights within each level of the hierarchy tree, but different weights between the levels.", + "bbox": [ + 75, + 571, + 468, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reconstruction Loss The total reconstruction loss of shape and color is calculated by summing the individual reconstruction losses at each level using Mean Squared Error (MSE), as shown below:", + "bbox": [ + 75, + 681, + 468, + 742 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {r e c o n} = \\lambda_ {r e c o n} \\sum_ {l = 1} ^ {L} \\left(S _ {g t} - S _ {l}\\right) ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 756, + 468, + 797 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c o l o r} = \\lambda_ {c o l o r} \\sum_ {l = 1} ^ {L} \\left(C _ {g t} - C _ {l}\\right) ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 806, + 468, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S_{gt}$ and $C_{gt}$ represent the ground truth for shape and color, respectively. $S_{l}$ and $C_{l}$ are the assembled part's occupancy and color value at level $l$ .", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hierarchical Bond Loss The generated shape parts, following the hierarchical tree structure, need to be constrained by the relationships between parent and child parts in terms of geometric composition. To ensure this, a bond loss is calculated between level $l$ and the next level $l + 1$ . The bond loss is defined as the error between the occupancy value of each parent part and the maximum of the occupancy values of its two derived parts:", + "bbox": [ + 496, + 571, + 890, + 693 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {b o n d} ^ {l} = \\sum_ {i = 1} ^ {2 ^ {l}} \\left(S _ {i} ^ {l} - \\max \\left(S _ {2 i - 1} ^ {l + 1}, S _ {2 i} ^ {l + 1}\\right)\\right) ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 705, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S_{i}^{l}$ is the occupancy value of the parent part at position $i$ and level $l$ , and $S_{2i-1}^{l+1}$ and $S_{2i}^{l+1}$ are the occupancy values of the two derived child parts. The total bond loss is computed by summing the bond loss across all levels when", + "bbox": [ + 496, + 757, + 890, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nl < L \\colon L _ {b o n d} = \\lambda_ {b o n d} \\sum_ {l = 1} ^ {l < L} L _ {b o n d} ^ {l}.\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 816, + 740, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Regularization Loss To reduce the complexity of the network and avoid overfitting, we introduce regularization loss $L_{reg}$ at the last layer of each Structure Divider at each level.", + "bbox": [ + 500, + 854, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2718", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This loss is calculated as the sum of the absolute values of all weight parameters, multiplied by a small regularization parameter $\\lambda_{reg}$ . The regularization loss is given by: $L_{reg} = \\lambda_{reg}\\sum_{i=1}^{n}|w_i|$ , where $w_i$ is the $i$ -th weight parameter in the layer, and $n$ is the total number of weight parameters. Our experiments show that without using regularization loss, the generated shapes may contain parts that are over-segmented or lack reasonable shape.", + "bbox": [ + 76, + 90, + 472, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison with Existing Method Our method differs from previous approaches that utilize (hierarchical) decoders for unsupervised shape part learning (such as in [9, 29]) in several ways. Firstly, our hierarchical decoder aims to learn semantic shape part-level implicit field rather than simply segmenting the shape into more parts. For example, when generating a chair with four legs, our approach considers the legs as a whole unit rather than segmenting them individually, as we typically describe the legs as a whole rather than as individual parts. To balance the network's fitting ability with its complexity, we utilize a regularization loss. Secondly, we also construct hierarchical color latent fields, which are often overlooked in other methods, to represent colored shapes.", + "bbox": [ + 76, + 241, + 472, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Text-Guided Shape Generation", + "text_level": 1, + "bbox": [ + 76, + 468, + 352, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 2 (B) shows the text-guided shape generation network, comprising three modules: text encoder $E_{T}$ , textual graph encoder $E_{G}$ , and hierarchical shape decoder $D_{T}$ . The $D_{T}$ module is similar to the pre-trained shape-specific decoder $D_{S}$ (as described in Section 3.2), with the addition of a new part-node attention module and a refiner. This attention module allows the network to learn the local connections between shape parts/colors and textual graph entities, followed by a refiner module $R$ for final results generation. $D_{T}$ module is initialized from the pre-trained $D_{S}$ to leverage its predefined structural implicit fields and facilitate the generation process in a hierarchical manner. We provide a detailed description of each part in the following sections.", + "bbox": [ + 76, + 493, + 472, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text Global Encoding We utilize a pre-trained BERT model [11] as the text encoder base. The encoded [CLS] token feature is used to extract global-level latent features $\\bar{t}_s$ and $\\bar{t}_c$ for shape and color, respectively, which serve as the input for the hierarchical shape decoder $D$ . Most existing approaches employ word-level features $w$ as fine-grained guidance in text-conditioned image or 3D shape generation tasks [40, 25]. However, we argue that the use of structured textual representation, such as the textual graph, would be more appropriate in aligning with shape structures. As a result, we opt to use the word features to initialize the graph node representation, as described below.", + "bbox": [ + 76, + 719, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/11013f26976ef3e04cbef40cc29a0dac99ed76c9b3bca4b5861b11a5ca457b6e.jpg", + "image_caption": [ + "Figure 3. The dashed box shows a triplet $(v_{s}, v_{r}, v_{o})$ in the graph, denoting the subjective, relation, and objective nodes. The three nodes are first concatenated and then jointly processed using MLP, which outputs a new value for each node. The final updated state $v_{s}^{\\prime}$ of $v_{s}$ is the result of pooling the corresponding values from all triplets in which the node $v_{s}$ is involved." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 890, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Shape-Oriented Text Graph We use a unique text graph representation designed specifically for shapes. The graph nodes are classified into root, part, and attribute categories, which represent the class of the shape, the parts, and the attribute for root/part nodes, respectively. We use a special type node relation to connect these nodes, forming the edges. For example, the text \"brown wooden chair with curved armrests\" would be parsed as (chair: brown, wooden), (armrests: curved), and their relation triplet (chair, with, armrests) in the (subjective, relation, objective) convention. To construct graphs from the original text, we first use the spaCy library [15] to parse the text's semantic dependencies. These raw dependencies contain many linguistic terms and do not correspond well with real-world structures. To obtain more clear and shape-oriented graph formation, we introduce a custom parser targeting 3D shapes based on [36]. To ensure that the graph remains connected, we connect all part nodes with the root node using the with relation if no linguistic relation is found. The lower left of Fig. 2 illustrates these two forms of text. Finally, we use the resulting relation triplets for training. Further details can be found in the supplementary materials.", + "bbox": [ + 496, + 344, + 893, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial-Aware Graph Encoding One challenge in using a shape-oriented graph, where entity or attribute nodes are connected through relation nodes, is the lack of a canonical grid for embedding the graph since nodes in such a graph exist in a non-Euclidean space. Existing methods based on Graph Convolution Networks (GCN) [21] typically rely on an adjacency matrix to represent the relations between nodes, which reduces their ability to represent spatial relations. To address the issue, we designed a novel Spatial-Aware Graph Encoder (SAGE) that differentiably encodes relation nodes in order to capture the structural relations between nodes. Specifically, we represent each graph triplet $g$ in the form $(v_{s}, v_{r}, v_{o})$ , where $v_{s}$ , $v_{o}$ , and $v_{r}$ represent", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2719", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/82c3b4940e0d9e0d8cb72f13adbfd6a08763ad5fb401f2817363ec464ed30be2.jpg", + "image_caption": [ + "Figure 4. The Part-Node attention. The attention map is obtained by the dot-product between the features from the intermediate part latent feature $(f_s^l)$ at level $l$ and the graph node features $g$ ." + ], + "image_footnote": [], + "bbox": [ + 117, + 89, + 454, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the values of subjective, objective, and relation nodes in a relationship, respectively. This makes the graph directional through the order of the nodes between the relation node, like chair has arms. As shown in Fig. 3, the computational layer in SAGE jointly encodes the nodes and their relation in each triplet, followed by a pooling layer that fuses the features of each node involved in more than one triplet.", + "bbox": [ + 75, + 305, + 468, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Part-Node Attention We introduce the Part-Node Attention modules for learning the relationship between intermediate parts/colors with textual graph nodes. These modules are based on the multi-head attention mechanism presented in [38]. For level $l$ , given the intermediate shape latents $f_{s}^{l} \\in \\mathbb{R}^{2^{l} \\times d}$ as query, and the encoded graph node features, $g \\in \\mathbb{R}^{n \\times d}$ ( $n$ is node number), as key and value, the calculation inside an attention head is:", + "bbox": [ + 75, + 431, + 468, + 551 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} _ {s} ^ {l} = \\operatorname {s o f t m a x} \\left(\\frac {Q K}{\\sqrt {d _ {h}}}\\right) V, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 566, + 468, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $Q = F_{Q}(f_{s}^{l})$ , $K = F_{K}(g)$ , $V = F_{V}(g)$ , and $F_{Q}, F_{K}$ , and $F_{V}$ are functions implemented using MLPs. $d$ is the feature dimension for $g$ and $f_{s}^{l}$ , while $d_{h}$ represents the dimension used within an attention head. The SoftMax operation is applied across the graph nodes, resulting in a weighted aggregation of the graph node features as the new intermediate part latents, $\\hat{f}_{s}^{l}$ . This process is visualized in Fig. 4. Also, attention is applied between the paired color features $f_{c}^{l}$ with the graph nodes using a separate module. In practice, we use the attention module at level 2 of the hierarchical decoder due to the generated semantically detailed shape components aligned with graph nodes.", + "bbox": [ + 75, + 607, + 468, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Joint Refiner The hierarchical bond loss, introduced in Section 3.2, serves as the foundation for structure-aware generation by constraining the composition of shape parts to follow a hierarchical tree structure. However, we posit that this also negatively impacts the generated results, as demonstrated in the perception-distortion tradeoff [5]. To", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mitigate this issue, we introduce a refiner $R$ in one level that uses the level latents jointly to generate the final shape. The refiner is also an implicit shape decoder with six layers of MLPs. The final shape output is accompanied by an additional refiner reconstruction loss $L_{R}$ but without the need to consider the hierarchical bond, which allows for the incorporation of additional detail in the final shape.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Strategy and Losses Training text-guided hierarchical shape generation directly presents significant challenges considering the big discrepancy between text and shape modalities. To address this, we propose a three-step strategy to ensure stable and effective training. First, we collect the shape and color latents from the shape encoder $E_{S}$ as guidance and train the text global feature encoder $E_{T}$ alone. The loss is using MSE: $L_{\\text{latent}} = \\lambda_{\\text{latent}}^{s} (\\bar{t}_{s} - \\bar{f}_{s})^{2} + \\lambda_{\\text{latent}}^{c} (\\bar{t}_{c} - \\bar{f}_{c})^{2}$ , where $\\bar{t}_{s}$ and $\\bar{t}_{c}$ are encoded by text encoder while $\\bar{f}_{s}$ and $\\bar{f}_{c}$ are encoded by shape encoder. Second, we train the decoder with attention modules, freezing the decoder during the first ten epochs to allow for warmed-up training. Finally, we train the entire network together. The total loss then includes the aforementioned shape/color reconstruction losses and the bond loss in shape-specific pre-training (see Sec. 3.2), and the refiner loss.", + "bbox": [ + 496, + 224, + 890, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 484, + 632, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section introduces our experimental design and implementation, followed by a comprehensive analysis of the results from various aspects and an ablation study. Supplementary material provides additional results and analysis.", + "bbox": [ + 496, + 512, + 890, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings and Implementation", + "text_level": 1, + "bbox": [ + 498, + 587, + 867, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments using the text-shape dataset [8], which is derived from the 3D shape dataset, ShapeNet [7], and augmented with textual annotations. The dataset comprises approximately 15,000 shapes and 75,000 textual descriptions, encompassing the table and chair shape classes. We also adhere to the same training and testing splits as previously established.", + "bbox": [ + 496, + 611, + 890, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement our framework in PyTorch [32]. We set $N = 4096$ as the sampled points. We pre-train the shape-specific decoder for 1000 epochs and the text-guided model for another 500 epochs. The learning rate is constant $1e^{-4}$ using Adam [20] for all phases. The loss weights are set as follows: $\\lambda_{\\text{recon}} = 10.0$ , $\\lambda_{\\text{color}} = 1.0$ , $\\lambda_{\\text{bond}} = 1.0$ , $\\lambda_{\\text{reg}} = 1e^{-4}$ , $\\lambda_{\\text{latent}}^s = 10.0$ and $\\lambda_{\\text{latent}}^c = 1.0$ . Detailed training strategy can be found in Sec. 3.3. Additionally, as our model outputs shapes at different levels, we select the results from level 2, as it is at this level that most shapes' parts have been parsed and refined. Level 3 is used as a spare level for monitoring no further parts are parsed.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2720", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a39f94ae94af3074bdd582afa710acfdb21f8f2c1123064e0bb4e9809bab8adc.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Quantitative comparison with state-of-the-art. $\\uparrow$ and $\\downarrow$ indicate higher or lower is better, respectively." + ], + "table_body": "
IoU (↑)IS (↑)EMD (↓)Acc (↑)
Text2Shape [8]9.641.960.444397.37
T2S-Implicit [25]12.211.970.207197.48
ours13.651.980.176797.5
", + "bbox": [ + 81, + 89, + 467, + 152 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Text-Guided Shape Generation", + "text_level": 1, + "bbox": [ + 76, + 213, + 352, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with the State-of-the-art We compare our method to two existing text-conditioned colored shape generation methods, Text2Shape [8] and T2S-Implicit [25]. We employ the same quantitative metrics for evaluation, including (1) Mean intersection-over-union (IoU) for shape occupancy measurement, regardless of color; (2) Earth Mover's Distance (EMD) for color distribution evaluation; (3) Inception Score (IS) to quantify the realism of the results; and (4) Accuracy (Acc) computed by a pre-trained shape classifier to indicate whether the predicted class matches the ground truth. To ensure fairness in comparison, we convert the generated shapes to volumes of the same size as in [8, 25] before calculating the quantitative values. The comparison results are presented in Table 1, where the numbers of the existing methods are directly taken from the respective papers. Our method exhibits superior performance across all metrics. The qualitative result presented in Fig. 5 provides further evidence of the effectiveness of our approach. As can be observed, our method achieves a higher degree of correspondence between the predicted shape structure and the text description, as demonstrated in Row 1 of the figure with the example of \"wood armrests\". Additionally, our approach exhibits an increased level of completeness in part generation, as evidenced by the absence of incomplete parts or extra artifacts, which are clearly labeled with red circles in the results generated by the existing methods.", + "bbox": [ + 76, + 241, + 472, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Part-level Refinement As T2S-Implicit [25] also employs an implicit decoder similar to our approach, we conduct a comparison with it to highlight the distinctions, particularly with regard to part-level details. The results are presented in high resolution in Fig. 6. We observed that [25] tends to struggle with smaller or complex parts, such as the legs of a sofa or the wheels of an office chair. In contrast, our method is capable of interpreting these parts explicitly, resulting in more realistic part-level generation. Another advantage of our approach is that the generation of a single part does not affect the other parts, which is a common issue in previous methods. For example, generating \"silver legs\" may result in the bottom of the sofa appearing slightly silver in the connection area, creating a noticeable color artifact.", + "bbox": [ + 76, + 672, + 472, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/600bd8836ac32bd08e535769f6dbe9df18258f8108306b5fc709ea2e18933ab8.jpg", + "image_caption": [ + "Figure 5. Visualization of the results of our method in comparison to Text2Shape [8] and T2S-Implicit [25]. Red circles indicate incomplete parts or extra artifacts." + ], + "image_footnote": [], + "bbox": [ + 498, + 89, + 890, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/99197c7aa1e4264c34473f139217f16785164e8c1644bc0b8ba05c3dde6312e1.jpg", + "image_caption": [ + "Figure 6. Comparison of part-level refinement." + ], + "image_footnote": [], + "bbox": [ + 500, + 400, + 893, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Part-level Manipulation The ability to learn structure-based 3D shape representations and align them with textual descriptions enables precise and convenient manipulation of shapes. In the process of generating a shape from text, our method allows for manipulation at various levels, including modification of the part-level structure, attributes, or colors. The results of this manipulation are illustrated in Fig. 7. The changes made in the text appear to reflect the changes in part structure (a), part attribute (c), and part-related color refinement (b), suggesting a possible consistency between the constituent parts of the shape and the corresponding text.", + "bbox": [ + 496, + 577, + 890, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 755, + 663, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct extensive ablation studies to validate the effectiveness of the proposed hierarchical structure-based representation, graph-part attention, and joint refiner. We also use Frechet Inception Distance (FID) and CLIP [34] to evaluate the quality of the generated shapes and the similarity between the generated shapes and their corresponding text descriptions, respectively. More detailed information on the metrics can be found in the supplementary material. The re", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2721", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/07a74e9bcb463d9fc3aade9582e90d56c1640aca27fcbe5497dc5d57e4caefe4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLevel 1Level 2Level 3
IOU↑FID↓CLIP↑IOU↑FID↓CLIP↑IOU↑FID↓CLIP↑
HierD [base]11.2017.4745.4311.0717.5245.3610.8517.6043.92
FlatD10.3819.5540.72------
HierD + WordAttn (L2)11.1317.4244.8811.6616.4647.4310.7217.6946.10
HierD + GraphAttn (L2)11.2917.3645.0112.6316.3150.6412.4116.7050.01
HierD + GraphAttn (L1)11.5416.8652.7511.3017.0951.4011.2717.1250.56
HierD + GraphAttn (L2) + Refiner(L2) [Full]11.1917.3847.2713.6515.8154.8912.2216.3452.20
", + "bbox": [ + 91, + 88, + 880, + 207 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Results of the ablation study on output levels. The base model, HierD, is the proposed hierarchical decoder using sentence-level guidance only. L1/L2 indicates component application at Level 1/Level 2. ↑/↓ denote better higher/ lower results, respectively.", + "bbox": [ + 76, + 208, + 892, + 238 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/427d5cbf36a613a16552bc79389deeb980b69b58291ef22749439213715305f5.jpg", + "image_caption": [ + "Figure 7. Manipulation of text results in corresponding changes at the part level. The second row shows the predicted structure from Level 1 to 2. (a) illustrates the removal of a part, (b) shows a change in part color, and (c) depicts a change in the part attribute." + ], + "image_footnote": [], + "bbox": [ + 78, + 265, + 467, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sults of our ablation studies are presented in Table 2. We use the Hierarchical Decoder with sentence-level guidance (HierD) as a base model, and then add various modules such as word-level attention (WordAttn), our proposed graph-part attention (GraphAttn), and the joint refiner to different levels of the hierarchy (L1/L2 for Level 1 and 2). We present the results for each level for an in-depth analysis of the impact of the hierarchy level on the performance of the model.", + "bbox": [ + 75, + 513, + 468, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hierarchical Representation We compare HierD to a flattened decoder (FlatD) which directly outputs 8 parts at Level 1. Utilizing the same settings as HierD, FlatD demonstrates inferior performance. Our qualitative analysis reveals that FlatD is more prone to overfitting, resulting in the generation of fewer parts that are overly complex geometrically, instead of generating finer parts.", + "bbox": [ + 75, + 654, + 468, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Graph-based Attention We have observed the clear advantages of utilizing graph-based attention over word-based attention. Specifically, when applied to Level 2, the use of GraphAttn results in superior performance across all metrics. Furthermore, the results from Level 3 also demonstrate improved performance. This can be attributed to the semantic alignment of the entity nodes in the graph with the intermediate parts. However, when applied to Level 1, the use of", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GraphAttn does not yield similar results as in Level 2. We speculate that this discrepancy may be due to the coarser nature of the parts at Level 1, which hinders the ability to establish effective connections between shape parts and textual graph entities.", + "bbox": [ + 496, + 261, + 892, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Joint Refiner The utilization of a refiner as the final shape output in our method is driven by the observation of the perception-distortion tradeoff [5]. As the level increases, we have noted a slight decline in performance as compared to previous levels. For instance, the IOU in Level 2 is consistently lower than that in Level 1. We posit that the hierarchical bond loss employed in our approach results in a stronger structural connection, but may impede the generation of detailed features on top of the structure. The incorporation of a refiner, on the other hand, has resulted in a notable improvement in all metrics.", + "bbox": [ + 496, + 359, + 892, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion, limitation, and future work", + "text_level": 1, + "bbox": [ + 498, + 541, + 856, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The ShapeScaffold is a method for generating shapes guided by text, with the understanding that both shape and text possess internal structures that can be aligned to improve the outcome. We represent shapes as hierarchical structures and model text as scene graphs. Our part-node attention mechanism allows us to learn correspondences between shape and text at various hierarchical levels, which has been shown to effectively establish connections between shape parts/colors and linguistic entities. The experiments have demonstrated the superiority of our approach in both shape generation and text-shape/color consistency.", + "bbox": [ + 496, + 566, + 890, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Several limitations should be acknowledged. Firstly, the current text-to-shape dataset is limited to chairs and tables, so further research is needed to explore the generalization of our method to other shape classes. Secondly, the results may be influenced by the parts priority problem, where incorrect occupancy state of parts junctions may result in unreasonable visualization. Therefore, further incorporation of prior interpretations of shapes should be considered in future studies. Lastly, it would be beneficial to explore scene generation that takes structure into account, with a focus on the composition of different shapes as an outer structure.", + "bbox": [ + 496, + 734, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2722", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by RCUK grant CAMERA (EP/M023281/1, EP/T022523/1), Centre for Augmented Reasoning (CAR) at the Australian Institute for Machine Learning, and a gift from Adobe.", + "bbox": [ + 76, + 90, + 472, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 162, + 171, + 178 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Judy Fan, Robert Hawkins, Noah Goodman, and Leonidas J Guibas. Shapeplot: Learning language for shape differentiation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8938-8947, 2019. 2", + "[2] Oron Ashual and Lior Wolf. Specifying object attributes and relations in interactive scene generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4561-4569, 2019. 3", + "[3] Renée Baillargeon. Infants' physical world. Current directions in psychological science, 13(3):89-94, 2004. 1", + "[4] Irving Biederman. Human image understanding: Recent research and a theory. Computer vision, graphics, and image processing, 32(1):29-73, 1985. 1", + "[5] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 6, 8", + "[6] Angel Chang, Manolis Savva, and Christopher D Manning. Learning spatial knowledge for text to 3d scene generation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 2028-2038, 2014. 2", + "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 6", + "[8] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian conference on computer vision, pages 100-116. Springer, 2018. 2, 6, 7", + "[9] Zhiqin Chen, Kangxue Yin, Matthew Fisher, Siddhartha Chaudhuri, and Hao Zhang. Bae-net: Branched autoencoder for shape co-segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8490-8499, 2019. 2, 5", + "[10] Morten H Christiansen and Nick Chater. Language as shaped by the brain. Behavioral and brain sciences, 31(5):489-509, 2008. 1", + "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5", + "[12] Helisa Dhamo, Azade Farshad, Iro Laina, Nassir Navab, Gregory D Hager, Federico Tombari, and Christian Rupprecht. Semantic image manipulation using scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5213-5222, 2020. 3", + "[13] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecraft: A recursive text-conditioned 3d shape" + ], + "bbox": [ + 78, + 188, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "generation model. arXiv preprint arXiv:2207.09446, 2022. 2", + "[14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proceedings of the 27th International Conference on Neural Information Processing Systems - Volume 2, NIPS'14, page 2672-2680, 2014. 2", + "[15] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. spaCy: Industrial-strength Natural Language Processing in Python, 2020. 5", + "[16] Fangqiao Hu, Jin Zhao, Yong Huang, and Hui Li. Learning structural graph layouts and 3d shapes for long span bridges 3d reconstruction. arXiv preprint arXiv:1907.03387, 2019. 2", + "[17] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2", + "[18] R. Kenny Jones, Aalia Habib, Rana Hanocka, and Daniel Ritchie. The neurally-guided shape parser: Grammar-based labeling of 3d shape regions with approximate inference. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[19] R Kenny Jones, Aalia Habib, and Daniel Ritchie. Shred: 3d shape region decomposition with learned local operations. arXiv preprint arXiv:2206.03480, 2022. 2", + "[20] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 6", + "[21] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 5", + "[22] Jun Li, Kai Xu, Siddhartha Chaudhuri, Ersin Yumer, Hao Zhang, and Leonidas Guibas. Grass: Generative recursive autoencoders for shape structures. ACM Transactions on Graphics (TOG), 36(4):1-14, 2017. 2", + "[23] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440, 2022. 2", + "[24] Chunxiao Liu, Zhendong Mao, Tianzhu Zhang, Hongtao Xie, Bin Wang, and Yongdong Zhang. Graph structured network for image-text matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10921-10930, 2020. 3", + "[25] Zhengzhe Liu, Yi Wang, Xiaojuan Qi, and Chi-Wing Fu. Towards implicit text-guided 3d shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17896-17906, 2022. 2, 5, 7", + "[26] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2723", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2", + "[27] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 306-315, 2022. 2", + "[28] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 2", + "[29] Chengjie Niu, Manyi Li, Kai Xu, and Hao Zhang. Rim-net: Recursive implicit fields for unsupervised learning of hierarchical shape structures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11779-11788, 2022. 2, 5", + "[30] Despoina Paschalidou, Luc Van Gool, and Andreas Geiger. Learning unsupervised hierarchical part decomposition of 3d objects from a single rgb image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1060-1070, 2020. 2", + "[31] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics revisited: Learning 3d shape parsing beyond cuboids. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10344–10353, 2019. 2", + "[32] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 6", + "[33] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 7", + "[35] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 2", + "[36] Richard Socher, Cliff C Lin, Chris Manning, and Andrew Y Ng. Parsing natural scenes and natural language with recursive neural networks. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 129–136, 2011. 5" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Shubham Tulsiani, Hao Su, Leonidas J Guibas, Alexei A Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2635-2643, 2017. 2", + "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 6", + "[39] Yanzhen Wang, Kai Xu, Jun Li, Hao Zhang, Ariel Shamir, Ligang Liu, Zhiquan Cheng, and Yueshan Xiong. Symmetry hierarchy of man-made objects. In Computer graphics forum, volume 30, pages 287-296. Wiley Online Library, 2011. 2", + "[40] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 3, 5" + ], + "bbox": [ + 501, + 92, + 893, + 372 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2724", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/ShapeScaffolder_ Structure-Aware 3D Shape Generation from Text/3212f743-a520-438c-b4e2-ff41229e0a86_model.json b/2023/ShapeScaffolder_ Structure-Aware 3D Shape Generation from Text/3212f743-a520-438c-b4e2-ff41229e0a86_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b390356b9bc10f2a1964a898e7626a32eb4f3b72 --- /dev/null +++ b/2023/ShapeScaffolder_ Structure-Aware 3D Shape Generation from Text/3212f743-a520-438c-b4e2-ff41229e0a86_model.json @@ -0,0 +1,1771 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.146, + 0.131, + 0.825, + 0.154 + ], + "angle": 0, + "content": "ShapeScaffold: Structure-Aware 3D Shape Generation from Text" + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.182, + 0.318, + 0.251 + ], + "angle": 0, + "content": "Xi Tian \nUniversity of Bath \nBath, UK \nxt275@bath.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.366, + 0.182, + 0.546, + 0.251 + ], + "angle": 0, + "content": "Yong-Liang Yang University of Bath Bath, UK y.yang@cs.bath.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.594, + 0.182, + 0.803, + 0.251 + ], + "angle": 0, + "content": "Qi Wu \nUniversity of Adelaide \nAdelaide, Australia \nqi.wu01@adelaide.edu.au" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.474, + 0.728 + ], + "angle": 0, + "content": "We present ShapeScaffold, a structure-based neural network for generating colored 3D shapes based on text input. The approach, similar to providing scaffolds as internal structural supports and adding more details to them, aims to capture finer text-shape connections and improve the quality of generated shapes. Traditional text-to-shape methods often generate 3D shapes as a whole. However, humans tend to understand both shape and text as being structure-based. For example, a table is interpreted as being composed of legs, a seat, and a back; similarly, texts possess inherent linguistic structures that can be analyzed as dependency graphs, depicting the relationships between entities within the text. We believe structure-aware shape generation can bring finer text-shape connections and improve shape generation quality. However, the lack of explicit shape structure and the high freedom of text structure make cross-modality learning challenging. To address these challenges, we first build the structured shape implicit fields in an unsupervised manner. We then propose the part-level attention mechanism between shape parts and textual graph nodes to align the two modalities at the structural level. Finally, we employ a shape refiner to add further detail to the predicted structure, yielding the final results. Extensive experimentation demonstrates that our approaches outperform state-of-the-art methods in terms of both shape fidelity and shape-text matching. Our methods also allow for part-level manipulation and improved part-level completeness." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.755, + 0.21, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The advance of 3D representation learning and generative models has sparked increasing interest in 3D shape generation. However, text-guided shape generation remains a challenging task. Many current approaches generate the 3D shape as a whole when no part information is utilized, while others treat text as a simple collection of words in order to provide finer guidance. Both shape and text, however, possess inherent internal structures that can be leveraged to" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.289, + 0.889, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.469, + 0.893, + 0.525 + ], + "angle": 0, + "content": "Figure 1. Text to shape generation with structure awareness. Raw description generates the initial coarse shape with structures. Further correspondence is built between shape parts and text graph nodes for fine-grained refinement." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.552, + 0.84, + 0.567 + ], + "angle": 0, + "content": "improve the alignment between the two modalities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.893, + 0.779 + ], + "angle": 0, + "content": "Humans are born with a sense of physical understanding of the world [3]. Studies have shown that the human visual system perceives objects as hierarchical arrangements of parts [4], comprising high-level structural properties and low-level details. In addition, humans use language to describe objects, including their appearance, structure, and functions. It has been discovered that language, shaped by the human brain over time, has evolved to have hierarchical linguistic structures in order to convey more complex meanings [10]. This suggests that structure-based reasoning is present in both the visual and language modalities. While this is intuitive and effortless for humans, it is a challenge for intelligent algorithms to understand the structural connection between these two modalities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.902 + ], + "angle": 0, + "content": "In this research, we aim to augment shape generation from text with an awareness of structure in both shape and language modalities. This would allow for shape generation to be guided not only by high-level descriptions, such as a sentence but also refined with part-level guidance from semantic text entities. This process is similar to how humans create visual artworks: sketching an idea's gist and then progressively refining the details. However, there are" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2715" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.242 + ], + "angle": 0, + "content": "several challenges to overcome. One challenge is that the shape structures are unknown unless annotated segmentation labels are provided. Even with annotation, where the process is time-consuming and labor-intensive, the definition of \"parts\" in shapes is subjective and ambiguous at multiple levels. Another challenge is that text descriptions of shapes are diverse and complex, making it difficult to optimize learning. In addition, for colored shapes, the alignment between structure and color representation must also be taken into consideration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.471, + 0.544 + ], + "angle": 0, + "content": "To address the challenges above, we propose a two-stage approach. First, we leverage an unsupervised hierarchical shape structure given part hierarchy is a prevalent and effective approach for representing shape structures, as demonstrated in previous research [22, 28]. To achieve this, we introduce a decomposition model to progressively decompose the shape into parts at multiple levels of granularity, from coarse to fine, such that each level can reconstruct the entire shape. Additionally, we learn color reconstruction at the part level. Fig. 1 illustrates the process. While the structures and parts identified in the first stage hold significance within the shape domain, they do not have a corresponding relationship with the text. In the second step, we establish a connection between the shape structure and the text structure. We utilize linguistic dependency graphs, which can be used to represent the structural elements present in the text, as attention signals to form the structure-level connection. Through these two steps, we are able to improve the generation results at the part level by building upon the established structure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.47, + 0.741 + ], + "angle": 0, + "content": "In our experimentation on the ShapeNet dataset [7], we have demonstrated the efficacy of utilizing a structure-aware representation in improving the correspondence between text and shape. As a result of the part-aware refinement, we observe a significant improvement in the quality of generated shapes, particularly in terms of their constituent parts. Additionally, our approach attains state-of-the-art results regarding several commonly used metrics such as Intersection over Union (IOU), Accuracy, and Inception Score (IS). Furthermore, through further exploration of the latent shape structure space, it is demonstrated that our method has the capability of generating previously unseen shapes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.741, + 0.47, + 0.831 + ], + "angle": 0, + "content": "The main contributions of this work are: (1) An unsupervised hierarchical decoder for building shape structural implicit field. (2) A structure-aware text-to-shape generation method that improves part-level correspondence and refinement. (3) Comprehensive experiments that demonstrate the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.22, + 0.86 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Text Guided 3D Shape Generation. Deep learning has led to a surge in works exploring text-guided 3D shape or" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "scene generation [6, 8, 1, 25, 27, 13]. Recently, to achieve faithful results, some works tend to adopt pre-trained CLIP models [35, 17, 26] or diffusion-based methods [33, 23] for large-scale learning on open datasets. Our work, however, focuses on the classic ShapeNet dataset [7], examining the generation of structure-aware, colored shapes from text and comparing our results to previous baselines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.893, + 0.516 + ], + "angle": 0, + "content": "Generating 3D shapes directly from text remains a challenge, particularly when attempting to concurrently generate colors. [8] proposes a method that learns a joint text-shape embedding space and then uses a generative adversarial network (GAN) [14] to generate colored voxels from text. However, this approach has not been successful in terms of shape resolution, color quality, and cross-modality consistency. To address these issues, [25] proposes a method that uses an implicit 3D representation with word-level attention to improve control over generated shape structures and colors. [27] and [13] both adopt a discrete autoencoder to capture patch-based shape priors and subsequently employ a transformer for autoregressive generation. However, it is worth noting that these methods do not take into account color information, which is an essential attribute of object appearance. Also, all these methods have treated the shape as a whole, ignoring the internal structure within the shape. Additionally, text modeling has been limited to language-level features, such as sentence and word features, rather than considering the structure of the text for structure-level matching with the shape domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.893, + 0.903 + ], + "angle": 0, + "content": "3D Structure-Aware Representation and Parsing. Although structure-based shape generation with text guidance has not been previously explored, there is a significant amount of research on learning structure-aware 3D shape representation or parsing 3D shapes into parts. For supervised methods, [22] proposes the use of a symmetry hierarchy [39] for hierarchical shape structure representation, while StructureNet [28] represents shapes using a graph neural network that considers both primitives and hierarchies. Similarly, [16] utilizes a binary parsing tree to represent the 3D structure of a cable-stayed bridge. Recently, semantic-based shape decomposition or parsing methods have been proposed in [19, 18], which use learned operations to obtain grammar-level shape parts. Unsupervised methods such as [37, 31] use convolutional neural networks (CNNs) to generate primitive shapes, e.g., cuboids or superquadrics, which abstract the input shape. BAE-NET [9] attempts to co-segment shapes into parts using a branched autoencoder architecture that disentangles common features for segmentation. However, these methods only predict the parts without considering their dependencies in the shape structure. [30] recovers the 3D objects into a hierarchy of parts directly from an RGB image using CNN, while the parts are primitive and lack details. RIM-NET [29] infers" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2716" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "hierarchical structures of 3D shapes using recursive implicit fields. It decomposes an input shape into two parts at each level, resulting in a binary tree hierarchy, where each level of the tree corresponds to an assembly of shape parts represented as implicit functions. However, the method does not consider color information and only processes one shape at a time during training. Our work uses a hierarchical tree structure for shape representation, but also includes color information for part-level implicit representation. In addition, our model is able to learn a shared latent space for a set of shapes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.278, + 0.473, + 0.506 + ], + "angle": 0, + "content": "Text as Textual Graph Guidance. Text-to-image (T2I) generation has seen significant progress in recent years, with a focus on both generating realistic images and increasing consistency between texts and images. Some approaches use attention-based methods [40] for fine-grained alignment, while others leverage shared structures in images and texts. Scene graphs, which are derived from texts and contain expressive structural relationships between entities, have been widely used to reduce the cross-modality gap between texts and images for tasks such as text-image matching [24], fine-grained image generation [2], and image manipulation [12]. Given that both shapes and texts contain structural elements, we opt to employ textual graphs as additional guidance in conjunction with the generation of structural parts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.52, + 0.17, + 0.535 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.545, + 0.188, + 0.56 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.645 + ], + "angle": 0, + "content": "The task is to generate 3D shapes with colors based on the given text description \\(T\\). The predicted object is denoted as \\(O \\in \\mathbb{R}^{N \\times 4}\\), comprising the shape occupancy values \\(S \\in \\mathbb{R}^{N \\times 1}\\) and the color RGB values \\(C \\in \\mathbb{R}^{N \\times 3}\\), where \\(N\\) is the number of sample points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The overview of our method is illustrated in Fig. 2. Our framework consists of a text encoder \\( E_{T} \\), textual graph encoder \\( E_{G} \\), hierarchical structure-aware shape decoder \\( D_{T} \\) and shape encoder \\( E_{S} \\). In contrast to general methods that generate the whole shape at once, our approach, called ShapeScaffold, aims to generate the 3D shape by considering its internal structures, enabling refinement at the part level and improvement of text-shape correspondence. To achieve this, we propose to capture the part-level information from both shape and text. For the shape, we introduce the concept of structural implicit field in the shape decoder to capture the underlying structure and relationships of shape parts and colors. As shown in Fig. 2 (middle right), the structure latent fields capture the semantic shape parts for a chair, including arm, cushion, legs, etc. For the text, in addition to encoding it as a global feature such as a sentence vector, we also propose using a linguistic tex" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.289 + ], + "angle": 0, + "content": "tual graph parsed from the text that includes explicit entities with relations. This enables the structure latent fields to interact with textual entities at a finer level, resulting in more direct and explicit correspondence between the text description, such as \"curved armrests\" and the corresponding arm part of the generated shape. However, it is non-trivial to train a hierarchical structure-aware generator conditioned on text. We approach this task by dividing the training process into two stages: (A) for learning the structural implicit field using an autoencoder, and (B) for generating shapes guided by text, as summarized in Fig. 2. Stage B can be further divided into text encoding (B-1) and shape generation (B-2). The following describes each stage in detail." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.299, + 0.75, + 0.315 + ], + "angle": 0, + "content": "3.2. Shape-Specific Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.895, + 0.534 + ], + "angle": 0, + "content": "To capture the internal semantic structures of shapes and facilitate shape decoding conditioned on text, we employ a pre-training technique using a shape-specific hierarchical decoder \\(D_{S}\\) and a shape encoder \\(E_{S}\\), forming an autoencoder (see Fig. 2 (A)). The encoder \\(E_{S}\\), which is a convolutional neural network (CNN), takes the volume of an object \\(O\\) as input and produces latent shape and color features to be used by the decoder. The decoder then creates an implicit field by partitioning the shape \\((\\bar{f}_s)\\) and color \\((\\bar{f}_c)\\) latents at each hierarchical level, reconstructing the latent features into parts and assembling them into the final object. An example of a successfully learned structural implicit field is shown in Fig. 2 (right), where semantic shape parts can be reconstructed from their corresponding latent positions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Hierarchical Shape Decoder The Hierarchical Decoder consists of two modules at each level: the Structure Divider (SD) and the Part Generator (PG). Given a pair of shape/color latent features \\((f_s^{l,i} \\in \\mathbb{R}^d, f_c^{l,i} \\in \\mathbb{R}^d)\\) at level \\(l\\) and position \\(i(1 \\leq i \\leq 2^l)\\), the SD first divides the shape latent into two child shape features and then generates the color latent for each child conditioned on the corresponding child shape latent. The SD constructs the latent space for both shape and color, as depicted by the grey and yellow squares in Fig. 2. The PG then generates shape parts one by one from its shape/color latents at a given set of point coordinates \\(\\mathbf{p} = \\{(x,y,z) \\mid x,y,z \\in \\mathbb{R}\\}\\). For each point \\(p\\), the input is the concatenation of \\(p\\) with the latent value, and SD outputs one occupancy probability for the shape channel and three values (RGB) for the color channel. The resulting generated shape \\(O_l\\) at level \\(l\\) is a composite of generated parts, where a point is occupied by a part with the highest occupancy probability, provided that this probability exceeds a pre-defined threshold. The color of each point is determined based on the corresponding color values of the occupying part. Both SD and PG are implemented using multi-layer perception (MLPs). The SD is implemented using a two-layer MLP for partitioning the shape latent space" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2717" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.095, + 0.425, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.43, + 0.094, + 0.674, + 0.107 + ], + "angle": 0, + "content": "(B-2) T2S: Hierarchical Generation" + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.111, + 0.677, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.093, + 0.882, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.238, + 0.274, + 0.254 + ], + "angle": 0, + "content": "(B-1) T2S: Text Encoding" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.258, + 0.43, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.167, + 0.882, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.357, + 0.421, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.375, + 0.882, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.464, + 0.892, + 0.548 + ], + "angle": 0, + "content": "Figure 2. Method overview. The proposed method includes two phases: shape autoencoder pre-training (A) and text-to-shape (T2S) phase including (B-1) and (B-2). Phase (A) trains the shape-specific decoder using shape/color reconstruction loss and hierarchical bond loss. Phase (B-1) utilizes the text encoder \\( E_{T} \\) to convert text into global-level shape \\( (\\bar{t}_s) \\) and color \\( (\\bar{t}_c) \\) representations, as well as local-level features in the form of a textual graph \\( (g) \\). In Phase (B-2), the shape decoder generates the object hierarchically using the global input from Phase (B-1), with further use of a part-node attention module and refiner \\( (R) \\) to connect part latents and graph node features for more precise guidance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.47, + 0.664 + ], + "angle": 0, + "content": "and a one-layer MLP for partitioning the color latent space. The PG has three layers for the shape latents and two layers for the color latents. Both SD and PG use LeakyReLU as intermediate activation layers and Sigmoid as the final one. Note SD and PG have shared weights within each level of the hierarchy tree, but different weights between the levels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.47, + 0.743 + ], + "angle": 0, + "content": "Reconstruction Loss The total reconstruction loss of shape and color is calculated by summing the individual reconstruction losses at each level using Mean Squared Error (MSE), as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.757, + 0.469, + 0.798 + ], + "angle": 0, + "content": "\\[\nL _ {r e c o n} = \\lambda_ {r e c o n} \\sum_ {l = 1} ^ {L} \\left(S _ {g t} - S _ {l}\\right) ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.808, + 0.469, + 0.849 + ], + "angle": 0, + "content": "\\[\nL _ {c o l o r} = \\lambda_ {c o l o r} \\sum_ {l = 1} ^ {L} \\left(C _ {g t} - C _ {l}\\right) ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\( S_{gt} \\) and \\( C_{gt} \\) represent the ground truth for shape and color, respectively. \\( S_{l} \\) and \\( C_{l} \\) are the assembled part's occupancy and color value at level \\( l \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.694 + ], + "angle": 0, + "content": "Hierarchical Bond Loss The generated shape parts, following the hierarchical tree structure, need to be constrained by the relationships between parent and child parts in terms of geometric composition. To ensure this, a bond loss is calculated between level \\( l \\) and the next level \\( l + 1 \\). The bond loss is defined as the error between the occupancy value of each parent part and the maximum of the occupancy values of its two derived parts:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.706, + 0.892, + 0.75 + ], + "angle": 0, + "content": "\\[\nL _ {b o n d} ^ {l} = \\sum_ {i = 1} ^ {2 ^ {l}} \\left(S _ {i} ^ {l} - \\max \\left(S _ {2 i - 1} ^ {l + 1}, S _ {2 i} ^ {l + 1}\\right)\\right) ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.758, + 0.892, + 0.818 + ], + "angle": 0, + "content": "where \\( S_{i}^{l} \\) is the occupancy value of the parent part at position \\( i \\) and level \\( l \\), and \\( S_{2i-1}^{l+1} \\) and \\( S_{2i}^{l+1} \\) are the occupancy values of the two derived child parts. The total bond loss is computed by summing the bond loss across all levels when" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.818, + 0.741, + 0.838 + ], + "angle": 0, + "content": "\\[\nl < L \\colon L _ {b o n d} = \\lambda_ {b o n d} \\sum_ {l = 1} ^ {l < L} L _ {b o n d} ^ {l}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Regularization Loss To reduce the complexity of the network and avoid overfitting, we introduce regularization loss \\( L_{reg} \\) at the last layer of each Structure Divider at each level." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2718" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.214 + ], + "angle": 0, + "content": "This loss is calculated as the sum of the absolute values of all weight parameters, multiplied by a small regularization parameter \\(\\lambda_{reg}\\). The regularization loss is given by: \\(L_{reg} = \\lambda_{reg}\\sum_{i=1}^{n}|w_i|\\), where \\(w_i\\) is the \\(i\\)-th weight parameter in the layer, and \\(n\\) is the total number of weight parameters. Our experiments show that without using regularization loss, the generated shapes may contain parts that are over-segmented or lack reasonable shape." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.242, + 0.473, + 0.455 + ], + "angle": 0, + "content": "Comparison with Existing Method Our method differs from previous approaches that utilize (hierarchical) decoders for unsupervised shape part learning (such as in [9, 29]) in several ways. Firstly, our hierarchical decoder aims to learn semantic shape part-level implicit field rather than simply segmenting the shape into more parts. For example, when generating a chair with four legs, our approach considers the legs as a whole unit rather than segmenting them individually, as we typically describe the legs as a whole rather than as individual parts. To balance the network's fitting ability with its complexity, we utilize a regularization loss. Secondly, we also construct hierarchical color latent fields, which are often overlooked in other methods, to represent colored shapes." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.469, + 0.353, + 0.485 + ], + "angle": 0, + "content": "3.3. Text-Guided Shape Generation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.494, + 0.473, + 0.692 + ], + "angle": 0, + "content": "Fig. 2 (B) shows the text-guided shape generation network, comprising three modules: text encoder \\( E_{T} \\), textual graph encoder \\( E_{G} \\), and hierarchical shape decoder \\( D_{T} \\). The \\( D_{T} \\) module is similar to the pre-trained shape-specific decoder \\( D_{S} \\) (as described in Section 3.2), with the addition of a new part-node attention module and a refiner. This attention module allows the network to learn the local connections between shape parts/colors and textual graph entities, followed by a refiner module \\( R \\) for final results generation. \\( D_{T} \\) module is initialized from the pre-trained \\( D_{S} \\) to leverage its predefined structural implicit fields and facilitate the generation process in a hierarchical manner. We provide a detailed description of each part in the following sections." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Text Global Encoding We utilize a pre-trained BERT model [11] as the text encoder base. The encoded [CLS] token feature is used to extract global-level latent features \\(\\bar{t}_s\\) and \\(\\bar{t}_c\\) for shape and color, respectively, which serve as the input for the hierarchical shape decoder \\(D\\). Most existing approaches employ word-level features \\(w\\) as fine-grained guidance in text-conditioned image or 3D shape generation tasks [40, 25]. However, we argue that the use of structured textual representation, such as the textual graph, would be more appropriate in aligning with shape structures. As a result, we opt to use the word features to initialize the graph node representation, as described below." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.891, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.234, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Figure 3. The dashed box shows a triplet \\((v_{s}, v_{r}, v_{o})\\) in the graph, denoting the subjective, relation, and objective nodes. The three nodes are first concatenated and then jointly processed using MLP, which outputs a new value for each node. The final updated state \\(v_{s}^{\\prime}\\) of \\(v_{s}\\) is the result of pooling the corresponding values from all triplets in which the node \\(v_{s}\\) is involved." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.345, + 0.895, + 0.68 + ], + "angle": 0, + "content": "Shape-Oriented Text Graph We use a unique text graph representation designed specifically for shapes. The graph nodes are classified into root, part, and attribute categories, which represent the class of the shape, the parts, and the attribute for root/part nodes, respectively. We use a special type node relation to connect these nodes, forming the edges. For example, the text \"brown wooden chair with curved armrests\" would be parsed as (chair: brown, wooden), (armrests: curved), and their relation triplet (chair, with, armrests) in the (subjective, relation, objective) convention. To construct graphs from the original text, we first use the spaCy library [15] to parse the text's semantic dependencies. These raw dependencies contain many linguistic terms and do not correspond well with real-world structures. To obtain more clear and shape-oriented graph formation, we introduce a custom parser targeting 3D shapes based on [36]. To ensure that the graph remains connected, we connect all part nodes with the root node using the with relation if no linguistic relation is found. The lower left of Fig. 2 illustrates these two forms of text. Finally, we use the resulting relation triplets for training. Further details can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Spatial-Aware Graph Encoding One challenge in using a shape-oriented graph, where entity or attribute nodes are connected through relation nodes, is the lack of a canonical grid for embedding the graph since nodes in such a graph exist in a non-Euclidean space. Existing methods based on Graph Convolution Networks (GCN) [21] typically rely on an adjacency matrix to represent the relations between nodes, which reduces their ability to represent spatial relations. To address the issue, we designed a novel Spatial-Aware Graph Encoder (SAGE) that differentiably encodes relation nodes in order to capture the structural relations between nodes. Specifically, we represent each graph triplet \\( g \\) in the form \\( (v_{s}, v_{r}, v_{o}) \\), where \\( v_{s} \\), \\( v_{o} \\), and \\( v_{r} \\) represent" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2719" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.09, + 0.455, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.238, + 0.472, + 0.281 + ], + "angle": 0, + "content": "Figure 4. The Part-Node attention. The attention map is obtained by the dot-product between the features from the intermediate part latent feature \\((f_s^l)\\) at level \\(l\\) and the graph node features \\(g\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.47, + 0.412 + ], + "angle": 0, + "content": "the values of subjective, objective, and relation nodes in a relationship, respectively. This makes the graph directional through the order of the nodes between the relation node, like chair has arms. As shown in Fig. 3, the computational layer in SAGE jointly encodes the nodes and their relation in each triplet, followed by a pooling layer that fuses the features of each node involved in more than one triplet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.47, + 0.553 + ], + "angle": 0, + "content": "Part-Node Attention We introduce the Part-Node Attention modules for learning the relationship between intermediate parts/colors with textual graph nodes. These modules are based on the multi-head attention mechanism presented in [38]. For level \\( l \\), given the intermediate shape latents \\( f_{s}^{l} \\in \\mathbb{R}^{2^{l} \\times d} \\) as query, and the encoded graph node features, \\( g \\in \\mathbb{R}^{n \\times d} \\) (\\( n \\) is node number), as key and value, the calculation inside an attention head is:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.567, + 0.469, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\hat {f} _ {s} ^ {l} = \\operatorname {s o f t m a x} \\left(\\frac {Q K}{\\sqrt {d _ {h}}}\\right) V, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.79 + ], + "angle": 0, + "content": "where \\( Q = F_{Q}(f_{s}^{l}) \\), \\( K = F_{K}(g) \\), \\( V = F_{V}(g) \\), and \\( F_{Q}, F_{K} \\), and \\( F_{V} \\) are functions implemented using MLPs. \\( d \\) is the feature dimension for \\( g \\) and \\( f_{s}^{l} \\), while \\( d_{h} \\) represents the dimension used within an attention head. The SoftMax operation is applied across the graph nodes, resulting in a weighted aggregation of the graph node features as the new intermediate part latents, \\( \\hat{f}_{s}^{l} \\). This process is visualized in Fig. 4. Also, attention is applied between the paired color features \\( f_{c}^{l} \\) with the graph nodes using a separate module. In practice, we use the attention module at level 2 of the hierarchical decoder due to the generated semantically detailed shape components aligned with graph nodes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Joint Refiner The hierarchical bond loss, introduced in Section 3.2, serves as the foundation for structure-aware generation by constraining the composition of shape parts to follow a hierarchical tree structure. However, we posit that this also negatively impacts the generated results, as demonstrated in the perception-distortion tradeoff [5]. To" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "mitigate this issue, we introduce a refiner \\( R \\) in one level that uses the level latents jointly to generate the final shape. The refiner is also an implicit shape decoder with six layers of MLPs. The final shape output is accompanied by an additional refiner reconstruction loss \\( L_{R} \\) but without the need to consider the hierarchical bond, which allows for the incorporation of additional detail in the final shape." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.225, + 0.892, + 0.468 + ], + "angle": 0, + "content": "Training Strategy and Losses Training text-guided hierarchical shape generation directly presents significant challenges considering the big discrepancy between text and shape modalities. To address this, we propose a three-step strategy to ensure stable and effective training. First, we collect the shape and color latents from the shape encoder \\( E_{S} \\) as guidance and train the text global feature encoder \\( E_{T} \\) alone. The loss is using MSE: \\( L_{\\text{latent}} = \\lambda_{\\text{latent}}^{s} (\\bar{t}_{s} - \\bar{f}_{s})^{2} + \\lambda_{\\text{latent}}^{c} (\\bar{t}_{c} - \\bar{f}_{c})^{2} \\), where \\( \\bar{t}_{s} \\) and \\( \\bar{t}_{c} \\) are encoded by text encoder while \\( \\bar{f}_{s} \\) and \\( \\bar{f}_{c} \\) are encoded by shape encoder. Second, we train the decoder with attention modules, freezing the decoder during the first ten epochs to allow for warmed-up training. Finally, we train the entire network together. The total loss then includes the aforementioned shape/color reconstruction losses and the bond loss in shape-specific pre-training (see Sec. 3.2), and the refiner loss." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.485, + 0.633, + 0.503 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.513, + 0.892, + 0.574 + ], + "angle": 0, + "content": "This section introduces our experimental design and implementation, followed by a comprehensive analysis of the results from various aspects and an ablation study. Supplementary material provides additional results and analysis." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.588, + 0.868, + 0.604 + ], + "angle": 0, + "content": "4.1. Experimental Settings and Implementation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.718 + ], + "angle": 0, + "content": "We conduct experiments using the text-shape dataset [8], which is derived from the 3D shape dataset, ShapeNet [7], and augmented with textual annotations. The dataset comprises approximately 15,000 shapes and 75,000 textual descriptions, encompassing the table and chair shape classes. We also adhere to the same training and testing splits as previously established." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We implement our framework in PyTorch [32]. We set \\( N = 4096 \\) as the sampled points. We pre-train the shape-specific decoder for 1000 epochs and the text-guided model for another 500 epochs. The learning rate is constant \\( 1e^{-4} \\) using Adam [20] for all phases. The loss weights are set as follows: \\( \\lambda_{\\text{recon}} = 10.0 \\), \\( \\lambda_{\\text{color}} = 1.0 \\), \\( \\lambda_{\\text{bond}} = 1.0 \\), \\( \\lambda_{\\text{reg}} = 1e^{-4} \\), \\( \\lambda_{\\text{latent}}^s = 10.0 \\) and \\( \\lambda_{\\text{latent}}^c = 1.0 \\). Detailed training strategy can be found in Sec. 3.3. Additionally, as our model outputs shapes at different levels, we select the results from level 2, as it is at this level that most shapes' parts have been parsed and refined. Level 3 is used as a spare level for monitoring no further parts are parsed." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2720" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.153 + ], + "angle": 0, + "content": "
IoU (↑)IS (↑)EMD (↓)Acc (↑)
Text2Shape [8]9.641.960.444397.37
T2S-Implicit [25]12.211.970.207197.48
ours13.651.980.176797.5
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.078, + 0.154, + 0.47, + 0.183 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with state-of-the-art. \\( \\uparrow \\) and \\( \\downarrow \\) indicate higher or lower is better, respectively." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.214, + 0.353, + 0.231 + ], + "angle": 0, + "content": "4.2. Text-Guided Shape Generation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.242, + 0.473, + 0.635 + ], + "angle": 0, + "content": "Comparison with the State-of-the-art We compare our method to two existing text-conditioned colored shape generation methods, Text2Shape [8] and T2S-Implicit [25]. We employ the same quantitative metrics for evaluation, including (1) Mean intersection-over-union (IoU) for shape occupancy measurement, regardless of color; (2) Earth Mover's Distance (EMD) for color distribution evaluation; (3) Inception Score (IS) to quantify the realism of the results; and (4) Accuracy (Acc) computed by a pre-trained shape classifier to indicate whether the predicted class matches the ground truth. To ensure fairness in comparison, we convert the generated shapes to volumes of the same size as in [8, 25] before calculating the quantitative values. The comparison results are presented in Table 1, where the numbers of the existing methods are directly taken from the respective papers. Our method exhibits superior performance across all metrics. The qualitative result presented in Fig. 5 provides further evidence of the effectiveness of our approach. As can be observed, our method achieves a higher degree of correspondence between the predicted shape structure and the text description, as demonstrated in Row 1 of the figure with the example of \"wood armrests\". Additionally, our approach exhibits an increased level of completeness in part generation, as evidenced by the absence of incomplete parts or extra artifacts, which are clearly labeled with red circles in the results generated by the existing methods." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.674, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Part-level Refinement As T2S-Implicit [25] also employs an implicit decoder similar to our approach, we conduct a comparison with it to highlight the distinctions, particularly with regard to part-level details. The results are presented in high resolution in Fig. 6. We observed that [25] tends to struggle with smaller or complex parts, such as the legs of a sofa or the wheels of an office chair. In contrast, our method is capable of interpreting these parts explicitly, resulting in more realistic part-level generation. Another advantage of our approach is that the generation of a single part does not affect the other parts, which is a common issue in previous methods. For example, generating \"silver legs\" may result in the bottom of the sofa appearing slightly silver in the connection area, creating a noticeable color artifact." + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.34, + 0.894, + 0.382 + ], + "angle": 0, + "content": "Figure 5. Visualization of the results of our method in comparison to Text2Shape [8] and T2S-Implicit [25]. Red circles indicate incomplete parts or extra artifacts." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.401, + 0.895, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.538, + 0.835, + 0.552 + ], + "angle": 0, + "content": "Figure 6. Comparison of part-level refinement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.745 + ], + "angle": 0, + "content": "Part-level Manipulation The ability to learn structure-based 3D shape representations and align them with textual descriptions enables precise and convenient manipulation of shapes. In the process of generating a shape from text, our method allows for manipulation at various levels, including modification of the part-level structure, attributes, or colors. The results of this manipulation are illustrated in Fig. 7. The changes made in the text appear to reflect the changes in part structure (a), part attribute (c), and part-related color refinement (b), suggesting a possible consistency between the constituent parts of the shape and the corresponding text." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.756, + 0.665, + 0.771 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We conduct extensive ablation studies to validate the effectiveness of the proposed hierarchical structure-based representation, graph-part attention, and joint refiner. We also use Frechet Inception Distance (FID) and CLIP [34] to evaluate the quality of the generated shapes and the similarity between the generated shapes and their corresponding text descriptions, respectively. More detailed information on the metrics can be found in the supplementary material. The re" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2721" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.882, + 0.208 + ], + "angle": 0, + "content": "
MethodLevel 1Level 2Level 3
IOU↑FID↓CLIP↑IOU↑FID↓CLIP↑IOU↑FID↓CLIP↑
HierD [base]11.2017.4745.4311.0717.5245.3610.8517.6043.92
FlatD10.3819.5540.72------
HierD + WordAttn (L2)11.1317.4244.8811.6616.4647.4310.7217.6946.10
HierD + GraphAttn (L2)11.2917.3645.0112.6316.3150.6412.4116.7050.01
HierD + GraphAttn (L1)11.5416.8652.7511.3017.0951.4011.2717.1250.56
HierD + GraphAttn (L2) + Refiner(L2) [Full]11.1917.3847.2713.6515.8154.8912.2216.3452.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.209, + 0.893, + 0.239 + ], + "angle": 0, + "content": "Table 2. Results of the ablation study on output levels. The base model, HierD, is the proposed hierarchical decoder using sentence-level guidance only. L1/L2 indicates component application at Level 1/Level 2. ↑/↓ denote better higher/ lower results, respectively." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.266, + 0.468, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.433, + 0.472, + 0.492 + ], + "angle": 0, + "content": "Figure 7. Manipulation of text results in corresponding changes at the part level. The second row shows the predicted structure from Level 1 to 2. (a) illustrates the removal of a part, (b) shows a change in part color, and (c) depicts a change in the part attribute." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.514, + 0.47, + 0.637 + ], + "angle": 0, + "content": "sults of our ablation studies are presented in Table 2. We use the Hierarchical Decoder with sentence-level guidance (HierD) as a base model, and then add various modules such as word-level attention (WordAttn), our proposed graph-part attention (GraphAttn), and the joint refiner to different levels of the hierarchy (L1/L2 for Level 1 and 2). We present the results for each level for an in-depth analysis of the impact of the hierarchy level on the performance of the model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.47, + 0.761 + ], + "angle": 0, + "content": "Hierarchical Representation We compare HierD to a flattened decoder (FlatD) which directly outputs 8 parts at Level 1. Utilizing the same settings as HierD, FlatD demonstrates inferior performance. Our qualitative analysis reveals that FlatD is more prone to overfitting, resulting in the generation of fewer parts that are overly complex geometrically, instead of generating finer parts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Graph-based Attention We have observed the clear advantages of utilizing graph-based attention over word-based attention. Specifically, when applied to Level 2, the use of GraphAttn results in superior performance across all metrics. Furthermore, the results from Level 3 also demonstrate improved performance. This can be attributed to the semantic alignment of the entity nodes in the graph with the intermediate parts. However, when applied to Level 1, the use of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.893, + 0.339 + ], + "angle": 0, + "content": "GraphAttn does not yield similar results as in Level 2. We speculate that this discrepancy may be due to the coarser nature of the parts at Level 1, which hinders the ability to establish effective connections between shape parts and textual graph entities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.893, + 0.527 + ], + "angle": 0, + "content": "Joint Refiner The utilization of a refiner as the final shape output in our method is driven by the observation of the perception-distortion tradeoff [5]. As the level increases, we have noted a slight decline in performance as compared to previous levels. For instance, the IOU in Level 2 is consistently lower than that in Level 1. We posit that the hierarchical bond loss employed in our approach results in a stronger structural connection, but may impede the generation of detailed features on top of the structure. The incorporation of a refiner, on the other hand, has resulted in a notable improvement in all metrics." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.542, + 0.857, + 0.56 + ], + "angle": 0, + "content": "5. Conclusion, limitation, and future work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.734 + ], + "angle": 0, + "content": "The ShapeScaffold is a method for generating shapes guided by text, with the understanding that both shape and text possess internal structures that can be aligned to improve the outcome. We represent shapes as hierarchical structures and model text as scene graphs. Our part-node attention mechanism allows us to learn correspondences between shape and text at various hierarchical levels, which has been shown to effectively establish connections between shape parts/colors and linguistic entities. The experiments have demonstrated the superiority of our approach in both shape generation and text-shape/color consistency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Several limitations should be acknowledged. Firstly, the current text-to-shape dataset is limited to chairs and tables, so further research is needed to explore the generalization of our method to other shape classes. Secondly, the results may be influenced by the parts priority problem, where incorrect occupancy state of parts junctions may result in unreasonable visualization. Therefore, further incorporation of prior interpretations of shapes should be considered in future studies. Lastly, it would be beneficial to explore scene generation that takes structure into account, with a focus on the composition of different shapes as an outer structure." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2722" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.152 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by RCUK grant CAMERA (EP/M023281/1, EP/T022523/1), Centre for Augmented Reasoning (CAR) at the Australian Institute for Machine Learning, and a gift from Adobe." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.164, + 0.173, + 0.179 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.189, + 0.47, + 0.258 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Judy Fan, Robert Hawkins, Noah Goodman, and Leonidas J Guibas. Shapeplot: Learning language for shape differentiation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8938-8947, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.314 + ], + "angle": 0, + "content": "[2] Oron Ashual and Lior Wolf. Specifying object attributes and relations in interactive scene generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4561-4569, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.315, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[3] Renée Baillargeon. Infants' physical world. Current directions in psychological science, 13(3):89-94, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.343, + 0.471, + 0.384 + ], + "angle": 0, + "content": "[4] Irving Biederman. Human image understanding: Recent research and a theory. Computer vision, graphics, and image processing, 32(1):29-73, 1985. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.385, + 0.471, + 0.426 + ], + "angle": 0, + "content": "[5] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.427, + 0.471, + 0.495 + ], + "angle": 0, + "content": "[6] Angel Chang, Manolis Savva, and Christopher D Manning. Learning spatial knowledge for text to 3d scene generation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 2028-2038, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.497, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.567, + 0.471, + 0.635 + ], + "angle": 0, + "content": "[8] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian conference on computer vision, pages 100-116. Springer, 2018. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.637, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[9] Zhiqin Chen, Kangxue Yin, Matthew Fisher, Siddhartha Chaudhuri, and Hao Zhang. Bae-net: Branched autoencoder for shape co-segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8490-8499, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.47, + 0.746 + ], + "angle": 0, + "content": "[10] Morten H Christiansen and Nick Chater. Language as shaped by the brain. Behavioral and brain sciences, 31(5):489-509, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.873 + ], + "angle": 0, + "content": "[12] Helisa Dhamo, Azade Farshad, Iro Laina, Nassir Navab, Gregory D Hager, Federico Tombari, and Christian Rupprecht. Semantic image manipulation using scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5213-5222, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecraft: A recursive text-conditioned 3d shape" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.189, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "generation model. arXiv preprint arXiv:2207.09446, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proceedings of the 27th International Conference on Neural Information Processing Systems - Volume 2, NIPS'14, page 2672-2680, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[15] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. spaCy: Industrial-strength Natural Language Processing in Python, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[16] Fangqiao Hu, Jin Zhao, Yong Huang, and Hui Li. Learning structural graph layouts and 3d shapes for long span bridges 3d reconstruction. arXiv preprint arXiv:1907.03387, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[17] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[18] R. Kenny Jones, Aalia Habib, Rana Hanocka, and Daniel Ritchie. The neurally-guided shape parser: Grammar-based labeling of 3d shape regions with approximate inference. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[19] R Kenny Jones, Aalia Habib, and Daniel Ritchie. Shred: 3d shape region decomposition with learned local operations. arXiv preprint arXiv:2206.03480, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[20] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[21] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[22] Jun Li, Kai Xu, Siddhartha Chaudhuri, Ersin Yumer, Hao Zhang, and Leonidas Guibas. Grass: Generative recursive autoencoders for shape structures. ACM Transactions on Graphics (TOG), 36(4):1-14, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[23] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[24] Chunxiao Liu, Zhendong Mao, Tianzhu Zhang, Hongtao Xie, Bin Wang, and Yongdong Zhang. Graph structured network for image-text matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10921-10930, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[25] Zhengzhe Liu, Yi Wang, Xiaojuan Qi, and Chi-Wing Fu. Towards implicit text-guided 3d shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17896-17906, 2022. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "2723" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.191 + ], + "angle": 0, + "content": "[27] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 306-315, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[28] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.251, + 0.47, + 0.32 + ], + "angle": 0, + "content": "[29] Chengjie Niu, Manyi Li, Kai Xu, and Hao Zhang. Rim-net: Recursive implicit fields for unsupervised learning of hierarchical shape structures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11779-11788, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.322, + 0.471, + 0.391 + ], + "angle": 0, + "content": "[30] Despoina Paschalidou, Luc Van Gool, and Andreas Geiger. Learning unsupervised hierarchical part decomposition of 3d objects from a single rgb image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1060-1070, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.393, + 0.47, + 0.461 + ], + "angle": 0, + "content": "[31] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics revisited: Learning 3d shape parsing beyond cuboids. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10344–10353, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.464, + 0.47, + 0.616 + ], + "angle": 0, + "content": "[32] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.47, + 0.66 + ], + "angle": 0, + "content": "[33] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.47, + 0.745 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[35] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[36] Richard Socher, Cliff C Lin, Chris Manning, and Andrew Y Ng. Parsing natural scenes and natural language with recursive neural networks. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 129–136, 2011. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[37] Shubham Tulsiani, Hao Su, Leonidas J Guibas, Alexei A Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2635-2643, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.219 + ], + "angle": 0, + "content": "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.893, + 0.288 + ], + "angle": 0, + "content": "[39] Yanzhen Wang, Kai Xu, Jun Li, Hao Zhang, Ariel Shamir, Ligang Liu, Zhiquan Cheng, and Yueshan Xiong. Symmetry hierarchy of man-made objects. In Computer graphics forum, volume 30, pages 287-296. Wiley Online Library, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.29, + 0.893, + 0.373 + ], + "angle": 0, + "content": "[40] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2724" + } + ] +] \ No newline at end of file diff --git a/2023/ShapeScaffolder_ Structure-Aware 3D Shape Generation from Text/3212f743-a520-438c-b4e2-ff41229e0a86_origin.pdf b/2023/ShapeScaffolder_ Structure-Aware 3D Shape Generation from Text/3212f743-a520-438c-b4e2-ff41229e0a86_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2709922c36c9b6b8bb4a76e25c4f842b1d0d8440 --- /dev/null +++ b/2023/ShapeScaffolder_ Structure-Aware 3D Shape Generation from Text/3212f743-a520-438c-b4e2-ff41229e0a86_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:abb66fcc60504ee7eaaae287be6147d71a331bf86d35e4df63a54e3f48cf9a8d +size 4695596 diff --git a/2023/ShapeScaffolder_ Structure-Aware 3D Shape Generation from Text/full.md b/2023/ShapeScaffolder_ Structure-Aware 3D Shape Generation from Text/full.md new file mode 100644 index 0000000000000000000000000000000000000000..843df22b3d8018f8a68aef4b5fc48c70401b1ee2 --- /dev/null +++ b/2023/ShapeScaffolder_ Structure-Aware 3D Shape Generation from Text/full.md @@ -0,0 +1,254 @@ +# ShapeScaffold: Structure-Aware 3D Shape Generation from Text + +Xi Tian +University of Bath +Bath, UK +xt275@bath.ac.uk + +Yong-Liang Yang University of Bath Bath, UK y.yang@cs.bath.ac.uk + +Qi Wu +University of Adelaide +Adelaide, Australia +qi.wu01@adelaide.edu.au + +# Abstract + +We present ShapeScaffold, a structure-based neural network for generating colored 3D shapes based on text input. The approach, similar to providing scaffolds as internal structural supports and adding more details to them, aims to capture finer text-shape connections and improve the quality of generated shapes. Traditional text-to-shape methods often generate 3D shapes as a whole. However, humans tend to understand both shape and text as being structure-based. For example, a table is interpreted as being composed of legs, a seat, and a back; similarly, texts possess inherent linguistic structures that can be analyzed as dependency graphs, depicting the relationships between entities within the text. We believe structure-aware shape generation can bring finer text-shape connections and improve shape generation quality. However, the lack of explicit shape structure and the high freedom of text structure make cross-modality learning challenging. To address these challenges, we first build the structured shape implicit fields in an unsupervised manner. We then propose the part-level attention mechanism between shape parts and textual graph nodes to align the two modalities at the structural level. Finally, we employ a shape refiner to add further detail to the predicted structure, yielding the final results. Extensive experimentation demonstrates that our approaches outperform state-of-the-art methods in terms of both shape fidelity and shape-text matching. Our methods also allow for part-level manipulation and improved part-level completeness. + +# 1. Introduction + +The advance of 3D representation learning and generative models has sparked increasing interest in 3D shape generation. However, text-guided shape generation remains a challenging task. Many current approaches generate the 3D shape as a whole when no part information is utilized, while others treat text as a simple collection of words in order to provide finer guidance. Both shape and text, however, possess inherent internal structures that can be leveraged to + +![](images/88baeeeb684e01923109f07d576ce08639792c69feaf7abf5b15f6997ef3dc84.jpg) +Figure 1. Text to shape generation with structure awareness. Raw description generates the initial coarse shape with structures. Further correspondence is built between shape parts and text graph nodes for fine-grained refinement. + +improve the alignment between the two modalities. + +Humans are born with a sense of physical understanding of the world [3]. Studies have shown that the human visual system perceives objects as hierarchical arrangements of parts [4], comprising high-level structural properties and low-level details. In addition, humans use language to describe objects, including their appearance, structure, and functions. It has been discovered that language, shaped by the human brain over time, has evolved to have hierarchical linguistic structures in order to convey more complex meanings [10]. This suggests that structure-based reasoning is present in both the visual and language modalities. While this is intuitive and effortless for humans, it is a challenge for intelligent algorithms to understand the structural connection between these two modalities. + +In this research, we aim to augment shape generation from text with an awareness of structure in both shape and language modalities. This would allow for shape generation to be guided not only by high-level descriptions, such as a sentence but also refined with part-level guidance from semantic text entities. This process is similar to how humans create visual artworks: sketching an idea's gist and then progressively refining the details. However, there are + +several challenges to overcome. One challenge is that the shape structures are unknown unless annotated segmentation labels are provided. Even with annotation, where the process is time-consuming and labor-intensive, the definition of "parts" in shapes is subjective and ambiguous at multiple levels. Another challenge is that text descriptions of shapes are diverse and complex, making it difficult to optimize learning. In addition, for colored shapes, the alignment between structure and color representation must also be taken into consideration. + +To address the challenges above, we propose a two-stage approach. First, we leverage an unsupervised hierarchical shape structure given part hierarchy is a prevalent and effective approach for representing shape structures, as demonstrated in previous research [22, 28]. To achieve this, we introduce a decomposition model to progressively decompose the shape into parts at multiple levels of granularity, from coarse to fine, such that each level can reconstruct the entire shape. Additionally, we learn color reconstruction at the part level. Fig. 1 illustrates the process. While the structures and parts identified in the first stage hold significance within the shape domain, they do not have a corresponding relationship with the text. In the second step, we establish a connection between the shape structure and the text structure. We utilize linguistic dependency graphs, which can be used to represent the structural elements present in the text, as attention signals to form the structure-level connection. Through these two steps, we are able to improve the generation results at the part level by building upon the established structure. + +In our experimentation on the ShapeNet dataset [7], we have demonstrated the efficacy of utilizing a structure-aware representation in improving the correspondence between text and shape. As a result of the part-aware refinement, we observe a significant improvement in the quality of generated shapes, particularly in terms of their constituent parts. Additionally, our approach attains state-of-the-art results regarding several commonly used metrics such as Intersection over Union (IOU), Accuracy, and Inception Score (IS). Furthermore, through further exploration of the latent shape structure space, it is demonstrated that our method has the capability of generating previously unseen shapes. + +The main contributions of this work are: (1) An unsupervised hierarchical decoder for building shape structural implicit field. (2) A structure-aware text-to-shape generation method that improves part-level correspondence and refinement. (3) Comprehensive experiments that demonstrate the effectiveness of our method. + +# 2. Related Work + +Text Guided 3D Shape Generation. Deep learning has led to a surge in works exploring text-guided 3D shape or + +scene generation [6, 8, 1, 25, 27, 13]. Recently, to achieve faithful results, some works tend to adopt pre-trained CLIP models [35, 17, 26] or diffusion-based methods [33, 23] for large-scale learning on open datasets. Our work, however, focuses on the classic ShapeNet dataset [7], examining the generation of structure-aware, colored shapes from text and comparing our results to previous baselines. + +Generating 3D shapes directly from text remains a challenge, particularly when attempting to concurrently generate colors. [8] proposes a method that learns a joint text-shape embedding space and then uses a generative adversarial network (GAN) [14] to generate colored voxels from text. However, this approach has not been successful in terms of shape resolution, color quality, and cross-modality consistency. To address these issues, [25] proposes a method that uses an implicit 3D representation with word-level attention to improve control over generated shape structures and colors. [27] and [13] both adopt a discrete autoencoder to capture patch-based shape priors and subsequently employ a transformer for autoregressive generation. However, it is worth noting that these methods do not take into account color information, which is an essential attribute of object appearance. Also, all these methods have treated the shape as a whole, ignoring the internal structure within the shape. Additionally, text modeling has been limited to language-level features, such as sentence and word features, rather than considering the structure of the text for structure-level matching with the shape domain. + +3D Structure-Aware Representation and Parsing. Although structure-based shape generation with text guidance has not been previously explored, there is a significant amount of research on learning structure-aware 3D shape representation or parsing 3D shapes into parts. For supervised methods, [22] proposes the use of a symmetry hierarchy [39] for hierarchical shape structure representation, while StructureNet [28] represents shapes using a graph neural network that considers both primitives and hierarchies. Similarly, [16] utilizes a binary parsing tree to represent the 3D structure of a cable-stayed bridge. Recently, semantic-based shape decomposition or parsing methods have been proposed in [19, 18], which use learned operations to obtain grammar-level shape parts. Unsupervised methods such as [37, 31] use convolutional neural networks (CNNs) to generate primitive shapes, e.g., cuboids or superquadrics, which abstract the input shape. BAE-NET [9] attempts to co-segment shapes into parts using a branched autoencoder architecture that disentangles common features for segmentation. However, these methods only predict the parts without considering their dependencies in the shape structure. [30] recovers the 3D objects into a hierarchy of parts directly from an RGB image using CNN, while the parts are primitive and lack details. RIM-NET [29] infers + +hierarchical structures of 3D shapes using recursive implicit fields. It decomposes an input shape into two parts at each level, resulting in a binary tree hierarchy, where each level of the tree corresponds to an assembly of shape parts represented as implicit functions. However, the method does not consider color information and only processes one shape at a time during training. Our work uses a hierarchical tree structure for shape representation, but also includes color information for part-level implicit representation. In addition, our model is able to learn a shared latent space for a set of shapes. + +Text as Textual Graph Guidance. Text-to-image (T2I) generation has seen significant progress in recent years, with a focus on both generating realistic images and increasing consistency between texts and images. Some approaches use attention-based methods [40] for fine-grained alignment, while others leverage shared structures in images and texts. Scene graphs, which are derived from texts and contain expressive structural relationships between entities, have been widely used to reduce the cross-modality gap between texts and images for tasks such as text-image matching [24], fine-grained image generation [2], and image manipulation [12]. Given that both shapes and texts contain structural elements, we opt to employ textual graphs as additional guidance in conjunction with the generation of structural parts. + +# 3. Method + +# 3.1. Overview + +The task is to generate 3D shapes with colors based on the given text description $T$ . The predicted object is denoted as $O \in \mathbb{R}^{N \times 4}$ , comprising the shape occupancy values $S \in \mathbb{R}^{N \times 1}$ and the color RGB values $C \in \mathbb{R}^{N \times 3}$ , where $N$ is the number of sample points. + +The overview of our method is illustrated in Fig. 2. Our framework consists of a text encoder $E_{T}$ , textual graph encoder $E_{G}$ , hierarchical structure-aware shape decoder $D_{T}$ and shape encoder $E_{S}$ . In contrast to general methods that generate the whole shape at once, our approach, called ShapeScaffold, aims to generate the 3D shape by considering its internal structures, enabling refinement at the part level and improvement of text-shape correspondence. To achieve this, we propose to capture the part-level information from both shape and text. For the shape, we introduce the concept of structural implicit field in the shape decoder to capture the underlying structure and relationships of shape parts and colors. As shown in Fig. 2 (middle right), the structure latent fields capture the semantic shape parts for a chair, including arm, cushion, legs, etc. For the text, in addition to encoding it as a global feature such as a sentence vector, we also propose using a linguistic tex + +tual graph parsed from the text that includes explicit entities with relations. This enables the structure latent fields to interact with textual entities at a finer level, resulting in more direct and explicit correspondence between the text description, such as "curved armrests" and the corresponding arm part of the generated shape. However, it is non-trivial to train a hierarchical structure-aware generator conditioned on text. We approach this task by dividing the training process into two stages: (A) for learning the structural implicit field using an autoencoder, and (B) for generating shapes guided by text, as summarized in Fig. 2. Stage B can be further divided into text encoding (B-1) and shape generation (B-2). The following describes each stage in detail. + +# 3.2. Shape-Specific Pre-training + +To capture the internal semantic structures of shapes and facilitate shape decoding conditioned on text, we employ a pre-training technique using a shape-specific hierarchical decoder $D_{S}$ and a shape encoder $E_{S}$ , forming an autoencoder (see Fig. 2 (A)). The encoder $E_{S}$ , which is a convolutional neural network (CNN), takes the volume of an object $O$ as input and produces latent shape and color features to be used by the decoder. The decoder then creates an implicit field by partitioning the shape $(\bar{f}_s)$ and color $(\bar{f}_c)$ latents at each hierarchical level, reconstructing the latent features into parts and assembling them into the final object. An example of a successfully learned structural implicit field is shown in Fig. 2 (right), where semantic shape parts can be reconstructed from their corresponding latent positions. + +Hierarchical Shape Decoder The Hierarchical Decoder consists of two modules at each level: the Structure Divider (SD) and the Part Generator (PG). Given a pair of shape/color latent features $(f_s^{l,i} \in \mathbb{R}^d, f_c^{l,i} \in \mathbb{R}^d)$ at level $l$ and position $i(1 \leq i \leq 2^l)$ , the SD first divides the shape latent into two child shape features and then generates the color latent for each child conditioned on the corresponding child shape latent. The SD constructs the latent space for both shape and color, as depicted by the grey and yellow squares in Fig. 2. The PG then generates shape parts one by one from its shape/color latents at a given set of point coordinates $\mathbf{p} = \{(x,y,z) \mid x,y,z \in \mathbb{R}\}$ . For each point $p$ , the input is the concatenation of $p$ with the latent value, and SD outputs one occupancy probability for the shape channel and three values (RGB) for the color channel. The resulting generated shape $O_l$ at level $l$ is a composite of generated parts, where a point is occupied by a part with the highest occupancy probability, provided that this probability exceeds a pre-defined threshold. The color of each point is determined based on the corresponding color values of the occupying part. Both SD and PG are implemented using multi-layer perception (MLPs). The SD is implemented using a two-layer MLP for partitioning the shape latent space + +![](images/6d54f528674ff29953f5258de83556ceed6e92a624e69882d34b7d76aea344a7.jpg) +(B-2) T2S: Hierarchical Generation + +![](images/a1b3617f7ad718a23a1aa3393332ec78b391dbc9a9f69a24902cda075583b42b.jpg) + +![](images/33116be9703471b80db4230b10f9ea70463e80f592ba955af3fcc77e720a6eca.jpg) + +![](images/9555a389380805c697c84df83a9ee72ac5495f30b273f409b10af1ea303c4e97.jpg) +(B-1) T2S: Text Encoding + +![](images/3b1a71df0864ffd4d10b5db68800b64f4c7fca8b4eca02d5128d532bdf35cdfd.jpg) + +![](images/45d8c22361b1a09b010ff5a3fb38fe2d2a7932e6d424d5344312b70fd1e6afe9.jpg) +Figure 2. Method overview. The proposed method includes two phases: shape autoencoder pre-training (A) and text-to-shape (T2S) phase including (B-1) and (B-2). Phase (A) trains the shape-specific decoder using shape/color reconstruction loss and hierarchical bond loss. Phase (B-1) utilizes the text encoder $E_{T}$ to convert text into global-level shape $(\bar{t}_s)$ and color $(\bar{t}_c)$ representations, as well as local-level features in the form of a textual graph $(g)$ . In Phase (B-2), the shape decoder generates the object hierarchically using the global input from Phase (B-1), with further use of a part-node attention module and refiner $(R)$ to connect part latents and graph node features for more precise guidance. + +![](images/9ad6853aa55a15ed8e6e18d0b2c15f2ba8b2e1307a08e61505bec802aa8c3f67.jpg) + +and a one-layer MLP for partitioning the color latent space. The PG has three layers for the shape latents and two layers for the color latents. Both SD and PG use LeakyReLU as intermediate activation layers and Sigmoid as the final one. Note SD and PG have shared weights within each level of the hierarchy tree, but different weights between the levels. + +Reconstruction Loss The total reconstruction loss of shape and color is calculated by summing the individual reconstruction losses at each level using Mean Squared Error (MSE), as shown below: + +$$ +L _ {r e c o n} = \lambda_ {r e c o n} \sum_ {l = 1} ^ {L} \left(S _ {g t} - S _ {l}\right) ^ {2}, \tag {1} +$$ + +$$ +L _ {c o l o r} = \lambda_ {c o l o r} \sum_ {l = 1} ^ {L} \left(C _ {g t} - C _ {l}\right) ^ {2}, \tag {2} +$$ + +where $S_{gt}$ and $C_{gt}$ represent the ground truth for shape and color, respectively. $S_{l}$ and $C_{l}$ are the assembled part's occupancy and color value at level $l$ . + +Hierarchical Bond Loss The generated shape parts, following the hierarchical tree structure, need to be constrained by the relationships between parent and child parts in terms of geometric composition. To ensure this, a bond loss is calculated between level $l$ and the next level $l + 1$ . The bond loss is defined as the error between the occupancy value of each parent part and the maximum of the occupancy values of its two derived parts: + +$$ +L _ {b o n d} ^ {l} = \sum_ {i = 1} ^ {2 ^ {l}} \left(S _ {i} ^ {l} - \max \left(S _ {2 i - 1} ^ {l + 1}, S _ {2 i} ^ {l + 1}\right)\right) ^ {2}, \tag {3} +$$ + +where $S_{i}^{l}$ is the occupancy value of the parent part at position $i$ and level $l$ , and $S_{2i-1}^{l+1}$ and $S_{2i}^{l+1}$ are the occupancy values of the two derived child parts. The total bond loss is computed by summing the bond loss across all levels when + +$$ +l < L \colon L _ {b o n d} = \lambda_ {b o n d} \sum_ {l = 1} ^ {l < L} L _ {b o n d} ^ {l}. +$$ + +Regularization Loss To reduce the complexity of the network and avoid overfitting, we introduce regularization loss $L_{reg}$ at the last layer of each Structure Divider at each level. + +This loss is calculated as the sum of the absolute values of all weight parameters, multiplied by a small regularization parameter $\lambda_{reg}$ . The regularization loss is given by: $L_{reg} = \lambda_{reg}\sum_{i=1}^{n}|w_i|$ , where $w_i$ is the $i$ -th weight parameter in the layer, and $n$ is the total number of weight parameters. Our experiments show that without using regularization loss, the generated shapes may contain parts that are over-segmented or lack reasonable shape. + +Comparison with Existing Method Our method differs from previous approaches that utilize (hierarchical) decoders for unsupervised shape part learning (such as in [9, 29]) in several ways. Firstly, our hierarchical decoder aims to learn semantic shape part-level implicit field rather than simply segmenting the shape into more parts. For example, when generating a chair with four legs, our approach considers the legs as a whole unit rather than segmenting them individually, as we typically describe the legs as a whole rather than as individual parts. To balance the network's fitting ability with its complexity, we utilize a regularization loss. Secondly, we also construct hierarchical color latent fields, which are often overlooked in other methods, to represent colored shapes. + +# 3.3. Text-Guided Shape Generation + +Fig. 2 (B) shows the text-guided shape generation network, comprising three modules: text encoder $E_{T}$ , textual graph encoder $E_{G}$ , and hierarchical shape decoder $D_{T}$ . The $D_{T}$ module is similar to the pre-trained shape-specific decoder $D_{S}$ (as described in Section 3.2), with the addition of a new part-node attention module and a refiner. This attention module allows the network to learn the local connections between shape parts/colors and textual graph entities, followed by a refiner module $R$ for final results generation. $D_{T}$ module is initialized from the pre-trained $D_{S}$ to leverage its predefined structural implicit fields and facilitate the generation process in a hierarchical manner. We provide a detailed description of each part in the following sections. + +Text Global Encoding We utilize a pre-trained BERT model [11] as the text encoder base. The encoded [CLS] token feature is used to extract global-level latent features $\bar{t}_s$ and $\bar{t}_c$ for shape and color, respectively, which serve as the input for the hierarchical shape decoder $D$ . Most existing approaches employ word-level features $w$ as fine-grained guidance in text-conditioned image or 3D shape generation tasks [40, 25]. However, we argue that the use of structured textual representation, such as the textual graph, would be more appropriate in aligning with shape structures. As a result, we opt to use the word features to initialize the graph node representation, as described below. + +![](images/11013f26976ef3e04cbef40cc29a0dac99ed76c9b3bca4b5861b11a5ca457b6e.jpg) +Figure 3. The dashed box shows a triplet $(v_{s}, v_{r}, v_{o})$ in the graph, denoting the subjective, relation, and objective nodes. The three nodes are first concatenated and then jointly processed using MLP, which outputs a new value for each node. The final updated state $v_{s}^{\prime}$ of $v_{s}$ is the result of pooling the corresponding values from all triplets in which the node $v_{s}$ is involved. + +Shape-Oriented Text Graph We use a unique text graph representation designed specifically for shapes. The graph nodes are classified into root, part, and attribute categories, which represent the class of the shape, the parts, and the attribute for root/part nodes, respectively. We use a special type node relation to connect these nodes, forming the edges. For example, the text "brown wooden chair with curved armrests" would be parsed as (chair: brown, wooden), (armrests: curved), and their relation triplet (chair, with, armrests) in the (subjective, relation, objective) convention. To construct graphs from the original text, we first use the spaCy library [15] to parse the text's semantic dependencies. These raw dependencies contain many linguistic terms and do not correspond well with real-world structures. To obtain more clear and shape-oriented graph formation, we introduce a custom parser targeting 3D shapes based on [36]. To ensure that the graph remains connected, we connect all part nodes with the root node using the with relation if no linguistic relation is found. The lower left of Fig. 2 illustrates these two forms of text. Finally, we use the resulting relation triplets for training. Further details can be found in the supplementary materials. + +Spatial-Aware Graph Encoding One challenge in using a shape-oriented graph, where entity or attribute nodes are connected through relation nodes, is the lack of a canonical grid for embedding the graph since nodes in such a graph exist in a non-Euclidean space. Existing methods based on Graph Convolution Networks (GCN) [21] typically rely on an adjacency matrix to represent the relations between nodes, which reduces their ability to represent spatial relations. To address the issue, we designed a novel Spatial-Aware Graph Encoder (SAGE) that differentiably encodes relation nodes in order to capture the structural relations between nodes. Specifically, we represent each graph triplet $g$ in the form $(v_{s}, v_{r}, v_{o})$ , where $v_{s}$ , $v_{o}$ , and $v_{r}$ represent + +![](images/82c3b4940e0d9e0d8cb72f13adbfd6a08763ad5fb401f2817363ec464ed30be2.jpg) +Figure 4. The Part-Node attention. The attention map is obtained by the dot-product between the features from the intermediate part latent feature $(f_s^l)$ at level $l$ and the graph node features $g$ . + +the values of subjective, objective, and relation nodes in a relationship, respectively. This makes the graph directional through the order of the nodes between the relation node, like chair has arms. As shown in Fig. 3, the computational layer in SAGE jointly encodes the nodes and their relation in each triplet, followed by a pooling layer that fuses the features of each node involved in more than one triplet. + +Part-Node Attention We introduce the Part-Node Attention modules for learning the relationship between intermediate parts/colors with textual graph nodes. These modules are based on the multi-head attention mechanism presented in [38]. For level $l$ , given the intermediate shape latents $f_{s}^{l} \in \mathbb{R}^{2^{l} \times d}$ as query, and the encoded graph node features, $g \in \mathbb{R}^{n \times d}$ ( $n$ is node number), as key and value, the calculation inside an attention head is: + +$$ +\hat {f} _ {s} ^ {l} = \operatorname {s o f t m a x} \left(\frac {Q K}{\sqrt {d _ {h}}}\right) V, \tag {4} +$$ + +where $Q = F_{Q}(f_{s}^{l})$ , $K = F_{K}(g)$ , $V = F_{V}(g)$ , and $F_{Q}, F_{K}$ , and $F_{V}$ are functions implemented using MLPs. $d$ is the feature dimension for $g$ and $f_{s}^{l}$ , while $d_{h}$ represents the dimension used within an attention head. The SoftMax operation is applied across the graph nodes, resulting in a weighted aggregation of the graph node features as the new intermediate part latents, $\hat{f}_{s}^{l}$ . This process is visualized in Fig. 4. Also, attention is applied between the paired color features $f_{c}^{l}$ with the graph nodes using a separate module. In practice, we use the attention module at level 2 of the hierarchical decoder due to the generated semantically detailed shape components aligned with graph nodes. + +Joint Refiner The hierarchical bond loss, introduced in Section 3.2, serves as the foundation for structure-aware generation by constraining the composition of shape parts to follow a hierarchical tree structure. However, we posit that this also negatively impacts the generated results, as demonstrated in the perception-distortion tradeoff [5]. To + +mitigate this issue, we introduce a refiner $R$ in one level that uses the level latents jointly to generate the final shape. The refiner is also an implicit shape decoder with six layers of MLPs. The final shape output is accompanied by an additional refiner reconstruction loss $L_{R}$ but without the need to consider the hierarchical bond, which allows for the incorporation of additional detail in the final shape. + +Training Strategy and Losses Training text-guided hierarchical shape generation directly presents significant challenges considering the big discrepancy between text and shape modalities. To address this, we propose a three-step strategy to ensure stable and effective training. First, we collect the shape and color latents from the shape encoder $E_{S}$ as guidance and train the text global feature encoder $E_{T}$ alone. The loss is using MSE: $L_{\text{latent}} = \lambda_{\text{latent}}^{s} (\bar{t}_{s} - \bar{f}_{s})^{2} + \lambda_{\text{latent}}^{c} (\bar{t}_{c} - \bar{f}_{c})^{2}$ , where $\bar{t}_{s}$ and $\bar{t}_{c}$ are encoded by text encoder while $\bar{f}_{s}$ and $\bar{f}_{c}$ are encoded by shape encoder. Second, we train the decoder with attention modules, freezing the decoder during the first ten epochs to allow for warmed-up training. Finally, we train the entire network together. The total loss then includes the aforementioned shape/color reconstruction losses and the bond loss in shape-specific pre-training (see Sec. 3.2), and the refiner loss. + +# 4. Experiments + +This section introduces our experimental design and implementation, followed by a comprehensive analysis of the results from various aspects and an ablation study. Supplementary material provides additional results and analysis. + +# 4.1. Experimental Settings and Implementation + +We conduct experiments using the text-shape dataset [8], which is derived from the 3D shape dataset, ShapeNet [7], and augmented with textual annotations. The dataset comprises approximately 15,000 shapes and 75,000 textual descriptions, encompassing the table and chair shape classes. We also adhere to the same training and testing splits as previously established. + +We implement our framework in PyTorch [32]. We set $N = 4096$ as the sampled points. We pre-train the shape-specific decoder for 1000 epochs and the text-guided model for another 500 epochs. The learning rate is constant $1e^{-4}$ using Adam [20] for all phases. The loss weights are set as follows: $\lambda_{\text{recon}} = 10.0$ , $\lambda_{\text{color}} = 1.0$ , $\lambda_{\text{bond}} = 1.0$ , $\lambda_{\text{reg}} = 1e^{-4}$ , $\lambda_{\text{latent}}^s = 10.0$ and $\lambda_{\text{latent}}^c = 1.0$ . Detailed training strategy can be found in Sec. 3.3. Additionally, as our model outputs shapes at different levels, we select the results from level 2, as it is at this level that most shapes' parts have been parsed and refined. Level 3 is used as a spare level for monitoring no further parts are parsed. + +
IoU (↑)IS (↑)EMD (↓)Acc (↑)
Text2Shape [8]9.641.960.444397.37
T2S-Implicit [25]12.211.970.207197.48
ours13.651.980.176797.5
+ +Table 1. Quantitative comparison with state-of-the-art. $\uparrow$ and $\downarrow$ indicate higher or lower is better, respectively. + +# 4.2. Text-Guided Shape Generation + +Comparison with the State-of-the-art We compare our method to two existing text-conditioned colored shape generation methods, Text2Shape [8] and T2S-Implicit [25]. We employ the same quantitative metrics for evaluation, including (1) Mean intersection-over-union (IoU) for shape occupancy measurement, regardless of color; (2) Earth Mover's Distance (EMD) for color distribution evaluation; (3) Inception Score (IS) to quantify the realism of the results; and (4) Accuracy (Acc) computed by a pre-trained shape classifier to indicate whether the predicted class matches the ground truth. To ensure fairness in comparison, we convert the generated shapes to volumes of the same size as in [8, 25] before calculating the quantitative values. The comparison results are presented in Table 1, where the numbers of the existing methods are directly taken from the respective papers. Our method exhibits superior performance across all metrics. The qualitative result presented in Fig. 5 provides further evidence of the effectiveness of our approach. As can be observed, our method achieves a higher degree of correspondence between the predicted shape structure and the text description, as demonstrated in Row 1 of the figure with the example of "wood armrests". Additionally, our approach exhibits an increased level of completeness in part generation, as evidenced by the absence of incomplete parts or extra artifacts, which are clearly labeled with red circles in the results generated by the existing methods. + +Part-level Refinement As T2S-Implicit [25] also employs an implicit decoder similar to our approach, we conduct a comparison with it to highlight the distinctions, particularly with regard to part-level details. The results are presented in high resolution in Fig. 6. We observed that [25] tends to struggle with smaller or complex parts, such as the legs of a sofa or the wheels of an office chair. In contrast, our method is capable of interpreting these parts explicitly, resulting in more realistic part-level generation. Another advantage of our approach is that the generation of a single part does not affect the other parts, which is a common issue in previous methods. For example, generating "silver legs" may result in the bottom of the sofa appearing slightly silver in the connection area, creating a noticeable color artifact. + +![](images/600bd8836ac32bd08e535769f6dbe9df18258f8108306b5fc709ea2e18933ab8.jpg) +Figure 5. Visualization of the results of our method in comparison to Text2Shape [8] and T2S-Implicit [25]. Red circles indicate incomplete parts or extra artifacts. + +![](images/99197c7aa1e4264c34473f139217f16785164e8c1644bc0b8ba05c3dde6312e1.jpg) +Figure 6. Comparison of part-level refinement. + +Part-level Manipulation The ability to learn structure-based 3D shape representations and align them with textual descriptions enables precise and convenient manipulation of shapes. In the process of generating a shape from text, our method allows for manipulation at various levels, including modification of the part-level structure, attributes, or colors. The results of this manipulation are illustrated in Fig. 7. The changes made in the text appear to reflect the changes in part structure (a), part attribute (c), and part-related color refinement (b), suggesting a possible consistency between the constituent parts of the shape and the corresponding text. + +# 4.3. Ablation Studies + +We conduct extensive ablation studies to validate the effectiveness of the proposed hierarchical structure-based representation, graph-part attention, and joint refiner. We also use Frechet Inception Distance (FID) and CLIP [34] to evaluate the quality of the generated shapes and the similarity between the generated shapes and their corresponding text descriptions, respectively. More detailed information on the metrics can be found in the supplementary material. The re + +
MethodLevel 1Level 2Level 3
IOU↑FID↓CLIP↑IOU↑FID↓CLIP↑IOU↑FID↓CLIP↑
HierD [base]11.2017.4745.4311.0717.5245.3610.8517.6043.92
FlatD10.3819.5540.72------
HierD + WordAttn (L2)11.1317.4244.8811.6616.4647.4310.7217.6946.10
HierD + GraphAttn (L2)11.2917.3645.0112.6316.3150.6412.4116.7050.01
HierD + GraphAttn (L1)11.5416.8652.7511.3017.0951.4011.2717.1250.56
HierD + GraphAttn (L2) + Refiner(L2) [Full]11.1917.3847.2713.6515.8154.8912.2216.3452.20
+ +Table 2. Results of the ablation study on output levels. The base model, HierD, is the proposed hierarchical decoder using sentence-level guidance only. L1/L2 indicates component application at Level 1/Level 2. ↑/↓ denote better higher/ lower results, respectively. + +![](images/427d5cbf36a613a16552bc79389deeb980b69b58291ef22749439213715305f5.jpg) +Figure 7. Manipulation of text results in corresponding changes at the part level. The second row shows the predicted structure from Level 1 to 2. (a) illustrates the removal of a part, (b) shows a change in part color, and (c) depicts a change in the part attribute. + +sults of our ablation studies are presented in Table 2. We use the Hierarchical Decoder with sentence-level guidance (HierD) as a base model, and then add various modules such as word-level attention (WordAttn), our proposed graph-part attention (GraphAttn), and the joint refiner to different levels of the hierarchy (L1/L2 for Level 1 and 2). We present the results for each level for an in-depth analysis of the impact of the hierarchy level on the performance of the model. + +Hierarchical Representation We compare HierD to a flattened decoder (FlatD) which directly outputs 8 parts at Level 1. Utilizing the same settings as HierD, FlatD demonstrates inferior performance. Our qualitative analysis reveals that FlatD is more prone to overfitting, resulting in the generation of fewer parts that are overly complex geometrically, instead of generating finer parts. + +Graph-based Attention We have observed the clear advantages of utilizing graph-based attention over word-based attention. Specifically, when applied to Level 2, the use of GraphAttn results in superior performance across all metrics. Furthermore, the results from Level 3 also demonstrate improved performance. This can be attributed to the semantic alignment of the entity nodes in the graph with the intermediate parts. However, when applied to Level 1, the use of + +GraphAttn does not yield similar results as in Level 2. We speculate that this discrepancy may be due to the coarser nature of the parts at Level 1, which hinders the ability to establish effective connections between shape parts and textual graph entities. + +Joint Refiner The utilization of a refiner as the final shape output in our method is driven by the observation of the perception-distortion tradeoff [5]. As the level increases, we have noted a slight decline in performance as compared to previous levels. For instance, the IOU in Level 2 is consistently lower than that in Level 1. We posit that the hierarchical bond loss employed in our approach results in a stronger structural connection, but may impede the generation of detailed features on top of the structure. The incorporation of a refiner, on the other hand, has resulted in a notable improvement in all metrics. + +# 5. Conclusion, limitation, and future work + +The ShapeScaffold is a method for generating shapes guided by text, with the understanding that both shape and text possess internal structures that can be aligned to improve the outcome. We represent shapes as hierarchical structures and model text as scene graphs. Our part-node attention mechanism allows us to learn correspondences between shape and text at various hierarchical levels, which has been shown to effectively establish connections between shape parts/colors and linguistic entities. The experiments have demonstrated the superiority of our approach in both shape generation and text-shape/color consistency. + +Several limitations should be acknowledged. Firstly, the current text-to-shape dataset is limited to chairs and tables, so further research is needed to explore the generalization of our method to other shape classes. Secondly, the results may be influenced by the parts priority problem, where incorrect occupancy state of parts junctions may result in unreasonable visualization. Therefore, further incorporation of prior interpretations of shapes should be considered in future studies. Lastly, it would be beneficial to explore scene generation that takes structure into account, with a focus on the composition of different shapes as an outer structure. + +Acknowledgements. This work is supported by RCUK grant CAMERA (EP/M023281/1, EP/T022523/1), Centre for Augmented Reasoning (CAR) at the Australian Institute for Machine Learning, and a gift from Adobe. + +# References + +[1] Panos Achlioptas, Judy Fan, Robert Hawkins, Noah Goodman, and Leonidas J Guibas. Shapeplot: Learning language for shape differentiation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8938-8947, 2019. 2 +[2] Oron Ashual and Lior Wolf. Specifying object attributes and relations in interactive scene generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4561-4569, 2019. 3 +[3] Renée Baillargeon. Infants' physical world. Current directions in psychological science, 13(3):89-94, 2004. 1 +[4] Irving Biederman. Human image understanding: Recent research and a theory. Computer vision, graphics, and image processing, 32(1):29-73, 1985. 1 +[5] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 6, 8 +[6] Angel Chang, Manolis Savva, and Christopher D Manning. Learning spatial knowledge for text to 3d scene generation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 2028-2038, 2014. 2 +[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 6 +[8] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian conference on computer vision, pages 100-116. Springer, 2018. 2, 6, 7 +[9] Zhiqin Chen, Kangxue Yin, Matthew Fisher, Siddhartha Chaudhuri, and Hao Zhang. Bae-net: Branched autoencoder for shape co-segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8490-8499, 2019. 2, 5 +[10] Morten H Christiansen and Nick Chater. Language as shaped by the brain. Behavioral and brain sciences, 31(5):489-509, 2008. 1 +[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5 +[12] Helisa Dhamo, Azade Farshad, Iro Laina, Nassir Navab, Gregory D Hager, Federico Tombari, and Christian Rupprecht. Semantic image manipulation using scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5213-5222, 2020. 3 +[13] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecraft: A recursive text-conditioned 3d shape + +generation model. arXiv preprint arXiv:2207.09446, 2022. 2 +[14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proceedings of the 27th International Conference on Neural Information Processing Systems - Volume 2, NIPS'14, page 2672-2680, 2014. 2 +[15] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. spaCy: Industrial-strength Natural Language Processing in Python, 2020. 5 +[16] Fangqiao Hu, Jin Zhao, Yong Huang, and Hui Li. Learning structural graph layouts and 3d shapes for long span bridges 3d reconstruction. arXiv preprint arXiv:1907.03387, 2019. 2 +[17] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2 +[18] R. Kenny Jones, Aalia Habib, Rana Hanocka, and Daniel Ritchie. The neurally-guided shape parser: Grammar-based labeling of 3d shape regions with approximate inference. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[19] R Kenny Jones, Aalia Habib, and Daniel Ritchie. Shred: 3d shape region decomposition with learned local operations. arXiv preprint arXiv:2206.03480, 2022. 2 +[20] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 6 +[21] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 5 +[22] Jun Li, Kai Xu, Siddhartha Chaudhuri, Ersin Yumer, Hao Zhang, and Leonidas Guibas. Grass: Generative recursive autoencoders for shape structures. ACM Transactions on Graphics (TOG), 36(4):1-14, 2017. 2 +[23] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440, 2022. 2 +[24] Chunxiao Liu, Zhendong Mao, Tianzhu Zhang, Hongtao Xie, Bin Wang, and Yongdong Zhang. Graph structured network for image-text matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10921-10930, 2020. 3 +[25] Zhengzhe Liu, Yi Wang, Xiaojuan Qi, and Chi-Wing Fu. Towards implicit text-guided 3d shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17896-17906, 2022. 2, 5, 7 +[26] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2 +[27] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 306-315, 2022. 2 +[28] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 2 +[29] Chengjie Niu, Manyi Li, Kai Xu, and Hao Zhang. Rim-net: Recursive implicit fields for unsupervised learning of hierarchical shape structures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11779-11788, 2022. 2, 5 +[30] Despoina Paschalidou, Luc Van Gool, and Andreas Geiger. Learning unsupervised hierarchical part decomposition of 3d objects from a single rgb image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1060-1070, 2020. 2 +[31] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics revisited: Learning 3d shape parsing beyond cuboids. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10344–10353, 2019. 2 +[32] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 6 +[33] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 7 +[35] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 2 +[36] Richard Socher, Cliff C Lin, Chris Manning, and Andrew Y Ng. Parsing natural scenes and natural language with recursive neural networks. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 129–136, 2011. 5 + +[37] Shubham Tulsiani, Hao Su, Leonidas J Guibas, Alexei A Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2635-2643, 2017. 2 +[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 6 +[39] Yanzhen Wang, Kai Xu, Jun Li, Hao Zhang, Ariel Shamir, Ligang Liu, Zhiquan Cheng, and Yueshan Xiong. Symmetry hierarchy of man-made objects. In Computer graphics forum, volume 30, pages 287-296. Wiley Online Library, 2011. 2 +[40] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 3, 5 \ No newline at end of file diff --git a/2023/ShapeScaffolder_ Structure-Aware 3D Shape Generation from Text/images.zip b/2023/ShapeScaffolder_ Structure-Aware 3D Shape Generation from Text/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1a8063a980144e3bbb920f8e97344982988bb500 --- /dev/null +++ b/2023/ShapeScaffolder_ Structure-Aware 3D Shape Generation from Text/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db7df8734cab50c127e9bba4f93f2bae567969e886c2f3a4d62b15f2f6d2b176 +size 476131 diff --git a/2023/ShapeScaffolder_ Structure-Aware 3D Shape Generation from Text/layout.json b/2023/ShapeScaffolder_ Structure-Aware 3D Shape Generation from Text/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f5feb9082149940ef1ab89ca86367937eab4f9d1 --- /dev/null +++ b/2023/ShapeScaffolder_ Structure-Aware 3D Shape Generation from Text/layout.json @@ -0,0 +1,7554 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 504, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 504, + 121 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 504, + 121 + ], + "type": "text", + "content": "ShapeScaffold: Structure-Aware 3D Shape Generation from Text" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 103, + 144, + 194, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 144, + 194, + 198 + ], + "spans": [ + { + "bbox": [ + 103, + 144, + 194, + 198 + ], + "type": "text", + "content": "Xi Tian \nUniversity of Bath \nBath, UK \nxt275@bath.ac.uk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 223, + 144, + 334, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 144, + 334, + 198 + ], + "spans": [ + { + "bbox": [ + 223, + 144, + 334, + 198 + ], + "type": "text", + "content": "Yong-Liang Yang University of Bath Bath, UK y.yang@cs.bath.ac.uk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 363, + 144, + 491, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 144, + 491, + 198 + ], + "spans": [ + { + "bbox": [ + 363, + 144, + 491, + 198 + ], + "type": "text", + "content": "Qi Wu \nUniversity of Adelaide \nAdelaide, Australia \nqi.wu01@adelaide.edu.au" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 251, + 290, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 290, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 290, + 576 + ], + "type": "text", + "content": "We present ShapeScaffold, a structure-based neural network for generating colored 3D shapes based on text input. The approach, similar to providing scaffolds as internal structural supports and adding more details to them, aims to capture finer text-shape connections and improve the quality of generated shapes. Traditional text-to-shape methods often generate 3D shapes as a whole. However, humans tend to understand both shape and text as being structure-based. For example, a table is interpreted as being composed of legs, a seat, and a back; similarly, texts possess inherent linguistic structures that can be analyzed as dependency graphs, depicting the relationships between entities within the text. We believe structure-aware shape generation can bring finer text-shape connections and improve shape generation quality. However, the lack of explicit shape structure and the high freedom of text structure make cross-modality learning challenging. To address these challenges, we first build the structured shape implicit fields in an unsupervised manner. We then propose the part-level attention mechanism between shape parts and textual graph nodes to align the two modalities at the structural level. Finally, we employ a shape refiner to add further detail to the predicted structure, yielding the final results. Extensive experimentation demonstrates that our approaches outperform state-of-the-art methods in terms of both shape fidelity and shape-text matching. Our methods also allow for part-level manipulation and improved part-level completeness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": "The advance of 3D representation learning and generative models has sparked increasing interest in 3D shape generation. However, text-guided shape generation remains a challenging task. Many current approaches generate the 3D shape as a whole when no part information is utilized, while others treat text as a simple collection of words in order to provide finer guidance. Both shape and text, however, possess inherent internal structures that can be leveraged to" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 228, + 544, + 366 + ], + "blocks": [ + { + "bbox": [ + 310, + 228, + 544, + 366 + ], + "lines": [ + { + "bbox": [ + 310, + 228, + 544, + 366 + ], + "spans": [ + { + "bbox": [ + 310, + 228, + 544, + 366 + ], + "type": "image", + "image_path": "88baeeeb684e01923109f07d576ce08639792c69feaf7abf5b15f6997ef3dc84.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 371, + 546, + 415 + ], + "lines": [ + { + "bbox": [ + 305, + 371, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 371, + 546, + 415 + ], + "type": "text", + "content": "Figure 1. Text to shape generation with structure awareness. Raw description generates the initial coarse shape with structures. Further correspondence is built between shape parts and text graph nodes for fine-grained refinement." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 437, + 514, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 514, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 514, + 449 + ], + "type": "text", + "content": "improve the alignment between the two modalities." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 449, + 546, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 546, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 546, + 616 + ], + "type": "text", + "content": "Humans are born with a sense of physical understanding of the world [3]. Studies have shown that the human visual system perceives objects as hierarchical arrangements of parts [4], comprising high-level structural properties and low-level details. In addition, humans use language to describe objects, including their appearance, structure, and functions. It has been discovered that language, shaped by the human brain over time, has evolved to have hierarchical linguistic structures in order to convey more complex meanings [10]. This suggests that structure-based reasoning is present in both the visual and language modalities. While this is intuitive and effortless for humans, it is a challenge for intelligent algorithms to understand the structural connection between these two modalities." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": "In this research, we aim to augment shape generation from text with an awareness of structure in both shape and language modalities. This would allow for shape generation to be guided not only by high-level descriptions, such as a sentence but also refined with part-level guidance from semantic text entities. This process is similar to how humans create visual artworks: sketching an idea's gist and then progressively refining the details. However, there are" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2715" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 191 + ], + "type": "text", + "content": "several challenges to overcome. One challenge is that the shape structures are unknown unless annotated segmentation labels are provided. Even with annotation, where the process is time-consuming and labor-intensive, the definition of \"parts\" in shapes is subjective and ambiguous at multiple levels. Another challenge is that text descriptions of shapes are diverse and complex, making it difficult to optimize learning. In addition, for colored shapes, the alignment between structure and color representation must also be taken into consideration." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 288, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 288, + 430 + ], + "type": "text", + "content": "To address the challenges above, we propose a two-stage approach. First, we leverage an unsupervised hierarchical shape structure given part hierarchy is a prevalent and effective approach for representing shape structures, as demonstrated in previous research [22, 28]. To achieve this, we introduce a decomposition model to progressively decompose the shape into parts at multiple levels of granularity, from coarse to fine, such that each level can reconstruct the entire shape. Additionally, we learn color reconstruction at the part level. Fig. 1 illustrates the process. While the structures and parts identified in the first stage hold significance within the shape domain, they do not have a corresponding relationship with the text. In the second step, we establish a connection between the shape structure and the text structure. We utilize linguistic dependency graphs, which can be used to represent the structural elements present in the text, as attention signals to form the structure-level connection. Through these two steps, we are able to improve the generation results at the part level by building upon the established structure." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 431, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 287, + 586 + ], + "type": "text", + "content": "In our experimentation on the ShapeNet dataset [7], we have demonstrated the efficacy of utilizing a structure-aware representation in improving the correspondence between text and shape. As a result of the part-aware refinement, we observe a significant improvement in the quality of generated shapes, particularly in terms of their constituent parts. Additionally, our approach attains state-of-the-art results regarding several commonly used metrics such as Intersection over Union (IOU), Accuracy, and Inception Score (IS). Furthermore, through further exploration of the latent shape structure space, it is demonstrated that our method has the capability of generating previously unseen shapes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 586, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 287, + 658 + ], + "type": "text", + "content": "The main contributions of this work are: (1) An unsupervised hierarchical decoder for building shape structural implicit field. (2) A structure-aware text-to-shape generation method that improves part-level correspondence and refinement. (3) Comprehensive experiments that demonstrate the effectiveness of our method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 669, + 134, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 134, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 134, + 681 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Text Guided 3D Shape Generation. Deep learning has led to a surge in works exploring text-guided 3D shape or" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "scene generation [6, 8, 1, 25, 27, 13]. Recently, to achieve faithful results, some works tend to adopt pre-trained CLIP models [35, 17, 26] or diffusion-based methods [33, 23] for large-scale learning on open datasets. Our work, however, focuses on the classic ShapeNet dataset [7], examining the generation of structure-aware, colored shapes from text and comparing our results to previous baselines." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 157, + 546, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 546, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 546, + 408 + ], + "type": "text", + "content": "Generating 3D shapes directly from text remains a challenge, particularly when attempting to concurrently generate colors. [8] proposes a method that learns a joint text-shape embedding space and then uses a generative adversarial network (GAN) [14] to generate colored voxels from text. However, this approach has not been successful in terms of shape resolution, color quality, and cross-modality consistency. To address these issues, [25] proposes a method that uses an implicit 3D representation with word-level attention to improve control over generated shape structures and colors. [27] and [13] both adopt a discrete autoencoder to capture patch-based shape priors and subsequently employ a transformer for autoregressive generation. However, it is worth noting that these methods do not take into account color information, which is an essential attribute of object appearance. Also, all these methods have treated the shape as a whole, ignoring the internal structure within the shape. Additionally, text modeling has been limited to language-level features, such as sentence and word features, rather than considering the structure of the text for structure-level matching with the shape domain." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 426, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 546, + 715 + ], + "type": "text", + "content": "3D Structure-Aware Representation and Parsing. Although structure-based shape generation with text guidance has not been previously explored, there is a significant amount of research on learning structure-aware 3D shape representation or parsing 3D shapes into parts. For supervised methods, [22] proposes the use of a symmetry hierarchy [39] for hierarchical shape structure representation, while StructureNet [28] represents shapes using a graph neural network that considers both primitives and hierarchies. Similarly, [16] utilizes a binary parsing tree to represent the 3D structure of a cable-stayed bridge. Recently, semantic-based shape decomposition or parsing methods have been proposed in [19, 18], which use learned operations to obtain grammar-level shape parts. Unsupervised methods such as [37, 31] use convolutional neural networks (CNNs) to generate primitive shapes, e.g., cuboids or superquadrics, which abstract the input shape. BAE-NET [9] attempts to co-segment shapes into parts using a branched autoencoder architecture that disentangles common features for segmentation. However, these methods only predict the parts without considering their dependencies in the shape structure. [30] recovers the 3D objects into a hierarchy of parts directly from an RGB image using CNN, while the parts are primitive and lack details. RIM-NET [29] infers" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2716" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "hierarchical structures of 3D shapes using recursive implicit fields. It decomposes an input shape into two parts at each level, resulting in a binary tree hierarchy, where each level of the tree corresponds to an assembly of shape parts represented as implicit functions. However, the method does not consider color information and only processes one shape at a time during training. Our work uses a hierarchical tree structure for shape representation, but also includes color information for part-level implicit representation. In addition, our model is able to learn a shared latent space for a set of shapes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "text", + "content": "Text as Textual Graph Guidance. Text-to-image (T2I) generation has seen significant progress in recent years, with a focus on both generating realistic images and increasing consistency between texts and images. Some approaches use attention-based methods [40] for fine-grained alignment, while others leverage shared structures in images and texts. Scene graphs, which are derived from texts and contain expressive structural relationships between entities, have been widely used to reduce the cross-modality gap between texts and images for tasks such as text-image matching [24], fine-grained image generation [2], and image manipulation [12]. Given that both shapes and texts contain structural elements, we opt to employ textual graphs as additional guidance in conjunction with the generation of structural parts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 411, + 104, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 104, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 104, + 423 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 431, + 115, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 115, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 115, + 443 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "content": "The task is to generate 3D shapes with colors based on the given text description " + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "content": ". The predicted object is denoted as " + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "inline_equation", + "content": "O \\in \\mathbb{R}^{N \\times 4}" + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "content": ", comprising the shape occupancy values " + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "content": " and the color RGB values " + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "inline_equation", + "content": "C \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "content": " is the number of sample points." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "The overview of our method is illustrated in Fig. 2. Our framework consists of a text encoder " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": ", textual graph encoder " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "E_{G}" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": ", hierarchical structure-aware shape decoder " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " and shape encoder " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "E_{S}" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": ". In contrast to general methods that generate the whole shape at once, our approach, called ShapeScaffold, aims to generate the 3D shape by considering its internal structures, enabling refinement at the part level and improvement of text-shape correspondence. To achieve this, we propose to capture the part-level information from both shape and text. For the shape, we introduce the concept of structural implicit field in the shape decoder to capture the underlying structure and relationships of shape parts and colors. As shown in Fig. 2 (middle right), the structure latent fields capture the semantic shape parts for a chair, including arm, cushion, legs, etc. For the text, in addition to encoding it as a global feature such as a sentence vector, we also propose using a linguistic tex" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "tual graph parsed from the text that includes explicit entities with relations. This enables the structure latent fields to interact with textual entities at a finer level, resulting in more direct and explicit correspondence between the text description, such as \"curved armrests\" and the corresponding arm part of the generated shape. However, it is non-trivial to train a hierarchical structure-aware generator conditioned on text. We approach this task by dividing the training process into two stages: (A) for learning the structural implicit field using an autoencoder, and (B) for generating shapes guided by text, as summarized in Fig. 2. Stage B can be further divided into text encoding (B-1) and shape generation (B-2). The following describes each stage in detail." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 236, + 459, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 236, + 459, + 249 + ], + "spans": [ + { + "bbox": [ + 305, + 236, + 459, + 249 + ], + "type": "text", + "content": "3.2. Shape-Specific Pre-training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 255, + 547, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 547, + 422 + ], + "type": "text", + "content": "To capture the internal semantic structures of shapes and facilitate shape decoding conditioned on text, we employ a pre-training technique using a shape-specific hierarchical decoder " + }, + { + "bbox": [ + 304, + 255, + 547, + 422 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 304, + 255, + 547, + 422 + ], + "type": "text", + "content": " and a shape encoder " + }, + { + "bbox": [ + 304, + 255, + 547, + 422 + ], + "type": "inline_equation", + "content": "E_{S}" + }, + { + "bbox": [ + 304, + 255, + 547, + 422 + ], + "type": "text", + "content": ", forming an autoencoder (see Fig. 2 (A)). The encoder " + }, + { + "bbox": [ + 304, + 255, + 547, + 422 + ], + "type": "inline_equation", + "content": "E_{S}" + }, + { + "bbox": [ + 304, + 255, + 547, + 422 + ], + "type": "text", + "content": ", which is a convolutional neural network (CNN), takes the volume of an object " + }, + { + "bbox": [ + 304, + 255, + 547, + 422 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 255, + 547, + 422 + ], + "type": "text", + "content": " as input and produces latent shape and color features to be used by the decoder. The decoder then creates an implicit field by partitioning the shape " + }, + { + "bbox": [ + 304, + 255, + 547, + 422 + ], + "type": "inline_equation", + "content": "(\\bar{f}_s)" + }, + { + "bbox": [ + 304, + 255, + 547, + 422 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 304, + 255, + 547, + 422 + ], + "type": "inline_equation", + "content": "(\\bar{f}_c)" + }, + { + "bbox": [ + 304, + 255, + 547, + 422 + ], + "type": "text", + "content": " latents at each hierarchical level, reconstructing the latent features into parts and assembling them into the final object. An example of a successfully learned structural implicit field is shown in Fig. 2 (right), where semantic shape parts can be reconstructed from their corresponding latent positions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "text", + "content": "Hierarchical Shape Decoder The Hierarchical Decoder consists of two modules at each level: the Structure Divider (SD) and the Part Generator (PG). Given a pair of shape/color latent features " + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "inline_equation", + "content": "(f_s^{l,i} \\in \\mathbb{R}^d, f_c^{l,i} \\in \\mathbb{R}^d)" + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "text", + "content": " at level " + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "text", + "content": " and position " + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "inline_equation", + "content": "i(1 \\leq i \\leq 2^l)" + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "text", + "content": ", the SD first divides the shape latent into two child shape features and then generates the color latent for each child conditioned on the corresponding child shape latent. The SD constructs the latent space for both shape and color, as depicted by the grey and yellow squares in Fig. 2. The PG then generates shape parts one by one from its shape/color latents at a given set of point coordinates " + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = \\{(x,y,z) \\mid x,y,z \\in \\mathbb{R}\\}" + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "text", + "content": ". For each point " + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "text", + "content": ", the input is the concatenation of " + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "text", + "content": " with the latent value, and SD outputs one occupancy probability for the shape channel and three values (RGB) for the color channel. The resulting generated shape " + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "inline_equation", + "content": "O_l" + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "text", + "content": " at level " + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "text", + "content": " is a composite of generated parts, where a point is occupied by a part with the highest occupancy probability, provided that this probability exceeds a pre-defined threshold. The color of each point is determined based on the corresponding color values of the occupying part. Both SD and PG are implemented using multi-layer perception (MLPs). The SD is implemented using a two-layer MLP for partitioning the shape latent space" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2717" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 75, + 260, + 182 + ], + "blocks": [ + { + "bbox": [ + 56, + 75, + 260, + 182 + ], + "lines": [ + { + "bbox": [ + 56, + 75, + 260, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 75, + 260, + 182 + ], + "type": "image", + "image_path": "6d54f528674ff29953f5258de83556ceed6e92a624e69882d34b7d76aea344a7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 74, + 412, + 84 + ], + "lines": [ + { + "bbox": [ + 263, + 74, + 412, + 84 + ], + "spans": [ + { + "bbox": [ + 263, + 74, + 412, + 84 + ], + "type": "text", + "content": "(B-2) T2S: Hierarchical Generation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 263, + 87, + 414, + 131 + ], + "blocks": [ + { + "bbox": [ + 263, + 87, + 414, + 131 + ], + "lines": [ + { + "bbox": [ + 263, + 87, + 414, + 131 + ], + "spans": [ + { + "bbox": [ + 263, + 87, + 414, + 131 + ], + "type": "image", + "image_path": "a1b3617f7ad718a23a1aa3393332ec78b391dbc9a9f69a24902cda075583b42b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 417, + 73, + 539, + 148 + ], + "blocks": [ + { + "bbox": [ + 417, + 73, + 539, + 148 + ], + "lines": [ + { + "bbox": [ + 417, + 73, + 539, + 148 + ], + "spans": [ + { + "bbox": [ + 417, + 73, + 539, + 148 + ], + "type": "image", + "image_path": "33116be9703471b80db4230b10f9ea70463e80f592ba955af3fcc77e720a6eca.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 204, + 263, + 281 + ], + "blocks": [ + { + "bbox": [ + 58, + 188, + 167, + 201 + ], + "lines": [ + { + "bbox": [ + 58, + 188, + 167, + 201 + ], + "spans": [ + { + "bbox": [ + 58, + 188, + 167, + 201 + ], + "type": "text", + "content": "(B-1) T2S: Text Encoding" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 204, + 263, + 281 + ], + "lines": [ + { + "bbox": [ + 56, + 204, + 263, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 204, + 263, + 281 + ], + "type": "image", + "image_path": "9555a389380805c697c84df83a9ee72ac5495f30b273f409b10af1ea303c4e97.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 263, + 132, + 539, + 297 + ], + "blocks": [ + { + "bbox": [ + 263, + 132, + 539, + 297 + ], + "lines": [ + { + "bbox": [ + 263, + 132, + 539, + 297 + ], + "spans": [ + { + "bbox": [ + 263, + 132, + 539, + 297 + ], + "type": "image", + "image_path": "3b1a71df0864ffd4d10b5db68800b64f4c7fca8b4eca02d5128d532bdf35cdfd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 57, + 282, + 257, + 363 + ], + "blocks": [ + { + "bbox": [ + 57, + 282, + 257, + 363 + ], + "lines": [ + { + "bbox": [ + 57, + 282, + 257, + 363 + ], + "spans": [ + { + "bbox": [ + 57, + 282, + 257, + 363 + ], + "type": "image", + "image_path": "45d8c22361b1a09b010ff5a3fb38fe2d2a7932e6d424d5344312b70fd1e6afe9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 367, + 545, + 434 + ], + "lines": [ + { + "bbox": [ + 46, + 367, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 545, + 434 + ], + "type": "text", + "content": "Figure 2. Method overview. The proposed method includes two phases: shape autoencoder pre-training (A) and text-to-shape (T2S) phase including (B-1) and (B-2). Phase (A) trains the shape-specific decoder using shape/color reconstruction loss and hierarchical bond loss. Phase (B-1) utilizes the text encoder " + }, + { + "bbox": [ + 46, + 367, + 545, + 434 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 46, + 367, + 545, + 434 + ], + "type": "text", + "content": " to convert text into global-level shape " + }, + { + "bbox": [ + 46, + 367, + 545, + 434 + ], + "type": "inline_equation", + "content": "(\\bar{t}_s)" + }, + { + "bbox": [ + 46, + 367, + 545, + 434 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 46, + 367, + 545, + 434 + ], + "type": "inline_equation", + "content": "(\\bar{t}_c)" + }, + { + "bbox": [ + 46, + 367, + 545, + 434 + ], + "type": "text", + "content": " representations, as well as local-level features in the form of a textual graph " + }, + { + "bbox": [ + 46, + 367, + 545, + 434 + ], + "type": "inline_equation", + "content": "(g)" + }, + { + "bbox": [ + 46, + 367, + 545, + 434 + ], + "type": "text", + "content": ". In Phase (B-2), the shape decoder generates the object hierarchically using the global input from Phase (B-1), with further use of a part-node attention module and refiner " + }, + { + "bbox": [ + 46, + 367, + 545, + 434 + ], + "type": "inline_equation", + "content": "(R)" + }, + { + "bbox": [ + 46, + 367, + 545, + 434 + ], + "type": "text", + "content": " to connect part latents and graph node features for more precise guidance." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 263, + 297, + 539, + 361 + ], + "blocks": [ + { + "bbox": [ + 263, + 297, + 539, + 361 + ], + "lines": [ + { + "bbox": [ + 263, + 297, + 539, + 361 + ], + "spans": [ + { + "bbox": [ + 263, + 297, + 539, + 361 + ], + "type": "image", + "image_path": "9ad6853aa55a15ed8e6e18d0b2c15f2ba8b2e1307a08e61505bec802aa8c3f67.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 453, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 525 + ], + "type": "text", + "content": "and a one-layer MLP for partitioning the color latent space. The PG has three layers for the shape latents and two layers for the color latents. Both SD and PG use LeakyReLU as intermediate activation layers and Sigmoid as the final one. Note SD and PG have shared weights within each level of the hierarchy tree, but different weights between the levels." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 540, + 287, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 588 + ], + "type": "text", + "content": "Reconstruction Loss The total reconstruction loss of shape and color is calculated by summing the individual reconstruction losses at each level using Mean Squared Error (MSE), as shown below:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 97, + 599, + 287, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 599, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 97, + 599, + 287, + 632 + ], + "type": "interline_equation", + "content": "L _ {r e c o n} = \\lambda_ {r e c o n} \\sum_ {l = 1} ^ {L} \\left(S _ {g t} - S _ {l}\\right) ^ {2}, \\tag {1}", + "image_path": "0a707beda8703881a3042e0e99a28cd299ffe40dde850b8644c86aaa7b4a325f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 99, + 639, + 287, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 639, + 287, + 672 + ], + "spans": [ + { + "bbox": [ + 99, + 639, + 287, + 672 + ], + "type": "interline_equation", + "content": "L _ {c o l o r} = \\lambda_ {c o l o r} \\sum_ {l = 1} ^ {L} \\left(C _ {g t} - C _ {l}\\right) ^ {2}, \\tag {2}", + "image_path": "32873e57a9a0f4e367e6d073c7964fd4ac82fe5319a312657404acc6da59ff80.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "S_{gt}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "C_{gt}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " represent the ground truth for shape and color, respectively. " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "S_{l}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "C_{l}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " are the assembled part's occupancy and color value at level " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 453, + 545, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 549 + ], + "type": "text", + "content": "Hierarchical Bond Loss The generated shape parts, following the hierarchical tree structure, need to be constrained by the relationships between parent and child parts in terms of geometric composition. To ensure this, a bond loss is calculated between level " + }, + { + "bbox": [ + 304, + 453, + 545, + 549 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 453, + 545, + 549 + ], + "type": "text", + "content": " and the next level " + }, + { + "bbox": [ + 304, + 453, + 545, + 549 + ], + "type": "inline_equation", + "content": "l + 1" + }, + { + "bbox": [ + 304, + 453, + 545, + 549 + ], + "type": "text", + "content": ". The bond loss is defined as the error between the occupancy value of each parent part and the maximum of the occupancy values of its two derived parts:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 343, + 559, + 545, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 559, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 343, + 559, + 545, + 594 + ], + "type": "interline_equation", + "content": "L _ {b o n d} ^ {l} = \\sum_ {i = 1} ^ {2 ^ {l}} \\left(S _ {i} ^ {l} - \\max \\left(S _ {2 i - 1} ^ {l + 1}, S _ {2 i} ^ {l + 1}\\right)\\right) ^ {2}, \\tag {3}", + "image_path": "e20ec544daa135585a261a4149b8c62b2e4f9650b405ee7b59656c0d09aaca04.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 600, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 600, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 600, + 545, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 600, + 545, + 647 + ], + "type": "inline_equation", + "content": "S_{i}^{l}" + }, + { + "bbox": [ + 304, + 600, + 545, + 647 + ], + "type": "text", + "content": " is the occupancy value of the parent part at position " + }, + { + "bbox": [ + 304, + 600, + 545, + 647 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 600, + 545, + 647 + ], + "type": "text", + "content": " and level " + }, + { + "bbox": [ + 304, + 600, + 545, + 647 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 600, + 545, + 647 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 600, + 545, + 647 + ], + "type": "inline_equation", + "content": "S_{2i-1}^{l+1}" + }, + { + "bbox": [ + 304, + 600, + 545, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 600, + 545, + 647 + ], + "type": "inline_equation", + "content": "S_{2i}^{l+1}" + }, + { + "bbox": [ + 304, + 600, + 545, + 647 + ], + "type": "text", + "content": " are the occupancy values of the two derived child parts. The total bond loss is computed by summing the bond loss across all levels when" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 647, + 453, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 453, + 663 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 453, + 663 + ], + "type": "interline_equation", + "content": "l < L \\colon L _ {b o n d} = \\lambda_ {b o n d} \\sum_ {l = 1} ^ {l < L} L _ {b o n d} ^ {l}.", + "image_path": "fadfa34a02c64f7c77a70a84c2a163cb0c9959fb503dc3acfbe0fb1140c96827.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "content": "Regularization Loss To reduce the complexity of the network and avoid overfitting, we introduce regularization loss " + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "L_{reg}" + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "content": " at the last layer of each Structure Divider at each level." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2718" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "text", + "content": "This loss is calculated as the sum of the absolute values of all weight parameters, multiplied by a small regularization parameter " + }, + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "inline_equation", + "content": "\\lambda_{reg}" + }, + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "text", + "content": ". The regularization loss is given by: " + }, + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "inline_equation", + "content": "L_{reg} = \\lambda_{reg}\\sum_{i=1}^{n}|w_i|" + }, + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "text", + "content": "-th weight parameter in the layer, and " + }, + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 72, + 289, + 169 + ], + "type": "text", + "content": " is the total number of weight parameters. Our experiments show that without using regularization loss, the generated shapes may contain parts that are over-segmented or lack reasonable shape." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 191, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 191, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 191, + 289, + 360 + ], + "type": "text", + "content": "Comparison with Existing Method Our method differs from previous approaches that utilize (hierarchical) decoders for unsupervised shape part learning (such as in [9, 29]) in several ways. Firstly, our hierarchical decoder aims to learn semantic shape part-level implicit field rather than simply segmenting the shape into more parts. For example, when generating a chair with four legs, our approach considers the legs as a whole unit rather than segmenting them individually, as we typically describe the legs as a whole rather than as individual parts. To balance the network's fitting ability with its complexity, we utilize a regularization loss. Secondly, we also construct hierarchical color latent fields, which are often overlooked in other methods, to represent colored shapes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 371, + 216, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 371, + 216, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 216, + 384 + ], + "type": "text", + "content": "3.3. Text-Guided Shape Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 391, + 289, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 391, + 289, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 391, + 289, + 548 + ], + "type": "text", + "content": "Fig. 2 (B) shows the text-guided shape generation network, comprising three modules: text encoder " + }, + { + "bbox": [ + 47, + 391, + 289, + 548 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 47, + 391, + 289, + 548 + ], + "type": "text", + "content": ", textual graph encoder " + }, + { + "bbox": [ + 47, + 391, + 289, + 548 + ], + "type": "inline_equation", + "content": "E_{G}" + }, + { + "bbox": [ + 47, + 391, + 289, + 548 + ], + "type": "text", + "content": ", and hierarchical shape decoder " + }, + { + "bbox": [ + 47, + 391, + 289, + 548 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 47, + 391, + 289, + 548 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 47, + 391, + 289, + 548 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 47, + 391, + 289, + 548 + ], + "type": "text", + "content": " module is similar to the pre-trained shape-specific decoder " + }, + { + "bbox": [ + 47, + 391, + 289, + 548 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 47, + 391, + 289, + 548 + ], + "type": "text", + "content": " (as described in Section 3.2), with the addition of a new part-node attention module and a refiner. This attention module allows the network to learn the local connections between shape parts/colors and textual graph entities, followed by a refiner module " + }, + { + "bbox": [ + 47, + 391, + 289, + 548 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 47, + 391, + 289, + 548 + ], + "type": "text", + "content": " for final results generation. " + }, + { + "bbox": [ + 47, + 391, + 289, + 548 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 47, + 391, + 289, + 548 + ], + "type": "text", + "content": " module is initialized from the pre-trained " + }, + { + "bbox": [ + 47, + 391, + 289, + 548 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 47, + 391, + 289, + 548 + ], + "type": "text", + "content": " to leverage its predefined structural implicit fields and facilitate the generation process in a hierarchical manner. We provide a detailed description of each part in the following sections." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 289, + 715 + ], + "type": "text", + "content": "Text Global Encoding We utilize a pre-trained BERT model [11] as the text encoder base. The encoded [CLS] token feature is used to extract global-level latent features " + }, + { + "bbox": [ + 47, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\bar{t}_s" + }, + { + "bbox": [ + 47, + 570, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\bar{t}_c" + }, + { + "bbox": [ + 47, + 570, + 289, + 715 + ], + "type": "text", + "content": " for shape and color, respectively, which serve as the input for the hierarchical shape decoder " + }, + { + "bbox": [ + 47, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 570, + 289, + 715 + ], + "type": "text", + "content": ". Most existing approaches employ word-level features " + }, + { + "bbox": [ + 47, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 47, + 570, + 289, + 715 + ], + "type": "text", + "content": " as fine-grained guidance in text-conditioned image or 3D shape generation tasks [40, 25]. However, we argue that the use of structured textual representation, such as the textual graph, would be more appropriate in aligning with shape structures. As a result, we opt to use the word features to initialize the graph node representation, as described below." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 545, + 180 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 545, + 180 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 545, + 180 + ], + "type": "image", + "image_path": "11013f26976ef3e04cbef40cc29a0dac99ed76c9b3bca4b5861b11a5ca457b6e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 185, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 304, + 185, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 547, + 251 + ], + "type": "text", + "content": "Figure 3. The dashed box shows a triplet " + }, + { + "bbox": [ + 304, + 185, + 547, + 251 + ], + "type": "inline_equation", + "content": "(v_{s}, v_{r}, v_{o})" + }, + { + "bbox": [ + 304, + 185, + 547, + 251 + ], + "type": "text", + "content": " in the graph, denoting the subjective, relation, and objective nodes. The three nodes are first concatenated and then jointly processed using MLP, which outputs a new value for each node. The final updated state " + }, + { + "bbox": [ + 304, + 185, + 547, + 251 + ], + "type": "inline_equation", + "content": "v_{s}^{\\prime}" + }, + { + "bbox": [ + 304, + 185, + 547, + 251 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 185, + 547, + 251 + ], + "type": "inline_equation", + "content": "v_{s}" + }, + { + "bbox": [ + 304, + 185, + 547, + 251 + ], + "type": "text", + "content": " is the result of pooling the corresponding values from all triplets in which the node " + }, + { + "bbox": [ + 304, + 185, + 547, + 251 + ], + "type": "inline_equation", + "content": "v_{s}" + }, + { + "bbox": [ + 304, + 185, + 547, + 251 + ], + "type": "text", + "content": " is involved." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 273, + 547, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 273, + 547, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 547, + 538 + ], + "type": "text", + "content": "Shape-Oriented Text Graph We use a unique text graph representation designed specifically for shapes. The graph nodes are classified into root, part, and attribute categories, which represent the class of the shape, the parts, and the attribute for root/part nodes, respectively. We use a special type node relation to connect these nodes, forming the edges. For example, the text \"brown wooden chair with curved armrests\" would be parsed as (chair: brown, wooden), (armrests: curved), and their relation triplet (chair, with, armrests) in the (subjective, relation, objective) convention. To construct graphs from the original text, we first use the spaCy library [15] to parse the text's semantic dependencies. These raw dependencies contain many linguistic terms and do not correspond well with real-world structures. To obtain more clear and shape-oriented graph formation, we introduce a custom parser targeting 3D shapes based on [36]. To ensure that the graph remains connected, we connect all part nodes with the root node using the with relation if no linguistic relation is found. The lower left of Fig. 2 illustrates these two forms of text. Finally, we use the resulting relation triplets for training. Further details can be found in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "Spatial-Aware Graph Encoding One challenge in using a shape-oriented graph, where entity or attribute nodes are connected through relation nodes, is the lack of a canonical grid for embedding the graph since nodes in such a graph exist in a non-Euclidean space. Existing methods based on Graph Convolution Networks (GCN) [21] typically rely on an adjacency matrix to represent the relations between nodes, which reduces their ability to represent spatial relations. To address the issue, we designed a novel Spatial-Aware Graph Encoder (SAGE) that differentiably encodes relation nodes in order to capture the structural relations between nodes. Specifically, we represent each graph triplet " + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": " in the form " + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "inline_equation", + "content": "(v_{s}, v_{r}, v_{o})" + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "inline_equation", + "content": "v_{s}" + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "inline_equation", + "content": "v_{o}" + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "inline_equation", + "content": "v_{r}" + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": " represent" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2719" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 71, + 278, + 186 + ], + "blocks": [ + { + "bbox": [ + 72, + 71, + 278, + 186 + ], + "lines": [ + { + "bbox": [ + 72, + 71, + 278, + 186 + ], + "spans": [ + { + "bbox": [ + 72, + 71, + 278, + 186 + ], + "type": "image", + "image_path": "82c3b4940e0d9e0d8cb72f13adbfd6a08763ad5fb401f2817363ec464ed30be2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 188, + 288, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 188, + 288, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 288, + 222 + ], + "type": "text", + "content": "Figure 4. The Part-Node attention. The attention map is obtained by the dot-product between the features from the intermediate part latent feature " + }, + { + "bbox": [ + 46, + 188, + 288, + 222 + ], + "type": "inline_equation", + "content": "(f_s^l)" + }, + { + "bbox": [ + 46, + 188, + 288, + 222 + ], + "type": "text", + "content": " at level " + }, + { + "bbox": [ + 46, + 188, + 288, + 222 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 188, + 288, + 222 + ], + "type": "text", + "content": " and the graph node features " + }, + { + "bbox": [ + 46, + 188, + 288, + 222 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 188, + 288, + 222 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 242, + 287, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 287, + 326 + ], + "type": "text", + "content": "the values of subjective, objective, and relation nodes in a relationship, respectively. This makes the graph directional through the order of the nodes between the relation node, like chair has arms. As shown in Fig. 3, the computational layer in SAGE jointly encodes the nodes and their relation in each triplet, followed by a pooling layer that fuses the features of each node involved in more than one triplet." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 342, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 437 + ], + "type": "text", + "content": "Part-Node Attention We introduce the Part-Node Attention modules for learning the relationship between intermediate parts/colors with textual graph nodes. These modules are based on the multi-head attention mechanism presented in [38]. For level " + }, + { + "bbox": [ + 46, + 342, + 287, + 437 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 342, + 287, + 437 + ], + "type": "text", + "content": ", given the intermediate shape latents " + }, + { + "bbox": [ + 46, + 342, + 287, + 437 + ], + "type": "inline_equation", + "content": "f_{s}^{l} \\in \\mathbb{R}^{2^{l} \\times d}" + }, + { + "bbox": [ + 46, + 342, + 287, + 437 + ], + "type": "text", + "content": " as query, and the encoded graph node features, " + }, + { + "bbox": [ + 46, + 342, + 287, + 437 + ], + "type": "inline_equation", + "content": "g \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 46, + 342, + 287, + 437 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 342, + 287, + 437 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 342, + 287, + 437 + ], + "type": "text", + "content": " is node number), as key and value, the calculation inside an attention head is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 449, + 287, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 449, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 113, + 449, + 287, + 475 + ], + "type": "interline_equation", + "content": "\\hat {f} _ {s} ^ {l} = \\operatorname {s o f t m a x} \\left(\\frac {Q K}{\\sqrt {d _ {h}}}\\right) V, \\tag {4}", + "image_path": "e711581c668e0d81d796aa8471cb04df6445bc00b3a93af8e65b95bc4c7b11e0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "inline_equation", + "content": "Q = F_{Q}(f_{s}^{l})" + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "inline_equation", + "content": "K = F_{K}(g)" + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "inline_equation", + "content": "V = F_{V}(g)" + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "inline_equation", + "content": "F_{Q}, F_{K}" + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "inline_equation", + "content": "F_{V}" + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "text", + "content": " are functions implemented using MLPs. " + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "text", + "content": " is the feature dimension for " + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "inline_equation", + "content": "f_{s}^{l}" + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "inline_equation", + "content": "d_{h}" + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "text", + "content": " represents the dimension used within an attention head. The SoftMax operation is applied across the graph nodes, resulting in a weighted aggregation of the graph node features as the new intermediate part latents, " + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "inline_equation", + "content": "\\hat{f}_{s}^{l}" + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "text", + "content": ". This process is visualized in Fig. 4. Also, attention is applied between the paired color features " + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "inline_equation", + "content": "f_{c}^{l}" + }, + { + "bbox": [ + 46, + 481, + 287, + 625 + ], + "type": "text", + "content": " with the graph nodes using a separate module. In practice, we use the attention module at level 2 of the hierarchical decoder due to the generated semantically detailed shape components aligned with graph nodes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "Joint Refiner The hierarchical bond loss, introduced in Section 3.2, serves as the foundation for structure-aware generation by constraining the composition of shape parts to follow a hierarchical tree structure. However, we posit that this also negatively impacts the generated results, as demonstrated in the perception-distortion tradeoff [5]. To" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "mitigate this issue, we introduce a refiner " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " in one level that uses the level latents jointly to generate the final shape. The refiner is also an implicit shape decoder with six layers of MLPs. The final shape output is accompanied by an additional refiner reconstruction loss " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "L_{R}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " but without the need to consider the hierarchical bond, which allows for the incorporation of additional detail in the final shape." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 178, + 545, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 178, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 178, + 545, + 370 + ], + "type": "text", + "content": "Training Strategy and Losses Training text-guided hierarchical shape generation directly presents significant challenges considering the big discrepancy between text and shape modalities. To address this, we propose a three-step strategy to ensure stable and effective training. First, we collect the shape and color latents from the shape encoder " + }, + { + "bbox": [ + 304, + 178, + 545, + 370 + ], + "type": "inline_equation", + "content": "E_{S}" + }, + { + "bbox": [ + 304, + 178, + 545, + 370 + ], + "type": "text", + "content": " as guidance and train the text global feature encoder " + }, + { + "bbox": [ + 304, + 178, + 545, + 370 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 304, + 178, + 545, + 370 + ], + "type": "text", + "content": " alone. The loss is using MSE: " + }, + { + "bbox": [ + 304, + 178, + 545, + 370 + ], + "type": "inline_equation", + "content": "L_{\\text{latent}} = \\lambda_{\\text{latent}}^{s} (\\bar{t}_{s} - \\bar{f}_{s})^{2} + \\lambda_{\\text{latent}}^{c} (\\bar{t}_{c} - \\bar{f}_{c})^{2}" + }, + { + "bbox": [ + 304, + 178, + 545, + 370 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 178, + 545, + 370 + ], + "type": "inline_equation", + "content": "\\bar{t}_{s}" + }, + { + "bbox": [ + 304, + 178, + 545, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 178, + 545, + 370 + ], + "type": "inline_equation", + "content": "\\bar{t}_{c}" + }, + { + "bbox": [ + 304, + 178, + 545, + 370 + ], + "type": "text", + "content": " are encoded by text encoder while " + }, + { + "bbox": [ + 304, + 178, + 545, + 370 + ], + "type": "inline_equation", + "content": "\\bar{f}_{s}" + }, + { + "bbox": [ + 304, + 178, + 545, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 178, + 545, + 370 + ], + "type": "inline_equation", + "content": "\\bar{f}_{c}" + }, + { + "bbox": [ + 304, + 178, + 545, + 370 + ], + "type": "text", + "content": " are encoded by shape encoder. Second, we train the decoder with attention modules, freezing the decoder during the first ten epochs to allow for warmed-up training. Finally, we train the entire network together. The total loss then includes the aforementioned shape/color reconstruction losses and the bond loss in shape-specific pre-training (see Sec. 3.2), and the refiner loss." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 384, + 387, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 387, + 398 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 387, + 398 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 406, + 545, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 406, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 545, + 454 + ], + "type": "text", + "content": "This section introduces our experimental design and implementation, followed by a comprehensive analysis of the results from various aspects and an ablation study. Supplementary material provides additional results and analysis." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 465, + 531, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 465, + 531, + 478 + ], + "spans": [ + { + "bbox": [ + 305, + 465, + 531, + 478 + ], + "type": "text", + "content": "4.1. Experimental Settings and Implementation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 484, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 568 + ], + "type": "text", + "content": "We conduct experiments using the text-shape dataset [8], which is derived from the 3D shape dataset, ShapeNet [7], and augmented with textual annotations. The dataset comprises approximately 15,000 shapes and 75,000 textual descriptions, encompassing the table and chair shape classes. We also adhere to the same training and testing splits as previously established." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "We implement our framework in PyTorch [32]. We set " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "N = 4096" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " as the sampled points. We pre-train the shape-specific decoder for 1000 epochs and the text-guided model for another 500 epochs. The learning rate is constant " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " using Adam [20] for all phases. The loss weights are set as follows: " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{recon}} = 10.0" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{color}} = 1.0" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{bond}} = 1.0" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{reg}} = 1e^{-4}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{latent}}^s = 10.0" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{latent}}^c = 1.0" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ". Detailed training strategy can be found in Sec. 3.3. Additionally, as our model outputs shapes at different levels, we select the results from level 2, as it is at this level that most shapes' parts have been parsed and refined. Level 3 is used as a spare level for monitoring no further parts are parsed." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2720" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 286, + 121 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 121 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 121 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 121 + ], + "type": "table", + "html": "
IoU (↑)IS (↑)EMD (↓)Acc (↑)
Text2Shape [8]9.641.960.444397.37
T2S-Implicit [25]12.211.970.207197.48
ours13.651.980.176797.5
", + "image_path": "a39f94ae94af3074bdd582afa710acfdb21f8f2c1123064e0bb4e9809bab8adc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 121, + 287, + 144 + ], + "lines": [ + { + "bbox": [ + 47, + 121, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 287, + 144 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with state-of-the-art. " + }, + { + "bbox": [ + 47, + 121, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 47, + 121, + 287, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 121, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 47, + 121, + 287, + 144 + ], + "type": "text", + "content": " indicate higher or lower is better, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 169, + 216, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 169, + 216, + 182 + ], + "spans": [ + { + "bbox": [ + 47, + 169, + 216, + 182 + ], + "type": "text", + "content": "4.2. Text-Guided Shape Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 191, + 289, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 191, + 289, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 191, + 289, + 502 + ], + "type": "text", + "content": "Comparison with the State-of-the-art We compare our method to two existing text-conditioned colored shape generation methods, Text2Shape [8] and T2S-Implicit [25]. We employ the same quantitative metrics for evaluation, including (1) Mean intersection-over-union (IoU) for shape occupancy measurement, regardless of color; (2) Earth Mover's Distance (EMD) for color distribution evaluation; (3) Inception Score (IS) to quantify the realism of the results; and (4) Accuracy (Acc) computed by a pre-trained shape classifier to indicate whether the predicted class matches the ground truth. To ensure fairness in comparison, we convert the generated shapes to volumes of the same size as in [8, 25] before calculating the quantitative values. The comparison results are presented in Table 1, where the numbers of the existing methods are directly taken from the respective papers. Our method exhibits superior performance across all metrics. The qualitative result presented in Fig. 5 provides further evidence of the effectiveness of our approach. As can be observed, our method achieves a higher degree of correspondence between the predicted shape structure and the text description, as demonstrated in Row 1 of the figure with the example of \"wood armrests\". Additionally, our approach exhibits an increased level of completeness in part generation, as evidenced by the absence of incomplete parts or extra artifacts, which are clearly labeled with red circles in the results generated by the existing methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 533, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 289, + 713 + ], + "type": "text", + "content": "Part-level Refinement As T2S-Implicit [25] also employs an implicit decoder similar to our approach, we conduct a comparison with it to highlight the distinctions, particularly with regard to part-level details. The results are presented in high resolution in Fig. 6. We observed that [25] tends to struggle with smaller or complex parts, such as the legs of a sofa or the wheels of an office chair. In contrast, our method is capable of interpreting these parts explicitly, resulting in more realistic part-level generation. Another advantage of our approach is that the generation of a single part does not affect the other parts, which is a common issue in previous methods. For example, generating \"silver legs\" may result in the bottom of the sofa appearing slightly silver in the connection area, creating a noticeable color artifact." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 305, + 71, + 545, + 266 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 266 + ], + "type": "image", + "image_path": "600bd8836ac32bd08e535769f6dbe9df18258f8108306b5fc709ea2e18933ab8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 269, + 547, + 302 + ], + "lines": [ + { + "bbox": [ + 305, + 269, + 547, + 302 + ], + "spans": [ + { + "bbox": [ + 305, + 269, + 547, + 302 + ], + "type": "text", + "content": "Figure 5. Visualization of the results of our method in comparison to Text2Shape [8] and T2S-Implicit [25]. Red circles indicate incomplete parts or extra artifacts." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 317, + 547, + 423 + ], + "blocks": [ + { + "bbox": [ + 306, + 317, + 547, + 423 + ], + "lines": [ + { + "bbox": [ + 306, + 317, + 547, + 423 + ], + "spans": [ + { + "bbox": [ + 306, + 317, + 547, + 423 + ], + "type": "image", + "image_path": "99197c7aa1e4264c34473f139217f16785164e8c1644bc0b8ba05c3dde6312e1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 426, + 511, + 437 + ], + "lines": [ + { + "bbox": [ + 340, + 426, + 511, + 437 + ], + "spans": [ + { + "bbox": [ + 340, + 426, + 511, + 437 + ], + "type": "text", + "content": "Figure 6. Comparison of part-level refinement." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 457, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 590 + ], + "type": "text", + "content": "Part-level Manipulation The ability to learn structure-based 3D shape representations and align them with textual descriptions enables precise and convenient manipulation of shapes. In the process of generating a shape from text, our method allows for manipulation at various levels, including modification of the part-level structure, attributes, or colors. The results of this manipulation are illustrated in Fig. 7. The changes made in the text appear to reflect the changes in part structure (a), part attribute (c), and part-related color refinement (b), suggesting a possible consistency between the constituent parts of the shape and the corresponding text." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 598, + 406, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 406, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 406, + 610 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "We conduct extensive ablation studies to validate the effectiveness of the proposed hierarchical structure-based representation, graph-part attention, and joint refiner. We also use Frechet Inception Distance (FID) and CLIP [34] to evaluate the quality of the generated shapes and the similarity between the generated shapes and their corresponding text descriptions, respectively. More detailed information on the metrics can be found in the supplementary material. The re" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2721" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 539, + 164 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 539, + 164 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 539, + 164 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 539, + 164 + ], + "type": "table", + "html": "
MethodLevel 1Level 2Level 3
IOU↑FID↓CLIP↑IOU↑FID↓CLIP↑IOU↑FID↓CLIP↑
HierD [base]11.2017.4745.4311.0717.5245.3610.8517.6043.92
FlatD10.3819.5540.72------
HierD + WordAttn (L2)11.1317.4244.8811.6616.4647.4310.7217.6946.10
HierD + GraphAttn (L2)11.2917.3645.0112.6316.3150.6412.4116.7050.01
HierD + GraphAttn (L1)11.5416.8652.7511.3017.0951.4011.2717.1250.56
HierD + GraphAttn (L2) + Refiner(L2) [Full]11.1917.3847.2713.6515.8154.8912.2216.3452.20
", + "image_path": "07a74e9bcb463d9fc3aade9582e90d56c1640aca27fcbe5497dc5d57e4caefe4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 165, + 546, + 189 + ], + "lines": [ + { + "bbox": [ + 47, + 165, + 546, + 189 + ], + "spans": [ + { + "bbox": [ + 47, + 165, + 546, + 189 + ], + "type": "text", + "content": "Table 2. Results of the ablation study on output levels. The base model, HierD, is the proposed hierarchical decoder using sentence-level guidance only. L1/L2 indicates component application at Level 1/Level 2. ↑/↓ denote better higher/ lower results, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 48, + 210, + 286, + 335 + ], + "blocks": [ + { + "bbox": [ + 48, + 210, + 286, + 335 + ], + "lines": [ + { + "bbox": [ + 48, + 210, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 286, + 335 + ], + "type": "image", + "image_path": "427d5cbf36a613a16552bc79389deeb980b69b58291ef22749439213715305f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 342, + 288, + 389 + ], + "lines": [ + { + "bbox": [ + 46, + 342, + 288, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 288, + 389 + ], + "type": "text", + "content": "Figure 7. Manipulation of text results in corresponding changes at the part level. The second row shows the predicted structure from Level 1 to 2. (a) illustrates the removal of a part, (b) shows a change in part color, and (c) depicts a change in the part attribute." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 407, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 504 + ], + "type": "text", + "content": "sults of our ablation studies are presented in Table 2. We use the Hierarchical Decoder with sentence-level guidance (HierD) as a base model, and then add various modules such as word-level attention (WordAttn), our proposed graph-part attention (GraphAttn), and the joint refiner to different levels of the hierarchy (L1/L2 for Level 1 and 2). We present the results for each level for an in-depth analysis of the impact of the hierarchy level on the performance of the model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 518, + 287, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 602 + ], + "type": "text", + "content": "Hierarchical Representation We compare HierD to a flattened decoder (FlatD) which directly outputs 8 parts at Level 1. Utilizing the same settings as HierD, FlatD demonstrates inferior performance. Our qualitative analysis reveals that FlatD is more prone to overfitting, resulting in the generation of fewer parts that are overly complex geometrically, instead of generating finer parts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "Graph-based Attention We have observed the clear advantages of utilizing graph-based attention over word-based attention. Specifically, when applied to Level 2, the use of GraphAttn results in superior performance across all metrics. Furthermore, the results from Level 3 also demonstrate improved performance. This can be attributed to the semantic alignment of the entity nodes in the graph with the intermediate parts. However, when applied to Level 1, the use of" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 207, + 546, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 546, + 268 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 546, + 268 + ], + "type": "text", + "content": "GraphAttn does not yield similar results as in Level 2. We speculate that this discrepancy may be due to the coarser nature of the parts at Level 1, which hinders the ability to establish effective connections between shape parts and textual graph entities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 285, + 546, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 546, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 546, + 417 + ], + "type": "text", + "content": "Joint Refiner The utilization of a refiner as the final shape output in our method is driven by the observation of the perception-distortion tradeoff [5]. As the level increases, we have noted a slight decline in performance as compared to previous levels. For instance, the IOU in Level 2 is consistently lower than that in Level 1. We posit that the hierarchical bond loss employed in our approach results in a stronger structural connection, but may impede the generation of detailed features on top of the structure. The incorporation of a refiner, on the other hand, has resulted in a notable improvement in all metrics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 429, + 524, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 429, + 524, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 429, + 524, + 443 + ], + "type": "text", + "content": "5. Conclusion, limitation, and future work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 449, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 581 + ], + "type": "text", + "content": "The ShapeScaffold is a method for generating shapes guided by text, with the understanding that both shape and text possess internal structures that can be aligned to improve the outcome. We represent shapes as hierarchical structures and model text as scene graphs. Our part-node attention mechanism allows us to learn correspondences between shape and text at various hierarchical levels, which has been shown to effectively establish connections between shape parts/colors and linguistic entities. The experiments have demonstrated the superiority of our approach in both shape generation and text-shape/color consistency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": "Several limitations should be acknowledged. Firstly, the current text-to-shape dataset is limited to chairs and tables, so further research is needed to explore the generalization of our method to other shape classes. Secondly, the results may be influenced by the parts priority problem, where incorrect occupancy state of parts junctions may result in unreasonable visualization. Therefore, further incorporation of prior interpretations of shapes should be considered in future studies. Lastly, it would be beneficial to explore scene generation that takes structure into account, with a focus on the composition of different shapes as an outer structure." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2722" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 120 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by RCUK grant CAMERA (EP/M023281/1, EP/T022523/1), Centre for Augmented Reasoning (CAR) at the Australian Institute for Machine Learning, and a gift from Adobe." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 105, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 105, + 141 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 105, + 141 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 149, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 149, + 287, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 149, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 149, + 287, + 204 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Judy Fan, Robert Hawkins, Noah Goodman, and Leonidas J Guibas. Shapeplot: Learning language for shape differentiation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8938-8947, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 248 + ], + "type": "text", + "content": "[2] Oron Ashual and Lior Wolf. Specifying object attributes and relations in interactive scene generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4561-4569, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 249, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 249, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 249, + 288, + 270 + ], + "type": "text", + "content": "[3] Renée Baillargeon. Infants' physical world. Current directions in psychological science, 13(3):89-94, 2004. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 271, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 271, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 271, + 288, + 304 + ], + "type": "text", + "content": "[4] Irving Biederman. Human image understanding: Recent research and a theory. Computer vision, graphics, and image processing, 32(1):29-73, 1985. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 304, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 288, + 337 + ], + "type": "text", + "content": "[5] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 338, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 338, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 338, + 288, + 392 + ], + "type": "text", + "content": "[6] Angel Chang, Manolis Savva, and Christopher D Manning. Learning spatial knowledge for text to 3d scene generation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 2028-2038, 2014. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 393, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 393, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 393, + 288, + 446 + ], + "type": "text", + "content": "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 449, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 449, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 449, + 288, + 502 + ], + "type": "text", + "content": "[8] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian conference on computer vision, pages 100-116. Springer, 2018. 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 504, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 504, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 504, + 288, + 557 + ], + "type": "text", + "content": "[9] Zhiqin Chen, Kangxue Yin, Matthew Fisher, Siddhartha Chaudhuri, and Hao Zhang. Bae-net: Branched autoencoder for shape co-segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8490-8499, 2019. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "type": "text", + "content": "[10] Morten H Christiansen and Nick Chater. Language as shaped by the brain. Behavioral and brain sciences, 31(5):489-509, 2008. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "text", + "content": "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 691 + ], + "type": "text", + "content": "[12] Helisa Dhamo, Azade Farshad, Iro Laina, Nassir Navab, Gregory D Hager, Federico Tombari, and Christian Rupprecht. Semantic image manipulation using scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5213-5222, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[13] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecraft: A recursive text-conditioned 3d shape" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "generation model. arXiv preprint arXiv:2207.09446, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 162 + ], + "type": "text", + "content": "[14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proceedings of the 27th International Conference on Neural Information Processing Systems - Volume 2, NIPS'14, page 2672-2680, 2014. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "type": "text", + "content": "[15] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. spaCy: Industrial-strength Natural Language Processing in Python, 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 198, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 239 + ], + "type": "text", + "content": "[16] Fangqiao Hu, Jin Zhao, Yong Huang, and Hui Li. Learning structural graph layouts and 3d shapes for long span bridges 3d reconstruction. arXiv preprint arXiv:1907.03387, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 243, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 297 + ], + "type": "text", + "content": "[17] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 298, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 353 + ], + "type": "text", + "content": "[18] R. Kenny Jones, Aalia Habib, Rana Hanocka, and Daniel Ritchie. The neurally-guided shape parser: Grammar-based labeling of 3d shape regions with approximate inference. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 354, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 387 + ], + "type": "text", + "content": "[19] R Kenny Jones, Aalia Habib, and Daniel Ritchie. Shred: 3d shape region decomposition with learned local operations. arXiv preprint arXiv:2206.03480, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 388, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 443 + ], + "type": "text", + "content": "[20] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 445, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 476 + ], + "type": "text", + "content": "[21] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "type": "text", + "content": "[22] Jun Li, Kai Xu, Siddhartha Chaudhuri, Ersin Yumer, Hao Zhang, and Leonidas Guibas. Grass: Generative recursive autoencoders for shape structures. ACM Transactions on Graphics (TOG), 36(4):1-14, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "type": "text", + "content": "[23] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 579, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 634 + ], + "type": "text", + "content": "[24] Chunxiao Liu, Zhendong Mao, Tianzhu Zhang, Hongtao Xie, Bin Wang, and Yongdong Zhang. Graph structured network for image-text matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10921-10930, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "text", + "content": "[25] Zhengzhe Liu, Yi Wang, Xiaojuan Qi, and Chi-Wing Fu. Towards implicit text-guided 3d shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17896-17906, 2022. 2, 5, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[26] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2723" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "type": "text", + "content": "[27] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 306-315, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "type": "text", + "content": "[28] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 253 + ], + "type": "text", + "content": "[29] Chengjie Niu, Manyi Li, Kai Xu, and Hao Zhang. Rim-net: Recursive implicit fields for unsupervised learning of hierarchical shape structures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11779-11788, 2022. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 255, + 288, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 288, + 309 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 288, + 309 + ], + "type": "text", + "content": "[30] Despoina Paschalidou, Luc Van Gool, and Andreas Geiger. Learning unsupervised hierarchical part decomposition of 3d objects from a single rgb image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1060-1070, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 311, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 311, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 311, + 287, + 365 + ], + "type": "text", + "content": "[31] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics revisited: Learning 3d shape parsing beyond cuboids. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10344–10353, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 367, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 367, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 367, + 287, + 487 + ], + "type": "text", + "content": "[32] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 489, + 287, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 522 + ], + "type": "text", + "content": "[33] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 523, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 590 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 591, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 656 + ], + "type": "text", + "content": "[35] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[36] Richard Socher, Cliff C Lin, Chris Manning, and Andrew Y Ng. Parsing natural scenes and natural language with recursive neural networks. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 129–136, 2011. 5" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 295 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "text", + "content": "[37] Shubham Tulsiani, Hao Su, Leonidas J Guibas, Alexei A Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2635-2643, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 129, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 173 + ], + "type": "text", + "content": "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 175, + 546, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 546, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 546, + 228 + ], + "type": "text", + "content": "[39] Yanzhen Wang, Kai Xu, Jun Li, Hao Zhang, Ariel Shamir, Ligang Liu, Zhiquan Cheng, and Yueshan Xiong. Symmetry hierarchy of man-made objects. In Computer graphics forum, volume 30, pages 287-296. Wiley Online Library, 2011. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 229, + 546, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 229, + 546, + 295 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 546, + 295 + ], + "type": "text", + "content": "[40] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 3, 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2724" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Shatter and Gather_ Learning Referring Image Segmentation with Text Supervision/eba869e5-3a94-47e8-afcc-2f1d3dc46bfb_content_list.json b/2023/Shatter and Gather_ Learning Referring Image Segmentation with Text Supervision/eba869e5-3a94-47e8-afcc-2f1d3dc46bfb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ca6760223960107fa3920919cb8fb510ee36a5c8 --- /dev/null +++ b/2023/Shatter and Gather_ Learning Referring Image Segmentation with Text Supervision/eba869e5-3a94-47e8-afcc-2f1d3dc46bfb_content_list.json @@ -0,0 +1,1735 @@ +[ + { + "type": "text", + "text": "Shatter and Gather: Learning Referring Image Segmentation with Text Supervision", + "text_level": 1, + "bbox": [ + 163, + 130, + 799, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dongwon Kim $^{1*}$ Namyup Kim $^{1*}$ Cuiling Lan $^{2}$ Suha Kwak $^{1}$ $^{1}$ POSTECH Microsoft Research Asia \nhttp://cvlab.postech.ac.kr/research/sag", + "bbox": [ + 210, + 198, + 756, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 287, + 313, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring image segmentation, the task of segmenting any arbitrary entities described in free-form texts, opens up a variety of vision applications. However, manual labeling of training data for this task is prohibitively costly, leading to lack of labeled data for training. We address this issue by a weakly supervised learning approach using text descriptions of training images as the only source of supervision. To this end, we first present a new model that discovers semantic entities in input image and then combines such entities relevant to text query to predict the mask of the referent. We also present a new loss function that allows the model to be trained without any further supervision. Our method was evaluated on four public benchmarks for referring image segmentation, where it clearly outperformed the existing method for the same task and recent open-vocabulary segmentation models on all the benchmarks.", + "bbox": [ + 75, + 319, + 473, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 584, + 209, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring image segmentation is the task of segmenting the referent corresponding to a natural language expression given as a query. Unlike the conventional semantic segmentation that aims at segmenting a pre-defined set of classes, referring image segmentation enables segmentation of any arbitrary entities described in free-form texts and thus opens up a wide variety of applications such as human-computer interaction [8, 57] and text-based image editing [7, 17].", + "bbox": [ + 75, + 604, + 468, + 726 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thanks to the development of deep neural networks, recent studies have demonstrated remarkable results on referring image segmentation in the supervised learning setting [14, 22, 25, 27, 67, 68, 69]. However, obtaining manual annotation of training data for referring image segmentation is prohibitively costly since the data requires two types of labels for each image, i.e., natural language expressions describing the entities that appear in the image and segmentation labels corresponding to them. This typically demands tremendously more manpower than annotation for semantic", + "bbox": [ + 75, + 726, + 470, + 877 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dd9ca38fcbe60fa581656fa48f2190717303716cf506d35dc6a41acacf2ed70b.jpg", + "image_caption": [ + "Figure 1: Our model consists of two different attention modules: bottom-up attention that identifies individual entities in an image, and top-down attention that combines the entities based on a referring expression. It is trained in a contrastive manner to ensure consistency between matching image-text pairs. In inference, a segmentation mask is predicted by combining the entities found by bottom-up attention with weights derived by top-down attention." + ], + "image_footnote": [], + "bbox": [ + 503, + 286, + 890, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "segmentation, leading to lack of labeled data for training.", + "bbox": [ + 498, + 703, + 877, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A solution to this issue is weakly supervised learning, which trains a model on a dataset providing weaker forms of supervision than the conventional ones. In this paper, we particularly focus on learning referring image segmentation only with natural language expressions describing the associated images. Since these labels are readily available in many vision-language datasets [28, 47], the issue on lack of labeled training data can be alleviated by our approach. However, learning referring image segmentation in this setting introduces another challenge: Since natural language expressions often contain relations between visual entities such as instances and object parts (e.g., \"The cat next to the", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 101, + 887, + 212, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15547", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "table leg\"), a model to be trained should be aware of the inter-object relations to exploit the expressions as supervision. Thereby we need a method to discover individual entities and infer their relations without any supervision but the referring expression.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To fulfill this requirement, we introduce a new bottom-up and top-down attention framework, which is illustrated in Fig. 1. Individual entities existing in an image are first discovered by bottom-up attention, which solely exploits visual information. For the bottom-up attention, an entity discovery module progressively refines a set of embedding vectors named slots to capture distinct visual entities, following slot attention [42]. We moreover propose a novel slot formulation named entity slot, which enables fine-grained entity discovery in real-world images. As a result, entity slots after the refinement capture individual visual entities without any segmentation supervision. These entities are considered primitive units for composing the segmentation mask of a referent during inference.", + "bbox": [ + 75, + 167, + 470, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Afterward, based on the relational and structural information of referring expression, top-down attention selectively attends to the referred entities and combines them into the predicted mask. Top-down attention is implemented by a modality fusion module consists of cross-attention transformer [15]. This attention scheme enables the module to infer the relevance between discovered entities and the referring expression instead of relying on the cosine similarity or heuristic similarity functions used in previous work [18, 53, 64, 73].", + "bbox": [ + 75, + 381, + 468, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For training, we propose the contrastive cycle-consistency $(\\mathrm{C}^3)$ loss that enforces cycle-consistency between image-text pairs under the contrastive learning framework [47]. To establish such consistency, the model is trained to preserve the textual information after the fusion of entities and textual features. By doing so, we observe that latent relevance between the discovered entities and referring expression automatically emerges in the top-down attention without any explicit mask supervision.", + "bbox": [ + 75, + 534, + 468, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One may wonder the difference between our task and open-vocabulary segmentation [18, 64, 73], i.e., semantic segmentation of arbitrary categories without explicit mask supervision. The key difference is that a referent to segment in referring image segmentation is given by a complex free-form text, which introduces additional challenges.", + "bbox": [ + 75, + 671, + 468, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed method was evaluated and compared on four public benchmarks for referring image segmentation, where it substantially outperformed the previous weakly supervised learning method [53]. Moreover, we reproduced recent open-vocabulary segmentation models [64, 73] and evaluated them on referring image segmentation benchmarks; the results show that the proposed model surpassed them even without pre-training on large-scale image-text data [5, 47, 54].", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contribution is three-fold as follows:", + "bbox": [ + 517, + 90, + 872, + 107 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a new weakly supervised learning model for referring image segmentation. Our model employs bottom-up and top-down attention to discover individual entities and infer their relations, which enables to exploit natural language expressions as supervision effectively during training.", + "- We propose a new loss function called contrastive cycle-consistency loss, which allows latent relevance between discovered entities and referring expressions to emerge without requiring further supervision.", + "- Our model clearly outperformed the existing method for the same task and recent open-vocabulary segmentation models on all the four benchmarks." + ], + "bbox": [ + 517, + 109, + 890, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 333, + 640, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Referring image segmentation: Early approaches [22, 38, 40, 51] obtain multimodal feature maps by simple concatenation-convolution operation. With this operation as a basis, most existing studies have suggested attention mechanisms and multi-level feature aggregations [6, 16, 23, 25, 46, 68]. Various approaches have been employed for calculating fine interactions between the modalities, including multimodal feature fusion [14, 29, 33] using transformers [15], a large-scale vision-language pretrained framework [58], and holistically multimodal fusion [67]. The existing referring image segmentation methods have been developed in a fully supervised learning setting. Recently, a weakly supervised learning method [53] using only imagedtext pairs for training is introduced to alleviate the high annotation cost. However, this study relies on a heuristic similarity function to infer relations between patches and texts, which suffers from comprehending the relational and structural information from image and text modalities. Therefore, we propose a new weakly supervised model consisting of bottom-up and top-down attention to discover visual entities and infer their relations effectively.", + "bbox": [ + 496, + 354, + 890, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Phrase grounding: Phrase grounding [12, 24, 29, 49, 66, 69] is the task of learning the association between image regions and words in caption describing the image. Weakly supervised approaches have also been proposed for this task [10, 20, 41, 56, 63], where the correspondence between words and regions are not given. It is worth noting that the contrastive learning framework introduced in [20] is closely related to the proposed $C^3$ loss by enforcing mutual information between regions and words. However, these studies rely on an off-the-shelf object detector [19], where the final prediction is restricted by the predefined set of classes with which the detector is trained. On the contrary, our method discovers visual entities in an image without an auxiliary model trained with external labeled data, which is more suitable for referring image segmentation.", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15548", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/47b618ab2f05409361cba42c53cac8d18a2060131b5c6f2b3448f65025736f6f.jpg", + "image_caption": [ + "Figure 2: Illustration of the overall architecture of our model along with its behavior during training and inference. (a) Feature extractors: The visual and textual features are extracted by transformer-based encoders of the two modalities respectively. (b) Entity discovery module: A set of visual entities are discovered from the visual features through the bottom-up attention. (c) Modality fusion module: The visual entities and a referring expression are fused considering their relevance, which is estimated by the top-down attention. Our model is trained with the proposed contrastive cycle-consistency $(\\mathbf{C}^3)$ loss between the cross-modal embeddings and textual features. In inference, the segmentation mask is predicted by jointly considering the visual entities $A^{\\mathrm{slot}}$ from the entity discovery module and the relevance scores $A^{\\mathrm{fuse}}$ from the modality fusion module." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 890, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Weakly supervised learning for segmentation: Weakly supervised learning has become a favored solution for addressing the lack of annotated data in training, especially for semantic segmentation [1, 2, 21, 26, 34, 36, 65, 71, 76] and instance segmentation [30, 31, 48, 75]. These methods refine coarse localization cues obtained by CAMs [50, 59, 72], and use the results as pseudo labels for self-training. However, they cannot be directly applied to the referring image segmentation due to the following challenges. First, there are no predefined classes in this task, making it difficult to learn shared semantics within CAMs by multiple instance learning. Second, there is no mechanism to discover visual entities and combine them with textual information.", + "bbox": [ + 75, + 438, + 472, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Slot attention: Slot attention [42] is proposed for the learning object-centric representation. It can discover the objects within an image in an unsupervised manner and be used in the various downstream task such as segmentation [64, 74] and retrieval [32, 60]. In specific, slot attention introduces an iterative attention mechanism that operates on the slots, which are the set of embedding vectors corresponding to the objects that appear in the input data. In previous work, slots are initialized as a set of learnable embeddings [32, 60, 64, 74] or a set of vectors sampled from a learnable Gaussian distribution [42, 52]. However, we empirically observed that both types of initialization are inapplicable to visual entity discovery for referring image segmentation, which we will elaborate on in Sec. 3.2.", + "bbox": [ + 75, + 638, + 472, + 851 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Open-vocabulary segmentation: The open-vocabulary segmentation is proposed to address the limitation of semantic segmentation that it is constrained to the pre", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "defined set of classes. Unlike semantic segmentation, open-vocabulary segmentation models are evaluated with classes unseen during training. The early approaches propose a generative model [4] and a projection model [62] to generate or project pixel-level semantic features from visual features for the unseen classes through word embedding space. Recent studies [18, 37, 64, 73] largely improve the quality of semantic features for the unseen classes, and further introduce the methods that do not require segmentation labels for seen classes using large-scale image-text pretrained models (e.g.CLIP [47], ALIGN [28]). However, since they are trained to predict masks for semantic categories, not for free-form texts describing arbitrary entities in an image, their applicability is still limited. We thus propose a new method for referring image segmentation that only requires image-text pairs during training.", + "bbox": [ + 496, + 438, + 893, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 500, + 696, + 671, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, our model consists of three main components: feature extractors, entity discovery module, and modality fusion module. The visual and textual features are extracted from the input using the backbone networks of the two modalities respectively. (Sec. 3.1). The visual features are then given as input to the entity discovery module, which aggregates the features into multiple entity embeddings using the bottom-up attention (Sec. 3.2). The entity embeddings correspond to distinct entities appearing in the image, which are considered as the smallest units of building the segmentation mask. The modality fusion module fuses the entity embeddings and the textual feature into", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15549", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cross-modal embeddings according to their relevance using the top-down attention (Sec. 3.3). The produced cross-modal embedding is used only for training, with a learning objective enforcing cycle-consistency and reconstruction of visual features (Sec. 3.4). The segmentation mask is predicted by combining the entities discovered by the entity discovery module with the relevance scores estimated by the modality fusion module (Sec. 3.5).", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Visual and Textual Feature Extraction", + "text_level": 1, + "bbox": [ + 76, + 223, + 406, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, we provide the details of the transformer-based backbone network for each modality.", + "bbox": [ + 75, + 243, + 468, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual feature extraction: We employ the vision transformer (ViT) [15] as the visual backbone network. An input image is first split into $N$ patches and then linearly projected. The patches are processed by the multiple self-attention blocks of ViT, where we refer to the output tokens of the last block as the visual features $\\mathbf{x}^{\\mathcal{V}}\\in \\mathbb{R}^{N\\times D}$ .", + "bbox": [ + 75, + 275, + 468, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Textual feature extraction: We consider the transformer-based text encoder, BERT [13]. A sentence is first tokenized into the token sequence, where the [CLS] and [SEP] tokens are added to the beginning and the end of the sequence. The activation of the last layer of the transformer for the [CLS] token is employed as the textual feature $\\mathbf{x}^{\\mathcal{T}}\\in \\mathbb{R}^{D}$ .", + "bbox": [ + 75, + 368, + 468, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Bottom-up Attention: Entity Discovery Module", + "text_level": 1, + "bbox": [ + 76, + 469, + 468, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The entity discovery module applies a bottom-up attention mechanism to aggregate patch-based visual features into multiple entity embeddings that represent instances or object parts in the image. The discovered entities are then treated as primitive units of the segmentation mask, where the mask is predicted by selectively combining them. This approach is useful since the textual information is often built upon the relations between such visual entities.", + "bbox": [ + 75, + 489, + 468, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the entity discovery module consists of two blocks: aggregation block and interaction block. The entity slots, which are embedding vectors that corresponds to various visual entities, are updated in an iterative process alternating between these two blocks. These slots eventually become entity embeddings. Let $\\Phi : \\mathbb{R}^{N \\times D} \\mapsto \\mathbb{R}^{K \\times D}$ be the entity discovery module that takes visual features as input and outputs the set of $K$ entity embeddings. The entity embeddings are obtained by iteratively updating initial entity slots $\\mathbf{S}^0 \\in \\mathbb{R}^{K \\times D}$ $T$ -times:", + "bbox": [ + 75, + 611, + 468, + 762 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{r l} & \\Phi (\\mathbf {x} ^ {\\mathcal {V}}) := \\mathbf {S} ^ {T}, \\text {w h e r e} \\\\ & \\quad \\mathbf {S} ^ {t + 1} = \\operatorname {I n t e r} \\left(\\operatorname {A g g} \\left(\\mathbf {x} ^ {\\mathcal {V}}; \\mathbf {S} ^ {t}\\right)\\right), \\end{array} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 773, + 468, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and Agg and Inter denote the aggregation and interaction blocks, respectively.", + "bbox": [ + 75, + 824, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the literature, the initial slots $\\mathbf{S}^0$ are often defined as either vectors sampled from a learnable Gaussian distribution [42, 52] or learnable embeddings [32, 60, 64, 74],", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/656824a77cfeba5f159c67838a6b7b9b2b9eb64851781382dec522c4f14f99e1.jpg", + "image_caption": [ + "Figure 3: Comparison between the proposed entity slot and previous slot initialization methods. The boundary color indicates the Gaussian distribution that each slot is sampled from. The entity slot successfully discovers fine-grained entities such as head and food, while the others often fail to capture individual entities of specific semantics." + ], + "image_footnote": [], + "bbox": [ + 514, + 89, + 877, + 361 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where such types of slots are referred as random slot and query slot in the remainder of this paper, respectively. However, we have empirically found that both of these initialization strategies are unsuitable for our task as shown in Fig. 3. The random slot suffers from the lack of semantic specificity since every slot is randomly sampled from a single distribution. This can lead to suboptimal performance since entities in real-world images are diverse in terms of semantics. The query slot, on the other hand, suffers from the limitation of being bound to specific semantic categories, such as human, animal, and vehicle. This limitation leads to entity discovery that is not sufficiently fine-grained, e.g., given an image of multiple people, a single slot that binds to \"human\" captures all of them and thus failed to identify the individual specified by the text query. This is problematic particularly in our model since entities act as the smallest units for producing segmentation masks. It is highly desired to have a mechanism that could guide the differentiation of slots to be semantically specific.", + "bbox": [ + 496, + 470, + 892, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To overcome these limitations, we introduce a new initialization strategy for the entity slot, which enables the discovery of fine-grained visual entities while learning diverse semantics. Formally, the initial entity slots $\\mathbf{S}^0$ are obtained by sampling $K_{s}$ slots from each of $K_{g}$ different Gaussian distributions, resulting in a total of $K = K_{s}\\cdot K_{g}$ slots:", + "bbox": [ + 496, + 758, + 890, + 849 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} ^ {0} := \\left[ \\mathbf {S} _ {1} ^ {0}; \\dots ; \\mathbf {S} _ {K _ {g}} ^ {0} \\right] \\in \\mathbb {R} ^ {K \\times D}, \\text {w h e r e} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 856, + 890, + 883 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {i} ^ {0} \\sim \\mathcal {N} (\\mu_ {i}, \\operatorname {d i a g} (\\sigma_ {i})) \\in \\mathbb {R} ^ {K _ {s} \\times D}.\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 878, + 792, + 897 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15550", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The idea is to initialize slots by sampling them from different Gaussian distributions, each specifying a distinct semantic property to be captured. Individual entities sharing the same semantics are differentiated by slots sampled from the same distribution, which enables fine-grained entity discovery.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From now, we present the aggregation block and interaction block at $t$ -th iteration. As described in Eq. (1), the slots of previous iteration $\\mathbf{S}^{t-1}$ are fed to the aggregation block with the visual features $\\mathbf{x}^{\\nu}$ . The role of the aggregation block is to let each slot aggregate visual features of distinct visual entities, and it is achieved by employing the competitive attention mechanism introduced in the slot attention [42]. Let $q(\\cdot)$ , $k(\\cdot)$ , and $v(\\cdot)$ are linear projection with layer normalization [3] that map inputs to dimension $D_h$ . Then, the attention map $A^{\\mathrm{slot}} \\in \\mathbb{R}^{N \\times K}$ between $\\mathbf{x}^{\\nu}$ and $\\mathbf{S}^{t-1}$ is computed as follows:", + "bbox": [ + 75, + 181, + 470, + 348 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {i, j} ^ {\\text {s l o t}} = \\frac {e ^ {M _ {i , j}}}{\\sum_ {j = 1} ^ {K} e ^ {M _ {i , j}}}, \\text {w h e r e} M = \\frac {k \\left(\\mathbf {x} ^ {\\mathcal {V}}\\right) q \\left(\\mathbf {S} ^ {t - 1}\\right) ^ {\\top}}{\\sqrt {D _ {h}}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 357, + 468, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike the cross-attention transformer [55] that normalized attention over the keys, we normalize attention map $A^{\\mathrm{slot}}$ over the slots. This normalization scheme encourages slots to compete for aggregating the visual features, which results in the set of slots binding to different visual entities in the input data. It is worth noting that $A^{\\mathrm{slot}}$ provides the localizations of entities discovered, which is used as primitive units of the segmentation mask during inference. Using $A^{\\mathrm{slot}}$ , the slots are updated in a similar manner to the transformer:", + "bbox": [ + 75, + 409, + 468, + 545 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\widehat {\\mathbf {S}} ^ {t} = \\operatorname {A g g} \\left(\\mathbf {x} ^ {\\nu}; \\mathbf {S} ^ {t}\\right) \\tag {4} \\\\ = \\operatorname {n o r m a l i z e} \\left(A ^ {\\text {s l o t}}\\right) ^ {\\top} v \\left(\\mathbf {x} ^ {\\nu}\\right) W _ {o} + \\mathbf {S} ^ {t - 1}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 566, + 468, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where normalize $(\\cdot)$ denotes $\\ell_1$ -normalization of the column vectors and $W_{o}\\in \\mathbb{R}^{D_{h}\\times D}$ is linear projection. Finally, slots are further refined by $\\widehat{\\mathbf{S}}^t\\gets \\mathrm{MLP}\\big(\\widehat{\\mathbf{S}}^t\\big) + \\widehat{\\mathbf{S}}^t$ where MLP is multi-layer perceptron (MLP) with layer normalization.", + "bbox": [ + 75, + 618, + 468, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The following interaction block consists of multiple self-attention transformers [55]. The weights of these transformers are shared, and each transformer is applied to slots sampled from the same distribution. We denote the transformer layer as Transformer(Q; K; V), where (Q; K; V) are for query, key, and value. The interaction block is denoted as follows:", + "bbox": [ + 75, + 679, + 468, + 784 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {I n t e r} \\left(\\widehat {\\mathbf {S}} ^ {t}\\right) = \\left[ \\operatorname {S A} \\left(\\widehat {\\mathbf {S}} _ {1} ^ {t}\\right); \\dots ; \\operatorname {S A} \\left(\\widehat {\\mathbf {S}} _ {K _ {g}} ^ {t}\\right) \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 794, + 468, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{SA}(\\mathbf{X})$ is for the Transformer(X; X; X). This self-attention scheme models information exchange between visual entities sharing the same semantic property, which helps slots within a single semantic category to capture different entities without redundancy.", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Top-down Attention: Modality Fusion Module", + "text_level": 1, + "bbox": [ + 500, + 90, + 890, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The modality fusion module is proposed for top-down attention mechanism, which combines discovered visual entities based on the textual feature. Let $\\Psi: \\mathbb{R}^{K \\times D} \\times \\mathbb{R}^D \\mapsto \\mathbb{R}^D$ be the modality fusion module, which takes entity embeddings $\\Phi(\\mathbf{x}^\\nu)$ and textual furret $\\mathbf{x}^T$ as inputs. Then, the output cross-modal embedding $\\mathbf{z} \\in \\mathbb{R}^D$ is computed using cross-attention transformer:", + "bbox": [ + 498, + 109, + 890, + 215 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} = \\Psi \\left(\\Phi \\left(\\mathbf {x} ^ {\\mathcal {V}}\\right); \\mathbf {x} ^ {\\mathcal {T}}\\right) = \\frac {\\operatorname {C A} \\left(\\Phi \\left(\\mathbf {x} ^ {\\mathcal {V}}\\right) ; \\mathbf {x} ^ {\\mathcal {T}}\\right)}{\\| \\operatorname {C A} \\left(\\Phi \\left(\\mathbf {x} ^ {\\mathcal {V}}\\right) ; \\mathbf {x} ^ {\\mathcal {T}}\\right) \\| _ {2}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 224, + 890, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{CA}(\\mathrm{X};\\mathrm{Y})$ is for the Transformer(Y;X;X) and $\\| \\cdot \\| _2$ denotes $\\ell_2$ norm. Moreover, we can obtain the attention map $A^{\\mathrm{fuse}}\\in \\mathbb{R}^{K}$ as a byproduct in cross-attention:", + "bbox": [ + 498, + 273, + 890, + 319 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {i} ^ {\\text {f u s e}} = \\frac {e ^ {M _ {i}}}{\\sum_ {i = 1} ^ {K} e ^ {M _ {i}}} \\text {, w h e r e} M = \\frac {k \\left(\\Phi \\left(\\mathbf {x} ^ {\\mathcal {V}}\\right)\\right) q \\left(\\mathbf {x} ^ {\\mathcal {T}}\\right)}{\\sqrt {D _ {h}}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 329, + 890, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using cross-attention, the module effectively reasons the relevance between entities and the textual feature. Inferred relevance is described in the $A^{\\mathrm{fuse}}$ , which is the probability of each entity being referred to in the textual feature.", + "bbox": [ + 498, + 378, + 890, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training", + "text_level": 1, + "bbox": [ + 500, + 449, + 604, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our model is trained with two different loss functions: contrastive cycle-consistency $(\\mathbf{C}^3)$ and reconstruction loss.", + "bbox": [ + 498, + 469, + 888, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.1 Contrastive cycle-consistency $(\\mathbf{C}^3)$ Loss", + "text_level": 1, + "bbox": [ + 498, + 512, + 828, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The output of our model is the cross-modal embedding, which is obtained by fusing information from visual and textual modalities. The fusion should selectively combine discovered entities based on textual information, but a segmentation label for which entities are referred to is not given in our weakly-supervised setting. To resolve this issue, we add an additional constraint to the contrastive learning objective: establishing cycle-consistency between each ground-truth pair of image and text.", + "bbox": [ + 496, + 529, + 890, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $\\{(\\mathbf{x}_i^\\nu, \\mathbf{x}_i^\\mathcal{T})\\}_{i=1}^B$ represent a batch of visual and textual features, and $\\mathbf{z}_{ij} = \\Psi \\big(\\Phi(\\mathbf{x}_i^\\nu); \\mathbf{x}_j^\\mathcal{T}\\big)$ denote the cross-modal embedding produced with $\\mathbf{x}_i^\\nu$ and $\\mathbf{x}_j^\\mathcal{T}$ . For simplicity, we refer to the cross-modal embedding produced using the ground-truth image-text pair as matching embedding, and all other cross-modal embeddings as non-matching embeddings. Then $\\mathbf{C}^3$ loss is computed as follows:", + "bbox": [ + 496, + 665, + 890, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C} ^ {3}} = - \\frac {1}{B} \\sum_ {j = 1} ^ {B} \\log \\frac {e ^ {\\langle \\mathbf {z} _ {j j} , \\mathbf {s g} (\\mathbf {x} _ {j} ^ {\\mathcal {T}}) \\rangle}}{\\sum_ {i = 1} ^ {B} e ^ {\\langle \\mathbf {z} _ {i j} , \\mathbf {s g} (\\mathbf {x} _ {j} ^ {\\mathcal {T}}) \\rangle}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 785, + 890, + 828 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\langle \\cdot, \\cdot \\rangle$ and $\\mathfrak{sg}(\\cdot)$ denote inner product and stop-gradient operation, respectively. Stop-gradient operation is used to prevent the collapsing solutions where every embedding becomes similar, as discussed in [9].", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15551", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed loss function encourages the matching embeddings to be similar to the input textual feature, while making the non-matching embeddings to be dissimilar. This approach ensures that the modality fusion module preserves the inherent information of the referring expression only when it is paired with the matching image (cycle-consistency), achieved by attending to the referred entities that may be absent in non-matching images. Therefore, an attention map $A^{\\mathrm{fuse}}$ is trained to reflect the referring expression accurately, where relevance between the expression and entities automatically emerges.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4.2 Reconstruction Loss", + "text_level": 1, + "bbox": [ + 76, + 268, + 272, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "If we can reconstruct the original data using only the set of slots, then the slots will correspond to distinctive individual entities in the input image. Therefore, we leverage the reconstruction objective $\\mathcal{L}_{\\mathrm{recon}}$ , which is given by", + "bbox": [ + 75, + 284, + 470, + 345 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n}} = \\frac {1}{B} \\sum_ {i = 0} ^ {B} \\left\\| \\operatorname {s g} \\left(\\mathbf {x} _ {i} ^ {\\mathcal {Y}}\\right) - f _ {\\mathrm {d e c}} \\left(\\Phi \\left(\\mathbf {x} _ {i} ^ {\\mathcal {Y}}\\right)\\right) \\right\\| ^ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 352, + 468, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $f_{\\mathrm{dec}}(\\cdot)$ is a MLP based decoder. Unlike the original slot attention [42], our reconstruction is performed in the feature space, rather than the pixel space. This feature reconstruction scheme enables an efficient decoder design. Detailed descriptions are provided in the supplementary material.", + "bbox": [ + 75, + 400, + 468, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Inference", + "text_level": 1, + "bbox": [ + 76, + 500, + 187, + 512 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The segmentation mask is computed with the two attention maps: $A^{\\mathrm{slot}} \\in \\mathbb{R}^{N \\times K}$ and $A^{\\mathrm{fuse}} \\in \\mathbb{R}^{K}$ . $A^{\\mathrm{slot}}$ is patchwise attention map that contains information about the discovered entities (Eq. (3)), while $A^{\\mathrm{fuse}}$ is entity-wise attention map denoting relevance between the referring expression and the entities (Eq. (7)). To compute the segmentation mask, we first compute $A^{\\mathrm{slot}} \\cdot A^{\\mathrm{fuse}}$ , which results in a 1-dimensional vector of size $N$ . We then reshape this vector into a 2-dimensional map. Next, we bilinearly interpolate the map to the original image size and rescale it using min-max normalization. Finally, we binarize the map using a threshold value $\\tau$ to obtain the segmentation mask.", + "bbox": [ + 75, + 518, + 468, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 709, + 209, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 76, + 731, + 277, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets: Our experiments utilize four datasets commonly used in referring image segmentation task: RefCOCO [70], RefCOCO+ [70], Gref [45], and PhraseCut [61]. RefCOCO, RefCOCO+, and Gref are obtained from COCO [39] dataset. RefCOCO contains 142,209 image-text pairs, while RefCOCO+ has 141,564 pairs and does not include expressions indicating orientation properties (e.g., top, right). Gref includes 104,560 pairs for 54,822 objects. PhraseCut comprises 345,486 image-text pairs, which are derived from Visual Genome [35].", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details: For visual and textual backbone, we employ ViT-S-16 pretrained on ImageNet-21K [11] and pretrained BERT [13], respectively. We make each modality feature have the same channel dimension $D$ by applying a $1 \\times 1$ convolution layer. We set $D$ and $D_h$ to 512 and 1024, respectively. In the aggregation block, the number of iteration $T$ is set to 6, and the number of slots $K$ is set to 36 except for PhraseCut, which is set to 60. For the entity slot, the number of slots sampled from each Gaussian distributions $K_s$ is set to 3 in RefCOCO+ and Gref, while 2 in RefCOCO and PhraseCut. During training, we use AdamW [44] optimizer, with the initial learning rate of 1e-4 scheduled with cosine annealing [43]. Batch size is set to 32 and model is trained for 50 epochs. We resize input images to $384 \\times 384$ for both training and evaluation. Threshold value $\\tau$ is set to 0.5 for all experiments.", + "bbox": [ + 496, + 90, + 890, + 332 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metric: Following [67], we employ the cumulative intersection-over-union (cIoU) and mean intersection-over-Union (mIoU) metric; cIoU divides the total intersections by the total unions across all test images, and mIoU computes IoU for each image and averages them. Also, we evaluate the accuracy at the IoU thresholds of $\\{0.3, 0.5, 0.7\\}$ , denote as $\\mathcal{A} @ \\{0.3, 0.5, 0.7\\}$ in Table 3.", + "bbox": [ + 496, + 334, + 890, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons with the State of the Art", + "text_level": 1, + "bbox": [ + 498, + 455, + 826, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Weakly supervised methods: As presented in Table 1, our method outperforms TSEG [53], the previous work on weakly-supervised referring image segmentation, in all benchmarks. It shows that the proposed bottom-up and top-down attention framework is more effective than using patch-based visual features with a carefully designed similarity function of TSEG.", + "bbox": [ + 496, + 474, + 890, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Open-vocabulary segmentation: In Table 1, we also compare our model with the recent open-sourced open-vocabulary segmentation models: GroupViT [64] and MaskCLIP [73], on the referring image segmentation benchmarks. Note that these methods are pre-trained with large-scale image-text datasets [5, 47, 54]. We also report their fine-tuned version for fair comparisons: models are additionally trained with the image-text pairs of each benchmark. Further details for the reproductions are provided in the supplementary material. Ours outperforms GroupViT and MaskCLIP in all settings and benchmarks, even without large-scale image-text pretraining. This result shows that open-vocabulary segmentation models [64, 73] have difficulty comprehending the complex free-from text involving the relation between entities, unlike ours. Interestingly, our model without fine-tuning, which pre-trained on the PhraseCut dataset and evaluated on other benchmarks, still achieved better results. This setting is identical to the open-vocabulary segmentation models, except it leverages a far smaller dataset. These results demonstrate the feasibility of our method in a zero-shot transfer setting.", + "bbox": [ + 496, + 583, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15552", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5951cccd42e601f5ae60924e30314009ec498f5f4d9d314ec6492ccd87d22fcd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsImage-text pretraining datasetFine tuningRefCOCORefCOCO+Gref valPhraseCut val
valtestAtestBvaltestAtestB
GroupViT [64]CC12M+YFCC (26M)X7.996.1610.518.496.7910.5910.685.82
✓10.8211.1111.2911.1410.7811.8412.779.41
MaskCLIP [73]WIT (400M)X11.5211.8512.0611.8712.0112.5712.7412.13
✓19.4518.6921.3719.9718.9321.4821.1123.80
TSEG [53]-✓25.44--22.01--22.0528.77
OursVisual Genome (0.31M)X21.8019.0024.9622.2019.8624.8525.89-
-✓34.7634.5835.0128.4828.6027.9828.8733.45
", + "bbox": [ + 116, + 88, + 849, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14f576d7b7ed4a9f9d9f63607e371c8fd82debd670a289241a43d13bdece0c3b.jpg", + "image_caption": [ + "Query: \"Guy on left with beard and guitar\"", + "Figure 4: Qualitative results of our framework with entity slot (Ours), random slot, and query slot on RefCOCO val set. In the case of entity slots, the boundary color indicates the Gaussian distribution that each slot is sampled from. For each slot type, we present 10 discovered entities from $A^{\\mathrm{slot}}$ and the final predictions." + ], + "image_footnote": [], + "bbox": [ + 76, + 290, + 890, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 492, + 230, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform ablation studies on the val set of RefCOCO to analyze the contributions of the entity discovery module and the modality fusion module.", + "bbox": [ + 75, + 512, + 470, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Entity discovery module: In Table 2(a), we ablate the entity discovery module and use visual feature directly, which leads to the drastical performance drop. This result demonstrates that contribution of our bottom-up attention framework is significant. Moreover, we ablate the interaction block, replace slot-attention of aggregation block to transformer, and train the model without reconstruction loss. The results show that all of the design choices for the module is important for effective entity discovery.", + "bbox": [ + 75, + 560, + 468, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Modality fusion module: In Table 2(b), the modality fusion module is ablated, and the model is trained with image-text contrastive learning [47], where the image feature is obtained by pooling the entity embeddings. The result demonstrates that inferring relevance between entities and referring expression with modality fusion module, trained with $\\mathrm{C}^3$ loss, is crucial for accurate segmentation.", + "bbox": [ + 75, + 700, + 468, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. In-depth Analysis", + "text_level": 1, + "bbox": [ + 76, + 820, + 250, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of slot initialization: To verify our design choice for entity slots, we investigate variants of slots; query, random, and entity slot. As summarized in Table 3, our entity slot outperforms other slot types across the different", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a4d72222ca8e02c37425656de18b78a940cb30d59f4ee316cdd862645f54eced.jpg", + "table_caption": [ + "Table 1: Comparison with weakly-supervised method (TSEG) and open-vocabulary segmentation method (GroupViT and MaskCLIP). The results on four datasets are reported in mIoU (\\%). Fine-tuning $\\checkmark$ means that the model is trained with the image-text pairs of the target benchmark; otherwise, it just relies on the pre-trained weight without further training." + ], + "table_footnote": [], + "table_body": "
cIoUmIoU
(a) Analysis on entity discovery module
Ours30.4034.76
w/o entity discovery module5.405.91
w/o interaction block29.5933.69
aggregation block w/ transformer11.2912.61
w/o reconstruction loss18.6121.84
(b) Analysis on modality fusion module
Ours30.4034.76
w/o modality fusion module and C3loss13.8114.65
", + "bbox": [ + 513, + 489, + 875, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2: Ablation studies of model components and loss function on RefCOCO val set.", + "bbox": [ + 500, + 660, + 890, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "numbers of slots. Specifically, as the number of slots is increased, the performance of the query slot declines, and that of the random slot converges. This result implies that simply increasing the number of slots can not handle the discovery of fine-grained visual entities in real-world images, and thus gains from the proposed entity slot type are non-trivial. In Fig. 4, we also present several visual entities discovered using each slot type, which show the superiority of the entity slot against others as explained in Sec. 3.2. Specifically, the entity slots from the same distribution share the same semantic property (e.g., head, body, or window) while capturing different fine-grained entities. The pro", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15553", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d4bd5fa503bc668a63907b6248fc962f30b8551aea5ddad4eb76ef3df679417.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 148, + 90, + 349, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3afde1d371f944240b051ee83bfb66269cfe19c9242be78ea73a44e183b3a864.jpg", + "image_caption": [ + "Prediction Ground Truth", + "Prediction Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 486, + 90, + 754, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3876b1fdd254e5d17cd010b355926aa0b6a6cf817368d1732c573d054817d73b.jpg", + "image_caption": [ + "Input Image", + "on left\"" + ], + "image_footnote": [], + "bbox": [ + 81, + 99, + 148, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c2e2055c93d7bd30bc89a7bce71af6290a47e9f92c8017fd9713a7c3b1a2200f.jpg", + "image_caption": [ + "Query: \"Right person glasses\"", + "Figure 5: Qualitative results of our framework on RefCOCO val set. We present the discovered entities from $A^{\\mathrm{slot}}$ and their relevance scores from $A^{\\mathrm{slot}}$ . Top-3 entities in terms of relevance to query expression are presented." + ], + "image_footnote": [], + "bbox": [ + 148, + 207, + 215, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c6aee612e77cd84364faccf5ac4672f5864255d047d7a90f37b3d79898198147.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 208, + 282, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/175c1b3fa9efcb4d93edab29328c248837500432843d6bf3637db927d00989e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 208, + 349, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/42dc45321ec730e96b3588542271f6b5a7121fd952f25bf07cf55bc86eef8615.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 208, + 416, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e9ac7d06d18138714c5a7b575e275b5de39907dc979823778807870c43b939c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 208, + 483, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e57ad991d52bab9adc68a4eaf4cdd9961b9b900f3033ff3f0437dd93eea64566.jpg", + "image_caption": [ + "Query: \"Jeans on right\"", + "Table 4: Impact of hyperparameters $K_{g}$ and $K_{s}$ (a) and inference schemes (b) on RefCOCO val set." + ], + "image_footnote": [], + "bbox": [ + 486, + 208, + 552, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/509784f85b9678d1583c72ce5453ab7a950f233d63083e48dab53dd6eb99d302.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 553, + 208, + 609, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0e953d3131db8f854a3fdcbebbbcfb0f0f55a1dc23a97e4339d30654543bf73f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 208, + 687, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/32446252ddc2a20e3b8394db36d4563ea5a92dc3fce8d448abc599b340c08e82.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 208, + 754, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/432a42f85675f372253e8afeb12cfab213b85484c93f65338685787316c95333.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 208, + 821, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c1573154031eb0f536c2de8ca95a455eb49c1ac14f6c1ac7366e16ab3449ebbd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 821, + 208, + 888, + 246 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b896815906e2352274e45bb4f5f1ed1bdccb5062e49c40f9125e741be9fcffb2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Slot type# slotsA@0.3A@0.5A@0.7cIoUmIoU
Query1227.577.170.9622.0421.44
2436.7310.611.9323.5924.78
3619.303.480.3116.6618.16
Random1236.3011.272.6924.2026.65
2445.1916.013.6626.6029.87
3645.6216.883.7626.5530.15
Entity(Ours)1239.4313.873.5624.7227.84
2452.3322.976.2329.2233.54
3655.0224.996.3530.4034.76
", + "bbox": [ + 84, + 306, + 457, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "posed entity slots effectively discover fine-grained visual entities, which results in an accurate prediction. In contrast, the prediction using random slots leads to a noisy mask, and using query slots leads to the mask of every head that appears in the image, respectively.", + "bbox": [ + 75, + 502, + 468, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of hyperparameters $K_{g}$ and $K_{s}$ : In Table 4(a), we investigate the impact of hyperparameters for the entity slots: $K_{g}$ and $K_{s}$ , which are the number of Gaussian distributions and slots sampled from each distribution, respectively. We keep the total number of entity slots to 36, only changing $K_{g}$ and $K_{s}$ . The optimal configurations are 18 and 2 for $K_{g}$ and $K_{s}$ , respectively. Interestingly, we observed that even entity slots with $K_{s} = 1$ , where a single slot is sampled from each distribution, vastly surpass the query slots in Table 3; it shows that random sampling of slots prevents the bounding to specific semantic category.", + "bbox": [ + 75, + 580, + 468, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of $A^{\\mathrm{fuse}}$ : To verify the efficacy of $A^{\\mathrm{fuse}}$ quantitatively, we evaluate our model with the following inference scheme variants: (1) AVG that replaces $A^{\\mathrm{fuse}}$ with the uniform probability distribution, (2) MAX that picks a single attention map from $A^{\\mathrm{slot}}$ with the maximum probability in $A^{\\mathrm{fuse}}$ , and (3) MIN that picks the attention map with the minimum probability. As summarized in Table 4(b), MAX largely outperforms AVG even though it only uses a single attention map for the final prediction, while MIN results in a near-zero cIoU and mIoU. This tendency shows that $A^{\\mathrm{fuse}}$", + "bbox": [ + 75, + 748, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c6a627b03b40b4e9b217290888f875808b1be263246e52b9f985bf1f8fa7e26c.jpg", + "table_caption": [ + "Table 3: Comparison between different slot types on Ref-COCO val set, where # slots is the total number of slots." + ], + "table_footnote": [], + "table_body": "
KgKscIoUmIoU
36129.3633.41
18230.4034.76
12329.6334.03
9428.3132.34
6627.2331.30
", + "bbox": [ + 504, + 306, + 658, + 388 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2653be3e59be38b6aa7412e54da24af530b1bb934cc553f834fbc17bb07e5cca.jpg", + "table_caption": [ + "(a)" + ], + "table_footnote": [], + "table_body": "
Inference schemecIoUmIoU
AVG9.018.16
MAX14.9319.54
MIN0.160.14
A fuse · A slot (Ours)30.4034.76
", + "bbox": [ + 668, + 308, + 879, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(b)", + "bbox": [ + 767, + 388, + 785, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "effectively assigns probability based on the relevance between reference and entities, even without direct supervision.", + "bbox": [ + 496, + 454, + 890, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Analysis: In Fig. 5, predicted segmentation mask and discovered entities with top-3 relevance are presented. Results show that the entity discovery module effectively discovers primitive units of segmentation masks, and the modality fusion module accurately infers their relevance to query referring expression.", + "bbox": [ + 496, + 500, + 890, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 601, + 617, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we address the problem of learning referring image segmentation with only image-text pairs. We present a new model based on a bottom-up and top-down attention framework, which first discovers entities in the input image and then combines such entities based on their relevance to the text query. Moreover, we present a new loss function enforcing cycle-consistency between image-text pairs, which enables the effective training of the model without any further supervision. As a result, the proposed method clearly outperforms an existing weakly-supervised method and recent open-vocabulary segmentation models. For future work, we will explore expanding our model to different modalities such as video and audio.", + "bbox": [ + 496, + 621, + 890, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by the NRF grant and the IITP grants funded by Ministry of Science and ICT, Korea (NRF-2021R1A2C3012728-20%, IITP-2020-0-00842-20%, IITP-2022-0-00290-20%, IITP-2022-0-00926-20%, IITP-2021-0-02068-10%, IITP-2019-0-01906-10%).", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15554", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jiwoon Ahn, Sunghyun Cho, and Suha Kwak. Weakly supervised learning of instance segmentation with inter-pixel relations. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[2] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3", + "[3] Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. volume abs/1607.06450, 2016. 5", + "[4] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In Proc. Neural Information Processing Systems (NeurIPS), 2019. 3", + "[5] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 6", + "[6] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In Proc. IEEE International Conference on Computer Vision (ICCV), 2019. 2", + "[7] Jianbo Chen, Yelong Shen, Jianfeng Gao, Jingjing Liu, and Xiaodong Liu. Language-based image editing with recurrent attentive models. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1", + "[8] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In Proc. Neural Information Processing Systems (NeurIPS), 2021. 1", + "[9] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5", + "[10] Samyak Datta, Karan Sikka, Anirban Roy, Karuna Ahuja, Devi Parikh, and Ajay Divakaran. Align2ground: Weakly supervised phrase grounding guided by image-caption alignment. In Proc. IEEE International Conference on Computer Vision (ICCV), 2019. 2", + "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: a large-scale hierarchical image database. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2009. 6", + "[12] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers. In Proc. IEEE International Conference on Computer Vision (ICCV), 2021. 2", + "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Annual Conference of the North American Chapter of the Association for Computational Linguistics, 2019. 4, 6", + "[14] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proc. IEEE International Conference on Computer Vision (ICCV), 2021. 1, 2", + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl-" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "vain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In Proc. International Conference on Learning Representations (ICLR), 2021. 2, 4", + "[16] Guang Feng, Zhiwei Hu, Lihe Zhang, and Hutchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[17] Tsu-Jui Fu, Xin Eric Wang, Scott T. Grafton, Miguel P. Eckstein, and William Yang Wang. M3l: Language-based video editing via multi-modal multi-level transformers. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1", + "[18] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In Proc. European Conference on Computer Vision (ECCV), 2022. 2, 3", + "[19] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 2", + "[20] Tanmay Gupta, Arash Vahdat, Gal Chechik, Xiaodong Yang, Jan Kautz, and Derek Hoiem. Contrastive learning for weakly supervised phrase grounding. In European Conference on Computer Vision. Springer, 2020. 2", + "[21] S. Hong, D. Yeo, S. Kwak, H. Lee, and B. Han. Weakly supervised semantic segmentation using web-crawled videos. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2224-2232, 2017. 3", + "[22] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In Proc. European Conference on Computer Vision (ECCV), 2016. 1, 2", + "[23] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[24] Binbin Huang, Dongze Lian, Weixin Luo, and Shenghua Gao. Look before you leap: Learning landmark features for one-stage visual grounding. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[25] Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoqi Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2", + "[26] Zilong Huang, Xinggang Wang, Jiasi Wang, Wenyu Liu, and Jingdong Wang. Weakly-supervised semantic segmentation network with deep seeded region growing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3", + "[27] Tianrui Hui, Si Liu, Shaofei Huang, Guanbin Li, Sansi Yu, Faxi Zhang, and Jizhong Han. Linguistic structure guided context modeling for referring image segmentation. In Proc. European Conference on Computer Vision (ECCV), 2020. 1", + "[28] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proc. International" + ], + "bbox": [ + 501, + 92, + 890, + 892 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15555", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Machine Learning (ICML), 2021. 1, 3", + "[29] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proc. IEEE International Conference on Computer Vision (ICCV), 2021. 2", + "[30] A. Khoreva, R. Benenson, J. Hosang, M. Hein, and B. Schiele. Simple does it: Weakly supervised instance and semantic segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3", + "[31] Beomyoung Kim, Youngjoon Yoo, Chae Eun Rhee, and Junmo Kim. Beyond semantic to instance segmentation: Weakly-supervised instance segmentation via semantic knowledge transfer and self-refinement. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[32] Dongwon Kim, Namyup Kim, and Suha Kwak. Improving cross-modal retrieval with set of diverse embeddings. 2022. 3, 4", + "[33] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Restr: Convolution-free referring image segmentation using transformers. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[34] Alexander Kolesnikov and Christoph H. Lampert. Seed, expand and constrain: Three principles for weakly-supervised image segmentation. In Proc. European Conference on Computer Vision (ECCV), 2016. 3", + "[35] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, Michael Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision (IJCV), 2017. 6", + "[36] Suha Kwak, Seunghoon Hong, and Bohyung Han. Weakly supervised semantic segmentation using superpixel pooling network. In Proc. AAAI Conference on Artificial Intelligence (AAAI), 2017. 3", + "[37] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In Proc. International Conference on Learning Representations (ICLR), 2022. 3", + "[38] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: common objects in context. In Proc. European Conference on Computer Vision (ECCV), 2014. 6", + "[40] Chenxi Liu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, and Alan Yuille. Recurrent multimodal interaction for referring image segmentation. In Proc. IEEE International Conference on Computer Vision (ICCV), 2017. 2", + "[41] Yongfei Liu, Bo Wan, Lin Ma, and Xuming He. Relation-aware instance refinement for weakly supervised visual grounding. In Proc. IEEE Conference on Computer Vision" + ], + "bbox": [ + 78, + 92, + 468, + 892 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Pattern Recognition (CVPR), 2021. 2", + "[42] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. In Proc. Neural Information Processing Systems (NeurIPS), 2020. 2, 3, 4, 5, 6", + "[43] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In Proc. International Conference on Learning Representations (ICLR), 2017. 6", + "[44] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. Proc. International Conference on Learning Representations (ICLR), 2019. 6", + "[45] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6", + "[46] Edgar Margffoy-Tuay, Juan C Pérez, Emilio Botero, and Pablo Arbeláez. Dynamic multimodal instance segmentation guided by natural language queries. In Proc. European Conference on Computer Vision (ECCV), 2018. 2", + "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. International Conference on Machine Learning (ICML), 2021. 1, 2, 3, 6, 7", + "[48] Tal Remez, Jonathan Huang, and Matthew Brown. Learning to segment via cut-and-paste. In Proc. European Conference on Computer Vision (ECCV), 2018. 3", + "[49] Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, and Bernt Schiele. Grounding of textual phrases in images by reconstruction. In Proc. European Conference on Computer Vision (ECCV), 2016. 2", + "[50] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3", + "[51] Hengcan Shi, Hongliang Li, Fanman Meng, and Qingbo Wu. Key-word-aware network for referring expression image segmentation. In Proc. European Conference on Computer Vision (ECCV), 2018. 2", + "[52] Gautam Singh, Fei Deng, and Sungjin Ahn. Illiterate dalle learns to compose. In Proc. International Conference on Learning Representations (ICLR), 2021. 3, 4", + "[53] Robin Strudel, Ivan Laptev, and Cordelia Schmid. Weakly-supervised segmentation of referring expressions. arXiv preprint arXiv:2205.04725, 2022. 2, 6, 7", + "[54] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 2016. 2, 6", + "[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. Neural Information Processing Systems (NeurIPS), 2017. 5", + "[56] Liwei Wang, Jing Huang, Yin Li, Kun Xu, Zhengyuan Yang, and Dong Yu. Improving weakly supervised visual ground" + ], + "bbox": [ + 503, + 92, + 890, + 890 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15556", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing by contrastive knowledge distillation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[57] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1", + "[58] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[59] Yunchao Wei, Jiashi Feng, Xiaodan Liang, Ming-Ming Cheng, Yao Zhao, and Shuicheng Yan. Object region mining with adversarial erasing: A simple classification to semantic segmentation approach. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3", + "[60] Philippe Weinzaepfel, Thomas Lucas, Diane Larlus, and Yannis Kalantidis. Learning super-features for image retrieval. In Proc. International Conference on Learning Representations (ICLR), 2022. 3, 4", + "[61] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasecut: Language-based image segmentation in the wild. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6", + "[62] Yongqin Xian, Subhabrata Choudhury, Yang He, Bernt Schiele, and Zeynep Akata. Semantic projection network for zero-and few-label semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8256-8265, 2019. 3", + "[63] Fanyi Xiao, Leonid Sigal, and Yong Jae Lee. Weakly-supervised visual grounding of phrases with linguistic structures. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2", + "[64] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4, 6, 7", + "[65] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310-4319, 2022. 3", + "[66] Li Yang, Yan Xu, Chunfeng Yuan, Wei Liu, Bing Li, and Weiming Hu. Improving visual grounding with visual-linguistic verification and iterative reasoning. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[67] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6", + "[68] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 893 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[69] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1, 2", + "[70] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Proc. European Conference on Computer Vision (ECCV), 2016. 6", + "[71] Dong Zhang, Hanwang Zhang, Jinhui Tang, Xian-Sheng Hua, and Qianru Sun. Causal intervention for weakly-supervised semantic segmentation. 2020. 3", + "[72] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba. Learning deep features for discriminative localization. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3", + "[73] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In Proc. European Conference on Computer Vision (ECCV), 2022. 2, 3, 6, 7", + "[74] Yi Zhou, Hui Zhang, Hana Lee, Shuyang Sun, Pingjun Li, Yangguang Zhu, ByungIn Yoo, Xiaojuan Qi, and Jae-Joon Han. Slot-vps: Object-centric representation learning for video panoptic segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 4", + "[75] Yanzhao Zhou, Yi Zhu, Qixiang Ye, Qiang Qiu, and Jianbin Jiao. Weakly supervised instance segmentation using class peak response. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3", + "[76] Yuliang Zou, Zizhao Zhang, Han Zhang, Chun-Liang Li, Xiao Bian, Jia-Bin Huang, and Tomas Pfister. Pseudoseg: Designing pseudo labels for semantic segmentation. In Proc. International Conference on Learning Representations (ICLR), 2021. 3" + ], + "bbox": [ + 501, + 92, + 890, + 553 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15557", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Shatter and Gather_ Learning Referring Image Segmentation with Text Supervision/eba869e5-3a94-47e8-afcc-2f1d3dc46bfb_model.json b/2023/Shatter and Gather_ Learning Referring Image Segmentation with Text Supervision/eba869e5-3a94-47e8-afcc-2f1d3dc46bfb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9982d2c36c92452767460cca117e2528c68a472c --- /dev/null +++ b/2023/Shatter and Gather_ Learning Referring Image Segmentation with Text Supervision/eba869e5-3a94-47e8-afcc-2f1d3dc46bfb_model.json @@ -0,0 +1,2642 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.165, + 0.131, + 0.8, + 0.177 + ], + "angle": 0, + "content": "Shatter and Gather: Learning Referring Image Segmentation with Text Supervision" + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.199, + 0.757, + 0.253 + ], + "angle": 0, + "content": "Dongwon Kim\\(^{1*}\\) Namyup Kim\\(^{1*}\\) Cuiling Lan\\(^{2}\\) Suha Kwak\\(^{1}\\) \n\\(^{1}\\)POSTECH Microsoft Research Asia \nhttp://cvlab.postech.ac.kr/research/sag" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.288, + 0.314, + 0.304 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.474, + 0.562 + ], + "angle": 0, + "content": "Referring image segmentation, the task of segmenting any arbitrary entities described in free-form texts, opens up a variety of vision applications. However, manual labeling of training data for this task is prohibitively costly, leading to lack of labeled data for training. We address this issue by a weakly supervised learning approach using text descriptions of training images as the only source of supervision. To this end, we first present a new model that discovers semantic entities in input image and then combines such entities relevant to text query to predict the mask of the referent. We also present a new loss function that allows the model to be trained without any further supervision. Our method was evaluated on four public benchmarks for referring image segmentation, where it clearly outperformed the existing method for the same task and recent open-vocabulary segmentation models on all the benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.585, + 0.21, + 0.601 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.47, + 0.727 + ], + "angle": 0, + "content": "Referring image segmentation is the task of segmenting the referent corresponding to a natural language expression given as a query. Unlike the conventional semantic segmentation that aims at segmenting a pre-defined set of classes, referring image segmentation enables segmentation of any arbitrary entities described in free-form texts and thus opens up a wide variety of applications such as human-computer interaction [8, 57] and text-based image editing [7, 17]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.471, + 0.878 + ], + "angle": 0, + "content": "Thanks to the development of deep neural networks, recent studies have demonstrated remarkable results on referring image segmentation in the supervised learning setting [14, 22, 25, 27, 67, 68, 69]. However, obtaining manual annotation of training data for referring image segmentation is prohibitively costly since the data requires two types of labels for each image, i.e., natural language expressions describing the entities that appear in the image and segmentation labels corresponding to them. This typically demands tremendously more manpower than annotation for semantic" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.287, + 0.892, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.564, + 0.895, + 0.686 + ], + "angle": 0, + "content": "Figure 1: Our model consists of two different attention modules: bottom-up attention that identifies individual entities in an image, and top-down attention that combines the entities based on a referring expression. It is trained in a contrastive manner to ensure consistency between matching image-text pairs. In inference, a segmentation mask is predicted by combining the entities found by bottom-up attention with weights derived by top-down attention." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.704, + 0.879, + 0.719 + ], + "angle": 0, + "content": "segmentation, leading to lack of labeled data for training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "A solution to this issue is weakly supervised learning, which trains a model on a dataset providing weaker forms of supervision than the conventional ones. In this paper, we particularly focus on learning referring image segmentation only with natural language expressions describing the associated images. Since these labels are readily available in many vision-language datasets [28, 47], the issue on lack of labeled training data can be alleviated by our approach. However, learning referring image segmentation in this setting introduces another challenge: Since natural language expressions often contain relations between visual entities such as instances and object parts (e.g., \"The cat next to the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.214, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15547" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "table leg\"), a model to be trained should be aware of the inter-object relations to exploit the expressions as supervision. Thereby we need a method to discover individual entities and infer their relations without any supervision but the referring expression." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.169, + 0.471, + 0.381 + ], + "angle": 0, + "content": "To fulfill this requirement, we introduce a new bottom-up and top-down attention framework, which is illustrated in Fig. 1. Individual entities existing in an image are first discovered by bottom-up attention, which solely exploits visual information. For the bottom-up attention, an entity discovery module progressively refines a set of embedding vectors named slots to capture distinct visual entities, following slot attention [42]. We moreover propose a novel slot formulation named entity slot, which enables fine-grained entity discovery in real-world images. As a result, entity slots after the refinement capture individual visual entities without any segmentation supervision. These entities are considered primitive units for composing the segmentation mask of a referent during inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.47, + 0.534 + ], + "angle": 0, + "content": "Afterward, based on the relational and structural information of referring expression, top-down attention selectively attends to the referred entities and combines them into the predicted mask. Top-down attention is implemented by a modality fusion module consists of cross-attention transformer [15]. This attention scheme enables the module to infer the relevance between discovered entities and the referring expression instead of relying on the cosine similarity or heuristic similarity functions used in previous work [18, 53, 64, 73]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.672 + ], + "angle": 0, + "content": "For training, we propose the contrastive cycle-consistency \\((\\mathrm{C}^3)\\) loss that enforces cycle-consistency between image-text pairs under the contrastive learning framework [47]. To establish such consistency, the model is trained to preserve the textual information after the fusion of entities and textual features. By doing so, we observe that latent relevance between the discovered entities and referring expression automatically emerges in the top-down attention without any explicit mask supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.47, + 0.764 + ], + "angle": 0, + "content": "One may wonder the difference between our task and open-vocabulary segmentation [18, 64, 73], i.e., semantic segmentation of arbitrary categories without explicit mask supervision. The key difference is that a referent to segment in referring image segmentation is given by a complex free-form text, which introduces additional challenges." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The proposed method was evaluated and compared on four public benchmarks for referring image segmentation, where it substantially outperformed the previous weakly supervised learning method [53]. Moreover, we reproduced recent open-vocabulary segmentation models [64, 73] and evaluated them on referring image segmentation benchmarks; the results show that the proposed model surpassed them even without pre-training on large-scale image-text data [5, 47, 54]." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.874, + 0.108 + ], + "angle": 0, + "content": "In summary, our contribution is three-fold as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.111, + 0.892, + 0.203 + ], + "angle": 0, + "content": "- We present a new weakly supervised learning model for referring image segmentation. Our model employs bottom-up and top-down attention to discover individual entities and infer their relations, which enables to exploit natural language expressions as supervision effectively during training." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.206, + 0.892, + 0.267 + ], + "angle": 0, + "content": "- We propose a new loss function called contrastive cycle-consistency loss, which allows latent relevance between discovered entities and referring expressions to emerge without requiring further supervision." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.269, + 0.892, + 0.313 + ], + "angle": 0, + "content": "- Our model clearly outperformed the existing method for the same task and recent open-vocabulary segmentation models on all the four benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.111, + 0.892, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.334, + 0.642, + 0.35 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.672 + ], + "angle": 0, + "content": "Referring image segmentation: Early approaches [22, 38, 40, 51] obtain multimodal feature maps by simple concatenation-convolution operation. With this operation as a basis, most existing studies have suggested attention mechanisms and multi-level feature aggregations [6, 16, 23, 25, 46, 68]. Various approaches have been employed for calculating fine interactions between the modalities, including multimodal feature fusion [14, 29, 33] using transformers [15], a large-scale vision-language pretrained framework [58], and holistically multimodal fusion [67]. The existing referring image segmentation methods have been developed in a fully supervised learning setting. Recently, a weakly supervised learning method [53] using only imagedtext pairs for training is introduced to alleviate the high annotation cost. However, this study relies on a heuristic similarity function to infer relations between patches and texts, which suffers from comprehending the relational and structural information from image and text modalities. Therefore, we propose a new weakly supervised model consisting of bottom-up and top-down attention to discover visual entities and infer their relations effectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Phrase grounding: Phrase grounding [12, 24, 29, 49, 66, 69] is the task of learning the association between image regions and words in caption describing the image. Weakly supervised approaches have also been proposed for this task [10, 20, 41, 56, 63], where the correspondence between words and regions are not given. It is worth noting that the contrastive learning framework introduced in [20] is closely related to the proposed \\(C^3\\) loss by enforcing mutual information between regions and words. However, these studies rely on an off-the-shelf object detector [19], where the final prediction is restricted by the predefined set of classes with which the detector is trained. On the contrary, our method discovers visual entities in an image without an auxiliary model trained with external labeled data, which is more suitable for referring image segmentation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15548" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.891, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.318, + 0.896, + 0.425 + ], + "angle": 0, + "content": "Figure 2: Illustration of the overall architecture of our model along with its behavior during training and inference. (a) Feature extractors: The visual and textual features are extracted by transformer-based encoders of the two modalities respectively. (b) Entity discovery module: A set of visual entities are discovered from the visual features through the bottom-up attention. (c) Modality fusion module: The visual entities and a referring expression are fused considering their relevance, which is estimated by the top-down attention. Our model is trained with the proposed contrastive cycle-consistency \\((\\mathbf{C}^3)\\) loss between the cross-modal embeddings and textual features. In inference, the segmentation mask is predicted by jointly considering the visual entities \\(A^{\\mathrm{slot}}\\) from the entity discovery module and the relevance scores \\(A^{\\mathrm{fuse}}\\) from the modality fusion module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.473, + 0.637 + ], + "angle": 0, + "content": "Weakly supervised learning for segmentation: Weakly supervised learning has become a favored solution for addressing the lack of annotated data in training, especially for semantic segmentation [1, 2, 21, 26, 34, 36, 65, 71, 76] and instance segmentation [30, 31, 48, 75]. These methods refine coarse localization cues obtained by CAMs [50, 59, 72], and use the results as pseudo labels for self-training. However, they cannot be directly applied to the referring image segmentation due to the following challenges. First, there are no predefined classes in this task, making it difficult to learn shared semantics within CAMs by multiple instance learning. Second, there is no mechanism to discover visual entities and combine them with textual information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.473, + 0.852 + ], + "angle": 0, + "content": "Slot attention: Slot attention [42] is proposed for the learning object-centric representation. It can discover the objects within an image in an unsupervised manner and be used in the various downstream task such as segmentation [64, 74] and retrieval [32, 60]. In specific, slot attention introduces an iterative attention mechanism that operates on the slots, which are the set of embedding vectors corresponding to the objects that appear in the input data. In previous work, slots are initialized as a set of learnable embeddings [32, 60, 64, 74] or a set of vectors sampled from a learnable Gaussian distribution [42, 52]. However, we empirically observed that both types of initialization are inapplicable to visual entity discovery for referring image segmentation, which we will elaborate on in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Open-vocabulary segmentation: The open-vocabulary segmentation is proposed to address the limitation of semantic segmentation that it is constrained to the pre" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.895, + 0.683 + ], + "angle": 0, + "content": "defined set of classes. Unlike semantic segmentation, open-vocabulary segmentation models are evaluated with classes unseen during training. The early approaches propose a generative model [4] and a projection model [62] to generate or project pixel-level semantic features from visual features for the unseen classes through word embedding space. Recent studies [18, 37, 64, 73] largely improve the quality of semantic features for the unseen classes, and further introduce the methods that do not require segmentation labels for seen classes using large-scale image-text pretrained models (e.g.CLIP [47], ALIGN [28]). However, since they are trained to predict masks for semantic categories, not for free-form texts describing arbitrary entities in an image, their applicability is still limited. We thus propose a new method for referring image segmentation that only requires image-text pairs during training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.698, + 0.673, + 0.715 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 2, our model consists of three main components: feature extractors, entity discovery module, and modality fusion module. The visual and textual features are extracted from the input using the backbone networks of the two modalities respectively. (Sec. 3.1). The visual features are then given as input to the entity discovery module, which aggregates the features into multiple entity embeddings using the bottom-up attention (Sec. 3.2). The entity embeddings correspond to distinct entities appearing in the image, which are considered as the smallest units of building the segmentation mask. The modality fusion module fuses the entity embeddings and the textual feature into" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15549" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "cross-modal embeddings according to their relevance using the top-down attention (Sec. 3.3). The produced cross-modal embedding is used only for training, with a learning objective enforcing cycle-consistency and reconstruction of visual features (Sec. 3.4). The segmentation mask is predicted by combining the entities discovered by the entity discovery module with the relevance scores estimated by the modality fusion module (Sec. 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.224, + 0.408, + 0.239 + ], + "angle": 0, + "content": "3.1. Visual and Textual Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.47, + 0.273 + ], + "angle": 0, + "content": "Here, we provide the details of the transformer-based backbone network for each modality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.469, + 0.366 + ], + "angle": 0, + "content": "Visual feature extraction: We employ the vision transformer (ViT) [15] as the visual backbone network. An input image is first split into \\(N\\) patches and then linearly projected. The patches are processed by the multiple self-attention blocks of ViT, where we refer to the output tokens of the last block as the visual features \\(\\mathbf{x}^{\\mathcal{V}}\\in \\mathbb{R}^{N\\times D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.47, + 0.46 + ], + "angle": 0, + "content": "Textual feature extraction: We consider the transformer-based text encoder, BERT [13]. A sentence is first tokenized into the token sequence, where the [CLS] and [SEP] tokens are added to the beginning and the end of the sequence. The activation of the last layer of the transformer for the [CLS] token is employed as the textual feature \\(\\mathbf{x}^{\\mathcal{T}}\\in \\mathbb{R}^{D}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.47, + 0.469, + 0.485 + ], + "angle": 0, + "content": "3.2. Bottom-up Attention: Entity Discovery Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.469, + 0.61 + ], + "angle": 0, + "content": "The entity discovery module applies a bottom-up attention mechanism to aggregate patch-based visual features into multiple entity embeddings that represent instances or object parts in the image. The discovered entities are then treated as primitive units of the segmentation mask, where the mask is predicted by selectively combining them. This approach is useful since the textual information is often built upon the relations between such visual entities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.763 + ], + "angle": 0, + "content": "Specifically, the entity discovery module consists of two blocks: aggregation block and interaction block. The entity slots, which are embedding vectors that corresponds to various visual entities, are updated in an iterative process alternating between these two blocks. These slots eventually become entity embeddings. Let \\(\\Phi : \\mathbb{R}^{N \\times D} \\mapsto \\mathbb{R}^{K \\times D}\\) be the entity discovery module that takes visual features as input and outputs the set of \\(K\\) entity embeddings. The entity embeddings are obtained by iteratively updating initial entity slots \\(\\mathbf{S}^0 \\in \\mathbb{R}^{K \\times D}\\) \\(T\\)-times:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.774, + 0.469, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r l} & \\Phi (\\mathbf {x} ^ {\\mathcal {V}}) := \\mathbf {S} ^ {T}, \\text {w h e r e} \\\\ & \\quad \\mathbf {S} ^ {t + 1} = \\operatorname {I n t e r} \\left(\\operatorname {A g g} \\left(\\mathbf {x} ^ {\\mathcal {V}}; \\mathbf {S} ^ {t}\\right)\\right), \\end{array} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.855 + ], + "angle": 0, + "content": "and Agg and Inter denote the aggregation and interaction blocks, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In the literature, the initial slots \\(\\mathbf{S}^0\\) are often defined as either vectors sampled from a learnable Gaussian distribution [42, 52] or learnable embeddings [32, 60, 64, 74]," + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.09, + 0.878, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.366, + 0.893, + 0.457 + ], + "angle": 0, + "content": "Figure 3: Comparison between the proposed entity slot and previous slot initialization methods. The boundary color indicates the Gaussian distribution that each slot is sampled from. The entity slot successfully discovers fine-grained entities such as head and food, while the others often fail to capture individual entities of specific semantics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.893, + 0.759 + ], + "angle": 0, + "content": "where such types of slots are referred as random slot and query slot in the remainder of this paper, respectively. However, we have empirically found that both of these initialization strategies are unsuitable for our task as shown in Fig. 3. The random slot suffers from the lack of semantic specificity since every slot is randomly sampled from a single distribution. This can lead to suboptimal performance since entities in real-world images are diverse in terms of semantics. The query slot, on the other hand, suffers from the limitation of being bound to specific semantic categories, such as human, animal, and vehicle. This limitation leads to entity discovery that is not sufficiently fine-grained, e.g., given an image of multiple people, a single slot that binds to \"human\" captures all of them and thus failed to identify the individual specified by the text query. This is problematic particularly in our model since entities act as the smallest units for producing segmentation masks. It is highly desired to have a mechanism that could guide the differentiation of slots to be semantically specific." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.759, + 0.892, + 0.851 + ], + "angle": 0, + "content": "To overcome these limitations, we introduce a new initialization strategy for the entity slot, which enables the discovery of fine-grained visual entities while learning diverse semantics. Formally, the initial entity slots \\(\\mathbf{S}^0\\) are obtained by sampling \\(K_{s}\\) slots from each of \\(K_{g}\\) different Gaussian distributions, resulting in a total of \\(K = K_{s}\\cdot K_{g}\\) slots:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.857, + 0.891, + 0.884 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} ^ {0} := \\left[ \\mathbf {S} _ {1} ^ {0}; \\dots ; \\mathbf {S} _ {K _ {g}} ^ {0} \\right] \\in \\mathbb {R} ^ {K \\times D}, \\text {w h e r e} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.879, + 0.793, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {i} ^ {0} \\sim \\mathcal {N} (\\mu_ {i}, \\operatorname {d i a g} (\\sigma_ {i})) \\in \\mathbb {R} ^ {K _ {s} \\times D}.\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15550" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "The idea is to initialize slots by sampling them from different Gaussian distributions, each specifying a distinct semantic property to be captured. Individual entities sharing the same semantics are differentiated by slots sampled from the same distribution, which enables fine-grained entity discovery." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.471, + 0.349 + ], + "angle": 0, + "content": "From now, we present the aggregation block and interaction block at \\( t \\)-th iteration. As described in Eq. (1), the slots of previous iteration \\( \\mathbf{S}^{t-1} \\) are fed to the aggregation block with the visual features \\( \\mathbf{x}^{\\nu} \\). The role of the aggregation block is to let each slot aggregate visual features of distinct visual entities, and it is achieved by employing the competitive attention mechanism introduced in the slot attention [42]. Let \\( q(\\cdot) \\), \\( k(\\cdot) \\), and \\( v(\\cdot) \\) are linear projection with layer normalization [3] that map inputs to dimension \\( D_h \\). Then, the attention map \\( A^{\\mathrm{slot}} \\in \\mathbb{R}^{N \\times K} \\) between \\( \\mathbf{x}^{\\nu} \\) and \\( \\mathbf{S}^{t-1} \\) is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.358, + 0.47, + 0.401 + ], + "angle": 0, + "content": "\\[\nA _ {i, j} ^ {\\text {s l o t}} = \\frac {e ^ {M _ {i , j}}}{\\sum_ {j = 1} ^ {K} e ^ {M _ {i , j}}}, \\text {w h e r e} M = \\frac {k \\left(\\mathbf {x} ^ {\\mathcal {V}}\\right) q \\left(\\mathbf {S} ^ {t - 1}\\right) ^ {\\top}}{\\sqrt {D _ {h}}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.469, + 0.546 + ], + "angle": 0, + "content": "Unlike the cross-attention transformer [55] that normalized attention over the keys, we normalize attention map \\( A^{\\mathrm{slot}} \\) over the slots. This normalization scheme encourages slots to compete for aggregating the visual features, which results in the set of slots binding to different visual entities in the input data. It is worth noting that \\( A^{\\mathrm{slot}} \\) provides the localizations of entities discovered, which is used as primitive units of the segmentation mask during inference. Using \\( A^{\\mathrm{slot}} \\), the slots are updated in a similar manner to the transformer:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.568, + 0.469, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\widehat {\\mathbf {S}} ^ {t} = \\operatorname {A g g} \\left(\\mathbf {x} ^ {\\nu}; \\mathbf {S} ^ {t}\\right) \\tag {4} \\\\ = \\operatorname {n o r m a l i z e} \\left(A ^ {\\text {s l o t}}\\right) ^ {\\top} v \\left(\\mathbf {x} ^ {\\nu}\\right) W _ {o} + \\mathbf {S} ^ {t - 1}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.619, + 0.469, + 0.68 + ], + "angle": 0, + "content": "where normalize \\((\\cdot)\\) denotes \\(\\ell_1\\)-normalization of the column vectors and \\(W_{o}\\in \\mathbb{R}^{D_{h}\\times D}\\) is linear projection. Finally, slots are further refined by \\(\\widehat{\\mathbf{S}}^t\\gets \\mathrm{MLP}\\big(\\widehat{\\mathbf{S}}^t\\big) + \\widehat{\\mathbf{S}}^t\\) where MLP is multi-layer perceptron (MLP) with layer normalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.469, + 0.785 + ], + "angle": 0, + "content": "The following interaction block consists of multiple self-attention transformers [55]. The weights of these transformers are shared, and each transformer is applied to slots sampled from the same distribution. We denote the transformer layer as Transformer(Q; K; V), where (Q; K; V) are for query, key, and value. The interaction block is denoted as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.795, + 0.469, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\operatorname {I n t e r} \\left(\\widehat {\\mathbf {S}} ^ {t}\\right) = \\left[ \\operatorname {S A} \\left(\\widehat {\\mathbf {S}} _ {1} ^ {t}\\right); \\dots ; \\operatorname {S A} \\left(\\widehat {\\mathbf {S}} _ {K _ {g}} ^ {t}\\right) \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathrm{SA}(\\mathbf{X})\\) is for the Transformer(X; X; X). This self-attention scheme models information exchange between visual entities sharing the same semantic property, which helps slots within a single semantic category to capture different entities without redundancy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.892, + 0.108 + ], + "angle": 0, + "content": "3.3. Top-down Attention: Modality Fusion Module" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.111, + 0.892, + 0.216 + ], + "angle": 0, + "content": "The modality fusion module is proposed for top-down attention mechanism, which combines discovered visual entities based on the textual feature. Let \\(\\Psi: \\mathbb{R}^{K \\times D} \\times \\mathbb{R}^D \\mapsto \\mathbb{R}^D\\) be the modality fusion module, which takes entity embeddings \\(\\Phi(\\mathbf{x}^\\nu)\\) and textual furret \\(\\mathbf{x}^T\\) as inputs. Then, the output cross-modal embedding \\(\\mathbf{z} \\in \\mathbb{R}^D\\) is computed using cross-attention transformer:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.225, + 0.892, + 0.263 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} = \\Psi \\left(\\Phi \\left(\\mathbf {x} ^ {\\mathcal {V}}\\right); \\mathbf {x} ^ {\\mathcal {T}}\\right) = \\frac {\\operatorname {C A} \\left(\\Phi \\left(\\mathbf {x} ^ {\\mathcal {V}}\\right) ; \\mathbf {x} ^ {\\mathcal {T}}\\right)}{\\| \\operatorname {C A} \\left(\\Phi \\left(\\mathbf {x} ^ {\\mathcal {V}}\\right) ; \\mathbf {x} ^ {\\mathcal {T}}\\right) \\| _ {2}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.274, + 0.892, + 0.32 + ], + "angle": 0, + "content": "where \\(\\mathrm{CA}(\\mathrm{X};\\mathrm{Y})\\) is for the Transformer(Y;X;X) and \\(\\| \\cdot \\| _2\\) denotes \\(\\ell_2\\) norm. Moreover, we can obtain the attention map \\(A^{\\mathrm{fuse}}\\in \\mathbb{R}^{K}\\) as a byproduct in cross-attention:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.33, + 0.892, + 0.369 + ], + "angle": 0, + "content": "\\[\nA _ {i} ^ {\\text {f u s e}} = \\frac {e ^ {M _ {i}}}{\\sum_ {i = 1} ^ {K} e ^ {M _ {i}}} \\text {, w h e r e} M = \\frac {k \\left(\\Phi \\left(\\mathbf {x} ^ {\\mathcal {V}}\\right)\\right) q \\left(\\mathbf {x} ^ {\\mathcal {T}}\\right)}{\\sqrt {D _ {h}}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.379, + 0.892, + 0.44 + ], + "angle": 0, + "content": "Using cross-attention, the module effectively reasons the relevance between entities and the textual feature. Inferred relevance is described in the \\( A^{\\mathrm{fuse}} \\), which is the probability of each entity being referred to in the textual feature." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.45, + 0.605, + 0.466 + ], + "angle": 0, + "content": "3.4. Training" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.47, + 0.89, + 0.501 + ], + "angle": 0, + "content": "Our model is trained with two different loss functions: contrastive cycle-consistency \\((\\mathbf{C}^3)\\) and reconstruction loss." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.513, + 0.829, + 0.529 + ], + "angle": 0, + "content": "3.4.1 Contrastive cycle-consistency \\((\\mathbf{C}^3)\\) Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.892, + 0.666 + ], + "angle": 0, + "content": "The output of our model is the cross-modal embedding, which is obtained by fusing information from visual and textual modalities. The fusion should selectively combine discovered entities based on textual information, but a segmentation label for which entities are referred to is not given in our weakly-supervised setting. To resolve this issue, we add an additional constraint to the contrastive learning objective: establishing cycle-consistency between each ground-truth pair of image and text." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.666, + 0.892, + 0.774 + ], + "angle": 0, + "content": "Let \\(\\{(\\mathbf{x}_i^\\nu, \\mathbf{x}_i^\\mathcal{T})\\}_{i=1}^B\\) represent a batch of visual and textual features, and \\(\\mathbf{z}_{ij} = \\Psi \\big(\\Phi(\\mathbf{x}_i^\\nu); \\mathbf{x}_j^\\mathcal{T}\\big)\\) denote the cross-modal embedding produced with \\(\\mathbf{x}_i^\\nu\\) and \\(\\mathbf{x}_j^\\mathcal{T}\\). For simplicity, we refer to the cross-modal embedding produced using the ground-truth image-text pair as matching embedding, and all other cross-modal embeddings as non-matching embeddings. Then \\(\\mathbf{C}^3\\) loss is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.786, + 0.892, + 0.829 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C} ^ {3}} = - \\frac {1}{B} \\sum_ {j = 1} ^ {B} \\log \\frac {e ^ {\\langle \\mathbf {z} _ {j j} , \\mathbf {s g} (\\mathbf {x} _ {j} ^ {\\mathcal {T}}) \\rangle}}{\\sum_ {i = 1} ^ {B} e ^ {\\langle \\mathbf {z} _ {i j} , \\mathbf {s g} (\\mathbf {x} _ {j} ^ {\\mathcal {T}}) \\rangle}} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\langle \\cdot, \\cdot \\rangle\\) and \\(\\mathfrak{sg}(\\cdot)\\) denote inner product and stop-gradient operation, respectively. Stop-gradient operation is used to prevent the collapsing solutions where every embedding becomes similar, as discussed in [9]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15551" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "The proposed loss function encourages the matching embeddings to be similar to the input textual feature, while making the non-matching embeddings to be dissimilar. This approach ensures that the modality fusion module preserves the inherent information of the referring expression only when it is paired with the matching image (cycle-consistency), achieved by attending to the referred entities that may be absent in non-matching images. Therefore, an attention map \\( A^{\\mathrm{fuse}} \\) is trained to reflect the referring expression accurately, where relevance between the expression and entities automatically emerges." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.269, + 0.274, + 0.282 + ], + "angle": 0, + "content": "3.4.2 Reconstruction Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.285, + 0.471, + 0.346 + ], + "angle": 0, + "content": "If we can reconstruct the original data using only the set of slots, then the slots will correspond to distinctive individual entities in the input image. Therefore, we leverage the reconstruction objective \\(\\mathcal{L}_{\\mathrm{recon}}\\), which is given by" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.353, + 0.47, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n}} = \\frac {1}{B} \\sum_ {i = 0} ^ {B} \\left\\| \\operatorname {s g} \\left(\\mathbf {x} _ {i} ^ {\\mathcal {Y}}\\right) - f _ {\\mathrm {d e c}} \\left(\\Phi \\left(\\mathbf {x} _ {i} ^ {\\mathcal {Y}}\\right)\\right) \\right\\| ^ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.47, + 0.492 + ], + "angle": 0, + "content": "where \\( f_{\\mathrm{dec}}(\\cdot) \\) is a MLP based decoder. Unlike the original slot attention [42], our reconstruction is performed in the feature space, rather than the pixel space. This feature reconstruction scheme enables an efficient decoder design. Detailed descriptions are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.5, + 0.188, + 0.513 + ], + "angle": 0, + "content": "3.5. Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.47, + 0.701 + ], + "angle": 0, + "content": "The segmentation mask is computed with the two attention maps: \\( A^{\\mathrm{slot}} \\in \\mathbb{R}^{N \\times K} \\) and \\( A^{\\mathrm{fuse}} \\in \\mathbb{R}^{K} \\). \\( A^{\\mathrm{slot}} \\) is patchwise attention map that contains information about the discovered entities (Eq. (3)), while \\( A^{\\mathrm{fuse}} \\) is entity-wise attention map denoting relevance between the referring expression and the entities (Eq. (7)). To compute the segmentation mask, we first compute \\( A^{\\mathrm{slot}} \\cdot A^{\\mathrm{fuse}} \\), which results in a 1-dimensional vector of size \\( N \\). We then reshape this vector into a 2-dimensional map. Next, we bilinearly interpolate the map to the original image size and rescale it using min-max normalization. Finally, we binarize the map using a threshold value \\( \\tau \\) to obtain the segmentation mask." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.21, + 0.725 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.732, + 0.279, + 0.748 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Datasets: Our experiments utilize four datasets commonly used in referring image segmentation task: RefCOCO [70], RefCOCO+ [70], Gref [45], and PhraseCut [61]. RefCOCO, RefCOCO+, and Gref are obtained from COCO [39] dataset. RefCOCO contains 142,209 image-text pairs, while RefCOCO+ has 141,564 pairs and does not include expressions indicating orientation properties (e.g., top, right). Gref includes 104,560 pairs for 54,822 objects. PhraseCut comprises 345,486 image-text pairs, which are derived from Visual Genome [35]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.333 + ], + "angle": 0, + "content": "Implementation details: For visual and textual backbone, we employ ViT-S-16 pretrained on ImageNet-21K [11] and pretrained BERT [13], respectively. We make each modality feature have the same channel dimension \\(D\\) by applying a \\(1 \\times 1\\) convolution layer. We set \\(D\\) and \\(D_h\\) to 512 and 1024, respectively. In the aggregation block, the number of iteration \\(T\\) is set to 6, and the number of slots \\(K\\) is set to 36 except for PhraseCut, which is set to 60. For the entity slot, the number of slots sampled from each Gaussian distributions \\(K_s\\) is set to 3 in RefCOCO+ and Gref, while 2 in RefCOCO and PhraseCut. During training, we use AdamW [44] optimizer, with the initial learning rate of 1e-4 scheduled with cosine annealing [43]. Batch size is set to 32 and model is trained for 50 epochs. We resize input images to \\(384 \\times 384\\) for both training and evaluation. Threshold value \\(\\tau\\) is set to 0.5 for all experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Evaluation metric: Following [67], we employ the cumulative intersection-over-union (cIoU) and mean intersection-over-Union (mIoU) metric; cIoU divides the total intersections by the total unions across all test images, and mIoU computes IoU for each image and averages them. Also, we evaluate the accuracy at the IoU thresholds of \\(\\{0.3, 0.5, 0.7\\}\\), denote as \\(\\mathcal{A} @ \\{0.3, 0.5, 0.7\\}\\) in Table 3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.456, + 0.828, + 0.472 + ], + "angle": 0, + "content": "4.2. Comparisons with the State of the Art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.892, + 0.581 + ], + "angle": 0, + "content": "Weakly supervised methods: As presented in Table 1, our method outperforms TSEG [53], the previous work on weakly-supervised referring image segmentation, in all benchmarks. It shows that the proposed bottom-up and top-down attention framework is more effective than using patch-based visual features with a carefully designed similarity function of TSEG." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Open-vocabulary segmentation: In Table 1, we also compare our model with the recent open-sourced open-vocabulary segmentation models: GroupViT [64] and MaskCLIP [73], on the referring image segmentation benchmarks. Note that these methods are pre-trained with large-scale image-text datasets [5, 47, 54]. We also report their fine-tuned version for fair comparisons: models are additionally trained with the image-text pairs of each benchmark. Further details for the reproductions are provided in the supplementary material. Ours outperforms GroupViT and MaskCLIP in all settings and benchmarks, even without large-scale image-text pretraining. This result shows that open-vocabulary segmentation models [64, 73] have difficulty comprehending the complex free-from text involving the relation between entities, unlike ours. Interestingly, our model without fine-tuning, which pre-trained on the PhraseCut dataset and evaluated on other benchmarks, still achieved better results. This setting is identical to the open-vocabulary segmentation models, except it leverages a far smaller dataset. These results demonstrate the feasibility of our method in a zero-shot transfer setting." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15552" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.089, + 0.85, + 0.234 + ], + "angle": 0, + "content": "
MethodsImage-text pretraining datasetFine tuningRefCOCORefCOCO+Gref valPhraseCut val
valtestAtestBvaltestAtestB
GroupViT [64]CC12M+YFCC (26M)X7.996.1610.518.496.7910.5910.685.82
✓10.8211.1111.2911.1410.7811.8412.779.41
MaskCLIP [73]WIT (400M)X11.5211.8512.0611.8712.0112.5712.7412.13
✓19.4518.6921.3719.9718.9321.4821.1123.80
TSEG [53]-✓25.44--22.01--22.0528.77
OursVisual Genome (0.31M)X21.8019.0024.9622.2019.8624.8525.89-
-✓34.7634.5835.0128.4828.6027.9828.8733.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.237, + 0.895, + 0.285 + ], + "angle": 0, + "content": "Table 1: Comparison with weakly-supervised method (TSEG) and open-vocabulary segmentation method (GroupViT and MaskCLIP). The results on four datasets are reported in mIoU (\\%). Fine-tuning \\(\\checkmark\\)means that the model is trained with the image-text pairs of the target benchmark; otherwise, it just relies on the pre-trained weight without further training." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.291, + 0.891, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.412, + 0.333, + 0.426 + ], + "angle": 0, + "content": "Query: \"Guy on left with beard and guitar\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.428, + 0.893, + 0.475 + ], + "angle": 0, + "content": "Figure 4: Qualitative results of our framework with entity slot (Ours), random slot, and query slot on RefCOCO val set. In the case of entity slots, the boundary color indicates the Gaussian distribution that each slot is sampled from. For each slot type, we present 10 discovered entities from \\( A^{\\mathrm{slot}} \\) and the final predictions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.493, + 0.232, + 0.51 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.471, + 0.558 + ], + "angle": 0, + "content": "We perform ablation studies on the val set of RefCOCO to analyze the contributions of the entity discovery module and the modality fusion module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.47, + 0.698 + ], + "angle": 0, + "content": "Entity discovery module: In Table 2(a), we ablate the entity discovery module and use visual feature directly, which leads to the drastical performance drop. This result demonstrates that contribution of our bottom-up attention framework is significant. Moreover, we ablate the interaction block, replace slot-attention of aggregation block to transformer, and train the model without reconstruction loss. The results show that all of the design choices for the module is important for effective entity discovery." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.47, + 0.807 + ], + "angle": 0, + "content": "Modality fusion module: In Table 2(b), the modality fusion module is ablated, and the model is trained with image-text contrastive learning [47], where the image feature is obtained by pooling the entity embeddings. The result demonstrates that inferring relevance between entities and referring expression with modality fusion module, trained with \\( \\mathrm{C}^3 \\) loss, is crucial for accurate segmentation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.821, + 0.251, + 0.838 + ], + "angle": 0, + "content": "4.4. In-depth Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Analysis of slot initialization: To verify our design choice for entity slots, we investigate variants of slots; query, random, and entity slot. As summarized in Table 3, our entity slot outperforms other slot types across the different" + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.491, + 0.876, + 0.658 + ], + "angle": 0, + "content": "
cIoUmIoU
(a) Analysis on entity discovery module
Ours30.4034.76
w/o entity discovery module5.405.91
w/o interaction block29.5933.69
aggregation block w/ transformer11.2912.61
w/o reconstruction loss18.6121.84
(b) Analysis on modality fusion module
Ours30.4034.76
w/o modality fusion module and C3loss13.8114.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.661, + 0.892, + 0.691 + ], + "angle": 0, + "content": "Table 2: Ablation studies of model components and loss function on RefCOCO val set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.894, + 0.903 + ], + "angle": 0, + "content": "numbers of slots. Specifically, as the number of slots is increased, the performance of the query slot declines, and that of the random slot converges. This result implies that simply increasing the number of slots can not handle the discovery of fine-grained visual entities in real-world images, and thus gains from the proposed entity slot type are non-trivial. In Fig. 4, we also present several visual entities discovered using each slot type, which show the superiority of the entity slot against others as explained in Sec. 3.2. Specifically, the entity slots from the same distribution share the same semantic property (e.g., head, body, or window) while capturing different fine-grained entities. The pro" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15553" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.09, + 0.147, + 0.1 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.091, + 0.35, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.358, + 0.091, + 0.484, + 0.19 + ], + "angle": 0, + "content": "Prediction Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.091, + 0.755, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.764, + 0.091, + 0.89, + 0.1 + ], + "angle": 0, + "content": "Prediction Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.101, + 0.149, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.191, + 0.184, + 0.199 + ], + "angle": 0, + "content": "on left\"" + }, + { + "type": "image", + "bbox": [ + 0.15, + 0.208, + 0.216, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.209, + 0.283, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.209, + 0.35, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.209, + 0.417, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.209, + 0.485, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.209, + 0.553, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.209, + 0.61, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.611, + 0.209, + 0.688, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.209, + 0.755, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.209, + 0.822, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.823, + 0.209, + 0.89, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.248, + 0.235, + 0.259 + ], + "angle": 0, + "content": "Query: \"Right person glasses\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.248, + 0.605, + 0.259 + ], + "angle": 0, + "content": "Query: \"Jeans on right\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.264, + 0.893, + 0.294 + ], + "angle": 0, + "content": "Figure 5: Qualitative results of our framework on RefCOCO val set. We present the discovered entities from \\( A^{\\mathrm{slot}} \\) and their relevance scores from \\( A^{\\mathrm{slot}} \\). Top-3 entities in terms of relevance to query expression are presented." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.307, + 0.458, + 0.45 + ], + "angle": 0, + "content": "
Slot type# slotsA@0.3A@0.5A@0.7cIoUmIoU
Query1227.577.170.9622.0421.44
2436.7310.611.9323.5924.78
3619.303.480.3116.6618.16
Random1236.3011.272.6924.2026.65
2445.1916.013.6626.6029.87
3645.6216.883.7626.5530.15
Entity(Ours)1239.4313.873.5624.7227.84
2452.3322.976.2329.2233.54
3655.0224.996.3530.4034.76
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.453, + 0.468, + 0.482 + ], + "angle": 0, + "content": "Table 3: Comparison between different slot types on Ref-COCO val set, where # slots is the total number of slots." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.469, + 0.579 + ], + "angle": 0, + "content": "posed entity slots effectively discover fine-grained visual entities, which results in an accurate prediction. In contrast, the prediction using random slots leads to a noisy mask, and using query slots leads to the mask of every head that appears in the image, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.469, + 0.748 + ], + "angle": 0, + "content": "Effect of hyperparameters \\( K_{g} \\) and \\( K_{s} \\): In Table 4(a), we investigate the impact of hyperparameters for the entity slots: \\( K_{g} \\) and \\( K_{s} \\), which are the number of Gaussian distributions and slots sampled from each distribution, respectively. We keep the total number of entity slots to 36, only changing \\( K_{g} \\) and \\( K_{s} \\). The optimal configurations are 18 and 2 for \\( K_{g} \\) and \\( K_{s} \\), respectively. Interestingly, we observed that even entity slots with \\( K_{s} = 1 \\), where a single slot is sampled from each distribution, vastly surpass the query slots in Table 3; it shows that random sampling of slots prevents the bounding to specific semantic category." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of \\( A^{\\mathrm{fuse}} \\): To verify the efficacy of \\( A^{\\mathrm{fuse}} \\) quantitatively, we evaluate our model with the following inference scheme variants: (1) AVG that replaces \\( A^{\\mathrm{fuse}} \\) with the uniform probability distribution, (2) MAX that picks a single attention map from \\( A^{\\mathrm{slot}} \\) with the maximum probability in \\( A^{\\mathrm{fuse}} \\), and (3) MIN that picks the attention map with the minimum probability. As summarized in Table 4(b), MAX largely outperforms AVG even though it only uses a single attention map for the final prediction, while MIN results in a near-zero cIoU and mIoU. This tendency shows that \\( A^{\\mathrm{fuse}} \\)" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.307, + 0.66, + 0.389 + ], + "angle": 0, + "content": "
KgKscIoUmIoU
36129.3633.41
18230.4034.76
12329.6334.03
9428.3132.34
6627.2331.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.573, + 0.392, + 0.591, + 0.404 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "table", + "bbox": [ + 0.669, + 0.309, + 0.88, + 0.389 + ], + "angle": 0, + "content": "
Inference schemecIoUmIoU
AVG9.018.16
MAX14.9319.54
MIN0.160.14
A fuse · A slot (Ours)30.4034.76
" + }, + { + "type": "table_caption", + "bbox": [ + 0.768, + 0.39, + 0.787, + 0.402 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.408, + 0.891, + 0.437 + ], + "angle": 0, + "content": "Table 4: Impact of hyperparameters \\( K_{g} \\) and \\( K_{s} \\) (a) and inference schemes (b) on RefCOCO val set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.498 + ], + "angle": 0, + "content": "effectively assigns probability based on the relevance between reference and entities, even without direct supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.892, + 0.591 + ], + "angle": 0, + "content": "Qualitative Analysis: In Fig. 5, predicted segmentation mask and discovered entities with top-3 relevance are presented. Results show that the entity discovery module effectively discovers primitive units of segmentation masks, and the modality fusion module accurately infers their relevance to query referring expression." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.602, + 0.619, + 0.617 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.892, + 0.818 + ], + "angle": 0, + "content": "In this paper, we address the problem of learning referring image segmentation with only image-text pairs. We present a new model based on a bottom-up and top-down attention framework, which first discovers entities in the input image and then combines such entities based on their relevance to the text query. Moreover, we present a new loss function enforcing cycle-consistency between image-text pairs, which enables the effective training of the model without any further supervision. As a result, the proposed method clearly outperforms an existing weakly-supervised method and recent open-vocabulary segmentation models. For future work, we will explore expanding our model to different modalities such as video and audio." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the NRF grant and the IITP grants funded by Ministry of Science and ICT, Korea (NRF-2021R1A2C3012728-20%, IITP-2020-0-00842-20%, IITP-2022-0-00290-20%, IITP-2022-0-00926-20%, IITP-2021-0-02068-10%, IITP-2019-0-01906-10%)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15554" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Jiwoon Ahn, Sunghyun Cho, and Suha Kwak. Weakly supervised learning of instance segmentation with inter-pixel relations. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.17, + 0.47, + 0.224 + ], + "angle": 0, + "content": "[2] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.224, + 0.47, + 0.25 + ], + "angle": 0, + "content": "[3] Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. volume abs/1607.06450, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.25, + 0.47, + 0.291 + ], + "angle": 0, + "content": "[4] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In Proc. Neural Information Processing Systems (NeurIPS), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.291, + 0.47, + 0.358 + ], + "angle": 0, + "content": "[5] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[6] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In Proc. IEEE International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.413, + 0.47, + 0.467 + ], + "angle": 0, + "content": "[7] Jianbo Chen, Yelong Shen, Jianfeng Gao, Jingjing Liu, and Xiaodong Liu. Language-based image editing with recurrent attentive models. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.47, + 0.522 + ], + "angle": 0, + "content": "[8] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In Proc. Neural Information Processing Systems (NeurIPS), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.47, + 0.561 + ], + "angle": 0, + "content": "[9] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.47, + 0.629 + ], + "angle": 0, + "content": "[10] Samyak Datta, Karan Sikka, Anirban Roy, Karuna Ahuja, Devi Parikh, and Ajay Divakaran. Align2ground: Weakly supervised phrase grounding guided by image-caption alignment. In Proc. IEEE International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.47, + 0.683 + ], + "angle": 0, + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: a large-scale hierarchical image database. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.684, + 0.47, + 0.737 + ], + "angle": 0, + "content": "[12] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers. In Proc. IEEE International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.738, + 0.47, + 0.805 + ], + "angle": 0, + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Annual Conference of the North American Chapter of the Association for Computational Linguistics, 2019. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.806, + 0.47, + 0.859 + ], + "angle": 0, + "content": "[14] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proc. IEEE International Conference on Computer Vision (ICCV), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "vain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In Proc. International Conference on Learning Representations (ICLR), 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.134, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[16] Guang Feng, Zhiwei Hu, Lihe Zhang, and Hutchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.188, + 0.892, + 0.255 + ], + "angle": 0, + "content": "[17] Tsu-Jui Fu, Xin Eric Wang, Scott T. Grafton, Miguel P. Eckstein, and William Yang Wang. M3l: Language-based video editing via multi-modal multi-level transformers. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.255, + 0.892, + 0.309 + ], + "angle": 0, + "content": "[18] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In Proc. European Conference on Computer Vision (ECCV), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.31, + 0.892, + 0.364 + ], + "angle": 0, + "content": "[19] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.364, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[20] Tanmay Gupta, Arash Vahdat, Gal Chechik, Xiaodong Yang, Jan Kautz, and Derek Hoiem. Contrastive learning for weakly supervised phrase grounding. In European Conference on Computer Vision. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[21] S. Hong, D. Yeo, S. Kwak, H. Lee, and B. Han. Weakly supervised semantic segmentation using web-crawled videos. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2224-2232, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.472, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[22] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In Proc. European Conference on Computer Vision (ECCV), 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[23] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.594, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[24] Binbin Huang, Dongze Lian, Weixin Luo, and Shenghua Gao. Look before you leap: Learning landmark features for one-stage visual grounding. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.649, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[25] Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoqi Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.892, + 0.783 + ], + "angle": 0, + "content": "[26] Zilong Huang, Xinggang Wang, Jiasi Wang, Wenyu Liu, and Jingdong Wang. Weakly-supervised semantic segmentation network with deep seeded region growing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.784, + 0.892, + 0.838 + ], + "angle": 0, + "content": "[27] Tianrui Hui, Si Liu, Shaofei Huang, Guanbin Li, Sansi Yu, Faxi Zhang, and Jizhong Han. Linguistic structure guided context modeling for referring image segmentation. In Proc. European Conference on Computer Vision (ECCV), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.838, + 0.892, + 0.893 + ], + "angle": 0, + "content": "[28] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proc. International" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15555" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.112, + 0.093, + 0.429, + 0.106 + ], + "angle": 0, + "content": "Conference on Machine Learning (ICML), 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.106, + 0.47, + 0.173 + ], + "angle": 0, + "content": "[29] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proc. IEEE International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.173, + 0.469, + 0.228 + ], + "angle": 0, + "content": "[30] A. Khoreva, R. Benenson, J. Hosang, M. Hein, and B. Schiele. Simple does it: Weakly supervised instance and semantic segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.228, + 0.469, + 0.309 + ], + "angle": 0, + "content": "[31] Beomyoung Kim, Youngjoon Yoo, Chae Eun Rhee, and Junmo Kim. Beyond semantic to instance segmentation: Weakly-supervised instance segmentation via semantic knowledge transfer and self-refinement. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.31, + 0.469, + 0.349 + ], + "angle": 0, + "content": "[32] Dongwon Kim, Namyup Kim, and Suha Kwak. Improving cross-modal retrieval with set of diverse embeddings. 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[33] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Restr: Convolution-free referring image segmentation using transformers. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.469, + 0.471 + ], + "angle": 0, + "content": "[34] Alexander Kolesnikov and Christoph H. Lampert. Seed, expand and constrain: Three principles for weakly-supervised image segmentation. In Proc. European Conference on Computer Vision (ECCV), 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.471, + 0.469, + 0.554 + ], + "angle": 0, + "content": "[35] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, Michael Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision (IJCV), 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.554, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[36] Suha Kwak, Seunghoon Hong, and Bohyung Han. Weakly supervised semantic segmentation using superpixel pooling network. In Proc. AAAI Conference on Artificial Intelligence (AAAI), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.662 + ], + "angle": 0, + "content": "[37] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In Proc. International Conference on Learning Representations (ICLR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.662, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[38] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.729, + 0.469, + 0.797 + ], + "angle": 0, + "content": "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: common objects in context. In Proc. European Conference on Computer Vision (ECCV), 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.798, + 0.469, + 0.852 + ], + "angle": 0, + "content": "[40] Chenxi Liu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, and Alan Yuille. Recurrent multimodal interaction for referring image segmentation. In Proc. IEEE International Conference on Computer Vision (ICCV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.852, + 0.469, + 0.893 + ], + "angle": 0, + "content": "[41] Yongfei Liu, Bo Wan, Lin Ma, and Xuming He. Relation-aware instance refinement for weakly supervised visual grounding. In Proc. IEEE Conference on Computer Vision" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.534, + 0.093, + 0.785, + 0.106 + ], + "angle": 0, + "content": "and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.106, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[42] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. In Proc. Neural Information Processing Systems (NeurIPS), 2020. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.174, + 0.892, + 0.214 + ], + "angle": 0, + "content": "[43] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In Proc. International Conference on Learning Representations (ICLR), 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.214, + 0.892, + 0.254 + ], + "angle": 0, + "content": "[44] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. Proc. International Conference on Learning Representations (ICLR), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.254, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[45] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.322, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[46] Edgar Margffoy-Tuay, Juan C Pérez, Emilio Botero, and Pablo Arbeláez. Dynamic multimodal instance segmentation guided by natural language queries. In Proc. European Conference on Computer Vision (ECCV), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.376, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. International Conference on Machine Learning (ICML), 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.459, + 0.892, + 0.499 + ], + "angle": 0, + "content": "[48] Tal Remez, Jonathan Huang, and Matthew Brown. Learning to segment via cut-and-paste. In Proc. European Conference on Computer Vision (ECCV), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.499, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[49] Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, and Bernt Schiele. Grounding of textual phrases in images by reconstruction. In Proc. European Conference on Computer Vision (ECCV), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[50] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.621, + 0.892, + 0.675 + ], + "angle": 0, + "content": "[51] Hengcan Shi, Hongliang Li, Fanman Meng, and Qingbo Wu. Key-word-aware network for referring expression image segmentation. In Proc. European Conference on Computer Vision (ECCV), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[52] Gautam Singh, Fei Deng, and Sungjin Ahn. Illiterate dalle learns to compose. In Proc. International Conference on Learning Representations (ICLR), 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.715, + 0.892, + 0.756 + ], + "angle": 0, + "content": "[53] Robin Strudel, Ivan Laptev, and Cordelia Schmid. Weakly-supervised segmentation of referring expressions. arXiv preprint arXiv:2205.04725, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.756, + 0.892, + 0.809 + ], + "angle": 0, + "content": "[54] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.809, + 0.892, + 0.864 + ], + "angle": 0, + "content": "[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. Neural Information Processing Systems (NeurIPS), 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.864, + 0.892, + 0.891 + ], + "angle": 0, + "content": "[56] Liwei Wang, Jing Huang, Yin Li, Kun Xu, Zhengyuan Yang, and Dong Yu. Improving weakly supervised visual ground" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15556" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "ing by contrastive knowledge distillation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.215 + ], + "angle": 0, + "content": "[57] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.215, + 0.469, + 0.27 + ], + "angle": 0, + "content": "[58] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.27, + 0.469, + 0.338 + ], + "angle": 0, + "content": "[59] Yunchao Wei, Jiashi Feng, Xiaodan Liang, Ming-Ming Cheng, Yao Zhao, and Shuicheng Yan. Object region mining with adversarial erasing: A simple classification to semantic segmentation approach. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.338, + 0.469, + 0.392 + ], + "angle": 0, + "content": "[60] Philippe Weinzaepfel, Thomas Lucas, Diane Larlus, and Yannis Kalantidis. Learning super-features for image retrieval. In Proc. International Conference on Learning Representations (ICLR), 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[61] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasecut: Language-based image segmentation in the wild. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.469, + 0.514 + ], + "angle": 0, + "content": "[62] Yongqin Xian, Subhabrata Choudhury, Yang He, Bernt Schiele, and Zeynep Akata. Semantic projection network for zero-and few-label semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8256-8265, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.514, + 0.469, + 0.568 + ], + "angle": 0, + "content": "[63] Fanyi Xiao, Leonid Sigal, and Yong Jae Lee. Weakly-supervised visual grounding of phrases with linguistic structures. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.469, + 0.636 + ], + "angle": 0, + "content": "[64] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[65] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310-4319, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[66] Li Yang, Yan Xu, Chunfeng Yuan, Wei Liu, Bing Li, and Weiming Hu. Improving visual grounding with visual-linguistic verification and iterative reasoning. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.771, + 0.469, + 0.84 + ], + "angle": 0, + "content": "[67] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.84, + 0.469, + 0.894 + ], + "angle": 0, + "content": "[68] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.894 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[69] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.161, + 0.892, + 0.215 + ], + "angle": 0, + "content": "[70] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Proc. European Conference on Computer Vision (ECCV), 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.215, + 0.892, + 0.256 + ], + "angle": 0, + "content": "[71] Dong Zhang, Hanwang Zhang, Jinhui Tang, Xian-Sheng Hua, and Qianru Sun. Causal intervention for weakly-supervised semantic segmentation. 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.256, + 0.892, + 0.31 + ], + "angle": 0, + "content": "[72] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba. Learning deep features for discriminative localization. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.31, + 0.892, + 0.35 + ], + "angle": 0, + "content": "[73] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In Proc. European Conference on Computer Vision (ECCV), 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.35, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[74] Yi Zhou, Hui Zhang, Hana Lee, Shuyang Sun, Pingjun Li, Yangguang Zhu, ByungIn Yoo, Xiaojuan Qi, and Jae-Joon Han. Slot-vps: Object-centric representation learning for video panoptic segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.431, + 0.892, + 0.486 + ], + "angle": 0, + "content": "[75] Yanzhao Zhou, Yi Zhu, Qixiang Ye, Qiang Qiu, and Jianbin Jiao. Weakly supervised instance segmentation using class peak response. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.486, + 0.892, + 0.554 + ], + "angle": 0, + "content": "[76] Yuliang Zou, Zizhao Zhang, Han Zhang, Chun-Liang Li, Xiao Bian, Jia-Bin Huang, and Tomas Pfister. Pseudoseg: Designing pseudo labels for semantic segmentation. In Proc. International Conference on Learning Representations (ICLR), 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15557" + } + ] +] \ No newline at end of file diff --git a/2023/Shatter and Gather_ Learning Referring Image Segmentation with Text Supervision/eba869e5-3a94-47e8-afcc-2f1d3dc46bfb_origin.pdf b/2023/Shatter and Gather_ Learning Referring Image Segmentation with Text Supervision/eba869e5-3a94-47e8-afcc-2f1d3dc46bfb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d3a00e0327276917ac1f639317b41e9f3a5df614 --- /dev/null +++ b/2023/Shatter and Gather_ Learning Referring Image Segmentation with Text Supervision/eba869e5-3a94-47e8-afcc-2f1d3dc46bfb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dcb289fc94f7f7397e0bfd8dd26ee0c8b73996d69d58a3c77819bbc9e6456236 +size 2759646 diff --git a/2023/Shatter and Gather_ Learning Referring Image Segmentation with Text Supervision/full.md b/2023/Shatter and Gather_ Learning Referring Image Segmentation with Text Supervision/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4f9821e1182082ed23b9b95e9f5e66dfb3946871 --- /dev/null +++ b/2023/Shatter and Gather_ Learning Referring Image Segmentation with Text Supervision/full.md @@ -0,0 +1,365 @@ +# Shatter and Gather: Learning Referring Image Segmentation with Text Supervision + +Dongwon Kim $^{1*}$ Namyup Kim $^{1*}$ Cuiling Lan $^{2}$ Suha Kwak $^{1}$ $^{1}$ POSTECH Microsoft Research Asia +http://cvlab.postech.ac.kr/research/sag + +# Abstract + +Referring image segmentation, the task of segmenting any arbitrary entities described in free-form texts, opens up a variety of vision applications. However, manual labeling of training data for this task is prohibitively costly, leading to lack of labeled data for training. We address this issue by a weakly supervised learning approach using text descriptions of training images as the only source of supervision. To this end, we first present a new model that discovers semantic entities in input image and then combines such entities relevant to text query to predict the mask of the referent. We also present a new loss function that allows the model to be trained without any further supervision. Our method was evaluated on four public benchmarks for referring image segmentation, where it clearly outperformed the existing method for the same task and recent open-vocabulary segmentation models on all the benchmarks. + +# 1. Introduction + +Referring image segmentation is the task of segmenting the referent corresponding to a natural language expression given as a query. Unlike the conventional semantic segmentation that aims at segmenting a pre-defined set of classes, referring image segmentation enables segmentation of any arbitrary entities described in free-form texts and thus opens up a wide variety of applications such as human-computer interaction [8, 57] and text-based image editing [7, 17]. + +Thanks to the development of deep neural networks, recent studies have demonstrated remarkable results on referring image segmentation in the supervised learning setting [14, 22, 25, 27, 67, 68, 69]. However, obtaining manual annotation of training data for referring image segmentation is prohibitively costly since the data requires two types of labels for each image, i.e., natural language expressions describing the entities that appear in the image and segmentation labels corresponding to them. This typically demands tremendously more manpower than annotation for semantic + +![](images/dd9ca38fcbe60fa581656fa48f2190717303716cf506d35dc6a41acacf2ed70b.jpg) +Figure 1: Our model consists of two different attention modules: bottom-up attention that identifies individual entities in an image, and top-down attention that combines the entities based on a referring expression. It is trained in a contrastive manner to ensure consistency between matching image-text pairs. In inference, a segmentation mask is predicted by combining the entities found by bottom-up attention with weights derived by top-down attention. + +segmentation, leading to lack of labeled data for training. + +A solution to this issue is weakly supervised learning, which trains a model on a dataset providing weaker forms of supervision than the conventional ones. In this paper, we particularly focus on learning referring image segmentation only with natural language expressions describing the associated images. Since these labels are readily available in many vision-language datasets [28, 47], the issue on lack of labeled training data can be alleviated by our approach. However, learning referring image segmentation in this setting introduces another challenge: Since natural language expressions often contain relations between visual entities such as instances and object parts (e.g., "The cat next to the + +table leg"), a model to be trained should be aware of the inter-object relations to exploit the expressions as supervision. Thereby we need a method to discover individual entities and infer their relations without any supervision but the referring expression. + +To fulfill this requirement, we introduce a new bottom-up and top-down attention framework, which is illustrated in Fig. 1. Individual entities existing in an image are first discovered by bottom-up attention, which solely exploits visual information. For the bottom-up attention, an entity discovery module progressively refines a set of embedding vectors named slots to capture distinct visual entities, following slot attention [42]. We moreover propose a novel slot formulation named entity slot, which enables fine-grained entity discovery in real-world images. As a result, entity slots after the refinement capture individual visual entities without any segmentation supervision. These entities are considered primitive units for composing the segmentation mask of a referent during inference. + +Afterward, based on the relational and structural information of referring expression, top-down attention selectively attends to the referred entities and combines them into the predicted mask. Top-down attention is implemented by a modality fusion module consists of cross-attention transformer [15]. This attention scheme enables the module to infer the relevance between discovered entities and the referring expression instead of relying on the cosine similarity or heuristic similarity functions used in previous work [18, 53, 64, 73]. + +For training, we propose the contrastive cycle-consistency $(\mathrm{C}^3)$ loss that enforces cycle-consistency between image-text pairs under the contrastive learning framework [47]. To establish such consistency, the model is trained to preserve the textual information after the fusion of entities and textual features. By doing so, we observe that latent relevance between the discovered entities and referring expression automatically emerges in the top-down attention without any explicit mask supervision. + +One may wonder the difference between our task and open-vocabulary segmentation [18, 64, 73], i.e., semantic segmentation of arbitrary categories without explicit mask supervision. The key difference is that a referent to segment in referring image segmentation is given by a complex free-form text, which introduces additional challenges. + +The proposed method was evaluated and compared on four public benchmarks for referring image segmentation, where it substantially outperformed the previous weakly supervised learning method [53]. Moreover, we reproduced recent open-vocabulary segmentation models [64, 73] and evaluated them on referring image segmentation benchmarks; the results show that the proposed model surpassed them even without pre-training on large-scale image-text data [5, 47, 54]. + +In summary, our contribution is three-fold as follows: + +- We present a new weakly supervised learning model for referring image segmentation. Our model employs bottom-up and top-down attention to discover individual entities and infer their relations, which enables to exploit natural language expressions as supervision effectively during training. +- We propose a new loss function called contrastive cycle-consistency loss, which allows latent relevance between discovered entities and referring expressions to emerge without requiring further supervision. +- Our model clearly outperformed the existing method for the same task and recent open-vocabulary segmentation models on all the four benchmarks. + +# 2. Related Work + +Referring image segmentation: Early approaches [22, 38, 40, 51] obtain multimodal feature maps by simple concatenation-convolution operation. With this operation as a basis, most existing studies have suggested attention mechanisms and multi-level feature aggregations [6, 16, 23, 25, 46, 68]. Various approaches have been employed for calculating fine interactions between the modalities, including multimodal feature fusion [14, 29, 33] using transformers [15], a large-scale vision-language pretrained framework [58], and holistically multimodal fusion [67]. The existing referring image segmentation methods have been developed in a fully supervised learning setting. Recently, a weakly supervised learning method [53] using only imagedtext pairs for training is introduced to alleviate the high annotation cost. However, this study relies on a heuristic similarity function to infer relations between patches and texts, which suffers from comprehending the relational and structural information from image and text modalities. Therefore, we propose a new weakly supervised model consisting of bottom-up and top-down attention to discover visual entities and infer their relations effectively. + +Phrase grounding: Phrase grounding [12, 24, 29, 49, 66, 69] is the task of learning the association between image regions and words in caption describing the image. Weakly supervised approaches have also been proposed for this task [10, 20, 41, 56, 63], where the correspondence between words and regions are not given. It is worth noting that the contrastive learning framework introduced in [20] is closely related to the proposed $C^3$ loss by enforcing mutual information between regions and words. However, these studies rely on an off-the-shelf object detector [19], where the final prediction is restricted by the predefined set of classes with which the detector is trained. On the contrary, our method discovers visual entities in an image without an auxiliary model trained with external labeled data, which is more suitable for referring image segmentation. + +![](images/47b618ab2f05409361cba42c53cac8d18a2060131b5c6f2b3448f65025736f6f.jpg) +Figure 2: Illustration of the overall architecture of our model along with its behavior during training and inference. (a) Feature extractors: The visual and textual features are extracted by transformer-based encoders of the two modalities respectively. (b) Entity discovery module: A set of visual entities are discovered from the visual features through the bottom-up attention. (c) Modality fusion module: The visual entities and a referring expression are fused considering their relevance, which is estimated by the top-down attention. Our model is trained with the proposed contrastive cycle-consistency $(\mathbf{C}^3)$ loss between the cross-modal embeddings and textual features. In inference, the segmentation mask is predicted by jointly considering the visual entities $A^{\mathrm{slot}}$ from the entity discovery module and the relevance scores $A^{\mathrm{fuse}}$ from the modality fusion module. + +Weakly supervised learning for segmentation: Weakly supervised learning has become a favored solution for addressing the lack of annotated data in training, especially for semantic segmentation [1, 2, 21, 26, 34, 36, 65, 71, 76] and instance segmentation [30, 31, 48, 75]. These methods refine coarse localization cues obtained by CAMs [50, 59, 72], and use the results as pseudo labels for self-training. However, they cannot be directly applied to the referring image segmentation due to the following challenges. First, there are no predefined classes in this task, making it difficult to learn shared semantics within CAMs by multiple instance learning. Second, there is no mechanism to discover visual entities and combine them with textual information. + +Slot attention: Slot attention [42] is proposed for the learning object-centric representation. It can discover the objects within an image in an unsupervised manner and be used in the various downstream task such as segmentation [64, 74] and retrieval [32, 60]. In specific, slot attention introduces an iterative attention mechanism that operates on the slots, which are the set of embedding vectors corresponding to the objects that appear in the input data. In previous work, slots are initialized as a set of learnable embeddings [32, 60, 64, 74] or a set of vectors sampled from a learnable Gaussian distribution [42, 52]. However, we empirically observed that both types of initialization are inapplicable to visual entity discovery for referring image segmentation, which we will elaborate on in Sec. 3.2. + +Open-vocabulary segmentation: The open-vocabulary segmentation is proposed to address the limitation of semantic segmentation that it is constrained to the pre + +defined set of classes. Unlike semantic segmentation, open-vocabulary segmentation models are evaluated with classes unseen during training. The early approaches propose a generative model [4] and a projection model [62] to generate or project pixel-level semantic features from visual features for the unseen classes through word embedding space. Recent studies [18, 37, 64, 73] largely improve the quality of semantic features for the unseen classes, and further introduce the methods that do not require segmentation labels for seen classes using large-scale image-text pretrained models (e.g.CLIP [47], ALIGN [28]). However, since they are trained to predict masks for semantic categories, not for free-form texts describing arbitrary entities in an image, their applicability is still limited. We thus propose a new method for referring image segmentation that only requires image-text pairs during training. + +# 3. Proposed Method + +As shown in Fig. 2, our model consists of three main components: feature extractors, entity discovery module, and modality fusion module. The visual and textual features are extracted from the input using the backbone networks of the two modalities respectively. (Sec. 3.1). The visual features are then given as input to the entity discovery module, which aggregates the features into multiple entity embeddings using the bottom-up attention (Sec. 3.2). The entity embeddings correspond to distinct entities appearing in the image, which are considered as the smallest units of building the segmentation mask. The modality fusion module fuses the entity embeddings and the textual feature into + +cross-modal embeddings according to their relevance using the top-down attention (Sec. 3.3). The produced cross-modal embedding is used only for training, with a learning objective enforcing cycle-consistency and reconstruction of visual features (Sec. 3.4). The segmentation mask is predicted by combining the entities discovered by the entity discovery module with the relevance scores estimated by the modality fusion module (Sec. 3.5). + +# 3.1. Visual and Textual Feature Extraction + +Here, we provide the details of the transformer-based backbone network for each modality. + +Visual feature extraction: We employ the vision transformer (ViT) [15] as the visual backbone network. An input image is first split into $N$ patches and then linearly projected. The patches are processed by the multiple self-attention blocks of ViT, where we refer to the output tokens of the last block as the visual features $\mathbf{x}^{\mathcal{V}}\in \mathbb{R}^{N\times D}$ . + +Textual feature extraction: We consider the transformer-based text encoder, BERT [13]. A sentence is first tokenized into the token sequence, where the [CLS] and [SEP] tokens are added to the beginning and the end of the sequence. The activation of the last layer of the transformer for the [CLS] token is employed as the textual feature $\mathbf{x}^{\mathcal{T}}\in \mathbb{R}^{D}$ . + +# 3.2. Bottom-up Attention: Entity Discovery Module + +The entity discovery module applies a bottom-up attention mechanism to aggregate patch-based visual features into multiple entity embeddings that represent instances or object parts in the image. The discovered entities are then treated as primitive units of the segmentation mask, where the mask is predicted by selectively combining them. This approach is useful since the textual information is often built upon the relations between such visual entities. + +Specifically, the entity discovery module consists of two blocks: aggregation block and interaction block. The entity slots, which are embedding vectors that corresponds to various visual entities, are updated in an iterative process alternating between these two blocks. These slots eventually become entity embeddings. Let $\Phi : \mathbb{R}^{N \times D} \mapsto \mathbb{R}^{K \times D}$ be the entity discovery module that takes visual features as input and outputs the set of $K$ entity embeddings. The entity embeddings are obtained by iteratively updating initial entity slots $\mathbf{S}^0 \in \mathbb{R}^{K \times D}$ $T$ -times: + +$$ +\begin{array}{r l} & \Phi (\mathbf {x} ^ {\mathcal {V}}) := \mathbf {S} ^ {T}, \text {w h e r e} \\ & \quad \mathbf {S} ^ {t + 1} = \operatorname {I n t e r} \left(\operatorname {A g g} \left(\mathbf {x} ^ {\mathcal {V}}; \mathbf {S} ^ {t}\right)\right), \end{array} \tag {1} +$$ + +and Agg and Inter denote the aggregation and interaction blocks, respectively. + +In the literature, the initial slots $\mathbf{S}^0$ are often defined as either vectors sampled from a learnable Gaussian distribution [42, 52] or learnable embeddings [32, 60, 64, 74], + +![](images/656824a77cfeba5f159c67838a6b7b9b2b9eb64851781382dec522c4f14f99e1.jpg) +Figure 3: Comparison between the proposed entity slot and previous slot initialization methods. The boundary color indicates the Gaussian distribution that each slot is sampled from. The entity slot successfully discovers fine-grained entities such as head and food, while the others often fail to capture individual entities of specific semantics. + +where such types of slots are referred as random slot and query slot in the remainder of this paper, respectively. However, we have empirically found that both of these initialization strategies are unsuitable for our task as shown in Fig. 3. The random slot suffers from the lack of semantic specificity since every slot is randomly sampled from a single distribution. This can lead to suboptimal performance since entities in real-world images are diverse in terms of semantics. The query slot, on the other hand, suffers from the limitation of being bound to specific semantic categories, such as human, animal, and vehicle. This limitation leads to entity discovery that is not sufficiently fine-grained, e.g., given an image of multiple people, a single slot that binds to "human" captures all of them and thus failed to identify the individual specified by the text query. This is problematic particularly in our model since entities act as the smallest units for producing segmentation masks. It is highly desired to have a mechanism that could guide the differentiation of slots to be semantically specific. + +To overcome these limitations, we introduce a new initialization strategy for the entity slot, which enables the discovery of fine-grained visual entities while learning diverse semantics. Formally, the initial entity slots $\mathbf{S}^0$ are obtained by sampling $K_{s}$ slots from each of $K_{g}$ different Gaussian distributions, resulting in a total of $K = K_{s}\cdot K_{g}$ slots: + +$$ +\mathbf {S} ^ {0} := \left[ \mathbf {S} _ {1} ^ {0}; \dots ; \mathbf {S} _ {K _ {g}} ^ {0} \right] \in \mathbb {R} ^ {K \times D}, \text {w h e r e} \tag {2} +$$ + +$$ +\mathbf {S} _ {i} ^ {0} \sim \mathcal {N} (\mu_ {i}, \operatorname {d i a g} (\sigma_ {i})) \in \mathbb {R} ^ {K _ {s} \times D}. +$$ + +The idea is to initialize slots by sampling them from different Gaussian distributions, each specifying a distinct semantic property to be captured. Individual entities sharing the same semantics are differentiated by slots sampled from the same distribution, which enables fine-grained entity discovery. + +From now, we present the aggregation block and interaction block at $t$ -th iteration. As described in Eq. (1), the slots of previous iteration $\mathbf{S}^{t-1}$ are fed to the aggregation block with the visual features $\mathbf{x}^{\nu}$ . The role of the aggregation block is to let each slot aggregate visual features of distinct visual entities, and it is achieved by employing the competitive attention mechanism introduced in the slot attention [42]. Let $q(\cdot)$ , $k(\cdot)$ , and $v(\cdot)$ are linear projection with layer normalization [3] that map inputs to dimension $D_h$ . Then, the attention map $A^{\mathrm{slot}} \in \mathbb{R}^{N \times K}$ between $\mathbf{x}^{\nu}$ and $\mathbf{S}^{t-1}$ is computed as follows: + +$$ +A _ {i, j} ^ {\text {s l o t}} = \frac {e ^ {M _ {i , j}}}{\sum_ {j = 1} ^ {K} e ^ {M _ {i , j}}}, \text {w h e r e} M = \frac {k \left(\mathbf {x} ^ {\mathcal {V}}\right) q \left(\mathbf {S} ^ {t - 1}\right) ^ {\top}}{\sqrt {D _ {h}}}. \tag {3} +$$ + +Unlike the cross-attention transformer [55] that normalized attention over the keys, we normalize attention map $A^{\mathrm{slot}}$ over the slots. This normalization scheme encourages slots to compete for aggregating the visual features, which results in the set of slots binding to different visual entities in the input data. It is worth noting that $A^{\mathrm{slot}}$ provides the localizations of entities discovered, which is used as primitive units of the segmentation mask during inference. Using $A^{\mathrm{slot}}$ , the slots are updated in a similar manner to the transformer: + +$$ +\begin{array}{l} \widehat {\mathbf {S}} ^ {t} = \operatorname {A g g} \left(\mathbf {x} ^ {\nu}; \mathbf {S} ^ {t}\right) \tag {4} \\ = \operatorname {n o r m a l i z e} \left(A ^ {\text {s l o t}}\right) ^ {\top} v \left(\mathbf {x} ^ {\nu}\right) W _ {o} + \mathbf {S} ^ {t - 1}, \\ \end{array} +$$ + +where normalize $(\cdot)$ denotes $\ell_1$ -normalization of the column vectors and $W_{o}\in \mathbb{R}^{D_{h}\times D}$ is linear projection. Finally, slots are further refined by $\widehat{\mathbf{S}}^t\gets \mathrm{MLP}\big(\widehat{\mathbf{S}}^t\big) + \widehat{\mathbf{S}}^t$ where MLP is multi-layer perceptron (MLP) with layer normalization. + +The following interaction block consists of multiple self-attention transformers [55]. The weights of these transformers are shared, and each transformer is applied to slots sampled from the same distribution. We denote the transformer layer as Transformer(Q; K; V), where (Q; K; V) are for query, key, and value. The interaction block is denoted as follows: + +$$ +\operatorname {I n t e r} \left(\widehat {\mathbf {S}} ^ {t}\right) = \left[ \operatorname {S A} \left(\widehat {\mathbf {S}} _ {1} ^ {t}\right); \dots ; \operatorname {S A} \left(\widehat {\mathbf {S}} _ {K _ {g}} ^ {t}\right) \right], \tag {5} +$$ + +where $\mathrm{SA}(\mathbf{X})$ is for the Transformer(X; X; X). This self-attention scheme models information exchange between visual entities sharing the same semantic property, which helps slots within a single semantic category to capture different entities without redundancy. + +# 3.3. Top-down Attention: Modality Fusion Module + +The modality fusion module is proposed for top-down attention mechanism, which combines discovered visual entities based on the textual feature. Let $\Psi: \mathbb{R}^{K \times D} \times \mathbb{R}^D \mapsto \mathbb{R}^D$ be the modality fusion module, which takes entity embeddings $\Phi(\mathbf{x}^\nu)$ and textual furret $\mathbf{x}^T$ as inputs. Then, the output cross-modal embedding $\mathbf{z} \in \mathbb{R}^D$ is computed using cross-attention transformer: + +$$ +\mathbf {z} = \Psi \left(\Phi \left(\mathbf {x} ^ {\mathcal {V}}\right); \mathbf {x} ^ {\mathcal {T}}\right) = \frac {\operatorname {C A} \left(\Phi \left(\mathbf {x} ^ {\mathcal {V}}\right) ; \mathbf {x} ^ {\mathcal {T}}\right)}{\| \operatorname {C A} \left(\Phi \left(\mathbf {x} ^ {\mathcal {V}}\right) ; \mathbf {x} ^ {\mathcal {T}}\right) \| _ {2}}, \tag {6} +$$ + +where $\mathrm{CA}(\mathrm{X};\mathrm{Y})$ is for the Transformer(Y;X;X) and $\| \cdot \| _2$ denotes $\ell_2$ norm. Moreover, we can obtain the attention map $A^{\mathrm{fuse}}\in \mathbb{R}^{K}$ as a byproduct in cross-attention: + +$$ +A _ {i} ^ {\text {f u s e}} = \frac {e ^ {M _ {i}}}{\sum_ {i = 1} ^ {K} e ^ {M _ {i}}} \text {, w h e r e} M = \frac {k \left(\Phi \left(\mathbf {x} ^ {\mathcal {V}}\right)\right) q \left(\mathbf {x} ^ {\mathcal {T}}\right)}{\sqrt {D _ {h}}}. \tag {7} +$$ + +Using cross-attention, the module effectively reasons the relevance between entities and the textual feature. Inferred relevance is described in the $A^{\mathrm{fuse}}$ , which is the probability of each entity being referred to in the textual feature. + +# 3.4. Training + +Our model is trained with two different loss functions: contrastive cycle-consistency $(\mathbf{C}^3)$ and reconstruction loss. + +# 3.4.1 Contrastive cycle-consistency $(\mathbf{C}^3)$ Loss + +The output of our model is the cross-modal embedding, which is obtained by fusing information from visual and textual modalities. The fusion should selectively combine discovered entities based on textual information, but a segmentation label for which entities are referred to is not given in our weakly-supervised setting. To resolve this issue, we add an additional constraint to the contrastive learning objective: establishing cycle-consistency between each ground-truth pair of image and text. + +Let $\{(\mathbf{x}_i^\nu, \mathbf{x}_i^\mathcal{T})\}_{i=1}^B$ represent a batch of visual and textual features, and $\mathbf{z}_{ij} = \Psi \big(\Phi(\mathbf{x}_i^\nu); \mathbf{x}_j^\mathcal{T}\big)$ denote the cross-modal embedding produced with $\mathbf{x}_i^\nu$ and $\mathbf{x}_j^\mathcal{T}$ . For simplicity, we refer to the cross-modal embedding produced using the ground-truth image-text pair as matching embedding, and all other cross-modal embeddings as non-matching embeddings. Then $\mathbf{C}^3$ loss is computed as follows: + +$$ +\mathcal {L} _ {\mathrm {C} ^ {3}} = - \frac {1}{B} \sum_ {j = 1} ^ {B} \log \frac {e ^ {\langle \mathbf {z} _ {j j} , \mathbf {s g} (\mathbf {x} _ {j} ^ {\mathcal {T}}) \rangle}}{\sum_ {i = 1} ^ {B} e ^ {\langle \mathbf {z} _ {i j} , \mathbf {s g} (\mathbf {x} _ {j} ^ {\mathcal {T}}) \rangle}} \tag {8} +$$ + +where $\langle \cdot, \cdot \rangle$ and $\mathfrak{sg}(\cdot)$ denote inner product and stop-gradient operation, respectively. Stop-gradient operation is used to prevent the collapsing solutions where every embedding becomes similar, as discussed in [9]. + +The proposed loss function encourages the matching embeddings to be similar to the input textual feature, while making the non-matching embeddings to be dissimilar. This approach ensures that the modality fusion module preserves the inherent information of the referring expression only when it is paired with the matching image (cycle-consistency), achieved by attending to the referred entities that may be absent in non-matching images. Therefore, an attention map $A^{\mathrm{fuse}}$ is trained to reflect the referring expression accurately, where relevance between the expression and entities automatically emerges. + +# 3.4.2 Reconstruction Loss + +If we can reconstruct the original data using only the set of slots, then the slots will correspond to distinctive individual entities in the input image. Therefore, we leverage the reconstruction objective $\mathcal{L}_{\mathrm{recon}}$ , which is given by + +$$ +\mathcal {L} _ {\text {r e c o n}} = \frac {1}{B} \sum_ {i = 0} ^ {B} \left\| \operatorname {s g} \left(\mathbf {x} _ {i} ^ {\mathcal {Y}}\right) - f _ {\mathrm {d e c}} \left(\Phi \left(\mathbf {x} _ {i} ^ {\mathcal {Y}}\right)\right) \right\| ^ {2}, \tag {9} +$$ + +where $f_{\mathrm{dec}}(\cdot)$ is a MLP based decoder. Unlike the original slot attention [42], our reconstruction is performed in the feature space, rather than the pixel space. This feature reconstruction scheme enables an efficient decoder design. Detailed descriptions are provided in the supplementary material. + +# 3.5. Inference + +The segmentation mask is computed with the two attention maps: $A^{\mathrm{slot}} \in \mathbb{R}^{N \times K}$ and $A^{\mathrm{fuse}} \in \mathbb{R}^{K}$ . $A^{\mathrm{slot}}$ is patchwise attention map that contains information about the discovered entities (Eq. (3)), while $A^{\mathrm{fuse}}$ is entity-wise attention map denoting relevance between the referring expression and the entities (Eq. (7)). To compute the segmentation mask, we first compute $A^{\mathrm{slot}} \cdot A^{\mathrm{fuse}}$ , which results in a 1-dimensional vector of size $N$ . We then reshape this vector into a 2-dimensional map. Next, we bilinearly interpolate the map to the original image size and rescale it using min-max normalization. Finally, we binarize the map using a threshold value $\tau$ to obtain the segmentation mask. + +# 4. Experiments + +# 4.1. Experimental Setting + +Datasets: Our experiments utilize four datasets commonly used in referring image segmentation task: RefCOCO [70], RefCOCO+ [70], Gref [45], and PhraseCut [61]. RefCOCO, RefCOCO+, and Gref are obtained from COCO [39] dataset. RefCOCO contains 142,209 image-text pairs, while RefCOCO+ has 141,564 pairs and does not include expressions indicating orientation properties (e.g., top, right). Gref includes 104,560 pairs for 54,822 objects. PhraseCut comprises 345,486 image-text pairs, which are derived from Visual Genome [35]. + +Implementation details: For visual and textual backbone, we employ ViT-S-16 pretrained on ImageNet-21K [11] and pretrained BERT [13], respectively. We make each modality feature have the same channel dimension $D$ by applying a $1 \times 1$ convolution layer. We set $D$ and $D_h$ to 512 and 1024, respectively. In the aggregation block, the number of iteration $T$ is set to 6, and the number of slots $K$ is set to 36 except for PhraseCut, which is set to 60. For the entity slot, the number of slots sampled from each Gaussian distributions $K_s$ is set to 3 in RefCOCO+ and Gref, while 2 in RefCOCO and PhraseCut. During training, we use AdamW [44] optimizer, with the initial learning rate of 1e-4 scheduled with cosine annealing [43]. Batch size is set to 32 and model is trained for 50 epochs. We resize input images to $384 \times 384$ for both training and evaluation. Threshold value $\tau$ is set to 0.5 for all experiments. + +Evaluation metric: Following [67], we employ the cumulative intersection-over-union (cIoU) and mean intersection-over-Union (mIoU) metric; cIoU divides the total intersections by the total unions across all test images, and mIoU computes IoU for each image and averages them. Also, we evaluate the accuracy at the IoU thresholds of $\{0.3, 0.5, 0.7\}$ , denote as $\mathcal{A} @ \{0.3, 0.5, 0.7\}$ in Table 3. + +# 4.2. Comparisons with the State of the Art + +Weakly supervised methods: As presented in Table 1, our method outperforms TSEG [53], the previous work on weakly-supervised referring image segmentation, in all benchmarks. It shows that the proposed bottom-up and top-down attention framework is more effective than using patch-based visual features with a carefully designed similarity function of TSEG. + +Open-vocabulary segmentation: In Table 1, we also compare our model with the recent open-sourced open-vocabulary segmentation models: GroupViT [64] and MaskCLIP [73], on the referring image segmentation benchmarks. Note that these methods are pre-trained with large-scale image-text datasets [5, 47, 54]. We also report their fine-tuned version for fair comparisons: models are additionally trained with the image-text pairs of each benchmark. Further details for the reproductions are provided in the supplementary material. Ours outperforms GroupViT and MaskCLIP in all settings and benchmarks, even without large-scale image-text pretraining. This result shows that open-vocabulary segmentation models [64, 73] have difficulty comprehending the complex free-from text involving the relation between entities, unlike ours. Interestingly, our model without fine-tuning, which pre-trained on the PhraseCut dataset and evaluated on other benchmarks, still achieved better results. This setting is identical to the open-vocabulary segmentation models, except it leverages a far smaller dataset. These results demonstrate the feasibility of our method in a zero-shot transfer setting. + +
MethodsImage-text pretraining datasetFine tuningRefCOCORefCOCO+Gref valPhraseCut val
valtestAtestBvaltestAtestB
GroupViT [64]CC12M+YFCC (26M)X7.996.1610.518.496.7910.5910.685.82
✓10.8211.1111.2911.1410.7811.8412.779.41
MaskCLIP [73]WIT (400M)X11.5211.8512.0611.8712.0112.5712.7412.13
✓19.4518.6921.3719.9718.9321.4821.1123.80
TSEG [53]-✓25.44--22.01--22.0528.77
OursVisual Genome (0.31M)X21.8019.0024.9622.2019.8624.8525.89-
-✓34.7634.5835.0128.4828.6027.9828.8733.45
+ +![](images/14f576d7b7ed4a9f9d9f63607e371c8fd82debd670a289241a43d13bdece0c3b.jpg) +Query: "Guy on left with beard and guitar" +Figure 4: Qualitative results of our framework with entity slot (Ours), random slot, and query slot on RefCOCO val set. In the case of entity slots, the boundary color indicates the Gaussian distribution that each slot is sampled from. For each slot type, we present 10 discovered entities from $A^{\mathrm{slot}}$ and the final predictions. + +# 4.3. Ablation Study + +We perform ablation studies on the val set of RefCOCO to analyze the contributions of the entity discovery module and the modality fusion module. + +Entity discovery module: In Table 2(a), we ablate the entity discovery module and use visual feature directly, which leads to the drastical performance drop. This result demonstrates that contribution of our bottom-up attention framework is significant. Moreover, we ablate the interaction block, replace slot-attention of aggregation block to transformer, and train the model without reconstruction loss. The results show that all of the design choices for the module is important for effective entity discovery. + +Modality fusion module: In Table 2(b), the modality fusion module is ablated, and the model is trained with image-text contrastive learning [47], where the image feature is obtained by pooling the entity embeddings. The result demonstrates that inferring relevance between entities and referring expression with modality fusion module, trained with $\mathrm{C}^3$ loss, is crucial for accurate segmentation. + +# 4.4. In-depth Analysis + +Analysis of slot initialization: To verify our design choice for entity slots, we investigate variants of slots; query, random, and entity slot. As summarized in Table 3, our entity slot outperforms other slot types across the different + +Table 1: Comparison with weakly-supervised method (TSEG) and open-vocabulary segmentation method (GroupViT and MaskCLIP). The results on four datasets are reported in mIoU (\%). Fine-tuning $\checkmark$ means that the model is trained with the image-text pairs of the target benchmark; otherwise, it just relies on the pre-trained weight without further training. + +
cIoUmIoU
(a) Analysis on entity discovery module
Ours30.4034.76
w/o entity discovery module5.405.91
w/o interaction block29.5933.69
aggregation block w/ transformer11.2912.61
w/o reconstruction loss18.6121.84
(b) Analysis on modality fusion module
Ours30.4034.76
w/o modality fusion module and C3loss13.8114.65
+ +Table 2: Ablation studies of model components and loss function on RefCOCO val set. + +numbers of slots. Specifically, as the number of slots is increased, the performance of the query slot declines, and that of the random slot converges. This result implies that simply increasing the number of slots can not handle the discovery of fine-grained visual entities in real-world images, and thus gains from the proposed entity slot type are non-trivial. In Fig. 4, we also present several visual entities discovered using each slot type, which show the superiority of the entity slot against others as explained in Sec. 3.2. Specifically, the entity slots from the same distribution share the same semantic property (e.g., head, body, or window) while capturing different fine-grained entities. The pro + +![](images/9d4bd5fa503bc668a63907b6248fc962f30b8551aea5ddad4eb76ef3df679417.jpg) + +![](images/3afde1d371f944240b051ee83bfb66269cfe19c9242be78ea73a44e183b3a864.jpg) +Prediction Ground Truth +Prediction Ground Truth + +![](images/3876b1fdd254e5d17cd010b355926aa0b6a6cf817368d1732c573d054817d73b.jpg) +Input Image +on left" + +![](images/c2e2055c93d7bd30bc89a7bce71af6290a47e9f92c8017fd9713a7c3b1a2200f.jpg) +Query: "Right person glasses" +Figure 5: Qualitative results of our framework on RefCOCO val set. We present the discovered entities from $A^{\mathrm{slot}}$ and their relevance scores from $A^{\mathrm{slot}}$ . Top-3 entities in terms of relevance to query expression are presented. + +![](images/c6aee612e77cd84364faccf5ac4672f5864255d047d7a90f37b3d79898198147.jpg) + +![](images/175c1b3fa9efcb4d93edab29328c248837500432843d6bf3637db927d00989e2.jpg) + +![](images/42dc45321ec730e96b3588542271f6b5a7121fd952f25bf07cf55bc86eef8615.jpg) + +![](images/e9ac7d06d18138714c5a7b575e275b5de39907dc979823778807870c43b939c9.jpg) + +![](images/e57ad991d52bab9adc68a4eaf4cdd9961b9b900f3033ff3f0437dd93eea64566.jpg) +Query: "Jeans on right" +Table 4: Impact of hyperparameters $K_{g}$ and $K_{s}$ (a) and inference schemes (b) on RefCOCO val set. + +![](images/509784f85b9678d1583c72ce5453ab7a950f233d63083e48dab53dd6eb99d302.jpg) + +![](images/0e953d3131db8f854a3fdcbebbbcfb0f0f55a1dc23a97e4339d30654543bf73f.jpg) + +![](images/32446252ddc2a20e3b8394db36d4563ea5a92dc3fce8d448abc599b340c08e82.jpg) + +![](images/432a42f85675f372253e8afeb12cfab213b85484c93f65338685787316c95333.jpg) + +![](images/c1573154031eb0f536c2de8ca95a455eb49c1ac14f6c1ac7366e16ab3449ebbd.jpg) + +
Slot type# slotsA@0.3A@0.5A@0.7cIoUmIoU
Query1227.577.170.9622.0421.44
2436.7310.611.9323.5924.78
3619.303.480.3116.6618.16
Random1236.3011.272.6924.2026.65
2445.1916.013.6626.6029.87
3645.6216.883.7626.5530.15
Entity(Ours)1239.4313.873.5624.7227.84
2452.3322.976.2329.2233.54
3655.0224.996.3530.4034.76
+ +posed entity slots effectively discover fine-grained visual entities, which results in an accurate prediction. In contrast, the prediction using random slots leads to a noisy mask, and using query slots leads to the mask of every head that appears in the image, respectively. + +Effect of hyperparameters $K_{g}$ and $K_{s}$ : In Table 4(a), we investigate the impact of hyperparameters for the entity slots: $K_{g}$ and $K_{s}$ , which are the number of Gaussian distributions and slots sampled from each distribution, respectively. We keep the total number of entity slots to 36, only changing $K_{g}$ and $K_{s}$ . The optimal configurations are 18 and 2 for $K_{g}$ and $K_{s}$ , respectively. Interestingly, we observed that even entity slots with $K_{s} = 1$ , where a single slot is sampled from each distribution, vastly surpass the query slots in Table 3; it shows that random sampling of slots prevents the bounding to specific semantic category. + +Effectiveness of $A^{\mathrm{fuse}}$ : To verify the efficacy of $A^{\mathrm{fuse}}$ quantitatively, we evaluate our model with the following inference scheme variants: (1) AVG that replaces $A^{\mathrm{fuse}}$ with the uniform probability distribution, (2) MAX that picks a single attention map from $A^{\mathrm{slot}}$ with the maximum probability in $A^{\mathrm{fuse}}$ , and (3) MIN that picks the attention map with the minimum probability. As summarized in Table 4(b), MAX largely outperforms AVG even though it only uses a single attention map for the final prediction, while MIN results in a near-zero cIoU and mIoU. This tendency shows that $A^{\mathrm{fuse}}$ + +Table 3: Comparison between different slot types on Ref-COCO val set, where # slots is the total number of slots. + +
KgKscIoUmIoU
36129.3633.41
18230.4034.76
12329.6334.03
9428.3132.34
6627.2331.30
+ +(a) + +
Inference schemecIoUmIoU
AVG9.018.16
MAX14.9319.54
MIN0.160.14
A fuse · A slot (Ours)30.4034.76
+ +(b) + +effectively assigns probability based on the relevance between reference and entities, even without direct supervision. + +Qualitative Analysis: In Fig. 5, predicted segmentation mask and discovered entities with top-3 relevance are presented. Results show that the entity discovery module effectively discovers primitive units of segmentation masks, and the modality fusion module accurately infers their relevance to query referring expression. + +# 5. Conclusion + +In this paper, we address the problem of learning referring image segmentation with only image-text pairs. We present a new model based on a bottom-up and top-down attention framework, which first discovers entities in the input image and then combines such entities based on their relevance to the text query. Moreover, we present a new loss function enforcing cycle-consistency between image-text pairs, which enables the effective training of the model without any further supervision. As a result, the proposed method clearly outperforms an existing weakly-supervised method and recent open-vocabulary segmentation models. For future work, we will explore expanding our model to different modalities such as video and audio. + +Acknowledgement. This work was supported by the NRF grant and the IITP grants funded by Ministry of Science and ICT, Korea (NRF-2021R1A2C3012728-20%, IITP-2020-0-00842-20%, IITP-2022-0-00290-20%, IITP-2022-0-00926-20%, IITP-2021-0-02068-10%, IITP-2019-0-01906-10%). + +# References + +[1] Jiwoon Ahn, Sunghyun Cho, and Suha Kwak. Weakly supervised learning of instance segmentation with inter-pixel relations. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[2] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3 +[3] Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. volume abs/1607.06450, 2016. 5 +[4] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In Proc. Neural Information Processing Systems (NeurIPS), 2019. 3 +[5] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 6 +[6] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In Proc. IEEE International Conference on Computer Vision (ICCV), 2019. 2 +[7] Jianbo Chen, Yelong Shen, Jianfeng Gao, Jingjing Liu, and Xiaodong Liu. Language-based image editing with recurrent attentive models. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1 +[8] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In Proc. Neural Information Processing Systems (NeurIPS), 2021. 1 +[9] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5 +[10] Samyak Datta, Karan Sikka, Anirban Roy, Karuna Ahuja, Devi Parikh, and Ajay Divakaran. Align2ground: Weakly supervised phrase grounding guided by image-caption alignment. In Proc. IEEE International Conference on Computer Vision (ICCV), 2019. 2 +[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: a large-scale hierarchical image database. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2009. 6 +[12] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers. In Proc. IEEE International Conference on Computer Vision (ICCV), 2021. 2 +[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Annual Conference of the North American Chapter of the Association for Computational Linguistics, 2019. 4, 6 +[14] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proc. IEEE International Conference on Computer Vision (ICCV), 2021. 1, 2 +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- + +vain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In Proc. International Conference on Learning Representations (ICLR), 2021. 2, 4 +[16] Guang Feng, Zhiwei Hu, Lihe Zhang, and Hutchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[17] Tsu-Jui Fu, Xin Eric Wang, Scott T. Grafton, Miguel P. Eckstein, and William Yang Wang. M3l: Language-based video editing via multi-modal multi-level transformers. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1 +[18] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In Proc. European Conference on Computer Vision (ECCV), 2022. 2, 3 +[19] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 2 +[20] Tanmay Gupta, Arash Vahdat, Gal Chechik, Xiaodong Yang, Jan Kautz, and Derek Hoiem. Contrastive learning for weakly supervised phrase grounding. In European Conference on Computer Vision. Springer, 2020. 2 +[21] S. Hong, D. Yeo, S. Kwak, H. Lee, and B. Han. Weakly supervised semantic segmentation using web-crawled videos. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2224-2232, 2017. 3 +[22] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In Proc. European Conference on Computer Vision (ECCV), 2016. 1, 2 +[23] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[24] Binbin Huang, Dongze Lian, Weixin Luo, and Shenghua Gao. Look before you leap: Learning landmark features for one-stage visual grounding. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[25] Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoqi Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2 +[26] Zilong Huang, Xinggang Wang, Jiasi Wang, Wenyu Liu, and Jingdong Wang. Weakly-supervised semantic segmentation network with deep seeded region growing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3 +[27] Tianrui Hui, Si Liu, Shaofei Huang, Guanbin Li, Sansi Yu, Faxi Zhang, and Jizhong Han. Linguistic structure guided context modeling for referring image segmentation. In Proc. European Conference on Computer Vision (ECCV), 2020. 1 +[28] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proc. International + +Conference on Machine Learning (ICML), 2021. 1, 3 +[29] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proc. IEEE International Conference on Computer Vision (ICCV), 2021. 2 +[30] A. Khoreva, R. Benenson, J. Hosang, M. Hein, and B. Schiele. Simple does it: Weakly supervised instance and semantic segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3 +[31] Beomyoung Kim, Youngjoon Yoo, Chae Eun Rhee, and Junmo Kim. Beyond semantic to instance segmentation: Weakly-supervised instance segmentation via semantic knowledge transfer and self-refinement. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[32] Dongwon Kim, Namyup Kim, and Suha Kwak. Improving cross-modal retrieval with set of diverse embeddings. 2022. 3, 4 +[33] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Restr: Convolution-free referring image segmentation using transformers. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[34] Alexander Kolesnikov and Christoph H. Lampert. Seed, expand and constrain: Three principles for weakly-supervised image segmentation. In Proc. European Conference on Computer Vision (ECCV), 2016. 3 +[35] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, Michael Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision (IJCV), 2017. 6 +[36] Suha Kwak, Seunghoon Hong, and Bohyung Han. Weakly supervised semantic segmentation using superpixel pooling network. In Proc. AAAI Conference on Artificial Intelligence (AAAI), 2017. 3 +[37] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In Proc. International Conference on Learning Representations (ICLR), 2022. 3 +[38] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: common objects in context. In Proc. European Conference on Computer Vision (ECCV), 2014. 6 +[40] Chenxi Liu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, and Alan Yuille. Recurrent multimodal interaction for referring image segmentation. In Proc. IEEE International Conference on Computer Vision (ICCV), 2017. 2 +[41] Yongfei Liu, Bo Wan, Lin Ma, and Xuming He. Relation-aware instance refinement for weakly supervised visual grounding. In Proc. IEEE Conference on Computer Vision + +and Pattern Recognition (CVPR), 2021. 2 +[42] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. In Proc. Neural Information Processing Systems (NeurIPS), 2020. 2, 3, 4, 5, 6 +[43] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In Proc. International Conference on Learning Representations (ICLR), 2017. 6 +[44] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. Proc. International Conference on Learning Representations (ICLR), 2019. 6 +[45] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6 +[46] Edgar Margffoy-Tuay, Juan C Pérez, Emilio Botero, and Pablo Arbeláez. Dynamic multimodal instance segmentation guided by natural language queries. In Proc. European Conference on Computer Vision (ECCV), 2018. 2 +[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. International Conference on Machine Learning (ICML), 2021. 1, 2, 3, 6, 7 +[48] Tal Remez, Jonathan Huang, and Matthew Brown. Learning to segment via cut-and-paste. In Proc. European Conference on Computer Vision (ECCV), 2018. 3 +[49] Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, and Bernt Schiele. Grounding of textual phrases in images by reconstruction. In Proc. European Conference on Computer Vision (ECCV), 2016. 2 +[50] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3 +[51] Hengcan Shi, Hongliang Li, Fanman Meng, and Qingbo Wu. Key-word-aware network for referring expression image segmentation. In Proc. European Conference on Computer Vision (ECCV), 2018. 2 +[52] Gautam Singh, Fei Deng, and Sungjin Ahn. Illiterate dalle learns to compose. In Proc. International Conference on Learning Representations (ICLR), 2021. 3, 4 +[53] Robin Strudel, Ivan Laptev, and Cordelia Schmid. Weakly-supervised segmentation of referring expressions. arXiv preprint arXiv:2205.04725, 2022. 2, 6, 7 +[54] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 2016. 2, 6 +[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. Neural Information Processing Systems (NeurIPS), 2017. 5 +[56] Liwei Wang, Jing Huang, Yin Li, Kun Xu, Zhengyuan Yang, and Dong Yu. Improving weakly supervised visual ground + +ing by contrastive knowledge distillation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[57] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1 +[58] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[59] Yunchao Wei, Jiashi Feng, Xiaodan Liang, Ming-Ming Cheng, Yao Zhao, and Shuicheng Yan. Object region mining with adversarial erasing: A simple classification to semantic segmentation approach. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3 +[60] Philippe Weinzaepfel, Thomas Lucas, Diane Larlus, and Yannis Kalantidis. Learning super-features for image retrieval. In Proc. International Conference on Learning Representations (ICLR), 2022. 3, 4 +[61] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasecut: Language-based image segmentation in the wild. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6 +[62] Yongqin Xian, Subhabrata Choudhury, Yang He, Bernt Schiele, and Zeynep Akata. Semantic projection network for zero-and few-label semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8256-8265, 2019. 3 +[63] Fanyi Xiao, Leonid Sigal, and Yong Jae Lee. Weakly-supervised visual grounding of phrases with linguistic structures. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2 +[64] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4, 6, 7 +[65] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310-4319, 2022. 3 +[66] Li Yang, Yan Xu, Chunfeng Yuan, Wei Liu, Bing Li, and Weiming Hu. Improving visual grounding with visual-linguistic verification and iterative reasoning. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[67] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6 +[68] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2 + +[69] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1, 2 +[70] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Proc. European Conference on Computer Vision (ECCV), 2016. 6 +[71] Dong Zhang, Hanwang Zhang, Jinhui Tang, Xian-Sheng Hua, and Qianru Sun. Causal intervention for weakly-supervised semantic segmentation. 2020. 3 +[72] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba. Learning deep features for discriminative localization. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3 +[73] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In Proc. European Conference on Computer Vision (ECCV), 2022. 2, 3, 6, 7 +[74] Yi Zhou, Hui Zhang, Hana Lee, Shuyang Sun, Pingjun Li, Yangguang Zhu, ByungIn Yoo, Xiaojuan Qi, and Jae-Joon Han. Slot-vps: Object-centric representation learning for video panoptic segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 4 +[75] Yanzhao Zhou, Yi Zhu, Qixiang Ye, Qiang Qiu, and Jianbin Jiao. Weakly supervised instance segmentation using class peak response. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3 +[76] Yuliang Zou, Zizhao Zhang, Han Zhang, Chun-Liang Li, Xiao Bian, Jia-Bin Huang, and Tomas Pfister. Pseudoseg: Designing pseudo labels for semantic segmentation. In Proc. International Conference on Learning Representations (ICLR), 2021. 3 \ No newline at end of file diff --git a/2023/Shatter and Gather_ Learning Referring Image Segmentation with Text Supervision/images.zip b/2023/Shatter and Gather_ Learning Referring Image Segmentation with Text Supervision/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4d20b413778c3720f6bf8af017ec955af1b94cd1 --- /dev/null +++ b/2023/Shatter and Gather_ Learning Referring Image Segmentation with Text Supervision/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a633fc82cc4f778d25d8e44c8663ddb3c453ad43418984e032dfd9735b314b56 +size 643550 diff --git a/2023/Shatter and Gather_ Learning Referring Image Segmentation with Text Supervision/layout.json b/2023/Shatter and Gather_ Learning Referring Image Segmentation with Text Supervision/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a30a2a95fbf679c52e3c7568c0a5315847451273 --- /dev/null +++ b/2023/Shatter and Gather_ Learning Referring Image Segmentation with Text Supervision/layout.json @@ -0,0 +1,10129 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 103, + 489, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 103, + 489, + 140 + ], + "spans": [ + { + "bbox": [ + 100, + 103, + 489, + 140 + ], + "type": "text", + "content": "Shatter and Gather: Learning Referring Image Segmentation with Text Supervision" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 157, + 463, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 157, + 463, + 200 + ], + "spans": [ + { + "bbox": [ + 129, + 157, + 463, + 200 + ], + "type": "text", + "content": "Dongwon Kim" + }, + { + "bbox": [ + 129, + 157, + 463, + 200 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 129, + 157, + 463, + 200 + ], + "type": "text", + "content": " Namyup Kim" + }, + { + "bbox": [ + 129, + 157, + 463, + 200 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 129, + 157, + 463, + 200 + ], + "type": "text", + "content": " Cuiling Lan" + }, + { + "bbox": [ + 129, + 157, + 463, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 129, + 157, + 463, + 200 + ], + "type": "text", + "content": " Suha Kwak" + }, + { + "bbox": [ + 129, + 157, + 463, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 157, + 463, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 157, + 463, + 200 + ], + "type": "text", + "content": "POSTECH Microsoft Research Asia \nhttp://cvlab.postech.ac.kr/research/sag" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 228, + 192, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 228, + 192, + 240 + ], + "spans": [ + { + "bbox": [ + 143, + 228, + 192, + 240 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 253, + 290, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 290, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 290, + 445 + ], + "type": "text", + "content": "Referring image segmentation, the task of segmenting any arbitrary entities described in free-form texts, opens up a variety of vision applications. However, manual labeling of training data for this task is prohibitively costly, leading to lack of labeled data for training. We address this issue by a weakly supervised learning approach using text descriptions of training images as the only source of supervision. To this end, we first present a new model that discovers semantic entities in input image and then combines such entities relevant to text query to predict the mask of the referent. We also present a new loss function that allows the model to be trained without any further supervision. Our method was evaluated on four public benchmarks for referring image segmentation, where it clearly outperformed the existing method for the same task and recent open-vocabulary segmentation models on all the benchmarks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 463, + 128, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 463, + 128, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 463, + 128, + 475 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 479, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 575 + ], + "type": "text", + "content": "Referring image segmentation is the task of segmenting the referent corresponding to a natural language expression given as a query. Unlike the conventional semantic segmentation that aims at segmenting a pre-defined set of classes, referring image segmentation enables segmentation of any arbitrary entities described in free-form texts and thus opens up a wide variety of applications such as human-computer interaction [8, 57] and text-based image editing [7, 17]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 575, + 288, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 288, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 288, + 695 + ], + "type": "text", + "content": "Thanks to the development of deep neural networks, recent studies have demonstrated remarkable results on referring image segmentation in the supervised learning setting [14, 22, 25, 27, 67, 68, 69]. However, obtaining manual annotation of training data for referring image segmentation is prohibitively costly since the data requires two types of labels for each image, i.e., natural language expressions describing the entities that appear in the image and segmentation labels corresponding to them. This typically demands tremendously more manpower than annotation for semantic" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 227, + 545, + 445 + ], + "blocks": [ + { + "bbox": [ + 308, + 227, + 545, + 445 + ], + "lines": [ + { + "bbox": [ + 308, + 227, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 308, + 227, + 545, + 445 + ], + "type": "image", + "image_path": "dd9ca38fcbe60fa581656fa48f2190717303716cf506d35dc6a41acacf2ed70b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 446, + 547, + 543 + ], + "lines": [ + { + "bbox": [ + 304, + 446, + 547, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 547, + 543 + ], + "type": "text", + "content": "Figure 1: Our model consists of two different attention modules: bottom-up attention that identifies individual entities in an image, and top-down attention that combines the entities based on a referring expression. It is trained in a contrastive manner to ensure consistency between matching image-text pairs. In inference, a segmentation mask is predicted by combining the entities found by bottom-up attention with weights derived by top-down attention." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 557, + 537, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 557, + 537, + 569 + ], + "spans": [ + { + "bbox": [ + 305, + 557, + 537, + 569 + ], + "type": "text", + "content": "segmentation, leading to lack of labeled data for training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "A solution to this issue is weakly supervised learning, which trains a model on a dataset providing weaker forms of supervision than the conventional ones. In this paper, we particularly focus on learning referring image segmentation only with natural language expressions describing the associated images. Since these labels are readily available in many vision-language datasets [28, 47], the issue on lack of labeled training data can be alleviated by our approach. However, learning referring image segmentation in this setting introduces another challenge: Since natural language expressions often contain relations between visual entities such as instances and object parts (e.g., \"The cat next to the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 130, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 130, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 130, + 712 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15547" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "table leg\"), a model to be trained should be aware of the inter-object relations to exploit the expressions as supervision. Thereby we need a method to discover individual entities and infer their relations without any supervision but the referring expression." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 301 + ], + "type": "text", + "content": "To fulfill this requirement, we introduce a new bottom-up and top-down attention framework, which is illustrated in Fig. 1. Individual entities existing in an image are first discovered by bottom-up attention, which solely exploits visual information. For the bottom-up attention, an entity discovery module progressively refines a set of embedding vectors named slots to capture distinct visual entities, following slot attention [42]. We moreover propose a novel slot formulation named entity slot, which enables fine-grained entity discovery in real-world images. As a result, entity slots after the refinement capture individual visual entities without any segmentation supervision. These entities are considered primitive units for composing the segmentation mask of a referent during inference." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 302, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 287, + 422 + ], + "type": "text", + "content": "Afterward, based on the relational and structural information of referring expression, top-down attention selectively attends to the referred entities and combines them into the predicted mask. Top-down attention is implemented by a modality fusion module consists of cross-attention transformer [15]. This attention scheme enables the module to infer the relevance between discovered entities and the referring expression instead of relying on the cosine similarity or heuristic similarity functions used in previous work [18, 53, 64, 73]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 423, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 532 + ], + "type": "text", + "content": "For training, we propose the contrastive cycle-consistency " + }, + { + "bbox": [ + 46, + 423, + 287, + 532 + ], + "type": "inline_equation", + "content": "(\\mathrm{C}^3)" + }, + { + "bbox": [ + 46, + 423, + 287, + 532 + ], + "type": "text", + "content": " loss that enforces cycle-consistency between image-text pairs under the contrastive learning framework [47]. To establish such consistency, the model is trained to preserve the textual information after the fusion of entities and textual features. By doing so, we observe that latent relevance between the discovered entities and referring expression automatically emerges in the top-down attention without any explicit mask supervision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 532, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 605 + ], + "type": "text", + "content": "One may wonder the difference between our task and open-vocabulary segmentation [18, 64, 73], i.e., semantic segmentation of arbitrary categories without explicit mask supervision. The key difference is that a referent to segment in referring image segmentation is given by a complex free-form text, which introduces additional challenges." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "The proposed method was evaluated and compared on four public benchmarks for referring image segmentation, where it substantially outperformed the previous weakly supervised learning method [53]. Moreover, we reproduced recent open-vocabulary segmentation models [64, 73] and evaluated them on referring image segmentation benchmarks; the results show that the proposed model surpassed them even without pre-training on large-scale image-text data [5, 47, 54]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 72, + 534, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 534, + 85 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 534, + 85 + ], + "type": "text", + "content": "In summary, our contribution is three-fold as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 87, + 545, + 247 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 317, + 87, + 545, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 87, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 87, + 545, + 160 + ], + "type": "text", + "content": "- We present a new weakly supervised learning model for referring image segmentation. Our model employs bottom-up and top-down attention to discover individual entities and infer their relations, which enables to exploit natural language expressions as supervision effectively during training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 163, + 545, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 545, + 211 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 545, + 211 + ], + "type": "text", + "content": "- We propose a new loss function called contrastive cycle-consistency loss, which allows latent relevance between discovered entities and referring expressions to emerge without requiring further supervision." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 213, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 213, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 317, + 213, + 545, + 247 + ], + "type": "text", + "content": "- Our model clearly outperformed the existing method for the same task and recent open-vocabulary segmentation models on all the four benchmarks." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 264, + 392, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 264, + 392, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 264, + 392, + 277 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 281, + 545, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 532 + ], + "type": "text", + "content": "Referring image segmentation: Early approaches [22, 38, 40, 51] obtain multimodal feature maps by simple concatenation-convolution operation. With this operation as a basis, most existing studies have suggested attention mechanisms and multi-level feature aggregations [6, 16, 23, 25, 46, 68]. Various approaches have been employed for calculating fine interactions between the modalities, including multimodal feature fusion [14, 29, 33] using transformers [15], a large-scale vision-language pretrained framework [58], and holistically multimodal fusion [67]. The existing referring image segmentation methods have been developed in a fully supervised learning setting. Recently, a weakly supervised learning method [53] using only imagedtext pairs for training is introduced to alleviate the high annotation cost. However, this study relies on a heuristic similarity function to infer relations between patches and texts, which suffers from comprehending the relational and structural information from image and text modalities. Therefore, we propose a new weakly supervised model consisting of bottom-up and top-down attention to discover visual entities and infer their relations effectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "Phrase grounding: Phrase grounding [12, 24, 29, 49, 66, 69] is the task of learning the association between image regions and words in caption describing the image. Weakly supervised approaches have also been proposed for this task [10, 20, 41, 56, 63], where the correspondence between words and regions are not given. It is worth noting that the contrastive learning framework introduced in [20] is closely related to the proposed " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "C^3" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " loss by enforcing mutual information between regions and words. However, these studies rely on an off-the-shelf object detector [19], where the final prediction is restricted by the predefined set of classes with which the detector is trained. On the contrary, our method discovers visual entities in an image without an auxiliary model trained with external labeled data, which is more suitable for referring image segmentation." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15548" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 545, + 248 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 545, + 248 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 545, + 248 + ], + "type": "image", + "image_path": "47b618ab2f05409361cba42c53cac8d18a2060131b5c6f2b3448f65025736f6f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 548, + 336 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 548, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 548, + 336 + ], + "type": "text", + "content": "Figure 2: Illustration of the overall architecture of our model along with its behavior during training and inference. (a) Feature extractors: The visual and textual features are extracted by transformer-based encoders of the two modalities respectively. (b) Entity discovery module: A set of visual entities are discovered from the visual features through the bottom-up attention. (c) Modality fusion module: The visual entities and a referring expression are fused considering their relevance, which is estimated by the top-down attention. Our model is trained with the proposed contrastive cycle-consistency " + }, + { + "bbox": [ + 46, + 251, + 548, + 336 + ], + "type": "inline_equation", + "content": "(\\mathbf{C}^3)" + }, + { + "bbox": [ + 46, + 251, + 548, + 336 + ], + "type": "text", + "content": " loss between the cross-modal embeddings and textual features. In inference, the segmentation mask is predicted by jointly considering the visual entities " + }, + { + "bbox": [ + 46, + 251, + 548, + 336 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{slot}}" + }, + { + "bbox": [ + 46, + 251, + 548, + 336 + ], + "type": "text", + "content": " from the entity discovery module and the relevance scores " + }, + { + "bbox": [ + 46, + 251, + 548, + 336 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{fuse}}" + }, + { + "bbox": [ + 46, + 251, + 548, + 336 + ], + "type": "text", + "content": " from the modality fusion module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 347, + 289, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 289, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 289, + 504 + ], + "type": "text", + "content": "Weakly supervised learning for segmentation: Weakly supervised learning has become a favored solution for addressing the lack of annotated data in training, especially for semantic segmentation [1, 2, 21, 26, 34, 36, 65, 71, 76] and instance segmentation [30, 31, 48, 75]. These methods refine coarse localization cues obtained by CAMs [50, 59, 72], and use the results as pseudo labels for self-training. However, they cannot be directly applied to the referring image segmentation due to the following challenges. First, there are no predefined classes in this task, making it difficult to learn shared semantics within CAMs by multiple instance learning. Second, there is no mechanism to discover visual entities and combine them with textual information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 506, + 289, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 289, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 289, + 674 + ], + "type": "text", + "content": "Slot attention: Slot attention [42] is proposed for the learning object-centric representation. It can discover the objects within an image in an unsupervised manner and be used in the various downstream task such as segmentation [64, 74] and retrieval [32, 60]. In specific, slot attention introduces an iterative attention mechanism that operates on the slots, which are the set of embedding vectors corresponding to the objects that appear in the input data. In previous work, slots are initialized as a set of learnable embeddings [32, 60, 64, 74] or a set of vectors sampled from a learnable Gaussian distribution [42, 52]. However, we empirically observed that both types of initialization are inapplicable to visual entity discovery for referring image segmentation, which we will elaborate on in Sec. 3.2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "Open-vocabulary segmentation: The open-vocabulary segmentation is proposed to address the limitation of semantic segmentation that it is constrained to the pre" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 347, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 547, + 540 + ], + "type": "text", + "content": "defined set of classes. Unlike semantic segmentation, open-vocabulary segmentation models are evaluated with classes unseen during training. The early approaches propose a generative model [4] and a projection model [62] to generate or project pixel-level semantic features from visual features for the unseen classes through word embedding space. Recent studies [18, 37, 64, 73] largely improve the quality of semantic features for the unseen classes, and further introduce the methods that do not require segmentation labels for seen classes using large-scale image-text pretrained models (e.g.CLIP [47], ALIGN [28]). However, since they are trained to predict masks for semantic categories, not for free-form texts describing arbitrary entities in an image, their applicability is still limited. We thus propose a new method for referring image segmentation that only requires image-text pairs during training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 552, + 411, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 411, + 566 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 411, + 566 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "As shown in Fig. 2, our model consists of three main components: feature extractors, entity discovery module, and modality fusion module. The visual and textual features are extracted from the input using the backbone networks of the two modalities respectively. (Sec. 3.1). The visual features are then given as input to the entity discovery module, which aggregates the features into multiple entity embeddings using the bottom-up attention (Sec. 3.2). The entity embeddings correspond to distinct entities appearing in the image, which are considered as the smallest units of building the segmentation mask. The modality fusion module fuses the entity embeddings and the textual feature into" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15549" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "cross-modal embeddings according to their relevance using the top-down attention (Sec. 3.3). The produced cross-modal embedding is used only for training, with a learning objective enforcing cycle-consistency and reconstruction of visual features (Sec. 3.4). The segmentation mask is predicted by combining the entities discovered by the entity discovery module with the relevance scores estimated by the modality fusion module (Sec. 3.5)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 177, + 249, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 177, + 249, + 189 + ], + "spans": [ + { + "bbox": [ + 47, + 177, + 249, + 189 + ], + "type": "text", + "content": "3.1. Visual and Textual Feature Extraction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 193, + 287, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 287, + 216 + ], + "type": "text", + "content": "Here, we provide the details of the transformer-based backbone network for each modality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 218, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 287, + 289 + ], + "type": "text", + "content": "Visual feature extraction: We employ the vision transformer (ViT) [15] as the visual backbone network. An input image is first split into " + }, + { + "bbox": [ + 46, + 218, + 287, + 289 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 218, + 287, + 289 + ], + "type": "text", + "content": " patches and then linearly projected. The patches are processed by the multiple self-attention blocks of ViT, where we refer to the output tokens of the last block as the visual features " + }, + { + "bbox": [ + 46, + 218, + 287, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathcal{V}}\\in \\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 46, + 218, + 287, + 289 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 292, + 287, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 287, + 364 + ], + "type": "text", + "content": "Textual feature extraction: We consider the transformer-based text encoder, BERT [13]. A sentence is first tokenized into the token sequence, where the [CLS] and [SEP] tokens are added to the beginning and the end of the sequence. The activation of the last layer of the transformer for the [CLS] token is employed as the textual feature " + }, + { + "bbox": [ + 46, + 292, + 287, + 364 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathcal{T}}\\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 46, + 292, + 287, + 364 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 372, + 287, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 287, + 384 + ], + "type": "text", + "content": "3.2. Bottom-up Attention: Entity Discovery Module" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 388, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 483 + ], + "type": "text", + "content": "The entity discovery module applies a bottom-up attention mechanism to aggregate patch-based visual features into multiple entity embeddings that represent instances or object parts in the image. The discovered entities are then treated as primitive units of the segmentation mask, where the mask is predicted by selectively combining them. This approach is useful since the textual information is often built upon the relations between such visual entities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 484, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 604 + ], + "type": "text", + "content": "Specifically, the entity discovery module consists of two blocks: aggregation block and interaction block. The entity slots, which are embedding vectors that corresponds to various visual entities, are updated in an iterative process alternating between these two blocks. These slots eventually become entity embeddings. Let " + }, + { + "bbox": [ + 46, + 484, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\Phi : \\mathbb{R}^{N \\times D} \\mapsto \\mathbb{R}^{K \\times D}" + }, + { + "bbox": [ + 46, + 484, + 287, + 604 + ], + "type": "text", + "content": " be the entity discovery module that takes visual features as input and outputs the set of " + }, + { + "bbox": [ + 46, + 484, + 287, + 604 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 484, + 287, + 604 + ], + "type": "text", + "content": " entity embeddings. The entity embeddings are obtained by iteratively updating initial entity slots " + }, + { + "bbox": [ + 46, + 484, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^0 \\in \\mathbb{R}^{K \\times D}" + }, + { + "bbox": [ + 46, + 484, + 287, + 604 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 484, + 287, + 604 + ], + "type": "text", + "content": "-times:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 99, + 613, + 287, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 613, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 99, + 613, + 287, + 644 + ], + "type": "interline_equation", + "content": "\\begin{array}{r l} & \\Phi (\\mathbf {x} ^ {\\mathcal {V}}) := \\mathbf {S} ^ {T}, \\text {w h e r e} \\\\ & \\quad \\mathbf {S} ^ {t + 1} = \\operatorname {I n t e r} \\left(\\operatorname {A g g} \\left(\\mathbf {x} ^ {\\mathcal {V}}; \\mathbf {S} ^ {t}\\right)\\right), \\end{array} \\tag {1}", + "image_path": "9df553139575d463d521da70d9a3864965b99142d93eb3647b868021398ef9e4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 677 + ], + "type": "text", + "content": "and Agg and Inter denote the aggregation and interaction blocks, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "In the literature, the initial slots " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^0" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " are often defined as either vectors sampled from a learnable Gaussian distribution [42, 52] or learnable embeddings [32, 60, 64, 74]," + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 71, + 537, + 286 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 537, + 286 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 537, + 286 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 537, + 286 + ], + "type": "image", + "image_path": "656824a77cfeba5f159c67838a6b7b9b2b9eb64851781382dec522c4f14f99e1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 289, + 546, + 361 + ], + "lines": [ + { + "bbox": [ + 304, + 289, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 546, + 361 + ], + "type": "text", + "content": "Figure 3: Comparison between the proposed entity slot and previous slot initialization methods. The boundary color indicates the Gaussian distribution that each slot is sampled from. The entity slot successfully discovers fine-grained entities such as head and food, while the others often fail to capture individual entities of specific semantics." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 373, + 546, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 546, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 546, + 601 + ], + "type": "text", + "content": "where such types of slots are referred as random slot and query slot in the remainder of this paper, respectively. However, we have empirically found that both of these initialization strategies are unsuitable for our task as shown in Fig. 3. The random slot suffers from the lack of semantic specificity since every slot is randomly sampled from a single distribution. This can lead to suboptimal performance since entities in real-world images are diverse in terms of semantics. The query slot, on the other hand, suffers from the limitation of being bound to specific semantic categories, such as human, animal, and vehicle. This limitation leads to entity discovery that is not sufficiently fine-grained, e.g., given an image of multiple people, a single slot that binds to \"human\" captures all of them and thus failed to identify the individual specified by the text query. This is problematic particularly in our model since entities act as the smallest units for producing segmentation masks. It is highly desired to have a mechanism that could guide the differentiation of slots to be semantically specific." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 601, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 545, + 673 + ], + "type": "text", + "content": "To overcome these limitations, we introduce a new initialization strategy for the entity slot, which enables the discovery of fine-grained visual entities while learning diverse semantics. Formally, the initial entity slots " + }, + { + "bbox": [ + 304, + 601, + 545, + 673 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^0" + }, + { + "bbox": [ + 304, + 601, + 545, + 673 + ], + "type": "text", + "content": " are obtained by sampling " + }, + { + "bbox": [ + 304, + 601, + 545, + 673 + ], + "type": "inline_equation", + "content": "K_{s}" + }, + { + "bbox": [ + 304, + 601, + 545, + 673 + ], + "type": "text", + "content": " slots from each of " + }, + { + "bbox": [ + 304, + 601, + 545, + 673 + ], + "type": "inline_equation", + "content": "K_{g}" + }, + { + "bbox": [ + 304, + 601, + 545, + 673 + ], + "type": "text", + "content": " different Gaussian distributions, resulting in a total of " + }, + { + "bbox": [ + 304, + 601, + 545, + 673 + ], + "type": "inline_equation", + "content": "K = K_{s}\\cdot K_{g}" + }, + { + "bbox": [ + 304, + 601, + 545, + 673 + ], + "type": "text", + "content": " slots:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 341, + 678, + 545, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 678, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 341, + 678, + 545, + 700 + ], + "type": "interline_equation", + "content": "\\mathbf {S} ^ {0} := \\left[ \\mathbf {S} _ {1} ^ {0}; \\dots ; \\mathbf {S} _ {K _ {g}} ^ {0} \\right] \\in \\mathbb {R} ^ {K \\times D}, \\text {w h e r e} \\tag {2}", + "image_path": "1c233e857843cf604919c5424900bc4d0efc077fbbe85a4530a92dc5d5a7b890.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 343, + 696, + 485, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 696, + 485, + 711 + ], + "spans": [ + { + "bbox": [ + 343, + 696, + 485, + 711 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {i} ^ {0} \\sim \\mathcal {N} (\\mu_ {i}, \\operatorname {d i a g} (\\sigma_ {i})) \\in \\mathbb {R} ^ {K _ {s} \\times D}.", + "image_path": "078bb216ff82bdcae1f1d562e2e8d89443cae2377da5e76551649a2bba992bd9.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15550" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "The idea is to initialize slots by sampling them from different Gaussian distributions, each specifying a distinct semantic property to be captured. Individual entities sharing the same semantics are differentiated by slots sampled from the same distribution, which enables fine-grained entity discovery." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "text", + "content": "From now, we present the aggregation block and interaction block at " + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "text", + "content": "-th iteration. As described in Eq. (1), the slots of previous iteration " + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{t-1}" + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "text", + "content": " are fed to the aggregation block with the visual features " + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\nu}" + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "text", + "content": ". The role of the aggregation block is to let each slot aggregate visual features of distinct visual entities, and it is achieved by employing the competitive attention mechanism introduced in the slot attention [42]. Let " + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "inline_equation", + "content": "q(\\cdot)" + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "inline_equation", + "content": "k(\\cdot)" + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "inline_equation", + "content": "v(\\cdot)" + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "text", + "content": " are linear projection with layer normalization [3] that map inputs to dimension " + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "inline_equation", + "content": "D_h" + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "text", + "content": ". Then, the attention map " + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{slot}} \\in \\mathbb{R}^{N \\times K}" + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\nu}" + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{t-1}" + }, + { + "bbox": [ + 46, + 144, + 288, + 276 + ], + "type": "text", + "content": " is computed as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 283, + 287, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 287, + 317 + ], + "type": "interline_equation", + "content": "A _ {i, j} ^ {\\text {s l o t}} = \\frac {e ^ {M _ {i , j}}}{\\sum_ {j = 1} ^ {K} e ^ {M _ {i , j}}}, \\text {w h e r e} M = \\frac {k \\left(\\mathbf {x} ^ {\\mathcal {V}}\\right) q \\left(\\mathbf {S} ^ {t - 1}\\right) ^ {\\top}}{\\sqrt {D _ {h}}}. \\tag {3}", + "image_path": "4c1954c46664c6ca3609cb4f20f7de3d1ec3d98de639894995393deb47e61b23.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 324, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 432 + ], + "type": "text", + "content": "Unlike the cross-attention transformer [55] that normalized attention over the keys, we normalize attention map " + }, + { + "bbox": [ + 46, + 324, + 287, + 432 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{slot}}" + }, + { + "bbox": [ + 46, + 324, + 287, + 432 + ], + "type": "text", + "content": " over the slots. This normalization scheme encourages slots to compete for aggregating the visual features, which results in the set of slots binding to different visual entities in the input data. It is worth noting that " + }, + { + "bbox": [ + 46, + 324, + 287, + 432 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{slot}}" + }, + { + "bbox": [ + 46, + 324, + 287, + 432 + ], + "type": "text", + "content": " provides the localizations of entities discovered, which is used as primitive units of the segmentation mask during inference. Using " + }, + { + "bbox": [ + 46, + 324, + 287, + 432 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{slot}}" + }, + { + "bbox": [ + 46, + 324, + 287, + 432 + ], + "type": "text", + "content": ", the slots are updated in a similar manner to the transformer:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 449, + 287, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 449, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 75, + 449, + 287, + 482 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\widehat {\\mathbf {S}} ^ {t} = \\operatorname {A g g} \\left(\\mathbf {x} ^ {\\nu}; \\mathbf {S} ^ {t}\\right) \\tag {4} \\\\ = \\operatorname {n o r m a l i z e} \\left(A ^ {\\text {s l o t}}\\right) ^ {\\top} v \\left(\\mathbf {x} ^ {\\nu}\\right) W _ {o} + \\mathbf {S} ^ {t - 1}, \\\\ \\end{array}", + "image_path": "7f2bda616c9b2c23bf4b8d409decbc3c76d89986772d2eb0cf2805fd45f62148.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 490, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 490, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 490, + 287, + 538 + ], + "type": "text", + "content": "where normalize " + }, + { + "bbox": [ + 46, + 490, + 287, + 538 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 46, + 490, + 287, + 538 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 46, + 490, + 287, + 538 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 46, + 490, + 287, + 538 + ], + "type": "text", + "content": "-normalization of the column vectors and " + }, + { + "bbox": [ + 46, + 490, + 287, + 538 + ], + "type": "inline_equation", + "content": "W_{o}\\in \\mathbb{R}^{D_{h}\\times D}" + }, + { + "bbox": [ + 46, + 490, + 287, + 538 + ], + "type": "text", + "content": " is linear projection. Finally, slots are further refined by " + }, + { + "bbox": [ + 46, + 490, + 287, + 538 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{S}}^t\\gets \\mathrm{MLP}\\big(\\widehat{\\mathbf{S}}^t\\big) + \\widehat{\\mathbf{S}}^t" + }, + { + "bbox": [ + 46, + 490, + 287, + 538 + ], + "type": "text", + "content": " where MLP is multi-layer perceptron (MLP) with layer normalization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 538, + 287, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 287, + 621 + ], + "type": "text", + "content": "The following interaction block consists of multiple self-attention transformers [55]. The weights of these transformers are shared, and each transformer is applied to slots sampled from the same distribution. We denote the transformer layer as Transformer(Q; K; V), where (Q; K; V) are for query, key, and value. The interaction block is denoted as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 87, + 629, + 287, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 629, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 87, + 629, + 287, + 647 + ], + "type": "interline_equation", + "content": "\\operatorname {I n t e r} \\left(\\widehat {\\mathbf {S}} ^ {t}\\right) = \\left[ \\operatorname {S A} \\left(\\widehat {\\mathbf {S}} _ {1} ^ {t}\\right); \\dots ; \\operatorname {S A} \\left(\\widehat {\\mathbf {S}} _ {K _ {g}} ^ {t}\\right) \\right], \\tag {5}", + "image_path": "18c8634d62987c5d6d19c40a867f9c9e74c0c639920138fe9f05a3f66af80724.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{SA}(\\mathbf{X})" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " is for the Transformer(X; X; X). This self-attention scheme models information exchange between visual entities sharing the same semantic property, which helps slots within a single semantic category to capture different entities without redundancy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 545, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 85 + ], + "type": "text", + "content": "3.3. Top-down Attention: Modality Fusion Module" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 87, + 545, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 87, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 305, + 87, + 545, + 171 + ], + "type": "text", + "content": "The modality fusion module is proposed for top-down attention mechanism, which combines discovered visual entities based on the textual feature. Let " + }, + { + "bbox": [ + 305, + 87, + 545, + 171 + ], + "type": "inline_equation", + "content": "\\Psi: \\mathbb{R}^{K \\times D} \\times \\mathbb{R}^D \\mapsto \\mathbb{R}^D" + }, + { + "bbox": [ + 305, + 87, + 545, + 171 + ], + "type": "text", + "content": " be the modality fusion module, which takes entity embeddings " + }, + { + "bbox": [ + 305, + 87, + 545, + 171 + ], + "type": "inline_equation", + "content": "\\Phi(\\mathbf{x}^\\nu)" + }, + { + "bbox": [ + 305, + 87, + 545, + 171 + ], + "type": "text", + "content": " and textual furret " + }, + { + "bbox": [ + 305, + 87, + 545, + 171 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^T" + }, + { + "bbox": [ + 305, + 87, + 545, + 171 + ], + "type": "text", + "content": " as inputs. Then, the output cross-modal embedding " + }, + { + "bbox": [ + 305, + 87, + 545, + 171 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 305, + 87, + 545, + 171 + ], + "type": "text", + "content": " is computed using cross-attention transformer:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 335, + 178, + 545, + 208 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 178, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 335, + 178, + 545, + 208 + ], + "type": "interline_equation", + "content": "\\mathbf {z} = \\Psi \\left(\\Phi \\left(\\mathbf {x} ^ {\\mathcal {V}}\\right); \\mathbf {x} ^ {\\mathcal {T}}\\right) = \\frac {\\operatorname {C A} \\left(\\Phi \\left(\\mathbf {x} ^ {\\mathcal {V}}\\right) ; \\mathbf {x} ^ {\\mathcal {T}}\\right)}{\\| \\operatorname {C A} \\left(\\Phi \\left(\\mathbf {x} ^ {\\mathcal {V}}\\right) ; \\mathbf {x} ^ {\\mathcal {T}}\\right) \\| _ {2}}, \\tag {6}", + "image_path": "2f020d64795a6fb90037d9bb21463053fd47c485ab77397f1fc29e760bb091dc.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 217, + 545, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 217, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 545, + 253 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 217, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\mathrm{CA}(\\mathrm{X};\\mathrm{Y})" + }, + { + "bbox": [ + 305, + 217, + 545, + 253 + ], + "type": "text", + "content": " is for the Transformer(Y;X;X) and " + }, + { + "bbox": [ + 305, + 217, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _2" + }, + { + "bbox": [ + 305, + 217, + 545, + 253 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 305, + 217, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 305, + 217, + 545, + 253 + ], + "type": "text", + "content": " norm. Moreover, we can obtain the attention map " + }, + { + "bbox": [ + 305, + 217, + 545, + 253 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{fuse}}\\in \\mathbb{R}^{K}" + }, + { + "bbox": [ + 305, + 217, + 545, + 253 + ], + "type": "text", + "content": " as a byproduct in cross-attention:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 261, + 545, + 292 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 261, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 314, + 261, + 545, + 292 + ], + "type": "interline_equation", + "content": "A _ {i} ^ {\\text {f u s e}} = \\frac {e ^ {M _ {i}}}{\\sum_ {i = 1} ^ {K} e ^ {M _ {i}}} \\text {, w h e r e} M = \\frac {k \\left(\\Phi \\left(\\mathbf {x} ^ {\\mathcal {V}}\\right)\\right) q \\left(\\mathbf {x} ^ {\\mathcal {T}}\\right)}{\\sqrt {D _ {h}}}. \\tag {7}", + "image_path": "04a6f23a8d905489c0282ecfcce397e429a23dabeb1395c8e198dfe6f7503688.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 300, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 300, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 545, + 348 + ], + "type": "text", + "content": "Using cross-attention, the module effectively reasons the relevance between entities and the textual feature. Inferred relevance is described in the " + }, + { + "bbox": [ + 305, + 300, + 545, + 348 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{fuse}}" + }, + { + "bbox": [ + 305, + 300, + 545, + 348 + ], + "type": "text", + "content": ", which is the probability of each entity being referred to in the textual feature." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 356, + 370, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 356, + 370, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 356, + 370, + 369 + ], + "type": "text", + "content": "3.4. Training" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 372, + 544, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 372, + 544, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 372, + 544, + 396 + ], + "type": "text", + "content": "Our model is trained with two different loss functions: contrastive cycle-consistency " + }, + { + "bbox": [ + 305, + 372, + 544, + 396 + ], + "type": "inline_equation", + "content": "(\\mathbf{C}^3)" + }, + { + "bbox": [ + 305, + 372, + 544, + 396 + ], + "type": "text", + "content": " and reconstruction loss." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 406, + 507, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 406, + 507, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 406, + 507, + 418 + ], + "type": "text", + "content": "3.4.1 Contrastive cycle-consistency " + }, + { + "bbox": [ + 305, + 406, + 507, + 418 + ], + "type": "inline_equation", + "content": "(\\mathbf{C}^3)" + }, + { + "bbox": [ + 305, + 406, + 507, + 418 + ], + "type": "text", + "content": " Loss" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 419, + 545, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 545, + 527 + ], + "type": "text", + "content": "The output of our model is the cross-modal embedding, which is obtained by fusing information from visual and textual modalities. The fusion should selectively combine discovered entities based on textual information, but a segmentation label for which entities are referred to is not given in our weakly-supervised setting. To resolve this issue, we add an additional constraint to the contrastive learning objective: establishing cycle-consistency between each ground-truth pair of image and text." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 527, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 527, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 545, + 613 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 527, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{x}_i^\\nu, \\mathbf{x}_i^\\mathcal{T})\\}_{i=1}^B" + }, + { + "bbox": [ + 304, + 527, + 545, + 613 + ], + "type": "text", + "content": " represent a batch of visual and textual features, and " + }, + { + "bbox": [ + 304, + 527, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{ij} = \\Psi \\big(\\Phi(\\mathbf{x}_i^\\nu); \\mathbf{x}_j^\\mathcal{T}\\big)" + }, + { + "bbox": [ + 304, + 527, + 545, + 613 + ], + "type": "text", + "content": " denote the cross-modal embedding produced with " + }, + { + "bbox": [ + 304, + 527, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^\\nu" + }, + { + "bbox": [ + 304, + 527, + 545, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 527, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j^\\mathcal{T}" + }, + { + "bbox": [ + 304, + 527, + 545, + 613 + ], + "type": "text", + "content": ". For simplicity, we refer to the cross-modal embedding produced using the ground-truth image-text pair as matching embedding, and all other cross-modal embeddings as non-matching embeddings. Then " + }, + { + "bbox": [ + 304, + 527, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^3" + }, + { + "bbox": [ + 304, + 527, + 545, + 613 + ], + "type": "text", + "content": " loss is computed as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 345, + 622, + 545, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 622, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 345, + 622, + 545, + 656 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C} ^ {3}} = - \\frac {1}{B} \\sum_ {j = 1} ^ {B} \\log \\frac {e ^ {\\langle \\mathbf {z} _ {j j} , \\mathbf {s g} (\\mathbf {x} _ {j} ^ {\\mathcal {T}}) \\rangle}}{\\sum_ {i = 1} ^ {B} e ^ {\\langle \\mathbf {z} _ {i j} , \\mathbf {s g} (\\mathbf {x} _ {j} ^ {\\mathcal {T}}) \\rangle}} \\tag {8}", + "image_path": "8c8be9aa2a72d2c45869d39ba88deb82f1a70bb12dfc78978ad1c4a18834ae42.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\langle \\cdot, \\cdot \\rangle" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathfrak{sg}(\\cdot)" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " denote inner product and stop-gradient operation, respectively. Stop-gradient operation is used to prevent the collapsing solutions where every embedding becomes similar, as discussed in [9]." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15551" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "The proposed loss function encourages the matching embeddings to be similar to the input textual feature, while making the non-matching embeddings to be dissimilar. This approach ensures that the modality fusion module preserves the inherent information of the referring expression only when it is paired with the matching image (cycle-consistency), achieved by attending to the referred entities that may be absent in non-matching images. Therefore, an attention map " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{fuse}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": " is trained to reflect the referring expression accurately, where relevance between the expression and entities automatically emerges." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 213, + 167, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 167, + 223 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 167, + 223 + ], + "type": "text", + "content": "3.4.2 Reconstruction Loss" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 225, + 288, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 225, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 288, + 274 + ], + "type": "text", + "content": "If we can reconstruct the original data using only the set of slots, then the slots will correspond to distinctive individual entities in the input image. Therefore, we leverage the reconstruction objective " + }, + { + "bbox": [ + 46, + 225, + 288, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{recon}}" + }, + { + "bbox": [ + 46, + 225, + 288, + 274 + ], + "type": "text", + "content": ", which is given by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 279, + 287, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 279, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 77, + 279, + 287, + 312 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n}} = \\frac {1}{B} \\sum_ {i = 0} ^ {B} \\left\\| \\operatorname {s g} \\left(\\mathbf {x} _ {i} ^ {\\mathcal {Y}}\\right) - f _ {\\mathrm {d e c}} \\left(\\Phi \\left(\\mathbf {x} _ {i} ^ {\\mathcal {Y}}\\right)\\right) \\right\\| ^ {2}, \\tag {9}", + "image_path": "1ae27aa17599d1c088f2e05f8b80d2e8c4a0589b9982dccf0c8156fb88d7be9b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 317, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 287, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 317, + 287, + 389 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{dec}}(\\cdot)" + }, + { + "bbox": [ + 46, + 317, + 287, + 389 + ], + "type": "text", + "content": " is a MLP based decoder. Unlike the original slot attention [42], our reconstruction is performed in the feature space, rather than the pixel space. This feature reconstruction scheme enables an efficient decoder design. Detailed descriptions are provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 396, + 115, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 115, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 115, + 406 + ], + "type": "text", + "content": "3.5. Inference" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 411, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 555 + ], + "type": "text", + "content": "The segmentation mask is computed with the two attention maps: " + }, + { + "bbox": [ + 46, + 411, + 287, + 555 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{slot}} \\in \\mathbb{R}^{N \\times K}" + }, + { + "bbox": [ + 46, + 411, + 287, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 411, + 287, + 555 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{fuse}} \\in \\mathbb{R}^{K}" + }, + { + "bbox": [ + 46, + 411, + 287, + 555 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 411, + 287, + 555 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{slot}}" + }, + { + "bbox": [ + 46, + 411, + 287, + 555 + ], + "type": "text", + "content": " is patchwise attention map that contains information about the discovered entities (Eq. (3)), while " + }, + { + "bbox": [ + 46, + 411, + 287, + 555 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{fuse}}" + }, + { + "bbox": [ + 46, + 411, + 287, + 555 + ], + "type": "text", + "content": " is entity-wise attention map denoting relevance between the referring expression and the entities (Eq. (7)). To compute the segmentation mask, we first compute " + }, + { + "bbox": [ + 46, + 411, + 287, + 555 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{slot}} \\cdot A^{\\mathrm{fuse}}" + }, + { + "bbox": [ + 46, + 411, + 287, + 555 + ], + "type": "text", + "content": ", which results in a 1-dimensional vector of size " + }, + { + "bbox": [ + 46, + 411, + 287, + 555 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 411, + 287, + 555 + ], + "type": "text", + "content": ". We then reshape this vector into a 2-dimensional map. Next, we bilinearly interpolate the map to the original image size and rescale it using min-max normalization. Finally, we binarize the map using a threshold value " + }, + { + "bbox": [ + 46, + 411, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 411, + 287, + 555 + ], + "type": "text", + "content": " to obtain the segmentation mask." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 562, + 128, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 128, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 128, + 574 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 579, + 170, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 170, + 592 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 170, + 592 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "Datasets: Our experiments utilize four datasets commonly used in referring image segmentation task: RefCOCO [70], RefCOCO+ [70], Gref [45], and PhraseCut [61]. RefCOCO, RefCOCO+, and Gref are obtained from COCO [39] dataset. RefCOCO contains 142,209 image-text pairs, while RefCOCO+ has 141,564 pairs and does not include expressions indicating orientation properties (e.g., top, right). Gref includes 104,560 pairs for 54,822 objects. PhraseCut comprises 345,486 image-text pairs, which are derived from Visual Genome [35]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": "Implementation details: For visual and textual backbone, we employ ViT-S-16 pretrained on ImageNet-21K [11] and pretrained BERT [13], respectively. We make each modality feature have the same channel dimension " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": " by applying a " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": " convolution layer. We set " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "D_h" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": " to 512 and 1024, respectively. In the aggregation block, the number of iteration " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": " is set to 6, and the number of slots " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": " is set to 36 except for PhraseCut, which is set to 60. For the entity slot, the number of slots sampled from each Gaussian distributions " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "K_s" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": " is set to 3 in RefCOCO+ and Gref, while 2 in RefCOCO and PhraseCut. During training, we use AdamW [44] optimizer, with the initial learning rate of 1e-4 scheduled with cosine annealing [43]. Batch size is set to 32 and model is trained for 50 epochs. We resize input images to " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": " for both training and evaluation. Threshold value " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": " is set to 0.5 for all experiments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 265, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 545, + 350 + ], + "type": "text", + "content": "Evaluation metric: Following [67], we employ the cumulative intersection-over-union (cIoU) and mean intersection-over-Union (mIoU) metric; cIoU divides the total intersections by the total unions across all test images, and mIoU computes IoU for each image and averages them. Also, we evaluate the accuracy at the IoU thresholds of " + }, + { + "bbox": [ + 304, + 265, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\{0.3, 0.5, 0.7\\}" + }, + { + "bbox": [ + 304, + 265, + 545, + 350 + ], + "type": "text", + "content": ", denote as " + }, + { + "bbox": [ + 304, + 265, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{A} @ \\{0.3, 0.5, 0.7\\}" + }, + { + "bbox": [ + 304, + 265, + 545, + 350 + ], + "type": "text", + "content": " in Table 3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 361, + 506, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 361, + 506, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 506, + 373 + ], + "type": "text", + "content": "4.2. Comparisons with the State of the Art" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 376, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 460 + ], + "type": "text", + "content": "Weakly supervised methods: As presented in Table 1, our method outperforms TSEG [53], the previous work on weakly-supervised referring image segmentation, in all benchmarks. It shows that the proposed bottom-up and top-down attention framework is more effective than using patch-based visual features with a carefully designed similarity function of TSEG." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "content": "Open-vocabulary segmentation: In Table 1, we also compare our model with the recent open-sourced open-vocabulary segmentation models: GroupViT [64] and MaskCLIP [73], on the referring image segmentation benchmarks. Note that these methods are pre-trained with large-scale image-text datasets [5, 47, 54]. We also report their fine-tuned version for fair comparisons: models are additionally trained with the image-text pairs of each benchmark. Further details for the reproductions are provided in the supplementary material. Ours outperforms GroupViT and MaskCLIP in all settings and benchmarks, even without large-scale image-text pretraining. This result shows that open-vocabulary segmentation models [64, 73] have difficulty comprehending the complex free-from text involving the relation between entities, unlike ours. Interestingly, our model without fine-tuning, which pre-trained on the PhraseCut dataset and evaluated on other benchmarks, still achieved better results. This setting is identical to the open-vocabulary segmentation models, except it leverages a far smaller dataset. These results demonstrate the feasibility of our method in a zero-shot transfer setting." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15552" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 70, + 520, + 185 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 520, + 185 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 520, + 185 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 520, + 185 + ], + "type": "table", + "html": "
MethodsImage-text pretraining datasetFine tuningRefCOCORefCOCO+Gref valPhraseCut val
valtestAtestBvaltestAtestB
GroupViT [64]CC12M+YFCC (26M)X7.996.1610.518.496.7910.5910.685.82
✓10.8211.1111.2911.1410.7811.8412.779.41
MaskCLIP [73]WIT (400M)X11.5211.8512.0611.8712.0112.5712.7412.13
✓19.4518.6921.3719.9718.9321.4821.1123.80
TSEG [53]-✓25.44--22.01--22.0528.77
OursVisual Genome (0.31M)X21.8019.0024.9622.2019.8624.8525.89-
-✓34.7634.5835.0128.4828.6027.9828.8733.45
", + "image_path": "5951cccd42e601f5ae60924e30314009ec498f5f4d9d314ec6492ccd87d22fcd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 230, + 545, + 326 + ], + "blocks": [ + { + "bbox": [ + 47, + 230, + 545, + 326 + ], + "lines": [ + { + "bbox": [ + 47, + 230, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 47, + 230, + 545, + 326 + ], + "type": "image", + "image_path": "14f576d7b7ed4a9f9d9f63607e371c8fd82debd670a289241a43d13bdece0c3b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 326, + 203, + 337 + ], + "lines": [ + { + "bbox": [ + 47, + 326, + 203, + 337 + ], + "spans": [ + { + "bbox": [ + 47, + 326, + 203, + 337 + ], + "type": "text", + "content": "Query: \"Guy on left with beard and guitar\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 338, + 546, + 376 + ], + "lines": [ + { + "bbox": [ + 46, + 338, + 546, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 546, + 376 + ], + "type": "text", + "content": "Figure 4: Qualitative results of our framework with entity slot (Ours), random slot, and query slot on RefCOCO val set. In the case of entity slots, the boundary color indicates the Gaussian distribution that each slot is sampled from. For each slot type, we present 10 discovered entities from " + }, + { + "bbox": [ + 46, + 338, + 546, + 376 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{slot}}" + }, + { + "bbox": [ + 46, + 338, + 546, + 376 + ], + "type": "text", + "content": " and the final predictions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 390, + 141, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 390, + 141, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 390, + 141, + 403 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 406, + 288, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 288, + 441 + ], + "type": "text", + "content": "We perform ablation studies on the val set of RefCOCO to analyze the contributions of the entity discovery module and the modality fusion module." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 287, + 552 + ], + "type": "text", + "content": "Entity discovery module: In Table 2(a), we ablate the entity discovery module and use visual feature directly, which leads to the drastical performance drop. This result demonstrates that contribution of our bottom-up attention framework is significant. Moreover, we ablate the interaction block, replace slot-attention of aggregation block to transformer, and train the model without reconstruction loss. The results show that all of the design choices for the module is important for effective entity discovery." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 555, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 639 + ], + "type": "text", + "content": "Modality fusion module: In Table 2(b), the modality fusion module is ablated, and the model is trained with image-text contrastive learning [47], where the image feature is obtained by pooling the entity embeddings. The result demonstrates that inferring relevance between entities and referring expression with modality fusion module, trained with " + }, + { + "bbox": [ + 46, + 555, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\mathrm{C}^3" + }, + { + "bbox": [ + 46, + 555, + 287, + 639 + ], + "type": "text", + "content": " loss, is crucial for accurate segmentation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 650, + 153, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 650, + 153, + 663 + ], + "spans": [ + { + "bbox": [ + 47, + 650, + 153, + 663 + ], + "type": "text", + "content": "4.4. In-depth Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Analysis of slot initialization: To verify our design choice for entity slots, we investigate variants of slots; query, random, and entity slot. As summarized in Table 3, our entity slot outperforms other slot types across the different" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 314, + 388, + 536, + 521 + ], + "blocks": [ + { + "bbox": [ + 46, + 187, + 547, + 225 + ], + "lines": [ + { + "bbox": [ + 46, + 187, + 547, + 225 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 547, + 225 + ], + "type": "text", + "content": "Table 1: Comparison with weakly-supervised method (TSEG) and open-vocabulary segmentation method (GroupViT and MaskCLIP). The results on four datasets are reported in mIoU (\\%). Fine-tuning " + }, + { + "bbox": [ + 46, + 187, + 547, + 225 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 46, + 187, + 547, + 225 + ], + "type": "text", + "content": "means that the model is trained with the image-text pairs of the target benchmark; otherwise, it just relies on the pre-trained weight without further training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 388, + 536, + 521 + ], + "lines": [ + { + "bbox": [ + 314, + 388, + 536, + 521 + ], + "spans": [ + { + "bbox": [ + 314, + 388, + 536, + 521 + ], + "type": "table", + "html": "
cIoUmIoU
(a) Analysis on entity discovery module
Ours30.4034.76
w/o entity discovery module5.405.91
w/o interaction block29.5933.69
aggregation block w/ transformer11.2912.61
w/o reconstruction loss18.6121.84
(b) Analysis on modality fusion module
Ours30.4034.76
w/o modality fusion module and C3loss13.8114.65
", + "image_path": "a4d72222ca8e02c37425656de18b78a940cb30d59f4ee316cdd862645f54eced.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 523, + 545, + 547 + ], + "lines": [ + { + "bbox": [ + 306, + 523, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 306, + 523, + 545, + 547 + ], + "type": "text", + "content": "Table 2: Ablation studies of model components and loss function on RefCOCO val set." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "numbers of slots. Specifically, as the number of slots is increased, the performance of the query slot declines, and that of the random slot converges. This result implies that simply increasing the number of slots can not handle the discovery of fine-grained visual entities in real-world images, and thus gains from the proposed entity slot type are non-trivial. In Fig. 4, we also present several visual entities discovered using each slot type, which show the superiority of the entity slot against others as explained in Sec. 3.2. Specifically, the entity slots from the same distribution share the same semantic property (e.g., head, body, or window) while capturing different fine-grained entities. The pro" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15553" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 72, + 214, + 151 + ], + "blocks": [ + { + "bbox": [ + 91, + 72, + 214, + 151 + ], + "lines": [ + { + "bbox": [ + 91, + 72, + 214, + 151 + ], + "spans": [ + { + "bbox": [ + 91, + 72, + 214, + 151 + ], + "type": "image", + "image_path": "9d4bd5fa503bc668a63907b6248fc962f30b8551aea5ddad4eb76ef3df679417.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 298, + 72, + 462, + 155 + ], + "blocks": [ + { + "bbox": [ + 219, + 72, + 296, + 150 + ], + "lines": [ + { + "bbox": [ + 219, + 72, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 219, + 72, + 296, + 150 + ], + "type": "text", + "content": "Prediction Ground Truth" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 298, + 72, + 462, + 155 + ], + "lines": [ + { + "bbox": [ + 298, + 72, + 462, + 155 + ], + "spans": [ + { + "bbox": [ + 298, + 72, + 462, + 155 + ], + "type": "image", + "image_path": "3afde1d371f944240b051ee83bfb66269cfe19c9242be78ea73a44e183b3a864.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 467, + 72, + 544, + 79 + ], + "lines": [ + { + "bbox": [ + 467, + 72, + 544, + 79 + ], + "spans": [ + { + "bbox": [ + 467, + 72, + 544, + 79 + ], + "type": "text", + "content": "Prediction Ground Truth" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 79, + 91, + 151 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 89, + 79 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 89, + 79 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 89, + 79 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 79, + 91, + 151 + ], + "lines": [ + { + "bbox": [ + 50, + 79, + 91, + 151 + ], + "spans": [ + { + "bbox": [ + 50, + 79, + 91, + 151 + ], + "type": "image", + "image_path": "3876b1fdd254e5d17cd010b355926aa0b6a6cf817368d1732c573d054817d73b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 151, + 112, + 157 + ], + "lines": [ + { + "bbox": [ + 91, + 151, + 112, + 157 + ], + "spans": [ + { + "bbox": [ + 91, + 151, + 112, + 157 + ], + "type": "text", + "content": "on left\"" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 91, + 164, + 132, + 195 + ], + "blocks": [ + { + "bbox": [ + 91, + 164, + 132, + 195 + ], + "lines": [ + { + "bbox": [ + 91, + 164, + 132, + 195 + ], + "spans": [ + { + "bbox": [ + 91, + 164, + 132, + 195 + ], + "type": "image", + "image_path": "c2e2055c93d7bd30bc89a7bce71af6290a47e9f92c8017fd9713a7c3b1a2200f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 50, + 196, + 143, + 205 + ], + "lines": [ + { + "bbox": [ + 50, + 196, + 143, + 205 + ], + "spans": [ + { + "bbox": [ + 50, + 196, + 143, + 205 + ], + "type": "text", + "content": "Query: \"Right person glasses\"" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 209, + 546, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 546, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 546, + 232 + ], + "type": "text", + "content": "Figure 5: Qualitative results of our framework on RefCOCO val set. We present the discovered entities from " + }, + { + "bbox": [ + 46, + 209, + 546, + 232 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{slot}}" + }, + { + "bbox": [ + 46, + 209, + 546, + 232 + ], + "type": "text", + "content": " and their relevance scores from " + }, + { + "bbox": [ + 46, + 209, + 546, + 232 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{slot}}" + }, + { + "bbox": [ + 46, + 209, + 546, + 232 + ], + "type": "text", + "content": ". Top-3 entities in terms of relevance to query expression are presented." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 132, + 165, + 173, + 195 + ], + "blocks": [ + { + "bbox": [ + 132, + 165, + 173, + 195 + ], + "lines": [ + { + "bbox": [ + 132, + 165, + 173, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 165, + 173, + 195 + ], + "type": "image", + "image_path": "c6aee612e77cd84364faccf5ac4672f5864255d047d7a90f37b3d79898198147.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 173, + 165, + 214, + 195 + ], + "blocks": [ + { + "bbox": [ + 173, + 165, + 214, + 195 + ], + "lines": [ + { + "bbox": [ + 173, + 165, + 214, + 195 + ], + "spans": [ + { + "bbox": [ + 173, + 165, + 214, + 195 + ], + "type": "image", + "image_path": "175c1b3fa9efcb4d93edab29328c248837500432843d6bf3637db927d00989e2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 214, + 165, + 255, + 195 + ], + "blocks": [ + { + "bbox": [ + 214, + 165, + 255, + 195 + ], + "lines": [ + { + "bbox": [ + 214, + 165, + 255, + 195 + ], + "spans": [ + { + "bbox": [ + 214, + 165, + 255, + 195 + ], + "type": "image", + "image_path": "42dc45321ec730e96b3588542271f6b5a7121fd952f25bf07cf55bc86eef8615.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 255, + 165, + 296, + 195 + ], + "blocks": [ + { + "bbox": [ + 255, + 165, + 296, + 195 + ], + "lines": [ + { + "bbox": [ + 255, + 165, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 255, + 165, + 296, + 195 + ], + "type": "image", + "image_path": "e9ac7d06d18138714c5a7b575e275b5de39907dc979823778807870c43b939c9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 298, + 165, + 338, + 195 + ], + "blocks": [ + { + "bbox": [ + 298, + 165, + 338, + 195 + ], + "lines": [ + { + "bbox": [ + 298, + 165, + 338, + 195 + ], + "spans": [ + { + "bbox": [ + 298, + 165, + 338, + 195 + ], + "type": "image", + "image_path": "e57ad991d52bab9adc68a4eaf4cdd9961b9b900f3033ff3f0437dd93eea64566.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 196, + 370, + 205 + ], + "lines": [ + { + "bbox": [ + 299, + 196, + 370, + 205 + ], + "spans": [ + { + "bbox": [ + 299, + 196, + 370, + 205 + ], + "type": "text", + "content": "Query: \"Jeans on right\"" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 323, + 545, + 346 + ], + "lines": [ + { + "bbox": [ + 306, + 323, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 545, + 346 + ], + "type": "text", + "content": "Table 4: Impact of hyperparameters " + }, + { + "bbox": [ + 306, + 323, + 545, + 346 + ], + "type": "inline_equation", + "content": "K_{g}" + }, + { + "bbox": [ + 306, + 323, + 545, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 323, + 545, + 346 + ], + "type": "inline_equation", + "content": "K_{s}" + }, + { + "bbox": [ + 306, + 323, + 545, + 346 + ], + "type": "text", + "content": " (a) and inference schemes (b) on RefCOCO val set." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 339, + 165, + 373, + 195 + ], + "blocks": [ + { + "bbox": [ + 339, + 165, + 373, + 195 + ], + "lines": [ + { + "bbox": [ + 339, + 165, + 373, + 195 + ], + "spans": [ + { + "bbox": [ + 339, + 165, + 373, + 195 + ], + "type": "image", + "image_path": "509784f85b9678d1583c72ce5453ab7a950f233d63083e48dab53dd6eb99d302.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 373, + 165, + 421, + 195 + ], + "blocks": [ + { + "bbox": [ + 373, + 165, + 421, + 195 + ], + "lines": [ + { + "bbox": [ + 373, + 165, + 421, + 195 + ], + "spans": [ + { + "bbox": [ + 373, + 165, + 421, + 195 + ], + "type": "image", + "image_path": "0e953d3131db8f854a3fdcbebbbcfb0f0f55a1dc23a97e4339d30654543bf73f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 421, + 165, + 462, + 195 + ], + "blocks": [ + { + "bbox": [ + 421, + 165, + 462, + 195 + ], + "lines": [ + { + "bbox": [ + 421, + 165, + 462, + 195 + ], + "spans": [ + { + "bbox": [ + 421, + 165, + 462, + 195 + ], + "type": "image", + "image_path": "32446252ddc2a20e3b8394db36d4563ea5a92dc3fce8d448abc599b340c08e82.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 462, + 165, + 503, + 195 + ], + "blocks": [ + { + "bbox": [ + 462, + 165, + 503, + 195 + ], + "lines": [ + { + "bbox": [ + 462, + 165, + 503, + 195 + ], + "spans": [ + { + "bbox": [ + 462, + 165, + 503, + 195 + ], + "type": "image", + "image_path": "432a42f85675f372253e8afeb12cfab213b85484c93f65338685787316c95333.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 503, + 165, + 544, + 195 + ], + "blocks": [ + { + "bbox": [ + 503, + 165, + 544, + 195 + ], + "lines": [ + { + "bbox": [ + 503, + 165, + 544, + 195 + ], + "spans": [ + { + "bbox": [ + 503, + 165, + 544, + 195 + ], + "type": "image", + "image_path": "c1573154031eb0f536c2de8ca95a455eb49c1ac14f6c1ac7366e16ab3449ebbd.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 52, + 243, + 280, + 356 + ], + "blocks": [ + { + "bbox": [ + 52, + 243, + 280, + 356 + ], + "lines": [ + { + "bbox": [ + 52, + 243, + 280, + 356 + ], + "spans": [ + { + "bbox": [ + 52, + 243, + 280, + 356 + ], + "type": "table", + "html": "
Slot type# slotsA@0.3A@0.5A@0.7cIoUmIoU
Query1227.577.170.9622.0421.44
2436.7310.611.9323.5924.78
3619.303.480.3116.6618.16
Random1236.3011.272.6924.2026.65
2445.1916.013.6626.6029.87
3645.6216.883.7626.5530.15
Entity(Ours)1239.4313.873.5624.7227.84
2452.3322.976.2329.2233.54
3655.0224.996.3530.4034.76
", + "image_path": "b896815906e2352274e45bb4f5f1ed1bdccb5062e49c40f9125e741be9fcffb2.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 398, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 287, + 458 + ], + "type": "text", + "content": "posed entity slots effectively discover fine-grained visual entities, which results in an accurate prediction. In contrast, the prediction using random slots leads to a noisy mask, and using query slots leads to the mask of every head that appears in the image, respectively." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "type": "text", + "content": "Effect of hyperparameters " + }, + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "type": "inline_equation", + "content": "K_{g}" + }, + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "type": "inline_equation", + "content": "K_{s}" + }, + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "type": "text", + "content": ": In Table 4(a), we investigate the impact of hyperparameters for the entity slots: " + }, + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "type": "inline_equation", + "content": "K_{g}" + }, + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "type": "inline_equation", + "content": "K_{s}" + }, + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "type": "text", + "content": ", which are the number of Gaussian distributions and slots sampled from each distribution, respectively. We keep the total number of entity slots to 36, only changing " + }, + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "type": "inline_equation", + "content": "K_{g}" + }, + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "type": "inline_equation", + "content": "K_{s}" + }, + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "type": "text", + "content": ". The optimal configurations are 18 and 2 for " + }, + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "type": "inline_equation", + "content": "K_{g}" + }, + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "type": "inline_equation", + "content": "K_{s}" + }, + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "type": "text", + "content": ", respectively. Interestingly, we observed that even entity slots with " + }, + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "type": "inline_equation", + "content": "K_{s} = 1" + }, + { + "bbox": [ + 46, + 460, + 287, + 592 + ], + "type": "text", + "content": ", where a single slot is sampled from each distribution, vastly surpass the query slots in Table 3; it shows that random sampling of slots prevents the bounding to specific semantic category." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 593, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 713 + ], + "type": "text", + "content": "Effectiveness of " + }, + { + "bbox": [ + 46, + 593, + 287, + 713 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{fuse}}" + }, + { + "bbox": [ + 46, + 593, + 287, + 713 + ], + "type": "text", + "content": ": To verify the efficacy of " + }, + { + "bbox": [ + 46, + 593, + 287, + 713 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{fuse}}" + }, + { + "bbox": [ + 46, + 593, + 287, + 713 + ], + "type": "text", + "content": " quantitatively, we evaluate our model with the following inference scheme variants: (1) AVG that replaces " + }, + { + "bbox": [ + 46, + 593, + 287, + 713 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{fuse}}" + }, + { + "bbox": [ + 46, + 593, + 287, + 713 + ], + "type": "text", + "content": " with the uniform probability distribution, (2) MAX that picks a single attention map from " + }, + { + "bbox": [ + 46, + 593, + 287, + 713 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{slot}}" + }, + { + "bbox": [ + 46, + 593, + 287, + 713 + ], + "type": "text", + "content": " with the maximum probability in " + }, + { + "bbox": [ + 46, + 593, + 287, + 713 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{fuse}}" + }, + { + "bbox": [ + 46, + 593, + 287, + 713 + ], + "type": "text", + "content": ", and (3) MIN that picks the attention map with the minimum probability. As summarized in Table 4(b), MAX largely outperforms AVG even though it only uses a single attention map for the final prediction, while MIN results in a near-zero cIoU and mIoU. This tendency shows that " + }, + { + "bbox": [ + 46, + 593, + 287, + 713 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{fuse}}" + } + ] + } + ], + "index": 25 + }, + { + "type": "table", + "bbox": [ + 309, + 243, + 403, + 308 + ], + "blocks": [ + { + "bbox": [ + 47, + 358, + 286, + 381 + ], + "lines": [ + { + "bbox": [ + 47, + 358, + 286, + 381 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 286, + 381 + ], + "type": "text", + "content": "Table 3: Comparison between different slot types on Ref-COCO val set, where # slots is the total number of slots." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 243, + 403, + 308 + ], + "lines": [ + { + "bbox": [ + 309, + 243, + 403, + 308 + ], + "spans": [ + { + "bbox": [ + 309, + 243, + 403, + 308 + ], + "type": "table", + "html": "
KgKscIoUmIoU
36129.3633.41
18230.4034.76
12329.6334.03
9428.3132.34
6627.2331.30
", + "image_path": "c6a627b03b40b4e9b217290888f875808b1be263246e52b9f985bf1f8fa7e26c.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "type": "table", + "bbox": [ + 409, + 244, + 538, + 308 + ], + "blocks": [ + { + "bbox": [ + 350, + 310, + 361, + 319 + ], + "lines": [ + { + "bbox": [ + 350, + 310, + 361, + 319 + ], + "spans": [ + { + "bbox": [ + 350, + 310, + 361, + 319 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 409, + 244, + 538, + 308 + ], + "lines": [ + { + "bbox": [ + 409, + 244, + 538, + 308 + ], + "spans": [ + { + "bbox": [ + 409, + 244, + 538, + 308 + ], + "type": "table", + "html": "
Inference schemecIoUmIoU
AVG9.018.16
MAX14.9319.54
MIN0.160.14
A fuse · A slot (Ours)30.4034.76
", + "image_path": "2653be3e59be38b6aa7412e54da24af530b1bb934cc553f834fbc17bb07e5cca.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 470, + 308, + 481, + 318 + ], + "lines": [ + { + "bbox": [ + 470, + 308, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 470, + 308, + 481, + 318 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 360, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 394 + ], + "type": "text", + "content": "effectively assigns probability based on the relevance between reference and entities, even without direct supervision." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 396, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 468 + ], + "type": "text", + "content": "Qualitative Analysis: In Fig. 5, predicted segmentation mask and discovered entities with top-3 relevance are presented. Results show that the entity discovery module effectively discovers primitive units of segmentation masks, and the modality fusion module accurately infers their relevance to query referring expression." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 476, + 378, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 378, + 488 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 378, + 488 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 492, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 545, + 647 + ], + "type": "text", + "content": "In this paper, we address the problem of learning referring image segmentation with only image-text pairs. We present a new model based on a bottom-up and top-down attention framework, which first discovers entities in the input image and then combines such entities based on their relevance to the text query. Moreover, we present a new loss function enforcing cycle-consistency between image-text pairs, which enables the effective training of the model without any further supervision. As a result, the proposed method clearly outperforms an existing weakly-supervised method and recent open-vocabulary segmentation models. For future work, we will explore expanding our model to different modalities such as video and audio." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the NRF grant and the IITP grants funded by Ministry of Science and ICT, Korea (NRF-2021R1A2C3012728-20%, IITP-2020-0-00842-20%, IITP-2022-0-00290-20%, IITP-2022-0-00926-20%, IITP-2021-0-02068-10%, IITP-2019-0-01906-10%)." + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15554" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Jiwoon Ahn, Sunghyun Cho, and Suha Kwak. Weakly supervised learning of instance segmentation with inter-pixel relations. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 134, + 287, + 177 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 134, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 53, + 134, + 287, + 177 + ], + "type": "text", + "content": "[2] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 177, + 287, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 177, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 53, + 177, + 287, + 198 + ], + "type": "text", + "content": "[3] Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. volume abs/1607.06450, 2016. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 198, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 198, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 287, + 230 + ], + "type": "text", + "content": "[4] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In Proc. Neural Information Processing Systems (NeurIPS), 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 230, + 287, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 230, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 53, + 230, + 287, + 283 + ], + "type": "text", + "content": "[5] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "text", + "content": "[6] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In Proc. IEEE International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 327, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 327, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 327, + 287, + 369 + ], + "type": "text", + "content": "[7] Jianbo Chen, Yelong Shen, Jianfeng Gao, Jingjing Liu, and Xiaodong Liu. Language-based image editing with recurrent attentive models. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "text", + "content": "[8] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In Proc. Neural Information Processing Systems (NeurIPS), 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 413, + 287, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 287, + 444 + ], + "type": "text", + "content": "[9] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 444, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 498 + ], + "type": "text", + "content": "[10] Samyak Datta, Karan Sikka, Anirban Roy, Karuna Ahuja, Devi Parikh, and Ajay Divakaran. Align2ground: Weakly supervised phrase grounding guided by image-caption alignment. In Proc. IEEE International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 498, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 540 + ], + "type": "text", + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: a large-scale hierarchical image database. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2009. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 541, + 287, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 541, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 48, + 541, + 287, + 583 + ], + "type": "text", + "content": "[12] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers. In Proc. IEEE International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 584, + 287, + 637 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 584, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 48, + 584, + 287, + 637 + ], + "type": "text", + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Annual Conference of the North American Chapter of the Association for Computational Linguistics, 2019. 4, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 638, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 638, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 638, + 287, + 680 + ], + "type": "text", + "content": "[14] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proc. IEEE International Conference on Computer Vision (ICCV), 2021. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl-" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 707 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "vain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In Proc. International Conference on Learning Representations (ICLR), 2021. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 106, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 148 + ], + "type": "text", + "content": "[16] Guang Feng, Zhiwei Hu, Lihe Zhang, and Hutchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 148, + 545, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 148, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 307, + 148, + 545, + 201 + ], + "type": "text", + "content": "[17] Tsu-Jui Fu, Xin Eric Wang, Scott T. Grafton, Miguel P. Eckstein, and William Yang Wang. M3l: Language-based video editing via multi-modal multi-level transformers. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 201, + 545, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 201, + 545, + 244 + ], + "spans": [ + { + "bbox": [ + 307, + 201, + 545, + 244 + ], + "type": "text", + "content": "[18] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In Proc. European Conference on Computer Vision (ECCV), 2022. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 245, + 545, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 245, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 545, + 288 + ], + "type": "text", + "content": "[19] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 288, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 288, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 288, + 545, + 331 + ], + "type": "text", + "content": "[20] Tanmay Gupta, Arash Vahdat, Gal Chechik, Xiaodong Yang, Jan Kautz, and Derek Hoiem. Contrastive learning for weakly supervised phrase grounding. In European Conference on Computer Vision. Springer, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 331, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 373 + ], + "type": "text", + "content": "[21] S. Hong, D. Yeo, S. Kwak, H. Lee, and B. Han. Weakly supervised semantic segmentation using web-crawled videos. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2224-2232, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 373, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 373, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 373, + 545, + 415 + ], + "type": "text", + "content": "[22] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In Proc. European Conference on Computer Vision (ECCV), 2016. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 416, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 545, + 469 + ], + "type": "text", + "content": "[23] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 470, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 470, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 470, + 545, + 514 + ], + "type": "text", + "content": "[24] Binbin Huang, Dongze Lian, Weixin Luo, and Shenghua Gao. Look before you leap: Learning landmark features for one-stage visual grounding. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 514, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 545, + 567 + ], + "type": "text", + "content": "[25] Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoqi Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 567, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 620 + ], + "type": "text", + "content": "[26] Zilong Huang, Xinggang Wang, Jiasi Wang, Wenyu Liu, and Jingdong Wang. Weakly-supervised semantic segmentation network with deep seeded region growing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 620, + 545, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 620, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 307, + 620, + 545, + 663 + ], + "type": "text", + "content": "[27] Tianrui Hui, Si Liu, Shaofei Huang, Guanbin Li, Sansi Yu, Faxi Zhang, and Jizhong Han. Linguistic structure guided context modeling for referring image segmentation. In Proc. European Conference on Computer Vision (ECCV), 2020. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 663, + 545, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 663, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 307, + 663, + 545, + 707 + ], + "type": "text", + "content": "[28] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proc. International" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "15555" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 707 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 68, + 73, + 262, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 73, + 262, + 83 + ], + "spans": [ + { + "bbox": [ + 68, + 73, + 262, + 83 + ], + "type": "text", + "content": "Conference on Machine Learning (ICML), 2021. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 83, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 83, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 83, + 287, + 137 + ], + "type": "text", + "content": "[29] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proc. IEEE International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 137, + 287, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 137, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 48, + 137, + 287, + 180 + ], + "type": "text", + "content": "[30] A. Khoreva, R. Benenson, J. Hosang, M. Hein, and B. Schiele. Simple does it: Weakly supervised instance and semantic segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 180, + 287, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 180, + 287, + 244 + ], + "spans": [ + { + "bbox": [ + 48, + 180, + 287, + 244 + ], + "type": "text", + "content": "[31] Beomyoung Kim, Youngjoon Yoo, Chae Eun Rhee, and Junmo Kim. Beyond semantic to instance segmentation: Weakly-supervised instance segmentation via semantic knowledge transfer and self-refinement. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 245, + 287, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 245, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 48, + 245, + 287, + 276 + ], + "type": "text", + "content": "[32] Dongwon Kim, Namyup Kim, and Suha Kwak. Improving cross-modal retrieval with set of diverse embeddings. 2022. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 277, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 287, + 330 + ], + "type": "text", + "content": "[33] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Restr: Convolution-free referring image segmentation using transformers. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 373 + ], + "type": "text", + "content": "[34] Alexander Kolesnikov and Christoph H. Lampert. Seed, expand and constrain: Three principles for weakly-supervised image segmentation. In Proc. European Conference on Computer Vision (ECCV), 2016. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 373, + 287, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 373, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 48, + 373, + 287, + 438 + ], + "type": "text", + "content": "[35] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, Michael Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision (IJCV), 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 438, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 438, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 438, + 287, + 480 + ], + "type": "text", + "content": "[36] Suha Kwak, Seunghoon Hong, and Bohyung Han. Weakly supervised semantic segmentation using superpixel pooling network. In Proc. AAAI Conference on Artificial Intelligence (AAAI), 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 481, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 524 + ], + "type": "text", + "content": "[37] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In Proc. International Conference on Learning Representations (ICLR), 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 524, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 524, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 524, + 287, + 577 + ], + "type": "text", + "content": "[38] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 577, + 287, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 577, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 48, + 577, + 287, + 631 + ], + "type": "text", + "content": "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: common objects in context. In Proc. European Conference on Computer Vision (ECCV), 2014. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 632, + 287, + 674 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 632, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 48, + 632, + 287, + 674 + ], + "type": "text", + "content": "[40] Chenxi Liu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, and Alan Yuille. Recurrent multimodal interaction for referring image segmentation. In Proc. IEEE International Conference on Computer Vision (ICCV), 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 674, + 287, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 674, + 287, + 707 + ], + "spans": [ + { + "bbox": [ + 48, + 674, + 287, + 707 + ], + "type": "text", + "content": "[41] Yongfei Liu, Bo Wan, Lin Ma, and Xuming He. Relation-aware instance refinement for weakly supervised visual grounding. In Proc. IEEE Conference on Computer Vision" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 705 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 480, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 480, + 83 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 480, + 83 + ], + "type": "text", + "content": "and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 83, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 83, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 308, + 83, + 545, + 137 + ], + "type": "text", + "content": "[42] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. In Proc. Neural Information Processing Systems (NeurIPS), 2020. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 137, + 545, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 137, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 308, + 137, + 545, + 169 + ], + "type": "text", + "content": "[43] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In Proc. International Conference on Learning Representations (ICLR), 2017. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 169, + 545, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 169, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 308, + 169, + 545, + 201 + ], + "type": "text", + "content": "[44] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. Proc. International Conference on Learning Representations (ICLR), 2019. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 201, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 201, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 308, + 201, + 545, + 255 + ], + "type": "text", + "content": "[45] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 255, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 255, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 255, + 545, + 297 + ], + "type": "text", + "content": "[46] Edgar Margffoy-Tuay, Juan C Pérez, Emilio Botero, and Pablo Arbeláez. Dynamic multimodal instance segmentation guided by natural language queries. In Proc. European Conference on Computer Vision (ECCV), 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 297, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 363 + ], + "type": "text", + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. International Conference on Machine Learning (ICML), 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 363, + 545, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 363, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 308, + 363, + 545, + 395 + ], + "type": "text", + "content": "[48] Tal Remez, Jonathan Huang, and Matthew Brown. Learning to segment via cut-and-paste. In Proc. European Conference on Computer Vision (ECCV), 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 395, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 395, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 395, + 545, + 437 + ], + "type": "text", + "content": "[49] Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, and Bernt Schiele. Grounding of textual phrases in images by reconstruction. In Proc. European Conference on Computer Vision (ECCV), 2016. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "type": "text", + "content": "[50] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 491, + 545, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 491, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 308, + 491, + 545, + 534 + ], + "type": "text", + "content": "[51] Hengcan Shi, Hongliang Li, Fanman Meng, and Qingbo Wu. Key-word-aware network for referring expression image segmentation. In Proc. European Conference on Computer Vision (ECCV), 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 534, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 566 + ], + "type": "text", + "content": "[52] Gautam Singh, Fei Deng, and Sungjin Ahn. Illiterate dalle learns to compose. In Proc. International Conference on Learning Representations (ICLR), 2021. 3, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 566, + 545, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 566, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 308, + 566, + 545, + 598 + ], + "type": "text", + "content": "[53] Robin Strudel, Ivan Laptev, and Cordelia Schmid. Weakly-supervised segmentation of referring expressions. arXiv preprint arXiv:2205.04725, 2022. 2, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 598, + 545, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 598, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 308, + 598, + 545, + 640 + ], + "type": "text", + "content": "[54] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 2016. 2, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 640, + 545, + 684 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 640, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 308, + 640, + 545, + 684 + ], + "type": "text", + "content": "[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. Neural Information Processing Systems (NeurIPS), 2017. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 684, + 545, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 684, + 545, + 705 + ], + "spans": [ + { + "bbox": [ + 308, + 684, + 545, + 705 + ], + "type": "text", + "content": "[56] Liwei Wang, Jing Huang, Yin Li, Kun Xu, Zhengyuan Yang, and Dong Yu. Improving weakly supervised visual ground" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "15556" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 708 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "ing by contrastive knowledge distillation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 170 + ], + "type": "text", + "content": "[57] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 170, + 287, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 170, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 48, + 170, + 287, + 213 + ], + "type": "text", + "content": "[58] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 213, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 213, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 48, + 213, + 287, + 267 + ], + "type": "text", + "content": "[59] Yunchao Wei, Jiashi Feng, Xiaodan Liang, Ming-Ming Cheng, Yao Zhao, and Shuicheng Yan. Object region mining with adversarial erasing: A simple classification to semantic segmentation approach. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 267, + 287, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 267, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 48, + 267, + 287, + 310 + ], + "type": "text", + "content": "[60] Philippe Weinzaepfel, Thomas Lucas, Diane Larlus, and Yannis Kalantidis. Learning super-features for image retrieval. In Proc. International Conference on Learning Representations (ICLR), 2022. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "type": "text", + "content": "[61] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasecut: Language-based image segmentation in the wild. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "type": "text", + "content": "[62] Yongqin Xian, Subhabrata Choudhury, Yang He, Bernt Schiele, and Zeynep Akata. Semantic projection network for zero-and few-label semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8256-8265, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 407, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 407, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 407, + 287, + 449 + ], + "type": "text", + "content": "[63] Fanyi Xiao, Leonid Sigal, and Yong Jae Lee. Weakly-supervised visual grounding of phrases with linguistic structures. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 503 + ], + "type": "text", + "content": "[64] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "type": "text", + "content": "[65] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token transformer for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4310-4319, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 610 + ], + "type": "text", + "content": "[66] Li Yang, Yan Xu, Chunfeng Yuan, Wei Liu, Bing Li, and Weiming Hu. Improving visual grounding with visual-linguistic verification and iterative reasoning. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 610, + 287, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 610, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 48, + 610, + 287, + 665 + ], + "type": "text", + "content": "[67] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 665, + 287, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 665, + 287, + 708 + ], + "spans": [ + { + "bbox": [ + 48, + 665, + 287, + 708 + ], + "type": "text", + "content": "[68] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 438 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[69] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 127, + 545, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 127, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 307, + 127, + 545, + 170 + ], + "type": "text", + "content": "[70] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Proc. European Conference on Computer Vision (ECCV), 2016. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 170, + 545, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 170, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 307, + 170, + 545, + 202 + ], + "type": "text", + "content": "[71] Dong Zhang, Hanwang Zhang, Jinhui Tang, Xian-Sheng Hua, and Qianru Sun. Causal intervention for weakly-supervised semantic segmentation. 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 202, + 545, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 202, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 307, + 202, + 545, + 245 + ], + "type": "text", + "content": "[72] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba. Learning deep features for discriminative localization. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 245, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 245, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 545, + 277 + ], + "type": "text", + "content": "[73] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In Proc. European Conference on Computer Vision (ECCV), 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 277, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 277, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 545, + 341 + ], + "type": "text", + "content": "[74] Yi Zhou, Hui Zhang, Hana Lee, Shuyang Sun, Pingjun Li, Yangguang Zhu, ByungIn Yoo, Xiaojuan Qi, and Jae-Joon Han. Slot-vps: Object-centric representation learning for video panoptic segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 341, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 341, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 341, + 545, + 384 + ], + "type": "text", + "content": "[75] Yanzhao Zhou, Yi Zhu, Qixiang Ye, Qiang Qiu, and Jianbin Jiao. Weakly supervised instance segmentation using class peak response. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 384, + 545, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 384, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 307, + 384, + 545, + 438 + ], + "type": "text", + "content": "[76] Yuliang Zou, Zizhao Zhang, Han Zhang, Chun-Liang Li, Xiao Bian, Jia-Bin Huang, and Tomas Pfister. Pseudoseg: Designing pseudo labels for semantic segmentation. In Proc. International Conference on Learning Representations (ICLR), 2021. 3" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15557" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Shift from Texture-bias to Shape-bias_ Edge Deformation-based Augmentation for Robust Object Recognition/e071fbaf-3036-4ac7-9bfd-fb6983ccf03f_content_list.json b/2023/Shift from Texture-bias to Shape-bias_ Edge Deformation-based Augmentation for Robust Object Recognition/e071fbaf-3036-4ac7-9bfd-fb6983ccf03f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b50b287a1113b08be4a5690f5759c3c43e3f7496 --- /dev/null +++ b/2023/Shift from Texture-bias to Shape-bias_ Edge Deformation-based Augmentation for Robust Object Recognition/e071fbaf-3036-4ac7-9bfd-fb6983ccf03f_content_list.json @@ -0,0 +1,1808 @@ +[ + { + "type": "text", + "text": "Shift from Texture-bias to Shape-bias: Edge Deformation-based Augmentation for Robust Object Recognition", + "text_level": 1, + "bbox": [ + 88, + 130, + 880, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xilin He $^{1,2,3}$ , Qinliang Lin $^{1,2,3}$ , Cheng Luo $^{1,2,3}$ Weicheng Xie $^{1,2,3}$ , Siyang Song $^{4}$ , Feng Liu $^{1,2,3}$ , Linlin Shen $^{1,2,3}$", + "bbox": [ + 84, + 210, + 962, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Computer Vision Institute, School of Computer Science & Software Engineering, Shenzhen University", + "bbox": [ + 116, + 229, + 934, + 248 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2Shenzhen Institute of Artificial Intelligence & Robotics for Society", + "bbox": [ + 259, + 248, + 789, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Guangdong Key Laboratory of Intelligent Information Processing", + "bbox": [ + 259, + 266, + 787, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ University of Leicester", + "bbox": [ + 428, + 284, + 622, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{2020152115, linqinliang2021, luocheng2020}@email.szu.edu.cn", + "bbox": [ + 259, + 303, + 789, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wcxie,feng.liu,llshen}@szu.edu.cn,ss1535@leicester.ac.uk", + "bbox": [ + 267, + 320, + 781, + 335 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 369, + 313, + 386 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent studies have shown the vulnerability of CNNs under perturbation noises, which is partially caused by the reason that the well-trained CNNs are too biased toward the object texture, i.e., they make predictions mainly based on texture cues. To reduce this texture-bias, current studies resort to learning augmented samples with heavily perturbed texture to make networks be more biased toward relatively stable shape cues. However, such methods usually fail to achieve real shape-biased networks due to the insufficient diversity of the shape cues. In this paper, we propose to augment the training dataset by generating semantically meaningful shapes and samples, via a shape deformation-based online augmentation, namely as SDbOA. The samples generated by our SDbOA have two main merits. First, the augmented samples with more diverse shape variations enable networks to learn the shape cues more elaborately, which encourages the network to be shape-biased. Second, semantic-meaningful shape-augmentation samples could be produced by jointly regularizing the generator with object texture and edge-guidance soft constraint, where the edges are represented more robustly with a self information guided map to better against the noises on them. Extensive experiments under various perturbation noises demonstrate the obvious superiority of our shape-bias-motivated model over the state of the arts in terms of robustness performance. Code is available at https://github.com/C0notSilly/-ICCV-23-Edge-Deformation-based-Online-Augmentation.", + "bbox": [ + 75, + 412, + 472, + 820 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0050a9b4b00509de6a17cdc2e21f40d5e7776e6b9bc49a8c602fc0d4018057ad.jpg", + "image_caption": [ + "Figure 1. Comparison of different models' predictions on texture-shape cue conflict samples [11] between Style Augmentation [21] (CVPRW'19), deformation augmentation in [43] and ours. Specific to the testing image (a), the bottom row shows the original (b) and augmented (c)-(e) images for training." + ], + "image_footnote": [], + "bbox": [ + 506, + 375, + 890, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 609, + 632, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous studies [13, 15, 38] have shown that convolution neural networks (CNNs) are vulnerable to perturbation noises, e.g., adversarial noises. Such perturbation noises imposed on the benign image can easily falsify model's prediction [28]. To improve the network robustness against these noises, algorithms of texture augmentation have been long studied to enable network to learn robust and discriminative object texture features of different categories [11, 24].", + "bbox": [ + 496, + 638, + 893, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, recent studies revealed that the vulnerability of CNNs maybe caused by their 'texture-bias', i.e., the features learned for object recognition bias towards objects' textures instead of their shapes [2, 11]. As illustrated in Fig. 1, when the input image (a) contains a cat that is filled with an elephant's texture, standard CNN networks are more likely to predict it as an elephant despite the obvious shape difference between cats and elephants.", + "bbox": [ + 496, + 762, + 893, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Consequently, some studies devote to reducing such", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 101, + 887, + 230, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1526", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "texture-bias to improve network robustness. Zhang et al. [47] found that CNNs learned with adversarial training strategies are less texture-biased. Alternatively, shape-biased CNNs are revealed to be more robust under diverse adversarial attacks. Sun et al. [37] found that the network adversarial robustness can be improved by directly using edge maps to aid the training process. Carlucci et al. [4] proposed to train CNNs on jigsaw puzzles, which achieved superior performances by enabling the network to pay more attention on the cues contained in the global structure. These works illustrate that the shape and structure features, e.g. edge map, can reduce the texture-bias with a stronger shape-bias, i.e. CNNs make predictions based on more shape cues, rather than texture cues.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, current studies mainly learned the features of the object shape that is fixed, while the advantage of the shape diversity for devising a more shape-biased model is not well studied. Meanwhile, direct deformation augmentation on images may destroy the object semantics, especially on the image boundary. Given the above observations, following question naturally emerged: Can we implement data augmentation to the edge map to enhance the shape cues and reduce texture-bias?", + "bbox": [ + 75, + 304, + 472, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose a thin plate spline (TPS) [10]-based edge map deformation strategy to augment the shape representation of the training data, aiming to enhance the shape-bias of the well-trained network. To make the encoded object shape less sensitive to the noises, we extend the edge map specific to only the object edges to include broader boundary regions via the self-information guided map [34]. Furthermore, to build up the correlation between the deformed shape and the texture category to make the augmented samples more semantic-meaningful, we improve the generator with the additional supervision of denoised texture cues and a shape-preservation loss.", + "bbox": [ + 75, + 441, + 472, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fig. 1 sheds light on the above motivation of our shape augmentation for enhancing networks' shape-bias with a toy experiment, where the probabilities to the shape label of an example sample [11] are predicted by the standard ResNet-50, the Style Augmentation [21], the image deformation augmentation [43] and our method. Fig. 1 shows that the proposed algorithm achieves significantly larger probability predicted to the shape label, i.e. tabby cat over the baseline and the related approaches, which reveals the obvious tendency of the proposed SDbOA in applying shape cues for making predictions. For this observation, we argue that the proposed edge-deformation augmentation enables networks to learn the object shape cues more elaborately, while reducing the contribution of the texture cues for prediction.", + "bbox": [ + 75, + 626, + 472, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To the best of our knowledge, this work is one of the pioneer attempts to enhance shape-bias of networks based on the augmentation of object's edges. The main contributions", + "bbox": [ + 76, + 854, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of this paper are summarized as follows:", + "bbox": [ + 498, + 90, + 771, + 107 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a TPS-based edge deformation to augment the object shape, and apply it to the online data augmentation, which could help learning more shape-biased networks.", + "- We propose a semantic-meaningful generation paradigm to produce shape-augmentation samples by jointly regularizing the generator with object texture and edge-guidance soft constraint, where a self information guided map is introduced to represent these edge cues more accurately and robustly.", + "- Extensive results show the superiority of the proposed algorithm over state of the arts (SOTAs) in terms of universal robustness under adversarial, backdoor and common corruption attacks and the training overhead." + ], + "bbox": [ + 517, + 117, + 893, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 364, + 638, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Texture or shape-biased models associated with recognition network robustness", + "text_level": 1, + "bbox": [ + 500, + 388, + 890, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial attacks have exposed the vulnerability of networks to perturbation noises [3, 28, 29, 38], leading to a growing concern for network robustness. Recent studies have investigated the association between network robustness and object texture or shape. Geirhos et al. [11] discovered that CNNs trained on ImageNet rely heavily on texture cues, leading to a texture bias in predictions. To address this issue, Hermann et al. [18] proposed additive augmentation techniques such as color distortion or blurring. However, the contribution of shape cues to recognition robustness has not received sufficient attention in their works.", + "bbox": [ + 496, + 428, + 893, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, Shi et al. [34] proposed an informative dropout approach to help models learn robust representation from a shape-bias perspective. Sun et al. [37] proposed to explicitly use object edges as semantically robust features to against adversarial attacks. Since the object shapes used for enhancing shape-bias are fixed, both approaches are unable to learn a real shape-biased network.", + "bbox": [ + 496, + 595, + 893, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we thus resort to the edge deformation and augmentation to improve the diversity of object shapes. Based on this, we attempt to enhance shape-bias, thus enabling the network to shift its attention from texture to more on the shapes, and better discriminating the samples that have different textures but the similar shape.", + "bbox": [ + 496, + 700, + 893, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Data augmentation for recognition robustness improvement", + "text_level": 1, + "bbox": [ + 500, + 801, + 890, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data augmentation is a popular paradigm for improving the robustness of recognition network against various perturbation noises [16, 17, 21, 24, 27, 32, 41]. This strategy can be applied to generic computer vision tasks, which differs", + "bbox": [ + 496, + 840, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1527", + "bbox": [ + 483, + 944, + 514, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/11a718bfe8ea60788098a431444de5e31a3b9896b2ee1208da66f63e98398238.jpg", + "image_caption": [ + "Figure 2. Pipeline of the proposed SDbOA. For the training, all the three modules, i.e. Edge map-based shape encoding (EMSE), TPS-based shape deformation (TSD) and Texture and shape-based generation (TSG) are used. For the testing, the TSD module is not needed. $\\oslash$ denotes the 'OR' operation." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 872, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "from some task-specific robustness improvement strategies such as adversarial defense algorithms [1, 28, 31, 33, 36, 39] that are designed for a specific attack scenario.", + "bbox": [ + 75, + 337, + 468, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lopes et al. [26] augmented the samples to trade off robustness against recognition accuracy via adding noises on the selected patches of a benign image. Hendrycks et al. [16] proposed the AugMix augmentation, i.e. mixing randomly generated augmentations to improve model robustness under common corruptions. Jackson et al. [21] proposed to augment the data with the style transfer to mitigate domain bias and reduce overfitting. Lee et al. [24] proposed to largely perturb the object texture to help model concentrate on the invariant shape structure.", + "bbox": [ + 75, + 383, + 468, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most augmentation methods [16, 21, 24, 26, 32] either focus on augmenting texture cues or fail to modify object shape regarding edges or structure, such as geometric transformations [32, 35]. However, compared to image deformation-based augmentation methods [10, 43, 48], which strictly adhere to shape constraints, this hard guidance may disregard the semantic coherence of the deformed sample, as shown in Fig. 1.", + "bbox": [ + 75, + 537, + 468, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we propose an object shape augmentation approach based on the TPS-based [10] edge deformation, with the aid of an extended edge encoding and a generator with a soft constraint loss for the shape representation, to dynamically enrich the object shapes.", + "bbox": [ + 75, + 660, + 468, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 752, + 210, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The algorithm architecture, shown in Fig. 2, comprises three main modules: (i) the edge map-based shape encoding (EMSE) module used for robustly representing object shape; (ii) the TPS-based shape deformation (TSD) module used for providing flexible shape augmentation; and (iii) the texture and shape-based generation (TSG) module used for augmenting the samples dynamically, guided by the deformed shape.", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Edge map-based shape encoding (EMSE)", + "text_level": 1, + "bbox": [ + 498, + 335, + 852, + 353 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To encode the object shape, the EMSE consists of two blocks. Firstly, a Robust Canny [37] block that generates an edge map for each image, describing the raw shape of the object in the image. While the Robust Canny may disregard some important shape cues of the object, we then additionally encode a self-information guided map [34] to provide complementary shape structure cues in addition to the Canny operator. Specifically, an edge map $\\hat{E}$ that encodes the pixel-wise self information is first computed as:", + "bbox": [ + 496, + 359, + 892, + 496 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {E} (p) = - \\log \\left(\\sum_ {p ^ {\\prime} \\in \\mathcal {N} _ {p}} e ^ {- \\left(p - p ^ {\\prime}\\right) ^ {2}}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 506, + 890, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{E}(p)$ denotes the estimated self-information at the pixel $p = (x, y)$ of the original image $I$ ; and $\\mathcal{N}_p$ denotes the neighbouring pixels of $p$ .", + "bbox": [ + 496, + 554, + 890, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, we obtain the average pixel-wise self-information $mi$ and perform a thresholding operation to obtain the self-information guided map $E_{info}$ as follows", + "bbox": [ + 496, + 599, + 890, + 647 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nE _ {i n f o} (p) = \\left\\{ \\begin{array}{l l} 1 & \\hat {E} (p) \\geq m i \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 657, + 890, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Eq. (2) indicates that the greater the difference between a pixel and its neighbors, the more shape cues this pixel may contain. That is, for pixels close to the object boundary, their self-information values tend to be relatively large. Consequently, we use this self-information map to extend the edge map extracted by Robust Canny, to leverage the edge-surrounding cues to against the possible noises imposed on the edges, as follows", + "bbox": [ + 496, + 708, + 890, + 830 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\text {e x t e n d}} = E _ {R C} + E _ {\\text {i n f o}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 842, + 890, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $E_{RC}$ denotes the edge maps extracted by the robust Canny.", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1528", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. TPS-based shape deformation (TSD)", + "text_level": 1, + "bbox": [ + 76, + 90, + 395, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To augment the shape representation, we propose TPS-based deformation on the meshing of the self-information edge map in Eq. (3).", + "bbox": [ + 76, + 113, + 468, + 159 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the shape deformation, Dabouei et al. [7] suggested to use the facial landmarks as control points, while this fashion brings about an overhead of the additional detection module and may damage semantic reasonability. By contrast, we propose to uniformly mesh the edge map, and use the endpoints of the grids as the control points to guide the edge map deformation for shape augmentation.", + "bbox": [ + 75, + 159, + 468, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the edge map $E_{\\mathrm{extend}}$ is equally divided into $n = 16$ grids, and the endpoints are denoted as $\\{P_i, i = 1, \\dots, n\\}$ , which serve as the original control points and are randomly perturbed to generate target control points $\\{Q_i, 1 \\leq i \\leq n\\}$ as follows", + "bbox": [ + 76, + 265, + 468, + 340 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {i} = P _ {i} + \\lambda \\mathcal {N} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 349, + 468, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where random noise $\\mathcal{N}$ obeys $N(0,1)$ , i.e. the normal distribution with mean 0 and variance 1, and the $\\lambda$ controls the intensity of image deformation, which is set as 0.1. The edge map meshing is visualized in Fig. 2.", + "bbox": [ + 75, + 377, + 468, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the original and target control points, the edge map deformation is then performed based on TPS. In particular, a transformation $\\Phi$ is found to maximize the smoothness of the edge map deformation while minimizing the gap between two sets of control points. In this work, we use two transformations $\\Phi^x, \\Phi^y$ [10], specific to the translation of the $x, y$ coordinates of the control points, to obtain the deformed edge map $E_{\\text{deform}}$ as:", + "bbox": [ + 75, + 438, + 468, + 559 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{c} E _ {\\text {d e f o r m}} (p) = \\left(p ^ {x} + \\Phi^ {x} (p), p ^ {y} + \\Phi^ {y} (p)\\right) \\\\ \\Phi^ {x} (p) = M ^ {x} \\cdot p + m _ {0} ^ {x} + \\sum_ {i} w _ {i} ^ {x} U (\\| p - P _ {i} \\| _ {2}) \\\\ M ^ {x}, m _ {0} ^ {x}, \\left\\{w _ {i} ^ {x} \\right\\} = \\\\ \\operatorname {a r g m i n} \\sum_ {i} \\left(\\Phi^ {x} \\left(P _ {i}\\right) - \\left(Q _ {i} ^ {x} - P _ {i} ^ {x}\\right)\\right) ^ {2} + \\mathcal {E} _ {\\Phi^ {x}} \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 569, + 468, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $U(\\cdot)$ denotes the function of radial basis kernel, $M^x\\in \\mathbb{R}^2$ $m_0^x,w_i^x\\in \\mathbb{R}$ are the parameters specific to the $x$ coordinate. $\\mathcal{E}_{\\Phi^x}$ is the second-order smoothness term specific to the transformation $\\Phi^x$ . For the $y$ coordinate, the transformation $\\Phi^y$ is similarly derived.", + "bbox": [ + 76, + 641, + 468, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Texture and shape-based generation (TSG)", + "text_level": 1, + "bbox": [ + 76, + 726, + 442, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the edge map deformation may cause the texture distorted on the boundary regions, it may deteriorate the fidelity of the deformed shape and the shape-based recognition performance. Thus, we propose a new generative model based on a two-stage training to inpaint the texture of the deformed edge map, i.e. reconstructing this map into the original image domain for the data augmentation.", + "bbox": [ + 75, + 750, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the Pix2Pix network [20] can achieve the inpainting from an edge map to the original image representation, it can not well build up the correlation between the edge", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "map and the texture category, i.e. an image with correct shape yet mismatched texture category may be generated. Therefore, we propose to use both shape and texture cues in the generator to build up this correlation.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, to provide texture cues for the generator, the original image is smoothed with an easy-to-implement denoising step:", + "bbox": [ + 496, + 152, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {t x t} = \\operatorname {I n t e r p} (D S (I)) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 200, + 890, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where Interp and $DS$ denote interpolation and down-sampling methods, both of which are based on bilinear interpolation. $I_{txt}$ denotes the resulted texture map. This denoised texture map aims to provide rough guidance cues for the synthesis of fine texture. To further provide the shape cues in the edge encoder of the generator, we use $E_{extend}$ in Eq. (3) in the 1st-stage training, and $E_{deform}$ in Eq. (5) in the 2nd-stage training.", + "bbox": [ + 496, + 227, + 890, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To train the generator, a two-stage strategy is used. While a rough image is generated based on $E_{\\text{extend}}$ and the denoised texture $I_{\\text{txt}}$ in the first stage, which is fine-tuned with the input of the deformed edge map $E_{\\text{deform}}$ , and the additional supervision of an object classifier and a newly introduced shape-preservation loss.", + "bbox": [ + 496, + 349, + 890, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the generator is trained in the 1st stage with the loss $\\mathcal{L}_{1st}$ , i.e. the sum of the adversarial loss $\\mathcal{L}_{gan}$ [12], a feature matching loss [22] and a auxiliary classification loss referred from ACGAN [30], which is used to generate rough image without shape deformation. During the second stage, we jointly fine-tune the generator and the classifier with the supervision of the cross-entropy classification loss $\\mathcal{L}_{cls}$ [37] in addition to $\\mathcal{L}_{gan}$ , to better represent the correlation between edge maps and textures.", + "bbox": [ + 496, + 441, + 890, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition, to decrease the sensitivity of the generator against the deformed edge map and stabilize the training, we introduce a $l_{1}$ -norm shape-preservation loss between the deformed edge map and that extracted from the generated image as:", + "bbox": [ + 496, + 578, + 890, + 654 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e d g e}} = \\left\\| \\mathrm {e C N N} \\left(I _ {\\text {s y n}}\\right), E _ {\\text {d e f o r m}} \\right\\| _ {1} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 667, + 890, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I_{syn}$ and eCNN denotes the synthesized image and convolution network to extract edges in [37]. It's worth noting that the deformed shape $E_{deform}$ serves as a soft guidance for generating the shape of the augmented sample, it differs from the hard constraint in the direct image deformation, where the object shape is changed strictly according to the target control points. Consequently, the training specific to the loss $\\mathcal{L}_{2nd}$ in the fine tuning stage is formulated as:", + "bbox": [ + 496, + 698, + 890, + 819 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {G, \\theta} \\max _ {D} \\mathcal {L} _ {2 n d} = \\mathcal {L} _ {g a n} + \\mathcal {L} _ {c l s} + \\mathcal {L} _ {e d g e} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 834, + 890, + 856 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta$ denotes the parameters of the classifier. $G$ and $D$ denote the generator and the discriminator.", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1529", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Overall training", + "text_level": 1, + "bbox": [ + 76, + 90, + 238, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For clarity, our sample synthesis with edge deformation is also presented in Alg. 1, which is further used for online shape augmentation in each epoch of the network training.", + "bbox": [ + 76, + 114, + 468, + 159 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 Edge deformation-based sample generation." + ], + "code_body": "Input: input an image $I$ and initialize the network Output: the shape-deformed image", + "guess_lang": "txt", + "bbox": [ + 76, + 191, + 413, + 220 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Obtain the extended edge map with Eq. (3);", + "2: Mesh the edge map and perturb the control points with Eq. (4);", + "3: Perform edge map deformation with Eq. (5);", + "4: Obtain the denoised texture map with Eq. (6);", + "5: Perform the two-stage training for the generator based on $\\mathcal{L}_{1st}$ and $\\mathcal{L}_{2nd}$ in Eq. (8);" + ], + "bbox": [ + 86, + 222, + 467, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Regarding the training complexity, our algorithm requires only the similar runtime cost as the vanilla training, which is significantly lower than that of offline augmentation. This statement, together with the adversarial robustness of the online and offline augmentation for our method are studied with a toy experiment in the supplementary material.", + "bbox": [ + 75, + 342, + 468, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 460, + 209, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset and experimental settings", + "text_level": 1, + "bbox": [ + 76, + 484, + 372, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets We evaluate the robustness performance of the proposed algorithm under five attacks, i.e. [3, 5, 28, 29, 38] on four databases, i.e. Fashion MNIST (FM) [42], CelebA (CA) [25], CIFAR-10 (C10) [23] and ImageNet (IN) [8].", + "bbox": [ + 75, + 507, + 468, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental setting ResNet50 [14] after the pretraining or ResNet18 is used as the backbone for ImageNet or the other three datasets, unless otherwise specified.", + "bbox": [ + 75, + 568, + 468, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We follow the protocol in EdgeNetRob [37] to evaluate the robustness performance under different adversarial attacks. For adversarial robustness evaluation, we use the $l_{\\infty}$ -norm perturbation constrained in the range [0, 1], and the perturbation budgets, i.e. 25/255 for Fashion MNIST, and 8/255 for CIFAR-10, CelebA and ImageNet. CelebA is used for the task of gender recognition.", + "bbox": [ + 75, + 613, + 468, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics The performance is evaluated with the shape-bias metrics [11, 19], the clean accuracy and the accuracies under diverse attack scenarios. We follow EdgeNetRob [37] to evaluate the clean accuracy, i.e. the accuracy on the testing set generated by EMSE and TSG modules for SDbOA, and the performance on the original testing set for non-generation algorithms.", + "bbox": [ + 75, + 719, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the capacity of our method in terms of clean accuracy preservation compared with state-of-the-art (SOTA) defense strategies and related data augmentation approaches, we use the performance of the baseline as the benchmark.", + "bbox": [ + 75, + 824, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/21a4d66a410e73915318eeebb660238d2713b9eeda866200db9e51806036b7a5.jpg", + "image_caption": [ + "Figure 3. Shape and texture sensitivity of the valilla net and ours against adversarial noises. (a) and (d) show a clean sample and an attacked sample generated by PGD-40 [15], with a texture patch and an edge map extracted by Robust Canny, (b) and (e) show the feature maps (FMs) with the highest $L_{2}$ -norm activation representing these samples based on ResNet-50, and (c) and (d) show the FMs by ours. 'Dist' denotes the $L_{2}$ -norm distance." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 883, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Function Analysis of the proposed modules", + "text_level": 1, + "bbox": [ + 498, + 391, + 864, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1 Function of shape bias for robustness", + "text_level": 1, + "bbox": [ + 500, + 416, + 812, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since adversarial noises are generated within a limited adversarial budget and sparsely distributed in perturbed regions, Fig. 3 shows that the object's shape of an image is less perturbed compared with its texture. Specifically, the edge maps reflecting the shape remain similar for the benign and attacked samples, while the feature maps of these samples with the vanilla net are largely different, i.e. their distance is large, reflecting that the yielded feature representation of object texture has already largely perturbed by the attack. By contrast, our algorithm yields significantly smaller feature map bias, suggesting a more shape-biased model can make feature representation more robust against adversarial perturbations.", + "bbox": [ + 496, + 441, + 890, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2 Function of the proposed modules", + "text_level": 1, + "bbox": [ + 498, + 662, + 790, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The performance of TSD Fig. 4 illustrates the fundamental contribution of the TSD module for shifting from texture-bias to shape-bias. For a moderate magnitude of $\\lambda$ in Eq. (4), e.g. $\\lambda = 0.1$ , augmented samples with diverse shapes and the similar texture are generated. To be specific, the constraint loss in Eq. (7) encourages the diversity of the generated shapes, and our GAN generator enhances the semantic rationality of the generated samples and the similarity of their textures. Consequently, these shape-augmented and texture semantic-preserved data could induce the networks to use more neurons to remember the diverse variations of an object shape, and encourage learned networks to make predictions mainly based on the shape cues, i.e. causing the learned networks to be more shape-biased.", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1530", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1e046d9c632e2bc8be9f9862270136a9cfb6b9d3f6ec9a46901b2d6090610582.jpg", + "image_caption": [ + "Figure 4. The t-SNE visualizations of features with varying $\\lambda$ ." + ], + "image_footnote": [], + "bbox": [ + 81, + 93, + 467, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The performance of EMSE and TSG To shed light on how the introduced self-information edge map and TSG work, we visualize the produced edge maps with the Robust Canny [37] and the proposed self-information in Fig. 5, together with the generated samples with Pix2Pix [20] and our generator.", + "bbox": [ + 75, + 291, + 467, + 381 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The first row of Fig. 5 shows that the proposed self-information guided map can encode richer shape cues and represent shapes more accurately and robustly than those with mere Robust Canny. The second row of Fig. 5 shows that the proposed TSG generates images with well-preserved sharpness of object shape and realistic texture, i.e. better building up the connection between shape and the texture label, compared with Pix2Pix [20].", + "bbox": [ + 75, + 382, + 467, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.Quantitative evaluation", + "text_level": 1, + "bbox": [ + 76, + 512, + 294, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1 Results of shape-bias metrics", + "text_level": 1, + "bbox": [ + 76, + 535, + 333, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Geirhos et al. [11] proposed the first shape-bias evaluation metric based on the prediction accuracy for a texture-shape cue conflict dataset [11] as:", + "bbox": [ + 75, + 559, + 468, + 604 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns b _ {G E} = \\frac {\\# \\text {C o r r e c t e d P r e d i c t i o n t o S h a p e L a b e l}}{\\# \\text {A l l t h e S a m p l e s}} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 611, + 468, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\#$ denotes counting the number. This dataset [11] consists of images with shape and texture from different", + "bbox": [ + 75, + 652, + 468, + 683 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/63ae311d3f9ab30a3a4efe53a84e807e97cc1bdfb3b825229d6c11b1c0413d19.jpg", + "image_caption": [ + "Figure 5. Visualizations of samples from EMSE and TSG. (a) represents a vanilla sample. (b) and (c) represent edge maps extracted by Robust Canny [37] and EMSE. (d) and (e) represent images generated by Pix2Pix [20] and our TSG." + ], + "image_footnote": [], + "bbox": [ + 142, + 704, + 405, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "classes to misguide the CNN. Specifically, the object textures of the images [11] are largely interfered with category-irrelevant styles, which places big challenge for current texture-biased networks.", + "bbox": [ + 496, + 90, + 890, + 150 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Islam et al. [19] proposed another shape-bias metric as:", + "bbox": [ + 517, + 152, + 883, + 167 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns b _ {I S} = \\frac {\\sum_ {i = 1} ^ {n f} \\rho_ {i}}{n f}, \\text {w h e r e} \\rho_ {i} = \\frac {\\operatorname {C o v} \\left(z _ {i} ^ {a} , z _ {i} ^ {b}\\right)}{\\sqrt {\\operatorname {V a r} \\left(z _ {i} ^ {a}\\right) \\operatorname {V a r} \\left(z _ {i} ^ {b}\\right)}} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 178, + 890, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $nf$ is the number of feature maps. $z_{i}^{a}$ and $z_{i}^{b}$ represents the $i$ -th flattened feature maps corresponding to the image pair $(a, b)$ from the texture-shape cue conflict dataset [11], while the two images contain the similar shape but have different textures. The normalized correlation coefficient $\\rho_{i}$ measures the proportion of feature maps to represent the shape cues. The quantitative results of the standard network, StyleAug [21], ShapeAug [24] and our SDbOA in terms of the two quantitative metrics are shown in Tab. 1.", + "bbox": [ + 496, + 229, + 890, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 shows that our SDbOA consistently achieves the larger shape-bias values compared with the standard network and the related augmentation algorithms [21, 24] in term of both quantitative metrics for the training databases of ImageNet (IN), Stylized IN [11] (SIN) and SIN+IN.", + "bbox": [ + 496, + 366, + 890, + 441 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5d336831a5a5c64fb7d01bea4958181dcb5db401bb2d6e0efa17db046eacab32.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDatasetsbGE[11]sbIS[19]
StandardIN21.3917.0
StyleAug [21]IN67.3128.5
ShapeAug [24]IN38.4621.3
SDbOAIN71.2831.2
StandardSIN81.3726.2
StyleAug [21]SIN73.7232.7
ShapeAug [24]SIN67.7529.8
SDbOASIN82.4039.9
StandardSIN+IN34.6518.4
StyleAug [21]SIN+IN71.0430.4
ShapeAug [24]SIN+IN63.6723.1
SDbOASIN+IN78.7937.5
", + "bbox": [ + 501, + 454, + 890, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative results in terms of two shape-bias metrics $(\\%)$ with two shape-bias enhancement methods, i.e. StyleAug [21] (CVPRW'19) and ShapeAug [24] (CVPRW'22).", + "bbox": [ + 498, + 659, + 890, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.2 Robustness against various perturbations", + "text_level": 1, + "bbox": [ + 500, + 709, + 841, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Robustness against adversarial attacks: To study the adversarial robustness of our SDbOA compared with the SO-TAs, we present the performances of seven algorithms in terms of clean accuracy and adversarial robustness against FGSM [13], PGD [28], C&W [3] and DeepFool [29] in Tab. 2.", + "bbox": [ + 496, + 734, + 890, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 2, the models trained with SDbOA maintain relatively stronger robustness against various adversarial attacks, than those with the SOTAs. Specifically, SDbOA outperforms NuAT [36] by the margins of $14.10\\%$ and $27.72\\%$ in terms of the robustness under C&W and", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1531", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DeepFool on CIFAR-10, and outperforms EdgeNetRob [37] by a margin of $6.98\\%$ under C&W on ImageNet.", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While Tab. 2 includes generic adversarial training (AT)-based SOTAs for the comparison, we further compare our algorithm with additional two AT-based algorithms [31,33] that are also based on generative model as similar as ours in Tab. 3, where the same protocol (evaluation with AutoAttack [5]) and network architectures as [31,33] are employed.", + "bbox": [ + 75, + 122, + 467, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since the works [31,33] require adversarial training, tens of millions of adversarial samples in addition to the vanilla images are generated for the training, resulting in a significantly larger runtime complexity. Despite this, Tab. 3 shows that our online augmentation model can still outperform the listed adversarial-training-based methods in terms of the clean accuracy and adversarial robustness with reasonable training overhead.", + "bbox": [ + 75, + 214, + 467, + 335 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8d69625fcb2d540d8dcf11a55b67b482a9cba4d79bb9f5a4872ce40b664f3551.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Data.MethodClean Acc.FGSM PGD CWDeep Fool
FMVanilla Net92.4227.420.5810.0039.5
PGD [28]86.5878.8481.2617.540.56
AT+CAS [1]87.4179.2378.4232.9531.46
NuAT [36]88.0281.4382.3767.0742.19
AT+CR [39]87.4680.8281.9534.5737.04
EdgeNetRob [37]85.0178.3075.7582.4771.43
SDbOA88.5280.3283.2687.2476.12
CAVanilla Net99.1848.860.0023.5720.80
PGD [28]92.8585.4083.3542.7819.84
AT+CAS [1]93.0689.8887.6865.3637.05
NuAT [36]94.2693.8089.5267.2529.42
AT+CR [39]95.3294.5890.7464.3839.06
EdgeNetRob [37]95.5292.6455.0875.9283.72
SDbOA98.7495.3488.2879.8491.50
C10Vanilla Net90.707.840.0016.7814.30
PGD [28]75.8254.5644.6618.020.38
AT+CAS [1]76.5162.3247.8151.4337.42
NuAT [36]79.5162.1246.4541.9634.66
AT+CR [39]82.8064.8153.3950.6643.19
EdgeNetRob [37]79.2162.4839.6643.4654.45
SDbOA83.2767.8250.9156.0662.38
INVanilla Net75.825.400.00-5.44
PGD [28]55.3440.4630.8316.0418.36
AT+CAS [1]57.0347.5836.7223.6727.92
NuAT [36]58.3848.1240.3021.7125.40
AT+CR [39]52.0341.7227.9321.4021.07
EdgeNetRob [37]65.7345.8733.6220.8523.40
SDbOA68.5649.1035.7627.8328.89
", + "bbox": [ + 86, + 349, + 455, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness of augmentation approaches against", + "bbox": [ + 96, + 885, + 467, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9df12a63ac885ce226588b554651e69c67e420fb64c8cf1d18aeeef20c8b2692.jpg", + "table_caption": [ + "Table 2. Clean accuracy and adversarial robustness (%) with vanilla net, PGD Training [28], AT+CAS [1] (ICLR'21), NuAT [36] (NeurIPS'21), AT+CR [39] (AAAI'22), EdgeNetRob [37] (ICCV'21) and ours. The best and 2nd best performances are labeled with bold and underline." + ], + "table_footnote": [], + "table_body": "
MethodCleanRobustRuntimeAT
FDA28 [31]85.9760.73Days✓
SDbOA2885.2768.153-5 Hours✗
PORT34 [33]87.0060.60Days✓
SDbOA3487.1273.363-5 Hours✗
", + "bbox": [ + 501, + 88, + 898, + 175 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "common corruptions and adversarial attacks: To evaluate the performances of our data augmentation and related augmentation techniques under common corruptions [15] and adversarial attacks, we present the results of nine SO-TAs in term of mean corruption error (mCE) and adversarial robustness in Tab. 4, where mCE is evaluated under all 15 corruptions and each corruption has 5 severities [15].", + "bbox": [ + 498, + 277, + 890, + 383 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2979cb79af8cb3198f51c1c1d8907f5fe2a9b35a95de4efbf5ff5c4e06ecdd26.jpg", + "table_caption": [ + "Table 3. Adversarial robustness (\\%) of the generative-model-based adversarial training approaches, i.e. FDA [31] (NeurIPS'21) and PORT [33] (ICLR'22) on CIFAR-10. We follow the robustness metric in [31] measured by AutoAttack [5]. AT represents 'adversarial training'. '28' or '34' stands for using WideResNet-28-10 or WideResNet-34-10 [45] as the classifier network, following [31, 33]." + ], + "table_footnote": [], + "table_body": "
MethodsCorruptions↓Adversaries↓
Baseline26.491.3
Cutout [9]arXiv25.996.0
Mixup [46]ICLR'1821.093.3
CutMix [44]ICCV'1926.592.1
AutoAugment [6]CVPR'1922.295.1
AugMix [16]ICLR'1912.486.8
AugMax [41]NeurIPS'2119.939.1
PixMix [17]CVPR'229.582.1
TPS-Deform [43]MM'2222.444.0
SDbOA18.827.7
", + "bbox": [ + 501, + 397, + 890, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Robustness performances (\\%) of nine data augmentation techniques on CIFAR-10-C, i.e. the corruption set of CIFAR-10, based on the baseline of WideResNet-40-4 [45] under PGD-20 attack. The adversaries metric is the misclassification rate under PGD-20 with the $l_{\\infty}$ budget of 2/255.", + "bbox": [ + 498, + 571, + 890, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 4 shows that large majority of SOTA data augmentation techniques improve robustness against corruptions without paying enough attention on the adversarial robustness. It's worth noting that our algorithm achieves much better adversarial robustness, i.e. $27.7\\%$ with large margins over those, i.e. $82.1\\%$ and $86.8\\%$ achieved by PixMix [17] and AugMax [41] that perform better than ours in terms of mCE. This maybe because that these two augmentations are mainly focused on robustness against a specified type of samples, while ours is developed for universal robustness.", + "bbox": [ + 496, + 657, + 890, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness against backdoor attacks: To study the robustness of the proposed algorithm under backdoor attacks [40] compared with the related data augmentation methods, we follow the protocol in [37], and present the clean accuracy on the standard testing data and the attack success rate on the poisoned data (Pois. ASR) of seven algorithms, un", + "bbox": [ + 496, + 810, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1532", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "der backdoor attacks with two patterns in Tab. 5. Poisoning ratio is set as $20\\%$ for Fashion MNIST and $5\\%$ for CelebA.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5d2c9a74134b833378c49642186cfdb3c3fb322016b858f413b744f5ec929f35.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5. Performances (\\%) of vanilla net, Spectral Signature (SpecSign) [40] (NeurIPS'18), EdgeNetRob [37] (ICCV'21), AugMax [41] (NeurIPS'21), PixMix [17] (CVPR'22), TPS-Deform [43] (MM'22) and SDbOA under backdoor attacks." + ], + "table_body": "
Data.MethodBackdoor Pattern
PixelPattern
Clean Acc.↑Pois. ASR↓Clean Acc.↑Pois. ASR↓
FMVanilla Net87.4394.3087.1295.22
SpecSign [40]86.2345.6285.9352.31
EdgeNetRob [37]83.480.1282.212.74
AugMax [41]89.8354.6089.2773.14
PixMix [17]89.2592.0488.0194.78
TPS-Deform [43]86.3547.4386.2762.21
SDbOA87.250.0487.070.98
CAVanilla Net98.3097.2098.0097.40
SpecSign [40]98.4564.7897.9854.23
EdgeNetRob [37]92.8010.9093.1012.50
AugMax [41]98.8445.9798.2374.10
PixMix [17]98.1878.3297.9889.28
TPS-Deform [43]97.4371.3197.4774.52
SDbOA95.203.4994.874.21
", + "bbox": [ + 81, + 132, + 465, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 5 shows that SDbOA achieves the lowest ASR values against backdoor attacks on the two datasets, e.g. it significantly outperforms AugMax [41] that performs best in terms of clean accuracy, by the margins of $54.56\\%$ and $72.16\\%$ specific to the pixel and pattern attacks on Fashion MNIST. Meanwhile, the clean accuracy is reasonably preserved compared with the most related algorithm, i.e. EdgeNetRob [37].", + "bbox": [ + 75, + 424, + 468, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.3 Hyperparameter analysis and ablation study", + "text_level": 1, + "bbox": [ + 76, + 565, + 441, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of hyperparameter $\\lambda$ : To analyze the sensitivity of the learned model's shape-bias against the deformation intensity, i.e. $\\lambda$ in Eq. (4), we conduct a hyperparameter analysis and present the results in terms of the two shape-bias metrics [11, 19] in Fig. 6.", + "bbox": [ + 75, + 589, + 468, + 665 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eee9cd0ebc73bc818425b08d38d5c43bfeaa832cc5aff415f36de7e5ccbd66b6.jpg", + "image_caption": [ + "Figure 6. Sensitivity of shape-bias metric values [11, 19] against different deformation intensities $(\\lambda)$ ." + ], + "image_footnote": [], + "bbox": [ + 109, + 681, + 434, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 6 shows that a greater deformation degree does not always result in a more shape-biased model, i.e. the shape-bias of the model learned by SDbOA gradually increases", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/11db19e4450bd264704362e0f8c4df45c148b44a6bcc34bf4f9eaedd00b3ba6c.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6. Ablation study in terms of the clean accuracy and adversarial robustness (%) on CIFAR 10." + ], + "table_body": "
EMSETSDTSGCleanFGSMPGDC&WDeepFool
74.3260.0646.7845.4855.82
✓76.5062.0847.0847.0554.25
✓✓78.2465.2649.8249.6161.35
✓✓87.6715.830.0021.8419.02
✓✓81.0866.8448.4052.3159.46
✓✓✓83.2767.8250.9156.0662.38
", + "bbox": [ + 501, + 75, + 879, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "when $\\lambda$ is smaller than the threshold of 0.1, while it is decreased when $\\lambda$ exceeds 0.1, i.e. the learned model gradually shifts from shape-bias to be more texture-biased.", + "bbox": [ + 500, + 215, + 890, + 260 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To shed light on this observation, we argue that a moderate deformation degree, e.g. $\\lambda = 0.1$ in Fig. 4, is beneficial to reasonably improve the diversity of shape representations without largely impairing shape semantic, and a model learned from these valid shape variations can be made better against the perturbation noises. However, as shown in Fig. 4, excessive deformation may be caused when a large $\\lambda$ , e.g. $\\lambda = 0.2$ is used, which may damage the global shape structure and original shape semantic, causing deformed samples to even cross classification boundaries. When it happens, these augmented samples with damaged shapes may instead encourage the learned network to give prediction based on the relatively stable texture, i.e. making the model to be more texture-biased.", + "bbox": [ + 496, + 262, + 890, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study: To study the performance of each module in our SDbOA, we perform the ablation study in Tab. 6, where Robust Canny [37] and Pix2Pix [20] are used as the baseline shape encoding and generator, respectively.", + "bbox": [ + 496, + 474, + 890, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By comparing the 5th and 7th rows of Tab. 6, it shows that the EMSE module is beneficial to the network robustness, reflecting the superiority of our edge augmentation over direct image deformation. By comparing the 6th and 7th rows, it shows that the proposed TSD enables the model to consistently improve the clean accuracy and the robustness performances under four attacks. By comparing the 4th and 7th rows, it shows that the improved generator with the supervision of texture cues and the shape-preservation loss largely improves the Pix2Pix in terms of the clean accuracy.", + "bbox": [ + 496, + 537, + 890, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 691, + 625, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Considering the network robustness maybe impaired by the excessive texture-bias, we shed light on how to shift from texture-bias to shape-bias for CNN models, i.e. we propose a novel edge deformation-based data augmentation that enables networks to elaborately learn the shape variations, allowing the well-learned networks to make predictions mainly based on the targets' shapes. Our data augmentation mainly differs from the related SOTAs that it is developed based on the deformation of object's edge map rather than the image, and can reasonably preserve the semantic rationality of the reconstructed samples with the joint supervision of object texture and soft constraint of de", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1533", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "formed shape cues. Experimental results reveal that our algorithm can largely enhance model's shape-bias in terms of two quantitative metrics, and rather competitive robustness against various perturbations compared with SOTAs. Considering the recognition system trained with the proposed shape augmentation is mainly applicable for pixelwise perturbation, yet may not work for patch-wise corruption where the shape structure is severely damaged, our future work will be devoted to solving it.", + "bbox": [ + 75, + 90, + 472, + 226 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements. The work was supported by Natural Science Foundation of China under grants no. 62276170, 82261138629, the Science and Technology Project of Guangdong Province under grants no. 2023A1515011549, 2023A1515010688, the Science and Technology Innovation Commission of Shenzhen under grants no. JCYJ20190808165203670, JCYJ20220531101412030.", + "bbox": [ + 75, + 227, + 472, + 348 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 362, + 174, + 378 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yang Bai, Yuyuan Zeng, Yong Jiang, Shu-Tao Xia, Xingjun Ma, and Yisen Wang. Improving adversarial robustness via channel-wise activation suppressing. In International Conference on Learning Representations, 2020. 3, 7", + "[2] Nicholas Baker, Hongjing Lu, Gennady Erlikhman, and Philip J Kellman. Deep convolutional networks do not classify based on global object shape. PLoS computational biology, 14(12):e1006613, 2018. 1", + "[3] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE symposium on security and privacy (sp), pages 39-57. IEEE, 2017. 2, 5, 6", + "[4] Fabio M Carlucci, Antonio D'Innocente, Silvia Bucci, Barbara Caputo, and Tatiana Tommasi. Domain generalization by solving jigsaw puzzles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2229-2238, 2019. 2", + "[5] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 5, 7", + "[6] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 113-123, 2019. 7", + "[7] Ali Dabouei, Sobhan Soleymani, Jeremy Dawson, and Nasser Nasrabadi. Fast geometrically-perturbed adversarial faces. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1979-1988. IEEE, 2019. 4", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5", + "[9] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 7" + ], + "bbox": [ + 86, + 388, + 472, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Gianluca Donato and Serge Belongie. Approximate thin plate spline mappings. In European conference on computer vision, pages 21-31. Springer, 2002. 2, 3, 4", + "[11] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations, 2018. 1, 2, 5, 6, 8", + "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 4", + "[13] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 6", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[15] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018. 1, 5, 7", + "[16] Dan Hendrycks, Norman Mu, Ekin D Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781, 2019. 2, 3, 7", + "[17] Dan Hendrycks, Andy Zou, Mantas Mazeika, Leonard Tang, Bo Li, Dawn Song, and Jacob Steinhardt. Pixmix: Dreamlike pictures comprehensively improve safety measures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16783-16792, 2022. 2, 7, 8", + "[18] Katherine Hermann, Ting Chen, and Simon Kornblith. The origins and prevalence of texture bias in convolutional neural networks. Advances in Neural Information Processing Systems, 33:19000-19015, 2020. 2", + "[19] Md Amirul Islam, Matthew Kowal, Patrick Esser, Sen Jia, Bjorn Ommer, Konstantinos G Derpanis, and Neil Bruce. Shape or texture: Understanding discriminative features in cnns. In International Conference on Learning Representations, 2021. 5, 6, 8", + "[20] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 4, 6, 8", + "[21] Philip TG Jackson, Amir Atapour Abarghouei, Stephen Bonner, Toby P Breckon, and Boguslaw Obara. Style augmentation: data augmentation via style randomization. In CVPR workshops, volume 6, pages 10-11, 2019. 1, 2, 3, 6", + "[22] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694-711. Springer, 2016. 4", + "[23] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1534", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Sangjun Lee, Inwoo Hwang, Gi-Cheon Kang, and Byoung-Tak Zhang. Improving robustness to texture bias via shape-focused augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 4323-4331, June 2022. 1, 2, 3, 6", + "[25] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 5", + "[26] Raphael Gontijo Lopes, Dong Yin, Ben Poole, Justin Gilmer, and Ekin D Cubuk. Improving robustness without sacrificing accuracy with patch gaussian augmentation. arXiv preprint arXiv:1906.02611, 2019. 3", + "[27] Cheng Luo, Qinliang Lin, Weicheng Xie, Bizhu Wu, Jinheng Xie, and Linlin Shen. Frequency-driven imperceptible adversarial attack on semantic similarity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15315-15324, 2022. 2", + "[28] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1, 2, 3, 5, 6, 7", + "[29] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2574–2582, 2016. 2, 5, 6", + "[30] Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier gans. In International conference on machine learning, pages 2642-2651. PMLR, 2017. 4", + "[31] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan A Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Fixing data augmentation to improve adversarial robustness. In Advances in Neural Information Processing Systems, 2021. 3, 7", + "[32] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan Andrei Calian, Florian Stimberg, Olivia Wiles, and Timothy A Mann. Data augmentation can improve robustness. Advances in Neural Information Processing Systems, 34:29935-29948, 2021. 2, 3", + "[33] Vikash Sehwag, Saeed Mahloujifar, Tinashe Handina, Sihui Dai, Chong Xiang, Mung Chiang, and Prateek Mittal. Robust learning meets generative models: Can proxy distributions improve adversarial robustness? In International Conference on Learning Representations, 2022. 3, 7", + "[34] Baifeng Shi, Dinghuai Zhang, Qi Dai, Zhanxing Zhu, Yadong Mu, and Jingdong Wang. Informative dropout for robust representation learning: A shape-bias perspective. In International Conference on Machine Learning, pages 8828-8839. PMLR, 2020. 2, 3", + "[35] Connor Shorten and Taghi M Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of big data, 6(1):1-48, 2019. 3", + "[36] Gaurang Sriramanan, Sravanti Addepalli, Arya Baburaj, et al. Towards efficient and effective adversarial train" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing. Advances in Neural Information Processing Systems, 34:11821-11833, 2021. 3, 6, 7", + "[37] Mingjie Sun, Zichao Li, Chaowei Xiao, Haonan Qiu, Bhavya Kailkhura, Mingyan Liu, and Bo Li. Can shape structure features improve model robustness under diverse adversarial settings? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7526-7535, 2021. 2, 3, 4, 5, 6, 7, 8", + "[38] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1, 2, 5", + "[39] Jihoon Tack, Sihyun Yu, Jongheon Jeong, Minseon Kim, Sung Ju Hwang, and Jinwoo Shin. Consistency regularization for adversarial robustness. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 8414-8422, 2022. 3, 7", + "[40] Brandon Tran, Jerry Li, and Aleksander Madry. Spectral signatures in backdoor attacks. Advances in neural information processing systems, 31, 2018. 7, 8", + "[41] Haotao Wang, Chaowei Xiao, Jean Kossaifi, Zhiding Yu, Anima Anandkumar, and Zhangyang Wang. Augmax: Adversarial composition of random augmentations for robust training. Advances in neural information processing systems, 34:237-250, 2021. 2, 7, 8", + "[42] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. 5", + "[43] Jun Yu, Guochen Xie, Zhongpeng Cai, Peng He, Fang Gao, and Qiang Ling. Micro expression generation with thin-plate spline motion model and face parsing. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7210–7214, 2022. 1, 2, 3, 7, 8", + "[44] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 7", + "[45] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In *British Machine Vision Conference* 2016. British Machine Vision Association, 2016. 7", + "[46] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 7", + "[47] Tianyuan Zhang and Zhanxing Zhu. Interpreting adversarially trained convolutional neural networks. In International Conference on Machine Learning, pages 7502-7511. PMLR, 2019. 2", + "[48] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3657–3666, 2022. 3" + ], + "bbox": [ + 501, + 92, + 893, + 854 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1535", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Shift from Texture-bias to Shape-bias_ Edge Deformation-based Augmentation for Robust Object Recognition/e071fbaf-3036-4ac7-9bfd-fb6983ccf03f_model.json b/2023/Shift from Texture-bias to Shape-bias_ Edge Deformation-based Augmentation for Robust Object Recognition/e071fbaf-3036-4ac7-9bfd-fb6983ccf03f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e09729238a89d990fda64c30307e86f782118079 --- /dev/null +++ b/2023/Shift from Texture-bias to Shape-bias_ Edge Deformation-based Augmentation for Robust Object Recognition/e071fbaf-3036-4ac7-9bfd-fb6983ccf03f_model.json @@ -0,0 +1,2420 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.089, + 0.131, + 0.882, + 0.177 + ], + "angle": 0, + "content": "Shift from Texture-bias to Shape-bias: Edge Deformation-based Augmentation for Robust Object Recognition" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.211, + 0.964, + 0.231 + ], + "angle": 0, + "content": "Xilin He\\(^{1,2,3}\\), Qinliang Lin\\(^{1,2,3}\\), Cheng Luo\\(^{1,2,3}\\) Weicheng Xie\\(^{1,2,3}\\), Siyang Song\\(^{4}\\), Feng Liu\\(^{1,2,3}\\), Linlin Shen\\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.231, + 0.935, + 0.249 + ], + "angle": 0, + "content": "1Computer Vision Institute, School of Computer Science & Software Engineering, Shenzhen University" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.249, + 0.79, + 0.267 + ], + "angle": 0, + "content": "2Shenzhen Institute of Artificial Intelligence & Robotics for Society" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.267, + 0.788, + 0.285 + ], + "angle": 0, + "content": "3Guangdong Key Laboratory of Intelligent Information Processing" + }, + { + "type": "text", + "bbox": [ + 0.429, + 0.285, + 0.623, + 0.302 + ], + "angle": 0, + "content": "\\(^{4}\\)University of Leicester" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.304, + 0.79, + 0.319 + ], + "angle": 0, + "content": "{2020152115, linqinliang2021, luocheng2020}@email.szu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.321, + 0.782, + 0.336 + ], + "angle": 0, + "content": "{wcxie,feng.liu,llshen}@szu.edu.cn,ss1535@leicester.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.371, + 0.314, + 0.387 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.473, + 0.821 + ], + "angle": 0, + "content": "Recent studies have shown the vulnerability of CNNs under perturbation noises, which is partially caused by the reason that the well-trained CNNs are too biased toward the object texture, i.e., they make predictions mainly based on texture cues. To reduce this texture-bias, current studies resort to learning augmented samples with heavily perturbed texture to make networks be more biased toward relatively stable shape cues. However, such methods usually fail to achieve real shape-biased networks due to the insufficient diversity of the shape cues. In this paper, we propose to augment the training dataset by generating semantically meaningful shapes and samples, via a shape deformation-based online augmentation, namely as SDbOA. The samples generated by our SDbOA have two main merits. First, the augmented samples with more diverse shape variations enable networks to learn the shape cues more elaborately, which encourages the network to be shape-biased. Second, semantic-meaningful shape-augmentation samples could be produced by jointly regularizing the generator with object texture and edge-guidance soft constraint, where the edges are represented more robustly with a self information guided map to better against the noises on them. Extensive experiments under various perturbation noises demonstrate the obvious superiority of our shape-bias-motivated model over the state of the arts in terms of robustness performance. Code is available at https://github.com/C0notSilly/-ICCV-23-Edge-Deformation-based-Online-Augmentation." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.376, + 0.891, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.525, + 0.895, + 0.596 + ], + "angle": 0, + "content": "Figure 1. Comparison of different models' predictions on texture-shape cue conflict samples [11] between Style Augmentation [21] (CVPRW'19), deformation augmentation in [43] and ours. Specific to the testing image (a), the bottom row shows the original (b) and augmented (c)-(e) images for training." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.611, + 0.633, + 0.627 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.894, + 0.761 + ], + "angle": 0, + "content": "Previous studies [13, 15, 38] have shown that convolution neural networks (CNNs) are vulnerable to perturbation noises, e.g., adversarial noises. Such perturbation noises imposed on the benign image can easily falsify model's prediction [28]. To improve the network robustness against these noises, algorithms of texture augmentation have been long studied to enable network to learn robust and discriminative object texture features of different categories [11, 24]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.894, + 0.884 + ], + "angle": 0, + "content": "However, recent studies revealed that the vulnerability of CNNs maybe caused by their 'texture-bias', i.e., the features learned for object recognition bias towards objects' textures instead of their shapes [2, 11]. As illustrated in Fig. 1, when the input image (a) contains a cat that is filled with an elephant's texture, standard CNN networks are more likely to predict it as an elephant despite the obvious shape difference between cats and elephants." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Consequently, some studies devote to reducing such" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.231, + 0.9 + ], + "angle": 0, + "content": "* Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1526" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "texture-bias to improve network robustness. Zhang et al. [47] found that CNNs learned with adversarial training strategies are less texture-biased. Alternatively, shape-biased CNNs are revealed to be more robust under diverse adversarial attacks. Sun et al. [37] found that the network adversarial robustness can be improved by directly using edge maps to aid the training process. Carlucci et al. [4] proposed to train CNNs on jigsaw puzzles, which achieved superior performances by enabling the network to pay more attention on the cues contained in the global structure. These works illustrate that the shape and structure features, e.g. edge map, can reduce the texture-bias with a stronger shape-bias, i.e. CNNs make predictions based on more shape cues, rather than texture cues." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.473, + 0.442 + ], + "angle": 0, + "content": "However, current studies mainly learned the features of the object shape that is fixed, while the advantage of the shape diversity for devising a more shape-biased model is not well studied. Meanwhile, direct deformation augmentation on images may destroy the object semantics, especially on the image boundary. Given the above observations, following question naturally emerged: Can we implement data augmentation to the edge map to enhance the shape cues and reduce texture-bias?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.443, + 0.473, + 0.626 + ], + "angle": 0, + "content": "To this end, we propose a thin plate spline (TPS) [10]-based edge map deformation strategy to augment the shape representation of the training data, aiming to enhance the shape-bias of the well-trained network. To make the encoded object shape less sensitive to the noises, we extend the edge map specific to only the object edges to include broader boundary regions via the self-information guided map [34]. Furthermore, to build up the correlation between the deformed shape and the texture category to make the augmented samples more semantic-meaningful, we improve the generator with the additional supervision of denoised texture cues and a shape-preservation loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.473, + 0.853 + ], + "angle": 0, + "content": "Fig. 1 sheds light on the above motivation of our shape augmentation for enhancing networks' shape-bias with a toy experiment, where the probabilities to the shape label of an example sample [11] are predicted by the standard ResNet-50, the Style Augmentation [21], the image deformation augmentation [43] and our method. Fig. 1 shows that the proposed algorithm achieves significantly larger probability predicted to the shape label, i.e. tabby cat over the baseline and the related approaches, which reveals the obvious tendency of the proposed SDbOA in applying shape cues for making predictions. For this observation, we argue that the proposed edge-deformation augmentation enables networks to learn the object shape cues more elaborately, while reducing the contribution of the texture cues for prediction." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To the best of our knowledge, this work is one of the pioneer attempts to enhance shape-bias of networks based on the augmentation of object's edges. The main contributions" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.772, + 0.108 + ], + "angle": 0, + "content": "of this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.118, + 0.892, + 0.178 + ], + "angle": 0, + "content": "- We propose a TPS-based edge deformation to augment the object shape, and apply it to the online data augmentation, which could help learning more shape-biased networks." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.189, + 0.895, + 0.281 + ], + "angle": 0, + "content": "- We propose a semantic-meaningful generation paradigm to produce shape-augmentation samples by jointly regularizing the generator with object texture and edge-guidance soft constraint, where a self information guided map is introduced to represent these edge cues more accurately and robustly." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.291, + 0.895, + 0.352 + ], + "angle": 0, + "content": "- Extensive results show the superiority of the proposed algorithm over state of the arts (SOTAs) in terms of universal robustness under adversarial, backdoor and common corruption attacks and the training overhead." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.118, + 0.895, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.365, + 0.64, + 0.381 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.39, + 0.892, + 0.422 + ], + "angle": 0, + "content": "2.1. Texture or shape-biased models associated with recognition network robustness" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.895, + 0.595 + ], + "angle": 0, + "content": "Adversarial attacks have exposed the vulnerability of networks to perturbation noises [3, 28, 29, 38], leading to a growing concern for network robustness. Recent studies have investigated the association between network robustness and object texture or shape. Geirhos et al. [11] discovered that CNNs trained on ImageNet rely heavily on texture cues, leading to a texture bias in predictions. To address this issue, Hermann et al. [18] proposed additive augmentation techniques such as color distortion or blurring. However, the contribution of shape cues to recognition robustness has not received sufficient attention in their works." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.895, + 0.701 + ], + "angle": 0, + "content": "Recently, Shi et al. [34] proposed an informative dropout approach to help models learn robust representation from a shape-bias perspective. Sun et al. [37] proposed to explicitly use object edges as semantically robust features to against adversarial attacks. Since the object shapes used for enhancing shape-bias are fixed, both approaches are unable to learn a real shape-biased network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.895, + 0.793 + ], + "angle": 0, + "content": "In this work, we thus resort to the edge deformation and augmentation to improve the diversity of object shapes. Based on this, we attempt to enhance shape-bias, thus enabling the network to shift its attention from texture to more on the shapes, and better discriminating the samples that have different textures but the similar shape." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.892, + 0.833 + ], + "angle": 0, + "content": "2.2. Data augmentation for recognition robustness improvement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Data augmentation is a popular paradigm for improving the robustness of recognition network against various perturbation noises [16, 17, 21, 24, 27, 32, 41]. This strategy can be applied to generic computer vision tasks, which differs" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "1527" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.089, + 0.873, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.283, + 0.893, + 0.327 + ], + "angle": 0, + "content": "Figure 2. Pipeline of the proposed SDbOA. For the training, all the three modules, i.e. Edge map-based shape encoding (EMSE), TPS-based shape deformation (TSD) and Texture and shape-based generation (TSG) are used. For the testing, the TSD module is not needed. \\(\\oslash\\) denotes the 'OR' operation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.47, + 0.384 + ], + "angle": 0, + "content": "from some task-specific robustness improvement strategies such as adversarial defense algorithms [1, 28, 31, 33, 36, 39] that are designed for a specific attack scenario." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.47, + 0.537 + ], + "angle": 0, + "content": "Lopes et al. [26] augmented the samples to trade off robustness against recognition accuracy via adding noises on the selected patches of a benign image. Hendrycks et al. [16] proposed the AugMix augmentation, i.e. mixing randomly generated augmentations to improve model robustness under common corruptions. Jackson et al. [21] proposed to augment the data with the style transfer to mitigate domain bias and reduce overfitting. Lee et al. [24] proposed to largely perturb the object texture to help model concentrate on the invariant shape structure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.659 + ], + "angle": 0, + "content": "Most augmentation methods [16, 21, 24, 26, 32] either focus on augmenting texture cues or fail to modify object shape regarding edges or structure, such as geometric transformations [32, 35]. However, compared to image deformation-based augmentation methods [10, 43, 48], which strictly adhere to shape constraints, this hard guidance may disregard the semantic coherence of the deformed sample, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.661, + 0.47, + 0.737 + ], + "angle": 0, + "content": "In this work, we propose an object shape augmentation approach based on the TPS-based [10] edge deformation, with the aid of an extended edge encoding and a generator with a soft constraint loss for the shape representation, to dynamically enrich the object shapes." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.753, + 0.212, + 0.77 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The algorithm architecture, shown in Fig. 2, comprises three main modules: (i) the edge map-based shape encoding (EMSE) module used for robustly representing object shape; (ii) the TPS-based shape deformation (TSD) module used for providing flexible shape augmentation; and (iii) the texture and shape-based generation (TSG) module used for augmenting the samples dynamically, guided by the deformed shape." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.337, + 0.854, + 0.354 + ], + "angle": 0, + "content": "3.1. Edge map-based shape encoding (EMSE)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.893, + 0.497 + ], + "angle": 0, + "content": "To encode the object shape, the EMSE consists of two blocks. Firstly, a Robust Canny [37] block that generates an edge map for each image, describing the raw shape of the object in the image. While the Robust Canny may disregard some important shape cues of the object, we then additionally encode a self-information guided map [34] to provide complementary shape structure cues in addition to the Canny operator. Specifically, an edge map \\(\\hat{E}\\) that encodes the pixel-wise self information is first computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.507, + 0.892, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\hat {E} (p) = - \\log \\left(\\sum_ {p ^ {\\prime} \\in \\mathcal {N} _ {p}} e ^ {- \\left(p - p ^ {\\prime}\\right) ^ {2}}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.555, + 0.892, + 0.601 + ], + "angle": 0, + "content": "where \\(\\hat{E}(p)\\) denotes the estimated self-information at the pixel \\(p = (x, y)\\) of the original image \\(I\\); and \\(\\mathcal{N}_p\\) denotes the neighbouring pixels of \\(p\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.601, + 0.892, + 0.648 + ], + "angle": 0, + "content": "Then, we obtain the average pixel-wise self-information \\(mi\\) and perform a thresholding operation to obtain the self-information guided map \\(E_{info}\\) as follows" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.659, + 0.892, + 0.7 + ], + "angle": 0, + "content": "\\[\nE _ {i n f o} (p) = \\left\\{ \\begin{array}{l l} 1 & \\hat {E} (p) \\geq m i \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.709, + 0.892, + 0.831 + ], + "angle": 0, + "content": "Eq. (2) indicates that the greater the difference between a pixel and its neighbors, the more shape cues this pixel may contain. That is, for pixels close to the object boundary, their self-information values tend to be relatively large. Consequently, we use this self-information map to extend the edge map extracted by Robust Canny, to leverage the edge-surrounding cues to against the possible noises imposed on the edges, as follows" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.843, + 0.892, + 0.861 + ], + "angle": 0, + "content": "\\[\nE _ {\\text {e x t e n d}} = E _ {R C} + E _ {\\text {i n f o}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(E_{RC}\\) denotes the edge maps extracted by the robust Canny." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1528" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.396, + 0.108 + ], + "angle": 0, + "content": "3.2. TPS-based shape deformation (TSD)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.114, + 0.47, + 0.16 + ], + "angle": 0, + "content": "To augment the shape representation, we propose TPS-based deformation on the meshing of the self-information edge map in Eq. (3)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.16, + 0.47, + 0.265 + ], + "angle": 0, + "content": "For the shape deformation, Dabouei et al. [7] suggested to use the facial landmarks as control points, while this fashion brings about an overhead of the additional detection module and may damage semantic reasonability. By contrast, we propose to uniformly mesh the edge map, and use the endpoints of the grids as the control points to guide the edge map deformation for shape augmentation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.266, + 0.469, + 0.342 + ], + "angle": 0, + "content": "Specifically, the edge map \\( E_{\\mathrm{extend}} \\) is equally divided into \\( n = 16 \\) grids, and the endpoints are denoted as \\( \\{P_i, i = 1, \\dots, n\\} \\), which serve as the original control points and are randomly perturbed to generate target control points \\( \\{Q_i, 1 \\leq i \\leq n\\} \\) as follows" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.351, + 0.469, + 0.368 + ], + "angle": 0, + "content": "\\[\nQ _ {i} = P _ {i} + \\lambda \\mathcal {N} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.378, + 0.469, + 0.439 + ], + "angle": 0, + "content": "where random noise \\(\\mathcal{N}\\) obeys \\(N(0,1)\\), i.e. the normal distribution with mean 0 and variance 1, and the \\(\\lambda\\) controls the intensity of image deformation, which is set as 0.1. The edge map meshing is visualized in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.469, + 0.56 + ], + "angle": 0, + "content": "Given the original and target control points, the edge map deformation is then performed based on TPS. In particular, a transformation \\(\\Phi\\) is found to maximize the smoothness of the edge map deformation while minimizing the gap between two sets of control points. In this work, we use two transformations \\(\\Phi^x, \\Phi^y\\) [10], specific to the translation of the \\(x, y\\) coordinates of the control points, to obtain the deformed edge map \\(E_{\\text{deform}}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.57, + 0.469, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{c} E _ {\\text {d e f o r m}} (p) = \\left(p ^ {x} + \\Phi^ {x} (p), p ^ {y} + \\Phi^ {y} (p)\\right) \\\\ \\Phi^ {x} (p) = M ^ {x} \\cdot p + m _ {0} ^ {x} + \\sum_ {i} w _ {i} ^ {x} U (\\| p - P _ {i} \\| _ {2}) \\\\ M ^ {x}, m _ {0} ^ {x}, \\left\\{w _ {i} ^ {x} \\right\\} = \\\\ \\operatorname {a r g m i n} \\sum_ {i} \\left(\\Phi^ {x} \\left(P _ {i}\\right) - \\left(Q _ {i} ^ {x} - P _ {i} ^ {x}\\right)\\right) ^ {2} + \\mathcal {E} _ {\\Phi^ {x}} \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.642, + 0.469, + 0.718 + ], + "angle": 0, + "content": "where \\(U(\\cdot)\\) denotes the function of radial basis kernel, \\(M^x\\in \\mathbb{R}^2\\) \\(m_0^x,w_i^x\\in \\mathbb{R}\\) are the parameters specific to the \\(x\\) coordinate. \\(\\mathcal{E}_{\\Phi^x}\\) is the second-order smoothness term specific to the transformation \\(\\Phi^x\\) . For the \\(y\\) coordinate, the transformation \\(\\Phi^y\\) is similarly derived." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.443, + 0.743 + ], + "angle": 0, + "content": "3.3. Texture and shape-based generation (TSG)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.856 + ], + "angle": 0, + "content": "Since the edge map deformation may cause the texture distorted on the boundary regions, it may deteriorate the fidelity of the deformed shape and the shape-based recognition performance. Thus, we propose a new generative model based on a two-stage training to inpaint the texture of the deformed edge map, i.e. reconstructing this map into the original image domain for the data augmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "While the Pix2Pix network [20] can achieve the inpainting from an edge map to the original image representation, it can not well build up the correlation between the edge" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "map and the texture category, i.e. an image with correct shape yet mismatched texture category may be generated. Therefore, we propose to use both shape and texture cues in the generator to build up this correlation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.892, + 0.198 + ], + "angle": 0, + "content": "Specifically, to provide texture cues for the generator, the original image is smoothed with an easy-to-implement denoising step:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.202, + 0.892, + 0.218 + ], + "angle": 0, + "content": "\\[\nI _ {t x t} = \\operatorname {I n t e r p} (D S (I)) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.349 + ], + "angle": 0, + "content": "where Interp and \\( DS \\) denote interpolation and down-sampling methods, both of which are based on bilinear interpolation. \\( I_{txt} \\) denotes the resulted texture map. This denoised texture map aims to provide rough guidance cues for the synthesis of fine texture. To further provide the shape cues in the edge encoder of the generator, we use \\( E_{extend} \\) in Eq. (3) in the 1st-stage training, and \\( E_{deform} \\) in Eq. (5) in the 2nd-stage training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.44 + ], + "angle": 0, + "content": "To train the generator, a two-stage strategy is used. While a rough image is generated based on \\( E_{\\text{extend}} \\) and the denoised texture \\( I_{\\text{txt}} \\) in the first stage, which is fine-tuned with the input of the deformed edge map \\( E_{\\text{deform}} \\), and the additional supervision of an object classifier and a newly introduced shape-preservation loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.892, + 0.577 + ], + "angle": 0, + "content": "Specifically, the generator is trained in the 1st stage with the loss \\(\\mathcal{L}_{1st}\\), i.e. the sum of the adversarial loss \\(\\mathcal{L}_{gan}\\) [12], a feature matching loss [22] and a auxiliary classification loss referred from ACGAN [30], which is used to generate rough image without shape deformation. During the second stage, we jointly fine-tune the generator and the classifier with the supervision of the cross-entropy classification loss \\(\\mathcal{L}_{cls}\\) [37] in addition to \\(\\mathcal{L}_{gan}\\), to better represent the correlation between edge maps and textures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.655 + ], + "angle": 0, + "content": "In addition, to decrease the sensitivity of the generator against the deformed edge map and stabilize the training, we introduce a \\(l_{1}\\)-norm shape-preservation loss between the deformed edge map and that extracted from the generated image as:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.669, + 0.891, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e d g e}} = \\left\\| \\mathrm {e C N N} \\left(I _ {\\text {s y n}}\\right), E _ {\\text {d e f o r m}} \\right\\| _ {1} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.699, + 0.892, + 0.82 + ], + "angle": 0, + "content": "where \\( I_{syn} \\) and eCNN denotes the synthesized image and convolution network to extract edges in [37]. It's worth noting that the deformed shape \\( E_{deform} \\) serves as a soft guidance for generating the shape of the augmented sample, it differs from the hard constraint in the direct image deformation, where the object shape is changed strictly according to the target control points. Consequently, the training specific to the loss \\( \\mathcal{L}_{2nd} \\) in the fine tuning stage is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.835, + 0.891, + 0.857 + ], + "angle": 0, + "content": "\\[\n\\min _ {G, \\theta} \\max _ {D} \\mathcal {L} _ {2 n d} = \\mathcal {L} _ {g a n} + \\mathcal {L} _ {c l s} + \\mathcal {L} _ {e d g e} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\theta\\) denotes the parameters of the classifier. \\(G\\) and \\(D\\) denote the generator and the discriminator." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1529" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.24, + 0.108 + ], + "angle": 0, + "content": "3.4. Overall training" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.47, + 0.16 + ], + "angle": 0, + "content": "For clarity, our sample synthesis with edge deformation is also presented in Alg. 1, which is further used for online shape augmentation in each epoch of the network training." + }, + { + "type": "code_caption", + "bbox": [ + 0.078, + 0.174, + 0.46, + 0.19 + ], + "angle": 0, + "content": "Algorithm 1 Edge deformation-based sample generation." + }, + { + "type": "code", + "bbox": [ + 0.078, + 0.192, + 0.415, + 0.222 + ], + "angle": 0, + "content": "Input: input an image \\(I\\) and initialize the network Output: the shape-deformed image" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.223, + 0.397, + 0.237 + ], + "angle": 0, + "content": "1: Obtain the extended edge map with Eq. (3);" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.238, + 0.468, + 0.266 + ], + "angle": 0, + "content": "2: Mesh the edge map and perturb the control points with Eq. (4);" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.268, + 0.403, + 0.282 + ], + "angle": 0, + "content": "3: Perform edge map deformation with Eq. (5);" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.283, + 0.409, + 0.297 + ], + "angle": 0, + "content": "4: Obtain the denoised texture map with Eq. (6);" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.298, + 0.468, + 0.327 + ], + "angle": 0, + "content": "5: Perform the two-stage training for the generator based on \\(\\mathcal{L}_{1st}\\) and \\(\\mathcal{L}_{2nd}\\) in Eq. (8);" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.223, + 0.468, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.47, + 0.448 + ], + "angle": 0, + "content": "Regarding the training complexity, our algorithm requires only the similar runtime cost as the vanilla training, which is significantly lower than that of offline augmentation. This statement, together with the adversarial robustness of the online and offline augmentation for our method are studied with a toy experiment in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.461, + 0.21, + 0.478 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.485, + 0.373, + 0.501 + ], + "angle": 0, + "content": "4.1. Dataset and experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.469, + 0.569 + ], + "angle": 0, + "content": "Datasets We evaluate the robustness performance of the proposed algorithm under five attacks, i.e. [3, 5, 28, 29, 38] on four databases, i.e. Fashion MNIST (FM) [42], CelebA (CA) [25], CIFAR-10 (C10) [23] and ImageNet (IN) [8]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.469, + 0.614 + ], + "angle": 0, + "content": "Experimental setting ResNet50 [14] after the pretraining or ResNet18 is used as the backbone for ImageNet or the other three datasets, unless otherwise specified." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.469, + 0.72 + ], + "angle": 0, + "content": "We follow the protocol in EdgeNetRob [37] to evaluate the robustness performance under different adversarial attacks. For adversarial robustness evaluation, we use the \\( l_{\\infty} \\)-norm perturbation constrained in the range [0, 1], and the perturbation budgets, i.e. 25/255 for Fashion MNIST, and 8/255 for CIFAR-10, CelebA and ImageNet. CelebA is used for the task of gender recognition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.825 + ], + "angle": 0, + "content": "Evaluation metrics The performance is evaluated with the shape-bias metrics [11, 19], the clean accuracy and the accuracies under diverse attack scenarios. We follow EdgeNetRob [37] to evaluate the clean accuracy, i.e. the accuracy on the testing set generated by EMSE and TSG modules for SDbOA, and the performance on the original testing set for non-generation algorithms." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.9 + ], + "angle": 0, + "content": "To evaluate the capacity of our method in terms of clean accuracy preservation compared with state-of-the-art (SOTA) defense strategies and related data augmentation approaches, we use the performance of the baseline as the benchmark." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.091, + 0.885, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.287, + 0.892, + 0.385 + ], + "angle": 0, + "content": "Figure 3. Shape and texture sensitivity of the valilla net and ours against adversarial noises. (a) and (d) show a clean sample and an attacked sample generated by PGD-40 [15], with a texture patch and an edge map extracted by Robust Canny, (b) and (e) show the feature maps (FMs) with the highest \\( L_{2} \\)-norm activation representing these samples based on ResNet-50, and (c) and (d) show the FMs by ours. 'Dist' denotes the \\( L_{2} \\)-norm distance." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.392, + 0.866, + 0.409 + ], + "angle": 0, + "content": "4.2. Function Analysis of the proposed modules" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.417, + 0.813, + 0.432 + ], + "angle": 0, + "content": "4.2.1 Function of shape bias for robustness" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.892, + 0.64 + ], + "angle": 0, + "content": "Since adversarial noises are generated within a limited adversarial budget and sparsely distributed in perturbed regions, Fig. 3 shows that the object's shape of an image is less perturbed compared with its texture. Specifically, the edge maps reflecting the shape remain similar for the benign and attacked samples, while the feature maps of these samples with the vanilla net are largely different, i.e. their distance is large, reflecting that the yielded feature representation of object texture has already largely perturbed by the attack. By contrast, our algorithm yields significantly smaller feature map bias, suggesting a more shape-biased model can make feature representation more robust against adversarial perturbations." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.664, + 0.791, + 0.68 + ], + "angle": 0, + "content": "4.2.2 Function of the proposed modules" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The performance of TSD Fig. 4 illustrates the fundamental contribution of the TSD module for shifting from texture-bias to shape-bias. For a moderate magnitude of \\(\\lambda\\) in Eq. (4), e.g. \\(\\lambda = 0.1\\), augmented samples with diverse shapes and the similar texture are generated. To be specific, the constraint loss in Eq. (7) encourages the diversity of the generated shapes, and our GAN generator enhances the semantic rationality of the generated samples and the similarity of their textures. Consequently, these shape-augmented and texture semantic-preserved data could induce the networks to use more neurons to remember the diverse variations of an object shape, and encourage learned networks to make predictions mainly based on the shape cues, i.e. causing the learned networks to be more shape-biased." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1530" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.094, + 0.468, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.268, + 0.446, + 0.282 + ], + "angle": 0, + "content": "Figure 4. The t-SNE visualizations of features with varying \\(\\lambda\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.468, + 0.382 + ], + "angle": 0, + "content": "The performance of EMSE and TSG To shed light on how the introduced self-information edge map and TSG work, we visualize the produced edge maps with the Robust Canny [37] and the proposed self-information in Fig. 5, together with the generated samples with Pix2Pix [20] and our generator." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.468, + 0.504 + ], + "angle": 0, + "content": "The first row of Fig. 5 shows that the proposed self-information guided map can encode richer shape cues and represent shapes more accurately and robustly than those with mere Robust Canny. The second row of Fig. 5 shows that the proposed TSG generates images with well-preserved sharpness of object shape and realistic texture, i.e. better building up the connection between shape and the texture label, compared with Pix2Pix [20]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.513, + 0.295, + 0.528 + ], + "angle": 0, + "content": "4.3.Quantitative evaluation" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.536, + 0.334, + 0.551 + ], + "angle": 0, + "content": "4.3.1 Results of shape-bias metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.469, + 0.605 + ], + "angle": 0, + "content": "Geirhos et al. [11] proposed the first shape-bias evaluation metric based on the prediction accuracy for a texture-shape cue conflict dataset [11] as:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.612, + 0.469, + 0.645 + ], + "angle": 0, + "content": "\\[\ns b _ {G E} = \\frac {\\# \\text {C o r r e c t e d P r e d i c t i o n t o S h a p e L a b e l}}{\\# \\text {A l l t h e S a m p l e s}} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.469, + 0.684 + ], + "angle": 0, + "content": "where \\(\\#\\) denotes counting the number. This dataset [11] consists of images with shape and texture from different" + }, + { + "type": "image", + "bbox": [ + 0.143, + 0.705, + 0.406, + 0.852 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.855, + 0.469, + 0.911 + ], + "angle": 0, + "content": "Figure 5. Visualizations of samples from EMSE and TSG. (a) represents a vanilla sample. (b) and (c) represent edge maps extracted by Robust Canny [37] and EMSE. (d) and (e) represent images generated by Pix2Pix [20] and our TSG." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "classes to misguide the CNN. Specifically, the object textures of the images [11] are largely interfered with category-irrelevant styles, which places big challenge for current texture-biased networks." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.153, + 0.885, + 0.168 + ], + "angle": 0, + "content": "Islam et al. [19] proposed another shape-bias metric as:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.179, + 0.892, + 0.218 + ], + "angle": 0, + "content": "\\[\ns b _ {I S} = \\frac {\\sum_ {i = 1} ^ {n f} \\rho_ {i}}{n f}, \\text {w h e r e} \\rho_ {i} = \\frac {\\operatorname {C o v} \\left(z _ {i} ^ {a} , z _ {i} ^ {b}\\right)}{\\sqrt {\\operatorname {V a r} \\left(z _ {i} ^ {a}\\right) \\operatorname {V a r} \\left(z _ {i} ^ {b}\\right)}} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.892, + 0.366 + ], + "angle": 0, + "content": "where \\( nf \\) is the number of feature maps. \\( z_{i}^{a} \\) and \\( z_{i}^{b} \\) represents the \\( i \\)-th flattened feature maps corresponding to the image pair \\( (a, b) \\) from the texture-shape cue conflict dataset [11], while the two images contain the similar shape but have different textures. The normalized correlation coefficient \\( \\rho_{i} \\) measures the proportion of feature maps to represent the shape cues. The quantitative results of the standard network, StyleAug [21], ShapeAug [24] and our SDbOA in terms of the two quantitative metrics are shown in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Tab. 1 shows that our SDbOA consistently achieves the larger shape-bias values compared with the standard network and the related augmentation algorithms [21, 24] in term of both quantitative metrics for the training databases of ImageNet (IN), Stylized IN [11] (SIN) and SIN+IN." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.455, + 0.892, + 0.657 + ], + "angle": 0, + "content": "
MethodDatasetsbGE[11]sbIS[19]
StandardIN21.3917.0
StyleAug [21]IN67.3128.5
ShapeAug [24]IN38.4621.3
SDbOAIN71.2831.2
StandardSIN81.3726.2
StyleAug [21]SIN73.7232.7
ShapeAug [24]SIN67.7529.8
SDbOASIN82.4039.9
StandardSIN+IN34.6518.4
StyleAug [21]SIN+IN71.0430.4
ShapeAug [24]SIN+IN63.6723.1
SDbOASIN+IN78.7937.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.66, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Table 1. Quantitative results in terms of two shape-bias metrics \\((\\%)\\) with two shape-bias enhancement methods, i.e. StyleAug [21] (CVPRW'19) and ShapeAug [24] (CVPRW'22)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.842, + 0.726 + ], + "angle": 0, + "content": "4.3.2 Robustness against various perturbations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Robustness against adversarial attacks: To study the adversarial robustness of our SDbOA compared with the SO-TAs, we present the performances of seven algorithms in terms of clean accuracy and adversarial robustness against FGSM [13], PGD [28], C&W [3] and DeepFool [29] in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.899 + ], + "angle": 0, + "content": "As shown in Tab. 2, the models trained with SDbOA maintain relatively stronger robustness against various adversarial attacks, than those with the SOTAs. Specifically, SDbOA outperforms NuAT [36] by the margins of \\(14.10\\%\\) and \\(27.72\\%\\) in terms of the robustness under C&W and" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1531" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.47, + 0.121 + ], + "angle": 0, + "content": "DeepFool on CIFAR-10, and outperforms EdgeNetRob [37] by a margin of \\(6.98\\%\\) under C&W on ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.123, + 0.468, + 0.213 + ], + "angle": 0, + "content": "While Tab. 2 includes generic adversarial training (AT)-based SOTAs for the comparison, we further compare our algorithm with additional two AT-based algorithms [31,33] that are also based on generative model as similar as ours in Tab. 3, where the same protocol (evaluation with AutoAttack [5]) and network architectures as [31,33] are employed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.468, + 0.336 + ], + "angle": 0, + "content": "Since the works [31,33] require adversarial training, tens of millions of adversarial samples in addition to the vanilla images are generated for the training, resulting in a significantly larger runtime complexity. Despite this, Tab. 3 shows that our online augmentation model can still outperform the listed adversarial-training-based methods in terms of the clean accuracy and adversarial robustness with reasonable training overhead." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.351, + 0.456, + 0.809 + ], + "angle": 0, + "content": "
Data.MethodClean Acc.FGSM PGD CWDeep Fool
FMVanilla Net92.4227.420.5810.0039.5
PGD [28]86.5878.8481.2617.540.56
AT+CAS [1]87.4179.2378.4232.9531.46
NuAT [36]88.0281.4382.3767.0742.19
AT+CR [39]87.4680.8281.9534.5737.04
EdgeNetRob [37]85.0178.3075.7582.4771.43
SDbOA88.5280.3283.2687.2476.12
CAVanilla Net99.1848.860.0023.5720.80
PGD [28]92.8585.4083.3542.7819.84
AT+CAS [1]93.0689.8887.6865.3637.05
NuAT [36]94.2693.8089.5267.2529.42
AT+CR [39]95.3294.5890.7464.3839.06
EdgeNetRob [37]95.5292.6455.0875.9283.72
SDbOA98.7495.3488.2879.8491.50
C10Vanilla Net90.707.840.0016.7814.30
PGD [28]75.8254.5644.6618.020.38
AT+CAS [1]76.5162.3247.8151.4337.42
NuAT [36]79.5162.1246.4541.9634.66
AT+CR [39]82.8064.8153.3950.6643.19
EdgeNetRob [37]79.2162.4839.6643.4654.45
SDbOA83.2767.8250.9156.0662.38
INVanilla Net75.825.400.00-5.44
PGD [28]55.3440.4630.8316.0418.36
AT+CAS [1]57.0347.5836.7223.6727.92
NuAT [36]58.3848.1240.3021.7125.40
AT+CR [39]52.0341.7227.9321.4021.07
EdgeNetRob [37]65.7345.8733.6220.8523.40
SDbOA68.5649.1035.7627.8328.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.808, + 0.468, + 0.876 + ], + "angle": 0, + "content": "Table 2. Clean accuracy and adversarial robustness (%) with vanilla net, PGD Training [28], AT+CAS [1] (ICLR'21), NuAT [36] (NeurIPS'21), AT+CR [39] (AAAI'22), EdgeNetRob [37] (ICCV'21) and ours. The best and 2nd best performances are labeled with bold and underline." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Robustness of augmentation approaches against" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.089, + 0.899, + 0.176 + ], + "angle": 0, + "content": "
MethodCleanRobustRuntimeAT
FDA28 [31]85.9760.73Days✓
SDbOA2885.2768.153-5 Hours✗
PORT34 [33]87.0060.60Days✓
SDbOA3487.1273.363-5 Hours✗
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.18, + 0.892, + 0.277 + ], + "angle": 0, + "content": "Table 3. Adversarial robustness (\\%) of the generative-model-based adversarial training approaches, i.e. FDA [31] (NeurIPS'21) and PORT [33] (ICLR'22) on CIFAR-10. We follow the robustness metric in [31] measured by AutoAttack [5]. AT represents 'adversarial training'. '28' or '34' stands for using WideResNet-28-10 or WideResNet-34-10 [45] as the classifier network, following [31, 33]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.279, + 0.892, + 0.385 + ], + "angle": 0, + "content": "common corruptions and adversarial attacks: To evaluate the performances of our data augmentation and related augmentation techniques under common corruptions [15] and adversarial attacks, we present the results of nine SO-TAs in term of mean corruption error (mCE) and adversarial robustness in Tab. 4, where mCE is evaluated under all 15 corruptions and each corruption has 5 severities [15]." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.398, + 0.892, + 0.569 + ], + "angle": 0, + "content": "
MethodsCorruptions↓Adversaries↓
Baseline26.491.3
Cutout [9]arXiv25.996.0
Mixup [46]ICLR'1821.093.3
CutMix [44]ICCV'1926.592.1
AutoAugment [6]CVPR'1922.295.1
AugMix [16]ICLR'1912.486.8
AugMax [41]NeurIPS'2119.939.1
PixMix [17]CVPR'229.582.1
TPS-Deform [43]MM'2222.444.0
SDbOA18.827.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.573, + 0.892, + 0.643 + ], + "angle": 0, + "content": "Table 4. Robustness performances (\\%) of nine data augmentation techniques on CIFAR-10-C, i.e. the corruption set of CIFAR-10, based on the baseline of WideResNet-40-4 [45] under PGD-20 attack. The adversaries metric is the misclassification rate under PGD-20 with the \\( l_{\\infty} \\) budget of 2/255." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Tab. 4 shows that large majority of SOTA data augmentation techniques improve robustness against corruptions without paying enough attention on the adversarial robustness. It's worth noting that our algorithm achieves much better adversarial robustness, i.e. \\(27.7\\%\\) with large margins over those, i.e. \\(82.1\\%\\) and \\(86.8\\%\\) achieved by PixMix [17] and AugMax [41] that perform better than ours in terms of mCE. This maybe because that these two augmentations are mainly focused on robustness against a specified type of samples, while ours is developed for universal robustness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Robustness against backdoor attacks: To study the robustness of the proposed algorithm under backdoor attacks [40] compared with the related data augmentation methods, we follow the protocol in [37], and present the clean accuracy on the standard testing data and the attack success rate on the poisoned data (Pois. ASR) of seven algorithms, un" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1532" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "der backdoor attacks with two patterns in Tab. 5. Poisoning ratio is set as \\(20\\%\\) for Fashion MNIST and \\(5\\%\\) for CelebA." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.133, + 0.466, + 0.365 + ], + "angle": 0, + "content": "
Data.MethodBackdoor Pattern
PixelPattern
Clean Acc.↑Pois. ASR↓Clean Acc.↑Pois. ASR↓
FMVanilla Net87.4394.3087.1295.22
SpecSign [40]86.2345.6285.9352.31
EdgeNetRob [37]83.480.1282.212.74
AugMax [41]89.8354.6089.2773.14
PixMix [17]89.2592.0488.0194.78
TPS-Deform [43]86.3547.4386.2762.21
SDbOA87.250.0487.070.98
CAVanilla Net98.3097.2098.0097.40
SpecSign [40]98.4564.7897.9854.23
EdgeNetRob [37]92.8010.9093.1012.50
AugMax [41]98.8445.9798.2374.10
PixMix [17]98.1878.3297.9889.28
TPS-Deform [43]97.4371.3197.4774.52
SDbOA95.203.4994.874.21
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.365, + 0.469, + 0.42 + ], + "angle": 0, + "content": "Table 5. Performances (\\%) of vanilla net, Spectral Signature (SpecSign) [40] (NeurIPS'18), EdgeNetRob [37] (ICCV'21), AugMax [41] (NeurIPS'21), PixMix [17] (CVPR'22), TPS-Deform [43] (MM'22) and SDbOA under backdoor attacks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.47, + 0.546 + ], + "angle": 0, + "content": "Tab. 5 shows that SDbOA achieves the lowest ASR values against backdoor attacks on the two datasets, e.g. it significantly outperforms AugMax [41] that performs best in terms of clean accuracy, by the margins of \\(54.56\\%\\) and \\(72.16\\%\\) specific to the pixel and pattern attacks on Fashion MNIST. Meanwhile, the clean accuracy is reasonably preserved compared with the most related algorithm, i.e. EdgeNetRob [37]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.566, + 0.442, + 0.581 + ], + "angle": 0, + "content": "4.3.3 Hyperparameter analysis and ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.47, + 0.666 + ], + "angle": 0, + "content": "Analysis of hyperparameter \\(\\lambda\\): To analyze the sensitivity of the learned model's shape-bias against the deformation intensity, i.e. \\(\\lambda\\) in Eq. (4), we conduct a hyperparameter analysis and present the results in terms of the two shape-bias metrics [11, 19] in Fig. 6." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.683, + 0.436, + 0.819 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.825, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Figure 6. Sensitivity of shape-bias metric values [11, 19] against different deformation intensities \\((\\lambda)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Fig. 6 shows that a greater deformation degree does not always result in a more shape-biased model, i.e. the shape-bias of the model learned by SDbOA gradually increases" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.077, + 0.88, + 0.187 + ], + "angle": 0, + "content": "
EMSETSDTSGCleanFGSMPGDC&WDeepFool
74.3260.0646.7845.4855.82
✓76.5062.0847.0847.0554.25
✓✓78.2465.2649.8249.6161.35
✓✓87.6715.830.0021.8419.02
✓✓81.0866.8448.4052.3159.46
✓✓✓83.2767.8250.9156.0662.38
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.187, + 0.89, + 0.213 + ], + "angle": 0, + "content": "Table 6. Ablation study in terms of the clean accuracy and adversarial robustness (%) on CIFAR 10." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.216, + 0.892, + 0.261 + ], + "angle": 0, + "content": "when \\(\\lambda\\) is smaller than the threshold of 0.1, while it is decreased when \\(\\lambda\\) exceeds 0.1, i.e. the learned model gradually shifts from shape-bias to be more texture-biased." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.263, + 0.892, + 0.473 + ], + "angle": 0, + "content": "To shed light on this observation, we argue that a moderate deformation degree, e.g. \\(\\lambda = 0.1\\) in Fig. 4, is beneficial to reasonably improve the diversity of shape representations without largely impairing shape semantic, and a model learned from these valid shape variations can be made better against the perturbation noises. However, as shown in Fig. 4, excessive deformation may be caused when a large \\(\\lambda\\), e.g. \\(\\lambda = 0.2\\) is used, which may damage the global shape structure and original shape semantic, causing deformed samples to even cross classification boundaries. When it happens, these augmented samples with damaged shapes may instead encourage the learned network to give prediction based on the relatively stable texture, i.e. making the model to be more texture-biased." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.892, + 0.536 + ], + "angle": 0, + "content": "Ablation study: To study the performance of each module in our SDbOA, we perform the ablation study in Tab. 6, where Robust Canny [37] and Pix2Pix [20] are used as the baseline shape encoding and generator, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.689 + ], + "angle": 0, + "content": "By comparing the 5th and 7th rows of Tab. 6, it shows that the EMSE module is beneficial to the network robustness, reflecting the superiority of our edge augmentation over direct image deformation. By comparing the 6th and 7th rows, it shows that the proposed TSD enables the model to consistently improve the clean accuracy and the robustness performances under four attacks. By comparing the 4th and 7th rows, it shows that the improved generator with the supervision of texture cues and the shape-preservation loss largely improves the Pix2Pix in terms of the clean accuracy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.693, + 0.627, + 0.708 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Considering the network robustness maybe impaired by the excessive texture-bias, we shed light on how to shift from texture-bias to shape-bias for CNN models, i.e. we propose a novel edge deformation-based data augmentation that enables networks to elaborately learn the shape variations, allowing the well-learned networks to make predictions mainly based on the targets' shapes. Our data augmentation mainly differs from the related SOTAs that it is developed based on the deformation of object's edge map rather than the image, and can reasonably preserve the semantic rationality of the reconstructed samples with the joint supervision of object texture and soft constraint of de" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1533" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.227 + ], + "angle": 0, + "content": "formed shape cues. Experimental results reveal that our algorithm can largely enhance model's shape-bias in terms of two quantitative metrics, and rather competitive robustness against various perturbations compared with SOTAs. Considering the recognition system trained with the proposed shape augmentation is mainly applicable for pixelwise perturbation, yet may not work for patch-wise corruption where the shape structure is severely damaged, our future work will be devoted to solving it." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.473, + 0.349 + ], + "angle": 0, + "content": "Acknowledgements. The work was supported by Natural Science Foundation of China under grants no. 62276170, 82261138629, the Science and Technology Project of Guangdong Province under grants no. 2023A1515011549, 2023A1515010688, the Science and Technology Innovation Commission of Shenzhen under grants no. JCYJ20190808165203670, JCYJ20220531101412030." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.363, + 0.175, + 0.379 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.389, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[1] Yang Bai, Yuyuan Zeng, Yong Jiang, Shu-Tao Xia, Xingjun Ma, and Yisen Wang. Improving adversarial robustness via channel-wise activation suppressing. In International Conference on Learning Representations, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.446, + 0.472, + 0.502 + ], + "angle": 0, + "content": "[2] Nicholas Baker, Hongjing Lu, Gennady Erlikhman, and Philip J Kellman. Deep convolutional networks do not classify based on global object shape. PLoS computational biology, 14(12):e1006613, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.504, + 0.473, + 0.545 + ], + "angle": 0, + "content": "[3] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE symposium on security and privacy (sp), pages 39-57. IEEE, 2017. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.547, + 0.473, + 0.616 + ], + "angle": 0, + "content": "[4] Fabio M Carlucci, Antonio D'Innocente, Silvia Bucci, Barbara Caputo, and Tatiana Tommasi. Domain generalization by solving jigsaw puzzles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2229-2238, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.618, + 0.473, + 0.673 + ], + "angle": 0, + "content": "[5] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.675, + 0.473, + 0.743 + ], + "angle": 0, + "content": "[6] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 113-123, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.746, + 0.473, + 0.802 + ], + "angle": 0, + "content": "[7] Ali Dabouei, Sobhan Soleymani, Jeremy Dawson, and Nasser Nasrabadi. Fast geometrically-perturbed adversarial faces. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1979-1988. IEEE, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.803, + 0.473, + 0.858 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.86, + 0.473, + 0.901 + ], + "angle": 0, + "content": "[9] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 7" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.389, + 0.473, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.135 + ], + "angle": 0, + "content": "[10] Gianluca Donato and Serge Belongie. Approximate thin plate spline mappings. In European conference on computer vision, pages 21-31. Springer, 2002. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.895, + 0.217 + ], + "angle": 0, + "content": "[11] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations, 2018. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.895, + 0.273 + ], + "angle": 0, + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.895, + 0.315 + ], + "angle": 0, + "content": "[13] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.895, + 0.372 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.894, + 0.426 + ], + "angle": 0, + "content": "[15] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.895, + 0.483 + ], + "angle": 0, + "content": "[16] Dan Hendrycks, Norman Mu, Ekin D Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781, 2019. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.895, + 0.565 + ], + "angle": 0, + "content": "[17] Dan Hendrycks, Andy Zou, Mantas Mazeika, Leonard Tang, Bo Li, Dawn Song, and Jacob Steinhardt. Pixmix: Dreamlike pictures comprehensively improve safety measures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16783-16792, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.895, + 0.621 + ], + "angle": 0, + "content": "[18] Katherine Hermann, Ting Chen, and Simon Kornblith. The origins and prevalence of texture bias in convolutional neural networks. Advances in Neural Information Processing Systems, 33:19000-19015, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.895, + 0.691 + ], + "angle": 0, + "content": "[19] Md Amirul Islam, Matthew Kowal, Patrick Esser, Sen Jia, Bjorn Ommer, Konstantinos G Derpanis, and Neil Bruce. Shape or texture: Understanding discriminative features in cnns. In International Conference on Learning Representations, 2021. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.895, + 0.761 + ], + "angle": 0, + "content": "[20] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.895, + 0.817 + ], + "angle": 0, + "content": "[21] Philip TG Jackson, Amir Atapour Abarghouei, Stephen Bonner, Toby P Breckon, and Boguslaw Obara. Style augmentation: data augmentation via style randomization. In CVPR workshops, volume 6, pages 10-11, 2019. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.895, + 0.872 + ], + "angle": 0, + "content": "[22] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694-711. Springer, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[23] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1534" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[24] Sangjun Lee, Inwoo Hwang, Gi-Cheon Kang, and Byoung-Tak Zhang. Improving robustness to texture bias via shape-focused augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 4323-4331, June 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.176, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[25] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.471, + 0.287 + ], + "angle": 0, + "content": "[26] Raphael Gontijo Lopes, Dong Yin, Ben Poole, Justin Gilmer, and Ekin D Cubuk. Improving robustness without sacrificing accuracy with patch gaussian augmentation. arXiv preprint arXiv:1906.02611, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.472, + 0.357 + ], + "angle": 0, + "content": "[27] Cheng Luo, Qinliang Lin, Weicheng Xie, Bizhu Wu, Jinheng Xie, and Linlin Shen. Frequency-driven imperceptible adversarial attack on semantic similarity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15315-15324, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.47, + 0.424 + ], + "angle": 0, + "content": "[28] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.47, + 0.496 + ], + "angle": 0, + "content": "[29] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2574–2582, 2016. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.47, + 0.552 + ], + "angle": 0, + "content": "[30] Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier gans. In International conference on machine learning, pages 2642-2651. PMLR, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.469, + 0.62 + ], + "angle": 0, + "content": "[31] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan A Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Fixing data augmentation to improve adversarial robustness. In Advances in Neural Information Processing Systems, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[32] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan Andrei Calian, Florian Stimberg, Olivia Wiles, and Timothy A Mann. Data augmentation can improve robustness. Advances in Neural Information Processing Systems, 34:29935-29948, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.762 + ], + "angle": 0, + "content": "[33] Vikash Sehwag, Saeed Mahloujifar, Tinashe Handina, Sihui Dai, Chong Xiang, Mung Chiang, and Prateek Mittal. Robust learning meets generative models: Can proxy distributions improve adversarial robustness? In International Conference on Learning Representations, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[34] Baifeng Shi, Dinghuai Zhang, Qi Dai, Zhanxing Zhu, Yadong Mu, and Jingdong Wang. Informative dropout for robust representation learning: A shape-bias perspective. In International Conference on Machine Learning, pages 8828-8839. PMLR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.873 + ], + "angle": 0, + "content": "[35] Connor Shorten and Taghi M Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of big data, 6(1):1-48, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[36] Gaurang Sriramanan, Sravanti Addepalli, Arya Baburaj, et al. Towards efficient and effective adversarial train" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ing. Advances in Neural Information Processing Systems, 34:11821-11833, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[37] Mingjie Sun, Zichao Li, Chaowei Xiao, Haonan Qiu, Bhavya Kailkhura, Mingyan Liu, and Bo Li. Can shape structure features improve model robustness under diverse adversarial settings? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7526-7535, 2021. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[38] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[39] Jihoon Tack, Sihyun Yu, Jongheon Jeong, Minseon Kim, Sung Ju Hwang, and Jinwoo Shin. Consistency regularization for adversarial robustness. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 8414-8422, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.333, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[40] Brandon Tran, Jerry Li, and Aleksander Madry. Spectral signatures in backdoor attacks. Advances in neural information processing systems, 31, 2018. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[41] Haotao Wang, Chaowei Xiao, Jean Kossaifi, Zhiding Yu, Anima Anandkumar, and Zhangyang Wang. Augmax: Adversarial composition of random augmentations for robust training. Advances in neural information processing systems, 34:237-250, 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[42] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.502, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[43] Jun Yu, Guochen Xie, Zhongpeng Cai, Peng He, Fang Gao, and Qiang Ling. Micro expression generation with thin-plate spline motion model and face parsing. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7210–7214, 2022. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.574, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[44] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.644, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[45] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In *British Machine Vision Conference* 2016. British Machine Vision Association, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.686, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[46] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.742, + 0.892, + 0.798 + ], + "angle": 0, + "content": "[47] Tianyuan Zhang and Zhanxing Zhu. Interpreting adversarially trained convolutional neural networks. In International Conference on Machine Learning, pages 7502-7511. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.8, + 0.892, + 0.855 + ], + "angle": 0, + "content": "[48] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3657–3666, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.855 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1535" + } + ] +] \ No newline at end of file diff --git a/2023/Shift from Texture-bias to Shape-bias_ Edge Deformation-based Augmentation for Robust Object Recognition/e071fbaf-3036-4ac7-9bfd-fb6983ccf03f_origin.pdf b/2023/Shift from Texture-bias to Shape-bias_ Edge Deformation-based Augmentation for Robust Object Recognition/e071fbaf-3036-4ac7-9bfd-fb6983ccf03f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..272b6d558dd5c79f42e7897da21b6db5c6429f5c --- /dev/null +++ b/2023/Shift from Texture-bias to Shape-bias_ Edge Deformation-based Augmentation for Robust Object Recognition/e071fbaf-3036-4ac7-9bfd-fb6983ccf03f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7feb5c5124c537c65a1467d1e3fa35f599b2111089f4c81072f3bf63a0c02f54 +size 2557540 diff --git a/2023/Shift from Texture-bias to Shape-bias_ Edge Deformation-based Augmentation for Robust Object Recognition/full.md b/2023/Shift from Texture-bias to Shape-bias_ Edge Deformation-based Augmentation for Robust Object Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fc3c91ac53f321433b23911fa1b70f500549c893 --- /dev/null +++ b/2023/Shift from Texture-bias to Shape-bias_ Edge Deformation-based Augmentation for Robust Object Recognition/full.md @@ -0,0 +1,361 @@ +# Shift from Texture-bias to Shape-bias: Edge Deformation-based Augmentation for Robust Object Recognition + +Xilin He $^{1,2,3}$ , Qinliang Lin $^{1,2,3}$ , Cheng Luo $^{1,2,3}$ Weicheng Xie $^{1,2,3}$ , Siyang Song $^{4}$ , Feng Liu $^{1,2,3}$ , Linlin Shen $^{1,2,3}$ + +1Computer Vision Institute, School of Computer Science & Software Engineering, Shenzhen University + +2Shenzhen Institute of Artificial Intelligence & Robotics for Society + +3Guangdong Key Laboratory of Intelligent Information Processing + +$^{4}$ University of Leicester + +{2020152115, linqinliang2021, luocheng2020}@email.szu.edu.cn + +{wcxie,feng.liu,llshen}@szu.edu.cn,ss1535@leicester.ac.uk + +# Abstract + +Recent studies have shown the vulnerability of CNNs under perturbation noises, which is partially caused by the reason that the well-trained CNNs are too biased toward the object texture, i.e., they make predictions mainly based on texture cues. To reduce this texture-bias, current studies resort to learning augmented samples with heavily perturbed texture to make networks be more biased toward relatively stable shape cues. However, such methods usually fail to achieve real shape-biased networks due to the insufficient diversity of the shape cues. In this paper, we propose to augment the training dataset by generating semantically meaningful shapes and samples, via a shape deformation-based online augmentation, namely as SDbOA. The samples generated by our SDbOA have two main merits. First, the augmented samples with more diverse shape variations enable networks to learn the shape cues more elaborately, which encourages the network to be shape-biased. Second, semantic-meaningful shape-augmentation samples could be produced by jointly regularizing the generator with object texture and edge-guidance soft constraint, where the edges are represented more robustly with a self information guided map to better against the noises on them. Extensive experiments under various perturbation noises demonstrate the obvious superiority of our shape-bias-motivated model over the state of the arts in terms of robustness performance. Code is available at https://github.com/C0notSilly/-ICCV-23-Edge-Deformation-based-Online-Augmentation. + +![](images/0050a9b4b00509de6a17cdc2e21f40d5e7776e6b9bc49a8c602fc0d4018057ad.jpg) +Figure 1. Comparison of different models' predictions on texture-shape cue conflict samples [11] between Style Augmentation [21] (CVPRW'19), deformation augmentation in [43] and ours. Specific to the testing image (a), the bottom row shows the original (b) and augmented (c)-(e) images for training. + +# 1. Introduction + +Previous studies [13, 15, 38] have shown that convolution neural networks (CNNs) are vulnerable to perturbation noises, e.g., adversarial noises. Such perturbation noises imposed on the benign image can easily falsify model's prediction [28]. To improve the network robustness against these noises, algorithms of texture augmentation have been long studied to enable network to learn robust and discriminative object texture features of different categories [11, 24]. + +However, recent studies revealed that the vulnerability of CNNs maybe caused by their 'texture-bias', i.e., the features learned for object recognition bias towards objects' textures instead of their shapes [2, 11]. As illustrated in Fig. 1, when the input image (a) contains a cat that is filled with an elephant's texture, standard CNN networks are more likely to predict it as an elephant despite the obvious shape difference between cats and elephants. + +Consequently, some studies devote to reducing such + +texture-bias to improve network robustness. Zhang et al. [47] found that CNNs learned with adversarial training strategies are less texture-biased. Alternatively, shape-biased CNNs are revealed to be more robust under diverse adversarial attacks. Sun et al. [37] found that the network adversarial robustness can be improved by directly using edge maps to aid the training process. Carlucci et al. [4] proposed to train CNNs on jigsaw puzzles, which achieved superior performances by enabling the network to pay more attention on the cues contained in the global structure. These works illustrate that the shape and structure features, e.g. edge map, can reduce the texture-bias with a stronger shape-bias, i.e. CNNs make predictions based on more shape cues, rather than texture cues. + +However, current studies mainly learned the features of the object shape that is fixed, while the advantage of the shape diversity for devising a more shape-biased model is not well studied. Meanwhile, direct deformation augmentation on images may destroy the object semantics, especially on the image boundary. Given the above observations, following question naturally emerged: Can we implement data augmentation to the edge map to enhance the shape cues and reduce texture-bias? + +To this end, we propose a thin plate spline (TPS) [10]-based edge map deformation strategy to augment the shape representation of the training data, aiming to enhance the shape-bias of the well-trained network. To make the encoded object shape less sensitive to the noises, we extend the edge map specific to only the object edges to include broader boundary regions via the self-information guided map [34]. Furthermore, to build up the correlation between the deformed shape and the texture category to make the augmented samples more semantic-meaningful, we improve the generator with the additional supervision of denoised texture cues and a shape-preservation loss. + +Fig. 1 sheds light on the above motivation of our shape augmentation for enhancing networks' shape-bias with a toy experiment, where the probabilities to the shape label of an example sample [11] are predicted by the standard ResNet-50, the Style Augmentation [21], the image deformation augmentation [43] and our method. Fig. 1 shows that the proposed algorithm achieves significantly larger probability predicted to the shape label, i.e. tabby cat over the baseline and the related approaches, which reveals the obvious tendency of the proposed SDbOA in applying shape cues for making predictions. For this observation, we argue that the proposed edge-deformation augmentation enables networks to learn the object shape cues more elaborately, while reducing the contribution of the texture cues for prediction. + +To the best of our knowledge, this work is one of the pioneer attempts to enhance shape-bias of networks based on the augmentation of object's edges. The main contributions + +of this paper are summarized as follows: + +- We propose a TPS-based edge deformation to augment the object shape, and apply it to the online data augmentation, which could help learning more shape-biased networks. +- We propose a semantic-meaningful generation paradigm to produce shape-augmentation samples by jointly regularizing the generator with object texture and edge-guidance soft constraint, where a self information guided map is introduced to represent these edge cues more accurately and robustly. +- Extensive results show the superiority of the proposed algorithm over state of the arts (SOTAs) in terms of universal robustness under adversarial, backdoor and common corruption attacks and the training overhead. + +# 2. Related work + +# 2.1. Texture or shape-biased models associated with recognition network robustness + +Adversarial attacks have exposed the vulnerability of networks to perturbation noises [3, 28, 29, 38], leading to a growing concern for network robustness. Recent studies have investigated the association between network robustness and object texture or shape. Geirhos et al. [11] discovered that CNNs trained on ImageNet rely heavily on texture cues, leading to a texture bias in predictions. To address this issue, Hermann et al. [18] proposed additive augmentation techniques such as color distortion or blurring. However, the contribution of shape cues to recognition robustness has not received sufficient attention in their works. + +Recently, Shi et al. [34] proposed an informative dropout approach to help models learn robust representation from a shape-bias perspective. Sun et al. [37] proposed to explicitly use object edges as semantically robust features to against adversarial attacks. Since the object shapes used for enhancing shape-bias are fixed, both approaches are unable to learn a real shape-biased network. + +In this work, we thus resort to the edge deformation and augmentation to improve the diversity of object shapes. Based on this, we attempt to enhance shape-bias, thus enabling the network to shift its attention from texture to more on the shapes, and better discriminating the samples that have different textures but the similar shape. + +# 2.2. Data augmentation for recognition robustness improvement + +Data augmentation is a popular paradigm for improving the robustness of recognition network against various perturbation noises [16, 17, 21, 24, 27, 32, 41]. This strategy can be applied to generic computer vision tasks, which differs + +![](images/11a718bfe8ea60788098a431444de5e31a3b9896b2ee1208da66f63e98398238.jpg) +Figure 2. Pipeline of the proposed SDbOA. For the training, all the three modules, i.e. Edge map-based shape encoding (EMSE), TPS-based shape deformation (TSD) and Texture and shape-based generation (TSG) are used. For the testing, the TSD module is not needed. $\oslash$ denotes the 'OR' operation. + +from some task-specific robustness improvement strategies such as adversarial defense algorithms [1, 28, 31, 33, 36, 39] that are designed for a specific attack scenario. + +Lopes et al. [26] augmented the samples to trade off robustness against recognition accuracy via adding noises on the selected patches of a benign image. Hendrycks et al. [16] proposed the AugMix augmentation, i.e. mixing randomly generated augmentations to improve model robustness under common corruptions. Jackson et al. [21] proposed to augment the data with the style transfer to mitigate domain bias and reduce overfitting. Lee et al. [24] proposed to largely perturb the object texture to help model concentrate on the invariant shape structure. + +Most augmentation methods [16, 21, 24, 26, 32] either focus on augmenting texture cues or fail to modify object shape regarding edges or structure, such as geometric transformations [32, 35]. However, compared to image deformation-based augmentation methods [10, 43, 48], which strictly adhere to shape constraints, this hard guidance may disregard the semantic coherence of the deformed sample, as shown in Fig. 1. + +In this work, we propose an object shape augmentation approach based on the TPS-based [10] edge deformation, with the aid of an extended edge encoding and a generator with a soft constraint loss for the shape representation, to dynamically enrich the object shapes. + +# 3. Methodology + +The algorithm architecture, shown in Fig. 2, comprises three main modules: (i) the edge map-based shape encoding (EMSE) module used for robustly representing object shape; (ii) the TPS-based shape deformation (TSD) module used for providing flexible shape augmentation; and (iii) the texture and shape-based generation (TSG) module used for augmenting the samples dynamically, guided by the deformed shape. + +# 3.1. Edge map-based shape encoding (EMSE) + +To encode the object shape, the EMSE consists of two blocks. Firstly, a Robust Canny [37] block that generates an edge map for each image, describing the raw shape of the object in the image. While the Robust Canny may disregard some important shape cues of the object, we then additionally encode a self-information guided map [34] to provide complementary shape structure cues in addition to the Canny operator. Specifically, an edge map $\hat{E}$ that encodes the pixel-wise self information is first computed as: + +$$ +\hat {E} (p) = - \log \left(\sum_ {p ^ {\prime} \in \mathcal {N} _ {p}} e ^ {- \left(p - p ^ {\prime}\right) ^ {2}}\right) \tag {1} +$$ + +where $\hat{E}(p)$ denotes the estimated self-information at the pixel $p = (x, y)$ of the original image $I$ ; and $\mathcal{N}_p$ denotes the neighbouring pixels of $p$ . + +Then, we obtain the average pixel-wise self-information $mi$ and perform a thresholding operation to obtain the self-information guided map $E_{info}$ as follows + +$$ +E _ {i n f o} (p) = \left\{ \begin{array}{l l} 1 & \hat {E} (p) \geq m i \\ 0 & \text {o t h e r w i s e} \end{array} \right. \tag {2} +$$ + +Eq. (2) indicates that the greater the difference between a pixel and its neighbors, the more shape cues this pixel may contain. That is, for pixels close to the object boundary, their self-information values tend to be relatively large. Consequently, we use this self-information map to extend the edge map extracted by Robust Canny, to leverage the edge-surrounding cues to against the possible noises imposed on the edges, as follows + +$$ +E _ {\text {e x t e n d}} = E _ {R C} + E _ {\text {i n f o}} \tag {3} +$$ + +where $E_{RC}$ denotes the edge maps extracted by the robust Canny. + +# 3.2. TPS-based shape deformation (TSD) + +To augment the shape representation, we propose TPS-based deformation on the meshing of the self-information edge map in Eq. (3). + +For the shape deformation, Dabouei et al. [7] suggested to use the facial landmarks as control points, while this fashion brings about an overhead of the additional detection module and may damage semantic reasonability. By contrast, we propose to uniformly mesh the edge map, and use the endpoints of the grids as the control points to guide the edge map deformation for shape augmentation. + +Specifically, the edge map $E_{\mathrm{extend}}$ is equally divided into $n = 16$ grids, and the endpoints are denoted as $\{P_i, i = 1, \dots, n\}$ , which serve as the original control points and are randomly perturbed to generate target control points $\{Q_i, 1 \leq i \leq n\}$ as follows + +$$ +Q _ {i} = P _ {i} + \lambda \mathcal {N} \tag {4} +$$ + +where random noise $\mathcal{N}$ obeys $N(0,1)$ , i.e. the normal distribution with mean 0 and variance 1, and the $\lambda$ controls the intensity of image deformation, which is set as 0.1. The edge map meshing is visualized in Fig. 2. + +Given the original and target control points, the edge map deformation is then performed based on TPS. In particular, a transformation $\Phi$ is found to maximize the smoothness of the edge map deformation while minimizing the gap between two sets of control points. In this work, we use two transformations $\Phi^x, \Phi^y$ [10], specific to the translation of the $x, y$ coordinates of the control points, to obtain the deformed edge map $E_{\text{deform}}$ as: + +$$ +\left\{ \begin{array}{c} E _ {\text {d e f o r m}} (p) = \left(p ^ {x} + \Phi^ {x} (p), p ^ {y} + \Phi^ {y} (p)\right) \\ \Phi^ {x} (p) = M ^ {x} \cdot p + m _ {0} ^ {x} + \sum_ {i} w _ {i} ^ {x} U (\| p - P _ {i} \| _ {2}) \\ M ^ {x}, m _ {0} ^ {x}, \left\{w _ {i} ^ {x} \right\} = \\ \operatorname {a r g m i n} \sum_ {i} \left(\Phi^ {x} \left(P _ {i}\right) - \left(Q _ {i} ^ {x} - P _ {i} ^ {x}\right)\right) ^ {2} + \mathcal {E} _ {\Phi^ {x}} \end{array} \right. \tag {5} +$$ + +where $U(\cdot)$ denotes the function of radial basis kernel, $M^x\in \mathbb{R}^2$ $m_0^x,w_i^x\in \mathbb{R}$ are the parameters specific to the $x$ coordinate. $\mathcal{E}_{\Phi^x}$ is the second-order smoothness term specific to the transformation $\Phi^x$ . For the $y$ coordinate, the transformation $\Phi^y$ is similarly derived. + +# 3.3. Texture and shape-based generation (TSG) + +Since the edge map deformation may cause the texture distorted on the boundary regions, it may deteriorate the fidelity of the deformed shape and the shape-based recognition performance. Thus, we propose a new generative model based on a two-stage training to inpaint the texture of the deformed edge map, i.e. reconstructing this map into the original image domain for the data augmentation. + +While the Pix2Pix network [20] can achieve the inpainting from an edge map to the original image representation, it can not well build up the correlation between the edge + +map and the texture category, i.e. an image with correct shape yet mismatched texture category may be generated. Therefore, we propose to use both shape and texture cues in the generator to build up this correlation. + +Specifically, to provide texture cues for the generator, the original image is smoothed with an easy-to-implement denoising step: + +$$ +I _ {t x t} = \operatorname {I n t e r p} (D S (I)) \tag {6} +$$ + +where Interp and $DS$ denote interpolation and down-sampling methods, both of which are based on bilinear interpolation. $I_{txt}$ denotes the resulted texture map. This denoised texture map aims to provide rough guidance cues for the synthesis of fine texture. To further provide the shape cues in the edge encoder of the generator, we use $E_{extend}$ in Eq. (3) in the 1st-stage training, and $E_{deform}$ in Eq. (5) in the 2nd-stage training. + +To train the generator, a two-stage strategy is used. While a rough image is generated based on $E_{\text{extend}}$ and the denoised texture $I_{\text{txt}}$ in the first stage, which is fine-tuned with the input of the deformed edge map $E_{\text{deform}}$ , and the additional supervision of an object classifier and a newly introduced shape-preservation loss. + +Specifically, the generator is trained in the 1st stage with the loss $\mathcal{L}_{1st}$ , i.e. the sum of the adversarial loss $\mathcal{L}_{gan}$ [12], a feature matching loss [22] and a auxiliary classification loss referred from ACGAN [30], which is used to generate rough image without shape deformation. During the second stage, we jointly fine-tune the generator and the classifier with the supervision of the cross-entropy classification loss $\mathcal{L}_{cls}$ [37] in addition to $\mathcal{L}_{gan}$ , to better represent the correlation between edge maps and textures. + +In addition, to decrease the sensitivity of the generator against the deformed edge map and stabilize the training, we introduce a $l_{1}$ -norm shape-preservation loss between the deformed edge map and that extracted from the generated image as: + +$$ +\mathcal {L} _ {\text {e d g e}} = \left\| \mathrm {e C N N} \left(I _ {\text {s y n}}\right), E _ {\text {d e f o r m}} \right\| _ {1} \tag {7} +$$ + +where $I_{syn}$ and eCNN denotes the synthesized image and convolution network to extract edges in [37]. It's worth noting that the deformed shape $E_{deform}$ serves as a soft guidance for generating the shape of the augmented sample, it differs from the hard constraint in the direct image deformation, where the object shape is changed strictly according to the target control points. Consequently, the training specific to the loss $\mathcal{L}_{2nd}$ in the fine tuning stage is formulated as: + +$$ +\min _ {G, \theta} \max _ {D} \mathcal {L} _ {2 n d} = \mathcal {L} _ {g a n} + \mathcal {L} _ {c l s} + \mathcal {L} _ {e d g e} \tag {8} +$$ + +where $\theta$ denotes the parameters of the classifier. $G$ and $D$ denote the generator and the discriminator. + +# 3.4. Overall training + +For clarity, our sample synthesis with edge deformation is also presented in Alg. 1, which is further used for online shape augmentation in each epoch of the network training. + +Algorithm 1 Edge deformation-based sample generation. +```txt +Input: input an image $I$ and initialize the network Output: the shape-deformed image +``` + +1: Obtain the extended edge map with Eq. (3); +2: Mesh the edge map and perturb the control points with Eq. (4); +3: Perform edge map deformation with Eq. (5); +4: Obtain the denoised texture map with Eq. (6); +5: Perform the two-stage training for the generator based on $\mathcal{L}_{1st}$ and $\mathcal{L}_{2nd}$ in Eq. (8); + +Regarding the training complexity, our algorithm requires only the similar runtime cost as the vanilla training, which is significantly lower than that of offline augmentation. This statement, together with the adversarial robustness of the online and offline augmentation for our method are studied with a toy experiment in the supplementary material. + +# 4. Experiments + +# 4.1. Dataset and experimental settings + +Datasets We evaluate the robustness performance of the proposed algorithm under five attacks, i.e. [3, 5, 28, 29, 38] on four databases, i.e. Fashion MNIST (FM) [42], CelebA (CA) [25], CIFAR-10 (C10) [23] and ImageNet (IN) [8]. + +Experimental setting ResNet50 [14] after the pretraining or ResNet18 is used as the backbone for ImageNet or the other three datasets, unless otherwise specified. + +We follow the protocol in EdgeNetRob [37] to evaluate the robustness performance under different adversarial attacks. For adversarial robustness evaluation, we use the $l_{\infty}$ -norm perturbation constrained in the range [0, 1], and the perturbation budgets, i.e. 25/255 for Fashion MNIST, and 8/255 for CIFAR-10, CelebA and ImageNet. CelebA is used for the task of gender recognition. + +Evaluation metrics The performance is evaluated with the shape-bias metrics [11, 19], the clean accuracy and the accuracies under diverse attack scenarios. We follow EdgeNetRob [37] to evaluate the clean accuracy, i.e. the accuracy on the testing set generated by EMSE and TSG modules for SDbOA, and the performance on the original testing set for non-generation algorithms. + +To evaluate the capacity of our method in terms of clean accuracy preservation compared with state-of-the-art (SOTA) defense strategies and related data augmentation approaches, we use the performance of the baseline as the benchmark. + +![](images/21a4d66a410e73915318eeebb660238d2713b9eeda866200db9e51806036b7a5.jpg) +Figure 3. Shape and texture sensitivity of the valilla net and ours against adversarial noises. (a) and (d) show a clean sample and an attacked sample generated by PGD-40 [15], with a texture patch and an edge map extracted by Robust Canny, (b) and (e) show the feature maps (FMs) with the highest $L_{2}$ -norm activation representing these samples based on ResNet-50, and (c) and (d) show the FMs by ours. 'Dist' denotes the $L_{2}$ -norm distance. + +# 4.2. Function Analysis of the proposed modules + +# 4.2.1 Function of shape bias for robustness + +Since adversarial noises are generated within a limited adversarial budget and sparsely distributed in perturbed regions, Fig. 3 shows that the object's shape of an image is less perturbed compared with its texture. Specifically, the edge maps reflecting the shape remain similar for the benign and attacked samples, while the feature maps of these samples with the vanilla net are largely different, i.e. their distance is large, reflecting that the yielded feature representation of object texture has already largely perturbed by the attack. By contrast, our algorithm yields significantly smaller feature map bias, suggesting a more shape-biased model can make feature representation more robust against adversarial perturbations. + +# 4.2.2 Function of the proposed modules + +The performance of TSD Fig. 4 illustrates the fundamental contribution of the TSD module for shifting from texture-bias to shape-bias. For a moderate magnitude of $\lambda$ in Eq. (4), e.g. $\lambda = 0.1$ , augmented samples with diverse shapes and the similar texture are generated. To be specific, the constraint loss in Eq. (7) encourages the diversity of the generated shapes, and our GAN generator enhances the semantic rationality of the generated samples and the similarity of their textures. Consequently, these shape-augmented and texture semantic-preserved data could induce the networks to use more neurons to remember the diverse variations of an object shape, and encourage learned networks to make predictions mainly based on the shape cues, i.e. causing the learned networks to be more shape-biased. + +![](images/1e046d9c632e2bc8be9f9862270136a9cfb6b9d3f6ec9a46901b2d6090610582.jpg) +Figure 4. The t-SNE visualizations of features with varying $\lambda$ . + +The performance of EMSE and TSG To shed light on how the introduced self-information edge map and TSG work, we visualize the produced edge maps with the Robust Canny [37] and the proposed self-information in Fig. 5, together with the generated samples with Pix2Pix [20] and our generator. + +The first row of Fig. 5 shows that the proposed self-information guided map can encode richer shape cues and represent shapes more accurately and robustly than those with mere Robust Canny. The second row of Fig. 5 shows that the proposed TSG generates images with well-preserved sharpness of object shape and realistic texture, i.e. better building up the connection between shape and the texture label, compared with Pix2Pix [20]. + +# 4.3.Quantitative evaluation + +# 4.3.1 Results of shape-bias metrics + +Geirhos et al. [11] proposed the first shape-bias evaluation metric based on the prediction accuracy for a texture-shape cue conflict dataset [11] as: + +$$ +s b _ {G E} = \frac {\# \text {C o r r e c t e d P r e d i c t i o n t o S h a p e L a b e l}}{\# \text {A l l t h e S a m p l e s}} \tag {9} +$$ + +where $\#$ denotes counting the number. This dataset [11] consists of images with shape and texture from different + +![](images/63ae311d3f9ab30a3a4efe53a84e807e97cc1bdfb3b825229d6c11b1c0413d19.jpg) +Figure 5. Visualizations of samples from EMSE and TSG. (a) represents a vanilla sample. (b) and (c) represent edge maps extracted by Robust Canny [37] and EMSE. (d) and (e) represent images generated by Pix2Pix [20] and our TSG. + +classes to misguide the CNN. Specifically, the object textures of the images [11] are largely interfered with category-irrelevant styles, which places big challenge for current texture-biased networks. + +Islam et al. [19] proposed another shape-bias metric as: + +$$ +s b _ {I S} = \frac {\sum_ {i = 1} ^ {n f} \rho_ {i}}{n f}, \text {w h e r e} \rho_ {i} = \frac {\operatorname {C o v} \left(z _ {i} ^ {a} , z _ {i} ^ {b}\right)}{\sqrt {\operatorname {V a r} \left(z _ {i} ^ {a}\right) \operatorname {V a r} \left(z _ {i} ^ {b}\right)}} \tag {10} +$$ + +where $nf$ is the number of feature maps. $z_{i}^{a}$ and $z_{i}^{b}$ represents the $i$ -th flattened feature maps corresponding to the image pair $(a, b)$ from the texture-shape cue conflict dataset [11], while the two images contain the similar shape but have different textures. The normalized correlation coefficient $\rho_{i}$ measures the proportion of feature maps to represent the shape cues. The quantitative results of the standard network, StyleAug [21], ShapeAug [24] and our SDbOA in terms of the two quantitative metrics are shown in Tab. 1. + +Tab. 1 shows that our SDbOA consistently achieves the larger shape-bias values compared with the standard network and the related augmentation algorithms [21, 24] in term of both quantitative metrics for the training databases of ImageNet (IN), Stylized IN [11] (SIN) and SIN+IN. + +
MethodDatasetsbGE[11]sbIS[19]
StandardIN21.3917.0
StyleAug [21]IN67.3128.5
ShapeAug [24]IN38.4621.3
SDbOAIN71.2831.2
StandardSIN81.3726.2
StyleAug [21]SIN73.7232.7
ShapeAug [24]SIN67.7529.8
SDbOASIN82.4039.9
StandardSIN+IN34.6518.4
StyleAug [21]SIN+IN71.0430.4
ShapeAug [24]SIN+IN63.6723.1
SDbOASIN+IN78.7937.5
+ +Table 1. Quantitative results in terms of two shape-bias metrics $(\%)$ with two shape-bias enhancement methods, i.e. StyleAug [21] (CVPRW'19) and ShapeAug [24] (CVPRW'22). + +# 4.3.2 Robustness against various perturbations + +Robustness against adversarial attacks: To study the adversarial robustness of our SDbOA compared with the SO-TAs, we present the performances of seven algorithms in terms of clean accuracy and adversarial robustness against FGSM [13], PGD [28], C&W [3] and DeepFool [29] in Tab. 2. + +As shown in Tab. 2, the models trained with SDbOA maintain relatively stronger robustness against various adversarial attacks, than those with the SOTAs. Specifically, SDbOA outperforms NuAT [36] by the margins of $14.10\%$ and $27.72\%$ in terms of the robustness under C&W and + +DeepFool on CIFAR-10, and outperforms EdgeNetRob [37] by a margin of $6.98\%$ under C&W on ImageNet. + +While Tab. 2 includes generic adversarial training (AT)-based SOTAs for the comparison, we further compare our algorithm with additional two AT-based algorithms [31,33] that are also based on generative model as similar as ours in Tab. 3, where the same protocol (evaluation with AutoAttack [5]) and network architectures as [31,33] are employed. + +Since the works [31,33] require adversarial training, tens of millions of adversarial samples in addition to the vanilla images are generated for the training, resulting in a significantly larger runtime complexity. Despite this, Tab. 3 shows that our online augmentation model can still outperform the listed adversarial-training-based methods in terms of the clean accuracy and adversarial robustness with reasonable training overhead. + +
Data.MethodClean Acc.FGSM PGD CWDeep Fool
FMVanilla Net92.4227.420.5810.0039.5
PGD [28]86.5878.8481.2617.540.56
AT+CAS [1]87.4179.2378.4232.9531.46
NuAT [36]88.0281.4382.3767.0742.19
AT+CR [39]87.4680.8281.9534.5737.04
EdgeNetRob [37]85.0178.3075.7582.4771.43
SDbOA88.5280.3283.2687.2476.12
CAVanilla Net99.1848.860.0023.5720.80
PGD [28]92.8585.4083.3542.7819.84
AT+CAS [1]93.0689.8887.6865.3637.05
NuAT [36]94.2693.8089.5267.2529.42
AT+CR [39]95.3294.5890.7464.3839.06
EdgeNetRob [37]95.5292.6455.0875.9283.72
SDbOA98.7495.3488.2879.8491.50
C10Vanilla Net90.707.840.0016.7814.30
PGD [28]75.8254.5644.6618.020.38
AT+CAS [1]76.5162.3247.8151.4337.42
NuAT [36]79.5162.1246.4541.9634.66
AT+CR [39]82.8064.8153.3950.6643.19
EdgeNetRob [37]79.2162.4839.6643.4654.45
SDbOA83.2767.8250.9156.0662.38
INVanilla Net75.825.400.00-5.44
PGD [28]55.3440.4630.8316.0418.36
AT+CAS [1]57.0347.5836.7223.6727.92
NuAT [36]58.3848.1240.3021.7125.40
AT+CR [39]52.0341.7227.9321.4021.07
EdgeNetRob [37]65.7345.8733.6220.8523.40
SDbOA68.5649.1035.7627.8328.89
+ +Robustness of augmentation approaches against + +Table 2. Clean accuracy and adversarial robustness (%) with vanilla net, PGD Training [28], AT+CAS [1] (ICLR'21), NuAT [36] (NeurIPS'21), AT+CR [39] (AAAI'22), EdgeNetRob [37] (ICCV'21) and ours. The best and 2nd best performances are labeled with bold and underline. + +
MethodCleanRobustRuntimeAT
FDA28 [31]85.9760.73Days✓
SDbOA2885.2768.153-5 Hours✗
PORT34 [33]87.0060.60Days✓
SDbOA3487.1273.363-5 Hours✗
+ +common corruptions and adversarial attacks: To evaluate the performances of our data augmentation and related augmentation techniques under common corruptions [15] and adversarial attacks, we present the results of nine SO-TAs in term of mean corruption error (mCE) and adversarial robustness in Tab. 4, where mCE is evaluated under all 15 corruptions and each corruption has 5 severities [15]. + +Table 3. Adversarial robustness (\%) of the generative-model-based adversarial training approaches, i.e. FDA [31] (NeurIPS'21) and PORT [33] (ICLR'22) on CIFAR-10. We follow the robustness metric in [31] measured by AutoAttack [5]. AT represents 'adversarial training'. '28' or '34' stands for using WideResNet-28-10 or WideResNet-34-10 [45] as the classifier network, following [31, 33]. + +
MethodsCorruptions↓Adversaries↓
Baseline26.491.3
Cutout [9]arXiv25.996.0
Mixup [46]ICLR'1821.093.3
CutMix [44]ICCV'1926.592.1
AutoAugment [6]CVPR'1922.295.1
AugMix [16]ICLR'1912.486.8
AugMax [41]NeurIPS'2119.939.1
PixMix [17]CVPR'229.582.1
TPS-Deform [43]MM'2222.444.0
SDbOA18.827.7
+ +Table 4. Robustness performances (\%) of nine data augmentation techniques on CIFAR-10-C, i.e. the corruption set of CIFAR-10, based on the baseline of WideResNet-40-4 [45] under PGD-20 attack. The adversaries metric is the misclassification rate under PGD-20 with the $l_{\infty}$ budget of 2/255. + +Tab. 4 shows that large majority of SOTA data augmentation techniques improve robustness against corruptions without paying enough attention on the adversarial robustness. It's worth noting that our algorithm achieves much better adversarial robustness, i.e. $27.7\%$ with large margins over those, i.e. $82.1\%$ and $86.8\%$ achieved by PixMix [17] and AugMax [41] that perform better than ours in terms of mCE. This maybe because that these two augmentations are mainly focused on robustness against a specified type of samples, while ours is developed for universal robustness. + +Robustness against backdoor attacks: To study the robustness of the proposed algorithm under backdoor attacks [40] compared with the related data augmentation methods, we follow the protocol in [37], and present the clean accuracy on the standard testing data and the attack success rate on the poisoned data (Pois. ASR) of seven algorithms, un + +der backdoor attacks with two patterns in Tab. 5. Poisoning ratio is set as $20\%$ for Fashion MNIST and $5\%$ for CelebA. + +
Data.MethodBackdoor Pattern
PixelPattern
Clean Acc.↑Pois. ASR↓Clean Acc.↑Pois. ASR↓
FMVanilla Net87.4394.3087.1295.22
SpecSign [40]86.2345.6285.9352.31
EdgeNetRob [37]83.480.1282.212.74
AugMax [41]89.8354.6089.2773.14
PixMix [17]89.2592.0488.0194.78
TPS-Deform [43]86.3547.4386.2762.21
SDbOA87.250.0487.070.98
CAVanilla Net98.3097.2098.0097.40
SpecSign [40]98.4564.7897.9854.23
EdgeNetRob [37]92.8010.9093.1012.50
AugMax [41]98.8445.9798.2374.10
PixMix [17]98.1878.3297.9889.28
TPS-Deform [43]97.4371.3197.4774.52
SDbOA95.203.4994.874.21
+ +Table 5. Performances (\%) of vanilla net, Spectral Signature (SpecSign) [40] (NeurIPS'18), EdgeNetRob [37] (ICCV'21), AugMax [41] (NeurIPS'21), PixMix [17] (CVPR'22), TPS-Deform [43] (MM'22) and SDbOA under backdoor attacks. + +Tab. 5 shows that SDbOA achieves the lowest ASR values against backdoor attacks on the two datasets, e.g. it significantly outperforms AugMax [41] that performs best in terms of clean accuracy, by the margins of $54.56\%$ and $72.16\%$ specific to the pixel and pattern attacks on Fashion MNIST. Meanwhile, the clean accuracy is reasonably preserved compared with the most related algorithm, i.e. EdgeNetRob [37]. + +# 4.3.3 Hyperparameter analysis and ablation study + +Analysis of hyperparameter $\lambda$ : To analyze the sensitivity of the learned model's shape-bias against the deformation intensity, i.e. $\lambda$ in Eq. (4), we conduct a hyperparameter analysis and present the results in terms of the two shape-bias metrics [11, 19] in Fig. 6. + +![](images/eee9cd0ebc73bc818425b08d38d5c43bfeaa832cc5aff415f36de7e5ccbd66b6.jpg) +Figure 6. Sensitivity of shape-bias metric values [11, 19] against different deformation intensities $(\lambda)$ . + +Fig. 6 shows that a greater deformation degree does not always result in a more shape-biased model, i.e. the shape-bias of the model learned by SDbOA gradually increases + +
EMSETSDTSGCleanFGSMPGDC&WDeepFool
74.3260.0646.7845.4855.82
✓76.5062.0847.0847.0554.25
✓✓78.2465.2649.8249.6161.35
✓✓87.6715.830.0021.8419.02
✓✓81.0866.8448.4052.3159.46
✓✓✓83.2767.8250.9156.0662.38
+ +Table 6. Ablation study in terms of the clean accuracy and adversarial robustness (%) on CIFAR 10. + +when $\lambda$ is smaller than the threshold of 0.1, while it is decreased when $\lambda$ exceeds 0.1, i.e. the learned model gradually shifts from shape-bias to be more texture-biased. + +To shed light on this observation, we argue that a moderate deformation degree, e.g. $\lambda = 0.1$ in Fig. 4, is beneficial to reasonably improve the diversity of shape representations without largely impairing shape semantic, and a model learned from these valid shape variations can be made better against the perturbation noises. However, as shown in Fig. 4, excessive deformation may be caused when a large $\lambda$ , e.g. $\lambda = 0.2$ is used, which may damage the global shape structure and original shape semantic, causing deformed samples to even cross classification boundaries. When it happens, these augmented samples with damaged shapes may instead encourage the learned network to give prediction based on the relatively stable texture, i.e. making the model to be more texture-biased. + +Ablation study: To study the performance of each module in our SDbOA, we perform the ablation study in Tab. 6, where Robust Canny [37] and Pix2Pix [20] are used as the baseline shape encoding and generator, respectively. + +By comparing the 5th and 7th rows of Tab. 6, it shows that the EMSE module is beneficial to the network robustness, reflecting the superiority of our edge augmentation over direct image deformation. By comparing the 6th and 7th rows, it shows that the proposed TSD enables the model to consistently improve the clean accuracy and the robustness performances under four attacks. By comparing the 4th and 7th rows, it shows that the improved generator with the supervision of texture cues and the shape-preservation loss largely improves the Pix2Pix in terms of the clean accuracy. + +# 5. Conclusions + +Considering the network robustness maybe impaired by the excessive texture-bias, we shed light on how to shift from texture-bias to shape-bias for CNN models, i.e. we propose a novel edge deformation-based data augmentation that enables networks to elaborately learn the shape variations, allowing the well-learned networks to make predictions mainly based on the targets' shapes. Our data augmentation mainly differs from the related SOTAs that it is developed based on the deformation of object's edge map rather than the image, and can reasonably preserve the semantic rationality of the reconstructed samples with the joint supervision of object texture and soft constraint of de + +formed shape cues. Experimental results reveal that our algorithm can largely enhance model's shape-bias in terms of two quantitative metrics, and rather competitive robustness against various perturbations compared with SOTAs. Considering the recognition system trained with the proposed shape augmentation is mainly applicable for pixelwise perturbation, yet may not work for patch-wise corruption where the shape structure is severely damaged, our future work will be devoted to solving it. + +Acknowledgements. The work was supported by Natural Science Foundation of China under grants no. 62276170, 82261138629, the Science and Technology Project of Guangdong Province under grants no. 2023A1515011549, 2023A1515010688, the Science and Technology Innovation Commission of Shenzhen under grants no. JCYJ20190808165203670, JCYJ20220531101412030. + +# References + +[1] Yang Bai, Yuyuan Zeng, Yong Jiang, Shu-Tao Xia, Xingjun Ma, and Yisen Wang. Improving adversarial robustness via channel-wise activation suppressing. In International Conference on Learning Representations, 2020. 3, 7 +[2] Nicholas Baker, Hongjing Lu, Gennady Erlikhman, and Philip J Kellman. Deep convolutional networks do not classify based on global object shape. PLoS computational biology, 14(12):e1006613, 2018. 1 +[3] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE symposium on security and privacy (sp), pages 39-57. IEEE, 2017. 2, 5, 6 +[4] Fabio M Carlucci, Antonio D'Innocente, Silvia Bucci, Barbara Caputo, and Tatiana Tommasi. Domain generalization by solving jigsaw puzzles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2229-2238, 2019. 2 +[5] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 5, 7 +[6] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 113-123, 2019. 7 +[7] Ali Dabouei, Sobhan Soleymani, Jeremy Dawson, and Nasser Nasrabadi. Fast geometrically-perturbed adversarial faces. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1979-1988. IEEE, 2019. 4 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5 +[9] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 7 + +[10] Gianluca Donato and Serge Belongie. Approximate thin plate spline mappings. In European conference on computer vision, pages 21-31. Springer, 2002. 2, 3, 4 +[11] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations, 2018. 1, 2, 5, 6, 8 +[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 4 +[13] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 6 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[15] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018. 1, 5, 7 +[16] Dan Hendrycks, Norman Mu, Ekin D Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781, 2019. 2, 3, 7 +[17] Dan Hendrycks, Andy Zou, Mantas Mazeika, Leonard Tang, Bo Li, Dawn Song, and Jacob Steinhardt. Pixmix: Dreamlike pictures comprehensively improve safety measures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16783-16792, 2022. 2, 7, 8 +[18] Katherine Hermann, Ting Chen, and Simon Kornblith. The origins and prevalence of texture bias in convolutional neural networks. Advances in Neural Information Processing Systems, 33:19000-19015, 2020. 2 +[19] Md Amirul Islam, Matthew Kowal, Patrick Esser, Sen Jia, Bjorn Ommer, Konstantinos G Derpanis, and Neil Bruce. Shape or texture: Understanding discriminative features in cnns. In International Conference on Learning Representations, 2021. 5, 6, 8 +[20] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 4, 6, 8 +[21] Philip TG Jackson, Amir Atapour Abarghouei, Stephen Bonner, Toby P Breckon, and Boguslaw Obara. Style augmentation: data augmentation via style randomization. In CVPR workshops, volume 6, pages 10-11, 2019. 1, 2, 3, 6 +[22] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694-711. Springer, 2016. 4 +[23] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5 + +[24] Sangjun Lee, Inwoo Hwang, Gi-Cheon Kang, and Byoung-Tak Zhang. Improving robustness to texture bias via shape-focused augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 4323-4331, June 2022. 1, 2, 3, 6 +[25] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 5 +[26] Raphael Gontijo Lopes, Dong Yin, Ben Poole, Justin Gilmer, and Ekin D Cubuk. Improving robustness without sacrificing accuracy with patch gaussian augmentation. arXiv preprint arXiv:1906.02611, 2019. 3 +[27] Cheng Luo, Qinliang Lin, Weicheng Xie, Bizhu Wu, Jinheng Xie, and Linlin Shen. Frequency-driven imperceptible adversarial attack on semantic similarity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15315-15324, 2022. 2 +[28] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1, 2, 3, 5, 6, 7 +[29] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2574–2582, 2016. 2, 5, 6 +[30] Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier gans. In International conference on machine learning, pages 2642-2651. PMLR, 2017. 4 +[31] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan A Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Fixing data augmentation to improve adversarial robustness. In Advances in Neural Information Processing Systems, 2021. 3, 7 +[32] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan Andrei Calian, Florian Stimberg, Olivia Wiles, and Timothy A Mann. Data augmentation can improve robustness. Advances in Neural Information Processing Systems, 34:29935-29948, 2021. 2, 3 +[33] Vikash Sehwag, Saeed Mahloujifar, Tinashe Handina, Sihui Dai, Chong Xiang, Mung Chiang, and Prateek Mittal. Robust learning meets generative models: Can proxy distributions improve adversarial robustness? In International Conference on Learning Representations, 2022. 3, 7 +[34] Baifeng Shi, Dinghuai Zhang, Qi Dai, Zhanxing Zhu, Yadong Mu, and Jingdong Wang. Informative dropout for robust representation learning: A shape-bias perspective. In International Conference on Machine Learning, pages 8828-8839. PMLR, 2020. 2, 3 +[35] Connor Shorten and Taghi M Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of big data, 6(1):1-48, 2019. 3 +[36] Gaurang Sriramanan, Sravanti Addepalli, Arya Baburaj, et al. Towards efficient and effective adversarial train + +ing. Advances in Neural Information Processing Systems, 34:11821-11833, 2021. 3, 6, 7 +[37] Mingjie Sun, Zichao Li, Chaowei Xiao, Haonan Qiu, Bhavya Kailkhura, Mingyan Liu, and Bo Li. Can shape structure features improve model robustness under diverse adversarial settings? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7526-7535, 2021. 2, 3, 4, 5, 6, 7, 8 +[38] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1, 2, 5 +[39] Jihoon Tack, Sihyun Yu, Jongheon Jeong, Minseon Kim, Sung Ju Hwang, and Jinwoo Shin. Consistency regularization for adversarial robustness. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 8414-8422, 2022. 3, 7 +[40] Brandon Tran, Jerry Li, and Aleksander Madry. Spectral signatures in backdoor attacks. Advances in neural information processing systems, 31, 2018. 7, 8 +[41] Haotao Wang, Chaowei Xiao, Jean Kossaifi, Zhiding Yu, Anima Anandkumar, and Zhangyang Wang. Augmax: Adversarial composition of random augmentations for robust training. Advances in neural information processing systems, 34:237-250, 2021. 2, 7, 8 +[42] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. 5 +[43] Jun Yu, Guochen Xie, Zhongpeng Cai, Peng He, Fang Gao, and Qiang Ling. Micro expression generation with thin-plate spline motion model and face parsing. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7210–7214, 2022. 1, 2, 3, 7, 8 +[44] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 7 +[45] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In *British Machine Vision Conference* 2016. British Machine Vision Association, 2016. 7 +[46] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 7 +[47] Tianyuan Zhang and Zhanxing Zhu. Interpreting adversarially trained convolutional neural networks. In International Conference on Machine Learning, pages 7502-7511. PMLR, 2019. 2 +[48] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3657–3666, 2022. 3 \ No newline at end of file diff --git a/2023/Shift from Texture-bias to Shape-bias_ Edge Deformation-based Augmentation for Robust Object Recognition/images.zip b/2023/Shift from Texture-bias to Shape-bias_ Edge Deformation-based Augmentation for Robust Object Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b132db9c26a1e7ec0dc044839586711631f3e8ed --- /dev/null +++ b/2023/Shift from Texture-bias to Shape-bias_ Edge Deformation-based Augmentation for Robust Object Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26d1975ebc086afb7ef52c9ea8b381554ed3bf7d322e046e2002077de692d38e +size 653520 diff --git a/2023/Shift from Texture-bias to Shape-bias_ Edge Deformation-based Augmentation for Robust Object Recognition/layout.json b/2023/Shift from Texture-bias to Shape-bias_ Edge Deformation-based Augmentation for Robust Object Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7d9b6d32ef5040d9022f248b8def1df6efea8b3d --- /dev/null +++ b/2023/Shift from Texture-bias to Shape-bias_ Edge Deformation-based Augmentation for Robust Object Recognition/layout.json @@ -0,0 +1,9160 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 54, + 103, + 539, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 103, + 539, + 140 + ], + "spans": [ + { + "bbox": [ + 54, + 103, + 539, + 140 + ], + "type": "text", + "content": "Shift from Texture-bias to Shape-bias: Edge Deformation-based Augmentation for Robust Object Recognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 167, + 589, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 167, + 589, + 182 + ], + "spans": [ + { + "bbox": [ + 52, + 167, + 589, + 182 + ], + "type": "text", + "content": "Xilin He" + }, + { + "bbox": [ + 52, + 167, + 589, + 182 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 52, + 167, + 589, + 182 + ], + "type": "text", + "content": ", Qinliang Lin" + }, + { + "bbox": [ + 52, + 167, + 589, + 182 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 52, + 167, + 589, + 182 + ], + "type": "text", + "content": ", Cheng Luo" + }, + { + "bbox": [ + 52, + 167, + 589, + 182 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 52, + 167, + 589, + 182 + ], + "type": "text", + "content": " Weicheng Xie" + }, + { + "bbox": [ + 52, + 167, + 589, + 182 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 52, + 167, + 589, + 182 + ], + "type": "text", + "content": ", Siyang Song" + }, + { + "bbox": [ + 52, + 167, + 589, + 182 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 52, + 167, + 589, + 182 + ], + "type": "text", + "content": ", Feng Liu" + }, + { + "bbox": [ + 52, + 167, + 589, + 182 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 52, + 167, + 589, + 182 + ], + "type": "text", + "content": ", Linlin Shen" + }, + { + "bbox": [ + 52, + 167, + 589, + 182 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 182, + 572, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 182, + 572, + 197 + ], + "spans": [ + { + "bbox": [ + 71, + 182, + 572, + 197 + ], + "type": "text", + "content": "1Computer Vision Institute, School of Computer Science & Software Engineering, Shenzhen University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 197, + 483, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 197, + 483, + 211 + ], + "spans": [ + { + "bbox": [ + 159, + 197, + 483, + 211 + ], + "type": "text", + "content": "2Shenzhen Institute of Artificial Intelligence & Robotics for Society" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 211, + 482, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 211, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 159, + 211, + 482, + 225 + ], + "type": "text", + "content": "3Guangdong Key Laboratory of Intelligent Information Processing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 262, + 225, + 381, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 225, + 381, + 239 + ], + "spans": [ + { + "bbox": [ + 262, + 225, + 381, + 239 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 262, + 225, + 381, + 239 + ], + "type": "text", + "content": "University of Leicester" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 240, + 483, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 240, + 483, + 252 + ], + "spans": [ + { + "bbox": [ + 159, + 240, + 483, + 252 + ], + "type": "text", + "content": "{2020152115, linqinliang2021, luocheng2020}@email.szu.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 164, + 254, + 478, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 254, + 478, + 266 + ], + "spans": [ + { + "bbox": [ + 164, + 254, + 478, + 266 + ], + "type": "text", + "content": "{wcxie,feng.liu,llshen}@szu.edu.cn,ss1535@leicester.ac.uk" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 293, + 192, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 293, + 192, + 306 + ], + "spans": [ + { + "bbox": [ + 143, + 293, + 192, + 306 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 327, + 289, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 289, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 289, + 650 + ], + "type": "text", + "content": "Recent studies have shown the vulnerability of CNNs under perturbation noises, which is partially caused by the reason that the well-trained CNNs are too biased toward the object texture, i.e., they make predictions mainly based on texture cues. To reduce this texture-bias, current studies resort to learning augmented samples with heavily perturbed texture to make networks be more biased toward relatively stable shape cues. However, such methods usually fail to achieve real shape-biased networks due to the insufficient diversity of the shape cues. In this paper, we propose to augment the training dataset by generating semantically meaningful shapes and samples, via a shape deformation-based online augmentation, namely as SDbOA. The samples generated by our SDbOA have two main merits. First, the augmented samples with more diverse shape variations enable networks to learn the shape cues more elaborately, which encourages the network to be shape-biased. Second, semantic-meaningful shape-augmentation samples could be produced by jointly regularizing the generator with object texture and edge-guidance soft constraint, where the edges are represented more robustly with a self information guided map to better against the noises on them. Extensive experiments under various perturbation noises demonstrate the obvious superiority of our shape-bias-motivated model over the state of the arts in terms of robustness performance. Code is available at https://github.com/C0notSilly/-ICCV-23-Edge-Deformation-based-Online-Augmentation." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 297, + 545, + 413 + ], + "blocks": [ + { + "bbox": [ + 310, + 297, + 545, + 413 + ], + "lines": [ + { + "bbox": [ + 310, + 297, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 310, + 297, + 545, + 413 + ], + "type": "image", + "image_path": "0050a9b4b00509de6a17cdc2e21f40d5e7776e6b9bc49a8c602fc0d4018057ad.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 415, + 547, + 472 + ], + "lines": [ + { + "bbox": [ + 305, + 415, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 547, + 472 + ], + "type": "text", + "content": "Figure 1. Comparison of different models' predictions on texture-shape cue conflict samples [11] between Style Augmentation [21] (CVPRW'19), deformation augmentation in [43] and ours. Specific to the testing image (a), the bottom row shows the original (b) and augmented (c)-(e) images for training." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 483, + 387, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 483, + 387, + 496 + ], + "spans": [ + { + "bbox": [ + 307, + 483, + 387, + 496 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 506, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 602 + ], + "type": "text", + "content": "Previous studies [13, 15, 38] have shown that convolution neural networks (CNNs) are vulnerable to perturbation noises, e.g., adversarial noises. Such perturbation noises imposed on the benign image can easily falsify model's prediction [28]. To improve the network robustness against these noises, algorithms of texture augmentation have been long studied to enable network to learn robust and discriminative object texture features of different categories [11, 24]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 604, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 547, + 700 + ], + "type": "text", + "content": "However, recent studies revealed that the vulnerability of CNNs maybe caused by their 'texture-bias', i.e., the features learned for object recognition bias towards objects' textures instead of their shapes [2, 11]. As illustrated in Fig. 1, when the input image (a) contains a cat that is filled with an elephant's texture, standard CNN networks are more likely to predict it as an elephant despite the obvious shape difference between cats and elephants." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "Consequently, some studies devote to reducing such" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 141, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 141, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 141, + 712 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1526" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "texture-bias to improve network robustness. Zhang et al. [47] found that CNNs learned with adversarial training strategies are less texture-biased. Alternatively, shape-biased CNNs are revealed to be more robust under diverse adversarial attacks. Sun et al. [37] found that the network adversarial robustness can be improved by directly using edge maps to aid the training process. Carlucci et al. [4] proposed to train CNNs on jigsaw puzzles, which achieved superior performances by enabling the network to pay more attention on the cues contained in the global structure. These works illustrate that the shape and structure features, e.g. edge map, can reduce the texture-bias with a stronger shape-bias, i.e. CNNs make predictions based on more shape cues, rather than texture cues." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 241, + 289, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 289, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 289, + 350 + ], + "type": "text", + "content": "However, current studies mainly learned the features of the object shape that is fixed, while the advantage of the shape diversity for devising a more shape-biased model is not well studied. Meanwhile, direct deformation augmentation on images may destroy the object semantics, especially on the image boundary. Given the above observations, following question naturally emerged: Can we implement data augmentation to the edge map to enhance the shape cues and reduce texture-bias?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 350, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 289, + 495 + ], + "type": "text", + "content": "To this end, we propose a thin plate spline (TPS) [10]-based edge map deformation strategy to augment the shape representation of the training data, aiming to enhance the shape-bias of the well-trained network. To make the encoded object shape less sensitive to the noises, we extend the edge map specific to only the object edges to include broader boundary regions via the self-information guided map [34]. Furthermore, to build up the correlation between the deformed shape and the texture category to make the augmented samples more semantic-meaningful, we improve the generator with the additional supervision of denoised texture cues and a shape-preservation loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 496, + 289, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 289, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 289, + 675 + ], + "type": "text", + "content": "Fig. 1 sheds light on the above motivation of our shape augmentation for enhancing networks' shape-bias with a toy experiment, where the probabilities to the shape label of an example sample [11] are predicted by the standard ResNet-50, the Style Augmentation [21], the image deformation augmentation [43] and our method. Fig. 1 shows that the proposed algorithm achieves significantly larger probability predicted to the shape label, i.e. tabby cat over the baseline and the related approaches, which reveals the obvious tendency of the proposed SDbOA in applying shape cues for making predictions. For this observation, we argue that the proposed edge-deformation augmentation enables networks to learn the object shape cues more elaborately, while reducing the contribution of the texture cues for prediction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": "To the best of our knowledge, this work is one of the pioneer attempts to enhance shape-bias of networks based on the augmentation of object's edges. The main contributions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 472, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 472, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 472, + 85 + ], + "type": "text", + "content": "of this paper are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 93, + 547, + 278 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 317, + 93, + 545, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 93, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 93, + 545, + 140 + ], + "type": "text", + "content": "- We propose a TPS-based edge deformation to augment the object shape, and apply it to the online data augmentation, which could help learning more shape-biased networks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 149, + 547, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 149, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 317, + 149, + 547, + 222 + ], + "type": "text", + "content": "- We propose a semantic-meaningful generation paradigm to produce shape-augmentation samples by jointly regularizing the generator with object texture and edge-guidance soft constraint, where a self information guided map is introduced to represent these edge cues more accurately and robustly." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 230, + 547, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 547, + 278 + ], + "type": "text", + "content": "- Extensive results show the superiority of the proposed algorithm over state of the arts (SOTAs) in terms of universal robustness under adversarial, backdoor and common corruption attacks and the training overhead." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 289, + 391, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 289, + 391, + 301 + ], + "spans": [ + { + "bbox": [ + 306, + 289, + 391, + 301 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 308, + 545, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 545, + 334 + ], + "type": "text", + "content": "2.1. Texture or shape-biased models associated with recognition network robustness" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 339, + 547, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 547, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 547, + 471 + ], + "type": "text", + "content": "Adversarial attacks have exposed the vulnerability of networks to perturbation noises [3, 28, 29, 38], leading to a growing concern for network robustness. Recent studies have investigated the association between network robustness and object texture or shape. Geirhos et al. [11] discovered that CNNs trained on ImageNet rely heavily on texture cues, leading to a texture bias in predictions. To address this issue, Hermann et al. [18] proposed additive augmentation techniques such as color distortion or blurring. However, the contribution of shape cues to recognition robustness has not received sufficient attention in their works." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 472, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 555 + ], + "type": "text", + "content": "Recently, Shi et al. [34] proposed an informative dropout approach to help models learn robust representation from a shape-bias perspective. Sun et al. [37] proposed to explicitly use object edges as semantically robust features to against adversarial attacks. Since the object shapes used for enhancing shape-bias are fixed, both approaches are unable to learn a real shape-biased network." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 555, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 547, + 628 + ], + "type": "text", + "content": "In this work, we thus resort to the edge deformation and augmentation to improve the diversity of object shapes. Based on this, we attempt to enhance shape-bias, thus enabling the network to shift its attention from texture to more on the shapes, and better discriminating the samples that have different textures but the similar shape." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 635, + 545, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 545, + 659 + ], + "type": "text", + "content": "2.2. Data augmentation for recognition robustness improvement" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "content": "Data augmentation is a popular paradigm for improving the robustness of recognition network against various perturbation noises [16, 17, 21, 24, 27, 32, 41]. This strategy can be applied to generic computer vision tasks, which differs" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "text", + "content": "1527" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 534, + 218 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 534, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 534, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 534, + 218 + ], + "type": "image", + "image_path": "11a718bfe8ea60788098a431444de5e31a3b9896b2ee1208da66f63e98398238.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 224, + 546, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 224, + 546, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 546, + 258 + ], + "type": "text", + "content": "Figure 2. Pipeline of the proposed SDbOA. For the training, all the three modules, i.e. Edge map-based shape encoding (EMSE), TPS-based shape deformation (TSD) and Texture and shape-based generation (TSG) are used. For the testing, the TSD module is not needed. " + }, + { + "bbox": [ + 46, + 224, + 546, + 258 + ], + "type": "inline_equation", + "content": "\\oslash" + }, + { + "bbox": [ + 46, + 224, + 546, + 258 + ], + "type": "text", + "content": " denotes the 'OR' operation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 267, + 287, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 287, + 304 + ], + "type": "text", + "content": "from some task-specific robustness improvement strategies such as adversarial defense algorithms [1, 28, 31, 33, 36, 39] that are designed for a specific attack scenario." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 304, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 425 + ], + "type": "text", + "content": "Lopes et al. [26] augmented the samples to trade off robustness against recognition accuracy via adding noises on the selected patches of a benign image. Hendrycks et al. [16] proposed the AugMix augmentation, i.e. mixing randomly generated augmentations to improve model robustness under common corruptions. Jackson et al. [21] proposed to augment the data with the style transfer to mitigate domain bias and reduce overfitting. Lee et al. [24] proposed to largely perturb the object texture to help model concentrate on the invariant shape structure." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 521 + ], + "type": "text", + "content": "Most augmentation methods [16, 21, 24, 26, 32] either focus on augmenting texture cues or fail to modify object shape regarding edges or structure, such as geometric transformations [32, 35]. However, compared to image deformation-based augmentation methods [10, 43, 48], which strictly adhere to shape constraints, this hard guidance may disregard the semantic coherence of the deformed sample, as shown in Fig. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 523, + 287, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 523, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 523, + 287, + 583 + ], + "type": "text", + "content": "In this work, we propose an object shape augmentation approach based on the TPS-based [10] edge deformation, with the aid of an extended edge encoding and a generator with a soft constraint loss for the shape representation, to dynamically enrich the object shapes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 596, + 129, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 129, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 129, + 609 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "The algorithm architecture, shown in Fig. 2, comprises three main modules: (i) the edge map-based shape encoding (EMSE) module used for robustly representing object shape; (ii) the TPS-based shape deformation (TSD) module used for providing flexible shape augmentation; and (iii) the texture and shape-based generation (TSG) module used for augmenting the samples dynamically, guided by the deformed shape." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 266, + 522, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 266, + 522, + 280 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 522, + 280 + ], + "type": "text", + "content": "3.1. Edge map-based shape encoding (EMSE)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 285, + 546, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 546, + 393 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 546, + 393 + ], + "type": "text", + "content": "To encode the object shape, the EMSE consists of two blocks. Firstly, a Robust Canny [37] block that generates an edge map for each image, describing the raw shape of the object in the image. While the Robust Canny may disregard some important shape cues of the object, we then additionally encode a self-information guided map [34] to provide complementary shape structure cues in addition to the Canny operator. Specifically, an edge map " + }, + { + "bbox": [ + 304, + 285, + 546, + 393 + ], + "type": "inline_equation", + "content": "\\hat{E}" + }, + { + "bbox": [ + 304, + 285, + 546, + 393 + ], + "type": "text", + "content": " that encodes the pixel-wise self information is first computed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 361, + 401, + 545, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 401, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 361, + 401, + 545, + 429 + ], + "type": "interline_equation", + "content": "\\hat {E} (p) = - \\log \\left(\\sum_ {p ^ {\\prime} \\in \\mathcal {N} _ {p}} e ^ {- \\left(p - p ^ {\\prime}\\right) ^ {2}}\\right) \\tag {1}", + "image_path": "ef3f9368fe9cbb040cdf2cf1c84a13c41b70b076063b298b282651950ee26431.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 439, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 439, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 439, + 545, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 439, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\hat{E}(p)" + }, + { + "bbox": [ + 304, + 439, + 545, + 475 + ], + "type": "text", + "content": " denotes the estimated self-information at the pixel " + }, + { + "bbox": [ + 304, + 439, + 545, + 475 + ], + "type": "inline_equation", + "content": "p = (x, y)" + }, + { + "bbox": [ + 304, + 439, + 545, + 475 + ], + "type": "text", + "content": " of the original image " + }, + { + "bbox": [ + 304, + 439, + 545, + 475 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 439, + 545, + 475 + ], + "type": "text", + "content": "; and " + }, + { + "bbox": [ + 304, + 439, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_p" + }, + { + "bbox": [ + 304, + 439, + 545, + 475 + ], + "type": "text", + "content": " denotes the neighbouring pixels of " + }, + { + "bbox": [ + 304, + 439, + 545, + 475 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 439, + 545, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 475, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 513 + ], + "type": "text", + "content": "Then, we obtain the average pixel-wise self-information " + }, + { + "bbox": [ + 304, + 475, + 545, + 513 + ], + "type": "inline_equation", + "content": "mi" + }, + { + "bbox": [ + 304, + 475, + 545, + 513 + ], + "type": "text", + "content": " and perform a thresholding operation to obtain the self-information guided map " + }, + { + "bbox": [ + 304, + 475, + 545, + 513 + ], + "type": "inline_equation", + "content": "E_{info}" + }, + { + "bbox": [ + 304, + 475, + 545, + 513 + ], + "type": "text", + "content": " as follows" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 364, + 521, + 545, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 521, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 364, + 521, + 545, + 554 + ], + "type": "interline_equation", + "content": "E _ {i n f o} (p) = \\left\\{ \\begin{array}{l l} 1 & \\hat {E} (p) \\geq m i \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}", + "image_path": "82d642274929bfc30394a40cbd597328779113ec12ba5450ba5a8f33d9932a55.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 561, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 561, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 561, + 545, + 658 + ], + "type": "text", + "content": "Eq. (2) indicates that the greater the difference between a pixel and its neighbors, the more shape cues this pixel may contain. That is, for pixels close to the object boundary, their self-information values tend to be relatively large. Consequently, we use this self-information map to extend the edge map extracted by Robust Canny, to leverage the edge-surrounding cues to against the possible noises imposed on the edges, as follows" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 373, + 667, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 667, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 373, + 667, + 545, + 681 + ], + "type": "interline_equation", + "content": "E _ {\\text {e x t e n d}} = E _ {R C} + E _ {\\text {i n f o}} \\tag {3}", + "image_path": "7483ca5ad3d7758bd313a622f3c8356a1b08c8330aa2c280ea724944892f3b7b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "E_{RC}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " denotes the edge maps extracted by the robust Canny." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1528" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 242, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 242, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 242, + 85 + ], + "type": "text", + "content": "3.2. TPS-based shape deformation (TSD)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 90, + 287, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 90, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 47, + 90, + 287, + 126 + ], + "type": "text", + "content": "To augment the shape representation, we propose TPS-based deformation on the meshing of the self-information edge map in Eq. (3)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 126, + 287, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 126, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 126, + 287, + 209 + ], + "type": "text", + "content": "For the shape deformation, Dabouei et al. [7] suggested to use the facial landmarks as control points, while this fashion brings about an overhead of the additional detection module and may damage semantic reasonability. By contrast, we propose to uniformly mesh the edge map, and use the endpoints of the grids as the control points to guide the edge map deformation for shape augmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 210, + 287, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 210, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 210, + 287, + 270 + ], + "type": "text", + "content": "Specifically, the edge map " + }, + { + "bbox": [ + 47, + 210, + 287, + 270 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{extend}}" + }, + { + "bbox": [ + 47, + 210, + 287, + 270 + ], + "type": "text", + "content": " is equally divided into " + }, + { + "bbox": [ + 47, + 210, + 287, + 270 + ], + "type": "inline_equation", + "content": "n = 16" + }, + { + "bbox": [ + 47, + 210, + 287, + 270 + ], + "type": "text", + "content": " grids, and the endpoints are denoted as " + }, + { + "bbox": [ + 47, + 210, + 287, + 270 + ], + "type": "inline_equation", + "content": "\\{P_i, i = 1, \\dots, n\\}" + }, + { + "bbox": [ + 47, + 210, + 287, + 270 + ], + "type": "text", + "content": ", which serve as the original control points and are randomly perturbed to generate target control points " + }, + { + "bbox": [ + 47, + 210, + 287, + 270 + ], + "type": "inline_equation", + "content": "\\{Q_i, 1 \\leq i \\leq n\\}" + }, + { + "bbox": [ + 47, + 210, + 287, + 270 + ], + "type": "text", + "content": " as follows" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 277, + 287, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 277, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 134, + 277, + 287, + 291 + ], + "type": "interline_equation", + "content": "Q _ {i} = P _ {i} + \\lambda \\mathcal {N} \\tag {4}", + "image_path": "bb22e3bab45fd6c5365b84a50332e50c91686c8eb3b952fcba8cb4f3d43e7aef.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 299, + 287, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 299, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 287, + 347 + ], + "type": "text", + "content": "where random noise " + }, + { + "bbox": [ + 46, + 299, + 287, + 347 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 299, + 287, + 347 + ], + "type": "text", + "content": " obeys " + }, + { + "bbox": [ + 46, + 299, + 287, + 347 + ], + "type": "inline_equation", + "content": "N(0,1)" + }, + { + "bbox": [ + 46, + 299, + 287, + 347 + ], + "type": "text", + "content": ", i.e. the normal distribution with mean 0 and variance 1, and the " + }, + { + "bbox": [ + 46, + 299, + 287, + 347 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 299, + 287, + 347 + ], + "type": "text", + "content": " controls the intensity of image deformation, which is set as 0.1. The edge map meshing is visualized in Fig. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "type": "text", + "content": "Given the original and target control points, the edge map deformation is then performed based on TPS. In particular, a transformation " + }, + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "type": "text", + "content": " is found to maximize the smoothness of the edge map deformation while minimizing the gap between two sets of control points. In this work, we use two transformations " + }, + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "type": "inline_equation", + "content": "\\Phi^x, \\Phi^y" + }, + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "type": "text", + "content": " [10], specific to the translation of the " + }, + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "type": "inline_equation", + "content": "x, y" + }, + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "type": "text", + "content": " coordinates of the control points, to obtain the deformed edge map " + }, + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "type": "inline_equation", + "content": "E_{\\text{deform}}" + }, + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 451, + 287, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 287, + 502 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{c} E _ {\\text {d e f o r m}} (p) = \\left(p ^ {x} + \\Phi^ {x} (p), p ^ {y} + \\Phi^ {y} (p)\\right) \\\\ \\Phi^ {x} (p) = M ^ {x} \\cdot p + m _ {0} ^ {x} + \\sum_ {i} w _ {i} ^ {x} U (\\| p - P _ {i} \\| _ {2}) \\\\ M ^ {x}, m _ {0} ^ {x}, \\left\\{w _ {i} ^ {x} \\right\\} = \\\\ \\operatorname {a r g m i n} \\sum_ {i} \\left(\\Phi^ {x} \\left(P _ {i}\\right) - \\left(Q _ {i} ^ {x} - P _ {i} ^ {x}\\right)\\right) ^ {2} + \\mathcal {E} _ {\\Phi^ {x}} \\end{array} \\right. \\tag {5}", + "image_path": "fe2b5e94a0c1ed4e669bc241696196f2bfa8e72fd57ce2e1e181e733696da2f1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 508, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 508, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 508, + 287, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 508, + 287, + 568 + ], + "type": "inline_equation", + "content": "U(\\cdot)" + }, + { + "bbox": [ + 47, + 508, + 287, + 568 + ], + "type": "text", + "content": " denotes the function of radial basis kernel, " + }, + { + "bbox": [ + 47, + 508, + 287, + 568 + ], + "type": "inline_equation", + "content": "M^x\\in \\mathbb{R}^2" + }, + { + "bbox": [ + 47, + 508, + 287, + 568 + ], + "type": "inline_equation", + "content": "m_0^x,w_i^x\\in \\mathbb{R}" + }, + { + "bbox": [ + 47, + 508, + 287, + 568 + ], + "type": "text", + "content": " are the parameters specific to the " + }, + { + "bbox": [ + 47, + 508, + 287, + 568 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 508, + 287, + 568 + ], + "type": "text", + "content": " coordinate. " + }, + { + "bbox": [ + 47, + 508, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\Phi^x}" + }, + { + "bbox": [ + 47, + 508, + 287, + 568 + ], + "type": "text", + "content": " is the second-order smoothness term specific to the transformation " + }, + { + "bbox": [ + 47, + 508, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\Phi^x" + }, + { + "bbox": [ + 47, + 508, + 287, + 568 + ], + "type": "text", + "content": " . For the " + }, + { + "bbox": [ + 47, + 508, + 287, + 568 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 508, + 287, + 568 + ], + "type": "text", + "content": " coordinate, the transformation " + }, + { + "bbox": [ + 47, + 508, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\Phi^y" + }, + { + "bbox": [ + 47, + 508, + 287, + 568 + ], + "type": "text", + "content": " is similarly derived." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 575, + 271, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 271, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 271, + 588 + ], + "type": "text", + "content": "3.3. Texture and shape-based generation (TSG)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "type": "text", + "content": "Since the edge map deformation may cause the texture distorted on the boundary regions, it may deteriorate the fidelity of the deformed shape and the shape-based recognition performance. Thus, we propose a new generative model based on a two-stage training to inpaint the texture of the deformed edge map, i.e. reconstructing this map into the original image domain for the data augmentation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "While the Pix2Pix network [20] can achieve the inpainting from an edge map to the original image representation, it can not well build up the correlation between the edge" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "map and the texture category, i.e. an image with correct shape yet mismatched texture category may be generated. Therefore, we propose to use both shape and texture cues in the generator to build up this correlation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 121, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 545, + 156 + ], + "type": "text", + "content": "Specifically, to provide texture cues for the generator, the original image is smoothed with an easy-to-implement denoising step:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 377, + 159, + 545, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 159, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 377, + 159, + 545, + 172 + ], + "type": "interline_equation", + "content": "I _ {t x t} = \\operatorname {I n t e r p} (D S (I)) \\tag {6}", + "image_path": "ac3820738bd50bdcd3dcebcd65960927a9d26a30476d124d2cb7d03c75ce1d5f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "type": "text", + "content": "where Interp and " + }, + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "type": "inline_equation", + "content": "DS" + }, + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "type": "text", + "content": " denote interpolation and down-sampling methods, both of which are based on bilinear interpolation. " + }, + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "type": "inline_equation", + "content": "I_{txt}" + }, + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "type": "text", + "content": " denotes the resulted texture map. This denoised texture map aims to provide rough guidance cues for the synthesis of fine texture. To further provide the shape cues in the edge encoder of the generator, we use " + }, + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "type": "inline_equation", + "content": "E_{extend}" + }, + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "type": "text", + "content": " in Eq. (3) in the 1st-stage training, and " + }, + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "type": "inline_equation", + "content": "E_{deform}" + }, + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "type": "text", + "content": " in Eq. (5) in the 2nd-stage training." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 277, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 348 + ], + "type": "text", + "content": "To train the generator, a two-stage strategy is used. While a rough image is generated based on " + }, + { + "bbox": [ + 304, + 277, + 545, + 348 + ], + "type": "inline_equation", + "content": "E_{\\text{extend}}" + }, + { + "bbox": [ + 304, + 277, + 545, + 348 + ], + "type": "text", + "content": " and the denoised texture " + }, + { + "bbox": [ + 304, + 277, + 545, + 348 + ], + "type": "inline_equation", + "content": "I_{\\text{txt}}" + }, + { + "bbox": [ + 304, + 277, + 545, + 348 + ], + "type": "text", + "content": " in the first stage, which is fine-tuned with the input of the deformed edge map " + }, + { + "bbox": [ + 304, + 277, + 545, + 348 + ], + "type": "inline_equation", + "content": "E_{\\text{deform}}" + }, + { + "bbox": [ + 304, + 277, + 545, + 348 + ], + "type": "text", + "content": ", and the additional supervision of an object classifier and a newly introduced shape-preservation loss." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 350, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 456 + ], + "type": "text", + "content": "Specifically, the generator is trained in the 1st stage with the loss " + }, + { + "bbox": [ + 304, + 350, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{1st}" + }, + { + "bbox": [ + 304, + 350, + 545, + 456 + ], + "type": "text", + "content": ", i.e. the sum of the adversarial loss " + }, + { + "bbox": [ + 304, + 350, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{gan}" + }, + { + "bbox": [ + 304, + 350, + 545, + 456 + ], + "type": "text", + "content": " [12], a feature matching loss [22] and a auxiliary classification loss referred from ACGAN [30], which is used to generate rough image without shape deformation. During the second stage, we jointly fine-tune the generator and the classifier with the supervision of the cross-entropy classification loss " + }, + { + "bbox": [ + 304, + 350, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 304, + 350, + 545, + 456 + ], + "type": "text", + "content": " [37] in addition to " + }, + { + "bbox": [ + 304, + 350, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{gan}" + }, + { + "bbox": [ + 304, + 350, + 545, + 456 + ], + "type": "text", + "content": ", to better represent the correlation between edge maps and textures." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 458, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 518 + ], + "type": "text", + "content": "In addition, to decrease the sensitivity of the generator against the deformed edge map and stabilize the training, we introduce a " + }, + { + "bbox": [ + 304, + 458, + 545, + 518 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 304, + 458, + 545, + 518 + ], + "type": "text", + "content": "-norm shape-preservation loss between the deformed edge map and that extracted from the generated image as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 353, + 529, + 545, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 529, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 353, + 529, + 545, + 543 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e d g e}} = \\left\\| \\mathrm {e C N N} \\left(I _ {\\text {s y n}}\\right), E _ {\\text {d e f o r m}} \\right\\| _ {1} \\tag {7}", + "image_path": "5f834d4e01aa4ddc5aa8665490cc14a8d6c2d5a19d6df085b07b9d75a2f28057.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 553, + 545, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 553, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 553, + 545, + 649 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 553, + 545, + 649 + ], + "type": "inline_equation", + "content": "I_{syn}" + }, + { + "bbox": [ + 304, + 553, + 545, + 649 + ], + "type": "text", + "content": " and eCNN denotes the synthesized image and convolution network to extract edges in [37]. It's worth noting that the deformed shape " + }, + { + "bbox": [ + 304, + 553, + 545, + 649 + ], + "type": "inline_equation", + "content": "E_{deform}" + }, + { + "bbox": [ + 304, + 553, + 545, + 649 + ], + "type": "text", + "content": " serves as a soft guidance for generating the shape of the augmented sample, it differs from the hard constraint in the direct image deformation, where the object shape is changed strictly according to the target control points. Consequently, the training specific to the loss " + }, + { + "bbox": [ + 304, + 553, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{2nd}" + }, + { + "bbox": [ + 304, + 553, + 545, + 649 + ], + "type": "text", + "content": " in the fine tuning stage is formulated as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 345, + 661, + 545, + 678 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 661, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 345, + 661, + 545, + 678 + ], + "type": "interline_equation", + "content": "\\min _ {G, \\theta} \\max _ {D} \\mathcal {L} _ {2 n d} = \\mathcal {L} _ {g a n} + \\mathcal {L} _ {c l s} + \\mathcal {L} _ {e d g e} \\tag {8}", + "image_path": "335c17b2ba53c0c0ee0fad732eb40babd8b7456ad7f19d47d336a7ad6164a25f.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " denotes the parameters of the classifier. " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " denote the generator and the discriminator." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1529" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 146, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 146, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 146, + 85 + ], + "type": "text", + "content": "3.4. Overall training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 126 + ], + "type": "text", + "content": "For clarity, our sample synthesis with edge deformation is also presented in Alg. 1, which is further used for online shape augmentation in each epoch of the network training." + } + ] + } + ], + "index": 1 + }, + { + "type": "code", + "bbox": [ + 47, + 152, + 253, + 175 + ], + "blocks": [ + { + "bbox": [ + 47, + 137, + 281, + 150 + ], + "lines": [ + { + "bbox": [ + 47, + 137, + 281, + 150 + ], + "spans": [ + { + "bbox": [ + 47, + 137, + 281, + 150 + ], + "type": "text", + "content": "Algorithm 1 Edge deformation-based sample generation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 47, + 152, + 253, + 175 + ], + "lines": [ + { + "bbox": [ + 47, + 152, + 253, + 175 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 253, + 175 + ], + "type": "text", + "content": "Input: input an image " + }, + { + "bbox": [ + 47, + 152, + 253, + 175 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 152, + 253, + 175 + ], + "type": "text", + "content": " and initialize the network Output: the shape-deformed image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "bbox": [ + 53, + 176, + 286, + 258 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 53, + 176, + 242, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 176, + 242, + 187 + ], + "spans": [ + { + "bbox": [ + 53, + 176, + 242, + 187 + ], + "type": "text", + "content": "1: Obtain the extended edge map with Eq. (3);" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 188, + 286, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 188, + 286, + 210 + ], + "spans": [ + { + "bbox": [ + 53, + 188, + 286, + 210 + ], + "type": "text", + "content": "2: Mesh the edge map and perturb the control points with Eq. (4);" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 212, + 246, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 212, + 246, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 212, + 246, + 223 + ], + "type": "text", + "content": "3: Perform edge map deformation with Eq. (5);" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 224, + 250, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 250, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 250, + 235 + ], + "type": "text", + "content": "4: Obtain the denoised texture map with Eq. (6);" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 236, + 286, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 286, + 258 + ], + "type": "text", + "content": "5: Perform the two-stage training for the generator based on " + }, + { + "bbox": [ + 53, + 236, + 286, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{1st}" + }, + { + "bbox": [ + 53, + 236, + 286, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 53, + 236, + 286, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{2nd}" + }, + { + "bbox": [ + 53, + 236, + 286, + 258 + ], + "type": "text", + "content": " in Eq. (8);" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 271, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 287, + 354 + ], + "type": "text", + "content": "Regarding the training complexity, our algorithm requires only the similar runtime cost as the vanilla training, which is significantly lower than that of offline augmentation. This statement, together with the adversarial robustness of the online and offline augmentation for our method are studied with a toy experiment in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 365, + 128, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 128, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 128, + 378 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 384, + 228, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 228, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 228, + 396 + ], + "type": "text", + "content": "4.1. Dataset and experimental settings" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 402, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 450 + ], + "type": "text", + "content": "Datasets We evaluate the robustness performance of the proposed algorithm under five attacks, i.e. [3, 5, 28, 29, 38] on four databases, i.e. Fashion MNIST (FM) [42], CelebA (CA) [25], CIFAR-10 (C10) [23] and ImageNet (IN) [8]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 450, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 486 + ], + "type": "text", + "content": "Experimental setting ResNet50 [14] after the pretraining or ResNet18 is used as the backbone for ImageNet or the other three datasets, unless otherwise specified." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "text", + "content": "We follow the protocol in EdgeNetRob [37] to evaluate the robustness performance under different adversarial attacks. For adversarial robustness evaluation, we use the " + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "text", + "content": "-norm perturbation constrained in the range [0, 1], and the perturbation budgets, i.e. 25/255 for Fashion MNIST, and 8/255 for CIFAR-10, CelebA and ImageNet. CelebA is used for the task of gender recognition." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "content": "Evaluation metrics The performance is evaluated with the shape-bias metrics [11, 19], the clean accuracy and the accuracies under diverse attack scenarios. We follow EdgeNetRob [37] to evaluate the clean accuracy, i.e. the accuracy on the testing set generated by EMSE and TSG modules for SDbOA, and the performance on the original testing set for non-generation algorithms." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": "To evaluate the capacity of our method in terms of clean accuracy preservation compared with state-of-the-art (SOTA) defense strategies and related data augmentation approaches, we use the performance of the baseline as the benchmark." + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 541, + 223 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 541, + 223 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 541, + 223 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 541, + 223 + ], + "type": "image", + "image_path": "21a4d66a410e73915318eeebb660238d2713b9eeda866200db9e51806036b7a5.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 227, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 304, + 227, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 545, + 304 + ], + "type": "text", + "content": "Figure 3. Shape and texture sensitivity of the valilla net and ours against adversarial noises. (a) and (d) show a clean sample and an attacked sample generated by PGD-40 [15], with a texture patch and an edge map extracted by Robust Canny, (b) and (e) show the feature maps (FMs) with the highest " + }, + { + "bbox": [ + 304, + 227, + 545, + 304 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 227, + 545, + 304 + ], + "type": "text", + "content": "-norm activation representing these samples based on ResNet-50, and (c) and (d) show the FMs by ours. 'Dist' denotes the " + }, + { + "bbox": [ + 304, + 227, + 545, + 304 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 227, + 545, + 304 + ], + "type": "text", + "content": "-norm distance." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 310, + 529, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 310, + 529, + 323 + ], + "spans": [ + { + "bbox": [ + 305, + 310, + 529, + 323 + ], + "type": "text", + "content": "4.2. Function Analysis of the proposed modules" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 330, + 497, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 330, + 497, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 330, + 497, + 342 + ], + "type": "text", + "content": "4.2.1 Function of shape bias for robustness" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 350, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 506 + ], + "type": "text", + "content": "Since adversarial noises are generated within a limited adversarial budget and sparsely distributed in perturbed regions, Fig. 3 shows that the object's shape of an image is less perturbed compared with its texture. Specifically, the edge maps reflecting the shape remain similar for the benign and attacked samples, while the feature maps of these samples with the vanilla net are largely different, i.e. their distance is large, reflecting that the yielded feature representation of object texture has already largely perturbed by the attack. By contrast, our algorithm yields significantly smaller feature map bias, suggesting a more shape-biased model can make feature representation more robust against adversarial perturbations." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 525, + 484, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 525, + 484, + 538 + ], + "spans": [ + { + "bbox": [ + 305, + 525, + 484, + 538 + ], + "type": "text", + "content": "4.2.2 Function of the proposed modules" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "The performance of TSD Fig. 4 illustrates the fundamental contribution of the TSD module for shifting from texture-bias to shape-bias. For a moderate magnitude of " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " in Eq. (4), e.g. " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda = 0.1" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ", augmented samples with diverse shapes and the similar texture are generated. To be specific, the constraint loss in Eq. (7) encourages the diversity of the generated shapes, and our GAN generator enhances the semantic rationality of the generated samples and the similarity of their textures. Consequently, these shape-augmented and texture semantic-preserved data could induce the networks to use more neurons to remember the diverse variations of an object shape, and encourage learned networks to make predictions mainly based on the shape cues, i.e. causing the learned networks to be more shape-biased." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1530" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 74, + 286, + 208 + ], + "blocks": [ + { + "bbox": [ + 50, + 74, + 286, + 208 + ], + "lines": [ + { + "bbox": [ + 50, + 74, + 286, + 208 + ], + "spans": [ + { + "bbox": [ + 50, + 74, + 286, + 208 + ], + "type": "image", + "image_path": "1e046d9c632e2bc8be9f9862270136a9cfb6b9d3f6ec9a46901b2d6090610582.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 212, + 272, + 223 + ], + "lines": [ + { + "bbox": [ + 47, + 212, + 272, + 223 + ], + "spans": [ + { + "bbox": [ + 47, + 212, + 272, + 223 + ], + "type": "text", + "content": "Figure 4. The t-SNE visualizations of features with varying " + }, + { + "bbox": [ + 47, + 212, + 272, + 223 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 212, + 272, + 223 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 231, + 286, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 286, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 286, + 302 + ], + "type": "text", + "content": "The performance of EMSE and TSG To shed light on how the introduced self-information edge map and TSG work, we visualize the produced edge maps with the Robust Canny [37] and the proposed self-information in Fig. 5, together with the generated samples with Pix2Pix [20] and our generator." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 303, + 286, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 286, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 286, + 399 + ], + "type": "text", + "content": "The first row of Fig. 5 shows that the proposed self-information guided map can encode richer shape cues and represent shapes more accurately and robustly than those with mere Robust Canny. The second row of Fig. 5 shows that the proposed TSG generates images with well-preserved sharpness of object shape and realistic texture, i.e. better building up the connection between shape and the texture label, compared with Pix2Pix [20]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 406, + 180, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 180, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 180, + 418 + ], + "type": "text", + "content": "4.3.Quantitative evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 424, + 204, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 424, + 204, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 424, + 204, + 436 + ], + "type": "text", + "content": "4.3.1 Results of shape-bias metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 443, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 479 + ], + "type": "text", + "content": "Geirhos et al. [11] proposed the first shape-bias evaluation metric based on the prediction accuracy for a texture-shape cue conflict dataset [11] as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 484, + 287, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 484, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 64, + 484, + 287, + 510 + ], + "type": "interline_equation", + "content": "s b _ {G E} = \\frac {\\# \\text {C o r r e c t e d P r e d i c t i o n t o S h a p e L a b e l}}{\\# \\text {A l l t h e S a m p l e s}} \\tag {9}", + "image_path": "355fb33b24ddd03077794d212329bdcf64b76f793c240375f3e61c2b9642a435.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 517, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 541 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 517, + 287, + 541 + ], + "type": "inline_equation", + "content": "\\#" + }, + { + "bbox": [ + 46, + 517, + 287, + 541 + ], + "type": "text", + "content": " denotes counting the number. This dataset [11] consists of images with shape and texture from different" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 87, + 558, + 248, + 674 + ], + "blocks": [ + { + "bbox": [ + 87, + 558, + 248, + 674 + ], + "lines": [ + { + "bbox": [ + 87, + 558, + 248, + 674 + ], + "spans": [ + { + "bbox": [ + 87, + 558, + 248, + 674 + ], + "type": "image", + "image_path": "63ae311d3f9ab30a3a4efe53a84e807e97cc1bdfb3b825229d6c11b1c0413d19.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 677, + 287, + 721 + ], + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 721 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 721 + ], + "type": "text", + "content": "Figure 5. Visualizations of samples from EMSE and TSG. (a) represents a vanilla sample. (b) and (c) represent edge maps extracted by Robust Canny [37] and EMSE. (d) and (e) represent images generated by Pix2Pix [20] and our TSG." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": "classes to misguide the CNN. Specifically, the object textures of the images [11] are largely interfered with category-irrelevant styles, which places big challenge for current texture-biased networks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 121, + 541, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 121, + 541, + 133 + ], + "spans": [ + { + "bbox": [ + 317, + 121, + 541, + 133 + ], + "type": "text", + "content": "Islam et al. [19] proposed another shape-bias metric as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 141, + 545, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 141, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 313, + 141, + 545, + 172 + ], + "type": "interline_equation", + "content": "s b _ {I S} = \\frac {\\sum_ {i = 1} ^ {n f} \\rho_ {i}}{n f}, \\text {w h e r e} \\rho_ {i} = \\frac {\\operatorname {C o v} \\left(z _ {i} ^ {a} , z _ {i} ^ {b}\\right)}{\\sqrt {\\operatorname {V a r} \\left(z _ {i} ^ {a}\\right) \\operatorname {V a r} \\left(z _ {i} ^ {b}\\right)}} \\tag {10}", + "image_path": "d29a82f731280e90e3595ef895a5781e29fee4fac8d684e3581f93b956faa2ad.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 182, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 289 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 182, + 545, + 289 + ], + "type": "inline_equation", + "content": "nf" + }, + { + "bbox": [ + 304, + 182, + 545, + 289 + ], + "type": "text", + "content": " is the number of feature maps. " + }, + { + "bbox": [ + 304, + 182, + 545, + 289 + ], + "type": "inline_equation", + "content": "z_{i}^{a}" + }, + { + "bbox": [ + 304, + 182, + 545, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 182, + 545, + 289 + ], + "type": "inline_equation", + "content": "z_{i}^{b}" + }, + { + "bbox": [ + 304, + 182, + 545, + 289 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 304, + 182, + 545, + 289 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 182, + 545, + 289 + ], + "type": "text", + "content": "-th flattened feature maps corresponding to the image pair " + }, + { + "bbox": [ + 304, + 182, + 545, + 289 + ], + "type": "inline_equation", + "content": "(a, b)" + }, + { + "bbox": [ + 304, + 182, + 545, + 289 + ], + "type": "text", + "content": " from the texture-shape cue conflict dataset [11], while the two images contain the similar shape but have different textures. The normalized correlation coefficient " + }, + { + "bbox": [ + 304, + 182, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\rho_{i}" + }, + { + "bbox": [ + 304, + 182, + 545, + 289 + ], + "type": "text", + "content": " measures the proportion of feature maps to represent the shape cues. The quantitative results of the standard network, StyleAug [21], ShapeAug [24] and our SDbOA in terms of the two quantitative metrics are shown in Tab. 1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 290, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 350 + ], + "type": "text", + "content": "Tab. 1 shows that our SDbOA consistently achieves the larger shape-bias values compared with the standard network and the related augmentation algorithms [21, 24] in term of both quantitative metrics for the training databases of ImageNet (IN), Stylized IN [11] (SIN) and SIN+IN." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 307, + 360, + 545, + 520 + ], + "blocks": [ + { + "bbox": [ + 307, + 360, + 545, + 520 + ], + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 520 + ], + "type": "table", + "html": "
MethodDatasetsbGE[11]sbIS[19]
StandardIN21.3917.0
StyleAug [21]IN67.3128.5
ShapeAug [24]IN38.4621.3
SDbOAIN71.2831.2
StandardSIN81.3726.2
StyleAug [21]SIN73.7232.7
ShapeAug [24]SIN67.7529.8
SDbOASIN82.4039.9
StandardSIN+IN34.6518.4
StyleAug [21]SIN+IN71.0430.4
ShapeAug [24]SIN+IN63.6723.1
SDbOASIN+IN78.7937.5
", + "image_path": "5d336831a5a5c64fb7d01bea4958181dcb5db401bb2d6e0efa17db046eacab32.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 522, + 545, + 556 + ], + "lines": [ + { + "bbox": [ + 305, + 522, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 305, + 522, + 545, + 556 + ], + "type": "text", + "content": "Table 1. Quantitative results in terms of two shape-bias metrics " + }, + { + "bbox": [ + 305, + 522, + 545, + 556 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 305, + 522, + 545, + 556 + ], + "type": "text", + "content": " with two shape-bias enhancement methods, i.e. StyleAug [21] (CVPRW'19) and ShapeAug [24] (CVPRW'22)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 562, + 515, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 515, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 515, + 574 + ], + "type": "text", + "content": "4.3.2 Robustness against various perturbations" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 582, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 651 + ], + "type": "text", + "content": "Robustness against adversarial attacks: To study the adversarial robustness of our SDbOA compared with the SO-TAs, we present the performances of seven algorithms in terms of clean accuracy and adversarial robustness against FGSM [13], PGD [28], C&W [3] and DeepFool [29] in Tab. 2." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "As shown in Tab. 2, the models trained with SDbOA maintain relatively stronger robustness against various adversarial attacks, than those with the SOTAs. Specifically, SDbOA outperforms NuAT [36] by the margins of " + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "inline_equation", + "content": "14.10\\%" + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "inline_equation", + "content": "27.72\\%" + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": " in terms of the robustness under C&W and" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1531" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "DeepFool on CIFAR-10, and outperforms EdgeNetRob [37] by a margin of " + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "6.98\\%" + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": " under C&W on ImageNet." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 97, + 286, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 97, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 97, + 286, + 168 + ], + "type": "text", + "content": "While Tab. 2 includes generic adversarial training (AT)-based SOTAs for the comparison, we further compare our algorithm with additional two AT-based algorithms [31,33] that are also based on generative model as similar as ours in Tab. 3, where the same protocol (evaluation with AutoAttack [5]) and network architectures as [31,33] are employed." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 170, + 286, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 286, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 286, + 266 + ], + "type": "text", + "content": "Since the works [31,33] require adversarial training, tens of millions of adversarial samples in addition to the vanilla images are generated for the training, resulting in a significantly larger runtime complexity. Despite this, Tab. 3 shows that our online augmentation model can still outperform the listed adversarial-training-based methods in terms of the clean accuracy and adversarial robustness with reasonable training overhead." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 53, + 277, + 279, + 640 + ], + "blocks": [ + { + "bbox": [ + 53, + 277, + 279, + 640 + ], + "lines": [ + { + "bbox": [ + 53, + 277, + 279, + 640 + ], + "spans": [ + { + "bbox": [ + 53, + 277, + 279, + 640 + ], + "type": "table", + "html": "
Data.MethodClean Acc.FGSM PGD CWDeep Fool
FMVanilla Net92.4227.420.5810.0039.5
PGD [28]86.5878.8481.2617.540.56
AT+CAS [1]87.4179.2378.4232.9531.46
NuAT [36]88.0281.4382.3767.0742.19
AT+CR [39]87.4680.8281.9534.5737.04
EdgeNetRob [37]85.0178.3075.7582.4771.43
SDbOA88.5280.3283.2687.2476.12
CAVanilla Net99.1848.860.0023.5720.80
PGD [28]92.8585.4083.3542.7819.84
AT+CAS [1]93.0689.8887.6865.3637.05
NuAT [36]94.2693.8089.5267.2529.42
AT+CR [39]95.3294.5890.7464.3839.06
EdgeNetRob [37]95.5292.6455.0875.9283.72
SDbOA98.7495.3488.2879.8491.50
C10Vanilla Net90.707.840.0016.7814.30
PGD [28]75.8254.5644.6618.020.38
AT+CAS [1]76.5162.3247.8151.4337.42
NuAT [36]79.5162.1246.4541.9634.66
AT+CR [39]82.8064.8153.3950.6643.19
EdgeNetRob [37]79.2162.4839.6643.4654.45
SDbOA83.2767.8250.9156.0662.38
INVanilla Net75.825.400.00-5.44
PGD [28]55.3440.4630.8316.0418.36
AT+CAS [1]57.0347.5836.7223.6727.92
NuAT [36]58.3848.1240.3021.7125.40
AT+CR [39]52.0341.7227.9321.4021.07
EdgeNetRob [37]65.7345.8733.6220.8523.40
SDbOA68.5649.1035.7627.8328.89
", + "image_path": "8d69625fcb2d540d8dcf11a55b67b482a9cba4d79bb9f5a4872ce40b664f3551.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 701, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 714 + ], + "type": "text", + "content": "Robustness of augmentation approaches against" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 550, + 139 + ], + "blocks": [ + { + "bbox": [ + 47, + 639, + 286, + 693 + ], + "lines": [ + { + "bbox": [ + 47, + 639, + 286, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 286, + 693 + ], + "type": "text", + "content": "Table 2. Clean accuracy and adversarial robustness (%) with vanilla net, PGD Training [28], AT+CAS [1] (ICLR'21), NuAT [36] (NeurIPS'21), AT+CR [39] (AAAI'22), EdgeNetRob [37] (ICCV'21) and ours. The best and 2nd best performances are labeled with bold and underline." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 70, + 550, + 139 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 550, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 550, + 139 + ], + "type": "table", + "html": "
MethodCleanRobustRuntimeAT
FDA28 [31]85.9760.73Days✓
SDbOA2885.2768.153-5 Hours✗
PORT34 [33]87.0060.60Days✓
SDbOA3487.1273.363-5 Hours✗
", + "image_path": "9df12a63ac885ce226588b554651e69c67e420fb64c8cf1d18aeeef20c8b2692.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 220, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 220, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 220, + 545, + 304 + ], + "type": "text", + "content": "common corruptions and adversarial attacks: To evaluate the performances of our data augmentation and related augmentation techniques under common corruptions [15] and adversarial attacks, we present the results of nine SO-TAs in term of mean corruption error (mCE) and adversarial robustness in Tab. 4, where mCE is evaluated under all 15 corruptions and each corruption has 5 severities [15]." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 315, + 545, + 450 + ], + "blocks": [ + { + "bbox": [ + 305, + 142, + 545, + 219 + ], + "lines": [ + { + "bbox": [ + 305, + 142, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 305, + 142, + 545, + 219 + ], + "type": "text", + "content": "Table 3. Adversarial robustness (\\%) of the generative-model-based adversarial training approaches, i.e. FDA [31] (NeurIPS'21) and PORT [33] (ICLR'22) on CIFAR-10. We follow the robustness metric in [31] measured by AutoAttack [5]. AT represents 'adversarial training'. '28' or '34' stands for using WideResNet-28-10 or WideResNet-34-10 [45] as the classifier network, following [31, 33]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 315, + 545, + 450 + ], + "lines": [ + { + "bbox": [ + 307, + 315, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 307, + 315, + 545, + 450 + ], + "type": "table", + "html": "
MethodsCorruptions↓Adversaries↓
Baseline26.491.3
Cutout [9]arXiv25.996.0
Mixup [46]ICLR'1821.093.3
CutMix [44]ICCV'1926.592.1
AutoAugment [6]CVPR'1922.295.1
AugMix [16]ICLR'1912.486.8
AugMax [41]NeurIPS'2119.939.1
PixMix [17]CVPR'229.582.1
TPS-Deform [43]MM'2222.444.0
SDbOA18.827.7
", + "image_path": "2979cb79af8cb3198f51c1c1d8907f5fe2a9b35a95de4efbf5ff5c4e06ecdd26.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 453, + 545, + 509 + ], + "lines": [ + { + "bbox": [ + 305, + 453, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 305, + 453, + 545, + 509 + ], + "type": "text", + "content": "Table 4. Robustness performances (\\%) of nine data augmentation techniques on CIFAR-10-C, i.e. the corruption set of CIFAR-10, based on the baseline of WideResNet-40-4 [45] under PGD-20 attack. The adversaries metric is the misclassification rate under PGD-20 with the " + }, + { + "bbox": [ + 305, + 453, + 545, + 509 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 305, + 453, + 545, + 509 + ], + "type": "text", + "content": " budget of 2/255." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "text", + "content": "Tab. 4 shows that large majority of SOTA data augmentation techniques improve robustness against corruptions without paying enough attention on the adversarial robustness. It's worth noting that our algorithm achieves much better adversarial robustness, i.e. " + }, + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "inline_equation", + "content": "27.7\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "text", + "content": " with large margins over those, i.e. " + }, + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "inline_equation", + "content": "82.1\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "inline_equation", + "content": "86.8\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "text", + "content": " achieved by PixMix [17] and AugMax [41] that perform better than ours in terms of mCE. This maybe because that these two augmentations are mainly focused on robustness against a specified type of samples, while ours is developed for universal robustness." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": "Robustness against backdoor attacks: To study the robustness of the proposed algorithm under backdoor attacks [40] compared with the related data augmentation methods, we follow the protocol in [37], and present the clean accuracy on the standard testing data and the attack success rate on the poisoned data (Pois. ASR) of seven algorithms, un" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1532" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "der backdoor attacks with two patterns in Tab. 5. Poisoning ratio is set as " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " for Fashion MNIST and " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " for CelebA." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 105, + 285, + 289 + ], + "blocks": [ + { + "bbox": [ + 50, + 105, + 285, + 289 + ], + "lines": [ + { + "bbox": [ + 50, + 105, + 285, + 289 + ], + "spans": [ + { + "bbox": [ + 50, + 105, + 285, + 289 + ], + "type": "table", + "html": "
Data.MethodBackdoor Pattern
PixelPattern
Clean Acc.↑Pois. ASR↓Clean Acc.↑Pois. ASR↓
FMVanilla Net87.4394.3087.1295.22
SpecSign [40]86.2345.6285.9352.31
EdgeNetRob [37]83.480.1282.212.74
AugMax [41]89.8354.6089.2773.14
PixMix [17]89.2592.0488.0194.78
TPS-Deform [43]86.3547.4386.2762.21
SDbOA87.250.0487.070.98
CAVanilla Net98.3097.2098.0097.40
SpecSign [40]98.4564.7897.9854.23
EdgeNetRob [37]92.8010.9093.1012.50
AugMax [41]98.8445.9798.2374.10
PixMix [17]98.1878.3297.9889.28
TPS-Deform [43]97.4371.3197.4774.52
SDbOA95.203.4994.874.21
", + "image_path": "5d2c9a74134b833378c49642186cfdb3c3fb322016b858f413b744f5ec929f35.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 289, + 287, + 332 + ], + "lines": [ + { + "bbox": [ + 47, + 289, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 287, + 332 + ], + "type": "text", + "content": "Table 5. Performances (\\%) of vanilla net, Spectral Signature (SpecSign) [40] (NeurIPS'18), EdgeNetRob [37] (ICCV'21), AugMax [41] (NeurIPS'21), PixMix [17] (CVPR'22), TPS-Deform [43] (MM'22) and SDbOA under backdoor attacks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 336, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 287, + 432 + ], + "type": "text", + "content": "Tab. 5 shows that SDbOA achieves the lowest ASR values against backdoor attacks on the two datasets, e.g. it significantly outperforms AugMax [41] that performs best in terms of clean accuracy, by the margins of " + }, + { + "bbox": [ + 46, + 336, + 287, + 432 + ], + "type": "inline_equation", + "content": "54.56\\%" + }, + { + "bbox": [ + 46, + 336, + 287, + 432 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 336, + 287, + 432 + ], + "type": "inline_equation", + "content": "72.16\\%" + }, + { + "bbox": [ + 46, + 336, + 287, + 432 + ], + "type": "text", + "content": " specific to the pixel and pattern attacks on Fashion MNIST. Meanwhile, the clean accuracy is reasonably preserved compared with the most related algorithm, i.e. EdgeNetRob [37]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 448, + 270, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 448, + 270, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 448, + 270, + 460 + ], + "type": "text", + "content": "4.3.3 Hyperparameter analysis and ablation study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 467, + 287, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 287, + 527 + ], + "type": "text", + "content": "Analysis of hyperparameter " + }, + { + "bbox": [ + 46, + 467, + 287, + 527 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 467, + 287, + 527 + ], + "type": "text", + "content": ": To analyze the sensitivity of the learned model's shape-bias against the deformation intensity, i.e. " + }, + { + "bbox": [ + 46, + 467, + 287, + 527 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 467, + 287, + 527 + ], + "type": "text", + "content": " in Eq. (4), we conduct a hyperparameter analysis and present the results in terms of the two shape-bias metrics [11, 19] in Fig. 6." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 67, + 540, + 266, + 648 + ], + "blocks": [ + { + "bbox": [ + 67, + 540, + 266, + 648 + ], + "lines": [ + { + "bbox": [ + 67, + 540, + 266, + 648 + ], + "spans": [ + { + "bbox": [ + 67, + 540, + 266, + 648 + ], + "type": "image", + "image_path": "eee9cd0ebc73bc818425b08d38d5c43bfeaa832cc5aff415f36de7e5ccbd66b6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 653, + 287, + 676 + ], + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 676 + ], + "type": "text", + "content": "Figure 6. Sensitivity of shape-bias metric values [11, 19] against different deformation intensities " + }, + { + "bbox": [ + 47, + 653, + 287, + 676 + ], + "type": "inline_equation", + "content": "(\\lambda)" + }, + { + "bbox": [ + 47, + 653, + 287, + 676 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "Fig. 6 shows that a greater deformation degree does not always result in a more shape-biased model, i.e. the shape-bias of the model learned by SDbOA gradually increases" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 60, + 538, + 148 + ], + "blocks": [ + { + "bbox": [ + 307, + 60, + 538, + 148 + ], + "lines": [ + { + "bbox": [ + 307, + 60, + 538, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 60, + 538, + 148 + ], + "type": "table", + "html": "
EMSETSDTSGCleanFGSMPGDC&WDeepFool
74.3260.0646.7845.4855.82
✓76.5062.0847.0847.0554.25
✓✓78.2465.2649.8249.6161.35
✓✓87.6715.830.0021.8419.02
✓✓81.0866.8448.4052.3159.46
✓✓✓83.2767.8250.9156.0662.38
", + "image_path": "11db19e4450bd264704362e0f8c4df45c148b44a6bcc34bf4f9eaedd00b3ba6c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 148, + 544, + 168 + ], + "lines": [ + { + "bbox": [ + 306, + 148, + 544, + 168 + ], + "spans": [ + { + "bbox": [ + 306, + 148, + 544, + 168 + ], + "type": "text", + "content": "Table 6. Ablation study in terms of the clean accuracy and adversarial robustness (%) on CIFAR 10." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 171, + 545, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 171, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 306, + 171, + 545, + 206 + ], + "type": "text", + "content": "when " + }, + { + "bbox": [ + 306, + 171, + 545, + 206 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 306, + 171, + 545, + 206 + ], + "type": "text", + "content": " is smaller than the threshold of 0.1, while it is decreased when " + }, + { + "bbox": [ + 306, + 171, + 545, + 206 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 306, + 171, + 545, + 206 + ], + "type": "text", + "content": " exceeds 0.1, i.e. the learned model gradually shifts from shape-bias to be more texture-biased." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 208, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 208, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 208, + 545, + 374 + ], + "type": "text", + "content": "To shed light on this observation, we argue that a moderate deformation degree, e.g. " + }, + { + "bbox": [ + 304, + 208, + 545, + 374 + ], + "type": "inline_equation", + "content": "\\lambda = 0.1" + }, + { + "bbox": [ + 304, + 208, + 545, + 374 + ], + "type": "text", + "content": " in Fig. 4, is beneficial to reasonably improve the diversity of shape representations without largely impairing shape semantic, and a model learned from these valid shape variations can be made better against the perturbation noises. However, as shown in Fig. 4, excessive deformation may be caused when a large " + }, + { + "bbox": [ + 304, + 208, + 545, + 374 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 208, + 545, + 374 + ], + "type": "text", + "content": ", e.g. " + }, + { + "bbox": [ + 304, + 208, + 545, + 374 + ], + "type": "inline_equation", + "content": "\\lambda = 0.2" + }, + { + "bbox": [ + 304, + 208, + 545, + 374 + ], + "type": "text", + "content": " is used, which may damage the global shape structure and original shape semantic, causing deformed samples to even cross classification boundaries. When it happens, these augmented samples with damaged shapes may instead encourage the learned network to give prediction based on the relatively stable texture, i.e. making the model to be more texture-biased." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 376, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 424 + ], + "type": "text", + "content": "Ablation study: To study the performance of each module in our SDbOA, we perform the ablation study in Tab. 6, where Robust Canny [37] and Pix2Pix [20] are used as the baseline shape encoding and generator, respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 426, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 545 + ], + "type": "text", + "content": "By comparing the 5th and 7th rows of Tab. 6, it shows that the EMSE module is beneficial to the network robustness, reflecting the superiority of our edge augmentation over direct image deformation. By comparing the 6th and 7th rows, it shows that the proposed TSD enables the model to consistently improve the clean accuracy and the robustness performances under four attacks. By comparing the 4th and 7th rows, it shows that the improved generator with the supervision of texture cues and the shape-preservation loss largely improves the Pix2Pix in terms of the clean accuracy." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 548, + 383, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 383, + 560 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 383, + 560 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Considering the network robustness maybe impaired by the excessive texture-bias, we shed light on how to shift from texture-bias to shape-bias for CNN models, i.e. we propose a novel edge deformation-based data augmentation that enables networks to elaborately learn the shape variations, allowing the well-learned networks to make predictions mainly based on the targets' shapes. Our data augmentation mainly differs from the related SOTAs that it is developed based on the deformation of object's edge map rather than the image, and can reasonably preserve the semantic rationality of the reconstructed samples with the joint supervision of object texture and soft constraint of de" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1533" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "content": "formed shape cues. Experimental results reveal that our algorithm can largely enhance model's shape-bias in terms of two quantitative metrics, and rather competitive robustness against various perturbations compared with SOTAs. Considering the recognition system trained with the proposed shape augmentation is mainly applicable for pixelwise perturbation, yet may not work for patch-wise corruption where the shape structure is severely damaged, our future work will be devoted to solving it." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 289, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 289, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 289, + 276 + ], + "type": "text", + "content": "Acknowledgements. The work was supported by Natural Science Foundation of China under grants no. 62276170, 82261138629, the Science and Technology Project of Guangdong Province under grants no. 2023A1515011549, 2023A1515010688, the Science and Technology Innovation Commission of Shenzhen under grants no. JCYJ20190808165203670, JCYJ20220531101412030." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 287, + 107, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 107, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 107, + 300 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 308, + 289, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 308, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 308, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 53, + 308, + 287, + 352 + ], + "type": "text", + "content": "[1] Yang Bai, Yuyuan Zeng, Yong Jiang, Shu-Tao Xia, Xingjun Ma, and Yisen Wang. Improving adversarial robustness via channel-wise activation suppressing. In International Conference on Learning Representations, 2020. 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 353, + 288, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 353, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 53, + 353, + 288, + 397 + ], + "type": "text", + "content": "[2] Nicholas Baker, Hongjing Lu, Gennady Erlikhman, and Philip J Kellman. Deep convolutional networks do not classify based on global object shape. PLoS computational biology, 14(12):e1006613, 2018. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 399, + 289, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 289, + 431 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 289, + 431 + ], + "type": "text", + "content": "[3] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE symposium on security and privacy (sp), pages 39-57. IEEE, 2017. 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 433, + 289, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 433, + 289, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 433, + 289, + 487 + ], + "type": "text", + "content": "[4] Fabio M Carlucci, Antonio D'Innocente, Silvia Bucci, Barbara Caputo, and Tatiana Tommasi. Domain generalization by solving jigsaw puzzles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2229-2238, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 489, + 289, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 489, + 289, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 489, + 289, + 533 + ], + "type": "text", + "content": "[5] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 534, + 289, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 534, + 289, + 588 + ], + "spans": [ + { + "bbox": [ + 53, + 534, + 289, + 588 + ], + "type": "text", + "content": "[6] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 113-123, 2019. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 590, + 289, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 590, + 289, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 590, + 289, + 635 + ], + "type": "text", + "content": "[7] Ali Dabouei, Sobhan Soleymani, Jeremy Dawson, and Nasser Nasrabadi. Fast geometrically-perturbed adversarial faces. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1979-1988. IEEE, 2019. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 635, + 289, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 289, + 679 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 289, + 679 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 681, + 289, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 681, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 681, + 289, + 713 + ], + "type": "text", + "content": "[9] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[10] Gianluca Donato and Serge Belongie. Approximate thin plate spline mappings. In European conference on computer vision, pages 21-31. Springer, 2002. 2, 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 107, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 171 + ], + "type": "text", + "content": "[11] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations, 2018. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 173, + 547, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 547, + 216 + ], + "type": "text", + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 217, + 547, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 547, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 547, + 249 + ], + "type": "text", + "content": "[13] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 250, + 547, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 547, + 294 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 547, + 294 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 294, + 547, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 547, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 547, + 337 + ], + "type": "text", + "content": "[15] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018. 1, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 338, + 547, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 547, + 382 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 547, + 382 + ], + "type": "text", + "content": "[16] Dan Hendrycks, Norman Mu, Ekin D Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781, 2019. 2, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 383, + 547, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 547, + 447 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 547, + 447 + ], + "type": "text", + "content": "[17] Dan Hendrycks, Andy Zou, Mantas Mazeika, Leonard Tang, Bo Li, Dawn Song, and Jacob Steinhardt. Pixmix: Dreamlike pictures comprehensively improve safety measures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16783-16792, 2022. 2, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 449, + 547, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 547, + 491 + ], + "type": "text", + "content": "[18] Katherine Hermann, Ting Chen, and Simon Kornblith. The origins and prevalence of texture bias in convolutional neural networks. Advances in Neural Information Processing Systems, 33:19000-19015, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 492, + 547, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 547, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 547, + 547 + ], + "type": "text", + "content": "[19] Md Amirul Islam, Matthew Kowal, Patrick Esser, Sen Jia, Bjorn Ommer, Konstantinos G Derpanis, and Neil Bruce. Shape or texture: Understanding discriminative features in cnns. In International Conference on Learning Representations, 2021. 5, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 548, + 547, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 547, + 602 + ], + "type": "text", + "content": "[20] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 4, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 603, + 547, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 547, + 647 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 547, + 647 + ], + "type": "text", + "content": "[21] Philip TG Jackson, Amir Atapour Abarghouei, Stephen Bonner, Toby P Breckon, and Boguslaw Obara. Style augmentation: data augmentation via style randomization. In CVPR workshops, volume 6, pages 10-11, 2019. 1, 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 647, + 547, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 547, + 690 + ], + "type": "text", + "content": "[22] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694-711. Springer, 2016. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "type": "text", + "content": "[23] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1534" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "text", + "content": "[24] Sangjun Lee, Inwoo Hwang, Gi-Cheon Kang, and Byoung-Tak Zhang. Improving robustness to texture bias via shape-focused augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 4323-4331, June 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 139, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 139, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 49, + 139, + 288, + 183 + ], + "type": "text", + "content": "[25] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "type": "text", + "content": "[26] Raphael Gontijo Lopes, Dong Yin, Ben Poole, Justin Gilmer, and Ekin D Cubuk. Improving robustness without sacrificing accuracy with patch gaussian augmentation. arXiv preprint arXiv:1906.02611, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 282 + ], + "type": "text", + "content": "[27] Cheng Luo, Qinliang Lin, Weicheng Xie, Bizhu Wu, Jinheng Xie, and Linlin Shen. Frequency-driven imperceptible adversarial attack on semantic similarity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15315-15324, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 283, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 335 + ], + "type": "text", + "content": "[28] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 337, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 337, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 287, + 392 + ], + "type": "text", + "content": "[29] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2574–2582, 2016. 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "text", + "content": "[30] Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier gans. In International conference on machine learning, pages 2642-2651. PMLR, 2017. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "text", + "content": "[31] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan A Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Fixing data augmentation to improve adversarial robustness. In Advances in Neural Information Processing Systems, 2021. 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 492, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 546 + ], + "type": "text", + "content": "[32] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan Andrei Calian, Florian Stimberg, Olivia Wiles, and Timothy A Mann. Data augmentation can improve robustness. Advances in Neural Information Processing Systems, 34:29935-29948, 2021. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 548, + 287, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 603 + ], + "type": "text", + "content": "[33] Vikash Sehwag, Saeed Mahloujifar, Tinashe Handina, Sihui Dai, Chong Xiang, Mung Chiang, and Prateek Mittal. Robust learning meets generative models: Can proxy distributions improve adversarial robustness? In International Conference on Learning Representations, 2022. 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "type": "text", + "content": "[34] Baifeng Shi, Dinghuai Zhang, Qi Dai, Zhanxing Zhu, Yadong Mu, and Jingdong Wang. Informative dropout for robust representation learning: A shape-bias perspective. In International Conference on Machine Learning, pages 8828-8839. PMLR, 2020. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "type": "text", + "content": "[35] Connor Shorten and Taghi M Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of big data, 6(1):1-48, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "text", + "content": "[36] Gaurang Sriramanan, Sravanti Addepalli, Arya Baburaj, et al. Towards efficient and effective adversarial train" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 677 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "ing. Advances in Neural Information Processing Systems, 34:11821-11833, 2021. 3, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 162 + ], + "type": "text", + "content": "[37] Mingjie Sun, Zichao Li, Chaowei Xiao, Haonan Qiu, Bhavya Kailkhura, Mingyan Liu, and Bo Li. Can shape structure features improve model robustness under diverse adversarial settings? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7526-7535, 2021. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[38] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1, 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 207, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 262 + ], + "type": "text", + "content": "[39] Jihoon Tack, Sihyun Yu, Jongheon Jeong, Minseon Kim, Sung Ju Hwang, and Jinwoo Shin. Consistency regularization for adversarial robustness. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 8414-8422, 2022. 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 263, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 263, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 545, + 297 + ], + "type": "text", + "content": "[40] Brandon Tran, Jerry Li, and Aleksander Madry. Spectral signatures in backdoor attacks. Advances in neural information processing systems, 31, 2018. 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "type": "text", + "content": "[41] Haotao Wang, Chaowei Xiao, Jean Kossaifi, Zhiding Yu, Anima Anandkumar, and Zhangyang Wang. Augmax: Adversarial composition of random augmentations for robust training. Advances in neural information processing systems, 34:237-250, 2021. 2, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "type": "text", + "content": "[42] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 397, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 397, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 397, + 545, + 453 + ], + "type": "text", + "content": "[43] Jun Yu, Guochen Xie, Zhongpeng Cai, Peng He, Fang Gao, and Qiang Ling. Micro expression generation with thin-plate spline motion model and face parsing. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7210–7214, 2022. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 454, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 454, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 454, + 545, + 509 + ], + "type": "text", + "content": "[44] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 510, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 542 + ], + "type": "text", + "content": "[45] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In *British Machine Vision Conference* 2016. British Machine Vision Association, 2016. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 543, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 543, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 543, + 545, + 586 + ], + "type": "text", + "content": "[46] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 587, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 587, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 587, + 545, + 632 + ], + "type": "text", + "content": "[47] Tianyuan Zhang and Zhanxing Zhu. Interpreting adversarially trained convolutional neural networks. In International Conference on Machine Learning, pages 7502-7511. PMLR, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 633, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 633, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 633, + 545, + 677 + ], + "type": "text", + "content": "[48] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3657–3666, 2022. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1535" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/ShiftNAS_ Improving One-shot NAS via Probability Shift/9d0cd6fb-8343-4e85-b19e-ff450eb2e5e4_content_list.json b/2023/ShiftNAS_ Improving One-shot NAS via Probability Shift/9d0cd6fb-8343-4e85-b19e-ff450eb2e5e4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ea1e92ee5f75dd3006128dbdfc349d07b3217730 --- /dev/null +++ b/2023/ShiftNAS_ Improving One-shot NAS via Probability Shift/9d0cd6fb-8343-4e85-b19e-ff450eb2e5e4_content_list.json @@ -0,0 +1,1580 @@ +[ + { + "type": "text", + "text": "ShiftNAS: Improving One-shot NAS via Probability Shift", + "text_level": 1, + "bbox": [ + 194, + 130, + 774, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingyang Zhang, Xinyi Yu* Haodong Zhao, Linlin Ou* Zhejiang University of Technology Hangzhou, Zhejiang, China", + "bbox": [ + 264, + 180, + 704, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yuxy@zjut.edu.cn, linlinou@zjut.edu.cn", + "bbox": [ + 313, + 237, + 653, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One-shot Neural architecture search (One-shot NAS) has been proposed as a time-efficient approach to obtain optimal subnet architectures and weights under different complexity cases by training only once. However, the subnet performance obtained by weight sharing is often inferior to the performance achieved by retraining. In this paper, we investigate the performance gap and attribute it to the use of uniform sampling, which is a common approach in supernet training. Uniform sampling concentrates training resources on subnets with intermediate computational resources, which are sampled with high probability. However, subnets with different complexity regions require different optimal training strategies for optimal performance.", + "bbox": [ + 75, + 318, + 472, + 515 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the problem of uniform sampling, we propose ShiftNAS, a method that can adjust the sampling probability based on the complexity of subnets. We achieve this by evaluating the performance variation of subnets with different complexity and designing an architecture generator that can accurately and efficiently provide subnets with the desired complexity. Both the sampling probability and the architecture generator can be trained end-to-end in a gradient-based manner. With ShiftNAS, we can directly obtain the optimal model architecture and parameters for a given computational complexity. We evaluate our approach on multiple visual network models, including convolutional neural networks (CNNs) and vision transformers (ViTs), and demonstrate that ShiftNAS is model-agnostic. Experimental results on ImageNet show that ShiftNAS can improve the performance of one-shot NAS without additional consumption. Source codes are available at GitHub.", + "bbox": [ + 75, + 515, + 472, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 803, + 209, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) have been widely applied to the field of computer vision with remarkable success [6, 8]. However, the deployment of these vision net", + "bbox": [ + 76, + 829, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9e326b61809dda24f08e4295dbc212bcc6ebd11bb228479e249d6a1220227467.jpg", + "image_caption": [ + "Figure 1: A conceptual overview of our ShiftNAS. Focusing on the computational resource (e.g. FLOPs), ShiftNAS split a search space into several parts where subnets have close computational complexity. In stochastic supernet training, ShiftNAS first samples computational complexity according to probability, then an architecture generator samples the subnet with desired computational complexity. The probability distribution is dynamically updated with stochastic training by estimating the performance variation of each sub-space." + ], + "image_footnote": [], + "bbox": [ + 506, + 287, + 890, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "works on edge devices still has some limitations, such as the massive model sizes and excessive computation overhead [9, 10, 20, 24]. In addition, designing architectures artificially in a trial-and-error manner is a resource-consuming task that requires not only architectural skills but also domain expertise. Consequently, how to acquire optimal architectures that balance latency and accuracy efficiently is of paramount importance.", + "bbox": [ + 496, + 657, + 893, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in neural architecture search (NAS) methods, such as [22, 7, 30, 35], have led to significant improvements in the performance of practical applications by automatically searching for optimal architectures within a defined search space. However, traditional NAS methods typically require a substantial computation budget [23, 35]. In order to speed up training and reduce the resource consumption of the training process, one-shot NAS", + "bbox": [ + 496, + 779, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Xinyi Yu and Linlin Ou are corresponding authors.", + "bbox": [ + 94, + 886, + 370, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5919", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "methods [18, 7] have adopted a two-stage training approach based on weight sharing. Specifically, a supernet is trained in the first stage, and subnets with better performance are searched for in the second stage. In some cases, post-processing methods (e.g., retraining and finetuning) are also necessary in the second stage since the performance of these subnets inherited from the supernet is often inferior to that of models trained from scratch. However, the downside of such methods is that the training consumption increases linearly with the number of architectures, which can be problematic. To address this issue, some one-shot NAS methods [3, 4, 29, 33] have utilized a weight entanglement training strategy to share the weights in each operation, eliminating the need for additional finetuning or retraining. Furthermore, a high-quality supernet is essential for candidate architectures to inherit weights directly [33], as well as for accurately ranking candidate architectures [14].", + "bbox": [ + 75, + 90, + 472, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, there is still potential for breakthroughs in training a better supernet. In the first stage, previous methods [7, 4] assume that all candidate architectures are equally important and should be sampled with equal probability during training. However, subnets with different numbers of parameters require different amounts of training resources [2]. For instance, subnets with 1.0 GFLOPs may converge after 30000 iterations, while subnets with 2.0 GFLOPs may require 50000 iterations. Additionally, only the optimal subnet will be deployed, while others will be ignored at the same computational complexity. Therefore, subnets that occupy more training resources may not be distributed in regions corresponding to resource constraints, resulting in sub-optimal performance of the final deployed model. We find that when all subnets are sampled with equal probability, the resulting computational resource distribution is approximately normal. Consequently, subnets trained under this distribution may appear to be under-fitting or over-fitting in different regions.", + "bbox": [ + 75, + 349, + 472, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the challenge of efficiently training a high-performance supernet, we propose a novel method, probability-Shift Neural Architecture search (ShiftNAS). In ShiftNAS, the sampling probability of subnets is not uniform and can be dynamically adjusted during the training process. The training sufficiency of each subnet is measured by evaluating its performance variation under different computational constraints. The subnets with high-performance variance are identified as undertrained, and their sampling probabilities are increased to tilt training resources dynamically towards them. This enables us to allocate resources more effectively and efficiently to achieve better performance for the subnets that need more training resources.", + "bbox": [ + 75, + 641, + 472, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In spite of having an optimal sampling distribution, efficiently and accurately sampling subnets with a desired computational constraint still poses a challenge. To address this", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "issue, we propose an LSTM-based architecture generator (AG) that can be optimized differentiably with a resource constraint loss function. The AG's output is then processed by Gumbel Softmax [12] to generate a one-hot vector policy for each searched operation. To suit the weight-entangled search space [32, 4], we employ a matrix mapping technique that can convert the one-hot vector into a differentiable mask. The mask is multiplied by the operation to obtain a differentiable subnet. The AG and supernet can be jointly trained to learn how to generate the best subnets with desired computational constraints. During evaluation, the AG can generate a corresponding subnet immediately for any given computational constraint. The weights of the searched subnet can be directly inherited from the well-trained supernet, making ShiftNAS free from any additional search or retrain costs.", + "bbox": [ + 496, + 90, + 893, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The overall contribution can be summarized as follows:", + "bbox": [ + 517, + 332, + 885, + 347 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A learnable sampling strategy, called probability shift, is proposed to relief the bias of uniform sampling which leads to performance gap between supernet training and subnet deployment.", + "- We propose an LSTM-based AG to precisely and efficiently offer the best subnet with desired resource constraints. AG training can be differentiably trained with supernet under weight-entangled search space.", + "- We achieve state-of-the-art or competitive results on both CNN and ViT models. Therefore, ShiftNAS is a model-agnostic search method." + ], + "bbox": [ + 517, + 354, + 890, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 549, + 640, + 565 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The one-shot NAS method [3, 29, 33] can automatically search for the optimal architecture in a predefined search space [4, 9, 18], which can be formulated by maximizing an expected accuracy over the space $A$ , i.e.,", + "bbox": [ + 496, + 574, + 890, + 635 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\alpha \\in A} L _ {v a l} \\left(w ^ {*} \\mid \\alpha\\right) \\\\ \\text {s . t .} w ^ {*} = \\underset {w _ {\\alpha}} {\\arg \\min } L _ {\\text {t r a i n}} (w \\mid \\alpha) \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 642, + 890, + 693 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nc o m p l e x i t y (\\alpha) < \\tau\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 696, + 728, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $w$ and $\\alpha$ represent the weight and architecture of subnets. $L_{val}$ and $L_{train}$ are loss functions in the validation and train dataset, respectively. The computational complexity of $\\alpha$ can be calculated by the complexity function. $\\tau$ denotes the complexity threshold for alpha. To solve this optimization problem, a two-stage approach based on weight sharing is usually employed.", + "bbox": [ + 496, + 719, + 890, + 825 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the first stage, a high-quality supernet is trained by sampling a large number of subnets in the defined search space. Notably, in order to obtain subnets that need to be incorporated into the training phase, the following approaches are commonly used in recent one-shot NAS works [4, 7].", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5920", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First is the uniform sampling method, which is also the most common strategy. The Uniform sample method considers that all architectures in the supernet are equally important, that is, they should be sampled with exactly equal probability. The weight optimization equation in the first stage can be formulated as:", + "bbox": [ + 76, + 90, + 472, + 183 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {w} E _ {\\alpha \\sim A} \\left[ L _ {\\text {t r a i n}} (w \\mid \\alpha) \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 191, + 361, + 213 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha = o ^ {1} \\cup o ^ {2} \\cup \\dots \\cup o ^ {D} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 215, + 468, + 231 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\no ^ {i} \\sim U (0, n _ {i})\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 237, + 282, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $D, n$ denotes the total operation and candidate number in certain operations, respectively. $o^i$ denotes the $i$ th selected operation, which follows a uniform distribution.", + "bbox": [ + 76, + 263, + 468, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since there is a gap between the training process and practical deployment, AttentiveNAS [29] proposes the Pareto-aware sampling method. Specifically, models deployed on edge devices are generally around the Pareto frontier, but they are uniformly sampled during training, which leads to the waste of training resources to a certain extent. To bridge the gap between training and deploying, each iteration extracts multiple subnets with equal size and selects the best or worst architecture for training.", + "bbox": [ + 75, + 308, + 468, + 443 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to AttentiveNAS, GreedyNAS [31] screen the weak subnets and just sample from the potentially-good subnets instead of all subnets, thus capturing another opportunity to improve the accuracy of target models.", + "bbox": [ + 76, + 443, + 468, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, these uniformly sampled architectures appear to be an approximately normal distribution of the computational cost. To further reduce the gap between training and deployment, Focusformer [19] proposes that all the sampled subnets should be based on resource distribution and focuses more on the Pareto frontier architectures through an architecture sampler. In addition, FairNAS [5] also reinforces the sampling process with a stricter standard of fairness, that is, the parameters of each choice block are updated the same number of times at any stage.", + "bbox": [ + 75, + 505, + 468, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When the first stage is over, this well-trained supernet can be used as the performance estimator of the candidate architecture. And the weight of candidate architectures can be directly inherited from the supernet.", + "bbox": [ + 75, + 655, + 468, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the second stage, we need to search for the best candidate architecture under different resource constraints in the supernet, which can be formulated as", + "bbox": [ + 75, + 715, + 468, + 760 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha = \\underset {\\alpha \\in A} {\\operatorname {a r g m i n}} L _ {v a l} (w \\mid \\alpha) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 771, + 468, + 800 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns. t. c o m p l e x i t y (\\alpha) < \\tau\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 800, + 354, + 815 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since there are so many architectures in the supernet that need to be evaluated, the consumption of computing resources is also unacceptable. Therefore, recent works resort to random search [17], evolution algorithms [4, 7, 23] or reinforcement learning [35] to find the most promising", + "bbox": [ + 76, + 825, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "architecture among all architectures. In the end, these selected architectures need some post-processing operations such as retraining to improve the architecture performance, but this also brings additional computational consumption.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to alleviate the above-mentioned resource consumption problem, some one-shot NAS methods train a high-quality supernet so that candidate architectures can directly inherit supernet weights without retraining, such as Autoformer [4], attentiveNAS [29], BigNAS [33], FocusFormer [19].", + "bbox": [ + 496, + 152, + 890, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Nerual Architecture Search with Probability Shift", + "text_level": 1, + "bbox": [ + 498, + 258, + 890, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first discuss the training resources allocation problem caused by uniform sampling in one-shot NAS. To solve the above-mentioned problem, we present ShiftNAS, an end-to-end supernet training framework. In ShiftNAS, a learnable sampling strategy is proposed, which can dynamically adjust the training resources allocation by shifting the sampling probability. Then, an LSTM-based AG is designed to accurately obtain the expected subnet for each sampling. We show the overview of our method in Figure 1.", + "bbox": [ + 496, + 303, + 890, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Rethinking the Sampling of One-shot NAS", + "text_level": 1, + "bbox": [ + 498, + 465, + 861, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In one-shot NAS, an overparameterized supernet $S$ is formed with multiple operations, where each operation contains several choices. The supernet $S$ contains $D$ operations, and the $d$ th operation $o^d$ can be selected from $n$ candidates $o_1, \\ldots, o_n$ , which represent various design choices, such as kernel size and channels in CNN search space [33], or heads number and MLP ratio in ViT search space [4]. A subnet $a$ sampled from the supernet can be represented by a tuple of size $D$ , i.e., $a = (o^1, \\ldots, o^D)$ .", + "bbox": [ + 496, + 489, + 890, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In previous one-shot NAS methods, the sampling probability of each operation candidate $o_i$ is given by $p(o_i) = \\frac{1}{n}$ , which assumes that all candidates are equally important. However, in reality, one-shot NAS is more concerned with the computational resources of the subnets. To address this, we introduce the notion of computational resource for each operation candidate $o^d$ , denoted by $b^d \\in b_1, \\ldots, b_n$ . The computational resource of a subnet $a$ randomly sampled from the supernet can then be computed as $B_a = \\sum_{d=1}^{D} b^d$ , where $b^d$ follows a uniform distribution.", + "bbox": [ + 496, + 627, + 890, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Remark 1 If we assume that the computational resource of each operation $b^{d}$ is independently sampled from a uniform distribution, then the total computational resource $B_{a}$ of a subnet $a$ sampled from the supernet follows an Irwin-Hall distribution. As the number of operations $D$ in the supernet increases, the Irwin-Hall distribution converges to a normal distribution [21].", + "bbox": [ + 496, + 795, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5921", + "bbox": [ + 482, + 924, + 513, + 936 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5060003a72932d560ab426fa4fed7db9cf51f1a7cb2aa193565dd52081eddf54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 117, + 90, + 243, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f3adc5a47f36faa731ecb92ce7adf7f26dd5a41c0966ffe2a99d46c84b89e85d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 90, + 429, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ad74b05bf2dcf2d2b8462c9e21c21b1eb1f2b51c2682900b27c7b804b5af1b7f.jpg", + "image_caption": [ + "(a) Uniform distribution of each operation" + ], + "image_footnote": [], + "bbox": [ + 114, + 195, + 243, + 292 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/17c40fd0c370ca01fcb2cb3e69be4e0588b60ac8784c68f29c6164208feb8dbc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 196, + 428, + 292 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1d1366ad1a27c4a1479c1679c671b708734673d3612c6981c9738afad63f2250.jpg", + "image_caption": [ + "(b) Normal distribution of subnets computational resources", + "Figure 2: During supernet training, operations are sampled from a uniform distribution in each iteration. The subnet consists of the sampled operations and its computational resource follows an approximated normal distribution." + ], + "image_footnote": [], + "bbox": [ + 116, + 315, + 431, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Upon analyzing Remark 1, it is evident that the uniform sampling strategy results in subnets being sampled with moderate computational resources, as depicted in Figure 2. This observation implies that subnets with moderate computational resources can be trained effectively, whereas subnets with large or small computational resources cannot be fully trained. Insufficiently-trained subnets lead to inaccurate ranking [14, 15] and inherit unreasonable weights from the supernet without retraining [1].", + "bbox": [ + 75, + 650, + 468, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Shifting Probability with Greedy Sample", + "text_level": 1, + "bbox": [ + 76, + 800, + 426, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One-shot NAS methods rely on the assumption that the supernet can effectively rank the performance of subnets with arbitrary computational resources, and therefore require each subnet to be fully trained with optimal training strategies. However, in practice, subnets with differ", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ent computational resources may require different training strategies, and the uniform sampling strategy used in previous one-shot NAS methods may lead to insufficient training of subnets in certain computational regions. For example, subnets with extreme computational resources may not be fully trained even when subnets with moderate computational resources are fully trained. To address this issue, we propose a learnable greedy sampling strategy that takes into account the training sufficiency of subnets in different computational regions.", + "bbox": [ + 496, + 90, + 890, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For any subnet, we judge whether the subnet has converged by calculating the following gradient of the subnet.", + "bbox": [ + 498, + 242, + 890, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {w} = \\nabla_ {w} L _ {\\text {v a l}} (w, \\alpha \\mid b) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 286, + 890, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{val}$ represents the task loss (e.g. cross entropy) in the validation dataset, $w, \\alpha$ and $b$ denote the weight, architecture and sampled computational resource of the subnet, respectively. When the subnet with $b$ computational resource converges, $\\nabla_w$ will tend to zero. In other words, insufficiency-trained subnets have a small gradient (probably negative) and should be sampled more times. Therefore, we can greedily adjust the sampling distribution of $B$ by optimizing the computational resource distribution at each training step", + "bbox": [ + 496, + 306, + 892, + 458 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {B} {\\arg \\min } E _ {b \\sim B} \\left[ \\nabla_ {w} L _ {v a l} (w, \\alpha \\mid b) \\right] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 465, + 890, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To optimize $B$ end to end based on gradient-based methods, we discrete computational resources into several parts, e.g. $[1.3G, 1.4G, \\dots, 2.0G]$ FLOPs, and uniformly initialize its probability distribution. In each training step, one computational resource can be sampled by Gumbel Softmax $G$ . Formally, Eq. 5 can be rewritten as", + "bbox": [ + 498, + 497, + 890, + 588 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {B} {\\arg \\min } \\nabla_ {w} L _ {v a l} (w, \\alpha \\mid G (B)) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 595, + 890, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this way, the gradient of $B$ in $t$ th update can be obtained by applying chain rule to Eq. 6", + "bbox": [ + 498, + 627, + 890, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {B _ {t}} = \\nabla_ {w _ {t}, B _ {t}} ^ {2} L _ {\\text {v a l}} \\left(w _ {t}, \\alpha \\mid G (B _ {t})\\right) \\nabla_ {B _ {t}} G (B _ {t}) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 665, + 890, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t$ denotes the update times of $B$ . $w_{t}$ represents the supernet weight when $B$ has been updated $t$ times. However, evaluating the gradient of $B_{t}$ requires an expensive matrix-vector product in its second term. Inspired by darts [18], the complexity can be substantially reduced using the finite difference approximation. We use two training steps to approximate the gradient of $w$ , therefore, Eq. 7 can be rewritten as", + "bbox": [ + 496, + 690, + 890, + 809 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {B _ {t}} = \\nabla_ {B _ {t}} L _ {\\text {v a l}} \\left(w _ {t}, \\alpha \\mid G \\left(B _ {t}\\right)\\right) \\nabla_ {B _ {t}} G \\left(B _ {t}\\right) \\tag {8} \\\\ - \\nabla_ {B _ {t}} L _ {v a l} \\left(w _ {t - 1}, \\alpha \\mid G (B _ {t})\\right) \\nabla_ {B _ {t}} G (B _ {t}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 815, + 890, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the implementation, the distribution $B$ is only updated every few iterations, so there is no excessive computational overhead. The iterative procedure is outlined in Appendix.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5922", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dab46cc2834897b0ae8093d2cb84654c0cd98c1b055fac3d9008e73b5bb398ee.jpg", + "image_caption": [ + "Figure 3: An overview of architecture generator training process. AG is given a sampled resource constraint B and generates several one-hot policies sequence by sequence. Then, one-hot policies will be mapped to corresponding masks to keep differentiable." + ], + "image_footnote": [], + "bbox": [ + 117, + 94, + 867, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Generating Architectures with Arbitray Computational Resources", + "text_level": 1, + "bbox": [ + 76, + 439, + 470, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The probability shift is able to sample subnets according to their corresponding training sufficiency. However, a crucial question emerges: given a certain computational resource $b$ , how can we rapidly sample a subnet that satisfies the desired resource constraint? A straightforward approach would be to keep sampling subnets until one is found that meets the resource constraint. Nevertheless, as illustrated in Figure 2, subnets with large or small computational resources have an extremely low probability of being sampled, which makes sampling these subnets computationally inefficient.", + "bbox": [ + 75, + 482, + 468, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this issue, we propose an architecture generator (AG) that can provide the corresponding subnet architecture according to any resource constraint. The AG is designed to stably generate a subnet architecture that satisfies the desired resource constraint. Inspired by previous works [35], which formulate the NAS problem as a sequence prediction problem, we also use an LSTM network to generate each operation sequence by sequence. To update the sampling probability $B$ , we use differentiable neural architecture search (DNAS) [11, 18] to jointly train the AG and the supernet. DNAS is preferred over other search methods, such as RL-based or evolutionary-based methods, as it converges faster.", + "bbox": [ + 75, + 654, + 468, + 851 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section presents an overview of the AG (Architecture Generator) training process with clarity and academic rigor. Figure 3 provides a visual representation of the AG", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "training process. During the training of AG and supernet, the sample distribution $B$ remains static and is not updated. To enable end-to-end updates of AG, the policy of each operation is generated using Gumbel Softmax with a one-hot vector, such as $[1,0,0]$ .", + "bbox": [ + 496, + 440, + 890, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is important to note that the focus of this paper is on weight entanglement. This means that each operation shares weights for their common parts, and weights with small indices are always activated. As a result, policies cannot be directly involved in forward and backward processes. To overcome this issue, a matrix map trick is employed. More specifically, given an one-hot policy $p \\in 1 \\times n$ where $n$ denotes the candidate number in this operation, the mapped policy $p'$ can be computed using the following", + "bbox": [ + 496, + 516, + 892, + 652 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np ^ {\\prime} = M \\cdot p ^ {T} = \\left[ m _ {0} ^ {T}, m _ {1} ^ {T}, \\dots , m _ {n} ^ {T} \\right] \\cdot p ^ {T} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 660, + 890, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M$ consists of several masks and $m_i^T$ represents the mask of $i$ th candidate. We show the matrix map example in Figure 3. Assume a ViT block has three heads and the heads number can be selected in [1,2,3], their weight-entangle masks can be designed as $m^0 = [1,0,0], m^1 = [1,1,0]$ and $m^2 = [1,1,1]$ , respectively. With Eq. 9, the one-hot policy $p$ participates in the forward of the supernet while the gradient of AG is automatically calculated by the chain rule.", + "bbox": [ + 496, + 688, + 890, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since our goal is to make AG learn how to accurately generate a subnet architecture with arbitrary resource constraints, the objective function of AG is designed to reduce task loss $L_{task}$ while minimizing the gap $L_{RC}$ between the resource constraint of the sampled subnet and the target.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5923", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/422fe1bb55e181b8cecce40a9c07ee58852ce9dfd8954252d41dc45a77262bf5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelParameters(M)FLOPs(G)Cost(GPU Days)Top-1 Acc.(%)Top-5 Acc.(%)
DeiT-tiny[28]5.71.22472.291.1
AutoFormer-tiny[4]5.71.33074.792.3
FocusFormer-tiny[19]6.21.42675.193.1
ShiftFormer-T(Ours)5.81.32476.093.1
DeiT-S[28]22.14.73079.995.0
T2T-ViT-14[34]21.56.13281.7-
ViT-S/16[6]22.95.1-78.8-
BoTNet-S1-59[26]33.57.3-81.795.8
AutoFormer-small[4]22.95.13581.495.6
FocusFormer-small[19]23.75.03281.695.6
ShiftFormer-S(Ours)23.65.03082.295.8
DeiT-B[28]86.617.64381.895.6
ViT-B/16[6]86.617.6-79.7-
AutoFormer-base[4]52.811.04381.495.7
FocusFormer-base[19]52.811.04181.995.6
ShiftFormer-B(Ours)52.811.04082.896.1
MobileNetV2 0.75×[24]2.60.211869.8-
MobileNetV3 1.0×[9]5.40.22379175.2-
DS-MBNet-M[16]-0.332473.2-
BigNAS-S[33]4.50.2411276.5-
Once-For-All[3]4.40.2310576.4-
GreedyNAS-C[31]4.70.283276.292.5
ShiftCNN-S(Ours)4.50.243277.293.1
MobileNetV2 1.3×[24]5.30.502172.8-
MobileNetV3 1.25×[9]8.10.35379176.7-
EfficientNet-B0[27]5.40.39379177.193.3
DS-MBNet-S[16]-0.572474.8-
BigNAS-M[33]5.50.4211278.9-
GreedyNAS-A[31]6.50.374077.193.3
ShiftCNN-B(Ours)5.60.423279.693.6
", + "bbox": [ + 106, + 87, + 867, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: ShiftNAS models performance on ImageNet with comparisons to other models. We group the models according to their FLOPs. We use ShiftFormer and ShiftCNN to denote the models searched by ShiftNAS. Cost represents the total GPU days including training, searching and retraining.", + "bbox": [ + 75, + 603, + 893, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, the joint loss $L$ is given by", + "bbox": [ + 76, + 688, + 328, + 704 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {\\text {t a s k}} + \\lambda L _ {R C}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 713, + 302, + 729 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {R C} = \\left(\\sum_ {i = 1} ^ {D} \\sum_ {j = 1} ^ {n} b _ {j} ^ {i} p _ {j} ^ {\\prime i} - C\\right) ^ {2} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 732, + 468, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $b_{j}^{i}$ and $p_{j}^{\\prime i}$ represents the computational resource and policy of $j$ th candidate in $i$ th operation, $C$ indicates the target resource constraint sampled from sampling distribution B. $\\lambda$ is the coefficient of $L_{RC}$ . In the implementation, we jointly optimize the AG and the supernet on the train dataset with Eq. 10. After few epochs, AG can generate the corresponding subnets with given resource constraint.", + "bbox": [ + 75, + 794, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 686, + 624, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 715, + 689, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset and evaluation metrics. We conduct main experiments on ImageNet [13] which contains about 1.2M images for training and $50\\mathrm{K}$ images for validation. We use the Top-1 accuracy and the number of giga floating-point operations(GFLOPs) to measure the performance and efficiency of networks.", + "bbox": [ + 498, + 744, + 890, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. ShiftNAS is a model-agnostic search method. Therefore, we search the models with various resource constraints on both CNN and ViT search space. We split the computational resource range into $K$", + "bbox": [ + 498, + 839, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5924", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b401a713a63d725250a45f8a463b5e734c3615e2bdcdf814783740b04aac9a72.jpg", + "image_caption": [ + "(a) ViT-tiny" + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 285, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/07fb239bd32efec950164da70ba4c0e187ded04c6becaad9d38ca2858f023a51.jpg", + "image_caption": [ + "(b) ViT-small", + "Figure 4: The effectiveness of probability shift. FLOPs/Accuracy tradeoffs of ShiftNAS with and without probability shift on search space (a) ViT-tiny; (b) ViT-small; (c) ViT-base; (d) CNN." + ], + "image_footnote": [], + "bbox": [ + 285, + 88, + 493, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b653b1a2a2b831ab127af11c88c669f6a8a61b4f89a30e5c42da785b9cf183b2.jpg", + "image_caption": [ + "(c) ViT-base" + ], + "image_footnote": [], + "bbox": [ + 493, + 88, + 697, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a503801ba6bd3e394c0218737a52fafacef978a2119ae084f9d7038bf697042.jpg", + "image_caption": [ + "(d) CNN" + ], + "image_footnote": [], + "bbox": [ + 697, + 88, + 903, + 207 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/eafdb08de79c8063608b557a6ef70113e106e513564f446fd150b735cb551bf7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelInherit(%)Finetune(%)Retrain(%)
ShiftFormer-T76.075.9(-0.1)76.1(+0.1)
ShiftFormer-S82.082.0(0.0)82.0(0.0)
ShiftFormer-B82.482.3(-0.1)82.6(+0.2)
ShiftCNN-S77.277.0(-0.2)77.3(+0.1)
ShiftCNN-B79.679.5(-0.1)79.6(0.0)
", + "bbox": [ + 78, + 291, + 478, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2: Comparison of subnets Top-1 accuracy with inherited, finetuned and retrained weights.", + "bbox": [ + 75, + 407, + 467, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "parts with 0.1 GFLOPs step. The sampling distribution will be represented by a $1 \\times K$ learnable vector. We randomly split $50\\mathrm{K}$ images from ImageNet and use Adam optimizer with 1e-3 learning rate to update the sampling distribution vector. The AG consists of an LSTM cell with 64 hidden numbers and 4/10 fully connected layers for ViT/CNN supernet and will be updated jointly with the supernet. In the first 50 training epochs, we only optimize the supernet weight and the AG since the AG cannot generate the corresponding subnets at the beginning. After 50 epochs, the sampling distribution vector is updated per 100 iteration. We discuss more hyperparameter settings in the Appendix, such as the split step of the computational resource and the updated frequency of the sampling distribution vector. The training strategies for both CNN and ViT are given in Appendix. All the models are trained on 8 Nvidia Tesla A100 GPUs.", + "bbox": [ + 75, + 457, + 468, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For CNN search space, we follow BigNAS [33] where the search space contains kernel size, channel number, depth and input resolution. The supernet will be split into 7 stages. Each stage has multiple choices of the block number and the first block of each stage has no residual path. The policy of kernel size will be generated block-wise, where the small kernel size is center cropped from the large kernel size. The policies of channel number and depth will be predicted stage-wise, where the lower-index channels and blocks are preferentially kept.", + "bbox": [ + 75, + 715, + 467, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For ViT search space, we follow AutoFormer [4] which searches embed dim, head number, MLP ratio and depth", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with three different scale settings. The policies of head number and MLP ratio will be given block-wise. As CNN, the lower-index heads and neurons are preferentially kept.", + "bbox": [ + 496, + 292, + 890, + 337 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After training, we directly sample the candidate with the max probability given by the AG. It is noted that the obtained model inherits weights from the trained supernet without retraining or fine-tuning. Therefore, no search or retraining cost exists in ShiftNAS.", + "bbox": [ + 496, + 337, + 890, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 500, + 422, + 637, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We search ViT and CNN architectures on ImageNet with different FLOPs constraints. We compare searched models with multiple ViT and CNN models on model performance (Top-1 Acc., Top-5 Acc.) and efficiency (FLOPs, parameters). It is seen from Table 1 that our ShiftFormer models surpass the recent manual and autoML-based transformer models under diverse model sizes. It is worth noting that ShiftFormer-tiny achieves a top-1 Acc. of 76.0 with only 1.3 GFLOPs, being $1.3\\%$ and $0.9\\%$ better than AutoFormer-tiny and FocusFormer-T which are also autoML-based methods, respectively. Compared to CNN models, ShiftCNN models also achieve better accuracy than all compared networks under similar FLOPs restrictions. For example, ShiftCNN-S and ShiftCNN-B achieve $0.7\\%$ and $0.7\\%$ higher top-1 accuracy than BigNAS-S and BigNAS-M, respectively. Specifically, ShiftCNN and BigNAS models are searched from the same search space. BigNAS models spend more than $2\\times$ GPU time than ShiftCNN models since the sandwich rule is used to train BigNAS supernet, which demonstrates the efficiency and outstanding performance of ShiftNAS. For fairness, we do not compare our models with others obtained from better search spaces [15, 29] in Table 1. The experiments on AttentiveNAS search space can be found in the Appendix.", + "bbox": [ + 496, + 445, + 890, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 816, + 651, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The effectiveness of probability shift. To demonstrate the effectiveness of the proposed sampling probability shift, we trained a ShiftNAS supernet and its baseline counterpart under ViT-tiny, ViT-small, ViT-base, and CNN search", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5925", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5562594818d2f7f6de501150e4e7e4a00017b0ddeedb956bc27cf742d07e415.jpg", + "image_caption": [ + "Figure 5: The sampling probability with different FLOPs groups on search space (a) ViT-tiny; (b) ViT-small; (c) ViT-base; (d) CNN." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 282, + 217 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dd20aabde47e64996b0068bd39f23c97ad57c5c6282821fc0856fca9036bfa4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 88, + 488, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/929abfbae529bd8c1aa5c83ed4c5bbb247a9fcd86a116d37dbf9cf6076da51c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 88, + 696, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2d79223fc352330bc002925a085337fd9186e29180aff94bdca2e2adb5439be8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 88, + 903, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/78bd1c72ac16b7d02f9ef833409f3a6a7e46e1341186558701ae38c350393970.jpg", + "image_caption": [ + "Figure 6: The ranking correlation of architecture generator." + ], + "image_footnote": [], + "bbox": [ + 106, + 282, + 439, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "spaces. In the baseline supernet training, we uniformly sampled FLOPs in each step and did not update FLOPs probability. The training settings for all supernets are provided in the Appendix. As illustrated in Figure 4, the supernet trained with probability shift is capable of covering a wide range of accuracy-FLOPs curves and outperforms the baseline supernet trained without probability shift by a significant margin, demonstrating the effectiveness of the proposed sampling probability shift mechanism.", + "bbox": [ + 75, + 535, + 467, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The ranking correlation of architecture generator. The experiment is based on ShiftFormer-tiny supernet. We measure the ranking correlation by randomly sampling 25 subnets from the ShiftFormer-tiny supernet. For fast evaluation, The sampled subnets are inherited weights from the supernet and are further finetuned to reach the optimal weights. As for the model ranking test, we evaluate sampled subnets in both supernets to get their inherited performance. As shown in Figure 6, the inherited performance is highly correlated with finetuned performance in good subnets (Inherited Top1-Acc > 65%). Despite the correlation with bad subnets is not obvious, we focus more attention on good subnets since bad subnets are not deployed in practice.", + "bbox": [ + 75, + 672, + 467, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Subnet performance without retraining. We further study the performance of ShiftNAS models when models", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "are independently fine-tuned or trained from scratch. Following Autoformer, we finetune or retrain the selected subnet with 40 epochs or 300 epochs on ImageNet. Table 2 shows that finetuning cannot improve or even harm the performance of subnets inherited from supernet. Besides, retraining can slightly improve the subnet performance but cost extra training resources which prevents the implementation of NAS.", + "bbox": [ + 496, + 286, + 890, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The optimal sampling strategy. ShiftNAS learns the optimal sampling probability for each subnet with different FLOPs. To figure out what sampling strategy is the best, we visualize the various sampling probability in different search spaces (ViT-tiny, ViT-small, ViT-base and CNN) at different training steps. As can be seen in Figure 5, our method learns to sample more subnets with large computational complexity. For example, the sampling probabilities of 1.6 to 1.9 GFLOPs subnets are clearly higher than 1.3 and 1.4 GFLOPs subnets in the ViT-tiny search space. This is intuitive since the large subnet contains more parameters and needs more computational resources to update.", + "bbox": [ + 496, + 407, + 890, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 603, + 617, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we presented ShiftNAS, a one-shot supernet training framework that can complexity-wise adjust the sampling probability. To automatically adjust the sampling probability, we proposed probability shift that can be learned according to the subnets training sufficiency. To fast obtain the subnets with desired complexity, AG is designed that can be jointly trained with the supernet in an end-to-end manner. Experiments on ImageNet showed that our method achieves SOTA results on ViT/CNN search spaces, and covers a wide range of efficiency/accuracy trade-offs without any extra retraining.", + "bbox": [ + 496, + 628, + 890, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation. ShiftNAS is designed for efficient search on weight-entanglement search spaces, so it cannot work on darts-like search spaces [18, 25].", + "bbox": [ + 496, + 794, + 890, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: This work was supported by the National Natural Science Foundation of China (62203392) and Natural Science Foundation of Zhejiang province (LY21F030018).", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5926", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Gabriel Bender, Pieter-Jan Kindermans, Barret Zoph, Vijay Vasudevan, and Quoc Le. Understanding and simplifying one-shot architecture search. In International conference on machine learning, pages 550-559. PMLR, 2018. 4", + "[2] Han Cai, Chuang Gan, Ji Lin, and Song Han. Network augmentation for tiny deep learning. arXiv preprint arXiv:2110.08890, 2021. 2", + "[3] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. In International Conference on Learning Representations, 2019. 2, 6", + "[4] Minghao Chen, Houwen Peng, Jianlong Fu, and Haibin Ling. Autoformer: Searching transformers for visual recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12270-12280, 2021. 2, 3, 6, 7", + "[5] Xiangxiang Chu, Bo Zhang, Ruijun Xu, and Jixiang Li. Fairnas: Rethinking evaluation fairness of weight sharing neural architecture search. arXiv preprint arXiv:1907.01845, 2019. 3", + "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 6", + "[7] Zichao Guo, Xiangyu Zhang, Haoyuan Mu, Wen Heng, Zechun Liu, Yichen Wei, and Jian Sun. Single path one-shot neural architecture search with uniform sampling. In European Conference on Computer Vision, pages 544-560. Springer, 2020. 1, 2, 3", + "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1", + "[9] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 1, 2, 6", + "[10] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 1", + "[11] Shoukang Hu, Sirui Xie, Hehui Zheng, Chunxiao Liu, Jianping Shi, Xunying Liu, and Dahua Lin. Dsnas: Direct neural architecture search without parameter retraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12084-12092, 2020. 5", + "[12] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. 2016. 2", + "[13] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural net" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "works. Advances in neural information processing systems, 25:1097-1105, 2012. 6", + "[14] Changlin Li, Jiefeng Peng, Liuchun Yuan, Guangrun Wang, Xiaodan Liang, Liang Lin, and Xiaojun Chang. Blockwisely supervised neural architecture search with knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1989-1998, 2020. 2, 4", + "[15] Changlin Li, Tao Tang, Guangrun Wang, Jiefeng Peng, Bing Wang, Xiaodan Liang, and Xiaojun Chang. Bossnas: Exploring hybrid cnn-transformers with block-wisely self-supervised neural architecture search. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12281-12291, 2021. 4, 7", + "[16] Changlin Li, Guangrun Wang, Bing Wang, Xiaodan Liang, Zhihui Li, and Xiaojun Chang. Ds-net++: Dynamic weight slicing for efficient inference in cnns and vision transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-16, 2022. 6", + "[17] Liam Li and Ameet Talwalkar. Random search and reproducibility for neural architecture search. In Uncertainty in artificial intelligence, pages 367-377. PMLR, 2020. 3", + "[18] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. In International Conference on Learning Representations, 2018. 2, 4, 5, 8", + "[19] Jing Liu, Jianfei Cai, and Bohan Zhuang. Focusformer: Focusing on what we need via architecture sampler. arXiv preprint arXiv:2208.10861, 2022. 3, 6", + "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1", + "[21] James E Marengo, David L Farnsworth, and Lucas Stefanic. A geometric derivation of the irwin-hall distribution. International Journal of Mathematics and Mathematical Sciences, 2017, 2017. 3", + "[22] Yameng Peng, Andy Song, Vic Ciesielski, Haytham M. Fayek, and Xiaojun Chang. Pre-nas: Evolutionary neural architecture search with predictor. IEEE Transactions on Evolutionary Computation, 27(1):26-36, 2023. 1", + "[23] Hieu Pham, Melody Guan, Barret Zoph, Quoc Le, and Jeff Dean. Efficient neural architecture search via parameters sharing. In International conference on machine learning, pages 4095-4104. PMLR, 2018. 1, 3", + "[24] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 1, 6", + "[25] Julien Siems, Lucas Zimmer, Arber Zela, Jovita Lukasik, Margret Keuper, and Frank Hutter. Nas-bench-301 and the case for surrogate benchmarks for neural architecture search. arXiv preprint arXiv:2008.09777, 2020. 8", + "[26] Aravind Srinivas, Tsung-Yi Lin, Niki Parmar, Jonathon Shlens, Pieter Abbeel, and Ashish Vaswani. Bottleneck transformers for visual recognition. In Proceedings of" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5927", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF conference on computer vision and pattern recognition, pages 16519-16529, 2021. 6", + "[27] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. 6", + "[28] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 6", + "[29] Dilin Wang, Meng Li, Chengyue Gong, and Vikas Chandra. Attentivenas: Improving neural architecture search via attentive sampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6418-6427, 2021. 2, 3, 7", + "[30] Bichen Wu, Xiaoliang Dai, Peizhao Zhang, Yanghan Wang, Fei Sun, Yiming Wu, Yuandong Tian, Peter Vajda, Yangqing Jia, and Kurt Keutzer. Fbnet: Hardware-aware efficient convnet design via differentiable neural architecture search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10734-10742, 2019. 1", + "[31] Shan You, Tao Huang, Mingmin Yang, Fei Wang, Chen Qian, and Changshui Zhang. Greedynas: Towards fast one-shot nas with greedy supernet. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1999-2008, 2020. 3, 6", + "[32] Jiahui Yu and Thomas Huang. Autoslim: Towards one-shot architecture search for channel numbers. arXiv preprint arXiv:1903.11728, 2019. 2", + "[33] Jiahui Yu, Pengchong Jin, Hanxiao Liu, Gabriel Bender, Pieter-Jan Kindermans, Mingxing Tan, Thomas Huang, Xiaodan Song, Ruoming Pang, and Quoc Le. Bignas: Scaling up neural architecture search with big single-stage models. In European Conference on Computer Vision, pages 702-717. Springer, 2020. 2, 3, 6, 7", + "[34] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF international conference on computer vision, pages 558-567, 2021. 6", + "[35] Barret Zoph and Quoc V Le. Neural architecture search with reinforcement learning. arXiv preprint arXiv:1611.01578, 2016. 1, 3, 5" + ], + "bbox": [ + 78, + 92, + 468, + 724 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5928", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/ShiftNAS_ Improving One-shot NAS via Probability Shift/9d0cd6fb-8343-4e85-b19e-ff450eb2e5e4_model.json b/2023/ShiftNAS_ Improving One-shot NAS via Probability Shift/9d0cd6fb-8343-4e85-b19e-ff450eb2e5e4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..283205b904bf4a4c22cc5e4e2951b6a3e4908d74 --- /dev/null +++ b/2023/ShiftNAS_ Improving One-shot NAS via Probability Shift/9d0cd6fb-8343-4e85-b19e-ff450eb2e5e4_model.json @@ -0,0 +1,2035 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.196, + 0.131, + 0.776, + 0.154 + ], + "angle": 0, + "content": "ShiftNAS: Improving One-shot NAS via Probability Shift" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.181, + 0.705, + 0.234 + ], + "angle": 0, + "content": "Mingyang Zhang, Xinyi Yu* Haodong Zhao, Linlin Ou* Zhejiang University of Technology Hangzhou, Zhejiang, China" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.238, + 0.655, + 0.253 + ], + "angle": 0, + "content": "yuxy@zjut.edu.cn, linlinou@zjut.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.473, + 0.516 + ], + "angle": 0, + "content": "One-shot Neural architecture search (One-shot NAS) has been proposed as a time-efficient approach to obtain optimal subnet architectures and weights under different complexity cases by training only once. However, the subnet performance obtained by weight sharing is often inferior to the performance achieved by retraining. In this paper, we investigate the performance gap and attribute it to the use of uniform sampling, which is a common approach in supernet training. Uniform sampling concentrates training resources on subnets with intermediate computational resources, which are sampled with high probability. However, subnets with different complexity regions require different optimal training strategies for optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.473, + 0.775 + ], + "angle": 0, + "content": "To address the problem of uniform sampling, we propose ShiftNAS, a method that can adjust the sampling probability based on the complexity of subnets. We achieve this by evaluating the performance variation of subnets with different complexity and designing an architecture generator that can accurately and efficiently provide subnets with the desired complexity. Both the sampling probability and the architecture generator can be trained end-to-end in a gradient-based manner. With ShiftNAS, we can directly obtain the optimal model architecture and parameters for a given computational complexity. We evaluate our approach on multiple visual network models, including convolutional neural networks (CNNs) and vision transformers (ViTs), and demonstrate that ShiftNAS is model-agnostic. Experimental results on ImageNet show that ShiftNAS can improve the performance of one-shot NAS without additional consumption. Source codes are available at GitHub." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.804, + 0.21, + 0.82 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.83, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) have been widely applied to the field of computer vision with remarkable success [6, 8]. However, the deployment of these vision net" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.288, + 0.891, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.475, + 0.895, + 0.627 + ], + "angle": 0, + "content": "Figure 1: A conceptual overview of our ShiftNAS. Focusing on the computational resource (e.g. FLOPs), ShiftNAS split a search space into several parts where subnets have close computational complexity. In stochastic supernet training, ShiftNAS first samples computational complexity according to probability, then an architecture generator samples the subnet with desired computational complexity. The probability distribution is dynamically updated with stochastic training by estimating the performance variation of each sub-space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.895, + 0.779 + ], + "angle": 0, + "content": "works on edge devices still has some limitations, such as the massive model sizes and excessive computation overhead [9, 10, 20, 24]. In addition, designing architectures artificially in a trial-and-error manner is a resource-consuming task that requires not only architectural skills but also domain expertise. Consequently, how to acquire optimal architectures that balance latency and accuracy efficiently is of paramount importance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Recent advancements in neural architecture search (NAS) methods, such as [22, 7, 30, 35], have led to significant improvements in the performance of practical applications by automatically searching for optimal architectures within a defined search space. However, traditional NAS methods typically require a substantial computation budget [23, 35]. In order to speed up training and reduce the resource consumption of the training process, one-shot NAS" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.372, + 0.901 + ], + "angle": 0, + "content": "* Xinyi Yu and Linlin Ou are corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "5919" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.349 + ], + "angle": 0, + "content": "methods [18, 7] have adopted a two-stage training approach based on weight sharing. Specifically, a supernet is trained in the first stage, and subnets with better performance are searched for in the second stage. In some cases, post-processing methods (e.g., retraining and finetuning) are also necessary in the second stage since the performance of these subnets inherited from the supernet is often inferior to that of models trained from scratch. However, the downside of such methods is that the training consumption increases linearly with the number of architectures, which can be problematic. To address this issue, some one-shot NAS methods [3, 4, 29, 33] have utilized a weight entanglement training strategy to share the weights in each operation, eliminating the need for additional finetuning or retraining. Furthermore, a high-quality supernet is essential for candidate architectures to inherit weights directly [33], as well as for accurately ranking candidate architectures [14]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.473, + 0.639 + ], + "angle": 0, + "content": "However, there is still potential for breakthroughs in training a better supernet. In the first stage, previous methods [7, 4] assume that all candidate architectures are equally important and should be sampled with equal probability during training. However, subnets with different numbers of parameters require different amounts of training resources [2]. For instance, subnets with 1.0 GFLOPs may converge after 30000 iterations, while subnets with 2.0 GFLOPs may require 50000 iterations. Additionally, only the optimal subnet will be deployed, while others will be ignored at the same computational complexity. Therefore, subnets that occupy more training resources may not be distributed in regions corresponding to resource constraints, resulting in sub-optimal performance of the final deployed model. We find that when all subnets are sampled with equal probability, the resulting computational resource distribution is approximately normal. Consequently, subnets trained under this distribution may appear to be under-fitting or over-fitting in different regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.473, + 0.852 + ], + "angle": 0, + "content": "To address the challenge of efficiently training a high-performance supernet, we propose a novel method, probability-Shift Neural Architecture search (ShiftNAS). In ShiftNAS, the sampling probability of subnets is not uniform and can be dynamically adjusted during the training process. The training sufficiency of each subnet is measured by evaluating its performance variation under different computational constraints. The subnets with high-performance variance are identified as undertrained, and their sampling probabilities are increased to tilt training resources dynamically towards them. This enables us to allocate resources more effectively and efficiently to achieve better performance for the subnets that need more training resources." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "In spite of having an optimal sampling distribution, efficiently and accurately sampling subnets with a desired computational constraint still poses a challenge. To address this" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.332 + ], + "angle": 0, + "content": "issue, we propose an LSTM-based architecture generator (AG) that can be optimized differentiably with a resource constraint loss function. The AG's output is then processed by Gumbel Softmax [12] to generate a one-hot vector policy for each searched operation. To suit the weight-entangled search space [32, 4], we employ a matrix mapping technique that can convert the one-hot vector into a differentiable mask. The mask is multiplied by the operation to obtain a differentiable subnet. The AG and supernet can be jointly trained to learn how to generate the best subnets with desired computational constraints. During evaluation, the AG can generate a corresponding subnet immediately for any given computational constraint. The weights of the searched subnet can be directly inherited from the well-trained supernet, making ShiftNAS free from any additional search or retrain costs." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.333, + 0.887, + 0.348 + ], + "angle": 0, + "content": "The overall contribution can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.355, + 0.892, + 0.416 + ], + "angle": 0, + "content": "- A learnable sampling strategy, called probability shift, is proposed to relief the bias of uniform sampling which leads to performance gap between supernet training and subnet deployment." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.424, + 0.892, + 0.485 + ], + "angle": 0, + "content": "- We propose an LSTM-based AG to precisely and efficiently offer the best subnet with desired resource constraints. AG training can be differentiably trained with supernet under weight-entangled search space." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.493, + 0.892, + 0.537 + ], + "angle": 0, + "content": "- We achieve state-of-the-art or competitive results on both CNN and ViT models. Therefore, ShiftNAS is a model-agnostic search method." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.355, + 0.892, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.55, + 0.642, + 0.566 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.892, + 0.636 + ], + "angle": 0, + "content": "The one-shot NAS method [3, 29, 33] can automatically search for the optimal architecture in a predefined search space [4, 9, 18], which can be formulated by maximizing an expected accuracy over the space \\(A\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.643, + 0.892, + 0.694 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\alpha \\in A} L _ {v a l} \\left(w ^ {*} \\mid \\alpha\\right) \\\\ \\text {s . t .} w ^ {*} = \\underset {w _ {\\alpha}} {\\arg \\min } L _ {\\text {t r a i n}} (w \\mid \\alpha) \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.698, + 0.729, + 0.713 + ], + "angle": 0, + "content": "\\[\nc o m p l e x i t y (\\alpha) < \\tau\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.826 + ], + "angle": 0, + "content": "where \\( w \\) and \\( \\alpha \\) represent the weight and architecture of subnets. \\( L_{val} \\) and \\( L_{train} \\) are loss functions in the validation and train dataset, respectively. The computational complexity of \\( \\alpha \\) can be calculated by the complexity function. \\( \\tau \\) denotes the complexity threshold for alpha. To solve this optimization problem, a two-stage approach based on weight sharing is usually employed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the first stage, a high-quality supernet is trained by sampling a large number of subnets in the defined search space. Notably, in order to obtain subnets that need to be incorporated into the training phase, the following approaches are commonly used in recent one-shot NAS works [4, 7]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "5920" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "First is the uniform sampling method, which is also the most common strategy. The Uniform sample method considers that all architectures in the supernet are equally important, that is, they should be sampled with exactly equal probability. The weight optimization equation in the first stage can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.192, + 0.362, + 0.214 + ], + "angle": 0, + "content": "\\[\n\\min _ {w} E _ {\\alpha \\sim A} \\left[ L _ {\\text {t r a i n}} (w \\mid \\alpha) \\right]\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.216, + 0.469, + 0.232 + ], + "angle": 0, + "content": "\\[\n\\alpha = o ^ {1} \\cup o ^ {2} \\cup \\dots \\cup o ^ {D} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.238, + 0.283, + 0.254 + ], + "angle": 0, + "content": "\\[\no ^ {i} \\sim U (0, n _ {i})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.264, + 0.47, + 0.308 + ], + "angle": 0, + "content": "where \\(D, n\\) denotes the total operation and candidate number in certain operations, respectively. \\(o^i\\) denotes the \\(i\\)th selected operation, which follows a uniform distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.309, + 0.469, + 0.444 + ], + "angle": 0, + "content": "Since there is a gap between the training process and practical deployment, AttentiveNAS [29] proposes the Pareto-aware sampling method. Specifically, models deployed on edge devices are generally around the Pareto frontier, but they are uniformly sampled during training, which leads to the waste of training resources to a certain extent. To bridge the gap between training and deploying, each iteration extracts multiple subnets with equal size and selects the best or worst architecture for training." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.444, + 0.469, + 0.505 + ], + "angle": 0, + "content": "Similar to AttentiveNAS, GreedyNAS [31] screen the weak subnets and just sample from the potentially-good subnets instead of all subnets, thus capturing another opportunity to improve the accuracy of target models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.469, + 0.655 + ], + "angle": 0, + "content": "However, these uniformly sampled architectures appear to be an approximately normal distribution of the computational cost. To further reduce the gap between training and deployment, Focusformer [19] proposes that all the sampled subnets should be based on resource distribution and focuses more on the Pareto frontier architectures through an architecture sampler. In addition, FairNAS [5] also reinforces the sampling process with a stricter standard of fairness, that is, the parameters of each choice block are updated the same number of times at any stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.469, + 0.716 + ], + "angle": 0, + "content": "When the first stage is over, this well-trained supernet can be used as the performance estimator of the candidate architecture. And the weight of candidate architectures can be directly inherited from the supernet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.469, + 0.761 + ], + "angle": 0, + "content": "In the second stage, we need to search for the best candidate architecture under different resource constraints in the supernet, which can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.772, + 0.469, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\alpha = \\underset {\\alpha \\in A} {\\operatorname {a r g m i n}} L _ {v a l} (w \\mid \\alpha) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.801, + 0.355, + 0.816 + ], + "angle": 0, + "content": "\\[\ns. t. c o m p l e x i t y (\\alpha) < \\tau\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Since there are so many architectures in the supernet that need to be evaluated, the consumption of computing resources is also unacceptable. Therefore, recent works resort to random search [17], evolution algorithms [4, 7, 23] or reinforcement learning [35] to find the most promising" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "architecture among all architectures. In the end, these selected architectures need some post-processing operations such as retraining to improve the architecture performance, but this also brings additional computational consumption." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.892, + 0.243 + ], + "angle": 0, + "content": "In order to alleviate the above-mentioned resource consumption problem, some one-shot NAS methods train a high-quality supernet so that candidate architectures can directly inherit supernet weights without retraining, such as Autoformer [4], attentiveNAS [29], BigNAS [33], FocusFormer [19]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.26, + 0.892, + 0.293 + ], + "angle": 0, + "content": "3. Nerual Architecture Search with Probability Shift" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.456 + ], + "angle": 0, + "content": "In this section, we first discuss the training resources allocation problem caused by uniform sampling in one-shot NAS. To solve the above-mentioned problem, we present ShiftNAS, an end-to-end supernet training framework. In ShiftNAS, a learnable sampling strategy is proposed, which can dynamically adjust the training resources allocation by shifting the sampling probability. Then, an LSTM-based AG is designed to accurately obtain the expected subnet for each sampling. We show the overview of our method in Figure 1." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.467, + 0.862, + 0.483 + ], + "angle": 0, + "content": "3.1. Rethinking the Sampling of One-shot NAS" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.892, + 0.627 + ], + "angle": 0, + "content": "In one-shot NAS, an overparameterized supernet \\( S \\) is formed with multiple operations, where each operation contains several choices. The supernet \\( S \\) contains \\( D \\) operations, and the \\( d \\)th operation \\( o^d \\) can be selected from \\( n \\) candidates \\( o_1, \\ldots, o_n \\), which represent various design choices, such as kernel size and channels in CNN search space [33], or heads number and MLP ratio in ViT search space [4]. A subnet \\( a \\) sampled from the supernet can be represented by a tuple of size \\( D \\), i.e., \\( a = (o^1, \\ldots, o^D) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.779 + ], + "angle": 0, + "content": "In previous one-shot NAS methods, the sampling probability of each operation candidate \\( o_i \\) is given by \\( p(o_i) = \\frac{1}{n} \\), which assumes that all candidates are equally important. However, in reality, one-shot NAS is more concerned with the computational resources of the subnets. To address this, we introduce the notion of computational resource for each operation candidate \\( o^d \\), denoted by \\( b^d \\in b_1, \\ldots, b_n \\). The computational resource of a subnet \\( a \\) randomly sampled from the supernet can then be computed as \\( B_a = \\sum_{d=1}^{D} b^d \\), where \\( b^d \\) follows a uniform distribution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Remark 1 If we assume that the computational resource of each operation \\( b^{d} \\) is independently sampled from a uniform distribution, then the total computational resource \\( B_{a} \\) of a subnet \\( a \\) sampled from the supernet follows an Irwin-Hall distribution. As the number of operations \\( D \\) in the supernet increases, the Irwin-Hall distribution converges to a normal distribution [21]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.514, + 0.937 + ], + "angle": 0, + "content": "5921" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.092, + 0.245, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.092, + 0.431, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.196, + 0.244, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.197, + 0.429, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.299, + 0.4, + 0.313 + ], + "angle": 0, + "content": "(a) Uniform distribution of each operation" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.316, + 0.432, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.534, + 0.449, + 0.548 + ], + "angle": 0, + "content": "(b) Normal distribution of subnets computational resources" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.56, + 0.47, + 0.621 + ], + "angle": 0, + "content": "Figure 2: During supernet training, operations are sampled from a uniform distribution in each iteration. The subnet consists of the sampled operations and its computational resource follows an approximated normal distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.469, + 0.788 + ], + "angle": 0, + "content": "Upon analyzing Remark 1, it is evident that the uniform sampling strategy results in subnets being sampled with moderate computational resources, as depicted in Figure 2. This observation implies that subnets with moderate computational resources can be trained effectively, whereas subnets with large or small computational resources cannot be fully trained. Insufficiently-trained subnets lead to inaccurate ranking [14, 15] and inherit unreasonable weights from the supernet without retraining [1]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.427, + 0.817 + ], + "angle": 0, + "content": "3.2. Shifting Probability with Greedy Sample" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "One-shot NAS methods rely on the assumption that the supernet can effectively rank the performance of subnets with arbitrary computational resources, and therefore require each subnet to be fully trained with optimal training strategies. However, in practice, subnets with differ" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.242 + ], + "angle": 0, + "content": "ent computational resources may require different training strategies, and the uniform sampling strategy used in previous one-shot NAS methods may lead to insufficient training of subnets in certain computational regions. For example, subnets with extreme computational resources may not be fully trained even when subnets with moderate computational resources are fully trained. To address this issue, we propose a learnable greedy sampling strategy that takes into account the training sufficiency of subnets in different computational regions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.243, + 0.892, + 0.273 + ], + "angle": 0, + "content": "For any subnet, we judge whether the subnet has converged by calculating the following gradient of the subnet." + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.287, + 0.892, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {w} = \\nabla_ {w} L _ {\\text {v a l}} (w, \\alpha \\mid b) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.893, + 0.459 + ], + "angle": 0, + "content": "where \\( L_{val} \\) represents the task loss (e.g. cross entropy) in the validation dataset, \\( w, \\alpha \\) and \\( b \\) denote the weight, architecture and sampled computational resource of the subnet, respectively. When the subnet with \\( b \\) computational resource converges, \\( \\nabla_w \\) will tend to zero. In other words, insufficiency-trained subnets have a small gradient (probably negative) and should be sampled more times. Therefore, we can greedily adjust the sampling distribution of \\( B \\) by optimizing the computational resource distribution at each training step" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.467, + 0.892, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\underset {B} {\\arg \\min } E _ {b \\sim B} \\left[ \\nabla_ {w} L _ {v a l} (w, \\alpha \\mid b) \\right] \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.498, + 0.892, + 0.589 + ], + "angle": 0, + "content": "To optimize \\( B \\) end to end based on gradient-based methods, we discrete computational resources into several parts, e.g. \\([1.3G, 1.4G, \\dots, 2.0G]\\) FLOPs, and uniformly initialize its probability distribution. In each training step, one computational resource can be sampled by Gumbel Softmax \\( G \\). Formally, Eq. 5 can be rewritten as" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.597, + 0.891, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\underset {B} {\\arg \\min } \\nabla_ {w} L _ {v a l} (w, \\alpha \\mid G (B)) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.628, + 0.892, + 0.658 + ], + "angle": 0, + "content": "In this way, the gradient of \\(B\\) in \\(t\\)th update can be obtained by applying chain rule to Eq. 6" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.666, + 0.891, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {B _ {t}} = \\nabla_ {w _ {t}, B _ {t}} ^ {2} L _ {\\text {v a l}} \\left(w _ {t}, \\alpha \\mid G (B _ {t})\\right) \\nabla_ {B _ {t}} G (B _ {t}) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.691, + 0.892, + 0.81 + ], + "angle": 0, + "content": "where \\( t \\) denotes the update times of \\( B \\). \\( w_{t} \\) represents the supernet weight when \\( B \\) has been updated \\( t \\) times. However, evaluating the gradient of \\( B_{t} \\) requires an expensive matrix-vector product in its second term. Inspired by darts [18], the complexity can be substantially reduced using the finite difference approximation. We use two training steps to approximate the gradient of \\( w \\), therefore, Eq. 7 can be rewritten as" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.816, + 0.891, + 0.849 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {B _ {t}} = \\nabla_ {B _ {t}} L _ {\\text {v a l}} \\left(w _ {t}, \\alpha \\mid G \\left(B _ {t}\\right)\\right) \\nabla_ {B _ {t}} G \\left(B _ {t}\\right) \\tag {8} \\\\ - \\nabla_ {B _ {t}} L _ {v a l} \\left(w _ {t - 1}, \\alpha \\mid G (B _ {t})\\right) \\nabla_ {B _ {t}} G (B _ {t}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the implementation, the distribution \\( B \\) is only updated every few iterations, so there is no excessive computational overhead. The iterative procedure is outlined in Appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "5922" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.095, + 0.868, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.368, + 0.895, + 0.415 + ], + "angle": 0, + "content": "Figure 3: An overview of architecture generator training process. AG is given a sampled resource constraint B and generates several one-hot policies sequence by sequence. Then, one-hot policies will be mapped to corresponding masks to keep differentiable." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.44, + 0.471, + 0.472 + ], + "angle": 0, + "content": "3.3. Generating Architectures with Arbitray Computational Resources" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.47, + 0.649 + ], + "angle": 0, + "content": "The probability shift is able to sample subnets according to their corresponding training sufficiency. However, a crucial question emerges: given a certain computational resource \\( b \\), how can we rapidly sample a subnet that satisfies the desired resource constraint? A straightforward approach would be to keep sampling subnets until one is found that meets the resource constraint. Nevertheless, as illustrated in Figure 2, subnets with large or small computational resources have an extremely low probability of being sampled, which makes sampling these subnets computationally inefficient." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.47, + 0.852 + ], + "angle": 0, + "content": "To address this issue, we propose an architecture generator (AG) that can provide the corresponding subnet architecture according to any resource constraint. The AG is designed to stably generate a subnet architecture that satisfies the desired resource constraint. Inspired by previous works [35], which formulate the NAS problem as a sequence prediction problem, we also use an LSTM network to generate each operation sequence by sequence. To update the sampling probability \\( B \\), we use differentiable neural architecture search (DNAS) [11, 18] to jointly train the AG and the supernet. DNAS is preferred over other search methods, such as RL-based or evolutionary-based methods, as it converges faster." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "This section presents an overview of the AG (Architecture Generator) training process with clarity and academic rigor. Figure 3 provides a visual representation of the AG" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.441, + 0.892, + 0.516 + ], + "angle": 0, + "content": "training process. During the training of AG and supernet, the sample distribution \\(B\\) remains static and is not updated. To enable end-to-end updates of AG, the policy of each operation is generated using Gumbel Softmax with a one-hot vector, such as \\([1,0,0]\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.893, + 0.653 + ], + "angle": 0, + "content": "It is important to note that the focus of this paper is on weight entanglement. This means that each operation shares weights for their common parts, and weights with small indices are always activated. As a result, policies cannot be directly involved in forward and backward processes. To overcome this issue, a matrix map trick is employed. More specifically, given an one-hot policy \\( p \\in 1 \\times n \\) where \\( n \\) denotes the candidate number in this operation, the mapped policy \\( p' \\) can be computed using the following" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.661, + 0.892, + 0.681 + ], + "angle": 0, + "content": "\\[\np ^ {\\prime} = M \\cdot p ^ {T} = \\left[ m _ {0} ^ {T}, m _ {1} ^ {T}, \\dots , m _ {n} ^ {T} \\right] \\cdot p ^ {T} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.824 + ], + "angle": 0, + "content": "where \\(M\\) consists of several masks and \\(m_i^T\\) represents the mask of \\(i\\) th candidate. We show the matrix map example in Figure 3. Assume a ViT block has three heads and the heads number can be selected in [1,2,3], their weight-entangle masks can be designed as \\(m^0 = [1,0,0], m^1 = [1,1,0]\\) and \\(m^2 = [1,1,1]\\), respectively. With Eq. 9, the one-hot policy \\(p\\) participates in the forward of the supernet while the gradient of AG is automatically calculated by the chain rule." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Since our goal is to make AG learn how to accurately generate a subnet architecture with arbitrary resource constraints, the objective function of AG is designed to reduce task loss \\( L_{task} \\) while minimizing the gap \\( L_{RC} \\) between the resource constraint of the sampled subnet and the target." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "5923" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.088, + 0.868, + 0.591 + ], + "angle": 0, + "content": "
ModelParameters(M)FLOPs(G)Cost(GPU Days)Top-1 Acc.(%)Top-5 Acc.(%)
DeiT-tiny[28]5.71.22472.291.1
AutoFormer-tiny[4]5.71.33074.792.3
FocusFormer-tiny[19]6.21.42675.193.1
ShiftFormer-T(Ours)5.81.32476.093.1
DeiT-S[28]22.14.73079.995.0
T2T-ViT-14[34]21.56.13281.7-
ViT-S/16[6]22.95.1-78.8-
BoTNet-S1-59[26]33.57.3-81.795.8
AutoFormer-small[4]22.95.13581.495.6
FocusFormer-small[19]23.75.03281.695.6
ShiftFormer-S(Ours)23.65.03082.295.8
DeiT-B[28]86.617.64381.895.6
ViT-B/16[6]86.617.6-79.7-
AutoFormer-base[4]52.811.04381.495.7
FocusFormer-base[19]52.811.04181.995.6
ShiftFormer-B(Ours)52.811.04082.896.1
MobileNetV2 0.75×[24]2.60.211869.8-
MobileNetV3 1.0×[9]5.40.22379175.2-
DS-MBNet-M[16]-0.332473.2-
BigNAS-S[33]4.50.2411276.5-
Once-For-All[3]4.40.2310576.4-
GreedyNAS-C[31]4.70.283276.292.5
ShiftCNN-S(Ours)4.50.243277.293.1
MobileNetV2 1.3×[24]5.30.502172.8-
MobileNetV3 1.25×[9]8.10.35379176.7-
EfficientNet-B0[27]5.40.39379177.193.3
DS-MBNet-S[16]-0.572474.8-
BigNAS-M[33]5.50.4211278.9-
GreedyNAS-A[31]6.50.374077.193.3
ShiftCNN-B(Ours)5.60.423279.693.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.604, + 0.895, + 0.651 + ], + "angle": 0, + "content": "Table 1: ShiftNAS models performance on ImageNet with comparisons to other models. We group the models according to their FLOPs. We use ShiftFormer and ShiftCNN to denote the models searched by ShiftNAS. Cost represents the total GPU days including training, searching and retraining." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.689, + 0.33, + 0.705 + ], + "angle": 0, + "content": "Therefore, the joint loss \\(L\\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.714, + 0.303, + 0.731 + ], + "angle": 0, + "content": "\\[\nL = L _ {\\text {t a s k}} + \\lambda L _ {R C}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.733, + 0.47, + 0.786 + ], + "angle": 0, + "content": "\\[\nL _ {R C} = \\left(\\sum_ {i = 1} ^ {D} \\sum_ {j = 1} ^ {n} b _ {j} ^ {i} p _ {j} ^ {\\prime i} - C\\right) ^ {2} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.473, + 0.903 + ], + "angle": 0, + "content": "where \\( b_{j}^{i} \\) and \\( p_{j}^{\\prime i} \\) represents the computational resource and policy of \\( j \\) th candidate in \\( i \\) th operation, \\( C \\) indicates the target resource constraint sampled from sampling distribution B. \\( \\lambda \\) is the coefficient of \\( L_{RC} \\). In the implementation, we jointly optimize the AG and the supernet on the train dataset with Eq. 10. After few epochs, AG can generate the corresponding subnets with given resource constraint." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.687, + 0.625, + 0.705 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.717, + 0.691, + 0.734 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.745, + 0.892, + 0.835 + ], + "angle": 0, + "content": "Dataset and evaluation metrics. We conduct main experiments on ImageNet [13] which contains about 1.2M images for training and \\(50\\mathrm{K}\\) images for validation. We use the Top-1 accuracy and the number of giga floating-point operations(GFLOPs) to measure the performance and efficiency of networks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Implementation details. ShiftNAS is a model-agnostic search method. Therefore, we search the models with various resource constraints on both CNN and ViT search space. We split the computational resource range into \\( K \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "5924" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.287, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.211, + 0.213, + 0.221 + ], + "angle": 0, + "content": "(a) ViT-tiny" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.089, + 0.495, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.211, + 0.428, + 0.221 + ], + "angle": 0, + "content": "(b) ViT-small" + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.089, + 0.699, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.211, + 0.64, + 0.221 + ], + "angle": 0, + "content": "(c) ViT-base" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.089, + 0.905, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.211, + 0.832, + 0.221 + ], + "angle": 0, + "content": "(d) CNN" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.235, + 0.892, + 0.266 + ], + "angle": 0, + "content": "Figure 4: The effectiveness of probability shift. FLOPs/Accuracy tradeoffs of ShiftNAS with and without probability shift on search space (a) ViT-tiny; (b) ViT-small; (c) ViT-base; (d) CNN." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.292, + 0.48, + 0.395 + ], + "angle": 0, + "content": "
ModelInherit(%)Finetune(%)Retrain(%)
ShiftFormer-T76.075.9(-0.1)76.1(+0.1)
ShiftFormer-S82.082.0(0.0)82.0(0.0)
ShiftFormer-B82.482.3(-0.1)82.6(+0.2)
ShiftCNN-S77.277.0(-0.2)77.3(+0.1)
ShiftCNN-B79.679.5(-0.1)79.6(0.0)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.409, + 0.468, + 0.44 + ], + "angle": 0, + "content": "Table 2: Comparison of subnets Top-1 accuracy with inherited, finetuned and retrained weights." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.47, + 0.715 + ], + "angle": 0, + "content": "parts with 0.1 GFLOPs step. The sampling distribution will be represented by a \\(1 \\times K\\) learnable vector. We randomly split \\(50\\mathrm{K}\\) images from ImageNet and use Adam optimizer with 1e-3 learning rate to update the sampling distribution vector. The AG consists of an LSTM cell with 64 hidden numbers and 4/10 fully connected layers for ViT/CNN supernet and will be updated jointly with the supernet. In the first 50 training epochs, we only optimize the supernet weight and the AG since the AG cannot generate the corresponding subnets at the beginning. After 50 epochs, the sampling distribution vector is updated per 100 iteration. We discuss more hyperparameter settings in the Appendix, such as the split step of the computational resource and the updated frequency of the sampling distribution vector. The training strategies for both CNN and ViT are given in Appendix. All the models are trained on 8 Nvidia Tesla A100 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.468, + 0.868 + ], + "angle": 0, + "content": "For CNN search space, we follow BigNAS [33] where the search space contains kernel size, channel number, depth and input resolution. The supernet will be split into 7 stages. Each stage has multiple choices of the block number and the first block of each stage has no residual path. The policy of kernel size will be generated block-wise, where the small kernel size is center cropped from the large kernel size. The policies of channel number and depth will be predicted stage-wise, where the lower-index channels and blocks are preferentially kept." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "For ViT search space, we follow AutoFormer [4] which searches embed dim, head number, MLP ratio and depth" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.293, + 0.892, + 0.338 + ], + "angle": 0, + "content": "with three different scale settings. The policies of head number and MLP ratio will be given block-wise. As CNN, the lower-index heads and neurons are preferentially kept." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.413 + ], + "angle": 0, + "content": "After training, we directly sample the candidate with the max probability given by the AG. It is noted that the obtained model inherits weights from the trained supernet without retraining or fine-tuning. Therefore, no search or retraining cost exists in ShiftNAS." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.423, + 0.638, + 0.437 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.892, + 0.809 + ], + "angle": 0, + "content": "We search ViT and CNN architectures on ImageNet with different FLOPs constraints. We compare searched models with multiple ViT and CNN models on model performance (Top-1 Acc., Top-5 Acc.) and efficiency (FLOPs, parameters). It is seen from Table 1 that our ShiftFormer models surpass the recent manual and autoML-based transformer models under diverse model sizes. It is worth noting that ShiftFormer-tiny achieves a top-1 Acc. of 76.0 with only 1.3 GFLOPs, being \\(1.3\\%\\) and \\(0.9\\%\\) better than AutoFormer-tiny and FocusFormer-T which are also autoML-based methods, respectively. Compared to CNN models, ShiftCNN models also achieve better accuracy than all compared networks under similar FLOPs restrictions. For example, ShiftCNN-S and ShiftCNN-B achieve \\(0.7\\%\\) and \\(0.7\\%\\) higher top-1 accuracy than BigNAS-S and BigNAS-M, respectively. Specifically, ShiftCNN and BigNAS models are searched from the same search space. BigNAS models spend more than \\(2\\times\\) GPU time than ShiftCNN models since the sandwich rule is used to train BigNAS supernet, which demonstrates the efficiency and outstanding performance of ShiftNAS. For fairness, we do not compare our models with others obtained from better search spaces [15, 29] in Table 1. The experiments on AttentiveNAS search space can be found in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.652, + 0.833 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The effectiveness of probability shift. To demonstrate the effectiveness of the proposed sampling probability shift, we trained a ShiftNAS supernet and its baseline counterpart under ViT-tiny, ViT-small, ViT-base, and CNN search" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.515, + 0.937 + ], + "angle": 0, + "content": "5925" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.283, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.089, + 0.49, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.089, + 0.697, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.089, + 0.905, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.23, + 0.892, + 0.26 + ], + "angle": 0, + "content": "Figure 5: The sampling probability with different FLOPs groups on search space (a) ViT-tiny; (b) ViT-small; (c) ViT-base; (d) CNN." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.284, + 0.441, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.491, + 0.468, + 0.506 + ], + "angle": 0, + "content": "Figure 6: The ranking correlation of architecture generator." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.468, + 0.672 + ], + "angle": 0, + "content": "spaces. In the baseline supernet training, we uniformly sampled FLOPs in each step and did not update FLOPs probability. The training settings for all supernets are provided in the Appendix. As illustrated in Figure 4, the supernet trained with probability shift is capable of covering a wide range of accuracy-FLOPs curves and outperforms the baseline supernet trained without probability shift by a significant margin, demonstrating the effectiveness of the proposed sampling probability shift mechanism." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.468, + 0.869 + ], + "angle": 0, + "content": "The ranking correlation of architecture generator. The experiment is based on ShiftFormer-tiny supernet. We measure the ranking correlation by randomly sampling 25 subnets from the ShiftFormer-tiny supernet. For fast evaluation, The sampled subnets are inherited weights from the supernet and are further finetuned to reach the optimal weights. As for the model ranking test, we evaluate sampled subnets in both supernets to get their inherited performance. As shown in Figure 6, the inherited performance is highly correlated with finetuned performance in good subnets (Inherited Top1-Acc > 65%). Despite the correlation with bad subnets is not obvious, we focus more attention on good subnets since bad subnets are not deployed in practice." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Subnet performance without retraining. We further study the performance of ShiftNAS models when models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.287, + 0.892, + 0.406 + ], + "angle": 0, + "content": "are independently fine-tuned or trained from scratch. Following Autoformer, we finetune or retrain the selected subnet with 40 epochs or 300 epochs on ImageNet. Table 2 shows that finetuning cannot improve or even harm the performance of subnets inherited from supernet. Besides, retraining can slightly improve the subnet performance but cost extra training resources which prevents the implementation of NAS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.59 + ], + "angle": 0, + "content": "The optimal sampling strategy. ShiftNAS learns the optimal sampling probability for each subnet with different FLOPs. To figure out what sampling strategy is the best, we visualize the various sampling probability in different search spaces (ViT-tiny, ViT-small, ViT-base and CNN) at different training steps. As can be seen in Figure 5, our method learns to sample more subnets with large computational complexity. For example, the sampling probabilities of 1.6 to 1.9 GFLOPs subnets are clearly higher than 1.3 and 1.4 GFLOPs subnets in the ViT-tiny search space. This is intuitive since the large subnet contains more parameters and needs more computational resources to update." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.619, + 0.619 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.795 + ], + "angle": 0, + "content": "In this paper, we presented ShiftNAS, a one-shot supernet training framework that can complexity-wise adjust the sampling probability. To automatically adjust the sampling probability, we proposed probability shift that can be learned according to the subnets training sufficiency. To fast obtain the subnets with desired complexity, AG is designed that can be jointly trained with the supernet in an end-to-end manner. Experiments on ImageNet showed that our method achieves SOTA results on ViT/CNN search spaces, and covers a wide range of efficiency/accuracy trade-offs without any extra retraining." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.841 + ], + "angle": 0, + "content": "Limitation. ShiftNAS is designed for efficient search on weight-entanglement search spaces, so it cannot work on darts-like search spaces [18, 25]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements: This work was supported by the National Natural Science Foundation of China (62203392) and Natural Science Foundation of Zhejiang province (LY21F030018)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "5926" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Gabriel Bender, Pieter-Jan Kindermans, Barret Zoph, Vijay Vasudevan, and Quoc Le. Understanding and simplifying one-shot architecture search. In International conference on machine learning, pages 550-559. PMLR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.214 + ], + "angle": 0, + "content": "[2] Han Cai, Chuang Gan, Ji Lin, and Song Han. Network augmentation for tiny deep learning. arXiv preprint arXiv:2110.08890, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[3] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. In International Conference on Learning Representations, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[4] Minghao Chen, Houwen Peng, Jianlong Fu, and Haibin Ling. Autoformer: Searching transformers for visual recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12270-12280, 2021. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[5] Xiangxiang Chu, Bo Zhang, Ruijun Xu, and Jixiang Li. Fairnas: Rethinking evaluation fairness of weight sharing neural architecture search. arXiv preprint arXiv:1907.01845, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.403, + 0.471, + 0.485 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.471, + 0.557 + ], + "angle": 0, + "content": "[7] Zichao Guo, Xiangyu Zhang, Haoyuan Mu, Wen Heng, Zechun Liu, Yichen Wei, and Jian Sun. Single path one-shot neural architecture search with uniform sampling. In European Conference on Computer Vision, pages 544-560. Springer, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.471, + 0.614 + ], + "angle": 0, + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.471, + 0.698 + ], + "angle": 0, + "content": "[9] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.701, + 0.471, + 0.77 + ], + "angle": 0, + "content": "[10] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.772, + 0.471, + 0.841 + ], + "angle": 0, + "content": "[11] Shoukang Hu, Sirui Xie, Hehui Zheng, Chunxiao Liu, Jianping Shi, Xunying Liu, and Dahua Lin. Dsnas: Direct neural architecture search without parameter retraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12084-12092, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.843, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[12] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[13] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural net" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "works. Advances in neural information processing systems, 25:1097-1105, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[14] Changlin Li, Jiefeng Peng, Liuchun Yuan, Guangrun Wang, Xiaodan Liang, Liang Lin, and Xiaojun Chang. Blockwisely supervised neural architecture search with knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1989-1998, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.894, + 0.289 + ], + "angle": 0, + "content": "[15] Changlin Li, Tao Tang, Guangrun Wang, Jiefeng Peng, Bing Wang, Xiaodan Liang, and Xiaojun Chang. Bossnas: Exploring hybrid cnn-transformers with block-wisely self-supervised neural architecture search. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12281-12291, 2021. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.894, + 0.36 + ], + "angle": 0, + "content": "[16] Changlin Li, Guangrun Wang, Bing Wang, Xiaodan Liang, Zhihui Li, and Xiaojun Chang. Ds-net++: Dynamic weight slicing for efficient inference in cnns and vision transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-16, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.894, + 0.403 + ], + "angle": 0, + "content": "[17] Liam Li and Ameet Talwalkar. Random search and reproducibility for neural architecture search. In Uncertainty in artificial intelligence, pages 367-377. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[18] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. In International Conference on Learning Representations, 2018. 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[19] Jing Liu, Jianfei Cai, and Bohan Zhuang. Focusformer: Focusing on what we need via architecture sampler. arXiv preprint arXiv:2208.10861, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[21] James E Marengo, David L Farnsworth, and Lucas Stefanic. A geometric derivation of the irwin-hall distribution. International Journal of Mathematics and Mathematical Sciences, 2017, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[22] Yameng Peng, Andy Song, Vic Ciesielski, Haytham M. Fayek, and Xiaojun Chang. Pre-nas: Evolutionary neural architecture search with predictor. IEEE Transactions on Evolutionary Computation, 27(1):26-36, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[23] Hieu Pham, Melody Guan, Barret Zoph, Quoc Le, and Jeff Dean. Efficient neural architecture search via parameters sharing. In International conference on machine learning, pages 4095-4104. PMLR, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[24] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[25] Julien Siems, Lucas Zimmer, Arber Zela, Jovita Lukasik, Margret Keuper, and Frank Hutter. Nas-bench-301 and the case for surrogate benchmarks for neural architecture search. arXiv preprint arXiv:2008.09777, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[26] Aravind Srinivas, Tsung-Yi Lin, Niki Parmar, Jonathon Shlens, Pieter Abbeel, and Ashish Vaswani. Bottleneck transformers for visual recognition. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "5927" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 16519-16529, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[27] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[28] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.317 + ], + "angle": 0, + "content": "[29] Dilin Wang, Meng Li, Chengyue Gong, and Vikas Chandra. Attentivenas: Improving neural architecture search via attentive sampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6418-6427, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[30] Bichen Wu, Xiaoliang Dai, Peizhao Zhang, Yanghan Wang, Fei Sun, Yiming Wu, Yuandong Tian, Peter Vajda, Yangqing Jia, and Kurt Keutzer. Fbnet: Hardware-aware efficient convnet design via differentiable neural architecture search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10734-10742, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.469, + 0.471 + ], + "angle": 0, + "content": "[31] Shan You, Tao Huang, Mingmin Yang, Fei Wang, Chen Qian, and Changshui Zhang. Greedynas: Towards fast one-shot nas with greedy supernet. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1999-2008, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.474, + 0.469, + 0.514 + ], + "angle": 0, + "content": "[32] Jiahui Yu and Thomas Huang. Autoslim: Towards one-shot architecture search for channel numbers. arXiv preprint arXiv:1903.11728, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.469, + 0.599 + ], + "angle": 0, + "content": "[33] Jiahui Yu, Pengchong Jin, Hanxiao Liu, Gabriel Bender, Pieter-Jan Kindermans, Mingxing Tan, Thomas Huang, Xiaodan Song, Ruoming Pang, and Quoc Le. Bignas: Scaling up neural architecture search with big single-stage models. In European Conference on Computer Vision, pages 702-717. Springer, 2020. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.469, + 0.683 + ], + "angle": 0, + "content": "[34] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF international conference on computer vision, pages 558-567, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.685, + 0.469, + 0.725 + ], + "angle": 0, + "content": "[35] Barret Zoph and Quoc V Le. Neural architecture search with reinforcement learning. arXiv preprint arXiv:1611.01578, 2016. 1, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "5928" + } + ] +] \ No newline at end of file diff --git a/2023/ShiftNAS_ Improving One-shot NAS via Probability Shift/9d0cd6fb-8343-4e85-b19e-ff450eb2e5e4_origin.pdf b/2023/ShiftNAS_ Improving One-shot NAS via Probability Shift/9d0cd6fb-8343-4e85-b19e-ff450eb2e5e4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..630643b3c32b6faa171c1787bc58597214f14eee --- /dev/null +++ b/2023/ShiftNAS_ Improving One-shot NAS via Probability Shift/9d0cd6fb-8343-4e85-b19e-ff450eb2e5e4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3c5b47737327d035a32d6b4095e97f2ee09384acba262ab90c5cce0fb20c580 +size 3399272 diff --git a/2023/ShiftNAS_ Improving One-shot NAS via Probability Shift/full.md b/2023/ShiftNAS_ Improving One-shot NAS via Probability Shift/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9f3ed294e93867da607688d59a304a71ff8ddfc7 --- /dev/null +++ b/2023/ShiftNAS_ Improving One-shot NAS via Probability Shift/full.md @@ -0,0 +1,313 @@ +# ShiftNAS: Improving One-shot NAS via Probability Shift + +Mingyang Zhang, Xinyi Yu* Haodong Zhao, Linlin Ou* Zhejiang University of Technology Hangzhou, Zhejiang, China + +yuxy@zjut.edu.cn, linlinou@zjut.edu.cn + +# Abstract + +One-shot Neural architecture search (One-shot NAS) has been proposed as a time-efficient approach to obtain optimal subnet architectures and weights under different complexity cases by training only once. However, the subnet performance obtained by weight sharing is often inferior to the performance achieved by retraining. In this paper, we investigate the performance gap and attribute it to the use of uniform sampling, which is a common approach in supernet training. Uniform sampling concentrates training resources on subnets with intermediate computational resources, which are sampled with high probability. However, subnets with different complexity regions require different optimal training strategies for optimal performance. + +To address the problem of uniform sampling, we propose ShiftNAS, a method that can adjust the sampling probability based on the complexity of subnets. We achieve this by evaluating the performance variation of subnets with different complexity and designing an architecture generator that can accurately and efficiently provide subnets with the desired complexity. Both the sampling probability and the architecture generator can be trained end-to-end in a gradient-based manner. With ShiftNAS, we can directly obtain the optimal model architecture and parameters for a given computational complexity. We evaluate our approach on multiple visual network models, including convolutional neural networks (CNNs) and vision transformers (ViTs), and demonstrate that ShiftNAS is model-agnostic. Experimental results on ImageNet show that ShiftNAS can improve the performance of one-shot NAS without additional consumption. Source codes are available at GitHub. + +# 1. Introduction + +Deep neural networks (DNNs) have been widely applied to the field of computer vision with remarkable success [6, 8]. However, the deployment of these vision net + +![](images/9e326b61809dda24f08e4295dbc212bcc6ebd11bb228479e249d6a1220227467.jpg) +Figure 1: A conceptual overview of our ShiftNAS. Focusing on the computational resource (e.g. FLOPs), ShiftNAS split a search space into several parts where subnets have close computational complexity. In stochastic supernet training, ShiftNAS first samples computational complexity according to probability, then an architecture generator samples the subnet with desired computational complexity. The probability distribution is dynamically updated with stochastic training by estimating the performance variation of each sub-space. + +works on edge devices still has some limitations, such as the massive model sizes and excessive computation overhead [9, 10, 20, 24]. In addition, designing architectures artificially in a trial-and-error manner is a resource-consuming task that requires not only architectural skills but also domain expertise. Consequently, how to acquire optimal architectures that balance latency and accuracy efficiently is of paramount importance. + +Recent advancements in neural architecture search (NAS) methods, such as [22, 7, 30, 35], have led to significant improvements in the performance of practical applications by automatically searching for optimal architectures within a defined search space. However, traditional NAS methods typically require a substantial computation budget [23, 35]. In order to speed up training and reduce the resource consumption of the training process, one-shot NAS + +methods [18, 7] have adopted a two-stage training approach based on weight sharing. Specifically, a supernet is trained in the first stage, and subnets with better performance are searched for in the second stage. In some cases, post-processing methods (e.g., retraining and finetuning) are also necessary in the second stage since the performance of these subnets inherited from the supernet is often inferior to that of models trained from scratch. However, the downside of such methods is that the training consumption increases linearly with the number of architectures, which can be problematic. To address this issue, some one-shot NAS methods [3, 4, 29, 33] have utilized a weight entanglement training strategy to share the weights in each operation, eliminating the need for additional finetuning or retraining. Furthermore, a high-quality supernet is essential for candidate architectures to inherit weights directly [33], as well as for accurately ranking candidate architectures [14]. + +However, there is still potential for breakthroughs in training a better supernet. In the first stage, previous methods [7, 4] assume that all candidate architectures are equally important and should be sampled with equal probability during training. However, subnets with different numbers of parameters require different amounts of training resources [2]. For instance, subnets with 1.0 GFLOPs may converge after 30000 iterations, while subnets with 2.0 GFLOPs may require 50000 iterations. Additionally, only the optimal subnet will be deployed, while others will be ignored at the same computational complexity. Therefore, subnets that occupy more training resources may not be distributed in regions corresponding to resource constraints, resulting in sub-optimal performance of the final deployed model. We find that when all subnets are sampled with equal probability, the resulting computational resource distribution is approximately normal. Consequently, subnets trained under this distribution may appear to be under-fitting or over-fitting in different regions. + +To address the challenge of efficiently training a high-performance supernet, we propose a novel method, probability-Shift Neural Architecture search (ShiftNAS). In ShiftNAS, the sampling probability of subnets is not uniform and can be dynamically adjusted during the training process. The training sufficiency of each subnet is measured by evaluating its performance variation under different computational constraints. The subnets with high-performance variance are identified as undertrained, and their sampling probabilities are increased to tilt training resources dynamically towards them. This enables us to allocate resources more effectively and efficiently to achieve better performance for the subnets that need more training resources. + +In spite of having an optimal sampling distribution, efficiently and accurately sampling subnets with a desired computational constraint still poses a challenge. To address this + +issue, we propose an LSTM-based architecture generator (AG) that can be optimized differentiably with a resource constraint loss function. The AG's output is then processed by Gumbel Softmax [12] to generate a one-hot vector policy for each searched operation. To suit the weight-entangled search space [32, 4], we employ a matrix mapping technique that can convert the one-hot vector into a differentiable mask. The mask is multiplied by the operation to obtain a differentiable subnet. The AG and supernet can be jointly trained to learn how to generate the best subnets with desired computational constraints. During evaluation, the AG can generate a corresponding subnet immediately for any given computational constraint. The weights of the searched subnet can be directly inherited from the well-trained supernet, making ShiftNAS free from any additional search or retrain costs. + +The overall contribution can be summarized as follows: + +- A learnable sampling strategy, called probability shift, is proposed to relief the bias of uniform sampling which leads to performance gap between supernet training and subnet deployment. +- We propose an LSTM-based AG to precisely and efficiently offer the best subnet with desired resource constraints. AG training can be differentiably trained with supernet under weight-entangled search space. +- We achieve state-of-the-art or competitive results on both CNN and ViT models. Therefore, ShiftNAS is a model-agnostic search method. + +# 2. Related Work + +The one-shot NAS method [3, 29, 33] can automatically search for the optimal architecture in a predefined search space [4, 9, 18], which can be formulated by maximizing an expected accuracy over the space $A$ , i.e., + +$$ +\begin{array}{l} \min _ {\alpha \in A} L _ {v a l} \left(w ^ {*} \mid \alpha\right) \\ \text {s . t .} w ^ {*} = \underset {w _ {\alpha}} {\arg \min } L _ {\text {t r a i n}} (w \mid \alpha) \tag {1} \\ \end{array} +$$ + +$$ +c o m p l e x i t y (\alpha) < \tau +$$ + +where $w$ and $\alpha$ represent the weight and architecture of subnets. $L_{val}$ and $L_{train}$ are loss functions in the validation and train dataset, respectively. The computational complexity of $\alpha$ can be calculated by the complexity function. $\tau$ denotes the complexity threshold for alpha. To solve this optimization problem, a two-stage approach based on weight sharing is usually employed. + +In the first stage, a high-quality supernet is trained by sampling a large number of subnets in the defined search space. Notably, in order to obtain subnets that need to be incorporated into the training phase, the following approaches are commonly used in recent one-shot NAS works [4, 7]. + +First is the uniform sampling method, which is also the most common strategy. The Uniform sample method considers that all architectures in the supernet are equally important, that is, they should be sampled with exactly equal probability. The weight optimization equation in the first stage can be formulated as: + +$$ +\min _ {w} E _ {\alpha \sim A} \left[ L _ {\text {t r a i n}} (w \mid \alpha) \right] +$$ + +$$ +\alpha = o ^ {1} \cup o ^ {2} \cup \dots \cup o ^ {D} \tag {2} +$$ + +$$ +o ^ {i} \sim U (0, n _ {i}) +$$ + +where $D, n$ denotes the total operation and candidate number in certain operations, respectively. $o^i$ denotes the $i$ th selected operation, which follows a uniform distribution. + +Since there is a gap between the training process and practical deployment, AttentiveNAS [29] proposes the Pareto-aware sampling method. Specifically, models deployed on edge devices are generally around the Pareto frontier, but they are uniformly sampled during training, which leads to the waste of training resources to a certain extent. To bridge the gap between training and deploying, each iteration extracts multiple subnets with equal size and selects the best or worst architecture for training. + +Similar to AttentiveNAS, GreedyNAS [31] screen the weak subnets and just sample from the potentially-good subnets instead of all subnets, thus capturing another opportunity to improve the accuracy of target models. + +However, these uniformly sampled architectures appear to be an approximately normal distribution of the computational cost. To further reduce the gap between training and deployment, Focusformer [19] proposes that all the sampled subnets should be based on resource distribution and focuses more on the Pareto frontier architectures through an architecture sampler. In addition, FairNAS [5] also reinforces the sampling process with a stricter standard of fairness, that is, the parameters of each choice block are updated the same number of times at any stage. + +When the first stage is over, this well-trained supernet can be used as the performance estimator of the candidate architecture. And the weight of candidate architectures can be directly inherited from the supernet. + +In the second stage, we need to search for the best candidate architecture under different resource constraints in the supernet, which can be formulated as + +$$ +\alpha = \underset {\alpha \in A} {\operatorname {a r g m i n}} L _ {v a l} (w \mid \alpha) \tag {3} +$$ + +$$ +s. t. c o m p l e x i t y (\alpha) < \tau +$$ + +Since there are so many architectures in the supernet that need to be evaluated, the consumption of computing resources is also unacceptable. Therefore, recent works resort to random search [17], evolution algorithms [4, 7, 23] or reinforcement learning [35] to find the most promising + +architecture among all architectures. In the end, these selected architectures need some post-processing operations such as retraining to improve the architecture performance, but this also brings additional computational consumption. + +In order to alleviate the above-mentioned resource consumption problem, some one-shot NAS methods train a high-quality supernet so that candidate architectures can directly inherit supernet weights without retraining, such as Autoformer [4], attentiveNAS [29], BigNAS [33], FocusFormer [19]. + +# 3. Nerual Architecture Search with Probability Shift + +In this section, we first discuss the training resources allocation problem caused by uniform sampling in one-shot NAS. To solve the above-mentioned problem, we present ShiftNAS, an end-to-end supernet training framework. In ShiftNAS, a learnable sampling strategy is proposed, which can dynamically adjust the training resources allocation by shifting the sampling probability. Then, an LSTM-based AG is designed to accurately obtain the expected subnet for each sampling. We show the overview of our method in Figure 1. + +# 3.1. Rethinking the Sampling of One-shot NAS + +In one-shot NAS, an overparameterized supernet $S$ is formed with multiple operations, where each operation contains several choices. The supernet $S$ contains $D$ operations, and the $d$ th operation $o^d$ can be selected from $n$ candidates $o_1, \ldots, o_n$ , which represent various design choices, such as kernel size and channels in CNN search space [33], or heads number and MLP ratio in ViT search space [4]. A subnet $a$ sampled from the supernet can be represented by a tuple of size $D$ , i.e., $a = (o^1, \ldots, o^D)$ . + +In previous one-shot NAS methods, the sampling probability of each operation candidate $o_i$ is given by $p(o_i) = \frac{1}{n}$ , which assumes that all candidates are equally important. However, in reality, one-shot NAS is more concerned with the computational resources of the subnets. To address this, we introduce the notion of computational resource for each operation candidate $o^d$ , denoted by $b^d \in b_1, \ldots, b_n$ . The computational resource of a subnet $a$ randomly sampled from the supernet can then be computed as $B_a = \sum_{d=1}^{D} b^d$ , where $b^d$ follows a uniform distribution. + +Remark 1 If we assume that the computational resource of each operation $b^{d}$ is independently sampled from a uniform distribution, then the total computational resource $B_{a}$ of a subnet $a$ sampled from the supernet follows an Irwin-Hall distribution. As the number of operations $D$ in the supernet increases, the Irwin-Hall distribution converges to a normal distribution [21]. + +![](images/5060003a72932d560ab426fa4fed7db9cf51f1a7cb2aa193565dd52081eddf54.jpg) + +![](images/f3adc5a47f36faa731ecb92ce7adf7f26dd5a41c0966ffe2a99d46c84b89e85d.jpg) + +![](images/ad74b05bf2dcf2d2b8462c9e21c21b1eb1f2b51c2682900b27c7b804b5af1b7f.jpg) +(a) Uniform distribution of each operation + +![](images/17c40fd0c370ca01fcb2cb3e69be4e0588b60ac8784c68f29c6164208feb8dbc.jpg) + +![](images/1d1366ad1a27c4a1479c1679c671b708734673d3612c6981c9738afad63f2250.jpg) +(b) Normal distribution of subnets computational resources +Figure 2: During supernet training, operations are sampled from a uniform distribution in each iteration. The subnet consists of the sampled operations and its computational resource follows an approximated normal distribution. + +Upon analyzing Remark 1, it is evident that the uniform sampling strategy results in subnets being sampled with moderate computational resources, as depicted in Figure 2. This observation implies that subnets with moderate computational resources can be trained effectively, whereas subnets with large or small computational resources cannot be fully trained. Insufficiently-trained subnets lead to inaccurate ranking [14, 15] and inherit unreasonable weights from the supernet without retraining [1]. + +# 3.2. Shifting Probability with Greedy Sample + +One-shot NAS methods rely on the assumption that the supernet can effectively rank the performance of subnets with arbitrary computational resources, and therefore require each subnet to be fully trained with optimal training strategies. However, in practice, subnets with differ + +ent computational resources may require different training strategies, and the uniform sampling strategy used in previous one-shot NAS methods may lead to insufficient training of subnets in certain computational regions. For example, subnets with extreme computational resources may not be fully trained even when subnets with moderate computational resources are fully trained. To address this issue, we propose a learnable greedy sampling strategy that takes into account the training sufficiency of subnets in different computational regions. + +For any subnet, we judge whether the subnet has converged by calculating the following gradient of the subnet. + +$$ +\nabla_ {w} = \nabla_ {w} L _ {\text {v a l}} (w, \alpha \mid b) \tag {4} +$$ + +where $L_{val}$ represents the task loss (e.g. cross entropy) in the validation dataset, $w, \alpha$ and $b$ denote the weight, architecture and sampled computational resource of the subnet, respectively. When the subnet with $b$ computational resource converges, $\nabla_w$ will tend to zero. In other words, insufficiency-trained subnets have a small gradient (probably negative) and should be sampled more times. Therefore, we can greedily adjust the sampling distribution of $B$ by optimizing the computational resource distribution at each training step + +$$ +\underset {B} {\arg \min } E _ {b \sim B} \left[ \nabla_ {w} L _ {v a l} (w, \alpha \mid b) \right] \tag {5} +$$ + +To optimize $B$ end to end based on gradient-based methods, we discrete computational resources into several parts, e.g. $[1.3G, 1.4G, \dots, 2.0G]$ FLOPs, and uniformly initialize its probability distribution. In each training step, one computational resource can be sampled by Gumbel Softmax $G$ . Formally, Eq. 5 can be rewritten as + +$$ +\underset {B} {\arg \min } \nabla_ {w} L _ {v a l} (w, \alpha \mid G (B)) \tag {6} +$$ + +In this way, the gradient of $B$ in $t$ th update can be obtained by applying chain rule to Eq. 6 + +$$ +\nabla_ {B _ {t}} = \nabla_ {w _ {t}, B _ {t}} ^ {2} L _ {\text {v a l}} \left(w _ {t}, \alpha \mid G (B _ {t})\right) \nabla_ {B _ {t}} G (B _ {t}) \tag {7} +$$ + +where $t$ denotes the update times of $B$ . $w_{t}$ represents the supernet weight when $B$ has been updated $t$ times. However, evaluating the gradient of $B_{t}$ requires an expensive matrix-vector product in its second term. Inspired by darts [18], the complexity can be substantially reduced using the finite difference approximation. We use two training steps to approximate the gradient of $w$ , therefore, Eq. 7 can be rewritten as + +$$ +\begin{array}{l} \nabla_ {B _ {t}} = \nabla_ {B _ {t}} L _ {\text {v a l}} \left(w _ {t}, \alpha \mid G \left(B _ {t}\right)\right) \nabla_ {B _ {t}} G \left(B _ {t}\right) \tag {8} \\ - \nabla_ {B _ {t}} L _ {v a l} \left(w _ {t - 1}, \alpha \mid G (B _ {t})\right) \nabla_ {B _ {t}} G (B _ {t}) \\ \end{array} +$$ + +In the implementation, the distribution $B$ is only updated every few iterations, so there is no excessive computational overhead. The iterative procedure is outlined in Appendix. + +![](images/dab46cc2834897b0ae8093d2cb84654c0cd98c1b055fac3d9008e73b5bb398ee.jpg) +Figure 3: An overview of architecture generator training process. AG is given a sampled resource constraint B and generates several one-hot policies sequence by sequence. Then, one-hot policies will be mapped to corresponding masks to keep differentiable. + +# 3.3. Generating Architectures with Arbitray Computational Resources + +The probability shift is able to sample subnets according to their corresponding training sufficiency. However, a crucial question emerges: given a certain computational resource $b$ , how can we rapidly sample a subnet that satisfies the desired resource constraint? A straightforward approach would be to keep sampling subnets until one is found that meets the resource constraint. Nevertheless, as illustrated in Figure 2, subnets with large or small computational resources have an extremely low probability of being sampled, which makes sampling these subnets computationally inefficient. + +To address this issue, we propose an architecture generator (AG) that can provide the corresponding subnet architecture according to any resource constraint. The AG is designed to stably generate a subnet architecture that satisfies the desired resource constraint. Inspired by previous works [35], which formulate the NAS problem as a sequence prediction problem, we also use an LSTM network to generate each operation sequence by sequence. To update the sampling probability $B$ , we use differentiable neural architecture search (DNAS) [11, 18] to jointly train the AG and the supernet. DNAS is preferred over other search methods, such as RL-based or evolutionary-based methods, as it converges faster. + +This section presents an overview of the AG (Architecture Generator) training process with clarity and academic rigor. Figure 3 provides a visual representation of the AG + +training process. During the training of AG and supernet, the sample distribution $B$ remains static and is not updated. To enable end-to-end updates of AG, the policy of each operation is generated using Gumbel Softmax with a one-hot vector, such as $[1,0,0]$ . + +It is important to note that the focus of this paper is on weight entanglement. This means that each operation shares weights for their common parts, and weights with small indices are always activated. As a result, policies cannot be directly involved in forward and backward processes. To overcome this issue, a matrix map trick is employed. More specifically, given an one-hot policy $p \in 1 \times n$ where $n$ denotes the candidate number in this operation, the mapped policy $p'$ can be computed using the following + +$$ +p ^ {\prime} = M \cdot p ^ {T} = \left[ m _ {0} ^ {T}, m _ {1} ^ {T}, \dots , m _ {n} ^ {T} \right] \cdot p ^ {T} \tag {9} +$$ + +where $M$ consists of several masks and $m_i^T$ represents the mask of $i$ th candidate. We show the matrix map example in Figure 3. Assume a ViT block has three heads and the heads number can be selected in [1,2,3], their weight-entangle masks can be designed as $m^0 = [1,0,0], m^1 = [1,1,0]$ and $m^2 = [1,1,1]$ , respectively. With Eq. 9, the one-hot policy $p$ participates in the forward of the supernet while the gradient of AG is automatically calculated by the chain rule. + +Since our goal is to make AG learn how to accurately generate a subnet architecture with arbitrary resource constraints, the objective function of AG is designed to reduce task loss $L_{task}$ while minimizing the gap $L_{RC}$ between the resource constraint of the sampled subnet and the target. + +
ModelParameters(M)FLOPs(G)Cost(GPU Days)Top-1 Acc.(%)Top-5 Acc.(%)
DeiT-tiny[28]5.71.22472.291.1
AutoFormer-tiny[4]5.71.33074.792.3
FocusFormer-tiny[19]6.21.42675.193.1
ShiftFormer-T(Ours)5.81.32476.093.1
DeiT-S[28]22.14.73079.995.0
T2T-ViT-14[34]21.56.13281.7-
ViT-S/16[6]22.95.1-78.8-
BoTNet-S1-59[26]33.57.3-81.795.8
AutoFormer-small[4]22.95.13581.495.6
FocusFormer-small[19]23.75.03281.695.6
ShiftFormer-S(Ours)23.65.03082.295.8
DeiT-B[28]86.617.64381.895.6
ViT-B/16[6]86.617.6-79.7-
AutoFormer-base[4]52.811.04381.495.7
FocusFormer-base[19]52.811.04181.995.6
ShiftFormer-B(Ours)52.811.04082.896.1
MobileNetV2 0.75×[24]2.60.211869.8-
MobileNetV3 1.0×[9]5.40.22379175.2-
DS-MBNet-M[16]-0.332473.2-
BigNAS-S[33]4.50.2411276.5-
Once-For-All[3]4.40.2310576.4-
GreedyNAS-C[31]4.70.283276.292.5
ShiftCNN-S(Ours)4.50.243277.293.1
MobileNetV2 1.3×[24]5.30.502172.8-
MobileNetV3 1.25×[9]8.10.35379176.7-
EfficientNet-B0[27]5.40.39379177.193.3
DS-MBNet-S[16]-0.572474.8-
BigNAS-M[33]5.50.4211278.9-
GreedyNAS-A[31]6.50.374077.193.3
ShiftCNN-B(Ours)5.60.423279.693.6
+ +Table 1: ShiftNAS models performance on ImageNet with comparisons to other models. We group the models according to their FLOPs. We use ShiftFormer and ShiftCNN to denote the models searched by ShiftNAS. Cost represents the total GPU days including training, searching and retraining. + +Therefore, the joint loss $L$ is given by + +$$ +L = L _ {\text {t a s k}} + \lambda L _ {R C} +$$ + +$$ +L _ {R C} = \left(\sum_ {i = 1} ^ {D} \sum_ {j = 1} ^ {n} b _ {j} ^ {i} p _ {j} ^ {\prime i} - C\right) ^ {2} \tag {10} +$$ + +where $b_{j}^{i}$ and $p_{j}^{\prime i}$ represents the computational resource and policy of $j$ th candidate in $i$ th operation, $C$ indicates the target resource constraint sampled from sampling distribution B. $\lambda$ is the coefficient of $L_{RC}$ . In the implementation, we jointly optimize the AG and the supernet on the train dataset with Eq. 10. After few epochs, AG can generate the corresponding subnets with given resource constraint. + +# 4. Experiment + +# 4.1. Experimental Setup + +Dataset and evaluation metrics. We conduct main experiments on ImageNet [13] which contains about 1.2M images for training and $50\mathrm{K}$ images for validation. We use the Top-1 accuracy and the number of giga floating-point operations(GFLOPs) to measure the performance and efficiency of networks. + +Implementation details. ShiftNAS is a model-agnostic search method. Therefore, we search the models with various resource constraints on both CNN and ViT search space. We split the computational resource range into $K$ + +![](images/b401a713a63d725250a45f8a463b5e734c3615e2bdcdf814783740b04aac9a72.jpg) +(a) ViT-tiny + +![](images/07fb239bd32efec950164da70ba4c0e187ded04c6becaad9d38ca2858f023a51.jpg) +(b) ViT-small +Figure 4: The effectiveness of probability shift. FLOPs/Accuracy tradeoffs of ShiftNAS with and without probability shift on search space (a) ViT-tiny; (b) ViT-small; (c) ViT-base; (d) CNN. + +![](images/b653b1a2a2b831ab127af11c88c669f6a8a61b4f89a30e5c42da785b9cf183b2.jpg) +(c) ViT-base + +![](images/6a503801ba6bd3e394c0218737a52fafacef978a2119ae084f9d7038bf697042.jpg) +(d) CNN + +
ModelInherit(%)Finetune(%)Retrain(%)
ShiftFormer-T76.075.9(-0.1)76.1(+0.1)
ShiftFormer-S82.082.0(0.0)82.0(0.0)
ShiftFormer-B82.482.3(-0.1)82.6(+0.2)
ShiftCNN-S77.277.0(-0.2)77.3(+0.1)
ShiftCNN-B79.679.5(-0.1)79.6(0.0)
+ +Table 2: Comparison of subnets Top-1 accuracy with inherited, finetuned and retrained weights. + +parts with 0.1 GFLOPs step. The sampling distribution will be represented by a $1 \times K$ learnable vector. We randomly split $50\mathrm{K}$ images from ImageNet and use Adam optimizer with 1e-3 learning rate to update the sampling distribution vector. The AG consists of an LSTM cell with 64 hidden numbers and 4/10 fully connected layers for ViT/CNN supernet and will be updated jointly with the supernet. In the first 50 training epochs, we only optimize the supernet weight and the AG since the AG cannot generate the corresponding subnets at the beginning. After 50 epochs, the sampling distribution vector is updated per 100 iteration. We discuss more hyperparameter settings in the Appendix, such as the split step of the computational resource and the updated frequency of the sampling distribution vector. The training strategies for both CNN and ViT are given in Appendix. All the models are trained on 8 Nvidia Tesla A100 GPUs. + +For CNN search space, we follow BigNAS [33] where the search space contains kernel size, channel number, depth and input resolution. The supernet will be split into 7 stages. Each stage has multiple choices of the block number and the first block of each stage has no residual path. The policy of kernel size will be generated block-wise, where the small kernel size is center cropped from the large kernel size. The policies of channel number and depth will be predicted stage-wise, where the lower-index channels and blocks are preferentially kept. + +For ViT search space, we follow AutoFormer [4] which searches embed dim, head number, MLP ratio and depth + +with three different scale settings. The policies of head number and MLP ratio will be given block-wise. As CNN, the lower-index heads and neurons are preferentially kept. + +After training, we directly sample the candidate with the max probability given by the AG. It is noted that the obtained model inherits weights from the trained supernet without retraining or fine-tuning. Therefore, no search or retraining cost exists in ShiftNAS. + +# 4.2. Main Results + +We search ViT and CNN architectures on ImageNet with different FLOPs constraints. We compare searched models with multiple ViT and CNN models on model performance (Top-1 Acc., Top-5 Acc.) and efficiency (FLOPs, parameters). It is seen from Table 1 that our ShiftFormer models surpass the recent manual and autoML-based transformer models under diverse model sizes. It is worth noting that ShiftFormer-tiny achieves a top-1 Acc. of 76.0 with only 1.3 GFLOPs, being $1.3\%$ and $0.9\%$ better than AutoFormer-tiny and FocusFormer-T which are also autoML-based methods, respectively. Compared to CNN models, ShiftCNN models also achieve better accuracy than all compared networks under similar FLOPs restrictions. For example, ShiftCNN-S and ShiftCNN-B achieve $0.7\%$ and $0.7\%$ higher top-1 accuracy than BigNAS-S and BigNAS-M, respectively. Specifically, ShiftCNN and BigNAS models are searched from the same search space. BigNAS models spend more than $2\times$ GPU time than ShiftCNN models since the sandwich rule is used to train BigNAS supernet, which demonstrates the efficiency and outstanding performance of ShiftNAS. For fairness, we do not compare our models with others obtained from better search spaces [15, 29] in Table 1. The experiments on AttentiveNAS search space can be found in the Appendix. + +# 4.3. Ablation Study + +The effectiveness of probability shift. To demonstrate the effectiveness of the proposed sampling probability shift, we trained a ShiftNAS supernet and its baseline counterpart under ViT-tiny, ViT-small, ViT-base, and CNN search + +![](images/b5562594818d2f7f6de501150e4e7e4a00017b0ddeedb956bc27cf742d07e415.jpg) +Figure 5: The sampling probability with different FLOPs groups on search space (a) ViT-tiny; (b) ViT-small; (c) ViT-base; (d) CNN. + +![](images/dd20aabde47e64996b0068bd39f23c97ad57c5c6282821fc0856fca9036bfa4d.jpg) + +![](images/929abfbae529bd8c1aa5c83ed4c5bbb247a9fcd86a116d37dbf9cf6076da51c8.jpg) + +![](images/2d79223fc352330bc002925a085337fd9186e29180aff94bdca2e2adb5439be8.jpg) + +![](images/78bd1c72ac16b7d02f9ef833409f3a6a7e46e1341186558701ae38c350393970.jpg) +Figure 6: The ranking correlation of architecture generator. + +spaces. In the baseline supernet training, we uniformly sampled FLOPs in each step and did not update FLOPs probability. The training settings for all supernets are provided in the Appendix. As illustrated in Figure 4, the supernet trained with probability shift is capable of covering a wide range of accuracy-FLOPs curves and outperforms the baseline supernet trained without probability shift by a significant margin, demonstrating the effectiveness of the proposed sampling probability shift mechanism. + +The ranking correlation of architecture generator. The experiment is based on ShiftFormer-tiny supernet. We measure the ranking correlation by randomly sampling 25 subnets from the ShiftFormer-tiny supernet. For fast evaluation, The sampled subnets are inherited weights from the supernet and are further finetuned to reach the optimal weights. As for the model ranking test, we evaluate sampled subnets in both supernets to get their inherited performance. As shown in Figure 6, the inherited performance is highly correlated with finetuned performance in good subnets (Inherited Top1-Acc > 65%). Despite the correlation with bad subnets is not obvious, we focus more attention on good subnets since bad subnets are not deployed in practice. + +Subnet performance without retraining. We further study the performance of ShiftNAS models when models + +are independently fine-tuned or trained from scratch. Following Autoformer, we finetune or retrain the selected subnet with 40 epochs or 300 epochs on ImageNet. Table 2 shows that finetuning cannot improve or even harm the performance of subnets inherited from supernet. Besides, retraining can slightly improve the subnet performance but cost extra training resources which prevents the implementation of NAS. + +The optimal sampling strategy. ShiftNAS learns the optimal sampling probability for each subnet with different FLOPs. To figure out what sampling strategy is the best, we visualize the various sampling probability in different search spaces (ViT-tiny, ViT-small, ViT-base and CNN) at different training steps. As can be seen in Figure 5, our method learns to sample more subnets with large computational complexity. For example, the sampling probabilities of 1.6 to 1.9 GFLOPs subnets are clearly higher than 1.3 and 1.4 GFLOPs subnets in the ViT-tiny search space. This is intuitive since the large subnet contains more parameters and needs more computational resources to update. + +# 5. Conclusion + +In this paper, we presented ShiftNAS, a one-shot supernet training framework that can complexity-wise adjust the sampling probability. To automatically adjust the sampling probability, we proposed probability shift that can be learned according to the subnets training sufficiency. To fast obtain the subnets with desired complexity, AG is designed that can be jointly trained with the supernet in an end-to-end manner. Experiments on ImageNet showed that our method achieves SOTA results on ViT/CNN search spaces, and covers a wide range of efficiency/accuracy trade-offs without any extra retraining. + +Limitation. ShiftNAS is designed for efficient search on weight-entanglement search spaces, so it cannot work on darts-like search spaces [18, 25]. + +Acknowledgements: This work was supported by the National Natural Science Foundation of China (62203392) and Natural Science Foundation of Zhejiang province (LY21F030018). + +# References + +[1] Gabriel Bender, Pieter-Jan Kindermans, Barret Zoph, Vijay Vasudevan, and Quoc Le. Understanding and simplifying one-shot architecture search. In International conference on machine learning, pages 550-559. PMLR, 2018. 4 +[2] Han Cai, Chuang Gan, Ji Lin, and Song Han. Network augmentation for tiny deep learning. arXiv preprint arXiv:2110.08890, 2021. 2 +[3] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. In International Conference on Learning Representations, 2019. 2, 6 +[4] Minghao Chen, Houwen Peng, Jianlong Fu, and Haibin Ling. Autoformer: Searching transformers for visual recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12270-12280, 2021. 2, 3, 6, 7 +[5] Xiangxiang Chu, Bo Zhang, Ruijun Xu, and Jixiang Li. Fairnas: Rethinking evaluation fairness of weight sharing neural architecture search. arXiv preprint arXiv:1907.01845, 2019. 3 +[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 6 +[7] Zichao Guo, Xiangyu Zhang, Haoyuan Mu, Wen Heng, Zechun Liu, Yichen Wei, and Jian Sun. Single path one-shot neural architecture search with uniform sampling. In European Conference on Computer Vision, pages 544-560. Springer, 2020. 1, 2, 3 +[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1 +[9] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 1, 2, 6 +[10] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 1 +[11] Shoukang Hu, Sirui Xie, Hehui Zheng, Chunxiao Liu, Jianping Shi, Xunying Liu, and Dahua Lin. Dsnas: Direct neural architecture search without parameter retraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12084-12092, 2020. 5 +[12] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. 2016. 2 +[13] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural net + +works. Advances in neural information processing systems, 25:1097-1105, 2012. 6 +[14] Changlin Li, Jiefeng Peng, Liuchun Yuan, Guangrun Wang, Xiaodan Liang, Liang Lin, and Xiaojun Chang. Blockwisely supervised neural architecture search with knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1989-1998, 2020. 2, 4 +[15] Changlin Li, Tao Tang, Guangrun Wang, Jiefeng Peng, Bing Wang, Xiaodan Liang, and Xiaojun Chang. Bossnas: Exploring hybrid cnn-transformers with block-wisely self-supervised neural architecture search. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12281-12291, 2021. 4, 7 +[16] Changlin Li, Guangrun Wang, Bing Wang, Xiaodan Liang, Zhihui Li, and Xiaojun Chang. Ds-net++: Dynamic weight slicing for efficient inference in cnns and vision transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-16, 2022. 6 +[17] Liam Li and Ameet Talwalkar. Random search and reproducibility for neural architecture search. In Uncertainty in artificial intelligence, pages 367-377. PMLR, 2020. 3 +[18] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. In International Conference on Learning Representations, 2018. 2, 4, 5, 8 +[19] Jing Liu, Jianfei Cai, and Bohan Zhuang. Focusformer: Focusing on what we need via architecture sampler. arXiv preprint arXiv:2208.10861, 2022. 3, 6 +[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1 +[21] James E Marengo, David L Farnsworth, and Lucas Stefanic. A geometric derivation of the irwin-hall distribution. International Journal of Mathematics and Mathematical Sciences, 2017, 2017. 3 +[22] Yameng Peng, Andy Song, Vic Ciesielski, Haytham M. Fayek, and Xiaojun Chang. Pre-nas: Evolutionary neural architecture search with predictor. IEEE Transactions on Evolutionary Computation, 27(1):26-36, 2023. 1 +[23] Hieu Pham, Melody Guan, Barret Zoph, Quoc Le, and Jeff Dean. Efficient neural architecture search via parameters sharing. In International conference on machine learning, pages 4095-4104. PMLR, 2018. 1, 3 +[24] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 1, 6 +[25] Julien Siems, Lucas Zimmer, Arber Zela, Jovita Lukasik, Margret Keuper, and Frank Hutter. Nas-bench-301 and the case for surrogate benchmarks for neural architecture search. arXiv preprint arXiv:2008.09777, 2020. 8 +[26] Aravind Srinivas, Tsung-Yi Lin, Niki Parmar, Jonathon Shlens, Pieter Abbeel, and Ashish Vaswani. Bottleneck transformers for visual recognition. In Proceedings of + +the IEEE/CVF conference on computer vision and pattern recognition, pages 16519-16529, 2021. 6 +[27] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. 6 +[28] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 6 +[29] Dilin Wang, Meng Li, Chengyue Gong, and Vikas Chandra. Attentivenas: Improving neural architecture search via attentive sampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6418-6427, 2021. 2, 3, 7 +[30] Bichen Wu, Xiaoliang Dai, Peizhao Zhang, Yanghan Wang, Fei Sun, Yiming Wu, Yuandong Tian, Peter Vajda, Yangqing Jia, and Kurt Keutzer. Fbnet: Hardware-aware efficient convnet design via differentiable neural architecture search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10734-10742, 2019. 1 +[31] Shan You, Tao Huang, Mingmin Yang, Fei Wang, Chen Qian, and Changshui Zhang. Greedynas: Towards fast one-shot nas with greedy supernet. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1999-2008, 2020. 3, 6 +[32] Jiahui Yu and Thomas Huang. Autoslim: Towards one-shot architecture search for channel numbers. arXiv preprint arXiv:1903.11728, 2019. 2 +[33] Jiahui Yu, Pengchong Jin, Hanxiao Liu, Gabriel Bender, Pieter-Jan Kindermans, Mingxing Tan, Thomas Huang, Xiaodan Song, Ruoming Pang, and Quoc Le. Bignas: Scaling up neural architecture search with big single-stage models. In European Conference on Computer Vision, pages 702-717. Springer, 2020. 2, 3, 6, 7 +[34] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF international conference on computer vision, pages 558-567, 2021. 6 +[35] Barret Zoph and Quoc V Le. Neural architecture search with reinforcement learning. arXiv preprint arXiv:1611.01578, 2016. 1, 3, 5 \ No newline at end of file diff --git a/2023/ShiftNAS_ Improving One-shot NAS via Probability Shift/images.zip b/2023/ShiftNAS_ Improving One-shot NAS via Probability Shift/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2eaddc96c2e9d1d02f034ef714e55ad026d5d5cc --- /dev/null +++ b/2023/ShiftNAS_ Improving One-shot NAS via Probability Shift/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57a608dfedc7c850be60daefa13fa9406b033611e8c5ed0555553c04f05482b1 +size 536962 diff --git a/2023/ShiftNAS_ Improving One-shot NAS via Probability Shift/layout.json b/2023/ShiftNAS_ Improving One-shot NAS via Probability Shift/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..15056a03a94f3b57ce8d7f98295f965be4ffe643 --- /dev/null +++ b/2023/ShiftNAS_ Improving One-shot NAS via Probability Shift/layout.json @@ -0,0 +1,7942 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 474, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 474, + 121 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 474, + 121 + ], + "type": "text", + "content": "ShiftNAS: Improving One-shot NAS via Probability Shift" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 143, + 431, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 143, + 431, + 185 + ], + "spans": [ + { + "bbox": [ + 162, + 143, + 431, + 185 + ], + "type": "text", + "content": "Mingyang Zhang, Xinyi Yu* Haodong Zhao, Linlin Ou* Zhejiang University of Technology Hangzhou, Zhejiang, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 192, + 188, + 400, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 188, + 400, + 200 + ], + "spans": [ + { + "bbox": [ + 192, + 188, + 400, + 200 + ], + "type": "text", + "content": "yuxy@zjut.edu.cn, linlinou@zjut.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 252, + 289, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 289, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 289, + 408 + ], + "type": "text", + "content": "One-shot Neural architecture search (One-shot NAS) has been proposed as a time-efficient approach to obtain optimal subnet architectures and weights under different complexity cases by training only once. However, the subnet performance obtained by weight sharing is often inferior to the performance achieved by retraining. In this paper, we investigate the performance gap and attribute it to the use of uniform sampling, which is a common approach in supernet training. Uniform sampling concentrates training resources on subnets with intermediate computational resources, which are sampled with high probability. However, subnets with different complexity regions require different optimal training strategies for optimal performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 408, + 289, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 289, + 613 + ], + "type": "text", + "content": "To address the problem of uniform sampling, we propose ShiftNAS, a method that can adjust the sampling probability based on the complexity of subnets. We achieve this by evaluating the performance variation of subnets with different complexity and designing an architecture generator that can accurately and efficiently provide subnets with the desired complexity. Both the sampling probability and the architecture generator can be trained end-to-end in a gradient-based manner. With ShiftNAS, we can directly obtain the optimal model architecture and parameters for a given computational complexity. We evaluate our approach on multiple visual network models, including convolutional neural networks (CNNs) and vision transformers (ViTs), and demonstrate that ShiftNAS is model-agnostic. Experimental results on ImageNet show that ShiftNAS can improve the performance of one-shot NAS without additional consumption. Source codes are available at GitHub." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 636, + 128, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 636, + 128, + 649 + ], + "spans": [ + { + "bbox": [ + 47, + 636, + 128, + 649 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 657, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 287, + 693 + ], + "type": "text", + "content": "Deep neural networks (DNNs) have been widely applied to the field of computer vision with remarkable success [6, 8]. However, the deployment of these vision net" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 228, + 545, + 369 + ], + "blocks": [ + { + "bbox": [ + 310, + 228, + 545, + 369 + ], + "lines": [ + { + "bbox": [ + 310, + 228, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 310, + 228, + 545, + 369 + ], + "type": "image", + "image_path": "9e326b61809dda24f08e4295dbc212bcc6ebd11bb228479e249d6a1220227467.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 376, + 547, + 496 + ], + "lines": [ + { + "bbox": [ + 304, + 376, + 547, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 547, + 496 + ], + "type": "text", + "content": "Figure 1: A conceptual overview of our ShiftNAS. Focusing on the computational resource (e.g. FLOPs), ShiftNAS split a search space into several parts where subnets have close computational complexity. In stochastic supernet training, ShiftNAS first samples computational complexity according to probability, then an architecture generator samples the subnet with desired computational complexity. The probability distribution is dynamically updated with stochastic training by estimating the performance variation of each sub-space." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 521, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 616 + ], + "type": "text", + "content": "works on edge devices still has some limitations, such as the massive model sizes and excessive computation overhead [9, 10, 20, 24]. In addition, designing architectures artificially in a trial-and-error manner is a resource-consuming task that requires not only architectural skills but also domain expertise. Consequently, how to acquire optimal architectures that balance latency and accuracy efficiently is of paramount importance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "content": "Recent advancements in neural architecture search (NAS) methods, such as [22, 7, 30, 35], have led to significant improvements in the performance of practical applications by automatically searching for optimal architectures within a defined search space. However, traditional NAS methods typically require a substantial computation budget [23, 35]. In order to speed up training and reduce the resource consumption of the training process, one-shot NAS" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 227, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 227, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 227, + 713 + ], + "type": "text", + "content": "* Xinyi Yu and Linlin Ou are corresponding authors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "5919" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 276 + ], + "type": "text", + "content": "methods [18, 7] have adopted a two-stage training approach based on weight sharing. Specifically, a supernet is trained in the first stage, and subnets with better performance are searched for in the second stage. In some cases, post-processing methods (e.g., retraining and finetuning) are also necessary in the second stage since the performance of these subnets inherited from the supernet is often inferior to that of models trained from scratch. However, the downside of such methods is that the training consumption increases linearly with the number of architectures, which can be problematic. To address this issue, some one-shot NAS methods [3, 4, 29, 33] have utilized a weight entanglement training strategy to share the weights in each operation, eliminating the need for additional finetuning or retraining. Furthermore, a high-quality supernet is essential for candidate architectures to inherit weights directly [33], as well as for accurately ranking candidate architectures [14]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 277, + 289, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 289, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 289, + 506 + ], + "type": "text", + "content": "However, there is still potential for breakthroughs in training a better supernet. In the first stage, previous methods [7, 4] assume that all candidate architectures are equally important and should be sampled with equal probability during training. However, subnets with different numbers of parameters require different amounts of training resources [2]. For instance, subnets with 1.0 GFLOPs may converge after 30000 iterations, while subnets with 2.0 GFLOPs may require 50000 iterations. Additionally, only the optimal subnet will be deployed, while others will be ignored at the same computational complexity. Therefore, subnets that occupy more training resources may not be distributed in regions corresponding to resource constraints, resulting in sub-optimal performance of the final deployed model. We find that when all subnets are sampled with equal probability, the resulting computational resource distribution is approximately normal. Consequently, subnets trained under this distribution may appear to be under-fitting or over-fitting in different regions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 508, + 289, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 289, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 289, + 674 + ], + "type": "text", + "content": "To address the challenge of efficiently training a high-performance supernet, we propose a novel method, probability-Shift Neural Architecture search (ShiftNAS). In ShiftNAS, the sampling probability of subnets is not uniform and can be dynamically adjusted during the training process. The training sufficiency of each subnet is measured by evaluating its performance variation under different computational constraints. The subnets with high-performance variance are identified as undertrained, and their sampling probabilities are increased to tilt training resources dynamically towards them. This enables us to allocate resources more effectively and efficiently to achieve better performance for the subnets that need more training resources." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "In spite of having an optimal sampling distribution, efficiently and accurately sampling subnets with a desired computational constraint still poses a challenge. To address this" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 262 + ], + "type": "text", + "content": "issue, we propose an LSTM-based architecture generator (AG) that can be optimized differentiably with a resource constraint loss function. The AG's output is then processed by Gumbel Softmax [12] to generate a one-hot vector policy for each searched operation. To suit the weight-entangled search space [32, 4], we employ a matrix mapping technique that can convert the one-hot vector into a differentiable mask. The mask is multiplied by the operation to obtain a differentiable subnet. The AG and supernet can be jointly trained to learn how to generate the best subnets with desired computational constraints. During evaluation, the AG can generate a corresponding subnet immediately for any given computational constraint. The weights of the searched subnet can be directly inherited from the well-trained supernet, making ShiftNAS free from any additional search or retrain costs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 263, + 542, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 263, + 542, + 275 + ], + "spans": [ + { + "bbox": [ + 317, + 263, + 542, + 275 + ], + "type": "text", + "content": "The overall contribution can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 281, + 545, + 425 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 317, + 281, + 545, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 281, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 317, + 281, + 545, + 329 + ], + "type": "text", + "content": "- A learnable sampling strategy, called probability shift, is proposed to relief the bias of uniform sampling which leads to performance gap between supernet training and subnet deployment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 335, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 335, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 317, + 335, + 545, + 384 + ], + "type": "text", + "content": "- We propose an LSTM-based AG to precisely and efficiently offer the best subnet with desired resource constraints. AG training can be differentiably trained with supernet under weight-entangled search space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 390, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 390, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 390, + 545, + 425 + ], + "type": "text", + "content": "- We achieve state-of-the-art or competitive results on both CNN and ViT models. Therefore, ShiftNAS is a model-agnostic search method." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 435, + 392, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 435, + 392, + 448 + ], + "spans": [ + { + "bbox": [ + 306, + 435, + 392, + 448 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 455, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 545, + 503 + ], + "type": "text", + "content": "The one-shot NAS method [3, 29, 33] can automatically search for the optimal architecture in a predefined search space [4, 9, 18], which can be formulated by maximizing an expected accuracy over the space " + }, + { + "bbox": [ + 304, + 455, + 545, + 503 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 455, + 545, + 503 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 361, + 509, + 545, + 549 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 509, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 361, + 509, + 545, + 549 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\alpha \\in A} L _ {v a l} \\left(w ^ {*} \\mid \\alpha\\right) \\\\ \\text {s . t .} w ^ {*} = \\underset {w _ {\\alpha}} {\\arg \\min } L _ {\\text {t r a i n}} (w \\mid \\alpha) \\tag {1} \\\\ \\end{array}", + "image_path": "bbabd251cd60f8f0a883e8aabae3c494efaa789366844c123b5428f3cc7f6154.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 361, + 552, + 446, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 552, + 446, + 564 + ], + "spans": [ + { + "bbox": [ + 361, + 552, + 446, + 564 + ], + "type": "interline_equation", + "content": "c o m p l e x i t y (\\alpha) < \\tau", + "image_path": "9554ac3631f95fda5b7ad35df9f30fdbfd478362eee67ab5e72157a530d6d8d7.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 570, + 545, + 654 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 570, + 545, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 570, + 545, + 654 + ], + "type": "text", + "content": " represent the weight and architecture of subnets. " + }, + { + "bbox": [ + 304, + 570, + 545, + 654 + ], + "type": "inline_equation", + "content": "L_{val}" + }, + { + "bbox": [ + 304, + 570, + 545, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 545, + 654 + ], + "type": "inline_equation", + "content": "L_{train}" + }, + { + "bbox": [ + 304, + 570, + 545, + 654 + ], + "type": "text", + "content": " are loss functions in the validation and train dataset, respectively. The computational complexity of " + }, + { + "bbox": [ + 304, + 570, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 570, + 545, + 654 + ], + "type": "text", + "content": " can be calculated by the complexity function. " + }, + { + "bbox": [ + 304, + 570, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 570, + 545, + 654 + ], + "type": "text", + "content": " denotes the complexity threshold for alpha. To solve this optimization problem, a two-stage approach based on weight sharing is usually employed." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "In the first stage, a high-quality supernet is trained by sampling a large number of subnets in the defined search space. Notably, in order to obtain subnets that need to be incorporated into the training phase, the following approaches are commonly used in recent one-shot NAS works [4, 7]." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "5920" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": "First is the uniform sampling method, which is also the most common strategy. The Uniform sample method considers that all architectures in the supernet are equally important, that is, they should be sampled with exactly equal probability. The weight optimization equation in the first stage can be formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 113, + 152, + 221, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 152, + 221, + 169 + ], + "spans": [ + { + "bbox": [ + 113, + 152, + 221, + 169 + ], + "type": "interline_equation", + "content": "\\min _ {w} E _ {\\alpha \\sim A} \\left[ L _ {\\text {t r a i n}} (w \\mid \\alpha) \\right]", + "image_path": "49297883f456cf5aecab9ffd57533063cf49a8d0945946f4813267ae815486eb.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 113, + 171, + 287, + 183 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 171, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 113, + 171, + 287, + 183 + ], + "type": "interline_equation", + "content": "\\alpha = o ^ {1} \\cup o ^ {2} \\cup \\dots \\cup o ^ {D} \\tag {2}", + "image_path": "ada3397aacd10730125857ca1f62fe4d203d03d7ed26b0941684be75a8a76c03.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 188, + 173, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 188, + 173, + 201 + ], + "spans": [ + { + "bbox": [ + 113, + 188, + 173, + 201 + ], + "type": "interline_equation", + "content": "o ^ {i} \\sim U (0, n _ {i})", + "image_path": "44586a245de3865fc9dbacef09e9bc52546ee7f2fd1681aa143ff361b5afd854.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 209, + 287, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 209, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 287, + 243 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 209, + 287, + 243 + ], + "type": "inline_equation", + "content": "D, n" + }, + { + "bbox": [ + 47, + 209, + 287, + 243 + ], + "type": "text", + "content": " denotes the total operation and candidate number in certain operations, respectively. " + }, + { + "bbox": [ + 47, + 209, + 287, + 243 + ], + "type": "inline_equation", + "content": "o^i" + }, + { + "bbox": [ + 47, + 209, + 287, + 243 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 47, + 209, + 287, + 243 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 209, + 287, + 243 + ], + "type": "text", + "content": "th selected operation, which follows a uniform distribution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 244, + 287, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 244, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 287, + 351 + ], + "type": "text", + "content": "Since there is a gap between the training process and practical deployment, AttentiveNAS [29] proposes the Pareto-aware sampling method. Specifically, models deployed on edge devices are generally around the Pareto frontier, but they are uniformly sampled during training, which leads to the waste of training resources to a certain extent. To bridge the gap between training and deploying, each iteration extracts multiple subnets with equal size and selects the best or worst architecture for training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 351, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 351, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 351, + 287, + 399 + ], + "type": "text", + "content": "Similar to AttentiveNAS, GreedyNAS [31] screen the weak subnets and just sample from the potentially-good subnets instead of all subnets, thus capturing another opportunity to improve the accuracy of target models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 400, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 518 + ], + "type": "text", + "content": "However, these uniformly sampled architectures appear to be an approximately normal distribution of the computational cost. To further reduce the gap between training and deployment, Focusformer [19] proposes that all the sampled subnets should be based on resource distribution and focuses more on the Pareto frontier architectures through an architecture sampler. In addition, FairNAS [5] also reinforces the sampling process with a stricter standard of fairness, that is, the parameters of each choice block are updated the same number of times at any stage." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 519, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 567 + ], + "type": "text", + "content": "When the first stage is over, this well-trained supernet can be used as the performance estimator of the candidate architecture. And the weight of candidate architectures can be directly inherited from the supernet." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 567, + 287, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 602 + ], + "type": "text", + "content": "In the second stage, we need to search for the best candidate architecture under different resource constraints in the supernet, which can be formulated as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 611, + 287, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 611, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 116, + 611, + 287, + 634 + ], + "type": "interline_equation", + "content": "\\alpha = \\underset {\\alpha \\in A} {\\operatorname {a r g m i n}} L _ {v a l} (w \\mid \\alpha) \\tag {3}", + "image_path": "df87d2af0d4ca7c023651a6202bd468ef8fc5dd913d02643a7ea63053ddb0ea4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 117, + 634, + 217, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 634, + 217, + 646 + ], + "spans": [ + { + "bbox": [ + 117, + 634, + 217, + 646 + ], + "type": "interline_equation", + "content": "s. t. c o m p l e x i t y (\\alpha) < \\tau", + "image_path": "e2e081af0e2be59525779d0b822b72e337cd9296aa84bdb0e060343740a94bd6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "text", + "content": "Since there are so many architectures in the supernet that need to be evaluated, the consumption of computing resources is also unacceptable. Therefore, recent works resort to random search [17], evolution algorithms [4, 7, 23] or reinforcement learning [35] to find the most promising" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "architecture among all architectures. In the end, these selected architectures need some post-processing operations such as retraining to improve the architecture performance, but this also brings additional computational consumption." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 121, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 545, + 192 + ], + "type": "text", + "content": "In order to alleviate the above-mentioned resource consumption problem, some one-shot NAS methods train a high-quality supernet so that candidate architectures can directly inherit supernet weights without retraining, such as Autoformer [4], attentiveNAS [29], BigNAS [33], FocusFormer [19]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 205, + 545, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 205, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 545, + 232 + ], + "type": "text", + "content": "3. Nerual Architecture Search with Probability Shift" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 240, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 361 + ], + "type": "text", + "content": "In this section, we first discuss the training resources allocation problem caused by uniform sampling in one-shot NAS. To solve the above-mentioned problem, we present ShiftNAS, an end-to-end supernet training framework. In ShiftNAS, a learnable sampling strategy is proposed, which can dynamically adjust the training resources allocation by shifting the sampling probability. Then, an LSTM-based AG is designed to accurately obtain the expected subnet for each sampling. We show the overview of our method in Figure 1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 369, + 527, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 369, + 527, + 382 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 527, + 382 + ], + "type": "text", + "content": "3.1. Rethinking the Sampling of One-shot NAS" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": "In one-shot NAS, an overparameterized supernet " + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": " is formed with multiple operations, where each operation contains several choices. The supernet " + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": " operations, and the " + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": "th operation " + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "inline_equation", + "content": "o^d" + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": " can be selected from " + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": " candidates " + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "inline_equation", + "content": "o_1, \\ldots, o_n" + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": ", which represent various design choices, such as kernel size and channels in CNN search space [33], or heads number and MLP ratio in ViT search space [4]. A subnet " + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": " sampled from the supernet can be represented by a tuple of size " + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "inline_equation", + "content": "a = (o^1, \\ldots, o^D)" + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "type": "text", + "content": "In previous one-shot NAS methods, the sampling probability of each operation candidate " + }, + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "type": "inline_equation", + "content": "p(o_i) = \\frac{1}{n}" + }, + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "type": "text", + "content": ", which assumes that all candidates are equally important. However, in reality, one-shot NAS is more concerned with the computational resources of the subnets. To address this, we introduce the notion of computational resource for each operation candidate " + }, + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "type": "inline_equation", + "content": "o^d" + }, + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "type": "text", + "content": ", denoted by " + }, + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "type": "inline_equation", + "content": "b^d \\in b_1, \\ldots, b_n" + }, + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "type": "text", + "content": ". The computational resource of a subnet " + }, + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "type": "text", + "content": " randomly sampled from the supernet can then be computed as " + }, + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "type": "inline_equation", + "content": "B_a = \\sum_{d=1}^{D} b^d" + }, + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "type": "inline_equation", + "content": "b^d" + }, + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "type": "text", + "content": " follows a uniform distribution." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "Remark 1 If we assume that the computational resource of each operation " + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "b^{d}" + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": " is independently sampled from a uniform distribution, then the total computational resource " + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "B_{a}" + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": " of a subnet " + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": " sampled from the supernet follows an Irwin-Hall distribution. As the number of operations " + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": " in the supernet increases, the Irwin-Hall distribution converges to a normal distribution [21]." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 314, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 314, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 314, + 742 + ], + "type": "text", + "content": "5921" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 72, + 149, + 148 + ], + "blocks": [ + { + "bbox": [ + 72, + 72, + 149, + 148 + ], + "lines": [ + { + "bbox": [ + 72, + 72, + 149, + 148 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 149, + 148 + ], + "type": "image", + "image_path": "5060003a72932d560ab426fa4fed7db9cf51f1a7cb2aa193565dd52081eddf54.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 186, + 72, + 263, + 148 + ], + "blocks": [ + { + "bbox": [ + 186, + 72, + 263, + 148 + ], + "lines": [ + { + "bbox": [ + 186, + 72, + 263, + 148 + ], + "spans": [ + { + "bbox": [ + 186, + 72, + 263, + 148 + ], + "type": "image", + "image_path": "f3adc5a47f36faa731ecb92ce7adf7f26dd5a41c0966ffe2a99d46c84b89e85d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 70, + 155, + 149, + 232 + ], + "blocks": [ + { + "bbox": [ + 70, + 155, + 149, + 232 + ], + "lines": [ + { + "bbox": [ + 70, + 155, + 149, + 232 + ], + "spans": [ + { + "bbox": [ + 70, + 155, + 149, + 232 + ], + "type": "image", + "image_path": "ad74b05bf2dcf2d2b8462c9e21c21b1eb1f2b51c2682900b27c7b804b5af1b7f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 236, + 244, + 247 + ], + "lines": [ + { + "bbox": [ + 90, + 236, + 244, + 247 + ], + "spans": [ + { + "bbox": [ + 90, + 236, + 244, + 247 + ], + "type": "text", + "content": "(a) Uniform distribution of each operation" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 184, + 156, + 262, + 232 + ], + "blocks": [ + { + "bbox": [ + 184, + 156, + 262, + 232 + ], + "lines": [ + { + "bbox": [ + 184, + 156, + 262, + 232 + ], + "spans": [ + { + "bbox": [ + 184, + 156, + 262, + 232 + ], + "type": "image", + "image_path": "17c40fd0c370ca01fcb2cb3e69be4e0588b60ac8784c68f29c6164208feb8dbc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 71, + 250, + 264, + 418 + ], + "blocks": [ + { + "bbox": [ + 71, + 250, + 264, + 418 + ], + "lines": [ + { + "bbox": [ + 71, + 250, + 264, + 418 + ], + "spans": [ + { + "bbox": [ + 71, + 250, + 264, + 418 + ], + "type": "image", + "image_path": "1d1366ad1a27c4a1479c1679c671b708734673d3612c6981c9738afad63f2250.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 422, + 274, + 434 + ], + "lines": [ + { + "bbox": [ + 59, + 422, + 274, + 434 + ], + "spans": [ + { + "bbox": [ + 59, + 422, + 274, + 434 + ], + "type": "text", + "content": "(b) Normal distribution of subnets computational resources" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 443, + 287, + 491 + ], + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 491 + ], + "type": "text", + "content": "Figure 2: During supernet training, operations are sampled from a uniform distribution in each iteration. The subnet consists of the sampled operations and its computational resource follows an approximated normal distribution." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 515, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 287, + 624 + ], + "type": "text", + "content": "Upon analyzing Remark 1, it is evident that the uniform sampling strategy results in subnets being sampled with moderate computational resources, as depicted in Figure 2. This observation implies that subnets with moderate computational resources can be trained effectively, whereas subnets with large or small computational resources cannot be fully trained. Insufficiently-trained subnets lead to inaccurate ranking [14, 15] and inherit unreasonable weights from the supernet without retraining [1]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 634, + 261, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 261, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 261, + 647 + ], + "type": "text", + "content": "3.2. Shifting Probability with Greedy Sample" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "One-shot NAS methods rely on the assumption that the supernet can effectively rank the performance of subnets with arbitrary computational resources, and therefore require each subnet to be fully trained with optimal training strategies. However, in practice, subnets with differ" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": "ent computational resources may require different training strategies, and the uniform sampling strategy used in previous one-shot NAS methods may lead to insufficient training of subnets in certain computational regions. For example, subnets with extreme computational resources may not be fully trained even when subnets with moderate computational resources are fully trained. To address this issue, we propose a learnable greedy sampling strategy that takes into account the training sufficiency of subnets in different computational regions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 192, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 192, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 305, + 192, + 545, + 216 + ], + "type": "text", + "content": "For any subnet, we judge whether the subnet has converged by calculating the following gradient of the subnet." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 374, + 227, + 545, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 227, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 374, + 227, + 545, + 240 + ], + "type": "interline_equation", + "content": "\\nabla_ {w} = \\nabla_ {w} L _ {\\text {v a l}} (w, \\alpha \\mid b) \\tag {4}", + "image_path": "b665aed31a75ade3e1838b8e52066757760b7f1a47832f8a3760bff43fd203e0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 243, + 546, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 546, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 546, + 363 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 243, + 546, + 363 + ], + "type": "inline_equation", + "content": "L_{val}" + }, + { + "bbox": [ + 304, + 243, + 546, + 363 + ], + "type": "text", + "content": " represents the task loss (e.g. cross entropy) in the validation dataset, " + }, + { + "bbox": [ + 304, + 243, + 546, + 363 + ], + "type": "inline_equation", + "content": "w, \\alpha" + }, + { + "bbox": [ + 304, + 243, + 546, + 363 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 243, + 546, + 363 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 243, + 546, + 363 + ], + "type": "text", + "content": " denote the weight, architecture and sampled computational resource of the subnet, respectively. When the subnet with " + }, + { + "bbox": [ + 304, + 243, + 546, + 363 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 243, + 546, + 363 + ], + "type": "text", + "content": " computational resource converges, " + }, + { + "bbox": [ + 304, + 243, + 546, + 363 + ], + "type": "inline_equation", + "content": "\\nabla_w" + }, + { + "bbox": [ + 304, + 243, + 546, + 363 + ], + "type": "text", + "content": " will tend to zero. In other words, insufficiency-trained subnets have a small gradient (probably negative) and should be sampled more times. Therefore, we can greedily adjust the sampling distribution of " + }, + { + "bbox": [ + 304, + 243, + 546, + 363 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 243, + 546, + 363 + ], + "type": "text", + "content": " by optimizing the computational resource distribution at each training step" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 356, + 369, + 545, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 369, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 356, + 369, + 545, + 389 + ], + "type": "interline_equation", + "content": "\\underset {B} {\\arg \\min } E _ {b \\sim B} \\left[ \\nabla_ {w} L _ {v a l} (w, \\alpha \\mid b) \\right] \\tag {5}", + "image_path": "2ad4d4ebf79050fd9540692cdf78d5c1427d8608e23d4e026b0f7559d628e3d8.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 394, + 545, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 394, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 305, + 394, + 545, + 466 + ], + "type": "text", + "content": "To optimize " + }, + { + "bbox": [ + 305, + 394, + 545, + 466 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 305, + 394, + 545, + 466 + ], + "type": "text", + "content": " end to end based on gradient-based methods, we discrete computational resources into several parts, e.g. " + }, + { + "bbox": [ + 305, + 394, + 545, + 466 + ], + "type": "inline_equation", + "content": "[1.3G, 1.4G, \\dots, 2.0G]" + }, + { + "bbox": [ + 305, + 394, + 545, + 466 + ], + "type": "text", + "content": " FLOPs, and uniformly initialize its probability distribution. In each training step, one computational resource can be sampled by Gumbel Softmax " + }, + { + "bbox": [ + 305, + 394, + 545, + 466 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 305, + 394, + 545, + 466 + ], + "type": "text", + "content": ". Formally, Eq. 5 can be rewritten as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 362, + 472, + 545, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 472, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 362, + 472, + 545, + 491 + ], + "type": "interline_equation", + "content": "\\underset {B} {\\arg \\min } \\nabla_ {w} L _ {v a l} (w, \\alpha \\mid G (B)) \\tag {6}", + "image_path": "79503d443d8f24507ea574c22c5a3e01228dd766fe49357ce576b701f2174323.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 497, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 497, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 305, + 497, + 545, + 521 + ], + "type": "text", + "content": "In this way, the gradient of " + }, + { + "bbox": [ + 305, + 497, + 545, + 521 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 305, + 497, + 545, + 521 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 305, + 497, + 545, + 521 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 497, + 545, + 521 + ], + "type": "text", + "content": "th update can be obtained by applying chain rule to Eq. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 329, + 527, + 545, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 527, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 329, + 527, + 545, + 541 + ], + "type": "interline_equation", + "content": "\\nabla_ {B _ {t}} = \\nabla_ {w _ {t}, B _ {t}} ^ {2} L _ {\\text {v a l}} \\left(w _ {t}, \\alpha \\mid G (B _ {t})\\right) \\nabla_ {B _ {t}} G (B _ {t}) \\tag {7}", + "image_path": "8827f55b6dd464506662939c5606a9ee9104b8675109e5a9547f48e2ddf0ccb0.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 547, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 547, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 547, + 545, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 547, + 545, + 641 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 547, + 545, + 641 + ], + "type": "text", + "content": " denotes the update times of " + }, + { + "bbox": [ + 304, + 547, + 545, + 641 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 547, + 545, + 641 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 547, + 545, + 641 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 304, + 547, + 545, + 641 + ], + "type": "text", + "content": " represents the supernet weight when " + }, + { + "bbox": [ + 304, + 547, + 545, + 641 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 547, + 545, + 641 + ], + "type": "text", + "content": " has been updated " + }, + { + "bbox": [ + 304, + 547, + 545, + 641 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 547, + 545, + 641 + ], + "type": "text", + "content": " times. However, evaluating the gradient of " + }, + { + "bbox": [ + 304, + 547, + 545, + 641 + ], + "type": "inline_equation", + "content": "B_{t}" + }, + { + "bbox": [ + 304, + 547, + 545, + 641 + ], + "type": "text", + "content": " requires an expensive matrix-vector product in its second term. Inspired by darts [18], the complexity can be substantially reduced using the finite difference approximation. We use two training steps to approximate the gradient of " + }, + { + "bbox": [ + 304, + 547, + 545, + 641 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 547, + 545, + 641 + ], + "type": "text", + "content": ", therefore, Eq. 7 can be rewritten as" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 328, + 646, + 545, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 646, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 328, + 646, + 545, + 672 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {B _ {t}} = \\nabla_ {B _ {t}} L _ {\\text {v a l}} \\left(w _ {t}, \\alpha \\mid G \\left(B _ {t}\\right)\\right) \\nabla_ {B _ {t}} G \\left(B _ {t}\\right) \\tag {8} \\\\ - \\nabla_ {B _ {t}} L _ {v a l} \\left(w _ {t - 1}, \\alpha \\mid G (B _ {t})\\right) \\nabla_ {B _ {t}} G (B _ {t}) \\\\ \\end{array}", + "image_path": "364734ef1f4d3528831468ef7099facc398eecd93035c3c83452928bdae7fcb6.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "In the implementation, the distribution " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " is only updated every few iterations, so there is no excessive computational overhead. The iterative procedure is outlined in Appendix." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "5922" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 75, + 531, + 281 + ], + "blocks": [ + { + "bbox": [ + 72, + 75, + 531, + 281 + ], + "lines": [ + { + "bbox": [ + 72, + 75, + 531, + 281 + ], + "spans": [ + { + "bbox": [ + 72, + 75, + 531, + 281 + ], + "type": "image", + "image_path": "dab46cc2834897b0ae8093d2cb84654c0cd98c1b055fac3d9008e73b5bb398ee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 291, + 547, + 328 + ], + "lines": [ + { + "bbox": [ + 46, + 291, + 547, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 547, + 328 + ], + "type": "text", + "content": "Figure 3: An overview of architecture generator training process. AG is given a sampled resource constraint B and generates several one-hot policies sequence by sequence. Then, one-hot policies will be mapped to corresponding masks to keep differentiable." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 348, + 288, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 288, + 373 + ], + "type": "text", + "content": "3.3. Generating Architectures with Arbitray Computational Resources" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 382, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 287, + 514 + ], + "type": "text", + "content": "The probability shift is able to sample subnets according to their corresponding training sufficiency. However, a crucial question emerges: given a certain computational resource " + }, + { + "bbox": [ + 46, + 382, + 287, + 514 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 382, + 287, + 514 + ], + "type": "text", + "content": ", how can we rapidly sample a subnet that satisfies the desired resource constraint? A straightforward approach would be to keep sampling subnets until one is found that meets the resource constraint. Nevertheless, as illustrated in Figure 2, subnets with large or small computational resources have an extremely low probability of being sampled, which makes sampling these subnets computationally inefficient." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 518, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 674 + ], + "type": "text", + "content": "To address this issue, we propose an architecture generator (AG) that can provide the corresponding subnet architecture according to any resource constraint. The AG is designed to stably generate a subnet architecture that satisfies the desired resource constraint. Inspired by previous works [35], which formulate the NAS problem as a sequence prediction problem, we also use an LSTM network to generate each operation sequence by sequence. To update the sampling probability " + }, + { + "bbox": [ + 46, + 518, + 287, + 674 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 518, + 287, + 674 + ], + "type": "text", + "content": ", we use differentiable neural architecture search (DNAS) [11, 18] to jointly train the AG and the supernet. DNAS is preferred over other search methods, such as RL-based or evolutionary-based methods, as it converges faster." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "This section presents an overview of the AG (Architecture Generator) training process with clarity and academic rigor. Figure 3 provides a visual representation of the AG" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 349, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 545, + 408 + ], + "type": "text", + "content": "training process. During the training of AG and supernet, the sample distribution " + }, + { + "bbox": [ + 304, + 349, + 545, + 408 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 349, + 545, + 408 + ], + "type": "text", + "content": " remains static and is not updated. To enable end-to-end updates of AG, the policy of each operation is generated using Gumbel Softmax with a one-hot vector, such as " + }, + { + "bbox": [ + 304, + 349, + 545, + 408 + ], + "type": "inline_equation", + "content": "[1,0,0]" + }, + { + "bbox": [ + 304, + 349, + 545, + 408 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 409, + 546, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 546, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 546, + 517 + ], + "type": "text", + "content": "It is important to note that the focus of this paper is on weight entanglement. This means that each operation shares weights for their common parts, and weights with small indices are always activated. As a result, policies cannot be directly involved in forward and backward processes. To overcome this issue, a matrix map trick is employed. More specifically, given an one-hot policy " + }, + { + "bbox": [ + 304, + 409, + 546, + 517 + ], + "type": "inline_equation", + "content": "p \\in 1 \\times n" + }, + { + "bbox": [ + 304, + 409, + 546, + 517 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 409, + 546, + 517 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 409, + 546, + 517 + ], + "type": "text", + "content": " denotes the candidate number in this operation, the mapped policy " + }, + { + "bbox": [ + 304, + 409, + 546, + 517 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 304, + 409, + 546, + 517 + ], + "type": "text", + "content": " can be computed using the following" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 342, + 523, + 545, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 523, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 342, + 523, + 545, + 539 + ], + "type": "interline_equation", + "content": "p ^ {\\prime} = M \\cdot p ^ {T} = \\left[ m _ {0} ^ {T}, m _ {1} ^ {T}, \\dots , m _ {n} ^ {T} \\right] \\cdot p ^ {T} \\tag {9}", + "image_path": "25725cdac913337f4a73dc59ed4bd8af79570875b111cae1a77ed9b0017d3c86.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 545, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 545, + 545, + 652 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 545, + 545, + 652 + ], + "type": "text", + "content": " consists of several masks and " + }, + { + "bbox": [ + 304, + 545, + 545, + 652 + ], + "type": "inline_equation", + "content": "m_i^T" + }, + { + "bbox": [ + 304, + 545, + 545, + 652 + ], + "type": "text", + "content": " represents the mask of " + }, + { + "bbox": [ + 304, + 545, + 545, + 652 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 545, + 545, + 652 + ], + "type": "text", + "content": " th candidate. We show the matrix map example in Figure 3. Assume a ViT block has three heads and the heads number can be selected in [1,2,3], their weight-entangle masks can be designed as " + }, + { + "bbox": [ + 304, + 545, + 545, + 652 + ], + "type": "inline_equation", + "content": "m^0 = [1,0,0], m^1 = [1,1,0]" + }, + { + "bbox": [ + 304, + 545, + 545, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 545, + 545, + 652 + ], + "type": "inline_equation", + "content": "m^2 = [1,1,1]" + }, + { + "bbox": [ + 304, + 545, + 545, + 652 + ], + "type": "text", + "content": ", respectively. With Eq. 9, the one-hot policy " + }, + { + "bbox": [ + 304, + 545, + 545, + 652 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 545, + 545, + 652 + ], + "type": "text", + "content": " participates in the forward of the supernet while the gradient of AG is automatically calculated by the chain rule." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Since our goal is to make AG learn how to accurately generate a subnet architecture with arbitrary resource constraints, the objective function of AG is designed to reduce task loss " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{task}" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " while minimizing the gap " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{RC}" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " between the resource constraint of the sampled subnet and the target." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "5923" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 69, + 531, + 468 + ], + "blocks": [ + { + "bbox": [ + 65, + 69, + 531, + 468 + ], + "lines": [ + { + "bbox": [ + 65, + 69, + 531, + 468 + ], + "spans": [ + { + "bbox": [ + 65, + 69, + 531, + 468 + ], + "type": "table", + "html": "
ModelParameters(M)FLOPs(G)Cost(GPU Days)Top-1 Acc.(%)Top-5 Acc.(%)
DeiT-tiny[28]5.71.22472.291.1
AutoFormer-tiny[4]5.71.33074.792.3
FocusFormer-tiny[19]6.21.42675.193.1
ShiftFormer-T(Ours)5.81.32476.093.1
DeiT-S[28]22.14.73079.995.0
T2T-ViT-14[34]21.56.13281.7-
ViT-S/16[6]22.95.1-78.8-
BoTNet-S1-59[26]33.57.3-81.795.8
AutoFormer-small[4]22.95.13581.495.6
FocusFormer-small[19]23.75.03281.695.6
ShiftFormer-S(Ours)23.65.03082.295.8
DeiT-B[28]86.617.64381.895.6
ViT-B/16[6]86.617.6-79.7-
AutoFormer-base[4]52.811.04381.495.7
FocusFormer-base[19]52.811.04181.995.6
ShiftFormer-B(Ours)52.811.04082.896.1
MobileNetV2 0.75×[24]2.60.211869.8-
MobileNetV3 1.0×[9]5.40.22379175.2-
DS-MBNet-M[16]-0.332473.2-
BigNAS-S[33]4.50.2411276.5-
Once-For-All[3]4.40.2310576.4-
GreedyNAS-C[31]4.70.283276.292.5
ShiftCNN-S(Ours)4.50.243277.293.1
MobileNetV2 1.3×[24]5.30.502172.8-
MobileNetV3 1.25×[9]8.10.35379176.7-
EfficientNet-B0[27]5.40.39379177.193.3
DS-MBNet-S[16]-0.572474.8-
BigNAS-M[33]5.50.4211278.9-
GreedyNAS-A[31]6.50.374077.193.3
ShiftCNN-B(Ours)5.60.423279.693.6
", + "image_path": "422fe1bb55e181b8cecce40a9c07ee58852ce9dfd8954252d41dc45a77262bf5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 478, + 547, + 515 + ], + "lines": [ + { + "bbox": [ + 46, + 478, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 547, + 515 + ], + "type": "text", + "content": "Table 1: ShiftNAS models performance on ImageNet with comparisons to other models. We group the models according to their FLOPs. We use ShiftFormer and ShiftCNN to denote the models searched by ShiftNAS. Cost represents the total GPU days including training, searching and retraining." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 545, + 201, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 545, + 201, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 545, + 201, + 558 + ], + "type": "text", + "content": "Therefore, the joint loss " + }, + { + "bbox": [ + 47, + 545, + 201, + 558 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 545, + 201, + 558 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 101, + 565, + 185, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 565, + 185, + 578 + ], + "spans": [ + { + "bbox": [ + 101, + 565, + 185, + 578 + ], + "type": "interline_equation", + "content": "L = L _ {\\text {t a s k}} + \\lambda L _ {R C}", + "image_path": "1a0d1a01d048bc3341cab1eafb13b68129c1cdc1803c3ae34774578b49476ccc.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 103, + 580, + 287, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 580, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 103, + 580, + 287, + 622 + ], + "type": "interline_equation", + "content": "L _ {R C} = \\left(\\sum_ {i = 1} ^ {D} \\sum_ {j = 1} ^ {n} b _ {j} ^ {i} p _ {j} ^ {\\prime i} - C\\right) ^ {2} \\tag {10}", + "image_path": "5809b07cd740f99fa47e57beb5d52e185baf4e8818468c69446d830a293108e4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "b_{j}^{i}" + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "p_{j}^{\\prime i}" + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": " represents the computational resource and policy of " + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": " th candidate in " + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": " th operation, " + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": " indicates the target resource constraint sampled from sampling distribution B. " + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": " is the coefficient of " + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "L_{RC}" + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": ". In the implementation, we jointly optimize the AG and the supernet on the train dataset with Eq. 10. After few epochs, AG can generate the corresponding subnets with given resource constraint." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 544, + 382, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 382, + 558 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 382, + 558 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 567, + 422, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 567, + 422, + 581 + ], + "spans": [ + { + "bbox": [ + 306, + 567, + 422, + 581 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 590, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 590, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 305, + 590, + 545, + 661 + ], + "type": "text", + "content": "Dataset and evaluation metrics. We conduct main experiments on ImageNet [13] which contains about 1.2M images for training and " + }, + { + "bbox": [ + 305, + 590, + 545, + 661 + ], + "type": "inline_equation", + "content": "50\\mathrm{K}" + }, + { + "bbox": [ + 305, + 590, + 545, + 661 + ], + "type": "text", + "content": " images for validation. We use the Top-1 accuracy and the number of giga floating-point operations(GFLOPs) to measure the performance and efficiency of networks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": "Implementation details. ShiftNAS is a model-agnostic search method. Therefore, we search the models with various resource constraints on both CNN and ViT search space. We split the computational resource range into " + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "K" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "5924" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 175, + 164 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 175, + 164 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 175, + 164 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 175, + 164 + ], + "type": "image", + "image_path": "b401a713a63d725250a45f8a463b5e734c3615e2bdcdf814783740b04aac9a72.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 167, + 130, + 175 + ], + "lines": [ + { + "bbox": [ + 94, + 167, + 130, + 175 + ], + "spans": [ + { + "bbox": [ + 94, + 167, + 130, + 175 + ], + "type": "text", + "content": "(a) ViT-tiny" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 175, + 70, + 302, + 164 + ], + "blocks": [ + { + "bbox": [ + 175, + 70, + 302, + 164 + ], + "lines": [ + { + "bbox": [ + 175, + 70, + 302, + 164 + ], + "spans": [ + { + "bbox": [ + 175, + 70, + 302, + 164 + ], + "type": "image", + "image_path": "07fb239bd32efec950164da70ba4c0e187ded04c6becaad9d38ca2858f023a51.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 167, + 261, + 175 + ], + "lines": [ + { + "bbox": [ + 220, + 167, + 261, + 175 + ], + "spans": [ + { + "bbox": [ + 220, + 167, + 261, + 175 + ], + "type": "text", + "content": "(b) ViT-small" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 186, + 545, + 210 + ], + "lines": [ + { + "bbox": [ + 46, + 186, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 545, + 210 + ], + "type": "text", + "content": "Figure 4: The effectiveness of probability shift. FLOPs/Accuracy tradeoffs of ShiftNAS with and without probability shift on search space (a) ViT-tiny; (b) ViT-small; (c) ViT-base; (d) CNN." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 302, + 70, + 427, + 164 + ], + "blocks": [ + { + "bbox": [ + 302, + 70, + 427, + 164 + ], + "lines": [ + { + "bbox": [ + 302, + 70, + 427, + 164 + ], + "spans": [ + { + "bbox": [ + 302, + 70, + 427, + 164 + ], + "type": "image", + "image_path": "b653b1a2a2b831ab127af11c88c669f6a8a61b4f89a30e5c42da785b9cf183b2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 167, + 391, + 175 + ], + "lines": [ + { + "bbox": [ + 352, + 167, + 391, + 175 + ], + "spans": [ + { + "bbox": [ + 352, + 167, + 391, + 175 + ], + "type": "text", + "content": "(c) ViT-base" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 427, + 70, + 553, + 164 + ], + "blocks": [ + { + "bbox": [ + 427, + 70, + 553, + 164 + ], + "lines": [ + { + "bbox": [ + 427, + 70, + 553, + 164 + ], + "spans": [ + { + "bbox": [ + 427, + 70, + 553, + 164 + ], + "type": "image", + "image_path": "6a503801ba6bd3e394c0218737a52fafacef978a2119ae084f9d7038bf697042.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 167, + 509, + 175 + ], + "lines": [ + { + "bbox": [ + 482, + 167, + 509, + 175 + ], + "spans": [ + { + "bbox": [ + 482, + 167, + 509, + 175 + ], + "type": "text", + "content": "(d) CNN" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 48, + 231, + 293, + 312 + ], + "blocks": [ + { + "bbox": [ + 48, + 231, + 293, + 312 + ], + "lines": [ + { + "bbox": [ + 48, + 231, + 293, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 293, + 312 + ], + "type": "table", + "html": "
ModelInherit(%)Finetune(%)Retrain(%)
ShiftFormer-T76.075.9(-0.1)76.1(+0.1)
ShiftFormer-S82.082.0(0.0)82.0(0.0)
ShiftFormer-B82.482.3(-0.1)82.6(+0.2)
ShiftCNN-S77.277.0(-0.2)77.3(+0.1)
ShiftCNN-B79.679.5(-0.1)79.6(0.0)
", + "image_path": "eafdb08de79c8063608b557a6ef70113e106e513564f446fd150b735cb551bf7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 323, + 286, + 348 + ], + "lines": [ + { + "bbox": [ + 46, + 323, + 286, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 286, + 348 + ], + "type": "text", + "content": "Table 2: Comparison of subnets Top-1 accuracy with inherited, finetuned and retrained weights." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 362, + 287, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 287, + 566 + ], + "type": "text", + "content": "parts with 0.1 GFLOPs step. The sampling distribution will be represented by a " + }, + { + "bbox": [ + 46, + 362, + 287, + 566 + ], + "type": "inline_equation", + "content": "1 \\times K" + }, + { + "bbox": [ + 46, + 362, + 287, + 566 + ], + "type": "text", + "content": " learnable vector. We randomly split " + }, + { + "bbox": [ + 46, + 362, + 287, + 566 + ], + "type": "inline_equation", + "content": "50\\mathrm{K}" + }, + { + "bbox": [ + 46, + 362, + 287, + 566 + ], + "type": "text", + "content": " images from ImageNet and use Adam optimizer with 1e-3 learning rate to update the sampling distribution vector. The AG consists of an LSTM cell with 64 hidden numbers and 4/10 fully connected layers for ViT/CNN supernet and will be updated jointly with the supernet. In the first 50 training epochs, we only optimize the supernet weight and the AG since the AG cannot generate the corresponding subnets at the beginning. After 50 epochs, the sampling distribution vector is updated per 100 iteration. We discuss more hyperparameter settings in the Appendix, such as the split step of the computational resource and the updated frequency of the sampling distribution vector. The training strategies for both CNN and ViT are given in Appendix. All the models are trained on 8 Nvidia Tesla A100 GPUs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 567, + 286, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 286, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 286, + 687 + ], + "type": "text", + "content": "For CNN search space, we follow BigNAS [33] where the search space contains kernel size, channel number, depth and input resolution. The supernet will be split into 7 stages. Each stage has multiple choices of the block number and the first block of each stage has no residual path. The policy of kernel size will be generated block-wise, where the small kernel size is center cropped from the large kernel size. The policies of channel number and depth will be predicted stage-wise, where the lower-index channels and blocks are preferentially kept." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "For ViT search space, we follow AutoFormer [4] which searches embed dim, head number, MLP ratio and depth" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 232, + 545, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 267 + ], + "type": "text", + "content": "with three different scale settings. The policies of head number and MLP ratio will be given block-wise. As CNN, the lower-index heads and neurons are preferentially kept." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 267, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 327 + ], + "type": "text", + "content": "After training, we directly sample the candidate with the max probability given by the AG. It is noted that the obtained model inherits weights from the trained supernet without retraining or fine-tuning. Therefore, no search or retraining cost exists in ShiftNAS." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 335, + 390, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 390, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 390, + 346 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 353, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 545, + 640 + ], + "type": "text", + "content": "We search ViT and CNN architectures on ImageNet with different FLOPs constraints. We compare searched models with multiple ViT and CNN models on model performance (Top-1 Acc., Top-5 Acc.) and efficiency (FLOPs, parameters). It is seen from Table 1 that our ShiftFormer models surpass the recent manual and autoML-based transformer models under diverse model sizes. It is worth noting that ShiftFormer-tiny achieves a top-1 Acc. of 76.0 with only 1.3 GFLOPs, being " + }, + { + "bbox": [ + 304, + 353, + 545, + 640 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 304, + 353, + 545, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 353, + 545, + 640 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 304, + 353, + 545, + 640 + ], + "type": "text", + "content": " better than AutoFormer-tiny and FocusFormer-T which are also autoML-based methods, respectively. Compared to CNN models, ShiftCNN models also achieve better accuracy than all compared networks under similar FLOPs restrictions. For example, ShiftCNN-S and ShiftCNN-B achieve " + }, + { + "bbox": [ + 304, + 353, + 545, + 640 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 304, + 353, + 545, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 353, + 545, + 640 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 304, + 353, + 545, + 640 + ], + "type": "text", + "content": " higher top-1 accuracy than BigNAS-S and BigNAS-M, respectively. Specifically, ShiftCNN and BigNAS models are searched from the same search space. BigNAS models spend more than " + }, + { + "bbox": [ + 304, + 353, + 545, + 640 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 304, + 353, + 545, + 640 + ], + "type": "text", + "content": " GPU time than ShiftCNN models since the sandwich rule is used to train BigNAS supernet, which demonstrates the efficiency and outstanding performance of ShiftNAS. For fairness, we do not compare our models with others obtained from better search spaces [15, 29] in Table 1. The experiments on AttentiveNAS search space can be found in the Appendix." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 647, + 399, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 399, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 399, + 659 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "The effectiveness of probability shift. To demonstrate the effectiveness of the proposed sampling probability shift, we trained a ShiftNAS supernet and its baseline counterpart under ViT-tiny, ViT-small, ViT-base, and CNN search" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "5925" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 173, + 172 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 173, + 172 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 173, + 172 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 173, + 172 + ], + "type": "image", + "image_path": "b5562594818d2f7f6de501150e4e7e4a00017b0ddeedb956bc27cf742d07e415.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 182, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 545, + 205 + ], + "type": "text", + "content": "Figure 5: The sampling probability with different FLOPs groups on search space (a) ViT-tiny; (b) ViT-small; (c) ViT-base; (d) CNN." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 299, + 173 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 299, + 173 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 299, + 173 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 299, + 173 + ], + "type": "image", + "image_path": "dd20aabde47e64996b0068bd39f23c97ad57c5c6282821fc0856fca9036bfa4d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 299, + 70, + 426, + 173 + ], + "blocks": [ + { + "bbox": [ + 299, + 70, + 426, + 173 + ], + "lines": [ + { + "bbox": [ + 299, + 70, + 426, + 173 + ], + "spans": [ + { + "bbox": [ + 299, + 70, + 426, + 173 + ], + "type": "image", + "image_path": "929abfbae529bd8c1aa5c83ed4c5bbb247a9fcd86a116d37dbf9cf6076da51c8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 427, + 70, + 553, + 173 + ], + "blocks": [ + { + "bbox": [ + 427, + 70, + 553, + 173 + ], + "lines": [ + { + "bbox": [ + 427, + 70, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 427, + 70, + 553, + 173 + ], + "type": "image", + "image_path": "2d79223fc352330bc002925a085337fd9186e29180aff94bdca2e2adb5439be8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 65, + 224, + 269, + 380 + ], + "blocks": [ + { + "bbox": [ + 65, + 224, + 269, + 380 + ], + "lines": [ + { + "bbox": [ + 65, + 224, + 269, + 380 + ], + "spans": [ + { + "bbox": [ + 65, + 224, + 269, + 380 + ], + "type": "image", + "image_path": "78bd1c72ac16b7d02f9ef833409f3a6a7e46e1341186558701ae38c350393970.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 388, + 286, + 400 + ], + "lines": [ + { + "bbox": [ + 47, + 388, + 286, + 400 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 286, + 400 + ], + "type": "text", + "content": "Figure 6: The ranking correlation of architecture generator." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 424, + 286, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 286, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 286, + 532 + ], + "type": "text", + "content": "spaces. In the baseline supernet training, we uniformly sampled FLOPs in each step and did not update FLOPs probability. The training settings for all supernets are provided in the Appendix. As illustrated in Figure 4, the supernet trained with probability shift is capable of covering a wide range of accuracy-FLOPs curves and outperforms the baseline supernet trained without probability shift by a significant margin, demonstrating the effectiveness of the proposed sampling probability shift mechanism." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 533, + 286, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 286, + 688 + ], + "type": "text", + "content": "The ranking correlation of architecture generator. The experiment is based on ShiftFormer-tiny supernet. We measure the ranking correlation by randomly sampling 25 subnets from the ShiftFormer-tiny supernet. For fast evaluation, The sampled subnets are inherited weights from the supernet and are further finetuned to reach the optimal weights. As for the model ranking test, we evaluate sampled subnets in both supernets to get their inherited performance. As shown in Figure 6, the inherited performance is highly correlated with finetuned performance in good subnets (Inherited Top1-Acc > 65%). Despite the correlation with bad subnets is not obvious, we focus more attention on good subnets since bad subnets are not deployed in practice." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "Subnet performance without retraining. We further study the performance of ShiftNAS models when models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 227, + 545, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 227, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 545, + 321 + ], + "type": "text", + "content": "are independently fine-tuned or trained from scratch. Following Autoformer, we finetune or retrain the selected subnet with 40 epochs or 300 epochs on ImageNet. Table 2 shows that finetuning cannot improve or even harm the performance of subnets inherited from supernet. Besides, retraining can slightly improve the subnet performance but cost extra training resources which prevents the implementation of NAS." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 323, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 467 + ], + "type": "text", + "content": "The optimal sampling strategy. ShiftNAS learns the optimal sampling probability for each subnet with different FLOPs. To figure out what sampling strategy is the best, we visualize the various sampling probability in different search spaces (ViT-tiny, ViT-small, ViT-base and CNN) at different training steps. As can be seen in Figure 5, our method learns to sample more subnets with large computational complexity. For example, the sampling probabilities of 1.6 to 1.9 GFLOPs subnets are clearly higher than 1.3 and 1.4 GFLOPs subnets in the ViT-tiny search space. This is intuitive since the large subnet contains more parameters and needs more computational resources to update." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 478, + 378, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 378, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 378, + 490 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 498, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 629 + ], + "type": "text", + "content": "In this paper, we presented ShiftNAS, a one-shot supernet training framework that can complexity-wise adjust the sampling probability. To automatically adjust the sampling probability, we proposed probability shift that can be learned according to the subnets training sufficiency. To fast obtain the subnets with desired complexity, AG is designed that can be jointly trained with the supernet in an end-to-end manner. Experiments on ImageNet showed that our method achieves SOTA results on ViT/CNN search spaces, and covers a wide range of efficiency/accuracy trade-offs without any extra retraining." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 666 + ], + "type": "text", + "content": "Limitation. ShiftNAS is designed for efficient search on weight-entanglement search spaces, so it cannot work on darts-like search spaces [18, 25]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements: This work was supported by the National Natural Science Foundation of China (62203392) and Natural Science Foundation of Zhejiang province (LY21F030018)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "5926" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Gabriel Bender, Pieter-Jan Kindermans, Barret Zoph, Vijay Vasudevan, and Quoc Le. Understanding and simplifying one-shot architecture search. In International conference on machine learning, pages 550-559. PMLR, 2018. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "type": "text", + "content": "[2] Han Cai, Chuang Gan, Ji Lin, and Song Han. Network augmentation for tiny deep learning. arXiv preprint arXiv:2110.08890, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 288, + 215 + ], + "type": "text", + "content": "[3] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. In International Conference on Learning Representations, 2019. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 271 + ], + "type": "text", + "content": "[4] Minghao Chen, Houwen Peng, Jianlong Fu, and Haibin Ling. Autoformer: Searching transformers for visual recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12270-12280, 2021. 2, 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 288, + 316 + ], + "type": "text", + "content": "[5] Xiangxiang Chu, Bo Zhang, Ruijun Xu, and Jixiang Li. Fairnas: Rethinking evaluation fairness of weight sharing neural architecture search. arXiv preprint arXiv:1907.01845, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 319, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 288, + 384 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 386, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 288, + 441 + ], + "type": "text", + "content": "[7] Zichao Guo, Xiangyu Zhang, Haoyuan Mu, Wen Heng, Zechun Liu, Yichen Wei, and Jian Sun. Single path one-shot neural architecture search with uniform sampling. In European Conference on Computer Vision, pages 544-560. Springer, 2020. 1, 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 288, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 288, + 486 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 288, + 486 + ], + "type": "text", + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 288, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 288, + 552 + ], + "type": "text", + "content": "[9] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 1, 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 288, + 609 + ], + "type": "text", + "content": "[10] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 611, + 288, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 288, + 666 + ], + "type": "text", + "content": "[11] Shoukang Hu, Sirui Xie, Hehui Zheng, Chunxiao Liu, Jianping Shi, Xunying Liu, and Dahua Lin. Dsnas: Direct neural architecture search without parameter retraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12084-12092, 2020. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 667, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 667, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 667, + 288, + 689 + ], + "type": "text", + "content": "[12] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. 2016. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[13] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural net" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "works. Advances in neural information processing systems, 25:1097-1105, 2012. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 161 + ], + "type": "text", + "content": "[14] Changlin Li, Jiefeng Peng, Liuchun Yuan, Guangrun Wang, Xiaodan Liang, Liang Lin, and Xiaojun Chang. Blockwisely supervised neural architecture search with knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1989-1998, 2020. 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 228 + ], + "type": "text", + "content": "[15] Changlin Li, Tao Tang, Guangrun Wang, Jiefeng Peng, Bing Wang, Xiaodan Liang, and Xiaojun Chang. Bossnas: Exploring hybrid cnn-transformers with block-wisely self-supervised neural architecture search. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12281-12291, 2021. 4, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 230, + 547, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 547, + 285 + ], + "type": "text", + "content": "[16] Changlin Li, Guangrun Wang, Bing Wang, Xiaodan Liang, Zhihui Li, and Xiaojun Chang. Ds-net++: Dynamic weight slicing for efficient inference in cnns and vision transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-16, 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 286, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 547, + 319 + ], + "type": "text", + "content": "[17] Liam Li and Ameet Talwalkar. Random search and reproducibility for neural architecture search. In Uncertainty in artificial intelligence, pages 367-377. PMLR, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "type": "text", + "content": "[18] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. In International Conference on Learning Representations, 2018. 2, 4, 5, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "text", + "content": "[19] Jing Liu, Jianfei Cai, and Bohan Zhuang. Focusformer: Focusing on what we need via architecture sampler. arXiv preprint arXiv:2208.10861, 2022. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 388, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 443 + ], + "type": "text", + "content": "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "text", + "content": "[21] James E Marengo, David L Farnsworth, and Lucas Stefanic. A geometric derivation of the irwin-hall distribution. International Journal of Mathematics and Mathematical Sciences, 2017, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "type": "text", + "content": "[22] Yameng Peng, Andy Song, Vic Ciesielski, Haytham M. Fayek, and Xiaojun Chang. Pre-nas: Evolutionary neural architecture search with predictor. IEEE Transactions on Evolutionary Computation, 27(1):26-36, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 534, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 578 + ], + "type": "text", + "content": "[23] Hieu Pham, Melody Guan, Barret Zoph, Quoc Le, and Jeff Dean. Efficient neural architecture search via parameters sharing. In International conference on machine learning, pages 4095-4104. PMLR, 2018. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 579, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 634 + ], + "type": "text", + "content": "[24] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 1, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "text", + "content": "[25] Julien Siems, Lucas Zimmer, Arber Zela, Jovita Lukasik, Margret Keuper, and Frank Hutter. Nas-bench-301 and the case for surrogate benchmarks for neural architecture search. arXiv preprint arXiv:2008.09777, 2020. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 681, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 547, + 713 + ], + "type": "text", + "content": "[26] Aravind Srinivas, Tsung-Yi Lin, Niki Parmar, Jonathon Shlens, Pieter Abbeel, and Ashish Vaswani. Bottleneck transformers for visual recognition. In Proceedings of" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "5927" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 574 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 16519-16529, 2021. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[27] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "type": "text", + "content": "[28] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "type": "text", + "content": "[29] Dilin Wang, Meng Li, Chengyue Gong, and Vikas Chandra. Attentivenas: Improving neural architecture search via attentive sampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6418-6427, 2021. 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 319 + ], + "type": "text", + "content": "[30] Bichen Wu, Xiaoliang Dai, Peizhao Zhang, Yanghan Wang, Fei Sun, Yiming Wu, Yuandong Tian, Peter Vajda, Yangqing Jia, and Kurt Keutzer. Fbnet: Hardware-aware efficient convnet design via differentiable neural architecture search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10734-10742, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 373 + ], + "type": "text", + "content": "[31] Shan You, Tao Huang, Mingmin Yang, Fei Wang, Chen Qian, and Changshui Zhang. Greedynas: Towards fast one-shot nas with greedy supernet. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1999-2008, 2020. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 375, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 375, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 375, + 287, + 407 + ], + "type": "text", + "content": "[32] Jiahui Yu and Thomas Huang. Autoslim: Towards one-shot architecture search for channel numbers. arXiv preprint arXiv:1903.11728, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 409, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 287, + 474 + ], + "type": "text", + "content": "[33] Jiahui Yu, Pengchong Jin, Hanxiao Liu, Gabriel Bender, Pieter-Jan Kindermans, Mingxing Tan, Thomas Huang, Xiaodan Song, Ruoming Pang, and Quoc Le. Bignas: Scaling up neural architecture search with big single-stage models. In European Conference on Computer Vision, pages 702-717. Springer, 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 475, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 287, + 540 + ], + "type": "text", + "content": "[34] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF international conference on computer vision, pages 558-567, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 542, + 287, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 542, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 542, + 287, + 574 + ], + "type": "text", + "content": "[35] Barret Zoph and Quoc V Le. Neural architecture search with reinforcement learning. arXiv preprint arXiv:1611.01578, 2016. 1, 3, 5" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "5928" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Shortcut-V2V_ Compression Framework for Video-to-Video Translation Based on Temporal Redundancy Reduction/99da4ac0-6f5c-433f-bc7b-caa55aa672d5_content_list.json b/2023/Shortcut-V2V_ Compression Framework for Video-to-Video Translation Based on Temporal Redundancy Reduction/99da4ac0-6f5c-433f-bc7b-caa55aa672d5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..39eae34b1690188f58914e4b524b17388b4a4264 --- /dev/null +++ b/2023/Shortcut-V2V_ Compression Framework for Video-to-Video Translation Based on Temporal Redundancy Reduction/99da4ac0-6f5c-433f-bc7b-caa55aa672d5_content_list.json @@ -0,0 +1,1574 @@ +[ + { + "type": "text", + "text": "Shortcut-V2V: Compression Framework for Video-to-Video Translation based on Temporal Redundancy Reduction", + "text_level": 1, + "bbox": [ + 122, + 128, + 848, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chaeyeon Chung\\*, Yeojeong Park\\*, Seunghwan Choi, Munkhsoyol Ganbat, and Jaegul Choo", + "bbox": [ + 91, + 214, + 877, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ KAIST AI $^{2}$ KT Research & Development Center, KT Corporation {cy-chung, indigopyj, shadow2496, soyol, jchoo}@kaist.ac.kr", + "bbox": [ + 228, + 244, + 746, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 340, + 313, + 356 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video-to-video translation aims to generate video frames of a target domain from an input video. Despite its usefulness, the existing networks require enormous computations, necessitating their model compression for wide use. While there exist compression methods that improve computational efficiency in various image/video tasks, a generally-applicable compression method for video-to-video translation has not been studied much. In response, we present Shortcut-V2V, a general-purpose compression framework for video-to-video translation. Shortcut-V2V avoids full inference for every neighboring video frame by approximating the intermediate features of a current frame from those of the previous frame. Moreover, in our framework, a newly-proposed block called AdaBD adaptively blends and deforms features of neighboring frames, which makes more accurate predictions of the intermediate features possible. We conduct quantitative and qualitative evaluations using well-known video-to-video translation models on various tasks to demonstrate the general applicability of our framework. The results show that Shortcut-V2V achieves comparable performance compared to the original video-to-video translation model while saving $3.2 - 5.7 \\times$ computational cost and $7.8 - 44 \\times$ memory at test time. Our code and videos are available at https://shortcut-v2v.github.io/.", + "bbox": [ + 75, + 373, + 472, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 768, + 209, + 786 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video-to-video translation is a task of generating temporally consistent and realistic video frames of a target domain from a given input video. Recent studies on video-to-video translation present promising performance in various domains such as inter-modality translation between labels", + "bbox": [ + 75, + 796, + 468, + 872 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and videos [34, 33, 18], and intra-modality translation between driving scene videos [32] or face videos [1].", + "bbox": [ + 500, + 340, + 890, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite enhanced usefulness, video-to-video translation networks usually require substantial computational cost and memory usage, which limits their applicability. For instance, multiply-accumulates (MACs) of a widely-used video translation model, vid2vid [34], is 2066.69G, while the basic convolutional neural networks, ResNet v2 50 [10] and Inception v3 [28], are 4.12G and 6G, respectively. Furthermore, temporally redundant computations for adjacent video frames also harm the cost efficiency of a video-to-video translation network. Performing full inference for every neighboring video frame that contains common visual features inevitably entails redundant operations [20, 19].", + "bbox": [ + 496, + 373, + 892, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this regard, Fast-Vid2Vid [40] proposes a compression framework for vid2vid [34] based on spatial input compression and temporal redundancy reduction. However, it cannot be applied to other video-to-video translation models since it is designed specifically for vid2vid. Moreover, Fast-Vid2Vid does not support real-time inference since it requires future frames to infer a current one. Alternatively, one can apply model compression approaches for image-to-image translation [13, 22, 12] directly to video-to-video translation, considering video frames as separate images. However, these approaches are not designed to consider the correlation among adjacent video frames during the compression. This may result in unrealistic output quality in video-to-video translation, where the inherent temporal coherence of an input video needs to be preserved in the outputs. Also, frame-by-frame inference without temporal redundancy reduction involves unnecessary computations, resulting in computational inefficiency.", + "bbox": [ + 496, + 553, + 893, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose Shortcut-V2V, a general-purpose framework for improving the computational efficiency of video-to-video translation based on temporal redundancy reduction. Shortcut-V2V allows the original video-to-video translation model to avoid temporally re", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*indicates equal contributions.", + "bbox": [ + 94, + 886, + 258, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7612", + "bbox": [ + 482, + 944, + 516, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ccd7fcbcdba55580a15d96a0c2daea7d8ec466b9c324b414c912480b8bce99f5.jpg", + "image_caption": [ + "Figure 1: Overview of Shortcut-V2V. (a) is an overall framework of Shorcut-V2V, and (b) shows a detailed architecture of Shortcut block. $\\uparrow 2$ and $\\downarrow 2$ refer to upsampling and downsampling by a factor of 2, respectively. G in Offset G and Offset/Mask G indicates a generator." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 436, + 255 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e709cafb05b43f5baf3babe234028bcd445d0aee9731a366389b0916160169dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 90, + 890, + 255 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dundant computations by approximating the decoding layer features of the current frame with largely reduced computations. To enable lightweight estimation, our framework leverages features from the previous frame (i.e., reference features), which have high visual similarity with the current frame. We also exploit current frame features from the encoding layer to handle newly-appeared regions in the current frame. Specifically, we first globally align the previous frame features with the current frame features, and our novel Adaptive Blending and Deformation block (AdaBD) in Shortcut-V2V blends features of neighboring frames while performing detailed deformation. AdaBD adaptively integrates the features regarding their redundancy in a lightweight manner.", + "bbox": [ + 75, + 340, + 468, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this way, our model significantly improves the test-time efficiency of the original network while preserving its original performance. Shortcut-V2V is easily applicable to a pretrained video-to-video translation model to save computational cost and memory usage. Our framework is also suitable for real-time inference since we do not require future frames for the current frame inference. To the best of our knowledge, this is the first attempt at a general-purpose model compression approach for video-to-video translation. We demonstrate the effectiveness of our approach using well-known video-to-video translation models, Unsupervised RecycleGAN [32] (Unsup) and vid2vid [34]. Shortcut-V2V reduces $3.2 - 5.7 \\times$ computational cost and $7.8 - 44 \\times$ memory usage while achieving comparable performance to the original model. Since there is no existing general-purpose compression method, we compare our method with Fast-Vid2Vid [40] and the compression methods for image-to-image translation. Our model presents superiority over the existing approaches in both quantitative and qualitative evaluations. Our contributions are summarized as follows:", + "bbox": [ + 75, + 553, + 470, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We introduce a novel, general-purpose model com", + "bbox": [ + 94, + 885, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pression framework for video-to-video translation, Shortcut-V2V, that enables the original network to avoid temporally redundant computations.", + "bbox": [ + 529, + 340, + 890, + 385 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present AdaBD that exploits features from neighboring frames via adaptive blending and deformation in a lightweight manner.", + "- Our framework saves up to $5.7 \\times$ MACs and $44 \\times$ parameters across various video-to-video translation tasks, achieving comparable performance to the original networks." + ], + "bbox": [ + 517, + 398, + 890, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 532, + 640, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Video-to-Video Translation", + "text_level": 1, + "bbox": [ + 500, + 559, + 743, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent video-to-video translation networks are generally classified into pix2pixHD-based [35] and CycleGAN-based [38] generator. vid2vid [34] proposes a pix2pixHD-based sequential generation framework that synthesizes a current output given the previous outputs as additional guidance. As the following work, few-shot vid2vid [33] achieves few-shot generalization of vid2vid based on attention modules, and world-consistent vid2vid [18] is proposed to improve long-term temporal consistency of vid2vid.", + "bbox": [ + 496, + 582, + 890, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While pix2pixHD-based models require paired annotated videos, RecycleGAN [1] and MocycleGAN [4] propose CycleGAN-based video translation models. They exploit spatio-temporal consistency losses to generate realistic videos using unpaired datasets. STC-V2V [21] leverages optical flow for semantic/temporal consistency to improve the output quality of the existing models. Unsupervised RecycleGAN [32] achieves state-of-the-art performance among CycleGAN-based frameworks with a pseudosupervision by the synthetic flow. Although the existing video-to-video translation networks achieve decent performance, they commonly demand a non-trivial amount of", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7613", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "computational costs and memory usage. Also, frame-by-frame inference necessarily causes redundant operations due to temporal redundancy among adjacent frames.", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Model Compression", + "text_level": 1, + "bbox": [ + 76, + 147, + 266, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model compression for video-related tasks has been actively proposed in various domains [19, 20, 16, 9, 27, 8, 14, 29], such as object detection, action recognition, semantic segmentation, and super-resolution. Several studies [19, 9, 20] exploit temporal redundancy among video frames to improve efficiency during training or inference. For instance, Habibian et al. [9] distill only the residual between adjacent frames from a teacher model to a student to speed up the inference. Also, Fast-Vid2Vid [40] firstly proposes a compression framework for video-to-video translation based on spatial and temporal compression. However, Fast-Vid2Vid focuses on vid2vid [34], limiting its application to other video-to-video translation networks. Also, temporal redundancy reduction via motion compensation in Fast-Vid2Vid requires the future frame to infer the current frame, which is not suitable for real-time inference.", + "bbox": [ + 75, + 170, + 468, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Alternatively, model-agnostic compression methods for image-to-image translation can also be applied to video-to-video translation models. The existing approaches for image synthesis mainly tackle channel pruning [13], knowledge distillation [12, 22], NAS [13], etc. For instance, CAT [12] compresses the teacher network with one-step pruning to satisfy the target computation budget, while OMGD [22] conducts a single-stage online distillation in which the teacher generator supports the student generator to be refined progressively. However, image-based compression methods cannot consider temporal coherence among neighboring video frames, which may induce unrealistic results in video translation tasks. Also, performing full model inference for each video frame still poses computational inefficiency due to the temporal redundancy.", + "bbox": [ + 75, + 412, + 468, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Deformable Convolution", + "text_level": 1, + "bbox": [ + 76, + 650, + 302, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deformable convolution [6] is originally proposed to enhance the transformation capability of a convolutional layer by adding the estimated offsets to a regular convolutional kernel in vision tasks such as object detection or semantic segmentation. Besides its original application, recent studies [30, 7, 36, 2, 11, 15] demonstrate that deformable convolution is also capable of aligning adjacent video frames. TDAN [30] utilizes deformable convolution to capture implicit motion cues between consecutive frames by dynamically predicted offsets in video super-resolution. In addition, EDVR [36] stacks several deformable convolution blocks to estimate large and complex motions in video restoration. In this paper, we also leverage a deformable convolution to adaptively align adjacent video frames in a lightweight manner.", + "bbox": [ + 75, + 674, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Shortcut-V2V Inference" + ], + "code_body": "1: Input: Input video $\\{\\mathbf{I}_t\\}_{t=0}^{N_T-1}$ of length $N_T$ , teacher model $T$ , layer index of encoder $l_e$ and decoder $l_d$ , Shortcut block $S$ , max interval $\\alpha$ \n2: Output: Output video $\\{\\mathbf{O}_t\\}_{t=0}^{N_T-1}$ \n3: for $t = 0$ to $N_T - 1$ do \n4: $\\mathbf{a}_t = T_{[le]}(\\mathbf{I}_t)$ \n5: if $t\\% \\alpha = 0$ then \n6: $\\mathbf{f}_t = T_{[le + 1;l_d - 1]}(\\mathbf{a}_t)$ \n7: $\\mathbf{a}_{\\mathrm{ref}}, \\mathbf{f}_{\\mathrm{ref}} \\gets \\mathbf{a}_t, \\mathbf{f}_t \\triangleright$ Update the reference features \n8: else \n9: $\\mathbf{f}_t = S(\\mathbf{f}_{\\mathrm{ref}}, \\mathbf{a}_{\\mathrm{ref}}, \\mathbf{a}_t)$ \n10: end if \n11: $\\mathbf{O}_t = T_{[l_d,]}(\\mathbf{f}_t)$ \n12: end for", + "bbox": [ + 506, + 125, + 890, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Shortcut-V2V", + "text_level": 1, + "bbox": [ + 500, + 366, + 640, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we propose Shortcut-V2V, a general compression framework to improve the test-time efficiency in video-to-video translation. As illustrated in Fig. 1(a), given $\\{\\mathbf{I}_t\\}_{t=0}^{N_T-1}$ as input video frames, we first use full teacher model $T$ to synthesize the output of the first frame. Then, for the next frames, our newly-proposed Shortcut block efficiently approximates $\\mathbf{f}_t$ , the features from the $l_d$ -th decoding layer of the teacher model. This is achieved by leveraging the $l_e$ -th encoding layer features $\\mathbf{a}_t$ along with reference features, $\\mathbf{a}_{ref}$ and $\\mathbf{f}_{ref}$ , from the previous frame. Here, $l_d$ and $l_e$ correspond to layer indices of the teacher model. Lastly, predicted features $\\hat{\\mathbf{f}}_t$ are injected into the following layers of the teacher model to synthesize the final output $\\hat{\\mathbf{O}}_t$ . To avoid error accumulation, we conduct full teacher inference and update the reference features at every max interval $\\alpha$ . We provide the detailed inference process of Shortcut-V2V in Algorithm 1.", + "bbox": [ + 496, + 392, + 890, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The architecture of our model is mainly inspired by Deformable Convolutional Network (DCN) [6, 39], which we explain in the next section.", + "bbox": [ + 496, + 650, + 890, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Deformable Convolutional Network", + "text_level": 1, + "bbox": [ + 498, + 709, + 808, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DCN [6, 39] is initially introduced to improve the transformation capability of a convolutional layer in image-based vision tasks, e.g., object detection and semantic segmentation. In the standard convolutional layer, a $3 \\times 3$ kernel with dilation 1 samples points over input features using a sampling position $\\mathbf{p}_k \\in \\{(-1, -1), (-1, 0), \\dots, (0, 1), (1, 1)\\}$ . Given input feature maps $\\mathbf{x}$ , DCN predicts additional offsets $\\Delta \\mathbf{p} \\in \\mathbb{R}^{2N_{\\mathbf{p}} \\times H \\times W}$ to augment each sampling position along x-axis and y-axis. Here, $N_{\\mathbf{p}}$ is the number of sampling positions in a kernel, and $H$ and $W$ are the height and width", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7614", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of output feature maps, respectively. For further manipulation of input feature amplitudes over the sampled points, DCNv2 [39] introduces a modulated deformable convolution with $\\mathbf{m} \\in \\mathbb{R}^{N_{\\mathbf{P}} \\times H \\times W}$ consisting of learnable modulation scalars. The deformed output feature maps $\\mathbf{x}'$ by DCNv2 are defined as:", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {\\prime} = f _ {d c} (\\mathbf {w}, \\mathbf {x}, \\Delta \\mathbf {p}, \\mathbf {m}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 191, + 468, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{dc}$ indicates a deformable convolution. Specifically, a single point $\\mathbf{p_o}$ of $\\mathbf{x}'$ is obtained as:", + "bbox": [ + 76, + 220, + 467, + 251 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {\\prime} \\left(\\mathbf {p} _ {\\mathbf {o}}\\right) = \\sum_ {k = 1} ^ {N _ {\\mathbf {p}}} \\mathbf {w} \\left(\\mathbf {p} _ {k}\\right) \\cdot \\mathbf {x} \\left(\\mathbf {p} _ {\\mathbf {o}} + \\mathbf {p} _ {k} + \\Delta \\mathbf {p} \\left(\\mathbf {p} _ {k}\\right)\\right) \\cdot \\mathbf {m} \\left(\\mathbf {p} _ {k}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 261, + 468, + 304 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{w}(\\mathbf{p}_k),\\mathbf{p}(\\mathbf{p}_k)$ , and $\\mathbf{m}(\\mathbf{p}_k)$ are convolutional layer weights, offsets, and modulation scalars between 0 and 1, respectively, for the $k$ -th sampling position.", + "bbox": [ + 76, + 314, + 468, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Taking advantage of the enhanced transformation capability, we also leverage deformable convolution to align features of adjacent frames only with a few convolution-like operations, instead of using heavy flow estimation networks.", + "bbox": [ + 75, + 359, + 468, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Shortcut Block", + "text_level": 1, + "bbox": [ + 76, + 444, + 228, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As described in Fig. 1(b), Shortcut block $S$ estimates the current frame features $\\hat{\\mathbf{f}}_t$ given $\\mathbf{a}_t$ and the reference frame features $\\mathbf{f}_{ref}$ and $\\mathbf{a}_{ref}$ as inputs:", + "bbox": [ + 76, + 467, + 468, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {f}} _ {t} = S \\left(\\mathbf {f} _ {r e f}, \\mathbf {a} _ {r e f}, \\mathbf {a} _ {t}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 523, + 468, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our block effectively obtains rich information from $\\mathbf{f}_{ref}$ via coarse-to-fine alignment referring to alignment between $\\mathbf{a}_{ref}$ and $\\mathbf{a}_t$ . Also, during the fine alignment, our newly-proposed AdaBD simultaneously performs adaptive blending of the aligned $\\mathbf{f}_{ref}$ and the current frame feature $\\mathbf{a}_t$ . Here, $\\mathbf{a}_t$ supports the synthesis of newly-appeared areas in the current frame.", + "bbox": [ + 76, + 551, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Coarse-to-Fine Alignment. To handle a wide range of misalignments between the frames, our model aligns $\\mathbf{f}_{ref}$ with the current frame in a coarse-to-fine manner. Our global/local alignment module consists of an offset generator to estimate offsets, and deformable convolution layers to deform features based on the predicted offsets. Following TDAN [30], an offset generator estimates sampling offsets given the adjacent frame features. For global alignment, we first downsample the given inputs to enlarge the receptive fields of the corresponding convolutional layers in a lightweight manner. The downsampled $\\mathbf{a}_{ref}$ and $\\mathbf{a}_t$ are concatenated and fed into a global offset generator to generate global offsets $\\Delta \\mathbf{p}_g\\in \\mathbb{R}^{2\\times \\frac{H}{2}\\times \\frac{W}{2}}$ . Since we only need to capture coarse movement, $\\Delta \\mathbf{p}_g$ includes a single offset for each kernel, unlike the original DCN. Each offset is identically applied to all the sampling positions within the", + "bbox": [ + 75, + 659, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "kernel. Then, the deformed features are upsampled back to the original size to obtain $\\mathbf{f}_{ref}^{\\prime}$ as follows:", + "bbox": [ + 498, + 90, + 890, + 123 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {r e f} ^ {\\prime} = \\left(f _ {d c} \\left(\\mathbf {w} _ {g}, \\left(\\mathbf {f} _ {r e f}\\right) ^ {\\downarrow 2}, \\Delta \\mathbf {p} _ {g}, \\mathbf {1})\\right) ^ {\\uparrow 2}, \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 133, + 890, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{w}_g$ denotes weights of global deformable convolution, and $(\\cdot)^{\\uparrow 2}$ and $(\\cdot)^{\\downarrow 2}$ refer to upsampling and downsampling by a factor of 2 through bilinear interpolation, respectively. 1 indicates a vector filled with 1 so that no modulation is applied here. While local alignment of the coarsely-aligned feature $\\mathbf{f}_{ref}^{\\prime}$ follows the process of global alignment, the difference lies in that each sampling point of each kernel has a unique offset, and the alignment operation is conducted in the original resolution. We leverage $\\mathbf{a}_{ref}^{\\prime}$ and $\\mathbf{a}_t$ to estimate local offsets, where $\\mathbf{a}_{ref}^{\\prime}$ is the aligned $\\mathbf{a}_{ref}$ which is downsampled, deformed, and upsampled with the same weights $\\mathbf{w}_g$ used to synthesize $\\mathbf{f}_{ref}^{\\prime}$ .", + "bbox": [ + 496, + 164, + 890, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We estimate the offsets for the decoding layer features $\\mathbf{f}$ using the encoding layer features a under the assumption that $\\mathbf{a}_t$ and $\\mathbf{f}_t$ have the same structural information. In video-to-video translation, input and output frames share the same underlying structure. Thus, it is natural for the network to learn to maintain the structural information of an input frame throughout the encoding and decoding process. More details are described in our supplementary materials. Adaptive Blending and Deformation. During the local alignment, we also take advantage of the current frame features $\\mathbf{a}_t$ from the encoding layer to handle the regions with large motion differences and new objects. To achieve this in a cost-efficient way, we introduce AdaBD, which simultaneously aligns $\\mathbf{f}_{ref}^{\\prime}$ and blends it with $\\mathbf{a}_t$ in an adaptive manner, as illustrated in Fig. 1(b) AdaBD.", + "bbox": [ + 496, + 345, + 890, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, our local offset/mask generator predicts a blending mask $\\mathbf{m}_b\\in \\mathbb{R}^{N_{\\mathbf{P}}\\times H\\times W}$ in addition to the local offsets $\\Delta \\mathbf{p}_l\\in \\mathbb{R}^{2N_{\\mathbf{P}}\\times H\\times W}$ . A learnable mask $\\mathbf{m}_b$ is composed of modulation scalars ranging from 0 to 1, each of which indicates the blending ratio of the current features $\\mathbf{a}_t$ to the aligned reference features. While DCNv2 [39] originally introduces the modulation scalars to control feature amplitudes of a single input, we leverage the scalars to adaptively blend the features from two adjacent frames considering their redundant areas.", + "bbox": [ + 496, + 571, + 890, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We apply deformable convolution by adding local offsets $\\Delta \\mathbf{p}_l$ to sampling positions of the coarsely-aligned reference features $\\mathbf{f}_{ref}^{\\prime}$ , while the current frame features $\\mathbf{a}_t$ are fed into standard convolutional operations. Concurrently, blending mask $\\mathbf{m}_b$ adaptively combines the two feature maps. In detail, an output point $\\mathbf{p_o}$ of $\\hat{\\mathbf{f}}_t$ is calculated as follows:", + "bbox": [ + 496, + 722, + 890, + 813 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\mathbf {f}} _ {t} (\\mathbf {p} _ {\\mathbf {o}}) = \\sum_ {k = 1} ^ {N _ {\\mathbf {p}}} \\mathbf {w} _ {l} (\\mathbf {p} _ {k}) \\cdot \\left\\{\\mathbf {a} _ {t} (\\mathbf {p} _ {\\mathbf {o}} + \\mathbf {p} _ {k}) \\cdot \\mathbf {m} _ {b} (\\mathbf {p} _ {k}) + \\right. \\\\ \\mathbf {f} _ {r e f} ^ {\\prime} \\left(\\mathbf {p} _ {\\mathbf {o}} + \\mathbf {p} _ {k} + \\Delta \\mathbf {p} _ {l} (\\mathbf {p} _ {k})\\right) \\cdot (1 - \\mathbf {m} _ {b} (\\mathbf {p} _ {k})) \\}, \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 824, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7615", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{w}_l$ indicates weights of local deformable convolution.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, the higher values of $\\mathbf{m}_b$ indicate the regions where current frame features are more required. In other words, Eq. 5 can be rewritten as a summation of standard convolution and deformable convolution:", + "bbox": [ + 75, + 121, + 468, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {f}} _ {t} = f _ {d c} \\left(\\mathbf {w} _ {l}, \\mathbf {a} _ {t}, \\mathbf {0}, \\mathbf {m} _ {b}\\right) + f _ {d c} \\left(\\mathbf {w} _ {l}, \\mathbf {f} _ {r e f} ^ {\\prime}, \\Delta \\mathbf {p} _ {l}, 1 - \\mathbf {m} _ {b}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 193, + 468, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this equation, the convolutional weights $\\mathbf{w}_l$ are shared between $\\mathbf{a}_t$ and $\\mathbf{f}_{ref}^{\\prime}$ . $f_{dc}$ with 0 indicates DCN with zero offsets, illustrating standard convolutional operation. To save computational costs, we decrease the channel dimension of all input features and reconstruct the original channel size before injecting the output features into the remaining layers of the teacher network.", + "bbox": [ + 75, + 226, + 470, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training Objectives", + "text_level": 1, + "bbox": [ + 76, + 340, + 264, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train Shortcut-V2V, we mainly leverage alignment loss, distillation loss, and GAN losses widely used in image/video translation networks [35, 34, 38]. First, we adopt alignment loss $L_{\\text{align}}$ to train the deformation layers in Shortcut-V2V. Since Shortcut-V2V aims to align the reference frame features $\\mathbf{f}_{ref}$ with the current frame, $L_{\\text{align}}$ computes L1 loss between the aligned feature $\\mathbf{f}_{ref}^*$ and the current frame features $\\mathbf{f}_t$ extracted from the teacher model. To obtain $\\mathbf{f}_{ref}^*$ , we align $\\mathbf{f}_{ref}$ in a coarse-to-fine manner without an intervention of $\\mathbf{m}_b$ or blending with the current features. The alignment loss $L_{\\text{align}}$ is formulated as follows:", + "bbox": [ + 75, + 364, + 468, + 545 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {r e f} ^ {*} = f _ {d c} \\left(\\mathbf {w} _ {l}, \\mathbf {f} _ {r e f} ^ {\\prime}, \\Delta \\mathbf {p} _ {l}, 1\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 547, + 468, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {a l i g n}} = \\left\\| \\mathbf {f} _ {t} - \\mathbf {f} _ {\\text {r e f}} ^ {*} \\right\\| _ {1}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 585, + 468, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, we employ knowledge distillation losses at the feature and output levels. A feature-level distillation loss $L_{feat}$ is applied between the estimated feature $\\hat{\\mathbf{f}}_t$ and the ground truth feature $\\mathbf{f}_t$ , while an output-level distillation loss $L_{out}$ compares the approximated output $\\hat{\\mathbf{O}}_t$ to the output $\\mathbf{O}_t$ generated by the teacher network. The perceptual loss [37] $L_{perc}$ is also incorporated to distill the high-frequency information of the outputs.", + "bbox": [ + 75, + 612, + 468, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lastly, we utilize a typical GAN loss $L_{GAN}$ and a temporal GAN loss $L_{T - GAN}$ , following the existing video-based frameworks [34, 3]. Temporal GAN loss $L_{T - GAN}$ encourages both temporal consistency and realisticity of the output frames. For the GAN losses, we consider the outputs of the teacher network as real images.", + "bbox": [ + 75, + 733, + 468, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall objective function $L_{total}$ is as follows:", + "bbox": [ + 96, + 824, + 429, + 839 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {t o t a l} = \\lambda_ {\\text {a l i g n}} L _ {\\text {a l i g n}} + \\lambda_ {\\text {f e a t}} L _ {\\text {f e a t}} + \\lambda_ {\\text {o u t}} L _ {\\text {o u t}} \\\\ + \\lambda_ {p e r c} L _ {p e r c} + \\lambda_ {G A N} L _ {G A N} + \\lambda_ {T - G A N} L _ {T - G A N}, \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 851, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8232a1b57b3336dff02584fe1e3648acb3033021ad182806c823e586d6f63650.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelDatasetMethodFVDMACs (G)Param. (M)
UnsupV2COriginal1.25384.617.84
CAT1.31022.831.48
OMGD1.19925.851.21
Ours1.18018.190.32
L2VOriginal1.66342.367.84
CAT1.95712.021.71
OMGD2.46712.931.21
Ours1.7549.090.32
vid2vidE2FOriginal0.1862066.69365
CAT0.288380.5524.67
OMGD0.264485.8435.81
Fast-Vid2Vid0.223767.6363.17
Ours0.209359.998.29
L2COriginal0.1461254.17411.34
CAT0.304399.3671.11
OMGD0.346391.5947.83
Fast-Vid2Vid0.310455.2971.8
Ours0.165389.1652.58
", + "bbox": [ + 501, + 87, + 890, + 348 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f3540ad71f0463c600d298ca225537ede0017e507b7c24350e1029a38d83756b.jpg", + "table_caption": [ + "Table 1: Quantitative comparison with baselines." + ], + "table_footnote": [], + "table_body": "
DatasetMethodmIoUACMP
V2LOriginal12.216.062.8
CAT5.998.8244.2
OMGD9.7613.058.7
Ours11.515.261.4
L2VOriginal10.015.647.3
CAT4.107.8127.6
OMGD3.426.8729.5
Ours9.2414.943.5
", + "bbox": [ + 519, + 387, + 874, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2: Comparison of segmentation score for V2L and FCN-score for L2V on Unsup.", + "bbox": [ + 498, + 537, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\text{align}}$ , $\\lambda_{\\text{feat}}$ , $\\lambda_{\\text{out}}$ , $\\lambda_{\\text{perc}}$ , $\\lambda_{\\text{GAN}}$ , and $\\lambda_{T - \\text{GAN}}$ are hyperparameters to control relative significance among the losses. More details on training objectives are described in the supplementary materials.", + "bbox": [ + 498, + 595, + 890, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 670, + 633, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 695, + 705, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models and Datasets. To demonstrate the generality of our model, we conduct experiments on various tasks with widely-used video-to-video translation models, Unsupervised RecycleGAN [32] and vid2vid [34].", + "bbox": [ + 496, + 718, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unsupervised RecycleGAN (Unsup) is the state-of-art video-to-video translation model among CycleGAN-based [38] ones. We conduct experiments for Unsup on Viper $\\rightarrow$ Cityscapes (V2C), which involves translating inter-modality from game driving scenes (Viper [23]) to real-world driving scenes (Cityscapes [5]). The Viper dataset consists of driving scenes collected from the GTA-V game engine, with 56 training videos and 21 test videos.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7616", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c3dd9ff3419167f8d9f04100039f10cd97a943373d8832faff4151022a229873.jpg", + "image_caption": [ + "Figure 2: Qualitative comparison. The upper rows represent the results of Unsup V2C, L2V, and V2L. The bottom rows show the results of vid2vid E2F and L2C." + ], + "image_footnote": [], + "bbox": [ + 86, + 92, + 880, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cityscapes is a dataset composed of 2,975 training videos and 500 validation videos. The input images are resized into $256 \\times 512$ for the experiments. We also tackle the translation of the videos from the Viper dataset to their corresponding segmentation label maps, $Viper \\rightarrow Label (V2L)$ , and vice versa, $Label \\rightarrow Viper (L2V)$ . We resize the images and the labels into $256 \\times 256$ following the previous work [1, 32].", + "bbox": [ + 75, + 445, + 468, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "vid2vid is a widely-used pix2pixHD-based video-to-video translation network that serves as a base architecture for various recent video-to-video translation models [33, 18]. Following vid2vid, we evaluate ShortcutV2V on Edge $\\rightarrow$ Face (E2F) and Label $\\rightarrow$ Cityscapes (L2C). E2F translates edge maps into facial videos from the FaceForensics [24], containing 704 videos for training and 150 videos for validation with various lengths. The images are cropped and resized to $512 \\times 512$ for the experiments. The edge maps are extracted using the estimated facial landmarks and Canny edge detector. Also, L2C synthesizes videos of driving scenes from segmentation label maps using the Cityscapes. We generate segmentation maps using pretrained networks following the previous studies [34, 40]. The images and the labels are resized into $256 \\times 512$ .", + "bbox": [ + 75, + 551, + 468, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Primarily, we adopt the Fréchet video distance (FVD) score [31] to evaluate the performance of Shortcut-V2V quantitatively. The FVD score measures the Fréchet Inception distance between the distribution of video-level features extracted from generated and real videos. The lower the FVD score is, the better visual quality and temporal consistency the generated video frames have. For V2L and L2V, we follow the measure", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ment of the evaluation metrics in the teacher model, Unsup [32]. We measure the segmentation scores, mean intersection over union (mIoU), mean pixel accuracy (MP), and average class accuracy (AC), to validate the performance of V2L. L2V is evaluated using FCN-score [17]. For the evaluation, we first estimate label maps from the generated videos using the FCN model pretrained with the Viper dataset. Then, we measure how accurately the estimated label maps are mapped to the ground truth segmentation labels. Higher FCN-scores refer to better output quality.", + "bbox": [ + 496, + 445, + 892, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We attach our Shortcut block to the fixed teacher networks implemented based on the official codes and pretrained by the authors, except for Unsup V2C. The teacher network of Unsup V2C is trained from scratch in the same way the original paper described [32].", + "bbox": [ + 496, + 597, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Also, standard convolutional kernels in the Shortcut block are replaced with HetConv [26] to further enhance computational efficiency without a performance drop. For the convenience of implementation and training stability, we intend a and $\\mathbf{f}$ to have the same spatial size. Furthermore, we set the max interval $\\alpha$ for each dataset considering the factors that reflect motion differences between frames, such as the frame per second (FPS) of training videos. Lower FPS usually results in larger motion differences between the adjacent frames, requiring a shorter max interval $\\alpha$ and vice versa. Specifically, we set $\\alpha$ as 3 on V2C, V2L, L2V, and L2C, since the FPS of Viper [23] and Cityscapes [5] is 15 and 17, respectively. Also, $\\alpha$ of E2F is set as 6, where the FPS of FaceForensics [24] is 30. Additional details are included in the supplementary materials.", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7617", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e546d61d4dacb5725ce8faca634d6ce788d30492b5f363b1eebf20b79fca9278.jpg", + "image_caption": [ + "(a) Unsup V2C" + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 272, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fa37ab61bc3da2787fec80a120ee3f0c45ca314a1a9515b0472ed0ad23ea1a8a.jpg", + "image_caption": [ + "(b) vid2vid E2F", + "Figure 3: Performance-efficiency trade-off of the original model, Shortcut-V2V, and the existing compression methods including OMGD [22], CAT [12], and FastVid2Vid [40]. We measure the FVD score and MACs, where the lower FVD score indicates better quality. Red points and stars denote ours with various model sizes." + ], + "image_footnote": [], + "bbox": [ + 274, + 89, + 467, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison to Baselines", + "text_level": 1, + "bbox": [ + 76, + 354, + 302, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our framework, we conduct qualitative and quantitative evaluations compared to the original model and other baselines. Since this is the first work that tackles a generally applicable compression framework for video-to-video translation, we compare Shortcut-V2V to the existing compression methods for image-to-image translation, CAT [12] and OMGD [22], regarding video frames as individual images. In the case of vid2vid, we additionally conduct a comparison to FastVid2Vid [40], which is the compression method designed specifically for vid2vid. For a fair comparison, we compress the student networks of the baselines to have similar or higher MACs compared to our model.", + "bbox": [ + 75, + 378, + 468, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Evaluation. According to Fig. 2, our method presents outputs of comparable visual quality to the original model with much fewer computations. In contrast, CAT on Unsup V2C generates undesirable buildings in the sky, and OMGD on Unsup V2C struggles with noticeable artifacts. For Unsup L2V, CAT and OMGD generate unrealistic textures for the terrain in the middle or the vegetation on the right. Moreover, in V2L, CAT estimates inappropriate labels on the sky and OMGD on the vegetation.", + "bbox": [ + 75, + 578, + 468, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For vid2vid E2F, CAT shows unwanted artifacts on the mouth, and OMGD presents inconsistent outputs. For vid2vid L2C, the outputs of CAT are blurry, especially for the trees in the background, and OMGD generates artifacts at the bottom of the images. Although Fast-Vid2Vid shows reasonable image quality, the output frames are inaccurately aligned with the input (e.g., a head pose of a person) and suffer from ghost effects (e.g., the black car on the left) due to motion compensation using interpolation.", + "bbox": [ + 75, + 715, + 467, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Evaluation. As shown in Table 1 and Table 2, our framework successfully improves the computational efficiency of the original network without a significant perfor", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/abab91fb97af0fddb89999d21828f7017aefc092d608bfa2d3bb628be4ba83f9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ConfigurationsUnsup V2Cvid2vid E2F
(a) w/o reference features1.2560.398
(b) w/o current features1.2490.218
(c) Single-stage alignment1.1950.213
(d) w/o adaptive blending1.2080.244
(e) Ours1.1800.209
", + "bbox": [ + 500, + 88, + 890, + 186 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3: An ablation study on Unsup V2C and vid2vid E2F. We measure the FVD scores.", + "bbox": [ + 500, + 196, + 888, + 226 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mance drop. In the case of Unsup, Shortcut-V2V reduces the MACs by $4.7 \\times$ and the number of parameters by $24.5 \\times$ . In addition, our approach saves vid2vid's MACs by 3.2 and $5.7 \\times$ and the number of parameters by 7.8 and $44 \\times$ on each task. Fig. 3 visualizes the performance-efficiency trade-off of our framework and other baselines.", + "bbox": [ + 496, + 256, + 890, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "According to Table 1, we also outperform other compression methods for image-to-image translation, CAT [12] and OMGD [22], even with fewer computations. Image-based approaches cannot consider temporal coherence among the frames during the compression, leading to a loss of quality. Meanwhile, we effectively preserve the original performance by exploiting rich information in the previous features during inference. Shortcut-V2V even shows superiority over Fast-Vid2Vid [40], which is a compression method specifically designed for vid2vid. Table 2 also demonstrates that ours on Unsup surpasses the existing compression models by a large margin.", + "bbox": [ + 496, + 348, + 892, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 541, + 651, + 558 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct an ablation study to evaluate the effect of each component of Shortcut-V2V. As described in Table 3, we compare the FVD scores of 5 different configurations, (a) w/o reference features, (b) w/o current features, (c) single-stage alignment, (d) w/o adaptive blending, and (e) ours, on Unsup V2C and vid2vid E2F.", + "bbox": [ + 496, + 566, + 890, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, while our model originally exploits both the reference features $\\mathbf{f}_{ref}$ and the current frame features $\\mathbf{a}_t$ to synthesize $\\hat{\\mathbf{f}}_t$ , (a) and (b) are designed to leverage only either of them. To be specific, (a) synthesizes $\\hat{\\mathbf{f}}_t$ by processing standard convolutions on $\\mathbf{a}_t$ without integrating the deformed $\\mathbf{f}_{ref}$ . Meanwhile, (b) estimates $\\hat{\\mathbf{f}}_t$ using only $\\mathbf{f}_{ref}$ aligned in a coarse-to-fine manner without blending $\\mathbf{a}_t$ . The results demonstrate that the absence of either $\\mathbf{f}_{ref}$ or $\\mathbf{a}_t$ leads to performance degradation. Next, (c) performs a single-stage alignment on $\\mathbf{f}_{ref}$ to show the necessity of a coarse-to-fine alignment. Specifically, (c) leverages a single offset/mask generator to predict $\\Delta \\mathbf{p}$ and $\\mathbf{m}_b$ which are then used in AdaBD for blending and deformation. The result indicates that coarse-to-fine alignment effectively assists the estimation of $\\hat{\\mathbf{f}}_t$ by minimizing misalignment between the features from adjacent frames. This makes Shortcut-V2V suitable for", + "bbox": [ + 496, + 657, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7618", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a7aa8d7781e173f499019437ea5ee417891dc2af3c13019a568244b0918fb35f.jpg", + "image_caption": [ + "Reference frame" + ], + "image_footnote": [], + "bbox": [ + 107, + 108, + 240, + 160 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/75e3ad8902851e6a2f44c7d02a2c9b9080b8e379286eae79b7d2aa9117ede368.jpg", + "image_caption": [ + "Current frame" + ], + "image_footnote": [], + "bbox": [ + 307, + 109, + 437, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/65d60aec67bff2e75167dcc2383c061e01dce23370e90af142008e946e409249.jpg", + "image_caption": [ + "Figure 4: Visualization of global/local offsets and blending masks. The green points denote output points of the current frame, and the red points around them are each output point's sampling positions modified by global offsets $\\Delta \\mathbf{p}_g$ and local offsets $\\Delta \\mathbf{p}_l$ . The values in a blending mask $\\mathbf{m}_b$ are averaged by the kernel size for visualization. The brighter area indicates higher mask values." + ], + "image_footnote": [], + "bbox": [ + 78, + 171, + 269, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/89ad3ad3142d4dd2756198857948ea712a8771cc7dc4c84a1c10220014674896.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 172, + 372, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/98ad9cf051ce8d7a056ac8da6db177267d7df905ed99ff1f41e7a03fbf79bf59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 172, + 468, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "generating videos with diverse motion differences. Lastly, (d) blends $\\mathbf{f}_{ref}^{*}$ and $\\mathbf{a}_t$ simply by element-wise addition instead of adaptive blending with $\\mathbf{m}_b$ in AdaBD. The result demonstrates that $\\mathbf{m}_b$ also encourages better estimation of $\\hat{\\mathbf{f}}_t$ by selectively exploiting features from temporally adjacent features.", + "bbox": [ + 75, + 398, + 468, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Offset/Mask Visualization", + "text_level": 1, + "bbox": [ + 76, + 498, + 313, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We provide a qualitative analysis of the generated global/local offsets and blending masks. Fig. 4 visualizes the sampling positions (red points) for each output position $\\mathbf{p}_o$ of the current frame (green points), where the predicted offsets $\\Delta \\mathbf{p}_g$ and $\\Delta \\mathbf{p}_l$ are added to the original sampling positions. According to the result, global offsets $\\Delta \\mathbf{p}_g$ effectively reflect the global movement of the objects such as the bridge. Also, the summation of the global and local offsets indicates the sampling points refined by local offsets enabling fine alignment. This shows that the estimated offsets effectively support the utilization of common features in the reference frame. For the blending masks $\\mathbf{m}_b$ , Fig. 4 presents that the regions with significant motion differences (e.g., trees) have large mask values compared to regions with little change (e.g., road, sky). That is, our model relies more on the current features rather than the reference features when the deformation is challenging, which aligns with our intention and leads to robust performance.", + "bbox": [ + 75, + 521, + 468, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Performance-Efficiency Trade-off", + "text_level": 1, + "bbox": [ + 76, + 801, + 372, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a performance-efficiency trade-off of our framework with respect to varying model sizes of Shortcut block and teacher model dependence.", + "bbox": [ + 75, + 825, + 468, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Shortcut Block Size. We construct our models of different sizes by reducing the number of the output chan", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a302dc8e92e275a0e806e8f74848f9802cf63fadb3e5affbd6079277761f2a86.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DependenceUnsup V2Cvid2vid E2F
FVDMACs (G)Param. (M)FVDMACs (G)Param. (M)
Low1.22113.790.140.277243.972.11
Medium1.18018.190.320.209359.998.29
High1.16627.971.380.193475.9932.98
", + "bbox": [ + 500, + 88, + 890, + 147 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4: Shortcut-V2V performance with different teacher model dependence. High dependency denotes using more teacher network layers during the inference.", + "bbox": [ + 498, + 157, + 890, + 203 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "nel of $1 \\times 1$ channel reduction layer by half (Ours-1/2) and a quarter (Ours-1/4). As shown in Fig. 3, despite the performance-efficiency trade-off depending on channel dimension, Ours-1/2 and Ours-1/4 still achieve comparable FVD to the baselines with less computation costs. On a single RTX 3090 GPU, our models with three different sizes on Unsup achieve 99.67, 109.69, and 111.80 FPS with an $11.6 - 21.2\\%$ reduction in inference time compared to the original model's speed of 88.10 FPS. Additionally, Ours, Ours-1/2, and Ours-1/4 on vid2vid contribute to a significant improvement in speed from 5.63 FPS to 12.86, 12.93, and 13.16 FPS, resulting in a $56.2 - 57.2\\%$ reduction in inference time. Our method is capable of a real-time inference, unlike the previous work [40] which requires future frames for the current frame inference due to motion compensation. More details on channel manipulation are described in the supplementary materials.", + "bbox": [ + 496, + 233, + 890, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Teacher Model Dependence. Since Shortcut-V2V leverages a subset of the teacher network during inference, the computational costs and memory usage may vary depending on the amount of teacher network we use. In this regard, we present an analysis of the computational efficiency and performance of our model with respect to various levels of teacher model dependence. We categorize the dependence on the teacher model into three levels, low, medium, and high, where high dependence indicates using more teacher network layers. According to Table 4, our model of higher dependence achieves better performance with larger MACs and the number of parameters. The results also demonstrate that Shortcut-V2V can leverage temporal redundancy in features extracted from various layers of the teacher network. Further details for teacher model dependence and experiments for temporal redundancy are included in the supplementary materials.", + "bbox": [ + 496, + 491, + 890, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 500, + 767, + 746, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although Shortcut-V2V significantly improves the test-time efficiency of video-to-video translation, it still poses several limitations. First, a constant max interval may induce unsatisfactory outputs when a degree of temporal redundancy largely varies between frames. Following recent studies [25, 20, 19], applying an adaptive interval based on frame selection algorithms or a learnable policy net-", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7619", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "work could be promising future research. In addition, we need to manually configure various hyperparameters such as channel dimension, max interval, and teacher model dependence, which can be automated by NAS [13]. Lastly, our method still has the potential to further improvement of computational efficiency by compressing the teacher model using other methods before applying our framework.", + "bbox": [ + 80, + 90, + 467, + 196 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Despite the limitations, Shortcut-V2V presents a significant improvement in test-time efficiency of video-to-video translation networks based on temporal redundancy reduction, while preserving the original model performance. Extensive experiments with widely-used video-to-video translation models successfully demonstrate the general applicability of our framework. To the best of our knowledge, this is the first work for a general model compression in the domain of video-to-video translation. We hope our work facilitates research on video-to-video translation and extends the range of its application.", + "bbox": [ + 80, + 198, + 467, + 362 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2021-0-02068, Artificial Intelligence Innovation Hub) and the Ministry of Culture, Sports and Tourism and Korea Creative Content Agency (Project Number: R2021040097, Contribution Rate: 50).", + "bbox": [ + 80, + 369, + 467, + 474 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7620", + "bbox": [ + 483, + 946, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aayush Bansal, Shugao Ma, Deva Ramanan, and Yaser Sheikh. Recycle-gan: Unsupervised video retargeting. Proc. of the European Conference on Computer Vision (ECCV), 2018. 1, 2, 6", + "[2] Kelvin C. K. Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Understanding deformable alignment in video super-resolution. Proc. the AAAI Conference on Artificial Intelligence (AAAI), 2021. 3", + "[3] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston H. Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. Proc. of the IEEE International Conference on Computer Vision (ICCV), 2019. 5", + "[4] Yang Chen, Yingwei Pan, Ting Yao, Xinmei Tian, and Tao Mei. Mocycle-gan: Unpaired video-to-video translation. In ACM Multimedia (ACM MM), 2019. 2", + "[5] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5, 6", + "[6] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. Proc. of the IEEE International Conference on Computer Vision (ICCV), 2017. 3", + "[7] Jianing Deng, Li Wang, Shiliang Pu, and Cheng Zhuo. Spatio-temporal deformable convolution for compressed video quality enhancement. Proc. the AAAI Conference on Artificial Intelligence (AAAI), 2020. 3", + "[8] Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation, 2019. 3", + "[9] Amirhossein Habibian, Haitam Ben Yahia, Davide Abati, Efstratos Gavves, and Fatih Porikli. Delta distillation for efficient video processing, 2022. 3", + "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1", + "[11] Bangrui Jiang, Zhihuai Xie, Zhen Xia, Songnan Li, and Shan Liu. Erdn: Equivalent receptive field deformable network for video deblurring. In Proc. of the European Conference on Computer Vision (ECCV), 2022. 3", + "[12] Qing Jin, Jian Ren, Oliver J. Woodford, Jiazhuo Wang, Geng Yuan, Yanzhi Wang, and Sergey Tulyakov. Teachers do more than teach: Compressing image-to-image models. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3, 7", + "[13] Muyang Li, Ji Lin, Yaoyao Ding, Zhijian Liu, Jun-Yan Zhu, and Song Han. GAN compression: Efficient architectures for interactive conditional gans. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 3, 9", + "[14] Ji Lin, Chuang Gan, and Song Han. Temporal shift module for efficient video understanding. Proc. of the IEEE International Conference on Computer Vision (ICCV), 2019. 3" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Jiayi Lin, Yan Huang, and Liang Wang. Fdan: Flow-guided deformable alignment network for video super-resolution. arXiv preprint arXiv:2105.05640, 2021. 3", + "[16] Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient semantic video segmentation with per-frame inference. Proc. of the European Conference on Computer Vision (ECCV), 2020. 3", + "[17] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 6", + "[18] Arun Mallya, Ting-Chun Wang, Karan Sapra, and Ming-Yu Liu. World-consistent video-to-video synthesis. Proc. of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 6", + "[19] Yue Meng, Rameswar Panda, Chung-Ching Lin, Prasanna Sattigeri, Leonid Karlinsky, Kate Saenko, Aude Oliva, and Rogério Feris. Adafuse: Adaptive temporal fusion network for efficient action recognition. Proc. the International Conference on Learning Representations (ICLR), 2021. 1, 3, 8", + "[20] Bowen Pan, Rameswar Panda, Camilo Fosco, Chung-Ching Lin, Alex Andonian, Yue Meng, Kate Saenko, Aude Oliva, and Rogerio Feris. Va-red²: Video adaptive redundancy reduction. Proc. the International Conference on Learning Representations (ICLR), 2021. 1, 3, 8", + "[21] KwanYong Park, Sanghyun Woo, Dahun Kim, Donghyeon Cho, and In So Kweon. Preserving semantic and temporal consistency for unpaired video-to-video translation. ACM Multimedia (ACMM), 2019. 2", + "[22] Yuxi Ren, Jie Wu, Xuefeng Xiao, and Jianchao Yang. Online multi-granularity distillation for gan compression. Proc. of the IEEE International Conference on Computer Vision (ICCV), 2021. 1, 3, 7", + "[23] Stephan R. Richter, Zeeshan Hayden, and Vladlen Koltun. Playing for benchmarks. In Proc. of the IEEE International Conference on Computer Vision (ICCV), 2017. 5, 6", + "[24] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. FaceForensics++: Learning to detect manipulated facial images. In Proc. of the IEEE International Conference on Computer Vision (ICCV), 2019. 6", + "[25] Jonghyeon Seon, Jaedong Hwang, Jonghwan Mun, and Bohyung Han. Stop or forward: Dynamic layer skipping for efficient action recognition. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2023. 8", + "[26] Pravendra Singh, Vinay Kumar Verma, Piyush Rai, and Vinay P. Namboodiri. Hetconv: Heterogeneous kernel-based convolutions for deep cnns. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6", + "[27] Ximeng Sun, Rameswar Panda, Chun-Fu Chen, Aude Oliva, Rogério Feris, and Kate Saenko. Dynamic network quantization for efficient video inference. Proc. of the IEEE International Conference on Computer Vision (ICCV), 2021. 3", + "[28] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7621", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Ryan Szeto, Mostafa El-Khamy, Jungwon Lee, and Jason J. Corso. Hypercon: Image-to-video model transfer for video-to-video translation tasks. IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2021. 3", + "[30] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. TDAN: temporally deformable alignment network for video super-resolution. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3, 4", + "[31] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. In Proc. the International Conference on Learning Representations Workshop (ICLRW), 2019. 6", + "[32] Kaihong Wang, Kumar Akash, and Teruhisa Mitsu. Learning temporally and semantically consistent unpaired video-to-video translation through pseudo-supervision from synthetic optical flow. Proc. the AAAI Conference on Artificial Intelligence (AAAI), 2022. 1, 2, 5, 6", + "[33] Ting-Chun Wang, Ming-Yu Liu, Andrew Tao, Guilin Liu, Jan Kautz, and Bryan Catanzaro. Few-shot video-to-video synthesis. Proc. the Advances in Neural Information Processing Systems (NeurIPS), 2019. 1, 2, 6", + "[34] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Proc. the Advances in Neural Information Processing Systems (NeurIPS), 2018. 1, 2, 3, 5, 6", + "[35] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 5", + "[36] Xintao Wang, Kelvin C. K. Chan, Ke Yu, Chao Dong, and Chen Change Loy. EDVR: video restoration with enhanced deformable convolutional networks. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2019. 3", + "[37] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5", + "[38] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proc. of the IEEE International Conference on Computer Vision (ICCV), 2017. 2, 5", + "[39] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3, 4", + "[40] Long Zhuo, Guangcong Wang, Shikai Li, Wayne Wu, and Ziwei Liu. Fast-vid2vid: Spatial-temporal compression for video-to-video synthesis. Proc. of the European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 6, 7, 8" + ], + "bbox": [ + 78, + 90, + 470, + 853 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7622", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Shortcut-V2V_ Compression Framework for Video-to-Video Translation Based on Temporal Redundancy Reduction/99da4ac0-6f5c-433f-bc7b-caa55aa672d5_model.json b/2023/Shortcut-V2V_ Compression Framework for Video-to-Video Translation Based on Temporal Redundancy Reduction/99da4ac0-6f5c-433f-bc7b-caa55aa672d5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c904f541cef43b1e38905b65964da1567c51c09b --- /dev/null +++ b/2023/Shortcut-V2V_ Compression Framework for Video-to-Video Translation Based on Temporal Redundancy Reduction/99da4ac0-6f5c-433f-bc7b-caa55aa672d5_model.json @@ -0,0 +1,2037 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.123, + 0.13, + 0.849, + 0.177 + ], + "angle": 0, + "content": "Shortcut-V2V: Compression Framework for Video-to-Video Translation based on Temporal Redundancy Reduction" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.215, + 0.878, + 0.235 + ], + "angle": 0, + "content": "Chaeyeon Chung\\*, Yeojeong Park\\*, Seunghwan Choi, Munkhsoyol Ganbat, and Jaegul Choo" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.246, + 0.748, + 0.298 + ], + "angle": 0, + "content": "\\(^{1}\\)KAIST AI \\(^{2}\\)KT Research & Development Center, KT Corporation {cy-chung, indigopyj, shadow2496, soyol, jchoo}@kaist.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.341, + 0.314, + 0.357 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.374, + 0.473, + 0.737 + ], + "angle": 0, + "content": "Video-to-video translation aims to generate video frames of a target domain from an input video. Despite its usefulness, the existing networks require enormous computations, necessitating their model compression for wide use. While there exist compression methods that improve computational efficiency in various image/video tasks, a generally-applicable compression method for video-to-video translation has not been studied much. In response, we present Shortcut-V2V, a general-purpose compression framework for video-to-video translation. Shortcut-V2V avoids full inference for every neighboring video frame by approximating the intermediate features of a current frame from those of the previous frame. Moreover, in our framework, a newly-proposed block called AdaBD adaptively blends and deforms features of neighboring frames, which makes more accurate predictions of the intermediate features possible. We conduct quantitative and qualitative evaluations using well-known video-to-video translation models on various tasks to demonstrate the general applicability of our framework. The results show that Shortcut-V2V achieves comparable performance compared to the original video-to-video translation model while saving \\(3.2 - 5.7 \\times\\) computational cost and \\(7.8 - 44 \\times\\) memory at test time. Our code and videos are available at https://shortcut-v2v.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.77, + 0.21, + 0.787 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.797, + 0.47, + 0.873 + ], + "angle": 0, + "content": "Video-to-video translation is a task of generating temporally consistent and realistic video frames of a target domain from a given input video. Recent studies on video-to-video translation present promising performance in various domains such as inter-modality translation between labels" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.342, + 0.892, + 0.373 + ], + "angle": 0, + "content": "and videos [34, 33, 18], and intra-modality translation between driving scene videos [32] or face videos [1]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.374, + 0.893, + 0.553 + ], + "angle": 0, + "content": "Despite enhanced usefulness, video-to-video translation networks usually require substantial computational cost and memory usage, which limits their applicability. For instance, multiply-accumulates (MACs) of a widely-used video translation model, vid2vid [34], is 2066.69G, while the basic convolutional neural networks, ResNet v2 50 [10] and Inception v3 [28], are 4.12G and 6G, respectively. Furthermore, temporally redundant computations for adjacent video frames also harm the cost efficiency of a video-to-video translation network. Performing full inference for every neighboring video frame that contains common visual features inevitably entails redundant operations [20, 19]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.895, + 0.825 + ], + "angle": 0, + "content": "In this regard, Fast-Vid2Vid [40] proposes a compression framework for vid2vid [34] based on spatial input compression and temporal redundancy reduction. However, it cannot be applied to other video-to-video translation models since it is designed specifically for vid2vid. Moreover, Fast-Vid2Vid does not support real-time inference since it requires future frames to infer a current one. Alternatively, one can apply model compression approaches for image-to-image translation [13, 22, 12] directly to video-to-video translation, considering video frames as separate images. However, these approaches are not designed to consider the correlation among adjacent video frames during the compression. This may result in unrealistic output quality in video-to-video translation, where the inherent temporal coherence of an input video needs to be preserved in the outputs. Also, frame-by-frame inference without temporal redundancy reduction involves unnecessary computations, resulting in computational inefficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose Shortcut-V2V, a general-purpose framework for improving the computational efficiency of video-to-video translation based on temporal redundancy reduction. Shortcut-V2V allows the original video-to-video translation model to avoid temporally re" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.259, + 0.901 + ], + "angle": 0, + "content": "*indicates equal contributions." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.958 + ], + "angle": 0, + "content": "7612" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.437, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.091, + 0.891, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.268, + 0.894, + 0.314 + ], + "angle": 0, + "content": "Figure 1: Overview of Shortcut-V2V. (a) is an overall framework of Shorcut-V2V, and (b) shows a detailed architecture of Shortcut block. \\(\\uparrow 2\\) and \\(\\downarrow 2\\) refer to upsampling and downsampling by a factor of 2, respectively. G in Offset G and Offset/Mask G indicates a generator." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.47, + 0.552 + ], + "angle": 0, + "content": "dundant computations by approximating the decoding layer features of the current frame with largely reduced computations. To enable lightweight estimation, our framework leverages features from the previous frame (i.e., reference features), which have high visual similarity with the current frame. We also exploit current frame features from the encoding layer to handle newly-appeared regions in the current frame. Specifically, we first globally align the previous frame features with the current frame features, and our novel Adaptive Blending and Deformation block (AdaBD) in Shortcut-V2V blends features of neighboring frames while performing detailed deformation. AdaBD adaptively integrates the features regarding their redundancy in a lightweight manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.471, + 0.87 + ], + "angle": 0, + "content": "In this way, our model significantly improves the test-time efficiency of the original network while preserving its original performance. Shortcut-V2V is easily applicable to a pretrained video-to-video translation model to save computational cost and memory usage. Our framework is also suitable for real-time inference since we do not require future frames for the current frame inference. To the best of our knowledge, this is the first attempt at a general-purpose model compression approach for video-to-video translation. We demonstrate the effectiveness of our approach using well-known video-to-video translation models, Unsupervised RecycleGAN [32] (Unsup) and vid2vid [34]. Shortcut-V2V reduces \\(3.2 - 5.7 \\times\\) computational cost and \\(7.8 - 44 \\times\\) memory usage while achieving comparable performance to the original model. Since there is no existing general-purpose compression method, we compare our method with Fast-Vid2Vid [40] and the compression methods for image-to-image translation. Our model presents superiority over the existing approaches in both quantitative and qualitative evaluations. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "- We introduce a novel, general-purpose model com" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.341, + 0.892, + 0.386 + ], + "angle": 0, + "content": "pression framework for video-to-video translation, Shortcut-V2V, that enables the original network to avoid temporally redundant computations." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.399, + 0.891, + 0.444 + ], + "angle": 0, + "content": "- We present AdaBD that exploits features from neighboring frames via adaptive blending and deformation in a lightweight manner." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.457, + 0.891, + 0.516 + ], + "angle": 0, + "content": "- Our framework saves up to \\(5.7 \\times\\) MACs and \\(44 \\times\\) parameters across various video-to-video translation tasks, achieving comparable performance to the original networks." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.399, + 0.891, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.534, + 0.641, + 0.549 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.744, + 0.574 + ], + "angle": 0, + "content": "2.1. Video-to-Video Translation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Recent video-to-video translation networks are generally classified into pix2pixHD-based [35] and CycleGAN-based [38] generator. vid2vid [34] proposes a pix2pixHD-based sequential generation framework that synthesizes a current output given the previous outputs as additional guidance. As the following work, few-shot vid2vid [33] achieves few-shot generalization of vid2vid based on attention modules, and world-consistent vid2vid [18] is proposed to improve long-term temporal consistency of vid2vid." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "While pix2pixHD-based models require paired annotated videos, RecycleGAN [1] and MocycleGAN [4] propose CycleGAN-based video translation models. They exploit spatio-temporal consistency losses to generate realistic videos using unpaired datasets. STC-V2V [21] leverages optical flow for semantic/temporal consistency to improve the output quality of the existing models. Unsupervised RecycleGAN [32] achieves state-of-the-art performance among CycleGAN-based frameworks with a pseudosupervision by the synthetic flow. Although the existing video-to-video translation networks achieve decent performance, they commonly demand a non-trivial amount of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7613" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "computational costs and memory usage. Also, frame-by-frame inference necessarily causes redundant operations due to temporal redundancy among adjacent frames." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.148, + 0.267, + 0.166 + ], + "angle": 0, + "content": "2.2. Model Compression" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.171, + 0.469, + 0.413 + ], + "angle": 0, + "content": "Model compression for video-related tasks has been actively proposed in various domains [19, 20, 16, 9, 27, 8, 14, 29], such as object detection, action recognition, semantic segmentation, and super-resolution. Several studies [19, 9, 20] exploit temporal redundancy among video frames to improve efficiency during training or inference. For instance, Habibian et al. [9] distill only the residual between adjacent frames from a teacher model to a student to speed up the inference. Also, Fast-Vid2Vid [40] firstly proposes a compression framework for video-to-video translation based on spatial and temporal compression. However, Fast-Vid2Vid focuses on vid2vid [34], limiting its application to other video-to-video translation networks. Also, temporal redundancy reduction via motion compensation in Fast-Vid2Vid requires the future frame to infer the current frame, which is not suitable for real-time inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.469, + 0.641 + ], + "angle": 0, + "content": "Alternatively, model-agnostic compression methods for image-to-image translation can also be applied to video-to-video translation models. The existing approaches for image synthesis mainly tackle channel pruning [13], knowledge distillation [12, 22], NAS [13], etc. For instance, CAT [12] compresses the teacher network with one-step pruning to satisfy the target computation budget, while OMGD [22] conducts a single-stage online distillation in which the teacher generator supports the student generator to be refined progressively. However, image-based compression methods cannot consider temporal coherence among neighboring video frames, which may induce unrealistic results in video translation tasks. Also, performing full model inference for each video frame still poses computational inefficiency due to the temporal redundancy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.303, + 0.667 + ], + "angle": 0, + "content": "2.3. Deformable Convolution" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Deformable convolution [6] is originally proposed to enhance the transformation capability of a convolutional layer by adding the estimated offsets to a regular convolutional kernel in vision tasks such as object detection or semantic segmentation. Besides its original application, recent studies [30, 7, 36, 2, 11, 15] demonstrate that deformable convolution is also capable of aligning adjacent video frames. TDAN [30] utilizes deformable convolution to capture implicit motion cues between consecutive frames by dynamically predicted offsets in video super-resolution. In addition, EDVR [36] stacks several deformable convolution blocks to estimate large and complex motions in video restoration. In this paper, we also leverage a deformable convolution to adaptively align adjacent video frames in a lightweight manner." + }, + { + "type": "code_caption", + "bbox": [ + 0.502, + 0.106, + 0.753, + 0.121 + ], + "angle": 0, + "content": "Algorithm 1 Shortcut-V2V Inference" + }, + { + "type": "algorithm", + "bbox": [ + 0.507, + 0.125, + 0.892, + 0.325 + ], + "angle": 0, + "content": "1: Input: Input video \\(\\{\\mathbf{I}_t\\}_{t=0}^{N_T-1}\\) of length \\(N_T\\), teacher model \\(T\\), layer index of encoder \\(l_e\\) and decoder \\(l_d\\), Shortcut block \\(S\\), max interval \\(\\alpha\\) \n2: Output: Output video \\(\\{\\mathbf{O}_t\\}_{t=0}^{N_T-1}\\) \n3: for \\(t = 0\\) to \\(N_T - 1\\) do \n4: \\(\\mathbf{a}_t = T_{[le]}(\\mathbf{I}_t)\\) \n5: if \\(t\\% \\alpha = 0\\) then \n6: \\(\\mathbf{f}_t = T_{[le + 1;l_d - 1]}(\\mathbf{a}_t)\\) \n7: \\(\\mathbf{a}_{\\mathrm{ref}}, \\mathbf{f}_{\\mathrm{ref}} \\gets \\mathbf{a}_t, \\mathbf{f}_t \\triangleright\\) Update the reference features \n8: else \n9: \\(\\mathbf{f}_t = S(\\mathbf{f}_{\\mathrm{ref}}, \\mathbf{a}_{\\mathrm{ref}}, \\mathbf{a}_t)\\) \n10: end if \n11: \\(\\mathbf{O}_t = T_{[l_d,]}(\\mathbf{f}_t)\\) \n12: end for" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.367, + 0.642, + 0.383 + ], + "angle": 0, + "content": "3. Shortcut-V2V" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.892, + 0.65 + ], + "angle": 0, + "content": "In this paper, we propose Shortcut-V2V, a general compression framework to improve the test-time efficiency in video-to-video translation. As illustrated in Fig. 1(a), given \\(\\{\\mathbf{I}_t\\}_{t=0}^{N_T-1}\\) as input video frames, we first use full teacher model \\(T\\) to synthesize the output of the first frame. Then, for the next frames, our newly-proposed Shortcut block efficiently approximates \\(\\mathbf{f}_t\\), the features from the \\(l_d\\)-th decoding layer of the teacher model. This is achieved by leveraging the \\(l_e\\)-th encoding layer features \\(\\mathbf{a}_t\\) along with reference features, \\(\\mathbf{a}_{ref}\\) and \\(\\mathbf{f}_{ref}\\), from the previous frame. Here, \\(l_d\\) and \\(l_e\\) correspond to layer indices of the teacher model. Lastly, predicted features \\(\\hat{\\mathbf{f}}_t\\) are injected into the following layers of the teacher model to synthesize the final output \\(\\hat{\\mathbf{O}}_t\\). To avoid error accumulation, we conduct full teacher inference and update the reference features at every max interval \\(\\alpha\\). We provide the detailed inference process of Shortcut-V2V in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.892, + 0.697 + ], + "angle": 0, + "content": "The architecture of our model is mainly inspired by Deformable Convolutional Network (DCN) [6, 39], which we explain in the next section." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.71, + 0.81, + 0.725 + ], + "angle": 0, + "content": "3.1. Deformable Convolutional Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "DCN [6, 39] is initially introduced to improve the transformation capability of a convolutional layer in image-based vision tasks, e.g., object detection and semantic segmentation. In the standard convolutional layer, a \\(3 \\times 3\\) kernel with dilation 1 samples points over input features using a sampling position \\(\\mathbf{p}_k \\in \\{(-1, -1), (-1, 0), \\dots, (0, 1), (1, 1)\\}\\). Given input feature maps \\(\\mathbf{x}\\), DCN predicts additional offsets \\(\\Delta \\mathbf{p} \\in \\mathbb{R}^{2N_{\\mathbf{p}} \\times H \\times W}\\) to augment each sampling position along x-axis and y-axis. Here, \\(N_{\\mathbf{p}}\\) is the number of sampling positions in a kernel, and \\(H\\) and \\(W\\) are the height and width" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7614" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "of output feature maps, respectively. For further manipulation of input feature amplitudes over the sampled points, DCNv2 [39] introduces a modulated deformable convolution with \\(\\mathbf{m} \\in \\mathbb{R}^{N_{\\mathbf{P}} \\times H \\times W}\\) consisting of learnable modulation scalars. The deformed output feature maps \\(\\mathbf{x}'\\) by DCNv2 are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.193, + 0.47, + 0.211 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {\\prime} = f _ {d c} (\\mathbf {w}, \\mathbf {x}, \\Delta \\mathbf {p}, \\mathbf {m}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.221, + 0.468, + 0.252 + ], + "angle": 0, + "content": "where \\( f_{dc} \\) indicates a deformable convolution. Specifically, a single point \\( \\mathbf{p_o} \\) of \\( \\mathbf{x}' \\) is obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.262, + 0.469, + 0.305 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {\\prime} \\left(\\mathbf {p} _ {\\mathbf {o}}\\right) = \\sum_ {k = 1} ^ {N _ {\\mathbf {p}}} \\mathbf {w} \\left(\\mathbf {p} _ {k}\\right) \\cdot \\mathbf {x} \\left(\\mathbf {p} _ {\\mathbf {o}} + \\mathbf {p} _ {k} + \\Delta \\mathbf {p} \\left(\\mathbf {p} _ {k}\\right)\\right) \\cdot \\mathbf {m} \\left(\\mathbf {p} _ {k}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.315, + 0.469, + 0.36 + ], + "angle": 0, + "content": "where \\(\\mathbf{w}(\\mathbf{p}_k),\\mathbf{p}(\\mathbf{p}_k)\\) , and \\(\\mathbf{m}(\\mathbf{p}_k)\\) are convolutional layer weights, offsets, and modulation scalars between 0 and 1, respectively, for the \\(k\\) -th sampling position." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.361, + 0.469, + 0.435 + ], + "angle": 0, + "content": "Taking advantage of the enhanced transformation capability, we also leverage deformable convolution to align features of adjacent frames only with a few convolution-like operations, instead of using heavy flow estimation networks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.445, + 0.229, + 0.459 + ], + "angle": 0, + "content": "3.2. Shortcut Block" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.468, + 0.469, + 0.515 + ], + "angle": 0, + "content": "As described in Fig. 1(b), Shortcut block \\( S \\) estimates the current frame features \\( \\hat{\\mathbf{f}}_t \\) given \\( \\mathbf{a}_t \\) and the reference frame features \\( \\mathbf{f}_{ref} \\) and \\( \\mathbf{a}_{ref} \\) as inputs:" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.525, + 0.469, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {f}} _ {t} = S \\left(\\mathbf {f} _ {r e f}, \\mathbf {a} _ {r e f}, \\mathbf {a} _ {t}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.553, + 0.469, + 0.658 + ], + "angle": 0, + "content": "Our block effectively obtains rich information from \\(\\mathbf{f}_{ref}\\) via coarse-to-fine alignment referring to alignment between \\(\\mathbf{a}_{ref}\\) and \\(\\mathbf{a}_t\\). Also, during the fine alignment, our newly-proposed AdaBD simultaneously performs adaptive blending of the aligned \\(\\mathbf{f}_{ref}\\) and the current frame feature \\(\\mathbf{a}_t\\). Here, \\(\\mathbf{a}_t\\) supports the synthesis of newly-appeared areas in the current frame." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Coarse-to-Fine Alignment. To handle a wide range of misalignments between the frames, our model aligns \\(\\mathbf{f}_{ref}\\) with the current frame in a coarse-to-fine manner. Our global/local alignment module consists of an offset generator to estimate offsets, and deformable convolution layers to deform features based on the predicted offsets. Following TDAN [30], an offset generator estimates sampling offsets given the adjacent frame features. For global alignment, we first downsample the given inputs to enlarge the receptive fields of the corresponding convolutional layers in a lightweight manner. The downsampled \\(\\mathbf{a}_{ref}\\) and \\(\\mathbf{a}_t\\) are concatenated and fed into a global offset generator to generate global offsets \\(\\Delta \\mathbf{p}_g\\in \\mathbb{R}^{2\\times \\frac{H}{2}\\times \\frac{W}{2}}\\). Since we only need to capture coarse movement, \\(\\Delta \\mathbf{p}_g\\) includes a single offset for each kernel, unlike the original DCN. Each offset is identically applied to all the sampling positions within the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.124 + ], + "angle": 0, + "content": "kernel. Then, the deformed features are upsampled back to the original size to obtain \\(\\mathbf{f}_{ref}^{\\prime}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.135, + 0.892, + 0.154 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {r e f} ^ {\\prime} = \\left(f _ {d c} \\left(\\mathbf {w} _ {g}, \\left(\\mathbf {f} _ {r e f}\\right) ^ {\\downarrow 2}, \\Delta \\mathbf {p} _ {g}, \\mathbf {1})\\right) ^ {\\uparrow 2}, \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.165, + 0.892, + 0.346 + ], + "angle": 0, + "content": "where \\(\\mathbf{w}_g\\) denotes weights of global deformable convolution, and \\((\\cdot)^{\\uparrow 2}\\) and \\((\\cdot)^{\\downarrow 2}\\) refer to upsampling and downsampling by a factor of 2 through bilinear interpolation, respectively. 1 indicates a vector filled with 1 so that no modulation is applied here. While local alignment of the coarsely-aligned feature \\(\\mathbf{f}_{ref}^{\\prime}\\) follows the process of global alignment, the difference lies in that each sampling point of each kernel has a unique offset, and the alignment operation is conducted in the original resolution. We leverage \\(\\mathbf{a}_{ref}^{\\prime}\\) and \\(\\mathbf{a}_t\\) to estimate local offsets, where \\(\\mathbf{a}_{ref}^{\\prime}\\) is the aligned \\(\\mathbf{a}_{ref}\\) which is downsampled, deformed, and upsampled with the same weights \\(\\mathbf{w}_g\\) used to synthesize \\(\\mathbf{f}_{ref}^{\\prime}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.346, + 0.892, + 0.572 + ], + "angle": 0, + "content": "We estimate the offsets for the decoding layer features \\(\\mathbf{f}\\) using the encoding layer features a under the assumption that \\(\\mathbf{a}_t\\) and \\(\\mathbf{f}_t\\) have the same structural information. In video-to-video translation, input and output frames share the same underlying structure. Thus, it is natural for the network to learn to maintain the structural information of an input frame throughout the encoding and decoding process. More details are described in our supplementary materials. Adaptive Blending and Deformation. During the local alignment, we also take advantage of the current frame features \\(\\mathbf{a}_t\\) from the encoding layer to handle the regions with large motion differences and new objects. To achieve this in a cost-efficient way, we introduce AdaBD, which simultaneously aligns \\(\\mathbf{f}_{ref}^{\\prime}\\) and blends it with \\(\\mathbf{a}_t\\) in an adaptive manner, as illustrated in Fig. 1(b) AdaBD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.721 + ], + "angle": 0, + "content": "First, our local offset/mask generator predicts a blending mask \\(\\mathbf{m}_b\\in \\mathbb{R}^{N_{\\mathbf{P}}\\times H\\times W}\\) in addition to the local offsets \\(\\Delta \\mathbf{p}_l\\in \\mathbb{R}^{2N_{\\mathbf{P}}\\times H\\times W}\\). A learnable mask \\(\\mathbf{m}_b\\) is composed of modulation scalars ranging from 0 to 1, each of which indicates the blending ratio of the current features \\(\\mathbf{a}_t\\) to the aligned reference features. While DCNv2 [39] originally introduces the modulation scalars to control feature amplitudes of a single input, we leverage the scalars to adaptively blend the features from two adjacent frames considering their redundant areas." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.723, + 0.892, + 0.814 + ], + "angle": 0, + "content": "We apply deformable convolution by adding local offsets \\(\\Delta \\mathbf{p}_l\\) to sampling positions of the coarsely-aligned reference features \\(\\mathbf{f}_{ref}^{\\prime}\\), while the current frame features \\(\\mathbf{a}_t\\) are fed into standard convolutional operations. Concurrently, blending mask \\(\\mathbf{m}_b\\) adaptively combines the two feature maps. In detail, an output point \\(\\mathbf{p_o}\\) of \\(\\hat{\\mathbf{f}}_t\\) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\mathbf {f}} _ {t} (\\mathbf {p} _ {\\mathbf {o}}) = \\sum_ {k = 1} ^ {N _ {\\mathbf {p}}} \\mathbf {w} _ {l} (\\mathbf {p} _ {k}) \\cdot \\left\\{\\mathbf {a} _ {t} (\\mathbf {p} _ {\\mathbf {o}} + \\mathbf {p} _ {k}) \\cdot \\mathbf {m} _ {b} (\\mathbf {p} _ {k}) + \\right. \\\\ \\mathbf {f} _ {r e f} ^ {\\prime} \\left(\\mathbf {p} _ {\\mathbf {o}} + \\mathbf {p} _ {k} + \\Delta \\mathbf {p} _ {l} (\\mathbf {p} _ {k})\\right) \\cdot (1 - \\mathbf {m} _ {b} (\\mathbf {p} _ {k})) \\}, \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7615" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "where \\(\\mathbf{w}_l\\) indicates weights of local deformable convolution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.47, + 0.182 + ], + "angle": 0, + "content": "Intuitively, the higher values of \\(\\mathbf{m}_b\\) indicate the regions where current frame features are more required. In other words, Eq. 5 can be rewritten as a summation of standard convolution and deformable convolution:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.194, + 0.469, + 0.226 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {f}} _ {t} = f _ {d c} \\left(\\mathbf {w} _ {l}, \\mathbf {a} _ {t}, \\mathbf {0}, \\mathbf {m} _ {b}\\right) + f _ {d c} \\left(\\mathbf {w} _ {l}, \\mathbf {f} _ {r e f} ^ {\\prime}, \\Delta \\mathbf {p} _ {l}, 1 - \\mathbf {m} _ {b}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.227, + 0.471, + 0.332 + ], + "angle": 0, + "content": "In this equation, the convolutional weights \\(\\mathbf{w}_l\\) are shared between \\(\\mathbf{a}_t\\) and \\(\\mathbf{f}_{ref}^{\\prime}\\). \\(f_{dc}\\) with 0 indicates DCN with zero offsets, illustrating standard convolutional operation. To save computational costs, we decrease the channel dimension of all input features and reconstruct the original channel size before injecting the output features into the remaining layers of the teacher network." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.342, + 0.266, + 0.359 + ], + "angle": 0, + "content": "3.3. Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.47, + 0.546 + ], + "angle": 0, + "content": "To train Shortcut-V2V, we mainly leverage alignment loss, distillation loss, and GAN losses widely used in image/video translation networks [35, 34, 38]. First, we adopt alignment loss \\( L_{\\text{align}} \\) to train the deformation layers in Shortcut-V2V. Since Shortcut-V2V aims to align the reference frame features \\( \\mathbf{f}_{ref} \\) with the current frame, \\( L_{\\text{align}} \\) computes L1 loss between the aligned feature \\( \\mathbf{f}_{ref}^* \\) and the current frame features \\( \\mathbf{f}_t \\) extracted from the teacher model. To obtain \\( \\mathbf{f}_{ref}^* \\), we align \\( \\mathbf{f}_{ref} \\) in a coarse-to-fine manner without an intervention of \\( \\mathbf{m}_b \\) or blending with the current features. The alignment loss \\( L_{\\text{align}} \\) is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.548, + 0.469, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {r e f} ^ {*} = f _ {d c} \\left(\\mathbf {w} _ {l}, \\mathbf {f} _ {r e f} ^ {\\prime}, \\Delta \\mathbf {p} _ {l}, 1\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.587, + 0.469, + 0.607 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {a l i g n}} = \\left\\| \\mathbf {f} _ {t} - \\mathbf {f} _ {\\text {r e f}} ^ {*} \\right\\| _ {1}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Additionally, we employ knowledge distillation losses at the feature and output levels. A feature-level distillation loss \\( L_{feat} \\) is applied between the estimated feature \\( \\hat{\\mathbf{f}}_t \\) and the ground truth feature \\( \\mathbf{f}_t \\), while an output-level distillation loss \\( L_{out} \\) compares the approximated output \\( \\hat{\\mathbf{O}}_t \\) to the output \\( \\mathbf{O}_t \\) generated by the teacher network. The perceptual loss [37] \\( L_{perc} \\) is also incorporated to distill the high-frequency information of the outputs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Lastly, we utilize a typical GAN loss \\( L_{GAN} \\) and a temporal GAN loss \\( L_{T - GAN} \\), following the existing video-based frameworks [34, 3]. Temporal GAN loss \\( L_{T - GAN} \\) encourages both temporal consistency and realisticity of the output frames. For the GAN losses, we consider the outputs of the teacher network as real images." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.825, + 0.431, + 0.84 + ], + "angle": 0, + "content": "The overall objective function \\( L_{total} \\) is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.852, + 0.484, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {t o t a l} = \\lambda_ {\\text {a l i g n}} L _ {\\text {a l i g n}} + \\lambda_ {\\text {f e a t}} L _ {\\text {f e a t}} + \\lambda_ {\\text {o u t}} L _ {\\text {o u t}} \\\\ + \\lambda_ {p e r c} L _ {p e r c} + \\lambda_ {G A N} L _ {G A N} + \\lambda_ {T - G A N} L _ {T - G A N}, \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.088, + 0.892, + 0.349 + ], + "angle": 0, + "content": "
ModelDatasetMethodFVDMACs (G)Param. (M)
UnsupV2COriginal1.25384.617.84
CAT1.31022.831.48
OMGD1.19925.851.21
Ours1.18018.190.32
L2VOriginal1.66342.367.84
CAT1.95712.021.71
OMGD2.46712.931.21
Ours1.7549.090.32
vid2vidE2FOriginal0.1862066.69365
CAT0.288380.5524.67
OMGD0.264485.8435.81
Fast-Vid2Vid0.223767.6363.17
Ours0.209359.998.29
L2COriginal0.1461254.17411.34
CAT0.304399.3671.11
OMGD0.346391.5947.83
Fast-Vid2Vid0.310455.2971.8
Ours0.165389.1652.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.534, + 0.359, + 0.857, + 0.374 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison with baselines." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.388, + 0.875, + 0.527 + ], + "angle": 0, + "content": "
DatasetMethodmIoUACMP
V2LOriginal12.216.062.8
CAT5.998.8244.2
OMGD9.7613.058.7
Ours11.515.261.4
L2VOriginal10.015.647.3
CAT4.107.8127.6
OMGD3.426.8729.5
Ours9.2414.943.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.538, + 0.892, + 0.568 + ], + "angle": 0, + "content": "Table 2: Comparison of segmentation score for V2L and FCN-score for L2V on Unsup." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.596, + 0.892, + 0.657 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\text{align}}\\), \\(\\lambda_{\\text{feat}}\\), \\(\\lambda_{\\text{out}}\\), \\(\\lambda_{\\text{perc}}\\), \\(\\lambda_{\\text{GAN}}\\), and \\(\\lambda_{T - \\text{GAN}}\\) are hyperparameters to control relative significance among the losses. More details on training objectives are described in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.671, + 0.634, + 0.687 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.707, + 0.712 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Models and Datasets. To demonstrate the generality of our model, we conduct experiments on various tasks with widely-used video-to-video translation models, Unsupervised RecycleGAN [32] and vid2vid [34]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Unsupervised RecycleGAN (Unsup) is the state-of-art video-to-video translation model among CycleGAN-based [38] ones. We conduct experiments for Unsup on Viper \\(\\rightarrow\\) Cityscapes (V2C), which involves translating inter-modality from game driving scenes (Viper [23]) to real-world driving scenes (Cityscapes [5]). The Viper dataset consists of driving scenes collected from the GTA-V game engine, with 56 training videos and 21 test videos." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7616" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.093, + 0.881, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.389, + 0.893, + 0.419 + ], + "angle": 0, + "content": "Figure 2: Qualitative comparison. The upper rows represent the results of Unsup V2C, L2V, and V2L. The bottom rows show the results of vid2vid E2F and L2C." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.47, + 0.552 + ], + "angle": 0, + "content": "Cityscapes is a dataset composed of 2,975 training videos and 500 validation videos. The input images are resized into \\(256 \\times 512\\) for the experiments. We also tackle the translation of the videos from the Viper dataset to their corresponding segmentation label maps, \\(Viper \\rightarrow Label (V2L)\\), and vice versa, \\(Label \\rightarrow Viper (L2V)\\). We resize the images and the labels into \\(256 \\times 256\\) following the previous work [1, 32]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.779 + ], + "angle": 0, + "content": "vid2vid is a widely-used pix2pixHD-based video-to-video translation network that serves as a base architecture for various recent video-to-video translation models [33, 18]. Following vid2vid, we evaluate ShortcutV2V on Edge \\(\\rightarrow\\) Face (E2F) and Label \\(\\rightarrow\\) Cityscapes (L2C). E2F translates edge maps into facial videos from the FaceForensics [24], containing 704 videos for training and 150 videos for validation with various lengths. The images are cropped and resized to \\(512 \\times 512\\) for the experiments. The edge maps are extracted using the estimated facial landmarks and Canny edge detector. Also, L2C synthesizes videos of driving scenes from segmentation label maps using the Cityscapes. We generate segmentation maps using pretrained networks following the previous studies [34, 40]. The images and the labels are resized into \\(256 \\times 512\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. Primarily, we adopt the Fréchet video distance (FVD) score [31] to evaluate the performance of Shortcut-V2V quantitatively. The FVD score measures the Fréchet Inception distance between the distribution of video-level features extracted from generated and real videos. The lower the FVD score is, the better visual quality and temporal consistency the generated video frames have. For V2L and L2V, we follow the measure" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.893, + 0.598 + ], + "angle": 0, + "content": "ment of the evaluation metrics in the teacher model, Unsup [32]. We measure the segmentation scores, mean intersection over union (mIoU), mean pixel accuracy (MP), and average class accuracy (AC), to validate the performance of V2L. L2V is evaluated using FCN-score [17]. For the evaluation, we first estimate label maps from the generated videos using the FCN model pretrained with the Viper dataset. Then, we measure how accurately the estimated label maps are mapped to the ground truth segmentation labels. Higher FCN-scores refer to better output quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Implementation Details. We attach our Shortcut block to the fixed teacher networks implemented based on the official codes and pretrained by the authors, except for Unsup V2C. The teacher network of Unsup V2C is trained from scratch in the same way the original paper described [32]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Also, standard convolutional kernels in the Shortcut block are replaced with HetConv [26] to further enhance computational efficiency without a performance drop. For the convenience of implementation and training stability, we intend a and \\( \\mathbf{f} \\) to have the same spatial size. Furthermore, we set the max interval \\( \\alpha \\) for each dataset considering the factors that reflect motion differences between frames, such as the frame per second (FPS) of training videos. Lower FPS usually results in larger motion differences between the adjacent frames, requiring a shorter max interval \\( \\alpha \\) and vice versa. Specifically, we set \\( \\alpha \\) as 3 on V2C, V2L, L2V, and L2C, since the FPS of Viper [23] and Cityscapes [5] is 15 and 17, respectively. Also, \\( \\alpha \\) of E2F is set as 6, where the FPS of FaceForensics [24] is 30. Additional details are included in the supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7617" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.273, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.203, + 0.232, + 0.216 + ], + "angle": 0, + "content": "(a) Unsup V2C" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.09, + 0.468, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.203, + 0.431, + 0.215 + ], + "angle": 0, + "content": "(b) vid2vid E2F" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.233, + 0.47, + 0.324 + ], + "angle": 0, + "content": "Figure 3: Performance-efficiency trade-off of the original model, Shortcut-V2V, and the existing compression methods including OMGD [22], CAT [12], and FastVid2Vid [40]. We measure the FVD score and MACs, where the lower FVD score indicates better quality. Red points and stars denote ours with various model sizes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.355, + 0.303, + 0.371 + ], + "angle": 0, + "content": "4.2. Comparison to Baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.47, + 0.576 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our framework, we conduct qualitative and quantitative evaluations compared to the original model and other baselines. Since this is the first work that tackles a generally applicable compression framework for video-to-video translation, we compare Shortcut-V2V to the existing compression methods for image-to-image translation, CAT [12] and OMGD [22], regarding video frames as individual images. In the case of vid2vid, we additionally conduct a comparison to FastVid2Vid [40], which is the compression method designed specifically for vid2vid. For a fair comparison, we compress the student networks of the baselines to have similar or higher MACs compared to our model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.47, + 0.716 + ], + "angle": 0, + "content": "Qualitative Evaluation. According to Fig. 2, our method presents outputs of comparable visual quality to the original model with much fewer computations. In contrast, CAT on Unsup V2C generates undesirable buildings in the sky, and OMGD on Unsup V2C struggles with noticeable artifacts. For Unsup L2V, CAT and OMGD generate unrealistic textures for the terrain in the middle or the vegetation on the right. Moreover, in V2L, CAT estimates inappropriate labels on the sky and OMGD on the vegetation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.468, + 0.853 + ], + "angle": 0, + "content": "For vid2vid E2F, CAT shows unwanted artifacts on the mouth, and OMGD presents inconsistent outputs. For vid2vid L2C, the outputs of CAT are blurry, especially for the trees in the background, and OMGD generates artifacts at the bottom of the images. Although Fast-Vid2Vid shows reasonable image quality, the output frames are inaccurately aligned with the input (e.g., a head pose of a person) and suffer from ghost effects (e.g., the black car on the left) due to motion compensation using interpolation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Quantitative Evaluation. As shown in Table 1 and Table 2, our framework successfully improves the computational efficiency of the original network without a significant perfor" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.892, + 0.188 + ], + "angle": 0, + "content": "
ConfigurationsUnsup V2Cvid2vid E2F
(a) w/o reference features1.2560.398
(b) w/o current features1.2490.218
(c) Single-stage alignment1.1950.213
(d) w/o adaptive blending1.2080.244
(e) Ours1.1800.209
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.198, + 0.89, + 0.227 + ], + "angle": 0, + "content": "Table 3: An ablation study on Unsup V2C and vid2vid E2F. We measure the FVD scores." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.257, + 0.892, + 0.346 + ], + "angle": 0, + "content": "mance drop. In the case of Unsup, Shortcut-V2V reduces the MACs by \\(4.7 \\times\\) and the number of parameters by \\(24.5 \\times\\). In addition, our approach saves vid2vid's MACs by 3.2 and \\(5.7 \\times\\) and the number of parameters by 7.8 and \\(44 \\times\\) on each task. Fig. 3 visualizes the performance-efficiency trade-off of our framework and other baselines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.893, + 0.531 + ], + "angle": 0, + "content": "According to Table 1, we also outperform other compression methods for image-to-image translation, CAT [12] and OMGD [22], even with fewer computations. Image-based approaches cannot consider temporal coherence among the frames during the compression, leading to a loss of quality. Meanwhile, we effectively preserve the original performance by exploiting rich information in the previous features during inference. Shortcut-V2V even shows superiority over Fast-Vid2Vid [40], which is a compression method specifically designed for vid2vid. Table 2 also demonstrates that ours on Unsup surpasses the existing compression models by a large margin." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.542, + 0.653, + 0.559 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.656 + ], + "angle": 0, + "content": "We conduct an ablation study to evaluate the effect of each component of Shortcut-V2V. As described in Table 3, we compare the FVD scores of 5 different configurations, (a) w/o reference features, (b) w/o current features, (c) single-stage alignment, (d) w/o adaptive blending, and (e) ours, on Unsup V2C and vid2vid E2F." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.893, + 0.901 + ], + "angle": 0, + "content": "First, while our model originally exploits both the reference features \\(\\mathbf{f}_{ref}\\) and the current frame features \\(\\mathbf{a}_t\\) to synthesize \\(\\hat{\\mathbf{f}}_t\\), (a) and (b) are designed to leverage only either of them. To be specific, (a) synthesizes \\(\\hat{\\mathbf{f}}_t\\) by processing standard convolutions on \\(\\mathbf{a}_t\\) without integrating the deformed \\(\\mathbf{f}_{ref}\\). Meanwhile, (b) estimates \\(\\hat{\\mathbf{f}}_t\\) using only \\(\\mathbf{f}_{ref}\\) aligned in a coarse-to-fine manner without blending \\(\\mathbf{a}_t\\). The results demonstrate that the absence of either \\(\\mathbf{f}_{ref}\\) or \\(\\mathbf{a}_t\\) leads to performance degradation. Next, (c) performs a single-stage alignment on \\(\\mathbf{f}_{ref}\\) to show the necessity of a coarse-to-fine alignment. Specifically, (c) leverages a single offset/mask generator to predict \\(\\Delta \\mathbf{p}\\) and \\(\\mathbf{m}_b\\) which are then used in AdaBD for blending and deformation. The result indicates that coarse-to-fine alignment effectively assists the estimation of \\(\\hat{\\mathbf{f}}_t\\) by minimizing misalignment between the features from adjacent frames. This makes Shortcut-V2V suitable for" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7618" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.096, + 0.227, + 0.108 + ], + "angle": 0, + "content": "Reference frame" + }, + { + "type": "image", + "bbox": [ + 0.109, + 0.109, + 0.241, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.096, + 0.418, + 0.108 + ], + "angle": 0, + "content": "Current frame" + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.11, + 0.439, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.172, + 0.27, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.173, + 0.374, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.173, + 0.469, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.269, + 0.47, + 0.375 + ], + "angle": 0, + "content": "Figure 4: Visualization of global/local offsets and blending masks. The green points denote output points of the current frame, and the red points around them are each output point's sampling positions modified by global offsets \\(\\Delta \\mathbf{p}_g\\) and local offsets \\(\\Delta \\mathbf{p}_l\\). The values in a blending mask \\(\\mathbf{m}_b\\) are averaged by the kernel size for visualization. The brighter area indicates higher mask values." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.47, + 0.492 + ], + "angle": 0, + "content": "generating videos with diverse motion differences. Lastly, (d) blends \\(\\mathbf{f}_{ref}^{*}\\) and \\(\\mathbf{a}_t\\) simply by element-wise addition instead of adaptive blending with \\(\\mathbf{m}_b\\) in AdaBD. The result demonstrates that \\(\\mathbf{m}_b\\) also encourages better estimation of \\(\\hat{\\mathbf{f}}_t\\) by selectively exploiting features from temporally adjacent features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.499, + 0.314, + 0.513 + ], + "angle": 0, + "content": "4.4. Offset/Mask Visualization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.794 + ], + "angle": 0, + "content": "We provide a qualitative analysis of the generated global/local offsets and blending masks. Fig. 4 visualizes the sampling positions (red points) for each output position \\(\\mathbf{p}_o\\) of the current frame (green points), where the predicted offsets \\(\\Delta \\mathbf{p}_g\\) and \\(\\Delta \\mathbf{p}_l\\) are added to the original sampling positions. According to the result, global offsets \\(\\Delta \\mathbf{p}_g\\) effectively reflect the global movement of the objects such as the bridge. Also, the summation of the global and local offsets indicates the sampling points refined by local offsets enabling fine alignment. This shows that the estimated offsets effectively support the utilization of common features in the reference frame. For the blending masks \\(\\mathbf{m}_b\\), Fig. 4 presents that the regions with significant motion differences (e.g., trees) have large mask values compared to regions with little change (e.g., road, sky). That is, our model relies more on the current features rather than the reference features when the deformation is challenging, which aligns with our intention and leads to robust performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.373, + 0.818 + ], + "angle": 0, + "content": "4.5. Performance-Efficiency Trade-off" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.871 + ], + "angle": 0, + "content": "We present a performance-efficiency trade-off of our framework with respect to varying model sizes of Shortcut block and teacher model dependence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Shortcut Block Size. We construct our models of different sizes by reducing the number of the output chan" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.892, + 0.148 + ], + "angle": 0, + "content": "
DependenceUnsup V2Cvid2vid E2F
FVDMACs (G)Param. (M)FVDMACs (G)Param. (M)
Low1.22113.790.140.277243.972.11
Medium1.18018.190.320.209359.998.29
High1.16627.971.380.193475.9932.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.158, + 0.892, + 0.204 + ], + "angle": 0, + "content": "Table 4: Shortcut-V2V performance with different teacher model dependence. High dependency denotes using more teacher network layers during the inference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.892, + 0.491 + ], + "angle": 0, + "content": "nel of \\(1 \\times 1\\) channel reduction layer by half (Ours-1/2) and a quarter (Ours-1/4). As shown in Fig. 3, despite the performance-efficiency trade-off depending on channel dimension, Ours-1/2 and Ours-1/4 still achieve comparable FVD to the baselines with less computation costs. On a single RTX 3090 GPU, our models with three different sizes on Unsup achieve 99.67, 109.69, and 111.80 FPS with an \\(11.6 - 21.2\\%\\) reduction in inference time compared to the original model's speed of 88.10 FPS. Additionally, Ours, Ours-1/2, and Ours-1/4 on vid2vid contribute to a significant improvement in speed from 5.63 FPS to 12.86, 12.93, and 13.16 FPS, resulting in a \\(56.2 - 57.2\\%\\) reduction in inference time. Our method is capable of a real-time inference, unlike the previous work [40] which requires future frames for the current frame inference due to motion compensation. More details on channel manipulation are described in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Teacher Model Dependence. Since Shortcut-V2V leverages a subset of the teacher network during inference, the computational costs and memory usage may vary depending on the amount of teacher network we use. In this regard, we present an analysis of the computational efficiency and performance of our model with respect to various levels of teacher model dependence. We categorize the dependence on the teacher model into three levels, low, medium, and high, where high dependence indicates using more teacher network layers. According to Table 4, our model of higher dependence achieves better performance with larger MACs and the number of parameters. The results also demonstrate that Shortcut-V2V can leverage temporal redundancy in features extracted from various layers of the teacher network. Further details for teacher model dependence and experiments for temporal redundancy are included in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.768, + 0.748, + 0.784 + ], + "angle": 0, + "content": "5. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Although Shortcut-V2V significantly improves the test-time efficiency of video-to-video translation, it still poses several limitations. First, a constant max interval may induce unsatisfactory outputs when a degree of temporal redundancy largely varies between frames. Following recent studies [25, 20, 19], applying an adaptive interval based on frame selection algorithms or a learnable policy net-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7619" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.197 + ], + "angle": 0, + "content": "work could be promising future research. In addition, we need to manually configure various hyperparameters such as channel dimension, max interval, and teacher model dependence, which can be automated by NAS [13]. Lastly, our method still has the potential to further improvement of computational efficiency by compressing the teacher model using other methods before applying our framework." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.199, + 0.468, + 0.363 + ], + "angle": 0, + "content": "Despite the limitations, Shortcut-V2V presents a significant improvement in test-time efficiency of video-to-video translation networks based on temporal redundancy reduction, while preserving the original model performance. Extensive experiments with widely-used video-to-video translation models successfully demonstrate the general applicability of our framework. To the best of our knowledge, this is the first work for a general model compression in the domain of video-to-video translation. We hope our work facilitates research on video-to-video translation and extends the range of its application." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.371, + 0.468, + 0.475 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2021-0-02068, Artificial Intelligence Innovation Hub) and the Ministry of Culture, Sports and Tourism and Korea Creative Content Agency (Project Number: R2021040097, Contribution Rate: 50)." + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.515, + 0.957 + ], + "angle": 0, + "content": "7620" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Aayush Bansal, Shugao Ma, Deva Ramanan, and Yaser Sheikh. Recycle-gan: Unsupervised video retargeting. Proc. of the European Conference on Computer Vision (ECCV), 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.47, + 0.226 + ], + "angle": 0, + "content": "[2] Kelvin C. K. Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Understanding deformable alignment in video super-resolution. Proc. the AAAI Conference on Artificial Intelligence (AAAI), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.47, + 0.282 + ], + "angle": 0, + "content": "[3] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston H. Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. Proc. of the IEEE International Conference on Computer Vision (ICCV), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.47, + 0.325 + ], + "angle": 0, + "content": "[4] Yang Chen, Yingwei Pan, Ting Yao, Xinmei Tian, and Tao Mei. Mocycle-gan: Unpaired video-to-video translation. In ACM Multimedia (ACM MM), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.47, + 0.408 + ], + "angle": 0, + "content": "[5] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.47, + 0.464 + ], + "angle": 0, + "content": "[6] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. Proc. of the IEEE International Conference on Computer Vision (ICCV), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.47, + 0.522 + ], + "angle": 0, + "content": "[7] Jianing Deng, Li Wang, Shiliang Pu, and Cheng Zhuo. Spatio-temporal deformable convolution for compressed video quality enhancement. Proc. the AAAI Conference on Artificial Intelligence (AAAI), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.47, + 0.563 + ], + "angle": 0, + "content": "[8] Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.47, + 0.606 + ], + "angle": 0, + "content": "[9] Amirhossein Habibian, Haitam Ben Yahia, Davide Abati, Efstratos Gavves, and Fatih Porikli. Delta distillation for efficient video processing, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.607, + 0.47, + 0.661 + ], + "angle": 0, + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.664, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[11] Bangrui Jiang, Zhihuai Xie, Zhen Xia, Songnan Li, and Shan Liu. Erdn: Equivalent receptive field deformable network for video deblurring. In Proc. of the European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.72, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[12] Qing Jin, Jian Ren, Oliver J. Woodford, Jiazhuo Wang, Geng Yuan, Yanzhi Wang, and Sergey Tulyakov. Teachers do more than teach: Compressing image-to-image models. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.789, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[13] Muyang Li, Ji Lin, Yaoyao Ding, Zhijian Liu, Jun-Yan Zhu, and Song Han. GAN compression: Efficient architectures for interactive conditional gans. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Ji Lin, Chuang Gan, and Song Han. Temporal shift module for efficient video understanding. Proc. of the IEEE International Conference on Computer Vision (ICCV), 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[15] Jiayi Lin, Yan Huang, and Liang Wang. Fdan: Flow-guided deformable alignment network for video super-resolution. arXiv preprint arXiv:2105.05640, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[16] Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient semantic video segmentation with per-frame inference. Proc. of the European Conference on Computer Vision (ECCV), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[17] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[18] Arun Mallya, Ting-Chun Wang, Karan Sapra, and Ming-Yu Liu. World-consistent video-to-video synthesis. Proc. of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[19] Yue Meng, Rameswar Panda, Chung-Ching Lin, Prasanna Sattigeri, Leonid Karlinsky, Kate Saenko, Aude Oliva, and Rogério Feris. Adafuse: Adaptive temporal fusion network for efficient action recognition. Proc. the International Conference on Learning Representations (ICLR), 2021. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[20] Bowen Pan, Rameswar Panda, Camilo Fosco, Chung-Ching Lin, Alex Andonian, Yue Meng, Kate Saenko, Aude Oliva, and Rogerio Feris. Va-red²: Video adaptive redundancy reduction. Proc. the International Conference on Learning Representations (ICLR), 2021. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[21] KwanYong Park, Sanghyun Woo, Dahun Kim, Donghyeon Cho, and In So Kweon. Preserving semantic and temporal consistency for unpaired video-to-video translation. ACM Multimedia (ACMM), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[22] Yuxi Ren, Jie Wu, Xuefeng Xiao, and Jianchao Yang. Online multi-granularity distillation for gan compression. Proc. of the IEEE International Conference on Computer Vision (ICCV), 2021. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.554, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[23] Stephan R. Richter, Zeeshan Hayden, and Vladlen Koltun. Playing for benchmarks. In Proc. of the IEEE International Conference on Computer Vision (ICCV), 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.596, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[24] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. FaceForensics++: Learning to detect manipulated facial images. In Proc. of the IEEE International Conference on Computer Vision (ICCV), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[25] Jonghyeon Seon, Jaedong Hwang, Jonghwan Mun, and Bohyung Han. Stop or forward: Dynamic layer skipping for efficient action recognition. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.721, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[26] Pravendra Singh, Vinay Kumar Verma, Piyush Rai, and Vinay P. Namboodiri. Hetconv: Heterogeneous kernel-based convolutions for deep cnns. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.777, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[27] Ximeng Sun, Rameswar Panda, Chun-Fu Chen, Aude Oliva, Rogério Feris, and Kate Saenko. Dynamic network quantization for efficient video inference. Proc. of the IEEE International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7621" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[29] Ryan Szeto, Mostafa El-Khamy, Jungwon Lee, and Jason J. Corso. Hypercon: Image-to-video model transfer for video-to-video translation tasks. IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.471, + 0.205 + ], + "angle": 0, + "content": "[30] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. TDAN: temporally deformable alignment network for video super-resolution. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.275 + ], + "angle": 0, + "content": "[31] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. In Proc. the International Conference on Learning Representations Workshop (ICLRW), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.471, + 0.346 + ], + "angle": 0, + "content": "[32] Kaihong Wang, Kumar Akash, and Teruhisa Mitsu. Learning temporally and semantically consistent unpaired video-to-video translation through pseudo-supervision from synthetic optical flow. Proc. the AAAI Conference on Artificial Intelligence (AAAI), 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.471, + 0.403 + ], + "angle": 0, + "content": "[33] Ting-Chun Wang, Ming-Yu Liu, Andrew Tao, Guilin Liu, Jan Kautz, and Bryan Catanzaro. Few-shot video-to-video synthesis. Proc. the Advances in Neural Information Processing Systems (NeurIPS), 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.471, + 0.459 + ], + "angle": 0, + "content": "[34] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Proc. the Advances in Neural Information Processing Systems (NeurIPS), 2018. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.471, + 0.53 + ], + "angle": 0, + "content": "[35] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.471, + 0.599 + ], + "angle": 0, + "content": "[36] Xintao Wang, Kelvin C. K. Chan, Ke Yu, Chao Dong, and Chen Change Loy. EDVR: video restoration with enhanced deformable convolutional networks. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.471, + 0.669 + ], + "angle": 0, + "content": "[37] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.471, + 0.739 + ], + "angle": 0, + "content": "[38] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proc. of the IEEE International Conference on Computer Vision (ICCV), 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.742, + 0.471, + 0.797 + ], + "angle": 0, + "content": "[39] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.799, + 0.471, + 0.854 + ], + "angle": 0, + "content": "[40] Long Zhuo, Guangcong Wang, Shikai Li, Wayne Wu, and Ziwei Liu. Fast-vid2vid: Spatial-temporal compression for video-to-video synthesis. Proc. of the European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.854 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7622" + } + ] +] \ No newline at end of file diff --git a/2023/Shortcut-V2V_ Compression Framework for Video-to-Video Translation Based on Temporal Redundancy Reduction/99da4ac0-6f5c-433f-bc7b-caa55aa672d5_origin.pdf b/2023/Shortcut-V2V_ Compression Framework for Video-to-Video Translation Based on Temporal Redundancy Reduction/99da4ac0-6f5c-433f-bc7b-caa55aa672d5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2282b8f9bc5966018d268dd415214a08d75cbe0a --- /dev/null +++ b/2023/Shortcut-V2V_ Compression Framework for Video-to-Video Translation Based on Temporal Redundancy Reduction/99da4ac0-6f5c-433f-bc7b-caa55aa672d5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c5c32ea2ea5b2b1c58e4060b07650b61953a64fa0954197efb14bdba2a73db5 +size 4346831 diff --git a/2023/Shortcut-V2V_ Compression Framework for Video-to-Video Translation Based on Temporal Redundancy Reduction/full.md b/2023/Shortcut-V2V_ Compression Framework for Video-to-Video Translation Based on Temporal Redundancy Reduction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ef0d675bd8506415d9cfc01ea7a32cd5a13d3723 --- /dev/null +++ b/2023/Shortcut-V2V_ Compression Framework for Video-to-Video Translation Based on Temporal Redundancy Reduction/full.md @@ -0,0 +1,311 @@ +# Shortcut-V2V: Compression Framework for Video-to-Video Translation based on Temporal Redundancy Reduction + +Chaeyeon Chung\*, Yeojeong Park\*, Seunghwan Choi, Munkhsoyol Ganbat, and Jaegul Choo + +$^{1}$ KAIST AI $^{2}$ KT Research & Development Center, KT Corporation {cy-chung, indigopyj, shadow2496, soyol, jchoo}@kaist.ac.kr + +# Abstract + +Video-to-video translation aims to generate video frames of a target domain from an input video. Despite its usefulness, the existing networks require enormous computations, necessitating their model compression for wide use. While there exist compression methods that improve computational efficiency in various image/video tasks, a generally-applicable compression method for video-to-video translation has not been studied much. In response, we present Shortcut-V2V, a general-purpose compression framework for video-to-video translation. Shortcut-V2V avoids full inference for every neighboring video frame by approximating the intermediate features of a current frame from those of the previous frame. Moreover, in our framework, a newly-proposed block called AdaBD adaptively blends and deforms features of neighboring frames, which makes more accurate predictions of the intermediate features possible. We conduct quantitative and qualitative evaluations using well-known video-to-video translation models on various tasks to demonstrate the general applicability of our framework. The results show that Shortcut-V2V achieves comparable performance compared to the original video-to-video translation model while saving $3.2 - 5.7 \times$ computational cost and $7.8 - 44 \times$ memory at test time. Our code and videos are available at https://shortcut-v2v.github.io/. + +# 1. Introduction + +Video-to-video translation is a task of generating temporally consistent and realistic video frames of a target domain from a given input video. Recent studies on video-to-video translation present promising performance in various domains such as inter-modality translation between labels + +and videos [34, 33, 18], and intra-modality translation between driving scene videos [32] or face videos [1]. + +Despite enhanced usefulness, video-to-video translation networks usually require substantial computational cost and memory usage, which limits their applicability. For instance, multiply-accumulates (MACs) of a widely-used video translation model, vid2vid [34], is 2066.69G, while the basic convolutional neural networks, ResNet v2 50 [10] and Inception v3 [28], are 4.12G and 6G, respectively. Furthermore, temporally redundant computations for adjacent video frames also harm the cost efficiency of a video-to-video translation network. Performing full inference for every neighboring video frame that contains common visual features inevitably entails redundant operations [20, 19]. + +In this regard, Fast-Vid2Vid [40] proposes a compression framework for vid2vid [34] based on spatial input compression and temporal redundancy reduction. However, it cannot be applied to other video-to-video translation models since it is designed specifically for vid2vid. Moreover, Fast-Vid2Vid does not support real-time inference since it requires future frames to infer a current one. Alternatively, one can apply model compression approaches for image-to-image translation [13, 22, 12] directly to video-to-video translation, considering video frames as separate images. However, these approaches are not designed to consider the correlation among adjacent video frames during the compression. This may result in unrealistic output quality in video-to-video translation, where the inherent temporal coherence of an input video needs to be preserved in the outputs. Also, frame-by-frame inference without temporal redundancy reduction involves unnecessary computations, resulting in computational inefficiency. + +In this paper, we propose Shortcut-V2V, a general-purpose framework for improving the computational efficiency of video-to-video translation based on temporal redundancy reduction. Shortcut-V2V allows the original video-to-video translation model to avoid temporally re + +![](images/ccd7fcbcdba55580a15d96a0c2daea7d8ec466b9c324b414c912480b8bce99f5.jpg) +Figure 1: Overview of Shortcut-V2V. (a) is an overall framework of Shorcut-V2V, and (b) shows a detailed architecture of Shortcut block. $\uparrow 2$ and $\downarrow 2$ refer to upsampling and downsampling by a factor of 2, respectively. G in Offset G and Offset/Mask G indicates a generator. + +![](images/e709cafb05b43f5baf3babe234028bcd445d0aee9731a366389b0916160169dd.jpg) + +dundant computations by approximating the decoding layer features of the current frame with largely reduced computations. To enable lightweight estimation, our framework leverages features from the previous frame (i.e., reference features), which have high visual similarity with the current frame. We also exploit current frame features from the encoding layer to handle newly-appeared regions in the current frame. Specifically, we first globally align the previous frame features with the current frame features, and our novel Adaptive Blending and Deformation block (AdaBD) in Shortcut-V2V blends features of neighboring frames while performing detailed deformation. AdaBD adaptively integrates the features regarding their redundancy in a lightweight manner. + +In this way, our model significantly improves the test-time efficiency of the original network while preserving its original performance. Shortcut-V2V is easily applicable to a pretrained video-to-video translation model to save computational cost and memory usage. Our framework is also suitable for real-time inference since we do not require future frames for the current frame inference. To the best of our knowledge, this is the first attempt at a general-purpose model compression approach for video-to-video translation. We demonstrate the effectiveness of our approach using well-known video-to-video translation models, Unsupervised RecycleGAN [32] (Unsup) and vid2vid [34]. Shortcut-V2V reduces $3.2 - 5.7 \times$ computational cost and $7.8 - 44 \times$ memory usage while achieving comparable performance to the original model. Since there is no existing general-purpose compression method, we compare our method with Fast-Vid2Vid [40] and the compression methods for image-to-image translation. Our model presents superiority over the existing approaches in both quantitative and qualitative evaluations. Our contributions are summarized as follows: + +- We introduce a novel, general-purpose model com + +pression framework for video-to-video translation, Shortcut-V2V, that enables the original network to avoid temporally redundant computations. + +- We present AdaBD that exploits features from neighboring frames via adaptive blending and deformation in a lightweight manner. +- Our framework saves up to $5.7 \times$ MACs and $44 \times$ parameters across various video-to-video translation tasks, achieving comparable performance to the original networks. + +# 2. Related Work + +# 2.1. Video-to-Video Translation + +Recent video-to-video translation networks are generally classified into pix2pixHD-based [35] and CycleGAN-based [38] generator. vid2vid [34] proposes a pix2pixHD-based sequential generation framework that synthesizes a current output given the previous outputs as additional guidance. As the following work, few-shot vid2vid [33] achieves few-shot generalization of vid2vid based on attention modules, and world-consistent vid2vid [18] is proposed to improve long-term temporal consistency of vid2vid. + +While pix2pixHD-based models require paired annotated videos, RecycleGAN [1] and MocycleGAN [4] propose CycleGAN-based video translation models. They exploit spatio-temporal consistency losses to generate realistic videos using unpaired datasets. STC-V2V [21] leverages optical flow for semantic/temporal consistency to improve the output quality of the existing models. Unsupervised RecycleGAN [32] achieves state-of-the-art performance among CycleGAN-based frameworks with a pseudosupervision by the synthetic flow. Although the existing video-to-video translation networks achieve decent performance, they commonly demand a non-trivial amount of + +computational costs and memory usage. Also, frame-by-frame inference necessarily causes redundant operations due to temporal redundancy among adjacent frames. + +# 2.2. Model Compression + +Model compression for video-related tasks has been actively proposed in various domains [19, 20, 16, 9, 27, 8, 14, 29], such as object detection, action recognition, semantic segmentation, and super-resolution. Several studies [19, 9, 20] exploit temporal redundancy among video frames to improve efficiency during training or inference. For instance, Habibian et al. [9] distill only the residual between adjacent frames from a teacher model to a student to speed up the inference. Also, Fast-Vid2Vid [40] firstly proposes a compression framework for video-to-video translation based on spatial and temporal compression. However, Fast-Vid2Vid focuses on vid2vid [34], limiting its application to other video-to-video translation networks. Also, temporal redundancy reduction via motion compensation in Fast-Vid2Vid requires the future frame to infer the current frame, which is not suitable for real-time inference. + +Alternatively, model-agnostic compression methods for image-to-image translation can also be applied to video-to-video translation models. The existing approaches for image synthesis mainly tackle channel pruning [13], knowledge distillation [12, 22], NAS [13], etc. For instance, CAT [12] compresses the teacher network with one-step pruning to satisfy the target computation budget, while OMGD [22] conducts a single-stage online distillation in which the teacher generator supports the student generator to be refined progressively. However, image-based compression methods cannot consider temporal coherence among neighboring video frames, which may induce unrealistic results in video translation tasks. Also, performing full model inference for each video frame still poses computational inefficiency due to the temporal redundancy. + +# 2.3. Deformable Convolution + +Deformable convolution [6] is originally proposed to enhance the transformation capability of a convolutional layer by adding the estimated offsets to a regular convolutional kernel in vision tasks such as object detection or semantic segmentation. Besides its original application, recent studies [30, 7, 36, 2, 11, 15] demonstrate that deformable convolution is also capable of aligning adjacent video frames. TDAN [30] utilizes deformable convolution to capture implicit motion cues between consecutive frames by dynamically predicted offsets in video super-resolution. In addition, EDVR [36] stacks several deformable convolution blocks to estimate large and complex motions in video restoration. In this paper, we also leverage a deformable convolution to adaptively align adjacent video frames in a lightweight manner. + +Algorithm 1 Shortcut-V2V Inference +1: Input: Input video $\{\mathbf{I}_t\}_{t=0}^{N_T-1}$ of length $N_T$ , teacher model $T$ , layer index of encoder $l_e$ and decoder $l_d$ , Shortcut block $S$ , max interval $\alpha$ +2: Output: Output video $\{\mathbf{O}_t\}_{t=0}^{N_T-1}$ +3: for $t = 0$ to $N_T - 1$ do +4: $\mathbf{a}_t = T_{[le]}(\mathbf{I}_t)$ +5: if $t\% \alpha = 0$ then +6: $\mathbf{f}_t = T_{[le + 1;l_d - 1]}(\mathbf{a}_t)$ +7: $\mathbf{a}_{\mathrm{ref}}, \mathbf{f}_{\mathrm{ref}} \gets \mathbf{a}_t, \mathbf{f}_t \triangleright$ Update the reference features +8: else +9: $\mathbf{f}_t = S(\mathbf{f}_{\mathrm{ref}}, \mathbf{a}_{\mathrm{ref}}, \mathbf{a}_t)$ +10: end if +11: $\mathbf{O}_t = T_{[l_d,]}(\mathbf{f}_t)$ +12: end for + +# 3. Shortcut-V2V + +In this paper, we propose Shortcut-V2V, a general compression framework to improve the test-time efficiency in video-to-video translation. As illustrated in Fig. 1(a), given $\{\mathbf{I}_t\}_{t=0}^{N_T-1}$ as input video frames, we first use full teacher model $T$ to synthesize the output of the first frame. Then, for the next frames, our newly-proposed Shortcut block efficiently approximates $\mathbf{f}_t$ , the features from the $l_d$ -th decoding layer of the teacher model. This is achieved by leveraging the $l_e$ -th encoding layer features $\mathbf{a}_t$ along with reference features, $\mathbf{a}_{ref}$ and $\mathbf{f}_{ref}$ , from the previous frame. Here, $l_d$ and $l_e$ correspond to layer indices of the teacher model. Lastly, predicted features $\hat{\mathbf{f}}_t$ are injected into the following layers of the teacher model to synthesize the final output $\hat{\mathbf{O}}_t$ . To avoid error accumulation, we conduct full teacher inference and update the reference features at every max interval $\alpha$ . We provide the detailed inference process of Shortcut-V2V in Algorithm 1. + +The architecture of our model is mainly inspired by Deformable Convolutional Network (DCN) [6, 39], which we explain in the next section. + +# 3.1. Deformable Convolutional Network + +DCN [6, 39] is initially introduced to improve the transformation capability of a convolutional layer in image-based vision tasks, e.g., object detection and semantic segmentation. In the standard convolutional layer, a $3 \times 3$ kernel with dilation 1 samples points over input features using a sampling position $\mathbf{p}_k \in \{(-1, -1), (-1, 0), \dots, (0, 1), (1, 1)\}$ . Given input feature maps $\mathbf{x}$ , DCN predicts additional offsets $\Delta \mathbf{p} \in \mathbb{R}^{2N_{\mathbf{p}} \times H \times W}$ to augment each sampling position along x-axis and y-axis. Here, $N_{\mathbf{p}}$ is the number of sampling positions in a kernel, and $H$ and $W$ are the height and width + +of output feature maps, respectively. For further manipulation of input feature amplitudes over the sampled points, DCNv2 [39] introduces a modulated deformable convolution with $\mathbf{m} \in \mathbb{R}^{N_{\mathbf{P}} \times H \times W}$ consisting of learnable modulation scalars. The deformed output feature maps $\mathbf{x}'$ by DCNv2 are defined as: + +$$ +\mathbf {x} ^ {\prime} = f _ {d c} (\mathbf {w}, \mathbf {x}, \Delta \mathbf {p}, \mathbf {m}), \tag {1} +$$ + +where $f_{dc}$ indicates a deformable convolution. Specifically, a single point $\mathbf{p_o}$ of $\mathbf{x}'$ is obtained as: + +$$ +\mathbf {x} ^ {\prime} \left(\mathbf {p} _ {\mathbf {o}}\right) = \sum_ {k = 1} ^ {N _ {\mathbf {p}}} \mathbf {w} \left(\mathbf {p} _ {k}\right) \cdot \mathbf {x} \left(\mathbf {p} _ {\mathbf {o}} + \mathbf {p} _ {k} + \Delta \mathbf {p} \left(\mathbf {p} _ {k}\right)\right) \cdot \mathbf {m} \left(\mathbf {p} _ {k}\right), \tag {2} +$$ + +where $\mathbf{w}(\mathbf{p}_k),\mathbf{p}(\mathbf{p}_k)$ , and $\mathbf{m}(\mathbf{p}_k)$ are convolutional layer weights, offsets, and modulation scalars between 0 and 1, respectively, for the $k$ -th sampling position. + +Taking advantage of the enhanced transformation capability, we also leverage deformable convolution to align features of adjacent frames only with a few convolution-like operations, instead of using heavy flow estimation networks. + +# 3.2. Shortcut Block + +As described in Fig. 1(b), Shortcut block $S$ estimates the current frame features $\hat{\mathbf{f}}_t$ given $\mathbf{a}_t$ and the reference frame features $\mathbf{f}_{ref}$ and $\mathbf{a}_{ref}$ as inputs: + +$$ +\hat {\mathbf {f}} _ {t} = S \left(\mathbf {f} _ {r e f}, \mathbf {a} _ {r e f}, \mathbf {a} _ {t}\right). \tag {3} +$$ + +Our block effectively obtains rich information from $\mathbf{f}_{ref}$ via coarse-to-fine alignment referring to alignment between $\mathbf{a}_{ref}$ and $\mathbf{a}_t$ . Also, during the fine alignment, our newly-proposed AdaBD simultaneously performs adaptive blending of the aligned $\mathbf{f}_{ref}$ and the current frame feature $\mathbf{a}_t$ . Here, $\mathbf{a}_t$ supports the synthesis of newly-appeared areas in the current frame. + +Coarse-to-Fine Alignment. To handle a wide range of misalignments between the frames, our model aligns $\mathbf{f}_{ref}$ with the current frame in a coarse-to-fine manner. Our global/local alignment module consists of an offset generator to estimate offsets, and deformable convolution layers to deform features based on the predicted offsets. Following TDAN [30], an offset generator estimates sampling offsets given the adjacent frame features. For global alignment, we first downsample the given inputs to enlarge the receptive fields of the corresponding convolutional layers in a lightweight manner. The downsampled $\mathbf{a}_{ref}$ and $\mathbf{a}_t$ are concatenated and fed into a global offset generator to generate global offsets $\Delta \mathbf{p}_g\in \mathbb{R}^{2\times \frac{H}{2}\times \frac{W}{2}}$ . Since we only need to capture coarse movement, $\Delta \mathbf{p}_g$ includes a single offset for each kernel, unlike the original DCN. Each offset is identically applied to all the sampling positions within the + +kernel. Then, the deformed features are upsampled back to the original size to obtain $\mathbf{f}_{ref}^{\prime}$ as follows: + +$$ +\mathbf {f} _ {r e f} ^ {\prime} = \left(f _ {d c} \left(\mathbf {w} _ {g}, \left(\mathbf {f} _ {r e f}\right) ^ {\downarrow 2}, \Delta \mathbf {p} _ {g}, \mathbf {1})\right) ^ {\uparrow 2}, \right. \tag {4} +$$ + +where $\mathbf{w}_g$ denotes weights of global deformable convolution, and $(\cdot)^{\uparrow 2}$ and $(\cdot)^{\downarrow 2}$ refer to upsampling and downsampling by a factor of 2 through bilinear interpolation, respectively. 1 indicates a vector filled with 1 so that no modulation is applied here. While local alignment of the coarsely-aligned feature $\mathbf{f}_{ref}^{\prime}$ follows the process of global alignment, the difference lies in that each sampling point of each kernel has a unique offset, and the alignment operation is conducted in the original resolution. We leverage $\mathbf{a}_{ref}^{\prime}$ and $\mathbf{a}_t$ to estimate local offsets, where $\mathbf{a}_{ref}^{\prime}$ is the aligned $\mathbf{a}_{ref}$ which is downsampled, deformed, and upsampled with the same weights $\mathbf{w}_g$ used to synthesize $\mathbf{f}_{ref}^{\prime}$ . + +We estimate the offsets for the decoding layer features $\mathbf{f}$ using the encoding layer features a under the assumption that $\mathbf{a}_t$ and $\mathbf{f}_t$ have the same structural information. In video-to-video translation, input and output frames share the same underlying structure. Thus, it is natural for the network to learn to maintain the structural information of an input frame throughout the encoding and decoding process. More details are described in our supplementary materials. Adaptive Blending and Deformation. During the local alignment, we also take advantage of the current frame features $\mathbf{a}_t$ from the encoding layer to handle the regions with large motion differences and new objects. To achieve this in a cost-efficient way, we introduce AdaBD, which simultaneously aligns $\mathbf{f}_{ref}^{\prime}$ and blends it with $\mathbf{a}_t$ in an adaptive manner, as illustrated in Fig. 1(b) AdaBD. + +First, our local offset/mask generator predicts a blending mask $\mathbf{m}_b\in \mathbb{R}^{N_{\mathbf{P}}\times H\times W}$ in addition to the local offsets $\Delta \mathbf{p}_l\in \mathbb{R}^{2N_{\mathbf{P}}\times H\times W}$ . A learnable mask $\mathbf{m}_b$ is composed of modulation scalars ranging from 0 to 1, each of which indicates the blending ratio of the current features $\mathbf{a}_t$ to the aligned reference features. While DCNv2 [39] originally introduces the modulation scalars to control feature amplitudes of a single input, we leverage the scalars to adaptively blend the features from two adjacent frames considering their redundant areas. + +We apply deformable convolution by adding local offsets $\Delta \mathbf{p}_l$ to sampling positions of the coarsely-aligned reference features $\mathbf{f}_{ref}^{\prime}$ , while the current frame features $\mathbf{a}_t$ are fed into standard convolutional operations. Concurrently, blending mask $\mathbf{m}_b$ adaptively combines the two feature maps. In detail, an output point $\mathbf{p_o}$ of $\hat{\mathbf{f}}_t$ is calculated as follows: + +$$ +\begin{array}{l} \hat {\mathbf {f}} _ {t} (\mathbf {p} _ {\mathbf {o}}) = \sum_ {k = 1} ^ {N _ {\mathbf {p}}} \mathbf {w} _ {l} (\mathbf {p} _ {k}) \cdot \left\{\mathbf {a} _ {t} (\mathbf {p} _ {\mathbf {o}} + \mathbf {p} _ {k}) \cdot \mathbf {m} _ {b} (\mathbf {p} _ {k}) + \right. \\ \mathbf {f} _ {r e f} ^ {\prime} \left(\mathbf {p} _ {\mathbf {o}} + \mathbf {p} _ {k} + \Delta \mathbf {p} _ {l} (\mathbf {p} _ {k})\right) \cdot (1 - \mathbf {m} _ {b} (\mathbf {p} _ {k})) \}, \tag {5} \\ \end{array} +$$ + +where $\mathbf{w}_l$ indicates weights of local deformable convolution. + +Intuitively, the higher values of $\mathbf{m}_b$ indicate the regions where current frame features are more required. In other words, Eq. 5 can be rewritten as a summation of standard convolution and deformable convolution: + +$$ +\hat {\mathbf {f}} _ {t} = f _ {d c} \left(\mathbf {w} _ {l}, \mathbf {a} _ {t}, \mathbf {0}, \mathbf {m} _ {b}\right) + f _ {d c} \left(\mathbf {w} _ {l}, \mathbf {f} _ {r e f} ^ {\prime}, \Delta \mathbf {p} _ {l}, 1 - \mathbf {m} _ {b}\right), \tag {6} +$$ + +In this equation, the convolutional weights $\mathbf{w}_l$ are shared between $\mathbf{a}_t$ and $\mathbf{f}_{ref}^{\prime}$ . $f_{dc}$ with 0 indicates DCN with zero offsets, illustrating standard convolutional operation. To save computational costs, we decrease the channel dimension of all input features and reconstruct the original channel size before injecting the output features into the remaining layers of the teacher network. + +# 3.3. Training Objectives + +To train Shortcut-V2V, we mainly leverage alignment loss, distillation loss, and GAN losses widely used in image/video translation networks [35, 34, 38]. First, we adopt alignment loss $L_{\text{align}}$ to train the deformation layers in Shortcut-V2V. Since Shortcut-V2V aims to align the reference frame features $\mathbf{f}_{ref}$ with the current frame, $L_{\text{align}}$ computes L1 loss between the aligned feature $\mathbf{f}_{ref}^*$ and the current frame features $\mathbf{f}_t$ extracted from the teacher model. To obtain $\mathbf{f}_{ref}^*$ , we align $\mathbf{f}_{ref}$ in a coarse-to-fine manner without an intervention of $\mathbf{m}_b$ or blending with the current features. The alignment loss $L_{\text{align}}$ is formulated as follows: + +$$ +\mathbf {f} _ {r e f} ^ {*} = f _ {d c} \left(\mathbf {w} _ {l}, \mathbf {f} _ {r e f} ^ {\prime}, \Delta \mathbf {p} _ {l}, 1\right), \tag {7} +$$ + +$$ +L _ {\text {a l i g n}} = \left\| \mathbf {f} _ {t} - \mathbf {f} _ {\text {r e f}} ^ {*} \right\| _ {1}. \tag {8} +$$ + +Additionally, we employ knowledge distillation losses at the feature and output levels. A feature-level distillation loss $L_{feat}$ is applied between the estimated feature $\hat{\mathbf{f}}_t$ and the ground truth feature $\mathbf{f}_t$ , while an output-level distillation loss $L_{out}$ compares the approximated output $\hat{\mathbf{O}}_t$ to the output $\mathbf{O}_t$ generated by the teacher network. The perceptual loss [37] $L_{perc}$ is also incorporated to distill the high-frequency information of the outputs. + +Lastly, we utilize a typical GAN loss $L_{GAN}$ and a temporal GAN loss $L_{T - GAN}$ , following the existing video-based frameworks [34, 3]. Temporal GAN loss $L_{T - GAN}$ encourages both temporal consistency and realisticity of the output frames. For the GAN losses, we consider the outputs of the teacher network as real images. + +The overall objective function $L_{total}$ is as follows: + +$$ +\begin{array}{l} L _ {t o t a l} = \lambda_ {\text {a l i g n}} L _ {\text {a l i g n}} + \lambda_ {\text {f e a t}} L _ {\text {f e a t}} + \lambda_ {\text {o u t}} L _ {\text {o u t}} \\ + \lambda_ {p e r c} L _ {p e r c} + \lambda_ {G A N} L _ {G A N} + \lambda_ {T - G A N} L _ {T - G A N}, \tag {9} \\ \end{array} +$$ + +
ModelDatasetMethodFVDMACs (G)Param. (M)
UnsupV2COriginal1.25384.617.84
CAT1.31022.831.48
OMGD1.19925.851.21
Ours1.18018.190.32
L2VOriginal1.66342.367.84
CAT1.95712.021.71
OMGD2.46712.931.21
Ours1.7549.090.32
vid2vidE2FOriginal0.1862066.69365
CAT0.288380.5524.67
OMGD0.264485.8435.81
Fast-Vid2Vid0.223767.6363.17
Ours0.209359.998.29
L2COriginal0.1461254.17411.34
CAT0.304399.3671.11
OMGD0.346391.5947.83
Fast-Vid2Vid0.310455.2971.8
Ours0.165389.1652.58
+ +Table 1: Quantitative comparison with baselines. + +
DatasetMethodmIoUACMP
V2LOriginal12.216.062.8
CAT5.998.8244.2
OMGD9.7613.058.7
Ours11.515.261.4
L2VOriginal10.015.647.3
CAT4.107.8127.6
OMGD3.426.8729.5
Ours9.2414.943.5
+ +Table 2: Comparison of segmentation score for V2L and FCN-score for L2V on Unsup. + +where $\lambda_{\text{align}}$ , $\lambda_{\text{feat}}$ , $\lambda_{\text{out}}$ , $\lambda_{\text{perc}}$ , $\lambda_{\text{GAN}}$ , and $\lambda_{T - \text{GAN}}$ are hyperparameters to control relative significance among the losses. More details on training objectives are described in the supplementary materials. + +# 4. Experiments + +# 4.1. Experimental Settings + +Models and Datasets. To demonstrate the generality of our model, we conduct experiments on various tasks with widely-used video-to-video translation models, Unsupervised RecycleGAN [32] and vid2vid [34]. + +Unsupervised RecycleGAN (Unsup) is the state-of-art video-to-video translation model among CycleGAN-based [38] ones. We conduct experiments for Unsup on Viper $\rightarrow$ Cityscapes (V2C), which involves translating inter-modality from game driving scenes (Viper [23]) to real-world driving scenes (Cityscapes [5]). The Viper dataset consists of driving scenes collected from the GTA-V game engine, with 56 training videos and 21 test videos. + +![](images/c3dd9ff3419167f8d9f04100039f10cd97a943373d8832faff4151022a229873.jpg) +Figure 2: Qualitative comparison. The upper rows represent the results of Unsup V2C, L2V, and V2L. The bottom rows show the results of vid2vid E2F and L2C. + +Cityscapes is a dataset composed of 2,975 training videos and 500 validation videos. The input images are resized into $256 \times 512$ for the experiments. We also tackle the translation of the videos from the Viper dataset to their corresponding segmentation label maps, $Viper \rightarrow Label (V2L)$ , and vice versa, $Label \rightarrow Viper (L2V)$ . We resize the images and the labels into $256 \times 256$ following the previous work [1, 32]. + +vid2vid is a widely-used pix2pixHD-based video-to-video translation network that serves as a base architecture for various recent video-to-video translation models [33, 18]. Following vid2vid, we evaluate ShortcutV2V on Edge $\rightarrow$ Face (E2F) and Label $\rightarrow$ Cityscapes (L2C). E2F translates edge maps into facial videos from the FaceForensics [24], containing 704 videos for training and 150 videos for validation with various lengths. The images are cropped and resized to $512 \times 512$ for the experiments. The edge maps are extracted using the estimated facial landmarks and Canny edge detector. Also, L2C synthesizes videos of driving scenes from segmentation label maps using the Cityscapes. We generate segmentation maps using pretrained networks following the previous studies [34, 40]. The images and the labels are resized into $256 \times 512$ . + +Evaluation Metrics. Primarily, we adopt the Fréchet video distance (FVD) score [31] to evaluate the performance of Shortcut-V2V quantitatively. The FVD score measures the Fréchet Inception distance between the distribution of video-level features extracted from generated and real videos. The lower the FVD score is, the better visual quality and temporal consistency the generated video frames have. For V2L and L2V, we follow the measure + +ment of the evaluation metrics in the teacher model, Unsup [32]. We measure the segmentation scores, mean intersection over union (mIoU), mean pixel accuracy (MP), and average class accuracy (AC), to validate the performance of V2L. L2V is evaluated using FCN-score [17]. For the evaluation, we first estimate label maps from the generated videos using the FCN model pretrained with the Viper dataset. Then, we measure how accurately the estimated label maps are mapped to the ground truth segmentation labels. Higher FCN-scores refer to better output quality. + +Implementation Details. We attach our Shortcut block to the fixed teacher networks implemented based on the official codes and pretrained by the authors, except for Unsup V2C. The teacher network of Unsup V2C is trained from scratch in the same way the original paper described [32]. + +Also, standard convolutional kernels in the Shortcut block are replaced with HetConv [26] to further enhance computational efficiency without a performance drop. For the convenience of implementation and training stability, we intend a and $\mathbf{f}$ to have the same spatial size. Furthermore, we set the max interval $\alpha$ for each dataset considering the factors that reflect motion differences between frames, such as the frame per second (FPS) of training videos. Lower FPS usually results in larger motion differences between the adjacent frames, requiring a shorter max interval $\alpha$ and vice versa. Specifically, we set $\alpha$ as 3 on V2C, V2L, L2V, and L2C, since the FPS of Viper [23] and Cityscapes [5] is 15 and 17, respectively. Also, $\alpha$ of E2F is set as 6, where the FPS of FaceForensics [24] is 30. Additional details are included in the supplementary materials. + +![](images/e546d61d4dacb5725ce8faca634d6ce788d30492b5f363b1eebf20b79fca9278.jpg) +(a) Unsup V2C + +![](images/fa37ab61bc3da2787fec80a120ee3f0c45ca314a1a9515b0472ed0ad23ea1a8a.jpg) +(b) vid2vid E2F +Figure 3: Performance-efficiency trade-off of the original model, Shortcut-V2V, and the existing compression methods including OMGD [22], CAT [12], and FastVid2Vid [40]. We measure the FVD score and MACs, where the lower FVD score indicates better quality. Red points and stars denote ours with various model sizes. + +# 4.2. Comparison to Baselines + +To demonstrate the effectiveness of our framework, we conduct qualitative and quantitative evaluations compared to the original model and other baselines. Since this is the first work that tackles a generally applicable compression framework for video-to-video translation, we compare Shortcut-V2V to the existing compression methods for image-to-image translation, CAT [12] and OMGD [22], regarding video frames as individual images. In the case of vid2vid, we additionally conduct a comparison to FastVid2Vid [40], which is the compression method designed specifically for vid2vid. For a fair comparison, we compress the student networks of the baselines to have similar or higher MACs compared to our model. + +Qualitative Evaluation. According to Fig. 2, our method presents outputs of comparable visual quality to the original model with much fewer computations. In contrast, CAT on Unsup V2C generates undesirable buildings in the sky, and OMGD on Unsup V2C struggles with noticeable artifacts. For Unsup L2V, CAT and OMGD generate unrealistic textures for the terrain in the middle or the vegetation on the right. Moreover, in V2L, CAT estimates inappropriate labels on the sky and OMGD on the vegetation. + +For vid2vid E2F, CAT shows unwanted artifacts on the mouth, and OMGD presents inconsistent outputs. For vid2vid L2C, the outputs of CAT are blurry, especially for the trees in the background, and OMGD generates artifacts at the bottom of the images. Although Fast-Vid2Vid shows reasonable image quality, the output frames are inaccurately aligned with the input (e.g., a head pose of a person) and suffer from ghost effects (e.g., the black car on the left) due to motion compensation using interpolation. + +Quantitative Evaluation. As shown in Table 1 and Table 2, our framework successfully improves the computational efficiency of the original network without a significant perfor + +
ConfigurationsUnsup V2Cvid2vid E2F
(a) w/o reference features1.2560.398
(b) w/o current features1.2490.218
(c) Single-stage alignment1.1950.213
(d) w/o adaptive blending1.2080.244
(e) Ours1.1800.209
+ +Table 3: An ablation study on Unsup V2C and vid2vid E2F. We measure the FVD scores. + +mance drop. In the case of Unsup, Shortcut-V2V reduces the MACs by $4.7 \times$ and the number of parameters by $24.5 \times$ . In addition, our approach saves vid2vid's MACs by 3.2 and $5.7 \times$ and the number of parameters by 7.8 and $44 \times$ on each task. Fig. 3 visualizes the performance-efficiency trade-off of our framework and other baselines. + +According to Table 1, we also outperform other compression methods for image-to-image translation, CAT [12] and OMGD [22], even with fewer computations. Image-based approaches cannot consider temporal coherence among the frames during the compression, leading to a loss of quality. Meanwhile, we effectively preserve the original performance by exploiting rich information in the previous features during inference. Shortcut-V2V even shows superiority over Fast-Vid2Vid [40], which is a compression method specifically designed for vid2vid. Table 2 also demonstrates that ours on Unsup surpasses the existing compression models by a large margin. + +# 4.3. Ablation Study + +We conduct an ablation study to evaluate the effect of each component of Shortcut-V2V. As described in Table 3, we compare the FVD scores of 5 different configurations, (a) w/o reference features, (b) w/o current features, (c) single-stage alignment, (d) w/o adaptive blending, and (e) ours, on Unsup V2C and vid2vid E2F. + +First, while our model originally exploits both the reference features $\mathbf{f}_{ref}$ and the current frame features $\mathbf{a}_t$ to synthesize $\hat{\mathbf{f}}_t$ , (a) and (b) are designed to leverage only either of them. To be specific, (a) synthesizes $\hat{\mathbf{f}}_t$ by processing standard convolutions on $\mathbf{a}_t$ without integrating the deformed $\mathbf{f}_{ref}$ . Meanwhile, (b) estimates $\hat{\mathbf{f}}_t$ using only $\mathbf{f}_{ref}$ aligned in a coarse-to-fine manner without blending $\mathbf{a}_t$ . The results demonstrate that the absence of either $\mathbf{f}_{ref}$ or $\mathbf{a}_t$ leads to performance degradation. Next, (c) performs a single-stage alignment on $\mathbf{f}_{ref}$ to show the necessity of a coarse-to-fine alignment. Specifically, (c) leverages a single offset/mask generator to predict $\Delta \mathbf{p}$ and $\mathbf{m}_b$ which are then used in AdaBD for blending and deformation. The result indicates that coarse-to-fine alignment effectively assists the estimation of $\hat{\mathbf{f}}_t$ by minimizing misalignment between the features from adjacent frames. This makes Shortcut-V2V suitable for + +![](images/a7aa8d7781e173f499019437ea5ee417891dc2af3c13019a568244b0918fb35f.jpg) +Reference frame + +![](images/75e3ad8902851e6a2f44c7d02a2c9b9080b8e379286eae79b7d2aa9117ede368.jpg) +Current frame + +![](images/65d60aec67bff2e75167dcc2383c061e01dce23370e90af142008e946e409249.jpg) +Figure 4: Visualization of global/local offsets and blending masks. The green points denote output points of the current frame, and the red points around them are each output point's sampling positions modified by global offsets $\Delta \mathbf{p}_g$ and local offsets $\Delta \mathbf{p}_l$ . The values in a blending mask $\mathbf{m}_b$ are averaged by the kernel size for visualization. The brighter area indicates higher mask values. + +![](images/89ad3ad3142d4dd2756198857948ea712a8771cc7dc4c84a1c10220014674896.jpg) + +![](images/98ad9cf051ce8d7a056ac8da6db177267d7df905ed99ff1f41e7a03fbf79bf59.jpg) + +generating videos with diverse motion differences. Lastly, (d) blends $\mathbf{f}_{ref}^{*}$ and $\mathbf{a}_t$ simply by element-wise addition instead of adaptive blending with $\mathbf{m}_b$ in AdaBD. The result demonstrates that $\mathbf{m}_b$ also encourages better estimation of $\hat{\mathbf{f}}_t$ by selectively exploiting features from temporally adjacent features. + +# 4.4. Offset/Mask Visualization + +We provide a qualitative analysis of the generated global/local offsets and blending masks. Fig. 4 visualizes the sampling positions (red points) for each output position $\mathbf{p}_o$ of the current frame (green points), where the predicted offsets $\Delta \mathbf{p}_g$ and $\Delta \mathbf{p}_l$ are added to the original sampling positions. According to the result, global offsets $\Delta \mathbf{p}_g$ effectively reflect the global movement of the objects such as the bridge. Also, the summation of the global and local offsets indicates the sampling points refined by local offsets enabling fine alignment. This shows that the estimated offsets effectively support the utilization of common features in the reference frame. For the blending masks $\mathbf{m}_b$ , Fig. 4 presents that the regions with significant motion differences (e.g., trees) have large mask values compared to regions with little change (e.g., road, sky). That is, our model relies more on the current features rather than the reference features when the deformation is challenging, which aligns with our intention and leads to robust performance. + +# 4.5. Performance-Efficiency Trade-off + +We present a performance-efficiency trade-off of our framework with respect to varying model sizes of Shortcut block and teacher model dependence. + +Shortcut Block Size. We construct our models of different sizes by reducing the number of the output chan + +
DependenceUnsup V2Cvid2vid E2F
FVDMACs (G)Param. (M)FVDMACs (G)Param. (M)
Low1.22113.790.140.277243.972.11
Medium1.18018.190.320.209359.998.29
High1.16627.971.380.193475.9932.98
+ +Table 4: Shortcut-V2V performance with different teacher model dependence. High dependency denotes using more teacher network layers during the inference. + +nel of $1 \times 1$ channel reduction layer by half (Ours-1/2) and a quarter (Ours-1/4). As shown in Fig. 3, despite the performance-efficiency trade-off depending on channel dimension, Ours-1/2 and Ours-1/4 still achieve comparable FVD to the baselines with less computation costs. On a single RTX 3090 GPU, our models with three different sizes on Unsup achieve 99.67, 109.69, and 111.80 FPS with an $11.6 - 21.2\%$ reduction in inference time compared to the original model's speed of 88.10 FPS. Additionally, Ours, Ours-1/2, and Ours-1/4 on vid2vid contribute to a significant improvement in speed from 5.63 FPS to 12.86, 12.93, and 13.16 FPS, resulting in a $56.2 - 57.2\%$ reduction in inference time. Our method is capable of a real-time inference, unlike the previous work [40] which requires future frames for the current frame inference due to motion compensation. More details on channel manipulation are described in the supplementary materials. + +Teacher Model Dependence. Since Shortcut-V2V leverages a subset of the teacher network during inference, the computational costs and memory usage may vary depending on the amount of teacher network we use. In this regard, we present an analysis of the computational efficiency and performance of our model with respect to various levels of teacher model dependence. We categorize the dependence on the teacher model into three levels, low, medium, and high, where high dependence indicates using more teacher network layers. According to Table 4, our model of higher dependence achieves better performance with larger MACs and the number of parameters. The results also demonstrate that Shortcut-V2V can leverage temporal redundancy in features extracted from various layers of the teacher network. Further details for teacher model dependence and experiments for temporal redundancy are included in the supplementary materials. + +# 5. Discussion and Conclusion + +Although Shortcut-V2V significantly improves the test-time efficiency of video-to-video translation, it still poses several limitations. First, a constant max interval may induce unsatisfactory outputs when a degree of temporal redundancy largely varies between frames. Following recent studies [25, 20, 19], applying an adaptive interval based on frame selection algorithms or a learnable policy net- + +work could be promising future research. In addition, we need to manually configure various hyperparameters such as channel dimension, max interval, and teacher model dependence, which can be automated by NAS [13]. Lastly, our method still has the potential to further improvement of computational efficiency by compressing the teacher model using other methods before applying our framework. + +Despite the limitations, Shortcut-V2V presents a significant improvement in test-time efficiency of video-to-video translation networks based on temporal redundancy reduction, while preserving the original model performance. Extensive experiments with widely-used video-to-video translation models successfully demonstrate the general applicability of our framework. To the best of our knowledge, this is the first work for a general model compression in the domain of video-to-video translation. We hope our work facilitates research on video-to-video translation and extends the range of its application. + +Acknowledgments. This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2021-0-02068, Artificial Intelligence Innovation Hub) and the Ministry of Culture, Sports and Tourism and Korea Creative Content Agency (Project Number: R2021040097, Contribution Rate: 50). + +# References + +[1] Aayush Bansal, Shugao Ma, Deva Ramanan, and Yaser Sheikh. Recycle-gan: Unsupervised video retargeting. Proc. of the European Conference on Computer Vision (ECCV), 2018. 1, 2, 6 +[2] Kelvin C. K. Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Understanding deformable alignment in video super-resolution. Proc. the AAAI Conference on Artificial Intelligence (AAAI), 2021. 3 +[3] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston H. Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. Proc. of the IEEE International Conference on Computer Vision (ICCV), 2019. 5 +[4] Yang Chen, Yingwei Pan, Ting Yao, Xinmei Tian, and Tao Mei. Mocycle-gan: Unpaired video-to-video translation. In ACM Multimedia (ACM MM), 2019. 2 +[5] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5, 6 +[6] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. Proc. of the IEEE International Conference on Computer Vision (ICCV), 2017. 3 +[7] Jianing Deng, Li Wang, Shiliang Pu, and Cheng Zhuo. Spatio-temporal deformable convolution for compressed video quality enhancement. Proc. the AAAI Conference on Artificial Intelligence (AAAI), 2020. 3 +[8] Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation, 2019. 3 +[9] Amirhossein Habibian, Haitam Ben Yahia, Davide Abati, Efstratos Gavves, and Fatih Porikli. Delta distillation for efficient video processing, 2022. 3 +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1 +[11] Bangrui Jiang, Zhihuai Xie, Zhen Xia, Songnan Li, and Shan Liu. Erdn: Equivalent receptive field deformable network for video deblurring. In Proc. of the European Conference on Computer Vision (ECCV), 2022. 3 +[12] Qing Jin, Jian Ren, Oliver J. Woodford, Jiazhuo Wang, Geng Yuan, Yanzhi Wang, and Sergey Tulyakov. Teachers do more than teach: Compressing image-to-image models. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3, 7 +[13] Muyang Li, Ji Lin, Yaoyao Ding, Zhijian Liu, Jun-Yan Zhu, and Song Han. GAN compression: Efficient architectures for interactive conditional gans. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 3, 9 +[14] Ji Lin, Chuang Gan, and Song Han. Temporal shift module for efficient video understanding. Proc. of the IEEE International Conference on Computer Vision (ICCV), 2019. 3 + +[15] Jiayi Lin, Yan Huang, and Liang Wang. Fdan: Flow-guided deformable alignment network for video super-resolution. arXiv preprint arXiv:2105.05640, 2021. 3 +[16] Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient semantic video segmentation with per-frame inference. Proc. of the European Conference on Computer Vision (ECCV), 2020. 3 +[17] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 6 +[18] Arun Mallya, Ting-Chun Wang, Karan Sapra, and Ming-Yu Liu. World-consistent video-to-video synthesis. Proc. of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 6 +[19] Yue Meng, Rameswar Panda, Chung-Ching Lin, Prasanna Sattigeri, Leonid Karlinsky, Kate Saenko, Aude Oliva, and Rogério Feris. Adafuse: Adaptive temporal fusion network for efficient action recognition. Proc. the International Conference on Learning Representations (ICLR), 2021. 1, 3, 8 +[20] Bowen Pan, Rameswar Panda, Camilo Fosco, Chung-Ching Lin, Alex Andonian, Yue Meng, Kate Saenko, Aude Oliva, and Rogerio Feris. Va-red²: Video adaptive redundancy reduction. Proc. the International Conference on Learning Representations (ICLR), 2021. 1, 3, 8 +[21] KwanYong Park, Sanghyun Woo, Dahun Kim, Donghyeon Cho, and In So Kweon. Preserving semantic and temporal consistency for unpaired video-to-video translation. ACM Multimedia (ACMM), 2019. 2 +[22] Yuxi Ren, Jie Wu, Xuefeng Xiao, and Jianchao Yang. Online multi-granularity distillation for gan compression. Proc. of the IEEE International Conference on Computer Vision (ICCV), 2021. 1, 3, 7 +[23] Stephan R. Richter, Zeeshan Hayden, and Vladlen Koltun. Playing for benchmarks. In Proc. of the IEEE International Conference on Computer Vision (ICCV), 2017. 5, 6 +[24] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. FaceForensics++: Learning to detect manipulated facial images. In Proc. of the IEEE International Conference on Computer Vision (ICCV), 2019. 6 +[25] Jonghyeon Seon, Jaedong Hwang, Jonghwan Mun, and Bohyung Han. Stop or forward: Dynamic layer skipping for efficient action recognition. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2023. 8 +[26] Pravendra Singh, Vinay Kumar Verma, Piyush Rai, and Vinay P. Namboodiri. Hetconv: Heterogeneous kernel-based convolutions for deep cnns. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6 +[27] Ximeng Sun, Rameswar Panda, Chun-Fu Chen, Aude Oliva, Rogério Feris, and Kate Saenko. Dynamic network quantization for efficient video inference. Proc. of the IEEE International Conference on Computer Vision (ICCV), 2021. 3 +[28] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1 + +[29] Ryan Szeto, Mostafa El-Khamy, Jungwon Lee, and Jason J. Corso. Hypercon: Image-to-video model transfer for video-to-video translation tasks. IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2021. 3 +[30] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. TDAN: temporally deformable alignment network for video super-resolution. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3, 4 +[31] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. In Proc. the International Conference on Learning Representations Workshop (ICLRW), 2019. 6 +[32] Kaihong Wang, Kumar Akash, and Teruhisa Mitsu. Learning temporally and semantically consistent unpaired video-to-video translation through pseudo-supervision from synthetic optical flow. Proc. the AAAI Conference on Artificial Intelligence (AAAI), 2022. 1, 2, 5, 6 +[33] Ting-Chun Wang, Ming-Yu Liu, Andrew Tao, Guilin Liu, Jan Kautz, and Bryan Catanzaro. Few-shot video-to-video synthesis. Proc. the Advances in Neural Information Processing Systems (NeurIPS), 2019. 1, 2, 6 +[34] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Proc. the Advances in Neural Information Processing Systems (NeurIPS), 2018. 1, 2, 3, 5, 6 +[35] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 5 +[36] Xintao Wang, Kelvin C. K. Chan, Ke Yu, Chao Dong, and Chen Change Loy. EDVR: video restoration with enhanced deformable convolutional networks. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2019. 3 +[37] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5 +[38] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proc. of the IEEE International Conference on Computer Vision (ICCV), 2017. 2, 5 +[39] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3, 4 +[40] Long Zhuo, Guangcong Wang, Shikai Li, Wayne Wu, and Ziwei Liu. Fast-vid2vid: Spatial-temporal compression for video-to-video synthesis. Proc. of the European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 6, 7, 8 \ No newline at end of file diff --git a/2023/Shortcut-V2V_ Compression Framework for Video-to-Video Translation Based on Temporal Redundancy Reduction/images.zip b/2023/Shortcut-V2V_ Compression Framework for Video-to-Video Translation Based on Temporal Redundancy Reduction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9854240eb21d85e0b40cf6f17e55ded7d5f8c328 --- /dev/null +++ b/2023/Shortcut-V2V_ Compression Framework for Video-to-Video Translation Based on Temporal Redundancy Reduction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea627ef2db23c2e7f6ff16736735bbcf6d0001232b3a7a9d0605eb8c0eb043cf +size 453306 diff --git a/2023/Shortcut-V2V_ Compression Framework for Video-to-Video Translation Based on Temporal Redundancy Reduction/layout.json b/2023/Shortcut-V2V_ Compression Framework for Video-to-Video Translation Based on Temporal Redundancy Reduction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..75f670287c06c1aa63321e2019a8616fcf12b97b --- /dev/null +++ b/2023/Shortcut-V2V_ Compression Framework for Video-to-Video Translation Based on Temporal Redundancy Reduction/layout.json @@ -0,0 +1,9834 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 102, + 519, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 102, + 519, + 140 + ], + "spans": [ + { + "bbox": [ + 75, + 102, + 519, + 140 + ], + "type": "text", + "content": "Shortcut-V2V: Compression Framework for Video-to-Video Translation based on Temporal Redundancy Reduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 170, + 537, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 170, + 537, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 170, + 537, + 186 + ], + "type": "text", + "content": "Chaeyeon Chung\\*, Yeojeong Park\\*, Seunghwan Choi, Munkhsoyol Ganbat, and Jaegul Choo" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 194, + 457, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 194, + 457, + 236 + ], + "spans": [ + { + "bbox": [ + 140, + 194, + 457, + 236 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 194, + 457, + 236 + ], + "type": "text", + "content": "KAIST AI " + }, + { + "bbox": [ + 140, + 194, + 457, + 236 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 194, + 457, + 236 + ], + "type": "text", + "content": "KT Research & Development Center, KT Corporation {cy-chung, indigopyj, shadow2496, soyol, jchoo}@kaist.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 270, + 192, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 270, + 192, + 282 + ], + "spans": [ + { + "bbox": [ + 143, + 270, + 192, + 282 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 296, + 289, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 296, + 289, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 289, + 583 + ], + "type": "text", + "content": "Video-to-video translation aims to generate video frames of a target domain from an input video. Despite its usefulness, the existing networks require enormous computations, necessitating their model compression for wide use. While there exist compression methods that improve computational efficiency in various image/video tasks, a generally-applicable compression method for video-to-video translation has not been studied much. In response, we present Shortcut-V2V, a general-purpose compression framework for video-to-video translation. Shortcut-V2V avoids full inference for every neighboring video frame by approximating the intermediate features of a current frame from those of the previous frame. Moreover, in our framework, a newly-proposed block called AdaBD adaptively blends and deforms features of neighboring frames, which makes more accurate predictions of the intermediate features possible. We conduct quantitative and qualitative evaluations using well-known video-to-video translation models on various tasks to demonstrate the general applicability of our framework. The results show that Shortcut-V2V achieves comparable performance compared to the original video-to-video translation model while saving " + }, + { + "bbox": [ + 46, + 296, + 289, + 583 + ], + "type": "inline_equation", + "content": "3.2 - 5.7 \\times" + }, + { + "bbox": [ + 46, + 296, + 289, + 583 + ], + "type": "text", + "content": " computational cost and " + }, + { + "bbox": [ + 46, + 296, + 289, + 583 + ], + "type": "inline_equation", + "content": "7.8 - 44 \\times" + }, + { + "bbox": [ + 46, + 296, + 289, + 583 + ], + "type": "text", + "content": " memory at test time. Our code and videos are available at https://shortcut-v2v.github.io/." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 609, + 128, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 128, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 128, + 623 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 631, + 287, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 631, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 46, + 631, + 287, + 691 + ], + "type": "text", + "content": "Video-to-video translation is a task of generating temporally consistent and realistic video frames of a target domain from a given input video. Recent studies on video-to-video translation present promising performance in various domains such as inter-modality translation between labels" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 270, + 545, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 270, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 545, + 295 + ], + "type": "text", + "content": "and videos [34, 33, 18], and intra-modality translation between driving scene videos [32] or face videos [1]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 296, + 546, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 296, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 296, + 546, + 437 + ], + "type": "text", + "content": "Despite enhanced usefulness, video-to-video translation networks usually require substantial computational cost and memory usage, which limits their applicability. For instance, multiply-accumulates (MACs) of a widely-used video translation model, vid2vid [34], is 2066.69G, while the basic convolutional neural networks, ResNet v2 50 [10] and Inception v3 [28], are 4.12G and 6G, respectively. Furthermore, temporally redundant computations for adjacent video frames also harm the cost efficiency of a video-to-video translation network. Performing full inference for every neighboring video frame that contains common visual features inevitably entails redundant operations [20, 19]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 438, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 547, + 653 + ], + "type": "text", + "content": "In this regard, Fast-Vid2Vid [40] proposes a compression framework for vid2vid [34] based on spatial input compression and temporal redundancy reduction. However, it cannot be applied to other video-to-video translation models since it is designed specifically for vid2vid. Moreover, Fast-Vid2Vid does not support real-time inference since it requires future frames to infer a current one. Alternatively, one can apply model compression approaches for image-to-image translation [13, 22, 12] directly to video-to-video translation, considering video frames as separate images. However, these approaches are not designed to consider the correlation among adjacent video frames during the compression. This may result in unrealistic output quality in video-to-video translation, where the inherent temporal coherence of an input video needs to be preserved in the outputs. Also, frame-by-frame inference without temporal redundancy reduction involves unnecessary computations, resulting in computational inefficiency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we propose Shortcut-V2V, a general-purpose framework for improving the computational efficiency of video-to-video translation based on temporal redundancy reduction. Shortcut-V2V allows the original video-to-video translation model to avoid temporally re" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 158, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 158, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 158, + 713 + ], + "type": "text", + "content": "*indicates equal contributions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "type": "text", + "content": "7612" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 267, + 202 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 267, + 202 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 267, + 202 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 267, + 202 + ], + "type": "image", + "image_path": "ccd7fcbcdba55580a15d96a0c2daea7d8ec466b9c324b414c912480b8bce99f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 212, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 45, + 212, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 45, + 212, + 547, + 248 + ], + "type": "text", + "content": "Figure 1: Overview of Shortcut-V2V. (a) is an overall framework of Shorcut-V2V, and (b) shows a detailed architecture of Shortcut block. " + }, + { + "bbox": [ + 45, + 212, + 547, + 248 + ], + "type": "inline_equation", + "content": "\\uparrow 2" + }, + { + "bbox": [ + 45, + 212, + 547, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 212, + 547, + 248 + ], + "type": "inline_equation", + "content": "\\downarrow 2" + }, + { + "bbox": [ + 45, + 212, + 547, + 248 + ], + "type": "text", + "content": " refer to upsampling and downsampling by a factor of 2, respectively. G in Offset G and Offset/Mask G indicates a generator." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 271, + 72, + 545, + 202 + ], + "blocks": [ + { + "bbox": [ + 271, + 72, + 545, + 202 + ], + "lines": [ + { + "bbox": [ + 271, + 72, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 271, + 72, + 545, + 202 + ], + "type": "image", + "image_path": "e709cafb05b43f5baf3babe234028bcd445d0aee9731a366389b0916160169dd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 270, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 437 + ], + "type": "text", + "content": "dundant computations by approximating the decoding layer features of the current frame with largely reduced computations. To enable lightweight estimation, our framework leverages features from the previous frame (i.e., reference features), which have high visual similarity with the current frame. We also exploit current frame features from the encoding layer to handle newly-appeared regions in the current frame. Specifically, we first globally align the previous frame features with the current frame features, and our novel Adaptive Blending and Deformation block (AdaBD) in Shortcut-V2V blends features of neighboring frames while performing detailed deformation. AdaBD adaptively integrates the features regarding their redundancy in a lightweight manner." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 438, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 288, + 689 + ], + "type": "text", + "content": "In this way, our model significantly improves the test-time efficiency of the original network while preserving its original performance. Shortcut-V2V is easily applicable to a pretrained video-to-video translation model to save computational cost and memory usage. Our framework is also suitable for real-time inference since we do not require future frames for the current frame inference. To the best of our knowledge, this is the first attempt at a general-purpose model compression approach for video-to-video translation. We demonstrate the effectiveness of our approach using well-known video-to-video translation models, Unsupervised RecycleGAN [32] (Unsup) and vid2vid [34]. Shortcut-V2V reduces " + }, + { + "bbox": [ + 46, + 438, + 288, + 689 + ], + "type": "inline_equation", + "content": "3.2 - 5.7 \\times" + }, + { + "bbox": [ + 46, + 438, + 288, + 689 + ], + "type": "text", + "content": " computational cost and " + }, + { + "bbox": [ + 46, + 438, + 288, + 689 + ], + "type": "inline_equation", + "content": "7.8 - 44 \\times" + }, + { + "bbox": [ + 46, + 438, + 288, + 689 + ], + "type": "text", + "content": " memory usage while achieving comparable performance to the original model. Since there is no existing general-purpose compression method, we compare our method with Fast-Vid2Vid [40] and the compression methods for image-to-image translation. Our model presents superiority over the existing approaches in both quantitative and qualitative evaluations. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 701, + 287, + 713 + ], + "type": "text", + "content": "- We introduce a novel, general-purpose model com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 324, + 270, + 545, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 270, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 324, + 270, + 545, + 305 + ], + "type": "text", + "content": "pression framework for video-to-video translation, Shortcut-V2V, that enables the original network to avoid temporally redundant computations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 316, + 545, + 408 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 317, + 316, + 545, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 545, + 351 + ], + "type": "text", + "content": "- We present AdaBD that exploits features from neighboring frames via adaptive blending and deformation in a lightweight manner." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 361, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 545, + 408 + ], + "type": "text", + "content": "- Our framework saves up to " + }, + { + "bbox": [ + 317, + 361, + 545, + 408 + ], + "type": "inline_equation", + "content": "5.7 \\times" + }, + { + "bbox": [ + 317, + 361, + 545, + 408 + ], + "type": "text", + "content": " MACs and " + }, + { + "bbox": [ + 317, + 361, + 545, + 408 + ], + "type": "inline_equation", + "content": "44 \\times" + }, + { + "bbox": [ + 317, + 361, + 545, + 408 + ], + "type": "text", + "content": " parameters across various video-to-video translation tasks, achieving comparable performance to the original networks." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 422, + 392, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 422, + 392, + 434 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 392, + 434 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 443, + 455, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 455, + 454 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 455, + 454 + ], + "type": "text", + "content": "2.1. Video-to-Video Translation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 461, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 569 + ], + "type": "text", + "content": "Recent video-to-video translation networks are generally classified into pix2pixHD-based [35] and CycleGAN-based [38] generator. vid2vid [34] proposes a pix2pixHD-based sequential generation framework that synthesizes a current output given the previous outputs as additional guidance. As the following work, few-shot vid2vid [33] achieves few-shot generalization of vid2vid based on attention modules, and world-consistent vid2vid [18] is proposed to improve long-term temporal consistency of vid2vid." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "While pix2pixHD-based models require paired annotated videos, RecycleGAN [1] and MocycleGAN [4] propose CycleGAN-based video translation models. They exploit spatio-temporal consistency losses to generate realistic videos using unpaired datasets. STC-V2V [21] leverages optical flow for semantic/temporal consistency to improve the output quality of the existing models. Unsupervised RecycleGAN [32] achieves state-of-the-art performance among CycleGAN-based frameworks with a pseudosupervision by the synthetic flow. Although the existing video-to-video translation networks achieve decent performance, they commonly demand a non-trivial amount of" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7613" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "computational costs and memory usage. Also, frame-by-frame inference necessarily causes redundant operations due to temporal redundancy among adjacent frames." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 117, + 163, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 117, + 163, + 131 + ], + "spans": [ + { + "bbox": [ + 47, + 117, + 163, + 131 + ], + "type": "text", + "content": "2.2. Model Compression" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 135, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 135, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 135, + 287, + 327 + ], + "type": "text", + "content": "Model compression for video-related tasks has been actively proposed in various domains [19, 20, 16, 9, 27, 8, 14, 29], such as object detection, action recognition, semantic segmentation, and super-resolution. Several studies [19, 9, 20] exploit temporal redundancy among video frames to improve efficiency during training or inference. For instance, Habibian et al. [9] distill only the residual between adjacent frames from a teacher model to a student to speed up the inference. Also, Fast-Vid2Vid [40] firstly proposes a compression framework for video-to-video translation based on spatial and temporal compression. However, Fast-Vid2Vid focuses on vid2vid [34], limiting its application to other video-to-video translation networks. Also, temporal redundancy reduction via motion compensation in Fast-Vid2Vid requires the future frame to infer the current frame, which is not suitable for real-time inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 327, + 287, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 507 + ], + "type": "text", + "content": "Alternatively, model-agnostic compression methods for image-to-image translation can also be applied to video-to-video translation models. The existing approaches for image synthesis mainly tackle channel pruning [13], knowledge distillation [12, 22], NAS [13], etc. For instance, CAT [12] compresses the teacher network with one-step pruning to satisfy the target computation budget, while OMGD [22] conducts a single-stage online distillation in which the teacher generator supports the student generator to be refined progressively. However, image-based compression methods cannot consider temporal coherence among neighboring video frames, which may induce unrealistic results in video translation tasks. Also, performing full model inference for each video frame still poses computational inefficiency due to the temporal redundancy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 515, + 185, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 185, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 185, + 528 + ], + "type": "text", + "content": "2.3. Deformable Convolution" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": "Deformable convolution [6] is originally proposed to enhance the transformation capability of a convolutional layer by adding the estimated offsets to a regular convolutional kernel in vision tasks such as object detection or semantic segmentation. Besides its original application, recent studies [30, 7, 36, 2, 11, 15] demonstrate that deformable convolution is also capable of aligning adjacent video frames. TDAN [30] utilizes deformable convolution to capture implicit motion cues between consecutive frames by dynamically predicted offsets in video super-resolution. In addition, EDVR [36] stacks several deformable convolution blocks to estimate large and complex motions in video restoration. In this paper, we also leverage a deformable convolution to adaptively align adjacent video frames in a lightweight manner." + } + ] + } + ], + "index": 5 + }, + { + "type": "code", + "bbox": [ + 310, + 99, + 545, + 257 + ], + "blocks": [ + { + "bbox": [ + 307, + 83, + 460, + 95 + ], + "lines": [ + { + "bbox": [ + 307, + 83, + 460, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 83, + 460, + 95 + ], + "type": "text", + "content": "Algorithm 1 Shortcut-V2V Inference" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "lines": [ + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "text", + "content": "1: Input: Input video " + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_t\\}_{t=0}^{N_T-1}" + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "text", + "content": " of length " + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "inline_equation", + "content": "N_T" + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "text", + "content": ", teacher model " + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "text", + "content": ", layer index of encoder " + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "inline_equation", + "content": "l_e" + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "inline_equation", + "content": "l_d" + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "text", + "content": ", Shortcut block " + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "text", + "content": ", max interval " + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "text", + "content": " \n2: Output: Output video " + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{O}_t\\}_{t=0}^{N_T-1}" + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "text", + "content": " \n3: for " + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "inline_equation", + "content": "N_T - 1" + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "text", + "content": " do \n4: " + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t = T_{[le]}(\\mathbf{I}_t)" + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "text", + "content": " \n5: if " + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "inline_equation", + "content": "t\\% \\alpha = 0" + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "text", + "content": " then \n6: " + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t = T_{[le + 1;l_d - 1]}(\\mathbf{a}_t)" + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{\\mathrm{ref}}, \\mathbf{f}_{\\mathrm{ref}} \\gets \\mathbf{a}_t, \\mathbf{f}_t \\triangleright" + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "text", + "content": " Update the reference features \n8: else \n9: " + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t = S(\\mathbf{f}_{\\mathrm{ref}}, \\mathbf{a}_{\\mathrm{ref}}, \\mathbf{a}_t)" + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "text", + "content": " \n10: end if \n11: " + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_t = T_{[l_d,]}(\\mathbf{f}_t)" + }, + { + "bbox": [ + 310, + 99, + 545, + 257 + ], + "type": "text", + "content": " \n12: end for" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "code_body" + } + ], + "index": 7, + "sub_type": "algorithm" + }, + { + "bbox": [ + 306, + 290, + 392, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 290, + 392, + 303 + ], + "spans": [ + { + "bbox": [ + 306, + 290, + 392, + 303 + ], + "type": "text", + "content": "3. Shortcut-V2V" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "text", + "content": "In this paper, we propose Shortcut-V2V, a general compression framework to improve the test-time efficiency in video-to-video translation. As illustrated in Fig. 1(a), given " + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_t\\}_{t=0}^{N_T-1}" + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "text", + "content": " as input video frames, we first use full teacher model " + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "text", + "content": " to synthesize the output of the first frame. Then, for the next frames, our newly-proposed Shortcut block efficiently approximates " + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t" + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "text", + "content": ", the features from the " + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "inline_equation", + "content": "l_d" + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "text", + "content": "-th decoding layer of the teacher model. This is achieved by leveraging the " + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "inline_equation", + "content": "l_e" + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "text", + "content": "-th encoding layer features " + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "text", + "content": " along with reference features, " + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{ref}" + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}" + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "text", + "content": ", from the previous frame. Here, " + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "inline_equation", + "content": "l_d" + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "inline_equation", + "content": "l_e" + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "text", + "content": " correspond to layer indices of the teacher model. Lastly, predicted features " + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_t" + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "text", + "content": " are injected into the following layers of the teacher model to synthesize the final output " + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{O}}_t" + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "text", + "content": ". To avoid error accumulation, we conduct full teacher inference and update the reference features at every max interval " + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 311, + 545, + 514 + ], + "type": "text", + "content": ". We provide the detailed inference process of Shortcut-V2V in Algorithm 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 515, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 545, + 552 + ], + "type": "text", + "content": "The architecture of our model is mainly inspired by Deformable Convolutional Network (DCN) [6, 39], which we explain in the next section." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 562, + 495, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 562, + 495, + 574 + ], + "spans": [ + { + "bbox": [ + 305, + 562, + 495, + 574 + ], + "type": "text", + "content": "3.1. Deformable Convolutional Network" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "DCN [6, 39] is initially introduced to improve the transformation capability of a convolutional layer in image-based vision tasks, e.g., object detection and semantic segmentation. In the standard convolutional layer, a " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " kernel with dilation 1 samples points over input features using a sampling position " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_k \\in \\{(-1, -1), (-1, 0), \\dots, (0, 1), (1, 1)\\}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ". Given input feature maps " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", DCN predicts additional offsets " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{p} \\in \\mathbb{R}^{2N_{\\mathbf{p}} \\times H \\times W}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " to augment each sampling position along x-axis and y-axis. Here, " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "N_{\\mathbf{p}}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " is the number of sampling positions in a kernel, and " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " are the height and width" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7614" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "of output feature maps, respectively. For further manipulation of input feature amplitudes over the sampled points, DCNv2 [39] introduces a modulated deformable convolution with " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{m} \\in \\mathbb{R}^{N_{\\mathbf{P}} \\times H \\times W}" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " consisting of learnable modulation scalars. The deformed output feature maps " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " by DCNv2 are defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 152, + 287, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 152, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 116, + 152, + 287, + 167 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {\\prime} = f _ {d c} (\\mathbf {w}, \\mathbf {x}, \\Delta \\mathbf {p}, \\mathbf {m}), \\tag {1}", + "image_path": "28fae1a8bb4c1b3a697498d6dff5b2fc01a0b2e9227c22f76973637c31c564ac.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 175, + 286, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 175, + 286, + 199 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 286, + 199 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 175, + 286, + 199 + ], + "type": "inline_equation", + "content": "f_{dc}" + }, + { + "bbox": [ + 47, + 175, + 286, + 199 + ], + "type": "text", + "content": " indicates a deformable convolution. Specifically, a single point " + }, + { + "bbox": [ + 47, + 175, + 286, + 199 + ], + "type": "inline_equation", + "content": "\\mathbf{p_o}" + }, + { + "bbox": [ + 47, + 175, + 286, + 199 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 175, + 286, + 199 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 47, + 175, + 286, + 199 + ], + "type": "text", + "content": " is obtained as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 207, + 287, + 241 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 207, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 51, + 207, + 287, + 241 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {\\prime} \\left(\\mathbf {p} _ {\\mathbf {o}}\\right) = \\sum_ {k = 1} ^ {N _ {\\mathbf {p}}} \\mathbf {w} \\left(\\mathbf {p} _ {k}\\right) \\cdot \\mathbf {x} \\left(\\mathbf {p} _ {\\mathbf {o}} + \\mathbf {p} _ {k} + \\Delta \\mathbf {p} \\left(\\mathbf {p} _ {k}\\right)\\right) \\cdot \\mathbf {m} \\left(\\mathbf {p} _ {k}\\right), \\tag {2}", + "image_path": "0cf9119157b73a05c027d0817ca837169870de54d6f47f65639771800c293f46.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 249, + 287, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 249, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 249, + 287, + 285 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 249, + 287, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{w}(\\mathbf{p}_k),\\mathbf{p}(\\mathbf{p}_k)" + }, + { + "bbox": [ + 47, + 249, + 287, + 285 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 47, + 249, + 287, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{m}(\\mathbf{p}_k)" + }, + { + "bbox": [ + 47, + 249, + 287, + 285 + ], + "type": "text", + "content": " are convolutional layer weights, offsets, and modulation scalars between 0 and 1, respectively, for the " + }, + { + "bbox": [ + 47, + 249, + 287, + 285 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 249, + 287, + 285 + ], + "type": "text", + "content": " -th sampling position." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 285, + 287, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 344 + ], + "type": "text", + "content": "Taking advantage of the enhanced transformation capability, we also leverage deformable convolution to align features of adjacent frames only with a few convolution-like operations, instead of using heavy flow estimation networks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 352, + 140, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 352, + 140, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 352, + 140, + 363 + ], + "type": "text", + "content": "3.2. Shortcut Block" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 370, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 370, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 287, + 407 + ], + "type": "text", + "content": "As described in Fig. 1(b), Shortcut block " + }, + { + "bbox": [ + 47, + 370, + 287, + 407 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 370, + 287, + 407 + ], + "type": "text", + "content": " estimates the current frame features " + }, + { + "bbox": [ + 47, + 370, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_t" + }, + { + "bbox": [ + 47, + 370, + 287, + 407 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 47, + 370, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 47, + 370, + 287, + 407 + ], + "type": "text", + "content": " and the reference frame features " + }, + { + "bbox": [ + 47, + 370, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}" + }, + { + "bbox": [ + 47, + 370, + 287, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 370, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{ref}" + }, + { + "bbox": [ + 47, + 370, + 287, + 407 + ], + "type": "text", + "content": " as inputs:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 119, + 415, + 287, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 415, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 119, + 415, + 287, + 430 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {f}} _ {t} = S \\left(\\mathbf {f} _ {r e f}, \\mathbf {a} _ {r e f}, \\mathbf {a} _ {t}\\right). \\tag {3}", + "image_path": "d225acd3e873f3278f38f12d47f7f73aeb74e2251f3219b4ff36370d857ecf7a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 437, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 287, + 521 + ], + "type": "text", + "content": "Our block effectively obtains rich information from " + }, + { + "bbox": [ + 47, + 437, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}" + }, + { + "bbox": [ + 47, + 437, + 287, + 521 + ], + "type": "text", + "content": " via coarse-to-fine alignment referring to alignment between " + }, + { + "bbox": [ + 47, + 437, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{ref}" + }, + { + "bbox": [ + 47, + 437, + 287, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 437, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 47, + 437, + 287, + 521 + ], + "type": "text", + "content": ". Also, during the fine alignment, our newly-proposed AdaBD simultaneously performs adaptive blending of the aligned " + }, + { + "bbox": [ + 47, + 437, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}" + }, + { + "bbox": [ + 47, + 437, + 287, + 521 + ], + "type": "text", + "content": " and the current frame feature " + }, + { + "bbox": [ + 47, + 437, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 47, + 437, + 287, + 521 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 47, + 437, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 47, + 437, + 287, + 521 + ], + "type": "text", + "content": " supports the synthesis of newly-appeared areas in the current frame." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": "Coarse-to-Fine Alignment. To handle a wide range of misalignments between the frames, our model aligns " + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}" + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": " with the current frame in a coarse-to-fine manner. Our global/local alignment module consists of an offset generator to estimate offsets, and deformable convolution layers to deform features based on the predicted offsets. Following TDAN [30], an offset generator estimates sampling offsets given the adjacent frame features. For global alignment, we first downsample the given inputs to enlarge the receptive fields of the corresponding convolutional layers in a lightweight manner. The downsampled " + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{ref}" + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": " are concatenated and fed into a global offset generator to generate global offsets " + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{p}_g\\in \\mathbb{R}^{2\\times \\frac{H}{2}\\times \\frac{W}{2}}" + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": ". Since we only need to capture coarse movement, " + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{p}_g" + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": " includes a single offset for each kernel, unlike the original DCN. Each offset is identically applied to all the sampling positions within the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 98 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 98 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 98 + ], + "type": "text", + "content": "kernel. Then, the deformed features are upsampled back to the original size to obtain " + }, + { + "bbox": [ + 305, + 72, + 545, + 98 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}^{\\prime}" + }, + { + "bbox": [ + 305, + 72, + 545, + 98 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 347, + 106, + 545, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 106, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 347, + 106, + 545, + 121 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {r e f} ^ {\\prime} = \\left(f _ {d c} \\left(\\mathbf {w} _ {g}, \\left(\\mathbf {f} _ {r e f}\\right) ^ {\\downarrow 2}, \\Delta \\mathbf {p} _ {g}, \\mathbf {1})\\right) ^ {\\uparrow 2}, \\right. \\tag {4}", + "image_path": "1e3d52141391f5f94125063a8b7752465e587f1d5b7e3c1dbe64b6fe5d9b1e52.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_g" + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "text", + "content": " denotes weights of global deformable convolution, and " + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "inline_equation", + "content": "(\\cdot)^{\\uparrow 2}" + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "inline_equation", + "content": "(\\cdot)^{\\downarrow 2}" + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "text", + "content": " refer to upsampling and downsampling by a factor of 2 through bilinear interpolation, respectively. 1 indicates a vector filled with 1 so that no modulation is applied here. While local alignment of the coarsely-aligned feature " + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}^{\\prime}" + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "text", + "content": " follows the process of global alignment, the difference lies in that each sampling point of each kernel has a unique offset, and the alignment operation is conducted in the original resolution. We leverage " + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{ref}^{\\prime}" + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "text", + "content": " to estimate local offsets, where " + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{ref}^{\\prime}" + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "text", + "content": " is the aligned " + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{ref}" + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "text", + "content": " which is downsampled, deformed, and upsampled with the same weights " + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_g" + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "text", + "content": " used to synthesize " + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}^{\\prime}" + }, + { + "bbox": [ + 304, + 130, + 545, + 274 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 274, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 545, + 453 + ], + "type": "text", + "content": "We estimate the offsets for the decoding layer features " + }, + { + "bbox": [ + 304, + 274, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 304, + 274, + 545, + 453 + ], + "type": "text", + "content": " using the encoding layer features a under the assumption that " + }, + { + "bbox": [ + 304, + 274, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 304, + 274, + 545, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 274, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t" + }, + { + "bbox": [ + 304, + 274, + 545, + 453 + ], + "type": "text", + "content": " have the same structural information. In video-to-video translation, input and output frames share the same underlying structure. Thus, it is natural for the network to learn to maintain the structural information of an input frame throughout the encoding and decoding process. More details are described in our supplementary materials. Adaptive Blending and Deformation. During the local alignment, we also take advantage of the current frame features " + }, + { + "bbox": [ + 304, + 274, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 304, + 274, + 545, + 453 + ], + "type": "text", + "content": " from the encoding layer to handle the regions with large motion differences and new objects. To achieve this in a cost-efficient way, we introduce AdaBD, which simultaneously aligns " + }, + { + "bbox": [ + 304, + 274, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}^{\\prime}" + }, + { + "bbox": [ + 304, + 274, + 545, + 453 + ], + "type": "text", + "content": " and blends it with " + }, + { + "bbox": [ + 304, + 274, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 304, + 274, + 545, + 453 + ], + "type": "text", + "content": " in an adaptive manner, as illustrated in Fig. 1(b) AdaBD." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 453, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 571 + ], + "type": "text", + "content": "First, our local offset/mask generator predicts a blending mask " + }, + { + "bbox": [ + 304, + 453, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_b\\in \\mathbb{R}^{N_{\\mathbf{P}}\\times H\\times W}" + }, + { + "bbox": [ + 304, + 453, + 545, + 571 + ], + "type": "text", + "content": " in addition to the local offsets " + }, + { + "bbox": [ + 304, + 453, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{p}_l\\in \\mathbb{R}^{2N_{\\mathbf{P}}\\times H\\times W}" + }, + { + "bbox": [ + 304, + 453, + 545, + 571 + ], + "type": "text", + "content": ". A learnable mask " + }, + { + "bbox": [ + 304, + 453, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_b" + }, + { + "bbox": [ + 304, + 453, + 545, + 571 + ], + "type": "text", + "content": " is composed of modulation scalars ranging from 0 to 1, each of which indicates the blending ratio of the current features " + }, + { + "bbox": [ + 304, + 453, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 304, + 453, + 545, + 571 + ], + "type": "text", + "content": " to the aligned reference features. While DCNv2 [39] originally introduces the modulation scalars to control feature amplitudes of a single input, we leverage the scalars to adaptively blend the features from two adjacent frames considering their redundant areas." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 572, + 545, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 572, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 572, + 545, + 644 + ], + "type": "text", + "content": "We apply deformable convolution by adding local offsets " + }, + { + "bbox": [ + 304, + 572, + 545, + 644 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{p}_l" + }, + { + "bbox": [ + 304, + 572, + 545, + 644 + ], + "type": "text", + "content": " to sampling positions of the coarsely-aligned reference features " + }, + { + "bbox": [ + 304, + 572, + 545, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}^{\\prime}" + }, + { + "bbox": [ + 304, + 572, + 545, + 644 + ], + "type": "text", + "content": ", while the current frame features " + }, + { + "bbox": [ + 304, + 572, + 545, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 304, + 572, + 545, + 644 + ], + "type": "text", + "content": " are fed into standard convolutional operations. Concurrently, blending mask " + }, + { + "bbox": [ + 304, + 572, + 545, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_b" + }, + { + "bbox": [ + 304, + 572, + 545, + 644 + ], + "type": "text", + "content": " adaptively combines the two feature maps. In detail, an output point " + }, + { + "bbox": [ + 304, + 572, + 545, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{p_o}" + }, + { + "bbox": [ + 304, + 572, + 545, + 644 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 572, + 545, + 644 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_t" + }, + { + "bbox": [ + 304, + 572, + 545, + 644 + ], + "type": "text", + "content": " is calculated as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\mathbf {f}} _ {t} (\\mathbf {p} _ {\\mathbf {o}}) = \\sum_ {k = 1} ^ {N _ {\\mathbf {p}}} \\mathbf {w} _ {l} (\\mathbf {p} _ {k}) \\cdot \\left\\{\\mathbf {a} _ {t} (\\mathbf {p} _ {\\mathbf {o}} + \\mathbf {p} _ {k}) \\cdot \\mathbf {m} _ {b} (\\mathbf {p} _ {k}) + \\right. \\\\ \\mathbf {f} _ {r e f} ^ {\\prime} \\left(\\mathbf {p} _ {\\mathbf {o}} + \\mathbf {p} _ {k} + \\Delta \\mathbf {p} _ {l} (\\mathbf {p} _ {k})\\right) \\cdot (1 - \\mathbf {m} _ {b} (\\mathbf {p} _ {k})) \\}, \\tag {5} \\\\ \\end{array}", + "image_path": "f61824940f57697a4c1f7f5fda03969b2b74c898b2960e65900a1723d1384c54.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7615" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_l" + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": " indicates weights of local deformable convolution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 287, + 144 + ], + "type": "text", + "content": "Intuitively, the higher values of " + }, + { + "bbox": [ + 46, + 96, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_b" + }, + { + "bbox": [ + 46, + 96, + 287, + 144 + ], + "type": "text", + "content": " indicate the regions where current frame features are more required. In other words, Eq. 5 can be rewritten as a summation of standard convolution and deformable convolution:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 153, + 287, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 153, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 51, + 153, + 287, + 178 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {f}} _ {t} = f _ {d c} \\left(\\mathbf {w} _ {l}, \\mathbf {a} _ {t}, \\mathbf {0}, \\mathbf {m} _ {b}\\right) + f _ {d c} \\left(\\mathbf {w} _ {l}, \\mathbf {f} _ {r e f} ^ {\\prime}, \\Delta \\mathbf {p} _ {l}, 1 - \\mathbf {m} _ {b}\\right), \\tag {6}", + "image_path": "61065cd14b7d0a4b93e44c3146c135f9a83de87c2c3eece2b86cf4020a82ad88.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 179, + 288, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 179, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 288, + 262 + ], + "type": "text", + "content": "In this equation, the convolutional weights " + }, + { + "bbox": [ + 46, + 179, + 288, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_l" + }, + { + "bbox": [ + 46, + 179, + 288, + 262 + ], + "type": "text", + "content": " are shared between " + }, + { + "bbox": [ + 46, + 179, + 288, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 46, + 179, + 288, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 179, + 288, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}^{\\prime}" + }, + { + "bbox": [ + 46, + 179, + 288, + 262 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 179, + 288, + 262 + ], + "type": "inline_equation", + "content": "f_{dc}" + }, + { + "bbox": [ + 46, + 179, + 288, + 262 + ], + "type": "text", + "content": " with 0 indicates DCN with zero offsets, illustrating standard convolutional operation. To save computational costs, we decrease the channel dimension of all input features and reconstruct the original channel size before injecting the output features into the remaining layers of the teacher network." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 270, + 162, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 162, + 284 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 162, + 284 + ], + "type": "text", + "content": "3.3. Training Objectives" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "type": "text", + "content": "To train Shortcut-V2V, we mainly leverage alignment loss, distillation loss, and GAN losses widely used in image/video translation networks [35, 34, 38]. First, we adopt alignment loss " + }, + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "type": "inline_equation", + "content": "L_{\\text{align}}" + }, + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "type": "text", + "content": " to train the deformation layers in Shortcut-V2V. Since Shortcut-V2V aims to align the reference frame features " + }, + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}" + }, + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "type": "text", + "content": " with the current frame, " + }, + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "type": "inline_equation", + "content": "L_{\\text{align}}" + }, + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "type": "text", + "content": " computes L1 loss between the aligned feature " + }, + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}^*" + }, + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "type": "text", + "content": " and the current frame features " + }, + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t" + }, + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "type": "text", + "content": " extracted from the teacher model. To obtain " + }, + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}^*" + }, + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "type": "text", + "content": ", we align " + }, + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}" + }, + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "type": "text", + "content": " in a coarse-to-fine manner without an intervention of " + }, + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_b" + }, + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "type": "text", + "content": " or blending with the current features. The alignment loss " + }, + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "type": "inline_equation", + "content": "L_{\\text{align}}" + }, + { + "bbox": [ + 46, + 289, + 287, + 432 + ], + "type": "text", + "content": " is formulated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 434, + 287, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 434, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 287, + 449 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {r e f} ^ {*} = f _ {d c} \\left(\\mathbf {w} _ {l}, \\mathbf {f} _ {r e f} ^ {\\prime}, \\Delta \\mathbf {p} _ {l}, 1\\right), \\tag {7}", + "image_path": "4ef31397fbf4162b8486f1571f710aa19942f819c6799248c785273db1b66d10.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 118, + 464, + 287, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 464, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 118, + 464, + 287, + 480 + ], + "type": "interline_equation", + "content": "L _ {\\text {a l i g n}} = \\left\\| \\mathbf {f} _ {t} - \\mathbf {f} _ {\\text {r e f}} ^ {*} \\right\\| _ {1}. \\tag {8}", + "image_path": "50ae9095c6f1afa9dae3abde2e3b4747ce98a5ff3251819326c07ea51ced3311.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 485, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 581 + ], + "type": "text", + "content": "Additionally, we employ knowledge distillation losses at the feature and output levels. A feature-level distillation loss " + }, + { + "bbox": [ + 46, + 485, + 287, + 581 + ], + "type": "inline_equation", + "content": "L_{feat}" + }, + { + "bbox": [ + 46, + 485, + 287, + 581 + ], + "type": "text", + "content": " is applied between the estimated feature " + }, + { + "bbox": [ + 46, + 485, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_t" + }, + { + "bbox": [ + 46, + 485, + 287, + 581 + ], + "type": "text", + "content": " and the ground truth feature " + }, + { + "bbox": [ + 46, + 485, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t" + }, + { + "bbox": [ + 46, + 485, + 287, + 581 + ], + "type": "text", + "content": ", while an output-level distillation loss " + }, + { + "bbox": [ + 46, + 485, + 287, + 581 + ], + "type": "inline_equation", + "content": "L_{out}" + }, + { + "bbox": [ + 46, + 485, + 287, + 581 + ], + "type": "text", + "content": " compares the approximated output " + }, + { + "bbox": [ + 46, + 485, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{O}}_t" + }, + { + "bbox": [ + 46, + 485, + 287, + 581 + ], + "type": "text", + "content": " to the output " + }, + { + "bbox": [ + 46, + 485, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_t" + }, + { + "bbox": [ + 46, + 485, + 287, + 581 + ], + "type": "text", + "content": " generated by the teacher network. The perceptual loss [37] " + }, + { + "bbox": [ + 46, + 485, + 287, + 581 + ], + "type": "inline_equation", + "content": "L_{perc}" + }, + { + "bbox": [ + 46, + 485, + 287, + 581 + ], + "type": "text", + "content": " is also incorporated to distill the high-frequency information of the outputs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 581, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 652 + ], + "type": "text", + "content": "Lastly, we utilize a typical GAN loss " + }, + { + "bbox": [ + 46, + 581, + 287, + 652 + ], + "type": "inline_equation", + "content": "L_{GAN}" + }, + { + "bbox": [ + 46, + 581, + 287, + 652 + ], + "type": "text", + "content": " and a temporal GAN loss " + }, + { + "bbox": [ + 46, + 581, + 287, + 652 + ], + "type": "inline_equation", + "content": "L_{T - GAN}" + }, + { + "bbox": [ + 46, + 581, + 287, + 652 + ], + "type": "text", + "content": ", following the existing video-based frameworks [34, 3]. Temporal GAN loss " + }, + { + "bbox": [ + 46, + 581, + 287, + 652 + ], + "type": "inline_equation", + "content": "L_{T - GAN}" + }, + { + "bbox": [ + 46, + 581, + 287, + 652 + ], + "type": "text", + "content": " encourages both temporal consistency and realisticity of the output frames. For the GAN losses, we consider the outputs of the teacher network as real images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 653, + 263, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 653, + 263, + 665 + ], + "spans": [ + { + "bbox": [ + 59, + 653, + 263, + 665 + ], + "type": "text", + "content": "The overall objective function " + }, + { + "bbox": [ + 59, + 653, + 263, + 665 + ], + "type": "inline_equation", + "content": "L_{total}" + }, + { + "bbox": [ + 59, + 653, + 263, + 665 + ], + "type": "text", + "content": " is as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 674, + 296, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 674, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 674, + 296, + 713 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {t o t a l} = \\lambda_ {\\text {a l i g n}} L _ {\\text {a l i g n}} + \\lambda_ {\\text {f e a t}} L _ {\\text {f e a t}} + \\lambda_ {\\text {o u t}} L _ {\\text {o u t}} \\\\ + \\lambda_ {p e r c} L _ {p e r c} + \\lambda_ {G A N} L _ {G A N} + \\lambda_ {T - G A N} L _ {T - G A N}, \\tag {9} \\\\ \\end{array}", + "image_path": "ad5cbbe898e0239cd0d692fadf5ab658ce251b96639913b12b120f99cc042fa7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 69, + 545, + 276 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 545, + 276 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 545, + 276 + ], + "type": "table", + "html": "
ModelDatasetMethodFVDMACs (G)Param. (M)
UnsupV2COriginal1.25384.617.84
CAT1.31022.831.48
OMGD1.19925.851.21
Ours1.18018.190.32
L2VOriginal1.66342.367.84
CAT1.95712.021.71
OMGD2.46712.931.21
Ours1.7549.090.32
vid2vidE2FOriginal0.1862066.69365
CAT0.288380.5524.67
OMGD0.264485.8435.81
Fast-Vid2Vid0.223767.6363.17
Ours0.209359.998.29
L2COriginal0.1461254.17411.34
CAT0.304399.3671.11
OMGD0.346391.5947.83
Fast-Vid2Vid0.310455.2971.8
Ours0.165389.1652.58
", + "image_path": "8232a1b57b3336dff02584fe1e3648acb3033021ad182806c823e586d6f63650.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 318, + 307, + 535, + 417 + ], + "blocks": [ + { + "bbox": [ + 326, + 284, + 524, + 296 + ], + "lines": [ + { + "bbox": [ + 326, + 284, + 524, + 296 + ], + "spans": [ + { + "bbox": [ + 326, + 284, + 524, + 296 + ], + "type": "text", + "content": "Table 1: Quantitative comparison with baselines." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 307, + 535, + 417 + ], + "lines": [ + { + "bbox": [ + 318, + 307, + 535, + 417 + ], + "spans": [ + { + "bbox": [ + 318, + 307, + 535, + 417 + ], + "type": "table", + "html": "
DatasetMethodmIoUACMP
V2LOriginal12.216.062.8
CAT5.998.8244.2
OMGD9.7613.058.7
Ours11.515.261.4
L2VOriginal10.015.647.3
CAT4.107.8127.6
OMGD3.426.8729.5
Ours9.2414.943.5
", + "image_path": "f3540ad71f0463c600d298ca225537ede0017e507b7c24350e1029a38d83756b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 426, + 545, + 449 + ], + "lines": [ + { + "bbox": [ + 305, + 426, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 545, + 449 + ], + "type": "text", + "content": "Table 2: Comparison of segmentation score for V2L and FCN-score for L2V on Unsup." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 472, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 472, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 545, + 520 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 472, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{align}}" + }, + { + "bbox": [ + 305, + 472, + 545, + 520 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 472, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{feat}}" + }, + { + "bbox": [ + 305, + 472, + 545, + 520 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 472, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{out}}" + }, + { + "bbox": [ + 305, + 472, + 545, + 520 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 472, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{perc}}" + }, + { + "bbox": [ + 305, + 472, + 545, + 520 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 472, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{GAN}}" + }, + { + "bbox": [ + 305, + 472, + 545, + 520 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 472, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\lambda_{T - \\text{GAN}}" + }, + { + "bbox": [ + 305, + 472, + 545, + 520 + ], + "type": "text", + "content": " are hyperparameters to control relative significance among the losses. More details on training objectives are described in the supplementary materials." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 531, + 388, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 531, + 388, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 531, + 388, + 544 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 551, + 432, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 432, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 432, + 563 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 569, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 617 + ], + "type": "text", + "content": "Models and Datasets. To demonstrate the generality of our model, we conduct experiments on various tasks with widely-used video-to-video translation models, Unsupervised RecycleGAN [32] and vid2vid [34]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Unsupervised RecycleGAN (Unsup) is the state-of-art video-to-video translation model among CycleGAN-based [38] ones. We conduct experiments for Unsup on Viper " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " Cityscapes (V2C), which involves translating inter-modality from game driving scenes (Viper [23]) to real-world driving scenes (Cityscapes [5]). The Viper dataset consists of driving scenes collected from the GTA-V game engine, with 56 training videos and 21 test videos." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7616" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 73, + 539, + 297 + ], + "blocks": [ + { + "bbox": [ + 53, + 73, + 539, + 297 + ], + "lines": [ + { + "bbox": [ + 53, + 73, + 539, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 539, + 297 + ], + "type": "image", + "image_path": "c3dd9ff3419167f8d9f04100039f10cd97a943373d8832faff4151022a229873.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 308, + 546, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 308, + 546, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 546, + 331 + ], + "type": "text", + "content": "Figure 2: Qualitative comparison. The upper rows represent the results of Unsup V2C, L2V, and V2L. The bottom rows show the results of vid2vid E2F and L2C." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 353, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 287, + 437 + ], + "type": "text", + "content": "Cityscapes is a dataset composed of 2,975 training videos and 500 validation videos. The input images are resized into " + }, + { + "bbox": [ + 46, + 353, + 287, + 437 + ], + "type": "inline_equation", + "content": "256 \\times 512" + }, + { + "bbox": [ + 46, + 353, + 287, + 437 + ], + "type": "text", + "content": " for the experiments. We also tackle the translation of the videos from the Viper dataset to their corresponding segmentation label maps, " + }, + { + "bbox": [ + 46, + 353, + 287, + 437 + ], + "type": "inline_equation", + "content": "Viper \\rightarrow Label (V2L)" + }, + { + "bbox": [ + 46, + 353, + 287, + 437 + ], + "type": "text", + "content": ", and vice versa, " + }, + { + "bbox": [ + 46, + 353, + 287, + 437 + ], + "type": "inline_equation", + "content": "Label \\rightarrow Viper (L2V)" + }, + { + "bbox": [ + 46, + 353, + 287, + 437 + ], + "type": "text", + "content": ". We resize the images and the labels into " + }, + { + "bbox": [ + 46, + 353, + 287, + 437 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 353, + 287, + 437 + ], + "type": "text", + "content": " following the previous work [1, 32]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 437, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 616 + ], + "type": "text", + "content": "vid2vid is a widely-used pix2pixHD-based video-to-video translation network that serves as a base architecture for various recent video-to-video translation models [33, 18]. Following vid2vid, we evaluate ShortcutV2V on Edge " + }, + { + "bbox": [ + 46, + 437, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 437, + 287, + 616 + ], + "type": "text", + "content": " Face (E2F) and Label " + }, + { + "bbox": [ + 46, + 437, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 437, + 287, + 616 + ], + "type": "text", + "content": " Cityscapes (L2C). E2F translates edge maps into facial videos from the FaceForensics [24], containing 704 videos for training and 150 videos for validation with various lengths. The images are cropped and resized to " + }, + { + "bbox": [ + 46, + 437, + 287, + 616 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 437, + 287, + 616 + ], + "type": "text", + "content": " for the experiments. The edge maps are extracted using the estimated facial landmarks and Canny edge detector. Also, L2C synthesizes videos of driving scenes from segmentation label maps using the Cityscapes. We generate segmentation maps using pretrained networks following the previous studies [34, 40]. The images and the labels are resized into " + }, + { + "bbox": [ + 46, + 437, + 287, + 616 + ], + "type": "inline_equation", + "content": "256 \\times 512" + }, + { + "bbox": [ + 46, + 437, + 287, + 616 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. Primarily, we adopt the Fréchet video distance (FVD) score [31] to evaluate the performance of Shortcut-V2V quantitatively. The FVD score measures the Fréchet Inception distance between the distribution of video-level features extracted from generated and real videos. The lower the FVD score is, the better visual quality and temporal consistency the generated video frames have. For V2L and L2V, we follow the measure" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 353, + 546, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 546, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 546, + 473 + ], + "type": "text", + "content": "ment of the evaluation metrics in the teacher model, Unsup [32]. We measure the segmentation scores, mean intersection over union (mIoU), mean pixel accuracy (MP), and average class accuracy (AC), to validate the performance of V2L. L2V is evaluated using FCN-score [17]. For the evaluation, we first estimate label maps from the generated videos using the FCN model pretrained with the Viper dataset. Then, we measure how accurately the estimated label maps are mapped to the ground truth segmentation labels. Higher FCN-scores refer to better output quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 473, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 533 + ], + "type": "text", + "content": "Implementation Details. We attach our Shortcut block to the fixed teacher networks implemented based on the official codes and pretrained by the authors, except for Unsup V2C. The teacher network of Unsup V2C is trained from scratch in the same way the original paper described [32]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "Also, standard convolutional kernels in the Shortcut block are replaced with HetConv [26] to further enhance computational efficiency without a performance drop. For the convenience of implementation and training stability, we intend a and " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " to have the same spatial size. Furthermore, we set the max interval " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " for each dataset considering the factors that reflect motion differences between frames, such as the frame per second (FPS) of training videos. Lower FPS usually results in larger motion differences between the adjacent frames, requiring a shorter max interval " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " and vice versa. Specifically, we set " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " as 3 on V2C, V2L, L2V, and L2C, since the FPS of Viper [23] and Cityscapes [5] is 15 and 17, respectively. Also, " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " of E2F is set as 6, where the FPS of FaceForensics [24] is 30. Additional details are included in the supplementary materials." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7617" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 167, + 159 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 167, + 159 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 167, + 159 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 167, + 159 + ], + "type": "image", + "image_path": "e546d61d4dacb5725ce8faca634d6ce788d30492b5f363b1eebf20b79fca9278.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 160, + 141, + 171 + ], + "lines": [ + { + "bbox": [ + 89, + 160, + 141, + 171 + ], + "spans": [ + { + "bbox": [ + 89, + 160, + 141, + 171 + ], + "type": "text", + "content": "(a) Unsup V2C" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 168, + 71, + 286, + 159 + ], + "blocks": [ + { + "bbox": [ + 168, + 71, + 286, + 159 + ], + "lines": [ + { + "bbox": [ + 168, + 71, + 286, + 159 + ], + "spans": [ + { + "bbox": [ + 168, + 71, + 286, + 159 + ], + "type": "image", + "image_path": "fa37ab61bc3da2787fec80a120ee3f0c45ca314a1a9515b0472ed0ad23ea1a8a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 160, + 263, + 170 + ], + "lines": [ + { + "bbox": [ + 209, + 160, + 263, + 170 + ], + "spans": [ + { + "bbox": [ + 209, + 160, + 263, + 170 + ], + "type": "text", + "content": "(b) vid2vid E2F" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 184, + 287, + 256 + ], + "lines": [ + { + "bbox": [ + 46, + 184, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 46, + 184, + 287, + 256 + ], + "type": "text", + "content": "Figure 3: Performance-efficiency trade-off of the original model, Shortcut-V2V, and the existing compression methods including OMGD [22], CAT [12], and FastVid2Vid [40]. We measure the FVD score and MACs, where the lower FVD score indicates better quality. Red points and stars denote ours with various model sizes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 281, + 185, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 185, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 185, + 293 + ], + "type": "text", + "content": "4.2. Comparison to Baselines" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 300, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 456 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our framework, we conduct qualitative and quantitative evaluations compared to the original model and other baselines. Since this is the first work that tackles a generally applicable compression framework for video-to-video translation, we compare Shortcut-V2V to the existing compression methods for image-to-image translation, CAT [12] and OMGD [22], regarding video frames as individual images. In the case of vid2vid, we additionally conduct a comparison to FastVid2Vid [40], which is the compression method designed specifically for vid2vid. For a fair comparison, we compress the student networks of the baselines to have similar or higher MACs compared to our model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 458, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 567 + ], + "type": "text", + "content": "Qualitative Evaluation. According to Fig. 2, our method presents outputs of comparable visual quality to the original model with much fewer computations. In contrast, CAT on Unsup V2C generates undesirable buildings in the sky, and OMGD on Unsup V2C struggles with noticeable artifacts. For Unsup L2V, CAT and OMGD generate unrealistic textures for the terrain in the middle or the vegetation on the right. Moreover, in V2L, CAT estimates inappropriate labels on the sky and OMGD on the vegetation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 567, + 286, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 286, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 286, + 675 + ], + "type": "text", + "content": "For vid2vid E2F, CAT shows unwanted artifacts on the mouth, and OMGD presents inconsistent outputs. For vid2vid L2C, the outputs of CAT are blurry, especially for the trees in the background, and OMGD generates artifacts at the bottom of the images. Although Fast-Vid2Vid shows reasonable image quality, the output frames are inaccurately aligned with the input (e.g., a head pose of a person) and suffer from ghost effects (e.g., the black car on the left) due to motion compensation using interpolation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "Quantitative Evaluation. As shown in Table 1 and Table 2, our framework successfully improves the computational efficiency of the original network without a significant perfor" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 545, + 148 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 545, + 148 + ], + "type": "table", + "html": "
ConfigurationsUnsup V2Cvid2vid E2F
(a) w/o reference features1.2560.398
(b) w/o current features1.2490.218
(c) Single-stage alignment1.1950.213
(d) w/o adaptive blending1.2080.244
(e) Ours1.1800.209
", + "image_path": "abab91fb97af0fddb89999d21828f7017aefc092d608bfa2d3bb628be4ba83f9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 156, + 544, + 179 + ], + "lines": [ + { + "bbox": [ + 306, + 156, + 544, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 544, + 179 + ], + "type": "text", + "content": "Table 3: An ablation study on Unsup V2C and vid2vid E2F. We measure the FVD scores." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 203, + 545, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 203, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 304, + 203, + 545, + 274 + ], + "type": "text", + "content": "mance drop. In the case of Unsup, Shortcut-V2V reduces the MACs by " + }, + { + "bbox": [ + 304, + 203, + 545, + 274 + ], + "type": "inline_equation", + "content": "4.7 \\times" + }, + { + "bbox": [ + 304, + 203, + 545, + 274 + ], + "type": "text", + "content": " and the number of parameters by " + }, + { + "bbox": [ + 304, + 203, + 545, + 274 + ], + "type": "inline_equation", + "content": "24.5 \\times" + }, + { + "bbox": [ + 304, + 203, + 545, + 274 + ], + "type": "text", + "content": ". In addition, our approach saves vid2vid's MACs by 3.2 and " + }, + { + "bbox": [ + 304, + 203, + 545, + 274 + ], + "type": "inline_equation", + "content": "5.7 \\times" + }, + { + "bbox": [ + 304, + 203, + 545, + 274 + ], + "type": "text", + "content": " and the number of parameters by 7.8 and " + }, + { + "bbox": [ + 304, + 203, + 545, + 274 + ], + "type": "inline_equation", + "content": "44 \\times" + }, + { + "bbox": [ + 304, + 203, + 545, + 274 + ], + "type": "text", + "content": " on each task. Fig. 3 visualizes the performance-efficiency trade-off of our framework and other baselines." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 276, + 546, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 546, + 420 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 546, + 420 + ], + "type": "text", + "content": "According to Table 1, we also outperform other compression methods for image-to-image translation, CAT [12] and OMGD [22], even with fewer computations. Image-based approaches cannot consider temporal coherence among the frames during the compression, leading to a loss of quality. Meanwhile, we effectively preserve the original performance by exploiting rich information in the previous features during inference. Shortcut-V2V even shows superiority over Fast-Vid2Vid [40], which is a compression method specifically designed for vid2vid. Table 2 also demonstrates that ours on Unsup surpasses the existing compression models by a large margin." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 429, + 399, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 429, + 399, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 429, + 399, + 442 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 449, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 519 + ], + "type": "text", + "content": "We conduct an ablation study to evaluate the effect of each component of Shortcut-V2V. As described in Table 3, we compare the FVD scores of 5 different configurations, (a) w/o reference features, (b) w/o current features, (c) single-stage alignment, (d) w/o adaptive blending, and (e) ours, on Unsup V2C and vid2vid E2F." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "content": "First, while our model originally exploits both the reference features " + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}" + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "content": " and the current frame features " + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "content": " to synthesize " + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_t" + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "content": ", (a) and (b) are designed to leverage only either of them. To be specific, (a) synthesizes " + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_t" + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "content": " by processing standard convolutions on " + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "content": " without integrating the deformed " + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}" + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "content": ". Meanwhile, (b) estimates " + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_t" + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "content": " using only " + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}" + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "content": " aligned in a coarse-to-fine manner without blending " + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "content": ". The results demonstrate that the absence of either " + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}" + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "content": " leads to performance degradation. Next, (c) performs a single-stage alignment on " + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}" + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "content": " to show the necessity of a coarse-to-fine alignment. Specifically, (c) leverages a single offset/mask generator to predict " + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{p}" + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_b" + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "content": " which are then used in AdaBD for blending and deformation. The result indicates that coarse-to-fine alignment effectively assists the estimation of " + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_t" + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "content": " by minimizing misalignment between the features from adjacent frames. This makes Shortcut-V2V suitable for" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7618" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 86, + 147, + 127 + ], + "blocks": [ + { + "bbox": [ + 75, + 76, + 138, + 85 + ], + "lines": [ + { + "bbox": [ + 75, + 76, + 138, + 85 + ], + "spans": [ + { + "bbox": [ + 75, + 76, + 138, + 85 + ], + "type": "text", + "content": "Reference frame" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 86, + 147, + 127 + ], + "lines": [ + { + "bbox": [ + 66, + 86, + 147, + 127 + ], + "spans": [ + { + "bbox": [ + 66, + 86, + 147, + 127 + ], + "type": "image", + "image_path": "a7aa8d7781e173f499019437ea5ee417891dc2af3c13019a568244b0918fb35f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 188, + 87, + 268, + 128 + ], + "blocks": [ + { + "bbox": [ + 200, + 76, + 255, + 85 + ], + "lines": [ + { + "bbox": [ + 200, + 76, + 255, + 85 + ], + "spans": [ + { + "bbox": [ + 200, + 76, + 255, + 85 + ], + "type": "text", + "content": "Current frame" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 188, + 87, + 268, + 128 + ], + "lines": [ + { + "bbox": [ + 188, + 87, + 268, + 128 + ], + "spans": [ + { + "bbox": [ + 188, + 87, + 268, + 128 + ], + "type": "image", + "image_path": "75e3ad8902851e6a2f44c7d02a2c9b9080b8e379286eae79b7d2aa9117ede368.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 136, + 165, + 202 + ], + "blocks": [ + { + "bbox": [ + 48, + 136, + 165, + 202 + ], + "lines": [ + { + "bbox": [ + 48, + 136, + 165, + 202 + ], + "spans": [ + { + "bbox": [ + 48, + 136, + 165, + 202 + ], + "type": "image", + "image_path": "65d60aec67bff2e75167dcc2383c061e01dce23370e90af142008e946e409249.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 287, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 287, + 297 + ], + "type": "text", + "content": "Figure 4: Visualization of global/local offsets and blending masks. The green points denote output points of the current frame, and the red points around them are each output point's sampling positions modified by global offsets " + }, + { + "bbox": [ + 46, + 213, + 287, + 297 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{p}_g" + }, + { + "bbox": [ + 46, + 213, + 287, + 297 + ], + "type": "text", + "content": " and local offsets " + }, + { + "bbox": [ + 46, + 213, + 287, + 297 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{p}_l" + }, + { + "bbox": [ + 46, + 213, + 287, + 297 + ], + "type": "text", + "content": ". The values in a blending mask " + }, + { + "bbox": [ + 46, + 213, + 287, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_b" + }, + { + "bbox": [ + 46, + 213, + 287, + 297 + ], + "type": "text", + "content": " are averaged by the kernel size for visualization. The brighter area indicates higher mask values." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 170, + 137, + 228, + 201 + ], + "blocks": [ + { + "bbox": [ + 170, + 137, + 228, + 201 + ], + "lines": [ + { + "bbox": [ + 170, + 137, + 228, + 201 + ], + "spans": [ + { + "bbox": [ + 170, + 137, + 228, + 201 + ], + "type": "image", + "image_path": "89ad3ad3142d4dd2756198857948ea712a8771cc7dc4c84a1c10220014674896.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 228, + 137, + 287, + 201 + ], + "blocks": [ + { + "bbox": [ + 228, + 137, + 287, + 201 + ], + "lines": [ + { + "bbox": [ + 228, + 137, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 228, + 137, + 287, + 201 + ], + "type": "image", + "image_path": "98ad9cf051ce8d7a056ac8da6db177267d7df905ed99ff1f41e7a03fbf79bf59.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 316, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 389 + ], + "type": "text", + "content": "generating videos with diverse motion differences. Lastly, (d) blends " + }, + { + "bbox": [ + 46, + 316, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ref}^{*}" + }, + { + "bbox": [ + 46, + 316, + 287, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 316, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 46, + 316, + 287, + 389 + ], + "type": "text", + "content": " simply by element-wise addition instead of adaptive blending with " + }, + { + "bbox": [ + 46, + 316, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_b" + }, + { + "bbox": [ + 46, + 316, + 287, + 389 + ], + "type": "text", + "content": " in AdaBD. The result demonstrates that " + }, + { + "bbox": [ + 46, + 316, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_b" + }, + { + "bbox": [ + 46, + 316, + 287, + 389 + ], + "type": "text", + "content": " also encourages better estimation of " + }, + { + "bbox": [ + 46, + 316, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_t" + }, + { + "bbox": [ + 46, + 316, + 287, + 389 + ], + "type": "text", + "content": " by selectively exploiting features from temporally adjacent features." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 395, + 192, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 395, + 192, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 395, + 192, + 406 + ], + "type": "text", + "content": "4.4. Offset/Mask Visualization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 413, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 628 + ], + "type": "text", + "content": "We provide a qualitative analysis of the generated global/local offsets and blending masks. Fig. 4 visualizes the sampling positions (red points) for each output position " + }, + { + "bbox": [ + 46, + 413, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_o" + }, + { + "bbox": [ + 46, + 413, + 287, + 628 + ], + "type": "text", + "content": " of the current frame (green points), where the predicted offsets " + }, + { + "bbox": [ + 46, + 413, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{p}_g" + }, + { + "bbox": [ + 46, + 413, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 413, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{p}_l" + }, + { + "bbox": [ + 46, + 413, + 287, + 628 + ], + "type": "text", + "content": " are added to the original sampling positions. According to the result, global offsets " + }, + { + "bbox": [ + 46, + 413, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{p}_g" + }, + { + "bbox": [ + 46, + 413, + 287, + 628 + ], + "type": "text", + "content": " effectively reflect the global movement of the objects such as the bridge. Also, the summation of the global and local offsets indicates the sampling points refined by local offsets enabling fine alignment. This shows that the estimated offsets effectively support the utilization of common features in the reference frame. For the blending masks " + }, + { + "bbox": [ + 46, + 413, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_b" + }, + { + "bbox": [ + 46, + 413, + 287, + 628 + ], + "type": "text", + "content": ", Fig. 4 presents that the regions with significant motion differences (e.g., trees) have large mask values compared to regions with little change (e.g., road, sky). That is, our model relies more on the current features rather than the reference features when the deformation is challenging, which aligns with our intention and leads to robust performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 635, + 228, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 228, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 228, + 647 + ], + "type": "text", + "content": "4.5. Performance-Efficiency Trade-off" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "text", + "content": "We present a performance-efficiency trade-off of our framework with respect to varying model sizes of Shortcut block and teacher model dependence." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Shortcut Block Size. We construct our models of different sizes by reducing the number of the output chan" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 545, + 117 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 545, + 117 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 545, + 117 + ], + "type": "table", + "html": "
DependenceUnsup V2Cvid2vid E2F
FVDMACs (G)Param. (M)FVDMACs (G)Param. (M)
Low1.22113.790.140.277243.972.11
Medium1.18018.190.320.209359.998.29
High1.16627.971.380.193475.9932.98
", + "image_path": "a302dc8e92e275a0e806e8f74848f9802cf63fadb3e5affbd6079277761f2a86.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 125, + 545, + 161 + ], + "lines": [ + { + "bbox": [ + 305, + 125, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 305, + 125, + 545, + 161 + ], + "type": "text", + "content": "Table 4: Shortcut-V2V performance with different teacher model dependence. High dependency denotes using more teacher network layers during the inference." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 185, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 545, + 388 + ], + "type": "text", + "content": "nel of " + }, + { + "bbox": [ + 304, + 185, + 545, + 388 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 185, + 545, + 388 + ], + "type": "text", + "content": " channel reduction layer by half (Ours-1/2) and a quarter (Ours-1/4). As shown in Fig. 3, despite the performance-efficiency trade-off depending on channel dimension, Ours-1/2 and Ours-1/4 still achieve comparable FVD to the baselines with less computation costs. On a single RTX 3090 GPU, our models with three different sizes on Unsup achieve 99.67, 109.69, and 111.80 FPS with an " + }, + { + "bbox": [ + 304, + 185, + 545, + 388 + ], + "type": "inline_equation", + "content": "11.6 - 21.2\\%" + }, + { + "bbox": [ + 304, + 185, + 545, + 388 + ], + "type": "text", + "content": " reduction in inference time compared to the original model's speed of 88.10 FPS. Additionally, Ours, Ours-1/2, and Ours-1/4 on vid2vid contribute to a significant improvement in speed from 5.63 FPS to 12.86, 12.93, and 13.16 FPS, resulting in a " + }, + { + "bbox": [ + 304, + 185, + 545, + 388 + ], + "type": "inline_equation", + "content": "56.2 - 57.2\\%" + }, + { + "bbox": [ + 304, + 185, + 545, + 388 + ], + "type": "text", + "content": " reduction in inference time. Our method is capable of a real-time inference, unlike the previous work [40] which requires future frames for the current frame inference due to motion compensation. More details on channel manipulation are described in the supplementary materials." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 389, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 594 + ], + "type": "text", + "content": "Teacher Model Dependence. Since Shortcut-V2V leverages a subset of the teacher network during inference, the computational costs and memory usage may vary depending on the amount of teacher network we use. In this regard, we present an analysis of the computational efficiency and performance of our model with respect to various levels of teacher model dependence. We categorize the dependence on the teacher model into three levels, low, medium, and high, where high dependence indicates using more teacher network layers. According to Table 4, our model of higher dependence achieves better performance with larger MACs and the number of parameters. The results also demonstrate that Shortcut-V2V can leverage temporal redundancy in features extracted from various layers of the teacher network. Further details for teacher model dependence and experiments for temporal redundancy are included in the supplementary materials." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 608, + 457, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 608, + 457, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 608, + 457, + 620 + ], + "type": "text", + "content": "5. Discussion and Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Although Shortcut-V2V significantly improves the test-time efficiency of video-to-video translation, it still poses several limitations. First, a constant max interval may induce unsatisfactory outputs when a degree of temporal redundancy largely varies between frames. Following recent studies [25, 20, 19], applying an adaptive interval based on frame selection algorithms or a learnable policy net-" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7619" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 156 + ], + "type": "text", + "content": "work could be promising future research. In addition, we need to manually configure various hyperparameters such as channel dimension, max interval, and teacher model dependence, which can be automated by NAS [13]. Lastly, our method still has the potential to further improvement of computational efficiency by compressing the teacher model using other methods before applying our framework." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 157, + 286, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 157, + 286, + 287 + ], + "spans": [ + { + "bbox": [ + 49, + 157, + 286, + 287 + ], + "type": "text", + "content": "Despite the limitations, Shortcut-V2V presents a significant improvement in test-time efficiency of video-to-video translation networks based on temporal redundancy reduction, while preserving the original model performance. Extensive experiments with widely-used video-to-video translation models successfully demonstrate the general applicability of our framework. To the best of our knowledge, this is the first work for a general model compression in the domain of video-to-video translation. We hope our work facilitates research on video-to-video translation and extends the range of its application." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 293, + 286, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 293, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 49, + 293, + 286, + 376 + ], + "type": "text", + "content": "Acknowledgments. This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2021-0-02068, Artificial Intelligence Innovation Hub) and the Ministry of Culture, Sports and Tourism and Korea Creative Content Agency (Project Number: R2021040097, Contribution Rate: 50)." + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "type": "text", + "content": "7620" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Aayush Bansal, Shugao Ma, Deva Ramanan, and Yaser Sheikh. Recycle-gan: Unsupervised video retargeting. Proc. of the European Conference on Computer Vision (ECCV), 2018. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 287, + 178 + ], + "type": "text", + "content": "[2] Kelvin C. K. Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Understanding deformable alignment in video super-resolution. Proc. the AAAI Conference on Artificial Intelligence (AAAI), 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 223 + ], + "type": "text", + "content": "[3] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston H. Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. Proc. of the IEEE International Conference on Computer Vision (ICCV), 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 257 + ], + "type": "text", + "content": "[4] Yang Chen, Yingwei Pan, Ting Yao, Xinmei Tian, and Tao Mei. Mocycle-gan: Unpaired video-to-video translation. In ACM Multimedia (ACM MM), 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 323 + ], + "type": "text", + "content": "[5] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 367 + ], + "type": "text", + "content": "[6] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. Proc. of the IEEE International Conference on Computer Vision (ICCV), 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "text", + "content": "[7] Jianing Deng, Li Wang, Shiliang Pu, and Cheng Zhuo. Spatio-temporal deformable convolution for compressed video quality enhancement. Proc. the AAAI Conference on Artificial Intelligence (AAAI), 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 287, + 445 + ], + "type": "text", + "content": "[8] Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 479 + ], + "type": "text", + "content": "[9] Amirhossein Habibian, Haitam Ben Yahia, Davide Abati, Efstratos Gavves, and Fatih Porikli. Delta distillation for efficient video processing, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 480, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 287, + 523 + ], + "type": "text", + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "text", + "content": "[11] Bangrui Jiang, Zhihuai Xie, Zhen Xia, Songnan Li, and Shan Liu. Erdn: Equivalent receptive field deformable network for video deblurring. In Proc. of the European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "type": "text", + "content": "[12] Qing Jin, Jian Ren, Oliver J. Woodford, Jiazhuo Wang, Geng Yuan, Yanzhi Wang, and Sergey Tulyakov. Teachers do more than teach: Compressing image-to-image models. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "text", + "content": "[13] Muyang Li, Ji Lin, Yaoyao Ding, Zhijian Liu, Jun-Yan Zhu, and Song Han. GAN compression: Efficient architectures for interactive conditional gans. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 3, 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[14] Ji Lin, Chuang Gan, and Song Han. Temporal shift module for efficient video understanding. Proc. of the IEEE International Conference on Computer Vision (ICCV), 2019. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[15] Jiayi Lin, Yan Huang, and Liang Wang. Fdan: Flow-guided deformable alignment network for video super-resolution. arXiv preprint arXiv:2105.05640, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "text", + "content": "[16] Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient semantic video segmentation with per-frame inference. Proc. of the European Conference on Computer Vision (ECCV), 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "text", + "content": "[17] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 195, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 237 + ], + "type": "text", + "content": "[18] Arun Mallya, Ting-Chun Wang, Karan Sapra, and Ming-Yu Liu. World-consistent video-to-video synthesis. Proc. of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "type": "text", + "content": "[19] Yue Meng, Rameswar Panda, Chung-Ching Lin, Prasanna Sattigeri, Leonid Karlinsky, Kate Saenko, Aude Oliva, and Rogério Feris. Adafuse: Adaptive temporal fusion network for efficient action recognition. Proc. the International Conference on Learning Representations (ICLR), 2021. 1, 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 294, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 348 + ], + "type": "text", + "content": "[20] Bowen Pan, Rameswar Panda, Camilo Fosco, Chung-Ching Lin, Alex Andonian, Yue Meng, Kate Saenko, Aude Oliva, and Rogerio Feris. Va-red²: Video adaptive redundancy reduction. Proc. the International Conference on Learning Representations (ICLR), 2021. 1, 3, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "text", + "content": "[21] KwanYong Park, Sanghyun Woo, Dahun Kim, Donghyeon Cho, and In So Kweon. Preserving semantic and temporal consistency for unpaired video-to-video translation. ACM Multimedia (ACMM), 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 394, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 437 + ], + "type": "text", + "content": "[22] Yuxi Ren, Jie Wu, Xuefeng Xiao, and Jianchao Yang. Online multi-granularity distillation for gan compression. Proc. of the IEEE International Conference on Computer Vision (ICCV), 2021. 1, 3, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 438, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 438, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 438, + 545, + 470 + ], + "type": "text", + "content": "[23] Stephan R. Richter, Zeeshan Hayden, and Vladlen Koltun. Playing for benchmarks. In Proc. of the IEEE International Conference on Computer Vision (ICCV), 2017. 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 472, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 472, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 472, + 545, + 525 + ], + "type": "text", + "content": "[24] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. FaceForensics++: Learning to detect manipulated facial images. In Proc. of the IEEE International Conference on Computer Vision (ICCV), 2019. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 526, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 545, + 569 + ], + "type": "text", + "content": "[25] Jonghyeon Seon, Jaedong Hwang, Jonghwan Mun, and Bohyung Han. Stop or forward: Dynamic layer skipping for efficient action recognition. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2023. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 571, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 571, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 571, + 545, + 613 + ], + "type": "text", + "content": "[26] Pravendra Singh, Vinay Kumar Verma, Piyush Rai, and Vinay P. Namboodiri. Hetconv: Heterogeneous kernel-based convolutions for deep cnns. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 615, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 615, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 615, + 545, + 658 + ], + "type": "text", + "content": "[27] Ximeng Sun, Rameswar Panda, Chun-Fu Chen, Aude Oliva, Rogério Feris, and Kate Saenko. Dynamic network quantization for efficient video inference. Proc. of the IEEE International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[28] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7621" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 676 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[29] Ryan Szeto, Mostafa El-Khamy, Jungwon Lee, and Jason J. Corso. Hypercon: Image-to-video model transfer for video-to-video translation tasks. IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "text", + "content": "[30] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. TDAN: temporally deformable alignment network for video super-resolution. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "text", + "content": "[31] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. In Proc. the International Conference on Learning Representations Workshop (ICLRW), 2019. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 288, + 274 + ], + "type": "text", + "content": "[32] Kaihong Wang, Kumar Akash, and Teruhisa Mitsu. Learning temporally and semantically consistent unpaired video-to-video translation through pseudo-supervision from synthetic optical flow. Proc. the AAAI Conference on Artificial Intelligence (AAAI), 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 288, + 319 + ], + "type": "text", + "content": "[33] Ting-Chun Wang, Ming-Yu Liu, Andrew Tao, Guilin Liu, Jan Kautz, and Bryan Catanzaro. Few-shot video-to-video synthesis. Proc. the Advances in Neural Information Processing Systems (NeurIPS), 2019. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 288, + 363 + ], + "type": "text", + "content": "[34] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Proc. the Advances in Neural Information Processing Systems (NeurIPS), 2018. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 288, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 288, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 288, + 419 + ], + "type": "text", + "content": "[35] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 420, + 288, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 420, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 48, + 420, + 288, + 474 + ], + "type": "text", + "content": "[36] Xintao Wang, Kelvin C. K. Chan, Ke Yu, Chao Dong, and Chen Change Loy. EDVR: video restoration with enhanced deformable convolutional networks. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 475, + 288, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 288, + 529 + ], + "type": "text", + "content": "[37] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 531, + 288, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 531, + 288, + 585 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 288, + 585 + ], + "type": "text", + "content": "[38] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proc. of the IEEE International Conference on Computer Vision (ICCV), 2017. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 587, + 288, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 587, + 288, + 631 + ], + "spans": [ + { + "bbox": [ + 48, + 587, + 288, + 631 + ], + "type": "text", + "content": "[39] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 632, + 288, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 632, + 288, + 676 + ], + "spans": [ + { + "bbox": [ + 48, + 632, + 288, + 676 + ], + "type": "text", + "content": "[40] Long Zhuo, Guangcong Wang, Shikai Li, Wayne Wu, and Ziwei Liu. Fast-vid2vid: Spatial-temporal compression for video-to-video synthesis. Proc. of the European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7622" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning/34917a9e-765e-47a5-bd3d-8044ab9de911_content_list.json b/2023/Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning/34917a9e-765e-47a5-bd3d-8044ab9de911_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ea624bd884d00a55d912cbd77c70868e6125b74a --- /dev/null +++ b/2023/Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning/34917a9e-765e-47a5-bd3d-8044ab9de911_content_list.json @@ -0,0 +1,1658 @@ +[ + { + "type": "text", + "text": "Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning", + "text_level": 1, + "bbox": [ + 99, + 130, + 870, + 155 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lihe Yang $^{1}$ Zhen Zhao $^{4}$ Lei Qi $^{5}$ Yu Qiao $^{3}$ Yinghuan Shi $^{2*}$ Hengshuang Zhao $^{1*}$", + "bbox": [ + 138, + 178, + 836, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The University of Hong Kong 2Nanjing University", + "bbox": [ + 272, + 200, + 696, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Shanghai AI Laboratory 4The University of Sydney 5Southeast University", + "bbox": [ + 174, + 220, + 795, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/LiheYoung/ShrinkMatch", + "bbox": [ + 305, + 239, + 661, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 290, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semi-supervised learning is attracting blooming attention, due to its success in combining unlabeled data. To mitigate potentially incorrect pseudo labels, recent frameworks mostly set a fixed confidence threshold to discard uncertain samples. This practice ensures high-quality pseudo labels, but incurs a relatively low utilization of the whole unlabeled set. In this work, our key insight is that these uncertain samples can be turned into certain ones, as long as the confusion classes for the top-1 class are detected and removed. Invoked by this, we propose a novel method dubbed ShrinkMatch to learn uncertain samples. For each uncertain sample, it adaptively seeks a shrunk class space, which merely contains the original top-1 class, as well as remaining less likely classes. Since the confusion ones are removed in this space, the re-calculated top-1 confidence can satisfy the pre-defined threshold. We then impose a consistency regularization between a pair of strongly and weakly augmented samples in the shrunk space to strive for discriminative representations. Furthermore, considering the varied reliability among uncertain samples and the gradually improved model during training, we correspondingly design two reweighting principles for our uncertain loss. Our method exhibits impressive performance on widely adopted benchmarks.", + "bbox": [ + 75, + 321, + 473, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 699, + 209, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the last decade, our computer vision community has witnessed inspiring progress, thanks to large-scale datasets [21, 10]. Nevertheless, it is laborious and costly to annotate massive images, hindering the progress to benefit a broader range of real-world scenarios. Inspired by this, semi-supervised learning (SSL) was proposed to utilize the unlabeled data under the assistance of limited labeled data.", + "bbox": [ + 75, + 724, + 470, + 829 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The frameworks in SSL are typically based on the strategy of pseudo labeling. Briefly, the model acquires knowledge from the labeled data, and then assigns predictions on the", + "bbox": [ + 75, + 830, + 470, + 877 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ae66fb03b8076858a055ecce8796fd1d19d44bae77fc18af83490942b6da113e.jpg", + "image_caption": [ + "Figure 1: Illustration of our motivation. Due to confusion classes for the top-1 class, the certainty fails to reach the pre-defined threshold (gray dotted line). FixMatch discards such uncertain samples. Our method, however, detects and removes confusion classes to enhance certainty, then enjoying full and safe utilization of all unlabeled images." + ], + "image_footnote": [], + "bbox": [ + 540, + 287, + 854, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "unlabeled data. The two sources of data are finally combined to train a better model. During this process, it is obvious that predictions on unlabeled data are not reliable. If the model is iteratively trained with incorrect pseudo labels, it will suffer the confirmation bias issue [1]. To address this dilemma, recent works [28] simply set a fixed confidence threshold to discard potentially unreliable samples. This simple strategy effectively retains high-quality pseudo labels, however, it also incurs a low utilization of the whole unlabeled set. As evidenced by our pilot study on CIFAR-100 [17], nearly $20\\%$ unlabeled images are filtered out for not satisfying the threshold of 0.95. Instead of blindly throwing them away, we believe there should exist a more promising approach. This work is just aimed to fully leverage previously uncertain samples in an informative but also safe manner.", + "bbox": [ + 496, + 595, + 893, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "So first, why are these samples uncertain? According to our observations on CIFAR-100 and ImageNet, although the top-1 accuracy could be low, the top-5 accuracy is much higher. This indicates in most cases, the model struggles to discriminate among a small portion of classes. As illustrated", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 47 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors", + "bbox": [ + 94, + 887, + 225, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16187", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/86eafff981b9f29c8572753a51af82349256cd27529a4ac7959752ef39fbf053.jpg", + "image_caption": [ + "Figure 2: Pseudo label accuracy on CIFAR-100 with 400 labels. We highlight that even for uncertain samples, their top-1 predictions are of high accuracy in the superclass space (20 classes). This accuracy can even be comparable to the delicately selected certain samples in the original class space." + ], + "image_footnote": [], + "bbox": [ + 107, + 88, + 441, + 237 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in Fig. 1, given a cat image, the model is not sure whether it belongs to tabby cat, tiger cat, or other cats. On the other hand, however, it is absolutely certain that the object is more like a tabby cat, rather than a table or anything else. In other words, it is reliable for the model to distinguish the top class from the remaining less likely classes.", + "bbox": [ + 75, + 334, + 468, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Invoked by these, we propose a novel method dubbed ShrinkMatch to learn uncertain samples. The prediction fails to satisfy the pre-defined threshold due to the existence of confusion classes for the top-1 class. Hence, our approach adaptively seeks a shrunk class space where the confusion classes are removed, to enable the re-calculated confidence to reach the original threshold. Moreover, the obtained shrunk class space is also required to be the largest among those that can satisfy the threshold. In a word, we seek a certain and largest shrunk space for uncertain samples. Then, logits of the strongly augmented image are correspondingly gathered in the new space. And a consistency regularization is imposed between the shrunk weak-strong predictions.", + "bbox": [ + 75, + 426, + 470, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Note that, the confusion classes are detected and removed in a fully automatic and instance-adaptive fashion. Moreover, even if the predicted top-1 class is not fully in line with the groundtruth label, they mostly belong to the same superclass. To prove this, as shown in Fig. 2, uncertain samples exhibit much higher pseudo label accuracy in the superclass space than the original space. This accuracy is even comparable to that of certain samples in the original class space. Therefore, contrasting these groundtruth-related classes against remaining unlikely classes is still highly beneficial, yielding more discriminative representations of our model. To avoid affecting the main classifier, we further adopt an auxiliary classifier to disentangle the learning in the shrunk space.", + "bbox": [ + 75, + 626, + 468, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the effectiveness, there still exist two main drawbacks to the above optimization target. (1) First, it treats all uncertain samples equally. The truth, however, is that in the original class space, the top-1 confidence of different uncertain samples can vary dramatically. And it is clear", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that samples with larger confidence should be attached more importance. To this end, we propose to balance different uncertain samples by their confidence in the original space. (2) Moreover, the regularization term also overlooks the gradually improved model state during training. At the start of training, there are abundant uncertain samples, but their predictions are extremely noisy, or even random. So even the highest scored class may share no relationship with the true class. Then as the training proceeds, the top classes become reliable. Considering this, we further propose to adaptively reweight the uncertain loss according to the model state. The model state is tracked and approximately estimated via performing exponential moving average on the proportion of certain samples in each mini-batch. With the two reweighting principles, the model turns out more stable, and avoids accumulating much noise from uncertain samples, especially at early training iterations.", + "bbox": [ + 496, + 90, + 893, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions lie in three aspects:", + "bbox": [ + 517, + 348, + 864, + 363 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We first point out that low certainty is typically caused by a small portion of confusion classes. To enhance the certainty, we propose to shrink the original class space by adaptively detecting and removing confusion ones for the top-1 class to turn it certain in the new space.", + "- We manage to reweight the uncertain loss from two perspectives: the image-based varied reliability among different uncertain samples, and the model-based gradually improved state as the training proceeds.", + "- Our proposed ShrinkMatch establishes new state-of-the-art results on widely acknowledged benchmarks." + ], + "bbox": [ + 517, + 375, + 893, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 571, + 640, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semi-supervised learning (SSL). The primary concern in SSL [19, 28, 36, 20, 25, 16, 37, 35, 41, 43, 14, 24, 29, 23, 4, 30, 11, 33, 2, 42] is to design effective constraints for unlabeled samples. Dating back to decades ago, pioneering works [19, 38] integrate unlabeled data via assigning pseudo labels to them, with the knowledge acquired from labeled data. In the era of deep learning, subsequent methods mainly follow such a bootstrapping fashion, but greatly boost it with some key components. Specifically, to enhance the quality of pseudo labels, $\\Pi$ -model [18] and Mean Teacher [31] ensemble model predictions and model parameters respectively. Later works start to exploit the role of perturbations. During this trend, [27] proposes to apply stochastic perturbations on inputs or features, and enforce consistency across these predictions. Then, UDA [34] emphasizes the necessity of strengthening the perturbation pool. It also follows VAT [22] and MixMatch [3] to supervise the prediction under strong perturbations with that under weak perturbations. Since then, weak-to-strong consistency regularization has become a standard practice in SSL. Eventually, the milestone", + "bbox": [ + 496, + 598, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16188", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "work FixMatch [28] presents a simplified framework using a fixed confidence threshold to discard uncertain samples. Our ShrinkMatch is built upon FixMatch. But, we highlight the value of previously neglected uncertain samples, and leverage them in an informative but also safe manner.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More recently, DST [7] decouples the generation and utilization of pseudo labels with a main and an auxiliary head respectively. Besides, SimMatch [45] explores instance-level relationships with labeled embeddings to supplement original class-level relations. Compared with them, our ShrinkMatch achieves larger improvements.", + "bbox": [ + 75, + 167, + 470, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Defining uncertain samples. Earlier works estimate the uncertainty with Bayesian Neural Networks [15], or its faster approximation, e.g., Monte Carlo Dropout [12]. Some other works measure the prediction disagreement among multiple randomly augmented inputs [32]. The latest trend is to directly use the entropy of predictions [39], cross entropy [36], or softmax confidence [28] as a measurement for uncertainty. Our work is not aimed at the optimal uncertainty estimation strategy, so we adopt the simplest solution from FixMatch, i.e., using the maximum softmax output as the certainty.", + "bbox": [ + 75, + 263, + 470, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Utilizing uncertain samples. UPS [26] leverages negative class labels whose confidence is below a pre-defined threshold, from a reversed multi-label classification perspective. In comparison, our model is enforced to tell the most likely class without being cheated by the less likely ones. So our supervision on uncertain samples is more informative and produces more discriminative representations. Moreover, we do not introduce any extra hyper-parameters, e.g., the lower threshold in [26], into our framework.", + "bbox": [ + 75, + 422, + 468, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 571, + 166, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We primarily provide some notations and review a common practice in semi-supervised learning (SSL) (Sec. 3.1). Next, we present our ShrinkMatch in detail (Sec. 3.2 and Sec. 3.3). Finally, we summarize our approach and provide a further discussion (Sec. 3.4 and Sec. 3.5).", + "bbox": [ + 75, + 595, + 470, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 681, + 217, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semi-supervised learning aims to learn a model with limited labeled images $\\mathcal{D}^l = \\{(x_k, y_k)\\}$ , aided by a large number of unlabeled images $D^u = \\{u_k\\}$ . Recent frameworks commonly follow the FixMatch practice. Concretely, an unlabeled image $u$ is first transformed by a weak augmentation pool $\\mathcal{A}^w$ and a strong augmentation pool $\\mathcal{A}^s$ to yield a pair of weakly and strongly augmented images $(u^w, u^s)$ . Then, they are fed into the model together to produce corresponding predictions $(p^w, p^s)$ . Typically, $p^w$ is of much higher accuracy than $p^s$ , while $p^s$ is beneficial to learning. Therefore, $p^w$ serves as the pseudo label for $p^s$ . Moreover, a core practice introduced by FixMatch is that, to improve the quality of selected pseudo labels, a fixed confidence threshold is", + "bbox": [ + 75, + 704, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pre-defined to abandon uncertain ones. So the unsupervised loss $\\mathcal{L}_u$ can be formulated as:", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {u} = \\frac {1}{B _ {u}} \\sum_ {k = 1} ^ {B _ {u}} \\mathbb {1} \\left(\\xi \\left(p _ {k} ^ {w}\\right) \\geq \\tau\\right) \\cdot \\mathrm {H} \\left(p _ {k} ^ {w}, p _ {k} ^ {s}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 121, + 890, + 159 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $B^{u}$ is the batch size of unlabeled images and $\\tau$ is the pre-defined threshold. $\\xi(p_{k}^{w})$ computes the confidence of logits $p_{k}^{w}$ by $\\xi(\\cdot) = \\max(\\sigma(\\cdot))$ , where $\\sigma$ is the softmax function. The H denotes the consistency regularization between the two distributions. It is typically the cross entropy loss.", + "bbox": [ + 496, + 160, + 893, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition, the labeled images are learned with a regular cross entropy loss to obtain the supervised loss $\\mathcal{L}_x$ . The overall loss in each mini-batch then will be:", + "bbox": [ + 496, + 234, + 890, + 279 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {x} + \\lambda_ {u} \\cdot \\mathcal {L} _ {u}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 281, + 890, + 295 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\lambda_{u}$ acts as a trade-off term between the two losses.", + "bbox": [ + 498, + 296, + 874, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Shrinking the Class Space for Certainty", + "text_level": 1, + "bbox": [ + 498, + 318, + 841, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our motivation. As reviewed above, FixMatch discards the samples whose confidence is lower than a pre-defined threshold, because their pseudo labels are empirically found relatively noisier. These samples are named uncertain samples in this work. Although this practice retains high-quality pseudo labels, it incurs a low utilization of the whole unlabeled set, especially when the scenario is challenging and the selection criterion is strict. Take the CIFAR-100 dataset as an instance, with a common threshold of 0.95 and 4 labels per class, there will be nearly $20\\%$ unlabeled samples being ignored due to their low certainty. We argue that, these uncertain samples can still benefit the model optimization, as long as we can design appropriate constraints (loss functions) on them. So we first investigate the cause of low certainty to gain some better intuitions.", + "bbox": [ + 496, + 340, + 893, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The reason for the low certainty of an unlabeled image is that, the model tends to be confused among some top classes. For example, given a cat image, the score of the class tabby cat and class tiger cat may be both high, so the model is not absolutely certain what the concrete class is. Motivated by this observation, we propose to shrink the original class space via adaptively detecting and removing the confusion classes for the top-1 class. Then the shrunk space is only composed of the original top-1 class, as well as the remaining less likely classes. After this process, re-calculated confidence of the top-1 class will satisfy the pre-defined threshold. Thereby, we can enforce the model to learn the previously uncertain samples in this new certain space, as shown in Fig. 3. Following the previous cat example, if the top-1 class is tabby cat, our method will scan scores of all other classes, and remove confusion ones (e.g., tiger cat and siamese cat) to construct a confident shrunk space, where the model is sure that the image is a tabby cat, rather than a table. Since we do not ask the model to discriminate among several top classes, it will avoid suffering from the noise when it makes a wrong judgment in the original space.", + "bbox": [ + 496, + 569, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16189", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/03d015f6bb40c84a03b58c2f1731515a2b38f84620b092cbcfa307c979e11d3d.jpg", + "image_caption": [ + "Figure 3: An overview of our proposed ShrinkMatch. Our motivation is to fully leverage the originally uncertain samples. \"S\" denotes shrinking the class space. The confusion classes for the top-1 class are detected and removed in a fully automatic and instance-adaptive fashion, to construct a shrunk space where the top-1 class is turned certain. $\\mathcal{L}_u$ is the original certain loss, while $\\mathcal{L}_s$ calculates the uncertain loss in the shrunk class space. We add an auxiliary head $h_{aux}$ to learn in the new space. On the right, we further reweight $\\mathcal{L}_s$ based on two principles. Principle 1 (image-based): image predictions with larger reliability are attached more importance. Principle 2 (model-based): We track the model state during training for reweighting." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 669, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bf167c4db3d499a9caeb5bf1f0765d7fd89760688be2bd5182872fdf52d228cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 88, + 867, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "How to seek the shrunk class space? Now, how can we enable our method to automatically seek the optimal shrunk space of an uncertain unlabeled image? Ideally, we hope this seeking process is free from any prior knowledge from humans, e.g., class relationships, and also does not require any extra hyper-parameters. Considering these, we opt to inherit the pre-defined confidence threshold as a criterion. To be specific, we first sort the predicted logits in a descending order to obtain $p^w = \\{s_{n_i}^w\\}_{i=1}^C$ for classes $\\{n_i\\}_{i=1}^C$ , where $s_{n_i}^w \\geq s_{n_i+1}^w$ . In the shrunk space, we will retain the original top-1 class, because it is still the most likely one to be true. Then, we find a set of less likely classes $\\{n_i\\}_{i=K}^C$ by enforcing two constraints on the $K$ :", + "bbox": [ + 75, + 364, + 472, + 561 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\xi \\left(\\left\\{s _ {n _ {1}} ^ {w} \\right\\} \\cup \\left\\{s _ {n _ {i}} ^ {w} \\right\\} _ {i = K} ^ {C}\\right) \\geq \\tau , \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 573, + 468, + 592 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\xi \\left(\\left\\{s _ {n _ {1}} ^ {w} \\right\\} \\cup \\left\\{s _ {n _ {i}} ^ {w} \\right\\} _ {i = K - 1} ^ {C}\\right) < \\tau , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 594, + 468, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\xi$ is defined the same as that in Eq. (1), calculating the confidence of the re-assembled logits. The final shrunk space is composed of the re-assembled classes $\\{n_1\\} \\cup \\{n_i\\}_{i = K}^C$ . The two constraints on $K$ not only ensure the top-1 class is turned certain in the new space (Eq. (3)), but also select the largest space among all candidates (Eq. (4)).", + "bbox": [ + 75, + 625, + 468, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "How to learn in the shrunk class space? For a weakly augmented uncertain image $x^w$ , the model is certain about its top-1 class in the shrunk class space. To learn effectively in this space, we follow the popular practice of weak-to-strong consistency regularization. The correspondingly shrunk prediction on the strongly augmented image is enforced to match that on the weakly augmented one. Concretely, the re-assembled logits from $p^w$ is denoted as $\\hat{p}^w$ , which means $\\hat{p}^w = \\{s_{n_1}^w\\} \\cup \\{s_{n_i}^w\\}_K^C$ . We use the re-assembled classes $\\{n_1\\} \\cup \\{n_i\\}_K^C$ in the shrunk space to correspondingly gather the logits $p^s$ on $x^s$ , yielding $\\hat{p}^s = \\{s_{n_1}^s\\} \\cup \\{s_{n_i}^s\\}_K^C$ . Then we can regularize the consistency between the two shrunk", + "bbox": [ + 75, + 719, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "distributions $\\hat{p}^s$ and $\\hat{p}^w$ , similar to that in Eq. (1):", + "bbox": [ + 500, + 364, + 823, + 380 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\frac {1}{B _ {u}} \\sum_ {k = 1} ^ {B _ {u}} \\mathbb {1} \\left(\\xi \\left(p _ {k} ^ {w}\\right) < \\tau\\right) \\cdot \\hat {\\mathrm {H}} \\left(\\hat {p} _ {k} ^ {w}, \\hat {p} _ {k} ^ {s}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 392, + 890, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the indicator function is to find uncertain samples.", + "bbox": [ + 498, + 444, + 874, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Empirically, we observe that if we use the original linear head $h_{main}$ to learn this auxiliary supervision, it will make the confidence of our model increase aggressively. Most noisy unlabeled samples are blindly judged as certain ones. We conjecture that it is because the $\\mathcal{L}_s$ strengthens weights of the classes that are frequently uncertain, then these classes will be incorrectly turned certain. With this in mind, our solution is simple. We adopt an auxiliary MLP head $h_{aux}$ that shares the backbone with $h_{main}$ , to deal with this auxiliary optimization target, as shown in Fig. 3. So the $\\hat{p}^s$ is indeed gathered from predictions of $h_{aux}$ ( $\\hat{p}^w$ is still from $h_{main}$ ). This modification enables our feature extractor to acquire more discriminative representations, and meantime does not affect predictions of the main head. Note that $h_{aux}$ is only applied for training, bringing no burden to the test stage.", + "bbox": [ + 496, + 459, + 892, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Reweighting the Uncertain Loss", + "text_level": 1, + "bbox": [ + 500, + 696, + 781, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite the effectiveness of the above uncertain loss, there still exist two main drawbacks. (1) On one hand, it overlooks the varied reliability of the top-1 class among different uncertain images. For example, suppose two uncertain images $u_{1}$ and $u_{2}$ with softmax predictions [0.8, 0.1, 0.1] and [0.5, 0.3, 0.2] in the original space, they should not be treated equally in the shrunk space. The $u_{1}$ with top-1 confidence of 0.8 is more likely to be true than $u_{2}$ , and thereby should be attached more attention to. (2) On the other, it ignores the gradually improved model performance as the training proceeds. To be concrete, at the very start of training, the predictions are extremely noisy or even random. And then", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16190", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "at later stages, the predictions become more and more reliable. So the uncertain predictions at different training stages should not be treated equally. Therefore, we further design two reweighting principles for the two concerns.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Principle 1: Reweighting with image-based varied reliability. According to the above intuition, we directly reweight the uncertain loss of each uncertain image by its top-1 confidence $\\xi(p_k^w)$ in the original class space, which is:", + "bbox": [ + 75, + 154, + 472, + 217 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} ^ {1} = \\frac {1}{B _ {u}} \\sum_ {k = 1} ^ {B _ {u}} \\mathbb {1} \\left(\\xi \\left(p _ {k} ^ {w}\\right) < \\tau\\right) \\cdot \\hat {\\mathrm {H}} \\left(\\hat {p} _ {k} ^ {w}, \\hat {p} _ {k} ^ {s}\\right) \\cdot \\xi \\left(p _ {k} ^ {w}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 220, + 470, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We do not use the top-1 confidence $\\xi(\\hat{p}_k^w)$ in the shrunk space as the weight, because after shrinking, this value of different predictions is very close to each other. So generally, $\\xi(p_k^w)$ is more discriminative than $\\xi(\\hat{p}_k^w)$ as the weight.", + "bbox": [ + 75, + 263, + 472, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Principle 2: Reweighting with model-based gradually improved state. One naive solution is to linearly increase the loss weight of $\\mathcal{L}_s$ from 0 at the beginning to $\\mu$ at iteration $T$ , and then keep $\\mu$ until the end. However, this practice has two severe disadvantages. First, the two additional hyperparameters $\\mu$ and $T$ are not easy to determine, and could be sensitive. More importantly, the linear scheduling criterion simply assumes the model state also improves linearly. Indeed, it can not be true. Thus, we here present a more promising principle to perform reweighting, that is free from any extra hyper-parameters. It can adjust the loss weight according to the model state in a fully adaptive fashion. To be specific, we use the certain ratio of the unlabeled set as an indicator for the model state. The certain ratio is traced at each iteration and accumulated globally in an exponential moving average (EMA) manner. Formally, the certain ratio in a single mini-batch is given by:", + "bbox": [ + 75, + 329, + 472, + 585 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm = \\frac {1}{B _ {u}} \\sum_ {k = 1} ^ {B _ {u}} \\mathbb {1} \\left(\\xi \\left(p _ {k} ^ {w}\\right) \\geq \\tau\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 592, + 468, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The global certain ratio $m^g$ is initialized as 0, and accumulated at each training iteration by:", + "bbox": [ + 76, + 633, + 472, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm ^ {g} \\leftarrow \\gamma \\cdot m ^ {g} + (1 - \\gamma) \\cdot m, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 670, + 468, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma$ is the momentum coefficient. It is a hyper-parameter already defined in FixMatch (baseline), where it is used to update the teacher parameters for final evaluation.", + "bbox": [ + 75, + 694, + 468, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Obviously, $m^g$ falls between 0 and 1. And it will approximately increase from 0 to a nearly saturated value. Then, the reweighted uncertain loss is given by:", + "bbox": [ + 75, + 739, + 470, + 784 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} ^ {2} = m ^ {g} \\cdot \\frac {1}{B _ {u}} \\sum_ {k = 1} ^ {B _ {u}} \\mathbb {1} \\left(\\xi \\left(p _ {k} ^ {w}\\right) < \\tau\\right) \\cdot \\hat {\\mathrm {H}} \\left(\\hat {p} _ {k} ^ {w}, \\hat {p} _ {k} ^ {s}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 785, + 468, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Integrating the above two intuitions and principles, the final reweighted uncertain loss will be:", + "bbox": [ + 76, + 825, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = m ^ {g} \\cdot \\frac {1}{B _ {u}} \\sum_ {k = 1} ^ {B _ {u}} \\mathbb {1} \\left(\\xi \\left(p _ {k} ^ {w}\\right) < \\tau\\right) \\cdot \\hat {\\mathrm {H}} \\left(\\hat {p} _ {k} ^ {w}, \\hat {p} _ {k} ^ {s}\\right) \\cdot \\xi \\left(p _ {k} ^ {w}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 862, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Summary", + "text_level": 1, + "bbox": [ + 500, + 90, + 612, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To summarize, the final loss in a mini-batch is a combination of the supervised loss $(\\mathcal{L}_x)$ , certain loss $(\\mathcal{L}_u$ , Eq. (1)), and uncertain loss in the shrunk class space $(\\mathcal{L}_s$ , Eq. (10)):", + "bbox": [ + 498, + 116, + 893, + 162 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {x} + \\lambda_ {u} \\cdot \\left(\\mathcal {L} _ {u} + \\mathcal {L} _ {s}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 179, + 892, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We do not carefully fine-tune the fusion weight between $\\mathcal{L}_u$ and $\\mathcal{L}_s$ , but use 1:1 by default to avoid hyper-parameters.", + "bbox": [ + 500, + 214, + 893, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Discussions", + "text_level": 1, + "bbox": [ + 500, + 262, + 624, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our uncertain loss in the shrunk space owns two properties: informative and also safe. The first property is because we manage to find the largest shrunk space that reaches the confidence threshold. Besides, we also adopt the weak-to-strong consistency regularization to pose a challenging optimization target. Both constraints ensure the learning in the shrunk space is not trivial and still quite informative. Next, we wish to explain the second property \"safe\", especially about how we avoid noise in the shrunk space.", + "bbox": [ + 496, + 287, + 893, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Noises in pseudo labels distinguish the semi-supervised paradigm from the fully-supervised one. So designing a safe optimization target for unlabeled data is crucial. Typically, the cross entropy loss will maximize the softmax probability $\\exp (s_t) / \\sum_{i = 1}^{C}\\exp (s_i)\\rightarrow 1$ for the target class $t$ . It inevitably suppresses scores of all other classes except $t$ . However, the true class could not be $t$ , and may be the $2^{\\mathrm{nd}}$ or $3^{\\mathrm{rd}}$ largest class, etc., which is wrongly restrained. This is frequently observed when the confidence of class $t$ is not high enough. As a milder alternative, the soft labeling still encounters a similar problem. In contrast, our shrunk target directly excludes these confusion classes. So only the almost unlikely classes are suppressed.", + "bbox": [ + 496, + 426, + 893, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is worth noting, even if the predicted top-1 class is not in line with the human label, it is probably one of the closest semantics, e.g., belonging to the same superclass as the groundtruth label, as shown in Fig. 2. Thus, contrasting these relevant classes against other less likely classes with our noise-robust shrunk loss is still beneficial. The model is encouraged to make discriminative predictions closer to top classes compared to tail classes. In addition, our disentangled auxiliary head can effectively leverage such supervision while not affecting the main classification tasks [7].", + "bbox": [ + 496, + 625, + 893, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 797, + 624, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first describe the implementation details of our framework. Then, we compare our ShrinkMatch with previous state-of-the-art methods (SOTAs) under extensive evaluation protocols. Lastly, we conduct comprehensive ablation studies on each component to validate the necessity.", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16191", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9b06fe82cd066293cf991aa656653e12a30db08a90cc6308e9a97463ebd5ca0a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Seed01234Mean
SimMatch [45]95.3495.1692.6393.7695.1094.39
ShrinkMatch | 4095.0994.6695.1294.7894.9594.92
SimMatch [45]95.5895.5095.3494.0695.2695.15
ShrinkMatch | 25095.3995.4495.3694.7695.3595.26
", + "bbox": [ + 80, + 88, + 467, + 186 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f93caf2b105632bb6645ae0dc5e6db66b25e5ecfa831cbbe932f02fe034d0b11.jpg", + "table_caption": [ + "Table 1: Comparison with SOTAs on CIFAR-10. The same seed ensures exactly the same data split. The 400 or 2500 denotes the number of labels." + ], + "table_footnote": [], + "table_body": "
Seed01234Mean
SimMatch [45]62.0660.1959.8964.8863.9262.19
ShrinkMatch | 40065.0063.4763.7766.4264.5264.64
SimMatch [45]74.6475.1974.5375.0375.2474.93
ShrinkMatch | 250075.0075.1174.5474.7874.7274.83
", + "bbox": [ + 78, + 250, + 468, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 393, + 294, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We use FixMatch + distribution alignment (DA) as our baseline on all datasets except ImageNet and SVHN. On ImageNet, we build our method on SimMatch. On SVHN, we discard DA. To be more convincing, we adopt the same codebase as our compared methods.", + "bbox": [ + 75, + 421, + 470, + 497 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hyper-parameters. Following prior arts, Wide ResNet-28-2 [40], WRN-28-8, WRN-37-2, and WRN-28-2 are used for CIFAR-10, CIFAR-100, STL-10, and SVHN respectively. A ResNet-50 [13] is used for ImageNet. The auxiliary head $h_{aux}$ for uncertain samples is a 3-layer MLP. On the ImageNet, we set $B_u = 64 \\times 5$ , but for other datasets, $B_u = 64 \\times 7$ . The labeled batch size is 64 for all datasets. On the ImageNet, the model is trained for 400 epochs, while on the others, it is trained for $2^{20}$ iterations. The initial learning rate is set as 0.03 for all datasets with a cosine scheduler. Specifically, on the ImageNet, it is warmed up for 5 epochs. The consistency regularization $H$ in $\\mathcal{L}_u$ on STL-10 and SVHN is a hard cross entropy (CE) loss, while on the CIFAR-10/100 and ImageNet, following the SimMatch and CoMatch [20], it is modified to a soft CE loss. The $\\hat{H}$ in our proposed $\\mathcal{L}_s$ is simply a hard CE loss. The weight $\\lambda_u$ for the two unsupervised losses is set as 10 on the ImageNet and 1 on others. The confidence threshold $\\tau$ is 0.7 on the ImageNet and 0.95 for others. The shared momentum coefficient $\\gamma$ between our global certain ratio $m^g$ and the teacher parameters is 0.999. Following the common practice, the teacher model is only maintained for final evaluation.", + "bbox": [ + 75, + 501, + 470, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. CIFAR-10 and CIFAR-100", + "text_level": 1, + "bbox": [ + 76, + 845, + 318, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The CIFAR dataset is composed of 50000/10000 training/validation images of size $32 \\times 32$ . The CIFAR-10 con", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/060124de43d2796039209300d53d5e2f2c8e2fda878526b51bc4ee3b875137ca.jpg", + "table_caption": [ + "Table 2: Comparison with SOTAs on CIFAR-100." + ], + "table_footnote": [], + "table_body": "
Seed012Mean
FixMatch [28]65.8567.9458.3064.03
FlexMatch [42]76.7168.2867.5570.85
ShrinkMatch | 4085.7585.6486.5585.98
FixMatch [28]90.9188.7190.9490.19
FlexMatch [42]91.3592.2991.6691.77
ShrinkMatch | 25091.1392.4391.1091.55
", + "bbox": [ + 501, + 88, + 887, + 208 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7db40f2d93af10e02d2d7154bf597f4190a2f6a89eefbae559129bec3046e78b.jpg", + "table_caption": [ + "Table 3: Comparison with SOTAs on STL-10." + ], + "table_footnote": [], + "table_body": "
Seed012Mean
FixMatch [28]94.5396.9097.1496.19
FlexMatch [42]89.1989.9396.3291.81
ShrinkMatch | 4097.9697.8196.7097.49
FixMatch [28]98.0097.9997.9497.98
FlexMatch [42]91.7691.8396.6593.41
ShrinkMatch | 25098.0898.0697.9898.04
", + "bbox": [ + 501, + 241, + 887, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4: Comparison with SOTAs on SVHN.", + "bbox": [ + 545, + 369, + 844, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tains 10 balanced classes to recognize, while CIFAR-10 is much more challenging, containing 100 classes. We run our ShrinkMatch on five different seeds and report each result as well as the average result. On the CIFAR-10 of Tab. 1 where performance is almost saturated, we still obtain a nontrivial improvement of $0.53\\%$ ( $94.39\\% \\to 94.92\\%$ ) with 4 labels per class (400 labels in total). In addition, notably, as shown in Tab. 2, with only 4 labels per class on the challenging CIFAR-100, our ShrinkMatch remarkably outperforms SimMatch by $2.45\\%$ ( $62.19\\% \\to 64.64\\%$ ) on average.", + "bbox": [ + 496, + 407, + 893, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. STL-10 and SVHN", + "text_level": 1, + "bbox": [ + 500, + 575, + 683, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The STL-10 dataset originally consists of 5K/100K/8K labeled/unlabeled/validation images. We follow FlexMatch to only select a subset of 40/250 labeled images, while the unlabeled set remains unchanged. As shown in Tab. 3, with 4 labels per class, our ShrinkMatch surpasses FlexMatch tremendously from $70.85\\%$ to $85.98\\%$ $(+15.13\\%)$ .", + "bbox": [ + 496, + 599, + 892, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The SVHN dataset is suitable to reveal the ability of different semi-supervised algorithms in the presence of class imbalance issue. As demonstrated in Tab. 4, although the performance of the original FixMatch (our baseline) has almost touched the upper bound, we still further boost it from $96.19\\%$ to $97.49\\%$ $(+1.30\\%)$ with 4 labels per class.", + "bbox": [ + 496, + 691, + 893, + 784 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. ImageNet-1K", + "text_level": 1, + "bbox": [ + 500, + 799, + 640, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The ImageNet dataset is rather challenging, containing 1.28M/50K training/validation images of 1000 classes. We exactly follow the codebase of SimMatch [45]. As shown in Tab. 5, our ShrinkMatch further boosts previous best results on both settings of $1\\%$ and $10\\%$ labeled images.", + "bbox": [ + 496, + 825, + 892, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16192", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e1afa3d50119300c4ef15f2032ec884bd3dd44b08c1dee1935e68bf5a208f884.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pre-trainingMethodEpochsParams (train / test)Top-1Top-5
1%10%1%10%
SimCLR v2 [8]80034.2M / 25.6M57.968.482.589.2
SwAV [6]Fine-tune80030.4M / 25.6M53.970.278.589.9
WCL [44]80034.2M / 25.6M65.072.086.391.2
MoCo v2 [9]Fine-tune80030.0M / 25.6M49.866.177.287.9
CoMatch [20]120030.0M / 25.6M67.173.787.191.4
MoCo-EMAN [5]FixMatch-EMAN [5]110030.0M / 25.6M63.074.083.490.9
NoneCoMatch [20]40030.0M / 25.6M66.073.686.491.6
SimMatch† [45]40030.0M / 25.6M67.074.186.991.5
ShrinkMatch40031.8M / 25.6M67.574.587.491.9
", + "bbox": [ + 116, + 88, + 849, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c87cc982bf63139eb4e6b1c2d4e25af9ce32ab22746ff03050467e7f250b60cb.jpg", + "image_caption": [ + "(a) Number of removed confusion classes." + ], + "image_footnote": [], + "bbox": [ + 84, + 306, + 343, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d4ac0e5765c69a424b6aaa025f86a60ac283a5d0a505573764e01c53cb7f3fa.jpg", + "image_caption": [ + "Figure 4: (a) The number of removed confusion classes for each uncertain image as the training proceeds. (b) Value of different loss functions, i.e., Eq. (1), Eq. (6), and Eq. (10). (c) The ratio of uncertain samples (certainty $< \\tau$ ) in each mini-batch." + ], + "image_footnote": [], + "bbox": [ + 354, + 305, + 609, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/99f0782972656e026945d310b4b2bac11410c18fb6fcae5fbdcba80b0ef1f0d7.jpg", + "image_caption": [ + "(b) Certain loss $\\mathcal{L}_u$ and uncertain loss $\\mathcal{L}_s$ .", + "(c) Ratio of uncertain samples." + ], + "image_footnote": [], + "bbox": [ + 620, + 306, + 879, + 438 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2f29df7cb38d810a87ec7d7a76f90fcc926129aea18c7f8f1beadf5dd8256cb1.jpg", + "table_caption": [ + "Table 5: Accuracy (%) on the ImageNet-1K with $1\\%$ and ${10}\\%$ labeled images. †: Reproduced in our environment." + ], + "table_footnote": [], + "table_body": "
Seed01234Mean
Baseline64.4057.5163.0764.7762.5362.46
ShrinkMatch65.0063.4763.7766.4264.5264.64
", + "bbox": [ + 80, + 510, + 464, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 611, + 240, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unless otherwise specified, we conduct our ablation studies on CIFAR-100 with 4 labels per class.", + "bbox": [ + 75, + 635, + 470, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of our holistic ShrinkMatch. We first view our ShrinkMatch as a holistic component added to our baseline. As displayed in Tab. 6 for CIFAR-100, our ShrinkMatch boosts the strong baseline significantly by $2.18\\%$ ( $62.46\\% \\rightarrow 64.64\\%$ ). And on the STL-10 of Tab. 7, the baseline is also improved evidently from $84.85\\%$ to $85.98\\%$ ( $+1.13\\%$ ).", + "bbox": [ + 75, + 669, + 470, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of two reweighting principles. In Sec. 3.3, considering the varied prediction reliability of different uncertain samples and different training stages, we propose two principles to reweight our uncertain loss $\\mathcal{L}_s$ . So here we carefully examine their necessity in Tab. 8. It can be observed that the two principles are both beneficial, jointly improving the basic shrinking practice from $63.16\\%$ to $64.24\\%$ . We also attempt a simple alternative of linearly increasing the uncertain loss weight from 0 (start) to 1 (end). But as evi", + "bbox": [ + 75, + 763, + 472, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/142460578d8a188a0b6a38c3dcbedb9609193744459980548a108b6eefdb8fb7.jpg", + "table_caption": [ + "Table 6: Ablation studies of ShrinkMatch on CIFAR-100." + ], + "table_footnote": [], + "table_body": "
Seed012Mean
Baseline84.8684.2185.4784.85
ShrinkMatch85.7585.6486.5585.98
", + "bbox": [ + 514, + 510, + 874, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7: Ablation studies of ShrinkMatch on STL-10.", + "bbox": [ + 516, + 575, + 874, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "denced in the Exp 3 & 4 of Tab. 8, it is obviously inferior to our designed reweighting strategy. We visualize our $m^g$ as training proceeds on ImageNet in Fig. 5. It looks essentially different from linear scheduling. Moreover, we visualize the uncertain loss value with or without the second reweighting principle in Fig. 4b. After reweighting, the uncertain loss is of a similar magnitude as the certain loss, which can avoid dominating the gradient at early training stages.", + "bbox": [ + 496, + 614, + 890, + 737 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The number of removed confusion classes. As introduced before, our approach detects and removes the confusion classes for the top-1 class in a fully automatic and instance-adaptive fashion. So we visualize the number of removed confusion classes for each uncertain sample in Fig. 4a. At the very start of training, model predictions are almost uniform, so abundant classes are removed for an uncertain sample. But as the training proceeds, much fewer confusion classes need to be removed to form a confident shrunk class space.", + "bbox": [ + 496, + 742, + 893, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hard label or soft label in the shrunk class space? By", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16193", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8aa678a1b0c5f7fe55e9163a040285f2ab1095b1b0ebdeb2cfa7deeb1cb02df9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ExpPrinciple 1Principle 2LSSeed 0Seed 1Mean
165.6260.7063.16
2✓65.6862.0863.88
3✓65.0962.9064.00
4✓63.0159.4561.23
5✓✓65.0063.4764.24
", + "bbox": [ + 78, + 88, + 470, + 188 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0bc0dd1bdc2a31d1e86aa6478e14619bf6261b6934c02e0052265ba1d105939f.jpg", + "table_caption": [ + "Table 8: Ablation studies on the effectiveness of the two reweighting principles. Due to randomness, the \"Mean\" results are more convincing. LS is short for linear scheduling, which linearly increases the loss weight." + ], + "table_footnote": [], + "table_body": "
ExpShrinkAux HeadLabelSeed 0Seed 1Mean
1Our Strong Baseline64.4057.5160.96
2✓✓H65.0063.4764.24
3✓✓S62.9259.7661.34
4✓H62.4358.8660.65
5✓S60.9260.2260.57
6✓H64.2159.1761.69
7✓S64.7560.8962.82
8H52.9051.8852.39
9S54.1055.7554.93
", + "bbox": [ + 78, + 265, + 465, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "default, we use the hard label in the shrunk space, following FixMatch. But we also ablate the choice of using the soft label as supervision. As shown in Exp 2 & 3 of Tab. 9, the hard label is much better than the soft label. We conjecture that this is because we have shrunk the class space to a safe one. The soft label will incur some unnecessary noise.", + "bbox": [ + 75, + 563, + 468, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of the auxiliary head. To prevent the main classifier from being affected, we use an auxiliary head to learn the uncertain samples in the shrunk class space only for discriminative representations. So we first provide the results when still using the main head for the shrunk space. As shown in Exp 2 & 4 of Tab. 9, the auxiliary head is indispensable for our shrinking practice. So we continue to figure out whether the improvement of our ShrinkMatch merely comes from the auxiliary head. We attempt to use the auxiliary head to directly learn the uncertain samples in the original space without shrinking. It can be concluded from Exp 2 & 6 & 7 of Tab. 9, the original space is indeed much inferior to our shrunk space for uncertain samples, although the auxiliary head still helps the baseline to some extent. Lastly, in Tab. 9, we also include two additional results for readers, i.e., directly learning uncertain samples without shrinking or", + "bbox": [ + 75, + 657, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/101689e4fbc26cfa05b514b02839eedd0961c52b84eaa44148e9c2bdecf24048.jpg", + "image_caption": [ + "Figure 5: Visualization of our global certain ratio $m^g$ ." + ], + "image_footnote": [], + "bbox": [ + 506, + 88, + 883, + 227 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6281296ced9011fa9544d89932055f909c74498bdea3e3df13c02b36c0ca38fc.jpg", + "table_caption": [ + "Table 9: Ablation studies on different options to learn uncertain sample. (1) whether to shrink the class space (Shrink), (2) whether to use an auxiliary head (Aux Head), and (3) use hard (H) or soft label (S). Exp 1 is our baseline without using uncertain images. Exp 8 & 9 are experiments of directly learning uncertain samples, which are quite noisy." + ], + "table_footnote": [], + "table_body": "
Threshold0.980.950.900.800.70
ShrinkMatch65.8065.0064.5562.7459.58
", + "bbox": [ + 504, + 253, + 883, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 10: Ablation studies of the confidence threshold.", + "bbox": [ + 513, + 305, + 875, + 320 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "auxiliary head (Exp 8 & 9), which are quite terrible because of introducing the abundant noise to our main classifier.", + "bbox": [ + 498, + 340, + 890, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The ratio of uncertain samples. To further justify our motivation, in Fig. 4c, we display the ratio of uncertain samples in each mini-batch. It can be seen that even in the middle of the whole training course, there are still around $40\\%$ and $30\\%$ uncertain samples on ImageNet and CIFAR-100 respectively. Therefore, an appropriate approach to utilizing these samples is necessary and definitely beneficial.", + "bbox": [ + 498, + 375, + 893, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different confidence thresholds $\\tau$ . We also try different thresholds for uncertain samples, as shown in Tab. 10. When increasing $\\tau$ from widely adopted 0.95 to 0.98 (more uncertain samples), our ShrinkMatch can be further improved. These results clearly highlight the effective and safe utilization of uncertain samples with our ShrinkMatch.", + "bbox": [ + 498, + 483, + 893, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 597, + 617, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we aim to fully leverage the uncertain samples in semi-supervised learning. We point out that the low certainty is typically caused by a small portion of confusion classes. Invoked by this, we propose a novel method dubbed ShrinkMatch to automatically detect and remove the confusion classes to construct a shrunk class space, where the top-1 class is turned certain. A weak-to-strong consistency regularization is enforced in the confident new space. Furthermore, we design two reweighting principles for the auxiliary uncertain loss, according to the reliability of different uncertain samples and the gradually improved state of the model. Consequently, our method establishes new state-of-the-art results on widely acknowledged benchmarks.", + "bbox": [ + 496, + 622, + 893, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported in part by the National Natural Science Foundation of China (62201484, 62222604, 62192783, 62206052), HKU Startup Fund, and HKU Seed Fund for Basic Research.", + "bbox": [ + 498, + 835, + 893, + 897 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16194", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eric Arazo, Diego Ortego, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In IJCNN, 2020. 1", + "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. In ICCV, 2021. 2", + "[3] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. In NeurIPS, 2019. 2", + "[4] Zhaowei Cai, Avinash Ravichandran, Paolo Favaro, Manchen Wang, Davide Modolo, Rahul Bhotika, Zhuowen Tu, and Stefano Soatto. Semi-supervised vision transformers at scale. In NeurIPS, 2022. 2", + "[5] Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Charless Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi-supervised learning. In CVPR, 2021. 7", + "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 7", + "[7] Baixu Chen, Junguang Jiang, Ximei Wang, Jianmin Wang, and Mingsheng Long. Debiased pseudo labeling in self-training. In NeurIPS, 2022. 3, 5", + "[8] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. In NeurIPS, 2020. 7", + "[9] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv:2003.04297, 2020. 7", + "[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 1", + "[11] Yue Fan, Dengxin Dai, Anna Kukleva, and Bernt Schiele. Cossl: Co-learning of representation and classifier for imbalanced semi-supervised learning. In CVPR, 2022. 2", + "[12] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In ICML, 2016. 3", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6", + "[14] Zijian Hu, Zhengyu Yang, Xuefeng Hu, and Ram Nevatia. Simple: similar pseudo label exploitation for semi-supervised classification. In CVPR, 2021. 2", + "[15] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? In NeurIPS, 2017. 3", + "[16] Jiwon Kim, Youngjo Min, Daehwan Kim, Gyuseong Lee, Junyoung Seo, Kwangrok Ryoo, and Seungryong Kim. Conmatch: Semi-supervised learning with confidence-guided consistency regularization. In ECCV, 2022. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1", + "[18] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In ICLR, 2017. 2", + "[19] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In ICML Workshop, 2013. 2", + "[20] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regularization. In ICCV, 2021. 2, 6, 7", + "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 1", + "[22] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. TPAMI, 2018. 2", + "[23] Islam Nassar, Samitha Herath, Ehsan Abbasnejad, Wray Buntine, and Gholamreza Haffari. All labels are not created equal: Enhancing semi-supervision via label grouping and co-training. In CVPR, 2021. 2", + "[24] Yassine Ouali, Céline Hudelot, and Myriam Tami. An overview of deep semi-supervised learning. arXiv:2006.05278, 2020. 2", + "[25] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V Le. Meta pseudo labels. In CVPR, 2021. 2", + "[26] Mamshad Nayeem Rizve, Kevin Duarte, Yogesh S Rawat, and Mubarak Shah. In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. In ICLR, 2021. 3", + "[27] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In NeurIPS, 2016. 2", + "[28] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 1, 2, 3, 6", + "[29] Fariborz Taherkhani, Ali Dabouei, Sobhan Soleymani, Jeremy Dawson, and Nasser M Nasrabadi. Self-supervised Wasserstein pseudo-labeling for semi-supervised image classification. In CVPR, 2021. 2", + "[30] Hui Tang and Kui Jia. Towards discovering the effectiveness of moderately confident samples for semi-supervised learning. In CVPR, 2022. 2", + "[31] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In NeurIPS, 2017. 2", + "[32] Guotai Wang, Wenqi Li, Michael Aertsen, Jan Deprest, Sébastien Ourselin, and Tom Vercauteren. Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolutional neural networks. Neurocomputing, 2019. 3", + "[33] Xudong Wang, Long Lian, and Stella X Yu. Unsupervised selective labeling for more effective semi-supervised learning. In ECCV, 2022. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16195", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V Le. Unsupervised data augmentation for consistency training. In NeurIPS, 2020. 2", + "[35] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020. 2", + "[36] Yi Xu, Lei Shang, Jinxing Ye, Qi Qian, Yu-Feng Li, Baigui Sun, Hao Li, and Rong Jin. Dash: Semi-supervised learning with dynamic thresholding. In ICML, 2021. 2, 3", + "[37] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In CVPR, 2022. 2", + "[38] David Yarowsky. Unsupervised word sense disambiguation, rivaling supervised methods. In ACL, 1995. 2", + "[39] Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In MICCAI, 2019. 3", + "[40] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC, 2016. 6", + "[41] Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. In ICCV, 2019. 2", + "[42] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In NeurIPS, 2021. 2, 6", + "[43] Zhen Zhao, Luping Zhou, Lei Wang, Yinghuan Shi, and Yang Gao. Lassl: Label-guided self-training for semi-supervised learning. In AAAI, 2022. 2", + "[44] Mingkai Zheng, Fei Wang, Shan You, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Weakly supervised contrastive learning. In ICCV, 2021. 7", + "[45] Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In CVPR, 2022. 3, 6, 7" + ], + "bbox": [ + 78, + 90, + 470, + 618 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16196", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning/34917a9e-765e-47a5-bd3d-8044ab9de911_model.json b/2023/Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning/34917a9e-765e-47a5-bd3d-8044ab9de911_model.json new file mode 100644 index 0000000000000000000000000000000000000000..20ccc38738980e7cb4cec492c84379d6a69b92f5 --- /dev/null +++ b/2023/Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning/34917a9e-765e-47a5-bd3d-8044ab9de911_model.json @@ -0,0 +1,2211 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.048 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.1, + 0.131, + 0.872, + 0.156 + ], + "angle": 0, + "content": "Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.179, + 0.837, + 0.201 + ], + "angle": 0, + "content": "Lihe Yang\\(^{1}\\) Zhen Zhao\\(^{4}\\) Lei Qi\\(^{5}\\) Yu Qiao\\(^{3}\\) Yinghuan Shi\\(^{2*}\\) Hengshuang Zhao\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.202, + 0.697, + 0.221 + ], + "angle": 0, + "content": "1The University of Hong Kong 2Nanjing University" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.221, + 0.796, + 0.239 + ], + "angle": 0, + "content": "3Shanghai AI Laboratory 4The University of Sydney 5Southeast University" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.241, + 0.663, + 0.255 + ], + "angle": 0, + "content": "https://github.com/LiheYoung/ShrinkMatch" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.291, + 0.314, + 0.307 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.474, + 0.671 + ], + "angle": 0, + "content": "Semi-supervised learning is attracting blooming attention, due to its success in combining unlabeled data. To mitigate potentially incorrect pseudo labels, recent frameworks mostly set a fixed confidence threshold to discard uncertain samples. This practice ensures high-quality pseudo labels, but incurs a relatively low utilization of the whole unlabeled set. In this work, our key insight is that these uncertain samples can be turned into certain ones, as long as the confusion classes for the top-1 class are detected and removed. Invoked by this, we propose a novel method dubbed ShrinkMatch to learn uncertain samples. For each uncertain sample, it adaptively seeks a shrunk class space, which merely contains the original top-1 class, as well as remaining less likely classes. Since the confusion ones are removed in this space, the re-calculated top-1 confidence can satisfy the pre-defined threshold. We then impose a consistency regularization between a pair of strongly and weakly augmented samples in the shrunk space to strive for discriminative representations. Furthermore, considering the varied reliability among uncertain samples and the gradually improved model during training, we correspondingly design two reweighting principles for our uncertain loss. Our method exhibits impressive performance on widely adopted benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.7, + 0.21, + 0.715 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.725, + 0.471, + 0.83 + ], + "angle": 0, + "content": "In the last decade, our computer vision community has witnessed inspiring progress, thanks to large-scale datasets [21, 10]. Nevertheless, it is laborious and costly to annotate massive images, hindering the progress to benefit a broader range of real-world scenarios. Inspired by this, semi-supervised learning (SSL) was proposed to utilize the unlabeled data under the assistance of limited labeled data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.831, + 0.471, + 0.878 + ], + "angle": 0, + "content": "The frameworks in SSL are typically based on the strategy of pseudo labeling. Briefly, the model acquires knowledge from the labeled data, and then assigns predictions on the" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.288, + 0.856, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.49, + 0.896, + 0.583 + ], + "angle": 0, + "content": "Figure 1: Illustration of our motivation. Due to confusion classes for the top-1 class, the certainty fails to reach the pre-defined threshold (gray dotted line). FixMatch discards such uncertain samples. Our method, however, detects and removes confusion classes to enhance certainty, then enjoying full and safe utilization of all unlabeled images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.895, + 0.824 + ], + "angle": 0, + "content": "unlabeled data. The two sources of data are finally combined to train a better model. During this process, it is obvious that predictions on unlabeled data are not reliable. If the model is iteratively trained with incorrect pseudo labels, it will suffer the confirmation bias issue [1]. To address this dilemma, recent works [28] simply set a fixed confidence threshold to discard potentially unreliable samples. This simple strategy effectively retains high-quality pseudo labels, however, it also incurs a low utilization of the whole unlabeled set. As evidenced by our pilot study on CIFAR-100 [17], nearly \\(20\\%\\) unlabeled images are filtered out for not satisfying the threshold of 0.95. Instead of blindly throwing them away, we believe there should exist a more promising approach. This work is just aimed to fully leverage previously uncertain samples in an informative but also safe manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "So first, why are these samples uncertain? According to our observations on CIFAR-100 and ImageNet, although the top-1 accuracy could be low, the top-5 accuracy is much higher. This indicates in most cases, the model struggles to discriminate among a small portion of classes. As illustrated" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.227, + 0.901 + ], + "angle": 0, + "content": "*Corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16187" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.089, + 0.442, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.249, + 0.472, + 0.326 + ], + "angle": 0, + "content": "Figure 2: Pseudo label accuracy on CIFAR-100 with 400 labels. We highlight that even for uncertain samples, their top-1 predictions are of high accuracy in the superclass space (20 classes). This accuracy can even be comparable to the delicately selected certain samples in the original class space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.47, + 0.426 + ], + "angle": 0, + "content": "in Fig. 1, given a cat image, the model is not sure whether it belongs to tabby cat, tiger cat, or other cats. On the other hand, however, it is absolutely certain that the object is more like a tabby cat, rather than a table or anything else. In other words, it is reliable for the model to distinguish the top class from the remaining less likely classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.428, + 0.471, + 0.625 + ], + "angle": 0, + "content": "Invoked by these, we propose a novel method dubbed ShrinkMatch to learn uncertain samples. The prediction fails to satisfy the pre-defined threshold due to the existence of confusion classes for the top-1 class. Hence, our approach adaptively seeks a shrunk class space where the confusion classes are removed, to enable the re-calculated confidence to reach the original threshold. Moreover, the obtained shrunk class space is also required to be the largest among those that can satisfy the threshold. In a word, we seek a certain and largest shrunk space for uncertain samples. Then, logits of the strongly augmented image are correspondingly gathered in the new space. And a consistency regularization is imposed between the shrunk weak-strong predictions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.824 + ], + "angle": 0, + "content": "Note that, the confusion classes are detected and removed in a fully automatic and instance-adaptive fashion. Moreover, even if the predicted top-1 class is not fully in line with the groundtruth label, they mostly belong to the same superclass. To prove this, as shown in Fig. 2, uncertain samples exhibit much higher pseudo label accuracy in the superclass space than the original space. This accuracy is even comparable to that of certain samples in the original class space. Therefore, contrasting these groundtruth-related classes against remaining unlikely classes is still highly beneficial, yielding more discriminative representations of our model. To avoid affecting the main classifier, we further adopt an auxiliary classifier to disentangle the learning in the shrunk space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Despite the effectiveness, there still exist two main drawbacks to the above optimization target. (1) First, it treats all uncertain samples equally. The truth, however, is that in the original class space, the top-1 confidence of different uncertain samples can vary dramatically. And it is clear" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.348 + ], + "angle": 0, + "content": "that samples with larger confidence should be attached more importance. To this end, we propose to balance different uncertain samples by their confidence in the original space. (2) Moreover, the regularization term also overlooks the gradually improved model state during training. At the start of training, there are abundant uncertain samples, but their predictions are extremely noisy, or even random. So even the highest scored class may share no relationship with the true class. Then as the training proceeds, the top classes become reliable. Considering this, we further propose to adaptively reweight the uncertain loss according to the model state. The model state is tracked and approximately estimated via performing exponential moving average on the proportion of certain samples in each mini-batch. With the two reweighting principles, the model turns out more stable, and avoids accumulating much noise from uncertain samples, especially at early training iterations." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.349, + 0.866, + 0.364 + ], + "angle": 0, + "content": "To summarize, our contributions lie in three aspects:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.375, + 0.892, + 0.452 + ], + "angle": 0, + "content": "- We first point out that low certainty is typically caused by a small portion of confusion classes. To enhance the certainty, we propose to shrink the original class space by adaptively detecting and removing confusion ones for the top-1 class to turn it certain in the new space." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.46, + 0.892, + 0.522 + ], + "angle": 0, + "content": "- We manage to reweight the uncertain loss from two perspectives: the image-based varied reliability among different uncertain samples, and the model-based gradually improved state as the training proceeds." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.53, + 0.894, + 0.562 + ], + "angle": 0, + "content": "- Our proposed ShrinkMatch establishes new state-of-the-art results on widely acknowledged benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.375, + 0.894, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.573, + 0.642, + 0.59 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Semi-supervised learning (SSL). The primary concern in SSL [19, 28, 36, 20, 25, 16, 37, 35, 41, 43, 14, 24, 29, 23, 4, 30, 11, 33, 2, 42] is to design effective constraints for unlabeled samples. Dating back to decades ago, pioneering works [19, 38] integrate unlabeled data via assigning pseudo labels to them, with the knowledge acquired from labeled data. In the era of deep learning, subsequent methods mainly follow such a bootstrapping fashion, but greatly boost it with some key components. Specifically, to enhance the quality of pseudo labels, \\(\\Pi\\)-model [18] and Mean Teacher [31] ensemble model predictions and model parameters respectively. Later works start to exploit the role of perturbations. During this trend, [27] proposes to apply stochastic perturbations on inputs or features, and enforce consistency across these predictions. Then, UDA [34] emphasizes the necessity of strengthening the perturbation pool. It also follows VAT [22] and MixMatch [3] to supervise the prediction under strong perturbations with that under weak perturbations. Since then, weak-to-strong consistency regularization has become a standard practice in SSL. Eventually, the milestone" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16188" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "work FixMatch [28] presents a simplified framework using a fixed confidence threshold to discard uncertain samples. Our ShrinkMatch is built upon FixMatch. But, we highlight the value of previously neglected uncertain samples, and leverage them in an informative but also safe manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.471, + 0.258 + ], + "angle": 0, + "content": "More recently, DST [7] decouples the generation and utilization of pseudo labels with a main and an auxiliary head respectively. Besides, SimMatch [45] explores instance-level relationships with labeled embeddings to supplement original class-level relations. Compared with them, our ShrinkMatch achieves larger improvements." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.265, + 0.471, + 0.416 + ], + "angle": 0, + "content": "Defining uncertain samples. Earlier works estimate the uncertainty with Bayesian Neural Networks [15], or its faster approximation, e.g., Monte Carlo Dropout [12]. Some other works measure the prediction disagreement among multiple randomly augmented inputs [32]. The latest trend is to directly use the entropy of predictions [39], cross entropy [36], or softmax confidence [28] as a measurement for uncertainty. Our work is not aimed at the optimal uncertainty estimation strategy, so we adopt the simplest solution from FixMatch, i.e., using the maximum softmax output as the certainty." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.423, + 0.47, + 0.558 + ], + "angle": 0, + "content": "Utilizing uncertain samples. UPS [26] leverages negative class labels whose confidence is below a pre-defined threshold, from a reversed multi-label classification perspective. In comparison, our model is enforced to tell the most likely class without being cheated by the less likely ones. So our supervision on uncertain samples is more informative and produces more discriminative representations. Moreover, we do not introduce any extra hyper-parameters, e.g., the lower threshold in [26], into our framework." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.572, + 0.168, + 0.587 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.471, + 0.673 + ], + "angle": 0, + "content": "We primarily provide some notations and review a common practice in semi-supervised learning (SSL) (Sec. 3.1). Next, we present our ShrinkMatch in detail (Sec. 3.2 and Sec. 3.3). Finally, we summarize our approach and provide a further discussion (Sec. 3.4 and Sec. 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.682, + 0.218, + 0.696 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Semi-supervised learning aims to learn a model with limited labeled images \\(\\mathcal{D}^l = \\{(x_k, y_k)\\}\\), aided by a large number of unlabeled images \\(D^u = \\{u_k\\}\\). Recent frameworks commonly follow the FixMatch practice. Concretely, an unlabeled image \\(u\\) is first transformed by a weak augmentation pool \\(\\mathcal{A}^w\\) and a strong augmentation pool \\(\\mathcal{A}^s\\) to yield a pair of weakly and strongly augmented images \\((u^w, u^s)\\). Then, they are fed into the model together to produce corresponding predictions \\((p^w, p^s)\\). Typically, \\(p^w\\) is of much higher accuracy than \\(p^s\\), while \\(p^s\\) is beneficial to learning. Therefore, \\(p^w\\) serves as the pseudo label for \\(p^s\\). Moreover, a core practice introduced by FixMatch is that, to improve the quality of selected pseudo labels, a fixed confidence threshold is" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "pre-defined to abandon uncertain ones. So the unsupervised loss \\(\\mathcal{L}_u\\) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.122, + 0.892, + 0.16 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {u} = \\frac {1}{B _ {u}} \\sum_ {k = 1} ^ {B _ {u}} \\mathbb {1} \\left(\\xi \\left(p _ {k} ^ {w}\\right) \\geq \\tau\\right) \\cdot \\mathrm {H} \\left(p _ {k} ^ {w}, p _ {k} ^ {s}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.161, + 0.894, + 0.235 + ], + "angle": 0, + "content": "where \\( B^{u} \\) is the batch size of unlabeled images and \\( \\tau \\) is the pre-defined threshold. \\( \\xi(p_{k}^{w}) \\) computes the confidence of logits \\( p_{k}^{w} \\) by \\( \\xi(\\cdot) = \\max(\\sigma(\\cdot)) \\), where \\( \\sigma \\) is the softmax function. The H denotes the consistency regularization between the two distributions. It is typically the cross entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.236, + 0.892, + 0.28 + ], + "angle": 0, + "content": "In addition, the labeled images are learned with a regular cross entropy loss to obtain the supervised loss \\(\\mathcal{L}_x\\). The overall loss in each mini-batch then will be:" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.282, + 0.892, + 0.296 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {x} + \\lambda_ {u} \\cdot \\mathcal {L} _ {u}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.297, + 0.875, + 0.312 + ], + "angle": 0, + "content": "where \\(\\lambda_{u}\\) acts as a trade-off term between the two losses." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.319, + 0.842, + 0.336 + ], + "angle": 0, + "content": "3.2. Shrinking the Class Space for Certainty" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.894, + 0.568 + ], + "angle": 0, + "content": "Our motivation. As reviewed above, FixMatch discards the samples whose confidence is lower than a pre-defined threshold, because their pseudo labels are empirically found relatively noisier. These samples are named uncertain samples in this work. Although this practice retains high-quality pseudo labels, it incurs a low utilization of the whole unlabeled set, especially when the scenario is challenging and the selection criterion is strict. Take the CIFAR-100 dataset as an instance, with a common threshold of 0.95 and 4 labels per class, there will be nearly \\(20\\%\\) unlabeled samples being ignored due to their low certainty. We argue that, these uncertain samples can still benefit the model optimization, as long as we can design appropriate constraints (loss functions) on them. So we first investigate the cause of low certainty to gain some better intuitions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.895, + 0.901 + ], + "angle": 0, + "content": "The reason for the low certainty of an unlabeled image is that, the model tends to be confused among some top classes. For example, given a cat image, the score of the class tabby cat and class tiger cat may be both high, so the model is not absolutely certain what the concrete class is. Motivated by this observation, we propose to shrink the original class space via adaptively detecting and removing the confusion classes for the top-1 class. Then the shrunk space is only composed of the original top-1 class, as well as the remaining less likely classes. After this process, re-calculated confidence of the top-1 class will satisfy the pre-defined threshold. Thereby, we can enforce the model to learn the previously uncertain samples in this new certain space, as shown in Fig. 3. Following the previous cat example, if the top-1 class is tabby cat, our method will scan scores of all other classes, and remove confusion ones (e.g., tiger cat and siamese cat) to construct a confident shrunk space, where the model is sure that the image is a tabby cat, rather than a table. Since we do not ask the model to discriminate among several top classes, it will avoid suffering from the noise when it makes a wrong judgment in the original space." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16189" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.67, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.089, + 0.868, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.256, + 0.894, + 0.348 + ], + "angle": 0, + "content": "Figure 3: An overview of our proposed ShrinkMatch. Our motivation is to fully leverage the originally uncertain samples. \"S\" denotes shrinking the class space. The confusion classes for the top-1 class are detected and removed in a fully automatic and instance-adaptive fashion, to construct a shrunk space where the top-1 class is turned certain. \\(\\mathcal{L}_u\\) is the original certain loss, while \\(\\mathcal{L}_s\\) calculates the uncertain loss in the shrunk class space. We add an auxiliary head \\(h_{aux}\\) to learn in the new space. On the right, we further reweight \\(\\mathcal{L}_s\\) based on two principles. Principle 1 (image-based): image predictions with larger reliability are attached more importance. Principle 2 (model-based): We track the model state during training for reweighting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.473, + 0.562 + ], + "angle": 0, + "content": "How to seek the shrunk class space? Now, how can we enable our method to automatically seek the optimal shrunk space of an uncertain unlabeled image? Ideally, we hope this seeking process is free from any prior knowledge from humans, e.g., class relationships, and also does not require any extra hyper-parameters. Considering these, we opt to inherit the pre-defined confidence threshold as a criterion. To be specific, we first sort the predicted logits in a descending order to obtain \\( p^w = \\{s_{n_i}^w\\}_{i=1}^C \\) for classes \\( \\{n_i\\}_{i=1}^C \\), where \\( s_{n_i}^w \\geq s_{n_i+1}^w \\). In the shrunk space, we will retain the original top-1 class, because it is still the most likely one to be true. Then, we find a set of less likely classes \\( \\{n_i\\}_{i=K}^C \\) by enforcing two constraints on the \\( K \\):" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.574, + 0.47, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\xi \\left(\\left\\{s _ {n _ {1}} ^ {w} \\right\\} \\cup \\left\\{s _ {n _ {i}} ^ {w} \\right\\} _ {i = K} ^ {C}\\right) \\geq \\tau , \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.595, + 0.47, + 0.614 + ], + "angle": 0, + "content": "\\[\n\\xi \\left(\\left\\{s _ {n _ {1}} ^ {w} \\right\\} \\cup \\left\\{s _ {n _ {i}} ^ {w} \\right\\} _ {i = K - 1} ^ {C}\\right) < \\tau , \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.47, + 0.716 + ], + "angle": 0, + "content": "where \\(\\xi\\) is defined the same as that in Eq. (1), calculating the confidence of the re-assembled logits. The final shrunk space is composed of the re-assembled classes \\(\\{n_1\\} \\cup \\{n_i\\}_{i = K}^C\\). The two constraints on \\(K\\) not only ensure the top-1 class is turned certain in the new space (Eq. (3)), but also select the largest space among all candidates (Eq. (4))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.901 + ], + "angle": 0, + "content": "How to learn in the shrunk class space? For a weakly augmented uncertain image \\( x^w \\), the model is certain about its top-1 class in the shrunk class space. To learn effectively in this space, we follow the popular practice of weak-to-strong consistency regularization. The correspondingly shrunk prediction on the strongly augmented image is enforced to match that on the weakly augmented one. Concretely, the re-assembled logits from \\( p^w \\) is denoted as \\( \\hat{p}^w \\), which means \\( \\hat{p}^w = \\{s_{n_1}^w\\} \\cup \\{s_{n_i}^w\\}_K^C \\). We use the re-assembled classes \\( \\{n_1\\} \\cup \\{n_i\\}_K^C \\) in the shrunk space to correspondingly gather the logits \\( p^s \\) on \\( x^s \\), yielding \\( \\hat{p}^s = \\{s_{n_1}^s\\} \\cup \\{s_{n_i}^s\\}_K^C \\). Then we can regularize the consistency between the two shrunk" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.366, + 0.825, + 0.381 + ], + "angle": 0, + "content": "distributions \\(\\hat{p}^s\\) and \\(\\hat{p}^w\\), similar to that in Eq. (1):" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.393, + 0.892, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\frac {1}{B _ {u}} \\sum_ {k = 1} ^ {B _ {u}} \\mathbb {1} \\left(\\xi \\left(p _ {k} ^ {w}\\right) < \\tau\\right) \\cdot \\hat {\\mathrm {H}} \\left(\\hat {p} _ {k} ^ {w}, \\hat {p} _ {k} ^ {s}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.445, + 0.875, + 0.459 + ], + "angle": 0, + "content": "where the indicator function is to find uncertain samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.893, + 0.687 + ], + "angle": 0, + "content": "Empirically, we observe that if we use the original linear head \\( h_{main} \\) to learn this auxiliary supervision, it will make the confidence of our model increase aggressively. Most noisy unlabeled samples are blindly judged as certain ones. We conjecture that it is because the \\( \\mathcal{L}_s \\) strengthens weights of the classes that are frequently uncertain, then these classes will be incorrectly turned certain. With this in mind, our solution is simple. We adopt an auxiliary MLP head \\( h_{aux} \\) that shares the backbone with \\( h_{main} \\), to deal with this auxiliary optimization target, as shown in Fig. 3. So the \\( \\hat{p}^s \\) is indeed gathered from predictions of \\( h_{aux} \\) (\\( \\hat{p}^w \\) is still from \\( h_{main} \\)). This modification enables our feature extractor to acquire more discriminative representations, and meantime does not affect predictions of the main head. Note that \\( h_{aux} \\) is only applied for training, bringing no burden to the test stage." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.782, + 0.712 + ], + "angle": 0, + "content": "3.3. Reweighting the Uncertain Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Despite the effectiveness of the above uncertain loss, there still exist two main drawbacks. (1) On one hand, it overlooks the varied reliability of the top-1 class among different uncertain images. For example, suppose two uncertain images \\( u_{1} \\) and \\( u_{2} \\) with softmax predictions [0.8, 0.1, 0.1] and [0.5, 0.3, 0.2] in the original space, they should not be treated equally in the shrunk space. The \\( u_{1} \\) with top-1 confidence of 0.8 is more likely to be true than \\( u_{2} \\), and thereby should be attached more attention to. (2) On the other, it ignores the gradually improved model performance as the training proceeds. To be concrete, at the very start of training, the predictions are extremely noisy or even random. And then" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16190" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "at later stages, the predictions become more and more reliable. So the uncertain predictions at different training stages should not be treated equally. Therefore, we further design two reweighting principles for the two concerns." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.155, + 0.473, + 0.218 + ], + "angle": 0, + "content": "Principle 1: Reweighting with image-based varied reliability. According to the above intuition, we directly reweight the uncertain loss of each uncertain image by its top-1 confidence \\(\\xi(p_k^w)\\) in the original class space, which is:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.221, + 0.472, + 0.262 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} ^ {1} = \\frac {1}{B _ {u}} \\sum_ {k = 1} ^ {B _ {u}} \\mathbb {1} \\left(\\xi \\left(p _ {k} ^ {w}\\right) < \\tau\\right) \\cdot \\hat {\\mathrm {H}} \\left(\\hat {p} _ {k} ^ {w}, \\hat {p} _ {k} ^ {s}\\right) \\cdot \\xi \\left(p _ {k} ^ {w}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.265, + 0.473, + 0.327 + ], + "angle": 0, + "content": "We do not use the top-1 confidence \\(\\xi(\\hat{p}_k^w)\\) in the shrunk space as the weight, because after shrinking, this value of different predictions is very close to each other. So generally, \\(\\xi(p_k^w)\\) is more discriminative than \\(\\xi(\\hat{p}_k^w)\\) as the weight." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.33, + 0.473, + 0.587 + ], + "angle": 0, + "content": "Principle 2: Reweighting with model-based gradually improved state. One naive solution is to linearly increase the loss weight of \\(\\mathcal{L}_s\\) from 0 at the beginning to \\(\\mu\\) at iteration \\(T\\), and then keep \\(\\mu\\) until the end. However, this practice has two severe disadvantages. First, the two additional hyperparameters \\(\\mu\\) and \\(T\\) are not easy to determine, and could be sensitive. More importantly, the linear scheduling criterion simply assumes the model state also improves linearly. Indeed, it can not be true. Thus, we here present a more promising principle to perform reweighting, that is free from any extra hyper-parameters. It can adjust the loss weight according to the model state in a fully adaptive fashion. To be specific, we use the certain ratio of the unlabeled set as an indicator for the model state. The certain ratio is traced at each iteration and accumulated globally in an exponential moving average (EMA) manner. Formally, the certain ratio in a single mini-batch is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.593, + 0.47, + 0.634 + ], + "angle": 0, + "content": "\\[\nm = \\frac {1}{B _ {u}} \\sum_ {k = 1} ^ {B _ {u}} \\mathbb {1} \\left(\\xi \\left(p _ {k} ^ {w}\\right) \\geq \\tau\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.634, + 0.473, + 0.664 + ], + "angle": 0, + "content": "The global certain ratio \\( m^g \\) is initialized as 0, and accumulated at each training iteration by:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.671, + 0.47, + 0.688 + ], + "angle": 0, + "content": "\\[\nm ^ {g} \\leftarrow \\gamma \\cdot m ^ {g} + (1 - \\gamma) \\cdot m, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.695, + 0.47, + 0.739 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is the momentum coefficient. It is a hyper-parameter already defined in FixMatch (baseline), where it is used to update the teacher parameters for final evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.741, + 0.472, + 0.785 + ], + "angle": 0, + "content": "Obviously, \\( m^g \\) falls between 0 and 1. And it will approximately increase from 0 to a nearly saturated value. Then, the reweighted uncertain loss is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.786, + 0.47, + 0.826 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} ^ {2} = m ^ {g} \\cdot \\frac {1}{B _ {u}} \\sum_ {k = 1} ^ {B _ {u}} \\mathbb {1} \\left(\\xi \\left(p _ {k} ^ {w}\\right) < \\tau\\right) \\cdot \\hat {\\mathrm {H}} \\left(\\hat {p} _ {k} ^ {w}, \\hat {p} _ {k} ^ {s}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.47, + 0.856 + ], + "angle": 0, + "content": "Integrating the above two intuitions and principles, the final reweighted uncertain loss will be:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.863, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = m ^ {g} \\cdot \\frac {1}{B _ {u}} \\sum_ {k = 1} ^ {B _ {u}} \\mathbb {1} \\left(\\xi \\left(p _ {k} ^ {w}\\right) < \\tau\\right) \\cdot \\hat {\\mathrm {H}} \\left(\\hat {p} _ {k} ^ {w}, \\hat {p} _ {k} ^ {s}\\right) \\cdot \\xi \\left(p _ {k} ^ {w}\\right). \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.092, + 0.613, + 0.107 + ], + "angle": 0, + "content": "3.4. Summary" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.117, + 0.894, + 0.163 + ], + "angle": 0, + "content": "To summarize, the final loss in a mini-batch is a combination of the supervised loss \\((\\mathcal{L}_x)\\), certain loss \\((\\mathcal{L}_u\\), Eq. (1)), and uncertain loss in the shrunk class space \\((\\mathcal{L}_s\\), Eq. (10)):" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.18, + 0.893, + 0.196 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {x} + \\lambda_ {u} \\cdot \\left(\\mathcal {L} _ {u} + \\mathcal {L} _ {s}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.215, + 0.894, + 0.246 + ], + "angle": 0, + "content": "We do not carefully fine-tune the fusion weight between \\(\\mathcal{L}_u\\) and \\(\\mathcal{L}_s\\), but use 1:1 by default to avoid hyper-parameters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.263, + 0.625, + 0.277 + ], + "angle": 0, + "content": "3.5. Discussions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.895, + 0.424 + ], + "angle": 0, + "content": "Our uncertain loss in the shrunk space owns two properties: informative and also safe. The first property is because we manage to find the largest shrunk space that reaches the confidence threshold. Besides, we also adopt the weak-to-strong consistency regularization to pose a challenging optimization target. Both constraints ensure the learning in the shrunk space is not trivial and still quite informative. Next, we wish to explain the second property \"safe\", especially about how we avoid noise in the shrunk space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.895, + 0.623 + ], + "angle": 0, + "content": "Noises in pseudo labels distinguish the semi-supervised paradigm from the fully-supervised one. So designing a safe optimization target for unlabeled data is crucial. Typically, the cross entropy loss will maximize the softmax probability \\(\\exp (s_t) / \\sum_{i = 1}^{C}\\exp (s_i)\\rightarrow 1\\) for the target class \\(t\\). It inevitably suppresses scores of all other classes except \\(t\\). However, the true class could not be \\(t\\), and may be the \\(2^{\\mathrm{nd}}\\) or \\(3^{\\mathrm{rd}}\\) largest class, etc., which is wrongly restrained. This is frequently observed when the confidence of class \\(t\\) is not high enough. As a milder alternative, the soft labeling still encounters a similar problem. In contrast, our shrunk target directly excludes these confusion classes. So only the almost unlikely classes are suppressed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.894, + 0.777 + ], + "angle": 0, + "content": "It is worth noting, even if the predicted top-1 class is not in line with the human label, it is probably one of the closest semantics, e.g., belonging to the same superclass as the groundtruth label, as shown in Fig. 2. Thus, contrasting these relevant classes against other less likely classes with our noise-robust shrunk loss is still beneficial. The model is encouraged to make discriminative predictions closer to top classes compared to tail classes. In addition, our disentangled auxiliary head can effectively leverage such supervision while not affecting the main classification tasks [7]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.798, + 0.625, + 0.815 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In this section, we first describe the implementation details of our framework. Then, we compare our ShrinkMatch with previous state-of-the-art methods (SOTAs) under extensive evaluation protocols. Lastly, we conduct comprehensive ablation studies on each component to validate the necessity." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16191" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.187 + ], + "angle": 0, + "content": "
Seed01234Mean
SimMatch [45]95.3495.1692.6393.7695.1094.39
ShrinkMatch | 4095.0994.6695.1294.7894.9594.92
SimMatch [45]95.5895.5095.3494.0695.2695.15
ShrinkMatch | 25095.3995.4495.3694.7695.3595.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.196, + 0.47, + 0.242 + ], + "angle": 0, + "content": "Table 1: Comparison with SOTAs on CIFAR-10. The same seed ensures exactly the same data split. The 400 or 2500 denotes the number of labels." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.25, + 0.47, + 0.35 + ], + "angle": 0, + "content": "
Seed01234Mean
SimMatch [45]62.0660.1959.8964.8863.9262.19
ShrinkMatch | 40065.0063.4763.7766.4264.5264.64
SimMatch [45]74.6475.1974.5375.0375.2474.93
ShrinkMatch | 250075.0075.1174.5474.7874.7274.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.106, + 0.358, + 0.441, + 0.373 + ], + "angle": 0, + "content": "Table 2: Comparison with SOTAs on CIFAR-100." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.394, + 0.295, + 0.41 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.471, + 0.498 + ], + "angle": 0, + "content": "Baselines. We use FixMatch + distribution alignment (DA) as our baseline on all datasets except ImageNet and SVHN. On ImageNet, we build our method on SimMatch. On SVHN, we discard DA. To be more convincing, we adopt the same codebase as our compared methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.471, + 0.835 + ], + "angle": 0, + "content": "Hyper-parameters. Following prior arts, Wide ResNet-28-2 [40], WRN-28-8, WRN-37-2, and WRN-28-2 are used for CIFAR-10, CIFAR-100, STL-10, and SVHN respectively. A ResNet-50 [13] is used for ImageNet. The auxiliary head \\( h_{aux} \\) for uncertain samples is a 3-layer MLP. On the ImageNet, we set \\( B_u = 64 \\times 5 \\), but for other datasets, \\( B_u = 64 \\times 7 \\). The labeled batch size is 64 for all datasets. On the ImageNet, the model is trained for 400 epochs, while on the others, it is trained for \\( 2^{20} \\) iterations. The initial learning rate is set as 0.03 for all datasets with a cosine scheduler. Specifically, on the ImageNet, it is warmed up for 5 epochs. The consistency regularization \\( H \\) in \\( \\mathcal{L}_u \\) on STL-10 and SVHN is a hard cross entropy (CE) loss, while on the CIFAR-10/100 and ImageNet, following the SimMatch and CoMatch [20], it is modified to a soft CE loss. The \\( \\hat{H} \\) in our proposed \\( \\mathcal{L}_s \\) is simply a hard CE loss. The weight \\( \\lambda_u \\) for the two unsupervised losses is set as 10 on the ImageNet and 1 on others. The confidence threshold \\( \\tau \\) is 0.7 on the ImageNet and 0.95 for others. The shared momentum coefficient \\( \\gamma \\) between our global certain ratio \\( m^g \\) and the teacher parameters is 0.999. Following the common practice, the teacher model is only maintained for final evaluation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.32, + 0.862 + ], + "angle": 0, + "content": "4.2. CIFAR-10 and CIFAR-100" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "The CIFAR dataset is composed of 50000/10000 training/validation images of size \\(32 \\times 32\\). The CIFAR-10 con" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.089, + 0.888, + 0.209 + ], + "angle": 0, + "content": "
Seed012Mean
FixMatch [28]65.8567.9458.3064.03
FlexMatch [42]76.7168.2867.5570.85
ShrinkMatch | 4085.7585.6486.5585.98
FixMatch [28]90.9188.7190.9490.19
FlexMatch [42]91.3592.2991.6691.77
ShrinkMatch | 25091.1392.4391.1091.55
" + }, + { + "type": "table_caption", + "bbox": [ + 0.542, + 0.216, + 0.849, + 0.232 + ], + "angle": 0, + "content": "Table 3: Comparison with SOTAs on STL-10." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.242, + 0.888, + 0.362 + ], + "angle": 0, + "content": "
Seed012Mean
FixMatch [28]94.5396.9097.1496.19
FlexMatch [42]89.1989.9396.3291.81
ShrinkMatch | 4097.9697.8196.7097.49
FixMatch [28]98.0097.9997.9497.98
FlexMatch [42]91.7691.8396.6593.41
ShrinkMatch | 25098.0898.0697.9898.04
" + }, + { + "type": "table_caption", + "bbox": [ + 0.546, + 0.37, + 0.845, + 0.385 + ], + "angle": 0, + "content": "Table 4: Comparison with SOTAs on SVHN." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.409, + 0.895, + 0.56 + ], + "angle": 0, + "content": "tains 10 balanced classes to recognize, while CIFAR-10 is much more challenging, containing 100 classes. We run our ShrinkMatch on five different seeds and report each result as well as the average result. On the CIFAR-10 of Tab. 1 where performance is almost saturated, we still obtain a nontrivial improvement of \\(0.53\\%\\) (\\(94.39\\% \\to 94.92\\%\\)) with 4 labels per class (400 labels in total). In addition, notably, as shown in Tab. 2, with only 4 labels per class on the challenging CIFAR-100, our ShrinkMatch remarkably outperforms SimMatch by \\(2.45\\%\\) (\\(62.19\\% \\to 64.64\\%\\)) on average." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.576, + 0.684, + 0.59 + ], + "angle": 0, + "content": "4.3. STL-10 and SVHN" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.601, + 0.893, + 0.692 + ], + "angle": 0, + "content": "The STL-10 dataset originally consists of 5K/100K/8K labeled/unlabeled/validation images. We follow FlexMatch to only select a subset of 40/250 labeled images, while the unlabeled set remains unchanged. As shown in Tab. 3, with 4 labels per class, our ShrinkMatch surpasses FlexMatch tremendously from \\(70.85\\%\\) to \\(85.98\\%\\) \\((+15.13\\%)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.693, + 0.894, + 0.785 + ], + "angle": 0, + "content": "The SVHN dataset is suitable to reveal the ability of different semi-supervised algorithms in the presence of class imbalance issue. As demonstrated in Tab. 4, although the performance of the original FixMatch (our baseline) has almost touched the upper bound, we still further boost it from \\(96.19\\%\\) to \\(97.49\\%\\) \\((+1.30\\%)\\) with 4 labels per class." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.642, + 0.817 + ], + "angle": 0, + "content": "4.4. ImageNet-1K" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.893, + 0.903 + ], + "angle": 0, + "content": "The ImageNet dataset is rather challenging, containing 1.28M/50K training/validation images of 1000 classes. We exactly follow the codebase of SimMatch [45]. As shown in Tab. 5, our ShrinkMatch further boosts previous best results on both settings of \\(1\\%\\) and \\(10\\%\\) labeled images." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16192" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.089, + 0.851, + 0.272 + ], + "angle": 0, + "content": "
Pre-trainingMethodEpochsParams (train / test)Top-1Top-5
1%10%1%10%
SimCLR v2 [8]80034.2M / 25.6M57.968.482.589.2
SwAV [6]Fine-tune80030.4M / 25.6M53.970.278.589.9
WCL [44]80034.2M / 25.6M65.072.086.391.2
MoCo v2 [9]Fine-tune80030.0M / 25.6M49.866.177.287.9
CoMatch [20]120030.0M / 25.6M67.173.787.191.4
MoCo-EMAN [5]FixMatch-EMAN [5]110030.0M / 25.6M63.074.083.490.9
NoneCoMatch [20]40030.0M / 25.6M66.073.686.491.6
SimMatch† [45]40030.0M / 25.6M67.074.186.991.5
ShrinkMatch40031.8M / 25.6M67.574.587.491.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.107, + 0.278, + 0.86, + 0.294 + ], + "angle": 0, + "content": "Table 5: Accuracy (%) on the ImageNet-1K with \\( 1\\% \\) and \\( {10}\\% \\) labeled images. †: Reproduced in our environment." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.307, + 0.344, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.444, + 0.34, + 0.458 + ], + "angle": 0, + "content": "(a) Number of removed confusion classes." + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.306, + 0.611, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.359, + 0.445, + 0.606, + 0.458 + ], + "angle": 0, + "content": "(b) Certain loss \\(\\mathcal{L}_u\\) and uncertain loss \\(\\mathcal{L}_s\\)." + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.308, + 0.88, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.445, + 0.842, + 0.458 + ], + "angle": 0, + "content": "(c) Ratio of uncertain samples." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.464, + 0.893, + 0.495 + ], + "angle": 0, + "content": "Figure 4: (a) The number of removed confusion classes for each uncertain image as the training proceeds. (b) Value of different loss functions, i.e., Eq. (1), Eq. (6), and Eq. (10). (c) The ratio of uncertain samples (certainty \\(< \\tau\\)) in each mini-batch." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.511, + 0.465, + 0.569 + ], + "angle": 0, + "content": "
Seed01234Mean
Baseline64.4057.5163.0764.7762.5362.46
ShrinkMatch65.0063.4763.7766.4264.5264.64
" + }, + { + "type": "table_caption", + "bbox": [ + 0.081, + 0.576, + 0.465, + 0.591 + ], + "angle": 0, + "content": "Table 6: Ablation studies of ShrinkMatch on CIFAR-100." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.612, + 0.241, + 0.627 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.636, + 0.471, + 0.666 + ], + "angle": 0, + "content": "Unless otherwise specified, we conduct our ablation studies on CIFAR-100 with 4 labels per class." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.471, + 0.761 + ], + "angle": 0, + "content": "Effectiveness of our holistic ShrinkMatch. We first view our ShrinkMatch as a holistic component added to our baseline. As displayed in Tab. 6 for CIFAR-100, our ShrinkMatch boosts the strong baseline significantly by \\(2.18\\%\\) (\\(62.46\\% \\rightarrow 64.64\\%\\)). And on the STL-10 of Tab. 7, the baseline is also improved evidently from \\(84.85\\%\\) to \\(85.98\\%\\) (\\(+1.13\\%\\))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of two reweighting principles. In Sec. 3.3, considering the varied prediction reliability of different uncertain samples and different training stages, we propose two principles to reweight our uncertain loss \\(\\mathcal{L}_s\\). So here we carefully examine their necessity in Tab. 8. It can be observed that the two principles are both beneficial, jointly improving the basic shrinking practice from \\(63.16\\%\\) to \\(64.24\\%\\). We also attempt a simple alternative of linearly increasing the uncertain loss weight from 0 (start) to 1 (end). But as evi" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.511, + 0.875, + 0.57 + ], + "angle": 0, + "content": "
Seed012Mean
Baseline84.8684.2185.4784.85
ShrinkMatch85.7585.6486.5585.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.517, + 0.576, + 0.875, + 0.591 + ], + "angle": 0, + "content": "Table 7: Ablation studies of ShrinkMatch on STL-10." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.616, + 0.892, + 0.738 + ], + "angle": 0, + "content": "denced in the Exp 3 & 4 of Tab. 8, it is obviously inferior to our designed reweighting strategy. We visualize our \\( m^g \\) as training proceeds on ImageNet in Fig. 5. It looks essentially different from linear scheduling. Moreover, we visualize the uncertain loss value with or without the second reweighting principle in Fig. 4b. After reweighting, the uncertain loss is of a similar magnitude as the certain loss, which can avoid dominating the gradient at early training stages." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.743, + 0.894, + 0.88 + ], + "angle": 0, + "content": "The number of removed confusion classes. As introduced before, our approach detects and removes the confusion classes for the top-1 class in a fully automatic and instance-adaptive fashion. So we visualize the number of removed confusion classes for each uncertain sample in Fig. 4a. At the very start of training, model predictions are almost uniform, so abundant classes are removed for an uncertain sample. But as the training proceeds, much fewer confusion classes need to be removed to form a confident shrunk class space." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Hard label or soft label in the shrunk class space? By" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16193" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.472, + 0.189 + ], + "angle": 0, + "content": "
ExpPrinciple 1Principle 2LSSeed 0Seed 1Mean
165.6260.7063.16
2✓65.6862.0863.88
3✓65.0962.9064.00
4✓63.0159.4561.23
5✓✓65.0063.4764.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.197, + 0.472, + 0.258 + ], + "angle": 0, + "content": "Table 8: Ablation studies on the effectiveness of the two reweighting principles. Due to randomness, the \"Mean\" results are more convincing. LS is short for linear scheduling, which linearly increases the loss weight." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.266, + 0.466, + 0.445 + ], + "angle": 0, + "content": "
ExpShrinkAux HeadLabelSeed 0Seed 1Mean
1Our Strong Baseline64.4057.5160.96
2✓✓H65.0063.4764.24
3✓✓S62.9259.7661.34
4✓H62.4358.8660.65
5✓S60.9260.2260.57
6✓H64.2159.1761.69
7✓S64.7560.8962.82
8H52.9051.8852.39
9S54.1055.7554.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.453, + 0.472, + 0.545 + ], + "angle": 0, + "content": "Table 9: Ablation studies on different options to learn uncertain sample. (1) whether to shrink the class space (Shrink), (2) whether to use an auxiliary head (Aux Head), and (3) use hard (H) or soft label (S). Exp 1 is our baseline without using uncertain images. Exp 8 & 9 are experiments of directly learning uncertain samples, which are quite noisy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.47, + 0.656 + ], + "angle": 0, + "content": "default, we use the hard label in the shrunk space, following FixMatch. But we also ablate the choice of using the soft label as supervision. As shown in Exp 2 & 3 of Tab. 9, the hard label is much better than the soft label. We conjecture that this is because we have shrunk the class space to a safe one. The soft label will incur some unnecessary noise." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Effects of the auxiliary head. To prevent the main classifier from being affected, we use an auxiliary head to learn the uncertain samples in the shrunk class space only for discriminative representations. So we first provide the results when still using the main head for the shrunk space. As shown in Exp 2 & 4 of Tab. 9, the auxiliary head is indispensable for our shrinking practice. So we continue to figure out whether the improvement of our ShrinkMatch merely comes from the auxiliary head. We attempt to use the auxiliary head to directly learn the uncertain samples in the original space without shrinking. It can be concluded from Exp 2 & 6 & 7 of Tab. 9, the original space is indeed much inferior to our shrunk space for uncertain samples, although the auxiliary head still helps the baseline to some extent. Lastly, in Tab. 9, we also include two additional results for readers, i.e., directly learning uncertain samples without shrinking or" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.089, + 0.885, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.232, + 0.874, + 0.248 + ], + "angle": 0, + "content": "Figure 5: Visualization of our global certain ratio \\( m^g \\)." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.254, + 0.885, + 0.299 + ], + "angle": 0, + "content": "
Threshold0.980.950.900.800.70
ShrinkMatch65.8065.0064.5562.7459.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.306, + 0.877, + 0.321 + ], + "angle": 0, + "content": "Table 10: Ablation studies of the confidence threshold." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.341, + 0.892, + 0.371 + ], + "angle": 0, + "content": "auxiliary head (Exp 8 & 9), which are quite terrible because of introducing the abundant noise to our main classifier." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.375, + 0.895, + 0.481 + ], + "angle": 0, + "content": "The ratio of uncertain samples. To further justify our motivation, in Fig. 4c, we display the ratio of uncertain samples in each mini-batch. It can be seen that even in the middle of the whole training course, there are still around \\(40\\%\\) and \\(30\\%\\) uncertain samples on ImageNet and CIFAR-100 respectively. Therefore, an appropriate approach to utilizing these samples is necessary and definitely beneficial." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.484, + 0.895, + 0.575 + ], + "angle": 0, + "content": "Different confidence thresholds \\(\\tau\\). We also try different thresholds for uncertain samples, as shown in Tab. 10. When increasing \\(\\tau\\) from widely adopted 0.95 to 0.98 (more uncertain samples), our ShrinkMatch can be further improved. These results clearly highlight the effective and safe utilization of uncertain samples with our ShrinkMatch." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.598, + 0.619, + 0.613 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.895, + 0.82 + ], + "angle": 0, + "content": "In this work, we aim to fully leverage the uncertain samples in semi-supervised learning. We point out that the low certainty is typically caused by a small portion of confusion classes. Invoked by this, we propose a novel method dubbed ShrinkMatch to automatically detect and remove the confusion classes to construct a shrunk class space, where the top-1 class is turned certain. A weak-to-strong consistency regularization is enforced in the confident new space. Furthermore, we design two reweighting principles for the auxiliary uncertain loss, according to the reliability of different uncertain samples and the gradually improved state of the model. Consequently, our method establishes new state-of-the-art results on widely acknowledged benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.837, + 0.894, + 0.898 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported in part by the National Natural Science Foundation of China (62201484, 62222604, 62192783, 62206052), HKU Startup Fund, and HKU Seed Fund for Basic Research." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16194" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Eric Arazo, Diego Ortego, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In IJCNN, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.227, + 0.471, + 0.282 + ], + "angle": 0, + "content": "[3] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.283, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[4] Zhaowei Cai, Avinash Ravichandran, Paolo Favaro, Manchen Wang, Davide Modolo, Rahul Bhotika, Zhuowen Tu, and Stefano Soatto. Semi-supervised vision transformers at scale. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.339, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[5] Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Charless Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi-supervised learning. In CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.471, + 0.45 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.451, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[7] Baixu Chen, Junguang Jiang, Ximei Wang, Jianmin Wang, and Mingsheng Long. Debiased pseudo labeling in self-training. In NeurIPS, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.471, + 0.536 + ], + "angle": 0, + "content": "[8] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. In NeurIPS, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.471, + 0.577 + ], + "angle": 0, + "content": "[9] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv:2003.04297, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.578, + 0.471, + 0.634 + ], + "angle": 0, + "content": "[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.634, + 0.471, + 0.676 + ], + "angle": 0, + "content": "[11] Yue Fan, Dengxin Dai, Anna Kukleva, and Bernt Schiele. Cossl: Co-learning of representation and classifier for imbalanced semi-supervised learning. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[12] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In ICML, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.471, + 0.802 + ], + "angle": 0, + "content": "[14] Zijian Hu, Zhengyu Yang, Xuefeng Hu, and Ram Nevatia. Simple: similar pseudo label exploitation for semi-supervised classification. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[15] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? In NeurIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[16] Jiwon Kim, Youngjo Min, Daehwan Kim, Gyuseong Lee, Junyoung Seo, Kwangrok Ryoo, and Seungryong Kim. Conmatch: Semi-supervised learning with confidence-guided consistency regularization. In ECCV, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "[17] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.15 + ], + "angle": 0, + "content": "[18] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In ICLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.151, + 0.894, + 0.193 + ], + "angle": 0, + "content": "[19] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In ICML Workshop, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.194, + 0.894, + 0.237 + ], + "angle": 0, + "content": "[20] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regularization. In ICCV, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.238, + 0.894, + 0.294 + ], + "angle": 0, + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.296, + 0.892, + 0.339 + ], + "angle": 0, + "content": "[22] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. TPAMI, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.34, + 0.894, + 0.396 + ], + "angle": 0, + "content": "[23] Islam Nassar, Samitha Herath, Ehsan Abbasnejad, Wray Buntine, and Gholamreza Haffari. All labels are not created equal: Enhancing semi-supervision via label grouping and co-training. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.397, + 0.894, + 0.439 + ], + "angle": 0, + "content": "[24] Yassine Ouali, Céline Hudelot, and Myriam Tami. An overview of deep semi-supervised learning. arXiv:2006.05278, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.44, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[25] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V Le. Meta pseudo labels. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.894, + 0.527 + ], + "angle": 0, + "content": "[26] Mamshad Nayeem Rizve, Kevin Duarte, Yogesh S Rawat, and Mubarak Shah. In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.528, + 0.894, + 0.571 + ], + "angle": 0, + "content": "[27] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In NeurIPS, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.571, + 0.894, + 0.641 + ], + "angle": 0, + "content": "[28] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.642, + 0.894, + 0.698 + ], + "angle": 0, + "content": "[29] Fariborz Taherkhani, Ali Dabouei, Sobhan Soleymani, Jeremy Dawson, and Nasser M Nasrabadi. Self-supervised Wasserstein pseudo-labeling for semi-supervised image classification. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.7, + 0.894, + 0.742 + ], + "angle": 0, + "content": "[30] Hui Tang and Kui Jia. Towards discovering the effectiveness of moderately confident samples for semi-supervised learning. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.743, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[31] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.894, + 0.857 + ], + "angle": 0, + "content": "[32] Guotai Wang, Wenqi Li, Michael Aertsen, Jan Deprest, Sébastien Ourselin, and Tom Vercauteren. Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolutional neural networks. Neurocomputing, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.858, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[33] Xudong Wang, Long Lian, and Stella X Yu. Unsupervised selective labeling for more effective semi-supervised learning. In ECCV, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16195" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.134 + ], + "angle": 0, + "content": "[34] Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V Le. Unsupervised data augmentation for consistency training. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.176 + ], + "angle": 0, + "content": "[35] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.472, + 0.22 + ], + "angle": 0, + "content": "[36] Yi Xu, Lei Shang, Jinxing Ye, Qi Qian, Yu-Feng Li, Baigui Sun, Hao Li, and Rong Jin. Dash: Semi-supervised learning with dynamic thresholding. In ICML, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.472, + 0.275 + ], + "angle": 0, + "content": "[37] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.472, + 0.305 + ], + "angle": 0, + "content": "[38] David Yarowsky. Unsupervised word sense disambiguation, rivaling supervised methods. In ACL, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.472, + 0.36 + ], + "angle": 0, + "content": "[39] Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In MICCAI, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.472, + 0.39 + ], + "angle": 0, + "content": "[40] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.472, + 0.433 + ], + "angle": 0, + "content": "[41] Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.472, + 0.49 + ], + "angle": 0, + "content": "[42] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In NeurIPS, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.492, + 0.472, + 0.533 + ], + "angle": 0, + "content": "[43] Zhen Zhao, Luping Zhou, Lei Wang, Yinghuan Shi, and Yang Gao. Lassl: Label-guided self-training for semi-supervised learning. In AAAI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.535, + 0.472, + 0.576 + ], + "angle": 0, + "content": "[44] Mingkai Zheng, Fei Wang, Shan You, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Weakly supervised contrastive learning. In ICCV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.578, + 0.472, + 0.619 + ], + "angle": 0, + "content": "[45] Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In CVPR, 2022. 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16196" + } + ] +] \ No newline at end of file diff --git a/2023/Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning/34917a9e-765e-47a5-bd3d-8044ab9de911_origin.pdf b/2023/Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning/34917a9e-765e-47a5-bd3d-8044ab9de911_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ed2ed919fab71b678b925e8224feebceb62bb0ee --- /dev/null +++ b/2023/Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning/34917a9e-765e-47a5-bd3d-8044ab9de911_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f6d57768d5c8e58f97da60b0dfed1f501bf748eed3096b1814e2fa358f4ba53 +size 675967 diff --git a/2023/Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning/full.md b/2023/Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4cf6573e164b65fe71a9432825018ad8e6daea2d --- /dev/null +++ b/2023/Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning/full.md @@ -0,0 +1,333 @@ +# Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning + +Lihe Yang $^{1}$ Zhen Zhao $^{4}$ Lei Qi $^{5}$ Yu Qiao $^{3}$ Yinghuan Shi $^{2*}$ Hengshuang Zhao $^{1*}$ + +1The University of Hong Kong 2Nanjing University + +3Shanghai AI Laboratory 4The University of Sydney 5Southeast University + +https://github.com/LiheYoung/ShrinkMatch + +# Abstract + +Semi-supervised learning is attracting blooming attention, due to its success in combining unlabeled data. To mitigate potentially incorrect pseudo labels, recent frameworks mostly set a fixed confidence threshold to discard uncertain samples. This practice ensures high-quality pseudo labels, but incurs a relatively low utilization of the whole unlabeled set. In this work, our key insight is that these uncertain samples can be turned into certain ones, as long as the confusion classes for the top-1 class are detected and removed. Invoked by this, we propose a novel method dubbed ShrinkMatch to learn uncertain samples. For each uncertain sample, it adaptively seeks a shrunk class space, which merely contains the original top-1 class, as well as remaining less likely classes. Since the confusion ones are removed in this space, the re-calculated top-1 confidence can satisfy the pre-defined threshold. We then impose a consistency regularization between a pair of strongly and weakly augmented samples in the shrunk space to strive for discriminative representations. Furthermore, considering the varied reliability among uncertain samples and the gradually improved model during training, we correspondingly design two reweighting principles for our uncertain loss. Our method exhibits impressive performance on widely adopted benchmarks. + +# 1. Introduction + +In the last decade, our computer vision community has witnessed inspiring progress, thanks to large-scale datasets [21, 10]. Nevertheless, it is laborious and costly to annotate massive images, hindering the progress to benefit a broader range of real-world scenarios. Inspired by this, semi-supervised learning (SSL) was proposed to utilize the unlabeled data under the assistance of limited labeled data. + +The frameworks in SSL are typically based on the strategy of pseudo labeling. Briefly, the model acquires knowledge from the labeled data, and then assigns predictions on the + +![](images/ae66fb03b8076858a055ecce8796fd1d19d44bae77fc18af83490942b6da113e.jpg) +Figure 1: Illustration of our motivation. Due to confusion classes for the top-1 class, the certainty fails to reach the pre-defined threshold (gray dotted line). FixMatch discards such uncertain samples. Our method, however, detects and removes confusion classes to enhance certainty, then enjoying full and safe utilization of all unlabeled images. + +unlabeled data. The two sources of data are finally combined to train a better model. During this process, it is obvious that predictions on unlabeled data are not reliable. If the model is iteratively trained with incorrect pseudo labels, it will suffer the confirmation bias issue [1]. To address this dilemma, recent works [28] simply set a fixed confidence threshold to discard potentially unreliable samples. This simple strategy effectively retains high-quality pseudo labels, however, it also incurs a low utilization of the whole unlabeled set. As evidenced by our pilot study on CIFAR-100 [17], nearly $20\%$ unlabeled images are filtered out for not satisfying the threshold of 0.95. Instead of blindly throwing them away, we believe there should exist a more promising approach. This work is just aimed to fully leverage previously uncertain samples in an informative but also safe manner. + +So first, why are these samples uncertain? According to our observations on CIFAR-100 and ImageNet, although the top-1 accuracy could be low, the top-5 accuracy is much higher. This indicates in most cases, the model struggles to discriminate among a small portion of classes. As illustrated + +![](images/86eafff981b9f29c8572753a51af82349256cd27529a4ac7959752ef39fbf053.jpg) +Figure 2: Pseudo label accuracy on CIFAR-100 with 400 labels. We highlight that even for uncertain samples, their top-1 predictions are of high accuracy in the superclass space (20 classes). This accuracy can even be comparable to the delicately selected certain samples in the original class space. + +in Fig. 1, given a cat image, the model is not sure whether it belongs to tabby cat, tiger cat, or other cats. On the other hand, however, it is absolutely certain that the object is more like a tabby cat, rather than a table or anything else. In other words, it is reliable for the model to distinguish the top class from the remaining less likely classes. + +Invoked by these, we propose a novel method dubbed ShrinkMatch to learn uncertain samples. The prediction fails to satisfy the pre-defined threshold due to the existence of confusion classes for the top-1 class. Hence, our approach adaptively seeks a shrunk class space where the confusion classes are removed, to enable the re-calculated confidence to reach the original threshold. Moreover, the obtained shrunk class space is also required to be the largest among those that can satisfy the threshold. In a word, we seek a certain and largest shrunk space for uncertain samples. Then, logits of the strongly augmented image are correspondingly gathered in the new space. And a consistency regularization is imposed between the shrunk weak-strong predictions. + +Note that, the confusion classes are detected and removed in a fully automatic and instance-adaptive fashion. Moreover, even if the predicted top-1 class is not fully in line with the groundtruth label, they mostly belong to the same superclass. To prove this, as shown in Fig. 2, uncertain samples exhibit much higher pseudo label accuracy in the superclass space than the original space. This accuracy is even comparable to that of certain samples in the original class space. Therefore, contrasting these groundtruth-related classes against remaining unlikely classes is still highly beneficial, yielding more discriminative representations of our model. To avoid affecting the main classifier, we further adopt an auxiliary classifier to disentangle the learning in the shrunk space. + +Despite the effectiveness, there still exist two main drawbacks to the above optimization target. (1) First, it treats all uncertain samples equally. The truth, however, is that in the original class space, the top-1 confidence of different uncertain samples can vary dramatically. And it is clear + +that samples with larger confidence should be attached more importance. To this end, we propose to balance different uncertain samples by their confidence in the original space. (2) Moreover, the regularization term also overlooks the gradually improved model state during training. At the start of training, there are abundant uncertain samples, but their predictions are extremely noisy, or even random. So even the highest scored class may share no relationship with the true class. Then as the training proceeds, the top classes become reliable. Considering this, we further propose to adaptively reweight the uncertain loss according to the model state. The model state is tracked and approximately estimated via performing exponential moving average on the proportion of certain samples in each mini-batch. With the two reweighting principles, the model turns out more stable, and avoids accumulating much noise from uncertain samples, especially at early training iterations. + +To summarize, our contributions lie in three aspects: + +- We first point out that low certainty is typically caused by a small portion of confusion classes. To enhance the certainty, we propose to shrink the original class space by adaptively detecting and removing confusion ones for the top-1 class to turn it certain in the new space. +- We manage to reweight the uncertain loss from two perspectives: the image-based varied reliability among different uncertain samples, and the model-based gradually improved state as the training proceeds. +- Our proposed ShrinkMatch establishes new state-of-the-art results on widely acknowledged benchmarks. + +# 2. Related Work + +Semi-supervised learning (SSL). The primary concern in SSL [19, 28, 36, 20, 25, 16, 37, 35, 41, 43, 14, 24, 29, 23, 4, 30, 11, 33, 2, 42] is to design effective constraints for unlabeled samples. Dating back to decades ago, pioneering works [19, 38] integrate unlabeled data via assigning pseudo labels to them, with the knowledge acquired from labeled data. In the era of deep learning, subsequent methods mainly follow such a bootstrapping fashion, but greatly boost it with some key components. Specifically, to enhance the quality of pseudo labels, $\Pi$ -model [18] and Mean Teacher [31] ensemble model predictions and model parameters respectively. Later works start to exploit the role of perturbations. During this trend, [27] proposes to apply stochastic perturbations on inputs or features, and enforce consistency across these predictions. Then, UDA [34] emphasizes the necessity of strengthening the perturbation pool. It also follows VAT [22] and MixMatch [3] to supervise the prediction under strong perturbations with that under weak perturbations. Since then, weak-to-strong consistency regularization has become a standard practice in SSL. Eventually, the milestone + +work FixMatch [28] presents a simplified framework using a fixed confidence threshold to discard uncertain samples. Our ShrinkMatch is built upon FixMatch. But, we highlight the value of previously neglected uncertain samples, and leverage them in an informative but also safe manner. + +More recently, DST [7] decouples the generation and utilization of pseudo labels with a main and an auxiliary head respectively. Besides, SimMatch [45] explores instance-level relationships with labeled embeddings to supplement original class-level relations. Compared with them, our ShrinkMatch achieves larger improvements. + +Defining uncertain samples. Earlier works estimate the uncertainty with Bayesian Neural Networks [15], or its faster approximation, e.g., Monte Carlo Dropout [12]. Some other works measure the prediction disagreement among multiple randomly augmented inputs [32]. The latest trend is to directly use the entropy of predictions [39], cross entropy [36], or softmax confidence [28] as a measurement for uncertainty. Our work is not aimed at the optimal uncertainty estimation strategy, so we adopt the simplest solution from FixMatch, i.e., using the maximum softmax output as the certainty. + +Utilizing uncertain samples. UPS [26] leverages negative class labels whose confidence is below a pre-defined threshold, from a reversed multi-label classification perspective. In comparison, our model is enforced to tell the most likely class without being cheated by the less likely ones. So our supervision on uncertain samples is more informative and produces more discriminative representations. Moreover, we do not introduce any extra hyper-parameters, e.g., the lower threshold in [26], into our framework. + +# 3. Method + +We primarily provide some notations and review a common practice in semi-supervised learning (SSL) (Sec. 3.1). Next, we present our ShrinkMatch in detail (Sec. 3.2 and Sec. 3.3). Finally, we summarize our approach and provide a further discussion (Sec. 3.4 and Sec. 3.5). + +# 3.1. Preliminaries + +Semi-supervised learning aims to learn a model with limited labeled images $\mathcal{D}^l = \{(x_k, y_k)\}$ , aided by a large number of unlabeled images $D^u = \{u_k\}$ . Recent frameworks commonly follow the FixMatch practice. Concretely, an unlabeled image $u$ is first transformed by a weak augmentation pool $\mathcal{A}^w$ and a strong augmentation pool $\mathcal{A}^s$ to yield a pair of weakly and strongly augmented images $(u^w, u^s)$ . Then, they are fed into the model together to produce corresponding predictions $(p^w, p^s)$ . Typically, $p^w$ is of much higher accuracy than $p^s$ , while $p^s$ is beneficial to learning. Therefore, $p^w$ serves as the pseudo label for $p^s$ . Moreover, a core practice introduced by FixMatch is that, to improve the quality of selected pseudo labels, a fixed confidence threshold is + +pre-defined to abandon uncertain ones. So the unsupervised loss $\mathcal{L}_u$ can be formulated as: + +$$ +\mathcal {L} _ {u} = \frac {1}{B _ {u}} \sum_ {k = 1} ^ {B _ {u}} \mathbb {1} \left(\xi \left(p _ {k} ^ {w}\right) \geq \tau\right) \cdot \mathrm {H} \left(p _ {k} ^ {w}, p _ {k} ^ {s}\right), \tag {1} +$$ + +where $B^{u}$ is the batch size of unlabeled images and $\tau$ is the pre-defined threshold. $\xi(p_{k}^{w})$ computes the confidence of logits $p_{k}^{w}$ by $\xi(\cdot) = \max(\sigma(\cdot))$ , where $\sigma$ is the softmax function. The H denotes the consistency regularization between the two distributions. It is typically the cross entropy loss. + +In addition, the labeled images are learned with a regular cross entropy loss to obtain the supervised loss $\mathcal{L}_x$ . The overall loss in each mini-batch then will be: + +$$ +\mathcal {L} = \mathcal {L} _ {x} + \lambda_ {u} \cdot \mathcal {L} _ {u}, \tag {2} +$$ + +where $\lambda_{u}$ acts as a trade-off term between the two losses. + +# 3.2. Shrinking the Class Space for Certainty + +Our motivation. As reviewed above, FixMatch discards the samples whose confidence is lower than a pre-defined threshold, because their pseudo labels are empirically found relatively noisier. These samples are named uncertain samples in this work. Although this practice retains high-quality pseudo labels, it incurs a low utilization of the whole unlabeled set, especially when the scenario is challenging and the selection criterion is strict. Take the CIFAR-100 dataset as an instance, with a common threshold of 0.95 and 4 labels per class, there will be nearly $20\%$ unlabeled samples being ignored due to their low certainty. We argue that, these uncertain samples can still benefit the model optimization, as long as we can design appropriate constraints (loss functions) on them. So we first investigate the cause of low certainty to gain some better intuitions. + +The reason for the low certainty of an unlabeled image is that, the model tends to be confused among some top classes. For example, given a cat image, the score of the class tabby cat and class tiger cat may be both high, so the model is not absolutely certain what the concrete class is. Motivated by this observation, we propose to shrink the original class space via adaptively detecting and removing the confusion classes for the top-1 class. Then the shrunk space is only composed of the original top-1 class, as well as the remaining less likely classes. After this process, re-calculated confidence of the top-1 class will satisfy the pre-defined threshold. Thereby, we can enforce the model to learn the previously uncertain samples in this new certain space, as shown in Fig. 3. Following the previous cat example, if the top-1 class is tabby cat, our method will scan scores of all other classes, and remove confusion ones (e.g., tiger cat and siamese cat) to construct a confident shrunk space, where the model is sure that the image is a tabby cat, rather than a table. Since we do not ask the model to discriminate among several top classes, it will avoid suffering from the noise when it makes a wrong judgment in the original space. + +![](images/03d015f6bb40c84a03b58c2f1731515a2b38f84620b092cbcfa307c979e11d3d.jpg) +Figure 3: An overview of our proposed ShrinkMatch. Our motivation is to fully leverage the originally uncertain samples. "S" denotes shrinking the class space. The confusion classes for the top-1 class are detected and removed in a fully automatic and instance-adaptive fashion, to construct a shrunk space where the top-1 class is turned certain. $\mathcal{L}_u$ is the original certain loss, while $\mathcal{L}_s$ calculates the uncertain loss in the shrunk class space. We add an auxiliary head $h_{aux}$ to learn in the new space. On the right, we further reweight $\mathcal{L}_s$ based on two principles. Principle 1 (image-based): image predictions with larger reliability are attached more importance. Principle 2 (model-based): We track the model state during training for reweighting. + +![](images/bf167c4db3d499a9caeb5bf1f0765d7fd89760688be2bd5182872fdf52d228cd.jpg) + +How to seek the shrunk class space? Now, how can we enable our method to automatically seek the optimal shrunk space of an uncertain unlabeled image? Ideally, we hope this seeking process is free from any prior knowledge from humans, e.g., class relationships, and also does not require any extra hyper-parameters. Considering these, we opt to inherit the pre-defined confidence threshold as a criterion. To be specific, we first sort the predicted logits in a descending order to obtain $p^w = \{s_{n_i}^w\}_{i=1}^C$ for classes $\{n_i\}_{i=1}^C$ , where $s_{n_i}^w \geq s_{n_i+1}^w$ . In the shrunk space, we will retain the original top-1 class, because it is still the most likely one to be true. Then, we find a set of less likely classes $\{n_i\}_{i=K}^C$ by enforcing two constraints on the $K$ : + +$$ +\xi \left(\left\{s _ {n _ {1}} ^ {w} \right\} \cup \left\{s _ {n _ {i}} ^ {w} \right\} _ {i = K} ^ {C}\right) \geq \tau , \tag {3} +$$ + +$$ +\xi \left(\left\{s _ {n _ {1}} ^ {w} \right\} \cup \left\{s _ {n _ {i}} ^ {w} \right\} _ {i = K - 1} ^ {C}\right) < \tau , \tag {4} +$$ + +where $\xi$ is defined the same as that in Eq. (1), calculating the confidence of the re-assembled logits. The final shrunk space is composed of the re-assembled classes $\{n_1\} \cup \{n_i\}_{i = K}^C$ . The two constraints on $K$ not only ensure the top-1 class is turned certain in the new space (Eq. (3)), but also select the largest space among all candidates (Eq. (4)). + +How to learn in the shrunk class space? For a weakly augmented uncertain image $x^w$ , the model is certain about its top-1 class in the shrunk class space. To learn effectively in this space, we follow the popular practice of weak-to-strong consistency regularization. The correspondingly shrunk prediction on the strongly augmented image is enforced to match that on the weakly augmented one. Concretely, the re-assembled logits from $p^w$ is denoted as $\hat{p}^w$ , which means $\hat{p}^w = \{s_{n_1}^w\} \cup \{s_{n_i}^w\}_K^C$ . We use the re-assembled classes $\{n_1\} \cup \{n_i\}_K^C$ in the shrunk space to correspondingly gather the logits $p^s$ on $x^s$ , yielding $\hat{p}^s = \{s_{n_1}^s\} \cup \{s_{n_i}^s\}_K^C$ . Then we can regularize the consistency between the two shrunk + +distributions $\hat{p}^s$ and $\hat{p}^w$ , similar to that in Eq. (1): + +$$ +\mathcal {L} _ {s} = \frac {1}{B _ {u}} \sum_ {k = 1} ^ {B _ {u}} \mathbb {1} \left(\xi \left(p _ {k} ^ {w}\right) < \tau\right) \cdot \hat {\mathrm {H}} \left(\hat {p} _ {k} ^ {w}, \hat {p} _ {k} ^ {s}\right), \tag {5} +$$ + +where the indicator function is to find uncertain samples. + +Empirically, we observe that if we use the original linear head $h_{main}$ to learn this auxiliary supervision, it will make the confidence of our model increase aggressively. Most noisy unlabeled samples are blindly judged as certain ones. We conjecture that it is because the $\mathcal{L}_s$ strengthens weights of the classes that are frequently uncertain, then these classes will be incorrectly turned certain. With this in mind, our solution is simple. We adopt an auxiliary MLP head $h_{aux}$ that shares the backbone with $h_{main}$ , to deal with this auxiliary optimization target, as shown in Fig. 3. So the $\hat{p}^s$ is indeed gathered from predictions of $h_{aux}$ ( $\hat{p}^w$ is still from $h_{main}$ ). This modification enables our feature extractor to acquire more discriminative representations, and meantime does not affect predictions of the main head. Note that $h_{aux}$ is only applied for training, bringing no burden to the test stage. + +# 3.3. Reweighting the Uncertain Loss + +Despite the effectiveness of the above uncertain loss, there still exist two main drawbacks. (1) On one hand, it overlooks the varied reliability of the top-1 class among different uncertain images. For example, suppose two uncertain images $u_{1}$ and $u_{2}$ with softmax predictions [0.8, 0.1, 0.1] and [0.5, 0.3, 0.2] in the original space, they should not be treated equally in the shrunk space. The $u_{1}$ with top-1 confidence of 0.8 is more likely to be true than $u_{2}$ , and thereby should be attached more attention to. (2) On the other, it ignores the gradually improved model performance as the training proceeds. To be concrete, at the very start of training, the predictions are extremely noisy or even random. And then + +at later stages, the predictions become more and more reliable. So the uncertain predictions at different training stages should not be treated equally. Therefore, we further design two reweighting principles for the two concerns. + +Principle 1: Reweighting with image-based varied reliability. According to the above intuition, we directly reweight the uncertain loss of each uncertain image by its top-1 confidence $\xi(p_k^w)$ in the original class space, which is: + +$$ +\mathcal {L} _ {s} ^ {1} = \frac {1}{B _ {u}} \sum_ {k = 1} ^ {B _ {u}} \mathbb {1} \left(\xi \left(p _ {k} ^ {w}\right) < \tau\right) \cdot \hat {\mathrm {H}} \left(\hat {p} _ {k} ^ {w}, \hat {p} _ {k} ^ {s}\right) \cdot \xi \left(p _ {k} ^ {w}\right). \tag {6} +$$ + +We do not use the top-1 confidence $\xi(\hat{p}_k^w)$ in the shrunk space as the weight, because after shrinking, this value of different predictions is very close to each other. So generally, $\xi(p_k^w)$ is more discriminative than $\xi(\hat{p}_k^w)$ as the weight. + +Principle 2: Reweighting with model-based gradually improved state. One naive solution is to linearly increase the loss weight of $\mathcal{L}_s$ from 0 at the beginning to $\mu$ at iteration $T$ , and then keep $\mu$ until the end. However, this practice has two severe disadvantages. First, the two additional hyperparameters $\mu$ and $T$ are not easy to determine, and could be sensitive. More importantly, the linear scheduling criterion simply assumes the model state also improves linearly. Indeed, it can not be true. Thus, we here present a more promising principle to perform reweighting, that is free from any extra hyper-parameters. It can adjust the loss weight according to the model state in a fully adaptive fashion. To be specific, we use the certain ratio of the unlabeled set as an indicator for the model state. The certain ratio is traced at each iteration and accumulated globally in an exponential moving average (EMA) manner. Formally, the certain ratio in a single mini-batch is given by: + +$$ +m = \frac {1}{B _ {u}} \sum_ {k = 1} ^ {B _ {u}} \mathbb {1} \left(\xi \left(p _ {k} ^ {w}\right) \geq \tau\right). \tag {7} +$$ + +The global certain ratio $m^g$ is initialized as 0, and accumulated at each training iteration by: + +$$ +m ^ {g} \leftarrow \gamma \cdot m ^ {g} + (1 - \gamma) \cdot m, \tag {8} +$$ + +where $\gamma$ is the momentum coefficient. It is a hyper-parameter already defined in FixMatch (baseline), where it is used to update the teacher parameters for final evaluation. + +Obviously, $m^g$ falls between 0 and 1. And it will approximately increase from 0 to a nearly saturated value. Then, the reweighted uncertain loss is given by: + +$$ +\mathcal {L} _ {s} ^ {2} = m ^ {g} \cdot \frac {1}{B _ {u}} \sum_ {k = 1} ^ {B _ {u}} \mathbb {1} \left(\xi \left(p _ {k} ^ {w}\right) < \tau\right) \cdot \hat {\mathrm {H}} \left(\hat {p} _ {k} ^ {w}, \hat {p} _ {k} ^ {s}\right). \tag {9} +$$ + +Integrating the above two intuitions and principles, the final reweighted uncertain loss will be: + +$$ +\mathcal {L} _ {s} = m ^ {g} \cdot \frac {1}{B _ {u}} \sum_ {k = 1} ^ {B _ {u}} \mathbb {1} \left(\xi \left(p _ {k} ^ {w}\right) < \tau\right) \cdot \hat {\mathrm {H}} \left(\hat {p} _ {k} ^ {w}, \hat {p} _ {k} ^ {s}\right) \cdot \xi \left(p _ {k} ^ {w}\right). \tag {10} +$$ + +# 3.4. Summary + +To summarize, the final loss in a mini-batch is a combination of the supervised loss $(\mathcal{L}_x)$ , certain loss $(\mathcal{L}_u$ , Eq. (1)), and uncertain loss in the shrunk class space $(\mathcal{L}_s$ , Eq. (10)): + +$$ +\mathcal {L} = \mathcal {L} _ {x} + \lambda_ {u} \cdot \left(\mathcal {L} _ {u} + \mathcal {L} _ {s}\right). \tag {11} +$$ + +We do not carefully fine-tune the fusion weight between $\mathcal{L}_u$ and $\mathcal{L}_s$ , but use 1:1 by default to avoid hyper-parameters. + +# 3.5. Discussions + +Our uncertain loss in the shrunk space owns two properties: informative and also safe. The first property is because we manage to find the largest shrunk space that reaches the confidence threshold. Besides, we also adopt the weak-to-strong consistency regularization to pose a challenging optimization target. Both constraints ensure the learning in the shrunk space is not trivial and still quite informative. Next, we wish to explain the second property "safe", especially about how we avoid noise in the shrunk space. + +Noises in pseudo labels distinguish the semi-supervised paradigm from the fully-supervised one. So designing a safe optimization target for unlabeled data is crucial. Typically, the cross entropy loss will maximize the softmax probability $\exp (s_t) / \sum_{i = 1}^{C}\exp (s_i)\rightarrow 1$ for the target class $t$ . It inevitably suppresses scores of all other classes except $t$ . However, the true class could not be $t$ , and may be the $2^{\mathrm{nd}}$ or $3^{\mathrm{rd}}$ largest class, etc., which is wrongly restrained. This is frequently observed when the confidence of class $t$ is not high enough. As a milder alternative, the soft labeling still encounters a similar problem. In contrast, our shrunk target directly excludes these confusion classes. So only the almost unlikely classes are suppressed. + +It is worth noting, even if the predicted top-1 class is not in line with the human label, it is probably one of the closest semantics, e.g., belonging to the same superclass as the groundtruth label, as shown in Fig. 2. Thus, contrasting these relevant classes against other less likely classes with our noise-robust shrunk loss is still beneficial. The model is encouraged to make discriminative predictions closer to top classes compared to tail classes. In addition, our disentangled auxiliary head can effectively leverage such supervision while not affecting the main classification tasks [7]. + +# 4. Experiment + +In this section, we first describe the implementation details of our framework. Then, we compare our ShrinkMatch with previous state-of-the-art methods (SOTAs) under extensive evaluation protocols. Lastly, we conduct comprehensive ablation studies on each component to validate the necessity. + +
Seed01234Mean
SimMatch [45]95.3495.1692.6393.7695.1094.39
ShrinkMatch | 4095.0994.6695.1294.7894.9594.92
SimMatch [45]95.5895.5095.3494.0695.2695.15
ShrinkMatch | 25095.3995.4495.3694.7695.3595.26
+ +Table 1: Comparison with SOTAs on CIFAR-10. The same seed ensures exactly the same data split. The 400 or 2500 denotes the number of labels. + +
Seed01234Mean
SimMatch [45]62.0660.1959.8964.8863.9262.19
ShrinkMatch | 40065.0063.4763.7766.4264.5264.64
SimMatch [45]74.6475.1974.5375.0375.2474.93
ShrinkMatch | 250075.0075.1174.5474.7874.7274.83
+ +# 4.1. Implementation Details + +Baselines. We use FixMatch + distribution alignment (DA) as our baseline on all datasets except ImageNet and SVHN. On ImageNet, we build our method on SimMatch. On SVHN, we discard DA. To be more convincing, we adopt the same codebase as our compared methods. + +Hyper-parameters. Following prior arts, Wide ResNet-28-2 [40], WRN-28-8, WRN-37-2, and WRN-28-2 are used for CIFAR-10, CIFAR-100, STL-10, and SVHN respectively. A ResNet-50 [13] is used for ImageNet. The auxiliary head $h_{aux}$ for uncertain samples is a 3-layer MLP. On the ImageNet, we set $B_u = 64 \times 5$ , but for other datasets, $B_u = 64 \times 7$ . The labeled batch size is 64 for all datasets. On the ImageNet, the model is trained for 400 epochs, while on the others, it is trained for $2^{20}$ iterations. The initial learning rate is set as 0.03 for all datasets with a cosine scheduler. Specifically, on the ImageNet, it is warmed up for 5 epochs. The consistency regularization $H$ in $\mathcal{L}_u$ on STL-10 and SVHN is a hard cross entropy (CE) loss, while on the CIFAR-10/100 and ImageNet, following the SimMatch and CoMatch [20], it is modified to a soft CE loss. The $\hat{H}$ in our proposed $\mathcal{L}_s$ is simply a hard CE loss. The weight $\lambda_u$ for the two unsupervised losses is set as 10 on the ImageNet and 1 on others. The confidence threshold $\tau$ is 0.7 on the ImageNet and 0.95 for others. The shared momentum coefficient $\gamma$ between our global certain ratio $m^g$ and the teacher parameters is 0.999. Following the common practice, the teacher model is only maintained for final evaluation. + +# 4.2. CIFAR-10 and CIFAR-100 + +The CIFAR dataset is composed of 50000/10000 training/validation images of size $32 \times 32$ . The CIFAR-10 con + +Table 2: Comparison with SOTAs on CIFAR-100. + +
Seed012Mean
FixMatch [28]65.8567.9458.3064.03
FlexMatch [42]76.7168.2867.5570.85
ShrinkMatch | 4085.7585.6486.5585.98
FixMatch [28]90.9188.7190.9490.19
FlexMatch [42]91.3592.2991.6691.77
ShrinkMatch | 25091.1392.4391.1091.55
+ +Table 3: Comparison with SOTAs on STL-10. + +
Seed012Mean
FixMatch [28]94.5396.9097.1496.19
FlexMatch [42]89.1989.9396.3291.81
ShrinkMatch | 4097.9697.8196.7097.49
FixMatch [28]98.0097.9997.9497.98
FlexMatch [42]91.7691.8396.6593.41
ShrinkMatch | 25098.0898.0697.9898.04
+ +Table 4: Comparison with SOTAs on SVHN. + +tains 10 balanced classes to recognize, while CIFAR-10 is much more challenging, containing 100 classes. We run our ShrinkMatch on five different seeds and report each result as well as the average result. On the CIFAR-10 of Tab. 1 where performance is almost saturated, we still obtain a nontrivial improvement of $0.53\%$ ( $94.39\% \to 94.92\%$ ) with 4 labels per class (400 labels in total). In addition, notably, as shown in Tab. 2, with only 4 labels per class on the challenging CIFAR-100, our ShrinkMatch remarkably outperforms SimMatch by $2.45\%$ ( $62.19\% \to 64.64\%$ ) on average. + +# 4.3. STL-10 and SVHN + +The STL-10 dataset originally consists of 5K/100K/8K labeled/unlabeled/validation images. We follow FlexMatch to only select a subset of 40/250 labeled images, while the unlabeled set remains unchanged. As shown in Tab. 3, with 4 labels per class, our ShrinkMatch surpasses FlexMatch tremendously from $70.85\%$ to $85.98\%$ $(+15.13\%)$ . + +The SVHN dataset is suitable to reveal the ability of different semi-supervised algorithms in the presence of class imbalance issue. As demonstrated in Tab. 4, although the performance of the original FixMatch (our baseline) has almost touched the upper bound, we still further boost it from $96.19\%$ to $97.49\%$ $(+1.30\%)$ with 4 labels per class. + +# 4.4. ImageNet-1K + +The ImageNet dataset is rather challenging, containing 1.28M/50K training/validation images of 1000 classes. We exactly follow the codebase of SimMatch [45]. As shown in Tab. 5, our ShrinkMatch further boosts previous best results on both settings of $1\%$ and $10\%$ labeled images. + +
Pre-trainingMethodEpochsParams (train / test)Top-1Top-5
1%10%1%10%
SimCLR v2 [8]80034.2M / 25.6M57.968.482.589.2
SwAV [6]Fine-tune80030.4M / 25.6M53.970.278.589.9
WCL [44]80034.2M / 25.6M65.072.086.391.2
MoCo v2 [9]Fine-tune80030.0M / 25.6M49.866.177.287.9
CoMatch [20]120030.0M / 25.6M67.173.787.191.4
MoCo-EMAN [5]FixMatch-EMAN [5]110030.0M / 25.6M63.074.083.490.9
NoneCoMatch [20]40030.0M / 25.6M66.073.686.491.6
SimMatch† [45]40030.0M / 25.6M67.074.186.991.5
ShrinkMatch40031.8M / 25.6M67.574.587.491.9
+ +![](images/c87cc982bf63139eb4e6b1c2d4e25af9ce32ab22746ff03050467e7f250b60cb.jpg) +(a) Number of removed confusion classes. + +![](images/6d4ac0e5765c69a424b6aaa025f86a60ac283a5d0a505573764e01c53cb7f3fa.jpg) +Figure 4: (a) The number of removed confusion classes for each uncertain image as the training proceeds. (b) Value of different loss functions, i.e., Eq. (1), Eq. (6), and Eq. (10). (c) The ratio of uncertain samples (certainty $< \tau$ ) in each mini-batch. + +![](images/99f0782972656e026945d310b4b2bac11410c18fb6fcae5fbdcba80b0ef1f0d7.jpg) +(b) Certain loss $\mathcal{L}_u$ and uncertain loss $\mathcal{L}_s$ . +(c) Ratio of uncertain samples. + +Table 5: Accuracy (%) on the ImageNet-1K with $1\%$ and ${10}\%$ labeled images. †: Reproduced in our environment. + +
Seed01234Mean
Baseline64.4057.5163.0764.7762.5362.46
ShrinkMatch65.0063.4763.7766.4264.5264.64
+ +# 4.5. Ablation Studies + +Unless otherwise specified, we conduct our ablation studies on CIFAR-100 with 4 labels per class. + +Effectiveness of our holistic ShrinkMatch. We first view our ShrinkMatch as a holistic component added to our baseline. As displayed in Tab. 6 for CIFAR-100, our ShrinkMatch boosts the strong baseline significantly by $2.18\%$ ( $62.46\% \rightarrow 64.64\%$ ). And on the STL-10 of Tab. 7, the baseline is also improved evidently from $84.85\%$ to $85.98\%$ ( $+1.13\%$ ). + +Effectiveness of two reweighting principles. In Sec. 3.3, considering the varied prediction reliability of different uncertain samples and different training stages, we propose two principles to reweight our uncertain loss $\mathcal{L}_s$ . So here we carefully examine their necessity in Tab. 8. It can be observed that the two principles are both beneficial, jointly improving the basic shrinking practice from $63.16\%$ to $64.24\%$ . We also attempt a simple alternative of linearly increasing the uncertain loss weight from 0 (start) to 1 (end). But as evi + +Table 6: Ablation studies of ShrinkMatch on CIFAR-100. + +
Seed012Mean
Baseline84.8684.2185.4784.85
ShrinkMatch85.7585.6486.5585.98
+ +Table 7: Ablation studies of ShrinkMatch on STL-10. + +denced in the Exp 3 & 4 of Tab. 8, it is obviously inferior to our designed reweighting strategy. We visualize our $m^g$ as training proceeds on ImageNet in Fig. 5. It looks essentially different from linear scheduling. Moreover, we visualize the uncertain loss value with or without the second reweighting principle in Fig. 4b. After reweighting, the uncertain loss is of a similar magnitude as the certain loss, which can avoid dominating the gradient at early training stages. + +The number of removed confusion classes. As introduced before, our approach detects and removes the confusion classes for the top-1 class in a fully automatic and instance-adaptive fashion. So we visualize the number of removed confusion classes for each uncertain sample in Fig. 4a. At the very start of training, model predictions are almost uniform, so abundant classes are removed for an uncertain sample. But as the training proceeds, much fewer confusion classes need to be removed to form a confident shrunk class space. + +Hard label or soft label in the shrunk class space? By + +
ExpPrinciple 1Principle 2LSSeed 0Seed 1Mean
165.6260.7063.16
2✓65.6862.0863.88
3✓65.0962.9064.00
4✓63.0159.4561.23
5✓✓65.0063.4764.24
+ +Table 8: Ablation studies on the effectiveness of the two reweighting principles. Due to randomness, the "Mean" results are more convincing. LS is short for linear scheduling, which linearly increases the loss weight. + +
ExpShrinkAux HeadLabelSeed 0Seed 1Mean
1Our Strong Baseline64.4057.5160.96
2✓✓H65.0063.4764.24
3✓✓S62.9259.7661.34
4✓H62.4358.8660.65
5✓S60.9260.2260.57
6✓H64.2159.1761.69
7✓S64.7560.8962.82
8H52.9051.8852.39
9S54.1055.7554.93
+ +default, we use the hard label in the shrunk space, following FixMatch. But we also ablate the choice of using the soft label as supervision. As shown in Exp 2 & 3 of Tab. 9, the hard label is much better than the soft label. We conjecture that this is because we have shrunk the class space to a safe one. The soft label will incur some unnecessary noise. + +Effects of the auxiliary head. To prevent the main classifier from being affected, we use an auxiliary head to learn the uncertain samples in the shrunk class space only for discriminative representations. So we first provide the results when still using the main head for the shrunk space. As shown in Exp 2 & 4 of Tab. 9, the auxiliary head is indispensable for our shrinking practice. So we continue to figure out whether the improvement of our ShrinkMatch merely comes from the auxiliary head. We attempt to use the auxiliary head to directly learn the uncertain samples in the original space without shrinking. It can be concluded from Exp 2 & 6 & 7 of Tab. 9, the original space is indeed much inferior to our shrunk space for uncertain samples, although the auxiliary head still helps the baseline to some extent. Lastly, in Tab. 9, we also include two additional results for readers, i.e., directly learning uncertain samples without shrinking or + +![](images/101689e4fbc26cfa05b514b02839eedd0961c52b84eaa44148e9c2bdecf24048.jpg) +Figure 5: Visualization of our global certain ratio $m^g$ . + +Table 9: Ablation studies on different options to learn uncertain sample. (1) whether to shrink the class space (Shrink), (2) whether to use an auxiliary head (Aux Head), and (3) use hard (H) or soft label (S). Exp 1 is our baseline without using uncertain images. Exp 8 & 9 are experiments of directly learning uncertain samples, which are quite noisy. + +
Threshold0.980.950.900.800.70
ShrinkMatch65.8065.0064.5562.7459.58
+ +Table 10: Ablation studies of the confidence threshold. + +auxiliary head (Exp 8 & 9), which are quite terrible because of introducing the abundant noise to our main classifier. + +The ratio of uncertain samples. To further justify our motivation, in Fig. 4c, we display the ratio of uncertain samples in each mini-batch. It can be seen that even in the middle of the whole training course, there are still around $40\%$ and $30\%$ uncertain samples on ImageNet and CIFAR-100 respectively. Therefore, an appropriate approach to utilizing these samples is necessary and definitely beneficial. + +Different confidence thresholds $\tau$ . We also try different thresholds for uncertain samples, as shown in Tab. 10. When increasing $\tau$ from widely adopted 0.95 to 0.98 (more uncertain samples), our ShrinkMatch can be further improved. These results clearly highlight the effective and safe utilization of uncertain samples with our ShrinkMatch. + +# 5. Conclusion + +In this work, we aim to fully leverage the uncertain samples in semi-supervised learning. We point out that the low certainty is typically caused by a small portion of confusion classes. Invoked by this, we propose a novel method dubbed ShrinkMatch to automatically detect and remove the confusion classes to construct a shrunk class space, where the top-1 class is turned certain. A weak-to-strong consistency regularization is enforced in the confident new space. Furthermore, we design two reweighting principles for the auxiliary uncertain loss, according to the reliability of different uncertain samples and the gradually improved state of the model. Consequently, our method establishes new state-of-the-art results on widely acknowledged benchmarks. + +Acknowledgements. This work is supported in part by the National Natural Science Foundation of China (62201484, 62222604, 62192783, 62206052), HKU Startup Fund, and HKU Seed Fund for Basic Research. + +# References + +[1] Eric Arazo, Diego Ortego, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In IJCNN, 2020. 1 +[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. In ICCV, 2021. 2 +[3] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. In NeurIPS, 2019. 2 +[4] Zhaowei Cai, Avinash Ravichandran, Paolo Favaro, Manchen Wang, Davide Modolo, Rahul Bhotika, Zhuowen Tu, and Stefano Soatto. Semi-supervised vision transformers at scale. In NeurIPS, 2022. 2 +[5] Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Charless Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi-supervised learning. In CVPR, 2021. 7 +[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 7 +[7] Baixu Chen, Junguang Jiang, Ximei Wang, Jianmin Wang, and Mingsheng Long. Debiased pseudo labeling in self-training. In NeurIPS, 2022. 3, 5 +[8] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. In NeurIPS, 2020. 7 +[9] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv:2003.04297, 2020. 7 +[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 1 +[11] Yue Fan, Dengxin Dai, Anna Kukleva, and Bernt Schiele. Cossl: Co-learning of representation and classifier for imbalanced semi-supervised learning. In CVPR, 2022. 2 +[12] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In ICML, 2016. 3 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6 +[14] Zijian Hu, Zhengyu Yang, Xuefeng Hu, and Ram Nevatia. Simple: similar pseudo label exploitation for semi-supervised classification. In CVPR, 2021. 2 +[15] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? In NeurIPS, 2017. 3 +[16] Jiwon Kim, Youngjo Min, Daehwan Kim, Gyuseong Lee, Junyoung Seo, Kwangrok Ryoo, and Seungryong Kim. Conmatch: Semi-supervised learning with confidence-guided consistency regularization. In ECCV, 2022. 2 + +[17] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1 +[18] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In ICLR, 2017. 2 +[19] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In ICML Workshop, 2013. 2 +[20] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regularization. In ICCV, 2021. 2, 6, 7 +[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 1 +[22] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. TPAMI, 2018. 2 +[23] Islam Nassar, Samitha Herath, Ehsan Abbasnejad, Wray Buntine, and Gholamreza Haffari. All labels are not created equal: Enhancing semi-supervision via label grouping and co-training. In CVPR, 2021. 2 +[24] Yassine Ouali, Céline Hudelot, and Myriam Tami. An overview of deep semi-supervised learning. arXiv:2006.05278, 2020. 2 +[25] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V Le. Meta pseudo labels. In CVPR, 2021. 2 +[26] Mamshad Nayeem Rizve, Kevin Duarte, Yogesh S Rawat, and Mubarak Shah. In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. In ICLR, 2021. 3 +[27] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In NeurIPS, 2016. 2 +[28] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 1, 2, 3, 6 +[29] Fariborz Taherkhani, Ali Dabouei, Sobhan Soleymani, Jeremy Dawson, and Nasser M Nasrabadi. Self-supervised Wasserstein pseudo-labeling for semi-supervised image classification. In CVPR, 2021. 2 +[30] Hui Tang and Kui Jia. Towards discovering the effectiveness of moderately confident samples for semi-supervised learning. In CVPR, 2022. 2 +[31] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In NeurIPS, 2017. 2 +[32] Guotai Wang, Wenqi Li, Michael Aertsen, Jan Deprest, Sébastien Ourselin, and Tom Vercauteren. Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolutional neural networks. Neurocomputing, 2019. 3 +[33] Xudong Wang, Long Lian, and Stella X Yu. Unsupervised selective labeling for more effective semi-supervised learning. In ECCV, 2022. 2 + +[34] Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V Le. Unsupervised data augmentation for consistency training. In NeurIPS, 2020. 2 +[35] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020. 2 +[36] Yi Xu, Lei Shang, Jinxing Ye, Qi Qian, Yu-Feng Li, Baigui Sun, Hao Li, and Rong Jin. Dash: Semi-supervised learning with dynamic thresholding. In ICML, 2021. 2, 3 +[37] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In CVPR, 2022. 2 +[38] David Yarowsky. Unsupervised word sense disambiguation, rivaling supervised methods. In ACL, 1995. 2 +[39] Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In MICCAI, 2019. 3 +[40] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC, 2016. 6 +[41] Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. In ICCV, 2019. 2 +[42] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In NeurIPS, 2021. 2, 6 +[43] Zhen Zhao, Luping Zhou, Lei Wang, Yinghuan Shi, and Yang Gao. Lassl: Label-guided self-training for semi-supervised learning. In AAAI, 2022. 2 +[44] Mingkai Zheng, Fei Wang, Shan You, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Weakly supervised contrastive learning. In ICCV, 2021. 7 +[45] Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In CVPR, 2022. 3, 6, 7 \ No newline at end of file diff --git a/2023/Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning/images.zip b/2023/Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..702f9aa99681fd4bfccb65f6a1dc62c6d6be08c2 --- /dev/null +++ b/2023/Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca581d5a56d9c0f1737cb9e20a520c6567a40cdd8ce841b0800c51b15a477e0b +size 477373 diff --git a/2023/Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning/layout.json b/2023/Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ee1809c68ce70b46f8d4c37a8436f12fa1c1212d --- /dev/null +++ b/2023/Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning/layout.json @@ -0,0 +1,9494 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 533, + 123 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 533, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 533, + 123 + ], + "type": "text", + "content": "Shrinking Class Space for Enhanced Certainty in Semi-Supervised Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 141, + 512, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 141, + 512, + 159 + ], + "spans": [ + { + "bbox": [ + 85, + 141, + 512, + 159 + ], + "type": "text", + "content": "Lihe Yang" + }, + { + "bbox": [ + 85, + 141, + 512, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 141, + 512, + 159 + ], + "type": "text", + "content": " Zhen Zhao" + }, + { + "bbox": [ + 85, + 141, + 512, + 159 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 85, + 141, + 512, + 159 + ], + "type": "text", + "content": " Lei Qi" + }, + { + "bbox": [ + 85, + 141, + 512, + 159 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 85, + 141, + 512, + 159 + ], + "type": "text", + "content": " Yu Qiao" + }, + { + "bbox": [ + 85, + 141, + 512, + 159 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 85, + 141, + 512, + 159 + ], + "type": "text", + "content": " Yinghuan Shi" + }, + { + "bbox": [ + 85, + 141, + 512, + 159 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 85, + 141, + 512, + 159 + ], + "type": "text", + "content": " Hengshuang Zhao" + }, + { + "bbox": [ + 85, + 141, + 512, + 159 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 167, + 159, + 426, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 159, + 426, + 175 + ], + "spans": [ + { + "bbox": [ + 167, + 159, + 426, + 175 + ], + "type": "text", + "content": "1The University of Hong Kong 2Nanjing University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 175, + 487, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 175, + 487, + 189 + ], + "spans": [ + { + "bbox": [ + 107, + 175, + 487, + 189 + ], + "type": "text", + "content": "3Shanghai AI Laboratory 4The University of Sydney 5Southeast University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 187, + 190, + 405, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 190, + 405, + 201 + ], + "spans": [ + { + "bbox": [ + 187, + 190, + 405, + 201 + ], + "type": "text", + "content": "https://github.com/LiheYoung/ShrinkMatch" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 230, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 230, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 230, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 255, + 290, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 290, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 290, + 531 + ], + "type": "text", + "content": "Semi-supervised learning is attracting blooming attention, due to its success in combining unlabeled data. To mitigate potentially incorrect pseudo labels, recent frameworks mostly set a fixed confidence threshold to discard uncertain samples. This practice ensures high-quality pseudo labels, but incurs a relatively low utilization of the whole unlabeled set. In this work, our key insight is that these uncertain samples can be turned into certain ones, as long as the confusion classes for the top-1 class are detected and removed. Invoked by this, we propose a novel method dubbed ShrinkMatch to learn uncertain samples. For each uncertain sample, it adaptively seeks a shrunk class space, which merely contains the original top-1 class, as well as remaining less likely classes. Since the confusion ones are removed in this space, the re-calculated top-1 confidence can satisfy the pre-defined threshold. We then impose a consistency regularization between a pair of strongly and weakly augmented samples in the shrunk space to strive for discriminative representations. Furthermore, considering the varied reliability among uncertain samples and the gradually improved model during training, we correspondingly design two reweighting principles for our uncertain loss. Our method exhibits impressive performance on widely adopted benchmarks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 554, + 128, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 128, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 128, + 566 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 574, + 288, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 288, + 657 + ], + "type": "text", + "content": "In the last decade, our computer vision community has witnessed inspiring progress, thanks to large-scale datasets [21, 10]. Nevertheless, it is laborious and costly to annotate massive images, hindering the progress to benefit a broader range of real-world scenarios. Inspired by this, semi-supervised learning (SSL) was proposed to utilize the unlabeled data under the assistance of limited labeled data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 658, + 288, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 658, + 288, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 658, + 288, + 695 + ], + "type": "text", + "content": "The frameworks in SSL are typically based on the strategy of pseudo labeling. Briefly, the model acquires knowledge from the labeled data, and then assigns predictions on the" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 331, + 228, + 523, + 380 + ], + "blocks": [ + { + "bbox": [ + 331, + 228, + 523, + 380 + ], + "lines": [ + { + "bbox": [ + 331, + 228, + 523, + 380 + ], + "spans": [ + { + "bbox": [ + 331, + 228, + 523, + 380 + ], + "type": "image", + "image_path": "ae66fb03b8076858a055ecce8796fd1d19d44bae77fc18af83490942b6da113e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 388, + 548, + 461 + ], + "lines": [ + { + "bbox": [ + 304, + 388, + 548, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 548, + 461 + ], + "type": "text", + "content": "Figure 1: Illustration of our motivation. Due to confusion classes for the top-1 class, the certainty fails to reach the pre-defined threshold (gray dotted line). FixMatch discards such uncertain samples. Our method, however, detects and removes confusion classes to enhance certainty, then enjoying full and safe utilization of all unlabeled images." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 472, + 547, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 652 + ], + "type": "text", + "content": "unlabeled data. The two sources of data are finally combined to train a better model. During this process, it is obvious that predictions on unlabeled data are not reliable. If the model is iteratively trained with incorrect pseudo labels, it will suffer the confirmation bias issue [1]. To address this dilemma, recent works [28] simply set a fixed confidence threshold to discard potentially unreliable samples. This simple strategy effectively retains high-quality pseudo labels, however, it also incurs a low utilization of the whole unlabeled set. As evidenced by our pilot study on CIFAR-100 [17], nearly " + }, + { + "bbox": [ + 304, + 472, + 547, + 652 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 472, + 547, + 652 + ], + "type": "text", + "content": " unlabeled images are filtered out for not satisfying the threshold of 0.95. Instead of blindly throwing them away, we believe there should exist a more promising approach. This work is just aimed to fully leverage previously uncertain samples in an informative but also safe manner." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "So first, why are these samples uncertain? According to our observations on CIFAR-100 and ImageNet, although the top-1 accuracy could be low, the top-5 accuracy is much higher. This indicates in most cases, the model struggles to discriminate among a small portion of classes. As illustrated" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 138, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 138, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 138, + 713 + ], + "type": "text", + "content": "*Corresponding authors" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16187" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 270, + 188 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 270, + 188 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 270, + 188 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 270, + 188 + ], + "type": "image", + "image_path": "86eafff981b9f29c8572753a51af82349256cd27529a4ac7959752ef39fbf053.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 197, + 288, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 197, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 197, + 288, + 258 + ], + "type": "text", + "content": "Figure 2: Pseudo label accuracy on CIFAR-100 with 400 labels. We highlight that even for uncertain samples, their top-1 predictions are of high accuracy in the superclass space (20 classes). This accuracy can even be comparable to the delicately selected certain samples in the original class space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 265, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 287, + 337 + ], + "type": "text", + "content": "in Fig. 1, given a cat image, the model is not sure whether it belongs to tabby cat, tiger cat, or other cats. On the other hand, however, it is absolutely certain that the object is more like a tabby cat, rather than a table or anything else. In other words, it is reliable for the model to distinguish the top class from the remaining less likely classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 338, + 288, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 288, + 495 + ], + "type": "text", + "content": "Invoked by these, we propose a novel method dubbed ShrinkMatch to learn uncertain samples. The prediction fails to satisfy the pre-defined threshold due to the existence of confusion classes for the top-1 class. Hence, our approach adaptively seeks a shrunk class space where the confusion classes are removed, to enable the re-calculated confidence to reach the original threshold. Moreover, the obtained shrunk class space is also required to be the largest among those that can satisfy the threshold. In a word, we seek a certain and largest shrunk space for uncertain samples. Then, logits of the strongly augmented image are correspondingly gathered in the new space. And a consistency regularization is imposed between the shrunk weak-strong predictions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 496, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 652 + ], + "type": "text", + "content": "Note that, the confusion classes are detected and removed in a fully automatic and instance-adaptive fashion. Moreover, even if the predicted top-1 class is not fully in line with the groundtruth label, they mostly belong to the same superclass. To prove this, as shown in Fig. 2, uncertain samples exhibit much higher pseudo label accuracy in the superclass space than the original space. This accuracy is even comparable to that of certain samples in the original class space. Therefore, contrasting these groundtruth-related classes against remaining unlikely classes is still highly beneficial, yielding more discriminative representations of our model. To avoid affecting the main classifier, we further adopt an auxiliary classifier to disentangle the learning in the shrunk space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Despite the effectiveness, there still exist two main drawbacks to the above optimization target. (1) First, it treats all uncertain samples equally. The truth, however, is that in the original class space, the top-1 confidence of different uncertain samples can vary dramatically. And it is clear" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 275 + ], + "type": "text", + "content": "that samples with larger confidence should be attached more importance. To this end, we propose to balance different uncertain samples by their confidence in the original space. (2) Moreover, the regularization term also overlooks the gradually improved model state during training. At the start of training, there are abundant uncertain samples, but their predictions are extremely noisy, or even random. So even the highest scored class may share no relationship with the true class. Then as the training proceeds, the top classes become reliable. Considering this, we further propose to adaptively reweight the uncertain loss according to the model state. The model state is tracked and approximately estimated via performing exponential moving average on the proportion of certain samples in each mini-batch. With the two reweighting principles, the model turns out more stable, and avoids accumulating much noise from uncertain samples, especially at early training iterations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 276, + 529, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 529, + 288 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 529, + 288 + ], + "type": "text", + "content": "To summarize, our contributions lie in three aspects:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 297, + 547, + 445 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 317, + 297, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 297, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 317, + 297, + 545, + 357 + ], + "type": "text", + "content": "- We first point out that low certainty is typically caused by a small portion of confusion classes. To enhance the certainty, we propose to shrink the original class space by adaptively detecting and removing confusion ones for the top-1 class to turn it certain in the new space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 364, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 364, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 317, + 364, + 545, + 413 + ], + "type": "text", + "content": "- We manage to reweight the uncertain loss from two perspectives: the image-based varied reliability among different uncertain samples, and the model-based gradually improved state as the training proceeds." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 419, + 547, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 419, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 317, + 419, + 547, + 445 + ], + "type": "text", + "content": "- Our proposed ShrinkMatch establishes new state-of-the-art results on widely acknowledged benchmarks." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 453, + 392, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 392, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 392, + 467 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "content": "Semi-supervised learning (SSL). The primary concern in SSL [19, 28, 36, 20, 25, 16, 37, 35, 41, 43, 14, 24, 29, 23, 4, 30, 11, 33, 2, 42] is to design effective constraints for unlabeled samples. Dating back to decades ago, pioneering works [19, 38] integrate unlabeled data via assigning pseudo labels to them, with the knowledge acquired from labeled data. In the era of deep learning, subsequent methods mainly follow such a bootstrapping fashion, but greatly boost it with some key components. Specifically, to enhance the quality of pseudo labels, " + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "content": "-model [18] and Mean Teacher [31] ensemble model predictions and model parameters respectively. Later works start to exploit the role of perturbations. During this trend, [27] proposes to apply stochastic perturbations on inputs or features, and enforce consistency across these predictions. Then, UDA [34] emphasizes the necessity of strengthening the perturbation pool. It also follows VAT [22] and MixMatch [3] to supervise the prediction under strong perturbations with that under weak perturbations. Since then, weak-to-strong consistency regularization has become a standard practice in SSL. Eventually, the milestone" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16188" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "work FixMatch [28] presents a simplified framework using a fixed confidence threshold to discard uncertain samples. Our ShrinkMatch is built upon FixMatch. But, we highlight the value of previously neglected uncertain samples, and leverage them in an informative but also safe manner." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 204 + ], + "type": "text", + "content": "More recently, DST [7] decouples the generation and utilization of pseudo labels with a main and an auxiliary head respectively. Besides, SimMatch [45] explores instance-level relationships with labeled embeddings to supplement original class-level relations. Compared with them, our ShrinkMatch achieves larger improvements." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 209, + 288, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 288, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 288, + 329 + ], + "type": "text", + "content": "Defining uncertain samples. Earlier works estimate the uncertainty with Bayesian Neural Networks [15], or its faster approximation, e.g., Monte Carlo Dropout [12]. Some other works measure the prediction disagreement among multiple randomly augmented inputs [32]. The latest trend is to directly use the entropy of predictions [39], cross entropy [36], or softmax confidence [28] as a measurement for uncertainty. Our work is not aimed at the optimal uncertainty estimation strategy, so we adopt the simplest solution from FixMatch, i.e., using the maximum softmax output as the certainty." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 335, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 441 + ], + "type": "text", + "content": "Utilizing uncertain samples. UPS [26] leverages negative class labels whose confidence is below a pre-defined threshold, from a reversed multi-label classification perspective. In comparison, our model is enforced to tell the most likely class without being cheated by the less likely ones. So our supervision on uncertain samples is more informative and produces more discriminative representations. Moreover, we do not introduce any extra hyper-parameters, e.g., the lower threshold in [26], into our framework." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 453, + 102, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 102, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 102, + 464 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 472, + 288, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 288, + 533 + ], + "type": "text", + "content": "We primarily provide some notations and review a common practice in semi-supervised learning (SSL) (Sec. 3.1). Next, we present our ShrinkMatch in detail (Sec. 3.2 and Sec. 3.3). Finally, we summarize our approach and provide a further discussion (Sec. 3.4 and Sec. 3.5)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 540, + 133, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 133, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 133, + 551 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": "Semi-supervised learning aims to learn a model with limited labeled images " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^l = \\{(x_k, y_k)\\}" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": ", aided by a large number of unlabeled images " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "D^u = \\{u_k\\}" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": ". Recent frameworks commonly follow the FixMatch practice. Concretely, an unlabeled image " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": " is first transformed by a weak augmentation pool " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^w" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": " and a strong augmentation pool " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^s" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": " to yield a pair of weakly and strongly augmented images " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "(u^w, u^s)" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": ". Then, they are fed into the model together to produce corresponding predictions " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "(p^w, p^s)" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": ". Typically, " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "p^w" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": " is of much higher accuracy than " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "p^s" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "p^s" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": " is beneficial to learning. Therefore, " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "p^w" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": " serves as the pseudo label for " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "p^s" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": ". Moreover, a core practice introduced by FixMatch is that, to improve the quality of selected pseudo labels, a fixed confidence threshold is" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "pre-defined to abandon uncertain ones. So the unsupervised loss " + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": " can be formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 341, + 96, + 545, + 126 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 96, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 341, + 96, + 545, + 126 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {u} = \\frac {1}{B _ {u}} \\sum_ {k = 1} ^ {B _ {u}} \\mathbb {1} \\left(\\xi \\left(p _ {k} ^ {w}\\right) \\geq \\tau\\right) \\cdot \\mathrm {H} \\left(p _ {k} ^ {w}, p _ {k} ^ {s}\\right), \\tag {1}", + "image_path": "17556067a52bb7550c639d9a37b94aaeefd61f0159b2a043f916805c77fd9339.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 127, + 547, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 127, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 304, + 127, + 547, + 186 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 127, + 547, + 186 + ], + "type": "inline_equation", + "content": "B^{u}" + }, + { + "bbox": [ + 304, + 127, + 547, + 186 + ], + "type": "text", + "content": " is the batch size of unlabeled images and " + }, + { + "bbox": [ + 304, + 127, + 547, + 186 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 127, + 547, + 186 + ], + "type": "text", + "content": " is the pre-defined threshold. " + }, + { + "bbox": [ + 304, + 127, + 547, + 186 + ], + "type": "inline_equation", + "content": "\\xi(p_{k}^{w})" + }, + { + "bbox": [ + 304, + 127, + 547, + 186 + ], + "type": "text", + "content": " computes the confidence of logits " + }, + { + "bbox": [ + 304, + 127, + 547, + 186 + ], + "type": "inline_equation", + "content": "p_{k}^{w}" + }, + { + "bbox": [ + 304, + 127, + 547, + 186 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 304, + 127, + 547, + 186 + ], + "type": "inline_equation", + "content": "\\xi(\\cdot) = \\max(\\sigma(\\cdot))" + }, + { + "bbox": [ + 304, + 127, + 547, + 186 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 127, + 547, + 186 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 127, + 547, + 186 + ], + "type": "text", + "content": " is the softmax function. The H denotes the consistency regularization between the two distributions. It is typically the cross entropy loss." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 186, + 545, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 545, + 221 + ], + "type": "text", + "content": "In addition, the labeled images are learned with a regular cross entropy loss to obtain the supervised loss " + }, + { + "bbox": [ + 304, + 186, + 545, + 221 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_x" + }, + { + "bbox": [ + 304, + 186, + 545, + 221 + ], + "type": "text", + "content": ". The overall loss in each mini-batch then will be:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 385, + 223, + 545, + 234 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 223, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 385, + 223, + 545, + 234 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {x} + \\lambda_ {u} \\cdot \\mathcal {L} _ {u}, \\tag {2}", + "image_path": "2af3f9a588ef46d4bda3d1c82ca478ffe3f703a3f73cfc1d7c55a162fecc18a4.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 235, + 535, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 235, + 535, + 247 + ], + "spans": [ + { + "bbox": [ + 305, + 235, + 535, + 247 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 235, + 535, + 247 + ], + "type": "inline_equation", + "content": "\\lambda_{u}" + }, + { + "bbox": [ + 305, + 235, + 535, + 247 + ], + "type": "text", + "content": " acts as a trade-off term between the two losses." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 252, + 515, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 252, + 515, + 266 + ], + "spans": [ + { + "bbox": [ + 305, + 252, + 515, + 266 + ], + "type": "text", + "content": "3.2. Shrinking the Class Space for Certainty" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 270, + 547, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 547, + 449 + ], + "type": "text", + "content": "Our motivation. As reviewed above, FixMatch discards the samples whose confidence is lower than a pre-defined threshold, because their pseudo labels are empirically found relatively noisier. These samples are named uncertain samples in this work. Although this practice retains high-quality pseudo labels, it incurs a low utilization of the whole unlabeled set, especially when the scenario is challenging and the selection criterion is strict. Take the CIFAR-100 dataset as an instance, with a common threshold of 0.95 and 4 labels per class, there will be nearly " + }, + { + "bbox": [ + 304, + 270, + 547, + 449 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 270, + 547, + 449 + ], + "type": "text", + "content": " unlabeled samples being ignored due to their low certainty. We argue that, these uncertain samples can still benefit the model optimization, as long as we can design appropriate constraints (loss functions) on them. So we first investigate the cause of low certainty to gain some better intuitions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 451, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 547, + 713 + ], + "type": "text", + "content": "The reason for the low certainty of an unlabeled image is that, the model tends to be confused among some top classes. For example, given a cat image, the score of the class tabby cat and class tiger cat may be both high, so the model is not absolutely certain what the concrete class is. Motivated by this observation, we propose to shrink the original class space via adaptively detecting and removing the confusion classes for the top-1 class. Then the shrunk space is only composed of the original top-1 class, as well as the remaining less likely classes. After this process, re-calculated confidence of the top-1 class will satisfy the pre-defined threshold. Thereby, we can enforce the model to learn the previously uncertain samples in this new certain space, as shown in Fig. 3. Following the previous cat example, if the top-1 class is tabby cat, our method will scan scores of all other classes, and remove confusion ones (e.g., tiger cat and siamese cat) to construct a confident shrunk space, where the model is sure that the image is a tabby cat, rather than a table. Since we do not ask the model to discriminate among several top classes, it will avoid suffering from the noise when it makes a wrong judgment in the original space." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16189" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 410, + 198 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 410, + 198 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 410, + 198 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 410, + 198 + ], + "type": "image", + "image_path": "03d015f6bb40c84a03b58c2f1731515a2b38f84620b092cbcfa307c979e11d3d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 202, + 547, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 547, + 275 + ], + "type": "text", + "content": "Figure 3: An overview of our proposed ShrinkMatch. Our motivation is to fully leverage the originally uncertain samples. \"S\" denotes shrinking the class space. The confusion classes for the top-1 class are detected and removed in a fully automatic and instance-adaptive fashion, to construct a shrunk space where the top-1 class is turned certain. " + }, + { + "bbox": [ + 46, + 202, + 547, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 46, + 202, + 547, + 275 + ], + "type": "text", + "content": " is the original certain loss, while " + }, + { + "bbox": [ + 46, + 202, + 547, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 46, + 202, + 547, + 275 + ], + "type": "text", + "content": " calculates the uncertain loss in the shrunk class space. We add an auxiliary head " + }, + { + "bbox": [ + 46, + 202, + 547, + 275 + ], + "type": "inline_equation", + "content": "h_{aux}" + }, + { + "bbox": [ + 46, + 202, + 547, + 275 + ], + "type": "text", + "content": " to learn in the new space. On the right, we further reweight " + }, + { + "bbox": [ + 46, + 202, + 547, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 46, + 202, + 547, + 275 + ], + "type": "text", + "content": " based on two principles. Principle 1 (image-based): image predictions with larger reliability are attached more importance. Principle 2 (model-based): We track the model state during training for reweighting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 433, + 70, + 531, + 198 + ], + "blocks": [ + { + "bbox": [ + 433, + 70, + 531, + 198 + ], + "lines": [ + { + "bbox": [ + 433, + 70, + 531, + 198 + ], + "spans": [ + { + "bbox": [ + 433, + 70, + 531, + 198 + ], + "type": "image", + "image_path": "bf167c4db3d499a9caeb5bf1f0765d7fd89760688be2bd5182872fdf52d228cd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 289, + 289, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 289, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 289, + 445 + ], + "type": "text", + "content": "How to seek the shrunk class space? Now, how can we enable our method to automatically seek the optimal shrunk space of an uncertain unlabeled image? Ideally, we hope this seeking process is free from any prior knowledge from humans, e.g., class relationships, and also does not require any extra hyper-parameters. Considering these, we opt to inherit the pre-defined confidence threshold as a criterion. To be specific, we first sort the predicted logits in a descending order to obtain " + }, + { + "bbox": [ + 46, + 289, + 289, + 445 + ], + "type": "inline_equation", + "content": "p^w = \\{s_{n_i}^w\\}_{i=1}^C" + }, + { + "bbox": [ + 46, + 289, + 289, + 445 + ], + "type": "text", + "content": " for classes " + }, + { + "bbox": [ + 46, + 289, + 289, + 445 + ], + "type": "inline_equation", + "content": "\\{n_i\\}_{i=1}^C" + }, + { + "bbox": [ + 46, + 289, + 289, + 445 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 289, + 289, + 445 + ], + "type": "inline_equation", + "content": "s_{n_i}^w \\geq s_{n_i+1}^w" + }, + { + "bbox": [ + 46, + 289, + 289, + 445 + ], + "type": "text", + "content": ". In the shrunk space, we will retain the original top-1 class, because it is still the most likely one to be true. Then, we find a set of less likely classes " + }, + { + "bbox": [ + 46, + 289, + 289, + 445 + ], + "type": "inline_equation", + "content": "\\{n_i\\}_{i=K}^C" + }, + { + "bbox": [ + 46, + 289, + 289, + 445 + ], + "type": "text", + "content": " by enforcing two constraints on the " + }, + { + "bbox": [ + 46, + 289, + 289, + 445 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 289, + 289, + 445 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 454, + 287, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 454, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 104, + 454, + 287, + 469 + ], + "type": "interline_equation", + "content": "\\xi \\left(\\left\\{s _ {n _ {1}} ^ {w} \\right\\} \\cup \\left\\{s _ {n _ {i}} ^ {w} \\right\\} _ {i = K} ^ {C}\\right) \\geq \\tau , \\tag {3}", + "image_path": "a213d6016224316e5159f38e41751af2aa1d966de95acd8e3cca8e88023267b5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 471, + 287, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 471, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 287, + 486 + ], + "type": "interline_equation", + "content": "\\xi \\left(\\left\\{s _ {n _ {1}} ^ {w} \\right\\} \\cup \\left\\{s _ {n _ {i}} ^ {w} \\right\\} _ {i = K - 1} ^ {C}\\right) < \\tau , \\tag {4}", + "image_path": "bca60a20c09f14c11eafa5e4304d1e67ce8cd925c958afbcc9ee8a0b8b436042.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 495, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 495, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 46, + 495, + 287, + 567 + ], + "type": "text", + "content": " is defined the same as that in Eq. (1), calculating the confidence of the re-assembled logits. The final shrunk space is composed of the re-assembled classes " + }, + { + "bbox": [ + 46, + 495, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\{n_1\\} \\cup \\{n_i\\}_{i = K}^C" + }, + { + "bbox": [ + 46, + 495, + 287, + 567 + ], + "type": "text", + "content": ". The two constraints on " + }, + { + "bbox": [ + 46, + 495, + 287, + 567 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 495, + 287, + 567 + ], + "type": "text", + "content": " not only ensure the top-1 class is turned certain in the new space (Eq. (3)), but also select the largest space among all candidates (Eq. (4))." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": "How to learn in the shrunk class space? For a weakly augmented uncertain image " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "x^w" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": ", the model is certain about its top-1 class in the shrunk class space. To learn effectively in this space, we follow the popular practice of weak-to-strong consistency regularization. The correspondingly shrunk prediction on the strongly augmented image is enforced to match that on the weakly augmented one. Concretely, the re-assembled logits from " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "p^w" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\hat{p}^w" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": ", which means " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\hat{p}^w = \\{s_{n_1}^w\\} \\cup \\{s_{n_i}^w\\}_K^C" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": ". We use the re-assembled classes " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\{n_1\\} \\cup \\{n_i\\}_K^C" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": " in the shrunk space to correspondingly gather the logits " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "p^s" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "x^s" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": ", yielding " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\hat{p}^s = \\{s_{n_1}^s\\} \\cup \\{s_{n_i}^s\\}_K^C" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": ". Then we can regularize the consistency between the two shrunk" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 289, + 504, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 289, + 504, + 301 + ], + "spans": [ + { + "bbox": [ + 306, + 289, + 504, + 301 + ], + "type": "text", + "content": "distributions " + }, + { + "bbox": [ + 306, + 289, + 504, + 301 + ], + "type": "inline_equation", + "content": "\\hat{p}^s" + }, + { + "bbox": [ + 306, + 289, + 504, + 301 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 289, + 504, + 301 + ], + "type": "inline_equation", + "content": "\\hat{p}^w" + }, + { + "bbox": [ + 306, + 289, + 504, + 301 + ], + "type": "text", + "content": ", similar to that in Eq. (1):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 341, + 311, + 545, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 311, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 341, + 311, + 545, + 343 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\frac {1}{B _ {u}} \\sum_ {k = 1} ^ {B _ {u}} \\mathbb {1} \\left(\\xi \\left(p _ {k} ^ {w}\\right) < \\tau\\right) \\cdot \\hat {\\mathrm {H}} \\left(\\hat {p} _ {k} ^ {w}, \\hat {p} _ {k} ^ {s}\\right), \\tag {5}", + "image_path": "99e736b2fdc9a7853df76a728b91c88e93b1f6cff4c6b15ce4b0cdb70eed18ce.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 352, + 535, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 352, + 535, + 363 + ], + "spans": [ + { + "bbox": [ + 305, + 352, + 535, + 363 + ], + "type": "text", + "content": "where the indicator function is to find uncertain samples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "type": "text", + "content": "Empirically, we observe that if we use the original linear head " + }, + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "type": "inline_equation", + "content": "h_{main}" + }, + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "type": "text", + "content": " to learn this auxiliary supervision, it will make the confidence of our model increase aggressively. Most noisy unlabeled samples are blindly judged as certain ones. We conjecture that it is because the " + }, + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "type": "text", + "content": " strengthens weights of the classes that are frequently uncertain, then these classes will be incorrectly turned certain. With this in mind, our solution is simple. We adopt an auxiliary MLP head " + }, + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "type": "inline_equation", + "content": "h_{aux}" + }, + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "type": "text", + "content": " that shares the backbone with " + }, + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "type": "inline_equation", + "content": "h_{main}" + }, + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "type": "text", + "content": ", to deal with this auxiliary optimization target, as shown in Fig. 3. So the " + }, + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "type": "inline_equation", + "content": "\\hat{p}^s" + }, + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "type": "text", + "content": " is indeed gathered from predictions of " + }, + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "type": "inline_equation", + "content": "h_{aux}" + }, + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "type": "inline_equation", + "content": "\\hat{p}^w" + }, + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "type": "text", + "content": " is still from " + }, + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "type": "inline_equation", + "content": "h_{main}" + }, + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "type": "text", + "content": "). This modification enables our feature extractor to acquire more discriminative representations, and meantime does not affect predictions of the main head. Note that " + }, + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "type": "inline_equation", + "content": "h_{aux}" + }, + { + "bbox": [ + 304, + 364, + 546, + 544 + ], + "type": "text", + "content": " is only applied for training, bringing no burden to the test stage." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 552, + 478, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 478, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 478, + 563 + ], + "type": "text", + "content": "3.3. Reweighting the Uncertain Loss" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "Despite the effectiveness of the above uncertain loss, there still exist two main drawbacks. (1) On one hand, it overlooks the varied reliability of the top-1 class among different uncertain images. For example, suppose two uncertain images " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "u_{1}" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "u_{2}" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " with softmax predictions [0.8, 0.1, 0.1] and [0.5, 0.3, 0.2] in the original space, they should not be treated equally in the shrunk space. The " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "u_{1}" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " with top-1 confidence of 0.8 is more likely to be true than " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "u_{2}" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": ", and thereby should be attached more attention to. (2) On the other, it ignores the gradually improved model performance as the training proceeds. To be concrete, at the very start of training, the predictions are extremely noisy or even random. And then" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16190" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "at later stages, the predictions become more and more reliable. So the uncertain predictions at different training stages should not be treated equally. Therefore, we further design two reweighting principles for the two concerns." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 122, + 289, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 122, + 289, + 172 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 289, + 172 + ], + "type": "text", + "content": "Principle 1: Reweighting with image-based varied reliability. According to the above intuition, we directly reweight the uncertain loss of each uncertain image by its top-1 confidence " + }, + { + "bbox": [ + 46, + 122, + 289, + 172 + ], + "type": "inline_equation", + "content": "\\xi(p_k^w)" + }, + { + "bbox": [ + 46, + 122, + 289, + 172 + ], + "type": "text", + "content": " in the original class space, which is:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 175, + 288, + 207 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 175, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 61, + 175, + 288, + 207 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} ^ {1} = \\frac {1}{B _ {u}} \\sum_ {k = 1} ^ {B _ {u}} \\mathbb {1} \\left(\\xi \\left(p _ {k} ^ {w}\\right) < \\tau\\right) \\cdot \\hat {\\mathrm {H}} \\left(\\hat {p} _ {k} ^ {w}, \\hat {p} _ {k} ^ {s}\\right) \\cdot \\xi \\left(p _ {k} ^ {w}\\right). \\tag {6}", + "image_path": "e5a97d0fde124f459c726235190faaffda5c790d5b17a0e05d913ed2df651a20.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 209, + 289, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 289, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 289, + 258 + ], + "type": "text", + "content": "We do not use the top-1 confidence " + }, + { + "bbox": [ + 46, + 209, + 289, + 258 + ], + "type": "inline_equation", + "content": "\\xi(\\hat{p}_k^w)" + }, + { + "bbox": [ + 46, + 209, + 289, + 258 + ], + "type": "text", + "content": " in the shrunk space as the weight, because after shrinking, this value of different predictions is very close to each other. So generally, " + }, + { + "bbox": [ + 46, + 209, + 289, + 258 + ], + "type": "inline_equation", + "content": "\\xi(p_k^w)" + }, + { + "bbox": [ + 46, + 209, + 289, + 258 + ], + "type": "text", + "content": " is more discriminative than " + }, + { + "bbox": [ + 46, + 209, + 289, + 258 + ], + "type": "inline_equation", + "content": "\\xi(\\hat{p}_k^w)" + }, + { + "bbox": [ + 46, + 209, + 289, + 258 + ], + "type": "text", + "content": " as the weight." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 261, + 289, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 261, + 289, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 289, + 464 + ], + "type": "text", + "content": "Principle 2: Reweighting with model-based gradually improved state. One naive solution is to linearly increase the loss weight of " + }, + { + "bbox": [ + 46, + 261, + 289, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 46, + 261, + 289, + 464 + ], + "type": "text", + "content": " from 0 at the beginning to " + }, + { + "bbox": [ + 46, + 261, + 289, + 464 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 261, + 289, + 464 + ], + "type": "text", + "content": " at iteration " + }, + { + "bbox": [ + 46, + 261, + 289, + 464 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 261, + 289, + 464 + ], + "type": "text", + "content": ", and then keep " + }, + { + "bbox": [ + 46, + 261, + 289, + 464 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 261, + 289, + 464 + ], + "type": "text", + "content": " until the end. However, this practice has two severe disadvantages. First, the two additional hyperparameters " + }, + { + "bbox": [ + 46, + 261, + 289, + 464 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 261, + 289, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 261, + 289, + 464 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 261, + 289, + 464 + ], + "type": "text", + "content": " are not easy to determine, and could be sensitive. More importantly, the linear scheduling criterion simply assumes the model state also improves linearly. Indeed, it can not be true. Thus, we here present a more promising principle to perform reweighting, that is free from any extra hyper-parameters. It can adjust the loss weight according to the model state in a fully adaptive fashion. To be specific, we use the certain ratio of the unlabeled set as an indicator for the model state. The certain ratio is traced at each iteration and accumulated globally in an exponential moving average (EMA) manner. Formally, the certain ratio in a single mini-batch is given by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 469, + 287, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 469, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 107, + 469, + 287, + 502 + ], + "type": "interline_equation", + "content": "m = \\frac {1}{B _ {u}} \\sum_ {k = 1} ^ {B _ {u}} \\mathbb {1} \\left(\\xi \\left(p _ {k} ^ {w}\\right) \\geq \\tau\\right). \\tag {7}", + "image_path": "ce0c96f42c33cd0018a2f902abd463d781c335af60b905cfc45c031be396baa3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 502, + 289, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 289, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 289, + 525 + ], + "type": "text", + "content": "The global certain ratio " + }, + { + "bbox": [ + 47, + 502, + 289, + 525 + ], + "type": "inline_equation", + "content": "m^g" + }, + { + "bbox": [ + 47, + 502, + 289, + 525 + ], + "type": "text", + "content": " is initialized as 0, and accumulated at each training iteration by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 531, + 287, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 531, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 107, + 531, + 287, + 544 + ], + "type": "interline_equation", + "content": "m ^ {g} \\leftarrow \\gamma \\cdot m ^ {g} + (1 - \\gamma) \\cdot m, \\tag {8}", + "image_path": "95c3d39c57d10b07f9f9f5669420afa38f3f74c4fefbda16a3146ccb17d37089.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 550, + 287, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 550, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 46, + 550, + 287, + 585 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 550, + 287, + 585 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 550, + 287, + 585 + ], + "type": "text", + "content": " is the momentum coefficient. It is a hyper-parameter already defined in FixMatch (baseline), where it is used to update the teacher parameters for final evaluation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 586, + 288, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 288, + 621 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 288, + 621 + ], + "type": "text", + "content": "Obviously, " + }, + { + "bbox": [ + 46, + 586, + 288, + 621 + ], + "type": "inline_equation", + "content": "m^g" + }, + { + "bbox": [ + 46, + 586, + 288, + 621 + ], + "type": "text", + "content": " falls between 0 and 1. And it will approximately increase from 0 to a nearly saturated value. Then, the reweighted uncertain loss is given by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 72, + 622, + 287, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 622, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 72, + 622, + 287, + 654 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} ^ {2} = m ^ {g} \\cdot \\frac {1}{B _ {u}} \\sum_ {k = 1} ^ {B _ {u}} \\mathbb {1} \\left(\\xi \\left(p _ {k} ^ {w}\\right) < \\tau\\right) \\cdot \\hat {\\mathrm {H}} \\left(\\hat {p} _ {k} ^ {w}, \\hat {p} _ {k} ^ {s}\\right). \\tag {9}", + "image_path": "4462420452c21fe7bdd2e7430a75fb4b5ddd39779f20ab2df166b9ac1ed9b661.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 654, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 677 + ], + "type": "text", + "content": "Integrating the above two intuitions and principles, the final reweighted uncertain loss will be:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 52, + 683, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 683, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 52, + 683, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = m ^ {g} \\cdot \\frac {1}{B _ {u}} \\sum_ {k = 1} ^ {B _ {u}} \\mathbb {1} \\left(\\xi \\left(p _ {k} ^ {w}\\right) < \\tau\\right) \\cdot \\hat {\\mathrm {H}} \\left(\\hat {p} _ {k} ^ {w}, \\hat {p} _ {k} ^ {s}\\right) \\cdot \\xi \\left(p _ {k} ^ {w}\\right). \\tag {10}", + "image_path": "08cb28704a6ef7f6a2e613d83c8ad07264c8625f334f9e6788f4aca00113679f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 72, + 375, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 375, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 375, + 84 + ], + "type": "text", + "content": "3.4. Summary" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 92, + 547, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 92, + 547, + 129 + ], + "spans": [ + { + "bbox": [ + 305, + 92, + 547, + 129 + ], + "type": "text", + "content": "To summarize, the final loss in a mini-batch is a combination of the supervised loss " + }, + { + "bbox": [ + 305, + 92, + 547, + 129 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_x)" + }, + { + "bbox": [ + 305, + 92, + 547, + 129 + ], + "type": "text", + "content": ", certain loss " + }, + { + "bbox": [ + 305, + 92, + 547, + 129 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_u" + }, + { + "bbox": [ + 305, + 92, + 547, + 129 + ], + "type": "text", + "content": ", Eq. (1)), and uncertain loss in the shrunk class space " + }, + { + "bbox": [ + 305, + 92, + 547, + 129 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_s" + }, + { + "bbox": [ + 305, + 92, + 547, + 129 + ], + "type": "text", + "content": ", Eq. (10)):" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 369, + 142, + 546, + 155 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 142, + 546, + 155 + ], + "spans": [ + { + "bbox": [ + 369, + 142, + 546, + 155 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {x} + \\lambda_ {u} \\cdot \\left(\\mathcal {L} _ {u} + \\mathcal {L} _ {s}\\right). \\tag {11}", + "image_path": "f44c42bb9434dffa5d1c16fab73110bb82bd8a0e19716c2666c425ca29601f25.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 170, + 547, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 170, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 306, + 170, + 547, + 194 + ], + "type": "text", + "content": "We do not carefully fine-tune the fusion weight between " + }, + { + "bbox": [ + 306, + 170, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 306, + 170, + 547, + 194 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 170, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 306, + 170, + 547, + 194 + ], + "type": "text", + "content": ", but use 1:1 by default to avoid hyper-parameters." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 208, + 382, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 208, + 382, + 219 + ], + "spans": [ + { + "bbox": [ + 306, + 208, + 382, + 219 + ], + "type": "text", + "content": "3.5. Discussions" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 228, + 547, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 335 + ], + "type": "text", + "content": "Our uncertain loss in the shrunk space owns two properties: informative and also safe. The first property is because we manage to find the largest shrunk space that reaches the confidence threshold. Besides, we also adopt the weak-to-strong consistency regularization to pose a challenging optimization target. Both constraints ensure the learning in the shrunk space is not trivial and still quite informative. Next, we wish to explain the second property \"safe\", especially about how we avoid noise in the shrunk space." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 338, + 547, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 547, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 547, + 493 + ], + "type": "text", + "content": "Noises in pseudo labels distinguish the semi-supervised paradigm from the fully-supervised one. So designing a safe optimization target for unlabeled data is crucial. Typically, the cross entropy loss will maximize the softmax probability " + }, + { + "bbox": [ + 304, + 338, + 547, + 493 + ], + "type": "inline_equation", + "content": "\\exp (s_t) / \\sum_{i = 1}^{C}\\exp (s_i)\\rightarrow 1" + }, + { + "bbox": [ + 304, + 338, + 547, + 493 + ], + "type": "text", + "content": " for the target class " + }, + { + "bbox": [ + 304, + 338, + 547, + 493 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 338, + 547, + 493 + ], + "type": "text", + "content": ". It inevitably suppresses scores of all other classes except " + }, + { + "bbox": [ + 304, + 338, + 547, + 493 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 338, + 547, + 493 + ], + "type": "text", + "content": ". However, the true class could not be " + }, + { + "bbox": [ + 304, + 338, + 547, + 493 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 338, + 547, + 493 + ], + "type": "text", + "content": ", and may be the " + }, + { + "bbox": [ + 304, + 338, + 547, + 493 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 304, + 338, + 547, + 493 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 338, + 547, + 493 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 304, + 338, + 547, + 493 + ], + "type": "text", + "content": " largest class, etc., which is wrongly restrained. This is frequently observed when the confidence of class " + }, + { + "bbox": [ + 304, + 338, + 547, + 493 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 338, + 547, + 493 + ], + "type": "text", + "content": " is not high enough. As a milder alternative, the soft labeling still encounters a similar problem. In contrast, our shrunk target directly excludes these confusion classes. So only the almost unlikely classes are suppressed." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 495, + 547, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 615 + ], + "type": "text", + "content": "It is worth noting, even if the predicted top-1 class is not in line with the human label, it is probably one of the closest semantics, e.g., belonging to the same superclass as the groundtruth label, as shown in Fig. 2. Thus, contrasting these relevant classes against other less likely classes with our noise-robust shrunk loss is still beneficial. The model is encouraged to make discriminative predictions closer to top classes compared to tail classes. In addition, our disentangled auxiliary head can effectively leverage such supervision while not affecting the main classification tasks [7]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 632, + 382, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 632, + 382, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 632, + 382, + 645 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "In this section, we first describe the implementation details of our framework. Then, we compare our ShrinkMatch with previous state-of-the-art methods (SOTAs) under extensive evaluation protocols. Lastly, we conduct comprehensive ablation studies on each component to validate the necessity." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16191" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 286, + 148 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 148 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 148 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 148 + ], + "type": "table", + "html": "
Seed01234Mean
SimMatch [45]95.3495.1692.6393.7695.1094.39
ShrinkMatch | 4095.0994.6695.1294.7894.9594.92
SimMatch [45]95.5895.5095.3494.0695.2695.15
ShrinkMatch | 25095.3995.4495.3694.7695.3595.26
", + "image_path": "9b06fe82cd066293cf991aa656653e12a30db08a90cc6308e9a97463ebd5ca0a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 198, + 287, + 277 + ], + "blocks": [ + { + "bbox": [ + 47, + 155, + 287, + 191 + ], + "lines": [ + { + "bbox": [ + 47, + 155, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 47, + 155, + 287, + 191 + ], + "type": "text", + "content": "Table 1: Comparison with SOTAs on CIFAR-10. The same seed ensures exactly the same data split. The 400 or 2500 denotes the number of labels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 198, + 287, + 277 + ], + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 277 + ], + "type": "table", + "html": "
Seed01234Mean
SimMatch [45]62.0660.1959.8964.8863.9262.19
ShrinkMatch | 40065.0063.4763.7766.4264.5264.64
SimMatch [45]74.6475.1974.5375.0375.2474.93
ShrinkMatch | 250075.0075.1174.5474.7874.7274.83
", + "image_path": "f93caf2b105632bb6645ae0dc5e6db66b25e5ecfa831cbbe932f02fe034d0b11.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 312, + 180, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 180, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 180, + 324 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 334, + 288, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 288, + 394 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 288, + 394 + ], + "type": "text", + "content": "Baselines. We use FixMatch + distribution alignment (DA) as our baseline on all datasets except ImageNet and SVHN. On ImageNet, we build our method on SimMatch. On SVHN, we discard DA. To be more convincing, we adopt the same codebase as our compared methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "text", + "content": "Hyper-parameters. Following prior arts, Wide ResNet-28-2 [40], WRN-28-8, WRN-37-2, and WRN-28-2 are used for CIFAR-10, CIFAR-100, STL-10, and SVHN respectively. A ResNet-50 [13] is used for ImageNet. The auxiliary head " + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "inline_equation", + "content": "h_{aux}" + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "text", + "content": " for uncertain samples is a 3-layer MLP. On the ImageNet, we set " + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "inline_equation", + "content": "B_u = 64 \\times 5" + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "text", + "content": ", but for other datasets, " + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "inline_equation", + "content": "B_u = 64 \\times 7" + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "text", + "content": ". The labeled batch size is 64 for all datasets. On the ImageNet, the model is trained for 400 epochs, while on the others, it is trained for " + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "inline_equation", + "content": "2^{20}" + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "text", + "content": " iterations. The initial learning rate is set as 0.03 for all datasets with a cosine scheduler. Specifically, on the ImageNet, it is warmed up for 5 epochs. The consistency regularization " + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "text", + "content": " on STL-10 and SVHN is a hard cross entropy (CE) loss, while on the CIFAR-10/100 and ImageNet, following the SimMatch and CoMatch [20], it is modified to a soft CE loss. The " + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "inline_equation", + "content": "\\hat{H}" + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "text", + "content": " in our proposed " + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "text", + "content": " is simply a hard CE loss. The weight " + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "inline_equation", + "content": "\\lambda_u" + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "text", + "content": " for the two unsupervised losses is set as 10 on the ImageNet and 1 on others. The confidence threshold " + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "text", + "content": " is 0.7 on the ImageNet and 0.95 for others. The shared momentum coefficient " + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "text", + "content": " between our global certain ratio " + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "inline_equation", + "content": "m^g" + }, + { + "bbox": [ + 46, + 397, + 288, + 661 + ], + "type": "text", + "content": " and the teacher parameters is 0.999. Following the common practice, the teacher model is only maintained for final evaluation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 670, + 195, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 195, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 195, + 682 + ], + "type": "text", + "content": "4.2. CIFAR-10 and CIFAR-100" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "The CIFAR dataset is composed of 50000/10000 training/validation images of size " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": ". The CIFAR-10 con" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 543, + 165 + ], + "blocks": [ + { + "bbox": [ + 64, + 283, + 269, + 295 + ], + "lines": [ + { + "bbox": [ + 64, + 283, + 269, + 295 + ], + "spans": [ + { + "bbox": [ + 64, + 283, + 269, + 295 + ], + "type": "text", + "content": "Table 2: Comparison with SOTAs on CIFAR-100." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 70, + 543, + 165 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 543, + 165 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 543, + 165 + ], + "type": "table", + "html": "
Seed012Mean
FixMatch [28]65.8567.9458.3064.03
FlexMatch [42]76.7168.2867.5570.85
ShrinkMatch | 4085.7585.6486.5585.98
FixMatch [28]90.9188.7190.9490.19
FlexMatch [42]91.3592.2991.6691.77
ShrinkMatch | 25091.1392.4391.1091.55
", + "image_path": "060124de43d2796039209300d53d5e2f2c8e2fda878526b51bc4ee3b875137ca.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 191, + 543, + 286 + ], + "blocks": [ + { + "bbox": [ + 331, + 171, + 519, + 183 + ], + "lines": [ + { + "bbox": [ + 331, + 171, + 519, + 183 + ], + "spans": [ + { + "bbox": [ + 331, + 171, + 519, + 183 + ], + "type": "text", + "content": "Table 3: Comparison with SOTAs on STL-10." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 191, + 543, + 286 + ], + "lines": [ + { + "bbox": [ + 307, + 191, + 543, + 286 + ], + "spans": [ + { + "bbox": [ + 307, + 191, + 543, + 286 + ], + "type": "table", + "html": "
Seed012Mean
FixMatch [28]94.5396.9097.1496.19
FlexMatch [42]89.1989.9396.3291.81
ShrinkMatch | 4097.9697.8196.7097.49
FixMatch [28]98.0097.9997.9497.98
FlexMatch [42]91.7691.8396.6593.41
ShrinkMatch | 25098.0898.0697.9898.04
", + "image_path": "7db40f2d93af10e02d2d7154bf597f4190a2f6a89eefbae559129bec3046e78b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 334, + 293, + 517, + 304 + ], + "lines": [ + { + "bbox": [ + 334, + 293, + 517, + 304 + ], + "spans": [ + { + "bbox": [ + 334, + 293, + 517, + 304 + ], + "type": "text", + "content": "Table 4: Comparison with SOTAs on SVHN." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 323, + 547, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 547, + 443 + ], + "type": "text", + "content": "tains 10 balanced classes to recognize, while CIFAR-10 is much more challenging, containing 100 classes. We run our ShrinkMatch on five different seeds and report each result as well as the average result. On the CIFAR-10 of Tab. 1 where performance is almost saturated, we still obtain a nontrivial improvement of " + }, + { + "bbox": [ + 304, + 323, + 547, + 443 + ], + "type": "inline_equation", + "content": "0.53\\%" + }, + { + "bbox": [ + 304, + 323, + 547, + 443 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 323, + 547, + 443 + ], + "type": "inline_equation", + "content": "94.39\\% \\to 94.92\\%" + }, + { + "bbox": [ + 304, + 323, + 547, + 443 + ], + "type": "text", + "content": ") with 4 labels per class (400 labels in total). In addition, notably, as shown in Tab. 2, with only 4 labels per class on the challenging CIFAR-100, our ShrinkMatch remarkably outperforms SimMatch by " + }, + { + "bbox": [ + 304, + 323, + 547, + 443 + ], + "type": "inline_equation", + "content": "2.45\\%" + }, + { + "bbox": [ + 304, + 323, + 547, + 443 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 323, + 547, + 443 + ], + "type": "inline_equation", + "content": "62.19\\% \\to 64.64\\%" + }, + { + "bbox": [ + 304, + 323, + 547, + 443 + ], + "type": "text", + "content": ") on average." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 456, + 418, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 418, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 418, + 467 + ], + "type": "text", + "content": "4.3. STL-10 and SVHN" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 475, + 546, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 546, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 546, + 548 + ], + "type": "text", + "content": "The STL-10 dataset originally consists of 5K/100K/8K labeled/unlabeled/validation images. We follow FlexMatch to only select a subset of 40/250 labeled images, while the unlabeled set remains unchanged. As shown in Tab. 3, with 4 labels per class, our ShrinkMatch surpasses FlexMatch tremendously from " + }, + { + "bbox": [ + 304, + 475, + 546, + 548 + ], + "type": "inline_equation", + "content": "70.85\\%" + }, + { + "bbox": [ + 304, + 475, + 546, + 548 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 475, + 546, + 548 + ], + "type": "inline_equation", + "content": "85.98\\%" + }, + { + "bbox": [ + 304, + 475, + 546, + 548 + ], + "type": "inline_equation", + "content": "(+15.13\\%)" + }, + { + "bbox": [ + 304, + 475, + 546, + 548 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 548, + 547, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 547, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 547, + 621 + ], + "type": "text", + "content": "The SVHN dataset is suitable to reveal the ability of different semi-supervised algorithms in the presence of class imbalance issue. As demonstrated in Tab. 4, although the performance of the original FixMatch (our baseline) has almost touched the upper bound, we still further boost it from " + }, + { + "bbox": [ + 304, + 548, + 547, + 621 + ], + "type": "inline_equation", + "content": "96.19\\%" + }, + { + "bbox": [ + 304, + 548, + 547, + 621 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 548, + 547, + 621 + ], + "type": "inline_equation", + "content": "97.49\\%" + }, + { + "bbox": [ + 304, + 548, + 547, + 621 + ], + "type": "inline_equation", + "content": "(+1.30\\%)" + }, + { + "bbox": [ + 304, + 548, + 547, + 621 + ], + "type": "text", + "content": " with 4 labels per class." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 633, + 392, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 392, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 392, + 647 + ], + "type": "text", + "content": "4.4. ImageNet-1K" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 654, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 546, + 715 + ], + "type": "text", + "content": "The ImageNet dataset is rather challenging, containing 1.28M/50K training/validation images of 1000 classes. We exactly follow the codebase of SimMatch [45]. As shown in Tab. 5, our ShrinkMatch further boosts previous best results on both settings of " + }, + { + "bbox": [ + 304, + 654, + 546, + 715 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 654, + 546, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 654, + 546, + 715 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 654, + 546, + 715 + ], + "type": "text", + "content": " labeled images." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16192" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 70, + 520, + 215 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 520, + 215 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 520, + 215 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 520, + 215 + ], + "type": "table", + "html": "
Pre-trainingMethodEpochsParams (train / test)Top-1Top-5
1%10%1%10%
SimCLR v2 [8]80034.2M / 25.6M57.968.482.589.2
SwAV [6]Fine-tune80030.4M / 25.6M53.970.278.589.9
WCL [44]80034.2M / 25.6M65.072.086.391.2
MoCo v2 [9]Fine-tune80030.0M / 25.6M49.866.177.287.9
CoMatch [20]120030.0M / 25.6M67.173.787.191.4
MoCo-EMAN [5]FixMatch-EMAN [5]110030.0M / 25.6M63.074.083.490.9
NoneCoMatch [20]40030.0M / 25.6M66.073.686.491.6
SimMatch† [45]40030.0M / 25.6M67.074.186.991.5
ShrinkMatch40031.8M / 25.6M67.574.587.491.9
", + "image_path": "e1afa3d50119300c4ef15f2032ec884bd3dd44b08c1dee1935e68bf5a208f884.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 243, + 210, + 347 + ], + "blocks": [ + { + "bbox": [ + 52, + 243, + 210, + 347 + ], + "lines": [ + { + "bbox": [ + 52, + 243, + 210, + 347 + ], + "spans": [ + { + "bbox": [ + 52, + 243, + 210, + 347 + ], + "type": "image", + "image_path": "c87cc982bf63139eb4e6b1c2d4e25af9ce32ab22746ff03050467e7f250b60cb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 351, + 208, + 362 + ], + "lines": [ + { + "bbox": [ + 54, + 351, + 208, + 362 + ], + "spans": [ + { + "bbox": [ + 54, + 351, + 208, + 362 + ], + "type": "text", + "content": "(a) Number of removed confusion classes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 217, + 242, + 373, + 347 + ], + "blocks": [ + { + "bbox": [ + 217, + 242, + 373, + 347 + ], + "lines": [ + { + "bbox": [ + 217, + 242, + 373, + 347 + ], + "spans": [ + { + "bbox": [ + 217, + 242, + 373, + 347 + ], + "type": "image", + "image_path": "6d4ac0e5765c69a424b6aaa025f86a60ac283a5d0a505573764e01c53cb7f3fa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 367, + 546, + 392 + ], + "lines": [ + { + "bbox": [ + 46, + 367, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 546, + 392 + ], + "type": "text", + "content": "Figure 4: (a) The number of removed confusion classes for each uncertain image as the training proceeds. (b) Value of different loss functions, i.e., Eq. (1), Eq. (6), and Eq. (10). (c) The ratio of uncertain samples (certainty " + }, + { + "bbox": [ + 46, + 367, + 546, + 392 + ], + "type": "inline_equation", + "content": "< \\tau" + }, + { + "bbox": [ + 46, + 367, + 546, + 392 + ], + "type": "text", + "content": ") in each mini-batch." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 380, + 243, + 538, + 347 + ], + "blocks": [ + { + "bbox": [ + 219, + 352, + 370, + 362 + ], + "lines": [ + { + "bbox": [ + 219, + 352, + 370, + 362 + ], + "spans": [ + { + "bbox": [ + 219, + 352, + 370, + 362 + ], + "type": "text", + "content": "(b) Certain loss " + }, + { + "bbox": [ + 219, + 352, + 370, + 362 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 219, + 352, + 370, + 362 + ], + "type": "text", + "content": " and uncertain loss " + }, + { + "bbox": [ + 219, + 352, + 370, + 362 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 219, + 352, + 370, + 362 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 380, + 243, + 538, + 347 + ], + "lines": [ + { + "bbox": [ + 380, + 243, + 538, + 347 + ], + "spans": [ + { + "bbox": [ + 380, + 243, + 538, + 347 + ], + "type": "image", + "image_path": "99f0782972656e026945d310b4b2bac11410c18fb6fcae5fbdcba80b0ef1f0d7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 352, + 515, + 362 + ], + "lines": [ + { + "bbox": [ + 403, + 352, + 515, + 362 + ], + "spans": [ + { + "bbox": [ + 403, + 352, + 515, + 362 + ], + "type": "text", + "content": "(c) Ratio of uncertain samples." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 49, + 404, + 284, + 450 + ], + "blocks": [ + { + "bbox": [ + 65, + 220, + 526, + 232 + ], + "lines": [ + { + "bbox": [ + 65, + 220, + 526, + 232 + ], + "spans": [ + { + "bbox": [ + 65, + 220, + 526, + 232 + ], + "type": "text", + "content": "Table 5: Accuracy (%) on the ImageNet-1K with " + }, + { + "bbox": [ + 65, + 220, + 526, + 232 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 65, + 220, + 526, + 232 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 65, + 220, + 526, + 232 + ], + "type": "inline_equation", + "content": "{10}\\%" + }, + { + "bbox": [ + 65, + 220, + 526, + 232 + ], + "type": "text", + "content": " labeled images. †: Reproduced in our environment." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 404, + 284, + 450 + ], + "lines": [ + { + "bbox": [ + 49, + 404, + 284, + 450 + ], + "spans": [ + { + "bbox": [ + 49, + 404, + 284, + 450 + ], + "type": "table", + "html": "
Seed01234Mean
Baseline64.4057.5163.0764.7762.5362.46
ShrinkMatch65.0063.4763.7766.4264.5264.64
", + "image_path": "2f29df7cb38d810a87ec7d7a76f90fcc926129aea18c7f8f1beadf5dd8256cb1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 484, + 147, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 484, + 147, + 496 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 147, + 496 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 503, + 288, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 503, + 288, + 527 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 288, + 527 + ], + "type": "text", + "content": "Unless otherwise specified, we conduct our ablation studies on CIFAR-100 with 4 labels per class." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 530, + 288, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 288, + 602 + ], + "type": "text", + "content": "Effectiveness of our holistic ShrinkMatch. We first view our ShrinkMatch as a holistic component added to our baseline. As displayed in Tab. 6 for CIFAR-100, our ShrinkMatch boosts the strong baseline significantly by " + }, + { + "bbox": [ + 46, + 530, + 288, + 602 + ], + "type": "inline_equation", + "content": "2.18\\%" + }, + { + "bbox": [ + 46, + 530, + 288, + 602 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 530, + 288, + 602 + ], + "type": "inline_equation", + "content": "62.46\\% \\rightarrow 64.64\\%" + }, + { + "bbox": [ + 46, + 530, + 288, + 602 + ], + "type": "text", + "content": "). And on the STL-10 of Tab. 7, the baseline is also improved evidently from " + }, + { + "bbox": [ + 46, + 530, + 288, + 602 + ], + "type": "inline_equation", + "content": "84.85\\%" + }, + { + "bbox": [ + 46, + 530, + 288, + 602 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 530, + 288, + 602 + ], + "type": "inline_equation", + "content": "85.98\\%" + }, + { + "bbox": [ + 46, + 530, + 288, + 602 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 530, + 288, + 602 + ], + "type": "inline_equation", + "content": "+1.13\\%" + }, + { + "bbox": [ + 46, + 530, + 288, + 602 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "type": "text", + "content": "Effectiveness of two reweighting principles. In Sec. 3.3, considering the varied prediction reliability of different uncertain samples and different training stages, we propose two principles to reweight our uncertain loss " + }, + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "type": "text", + "content": ". So here we carefully examine their necessity in Tab. 8. It can be observed that the two principles are both beneficial, jointly improving the basic shrinking practice from " + }, + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "type": "inline_equation", + "content": "63.16\\%" + }, + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "type": "inline_equation", + "content": "64.24\\%" + }, + { + "bbox": [ + 46, + 605, + 289, + 714 + ], + "type": "text", + "content": ". We also attempt a simple alternative of linearly increasing the uncertain loss weight from 0 (start) to 1 (end). But as evi" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 315, + 404, + 535, + 451 + ], + "blocks": [ + { + "bbox": [ + 49, + 456, + 284, + 468 + ], + "lines": [ + { + "bbox": [ + 49, + 456, + 284, + 468 + ], + "spans": [ + { + "bbox": [ + 49, + 456, + 284, + 468 + ], + "type": "text", + "content": "Table 6: Ablation studies of ShrinkMatch on CIFAR-100." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 404, + 535, + 451 + ], + "lines": [ + { + "bbox": [ + 315, + 404, + 535, + 451 + ], + "spans": [ + { + "bbox": [ + 315, + 404, + 535, + 451 + ], + "type": "table", + "html": "
Seed012Mean
Baseline84.8684.2185.4784.85
ShrinkMatch85.7585.6486.5585.98
", + "image_path": "142460578d8a188a0b6a38c3dcbedb9609193744459980548a108b6eefdb8fb7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 456, + 535, + 468 + ], + "lines": [ + { + "bbox": [ + 316, + 456, + 535, + 468 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 535, + 468 + ], + "type": "text", + "content": "Table 7: Ablation studies of ShrinkMatch on STL-10." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 487, + 545, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 584 + ], + "type": "text", + "content": "denced in the Exp 3 & 4 of Tab. 8, it is obviously inferior to our designed reweighting strategy. We visualize our " + }, + { + "bbox": [ + 304, + 487, + 545, + 584 + ], + "type": "inline_equation", + "content": "m^g" + }, + { + "bbox": [ + 304, + 487, + 545, + 584 + ], + "type": "text", + "content": " as training proceeds on ImageNet in Fig. 5. It looks essentially different from linear scheduling. Moreover, we visualize the uncertain loss value with or without the second reweighting principle in Fig. 4b. After reweighting, the uncertain loss is of a similar magnitude as the certain loss, which can avoid dominating the gradient at early training stages." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 588, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 547, + 696 + ], + "type": "text", + "content": "The number of removed confusion classes. As introduced before, our approach detects and removes the confusion classes for the top-1 class in a fully automatic and instance-adaptive fashion. So we visualize the number of removed confusion classes for each uncertain sample in Fig. 4a. At the very start of training, model predictions are almost uniform, so abundant classes are removed for an uncertain sample. But as the training proceeds, much fewer confusion classes need to be removed to form a confident shrunk class space." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "Hard label or soft label in the shrunk class space? By" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16193" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 288, + 149 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 149 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 149 + ], + "type": "table", + "html": "
ExpPrinciple 1Principle 2LSSeed 0Seed 1Mean
165.6260.7063.16
2✓65.6862.0863.88
3✓65.0962.9064.00
4✓63.0159.4561.23
5✓✓65.0063.4764.24
", + "image_path": "8aa678a1b0c5f7fe55e9163a040285f2ab1095b1b0ebdeb2cfa7deeb1cb02df9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 210, + 285, + 352 + ], + "blocks": [ + { + "bbox": [ + 46, + 156, + 288, + 204 + ], + "lines": [ + { + "bbox": [ + 46, + 156, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 288, + 204 + ], + "type": "text", + "content": "Table 8: Ablation studies on the effectiveness of the two reweighting principles. Due to randomness, the \"Mean\" results are more convincing. LS is short for linear scheduling, which linearly increases the loss weight." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 210, + 285, + 352 + ], + "lines": [ + { + "bbox": [ + 48, + 210, + 285, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 285, + 352 + ], + "type": "table", + "html": "
ExpShrinkAux HeadLabelSeed 0Seed 1Mean
1Our Strong Baseline64.4057.5160.96
2✓✓H65.0063.4764.24
3✓✓S62.9259.7661.34
4✓H62.4358.8660.65
5✓S60.9260.2260.57
6✓H64.2159.1761.69
7✓S64.7560.8962.82
8H52.9051.8852.39
9S54.1055.7554.93
", + "image_path": "0bc0dd1bdc2a31d1e86aa6478e14619bf6261b6934c02e0052265ba1d105939f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 446, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 446, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 287, + 519 + ], + "type": "text", + "content": "default, we use the hard label in the shrunk space, following FixMatch. But we also ablate the choice of using the soft label as supervision. As shown in Exp 2 & 3 of Tab. 9, the hard label is much better than the soft label. We conjecture that this is because we have shrunk the class space to a safe one. The soft label will incur some unnecessary noise." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "content": "Effects of the auxiliary head. To prevent the main classifier from being affected, we use an auxiliary head to learn the uncertain samples in the shrunk class space only for discriminative representations. So we first provide the results when still using the main head for the shrunk space. As shown in Exp 2 & 4 of Tab. 9, the auxiliary head is indispensable for our shrinking practice. So we continue to figure out whether the improvement of our ShrinkMatch merely comes from the auxiliary head. We attempt to use the auxiliary head to directly learn the uncertain samples in the original space without shrinking. It can be concluded from Exp 2 & 6 & 7 of Tab. 9, the original space is indeed much inferior to our shrunk space for uncertain samples, although the auxiliary head still helps the baseline to some extent. Lastly, in Tab. 9, we also include two additional results for readers, i.e., directly learning uncertain samples without shrinking or" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 70, + 541, + 180 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 541, + 180 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 541, + 180 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 541, + 180 + ], + "type": "image", + "image_path": "101689e4fbc26cfa05b514b02839eedd0961c52b84eaa44148e9c2bdecf24048.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 183, + 534, + 196 + ], + "lines": [ + { + "bbox": [ + 316, + 183, + 534, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 183, + 534, + 196 + ], + "type": "text", + "content": "Figure 5: Visualization of our global certain ratio " + }, + { + "bbox": [ + 316, + 183, + 534, + 196 + ], + "type": "inline_equation", + "content": "m^g" + }, + { + "bbox": [ + 316, + 183, + 534, + 196 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 201, + 541, + 236 + ], + "blocks": [ + { + "bbox": [ + 46, + 358, + 288, + 431 + ], + "lines": [ + { + "bbox": [ + 46, + 358, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 288, + 431 + ], + "type": "text", + "content": "Table 9: Ablation studies on different options to learn uncertain sample. (1) whether to shrink the class space (Shrink), (2) whether to use an auxiliary head (Aux Head), and (3) use hard (H) or soft label (S). Exp 1 is our baseline without using uncertain images. Exp 8 & 9 are experiments of directly learning uncertain samples, which are quite noisy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 201, + 541, + 236 + ], + "lines": [ + { + "bbox": [ + 309, + 201, + 541, + 236 + ], + "spans": [ + { + "bbox": [ + 309, + 201, + 541, + 236 + ], + "type": "table", + "html": "
Threshold0.980.950.900.800.70
ShrinkMatch65.8065.0064.5562.7459.58
", + "image_path": "6281296ced9011fa9544d89932055f909c74498bdea3e3df13c02b36c0ca38fc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 242, + 536, + 254 + ], + "lines": [ + { + "bbox": [ + 314, + 242, + 536, + 254 + ], + "spans": [ + { + "bbox": [ + 314, + 242, + 536, + 254 + ], + "type": "text", + "content": "Table 10: Ablation studies of the confidence threshold." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 270, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 545, + 293 + ], + "type": "text", + "content": "auxiliary head (Exp 8 & 9), which are quite terrible because of introducing the abundant noise to our main classifier." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 297, + 547, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 297, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 547, + 380 + ], + "type": "text", + "content": "The ratio of uncertain samples. To further justify our motivation, in Fig. 4c, we display the ratio of uncertain samples in each mini-batch. It can be seen that even in the middle of the whole training course, there are still around " + }, + { + "bbox": [ + 305, + 297, + 547, + 380 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 305, + 297, + 547, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 297, + 547, + 380 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 305, + 297, + 547, + 380 + ], + "type": "text", + "content": " uncertain samples on ImageNet and CIFAR-100 respectively. Therefore, an appropriate approach to utilizing these samples is necessary and definitely beneficial." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 383, + 547, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 383, + 547, + 455 + ], + "spans": [ + { + "bbox": [ + 305, + 383, + 547, + 455 + ], + "type": "text", + "content": "Different confidence thresholds " + }, + { + "bbox": [ + 305, + 383, + 547, + 455 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 383, + 547, + 455 + ], + "type": "text", + "content": ". We also try different thresholds for uncertain samples, as shown in Tab. 10. When increasing " + }, + { + "bbox": [ + 305, + 383, + 547, + 455 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 383, + 547, + 455 + ], + "type": "text", + "content": " from widely adopted 0.95 to 0.98 (more uncertain samples), our ShrinkMatch can be further improved. These results clearly highlight the effective and safe utilization of uncertain samples with our ShrinkMatch." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 473, + 378, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 473, + 378, + 485 + ], + "spans": [ + { + "bbox": [ + 306, + 473, + 378, + 485 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 493, + 547, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 547, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 547, + 649 + ], + "type": "text", + "content": "In this work, we aim to fully leverage the uncertain samples in semi-supervised learning. We point out that the low certainty is typically caused by a small portion of confusion classes. Invoked by this, we propose a novel method dubbed ShrinkMatch to automatically detect and remove the confusion classes to construct a shrunk class space, where the top-1 class is turned certain. A weak-to-strong consistency regularization is enforced in the confident new space. Furthermore, we design two reweighting principles for the auxiliary uncertain loss, according to the reliability of different uncertain samples and the gradually improved state of the model. Consequently, our method establishes new state-of-the-art results on widely acknowledged benchmarks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 662, + 547, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 662, + 547, + 711 + ], + "spans": [ + { + "bbox": [ + 305, + 662, + 547, + 711 + ], + "type": "text", + "content": "Acknowledgements. This work is supported in part by the National Natural Science Foundation of China (62201484, 62222604, 62192783, 62206052), HKU Startup Fund, and HKU Seed Fund for Basic Research." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16194" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Eric Arazo, Diego Ortego, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In IJCNN, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 179 + ], + "type": "text", + "content": "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. In ICCV, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 179, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 179, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 179, + 288, + 223 + ], + "type": "text", + "content": "[3] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "type": "text", + "content": "[4] Zhaowei Cai, Avinash Ravichandran, Paolo Favaro, Manchen Wang, Davide Modolo, Rahul Bhotika, Zhuowen Tu, and Stefano Soatto. Semi-supervised vision transformers at scale. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 268, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 268, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 268, + 288, + 312 + ], + "type": "text", + "content": "[5] Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Charless Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi-supervised learning. In CVPR, 2021. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 313, + 288, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 288, + 356 + ], + "type": "text", + "content": "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 357, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 390 + ], + "type": "text", + "content": "[7] Baixu Chen, Junguang Jiang, Ximei Wang, Jianmin Wang, and Mingsheng Long. Debiased pseudo labeling in self-training. In NeurIPS, 2022. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 391, + 288, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 424 + ], + "type": "text", + "content": "[8] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. In NeurIPS, 2020. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 424, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 456 + ], + "type": "text", + "content": "[9] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv:2003.04297, 2020. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 457, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 457, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 457, + 288, + 502 + ], + "type": "text", + "content": "[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 502, + 288, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 288, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 288, + 535 + ], + "type": "text", + "content": "[11] Yue Fan, Dengxin Dai, Anna Kukleva, and Bernt Schiele. Cossl: Co-learning of representation and classifier for imbalanced semi-supervised learning. In CVPR, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 536, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 568 + ], + "type": "text", + "content": "[12] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In ICML, 2016. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 569, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 288, + 602 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 602, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 635 + ], + "type": "text", + "content": "[14] Zijian Hu, Zhengyu Yang, Xuefeng Hu, and Ram Nevatia. Simple: similar pseudo label exploitation for semi-supervised classification. In CVPR, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 635, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 668 + ], + "type": "text", + "content": "[15] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? In NeurIPS, 2017. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "type": "text", + "content": "[16] Jiwon Kim, Youngjo Min, Daehwan Kim, Gyuseong Lee, Junyoung Seo, Kwangrok Ryoo, and Seungryong Kim. Conmatch: Semi-supervised learning with confidence-guided consistency regularization. In ECCV, 2022. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[17] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 96, + 547, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 118 + ], + "type": "text", + "content": "[18] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In ICLR, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 119, + 547, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 119, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 308, + 119, + 547, + 152 + ], + "type": "text", + "content": "[19] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In ICML Workshop, 2013. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 153, + 547, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 153, + 547, + 187 + ], + "spans": [ + { + "bbox": [ + 308, + 153, + 547, + 187 + ], + "type": "text", + "content": "[20] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regularization. In ICCV, 2021. 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 188, + 547, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 188, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 308, + 188, + 547, + 232 + ], + "type": "text", + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 234, + 545, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 234, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 308, + 234, + 545, + 268 + ], + "type": "text", + "content": "[22] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. TPAMI, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 269, + 547, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 269, + 547, + 313 + ], + "spans": [ + { + "bbox": [ + 308, + 269, + 547, + 313 + ], + "type": "text", + "content": "[23] Islam Nassar, Samitha Herath, Ehsan Abbasnejad, Wray Buntine, and Gholamreza Haffari. All labels are not created equal: Enhancing semi-supervision via label grouping and co-training. In CVPR, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 314, + 547, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 314, + 547, + 347 + ], + "spans": [ + { + "bbox": [ + 308, + 314, + 547, + 347 + ], + "type": "text", + "content": "[24] Yassine Ouali, Céline Hudelot, and Myriam Tami. An overview of deep semi-supervised learning. arXiv:2006.05278, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 348, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 348, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 348, + 545, + 371 + ], + "type": "text", + "content": "[25] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V Le. Meta pseudo labels. In CVPR, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 372, + 547, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 547, + 417 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 547, + 417 + ], + "type": "text", + "content": "[26] Mamshad Nayeem Rizve, Kevin Duarte, Yogesh S Rawat, and Mubarak Shah. In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. In ICLR, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 418, + 547, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 418, + 547, + 452 + ], + "spans": [ + { + "bbox": [ + 308, + 418, + 547, + 452 + ], + "type": "text", + "content": "[27] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In NeurIPS, 2016. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 452, + 547, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 452, + 547, + 507 + ], + "spans": [ + { + "bbox": [ + 308, + 452, + 547, + 507 + ], + "type": "text", + "content": "[28] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 508, + 547, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 508, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 308, + 508, + 547, + 552 + ], + "type": "text", + "content": "[29] Fariborz Taherkhani, Ali Dabouei, Sobhan Soleymani, Jeremy Dawson, and Nasser M Nasrabadi. Self-supervised Wasserstein pseudo-labeling for semi-supervised image classification. In CVPR, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 554, + 547, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 554, + 547, + 587 + ], + "spans": [ + { + "bbox": [ + 308, + 554, + 547, + 587 + ], + "type": "text", + "content": "[30] Hui Tang and Kui Jia. Towards discovering the effectiveness of moderately confident samples for semi-supervised learning. In CVPR, 2022. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 588, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 588, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 588, + 545, + 622 + ], + "type": "text", + "content": "[31] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 624, + 547, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 547, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 547, + 678 + ], + "type": "text", + "content": "[32] Guotai Wang, Wenqi Li, Michael Aertsen, Jan Deprest, Sébastien Ourselin, and Tom Vercauteren. Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolutional neural networks. Neurocomputing, 2019. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 679, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 679, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 679, + 547, + 713 + ], + "type": "text", + "content": "[33] Xudong Wang, Long Lian, and Stella X Yu. Unsupervised selective labeling for more effective semi-supervised learning. In ECCV, 2022. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16195" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 490 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "type": "text", + "content": "[34] Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V Le. Unsupervised data augmentation for consistency training. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 139 + ], + "type": "text", + "content": "[35] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 288, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 288, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 288, + 174 + ], + "type": "text", + "content": "[36] Yi Xu, Lei Shang, Jinxing Ye, Qi Qian, Yu-Feng Li, Baigui Sun, Hao Li, and Rong Jin. Dash: Semi-supervised learning with dynamic thresholding. In ICML, 2021. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 217 + ], + "type": "text", + "content": "[37] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In CVPR, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 241 + ], + "type": "text", + "content": "[38] David Yarowsky. Unsupervised word sense disambiguation, rivaling supervised methods. In ACL, 1995. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 243, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 288, + 285 + ], + "type": "text", + "content": "[39] Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In MICCAI, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 287, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 288, + 308 + ], + "type": "text", + "content": "[40] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC, 2016. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 310, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 288, + 342 + ], + "type": "text", + "content": "[41] Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. In ICCV, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 344, + 288, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 288, + 388 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 288, + 388 + ], + "type": "text", + "content": "[42] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In NeurIPS, 2021. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 389, + 288, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 389, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 48, + 389, + 288, + 422 + ], + "type": "text", + "content": "[43] Zhen Zhao, Luping Zhou, Lei Wang, Yinghuan Shi, and Yang Gao. Lassl: Label-guided self-training for semi-supervised learning. In AAAI, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 423, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 423, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 48, + 423, + 288, + 456 + ], + "type": "text", + "content": "[44] Mingkai Zheng, Fei Wang, Shan You, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Weakly supervised contrastive learning. In ICCV, 2021. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 457, + 288, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 457, + 288, + 490 + ], + "spans": [ + { + "bbox": [ + 48, + 457, + 288, + 490 + ], + "type": "text", + "content": "[45] Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In CVPR, 2022. 3, 6, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16196" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SiLK_ Simple Learned Keypoints/545a50e4-a6f6-42d1-849e-c613959c21fd_content_list.json b/2023/SiLK_ Simple Learned Keypoints/545a50e4-a6f6-42d1-849e-c613959c21fd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3248b7fed80b9b7014bb7bc835b7c0d79e683be0 --- /dev/null +++ b/2023/SiLK_ Simple Learned Keypoints/545a50e4-a6f6-42d1-849e-c613959c21fd_content_list.json @@ -0,0 +1,1646 @@ +[ + { + "type": "text", + "text": "SiLK: Simple Learned Keypoints", + "text_level": 1, + "bbox": [ + 313, + 130, + 656, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pierre Gleize", + "bbox": [ + 104, + 181, + 212, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gleize@meta.com", + "bbox": [ + 89, + 202, + 227, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weiyao Wang", + "bbox": [ + 441, + 181, + 557, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "weiyaowang@meta.com", + "bbox": [ + 411, + 202, + 584, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Matt Feiszli", + "bbox": [ + 774, + 181, + 872, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mdf@meta.com", + "bbox": [ + 769, + 202, + 879, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meta AI", + "text_level": 1, + "bbox": [ + 460, + 234, + 532, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/facebookresearch/silk", + "bbox": [ + 259, + 252, + 733, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keypoint detection & descriptors are foundational technologies for computer vision tasks like image matching, 3D reconstruction and visual odometry. Hand-engineered methods like Harris corners, SIFT, and HOG descriptors have been used for decades; more recently, there has been a trend to introduce learning in an attempt to improve keypoint detectors. On inspection however, the results are difficult to interpret; recent learning-based methods employ a vast diversity of experimental setups and design choices: empirical results are often reported using different backbones, protocols, datasets, types of supervisions or tasks. Since these differences are often coupled together, it raises a natural question on what makes a good learned keypoint detector. In this work, we revisit the design of existing keypoint detectors by deconstructing their methodologies and identifying the key components. We re-design each component from first-principle and propose Simple Learned Keypoints (SiLK) that is fully-differentiable, lightweight, and flexible. Despite its simplicity, SiLK advances new state-of-the-art on Detection Repeatability and Homography Estimation tasks on HPatches and 3D Point-Cloud Registration task on ScanNet, and achieves competitive performance to state-of-the-art on camera pose estimation in 2022 Image Matching Challenge and ScanNet.", + "bbox": [ + 75, + 335, + 473, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 724, + 209, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keypoint detection and matching is a foundational computer vision technique to obtain a sparse yet informative representation of an image. Image stitching [7, 1], SLAM [14, 34], SfM [44], camera calibrations, tracking [37], and object detection [30] are important tasks built on keypoint correspondences [32]. A good keypoint model should be able to select a subset of points useful and informative to a specific task. One typically also wants robustness of the descriptor to some set of transformations (e.g. scale, viewpoint, or lighting variation).", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8a2961e714718bde680b4d90889dbc0bee62d73af33449ea5892171ef87d0224.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 303, + 697, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cc025ba189fd5a287d8539e754d893377492c81c99bae68160a6b3d1df2ebdb3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 303, + 883, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/81ab59d8cd198820ac22a4fa0a6e9258f6db0da5104c2516e9e2e8746ecb8b75.jpg", + "image_caption": [ + "Figure 1. The top image is an example of keypoint matching under viewpoint change; correct matches are green, incorrect ones are red. The bottom image shows keypoints which are cycle-consistent by SiLK. As can be observed, SiLK has learned to find distinctive geometric features (corners, curves, intersections,..); from single non-annotated images." + ], + "image_footnote": [], + "bbox": [ + 501, + 419, + 890, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing keypoint methods come in multiple forms and flavors (Tab. 1). However, those differences are often coupled together and not controlled for, which make it challenging to identify the source of gain. Our quest to finding the key-components of good keypoint detectors led to an in-depth review of alternative approaches, as well as the design, from first principle, of SiLK (Simple Learned Keypoints): A simple self-supervised approach to learn", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22499", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "distinctive and robust keypoints from arbitrary image data in a traditional \"detect-and-describe\" framework. Despite its simplicity, SiLK is competitive or surpasses SOTA in most settings.", + "bbox": [ + 78, + 90, + 467, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, leveraging SiLK's simple one-stage training protocol and modular architecture, we are able to ablate various dimensions of detector performance for different tasks. In particular, with an eye toward real-time performance, we identify tasks where extremely lightweight backbone architectures are sufficient.", + "bbox": [ + 78, + 152, + 467, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 78, + 256, + 215, + 270 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early work focused on carefully engineered methods to identify distinctive keypoints with descriptors which are robust to changes such as viewpoint and illumination. Hand-crafted techniques like Harriscorners[20], SIFT[31], ORB[41] and others [40, 8, 26, 2, 5, 13] have used explicit geometric notions like corners, gradients, and scale-space extrema to achieve results which remain both efficient and competitive to date [44, 34].", + "bbox": [ + 78, + 281, + 467, + 401 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recent work like SuperPoint[16] chose to learn to find corners; they generate a large set of synthetic shapes with annotated corners and train a model with this ground truth. While providing compelling evidence for learned methods, their training procedure is quite complex: it contains multiple training phases, a synthetic dataset, and employs a homography adaptation trick that can be difficult to tune (see our reproduced results in Tab. 2).", + "bbox": [ + 78, + 402, + 467, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the same spirit as more recent work [17, 39, 50, 49, 21, 11, 48], SiLK aims to learn keypoints in simple end-to-end fashion, without explicitly defining them as corners.", + "bbox": [ + 78, + 523, + 467, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several attempts have been made to learn keypoints implicitly; either by the careful design of the loss [17, 39, 11, 48]; or by directly predicting the matching success of descriptors [50, 21, 49]. SiLK falls in the second category, but with slight twist (cf. Sec. 3.4).", + "bbox": [ + 78, + 568, + 467, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To learn descriptors, contrastive losses are commonly used. Similar to [50, 47], SiLK adopts a probabilistic approach by modeling the matching probabilities in a double-softmax, cycle-consistency setting and optimizes the log likelihood. The probabilistic formulation, similar to InfoNCE [35], gives us a clean way to reason about matching, and the abundant supply of hard negative examples (from pixels in the same image) makes it an attractive choice.", + "bbox": [ + 78, + 643, + 467, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Context aggregation (CA) is a recent addition to the toolkit. Initiated by SuperGlue[42], CA aims to refine or transform descriptors from a pair of images before matching them. Implemented as a GNN[43] in [42], or as a Transformer[51] in [47], CA's predictions are conditional on all descriptors from the pair of images. In other words, for an image, the descriptors will be different when matching against different images. As a result, CA needs to run on every pair of images prior to matching, as opposed to", + "bbox": [ + 78, + 765, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "running on single images. The run-time implications render CA prohibitively expensive in some applications (quadratic versus linear complexity). SiLK does not use CA, but outperforms [42] and performs competitively with [47]. Incorporating CA is optional for future works if performance is paramount and computation cost is less of a concern.", + "bbox": [ + 501, + 90, + 890, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a postprocessing after CA, SuperGlue[42] introduced the concept of differentiable optimal transport (OT) to improve matching, using the Sinkhorn algorithm[46]. LoFTR [47] leveraged OT as well, but found little difference between OT and the simpler approach of mutual nearest neighbor (MNN) in some benchmarks.", + "bbox": [ + 503, + 181, + 890, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 503, + 285, + 632, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SiLK's contribution is simplicity and flexibility. Our solution is built on the traditional approach of identifying distinctive pixels via robust local descriptors. We use modern but established techniques to learn to localize and describe keypoints given an arbitrary source of unlabeled images. Unlike classical methods, our descriptors and invariances are learned, and unlike some modern methods, there is no particular complexity in the matching process (SiLK employs only cosine distances and mutual nearest neighbor); this leaves few structural hyperparameters to tune. The simple backbone+heads design is backbone-agnostic, allowing experimentation. The annotation-free SSL approach means SiLK can be trained on any image or video dataset. Finally, a simple, one-stage training pipeline allows us to easily train and ablate different architectures, datasets, and hyperparameters for different tasks.", + "bbox": [ + 503, + 310, + 890, + 550 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SiLK is trained to identify keypoints from single grayscale images. It provides both keypoint detections (location) and keypoint descriptors (for matching). Cycle consistency is employed for descriptor learning and a binary classifier identifies distinctive keypoints at pixel-level.", + "bbox": [ + 503, + 551, + 890, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To learn descriptors, we take a source image and a transformed copy, extract descriptors for each point, and use descriptor similarity to define transition probabilities from a source location to each transformed location (and vice-versa). We optimize the descriptors to maximize cycle-consistency; i.e. we maximize the probability of a round-trip from the source to its transformed location and back.", + "bbox": [ + 503, + 628, + 890, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To locate good keypoints, we train a binary classifier to identify points which will satisfy a matching criterion. A point and its transformed copy are positives when they are mutual nearest neighbors in the sense of transition probabilities, and they are negatives otherwise. We train both the cycle-consistency and classification losses jointly.", + "bbox": [ + 503, + 733, + 890, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We provide simple pseudo-code in Fig. 2.", + "bbox": [ + 522, + 824, + 794, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Architecture", + "text_level": 1, + "bbox": [ + 503, + 847, + 632, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The SiLK architecture (Fig. 3) follows the \"detect-and-describe\" architecture originally proposed by SuperPoint", + "bbox": [ + 503, + 869, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22500", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "def loss(image_0, model):\n # get warped image and pixel correspondences\n image_1, corr_0, corr_1 = rand_homo(image_0)\n # apply image augmentations\n image_0 = augment(image_0)\n image_1 = augment(image_1)\n # extract dense descriptors and keypoints\n desc_0, kpts_0 = model(image_0)\n desc_1, kpts_1 = model(image_1)\n # compute similarity matrix\n sim_mat = cosine desc_0, desc_1)\n # compute the descriptor loss\n # using ground truth correspondences\n loss_desc = nll(sim_mat, corr_0, corr_1)\n # measure matching success\n y = is_MATCH_success(sim_mat, corr_0, corr_1)\n # compute keypoint loss\n # using current matching success\n loss_kpts = bce(kpts_0, y, corr_0)\n loss_kpts += bce(kpts_1, y, corr_1)\n return loss_desc + loss_kpts", + "guess_lang": "python", + "bbox": [ + 63, + 291, + 468, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[16]. A dense feature map is first extracted by feeding an image to an encoder backbone. The shared feature map is then fed to two heads: The keypoint head extracts the logits (used to calculate the dense keypoint probabilities), while the descriptor head extracts a dense descriptor map (used to calculate keypoint similarities). The model is backbone-agnostic and can easily be swapped.", + "bbox": [ + 75, + 678, + 468, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. High Matching Probability Defines Keypoints", + "text_level": 1, + "bbox": [ + 76, + 799, + 464, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As mentioned above, the keypoint probability estimate predicts the probability that a pixel will be correctly matched (i.e. survive a round-trip). Points with the highest likelihood of matching correctly are exactly those which we select as keypoints.", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/46fd8418c095cb447fab4c346a278b0cd3e97c17a2250fa0497f6eb9fe5e564d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Keypoint DetectionDescriptors & MatchingModel & Training
LearnedSparseCell-basedNMSSupervisionCAMatcherSupervisionInputBackboneDataE2E
SIFTNo✓No✓-NoMNN-----
SuperPoint✓✓✓✓Homo. Adapt.NoMNNRand. Homo.GreyVGGCOCONo
SuperGlue-✓---GraphNNOTSfMGreyVGGOxford & ParisNo
GLAMPoints✓✓No✓Matching Succ.NoMNN-GreyUNetSlitLamp✓
D2-Net✓✓✓✓Triplet RankingNoMNNSfMRGBVGGMegaDepth✓
R2D2✓✓✓✓Matching APNoMNNSfMRGBL2-NetAachen✓
DISK✓✓✓✓Matching Succ.NoMNNSfMRGBUNetMegaDepth✓
URR-No----MNN3D RenderingRGBResNetScanNet✓
LoFTR-No---TransformerMNN,OTSfMGreyFPNScanNet,MegaDepth✓
SiLK✓OptionalNoNoMatching Succ.NoMNNRand. Homo.GreyGenericAny Image Set✓
", + "bbox": [ + 60, + 87, + 890, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Non-exhaustive deconstruction of keypoint detectors along different dimensions. On each, SiLK adopts the simplest choice or is flexible to different choices. In particular, SiLK does not depend strongly on backbone type or training data (Tab. 7 & Tab. 9).", + "bbox": [ + 76, + 234, + 890, + 262 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a16e87181d3c935e6110d68824ccea98d1a2a0d78bee4cf8a0b6ff98b04e153e.jpg", + "image_caption": [ + "Figure 2. Pseudo-code: learning keypoints from a single image.", + "Figure 3. Architecture for SiLK" + ], + "image_footnote": [], + "bbox": [ + 537, + 284, + 856, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A common approach [16, 50, 39, 17] for obtaining keypoint probabilities is to use a softmax cell-based approach. A cell is a $N \\times N$ patch in which the probability of each cell position is determined by a local softmax. The softmax operates on $N \\times N + 1$ bins; $+1$ being the dustbin, handling the case of cells devoid of keypoints.", + "bbox": [ + 496, + 527, + 890, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SiLK's approach is equivalent to a cell-based approach with a cell size of $N = 1$ . This has several consequences. First, the softmax formulation becomes a sigmoid $\\sigma(x) = \\frac{1}{1 + e^{-x}}$ . Second, it removes the sparsity constraint that keypoints are exclusive events inside a cell. And third, this removes a free parameter (the cell size $N$ ) that we do not have to later tune.", + "bbox": [ + 496, + 619, + 890, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the same spirit, SiLK does not use NMS during inference. Even though NMS is an established pruning technique that aims to spread out keypoints [16], we find SiLK doesn't need NMS to perform (cf. Tab. 2).", + "bbox": [ + 496, + 724, + 890, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Descriptors Define Matching Probability", + "text_level": 1, + "bbox": [ + 498, + 797, + 846, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to [47, 50], we model the cycle matching probability using a double softmax (i.e. the probability of matching $i$ to $j$ , and back).", + "bbox": [ + 496, + 821, + 890, + 867 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP _ {i \\leftrightarrow j} = P _ {i \\rightarrow j} P _ {i \\leftarrow j}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 885, + 767, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22501", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $P_{i\\rightarrow j}$ is the directional probability of matching the $i$ th descriptor from image $I$ to the $j$ th descriptor in image $I'$ . $P_{i\\leftarrow j}$ is similar, but in the reverse direction. Both forward and backward probabilities are modeled as a softmax with temperature over descriptor cosine similarities. For fixed $i$ , $P_{i\\rightarrow j}$ is a softmax over the $i$ -th row of the descriptor similarity matrix, and $P_{i\\leftarrow j}$ takes softmax over column $j$ .", + "bbox": [ + 76, + 90, + 472, + 198 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Training", + "text_level": 1, + "bbox": [ + 76, + 205, + 181, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Self-Supervision. Pixel-accurate correspondences are required during training. Similar to SuperPoint [16], we obtain image-pair correspondences by applying a random transformation (homography) to an image. However, the homography is a linear mapping that gives correspondences at a subpixel level.", + "bbox": [ + 75, + 228, + 468, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain pixel correspondences, we apply the sampled homography to all the pixel positions of image $I$ ; positions being the center of pixels (e.g. the top-left pixel has position $(0.5, 0.5)$ ). This first step establishes dense directional correspondences from $I$ to $I'$ . We then run the same process from $I'$ to $I$ (using the inverse homography this time). Once both directional correspondences are obtained, the resulting positions are discretized; out-of-bound and non-bijective correspondences are discarded.", + "bbox": [ + 75, + 319, + 468, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image Augmentations. As done in [16, 42], we employ image augmentation to improve robustness; augmentations include random brightness, contrast, gaussian noise, speckle noise and motion blur. We refer to supplementary materials for details.", + "bbox": [ + 75, + 455, + 468, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Negative and Positive Selection. One defining property of a keypoint is distinctiveness – the point can be reliably distinguished from its peers. In our case this means the point can be reliably identified in a matching algorithm, similar to [49, 50]. With that in mind, we adopt a similar supervision technique as [49]. Keypoints that are correctly matched (using the currently trained descriptors) are labelled as positive; otherwise negative.", + "bbox": [ + 75, + 530, + 468, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Descriptor and Keypoint Losses. Similar to [47, 50], the descriptor loss is the negative log-likelihood loss applied to the matching probabilities for the positive round-trips (i.e. paths from point $i$ to its location $i'$ in the transformed image and back again).", + "bbox": [ + 75, + 651, + 468, + 728 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d e s c} = - \\left(\\log P _ {i \\rightarrow i ^ {\\prime}} + \\log P _ {i \\leftarrow i ^ {\\prime}}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 738, + 393, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This implicitly penalizes non-positive paths via softmax. One might notice $\\mathcal{L}_{desc}$ requires the computation of a large matrix (size $HW\\times HW$ ). To handle potential GPU out-of-memory, we provide a simple, yet efficient implementation which computes the similarity matrix in a block-wise fashion, and recomputes block dot products instead of storing them for backpropagation (in the same vein as [36, 25]).", + "bbox": [ + 75, + 763, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The keypoint loss $\\mathcal{L}_{key}$ is a simple binary cross-entropy loss applied to a logistic sigmoid, in contrast to [49]. It", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is trained to identify keypoints with successful round-trip matches (defined by mutual-nearest-neighbor) among all others (unsuccessful).", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 148, + 633, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we empirically evaluate SiLK together with representative baselines and state-of-the-art methods. On HPatches we evaluate two complementary keypoint quality metrics (Repeatability, Mean Matching Accuracy [33]) and planar stereo estimation capabilities (Homography Estimation). In addition, we benchmark on three real-world stereo tasks: outdoor camera pose estimation on Image Matching Challenge (IMC) 2022, and both indoor camera pose estimation and 3D point-cloud registration on ScanNet. In these experiments, we study the following:", + "bbox": [ + 496, + 174, + 890, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(1) Many methods employ complex strategies to learn and predict good keypoints and descriptors, including elements like multi-stage training, cell-based priors, complicated post-processing, context aggregation, and groundtruth 3D pose supervision (see Tab. 1), in various combinations. What machinery is necessary? SiLK contains no such machinery, and can be viewed as a reduction from these methods. However, SiLK either achieves new SOTA or compares very favorably. This questions the need for complex schemes for the evaluated tasks.", + "bbox": [ + 496, + 325, + 890, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(2) We observed rather strong performance from engineered features (e.g. SIFT in Tab. 2&Tab. 6) vs learned methods. This motivates us to revisit design choices in a learned keypoint detector: What makes a good keypoint detector? We ablate each component (data, backbone, etc.) and test generalization performance under various conditions (e.g. input size, test data, task, etc). SiLK proves very robust to these choices (Sec. 4.6). In particular, a very lightweight version of SiLK (two 3x3 convolution layers) is competitive to SOTA on homography estimation, camera pose estimation and point cloud registration (Tab. 7&Tab. 8).", + "bbox": [ + 496, + 476, + 890, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We hope these results can serve the community and help adapt keypoint models to their tasks and needs. For example, labeling tasks (e.g. self-training [37], object pose [38]) might focus on high accuracy (i.e. larger backbone and denser keypoint selection), while tasks requiring speed (e.g. SLAM [34]) might find our lightweight backbone attractive.", + "bbox": [ + 496, + 642, + 890, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 741, + 718, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our own training pipeline has been used for all experiments with SiLK, as well as our reproduced SuperPoint (Tab. 2) results. Training time is ~5 hours on 2 Tesla V100-SXM2 GPUs using our default setup.", + "bbox": [ + 496, + 763, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Default Setup. Unless specified otherwise, all results use the following setup. Trained on COCO [29] images (randomly sampled), with Adam [24] optimizer with learning rate $1e^{-4}$ and betas $(0.9, 0.999)$ ; trained for $100\\mathrm{k}$ iterations; batch size of 1 per GPU; dense descriptor map", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22502", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is $146 \\times 146$ for all architectures and input resolutions; cosine similarities scaled by temperature $20^{-1}$ ; VGGnp-4 backbone (VGG architecture with max-pooling removed, details in Sec. 4.6); sparse keypoints obtained with top-k ( $k = 10000$ ); detection head is 13x3 convolution (128-dims), and 11x1 convolution; descriptor head is 13x3 convolution (128-dims), and 11x1 convolution (128-dims out); no padding in convolutions; ReLU and batchnorm used as non-linearity and normalization (see supplementary).", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. HPatches Homography Estimation", + "text_level": 1, + "bbox": [ + 76, + 239, + 382, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [16, 42, 39, 47], we evaluate homography estimation on HPatches [3]. HPatches contains 57 scenes (of 6 images) with significant illumination changes and 59 scenes with large viewpoint variations. Images in each scene are related by groundtruth homographies. We follow LoFTR [47], (currently SOTA on HPatches), and scale the shorter image edge to 480 at inference time.", + "bbox": [ + 75, + 263, + 468, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Protocol. For every image pairs, the model detects a set of keypoints. These are desired to be distinctive and thus repeatably detected across views. We use Repeatability to evaluate detection performance as in [16]. To test invariance of keypoint descriptors, each model's preferred matching algorithm establishes correspondence across images to obtain a subset of keypoints. We distinguish this subset as post-matching and the entire set as pre-matching. The accuracy of each correspondence is evaluated by Mean Matching Accuracy as in [17, 39, 50]. Finally, we use OpenCV RANSAC algorithm to estimate homography from matched keypoints, and evaluate Homography Estimation Accuracy [16] and Homography Estimation AUC [42, 47].", + "bbox": [ + 75, + 369, + 468, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare SiLK against both sparse detector-based methods and dense detector-free methods. The sparse detector-based methods generally follow \"detect then match\": the model first detects a sparse set of distinctive points, then matches features. We include SIFT [31] as well as learned detectors SuperPoint [16] (both official release and our repro), R2D2 [39] and DISK [50]. On the other hand, dense detector-free perform \"detection by matching\": the model first extracts features, then applies a learned prematching CA module (e.g. GNN [42] or transformer [47]) to adapt the features to a specific pair of images, and then finds matches. While SiLK does not employ CA, we still include comparisons to the SOTA detector-free LoFTR [47].", + "bbox": [ + 75, + 566, + 468, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results. Despite its simplicity, SiLK outperforms all methods on repeatability, homography accuracy and homography estimation (Tab. 2). In particular, SiLK has a strong margin when the error threshold is small ( $\\epsilon = 1$ ). This validates our pixel-accurate keypoint localization. SiLK lags only on the MMA@3 metric. This is caused by the pixel-accurate contrastive loss (Fig. 4). Consequently, SiLK does not benefit from increasing error threshold in MMA. In addition, even vs LoFTR (which uses dense features and CA)", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in Tab. 3, SiLK shows strong performance on Homography Estimation AUC and competitive performance on Homography accuracy. This questions the necessity of CA for these particular tasks; this may be valuable in applications which are particularly sensitive to runtime performance.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bf27379361492e49caf5812bdde8afd09ec6a78ad275f247f89f82258db8243d.jpg", + "image_caption": [ + "Figure 4. The pixel-accurate contrastive loss results in very discriminative local features, thus reshaping the error distribution to make fewer local errors (MMA@1). This differs from interpolated, cell-based descriptors used in SuperPoint, which produces less accurate keypoints (MMA@3+)." + ], + "image_footnote": [], + "bbox": [ + 514, + 183, + 696, + 303 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/35603335e9a0419b19c6b8ed9978fbde35c266597b4785a95adc87b775fd255c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 183, + 880, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. IMC 2022 outdoor pose estimation", + "text_level": 1, + "bbox": [ + 500, + 404, + 802, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Image Matching Challenge (IMC 2022) [22] provides pairs of outdoor images from different viewpoints; participants are required to estimate the fundamental matrix. Camera pose accuracy is then computed for ten thresholds of rotation and translation error (ranging from $(1^{\\circ}, 20\\mathrm{cm})$ to $(10^{\\circ}, 5\\mathrm{m}))$ . Mean average accuracy (mAA) is reported by averaging across thresholds and scenes.", + "bbox": [ + 496, + 428, + 890, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Protocol. Sparse methods (DISK, SiLK) detect keypoints in individual images; mutual nearest neighbor is used to select matches from each pair. Methods with CA (SuperGlue, LoFTR) directly identify matches from each image pair. In either case, the fundamental matrix is estimated using MAGSAC [4]. The challenge allows different image sizes and tuning MAGSAC parameters [22]. We use 30k keypoints and MAGSAC threshold .25.", + "bbox": [ + 496, + 534, + 890, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We consider the best leaderboard results from three baselines. (i) DISK is the winner of IMC 2020 and is SOTA among sparse methods. 2022 DISK results are from the IMC team (submission). We also take the best version of (ii) SuperGlue (submission) and (iii) LoFTR (submission) provided by the community.", + "bbox": [ + 496, + 655, + 890, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results. We coarsely tune SiLK for this task for 30 trials (compared to LoFTR's 200 trials). SiLK again performs competitively (c.f. Tab. 4), outperforming DISK by a significant margin $(+0.19 / +0.18\\mathrm{mAA})$ . SiLK also performs favorably compared to SuperGlue, which uses context aggregation and optimal transport matching.", + "bbox": [ + 496, + 746, + 890, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. ScanNet: Indoor Pose & Point Clouds", + "text_level": 1, + "bbox": [ + 500, + 845, + 826, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ScanNet [12] is a large-scale dataset of 1513 indoor scenes of RGB-D images and ground-truth camera poses.", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22503", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8fab6dd60fba888a9e5262f78c92d53474677457647400519f218342dbd10564.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RepeatabilityHom. Est. Acc.Hom. Est. AUCMMA# of keypoints
ε = 1ε = 3ε = 1ε = 3ε = 1ε = 3ε = 1ε = 3pre-matchpost-match
SuperPoint (MagicLeap)0.340.610.430.80.20.510.410.72847499
SuperPoint (Ours)0.330.520.480.750.260.520.380.531143474
SIFT0.310.520.60.840.340.610.410.552189910
R2D20.360.720.450.790.20.50.340.7560881967
DISK0.380.690.450.80.220.520.520.8433491794
SiLK (top-10k)0.620.810.620.870.40.660.590.71100004283
SiLK (top-5k)0.560.760.60.850.390.640.570.6950002074
SiLK (top-1k)0.430.610.530.810.320.580.520.631000389
", + "bbox": [ + 127, + 88, + 841, + 210 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3b69ceeea98bca40440c84051c76d3bedfb43ee0cd689fd3d4b75764c6c63c7a.jpg", + "table_caption": [ + "Table 2. SiLK achieves new SOTA on HPatches compared to other methods with sparse keypoints and features. Despite its simplicity, SiLK achieves higher performance on all metrics except MMA@3. We include the # of keypoints to ensure a fair comparison." + ], + "table_footnote": [], + "table_body": "
Hom. Est. Acc.Hom. Est. AUCMMA
ε = 1ε = 3ε = 1ε = 3ε = 1ε = 3
LoFTR (MegaDepth)0.650.870.370.650.640.91
LoFTR (ScanNet)0.240.570.070.330.360.76
SiLK (top-10k)0.620.870.40.660.590.71
SiLK (top-5k)0.600.850.390.640.570.69
", + "bbox": [ + 78, + 258, + 478, + 334 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ff25d48183a95c4c61011071acd1414ffda31e93c433f1d770865aad64fdcf2b.jpg", + "table_caption": [ + "Table 3. SiLK achieves competitive performance to SOTA LoFTR on HPatches, despite not using context aggregation. The top-5k SiLK has similar number of matches compared to outdoor LoFTR. We remark that LoFTR has large generalization gap when training on different types of dataset (indoor vs. outdoor)." + ], + "table_footnote": [], + "table_body": "
private mAApublic mAA
Sparse Features
DISK0.4910.502
SuperGlue0.6760.678
SiLK0.6850.684
Dense Features
LoFTR (MegaDepth)0.7350.721
", + "bbox": [ + 133, + 415, + 416, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using the official train/test split we evaluate both relative camera pose estimation and point-cloud registration. Relative camera pose estimation has been used in multiple previous works [52, 53, 6, 42, 47]. The task is to estimate the essential matrix with RANSAC from point matches in a pair of images. We report pose error AUC at thresholds $(5^{\\circ},10^{\\circ},20^{\\circ})$ using 20k keypoints and inlier threshold .5.", + "bbox": [ + 75, + 568, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This protocol measures translation error in degrees and is known to suffer from scale ambiguity [52]. Angular translation error may be unstable, in particular if the underlying translation (in meters) is small. In response, recent works [19, 18] have introduced a 3D point-cloud registration task, using the ground-truth depth provided in ScanNet. A pair of images 20 frames apart is first sampled. Given this pair, a model predicts point matches. After matching, relative camera pose is estimated. Different from the previous protocol, ground-truth depth and camera intrinsics are now used to align matches in 3D. In addition to relative pose errors (reported separately for translation (cm) and rotation (degree)), the Chamfer distance (in cm) is measured between the registered point cloud and the groundtruth point cloud. We refer to the original papers [19, 18] for details on", + "bbox": [ + 75, + 674, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3042c0c91468d03009ee439be58d1b6fbed2d4e993e3da199efee23e794ec834.jpg", + "table_caption": [ + "Table 4. SiLK achieves new SOTA for sparse methods on IMC2022 and performs competitively to dense methods with context aggregation." + ], + "table_footnote": [], + "table_body": "
Pose Estimation AUC@5°@10°@20°
Sparse Features
D2-Net [17] + MNN5.314.528.0
SuperPoint [16] + MNN9.421.536.4
SuperPoint + PointCN [52]11.425.541.4
SuperPoint + OANet [53]11.826.943.9
SuperPoint + SuperGlue [42]16.233.851.8
SiLK + MNN18.034.450.4
Dense Features
DRC-Net [27]7.717.930.5
LoFTR (MegaDepth)16.933.650.6
LoFTR (ScanNet)21.540.857.6
", + "bbox": [ + 549, + 258, + 844, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5. SiLK advances SOTA on relative pose estimation among sparse methods on ScanNet and performs competitively against dense method LoFTR.", + "bbox": [ + 498, + 405, + 890, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the evaluation setup and metrics.", + "bbox": [ + 500, + 470, + 718, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4.1 Relative pose estimation", + "text_level": 1, + "bbox": [ + 500, + 503, + 723, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare SiLK with both sparse detector-based methods and dense detector-free methods. For the sparse methods, we consider local feature descriptors (R2D2, SuperPoint) with mutual nearest neighbor (MNN) for matching. SiLK falls into this category. In addition, we consider multiple learned context aggregation methods for matching that operates on SuperPoint, including PointCN [52], OANet [53] and SuperGlue. For detector-free dense methods, we consider DRC-Net [27] and LoFTR. We include two versions of LoFTR: one trained on MegaDepth with optimal transport post-processing and one trained on ScanNet.", + "bbox": [ + 496, + 526, + 890, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. As summarized in Tab. 5, SiLK significantly outperforms D2-Net (+12.7) and SuperPoint (+8.6) when using mutual nearest neighbor matching. In addition, SiLK outperforms the previous SOTA sparse method SuperGlue, despite its simpler design without context aggregation. SiLK performs similarly to LoFTR trained with MegaDepth. SiLK is only outperformed by LoFTR trained on ScanNet, the same as evaluation data.", + "bbox": [ + 496, + 707, + 890, + 828 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4.2 Pairwise 3D point-cloud registration", + "text_level": 1, + "bbox": [ + 500, + 845, + 807, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We consider three main types of baselines. (i) Sparse Features + RANSAC We extract sparse key-", + "bbox": [ + 500, + 869, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22504", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "points and their features from off-the-shelf models, and use RANSAC to estimate alignment. This includes SIFT and SuperPoint, and 3D geometry model FCGF [10]. (ii) Dense Feature Matching We follow the [18] and select high-quality corresponding pairs using the ratio test, and then solve a weighted Procustes problem [9, 23] to produce alignment. We add a dense version of SuperPoint by discarding the keypoint prediction. We include the current state-of-the-art URR [18] that learns invariant point descriptors through cross-view synthesis. By ignoring the keypoint scoring prediction, SiLK also belongs to this category. The goal is to evaluate the quality of the dense point features. (iii) Pose/Geometry Supervised We consider methods that use groundtruth poses to supervise, which are not required in (i) and (ii). These include LoFTR, DGR [9] and 3D MV Registration [19]. For LoFTR, we include only the MegaDepth model, since it performed better in this task.", + "bbox": [ + 76, + 90, + 472, + 347 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. As shown in Tab. 6, SiLK achieves new SOTA across all metrics (except rotation accuracy at $45^{\\circ}$ ). In particular, SiLK achieves very high accuracy at small thresholds ( $5^{\\circ}$ angular, 5cm translation and 1cm for chamfer), validating SiLK's pixel-level precision. Comparing with DGR, 3D MV Reg and LoFTR that use groundtruth camera poses during training, SiLK significantly outperforms, indicating that groundtruth 3D supervision is not necessary to train good keypoint features. We did not include the chamfer results for LoFTR as the provided positions do not match the required resolution for correct chamfer evaluation. SiLK also achieves superior performance vs previous SOTA URR. We note that URR requires two different frames sampled from the same scene during training, and is supervised by a differentiable cross-view rendering process. In contrast, SiLK only requires a single 2D image and is trained with a simple point matching loss. Finally, we observe that SuperPoint performs competitively when evaluated in this dense fashion; this is an important difference from the results reported in URR using sparse features.", + "bbox": [ + 76, + 348, + 472, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Discussion. What makes SiLK perform?", + "text_level": 1, + "bbox": [ + 76, + 664, + 428, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multiple factors contribute to SiLK's performance. In particular, the cycle-consistent loss and the dense predictions. First, the cycle-consistent loss (Fig. 2, supplementary) enforces the natural properties of keypoints: distinctiveness & robustness to viewpoint / photometric changes, while existing methods rely on proxy objectives or additional complexity (i.e. \"corners\" in SuperPoint[16] / SuperGlue[42], \"peakiness\" in R2D2[39], \"keypoint weighting scheme\" in D2-Net[17], \"RL cycle-consistent reward\" in DISK[50]). Removing unnecessary complexity helps the model focus its learning on the essentials, and is supported by consistent, strong results across benchmarks and metrics (Tab. 2, Tab. 3, Tab. 4, Tab. 5, Tab. 6). Second, SiLK produces dense, pixel-accurate keypoints, while all", + "bbox": [ + 76, + 688, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "existing methods have structural constraints (downsampling layers, cell-based schemes, and NMS, c.f. Tab. 1). This leads to more robust contrastive learning (i.e. more negative pairs), allows producing more keypoints matches and boosts results in the accurate regime ( $\\epsilon = 1$ in Tab. 2).", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. What makes good keypoint detectors?", + "text_level": 1, + "bbox": [ + 498, + 178, + 828, + 195 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Leveraging SiLK's flexibility (Sec. 3), we comprehensively ablate a large pool of design choices such as model architecture and image resolution. Surprisingly, we found that reducing architecture size, compute cost, and training input size only mildly impact model performance on homography estimation, camera pose estimation and point cloud registration. This benefits many important applications, such as on-device inference.", + "bbox": [ + 496, + 202, + 893, + 323 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here we discuss the key findings. We use the metrics Repeatability@1 (R), Homography Estimation Accuracy@1 (HA), Mean Matching Accuracy@1 (MMA) for HPatches, and Rotation Accuracy@5° (RA), Translation Accuracy@5cm (TA) and Chamfer@1 (C) for ScanNet, all at lowest error thresholds. Additional results and analysis are included in supplementary.", + "bbox": [ + 496, + 324, + 893, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Agnostic to backbone. Existing methods use various backbones (Tab. 1); the effects on keypoint models are not well understood. We consider FPN from LoFTR and UNet from DISK; both are modern compared to SiLK's VGGnp [45] backbone. We find no empirical performance gain despite far greater parameter counts (Tab. 7). This questions the need for high-capacity models for these keypoint problems.", + "bbox": [ + 496, + 431, + 893, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next we reduce the complexity of the original Super-Point backbone VGG-4. Max-pooling and up-sampling layers are removed. Our VGGnp-4 contains four convolution blocks, each with two convolution layers followed by ReLU. We discard convolution blocks from VGGnp-4 to obtain VGGnp-3, VGGnp-2 and VGGnp-1. On top of VGGnp-1, we reduce channel and descriptor sizes to 64 and 32 respectively and obtain an ultra-lightweight model, VGGnp- $\\mu$ . Results on all metrics (except MMA) only drop mildly as we shrink the model. In particular, SiLK (VGGnp- $\\mu$ in Tab. 7) achieves very competitive performance (Tab. 2&Tab. 6). On the other hand, matching scores (MMA) drop significantly. We suggest two possible reasons: first, pointwise matching may benefit from the larger receptive field of deeper models. Second, homography estimation aggregates numerous pointwise measurements; homographies will improve if the noises cancel out, or if the outliers are removed by RANSAC.", + "bbox": [ + 496, + 537, + 893, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fast training on tiny images. By default, SiLK uses 146x146 descriptor map resolution during training. Higher resolution provides more points, which benefits the contrastive loss with more negatives, but also increases training time and GPU memory usage. Surprisingly, performance changes very little when varying resolution during", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22505", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6c3b4a2d4c56023d848163f27280dcc7d774309191f62333e7dd9bdff3c0773d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RotationTranslationChamfer
Accuracy ↑Error ↓Accuracy ↑Error ↓Accuracy ↑Error ↓
5°10°45°MeanMed.51025MeanMed.1510MeanMed.
Sparse Features + RANSAC
SIFT [31]55.275.789.218.64.317.744.579.826.511.238.170.678.342.61.7
SuperPoint [16]65.586.996.68.93.621.251.788.016.19.745.781.188.219.21.2
FCGF [10]70.287.796.29.53.327.558.382.923.68.352.078.083.724.40.9
Pose/Geometry Supervised
DGR [9]81.189.394.89.41.854.576.288.718.44.570.585.589.013.70.4
3D MV Reg [19]87.793.297.06.01.269.083.191.811.72.978.989.291.810.20.2
LoFTR(MegaDepth) [47]91.796.899.42.81.265.985.297.16.03.3-----
Dense feature matching
SuperPoint [16]93.098.499.82.51.656.884.798.26.54.377.396.198.44.50.3
URR [18]92.795.898.53.40.877.289.696.17.32.386.094.696.15.90.1
SiLK (VGGnp-4)98.199.099.61.70.882.994.899.04.12.192.898.399.14.30.1
", + "bbox": [ + 135, + 88, + 834, + 271 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2082f5310f50a0a57224c25026c0e04d004b5027ce55a58d0d8400ac675ff378.jpg", + "table_caption": [ + "Table 6. SiLK achieves state-of-the-art on camera pose estimation and point cloud registration on ScanNet." + ], + "table_footnote": [], + "table_body": "
HPatchesScanNetModel
RHAcMMARATACParamFPSGFLOP
VGGnp-40.620.620.5998.182.992.8942k12.2370
VGGnp-30.630.610.5898.084.293.5868k12.5330
VGGnp-20.630.570.5597.383.592.4757k14.3268
VGGnp-10.630.570.4494.682.190.0461k18.990
VGGnp-μ0.640.560.4093.581.189.076k36.523
FPN [47]0.580.550.52---6.6M17.5298
UNet [50]0.600.410.58---1.3M25.9198
", + "bbox": [ + 80, + 306, + 467, + 416 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3e16c288ef9ea6144d4759eec8a61354aa818bb57aec1e65113ecedb8522ec76.jpg", + "table_caption": [ + "Table 7. SiLK is backbone agnostic. Backbones from existing methods are trained and evaluated. Low-capacity model performs well on all metrics except MMA. FPS and GFLOPs measured on $480 \\times 640$ images with NVIDIA Quadro GP100 GPU." + ], + "table_footnote": [], + "table_body": "
SizeTimeHPatchesScanNet
RHAcMMARATAC
8221.7h0.600.580.5698.183.592.5
11422.7h0.620.620.5998.182.992.8
14625h0.630.590.5998.283.392.9
17829.5h0.630.620.5998.183.592.9
210218h0.630.610.6098.183.492.8
", + "bbox": [ + 124, + 484, + 423, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "training (Tab. 8), especially on ScanNet. Tiny feature maps (82x82) remain competitive on both HPatches and ScanNet, and trains in 1.7 hours on two GPUs. This enables applications like test-time finetuning, on-device finetuning, and rapid experimental iteration.", + "bbox": [ + 75, + 626, + 467, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness to training data. Existing methods use various training sets (Tab. 1); empirically we observe cases of poor generalization across datasets. For example, LoFTR[47], trained on indoor ScanNet data, drops significantly vs LoFTR trained on outdoor MegaDepth data (Tab. 3) and vice-versa (Tab. 5). This overfitting may be exacerbated by the high-capacity machinery used by these methods, e.g. LoFTR's Transformer contextualizer. We measure SiLK's robustness on training data choices, by using different images from COCO[29], ImageNet[15], MegaDepth[28] and ScanNet[12]. We also combine them to formulate a diversified set of training data.", + "bbox": [ + 75, + 703, + 467, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SiLK is quite robust to change in training set, with the", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0f8c7063ae775f8db9f2920530103e0850b4aa264b726c514a79f0ccd3db8083.jpg", + "table_caption": [ + "Table 8. Decreasing training image size has minimal impact. SiLK can be trained under 3h, with little performance drop." + ], + "table_footnote": [], + "table_body": "
HPatchesScanNet
RHAcMMARATAC
COCO0.620.620.5998.182.992.8
ImageNet0.630.60.5998.183.593.0
MegaDep.0.620.610.5797.983.592.9
ScanNet0.600.550.4897.682.892.7
C+I+M+S0.610.590.5497.783.092.6
C+I+M0.640.60.5998.082.992.8
", + "bbox": [ + 550, + 306, + 841, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. SiLK is robust to different training sets. A noticeable drop is observed only when training on ScanNet.", + "bbox": [ + 500, + 404, + 890, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "exception of ScanNet (Tab. 9). We hypothesize this is due to the significant amount of uniform surfaces (e.g. walls, doors) present in ScanNet. These featureless areas contain few keypoints to learn from. SiLK's drop agrees directionally with LoFTR's drop observed in Tab. 3, but the magnitude is smaller. This may be because SiLK's VGGnp backbone has lower capacity than LoFTR (FPN+Transformer), and hence is less susceptible to overfitting. Finally, we remark that SiLK trained with COCO is used for comparisons in Sec. 4.2, Sec. 4.3 and Sec. 4.4 for HPatches, IMC and ScanNet, whereas LoFTR requires different training data (MegaDepth or ScanNet) to achieve strong performance.", + "bbox": [ + 496, + 460, + 890, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 661, + 617, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents SiLK, a simple and flexible framework for keypoint detection and descriptors. SiLK is designed from the principles of distinctiveness and invariance, and achieves or advances SOTA on key low-level tasks for 3D visual perception. SiLK's simplicity questions the need for complex machinery for good keypoint detection in low-level applications. In addition, extensive ablations reveal SiLK's robustness to backbone, training data and training input size. These findings lead to a tiny version of SiLK that is lightweight, accurate, and trains quickly. We view this \"tiny and learned\" regime as very promising for applications where runtime and/or power consumption is critical. We hope SiLK can draw attention to the field and facilitate stronger solutions.", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22506", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ebtsam Adel, Mohammed Elmogy, and Hazem Elbakry. Image stitching based on feature extraction techniques: a survey. International Journal of Computer Applications, 99(6):1-8, 2014.", + "[2] Alexandre Alahi, Raphael Ortiz, and Pierre Vandergheynst. Freak: Fast retina keypoint. In 2012 IEEE conference on computer vision and pattern recognition, pages 510-517. IEEE, 2012.", + "[3] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5173-5182, 2017.", + "[4] Daniel Barath, Jiri Matas, and Jana Noskova. MAGSAC: marginalizing sample consensus. In _Conference on Computer Vision and Pattern Recognition_, 2019.", + "[5] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer vision and image understanding, 110(3):346-359, 2008.", + "[6] Eric Brachmann and Carsten Rother. Neural-guided ransac: Learning where to sample model hypotheses. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4321-4330, 2019.", + "[7] Matthew A. Brown and David G. Lowe. Automatic panoramic image stitching using invariant features. International Journal of Computer Vision, 74:59-73, 2006.", + "[8] Michael Calonder, Vincent Lepetit, Christoph Strecha, and Pascal Fua. Brief: Binary robust independent elementary features. In European conference on computer vision, pages 778-792. Springer, 2010.", + "[9] Christopher Choy, Wei Dong, and Vladlen Koltun. Deep global registration. In CVPR, 2020.", + "[10] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 8957-8965, 2019.", + "[11] Peter Hviid Christiansen, Mikkel Fly Kragh, Yury Brodskiy, and Henrik Karstoft. Unsupperpoint: End-to-end unsupervised interest point detector and descriptor. arXiv preprint arXiv:1907.04011, 2019.", + "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017.", + "[13] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), volume 1, pages 886-893. IEEE, 2005.", + "[14] Andrew J. Davison, Ian D. Reid, Nicholas Molton, and Olivier Stasse. Monoslam: Real-time single camera slam. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29:1052-1067, 2007.", + "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009.", + "[16] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018.", + "[17] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint detection and description of local features. arXiv preprint arXiv:1905.03561, 2019.", + "[18] Mohamed El Banani, Luya Gao, and Justin Johnson. Unsupervisedr&r: Unsupervised point cloud registration via differentiable rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7129-7139, 2021.", + "[19] Z. Gojcic, C. Zhou, J. D. Wegner, L. J. Guibas, and T. Birdal. Learning multiview 3d point cloud registration. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1756-1766, Los Alamitos, CA, USA, jun 2020. IEEE Computer Society.", + "[20] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In Alvey vision conference, volume 15, pages 10-5244. Manchester, UK, 1988.", + "[21] Wilfried Hartmann, Michal Havlena, and Konrad Schindler. Predicting matchability. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9-16, 2014.", + "[22] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image Matching across Wide Baselines: From Paper to Practice. International Journal of Computer Vision, 2020.", + "[23] W. Kabsch. A solution for the best rotation to relate two sets of vectors. Acta Crystallographica Section A, 32(5):922-923, Sep 1976.", + "[24] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2015.", + "[25] Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, and Daniel Haziza. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers, 2022.", + "[26] Stefan Leutenegger, Margarita Chli, and Roland Y Siegwart. Brisk: Binary robust invariant scalable keypoints. In 2011 International conference on computer vision, pages 2548-2555. Ieee, 2011.", + "[27] Xinghui Li, Kai Han, Shuda Li, and Victor Prisacariu. Dual-resolution correspondence networks. In Conference on Neural Information Processing Systems (NeurIPS), 2020.", + "[28] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018.", + "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22507", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "European conference on computer vision, pages 740-755. Springer, 2014.", + "[30] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. Ssd: Single shot multibox detector. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision - ECCV 2016, pages 21-37, Cham, 2016. Springer International Publishing.", + "[31] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60(2):91-110, 2004.", + "[32] Jiayi Ma, Xingyu Jiang, Aoxiang Fan, Junjun Jiang, and Junchi Yan. Image matching from handcrafted to deep features: A survey. International Journal of Computer Vision, 129(1):23-79, 2021.", + "[33] K. Mikolajczyk and C. Schmid. A performance evaluation of local descriptors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(10):1615-1630, 2005.", + "[34] Raul Mur-Artal, J. M. M. Montiel, and Juan D. Tardós. Orb-slam: A versatile and accurate monocular slam system. IEEE Transactions on Robotics, 31:1147-1163, 2015.", + "[35] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018.", + "[36] Markus N Rabe and Charles Staats. Self-attention does not need $o(n^{2})$ memory. arXiv preprint arXiv:2112.05682, 2021.", + "[37] Umer Rafi, Andreas Doering, Bastian Leibe, and Juergen Gall. Self-supervised keypoint correspondences for multi-person pose estimation and tracking in videos. In Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XX, page 36–52, Berlin, Heidelberg, 2020. Springer-Verlag.", + "[38] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10881-10891, 2021.", + "[39] Jerome Revaud, Philippe Weinzaepfel, César De Souza, Noe Pion, Gabriela Csurka, Yohann Cabon, and Martin Humenberger. R2d2: repeatable and reliable detector and descriptor. arXiv preprint arXiv:1906.06195, 2019.", + "[40] Edward Rosten and Tom Drummond. Machine learning for high-speed corner detection. In European conference on computer vision, pages 430-443. Springer, 2006.", + "[41] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision, pages 2564-2571. Ieee, 2011.", + "[42] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020.", + "[43] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural" + ], + "bbox": [ + 78, + 90, + 468, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "network model. IEEE transactions on neural networks, 20(1):61-80, 2008.", + "[44] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016.", + "[45] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2015.", + "[46] Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 21(2):343-348, 1967.", + "[47] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021.", + "[48] Jiexiong Tang, Hanme Kim, Vitor Guizilini, Sudeep Pillai, and Rares Ambrus. Neural outlier rejection for self-supervised keypoint learning. arXiv preprint arXiv:1912.10615, 2019.", + "[49] Prune Truong, Stefanos Apostolopoulos, Agata Mosinska, Samuel Stucky, Carlos Ciller, and Sandro De Zanet. Glampoints: Greedily learned accurate match points. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10732-10741, 2019.", + "[50] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020.", + "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.", + "[52] Kwang Moo Yi, Eduard Trulls, Yuki Ono, Vincent Lepetit, Mathieu Salzmann, and Pascal V. Fua. Learning to find good correspondences. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2666-2674, 2017.", + "[53] Jiahui Zhang, Dawei Sun, Zixin Luo, Anbang Yao, Lei Zhou, Tianwei Shen, Yurong Chen, Long Quan, and Honggen Liao. Learning two-view correspondences and geometry using order-aware network. International Conference on Computer Vision (ICCV), 2019." + ], + "bbox": [ + 501, + 92, + 890, + 700 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22508", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/SiLK_ Simple Learned Keypoints/545a50e4-a6f6-42d1-849e-c613959c21fd_model.json b/2023/SiLK_ Simple Learned Keypoints/545a50e4-a6f6-42d1-849e-c613959c21fd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4a4b3bc4e02f34c572042a6c5a00fac83ca977f6 --- /dev/null +++ b/2023/SiLK_ Simple Learned Keypoints/545a50e4-a6f6-42d1-849e-c613959c21fd_model.json @@ -0,0 +1,2244 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.315, + 0.131, + 0.658, + 0.154 + ], + "angle": 0, + "content": "SiLK: Simple Learned Keypoints" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.182, + 0.214, + 0.198 + ], + "angle": 0, + "content": "Pierre Gleize" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.203, + 0.228, + 0.216 + ], + "angle": 0, + "content": "gleize@meta.com" + }, + { + "type": "text", + "bbox": [ + 0.442, + 0.182, + 0.558, + 0.2 + ], + "angle": 0, + "content": "Weiyao Wang" + }, + { + "type": "text", + "bbox": [ + 0.413, + 0.203, + 0.585, + 0.217 + ], + "angle": 0, + "content": "weiyaowang@meta.com" + }, + { + "type": "text", + "bbox": [ + 0.775, + 0.182, + 0.874, + 0.198 + ], + "angle": 0, + "content": "Matt Feiszli" + }, + { + "type": "text", + "bbox": [ + 0.77, + 0.203, + 0.88, + 0.216 + ], + "angle": 0, + "content": "mdf@meta.com" + }, + { + "type": "title", + "bbox": [ + 0.462, + 0.235, + 0.534, + 0.25 + ], + "angle": 0, + "content": "Meta AI" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.253, + 0.735, + 0.27 + ], + "angle": 0, + "content": "https://github.com/facebookresearch/silk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.474, + 0.699 + ], + "angle": 0, + "content": "Keypoint detection & descriptors are foundational technologies for computer vision tasks like image matching, 3D reconstruction and visual odometry. Hand-engineered methods like Harris corners, SIFT, and HOG descriptors have been used for decades; more recently, there has been a trend to introduce learning in an attempt to improve keypoint detectors. On inspection however, the results are difficult to interpret; recent learning-based methods employ a vast diversity of experimental setups and design choices: empirical results are often reported using different backbones, protocols, datasets, types of supervisions or tasks. Since these differences are often coupled together, it raises a natural question on what makes a good learned keypoint detector. In this work, we revisit the design of existing keypoint detectors by deconstructing their methodologies and identifying the key components. We re-design each component from first-principle and propose Simple Learned Keypoints (SiLK) that is fully-differentiable, lightweight, and flexible. Despite its simplicity, SiLK advances new state-of-the-art on Detection Repeatability and Homography Estimation tasks on HPatches and 3D Point-Cloud Registration task on ScanNet, and achieves competitive performance to state-of-the-art on camera pose estimation in 2022 Image Matching Challenge and ScanNet." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.725, + 0.21, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Keypoint detection and matching is a foundational computer vision technique to obtain a sparse yet informative representation of an image. Image stitching [7, 1], SLAM [14, 34], SfM [44], camera calibrations, tracking [37], and object detection [30] are important tasks built on keypoint correspondences [32]. A good keypoint model should be able to select a subset of points useful and informative to a specific task. One typically also wants robustness of the descriptor to some set of transformations (e.g. scale, viewpoint, or lighting variation)." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.304, + 0.698, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.304, + 0.885, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.42, + 0.892, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.66, + 0.894, + 0.744 + ], + "angle": 0, + "content": "Figure 1. The top image is an example of keypoint matching under viewpoint change; correct matches are green, incorrect ones are red. The bottom image shows keypoints which are cycle-consistent by SiLK. As can be observed, SiLK has learned to find distinctive geometric features (corners, curves, intersections,..); from single non-annotated images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Existing keypoint methods come in multiple forms and flavors (Tab. 1). However, those differences are often coupled together and not controlled for, which make it challenging to identify the source of gain. Our quest to finding the key-components of good keypoint detectors led to an in-depth review of alternative approaches, as well as the design, from first principle, of SiLK (Simple Learned Keypoints): A simple self-supervised approach to learn" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22499" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.151 + ], + "angle": 0, + "content": "distinctive and robust keypoints from arbitrary image data in a traditional \"detect-and-describe\" framework. Despite its simplicity, SiLK is competitive or surpasses SOTA in most settings." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.153, + 0.468, + 0.242 + ], + "angle": 0, + "content": "Additionally, leveraging SiLK's simple one-stage training protocol and modular architecture, we are able to ablate various dimensions of detector performance for different tasks. In particular, with an eye toward real-time performance, we identify tasks where extremely lightweight backbone architectures are sufficient." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.257, + 0.217, + 0.271 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.282, + 0.468, + 0.402 + ], + "angle": 0, + "content": "Early work focused on carefully engineered methods to identify distinctive keypoints with descriptors which are robust to changes such as viewpoint and illumination. Hand-crafted techniques like Harriscorners[20], SIFT[31], ORB[41] and others [40, 8, 26, 2, 5, 13] have used explicit geometric notions like corners, gradients, and scale-space extrema to achieve results which remain both efficient and competitive to date [44, 34]." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.403, + 0.468, + 0.522 + ], + "angle": 0, + "content": "More recent work like SuperPoint[16] chose to learn to find corners; they generate a large set of synthetic shapes with annotated corners and train a model with this ground truth. While providing compelling evidence for learned methods, their training procedure is quite complex: it contains multiple training phases, a synthetic dataset, and employs a homography adaptation trick that can be difficult to tune (see our reproduced results in Tab. 2)." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.524, + 0.468, + 0.568 + ], + "angle": 0, + "content": "In the same spirit as more recent work [17, 39, 50, 49, 21, 11, 48], SiLK aims to learn keypoints in simple end-to-end fashion, without explicitly defining them as corners." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.569, + 0.468, + 0.643 + ], + "angle": 0, + "content": "Several attempts have been made to learn keypoints implicitly; either by the careful design of the loss [17, 39, 11, 48]; or by directly predicting the matching success of descriptors [50, 21, 49]. SiLK falls in the second category, but with slight twist (cf. Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.645, + 0.468, + 0.764 + ], + "angle": 0, + "content": "To learn descriptors, contrastive losses are commonly used. Similar to [50, 47], SiLK adopts a probabilistic approach by modeling the matching probabilities in a double-softmax, cycle-consistency setting and optimizes the log likelihood. The probabilistic formulation, similar to InfoNCE [35], gives us a clean way to reason about matching, and the abundant supply of hard negative examples (from pixels in the same image) makes it an attractive choice." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.766, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Context aggregation (CA) is a recent addition to the toolkit. Initiated by SuperGlue[42], CA aims to refine or transform descriptors from a pair of images before matching them. Implemented as a GNN[43] in [42], or as a Transformer[51] in [47], CA's predictions are conditional on all descriptors from the pair of images. In other words, for an image, the descriptors will be different when matching against different images. As a result, CA needs to run on every pair of images prior to matching, as opposed to" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.092, + 0.891, + 0.181 + ], + "angle": 0, + "content": "running on single images. The run-time implications render CA prohibitively expensive in some applications (quadratic versus linear complexity). SiLK does not use CA, but outperforms [42] and performs competitively with [47]. Incorporating CA is optional for future works if performance is paramount and computation cost is less of a concern." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.183, + 0.891, + 0.272 + ], + "angle": 0, + "content": "As a postprocessing after CA, SuperGlue[42] introduced the concept of differentiable optimal transport (OT) to improve matching, using the Sinkhorn algorithm[46]. LoFTR [47] leveraged OT as well, but found little difference between OT and the simpler approach of mutual nearest neighbor (MNN) in some benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.286, + 0.633, + 0.302 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.311, + 0.891, + 0.551 + ], + "angle": 0, + "content": "SiLK's contribution is simplicity and flexibility. Our solution is built on the traditional approach of identifying distinctive pixels via robust local descriptors. We use modern but established techniques to learn to localize and describe keypoints given an arbitrary source of unlabeled images. Unlike classical methods, our descriptors and invariances are learned, and unlike some modern methods, there is no particular complexity in the matching process (SiLK employs only cosine distances and mutual nearest neighbor); this leaves few structural hyperparameters to tune. The simple backbone+heads design is backbone-agnostic, allowing experimentation. The annotation-free SSL approach means SiLK can be trained on any image or video dataset. Finally, a simple, one-stage training pipeline allows us to easily train and ablate different architectures, datasets, and hyperparameters for different tasks." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.553, + 0.891, + 0.627 + ], + "angle": 0, + "content": "SiLK is trained to identify keypoints from single grayscale images. It provides both keypoint detections (location) and keypoint descriptors (for matching). Cycle consistency is employed for descriptor learning and a binary classifier identifies distinctive keypoints at pixel-level." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.629, + 0.891, + 0.732 + ], + "angle": 0, + "content": "To learn descriptors, we take a source image and a transformed copy, extract descriptors for each point, and use descriptor similarity to define transition probabilities from a source location to each transformed location (and vice-versa). We optimize the descriptors to maximize cycle-consistency; i.e. we maximize the probability of a round-trip from the source to its transformed location and back." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.734, + 0.891, + 0.824 + ], + "angle": 0, + "content": "To locate good keypoints, we train a binary classifier to identify points which will satisfy a matching criterion. A point and its transformed copy are positives when they are mutual nearest neighbors in the sense of transition probabilities, and they are negatives otherwise. We train both the cycle-consistency and classification losses jointly." + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.825, + 0.795, + 0.839 + ], + "angle": 0, + "content": "We provide simple pseudo-code in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.848, + 0.633, + 0.862 + ], + "angle": 0, + "content": "3.1. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.871, + 0.891, + 0.9 + ], + "angle": 0, + "content": "The SiLK architecture (Fig. 3) follows the \"detect-and-describe\" architecture originally proposed by SuperPoint" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22500" + } + ], + [ + { + "type": "code", + "bbox": [ + 0.065, + 0.292, + 0.47, + 0.629 + ], + "angle": 0, + "content": "def loss(image_0, model):\n # get warped image and pixel correspondences\n image_1, corr_0, corr_1 = rand_homo(image_0)\n # apply image augmentations\n image_0 = augment(image_0)\n image_1 = augment(image_1)\n # extract dense descriptors and keypoints\n desc_0, kpts_0 = model(image_0)\n desc_1, kpts_1 = model(image_1)\n # compute similarity matrix\n sim_mat = cosine desc_0, desc_1)\n # compute the descriptor loss\n # using ground truth correspondences\n loss_desc = nll(sim_mat, corr_0, corr_1)\n # measure matching success\n y = is_MATCH_success(sim_mat, corr_0, corr_1)\n # compute keypoint loss\n # using current matching success\n loss_kpts = bce(kpts_0, y, corr_0)\n loss_kpts += bce(kpts_1, y, corr_1)\n return loss_desc + loss_kpts" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.636, + 0.455, + 0.651 + ], + "angle": 0, + "content": "Figure 2. Pseudo-code: learning keypoints from a single image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.679, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[16]. A dense feature map is first extracted by feeding an image to an encoder backbone. The shared feature map is then fed to two heads: The keypoint head extracts the logits (used to calculate the dense keypoint probabilities), while the descriptor head extracts a dense descriptor map (used to calculate keypoint similarities). The model is backbone-agnostic and can easily be swapped." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.8, + 0.465, + 0.817 + ], + "angle": 0, + "content": "3.2. High Matching Probability Defines Keypoints" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "As mentioned above, the keypoint probability estimate predicts the probability that a pixel will be correctly matched (i.e. survive a round-trip). Points with the highest likelihood of matching correctly are exactly those which we select as keypoints." + }, + { + "type": "table", + "bbox": [ + 0.061, + 0.088, + 0.891, + 0.235 + ], + "angle": 0, + "content": "
Keypoint DetectionDescriptors & MatchingModel & Training
LearnedSparseCell-basedNMSSupervisionCAMatcherSupervisionInputBackboneDataE2E
SIFTNo✓No✓-NoMNN-----
SuperPoint✓✓✓✓Homo. Adapt.NoMNNRand. Homo.GreyVGGCOCONo
SuperGlue-✓---GraphNNOTSfMGreyVGGOxford & ParisNo
GLAMPoints✓✓No✓Matching Succ.NoMNN-GreyUNetSlitLamp✓
D2-Net✓✓✓✓Triplet RankingNoMNNSfMRGBVGGMegaDepth✓
R2D2✓✓✓✓Matching APNoMNNSfMRGBL2-NetAachen✓
DISK✓✓✓✓Matching Succ.NoMNNSfMRGBUNetMegaDepth✓
URR-No----MNN3D RenderingRGBResNetScanNet✓
LoFTR-No---TransformerMNN,OTSfMGreyFPNScanNet,MegaDepth✓
SiLK✓OptionalNoNoMatching Succ.NoMNNRand. Homo.GreyGenericAny Image Set✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.236, + 0.892, + 0.263 + ], + "angle": 0, + "content": "Table 1. Non-exhaustive deconstruction of keypoint detectors along different dimensions. On each, SiLK adopts the simplest choice or is flexible to different choices. In particular, SiLK does not depend strongly on backbone type or training data (Tab. 7 & Tab. 9)." + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.285, + 0.857, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.487, + 0.794, + 0.502 + ], + "angle": 0, + "content": "Figure 3. Architecture for SiLK" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.892, + 0.619 + ], + "angle": 0, + "content": "A common approach [16, 50, 39, 17] for obtaining keypoint probabilities is to use a softmax cell-based approach. A cell is a \\( N \\times N \\) patch in which the probability of each cell position is determined by a local softmax. The softmax operates on \\( N \\times N + 1 \\) bins; \\( +1 \\) being the dustbin, handling the case of cells devoid of keypoints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.892, + 0.724 + ], + "angle": 0, + "content": "SiLK's approach is equivalent to a cell-based approach with a cell size of \\( N = 1 \\). This has several consequences. First, the softmax formulation becomes a sigmoid \\( \\sigma(x) = \\frac{1}{1 + e^{-x}} \\). Second, it removes the sparsity constraint that keypoints are exclusive events inside a cell. And third, this removes a free parameter (the cell size \\( N \\)) that we do not have to later tune." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.726, + 0.892, + 0.787 + ], + "angle": 0, + "content": "In the same spirit, SiLK does not use NMS during inference. Even though NMS is an established pruning technique that aims to spread out keypoints [16], we find SiLK doesn't need NMS to perform (cf. Tab. 2)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.798, + 0.848, + 0.815 + ], + "angle": 0, + "content": "3.3. Descriptors Define Matching Probability" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.822, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Similar to [47, 50], we model the cycle matching probability using a double softmax (i.e. the probability of matching \\( i \\) to \\( j \\), and back)." + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.886, + 0.768, + 0.902 + ], + "angle": 0, + "content": "\\[\nP _ {i \\leftrightarrow j} = P _ {i \\rightarrow j} P _ {i \\leftarrow j}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22501" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.199 + ], + "angle": 0, + "content": "where \\( P_{i\\rightarrow j} \\) is the directional probability of matching the \\( i \\)th descriptor from image \\( I \\) to the \\( j \\)th descriptor in image \\( I' \\). \\( P_{i\\leftarrow j} \\) is similar, but in the reverse direction. Both forward and backward probabilities are modeled as a softmax with temperature over descriptor cosine similarities. For fixed \\( i \\), \\( P_{i\\rightarrow j} \\) is a softmax over the \\( i \\)-th row of the descriptor similarity matrix, and \\( P_{i\\leftarrow j} \\) takes softmax over column \\( j \\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.206, + 0.182, + 0.222 + ], + "angle": 0, + "content": "3.4. Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.229, + 0.47, + 0.319 + ], + "angle": 0, + "content": "Self-Supervision. Pixel-accurate correspondences are required during training. Similar to SuperPoint [16], we obtain image-pair correspondences by applying a random transformation (homography) to an image. However, the homography is a linear mapping that gives correspondences at a subpixel level." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.47, + 0.456 + ], + "angle": 0, + "content": "To obtain pixel correspondences, we apply the sampled homography to all the pixel positions of image \\(I\\); positions being the center of pixels (e.g. the top-left pixel has position \\((0.5, 0.5)\\)). This first step establishes dense directional correspondences from \\(I\\) to \\(I'\\). We then run the same process from \\(I'\\) to \\(I\\) (using the inverse homography this time). Once both directional correspondences are obtained, the resulting positions are discretized; out-of-bound and non-bijective correspondences are discarded." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.47, + 0.531 + ], + "angle": 0, + "content": "Image Augmentations. As done in [16, 42], we employ image augmentation to improve robustness; augmentations include random brightness, contrast, gaussian noise, speckle noise and motion blur. We refer to supplementary materials for details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.47, + 0.652 + ], + "angle": 0, + "content": "Negative and Positive Selection. One defining property of a keypoint is distinctiveness – the point can be reliably distinguished from its peers. In our case this means the point can be reliably identified in a matching algorithm, similar to [49, 50]. With that in mind, we adopt a similar supervision technique as [49]. Keypoints that are correctly matched (using the currently trained descriptors) are labelled as positive; otherwise negative." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.652, + 0.47, + 0.729 + ], + "angle": 0, + "content": "Descriptor and Keypoint Losses. Similar to [47, 50], the descriptor loss is the negative log-likelihood loss applied to the matching probabilities for the positive round-trips (i.e. paths from point \\( i \\) to its location \\( i' \\) in the transformed image and back again)." + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.739, + 0.395, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d e s c} = - \\left(\\log P _ {i \\rightarrow i ^ {\\prime}} + \\log P _ {i \\leftarrow i ^ {\\prime}}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.871 + ], + "angle": 0, + "content": "This implicitly penalizes non-positive paths via softmax. One might notice \\(\\mathcal{L}_{desc}\\) requires the computation of a large matrix (size \\(HW\\times HW\\)). To handle potential GPU out-of-memory, we provide a simple, yet efficient implementation which computes the similarity matrix in a block-wise fashion, and recomputes block dot products instead of storing them for backpropagation (in the same vein as [36, 25])." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The keypoint loss \\(\\mathcal{L}_{key}\\) is a simple binary cross-entropy loss applied to a logistic sigmoid, in contrast to [49]. It" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "is trained to identify keypoints with successful round-trip matches (defined by mutual-nearest-neighbor) among all others (unsuccessful)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.149, + 0.634, + 0.166 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.175, + 0.892, + 0.326 + ], + "angle": 0, + "content": "In this section, we empirically evaluate SiLK together with representative baselines and state-of-the-art methods. On HPatches we evaluate two complementary keypoint quality metrics (Repeatability, Mean Matching Accuracy [33]) and planar stereo estimation capabilities (Homography Estimation). In addition, we benchmark on three real-world stereo tasks: outdoor camera pose estimation on Image Matching Challenge (IMC) 2022, and both indoor camera pose estimation and 3D point-cloud registration on ScanNet. In these experiments, we study the following:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.892, + 0.476 + ], + "angle": 0, + "content": "(1) Many methods employ complex strategies to learn and predict good keypoints and descriptors, including elements like multi-stage training, cell-based priors, complicated post-processing, context aggregation, and groundtruth 3D pose supervision (see Tab. 1), in various combinations. What machinery is necessary? SiLK contains no such machinery, and can be viewed as a reduction from these methods. However, SiLK either achieves new SOTA or compares very favorably. This questions the need for complex schemes for the evaluated tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.643 + ], + "angle": 0, + "content": "(2) We observed rather strong performance from engineered features (e.g. SIFT in Tab. 2&Tab. 6) vs learned methods. This motivates us to revisit design choices in a learned keypoint detector: What makes a good keypoint detector? We ablate each component (data, backbone, etc.) and test generalization performance under various conditions (e.g. input size, test data, task, etc). SiLK proves very robust to these choices (Sec. 4.6). In particular, a very lightweight version of SiLK (two 3x3 convolution layers) is competitive to SOTA on homography estimation, camera pose estimation and point cloud registration (Tab. 7&Tab. 8)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.734 + ], + "angle": 0, + "content": "We hope these results can serve the community and help adapt keypoint models to their tasks and needs. For example, labeling tasks (e.g. self-training [37], object pose [38]) might focus on high accuracy (i.e. larger backbone and denser keypoint selection), while tasks requiring speed (e.g. SLAM [34]) might find our lightweight backbone attractive." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.742, + 0.719, + 0.758 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Our own training pipeline has been used for all experiments with SiLK, as well as our reproduced SuperPoint (Tab. 2) results. Training time is ~5 hours on 2 Tesla V100-SXM2 GPUs using our default setup." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Default Setup. Unless specified otherwise, all results use the following setup. Trained on COCO [29] images (randomly sampled), with Adam [24] optimizer with learning rate \\(1e^{-4}\\) and betas \\((0.9, 0.999)\\); trained for \\(100\\mathrm{k}\\) iterations; batch size of 1 per GPU; dense descriptor map" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22502" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "is \\(146 \\times 146\\) for all architectures and input resolutions; cosine similarities scaled by temperature \\(20^{-1}\\); VGGnp-4 backbone (VGG architecture with max-pooling removed, details in Sec. 4.6); sparse keypoints obtained with top-k (\\(k = 10000\\)); detection head is 13x3 convolution (128-dims), and 11x1 convolution; descriptor head is 13x3 convolution (128-dims), and 11x1 convolution (128-dims out); no padding in convolutions; ReLU and batchnorm used as non-linearity and normalization (see supplementary)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.24, + 0.383, + 0.258 + ], + "angle": 0, + "content": "4.2. HPatches Homography Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.264, + 0.47, + 0.369 + ], + "angle": 0, + "content": "Following [16, 42, 39, 47], we evaluate homography estimation on HPatches [3]. HPatches contains 57 scenes (of 6 images) with significant illumination changes and 59 scenes with large viewpoint variations. Images in each scene are related by groundtruth homographies. We follow LoFTR [47], (currently SOTA on HPatches), and scale the shorter image edge to 480 at inference time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.47, + 0.567 + ], + "angle": 0, + "content": "Evaluation Protocol. For every image pairs, the model detects a set of keypoints. These are desired to be distinctive and thus repeatably detected across views. We use Repeatability to evaluate detection performance as in [16]. To test invariance of keypoint descriptors, each model's preferred matching algorithm establishes correspondence across images to obtain a subset of keypoints. We distinguish this subset as post-matching and the entire set as pre-matching. The accuracy of each correspondence is evaluated by Mean Matching Accuracy as in [17, 39, 50]. Finally, we use OpenCV RANSAC algorithm to estimate homography from matched keypoints, and evaluate Homography Estimation Accuracy [16] and Homography Estimation AUC [42, 47]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.47, + 0.763 + ], + "angle": 0, + "content": "Baselines. We compare SiLK against both sparse detector-based methods and dense detector-free methods. The sparse detector-based methods generally follow \"detect then match\": the model first detects a sparse set of distinctive points, then matches features. We include SIFT [31] as well as learned detectors SuperPoint [16] (both official release and our repro), R2D2 [39] and DISK [50]. On the other hand, dense detector-free perform \"detection by matching\": the model first extracts features, then applies a learned prematching CA module (e.g. GNN [42] or transformer [47]) to adapt the features to a specific pair of images, and then finds matches. While SiLK does not employ CA, we still include comparisons to the SOTA detector-free LoFTR [47]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Results. Despite its simplicity, SiLK outperforms all methods on repeatability, homography accuracy and homography estimation (Tab. 2). In particular, SiLK has a strong margin when the error threshold is small (\\(\\epsilon = 1\\)). This validates our pixel-accurate keypoint localization. SiLK lags only on the MMA@3 metric. This is caused by the pixel-accurate contrastive loss (Fig. 4). Consequently, SiLK does not benefit from increasing error threshold in MMA. In addition, even vs LoFTR (which uses dense features and CA)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "in Tab. 3, SiLK shows strong performance on Homography Estimation AUC and competitive performance on Homography accuracy. This questions the necessity of CA for these particular tasks; this may be valuable in applications which are particularly sensitive to runtime performance." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.184, + 0.697, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.184, + 0.882, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.311, + 0.892, + 0.38 + ], + "angle": 0, + "content": "Figure 4. The pixel-accurate contrastive loss results in very discriminative local features, thus reshaping the error distribution to make fewer local errors (MMA@1). This differs from interpolated, cell-based descriptors used in SuperPoint, which produces less accurate keypoints (MMA@3+)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.405, + 0.803, + 0.422 + ], + "angle": 0, + "content": "4.3. IMC 2022 outdoor pose estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.892, + 0.534 + ], + "angle": 0, + "content": "The Image Matching Challenge (IMC 2022) [22] provides pairs of outdoor images from different viewpoints; participants are required to estimate the fundamental matrix. Camera pose accuracy is then computed for ten thresholds of rotation and translation error (ranging from \\((1^{\\circ}, 20\\mathrm{cm})\\) to \\((10^{\\circ}, 5\\mathrm{m}))\\). Mean average accuracy (mAA) is reported by averaging across thresholds and scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.655 + ], + "angle": 0, + "content": "Evaluation Protocol. Sparse methods (DISK, SiLK) detect keypoints in individual images; mutual nearest neighbor is used to select matches from each pair. Methods with CA (SuperGlue, LoFTR) directly identify matches from each image pair. In either case, the fundamental matrix is estimated using MAGSAC [4]. The challenge allows different image sizes and tuning MAGSAC parameters [22]. We use 30k keypoints and MAGSAC threshold .25." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.892, + 0.746 + ], + "angle": 0, + "content": "Baselines. We consider the best leaderboard results from three baselines. (i) DISK is the winner of IMC 2020 and is SOTA among sparse methods. 2022 DISK results are from the IMC team (submission). We also take the best version of (ii) SuperGlue (submission) and (iii) LoFTR (submission) provided by the community." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.892, + 0.838 + ], + "angle": 0, + "content": "Results. We coarsely tune SiLK for this task for 30 trials (compared to LoFTR's 200 trials). SiLK again performs competitively (c.f. Tab. 4), outperforming DISK by a significant margin \\((+0.19 / +0.18\\mathrm{mAA})\\). SiLK also performs favorably compared to SuperGlue, which uses context aggregation and optimal transport matching." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.828, + 0.862 + ], + "angle": 0, + "content": "4.4. ScanNet: Indoor Pose & Point Clouds" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "ScanNet [12] is a large-scale dataset of 1513 indoor scenes of RGB-D images and ground-truth camera poses." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22503" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.129, + 0.089, + 0.842, + 0.212 + ], + "angle": 0, + "content": "
RepeatabilityHom. Est. Acc.Hom. Est. AUCMMA# of keypoints
ε = 1ε = 3ε = 1ε = 3ε = 1ε = 3ε = 1ε = 3pre-matchpost-match
SuperPoint (MagicLeap)0.340.610.430.80.20.510.410.72847499
SuperPoint (Ours)0.330.520.480.750.260.520.380.531143474
SIFT0.310.520.60.840.340.610.410.552189910
R2D20.360.720.450.790.20.50.340.7560881967
DISK0.380.690.450.80.220.520.520.8433491794
SiLK (top-10k)0.620.810.620.870.40.660.590.71100004283
SiLK (top-5k)0.560.760.60.850.390.640.570.6950002074
SiLK (top-1k)0.430.610.530.810.320.580.520.631000389
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.211, + 0.894, + 0.24 + ], + "angle": 0, + "content": "Table 2. SiLK achieves new SOTA on HPatches compared to other methods with sparse keypoints and features. Despite its simplicity, SiLK achieves higher performance on all metrics except MMA@3. We include the # of keypoints to ensure a fair comparison." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.26, + 0.48, + 0.335 + ], + "angle": 0, + "content": "
Hom. Est. Acc.Hom. Est. AUCMMA
ε = 1ε = 3ε = 1ε = 3ε = 1ε = 3
LoFTR (MegaDepth)0.650.870.370.650.640.91
LoFTR (ScanNet)0.240.570.070.330.360.76
SiLK (top-10k)0.620.870.40.660.590.71
SiLK (top-5k)0.600.850.390.640.570.69
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.335, + 0.47, + 0.404 + ], + "angle": 0, + "content": "Table 3. SiLK achieves competitive performance to SOTA LoFTR on HPatches, despite not using context aggregation. The top-5k SiLK has similar number of matches compared to outdoor LoFTR. We remark that LoFTR has large generalization gap when training on different types of dataset (indoor vs. outdoor)." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.416, + 0.418, + 0.502 + ], + "angle": 0, + "content": "
private mAApublic mAA
Sparse Features
DISK0.4910.502
SuperGlue0.6760.678
SiLK0.6850.684
Dense Features
LoFTR (MegaDepth)0.7350.721
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.502, + 0.47, + 0.543 + ], + "angle": 0, + "content": "Table 4. SiLK achieves new SOTA for sparse methods on IMC2022 and performs competitively to dense methods with context aggregation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.469, + 0.674 + ], + "angle": 0, + "content": "Using the official train/test split we evaluate both relative camera pose estimation and point-cloud registration. Relative camera pose estimation has been used in multiple previous works [52, 53, 6, 42, 47]. The task is to estimate the essential matrix with RANSAC from point matches in a pair of images. We report pose error AUC at thresholds \\((5^{\\circ},10^{\\circ},20^{\\circ})\\) using 20k keypoints and inlier threshold .5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.902 + ], + "angle": 0, + "content": "This protocol measures translation error in degrees and is known to suffer from scale ambiguity [52]. Angular translation error may be unstable, in particular if the underlying translation (in meters) is small. In response, recent works [19, 18] have introduced a 3D point-cloud registration task, using the ground-truth depth provided in ScanNet. A pair of images 20 frames apart is first sampled. Given this pair, a model predicts point matches. After matching, relative camera pose is estimated. Different from the previous protocol, ground-truth depth and camera intrinsics are now used to align matches in 3D. In addition to relative pose errors (reported separately for translation (cm) and rotation (degree)), the Chamfer distance (in cm) is measured between the registered point cloud and the groundtruth point cloud. We refer to the original papers [19, 18] for details on" + }, + { + "type": "table", + "bbox": [ + 0.55, + 0.26, + 0.846, + 0.406 + ], + "angle": 0, + "content": "
Pose Estimation AUC@5°@10°@20°
Sparse Features
D2-Net [17] + MNN5.314.528.0
SuperPoint [16] + MNN9.421.536.4
SuperPoint + PointCN [52]11.425.541.4
SuperPoint + OANet [53]11.826.943.9
SuperPoint + SuperGlue [42]16.233.851.8
SiLK + MNN18.034.450.4
Dense Features
DRC-Net [27]7.717.930.5
LoFTR (MegaDepth)16.933.650.6
LoFTR (ScanNet)21.540.857.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.406, + 0.892, + 0.448 + ], + "angle": 0, + "content": "Table 5. SiLK advances SOTA on relative pose estimation among sparse methods on ScanNet and performs competitively against dense method LoFTR." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.471, + 0.719, + 0.486 + ], + "angle": 0, + "content": "the evaluation setup and metrics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.504, + 0.724, + 0.519 + ], + "angle": 0, + "content": "4.4.1 Relative pose estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.892, + 0.707 + ], + "angle": 0, + "content": "Baselines. We compare SiLK with both sparse detector-based methods and dense detector-free methods. For the sparse methods, we consider local feature descriptors (R2D2, SuperPoint) with mutual nearest neighbor (MNN) for matching. SiLK falls into this category. In addition, we consider multiple learned context aggregation methods for matching that operates on SuperPoint, including PointCN [52], OANet [53] and SuperGlue. For detector-free dense methods, we consider DRC-Net [27] and LoFTR. We include two versions of LoFTR: one trained on MegaDepth with optimal transport post-processing and one trained on ScanNet." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.708, + 0.892, + 0.829 + ], + "angle": 0, + "content": "Results. As summarized in Tab. 5, SiLK significantly outperforms D2-Net (+12.7) and SuperPoint (+8.6) when using mutual nearest neighbor matching. In addition, SiLK outperforms the previous SOTA sparse method SuperGlue, despite its simpler design without context aggregation. SiLK performs similarly to LoFTR trained with MegaDepth. SiLK is only outperformed by LoFTR trained on ScanNet, the same as evaluation data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.808, + 0.862 + ], + "angle": 0, + "content": "4.4.2 Pairwise 3D point-cloud registration" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Baselines. We consider three main types of baselines. (i) Sparse Features + RANSAC We extract sparse key-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22504" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.348 + ], + "angle": 0, + "content": "points and their features from off-the-shelf models, and use RANSAC to estimate alignment. This includes SIFT and SuperPoint, and 3D geometry model FCGF [10]. (ii) Dense Feature Matching We follow the [18] and select high-quality corresponding pairs using the ratio test, and then solve a weighted Procustes problem [9, 23] to produce alignment. We add a dense version of SuperPoint by discarding the keypoint prediction. We include the current state-of-the-art URR [18] that learns invariant point descriptors through cross-view synthesis. By ignoring the keypoint scoring prediction, SiLK also belongs to this category. The goal is to evaluate the quality of the dense point features. (iii) Pose/Geometry Supervised We consider methods that use groundtruth poses to supervise, which are not required in (i) and (ii). These include LoFTR, DGR [9] and 3D MV Registration [19]. For LoFTR, we include only the MegaDepth model, since it performed better in this task." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.349, + 0.473, + 0.653 + ], + "angle": 0, + "content": "Results. As shown in Tab. 6, SiLK achieves new SOTA across all metrics (except rotation accuracy at \\(45^{\\circ}\\)). In particular, SiLK achieves very high accuracy at small thresholds (\\(5^{\\circ}\\) angular, 5cm translation and 1cm for chamfer), validating SiLK's pixel-level precision. Comparing with DGR, 3D MV Reg and LoFTR that use groundtruth camera poses during training, SiLK significantly outperforms, indicating that groundtruth 3D supervision is not necessary to train good keypoint features. We did not include the chamfer results for LoFTR as the provided positions do not match the required resolution for correct chamfer evaluation. SiLK also achieves superior performance vs previous SOTA URR. We note that URR requires two different frames sampled from the same scene during training, and is supervised by a differentiable cross-view rendering process. In contrast, SiLK only requires a single 2D image and is trained with a simple point matching loss. Finally, we observe that SuperPoint performs competitively when evaluated in this dense fashion; this is an important difference from the results reported in URR using sparse features." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.665, + 0.429, + 0.682 + ], + "angle": 0, + "content": "4.5. Discussion. What makes SiLK perform?" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.689, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Multiple factors contribute to SiLK's performance. In particular, the cycle-consistent loss and the dense predictions. First, the cycle-consistent loss (Fig. 2, supplementary) enforces the natural properties of keypoints: distinctiveness & robustness to viewpoint / photometric changes, while existing methods rely on proxy objectives or additional complexity (i.e. \"corners\" in SuperPoint[16] / SuperGlue[42], \"peakiness\" in R2D2[39], \"keypoint weighting scheme\" in D2-Net[17], \"RL cycle-consistent reward\" in DISK[50]). Removing unnecessary complexity helps the model focus its learning on the essentials, and is supported by consistent, strong results across benchmarks and metrics (Tab. 2, Tab. 3, Tab. 4, Tab. 5, Tab. 6). Second, SiLK produces dense, pixel-accurate keypoints, while all" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.168 + ], + "angle": 0, + "content": "existing methods have structural constraints (downsampling layers, cell-based schemes, and NMS, c.f. Tab. 1). This leads to more robust contrastive learning (i.e. more negative pairs), allows producing more keypoints matches and boosts results in the accurate regime (\\(\\epsilon = 1\\) in Tab. 2)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.179, + 0.83, + 0.196 + ], + "angle": 0, + "content": "4.6. What makes good keypoint detectors?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.203, + 0.895, + 0.324 + ], + "angle": 0, + "content": "Leveraging SiLK's flexibility (Sec. 3), we comprehensively ablate a large pool of design choices such as model architecture and image resolution. Surprisingly, we found that reducing architecture size, compute cost, and training input size only mildly impact model performance on homography estimation, camera pose estimation and point cloud registration. This benefits many important applications, such as on-device inference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.895, + 0.431 + ], + "angle": 0, + "content": "Here we discuss the key findings. We use the metrics Repeatability@1 (R), Homography Estimation Accuracy@1 (HA), Mean Matching Accuracy@1 (MMA) for HPatches, and Rotation Accuracy@5° (RA), Translation Accuracy@5cm (TA) and Chamfer@1 (C) for ScanNet, all at lowest error thresholds. Additional results and analysis are included in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.895, + 0.537 + ], + "angle": 0, + "content": "Agnostic to backbone. Existing methods use various backbones (Tab. 1); the effects on keypoint models are not well understood. We consider FPN from LoFTR and UNet from DISK; both are modern compared to SiLK's VGGnp [45] backbone. We find no empirical performance gain despite far greater parameter counts (Tab. 7). This questions the need for high-capacity models for these keypoint problems." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.895, + 0.809 + ], + "angle": 0, + "content": "Next we reduce the complexity of the original Super-Point backbone VGG-4. Max-pooling and up-sampling layers are removed. Our VGGnp-4 contains four convolution blocks, each with two convolution layers followed by ReLU. We discard convolution blocks from VGGnp-4 to obtain VGGnp-3, VGGnp-2 and VGGnp-1. On top of VGGnp-1, we reduce channel and descriptor sizes to 64 and 32 respectively and obtain an ultra-lightweight model, VGGnp-\\(\\mu\\). Results on all metrics (except MMA) only drop mildly as we shrink the model. In particular, SiLK (VGGnp-\\(\\mu\\) in Tab. 7) achieves very competitive performance (Tab. 2&Tab. 6). On the other hand, matching scores (MMA) drop significantly. We suggest two possible reasons: first, pointwise matching may benefit from the larger receptive field of deeper models. Second, homography estimation aggregates numerous pointwise measurements; homographies will improve if the noises cancel out, or if the outliers are removed by RANSAC." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Fast training on tiny images. By default, SiLK uses 146x146 descriptor map resolution during training. Higher resolution provides more points, which benefits the contrastive loss with more negatives, but also increases training time and GPU memory usage. Surprisingly, performance changes very little when varying resolution during" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22505" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.136, + 0.089, + 0.836, + 0.272 + ], + "angle": 0, + "content": "
RotationTranslationChamfer
Accuracy ↑Error ↓Accuracy ↑Error ↓Accuracy ↑Error ↓
5°10°45°MeanMed.51025MeanMed.1510MeanMed.
Sparse Features + RANSAC
SIFT [31]55.275.789.218.64.317.744.579.826.511.238.170.678.342.61.7
SuperPoint [16]65.586.996.68.93.621.251.788.016.19.745.781.188.219.21.2
FCGF [10]70.287.796.29.53.327.558.382.923.68.352.078.083.724.40.9
Pose/Geometry Supervised
DGR [9]81.189.394.89.41.854.576.288.718.44.570.585.589.013.70.4
3D MV Reg [19]87.793.297.06.01.269.083.191.811.72.978.989.291.810.20.2
LoFTR(MegaDepth) [47]91.796.899.42.81.265.985.297.16.03.3-----
Dense feature matching
SuperPoint [16]93.098.499.82.51.656.884.798.26.54.377.396.198.44.50.3
URR [18]92.795.898.53.40.877.289.696.17.32.386.094.696.15.90.1
SiLK (VGGnp-4)98.199.099.61.70.882.994.899.04.12.192.898.399.14.30.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.15, + 0.272, + 0.819, + 0.285 + ], + "angle": 0, + "content": "Table 6. SiLK achieves state-of-the-art on camera pose estimation and point cloud registration on ScanNet." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.307, + 0.468, + 0.417 + ], + "angle": 0, + "content": "
HPatchesScanNetModel
RHAcMMARATACParamFPSGFLOP
VGGnp-40.620.620.5998.182.992.8942k12.2370
VGGnp-30.630.610.5898.084.293.5868k12.5330
VGGnp-20.630.570.5597.383.592.4757k14.3268
VGGnp-10.630.570.4494.682.190.0461k18.990
VGGnp-μ0.640.560.4093.581.189.076k36.523
FPN [47]0.580.550.52---6.6M17.5298
UNet [50]0.600.410.58---1.3M25.9198
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.418, + 0.468, + 0.471 + ], + "angle": 0, + "content": "Table 7. SiLK is backbone agnostic. Backbones from existing methods are trained and evaluated. Low-capacity model performs well on all metrics except MMA. FPS and GFLOPs measured on \\(480 \\times 640\\) images with NVIDIA Quadro GP100 GPU." + }, + { + "type": "table", + "bbox": [ + 0.125, + 0.486, + 0.424, + 0.572 + ], + "angle": 0, + "content": "
SizeTimeHPatchesScanNet
RHAcMMARATAC
8221.7h0.600.580.5698.183.592.5
11422.7h0.620.620.5998.182.992.8
14625h0.630.590.5998.283.392.9
17829.5h0.630.620.5998.183.592.9
210218h0.630.610.6098.183.492.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.572, + 0.468, + 0.599 + ], + "angle": 0, + "content": "Table 8. Decreasing training image size has minimal impact. SiLK can be trained under 3h, with little performance drop." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.468, + 0.702 + ], + "angle": 0, + "content": "training (Tab. 8), especially on ScanNet. Tiny feature maps (82x82) remain competitive on both HPatches and ScanNet, and trains in 1.7 hours on two GPUs. This enables applications like test-time finetuning, on-device finetuning, and rapid experimental iteration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.468, + 0.884 + ], + "angle": 0, + "content": "Robustness to training data. Existing methods use various training sets (Tab. 1); empirically we observe cases of poor generalization across datasets. For example, LoFTR[47], trained on indoor ScanNet data, drops significantly vs LoFTR trained on outdoor MegaDepth data (Tab. 3) and vice-versa (Tab. 5). This overfitting may be exacerbated by the high-capacity machinery used by these methods, e.g. LoFTR's Transformer contextualizer. We measure SiLK's robustness on training data choices, by using different images from COCO[29], ImageNet[15], MegaDepth[28] and ScanNet[12]. We also combine them to formulate a diversified set of training data." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "SiLK is quite robust to change in training set, with the" + }, + { + "type": "table", + "bbox": [ + 0.552, + 0.307, + 0.842, + 0.405 + ], + "angle": 0, + "content": "
HPatchesScanNet
RHAcMMARATAC
COCO0.620.620.5998.182.992.8
ImageNet0.630.60.5998.183.593.0
MegaDep.0.620.610.5797.983.592.9
ScanNet0.600.550.4897.682.892.7
C+I+M+S0.610.590.5497.783.092.6
C+I+M0.640.60.5998.082.992.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.405, + 0.892, + 0.431 + ], + "angle": 0, + "content": "Table 9. SiLK is robust to different training sets. A noticeable drop is observed only when training on ScanNet." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.643 + ], + "angle": 0, + "content": "exception of ScanNet (Tab. 9). We hypothesize this is due to the significant amount of uniform surfaces (e.g. walls, doors) present in ScanNet. These featureless areas contain few keypoints to learn from. SiLK's drop agrees directionally with LoFTR's drop observed in Tab. 3, but the magnitude is smaller. This may be because SiLK's VGGnp backbone has lower capacity than LoFTR (FPN+Transformer), and hence is less susceptible to overfitting. Finally, we remark that SiLK trained with COCO is used for comparisons in Sec. 4.2, Sec. 4.3 and Sec. 4.4 for HPatches, IMC and ScanNet, whereas LoFTR requires different training data (MegaDepth or ScanNet) to achieve strong performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.662, + 0.618, + 0.678 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This paper presents SiLK, a simple and flexible framework for keypoint detection and descriptors. SiLK is designed from the principles of distinctiveness and invariance, and achieves or advances SOTA on key low-level tasks for 3D visual perception. SiLK's simplicity questions the need for complex machinery for good keypoint detection in low-level applications. In addition, extensive ablations reveal SiLK's robustness to backbone, training data and training input size. These findings lead to a tiny version of SiLK that is lightweight, accurate, and trains quickly. We view this \"tiny and learned\" regime as very promising for applications where runtime and/or power consumption is critical. We hope SiLK can draw attention to the field and facilitate stronger solutions." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22506" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Ebtsam Adel, Mohammed Elmogy, and Hazem Elbakry. Image stitching based on feature extraction techniques: a survey. International Journal of Computer Applications, 99(6):1-8, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.47, + 0.227 + ], + "angle": 0, + "content": "[2] Alexandre Alahi, Raphael Ortiz, and Pierre Vandergheynst. Freak: Fast retina keypoint. In 2012 IEEE conference on computer vision and pattern recognition, pages 510-517. IEEE, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.471, + 0.299 + ], + "angle": 0, + "content": "[3] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5173-5182, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.301, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[4] Daniel Barath, Jiri Matas, and Jana Noskova. MAGSAC: marginalizing sample consensus. In _Conference on Computer Vision and Pattern Recognition_, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.47, + 0.386 + ], + "angle": 0, + "content": "[5] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer vision and image understanding, 110(3):346-359, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[6] Eric Brachmann and Carsten Rother. Neural-guided ransac: Learning where to sample model hypotheses. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4321-4330, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.469, + 0.486 + ], + "angle": 0, + "content": "[7] Matthew A. Brown and David G. Lowe. Automatic panoramic image stitching using invariant features. International Journal of Computer Vision, 74:59-73, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.469, + 0.543 + ], + "angle": 0, + "content": "[8] Michael Calonder, Vincent Lepetit, Christoph Strecha, and Pascal Fua. Brief: Binary robust independent elementary features. In European conference on computer vision, pages 778-792. Springer, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[9] Christopher Choy, Wei Dong, and Vladlen Koltun. Deep global registration. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.469, + 0.628 + ], + "angle": 0, + "content": "[10] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 8957-8965, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[11] Peter Hviid Christiansen, Mikkel Fly Kragh, Yury Brodskiy, and Henrik Karstoft. Unsupperpoint: End-to-end unsupervised interest point detector and descriptor. arXiv preprint arXiv:1907.04011, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[13] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), volume 1, pages 886-893. IEEE, 2005." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[14] Andrew J. Davison, Ian D. Reid, Nicholas Molton, and Olivier Stasse. Monoslam: Real-time single camera slam. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29:1052-1067, 2007." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[16] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[17] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint detection and description of local features. arXiv preprint arXiv:1905.03561, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[18] Mohamed El Banani, Luya Gao, and Justin Johnson. Unsupervisedr&r: Unsupervised point cloud registration via differentiable rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7129-7139, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[19] Z. Gojcic, C. Zhou, J. D. Wegner, L. J. Guibas, and T. Birdal. Learning multiview 3d point cloud registration. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1756-1766, Los Alamitos, CA, USA, jun 2020. IEEE Computer Society." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[20] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In Alvey vision conference, volume 15, pages 10-5244. Manchester, UK, 1988." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.434, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[21] Wilfried Hartmann, Michal Havlena, and Konrad Schindler. Predicting matchability. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9-16, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[22] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image Matching across Wide Baselines: From Paper to Practice. International Journal of Computer Vision, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[23] W. Kabsch. A solution for the best rotation to relate two sets of vectors. Acta Crystallographica Section A, 32(5):922-923, Sep 1976." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[24] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[25] Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, and Daniel Haziza. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[26] Stefan Leutenegger, Margarita Chli, and Roland Y Siegwart. Brisk: Binary robust invariant scalable keypoints. In 2011 International conference on computer vision, pages 2548-2555. Ieee, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[27] Xinghui Li, Kai Han, Shuda Li, and Victor Prisacariu. Dual-resolution correspondence networks. In Conference on Neural Information Processing Systems (NeurIPS), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[28] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22507" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "European conference on computer vision, pages 740-755. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[30] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. Ssd: Single shot multibox detector. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision - ECCV 2016, pages 21-37, Cham, 2016. Springer International Publishing." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[31] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60(2):91-110, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[32] Jiayi Ma, Xingyu Jiang, Aoxiang Fan, Junjun Jiang, and Junchi Yan. Image matching from handcrafted to deep features: A survey. International Journal of Computer Vision, 129(1):23-79, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.348 + ], + "angle": 0, + "content": "[33] K. Mikolajczyk and C. Schmid. A performance evaluation of local descriptors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(10):1615-1630, 2005." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.39 + ], + "angle": 0, + "content": "[34] Raul Mur-Artal, J. M. M. Montiel, and Juan D. Tardós. Orb-slam: A versatile and accurate monocular slam system. IEEE Transactions on Robotics, 31:1147-1163, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.469, + 0.434 + ], + "angle": 0, + "content": "[35] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.469, + 0.474 + ], + "angle": 0, + "content": "[36] Markus N Rabe and Charles Staats. Self-attention does not need \\( o(n^{2}) \\) memory. arXiv preprint arXiv:2112.05682, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.469, + 0.561 + ], + "angle": 0, + "content": "[37] Umer Rafi, Andreas Doering, Bastian Leibe, and Juergen Gall. Self-supervised keypoint correspondences for multi-person pose estimation and tracking in videos. In Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XX, page 36–52, Berlin, Heidelberg, 2020. Springer-Verlag." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.644 + ], + "angle": 0, + "content": "[38] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10881-10891, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[39] Jerome Revaud, Philippe Weinzaepfel, César De Souza, Noe Pion, Gabriela Csurka, Yohann Cabon, and Martin Humenberger. R2d2: repeatable and reliable detector and descriptor. arXiv preprint arXiv:1906.06195, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.469, + 0.745 + ], + "angle": 0, + "content": "[40] Edward Rosten and Tom Drummond. Machine learning for high-speed corner detection. In European conference on computer vision, pages 430-443. Springer, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[41] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision, pages 2564-2571. Ieee, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[42] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.903 + ], + "angle": 0, + "content": "[43] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.891, + 0.12 + ], + "angle": 0, + "content": "network model. IEEE transactions on neural networks, 20(1):61-80, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[44] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[45] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[46] Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 21(2):343-348, 1967." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[47] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[48] Jiexiong Tang, Hanme Kim, Vitor Guizilini, Sudeep Pillai, and Rares Ambrus. Neural outlier rejection for self-supervised keypoint learning. arXiv preprint arXiv:1912.10615, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[49] Prune Truong, Stefanos Apostolopoulos, Agata Mosinska, Samuel Stucky, Carlos Ciller, and Sandro De Zanet. Glampoints: Greedily learned accurate match points. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10732-10741, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.515 + ], + "angle": 0, + "content": "[50] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.517, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[52] Kwang Moo Yi, Eduard Trulls, Yuki Ono, Vincent Lepetit, Mathieu Salzmann, and Pascal V. Fua. Learning to find good correspondences. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2666-2674, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.631, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[53] Jiahui Zhang, Dawei Sun, Zixin Luo, Anbang Yao, Lei Zhou, Tianwei Shen, Yurong Chen, Long Quan, and Honggen Liao. Learning two-view correspondences and geometry using order-aware network. International Conference on Computer Vision (ICCV), 2019." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22508" + } + ] +] \ No newline at end of file diff --git a/2023/SiLK_ Simple Learned Keypoints/545a50e4-a6f6-42d1-849e-c613959c21fd_origin.pdf b/2023/SiLK_ Simple Learned Keypoints/545a50e4-a6f6-42d1-849e-c613959c21fd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d83fb2007fd4b18cad069f5f238b358800479be6 --- /dev/null +++ b/2023/SiLK_ Simple Learned Keypoints/545a50e4-a6f6-42d1-849e-c613959c21fd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0ecf1c07404a91202a8b33a96d3e251bdc2af46705bccea1761dae887e0aeb5 +size 1197489 diff --git a/2023/SiLK_ Simple Learned Keypoints/full.md b/2023/SiLK_ Simple Learned Keypoints/full.md new file mode 100644 index 0000000000000000000000000000000000000000..aa5f162bf229777adff89ca50cb046dc3f1a3c06 --- /dev/null +++ b/2023/SiLK_ Simple Learned Keypoints/full.md @@ -0,0 +1,341 @@ +# SiLK: Simple Learned Keypoints + +Pierre Gleize + +gleize@meta.com + +Weiyao Wang + +weiyaowang@meta.com + +Matt Feiszli + +mdf@meta.com + +# Meta AI + +https://github.com/facebookresearch/silk + +# Abstract + +Keypoint detection & descriptors are foundational technologies for computer vision tasks like image matching, 3D reconstruction and visual odometry. Hand-engineered methods like Harris corners, SIFT, and HOG descriptors have been used for decades; more recently, there has been a trend to introduce learning in an attempt to improve keypoint detectors. On inspection however, the results are difficult to interpret; recent learning-based methods employ a vast diversity of experimental setups and design choices: empirical results are often reported using different backbones, protocols, datasets, types of supervisions or tasks. Since these differences are often coupled together, it raises a natural question on what makes a good learned keypoint detector. In this work, we revisit the design of existing keypoint detectors by deconstructing their methodologies and identifying the key components. We re-design each component from first-principle and propose Simple Learned Keypoints (SiLK) that is fully-differentiable, lightweight, and flexible. Despite its simplicity, SiLK advances new state-of-the-art on Detection Repeatability and Homography Estimation tasks on HPatches and 3D Point-Cloud Registration task on ScanNet, and achieves competitive performance to state-of-the-art on camera pose estimation in 2022 Image Matching Challenge and ScanNet. + +# 1. Introduction + +Keypoint detection and matching is a foundational computer vision technique to obtain a sparse yet informative representation of an image. Image stitching [7, 1], SLAM [14, 34], SfM [44], camera calibrations, tracking [37], and object detection [30] are important tasks built on keypoint correspondences [32]. A good keypoint model should be able to select a subset of points useful and informative to a specific task. One typically also wants robustness of the descriptor to some set of transformations (e.g. scale, viewpoint, or lighting variation). + +![](images/8a2961e714718bde680b4d90889dbc0bee62d73af33449ea5892171ef87d0224.jpg) + +![](images/cc025ba189fd5a287d8539e754d893377492c81c99bae68160a6b3d1df2ebdb3.jpg) + +![](images/81ab59d8cd198820ac22a4fa0a6e9258f6db0da5104c2516e9e2e8746ecb8b75.jpg) +Figure 1. The top image is an example of keypoint matching under viewpoint change; correct matches are green, incorrect ones are red. The bottom image shows keypoints which are cycle-consistent by SiLK. As can be observed, SiLK has learned to find distinctive geometric features (corners, curves, intersections,..); from single non-annotated images. + +Existing keypoint methods come in multiple forms and flavors (Tab. 1). However, those differences are often coupled together and not controlled for, which make it challenging to identify the source of gain. Our quest to finding the key-components of good keypoint detectors led to an in-depth review of alternative approaches, as well as the design, from first principle, of SiLK (Simple Learned Keypoints): A simple self-supervised approach to learn + +distinctive and robust keypoints from arbitrary image data in a traditional "detect-and-describe" framework. Despite its simplicity, SiLK is competitive or surpasses SOTA in most settings. + +Additionally, leveraging SiLK's simple one-stage training protocol and modular architecture, we are able to ablate various dimensions of detector performance for different tasks. In particular, with an eye toward real-time performance, we identify tasks where extremely lightweight backbone architectures are sufficient. + +# 2. Related Work + +Early work focused on carefully engineered methods to identify distinctive keypoints with descriptors which are robust to changes such as viewpoint and illumination. Hand-crafted techniques like Harriscorners[20], SIFT[31], ORB[41] and others [40, 8, 26, 2, 5, 13] have used explicit geometric notions like corners, gradients, and scale-space extrema to achieve results which remain both efficient and competitive to date [44, 34]. + +More recent work like SuperPoint[16] chose to learn to find corners; they generate a large set of synthetic shapes with annotated corners and train a model with this ground truth. While providing compelling evidence for learned methods, their training procedure is quite complex: it contains multiple training phases, a synthetic dataset, and employs a homography adaptation trick that can be difficult to tune (see our reproduced results in Tab. 2). + +In the same spirit as more recent work [17, 39, 50, 49, 21, 11, 48], SiLK aims to learn keypoints in simple end-to-end fashion, without explicitly defining them as corners. + +Several attempts have been made to learn keypoints implicitly; either by the careful design of the loss [17, 39, 11, 48]; or by directly predicting the matching success of descriptors [50, 21, 49]. SiLK falls in the second category, but with slight twist (cf. Sec. 3.4). + +To learn descriptors, contrastive losses are commonly used. Similar to [50, 47], SiLK adopts a probabilistic approach by modeling the matching probabilities in a double-softmax, cycle-consistency setting and optimizes the log likelihood. The probabilistic formulation, similar to InfoNCE [35], gives us a clean way to reason about matching, and the abundant supply of hard negative examples (from pixels in the same image) makes it an attractive choice. + +Context aggregation (CA) is a recent addition to the toolkit. Initiated by SuperGlue[42], CA aims to refine or transform descriptors from a pair of images before matching them. Implemented as a GNN[43] in [42], or as a Transformer[51] in [47], CA's predictions are conditional on all descriptors from the pair of images. In other words, for an image, the descriptors will be different when matching against different images. As a result, CA needs to run on every pair of images prior to matching, as opposed to + +running on single images. The run-time implications render CA prohibitively expensive in some applications (quadratic versus linear complexity). SiLK does not use CA, but outperforms [42] and performs competitively with [47]. Incorporating CA is optional for future works if performance is paramount and computation cost is less of a concern. + +As a postprocessing after CA, SuperGlue[42] introduced the concept of differentiable optimal transport (OT) to improve matching, using the Sinkhorn algorithm[46]. LoFTR [47] leveraged OT as well, but found little difference between OT and the simpler approach of mutual nearest neighbor (MNN) in some benchmarks. + +# 3. Methodology + +SiLK's contribution is simplicity and flexibility. Our solution is built on the traditional approach of identifying distinctive pixels via robust local descriptors. We use modern but established techniques to learn to localize and describe keypoints given an arbitrary source of unlabeled images. Unlike classical methods, our descriptors and invariances are learned, and unlike some modern methods, there is no particular complexity in the matching process (SiLK employs only cosine distances and mutual nearest neighbor); this leaves few structural hyperparameters to tune. The simple backbone+heads design is backbone-agnostic, allowing experimentation. The annotation-free SSL approach means SiLK can be trained on any image or video dataset. Finally, a simple, one-stage training pipeline allows us to easily train and ablate different architectures, datasets, and hyperparameters for different tasks. + +SiLK is trained to identify keypoints from single grayscale images. It provides both keypoint detections (location) and keypoint descriptors (for matching). Cycle consistency is employed for descriptor learning and a binary classifier identifies distinctive keypoints at pixel-level. + +To learn descriptors, we take a source image and a transformed copy, extract descriptors for each point, and use descriptor similarity to define transition probabilities from a source location to each transformed location (and vice-versa). We optimize the descriptors to maximize cycle-consistency; i.e. we maximize the probability of a round-trip from the source to its transformed location and back. + +To locate good keypoints, we train a binary classifier to identify points which will satisfy a matching criterion. A point and its transformed copy are positives when they are mutual nearest neighbors in the sense of transition probabilities, and they are negatives otherwise. We train both the cycle-consistency and classification losses jointly. + +We provide simple pseudo-code in Fig. 2. + +# 3.1. Architecture + +The SiLK architecture (Fig. 3) follows the "detect-and-describe" architecture originally proposed by SuperPoint + +```python +def loss(image_0, model): + # get warped image and pixel correspondences + image_1, corr_0, corr_1 = rand_homo(image_0) + # apply image augmentations + image_0 = augment(image_0) + image_1 = augment(image_1) + # extract dense descriptors and keypoints + desc_0, kpts_0 = model(image_0) + desc_1, kpts_1 = model(image_1) + # compute similarity matrix + sim_mat = cosine desc_0, desc_1) + # compute the descriptor loss + # using ground truth correspondences + loss_desc = nll(sim_mat, corr_0, corr_1) + # measure matching success + y = is_MATCH_success(sim_mat, corr_0, corr_1) + # compute keypoint loss + # using current matching success + loss_kpts = bce(kpts_0, y, corr_0) + loss_kpts += bce(kpts_1, y, corr_1) + return loss_desc + loss_kpts +``` + +[16]. A dense feature map is first extracted by feeding an image to an encoder backbone. The shared feature map is then fed to two heads: The keypoint head extracts the logits (used to calculate the dense keypoint probabilities), while the descriptor head extracts a dense descriptor map (used to calculate keypoint similarities). The model is backbone-agnostic and can easily be swapped. + +# 3.2. High Matching Probability Defines Keypoints + +As mentioned above, the keypoint probability estimate predicts the probability that a pixel will be correctly matched (i.e. survive a round-trip). Points with the highest likelihood of matching correctly are exactly those which we select as keypoints. + +
Keypoint DetectionDescriptors & MatchingModel & Training
LearnedSparseCell-basedNMSSupervisionCAMatcherSupervisionInputBackboneDataE2E
SIFTNo✓No✓-NoMNN-----
SuperPoint✓✓✓✓Homo. Adapt.NoMNNRand. Homo.GreyVGGCOCONo
SuperGlue-✓---GraphNNOTSfMGreyVGGOxford & ParisNo
GLAMPoints✓✓No✓Matching Succ.NoMNN-GreyUNetSlitLamp✓
D2-Net✓✓✓✓Triplet RankingNoMNNSfMRGBVGGMegaDepth✓
R2D2✓✓✓✓Matching APNoMNNSfMRGBL2-NetAachen✓
DISK✓✓✓✓Matching Succ.NoMNNSfMRGBUNetMegaDepth✓
URR-No----MNN3D RenderingRGBResNetScanNet✓
LoFTR-No---TransformerMNN,OTSfMGreyFPNScanNet,MegaDepth✓
SiLK✓OptionalNoNoMatching Succ.NoMNNRand. Homo.GreyGenericAny Image Set✓
+ +Table 1. Non-exhaustive deconstruction of keypoint detectors along different dimensions. On each, SiLK adopts the simplest choice or is flexible to different choices. In particular, SiLK does not depend strongly on backbone type or training data (Tab. 7 & Tab. 9). + +![](images/a16e87181d3c935e6110d68824ccea98d1a2a0d78bee4cf8a0b6ff98b04e153e.jpg) +Figure 2. Pseudo-code: learning keypoints from a single image. +Figure 3. Architecture for SiLK + +A common approach [16, 50, 39, 17] for obtaining keypoint probabilities is to use a softmax cell-based approach. A cell is a $N \times N$ patch in which the probability of each cell position is determined by a local softmax. The softmax operates on $N \times N + 1$ bins; $+1$ being the dustbin, handling the case of cells devoid of keypoints. + +SiLK's approach is equivalent to a cell-based approach with a cell size of $N = 1$ . This has several consequences. First, the softmax formulation becomes a sigmoid $\sigma(x) = \frac{1}{1 + e^{-x}}$ . Second, it removes the sparsity constraint that keypoints are exclusive events inside a cell. And third, this removes a free parameter (the cell size $N$ ) that we do not have to later tune. + +In the same spirit, SiLK does not use NMS during inference. Even though NMS is an established pruning technique that aims to spread out keypoints [16], we find SiLK doesn't need NMS to perform (cf. Tab. 2). + +# 3.3. Descriptors Define Matching Probability + +Similar to [47, 50], we model the cycle matching probability using a double softmax (i.e. the probability of matching $i$ to $j$ , and back). + +$$ +P _ {i \leftrightarrow j} = P _ {i \rightarrow j} P _ {i \leftarrow j} +$$ + +where $P_{i\rightarrow j}$ is the directional probability of matching the $i$ th descriptor from image $I$ to the $j$ th descriptor in image $I'$ . $P_{i\leftarrow j}$ is similar, but in the reverse direction. Both forward and backward probabilities are modeled as a softmax with temperature over descriptor cosine similarities. For fixed $i$ , $P_{i\rightarrow j}$ is a softmax over the $i$ -th row of the descriptor similarity matrix, and $P_{i\leftarrow j}$ takes softmax over column $j$ . + +# 3.4. Training + +Self-Supervision. Pixel-accurate correspondences are required during training. Similar to SuperPoint [16], we obtain image-pair correspondences by applying a random transformation (homography) to an image. However, the homography is a linear mapping that gives correspondences at a subpixel level. + +To obtain pixel correspondences, we apply the sampled homography to all the pixel positions of image $I$ ; positions being the center of pixels (e.g. the top-left pixel has position $(0.5, 0.5)$ ). This first step establishes dense directional correspondences from $I$ to $I'$ . We then run the same process from $I'$ to $I$ (using the inverse homography this time). Once both directional correspondences are obtained, the resulting positions are discretized; out-of-bound and non-bijective correspondences are discarded. + +Image Augmentations. As done in [16, 42], we employ image augmentation to improve robustness; augmentations include random brightness, contrast, gaussian noise, speckle noise and motion blur. We refer to supplementary materials for details. + +Negative and Positive Selection. One defining property of a keypoint is distinctiveness – the point can be reliably distinguished from its peers. In our case this means the point can be reliably identified in a matching algorithm, similar to [49, 50]. With that in mind, we adopt a similar supervision technique as [49]. Keypoints that are correctly matched (using the currently trained descriptors) are labelled as positive; otherwise negative. + +Descriptor and Keypoint Losses. Similar to [47, 50], the descriptor loss is the negative log-likelihood loss applied to the matching probabilities for the positive round-trips (i.e. paths from point $i$ to its location $i'$ in the transformed image and back again). + +$$ +\mathcal {L} _ {d e s c} = - \left(\log P _ {i \rightarrow i ^ {\prime}} + \log P _ {i \leftarrow i ^ {\prime}}\right) +$$ + +This implicitly penalizes non-positive paths via softmax. One might notice $\mathcal{L}_{desc}$ requires the computation of a large matrix (size $HW\times HW$ ). To handle potential GPU out-of-memory, we provide a simple, yet efficient implementation which computes the similarity matrix in a block-wise fashion, and recomputes block dot products instead of storing them for backpropagation (in the same vein as [36, 25]). + +The keypoint loss $\mathcal{L}_{key}$ is a simple binary cross-entropy loss applied to a logistic sigmoid, in contrast to [49]. It + +is trained to identify keypoints with successful round-trip matches (defined by mutual-nearest-neighbor) among all others (unsuccessful). + +# 4. Experiments + +In this section, we empirically evaluate SiLK together with representative baselines and state-of-the-art methods. On HPatches we evaluate two complementary keypoint quality metrics (Repeatability, Mean Matching Accuracy [33]) and planar stereo estimation capabilities (Homography Estimation). In addition, we benchmark on three real-world stereo tasks: outdoor camera pose estimation on Image Matching Challenge (IMC) 2022, and both indoor camera pose estimation and 3D point-cloud registration on ScanNet. In these experiments, we study the following: + +(1) Many methods employ complex strategies to learn and predict good keypoints and descriptors, including elements like multi-stage training, cell-based priors, complicated post-processing, context aggregation, and groundtruth 3D pose supervision (see Tab. 1), in various combinations. What machinery is necessary? SiLK contains no such machinery, and can be viewed as a reduction from these methods. However, SiLK either achieves new SOTA or compares very favorably. This questions the need for complex schemes for the evaluated tasks. + +(2) We observed rather strong performance from engineered features (e.g. SIFT in Tab. 2&Tab. 6) vs learned methods. This motivates us to revisit design choices in a learned keypoint detector: What makes a good keypoint detector? We ablate each component (data, backbone, etc.) and test generalization performance under various conditions (e.g. input size, test data, task, etc). SiLK proves very robust to these choices (Sec. 4.6). In particular, a very lightweight version of SiLK (two 3x3 convolution layers) is competitive to SOTA on homography estimation, camera pose estimation and point cloud registration (Tab. 7&Tab. 8). + +We hope these results can serve the community and help adapt keypoint models to their tasks and needs. For example, labeling tasks (e.g. self-training [37], object pose [38]) might focus on high accuracy (i.e. larger backbone and denser keypoint selection), while tasks requiring speed (e.g. SLAM [34]) might find our lightweight backbone attractive. + +# 4.1. Implementation Details + +Our own training pipeline has been used for all experiments with SiLK, as well as our reproduced SuperPoint (Tab. 2) results. Training time is ~5 hours on 2 Tesla V100-SXM2 GPUs using our default setup. + +Default Setup. Unless specified otherwise, all results use the following setup. Trained on COCO [29] images (randomly sampled), with Adam [24] optimizer with learning rate $1e^{-4}$ and betas $(0.9, 0.999)$ ; trained for $100\mathrm{k}$ iterations; batch size of 1 per GPU; dense descriptor map + +is $146 \times 146$ for all architectures and input resolutions; cosine similarities scaled by temperature $20^{-1}$ ; VGGnp-4 backbone (VGG architecture with max-pooling removed, details in Sec. 4.6); sparse keypoints obtained with top-k ( $k = 10000$ ); detection head is 13x3 convolution (128-dims), and 11x1 convolution; descriptor head is 13x3 convolution (128-dims), and 11x1 convolution (128-dims out); no padding in convolutions; ReLU and batchnorm used as non-linearity and normalization (see supplementary). + +# 4.2. HPatches Homography Estimation + +Following [16, 42, 39, 47], we evaluate homography estimation on HPatches [3]. HPatches contains 57 scenes (of 6 images) with significant illumination changes and 59 scenes with large viewpoint variations. Images in each scene are related by groundtruth homographies. We follow LoFTR [47], (currently SOTA on HPatches), and scale the shorter image edge to 480 at inference time. + +Evaluation Protocol. For every image pairs, the model detects a set of keypoints. These are desired to be distinctive and thus repeatably detected across views. We use Repeatability to evaluate detection performance as in [16]. To test invariance of keypoint descriptors, each model's preferred matching algorithm establishes correspondence across images to obtain a subset of keypoints. We distinguish this subset as post-matching and the entire set as pre-matching. The accuracy of each correspondence is evaluated by Mean Matching Accuracy as in [17, 39, 50]. Finally, we use OpenCV RANSAC algorithm to estimate homography from matched keypoints, and evaluate Homography Estimation Accuracy [16] and Homography Estimation AUC [42, 47]. + +Baselines. We compare SiLK against both sparse detector-based methods and dense detector-free methods. The sparse detector-based methods generally follow "detect then match": the model first detects a sparse set of distinctive points, then matches features. We include SIFT [31] as well as learned detectors SuperPoint [16] (both official release and our repro), R2D2 [39] and DISK [50]. On the other hand, dense detector-free perform "detection by matching": the model first extracts features, then applies a learned prematching CA module (e.g. GNN [42] or transformer [47]) to adapt the features to a specific pair of images, and then finds matches. While SiLK does not employ CA, we still include comparisons to the SOTA detector-free LoFTR [47]. + +Results. Despite its simplicity, SiLK outperforms all methods on repeatability, homography accuracy and homography estimation (Tab. 2). In particular, SiLK has a strong margin when the error threshold is small ( $\epsilon = 1$ ). This validates our pixel-accurate keypoint localization. SiLK lags only on the MMA@3 metric. This is caused by the pixel-accurate contrastive loss (Fig. 4). Consequently, SiLK does not benefit from increasing error threshold in MMA. In addition, even vs LoFTR (which uses dense features and CA) + +in Tab. 3, SiLK shows strong performance on Homography Estimation AUC and competitive performance on Homography accuracy. This questions the necessity of CA for these particular tasks; this may be valuable in applications which are particularly sensitive to runtime performance. + +![](images/bf27379361492e49caf5812bdde8afd09ec6a78ad275f247f89f82258db8243d.jpg) +Figure 4. The pixel-accurate contrastive loss results in very discriminative local features, thus reshaping the error distribution to make fewer local errors (MMA@1). This differs from interpolated, cell-based descriptors used in SuperPoint, which produces less accurate keypoints (MMA@3+). + +![](images/35603335e9a0419b19c6b8ed9978fbde35c266597b4785a95adc87b775fd255c.jpg) + +# 4.3. IMC 2022 outdoor pose estimation + +The Image Matching Challenge (IMC 2022) [22] provides pairs of outdoor images from different viewpoints; participants are required to estimate the fundamental matrix. Camera pose accuracy is then computed for ten thresholds of rotation and translation error (ranging from $(1^{\circ}, 20\mathrm{cm})$ to $(10^{\circ}, 5\mathrm{m}))$ . Mean average accuracy (mAA) is reported by averaging across thresholds and scenes. + +Evaluation Protocol. Sparse methods (DISK, SiLK) detect keypoints in individual images; mutual nearest neighbor is used to select matches from each pair. Methods with CA (SuperGlue, LoFTR) directly identify matches from each image pair. In either case, the fundamental matrix is estimated using MAGSAC [4]. The challenge allows different image sizes and tuning MAGSAC parameters [22]. We use 30k keypoints and MAGSAC threshold .25. + +Baselines. We consider the best leaderboard results from three baselines. (i) DISK is the winner of IMC 2020 and is SOTA among sparse methods. 2022 DISK results are from the IMC team (submission). We also take the best version of (ii) SuperGlue (submission) and (iii) LoFTR (submission) provided by the community. + +Results. We coarsely tune SiLK for this task for 30 trials (compared to LoFTR's 200 trials). SiLK again performs competitively (c.f. Tab. 4), outperforming DISK by a significant margin $(+0.19 / +0.18\mathrm{mAA})$ . SiLK also performs favorably compared to SuperGlue, which uses context aggregation and optimal transport matching. + +# 4.4. ScanNet: Indoor Pose & Point Clouds + +ScanNet [12] is a large-scale dataset of 1513 indoor scenes of RGB-D images and ground-truth camera poses. + +
RepeatabilityHom. Est. Acc.Hom. Est. AUCMMA# of keypoints
ε = 1ε = 3ε = 1ε = 3ε = 1ε = 3ε = 1ε = 3pre-matchpost-match
SuperPoint (MagicLeap)0.340.610.430.80.20.510.410.72847499
SuperPoint (Ours)0.330.520.480.750.260.520.380.531143474
SIFT0.310.520.60.840.340.610.410.552189910
R2D20.360.720.450.790.20.50.340.7560881967
DISK0.380.690.450.80.220.520.520.8433491794
SiLK (top-10k)0.620.810.620.870.40.660.590.71100004283
SiLK (top-5k)0.560.760.60.850.390.640.570.6950002074
SiLK (top-1k)0.430.610.530.810.320.580.520.631000389
+ +Table 2. SiLK achieves new SOTA on HPatches compared to other methods with sparse keypoints and features. Despite its simplicity, SiLK achieves higher performance on all metrics except MMA@3. We include the # of keypoints to ensure a fair comparison. + +
Hom. Est. Acc.Hom. Est. AUCMMA
ε = 1ε = 3ε = 1ε = 3ε = 1ε = 3
LoFTR (MegaDepth)0.650.870.370.650.640.91
LoFTR (ScanNet)0.240.570.070.330.360.76
SiLK (top-10k)0.620.870.40.660.590.71
SiLK (top-5k)0.600.850.390.640.570.69
+ +Table 3. SiLK achieves competitive performance to SOTA LoFTR on HPatches, despite not using context aggregation. The top-5k SiLK has similar number of matches compared to outdoor LoFTR. We remark that LoFTR has large generalization gap when training on different types of dataset (indoor vs. outdoor). + +
private mAApublic mAA
Sparse Features
DISK0.4910.502
SuperGlue0.6760.678
SiLK0.6850.684
Dense Features
LoFTR (MegaDepth)0.7350.721
+ +Using the official train/test split we evaluate both relative camera pose estimation and point-cloud registration. Relative camera pose estimation has been used in multiple previous works [52, 53, 6, 42, 47]. The task is to estimate the essential matrix with RANSAC from point matches in a pair of images. We report pose error AUC at thresholds $(5^{\circ},10^{\circ},20^{\circ})$ using 20k keypoints and inlier threshold .5. + +This protocol measures translation error in degrees and is known to suffer from scale ambiguity [52]. Angular translation error may be unstable, in particular if the underlying translation (in meters) is small. In response, recent works [19, 18] have introduced a 3D point-cloud registration task, using the ground-truth depth provided in ScanNet. A pair of images 20 frames apart is first sampled. Given this pair, a model predicts point matches. After matching, relative camera pose is estimated. Different from the previous protocol, ground-truth depth and camera intrinsics are now used to align matches in 3D. In addition to relative pose errors (reported separately for translation (cm) and rotation (degree)), the Chamfer distance (in cm) is measured between the registered point cloud and the groundtruth point cloud. We refer to the original papers [19, 18] for details on + +Table 4. SiLK achieves new SOTA for sparse methods on IMC2022 and performs competitively to dense methods with context aggregation. + +
Pose Estimation AUC@5°@10°@20°
Sparse Features
D2-Net [17] + MNN5.314.528.0
SuperPoint [16] + MNN9.421.536.4
SuperPoint + PointCN [52]11.425.541.4
SuperPoint + OANet [53]11.826.943.9
SuperPoint + SuperGlue [42]16.233.851.8
SiLK + MNN18.034.450.4
Dense Features
DRC-Net [27]7.717.930.5
LoFTR (MegaDepth)16.933.650.6
LoFTR (ScanNet)21.540.857.6
+ +Table 5. SiLK advances SOTA on relative pose estimation among sparse methods on ScanNet and performs competitively against dense method LoFTR. + +the evaluation setup and metrics. + +# 4.4.1 Relative pose estimation + +Baselines. We compare SiLK with both sparse detector-based methods and dense detector-free methods. For the sparse methods, we consider local feature descriptors (R2D2, SuperPoint) with mutual nearest neighbor (MNN) for matching. SiLK falls into this category. In addition, we consider multiple learned context aggregation methods for matching that operates on SuperPoint, including PointCN [52], OANet [53] and SuperGlue. For detector-free dense methods, we consider DRC-Net [27] and LoFTR. We include two versions of LoFTR: one trained on MegaDepth with optimal transport post-processing and one trained on ScanNet. + +Results. As summarized in Tab. 5, SiLK significantly outperforms D2-Net (+12.7) and SuperPoint (+8.6) when using mutual nearest neighbor matching. In addition, SiLK outperforms the previous SOTA sparse method SuperGlue, despite its simpler design without context aggregation. SiLK performs similarly to LoFTR trained with MegaDepth. SiLK is only outperformed by LoFTR trained on ScanNet, the same as evaluation data. + +# 4.4.2 Pairwise 3D point-cloud registration + +Baselines. We consider three main types of baselines. (i) Sparse Features + RANSAC We extract sparse key- + +points and their features from off-the-shelf models, and use RANSAC to estimate alignment. This includes SIFT and SuperPoint, and 3D geometry model FCGF [10]. (ii) Dense Feature Matching We follow the [18] and select high-quality corresponding pairs using the ratio test, and then solve a weighted Procustes problem [9, 23] to produce alignment. We add a dense version of SuperPoint by discarding the keypoint prediction. We include the current state-of-the-art URR [18] that learns invariant point descriptors through cross-view synthesis. By ignoring the keypoint scoring prediction, SiLK also belongs to this category. The goal is to evaluate the quality of the dense point features. (iii) Pose/Geometry Supervised We consider methods that use groundtruth poses to supervise, which are not required in (i) and (ii). These include LoFTR, DGR [9] and 3D MV Registration [19]. For LoFTR, we include only the MegaDepth model, since it performed better in this task. + +Results. As shown in Tab. 6, SiLK achieves new SOTA across all metrics (except rotation accuracy at $45^{\circ}$ ). In particular, SiLK achieves very high accuracy at small thresholds ( $5^{\circ}$ angular, 5cm translation and 1cm for chamfer), validating SiLK's pixel-level precision. Comparing with DGR, 3D MV Reg and LoFTR that use groundtruth camera poses during training, SiLK significantly outperforms, indicating that groundtruth 3D supervision is not necessary to train good keypoint features. We did not include the chamfer results for LoFTR as the provided positions do not match the required resolution for correct chamfer evaluation. SiLK also achieves superior performance vs previous SOTA URR. We note that URR requires two different frames sampled from the same scene during training, and is supervised by a differentiable cross-view rendering process. In contrast, SiLK only requires a single 2D image and is trained with a simple point matching loss. Finally, we observe that SuperPoint performs competitively when evaluated in this dense fashion; this is an important difference from the results reported in URR using sparse features. + +# 4.5. Discussion. What makes SiLK perform? + +Multiple factors contribute to SiLK's performance. In particular, the cycle-consistent loss and the dense predictions. First, the cycle-consistent loss (Fig. 2, supplementary) enforces the natural properties of keypoints: distinctiveness & robustness to viewpoint / photometric changes, while existing methods rely on proxy objectives or additional complexity (i.e. "corners" in SuperPoint[16] / SuperGlue[42], "peakiness" in R2D2[39], "keypoint weighting scheme" in D2-Net[17], "RL cycle-consistent reward" in DISK[50]). Removing unnecessary complexity helps the model focus its learning on the essentials, and is supported by consistent, strong results across benchmarks and metrics (Tab. 2, Tab. 3, Tab. 4, Tab. 5, Tab. 6). Second, SiLK produces dense, pixel-accurate keypoints, while all + +existing methods have structural constraints (downsampling layers, cell-based schemes, and NMS, c.f. Tab. 1). This leads to more robust contrastive learning (i.e. more negative pairs), allows producing more keypoints matches and boosts results in the accurate regime ( $\epsilon = 1$ in Tab. 2). + +# 4.6. What makes good keypoint detectors? + +Leveraging SiLK's flexibility (Sec. 3), we comprehensively ablate a large pool of design choices such as model architecture and image resolution. Surprisingly, we found that reducing architecture size, compute cost, and training input size only mildly impact model performance on homography estimation, camera pose estimation and point cloud registration. This benefits many important applications, such as on-device inference. + +Here we discuss the key findings. We use the metrics Repeatability@1 (R), Homography Estimation Accuracy@1 (HA), Mean Matching Accuracy@1 (MMA) for HPatches, and Rotation Accuracy@5° (RA), Translation Accuracy@5cm (TA) and Chamfer@1 (C) for ScanNet, all at lowest error thresholds. Additional results and analysis are included in supplementary. + +Agnostic to backbone. Existing methods use various backbones (Tab. 1); the effects on keypoint models are not well understood. We consider FPN from LoFTR and UNet from DISK; both are modern compared to SiLK's VGGnp [45] backbone. We find no empirical performance gain despite far greater parameter counts (Tab. 7). This questions the need for high-capacity models for these keypoint problems. + +Next we reduce the complexity of the original Super-Point backbone VGG-4. Max-pooling and up-sampling layers are removed. Our VGGnp-4 contains four convolution blocks, each with two convolution layers followed by ReLU. We discard convolution blocks from VGGnp-4 to obtain VGGnp-3, VGGnp-2 and VGGnp-1. On top of VGGnp-1, we reduce channel and descriptor sizes to 64 and 32 respectively and obtain an ultra-lightweight model, VGGnp- $\mu$ . Results on all metrics (except MMA) only drop mildly as we shrink the model. In particular, SiLK (VGGnp- $\mu$ in Tab. 7) achieves very competitive performance (Tab. 2&Tab. 6). On the other hand, matching scores (MMA) drop significantly. We suggest two possible reasons: first, pointwise matching may benefit from the larger receptive field of deeper models. Second, homography estimation aggregates numerous pointwise measurements; homographies will improve if the noises cancel out, or if the outliers are removed by RANSAC. + +Fast training on tiny images. By default, SiLK uses 146x146 descriptor map resolution during training. Higher resolution provides more points, which benefits the contrastive loss with more negatives, but also increases training time and GPU memory usage. Surprisingly, performance changes very little when varying resolution during + +
RotationTranslationChamfer
Accuracy ↑Error ↓Accuracy ↑Error ↓Accuracy ↑Error ↓
5°10°45°MeanMed.51025MeanMed.1510MeanMed.
Sparse Features + RANSAC
SIFT [31]55.275.789.218.64.317.744.579.826.511.238.170.678.342.61.7
SuperPoint [16]65.586.996.68.93.621.251.788.016.19.745.781.188.219.21.2
FCGF [10]70.287.796.29.53.327.558.382.923.68.352.078.083.724.40.9
Pose/Geometry Supervised
DGR [9]81.189.394.89.41.854.576.288.718.44.570.585.589.013.70.4
3D MV Reg [19]87.793.297.06.01.269.083.191.811.72.978.989.291.810.20.2
LoFTR(MegaDepth) [47]91.796.899.42.81.265.985.297.16.03.3-----
Dense feature matching
SuperPoint [16]93.098.499.82.51.656.884.798.26.54.377.396.198.44.50.3
URR [18]92.795.898.53.40.877.289.696.17.32.386.094.696.15.90.1
SiLK (VGGnp-4)98.199.099.61.70.882.994.899.04.12.192.898.399.14.30.1
+ +Table 6. SiLK achieves state-of-the-art on camera pose estimation and point cloud registration on ScanNet. + +
HPatchesScanNetModel
RHAcMMARATACParamFPSGFLOP
VGGnp-40.620.620.5998.182.992.8942k12.2370
VGGnp-30.630.610.5898.084.293.5868k12.5330
VGGnp-20.630.570.5597.383.592.4757k14.3268
VGGnp-10.630.570.4494.682.190.0461k18.990
VGGnp-μ0.640.560.4093.581.189.076k36.523
FPN [47]0.580.550.52---6.6M17.5298
UNet [50]0.600.410.58---1.3M25.9198
+ +Table 7. SiLK is backbone agnostic. Backbones from existing methods are trained and evaluated. Low-capacity model performs well on all metrics except MMA. FPS and GFLOPs measured on $480 \times 640$ images with NVIDIA Quadro GP100 GPU. + +
SizeTimeHPatchesScanNet
RHAcMMARATAC
8221.7h0.600.580.5698.183.592.5
11422.7h0.620.620.5998.182.992.8
14625h0.630.590.5998.283.392.9
17829.5h0.630.620.5998.183.592.9
210218h0.630.610.6098.183.492.8
+ +training (Tab. 8), especially on ScanNet. Tiny feature maps (82x82) remain competitive on both HPatches and ScanNet, and trains in 1.7 hours on two GPUs. This enables applications like test-time finetuning, on-device finetuning, and rapid experimental iteration. + +Robustness to training data. Existing methods use various training sets (Tab. 1); empirically we observe cases of poor generalization across datasets. For example, LoFTR[47], trained on indoor ScanNet data, drops significantly vs LoFTR trained on outdoor MegaDepth data (Tab. 3) and vice-versa (Tab. 5). This overfitting may be exacerbated by the high-capacity machinery used by these methods, e.g. LoFTR's Transformer contextualizer. We measure SiLK's robustness on training data choices, by using different images from COCO[29], ImageNet[15], MegaDepth[28] and ScanNet[12]. We also combine them to formulate a diversified set of training data. + +SiLK is quite robust to change in training set, with the + +Table 8. Decreasing training image size has minimal impact. SiLK can be trained under 3h, with little performance drop. + +
HPatchesScanNet
RHAcMMARATAC
COCO0.620.620.5998.182.992.8
ImageNet0.630.60.5998.183.593.0
MegaDep.0.620.610.5797.983.592.9
ScanNet0.600.550.4897.682.892.7
C+I+M+S0.610.590.5497.783.092.6
C+I+M0.640.60.5998.082.992.8
+ +Table 9. SiLK is robust to different training sets. A noticeable drop is observed only when training on ScanNet. + +exception of ScanNet (Tab. 9). We hypothesize this is due to the significant amount of uniform surfaces (e.g. walls, doors) present in ScanNet. These featureless areas contain few keypoints to learn from. SiLK's drop agrees directionally with LoFTR's drop observed in Tab. 3, but the magnitude is smaller. This may be because SiLK's VGGnp backbone has lower capacity than LoFTR (FPN+Transformer), and hence is less susceptible to overfitting. Finally, we remark that SiLK trained with COCO is used for comparisons in Sec. 4.2, Sec. 4.3 and Sec. 4.4 for HPatches, IMC and ScanNet, whereas LoFTR requires different training data (MegaDepth or ScanNet) to achieve strong performance. + +# 5. Conclusion + +This paper presents SiLK, a simple and flexible framework for keypoint detection and descriptors. SiLK is designed from the principles of distinctiveness and invariance, and achieves or advances SOTA on key low-level tasks for 3D visual perception. SiLK's simplicity questions the need for complex machinery for good keypoint detection in low-level applications. In addition, extensive ablations reveal SiLK's robustness to backbone, training data and training input size. These findings lead to a tiny version of SiLK that is lightweight, accurate, and trains quickly. We view this "tiny and learned" regime as very promising for applications where runtime and/or power consumption is critical. We hope SiLK can draw attention to the field and facilitate stronger solutions. + +# References + +[1] Ebtsam Adel, Mohammed Elmogy, and Hazem Elbakry. Image stitching based on feature extraction techniques: a survey. International Journal of Computer Applications, 99(6):1-8, 2014. +[2] Alexandre Alahi, Raphael Ortiz, and Pierre Vandergheynst. Freak: Fast retina keypoint. In 2012 IEEE conference on computer vision and pattern recognition, pages 510-517. IEEE, 2012. +[3] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5173-5182, 2017. +[4] Daniel Barath, Jiri Matas, and Jana Noskova. MAGSAC: marginalizing sample consensus. In _Conference on Computer Vision and Pattern Recognition_, 2019. +[5] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer vision and image understanding, 110(3):346-359, 2008. +[6] Eric Brachmann and Carsten Rother. Neural-guided ransac: Learning where to sample model hypotheses. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4321-4330, 2019. +[7] Matthew A. Brown and David G. Lowe. Automatic panoramic image stitching using invariant features. International Journal of Computer Vision, 74:59-73, 2006. +[8] Michael Calonder, Vincent Lepetit, Christoph Strecha, and Pascal Fua. Brief: Binary robust independent elementary features. In European conference on computer vision, pages 778-792. Springer, 2010. +[9] Christopher Choy, Wei Dong, and Vladlen Koltun. Deep global registration. In CVPR, 2020. +[10] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 8957-8965, 2019. +[11] Peter Hviid Christiansen, Mikkel Fly Kragh, Yury Brodskiy, and Henrik Karstoft. Unsupperpoint: End-to-end unsupervised interest point detector and descriptor. arXiv preprint arXiv:1907.04011, 2019. +[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. +[13] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), volume 1, pages 886-893. IEEE, 2005. +[14] Andrew J. Davison, Ian D. Reid, Nicholas Molton, and Olivier Stasse. Monoslam: Real-time single camera slam. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29:1052-1067, 2007. +[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image + +database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. +[16] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. +[17] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint detection and description of local features. arXiv preprint arXiv:1905.03561, 2019. +[18] Mohamed El Banani, Luya Gao, and Justin Johnson. Unsupervisedr&r: Unsupervised point cloud registration via differentiable rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7129-7139, 2021. +[19] Z. Gojcic, C. Zhou, J. D. Wegner, L. J. Guibas, and T. Birdal. Learning multiview 3d point cloud registration. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1756-1766, Los Alamitos, CA, USA, jun 2020. IEEE Computer Society. +[20] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In Alvey vision conference, volume 15, pages 10-5244. Manchester, UK, 1988. +[21] Wilfried Hartmann, Michal Havlena, and Konrad Schindler. Predicting matchability. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9-16, 2014. +[22] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image Matching across Wide Baselines: From Paper to Practice. International Journal of Computer Vision, 2020. +[23] W. Kabsch. A solution for the best rotation to relate two sets of vectors. Acta Crystallographica Section A, 32(5):922-923, Sep 1976. +[24] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2015. +[25] Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, and Daniel Haziza. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers, 2022. +[26] Stefan Leutenegger, Margarita Chli, and Roland Y Siegwart. Brisk: Binary robust invariant scalable keypoints. In 2011 International conference on computer vision, pages 2548-2555. Ieee, 2011. +[27] Xinghui Li, Kai Han, Shuda Li, and Victor Prisacariu. Dual-resolution correspondence networks. In Conference on Neural Information Processing Systems (NeurIPS), 2020. +[28] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. +[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In + +European conference on computer vision, pages 740-755. Springer, 2014. +[30] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. Ssd: Single shot multibox detector. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision - ECCV 2016, pages 21-37, Cham, 2016. Springer International Publishing. +[31] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60(2):91-110, 2004. +[32] Jiayi Ma, Xingyu Jiang, Aoxiang Fan, Junjun Jiang, and Junchi Yan. Image matching from handcrafted to deep features: A survey. International Journal of Computer Vision, 129(1):23-79, 2021. +[33] K. Mikolajczyk and C. Schmid. A performance evaluation of local descriptors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(10):1615-1630, 2005. +[34] Raul Mur-Artal, J. M. M. Montiel, and Juan D. Tardós. Orb-slam: A versatile and accurate monocular slam system. IEEE Transactions on Robotics, 31:1147-1163, 2015. +[35] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. +[36] Markus N Rabe and Charles Staats. Self-attention does not need $o(n^{2})$ memory. arXiv preprint arXiv:2112.05682, 2021. +[37] Umer Rafi, Andreas Doering, Bastian Leibe, and Juergen Gall. Self-supervised keypoint correspondences for multi-person pose estimation and tracking in videos. In Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XX, page 36–52, Berlin, Heidelberg, 2020. Springer-Verlag. +[38] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10881-10891, 2021. +[39] Jerome Revaud, Philippe Weinzaepfel, César De Souza, Noe Pion, Gabriela Csurka, Yohann Cabon, and Martin Humenberger. R2d2: repeatable and reliable detector and descriptor. arXiv preprint arXiv:1906.06195, 2019. +[40] Edward Rosten and Tom Drummond. Machine learning for high-speed corner detection. In European conference on computer vision, pages 430-443. Springer, 2006. +[41] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision, pages 2564-2571. Ieee, 2011. +[42] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. +[43] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural + +network model. IEEE transactions on neural networks, 20(1):61-80, 2008. +[44] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. +[45] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2015. +[46] Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 21(2):343-348, 1967. +[47] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. +[48] Jiexiong Tang, Hanme Kim, Vitor Guizilini, Sudeep Pillai, and Rares Ambrus. Neural outlier rejection for self-supervised keypoint learning. arXiv preprint arXiv:1912.10615, 2019. +[49] Prune Truong, Stefanos Apostolopoulos, Agata Mosinska, Samuel Stucky, Carlos Ciller, and Sandro De Zanet. Glampoints: Greedily learned accurate match points. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10732-10741, 2019. +[50] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. +[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. +[52] Kwang Moo Yi, Eduard Trulls, Yuki Ono, Vincent Lepetit, Mathieu Salzmann, and Pascal V. Fua. Learning to find good correspondences. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2666-2674, 2017. +[53] Jiahui Zhang, Dawei Sun, Zixin Luo, Anbang Yao, Lei Zhou, Tianwei Shen, Yurong Chen, Long Quan, and Honggen Liao. Learning two-view correspondences and geometry using order-aware network. International Conference on Computer Vision (ICCV), 2019. \ No newline at end of file diff --git a/2023/SiLK_ Simple Learned Keypoints/images.zip b/2023/SiLK_ Simple Learned Keypoints/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6c45e19d5cbd4d5056e8a11503aea0f706c9e154 --- /dev/null +++ b/2023/SiLK_ Simple Learned Keypoints/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80dbfa3edbbf5eb8e362afc2010ac7e1a1bc52a001e8bbb590907b9f0e2e1ebc +size 561996 diff --git a/2023/SiLK_ Simple Learned Keypoints/layout.json b/2023/SiLK_ Simple Learned Keypoints/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..087f18384e1e98b6d4e79a8cda35bdcc0fbd611e --- /dev/null +++ b/2023/SiLK_ Simple Learned Keypoints/layout.json @@ -0,0 +1,7871 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 192, + 103, + 402, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 103, + 402, + 121 + ], + "spans": [ + { + "bbox": [ + 192, + 103, + 402, + 121 + ], + "type": "text", + "content": "SiLK: Simple Learned Keypoints" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 144, + 130, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 144, + 130, + 156 + ], + "spans": [ + { + "bbox": [ + 64, + 144, + 130, + 156 + ], + "type": "text", + "content": "Pierre Gleize" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 160, + 139, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 160, + 139, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 160, + 139, + 171 + ], + "type": "text", + "content": "gleize@meta.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 270, + 144, + 341, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 144, + 341, + 158 + ], + "spans": [ + { + "bbox": [ + 270, + 144, + 341, + 158 + ], + "type": "text", + "content": "Weiyao Wang" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 252, + 160, + 358, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 160, + 358, + 171 + ], + "spans": [ + { + "bbox": [ + 252, + 160, + 358, + 171 + ], + "type": "text", + "content": "weiyaowang@meta.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 474, + 144, + 534, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 144, + 534, + 156 + ], + "spans": [ + { + "bbox": [ + 474, + 144, + 534, + 156 + ], + "type": "text", + "content": "Matt Feiszli" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 471, + 160, + 538, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 160, + 538, + 171 + ], + "spans": [ + { + "bbox": [ + 471, + 160, + 538, + 171 + ], + "type": "text", + "content": "mdf@meta.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 282, + 186, + 326, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 186, + 326, + 198 + ], + "spans": [ + { + "bbox": [ + 282, + 186, + 326, + 198 + ], + "type": "text", + "content": "Meta AI" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 159, + 200, + 449, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 200, + 449, + 213 + ], + "spans": [ + { + "bbox": [ + 159, + 200, + 449, + 213 + ], + "type": "text", + "content": "https://github.com/facebookresearch/silk" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 266, + 290, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 553 + ], + "type": "text", + "content": "Keypoint detection & descriptors are foundational technologies for computer vision tasks like image matching, 3D reconstruction and visual odometry. Hand-engineered methods like Harris corners, SIFT, and HOG descriptors have been used for decades; more recently, there has been a trend to introduce learning in an attempt to improve keypoint detectors. On inspection however, the results are difficult to interpret; recent learning-based methods employ a vast diversity of experimental setups and design choices: empirical results are often reported using different backbones, protocols, datasets, types of supervisions or tasks. Since these differences are often coupled together, it raises a natural question on what makes a good learned keypoint detector. In this work, we revisit the design of existing keypoint detectors by deconstructing their methodologies and identifying the key components. We re-design each component from first-principle and propose Simple Learned Keypoints (SiLK) that is fully-differentiable, lightweight, and flexible. Despite its simplicity, SiLK advances new state-of-the-art on Detection Repeatability and Homography Estimation tasks on HPatches and 3D Point-Cloud Registration task on ScanNet, and achieves competitive performance to state-of-the-art on camera pose estimation in 2022 Image Matching Challenge and ScanNet." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 574, + 128, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 128, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 128, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "Keypoint detection and matching is a foundational computer vision technique to obtain a sparse yet informative representation of an image. Image stitching [7, 1], SLAM [14, 34], SfM [44], camera calibrations, tracking [37], and object detection [30] are important tasks built on keypoint correspondences [32]. A good keypoint model should be able to select a subset of points useful and informative to a specific task. One typically also wants robustness of the descriptor to some set of transformations (e.g. scale, viewpoint, or lighting variation)." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 240, + 427, + 329 + ], + "blocks": [ + { + "bbox": [ + 307, + 240, + 427, + 329 + ], + "lines": [ + { + "bbox": [ + 307, + 240, + 427, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 427, + 329 + ], + "type": "image", + "image_path": "8a2961e714718bde680b4d90889dbc0bee62d73af33449ea5892171ef87d0224.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 427, + 240, + 541, + 329 + ], + "blocks": [ + { + "bbox": [ + 427, + 240, + 541, + 329 + ], + "lines": [ + { + "bbox": [ + 427, + 240, + 541, + 329 + ], + "spans": [ + { + "bbox": [ + 427, + 240, + 541, + 329 + ], + "type": "image", + "image_path": "cc025ba189fd5a287d8539e754d893377492c81c99bae68160a6b3d1df2ebdb3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 307, + 332, + 545, + 508 + ], + "blocks": [ + { + "bbox": [ + 307, + 332, + 545, + 508 + ], + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 508 + ], + "type": "image", + "image_path": "81ab59d8cd198820ac22a4fa0a6e9258f6db0da5104c2516e9e2e8746ecb8b75.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 522, + 547, + 589 + ], + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 589 + ], + "type": "text", + "content": "Figure 1. The top image is an example of keypoint matching under viewpoint change; correct matches are green, incorrect ones are red. The bottom image shows keypoints which are cycle-consistent by SiLK. As can be observed, SiLK has learned to find distinctive geometric features (corners, curves, intersections,..); from single non-annotated images." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Existing keypoint methods come in multiple forms and flavors (Tab. 1). However, those differences are often coupled together and not controlled for, which make it challenging to identify the source of gain. Our quest to finding the key-components of good keypoint detectors led to an in-depth review of alternative approaches, as well as the design, from first principle, of SiLK (Simple Learned Keypoints): A simple self-supervised approach to learn" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22499" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 119 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 119 + ], + "type": "text", + "content": "distinctive and robust keypoints from arbitrary image data in a traditional \"detect-and-describe\" framework. Despite its simplicity, SiLK is competitive or surpasses SOTA in most settings." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 121, + 286, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 121, + 286, + 191 + ], + "spans": [ + { + "bbox": [ + 48, + 121, + 286, + 191 + ], + "type": "text", + "content": "Additionally, leveraging SiLK's simple one-stage training protocol and modular architecture, we are able to ablate various dimensions of detector performance for different tasks. In particular, with an eye toward real-time performance, we identify tasks where extremely lightweight backbone architectures are sufficient." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 203, + 132, + 214 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 203, + 132, + 214 + ], + "spans": [ + { + "bbox": [ + 48, + 203, + 132, + 214 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 223, + 286, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 223, + 286, + 318 + ], + "spans": [ + { + "bbox": [ + 48, + 223, + 286, + 318 + ], + "type": "text", + "content": "Early work focused on carefully engineered methods to identify distinctive keypoints with descriptors which are robust to changes such as viewpoint and illumination. Hand-crafted techniques like Harriscorners[20], SIFT[31], ORB[41] and others [40, 8, 26, 2, 5, 13] have used explicit geometric notions like corners, gradients, and scale-space extrema to achieve results which remain both efficient and competitive to date [44, 34]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 319, + 286, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 286, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 286, + 413 + ], + "type": "text", + "content": "More recent work like SuperPoint[16] chose to learn to find corners; they generate a large set of synthetic shapes with annotated corners and train a model with this ground truth. While providing compelling evidence for learned methods, their training procedure is quite complex: it contains multiple training phases, a synthetic dataset, and employs a homography adaptation trick that can be difficult to tune (see our reproduced results in Tab. 2)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 415, + 286, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 286, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 286, + 449 + ], + "type": "text", + "content": "In the same spirit as more recent work [17, 39, 50, 49, 21, 11, 48], SiLK aims to learn keypoints in simple end-to-end fashion, without explicitly defining them as corners." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 450, + 286, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 450, + 286, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 450, + 286, + 509 + ], + "type": "text", + "content": "Several attempts have been made to learn keypoints implicitly; either by the careful design of the loss [17, 39, 11, 48]; or by directly predicting the matching success of descriptors [50, 21, 49]. SiLK falls in the second category, but with slight twist (cf. Sec. 3.4)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 510, + 286, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 286, + 605 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 286, + 605 + ], + "type": "text", + "content": "To learn descriptors, contrastive losses are commonly used. Similar to [50, 47], SiLK adopts a probabilistic approach by modeling the matching probabilities in a double-softmax, cycle-consistency setting and optimizes the log likelihood. The probabilistic formulation, similar to InfoNCE [35], gives us a clean way to reason about matching, and the abundant supply of hard negative examples (from pixels in the same image) makes it an attractive choice." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 606, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 606, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 606, + 286, + 712 + ], + "type": "text", + "content": "Context aggregation (CA) is a recent addition to the toolkit. Initiated by SuperGlue[42], CA aims to refine or transform descriptors from a pair of images before matching them. Implemented as a GNN[43] in [42], or as a Transformer[51] in [47], CA's predictions are conditional on all descriptors from the pair of images. In other words, for an image, the descriptors will be different when matching against different images. As a result, CA needs to run on every pair of images prior to matching, as opposed to" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 72, + 545, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 143 + ], + "type": "text", + "content": "running on single images. The run-time implications render CA prohibitively expensive in some applications (quadratic versus linear complexity). SiLK does not use CA, but outperforms [42] and performs competitively with [47]. Incorporating CA is optional for future works if performance is paramount and computation cost is less of a concern." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 144, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 144, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 144, + 545, + 215 + ], + "type": "text", + "content": "As a postprocessing after CA, SuperGlue[42] introduced the concept of differentiable optimal transport (OT) to improve matching, using the Sinkhorn algorithm[46]. LoFTR [47] leveraged OT as well, but found little difference between OT and the simpler approach of mutual nearest neighbor (MNN) in some benchmarks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 308, + 226, + 387, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 226, + 387, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 226, + 387, + 239 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 308, + 246, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 246, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 246, + 545, + 436 + ], + "type": "text", + "content": "SiLK's contribution is simplicity and flexibility. Our solution is built on the traditional approach of identifying distinctive pixels via robust local descriptors. We use modern but established techniques to learn to localize and describe keypoints given an arbitrary source of unlabeled images. Unlike classical methods, our descriptors and invariances are learned, and unlike some modern methods, there is no particular complexity in the matching process (SiLK employs only cosine distances and mutual nearest neighbor); this leaves few structural hyperparameters to tune. The simple backbone+heads design is backbone-agnostic, allowing experimentation. The annotation-free SSL approach means SiLK can be trained on any image or video dataset. Finally, a simple, one-stage training pipeline allows us to easily train and ablate different architectures, datasets, and hyperparameters for different tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 437, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 496 + ], + "type": "text", + "content": "SiLK is trained to identify keypoints from single grayscale images. It provides both keypoint detections (location) and keypoint descriptors (for matching). Cycle consistency is employed for descriptor learning and a binary classifier identifies distinctive keypoints at pixel-level." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 498, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 498, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 308, + 498, + 545, + 579 + ], + "type": "text", + "content": "To learn descriptors, we take a source image and a transformed copy, extract descriptors for each point, and use descriptor similarity to define transition probabilities from a source location to each transformed location (and vice-versa). We optimize the descriptors to maximize cycle-consistency; i.e. we maximize the probability of a round-trip from the source to its transformed location and back." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 581, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 581, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 308, + 581, + 545, + 652 + ], + "type": "text", + "content": "To locate good keypoints, we train a binary classifier to identify points which will satisfy a matching criterion. A point and its transformed copy are positives when they are mutual nearest neighbors in the sense of transition probabilities, and they are negatives otherwise. We train both the cycle-consistency and classification losses jointly." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 653, + 486, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 653, + 486, + 664 + ], + "spans": [ + { + "bbox": [ + 320, + 653, + 486, + 664 + ], + "type": "text", + "content": "We provide simple pseudo-code in Fig. 2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 671, + 387, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 671, + 387, + 682 + ], + "spans": [ + { + "bbox": [ + 308, + 671, + 387, + 682 + ], + "type": "text", + "content": "3.1. Architecture" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 689, + 545, + 712 + ], + "type": "text", + "content": "The SiLK architecture (Fig. 3) follows the \"detect-and-describe\" architecture originally proposed by SuperPoint" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22500" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 39, + 231, + 287, + 498 + ], + "blocks": [ + { + "bbox": [ + 39, + 231, + 287, + 498 + ], + "lines": [ + { + "bbox": [ + 39, + 231, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 39, + 231, + 287, + 498 + ], + "type": "text", + "content": "def loss(image_0, model):\n # get warped image and pixel correspondences\n image_1, corr_0, corr_1 = rand_homo(image_0)\n # apply image augmentations\n image_0 = augment(image_0)\n image_1 = augment(image_1)\n # extract dense descriptors and keypoints\n desc_0, kpts_0 = model(image_0)\n desc_1, kpts_1 = model(image_1)\n # compute similarity matrix\n sim_mat = cosine desc_0, desc_1)\n # compute the descriptor loss\n # using ground truth correspondences\n loss_desc = nll(sim_mat, corr_0, corr_1)\n # measure matching success\n y = is_MATCH_success(sim_mat, corr_0, corr_1)\n # compute keypoint loss\n # using current matching success\n loss_kpts = bce(kpts_0, y, corr_0)\n loss_kpts += bce(kpts_1, y, corr_1)\n return loss_desc + loss_kpts" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_body" + } + ], + "index": 0, + "sub_type": "code", + "guess_lang": "python" + }, + { + "bbox": [ + 46, + 537, + 287, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 537, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 537, + 287, + 622 + ], + "type": "text", + "content": "[16]. A dense feature map is first extracted by feeding an image to an encoder backbone. The shared feature map is then fed to two heads: The keypoint head extracts the logits (used to calculate the dense keypoint probabilities), while the descriptor head extracts a dense descriptor map (used to calculate keypoint similarities). The model is backbone-agnostic and can easily be swapped." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 633, + 284, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 284, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 284, + 647 + ], + "type": "text", + "content": "3.2. High Matching Probability Defines Keypoints" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "As mentioned above, the keypoint probability estimate predicts the probability that a pixel will be correctly matched (i.e. survive a round-trip). Points with the highest likelihood of matching correctly are exactly those which we select as keypoints." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 37, + 69, + 545, + 186 + ], + "blocks": [ + { + "bbox": [ + 37, + 69, + 545, + 186 + ], + "lines": [ + { + "bbox": [ + 37, + 69, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 37, + 69, + 545, + 186 + ], + "type": "table", + "html": "
Keypoint DetectionDescriptors & MatchingModel & Training
LearnedSparseCell-basedNMSSupervisionCAMatcherSupervisionInputBackboneDataE2E
SIFTNo✓No✓-NoMNN-----
SuperPoint✓✓✓✓Homo. Adapt.NoMNNRand. Homo.GreyVGGCOCONo
SuperGlue-✓---GraphNNOTSfMGreyVGGOxford & ParisNo
GLAMPoints✓✓No✓Matching Succ.NoMNN-GreyUNetSlitLamp✓
D2-Net✓✓✓✓Triplet RankingNoMNNSfMRGBVGGMegaDepth✓
R2D2✓✓✓✓Matching APNoMNNSfMRGBL2-NetAachen✓
DISK✓✓✓✓Matching Succ.NoMNNSfMRGBUNetMegaDepth✓
URR-No----MNN3D RenderingRGBResNetScanNet✓
LoFTR-No---TransformerMNN,OTSfMGreyFPNScanNet,MegaDepth✓
SiLK✓OptionalNoNoMatching Succ.NoMNNRand. Homo.GreyGenericAny Image Set✓
", + "image_path": "46fd8418c095cb447fab4c346a278b0cd3e97c17a2250fa0497f6eb9fe5e564d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 186, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 47, + 186, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 545, + 208 + ], + "type": "text", + "content": "Table 1. Non-exhaustive deconstruction of keypoint detectors along different dimensions. On each, SiLK adopts the simplest choice or is flexible to different choices. In particular, SiLK does not depend strongly on backbone type or training data (Tab. 7 & Tab. 9)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 329, + 225, + 524, + 382 + ], + "blocks": [ + { + "bbox": [ + 47, + 503, + 278, + 515 + ], + "lines": [ + { + "bbox": [ + 47, + 503, + 278, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 278, + 515 + ], + "type": "text", + "content": "Figure 2. Pseudo-code: learning keypoints from a single image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 225, + 524, + 382 + ], + "lines": [ + { + "bbox": [ + 329, + 225, + 524, + 382 + ], + "spans": [ + { + "bbox": [ + 329, + 225, + 524, + 382 + ], + "type": "image", + "image_path": "a16e87181d3c935e6110d68824ccea98d1a2a0d78bee4cf8a0b6ff98b04e153e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 385, + 485, + 397 + ], + "lines": [ + { + "bbox": [ + 367, + 385, + 485, + 397 + ], + "spans": [ + { + "bbox": [ + 367, + 385, + 485, + 397 + ], + "type": "text", + "content": "Figure 3. Architecture for SiLK" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 418, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 490 + ], + "type": "text", + "content": "A common approach [16, 50, 39, 17] for obtaining keypoint probabilities is to use a softmax cell-based approach. A cell is a " + }, + { + "bbox": [ + 304, + 418, + 545, + 490 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 304, + 418, + 545, + 490 + ], + "type": "text", + "content": " patch in which the probability of each cell position is determined by a local softmax. The softmax operates on " + }, + { + "bbox": [ + 304, + 418, + 545, + 490 + ], + "type": "inline_equation", + "content": "N \\times N + 1" + }, + { + "bbox": [ + 304, + 418, + 545, + 490 + ], + "type": "text", + "content": " bins; " + }, + { + "bbox": [ + 304, + 418, + 545, + 490 + ], + "type": "inline_equation", + "content": "+1" + }, + { + "bbox": [ + 304, + 418, + 545, + 490 + ], + "type": "text", + "content": " being the dustbin, handling the case of cells devoid of keypoints." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 491, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 573 + ], + "type": "text", + "content": "SiLK's approach is equivalent to a cell-based approach with a cell size of " + }, + { + "bbox": [ + 304, + 491, + 545, + 573 + ], + "type": "inline_equation", + "content": "N = 1" + }, + { + "bbox": [ + 304, + 491, + 545, + 573 + ], + "type": "text", + "content": ". This has several consequences. First, the softmax formulation becomes a sigmoid " + }, + { + "bbox": [ + 304, + 491, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\sigma(x) = \\frac{1}{1 + e^{-x}}" + }, + { + "bbox": [ + 304, + 491, + 545, + 573 + ], + "type": "text", + "content": ". Second, it removes the sparsity constraint that keypoints are exclusive events inside a cell. And third, this removes a free parameter (the cell size " + }, + { + "bbox": [ + 304, + 491, + 545, + 573 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 491, + 545, + 573 + ], + "type": "text", + "content": ") that we do not have to later tune." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 574, + 545, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 574, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 304, + 574, + 545, + 623 + ], + "type": "text", + "content": "In the same spirit, SiLK does not use NMS during inference. Even though NMS is an established pruning technique that aims to spread out keypoints [16], we find SiLK doesn't need NMS to perform (cf. Tab. 2)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 632, + 518, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 632, + 518, + 645 + ], + "spans": [ + { + "bbox": [ + 305, + 632, + 518, + 645 + ], + "type": "text", + "content": "3.3. Descriptors Define Matching Probability" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 651, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 651, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 651, + 545, + 687 + ], + "type": "text", + "content": "Similar to [47, 50], we model the cycle matching probability using a double softmax (i.e. the probability of matching " + }, + { + "bbox": [ + 304, + 651, + 545, + 687 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 651, + 545, + 687 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 651, + 545, + 687 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 651, + 545, + 687 + ], + "type": "text", + "content": ", and back)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 381, + 701, + 470, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 701, + 470, + 714 + ], + "spans": [ + { + "bbox": [ + 381, + 701, + 470, + 714 + ], + "type": "interline_equation", + "content": "P _ {i \\leftrightarrow j} = P _ {i \\rightarrow j} P _ {i \\leftarrow j}", + "image_path": "0367a5d9af6cada909d5e1fbeb3ed5f763c60b8c670534f7a9ab7504d0b7de7b.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22501" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "P_{i\\rightarrow j}" + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "text", + "content": " is the directional probability of matching the " + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "text", + "content": "th descriptor from image " + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "text", + "content": " to the " + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "text", + "content": "th descriptor in image " + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "P_{i\\leftarrow j}" + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "text", + "content": " is similar, but in the reverse direction. Both forward and backward probabilities are modeled as a softmax with temperature over descriptor cosine similarities. For fixed " + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "P_{i\\rightarrow j}" + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "text", + "content": " is a softmax over the " + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "text", + "content": "-th row of the descriptor similarity matrix, and " + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "P_{i\\leftarrow j}" + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "text", + "content": " takes softmax over column " + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 72, + 289, + 157 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 163, + 111, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 111, + 175 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 111, + 175 + ], + "type": "text", + "content": "3.4. Training" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 181, + 287, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 181, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 287, + 252 + ], + "type": "text", + "content": "Self-Supervision. Pixel-accurate correspondences are required during training. Similar to SuperPoint [16], we obtain image-pair correspondences by applying a random transformation (homography) to an image. However, the homography is a linear mapping that gives correspondences at a subpixel level." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 253, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 287, + 361 + ], + "type": "text", + "content": "To obtain pixel correspondences, we apply the sampled homography to all the pixel positions of image " + }, + { + "bbox": [ + 46, + 253, + 287, + 361 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 253, + 287, + 361 + ], + "type": "text", + "content": "; positions being the center of pixels (e.g. the top-left pixel has position " + }, + { + "bbox": [ + 46, + 253, + 287, + 361 + ], + "type": "inline_equation", + "content": "(0.5, 0.5)" + }, + { + "bbox": [ + 46, + 253, + 287, + 361 + ], + "type": "text", + "content": "). This first step establishes dense directional correspondences from " + }, + { + "bbox": [ + 46, + 253, + 287, + 361 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 253, + 287, + 361 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 253, + 287, + 361 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 46, + 253, + 287, + 361 + ], + "type": "text", + "content": ". We then run the same process from " + }, + { + "bbox": [ + 46, + 253, + 287, + 361 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 46, + 253, + 287, + 361 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 253, + 287, + 361 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 253, + 287, + 361 + ], + "type": "text", + "content": " (using the inverse homography this time). Once both directional correspondences are obtained, the resulting positions are discretized; out-of-bound and non-bijective correspondences are discarded." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 361, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 420 + ], + "type": "text", + "content": "Image Augmentations. As done in [16, 42], we employ image augmentation to improve robustness; augmentations include random brightness, contrast, gaussian noise, speckle noise and motion blur. We refer to supplementary materials for details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 420, + 287, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 287, + 516 + ], + "type": "text", + "content": "Negative and Positive Selection. One defining property of a keypoint is distinctiveness – the point can be reliably distinguished from its peers. In our case this means the point can be reliably identified in a matching algorithm, similar to [49, 50]. With that in mind, we adopt a similar supervision technique as [49]. Keypoints that are correctly matched (using the currently trained descriptors) are labelled as positive; otherwise negative." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 516, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 516, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 287, + 577 + ], + "type": "text", + "content": "Descriptor and Keypoint Losses. Similar to [47, 50], the descriptor loss is the negative log-likelihood loss applied to the matching probabilities for the positive round-trips (i.e. paths from point " + }, + { + "bbox": [ + 46, + 516, + 287, + 577 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 516, + 287, + 577 + ], + "type": "text", + "content": " to its location " + }, + { + "bbox": [ + 46, + 516, + 287, + 577 + ], + "type": "inline_equation", + "content": "i'" + }, + { + "bbox": [ + 46, + 516, + 287, + 577 + ], + "type": "text", + "content": " in the transformed image and back again)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 93, + 585, + 241, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 585, + 241, + 597 + ], + "spans": [ + { + "bbox": [ + 93, + 585, + 241, + 597 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d e s c} = - \\left(\\log P _ {i \\rightarrow i ^ {\\prime}} + \\log P _ {i \\leftarrow i ^ {\\prime}}\\right)", + "image_path": "fa1072dcba32ec8903ecd9c26f04650043e794875d30947997e658272079e89d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 689 + ], + "type": "text", + "content": "This implicitly penalizes non-positive paths via softmax. One might notice " + }, + { + "bbox": [ + 46, + 605, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{desc}" + }, + { + "bbox": [ + 46, + 605, + 287, + 689 + ], + "type": "text", + "content": " requires the computation of a large matrix (size " + }, + { + "bbox": [ + 46, + 605, + 287, + 689 + ], + "type": "inline_equation", + "content": "HW\\times HW" + }, + { + "bbox": [ + 46, + 605, + 287, + 689 + ], + "type": "text", + "content": "). To handle potential GPU out-of-memory, we provide a simple, yet efficient implementation which computes the similarity matrix in a block-wise fashion, and recomputes block dot products instead of storing them for backpropagation (in the same vein as [36, 25])." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "The keypoint loss " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{key}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " is a simple binary cross-entropy loss applied to a logistic sigmoid, in contrast to [49]. It" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "is trained to identify keypoints with successful round-trip matches (defined by mutual-nearest-neighbor) among all others (unsuccessful)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 118, + 388, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 118, + 388, + 131 + ], + "spans": [ + { + "bbox": [ + 305, + 118, + 388, + 131 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 138, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 138, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 138, + 545, + 258 + ], + "type": "text", + "content": "In this section, we empirically evaluate SiLK together with representative baselines and state-of-the-art methods. On HPatches we evaluate two complementary keypoint quality metrics (Repeatability, Mean Matching Accuracy [33]) and planar stereo estimation capabilities (Homography Estimation). In addition, we benchmark on three real-world stereo tasks: outdoor camera pose estimation on Image Matching Challenge (IMC) 2022, and both indoor camera pose estimation and 3D point-cloud registration on ScanNet. In these experiments, we study the following:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 258, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 376 + ], + "type": "text", + "content": "(1) Many methods employ complex strategies to learn and predict good keypoints and descriptors, including elements like multi-stage training, cell-based priors, complicated post-processing, context aggregation, and groundtruth 3D pose supervision (see Tab. 1), in various combinations. What machinery is necessary? SiLK contains no such machinery, and can be viewed as a reduction from these methods. However, SiLK either achieves new SOTA or compares very favorably. This questions the need for complex schemes for the evaluated tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 377, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 509 + ], + "type": "text", + "content": "(2) We observed rather strong performance from engineered features (e.g. SIFT in Tab. 2&Tab. 6) vs learned methods. This motivates us to revisit design choices in a learned keypoint detector: What makes a good keypoint detector? We ablate each component (data, backbone, etc.) and test generalization performance under various conditions (e.g. input size, test data, task, etc). SiLK proves very robust to these choices (Sec. 4.6). In particular, a very lightweight version of SiLK (two 3x3 convolution layers) is competitive to SOTA on homography estimation, camera pose estimation and point cloud registration (Tab. 7&Tab. 8)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 509, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 581 + ], + "type": "text", + "content": "We hope these results can serve the community and help adapt keypoint models to their tasks and needs. For example, labeling tasks (e.g. self-training [37], object pose [38]) might focus on high accuracy (i.e. larger backbone and denser keypoint selection), while tasks requiring speed (e.g. SLAM [34]) might find our lightweight backbone attractive." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 587, + 440, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 587, + 440, + 600 + ], + "spans": [ + { + "bbox": [ + 305, + 587, + 440, + 600 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "type": "text", + "content": "Our own training pipeline has been used for all experiments with SiLK, as well as our reproduced SuperPoint (Tab. 2) results. Training time is ~5 hours on 2 Tesla V100-SXM2 GPUs using our default setup." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Default Setup. Unless specified otherwise, all results use the following setup. Trained on COCO [29] images (randomly sampled), with Adam [24] optimizer with learning rate " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " and betas " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "(0.9, 0.999)" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "; trained for " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "100\\mathrm{k}" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " iterations; batch size of 1 per GPU; dense descriptor map" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22502" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "is " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "146 \\times 146" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " for all architectures and input resolutions; cosine similarities scaled by temperature " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "20^{-1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "; VGGnp-4 backbone (VGG architecture with max-pooling removed, details in Sec. 4.6); sparse keypoints obtained with top-k (" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "k = 10000" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "); detection head is 13x3 convolution (128-dims), and 11x1 convolution; descriptor head is 13x3 convolution (128-dims), and 11x1 convolution (128-dims out); no padding in convolutions; ReLU and batchnorm used as non-linearity and normalization (see supplementary)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 190, + 234, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 190, + 234, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 234, + 204 + ], + "type": "text", + "content": "4.2. HPatches Homography Estimation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 209, + 287, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 287, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 287, + 292 + ], + "type": "text", + "content": "Following [16, 42, 39, 47], we evaluate homography estimation on HPatches [3]. HPatches contains 57 scenes (of 6 images) with significant illumination changes and 59 scenes with large viewpoint variations. Images in each scene are related by groundtruth homographies. We follow LoFTR [47], (currently SOTA on HPatches), and scale the shorter image edge to 480 at inference time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 293, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 449 + ], + "type": "text", + "content": "Evaluation Protocol. For every image pairs, the model detects a set of keypoints. These are desired to be distinctive and thus repeatably detected across views. We use Repeatability to evaluate detection performance as in [16]. To test invariance of keypoint descriptors, each model's preferred matching algorithm establishes correspondence across images to obtain a subset of keypoints. We distinguish this subset as post-matching and the entire set as pre-matching. The accuracy of each correspondence is evaluated by Mean Matching Accuracy as in [17, 39, 50]. Finally, we use OpenCV RANSAC algorithm to estimate homography from matched keypoints, and evaluate Homography Estimation Accuracy [16] and Homography Estimation AUC [42, 47]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 449, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 604 + ], + "type": "text", + "content": "Baselines. We compare SiLK against both sparse detector-based methods and dense detector-free methods. The sparse detector-based methods generally follow \"detect then match\": the model first detects a sparse set of distinctive points, then matches features. We include SIFT [31] as well as learned detectors SuperPoint [16] (both official release and our repro), R2D2 [39] and DISK [50]. On the other hand, dense detector-free perform \"detection by matching\": the model first extracts features, then applies a learned prematching CA module (e.g. GNN [42] or transformer [47]) to adapt the features to a specific pair of images, and then finds matches. While SiLK does not employ CA, we still include comparisons to the SOTA detector-free LoFTR [47]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "Results. Despite its simplicity, SiLK outperforms all methods on repeatability, homography accuracy and homography estimation (Tab. 2). In particular, SiLK has a strong margin when the error threshold is small (" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon = 1" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "). This validates our pixel-accurate keypoint localization. SiLK lags only on the MMA@3 metric. This is caused by the pixel-accurate contrastive loss (Fig. 4). Consequently, SiLK does not benefit from increasing error threshold in MMA. In addition, even vs LoFTR (which uses dense features and CA)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "in Tab. 3, SiLK shows strong performance on Homography Estimation AUC and competitive performance on Homography accuracy. This questions the necessity of CA for these particular tasks; this may be valuable in applications which are particularly sensitive to runtime performance." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 145, + 426, + 240 + ], + "blocks": [ + { + "bbox": [ + 315, + 145, + 426, + 240 + ], + "lines": [ + { + "bbox": [ + 315, + 145, + 426, + 240 + ], + "spans": [ + { + "bbox": [ + 315, + 145, + 426, + 240 + ], + "type": "image", + "image_path": "bf27379361492e49caf5812bdde8afd09ec6a78ad275f247f89f82258db8243d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 246, + 545, + 300 + ], + "lines": [ + { + "bbox": [ + 304, + 246, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 545, + 300 + ], + "type": "text", + "content": "Figure 4. The pixel-accurate contrastive loss results in very discriminative local features, thus reshaping the error distribution to make fewer local errors (MMA@1). This differs from interpolated, cell-based descriptors used in SuperPoint, which produces less accurate keypoints (MMA@3+)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 427, + 145, + 539, + 240 + ], + "blocks": [ + { + "bbox": [ + 427, + 145, + 539, + 240 + ], + "lines": [ + { + "bbox": [ + 427, + 145, + 539, + 240 + ], + "spans": [ + { + "bbox": [ + 427, + 145, + 539, + 240 + ], + "type": "image", + "image_path": "35603335e9a0419b19c6b8ed9978fbde35c266597b4785a95adc87b775fd255c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 320, + 491, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 320, + 491, + 334 + ], + "spans": [ + { + "bbox": [ + 306, + 320, + 491, + 334 + ], + "type": "text", + "content": "4.3. IMC 2022 outdoor pose estimation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 339, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 545, + 422 + ], + "type": "text", + "content": "The Image Matching Challenge (IMC 2022) [22] provides pairs of outdoor images from different viewpoints; participants are required to estimate the fundamental matrix. Camera pose accuracy is then computed for ten thresholds of rotation and translation error (ranging from " + }, + { + "bbox": [ + 304, + 339, + 545, + 422 + ], + "type": "inline_equation", + "content": "(1^{\\circ}, 20\\mathrm{cm})" + }, + { + "bbox": [ + 304, + 339, + 545, + 422 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 339, + 545, + 422 + ], + "type": "inline_equation", + "content": "(10^{\\circ}, 5\\mathrm{m}))" + }, + { + "bbox": [ + 304, + 339, + 545, + 422 + ], + "type": "text", + "content": ". Mean average accuracy (mAA) is reported by averaging across thresholds and scenes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 423, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 518 + ], + "type": "text", + "content": "Evaluation Protocol. Sparse methods (DISK, SiLK) detect keypoints in individual images; mutual nearest neighbor is used to select matches from each pair. Methods with CA (SuperGlue, LoFTR) directly identify matches from each image pair. In either case, the fundamental matrix is estimated using MAGSAC [4]. The challenge allows different image sizes and tuning MAGSAC parameters [22]. We use 30k keypoints and MAGSAC threshold .25." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 519, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 545, + 590 + ], + "type": "text", + "content": "Baselines. We consider the best leaderboard results from three baselines. (i) DISK is the winner of IMC 2020 and is SOTA among sparse methods. 2022 DISK results are from the IMC team (submission). We also take the best version of (ii) SuperGlue (submission) and (iii) LoFTR (submission) provided by the community." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 591, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 545, + 663 + ], + "type": "text", + "content": "Results. We coarsely tune SiLK for this task for 30 trials (compared to LoFTR's 200 trials). SiLK again performs competitively (c.f. Tab. 4), outperforming DISK by a significant margin " + }, + { + "bbox": [ + 304, + 591, + 545, + 663 + ], + "type": "inline_equation", + "content": "(+0.19 / +0.18\\mathrm{mAA})" + }, + { + "bbox": [ + 304, + 591, + 545, + 663 + ], + "type": "text", + "content": ". SiLK also performs favorably compared to SuperGlue, which uses context aggregation and optimal transport matching." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 670, + 506, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 506, + 682 + ], + "type": "text", + "content": "4.4. ScanNet: Indoor Pose & Point Clouds" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "ScanNet [12] is a large-scale dataset of 1513 indoor scenes of RGB-D images and ground-truth camera poses." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22503" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 78, + 70, + 515, + 167 + ], + "blocks": [ + { + "bbox": [ + 78, + 70, + 515, + 167 + ], + "lines": [ + { + "bbox": [ + 78, + 70, + 515, + 167 + ], + "spans": [ + { + "bbox": [ + 78, + 70, + 515, + 167 + ], + "type": "table", + "html": "
RepeatabilityHom. Est. Acc.Hom. Est. AUCMMA# of keypoints
ε = 1ε = 3ε = 1ε = 3ε = 1ε = 3ε = 1ε = 3pre-matchpost-match
SuperPoint (MagicLeap)0.340.610.430.80.20.510.410.72847499
SuperPoint (Ours)0.330.520.480.750.260.520.380.531143474
SIFT0.310.520.60.840.340.610.410.552189910
R2D20.360.720.450.790.20.50.340.7560881967
DISK0.380.690.450.80.220.520.520.8433491794
SiLK (top-10k)0.620.810.620.870.40.660.590.71100004283
SiLK (top-5k)0.560.760.60.850.390.640.570.6950002074
SiLK (top-1k)0.430.610.530.810.320.580.520.631000389
", + "image_path": "8fab6dd60fba888a9e5262f78c92d53474677457647400519f218342dbd10564.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 205, + 293, + 265 + ], + "blocks": [ + { + "bbox": [ + 46, + 167, + 547, + 190 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 547, + 190 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 547, + 190 + ], + "type": "text", + "content": "Table 2. SiLK achieves new SOTA on HPatches compared to other methods with sparse keypoints and features. Despite its simplicity, SiLK achieves higher performance on all metrics except MMA@3. We include the # of keypoints to ensure a fair comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 205, + 293, + 265 + ], + "lines": [ + { + "bbox": [ + 48, + 205, + 293, + 265 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 293, + 265 + ], + "type": "table", + "html": "
Hom. Est. Acc.Hom. Est. AUCMMA
ε = 1ε = 3ε = 1ε = 3ε = 1ε = 3
LoFTR (MegaDepth)0.650.870.370.650.640.91
LoFTR (ScanNet)0.240.570.070.330.360.76
SiLK (top-10k)0.620.870.40.660.590.71
SiLK (top-5k)0.600.850.390.640.570.69
", + "image_path": "3b69ceeea98bca40440c84051c76d3bedfb43ee0cd689fd3d4b75764c6c63c7a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 82, + 329, + 255, + 397 + ], + "blocks": [ + { + "bbox": [ + 46, + 265, + 287, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 265, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 287, + 319 + ], + "type": "text", + "content": "Table 3. SiLK achieves competitive performance to SOTA LoFTR on HPatches, despite not using context aggregation. The top-5k SiLK has similar number of matches compared to outdoor LoFTR. We remark that LoFTR has large generalization gap when training on different types of dataset (indoor vs. outdoor)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 329, + 255, + 397 + ], + "lines": [ + { + "bbox": [ + 82, + 329, + 255, + 397 + ], + "spans": [ + { + "bbox": [ + 82, + 329, + 255, + 397 + ], + "type": "table", + "html": "
private mAApublic mAA
Sparse Features
DISK0.4910.502
SuperGlue0.6760.678
SiLK0.6850.684
Dense Features
LoFTR (MegaDepth)0.7350.721
", + "image_path": "ff25d48183a95c4c61011071acd1414ffda31e93c433f1d770865aad64fdcf2b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 450, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 533 + ], + "type": "text", + "content": "Using the official train/test split we evaluate both relative camera pose estimation and point-cloud registration. Relative camera pose estimation has been used in multiple previous works [52, 53, 6, 42, 47]. The task is to estimate the essential matrix with RANSAC from point matches in a pair of images. We report pose error AUC at thresholds " + }, + { + "bbox": [ + 46, + 450, + 287, + 533 + ], + "type": "inline_equation", + "content": "(5^{\\circ},10^{\\circ},20^{\\circ})" + }, + { + "bbox": [ + 46, + 450, + 287, + 533 + ], + "type": "text", + "content": " using 20k keypoints and inlier threshold .5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": "This protocol measures translation error in degrees and is known to suffer from scale ambiguity [52]. Angular translation error may be unstable, in particular if the underlying translation (in meters) is small. In response, recent works [19, 18] have introduced a 3D point-cloud registration task, using the ground-truth depth provided in ScanNet. A pair of images 20 frames apart is first sampled. Given this pair, a model predicts point matches. After matching, relative camera pose is estimated. Different from the previous protocol, ground-truth depth and camera intrinsics are now used to align matches in 3D. In addition to relative pose errors (reported separately for translation (cm) and rotation (degree)), the Chamfer distance (in cm) is measured between the registered point cloud and the groundtruth point cloud. We refer to the original papers [19, 18] for details on" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 336, + 205, + 517, + 321 + ], + "blocks": [ + { + "bbox": [ + 46, + 397, + 287, + 430 + ], + "lines": [ + { + "bbox": [ + 46, + 397, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 287, + 430 + ], + "type": "text", + "content": "Table 4. SiLK achieves new SOTA for sparse methods on IMC2022 and performs competitively to dense methods with context aggregation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 336, + 205, + 517, + 321 + ], + "lines": [ + { + "bbox": [ + 336, + 205, + 517, + 321 + ], + "spans": [ + { + "bbox": [ + 336, + 205, + 517, + 321 + ], + "type": "table", + "html": "
Pose Estimation AUC@5°@10°@20°
Sparse Features
D2-Net [17] + MNN5.314.528.0
SuperPoint [16] + MNN9.421.536.4
SuperPoint + PointCN [52]11.425.541.4
SuperPoint + OANet [53]11.826.943.9
SuperPoint + SuperGlue [42]16.233.851.8
SiLK + MNN18.034.450.4
Dense Features
DRC-Net [27]7.717.930.5
LoFTR (MegaDepth)16.933.650.6
LoFTR (ScanNet)21.540.857.6
", + "image_path": "3042c0c91468d03009ee439be58d1b6fbed2d4e993e3da199efee23e794ec834.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 321, + 545, + 354 + ], + "lines": [ + { + "bbox": [ + 305, + 321, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 305, + 321, + 545, + 354 + ], + "type": "text", + "content": "Table 5. SiLK advances SOTA on relative pose estimation among sparse methods on ScanNet and performs competitively against dense method LoFTR." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 373, + 440, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 440, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 440, + 384 + ], + "type": "text", + "content": "the evaluation setup and metrics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 399, + 443, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 443, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 443, + 411 + ], + "type": "text", + "content": "4.4.1 Relative pose estimation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 417, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 545, + 559 + ], + "type": "text", + "content": "Baselines. We compare SiLK with both sparse detector-based methods and dense detector-free methods. For the sparse methods, we consider local feature descriptors (R2D2, SuperPoint) with mutual nearest neighbor (MNN) for matching. SiLK falls into this category. In addition, we consider multiple learned context aggregation methods for matching that operates on SuperPoint, including PointCN [52], OANet [53] and SuperGlue. For detector-free dense methods, we consider DRC-Net [27] and LoFTR. We include two versions of LoFTR: one trained on MegaDepth with optimal transport post-processing and one trained on ScanNet." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 560, + 545, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 560, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 560, + 545, + 656 + ], + "type": "text", + "content": "Results. As summarized in Tab. 5, SiLK significantly outperforms D2-Net (+12.7) and SuperPoint (+8.6) when using mutual nearest neighbor matching. In addition, SiLK outperforms the previous SOTA sparse method SuperGlue, despite its simpler design without context aggregation. SiLK performs similarly to LoFTR trained with MegaDepth. SiLK is only outperformed by LoFTR trained on ScanNet, the same as evaluation data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 670, + 494, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 494, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 494, + 682 + ], + "type": "text", + "content": "4.4.2 Pairwise 3D point-cloud registration" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "content": "Baselines. We consider three main types of baselines. (i) Sparse Features + RANSAC We extract sparse key-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22504" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": "points and their features from off-the-shelf models, and use RANSAC to estimate alignment. This includes SIFT and SuperPoint, and 3D geometry model FCGF [10]. (ii) Dense Feature Matching We follow the [18] and select high-quality corresponding pairs using the ratio test, and then solve a weighted Procustes problem [9, 23] to produce alignment. We add a dense version of SuperPoint by discarding the keypoint prediction. We include the current state-of-the-art URR [18] that learns invariant point descriptors through cross-view synthesis. By ignoring the keypoint scoring prediction, SiLK also belongs to this category. The goal is to evaluate the quality of the dense point features. (iii) Pose/Geometry Supervised We consider methods that use groundtruth poses to supervise, which are not required in (i) and (ii). These include LoFTR, DGR [9] and 3D MV Registration [19]. For LoFTR, we include only the MegaDepth model, since it performed better in this task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 276, + 289, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 276, + 289, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 276, + 289, + 517 + ], + "type": "text", + "content": "Results. As shown in Tab. 6, SiLK achieves new SOTA across all metrics (except rotation accuracy at " + }, + { + "bbox": [ + 47, + 276, + 289, + 517 + ], + "type": "inline_equation", + "content": "45^{\\circ}" + }, + { + "bbox": [ + 47, + 276, + 289, + 517 + ], + "type": "text", + "content": "). In particular, SiLK achieves very high accuracy at small thresholds (" + }, + { + "bbox": [ + 47, + 276, + 289, + 517 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 47, + 276, + 289, + 517 + ], + "type": "text", + "content": " angular, 5cm translation and 1cm for chamfer), validating SiLK's pixel-level precision. Comparing with DGR, 3D MV Reg and LoFTR that use groundtruth camera poses during training, SiLK significantly outperforms, indicating that groundtruth 3D supervision is not necessary to train good keypoint features. We did not include the chamfer results for LoFTR as the provided positions do not match the required resolution for correct chamfer evaluation. SiLK also achieves superior performance vs previous SOTA URR. We note that URR requires two different frames sampled from the same scene during training, and is supervised by a differentiable cross-view rendering process. In contrast, SiLK only requires a single 2D image and is trained with a simple point matching loss. Finally, we observe that SuperPoint performs competitively when evaluated in this dense fashion; this is an important difference from the results reported in URR using sparse features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 526, + 262, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 262, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 262, + 540 + ], + "type": "text", + "content": "4.5. Discussion. What makes SiLK perform?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 545, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 545, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 545, + 289, + 715 + ], + "type": "text", + "content": "Multiple factors contribute to SiLK's performance. In particular, the cycle-consistent loss and the dense predictions. First, the cycle-consistent loss (Fig. 2, supplementary) enforces the natural properties of keypoints: distinctiveness & robustness to viewpoint / photometric changes, while existing methods rely on proxy objectives or additional complexity (i.e. \"corners\" in SuperPoint[16] / SuperGlue[42], \"peakiness\" in R2D2[39], \"keypoint weighting scheme\" in D2-Net[17], \"RL cycle-consistent reward\" in DISK[50]). Removing unnecessary complexity helps the model focus its learning on the essentials, and is supported by consistent, strong results across benchmarks and metrics (Tab. 2, Tab. 3, Tab. 4, Tab. 5, Tab. 6). Second, SiLK produces dense, pixel-accurate keypoints, while all" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "existing methods have structural constraints (downsampling layers, cell-based schemes, and NMS, c.f. Tab. 1). This leads to more robust contrastive learning (i.e. more negative pairs), allows producing more keypoints matches and boosts results in the accurate regime (" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "inline_equation", + "content": "\\epsilon = 1" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": " in Tab. 2)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 141, + 507, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 141, + 507, + 155 + ], + "spans": [ + { + "bbox": [ + 305, + 141, + 507, + 155 + ], + "type": "text", + "content": "4.6. What makes good keypoint detectors?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 160, + 547, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 160, + 547, + 256 + ], + "spans": [ + { + "bbox": [ + 304, + 160, + 547, + 256 + ], + "type": "text", + "content": "Leveraging SiLK's flexibility (Sec. 3), we comprehensively ablate a large pool of design choices such as model architecture and image resolution. Surprisingly, we found that reducing architecture size, compute cost, and training input size only mildly impact model performance on homography estimation, camera pose estimation and point cloud registration. This benefits many important applications, such as on-device inference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 257, + 547, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 547, + 341 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 547, + 341 + ], + "type": "text", + "content": "Here we discuss the key findings. We use the metrics Repeatability@1 (R), Homography Estimation Accuracy@1 (HA), Mean Matching Accuracy@1 (MMA) for HPatches, and Rotation Accuracy@5° (RA), Translation Accuracy@5cm (TA) and Chamfer@1 (C) for ScanNet, all at lowest error thresholds. Additional results and analysis are included in supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 342, + 547, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 425 + ], + "type": "text", + "content": "Agnostic to backbone. Existing methods use various backbones (Tab. 1); the effects on keypoint models are not well understood. We consider FPN from LoFTR and UNet from DISK; both are modern compared to SiLK's VGGnp [45] backbone. We find no empirical performance gain despite far greater parameter counts (Tab. 7). This questions the need for high-capacity models for these keypoint problems." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 426, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 640 + ], + "type": "text", + "content": "Next we reduce the complexity of the original Super-Point backbone VGG-4. Max-pooling and up-sampling layers are removed. Our VGGnp-4 contains four convolution blocks, each with two convolution layers followed by ReLU. We discard convolution blocks from VGGnp-4 to obtain VGGnp-3, VGGnp-2 and VGGnp-1. On top of VGGnp-1, we reduce channel and descriptor sizes to 64 and 32 respectively and obtain an ultra-lightweight model, VGGnp-" + }, + { + "bbox": [ + 304, + 426, + 547, + 640 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 426, + 547, + 640 + ], + "type": "text", + "content": ". Results on all metrics (except MMA) only drop mildly as we shrink the model. In particular, SiLK (VGGnp-" + }, + { + "bbox": [ + 304, + 426, + 547, + 640 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 426, + 547, + 640 + ], + "type": "text", + "content": " in Tab. 7) achieves very competitive performance (Tab. 2&Tab. 6). On the other hand, matching scores (MMA) drop significantly. We suggest two possible reasons: first, pointwise matching may benefit from the larger receptive field of deeper models. Second, homography estimation aggregates numerous pointwise measurements; homographies will improve if the noises cancel out, or if the outliers are removed by RANSAC." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "Fast training on tiny images. By default, SiLK uses 146x146 descriptor map resolution during training. Higher resolution provides more points, which benefits the contrastive loss with more negatives, but also increases training time and GPU memory usage. Surprisingly, performance changes very little when varying resolution during" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22505" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 511, + 215 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 511, + 215 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 511, + 215 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 511, + 215 + ], + "type": "table", + "html": "
RotationTranslationChamfer
Accuracy ↑Error ↓Accuracy ↑Error ↓Accuracy ↑Error ↓
5°10°45°MeanMed.51025MeanMed.1510MeanMed.
Sparse Features + RANSAC
SIFT [31]55.275.789.218.64.317.744.579.826.511.238.170.678.342.61.7
SuperPoint [16]65.586.996.68.93.621.251.788.016.19.745.781.188.219.21.2
FCGF [10]70.287.796.29.53.327.558.382.923.68.352.078.083.724.40.9
Pose/Geometry Supervised
DGR [9]81.189.394.89.41.854.576.288.718.44.570.585.589.013.70.4
3D MV Reg [19]87.793.297.06.01.269.083.191.811.72.978.989.291.810.20.2
LoFTR(MegaDepth) [47]91.796.899.42.81.265.985.297.16.03.3-----
Dense feature matching
SuperPoint [16]93.098.499.82.51.656.884.798.26.54.377.396.198.44.50.3
URR [18]92.795.898.53.40.877.289.696.17.32.386.094.696.15.90.1
SiLK (VGGnp-4)98.199.099.61.70.882.994.899.04.12.192.898.399.14.30.1
", + "image_path": "6c3b4a2d4c56023d848163f27280dcc7d774309191f62333e7dd9bdff3c0773d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 243, + 286, + 330 + ], + "blocks": [ + { + "bbox": [ + 91, + 215, + 501, + 225 + ], + "lines": [ + { + "bbox": [ + 91, + 215, + 501, + 225 + ], + "spans": [ + { + "bbox": [ + 91, + 215, + 501, + 225 + ], + "type": "text", + "content": "Table 6. SiLK achieves state-of-the-art on camera pose estimation and point cloud registration on ScanNet." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 243, + 286, + 330 + ], + "lines": [ + { + "bbox": [ + 49, + 243, + 286, + 330 + ], + "spans": [ + { + "bbox": [ + 49, + 243, + 286, + 330 + ], + "type": "table", + "html": "
HPatchesScanNetModel
RHAcMMARATACParamFPSGFLOP
VGGnp-40.620.620.5998.182.992.8942k12.2370
VGGnp-30.630.610.5898.084.293.5868k12.5330
VGGnp-20.630.570.5597.383.592.4757k14.3268
VGGnp-10.630.570.4494.682.190.0461k18.990
VGGnp-μ0.640.560.4093.581.189.076k36.523
FPN [47]0.580.550.52---6.6M17.5298
UNet [50]0.600.410.58---1.3M25.9198
", + "image_path": "2082f5310f50a0a57224c25026c0e04d004b5027ce55a58d0d8400ac675ff378.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 76, + 384, + 259, + 453 + ], + "blocks": [ + { + "bbox": [ + 47, + 331, + 286, + 373 + ], + "lines": [ + { + "bbox": [ + 47, + 331, + 286, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 286, + 373 + ], + "type": "text", + "content": "Table 7. SiLK is backbone agnostic. Backbones from existing methods are trained and evaluated. Low-capacity model performs well on all metrics except MMA. FPS and GFLOPs measured on " + }, + { + "bbox": [ + 47, + 331, + 286, + 373 + ], + "type": "inline_equation", + "content": "480 \\times 640" + }, + { + "bbox": [ + 47, + 331, + 286, + 373 + ], + "type": "text", + "content": " images with NVIDIA Quadro GP100 GPU." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 76, + 384, + 259, + 453 + ], + "lines": [ + { + "bbox": [ + 76, + 384, + 259, + 453 + ], + "spans": [ + { + "bbox": [ + 76, + 384, + 259, + 453 + ], + "type": "table", + "html": "
SizeTimeHPatchesScanNet
RHAcMMARATAC
8221.7h0.600.580.5698.183.592.5
11422.7h0.620.620.5998.182.992.8
14625h0.630.590.5998.283.392.9
17829.5h0.630.620.5998.183.592.9
210218h0.630.610.6098.183.492.8
", + "image_path": "3e16c288ef9ea6144d4759eec8a61354aa818bb57aec1e65113ecedb8522ec76.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 496, + 286, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 286, + 555 + ], + "type": "text", + "content": "training (Tab. 8), especially on ScanNet. Tiny feature maps (82x82) remain competitive on both HPatches and ScanNet, and trains in 1.7 hours on two GPUs. This enables applications like test-time finetuning, on-device finetuning, and rapid experimental iteration." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 286, + 700 + ], + "type": "text", + "content": "Robustness to training data. Existing methods use various training sets (Tab. 1); empirically we observe cases of poor generalization across datasets. For example, LoFTR[47], trained on indoor ScanNet data, drops significantly vs LoFTR trained on outdoor MegaDepth data (Tab. 3) and vice-versa (Tab. 5). This overfitting may be exacerbated by the high-capacity machinery used by these methods, e.g. LoFTR's Transformer contextualizer. We measure SiLK's robustness on training data choices, by using different images from COCO[29], ImageNet[15], MegaDepth[28] and ScanNet[12]. We also combine them to formulate a diversified set of training data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "SiLK is quite robust to change in training set, with the" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 337, + 243, + 515, + 320 + ], + "blocks": [ + { + "bbox": [ + 47, + 453, + 286, + 474 + ], + "lines": [ + { + "bbox": [ + 47, + 453, + 286, + 474 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 286, + 474 + ], + "type": "text", + "content": "Table 8. Decreasing training image size has minimal impact. SiLK can be trained under 3h, with little performance drop." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 243, + 515, + 320 + ], + "lines": [ + { + "bbox": [ + 337, + 243, + 515, + 320 + ], + "spans": [ + { + "bbox": [ + 337, + 243, + 515, + 320 + ], + "type": "table", + "html": "
HPatchesScanNet
RHAcMMARATAC
COCO0.620.620.5998.182.992.8
ImageNet0.630.60.5998.183.593.0
MegaDep.0.620.610.5797.983.592.9
ScanNet0.600.550.4897.682.892.7
C+I+M+S0.610.590.5497.783.092.6
C+I+M0.640.60.5998.082.992.8
", + "image_path": "0f8c7063ae775f8db9f2920530103e0850b4aa264b726c514a79f0ccd3db8083.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 320, + 545, + 341 + ], + "lines": [ + { + "bbox": [ + 306, + 320, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 306, + 320, + 545, + 341 + ], + "type": "text", + "content": "Table 9. SiLK is robust to different training sets. A noticeable drop is observed only when training on ScanNet." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 365, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 509 + ], + "type": "text", + "content": "exception of ScanNet (Tab. 9). We hypothesize this is due to the significant amount of uniform surfaces (e.g. walls, doors) present in ScanNet. These featureless areas contain few keypoints to learn from. SiLK's drop agrees directionally with LoFTR's drop observed in Tab. 3, but the magnitude is smaller. This may be because SiLK's VGGnp backbone has lower capacity than LoFTR (FPN+Transformer), and hence is less susceptible to overfitting. Finally, we remark that SiLK trained with COCO is used for comparisons in Sec. 4.2, Sec. 4.3 and Sec. 4.4 for HPatches, IMC and ScanNet, whereas LoFTR requires different training data (MegaDepth or ScanNet) to achieve strong performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 524, + 378, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 524, + 378, + 536 + ], + "spans": [ + { + "bbox": [ + 306, + 524, + 378, + 536 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "This paper presents SiLK, a simple and flexible framework for keypoint detection and descriptors. SiLK is designed from the principles of distinctiveness and invariance, and achieves or advances SOTA on key low-level tasks for 3D visual perception. SiLK's simplicity questions the need for complex machinery for good keypoint detection in low-level applications. In addition, extensive ablations reveal SiLK's robustness to backbone, training data and training input size. These findings lead to a tiny version of SiLK that is lightweight, accurate, and trains quickly. We view this \"tiny and learned\" regime as very promising for applications where runtime and/or power consumption is critical. We hope SiLK can draw attention to the field and facilitate stronger solutions." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22506" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Ebtsam Adel, Mohammed Elmogy, and Hazem Elbakry. Image stitching based on feature extraction techniques: a survey. International Journal of Computer Applications, 99(6):1-8, 2014." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 179 + ], + "type": "text", + "content": "[2] Alexandre Alahi, Raphael Ortiz, and Pierre Vandergheynst. Freak: Fast retina keypoint. In 2012 IEEE conference on computer vision and pattern recognition, pages 510-517. IEEE, 2012." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 236 + ], + "type": "text", + "content": "[3] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5173-5182, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 238, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 287, + 270 + ], + "type": "text", + "content": "[4] Daniel Barath, Jiri Matas, and Jana Noskova. MAGSAC: marginalizing sample consensus. In _Conference on Computer Vision and Pattern Recognition_, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 272, + 287, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 287, + 305 + ], + "type": "text", + "content": "[5] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer vision and image understanding, 110(3):346-359, 2008." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 307, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 287, + 350 + ], + "type": "text", + "content": "[6] Eric Brachmann and Carsten Rother. Neural-guided ransac: Learning where to sample model hypotheses. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4321-4330, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 352, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 287, + 384 + ], + "type": "text", + "content": "[7] Matthew A. Brown and David G. Lowe. Automatic panoramic image stitching using invariant features. International Journal of Computer Vision, 74:59-73, 2006." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "type": "text", + "content": "[8] Michael Calonder, Vincent Lepetit, Christoph Strecha, and Pascal Fua. Brief: Binary robust independent elementary features. In European conference on computer vision, pages 778-792. Springer, 2010." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 431, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 287, + 453 + ], + "type": "text", + "content": "[9] Christopher Choy, Wei Dong, and Vladlen Koltun. Deep global registration. In CVPR, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 454, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 287, + 497 + ], + "type": "text", + "content": "[10] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 8957-8965, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 499, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 543 + ], + "type": "text", + "content": "[11] Peter Hviid Christiansen, Mikkel Fly Kragh, Yury Brodskiy, and Henrik Karstoft. Unsupperpoint: End-to-end unsupervised interest point detector and descriptor. arXiv preprint arXiv:1907.04011, 2019." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "text", + "content": "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "text", + "content": "[13] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), volume 1, pages 886-893. IEEE, 2005." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[14] Andrew J. Davison, Ian D. Reid, Nicholas Molton, and Olivier Stasse. Monoslam: Real-time single camera slam. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29:1052-1067, 2007." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 97, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 149 + ], + "type": "text", + "content": "[16] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[17] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint detection and description of local features. arXiv preprint arXiv:1905.03561, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "type": "text", + "content": "[18] Mohamed El Banani, Luya Gao, and Justin Johnson. Unsupervisedr&r: Unsupervised point cloud registration via differentiable rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7129-7139, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 253, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 307 + ], + "type": "text", + "content": "[19] Z. Gojcic, C. Zhou, J. D. Wegner, L. J. Guibas, and T. Birdal. Learning multiview 3d point cloud registration. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1756-1766, Los Alamitos, CA, USA, jun 2020. IEEE Computer Society." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 309, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 341 + ], + "type": "text", + "content": "[20] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In Alvey vision conference, volume 15, pages 10-5244. Manchester, UK, 1988." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 343, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 343, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 343, + 545, + 386 + ], + "type": "text", + "content": "[21] Wilfried Hartmann, Michal Havlena, and Konrad Schindler. Predicting matchability. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9-16, 2014." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "type": "text", + "content": "[22] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image Matching across Wide Baselines: From Paper to Practice. International Journal of Computer Vision, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 433, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 465 + ], + "type": "text", + "content": "[23] W. Kabsch. A solution for the best rotation to relate two sets of vectors. Acta Crystallographica Section A, 32(5):922-923, Sep 1976." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 468, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 488 + ], + "type": "text", + "content": "[24] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2015." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 490, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 555 + ], + "type": "text", + "content": "[25] Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, and Daniel Haziza. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers, 2022." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 557, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 599 + ], + "type": "text", + "content": "[26] Stefan Leutenegger, Margarita Chli, and Roland Y Siegwart. Brisk: Binary robust invariant scalable keypoints. In 2011 International conference on computer vision, pages 2548-2555. Ieee, 2011." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 601, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 634 + ], + "type": "text", + "content": "[27] Xinghui Li, Kai Han, Shuda Li, and Victor Prisacariu. Dual-resolution correspondence networks. In Conference on Neural Information Processing Systems (NeurIPS), 2020." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "text", + "content": "[28] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22507" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "European conference on computer vision, pages 740-755. Springer, 2014." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "type": "text", + "content": "[30] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. Ssd: Single shot multibox detector. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision - ECCV 2016, pages 21-37, Cham, 2016. Springer International Publishing." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 195 + ], + "type": "text", + "content": "[31] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60(2):91-110, 2004." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "type": "text", + "content": "[32] Jiayi Ma, Xingyu Jiang, Aoxiang Fan, Junjun Jiang, and Junchi Yan. Image matching from handcrafted to deep features: A survey. International Journal of Computer Vision, 129(1):23-79, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 275 + ], + "type": "text", + "content": "[33] K. Mikolajczyk and C. Schmid. A performance evaluation of local descriptors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(10):1615-1630, 2005." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "type": "text", + "content": "[34] Raul Mur-Artal, J. M. M. Montiel, and Juan D. Tardós. Orb-slam: A versatile and accurate monocular slam system. IEEE Transactions on Robotics, 31:1147-1163, 2015." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 310, + 287, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 343 + ], + "type": "text", + "content": "[35] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 344, + 287, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 287, + 375 + ], + "type": "text", + "content": "[36] Markus N Rabe and Charles Staats. Self-attention does not need " + }, + { + "bbox": [ + 48, + 344, + 287, + 375 + ], + "type": "inline_equation", + "content": "o(n^{2})" + }, + { + "bbox": [ + 48, + 344, + 287, + 375 + ], + "type": "text", + "content": " memory. arXiv preprint arXiv:2112.05682, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 377, + 287, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 444 + ], + "type": "text", + "content": "[37] Umer Rafi, Andreas Doering, Bastian Leibe, and Juergen Gall. Self-supervised keypoint correspondences for multi-person pose estimation and tracking in videos. In Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XX, page 36–52, Berlin, Heidelberg, 2020. Springer-Verlag." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 445, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 510 + ], + "type": "text", + "content": "[38] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10881-10891, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "type": "text", + "content": "[39] Jerome Revaud, Philippe Weinzaepfel, César De Souza, Noe Pion, Gabriela Csurka, Yohann Cabon, and Martin Humenberger. R2d2: repeatable and reliable detector and descriptor. arXiv preprint arXiv:1906.06195, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 556, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 590 + ], + "type": "text", + "content": "[40] Edward Rosten and Tom Drummond. Machine learning for high-speed corner detection. In European conference on computer vision, pages 430-443. Springer, 2006." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[41] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision, pages 2564-2571. Ieee, 2011." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "type": "text", + "content": "[42] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[43] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 555 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "network model. IEEE transactions on neural networks, 20(1):61-80, 2008." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[44] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[45] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2015." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "text", + "content": "[46] Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 21(2):343-348, 1967." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 209, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 263 + ], + "type": "text", + "content": "[47] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[48] Jiexiong Tang, Hanme Kim, Vitor Guizilini, Sudeep Pillai, and Rares Ambrus. Neural outlier rejection for self-supervised keypoint learning. arXiv preprint arXiv:1912.10615, 2019." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 309, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 364 + ], + "type": "text", + "content": "[49] Prune Truong, Stefanos Apostolopoulos, Agata Mosinska, Samuel Stucky, Carlos Ciller, and Sandro De Zanet. Glampoints: Greedily learned accurate match points. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10732-10741, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 365, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 407 + ], + "type": "text", + "content": "[50] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 409, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 409, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 409, + 545, + 453 + ], + "type": "text", + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 455, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 498 + ], + "type": "text", + "content": "[52] Kwang Moo Yi, Eduard Trulls, Yuki Ono, Vincent Lepetit, Mathieu Salzmann, and Pascal V. Fua. Learning to find good correspondences. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2666-2674, 2017." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 499, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 499, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 499, + 545, + 555 + ], + "type": "text", + "content": "[53] Jiahui Zhang, Dawei Sun, Zixin Luo, Anbang Yao, Lei Zhou, Tianwei Shen, Yurong Chen, Long Quan, and Honggen Liao. Learning two-view correspondences and geometry using order-aware network. International Conference on Computer Vision (ICCV), 2019." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22508" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Sigmoid Loss for Language Image Pre-Training/85b123ac-a2cf-49ce-9282-93b24961a80b_content_list.json b/2023/Sigmoid Loss for Language Image Pre-Training/85b123ac-a2cf-49ce-9282-93b24961a80b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0c16e215b83806d947c2c734222a99c50eda4eff --- /dev/null +++ b/2023/Sigmoid Loss for Language Image Pre-Training/85b123ac-a2cf-49ce-9282-93b24961a80b_content_list.json @@ -0,0 +1,1597 @@ +[ + { + "type": "text", + "text": "Sigmoid Loss for Language Image Pre-Training", + "text_level": 1, + "bbox": [ + 243, + 130, + 727, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaohua Zhai\\* Basil Mustafa Alexander Kolesnikov Lucas Beyer\\* Google DeepMind, Zurich, Switzerland", + "bbox": [ + 205, + 180, + 759, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xzhai, basilm, akolesnikov, lbeyer}@google.com", + "bbox": [ + 276, + 219, + 692, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 282, + 313, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a simple pairwise sigmoid loss for image-text pre-training. Unlike standard contrastive learning with softmax normalization, the sigmoid loss operates solely on image-text pairs and does not require a global view of the pairwise similarities for normalization. The sigmoid loss simultaneously allows further scaling up the batch size, while also performing better at smaller batch sizes. With only four TPUv4 chips, we can train a Base CLIP model at $4k$ batch size and a Large LiT model at $20k$ batch size, the latter achieves $84.5\\%$ ImageNet zero-shot accuracy in two days. This disentanglement of the batch size from the loss further allows us to study the impact of examples vs pairs and negative to positive ratio. Finally, we push the batch size to the extreme, up to one million, and find that the benefits of growing batch size quickly diminish, with a more reasonable batch size of $32k$ being sufficient. We hope our research motivates further explorations in improving the quality and efficiency of language-image pre-training.", + "bbox": [ + 75, + 316, + 472, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 619, + 209, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contrastive pre-training using weak supervision from image-text pairs found on the web is becoming the go-to method for obtaining generic computer vision backbones, slowly replacing pre-training on large labelled multi-class datasets. The high-level idea is to simultaneously learn an aligned representation space for images and texts using paired data. Seminal works CLIP [34] and ALIGN [23] established the viability of this approach at a large scale, and following their success, many large image-text datasets became available privately [55, 13, 21, 45] and publicly [37, 6, 15, 7, 38].", + "bbox": [ + 75, + 646, + 468, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The standard recipe to pre-train such models leverages the image-text contrastive objective. It aligns the image and text embeddings for matching (positive) image-text pairs while making sure that unrelated (negative) image-text pairs", + "bbox": [ + 75, + 813, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Table 1: SigLiT and SigLIP results. Sigmoid loss is memory efficient, allows larger batch sizes (BS) that unlocks language image pre-training with a small number of chips. SigLiT model with a frozen public L/16 checkpoint [39], trained on the LiT image-text dataset [55] using four TPU-v4 chips for one day, achieves $79.7\\%$ 0-shot accuracy on ImageNet. The same setup with a g/14 checkpoint [54] leads to $84.5\\%$ accuracy, trained for two days. With a public unlocked B/16 image checkpoint [39], trained on the WebLI dataset [13], SigLIP achieves $71.0\\%$ 0-shot accuracy using 16 TPU-v4 chips for three days. The last two rows show results with randomly initialized models.", + "bbox": [ + 496, + 267, + 893, + 450 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/c191f62bf6543067b0a5bb14e69cbea68eac3d58063fd0c8bfc53f5da8ff81c3.jpg", + "table_caption": [], + "table_footnote": [ + "* We use a variant of the $L$ model with 12 layers." + ], + "table_body": "
ImageTextBS#TPUv4DaysINet-0
SigLiTB/8L*32 k4179.7
SigLiTg/14L20 k4284.5
SigLIPB/16B16 k16371.0
SigLIPB/16B32 k32272.1
SigLIPB/16B32 k32573.4
", + "bbox": [ + 500, + 462, + 890, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are dissimilar in the embedding space. This is achieved via a batch-level softmax-based contrastive loss, applied twice to normalize the pairwise similarity scores across all images, then all texts. A naive implementation of the softmax is numerically unstable; it is usually stabilized by subtracting the maximum input value before applying the softmax [18], which requires another pass over the full batch.", + "bbox": [ + 496, + 638, + 893, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a simpler alternative: the sigmoid loss. It does not require any operation across the full batch and hence greatly simplifies the distributed loss implementation and boosts efficiency. Additionally, it conceptually decouples the batch size from the definition of the task. We compare the proposed sigmoid loss with the standard softmax loss across multiple setups. In particular, we investigate sigmoid-based loss with two prominent approaches for image-text learning: CLIP [34] and LiT [55], which we call sigmoid language image pre", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*equal contribution", + "bbox": [ + 101, + 887, + 207, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11975", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "training (SigLIP) and sigmoid LiT (SigLiT), respectively. We find that the sigmoid loss performs significantly better than the softmax loss when the batch size is smaller than $16\\mathrm{k}$ . As the train batch size grows, the gap closes. Importantly, the sigmoid loss is symmetric, requires just a single pass, and a typical implementation requires less memory than the softmax loss. This enables successful training of a SigLiT model at a batch size of one million. However, we find that the performance saturates with growing batch size, both for softmax and sigmoid. The good news is that a reasonable batch size, i.e. $32\\mathrm{k}$ , is sufficient for image-text pretraining. This conclusion also holds for multilingual SigLIP training on over 100 languages.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In Table 1, we present setups for image-text pre-training that require a moderate amount of TPUv4 chips for training. SigLiT is surprisingly efficient, reaching $79.7\\%$ zero-shot accuracy on ImageNet in just a single day on four chips. SigLIP's more demanding from-scratch training reaches $73.4\\%$ zero-shot accuracy in 5 days with 32 TPUv4 chips. This compares favorably to prior works such as FLIP [29] and CLIP [34], which require approximately 5 and 10 days respectively on 256 TPUv3 cores. When fine-tuning a pretrained vision backbone in SigLIP, denoted as $\\boxed{ \\begin{array}{r l} \\end{array} }$ in Table 1, we found that disabling the weight decay on the pre-trained backbone leads to better results (see Figure 4 for details). We hope our work paves the way for making the nascent language-image pre-training field more accessible.", + "bbox": [ + 75, + 287, + 473, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 512, + 218, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive learning with the sigmoid loss. One prior work proposes a similar sigmoid loss for the task of unsupervised dimensionality reduction [19]; in the scope of contrastive image-text learning, the vast majority of works rely on the softmax-based InfoNCE loss as popularized by [43]. In supervised classification, the sigmoid loss has already been shown to be slightly more effective and robust than the softmax loss [3, 47].", + "bbox": [ + 75, + 542, + 468, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive language-image pre-training has become popular since CLIP [34] and ALIGN [23] applied softmax contrastive learning [56, 43, 10, 24] to large-scale image-text datasets. Both models perform very well on zero-shot transfer tasks, including classification and retrieval. Follow-up works show that contrastively pre-trained models produce good representations for fine-tuning [49, 16], linear regression [23], object detection [30], semantic segmentation [31] and video tasks [53].", + "bbox": [ + 75, + 667, + 468, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative language-image pre-training Besides softmax contrastive pre-training, various alternatives have been proposed. GIT [45], SimVLM [46], and LEMON [21] successfully pre-train models using a generative text decoder instead, while CoCa [52] adds such a decoder to the discriminative CLIP/ALIGN setup, thus combining the pros", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 Sigmoid loss pseudo-implementation." + ], + "code_body": "1 # img_emb : image model embedding [n, dim] \n2 # txt_emb : text model embedding [n, dim] \n3 # t_prime, b : learnable temperature and bias \n4 # n : mini-batch size \n5 \n6 t = exp(t_prime) \n7 zimg = l2_normalize(img_emb) \n8 ztxt = l2_normalize(txt_emb) \n9 logits = dot(zimg, ztxt.T) * t + b \n10 labels = 2 * eye(n) - ones(n) # -1 with diagonal 1 \n11 l = -sum(log_sigmoid(labels * logits)) / n", + "guess_lang": "python", + "bbox": [ + 485, + 113, + 869, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and cons of both approaches into a single very capable model. BLIP [28] further proposes CapFilt which uses the generative decoder to create better captions and the discriminative part of the model to filter pairs. Language-Image pre-training is a very active field and surveys [8] rapidly become outdated.", + "bbox": [ + 496, + 263, + 890, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Efficient language-image pre-training On the other hand, few works have tried making language image pre-training more efficient. LiT [55] and FLIP [29] are notable attempts, the former requires a pre-trained and locked backbone, and the latter sacrifices quality by randomly dropping visual tokens. BASIC [33] and LAION [48] look at scaling batch-size but only go up to $16\\mathrm{k}$ and $160\\mathrm{k}$ respectively, by using many hundreds of chips, and for the former also mixing in a large private classification dataset [33, 51]. The recent Lion optimizer [12] claims to be able to reduce the training cost to reach similar quality.", + "bbox": [ + 496, + 361, + 892, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 544, + 589, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we first review the widely-used softmax-based contrastive loss. We then introduce the pairwise sigmoid loss and discuss its efficient implementation.", + "bbox": [ + 496, + 571, + 890, + 614 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a mini-batch $\\mathcal{B} = \\{(I_1, T_1), (I_2, T_2), \\ldots\\}$ of image-text pairs, the contrastive learning objective encourages embeddings of matching pairs $(I_i, T_i)$ to align with each other, while pushing embeddings of unmatched pairs $(I_i, T_{j \\neq i})$ apart. For practical purposes, it is assumed that for all images $i$ , the text associated with a different image $j$ is not related to $i$ , and vice-versa. This assumption is usually noisy and imperfect.", + "bbox": [ + 496, + 616, + 892, + 738 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Softmax loss for language image pre-training", + "text_level": 1, + "bbox": [ + 498, + 750, + 880, + 767 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "When using the softmax loss to formalize this objective, an image model $f(\\cdot)$ and a text model $g(\\cdot)$ are trained to minimize the following objective:", + "bbox": [ + 496, + 773, + 890, + 821 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n- \\frac {1}{2 | \\mathcal {B} |} \\sum_ {i = 1} ^ {| \\mathcal {B} |} \\left(\\overbrace {\\log \\frac {e ^ {t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {i}}}{\\sum_ {j = 1} ^ {| \\mathcal {B} |} e ^ {t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {j}}}} ^ {\\text {i m a g e} \\rightarrow \\text {t e x t s o f t m a x}} + \\overbrace {\\log \\frac {e ^ {t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {i}}}{\\sum_ {j = 1} ^ {| \\mathcal {B} |} e ^ {t \\mathbf {x} _ {j} \\cdot \\mathbf {y} _ {i}}}} ^ {\\text {t e x t} \\rightarrow \\text {i m a g e s o f t m a x}}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 832, + 870, + 905 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11976", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/40b902962f5bea4a83d847167bbf30f869720dcad723b3cfcf9d1f670646a50c.jpg", + "image_caption": [ + "Figure 1: Efficient loss implementation demonstrated via a mock setup with 3 devices and a global batch size of 12. There are no all-gathers, and at any point in time only the bright yellow square (size $4 \\times 4$ ) is materialized in memory." + ], + "image_footnote": [ + "(a) Initially each device holds 4 (b) They each compute the comimage and 4 text representations. component of the loss (highlighted) Each device needs to see the repre for their representations, which representations from other devices includes the positives. to calculate the full loss." + ], + "bbox": [ + 76, + 85, + 277, + 233 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f9c22c520c1dfb106788a3f98f65c3814546d131b705c71c53e667fdf7d80973.jpg", + "image_caption": [], + "image_footnote": [ + "(c) Texts are swapped across the devices, so device 1 now has $I_{1:4}$ and $T_{5:8}$ etc. The new loss is computed and accumulated with the previous." + ], + "bbox": [ + 282, + 85, + 480, + 270 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fbc90ed975367f418d02645cc1e6163dc3f70c8308d1fad6aede1cbd706cbedd.jpg", + "image_caption": [], + "image_footnote": [ + "(d) This repeats till every image & text pair have interacted, e.g. device 1 has the loss of $I_{1:4}$ and $T_{1:12}$ . A final cross-device sum brings everything together." + ], + "bbox": [ + 483, + 85, + 684, + 271 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7e79e3050f7c6ac6fdadeea3b3b6a16f4aba788a667d2823cd80049989fa559c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 85, + 888, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_i = \\frac{f(I_i)}{\\|f(I_i)\\|_2}$ and $\\mathbf{y}_i = \\frac{g(T_i)}{\\|g(T_i)\\|_2}$ . In this paper, we adopt the vision transformer architecture [17] for images and the transformer architecture [44] for texts. Note that due to the asymmetry of the softmax loss, the normalization is independently performed two times: across images and across texts [34]. The scalar $t$ is parametrized as $\\exp(t')$ , where $t'$ is a global freely learnable parameter.", + "bbox": [ + 75, + 434, + 468, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Sigmoid loss for language image pre-training", + "text_level": 1, + "bbox": [ + 76, + 554, + 457, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead of the softmax-based contrastive loss, we propose a simpler alternative that does not require computing global normalization factors. The sigmoid-based loss processes every image-text pair independently, effectively turning the learning problem into the standard binary classification on the dataset of all pair combinations, with a positive labels for the matching pairs $(I_i, T_i)$ and negative labels for all other pairs $(I_i, T_{j \\neq i})$ . It is defined as follows:", + "bbox": [ + 75, + 578, + 468, + 700 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n- \\frac {1}{| \\mathcal {B} |} \\sum_ {i = 1} ^ {| \\mathcal {B} |} \\sum_ {j = 1} ^ {| \\mathcal {B} |} \\underbrace {\\log \\frac {1}{1 + e ^ {z _ {i j} (- t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {j} + b)}}} _ {\\mathcal {L} _ {i j}}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 710, + 400, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $z_{ij}$ is the label for a given image and text input, which equals 1 if they are paired and -1 otherwise. Note that at initialization, the heavy imbalance coming from the many negatives dominates the loss, leading to large initial optimization steps attempting to correct this bias. To alleviate this, we introduce an additional learnable bias term $b$ similar to the temperature $t$ . We initialize $t'$ and $b$ to 10 and -10 respectively. This makes sure the training starts roughly close", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to the prior and does not require massive over-correction. Algorithm 1 presents a pseudocode implementation of the proposed sigmoid loss for language image pre-training.", + "bbox": [ + 498, + 436, + 890, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Efficient \"chunked\" implementation", + "text_level": 1, + "bbox": [ + 500, + 489, + 815, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive training typically utilizes data parallelism. Computing the loss when data is split across $D$ devices necessitates gathering all embeddings [55] with expensive all-gathers and, more importantly, the materialization of a memory-intensive $|\\mathcal{B}| \\times |\\mathcal{B}|$ matrix of pairwise similarities.", + "bbox": [ + 498, + 513, + 890, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The sigmoid loss, however, is particularly amenable to a memory efficient, fast, and numerically stable implementation that ameliorates both these issues. Denoting the perdevice batch size as $b = \\frac{|\\mathcal{B}|}{D}$ , the loss is reformulated as:", + "bbox": [ + 498, + 589, + 890, + 654 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/970912ada39eeb482401c9a2af5950c7c2af4acfff9bd0018fd6a049194b920b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 660, + 851, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This is particularly simple for the sigmoid loss as each pair is an independent term in the loss. Figure 1 illustrates this method. In words, we first compute the component of the loss corresponding to the positive pairs, and $b - 1$ negative pairs. We then permute representations across devices, so each device takes negatives from its neighbouring device (next iteration of sum B). The loss is then calculated with respect to this chunk (sum C). This is done independently in each device, such that each device computes the", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11977", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b6a0eca64575fcaa25a7711def362e40f517d55a837a8473122c285cc177fbbd.jpg", + "image_caption": [ + "Figure 2: The effect of pre-training batch size. Left: SigLiT results, trained for 18B seen examples. Sigmoid loss outperforms the softmax loss significantly with small batch sizes, and performs similarly at larger batch sizes. We successfully trained an SigLiT model with up to one million batch size. However, performance for both sigmoid and softmax saturate at around $32\\mathrm{k}$ batch size. Middle: SigLIP results, trained for 9B seen examples. Both sigmoid loss and softmax loss saturate at a reasonable batch size, while the peak of the sigmoid loss comes earlier and slightly outperforms the peak of the softmax loss. A very large batch size hurts both losses. Right: mSigLIP results, trained for 30B seen examples. With a multilingual setup using over 100 languages, $32\\mathrm{k}$ batch size is surprisingly sufficient and scaling beyond that hurts performance on a 36-language cross-modal retrieval task." + ], + "image_footnote": [], + "bbox": [ + 89, + 89, + 356, + 282 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d7d5f072cdf4d3412cacf41ede119e784e91cde95ebd583f4570da79f9548186.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 89, + 614, + 282 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9d01164e07d4eef84e6f33fa78d4eb81c626ee64cd20ac0d5ab0924f0fc47b48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 90, + 880, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "loss with respect to its local batch $b$ . Losses can then simply be summed across all devices (sum A). Individual collective permutes (for sum B) are fast (and indeed $D$ collective permutes is typically faster than two all-gathers between $D$ devices), and the memory cost at any given moment is reduced from $|\\mathcal{B}|^2$ to $b^2$ (for sum C). Usually $b$ is constant as scaling $|\\mathcal{B}|$ is achieved by increasing the number of accelerators. Due to being quadratic with respect to the batch size, the vanilla loss computation rapidly bottlenecks scaling up. This chunked approach enabled training with batch sizes over 1 million on relatively few devices.", + "bbox": [ + 75, + 446, + 472, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 76, + 638, + 163, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we evaluate the proposed SigLiT and SigLIP models across a wide range of batch sizes. We discuss what can be achieved with a small number of accelerator chips, using both SigLiT and SigLIP recipes. We also briefly discuss the impact of batch size on multilingual language image pre-training. We ablate the importance of our large-batch stabilization modification and the introduced learned bias term and present a study on the effect of positive and negative pairs ratio in the sigmoid loss. Lastly, we explore SigLIP's data noise robustness.", + "bbox": [ + 75, + 669, + 468, + 820 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To validate our models, we report zero-shot transfer results on the ImageNet dataset [14] and zero-shot retrieval results across 36 languages on the XM3600 dataset [41]. We use the ScalingViT-Adafactor optimizer [54] by default for all our experiments.", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. SigLiT: Scaling batch size to the limit", + "text_level": 1, + "bbox": [ + 498, + 445, + 823, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following [55], we use the same precomputed embeddings for the images using a ViT-g vision model, and train a base size text tower from scratch with the same hyperparameters using the LiT image-text dataset [55].", + "bbox": [ + 496, + 470, + 892, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We perform a study over a wide range of batch sizes, from 512 to $1M$ , demonstrating the impact of batch size for contrastive learning. Results are presented in Figure 2 (left). When the batch size is smaller than $16k$ , sigmoid loss outperforms softmax loss by a large margin. With growing batch sizes, we observe that softmax loss quickly catches up and potentially slightly underperforms sigmoid loss with a large enough batch size. Overall, we recommend using the SigLIP recipe for large batch sizes as well, due to the simplicity, compute savings, and straightforward memory efficient implementation.", + "bbox": [ + 496, + 535, + 893, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There is a consensus that contrastive learning benefits from large batch sizes, while most of the existing studies stop at $64\\mathrm{k}$ batch size [55, 33, 10]. We successfully trained an SigLiT model at one million batch size, to explore the limit of contrastive learning. To our surprise, the performance saturates at $32\\mathrm{k}$ batch size, further scaling up the batch size only gives a minor boost, and the model peaks at $256\\mathrm{k}$ batch size. Our best SigLiT with a $B$ -sized text mode achieves $84.7\\%$ zero-shot transfer accuracy on ImageNet, while the original LiT paper reports a slightly better $85.2\\%$ score with a 10 times larger $g$ -sized text model. Figure 3 presents the impact of training duration for different batch sizes. It demonstrates that large, $262k$ batch size signifi", + "bbox": [ + 496, + 704, + 895, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11978", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8e77149cddcaaa1e9ded9c72e0d68c055f9f4e30b49d6e13288a4ce900acd2da.jpg", + "image_caption": [ + "Figure 3: SigLiT ImageNet 0-shot transfer results with different training durations. Large batch size results in a big performance boost, but needs a sufficiently long schedule to ramp up, as for short schedules, very large batch size results in a small number of gradient update steps." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 452, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cantly outperforms smaller $8k$ batch size when trained for a sufficiently long time. Note, that for short training durations, large batch size leads to the fewer absolute number of update steps and thus needs more time to ramp up.", + "bbox": [ + 75, + 414, + 470, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. SigLIP: Sigmoid loss is beneficial for language-image pre-training", + "text_level": 1, + "bbox": [ + 76, + 502, + 468, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We pre-train SigLIP models on the WebLI dataset [13], using only English image and text pairs. We use moderately-sized models: B/16 ViT for image embeddings and B-sized transformer for text embeddings. The input images are resized to $224 \\times 224$ resolution. The text is tokenized by a $32\\mathrm{k}$ vocabulary sentencepiece tokenizer [27] trained on the English C4 dataset [35], and a maximum of 16 text tokens are kept. Figure 2 middle plot shows SigLIP results, With less than $32\\mathrm{k}$ batch size, SigLIP outperforms CLIP baselines with the standard softmax loss. On the other end of the scale, the memory efficiency of the sigmoid loss enabled much larger batch sizes. For example, with four TPU-v4 chips, we could fit a batch size of 4096 with a Base SigLIP but only 2048 with a corresponding CLIP model. The two advantages together demonstrate significant benefits of the sigmoid loss for language image pre-training with fixed resources, which will be discussed in Section 4.5.", + "bbox": [ + 75, + 546, + 468, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As batch size increases, the gap between the sigmoid and the softmax losses diminish. SigLIP performs best at batch size $32\\mathrm{k}$ , whereas the softmax loss required $98\\mathrm{k}$ for optimal performance and still didn't outperform the sigmoid based variant. Scaling further, a larger batch size like $307\\mathrm{k}$ hurts both losses.", + "bbox": [ + 75, + 809, + 470, + 898 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5b32bc4d8e4e01531b2edadbe7dba03567815f9783ca946f0f8fd9b10674228e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
16 k32 k64 k128 k240 k
INet-071.673.273.273.273.1
XM avg34.834.934.433.632.7
XM de54.754.855.454.354.7
XM en46.546.246.546.646.6
XM hi9.18.57.98.17.3
XM ru50.149.949.748.649.3
XM zh30.732.532.030.623.7
", + "bbox": [ + 498, + 88, + 893, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2: Multilingual SigLIP results with various batch sizes, pre-trained for 30 billion seen examples. We report zero-shot transfer results on ImageNet (INet-0) and averaged text to image retrieval results across 36 languages on the crossmodal 3600 dataset (XM). The full table on 36 languages can be found in Appendix.", + "bbox": [ + 496, + 256, + 893, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. mSigLIP: Multi-lingual pre-training", + "text_level": 1, + "bbox": [ + 500, + 402, + 816, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further scale up the training data by keeping all the 100 languages from the WebLI dataset [13]. With multilingual data, one usually needs to use a larger international vocabulary. We first verify the impact of two tokenizers: a small multilingual vocabulary with $32\\mathrm{k}$ tokens [35], and a large multilingual vocabulary with $250\\mathrm{k}$ tokens [50]. We train B-sized ViT and text models for $900M$ total examples seen, and observe slightly more than $1\\%$ improvement when using a larger vocabulary.", + "bbox": [ + 496, + 426, + 890, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, the token embeddings become huge for very large vocabulary sizes. Following the standard setup, we would need to store a $N \\times W$ token embedding lookup table to train the multilingual model, where $N$ is the vocabulary size mentioned above and $W$ is the embedding dimension of the text model. To save memory, we propose to use a \"bottlenecked\" token embedding. We use $N \\times K$ embedding matrix and additional $K \\times W$ projection, where the bottleneck $K$ is much smaller than $W$ .", + "bbox": [ + 496, + 565, + 890, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our experiments, we observed that using a large multilingual vocabulary with a bottleneck can be scaled up as efficiently as using a small multilingual vocabulary. Specifically, by enabling the bottleneck of size $K = 96$ for Base architecture with $W = 768$ , we only see about a half percent quality drop on ImageNet zero-shot transfer, compared to using the full $250k$ vocabulary.", + "bbox": [ + 496, + 702, + 890, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the memory improvements, we train mSigLIP models for various batch sizes, for a total of 30 billion examples seen. Table 2 and Figure 2 (right plot) show the results. We were expecting a large batch size to improve multilingual pre-training, where the model sees more examples from the same language as hard negatives in a sin", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11979", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a8d6942afea6fe1bd82be48abdd04d2c8e997bafd7cf91bed102f0dc5b56595e.jpg", + "image_caption": [ + "Figure 4: Top: SigLIP with pre-trained encoders ramps up quickly. However, only disabling weight decay on the pre-trained encoder weights leads to stable behavior and good ImageNet 0-shot transfer results. Bottom: ImageNet 10-shot transfer results, where decaying the pre-trained weights leads to deterioration of the pre-trained model visual representation quality. Disabling weight decay makes the curve flatter." + ], + "image_footnote": [], + "bbox": [ + 89, + 90, + 465, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "gle mini-batch. However, we didn't observe clear improvements with a batch size larger than $32\\mathrm{k}$ . A batch size of $32\\mathrm{k}$ is sufficient for a multilingual setup as well. On the XM3600 cross-modal retrieval tasks, we found that going beyond $32\\mathrm{k}$ batch size leads to worse results on average while on ImageNet zero-shot transfer it stays flat. mSigLIP sets the new state-of-the-art on XM3600 text to image retrieval task, with only a Base size model. Our best result is $34.9\\%$ , which is more than $6\\%$ higher than the previously reported result $28.5\\%$ [13] with a standard LiT model [55] using a much larger four billion ViT-e model.", + "bbox": [ + 75, + 526, + 472, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. SigLiT with four TPU-v4 chips", + "text_level": 1, + "bbox": [ + 76, + 707, + 352, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For many practitioners, the important question usually is \"what can be trained with a limited amount of resources?\" We explore the usage of SigLiT models in this section with only four TPU-v4 chips, as the memory efficient sigmoid loss is suitable for this application scenario.", + "bbox": [ + 75, + 732, + 468, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We follow the same setup as in section 4.1. We use the publicly available ViT-Augreg-B/8 [39] model as the frozen ( ) vision tower, and precompute embeddings to accelerate the training [55]. The text model is a Large Transformer, but with a depth of only 12 layers (instead of 24). It is trained using the LION [12] optimizer with decoupled weight de", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e87c68b11bc442c003f85424b4a8084854d3ccc5d0a5587f463252ad4ad1e0a4.jpg", + "image_caption": [ + "Figure 5: The effect of Adam and AdaFactor's $\\beta_{2}$ . As we increase batch-size, we observe more frequent training instability. This instability can mainly be seen in the loss curves (top) and is caused by spikes in the gradient norm (middle) which results in large parameter updates (bottom). Decreasing the $\\beta_{2}$ momentum value stabilizes the training. Even though occasional gradient spikes still happen (see step at 2B), they do not destabilize the training process." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 885, + 362 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cay $1 \\times 10^{-7}$ , linearly warm-up of learning rate over $6.5\\mathrm{k}$ steps up to a peak of $1 \\times 10^{-4}$ , followed by a cosine decay to 0. We train for a total of 65000 steps with a batch size of $32\\mathrm{k}$ - this leads to just under one day of training. Table 1 shows the results when training a model on four chips for one day, achieving $79.7\\%$ 0-shot ImageNet classification accuracy; very competitive in this limited resource regime. With a ViT-g/14 [54] model as the vision tower and a Large text tower, we can train at $20\\mathrm{k}$ batch size on four chips for $107\\mathrm{k}$ steps in under two days. This further pushes the 0-shot ImageNet classification accuracy up to $84.5\\%$ .", + "bbox": [ + 496, + 526, + 893, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. SigLIP with a small amount of TPU-v4 chips", + "text_level": 1, + "bbox": [ + 500, + 709, + 879, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It's resource demanding to train a CLIP model from scratch in general, with SigLIP it's possible to fit a larger train batch size with fewer amount of chips. In this section, we explore ways to train SigLIP models efficiently with pretrained weights. We use pre-trained weights to initialize the image model to accelerate the pre-training, which was originally discussed in [55]. We use the public and unlocked ViT-Augreg-B/16 [39] model to initialize our vision tower and fine-tune on the same WebLI English data as used for SigLIP. In all the experiments, we apply a 0.1 learning rate multiplier to the pre-trained image tower to make it suitable", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11980", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/98d1b8aa5b072bfb683bf8d0ab866d58f176ca50affa4e4cdf43d7f9209e646c.jpg", + "image_caption": [ + "Figure 6: The effect of batch composition. We simulate various batch compositions by masking out negatives, either randomly, keeping only the hardest, or the easiest. With no masking, we have $16\\mathrm{k}$ negatives for each positive in the batch $(1:16\\mathrm{k})$ and the strongest masking we apply (1:1.6) results in almost balanced minibatches. In one setting we match total pairs seen by training for significantly longer. We observe ImageNet 0-shot score, the final value of the learned bias, and the average logits of positive and negative pairs. Overall, the imbalance does not seem to be detrimental, but finding an efficient way of mining negatives might be beneficial." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 344, + 299 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84693baafeb673a4a06b73173a7d466c1058f5b6d800fd5ddd2c57e16b5d6884.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 348, + 90, + 612, + 297 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b3ac5c2d1fe1add17df1d7316c8d4529441fe14af09f3e7d9fe151ceda2b1e0c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 90, + 883, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for fine-tuning.", + "bbox": [ + 75, + 431, + 181, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4 presents unlocked fine-tuning results alongside from-scratch randomly initialized baselines. We used 16 TPU-v4 chips and train at $16\\mathrm{k}$ batch size for $2.4\\mathrm{B}$ examples seen. We found that the fine-tuning setup doesn't perform well out-of-the-box; this is consistent with prior works [55] where finetuning image models degraded visual representation quality. This is evidenced by ImageNet 10-shot linear classification, where in Figure 4 the fine-tuned setup is barely better than the from-scratch baseline.", + "bbox": [ + 75, + 446, + 468, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We hypothesize that the default weight decay applied to the pre-trained weights reduces their effectiveness. Motivated by the fine-tuning recipe from [17, 54, 25], that uses no weight decay, we also propose disabling weight decay on the pre-trained weights for SigLIP training. Weight decay is therefore only applied to the randomly initialized weights in the text model. This simple modification significantly improved SigLIP results. Figure 4 shows that with our improved recipe, SigLIP reaches $71\\%$ 0-shot accuracy on ImageNet, using $16k$ batch size, trained on 16 chips for three days. We also present from-scratch results in the bottom rows of Table 1: with 32 TPUv4 chips for only two days, SigLIP achieves $72.1\\%$ 0-shot accuracy. This presents a significant training cost reduction e.g. compared to CLIP (approx. 2500 TPUv3-days for $72.6\\%$ ) reported in [29].", + "bbox": [ + 75, + 582, + 468, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Scaling up SigLIP", + "text_level": 1, + "bbox": [ + 76, + 816, + 253, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we scale up SigLIP by \"overtraining\" the model [42, 1] for a long schedule. We present results in Table 3 using ViT-B or ViT-L as the vision encoder, while the text encoder is kept the same size (B or L respec", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tively). Following the recipe as described in Section 4.2, we train both models for 40 billion examples seen, but increase the number of image patches from $(224 / 16)^{2} = 196$ to $(256 / 16)^{2} = 256$ and the number of text tokens from 16 to 64. To get SigLIP results for different resolutions, we continue training the pre-trained model for 5 billion examples on the target resolutions (i.e. 224 or 384), using a $100\\mathrm{x}$ smaller learning rate without weight decay. In Table 3, we report zero-shot classification results on ImageNet [14], ObjectNet [2], ImageNetv2 [36], ImageNet Real [3], and zero-shot image-to-text $(\\mathrm{I}\\rightarrow \\mathrm{T})$ retrieval, text-to-image $(\\mathrm{I}\\rightarrow \\mathrm{T})$ retrieval results on MSCOCO [11]. Our SigLIP models outperform the previous models consistently with ViT-B and ViT-L architectures.", + "bbox": [ + 496, + 431, + 890, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.7. Stabilizing large-batch training", + "text_level": 1, + "bbox": [ + 498, + 650, + 776, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As we move to large batch sizes, the language image pretraining using transformers becomes increasingly more unstable, even when using a modestly-sized model (e.g. Base size). The reason for these instabilities is large spikes in the gradient norms, which translate to large-magnitude changes in the weights that may destabilize the training process, see Figure 5. We observe that reducing $\\beta_{2}$ in Adam and AdaFactor from its default 0.999 to 0.95 (which was suggested in [20, 9]) is enough to stabilize the training. Intuitively, this allows recovering from gradient spikes quicker. We opt for setting $\\beta_{2} = 0.95$ for all our experiments.", + "bbox": [ + 496, + 674, + 890, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.8. Negative ratio in sigmoid loss", + "text_level": 1, + "bbox": [ + 500, + 847, + 759, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "One question which arises when shifting the perspective from the softmax's \"pick the right class\" view to the sig-", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11981", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/512948c27dcfb58d9014b092a0d0b06d6e8aec92239559569990b9672142d5b5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodImage EncoderImageNet-1kCOCO
SizePatch #Validationv2RealObjectNetI→TT→I
CLIPB19668.361.9-55.352.433.1
OpenCLIPB19670.262.3-56.059.442.3
EVA-CLIPB19674.767.0-62.358.742.2
SigLIPB19676.369.682.870.764.447.2
SigLIPB25676.670.083.171.365.147.4
CLIPL25675.569.0-69.956.336.5
OpenCLIPL25674.061.1-66.462.146.1
EVA-CLIPL25679.875.3-72.963.747.5
SigLIPL25680.674.285.977.969.551.1
CLIPL57676.672.0-70.957.937.1
EVA-CLIPL57680.473.8-78.464.147.9
", + "bbox": [ + 76, + 88, + 890, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3: Comparison with other methods. Our SigLIP models outperform the previous models, e.g. OpenCLIP [22] and CLIP [34], by a significant margin on both the zero-shot classification tasks and the zero-shot retrieval tasks. Compared to the recently introduced EVA-CLIP [40] model, SigLIP-L outperforms EVA-CLIP-L on all the tasks with the same number of 256 patches. SigLIP-L (256 patches) also outperforms EVA-CLIP-L (576 patches) by a large margin on the retrieval tasks.", + "bbox": [ + 75, + 348, + 893, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "moid's \"rate this pair\" view, is the imbalance in positive versus negative pairs. For a batch size $|\\mathcal{B}|$ , the batch contains $|\\mathcal{B}|$ positive pairs, but $|\\mathcal{B}|^2 - |\\mathcal{B}|$ negative examples. In the modest batch-size of $16\\mathrm{k}$ , there are actually $268\\mathrm{M}$ negative examples for only $16\\mathrm{k}$ positive ones. At the same time, because the sigmoid loss decomposes into a sum of per-example losses, we can perform controlled experiments to study the effect of the mini-batch composition and distribution of examples visited. We run experiments in the SigLiT setup at batch-size $16\\mathrm{k}$ for $900\\mathrm{M}$ steps and vary the composition of the batch by masking out (i.e. ignoring) enough negative examples to reach a target \"positive: negative\" ratio, masking in the following ways:", + "bbox": [ + 75, + 436, + 468, + 633 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Random: Randomly choose negative pairs to mask.", + "- Hard: Keep hardest negative pairs (highest loss).", + "- Easy: Keep easiest negatives pairs (lowest loss).", + "- Hard + matching total pairs seen: Masking examples while training for a fixed number of steps does decrease the total number of pairs seen during training. Hence in the matched pairs setting, we increase the number of training steps by the masking ratio in order to keep the number of pairs seen constant." + ], + "bbox": [ + 94, + 638, + 467, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 6 shows the effect of the various masking strategies. Randomly removing negatives to rebalance does deteriorate performance. Keeping the easiest examples does not work at all, while keeping the hardest negatives does almost maintain the quality, indicating that, as could be expected, a lot of the learning on the negative side comes from the harder examples. This is further confirmed by the slightly", + "bbox": [ + 75, + 794, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "increased performance of training longer on the hardest examples in order to match the total pairs seen.", + "bbox": [ + 498, + 436, + 890, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also look at the value of the learned bias at the end of training as well as the average logit value for positive and negative examples across these settings, and find the result mostly follows what one would expect: as fewer negatives are present, the bias and logits become more positive overall. Interestingly, when training with more hard negative pairs, the average logits of positive pairs stays mostly flat.", + "bbox": [ + 496, + 481, + 890, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This study confirms that (1) the imbalance does not seem to be a major reason for concern, while at the same time (2) coming up with an efficient way of including more negative examples can be promising but is not trivial.", + "bbox": [ + 496, + 602, + 890, + 662 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d38ca012e323a40f75e0f1bb6e9119930055bcb53648f3d1c1555f4e73709fef.jpg", + "table_caption": [ + "Table 4: Bias (b) and temperature $(\\mathbf{t}^{\\prime})$ initialization. Results are reported using Base architecture, $8\\mathrm{k}$ batch size, trained for 900M examples. Enabling the bias term b with -10 initialization improves results consistently." + ], + "table_footnote": [], + "table_body": "
bt'INet-0Pet-0C100-0
n/alog 1062.081.859.9
-10log 1063.082.461.0
-10log 161.080.060.4
0log 1061.779.959.0
0log 153.773.253.8
", + "bbox": [ + 501, + 789, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11982", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3971470a9b627315836ee46a6b56f5162dba6ad57ca79f020f21438ca925d6f.jpg", + "image_caption": [ + "Figure 7: Sigmoid-training increases robustness to data noise. Titles show the type of corruption applied, and x-axes show the probability with which they are applied. With increasing corruption severity, M-scale models trained with sigmoid loss for 3.6 billion examples retain superiority over corresponding softmax baseline." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 890, + 234 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.9. Bias term in sigmoid loss", + "text_level": 1, + "bbox": [ + 76, + 321, + 305, + 338 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We ablate the bias term in the loss function, using the Base architecture with an $8\\mathrm{k}$ batch size, trained for 900M examples with the SigLIP setup. Zero-shot transfer results are reported on ImageNet [14], Oxford-iit pet [32] and Cifar100 [26]. Table 4 presents results with and without a bias term in the sigmoid loss.", + "bbox": [ + 75, + 345, + 470, + 436 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Enabling the bias term with a $-10$ initialization consistently improves performance across all tasks. This is because the bias term ensures that the training starts close to the prior, preventing dramatic over-correction in early optimization. In contrast, a randomly chosen bias term initialization, such as the 0 initialization in Table 4, fails to address the over-correction issue, leading to significantly worse results. This effect is particularly noticeable when using a small temperature $t^{\\prime}$ initialization. We set the bias and temperature initialization to $b = -10$ and $t^\\prime = \\log 10$ (hence $t = 10$ ) as the default for all experiments.", + "bbox": [ + 75, + 436, + 468, + 603 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.10. Label noise robustness", + "text_level": 1, + "bbox": [ + 76, + 614, + 295, + 628 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Prior works demonstrated improved robustness against label noise when using the sigmoid loss for classification models [3]. This property would be particularly useful here in the face of the famously noisy nature of popular large-scale image-text datasets. In order to study this for SigLIP, we train M/16 image models alongside an M text model at batch size 16384 for 3.6 billion seen examples. We corrupt the training data using one of the following methods:", + "bbox": [ + 75, + 638, + 468, + 758 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Image: With probability $p$ , replace the image with uniform random noise.", + "- Text: With probability $p$ , replace tokenized text with a new sequence of randomly sampled tokens, up to some (sampled) sequence length.", + "- Batch alignment: Randomly shuffle the ordering of $p\\%$ of the batch.", + "- Image & text: Apply both with probability $p$ each." + ], + "bbox": [ + 94, + 763, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "- Image, text & batch: Alongside (4), also shuffle fraction $p$ of alignments.", + "bbox": [ + 517, + 321, + 890, + 352 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Results from varying the likelihood of the corruption are shown in Figure 7. Models trained with sigmoid loss are increasingly robust to all kinds of added noise.", + "bbox": [ + 498, + 357, + 890, + 402 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 416, + 617, + 431 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conducted a study on two language-image pretraining instances that used the sigmoid loss: SigLiT and SigLIP. Our results demonstrate that the sigmoid loss performs better than the softmax baseline, particularly for small train batch sizes. This loss function is also more memory efficient, which allows larger train batch sizes without requiring additional resources. We performed a thorough investigation of the batch size in contrastive learning. Surprisingly, we found that a relatively modest batch size of $32\\mathrm{k}$ yielded nearly optimal performance. Further studies have been performed to understand better the introduced bias term in the sigmoid loss, robustness to data noises and the impact of positive and negative pairs ratio in the sigmoid loss. We hope this work will facilitate language-image pretraining research with limited resources.", + "bbox": [ + 496, + 441, + 893, + 667 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements. We thank Daniel Keysers, Ilya Tolstikhin, Olivier Bousquet and Michael Tschannen for their valuable feedback and discussions on this paper. We thank Joan Puigcerver, Josip Djolonga and Black Hechtman for discussions on efficient implementations of the chunked contrastive loss. We thank Kaiming He and Xinlei Chen for the discussion of $\\beta_{2}$ to stabilize the training. We also thank Ross Wightman for spotting a mistake in the pseudocode in the first version of this paper. Similarly, we thank Boris Dayma for spotting a typo making $t$ vs $t'$ confusing, which we fixed in the third version of this paper. We thank the Google Deepmind team for providing a supportive research environment. We use the big-vision codebase [5, 4] for all experiments in this project.", + "bbox": [ + 496, + 686, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11983", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 90, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ibrahim Alabdulmohsin, Xiaohua Zhai, Alexander Kolesnikov, and Lucas Beyer. Getting vit in shape: Scaling laws for compute-optimal model design. In NeurIPS, 2023. 7", + "[2] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. ObjectNet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. In NeurIPS, 2019. 7", + "[3] Lucas Beyer, Olivier J. Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with imagenet? CoRR, abs/2006.07159, 2020. 2, 7, 9", + "[4] Lucas Beyer, Xiaohua Zhai, and Alexander Kolesnikov. Better plain vit baselines for imagenet-1k, 2022. 9", + "[5] Lucas Beyer, Xiaohua Zhai, and Alexander Kolesnikov. Big vision. https://github.com/google-research/big_vision, 2022.9", + "[6] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.1", + "[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021. 1", + "[8] Feilong Chen, Duzhen Zhang, Minglun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. VLP: A survey on vision-language pre-training. Int. J. Autom. Comput., 20(1):38-56, 2023. 2", + "[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 1691-1703. PMLR, 2020. 7", + "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2, 4", + "[11] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. CoRR, abs/1504.00325, 2015. 7", + "[12] Xiangning Chen, Chen Liang, Da Huang, Esteban Real, Kaiyuan Wang, Yao Liu, Hieu Pham, Xuanyi Dong, Thang Luong, Cho-Jui Hsieh, Yifeng Lu, and Quoc V. Le. Symbolic discovery of optimization algorithms, 2023. 2, 6", + "[13] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Radu Soricut. Pali: A jointly-scaled multilingual language-image model. CoRR, abs/2209.06794, 2022. 1, 5, 6", + "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 4, 7, 9", + "[15] Karan Desai, Gaurav Kaul, Zubin Aysola, and Justin Johnson. Redcaps: Web-curated image-text data created by the people, for the people. In Joaquin Vanschoren and SaiKit Yeung, editors, Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks 1, NeurIPS Datasets and Benchmarks 2021, December 2021, virtual, 2021. 1", + "[16] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Shuyang Gu, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Clip itself is a strong fine-tuner: Achieving $85.7\\%$ and $88.0\\%$ top-1 accuracy with vit-b and vit-l onImagenet. CoRR, abs/2212.06138, 2022. 2", + "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth $16 \\times 16$ words: Transformers for image recognition at scale. In ICLR, 2021. 3, 7", + "[18] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016. http://www.deeplearningbook.org.1", + "[19] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In CVPR, volume 2, 2006. 2", + "[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15979-15988. IEEE, 2022. 7", + "[21] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. Scaling up vision-language pre-training for image captioning. CoRR, abs/2111.12233, 2021. 1, 2", + "[22] Gabriel Ilharco, Mitchell Wortsman, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. OpenCLIP. Zenodo, 2021. 8", + "[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1, 2", + "[24] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In Hugo Larochelle, Marc'Aurelio Ranzato, Raia Hadsell, MariaFlorina Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2", + "[25] Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, and Neil Houlsby." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11984", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Big transfer (BiT): General visual representation learning. In ECCV, 2020. 7", + "[26] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, Univ. of Toronto, 2009. 9", + "[27] Taku Kudo and John Richardson. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In EMNLP, 2018. 5, 13", + "[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors, International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 of Proceedings of Machine Learning Research, pages 12888-12900. PMLR, 2022. 2", + "[29] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. CoRR, abs/2212.00794, 2022. 2, 7", + "[30] Matthias Minderer, Alexey A. Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple open-vocabulary object detection. In Shai Avidan, Gabriel J. Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, volume 13670 of Lecture Notes in Computer Science, pages 728-755. Springer, 2022. 2", + "[31] Jishnu Mukhoti, Tsung-Yu Lin, Amid Poursaeed, Rui Wang, Ashish Shah, Philip H. S. Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning, 2022. 2", + "[32] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, 2012. 9", + "[33] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Hanxiao Liu, Adams Wei Yu, Minh-Thang Luong, Mingxing Tan, and Quoc V. Le. Combined scaling for zero-shot transfer learning. CoRR, abs/2111.10050, 2021. 2, 4", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 8", + "[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv e-prints, 2019. 5, 13", + "[36] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In ICML, 2019. 7", + "[37] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. CoRR, abs/2210.08402, 2022. 1", + "[38] Krishna Srinivasan, Karthik Raman, Jiecao Chen, Michael Bendersky, and Marc Najork. WIT: wikipedia-based image text dataset for multimodal multilingual machine learning. CoRR, abs/2103.01913, 2021. 1", + "[39] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your ViT? Data, augmentation, and regularization in vision transformers. CoRR, abs/2106.10270, 2021. 1, 6", + "[40] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. EVA-CLIP: improved training techniques for CLIP at scale. CoRR, abs/2303.15389, 2023. 8", + "[41] Ashish V. Thapliyal, Jordi Pont-Tuset, Xi Chen, and Radu Soricut. Crossmodal-3600: A massively multilingual multimodal evaluation dataset. In Yoav Goldberg, Zornitsa Kozareva, and Yue Zhang, editors, Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022, pages 715-729. Association for Computational Linguistics, 2022. 4", + "[42] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. CoRR, abs/2302.13971, 2023. 7", + "[43] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, abs/1807.03748, 2018. 2", + "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3", + "[45] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. GIT: A generative image-to-text transformer for vision and language. CoRR, abs/2205.14100, 2022. 1, 2", + "[46] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 2", + "[47] Ross Wightman, Hugo Touvron, and Hervé Jégou. Resnet strikes back: An improved training procedure in timm. CoRR, abs/2110.00476, 2021. 2", + "[48] Mitchell Wortsman. Reaching $80\\%$ zero-shot accuracy with OpenCLIP: VIT-G/14 trained on LAION-2B. https://web.archive.org/web/20230127012732/https://laion.ai/blog/giant-openclip/.2", + "[49] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. In IEEE/CVF Conference on Computer" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11985", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 7949-7961. IEEE, 2022. 2", + "[50] Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. mT5: A massively multilingual pre-trained text-to-text transformer. In NAACL-HLT, 2021. 5", + "[51] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Bin Xiao, Ce Liu, Lu Yuan, and Jianfeng Gao. Unified contrastive learning in image-text-label space. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 19141-19151. IEEE, 2022. 2", + "[52] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. CoRR, abs/2205.01917, 2022. 2", + "[53] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pengchuan Zhang. Florence: A new foundation model for computer vision. CoRR, abs/2111.11432, 2021. 2", + "[54] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. CVPR, 2022. 1, 4, 6, 7, 13", + "[55] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 18102-18112. IEEE, 2022. 1, 2, 3, 4, 6, 7, 13", + "[56] Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D. Manning, and Curtis P. Langlotz. Contrastive learning of medical visual representations from paired images and text. In Zachary C. Lipton, Rajesh Ranganath, Mark P. Sendak, Michael W. Sjoding, and Serena Yeung, editors, Proceedings of the Machine Learning for Healthcare Conference, MLHC 2022, 5-6 August 2022, Durham, NC, USA, volume 182 of Proceedings of Machine Learning Research, pages 2-25. PMLR, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 695 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "11986", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Sigmoid Loss for Language Image Pre-Training/85b123ac-a2cf-49ce-9282-93b24961a80b_model.json b/2023/Sigmoid Loss for Language Image Pre-Training/85b123ac-a2cf-49ce-9282-93b24961a80b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..47bf27387d65e848d438da3b209f498e488a206e --- /dev/null +++ b/2023/Sigmoid Loss for Language Image Pre-Training/85b123ac-a2cf-49ce-9282-93b24961a80b_model.json @@ -0,0 +1,2347 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.244, + 0.131, + 0.728, + 0.154 + ], + "angle": 0, + "content": "Sigmoid Loss for Language Image Pre-Training" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.181, + 0.761, + 0.217 + ], + "angle": 0, + "content": "Xiaohua Zhai\\* Basil Mustafa Alexander Kolesnikov Lucas Beyer\\* Google DeepMind, Zurich, Switzerland" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.22, + 0.694, + 0.234 + ], + "angle": 0, + "content": "{xzhai, basilm, akolesnikov, lbeyer}@google.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.284, + 0.314, + 0.301 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.317, + 0.473, + 0.591 + ], + "angle": 0, + "content": "We propose a simple pairwise sigmoid loss for image-text pre-training. Unlike standard contrastive learning with softmax normalization, the sigmoid loss operates solely on image-text pairs and does not require a global view of the pairwise similarities for normalization. The sigmoid loss simultaneously allows further scaling up the batch size, while also performing better at smaller batch sizes. With only four TPUv4 chips, we can train a Base CLIP model at \\(4k\\) batch size and a Large LiT model at \\(20k\\) batch size, the latter achieves \\(84.5\\%\\) ImageNet zero-shot accuracy in two days. This disentanglement of the batch size from the loss further allows us to study the impact of examples vs pairs and negative to positive ratio. Finally, we push the batch size to the extreme, up to one million, and find that the benefits of growing batch size quickly diminish, with a more reasonable batch size of \\(32k\\) being sufficient. We hope our research motivates further explorations in improving the quality and efficiency of language-image pre-training." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.62, + 0.21, + 0.636 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.647, + 0.47, + 0.812 + ], + "angle": 0, + "content": "Contrastive pre-training using weak supervision from image-text pairs found on the web is becoming the go-to method for obtaining generic computer vision backbones, slowly replacing pre-training on large labelled multi-class datasets. The high-level idea is to simultaneously learn an aligned representation space for images and texts using paired data. Seminal works CLIP [34] and ALIGN [23] established the viability of this approach at a large scale, and following their success, many large image-text datasets became available privately [55, 13, 21, 45] and publicly [37, 6, 15, 7, 38]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.814, + 0.47, + 0.876 + ], + "angle": 0, + "content": "The standard recipe to pre-train such models leverages the image-text contrastive objective. It aligns the image and text embeddings for matching (positive) image-text pairs while making sure that unrelated (negative) image-text pairs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.268, + 0.895, + 0.451 + ], + "angle": 0, + "content": "Table 1: SigLiT and SigLIP results. Sigmoid loss is memory efficient, allows larger batch sizes (BS) that unlocks language image pre-training with a small number of chips. SigLiT model with a frozen public L/16 checkpoint [39], trained on the LiT image-text dataset [55] using four TPU-v4 chips for one day, achieves \\(79.7\\%\\) 0-shot accuracy on ImageNet. The same setup with a g/14 checkpoint [54] leads to \\(84.5\\%\\) accuracy, trained for two days. With a public unlocked B/16 image checkpoint [39], trained on the WebLI dataset [13], SigLIP achieves \\(71.0\\%\\) 0-shot accuracy using 16 TPU-v4 chips for three days. The last two rows show results with randomly initialized models." + }, + { + "type": "table", + "bbox": [ + 0.5, + 0.463, + 0.892, + 0.581 + ], + "angle": 0, + "content": "
ImageTextBS#TPUv4DaysINet-0
SigLiTB/8L*32 k4179.7
SigLiTg/14L20 k4284.5
SigLIPB/16B16 k16371.0
SigLIPB/16B32 k32272.1
SigLIPB/16B32 k32573.4
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.632, + 0.583, + 0.892, + 0.596 + ], + "angle": 0, + "content": "* We use a variant of the \\( L \\) model with 12 layers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.894, + 0.746 + ], + "angle": 0, + "content": "are dissimilar in the embedding space. This is achieved via a batch-level softmax-based contrastive loss, applied twice to normalize the pairwise similarity scores across all images, then all texts. A naive implementation of the softmax is numerically unstable; it is usually stabilized by subtracting the maximum input value before applying the softmax [18], which requires another pass over the full batch." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose a simpler alternative: the sigmoid loss. It does not require any operation across the full batch and hence greatly simplifies the distributed loss implementation and boosts efficiency. Additionally, it conceptually decouples the batch size from the definition of the task. We compare the proposed sigmoid loss with the standard softmax loss across multiple setups. In particular, we investigate sigmoid-based loss with two prominent approaches for image-text learning: CLIP [34] and LiT [55], which we call sigmoid language image pre" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.209, + 0.9 + ], + "angle": 0, + "content": "*equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11975" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.288 + ], + "angle": 0, + "content": "training (SigLIP) and sigmoid LiT (SigLiT), respectively. We find that the sigmoid loss performs significantly better than the softmax loss when the batch size is smaller than \\(16\\mathrm{k}\\). As the train batch size grows, the gap closes. Importantly, the sigmoid loss is symmetric, requires just a single pass, and a typical implementation requires less memory than the softmax loss. This enables successful training of a SigLiT model at a batch size of one million. However, we find that the performance saturates with growing batch size, both for softmax and sigmoid. The good news is that a reasonable batch size, i.e. \\(32\\mathrm{k}\\), is sufficient for image-text pretraining. This conclusion also holds for multilingual SigLIP training on over 100 languages." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.474, + 0.501 + ], + "angle": 0, + "content": "In Table 1, we present setups for image-text pre-training that require a moderate amount of TPUv4 chips for training. SigLiT is surprisingly efficient, reaching \\(79.7\\%\\) zero-shot accuracy on ImageNet in just a single day on four chips. SigLIP's more demanding from-scratch training reaches \\(73.4\\%\\) zero-shot accuracy in 5 days with 32 TPUv4 chips. This compares favorably to prior works such as FLIP [29] and CLIP [34], which require approximately 5 and 10 days respectively on 256 TPUv3 cores. When fine-tuning a pretrained vision backbone in SigLIP, denoted as \\(\\boxed{ \\begin{array}{r l} \\end{array} }\\) in Table 1, we found that disabling the weight decay on the pre-trained backbone leads to better results (see Figure 4 for details). We hope our work paves the way for making the nascent language-image pre-training field more accessible." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.513, + 0.22, + 0.529 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.47, + 0.665 + ], + "angle": 0, + "content": "Contrastive learning with the sigmoid loss. One prior work proposes a similar sigmoid loss for the task of unsupervised dimensionality reduction [19]; in the scope of contrastive image-text learning, the vast majority of works rely on the softmax-based InfoNCE loss as popularized by [43]. In supervised classification, the sigmoid loss has already been shown to be slightly more effective and robust than the softmax loss [3, 47]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.47, + 0.806 + ], + "angle": 0, + "content": "Contrastive language-image pre-training has become popular since CLIP [34] and ALIGN [23] applied softmax contrastive learning [56, 43, 10, 24] to large-scale image-text datasets. Both models perform very well on zero-shot transfer tasks, including classification and retrieval. Follow-up works show that contrastively pre-trained models produce good representations for fine-tuning [49, 16], linear regression [23], object detection [30], semantic segmentation [31] and video tasks [53]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Generative language-image pre-training Besides softmax contrastive pre-training, various alternatives have been proposed. GIT [45], SimVLM [46], and LEMON [21] successfully pre-train models using a generative text decoder instead, while CoCa [52] adds such a decoder to the discriminative CLIP/ALIGN setup, thus combining the pros" + }, + { + "type": "code_caption", + "bbox": [ + 0.501, + 0.091, + 0.842, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Sigmoid loss pseudo-implementation." + }, + { + "type": "code", + "bbox": [ + 0.486, + 0.114, + 0.87, + 0.224 + ], + "angle": 0, + "content": "1 # img_emb : image model embedding [n, dim] \n2 # txt_emb : text model embedding [n, dim] \n3 # t_prime, b : learnable temperature and bias \n4 # n : mini-batch size \n5 \n6 t = exp(t_prime) \n7 zimg = l2_normalize(img_emb) \n8 ztxt = l2_normalize(txt_emb) \n9 logits = dot(zimg, ztxt.T) * t + b \n10 labels = 2 * eye(n) - ones(n) # -1 with diagonal 1 \n11 l = -sum(log_sigmoid(labels * logits)) / n" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.265, + 0.892, + 0.355 + ], + "angle": 0, + "content": "and cons of both approaches into a single very capable model. BLIP [28] further proposes CapFilt which uses the generative decoder to create better captions and the discriminative part of the model to filter pairs. Language-Image pre-training is a very active field and surveys [8] rapidly become outdated." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.893, + 0.528 + ], + "angle": 0, + "content": "Efficient language-image pre-training On the other hand, few works have tried making language image pre-training more efficient. LiT [55] and FLIP [29] are notable attempts, the former requires a pre-trained and locked backbone, and the latter sacrifices quality by randomly dropping visual tokens. BASIC [33] and LAION [48] look at scaling batch-size but only go up to \\(16\\mathrm{k}\\) and \\(160\\mathrm{k}\\) respectively, by using many hundreds of chips, and for the former also mixing in a large private classification dataset [33, 51]. The recent Lion optimizer [12] claims to be able to reduce the training cost to reach similar quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.545, + 0.591, + 0.56 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.616 + ], + "angle": 0, + "content": "In this section, we first review the widely-used softmax-based contrastive loss. We then introduce the pairwise sigmoid loss and discuss its efficient implementation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.617, + 0.893, + 0.739 + ], + "angle": 0, + "content": "Given a mini-batch \\(\\mathcal{B} = \\{(I_1, T_1), (I_2, T_2), \\ldots\\}\\) of image-text pairs, the contrastive learning objective encourages embeddings of matching pairs \\((I_i, T_i)\\) to align with each other, while pushing embeddings of unmatched pairs \\((I_i, T_{j \\neq i})\\) apart. For practical purposes, it is assumed that for all images \\(i\\), the text associated with a different image \\(j\\) is not related to \\(i\\), and vice-versa. This assumption is usually noisy and imperfect." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.751, + 0.882, + 0.768 + ], + "angle": 0, + "content": "3.1. Softmax loss for language image pre-training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.775, + 0.892, + 0.822 + ], + "angle": 0, + "content": "When using the softmax loss to formalize this objective, an image model \\( f(\\cdot) \\) and a text model \\( g(\\cdot) \\) are trained to minimize the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.833, + 0.871, + 0.906 + ], + "angle": 0, + "content": "\\[\n- \\frac {1}{2 | \\mathcal {B} |} \\sum_ {i = 1} ^ {| \\mathcal {B} |} \\left(\\overbrace {\\log \\frac {e ^ {t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {i}}}{\\sum_ {j = 1} ^ {| \\mathcal {B} |} e ^ {t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {j}}}} ^ {\\text {i m a g e} \\rightarrow \\text {t e x t s o f t m a x}} + \\overbrace {\\log \\frac {e ^ {t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {i}}}{\\sum_ {j = 1} ^ {| \\mathcal {B} |} e ^ {t \\mathbf {x} _ {j} \\cdot \\mathbf {y} _ {i}}}} ^ {\\text {t e x t} \\rightarrow \\text {i m a g e s o f t m a x}}\\right)\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11976" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.278, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.087, + 0.481, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.087, + 0.685, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.087, + 0.889, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.076, + 0.298, + 0.48, + 0.367 + ], + "angle": 0, + "content": "(a) Initially each device holds 4 (b) They each compute the comimage and 4 text representations. component of the loss (highlighted) Each device needs to see the repre for their representations, which representations from other devices includes the positives. to calculate the full loss." + }, + { + "type": "image_footnote", + "bbox": [ + 0.484, + 0.298, + 0.685, + 0.368 + ], + "angle": 0, + "content": "(c) Texts are swapped across the devices, so device 1 now has \\( I_{1:4} \\) and \\( T_{5:8} \\) etc. The new loss is computed and accumulated with the previous." + }, + { + "type": "image_footnote", + "bbox": [ + 0.688, + 0.299, + 0.889, + 0.369 + ], + "angle": 0, + "content": "(d) This repeats till every image & text pair have interacted, e.g. device 1 has the loss of \\( I_{1:4} \\) and \\( T_{1:12} \\). A final cross-device sum brings everything together." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.298, + 0.889, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.38, + 0.89, + 0.412 + ], + "angle": 0, + "content": "Figure 1: Efficient loss implementation demonstrated via a mock setup with 3 devices and a global batch size of 12. There are no all-gathers, and at any point in time only the bright yellow square (size \\(4 \\times 4\\)) is materialized in memory." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.435, + 0.47, + 0.545 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_i = \\frac{f(I_i)}{\\|f(I_i)\\|_2}\\) and \\(\\mathbf{y}_i = \\frac{g(T_i)}{\\|g(T_i)\\|_2}\\). In this paper, we adopt the vision transformer architecture [17] for images and the transformer architecture [44] for texts. Note that due to the asymmetry of the softmax loss, the normalization is independently performed two times: across images and across texts [34]. The scalar \\(t\\) is parametrized as \\(\\exp(t')\\), where \\(t'\\) is a global freely learnable parameter." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.555, + 0.458, + 0.573 + ], + "angle": 0, + "content": "3.2. Sigmoid loss for language image pre-training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.47, + 0.701 + ], + "angle": 0, + "content": "Instead of the softmax-based contrastive loss, we propose a simpler alternative that does not require computing global normalization factors. The sigmoid-based loss processes every image-text pair independently, effectively turning the learning problem into the standard binary classification on the dataset of all pair combinations, with a positive labels for the matching pairs \\((I_i, T_i)\\) and negative labels for all other pairs \\((I_i, T_{j \\neq i})\\). It is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.712, + 0.401, + 0.769 + ], + "angle": 0, + "content": "\\[\n- \\frac {1}{| \\mathcal {B} |} \\sum_ {i = 1} ^ {| \\mathcal {B} |} \\sum_ {j = 1} ^ {| \\mathcal {B} |} \\underbrace {\\log \\frac {1}{1 + e ^ {z _ {i j} (- t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {j} + b)}}} _ {\\mathcal {L} _ {i j}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\( z_{ij} \\) is the label for a given image and text input, which equals 1 if they are paired and -1 otherwise. Note that at initialization, the heavy imbalance coming from the many negatives dominates the loss, leading to large initial optimization steps attempting to correct this bias. To alleviate this, we introduce an additional learnable bias term \\( b \\) similar to the temperature \\( t \\). We initialize \\( t' \\) and \\( b \\) to 10 and -10 respectively. This makes sure the training starts roughly close" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.438, + 0.892, + 0.485 + ], + "angle": 0, + "content": "to the prior and does not require massive over-correction. Algorithm 1 presents a pseudocode implementation of the proposed sigmoid loss for language image pre-training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.491, + 0.816, + 0.507 + ], + "angle": 0, + "content": "3.3. Efficient \"chunked\" implementation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.514, + 0.892, + 0.59 + ], + "angle": 0, + "content": "Contrastive training typically utilizes data parallelism. Computing the loss when data is split across \\(D\\) devices necessitates gathering all embeddings [55] with expensive all-gathers and, more importantly, the materialization of a memory-intensive \\(|\\mathcal{B}| \\times |\\mathcal{B}|\\) matrix of pairwise similarities." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.59, + 0.892, + 0.655 + ], + "angle": 0, + "content": "The sigmoid loss, however, is particularly amenable to a memory efficient, fast, and numerically stable implementation that ameliorates both these issues. Denoting the perdevice batch size as \\( b = \\frac{|\\mathcal{B}|}{D} \\), the loss is reformulated as:" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.661, + 0.852, + 0.76 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This is particularly simple for the sigmoid loss as each pair is an independent term in the loss. Figure 1 illustrates this method. In words, we first compute the component of the loss corresponding to the positive pairs, and \\( b - 1 \\) negative pairs. We then permute representations across devices, so each device takes negatives from its neighbouring device (next iteration of sum B). The loss is then calculated with respect to this chunk (sum C). This is done independently in each device, such that each device computes the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11977" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.09, + 0.357, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.09, + 0.615, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.091, + 0.882, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.299, + 0.896, + 0.421 + ], + "angle": 0, + "content": "Figure 2: The effect of pre-training batch size. Left: SigLiT results, trained for 18B seen examples. Sigmoid loss outperforms the softmax loss significantly with small batch sizes, and performs similarly at larger batch sizes. We successfully trained an SigLiT model with up to one million batch size. However, performance for both sigmoid and softmax saturate at around \\(32\\mathrm{k}\\) batch size. Middle: SigLIP results, trained for 9B seen examples. Both sigmoid loss and softmax loss saturate at a reasonable batch size, while the peak of the sigmoid loss comes earlier and slightly outperforms the peak of the softmax loss. A very large batch size hurts both losses. Right: mSigLIP results, trained for 30B seen examples. With a multilingual setup using over 100 languages, \\(32\\mathrm{k}\\) batch size is surprisingly sufficient and scaling beyond that hurts performance on a 36-language cross-modal retrieval task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.473, + 0.614 + ], + "angle": 0, + "content": "loss with respect to its local batch \\( b \\). Losses can then simply be summed across all devices (sum A). Individual collective permutes (for sum B) are fast (and indeed \\( D \\) collective permutes is typically faster than two all-gathers between \\( D \\) devices), and the memory cost at any given moment is reduced from \\( |\\mathcal{B}|^2 \\) to \\( b^2 \\) (for sum C). Usually \\( b \\) is constant as scaling \\( |\\mathcal{B}| \\) is achieved by increasing the number of accelerators. Due to being quadratic with respect to the batch size, the vanilla loss computation rapidly bottlenecks scaling up. This chunked approach enabled training with batch sizes over 1 million on relatively few devices." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.64, + 0.165, + 0.657 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.47, + 0.821 + ], + "angle": 0, + "content": "In this section, we evaluate the proposed SigLiT and SigLIP models across a wide range of batch sizes. We discuss what can be achieved with a small number of accelerator chips, using both SigLiT and SigLIP recipes. We also briefly discuss the impact of batch size on multilingual language image pre-training. We ablate the importance of our large-batch stabilization modification and the introduced learned bias term and present a study on the effect of positive and negative pairs ratio in the sigmoid loss. Lastly, we explore SigLIP's data noise robustness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To validate our models, we report zero-shot transfer results on the ImageNet dataset [14] and zero-shot retrieval results across 36 languages on the XM3600 dataset [41]. We use the ScalingViT-Adafactor optimizer [54] by default for all our experiments." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.446, + 0.825, + 0.464 + ], + "angle": 0, + "content": "4.1. SigLiT: Scaling batch size to the limit" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.893, + 0.534 + ], + "angle": 0, + "content": "Following [55], we use the same precomputed embeddings for the images using a ViT-g vision model, and train a base size text tower from scratch with the same hyperparameters using the LiT image-text dataset [55]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.895, + 0.702 + ], + "angle": 0, + "content": "We perform a study over a wide range of batch sizes, from 512 to \\(1M\\), demonstrating the impact of batch size for contrastive learning. Results are presented in Figure 2 (left). When the batch size is smaller than \\(16k\\), sigmoid loss outperforms softmax loss by a large margin. With growing batch sizes, we observe that softmax loss quickly catches up and potentially slightly underperforms sigmoid loss with a large enough batch size. Overall, we recommend using the SigLIP recipe for large batch sizes as well, due to the simplicity, compute savings, and straightforward memory efficient implementation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.896, + 0.903 + ], + "angle": 0, + "content": "There is a consensus that contrastive learning benefits from large batch sizes, while most of the existing studies stop at \\(64\\mathrm{k}\\) batch size [55, 33, 10]. We successfully trained an SigLiT model at one million batch size, to explore the limit of contrastive learning. To our surprise, the performance saturates at \\(32\\mathrm{k}\\) batch size, further scaling up the batch size only gives a minor boost, and the model peaks at \\(256\\mathrm{k}\\) batch size. Our best SigLiT with a \\(B\\)-sized text mode achieves \\(84.7\\%\\) zero-shot transfer accuracy on ImageNet, while the original LiT paper reports a slightly better \\(85.2\\%\\) score with a 10 times larger \\(g\\)-sized text model. Figure 3 presents the impact of training duration for different batch sizes. It demonstrates that large, \\(262k\\) batch size signifi" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "11978" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.092, + 0.454, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.3, + 0.472, + 0.378 + ], + "angle": 0, + "content": "Figure 3: SigLiT ImageNet 0-shot transfer results with different training durations. Large batch size results in a big performance boost, but needs a sufficiently long schedule to ramp up, as for short schedules, very large batch size results in a small number of gradient update steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.472, + 0.477 + ], + "angle": 0, + "content": "cantly outperforms smaller \\(8k\\) batch size when trained for a sufficiently long time. Note, that for short training durations, large batch size leads to the fewer absolute number of update steps and thus needs more time to ramp up." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.503, + 0.47, + 0.536 + ], + "angle": 0, + "content": "4.2. SigLIP: Sigmoid loss is beneficial for language-image pre-training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.804 + ], + "angle": 0, + "content": "We pre-train SigLIP models on the WebLI dataset [13], using only English image and text pairs. We use moderately-sized models: B/16 ViT for image embeddings and B-sized transformer for text embeddings. The input images are resized to \\(224 \\times 224\\) resolution. The text is tokenized by a \\(32\\mathrm{k}\\) vocabulary sentencepiece tokenizer [27] trained on the English C4 dataset [35], and a maximum of 16 text tokens are kept. Figure 2 middle plot shows SigLIP results, With less than \\(32\\mathrm{k}\\) batch size, SigLIP outperforms CLIP baselines with the standard softmax loss. On the other end of the scale, the memory efficiency of the sigmoid loss enabled much larger batch sizes. For example, with four TPU-v4 chips, we could fit a batch size of 4096 with a Base SigLIP but only 2048 with a corresponding CLIP model. The two advantages together demonstrate significant benefits of the sigmoid loss for language image pre-training with fixed resources, which will be discussed in Section 4.5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.9 + ], + "angle": 0, + "content": "As batch size increases, the gap between the sigmoid and the softmax losses diminish. SigLIP performs best at batch size \\(32\\mathrm{k}\\), whereas the softmax loss required \\(98\\mathrm{k}\\) for optimal performance and still didn't outperform the sigmoid based variant. Scaling further, a larger batch size like \\(307\\mathrm{k}\\) hurts both losses." + }, + { + "type": "table", + "bbox": [ + 0.499, + 0.089, + 0.895, + 0.248 + ], + "angle": 0, + "content": "
16 k32 k64 k128 k240 k
INet-071.673.273.273.273.1
XM avg34.834.934.433.632.7
XM de54.754.855.454.354.7
XM en46.546.246.546.646.6
XM hi9.18.57.98.17.3
XM ru50.149.949.748.649.3
XM zh30.732.532.030.623.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.257, + 0.895, + 0.349 + ], + "angle": 0, + "content": "Table 2: Multilingual SigLIP results with various batch sizes, pre-trained for 30 billion seen examples. We report zero-shot transfer results on ImageNet (INet-0) and averaged text to image retrieval results across 36 languages on the crossmodal 3600 dataset (XM). The full table on 36 languages can be found in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.403, + 0.818, + 0.42 + ], + "angle": 0, + "content": "4.3. mSigLIP: Multi-lingual pre-training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.892, + 0.563 + ], + "angle": 0, + "content": "We further scale up the training data by keeping all the 100 languages from the WebLI dataset [13]. With multilingual data, one usually needs to use a larger international vocabulary. We first verify the impact of two tokenizers: a small multilingual vocabulary with \\(32\\mathrm{k}\\) tokens [35], and a large multilingual vocabulary with \\(250\\mathrm{k}\\) tokens [50]. We train B-sized ViT and text models for \\(900M\\) total examples seen, and observe slightly more than \\(1\\%\\) improvement when using a larger vocabulary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.892, + 0.701 + ], + "angle": 0, + "content": "However, the token embeddings become huge for very large vocabulary sizes. Following the standard setup, we would need to store a \\( N \\times W \\) token embedding lookup table to train the multilingual model, where \\( N \\) is the vocabulary size mentioned above and \\( W \\) is the embedding dimension of the text model. To save memory, we propose to use a \"bottlenecked\" token embedding. We use \\( N \\times K \\) embedding matrix and additional \\( K \\times W \\) projection, where the bottleneck \\( K \\) is much smaller than \\( W \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.809 + ], + "angle": 0, + "content": "In our experiments, we observed that using a large multilingual vocabulary with a bottleneck can be scaled up as efficiently as using a small multilingual vocabulary. Specifically, by enabling the bottleneck of size \\( K = 96 \\) for Base architecture with \\( W = 768 \\), we only see about a half percent quality drop on ImageNet zero-shot transfer, compared to using the full \\( 250k \\) vocabulary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "With the memory improvements, we train mSigLIP models for various batch sizes, for a total of 30 billion examples seen. Table 2 and Figure 2 (right plot) show the results. We were expecting a large batch size to improve multilingual pre-training, where the model sees more examples from the same language as hard negatives in a sin" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11979" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.092, + 0.466, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.376, + 0.473, + 0.496 + ], + "angle": 0, + "content": "Figure 4: Top: SigLIP with pre-trained encoders ramps up quickly. However, only disabling weight decay on the pre-trained encoder weights leads to stable behavior and good ImageNet 0-shot transfer results. Bottom: ImageNet 10-shot transfer results, where decaying the pre-trained weights leads to deterioration of the pre-trained model visual representation quality. Disabling weight decay makes the curve flatter." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.527, + 0.473, + 0.695 + ], + "angle": 0, + "content": "gle mini-batch. However, we didn't observe clear improvements with a batch size larger than \\(32\\mathrm{k}\\). A batch size of \\(32\\mathrm{k}\\) is sufficient for a multilingual setup as well. On the XM3600 cross-modal retrieval tasks, we found that going beyond \\(32\\mathrm{k}\\) batch size leads to worse results on average while on ImageNet zero-shot transfer it stays flat. mSigLIP sets the new state-of-the-art on XM3600 text to image retrieval task, with only a Base size model. Our best result is \\(34.9\\%\\), which is more than \\(6\\%\\) higher than the previously reported result \\(28.5\\%\\) [13] with a standard LiT model [55] using a much larger four billion ViT-e model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.708, + 0.353, + 0.725 + ], + "angle": 0, + "content": "4.4. SigLiT with four TPU-v4 chips" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.809 + ], + "angle": 0, + "content": "For many practitioners, the important question usually is \"what can be trained with a limited amount of resources?\" We explore the usage of SigLiT models in this section with only four TPU-v4 chips, as the memory efficient sigmoid loss is suitable for this application scenario." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We follow the same setup as in section 4.1. We use the publicly available ViT-Augreg-B/8 [39] model as the frozen ( ) vision tower, and precompute embeddings to accelerate the training [55]. The text model is a Large Transformer, but with a depth of only 12 layers (instead of 24). It is trained using the LION [12] optimizer with decoupled weight de" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.092, + 0.887, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.376, + 0.895, + 0.499 + ], + "angle": 0, + "content": "Figure 5: The effect of Adam and AdaFactor's \\(\\beta_{2}\\). As we increase batch-size, we observe more frequent training instability. This instability can mainly be seen in the loss curves (top) and is caused by spikes in the gradient norm (middle) which results in large parameter updates (bottom). Decreasing the \\(\\beta_{2}\\) momentum value stabilizes the training. Even though occasional gradient spikes still happen (see step at 2B), they do not destabilize the training process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.895, + 0.696 + ], + "angle": 0, + "content": "cay \\(1 \\times 10^{-7}\\), linearly warm-up of learning rate over \\(6.5\\mathrm{k}\\) steps up to a peak of \\(1 \\times 10^{-4}\\), followed by a cosine decay to 0. We train for a total of 65000 steps with a batch size of \\(32\\mathrm{k}\\) - this leads to just under one day of training. Table 1 shows the results when training a model on four chips for one day, achieving \\(79.7\\%\\) 0-shot ImageNet classification accuracy; very competitive in this limited resource regime. With a ViT-g/14 [54] model as the vision tower and a Large text tower, we can train at \\(20\\mathrm{k}\\) batch size on four chips for \\(107\\mathrm{k}\\) steps in under two days. This further pushes the 0-shot ImageNet classification accuracy up to \\(84.5\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.88, + 0.727 + ], + "angle": 0, + "content": "4.5. SigLIP with a small amount of TPU-v4 chips" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "It's resource demanding to train a CLIP model from scratch in general, with SigLIP it's possible to fit a larger train batch size with fewer amount of chips. In this section, we explore ways to train SigLIP models efficiently with pretrained weights. We use pre-trained weights to initialize the image model to accelerate the pre-training, which was originally discussed in [55]. We use the public and unlocked ViT-Augreg-B/16 [39] model to initialize our vision tower and fine-tune on the same WebLI English data as used for SigLIP. In all the experiments, we apply a 0.1 learning rate multiplier to the pre-trained image tower to make it suitable" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11980" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.345, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.092, + 0.614, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.092, + 0.885, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.314, + 0.892, + 0.405 + ], + "angle": 0, + "content": "Figure 6: The effect of batch composition. We simulate various batch compositions by masking out negatives, either randomly, keeping only the hardest, or the easiest. With no masking, we have \\(16\\mathrm{k}\\) negatives for each positive in the batch \\((1:16\\mathrm{k})\\) and the strongest masking we apply (1:1.6) results in almost balanced minibatches. In one setting we match total pairs seen by training for significantly longer. We observe ImageNet 0-shot score, the final value of the learned bias, and the average logits of positive and negative pairs. Overall, the imbalance does not seem to be detrimental, but finding an efficient way of mining negatives might be beneficial." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.182, + 0.447 + ], + "angle": 0, + "content": "for fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.47, + 0.582 + ], + "angle": 0, + "content": "Figure 4 presents unlocked fine-tuning results alongside from-scratch randomly initialized baselines. We used 16 TPU-v4 chips and train at \\(16\\mathrm{k}\\) batch size for \\(2.4\\mathrm{B}\\) examples seen. We found that the fine-tuning setup doesn't perform well out-of-the-box; this is consistent with prior works [55] where finetuning image models degraded visual representation quality. This is evidenced by ImageNet 10-shot linear classification, where in Figure 4 the fine-tuned setup is barely better than the from-scratch baseline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.81 + ], + "angle": 0, + "content": "We hypothesize that the default weight decay applied to the pre-trained weights reduces their effectiveness. Motivated by the fine-tuning recipe from [17, 54, 25], that uses no weight decay, we also propose disabling weight decay on the pre-trained weights for SigLIP training. Weight decay is therefore only applied to the randomly initialized weights in the text model. This simple modification significantly improved SigLIP results. Figure 4 shows that with our improved recipe, SigLIP reaches \\(71\\%\\) 0-shot accuracy on ImageNet, using \\(16k\\) batch size, trained on 16 chips for three days. We also present from-scratch results in the bottom rows of Table 1: with 32 TPUv4 chips for only two days, SigLIP achieves \\(72.1\\%\\) 0-shot accuracy. This presents a significant training cost reduction e.g. compared to CLIP (approx. 2500 TPUv3-days for \\(72.6\\%\\)) reported in [29]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.254, + 0.833 + ], + "angle": 0, + "content": "4.6. Scaling up SigLIP" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this section, we scale up SigLIP by \"overtraining\" the model [42, 1] for a long schedule. We present results in Table 3 using ViT-B or ViT-L as the vision encoder, while the text encoder is kept the same size (B or L respec" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.892, + 0.643 + ], + "angle": 0, + "content": "tively). Following the recipe as described in Section 4.2, we train both models for 40 billion examples seen, but increase the number of image patches from \\((224 / 16)^{2} = 196\\) to \\((256 / 16)^{2} = 256\\) and the number of text tokens from 16 to 64. To get SigLIP results for different resolutions, we continue training the pre-trained model for 5 billion examples on the target resolutions (i.e. 224 or 384), using a \\(100\\mathrm{x}\\) smaller learning rate without weight decay. In Table 3, we report zero-shot classification results on ImageNet [14], ObjectNet [2], ImageNetv2 [36], ImageNet Real [3], and zero-shot image-to-text \\((\\mathrm{I}\\rightarrow \\mathrm{T})\\) retrieval, text-to-image \\((\\mathrm{I}\\rightarrow \\mathrm{T})\\) retrieval results on MSCOCO [11]. Our SigLIP models outperform the previous models consistently with ViT-B and ViT-L architectures." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.651, + 0.777, + 0.667 + ], + "angle": 0, + "content": "4.7. Stabilizing large-batch training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.84 + ], + "angle": 0, + "content": "As we move to large batch sizes, the language image pretraining using transformers becomes increasingly more unstable, even when using a modestly-sized model (e.g. Base size). The reason for these instabilities is large spikes in the gradient norms, which translate to large-magnitude changes in the weights that may destabilize the training process, see Figure 5. We observe that reducing \\(\\beta_{2}\\) in Adam and AdaFactor from its default 0.999 to 0.95 (which was suggested in [20, 9]) is enough to stabilize the training. Intuitively, this allows recovering from gradient spikes quicker. We opt for setting \\(\\beta_{2} = 0.95\\) for all our experiments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.761, + 0.864 + ], + "angle": 0, + "content": "4.8. Negative ratio in sigmoid loss" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "One question which arises when shifting the perspective from the softmax's \"pick the right class\" view to the sig-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11981" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.892, + 0.339 + ], + "angle": 0, + "content": "
MethodImage EncoderImageNet-1kCOCO
SizePatch #Validationv2RealObjectNetI→TT→I
CLIPB19668.361.9-55.352.433.1
OpenCLIPB19670.262.3-56.059.442.3
EVA-CLIPB19674.767.0-62.358.742.2
SigLIPB19676.369.682.870.764.447.2
SigLIPB25676.670.083.171.365.147.4
CLIPL25675.569.0-69.956.336.5
OpenCLIPL25674.061.1-66.462.146.1
EVA-CLIPL25679.875.3-72.963.747.5
SigLIPL25680.674.285.977.969.551.1
CLIPL57676.672.0-70.957.937.1
EVA-CLIPL57680.473.8-78.464.147.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.349, + 0.894, + 0.412 + ], + "angle": 0, + "content": "Table 3: Comparison with other methods. Our SigLIP models outperform the previous models, e.g. OpenCLIP [22] and CLIP [34], by a significant margin on both the zero-shot classification tasks and the zero-shot retrieval tasks. Compared to the recently introduced EVA-CLIP [40] model, SigLIP-L outperforms EVA-CLIP-L on all the tasks with the same number of 256 patches. SigLIP-L (256 patches) also outperforms EVA-CLIP-L (576 patches) by a large margin on the retrieval tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.438, + 0.47, + 0.634 + ], + "angle": 0, + "content": "moid's \"rate this pair\" view, is the imbalance in positive versus negative pairs. For a batch size \\( |\\mathcal{B}| \\), the batch contains \\( |\\mathcal{B}| \\) positive pairs, but \\( |\\mathcal{B}|^2 - |\\mathcal{B}| \\) negative examples. In the modest batch-size of \\( 16\\mathrm{k} \\), there are actually \\( 268\\mathrm{M} \\) negative examples for only \\( 16\\mathrm{k} \\) positive ones. At the same time, because the sigmoid loss decomposes into a sum of per-example losses, we can perform controlled experiments to study the effect of the mini-batch composition and distribution of examples visited. We run experiments in the SigLiT setup at batch-size \\( 16\\mathrm{k} \\) for \\( 900\\mathrm{M} \\) steps and vary the composition of the batch by masking out (i.e. ignoring) enough negative examples to reach a target \"positive: negative\" ratio, masking in the following ways:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.639, + 0.455, + 0.653 + ], + "angle": 0, + "content": "- Random: Randomly choose negative pairs to mask." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.659, + 0.437, + 0.674 + ], + "angle": 0, + "content": "- Hard: Keep hardest negative pairs (highest loss)." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.68, + 0.431, + 0.694 + ], + "angle": 0, + "content": "- Easy: Keep easiest negatives pairs (lowest loss)." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.699, + 0.468, + 0.789 + ], + "angle": 0, + "content": "- Hard + matching total pairs seen: Masking examples while training for a fixed number of steps does decrease the total number of pairs seen during training. Hence in the matched pairs setting, we increase the number of training steps by the masking ratio in order to keep the number of pairs seen constant." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.639, + 0.468, + 0.789 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Figure 6 shows the effect of the various masking strategies. Randomly removing negatives to rebalance does deteriorate performance. Keeping the easiest examples does not work at all, while keeping the hardest negatives does almost maintain the quality, indicating that, as could be expected, a lot of the learning on the negative side comes from the harder examples. This is further confirmed by the slightly" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.438, + 0.892, + 0.468 + ], + "angle": 0, + "content": "increased performance of training longer on the hardest examples in order to match the total pairs seen." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.892, + 0.589 + ], + "angle": 0, + "content": "We also look at the value of the learned bias at the end of training as well as the average logit value for positive and negative examples across these settings, and find the result mostly follows what one would expect: as fewer negatives are present, the bias and logits become more positive overall. Interestingly, when training with more hard negative pairs, the average logits of positive pairs stays mostly flat." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.892, + 0.663 + ], + "angle": 0, + "content": "This study confirms that (1) the imbalance does not seem to be a major reason for concern, while at the same time (2) coming up with an efficient way of including more negative examples can be promising but is not trivial." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.716, + 0.892, + 0.777 + ], + "angle": 0, + "content": "Table 4: Bias (b) and temperature \\((\\mathbf{t}^{\\prime})\\) initialization. Results are reported using Base architecture, \\(8\\mathrm{k}\\) batch size, trained for 900M examples. Enabling the bias term b with -10 initialization improves results consistently." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.79, + 0.892, + 0.903 + ], + "angle": 0, + "content": "
bt'INet-0Pet-0C100-0
n/alog 1062.081.859.9
-10log 1063.082.461.0
-10log 161.080.060.4
0log 1061.779.959.0
0log 153.773.253.8
" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11982" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.891, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.25, + 0.895, + 0.298 + ], + "angle": 0, + "content": "Figure 7: Sigmoid-training increases robustness to data noise. Titles show the type of corruption applied, and x-axes show the probability with which they are applied. With increasing corruption severity, M-scale models trained with sigmoid loss for 3.6 billion examples retain superiority over corresponding softmax baseline." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.322, + 0.307, + 0.339 + ], + "angle": 0, + "content": "4.9. Bias term in sigmoid loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.346, + 0.471, + 0.437 + ], + "angle": 0, + "content": "We ablate the bias term in the loss function, using the Base architecture with an \\(8\\mathrm{k}\\) batch size, trained for 900M examples with the SigLIP setup. Zero-shot transfer results are reported on ImageNet [14], Oxford-iit pet [32] and Cifar100 [26]. Table 4 presents results with and without a bias term in the sigmoid loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.438, + 0.47, + 0.604 + ], + "angle": 0, + "content": "Enabling the bias term with a \\(-10\\) initialization consistently improves performance across all tasks. This is because the bias term ensures that the training starts close to the prior, preventing dramatic over-correction in early optimization. In contrast, a randomly chosen bias term initialization, such as the 0 initialization in Table 4, fails to address the over-correction issue, leading to significantly worse results. This effect is particularly noticeable when using a small temperature \\(t^{\\prime}\\) initialization. We set the bias and temperature initialization to \\(b = -10\\) and \\(t^\\prime = \\log 10\\) (hence \\(t = 10\\)) as the default for all experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.615, + 0.296, + 0.63 + ], + "angle": 0, + "content": "4.10. Label noise robustness" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.639, + 0.47, + 0.76 + ], + "angle": 0, + "content": "Prior works demonstrated improved robustness against label noise when using the sigmoid loss for classification models [3]. This property would be particularly useful here in the face of the famously noisy nature of popular large-scale image-text datasets. In order to study this for SigLIP, we train M/16 image models alongside an M text model at batch size 16384 for 3.6 billion seen examples. We corrupt the training data using one of the following methods:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.765, + 0.469, + 0.794 + ], + "angle": 0, + "content": "- Image: With probability \\( p \\), replace the image with uniform random noise." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.8, + 0.469, + 0.845 + ], + "angle": 0, + "content": "- Text: With probability \\( p \\), replace tokenized text with a new sequence of randomly sampled tokens, up to some (sampled) sequence length." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.851, + 0.469, + 0.88 + ], + "angle": 0, + "content": "- Batch alignment: Randomly shuffle the ordering of \\( p\\% \\) of the batch." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.886, + 0.448, + 0.901 + ], + "angle": 0, + "content": "- Image & text: Apply both with probability \\( p \\) each." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.323, + 0.892, + 0.353 + ], + "angle": 0, + "content": "- Image, text & batch: Alongside (4), also shuffle fraction \\( p \\) of alignments." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.358, + 0.892, + 0.404 + ], + "angle": 0, + "content": "Results from varying the likelihood of the corruption are shown in Figure 7. Models trained with sigmoid loss are increasingly robust to all kinds of added noise." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.417, + 0.619, + 0.433 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.895, + 0.669 + ], + "angle": 0, + "content": "We conducted a study on two language-image pretraining instances that used the sigmoid loss: SigLiT and SigLIP. Our results demonstrate that the sigmoid loss performs better than the softmax baseline, particularly for small train batch sizes. This loss function is also more memory efficient, which allows larger train batch sizes without requiring additional resources. We performed a thorough investigation of the batch size in contrastive learning. Surprisingly, we found that a relatively modest batch size of \\(32\\mathrm{k}\\) yielded nearly optimal performance. Further studies have been performed to understand better the introduced bias term in the sigmoid loss, robustness to data noises and the impact of positive and negative pairs ratio in the sigmoid loss. We hope this work will facilitate language-image pretraining research with limited resources." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.895, + 0.899 + ], + "angle": 0, + "content": "Acknowledgements. We thank Daniel Keysers, Ilya Tolstikhin, Olivier Bousquet and Michael Tschannen for their valuable feedback and discussions on this paper. We thank Joan Puigcerver, Josip Djolonga and Black Hechtman for discussions on efficient implementations of the chunked contrastive loss. We thank Kaiming He and Xinlei Chen for the discussion of \\(\\beta_{2}\\) to stabilize the training. We also thank Ross Wightman for spotting a mistake in the pseudocode in the first version of this paper. Similarly, we thank Boris Dayma for spotting a typo making \\(t\\) vs \\(t'\\) confusing, which we fixed in the third version of this paper. We thank the Google Deepmind team for providing a supportive research environment. We use the big-vision codebase [5, 4] for all experiments in this project." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11983" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.091, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Ibrahim Alabdulmohsin, Xiaohua Zhai, Alexander Kolesnikov, and Lucas Beyer. Getting vit in shape: Scaling laws for compute-optimal model design. In NeurIPS, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. ObjectNet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. In NeurIPS, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.286 + ], + "angle": 0, + "content": "[3] Lucas Beyer, Olivier J. Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with imagenet? CoRR, abs/2006.07159, 2020. 2, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.47, + 0.316 + ], + "angle": 0, + "content": "[4] Lucas Beyer, Xiaohua Zhai, and Alexander Kolesnikov. Better plain vit baselines for imagenet-1k, 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.318, + 0.47, + 0.358 + ], + "angle": 0, + "content": "[5] Lucas Beyer, Xiaohua Zhai, and Alexander Kolesnikov. Big vision. https://github.com/google-research/big_vision, 2022.9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.361, + 0.468, + 0.416 + ], + "angle": 0, + "content": "[6] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.419, + 0.468, + 0.473 + ], + "angle": 0, + "content": "[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.476, + 0.468, + 0.53 + ], + "angle": 0, + "content": "[8] Feilong Chen, Duzhen Zhang, Minglun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. VLP: A survey on vision-language pre-training. Int. J. Autom. Comput., 20(1):38-56, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.533, + 0.468, + 0.628 + ], + "angle": 0, + "content": "[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 1691-1703. PMLR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.468, + 0.673 + ], + "angle": 0, + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.468, + 0.73 + ], + "angle": 0, + "content": "[11] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. CoRR, abs/1504.00325, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.468, + 0.788 + ], + "angle": 0, + "content": "[12] Xiangning Chen, Chen Liang, Da Huang, Esteban Real, Kaiyuan Wang, Yao Liu, Hieu Pham, Xuanyi Dong, Thang Luong, Cho-Jui Hsieh, Yifeng Lu, and Quoc V. Le. Symbolic discovery of optimization algorithms, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[13] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Radu Soricut. Pali: A jointly-scaled multilingual language-image model. CoRR, abs/2209.06794, 2022. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 4, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[15] Karan Desai, Gaurav Kaul, Zubin Aysola, and Justin Johnson. Redcaps: Web-curated image-text data created by the people, for the people. In Joaquin Vanschoren and SaiKit Yeung, editors, Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks 1, NeurIPS Datasets and Benchmarks 2021, December 2021, virtual, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[16] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Shuyang Gu, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Clip itself is a strong fine-tuner: Achieving \\(85.7\\%\\) and \\(88.0\\%\\) top-1 accuracy with vit-b and vit-l onImagenet. CoRR, abs/2212.06138, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth \\(16 \\times 16\\) words: Transformers for image recognition at scale. In ICLR, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[18] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016. http://www.deeplearningbook.org.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[19] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In CVPR, volume 2, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15979-15988. IEEE, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[21] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. Scaling up vision-language pre-training for image captioning. CoRR, abs/2111.12233, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[22] Gabriel Ilharco, Mitchell Wortsman, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. OpenCLIP. Zenodo, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[24] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In Hugo Larochelle, Marc'Aurelio Ranzato, Raia Hadsell, MariaFlorina Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, and Neil Houlsby." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11984" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Big transfer (BiT): General visual representation learning. In ECCV, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.148 + ], + "angle": 0, + "content": "[26] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, Univ. of Toronto, 2009. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[27] Taku Kudo and John Richardson. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In EMNLP, 2018. 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.191, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors, International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 of Proceedings of Machine Learning Research, pages 12888-12900. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.316, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[29] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. CoRR, abs/2212.00794, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.359, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[30] Matthias Minderer, Alexey A. Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple open-vocabulary object detection. In Shai Avidan, Gabriel J. Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, volume 13670 of Lecture Notes in Computer Science, pages 728-755. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.511, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[31] Jishnu Mukhoti, Tsung-Yu Lin, Amid Poursaeed, Rui Wang, Ashish Shah, Philip H. S. Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.567, + 0.469, + 0.608 + ], + "angle": 0, + "content": "[32] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, 2012. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.609, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[33] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Hanxiao Liu, Adams Wei Yu, Minh-Thang Luong, Mingxing Tan, and Quoc V. Le. Combined scaling for zero-shot transfer learning. CoRR, abs/2111.10050, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.665, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.748, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv e-prints, 2019. 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.804, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[36] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In ICML, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[37] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. CoRR, abs/2210.08402, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.15, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[38] Krishna Srinivasan, Karthik Raman, Jiecao Chen, Michael Bendersky, and Marc Najork. WIT: wikipedia-based image text dataset for multimodal multilingual machine learning. CoRR, abs/2103.01913, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[39] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your ViT? Data, augmentation, and regularization in vision transformers. CoRR, abs/2106.10270, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[40] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. EVA-CLIP: improved training techniques for CLIP at scale. CoRR, abs/2303.15389, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[41] Ashish V. Thapliyal, Jordi Pont-Tuset, Xi Chen, and Radu Soricut. Crossmodal-3600: A massively multilingual multimodal evaluation dataset. In Yoav Goldberg, Zornitsa Kozareva, and Yue Zhang, editors, Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022, pages 715-729. Association for Computational Linguistics, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.501 + ], + "angle": 0, + "content": "[42] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. CoRR, abs/2302.13971, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[43] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, abs/1807.03748, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[45] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. GIT: A generative image-to-text transformer for vision and language. CoRR, abs/2205.14100, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[46] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[47] Ross Wightman, Hugo Touvron, and Hervé Jégou. Resnet strikes back: An improved training procedure in timm. CoRR, abs/2110.00476, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[48] Mitchell Wortsman. Reaching \\(80\\%\\) zero-shot accuracy with OpenCLIP: VIT-G/14 trained on LAION-2B. https://web.archive.org/web/20230127012732/https://laion.ai/blog/giant-openclip/.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[49] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. In IEEE/CVF Conference on Computer" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11985" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 7949-7961. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[50] Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. mT5: A massively multilingual pre-trained text-to-text transformer. In NAACL-HLT, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.274 + ], + "angle": 0, + "content": "[51] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Bin Xiao, Ce Liu, Lu Yuan, and Jianfeng Gao. Unified contrastive learning in image-text-label space. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 19141-19151. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[52] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. CoRR, abs/2205.01917, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.469, + 0.43 + ], + "angle": 0, + "content": "[53] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pengchuan Zhang. Florence: A new foundation model for computer vision. CoRR, abs/2111.11432, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.432, + 0.469, + 0.471 + ], + "angle": 0, + "content": "[54] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. CVPR, 2022. 1, 4, 6, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.474, + 0.469, + 0.569 + ], + "angle": 0, + "content": "[55] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 18102-18112. IEEE, 2022. 1, 2, 3, 4, 6, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.572, + 0.469, + 0.696 + ], + "angle": 0, + "content": "[56] Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D. Manning, and Curtis P. Langlotz. Contrastive learning of medical visual representations from paired images and text. In Zachary C. Lipton, Rajesh Ranganath, Mark P. Sendak, Michael W. Sjoding, and Serena Yeung, editors, Proceedings of the Machine Learning for Healthcare Conference, MLHC 2022, 5-6 August 2022, Durham, NC, USA, volume 182 of Proceedings of Machine Learning Research, pages 2-25. PMLR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11986" + } + ] +] \ No newline at end of file diff --git a/2023/Sigmoid Loss for Language Image Pre-Training/85b123ac-a2cf-49ce-9282-93b24961a80b_origin.pdf b/2023/Sigmoid Loss for Language Image Pre-Training/85b123ac-a2cf-49ce-9282-93b24961a80b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f4834920683daf8be41ac638589c353287971640 --- /dev/null +++ b/2023/Sigmoid Loss for Language Image Pre-Training/85b123ac-a2cf-49ce-9282-93b24961a80b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a8f6c411427aaf06e6e6061207d6b98bd5ec778f7d82768e89029d43951c201 +size 345872 diff --git a/2023/Sigmoid Loss for Language Image Pre-Training/full.md b/2023/Sigmoid Loss for Language Image Pre-Training/full.md new file mode 100644 index 0000000000000000000000000000000000000000..93a75ffab1d5e8946fe0a2c5ad8cf0940e1dd449 --- /dev/null +++ b/2023/Sigmoid Loss for Language Image Pre-Training/full.md @@ -0,0 +1,319 @@ +# Sigmoid Loss for Language Image Pre-Training + +Xiaohua Zhai\* Basil Mustafa Alexander Kolesnikov Lucas Beyer\* Google DeepMind, Zurich, Switzerland + +{xzhai, basilm, akolesnikov, lbeyer}@google.com + +# Abstract + +We propose a simple pairwise sigmoid loss for image-text pre-training. Unlike standard contrastive learning with softmax normalization, the sigmoid loss operates solely on image-text pairs and does not require a global view of the pairwise similarities for normalization. The sigmoid loss simultaneously allows further scaling up the batch size, while also performing better at smaller batch sizes. With only four TPUv4 chips, we can train a Base CLIP model at $4k$ batch size and a Large LiT model at $20k$ batch size, the latter achieves $84.5\%$ ImageNet zero-shot accuracy in two days. This disentanglement of the batch size from the loss further allows us to study the impact of examples vs pairs and negative to positive ratio. Finally, we push the batch size to the extreme, up to one million, and find that the benefits of growing batch size quickly diminish, with a more reasonable batch size of $32k$ being sufficient. We hope our research motivates further explorations in improving the quality and efficiency of language-image pre-training. + +# 1. Introduction + +Contrastive pre-training using weak supervision from image-text pairs found on the web is becoming the go-to method for obtaining generic computer vision backbones, slowly replacing pre-training on large labelled multi-class datasets. The high-level idea is to simultaneously learn an aligned representation space for images and texts using paired data. Seminal works CLIP [34] and ALIGN [23] established the viability of this approach at a large scale, and following their success, many large image-text datasets became available privately [55, 13, 21, 45] and publicly [37, 6, 15, 7, 38]. + +The standard recipe to pre-train such models leverages the image-text contrastive objective. It aligns the image and text embeddings for matching (positive) image-text pairs while making sure that unrelated (negative) image-text pairs + +Table 1: SigLiT and SigLIP results. Sigmoid loss is memory efficient, allows larger batch sizes (BS) that unlocks language image pre-training with a small number of chips. SigLiT model with a frozen public L/16 checkpoint [39], trained on the LiT image-text dataset [55] using four TPU-v4 chips for one day, achieves $79.7\%$ 0-shot accuracy on ImageNet. The same setup with a g/14 checkpoint [54] leads to $84.5\%$ accuracy, trained for two days. With a public unlocked B/16 image checkpoint [39], trained on the WebLI dataset [13], SigLIP achieves $71.0\%$ 0-shot accuracy using 16 TPU-v4 chips for three days. The last two rows show results with randomly initialized models. + +
ImageTextBS#TPUv4DaysINet-0
SigLiTB/8L*32 k4179.7
SigLiTg/14L20 k4284.5
SigLIPB/16B16 k16371.0
SigLIPB/16B32 k32272.1
SigLIPB/16B32 k32573.4
+ +* We use a variant of the $L$ model with 12 layers. + +are dissimilar in the embedding space. This is achieved via a batch-level softmax-based contrastive loss, applied twice to normalize the pairwise similarity scores across all images, then all texts. A naive implementation of the softmax is numerically unstable; it is usually stabilized by subtracting the maximum input value before applying the softmax [18], which requires another pass over the full batch. + +In this paper, we propose a simpler alternative: the sigmoid loss. It does not require any operation across the full batch and hence greatly simplifies the distributed loss implementation and boosts efficiency. Additionally, it conceptually decouples the batch size from the definition of the task. We compare the proposed sigmoid loss with the standard softmax loss across multiple setups. In particular, we investigate sigmoid-based loss with two prominent approaches for image-text learning: CLIP [34] and LiT [55], which we call sigmoid language image pre + +training (SigLIP) and sigmoid LiT (SigLiT), respectively. We find that the sigmoid loss performs significantly better than the softmax loss when the batch size is smaller than $16\mathrm{k}$ . As the train batch size grows, the gap closes. Importantly, the sigmoid loss is symmetric, requires just a single pass, and a typical implementation requires less memory than the softmax loss. This enables successful training of a SigLiT model at a batch size of one million. However, we find that the performance saturates with growing batch size, both for softmax and sigmoid. The good news is that a reasonable batch size, i.e. $32\mathrm{k}$ , is sufficient for image-text pretraining. This conclusion also holds for multilingual SigLIP training on over 100 languages. + +In Table 1, we present setups for image-text pre-training that require a moderate amount of TPUv4 chips for training. SigLiT is surprisingly efficient, reaching $79.7\%$ zero-shot accuracy on ImageNet in just a single day on four chips. SigLIP's more demanding from-scratch training reaches $73.4\%$ zero-shot accuracy in 5 days with 32 TPUv4 chips. This compares favorably to prior works such as FLIP [29] and CLIP [34], which require approximately 5 and 10 days respectively on 256 TPUv3 cores. When fine-tuning a pretrained vision backbone in SigLIP, denoted as $\boxed{ \begin{array}{r l} \end{array} }$ in Table 1, we found that disabling the weight decay on the pre-trained backbone leads to better results (see Figure 4 for details). We hope our work paves the way for making the nascent language-image pre-training field more accessible. + +# 2. Related Work + +Contrastive learning with the sigmoid loss. One prior work proposes a similar sigmoid loss for the task of unsupervised dimensionality reduction [19]; in the scope of contrastive image-text learning, the vast majority of works rely on the softmax-based InfoNCE loss as popularized by [43]. In supervised classification, the sigmoid loss has already been shown to be slightly more effective and robust than the softmax loss [3, 47]. + +Contrastive language-image pre-training has become popular since CLIP [34] and ALIGN [23] applied softmax contrastive learning [56, 43, 10, 24] to large-scale image-text datasets. Both models perform very well on zero-shot transfer tasks, including classification and retrieval. Follow-up works show that contrastively pre-trained models produce good representations for fine-tuning [49, 16], linear regression [23], object detection [30], semantic segmentation [31] and video tasks [53]. + +Generative language-image pre-training Besides softmax contrastive pre-training, various alternatives have been proposed. GIT [45], SimVLM [46], and LEMON [21] successfully pre-train models using a generative text decoder instead, while CoCa [52] adds such a decoder to the discriminative CLIP/ALIGN setup, thus combining the pros + +Algorithm 1 Sigmoid loss pseudo-implementation. +```python +1 # img_emb : image model embedding [n, dim] +2 # txt_emb : text model embedding [n, dim] +3 # t_prime, b : learnable temperature and bias +4 # n : mini-batch size +5 +6 t = exp(t_prime) +7 zimg = l2_normalize(img_emb) +8 ztxt = l2_normalize(txt_emb) +9 logits = dot(zimg, ztxt.T) * t + b +10 labels = 2 * eye(n) - ones(n) # -1 with diagonal 1 +11 l = -sum(log_sigmoid(labels * logits)) / n +``` + +and cons of both approaches into a single very capable model. BLIP [28] further proposes CapFilt which uses the generative decoder to create better captions and the discriminative part of the model to filter pairs. Language-Image pre-training is a very active field and surveys [8] rapidly become outdated. + +Efficient language-image pre-training On the other hand, few works have tried making language image pre-training more efficient. LiT [55] and FLIP [29] are notable attempts, the former requires a pre-trained and locked backbone, and the latter sacrifices quality by randomly dropping visual tokens. BASIC [33] and LAION [48] look at scaling batch-size but only go up to $16\mathrm{k}$ and $160\mathrm{k}$ respectively, by using many hundreds of chips, and for the former also mixing in a large private classification dataset [33, 51]. The recent Lion optimizer [12] claims to be able to reduce the training cost to reach similar quality. + +# 3. Method + +In this section, we first review the widely-used softmax-based contrastive loss. We then introduce the pairwise sigmoid loss and discuss its efficient implementation. + +Given a mini-batch $\mathcal{B} = \{(I_1, T_1), (I_2, T_2), \ldots\}$ of image-text pairs, the contrastive learning objective encourages embeddings of matching pairs $(I_i, T_i)$ to align with each other, while pushing embeddings of unmatched pairs $(I_i, T_{j \neq i})$ apart. For practical purposes, it is assumed that for all images $i$ , the text associated with a different image $j$ is not related to $i$ , and vice-versa. This assumption is usually noisy and imperfect. + +# 3.1. Softmax loss for language image pre-training + +When using the softmax loss to formalize this objective, an image model $f(\cdot)$ and a text model $g(\cdot)$ are trained to minimize the following objective: + +$$ +- \frac {1}{2 | \mathcal {B} |} \sum_ {i = 1} ^ {| \mathcal {B} |} \left(\overbrace {\log \frac {e ^ {t \mathbf {x} _ {i} \cdot \mathbf {y} _ {i}}}{\sum_ {j = 1} ^ {| \mathcal {B} |} e ^ {t \mathbf {x} _ {i} \cdot \mathbf {y} _ {j}}}} ^ {\text {i m a g e} \rightarrow \text {t e x t s o f t m a x}} + \overbrace {\log \frac {e ^ {t \mathbf {x} _ {i} \cdot \mathbf {y} _ {i}}}{\sum_ {j = 1} ^ {| \mathcal {B} |} e ^ {t \mathbf {x} _ {j} \cdot \mathbf {y} _ {i}}}} ^ {\text {t e x t} \rightarrow \text {i m a g e s o f t m a x}}\right) +$$ + +Figure 1: Efficient loss implementation demonstrated via a mock setup with 3 devices and a global batch size of 12. There are no all-gathers, and at any point in time only the bright yellow square (size $4 \times 4$ ) is materialized in memory. +![](images/40b902962f5bea4a83d847167bbf30f869720dcad723b3cfcf9d1f670646a50c.jpg) +(a) Initially each device holds 4 (b) They each compute the comimage and 4 text representations. component of the loss (highlighted) Each device needs to see the repre for their representations, which representations from other devices includes the positives. to calculate the full loss. + +![](images/f9c22c520c1dfb106788a3f98f65c3814546d131b705c71c53e667fdf7d80973.jpg) +(c) Texts are swapped across the devices, so device 1 now has $I_{1:4}$ and $T_{5:8}$ etc. The new loss is computed and accumulated with the previous. + +![](images/fbc90ed975367f418d02645cc1e6163dc3f70c8308d1fad6aede1cbd706cbedd.jpg) +(d) This repeats till every image & text pair have interacted, e.g. device 1 has the loss of $I_{1:4}$ and $T_{1:12}$ . A final cross-device sum brings everything together. + +![](images/7e79e3050f7c6ac6fdadeea3b3b6a16f4aba788a667d2823cd80049989fa559c.jpg) + +where $\mathbf{x}_i = \frac{f(I_i)}{\|f(I_i)\|_2}$ and $\mathbf{y}_i = \frac{g(T_i)}{\|g(T_i)\|_2}$ . In this paper, we adopt the vision transformer architecture [17] for images and the transformer architecture [44] for texts. Note that due to the asymmetry of the softmax loss, the normalization is independently performed two times: across images and across texts [34]. The scalar $t$ is parametrized as $\exp(t')$ , where $t'$ is a global freely learnable parameter. + +# 3.2. Sigmoid loss for language image pre-training + +Instead of the softmax-based contrastive loss, we propose a simpler alternative that does not require computing global normalization factors. The sigmoid-based loss processes every image-text pair independently, effectively turning the learning problem into the standard binary classification on the dataset of all pair combinations, with a positive labels for the matching pairs $(I_i, T_i)$ and negative labels for all other pairs $(I_i, T_{j \neq i})$ . It is defined as follows: + +$$ +- \frac {1}{| \mathcal {B} |} \sum_ {i = 1} ^ {| \mathcal {B} |} \sum_ {j = 1} ^ {| \mathcal {B} |} \underbrace {\log \frac {1}{1 + e ^ {z _ {i j} (- t \mathbf {x} _ {i} \cdot \mathbf {y} _ {j} + b)}}} _ {\mathcal {L} _ {i j}} +$$ + +where $z_{ij}$ is the label for a given image and text input, which equals 1 if they are paired and -1 otherwise. Note that at initialization, the heavy imbalance coming from the many negatives dominates the loss, leading to large initial optimization steps attempting to correct this bias. To alleviate this, we introduce an additional learnable bias term $b$ similar to the temperature $t$ . We initialize $t'$ and $b$ to 10 and -10 respectively. This makes sure the training starts roughly close + +to the prior and does not require massive over-correction. Algorithm 1 presents a pseudocode implementation of the proposed sigmoid loss for language image pre-training. + +# 3.3. Efficient "chunked" implementation + +Contrastive training typically utilizes data parallelism. Computing the loss when data is split across $D$ devices necessitates gathering all embeddings [55] with expensive all-gathers and, more importantly, the materialization of a memory-intensive $|\mathcal{B}| \times |\mathcal{B}|$ matrix of pairwise similarities. + +The sigmoid loss, however, is particularly amenable to a memory efficient, fast, and numerically stable implementation that ameliorates both these issues. Denoting the perdevice batch size as $b = \frac{|\mathcal{B}|}{D}$ , the loss is reformulated as: + +![](images/970912ada39eeb482401c9a2af5950c7c2af4acfff9bd0018fd6a049194b920b.jpg) + +This is particularly simple for the sigmoid loss as each pair is an independent term in the loss. Figure 1 illustrates this method. In words, we first compute the component of the loss corresponding to the positive pairs, and $b - 1$ negative pairs. We then permute representations across devices, so each device takes negatives from its neighbouring device (next iteration of sum B). The loss is then calculated with respect to this chunk (sum C). This is done independently in each device, such that each device computes the + +![](images/b6a0eca64575fcaa25a7711def362e40f517d55a837a8473122c285cc177fbbd.jpg) +Figure 2: The effect of pre-training batch size. Left: SigLiT results, trained for 18B seen examples. Sigmoid loss outperforms the softmax loss significantly with small batch sizes, and performs similarly at larger batch sizes. We successfully trained an SigLiT model with up to one million batch size. However, performance for both sigmoid and softmax saturate at around $32\mathrm{k}$ batch size. Middle: SigLIP results, trained for 9B seen examples. Both sigmoid loss and softmax loss saturate at a reasonable batch size, while the peak of the sigmoid loss comes earlier and slightly outperforms the peak of the softmax loss. A very large batch size hurts both losses. Right: mSigLIP results, trained for 30B seen examples. With a multilingual setup using over 100 languages, $32\mathrm{k}$ batch size is surprisingly sufficient and scaling beyond that hurts performance on a 36-language cross-modal retrieval task. + +![](images/d7d5f072cdf4d3412cacf41ede119e784e91cde95ebd583f4570da79f9548186.jpg) + +![](images/9d01164e07d4eef84e6f33fa78d4eb81c626ee64cd20ac0d5ab0924f0fc47b48.jpg) + +loss with respect to its local batch $b$ . Losses can then simply be summed across all devices (sum A). Individual collective permutes (for sum B) are fast (and indeed $D$ collective permutes is typically faster than two all-gathers between $D$ devices), and the memory cost at any given moment is reduced from $|\mathcal{B}|^2$ to $b^2$ (for sum C). Usually $b$ is constant as scaling $|\mathcal{B}|$ is achieved by increasing the number of accelerators. Due to being quadratic with respect to the batch size, the vanilla loss computation rapidly bottlenecks scaling up. This chunked approach enabled training with batch sizes over 1 million on relatively few devices. + +# 4. Results + +In this section, we evaluate the proposed SigLiT and SigLIP models across a wide range of batch sizes. We discuss what can be achieved with a small number of accelerator chips, using both SigLiT and SigLIP recipes. We also briefly discuss the impact of batch size on multilingual language image pre-training. We ablate the importance of our large-batch stabilization modification and the introduced learned bias term and present a study on the effect of positive and negative pairs ratio in the sigmoid loss. Lastly, we explore SigLIP's data noise robustness. + +To validate our models, we report zero-shot transfer results on the ImageNet dataset [14] and zero-shot retrieval results across 36 languages on the XM3600 dataset [41]. We use the ScalingViT-Adafactor optimizer [54] by default for all our experiments. + +# 4.1. SigLiT: Scaling batch size to the limit + +Following [55], we use the same precomputed embeddings for the images using a ViT-g vision model, and train a base size text tower from scratch with the same hyperparameters using the LiT image-text dataset [55]. + +We perform a study over a wide range of batch sizes, from 512 to $1M$ , demonstrating the impact of batch size for contrastive learning. Results are presented in Figure 2 (left). When the batch size is smaller than $16k$ , sigmoid loss outperforms softmax loss by a large margin. With growing batch sizes, we observe that softmax loss quickly catches up and potentially slightly underperforms sigmoid loss with a large enough batch size. Overall, we recommend using the SigLIP recipe for large batch sizes as well, due to the simplicity, compute savings, and straightforward memory efficient implementation. + +There is a consensus that contrastive learning benefits from large batch sizes, while most of the existing studies stop at $64\mathrm{k}$ batch size [55, 33, 10]. We successfully trained an SigLiT model at one million batch size, to explore the limit of contrastive learning. To our surprise, the performance saturates at $32\mathrm{k}$ batch size, further scaling up the batch size only gives a minor boost, and the model peaks at $256\mathrm{k}$ batch size. Our best SigLiT with a $B$ -sized text mode achieves $84.7\%$ zero-shot transfer accuracy on ImageNet, while the original LiT paper reports a slightly better $85.2\%$ score with a 10 times larger $g$ -sized text model. Figure 3 presents the impact of training duration for different batch sizes. It demonstrates that large, $262k$ batch size signifi + +![](images/8e77149cddcaaa1e9ded9c72e0d68c055f9f4e30b49d6e13288a4ce900acd2da.jpg) +Figure 3: SigLiT ImageNet 0-shot transfer results with different training durations. Large batch size results in a big performance boost, but needs a sufficiently long schedule to ramp up, as for short schedules, very large batch size results in a small number of gradient update steps. + +cantly outperforms smaller $8k$ batch size when trained for a sufficiently long time. Note, that for short training durations, large batch size leads to the fewer absolute number of update steps and thus needs more time to ramp up. + +# 4.2. SigLIP: Sigmoid loss is beneficial for language-image pre-training + +We pre-train SigLIP models on the WebLI dataset [13], using only English image and text pairs. We use moderately-sized models: B/16 ViT for image embeddings and B-sized transformer for text embeddings. The input images are resized to $224 \times 224$ resolution. The text is tokenized by a $32\mathrm{k}$ vocabulary sentencepiece tokenizer [27] trained on the English C4 dataset [35], and a maximum of 16 text tokens are kept. Figure 2 middle plot shows SigLIP results, With less than $32\mathrm{k}$ batch size, SigLIP outperforms CLIP baselines with the standard softmax loss. On the other end of the scale, the memory efficiency of the sigmoid loss enabled much larger batch sizes. For example, with four TPU-v4 chips, we could fit a batch size of 4096 with a Base SigLIP but only 2048 with a corresponding CLIP model. The two advantages together demonstrate significant benefits of the sigmoid loss for language image pre-training with fixed resources, which will be discussed in Section 4.5. + +As batch size increases, the gap between the sigmoid and the softmax losses diminish. SigLIP performs best at batch size $32\mathrm{k}$ , whereas the softmax loss required $98\mathrm{k}$ for optimal performance and still didn't outperform the sigmoid based variant. Scaling further, a larger batch size like $307\mathrm{k}$ hurts both losses. + +
16 k32 k64 k128 k240 k
INet-071.673.273.273.273.1
XM avg34.834.934.433.632.7
XM de54.754.855.454.354.7
XM en46.546.246.546.646.6
XM hi9.18.57.98.17.3
XM ru50.149.949.748.649.3
XM zh30.732.532.030.623.7
+ +Table 2: Multilingual SigLIP results with various batch sizes, pre-trained for 30 billion seen examples. We report zero-shot transfer results on ImageNet (INet-0) and averaged text to image retrieval results across 36 languages on the crossmodal 3600 dataset (XM). The full table on 36 languages can be found in Appendix. + +# 4.3. mSigLIP: Multi-lingual pre-training + +We further scale up the training data by keeping all the 100 languages from the WebLI dataset [13]. With multilingual data, one usually needs to use a larger international vocabulary. We first verify the impact of two tokenizers: a small multilingual vocabulary with $32\mathrm{k}$ tokens [35], and a large multilingual vocabulary with $250\mathrm{k}$ tokens [50]. We train B-sized ViT and text models for $900M$ total examples seen, and observe slightly more than $1\%$ improvement when using a larger vocabulary. + +However, the token embeddings become huge for very large vocabulary sizes. Following the standard setup, we would need to store a $N \times W$ token embedding lookup table to train the multilingual model, where $N$ is the vocabulary size mentioned above and $W$ is the embedding dimension of the text model. To save memory, we propose to use a "bottlenecked" token embedding. We use $N \times K$ embedding matrix and additional $K \times W$ projection, where the bottleneck $K$ is much smaller than $W$ . + +In our experiments, we observed that using a large multilingual vocabulary with a bottleneck can be scaled up as efficiently as using a small multilingual vocabulary. Specifically, by enabling the bottleneck of size $K = 96$ for Base architecture with $W = 768$ , we only see about a half percent quality drop on ImageNet zero-shot transfer, compared to using the full $250k$ vocabulary. + +With the memory improvements, we train mSigLIP models for various batch sizes, for a total of 30 billion examples seen. Table 2 and Figure 2 (right plot) show the results. We were expecting a large batch size to improve multilingual pre-training, where the model sees more examples from the same language as hard negatives in a sin + +![](images/a8d6942afea6fe1bd82be48abdd04d2c8e997bafd7cf91bed102f0dc5b56595e.jpg) +Figure 4: Top: SigLIP with pre-trained encoders ramps up quickly. However, only disabling weight decay on the pre-trained encoder weights leads to stable behavior and good ImageNet 0-shot transfer results. Bottom: ImageNet 10-shot transfer results, where decaying the pre-trained weights leads to deterioration of the pre-trained model visual representation quality. Disabling weight decay makes the curve flatter. + +gle mini-batch. However, we didn't observe clear improvements with a batch size larger than $32\mathrm{k}$ . A batch size of $32\mathrm{k}$ is sufficient for a multilingual setup as well. On the XM3600 cross-modal retrieval tasks, we found that going beyond $32\mathrm{k}$ batch size leads to worse results on average while on ImageNet zero-shot transfer it stays flat. mSigLIP sets the new state-of-the-art on XM3600 text to image retrieval task, with only a Base size model. Our best result is $34.9\%$ , which is more than $6\%$ higher than the previously reported result $28.5\%$ [13] with a standard LiT model [55] using a much larger four billion ViT-e model. + +# 4.4. SigLiT with four TPU-v4 chips + +For many practitioners, the important question usually is "what can be trained with a limited amount of resources?" We explore the usage of SigLiT models in this section with only four TPU-v4 chips, as the memory efficient sigmoid loss is suitable for this application scenario. + +We follow the same setup as in section 4.1. We use the publicly available ViT-Augreg-B/8 [39] model as the frozen ( ) vision tower, and precompute embeddings to accelerate the training [55]. The text model is a Large Transformer, but with a depth of only 12 layers (instead of 24). It is trained using the LION [12] optimizer with decoupled weight de + +![](images/e87c68b11bc442c003f85424b4a8084854d3ccc5d0a5587f463252ad4ad1e0a4.jpg) +Figure 5: The effect of Adam and AdaFactor's $\beta_{2}$ . As we increase batch-size, we observe more frequent training instability. This instability can mainly be seen in the loss curves (top) and is caused by spikes in the gradient norm (middle) which results in large parameter updates (bottom). Decreasing the $\beta_{2}$ momentum value stabilizes the training. Even though occasional gradient spikes still happen (see step at 2B), they do not destabilize the training process. + +cay $1 \times 10^{-7}$ , linearly warm-up of learning rate over $6.5\mathrm{k}$ steps up to a peak of $1 \times 10^{-4}$ , followed by a cosine decay to 0. We train for a total of 65000 steps with a batch size of $32\mathrm{k}$ - this leads to just under one day of training. Table 1 shows the results when training a model on four chips for one day, achieving $79.7\%$ 0-shot ImageNet classification accuracy; very competitive in this limited resource regime. With a ViT-g/14 [54] model as the vision tower and a Large text tower, we can train at $20\mathrm{k}$ batch size on four chips for $107\mathrm{k}$ steps in under two days. This further pushes the 0-shot ImageNet classification accuracy up to $84.5\%$ . + +# 4.5. SigLIP with a small amount of TPU-v4 chips + +It's resource demanding to train a CLIP model from scratch in general, with SigLIP it's possible to fit a larger train batch size with fewer amount of chips. In this section, we explore ways to train SigLIP models efficiently with pretrained weights. We use pre-trained weights to initialize the image model to accelerate the pre-training, which was originally discussed in [55]. We use the public and unlocked ViT-Augreg-B/16 [39] model to initialize our vision tower and fine-tune on the same WebLI English data as used for SigLIP. In all the experiments, we apply a 0.1 learning rate multiplier to the pre-trained image tower to make it suitable + +![](images/98d1b8aa5b072bfb683bf8d0ab866d58f176ca50affa4e4cdf43d7f9209e646c.jpg) +Figure 6: The effect of batch composition. We simulate various batch compositions by masking out negatives, either randomly, keeping only the hardest, or the easiest. With no masking, we have $16\mathrm{k}$ negatives for each positive in the batch $(1:16\mathrm{k})$ and the strongest masking we apply (1:1.6) results in almost balanced minibatches. In one setting we match total pairs seen by training for significantly longer. We observe ImageNet 0-shot score, the final value of the learned bias, and the average logits of positive and negative pairs. Overall, the imbalance does not seem to be detrimental, but finding an efficient way of mining negatives might be beneficial. + +![](images/84693baafeb673a4a06b73173a7d466c1058f5b6d800fd5ddd2c57e16b5d6884.jpg) + +![](images/b3ac5c2d1fe1add17df1d7316c8d4529441fe14af09f3e7d9fe151ceda2b1e0c.jpg) + +for fine-tuning. + +Figure 4 presents unlocked fine-tuning results alongside from-scratch randomly initialized baselines. We used 16 TPU-v4 chips and train at $16\mathrm{k}$ batch size for $2.4\mathrm{B}$ examples seen. We found that the fine-tuning setup doesn't perform well out-of-the-box; this is consistent with prior works [55] where finetuning image models degraded visual representation quality. This is evidenced by ImageNet 10-shot linear classification, where in Figure 4 the fine-tuned setup is barely better than the from-scratch baseline. + +We hypothesize that the default weight decay applied to the pre-trained weights reduces their effectiveness. Motivated by the fine-tuning recipe from [17, 54, 25], that uses no weight decay, we also propose disabling weight decay on the pre-trained weights for SigLIP training. Weight decay is therefore only applied to the randomly initialized weights in the text model. This simple modification significantly improved SigLIP results. Figure 4 shows that with our improved recipe, SigLIP reaches $71\%$ 0-shot accuracy on ImageNet, using $16k$ batch size, trained on 16 chips for three days. We also present from-scratch results in the bottom rows of Table 1: with 32 TPUv4 chips for only two days, SigLIP achieves $72.1\%$ 0-shot accuracy. This presents a significant training cost reduction e.g. compared to CLIP (approx. 2500 TPUv3-days for $72.6\%$ ) reported in [29]. + +# 4.6. Scaling up SigLIP + +In this section, we scale up SigLIP by "overtraining" the model [42, 1] for a long schedule. We present results in Table 3 using ViT-B or ViT-L as the vision encoder, while the text encoder is kept the same size (B or L respec + +tively). Following the recipe as described in Section 4.2, we train both models for 40 billion examples seen, but increase the number of image patches from $(224 / 16)^{2} = 196$ to $(256 / 16)^{2} = 256$ and the number of text tokens from 16 to 64. To get SigLIP results for different resolutions, we continue training the pre-trained model for 5 billion examples on the target resolutions (i.e. 224 or 384), using a $100\mathrm{x}$ smaller learning rate without weight decay. In Table 3, we report zero-shot classification results on ImageNet [14], ObjectNet [2], ImageNetv2 [36], ImageNet Real [3], and zero-shot image-to-text $(\mathrm{I}\rightarrow \mathrm{T})$ retrieval, text-to-image $(\mathrm{I}\rightarrow \mathrm{T})$ retrieval results on MSCOCO [11]. Our SigLIP models outperform the previous models consistently with ViT-B and ViT-L architectures. + +# 4.7. Stabilizing large-batch training + +As we move to large batch sizes, the language image pretraining using transformers becomes increasingly more unstable, even when using a modestly-sized model (e.g. Base size). The reason for these instabilities is large spikes in the gradient norms, which translate to large-magnitude changes in the weights that may destabilize the training process, see Figure 5. We observe that reducing $\beta_{2}$ in Adam and AdaFactor from its default 0.999 to 0.95 (which was suggested in [20, 9]) is enough to stabilize the training. Intuitively, this allows recovering from gradient spikes quicker. We opt for setting $\beta_{2} = 0.95$ for all our experiments. + +# 4.8. Negative ratio in sigmoid loss + +One question which arises when shifting the perspective from the softmax's "pick the right class" view to the sig- + +
MethodImage EncoderImageNet-1kCOCO
SizePatch #Validationv2RealObjectNetI→TT→I
CLIPB19668.361.9-55.352.433.1
OpenCLIPB19670.262.3-56.059.442.3
EVA-CLIPB19674.767.0-62.358.742.2
SigLIPB19676.369.682.870.764.447.2
SigLIPB25676.670.083.171.365.147.4
CLIPL25675.569.0-69.956.336.5
OpenCLIPL25674.061.1-66.462.146.1
EVA-CLIPL25679.875.3-72.963.747.5
SigLIPL25680.674.285.977.969.551.1
CLIPL57676.672.0-70.957.937.1
EVA-CLIPL57680.473.8-78.464.147.9
+ +Table 3: Comparison with other methods. Our SigLIP models outperform the previous models, e.g. OpenCLIP [22] and CLIP [34], by a significant margin on both the zero-shot classification tasks and the zero-shot retrieval tasks. Compared to the recently introduced EVA-CLIP [40] model, SigLIP-L outperforms EVA-CLIP-L on all the tasks with the same number of 256 patches. SigLIP-L (256 patches) also outperforms EVA-CLIP-L (576 patches) by a large margin on the retrieval tasks. + +moid's "rate this pair" view, is the imbalance in positive versus negative pairs. For a batch size $|\mathcal{B}|$ , the batch contains $|\mathcal{B}|$ positive pairs, but $|\mathcal{B}|^2 - |\mathcal{B}|$ negative examples. In the modest batch-size of $16\mathrm{k}$ , there are actually $268\mathrm{M}$ negative examples for only $16\mathrm{k}$ positive ones. At the same time, because the sigmoid loss decomposes into a sum of per-example losses, we can perform controlled experiments to study the effect of the mini-batch composition and distribution of examples visited. We run experiments in the SigLiT setup at batch-size $16\mathrm{k}$ for $900\mathrm{M}$ steps and vary the composition of the batch by masking out (i.e. ignoring) enough negative examples to reach a target "positive: negative" ratio, masking in the following ways: + +- Random: Randomly choose negative pairs to mask. +- Hard: Keep hardest negative pairs (highest loss). +- Easy: Keep easiest negatives pairs (lowest loss). +- Hard + matching total pairs seen: Masking examples while training for a fixed number of steps does decrease the total number of pairs seen during training. Hence in the matched pairs setting, we increase the number of training steps by the masking ratio in order to keep the number of pairs seen constant. + +Figure 6 shows the effect of the various masking strategies. Randomly removing negatives to rebalance does deteriorate performance. Keeping the easiest examples does not work at all, while keeping the hardest negatives does almost maintain the quality, indicating that, as could be expected, a lot of the learning on the negative side comes from the harder examples. This is further confirmed by the slightly + +increased performance of training longer on the hardest examples in order to match the total pairs seen. + +We also look at the value of the learned bias at the end of training as well as the average logit value for positive and negative examples across these settings, and find the result mostly follows what one would expect: as fewer negatives are present, the bias and logits become more positive overall. Interestingly, when training with more hard negative pairs, the average logits of positive pairs stays mostly flat. + +This study confirms that (1) the imbalance does not seem to be a major reason for concern, while at the same time (2) coming up with an efficient way of including more negative examples can be promising but is not trivial. + +Table 4: Bias (b) and temperature $(\mathbf{t}^{\prime})$ initialization. Results are reported using Base architecture, $8\mathrm{k}$ batch size, trained for 900M examples. Enabling the bias term b with -10 initialization improves results consistently. + +
bt'INet-0Pet-0C100-0
n/alog 1062.081.859.9
-10log 1063.082.461.0
-10log 161.080.060.4
0log 1061.779.959.0
0log 153.773.253.8
+ +![](images/f3971470a9b627315836ee46a6b56f5162dba6ad57ca79f020f21438ca925d6f.jpg) +Figure 7: Sigmoid-training increases robustness to data noise. Titles show the type of corruption applied, and x-axes show the probability with which they are applied. With increasing corruption severity, M-scale models trained with sigmoid loss for 3.6 billion examples retain superiority over corresponding softmax baseline. + +# 4.9. Bias term in sigmoid loss + +We ablate the bias term in the loss function, using the Base architecture with an $8\mathrm{k}$ batch size, trained for 900M examples with the SigLIP setup. Zero-shot transfer results are reported on ImageNet [14], Oxford-iit pet [32] and Cifar100 [26]. Table 4 presents results with and without a bias term in the sigmoid loss. + +Enabling the bias term with a $-10$ initialization consistently improves performance across all tasks. This is because the bias term ensures that the training starts close to the prior, preventing dramatic over-correction in early optimization. In contrast, a randomly chosen bias term initialization, such as the 0 initialization in Table 4, fails to address the over-correction issue, leading to significantly worse results. This effect is particularly noticeable when using a small temperature $t^{\prime}$ initialization. We set the bias and temperature initialization to $b = -10$ and $t^\prime = \log 10$ (hence $t = 10$ ) as the default for all experiments. + +# 4.10. Label noise robustness + +Prior works demonstrated improved robustness against label noise when using the sigmoid loss for classification models [3]. This property would be particularly useful here in the face of the famously noisy nature of popular large-scale image-text datasets. In order to study this for SigLIP, we train M/16 image models alongside an M text model at batch size 16384 for 3.6 billion seen examples. We corrupt the training data using one of the following methods: + +- Image: With probability $p$ , replace the image with uniform random noise. +- Text: With probability $p$ , replace tokenized text with a new sequence of randomly sampled tokens, up to some (sampled) sequence length. +- Batch alignment: Randomly shuffle the ordering of $p\%$ of the batch. +- Image & text: Apply both with probability $p$ each. + +- Image, text & batch: Alongside (4), also shuffle fraction $p$ of alignments. + +Results from varying the likelihood of the corruption are shown in Figure 7. Models trained with sigmoid loss are increasingly robust to all kinds of added noise. + +# 5. Conclusion + +We conducted a study on two language-image pretraining instances that used the sigmoid loss: SigLiT and SigLIP. Our results demonstrate that the sigmoid loss performs better than the softmax baseline, particularly for small train batch sizes. This loss function is also more memory efficient, which allows larger train batch sizes without requiring additional resources. We performed a thorough investigation of the batch size in contrastive learning. Surprisingly, we found that a relatively modest batch size of $32\mathrm{k}$ yielded nearly optimal performance. Further studies have been performed to understand better the introduced bias term in the sigmoid loss, robustness to data noises and the impact of positive and negative pairs ratio in the sigmoid loss. We hope this work will facilitate language-image pretraining research with limited resources. + +Acknowledgements. We thank Daniel Keysers, Ilya Tolstikhin, Olivier Bousquet and Michael Tschannen for their valuable feedback and discussions on this paper. We thank Joan Puigcerver, Josip Djolonga and Black Hechtman for discussions on efficient implementations of the chunked contrastive loss. We thank Kaiming He and Xinlei Chen for the discussion of $\beta_{2}$ to stabilize the training. We also thank Ross Wightman for spotting a mistake in the pseudocode in the first version of this paper. Similarly, we thank Boris Dayma for spotting a typo making $t$ vs $t'$ confusing, which we fixed in the third version of this paper. We thank the Google Deepmind team for providing a supportive research environment. We use the big-vision codebase [5, 4] for all experiments in this project. + +# References + +[1] Ibrahim Alabdulmohsin, Xiaohua Zhai, Alexander Kolesnikov, and Lucas Beyer. Getting vit in shape: Scaling laws for compute-optimal model design. In NeurIPS, 2023. 7 +[2] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. ObjectNet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. In NeurIPS, 2019. 7 +[3] Lucas Beyer, Olivier J. Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with imagenet? CoRR, abs/2006.07159, 2020. 2, 7, 9 +[4] Lucas Beyer, Xiaohua Zhai, and Alexander Kolesnikov. Better plain vit baselines for imagenet-1k, 2022. 9 +[5] Lucas Beyer, Xiaohua Zhai, and Alexander Kolesnikov. Big vision. https://github.com/google-research/big_vision, 2022.9 +[6] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.1 +[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021. 1 +[8] Feilong Chen, Duzhen Zhang, Minglun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. VLP: A survey on vision-language pre-training. Int. J. Autom. Comput., 20(1):38-56, 2023. 2 +[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 1691-1703. PMLR, 2020. 7 +[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2, 4 +[11] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. CoRR, abs/1504.00325, 2015. 7 +[12] Xiangning Chen, Chen Liang, Da Huang, Esteban Real, Kaiyuan Wang, Yao Liu, Hieu Pham, Xuanyi Dong, Thang Luong, Cho-Jui Hsieh, Yifeng Lu, and Quoc V. Le. Symbolic discovery of optimization algorithms, 2023. 2, 6 +[13] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and + +Radu Soricut. Pali: A jointly-scaled multilingual language-image model. CoRR, abs/2209.06794, 2022. 1, 5, 6 +[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 4, 7, 9 +[15] Karan Desai, Gaurav Kaul, Zubin Aysola, and Justin Johnson. Redcaps: Web-curated image-text data created by the people, for the people. In Joaquin Vanschoren and SaiKit Yeung, editors, Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks 1, NeurIPS Datasets and Benchmarks 2021, December 2021, virtual, 2021. 1 +[16] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Shuyang Gu, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Clip itself is a strong fine-tuner: Achieving $85.7\%$ and $88.0\%$ top-1 accuracy with vit-b and vit-l onImagenet. CoRR, abs/2212.06138, 2022. 2 +[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth $16 \times 16$ words: Transformers for image recognition at scale. In ICLR, 2021. 3, 7 +[18] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016. http://www.deeplearningbook.org.1 +[19] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In CVPR, volume 2, 2006. 2 +[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15979-15988. IEEE, 2022. 7 +[21] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. Scaling up vision-language pre-training for image captioning. CoRR, abs/2111.12233, 2021. 1, 2 +[22] Gabriel Ilharco, Mitchell Wortsman, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. OpenCLIP. Zenodo, 2021. 8 +[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1, 2 +[24] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In Hugo Larochelle, Marc'Aurelio Ranzato, Raia Hadsell, MariaFlorina Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2 +[25] Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, and Neil Houlsby. + +Big transfer (BiT): General visual representation learning. In ECCV, 2020. 7 +[26] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, Univ. of Toronto, 2009. 9 +[27] Taku Kudo and John Richardson. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In EMNLP, 2018. 5, 13 +[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors, International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 of Proceedings of Machine Learning Research, pages 12888-12900. PMLR, 2022. 2 +[29] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. CoRR, abs/2212.00794, 2022. 2, 7 +[30] Matthias Minderer, Alexey A. Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple open-vocabulary object detection. In Shai Avidan, Gabriel J. Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, volume 13670 of Lecture Notes in Computer Science, pages 728-755. Springer, 2022. 2 +[31] Jishnu Mukhoti, Tsung-Yu Lin, Amid Poursaeed, Rui Wang, Ashish Shah, Philip H. S. Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning, 2022. 2 +[32] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, 2012. 9 +[33] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Hanxiao Liu, Adams Wei Yu, Minh-Thang Luong, Mingxing Tan, and Quoc V. Le. Combined scaling for zero-shot transfer learning. CoRR, abs/2111.10050, 2021. 2, 4 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 8 +[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv e-prints, 2019. 5, 13 +[36] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In ICML, 2019. 7 +[37] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine + +Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. CoRR, abs/2210.08402, 2022. 1 +[38] Krishna Srinivasan, Karthik Raman, Jiecao Chen, Michael Bendersky, and Marc Najork. WIT: wikipedia-based image text dataset for multimodal multilingual machine learning. CoRR, abs/2103.01913, 2021. 1 +[39] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your ViT? Data, augmentation, and regularization in vision transformers. CoRR, abs/2106.10270, 2021. 1, 6 +[40] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. EVA-CLIP: improved training techniques for CLIP at scale. CoRR, abs/2303.15389, 2023. 8 +[41] Ashish V. Thapliyal, Jordi Pont-Tuset, Xi Chen, and Radu Soricut. Crossmodal-3600: A massively multilingual multimodal evaluation dataset. In Yoav Goldberg, Zornitsa Kozareva, and Yue Zhang, editors, Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022, pages 715-729. Association for Computational Linguistics, 2022. 4 +[42] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. CoRR, abs/2302.13971, 2023. 7 +[43] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, abs/1807.03748, 2018. 2 +[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3 +[45] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. GIT: A generative image-to-text transformer for vision and language. CoRR, abs/2205.14100, 2022. 1, 2 +[46] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 2 +[47] Ross Wightman, Hugo Touvron, and Hervé Jégou. Resnet strikes back: An improved training procedure in timm. CoRR, abs/2110.00476, 2021. 2 +[48] Mitchell Wortsman. Reaching $80\%$ zero-shot accuracy with OpenCLIP: VIT-G/14 trained on LAION-2B. https://web.archive.org/web/20230127012732/https://laion.ai/blog/giant-openclip/.2 +[49] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. In IEEE/CVF Conference on Computer + +Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 7949-7961. IEEE, 2022. 2 +[50] Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. mT5: A massively multilingual pre-trained text-to-text transformer. In NAACL-HLT, 2021. 5 +[51] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Bin Xiao, Ce Liu, Lu Yuan, and Jianfeng Gao. Unified contrastive learning in image-text-label space. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 19141-19151. IEEE, 2022. 2 +[52] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. CoRR, abs/2205.01917, 2022. 2 +[53] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pengchuan Zhang. Florence: A new foundation model for computer vision. CoRR, abs/2111.11432, 2021. 2 +[54] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. CVPR, 2022. 1, 4, 6, 7, 13 +[55] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 18102-18112. IEEE, 2022. 1, 2, 3, 4, 6, 7, 13 +[56] Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D. Manning, and Curtis P. Langlotz. Contrastive learning of medical visual representations from paired images and text. In Zachary C. Lipton, Rajesh Ranganath, Mark P. Sendak, Michael W. Sjoding, and Serena Yeung, editors, Proceedings of the Machine Learning for Healthcare Conference, MLHC 2022, 5-6 August 2022, Durham, NC, USA, volume 182 of Proceedings of Machine Learning Research, pages 2-25. PMLR, 2022. 2 \ No newline at end of file diff --git a/2023/Sigmoid Loss for Language Image Pre-Training/images.zip b/2023/Sigmoid Loss for Language Image Pre-Training/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e451d9d073a66877779c8a9753a505ef97032ae4 --- /dev/null +++ b/2023/Sigmoid Loss for Language Image Pre-Training/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cda96e896f0dfda6667a66d9de49cecc089c7213e406f1c27e112b9a51f45a86 +size 505805 diff --git a/2023/Sigmoid Loss for Language Image Pre-Training/layout.json b/2023/Sigmoid Loss for Language Image Pre-Training/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2749ca708173f1880c18de3d85379115fc6ed72b --- /dev/null +++ b/2023/Sigmoid Loss for Language Image Pre-Training/layout.json @@ -0,0 +1,10063 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 103, + 445, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 103, + 445, + 121 + ], + "spans": [ + { + "bbox": [ + 149, + 103, + 445, + 121 + ], + "type": "text", + "content": "Sigmoid Loss for Language Image Pre-Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 143, + 465, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 143, + 465, + 171 + ], + "spans": [ + { + "bbox": [ + 126, + 143, + 465, + 171 + ], + "type": "text", + "content": "Xiaohua Zhai\\* Basil Mustafa Alexander Kolesnikov Lucas Beyer\\* Google DeepMind, Zurich, Switzerland" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 169, + 174, + 424, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 174, + 424, + 185 + ], + "spans": [ + { + "bbox": [ + 169, + 174, + 424, + 185 + ], + "type": "text", + "content": "{xzhai, basilm, akolesnikov, lbeyer}@google.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 224, + 192, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 224, + 192, + 238 + ], + "spans": [ + { + "bbox": [ + 143, + 224, + 192, + 238 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 251, + 289, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 289, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 289, + 468 + ], + "type": "text", + "content": "We propose a simple pairwise sigmoid loss for image-text pre-training. Unlike standard contrastive learning with softmax normalization, the sigmoid loss operates solely on image-text pairs and does not require a global view of the pairwise similarities for normalization. The sigmoid loss simultaneously allows further scaling up the batch size, while also performing better at smaller batch sizes. With only four TPUv4 chips, we can train a Base CLIP model at " + }, + { + "bbox": [ + 46, + 251, + 289, + 468 + ], + "type": "inline_equation", + "content": "4k" + }, + { + "bbox": [ + 46, + 251, + 289, + 468 + ], + "type": "text", + "content": " batch size and a Large LiT model at " + }, + { + "bbox": [ + 46, + 251, + 289, + 468 + ], + "type": "inline_equation", + "content": "20k" + }, + { + "bbox": [ + 46, + 251, + 289, + 468 + ], + "type": "text", + "content": " batch size, the latter achieves " + }, + { + "bbox": [ + 46, + 251, + 289, + 468 + ], + "type": "inline_equation", + "content": "84.5\\%" + }, + { + "bbox": [ + 46, + 251, + 289, + 468 + ], + "type": "text", + "content": " ImageNet zero-shot accuracy in two days. This disentanglement of the batch size from the loss further allows us to study the impact of examples vs pairs and negative to positive ratio. Finally, we push the batch size to the extreme, up to one million, and find that the benefits of growing batch size quickly diminish, with a more reasonable batch size of " + }, + { + "bbox": [ + 46, + 251, + 289, + 468 + ], + "type": "inline_equation", + "content": "32k" + }, + { + "bbox": [ + 46, + 251, + 289, + 468 + ], + "type": "text", + "content": " being sufficient. We hope our research motivates further explorations in improving the quality and efficiency of language-image pre-training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 491, + 128, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 128, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 128, + 503 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 512, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 512, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 512, + 287, + 643 + ], + "type": "text", + "content": "Contrastive pre-training using weak supervision from image-text pairs found on the web is becoming the go-to method for obtaining generic computer vision backbones, slowly replacing pre-training on large labelled multi-class datasets. The high-level idea is to simultaneously learn an aligned representation space for images and texts using paired data. Seminal works CLIP [34] and ALIGN [23] established the viability of this approach at a large scale, and following their success, many large image-text datasets became available privately [55, 13, 21, 45] and publicly [37, 6, 15, 7, 38]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 644, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 644, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 644, + 287, + 693 + ], + "type": "text", + "content": "The standard recipe to pre-train such models leverages the image-text contrastive objective. It aligns the image and text embeddings for matching (positive) image-text pairs while making sure that unrelated (negative) image-text pairs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 212, + 547, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 212, + 547, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 212, + 547, + 357 + ], + "type": "text", + "content": "Table 1: SigLiT and SigLIP results. Sigmoid loss is memory efficient, allows larger batch sizes (BS) that unlocks language image pre-training with a small number of chips. SigLiT model with a frozen public L/16 checkpoint [39], trained on the LiT image-text dataset [55] using four TPU-v4 chips for one day, achieves " + }, + { + "bbox": [ + 304, + 212, + 547, + 357 + ], + "type": "inline_equation", + "content": "79.7\\%" + }, + { + "bbox": [ + 304, + 212, + 547, + 357 + ], + "type": "text", + "content": " 0-shot accuracy on ImageNet. The same setup with a g/14 checkpoint [54] leads to " + }, + { + "bbox": [ + 304, + 212, + 547, + 357 + ], + "type": "inline_equation", + "content": "84.5\\%" + }, + { + "bbox": [ + 304, + 212, + 547, + 357 + ], + "type": "text", + "content": " accuracy, trained for two days. With a public unlocked B/16 image checkpoint [39], trained on the WebLI dataset [13], SigLIP achieves " + }, + { + "bbox": [ + 304, + 212, + 547, + 357 + ], + "type": "inline_equation", + "content": "71.0\\%" + }, + { + "bbox": [ + 304, + 212, + 547, + 357 + ], + "type": "text", + "content": " 0-shot accuracy using 16 TPU-v4 chips for three days. The last two rows show results with randomly initialized models." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 306, + 366, + 545, + 460 + ], + "blocks": [ + { + "bbox": [ + 306, + 366, + 545, + 460 + ], + "lines": [ + { + "bbox": [ + 306, + 366, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 366, + 545, + 460 + ], + "type": "table", + "html": "
ImageTextBS#TPUv4DaysINet-0
SigLiTB/8L*32 k4179.7
SigLiTg/14L20 k4284.5
SigLIPB/16B16 k16371.0
SigLIPB/16B32 k32272.1
SigLIPB/16B32 k32573.4
", + "image_path": "c191f62bf6543067b0a5bb14e69cbea68eac3d58063fd0c8bfc53f5da8ff81c3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 386, + 461, + 545, + 472 + ], + "lines": [ + { + "bbox": [ + 386, + 461, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 386, + 461, + 545, + 472 + ], + "type": "text", + "content": "* We use a variant of the " + }, + { + "bbox": [ + 386, + 461, + 545, + 472 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 386, + 461, + 545, + 472 + ], + "type": "text", + "content": " model with 12 layers." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 506, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 590 + ], + "type": "text", + "content": "are dissimilar in the embedding space. This is achieved via a batch-level softmax-based contrastive loss, applied twice to normalize the pairwise similarity scores across all images, then all texts. A naive implementation of the softmax is numerically unstable; it is usually stabilized by subtracting the maximum input value before applying the softmax [18], which requires another pass over the full batch." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we propose a simpler alternative: the sigmoid loss. It does not require any operation across the full batch and hence greatly simplifies the distributed loss implementation and boosts efficiency. Additionally, it conceptually decouples the batch size from the definition of the task. We compare the proposed sigmoid loss with the standard softmax loss across multiple setups. In particular, we investigate sigmoid-based loss with two prominent approaches for image-text learning: CLIP [34] and LiT [55], which we call sigmoid language image pre" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 127, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 127, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 127, + 712 + ], + "type": "text", + "content": "*equal contribution" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11975" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "training (SigLIP) and sigmoid LiT (SigLiT), respectively. We find that the sigmoid loss performs significantly better than the softmax loss when the batch size is smaller than " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "16\\mathrm{k}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": ". As the train batch size grows, the gap closes. Importantly, the sigmoid loss is symmetric, requires just a single pass, and a typical implementation requires less memory than the softmax loss. This enables successful training of a SigLiT model at a batch size of one million. However, we find that the performance saturates with growing batch size, both for softmax and sigmoid. The good news is that a reasonable batch size, i.e. " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "32\\mathrm{k}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": ", is sufficient for image-text pretraining. This conclusion also holds for multilingual SigLIP training on over 100 languages." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 290, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 290, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 290, + 396 + ], + "type": "text", + "content": "In Table 1, we present setups for image-text pre-training that require a moderate amount of TPUv4 chips for training. SigLiT is surprisingly efficient, reaching " + }, + { + "bbox": [ + 46, + 228, + 290, + 396 + ], + "type": "inline_equation", + "content": "79.7\\%" + }, + { + "bbox": [ + 46, + 228, + 290, + 396 + ], + "type": "text", + "content": " zero-shot accuracy on ImageNet in just a single day on four chips. SigLIP's more demanding from-scratch training reaches " + }, + { + "bbox": [ + 46, + 228, + 290, + 396 + ], + "type": "inline_equation", + "content": "73.4\\%" + }, + { + "bbox": [ + 46, + 228, + 290, + 396 + ], + "type": "text", + "content": " zero-shot accuracy in 5 days with 32 TPUv4 chips. This compares favorably to prior works such as FLIP [29] and CLIP [34], which require approximately 5 and 10 days respectively on 256 TPUv3 cores. When fine-tuning a pretrained vision backbone in SigLIP, denoted as " + }, + { + "bbox": [ + 46, + 228, + 290, + 396 + ], + "type": "inline_equation", + "content": "\\boxed{ \\begin{array}{r l} \\end{array} }" + }, + { + "bbox": [ + 46, + 228, + 290, + 396 + ], + "type": "text", + "content": " in Table 1, we found that disabling the weight decay on the pre-trained backbone leads to better results (see Figure 4 for details). We hope our work paves the way for making the nascent language-image pre-training field more accessible." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 406, + 134, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 134, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 134, + 418 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 430, + 287, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 526 + ], + "type": "text", + "content": "Contrastive learning with the sigmoid loss. One prior work proposes a similar sigmoid loss for the task of unsupervised dimensionality reduction [19]; in the scope of contrastive image-text learning, the vast majority of works rely on the softmax-based InfoNCE loss as popularized by [43]. In supervised classification, the sigmoid loss has already been shown to be slightly more effective and robust than the softmax loss [3, 47]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 529, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 638 + ], + "type": "text", + "content": "Contrastive language-image pre-training has become popular since CLIP [34] and ALIGN [23] applied softmax contrastive learning [56, 43, 10, 24] to large-scale image-text datasets. Both models perform very well on zero-shot transfer tasks, including classification and retrieval. Follow-up works show that contrastively pre-trained models produce good representations for fine-tuning [49, 16], linear regression [23], object detection [30], semantic segmentation [31] and video tasks [53]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "Generative language-image pre-training Besides softmax contrastive pre-training, various alternatives have been proposed. GIT [45], SimVLM [46], and LEMON [21] successfully pre-train models using a generative text decoder instead, while CoCa [52] adds such a decoder to the discriminative CLIP/ALIGN setup, thus combining the pros" + } + ] + } + ], + "index": 5 + }, + { + "type": "code", + "bbox": [ + 297, + 90, + 532, + 177 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 515, + 84 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 515, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 515, + 84 + ], + "type": "text", + "content": "Algorithm 1 Sigmoid loss pseudo-implementation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 297, + 90, + 532, + 177 + ], + "lines": [ + { + "bbox": [ + 297, + 90, + 532, + 177 + ], + "spans": [ + { + "bbox": [ + 297, + 90, + 532, + 177 + ], + "type": "text", + "content": "1 # img_emb : image model embedding [n, dim] \n2 # txt_emb : text model embedding [n, dim] \n3 # t_prime, b : learnable temperature and bias \n4 # n : mini-batch size \n5 \n6 t = exp(t_prime) \n7 zimg = l2_normalize(img_emb) \n8 ztxt = l2_normalize(txt_emb) \n9 logits = dot(zimg, ztxt.T) * t + b \n10 labels = 2 * eye(n) - ones(n) # -1 with diagonal 1 \n11 l = -sum(log_sigmoid(labels * logits)) / n" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "code_body" + } + ], + "index": 7, + "sub_type": "code", + "guess_lang": "python" + }, + { + "bbox": [ + 304, + 209, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 281 + ], + "type": "text", + "content": "and cons of both approaches into a single very capable model. BLIP [28] further proposes CapFilt which uses the generative decoder to create better captions and the discriminative part of the model to filter pairs. Language-Image pre-training is a very active field and surveys [8] rapidly become outdated." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 286, + 546, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 546, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 546, + 418 + ], + "type": "text", + "content": "Efficient language-image pre-training On the other hand, few works have tried making language image pre-training more efficient. LiT [55] and FLIP [29] are notable attempts, the former requires a pre-trained and locked backbone, and the latter sacrifices quality by randomly dropping visual tokens. BASIC [33] and LAION [48] look at scaling batch-size but only go up to " + }, + { + "bbox": [ + 304, + 286, + 546, + 418 + ], + "type": "inline_equation", + "content": "16\\mathrm{k}" + }, + { + "bbox": [ + 304, + 286, + 546, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 286, + 546, + 418 + ], + "type": "inline_equation", + "content": "160\\mathrm{k}" + }, + { + "bbox": [ + 304, + 286, + 546, + 418 + ], + "type": "text", + "content": " respectively, by using many hundreds of chips, and for the former also mixing in a large private classification dataset [33, 51]. The recent Lion optimizer [12] claims to be able to reduce the training cost to reach similar quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 431, + 361, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 361, + 443 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 361, + 443 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 453, + 545, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 487 + ], + "type": "text", + "content": "In this section, we first review the widely-used softmax-based contrastive loss. We then introduce the pairwise sigmoid loss and discuss its efficient implementation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 488, + 546, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 488, + 546, + 585 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 546, + 585 + ], + "type": "text", + "content": "Given a mini-batch " + }, + { + "bbox": [ + 304, + 488, + 546, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{(I_1, T_1), (I_2, T_2), \\ldots\\}" + }, + { + "bbox": [ + 304, + 488, + 546, + 585 + ], + "type": "text", + "content": " of image-text pairs, the contrastive learning objective encourages embeddings of matching pairs " + }, + { + "bbox": [ + 304, + 488, + 546, + 585 + ], + "type": "inline_equation", + "content": "(I_i, T_i)" + }, + { + "bbox": [ + 304, + 488, + 546, + 585 + ], + "type": "text", + "content": " to align with each other, while pushing embeddings of unmatched pairs " + }, + { + "bbox": [ + 304, + 488, + 546, + 585 + ], + "type": "inline_equation", + "content": "(I_i, T_{j \\neq i})" + }, + { + "bbox": [ + 304, + 488, + 546, + 585 + ], + "type": "text", + "content": " apart. For practical purposes, it is assumed that for all images " + }, + { + "bbox": [ + 304, + 488, + 546, + 585 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 488, + 546, + 585 + ], + "type": "text", + "content": ", the text associated with a different image " + }, + { + "bbox": [ + 304, + 488, + 546, + 585 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 488, + 546, + 585 + ], + "type": "text", + "content": " is not related to " + }, + { + "bbox": [ + 304, + 488, + 546, + 585 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 488, + 546, + 585 + ], + "type": "text", + "content": ", and vice-versa. This assumption is usually noisy and imperfect." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 594, + 539, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 594, + 539, + 608 + ], + "spans": [ + { + "bbox": [ + 305, + 594, + 539, + 608 + ], + "type": "text", + "content": "3.1. Softmax loss for language image pre-training" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 613, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 613, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 613, + 545, + 651 + ], + "type": "text", + "content": "When using the softmax loss to formalize this objective, an image model " + }, + { + "bbox": [ + 304, + 613, + 545, + 651 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 304, + 613, + 545, + 651 + ], + "type": "text", + "content": " and a text model " + }, + { + "bbox": [ + 304, + 613, + 545, + 651 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 304, + 613, + 545, + 651 + ], + "type": "text", + "content": " are trained to minimize the following objective:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 659, + 533, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 533, + 717 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 533, + 717 + ], + "type": "interline_equation", + "content": "- \\frac {1}{2 | \\mathcal {B} |} \\sum_ {i = 1} ^ {| \\mathcal {B} |} \\left(\\overbrace {\\log \\frac {e ^ {t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {i}}}{\\sum_ {j = 1} ^ {| \\mathcal {B} |} e ^ {t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {j}}}} ^ {\\text {i m a g e} \\rightarrow \\text {t e x t s o f t m a x}} + \\overbrace {\\log \\frac {e ^ {t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {i}}}{\\sum_ {j = 1} ^ {| \\mathcal {B} |} e ^ {t \\mathbf {x} _ {j} \\cdot \\mathbf {y} _ {i}}}} ^ {\\text {t e x t} \\rightarrow \\text {i m a g e s o f t m a x}}\\right)", + "image_path": "3429424237d4f9e461ae9338e8feffb108877199ffba6c86b19983b536fd1a17.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11976" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 170, + 185 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 170, + 185 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 170, + 185 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 170, + 185 + ], + "type": "image", + "image_path": "40b902962f5bea4a83d847167bbf30f869720dcad723b3cfcf9d1f670646a50c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 293, + 290 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 293, + 290 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 293, + 290 + ], + "type": "text", + "content": "(a) Initially each device holds 4 (b) They each compute the comimage and 4 text representations. component of the loss (highlighted) Each device needs to see the repre for their representations, which representations from other devices includes the positives. to calculate the full loss." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 46, + 300, + 544, + 326 + ], + "lines": [ + { + "bbox": [ + 46, + 300, + 544, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 544, + 326 + ], + "type": "text", + "content": "Figure 1: Efficient loss implementation demonstrated via a mock setup with 3 devices and a global batch size of 12. There are no all-gathers, and at any point in time only the bright yellow square (size " + }, + { + "bbox": [ + 46, + 300, + 544, + 326 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 46, + 300, + 544, + 326 + ], + "type": "text", + "content": ") is materialized in memory." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 68, + 294, + 214 + ], + "blocks": [ + { + "bbox": [ + 173, + 68, + 294, + 214 + ], + "lines": [ + { + "bbox": [ + 173, + 68, + 294, + 214 + ], + "spans": [ + { + "bbox": [ + 173, + 68, + 294, + 214 + ], + "type": "image", + "image_path": "f9c22c520c1dfb106788a3f98f65c3814546d131b705c71c53e667fdf7d80973.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 236, + 419, + 291 + ], + "lines": [ + { + "bbox": [ + 296, + 236, + 419, + 291 + ], + "spans": [ + { + "bbox": [ + 296, + 236, + 419, + 291 + ], + "type": "text", + "content": "(c) Texts are swapped across the devices, so device 1 now has " + }, + { + "bbox": [ + 296, + 236, + 419, + 291 + ], + "type": "inline_equation", + "content": "I_{1:4}" + }, + { + "bbox": [ + 296, + 236, + 419, + 291 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 236, + 419, + 291 + ], + "type": "inline_equation", + "content": "T_{5:8}" + }, + { + "bbox": [ + 296, + 236, + 419, + 291 + ], + "type": "text", + "content": " etc. The new loss is computed and accumulated with the previous." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 296, + 68, + 419, + 215 + ], + "blocks": [ + { + "bbox": [ + 296, + 68, + 419, + 215 + ], + "lines": [ + { + "bbox": [ + 296, + 68, + 419, + 215 + ], + "spans": [ + { + "bbox": [ + 296, + 68, + 419, + 215 + ], + "type": "image", + "image_path": "fbc90ed975367f418d02645cc1e6163dc3f70c8308d1fad6aede1cbd706cbedd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 236, + 544, + 292 + ], + "lines": [ + { + "bbox": [ + 421, + 236, + 544, + 292 + ], + "spans": [ + { + "bbox": [ + 421, + 236, + 544, + 292 + ], + "type": "text", + "content": "(d) This repeats till every image & text pair have interacted, e.g. device 1 has the loss of " + }, + { + "bbox": [ + 421, + 236, + 544, + 292 + ], + "type": "inline_equation", + "content": "I_{1:4}" + }, + { + "bbox": [ + 421, + 236, + 544, + 292 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 421, + 236, + 544, + 292 + ], + "type": "inline_equation", + "content": "T_{1:12}" + }, + { + "bbox": [ + 421, + 236, + 544, + 292 + ], + "type": "text", + "content": ". A final cross-device sum brings everything together." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 421, + 68, + 544, + 227 + ], + "blocks": [ + { + "bbox": [ + 421, + 68, + 544, + 227 + ], + "lines": [ + { + "bbox": [ + 421, + 68, + 544, + 227 + ], + "spans": [ + { + "bbox": [ + 421, + 68, + 544, + 227 + ], + "type": "image", + "image_path": "7e79e3050f7c6ac6fdadeea3b3b6a16f4aba788a667d2823cd80049989fa559c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 344, + 287, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 344, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 287, + 431 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 344, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i = \\frac{f(I_i)}{\\|f(I_i)\\|_2}" + }, + { + "bbox": [ + 46, + 344, + 287, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 344, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i = \\frac{g(T_i)}{\\|g(T_i)\\|_2}" + }, + { + "bbox": [ + 46, + 344, + 287, + 431 + ], + "type": "text", + "content": ". In this paper, we adopt the vision transformer architecture [17] for images and the transformer architecture [44] for texts. Note that due to the asymmetry of the softmax loss, the normalization is independently performed two times: across images and across texts [34]. The scalar " + }, + { + "bbox": [ + 46, + 344, + 287, + 431 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 344, + 287, + 431 + ], + "type": "text", + "content": " is parametrized as " + }, + { + "bbox": [ + 46, + 344, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\exp(t')" + }, + { + "bbox": [ + 46, + 344, + 287, + 431 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 344, + 287, + 431 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 46, + 344, + 287, + 431 + ], + "type": "text", + "content": " is a global freely learnable parameter." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 439, + 280, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 439, + 280, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 439, + 280, + 453 + ], + "type": "text", + "content": "3.2. Sigmoid loss for language image pre-training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 458, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 555 + ], + "type": "text", + "content": "Instead of the softmax-based contrastive loss, we propose a simpler alternative that does not require computing global normalization factors. The sigmoid-based loss processes every image-text pair independently, effectively turning the learning problem into the standard binary classification on the dataset of all pair combinations, with a positive labels for the matching pairs " + }, + { + "bbox": [ + 46, + 458, + 287, + 555 + ], + "type": "inline_equation", + "content": "(I_i, T_i)" + }, + { + "bbox": [ + 46, + 458, + 287, + 555 + ], + "type": "text", + "content": " and negative labels for all other pairs " + }, + { + "bbox": [ + 46, + 458, + 287, + 555 + ], + "type": "inline_equation", + "content": "(I_i, T_{j \\neq i})" + }, + { + "bbox": [ + 46, + 458, + 287, + 555 + ], + "type": "text", + "content": ". It is defined as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 91, + 563, + 245, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 563, + 245, + 609 + ], + "spans": [ + { + "bbox": [ + 91, + 563, + 245, + 609 + ], + "type": "interline_equation", + "content": "- \\frac {1}{| \\mathcal {B} |} \\sum_ {i = 1} ^ {| \\mathcal {B} |} \\sum_ {j = 1} ^ {| \\mathcal {B} |} \\underbrace {\\log \\frac {1}{1 + e ^ {z _ {i j} (- t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {j} + b)}}} _ {\\mathcal {L} _ {i j}}", + "image_path": "3a7e068446670cb7f68560bd66223ec4919ac35efbf415848f4133e4d15a99b8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "z_{ij}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " is the label for a given image and text input, which equals 1 if they are paired and -1 otherwise. Note that at initialization, the heavy imbalance coming from the many negatives dominates the loss, leading to large initial optimization steps attempting to correct this bias. To alleviate this, we introduce an additional learnable bias term " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " similar to the temperature " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": ". We initialize " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " to 10 and -10 respectively. This makes sure the training starts roughly close" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 346, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 545, + 384 + ], + "type": "text", + "content": "to the prior and does not require massive over-correction. Algorithm 1 presents a pseudocode implementation of the proposed sigmoid loss for language image pre-training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 388, + 499, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 388, + 499, + 401 + ], + "spans": [ + { + "bbox": [ + 306, + 388, + 499, + 401 + ], + "type": "text", + "content": "3.3. Efficient \"chunked\" implementation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 407, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 545, + 467 + ], + "type": "text", + "content": "Contrastive training typically utilizes data parallelism. Computing the loss when data is split across " + }, + { + "bbox": [ + 305, + 407, + 545, + 467 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 407, + 545, + 467 + ], + "type": "text", + "content": " devices necessitates gathering all embeddings [55] with expensive all-gathers and, more importantly, the materialization of a memory-intensive " + }, + { + "bbox": [ + 305, + 407, + 545, + 467 + ], + "type": "inline_equation", + "content": "|\\mathcal{B}| \\times |\\mathcal{B}|" + }, + { + "bbox": [ + 305, + 407, + 545, + 467 + ], + "type": "text", + "content": " matrix of pairwise similarities." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 467, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 467, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 305, + 467, + 545, + 518 + ], + "type": "text", + "content": "The sigmoid loss, however, is particularly amenable to a memory efficient, fast, and numerically stable implementation that ameliorates both these issues. Denoting the perdevice batch size as " + }, + { + "bbox": [ + 305, + 467, + 545, + 518 + ], + "type": "inline_equation", + "content": "b = \\frac{|\\mathcal{B}|}{D}" + }, + { + "bbox": [ + 305, + 467, + 545, + 518 + ], + "type": "text", + "content": ", the loss is reformulated as:" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 332, + 523, + 521, + 601 + ], + "blocks": [ + { + "bbox": [ + 332, + 523, + 521, + 601 + ], + "lines": [ + { + "bbox": [ + 332, + 523, + 521, + 601 + ], + "spans": [ + { + "bbox": [ + 332, + 523, + 521, + 601 + ], + "type": "image", + "image_path": "970912ada39eeb482401c9a2af5950c7c2af4acfff9bd0018fd6a049194b920b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "This is particularly simple for the sigmoid loss as each pair is an independent term in the loss. Figure 1 illustrates this method. In words, we first compute the component of the loss corresponding to the positive pairs, and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "b - 1" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " negative pairs. We then permute representations across devices, so each device takes negatives from its neighbouring device (next iteration of sum B). The loss is then calculated with respect to this chunk (sum C). This is done independently in each device, such that each device computes the" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11977" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 71, + 218, + 224 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 218, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 218, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 218, + 224 + ], + "type": "image", + "image_path": "b6a0eca64575fcaa25a7711def362e40f517d55a837a8473122c285cc177fbbd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 548, + 333 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 548, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 548, + 333 + ], + "type": "text", + "content": "Figure 2: The effect of pre-training batch size. Left: SigLiT results, trained for 18B seen examples. Sigmoid loss outperforms the softmax loss significantly with small batch sizes, and performs similarly at larger batch sizes. We successfully trained an SigLiT model with up to one million batch size. However, performance for both sigmoid and softmax saturate at around " + }, + { + "bbox": [ + 46, + 236, + 548, + 333 + ], + "type": "inline_equation", + "content": "32\\mathrm{k}" + }, + { + "bbox": [ + 46, + 236, + 548, + 333 + ], + "type": "text", + "content": " batch size. Middle: SigLIP results, trained for 9B seen examples. Both sigmoid loss and softmax loss saturate at a reasonable batch size, while the peak of the sigmoid loss comes earlier and slightly outperforms the peak of the softmax loss. A very large batch size hurts both losses. Right: mSigLIP results, trained for 30B seen examples. With a multilingual setup using over 100 languages, " + }, + { + "bbox": [ + 46, + 236, + 548, + 333 + ], + "type": "inline_equation", + "content": "32\\mathrm{k}" + }, + { + "bbox": [ + 46, + 236, + 548, + 333 + ], + "type": "text", + "content": " batch size is surprisingly sufficient and scaling beyond that hurts performance on a 36-language cross-modal retrieval task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 71, + 376, + 224 + ], + "blocks": [ + { + "bbox": [ + 225, + 71, + 376, + 224 + ], + "lines": [ + { + "bbox": [ + 225, + 71, + 376, + 224 + ], + "spans": [ + { + "bbox": [ + 225, + 71, + 376, + 224 + ], + "type": "image", + "image_path": "d7d5f072cdf4d3412cacf41ede119e784e91cde95ebd583f4570da79f9548186.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 378, + 72, + 539, + 224 + ], + "blocks": [ + { + "bbox": [ + 378, + 72, + 539, + 224 + ], + "lines": [ + { + "bbox": [ + 378, + 72, + 539, + 224 + ], + "spans": [ + { + "bbox": [ + 378, + 72, + 539, + 224 + ], + "type": "image", + "image_path": "9d01164e07d4eef84e6f33fa78d4eb81c626ee64cd20ac0d5ab0924f0fc47b48.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 354, + 289, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 289, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 289, + 486 + ], + "type": "text", + "content": "loss with respect to its local batch " + }, + { + "bbox": [ + 46, + 354, + 289, + 486 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 354, + 289, + 486 + ], + "type": "text", + "content": ". Losses can then simply be summed across all devices (sum A). Individual collective permutes (for sum B) are fast (and indeed " + }, + { + "bbox": [ + 46, + 354, + 289, + 486 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 354, + 289, + 486 + ], + "type": "text", + "content": " collective permutes is typically faster than two all-gathers between " + }, + { + "bbox": [ + 46, + 354, + 289, + 486 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 354, + 289, + 486 + ], + "type": "text", + "content": " devices), and the memory cost at any given moment is reduced from " + }, + { + "bbox": [ + 46, + 354, + 289, + 486 + ], + "type": "inline_equation", + "content": "|\\mathcal{B}|^2" + }, + { + "bbox": [ + 46, + 354, + 289, + 486 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 354, + 289, + 486 + ], + "type": "inline_equation", + "content": "b^2" + }, + { + "bbox": [ + 46, + 354, + 289, + 486 + ], + "type": "text", + "content": " (for sum C). Usually " + }, + { + "bbox": [ + 46, + 354, + 289, + 486 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 354, + 289, + 486 + ], + "type": "text", + "content": " is constant as scaling " + }, + { + "bbox": [ + 46, + 354, + 289, + 486 + ], + "type": "inline_equation", + "content": "|\\mathcal{B}|" + }, + { + "bbox": [ + 46, + 354, + 289, + 486 + ], + "type": "text", + "content": " is achieved by increasing the number of accelerators. Due to being quadratic with respect to the batch size, the vanilla loss computation rapidly bottlenecks scaling up. This chunked approach enabled training with batch sizes over 1 million on relatively few devices." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 506, + 100, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 100, + 520 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 100, + 520 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 530, + 287, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 287, + 650 + ], + "type": "text", + "content": "In this section, we evaluate the proposed SigLiT and SigLIP models across a wide range of batch sizes. We discuss what can be achieved with a small number of accelerator chips, using both SigLiT and SigLIP recipes. We also briefly discuss the impact of batch size on multilingual language image pre-training. We ablate the importance of our large-batch stabilization modification and the introduced learned bias term and present a study on the effect of positive and negative pairs ratio in the sigmoid loss. Lastly, we explore SigLIP's data noise robustness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "To validate our models, we report zero-shot transfer results on the ImageNet dataset [14] and zero-shot retrieval results across 36 languages on the XM3600 dataset [41]. We use the ScalingViT-Adafactor optimizer [54] by default for all our experiments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 353, + 504, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 353, + 504, + 367 + ], + "spans": [ + { + "bbox": [ + 305, + 353, + 504, + 367 + ], + "type": "text", + "content": "4.1. SigLiT: Scaling batch size to the limit" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 373, + 546, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 546, + 422 + ], + "type": "text", + "content": "Following [55], we use the same precomputed embeddings for the images using a ViT-g vision model, and train a base size text tower from scratch with the same hyperparameters using the LiT image-text dataset [55]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 424, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 547, + 555 + ], + "type": "text", + "content": "We perform a study over a wide range of batch sizes, from 512 to " + }, + { + "bbox": [ + 304, + 424, + 547, + 555 + ], + "type": "inline_equation", + "content": "1M" + }, + { + "bbox": [ + 304, + 424, + 547, + 555 + ], + "type": "text", + "content": ", demonstrating the impact of batch size for contrastive learning. Results are presented in Figure 2 (left). When the batch size is smaller than " + }, + { + "bbox": [ + 304, + 424, + 547, + 555 + ], + "type": "inline_equation", + "content": "16k" + }, + { + "bbox": [ + 304, + 424, + 547, + 555 + ], + "type": "text", + "content": ", sigmoid loss outperforms softmax loss by a large margin. With growing batch sizes, we observe that softmax loss quickly catches up and potentially slightly underperforms sigmoid loss with a large enough batch size. Overall, we recommend using the SigLIP recipe for large batch sizes as well, due to the simplicity, compute savings, and straightforward memory efficient implementation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "content": "There is a consensus that contrastive learning benefits from large batch sizes, while most of the existing studies stop at " + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "inline_equation", + "content": "64\\mathrm{k}" + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "content": " batch size [55, 33, 10]. We successfully trained an SigLiT model at one million batch size, to explore the limit of contrastive learning. To our surprise, the performance saturates at " + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "inline_equation", + "content": "32\\mathrm{k}" + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "content": " batch size, further scaling up the batch size only gives a minor boost, and the model peaks at " + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "inline_equation", + "content": "256\\mathrm{k}" + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "content": " batch size. Our best SigLiT with a " + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "content": "-sized text mode achieves " + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "inline_equation", + "content": "84.7\\%" + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "content": " zero-shot transfer accuracy on ImageNet, while the original LiT paper reports a slightly better " + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "inline_equation", + "content": "85.2\\%" + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "content": " score with a 10 times larger " + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "content": "-sized text model. Figure 3 presents the impact of training duration for different batch sizes. It demonstrates that large, " + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "inline_equation", + "content": "262k" + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "content": " batch size signifi" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11978" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 277, + 224 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 277, + 224 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 277, + 224 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 277, + 224 + ], + "type": "image", + "image_path": "8e77149cddcaaa1e9ded9c72e0d68c055f9f4e30b49d6e13288a4ce900acd2da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 237, + 288, + 299 + ], + "lines": [ + { + "bbox": [ + 46, + 237, + 288, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 288, + 299 + ], + "type": "text", + "content": "Figure 3: SigLiT ImageNet 0-shot transfer results with different training durations. Large batch size results in a big performance boost, but needs a sufficiently long schedule to ramp up, as for short schedules, very large batch size results in a small number of gradient update steps." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 328, + 288, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 288, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 288, + 377 + ], + "type": "text", + "content": "cantly outperforms smaller " + }, + { + "bbox": [ + 46, + 328, + 288, + 377 + ], + "type": "inline_equation", + "content": "8k" + }, + { + "bbox": [ + 46, + 328, + 288, + 377 + ], + "type": "text", + "content": " batch size when trained for a sufficiently long time. Note, that for short training durations, large batch size leads to the fewer absolute number of update steps and thus needs more time to ramp up." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 398, + 287, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 398, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 47, + 398, + 287, + 424 + ], + "type": "text", + "content": "4.2. SigLIP: Sigmoid loss is beneficial for language-image pre-training" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 433, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 636 + ], + "type": "text", + "content": "We pre-train SigLIP models on the WebLI dataset [13], using only English image and text pairs. We use moderately-sized models: B/16 ViT for image embeddings and B-sized transformer for text embeddings. The input images are resized to " + }, + { + "bbox": [ + 46, + 433, + 287, + 636 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 433, + 287, + 636 + ], + "type": "text", + "content": " resolution. The text is tokenized by a " + }, + { + "bbox": [ + 46, + 433, + 287, + 636 + ], + "type": "inline_equation", + "content": "32\\mathrm{k}" + }, + { + "bbox": [ + 46, + 433, + 287, + 636 + ], + "type": "text", + "content": " vocabulary sentencepiece tokenizer [27] trained on the English C4 dataset [35], and a maximum of 16 text tokens are kept. Figure 2 middle plot shows SigLIP results, With less than " + }, + { + "bbox": [ + 46, + 433, + 287, + 636 + ], + "type": "inline_equation", + "content": "32\\mathrm{k}" + }, + { + "bbox": [ + 46, + 433, + 287, + 636 + ], + "type": "text", + "content": " batch size, SigLIP outperforms CLIP baselines with the standard softmax loss. On the other end of the scale, the memory efficiency of the sigmoid loss enabled much larger batch sizes. For example, with four TPU-v4 chips, we could fit a batch size of 4096 with a Base SigLIP but only 2048 with a corresponding CLIP model. The two advantages together demonstrate significant benefits of the sigmoid loss for language image pre-training with fixed resources, which will be discussed in Section 4.5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "text", + "content": "As batch size increases, the gap between the sigmoid and the softmax losses diminish. SigLIP performs best at batch size " + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "inline_equation", + "content": "32\\mathrm{k}" + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "text", + "content": ", whereas the softmax loss required " + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "inline_equation", + "content": "98\\mathrm{k}" + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "text", + "content": " for optimal performance and still didn't outperform the sigmoid based variant. Scaling further, a larger batch size like " + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "inline_equation", + "content": "307\\mathrm{k}" + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "text", + "content": " hurts both losses." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 305, + 70, + 547, + 196 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 547, + 196 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 547, + 196 + ], + "type": "table", + "html": "
16 k32 k64 k128 k240 k
INet-071.673.273.273.273.1
XM avg34.834.934.433.632.7
XM de54.754.855.454.354.7
XM en46.546.246.546.646.6
XM hi9.18.57.98.17.3
XM ru50.149.949.748.649.3
XM zh30.732.532.030.623.7
", + "image_path": "5b32bc4d8e4e01531b2edadbe7dba03567815f9783ca946f0f8fd9b10674228e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 203, + 547, + 276 + ], + "lines": [ + { + "bbox": [ + 304, + 203, + 547, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 203, + 547, + 276 + ], + "type": "text", + "content": "Table 2: Multilingual SigLIP results with various batch sizes, pre-trained for 30 billion seen examples. We report zero-shot transfer results on ImageNet (INet-0) and averaged text to image retrieval results across 36 languages on the crossmodal 3600 dataset (XM). The full table on 36 languages can be found in Appendix." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 319, + 500, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 319, + 500, + 332 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 500, + 332 + ], + "type": "text", + "content": "4.3. mSigLIP: Multi-lingual pre-training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 338, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 445 + ], + "type": "text", + "content": "We further scale up the training data by keeping all the 100 languages from the WebLI dataset [13]. With multilingual data, one usually needs to use a larger international vocabulary. We first verify the impact of two tokenizers: a small multilingual vocabulary with " + }, + { + "bbox": [ + 304, + 338, + 545, + 445 + ], + "type": "inline_equation", + "content": "32\\mathrm{k}" + }, + { + "bbox": [ + 304, + 338, + 545, + 445 + ], + "type": "text", + "content": " tokens [35], and a large multilingual vocabulary with " + }, + { + "bbox": [ + 304, + 338, + 545, + 445 + ], + "type": "inline_equation", + "content": "250\\mathrm{k}" + }, + { + "bbox": [ + 304, + 338, + 545, + 445 + ], + "type": "text", + "content": " tokens [50]. We train B-sized ViT and text models for " + }, + { + "bbox": [ + 304, + 338, + 545, + 445 + ], + "type": "inline_equation", + "content": "900M" + }, + { + "bbox": [ + 304, + 338, + 545, + 445 + ], + "type": "text", + "content": " total examples seen, and observe slightly more than " + }, + { + "bbox": [ + 304, + 338, + 545, + 445 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 338, + 545, + 445 + ], + "type": "text", + "content": " improvement when using a larger vocabulary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "type": "text", + "content": "However, the token embeddings become huge for very large vocabulary sizes. Following the standard setup, we would need to store a " + }, + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "type": "inline_equation", + "content": "N \\times W" + }, + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "type": "text", + "content": " token embedding lookup table to train the multilingual model, where " + }, + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "type": "text", + "content": " is the vocabulary size mentioned above and " + }, + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "type": "text", + "content": " is the embedding dimension of the text model. To save memory, we propose to use a \"bottlenecked\" token embedding. We use " + }, + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "type": "inline_equation", + "content": "N \\times K" + }, + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "type": "text", + "content": " embedding matrix and additional " + }, + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "type": "inline_equation", + "content": "K \\times W" + }, + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "type": "text", + "content": " projection, where the bottleneck " + }, + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "type": "text", + "content": " is much smaller than " + }, + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "content": "In our experiments, we observed that using a large multilingual vocabulary with a bottleneck can be scaled up as efficiently as using a small multilingual vocabulary. Specifically, by enabling the bottleneck of size " + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "inline_equation", + "content": "K = 96" + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "content": " for Base architecture with " + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "inline_equation", + "content": "W = 768" + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "content": ", we only see about a half percent quality drop on ImageNet zero-shot transfer, compared to using the full " + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "inline_equation", + "content": "250k" + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "content": " vocabulary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "With the memory improvements, we train mSigLIP models for various batch sizes, for a total of 30 billion examples seen. Table 2 and Figure 2 (right plot) show the results. We were expecting a large batch size to improve multilingual pre-training, where the model sees more examples from the same language as hard negatives in a sin" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11979" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 72, + 285, + 289 + ], + "blocks": [ + { + "bbox": [ + 55, + 72, + 285, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 72, + 285, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 285, + 289 + ], + "type": "image", + "image_path": "a8d6942afea6fe1bd82be48abdd04d2c8e997bafd7cf91bed102f0dc5b56595e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 297, + 289, + 392 + ], + "lines": [ + { + "bbox": [ + 46, + 297, + 289, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 289, + 392 + ], + "type": "text", + "content": "Figure 4: Top: SigLIP with pre-trained encoders ramps up quickly. However, only disabling weight decay on the pre-trained encoder weights leads to stable behavior and good ImageNet 0-shot transfer results. Bottom: ImageNet 10-shot transfer results, where decaying the pre-trained weights leads to deterioration of the pre-trained model visual representation quality. Disabling weight decay makes the curve flatter." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 417, + 289, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 417, + 289, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 417, + 289, + 550 + ], + "type": "text", + "content": "gle mini-batch. However, we didn't observe clear improvements with a batch size larger than " + }, + { + "bbox": [ + 46, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "32\\mathrm{k}" + }, + { + "bbox": [ + 46, + 417, + 289, + 550 + ], + "type": "text", + "content": ". A batch size of " + }, + { + "bbox": [ + 46, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "32\\mathrm{k}" + }, + { + "bbox": [ + 46, + 417, + 289, + 550 + ], + "type": "text", + "content": " is sufficient for a multilingual setup as well. On the XM3600 cross-modal retrieval tasks, we found that going beyond " + }, + { + "bbox": [ + 46, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "32\\mathrm{k}" + }, + { + "bbox": [ + 46, + 417, + 289, + 550 + ], + "type": "text", + "content": " batch size leads to worse results on average while on ImageNet zero-shot transfer it stays flat. mSigLIP sets the new state-of-the-art on XM3600 text to image retrieval task, with only a Base size model. Our best result is " + }, + { + "bbox": [ + 46, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "34.9\\%" + }, + { + "bbox": [ + 46, + 417, + 289, + 550 + ], + "type": "text", + "content": ", which is more than " + }, + { + "bbox": [ + 46, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 46, + 417, + 289, + 550 + ], + "type": "text", + "content": " higher than the previously reported result " + }, + { + "bbox": [ + 46, + 417, + 289, + 550 + ], + "type": "inline_equation", + "content": "28.5\\%" + }, + { + "bbox": [ + 46, + 417, + 289, + 550 + ], + "type": "text", + "content": " [13] with a standard LiT model [55] using a much larger four billion ViT-e model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 560, + 216, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 560, + 216, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 560, + 216, + 574 + ], + "type": "text", + "content": "4.4. SigLiT with four TPU-v4 chips" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 580, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 640 + ], + "type": "text", + "content": "For many practitioners, the important question usually is \"what can be trained with a limited amount of resources?\" We explore the usage of SigLiT models in this section with only four TPU-v4 chips, as the memory efficient sigmoid loss is suitable for this application scenario." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "We follow the same setup as in section 4.1. We use the publicly available ViT-Augreg-B/8 [39] model as the frozen ( ) vision tower, and precompute embeddings to accelerate the training [55]. The text model is a Large Transformer, but with a depth of only 12 layers (instead of 24). It is trained using the LION [12] optimizer with decoupled weight de" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 542, + 287 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 542, + 287 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 542, + 287 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 542, + 287 + ], + "type": "image", + "image_path": "e87c68b11bc442c003f85424b4a8084854d3ccc5d0a5587f463252ad4ad1e0a4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 297, + 547, + 395 + ], + "lines": [ + { + "bbox": [ + 304, + 297, + 547, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 547, + 395 + ], + "type": "text", + "content": "Figure 5: The effect of Adam and AdaFactor's " + }, + { + "bbox": [ + 304, + 297, + 547, + 395 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 304, + 297, + 547, + 395 + ], + "type": "text", + "content": ". As we increase batch-size, we observe more frequent training instability. This instability can mainly be seen in the loss curves (top) and is caused by spikes in the gradient norm (middle) which results in large parameter updates (bottom). Decreasing the " + }, + { + "bbox": [ + 304, + 297, + 547, + 395 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 304, + 297, + 547, + 395 + ], + "type": "text", + "content": " momentum value stabilizes the training. Even though occasional gradient spikes still happen (see step at 2B), they do not destabilize the training process." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 417, + 547, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 547, + 551 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 547, + 551 + ], + "type": "text", + "content": "cay " + }, + { + "bbox": [ + 304, + 417, + 547, + 551 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-7}" + }, + { + "bbox": [ + 304, + 417, + 547, + 551 + ], + "type": "text", + "content": ", linearly warm-up of learning rate over " + }, + { + "bbox": [ + 304, + 417, + 547, + 551 + ], + "type": "inline_equation", + "content": "6.5\\mathrm{k}" + }, + { + "bbox": [ + 304, + 417, + 547, + 551 + ], + "type": "text", + "content": " steps up to a peak of " + }, + { + "bbox": [ + 304, + 417, + 547, + 551 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 417, + 547, + 551 + ], + "type": "text", + "content": ", followed by a cosine decay to 0. We train for a total of 65000 steps with a batch size of " + }, + { + "bbox": [ + 304, + 417, + 547, + 551 + ], + "type": "inline_equation", + "content": "32\\mathrm{k}" + }, + { + "bbox": [ + 304, + 417, + 547, + 551 + ], + "type": "text", + "content": " - this leads to just under one day of training. Table 1 shows the results when training a model on four chips for one day, achieving " + }, + { + "bbox": [ + 304, + 417, + 547, + 551 + ], + "type": "inline_equation", + "content": "79.7\\%" + }, + { + "bbox": [ + 304, + 417, + 547, + 551 + ], + "type": "text", + "content": " 0-shot ImageNet classification accuracy; very competitive in this limited resource regime. With a ViT-g/14 [54] model as the vision tower and a Large text tower, we can train at " + }, + { + "bbox": [ + 304, + 417, + 547, + 551 + ], + "type": "inline_equation", + "content": "20\\mathrm{k}" + }, + { + "bbox": [ + 304, + 417, + 547, + 551 + ], + "type": "text", + "content": " batch size on four chips for " + }, + { + "bbox": [ + 304, + 417, + 547, + 551 + ], + "type": "inline_equation", + "content": "107\\mathrm{k}" + }, + { + "bbox": [ + 304, + 417, + 547, + 551 + ], + "type": "text", + "content": " steps in under two days. This further pushes the 0-shot ImageNet classification accuracy up to " + }, + { + "bbox": [ + 304, + 417, + 547, + 551 + ], + "type": "inline_equation", + "content": "84.5\\%" + }, + { + "bbox": [ + 304, + 417, + 547, + 551 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 562, + 538, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 538, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 538, + 575 + ], + "type": "text", + "content": "4.5. SigLIP with a small amount of TPU-v4 chips" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "It's resource demanding to train a CLIP model from scratch in general, with SigLIP it's possible to fit a larger train batch size with fewer amount of chips. In this section, we explore ways to train SigLIP models efficiently with pretrained weights. We use pre-trained weights to initialize the image model to accelerate the pre-training, which was originally discussed in [55]. We use the public and unlocked ViT-Augreg-B/16 [39] model to initialize our vision tower and fine-tune on the same WebLI English data as used for SigLIP. In all the experiments, we apply a 0.1 learning rate multiplier to the pre-trained image tower to make it suitable" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11980" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 211, + 237 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 211, + 237 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 211, + 237 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 211, + 237 + ], + "type": "image", + "image_path": "98d1b8aa5b072bfb683bf8d0ab866d58f176ca50affa4e4cdf43d7f9209e646c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 248, + 545, + 320 + ], + "lines": [ + { + "bbox": [ + 45, + 248, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 45, + 248, + 545, + 320 + ], + "type": "text", + "content": "Figure 6: The effect of batch composition. We simulate various batch compositions by masking out negatives, either randomly, keeping only the hardest, or the easiest. With no masking, we have " + }, + { + "bbox": [ + 45, + 248, + 545, + 320 + ], + "type": "inline_equation", + "content": "16\\mathrm{k}" + }, + { + "bbox": [ + 45, + 248, + 545, + 320 + ], + "type": "text", + "content": " negatives for each positive in the batch " + }, + { + "bbox": [ + 45, + 248, + 545, + 320 + ], + "type": "inline_equation", + "content": "(1:16\\mathrm{k})" + }, + { + "bbox": [ + 45, + 248, + 545, + 320 + ], + "type": "text", + "content": " and the strongest masking we apply (1:1.6) results in almost balanced minibatches. In one setting we match total pairs seen by training for significantly longer. We observe ImageNet 0-shot score, the final value of the learned bias, and the average logits of positive and negative pairs. Overall, the imbalance does not seem to be detrimental, but finding an efficient way of mining negatives might be beneficial." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 213, + 72, + 375, + 236 + ], + "blocks": [ + { + "bbox": [ + 213, + 72, + 375, + 236 + ], + "lines": [ + { + "bbox": [ + 213, + 72, + 375, + 236 + ], + "spans": [ + { + "bbox": [ + 213, + 72, + 375, + 236 + ], + "type": "image", + "image_path": "84693baafeb673a4a06b73173a7d466c1058f5b6d800fd5ddd2c57e16b5d6884.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 378, + 72, + 541, + 236 + ], + "blocks": [ + { + "bbox": [ + 378, + 72, + 541, + 236 + ], + "lines": [ + { + "bbox": [ + 378, + 72, + 541, + 236 + ], + "spans": [ + { + "bbox": [ + 378, + 72, + 541, + 236 + ], + "type": "image", + "image_path": "b3ac5c2d1fe1add17df1d7316c8d4529441fe14af09f3e7d9fe151ceda2b1e0c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 342, + 111, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 111, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 111, + 354 + ], + "type": "text", + "content": "for fine-tuning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 354, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 460 + ], + "type": "text", + "content": "Figure 4 presents unlocked fine-tuning results alongside from-scratch randomly initialized baselines. We used 16 TPU-v4 chips and train at " + }, + { + "bbox": [ + 46, + 354, + 287, + 460 + ], + "type": "inline_equation", + "content": "16\\mathrm{k}" + }, + { + "bbox": [ + 46, + 354, + 287, + 460 + ], + "type": "text", + "content": " batch size for " + }, + { + "bbox": [ + 46, + 354, + 287, + 460 + ], + "type": "inline_equation", + "content": "2.4\\mathrm{B}" + }, + { + "bbox": [ + 46, + 354, + 287, + 460 + ], + "type": "text", + "content": " examples seen. We found that the fine-tuning setup doesn't perform well out-of-the-box; this is consistent with prior works [55] where finetuning image models degraded visual representation quality. This is evidenced by ImageNet 10-shot linear classification, where in Figure 4 the fine-tuned setup is barely better than the from-scratch baseline." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 461, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 641 + ], + "type": "text", + "content": "We hypothesize that the default weight decay applied to the pre-trained weights reduces their effectiveness. Motivated by the fine-tuning recipe from [17, 54, 25], that uses no weight decay, we also propose disabling weight decay on the pre-trained weights for SigLIP training. Weight decay is therefore only applied to the randomly initialized weights in the text model. This simple modification significantly improved SigLIP results. Figure 4 shows that with our improved recipe, SigLIP reaches " + }, + { + "bbox": [ + 46, + 461, + 287, + 641 + ], + "type": "inline_equation", + "content": "71\\%" + }, + { + "bbox": [ + 46, + 461, + 287, + 641 + ], + "type": "text", + "content": " 0-shot accuracy on ImageNet, using " + }, + { + "bbox": [ + 46, + 461, + 287, + 641 + ], + "type": "inline_equation", + "content": "16k" + }, + { + "bbox": [ + 46, + 461, + 287, + 641 + ], + "type": "text", + "content": " batch size, trained on 16 chips for three days. We also present from-scratch results in the bottom rows of Table 1: with 32 TPUv4 chips for only two days, SigLIP achieves " + }, + { + "bbox": [ + 46, + 461, + 287, + 641 + ], + "type": "inline_equation", + "content": "72.1\\%" + }, + { + "bbox": [ + 46, + 461, + 287, + 641 + ], + "type": "text", + "content": " 0-shot accuracy. This presents a significant training cost reduction e.g. compared to CLIP (approx. 2500 TPUv3-days for " + }, + { + "bbox": [ + 46, + 461, + 287, + 641 + ], + "type": "inline_equation", + "content": "72.6\\%" + }, + { + "bbox": [ + 46, + 461, + 287, + 641 + ], + "type": "text", + "content": ") reported in [29]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 647, + 155, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 155, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 155, + 659 + ], + "type": "text", + "content": "4.6. Scaling up SigLIP" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "In this section, we scale up SigLIP by \"overtraining\" the model [42, 1] for a long schedule. We present results in Table 3 using ViT-B or ViT-L as the vision encoder, while the text encoder is kept the same size (B or L respec" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 342, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 509 + ], + "type": "text", + "content": "tively). Following the recipe as described in Section 4.2, we train both models for 40 billion examples seen, but increase the number of image patches from " + }, + { + "bbox": [ + 304, + 342, + 545, + 509 + ], + "type": "inline_equation", + "content": "(224 / 16)^{2} = 196" + }, + { + "bbox": [ + 304, + 342, + 545, + 509 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 342, + 545, + 509 + ], + "type": "inline_equation", + "content": "(256 / 16)^{2} = 256" + }, + { + "bbox": [ + 304, + 342, + 545, + 509 + ], + "type": "text", + "content": " and the number of text tokens from 16 to 64. To get SigLIP results for different resolutions, we continue training the pre-trained model for 5 billion examples on the target resolutions (i.e. 224 or 384), using a " + }, + { + "bbox": [ + 304, + 342, + 545, + 509 + ], + "type": "inline_equation", + "content": "100\\mathrm{x}" + }, + { + "bbox": [ + 304, + 342, + 545, + 509 + ], + "type": "text", + "content": " smaller learning rate without weight decay. In Table 3, we report zero-shot classification results on ImageNet [14], ObjectNet [2], ImageNetv2 [36], ImageNet Real [3], and zero-shot image-to-text " + }, + { + "bbox": [ + 304, + 342, + 545, + 509 + ], + "type": "inline_equation", + "content": "(\\mathrm{I}\\rightarrow \\mathrm{T})" + }, + { + "bbox": [ + 304, + 342, + 545, + 509 + ], + "type": "text", + "content": " retrieval, text-to-image " + }, + { + "bbox": [ + 304, + 342, + 545, + 509 + ], + "type": "inline_equation", + "content": "(\\mathrm{I}\\rightarrow \\mathrm{T})" + }, + { + "bbox": [ + 304, + 342, + 545, + 509 + ], + "type": "text", + "content": " retrieval results on MSCOCO [11]. Our SigLIP models outperform the previous models consistently with ViT-B and ViT-L architectures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 515, + 475, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 515, + 475, + 528 + ], + "spans": [ + { + "bbox": [ + 305, + 515, + 475, + 528 + ], + "type": "text", + "content": "4.7. Stabilizing large-batch training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 534, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 665 + ], + "type": "text", + "content": "As we move to large batch sizes, the language image pretraining using transformers becomes increasingly more unstable, even when using a modestly-sized model (e.g. Base size). The reason for these instabilities is large spikes in the gradient norms, which translate to large-magnitude changes in the weights that may destabilize the training process, see Figure 5. We observe that reducing " + }, + { + "bbox": [ + 304, + 534, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 304, + 534, + 545, + 665 + ], + "type": "text", + "content": " in Adam and AdaFactor from its default 0.999 to 0.95 (which was suggested in [20, 9]) is enough to stabilize the training. Intuitively, this allows recovering from gradient spikes quicker. We opt for setting " + }, + { + "bbox": [ + 304, + 534, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.95" + }, + { + "bbox": [ + 304, + 534, + 545, + 665 + ], + "type": "text", + "content": " for all our experiments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 671, + 465, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 465, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 465, + 684 + ], + "type": "text", + "content": "4.8. Negative ratio in sigmoid loss" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "One question which arises when shifting the perspective from the softmax's \"pick the right class\" view to the sig-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11981" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 545, + 268 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 268 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 268 + ], + "type": "table", + "html": "
MethodImage EncoderImageNet-1kCOCO
SizePatch #Validationv2RealObjectNetI→TT→I
CLIPB19668.361.9-55.352.433.1
OpenCLIPB19670.262.3-56.059.442.3
EVA-CLIPB19674.767.0-62.358.742.2
SigLIPB19676.369.682.870.764.447.2
SigLIPB25676.670.083.171.365.147.4
CLIPL25675.569.0-69.956.336.5
OpenCLIPL25674.061.1-66.462.146.1
EVA-CLIPL25679.875.3-72.963.747.5
SigLIPL25680.674.285.977.969.551.1
CLIPL57676.672.0-70.957.937.1
EVA-CLIPL57680.473.8-78.464.147.9
", + "image_path": "512948c27dcfb58d9014b092a0d0b06d6e8aec92239559569990b9672142d5b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 276, + 547, + 326 + ], + "lines": [ + { + "bbox": [ + 46, + 276, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 547, + 326 + ], + "type": "text", + "content": "Table 3: Comparison with other methods. Our SigLIP models outperform the previous models, e.g. OpenCLIP [22] and CLIP [34], by a significant margin on both the zero-shot classification tasks and the zero-shot retrieval tasks. Compared to the recently introduced EVA-CLIP [40] model, SigLIP-L outperforms EVA-CLIP-L on all the tasks with the same number of 256 patches. SigLIP-L (256 patches) also outperforms EVA-CLIP-L (576 patches) by a large margin on the retrieval tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 346, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 502 + ], + "type": "text", + "content": "moid's \"rate this pair\" view, is the imbalance in positive versus negative pairs. For a batch size " + }, + { + "bbox": [ + 46, + 346, + 287, + 502 + ], + "type": "inline_equation", + "content": "|\\mathcal{B}|" + }, + { + "bbox": [ + 46, + 346, + 287, + 502 + ], + "type": "text", + "content": ", the batch contains " + }, + { + "bbox": [ + 46, + 346, + 287, + 502 + ], + "type": "inline_equation", + "content": "|\\mathcal{B}|" + }, + { + "bbox": [ + 46, + 346, + 287, + 502 + ], + "type": "text", + "content": " positive pairs, but " + }, + { + "bbox": [ + 46, + 346, + 287, + 502 + ], + "type": "inline_equation", + "content": "|\\mathcal{B}|^2 - |\\mathcal{B}|" + }, + { + "bbox": [ + 46, + 346, + 287, + 502 + ], + "type": "text", + "content": " negative examples. In the modest batch-size of " + }, + { + "bbox": [ + 46, + 346, + 287, + 502 + ], + "type": "inline_equation", + "content": "16\\mathrm{k}" + }, + { + "bbox": [ + 46, + 346, + 287, + 502 + ], + "type": "text", + "content": ", there are actually " + }, + { + "bbox": [ + 46, + 346, + 287, + 502 + ], + "type": "inline_equation", + "content": "268\\mathrm{M}" + }, + { + "bbox": [ + 46, + 346, + 287, + 502 + ], + "type": "text", + "content": " negative examples for only " + }, + { + "bbox": [ + 46, + 346, + 287, + 502 + ], + "type": "inline_equation", + "content": "16\\mathrm{k}" + }, + { + "bbox": [ + 46, + 346, + 287, + 502 + ], + "type": "text", + "content": " positive ones. At the same time, because the sigmoid loss decomposes into a sum of per-example losses, we can perform controlled experiments to study the effect of the mini-batch composition and distribution of examples visited. We run experiments in the SigLiT setup at batch-size " + }, + { + "bbox": [ + 46, + 346, + 287, + 502 + ], + "type": "inline_equation", + "content": "16\\mathrm{k}" + }, + { + "bbox": [ + 46, + 346, + 287, + 502 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 346, + 287, + 502 + ], + "type": "inline_equation", + "content": "900\\mathrm{M}" + }, + { + "bbox": [ + 46, + 346, + 287, + 502 + ], + "type": "text", + "content": " steps and vary the composition of the batch by masking out (i.e. ignoring) enough negative examples to reach a target \"positive: negative\" ratio, masking in the following ways:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 506, + 286, + 624 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 58, + 506, + 278, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 506, + 278, + 517 + ], + "spans": [ + { + "bbox": [ + 58, + 506, + 278, + 517 + ], + "type": "text", + "content": "- Random: Randomly choose negative pairs to mask." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 521, + 267, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 521, + 267, + 533 + ], + "spans": [ + { + "bbox": [ + 58, + 521, + 267, + 533 + ], + "type": "text", + "content": "- Hard: Keep hardest negative pairs (highest loss)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 538, + 263, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 538, + 263, + 549 + ], + "spans": [ + { + "bbox": [ + 58, + 538, + 263, + 549 + ], + "type": "text", + "content": "- Easy: Keep easiest negatives pairs (lowest loss)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 553, + 286, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 553, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 58, + 553, + 286, + 624 + ], + "type": "text", + "content": "- Hard + matching total pairs seen: Masking examples while training for a fixed number of steps does decrease the total number of pairs seen during training. Hence in the matched pairs setting, we increase the number of training steps by the masking ratio in order to keep the number of pairs seen constant." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "Figure 6 shows the effect of the various masking strategies. Randomly removing negatives to rebalance does deteriorate performance. Keeping the easiest examples does not work at all, while keeping the hardest negatives does almost maintain the quality, indicating that, as could be expected, a lot of the learning on the negative side comes from the harder examples. This is further confirmed by the slightly" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 346, + 545, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 545, + 370 + ], + "type": "text", + "content": "increased performance of training longer on the hardest examples in order to match the total pairs seen." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 381, + 545, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 545, + 466 + ], + "type": "text", + "content": "We also look at the value of the learned bias at the end of training as well as the average logit value for positive and negative examples across these settings, and find the result mostly follows what one would expect: as fewer negatives are present, the bias and logits become more positive overall. Interestingly, when training with more hard negative pairs, the average logits of positive pairs stays mostly flat." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 477, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 545, + 525 + ], + "type": "text", + "content": "This study confirms that (1) the imbalance does not seem to be a major reason for concern, while at the same time (2) coming up with an efficient way of including more negative examples can be promising but is not trivial." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 625, + 545, + 715 + ], + "blocks": [ + { + "bbox": [ + 304, + 567, + 545, + 615 + ], + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 615 + ], + "type": "text", + "content": "Table 4: Bias (b) and temperature " + }, + { + "bbox": [ + 304, + 567, + 545, + 615 + ], + "type": "inline_equation", + "content": "(\\mathbf{t}^{\\prime})" + }, + { + "bbox": [ + 304, + 567, + 545, + 615 + ], + "type": "text", + "content": " initialization. Results are reported using Base architecture, " + }, + { + "bbox": [ + 304, + 567, + 545, + 615 + ], + "type": "inline_equation", + "content": "8\\mathrm{k}" + }, + { + "bbox": [ + 304, + 567, + 545, + 615 + ], + "type": "text", + "content": " batch size, trained for 900M examples. Enabling the bias term b with -10 initialization improves results consistently." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 625, + 545, + 715 + ], + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 715 + ], + "type": "table", + "html": "
bt'INet-0Pet-0C100-0
n/alog 1062.081.859.9
-10log 1063.082.461.0
-10log 161.080.060.4
0log 1061.779.959.0
0log 153.773.253.8
", + "image_path": "d38ca012e323a40f75e0f1bb6e9119930055bcb53648f3d1c1555f4e73709fef.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11982" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 545, + 186 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 545, + 186 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 545, + 186 + ], + "type": "image", + "image_path": "f3971470a9b627315836ee46a6b56f5162dba6ad57ca79f020f21438ca925d6f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 547, + 236 + ], + "type": "text", + "content": "Figure 7: Sigmoid-training increases robustness to data noise. Titles show the type of corruption applied, and x-axes show the probability with which they are applied. With increasing corruption severity, M-scale models trained with sigmoid loss for 3.6 billion examples retain superiority over corresponding softmax baseline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 255, + 187, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 187, + 268 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 187, + 268 + ], + "type": "text", + "content": "4.9. Bias term in sigmoid loss" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 274, + 288, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 288, + 346 + ], + "type": "text", + "content": "We ablate the bias term in the loss function, using the Base architecture with an " + }, + { + "bbox": [ + 46, + 274, + 288, + 346 + ], + "type": "inline_equation", + "content": "8\\mathrm{k}" + }, + { + "bbox": [ + 46, + 274, + 288, + 346 + ], + "type": "text", + "content": " batch size, trained for 900M examples with the SigLIP setup. Zero-shot transfer results are reported on ImageNet [14], Oxford-iit pet [32] and Cifar100 [26]. Table 4 presents results with and without a bias term in the sigmoid loss." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 346, + 287, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 478 + ], + "type": "text", + "content": "Enabling the bias term with a " + }, + { + "bbox": [ + 46, + 346, + 287, + 478 + ], + "type": "inline_equation", + "content": "-10" + }, + { + "bbox": [ + 46, + 346, + 287, + 478 + ], + "type": "text", + "content": " initialization consistently improves performance across all tasks. This is because the bias term ensures that the training starts close to the prior, preventing dramatic over-correction in early optimization. In contrast, a randomly chosen bias term initialization, such as the 0 initialization in Table 4, fails to address the over-correction issue, leading to significantly worse results. This effect is particularly noticeable when using a small temperature " + }, + { + "bbox": [ + 46, + 346, + 287, + 478 + ], + "type": "inline_equation", + "content": "t^{\\prime}" + }, + { + "bbox": [ + 46, + 346, + 287, + 478 + ], + "type": "text", + "content": " initialization. We set the bias and temperature initialization to " + }, + { + "bbox": [ + 46, + 346, + 287, + 478 + ], + "type": "inline_equation", + "content": "b = -10" + }, + { + "bbox": [ + 46, + 346, + 287, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 346, + 287, + 478 + ], + "type": "inline_equation", + "content": "t^\\prime = \\log 10" + }, + { + "bbox": [ + 46, + 346, + 287, + 478 + ], + "type": "text", + "content": " (hence " + }, + { + "bbox": [ + 46, + 346, + 287, + 478 + ], + "type": "inline_equation", + "content": "t = 10" + }, + { + "bbox": [ + 46, + 346, + 287, + 478 + ], + "type": "text", + "content": ") as the default for all experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 487, + 181, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 181, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 181, + 498 + ], + "type": "text", + "content": "4.10. Label noise robustness" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "type": "text", + "content": "Prior works demonstrated improved robustness against label noise when using the sigmoid loss for classification models [3]. This property would be particularly useful here in the face of the famously noisy nature of popular large-scale image-text datasets. In order to study this for SigLIP, we train M/16 image models alongside an M text model at batch size 16384 for 3.6 billion seen examples. We corrupt the training data using one of the following methods:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 605, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 58, + 605, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 605, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 58, + 605, + 287, + 628 + ], + "type": "text", + "content": "- Image: With probability " + }, + { + "bbox": [ + 58, + 605, + 287, + 628 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 58, + 605, + 287, + 628 + ], + "type": "text", + "content": ", replace the image with uniform random noise." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 633, + 287, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 633, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 58, + 633, + 287, + 669 + ], + "type": "text", + "content": "- Text: With probability " + }, + { + "bbox": [ + 58, + 633, + 287, + 669 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 58, + 633, + 287, + 669 + ], + "type": "text", + "content": ", replace tokenized text with a new sequence of randomly sampled tokens, up to some (sampled) sequence length." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 673, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 673, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 58, + 673, + 287, + 696 + ], + "type": "text", + "content": "- Batch alignment: Randomly shuffle the ordering of " + }, + { + "bbox": [ + 58, + 673, + 287, + 696 + ], + "type": "inline_equation", + "content": "p\\%" + }, + { + "bbox": [ + 58, + 673, + 287, + 696 + ], + "type": "text", + "content": " of the batch." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 701, + 274, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 701, + 274, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 701, + 274, + 713 + ], + "type": "text", + "content": "- Image & text: Apply both with probability " + }, + { + "bbox": [ + 58, + 701, + 274, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 58, + 701, + 274, + 713 + ], + "type": "text", + "content": " each." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 317, + 255, + 545, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 255, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 317, + 255, + 545, + 279 + ], + "type": "text", + "content": "- Image, text & batch: Alongside (4), also shuffle fraction " + }, + { + "bbox": [ + 317, + 255, + 545, + 279 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 317, + 255, + 545, + 279 + ], + "type": "text", + "content": " of alignments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 283, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 283, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 283, + 545, + 319 + ], + "type": "text", + "content": "Results from varying the likelihood of the corruption are shown in Figure 7. Models trained with sigmoid loss are increasingly robust to all kinds of added noise." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 330, + 378, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 330, + 378, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 330, + 378, + 342 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 350, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 547, + 529 + ], + "type": "text", + "content": "We conducted a study on two language-image pretraining instances that used the sigmoid loss: SigLiT and SigLIP. Our results demonstrate that the sigmoid loss performs better than the softmax baseline, particularly for small train batch sizes. This loss function is also more memory efficient, which allows larger train batch sizes without requiring additional resources. We performed a thorough investigation of the batch size in contrastive learning. Surprisingly, we found that a relatively modest batch size of " + }, + { + "bbox": [ + 304, + 350, + 547, + 529 + ], + "type": "inline_equation", + "content": "32\\mathrm{k}" + }, + { + "bbox": [ + 304, + 350, + 547, + 529 + ], + "type": "text", + "content": " yielded nearly optimal performance. Further studies have been performed to understand better the introduced bias term in the sigmoid loss, robustness to data noises and the impact of positive and negative pairs ratio in the sigmoid loss. We hope this work will facilitate language-image pretraining research with limited resources." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 544, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 712 + ], + "type": "text", + "content": "Acknowledgements. We thank Daniel Keysers, Ilya Tolstikhin, Olivier Bousquet and Michael Tschannen for their valuable feedback and discussions on this paper. We thank Joan Puigcerver, Josip Djolonga and Black Hechtman for discussions on efficient implementations of the chunked contrastive loss. We thank Kaiming He and Xinlei Chen for the discussion of " + }, + { + "bbox": [ + 304, + 544, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 304, + 544, + 547, + 712 + ], + "type": "text", + "content": " to stabilize the training. We also thank Ross Wightman for spotting a mistake in the pseudocode in the first version of this paper. Similarly, we thank Boris Dayma for spotting a typo making " + }, + { + "bbox": [ + 304, + 544, + 547, + 712 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 544, + 547, + 712 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 304, + 544, + 547, + 712 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 304, + 544, + 547, + 712 + ], + "type": "text", + "content": " confusing, which we fixed in the third version of this paper. We thank the Google Deepmind team for providing a supportive research environment. We use the big-vision codebase [5, 4] for all experiments in this project." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11983" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Ibrahim Alabdulmohsin, Xiaohua Zhai, Alexander Kolesnikov, and Lucas Beyer. Getting vit in shape: Scaling laws for compute-optimal model design. In NeurIPS, 2023. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "text", + "content": "[2] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. ObjectNet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. In NeurIPS, 2019. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 287, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 226 + ], + "type": "text", + "content": "[3] Lucas Beyer, Olivier J. Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with imagenet? CoRR, abs/2006.07159, 2020. 2, 7, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 287, + 250 + ], + "type": "text", + "content": "[4] Lucas Beyer, Xiaohua Zhai, and Alexander Kolesnikov. Better plain vit baselines for imagenet-1k, 2022. 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 287, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 287, + 283 + ], + "type": "text", + "content": "[5] Lucas Beyer, Xiaohua Zhai, and Alexander Kolesnikov. Big vision. https://github.com/google-research/big_vision, 2022.9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 285, + 286, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 286, + 329 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 286, + 329 + ], + "type": "text", + "content": "[6] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 331, + 286, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 286, + 374 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 286, + 374 + ], + "type": "text", + "content": "[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 376, + 286, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 286, + 419 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 286, + 419 + ], + "type": "text", + "content": "[8] Feilong Chen, Duzhen Zhang, Minglun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. VLP: A survey on vision-language pre-training. Int. J. Autom. Comput., 20(1):38-56, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 422, + 286, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 422, + 286, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 422, + 286, + 497 + ], + "type": "text", + "content": "[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 1691-1703. PMLR, 2020. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 500, + 286, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 286, + 533 + ], + "type": "text", + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 535, + 286, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 286, + 578 + ], + "type": "text", + "content": "[11] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. CoRR, abs/1504.00325, 2015. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 580, + 286, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 286, + 624 + ], + "type": "text", + "content": "[12] Xiangning Chen, Chen Liang, Da Huang, Esteban Real, Kaiyuan Wang, Yao Liu, Hieu Pham, Xuanyi Dong, Thang Luong, Cho-Jui Hsieh, Yifeng Lu, and Quoc V. Le. Symbolic discovery of optimization algorithms, 2023. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 625, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 286, + 713 + ], + "type": "text", + "content": "[13] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Radu Soricut. Pali: A jointly-scaled multilingual language-image model. CoRR, abs/2209.06794, 2022. 1, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "text", + "content": "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 4, 7, 9" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 130, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 206 + ], + "type": "text", + "content": "[15] Karan Desai, Gaurav Kaul, Zubin Aysola, and Justin Johnson. Redcaps: Web-curated image-text data created by the people, for the people. In Joaquin Vanschoren and SaiKit Yeung, editors, Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks 1, NeurIPS Datasets and Benchmarks 2021, December 2021, virtual, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "text", + "content": "[16] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Shuyang Gu, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Clip itself is a strong fine-tuner: Achieving " + }, + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "inline_equation", + "content": "85.7\\%" + }, + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "inline_equation", + "content": "88.0\\%" + }, + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "text", + "content": " top-1 accuracy with vit-b and vit-l onImagenet. CoRR, abs/2212.06138, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 264, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 330 + ], + "type": "text", + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth " + }, + { + "bbox": [ + 308, + 264, + 545, + 330 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 308, + 264, + 545, + 330 + ], + "type": "text", + "content": " words: Transformers for image recognition at scale. In ICLR, 2021. 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 332, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 364 + ], + "type": "text", + "content": "[18] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016. http://www.deeplearningbook.org.1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 365, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 398 + ], + "type": "text", + "content": "[19] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In CVPR, volume 2, 2006. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 399, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 464 + ], + "type": "text", + "content": "[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15979-15988. IEEE, 2022. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 467, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 510 + ], + "type": "text", + "content": "[21] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. Scaling up vision-language pre-training for image captioning. CoRR, abs/2111.12233, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "text", + "content": "[22] Gabriel Ilharco, Mitchell Wortsman, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. OpenCLIP. Zenodo, 2021. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 557, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 601 + ], + "type": "text", + "content": "[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 602, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 689 + ], + "type": "text", + "content": "[24] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In Hugo Larochelle, Marc'Aurelio Ranzato, Raia Hadsell, MariaFlorina Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[25] Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, and Neil Houlsby." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11984" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "Big transfer (BiT): General visual representation learning. In ECCV, 2020. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 117 + ], + "type": "text", + "content": "[26] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, Univ. of Toronto, 2009. 9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "text", + "content": "[27] Taku Kudo and John Richardson. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In EMNLP, 2018. 5, 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 151, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 151, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 49, + 151, + 287, + 249 + ], + "type": "text", + "content": "[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors, International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 of Proceedings of Machine Learning Research, pages 12888-12900. PMLR, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 250, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 250, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 49, + 250, + 287, + 282 + ], + "type": "text", + "content": "[29] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. CoRR, abs/2212.00794, 2022. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 284, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 284, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 49, + 284, + 287, + 403 + ], + "type": "text", + "content": "[30] Matthias Minderer, Alexey A. Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple open-vocabulary object detection. In Shai Avidan, Gabriel J. Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, volume 13670 of Lecture Notes in Computer Science, pages 728-755. Springer, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 404, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 404, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 49, + 404, + 287, + 448 + ], + "type": "text", + "content": "[31] Jishnu Mukhoti, Tsung-Yu Lin, Amid Poursaeed, Rui Wang, Ashish Shah, Philip H. S. Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 449, + 287, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 449, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 49, + 449, + 287, + 481 + ], + "type": "text", + "content": "[32] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, 2012. 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 482, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 482, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 49, + 482, + 287, + 525 + ], + "type": "text", + "content": "[33] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Hanxiao Liu, Adams Wei Yu, Minh-Thang Luong, Mingxing Tan, and Quoc V. Le. Combined scaling for zero-shot transfer learning. CoRR, abs/2111.10050, 2021. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 526, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 526, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 49, + 526, + 287, + 591 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 592, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 592, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 49, + 592, + 287, + 635 + ], + "type": "text", + "content": "[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv e-prints, 2019. 5, 13" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 636, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 636, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 49, + 636, + 287, + 669 + ], + "type": "text", + "content": "[36] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In ICML, 2019. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "text", + "content": "[37] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "text", + "content": "Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. CoRR, abs/2210.08402, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "type": "text", + "content": "[38] Krishna Srinivasan, Karthik Raman, Jiecao Chen, Michael Bendersky, and Marc Najork. WIT: wikipedia-based image text dataset for multimodal multilingual machine learning. CoRR, abs/2103.01913, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 163, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 207 + ], + "type": "text", + "content": "[39] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your ViT? Data, augmentation, and regularization in vision transformers. CoRR, abs/2106.10270, 2021. 1, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 209, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 241 + ], + "type": "text", + "content": "[40] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. EVA-CLIP: improved training techniques for CLIP at scale. CoRR, abs/2303.15389, 2023. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 243, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 330 + ], + "type": "text", + "content": "[41] Ashish V. Thapliyal, Jordi Pont-Tuset, Xi Chen, and Radu Soricut. Crossmodal-3600: A massively multilingual multimodal evaluation dataset. In Yoav Goldberg, Zornitsa Kozareva, and Yue Zhang, editors, Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022, pages 715-729. Association for Computational Linguistics, 2022. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 331, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 396 + ], + "type": "text", + "content": "[42] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. CoRR, abs/2302.13971, 2023. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 399, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 430 + ], + "type": "text", + "content": "[43] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, abs/1807.03748, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 433, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 466 + ], + "type": "text", + "content": "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 467, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 510 + ], + "type": "text", + "content": "[45] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. GIT: A generative image-to-text transformer for vision and language. CoRR, abs/2205.14100, 2022. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 512, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 577 + ], + "type": "text", + "content": "[46] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 579, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 611 + ], + "type": "text", + "content": "[47] Ross Wightman, Hugo Touvron, and Hervé Jégou. Resnet strikes back: An improved training procedure in timm. CoRR, abs/2110.00476, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "text", + "content": "[48] Mitchell Wortsman. Reaching " + }, + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "text", + "content": " zero-shot accuracy with OpenCLIP: VIT-G/14 trained on LAION-2B. https://web.archive.org/web/20230127012732/https://laion.ai/blog/giant-openclip/.2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[49] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. In IEEE/CVF Conference on Computer" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11985" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 551 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "text", + "content": "Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 7949-7961. IEEE, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[50] Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. mT5: A massively multilingual pre-trained text-to-text transformer. In NAACL-HLT, 2021. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 217 + ], + "type": "text", + "content": "[51] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Bin Xiao, Ce Liu, Lu Yuan, and Jianfeng Gao. Unified contrastive learning in image-text-label space. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 19141-19151. IEEE, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "text", + "content": "[52] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. CoRR, abs/2205.01917, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 287, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 340 + ], + "type": "text", + "content": "[53] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pengchuan Zhang. Florence: A new foundation model for computer vision. CoRR, abs/2111.11432, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 373 + ], + "type": "text", + "content": "[54] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. CVPR, 2022. 1, 4, 6, 7, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 375, + 287, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 375, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 48, + 375, + 287, + 450 + ], + "type": "text", + "content": "[55] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 18102-18112. IEEE, 2022. 1, 2, 3, 4, 6, 7, 13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 453, + 287, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 287, + 551 + ], + "type": "text", + "content": "[56] Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D. Manning, and Curtis P. Langlotz. Contrastive learning of medical visual representations from paired images and text. In Zachary C. Lipton, Rajesh Ranganath, Mark P. Sendak, Michael W. Sjoding, and Serena Yeung, editors, Proceedings of the Machine Learning for Healthcare Conference, MLHC 2022, 5-6 August 2022, Durham, NC, USA, volume 182 of Proceedings of Machine Learning Research, pages 2-25. PMLR, 2022. 2" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11986" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Sign Language Translation with Iterative Prototype/f2c1bf6d-f55b-40d7-9372-42bff493c6d3_content_list.json b/2023/Sign Language Translation with Iterative Prototype/f2c1bf6d-f55b-40d7-9372-42bff493c6d3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b1435caa30108fd06b73ee096f8499f5d746742a --- /dev/null +++ b/2023/Sign Language Translation with Iterative Prototype/f2c1bf6d-f55b-40d7-9372-42bff493c6d3_content_list.json @@ -0,0 +1,1620 @@ +[ + { + "type": "text", + "text": "Sign Language Translation with Iterative Prototype", + "text_level": 1, + "bbox": [ + 222, + 130, + 746, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huijie Yao $^{1}$ Wengang Zhou $^{1,2,*}$ Hao Feng $^{1}$ Hezhen Hu $^{1}$ Hao Zhou $^{1}$ Houqiang Li $^{1,2,*}$", + "bbox": [ + 104, + 178, + 866, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ CAS Key Laboratory of Technology in GIPAS, EEIS Department, University of Science and Technology of China $^{2}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center", + "bbox": [ + 104, + 199, + 864, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yaohuijie,haof,alexhu,zhouh156}@mail.ustc.edu.cn，{zhwg,lihq}@ustc.edu.cn", + "bbox": [ + 158, + 234, + 810, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper presents IP-SLT, a simple yet effective framework for sign language translation (SLT). Our IP-SLT adopts a recurrent structure and enhances the semantic representation (prototype) of the input sign language video via an iterative refinement manner. Our idea mimics the behavior of human reading, where a sentence can be digested repeatedly, till reaching accurate understanding. Technically, IP-SLT consists of feature extraction, prototype initialization, and iterative prototype refinement. The initialization module generates the initial prototype based on the visual feature extracted by the feature extraction module. Then, the iterative refinement module leverages the cross-attention mechanism to polish the previous prototype by aggregating it with the original video feature. Through repeated refinement, the prototype finally converges to a more stable and accurate state, leading to a fluent and appropriate translation. In addition, to leverage the sequential dependence of prototypes, we further propose an iterative distillation loss to compress the knowledge of the final iteration into previous ones. As the autoregressive decoding process is executed only once in inference, our IP-SLT is ready to improve various SLT systems with acceptable overhead. Extensive experiments are conducted on public benchmarks to demonstrate the effectiveness of the IP-SLT.", + "bbox": [ + 75, + 318, + 473, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 712, + 209, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sign language translation (SLT) aims to automatically generate spoken language translations based on sign language videos, which holds both social significance and academic value. On the one hand, a high-quality SLT system can greatly facilitate communication between deafmute and hearing individuals [4, 11, 12, 41]. On the other hand, SLT as an interdisciplinary research topic necessitates a comprehensive understanding of computer vision [31, 13, 36, 19] and natural language processing [43, 40], given its", + "bbox": [ + 75, + 738, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d3b242dc929ef28a6c798e2106b9377eacab9351036ce2d8ede668759cf7ef71.jpg", + "image_caption": [ + "(a) The Previous Pipeline" + ], + "image_footnote": [], + "bbox": [ + 516, + 287, + 877, + 323 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fb659078773baacc81bce84eef610a8c414e47deea8c8d3b6e8d2a152c3510af.jpg", + "image_caption": [ + "(b) Our IP-SLT Pipeline", + "Figure 1. Illustration of the pipeline of the previous works and our IP-SLT. (a) The previous studies rely on a one-pass forward process to generate the final translation. (b) To mitigate the vision-text gap, we introduce the iterative refinement module into the original SLT system. The refinement module updates the current prototype conditioned on the sign language video, which can be run iteratively to obtain a better representation of the semantic meaning of the sign language video." + ], + "image_footnote": [], + "bbox": [ + 516, + 349, + 877, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "involvement with vision and text modalities. As a result, SLT has emerged as a vital research topic, garnering increasing attention [5, 49, 31, 2, 21, 20, 42].", + "bbox": [ + 496, + 594, + 892, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SLT is a challenging task, which faces a tough domain gap between the input video and output text, as well as a limited dataset scale due to costly data collection and annotation [49, 7, 44, 21]. Since SLT is typically viewed as a sequence-to-sequence mapping problem, the existing SLT systems [23, 5, 7] commonly adopt the one-pass forward pipeline based on encoder-decoder architecture [40, 30] (as shown in Fig. 1 (a)). In such a framework, the encoder transforms the sign video into its semantic representation (prototype), which is then fed into the decoder to obtain the final translation. However, due to the inherent gap between vision and text, it may be hard to conduct such mapping within the vanilla one-pass architecture.", + "bbox": [ + 496, + 641, + 893, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this study, we present IP-SLT with the iterative prototype to boost sign language translation (as shown in Fig. 1 (b)), which is inspired by the human reading process. During this process, we note that repeatedly digging into the", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors: Wengang Zhou and Houqiang Li", + "bbox": [ + 94, + 886, + 405, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15592", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "source materials is necessary for accurate understanding. Similarly, when we are trying to translate a sign language video into a sentence, we commonly do not directly write it down. Instead, we would recall and go back to the original sign video to check our answers. To implement the above idea, our IP-SLT adopts a recurrent structure that enhances the semantic representation (prototype) of the input sign language video via an iterative refinement process. IP-SLT generally contains three main components, including feature extraction, prototype initialization, and iterative prototype refinement. Given a sign video to be translated, we first extract its visual representation, which is then used for generating an initial raw prototype. Subsequently, we iteratively leverage the attention mechanism [43] to update the prototype toward the semantic meaning of the sign video. At each iteration, we refine the previous prototype by aggregating it with the original visual representation. In this way, the network repeatedly digs the semantic context from the sign video to polish the prototype. Through iterative refinement, the prototype finally converges to a stable and accurate state, producing a high-quality translation.", + "bbox": [ + 76, + 90, + 472, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, our IP-SLT introduces a novel design discussed next. Firstly, to leverage the sequential dependence between different iterations, we further propose the iterative distillation loss which allows the previous prototypes to obtain supervision from the final one. Since the final prototype converges to a more stable and accurate state, it is possible for IP-SLT to achieve better performance. Secondly, during training, all predicted prototypes are transformed into their corresponding translations to provide guidance for each iteration. Our inference process is neat since only the final prototype is used for the autoregressive decoding process. Thirdly, our IP-SLT can easily work with different visual backbones. Through end-to-end optimization, our IP-SLT achieves significant performance improvements over the baselines.", + "bbox": [ + 76, + 407, + 472, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are three-fold:", + "bbox": [ + 96, + 633, + 395, + 648 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose IP-SLT, a novel framework to ameliorate sign language translation, which iteratively refines the prototypes by aggregating the previous translation progress and the original visual representation.", + "- We propose an iterative distillation loss to enhance the basic supervision, by leveraging the sequential dependence between the outputs at each iteration.", + "- We conduct extensive experiments to validate the proposed method, and show encouraging improved results on the two prevalent benchmarks, i.e., CSL-Daily [49] and PHOENIX-2014T [5]." + ], + "bbox": [ + 93, + 655, + 468, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 845, + 218, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we briefly review the related works, i.e., sign language translation and iterative refinement methods.", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sign Language Translation. Camgoz et al. [5] pioneer the neural SLT task and publish the neural dataset PHOENIX-2014T and regard the SLT as a sequence-to-sequence problem. They implement the neural SLT system using the encoder-decoder paradigm [3]. This paradigm is adopted by subsequent studies which focus on addressing the challenges of data scarcity and domain gap. Considering the lack of frame-level annotation in sign language datasets, Li et al. [27] design the temporal semantic pyramid structure to obtain more discriminative features. Camgoz et al. [7] explore the mutual benefits of SLT and continuous sign language recognition through joint optimization. Zhou et al. [49] leverage gloss annotation to transform the monolingual texts into pseudo-videos. According to the characteristic of sign language, several works [46, 6, 50] propose multi-channel SLT systems which explicitly extract and align the key parts of sign language expression. Jin et al. [23] leverage the additional prior knowledge to obtain high-quality translations. Chen et al. [9] propose a transfer learning baseline for SLT by leveraging external resources from related tasks. Chen et al. [10] further combine the raw videos and the keypoint sequences to achieve better semantic understanding with auxiliary supervision.", + "bbox": [ + 496, + 90, + 890, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, our proposed approach employs an iterative refinement process that utilizes the previous prototype as an additional clue to enhance the accuracy of the mapping between sign videos and their translations.", + "bbox": [ + 496, + 439, + 890, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Iterative Refinement Methods. The idea of iterative refinement is applied to various computer vision tasks, such as image generation [1, 18, 37], instance segmentation [34, 28, 29, 47], image rectification [15], etc., which shows promising performance improvements. CARN [1] maintains training stability in super-resolution tasks and improves the quality of output images. Ling et al. [28] regard the object instance segmentation task as a regression task and proposes the Curve-GCN to iteratively predict the locations of all vertices. DeepSanke [34] uses the deep network to iteratively enclose the object boundary based on an initial contour. The previous studies [48, 35, 14, 36, 31] in continuous sign language recognition task follow an iterative training scheme to enhance the discriminative power of feature extraction modules which use the convolutional neural network and their variants [8, 22, 39]. They leverage alignment proposals given by the connectionist temporal classification (CTC) [17] decoding as supervision at frame-wise granularity, which cannot be directly applied to the SLT.", + "bbox": [ + 496, + 506, + 890, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Different from the aforementioned methods, we explore how to reduce the vision-text gap by proposing an iterative refinement module to the existing SLT system. To reduce the complexity, we design the refinement module in a shared-weight manner. Moreover, we put forward the iterative distillation loss to leverage the sequential dependence between different iterations.", + "bbox": [ + 496, + 795, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15593", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/848452f5099d44eb4ba752e0c5e7717723377bd198fca8e83909ab95951806e3.jpg", + "image_caption": [ + "Figure 2. An overview of the proposed IP-SLT framework. Given a sign video $\\mathbf{X}$ , the feature extraction module is responsible for embedding the input into visual representation $\\mathbf{F}$ . The initialization module (the encoder $\\mathcal{E}_1$ and decoder $\\mathcal{D}_1$ ) generates the initial prototype $\\mathbf{E}^0$ and the raw translation $\\mathbf{Y}^0$ . The refinement module (the encoder $\\mathcal{E}_2$ and decoder $\\mathcal{D}_2$ ) first takes the initial prototype $\\mathbf{E}^0$ as input and generates a prototype for the current step by fusing it with the original visual representation $\\mathbf{F}$ . Through $K$ times refinement, the prototype sequence $\\mathbf{E} = \\{\\mathbf{E}^0, \\mathbf{E}^1, \\dots, \\mathbf{E}^K\\}$ and corresponding translation sequence $\\mathbf{Y} = \\{\\mathbf{Y}^0, \\mathbf{Y}^1, \\dots, \\mathbf{Y}^K\\}$ are obtained. In light of the fact that the decoding part of IP-SLT consists of $K + 1$ branches based on the iteration order, we introduce the iterative distillation loss to improve the underlying supervision. It should be noted that the parts enclosed in dashed boxes can be removed in inference." + ], + "image_footnote": [], + "bbox": [ + 143, + 88, + 831, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 382, + 212, + 400 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce the overall architecture of our IP-SLT, and then separately elaborate individual components. Finally, we propose the design of the training objective and inference strategy for the IP-SLT.", + "bbox": [ + 75, + 409, + 470, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Framework Overview", + "text_level": 1, + "bbox": [ + 76, + 479, + 282, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The primary objective of the SLT system is to acquire knowledge about the mapping $f: \\mathcal{X} \\mapsto \\mathcal{Y}$ , where $\\mathcal{X}$ and $\\mathcal{Y}$ denote the collections of $N$ sign language videos and spoken language sentences associated with vocabulary $\\mathcal{V}$ , respectively. Most SLT systems adopt the encoder-decoder architecture [40], where the input $X \\in \\mathcal{X}$ is first encoded to derive a high-level context representation. It is then passed to the decoder to generate the output $Y \\in \\mathcal{Y}$ . The encoder and decoder can be specialized using different types of neural networks, such as GRU [3], CNN [16], and Transformer [43]. Considering the performance of existing SLT systems, we adopt the Transformer as well.", + "bbox": [ + 75, + 505, + 468, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the goal of narrowing the domain gap between vision and text, we augment the original translation process with an iterative refinement step. Fig. 2 provides an overview of the proposed IP-SLT model, which consists of three stages, namely feature extraction, prototype initialization, and iterative prototype refinement. As with previous approaches, the initialization and iterative refinement module adopt the encoder-decoder architecture. Given the sign language video $\\mathbf{X} = \\{\\pmb{x}_t\\}_{t=1}^{T_x}$ with $T_x$ frames, the feature extraction module embeds it into the spatial-temporal feature $\\mathbf{F} = \\{\\pmb{f}_t\\}_{t=1}^{T_f}$ . Next, the encoder $\\mathcal{E}_1$ and decoder $\\mathcal{D}_1$ are employed in the initialization module to derive the initial prototype $\\mathbf{E}^0$ and initial translation $\\mathbf{Y}^0$ from the visual feature $\\mathbf{F}$ .", + "bbox": [ + 75, + 686, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Subsequently, the refinement module iteratively refines the previous prototype and generates the final translation $\\mathbf{Y}^{K} = \\{\\mathbf{y}_{t}^{K}\\}_{t=1}^{T_{y,K}}$ with $T_{y,K}$ words after total $K$ iterations. At the $k$ -th iteration, the encoder $\\mathcal{E}_2$ estimates the prototype $\\mathbf{E}^k$ for the current step by augmenting the original visual feature $\\mathbf{F}$ with the previous prototype $\\mathbf{E}^{k-1}$ . Next, the decoder $\\mathcal{D}_2$ predicts the corresponding translation $\\mathbf{Y}^k$ . Finally, the translation sequence $\\mathbf{Y} = \\{\\mathbf{Y}^0, \\mathbf{Y}^1, \\dots, \\mathbf{Y}^K\\}$ is obtained according to the prototype sequence $\\mathbf{E} = \\{\\mathbf{E}^0, \\mathbf{E}^1, \\dots, \\mathbf{E}^K\\}$ . For the IP-SLT optimization, by dividing the decoding part into $K+1$ branches, we add iterative distillation supervision from the final translation to the middle translations. Since the refinement process takes place in the encoder $\\mathcal{E}_2$ , in inference, the proposed IP-SLT can generate the translation directly from the $K$ -th prototype which causes acceptable overhead.", + "bbox": [ + 496, + 383, + 893, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Feature Extraction", + "text_level": 1, + "bbox": [ + 500, + 636, + 684, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The feature extraction module embeds a series of video frames $\\mathbf{X} \\in \\mathbb{R}^{T_x \\times H \\times W \\times 3}$ with width $W$ and height $H$ into its visual feature $\\mathbf{F} \\in \\mathbb{R}^{T_f \\times C}$ with the dimension of feature $C$ . Since its goal is to extract a distinguishable representation for SLT, we can draw on the visual backbone used in CSLR [14, 26, 25, 13] to extract the valid representation. Generally, with sliding window size $w$ and stride size $s$ , the sign video is split into $T_f = \\left\\lceil \\frac{T_x}{s} \\right\\rceil$ clips. By passing sign videos through it, the spatial-temporal embeddings $\\mathbf{F} = \\{\\mathbf{f}_t\\}_{t=1}^{T_f}$ are extracted as:", + "bbox": [ + 496, + 659, + 893, + 813 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\boldsymbol {f} _ {t} \\right\\} _ {t = 1} ^ {T _ {f}} = \\operatorname {E x t r a c t o r} \\left(\\left\\{\\boldsymbol {x} _ {t} \\right\\} _ {t = 1} ^ {T _ {x}}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 820, + 890, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Prototype Initialization", + "text_level": 1, + "bbox": [ + 500, + 847, + 718, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After a visual feature $\\pmb{F} \\in \\mathbb{R}^{T_f \\times C}$ is extracted by the feature extraction module, it is first fed into the initialization", + "bbox": [ + 498, + 869, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15594", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/381e7b884c3be4fac6d50fe0e688936241a4bfabfe3766ed68dd30a5c06a5e4d.jpg", + "image_caption": [ + "Figure 3. An illustration of the prototype update process in the $l$ -th layer of the encoder $\\mathcal{E}_2$ at the $k$ -th iteration." + ], + "image_footnote": [], + "bbox": [ + 129, + 90, + 415, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "module. The initialization module consists of an encoder $\\mathcal{E}_1$ and a decoder $\\mathcal{D}_1$ .", + "bbox": [ + 76, + 273, + 467, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The visual representation $\\pmb{F}$ is first fed into the encoder $\\mathcal{E}_1$ , and encoded into $T_{f}$ raw states $\\pmb{E}^{0} = \\{e_{t}^{0}\\}_{t=1}^{T_{f}} \\in \\mathbb{R}^{T_{f} \\times C}$ . Then the decoder $\\mathcal{D}_1$ reads the prototype $\\pmb{E}^{0}$ and produces the initial translation $\\pmb{Y}^{0} = \\{\\pmb{y}_{t}^{0}\\}_{t=1}^{T_{y,0}}$ according to the predicted logits $\\pmb{U}^{0} = \\{\\pmb{u}_{t}^{0}\\}_{t=1}^{T_{y,0}}$ . Specifically, it predicts the conditional probability of the translation sequence, which is formulated as:", + "bbox": [ + 76, + 304, + 467, + 415 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np ^ {0 - t h} \\left(\\boldsymbol {Y} ^ {0} | \\boldsymbol {X}\\right) = \\prod_ {t = 1} ^ {T _ {y, 0}} p \\left(\\boldsymbol {y} _ {t} ^ {0} \\mid \\boldsymbol {F}, \\boldsymbol {y} _ {0: t - 1} ^ {0}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 421, + 468, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{y}_{0:t-1}^{0} = \\{\\pmb{y}_{0}^{0}, \\pmb{y}_{1}^{0}, \\dots, \\pmb{y}_{t-1}^{0}\\}$ denotes the previous output sub-sequence at the $t$ -th step. The initial token $\\pmb{y}_{0}^{0}$ represents the beginning of a sentence. The predicted probability of each token in the translation is computed as:", + "bbox": [ + 76, + 474, + 468, + 536 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p \\left(\\boldsymbol {y} _ {t} ^ {0} \\mid \\boldsymbol {F}, \\boldsymbol {y} _ {1: t - 1} ^ {0}\\right) = \\operatorname {s o f t m a x} \\left(\\boldsymbol {u} _ {t} ^ {0}\\right) _ {\\boldsymbol {y} _ {t} ^ {0}} \\tag {3} \\\\ = \\operatorname {s o f t m a x} \\left(\\boldsymbol {h} _ {t} ^ {0} \\cdot \\boldsymbol {W}\\right) _ {\\boldsymbol {y} _ {t} ^ {0}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 542, + 468, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h_t^0 \\in \\mathbb{R}^C$ represents the output of the final layer at the $t$ -th step, and $W \\in \\mathbb{R}^{C \\times |\\mathcal{V}|}$ denotes a linear mapping to projects the hidden state $h_t^0$ into the predicted logits over the target vocabulary $\\mathcal{V}$ . The probability is calculated by applying the softmax(·) function to the logits. Notably, our goal is to obtain a more accurate prototype for SLT, thus, the decoder $\\mathcal{D}_1$ is only used in the training process to provide guidance for the initialization module.", + "bbox": [ + 76, + 594, + 468, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Iterative Prototype Refinement", + "text_level": 1, + "bbox": [ + 76, + 726, + 352, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once the initial prototype $\\pmb{E}^{0}$ is obtained through the initialization module, we feed it together with the original visual representation $\\pmb{F}$ into the iterative refinement module. The module maintains a single prototype which is iteratively refined. In this way, the coarse semantic feature finally converges to a stable state where the prototype best fits the sign language semantics. The refinement module is divided into two sub-processes, i.e., iterative prototype aggregation in the encoder $\\mathcal{E}_2$ and translation generation in the decoder $\\mathcal{D}_2$ .", + "bbox": [ + 76, + 750, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Iterative prototype aggregation. To utilize the prototype as a reference, each layer of the encoder $\\mathcal{E}_2$ attends over the maintained semantic features through the attention mechanism. As shown in Fig. 3, we illustrate the prototype aggregation process in the $l$ -th layer of the encoder $\\mathcal{E}_2$ at the $k$ -th iteration. The input of the refinement module consists of the original visual feature $\\pmb{F}$ and the $(k-1)$ -th prototype $\\pmb{E}^{k-1}$ . The attention mechanism $\\mathrm{attn}(\\pmb{q},\\pmb{K},\\pmb{V})$ is originally used in Transformer [43], which is formulated as:", + "bbox": [ + 498, + 90, + 890, + 226 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {a t t n} (\\boldsymbol {q}, \\boldsymbol {K}, \\boldsymbol {V}) = \\sum_ {i = 1} ^ {| V |} \\alpha_ {i} \\boldsymbol {W} _ {v} \\boldsymbol {v} _ {i}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 231, + 890, + 273 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} = \\operatorname {s o f t m a x} \\left(\\left(\\boldsymbol {W} _ {q} \\boldsymbol {q}\\right) ^ {T} \\left(\\boldsymbol {W} _ {k} \\boldsymbol {k} _ {i}\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 277, + 807, + 295 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{q}, W_{k}$ and $W_{v}$ are learnable parameters. To make better use of the $(k - 1)$ -th prototype, we consider the $(k - 1)$ -th prototype $E^{k - 1} = \\{e_t^{k - 1}\\}_{t = 1}^{T_f}$ as key-value pair, and the original visual feature $F = \\{f_t\\}_{t = 1}^{T_f}$ as the query. In this way, we inject the rich semantic information into the new prototype $E^k = \\{e_t^k\\}_{t = 1}^{T_f}$ .", + "bbox": [ + 500, + 301, + 890, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The encoder $\\mathcal{E}_2$ is composed of $L_{e}$ identical layers, where $\\pmb{E}_{l}^{k}$ denotes the output of the $l$ -th layer at the $k$ -th iteration. Similar to [51], the hidden state is computed as:", + "bbox": [ + 498, + 401, + 890, + 445 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\widetilde {\\boldsymbol {E}} _ {l} ^ {k} = \\beta \\cdot \\operatorname {a t t n} _ {s} \\left(\\boldsymbol {E} _ {l - 1} ^ {k}, \\boldsymbol {E} _ {l - 1} ^ {k}, \\boldsymbol {E} _ {l - 1} ^ {k}\\right) \\tag {5} \\\\ + (1 - \\beta) \\cdot \\operatorname {a t t n} _ {c} \\left(\\boldsymbol {E} _ {l - 1} ^ {k}, \\boldsymbol {E} ^ {k - 1}, \\boldsymbol {E} ^ {k - 1}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 453, + 890, + 493 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{attn}_s$ and $\\mathrm{attn}_c$ denote the self-attention sub-layer and cross-attention sub-layer used in Transformer [43], respectively. For the first layer of the encoder $\\mathcal{E}_2$ , $E_{l - 1}^k$ is equal to the visual feature $\\pmb{F}$ . The output of the encoder $\\mathcal{E}_2$ is $E_{L_e}^k$ (i.e., $E^k$ ). $\\beta \\in [0,1]$ is a hyperparameter that weights the importance of previous prototypes during training and inference. To further fuse and refine the prototype, it is linked with a fully connected sub-layer $\\mathrm{FFN}(\\cdot)$ using the residual connection. The output of the $l$ -th layer at the $k$ -th iteration is formulated as:", + "bbox": [ + 498, + 502, + 890, + 655 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {E} _ {l} ^ {k} = \\operatorname {F F N} \\left(\\operatorname {L N} \\left(\\widetilde {\\boldsymbol {E}} _ {l} ^ {k} + \\boldsymbol {E} _ {l - 1} ^ {k}\\right)\\right) + \\operatorname {L N} \\left(\\widetilde {\\boldsymbol {E}} _ {l} ^ {k} + \\boldsymbol {E} _ {l - 1} ^ {k}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 664, + 890, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{LN}(\\cdot)$ is the layer normalization operation.", + "bbox": [ + 498, + 691, + 826, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Translation generation. The decoder $\\mathcal{D}_2$ iteratively takes the prototype $E^k$ as input and generates the corresponding translation $\\pmb{Y}^{k}$ . Take the $k$ -th iteration as an example. The decoder $\\mathcal{D}_2$ predicts the conditional probability of translation $\\pmb{Y}^{k} = \\{\\pmb{y}_{t}^{k}\\}_{t = 1}^{T_{y,k}}$ based on the predicted logits $\\pmb{U}^{k} = \\{\\pmb{u}_{t}^{k}\\}_{t = 1}^{T_{y,k}}$ , which is computed as:", + "bbox": [ + 498, + 709, + 890, + 805 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np ^ {k - t h} \\left(\\boldsymbol {Y} ^ {k} \\mid \\boldsymbol {X}\\right) = \\prod_ {t = 1} ^ {T _ {y, k}} p \\left(\\boldsymbol {y} _ {t} ^ {k} \\mid \\boldsymbol {F}, \\boldsymbol {E} ^ {k - 1}, \\boldsymbol {y} _ {0: t - 1} ^ {k}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 811, + 890, + 852 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np (\\boldsymbol {y} _ {t} ^ {k} | \\boldsymbol {F}, \\boldsymbol {E} ^ {k - 1}, \\boldsymbol {y} _ {0: t - 1} ^ {k}) = s o f t m a x (\\boldsymbol {u} _ {t} ^ {k}) _ {\\boldsymbol {y} _ {t} ^ {k}},\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 854, + 849, + 876 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{u}_t^k$ is the predicted logits of the decoder $\\mathcal{D}_2$ at the $t$ -", + "bbox": [ + 503, + 883, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15595", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "th step. The decoder $\\mathcal{D}_2$ iteratively takes the prototype given by the encoder $\\mathcal{E}_2$ as input to generate the final translation $\\mathbf{Y}^K$ . With total $K$ iterations, the outputs of encoder $\\mathcal{E}_1$ and $\\mathcal{E}_2$ compose the prototype sequence $\\mathbf{E} = \\{\\mathbf{E}^k\\}_{k=0}^K$ , where $\\mathbf{E}^k = \\{\\mathbf{e}_t^k\\}_{t=1}^{T_f}$ is the semantic feature at the $k$ -th iteration. Accordingly, the decoder $\\mathcal{D}_1$ and $\\mathcal{D}_2$ generate $K+1$ translations $\\mathbf{Y} = \\{\\mathbf{Y}^k\\}_{k=0}^K$ , where $\\mathbf{Y}^k = \\{\\mathbf{y}_t^k\\}_{t=1}^{T_y,k}$ is the translation at the $k$ -th iteration during training. Note that after $K$ iterations, the decoder $\\mathcal{D}_2$ obtains the converged prototype $\\mathbf{E}^K$ and generates the translation only once in inference.", + "bbox": [ + 76, + 90, + 472, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training Objective", + "text_level": 1, + "bbox": [ + 76, + 257, + 261, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce two kinds of losses in the training period of the IP-SLT system. Firstly, the cross entropy loss is adopted to supervise the final generated sentence. Secondly, we put forward an iterative distillation loss. As the decoder $\\mathcal{D}_1$ and $\\mathcal{D}_2$ generate translation sequence $\\mathbf{Y} = \\{\\mathbf{Y}^0,\\mathbf{Y}^1,\\dots ,\\mathbf{Y}^K\\}$ , we naturally divide the sequence $\\mathbf{Y}$ into an initial prediction, $K - 1$ intermediate predictions, and the final prediction according to the order of iteration. Conceptually, the $K - 1$ intermediate predictions are regarded as the student model and distill knowledge from the final prediction which is regarded as the teacher model.", + "bbox": [ + 76, + 281, + 472, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cross entropy loss. As mentioned above, the IP-SLT generates the translation sequence based on the conditional probability provided by the decoder $\\mathcal{D}_1$ and $\\mathcal{D}_2$ . The cross-entropy loss [43] is computed with the ground truth from $N$ training samples and the outputs of the decoder $\\mathcal{D}_1$ and $\\mathcal{D}_2$ . Its training objective in the proposed approach is to maximize the log-likelihood which is equal to minimizing the cross entropy loss formulated as:", + "bbox": [ + 75, + 453, + 472, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {C E, k} = - \\log p ^ {k - t h} (\\hat {\\mathbf {Y}} | \\mathbf {X}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 582, + 468, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{Y}$ denotes translation annotation. We apply this at the output of the initialization module and the $K$ -th output of the refinement module.", + "bbox": [ + 75, + 614, + 468, + 661 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Iterative distillation loss. Since the KL (Kullback-Leibler) divergence loss can affect the teacher's networks to the student's networks, we compute it between the $K - 1$ intermediate predictions and the final prediction. As the final translation is based on the previous prototypes, we are able to get better translations by approximating more vital characterization capabilities of middle prototypes. The iterative distillation loss (IDL) is formulated as:", + "bbox": [ + 75, + 666, + 472, + 787 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {I D L} = \\sum_ {k = 1} ^ {K - 1} K L \\left(\\boldsymbol {U} ^ {k}, \\boldsymbol {U} ^ {K}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 800, + 468, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{U}^{k}$ and $\\pmb{U}^{K}$ are the predicted output of the decoder $\\mathcal{D}_2$ at the $k$ -th and $K$ -th iteration, respectively. By computing the Kullback-Leibler divergence between the $\\pmb{U}^{k}$ and", + "bbox": [ + 76, + 854, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\pmb{U}^{K}$ , the IP-SLT is encouraged to approximate the performance of the final iteration. We apply this at the $K - 1$ shallow outputs of the refinement module.", + "bbox": [ + 498, + 90, + 890, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall, the loss function of our IP-SLT is formulated as:", + "bbox": [ + 517, + 136, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {C E, 0} + L _ {C E, K} + \\lambda \\cdot L _ {I D L}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 162, + 890, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Under the guidance of CE loss, we first train the initial generation module until convergence as a warm start, and then apply the loss function as Equ. (10) for optimizing the IP-SLT system in an end-to-end manner.", + "bbox": [ + 498, + 191, + 890, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Inference Strategy", + "text_level": 1, + "bbox": [ + 500, + 261, + 679, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the refinement process only involves in the encoder $\\mathcal{E}_2$ , the initialization and iterative refinement modules behave differently during training and inference. The decoder $\\mathcal{D}_1$ of the initialization module is leveraged to provide guidance for the encoder $\\mathcal{E}_1$ during the training process, while it is not required in inference. Similarly, the autoregressive decoder $\\mathcal{D}_2$ of the iterative refinement module generates translations for each iteration during training, while it just decodes the final translation once in inference.", + "bbox": [ + 496, + 284, + 890, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In inference, given a sign language video to be translated, the feature extraction module first converts it into a visual feature. The encoder $\\mathcal{E}_1$ of the initialization module transforms it into the raw prototype. Then, the encoder $\\mathcal{E}_2$ of the refinement module further iteratively refines it by fusing it with the original visual feature. Finally, the decoder $\\mathcal{D}_2$ generates the translation based on the final prototype.", + "bbox": [ + 496, + 420, + 890, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 539, + 633, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 564, + 689, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We evaluate our approach on two pubic sign language translation datasets, i.e., PHOENIX14T [5] and CSL-Daily [49]. Both datasets provide gloss-level and spoken-sentence-level annotations. The PHOENIX14T dataset [5] is the first large-scale neural SLT dataset created with 9 German sign language interpreters. The dataset is split into a training set (7,096), a development set (519), and a test set (642). The CSL-Daily dataset [49] is a Chinese SLT dataset containing 20,654 annotated sign language videos. We follow the previous experimental setting [49] and split it into the training, development, and test set.", + "bbox": [ + 496, + 590, + 890, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. Following the work [49], we quantitatively assess the quality of translations according to the BLEU- $N$ [32] and ROUGE [38]. The BLEU- $N$ ( $N$ ranges from 1 to 4) cares more about the accuracy of the predicted translation while the ROUGE cares more about the consistency of sentences. For both evaluation metrics, a higher value indicates a better performance.", + "bbox": [ + 496, + 760, + 890, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training settings. We implement our approach on Pytorch [33]. The encoder and decoder of the initialization", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15596", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a32ea604528750f74c8623e2612ddc7b2b1eb5b26e595f01e38fa039004bc594.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsLLMDevTest
ROUGEBLEU-1BLEU-2BLEU-3BLEU-4ROUGEBLEU-1BLEU-2BLEU-3BLEU-4
Joint-SLRT [7]✘-47.2634.4027.0522.38-46.6133.7326.1921.32
PET [23]✘-----49.9749.5437.1929.3024.02
BN-TIN-Transf. [49]✘46.8746.9033.9826.4921.7846.9847.5734.6426.7821.68
STMC [50]✘48.2447.6036.4329.1824.0946.6546.9836.0928.7023.65
IP-SLT✘54.4354.1041.5633.6628.2253.7254.2541.5133.4527.97
MMTLB [9]✓53.1053.9541.1233.1427.6152.6553.9741.7533.8428.39
TwoStream-SLT [10]✓54.0854.3241.9934.1528.6653.4854.9042.4334.4628.95
", + "bbox": [ + 99, + 88, + 870, + 231 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7eb3b109c201c7c7aeceb933ce145efff43039c009f765088d640bead7dafe17.jpg", + "table_caption": [ + "Table 1. Performance comparison of IP-SLT with methods for SLT on PHOENIX-2014T. 'LLM' denotes adopting pre-trained large language models." + ], + "table_footnote": [], + "table_body": "
MethodsLLMDevTest
ROUGEBLEU-1BLEU-2BLEU-3BLEU-4ROUGEBLEU-1BLEU-2BLEU-3BLEU-4
SL-Luong [5]X34.2834.2219.7212.247.9634.5434.1619.5711.847.56
Joint-SLRT [7]X37.0637.4724.6716.8611.8836.7437.3824.3616.5511.79
BN-TIN-Transf. [49]X37.2940.6626.5618.0612.7337.6740.7426.9618.4813.19
IP-SLTX44.3345.2631.7722.8716.7444.0944.8531.5022.6616.72
MMTLB [9]✓53.3853.8140.8431.2924.4253.2553.3140.4130.8723.92
TwoStream-SLT [10]✓55.1055.2142.3132.7125.7655.7255.4442.5932.8725.79
", + "bbox": [ + 99, + 268, + 870, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Performance comparison of IP-SLT with methods for SLT on CSL-Daily. 'LLM' denotes adopting pre-trained large language models.", + "bbox": [ + 76, + 401, + 892, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and refinement module consist of 3 layers, respectively. The dimension of the feed-forward network is 2048. The visual feature of the STMC [50] is 1024-dimension while the visual feature of BN-TIN-Transf. [49] and VAC [31] are 512-dimension. To alleviate over-fitting, we set dropout and attention head to 0.1 and 8, respectively. The training optimizer is Adam [24]. During training, the learning rate is fixed to $5 \\times 10^{-5}$ . To ensure the features provided by the previous prototype and original sign video are fully utilized, we apply the drop-net [51] during training, which effects Equ. (5). During training, for any layer in the encoder $\\mathcal{E}_2$ , with probability $\\beta$ , the hidden state $\\widetilde{\\pmb{E}}_l^k$ in Equ. (5) is the output of the self-attention sub-layer $\\mathrm{attn}_s$ ; with probability $1 - \\beta$ , it is the output of the cross-attention sub-layer $\\mathrm{attn}_c$ . In inference, the hidden state $\\widetilde{\\pmb{E}}_l^k$ is computed as Equ. (5).", + "bbox": [ + 75, + 446, + 472, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference details. In inference, we use the beam search strategy [45] to improve the decoding accuracy. For the PHOENIX-2014T dataset and the CSL-Daily dataset, we set the beam search width and the length penalty to 3 and 1.0, respectively. To reduce the computational complexity, our IP-SLT just decodes once in inference for each input sign language video.", + "bbox": [ + 75, + 680, + 468, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 76, + 800, + 444, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the proposed IP-SLT with the previous SLT systems on two public benchmarks, i.e., PHOENIX14T [5] and CSL-Daily [49], and the performance of our IP-SLT is shown in Tab. 1 and Tab. 2, respectively. For PHOENIX14T and CSL-Daily dataset, we adopt the STMC [50] and BN-", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TIN-Transf. [49] as the baseline, respectively. Our IP-SLT follows the sign-to-text (S2T) paradigm, which directly transforms the sign language video into translation. Note MMTLB [9] and TwoStream-SLT [10] adopt pre-trained large-scale language models that leverage more model parameters and extra resources than IP-SLT.", + "bbox": [ + 496, + 446, + 890, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By combining all proposed components together, our IP-SLT achieves substantial improvements against the baseline. The IP-SLT achieves 28.22 and 16.74 BLEU-4 scores on the DEV set of PHOENIX14T and CSL-Daily, respectively. The quantitative results demonstrate that our IP-SLT achieves promising performance improvements. Our IP-SLT delivers promising performance gains on DEV and test sets by leveraging the iterative refinement process. The results prove the advantage of aggregating the previous translation progress and the original visual representation, which distinguishes our IP-SLT from previous SLT systems.", + "bbox": [ + 496, + 539, + 892, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 719, + 663, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we put forward several ablation studies on the DEV set of PHOENIX-2014T. Unless otherwise specified, we adopt the STMC [50] as the baseline for the following experiments.", + "bbox": [ + 496, + 743, + 890, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Impact of network architecture. The main difference between our proposed method and the existing work is to leverage the previous information as an additional clue to enhance the current prototype. To evaluate the effectiveness of each component, we gradually add the refinement module and the iterative distillation loss to the baseline SLT sys", + "bbox": [ + 496, + 809, + 892, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15597", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8522843bb86dc3c38234e11689b7c8e3dc34aafe9d6cc69f2a4b38924bcfd6a9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingROUGEBLEU-1BLEU-2BLEU-3BLEU-4
Baseline50.1050.2137.1229.4124.31
+Refinement51.2251.0438.4030.6125.39
+IDL54.4354.1041.5633.6628.22
6-6 Layers50.4950.5837.5729.6424.53
", + "bbox": [ + 81, + 89, + 465, + 178 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ee687f7bfe7cf367471e71842f47c433cfae205f17d847dc89021e61cc41d109.jpg", + "table_caption": [ + "Table 3. Effect of our proposed components. 'Refinement' denotes applying the refinement process. 'IDL' denotes applying the iterative distillation loss. '6-6 layers' denotes enlarging the encoder and decoder of the baseline system from 3 to 6 layers." + ], + "table_footnote": [], + "table_body": "
ModelRMROUGEBLEU-1BLEU-2BLEU-3BLEU-4
STMCw/o50.1050.2137.1229.4124.31
w/54.4354.1041.5633.6628.22
BN-TIN-Transf.w/o47.4147.9934.9427.3322.35
w/52.0652.0639.0131.0825.69
VAC-Transf.w/o49.4850.0137.0029.1223.91
w/53.6853.6041.2833.4728.07
", + "bbox": [ + 81, + 250, + 465, + 371 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b0e0250ea67f823d4e3d0376da2117064893bd6651bc116588e466ade07bcdcb.jpg", + "table_caption": [ + "Table 4. Generalization of IP-SLT. 'RM' denotes leveraging the refinement process. 'w' and 'w/o' denote the baseline SLT system with and without a refinement process, respectively." + ], + "table_footnote": [], + "table_body": "
λROUGEBLEU-1BLEU-2BLEU-3BLEU-4
051.3050.3238.0130.2925.14
553.0752.3140.0832.2927.02
1053.7153.5140.9232.9427.47
1554.4354.1041.5633.6628.22
2054.4254.1641.5133.4427.87
", + "bbox": [ + 81, + 428, + 465, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tem. Directly applying the refinement process to the baseline delivers a performance gain of 1.08 BLEU-4. We further apply the iterative distillation loss to improve the performance. The results suggest that adding distillation supervision can be helpful with a gain of 2.83 BLEU-4. Besides, to keep the number of parameters unchanged, we enlarge the depth from 3 to 6 layers and evaluate the performance. Naively enlarging the model scale slightly improves the performance $(+0.22$ BLEU-4). The results are shown in Tab. 3.", + "bbox": [ + 75, + 561, + 467, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalization of the IP-SLT. We conduct three sets of experiments by changing the visual backbone to evaluate the generalization of the proposed IP-SLT approach in Tab. 4. Specifically, the BN-TIN-Transf. [49] uses a basic CNN network to get the dense representation of sign video. The STMC [50] extracts and aligns the key parts of sign language expression to achieve better performance. The VAC [31] proposes two auxiliary supervision methods to enhance the feature extraction module. VAC-Transf. replaces the feature extractor in BN-TIN-Transf. [49] with the visual backbone of VAC [31]. Applying our proposed IP-SLT methods to the BN-TIN-Transf., VAC-Transf. and STMC, we achieve 25.69, 28.07, and 28.22 BLEU-4 scores", + "bbox": [ + 75, + 703, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5221bfb387a3cb48c974c2130ff2f8d239d6b8cc5688cfc91527a807b74e5f9d.jpg", + "table_caption": [ + "Table 5. The weight $\\lambda$ of iterative distillation loss to CE loss." + ], + "table_footnote": [], + "table_body": "
βROUGEBLEU-1BLEU-2BLEU-3BLEU-4
0.050.1050.2137.1229.4124.31
0.254.3553.8141.4033.5228.02
0.454.3954.0241.2433.1927.70
0.554.4354.1041.5633.6628.22
0.653.9653.6340.9433.1027.72
0.854.3153.3940.9232.9827.53
-50.9750.6738.2430.5225.41
", + "bbox": [ + 504, + 89, + 887, + 214 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f4117220fae99f97ae73457b2fe858ef5cd208afe9630d33f66e311107002845.jpg", + "table_caption": [ + "Table 6. The weight $\\beta$ of the original visual feature to the previous prototype. '0.0' denotes the assessment of the baseline. '-' denotes setting $\\beta$ as 0.5 without using the drop-net [51]." + ], + "table_footnote": [], + "table_body": "
SettingROUGEBLEU-1BLEU-2BLEU-3BLEU-4
None50.1050.2137.1229.4124.31
Con-input54.3753.8341.1032.9727.39
Con-feature53.9752.7640.5632.8127.47
Add-feature54.4354.1041.5633.6628.22
", + "bbox": [ + 504, + 271, + 887, + 353 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Effect of the different refinement methods. 'Con-input' denotes directly concatenating the original visual feature and previous prototype as input. 'Con-feature' denotes concatenating the original feature and the previous feature given by the cross-attention mechanism in each layer. On top of that, 'Add-feature' denotes changing it to an addition operation.", + "bbox": [ + 498, + 359, + 890, + 443 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on the DEV set, surpassing the baselines by 3.34, 4.16, 3.91, respectively. Using a high-quality visual backbone delivers further quality gains.", + "bbox": [ + 498, + 465, + 890, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of $\\lambda$ . In our experiments, the weight $\\lambda$ of iterative distillation loss is set to 15. It is a hyper-parameter that is designed to balance the effect of cross-entropy loss and the iterative distillation loss. We conduct experiments by varying the weight $\\lambda$ . Tab. 5 shows that our IP-SLT achieves the best performance when the weight $\\lambda$ is set to 15.", + "bbox": [ + 498, + 513, + 890, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of $\\beta$ . In the above experiments, the weight $\\beta$ is fixed to 0.5. The weight $\\beta$ represents the importance of the previous prototype compared with the original visual feature. As a hyper-parameter of our proposed methods, the weight $\\beta$ is examined with a set of different values in Tab. 6. When the weight is 0.5, the performance is the highest one. This indicates that the previous prototype to SLT is as important as the original visual feature. Besides, to fully use the previous prototype, the drop-net is required.", + "bbox": [ + 496, + 608, + 890, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of refinement method. We also examine a set of refinement methods for IP-SLT considering the fusion mechanism as a key part of our proposed method in Tab. 7. Directly concatenating the original visual feature and previous prototype in the time dimension improves the performance of SLT from 24.31 to 27.39. We further add the cross-attention mechanism to explicitly leverage the useful information from the previous prototype. And then, we concatenate the representation from the original visual feature and the representation from the previous prototype in fea", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15598", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/28bb906546f9678e6d09a5a5eb4a81444305ff86d49116319d6790e40dd06c1f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
KROUGEBLEU-1BLEU-2BLEU-3BLEU-4
050.1050.2137.1229.4124.31
151.2251.0438.4030.6125.39
253.7353.3940.7632.9827.66
354.4354.1041.5633.6628.22
454.6353.9141.4033.5328.01
", + "bbox": [ + 81, + 89, + 465, + 179 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1aa3a73c60ce73bfed4ac70203a609b9d505cb3687ee3038b216b2f6c6efcd69.jpg", + "table_caption": [ + "Table 8. Effect of the iteration number $K$ in the iterative refinement module." + ], + "table_footnote": [ + "Table 9. Comparing the baseline and non-shared weight prototype refinement method with IP-SLT. 'SW' denotes the sharing weight across all iterations. '-' denotes the baseline SLT system. 'w/' and 'w/o' denote the refinement module in different iterations with and without shared parameters, respectively. 'I-P' and 'T-P' denote the parameter amount calculated in inference and during training, respectively." + ], + "table_body": "
ModelSWI-P(M)T-P(M)FLOPs(B)ROUGEBLEU-4
STMC-92.692.628.750.1024.31
w/128.7172.632.654.4328.22
w/o238.6357.932.654.8128.48
BN-TIN-Transf.-28.328.39.647.4122.35
w/37.753.410.652.0625.69
w/o66.1110.010.652.2425.96
VAC-Transf.-28.328.39.649.4823.91
w/37.753.410.653.6828.07
w/o66.1110.010.654.8228.27
", + "bbox": [ + 81, + 220, + 465, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ture dimension, which delivers performance gains of 3.16 BLEU-4. On top of that, changing the concatenating process to an element-wise addition operation achieves a further quality gain of 3.91.", + "bbox": [ + 75, + 500, + 468, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of iteration number $K$ . The iteration number $K$ is an important hyper-parameter and is fixed to 3 in the previous experiments. We conduct experiments with different iteration numbers to explore the effect of iteration number $K$ . Tab. 8 shows that there is a culmination in the performance at iteration 3. Before this culmination, the performance improvement of our proposed method increases fast. When the iteration number is bigger than 3, the performance of our method is slightly weakened (-0.21 BLEU-4).", + "bbox": [ + 75, + 564, + 468, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of sharing weights between different iterations. In Tab. 9, we examine the storage and translation quality of our proposed IP-SLT with different baselines i.e., STMC, BN-TIN-Transf, and VAC-Transf. In order to ensure the efficiency of the model, our refinement module shares weight across all iterations. Since the feature extraction module of each group is identical, the parameters of the feature extraction module are not included in the calculation of the storage of different models. As in inference, several parts can be removed from IP-SLT without changing the performance, we report the number of parameters in both training and inference, respectively. The results indicate that our iterative process causes acceptable overhead with re", + "bbox": [ + 75, + 704, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b856cc2c08e3659ab59d4cb77ff09fe1a0205722d3a240b8c8cf55a897265b51.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeText
GTich=winsche ihnen einen tatsächien abend und machen sie es gut.
Baselinejetzt wünsche ich ihnen noch einen tatsächien abend.
Ourihnen einen tatsächien abend und machen sie es gut.
GTin der neuenWoche wird es milder aber es bleibt wechselhaft.
Baselinedann wird es wieder milder.
Ourin der neuenWoche wird es dann wieder milder.
", + "bbox": [ + 504, + 89, + 890, + 189 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 10. Qualitative evaluation. 'GT' denotes the spoken language translation annotation. 'Baseline' and 'Our' denote the translation result of baseline and our IP-SLT, respectively.", + "bbox": [ + 498, + 195, + 890, + 237 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "markable performance improvements. We further conduct experiments in which the parameters of each iteration are independent. The results demonstrate that the proposed parameter-shared method achieves near performance gain (28.22 v.s 28.48 BLEU-4, 25.69 v.s 25.96 BLEU-4, and 28.07 v.s 28.27 BLEU-4) with non-shared one but leveraging much less parameter number.", + "bbox": [ + 496, + 258, + 890, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computation comparison with baseline. The FLOPs is a key factor of computation efficiency. We conduct experiments to compare the computation of the proposed IP-SLT method with different baselines. Similarly, we exclude the FLOPs of the feature extraction module and report relevant computational costs in inference. Tab. 9 shows that leveraging the iterative refinement process can cause acceptable computation costs while achieving promising performance.", + "bbox": [ + 496, + 368, + 890, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Case Study. To provide a more intuitionistic view of our proposed method, we list some translation samples of the proposed IP-SLT method in Tab. 10. We observe that based on the previous prototype and original visual representation, the IP-SLT can generate more accurate and fluent sentences.", + "bbox": [ + 496, + 494, + 890, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 584, + 617, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a new framework IP-SLT which introduces the iterative refinement into a conventional SLT system. With the goal to polish the semantic representation by leveraging the previous results, we present IP-SLT to support the iterative refinement process. The proposed method is differentiable and optimized in an end-to-end manner to achieve its best performance. On top of it, we put forward the iterative distillation loss to further improve the translation quality by leveraging the sequential dependence between the outputs of each iteration. In inference, the autoregressive decoding process is required once to generate the translation based on the final prototype, applying IP-SLT does not significantly affect efficiency. The experimental results demonstrate the effectiveness of the IP-SLT.", + "bbox": [ + 496, + 609, + 890, + 820 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: This work was supported by NSFC under Contract U20A20183 and 62021001. It was also supported by the GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, and the Supercomputing Center of the USTC.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15599", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Image super-resolution via progressive cascading residual network. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2018. 2", + "[2] Samuel Albanie, Gül Varol, Liliane Momeni, Triantafyllos Afouras, Joon Son Chung, Neil Fox, and Andrew Zisserman. BSL-1K: Scaling up co-articulated sign language recognition using mouthing cues. In Computer Vision European Conference, 2020. 1", + "[3] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. 2, 3", + "[4] Danielle Bragg, Oscar Koller, Mary Bellard, Larwan Berke, Patrick Boudreault, Annelies Braffort, Naomi Caselli, Matt Huenerfauth, Hernisa Kacorri, Tessa Verhoef, et al. Sign language recognition, generation, and translation: An interdisciplinary perspective. In International ACM SIGACCESS Conference on Computers and Accessibility, 2019. 1", + "[5] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In IEEE Conference on Computer Vision and Pattern Recognition, 2018. 1, 2, 5, 6", + "[6] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Multi-channel transformers for multi-articulatory sign language translation. In Computer Vision-ECCV 2020 Workshops, 2020. 2", + "[7] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 2, 6", + "[8] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2", + "[9] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2, 6", + "[10] Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie Liu, and Brian Mak. Two-stream network for sign language recognition and translation. Advances in Neural Information Processing Systems, 2022. 2, 6", + "[11] Helen Cooper, Brian Holt, and Richard Bowden. Sign language recognition. Visual Analysis of Humans: Looking at People, 2011. 1", + "[12] Kearsy Cormier, Neil Fox, Bencie Woll, Andrew Zisserman, Necati Cihan Camgoz, and Richard Bowden. ExTOL: Automatic recognition of British sign language using the bsl corpus. In Workshop on Sign Language Translation and Avatar Technology, 2019. 1", + "[13] Runpeng Cui, Hu Liu, and Changshui Zhang. Recurrent convolutional neural networks for continuous sign language recognition by staged optimization. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1, 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Runpeng Cui, Hu Liu, and Changshui Zhang. A deep neural framework for continuous sign language recognition by iterative training. IEEE Transactions on Multimedia, 2019. 2, 3", + "[15] Hao Feng, Wengang Zhou, Jiajun Deng, Qi Tian, and Houqiang Li. DocScanner: robust document image rectification with progressive learning. arXiv preprint arXiv:2110.14968, 2021. 2", + "[16] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. Convolutional sequence to sequence learning. In International Conference on Machine Learning, 2017. 3", + "[17] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In International Conference on Machine learning, 2006. 2", + "[18] Karol Gregor, Ivo Danihelka, Alex Graves, Danilo Rezende, and Daan Wierstra. DRAW: A recurrent neural network for image generation. In International Conference on Machine Learning, 2015. 2", + "[19] Jun He, Richang Hong, Xueliang Liu, Mingliang Xu, and Qianru Sun. Revisiting local descriptor for improved few-shot classification. ACM Transactions on Multimedia Computing, Communications, and Applications, 2022. 1", + "[20] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. Signbert+: Hand-model-aware self-supervised pretraining for sign language understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1", + "[21] Jie Huang, Wengang Zhou, Qilin Zhang, Houqiang Li, and Weiping Li. Video-based sign language recognition without temporal segmentation. In AAAI Conference on Artificial Intelligence, 2018. 1", + "[22] Shuiwang Ji, Wei Xu, Ming Yang, and Kai Yu. 3D convolutional neural networks for human action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2012. 2", + "[23] Tao Jin, Zhou Zhao, Meng Zhang, and Xingshan Zeng. Prior knowledge and memory enriched transformer for sign language translation. In Findings of the Association for Computational Linguistics, 2022. 1, 2, 6", + "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[25] Oscar Koller, Necati Cihan Camgoz, Hermann Ney, and Richard Bowden. Weakly supervised learning with multistream cnn-lstm-hmms to discover sequential parallelism in sign language videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 3", + "[26] Oscar Koller, Hermann Ney, and Richard Bowden. Deep hand: How to train a CNN on 1 million hand images when your data is continuous and weakly labelled. In IEEE Conference on Computer Vision and Pattern Recognition, 2016. 3", + "[27] Dongxu Li, Chenchen Xu, Xin Yu, Kaihao Zhang, Benjamin Swift, Hanna Suominen, and Hongdong Li. TSPNet: Hierarchical feature learning via temporal semantic pyramid for" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15600", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sign language translation. Advances in Neural Information Processing Systems, 2020. 2", + "[28] Huan Ling, Jun Gao, Amlan Kar, Wenzheng Chen, and Sanja Fidler. Fast interactive object annotation with Curve-GCN. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2", + "[29] Zichen Liu, Jun Hao Liew, Xiangyu Chen, and Jiashi Feng. DANCE: A deep attentive contour model for efficient instance segmentation. In IEEE Winter Conference on Applications of Computer Vision, 2021. 2", + "[30] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. arXiv preprint arXiv:1508.04025, 2015. 1", + "[31] Yuecong Min, Aiming Hao, Xiujuan Chai, and Xilin Chen. Visual alignment constraint for continuous sign language recognition. In IEEE International Conference on Computer Vision, 2021. 1, 2, 6, 7", + "[32] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: a method for automatic evaluation of machine translation. In Annual Meeting of the Association for Computational Linguistics, 2002. 5", + "[33] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 5", + "[34] Sida Peng, Wen Jiang, Huaijin Pi, Xiuli Li, Hujun Bao, and Xiaowei Zhou. Deep snake for real-time instance segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[35] Junfu Pu, Wengang Zhou, and Houqiang Li. Dilated convolutional network with iterative optimization for continuous sign language recognition. In International Joint Conference on Artificial Intelligence, 2018. 2", + "[36] Junfu Pu, Wengang Zhou, and Houqiang Li. Iterative alignment network for continuous sign language recognition. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2", + "[37] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive Image Deraining Networks: A better and simpler baseline. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2", + "[38] Lin CY ROUGE. A package for automatic evaluation of summaries. In Workshop on Text Summarization of ACL, 2004. 5", + "[39] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. Advances in Neural Information Processing Systems, 2014. 2", + "[40] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. Advances in Neural Information Processing Systems, 2014. 1, 3", + "[41] Rachel Sutton-Spence and Bencie Woll. The linguistics of British Sign Language: an introduction. Cambridge University Press, 1999. 1", + "[42] Shengeng Tang, Dan Guo, Richang Hong, and Meng Wang. Graph-based multimodal sequential embedding for sign language translation. IEEE Transactions on Multimedia, 2021. 1" + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1, 2, 3, 4, 5", + "[44] Hanjie Wang, Xiujuan Chai, Xiaopeng Hong, Guoying Zhao, and Xilin Chen. Isolated sign language recognition with grassmann covariance matrices. ACM Transactions on Accessible Computing, 2016. 1", + "[45] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google's neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144, 2016. 6", + "[46] Kayo Yin and Jesse Read. Better sign language translation with stmc-transformer. arXiv preprint arXiv:2004.00588, 2020.2", + "[47] Tao Zhang, Shiqing Wei, and Shunping Ji. E2EC: An end-to-end contour-based method for high-quality high-speed instance segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[48] Hao Zhou, Wengang Zhou, and Houqiang Li. Dynamic pseudo label decoding for continuous sign language recognition. In IEEE International conference on Multimedia and Expo, 2019. 2", + "[49] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 5, 6, 7", + "[50] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for sign language recognition and translation. IEEE Transactions on Multimedia, 2021. 2, 6, 7", + "[51] Jinhua Zhu, Yingce Xia, Lijun Wu, Di He, Tao Qin, Wengang Zhou, Houqiang Li, and Tie-Yan Liu. Incorporating bert into neural machine translation. arXiv preprint arXiv:2002.06823, 2020. 4, 6, 7" + ], + "bbox": [ + 501, + 92, + 890, + 627 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15601", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Sign Language Translation with Iterative Prototype/f2c1bf6d-f55b-40d7-9372-42bff493c6d3_model.json b/2023/Sign Language Translation with Iterative Prototype/f2c1bf6d-f55b-40d7-9372-42bff493c6d3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5d1c5584ae305ce05799cfe9db49def2e7ecad69 --- /dev/null +++ b/2023/Sign Language Translation with Iterative Prototype/f2c1bf6d-f55b-40d7-9372-42bff493c6d3_model.json @@ -0,0 +1,2222 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.223, + 0.131, + 0.748, + 0.154 + ], + "angle": 0, + "content": "Sign Language Translation with Iterative Prototype" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.179, + 0.867, + 0.199 + ], + "angle": 0, + "content": "Huijie Yao\\(^{1}\\) Wengang Zhou\\(^{1,2,*}\\) Hao Feng\\(^{1}\\) Hezhen Hu\\(^{1}\\) Hao Zhou\\(^{1}\\) Houqiang Li\\(^{1,2,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.2, + 0.865, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\) CAS Key Laboratory of Technology in GIPAS, EEIS Department, University of Science and Technology of China \\(^{2}\\) Institute of Artificial Intelligence, Hefei Comprehensive National Science Center" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.236, + 0.812, + 0.253 + ], + "angle": 0, + "content": "{yaohuijie,haof,alexhu,zhouh156}@mail.ustc.edu.cn，{zhwg,lihq}@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.683 + ], + "angle": 0, + "content": "This paper presents IP-SLT, a simple yet effective framework for sign language translation (SLT). Our IP-SLT adopts a recurrent structure and enhances the semantic representation (prototype) of the input sign language video via an iterative refinement manner. Our idea mimics the behavior of human reading, where a sentence can be digested repeatedly, till reaching accurate understanding. Technically, IP-SLT consists of feature extraction, prototype initialization, and iterative prototype refinement. The initialization module generates the initial prototype based on the visual feature extracted by the feature extraction module. Then, the iterative refinement module leverages the cross-attention mechanism to polish the previous prototype by aggregating it with the original video feature. Through repeated refinement, the prototype finally converges to a more stable and accurate state, leading to a fluent and appropriate translation. In addition, to leverage the sequential dependence of prototypes, we further propose an iterative distillation loss to compress the knowledge of the final iteration into previous ones. As the autoregressive decoding process is executed only once in inference, our IP-SLT is ready to improve various SLT systems with acceptable overhead. Extensive experiments are conducted on public benchmarks to demonstrate the effectiveness of the IP-SLT." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.713, + 0.21, + 0.729 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.739, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Sign language translation (SLT) aims to automatically generate spoken language translations based on sign language videos, which holds both social significance and academic value. On the one hand, a high-quality SLT system can greatly facilitate communication between deafmute and hearing individuals [4, 11, 12, 41]. On the other hand, SLT as an interdisciplinary research topic necessitates a comprehensive understanding of computer vision [31, 13, 36, 19] and natural language processing [43, 40], given its" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.289, + 0.878, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.325, + 0.769, + 0.337 + ], + "angle": 0, + "content": "(a) The Previous Pipeline" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.351, + 0.878, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.435, + 0.768, + 0.448 + ], + "angle": 0, + "content": "(b) Our IP-SLT Pipeline" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.454, + 0.895, + 0.567 + ], + "angle": 0, + "content": "Figure 1. Illustration of the pipeline of the previous works and our IP-SLT. (a) The previous studies rely on a one-pass forward process to generate the final translation. (b) To mitigate the vision-text gap, we introduce the iterative refinement module into the original SLT system. The refinement module updates the current prototype conditioned on the sign language video, which can be run iteratively to obtain a better representation of the semantic meaning of the sign language video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.893, + 0.64 + ], + "angle": 0, + "content": "involvement with vision and text modalities. As a result, SLT has emerged as a vital research topic, garnering increasing attention [5, 49, 31, 2, 21, 20, 42]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.895, + 0.839 + ], + "angle": 0, + "content": "SLT is a challenging task, which faces a tough domain gap between the input video and output text, as well as a limited dataset scale due to costly data collection and annotation [49, 7, 44, 21]. Since SLT is typically viewed as a sequence-to-sequence mapping problem, the existing SLT systems [23, 5, 7] commonly adopt the one-pass forward pipeline based on encoder-decoder architecture [40, 30] (as shown in Fig. 1 (a)). In such a framework, the encoder transforms the sign video into its semantic representation (prototype), which is then fed into the decoder to obtain the final translation. However, due to the inherent gap between vision and text, it may be hard to conduct such mapping within the vanilla one-pass architecture." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this study, we present IP-SLT with the iterative prototype to boost sign language translation (as shown in Fig. 1 (b)), which is inspired by the human reading process. During this process, we note that repeatedly digging into the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.406, + 0.902 + ], + "angle": 0, + "content": "*Corresponding authors: Wengang Zhou and Houqiang Li" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15592" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.408 + ], + "angle": 0, + "content": "source materials is necessary for accurate understanding. Similarly, when we are trying to translate a sign language video into a sentence, we commonly do not directly write it down. Instead, we would recall and go back to the original sign video to check our answers. To implement the above idea, our IP-SLT adopts a recurrent structure that enhances the semantic representation (prototype) of the input sign language video via an iterative refinement process. IP-SLT generally contains three main components, including feature extraction, prototype initialization, and iterative prototype refinement. Given a sign video to be translated, we first extract its visual representation, which is then used for generating an initial raw prototype. Subsequently, we iteratively leverage the attention mechanism [43] to update the prototype toward the semantic meaning of the sign video. At each iteration, we refine the previous prototype by aggregating it with the original visual representation. In this way, the network repeatedly digs the semantic context from the sign video to polish the prototype. Through iterative refinement, the prototype finally converges to a stable and accurate state, producing a high-quality translation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.409, + 0.473, + 0.635 + ], + "angle": 0, + "content": "In addition, our IP-SLT introduces a novel design discussed next. Firstly, to leverage the sequential dependence between different iterations, we further propose the iterative distillation loss which allows the previous prototypes to obtain supervision from the final one. Since the final prototype converges to a more stable and accurate state, it is possible for IP-SLT to achieve better performance. Secondly, during training, all predicted prototypes are transformed into their corresponding translations to provide guidance for each iteration. Our inference process is neat since only the final prototype is used for the autoregressive decoding process. Thirdly, our IP-SLT can easily work with different visual backbones. Through end-to-end optimization, our IP-SLT achieves significant performance improvements over the baselines." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.635, + 0.396, + 0.65 + ], + "angle": 0, + "content": "In summary, our contributions are three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.656, + 0.469, + 0.716 + ], + "angle": 0, + "content": "- We propose IP-SLT, a novel framework to ameliorate sign language translation, which iteratively refines the prototypes by aggregating the previous translation progress and the original visual representation." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.724, + 0.469, + 0.768 + ], + "angle": 0, + "content": "- We propose an iterative distillation loss to enhance the basic supervision, by leveraging the sequential dependence between the outputs at each iteration." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.777, + 0.469, + 0.836 + ], + "angle": 0, + "content": "- We conduct extensive experiments to validate the proposed method, and show encouraging improved results on the two prevalent benchmarks, i.e., CSL-Daily [49] and PHOENIX-2014T [5]." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.656, + 0.469, + 0.836 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.846, + 0.22, + 0.86 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this section, we briefly review the related works, i.e., sign language translation and iterative refinement methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.439 + ], + "angle": 0, + "content": "Sign Language Translation. Camgoz et al. [5] pioneer the neural SLT task and publish the neural dataset PHOENIX-2014T and regard the SLT as a sequence-to-sequence problem. They implement the neural SLT system using the encoder-decoder paradigm [3]. This paradigm is adopted by subsequent studies which focus on addressing the challenges of data scarcity and domain gap. Considering the lack of frame-level annotation in sign language datasets, Li et al. [27] design the temporal semantic pyramid structure to obtain more discriminative features. Camgoz et al. [7] explore the mutual benefits of SLT and continuous sign language recognition through joint optimization. Zhou et al. [49] leverage gloss annotation to transform the monolingual texts into pseudo-videos. According to the characteristic of sign language, several works [46, 6, 50] propose multi-channel SLT systems which explicitly extract and align the key parts of sign language expression. Jin et al. [23] leverage the additional prior knowledge to obtain high-quality translations. Chen et al. [9] propose a transfer learning baseline for SLT by leveraging external resources from related tasks. Chen et al. [10] further combine the raw videos and the keypoint sequences to achieve better semantic understanding with auxiliary supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.501 + ], + "angle": 0, + "content": "In contrast, our proposed approach employs an iterative refinement process that utilizes the previous prototype as an additional clue to enhance the accuracy of the mapping between sign videos and their translations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Iterative Refinement Methods. The idea of iterative refinement is applied to various computer vision tasks, such as image generation [1, 18, 37], instance segmentation [34, 28, 29, 47], image rectification [15], etc., which shows promising performance improvements. CARN [1] maintains training stability in super-resolution tasks and improves the quality of output images. Ling et al. [28] regard the object instance segmentation task as a regression task and proposes the Curve-GCN to iteratively predict the locations of all vertices. DeepSanke [34] uses the deep network to iteratively enclose the object boundary based on an initial contour. The previous studies [48, 35, 14, 36, 31] in continuous sign language recognition task follow an iterative training scheme to enhance the discriminative power of feature extraction modules which use the convolutional neural network and their variants [8, 22, 39]. They leverage alignment proposals given by the connectionist temporal classification (CTC) [17] decoding as supervision at frame-wise granularity, which cannot be directly applied to the SLT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Different from the aforementioned methods, we explore how to reduce the vision-text gap by proposing an iterative refinement module to the existing SLT system. To reduce the complexity, we design the refinement module in a shared-weight manner. Moreover, we put forward the iterative distillation loss to leverage the sequential dependence between different iterations." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15593" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.145, + 0.089, + 0.833, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.262, + 0.895, + 0.364 + ], + "angle": 0, + "content": "Figure 2. An overview of the proposed IP-SLT framework. Given a sign video \\( \\mathbf{X} \\), the feature extraction module is responsible for embedding the input into visual representation \\( \\mathbf{F} \\). The initialization module (the encoder \\( \\mathcal{E}_1 \\) and decoder \\( \\mathcal{D}_1 \\)) generates the initial prototype \\( \\mathbf{E}^0 \\) and the raw translation \\( \\mathbf{Y}^0 \\). The refinement module (the encoder \\( \\mathcal{E}_2 \\) and decoder \\( \\mathcal{D}_2 \\)) first takes the initial prototype \\( \\mathbf{E}^0 \\) as input and generates a prototype for the current step by fusing it with the original visual representation \\( \\mathbf{F} \\). Through \\( K \\) times refinement, the prototype sequence \\( \\mathbf{E} = \\{\\mathbf{E}^0, \\mathbf{E}^1, \\dots, \\mathbf{E}^K\\} \\) and corresponding translation sequence \\( \\mathbf{Y} = \\{\\mathbf{Y}^0, \\mathbf{Y}^1, \\dots, \\mathbf{Y}^K\\} \\) are obtained. In light of the fact that the decoding part of IP-SLT consists of \\( K + 1 \\) branches based on the iteration order, we introduce the iterative distillation loss to improve the underlying supervision. It should be noted that the parts enclosed in dashed boxes can be removed in inference." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.383, + 0.214, + 0.401 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.471, + 0.471 + ], + "angle": 0, + "content": "In this section, we first introduce the overall architecture of our IP-SLT, and then separately elaborate individual components. Finally, we propose the design of the training objective and inference strategy for the IP-SLT." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.481, + 0.283, + 0.497 + ], + "angle": 0, + "content": "3.1. Framework Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.687 + ], + "angle": 0, + "content": "The primary objective of the SLT system is to acquire knowledge about the mapping \\( f: \\mathcal{X} \\mapsto \\mathcal{Y} \\), where \\( \\mathcal{X} \\) and \\( \\mathcal{Y} \\) denote the collections of \\( N \\) sign language videos and spoken language sentences associated with vocabulary \\( \\mathcal{V} \\), respectively. Most SLT systems adopt the encoder-decoder architecture [40], where the input \\( X \\in \\mathcal{X} \\) is first encoded to derive a high-level context representation. It is then passed to the decoder to generate the output \\( Y \\in \\mathcal{Y} \\). The encoder and decoder can be specialized using different types of neural networks, such as GRU [3], CNN [16], and Transformer [43]. Considering the performance of existing SLT systems, we adopt the Transformer as well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.471, + 0.901 + ], + "angle": 0, + "content": "With the goal of narrowing the domain gap between vision and text, we augment the original translation process with an iterative refinement step. Fig. 2 provides an overview of the proposed IP-SLT model, which consists of three stages, namely feature extraction, prototype initialization, and iterative prototype refinement. As with previous approaches, the initialization and iterative refinement module adopt the encoder-decoder architecture. Given the sign language video \\( \\mathbf{X} = \\{\\pmb{x}_t\\}_{t=1}^{T_x} \\) with \\( T_x \\) frames, the feature extraction module embeds it into the spatial-temporal feature \\( \\mathbf{F} = \\{\\pmb{f}_t\\}_{t=1}^{T_f} \\). Next, the encoder \\( \\mathcal{E}_1 \\) and decoder \\( \\mathcal{D}_1 \\) are employed in the initialization module to derive the initial prototype \\( \\mathbf{E}^0 \\) and initial translation \\( \\mathbf{Y}^0 \\) from the visual feature \\( \\mathbf{F} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.895, + 0.629 + ], + "angle": 0, + "content": "Subsequently, the refinement module iteratively refines the previous prototype and generates the final translation \\( \\mathbf{Y}^{K} = \\{\\mathbf{y}_{t}^{K}\\}_{t=1}^{T_{y,K}} \\) with \\( T_{y,K} \\) words after total \\( K \\) iterations. At the \\( k \\)-th iteration, the encoder \\( \\mathcal{E}_2 \\) estimates the prototype \\( \\mathbf{E}^k \\) for the current step by augmenting the original visual feature \\( \\mathbf{F} \\) with the previous prototype \\( \\mathbf{E}^{k-1} \\). Next, the decoder \\( \\mathcal{D}_2 \\) predicts the corresponding translation \\( \\mathbf{Y}^k \\). Finally, the translation sequence \\( \\mathbf{Y} = \\{\\mathbf{Y}^0, \\mathbf{Y}^1, \\dots, \\mathbf{Y}^K\\} \\) is obtained according to the prototype sequence \\( \\mathbf{E} = \\{\\mathbf{E}^0, \\mathbf{E}^1, \\dots, \\mathbf{E}^K\\} \\). For the IP-SLT optimization, by dividing the decoding part into \\( K+1 \\) branches, we add iterative distillation supervision from the final translation to the middle translations. Since the refinement process takes place in the encoder \\( \\mathcal{E}_2 \\), in inference, the proposed IP-SLT can generate the translation directly from the \\( K \\)-th prototype which causes acceptable overhead." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.637, + 0.685, + 0.652 + ], + "angle": 0, + "content": "3.2. Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.894, + 0.814 + ], + "angle": 0, + "content": "The feature extraction module embeds a series of video frames \\( \\mathbf{X} \\in \\mathbb{R}^{T_x \\times H \\times W \\times 3} \\) with width \\( W \\) and height \\( H \\) into its visual feature \\( \\mathbf{F} \\in \\mathbb{R}^{T_f \\times C} \\) with the dimension of feature \\( C \\). Since its goal is to extract a distinguishable representation for SLT, we can draw on the visual backbone used in CSLR [14, 26, 25, 13] to extract the valid representation. Generally, with sliding window size \\( w \\) and stride size \\( s \\), the sign video is split into \\( T_f = \\left\\lceil \\frac{T_x}{s} \\right\\rceil \\) clips. By passing sign videos through it, the spatial-temporal embeddings \\( \\mathbf{F} = \\{\\mathbf{f}_t\\}_{t=1}^{T_f} \\) are extracted as:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.821, + 0.892, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\boldsymbol {f} _ {t} \\right\\} _ {t = 1} ^ {T _ {f}} = \\operatorname {E x t r a c t o r} \\left(\\left\\{\\boldsymbol {x} _ {t} \\right\\} _ {t = 1} ^ {T _ {x}}\\right). \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.719, + 0.864 + ], + "angle": 0, + "content": "3.3. Prototype Initialization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.87, + 0.894, + 0.901 + ], + "angle": 0, + "content": "After a visual feature \\( \\pmb{F} \\in \\mathbb{R}^{T_f \\times C} \\) is extracted by the feature extraction module, it is first fed into the initialization" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15594" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.13, + 0.092, + 0.416, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.222, + 0.47, + 0.25 + ], + "angle": 0, + "content": "Figure 3. An illustration of the prototype update process in the \\(l\\)-th layer of the encoder \\(\\mathcal{E}_2\\) at the \\(k\\)-th iteration." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.275, + 0.468, + 0.304 + ], + "angle": 0, + "content": "module. The initialization module consists of an encoder \\(\\mathcal{E}_1\\) and a decoder \\(\\mathcal{D}_1\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.305, + 0.468, + 0.416 + ], + "angle": 0, + "content": "The visual representation \\(\\pmb{F}\\) is first fed into the encoder \\(\\mathcal{E}_1\\), and encoded into \\(T_{f}\\) raw states \\(\\pmb{E}^{0} = \\{e_{t}^{0}\\}_{t=1}^{T_{f}} \\in \\mathbb{R}^{T_{f} \\times C}\\). Then the decoder \\(\\mathcal{D}_1\\) reads the prototype \\(\\pmb{E}^{0}\\) and produces the initial translation \\(\\pmb{Y}^{0} = \\{\\pmb{y}_{t}^{0}\\}_{t=1}^{T_{y,0}}\\) according to the predicted logits \\(\\pmb{U}^{0} = \\{\\pmb{u}_{t}^{0}\\}_{t=1}^{T_{y,0}}\\). Specifically, it predicts the conditional probability of the translation sequence, which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.422, + 0.469, + 0.464 + ], + "angle": 0, + "content": "\\[\np ^ {0 - t h} \\left(\\boldsymbol {Y} ^ {0} | \\boldsymbol {X}\\right) = \\prod_ {t = 1} ^ {T _ {y, 0}} p \\left(\\boldsymbol {y} _ {t} ^ {0} \\mid \\boldsymbol {F}, \\boldsymbol {y} _ {0: t - 1} ^ {0}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.475, + 0.469, + 0.537 + ], + "angle": 0, + "content": "where \\(\\pmb{y}_{0:t-1}^{0} = \\{\\pmb{y}_{0}^{0}, \\pmb{y}_{1}^{0}, \\dots, \\pmb{y}_{t-1}^{0}\\}\\) denotes the previous output sub-sequence at the \\(t\\)-th step. The initial token \\(\\pmb{y}_{0}^{0}\\) represents the beginning of a sentence. The predicted probability of each token in the translation is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.544, + 0.469, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p \\left(\\boldsymbol {y} _ {t} ^ {0} \\mid \\boldsymbol {F}, \\boldsymbol {y} _ {1: t - 1} ^ {0}\\right) = \\operatorname {s o f t m a x} \\left(\\boldsymbol {u} _ {t} ^ {0}\\right) _ {\\boldsymbol {y} _ {t} ^ {0}} \\tag {3} \\\\ = \\operatorname {s o f t m a x} \\left(\\boldsymbol {h} _ {t} ^ {0} \\cdot \\boldsymbol {W}\\right) _ {\\boldsymbol {y} _ {t} ^ {0}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.595, + 0.469, + 0.716 + ], + "angle": 0, + "content": "where \\( h_t^0 \\in \\mathbb{R}^C \\) represents the output of the final layer at the \\( t \\)-th step, and \\( W \\in \\mathbb{R}^{C \\times |\\mathcal{V}|} \\) denotes a linear mapping to projects the hidden state \\( h_t^0 \\) into the predicted logits over the target vocabulary \\( \\mathcal{V} \\). The probability is calculated by applying the softmax(·) function to the logits. Notably, our goal is to obtain a more accurate prototype for SLT, thus, the decoder \\( \\mathcal{D}_1 \\) is only used in the training process to provide guidance for the initialization module." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.727, + 0.353, + 0.743 + ], + "angle": 0, + "content": "3.4. Iterative Prototype Refinement" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Once the initial prototype \\( \\pmb{E}^{0} \\) is obtained through the initialization module, we feed it together with the original visual representation \\( \\pmb{F} \\) into the iterative refinement module. The module maintains a single prototype which is iteratively refined. In this way, the coarse semantic feature finally converges to a stable state where the prototype best fits the sign language semantics. The refinement module is divided into two sub-processes, i.e., iterative prototype aggregation in the encoder \\( \\mathcal{E}_2 \\) and translation generation in the decoder \\( \\mathcal{D}_2 \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "Iterative prototype aggregation. To utilize the prototype as a reference, each layer of the encoder \\(\\mathcal{E}_2\\) attends over the maintained semantic features through the attention mechanism. As shown in Fig. 3, we illustrate the prototype aggregation process in the \\(l\\)-th layer of the encoder \\(\\mathcal{E}_2\\) at the \\(k\\)-th iteration. The input of the refinement module consists of the original visual feature \\(\\pmb{F}\\) and the \\((k-1)\\)-th prototype \\(\\pmb{E}^{k-1}\\). The attention mechanism \\(\\mathrm{attn}(\\pmb{q},\\pmb{K},\\pmb{V})\\) is originally used in Transformer [43], which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.232, + 0.892, + 0.275 + ], + "angle": 0, + "content": "\\[\n\\operatorname {a t t n} (\\boldsymbol {q}, \\boldsymbol {K}, \\boldsymbol {V}) = \\sum_ {i = 1} ^ {| V |} \\alpha_ {i} \\boldsymbol {W} _ {v} \\boldsymbol {v} _ {i}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.278, + 0.808, + 0.296 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} = \\operatorname {s o f t m a x} \\left(\\left(\\boldsymbol {W} _ {q} \\boldsymbol {q}\\right) ^ {T} \\left(\\boldsymbol {W} _ {k} \\boldsymbol {k} _ {i}\\right)\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.303, + 0.892, + 0.401 + ], + "angle": 0, + "content": "where \\(W_{q}, W_{k}\\) and \\(W_{v}\\) are learnable parameters. To make better use of the \\((k - 1)\\)-th prototype, we consider the \\((k - 1)\\)-th prototype \\(E^{k - 1} = \\{e_t^{k - 1}\\}_{t = 1}^{T_f}\\) as key-value pair, and the original visual feature \\(F = \\{f_t\\}_{t = 1}^{T_f}\\) as the query. In this way, we inject the rich semantic information into the new prototype \\(E^k = \\{e_t^k\\}_{t = 1}^{T_f}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.402, + 0.892, + 0.446 + ], + "angle": 0, + "content": "The encoder \\(\\mathcal{E}_2\\) is composed of \\(L_{e}\\) identical layers, where \\(\\pmb{E}_{l}^{k}\\) denotes the output of the \\(l\\)-th layer at the \\(k\\)-th iteration. Similar to [51], the hidden state is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.454, + 0.892, + 0.494 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\widetilde {\\boldsymbol {E}} _ {l} ^ {k} = \\beta \\cdot \\operatorname {a t t n} _ {s} \\left(\\boldsymbol {E} _ {l - 1} ^ {k}, \\boldsymbol {E} _ {l - 1} ^ {k}, \\boldsymbol {E} _ {l - 1} ^ {k}\\right) \\tag {5} \\\\ + (1 - \\beta) \\cdot \\operatorname {a t t n} _ {c} \\left(\\boldsymbol {E} _ {l - 1} ^ {k}, \\boldsymbol {E} ^ {k - 1}, \\boldsymbol {E} ^ {k - 1}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.503, + 0.892, + 0.656 + ], + "angle": 0, + "content": "where \\(\\mathrm{attn}_s\\) and \\(\\mathrm{attn}_c\\) denote the self-attention sub-layer and cross-attention sub-layer used in Transformer [43], respectively. For the first layer of the encoder \\(\\mathcal{E}_2\\), \\(E_{l - 1}^k\\) is equal to the visual feature \\(\\pmb{F}\\). The output of the encoder \\(\\mathcal{E}_2\\) is \\(E_{L_e}^k\\) (i.e., \\(E^k\\)). \\(\\beta \\in [0,1]\\) is a hyperparameter that weights the importance of previous prototypes during training and inference. To further fuse and refine the prototype, it is linked with a fully connected sub-layer \\(\\mathrm{FFN}(\\cdot)\\) using the residual connection. The output of the \\(l\\)-th layer at the \\(k\\)-th iteration is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.665, + 0.891, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {E} _ {l} ^ {k} = \\operatorname {F F N} \\left(\\operatorname {L N} \\left(\\widetilde {\\boldsymbol {E}} _ {l} ^ {k} + \\boldsymbol {E} _ {l - 1} ^ {k}\\right)\\right) + \\operatorname {L N} \\left(\\widetilde {\\boldsymbol {E}} _ {l} ^ {k} + \\boldsymbol {E} _ {l - 1} ^ {k}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.693, + 0.828, + 0.708 + ], + "angle": 0, + "content": "where \\(\\mathrm{LN}(\\cdot)\\) is the layer normalization operation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.71, + 0.892, + 0.806 + ], + "angle": 0, + "content": "Translation generation. The decoder \\(\\mathcal{D}_2\\) iteratively takes the prototype \\(E^k\\) as input and generates the corresponding translation \\(\\pmb{Y}^{k}\\). Take the \\(k\\)-th iteration as an example. The decoder \\(\\mathcal{D}_2\\) predicts the conditional probability of translation \\(\\pmb{Y}^{k} = \\{\\pmb{y}_{t}^{k}\\}_{t = 1}^{T_{y,k}}\\) based on the predicted logits \\(\\pmb{U}^{k} = \\{\\pmb{u}_{t}^{k}\\}_{t = 1}^{T_{y,k}}\\), which is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.812, + 0.892, + 0.853 + ], + "angle": 0, + "content": "\\[\np ^ {k - t h} \\left(\\boldsymbol {Y} ^ {k} \\mid \\boldsymbol {X}\\right) = \\prod_ {t = 1} ^ {T _ {y, k}} p \\left(\\boldsymbol {y} _ {t} ^ {k} \\mid \\boldsymbol {F}, \\boldsymbol {E} ^ {k - 1}, \\boldsymbol {y} _ {0: t - 1} ^ {k}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.856, + 0.85, + 0.877 + ], + "angle": 0, + "content": "\\[\np (\\boldsymbol {y} _ {t} ^ {k} | \\boldsymbol {F}, \\boldsymbol {E} ^ {k - 1}, \\boldsymbol {y} _ {0: t - 1} ^ {k}) = s o f t m a x (\\boldsymbol {u} _ {t} ^ {k}) _ {\\boldsymbol {y} _ {t} ^ {k}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.885, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\pmb{u}_t^k\\) is the predicted logits of the decoder \\(\\mathcal{D}_2\\) at the \\(t\\)-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15595" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.248 + ], + "angle": 0, + "content": "th step. The decoder \\(\\mathcal{D}_2\\) iteratively takes the prototype given by the encoder \\(\\mathcal{E}_2\\) as input to generate the final translation \\(\\mathbf{Y}^K\\). With total \\(K\\) iterations, the outputs of encoder \\(\\mathcal{E}_1\\) and \\(\\mathcal{E}_2\\) compose the prototype sequence \\(\\mathbf{E} = \\{\\mathbf{E}^k\\}_{k=0}^K\\), where \\(\\mathbf{E}^k = \\{\\mathbf{e}_t^k\\}_{t=1}^{T_f}\\) is the semantic feature at the \\(k\\)-th iteration. Accordingly, the decoder \\(\\mathcal{D}_1\\) and \\(\\mathcal{D}_2\\) generate \\(K+1\\) translations \\(\\mathbf{Y} = \\{\\mathbf{Y}^k\\}_{k=0}^K\\), where \\(\\mathbf{Y}^k = \\{\\mathbf{y}_t^k\\}_{t=1}^{T_y,k}\\) is the translation at the \\(k\\)-th iteration during training. Note that after \\(K\\) iterations, the decoder \\(\\mathcal{D}_2\\) obtains the converged prototype \\(\\mathbf{E}^K\\) and generates the translation only once in inference." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.258, + 0.262, + 0.274 + ], + "angle": 0, + "content": "3.5. Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.282, + 0.473, + 0.45 + ], + "angle": 0, + "content": "We introduce two kinds of losses in the training period of the IP-SLT system. Firstly, the cross entropy loss is adopted to supervise the final generated sentence. Secondly, we put forward an iterative distillation loss. As the decoder \\(\\mathcal{D}_1\\) and \\(\\mathcal{D}_2\\) generate translation sequence \\(\\mathbf{Y} = \\{\\mathbf{Y}^0,\\mathbf{Y}^1,\\dots ,\\mathbf{Y}^K\\}\\), we naturally divide the sequence \\(\\mathbf{Y}\\) into an initial prediction, \\(K - 1\\) intermediate predictions, and the final prediction according to the order of iteration. Conceptually, the \\(K - 1\\) intermediate predictions are regarded as the student model and distill knowledge from the final prediction which is regarded as the teacher model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.454, + 0.473, + 0.577 + ], + "angle": 0, + "content": "Cross entropy loss. As mentioned above, the IP-SLT generates the translation sequence based on the conditional probability provided by the decoder \\(\\mathcal{D}_1\\) and \\(\\mathcal{D}_2\\). The cross-entropy loss [43] is computed with the ground truth from \\(N\\) training samples and the outputs of the decoder \\(\\mathcal{D}_1\\) and \\(\\mathcal{D}_2\\). Its training objective in the proposed approach is to maximize the log-likelihood which is equal to minimizing the cross entropy loss formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.583, + 0.47, + 0.602 + ], + "angle": 0, + "content": "\\[\nL _ {C E, k} = - \\log p ^ {k - t h} (\\hat {\\mathbf {Y}} | \\mathbf {X}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.616, + 0.47, + 0.662 + ], + "angle": 0, + "content": "where \\(\\hat{Y}\\) denotes translation annotation. We apply this at the output of the initialization module and the \\(K\\)-th output of the refinement module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.473, + 0.789 + ], + "angle": 0, + "content": "Iterative distillation loss. Since the KL (Kullback-Leibler) divergence loss can affect the teacher's networks to the student's networks, we compute it between the \\( K - 1 \\) intermediate predictions and the final prediction. As the final translation is based on the previous prototypes, we are able to get better translations by approximating more vital characterization capabilities of middle prototypes. The iterative distillation loss (IDL) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.801, + 0.47, + 0.842 + ], + "angle": 0, + "content": "\\[\nL _ {I D L} = \\sum_ {k = 1} ^ {K - 1} K L \\left(\\boldsymbol {U} ^ {k}, \\boldsymbol {U} ^ {K}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.855, + 0.473, + 0.903 + ], + "angle": 0, + "content": "where \\(\\pmb{U}^{k}\\) and \\(\\pmb{U}^{K}\\) are the predicted output of the decoder \\(\\mathcal{D}_2\\) at the \\(k\\)-th and \\(K\\)-th iteration, respectively. By computing the Kullback-Leibler divergence between the \\(\\pmb{U}^{k}\\) and" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.136 + ], + "angle": 0, + "content": "\\(\\pmb{U}^{K}\\), the IP-SLT is encouraged to approximate the performance of the final iteration. We apply this at the \\(K - 1\\) shallow outputs of the refinement module." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.137, + 0.892, + 0.152 + ], + "angle": 0, + "content": "Overall, the loss function of our IP-SLT is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.164, + 0.892, + 0.181 + ], + "angle": 0, + "content": "\\[\nL = L _ {C E, 0} + L _ {C E, K} + \\lambda \\cdot L _ {I D L}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.192, + 0.892, + 0.253 + ], + "angle": 0, + "content": "Under the guidance of CE loss, we first train the initial generation module until convergence as a warm start, and then apply the loss function as Equ. (10) for optimizing the IP-SLT system in an end-to-end manner." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.262, + 0.68, + 0.278 + ], + "angle": 0, + "content": "3.6. Inference Strategy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.285, + 0.892, + 0.42 + ], + "angle": 0, + "content": "Since the refinement process only involves in the encoder \\(\\mathcal{E}_2\\), the initialization and iterative refinement modules behave differently during training and inference. The decoder \\(\\mathcal{D}_1\\) of the initialization module is leveraged to provide guidance for the encoder \\(\\mathcal{E}_1\\) during the training process, while it is not required in inference. Similarly, the autoregressive decoder \\(\\mathcal{D}_2\\) of the iterative refinement module generates translations for each iteration during training, while it just decodes the final translation once in inference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.421, + 0.892, + 0.527 + ], + "angle": 0, + "content": "In inference, given a sign language video to be translated, the feature extraction module first converts it into a visual feature. The encoder \\(\\mathcal{E}_1\\) of the initialization module transforms it into the raw prototype. Then, the encoder \\(\\mathcal{E}_2\\) of the refinement module further iteratively refines it by fusing it with the original visual feature. Finally, the decoder \\(\\mathcal{D}_2\\) generates the translation based on the final prototype." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.54, + 0.634, + 0.557 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.565, + 0.691, + 0.581 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.892, + 0.759 + ], + "angle": 0, + "content": "Dataset. We evaluate our approach on two pubic sign language translation datasets, i.e., PHOENIX14T [5] and CSL-Daily [49]. Both datasets provide gloss-level and spoken-sentence-level annotations. The PHOENIX14T dataset [5] is the first large-scale neural SLT dataset created with 9 German sign language interpreters. The dataset is split into a training set (7,096), a development set (519), and a test set (642). The CSL-Daily dataset [49] is a Chinese SLT dataset containing 20,654 annotated sign language videos. We follow the previous experimental setting [49] and split it into the training, development, and test set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.761, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Evaluation metrics. Following the work [49], we quantitatively assess the quality of translations according to the BLEU- \\(N\\) [32] and ROUGE [38]. The BLEU- \\(N\\) (\\(N\\) ranges from 1 to 4) cares more about the accuracy of the predicted translation while the ROUGE cares more about the consistency of sentences. For both evaluation metrics, a higher value indicates a better performance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Training settings. We implement our approach on Pytorch [33]. The encoder and decoder of the initialization" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15596" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.872, + 0.232 + ], + "angle": 0, + "content": "
MethodsLLMDevTest
ROUGEBLEU-1BLEU-2BLEU-3BLEU-4ROUGEBLEU-1BLEU-2BLEU-3BLEU-4
Joint-SLRT [7]✘-47.2634.4027.0522.38-46.6133.7326.1921.32
PET [23]✘-----49.9749.5437.1929.3024.02
BN-TIN-Transf. [49]✘46.8746.9033.9826.4921.7846.9847.5734.6426.7821.68
STMC [50]✘48.2447.6036.4329.1824.0946.6546.9836.0928.7023.65
IP-SLT✘54.4354.1041.5633.6628.2253.7254.2541.5133.4527.97
MMTLB [9]✓53.1053.9541.1233.1427.6152.6553.9741.7533.8428.39
TwoStream-SLT [10]✓54.0854.3241.9934.1528.6653.4854.9042.4334.4628.95
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.234, + 0.892, + 0.263 + ], + "angle": 0, + "content": "Table 1. Performance comparison of IP-SLT with methods for SLT on PHOENIX-2014T. 'LLM' denotes adopting pre-trained large language models." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.27, + 0.872, + 0.4 + ], + "angle": 0, + "content": "
MethodsLLMDevTest
ROUGEBLEU-1BLEU-2BLEU-3BLEU-4ROUGEBLEU-1BLEU-2BLEU-3BLEU-4
SL-Luong [5]X34.2834.2219.7212.247.9634.5434.1619.5711.847.56
Joint-SLRT [7]X37.0637.4724.6716.8611.8836.7437.3824.3616.5511.79
BN-TIN-Transf. [49]X37.2940.6626.5618.0612.7337.6740.7426.9618.4813.19
IP-SLTX44.3345.2631.7722.8716.7444.0944.8531.5022.6616.72
MMTLB [9]✓53.3853.8140.8431.2924.4253.2553.3140.4130.8723.92
TwoStream-SLT [10]✓55.1055.2142.3132.7125.7655.7255.4442.5932.8725.79
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.402, + 0.893, + 0.43 + ], + "angle": 0, + "content": "Table 2. Performance comparison of IP-SLT with methods for SLT on CSL-Daily. 'LLM' denotes adopting pre-trained large language models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.473, + 0.678 + ], + "angle": 0, + "content": "and refinement module consist of 3 layers, respectively. The dimension of the feed-forward network is 2048. The visual feature of the STMC [50] is 1024-dimension while the visual feature of BN-TIN-Transf. [49] and VAC [31] are 512-dimension. To alleviate over-fitting, we set dropout and attention head to 0.1 and 8, respectively. The training optimizer is Adam [24]. During training, the learning rate is fixed to \\(5 \\times 10^{-5}\\). To ensure the features provided by the previous prototype and original sign video are fully utilized, we apply the drop-net [51] during training, which effects Equ. (5). During training, for any layer in the encoder \\(\\mathcal{E}_2\\), with probability \\(\\beta\\), the hidden state \\(\\widetilde{\\pmb{E}}_l^k\\) in Equ. (5) is the output of the self-attention sub-layer \\(\\mathrm{attn}_s\\); with probability \\(1 - \\beta\\), it is the output of the cross-attention sub-layer \\(\\mathrm{attn}_c\\). In inference, the hidden state \\(\\widetilde{\\pmb{E}}_l^k\\) is computed as Equ. (5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.47, + 0.789 + ], + "angle": 0, + "content": "Inference details. In inference, we use the beam search strategy [45] to improve the decoding accuracy. For the PHOENIX-2014T dataset and the CSL-Daily dataset, we set the beam search width and the length penalty to 3 and 1.0, respectively. To reduce the computational complexity, our IP-SLT just decodes once in inference for each input sign language video." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.446, + 0.817 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We compare the proposed IP-SLT with the previous SLT systems on two public benchmarks, i.e., PHOENIX14T [5] and CSL-Daily [49], and the performance of our IP-SLT is shown in Tab. 1 and Tab. 2, respectively. For PHOENIX14T and CSL-Daily dataset, we adopt the STMC [50] and BN-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.892, + 0.538 + ], + "angle": 0, + "content": "TIN-Transf. [49] as the baseline, respectively. Our IP-SLT follows the sign-to-text (S2T) paradigm, which directly transforms the sign language video into translation. Note MMTLB [9] and TwoStream-SLT [10] adopt pre-trained large-scale language models that leverage more model parameters and extra resources than IP-SLT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.893, + 0.708 + ], + "angle": 0, + "content": "By combining all proposed components together, our IP-SLT achieves substantial improvements against the baseline. The IP-SLT achieves 28.22 and 16.74 BLEU-4 scores on the DEV set of PHOENIX14T and CSL-Daily, respectively. The quantitative results demonstrate that our IP-SLT achieves promising performance improvements. Our IP-SLT delivers promising performance gains on DEV and test sets by leveraging the iterative refinement process. The results prove the advantage of aggregating the previous translation progress and the original visual representation, which distinguishes our IP-SLT from previous SLT systems." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.72, + 0.665, + 0.735 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.744, + 0.892, + 0.805 + ], + "angle": 0, + "content": "In this section, we put forward several ablation studies on the DEV set of PHOENIX-2014T. Unless otherwise specified, we adopt the STMC [50] as the baseline for the following experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Impact of network architecture. The main difference between our proposed method and the existing work is to leverage the previous information as an additional clue to enhance the current prototype. To evaluate the effectiveness of each component, we gradually add the refinement module and the iterative distillation loss to the baseline SLT sys" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15597" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.466, + 0.179 + ], + "angle": 0, + "content": "
SettingROUGEBLEU-1BLEU-2BLEU-3BLEU-4
Baseline50.1050.2137.1229.4124.31
+Refinement51.2251.0438.4030.6125.39
+IDL54.4354.1041.5633.6628.22
6-6 Layers50.4950.5837.5729.6424.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.185, + 0.47, + 0.242 + ], + "angle": 0, + "content": "Table 3. Effect of our proposed components. 'Refinement' denotes applying the refinement process. 'IDL' denotes applying the iterative distillation loss. '6-6 layers' denotes enlarging the encoder and decoder of the baseline system from 3 to 6 layers." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.25, + 0.466, + 0.372 + ], + "angle": 0, + "content": "
ModelRMROUGEBLEU-1BLEU-2BLEU-3BLEU-4
STMCw/o50.1050.2137.1229.4124.31
w/54.4354.1041.5633.6628.22
BN-TIN-Transf.w/o47.4147.9934.9427.3322.35
w/52.0652.0639.0131.0825.69
VAC-Transf.w/o49.4850.0137.0029.1223.91
w/53.6853.6041.2833.4728.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.376, + 0.469, + 0.42 + ], + "angle": 0, + "content": "Table 4. Generalization of IP-SLT. 'RM' denotes leveraging the refinement process. 'w' and 'w/o' denote the baseline SLT system with and without a refinement process, respectively." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.429, + 0.466, + 0.521 + ], + "angle": 0, + "content": "
λROUGEBLEU-1BLEU-2BLEU-3BLEU-4
051.3050.3238.0130.2925.14
553.0752.3140.0832.2927.02
1053.7153.5140.9232.9427.47
1554.4354.1041.5633.6628.22
2054.4254.1641.5133.4427.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.526, + 0.439, + 0.541 + ], + "angle": 0, + "content": "Table 5. The weight \\( \\lambda \\) of iterative distillation loss to CE loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.468, + 0.7 + ], + "angle": 0, + "content": "tem. Directly applying the refinement process to the baseline delivers a performance gain of 1.08 BLEU-4. We further apply the iterative distillation loss to improve the performance. The results suggest that adding distillation supervision can be helpful with a gain of 2.83 BLEU-4. Besides, to keep the number of parameters unchanged, we enlarge the depth from 3 to 6 layers and evaluate the performance. Naively enlarging the model scale slightly improves the performance \\((+0.22\\) BLEU-4). The results are shown in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Generalization of the IP-SLT. We conduct three sets of experiments by changing the visual backbone to evaluate the generalization of the proposed IP-SLT approach in Tab. 4. Specifically, the BN-TIN-Transf. [49] uses a basic CNN network to get the dense representation of sign video. The STMC [50] extracts and aligns the key parts of sign language expression to achieve better performance. The VAC [31] proposes two auxiliary supervision methods to enhance the feature extraction module. VAC-Transf. replaces the feature extractor in BN-TIN-Transf. [49] with the visual backbone of VAC [31]. Applying our proposed IP-SLT methods to the BN-TIN-Transf., VAC-Transf. and STMC, we achieve 25.69, 28.07, and 28.22 BLEU-4 scores" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.09, + 0.888, + 0.215 + ], + "angle": 0, + "content": "
βROUGEBLEU-1BLEU-2BLEU-3BLEU-4
0.050.1050.2137.1229.4124.31
0.254.3553.8141.4033.5228.02
0.454.3954.0241.2433.1927.70
0.554.4354.1041.5633.6628.22
0.653.9653.6340.9433.1027.72
0.854.3153.3940.9232.9827.53
-50.9750.6738.2430.5225.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.221, + 0.892, + 0.263 + ], + "angle": 0, + "content": "Table 6. The weight \\(\\beta\\) of the original visual feature to the previous prototype. '0.0' denotes the assessment of the baseline. '-' denotes setting \\(\\beta\\) as 0.5 without using the drop-net [51]." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.272, + 0.888, + 0.354 + ], + "angle": 0, + "content": "
SettingROUGEBLEU-1BLEU-2BLEU-3BLEU-4
None50.1050.2137.1229.4124.31
Con-input54.3753.8341.1032.9727.39
Con-feature53.9752.7640.5632.8127.47
Add-feature54.4354.1041.5633.6628.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.36, + 0.892, + 0.444 + ], + "angle": 0, + "content": "Table 7. Effect of the different refinement methods. 'Con-input' denotes directly concatenating the original visual feature and previous prototype as input. 'Con-feature' denotes concatenating the original feature and the previous feature given by the cross-attention mechanism in each layer. On top of that, 'Add-feature' denotes changing it to an addition operation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.466, + 0.891, + 0.512 + ], + "angle": 0, + "content": "on the DEV set, surpassing the baselines by 3.34, 4.16, 3.91, respectively. Using a high-quality visual backbone delivers further quality gains." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.515, + 0.892, + 0.606 + ], + "angle": 0, + "content": "Impact of \\(\\lambda\\). In our experiments, the weight \\(\\lambda\\) of iterative distillation loss is set to 15. It is a hyper-parameter that is designed to balance the effect of cross-entropy loss and the iterative distillation loss. We conduct experiments by varying the weight \\(\\lambda\\). Tab. 5 shows that our IP-SLT achieves the best performance when the weight \\(\\lambda\\) is set to 15." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.892, + 0.746 + ], + "angle": 0, + "content": "Impact of \\(\\beta\\). In the above experiments, the weight \\(\\beta\\) is fixed to 0.5. The weight \\(\\beta\\) represents the importance of the previous prototype compared with the original visual feature. As a hyper-parameter of our proposed methods, the weight \\(\\beta\\) is examined with a set of different values in Tab. 6. When the weight is 0.5, the performance is the highest one. This indicates that the previous prototype to SLT is as important as the original visual feature. Besides, to fully use the previous prototype, the drop-net is required." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Impact of refinement method. We also examine a set of refinement methods for IP-SLT considering the fusion mechanism as a key part of our proposed method in Tab. 7. Directly concatenating the original visual feature and previous prototype in the time dimension improves the performance of SLT from 24.31 to 27.39. We further add the cross-attention mechanism to explicitly leverage the useful information from the previous prototype. And then, we concatenate the representation from the original visual feature and the representation from the previous prototype in fea" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15598" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.466, + 0.18 + ], + "angle": 0, + "content": "
KROUGEBLEU-1BLEU-2BLEU-3BLEU-4
050.1050.2137.1229.4124.31
151.2251.0438.4030.6125.39
253.7353.3940.7632.9827.66
354.4354.1041.5633.6628.22
454.6353.9141.4033.5328.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.186, + 0.468, + 0.214 + ], + "angle": 0, + "content": "Table 8. Effect of the iteration number \\( K \\) in the iterative refinement module." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.222, + 0.466, + 0.375 + ], + "angle": 0, + "content": "
ModelSWI-P(M)T-P(M)FLOPs(B)ROUGEBLEU-4
STMC-92.692.628.750.1024.31
w/128.7172.632.654.4328.22
w/o238.6357.932.654.8128.48
BN-TIN-Transf.-28.328.39.647.4122.35
w/37.753.410.652.0625.69
w/o66.1110.010.652.2425.96
VAC-Transf.-28.328.39.649.4823.91
w/37.753.410.653.6828.07
w/o66.1110.010.654.8228.27
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.381, + 0.469, + 0.479 + ], + "angle": 0, + "content": "Table 9. Comparing the baseline and non-shared weight prototype refinement method with IP-SLT. 'SW' denotes the sharing weight across all iterations. '-' denotes the baseline SLT system. 'w/' and 'w/o' denote the refinement module in different iterations with and without shared parameters, respectively. 'I-P' and 'T-P' denote the parameter amount calculated in inference and during training, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.469, + 0.56 + ], + "angle": 0, + "content": "ture dimension, which delivers performance gains of 3.16 BLEU-4. On top of that, changing the concatenating process to an element-wise addition operation achieves a further quality gain of 3.91." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.469, + 0.701 + ], + "angle": 0, + "content": "Impact of iteration number \\( K \\). The iteration number \\( K \\) is an important hyper-parameter and is fixed to 3 in the previous experiments. We conduct experiments with different iteration numbers to explore the effect of iteration number \\( K \\). Tab. 8 shows that there is a culmination in the performance at iteration 3. Before this culmination, the performance improvement of our proposed method increases fast. When the iteration number is bigger than 3, the performance of our method is slightly weakened (-0.21 BLEU-4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Impact of sharing weights between different iterations. In Tab. 9, we examine the storage and translation quality of our proposed IP-SLT with different baselines i.e., STMC, BN-TIN-Transf, and VAC-Transf. In order to ensure the efficiency of the model, our refinement module shares weight across all iterations. Since the feature extraction module of each group is identical, the parameters of the feature extraction module are not included in the calculation of the storage of different models. As in inference, several parts can be removed from IP-SLT without changing the performance, we report the number of parameters in both training and inference, respectively. The results indicate that our iterative process causes acceptable overhead with re" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.09, + 0.892, + 0.19 + ], + "angle": 0, + "content": "
TypeText
GTich=winsche ihnen einen tatsächien abend und machen sie es gut.
Baselinejetzt wünsche ich ihnen noch einen tatsächien abend.
Ourihnen einen tatsächien abend und machen sie es gut.
GTin der neuenWoche wird es milder aber es bleibt wechselhaft.
Baselinedann wird es wieder milder.
Ourin der neuenWoche wird es dann wieder milder.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.196, + 0.892, + 0.238 + ], + "angle": 0, + "content": "Table 10. Qualitative evaluation. 'GT' denotes the spoken language translation annotation. 'Baseline' and 'Our' denote the translation result of baseline and our IP-SLT, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.892, + 0.366 + ], + "angle": 0, + "content": "markable performance improvements. We further conduct experiments in which the parameters of each iteration are independent. The results demonstrate that the proposed parameter-shared method achieves near performance gain (28.22 v.s 28.48 BLEU-4, 25.69 v.s 25.96 BLEU-4, and 28.07 v.s 28.27 BLEU-4) with non-shared one but leveraging much less parameter number." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.892, + 0.491 + ], + "angle": 0, + "content": "Computation comparison with baseline. The FLOPs is a key factor of computation efficiency. We conduct experiments to compare the computation of the proposed IP-SLT method with different baselines. Similarly, we exclude the FLOPs of the feature extraction module and report relevant computational costs in inference. Tab. 9 shows that leveraging the iterative refinement process can cause acceptable computation costs while achieving promising performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.495, + 0.892, + 0.57 + ], + "angle": 0, + "content": "Case Study. To provide a more intuitionistic view of our proposed method, we list some translation samples of the proposed IP-SLT method in Tab. 10. We observe that based on the previous prototype and original visual representation, the IP-SLT can generate more accurate and fluent sentences." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.585, + 0.619, + 0.601 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.821 + ], + "angle": 0, + "content": "In this work, we propose a new framework IP-SLT which introduces the iterative refinement into a conventional SLT system. With the goal to polish the semantic representation by leveraging the previous results, we present IP-SLT to support the iterative refinement process. The proposed method is differentiable and optimized in an end-to-end manner to achieve its best performance. On top of it, we put forward the iterative distillation loss to further improve the translation quality by leveraging the sequential dependence between the outputs of each iteration. In inference, the autoregressive decoding process is required once to generate the translation based on the final prototype, applying IP-SLT does not significantly affect efficiency. The experimental results demonstrate the effectiveness of the IP-SLT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments: This work was supported by NSFC under Contract U20A20183 and 62021001. It was also supported by the GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, and the Supercomputing Center of the USTC." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15599" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Image super-resolution via progressive cascading residual network. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Samuel Albanie, Gül Varol, Liliane Momeni, Triantafyllos Afouras, Joon Son Chung, Neil Fox, and Andrew Zisserman. BSL-1K: Scaling up co-articulated sign language recognition using mouthing cues. In Computer Vision European Conference, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.286 + ], + "angle": 0, + "content": "[3] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.47, + 0.37 + ], + "angle": 0, + "content": "[4] Danielle Bragg, Oscar Koller, Mary Bellard, Larwan Berke, Patrick Boudreault, Annelies Braffort, Naomi Caselli, Matt Huenerfauth, Hernisa Kacorri, Tessa Verhoef, et al. Sign language recognition, generation, and translation: An interdisciplinary perspective. In International ACM SIGACCESS Conference on Computers and Accessibility, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.47, + 0.428 + ], + "angle": 0, + "content": "[5] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In IEEE Conference on Computer Vision and Pattern Recognition, 2018. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.43, + 0.47, + 0.486 + ], + "angle": 0, + "content": "[6] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Multi-channel transformers for multi-articulatory sign language translation. In Computer Vision-ECCV 2020 Workshops, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.47, + 0.556 + ], + "angle": 0, + "content": "[7] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.47, + 0.613 + ], + "angle": 0, + "content": "[8] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.47, + 0.671 + ], + "angle": 0, + "content": "[9] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.673, + 0.47, + 0.729 + ], + "angle": 0, + "content": "[10] Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie Liu, and Brian Mak. Two-stream network for sign language recognition and translation. Advances in Neural Information Processing Systems, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.731, + 0.47, + 0.772 + ], + "angle": 0, + "content": "[11] Helen Cooper, Brian Holt, and Richard Bowden. Sign language recognition. Visual Analysis of Humans: Looking at People, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[12] Kearsy Cormier, Neil Fox, Bencie Woll, Andrew Zisserman, Necati Cihan Camgoz, and Richard Bowden. ExTOL: Automatic recognition of British sign language using the bsl corpus. In Workshop on Sign Language Translation and Avatar Technology, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Runpeng Cui, Hu Liu, and Changshui Zhang. Recurrent convolutional neural networks for continuous sign language recognition by staged optimization. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[14] Runpeng Cui, Hu Liu, and Changshui Zhang. A deep neural framework for continuous sign language recognition by iterative training. IEEE Transactions on Multimedia, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[15] Hao Feng, Wengang Zhou, Jiajun Deng, Qi Tian, and Houqiang Li. DocScanner: robust document image rectification with progressive learning. arXiv preprint arXiv:2110.14968, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[16] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. Convolutional sequence to sequence learning. In International Conference on Machine Learning, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[17] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In International Conference on Machine learning, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[18] Karol Gregor, Ivo Danihelka, Alex Graves, Danilo Rezende, and Daan Wierstra. DRAW: A recurrent neural network for image generation. In International Conference on Machine Learning, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[19] Jun He, Richang Hong, Xueliang Liu, Mingliang Xu, and Qianru Sun. Revisiting local descriptor for improved few-shot classification. ACM Transactions on Multimedia Computing, Communications, and Applications, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[20] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. Signbert+: Hand-model-aware self-supervised pretraining for sign language understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[21] Jie Huang, Wengang Zhou, Qilin Zhang, Houqiang Li, and Weiping Li. Video-based sign language recognition without temporal segmentation. In AAAI Conference on Artificial Intelligence, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[22] Shuiwang Ji, Wei Xu, Ming Yang, and Kai Yu. 3D convolutional neural networks for human action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[23] Tao Jin, Zhou Zhao, Meng Zhang, and Xingshan Zeng. Prior knowledge and memory enriched transformer for sign language translation. In Findings of the Association for Computational Linguistics, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[25] Oscar Koller, Necati Cihan Camgoz, Hermann Ney, and Richard Bowden. Weakly supervised learning with multistream cnn-lstm-hmms to discover sequential parallelism in sign language videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[26] Oscar Koller, Hermann Ney, and Richard Bowden. Deep hand: How to train a CNN on 1 million hand images when your data is continuous and weakly labelled. In IEEE Conference on Computer Vision and Pattern Recognition, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Dongxu Li, Chenchen Xu, Xin Yu, Kaihao Zhang, Benjamin Swift, Hanna Suominen, and Hongdong Li. TSPNet: Hierarchical feature learning via temporal semantic pyramid for" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15600" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "sign language translation. Advances in Neural Information Processing Systems, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[28] Huan Ling, Jun Gao, Amlan Kar, Wenzheng Chen, and Sanja Fidler. Fast interactive object annotation with Curve-GCN. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[29] Zichen Liu, Jun Hao Liew, Xiangyu Chen, and Jiashi Feng. DANCE: A deep attentive contour model for efficient instance segmentation. In IEEE Winter Conference on Applications of Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.469, + 0.276 + ], + "angle": 0, + "content": "[30] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. arXiv preprint arXiv:1508.04025, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.279, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[31] Yuecong Min, Aiming Hao, Xiujuan Chai, and Xilin Chen. Visual alignment constraint for continuous sign language recognition. In IEEE International Conference on Computer Vision, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.335, + 0.469, + 0.39 + ], + "angle": 0, + "content": "[32] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: a method for automatic evaluation of machine translation. In Annual Meeting of the Association for Computational Linguistics, 2002. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[33] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[34] Sida Peng, Wen Jiang, Huaijin Pi, Xiuli Li, Hujun Bao, and Xiaowei Zhou. Deep snake for real-time instance segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[35] Junfu Pu, Wengang Zhou, and Houqiang Li. Dilated convolutional network with iterative optimization for continuous sign language recognition. In International Joint Conference on Artificial Intelligence, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.615 + ], + "angle": 0, + "content": "[36] Junfu Pu, Wengang Zhou, and Houqiang Li. Iterative alignment network for continuous sign language recognition. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[37] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive Image Deraining Networks: A better and simpler baseline. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[38] Lin CY ROUGE. A package for automatic evaluation of summaries. In Workshop on Text Summarization of ACL, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[39] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. Advances in Neural Information Processing Systems, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[40] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. Advances in Neural Information Processing Systems, 2014. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[41] Rachel Sutton-Spence and Bencie Woll. The linguistics of British Sign Language: an introduction. Cambridge University Press, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[42] Shengeng Tang, Dan Guo, Richang Hong, and Meng Wang. Graph-based multimodal sequential embedding for sign language translation. IEEE Transactions on Multimedia, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[44] Hanjie Wang, Xiujuan Chai, Xiaopeng Hong, Guoying Zhao, and Xilin Chen. Isolated sign language recognition with grassmann covariance matrices. ACM Transactions on Accessible Computing, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[45] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google's neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[46] Kayo Yin and Jesse Read. Better sign language translation with stmc-transformer. arXiv preprint arXiv:2004.00588, 2020.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.333, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[47] Tao Zhang, Shiqing Wei, and Shunping Ji. E2EC: An end-to-end contour-based method for high-quality high-speed instance segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.39, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[48] Hao Zhou, Wengang Zhou, and Houqiang Li. Dynamic pseudo label decoding for continuous sign language recognition. In IEEE International conference on Multimedia and Expo, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.447, + 0.892, + 0.514 + ], + "angle": 0, + "content": "[49] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.517, + 0.892, + 0.571 + ], + "angle": 0, + "content": "[50] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for sign language recognition and translation. IEEE Transactions on Multimedia, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.574, + 0.892, + 0.628 + ], + "angle": 0, + "content": "[51] Jinhua Zhu, Yingce Xia, Lijun Wu, Di He, Tao Qin, Wengang Zhou, Houqiang Li, and Tie-Yan Liu. Incorporating bert into neural machine translation. arXiv preprint arXiv:2002.06823, 2020. 4, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15601" + } + ] +] \ No newline at end of file diff --git a/2023/Sign Language Translation with Iterative Prototype/f2c1bf6d-f55b-40d7-9372-42bff493c6d3_origin.pdf b/2023/Sign Language Translation with Iterative Prototype/f2c1bf6d-f55b-40d7-9372-42bff493c6d3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1a593bd8c774f01f05bf510e3e14bf043f925b17 --- /dev/null +++ b/2023/Sign Language Translation with Iterative Prototype/f2c1bf6d-f55b-40d7-9372-42bff493c6d3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aef94d186d64928e68d01871095b5e5efdc1cd396ac61fd28bcc48968ced82de +size 806966 diff --git a/2023/Sign Language Translation with Iterative Prototype/full.md b/2023/Sign Language Translation with Iterative Prototype/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f37cf0b07b1b548c51d9f006a96023e3d3c6ec0a --- /dev/null +++ b/2023/Sign Language Translation with Iterative Prototype/full.md @@ -0,0 +1,332 @@ +# Sign Language Translation with Iterative Prototype + +Huijie Yao $^{1}$ Wengang Zhou $^{1,2,*}$ Hao Feng $^{1}$ Hezhen Hu $^{1}$ Hao Zhou $^{1}$ Houqiang Li $^{1,2,*}$ + +$^{1}$ CAS Key Laboratory of Technology in GIPAS, EEIS Department, University of Science and Technology of China $^{2}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center + +{yaohuijie,haof,alexhu,zhouh156}@mail.ustc.edu.cn，{zhwg,lihq}@ustc.edu.cn + +# Abstract + +This paper presents IP-SLT, a simple yet effective framework for sign language translation (SLT). Our IP-SLT adopts a recurrent structure and enhances the semantic representation (prototype) of the input sign language video via an iterative refinement manner. Our idea mimics the behavior of human reading, where a sentence can be digested repeatedly, till reaching accurate understanding. Technically, IP-SLT consists of feature extraction, prototype initialization, and iterative prototype refinement. The initialization module generates the initial prototype based on the visual feature extracted by the feature extraction module. Then, the iterative refinement module leverages the cross-attention mechanism to polish the previous prototype by aggregating it with the original video feature. Through repeated refinement, the prototype finally converges to a more stable and accurate state, leading to a fluent and appropriate translation. In addition, to leverage the sequential dependence of prototypes, we further propose an iterative distillation loss to compress the knowledge of the final iteration into previous ones. As the autoregressive decoding process is executed only once in inference, our IP-SLT is ready to improve various SLT systems with acceptable overhead. Extensive experiments are conducted on public benchmarks to demonstrate the effectiveness of the IP-SLT. + +# 1. Introduction + +Sign language translation (SLT) aims to automatically generate spoken language translations based on sign language videos, which holds both social significance and academic value. On the one hand, a high-quality SLT system can greatly facilitate communication between deafmute and hearing individuals [4, 11, 12, 41]. On the other hand, SLT as an interdisciplinary research topic necessitates a comprehensive understanding of computer vision [31, 13, 36, 19] and natural language processing [43, 40], given its + +![](images/d3b242dc929ef28a6c798e2106b9377eacab9351036ce2d8ede668759cf7ef71.jpg) +(a) The Previous Pipeline + +![](images/fb659078773baacc81bce84eef610a8c414e47deea8c8d3b6e8d2a152c3510af.jpg) +(b) Our IP-SLT Pipeline +Figure 1. Illustration of the pipeline of the previous works and our IP-SLT. (a) The previous studies rely on a one-pass forward process to generate the final translation. (b) To mitigate the vision-text gap, we introduce the iterative refinement module into the original SLT system. The refinement module updates the current prototype conditioned on the sign language video, which can be run iteratively to obtain a better representation of the semantic meaning of the sign language video. + +involvement with vision and text modalities. As a result, SLT has emerged as a vital research topic, garnering increasing attention [5, 49, 31, 2, 21, 20, 42]. + +SLT is a challenging task, which faces a tough domain gap between the input video and output text, as well as a limited dataset scale due to costly data collection and annotation [49, 7, 44, 21]. Since SLT is typically viewed as a sequence-to-sequence mapping problem, the existing SLT systems [23, 5, 7] commonly adopt the one-pass forward pipeline based on encoder-decoder architecture [40, 30] (as shown in Fig. 1 (a)). In such a framework, the encoder transforms the sign video into its semantic representation (prototype), which is then fed into the decoder to obtain the final translation. However, due to the inherent gap between vision and text, it may be hard to conduct such mapping within the vanilla one-pass architecture. + +In this study, we present IP-SLT with the iterative prototype to boost sign language translation (as shown in Fig. 1 (b)), which is inspired by the human reading process. During this process, we note that repeatedly digging into the + +source materials is necessary for accurate understanding. Similarly, when we are trying to translate a sign language video into a sentence, we commonly do not directly write it down. Instead, we would recall and go back to the original sign video to check our answers. To implement the above idea, our IP-SLT adopts a recurrent structure that enhances the semantic representation (prototype) of the input sign language video via an iterative refinement process. IP-SLT generally contains three main components, including feature extraction, prototype initialization, and iterative prototype refinement. Given a sign video to be translated, we first extract its visual representation, which is then used for generating an initial raw prototype. Subsequently, we iteratively leverage the attention mechanism [43] to update the prototype toward the semantic meaning of the sign video. At each iteration, we refine the previous prototype by aggregating it with the original visual representation. In this way, the network repeatedly digs the semantic context from the sign video to polish the prototype. Through iterative refinement, the prototype finally converges to a stable and accurate state, producing a high-quality translation. + +In addition, our IP-SLT introduces a novel design discussed next. Firstly, to leverage the sequential dependence between different iterations, we further propose the iterative distillation loss which allows the previous prototypes to obtain supervision from the final one. Since the final prototype converges to a more stable and accurate state, it is possible for IP-SLT to achieve better performance. Secondly, during training, all predicted prototypes are transformed into their corresponding translations to provide guidance for each iteration. Our inference process is neat since only the final prototype is used for the autoregressive decoding process. Thirdly, our IP-SLT can easily work with different visual backbones. Through end-to-end optimization, our IP-SLT achieves significant performance improvements over the baselines. + +In summary, our contributions are three-fold: + +- We propose IP-SLT, a novel framework to ameliorate sign language translation, which iteratively refines the prototypes by aggregating the previous translation progress and the original visual representation. +- We propose an iterative distillation loss to enhance the basic supervision, by leveraging the sequential dependence between the outputs at each iteration. +- We conduct extensive experiments to validate the proposed method, and show encouraging improved results on the two prevalent benchmarks, i.e., CSL-Daily [49] and PHOENIX-2014T [5]. + +# 2. Related Work + +In this section, we briefly review the related works, i.e., sign language translation and iterative refinement methods. + +Sign Language Translation. Camgoz et al. [5] pioneer the neural SLT task and publish the neural dataset PHOENIX-2014T and regard the SLT as a sequence-to-sequence problem. They implement the neural SLT system using the encoder-decoder paradigm [3]. This paradigm is adopted by subsequent studies which focus on addressing the challenges of data scarcity and domain gap. Considering the lack of frame-level annotation in sign language datasets, Li et al. [27] design the temporal semantic pyramid structure to obtain more discriminative features. Camgoz et al. [7] explore the mutual benefits of SLT and continuous sign language recognition through joint optimization. Zhou et al. [49] leverage gloss annotation to transform the monolingual texts into pseudo-videos. According to the characteristic of sign language, several works [46, 6, 50] propose multi-channel SLT systems which explicitly extract and align the key parts of sign language expression. Jin et al. [23] leverage the additional prior knowledge to obtain high-quality translations. Chen et al. [9] propose a transfer learning baseline for SLT by leveraging external resources from related tasks. Chen et al. [10] further combine the raw videos and the keypoint sequences to achieve better semantic understanding with auxiliary supervision. + +In contrast, our proposed approach employs an iterative refinement process that utilizes the previous prototype as an additional clue to enhance the accuracy of the mapping between sign videos and their translations. + +Iterative Refinement Methods. The idea of iterative refinement is applied to various computer vision tasks, such as image generation [1, 18, 37], instance segmentation [34, 28, 29, 47], image rectification [15], etc., which shows promising performance improvements. CARN [1] maintains training stability in super-resolution tasks and improves the quality of output images. Ling et al. [28] regard the object instance segmentation task as a regression task and proposes the Curve-GCN to iteratively predict the locations of all vertices. DeepSanke [34] uses the deep network to iteratively enclose the object boundary based on an initial contour. The previous studies [48, 35, 14, 36, 31] in continuous sign language recognition task follow an iterative training scheme to enhance the discriminative power of feature extraction modules which use the convolutional neural network and their variants [8, 22, 39]. They leverage alignment proposals given by the connectionist temporal classification (CTC) [17] decoding as supervision at frame-wise granularity, which cannot be directly applied to the SLT. + +Different from the aforementioned methods, we explore how to reduce the vision-text gap by proposing an iterative refinement module to the existing SLT system. To reduce the complexity, we design the refinement module in a shared-weight manner. Moreover, we put forward the iterative distillation loss to leverage the sequential dependence between different iterations. + +![](images/848452f5099d44eb4ba752e0c5e7717723377bd198fca8e83909ab95951806e3.jpg) +Figure 2. An overview of the proposed IP-SLT framework. Given a sign video $\mathbf{X}$ , the feature extraction module is responsible for embedding the input into visual representation $\mathbf{F}$ . The initialization module (the encoder $\mathcal{E}_1$ and decoder $\mathcal{D}_1$ ) generates the initial prototype $\mathbf{E}^0$ and the raw translation $\mathbf{Y}^0$ . The refinement module (the encoder $\mathcal{E}_2$ and decoder $\mathcal{D}_2$ ) first takes the initial prototype $\mathbf{E}^0$ as input and generates a prototype for the current step by fusing it with the original visual representation $\mathbf{F}$ . Through $K$ times refinement, the prototype sequence $\mathbf{E} = \{\mathbf{E}^0, \mathbf{E}^1, \dots, \mathbf{E}^K\}$ and corresponding translation sequence $\mathbf{Y} = \{\mathbf{Y}^0, \mathbf{Y}^1, \dots, \mathbf{Y}^K\}$ are obtained. In light of the fact that the decoding part of IP-SLT consists of $K + 1$ branches based on the iteration order, we introduce the iterative distillation loss to improve the underlying supervision. It should be noted that the parts enclosed in dashed boxes can be removed in inference. + +# 3. Methodology + +In this section, we first introduce the overall architecture of our IP-SLT, and then separately elaborate individual components. Finally, we propose the design of the training objective and inference strategy for the IP-SLT. + +# 3.1. Framework Overview + +The primary objective of the SLT system is to acquire knowledge about the mapping $f: \mathcal{X} \mapsto \mathcal{Y}$ , where $\mathcal{X}$ and $\mathcal{Y}$ denote the collections of $N$ sign language videos and spoken language sentences associated with vocabulary $\mathcal{V}$ , respectively. Most SLT systems adopt the encoder-decoder architecture [40], where the input $X \in \mathcal{X}$ is first encoded to derive a high-level context representation. It is then passed to the decoder to generate the output $Y \in \mathcal{Y}$ . The encoder and decoder can be specialized using different types of neural networks, such as GRU [3], CNN [16], and Transformer [43]. Considering the performance of existing SLT systems, we adopt the Transformer as well. + +With the goal of narrowing the domain gap between vision and text, we augment the original translation process with an iterative refinement step. Fig. 2 provides an overview of the proposed IP-SLT model, which consists of three stages, namely feature extraction, prototype initialization, and iterative prototype refinement. As with previous approaches, the initialization and iterative refinement module adopt the encoder-decoder architecture. Given the sign language video $\mathbf{X} = \{\pmb{x}_t\}_{t=1}^{T_x}$ with $T_x$ frames, the feature extraction module embeds it into the spatial-temporal feature $\mathbf{F} = \{\pmb{f}_t\}_{t=1}^{T_f}$ . Next, the encoder $\mathcal{E}_1$ and decoder $\mathcal{D}_1$ are employed in the initialization module to derive the initial prototype $\mathbf{E}^0$ and initial translation $\mathbf{Y}^0$ from the visual feature $\mathbf{F}$ . + +Subsequently, the refinement module iteratively refines the previous prototype and generates the final translation $\mathbf{Y}^{K} = \{\mathbf{y}_{t}^{K}\}_{t=1}^{T_{y,K}}$ with $T_{y,K}$ words after total $K$ iterations. At the $k$ -th iteration, the encoder $\mathcal{E}_2$ estimates the prototype $\mathbf{E}^k$ for the current step by augmenting the original visual feature $\mathbf{F}$ with the previous prototype $\mathbf{E}^{k-1}$ . Next, the decoder $\mathcal{D}_2$ predicts the corresponding translation $\mathbf{Y}^k$ . Finally, the translation sequence $\mathbf{Y} = \{\mathbf{Y}^0, \mathbf{Y}^1, \dots, \mathbf{Y}^K\}$ is obtained according to the prototype sequence $\mathbf{E} = \{\mathbf{E}^0, \mathbf{E}^1, \dots, \mathbf{E}^K\}$ . For the IP-SLT optimization, by dividing the decoding part into $K+1$ branches, we add iterative distillation supervision from the final translation to the middle translations. Since the refinement process takes place in the encoder $\mathcal{E}_2$ , in inference, the proposed IP-SLT can generate the translation directly from the $K$ -th prototype which causes acceptable overhead. + +# 3.2. Feature Extraction + +The feature extraction module embeds a series of video frames $\mathbf{X} \in \mathbb{R}^{T_x \times H \times W \times 3}$ with width $W$ and height $H$ into its visual feature $\mathbf{F} \in \mathbb{R}^{T_f \times C}$ with the dimension of feature $C$ . Since its goal is to extract a distinguishable representation for SLT, we can draw on the visual backbone used in CSLR [14, 26, 25, 13] to extract the valid representation. Generally, with sliding window size $w$ and stride size $s$ , the sign video is split into $T_f = \left\lceil \frac{T_x}{s} \right\rceil$ clips. By passing sign videos through it, the spatial-temporal embeddings $\mathbf{F} = \{\mathbf{f}_t\}_{t=1}^{T_f}$ are extracted as: + +$$ +\left\{\boldsymbol {f} _ {t} \right\} _ {t = 1} ^ {T _ {f}} = \operatorname {E x t r a c t o r} \left(\left\{\boldsymbol {x} _ {t} \right\} _ {t = 1} ^ {T _ {x}}\right). \tag {1} +$$ + +# 3.3. Prototype Initialization + +After a visual feature $\pmb{F} \in \mathbb{R}^{T_f \times C}$ is extracted by the feature extraction module, it is first fed into the initialization + +![](images/381e7b884c3be4fac6d50fe0e688936241a4bfabfe3766ed68dd30a5c06a5e4d.jpg) +Figure 3. An illustration of the prototype update process in the $l$ -th layer of the encoder $\mathcal{E}_2$ at the $k$ -th iteration. + +module. The initialization module consists of an encoder $\mathcal{E}_1$ and a decoder $\mathcal{D}_1$ . + +The visual representation $\pmb{F}$ is first fed into the encoder $\mathcal{E}_1$ , and encoded into $T_{f}$ raw states $\pmb{E}^{0} = \{e_{t}^{0}\}_{t=1}^{T_{f}} \in \mathbb{R}^{T_{f} \times C}$ . Then the decoder $\mathcal{D}_1$ reads the prototype $\pmb{E}^{0}$ and produces the initial translation $\pmb{Y}^{0} = \{\pmb{y}_{t}^{0}\}_{t=1}^{T_{y,0}}$ according to the predicted logits $\pmb{U}^{0} = \{\pmb{u}_{t}^{0}\}_{t=1}^{T_{y,0}}$ . Specifically, it predicts the conditional probability of the translation sequence, which is formulated as: + +$$ +p ^ {0 - t h} \left(\boldsymbol {Y} ^ {0} | \boldsymbol {X}\right) = \prod_ {t = 1} ^ {T _ {y, 0}} p \left(\boldsymbol {y} _ {t} ^ {0} \mid \boldsymbol {F}, \boldsymbol {y} _ {0: t - 1} ^ {0}\right), \tag {2} +$$ + +where $\pmb{y}_{0:t-1}^{0} = \{\pmb{y}_{0}^{0}, \pmb{y}_{1}^{0}, \dots, \pmb{y}_{t-1}^{0}\}$ denotes the previous output sub-sequence at the $t$ -th step. The initial token $\pmb{y}_{0}^{0}$ represents the beginning of a sentence. The predicted probability of each token in the translation is computed as: + +$$ +\begin{array}{l} p \left(\boldsymbol {y} _ {t} ^ {0} \mid \boldsymbol {F}, \boldsymbol {y} _ {1: t - 1} ^ {0}\right) = \operatorname {s o f t m a x} \left(\boldsymbol {u} _ {t} ^ {0}\right) _ {\boldsymbol {y} _ {t} ^ {0}} \tag {3} \\ = \operatorname {s o f t m a x} \left(\boldsymbol {h} _ {t} ^ {0} \cdot \boldsymbol {W}\right) _ {\boldsymbol {y} _ {t} ^ {0}}, \\ \end{array} +$$ + +where $h_t^0 \in \mathbb{R}^C$ represents the output of the final layer at the $t$ -th step, and $W \in \mathbb{R}^{C \times |\mathcal{V}|}$ denotes a linear mapping to projects the hidden state $h_t^0$ into the predicted logits over the target vocabulary $\mathcal{V}$ . The probability is calculated by applying the softmax(·) function to the logits. Notably, our goal is to obtain a more accurate prototype for SLT, thus, the decoder $\mathcal{D}_1$ is only used in the training process to provide guidance for the initialization module. + +# 3.4. Iterative Prototype Refinement + +Once the initial prototype $\pmb{E}^{0}$ is obtained through the initialization module, we feed it together with the original visual representation $\pmb{F}$ into the iterative refinement module. The module maintains a single prototype which is iteratively refined. In this way, the coarse semantic feature finally converges to a stable state where the prototype best fits the sign language semantics. The refinement module is divided into two sub-processes, i.e., iterative prototype aggregation in the encoder $\mathcal{E}_2$ and translation generation in the decoder $\mathcal{D}_2$ . + +Iterative prototype aggregation. To utilize the prototype as a reference, each layer of the encoder $\mathcal{E}_2$ attends over the maintained semantic features through the attention mechanism. As shown in Fig. 3, we illustrate the prototype aggregation process in the $l$ -th layer of the encoder $\mathcal{E}_2$ at the $k$ -th iteration. The input of the refinement module consists of the original visual feature $\pmb{F}$ and the $(k-1)$ -th prototype $\pmb{E}^{k-1}$ . The attention mechanism $\mathrm{attn}(\pmb{q},\pmb{K},\pmb{V})$ is originally used in Transformer [43], which is formulated as: + +$$ +\operatorname {a t t n} (\boldsymbol {q}, \boldsymbol {K}, \boldsymbol {V}) = \sum_ {i = 1} ^ {| V |} \alpha_ {i} \boldsymbol {W} _ {v} \boldsymbol {v} _ {i}, \tag {4} +$$ + +$$ +\alpha_ {i} = \operatorname {s o f t m a x} \left(\left(\boldsymbol {W} _ {q} \boldsymbol {q}\right) ^ {T} \left(\boldsymbol {W} _ {k} \boldsymbol {k} _ {i}\right)\right), +$$ + +where $W_{q}, W_{k}$ and $W_{v}$ are learnable parameters. To make better use of the $(k - 1)$ -th prototype, we consider the $(k - 1)$ -th prototype $E^{k - 1} = \{e_t^{k - 1}\}_{t = 1}^{T_f}$ as key-value pair, and the original visual feature $F = \{f_t\}_{t = 1}^{T_f}$ as the query. In this way, we inject the rich semantic information into the new prototype $E^k = \{e_t^k\}_{t = 1}^{T_f}$ . + +The encoder $\mathcal{E}_2$ is composed of $L_{e}$ identical layers, where $\pmb{E}_{l}^{k}$ denotes the output of the $l$ -th layer at the $k$ -th iteration. Similar to [51], the hidden state is computed as: + +$$ +\begin{array}{l} \widetilde {\boldsymbol {E}} _ {l} ^ {k} = \beta \cdot \operatorname {a t t n} _ {s} \left(\boldsymbol {E} _ {l - 1} ^ {k}, \boldsymbol {E} _ {l - 1} ^ {k}, \boldsymbol {E} _ {l - 1} ^ {k}\right) \tag {5} \\ + (1 - \beta) \cdot \operatorname {a t t n} _ {c} \left(\boldsymbol {E} _ {l - 1} ^ {k}, \boldsymbol {E} ^ {k - 1}, \boldsymbol {E} ^ {k - 1}\right), \\ \end{array} +$$ + +where $\mathrm{attn}_s$ and $\mathrm{attn}_c$ denote the self-attention sub-layer and cross-attention sub-layer used in Transformer [43], respectively. For the first layer of the encoder $\mathcal{E}_2$ , $E_{l - 1}^k$ is equal to the visual feature $\pmb{F}$ . The output of the encoder $\mathcal{E}_2$ is $E_{L_e}^k$ (i.e., $E^k$ ). $\beta \in [0,1]$ is a hyperparameter that weights the importance of previous prototypes during training and inference. To further fuse and refine the prototype, it is linked with a fully connected sub-layer $\mathrm{FFN}(\cdot)$ using the residual connection. The output of the $l$ -th layer at the $k$ -th iteration is formulated as: + +$$ +\boldsymbol {E} _ {l} ^ {k} = \operatorname {F F N} \left(\operatorname {L N} \left(\widetilde {\boldsymbol {E}} _ {l} ^ {k} + \boldsymbol {E} _ {l - 1} ^ {k}\right)\right) + \operatorname {L N} \left(\widetilde {\boldsymbol {E}} _ {l} ^ {k} + \boldsymbol {E} _ {l - 1} ^ {k}\right), \tag {6} +$$ + +where $\mathrm{LN}(\cdot)$ is the layer normalization operation. + +Translation generation. The decoder $\mathcal{D}_2$ iteratively takes the prototype $E^k$ as input and generates the corresponding translation $\pmb{Y}^{k}$ . Take the $k$ -th iteration as an example. The decoder $\mathcal{D}_2$ predicts the conditional probability of translation $\pmb{Y}^{k} = \{\pmb{y}_{t}^{k}\}_{t = 1}^{T_{y,k}}$ based on the predicted logits $\pmb{U}^{k} = \{\pmb{u}_{t}^{k}\}_{t = 1}^{T_{y,k}}$ , which is computed as: + +$$ +p ^ {k - t h} \left(\boldsymbol {Y} ^ {k} \mid \boldsymbol {X}\right) = \prod_ {t = 1} ^ {T _ {y, k}} p \left(\boldsymbol {y} _ {t} ^ {k} \mid \boldsymbol {F}, \boldsymbol {E} ^ {k - 1}, \boldsymbol {y} _ {0: t - 1} ^ {k}\right), \tag {7} +$$ + +$$ +p (\boldsymbol {y} _ {t} ^ {k} | \boldsymbol {F}, \boldsymbol {E} ^ {k - 1}, \boldsymbol {y} _ {0: t - 1} ^ {k}) = s o f t m a x (\boldsymbol {u} _ {t} ^ {k}) _ {\boldsymbol {y} _ {t} ^ {k}}, +$$ + +where $\pmb{u}_t^k$ is the predicted logits of the decoder $\mathcal{D}_2$ at the $t$ - + +th step. The decoder $\mathcal{D}_2$ iteratively takes the prototype given by the encoder $\mathcal{E}_2$ as input to generate the final translation $\mathbf{Y}^K$ . With total $K$ iterations, the outputs of encoder $\mathcal{E}_1$ and $\mathcal{E}_2$ compose the prototype sequence $\mathbf{E} = \{\mathbf{E}^k\}_{k=0}^K$ , where $\mathbf{E}^k = \{\mathbf{e}_t^k\}_{t=1}^{T_f}$ is the semantic feature at the $k$ -th iteration. Accordingly, the decoder $\mathcal{D}_1$ and $\mathcal{D}_2$ generate $K+1$ translations $\mathbf{Y} = \{\mathbf{Y}^k\}_{k=0}^K$ , where $\mathbf{Y}^k = \{\mathbf{y}_t^k\}_{t=1}^{T_y,k}$ is the translation at the $k$ -th iteration during training. Note that after $K$ iterations, the decoder $\mathcal{D}_2$ obtains the converged prototype $\mathbf{E}^K$ and generates the translation only once in inference. + +# 3.5. Training Objective + +We introduce two kinds of losses in the training period of the IP-SLT system. Firstly, the cross entropy loss is adopted to supervise the final generated sentence. Secondly, we put forward an iterative distillation loss. As the decoder $\mathcal{D}_1$ and $\mathcal{D}_2$ generate translation sequence $\mathbf{Y} = \{\mathbf{Y}^0,\mathbf{Y}^1,\dots ,\mathbf{Y}^K\}$ , we naturally divide the sequence $\mathbf{Y}$ into an initial prediction, $K - 1$ intermediate predictions, and the final prediction according to the order of iteration. Conceptually, the $K - 1$ intermediate predictions are regarded as the student model and distill knowledge from the final prediction which is regarded as the teacher model. + +Cross entropy loss. As mentioned above, the IP-SLT generates the translation sequence based on the conditional probability provided by the decoder $\mathcal{D}_1$ and $\mathcal{D}_2$ . The cross-entropy loss [43] is computed with the ground truth from $N$ training samples and the outputs of the decoder $\mathcal{D}_1$ and $\mathcal{D}_2$ . Its training objective in the proposed approach is to maximize the log-likelihood which is equal to minimizing the cross entropy loss formulated as: + +$$ +L _ {C E, k} = - \log p ^ {k - t h} (\hat {\mathbf {Y}} | \mathbf {X}), \tag {8} +$$ + +where $\hat{Y}$ denotes translation annotation. We apply this at the output of the initialization module and the $K$ -th output of the refinement module. + +Iterative distillation loss. Since the KL (Kullback-Leibler) divergence loss can affect the teacher's networks to the student's networks, we compute it between the $K - 1$ intermediate predictions and the final prediction. As the final translation is based on the previous prototypes, we are able to get better translations by approximating more vital characterization capabilities of middle prototypes. The iterative distillation loss (IDL) is formulated as: + +$$ +L _ {I D L} = \sum_ {k = 1} ^ {K - 1} K L \left(\boldsymbol {U} ^ {k}, \boldsymbol {U} ^ {K}\right), \tag {9} +$$ + +where $\pmb{U}^{k}$ and $\pmb{U}^{K}$ are the predicted output of the decoder $\mathcal{D}_2$ at the $k$ -th and $K$ -th iteration, respectively. By computing the Kullback-Leibler divergence between the $\pmb{U}^{k}$ and + +$\pmb{U}^{K}$ , the IP-SLT is encouraged to approximate the performance of the final iteration. We apply this at the $K - 1$ shallow outputs of the refinement module. + +Overall, the loss function of our IP-SLT is formulated as: + +$$ +L = L _ {C E, 0} + L _ {C E, K} + \lambda \cdot L _ {I D L}. \tag {10} +$$ + +Under the guidance of CE loss, we first train the initial generation module until convergence as a warm start, and then apply the loss function as Equ. (10) for optimizing the IP-SLT system in an end-to-end manner. + +# 3.6. Inference Strategy + +Since the refinement process only involves in the encoder $\mathcal{E}_2$ , the initialization and iterative refinement modules behave differently during training and inference. The decoder $\mathcal{D}_1$ of the initialization module is leveraged to provide guidance for the encoder $\mathcal{E}_1$ during the training process, while it is not required in inference. Similarly, the autoregressive decoder $\mathcal{D}_2$ of the iterative refinement module generates translations for each iteration during training, while it just decodes the final translation once in inference. + +In inference, given a sign language video to be translated, the feature extraction module first converts it into a visual feature. The encoder $\mathcal{E}_1$ of the initialization module transforms it into the raw prototype. Then, the encoder $\mathcal{E}_2$ of the refinement module further iteratively refines it by fusing it with the original visual feature. Finally, the decoder $\mathcal{D}_2$ generates the translation based on the final prototype. + +# 4. Experiments + +# 4.1. Experimental Setup + +Dataset. We evaluate our approach on two pubic sign language translation datasets, i.e., PHOENIX14T [5] and CSL-Daily [49]. Both datasets provide gloss-level and spoken-sentence-level annotations. The PHOENIX14T dataset [5] is the first large-scale neural SLT dataset created with 9 German sign language interpreters. The dataset is split into a training set (7,096), a development set (519), and a test set (642). The CSL-Daily dataset [49] is a Chinese SLT dataset containing 20,654 annotated sign language videos. We follow the previous experimental setting [49] and split it into the training, development, and test set. + +Evaluation metrics. Following the work [49], we quantitatively assess the quality of translations according to the BLEU- $N$ [32] and ROUGE [38]. The BLEU- $N$ ( $N$ ranges from 1 to 4) cares more about the accuracy of the predicted translation while the ROUGE cares more about the consistency of sentences. For both evaluation metrics, a higher value indicates a better performance. + +Training settings. We implement our approach on Pytorch [33]. The encoder and decoder of the initialization + +
MethodsLLMDevTest
ROUGEBLEU-1BLEU-2BLEU-3BLEU-4ROUGEBLEU-1BLEU-2BLEU-3BLEU-4
Joint-SLRT [7]✘-47.2634.4027.0522.38-46.6133.7326.1921.32
PET [23]✘-----49.9749.5437.1929.3024.02
BN-TIN-Transf. [49]✘46.8746.9033.9826.4921.7846.9847.5734.6426.7821.68
STMC [50]✘48.2447.6036.4329.1824.0946.6546.9836.0928.7023.65
IP-SLT✘54.4354.1041.5633.6628.2253.7254.2541.5133.4527.97
MMTLB [9]✓53.1053.9541.1233.1427.6152.6553.9741.7533.8428.39
TwoStream-SLT [10]✓54.0854.3241.9934.1528.6653.4854.9042.4334.4628.95
+ +Table 1. Performance comparison of IP-SLT with methods for SLT on PHOENIX-2014T. 'LLM' denotes adopting pre-trained large language models. + +
MethodsLLMDevTest
ROUGEBLEU-1BLEU-2BLEU-3BLEU-4ROUGEBLEU-1BLEU-2BLEU-3BLEU-4
SL-Luong [5]X34.2834.2219.7212.247.9634.5434.1619.5711.847.56
Joint-SLRT [7]X37.0637.4724.6716.8611.8836.7437.3824.3616.5511.79
BN-TIN-Transf. [49]X37.2940.6626.5618.0612.7337.6740.7426.9618.4813.19
IP-SLTX44.3345.2631.7722.8716.7444.0944.8531.5022.6616.72
MMTLB [9]✓53.3853.8140.8431.2924.4253.2553.3140.4130.8723.92
TwoStream-SLT [10]✓55.1055.2142.3132.7125.7655.7255.4442.5932.8725.79
+ +Table 2. Performance comparison of IP-SLT with methods for SLT on CSL-Daily. 'LLM' denotes adopting pre-trained large language models. + +and refinement module consist of 3 layers, respectively. The dimension of the feed-forward network is 2048. The visual feature of the STMC [50] is 1024-dimension while the visual feature of BN-TIN-Transf. [49] and VAC [31] are 512-dimension. To alleviate over-fitting, we set dropout and attention head to 0.1 and 8, respectively. The training optimizer is Adam [24]. During training, the learning rate is fixed to $5 \times 10^{-5}$ . To ensure the features provided by the previous prototype and original sign video are fully utilized, we apply the drop-net [51] during training, which effects Equ. (5). During training, for any layer in the encoder $\mathcal{E}_2$ , with probability $\beta$ , the hidden state $\widetilde{\pmb{E}}_l^k$ in Equ. (5) is the output of the self-attention sub-layer $\mathrm{attn}_s$ ; with probability $1 - \beta$ , it is the output of the cross-attention sub-layer $\mathrm{attn}_c$ . In inference, the hidden state $\widetilde{\pmb{E}}_l^k$ is computed as Equ. (5). + +Inference details. In inference, we use the beam search strategy [45] to improve the decoding accuracy. For the PHOENIX-2014T dataset and the CSL-Daily dataset, we set the beam search width and the length penalty to 3 and 1.0, respectively. To reduce the computational complexity, our IP-SLT just decodes once in inference for each input sign language video. + +# 4.2. Comparison with State-of-the-Art Methods + +We compare the proposed IP-SLT with the previous SLT systems on two public benchmarks, i.e., PHOENIX14T [5] and CSL-Daily [49], and the performance of our IP-SLT is shown in Tab. 1 and Tab. 2, respectively. For PHOENIX14T and CSL-Daily dataset, we adopt the STMC [50] and BN- + +TIN-Transf. [49] as the baseline, respectively. Our IP-SLT follows the sign-to-text (S2T) paradigm, which directly transforms the sign language video into translation. Note MMTLB [9] and TwoStream-SLT [10] adopt pre-trained large-scale language models that leverage more model parameters and extra resources than IP-SLT. + +By combining all proposed components together, our IP-SLT achieves substantial improvements against the baseline. The IP-SLT achieves 28.22 and 16.74 BLEU-4 scores on the DEV set of PHOENIX14T and CSL-Daily, respectively. The quantitative results demonstrate that our IP-SLT achieves promising performance improvements. Our IP-SLT delivers promising performance gains on DEV and test sets by leveraging the iterative refinement process. The results prove the advantage of aggregating the previous translation progress and the original visual representation, which distinguishes our IP-SLT from previous SLT systems. + +# 4.3. Ablation Studies + +In this section, we put forward several ablation studies on the DEV set of PHOENIX-2014T. Unless otherwise specified, we adopt the STMC [50] as the baseline for the following experiments. + +Impact of network architecture. The main difference between our proposed method and the existing work is to leverage the previous information as an additional clue to enhance the current prototype. To evaluate the effectiveness of each component, we gradually add the refinement module and the iterative distillation loss to the baseline SLT sys + +
SettingROUGEBLEU-1BLEU-2BLEU-3BLEU-4
Baseline50.1050.2137.1229.4124.31
+Refinement51.2251.0438.4030.6125.39
+IDL54.4354.1041.5633.6628.22
6-6 Layers50.4950.5837.5729.6424.53
+ +Table 3. Effect of our proposed components. 'Refinement' denotes applying the refinement process. 'IDL' denotes applying the iterative distillation loss. '6-6 layers' denotes enlarging the encoder and decoder of the baseline system from 3 to 6 layers. + +
ModelRMROUGEBLEU-1BLEU-2BLEU-3BLEU-4
STMCw/o50.1050.2137.1229.4124.31
w/54.4354.1041.5633.6628.22
BN-TIN-Transf.w/o47.4147.9934.9427.3322.35
w/52.0652.0639.0131.0825.69
VAC-Transf.w/o49.4850.0137.0029.1223.91
w/53.6853.6041.2833.4728.07
+ +Table 4. Generalization of IP-SLT. 'RM' denotes leveraging the refinement process. 'w' and 'w/o' denote the baseline SLT system with and without a refinement process, respectively. + +
λROUGEBLEU-1BLEU-2BLEU-3BLEU-4
051.3050.3238.0130.2925.14
553.0752.3140.0832.2927.02
1053.7153.5140.9232.9427.47
1554.4354.1041.5633.6628.22
2054.4254.1641.5133.4427.87
+ +tem. Directly applying the refinement process to the baseline delivers a performance gain of 1.08 BLEU-4. We further apply the iterative distillation loss to improve the performance. The results suggest that adding distillation supervision can be helpful with a gain of 2.83 BLEU-4. Besides, to keep the number of parameters unchanged, we enlarge the depth from 3 to 6 layers and evaluate the performance. Naively enlarging the model scale slightly improves the performance $(+0.22$ BLEU-4). The results are shown in Tab. 3. + +Generalization of the IP-SLT. We conduct three sets of experiments by changing the visual backbone to evaluate the generalization of the proposed IP-SLT approach in Tab. 4. Specifically, the BN-TIN-Transf. [49] uses a basic CNN network to get the dense representation of sign video. The STMC [50] extracts and aligns the key parts of sign language expression to achieve better performance. The VAC [31] proposes two auxiliary supervision methods to enhance the feature extraction module. VAC-Transf. replaces the feature extractor in BN-TIN-Transf. [49] with the visual backbone of VAC [31]. Applying our proposed IP-SLT methods to the BN-TIN-Transf., VAC-Transf. and STMC, we achieve 25.69, 28.07, and 28.22 BLEU-4 scores + +Table 5. The weight $\lambda$ of iterative distillation loss to CE loss. + +
βROUGEBLEU-1BLEU-2BLEU-3BLEU-4
0.050.1050.2137.1229.4124.31
0.254.3553.8141.4033.5228.02
0.454.3954.0241.2433.1927.70
0.554.4354.1041.5633.6628.22
0.653.9653.6340.9433.1027.72
0.854.3153.3940.9232.9827.53
-50.9750.6738.2430.5225.41
+ +Table 6. The weight $\beta$ of the original visual feature to the previous prototype. '0.0' denotes the assessment of the baseline. '-' denotes setting $\beta$ as 0.5 without using the drop-net [51]. + +
SettingROUGEBLEU-1BLEU-2BLEU-3BLEU-4
None50.1050.2137.1229.4124.31
Con-input54.3753.8341.1032.9727.39
Con-feature53.9752.7640.5632.8127.47
Add-feature54.4354.1041.5633.6628.22
+ +Table 7. Effect of the different refinement methods. 'Con-input' denotes directly concatenating the original visual feature and previous prototype as input. 'Con-feature' denotes concatenating the original feature and the previous feature given by the cross-attention mechanism in each layer. On top of that, 'Add-feature' denotes changing it to an addition operation. + +on the DEV set, surpassing the baselines by 3.34, 4.16, 3.91, respectively. Using a high-quality visual backbone delivers further quality gains. + +Impact of $\lambda$ . In our experiments, the weight $\lambda$ of iterative distillation loss is set to 15. It is a hyper-parameter that is designed to balance the effect of cross-entropy loss and the iterative distillation loss. We conduct experiments by varying the weight $\lambda$ . Tab. 5 shows that our IP-SLT achieves the best performance when the weight $\lambda$ is set to 15. + +Impact of $\beta$ . In the above experiments, the weight $\beta$ is fixed to 0.5. The weight $\beta$ represents the importance of the previous prototype compared with the original visual feature. As a hyper-parameter of our proposed methods, the weight $\beta$ is examined with a set of different values in Tab. 6. When the weight is 0.5, the performance is the highest one. This indicates that the previous prototype to SLT is as important as the original visual feature. Besides, to fully use the previous prototype, the drop-net is required. + +Impact of refinement method. We also examine a set of refinement methods for IP-SLT considering the fusion mechanism as a key part of our proposed method in Tab. 7. Directly concatenating the original visual feature and previous prototype in the time dimension improves the performance of SLT from 24.31 to 27.39. We further add the cross-attention mechanism to explicitly leverage the useful information from the previous prototype. And then, we concatenate the representation from the original visual feature and the representation from the previous prototype in fea + +
KROUGEBLEU-1BLEU-2BLEU-3BLEU-4
050.1050.2137.1229.4124.31
151.2251.0438.4030.6125.39
253.7353.3940.7632.9827.66
354.4354.1041.5633.6628.22
454.6353.9141.4033.5328.01
+ +Table 8. Effect of the iteration number $K$ in the iterative refinement module. + +
ModelSWI-P(M)T-P(M)FLOPs(B)ROUGEBLEU-4
STMC-92.692.628.750.1024.31
w/128.7172.632.654.4328.22
w/o238.6357.932.654.8128.48
BN-TIN-Transf.-28.328.39.647.4122.35
w/37.753.410.652.0625.69
w/o66.1110.010.652.2425.96
VAC-Transf.-28.328.39.649.4823.91
w/37.753.410.653.6828.07
w/o66.1110.010.654.8228.27
+ +Table 9. Comparing the baseline and non-shared weight prototype refinement method with IP-SLT. 'SW' denotes the sharing weight across all iterations. '-' denotes the baseline SLT system. 'w/' and 'w/o' denote the refinement module in different iterations with and without shared parameters, respectively. 'I-P' and 'T-P' denote the parameter amount calculated in inference and during training, respectively. + +ture dimension, which delivers performance gains of 3.16 BLEU-4. On top of that, changing the concatenating process to an element-wise addition operation achieves a further quality gain of 3.91. + +Impact of iteration number $K$ . The iteration number $K$ is an important hyper-parameter and is fixed to 3 in the previous experiments. We conduct experiments with different iteration numbers to explore the effect of iteration number $K$ . Tab. 8 shows that there is a culmination in the performance at iteration 3. Before this culmination, the performance improvement of our proposed method increases fast. When the iteration number is bigger than 3, the performance of our method is slightly weakened (-0.21 BLEU-4). + +Impact of sharing weights between different iterations. In Tab. 9, we examine the storage and translation quality of our proposed IP-SLT with different baselines i.e., STMC, BN-TIN-Transf, and VAC-Transf. In order to ensure the efficiency of the model, our refinement module shares weight across all iterations. Since the feature extraction module of each group is identical, the parameters of the feature extraction module are not included in the calculation of the storage of different models. As in inference, several parts can be removed from IP-SLT without changing the performance, we report the number of parameters in both training and inference, respectively. The results indicate that our iterative process causes acceptable overhead with re + +
TypeText
GTich=winsche ihnen einen tatsächien abend und machen sie es gut.
Baselinejetzt wünsche ich ihnen noch einen tatsächien abend.
Ourihnen einen tatsächien abend und machen sie es gut.
GTin der neuenWoche wird es milder aber es bleibt wechselhaft.
Baselinedann wird es wieder milder.
Ourin der neuenWoche wird es dann wieder milder.
+ +Table 10. Qualitative evaluation. 'GT' denotes the spoken language translation annotation. 'Baseline' and 'Our' denote the translation result of baseline and our IP-SLT, respectively. + +markable performance improvements. We further conduct experiments in which the parameters of each iteration are independent. The results demonstrate that the proposed parameter-shared method achieves near performance gain (28.22 v.s 28.48 BLEU-4, 25.69 v.s 25.96 BLEU-4, and 28.07 v.s 28.27 BLEU-4) with non-shared one but leveraging much less parameter number. + +Computation comparison with baseline. The FLOPs is a key factor of computation efficiency. We conduct experiments to compare the computation of the proposed IP-SLT method with different baselines. Similarly, we exclude the FLOPs of the feature extraction module and report relevant computational costs in inference. Tab. 9 shows that leveraging the iterative refinement process can cause acceptable computation costs while achieving promising performance. + +Case Study. To provide a more intuitionistic view of our proposed method, we list some translation samples of the proposed IP-SLT method in Tab. 10. We observe that based on the previous prototype and original visual representation, the IP-SLT can generate more accurate and fluent sentences. + +# 5. Conclusion + +In this work, we propose a new framework IP-SLT which introduces the iterative refinement into a conventional SLT system. With the goal to polish the semantic representation by leveraging the previous results, we present IP-SLT to support the iterative refinement process. The proposed method is differentiable and optimized in an end-to-end manner to achieve its best performance. On top of it, we put forward the iterative distillation loss to further improve the translation quality by leveraging the sequential dependence between the outputs of each iteration. In inference, the autoregressive decoding process is required once to generate the translation based on the final prototype, applying IP-SLT does not significantly affect efficiency. The experimental results demonstrate the effectiveness of the IP-SLT. + +Acknowledgments: This work was supported by NSFC under Contract U20A20183 and 62021001. It was also supported by the GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, and the Supercomputing Center of the USTC. + +# References + +[1] Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Image super-resolution via progressive cascading residual network. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2018. 2 +[2] Samuel Albanie, Gül Varol, Liliane Momeni, Triantafyllos Afouras, Joon Son Chung, Neil Fox, and Andrew Zisserman. BSL-1K: Scaling up co-articulated sign language recognition using mouthing cues. In Computer Vision European Conference, 2020. 1 +[3] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. 2, 3 +[4] Danielle Bragg, Oscar Koller, Mary Bellard, Larwan Berke, Patrick Boudreault, Annelies Braffort, Naomi Caselli, Matt Huenerfauth, Hernisa Kacorri, Tessa Verhoef, et al. Sign language recognition, generation, and translation: An interdisciplinary perspective. In International ACM SIGACCESS Conference on Computers and Accessibility, 2019. 1 +[5] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In IEEE Conference on Computer Vision and Pattern Recognition, 2018. 1, 2, 5, 6 +[6] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Multi-channel transformers for multi-articulatory sign language translation. In Computer Vision-ECCV 2020 Workshops, 2020. 2 +[7] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 2, 6 +[8] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2 +[9] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2, 6 +[10] Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie Liu, and Brian Mak. Two-stream network for sign language recognition and translation. Advances in Neural Information Processing Systems, 2022. 2, 6 +[11] Helen Cooper, Brian Holt, and Richard Bowden. Sign language recognition. Visual Analysis of Humans: Looking at People, 2011. 1 +[12] Kearsy Cormier, Neil Fox, Bencie Woll, Andrew Zisserman, Necati Cihan Camgoz, and Richard Bowden. ExTOL: Automatic recognition of British sign language using the bsl corpus. In Workshop on Sign Language Translation and Avatar Technology, 2019. 1 +[13] Runpeng Cui, Hu Liu, and Changshui Zhang. Recurrent convolutional neural networks for continuous sign language recognition by staged optimization. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1, 3 + +[14] Runpeng Cui, Hu Liu, and Changshui Zhang. A deep neural framework for continuous sign language recognition by iterative training. IEEE Transactions on Multimedia, 2019. 2, 3 +[15] Hao Feng, Wengang Zhou, Jiajun Deng, Qi Tian, and Houqiang Li. DocScanner: robust document image rectification with progressive learning. arXiv preprint arXiv:2110.14968, 2021. 2 +[16] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. Convolutional sequence to sequence learning. In International Conference on Machine Learning, 2017. 3 +[17] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In International Conference on Machine learning, 2006. 2 +[18] Karol Gregor, Ivo Danihelka, Alex Graves, Danilo Rezende, and Daan Wierstra. DRAW: A recurrent neural network for image generation. In International Conference on Machine Learning, 2015. 2 +[19] Jun He, Richang Hong, Xueliang Liu, Mingliang Xu, and Qianru Sun. Revisiting local descriptor for improved few-shot classification. ACM Transactions on Multimedia Computing, Communications, and Applications, 2022. 1 +[20] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. Signbert+: Hand-model-aware self-supervised pretraining for sign language understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1 +[21] Jie Huang, Wengang Zhou, Qilin Zhang, Houqiang Li, and Weiping Li. Video-based sign language recognition without temporal segmentation. In AAAI Conference on Artificial Intelligence, 2018. 1 +[22] Shuiwang Ji, Wei Xu, Ming Yang, and Kai Yu. 3D convolutional neural networks for human action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2012. 2 +[23] Tao Jin, Zhou Zhao, Meng Zhang, and Xingshan Zeng. Prior knowledge and memory enriched transformer for sign language translation. In Findings of the Association for Computational Linguistics, 2022. 1, 2, 6 +[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[25] Oscar Koller, Necati Cihan Camgoz, Hermann Ney, and Richard Bowden. Weakly supervised learning with multistream cnn-lstm-hmms to discover sequential parallelism in sign language videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 3 +[26] Oscar Koller, Hermann Ney, and Richard Bowden. Deep hand: How to train a CNN on 1 million hand images when your data is continuous and weakly labelled. In IEEE Conference on Computer Vision and Pattern Recognition, 2016. 3 +[27] Dongxu Li, Chenchen Xu, Xin Yu, Kaihao Zhang, Benjamin Swift, Hanna Suominen, and Hongdong Li. TSPNet: Hierarchical feature learning via temporal semantic pyramid for + +sign language translation. Advances in Neural Information Processing Systems, 2020. 2 +[28] Huan Ling, Jun Gao, Amlan Kar, Wenzheng Chen, and Sanja Fidler. Fast interactive object annotation with Curve-GCN. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2 +[29] Zichen Liu, Jun Hao Liew, Xiangyu Chen, and Jiashi Feng. DANCE: A deep attentive contour model for efficient instance segmentation. In IEEE Winter Conference on Applications of Computer Vision, 2021. 2 +[30] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. arXiv preprint arXiv:1508.04025, 2015. 1 +[31] Yuecong Min, Aiming Hao, Xiujuan Chai, and Xilin Chen. Visual alignment constraint for continuous sign language recognition. In IEEE International Conference on Computer Vision, 2021. 1, 2, 6, 7 +[32] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: a method for automatic evaluation of machine translation. In Annual Meeting of the Association for Computational Linguistics, 2002. 5 +[33] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 5 +[34] Sida Peng, Wen Jiang, Huaijin Pi, Xiuli Li, Hujun Bao, and Xiaowei Zhou. Deep snake for real-time instance segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2 +[35] Junfu Pu, Wengang Zhou, and Houqiang Li. Dilated convolutional network with iterative optimization for continuous sign language recognition. In International Joint Conference on Artificial Intelligence, 2018. 2 +[36] Junfu Pu, Wengang Zhou, and Houqiang Li. Iterative alignment network for continuous sign language recognition. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2 +[37] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive Image Deraining Networks: A better and simpler baseline. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2 +[38] Lin CY ROUGE. A package for automatic evaluation of summaries. In Workshop on Text Summarization of ACL, 2004. 5 +[39] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. Advances in Neural Information Processing Systems, 2014. 2 +[40] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. Advances in Neural Information Processing Systems, 2014. 1, 3 +[41] Rachel Sutton-Spence and Bencie Woll. The linguistics of British Sign Language: an introduction. Cambridge University Press, 1999. 1 +[42] Shengeng Tang, Dan Guo, Richang Hong, and Meng Wang. Graph-based multimodal sequential embedding for sign language translation. IEEE Transactions on Multimedia, 2021. 1 + +[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1, 2, 3, 4, 5 +[44] Hanjie Wang, Xiujuan Chai, Xiaopeng Hong, Guoying Zhao, and Xilin Chen. Isolated sign language recognition with grassmann covariance matrices. ACM Transactions on Accessible Computing, 2016. 1 +[45] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google's neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144, 2016. 6 +[46] Kayo Yin and Jesse Read. Better sign language translation with stmc-transformer. arXiv preprint arXiv:2004.00588, 2020.2 +[47] Tao Zhang, Shiqing Wei, and Shunping Ji. E2EC: An end-to-end contour-based method for high-quality high-speed instance segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2 +[48] Hao Zhou, Wengang Zhou, and Houqiang Li. Dynamic pseudo label decoding for continuous sign language recognition. In IEEE International conference on Multimedia and Expo, 2019. 2 +[49] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 5, 6, 7 +[50] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for sign language recognition and translation. IEEE Transactions on Multimedia, 2021. 2, 6, 7 +[51] Jinhua Zhu, Yingce Xia, Lijun Wu, Di He, Tao Qin, Wengang Zhou, Houqiang Li, and Tie-Yan Liu. Incorporating bert into neural machine translation. arXiv preprint arXiv:2002.06823, 2020. 4, 6, 7 \ No newline at end of file diff --git a/2023/Sign Language Translation with Iterative Prototype/images.zip b/2023/Sign Language Translation with Iterative Prototype/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..62d9516c07606cd01e90611e72e758320ce37153 --- /dev/null +++ b/2023/Sign Language Translation with Iterative Prototype/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2dfb06c7367cf7385f426db5b603ae3b61cc23070baeb71382e56d7078fc2e44 +size 496915 diff --git a/2023/Sign Language Translation with Iterative Prototype/layout.json b/2023/Sign Language Translation with Iterative Prototype/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..228c5c84225b54e1ec41cba03e2f8d19e4f8a927 --- /dev/null +++ b/2023/Sign Language Translation with Iterative Prototype/layout.json @@ -0,0 +1,10881 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 103, + 457, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 103, + 457, + 121 + ], + "spans": [ + { + "bbox": [ + 136, + 103, + 457, + 121 + ], + "type": "text", + "content": "Sign Language Translation with Iterative Prototype" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 141, + 530, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 141, + 530, + 157 + ], + "spans": [ + { + "bbox": [ + 64, + 141, + 530, + 157 + ], + "type": "text", + "content": "Huijie Yao" + }, + { + "bbox": [ + 64, + 141, + 530, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 141, + 530, + 157 + ], + "type": "text", + "content": " Wengang Zhou" + }, + { + "bbox": [ + 64, + 141, + 530, + 157 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 64, + 141, + 530, + 157 + ], + "type": "text", + "content": " Hao Feng" + }, + { + "bbox": [ + 64, + 141, + 530, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 141, + 530, + 157 + ], + "type": "text", + "content": " Hezhen Hu" + }, + { + "bbox": [ + 64, + 141, + 530, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 141, + 530, + 157 + ], + "type": "text", + "content": " Hao Zhou" + }, + { + "bbox": [ + 64, + 141, + 530, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 141, + 530, + 157 + ], + "type": "text", + "content": " Houqiang Li" + }, + { + "bbox": [ + 64, + 141, + 530, + 157 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 158, + 529, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 158, + 529, + 185 + ], + "spans": [ + { + "bbox": [ + 64, + 158, + 529, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 158, + 529, + 185 + ], + "type": "text", + "content": " CAS Key Laboratory of Technology in GIPAS, EEIS Department, University of Science and Technology of China " + }, + { + "bbox": [ + 64, + 158, + 529, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 158, + 529, + 185 + ], + "type": "text", + "content": " Institute of Artificial Intelligence, Hefei Comprehensive National Science Center" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 186, + 496, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 186, + 496, + 200 + ], + "spans": [ + { + "bbox": [ + 97, + 186, + 496, + 200 + ], + "type": "text", + "content": "{yaohuijie,haof,alexhu,zhouh156}@mail.ustc.edu.cn，{zhwg,lihq}@ustc.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 252, + 290, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 540 + ], + "type": "text", + "content": "This paper presents IP-SLT, a simple yet effective framework for sign language translation (SLT). Our IP-SLT adopts a recurrent structure and enhances the semantic representation (prototype) of the input sign language video via an iterative refinement manner. Our idea mimics the behavior of human reading, where a sentence can be digested repeatedly, till reaching accurate understanding. Technically, IP-SLT consists of feature extraction, prototype initialization, and iterative prototype refinement. The initialization module generates the initial prototype based on the visual feature extracted by the feature extraction module. Then, the iterative refinement module leverages the cross-attention mechanism to polish the previous prototype by aggregating it with the original video feature. Through repeated refinement, the prototype finally converges to a more stable and accurate state, leading to a fluent and appropriate translation. In addition, to leverage the sequential dependence of prototypes, we further propose an iterative distillation loss to compress the knowledge of the final iteration into previous ones. As the autoregressive decoding process is executed only once in inference, our IP-SLT is ready to improve various SLT systems with acceptable overhead. Extensive experiments are conducted on public benchmarks to demonstrate the effectiveness of the IP-SLT." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 564, + 128, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 128, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 128, + 577 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 585, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 585, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 585, + 287, + 693 + ], + "type": "text", + "content": "Sign language translation (SLT) aims to automatically generate spoken language translations based on sign language videos, which holds both social significance and academic value. On the one hand, a high-quality SLT system can greatly facilitate communication between deafmute and hearing individuals [4, 11, 12, 41]. On the other hand, SLT as an interdisciplinary research topic necessitates a comprehensive understanding of computer vision [31, 13, 36, 19] and natural language processing [43, 40], given its" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 228, + 537, + 256 + ], + "blocks": [ + { + "bbox": [ + 316, + 228, + 537, + 256 + ], + "lines": [ + { + "bbox": [ + 316, + 228, + 537, + 256 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 537, + 256 + ], + "type": "image", + "image_path": "d3b242dc929ef28a6c798e2106b9377eacab9351036ce2d8ede668759cf7ef71.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 257, + 470, + 266 + ], + "lines": [ + { + "bbox": [ + 389, + 257, + 470, + 266 + ], + "spans": [ + { + "bbox": [ + 389, + 257, + 470, + 266 + ], + "type": "text", + "content": "(a) The Previous Pipeline" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 277, + 537, + 342 + ], + "blocks": [ + { + "bbox": [ + 316, + 277, + 537, + 342 + ], + "lines": [ + { + "bbox": [ + 316, + 277, + 537, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 537, + 342 + ], + "type": "image", + "image_path": "fb659078773baacc81bce84eef610a8c414e47deea8c8d3b6e8d2a152c3510af.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 344, + 470, + 354 + ], + "lines": [ + { + "bbox": [ + 392, + 344, + 470, + 354 + ], + "spans": [ + { + "bbox": [ + 392, + 344, + 470, + 354 + ], + "type": "text", + "content": "(b) Our IP-SLT Pipeline" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 359, + 547, + 449 + ], + "lines": [ + { + "bbox": [ + 304, + 359, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 359, + 547, + 449 + ], + "type": "text", + "content": "Figure 1. Illustration of the pipeline of the previous works and our IP-SLT. (a) The previous studies rely on a one-pass forward process to generate the final translation. (b) To mitigate the vision-text gap, we introduce the iterative refinement module into the original SLT system. The refinement module updates the current prototype conditioned on the sign language video, which can be run iteratively to obtain a better representation of the semantic meaning of the sign language video." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 471, + 546, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 546, + 506 + ], + "type": "text", + "content": "involvement with vision and text modalities. As a result, SLT has emerged as a vital research topic, garnering increasing attention [5, 49, 31, 2, 21, 20, 42]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 508, + 547, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 547, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 547, + 664 + ], + "type": "text", + "content": "SLT is a challenging task, which faces a tough domain gap between the input video and output text, as well as a limited dataset scale due to costly data collection and annotation [49, 7, 44, 21]. Since SLT is typically viewed as a sequence-to-sequence mapping problem, the existing SLT systems [23, 5, 7] commonly adopt the one-pass forward pipeline based on encoder-decoder architecture [40, 30] (as shown in Fig. 1 (a)). In such a framework, the encoder transforms the sign video into its semantic representation (prototype), which is then fed into the decoder to obtain the final translation. However, due to the inherent gap between vision and text, it may be hard to conduct such mapping within the vanilla one-pass architecture." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "In this study, we present IP-SLT with the iterative prototype to boost sign language translation (as shown in Fig. 1 (b)), which is inspired by the human reading process. During this process, we note that repeatedly digging into the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 248, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 248, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 248, + 714 + ], + "type": "text", + "content": "*Corresponding authors: Wengang Zhou and Houqiang Li" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15592" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 323 + ], + "type": "text", + "content": "source materials is necessary for accurate understanding. Similarly, when we are trying to translate a sign language video into a sentence, we commonly do not directly write it down. Instead, we would recall and go back to the original sign video to check our answers. To implement the above idea, our IP-SLT adopts a recurrent structure that enhances the semantic representation (prototype) of the input sign language video via an iterative refinement process. IP-SLT generally contains three main components, including feature extraction, prototype initialization, and iterative prototype refinement. Given a sign video to be translated, we first extract its visual representation, which is then used for generating an initial raw prototype. Subsequently, we iteratively leverage the attention mechanism [43] to update the prototype toward the semantic meaning of the sign video. At each iteration, we refine the previous prototype by aggregating it with the original visual representation. In this way, the network repeatedly digs the semantic context from the sign video to polish the prototype. Through iterative refinement, the prototype finally converges to a stable and accurate state, producing a high-quality translation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 323, + 289, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 289, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 289, + 502 + ], + "type": "text", + "content": "In addition, our IP-SLT introduces a novel design discussed next. Firstly, to leverage the sequential dependence between different iterations, we further propose the iterative distillation loss which allows the previous prototypes to obtain supervision from the final one. Since the final prototype converges to a more stable and accurate state, it is possible for IP-SLT to achieve better performance. Secondly, during training, all predicted prototypes are transformed into their corresponding translations to provide guidance for each iteration. Our inference process is neat since only the final prototype is used for the autoregressive decoding process. Thirdly, our IP-SLT can easily work with different visual backbones. Through end-to-end optimization, our IP-SLT achieves significant performance improvements over the baselines." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 502, + 242, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 502, + 242, + 514 + ], + "spans": [ + { + "bbox": [ + 59, + 502, + 242, + 514 + ], + "type": "text", + "content": "In summary, our contributions are three-fold:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 519, + 287, + 662 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 57, + 519, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 519, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 519, + 287, + 567 + ], + "type": "text", + "content": "- We propose IP-SLT, a novel framework to ameliorate sign language translation, which iteratively refines the prototypes by aggregating the previous translation progress and the original visual representation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 573, + 287, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 573, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 57, + 573, + 287, + 608 + ], + "type": "text", + "content": "- We propose an iterative distillation loss to enhance the basic supervision, by leveraging the sequential dependence between the outputs at each iteration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 615, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 615, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 57, + 615, + 287, + 662 + ], + "type": "text", + "content": "- We conduct extensive experiments to validate the proposed method, and show encouraging improved results on the two prevalent benchmarks, i.e., CSL-Daily [49] and PHOENIX-2014T [5]." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 670, + 134, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 134, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 134, + 681 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "In this section, we briefly review the related works, i.e., sign language translation and iterative refinement methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 347 + ], + "type": "text", + "content": "Sign Language Translation. Camgoz et al. [5] pioneer the neural SLT task and publish the neural dataset PHOENIX-2014T and regard the SLT as a sequence-to-sequence problem. They implement the neural SLT system using the encoder-decoder paradigm [3]. This paradigm is adopted by subsequent studies which focus on addressing the challenges of data scarcity and domain gap. Considering the lack of frame-level annotation in sign language datasets, Li et al. [27] design the temporal semantic pyramid structure to obtain more discriminative features. Camgoz et al. [7] explore the mutual benefits of SLT and continuous sign language recognition through joint optimization. Zhou et al. [49] leverage gloss annotation to transform the monolingual texts into pseudo-videos. According to the characteristic of sign language, several works [46, 6, 50] propose multi-channel SLT systems which explicitly extract and align the key parts of sign language expression. Jin et al. [23] leverage the additional prior knowledge to obtain high-quality translations. Chen et al. [9] propose a transfer learning baseline for SLT by leveraging external resources from related tasks. Chen et al. [10] further combine the raw videos and the keypoint sequences to achieve better semantic understanding with auxiliary supervision." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 348, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 396 + ], + "type": "text", + "content": "In contrast, our proposed approach employs an iterative refinement process that utilizes the previous prototype as an additional clue to enhance the accuracy of the mapping between sign videos and their translations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 401, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 628 + ], + "type": "text", + "content": "Iterative Refinement Methods. The idea of iterative refinement is applied to various computer vision tasks, such as image generation [1, 18, 37], instance segmentation [34, 28, 29, 47], image rectification [15], etc., which shows promising performance improvements. CARN [1] maintains training stability in super-resolution tasks and improves the quality of output images. Ling et al. [28] regard the object instance segmentation task as a regression task and proposes the Curve-GCN to iteratively predict the locations of all vertices. DeepSanke [34] uses the deep network to iteratively enclose the object boundary based on an initial contour. The previous studies [48, 35, 14, 36, 31] in continuous sign language recognition task follow an iterative training scheme to enhance the discriminative power of feature extraction modules which use the convolutional neural network and their variants [8, 22, 39]. They leverage alignment proposals given by the connectionist temporal classification (CTC) [17] decoding as supervision at frame-wise granularity, which cannot be directly applied to the SLT." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "content": "Different from the aforementioned methods, we explore how to reduce the vision-text gap by proposing an iterative refinement module to the existing SLT system. To reduce the complexity, we design the refinement module in a shared-weight manner. Moreover, we put forward the iterative distillation loss to leverage the sequential dependence between different iterations." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15593" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 70, + 509, + 203 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 509, + 203 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 509, + 203 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 509, + 203 + ], + "type": "image", + "image_path": "848452f5099d44eb4ba752e0c5e7717723377bd198fca8e83909ab95951806e3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "text", + "content": "Figure 2. An overview of the proposed IP-SLT framework. Given a sign video " + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "text", + "content": ", the feature extraction module is responsible for embedding the input into visual representation " + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "text", + "content": ". The initialization module (the encoder " + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_1" + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "text", + "content": ") generates the initial prototype " + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^0" + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "text", + "content": " and the raw translation " + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^0" + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "text", + "content": ". The refinement module (the encoder " + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_2" + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "text", + "content": ") first takes the initial prototype " + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^0" + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "text", + "content": " as input and generates a prototype for the current step by fusing it with the original visual representation " + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "text", + "content": ". Through " + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "text", + "content": " times refinement, the prototype sequence " + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{E} = \\{\\mathbf{E}^0, \\mathbf{E}^1, \\dots, \\mathbf{E}^K\\}" + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "text", + "content": " and corresponding translation sequence " + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} = \\{\\mathbf{Y}^0, \\mathbf{Y}^1, \\dots, \\mathbf{Y}^K\\}" + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "text", + "content": " are obtained. In light of the fact that the decoding part of IP-SLT consists of " + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "inline_equation", + "content": "K + 1" + }, + { + "bbox": [ + 46, + 207, + 547, + 288 + ], + "type": "text", + "content": " branches based on the iteration order, we introduce the iterative distillation loss to improve the underlying supervision. It should be noted that the parts enclosed in dashed boxes can be removed in inference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 303, + 130, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 303, + 130, + 317 + ], + "spans": [ + { + "bbox": [ + 47, + 303, + 130, + 317 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 324, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 288, + 373 + ], + "type": "text", + "content": "In this section, we first introduce the overall architecture of our IP-SLT, and then separately elaborate individual components. Finally, we propose the design of the training objective and inference strategy for the IP-SLT." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 380, + 173, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 173, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 173, + 393 + ], + "type": "text", + "content": "3.1. Framework Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "type": "text", + "content": "The primary objective of the SLT system is to acquire knowledge about the mapping " + }, + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "type": "inline_equation", + "content": "f: \\mathcal{X} \\mapsto \\mathcal{Y}" + }, + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "type": "text", + "content": " denote the collections of " + }, + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "type": "text", + "content": " sign language videos and spoken language sentences associated with vocabulary " + }, + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "type": "text", + "content": ", respectively. Most SLT systems adopt the encoder-decoder architecture [40], where the input " + }, + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "type": "inline_equation", + "content": "X \\in \\mathcal{X}" + }, + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "type": "text", + "content": " is first encoded to derive a high-level context representation. It is then passed to the decoder to generate the output " + }, + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "type": "inline_equation", + "content": "Y \\in \\mathcal{Y}" + }, + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "type": "text", + "content": ". The encoder and decoder can be specialized using different types of neural networks, such as GRU [3], CNN [16], and Transformer [43]. Considering the performance of existing SLT systems, we adopt the Transformer as well." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 544, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 288, + 713 + ], + "type": "text", + "content": "With the goal of narrowing the domain gap between vision and text, we augment the original translation process with an iterative refinement step. Fig. 2 provides an overview of the proposed IP-SLT model, which consists of three stages, namely feature extraction, prototype initialization, and iterative prototype refinement. As with previous approaches, the initialization and iterative refinement module adopt the encoder-decoder architecture. Given the sign language video " + }, + { + "bbox": [ + 46, + 544, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = \\{\\pmb{x}_t\\}_{t=1}^{T_x}" + }, + { + "bbox": [ + 46, + 544, + 288, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 544, + 288, + 713 + ], + "type": "inline_equation", + "content": "T_x" + }, + { + "bbox": [ + 46, + 544, + 288, + 713 + ], + "type": "text", + "content": " frames, the feature extraction module embeds it into the spatial-temporal feature " + }, + { + "bbox": [ + 46, + 544, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F} = \\{\\pmb{f}_t\\}_{t=1}^{T_f}" + }, + { + "bbox": [ + 46, + 544, + 288, + 713 + ], + "type": "text", + "content": ". Next, the encoder " + }, + { + "bbox": [ + 46, + 544, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_1" + }, + { + "bbox": [ + 46, + 544, + 288, + 713 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 46, + 544, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 46, + 544, + 288, + 713 + ], + "type": "text", + "content": " are employed in the initialization module to derive the initial prototype " + }, + { + "bbox": [ + 46, + 544, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^0" + }, + { + "bbox": [ + 46, + 544, + 288, + 713 + ], + "type": "text", + "content": " and initial translation " + }, + { + "bbox": [ + 46, + 544, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^0" + }, + { + "bbox": [ + 46, + 544, + 288, + 713 + ], + "type": "text", + "content": " from the visual feature " + }, + { + "bbox": [ + 46, + 544, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 544, + 288, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "text", + "content": "Subsequently, the refinement module iteratively refines the previous prototype and generates the final translation " + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^{K} = \\{\\mathbf{y}_{t}^{K}\\}_{t=1}^{T_{y,K}}" + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "inline_equation", + "content": "T_{y,K}" + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "text", + "content": " words after total " + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "text", + "content": " iterations. At the " + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "text", + "content": "-th iteration, the encoder " + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_2" + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "text", + "content": " estimates the prototype " + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^k" + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "text", + "content": " for the current step by augmenting the original visual feature " + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "text", + "content": " with the previous prototype " + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^{k-1}" + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "text", + "content": ". Next, the decoder " + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "text", + "content": " predicts the corresponding translation " + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^k" + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "text", + "content": ". Finally, the translation sequence " + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} = \\{\\mathbf{Y}^0, \\mathbf{Y}^1, \\dots, \\mathbf{Y}^K\\}" + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "text", + "content": " is obtained according to the prototype sequence " + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{E} = \\{\\mathbf{E}^0, \\mathbf{E}^1, \\dots, \\mathbf{E}^K\\}" + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "text", + "content": ". For the IP-SLT optimization, by dividing the decoding part into " + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "inline_equation", + "content": "K+1" + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "text", + "content": " branches, we add iterative distillation supervision from the final translation to the middle translations. Since the refinement process takes place in the encoder " + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_2" + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "text", + "content": ", in inference, the proposed IP-SLT can generate the translation directly from the " + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 304, + 547, + 498 + ], + "type": "text", + "content": "-th prototype which causes acceptable overhead." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 504, + 419, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 504, + 419, + 516 + ], + "spans": [ + { + "bbox": [ + 306, + 504, + 419, + 516 + ], + "type": "text", + "content": "3.2. Feature Extraction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "type": "text", + "content": "The feature extraction module embeds a series of video frames " + }, + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{T_x \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "type": "text", + "content": " with width " + }, + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "type": "text", + "content": " and height " + }, + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "type": "text", + "content": " into its visual feature " + }, + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{T_f \\times C}" + }, + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "type": "text", + "content": " with the dimension of feature " + }, + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "type": "text", + "content": ". Since its goal is to extract a distinguishable representation for SLT, we can draw on the visual backbone used in CSLR [14, 26, 25, 13] to extract the valid representation. Generally, with sliding window size " + }, + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "type": "text", + "content": " and stride size " + }, + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "type": "text", + "content": ", the sign video is split into " + }, + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "type": "inline_equation", + "content": "T_f = \\left\\lceil \\frac{T_x}{s} \\right\\rceil" + }, + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "type": "text", + "content": " clips. By passing sign videos through it, the spatial-temporal embeddings " + }, + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{F} = \\{\\mathbf{f}_t\\}_{t=1}^{T_f}" + }, + { + "bbox": [ + 304, + 522, + 547, + 644 + ], + "type": "text", + "content": " are extracted as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 357, + 650, + 545, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 650, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 357, + 650, + 545, + 666 + ], + "type": "interline_equation", + "content": "\\left\\{\\boldsymbol {f} _ {t} \\right\\} _ {t = 1} ^ {T _ {f}} = \\operatorname {E x t r a c t o r} \\left(\\left\\{\\boldsymbol {x} _ {t} \\right\\} _ {t = 1} ^ {T _ {x}}\\right). \\tag {1}", + "image_path": "60549d8491de2d8124391709624e47e7017b82cb8654397ab6e582f4fac24312.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 671, + 440, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 440, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 440, + 684 + ], + "type": "text", + "content": "3.3. Prototype Initialization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "content": "After a visual feature " + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{F} \\in \\mathbb{R}^{T_f \\times C}" + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "content": " is extracted by the feature extraction module, it is first fed into the initialization" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15594" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 79, + 72, + 254, + 171 + ], + "blocks": [ + { + "bbox": [ + 79, + 72, + 254, + 171 + ], + "lines": [ + { + "bbox": [ + 79, + 72, + 254, + 171 + ], + "spans": [ + { + "bbox": [ + 79, + 72, + 254, + 171 + ], + "type": "image", + "image_path": "381e7b884c3be4fac6d50fe0e688936241a4bfabfe3766ed68dd30a5c06a5e4d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "lines": [ + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "type": "text", + "content": "Figure 3. An illustration of the prototype update process in the " + }, + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "type": "text", + "content": "-th layer of the encoder " + }, + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_2" + }, + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "type": "text", + "content": "-th iteration." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 217, + 286, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 286, + 240 + ], + "type": "text", + "content": "module. The initialization module consists of an encoder " + }, + { + "bbox": [ + 47, + 217, + 286, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_1" + }, + { + "bbox": [ + 47, + 217, + 286, + 240 + ], + "type": "text", + "content": " and a decoder " + }, + { + "bbox": [ + 47, + 217, + 286, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 47, + 217, + 286, + 240 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 241, + 286, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 241, + 286, + 329 + ], + "spans": [ + { + "bbox": [ + 47, + 241, + 286, + 329 + ], + "type": "text", + "content": "The visual representation " + }, + { + "bbox": [ + 47, + 241, + 286, + 329 + ], + "type": "inline_equation", + "content": "\\pmb{F}" + }, + { + "bbox": [ + 47, + 241, + 286, + 329 + ], + "type": "text", + "content": " is first fed into the encoder " + }, + { + "bbox": [ + 47, + 241, + 286, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_1" + }, + { + "bbox": [ + 47, + 241, + 286, + 329 + ], + "type": "text", + "content": ", and encoded into " + }, + { + "bbox": [ + 47, + 241, + 286, + 329 + ], + "type": "inline_equation", + "content": "T_{f}" + }, + { + "bbox": [ + 47, + 241, + 286, + 329 + ], + "type": "text", + "content": " raw states " + }, + { + "bbox": [ + 47, + 241, + 286, + 329 + ], + "type": "inline_equation", + "content": "\\pmb{E}^{0} = \\{e_{t}^{0}\\}_{t=1}^{T_{f}} \\in \\mathbb{R}^{T_{f} \\times C}" + }, + { + "bbox": [ + 47, + 241, + 286, + 329 + ], + "type": "text", + "content": ". Then the decoder " + }, + { + "bbox": [ + 47, + 241, + 286, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 47, + 241, + 286, + 329 + ], + "type": "text", + "content": " reads the prototype " + }, + { + "bbox": [ + 47, + 241, + 286, + 329 + ], + "type": "inline_equation", + "content": "\\pmb{E}^{0}" + }, + { + "bbox": [ + 47, + 241, + 286, + 329 + ], + "type": "text", + "content": " and produces the initial translation " + }, + { + "bbox": [ + 47, + 241, + 286, + 329 + ], + "type": "inline_equation", + "content": "\\pmb{Y}^{0} = \\{\\pmb{y}_{t}^{0}\\}_{t=1}^{T_{y,0}}" + }, + { + "bbox": [ + 47, + 241, + 286, + 329 + ], + "type": "text", + "content": " according to the predicted logits " + }, + { + "bbox": [ + 47, + 241, + 286, + 329 + ], + "type": "inline_equation", + "content": "\\pmb{U}^{0} = \\{\\pmb{u}_{t}^{0}\\}_{t=1}^{T_{y,0}}" + }, + { + "bbox": [ + 47, + 241, + 286, + 329 + ], + "type": "text", + "content": ". Specifically, it predicts the conditional probability of the translation sequence, which is formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 86, + 334, + 287, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 334, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 86, + 334, + 287, + 367 + ], + "type": "interline_equation", + "content": "p ^ {0 - t h} \\left(\\boldsymbol {Y} ^ {0} | \\boldsymbol {X}\\right) = \\prod_ {t = 1} ^ {T _ {y, 0}} p \\left(\\boldsymbol {y} _ {t} ^ {0} \\mid \\boldsymbol {F}, \\boldsymbol {y} _ {0: t - 1} ^ {0}\\right), \\tag {2}", + "image_path": "eb71182d3940c64018f30456c108a447d9c9d1c6f6666c001817d4eee81d5e1c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 376, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 287, + 425 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 376, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\pmb{y}_{0:t-1}^{0} = \\{\\pmb{y}_{0}^{0}, \\pmb{y}_{1}^{0}, \\dots, \\pmb{y}_{t-1}^{0}\\}" + }, + { + "bbox": [ + 47, + 376, + 287, + 425 + ], + "type": "text", + "content": " denotes the previous output sub-sequence at the " + }, + { + "bbox": [ + 47, + 376, + 287, + 425 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 376, + 287, + 425 + ], + "type": "text", + "content": "-th step. The initial token " + }, + { + "bbox": [ + 47, + 376, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\pmb{y}_{0}^{0}" + }, + { + "bbox": [ + 47, + 376, + 287, + 425 + ], + "type": "text", + "content": " represents the beginning of a sentence. The predicted probability of each token in the translation is computed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 430, + 287, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 430, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 82, + 430, + 287, + 462 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p \\left(\\boldsymbol {y} _ {t} ^ {0} \\mid \\boldsymbol {F}, \\boldsymbol {y} _ {1: t - 1} ^ {0}\\right) = \\operatorname {s o f t m a x} \\left(\\boldsymbol {u} _ {t} ^ {0}\\right) _ {\\boldsymbol {y} _ {t} ^ {0}} \\tag {3} \\\\ = \\operatorname {s o f t m a x} \\left(\\boldsymbol {h} _ {t} ^ {0} \\cdot \\boldsymbol {W}\\right) _ {\\boldsymbol {y} _ {t} ^ {0}}, \\\\ \\end{array}", + "image_path": "547988a7948a7b501005f60e10813ff143d832b4d705da5f00177096c2171c33.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 471, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 471, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 471, + 287, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 471, + 287, + 567 + ], + "type": "inline_equation", + "content": "h_t^0 \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 47, + 471, + 287, + 567 + ], + "type": "text", + "content": " represents the output of the final layer at the " + }, + { + "bbox": [ + 47, + 471, + 287, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 471, + 287, + 567 + ], + "type": "text", + "content": "-th step, and " + }, + { + "bbox": [ + 47, + 471, + 287, + 567 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{C \\times |\\mathcal{V}|}" + }, + { + "bbox": [ + 47, + 471, + 287, + 567 + ], + "type": "text", + "content": " denotes a linear mapping to projects the hidden state " + }, + { + "bbox": [ + 47, + 471, + 287, + 567 + ], + "type": "inline_equation", + "content": "h_t^0" + }, + { + "bbox": [ + 47, + 471, + 287, + 567 + ], + "type": "text", + "content": " into the predicted logits over the target vocabulary " + }, + { + "bbox": [ + 47, + 471, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 47, + 471, + 287, + 567 + ], + "type": "text", + "content": ". The probability is calculated by applying the softmax(·) function to the logits. Notably, our goal is to obtain a more accurate prototype for SLT, thus, the decoder " + }, + { + "bbox": [ + 47, + 471, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 47, + 471, + 287, + 567 + ], + "type": "text", + "content": " is only used in the training process to provide guidance for the initialization module." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 575, + 216, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 216, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 216, + 588 + ], + "type": "text", + "content": "3.4. Iterative Prototype Refinement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 594, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 287, + 712 + ], + "type": "text", + "content": "Once the initial prototype " + }, + { + "bbox": [ + 47, + 594, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\pmb{E}^{0}" + }, + { + "bbox": [ + 47, + 594, + 287, + 712 + ], + "type": "text", + "content": " is obtained through the initialization module, we feed it together with the original visual representation " + }, + { + "bbox": [ + 47, + 594, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\pmb{F}" + }, + { + "bbox": [ + 47, + 594, + 287, + 712 + ], + "type": "text", + "content": " into the iterative refinement module. The module maintains a single prototype which is iteratively refined. In this way, the coarse semantic feature finally converges to a stable state where the prototype best fits the sign language semantics. The refinement module is divided into two sub-processes, i.e., iterative prototype aggregation in the encoder " + }, + { + "bbox": [ + 47, + 594, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_2" + }, + { + "bbox": [ + 47, + 594, + 287, + 712 + ], + "type": "text", + "content": " and translation generation in the decoder " + }, + { + "bbox": [ + 47, + 594, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 47, + 594, + 287, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "text", + "content": "Iterative prototype aggregation. To utilize the prototype as a reference, each layer of the encoder " + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_2" + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "text", + "content": " attends over the maintained semantic features through the attention mechanism. As shown in Fig. 3, we illustrate the prototype aggregation process in the " + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "text", + "content": "-th layer of the encoder " + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_2" + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "text", + "content": "-th iteration. The input of the refinement module consists of the original visual feature " + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\pmb{F}" + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "(k-1)" + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "text", + "content": "-th prototype " + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\pmb{E}^{k-1}" + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "text", + "content": ". The attention mechanism " + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathrm{attn}(\\pmb{q},\\pmb{K},\\pmb{V})" + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "text", + "content": " is originally used in Transformer [43], which is formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 357, + 183, + 545, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 183, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 357, + 183, + 545, + 217 + ], + "type": "interline_equation", + "content": "\\operatorname {a t t n} (\\boldsymbol {q}, \\boldsymbol {K}, \\boldsymbol {V}) = \\sum_ {i = 1} ^ {| V |} \\alpha_ {i} \\boldsymbol {W} _ {v} \\boldsymbol {v} _ {i}, \\tag {4}", + "image_path": "88f9a9d636891c4b80b04582bb68748e45bc2f071826c978da76950adba2aa6e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 358, + 220, + 494, + 234 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 220, + 494, + 234 + ], + "spans": [ + { + "bbox": [ + 358, + 220, + 494, + 234 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} = \\operatorname {s o f t m a x} \\left(\\left(\\boldsymbol {W} _ {q} \\boldsymbol {q}\\right) ^ {T} \\left(\\boldsymbol {W} _ {k} \\boldsymbol {k} _ {i}\\right)\\right),", + "image_path": "b6ceb47dd2a27f3091c8a40157093b1a3378794928216fd900400295d3c4444e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 239, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 545, + 317 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 239, + 545, + 317 + ], + "type": "inline_equation", + "content": "W_{q}, W_{k}" + }, + { + "bbox": [ + 306, + 239, + 545, + 317 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 239, + 545, + 317 + ], + "type": "inline_equation", + "content": "W_{v}" + }, + { + "bbox": [ + 306, + 239, + 545, + 317 + ], + "type": "text", + "content": " are learnable parameters. To make better use of the " + }, + { + "bbox": [ + 306, + 239, + 545, + 317 + ], + "type": "inline_equation", + "content": "(k - 1)" + }, + { + "bbox": [ + 306, + 239, + 545, + 317 + ], + "type": "text", + "content": "-th prototype, we consider the " + }, + { + "bbox": [ + 306, + 239, + 545, + 317 + ], + "type": "inline_equation", + "content": "(k - 1)" + }, + { + "bbox": [ + 306, + 239, + 545, + 317 + ], + "type": "text", + "content": "-th prototype " + }, + { + "bbox": [ + 306, + 239, + 545, + 317 + ], + "type": "inline_equation", + "content": "E^{k - 1} = \\{e_t^{k - 1}\\}_{t = 1}^{T_f}" + }, + { + "bbox": [ + 306, + 239, + 545, + 317 + ], + "type": "text", + "content": " as key-value pair, and the original visual feature " + }, + { + "bbox": [ + 306, + 239, + 545, + 317 + ], + "type": "inline_equation", + "content": "F = \\{f_t\\}_{t = 1}^{T_f}" + }, + { + "bbox": [ + 306, + 239, + 545, + 317 + ], + "type": "text", + "content": " as the query. In this way, we inject the rich semantic information into the new prototype " + }, + { + "bbox": [ + 306, + 239, + 545, + 317 + ], + "type": "inline_equation", + "content": "E^k = \\{e_t^k\\}_{t = 1}^{T_f}" + }, + { + "bbox": [ + 306, + 239, + 545, + 317 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 318, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 318, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 305, + 318, + 545, + 353 + ], + "type": "text", + "content": "The encoder " + }, + { + "bbox": [ + 305, + 318, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_2" + }, + { + "bbox": [ + 305, + 318, + 545, + 353 + ], + "type": "text", + "content": " is composed of " + }, + { + "bbox": [ + 305, + 318, + 545, + 353 + ], + "type": "inline_equation", + "content": "L_{e}" + }, + { + "bbox": [ + 305, + 318, + 545, + 353 + ], + "type": "text", + "content": " identical layers, where " + }, + { + "bbox": [ + 305, + 318, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\pmb{E}_{l}^{k}" + }, + { + "bbox": [ + 305, + 318, + 545, + 353 + ], + "type": "text", + "content": " denotes the output of the " + }, + { + "bbox": [ + 305, + 318, + 545, + 353 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 318, + 545, + 353 + ], + "type": "text", + "content": "-th layer at the " + }, + { + "bbox": [ + 305, + 318, + 545, + 353 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 318, + 545, + 353 + ], + "type": "text", + "content": "-th iteration. Similar to [51], the hidden state is computed as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 333, + 359, + 545, + 391 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 359, + 545, + 391 + ], + "spans": [ + { + "bbox": [ + 333, + 359, + 545, + 391 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\widetilde {\\boldsymbol {E}} _ {l} ^ {k} = \\beta \\cdot \\operatorname {a t t n} _ {s} \\left(\\boldsymbol {E} _ {l - 1} ^ {k}, \\boldsymbol {E} _ {l - 1} ^ {k}, \\boldsymbol {E} _ {l - 1} ^ {k}\\right) \\tag {5} \\\\ + (1 - \\beta) \\cdot \\operatorname {a t t n} _ {c} \\left(\\boldsymbol {E} _ {l - 1} ^ {k}, \\boldsymbol {E} ^ {k - 1}, \\boldsymbol {E} ^ {k - 1}\\right), \\\\ \\end{array}", + "image_path": "3c8af0606fb3954a5b13cf58053a88b4437927c1df2637cadd33bf4dd5fd5e47.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{attn}_s" + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{attn}_c" + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "text", + "content": " denote the self-attention sub-layer and cross-attention sub-layer used in Transformer [43], respectively. For the first layer of the encoder " + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_2" + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "inline_equation", + "content": "E_{l - 1}^k" + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "text", + "content": " is equal to the visual feature " + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "inline_equation", + "content": "\\pmb{F}" + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "text", + "content": ". The output of the encoder " + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_2" + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "inline_equation", + "content": "E_{L_e}^k" + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "inline_equation", + "content": "E^k" + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "inline_equation", + "content": "\\beta \\in [0,1]" + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "text", + "content": " is a hyperparameter that weights the importance of previous prototypes during training and inference. To further fuse and refine the prototype, it is linked with a fully connected sub-layer " + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{FFN}(\\cdot)" + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "text", + "content": " using the residual connection. The output of the " + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "text", + "content": "-th layer at the " + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 398, + 545, + 519 + ], + "type": "text", + "content": "-th iteration is formulated as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 526, + 545, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 545, + 541 + ], + "type": "interline_equation", + "content": "\\boldsymbol {E} _ {l} ^ {k} = \\operatorname {F F N} \\left(\\operatorname {L N} \\left(\\widetilde {\\boldsymbol {E}} _ {l} ^ {k} + \\boldsymbol {E} _ {l - 1} ^ {k}\\right)\\right) + \\operatorname {L N} \\left(\\widetilde {\\boldsymbol {E}} _ {l} ^ {k} + \\boldsymbol {E} _ {l - 1} ^ {k}\\right), \\tag {6}", + "image_path": "6e3c35ded5083bb1d94eebb784993f5e0ee8bab048afed144b10a8ced127101b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 548, + 506, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 548, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 305, + 548, + 506, + 560 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 548, + 506, + 560 + ], + "type": "inline_equation", + "content": "\\mathrm{LN}(\\cdot)" + }, + { + "bbox": [ + 305, + 548, + 506, + 560 + ], + "type": "text", + "content": " is the layer normalization operation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 562, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 562, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 305, + 562, + 545, + 638 + ], + "type": "text", + "content": "Translation generation. The decoder " + }, + { + "bbox": [ + 305, + 562, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 305, + 562, + 545, + 638 + ], + "type": "text", + "content": " iteratively takes the prototype " + }, + { + "bbox": [ + 305, + 562, + 545, + 638 + ], + "type": "inline_equation", + "content": "E^k" + }, + { + "bbox": [ + 305, + 562, + 545, + 638 + ], + "type": "text", + "content": " as input and generates the corresponding translation " + }, + { + "bbox": [ + 305, + 562, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\pmb{Y}^{k}" + }, + { + "bbox": [ + 305, + 562, + 545, + 638 + ], + "type": "text", + "content": ". Take the " + }, + { + "bbox": [ + 305, + 562, + 545, + 638 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 562, + 545, + 638 + ], + "type": "text", + "content": "-th iteration as an example. The decoder " + }, + { + "bbox": [ + 305, + 562, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 305, + 562, + 545, + 638 + ], + "type": "text", + "content": " predicts the conditional probability of translation " + }, + { + "bbox": [ + 305, + 562, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\pmb{Y}^{k} = \\{\\pmb{y}_{t}^{k}\\}_{t = 1}^{T_{y,k}}" + }, + { + "bbox": [ + 305, + 562, + 545, + 638 + ], + "type": "text", + "content": " based on the predicted logits " + }, + { + "bbox": [ + 305, + 562, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\pmb{U}^{k} = \\{\\pmb{u}_{t}^{k}\\}_{t = 1}^{T_{y,k}}" + }, + { + "bbox": [ + 305, + 562, + 545, + 638 + ], + "type": "text", + "content": ", which is computed as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 330, + 643, + 545, + 675 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 643, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 330, + 643, + 545, + 675 + ], + "type": "interline_equation", + "content": "p ^ {k - t h} \\left(\\boldsymbol {Y} ^ {k} \\mid \\boldsymbol {X}\\right) = \\prod_ {t = 1} ^ {T _ {y, k}} p \\left(\\boldsymbol {y} _ {t} ^ {k} \\mid \\boldsymbol {F}, \\boldsymbol {E} ^ {k - 1}, \\boldsymbol {y} _ {0: t - 1} ^ {k}\\right), \\tag {7}", + "image_path": "acfb1bacad69ecb95fd1e303ae8c1d5bc2c43a96d0a8849a75eb1029ad080ada.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 339, + 677, + 520, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 677, + 520, + 694 + ], + "spans": [ + { + "bbox": [ + 339, + 677, + 520, + 694 + ], + "type": "interline_equation", + "content": "p (\\boldsymbol {y} _ {t} ^ {k} | \\boldsymbol {F}, \\boldsymbol {E} ^ {k - 1}, \\boldsymbol {y} _ {0: t - 1} ^ {k}) = s o f t m a x (\\boldsymbol {u} _ {t} ^ {k}) _ {\\boldsymbol {y} _ {t} ^ {k}},", + "image_path": "ac04710a46c9096aad560649933d5d0705beae87530c07780b21b2282a9f4212.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 700, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 700, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 700, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 308, + 700, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{u}_t^k" + }, + { + "bbox": [ + 308, + 700, + 545, + 714 + ], + "type": "text", + "content": " is the predicted logits of the decoder " + }, + { + "bbox": [ + 308, + 700, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 308, + 700, + 545, + 714 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 308, + 700, + 545, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 308, + 700, + 545, + 714 + ], + "type": "text", + "content": "-" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15595" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "text", + "content": "th step. The decoder " + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "text", + "content": " iteratively takes the prototype given by the encoder " + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_2" + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "text", + "content": " as input to generate the final translation " + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^K" + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "text", + "content": ". With total " + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "text", + "content": " iterations, the outputs of encoder " + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_1" + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_2" + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "text", + "content": " compose the prototype sequence " + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "inline_equation", + "content": "\\mathbf{E} = \\{\\mathbf{E}^k\\}_{k=0}^K" + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^k = \\{\\mathbf{e}_t^k\\}_{t=1}^{T_f}" + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "text", + "content": " is the semantic feature at the " + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "text", + "content": "-th iteration. Accordingly, the decoder " + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "text", + "content": " generate " + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "inline_equation", + "content": "K+1" + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "text", + "content": " translations " + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} = \\{\\mathbf{Y}^k\\}_{k=0}^K" + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^k = \\{\\mathbf{y}_t^k\\}_{t=1}^{T_y,k}" + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "text", + "content": " is the translation at the " + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "text", + "content": "-th iteration during training. Note that after " + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "text", + "content": " iterations, the decoder " + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "text", + "content": " obtains the converged prototype " + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^K" + }, + { + "bbox": [ + 47, + 72, + 289, + 196 + ], + "type": "text", + "content": " and generates the translation only once in inference." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 204, + 160, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 204, + 160, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 204, + 160, + 217 + ], + "type": "text", + "content": "3.5. Training Objective" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 223, + 289, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 223, + 289, + 356 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 289, + 356 + ], + "type": "text", + "content": "We introduce two kinds of losses in the training period of the IP-SLT system. Firstly, the cross entropy loss is adopted to supervise the final generated sentence. Secondly, we put forward an iterative distillation loss. As the decoder " + }, + { + "bbox": [ + 47, + 223, + 289, + 356 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 47, + 223, + 289, + 356 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 223, + 289, + 356 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 47, + 223, + 289, + 356 + ], + "type": "text", + "content": " generate translation sequence " + }, + { + "bbox": [ + 47, + 223, + 289, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} = \\{\\mathbf{Y}^0,\\mathbf{Y}^1,\\dots ,\\mathbf{Y}^K\\}" + }, + { + "bbox": [ + 47, + 223, + 289, + 356 + ], + "type": "text", + "content": ", we naturally divide the sequence " + }, + { + "bbox": [ + 47, + 223, + 289, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 47, + 223, + 289, + 356 + ], + "type": "text", + "content": " into an initial prediction, " + }, + { + "bbox": [ + 47, + 223, + 289, + 356 + ], + "type": "inline_equation", + "content": "K - 1" + }, + { + "bbox": [ + 47, + 223, + 289, + 356 + ], + "type": "text", + "content": " intermediate predictions, and the final prediction according to the order of iteration. Conceptually, the " + }, + { + "bbox": [ + 47, + 223, + 289, + 356 + ], + "type": "inline_equation", + "content": "K - 1" + }, + { + "bbox": [ + 47, + 223, + 289, + 356 + ], + "type": "text", + "content": " intermediate predictions are regarded as the student model and distill knowledge from the final prediction which is regarded as the teacher model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 359, + 289, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 359, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 289, + 456 + ], + "type": "text", + "content": "Cross entropy loss. As mentioned above, the IP-SLT generates the translation sequence based on the conditional probability provided by the decoder " + }, + { + "bbox": [ + 46, + 359, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 46, + 359, + 289, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 359, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 46, + 359, + 289, + 456 + ], + "type": "text", + "content": ". The cross-entropy loss [43] is computed with the ground truth from " + }, + { + "bbox": [ + 46, + 359, + 289, + 456 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 359, + 289, + 456 + ], + "type": "text", + "content": " training samples and the outputs of the decoder " + }, + { + "bbox": [ + 46, + 359, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 46, + 359, + 289, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 359, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 46, + 359, + 289, + 456 + ], + "type": "text", + "content": ". Its training objective in the proposed approach is to maximize the log-likelihood which is equal to minimizing the cross entropy loss formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 461, + 287, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 461, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 287, + 476 + ], + "type": "interline_equation", + "content": "L _ {C E, k} = - \\log p ^ {k - t h} (\\hat {\\mathbf {Y}} | \\mathbf {X}), \\tag {8}", + "image_path": "b5a71cccb0a42e41cd5426a00c09022f16e3030c3521731cb6cfc80617f0927d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 487, + 287, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 524 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 487, + 287, + 524 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 46, + 487, + 287, + 524 + ], + "type": "text", + "content": " denotes translation annotation. We apply this at the output of the initialization module and the " + }, + { + "bbox": [ + 46, + 487, + 287, + 524 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 487, + 287, + 524 + ], + "type": "text", + "content": "-th output of the refinement module." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 528, + 289, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 289, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 289, + 624 + ], + "type": "text", + "content": "Iterative distillation loss. Since the KL (Kullback-Leibler) divergence loss can affect the teacher's networks to the student's networks, we compute it between the " + }, + { + "bbox": [ + 46, + 528, + 289, + 624 + ], + "type": "inline_equation", + "content": "K - 1" + }, + { + "bbox": [ + 46, + 528, + 289, + 624 + ], + "type": "text", + "content": " intermediate predictions and the final prediction. As the final translation is based on the previous prototypes, we are able to get better translations by approximating more vital characterization capabilities of middle prototypes. The iterative distillation loss (IDL) is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 634, + 287, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 634, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 287, + 666 + ], + "type": "interline_equation", + "content": "L _ {I D L} = \\sum_ {k = 1} ^ {K - 1} K L \\left(\\boldsymbol {U} ^ {k}, \\boldsymbol {U} ^ {K}\\right), \\tag {9}", + "image_path": "30762e38398bd72d4541f8a755fba5d4db948f85757af6f4689d4be7a7f1c76e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{U}^{k}" + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{U}^{K}" + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": " are the predicted output of the decoder " + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": "-th iteration, respectively. By computing the Kullback-Leibler divergence between the " + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{U}^{k}" + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "\\pmb{U}^{K}" + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": ", the IP-SLT is encouraged to approximate the performance of the final iteration. We apply this at the " + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "K - 1" + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": " shallow outputs of the refinement module." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 108, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 545, + 120 + ], + "type": "text", + "content": "Overall, the loss function of our IP-SLT is formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 354, + 129, + 545, + 143 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 129, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 354, + 129, + 545, + 143 + ], + "type": "interline_equation", + "content": "L = L _ {C E, 0} + L _ {C E, K} + \\lambda \\cdot L _ {I D L}. \\tag {10}", + "image_path": "1fedb7caa8145d94321faa2c34ff1544d797a33edbbc299c7899147f4a546be4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 152, + 545, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 152, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 305, + 152, + 545, + 200 + ], + "type": "text", + "content": "Under the guidance of CE loss, we first train the initial generation module until convergence as a warm start, and then apply the loss function as Equ. (10) for optimizing the IP-SLT system in an end-to-end manner." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 207, + 416, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 207, + 416, + 220 + ], + "spans": [ + { + "bbox": [ + 306, + 207, + 416, + 220 + ], + "type": "text", + "content": "3.6. Inference Strategy" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 225, + 545, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 225, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 304, + 225, + 545, + 332 + ], + "type": "text", + "content": "Since the refinement process only involves in the encoder " + }, + { + "bbox": [ + 304, + 225, + 545, + 332 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_2" + }, + { + "bbox": [ + 304, + 225, + 545, + 332 + ], + "type": "text", + "content": ", the initialization and iterative refinement modules behave differently during training and inference. The decoder " + }, + { + "bbox": [ + 304, + 225, + 545, + 332 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 304, + 225, + 545, + 332 + ], + "type": "text", + "content": " of the initialization module is leveraged to provide guidance for the encoder " + }, + { + "bbox": [ + 304, + 225, + 545, + 332 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_1" + }, + { + "bbox": [ + 304, + 225, + 545, + 332 + ], + "type": "text", + "content": " during the training process, while it is not required in inference. Similarly, the autoregressive decoder " + }, + { + "bbox": [ + 304, + 225, + 545, + 332 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 304, + 225, + 545, + 332 + ], + "type": "text", + "content": " of the iterative refinement module generates translations for each iteration during training, while it just decodes the final translation once in inference." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 333, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 333, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 545, + 417 + ], + "type": "text", + "content": "In inference, given a sign language video to be translated, the feature extraction module first converts it into a visual feature. The encoder " + }, + { + "bbox": [ + 304, + 333, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_1" + }, + { + "bbox": [ + 304, + 333, + 545, + 417 + ], + "type": "text", + "content": " of the initialization module transforms it into the raw prototype. Then, the encoder " + }, + { + "bbox": [ + 304, + 333, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_2" + }, + { + "bbox": [ + 304, + 333, + 545, + 417 + ], + "type": "text", + "content": " of the refinement module further iteratively refines it by fusing it with the original visual feature. Finally, the decoder " + }, + { + "bbox": [ + 304, + 333, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 304, + 333, + 545, + 417 + ], + "type": "text", + "content": " generates the translation based on the final prototype." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 427, + 388, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 427, + 388, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 427, + 388, + 441 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 447, + 422, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 447, + 422, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 447, + 422, + 460 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 468, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 601 + ], + "type": "text", + "content": "Dataset. We evaluate our approach on two pubic sign language translation datasets, i.e., PHOENIX14T [5] and CSL-Daily [49]. Both datasets provide gloss-level and spoken-sentence-level annotations. The PHOENIX14T dataset [5] is the first large-scale neural SLT dataset created with 9 German sign language interpreters. The dataset is split into a training set (7,096), a development set (519), and a test set (642). The CSL-Daily dataset [49] is a Chinese SLT dataset containing 20,654 annotated sign language videos. We follow the previous experimental setting [49] and split it into the training, development, and test set." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "text", + "content": "Evaluation metrics. Following the work [49], we quantitatively assess the quality of translations according to the BLEU- " + }, + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "text", + "content": " [32] and ROUGE [38]. The BLEU- " + }, + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "text", + "content": " ranges from 1 to 4) cares more about the accuracy of the predicted translation while the ROUGE cares more about the consistency of sentences. For both evaluation metrics, a higher value indicates a better performance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Training settings. We implement our approach on Pytorch [33]. The encoder and decoder of the initialization" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15596" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 533, + 183 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 533, + 183 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 533, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 533, + 183 + ], + "type": "table", + "html": "
MethodsLLMDevTest
ROUGEBLEU-1BLEU-2BLEU-3BLEU-4ROUGEBLEU-1BLEU-2BLEU-3BLEU-4
Joint-SLRT [7]✘-47.2634.4027.0522.38-46.6133.7326.1921.32
PET [23]✘-----49.9749.5437.1929.3024.02
BN-TIN-Transf. [49]✘46.8746.9033.9826.4921.7846.9847.5734.6426.7821.68
STMC [50]✘48.2447.6036.4329.1824.0946.6546.9836.0928.7023.65
IP-SLT✘54.4354.1041.5633.6628.2253.7254.2541.5133.4527.97
MMTLB [9]✓53.1053.9541.1233.1427.6152.6553.9741.7533.8428.39
TwoStream-SLT [10]✓54.0854.3241.9934.1528.6653.4854.9042.4334.4628.95
", + "image_path": "a32ea604528750f74c8623e2612ddc7b2b1eb5b26e595f01e38fa039004bc594.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 213, + 533, + 316 + ], + "blocks": [ + { + "bbox": [ + 47, + 185, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 47, + 185, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 47, + 185, + 545, + 208 + ], + "type": "text", + "content": "Table 1. Performance comparison of IP-SLT with methods for SLT on PHOENIX-2014T. 'LLM' denotes adopting pre-trained large language models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 213, + 533, + 316 + ], + "lines": [ + { + "bbox": [ + 61, + 213, + 533, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 213, + 533, + 316 + ], + "type": "table", + "html": "
MethodsLLMDevTest
ROUGEBLEU-1BLEU-2BLEU-3BLEU-4ROUGEBLEU-1BLEU-2BLEU-3BLEU-4
SL-Luong [5]X34.2834.2219.7212.247.9634.5434.1619.5711.847.56
Joint-SLRT [7]X37.0637.4724.6716.8611.8836.7437.3824.3616.5511.79
BN-TIN-Transf. [49]X37.2940.6626.5618.0612.7337.6740.7426.9618.4813.19
IP-SLTX44.3345.2631.7722.8716.7444.0944.8531.5022.6616.72
MMTLB [9]✓53.3853.8140.8431.2924.4253.2553.3140.4130.8723.92
TwoStream-SLT [10]✓55.1055.2142.3132.7125.7655.7255.4442.5932.8725.79
", + "image_path": "7eb3b109c201c7c7aeceb933ce145efff43039c009f765088d640bead7dafe17.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 318, + 546, + 340 + ], + "lines": [ + { + "bbox": [ + 47, + 318, + 546, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 546, + 340 + ], + "type": "text", + "content": "Table 2. Performance comparison of IP-SLT with methods for SLT on CSL-Daily. 'LLM' denotes adopting pre-trained large language models." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 354, + 289, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 289, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 289, + 536 + ], + "type": "text", + "content": "and refinement module consist of 3 layers, respectively. The dimension of the feed-forward network is 2048. The visual feature of the STMC [50] is 1024-dimension while the visual feature of BN-TIN-Transf. [49] and VAC [31] are 512-dimension. To alleviate over-fitting, we set dropout and attention head to 0.1 and 8, respectively. The training optimizer is Adam [24]. During training, the learning rate is fixed to " + }, + { + "bbox": [ + 46, + 354, + 289, + 536 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 354, + 289, + 536 + ], + "type": "text", + "content": ". To ensure the features provided by the previous prototype and original sign video are fully utilized, we apply the drop-net [51] during training, which effects Equ. (5). During training, for any layer in the encoder " + }, + { + "bbox": [ + 46, + 354, + 289, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_2" + }, + { + "bbox": [ + 46, + 354, + 289, + 536 + ], + "type": "text", + "content": ", with probability " + }, + { + "bbox": [ + 46, + 354, + 289, + 536 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 354, + 289, + 536 + ], + "type": "text", + "content": ", the hidden state " + }, + { + "bbox": [ + 46, + 354, + 289, + 536 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{E}}_l^k" + }, + { + "bbox": [ + 46, + 354, + 289, + 536 + ], + "type": "text", + "content": " in Equ. (5) is the output of the self-attention sub-layer " + }, + { + "bbox": [ + 46, + 354, + 289, + 536 + ], + "type": "inline_equation", + "content": "\\mathrm{attn}_s" + }, + { + "bbox": [ + 46, + 354, + 289, + 536 + ], + "type": "text", + "content": "; with probability " + }, + { + "bbox": [ + 46, + 354, + 289, + 536 + ], + "type": "inline_equation", + "content": "1 - \\beta" + }, + { + "bbox": [ + 46, + 354, + 289, + 536 + ], + "type": "text", + "content": ", it is the output of the cross-attention sub-layer " + }, + { + "bbox": [ + 46, + 354, + 289, + 536 + ], + "type": "inline_equation", + "content": "\\mathrm{attn}_c" + }, + { + "bbox": [ + 46, + 354, + 289, + 536 + ], + "type": "text", + "content": ". In inference, the hidden state " + }, + { + "bbox": [ + 46, + 354, + 289, + 536 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{E}}_l^k" + }, + { + "bbox": [ + 46, + 354, + 289, + 536 + ], + "type": "text", + "content": " is computed as Equ. (5)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 539, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 624 + ], + "type": "text", + "content": "Inference details. In inference, we use the beam search strategy [45] to improve the decoding accuracy. For the PHOENIX-2014T dataset and the CSL-Daily dataset, we set the beam search width and the length penalty to 3 and 1.0, respectively. To reduce the computational complexity, our IP-SLT just decodes once in inference for each input sign language video." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 634, + 272, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 272, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 272, + 647 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "We compare the proposed IP-SLT with the previous SLT systems on two public benchmarks, i.e., PHOENIX14T [5] and CSL-Daily [49], and the performance of our IP-SLT is shown in Tab. 1 and Tab. 2, respectively. For PHOENIX14T and CSL-Daily dataset, we adopt the STMC [50] and BN-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 354, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 426 + ], + "type": "text", + "content": "TIN-Transf. [49] as the baseline, respectively. Our IP-SLT follows the sign-to-text (S2T) paradigm, which directly transforms the sign language video into translation. Note MMTLB [9] and TwoStream-SLT [10] adopt pre-trained large-scale language models that leverage more model parameters and extra resources than IP-SLT." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 427, + 546, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 546, + 560 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 546, + 560 + ], + "type": "text", + "content": "By combining all proposed components together, our IP-SLT achieves substantial improvements against the baseline. The IP-SLT achieves 28.22 and 16.74 BLEU-4 scores on the DEV set of PHOENIX14T and CSL-Daily, respectively. The quantitative results demonstrate that our IP-SLT achieves promising performance improvements. Our IP-SLT delivers promising performance gains on DEV and test sets by leveraging the iterative refinement process. The results prove the advantage of aggregating the previous translation progress and the original visual representation, which distinguishes our IP-SLT from previous SLT systems." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 570, + 406, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 570, + 406, + 582 + ], + "spans": [ + { + "bbox": [ + 306, + 570, + 406, + 582 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 589, + 545, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 589, + 545, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 589, + 545, + 637 + ], + "type": "text", + "content": "In this section, we put forward several ablation studies on the DEV set of PHOENIX-2014T. Unless otherwise specified, we adopt the STMC [50] as the baseline for the following experiments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "type": "text", + "content": "Impact of network architecture. The main difference between our proposed method and the existing work is to leverage the previous information as an additional clue to enhance the current prototype. To evaluate the effectiveness of each component, we gradually add the refinement module and the iterative distillation loss to the baseline SLT sys" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15597" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 285, + 141 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 285, + 141 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 285, + 141 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 285, + 141 + ], + "type": "table", + "html": "
SettingROUGEBLEU-1BLEU-2BLEU-3BLEU-4
Baseline50.1050.2137.1229.4124.31
+Refinement51.2251.0438.4030.6125.39
+IDL54.4354.1041.5633.6628.22
6-6 Layers50.4950.5837.5729.6424.53
", + "image_path": "8522843bb86dc3c38234e11689b7c8e3dc34aafe9d6cc69f2a4b38924bcfd6a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 198, + 285, + 294 + ], + "blocks": [ + { + "bbox": [ + 47, + 146, + 287, + 191 + ], + "lines": [ + { + "bbox": [ + 47, + 146, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 47, + 146, + 287, + 191 + ], + "type": "text", + "content": "Table 3. Effect of our proposed components. 'Refinement' denotes applying the refinement process. 'IDL' denotes applying the iterative distillation loss. '6-6 layers' denotes enlarging the encoder and decoder of the baseline system from 3 to 6 layers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 198, + 285, + 294 + ], + "lines": [ + { + "bbox": [ + 50, + 198, + 285, + 294 + ], + "spans": [ + { + "bbox": [ + 50, + 198, + 285, + 294 + ], + "type": "table", + "html": "
ModelRMROUGEBLEU-1BLEU-2BLEU-3BLEU-4
STMCw/o50.1050.2137.1229.4124.31
w/54.4354.1041.5633.6628.22
BN-TIN-Transf.w/o47.4147.9934.9427.3322.35
w/52.0652.0639.0131.0825.69
VAC-Transf.w/o49.4850.0137.0029.1223.91
w/53.6853.6041.2833.4728.07
", + "image_path": "ee687f7bfe7cf367471e71842f47c433cfae205f17d847dc89021e61cc41d109.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 339, + 285, + 412 + ], + "blocks": [ + { + "bbox": [ + 47, + 297, + 287, + 332 + ], + "lines": [ + { + "bbox": [ + 47, + 297, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 287, + 332 + ], + "type": "text", + "content": "Table 4. Generalization of IP-SLT. 'RM' denotes leveraging the refinement process. 'w' and 'w/o' denote the baseline SLT system with and without a refinement process, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 339, + 285, + 412 + ], + "lines": [ + { + "bbox": [ + 50, + 339, + 285, + 412 + ], + "spans": [ + { + "bbox": [ + 50, + 339, + 285, + 412 + ], + "type": "table", + "html": "
λROUGEBLEU-1BLEU-2BLEU-3BLEU-4
051.3050.3238.0130.2925.14
553.0752.3140.0832.2927.02
1053.7153.5140.9232.9427.47
1554.4354.1041.5633.6628.22
2054.4254.1641.5133.4427.87
", + "image_path": "b0e0250ea67f823d4e3d0376da2117064893bd6651bc116588e466ade07bcdcb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 445, + 286, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 286, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 286, + 554 + ], + "type": "text", + "content": "tem. Directly applying the refinement process to the baseline delivers a performance gain of 1.08 BLEU-4. We further apply the iterative distillation loss to improve the performance. The results suggest that adding distillation supervision can be helpful with a gain of 2.83 BLEU-4. Besides, to keep the number of parameters unchanged, we enlarge the depth from 3 to 6 layers and evaluate the performance. Naively enlarging the model scale slightly improves the performance " + }, + { + "bbox": [ + 46, + 445, + 286, + 554 + ], + "type": "inline_equation", + "content": "(+0.22" + }, + { + "bbox": [ + 46, + 445, + 286, + 554 + ], + "type": "text", + "content": " BLEU-4). The results are shown in Tab. 3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 713 + ], + "type": "text", + "content": "Generalization of the IP-SLT. We conduct three sets of experiments by changing the visual backbone to evaluate the generalization of the proposed IP-SLT approach in Tab. 4. Specifically, the BN-TIN-Transf. [49] uses a basic CNN network to get the dense representation of sign video. The STMC [50] extracts and aligns the key parts of sign language expression to achieve better performance. The VAC [31] proposes two auxiliary supervision methods to enhance the feature extraction module. VAC-Transf. replaces the feature extractor in BN-TIN-Transf. [49] with the visual backbone of VAC [31]. Applying our proposed IP-SLT methods to the BN-TIN-Transf., VAC-Transf. and STMC, we achieve 25.69, 28.07, and 28.22 BLEU-4 scores" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 543, + 170 + ], + "blocks": [ + { + "bbox": [ + 47, + 416, + 268, + 428 + ], + "lines": [ + { + "bbox": [ + 47, + 416, + 268, + 428 + ], + "spans": [ + { + "bbox": [ + 47, + 416, + 268, + 428 + ], + "type": "text", + "content": "Table 5. The weight " + }, + { + "bbox": [ + 47, + 416, + 268, + 428 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 416, + 268, + 428 + ], + "type": "text", + "content": " of iterative distillation loss to CE loss." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 71, + 543, + 170 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 543, + 170 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 543, + 170 + ], + "type": "table", + "html": "
βROUGEBLEU-1BLEU-2BLEU-3BLEU-4
0.050.1050.2137.1229.4124.31
0.254.3553.8141.4033.5228.02
0.454.3954.0241.2433.1927.70
0.554.4354.1041.5633.6628.22
0.653.9653.6340.9433.1027.72
0.854.3153.3940.9232.9827.53
-50.9750.6738.2430.5225.41
", + "image_path": "5221bfb387a3cb48c974c2130ff2f8d239d6b8cc5688cfc91527a807b74e5f9d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 215, + 543, + 280 + ], + "blocks": [ + { + "bbox": [ + 305, + 175, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 305, + 175, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 305, + 175, + 545, + 208 + ], + "type": "text", + "content": "Table 6. The weight " + }, + { + "bbox": [ + 305, + 175, + 545, + 208 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 175, + 545, + 208 + ], + "type": "text", + "content": " of the original visual feature to the previous prototype. '0.0' denotes the assessment of the baseline. '-' denotes setting " + }, + { + "bbox": [ + 305, + 175, + 545, + 208 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 175, + 545, + 208 + ], + "type": "text", + "content": " as 0.5 without using the drop-net [51]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 215, + 543, + 280 + ], + "lines": [ + { + "bbox": [ + 309, + 215, + 543, + 280 + ], + "spans": [ + { + "bbox": [ + 309, + 215, + 543, + 280 + ], + "type": "table", + "html": "
SettingROUGEBLEU-1BLEU-2BLEU-3BLEU-4
None50.1050.2137.1229.4124.31
Con-input54.3753.8341.1032.9727.39
Con-feature53.9752.7640.5632.8127.47
Add-feature54.4354.1041.5633.6628.22
", + "image_path": "f4117220fae99f97ae73457b2fe858ef5cd208afe9630d33f66e311107002845.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 285, + 545, + 351 + ], + "lines": [ + { + "bbox": [ + 305, + 285, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 545, + 351 + ], + "type": "text", + "content": "Table 7. Effect of the different refinement methods. 'Con-input' denotes directly concatenating the original visual feature and previous prototype as input. 'Con-feature' denotes concatenating the original feature and the previous feature given by the cross-attention mechanism in each layer. On top of that, 'Add-feature' denotes changing it to an addition operation." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 369, + 545, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 369, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 545, + 405 + ], + "type": "text", + "content": "on the DEV set, surpassing the baselines by 3.34, 4.16, 3.91, respectively. Using a high-quality visual backbone delivers further quality gains." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 407, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 545, + 479 + ], + "type": "text", + "content": "Impact of " + }, + { + "bbox": [ + 305, + 407, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 407, + 545, + 479 + ], + "type": "text", + "content": ". In our experiments, the weight " + }, + { + "bbox": [ + 305, + 407, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 407, + 545, + 479 + ], + "type": "text", + "content": " of iterative distillation loss is set to 15. It is a hyper-parameter that is designed to balance the effect of cross-entropy loss and the iterative distillation loss. We conduct experiments by varying the weight " + }, + { + "bbox": [ + 305, + 407, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 407, + 545, + 479 + ], + "type": "text", + "content": ". Tab. 5 shows that our IP-SLT achieves the best performance when the weight " + }, + { + "bbox": [ + 305, + 407, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 407, + 545, + 479 + ], + "type": "text", + "content": " is set to 15." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 482, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 545, + 590 + ], + "type": "text", + "content": "Impact of " + }, + { + "bbox": [ + 304, + 482, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 482, + 545, + 590 + ], + "type": "text", + "content": ". In the above experiments, the weight " + }, + { + "bbox": [ + 304, + 482, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 482, + 545, + 590 + ], + "type": "text", + "content": " is fixed to 0.5. The weight " + }, + { + "bbox": [ + 304, + 482, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 482, + 545, + 590 + ], + "type": "text", + "content": " represents the importance of the previous prototype compared with the original visual feature. As a hyper-parameter of our proposed methods, the weight " + }, + { + "bbox": [ + 304, + 482, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 482, + 545, + 590 + ], + "type": "text", + "content": " is examined with a set of different values in Tab. 6. When the weight is 0.5, the performance is the highest one. This indicates that the previous prototype to SLT is as important as the original visual feature. Besides, to fully use the previous prototype, the drop-net is required." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Impact of refinement method. We also examine a set of refinement methods for IP-SLT considering the fusion mechanism as a key part of our proposed method in Tab. 7. Directly concatenating the original visual feature and previous prototype in the time dimension improves the performance of SLT from 24.31 to 27.39. We further add the cross-attention mechanism to explicitly leverage the useful information from the previous prototype. And then, we concatenate the representation from the original visual feature and the representation from the previous prototype in fea" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15598" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 285, + 142 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 285, + 142 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 285, + 142 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 285, + 142 + ], + "type": "table", + "html": "
KROUGEBLEU-1BLEU-2BLEU-3BLEU-4
050.1050.2137.1229.4124.31
151.2251.0438.4030.6125.39
253.7353.3940.7632.9827.66
354.4354.1041.5633.6628.22
454.6353.9141.4033.5328.01
", + "image_path": "28bb906546f9678e6d09a5a5eb4a81444305ff86d49116319d6790e40dd06c1f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 175, + 285, + 297 + ], + "blocks": [ + { + "bbox": [ + 47, + 147, + 286, + 169 + ], + "lines": [ + { + "bbox": [ + 47, + 147, + 286, + 169 + ], + "spans": [ + { + "bbox": [ + 47, + 147, + 286, + 169 + ], + "type": "text", + "content": "Table 8. Effect of the iteration number " + }, + { + "bbox": [ + 47, + 147, + 286, + 169 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 147, + 286, + 169 + ], + "type": "text", + "content": " in the iterative refinement module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 175, + 285, + 297 + ], + "lines": [ + { + "bbox": [ + 50, + 175, + 285, + 297 + ], + "spans": [ + { + "bbox": [ + 50, + 175, + 285, + 297 + ], + "type": "table", + "html": "
ModelSWI-P(M)T-P(M)FLOPs(B)ROUGEBLEU-4
STMC-92.692.628.750.1024.31
w/128.7172.632.654.4328.22
w/o238.6357.932.654.8128.48
BN-TIN-Transf.-28.328.39.647.4122.35
w/37.753.410.652.0625.69
w/o66.1110.010.652.2425.96
VAC-Transf.-28.328.39.649.4823.91
w/37.753.410.653.6828.07
w/o66.1110.010.654.8228.27
", + "image_path": "1aa3a73c60ce73bfed4ac70203a609b9d505cb3687ee3038b216b2f6c6efcd69.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 301, + 287, + 379 + ], + "lines": [ + { + "bbox": [ + 46, + 301, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 287, + 379 + ], + "type": "text", + "content": "Table 9. Comparing the baseline and non-shared weight prototype refinement method with IP-SLT. 'SW' denotes the sharing weight across all iterations. '-' denotes the baseline SLT system. 'w/' and 'w/o' denote the refinement module in different iterations with and without shared parameters, respectively. 'I-P' and 'T-P' denote the parameter amount calculated in inference and during training, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 396, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 443 + ], + "type": "text", + "content": "ture dimension, which delivers performance gains of 3.16 BLEU-4. On top of that, changing the concatenating process to an element-wise addition operation achieves a further quality gain of 3.91." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 447, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 287, + 555 + ], + "type": "text", + "content": "Impact of iteration number " + }, + { + "bbox": [ + 46, + 447, + 287, + 555 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 447, + 287, + 555 + ], + "type": "text", + "content": ". The iteration number " + }, + { + "bbox": [ + 46, + 447, + 287, + 555 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 447, + 287, + 555 + ], + "type": "text", + "content": " is an important hyper-parameter and is fixed to 3 in the previous experiments. We conduct experiments with different iteration numbers to explore the effect of iteration number " + }, + { + "bbox": [ + 46, + 447, + 287, + 555 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 447, + 287, + 555 + ], + "type": "text", + "content": ". Tab. 8 shows that there is a culmination in the performance at iteration 3. Before this culmination, the performance improvement of our proposed method increases fast. When the iteration number is bigger than 3, the performance of our method is slightly weakened (-0.21 BLEU-4)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "content": "Impact of sharing weights between different iterations. In Tab. 9, we examine the storage and translation quality of our proposed IP-SLT with different baselines i.e., STMC, BN-TIN-Transf, and VAC-Transf. In order to ensure the efficiency of the model, our refinement module shares weight across all iterations. Since the feature extraction module of each group is identical, the parameters of the feature extraction module are not included in the calculation of the storage of different models. As in inference, several parts can be removed from IP-SLT without changing the performance, we report the number of parameters in both training and inference, respectively. The results indicate that our iterative process causes acceptable overhead with re" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 545, + 150 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 545, + 150 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 150 + ], + "type": "table", + "html": "
TypeText
GTich=winsche ihnen einen tatsächien abend und machen sie es gut.
Baselinejetzt wünsche ich ihnen noch einen tatsächien abend.
Ourihnen einen tatsächien abend und machen sie es gut.
GTin der neuenWoche wird es milder aber es bleibt wechselhaft.
Baselinedann wird es wieder milder.
Ourin der neuenWoche wird es dann wieder milder.
", + "image_path": "b856cc2c08e3659ab59d4cb77ff09fe1a0205722d3a240b8c8cf55a897265b51.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 155, + 545, + 188 + ], + "lines": [ + { + "bbox": [ + 305, + 155, + 545, + 188 + ], + "spans": [ + { + "bbox": [ + 305, + 155, + 545, + 188 + ], + "type": "text", + "content": "Table 10. Qualitative evaluation. 'GT' denotes the spoken language translation annotation. 'Baseline' and 'Our' denote the translation result of baseline and our IP-SLT, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 205, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 289 + ], + "type": "text", + "content": "markable performance improvements. We further conduct experiments in which the parameters of each iteration are independent. The results demonstrate that the proposed parameter-shared method achieves near performance gain (28.22 v.s 28.48 BLEU-4, 25.69 v.s 25.96 BLEU-4, and 28.07 v.s 28.27 BLEU-4) with non-shared one but leveraging much less parameter number." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 292, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 545, + 388 + ], + "type": "text", + "content": "Computation comparison with baseline. The FLOPs is a key factor of computation efficiency. We conduct experiments to compare the computation of the proposed IP-SLT method with different baselines. Similarly, we exclude the FLOPs of the feature extraction module and report relevant computational costs in inference. Tab. 9 shows that leveraging the iterative refinement process can cause acceptable computation costs while achieving promising performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 392, + 545, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 451 + ], + "type": "text", + "content": "Case Study. To provide a more intuitionistic view of our proposed method, we list some translation samples of the proposed IP-SLT method in Tab. 10. We observe that based on the previous prototype and original visual representation, the IP-SLT can generate more accurate and fluent sentences." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 463, + 378, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 463, + 378, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 463, + 378, + 475 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 483, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 650 + ], + "type": "text", + "content": "In this work, we propose a new framework IP-SLT which introduces the iterative refinement into a conventional SLT system. With the goal to polish the semantic representation by leveraging the previous results, we present IP-SLT to support the iterative refinement process. The proposed method is differentiable and optimized in an end-to-end manner to achieve its best performance. On top of it, we put forward the iterative distillation loss to further improve the translation quality by leveraging the sequential dependence between the outputs of each iteration. In inference, the autoregressive decoding process is required once to generate the translation based on the final prototype, applying IP-SLT does not significantly affect efficiency. The experimental results demonstrate the effectiveness of the IP-SLT." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgments: This work was supported by NSFC under Contract U20A20183 and 62021001. It was also supported by the GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, and the Supercomputing Center of the USTC." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15599" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Image super-resolution via progressive cascading residual network. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "text", + "content": "[2] Samuel Albanie, Gül Varol, Liliane Momeni, Triantafyllos Afouras, Joon Son Chung, Neil Fox, and Andrew Zisserman. BSL-1K: Scaling up co-articulated sign language recognition using mouthing cues. In Computer Vision European Conference, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 287, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 226 + ], + "type": "text", + "content": "[3] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 287, + 293 + ], + "type": "text", + "content": "[4] Danielle Bragg, Oscar Koller, Mary Bellard, Larwan Berke, Patrick Boudreault, Annelies Braffort, Naomi Caselli, Matt Huenerfauth, Hernisa Kacorri, Tessa Verhoef, et al. Sign language recognition, generation, and translation: An interdisciplinary perspective. In International ACM SIGACCESS Conference on Computers and Accessibility, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 287, + 338 + ], + "type": "text", + "content": "[5] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In IEEE Conference on Computer Vision and Pattern Recognition, 2018. 1, 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 340, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 287, + 384 + ], + "type": "text", + "content": "[6] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Multi-channel transformers for multi-articulatory sign language translation. In Computer Vision-ECCV 2020 Workshops, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 386, + 287, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 287, + 440 + ], + "type": "text", + "content": "[7] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "type": "text", + "content": "[8] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 531 + ], + "type": "text", + "content": "[9] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 577 + ], + "type": "text", + "content": "[10] Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie Liu, and Brian Mak. Two-stream network for sign language recognition and translation. Advances in Neural Information Processing Systems, 2022. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "type": "text", + "content": "[11] Helen Cooper, Brian Holt, and Richard Bowden. Sign language recognition. Visual Analysis of Humans: Looking at People, 2011. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "text", + "content": "[12] Kearsy Cormier, Neil Fox, Bencie Woll, Andrew Zisserman, Necati Cihan Camgoz, and Richard Bowden. ExTOL: Automatic recognition of British sign language using the bsl corpus. In Workshop on Sign Language Translation and Avatar Technology, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[13] Runpeng Cui, Hu Liu, and Changshui Zhang. Recurrent convolutional neural networks for continuous sign language recognition by staged optimization. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[14] Runpeng Cui, Hu Liu, and Changshui Zhang. A deep neural framework for continuous sign language recognition by iterative training. IEEE Transactions on Multimedia, 2019. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[15] Hao Feng, Wengang Zhou, Jiajun Deng, Qi Tian, and Houqiang Li. DocScanner: robust document image rectification with progressive learning. arXiv preprint arXiv:2110.14968, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "text", + "content": "[16] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. Convolutional sequence to sequence learning. In International Conference on Machine Learning, 2017. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 209, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 262 + ], + "type": "text", + "content": "[17] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In International Conference on Machine learning, 2006. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[18] Karol Gregor, Ivo Danihelka, Alex Graves, Danilo Rezende, and Daan Wierstra. DRAW: A recurrent neural network for image generation. In International Conference on Machine Learning, 2015. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "type": "text", + "content": "[19] Jun He, Richang Hong, Xueliang Liu, Mingliang Xu, and Qianru Sun. Revisiting local descriptor for improved few-shot classification. ACM Transactions on Multimedia Computing, Communications, and Applications, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 355, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 398 + ], + "type": "text", + "content": "[20] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. Signbert+: Hand-model-aware self-supervised pretraining for sign language understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "text", + "content": "[21] Jie Huang, Wengang Zhou, Qilin Zhang, Houqiang Li, and Weiping Li. Video-based sign language recognition without temporal segmentation. In AAAI Conference on Artificial Intelligence, 2018. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "type": "text", + "content": "[22] Shuiwang Ji, Wei Xu, Ming Yang, and Kai Yu. 3D convolutional neural networks for human action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2012. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "type": "text", + "content": "[23] Tao Jin, Zhou Zhao, Meng Zhang, and Xingshan Zeng. Prior knowledge and memory enriched transformer for sign language translation. In Findings of the Association for Computational Linguistics, 2022. 1, 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 535, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 566 + ], + "type": "text", + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 568, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 623 + ], + "type": "text", + "content": "[25] Oscar Koller, Necati Cihan Camgoz, Hermann Ney, and Richard Bowden. Weakly supervised learning with multistream cnn-lstm-hmms to discover sequential parallelism in sign language videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "type": "text", + "content": "[26] Oscar Koller, Hermann Ney, and Richard Bowden. Deep hand: How to train a CNN on 1 million hand images when your data is continuous and weakly labelled. In IEEE Conference on Computer Vision and Pattern Recognition, 2016. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[27] Dongxu Li, Chenchen Xu, Xin Yu, Kaihao Zhang, Benjamin Swift, Hanna Suominen, and Hongdong Li. TSPNet: Hierarchical feature learning via temporal semantic pyramid for" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15600" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "sign language translation. Advances in Neural Information Processing Systems, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[28] Huan Ling, Jun Gao, Amlan Kar, Wenzheng Chen, and Sanja Fidler. Fast interactive object annotation with Curve-GCN. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "text", + "content": "[29] Zichen Liu, Jun Hao Liew, Xiangyu Chen, and Jiashi Feng. DANCE: A deep attentive contour model for efficient instance segmentation. In IEEE Winter Conference on Applications of Computer Vision, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 218 + ], + "type": "text", + "content": "[30] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. arXiv preprint arXiv:1508.04025, 2015. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "text", + "content": "[31] Yuecong Min, Aiming Hao, Xiujuan Chai, and Xilin Chen. Visual alignment constraint for continuous sign language recognition. In IEEE International Conference on Computer Vision, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "type": "text", + "content": "[32] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: a method for automatic evaluation of machine translation. In Annual Meeting of the Association for Computational Linguistics, 2002. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "type": "text", + "content": "[33] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 355, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 398 + ], + "type": "text", + "content": "[34] Sida Peng, Wen Jiang, Huaijin Pi, Xiuli Li, Hujun Bao, and Xiaowei Zhou. Deep snake for real-time instance segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "type": "text", + "content": "[35] Junfu Pu, Wengang Zhou, and Houqiang Li. Dilated convolutional network with iterative optimization for continuous sign language recognition. In International Joint Conference on Artificial Intelligence, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "type": "text", + "content": "[36] Junfu Pu, Wengang Zhou, and Houqiang Li. Iterative alignment network for continuous sign language recognition. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 489, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 533 + ], + "type": "text", + "content": "[37] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive Image Deraining Networks: A better and simpler baseline. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 534, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 566 + ], + "type": "text", + "content": "[38] Lin CY ROUGE. A package for automatic evaluation of summaries. In Workshop on Text Summarization of ACL, 2004. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 568, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 601 + ], + "type": "text", + "content": "[39] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. Advances in Neural Information Processing Systems, 2014. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 602, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 634 + ], + "type": "text", + "content": "[40] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. Advances in Neural Information Processing Systems, 2014. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 636, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 667 + ], + "type": "text", + "content": "[41] Rachel Sutton-Spence and Bencie Woll. The linguistics of British Sign Language: an introduction. Cambridge University Press, 1999. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[42] Shengeng Tang, Dan Guo, Richang Hong, and Meng Wang. Graph-based multimodal sequential embedding for sign language translation. IEEE Transactions on Multimedia, 2021. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 497 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[44] Hanjie Wang, Xiujuan Chai, Xiaopeng Hong, Guoying Zhao, and Xilin Chen. Isolated sign language recognition with grassmann covariance matrices. ACM Transactions on Accessible Computing, 2016. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "type": "text", + "content": "[45] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google's neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144, 2016. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 230, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 261 + ], + "type": "text", + "content": "[46] Kayo Yin and Jesse Read. Better sign language translation with stmc-transformer. arXiv preprint arXiv:2004.00588, 2020.2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 263, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 545, + 307 + ], + "type": "text", + "content": "[47] Tao Zhang, Shiqing Wei, and Shunping Ji. E2EC: An end-to-end contour-based method for high-quality high-speed instance segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 308, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 545, + 352 + ], + "type": "text", + "content": "[48] Hao Zhou, Wengang Zhou, and Houqiang Li. Dynamic pseudo label decoding for continuous sign language recognition. In IEEE International conference on Multimedia and Expo, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 354, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 407 + ], + "type": "text", + "content": "[49] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 409, + 545, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 409, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 308, + 409, + 545, + 452 + ], + "type": "text", + "content": "[50] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for sign language recognition and translation. IEEE Transactions on Multimedia, 2021. 2, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 454, + 545, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 454, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 308, + 454, + 545, + 497 + ], + "type": "text", + "content": "[51] Jinhua Zhu, Yingce Xia, Lijun Wu, Di He, Tao Qin, Wengang Zhou, Houqiang Li, and Tie-Yan Liu. Incorporating bert into neural machine translation. arXiv preprint arXiv:2002.06823, 2020. 4, 6, 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15601" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SimFIR_ A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning/e93fe221-1f98-47dd-bbe6-409ea24f23b2_content_list.json b/2023/SimFIR_ A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning/e93fe221-1f98-47dd-bbe6-409ea24f23b2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..64b64ba1ed0fc101eb135b7a401b2f134c29244e --- /dev/null +++ b/2023/SimFIR_ A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning/e93fe221-1f98-47dd-bbe6-409ea24f23b2_content_list.json @@ -0,0 +1,1953 @@ +[ + { + "type": "text", + "text": "SimFIR: A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning", + "text_level": 1, + "bbox": [ + 143, + 130, + 826, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Feng $^{1,2}$ Wendi Wang $^{1}$ Jiajun Deng $^{3}$ Wengang Zhou $^{1,4,*}$ Li Li $^{1}$ Houqiang Li $^{1,4,*}$", + "bbox": [ + 109, + 202, + 862, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ CAS Key Laboratory of Technology in GIPAS, EEIS Department, University of Science and Technology of China", + "bbox": [ + 104, + 222, + 864, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Zhangjiang Laboratory, Shanghai, China $^{3}$ The University of Sydney", + "bbox": [ + 241, + 239, + 728, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Institute of Artificial Intelligence, Hefei Comprehensive National Science Center", + "bbox": [ + 210, + 257, + 756, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{haof,wendiwang}@mail.ustc.edu.cn，jiajun.deng@sydney.edu.au，{zhwg,lil1,lihq}@ustc.edu.cn", + "bbox": [ + 88, + 276, + 880, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In fisheye images, rich distinct distortion patterns are regularly distributed in the image plane. These distortion patterns are independent of the visual content and provide informative cues for rectification. To make the best of such rectification cues, we introduce SimFIR, a simple framework for fisheye image rectification based on self-supervised representation learning. Technically, we first split a fisheye image into multiple patches and extract their representations with a Vision Transformer (ViT). To learn fine-grained distortion representations, we then associate different image patches with their specific distortion patterns based on the fisheye model, and further subtly design an innovative unified distortion-aware pretext task for their learning. The transfer performance on the downstream rectification task is remarkably boosted, which verifies the effectiveness of the learned representations. Extensive experiments are conducted, and the quantitative and qualitative results demonstrate the superiority of our method over the state-of-the-art algorithms as well as its strong generalization ability on real-world fisheye images.", + "bbox": [ + 73, + 358, + 473, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 686, + 209, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fisheye lenses have been widely used in systems for video surveillance [41, 40, 22], autonomous driving [25, 23] and robotic applications [11, 12], thanks to their extremely wide field of view (FoV). However, images captured by the fisheye lenses are inevitably distorted, which are not directly applicable for the downstream analysis, such as object detection [43, 24], pose estimation [62, 1, 45] and scene segmentation [33, 13, 5]. Over the past few years, fisheye image rectification has become an emerging research topic.", + "bbox": [ + 75, + 712, + 468, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the literature, some classical methods [4, 30, 27, 32, 44, 26, 17] estimate epipolar geometry with images", + "bbox": [ + 76, + 849, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6e38eb2ecc873e801239b67b6afc8c0d0a673db477e89c907afe78bec953af75.jpg", + "image_caption": [ + "Figure 1: An illustration of fisheye distortions. The distortions in fisheye images are radially distributed and centrally symmetrical. The distortion degree increases with distance from the optical axis/image center." + ], + "image_footnote": [], + "bbox": [ + 508, + 325, + 888, + 491 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "from different viewpoints. However, multiview images are not always accessible, which limits the real applications of these methods. To overcome this limitation, other classical methods are based on the detection of plumb lines [53, 55, 7, 14, 61, 2]. Nevertheless, taking the detected plumb lines as the basis of fisheye image rectification inevitably causes error accumulation. Recently, deep learning based solutions [46, 6, 60, 38, 36, 58, 37, 39, 57] have become a promising alternative to the traditional methods. There are two directions among deep learning based methods. One direction [46, 6, 60, 57] attempts to predict the distortion parameters to infer pixel-wise displacement for rectification. The other direction [38, 58, 37, 39] directly predicts the rectified images with specifically designed architectures. Although these methods are reported with state-of-the-art results, the intrinsic distortion characteristics of fisheye images are largely ignored, described next.", + "bbox": [ + 496, + 594, + 893, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generally, as illustrated in Fig. 1, the distortions in fisheye images are radially distributed and centrally symmetrical. The distortion degree increases with distance from the", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 47 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors: Wengang Zhou and Houqiang Li", + "bbox": [ + 94, + 886, + 406, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12418", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "image center along the radial axis. It should be noted that here we follow the existing state-of-the-art rectification algorithms [46, 6, 60, 38, 36, 58, 37, 39, 57] to only consider the ideal fisheye images, in which the distortion center locate at the image center. Furthermore, when dividing a fisheye image into patches, the distortion patterns and degrees vary among patches at different distances from the image center. To improve fisheye image rectification, it is beneficial to model these fine-grained distortion patterns that provide cues for the rectification.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Formally, we present SimFIR, a new framework for fisheye image rectification based on self-supervised representation learning [59]. SimFIR introduces an effective way for encoding the fine-grained distortion patterns in fisheye images, by learning the associations between the local image region and their distinct distortion pattern. Specifically, we first split a fisheye image into multiple patches and label them with a position map (see Fig. 1) based on their specific distortion degrees. Then, these image patches are embedded as input tokens of Vision Transformer (ViT) [15] to abstract patch representations. Given these patch representations, we introduce a unified distortion-aware pretext task to optimize the network. On the one hand, we perform contrastive learning with InfoNCE loss [42] to narrow the representation of patches with the same distortion pattern while enlarging the feature distance with others. On the other hand, we devise a classification supervision to predict the patchwise quantified distortion degree, aiming to differentiate the distinct local distortion patterns. After pre-training, the network is capable of extracting the fine-grained distortion representations, regardless of the image textures. Thereafter, the learned distortion representations are transferred to the downstream rectification task.", + "bbox": [ + 75, + 242, + 472, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides, we introduce a new pipeline for fisheye image rectification, which unwarps fisheye images with a dense flow field, i.e., the warping flow, aiming to process images of arbitrary resolution without complex post-processing [36]. Then, to facilitate training, we construct a well-annotated synthesized dataset that covers various scenes and distortion levels. To validate the learned representations, we conduct extensive experiments by pretraining the network with our self-supervised paradigm and then transferring it to the downstream rectification task. The results demonstrate the effectiveness of our method as well as its superiority over existing state-of-the-art algorithms. In addition, we show the strong generalization ability of our method on processing real-world fisheye images.", + "bbox": [ + 75, + 590, + 472, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we make three-fold contributions:", + "bbox": [ + 96, + 801, + 406, + 816 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present SimFIR, a novel self-supervised based framework, to improve fisheye image rectification by extracting its fine-grained distortion patterns.", + "- We introduce a new pipeline for fisheye image recti" + ], + "bbox": [ + 94, + 828, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "fication that unwarps images with backward warping flow, aiming to process images with any resolutions.", + "bbox": [ + 531, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We conduct extensive experiments to verify the merits of our method and report the state-of-the-art performance with our proposed method.", + "bbox": [ + 517, + 132, + 890, + 178 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 191, + 640, + 208 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fisheye image rectification has been widely explored in the computer vision community. In the following, we separately discuss the methods of each group.", + "bbox": [ + 498, + 218, + 890, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rectification Based on Multi-view Calibration. Some traditional methods [4, 27, 32, 44, 29, 51, 49] calibrate fisheye images by finding the corresponding feature points from multi perspectives. Typically, FMC [4] proposes to estimate the epipolar geometry of two views with different distortion factors. AMS [32] explores the minimal solution that calibrates radial distortion with only eight-point correspondences in two images. However, such methods require special chessboards and multiple images are not always accessible, unfriendly to their practical applications.", + "bbox": [ + 496, + 267, + 890, + 419 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rectification Based on Curve Line Detection. The overcome these limitations, line-based camera calibration methods [53, 55, 7, 14, 3, 61, 8] propose to utilize the prior that straight lines in the 3D world must project to straight lines in the image plane. Typically, SC [48] first detects the longest distorted lines within an image by applying Hough transformation [16] and then estimates the distortion parameters from these lines. However, the curve detection is not robust and challenging, limiting their performance.", + "bbox": [ + 496, + 422, + 890, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rectification Based on Parameter Estimation. By predicting the distortion parameters of fisheye models, some methods compute the per-pixel coordinate displacement to perform rectification. RLDC [46] first employs a convolutional neural network to predict the distortion parameters. DeepCalib [6] predicts the distortion parameters and the focal length with a network. Given a single image, Blind [36] devises a unified framework to correct different types of geometric distortion by the estimation of the distortion parameter and forward warping flow [56]. LCSL [57] imposes an explicit geometry constraint onto the processes of the fisheye lens calibration and rectification.", + "bbox": [ + 496, + 564, + 890, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rectification Based on Image Estimation. Methods following this direction commonly attempt to directly predict the rectified images. DR-GAN [38] exploits a generative adversarial network (GAN) [34] for rectified image synthesis, and proposes a low-to-high perceptual loss to capture the mapping relation between different structural images. FE-GAN [9] learns pixel-level distortion flow from sets of fisheye distorted images and distortion-free ones with the cross-rotation and intra-warping consistency. SIR [18] devises a self-supervised learning idea for recti", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12419", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/633c14111b2496382355018af55e252138f30ac344882cda829189842819fbda.jpg", + "image_caption": [ + "(a) Stage 1: Self-supervised pre-training for fine-gained distortion learning" + ], + "image_footnote": [], + "bbox": [ + 143, + 88, + 813, + 188 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a6bdc4b49e9fac12dd742cb3d6cc3f500654e9ba8455e4bb156a97316217c01b.jpg", + "image_caption": [ + "(b) Stage 2: Fine-tuning for distortion rectification" + ], + "image_footnote": [], + "bbox": [ + 143, + 222, + 830, + 348 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6642815caaf99c5107fad157922149033b12cc536f55c594b1f3246169f16711.jpg", + "image_caption": [ + "Figure 2: Framework of our method for fisheye image rectification. It consists of two stages: (a) self-supervised pre-training that learns the fine-grained distortion representation of fisheye image $\\mathbf{I}_d$ with a unified distortion-aware pretext task; (b) fine-tuning for rectification which uses the learned representation to reconstruct the rectified image $\\mathbf{I}_r$ with a flow field. \"BRM\" and \"FPM\" denote the proposed boundary refinement module and flow prediction module, respectively." + ], + "image_footnote": [], + "bbox": [ + 153, + 381, + 813, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fication by exploiting intra-model and inter-model consistency. DDM [39] proposes a distortion distribution map that intuitively indicates the global distortion features of a distorted image. Recently, MLC [37] introduces a multilevel curriculum to train the rectification model hierarchically. PCN [58] develops an encoder-decoder architecture to generate multi-scale rectified images progressively.", + "bbox": [ + 75, + 506, + 470, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Fisheye Model and Dataset", + "text_level": 1, + "bbox": [ + 76, + 628, + 328, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first introduce the distortion model for fisheye images. Since obtaining ground truth is difficult for real fisheye images, we follow the common practices in the literature [60, 57, 6, 39, 37, 58] to exploit synthetic data for training. The polynomial model [30] of a fisheye image is formulated as follows,", + "bbox": [ + 75, + 654, + 468, + 744 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {c} = \\sum_ {i = 1} ^ {N _ {k}} \\lambda_ {i} \\theta_ {d} ^ {2 i - 1}, N _ {k} = 1, 2, 3, 4, \\dots , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 757, + 468, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta_{c}$ denotes the angle of incident light, $\\theta_{d}$ is the angle of the light after passing through the lens, $\\lambda_{i}$ is the $i^{th}$ distortion parameter, and $N_{k}$ is the number of distortion parameters. According to the projection model for general pinhole cameras, we have $r_c = f\\cdot \\tan \\theta_c\\approx f\\cdot \\theta_c$ , where $f$ is the focal length of a camera, $r_c$ is the $L_{2}$ distance between", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the image center and an arbitrary point $P$ in a distortion-free image. Hence, the distortion model can be rewritten as,", + "bbox": [ + 498, + 507, + 890, + 537 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nr _ {c} = f \\sum_ {i = 1} ^ {N _ {k}} \\lambda_ {i} \\theta_ {d} ^ {2 i - 1}, N _ {k} = 1, 2, 3, 4, \\dots . \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 551, + 890, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, based on the equidistant projection relation $r_d = f \\cdot \\theta_d$ , where $r_d$ is the $L_2$ distance between the image center and the mapping point $P'$ in the distorted image, we merge $f$ and $\\theta_d$ to obtain the fisheye distortion model as,", + "bbox": [ + 498, + 606, + 892, + 667 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nr _ {c} = \\sum_ {i = 1} ^ {N _ {k}} k _ {i} r _ {d} ^ {2 i - 1}, N _ {k} = 1, 2, 3, 4, \\dots . \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 684, + 890, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It depicts the radial distortion distribution of fisheye image and is used for the dataset generation (see Sec. 5.1).", + "bbox": [ + 498, + 734, + 890, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 500, + 782, + 635, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to learn useful visual representations of fisheye images to improve their rectification. Fig. 2 presents the overview of our method, which consists of two stages: (a) self-supervised pre-training for distortion representation learning, and (b) fine-tuning for distortion rectification. We elaborate on each stage in the following sections.", + "bbox": [ + 496, + 809, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12420", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Self-supervised Pre-training", + "text_level": 1, + "bbox": [ + 76, + 90, + 330, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Eq. (3), fisheye images have an inherent property where the distortions are radially symmetrical, and the degree of distortion in a particular region is dependent on its position on the image plane. A greater radius leads to a higher degree of distortion. To enhance rectification by capturing these detailed distortion patterns, we propose a unified distortion-aware pretext task.", + "bbox": [ + 75, + 113, + 468, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, on the one hand, we use contrastive learning [10] to narrow the patch features with the same distortion degree and push away the different ones. This approach allows us to encode local distortion features while eliminating texture features. On the other hand, we estimate the patch-wise quantified distortion degrees to learn fine-grained distortion features, by introducing a position loss (see Fig. 2). To implement the idea, we introduce a transformer-based model [54], described next.", + "bbox": [ + 75, + 220, + 470, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Patch Embedding. The standard transformer [54] takes a sequence of token embeddings as input. To adapt it, as shown in Fig. 2, given a fisheye image $\\mathbf{I}_d \\in \\mathbb{R}^{H \\times W \\times 3}$ , we first divide it into a sequence of 2D patches $\\mathbf{x}_p \\in \\mathbb{R}^{N \\times (P^2 \\cdot 3)}$ , where $H$ and $W$ are the height and width of the image $\\mathbf{I}_d$ , $P$ is the patch size, and $N = HW / P^2$ denotes the number of patches. Then, we flatten these patches into vectors and map them to $D$ dimensions with a linear projection. We refer to the output projection as the patch embeddings $\\mathbf{E}_o \\in \\mathbb{R}^{N \\times D}$ .", + "bbox": [ + 75, + 359, + 468, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Distortion Encoding Transformer. To retain positional information, positional embeddings are supplemented to the patch embeddings in transformer [54]. In our method, to avoid the simple mapping between the positional embedding and specific distortion pattern, we arbitrarily shuffle the patch embeddings $\\pmb{E}_{o} \\in \\mathbb{R}^{N \\times D}$ along the first dimension (token dimension). Then, positional embeddings [54] (the sine-cosine version) are added to the shuffled patch embeddings. The resulting sequence of embedding vectors are fed into the $N_{T}$ transformer encoder [54] layers. Thereafter, we obtain the abstracted patch representations $\\pmb{E}_{r} \\in \\mathbb{R}^{N \\times D}$ , as shown in Fig. 2.", + "bbox": [ + 75, + 518, + 468, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the meantime, we build a position map $P_{o}\\in \\mathbb{R}^{\\frac{H}{P}\\times \\frac{W}{P}}$ in which each value denotes the distortion degree of a patch. For clarity, we provide an example of the position map, as shown in Fig. 1. Note that it can be readily obtained from the patch radius. Besides, in Table 1, we show the category number $C_t$ of distortion degree, when dividing a $256\\times 256$ image with $P\\times P$ patches. Note that $C_t$ measures the level of granularity of the division on distortion degree. Then, $P_{o}$ is reshaped and shuffled in the same way as the patch embeddings. In this way, we obtain the per-patch distortion degree annotation $P_{r}\\in \\mathbb{R}^{N\\times 1}$ , corresponding to the patch representations $E_{r}\\in \\mathbb{R}^{N\\times D}$ .", + "bbox": [ + 75, + 696, + 468, + 881 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contrastive Loss. To learn a useful distortion represent", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/bb9aa4237d6d6fdb5839e28e4e536b27bc15122851c759cd1f5845af4a82ae8c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
P8×816×1632×3264×64
Ct13636103
", + "bbox": [ + 504, + 97, + 887, + 131 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1: The distortion degree number $C_t$ in the position map when dividing an $256 \\times 256$ image with $P \\times P$ patches.", + "bbox": [ + 498, + 142, + 890, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tation for rectification, we first introduce a contrastive loss to supervise the output representation $\\pmb{E}_r \\in \\mathbb{R}^{N \\times D}$ . The motivation of contrastive learning [10] is to associate the query with its positive examples and push away the negative examples. Considering that the distortion pattern of fisheye images is radially distributed (see Fig. 1) and the distortion degree increases with the distance from the image center, we regard the patches with the same distortion degree as positive examples and with different degrees as negative examples, respectively. In this way, the contrastive learning can help the network encode the local distinct distortion pattern in different image regions, regardless of their individual appearance. Concretely, the cross-entropy loss is calculated, representing the probability of the positive examples being selected over the negatives. For the $i^{th}(i = 1,2,\\dots ,N)$ patch of the input image, the loss function is defined as follows,", + "bbox": [ + 496, + 198, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {i} = - \\log \\frac {\\sum_ {j = 1} ^ {N} \\mathbb {1} _ {[ j \\in N _ {i} ]} \\exp \\left(\\boldsymbol {p} _ {i} \\cdot \\boldsymbol {p} _ {j} / \\tau\\right)}{\\sum_ {k = 1} ^ {N} \\mathbb {1} _ {[ k \\neq i ]} \\exp \\left(\\boldsymbol {p} _ {i} \\cdot \\boldsymbol {p} _ {k} / \\tau\\right)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 464, + 890, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{p}_i\\in \\mathbb{R}^{1\\times D}$ is the $i^{th}$ representation in $\\pmb{E}_{r}\\in \\mathbb{R}^{N\\times D}$ , $N_{i}$ is index set of positive examples for the $i^{th}$ patch, $\\mathbb{1}_{[f(\\cdot)]}\\in \\{0,1\\}$ represents an indicator function evaluating to 1 iff $f(\\cdot)$ , and $\\tau$ denotes a temperature parameter. The total contrastive loss $\\mathcal{L}_{con}$ is calculated over the $N$ patches,", + "bbox": [ + 498, + 515, + 890, + 593 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n}} = \\sum_ {i = 1} ^ {N} \\ell_ {i}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 603, + 890, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Position Loss. As mentioned above, the shuffled patch representations $\\pmb{E}_r \\in \\mathbb{R}^{N \\times D}$ have been labeled with $P_r \\in \\mathbb{R}^{N \\times 1}$ according to their distortion patterns (degrees). To model these local distinct distortion patterns, we further introduce a simple but effective position loss. Specifically, given the patch representation $\\pmb{E}_r \\in \\mathbb{R}^{N \\times D}$ , we apply a layer normalization [54] and a linear projection to estimate their distortion degrees. The loss is defined as a cross-entropy loss,", + "bbox": [ + 498, + 657, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p o s} = - \\sum_ {n = 1} ^ {N} \\sum_ {c = 1} ^ {C _ {t}} \\log \\frac {\\exp \\left(\\hat {\\boldsymbol {x}} _ {n , c}\\right)}{\\sum_ {i = 1} ^ {C _ {t}} \\exp \\left(\\hat {\\boldsymbol {x}} _ {n , i}\\right)} \\boldsymbol {x} _ {n, c}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 804, + 890, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{x}$ and $\\hat{\\pmb{x}}$ denote the ground truth and predicted confidence, respectively; $C_t$ is the category number of distortion degree in $P_r$ which measures the level of granularity of the", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12421", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "learned features. Note that this constraint on the network also facilitates the perception of positives and negatives in the above contrastive learning.", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At the pre-training stage, the overall architecture is end-to-end optimized with the following training objectives:", + "bbox": [ + 76, + 136, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p r e} = \\mathcal {L} _ {c o n} + \\mathcal {L} _ {p o s}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 179, + 468, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Fine-tuning for Rectification", + "text_level": 1, + "bbox": [ + 76, + 205, + 331, + 222 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After pre-training, the network is capable of extracting the fine-grained distortion features. Next, as illustrated in Fig. 2, we initialize the rectification network with the pretrained weights and perform fine-tuning.", + "bbox": [ + 75, + 229, + 468, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we adopt the same patch embedding module as the pre-training stage, and directly feed the embeddings $\\pmb{E}_o$ into $N_T$ transformer encoder layers without shuffling. To adapt to the rectification task, we only initialize the first $N_F(N_F < N_T)$ layers with pre-trained weights, and randomly initialize the remaining layers. The obtained patch representation $\\pmb{E}_r$ is then passed through two parallel modules for rectification, described below.", + "bbox": [ + 75, + 291, + 468, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Flow Prediction Module. Different from the existing parameter-based and image-based methods, we unwarp a fisheye image with a backward flow field. Given the representation $\\pmb{E}_r \\in \\mathbb{R}^{N \\times D}$ , we estimate a full scale flow $\\pmb{f}_b \\in \\mathbb{R}^{H \\times W \\times 2}$ . As illustrated in Fig. 3, with $\\pmb{f}_b = (\\pmb{f}_x, \\pmb{f}_y)$ , the rectified image $I_c \\in \\mathbb{R}^{H \\times W \\times 3}$ is obtained by the warping operation based on bilinear sampling as follows,", + "bbox": [ + 76, + 415, + 468, + 521 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {I} _ {c} (u, v) = \\boldsymbol {I} _ {d} (\\boldsymbol {f} _ {x} (u, v), \\boldsymbol {f} _ {y} (u, v)), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 534, + 468, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(u,v)$ is the integer pixel coordinate in rectified image $I_{c}$ , and $(\\pmb{f}_x(u,v),\\pmb{f}_y(u,v))$ is the predicted decimal pixel coordinate in distorted image $I_{d}$ .", + "bbox": [ + 75, + 561, + 468, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To this end, we introduce a learnable upsample module. We first reshape $\\pmb{E}_r$ to shape $\\frac{H}{P} \\times \\frac{W}{P} \\times D$ . Then, two convolutional layers (stride 1) are followed and produce a $1 / P$ scale warping flow $\\pmb{f}_m \\in \\mathbb{R}^{\\frac{H}{P} \\times \\frac{W}{P} \\times 2}$ . Next, We use other two convolutional layers to predict a $H / P \\times W / P \\times (P \\times P \\times 9)$ mask and perform softmax over the weights of the $3 \\times 3$ neighborhood of each pixel in $\\pmb{f}_m$ . Finally, the obtained $\\frac{H}{P} \\times \\frac{W}{P} \\times P \\times P \\times 2$ map is permuted and reshaped to the full resolution warping flow $\\pmb{f}_b \\in \\mathbb{R}^{H \\times W \\times 2}$ .", + "bbox": [ + 75, + 608, + 468, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Boundary Refinement Module. We find that the obtained $I_{c}$ struggles with the serrated boundaries and noisy backgrounds (see Fig. 7). To further improve the image quality, we propose a boundary refinement module (BRM).", + "bbox": [ + 75, + 750, + 468, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to the flow prediction module, BRM also estimates the weights of neighbor pixels for upsampling. Then, we obtain a confidence map $M \\in \\mathbb{R}^{H \\times W \\times 1}$ , in which each value denotes the confidence whether a pixel belongs to the rectified region. $M$ is further binarized with a threshold $\\sigma$ to obtain the binary boundary refinement mask $M_{b} \\in$", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/07f6194413800740c8e9b70a811cd1e5077fd904f945801add920b52669bfd4c.jpg", + "image_caption": [ + "Figure 3: Visualization of the rectification process of a certain pixel on the rectified image based on warping flow." + ], + "image_footnote": [], + "bbox": [ + 540, + 87, + 854, + 178 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbb{R}^{H\\times W\\times 1}$ , as illustrated in Fig. 2. The final rectified image $\\pmb{I}_r$ can be obtained by multiplying $\\pmb{I}_c$ with $M_b$ .", + "bbox": [ + 500, + 227, + 888, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Loss. At the fine-tuning stage, the model is end-to-end optimized with the following training objectives,", + "bbox": [ + 500, + 261, + 888, + 291 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f t} = \\mathcal {L} _ {\\text {f l o w}} + \\mathcal {L} _ {\\text {m a s k}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 300, + 888, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{flow}$ is the $L_{1}$ distance between the predicted warping flow $\\pmb{f}_b$ and its given ground truth $\\pmb{f}_{gt}$ as follows,", + "bbox": [ + 500, + 323, + 888, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f l o w}} = \\left\\| \\boldsymbol {f} _ {b} \\times \\boldsymbol {M} _ {g t} - \\boldsymbol {f} _ {g t} \\right\\| _ {1}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 362, + 888, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_{gt}$ denotes the ground truth valid foreground region of the rectified image.", + "bbox": [ + 498, + 385, + 888, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{mask}$ is the cross-entropy loss between the predicted boundary refinement confidence map $M$ and its given ground truth $M_{gt}$ as follows,", + "bbox": [ + 498, + 416, + 888, + 460 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = - \\sum_ {i = 1} ^ {N _ {d}} \\left[ \\boldsymbol {y} _ {i} \\log \\left(\\hat {\\boldsymbol {y}} _ {i}\\right) + \\left(1 - \\boldsymbol {y} _ {i}\\right) \\log \\left(1 - \\hat {\\boldsymbol {y}} _ {i}\\right) \\right], \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 470, + 888, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N_{d} = H\\times W$ is the number of pixels in image $\\pmb{I}_d$ , $\\pmb{y}_i\\in \\{0,1\\}$ and $\\hat{\\pmb{y}}_i\\in [0,1]$ denote the ground-truth and the predicted confidence value in $M_{gt}$ and $M$ , respectively.", + "bbox": [ + 498, + 523, + 888, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 580, + 632, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Dataset and Metrics", + "text_level": 1, + "bbox": [ + 500, + 604, + 689, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train and evaluate our method according to the conventions in the literature. Following method [46, 6, 58, 37, 38], we establish a dataset using original images from the Place365 dataset [63]. We use the polynomial model (Eq. (3)) to generate the dataset and the first four distortion parameters are used ( $N_{k} = 4$ ), which is enough to fit most of the real-world application scenarios. The synthetic dataset contains multiple annotations for each distorted fisheye image, including distortion-free image, warping flow, and distortion parameters. We synthesize $100\\mathrm{k}$ images for training and $5\\mathrm{k}$ images for evaluation, respectively.", + "bbox": [ + 498, + 628, + 890, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In fisheye image rectification, Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM) are the two most popular evaluation metrics in existing works. Given a rectified image and a reference distortion-free one, PSNR measures the detailed quality differences while SSIM assesses their structural differences. Besides, we use Frechet Inception Distance (FID) to evaluate the distribution difference.", + "bbox": [ + 498, + 795, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12422", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/981f97b73596b14fa21ea7c0e211bb1056e3241547db5ccca90f887fbc910c90.jpg", + "table_caption": [ + "(a) Pre-training. Joint pre-training obtains remarkable performance gain." + ], + "table_footnote": [], + "table_body": "
LconLposPSNRSSIM
20.330.82
✓20.700.83
✓21.300.84
✓✓22.470.86
", + "bbox": [ + 96, + 121, + 349, + 195 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/67298285341caf53a37e71f533d0835c7d41c314a3ad5ddca01506af82e26c26.jpg", + "table_caption": [ + "(b) Fine-tuning way. Both fine-tuning strategies work and training from-scratch is better." + ], + "table_footnote": [], + "table_body": "
settingsPSNRSSIM
w/o pre-training20.330.82
w/ pre-training: frozen encoder21.120.85
unfrozen encoder22.470.86
", + "bbox": [ + 364, + 121, + 630, + 195 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d41ce37276a035d7e6334913a3e65f94e0492e7b6e7b9003be717d1cd2c7caf6.jpg", + "table_caption": [ + "(c) Shuffle. Without the shuffle, the performance drops to the baseline level." + ], + "table_footnote": [], + "table_body": "
settingsPSNRSSIM
w/o shuffle20.280.82
w/ shuffle22.470.86
", + "bbox": [ + 645, + 121, + 875, + 167 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3a432a8f9e986b024948722e1d202e47320d9a39b5cb636c0f72428c4f2dddf8.jpg", + "table_caption": [ + "(d) Prediction target. Compared with distortion parameter and image, warping flow is the most effective target." + ], + "table_footnote": [], + "table_body": "
targetsPSNRSSIM
parameter21.090.82
image20.500.69
flow22.470.86
", + "bbox": [ + 125, + 250, + 348, + 308 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(e) Patch size. Reducing patch size works better. To strike a balance between accuracy and efficiency, we choose $16 \\times 16$ by default.", + "bbox": [ + 359, + 209, + 622, + 246 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9e6de42b160e38873f2e54fec4a26f7b5edb6aac745f35db9a519a2d2d826bb7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
patchPSNRSSIMFPS
825.820.9014.93
1622.470.8615.51
3220.230.8114.64
6418.650.7414.19
", + "bbox": [ + 364, + 250, + 624, + 319 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a95959e93fe9d11189ad90f007dcfc81112d6a49bd46f2092206350a7594d933.jpg", + "table_caption": [ + "(f) Architecture setting. Several different architectures are set up. Our scheme performs better." + ], + "table_footnote": [], + "table_body": "
settingsPSNRSSIM
w/o BRM18.010.68
w/o FPM22.110.85
full22.470.86
", + "bbox": [ + 635, + 250, + 857, + 308 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2: Ablation experiments of our method. The default setting is: the patch size is ${16} \\times {16}$ and we import 8 pre-trained layers for fine-tuning. Default settings are all marked in gray .", + "bbox": [ + 75, + 328, + 890, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation", + "text_level": 1, + "bbox": [ + 76, + 377, + 235, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Settings. The image size $(H,W)$ is (256, 256). The latent vector size $D$ in transformer encoder layer [54] is 256 and the total layer number $N_{T}$ is 10. The threshold $\\sigma$ for binarizing the confidence map in the boundary refinement module is empirically set as 0.5.", + "bbox": [ + 75, + 401, + 468, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Details. We use 100k images for pre-training and 50k images for fine-tuning. For contrastive learning, we set the temperature $\\tau = 0.07$ (Eq. (4)). We utilize Adam [31] as the optimizer and one-cycle policy [50, 52] with a maximum learning rate of $10^{-4}$ . Both stages are trained for 65 epochs with a batch size of 64. Two NVIDIA GeForce RTX 2080Ti GPUs are used to train the network.", + "bbox": [ + 75, + 482, + 468, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 599, + 228, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct ablations to verify the effectiveness of the core settings and components in our method.", + "bbox": [ + 75, + 623, + 468, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Impact of Self-supervised Pre-training. The key idea of this work is the self-supervised distortion representation learning for fisheye images. Table 2 (a) shows the studies on the proposed pretext task. Without any pre-training, we can still provide a strong baseline, with a PSNR of 20.33 and a SSIM of 0.82. This can be attributed to the powerful representative ability of transformer [54]. Then, compared with the baseline, the performance improves by pre-training with an independent contrastive loss $\\mathcal{L}_{con}$ or position loss $\\mathcal{L}_{pos}$ . Furthermore, the unified pre-training boosts the rectification performance significantly. These results verify the effectiveness of our self-supervised representation learning strategy. Note that the two pre-text tasks promote each other. In Fig. 4, we provide the training loss plots of fine-tuning under the setting w/ and w/o the pre-training. From the training perspective, we can observe that our pre-training", + "bbox": [ + 75, + 657, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0239745c14468bb5fe721bc3ea80f79ab6da1313d759164724481fbb74cd1610.jpg", + "image_caption": [ + "(a) flow loss" + ], + "image_footnote": [], + "bbox": [ + 509, + 383, + 689, + 489 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0115f2457a924ff0917837b2e1cb0280510c62da158ef65252b6f0fd8cee5721.jpg", + "image_caption": [ + "(b) mask loss", + "Figure 4: Training loss plots during the fine-tuning stage. With the distortion representation pre-training, the training errors of the rectification $(\\mathcal{L}_{flow}$ and $\\mathcal{L}_{mask})$ are reduced." + ], + "image_footnote": [], + "bbox": [ + 699, + 383, + 883, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "effectively reduces the losses of the baseline.", + "bbox": [ + 500, + 580, + 795, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Impact of Fine-tuning Way. During the fine-tuning stage, we train the model from scratch. An alternative way is to freeze the pre-trained encoder and only update the rest weights of the model. As shown in Table 2 (b), both fine-tuning ways produce better performance than the baseline (w/o pre-training), illustrating the effectiveness of our learned representations and our default fine-tuning strategy.", + "bbox": [ + 496, + 601, + 890, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Impact of Shuffle Manipulation. Shuffle strategy is important for our method. During the pre-training stage, we arbitrarily shuffle the patches, to avoid the directed association between the fixed positional embedding and the specific distortion pattern/degree. As shown in Table 2 (c), without the shuffle strategy, the performance degrades to the baseline level in Table 2 (a), which means that the pretraining does not learn any useful representation for the downstream rectification.", + "bbox": [ + 496, + 713, + 890, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Impact of Reconstruction Target. Our SimFIR estimates a warping flow [21, 19] to resample the pixels from distorted images for rectification. In contrast, other", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12423", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f9a603b1f68151d8be30e1952b56d89969d9178ee4adeddb30493762e622fc1e.jpg", + "image_caption": [ + "Distorted" + ], + "image_footnote": [], + "bbox": [ + 83, + 87, + 176, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a6fddc3642451e0de3e927afbcd858f8877a788454d221e6039fe3138d73f4af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 88, + 272, + 160 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c2e28f18320a65836fc77eb0f417a58c7a1d28d2d458fa308938c62de7a36d9.jpg", + "image_caption": [ + "Predict image" + ], + "image_footnote": [], + "bbox": [ + 179, + 162, + 271, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/969255af5d368595bc7389d17911c1c767857834e9f9833044df894c5506ff77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 88, + 367, + 160 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/22ced87002969c48ef0968bac6cc4f58f5be16b370d49f9c6ac47e86bed38b10.jpg", + "image_caption": [ + "Predict flow" + ], + "image_footnote": [], + "bbox": [ + 274, + 162, + 367, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fe00e8adac9797956dc0f5bc2969d2880bb48d5a25d1b933d230cde554297c99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 88, + 465, + 160 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/48e7e78695ec26fc399ec072364272131c4e5b11fa1a3d04611bdec73b5d88aa.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 370, + 162, + 464, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1ac6ee09ac80bf311091a481b9454fb01a74e55b449a8667a2943ba18b608a3f.jpg", + "image_caption": [ + "Figure 5: Examples to illustrate the impact of the reconstruction target. For our method, predicting flow to unwarp the distorted images preserves more image details.", + "Figure 6: Running efficiency comparison between our SimFIR and the state-of-the-art PCN [58] on processing different resolution fisheye images." + ], + "image_footnote": [], + "bbox": [ + 122, + 321, + 419, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pipelines predict the distortion parameters [46, 6, 60, 57] or regress the rectified images directly [39, 58, 38]. We implement the above two pipelines into our method with the following experiments: 1) meaning the output representations $\\pmb{E}_r \\in \\mathbb{R}^{N \\times D}$ into a vector and projecting it to regress the distortion parameters, and 2) reshaping the output $\\pmb{E}_r$ to form a reconstructed image, following MAE [28]. As shown in Table 2 (d), the performance drops with these two targets. This is because for distortion rectification, predicting the deformation instead of the target image itself is simpler [35, 20], and can preserve more image details (see Fig. 5). The network can focus on the spatial transformation between two images, without needing to restore the texture details. Besides, it is difficult to extract complex nonlinear model parameters in different intervals [58].", + "bbox": [ + 75, + 546, + 468, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Patch Size. Table 2 (e) studies the impact of patch size. We can observe that a smaller patch steadily produces a higher performance. The reason is that, as shown in Table 1, a smaller patch setting increases the subdivision of distortion patterns, which helps to extract more fine-grained distortion representations. However, a smaller patch will also introduce more patches when dividing an input image. To balance the computational cost, we choose a $16 \\times 16$", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a058b8b71b3d12852876e30a0c0d4080ff2e7f5aca9712954cead75bcc3ce29d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 87, + 599, + 160 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0bc61729e0d3d8bca1668c505114302f472ca9cec1c7b23b588af4ab44edf9a9.jpg", + "image_caption": [ + "Distorted" + ], + "image_footnote": [], + "bbox": [ + 506, + 162, + 599, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2061970f951d6a84b8f23d5581b9203e88cdae9d01618b67ead7edd2902f2ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 88, + 694, + 160 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/712eefc95a89ff2545eae150374b376884f971520e7c8a5eb64268616d61d52d.jpg", + "image_caption": [ + "w/o BRM", + "Figure 7: Examples to illustrate the impact of the boundary refinement module (abbreviated as \"BRM\"). With the help of BRM, the boundary regions are well rectified." + ], + "image_footnote": [], + "bbox": [ + 601, + 162, + 694, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78c6d3bc826a12e3a30179d6dc0ebcddd61965459e7fc95f96e8e451d515288a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 88, + 790, + 160 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/834b23cc52e9216db515b5e9ae0de3e2136c14f8a0566217440760c7feaee1b2.jpg", + "image_caption": [ + "w/BRM" + ], + "image_footnote": [], + "bbox": [ + 697, + 162, + 790, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/80ab87c44fdfa1d9e464fa09527d3f74dfd74e390c0e24f0e256dfba54a06e13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 88, + 887, + 160 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/27cb2f89cdf5c3f670fd791524e2428c0e3c9d25cd942d87ab164d88aba8fc56.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 792, + 162, + 887, + 234 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/24657aa374c5cadc4938019017e3e09a77aa340d59352396845611e2583b7664.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑FID ↓
SC [48]10.140.32229.60
DeepCalib [6]13.280.51136.73
DR-GAN [38]18.660.7074.56
DDM [39]17.510.66114.66
MLC [37]19.330.7355.97
PCN [58]20.530.7630.64
Ours22.470.8617.23
", + "bbox": [ + 513, + 318, + 879, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3: Comparison between the state-of-the-art methods and the proposed method on the test set. “↑” indicates the higher the better and “↓” means the opposite.", + "bbox": [ + 496, + 450, + 890, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "patch size as the default setting in our final model.", + "bbox": [ + 498, + 513, + 831, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Boundary Refinement Module. In our SimFIR, the BRM is used to smooth the serrated boundaries and remove the noisy backgrounds of the coarse rectified image $I_{c}$ (see Sec. 4.2). As shown in Table 2 (f), with the proposed BRM, performances are remarkably improved. Moreover, we provide two examples to visualize the improvement from the BRM in Fig. 7. As we can see, the boundary regions are well rectified with the help of BRM.", + "bbox": [ + 496, + 532, + 890, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Flow Prediction Module. We compare the bilinear upsampling with our learnable upsampling used for FPM. Table 2 (f) shows that our learnable FPM works better than the module based on the bilinear upsampling. This is likely because the coarse bilinear upsampling operation can not recover the small deformations.", + "bbox": [ + 496, + 656, + 890, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 500, + 756, + 862, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparison. In this section, we compare our SimFIR with the recent state-of-the-art rectification methods, including SC [48], DeepCalib [6], DR-GAN [38], DDM [39], MLC [37], and PCN [58]. The quantitative results are shown in Table 3. As we can see, our method achieves state-of-the-art performance on all metrics. Note that for SC [48], DeepCalib [6], and PCN [58], we calculate the results based on their released codes. For DR-", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12424", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a57ce5860ee7a3020db8c6f4e5a2d39cb4f66c77db246a9dd1fe044a76b7225c.jpg", + "image_caption": [ + "Figure 8: Qualitative results on the synthesized test set. For each comparison, we show the distorted image, the rectified results of SC [48], DeepCalib [6], DR-GAN [38], DDM [39], MLC [37], PCN [58], our method, and ground truth." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 872, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dfb1360f6737ba4235b694bee8ef69d7d51504e14cff0dec84ec488f54fdfc5f.jpg", + "image_caption": [ + "Figure 9: Qualitative results on real-world fisheye images. For each comparison, we show the distorted image, the rectified results of SC [48], Blind [36], DeepCalib [6], DR-GAN [38], DDM [39], MLC [37], PCN [58], and our method." + ], + "image_footnote": [], + "bbox": [ + 94, + 282, + 872, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GAN [38], DDM [39], and MLC [37], we obtain the rectified test set from the authors.", + "bbox": [ + 75, + 563, + 468, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Comparison. To better demonstrate the effectiveness of our proposed method, we further conduct qualitative comparisons with existing methods [48, 6, 38, 39, 37, 58] on the synthetic test set. As present in Fig. 8, the rectified images of our method show less distortion.", + "bbox": [ + 75, + 604, + 468, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Application to Real-world Fisheye Images. Additional experiments are conducted on real-world fisheye images. In this evaluation, we capture the images with several fisheye lenses of different FoV. As shown in Fig. 9, although there exists a gap between synthetic and real-world data, our method still shows strong generalization ability. Compared with SC [48], our method guarantees the integrity of the recovered content. Compared with Blind [36], DeepCalib [6], DR-GAN [38], DDM [39], MLC [37], and PCN [58], our rectified images show less distortion. Note that in Fig. 9, the resolution of the input distorted fisheye images is much higher (up to $2454 \\times 2454$ ) than that in the test set (256 $\\times$ 256). Compared with DR-GAN [38], DDM [39], and PCN [58], our method preserves more texture details.", + "bbox": [ + 75, + 689, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Efficiency Comparison. Next, we compare the running efficiency of our method and the state-of-the-art method PCN [58] on processing images with various resolutions. As shown in Fig. 6, on different resolutions, our method shows higher efficiency. The reason is that PCN [58] directly regresses the rectified image with a fully convolutional network [47], like DR-GAN [38], and DDM [39]. In contrast, our SimFIR infers a warping flow at the fixed resolution $(256 \\times 256)$ and interpolates it to the original high resolution same as the input image for unwarping. Such a strategy ensures efficiency and does not demand any extra training at a certain scale like image-based methods.", + "bbox": [ + 496, + 563, + 890, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 766, + 619, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present an effective self-supervised representation learning paradigm for fisheye images. We associate different image regions with their specific distortion patterns and design a distortion-aware pretext task for their learning. With the pre-training, the network extracts the fine-grained distortion representation. Extensive experiments are conducted on our proposed dataset, and the trans", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12425", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fer performance in the downstream rectification task verifies the effectiveness of the learned representations.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements This work was supported by NSFC under Contract 61836011 and 62021001. It was also supported by the GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, and the Supercomputing Center of the USTC.", + "bbox": [ + 76, + 127, + 468, + 203 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 215, + 173, + 231 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kohei Aso, Dong-Hyun Hwang, and Hideki Koike. Portable 3D human pose estimation for human-human interaction using a chest-mounted fisheye camera. In AHs, pages 116-120, 2021.", + "[2] Joao Barreto, Jose Roquette, Peter Sturm, and Fernando Fonseca. Automatic camera calibration applied to medical endoscopy. In BMVC, pages 1-10, 2009.", + "[3] João Pedro Barreto and Helder Araujo. Geometric properties of central catadioptric line images and their application in calibration. TPAMI, 27(8):1327-1333, 2005.", + "[4] João Pedro Barreto and Kostas Daniilidis. Fundamental matrix for cameras with radial distortion. In ICCV, volume 1, pages 625-632, 2005.", + "[5] Gregor Blott, Masato Takami, and Christian Heipke. Semantic segmentation of fisheye images. In ECCVW, pages 0-0, 2018.", + "[6] Oleksandr Bogdan, Viktor Eckstein, Francois Rameau, and Jean-Charles Bazin. DeepCalib: a deep learning approach for automatic intrinsic calibration of wide field-of-view cameras. In CVMP, pages 1–10, 2018.", + "[7] Faisal Bukhari and Matthew N Dailey. Automatic radial distortion estimation from a single image. Journal of Mathematical Imaging and Vision, 45(1):31-45, 2013.", + "[8] Gyanesh Chander, Brian L Markham, and Dennis L Helder. Summary of current radiometric calibration coefficients for Landsat MSS, TM, ETM+, and EO-1 ALI sensors. Remote Sensing of Environment, 113(5):893–903, 2009.", + "[9] Chun-Hao Chao, Pin-Lun Hsu, Hung-Yi Lee, and Yu-Chiang Frank Wang. Self-supervised deep learning for fisheye image rectification. In ICASSP, pages 2248-2252, 2020.", + "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020.", + "[11] Jonathan Courbon, Youcef Mezouar, Laurent Eckt, and Philippe Martinet. A generic fisheye camera model for robotic applications. In IROS, pages 1683-1688, 2007.", + "[12] Jonathan Courbon, Youcef Mezouar, and Philippe Martinet. Evaluation of the unified model of the sphere for fisheye cameras in robotic applications. Advanced Robotics, 26(8-9):947-967, 2012.", + "[13] Liuyuan Deng, Ming Yang, Yeqiang Qian, Chunxiang Wang, and Bing Wang. CNN based semantic segmentation for urban traffic scenes using fisheye camera. In IV, pages 231-236, 2017.", + "[14] Frederic Devernay and Olivier Faugeras. Straight lines have to be straight. Machine Vision and Applications, 13(1):14-24, 2001." + ], + "bbox": [ + 76, + 241, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.", + "[16] Richard O Duda and Peter E Hart. Use of the hough transformation to detect lines and curves in pictures. CACM, 15(1):11-15, 1972.", + "[17] Jinlong Fan, Jing Zhang, Stephen J Maybank, and Dacheng Tao. Wide-angle image rectification: a survey. IJCV, 130(3):747-776, 2022.", + "[18] Jinlong Fan, Jing Zhang, and Dacheng Tao. SIR: Self-supervised image rectification via seeing the same scene from multiple different lenses. arXiv preprint arXiv:2011.14611, 2020.", + "[19] Hao Feng, Shaokai Liu, Jiajun Deng, Wengang Zhou, and Houqiang Li. Deep unrestricted document image rectification. arXiv preprint arXiv:2304.08796, 2023.", + "[20] Hao Feng, Wengang Zhou, Jiajun Deng, Qi Tian, and Houqiang Li. DocScanner: Robust document image rectification with progressive learning. arXiv preprint arXiv:2110.14968, 2021.", + "[21] Hao Feng, Wengang Zhou, Jiajun Deng, Yuechen Wang, and Houqiang Li. Geometric representation learning for document image rectification. In ECCV, pages 475-492, 2022.", + "[22] James M Ferryman, Stephen J Maybank, and Anthony D Worrall. Visual surveillance for moving vehicles. *IJCV*, 37(2):187-197, 2000.", + "[23] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the KITTI vision benchmark suite. In CVPR, pages 3354-3361, 2012.", + "[24] Payam Goodarzi, Martin Stellmacher, Michael Paetzold, Ahmed Hussein, and Elmar Matthes. Optimization of a CNN-based object detector for fisheye cameras. In ICVES, pages 1-7, 2019.", + "[25] Sorin Grigorescu, Bogdan Trasnea, Tiberiu Cocias, and Gigel Macesanu. A survey of deep learning techniques for autonomous driving. Journal of Field Robotics, 37(3):362-386, 2020.", + "[26] Michael D Grossberg and Shree K Nayar. A general imaging model and a method for finding its parameters. In ICCV, volume 2, pages 108-115, 2001.", + "[27] Richard Hartley and Sing Bing Kang. Parameter-free radial distortion correction with center of distortion estimation. TPAMI, 29(8):1309-1321, 2007.", + "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022.", + "[29] Jose Henrique Brito, Roland Angst, Kevin Koser, and Marc Pollefeys. Radial distortion self-calibration. In CVPR, pages 1368-1375, 2013.", + "[30] Juho Kannala and Sami S Brandt. A generic camera model and calibration method for conventional, wide-angle, and fish-eye lenses. TPAMI, 28(8):1335-1340, 2006.", + "[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014." + ], + "bbox": [ + 501, + 92, + 892, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12426", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Zuzana Kukelova and Tomas Pajdla. A minimal solution to radial distortion autocalibration. TPAMI, 33(12):2410-2422, 2011.", + "[33] Varun Ravi Kumar, Marvin Klingner, Senthil Yogamani, Stefan Milz, Tim Fingscheidt, and Patrick Mader. SyndistNet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In IEEE Winter Conference on Applications of Computer Vision, pages 61-71, 2021.", + "[34] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In CVPR, pages 4681-4690, 2017.", + "[35] Xiaoyu Li, Bo Zhang, Jing Liao, and Pedro V Sander. Document rectification and illumination correction using a patch-based CNN. TOG, 38(6):1-11, 2019.", + "[36] Xiaoyu Li, Bo Zhang, Pedro V Sander, and Jing Liao. Blind geometric distortion correction on images through deep learning. In CVPR, pages 4855-4864, 2019.", + "[37] Kang Liao, Chunyu Lin, Lixin Liao, Yao Zhao, and Weiyao Lin. Multi-level curriculum for training a distortion-aware barrel distortion rectification model. In ICCV, pages 4389-4398, 2021.", + "[38] Kang Liao, Chunyu Lin, Yao Zhao, and Moncef Gabbouj. DR-GAN: Automatic radial distortion rectification using conditional gan in real-time. TCSVT, 30(3):725-733, 2019.", + "[39] Kang Liao, Chunyu Lin, Yao Zhao, and Mai Xu. Model-free distortion rectification framework bridged by distortion distribution map. TIP, 29:3707-3718, 2020.", + "[40] Liang Lin, Yongyi Lu, Yan Pan, and Xiaowu Chen. Integrating graph partitioning and matching for trajectory analysis in video surveillance. TIP, 21(12):4844-4857, 2012.", + "[41] Khan Muhammad, Jamil Ahmad, Zhihan Lv, Paolo Bellavista, Po Yang, and Sung Wook Baik. Efficient deep CNN-based fire detection and localization in video surveillance applications. TSMSC, 49(7):1419-1434, 2018.", + "[42] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018.", + "[43] Elad Plaut, Erez Ben Yaacov, and Bat El Shlomo. 3D object detection from a single fisheye image without a single fisheye training image. In CVPR, pages 3659-3667, 2021.", + "[44] Luis Puig, Yalin Bastanlar, Peter Sturm, José Jesús Guerrero, and Joao Barreto. Calibration of central catadioptric cameras using a dlt-like approach. IJCV, 93(1):101-114, 2011.", + "[45] Helge Rhodin, Christian Richardt, Dan Casas, Eldar Insafutdinov, Mohammad Shafiei, Hans-Peter Seidel, Bernt Schiele, and Christian Theobalt. EgoCap: Egocentric marker-less motion capture with two fisheye cameras. TOG, 35(6):1-11, 2016.", + "[46] Jiangpeng Rong, Shiyao Huang, Zeyu Shang, and Xianghua Ying. Radial lens distortion correction using convolutional neural networks trained with synthesized images. In ACCV, pages 35-49, 2016.", + "[47] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234–241, 2015." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Daniel Santana-Cedrés, Luis Gomez, Miguel Aleman-Flores, Agustín Salgado, Julio Esclarín, Luis Mazorra, and Luis Alvarez. An iterative optimization algorithm for lens distortion correction using two-parameter models. Image Processing On Line, 6:326-364, 2016.", + "[49] Davide Scaramuzza, Agostino Martinelli, and Roland Siegwart. A flexible technique for accurate omnidirectional camera calibration and structure from motion. In ICVS, pages 45-45, 2006.", + "[50] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial Intelligence and Machine Learning for Multi-domain Operations Applications, volume 11006, pages 369-386, 2019.", + "[51] Peter Sturm and Srikumar Ramalingam. A generic concept for camera calibration. In ECCV, pages 1-13, 2004.", + "[52] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419, 2020.", + "[53] Thorsten Thormählen, Hellward Broszio, and Ingolf Wassermann. Robust line-based calibration of lens distortion from a single view. *Mirage*, pages 105–112, 2003.", + "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017.", + "[55] Aiqi Wang, Tianshuang Qiu, and Longtan Shao. A simple method of radial distortion correction with centre of distortion estimation. Journal of Mathematical Imaging and Vision, 35(3):165-172, 2009.", + "[56] Guo-Wang Xie, Fei Yin, Xu-Yao Zhang, and Cheng-Lin Liu. Dewarping document image by displacement flow estimation with fully convolutional network. In DAS, pages 131-144, 2020.", + "[57] Zhucun Xue, Nan Xue, Gui-Song Xia, and Weiming Shen. Learning to calibrate straight lines for fisheye image rectification. In CVPR, pages 1643-1651, 2019.", + "[58] Shangrong Yang, Chunyu Lin, Kang Liao, Chunjie Zhang, and Yao Zhao. Progressively complementary network for fisheye image rectification using appearance flow. In CVPR, pages 6348-6357, 2021.", + "[59] Mingde Yao, Dongliang He, Xin Li, Fu Li, and Zhiwei Xiong. Towards interactive self-supervised denoising. IEEE Transactions on Circuits and Systems for Video Technology, 2023.", + "[60] Xiaoqing Yin, Xinchao Wang, Jun Yu, Maojun Zhang, Pascal Fua, and Dacheng Tao. FisheyeRecNet: A multi-context collaborative deep network for fisheye image rectification. In ECCV, pages 469-484, 2018.", + "[61] Mi Zhang, Jian Yao, Menghan Xia, Kai Li, Yi Zhang, and Yaping Liu. Line-based multi-label energy optimization for fisheye image rectification and calibration. In CVPR, pages 4137-4145, 2015.", + "[62] Yahui Zhang, Shaodi You, and Theo Gevers. Automatic calibration of the fisheye camera for egocentric 3D human pose estimation from a single image. In WACV, pages 1772-1781, 2021.", + "[63] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. TPAMI, 40(6):1452-1464, 2017." + ], + "bbox": [ + 501, + 90, + 890, + 896 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12427", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/SimFIR_ A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning/e93fe221-1f98-47dd-bbe6-409ea24f23b2_model.json b/2023/SimFIR_ A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning/e93fe221-1f98-47dd-bbe6-409ea24f23b2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..74c08493d4416d0957159a6017ee7f01f1856d4d --- /dev/null +++ b/2023/SimFIR_ A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning/e93fe221-1f98-47dd-bbe6-409ea24f23b2_model.json @@ -0,0 +1,2750 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.048 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.145, + 0.131, + 0.828, + 0.177 + ], + "angle": 0, + "content": "SimFIR: A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.203, + 0.864, + 0.223 + ], + "angle": 0, + "content": "Hao Feng\\(^{1,2}\\) Wendi Wang\\(^{1}\\) Jiajun Deng\\(^{3}\\) Wengang Zhou\\(^{1,4,*}\\) Li Li\\(^{1}\\) Houqiang Li\\(^{1,4,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.223, + 0.866, + 0.239 + ], + "angle": 0, + "content": "\\(^{1}\\) CAS Key Laboratory of Technology in GIPAS, EEIS Department, University of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.24, + 0.729, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\) Zhangjiang Laboratory, Shanghai, China \\(^{3}\\) The University of Sydney" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.258, + 0.758, + 0.275 + ], + "angle": 0, + "content": "4 Institute of Artificial Intelligence, Hefei Comprehensive National Science Center" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.277, + 0.882, + 0.294 + ], + "angle": 0, + "content": "{haof,wendiwang}@mail.ustc.edu.cn，jiajun.deng@sydney.edu.au，{zhwg,lil1,lihq}@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.359, + 0.474, + 0.661 + ], + "angle": 0, + "content": "In fisheye images, rich distinct distortion patterns are regularly distributed in the image plane. These distortion patterns are independent of the visual content and provide informative cues for rectification. To make the best of such rectification cues, we introduce SimFIR, a simple framework for fisheye image rectification based on self-supervised representation learning. Technically, we first split a fisheye image into multiple patches and extract their representations with a Vision Transformer (ViT). To learn fine-grained distortion representations, we then associate different image patches with their specific distortion patterns based on the fisheye model, and further subtly design an innovative unified distortion-aware pretext task for their learning. The transfer performance on the downstream rectification task is remarkably boosted, which verifies the effectiveness of the learned representations. Extensive experiments are conducted, and the quantitative and qualitative results demonstrate the superiority of our method over the state-of-the-art algorithms as well as its strong generalization ability on real-world fisheye images." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.688, + 0.21, + 0.704 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.713, + 0.47, + 0.85 + ], + "angle": 0, + "content": "Fisheye lenses have been widely used in systems for video surveillance [41, 40, 22], autonomous driving [25, 23] and robotic applications [11, 12], thanks to their extremely wide field of view (FoV). However, images captured by the fisheye lenses are inevitably distorted, which are not directly applicable for the downstream analysis, such as object detection [43, 24], pose estimation [62, 1, 45] and scene segmentation [33, 13, 5]. Over the past few years, fisheye image rectification has become an emerging research topic." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.85, + 0.47, + 0.88 + ], + "angle": 0, + "content": "In the literature, some classical methods [4, 30, 27, 32, 44, 26, 17] estimate epipolar geometry with images" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.326, + 0.889, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.499, + 0.895, + 0.561 + ], + "angle": 0, + "content": "Figure 1: An illustration of fisheye distortions. The distortions in fisheye images are radially distributed and centrally symmetrical. The distortion degree increases with distance from the optical axis/image center." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.895, + 0.852 + ], + "angle": 0, + "content": "from different viewpoints. However, multiview images are not always accessible, which limits the real applications of these methods. To overcome this limitation, other classical methods are based on the detection of plumb lines [53, 55, 7, 14, 61, 2]. Nevertheless, taking the detected plumb lines as the basis of fisheye image rectification inevitably causes error accumulation. Recently, deep learning based solutions [46, 6, 60, 38, 36, 58, 37, 39, 57] have become a promising alternative to the traditional methods. There are two directions among deep learning based methods. One direction [46, 6, 60, 57] attempts to predict the distortion parameters to infer pixel-wise displacement for rectification. The other direction [38, 58, 37, 39] directly predicts the rectified images with specifically designed architectures. Although these methods are reported with state-of-the-art results, the intrinsic distortion characteristics of fisheye images are largely ignored, described next." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Generally, as illustrated in Fig. 1, the distortions in fisheye images are radially distributed and centrally symmetrical. The distortion degree increases with distance from the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.407, + 0.902 + ], + "angle": 0, + "content": "*Corresponding authors: Wengang Zhou and Houqiang Li" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12418" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "image center along the radial axis. It should be noted that here we follow the existing state-of-the-art rectification algorithms [46, 6, 60, 38, 36, 58, 37, 39, 57] to only consider the ideal fisheye images, in which the distortion center locate at the image center. Furthermore, when dividing a fisheye image into patches, the distortion patterns and degrees vary among patches at different distances from the image center. To improve fisheye image rectification, it is beneficial to model these fine-grained distortion patterns that provide cues for the rectification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.473, + 0.59 + ], + "angle": 0, + "content": "Formally, we present SimFIR, a new framework for fisheye image rectification based on self-supervised representation learning [59]. SimFIR introduces an effective way for encoding the fine-grained distortion patterns in fisheye images, by learning the associations between the local image region and their distinct distortion pattern. Specifically, we first split a fisheye image into multiple patches and label them with a position map (see Fig. 1) based on their specific distortion degrees. Then, these image patches are embedded as input tokens of Vision Transformer (ViT) [15] to abstract patch representations. Given these patch representations, we introduce a unified distortion-aware pretext task to optimize the network. On the one hand, we perform contrastive learning with InfoNCE loss [42] to narrow the representation of patches with the same distortion pattern while enlarging the feature distance with others. On the other hand, we devise a classification supervision to predict the patchwise quantified distortion degree, aiming to differentiate the distinct local distortion patterns. After pre-training, the network is capable of extracting the fine-grained distortion representations, regardless of the image textures. Thereafter, the learned distortion representations are transferred to the downstream rectification task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.473, + 0.802 + ], + "angle": 0, + "content": "Besides, we introduce a new pipeline for fisheye image rectification, which unwarps fisheye images with a dense flow field, i.e., the warping flow, aiming to process images of arbitrary resolution without complex post-processing [36]. Then, to facilitate training, we construct a well-annotated synthesized dataset that covers various scenes and distortion levels. To validate the learned representations, we conduct extensive experiments by pretraining the network with our self-supervised paradigm and then transferring it to the downstream rectification task. The results demonstrate the effectiveness of our method as well as its superiority over existing state-of-the-art algorithms. In addition, we show the strong generalization ability of our method on processing real-world fisheye images." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.803, + 0.408, + 0.817 + ], + "angle": 0, + "content": "In summary, we make three-fold contributions:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.829, + 0.468, + 0.874 + ], + "angle": 0, + "content": "- We present SimFIR, a novel self-supervised based framework, to improve fisheye image rectification by extracting its fine-grained distortion patterns." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- We introduce a new pipeline for fisheye image recti" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.829, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "fication that unwarps images with backward warping flow, aiming to process images with any resolutions." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.133, + 0.892, + 0.179 + ], + "angle": 0, + "content": "- We conduct extensive experiments to verify the merits of our method and report the state-of-the-art performance with our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.193, + 0.642, + 0.209 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.219, + 0.892, + 0.265 + ], + "angle": 0, + "content": "Fisheye image rectification has been widely explored in the computer vision community. In the following, we separately discuss the methods of each group." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.268, + 0.892, + 0.42 + ], + "angle": 0, + "content": "Rectification Based on Multi-view Calibration. Some traditional methods [4, 27, 32, 44, 29, 51, 49] calibrate fisheye images by finding the corresponding feature points from multi perspectives. Typically, FMC [4] proposes to estimate the epipolar geometry of two views with different distortion factors. AMS [32] explores the minimal solution that calibrates radial distortion with only eight-point correspondences in two images. However, such methods require special chessboards and multiple images are not always accessible, unfriendly to their practical applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.56 + ], + "angle": 0, + "content": "Rectification Based on Curve Line Detection. The overcome these limitations, line-based camera calibration methods [53, 55, 7, 14, 3, 61, 8] propose to utilize the prior that straight lines in the 3D world must project to straight lines in the image plane. Typically, SC [48] first detects the longest distorted lines within an image by applying Hough transformation [16] and then estimates the distortion parameters from these lines. However, the curve detection is not robust and challenging, limiting their performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.745 + ], + "angle": 0, + "content": "Rectification Based on Parameter Estimation. By predicting the distortion parameters of fisheye models, some methods compute the per-pixel coordinate displacement to perform rectification. RLDC [46] first employs a convolutional neural network to predict the distortion parameters. DeepCalib [6] predicts the distortion parameters and the focal length with a network. Given a single image, Blind [36] devises a unified framework to correct different types of geometric distortion by the estimation of the distortion parameter and forward warping flow [56]. LCSL [57] imposes an explicit geometry constraint onto the processes of the fisheye lens calibration and rectification." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Rectification Based on Image Estimation. Methods following this direction commonly attempt to directly predict the rectified images. DR-GAN [38] exploits a generative adversarial network (GAN) [34] for rectified image synthesis, and proposes a low-to-high perceptual loss to capture the mapping relation between different structural images. FE-GAN [9] learns pixel-level distortion flow from sets of fisheye distorted images and distortion-free ones with the cross-rotation and intra-warping consistency. SIR [18] devises a self-supervised learning idea for recti" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12419" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.144, + 0.089, + 0.815, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.194, + 0.7, + 0.209 + ], + "angle": 0, + "content": "(a) Stage 1: Self-supervised pre-training for fine-gained distortion learning" + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.223, + 0.831, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.353, + 0.631, + 0.367 + ], + "angle": 0, + "content": "(b) Stage 2: Fine-tuning for distortion rectification" + }, + { + "type": "image", + "bbox": [ + 0.155, + 0.382, + 0.814, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.42, + 0.893, + 0.482 + ], + "angle": 0, + "content": "Figure 2: Framework of our method for fisheye image rectification. It consists of two stages: (a) self-supervised pre-training that learns the fine-grained distortion representation of fisheye image \\( \\mathbf{I}_d \\) with a unified distortion-aware pretext task; (b) fine-tuning for rectification which uses the learned representation to reconstruct the rectified image \\( \\mathbf{I}_r \\) with a flow field. \"BRM\" and \"FPM\" denote the proposed boundary refinement module and flow prediction module, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.471, + 0.614 + ], + "angle": 0, + "content": "fication by exploiting intra-model and inter-model consistency. DDM [39] proposes a distortion distribution map that intuitively indicates the global distortion features of a distorted image. Recently, MLC [37] introduces a multilevel curriculum to train the rectification model hierarchically. PCN [58] develops an encoder-decoder architecture to generate multi-scale rectified images progressively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.629, + 0.329, + 0.646 + ], + "angle": 0, + "content": "3. Fisheye Model and Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.47, + 0.745 + ], + "angle": 0, + "content": "We first introduce the distortion model for fisheye images. Since obtaining ground truth is difficult for real fisheye images, we follow the common practices in the literature [60, 57, 6, 39, 37, 58] to exploit synthetic data for training. The polynomial model [30] of a fisheye image is formulated as follows," + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.758, + 0.47, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\theta_ {c} = \\sum_ {i = 1} ^ {N _ {k}} \\lambda_ {i} \\theta_ {d} ^ {2 i - 1}, N _ {k} = 1, 2, 3, 4, \\dots , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\(\\theta_{c}\\) denotes the angle of incident light, \\(\\theta_{d}\\) is the angle of the light after passing through the lens, \\(\\lambda_{i}\\) is the \\(i^{th}\\) distortion parameter, and \\(N_{k}\\) is the number of distortion parameters. According to the projection model for general pinhole cameras, we have \\(r_c = f\\cdot \\tan \\theta_c\\approx f\\cdot \\theta_c\\) , where \\(f\\) is the focal length of a camera, \\(r_c\\) is the \\(L_{2}\\) distance between" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.508, + 0.892, + 0.539 + ], + "angle": 0, + "content": "the image center and an arbitrary point \\(P\\) in a distortion-free image. Hence, the distortion model can be rewritten as," + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.552, + 0.892, + 0.593 + ], + "angle": 0, + "content": "\\[\nr _ {c} = f \\sum_ {i = 1} ^ {N _ {k}} \\lambda_ {i} \\theta_ {d} ^ {2 i - 1}, N _ {k} = 1, 2, 3, 4, \\dots . \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.607, + 0.893, + 0.669 + ], + "angle": 0, + "content": "Furthermore, based on the equidistant projection relation \\(r_d = f \\cdot \\theta_d\\), where \\(r_d\\) is the \\(L_2\\) distance between the image center and the mapping point \\(P'\\) in the distorted image, we merge \\(f\\) and \\(\\theta_d\\) to obtain the fisheye distortion model as," + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.685, + 0.892, + 0.726 + ], + "angle": 0, + "content": "\\[\nr _ {c} = \\sum_ {i = 1} ^ {N _ {k}} k _ {i} r _ {d} ^ {2 i - 1}, N _ {k} = 1, 2, 3, 4, \\dots . \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.735, + 0.892, + 0.766 + ], + "angle": 0, + "content": "It depicts the radial distortion distribution of fisheye image and is used for the dataset generation (see Sec. 5.1)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.784, + 0.636, + 0.801 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Our goal is to learn useful visual representations of fisheye images to improve their rectification. Fig. 2 presents the overview of our method, which consists of two stages: (a) self-supervised pre-training for distortion representation learning, and (b) fine-tuning for distortion rectification. We elaborate on each stage in the following sections." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12420" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.331, + 0.108 + ], + "angle": 0, + "content": "4.1. Self-supervised Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.22 + ], + "angle": 0, + "content": "As shown in Eq. (3), fisheye images have an inherent property where the distortions are radially symmetrical, and the degree of distortion in a particular region is dependent on its position on the image plane. A greater radius leads to a higher degree of distortion. To enhance rectification by capturing these detailed distortion patterns, we propose a unified distortion-aware pretext task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.221, + 0.471, + 0.356 + ], + "angle": 0, + "content": "Specifically, on the one hand, we use contrastive learning [10] to narrow the patch features with the same distortion degree and push away the different ones. This approach allows us to encode local distortion features while eliminating texture features. On the other hand, we estimate the patch-wise quantified distortion degrees to learn fine-grained distortion features, by introducing a position loss (see Fig. 2). To implement the idea, we introduce a transformer-based model [54], described next." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.361, + 0.47, + 0.514 + ], + "angle": 0, + "content": "Patch Embedding. The standard transformer [54] takes a sequence of token embeddings as input. To adapt it, as shown in Fig. 2, given a fisheye image \\( \\mathbf{I}_d \\in \\mathbb{R}^{H \\times W \\times 3} \\), we first divide it into a sequence of 2D patches \\( \\mathbf{x}_p \\in \\mathbb{R}^{N \\times (P^2 \\cdot 3)} \\), where \\( H \\) and \\( W \\) are the height and width of the image \\( \\mathbf{I}_d \\), \\( P \\) is the patch size, and \\( N = HW / P^2 \\) denotes the number of patches. Then, we flatten these patches into vectors and map them to \\( D \\) dimensions with a linear projection. We refer to the output projection as the patch embeddings \\( \\mathbf{E}_o \\in \\mathbb{R}^{N \\times D} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.47, + 0.699 + ], + "angle": 0, + "content": "Distortion Encoding Transformer. To retain positional information, positional embeddings are supplemented to the patch embeddings in transformer [54]. In our method, to avoid the simple mapping between the positional embedding and specific distortion pattern, we arbitrarily shuffle the patch embeddings \\( \\pmb{E}_{o} \\in \\mathbb{R}^{N \\times D} \\) along the first dimension (token dimension). Then, positional embeddings [54] (the sine-cosine version) are added to the shuffled patch embeddings. The resulting sequence of embedding vectors are fed into the \\( N_{T} \\) transformer encoder [54] layers. Thereafter, we obtain the abstracted patch representations \\( \\pmb{E}_{r} \\in \\mathbb{R}^{N \\times D} \\), as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.698, + 0.47, + 0.882 + ], + "angle": 0, + "content": "In the meantime, we build a position map \\(P_{o}\\in \\mathbb{R}^{\\frac{H}{P}\\times \\frac{W}{P}}\\) in which each value denotes the distortion degree of a patch. For clarity, we provide an example of the position map, as shown in Fig. 1. Note that it can be readily obtained from the patch radius. Besides, in Table 1, we show the category number \\(C_t\\) of distortion degree, when dividing a \\(256\\times 256\\) image with \\(P\\times P\\) patches. Note that \\(C_t\\) measures the level of granularity of the division on distortion degree. Then, \\(P_{o}\\) is reshaped and shuffled in the same way as the patch embeddings. In this way, we obtain the per-patch distortion degree annotation \\(P_{r}\\in \\mathbb{R}^{N\\times 1}\\), corresponding to the patch representations \\(E_{r}\\in \\mathbb{R}^{N\\times D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Contrastive Loss. To learn a useful distortion represent" + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.098, + 0.888, + 0.132 + ], + "angle": 0, + "content": "
P8×816×1632×3264×64
Ct13636103
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.143, + 0.892, + 0.173 + ], + "angle": 0, + "content": "Table 1: The distortion degree number \\( C_t \\) in the position map when dividing an \\( 256 \\times 256 \\) image with \\( P \\times P \\) patches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.892, + 0.456 + ], + "angle": 0, + "content": "tation for rectification, we first introduce a contrastive loss to supervise the output representation \\( \\pmb{E}_r \\in \\mathbb{R}^{N \\times D} \\). The motivation of contrastive learning [10] is to associate the query with its positive examples and push away the negative examples. Considering that the distortion pattern of fisheye images is radially distributed (see Fig. 1) and the distortion degree increases with the distance from the image center, we regard the patches with the same distortion degree as positive examples and with different degrees as negative examples, respectively. In this way, the contrastive learning can help the network encode the local distinct distortion pattern in different image regions, regardless of their individual appearance. Concretely, the cross-entropy loss is calculated, representing the probability of the positive examples being selected over the negatives. For the \\( i^{th}(i = 1,2,\\dots ,N) \\) patch of the input image, the loss function is defined as follows," + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.465, + 0.891, + 0.507 + ], + "angle": 0, + "content": "\\[\n\\ell_ {i} = - \\log \\frac {\\sum_ {j = 1} ^ {N} \\mathbb {1} _ {[ j \\in N _ {i} ]} \\exp \\left(\\boldsymbol {p} _ {i} \\cdot \\boldsymbol {p} _ {j} / \\tau\\right)}{\\sum_ {k = 1} ^ {N} \\mathbb {1} _ {[ k \\neq i ]} \\exp \\left(\\boldsymbol {p} _ {i} \\cdot \\boldsymbol {p} _ {k} / \\tau\\right)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.516, + 0.892, + 0.594 + ], + "angle": 0, + "content": "where \\(\\pmb{p}_i\\in \\mathbb{R}^{1\\times D}\\) is the \\(i^{th}\\) representation in \\(\\pmb{E}_{r}\\in \\mathbb{R}^{N\\times D}\\), \\(N_{i}\\) is index set of positive examples for the \\(i^{th}\\) patch, \\(\\mathbb{1}_{[f(\\cdot)]}\\in \\{0,1\\}\\) represents an indicator function evaluating to 1 iff \\(f(\\cdot)\\), and \\(\\tau\\) denotes a temperature parameter. The total contrastive loss \\(\\mathcal{L}_{con}\\) is calculated over the \\(N\\) patches," + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.604, + 0.891, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n}} = \\sum_ {i = 1} ^ {N} \\ell_ {i}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.658, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Position Loss. As mentioned above, the shuffled patch representations \\( \\pmb{E}_r \\in \\mathbb{R}^{N \\times D} \\) have been labeled with \\( P_r \\in \\mathbb{R}^{N \\times 1} \\) according to their distortion patterns (degrees). To model these local distinct distortion patterns, we further introduce a simple but effective position loss. Specifically, given the patch representation \\( \\pmb{E}_r \\in \\mathbb{R}^{N \\times D} \\), we apply a layer normalization [54] and a linear projection to estimate their distortion degrees. The loss is defined as a cross-entropy loss," + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.805, + 0.891, + 0.846 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p o s} = - \\sum_ {n = 1} ^ {N} \\sum_ {c = 1} ^ {C _ {t}} \\log \\frac {\\exp \\left(\\hat {\\boldsymbol {x}} _ {n , c}\\right)}{\\sum_ {i = 1} ^ {C _ {t}} \\exp \\left(\\hat {\\boldsymbol {x}} _ {n , i}\\right)} \\boldsymbol {x} _ {n, c}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\( \\pmb{x} \\) and \\( \\hat{\\pmb{x}} \\) denote the ground truth and predicted confidence, respectively; \\( C_t \\) is the category number of distortion degree in \\( P_r \\) which measures the level of granularity of the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12421" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "learned features. Note that this constraint on the network also facilitates the perception of positives and negatives in the above contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.137, + 0.469, + 0.168 + ], + "angle": 0, + "content": "At the pre-training stage, the overall architecture is end-to-end optimized with the following training objectives:" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.18, + 0.47, + 0.197 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p r e} = \\mathcal {L} _ {c o n} + \\mathcal {L} _ {p o s}. \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.207, + 0.332, + 0.223 + ], + "angle": 0, + "content": "4.2. Fine-tuning for Rectification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.231, + 0.469, + 0.291 + ], + "angle": 0, + "content": "After pre-training, the network is capable of extracting the fine-grained distortion features. Next, as illustrated in Fig. 2, we initialize the rectification network with the pretrained weights and perform fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.469, + 0.412 + ], + "angle": 0, + "content": "Specifically, we adopt the same patch embedding module as the pre-training stage, and directly feed the embeddings \\( \\pmb{E}_o \\) into \\( N_T \\) transformer encoder layers without shuffling. To adapt to the rectification task, we only initialize the first \\( N_F(N_F < N_T) \\) layers with pre-trained weights, and randomly initialize the remaining layers. The obtained patch representation \\( \\pmb{E}_r \\) is then passed through two parallel modules for rectification, described below." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.416, + 0.469, + 0.522 + ], + "angle": 0, + "content": "Flow Prediction Module. Different from the existing parameter-based and image-based methods, we unwarp a fisheye image with a backward flow field. Given the representation \\( \\pmb{E}_r \\in \\mathbb{R}^{N \\times D} \\), we estimate a full scale flow \\( \\pmb{f}_b \\in \\mathbb{R}^{H \\times W \\times 2} \\). As illustrated in Fig. 3, with \\( \\pmb{f}_b = (\\pmb{f}_x, \\pmb{f}_y) \\), the rectified image \\( I_c \\in \\mathbb{R}^{H \\times W \\times 3} \\) is obtained by the warping operation based on bilinear sampling as follows," + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.535, + 0.469, + 0.552 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {I} _ {c} (u, v) = \\boldsymbol {I} _ {d} (\\boldsymbol {f} _ {x} (u, v), \\boldsymbol {f} _ {y} (u, v)), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.469, + 0.608 + ], + "angle": 0, + "content": "where \\((u,v)\\) is the integer pixel coordinate in rectified image \\(I_{c}\\), and \\((\\pmb{f}_x(u,v),\\pmb{f}_y(u,v))\\) is the predicted decimal pixel coordinate in distorted image \\(I_{d}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.469, + 0.746 + ], + "angle": 0, + "content": "To this end, we introduce a learnable upsample module. We first reshape \\( \\pmb{E}_r \\) to shape \\( \\frac{H}{P} \\times \\frac{W}{P} \\times D \\). Then, two convolutional layers (stride 1) are followed and produce a \\( 1 / P \\) scale warping flow \\( \\pmb{f}_m \\in \\mathbb{R}^{\\frac{H}{P} \\times \\frac{W}{P} \\times 2} \\). Next, We use other two convolutional layers to predict a \\( H / P \\times W / P \\times (P \\times P \\times 9) \\) mask and perform softmax over the weights of the \\( 3 \\times 3 \\) neighborhood of each pixel in \\( \\pmb{f}_m \\). Finally, the obtained \\( \\frac{H}{P} \\times \\frac{W}{P} \\times P \\times P \\times 2 \\) map is permuted and reshaped to the full resolution warping flow \\( \\pmb{f}_b \\in \\mathbb{R}^{H \\times W \\times 2} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.81 + ], + "angle": 0, + "content": "Boundary Refinement Module. We find that the obtained \\( I_{c} \\) struggles with the serrated boundaries and noisy backgrounds (see Fig. 7). To further improve the image quality, we propose a boundary refinement module (BRM)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Similar to the flow prediction module, BRM also estimates the weights of neighbor pixels for upsampling. Then, we obtain a confidence map \\( M \\in \\mathbb{R}^{H \\times W \\times 1} \\), in which each value denotes the confidence whether a pixel belongs to the rectified region. \\( M \\) is further binarized with a threshold \\( \\sigma \\) to obtain the binary boundary refinement mask \\( M_{b} \\in \\)" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.088, + 0.855, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.184, + 0.89, + 0.214 + ], + "angle": 0, + "content": "Figure 3: Visualization of the rectification process of a certain pixel on the rectified image based on warping flow." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.228, + 0.89, + 0.259 + ], + "angle": 0, + "content": "\\(\\mathbb{R}^{H\\times W\\times 1}\\), as illustrated in Fig. 2. The final rectified image \\(\\pmb{I}_r\\) can be obtained by multiplying \\(\\pmb{I}_c\\) with \\(M_b\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.262, + 0.89, + 0.292 + ], + "angle": 0, + "content": "Training Loss. At the fine-tuning stage, the model is end-to-end optimized with the following training objectives," + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.301, + 0.89, + 0.317 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f t} = \\mathcal {L} _ {\\text {f l o w}} + \\mathcal {L} _ {\\text {m a s k}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.324, + 0.89, + 0.355 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{flow}\\) is the \\(L_{1}\\) distance between the predicted warping flow \\(\\pmb{f}_b\\) and its given ground truth \\(\\pmb{f}_{gt}\\) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.363, + 0.89, + 0.38 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f l o w}} = \\left\\| \\boldsymbol {f} _ {b} \\times \\boldsymbol {M} _ {g t} - \\boldsymbol {f} _ {g t} \\right\\| _ {1}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.386, + 0.89, + 0.416 + ], + "angle": 0, + "content": "where \\(M_{gt}\\) denotes the ground truth valid foreground region of the rectified image." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.417, + 0.89, + 0.462 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{mask}\\) is the cross-entropy loss between the predicted boundary refinement confidence map \\(M\\) and its given ground truth \\(M_{gt}\\) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.471, + 0.89, + 0.524 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = - \\sum_ {i = 1} ^ {N _ {d}} \\left[ \\boldsymbol {y} _ {i} \\log \\left(\\hat {\\boldsymbol {y}} _ {i}\\right) + \\left(1 - \\boldsymbol {y} _ {i}\\right) \\log \\left(1 - \\hat {\\boldsymbol {y}} _ {i}\\right) \\right], \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.524, + 0.89, + 0.57 + ], + "angle": 0, + "content": "where \\(N_{d} = H\\times W\\) is the number of pixels in image \\(\\pmb{I}_d\\), \\(\\pmb{y}_i\\in \\{0,1\\}\\) and \\(\\hat{\\pmb{y}}_i\\in [0,1]\\) denote the ground-truth and the predicted confidence value in \\(M_{gt}\\) and \\(M\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.581, + 0.633, + 0.598 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.691, + 0.621 + ], + "angle": 0, + "content": "5.1. Dataset and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.629, + 0.892, + 0.795 + ], + "angle": 0, + "content": "We train and evaluate our method according to the conventions in the literature. Following method [46, 6, 58, 37, 38], we establish a dataset using original images from the Place365 dataset [63]. We use the polynomial model (Eq. (3)) to generate the dataset and the first four distortion parameters are used (\\(N_{k} = 4\\)), which is enough to fit most of the real-world application scenarios. The synthetic dataset contains multiple annotations for each distorted fisheye image, including distortion-free image, warping flow, and distortion parameters. We synthesize \\(100\\mathrm{k}\\) images for training and \\(5\\mathrm{k}\\) images for evaluation, respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.796, + 0.892, + 0.9 + ], + "angle": 0, + "content": "In fisheye image rectification, Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM) are the two most popular evaluation metrics in existing works. Given a rectified image and a reference distortion-free one, PSNR measures the detailed quality differences while SSIM assesses their structural differences. Besides, we use Frechet Inception Distance (FID) to evaluate the distribution difference." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12422" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.094, + 0.348, + 0.119 + ], + "angle": 0, + "content": "(a) Pre-training. Joint pre-training obtains remarkable performance gain." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.122, + 0.35, + 0.196 + ], + "angle": 0, + "content": "
LconLposPSNRSSIM
20.330.82
✓20.700.83
✓21.300.84
✓✓22.470.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.362, + 0.094, + 0.631, + 0.118 + ], + "angle": 0, + "content": "(b) Fine-tuning way. Both fine-tuning strategies work and training from-scratch is better." + }, + { + "type": "table", + "bbox": [ + 0.366, + 0.122, + 0.631, + 0.196 + ], + "angle": 0, + "content": "
settingsPSNRSSIM
w/o pre-training20.330.82
w/ pre-training: frozen encoder21.120.85
unfrozen encoder22.470.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.645, + 0.094, + 0.877, + 0.118 + ], + "angle": 0, + "content": "(c) Shuffle. Without the shuffle, the performance drops to the baseline level." + }, + { + "type": "table", + "bbox": [ + 0.647, + 0.122, + 0.877, + 0.168 + ], + "angle": 0, + "content": "
settingsPSNRSSIM
w/o shuffle20.280.82
w/ shuffle22.470.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.121, + 0.21, + 0.349, + 0.247 + ], + "angle": 0, + "content": "(d) Prediction target. Compared with distortion parameter and image, warping flow is the most effective target." + }, + { + "type": "table", + "bbox": [ + 0.127, + 0.251, + 0.349, + 0.309 + ], + "angle": 0, + "content": "
targetsPSNRSSIM
parameter21.090.82
image20.500.69
flow22.470.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.361, + 0.21, + 0.624, + 0.247 + ], + "angle": 0, + "content": "(e) Patch size. Reducing patch size works better. To strike a balance between accuracy and efficiency, we choose \\(16 \\times 16\\) by default." + }, + { + "type": "table", + "bbox": [ + 0.366, + 0.251, + 0.625, + 0.32 + ], + "angle": 0, + "content": "
patchPSNRSSIMFPS
825.820.9014.93
1622.470.8615.51
3220.230.8114.64
6418.650.7414.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.633, + 0.212, + 0.859, + 0.248 + ], + "angle": 0, + "content": "(f) Architecture setting. Several different architectures are set up. Our scheme performs better." + }, + { + "type": "table", + "bbox": [ + 0.637, + 0.251, + 0.859, + 0.309 + ], + "angle": 0, + "content": "
settingsPSNRSSIM
w/o BRM18.010.68
w/o FPM22.110.85
full22.470.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.329, + 0.892, + 0.361 + ], + "angle": 0, + "content": "Table 2: Ablation experiments of our method. The default setting is: the patch size is \\( {16} \\times {16} \\) and we import 8 pre-trained layers for fine-tuning. Default settings are all marked in gray ." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.378, + 0.236, + 0.394 + ], + "angle": 0, + "content": "5.2. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.469, + 0.478 + ], + "angle": 0, + "content": "Model Settings. The image size \\((H,W)\\) is (256, 256). The latent vector size \\(D\\) in transformer encoder layer [54] is 256 and the total layer number \\(N_{T}\\) is 10. The threshold \\(\\sigma\\) for binarizing the confidence map in the boundary refinement module is empirically set as 0.5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.469, + 0.589 + ], + "angle": 0, + "content": "Training Details. We use 100k images for pre-training and 50k images for fine-tuning. For contrastive learning, we set the temperature \\(\\tau = 0.07\\) (Eq. (4)). We utilize Adam [31] as the optimizer and one-cycle policy [50, 52] with a maximum learning rate of \\(10^{-4}\\). Both stages are trained for 65 epochs with a batch size of 64. Two NVIDIA GeForce RTX 2080Ti GPUs are used to train the network." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.6, + 0.23, + 0.617 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.469, + 0.655 + ], + "angle": 0, + "content": "We conduct ablations to verify the effectiveness of the core settings and components in our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Impact of Self-supervised Pre-training. The key idea of this work is the self-supervised distortion representation learning for fisheye images. Table 2 (a) shows the studies on the proposed pretext task. Without any pre-training, we can still provide a strong baseline, with a PSNR of 20.33 and a SSIM of 0.82. This can be attributed to the powerful representative ability of transformer [54]. Then, compared with the baseline, the performance improves by pre-training with an independent contrastive loss \\(\\mathcal{L}_{con}\\) or position loss \\(\\mathcal{L}_{pos}\\). Furthermore, the unified pre-training boosts the rectification performance significantly. These results verify the effectiveness of our self-supervised representation learning strategy. Note that the two pre-text tasks promote each other. In Fig. 4, we provide the training loss plots of fine-tuning under the setting w/ and w/o the pre-training. From the training perspective, we can observe that our pre-training" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.384, + 0.691, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.563, + 0.497, + 0.633, + 0.508 + ], + "angle": 0, + "content": "(a) flow loss" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.384, + 0.885, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.752, + 0.497, + 0.827, + 0.508 + ], + "angle": 0, + "content": "(b) mask loss" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.517, + 0.892, + 0.563 + ], + "angle": 0, + "content": "Figure 4: Training loss plots during the fine-tuning stage. With the distortion representation pre-training, the training errors of the rectification \\((\\mathcal{L}_{flow}\\) and \\(\\mathcal{L}_{mask})\\) are reduced." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.582, + 0.797, + 0.596 + ], + "angle": 0, + "content": "effectively reduces the losses of the baseline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.602, + 0.892, + 0.709 + ], + "angle": 0, + "content": "Impact of Fine-tuning Way. During the fine-tuning stage, we train the model from scratch. An alternative way is to freeze the pre-trained encoder and only update the rest weights of the model. As shown in Table 2 (b), both fine-tuning ways produce better performance than the baseline (w/o pre-training), illustrating the effectiveness of our learned representations and our default fine-tuning strategy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.85 + ], + "angle": 0, + "content": "Impact of Shuffle Manipulation. Shuffle strategy is important for our method. During the pre-training stage, we arbitrarily shuffle the patches, to avoid the directed association between the fixed positional embedding and the specific distortion pattern/degree. As shown in Table 2 (c), without the shuffle strategy, the performance degrades to the baseline level in Table 2 (a), which means that the pretraining does not learn any useful representation for the downstream rectification." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Impact of Reconstruction Target. Our SimFIR estimates a warping flow [21, 19] to resample the pixels from distorted images for rectification. In contrast, other" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12423" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.088, + 0.177, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.238, + 0.159, + 0.25 + ], + "angle": 0, + "content": "Distorted" + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.089, + 0.273, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.163, + 0.272, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.238, + 0.269, + 0.251 + ], + "angle": 0, + "content": "Predict image" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.089, + 0.369, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.163, + 0.368, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.286, + 0.238, + 0.361, + 0.251 + ], + "angle": 0, + "content": "Predict flow" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.089, + 0.466, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.163, + 0.465, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.238, + 0.43, + 0.25 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.263, + 0.47, + 0.308 + ], + "angle": 0, + "content": "Figure 5: Examples to illustrate the impact of the reconstruction target. For our method, predicting flow to unwarp the distorted images preserves more image details." + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.323, + 0.421, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.481, + 0.47, + 0.527 + ], + "angle": 0, + "content": "Figure 6: Running efficiency comparison between our SimFIR and the state-of-the-art PCN [58] on processing different resolution fisheye images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.775 + ], + "angle": 0, + "content": "pipelines predict the distortion parameters [46, 6, 60, 57] or regress the rectified images directly [39, 58, 38]. We implement the above two pipelines into our method with the following experiments: 1) meaning the output representations \\( \\pmb{E}_r \\in \\mathbb{R}^{N \\times D} \\) into a vector and projecting it to regress the distortion parameters, and 2) reshaping the output \\( \\pmb{E}_r \\) to form a reconstructed image, following MAE [28]. As shown in Table 2 (d), the performance drops with these two targets. This is because for distortion rectification, predicting the deformation instead of the target image itself is simpler [35, 20], and can preserve more image details (see Fig. 5). The network can focus on the spatial transformation between two images, without needing to restore the texture details. Besides, it is difficult to extract complex nonlinear model parameters in different intervals [58]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Impact of Patch Size. Table 2 (e) studies the impact of patch size. We can observe that a smaller patch steadily produces a higher performance. The reason is that, as shown in Table 1, a smaller patch setting increases the subdivision of distortion patterns, which helps to extract more fine-grained distortion representations. However, a smaller patch will also introduce more patches when dividing an input image. To balance the computational cost, we choose a \\(16 \\times 16\\)" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.088, + 0.6, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.163, + 0.6, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.238, + 0.582, + 0.25 + ], + "angle": 0, + "content": "Distorted" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.089, + 0.696, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.163, + 0.696, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.238, + 0.68, + 0.25 + ], + "angle": 0, + "content": "w/o BRM" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.089, + 0.791, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.163, + 0.791, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.238, + 0.771, + 0.25 + ], + "angle": 0, + "content": "w/BRM" + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.089, + 0.888, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.163, + 0.888, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.831, + 0.238, + 0.851, + 0.25 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.259, + 0.892, + 0.306 + ], + "angle": 0, + "content": "Figure 7: Examples to illustrate the impact of the boundary refinement module (abbreviated as \"BRM\"). With the help of BRM, the boundary regions are well rectified." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.319, + 0.88, + 0.44 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑FID ↓
SC [48]10.140.32229.60
DeepCalib [6]13.280.51136.73
DR-GAN [38]18.660.7074.56
DDM [39]17.510.66114.66
MLC [37]19.330.7355.97
PCN [58]20.530.7630.64
Ours22.470.8617.23
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.451, + 0.892, + 0.497 + ], + "angle": 0, + "content": "Table 3: Comparison between the state-of-the-art methods and the proposed method on the test set. “↑” indicates the higher the better and “↓” means the opposite." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.514, + 0.833, + 0.529 + ], + "angle": 0, + "content": "patch size as the default setting in our final model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.654 + ], + "angle": 0, + "content": "Impact of Boundary Refinement Module. In our SimFIR, the BRM is used to smooth the serrated boundaries and remove the noisy backgrounds of the coarse rectified image \\( I_{c} \\) (see Sec. 4.2). As shown in Table 2 (f), with the proposed BRM, performances are remarkably improved. Moreover, we provide two examples to visualize the improvement from the BRM in Fig. 7. As we can see, the boundary regions are well rectified with the help of BRM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.747 + ], + "angle": 0, + "content": "Impact of Flow Prediction Module. We compare the bilinear upsampling with our learnable upsampling used for FPM. Table 2 (f) shows that our learnable FPM works better than the module based on the bilinear upsampling. This is likely because the coarse bilinear upsampling operation can not recover the small deformations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.864, + 0.773 + ], + "angle": 0, + "content": "5.4. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Quantitative Comparison. In this section, we compare our SimFIR with the recent state-of-the-art rectification methods, including SC [48], DeepCalib [6], DR-GAN [38], DDM [39], MLC [37], and PCN [58]. The quantitative results are shown in Table 3. As we can see, our method achieves state-of-the-art performance on all metrics. Note that for SC [48], DeepCalib [6], and PCN [58], we calculate the results based on their released codes. For DR-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12424" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.087, + 0.874, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.242, + 0.895, + 0.274 + ], + "angle": 0, + "content": "Figure 8: Qualitative results on the synthesized test set. For each comparison, we show the distorted image, the rectified results of SC [48], DeepCalib [6], DR-GAN [38], DDM [39], MLC [37], PCN [58], our method, and ground truth." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.284, + 0.874, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.507, + 0.893, + 0.539 + ], + "angle": 0, + "content": "Figure 9: Qualitative results on real-world fisheye images. For each comparison, we show the distorted image, the rectified results of SC [48], Blind [36], DeepCalib [6], DR-GAN [38], DDM [39], MLC [37], PCN [58], and our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.47, + 0.595 + ], + "angle": 0, + "content": "GAN [38], DDM [39], and MLC [37], we obtain the rectified test set from the authors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.47, + 0.68 + ], + "angle": 0, + "content": "Qualitative Comparison. To better demonstrate the effectiveness of our proposed method, we further conduct qualitative comparisons with existing methods [48, 6, 38, 39, 37, 58] on the synthetic test set. As present in Fig. 8, the rectified images of our method show less distortion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Application to Real-world Fisheye Images. Additional experiments are conducted on real-world fisheye images. In this evaluation, we capture the images with several fisheye lenses of different FoV. As shown in Fig. 9, although there exists a gap between synthetic and real-world data, our method still shows strong generalization ability. Compared with SC [48], our method guarantees the integrity of the recovered content. Compared with Blind [36], DeepCalib [6], DR-GAN [38], DDM [39], MLC [37], and PCN [58], our rectified images show less distortion. Note that in Fig. 9, the resolution of the input distorted fisheye images is much higher (up to \\(2454 \\times 2454\\)) than that in the test set (256 \\(\\times\\) 256). Compared with DR-GAN [38], DDM [39], and PCN [58], our method preserves more texture details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.892, + 0.746 + ], + "angle": 0, + "content": "Efficiency Comparison. Next, we compare the running efficiency of our method and the state-of-the-art method PCN [58] on processing images with various resolutions. As shown in Fig. 6, on different resolutions, our method shows higher efficiency. The reason is that PCN [58] directly regresses the rectified image with a fully convolutional network [47], like DR-GAN [38], and DDM [39]. In contrast, our SimFIR infers a warping flow at the fixed resolution \\((256 \\times 256)\\) and interpolates it to the original high resolution same as the input image for unwarping. Such a strategy ensures efficiency and does not demand any extra training at a certain scale like image-based methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.767, + 0.62, + 0.783 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this work, we present an effective self-supervised representation learning paradigm for fisheye images. We associate different image regions with their specific distortion patterns and design a distortion-aware pretext task for their learning. With the pre-training, the network extracts the fine-grained distortion representation. Extensive experiments are conducted on our proposed dataset, and the trans" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12425" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "fer performance in the downstream rectification task verifies the effectiveness of the learned representations." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.128, + 0.47, + 0.204 + ], + "angle": 0, + "content": "Acknowledgements This work was supported by NSFC under Contract 61836011 and 62021001. It was also supported by the GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, and the Supercomputing Center of the USTC." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.216, + 0.174, + 0.232 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.47, + 0.296 + ], + "angle": 0, + "content": "[1] Kohei Aso, Dong-Hyun Hwang, and Hideki Koike. Portable 3D human pose estimation for human-human interaction using a chest-mounted fisheye camera. In AHs, pages 116-120, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.297, + 0.47, + 0.34 + ], + "angle": 0, + "content": "[2] Joao Barreto, Jose Roquette, Peter Sturm, and Fernando Fonseca. Automatic camera calibration applied to medical endoscopy. In BMVC, pages 1-10, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.47, + 0.382 + ], + "angle": 0, + "content": "[3] João Pedro Barreto and Helder Araujo. Geometric properties of central catadioptric line images and their application in calibration. TPAMI, 27(8):1327-1333, 2005." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.47, + 0.424 + ], + "angle": 0, + "content": "[4] João Pedro Barreto and Kostas Daniilidis. Fundamental matrix for cameras with radial distortion. In ICCV, volume 1, pages 625-632, 2005." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.47, + 0.465 + ], + "angle": 0, + "content": "[5] Gregor Blott, Masato Takami, and Christian Heipke. Semantic segmentation of fisheye images. In ECCVW, pages 0-0, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.47, + 0.522 + ], + "angle": 0, + "content": "[6] Oleksandr Bogdan, Viktor Eckstein, Francois Rameau, and Jean-Charles Bazin. DeepCalib: a deep learning approach for automatic intrinsic calibration of wide field-of-view cameras. In CVMP, pages 1–10, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[7] Faisal Bukhari and Matthew N Dailey. Automatic radial distortion estimation from a single image. Journal of Mathematical Imaging and Vision, 45(1):31-45, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[8] Gyanesh Chander, Brian L Markham, and Dennis L Helder. Summary of current radiometric calibration coefficients for Landsat MSS, TM, ETM+, and EO-1 ALI sensors. Remote Sensing of Environment, 113(5):893–903, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.47, + 0.663 + ], + "angle": 0, + "content": "[9] Chun-Hao Chao, Pin-Lun Hsu, Hung-Yi Lee, and Yu-Chiang Frank Wang. Self-supervised deep learning for fisheye image rectification. In ICASSP, pages 2248-2252, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.663, + 0.47, + 0.705 + ], + "angle": 0, + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.705, + 0.47, + 0.747 + ], + "angle": 0, + "content": "[11] Jonathan Courbon, Youcef Mezouar, Laurent Eckt, and Philippe Martinet. A generic fisheye camera model for robotic applications. In IROS, pages 1683-1688, 2007." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.747, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[12] Jonathan Courbon, Youcef Mezouar, and Philippe Martinet. Evaluation of the unified model of the sphere for fisheye cameras in robotic applications. Advanced Robotics, 26(8-9):947-967, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.803, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[13] Liuyuan Deng, Ming Yang, Yeqiang Qian, Chunxiang Wang, and Bing Wang. CNN based semantic segmentation for urban traffic scenes using fisheye camera. In IV, pages 231-236, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Frederic Devernay and Olivier Faugeras. Straight lines have to be straight. Machine Vision and Applications, 13(1):14-24, 2001." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.242, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.175 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.176, + 0.893, + 0.217 + ], + "angle": 0, + "content": "[16] Richard O Duda and Peter E Hart. Use of the hough transformation to detect lines and curves in pictures. CACM, 15(1):11-15, 1972." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.893, + 0.259 + ], + "angle": 0, + "content": "[17] Jinlong Fan, Jing Zhang, Stephen J Maybank, and Dacheng Tao. Wide-angle image rectification: a survey. IJCV, 130(3):747-776, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.26, + 0.893, + 0.315 + ], + "angle": 0, + "content": "[18] Jinlong Fan, Jing Zhang, and Dacheng Tao. SIR: Self-supervised image rectification via seeing the same scene from multiple different lenses. arXiv preprint arXiv:2011.14611, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.315, + 0.893, + 0.357 + ], + "angle": 0, + "content": "[19] Hao Feng, Shaokai Liu, Jiajun Deng, Wengang Zhou, and Houqiang Li. Deep unrestricted document image rectification. arXiv preprint arXiv:2304.08796, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.357, + 0.893, + 0.412 + ], + "angle": 0, + "content": "[20] Hao Feng, Wengang Zhou, Jiajun Deng, Qi Tian, and Houqiang Li. DocScanner: Robust document image rectification with progressive learning. arXiv preprint arXiv:2110.14968, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.412, + 0.893, + 0.455 + ], + "angle": 0, + "content": "[21] Hao Feng, Wengang Zhou, Jiajun Deng, Yuechen Wang, and Houqiang Li. Geometric representation learning for document image rectification. In ECCV, pages 475-492, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.893, + 0.496 + ], + "angle": 0, + "content": "[22] James M Ferryman, Stephen J Maybank, and Anthony D Worrall. Visual surveillance for moving vehicles. *IJCV*, 37(2):187-197, 2000." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.496, + 0.893, + 0.538 + ], + "angle": 0, + "content": "[23] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the KITTI vision benchmark suite. In CVPR, pages 3354-3361, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.893, + 0.594 + ], + "angle": 0, + "content": "[24] Payam Goodarzi, Martin Stellmacher, Michael Paetzold, Ahmed Hussein, and Elmar Matthes. Optimization of a CNN-based object detector for fisheye cameras. In ICVES, pages 1-7, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.893, + 0.649 + ], + "angle": 0, + "content": "[25] Sorin Grigorescu, Bogdan Trasnea, Tiberiu Cocias, and Gigel Macesanu. A survey of deep learning techniques for autonomous driving. Journal of Field Robotics, 37(3):362-386, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.893, + 0.692 + ], + "angle": 0, + "content": "[26] Michael D Grossberg and Shree K Nayar. A general imaging model and a method for finding its parameters. In ICCV, volume 2, pages 108-115, 2001." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.893, + 0.734 + ], + "angle": 0, + "content": "[27] Richard Hartley and Sing Bing Kang. Parameter-free radial distortion correction with center of distortion estimation. TPAMI, 29(8):1309-1321, 2007." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.734, + 0.893, + 0.776 + ], + "angle": 0, + "content": "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.893, + 0.817 + ], + "angle": 0, + "content": "[29] Jose Henrique Brito, Roland Angst, Kevin Koser, and Marc Pollefeys. Radial distortion self-calibration. In CVPR, pages 1368-1375, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.893, + 0.859 + ], + "angle": 0, + "content": "[30] Juho Kannala and Sami S Brandt. A generic camera model and calibration method for conventional, wide-angle, and fish-eye lenses. TPAMI, 28(8):1335-1340, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12426" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "[32] Zuzana Kukelova and Tomas Pajdla. A minimal solution to radial distortion autocalibration. TPAMI, 33(12):2410-2422, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.469, + 0.218 + ], + "angle": 0, + "content": "[33] Varun Ravi Kumar, Marvin Klingner, Senthil Yogamani, Stefan Milz, Tim Fingscheidt, and Patrick Mader. SyndistNet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In IEEE Winter Conference on Applications of Computer Vision, pages 61-71, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.469, + 0.287 + ], + "angle": 0, + "content": "[34] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In CVPR, pages 4681-4690, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.287, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[35] Xiaoyu Li, Bo Zhang, Jing Liao, and Pedro V Sander. Document rectification and illumination correction using a patch-based CNN. TOG, 38(6):1-11, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.328, + 0.469, + 0.371 + ], + "angle": 0, + "content": "[36] Xiaoyu Li, Bo Zhang, Pedro V Sander, and Jing Liao. Blind geometric distortion correction on images through deep learning. In CVPR, pages 4855-4864, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.469, + 0.425 + ], + "angle": 0, + "content": "[37] Kang Liao, Chunyu Lin, Lixin Liao, Yao Zhao, and Weiyao Lin. Multi-level curriculum for training a distortion-aware barrel distortion rectification model. In ICCV, pages 4389-4398, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[38] Kang Liao, Chunyu Lin, Yao Zhao, and Moncef Gabbouj. DR-GAN: Automatic radial distortion rectification using conditional gan in real-time. TCSVT, 30(3):725-733, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[39] Kang Liao, Chunyu Lin, Yao Zhao, and Mai Xu. Model-free distortion rectification framework bridged by distortion distribution map. TIP, 29:3707-3718, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.51, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[40] Liang Lin, Yongyi Lu, Yan Pan, and Xiaowu Chen. Integrating graph partitioning and matching for trajectory analysis in video surveillance. TIP, 21(12):4844-4857, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[41] Khan Muhammad, Jamil Ahmad, Zhihan Lv, Paolo Bellavista, Po Yang, and Sung Wook Baik. Efficient deep CNN-based fire detection and localization in video surveillance applications. TSMSC, 49(7):1419-1434, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.65 + ], + "angle": 0, + "content": "[42] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.692 + ], + "angle": 0, + "content": "[43] Elad Plaut, Erez Ben Yaacov, and Bat El Shlomo. 3D object detection from a single fisheye image without a single fisheye training image. In CVPR, pages 3659-3667, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[44] Luis Puig, Yalin Bastanlar, Peter Sturm, José Jesús Guerrero, and Joao Barreto. Calibration of central catadioptric cameras using a dlt-like approach. IJCV, 93(1):101-114, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[45] Helge Rhodin, Christian Richardt, Dan Casas, Eldar Insafutdinov, Mohammad Shafiei, Hans-Peter Seidel, Bernt Schiele, and Christian Theobalt. EgoCap: Egocentric marker-less motion capture with two fisheye cameras. TOG, 35(6):1-11, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[46] Jiangpeng Rong, Shiyao Huang, Zeyu Shang, and Xianghua Ying. Radial lens distortion correction using convolutional neural networks trained with synthesized images. In ACCV, pages 35-49, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[47] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234–241, 2015." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[48] Daniel Santana-Cedrés, Luis Gomez, Miguel Aleman-Flores, Agustín Salgado, Julio Esclarín, Luis Mazorra, and Luis Alvarez. An iterative optimization algorithm for lens distortion correction using two-parameter models. Image Processing On Line, 6:326-364, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.161, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[49] Davide Scaramuzza, Agostino Martinelli, and Roland Siegwart. A flexible technique for accurate omnidirectional camera calibration and structure from motion. In ICVS, pages 45-45, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.217, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[50] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial Intelligence and Machine Learning for Multi-domain Operations Applications, volume 11006, pages 369-386, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.286, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[51] Peter Sturm and Srikumar Ramalingam. A generic concept for camera calibration. In ECCV, pages 1-13, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.314, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[52] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.342, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[53] Thorsten Thormählen, Hellward Broszio, and Ingolf Wassermann. Robust line-based calibration of lens distortion from a single view. *Mirage*, pages 105–112, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.384, + 0.892, + 0.425 + ], + "angle": 0, + "content": "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.425, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[55] Aiqi Wang, Tianshuang Qiu, and Longtan Shao. A simple method of radial distortion correction with centre of distortion estimation. Journal of Mathematical Imaging and Vision, 35(3):165-172, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.481, + 0.892, + 0.535 + ], + "angle": 0, + "content": "[56] Guo-Wang Xie, Fei Yin, Xu-Yao Zhang, and Cheng-Lin Liu. Dewarping document image by displacement flow estimation with fully convolutional network. In DAS, pages 131-144, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.535, + 0.892, + 0.578 + ], + "angle": 0, + "content": "[57] Zhucun Xue, Nan Xue, Gui-Song Xia, and Weiming Shen. Learning to calibrate straight lines for fisheye image rectification. In CVPR, pages 1643-1651, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.578, + 0.892, + 0.633 + ], + "angle": 0, + "content": "[58] Shangrong Yang, Chunyu Lin, Kang Liao, Chunjie Zhang, and Yao Zhao. Progressively complementary network for fisheye image rectification using appearance flow. In CVPR, pages 6348-6357, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[59] Mingde Yao, Dongliang He, Xin Li, Fu Li, and Zhiwei Xiong. Towards interactive self-supervised denoising. IEEE Transactions on Circuits and Systems for Video Technology, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[60] Xiaoqing Yin, Xinchao Wang, Jun Yu, Maojun Zhang, Pascal Fua, and Dacheng Tao. FisheyeRecNet: A multi-context collaborative deep network for fisheye image rectification. In ECCV, pages 469-484, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[61] Mi Zhang, Jian Yao, Menghan Xia, Kai Li, Yi Zhang, and Yaping Liu. Line-based multi-label energy optimization for fisheye image rectification and calibration. In CVPR, pages 4137-4145, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.799, + 0.892, + 0.854 + ], + "angle": 0, + "content": "[62] Yahui Zhang, Shaodi You, and Theo Gevers. Automatic calibration of the fisheye camera for egocentric 3D human pose estimation from a single image. In WACV, pages 1772-1781, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.854, + 0.892, + 0.897 + ], + "angle": 0, + "content": "[63] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. TPAMI, 40(6):1452-1464, 2017." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12427" + } + ] +] \ No newline at end of file diff --git a/2023/SimFIR_ A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning/e93fe221-1f98-47dd-bbe6-409ea24f23b2_origin.pdf b/2023/SimFIR_ A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning/e93fe221-1f98-47dd-bbe6-409ea24f23b2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cde213e12c8afd5ca09d5c0e1e28596f6a277cb3 --- /dev/null +++ b/2023/SimFIR_ A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning/e93fe221-1f98-47dd-bbe6-409ea24f23b2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd6df56c81c1970bc2a3391c74e3e78154a0c1bc818ec41d4c47cce127f93b27 +size 1923728 diff --git a/2023/SimFIR_ A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning/full.md b/2023/SimFIR_ A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..345adcbbd7292cbda61ab6ec1d109d58a55cdb48 --- /dev/null +++ b/2023/SimFIR_ A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning/full.md @@ -0,0 +1,400 @@ +# SimFIR: A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning + +Hao Feng $^{1,2}$ Wendi Wang $^{1}$ Jiajun Deng $^{3}$ Wengang Zhou $^{1,4,*}$ Li Li $^{1}$ Houqiang Li $^{1,4,*}$ + +$^{1}$ CAS Key Laboratory of Technology in GIPAS, EEIS Department, University of Science and Technology of China + +$^{2}$ Zhangjiang Laboratory, Shanghai, China $^{3}$ The University of Sydney + +4 Institute of Artificial Intelligence, Hefei Comprehensive National Science Center + +{haof,wendiwang}@mail.ustc.edu.cn，jiajun.deng@sydney.edu.au，{zhwg,lil1,lihq}@ustc.edu.cn + +# Abstract + +In fisheye images, rich distinct distortion patterns are regularly distributed in the image plane. These distortion patterns are independent of the visual content and provide informative cues for rectification. To make the best of such rectification cues, we introduce SimFIR, a simple framework for fisheye image rectification based on self-supervised representation learning. Technically, we first split a fisheye image into multiple patches and extract their representations with a Vision Transformer (ViT). To learn fine-grained distortion representations, we then associate different image patches with their specific distortion patterns based on the fisheye model, and further subtly design an innovative unified distortion-aware pretext task for their learning. The transfer performance on the downstream rectification task is remarkably boosted, which verifies the effectiveness of the learned representations. Extensive experiments are conducted, and the quantitative and qualitative results demonstrate the superiority of our method over the state-of-the-art algorithms as well as its strong generalization ability on real-world fisheye images. + +# 1. Introduction + +Fisheye lenses have been widely used in systems for video surveillance [41, 40, 22], autonomous driving [25, 23] and robotic applications [11, 12], thanks to their extremely wide field of view (FoV). However, images captured by the fisheye lenses are inevitably distorted, which are not directly applicable for the downstream analysis, such as object detection [43, 24], pose estimation [62, 1, 45] and scene segmentation [33, 13, 5]. Over the past few years, fisheye image rectification has become an emerging research topic. + +In the literature, some classical methods [4, 30, 27, 32, 44, 26, 17] estimate epipolar geometry with images + +![](images/6e38eb2ecc873e801239b67b6afc8c0d0a673db477e89c907afe78bec953af75.jpg) +Figure 1: An illustration of fisheye distortions. The distortions in fisheye images are radially distributed and centrally symmetrical. The distortion degree increases with distance from the optical axis/image center. + +from different viewpoints. However, multiview images are not always accessible, which limits the real applications of these methods. To overcome this limitation, other classical methods are based on the detection of plumb lines [53, 55, 7, 14, 61, 2]. Nevertheless, taking the detected plumb lines as the basis of fisheye image rectification inevitably causes error accumulation. Recently, deep learning based solutions [46, 6, 60, 38, 36, 58, 37, 39, 57] have become a promising alternative to the traditional methods. There are two directions among deep learning based methods. One direction [46, 6, 60, 57] attempts to predict the distortion parameters to infer pixel-wise displacement for rectification. The other direction [38, 58, 37, 39] directly predicts the rectified images with specifically designed architectures. Although these methods are reported with state-of-the-art results, the intrinsic distortion characteristics of fisheye images are largely ignored, described next. + +Generally, as illustrated in Fig. 1, the distortions in fisheye images are radially distributed and centrally symmetrical. The distortion degree increases with distance from the + +image center along the radial axis. It should be noted that here we follow the existing state-of-the-art rectification algorithms [46, 6, 60, 38, 36, 58, 37, 39, 57] to only consider the ideal fisheye images, in which the distortion center locate at the image center. Furthermore, when dividing a fisheye image into patches, the distortion patterns and degrees vary among patches at different distances from the image center. To improve fisheye image rectification, it is beneficial to model these fine-grained distortion patterns that provide cues for the rectification. + +Formally, we present SimFIR, a new framework for fisheye image rectification based on self-supervised representation learning [59]. SimFIR introduces an effective way for encoding the fine-grained distortion patterns in fisheye images, by learning the associations between the local image region and their distinct distortion pattern. Specifically, we first split a fisheye image into multiple patches and label them with a position map (see Fig. 1) based on their specific distortion degrees. Then, these image patches are embedded as input tokens of Vision Transformer (ViT) [15] to abstract patch representations. Given these patch representations, we introduce a unified distortion-aware pretext task to optimize the network. On the one hand, we perform contrastive learning with InfoNCE loss [42] to narrow the representation of patches with the same distortion pattern while enlarging the feature distance with others. On the other hand, we devise a classification supervision to predict the patchwise quantified distortion degree, aiming to differentiate the distinct local distortion patterns. After pre-training, the network is capable of extracting the fine-grained distortion representations, regardless of the image textures. Thereafter, the learned distortion representations are transferred to the downstream rectification task. + +Besides, we introduce a new pipeline for fisheye image rectification, which unwarps fisheye images with a dense flow field, i.e., the warping flow, aiming to process images of arbitrary resolution without complex post-processing [36]. Then, to facilitate training, we construct a well-annotated synthesized dataset that covers various scenes and distortion levels. To validate the learned representations, we conduct extensive experiments by pretraining the network with our self-supervised paradigm and then transferring it to the downstream rectification task. The results demonstrate the effectiveness of our method as well as its superiority over existing state-of-the-art algorithms. In addition, we show the strong generalization ability of our method on processing real-world fisheye images. + +In summary, we make three-fold contributions: + +- We present SimFIR, a novel self-supervised based framework, to improve fisheye image rectification by extracting its fine-grained distortion patterns. +- We introduce a new pipeline for fisheye image recti + +fication that unwarps images with backward warping flow, aiming to process images with any resolutions. + +- We conduct extensive experiments to verify the merits of our method and report the state-of-the-art performance with our proposed method. + +# 2. Related Work + +Fisheye image rectification has been widely explored in the computer vision community. In the following, we separately discuss the methods of each group. + +Rectification Based on Multi-view Calibration. Some traditional methods [4, 27, 32, 44, 29, 51, 49] calibrate fisheye images by finding the corresponding feature points from multi perspectives. Typically, FMC [4] proposes to estimate the epipolar geometry of two views with different distortion factors. AMS [32] explores the minimal solution that calibrates radial distortion with only eight-point correspondences in two images. However, such methods require special chessboards and multiple images are not always accessible, unfriendly to their practical applications. + +Rectification Based on Curve Line Detection. The overcome these limitations, line-based camera calibration methods [53, 55, 7, 14, 3, 61, 8] propose to utilize the prior that straight lines in the 3D world must project to straight lines in the image plane. Typically, SC [48] first detects the longest distorted lines within an image by applying Hough transformation [16] and then estimates the distortion parameters from these lines. However, the curve detection is not robust and challenging, limiting their performance. + +Rectification Based on Parameter Estimation. By predicting the distortion parameters of fisheye models, some methods compute the per-pixel coordinate displacement to perform rectification. RLDC [46] first employs a convolutional neural network to predict the distortion parameters. DeepCalib [6] predicts the distortion parameters and the focal length with a network. Given a single image, Blind [36] devises a unified framework to correct different types of geometric distortion by the estimation of the distortion parameter and forward warping flow [56]. LCSL [57] imposes an explicit geometry constraint onto the processes of the fisheye lens calibration and rectification. + +Rectification Based on Image Estimation. Methods following this direction commonly attempt to directly predict the rectified images. DR-GAN [38] exploits a generative adversarial network (GAN) [34] for rectified image synthesis, and proposes a low-to-high perceptual loss to capture the mapping relation between different structural images. FE-GAN [9] learns pixel-level distortion flow from sets of fisheye distorted images and distortion-free ones with the cross-rotation and intra-warping consistency. SIR [18] devises a self-supervised learning idea for recti + +![](images/633c14111b2496382355018af55e252138f30ac344882cda829189842819fbda.jpg) +(a) Stage 1: Self-supervised pre-training for fine-gained distortion learning + +![](images/a6bdc4b49e9fac12dd742cb3d6cc3f500654e9ba8455e4bb156a97316217c01b.jpg) +(b) Stage 2: Fine-tuning for distortion rectification + +![](images/6642815caaf99c5107fad157922149033b12cc536f55c594b1f3246169f16711.jpg) +Figure 2: Framework of our method for fisheye image rectification. It consists of two stages: (a) self-supervised pre-training that learns the fine-grained distortion representation of fisheye image $\mathbf{I}_d$ with a unified distortion-aware pretext task; (b) fine-tuning for rectification which uses the learned representation to reconstruct the rectified image $\mathbf{I}_r$ with a flow field. "BRM" and "FPM" denote the proposed boundary refinement module and flow prediction module, respectively. + +fication by exploiting intra-model and inter-model consistency. DDM [39] proposes a distortion distribution map that intuitively indicates the global distortion features of a distorted image. Recently, MLC [37] introduces a multilevel curriculum to train the rectification model hierarchically. PCN [58] develops an encoder-decoder architecture to generate multi-scale rectified images progressively. + +# 3. Fisheye Model and Dataset + +We first introduce the distortion model for fisheye images. Since obtaining ground truth is difficult for real fisheye images, we follow the common practices in the literature [60, 57, 6, 39, 37, 58] to exploit synthetic data for training. The polynomial model [30] of a fisheye image is formulated as follows, + +$$ +\theta_ {c} = \sum_ {i = 1} ^ {N _ {k}} \lambda_ {i} \theta_ {d} ^ {2 i - 1}, N _ {k} = 1, 2, 3, 4, \dots , \tag {1} +$$ + +where $\theta_{c}$ denotes the angle of incident light, $\theta_{d}$ is the angle of the light after passing through the lens, $\lambda_{i}$ is the $i^{th}$ distortion parameter, and $N_{k}$ is the number of distortion parameters. According to the projection model for general pinhole cameras, we have $r_c = f\cdot \tan \theta_c\approx f\cdot \theta_c$ , where $f$ is the focal length of a camera, $r_c$ is the $L_{2}$ distance between + +the image center and an arbitrary point $P$ in a distortion-free image. Hence, the distortion model can be rewritten as, + +$$ +r _ {c} = f \sum_ {i = 1} ^ {N _ {k}} \lambda_ {i} \theta_ {d} ^ {2 i - 1}, N _ {k} = 1, 2, 3, 4, \dots . \tag {2} +$$ + +Furthermore, based on the equidistant projection relation $r_d = f \cdot \theta_d$ , where $r_d$ is the $L_2$ distance between the image center and the mapping point $P'$ in the distorted image, we merge $f$ and $\theta_d$ to obtain the fisheye distortion model as, + +$$ +r _ {c} = \sum_ {i = 1} ^ {N _ {k}} k _ {i} r _ {d} ^ {2 i - 1}, N _ {k} = 1, 2, 3, 4, \dots . \tag {3} +$$ + +It depicts the radial distortion distribution of fisheye image and is used for the dataset generation (see Sec. 5.1). + +# 4. Methodology + +Our goal is to learn useful visual representations of fisheye images to improve their rectification. Fig. 2 presents the overview of our method, which consists of two stages: (a) self-supervised pre-training for distortion representation learning, and (b) fine-tuning for distortion rectification. We elaborate on each stage in the following sections. + +# 4.1. Self-supervised Pre-training + +As shown in Eq. (3), fisheye images have an inherent property where the distortions are radially symmetrical, and the degree of distortion in a particular region is dependent on its position on the image plane. A greater radius leads to a higher degree of distortion. To enhance rectification by capturing these detailed distortion patterns, we propose a unified distortion-aware pretext task. + +Specifically, on the one hand, we use contrastive learning [10] to narrow the patch features with the same distortion degree and push away the different ones. This approach allows us to encode local distortion features while eliminating texture features. On the other hand, we estimate the patch-wise quantified distortion degrees to learn fine-grained distortion features, by introducing a position loss (see Fig. 2). To implement the idea, we introduce a transformer-based model [54], described next. + +Patch Embedding. The standard transformer [54] takes a sequence of token embeddings as input. To adapt it, as shown in Fig. 2, given a fisheye image $\mathbf{I}_d \in \mathbb{R}^{H \times W \times 3}$ , we first divide it into a sequence of 2D patches $\mathbf{x}_p \in \mathbb{R}^{N \times (P^2 \cdot 3)}$ , where $H$ and $W$ are the height and width of the image $\mathbf{I}_d$ , $P$ is the patch size, and $N = HW / P^2$ denotes the number of patches. Then, we flatten these patches into vectors and map them to $D$ dimensions with a linear projection. We refer to the output projection as the patch embeddings $\mathbf{E}_o \in \mathbb{R}^{N \times D}$ . + +Distortion Encoding Transformer. To retain positional information, positional embeddings are supplemented to the patch embeddings in transformer [54]. In our method, to avoid the simple mapping between the positional embedding and specific distortion pattern, we arbitrarily shuffle the patch embeddings $\pmb{E}_{o} \in \mathbb{R}^{N \times D}$ along the first dimension (token dimension). Then, positional embeddings [54] (the sine-cosine version) are added to the shuffled patch embeddings. The resulting sequence of embedding vectors are fed into the $N_{T}$ transformer encoder [54] layers. Thereafter, we obtain the abstracted patch representations $\pmb{E}_{r} \in \mathbb{R}^{N \times D}$ , as shown in Fig. 2. + +In the meantime, we build a position map $P_{o}\in \mathbb{R}^{\frac{H}{P}\times \frac{W}{P}}$ in which each value denotes the distortion degree of a patch. For clarity, we provide an example of the position map, as shown in Fig. 1. Note that it can be readily obtained from the patch radius. Besides, in Table 1, we show the category number $C_t$ of distortion degree, when dividing a $256\times 256$ image with $P\times P$ patches. Note that $C_t$ measures the level of granularity of the division on distortion degree. Then, $P_{o}$ is reshaped and shuffled in the same way as the patch embeddings. In this way, we obtain the per-patch distortion degree annotation $P_{r}\in \mathbb{R}^{N\times 1}$ , corresponding to the patch representations $E_{r}\in \mathbb{R}^{N\times D}$ . + +Contrastive Loss. To learn a useful distortion represent + +
P8×816×1632×3264×64
Ct13636103
+ +Table 1: The distortion degree number $C_t$ in the position map when dividing an $256 \times 256$ image with $P \times P$ patches. + +tation for rectification, we first introduce a contrastive loss to supervise the output representation $\pmb{E}_r \in \mathbb{R}^{N \times D}$ . The motivation of contrastive learning [10] is to associate the query with its positive examples and push away the negative examples. Considering that the distortion pattern of fisheye images is radially distributed (see Fig. 1) and the distortion degree increases with the distance from the image center, we regard the patches with the same distortion degree as positive examples and with different degrees as negative examples, respectively. In this way, the contrastive learning can help the network encode the local distinct distortion pattern in different image regions, regardless of their individual appearance. Concretely, the cross-entropy loss is calculated, representing the probability of the positive examples being selected over the negatives. For the $i^{th}(i = 1,2,\dots ,N)$ patch of the input image, the loss function is defined as follows, + +$$ +\ell_ {i} = - \log \frac {\sum_ {j = 1} ^ {N} \mathbb {1} _ {[ j \in N _ {i} ]} \exp \left(\boldsymbol {p} _ {i} \cdot \boldsymbol {p} _ {j} / \tau\right)}{\sum_ {k = 1} ^ {N} \mathbb {1} _ {[ k \neq i ]} \exp \left(\boldsymbol {p} _ {i} \cdot \boldsymbol {p} _ {k} / \tau\right)}, \tag {4} +$$ + +where $\pmb{p}_i\in \mathbb{R}^{1\times D}$ is the $i^{th}$ representation in $\pmb{E}_{r}\in \mathbb{R}^{N\times D}$ , $N_{i}$ is index set of positive examples for the $i^{th}$ patch, $\mathbb{1}_{[f(\cdot)]}\in \{0,1\}$ represents an indicator function evaluating to 1 iff $f(\cdot)$ , and $\tau$ denotes a temperature parameter. The total contrastive loss $\mathcal{L}_{con}$ is calculated over the $N$ patches, + +$$ +\mathcal {L} _ {\text {c o n}} = \sum_ {i = 1} ^ {N} \ell_ {i}. \tag {5} +$$ + +Position Loss. As mentioned above, the shuffled patch representations $\pmb{E}_r \in \mathbb{R}^{N \times D}$ have been labeled with $P_r \in \mathbb{R}^{N \times 1}$ according to their distortion patterns (degrees). To model these local distinct distortion patterns, we further introduce a simple but effective position loss. Specifically, given the patch representation $\pmb{E}_r \in \mathbb{R}^{N \times D}$ , we apply a layer normalization [54] and a linear projection to estimate their distortion degrees. The loss is defined as a cross-entropy loss, + +$$ +\mathcal {L} _ {p o s} = - \sum_ {n = 1} ^ {N} \sum_ {c = 1} ^ {C _ {t}} \log \frac {\exp \left(\hat {\boldsymbol {x}} _ {n , c}\right)}{\sum_ {i = 1} ^ {C _ {t}} \exp \left(\hat {\boldsymbol {x}} _ {n , i}\right)} \boldsymbol {x} _ {n, c}, \tag {6} +$$ + +where $\pmb{x}$ and $\hat{\pmb{x}}$ denote the ground truth and predicted confidence, respectively; $C_t$ is the category number of distortion degree in $P_r$ which measures the level of granularity of the + +learned features. Note that this constraint on the network also facilitates the perception of positives and negatives in the above contrastive learning. + +At the pre-training stage, the overall architecture is end-to-end optimized with the following training objectives: + +$$ +\mathcal {L} _ {p r e} = \mathcal {L} _ {c o n} + \mathcal {L} _ {p o s}. \tag {7} +$$ + +# 4.2. Fine-tuning for Rectification + +After pre-training, the network is capable of extracting the fine-grained distortion features. Next, as illustrated in Fig. 2, we initialize the rectification network with the pretrained weights and perform fine-tuning. + +Specifically, we adopt the same patch embedding module as the pre-training stage, and directly feed the embeddings $\pmb{E}_o$ into $N_T$ transformer encoder layers without shuffling. To adapt to the rectification task, we only initialize the first $N_F(N_F < N_T)$ layers with pre-trained weights, and randomly initialize the remaining layers. The obtained patch representation $\pmb{E}_r$ is then passed through two parallel modules for rectification, described below. + +Flow Prediction Module. Different from the existing parameter-based and image-based methods, we unwarp a fisheye image with a backward flow field. Given the representation $\pmb{E}_r \in \mathbb{R}^{N \times D}$ , we estimate a full scale flow $\pmb{f}_b \in \mathbb{R}^{H \times W \times 2}$ . As illustrated in Fig. 3, with $\pmb{f}_b = (\pmb{f}_x, \pmb{f}_y)$ , the rectified image $I_c \in \mathbb{R}^{H \times W \times 3}$ is obtained by the warping operation based on bilinear sampling as follows, + +$$ +\boldsymbol {I} _ {c} (u, v) = \boldsymbol {I} _ {d} (\boldsymbol {f} _ {x} (u, v), \boldsymbol {f} _ {y} (u, v)), \tag {8} +$$ + +where $(u,v)$ is the integer pixel coordinate in rectified image $I_{c}$ , and $(\pmb{f}_x(u,v),\pmb{f}_y(u,v))$ is the predicted decimal pixel coordinate in distorted image $I_{d}$ . + +To this end, we introduce a learnable upsample module. We first reshape $\pmb{E}_r$ to shape $\frac{H}{P} \times \frac{W}{P} \times D$ . Then, two convolutional layers (stride 1) are followed and produce a $1 / P$ scale warping flow $\pmb{f}_m \in \mathbb{R}^{\frac{H}{P} \times \frac{W}{P} \times 2}$ . Next, We use other two convolutional layers to predict a $H / P \times W / P \times (P \times P \times 9)$ mask and perform softmax over the weights of the $3 \times 3$ neighborhood of each pixel in $\pmb{f}_m$ . Finally, the obtained $\frac{H}{P} \times \frac{W}{P} \times P \times P \times 2$ map is permuted and reshaped to the full resolution warping flow $\pmb{f}_b \in \mathbb{R}^{H \times W \times 2}$ . + +Boundary Refinement Module. We find that the obtained $I_{c}$ struggles with the serrated boundaries and noisy backgrounds (see Fig. 7). To further improve the image quality, we propose a boundary refinement module (BRM). + +Similar to the flow prediction module, BRM also estimates the weights of neighbor pixels for upsampling. Then, we obtain a confidence map $M \in \mathbb{R}^{H \times W \times 1}$ , in which each value denotes the confidence whether a pixel belongs to the rectified region. $M$ is further binarized with a threshold $\sigma$ to obtain the binary boundary refinement mask $M_{b} \in$ + +![](images/07f6194413800740c8e9b70a811cd1e5077fd904f945801add920b52669bfd4c.jpg) +Figure 3: Visualization of the rectification process of a certain pixel on the rectified image based on warping flow. + +$\mathbb{R}^{H\times W\times 1}$ , as illustrated in Fig. 2. The final rectified image $\pmb{I}_r$ can be obtained by multiplying $\pmb{I}_c$ with $M_b$ . + +Training Loss. At the fine-tuning stage, the model is end-to-end optimized with the following training objectives, + +$$ +\mathcal {L} _ {f t} = \mathcal {L} _ {\text {f l o w}} + \mathcal {L} _ {\text {m a s k}}. \tag {9} +$$ + +$\mathcal{L}_{flow}$ is the $L_{1}$ distance between the predicted warping flow $\pmb{f}_b$ and its given ground truth $\pmb{f}_{gt}$ as follows, + +$$ +\mathcal {L} _ {\text {f l o w}} = \left\| \boldsymbol {f} _ {b} \times \boldsymbol {M} _ {g t} - \boldsymbol {f} _ {g t} \right\| _ {1}, \tag {10} +$$ + +where $M_{gt}$ denotes the ground truth valid foreground region of the rectified image. + +$\mathcal{L}_{mask}$ is the cross-entropy loss between the predicted boundary refinement confidence map $M$ and its given ground truth $M_{gt}$ as follows, + +$$ +\mathcal {L} _ {\text {m a s k}} = - \sum_ {i = 1} ^ {N _ {d}} \left[ \boldsymbol {y} _ {i} \log \left(\hat {\boldsymbol {y}} _ {i}\right) + \left(1 - \boldsymbol {y} _ {i}\right) \log \left(1 - \hat {\boldsymbol {y}} _ {i}\right) \right], \tag {11} +$$ + +where $N_{d} = H\times W$ is the number of pixels in image $\pmb{I}_d$ , $\pmb{y}_i\in \{0,1\}$ and $\hat{\pmb{y}}_i\in [0,1]$ denote the ground-truth and the predicted confidence value in $M_{gt}$ and $M$ , respectively. + +# 5. Experiments + +# 5.1. Dataset and Metrics + +We train and evaluate our method according to the conventions in the literature. Following method [46, 6, 58, 37, 38], we establish a dataset using original images from the Place365 dataset [63]. We use the polynomial model (Eq. (3)) to generate the dataset and the first four distortion parameters are used ( $N_{k} = 4$ ), which is enough to fit most of the real-world application scenarios. The synthetic dataset contains multiple annotations for each distorted fisheye image, including distortion-free image, warping flow, and distortion parameters. We synthesize $100\mathrm{k}$ images for training and $5\mathrm{k}$ images for evaluation, respectively. + +In fisheye image rectification, Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM) are the two most popular evaluation metrics in existing works. Given a rectified image and a reference distortion-free one, PSNR measures the detailed quality differences while SSIM assesses their structural differences. Besides, we use Frechet Inception Distance (FID) to evaluate the distribution difference. + +(a) Pre-training. Joint pre-training obtains remarkable performance gain. + +
LconLposPSNRSSIM
20.330.82
✓20.700.83
✓21.300.84
✓✓22.470.86
+ +(b) Fine-tuning way. Both fine-tuning strategies work and training from-scratch is better. + +
settingsPSNRSSIM
w/o pre-training20.330.82
w/ pre-training: frozen encoder21.120.85
unfrozen encoder22.470.86
+ +(c) Shuffle. Without the shuffle, the performance drops to the baseline level. + +
settingsPSNRSSIM
w/o shuffle20.280.82
w/ shuffle22.470.86
+ +(d) Prediction target. Compared with distortion parameter and image, warping flow is the most effective target. + +
targetsPSNRSSIM
parameter21.090.82
image20.500.69
flow22.470.86
+ +(e) Patch size. Reducing patch size works better. To strike a balance between accuracy and efficiency, we choose $16 \times 16$ by default. + +
patchPSNRSSIMFPS
825.820.9014.93
1622.470.8615.51
3220.230.8114.64
6418.650.7414.19
+ +(f) Architecture setting. Several different architectures are set up. Our scheme performs better. + +
settingsPSNRSSIM
w/o BRM18.010.68
w/o FPM22.110.85
full22.470.86
+ +Table 2: Ablation experiments of our method. The default setting is: the patch size is ${16} \times {16}$ and we import 8 pre-trained layers for fine-tuning. Default settings are all marked in gray . + +# 5.2. Implementation + +Model Settings. The image size $(H,W)$ is (256, 256). The latent vector size $D$ in transformer encoder layer [54] is 256 and the total layer number $N_{T}$ is 10. The threshold $\sigma$ for binarizing the confidence map in the boundary refinement module is empirically set as 0.5. + +Training Details. We use 100k images for pre-training and 50k images for fine-tuning. For contrastive learning, we set the temperature $\tau = 0.07$ (Eq. (4)). We utilize Adam [31] as the optimizer and one-cycle policy [50, 52] with a maximum learning rate of $10^{-4}$ . Both stages are trained for 65 epochs with a batch size of 64. Two NVIDIA GeForce RTX 2080Ti GPUs are used to train the network. + +# 5.3. Ablation Study + +We conduct ablations to verify the effectiveness of the core settings and components in our method. + +Impact of Self-supervised Pre-training. The key idea of this work is the self-supervised distortion representation learning for fisheye images. Table 2 (a) shows the studies on the proposed pretext task. Without any pre-training, we can still provide a strong baseline, with a PSNR of 20.33 and a SSIM of 0.82. This can be attributed to the powerful representative ability of transformer [54]. Then, compared with the baseline, the performance improves by pre-training with an independent contrastive loss $\mathcal{L}_{con}$ or position loss $\mathcal{L}_{pos}$ . Furthermore, the unified pre-training boosts the rectification performance significantly. These results verify the effectiveness of our self-supervised representation learning strategy. Note that the two pre-text tasks promote each other. In Fig. 4, we provide the training loss plots of fine-tuning under the setting w/ and w/o the pre-training. From the training perspective, we can observe that our pre-training + +![](images/0239745c14468bb5fe721bc3ea80f79ab6da1313d759164724481fbb74cd1610.jpg) +(a) flow loss + +![](images/0115f2457a924ff0917837b2e1cb0280510c62da158ef65252b6f0fd8cee5721.jpg) +(b) mask loss +Figure 4: Training loss plots during the fine-tuning stage. With the distortion representation pre-training, the training errors of the rectification $(\mathcal{L}_{flow}$ and $\mathcal{L}_{mask})$ are reduced. + +effectively reduces the losses of the baseline. + +Impact of Fine-tuning Way. During the fine-tuning stage, we train the model from scratch. An alternative way is to freeze the pre-trained encoder and only update the rest weights of the model. As shown in Table 2 (b), both fine-tuning ways produce better performance than the baseline (w/o pre-training), illustrating the effectiveness of our learned representations and our default fine-tuning strategy. + +Impact of Shuffle Manipulation. Shuffle strategy is important for our method. During the pre-training stage, we arbitrarily shuffle the patches, to avoid the directed association between the fixed positional embedding and the specific distortion pattern/degree. As shown in Table 2 (c), without the shuffle strategy, the performance degrades to the baseline level in Table 2 (a), which means that the pretraining does not learn any useful representation for the downstream rectification. + +Impact of Reconstruction Target. Our SimFIR estimates a warping flow [21, 19] to resample the pixels from distorted images for rectification. In contrast, other + +![](images/f9a603b1f68151d8be30e1952b56d89969d9178ee4adeddb30493762e622fc1e.jpg) +Distorted + +![](images/a6fddc3642451e0de3e927afbcd858f8877a788454d221e6039fe3138d73f4af.jpg) + +![](images/3c2e28f18320a65836fc77eb0f417a58c7a1d28d2d458fa308938c62de7a36d9.jpg) +Predict image + +![](images/969255af5d368595bc7389d17911c1c767857834e9f9833044df894c5506ff77.jpg) + +![](images/22ced87002969c48ef0968bac6cc4f58f5be16b370d49f9c6ac47e86bed38b10.jpg) +Predict flow + +![](images/fe00e8adac9797956dc0f5bc2969d2880bb48d5a25d1b933d230cde554297c99.jpg) + +![](images/48e7e78695ec26fc399ec072364272131c4e5b11fa1a3d04611bdec73b5d88aa.jpg) +GT + +![](images/1ac6ee09ac80bf311091a481b9454fb01a74e55b449a8667a2943ba18b608a3f.jpg) +Figure 5: Examples to illustrate the impact of the reconstruction target. For our method, predicting flow to unwarp the distorted images preserves more image details. +Figure 6: Running efficiency comparison between our SimFIR and the state-of-the-art PCN [58] on processing different resolution fisheye images. + +pipelines predict the distortion parameters [46, 6, 60, 57] or regress the rectified images directly [39, 58, 38]. We implement the above two pipelines into our method with the following experiments: 1) meaning the output representations $\pmb{E}_r \in \mathbb{R}^{N \times D}$ into a vector and projecting it to regress the distortion parameters, and 2) reshaping the output $\pmb{E}_r$ to form a reconstructed image, following MAE [28]. As shown in Table 2 (d), the performance drops with these two targets. This is because for distortion rectification, predicting the deformation instead of the target image itself is simpler [35, 20], and can preserve more image details (see Fig. 5). The network can focus on the spatial transformation between two images, without needing to restore the texture details. Besides, it is difficult to extract complex nonlinear model parameters in different intervals [58]. + +Impact of Patch Size. Table 2 (e) studies the impact of patch size. We can observe that a smaller patch steadily produces a higher performance. The reason is that, as shown in Table 1, a smaller patch setting increases the subdivision of distortion patterns, which helps to extract more fine-grained distortion representations. However, a smaller patch will also introduce more patches when dividing an input image. To balance the computational cost, we choose a $16 \times 16$ + +![](images/a058b8b71b3d12852876e30a0c0d4080ff2e7f5aca9712954cead75bcc3ce29d.jpg) + +![](images/0bc61729e0d3d8bca1668c505114302f472ca9cec1c7b23b588af4ab44edf9a9.jpg) +Distorted + +![](images/a2061970f951d6a84b8f23d5581b9203e88cdae9d01618b67ead7edd2902f2ed.jpg) + +![](images/712eefc95a89ff2545eae150374b376884f971520e7c8a5eb64268616d61d52d.jpg) +w/o BRM +Figure 7: Examples to illustrate the impact of the boundary refinement module (abbreviated as "BRM"). With the help of BRM, the boundary regions are well rectified. + +![](images/78c6d3bc826a12e3a30179d6dc0ebcddd61965459e7fc95f96e8e451d515288a.jpg) + +![](images/834b23cc52e9216db515b5e9ae0de3e2136c14f8a0566217440760c7feaee1b2.jpg) +w/BRM + +![](images/80ab87c44fdfa1d9e464fa09527d3f74dfd74e390c0e24f0e256dfba54a06e13.jpg) + +![](images/27cb2f89cdf5c3f670fd791524e2428c0e3c9d25cd942d87ab164d88aba8fc56.jpg) +GT + +
MethodPSNR ↑SSIM ↑FID ↓
SC [48]10.140.32229.60
DeepCalib [6]13.280.51136.73
DR-GAN [38]18.660.7074.56
DDM [39]17.510.66114.66
MLC [37]19.330.7355.97
PCN [58]20.530.7630.64
Ours22.470.8617.23
+ +Table 3: Comparison between the state-of-the-art methods and the proposed method on the test set. “↑” indicates the higher the better and “↓” means the opposite. + +patch size as the default setting in our final model. + +Impact of Boundary Refinement Module. In our SimFIR, the BRM is used to smooth the serrated boundaries and remove the noisy backgrounds of the coarse rectified image $I_{c}$ (see Sec. 4.2). As shown in Table 2 (f), with the proposed BRM, performances are remarkably improved. Moreover, we provide two examples to visualize the improvement from the BRM in Fig. 7. As we can see, the boundary regions are well rectified with the help of BRM. + +Impact of Flow Prediction Module. We compare the bilinear upsampling with our learnable upsampling used for FPM. Table 2 (f) shows that our learnable FPM works better than the module based on the bilinear upsampling. This is likely because the coarse bilinear upsampling operation can not recover the small deformations. + +# 5.4. Comparison with State-of-the-art Methods + +Quantitative Comparison. In this section, we compare our SimFIR with the recent state-of-the-art rectification methods, including SC [48], DeepCalib [6], DR-GAN [38], DDM [39], MLC [37], and PCN [58]. The quantitative results are shown in Table 3. As we can see, our method achieves state-of-the-art performance on all metrics. Note that for SC [48], DeepCalib [6], and PCN [58], we calculate the results based on their released codes. For DR- + +![](images/a57ce5860ee7a3020db8c6f4e5a2d39cb4f66c77db246a9dd1fe044a76b7225c.jpg) +Figure 8: Qualitative results on the synthesized test set. For each comparison, we show the distorted image, the rectified results of SC [48], DeepCalib [6], DR-GAN [38], DDM [39], MLC [37], PCN [58], our method, and ground truth. + +![](images/dfb1360f6737ba4235b694bee8ef69d7d51504e14cff0dec84ec488f54fdfc5f.jpg) +Figure 9: Qualitative results on real-world fisheye images. For each comparison, we show the distorted image, the rectified results of SC [48], Blind [36], DeepCalib [6], DR-GAN [38], DDM [39], MLC [37], PCN [58], and our method. + +GAN [38], DDM [39], and MLC [37], we obtain the rectified test set from the authors. + +Qualitative Comparison. To better demonstrate the effectiveness of our proposed method, we further conduct qualitative comparisons with existing methods [48, 6, 38, 39, 37, 58] on the synthetic test set. As present in Fig. 8, the rectified images of our method show less distortion. + +Application to Real-world Fisheye Images. Additional experiments are conducted on real-world fisheye images. In this evaluation, we capture the images with several fisheye lenses of different FoV. As shown in Fig. 9, although there exists a gap between synthetic and real-world data, our method still shows strong generalization ability. Compared with SC [48], our method guarantees the integrity of the recovered content. Compared with Blind [36], DeepCalib [6], DR-GAN [38], DDM [39], MLC [37], and PCN [58], our rectified images show less distortion. Note that in Fig. 9, the resolution of the input distorted fisheye images is much higher (up to $2454 \times 2454$ ) than that in the test set (256 $\times$ 256). Compared with DR-GAN [38], DDM [39], and PCN [58], our method preserves more texture details. + +Efficiency Comparison. Next, we compare the running efficiency of our method and the state-of-the-art method PCN [58] on processing images with various resolutions. As shown in Fig. 6, on different resolutions, our method shows higher efficiency. The reason is that PCN [58] directly regresses the rectified image with a fully convolutional network [47], like DR-GAN [38], and DDM [39]. In contrast, our SimFIR infers a warping flow at the fixed resolution $(256 \times 256)$ and interpolates it to the original high resolution same as the input image for unwarping. Such a strategy ensures efficiency and does not demand any extra training at a certain scale like image-based methods. + +# 6. Conclusion + +In this work, we present an effective self-supervised representation learning paradigm for fisheye images. We associate different image regions with their specific distortion patterns and design a distortion-aware pretext task for their learning. With the pre-training, the network extracts the fine-grained distortion representation. Extensive experiments are conducted on our proposed dataset, and the trans + +fer performance in the downstream rectification task verifies the effectiveness of the learned representations. + +Acknowledgements This work was supported by NSFC under Contract 61836011 and 62021001. It was also supported by the GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, and the Supercomputing Center of the USTC. + +# References + +[1] Kohei Aso, Dong-Hyun Hwang, and Hideki Koike. Portable 3D human pose estimation for human-human interaction using a chest-mounted fisheye camera. In AHs, pages 116-120, 2021. +[2] Joao Barreto, Jose Roquette, Peter Sturm, and Fernando Fonseca. Automatic camera calibration applied to medical endoscopy. In BMVC, pages 1-10, 2009. +[3] João Pedro Barreto and Helder Araujo. Geometric properties of central catadioptric line images and their application in calibration. TPAMI, 27(8):1327-1333, 2005. +[4] João Pedro Barreto and Kostas Daniilidis. Fundamental matrix for cameras with radial distortion. In ICCV, volume 1, pages 625-632, 2005. +[5] Gregor Blott, Masato Takami, and Christian Heipke. Semantic segmentation of fisheye images. In ECCVW, pages 0-0, 2018. +[6] Oleksandr Bogdan, Viktor Eckstein, Francois Rameau, and Jean-Charles Bazin. DeepCalib: a deep learning approach for automatic intrinsic calibration of wide field-of-view cameras. In CVMP, pages 1–10, 2018. +[7] Faisal Bukhari and Matthew N Dailey. Automatic radial distortion estimation from a single image. Journal of Mathematical Imaging and Vision, 45(1):31-45, 2013. +[8] Gyanesh Chander, Brian L Markham, and Dennis L Helder. Summary of current radiometric calibration coefficients for Landsat MSS, TM, ETM+, and EO-1 ALI sensors. Remote Sensing of Environment, 113(5):893–903, 2009. +[9] Chun-Hao Chao, Pin-Lun Hsu, Hung-Yi Lee, and Yu-Chiang Frank Wang. Self-supervised deep learning for fisheye image rectification. In ICASSP, pages 2248-2252, 2020. +[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020. +[11] Jonathan Courbon, Youcef Mezouar, Laurent Eckt, and Philippe Martinet. A generic fisheye camera model for robotic applications. In IROS, pages 1683-1688, 2007. +[12] Jonathan Courbon, Youcef Mezouar, and Philippe Martinet. Evaluation of the unified model of the sphere for fisheye cameras in robotic applications. Advanced Robotics, 26(8-9):947-967, 2012. +[13] Liuyuan Deng, Ming Yang, Yeqiang Qian, Chunxiang Wang, and Bing Wang. CNN based semantic segmentation for urban traffic scenes using fisheye camera. In IV, pages 231-236, 2017. +[14] Frederic Devernay and Olivier Faugeras. Straight lines have to be straight. Machine Vision and Applications, 13(1):14-24, 2001. + +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. +[16] Richard O Duda and Peter E Hart. Use of the hough transformation to detect lines and curves in pictures. CACM, 15(1):11-15, 1972. +[17] Jinlong Fan, Jing Zhang, Stephen J Maybank, and Dacheng Tao. Wide-angle image rectification: a survey. IJCV, 130(3):747-776, 2022. +[18] Jinlong Fan, Jing Zhang, and Dacheng Tao. SIR: Self-supervised image rectification via seeing the same scene from multiple different lenses. arXiv preprint arXiv:2011.14611, 2020. +[19] Hao Feng, Shaokai Liu, Jiajun Deng, Wengang Zhou, and Houqiang Li. Deep unrestricted document image rectification. arXiv preprint arXiv:2304.08796, 2023. +[20] Hao Feng, Wengang Zhou, Jiajun Deng, Qi Tian, and Houqiang Li. DocScanner: Robust document image rectification with progressive learning. arXiv preprint arXiv:2110.14968, 2021. +[21] Hao Feng, Wengang Zhou, Jiajun Deng, Yuechen Wang, and Houqiang Li. Geometric representation learning for document image rectification. In ECCV, pages 475-492, 2022. +[22] James M Ferryman, Stephen J Maybank, and Anthony D Worrall. Visual surveillance for moving vehicles. *IJCV*, 37(2):187-197, 2000. +[23] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the KITTI vision benchmark suite. In CVPR, pages 3354-3361, 2012. +[24] Payam Goodarzi, Martin Stellmacher, Michael Paetzold, Ahmed Hussein, and Elmar Matthes. Optimization of a CNN-based object detector for fisheye cameras. In ICVES, pages 1-7, 2019. +[25] Sorin Grigorescu, Bogdan Trasnea, Tiberiu Cocias, and Gigel Macesanu. A survey of deep learning techniques for autonomous driving. Journal of Field Robotics, 37(3):362-386, 2020. +[26] Michael D Grossberg and Shree K Nayar. A general imaging model and a method for finding its parameters. In ICCV, volume 2, pages 108-115, 2001. +[27] Richard Hartley and Sing Bing Kang. Parameter-free radial distortion correction with center of distortion estimation. TPAMI, 29(8):1309-1321, 2007. +[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. +[29] Jose Henrique Brito, Roland Angst, Kevin Koser, and Marc Pollefeys. Radial distortion self-calibration. In CVPR, pages 1368-1375, 2013. +[30] Juho Kannala and Sami S Brandt. A generic camera model and calibration method for conventional, wide-angle, and fish-eye lenses. TPAMI, 28(8):1335-1340, 2006. +[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. + +[32] Zuzana Kukelova and Tomas Pajdla. A minimal solution to radial distortion autocalibration. TPAMI, 33(12):2410-2422, 2011. +[33] Varun Ravi Kumar, Marvin Klingner, Senthil Yogamani, Stefan Milz, Tim Fingscheidt, and Patrick Mader. SyndistNet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In IEEE Winter Conference on Applications of Computer Vision, pages 61-71, 2021. +[34] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In CVPR, pages 4681-4690, 2017. +[35] Xiaoyu Li, Bo Zhang, Jing Liao, and Pedro V Sander. Document rectification and illumination correction using a patch-based CNN. TOG, 38(6):1-11, 2019. +[36] Xiaoyu Li, Bo Zhang, Pedro V Sander, and Jing Liao. Blind geometric distortion correction on images through deep learning. In CVPR, pages 4855-4864, 2019. +[37] Kang Liao, Chunyu Lin, Lixin Liao, Yao Zhao, and Weiyao Lin. Multi-level curriculum for training a distortion-aware barrel distortion rectification model. In ICCV, pages 4389-4398, 2021. +[38] Kang Liao, Chunyu Lin, Yao Zhao, and Moncef Gabbouj. DR-GAN: Automatic radial distortion rectification using conditional gan in real-time. TCSVT, 30(3):725-733, 2019. +[39] Kang Liao, Chunyu Lin, Yao Zhao, and Mai Xu. Model-free distortion rectification framework bridged by distortion distribution map. TIP, 29:3707-3718, 2020. +[40] Liang Lin, Yongyi Lu, Yan Pan, and Xiaowu Chen. Integrating graph partitioning and matching for trajectory analysis in video surveillance. TIP, 21(12):4844-4857, 2012. +[41] Khan Muhammad, Jamil Ahmad, Zhihan Lv, Paolo Bellavista, Po Yang, and Sung Wook Baik. Efficient deep CNN-based fire detection and localization in video surveillance applications. TSMSC, 49(7):1419-1434, 2018. +[42] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. +[43] Elad Plaut, Erez Ben Yaacov, and Bat El Shlomo. 3D object detection from a single fisheye image without a single fisheye training image. In CVPR, pages 3659-3667, 2021. +[44] Luis Puig, Yalin Bastanlar, Peter Sturm, José Jesús Guerrero, and Joao Barreto. Calibration of central catadioptric cameras using a dlt-like approach. IJCV, 93(1):101-114, 2011. +[45] Helge Rhodin, Christian Richardt, Dan Casas, Eldar Insafutdinov, Mohammad Shafiei, Hans-Peter Seidel, Bernt Schiele, and Christian Theobalt. EgoCap: Egocentric marker-less motion capture with two fisheye cameras. TOG, 35(6):1-11, 2016. +[46] Jiangpeng Rong, Shiyao Huang, Zeyu Shang, and Xianghua Ying. Radial lens distortion correction using convolutional neural networks trained with synthesized images. In ACCV, pages 35-49, 2016. +[47] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234–241, 2015. + +[48] Daniel Santana-Cedrés, Luis Gomez, Miguel Aleman-Flores, Agustín Salgado, Julio Esclarín, Luis Mazorra, and Luis Alvarez. An iterative optimization algorithm for lens distortion correction using two-parameter models. Image Processing On Line, 6:326-364, 2016. +[49] Davide Scaramuzza, Agostino Martinelli, and Roland Siegwart. A flexible technique for accurate omnidirectional camera calibration and structure from motion. In ICVS, pages 45-45, 2006. +[50] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial Intelligence and Machine Learning for Multi-domain Operations Applications, volume 11006, pages 369-386, 2019. +[51] Peter Sturm and Srikumar Ramalingam. A generic concept for camera calibration. In ECCV, pages 1-13, 2004. +[52] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419, 2020. +[53] Thorsten Thormählen, Hellward Broszio, and Ingolf Wassermann. Robust line-based calibration of lens distortion from a single view. *Mirage*, pages 105–112, 2003. +[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. +[55] Aiqi Wang, Tianshuang Qiu, and Longtan Shao. A simple method of radial distortion correction with centre of distortion estimation. Journal of Mathematical Imaging and Vision, 35(3):165-172, 2009. +[56] Guo-Wang Xie, Fei Yin, Xu-Yao Zhang, and Cheng-Lin Liu. Dewarping document image by displacement flow estimation with fully convolutional network. In DAS, pages 131-144, 2020. +[57] Zhucun Xue, Nan Xue, Gui-Song Xia, and Weiming Shen. Learning to calibrate straight lines for fisheye image rectification. In CVPR, pages 1643-1651, 2019. +[58] Shangrong Yang, Chunyu Lin, Kang Liao, Chunjie Zhang, and Yao Zhao. Progressively complementary network for fisheye image rectification using appearance flow. In CVPR, pages 6348-6357, 2021. +[59] Mingde Yao, Dongliang He, Xin Li, Fu Li, and Zhiwei Xiong. Towards interactive self-supervised denoising. IEEE Transactions on Circuits and Systems for Video Technology, 2023. +[60] Xiaoqing Yin, Xinchao Wang, Jun Yu, Maojun Zhang, Pascal Fua, and Dacheng Tao. FisheyeRecNet: A multi-context collaborative deep network for fisheye image rectification. In ECCV, pages 469-484, 2018. +[61] Mi Zhang, Jian Yao, Menghan Xia, Kai Li, Yi Zhang, and Yaping Liu. Line-based multi-label energy optimization for fisheye image rectification and calibration. In CVPR, pages 4137-4145, 2015. +[62] Yahui Zhang, Shaodi You, and Theo Gevers. Automatic calibration of the fisheye camera for egocentric 3D human pose estimation from a single image. In WACV, pages 1772-1781, 2021. +[63] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. TPAMI, 40(6):1452-1464, 2017. \ No newline at end of file diff --git a/2023/SimFIR_ A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning/images.zip b/2023/SimFIR_ A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0cd93cf95ae307b04329cd61f938779a5c3110da --- /dev/null +++ b/2023/SimFIR_ A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:221080cec01dbd2a94025c2c476e3e4fb44be7cc09e6506a54c4f1bf5fa10959 +size 597497 diff --git a/2023/SimFIR_ A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning/layout.json b/2023/SimFIR_ A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8ec5186675395744e727d70b33e68e44f86c65df --- /dev/null +++ b/2023/SimFIR_ A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning/layout.json @@ -0,0 +1,11240 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 103, + 506, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 103, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 88, + 103, + 506, + 140 + ], + "type": "text", + "content": "SimFIR: A Simple Framework for Fisheye Image Rectification with Self-supervised Representation Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 160, + 528, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 160, + 528, + 176 + ], + "spans": [ + { + "bbox": [ + 67, + 160, + 528, + 176 + ], + "type": "text", + "content": "Hao Feng" + }, + { + "bbox": [ + 67, + 160, + 528, + 176 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 67, + 160, + 528, + 176 + ], + "type": "text", + "content": " Wendi Wang" + }, + { + "bbox": [ + 67, + 160, + 528, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 160, + 528, + 176 + ], + "type": "text", + "content": " Jiajun Deng" + }, + { + "bbox": [ + 67, + 160, + 528, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 160, + 528, + 176 + ], + "type": "text", + "content": " Wengang Zhou" + }, + { + "bbox": [ + 67, + 160, + 528, + 176 + ], + "type": "inline_equation", + "content": "^{1,4,*}" + }, + { + "bbox": [ + 67, + 160, + 528, + 176 + ], + "type": "text", + "content": " Li Li" + }, + { + "bbox": [ + 67, + 160, + 528, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 160, + 528, + 176 + ], + "type": "text", + "content": " Houqiang Li" + }, + { + "bbox": [ + 67, + 160, + 528, + 176 + ], + "type": "inline_equation", + "content": "^{1,4,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 176, + 529, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 176, + 529, + 189 + ], + "spans": [ + { + "bbox": [ + 64, + 176, + 529, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 176, + 529, + 189 + ], + "type": "text", + "content": " CAS Key Laboratory of Technology in GIPAS, EEIS Department, University of Science and Technology of China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 148, + 190, + 446, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 190, + 446, + 203 + ], + "spans": [ + { + "bbox": [ + 148, + 190, + 446, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 148, + 190, + 446, + 203 + ], + "type": "text", + "content": " Zhangjiang Laboratory, Shanghai, China " + }, + { + "bbox": [ + 148, + 190, + 446, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 148, + 190, + 446, + 203 + ], + "type": "text", + "content": " The University of Sydney" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 204, + 463, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 204, + 463, + 217 + ], + "spans": [ + { + "bbox": [ + 129, + 204, + 463, + 217 + ], + "type": "text", + "content": "4 Institute of Artificial Intelligence, Hefei Comprehensive National Science Center" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 219, + 539, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 219, + 539, + 232 + ], + "spans": [ + { + "bbox": [ + 54, + 219, + 539, + 232 + ], + "type": "text", + "content": "{haof,wendiwang}@mail.ustc.edu.cn，jiajun.deng@sydney.edu.au，{zhwg,lil1,lihq}@ustc.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 284, + 290, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 284, + 290, + 523 + ], + "spans": [ + { + "bbox": [ + 45, + 284, + 290, + 523 + ], + "type": "text", + "content": "In fisheye images, rich distinct distortion patterns are regularly distributed in the image plane. These distortion patterns are independent of the visual content and provide informative cues for rectification. To make the best of such rectification cues, we introduce SimFIR, a simple framework for fisheye image rectification based on self-supervised representation learning. Technically, we first split a fisheye image into multiple patches and extract their representations with a Vision Transformer (ViT). To learn fine-grained distortion representations, we then associate different image patches with their specific distortion patterns based on the fisheye model, and further subtly design an innovative unified distortion-aware pretext task for their learning. The transfer performance on the downstream rectification task is remarkably boosted, which verifies the effectiveness of the learned representations. Extensive experiments are conducted, and the quantitative and qualitative results demonstrate the superiority of our method over the state-of-the-art algorithms as well as its strong generalization ability on real-world fisheye images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 544, + 128, + 557 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 128, + 557 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 128, + 557 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 564, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 287, + 673 + ], + "type": "text", + "content": "Fisheye lenses have been widely used in systems for video surveillance [41, 40, 22], autonomous driving [25, 23] and robotic applications [11, 12], thanks to their extremely wide field of view (FoV). However, images captured by the fisheye lenses are inevitably distorted, which are not directly applicable for the downstream analysis, such as object detection [43, 24], pose estimation [62, 1, 45] and scene segmentation [33, 13, 5]. Over the past few years, fisheye image rectification has become an emerging research topic." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 673, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 287, + 696 + ], + "type": "text", + "content": "In the literature, some classical methods [4, 30, 27, 32, 44, 26, 17] estimate epipolar geometry with images" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 311, + 258, + 544, + 389 + ], + "blocks": [ + { + "bbox": [ + 311, + 258, + 544, + 389 + ], + "lines": [ + { + "bbox": [ + 311, + 258, + 544, + 389 + ], + "spans": [ + { + "bbox": [ + 311, + 258, + 544, + 389 + ], + "type": "image", + "image_path": "6e38eb2ecc873e801239b67b6afc8c0d0a673db477e89c907afe78bec953af75.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 395, + 547, + 444 + ], + "lines": [ + { + "bbox": [ + 305, + 395, + 547, + 444 + ], + "spans": [ + { + "bbox": [ + 305, + 395, + 547, + 444 + ], + "type": "text", + "content": "Figure 1: An illustration of fisheye distortions. The distortions in fisheye images are radially distributed and centrally symmetrical. The distortion degree increases with distance from the optical axis/image center." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 471, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 547, + 674 + ], + "type": "text", + "content": "from different viewpoints. However, multiview images are not always accessible, which limits the real applications of these methods. To overcome this limitation, other classical methods are based on the detection of plumb lines [53, 55, 7, 14, 61, 2]. Nevertheless, taking the detected plumb lines as the basis of fisheye image rectification inevitably causes error accumulation. Recently, deep learning based solutions [46, 6, 60, 38, 36, 58, 37, 39, 57] have become a promising alternative to the traditional methods. There are two directions among deep learning based methods. One direction [46, 6, 60, 57] attempts to predict the distortion parameters to infer pixel-wise displacement for rectification. The other direction [38, 58, 37, 39] directly predicts the rectified images with specifically designed architectures. Although these methods are reported with state-of-the-art results, the intrinsic distortion characteristics of fisheye images are largely ignored, described next." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "Generally, as illustrated in Fig. 1, the distortions in fisheye images are radially distributed and centrally symmetrical. The distortion degree increases with distance from the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 249, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 249, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 249, + 714 + ], + "type": "text", + "content": "*Corresponding authors: Wengang Zhou and Houqiang Li" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12418" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "image center along the radial axis. It should be noted that here we follow the existing state-of-the-art rectification algorithms [46, 6, 60, 38, 36, 58, 37, 39, 57] to only consider the ideal fisheye images, in which the distortion center locate at the image center. Furthermore, when dividing a fisheye image into patches, the distortion patterns and degrees vary among patches at different distances from the image center. To improve fisheye image rectification, it is beneficial to model these fine-grained distortion patterns that provide cues for the rectification." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 289, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 289, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 289, + 467 + ], + "type": "text", + "content": "Formally, we present SimFIR, a new framework for fisheye image rectification based on self-supervised representation learning [59]. SimFIR introduces an effective way for encoding the fine-grained distortion patterns in fisheye images, by learning the associations between the local image region and their distinct distortion pattern. Specifically, we first split a fisheye image into multiple patches and label them with a position map (see Fig. 1) based on their specific distortion degrees. Then, these image patches are embedded as input tokens of Vision Transformer (ViT) [15] to abstract patch representations. Given these patch representations, we introduce a unified distortion-aware pretext task to optimize the network. On the one hand, we perform contrastive learning with InfoNCE loss [42] to narrow the representation of patches with the same distortion pattern while enlarging the feature distance with others. On the other hand, we devise a classification supervision to predict the patchwise quantified distortion degree, aiming to differentiate the distinct local distortion patterns. After pre-training, the network is capable of extracting the fine-grained distortion representations, regardless of the image textures. Thereafter, the learned distortion representations are transferred to the downstream rectification task." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 468, + 289, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 289, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 289, + 635 + ], + "type": "text", + "content": "Besides, we introduce a new pipeline for fisheye image rectification, which unwarps fisheye images with a dense flow field, i.e., the warping flow, aiming to process images of arbitrary resolution without complex post-processing [36]. Then, to facilitate training, we construct a well-annotated synthesized dataset that covers various scenes and distortion levels. To validate the learned representations, we conduct extensive experiments by pretraining the network with our self-supervised paradigm and then transferring it to the downstream rectification task. The results demonstrate the effectiveness of our method as well as its superiority over existing state-of-the-art algorithms. In addition, we show the strong generalization ability of our method on processing real-world fisheye images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 635, + 249, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 635, + 249, + 647 + ], + "spans": [ + { + "bbox": [ + 59, + 635, + 249, + 647 + ], + "type": "text", + "content": "In summary, we make three-fold contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 656, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 58, + 656, + 286, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 656, + 286, + 692 + ], + "spans": [ + { + "bbox": [ + 58, + 656, + 286, + 692 + ], + "type": "text", + "content": "- We present SimFIR, a novel self-supervised based framework, to improve fisheye image rectification by extracting its fine-grained distortion patterns." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 701, + 286, + 713 + ], + "type": "text", + "content": "- We introduce a new pipeline for fisheye image recti" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 325, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 545, + 96 + ], + "type": "text", + "content": "fication that unwarps images with backward warping flow, aiming to process images with any resolutions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 105, + 545, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 105, + 545, + 141 + ], + "spans": [ + { + "bbox": [ + 317, + 105, + 545, + 141 + ], + "type": "text", + "content": "- We conduct extensive experiments to verify the merits of our method and report the state-of-the-art performance with our proposed method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 152, + 392, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 152, + 392, + 165 + ], + "spans": [ + { + "bbox": [ + 306, + 152, + 392, + 165 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 173, + 545, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 173, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 305, + 173, + 545, + 209 + ], + "type": "text", + "content": "Fisheye image rectification has been widely explored in the computer vision community. In the following, we separately discuss the methods of each group." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 212, + 545, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 212, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 304, + 212, + 545, + 332 + ], + "type": "text", + "content": "Rectification Based on Multi-view Calibration. Some traditional methods [4, 27, 32, 44, 29, 51, 49] calibrate fisheye images by finding the corresponding feature points from multi perspectives. Typically, FMC [4] proposes to estimate the epipolar geometry of two views with different distortion factors. AMS [32] explores the minimal solution that calibrates radial distortion with only eight-point correspondences in two images. However, such methods require special chessboards and multiple images are not always accessible, unfriendly to their practical applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 335, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 443 + ], + "type": "text", + "content": "Rectification Based on Curve Line Detection. The overcome these limitations, line-based camera calibration methods [53, 55, 7, 14, 3, 61, 8] propose to utilize the prior that straight lines in the 3D world must project to straight lines in the image plane. Typically, SC [48] first detects the longest distorted lines within an image by applying Hough transformation [16] and then estimates the distortion parameters from these lines. However, the curve detection is not robust and challenging, limiting their performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 447, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 590 + ], + "type": "text", + "content": "Rectification Based on Parameter Estimation. By predicting the distortion parameters of fisheye models, some methods compute the per-pixel coordinate displacement to perform rectification. RLDC [46] first employs a convolutional neural network to predict the distortion parameters. DeepCalib [6] predicts the distortion parameters and the focal length with a network. Given a single image, Blind [36] devises a unified framework to correct different types of geometric distortion by the estimation of the distortion parameter and forward warping flow [56]. LCSL [57] imposes an explicit geometry constraint onto the processes of the fisheye lens calibration and rectification." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Rectification Based on Image Estimation. Methods following this direction commonly attempt to directly predict the rectified images. DR-GAN [38] exploits a generative adversarial network (GAN) [34] for rectified image synthesis, and proposes a low-to-high perceptual loss to capture the mapping relation between different structural images. FE-GAN [9] learns pixel-level distortion flow from sets of fisheye distorted images and distortion-free ones with the cross-rotation and intra-warping consistency. SIR [18] devises a self-supervised learning idea for recti" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12419" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 70, + 498, + 149 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 498, + 149 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 498, + 149 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 498, + 149 + ], + "type": "image", + "image_path": "633c14111b2496382355018af55e252138f30ac344882cda829189842819fbda.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 153, + 428, + 165 + ], + "lines": [ + { + "bbox": [ + 164, + 153, + 428, + 165 + ], + "spans": [ + { + "bbox": [ + 164, + 153, + 428, + 165 + ], + "type": "text", + "content": "(a) Stage 1: Self-supervised pre-training for fine-gained distortion learning" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 88, + 176, + 508, + 276 + ], + "blocks": [ + { + "bbox": [ + 88, + 176, + 508, + 276 + ], + "lines": [ + { + "bbox": [ + 88, + 176, + 508, + 276 + ], + "spans": [ + { + "bbox": [ + 88, + 176, + 508, + 276 + ], + "type": "image", + "image_path": "a6bdc4b49e9fac12dd742cb3d6cc3f500654e9ba8455e4bb156a97316217c01b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 279, + 386, + 290 + ], + "lines": [ + { + "bbox": [ + 205, + 279, + 386, + 290 + ], + "spans": [ + { + "bbox": [ + 205, + 279, + 386, + 290 + ], + "type": "text", + "content": "(b) Stage 2: Fine-tuning for distortion rectification" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 94, + 302, + 498, + 324 + ], + "blocks": [ + { + "bbox": [ + 94, + 302, + 498, + 324 + ], + "lines": [ + { + "bbox": [ + 94, + 302, + 498, + 324 + ], + "spans": [ + { + "bbox": [ + 94, + 302, + 498, + 324 + ], + "type": "image", + "image_path": "6642815caaf99c5107fad157922149033b12cc536f55c594b1f3246169f16711.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 332, + 546, + 381 + ], + "lines": [ + { + "bbox": [ + 46, + 332, + 546, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 546, + 381 + ], + "type": "text", + "content": "Figure 2: Framework of our method for fisheye image rectification. It consists of two stages: (a) self-supervised pre-training that learns the fine-grained distortion representation of fisheye image " + }, + { + "bbox": [ + 46, + 332, + 546, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_d" + }, + { + "bbox": [ + 46, + 332, + 546, + 381 + ], + "type": "text", + "content": " with a unified distortion-aware pretext task; (b) fine-tuning for rectification which uses the learned representation to reconstruct the rectified image " + }, + { + "bbox": [ + 46, + 332, + 546, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_r" + }, + { + "bbox": [ + 46, + 332, + 546, + 381 + ], + "type": "text", + "content": " with a flow field. \"BRM\" and \"FPM\" denote the proposed boundary refinement module and flow prediction module, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 401, + 288, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 288, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 288, + 486 + ], + "type": "text", + "content": "fication by exploiting intra-model and inter-model consistency. DDM [39] proposes a distortion distribution map that intuitively indicates the global distortion features of a distorted image. Recently, MLC [37] introduces a multilevel curriculum to train the rectification model hierarchically. PCN [58] develops an encoder-decoder architecture to generate multi-scale rectified images progressively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 498, + 201, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 201, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 201, + 511 + ], + "type": "text", + "content": "3. Fisheye Model and Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 518, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 590 + ], + "type": "text", + "content": "We first introduce the distortion model for fisheye images. Since obtaining ground truth is difficult for real fisheye images, we follow the common practices in the literature [60, 57, 6, 39, 37, 58] to exploit synthetic data for training. The polynomial model [30] of a fisheye image is formulated as follows," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 86, + 600, + 287, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 600, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 86, + 600, + 287, + 633 + ], + "type": "interline_equation", + "content": "\\theta_ {c} = \\sum_ {i = 1} ^ {N _ {k}} \\lambda_ {i} \\theta_ {d} ^ {2 i - 1}, N _ {k} = 1, 2, 3, 4, \\dots , \\tag {1}", + "image_path": "c6971106892b2a2fad8ea265c4abe6b0908c6d804fbd81776908628994888145.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{c}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " denotes the angle of incident light, " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{d}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " is the angle of the light after passing through the lens, " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " distortion parameter, and " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " is the number of distortion parameters. According to the projection model for general pinhole cameras, we have " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "r_c = f\\cdot \\tan \\theta_c\\approx f\\cdot \\theta_c" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " is the focal length of a camera, " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "r_c" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " distance between" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 402, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 402, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 402, + 545, + 426 + ], + "type": "text", + "content": "the image center and an arbitrary point " + }, + { + "bbox": [ + 305, + 402, + 545, + 426 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 305, + 402, + 545, + 426 + ], + "type": "text", + "content": " in a distortion-free image. Hence, the distortion model can be rewritten as," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 341, + 437, + 545, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 437, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 341, + 437, + 545, + 469 + ], + "type": "interline_equation", + "content": "r _ {c} = f \\sum_ {i = 1} ^ {N _ {k}} \\lambda_ {i} \\theta_ {d} ^ {2 i - 1}, N _ {k} = 1, 2, 3, 4, \\dots . \\tag {2}", + "image_path": "5d488f706163ed565fcaf9db7c7d83d917e3ac59a08d2a3582127e6ae3d9687d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 480, + 546, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 480, + 546, + 529 + ], + "spans": [ + { + "bbox": [ + 305, + 480, + 546, + 529 + ], + "type": "text", + "content": "Furthermore, based on the equidistant projection relation " + }, + { + "bbox": [ + 305, + 480, + 546, + 529 + ], + "type": "inline_equation", + "content": "r_d = f \\cdot \\theta_d" + }, + { + "bbox": [ + 305, + 480, + 546, + 529 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 480, + 546, + 529 + ], + "type": "inline_equation", + "content": "r_d" + }, + { + "bbox": [ + 305, + 480, + 546, + 529 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 305, + 480, + 546, + 529 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 305, + 480, + 546, + 529 + ], + "type": "text", + "content": " distance between the image center and the mapping point " + }, + { + "bbox": [ + 305, + 480, + 546, + 529 + ], + "type": "inline_equation", + "content": "P'" + }, + { + "bbox": [ + 305, + 480, + 546, + 529 + ], + "type": "text", + "content": " in the distorted image, we merge " + }, + { + "bbox": [ + 305, + 480, + 546, + 529 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 480, + 546, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 480, + 546, + 529 + ], + "type": "inline_equation", + "content": "\\theta_d" + }, + { + "bbox": [ + 305, + 480, + 546, + 529 + ], + "type": "text", + "content": " to obtain the fisheye distortion model as," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 346, + 542, + 545, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 542, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 346, + 542, + 545, + 574 + ], + "type": "interline_equation", + "content": "r _ {c} = \\sum_ {i = 1} ^ {N _ {k}} k _ {i} r _ {d} ^ {2 i - 1}, N _ {k} = 1, 2, 3, 4, \\dots . \\tag {3}", + "image_path": "b93b7981b4bc840fabc2f3b0ea04088c35822b76c11a320810300388e161a1cf.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 582, + 545, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 582, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 305, + 582, + 545, + 606 + ], + "type": "text", + "content": "It depicts the radial distortion distribution of fisheye image and is used for the dataset generation (see Sec. 5.1)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 620, + 389, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 620, + 389, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 620, + 389, + 634 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 641, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 546, + 714 + ], + "type": "text", + "content": "Our goal is to learn useful visual representations of fisheye images to improve their rectification. Fig. 2 presents the overview of our method, which consists of two stages: (a) self-supervised pre-training for distortion representation learning, and (b) fine-tuning for distortion rectification. We elaborate on each stage in the following sections." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12420" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 202, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 202, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 202, + 85 + ], + "type": "text", + "content": "4.1. Self-supervised Pre-training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 174 + ], + "type": "text", + "content": "As shown in Eq. (3), fisheye images have an inherent property where the distortions are radially symmetrical, and the degree of distortion in a particular region is dependent on its position on the image plane. A greater radius leads to a higher degree of distortion. To enhance rectification by capturing these detailed distortion patterns, we propose a unified distortion-aware pretext task." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 175, + 288, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 288, + 281 + ], + "type": "text", + "content": "Specifically, on the one hand, we use contrastive learning [10] to narrow the patch features with the same distortion degree and push away the different ones. This approach allows us to encode local distortion features while eliminating texture features. On the other hand, we estimate the patch-wise quantified distortion degrees to learn fine-grained distortion features, by introducing a position loss (see Fig. 2). To implement the idea, we introduce a transformer-based model [54], described next." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "type": "text", + "content": "Patch Embedding. The standard transformer [54] takes a sequence of token embeddings as input. To adapt it, as shown in Fig. 2, given a fisheye image " + }, + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_d \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "type": "text", + "content": ", we first divide it into a sequence of 2D patches " + }, + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_p \\in \\mathbb{R}^{N \\times (P^2 \\cdot 3)}" + }, + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "type": "text", + "content": " are the height and width of the image " + }, + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_d" + }, + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "type": "text", + "content": " is the patch size, and " + }, + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "type": "inline_equation", + "content": "N = HW / P^2" + }, + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "type": "text", + "content": " denotes the number of patches. Then, we flatten these patches into vectors and map them to " + }, + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "type": "text", + "content": " dimensions with a linear projection. We refer to the output projection as the patch embeddings " + }, + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_o \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 46, + 285, + 287, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 411, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 553 + ], + "type": "text", + "content": "Distortion Encoding Transformer. To retain positional information, positional embeddings are supplemented to the patch embeddings in transformer [54]. In our method, to avoid the simple mapping between the positional embedding and specific distortion pattern, we arbitrarily shuffle the patch embeddings " + }, + { + "bbox": [ + 46, + 411, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\pmb{E}_{o} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 46, + 411, + 287, + 553 + ], + "type": "text", + "content": " along the first dimension (token dimension). Then, positional embeddings [54] (the sine-cosine version) are added to the shuffled patch embeddings. The resulting sequence of embedding vectors are fed into the " + }, + { + "bbox": [ + 46, + 411, + 287, + 553 + ], + "type": "inline_equation", + "content": "N_{T}" + }, + { + "bbox": [ + 46, + 411, + 287, + 553 + ], + "type": "text", + "content": " transformer encoder [54] layers. Thereafter, we obtain the abstracted patch representations " + }, + { + "bbox": [ + 46, + 411, + 287, + 553 + ], + "type": "inline_equation", + "content": "\\pmb{E}_{r} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 46, + 411, + 287, + 553 + ], + "type": "text", + "content": ", as shown in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 552, + 287, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 698 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 698 + ], + "type": "text", + "content": "In the meantime, we build a position map " + }, + { + "bbox": [ + 46, + 552, + 287, + 698 + ], + "type": "inline_equation", + "content": "P_{o}\\in \\mathbb{R}^{\\frac{H}{P}\\times \\frac{W}{P}}" + }, + { + "bbox": [ + 46, + 552, + 287, + 698 + ], + "type": "text", + "content": " in which each value denotes the distortion degree of a patch. For clarity, we provide an example of the position map, as shown in Fig. 1. Note that it can be readily obtained from the patch radius. Besides, in Table 1, we show the category number " + }, + { + "bbox": [ + 46, + 552, + 287, + 698 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 46, + 552, + 287, + 698 + ], + "type": "text", + "content": " of distortion degree, when dividing a " + }, + { + "bbox": [ + 46, + 552, + 287, + 698 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 46, + 552, + 287, + 698 + ], + "type": "text", + "content": " image with " + }, + { + "bbox": [ + 46, + 552, + 287, + 698 + ], + "type": "inline_equation", + "content": "P\\times P" + }, + { + "bbox": [ + 46, + 552, + 287, + 698 + ], + "type": "text", + "content": " patches. Note that " + }, + { + "bbox": [ + 46, + 552, + 287, + 698 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 46, + 552, + 287, + 698 + ], + "type": "text", + "content": " measures the level of granularity of the division on distortion degree. Then, " + }, + { + "bbox": [ + 46, + 552, + 287, + 698 + ], + "type": "inline_equation", + "content": "P_{o}" + }, + { + "bbox": [ + 46, + 552, + 287, + 698 + ], + "type": "text", + "content": " is reshaped and shuffled in the same way as the patch embeddings. In this way, we obtain the per-patch distortion degree annotation " + }, + { + "bbox": [ + 46, + 552, + 287, + 698 + ], + "type": "inline_equation", + "content": "P_{r}\\in \\mathbb{R}^{N\\times 1}" + }, + { + "bbox": [ + 46, + 552, + 287, + 698 + ], + "type": "text", + "content": ", corresponding to the patch representations " + }, + { + "bbox": [ + 46, + 552, + 287, + 698 + ], + "type": "inline_equation", + "content": "E_{r}\\in \\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 46, + 552, + 287, + 698 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "Contrastive Loss. To learn a useful distortion represent" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 77, + 543, + 104 + ], + "blocks": [ + { + "bbox": [ + 309, + 77, + 543, + 104 + ], + "lines": [ + { + "bbox": [ + 309, + 77, + 543, + 104 + ], + "spans": [ + { + "bbox": [ + 309, + 77, + 543, + 104 + ], + "type": "table", + "html": "
P8×816×1632×3264×64
Ct13636103
", + "image_path": "bb9aa4237d6d6fdb5839e28e4e536b27bc15122851c759cd1f5845af4a82ae8c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 113, + 545, + 137 + ], + "lines": [ + { + "bbox": [ + 305, + 113, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 305, + 113, + 545, + 137 + ], + "type": "text", + "content": "Table 1: The distortion degree number " + }, + { + "bbox": [ + 305, + 113, + 545, + 137 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 305, + 113, + 545, + 137 + ], + "type": "text", + "content": " in the position map when dividing an " + }, + { + "bbox": [ + 305, + 113, + 545, + 137 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 305, + 113, + 545, + 137 + ], + "type": "text", + "content": " image with " + }, + { + "bbox": [ + 305, + 113, + 545, + 137 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 305, + 113, + 545, + 137 + ], + "type": "text", + "content": " patches." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 157, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 545, + 361 + ], + "type": "text", + "content": "tation for rectification, we first introduce a contrastive loss to supervise the output representation " + }, + { + "bbox": [ + 304, + 157, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{E}_r \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 304, + 157, + 545, + 361 + ], + "type": "text", + "content": ". The motivation of contrastive learning [10] is to associate the query with its positive examples and push away the negative examples. Considering that the distortion pattern of fisheye images is radially distributed (see Fig. 1) and the distortion degree increases with the distance from the image center, we regard the patches with the same distortion degree as positive examples and with different degrees as negative examples, respectively. In this way, the contrastive learning can help the network encode the local distinct distortion pattern in different image regions, regardless of their individual appearance. Concretely, the cross-entropy loss is calculated, representing the probability of the positive examples being selected over the negatives. For the " + }, + { + "bbox": [ + 304, + 157, + 545, + 361 + ], + "type": "inline_equation", + "content": "i^{th}(i = 1,2,\\dots ,N)" + }, + { + "bbox": [ + 304, + 157, + 545, + 361 + ], + "type": "text", + "content": " patch of the input image, the loss function is defined as follows," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 342, + 368, + 545, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 368, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 342, + 368, + 545, + 401 + ], + "type": "interline_equation", + "content": "\\ell_ {i} = - \\log \\frac {\\sum_ {j = 1} ^ {N} \\mathbb {1} _ {[ j \\in N _ {i} ]} \\exp \\left(\\boldsymbol {p} _ {i} \\cdot \\boldsymbol {p} _ {j} / \\tau\\right)}{\\sum_ {k = 1} ^ {N} \\mathbb {1} _ {[ k \\neq i ]} \\exp \\left(\\boldsymbol {p} _ {i} \\cdot \\boldsymbol {p} _ {k} / \\tau\\right)}, \\tag {4}", + "image_path": "b98c115408743f74a1e8b60ac0aa98bfc11e0673437d166054a7b7f4c31fca8e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i\\in \\mathbb{R}^{1\\times D}" + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "text", + "content": " representation in " + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "inline_equation", + "content": "\\pmb{E}_{r}\\in \\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "text", + "content": " is index set of positive examples for the " + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "text", + "content": " patch, " + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{[f(\\cdot)]}\\in \\{0,1\\}" + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "text", + "content": " represents an indicator function evaluating to 1 iff " + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "text", + "content": " denotes a temperature parameter. The total contrastive loss " + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "text", + "content": " is calculated over the " + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 408, + 545, + 470 + ], + "type": "text", + "content": " patches," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 394, + 478, + 545, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 478, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 394, + 478, + 545, + 510 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n}} = \\sum_ {i = 1} ^ {N} \\ell_ {i}. \\tag {5}", + "image_path": "92ee030a20d7ab77c781e51715c4be7921ee92d619d407b7a1322d786473178a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 521, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 521, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 305, + 521, + 545, + 628 + ], + "type": "text", + "content": "Position Loss. As mentioned above, the shuffled patch representations " + }, + { + "bbox": [ + 305, + 521, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{E}_r \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 305, + 521, + 545, + 628 + ], + "type": "text", + "content": " have been labeled with " + }, + { + "bbox": [ + 305, + 521, + 545, + 628 + ], + "type": "inline_equation", + "content": "P_r \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 305, + 521, + 545, + 628 + ], + "type": "text", + "content": " according to their distortion patterns (degrees). To model these local distinct distortion patterns, we further introduce a simple but effective position loss. Specifically, given the patch representation " + }, + { + "bbox": [ + 305, + 521, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{E}_r \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 305, + 521, + 545, + 628 + ], + "type": "text", + "content": ", we apply a layer normalization [54] and a linear projection to estimate their distortion degrees. The loss is defined as a cross-entropy loss," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 335, + 637, + 545, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 637, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 335, + 637, + 545, + 670 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p o s} = - \\sum_ {n = 1} ^ {N} \\sum_ {c = 1} ^ {C _ {t}} \\log \\frac {\\exp \\left(\\hat {\\boldsymbol {x}} _ {n , c}\\right)}{\\sum_ {i = 1} ^ {C _ {t}} \\exp \\left(\\hat {\\boldsymbol {x}} _ {n , i}\\right)} \\boldsymbol {x} _ {n, c}, \\tag {6}", + "image_path": "92bd34c5a19b2645f9a07938fd2de22c69b7d145b95e1c81a279777b0d57018c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " denote the ground truth and predicted confidence, respectively; " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " is the category number of distortion degree in " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "P_r" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " which measures the level of granularity of the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12421" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "learned features. Note that this constraint on the network also facilitates the perception of positives and negatives in the above contrastive learning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 108, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 108, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 108, + 287, + 133 + ], + "type": "text", + "content": "At the pre-training stage, the overall architecture is end-to-end optimized with the following training objectives:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 122, + 142, + 287, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 142, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 122, + 142, + 287, + 156 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p r e} = \\mathcal {L} _ {c o n} + \\mathcal {L} _ {p o s}. \\tag {7}", + "image_path": "95e9ce1d36420e6df3d118b6be78bc13e974ce566f2aa4648a1a523b3fb3d3fd.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 163, + 203, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 203, + 176 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 203, + 176 + ], + "type": "text", + "content": "4.2. Fine-tuning for Rectification" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 182, + 287, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 287, + 230 + ], + "type": "text", + "content": "After pre-training, the network is capable of extracting the fine-grained distortion features. Next, as illustrated in Fig. 2, we initialize the rectification network with the pretrained weights and perform fine-tuning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 231, + 287, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 287, + 326 + ], + "type": "text", + "content": "Specifically, we adopt the same patch embedding module as the pre-training stage, and directly feed the embeddings " + }, + { + "bbox": [ + 46, + 231, + 287, + 326 + ], + "type": "inline_equation", + "content": "\\pmb{E}_o" + }, + { + "bbox": [ + 46, + 231, + 287, + 326 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 231, + 287, + 326 + ], + "type": "inline_equation", + "content": "N_T" + }, + { + "bbox": [ + 46, + 231, + 287, + 326 + ], + "type": "text", + "content": " transformer encoder layers without shuffling. To adapt to the rectification task, we only initialize the first " + }, + { + "bbox": [ + 46, + 231, + 287, + 326 + ], + "type": "inline_equation", + "content": "N_F(N_F < N_T)" + }, + { + "bbox": [ + 46, + 231, + 287, + 326 + ], + "type": "text", + "content": " layers with pre-trained weights, and randomly initialize the remaining layers. The obtained patch representation " + }, + { + "bbox": [ + 46, + 231, + 287, + 326 + ], + "type": "inline_equation", + "content": "\\pmb{E}_r" + }, + { + "bbox": [ + 46, + 231, + 287, + 326 + ], + "type": "text", + "content": " is then passed through two parallel modules for rectification, described below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 329, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 329, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 329, + 287, + 413 + ], + "type": "text", + "content": "Flow Prediction Module. Different from the existing parameter-based and image-based methods, we unwarp a fisheye image with a backward flow field. Given the representation " + }, + { + "bbox": [ + 47, + 329, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\pmb{E}_r \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 47, + 329, + 287, + 413 + ], + "type": "text", + "content": ", we estimate a full scale flow " + }, + { + "bbox": [ + 47, + 329, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\pmb{f}_b \\in \\mathbb{R}^{H \\times W \\times 2}" + }, + { + "bbox": [ + 47, + 329, + 287, + 413 + ], + "type": "text", + "content": ". As illustrated in Fig. 3, with " + }, + { + "bbox": [ + 47, + 329, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\pmb{f}_b = (\\pmb{f}_x, \\pmb{f}_y)" + }, + { + "bbox": [ + 47, + 329, + 287, + 413 + ], + "type": "text", + "content": ", the rectified image " + }, + { + "bbox": [ + 47, + 329, + 287, + 413 + ], + "type": "inline_equation", + "content": "I_c \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 47, + 329, + 287, + 413 + ], + "type": "text", + "content": " is obtained by the warping operation based on bilinear sampling as follows," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 96, + 423, + 287, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 423, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 96, + 423, + 287, + 437 + ], + "type": "interline_equation", + "content": "\\boldsymbol {I} _ {c} (u, v) = \\boldsymbol {I} _ {d} (\\boldsymbol {f} _ {x} (u, v), \\boldsymbol {f} _ {y} (u, v)), \\tag {8}", + "image_path": "55ad69503776cdc98bdd27214b63a51af4d04966f7d4f211a34f880cc2b33db0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 445, + 287, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 481 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 445, + 287, + 481 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 46, + 445, + 287, + 481 + ], + "type": "text", + "content": " is the integer pixel coordinate in rectified image " + }, + { + "bbox": [ + 46, + 445, + 287, + 481 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 46, + 445, + 287, + 481 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 445, + 287, + 481 + ], + "type": "inline_equation", + "content": "(\\pmb{f}_x(u,v),\\pmb{f}_y(u,v))" + }, + { + "bbox": [ + 46, + 445, + 287, + 481 + ], + "type": "text", + "content": " is the predicted decimal pixel coordinate in distorted image " + }, + { + "bbox": [ + 46, + 445, + 287, + 481 + ], + "type": "inline_equation", + "content": "I_{d}" + }, + { + "bbox": [ + 46, + 445, + 287, + 481 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "type": "text", + "content": "To this end, we introduce a learnable upsample module. We first reshape " + }, + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\pmb{E}_r" + }, + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "type": "text", + "content": " to shape " + }, + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\frac{H}{P} \\times \\frac{W}{P} \\times D" + }, + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "type": "text", + "content": ". Then, two convolutional layers (stride 1) are followed and produce a " + }, + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "type": "inline_equation", + "content": "1 / P" + }, + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "type": "text", + "content": " scale warping flow " + }, + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\pmb{f}_m \\in \\mathbb{R}^{\\frac{H}{P} \\times \\frac{W}{P} \\times 2}" + }, + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "type": "text", + "content": ". Next, We use other two convolutional layers to predict a " + }, + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "type": "inline_equation", + "content": "H / P \\times W / P \\times (P \\times P \\times 9)" + }, + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "type": "text", + "content": " mask and perform softmax over the weights of the " + }, + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "type": "text", + "content": " neighborhood of each pixel in " + }, + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\pmb{f}_m" + }, + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "type": "text", + "content": ". Finally, the obtained " + }, + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\frac{H}{P} \\times \\frac{W}{P} \\times P \\times P \\times 2" + }, + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "type": "text", + "content": " map is permuted and reshaped to the full resolution warping flow " + }, + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\pmb{f}_b \\in \\mathbb{R}^{H \\times W \\times 2}" + }, + { + "bbox": [ + 46, + 482, + 287, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "type": "text", + "content": "Boundary Refinement Module. We find that the obtained " + }, + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "type": "text", + "content": " struggles with the serrated boundaries and noisy backgrounds (see Fig. 7). To further improve the image quality, we propose a boundary refinement module (BRM)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "Similar to the flow prediction module, BRM also estimates the weights of neighbor pixels for upsampling. Then, we obtain a confidence map " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{H \\times W \\times 1}" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": ", in which each value denotes the confidence whether a pixel belongs to the rectified region. " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " is further binarized with a threshold " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " to obtain the binary boundary refinement mask " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "M_{b} \\in" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 331, + 69, + 523, + 141 + ], + "blocks": [ + { + "bbox": [ + 331, + 69, + 523, + 141 + ], + "lines": [ + { + "bbox": [ + 331, + 69, + 523, + 141 + ], + "spans": [ + { + "bbox": [ + 331, + 69, + 523, + 141 + ], + "type": "image", + "image_path": "07f6194413800740c8e9b70a811cd1e5077fd904f945801add920b52669bfd4c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 145, + 544, + 169 + ], + "lines": [ + { + "bbox": [ + 305, + 145, + 544, + 169 + ], + "spans": [ + { + "bbox": [ + 305, + 145, + 544, + 169 + ], + "type": "text", + "content": "Figure 3: Visualization of the rectification process of a certain pixel on the rectified image based on warping flow." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 180, + 544, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 180, + 544, + 205 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 544, + 205 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H\\times W\\times 1}" + }, + { + "bbox": [ + 306, + 180, + 544, + 205 + ], + "type": "text", + "content": ", as illustrated in Fig. 2. The final rectified image " + }, + { + "bbox": [ + 306, + 180, + 544, + 205 + ], + "type": "inline_equation", + "content": "\\pmb{I}_r" + }, + { + "bbox": [ + 306, + 180, + 544, + 205 + ], + "type": "text", + "content": " can be obtained by multiplying " + }, + { + "bbox": [ + 306, + 180, + 544, + 205 + ], + "type": "inline_equation", + "content": "\\pmb{I}_c" + }, + { + "bbox": [ + 306, + 180, + 544, + 205 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 306, + 180, + 544, + 205 + ], + "type": "inline_equation", + "content": "M_b" + }, + { + "bbox": [ + 306, + 180, + 544, + 205 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 207, + 544, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 207, + 544, + 231 + ], + "spans": [ + { + "bbox": [ + 306, + 207, + 544, + 231 + ], + "type": "text", + "content": "Training Loss. At the fine-tuning stage, the model is end-to-end optimized with the following training objectives," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 376, + 238, + 544, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 238, + 544, + 251 + ], + "spans": [ + { + "bbox": [ + 376, + 238, + 544, + 251 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f t} = \\mathcal {L} _ {\\text {f l o w}} + \\mathcal {L} _ {\\text {m a s k}}. \\tag {9}", + "image_path": "3107f6462ff9748543dc80bf3d3866096f8c47d638b3a4a516896a1b790f1044.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 256, + 544, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 256, + 544, + 281 + ], + "spans": [ + { + "bbox": [ + 306, + 256, + 544, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{flow}" + }, + { + "bbox": [ + 306, + 256, + 544, + 281 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 306, + 256, + 544, + 281 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 306, + 256, + 544, + 281 + ], + "type": "text", + "content": " distance between the predicted warping flow " + }, + { + "bbox": [ + 306, + 256, + 544, + 281 + ], + "type": "inline_equation", + "content": "\\pmb{f}_b" + }, + { + "bbox": [ + 306, + 256, + 544, + 281 + ], + "type": "text", + "content": " and its given ground truth " + }, + { + "bbox": [ + 306, + 256, + 544, + 281 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{gt}" + }, + { + "bbox": [ + 306, + 256, + 544, + 281 + ], + "type": "text", + "content": " as follows," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 362, + 287, + 544, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 287, + 544, + 300 + ], + "spans": [ + { + "bbox": [ + 362, + 287, + 544, + 300 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f l o w}} = \\left\\| \\boldsymbol {f} _ {b} \\times \\boldsymbol {M} _ {g t} - \\boldsymbol {f} _ {g t} \\right\\| _ {1}, \\tag {10}", + "image_path": "faa45a3351f2ad1aab4cde6b50459b481518bfc776a97b3c406b1ccaa2248cae.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 305, + 544, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 305, + 544, + 329 + ], + "spans": [ + { + "bbox": [ + 305, + 305, + 544, + 329 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 305, + 544, + 329 + ], + "type": "inline_equation", + "content": "M_{gt}" + }, + { + "bbox": [ + 305, + 305, + 544, + 329 + ], + "type": "text", + "content": " denotes the ground truth valid foreground region of the rectified image." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 330, + 544, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 330, + 544, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 330, + 544, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 305, + 330, + 544, + 365 + ], + "type": "text", + "content": " is the cross-entropy loss between the predicted boundary refinement confidence map " + }, + { + "bbox": [ + 305, + 330, + 544, + 365 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 330, + 544, + 365 + ], + "type": "text", + "content": " and its given ground truth " + }, + { + "bbox": [ + 305, + 330, + 544, + 365 + ], + "type": "inline_equation", + "content": "M_{gt}" + }, + { + "bbox": [ + 305, + 330, + 544, + 365 + ], + "type": "text", + "content": " as follows," + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 318, + 373, + 544, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 373, + 544, + 415 + ], + "spans": [ + { + "bbox": [ + 318, + 373, + 544, + 415 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = - \\sum_ {i = 1} ^ {N _ {d}} \\left[ \\boldsymbol {y} _ {i} \\log \\left(\\hat {\\boldsymbol {y}} _ {i}\\right) + \\left(1 - \\boldsymbol {y} _ {i}\\right) \\log \\left(1 - \\hat {\\boldsymbol {y}} _ {i}\\right) \\right], \\tag {11}", + "image_path": "a0b0724db97ee92beb52a9d4b7c4f8d0548c8e20b026c54601b87f8272372340.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 415, + 544, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 415, + 544, + 451 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 544, + 451 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 415, + 544, + 451 + ], + "type": "inline_equation", + "content": "N_{d} = H\\times W" + }, + { + "bbox": [ + 305, + 415, + 544, + 451 + ], + "type": "text", + "content": " is the number of pixels in image " + }, + { + "bbox": [ + 305, + 415, + 544, + 451 + ], + "type": "inline_equation", + "content": "\\pmb{I}_d" + }, + { + "bbox": [ + 305, + 415, + 544, + 451 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 415, + 544, + 451 + ], + "type": "inline_equation", + "content": "\\pmb{y}_i\\in \\{0,1\\}" + }, + { + "bbox": [ + 305, + 415, + 544, + 451 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 415, + 544, + 451 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_i\\in [0,1]" + }, + { + "bbox": [ + 305, + 415, + 544, + 451 + ], + "type": "text", + "content": " denote the ground-truth and the predicted confidence value in " + }, + { + "bbox": [ + 305, + 415, + 544, + 451 + ], + "type": "inline_equation", + "content": "M_{gt}" + }, + { + "bbox": [ + 305, + 415, + 544, + 451 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 415, + 544, + 451 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 415, + 544, + 451 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 460, + 387, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 387, + 473 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 387, + 473 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 479, + 422, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 422, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 422, + 491 + ], + "type": "text", + "content": "5.1. Dataset and Metrics" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 498, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 498, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 305, + 498, + 545, + 629 + ], + "type": "text", + "content": "We train and evaluate our method according to the conventions in the literature. Following method [46, 6, 58, 37, 38], we establish a dataset using original images from the Place365 dataset [63]. We use the polynomial model (Eq. (3)) to generate the dataset and the first four distortion parameters are used (" + }, + { + "bbox": [ + 305, + 498, + 545, + 629 + ], + "type": "inline_equation", + "content": "N_{k} = 4" + }, + { + "bbox": [ + 305, + 498, + 545, + 629 + ], + "type": "text", + "content": "), which is enough to fit most of the real-world application scenarios. The synthetic dataset contains multiple annotations for each distorted fisheye image, including distortion-free image, warping flow, and distortion parameters. We synthesize " + }, + { + "bbox": [ + 305, + 498, + 545, + 629 + ], + "type": "inline_equation", + "content": "100\\mathrm{k}" + }, + { + "bbox": [ + 305, + 498, + 545, + 629 + ], + "type": "text", + "content": " images for training and " + }, + { + "bbox": [ + 305, + 498, + 545, + 629 + ], + "type": "inline_equation", + "content": "5\\mathrm{k}" + }, + { + "bbox": [ + 305, + 498, + 545, + 629 + ], + "type": "text", + "content": " images for evaluation, respectively." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 630, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 630, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 630, + 545, + 712 + ], + "type": "text", + "content": "In fisheye image rectification, Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM) are the two most popular evaluation metrics in existing works. Given a rectified image and a reference distortion-free one, PSNR measures the detailed quality differences while SSIM assesses their structural differences. Besides, we use Frechet Inception Distance (FID) to evaluate the distribution difference." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12422" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 96, + 214, + 155 + ], + "blocks": [ + { + "bbox": [ + 56, + 74, + 212, + 94 + ], + "lines": [ + { + "bbox": [ + 56, + 74, + 212, + 94 + ], + "spans": [ + { + "bbox": [ + 56, + 74, + 212, + 94 + ], + "type": "text", + "content": "(a) Pre-training. Joint pre-training obtains remarkable performance gain." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 96, + 214, + 155 + ], + "lines": [ + { + "bbox": [ + 59, + 96, + 214, + 155 + ], + "spans": [ + { + "bbox": [ + 59, + 96, + 214, + 155 + ], + "type": "table", + "html": "
LconLposPSNRSSIM
20.330.82
✓20.700.83
✓21.300.84
✓✓22.470.86
", + "image_path": "981f97b73596b14fa21ea7c0e211bb1056e3241547db5ccca90f887fbc910c90.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 223, + 96, + 386, + 155 + ], + "blocks": [ + { + "bbox": [ + 221, + 74, + 386, + 93 + ], + "lines": [ + { + "bbox": [ + 221, + 74, + 386, + 93 + ], + "spans": [ + { + "bbox": [ + 221, + 74, + 386, + 93 + ], + "type": "text", + "content": "(b) Fine-tuning way. Both fine-tuning strategies work and training from-scratch is better." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 223, + 96, + 386, + 155 + ], + "lines": [ + { + "bbox": [ + 223, + 96, + 386, + 155 + ], + "spans": [ + { + "bbox": [ + 223, + 96, + 386, + 155 + ], + "type": "table", + "html": "
settingsPSNRSSIM
w/o pre-training20.330.82
w/ pre-training: frozen encoder21.120.85
unfrozen encoder22.470.86
", + "image_path": "67298285341caf53a37e71f533d0835c7d41c314a3ad5ddca01506af82e26c26.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 395, + 96, + 536, + 133 + ], + "blocks": [ + { + "bbox": [ + 394, + 74, + 536, + 93 + ], + "lines": [ + { + "bbox": [ + 394, + 74, + 536, + 93 + ], + "spans": [ + { + "bbox": [ + 394, + 74, + 536, + 93 + ], + "type": "text", + "content": "(c) Shuffle. Without the shuffle, the performance drops to the baseline level." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 395, + 96, + 536, + 133 + ], + "lines": [ + { + "bbox": [ + 395, + 96, + 536, + 133 + ], + "spans": [ + { + "bbox": [ + 395, + 96, + 536, + 133 + ], + "type": "table", + "html": "
settingsPSNRSSIM
w/o shuffle20.280.82
w/ shuffle22.470.86
", + "image_path": "d41ce37276a035d7e6334913a3e65f94e0492e7b6e7b9003be717d1cd2c7caf6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 77, + 198, + 213, + 244 + ], + "blocks": [ + { + "bbox": [ + 74, + 166, + 213, + 195 + ], + "lines": [ + { + "bbox": [ + 74, + 166, + 213, + 195 + ], + "spans": [ + { + "bbox": [ + 74, + 166, + 213, + 195 + ], + "type": "text", + "content": "(d) Prediction target. Compared with distortion parameter and image, warping flow is the most effective target." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 198, + 213, + 244 + ], + "lines": [ + { + "bbox": [ + 77, + 198, + 213, + 244 + ], + "spans": [ + { + "bbox": [ + 77, + 198, + 213, + 244 + ], + "type": "table", + "html": "
targetsPSNRSSIM
parameter21.090.82
image20.500.69
flow22.470.86
", + "image_path": "3a432a8f9e986b024948722e1d202e47320d9a39b5cb636c0f72428c4f2dddf8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 220, + 166, + 381, + 195 + ], + "lines": [ + { + "bbox": [ + 220, + 166, + 381, + 195 + ], + "spans": [ + { + "bbox": [ + 220, + 166, + 381, + 195 + ], + "type": "text", + "content": "(e) Patch size. Reducing patch size works better. To strike a balance between accuracy and efficiency, we choose " + }, + { + "bbox": [ + 220, + 166, + 381, + 195 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 220, + 166, + 381, + 195 + ], + "type": "text", + "content": " by default." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 223, + 198, + 382, + 253 + ], + "blocks": [ + { + "bbox": [ + 223, + 198, + 382, + 253 + ], + "lines": [ + { + "bbox": [ + 223, + 198, + 382, + 253 + ], + "spans": [ + { + "bbox": [ + 223, + 198, + 382, + 253 + ], + "type": "table", + "html": "
patchPSNRSSIMFPS
825.820.9014.93
1622.470.8615.51
3220.230.8114.64
6418.650.7414.19
", + "image_path": "9e6de42b160e38873f2e54fec4a26f7b5edb6aac745f35db9a519a2d2d826bb7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 389, + 198, + 525, + 244 + ], + "blocks": [ + { + "bbox": [ + 387, + 167, + 525, + 196 + ], + "lines": [ + { + "bbox": [ + 387, + 167, + 525, + 196 + ], + "spans": [ + { + "bbox": [ + 387, + 167, + 525, + 196 + ], + "type": "text", + "content": "(f) Architecture setting. Several different architectures are set up. Our scheme performs better." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 389, + 198, + 525, + 244 + ], + "lines": [ + { + "bbox": [ + 389, + 198, + 525, + 244 + ], + "spans": [ + { + "bbox": [ + 389, + 198, + 525, + 244 + ], + "type": "table", + "html": "
settingsPSNRSSIM
w/o BRM18.010.68
w/o FPM22.110.85
full22.470.86
", + "image_path": "a95959e93fe9d11189ad90f007dcfc81112d6a49bd46f2092206350a7594d933.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 260, + 545, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 260, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 545, + 285 + ], + "type": "text", + "content": "Table 2: Ablation experiments of our method. The default setting is: the patch size is " + }, + { + "bbox": [ + 46, + 260, + 545, + 285 + ], + "type": "inline_equation", + "content": "{16} \\times {16}" + }, + { + "bbox": [ + 46, + 260, + 545, + 285 + ], + "type": "text", + "content": " and we import 8 pre-trained layers for fine-tuning. Default settings are all marked in gray ." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 299, + 144, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 299, + 144, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 299, + 144, + 312 + ], + "type": "text", + "content": "5.2. Implementation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 318, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 378 + ], + "type": "text", + "content": "Model Settings. The image size " + }, + { + "bbox": [ + 46, + 318, + 287, + 378 + ], + "type": "inline_equation", + "content": "(H,W)" + }, + { + "bbox": [ + 46, + 318, + 287, + 378 + ], + "type": "text", + "content": " is (256, 256). The latent vector size " + }, + { + "bbox": [ + 46, + 318, + 287, + 378 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 318, + 287, + 378 + ], + "type": "text", + "content": " in transformer encoder layer [54] is 256 and the total layer number " + }, + { + "bbox": [ + 46, + 318, + 287, + 378 + ], + "type": "inline_equation", + "content": "N_{T}" + }, + { + "bbox": [ + 46, + 318, + 287, + 378 + ], + "type": "text", + "content": " is 10. The threshold " + }, + { + "bbox": [ + 46, + 318, + 287, + 378 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 318, + 287, + 378 + ], + "type": "text", + "content": " for binarizing the confidence map in the boundary refinement module is empirically set as 0.5." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 382, + 287, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 287, + 466 + ], + "type": "text", + "content": "Training Details. We use 100k images for pre-training and 50k images for fine-tuning. For contrastive learning, we set the temperature " + }, + { + "bbox": [ + 46, + 382, + 287, + 466 + ], + "type": "inline_equation", + "content": "\\tau = 0.07" + }, + { + "bbox": [ + 46, + 382, + 287, + 466 + ], + "type": "text", + "content": " (Eq. (4)). We utilize Adam [31] as the optimizer and one-cycle policy [50, 52] with a maximum learning rate of " + }, + { + "bbox": [ + 46, + 382, + 287, + 466 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 382, + 287, + 466 + ], + "type": "text", + "content": ". Both stages are trained for 65 epochs with a batch size of 64. Two NVIDIA GeForce RTX 2080Ti GPUs are used to train the network." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 475, + 140, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 140, + 488 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 140, + 488 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 494, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 287, + 518 + ], + "type": "text", + "content": "We conduct ablations to verify the effectiveness of the core settings and components in our method." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 521, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 714 + ], + "type": "text", + "content": "Impact of Self-supervised Pre-training. The key idea of this work is the self-supervised distortion representation learning for fisheye images. Table 2 (a) shows the studies on the proposed pretext task. Without any pre-training, we can still provide a strong baseline, with a PSNR of 20.33 and a SSIM of 0.82. This can be attributed to the powerful representative ability of transformer [54]. Then, compared with the baseline, the performance improves by pre-training with an independent contrastive loss " + }, + { + "bbox": [ + 46, + 521, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 46, + 521, + 287, + 714 + ], + "type": "text", + "content": " or position loss " + }, + { + "bbox": [ + 46, + 521, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pos}" + }, + { + "bbox": [ + 46, + 521, + 287, + 714 + ], + "type": "text", + "content": ". Furthermore, the unified pre-training boosts the rectification performance significantly. These results verify the effectiveness of our self-supervised representation learning strategy. Note that the two pre-text tasks promote each other. In Fig. 4, we provide the training loss plots of fine-tuning under the setting w/ and w/o the pre-training. From the training perspective, we can observe that our pre-training" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 312, + 304, + 422, + 388 + ], + "blocks": [ + { + "bbox": [ + 312, + 304, + 422, + 388 + ], + "lines": [ + { + "bbox": [ + 312, + 304, + 422, + 388 + ], + "spans": [ + { + "bbox": [ + 312, + 304, + 422, + 388 + ], + "type": "image", + "image_path": "0239745c14468bb5fe721bc3ea80f79ab6da1313d759164724481fbb74cd1610.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 344, + 393, + 387, + 402 + ], + "lines": [ + { + "bbox": [ + 344, + 393, + 387, + 402 + ], + "spans": [ + { + "bbox": [ + 344, + 393, + 387, + 402 + ], + "type": "text", + "content": "(a) flow loss" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 428, + 304, + 541, + 388 + ], + "blocks": [ + { + "bbox": [ + 428, + 304, + 541, + 388 + ], + "lines": [ + { + "bbox": [ + 428, + 304, + 541, + 388 + ], + "spans": [ + { + "bbox": [ + 428, + 304, + 541, + 388 + ], + "type": "image", + "image_path": "0115f2457a924ff0917837b2e1cb0280510c62da158ef65252b6f0fd8cee5721.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 393, + 506, + 402 + ], + "lines": [ + { + "bbox": [ + 460, + 393, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 460, + 393, + 506, + 402 + ], + "type": "text", + "content": "(b) mask loss" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 409, + 545, + 445 + ], + "lines": [ + { + "bbox": [ + 305, + 409, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 409, + 545, + 445 + ], + "type": "text", + "content": "Figure 4: Training loss plots during the fine-tuning stage. With the distortion representation pre-training, the training errors of the rectification " + }, + { + "bbox": [ + 305, + 409, + 545, + 445 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{flow}" + }, + { + "bbox": [ + 305, + 409, + 545, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 409, + 545, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask})" + }, + { + "bbox": [ + 305, + 409, + 545, + 445 + ], + "type": "text", + "content": " are reduced." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 460, + 487, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 487, + 472 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 487, + 472 + ], + "type": "text", + "content": "effectively reduces the losses of the baseline." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 476, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 545, + 561 + ], + "type": "text", + "content": "Impact of Fine-tuning Way. During the fine-tuning stage, we train the model from scratch. An alternative way is to freeze the pre-trained encoder and only update the rest weights of the model. As shown in Table 2 (b), both fine-tuning ways produce better performance than the baseline (w/o pre-training), illustrating the effectiveness of our learned representations and our default fine-tuning strategy." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 565, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 545, + 673 + ], + "type": "text", + "content": "Impact of Shuffle Manipulation. Shuffle strategy is important for our method. During the pre-training stage, we arbitrarily shuffle the patches, to avoid the directed association between the fixed positional embedding and the specific distortion pattern/degree. As shown in Table 2 (c), without the shuffle strategy, the performance degrades to the baseline level in Table 2 (a), which means that the pretraining does not learn any useful representation for the downstream rectification." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Impact of Reconstruction Target. Our SimFIR estimates a warping flow [21, 19] to resample the pixels from distorted images for rectification. In contrast, other" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12423" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 69, + 108, + 186 + ], + "blocks": [ + { + "bbox": [ + 51, + 69, + 108, + 186 + ], + "lines": [ + { + "bbox": [ + 51, + 69, + 108, + 186 + ], + "spans": [ + { + "bbox": [ + 51, + 69, + 108, + 186 + ], + "type": "image", + "image_path": "f9a603b1f68151d8be30e1952b56d89969d9178ee4adeddb30493762e622fc1e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 188, + 97, + 198 + ], + "lines": [ + { + "bbox": [ + 61, + 188, + 97, + 198 + ], + "spans": [ + { + "bbox": [ + 61, + 188, + 97, + 198 + ], + "type": "text", + "content": "Distorted" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 110, + 70, + 167, + 127 + ], + "blocks": [ + { + "bbox": [ + 110, + 70, + 167, + 127 + ], + "lines": [ + { + "bbox": [ + 110, + 70, + 167, + 127 + ], + "spans": [ + { + "bbox": [ + 110, + 70, + 167, + 127 + ], + "type": "image", + "image_path": "a6fddc3642451e0de3e927afbcd858f8877a788454d221e6039fe3138d73f4af.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 110, + 129, + 166, + 186 + ], + "blocks": [ + { + "bbox": [ + 110, + 129, + 166, + 186 + ], + "lines": [ + { + "bbox": [ + 110, + 129, + 166, + 186 + ], + "spans": [ + { + "bbox": [ + 110, + 129, + 166, + 186 + ], + "type": "image", + "image_path": "3c2e28f18320a65836fc77eb0f417a58c7a1d28d2d458fa308938c62de7a36d9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 188, + 164, + 198 + ], + "lines": [ + { + "bbox": [ + 113, + 188, + 164, + 198 + ], + "spans": [ + { + "bbox": [ + 113, + 188, + 164, + 198 + ], + "type": "text", + "content": "Predict image" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 168, + 70, + 225, + 127 + ], + "blocks": [ + { + "bbox": [ + 168, + 70, + 225, + 127 + ], + "lines": [ + { + "bbox": [ + 168, + 70, + 225, + 127 + ], + "spans": [ + { + "bbox": [ + 168, + 70, + 225, + 127 + ], + "type": "image", + "image_path": "969255af5d368595bc7389d17911c1c767857834e9f9833044df894c5506ff77.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 168, + 129, + 225, + 185 + ], + "blocks": [ + { + "bbox": [ + 168, + 129, + 225, + 185 + ], + "lines": [ + { + "bbox": [ + 168, + 129, + 225, + 185 + ], + "spans": [ + { + "bbox": [ + 168, + 129, + 225, + 185 + ], + "type": "image", + "image_path": "22ced87002969c48ef0968bac6cc4f58f5be16b370d49f9c6ac47e86bed38b10.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 188, + 220, + 198 + ], + "lines": [ + { + "bbox": [ + 175, + 188, + 220, + 198 + ], + "spans": [ + { + "bbox": [ + 175, + 188, + 220, + 198 + ], + "type": "text", + "content": "Predict flow" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 227, + 70, + 285, + 127 + ], + "blocks": [ + { + "bbox": [ + 227, + 70, + 285, + 127 + ], + "lines": [ + { + "bbox": [ + 227, + 70, + 285, + 127 + ], + "spans": [ + { + "bbox": [ + 227, + 70, + 285, + 127 + ], + "type": "image", + "image_path": "fe00e8adac9797956dc0f5bc2969d2880bb48d5a25d1b933d230cde554297c99.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 227, + 129, + 284, + 185 + ], + "blocks": [ + { + "bbox": [ + 227, + 129, + 284, + 185 + ], + "lines": [ + { + "bbox": [ + 227, + 129, + 284, + 185 + ], + "spans": [ + { + "bbox": [ + 227, + 129, + 284, + 185 + ], + "type": "image", + "image_path": "48e7e78695ec26fc399ec072364272131c4e5b11fa1a3d04611bdec73b5d88aa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 188, + 263, + 198 + ], + "lines": [ + { + "bbox": [ + 249, + 188, + 263, + 198 + ], + "spans": [ + { + "bbox": [ + 249, + 188, + 263, + 198 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 75, + 255, + 257, + 373 + ], + "blocks": [ + { + "bbox": [ + 46, + 208, + 287, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 208, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 287, + 243 + ], + "type": "text", + "content": "Figure 5: Examples to illustrate the impact of the reconstruction target. For our method, predicting flow to unwarp the distorted images preserves more image details." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 255, + 257, + 373 + ], + "lines": [ + { + "bbox": [ + 75, + 255, + 257, + 373 + ], + "spans": [ + { + "bbox": [ + 75, + 255, + 257, + 373 + ], + "type": "image", + "image_path": "1ac6ee09ac80bf311091a481b9454fb01a74e55b449a8667a2943ba18b608a3f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 380, + 287, + 417 + ], + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 417 + ], + "type": "text", + "content": "Figure 6: Running efficiency comparison between our SimFIR and the state-of-the-art PCN [58] on processing different resolution fisheye images." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 433, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 613 + ], + "type": "text", + "content": "pipelines predict the distortion parameters [46, 6, 60, 57] or regress the rectified images directly [39, 58, 38]. We implement the above two pipelines into our method with the following experiments: 1) meaning the output representations " + }, + { + "bbox": [ + 46, + 433, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\pmb{E}_r \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 46, + 433, + 287, + 613 + ], + "type": "text", + "content": " into a vector and projecting it to regress the distortion parameters, and 2) reshaping the output " + }, + { + "bbox": [ + 46, + 433, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\pmb{E}_r" + }, + { + "bbox": [ + 46, + 433, + 287, + 613 + ], + "type": "text", + "content": " to form a reconstructed image, following MAE [28]. As shown in Table 2 (d), the performance drops with these two targets. This is because for distortion rectification, predicting the deformation instead of the target image itself is simpler [35, 20], and can preserve more image details (see Fig. 5). The network can focus on the spatial transformation between two images, without needing to restore the texture details. Besides, it is difficult to extract complex nonlinear model parameters in different intervals [58]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Impact of Patch Size. Table 2 (e) studies the impact of patch size. We can observe that a smaller patch steadily produces a higher performance. The reason is that, as shown in Table 1, a smaller patch setting increases the subdivision of distortion patterns, which helps to extract more fine-grained distortion representations. However, a smaller patch will also introduce more patches when dividing an input image. To balance the computational cost, we choose a " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "16 \\times 16" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 310, + 69, + 367, + 127 + ], + "blocks": [ + { + "bbox": [ + 310, + 69, + 367, + 127 + ], + "lines": [ + { + "bbox": [ + 310, + 69, + 367, + 127 + ], + "spans": [ + { + "bbox": [ + 310, + 69, + 367, + 127 + ], + "type": "image", + "image_path": "a058b8b71b3d12852876e30a0c0d4080ff2e7f5aca9712954cead75bcc3ce29d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 310, + 129, + 367, + 186 + ], + "blocks": [ + { + "bbox": [ + 310, + 129, + 367, + 186 + ], + "lines": [ + { + "bbox": [ + 310, + 129, + 367, + 186 + ], + "spans": [ + { + "bbox": [ + 310, + 129, + 367, + 186 + ], + "type": "image", + "image_path": "0bc61729e0d3d8bca1668c505114302f472ca9cec1c7b23b588af4ab44edf9a9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 188, + 356, + 198 + ], + "lines": [ + { + "bbox": [ + 321, + 188, + 356, + 198 + ], + "spans": [ + { + "bbox": [ + 321, + 188, + 356, + 198 + ], + "type": "text", + "content": "Distorted" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 368, + 70, + 425, + 127 + ], + "blocks": [ + { + "bbox": [ + 368, + 70, + 425, + 127 + ], + "lines": [ + { + "bbox": [ + 368, + 70, + 425, + 127 + ], + "spans": [ + { + "bbox": [ + 368, + 70, + 425, + 127 + ], + "type": "image", + "image_path": "a2061970f951d6a84b8f23d5581b9203e88cdae9d01618b67ead7edd2902f2ed.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 368, + 129, + 425, + 186 + ], + "blocks": [ + { + "bbox": [ + 368, + 129, + 425, + 186 + ], + "lines": [ + { + "bbox": [ + 368, + 129, + 425, + 186 + ], + "spans": [ + { + "bbox": [ + 368, + 129, + 425, + 186 + ], + "type": "image", + "image_path": "712eefc95a89ff2545eae150374b376884f971520e7c8a5eb64268616d61d52d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 188, + 416, + 198 + ], + "lines": [ + { + "bbox": [ + 378, + 188, + 416, + 198 + ], + "spans": [ + { + "bbox": [ + 378, + 188, + 416, + 198 + ], + "type": "text", + "content": "w/o BRM" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 205, + 545, + 242 + ], + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 242 + ], + "type": "text", + "content": "Figure 7: Examples to illustrate the impact of the boundary refinement module (abbreviated as \"BRM\"). With the help of BRM, the boundary regions are well rectified." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 427, + 70, + 484, + 127 + ], + "blocks": [ + { + "bbox": [ + 427, + 70, + 484, + 127 + ], + "lines": [ + { + "bbox": [ + 427, + 70, + 484, + 127 + ], + "spans": [ + { + "bbox": [ + 427, + 70, + 484, + 127 + ], + "type": "image", + "image_path": "78c6d3bc826a12e3a30179d6dc0ebcddd61965459e7fc95f96e8e451d515288a.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 427, + 129, + 484, + 186 + ], + "blocks": [ + { + "bbox": [ + 427, + 129, + 484, + 186 + ], + "lines": [ + { + "bbox": [ + 427, + 129, + 484, + 186 + ], + "spans": [ + { + "bbox": [ + 427, + 129, + 484, + 186 + ], + "type": "image", + "image_path": "834b23cc52e9216db515b5e9ae0de3e2136c14f8a0566217440760c7feaee1b2.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 188, + 471, + 198 + ], + "lines": [ + { + "bbox": [ + 440, + 188, + 471, + 198 + ], + "spans": [ + { + "bbox": [ + 440, + 188, + 471, + 198 + ], + "type": "text", + "content": "w/BRM" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 485, + 70, + 543, + 127 + ], + "blocks": [ + { + "bbox": [ + 485, + 70, + 543, + 127 + ], + "lines": [ + { + "bbox": [ + 485, + 70, + 543, + 127 + ], + "spans": [ + { + "bbox": [ + 485, + 70, + 543, + 127 + ], + "type": "image", + "image_path": "80ab87c44fdfa1d9e464fa09527d3f74dfd74e390c0e24f0e256dfba54a06e13.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 485, + 129, + 543, + 186 + ], + "blocks": [ + { + "bbox": [ + 485, + 129, + 543, + 186 + ], + "lines": [ + { + "bbox": [ + 485, + 129, + 543, + 186 + ], + "spans": [ + { + "bbox": [ + 485, + 129, + 543, + 186 + ], + "type": "image", + "image_path": "27cb2f89cdf5c3f670fd791524e2428c0e3c9d25cd942d87ab164d88aba8fc56.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 508, + 188, + 520, + 198 + ], + "lines": [ + { + "bbox": [ + 508, + 188, + 520, + 198 + ], + "spans": [ + { + "bbox": [ + 508, + 188, + 520, + 198 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "table", + "bbox": [ + 314, + 252, + 538, + 348 + ], + "blocks": [ + { + "bbox": [ + 314, + 252, + 538, + 348 + ], + "lines": [ + { + "bbox": [ + 314, + 252, + 538, + 348 + ], + "spans": [ + { + "bbox": [ + 314, + 252, + 538, + 348 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑FID ↓
SC [48]10.140.32229.60
DeepCalib [6]13.280.51136.73
DR-GAN [38]18.660.7074.56
DDM [39]17.510.66114.66
MLC [37]19.330.7355.97
PCN [58]20.530.7630.64
Ours22.470.8617.23
", + "image_path": "24657aa374c5cadc4938019017e3e09a77aa340d59352396845611e2583b7664.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "table_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 357, + 545, + 393 + ], + "lines": [ + { + "bbox": [ + 304, + 357, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 545, + 393 + ], + "type": "text", + "content": "Table 3: Comparison between the state-of-the-art methods and the proposed method on the test set. “↑” indicates the higher the better and “↓” means the opposite." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 407, + 509, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 509, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 509, + 418 + ], + "type": "text", + "content": "patch size as the default setting in our final model." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 422, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 517 + ], + "type": "text", + "content": "Impact of Boundary Refinement Module. In our SimFIR, the BRM is used to smooth the serrated boundaries and remove the noisy backgrounds of the coarse rectified image " + }, + { + "bbox": [ + 304, + 422, + 545, + 517 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 304, + 422, + 545, + 517 + ], + "type": "text", + "content": " (see Sec. 4.2). As shown in Table 2 (f), with the proposed BRM, performances are remarkably improved. Moreover, we provide two examples to visualize the improvement from the BRM in Fig. 7. As we can see, the boundary regions are well rectified with the help of BRM." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 520, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 591 + ], + "type": "text", + "content": "Impact of Flow Prediction Module. We compare the bilinear upsampling with our learnable upsampling used for FPM. Table 2 (f) shows that our learnable FPM works better than the module based on the bilinear upsampling. This is likely because the coarse bilinear upsampling operation can not recover the small deformations." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 306, + 599, + 528, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 528, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 528, + 612 + ], + "type": "text", + "content": "5.4. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Quantitative Comparison. In this section, we compare our SimFIR with the recent state-of-the-art rectification methods, including SC [48], DeepCalib [6], DR-GAN [38], DDM [39], MLC [37], and PCN [58]. The quantitative results are shown in Table 3. As we can see, our method achieves state-of-the-art performance on all metrics. Note that for SC [48], DeepCalib [6], and PCN [58], we calculate the results based on their released codes. For DR-" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12424" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 534, + 186 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 534, + 186 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 534, + 186 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 534, + 186 + ], + "type": "image", + "image_path": "a57ce5860ee7a3020db8c6f4e5a2d39cb4f66c77db246a9dd1fe044a76b7225c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 191, + 547, + 217 + ], + "lines": [ + { + "bbox": [ + 46, + 191, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 191, + 547, + 217 + ], + "type": "text", + "content": "Figure 8: Qualitative results on the synthesized test set. For each comparison, we show the distorted image, the rectified results of SC [48], DeepCalib [6], DR-GAN [38], DDM [39], MLC [37], PCN [58], our method, and ground truth." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 224, + 534, + 396 + ], + "blocks": [ + { + "bbox": [ + 58, + 224, + 534, + 396 + ], + "lines": [ + { + "bbox": [ + 58, + 224, + 534, + 396 + ], + "spans": [ + { + "bbox": [ + 58, + 224, + 534, + 396 + ], + "type": "image", + "image_path": "dfb1360f6737ba4235b694bee8ef69d7d51504e14cff0dec84ec488f54fdfc5f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 401, + 546, + 426 + ], + "lines": [ + { + "bbox": [ + 46, + 401, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 546, + 426 + ], + "type": "text", + "content": "Figure 9: Qualitative results on real-world fisheye images. For each comparison, we show the distorted image, the rectified results of SC [48], Blind [36], DeepCalib [6], DR-GAN [38], DDM [39], MLC [37], PCN [58], and our method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 446, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 446, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 287, + 471 + ], + "type": "text", + "content": "GAN [38], DDM [39], and MLC [37], we obtain the rectified test set from the authors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 479, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 538 + ], + "type": "text", + "content": "Qualitative Comparison. To better demonstrate the effectiveness of our proposed method, we further conduct qualitative comparisons with existing methods [48, 6, 38, 39, 37, 58] on the synthetic test set. As present in Fig. 8, the rectified images of our method show less distortion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": "Application to Real-world Fisheye Images. Additional experiments are conducted on real-world fisheye images. In this evaluation, we capture the images with several fisheye lenses of different FoV. As shown in Fig. 9, although there exists a gap between synthetic and real-world data, our method still shows strong generalization ability. Compared with SC [48], our method guarantees the integrity of the recovered content. Compared with Blind [36], DeepCalib [6], DR-GAN [38], DDM [39], MLC [37], and PCN [58], our rectified images show less distortion. Note that in Fig. 9, the resolution of the input distorted fisheye images is much higher (up to " + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "inline_equation", + "content": "2454 \\times 2454" + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": ") than that in the test set (256 " + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": " 256). Compared with DR-GAN [38], DDM [39], and PCN [58], our method preserves more texture details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 446, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 545, + 590 + ], + "type": "text", + "content": "Efficiency Comparison. Next, we compare the running efficiency of our method and the state-of-the-art method PCN [58] on processing images with various resolutions. As shown in Fig. 6, on different resolutions, our method shows higher efficiency. The reason is that PCN [58] directly regresses the rectified image with a fully convolutional network [47], like DR-GAN [38], and DDM [39]. In contrast, our SimFIR infers a warping flow at the fixed resolution " + }, + { + "bbox": [ + 304, + 446, + 545, + 590 + ], + "type": "inline_equation", + "content": "(256 \\times 256)" + }, + { + "bbox": [ + 304, + 446, + 545, + 590 + ], + "type": "text", + "content": " and interpolates it to the original high resolution same as the input image for unwarping. Such a strategy ensures efficiency and does not demand any extra training at a certain scale like image-based methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 607, + 379, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 607, + 379, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 607, + 379, + 620 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "In this work, we present an effective self-supervised representation learning paradigm for fisheye images. We associate different image regions with their specific distortion patterns and design a distortion-aware pretext task for their learning. With the pre-training, the network extracts the fine-grained distortion representation. Extensive experiments are conducted on our proposed dataset, and the trans" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12425" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "fer performance in the downstream rectification task verifies the effectiveness of the learned representations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 101, + 287, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 101, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 47, + 101, + 287, + 161 + ], + "type": "text", + "content": "Acknowledgements This work was supported by NSFC under Contract 61836011 and 62021001. It was also supported by the GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, and the Supercomputing Center of the USTC." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 171, + 106, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 171, + 106, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 171, + 106, + 183 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 191, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 191, + 287, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 287, + 234 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 287, + 234 + ], + "type": "text", + "content": "[1] Kohei Aso, Dong-Hyun Hwang, and Hideki Koike. Portable 3D human pose estimation for human-human interaction using a chest-mounted fisheye camera. In AHs, pages 116-120, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 235, + 287, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 235, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 235, + 287, + 269 + ], + "type": "text", + "content": "[2] Joao Barreto, Jose Roquette, Peter Sturm, and Fernando Fonseca. Automatic camera calibration applied to medical endoscopy. In BMVC, pages 1-10, 2009." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 287, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 287, + 302 + ], + "type": "text", + "content": "[3] João Pedro Barreto and Helder Araujo. Geometric properties of central catadioptric line images and their application in calibration. TPAMI, 27(8):1327-1333, 2005." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "type": "text", + "content": "[4] João Pedro Barreto and Kostas Daniilidis. Fundamental matrix for cameras with radial distortion. In ICCV, volume 1, pages 625-632, 2005." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 335, + 287, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 287, + 368 + ], + "type": "text", + "content": "[5] Gregor Blott, Masato Takami, and Christian Heipke. Semantic segmentation of fisheye images. In ECCVW, pages 0-0, 2018." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "text", + "content": "[6] Oleksandr Bogdan, Viktor Eckstein, Francois Rameau, and Jean-Charles Bazin. DeepCalib: a deep learning approach for automatic intrinsic calibration of wide field-of-view cameras. In CVMP, pages 1–10, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 413, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 287, + 446 + ], + "type": "text", + "content": "[7] Faisal Bukhari and Matthew N Dailey. Automatic radial distortion estimation from a single image. Journal of Mathematical Imaging and Vision, 45(1):31-45, 2013." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "type": "text", + "content": "[8] Gyanesh Chander, Brian L Markham, and Dennis L Helder. Summary of current radiometric calibration coefficients for Landsat MSS, TM, ETM+, and EO-1 ALI sensors. Remote Sensing of Environment, 113(5):893–903, 2009." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 491, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 287, + 525 + ], + "type": "text", + "content": "[9] Chun-Hao Chao, Pin-Lun Hsu, Hung-Yi Lee, and Yu-Chiang Frank Wang. Self-supervised deep learning for fisheye image rectification. In ICASSP, pages 2248-2252, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 525, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 287, + 558 + ], + "type": "text", + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 558, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 287, + 591 + ], + "type": "text", + "content": "[11] Jonathan Courbon, Youcef Mezouar, Laurent Eckt, and Philippe Martinet. A generic fisheye camera model for robotic applications. In IROS, pages 1683-1688, 2007." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 591, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 287, + 635 + ], + "type": "text", + "content": "[12] Jonathan Courbon, Youcef Mezouar, and Philippe Martinet. Evaluation of the unified model of the sphere for fisheye cameras in robotic applications. Advanced Robotics, 26(8-9):947-967, 2012." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 635, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 287, + 679 + ], + "type": "text", + "content": "[13] Liuyuan Deng, Ming Yang, Yeqiang Qian, Chunxiang Wang, and Bing Wang. CNN based semantic segmentation for urban traffic scenes using fisheye camera. In IV, pages 231-236, 2017." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 680, + 287, + 713 + ], + "type": "text", + "content": "[14] Frederic Devernay and Olivier Faugeras. Straight lines have to be straight. Machine Vision and Applications, 13(1):14-24, 2001." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 138 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 139, + 546, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 139, + 546, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 139, + 546, + 171 + ], + "type": "text", + "content": "[16] Richard O Duda and Peter E Hart. Use of the hough transformation to detect lines and curves in pictures. CACM, 15(1):11-15, 1972." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 172, + 546, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 546, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 546, + 205 + ], + "type": "text", + "content": "[17] Jinlong Fan, Jing Zhang, Stephen J Maybank, and Dacheng Tao. Wide-angle image rectification: a survey. IJCV, 130(3):747-776, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 205, + 546, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 205, + 546, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 205, + 546, + 249 + ], + "type": "text", + "content": "[18] Jinlong Fan, Jing Zhang, and Dacheng Tao. SIR: Self-supervised image rectification via seeing the same scene from multiple different lenses. arXiv preprint arXiv:2011.14611, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 249, + 546, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 249, + 546, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 249, + 546, + 282 + ], + "type": "text", + "content": "[19] Hao Feng, Shaokai Liu, Jiajun Deng, Wengang Zhou, and Houqiang Li. Deep unrestricted document image rectification. arXiv preprint arXiv:2304.08796, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 282, + 546, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 282, + 546, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 282, + 546, + 326 + ], + "type": "text", + "content": "[20] Hao Feng, Wengang Zhou, Jiajun Deng, Qi Tian, and Houqiang Li. DocScanner: Robust document image rectification with progressive learning. arXiv preprint arXiv:2110.14968, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 326, + 546, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 326, + 546, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 326, + 546, + 360 + ], + "type": "text", + "content": "[21] Hao Feng, Wengang Zhou, Jiajun Deng, Yuechen Wang, and Houqiang Li. Geometric representation learning for document image rectification. In ECCV, pages 475-492, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 360, + 546, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 546, + 392 + ], + "type": "text", + "content": "[22] James M Ferryman, Stephen J Maybank, and Anthony D Worrall. Visual surveillance for moving vehicles. *IJCV*, 37(2):187-197, 2000." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 392, + 546, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 392, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 392, + 546, + 426 + ], + "type": "text", + "content": "[23] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the KITTI vision benchmark suite. In CVPR, pages 3354-3361, 2012." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 426, + 546, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 546, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 546, + 470 + ], + "type": "text", + "content": "[24] Payam Goodarzi, Martin Stellmacher, Michael Paetzold, Ahmed Hussein, and Elmar Matthes. Optimization of a CNN-based object detector for fisheye cameras. In ICVES, pages 1-7, 2019." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 471, + 546, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 546, + 514 + ], + "type": "text", + "content": "[25] Sorin Grigorescu, Bogdan Trasnea, Tiberiu Cocias, and Gigel Macesanu. A survey of deep learning techniques for autonomous driving. Journal of Field Robotics, 37(3):362-386, 2020." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 514, + 546, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 546, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 546, + 548 + ], + "type": "text", + "content": "[26] Michael D Grossberg and Shree K Nayar. A general imaging model and a method for finding its parameters. In ICCV, volume 2, pages 108-115, 2001." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 548, + 546, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 546, + 581 + ], + "type": "text", + "content": "[27] Richard Hartley and Sing Bing Kang. Parameter-free radial distortion correction with center of distortion estimation. TPAMI, 29(8):1309-1321, 2007." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 581, + 546, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 581, + 546, + 614 + ], + "spans": [ + { + "bbox": [ + 307, + 581, + 546, + 614 + ], + "type": "text", + "content": "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 614, + 546, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 546, + 647 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 546, + 647 + ], + "type": "text", + "content": "[29] Jose Henrique Brito, Roland Angst, Kevin Koser, and Marc Pollefeys. Radial distortion self-calibration. In CVPR, pages 1368-1375, 2013." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 647, + 546, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 546, + 680 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 546, + 680 + ], + "type": "text", + "content": "[30] Juho Kannala and Sami S Brandt. A generic camera model and calibration method for conventional, wide-angle, and fish-eye lenses. TPAMI, 28(8):1335-1340, 2006." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "type": "text", + "content": "[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014." + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12426" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[32] Zuzana Kukelova and Tomas Pajdla. A minimal solution to radial distortion autocalibration. TPAMI, 33(12):2410-2422, 2011." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 172 + ], + "type": "text", + "content": "[33] Varun Ravi Kumar, Marvin Klingner, Senthil Yogamani, Stefan Milz, Tim Fingscheidt, and Patrick Mader. SyndistNet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In IEEE Winter Conference on Applications of Computer Vision, pages 61-71, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "type": "text", + "content": "[34] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In CVPR, pages 4681-4690, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 227, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 227, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 227, + 287, + 259 + ], + "type": "text", + "content": "[35] Xiaoyu Li, Bo Zhang, Jing Liao, and Pedro V Sander. Document rectification and illumination correction using a patch-based CNN. TOG, 38(6):1-11, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 259, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 259, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 259, + 287, + 293 + ], + "type": "text", + "content": "[36] Xiaoyu Li, Bo Zhang, Pedro V Sander, and Jing Liao. Blind geometric distortion correction on images through deep learning. In CVPR, pages 4855-4864, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 293, + 287, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 336 + ], + "type": "text", + "content": "[37] Kang Liao, Chunyu Lin, Lixin Liao, Yao Zhao, and Weiyao Lin. Multi-level curriculum for training a distortion-aware barrel distortion rectification model. In ICCV, pages 4389-4398, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 337, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 337, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 287, + 370 + ], + "type": "text", + "content": "[38] Kang Liao, Chunyu Lin, Yao Zhao, and Moncef Gabbouj. DR-GAN: Automatic radial distortion rectification using conditional gan in real-time. TCSVT, 30(3):725-733, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "type": "text", + "content": "[39] Kang Liao, Chunyu Lin, Yao Zhao, and Mai Xu. Model-free distortion rectification framework bridged by distortion distribution map. TIP, 29:3707-3718, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 403, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 403, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 403, + 287, + 437 + ], + "type": "text", + "content": "[40] Liang Lin, Yongyi Lu, Yan Pan, and Xiaowu Chen. Integrating graph partitioning and matching for trajectory analysis in video surveillance. TIP, 21(12):4844-4857, 2012." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "text", + "content": "[41] Khan Muhammad, Jamil Ahmad, Zhihan Lv, Paolo Bellavista, Po Yang, and Sung Wook Baik. Efficient deep CNN-based fire detection and localization in video surveillance applications. TSMSC, 49(7):1419-1434, 2018." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 481, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 514 + ], + "type": "text", + "content": "[42] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 514, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 548 + ], + "type": "text", + "content": "[43] Elad Plaut, Erez Ben Yaacov, and Bat El Shlomo. 3D object detection from a single fisheye image without a single fisheye training image. In CVPR, pages 3659-3667, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "type": "text", + "content": "[44] Luis Puig, Yalin Bastanlar, Peter Sturm, José Jesús Guerrero, and Joao Barreto. Calibration of central catadioptric cameras using a dlt-like approach. IJCV, 93(1):101-114, 2011." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 581, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 634 + ], + "type": "text", + "content": "[45] Helge Rhodin, Christian Richardt, Dan Casas, Eldar Insafutdinov, Mohammad Shafiei, Hans-Peter Seidel, Bernt Schiele, and Christian Theobalt. EgoCap: Egocentric marker-less motion capture with two fisheye cameras. TOG, 35(6):1-11, 2016." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 680 + ], + "type": "text", + "content": "[46] Jiangpeng Rong, Shiyao Huang, Zeyu Shang, and Xianghua Ying. Radial lens distortion correction using convolutional neural networks trained with synthesized images. In ACCV, pages 35-49, 2016." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[47] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234–241, 2015." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 710 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 127 + ], + "type": "text", + "content": "[48] Daniel Santana-Cedrés, Luis Gomez, Miguel Aleman-Flores, Agustín Salgado, Julio Esclarín, Luis Mazorra, and Luis Alvarez. An iterative optimization algorithm for lens distortion correction using two-parameter models. Image Processing On Line, 6:326-364, 2016." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 127, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 127, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 127, + 545, + 171 + ], + "type": "text", + "content": "[49] Davide Scaramuzza, Agostino Martinelli, and Roland Siegwart. A flexible technique for accurate omnidirectional camera calibration and structure from motion. In ICVS, pages 45-45, 2006." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 171, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 171, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 307, + 171, + 545, + 226 + ], + "type": "text", + "content": "[50] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial Intelligence and Machine Learning for Multi-domain Operations Applications, volume 11006, pages 369-386, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 226, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 226, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 307, + 226, + 545, + 248 + ], + "type": "text", + "content": "[51] Peter Sturm and Srikumar Ramalingam. A generic concept for camera calibration. In ECCV, pages 1-13, 2004." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 248, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 248, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 248, + 545, + 270 + ], + "type": "text", + "content": "[52] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 270, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 270, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 270, + 545, + 304 + ], + "type": "text", + "content": "[53] Thorsten Thormählen, Hellward Broszio, and Ingolf Wassermann. Robust line-based calibration of lens distortion from a single view. *Mirage*, pages 105–112, 2003." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 304, + 545, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 336 + ], + "type": "text", + "content": "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 336, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 336, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 336, + 545, + 380 + ], + "type": "text", + "content": "[55] Aiqi Wang, Tianshuang Qiu, and Longtan Shao. A simple method of radial distortion correction with centre of distortion estimation. Journal of Mathematical Imaging and Vision, 35(3):165-172, 2009." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 380, + 545, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 380, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 307, + 380, + 545, + 423 + ], + "type": "text", + "content": "[56] Guo-Wang Xie, Fei Yin, Xu-Yao Zhang, and Cheng-Lin Liu. Dewarping document image by displacement flow estimation with fully convolutional network. In DAS, pages 131-144, 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 423, + 545, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 423, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 307, + 423, + 545, + 457 + ], + "type": "text", + "content": "[57] Zhucun Xue, Nan Xue, Gui-Song Xia, and Weiming Shen. Learning to calibrate straight lines for fisheye image rectification. In CVPR, pages 1643-1651, 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 457, + 545, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 457, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 307, + 457, + 545, + 501 + ], + "type": "text", + "content": "[58] Shangrong Yang, Chunyu Lin, Kang Liao, Chunjie Zhang, and Yao Zhao. Progressively complementary network for fisheye image rectification using appearance flow. In CVPR, pages 6348-6357, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "type": "text", + "content": "[59] Mingde Yao, Dongliang He, Xin Li, Fu Li, and Zhiwei Xiong. Towards interactive self-supervised denoising. IEEE Transactions on Circuits and Systems for Video Technology, 2023." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 544, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 589 + ], + "type": "text", + "content": "[60] Xiaoqing Yin, Xinchao Wang, Jun Yu, Maojun Zhang, Pascal Fua, and Dacheng Tao. FisheyeRecNet: A multi-context collaborative deep network for fisheye image rectification. In ECCV, pages 469-484, 2018." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "type": "text", + "content": "[61] Mi Zhang, Jian Yao, Menghan Xia, Kai Li, Yi Zhang, and Yaping Liu. Line-based multi-label energy optimization for fisheye image rectification and calibration. In CVPR, pages 4137-4145, 2015." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 632, + 545, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 632, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 307, + 632, + 545, + 676 + ], + "type": "text", + "content": "[62] Yahui Zhang, Shaodi You, and Theo Gevers. Automatic calibration of the fisheye camera for egocentric 3D human pose estimation from a single image. In WACV, pages 1772-1781, 2021." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 676, + 545, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 676, + 545, + 710 + ], + "spans": [ + { + "bbox": [ + 307, + 676, + 545, + 710 + ], + "type": "text", + "content": "[63] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. TPAMI, 40(6):1452-1464, 2017." + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12427" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SimMatchV2_ Semi-Supervised Learning with Graph Consistency/4de24035-8115-4285-8e43-089c94dd354f_content_list.json b/2023/SimMatchV2_ Semi-Supervised Learning with Graph Consistency/4de24035-8115-4285-8e43-089c94dd354f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1c45f8a31afbd9e9e35a01260d70332f396d0d1d --- /dev/null +++ b/2023/SimMatchV2_ Semi-Supervised Learning with Graph Consistency/4de24035-8115-4285-8e43-089c94dd354f_content_list.json @@ -0,0 +1,1508 @@ +[ + { + "type": "text", + "text": "SimMatchV2: Semi-Supervised Learning with Graph Consistency", + "text_level": 1, + "bbox": [ + 150, + 130, + 818, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingkai Zheng $^{1}$ Shan You $^{2*}$ \nLang Huang $^{3}$ Chen Luo $^{5}$ Fei Wang $^{4}$ Chen Qian $^{2}$ Chang Xu $^{1}$ $^{1}$ School of Computer Science, Faculty of Engineering, The University of Sydney \n $^{2}$ SenseTime Research ${}^{3}$ The University of Tokyo \n $^{4}$ University of Science and Technology of China \n $^{5}$ State Grid Anhui Electric Power Research Institute", + "bbox": [ + 215, + 179, + 772, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semi-Supervised image classification is one of the most fundamental problem in computer vision, which significantly reduces the need for human labor. In this paper, we introduce a new semi-supervised learning algorithm - SimMatchV2, which formulates various consistency regularizations between labeled and unlabeled data from the graph perspective. In SimMatchV2, we regard the augmented view of a sample as a node, which consists of a label and its corresponding representation. Different nodes are connected with the edges, which are measured by the similarity of the node representations. Inspired by the message passing and node classification in graph theory, we propose four types of consistencies, namely 1) node-node consistency, 2) node-edge consistency, 3) edge-edge consistency, and 4) edge-node consistency. We also uncover that a simple feature normalization can reduce the gaps of the feature norm between different augmented views, significantly improving the performance of SimMatchV2. Our SimMatchV2 has been validated on multiple semi-supervised learning benchmarks. Notably, with ResNet-50 as our backbone and 300 epochs of training, SimMatchV2 achieves $71.9\\%$ and $76.2\\%$ Top-1 Accuracy with $1\\%$ and $10\\%$ labeled examples on ImageNet, which significantly outperforms the previous methods and achieves state-of-the-art performance.", + "bbox": [ + 75, + 353, + 473, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 744, + 209, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the past decades, deep learning has demonstrated its superior performance in various tasks [15, 36, 17, 59]. However, its outstanding performance is usually based on the fully-supervised setting, which requires a large amount of data and annotations. In a real-world scenario, high-quality data annotations are fairly hard to collect since it is generally expensive, time-consuming, and sometimes", + "bbox": [ + 75, + 771, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/453fecbdc9a8f91f96af2826d42b2b7def9719b7d570adfa980eabb118ec480b.jpg", + "image_caption": [ + "Figure 1: The Top-1 Accuracy of ResNet-50 on ImagetNet with $1\\%$ labeled data. Our proposed SimMatchV2 performs significantly better than previous methods with fewer training epochs. Most importantly, 100 epochs of SimMatchV2 training only takes 14.7 GPU days, but the performance is much higher than PAWS [1] with 300 epochs (68 GPU days)." + ], + "image_footnote": [], + "bbox": [ + 504, + 323, + 888, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "require expert knowledge ( e.g. medical images). Semi-supervised learning (SSL) has recently drawn much attention to the research community due to its efficiency in dealing with few labeled data and leveraging massive unlabeled data, significantly reducing the need for human labor.", + "bbox": [ + 496, + 612, + 893, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For the typical SSL algorithms, one standard paradigm is to train a semantic classifier on the labeled data, then use the classifier to make a prediction on the unlabeled data as the pseudo-label [33]. Besides, a class-level \"consistency regularization\" will be applied to enforce consistency between the class predictions of a model on different perturbations of the same input, which encourages the model to learn more robust and invariant features that generalize well to unseen data. Thus, the quality of pseudo-labels plays a crucial role in consistency regularization, and many works focus on generating stable and high-quality pseudo-labels. For example, MixMatch [3], ReMixMatch [2], and FixMatch [43] are three typical \"consistency regularization\" based methods that significantly improve the performance of the SSL", + "bbox": [ + 496, + 689, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Correspondence to: Shan You .", + "bbox": [ + 94, + 886, + 452, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16432", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tasks. However, these methods highly depend on the classifier trained on the labeled data. A limited number of labels will result in a poor classifier that generates inaccurate pseudo-labels and leads to bad performance.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the recent success of unsupervised visual representation learning [9, 24, 10, 11, 22, 7, 12, 27, 28, 23], the pre-training and fine-tuning paradigm has gained popularity for the semi-supervised learning. In general, the model will be first trained on a large amount of unlabeled data with a proxy task (e.g. instance discrimination), which also imposes the consistency regularization on the instance level (i.e. forcing the model to give similar representation for the different perturbations of the same instance). Then, a classifier will be attached to the pre-trained model, and the whole model will be fine-tuned with the limited labeled data. Many works [29, 35, 26, 67, 68, 13, 50, 47, 30, 54] have been proposed to design a better proxy task to improve the quality of the representation. Apparently, the biggest advantage of this approach is that it does not depend on the supervised trained classifier but still yields high-quality representations for unlabeled data. Nevertheless, the drawback of this approach is also very obvious, since no labels were used during the pre-training, making the model hard to converge. Generally, the pre-training requires more than 800 epochs to converge, which is extremely time-consuming.", + "bbox": [ + 75, + 152, + 470, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although many different types of consistency regularizations have been proposed, including class-level [3, 2, 43], instance-level [9, 24, 22], and those that consider both [66, 43], how to establish a complete and effective consistency regularizations remains unresolved. This paper introduces a new SSL algorithm - SimMatchV2, which provides a complete consistency regularization solution for various scenarios. The overall idea is inspired by the message passing and node classification in the graph theory. In SimMatchV2, we consider the augmented view of a sample as a node in the graph which consists of a node label and its corresponding representation. Different nodes are connected with the edges, which are measured by the similarity of the node representations. Thus, the information of different nodes can be efficiently propagated through the graph. To this end, we propose four types of consistency regularization, namely 1) node-node consistency, 2) node-edge consistency, 3) edge-edge consistency, and 4) edge-node consistency. Note that the 1) node-node consistency has been proposed in previous works, but our SimMatchV2 provides a complete and unified solution to design the consistencies. Moreover, we also propose a feature normalization technique to reduce the gaps between the feature norm between different augmented views and greatly improve the performance. Our contributions can be summarized as:", + "bbox": [ + 75, + 470, + 470, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose SimMatchV2 - a novel semi-supervised algorithm that models various consistency regularizations from the graph perspectives and proposes four", + "bbox": [ + 94, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "types of consistencies to provide a complete solution for the regularization in SSL.", + "bbox": [ + 529, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We show that a simple feature normalization reduces the gaps between the feature norm of different augmented views and greatly improves the performance.", + "- SimMatchV2 sets a new state-of-the-art performance on various SSL benchmarks. Notably, with ResNet50 as our backbone and 300 epochs of training, SimMatchV2 achieves $71.9\\%$ and $76.2\\%$ Top-1 accuracy with $1\\%$ and $10\\%$ labeled data on ImageNet. This result is $+4.7\\%$ and $+0.7\\%$ better than prior arts." + ], + "bbox": [ + 517, + 126, + 890, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 284, + 637, + 300 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Consistency Regularization with Pseudo-Label. Pseudo-label is a method that aims to generate labels for unlabeled data. As has been mentioned, it is often used in conjunction with consistency regularization, which constrains the model from making similar predictions on different perturbations of the same input. Most of these methods focus on generating high-quality pseudo-labels and exploring better perturbation strategies. For example, II-model [32] propose a temporal ensemble strategy for the pseudo-label to reduce the noise in the target. MeanTeacher [46] instead, maintain an exponential moving average (EMA) network to generate more reliable and stable pseudo-labels. Later, MixMatch [3] was proposed and greatly improved the SSL performance. Specifically, it adopts the averaged prediction of multiple strongly augmented views of the same image as the pseudo-label, then incorporates with the Mixup [63] augmentation that generates a convex combination of pairs of images and uses the corresponding mixed pseudo-label as the supervision target. ReMixMatch [2] improves MixMatch by using only a weakly augmented view to generate the pseudo-label and also introduces the distribution alignment to encourage the marginal distribution of predictions on unlabeled data to be consistent with the labeled data. FixMatch [43] abandons these complicated strategies and proposes a confidence-based threshold method to filter out the low-confidence pseudo-labels during the training.", + "bbox": [ + 496, + 310, + 890, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-Supervised Learning for SSL. Self-supervised learning has demonstrated its outstanding performance in learning visual representations without labels. It is intuitive and natural to leverage the proxy tasks from self-supervised learning to help with SSL tasks. As mentioned, the most straightforward solution is to pre-train the model in a self-supervised manner on a large amount of unlabeled data and then fine-tune the model with the limited labeled data. In this way, the design of the self-supervised task is particularly important. For example, instance discrimination [53] is one of the most popular proxy tasks, which aims to encourage different augmented views of the same image to", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16433", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7feb6baedfb99e70b404cc269aa65c21deb7b2be80231cd5f771684e470a29fc.jpg", + "image_caption": [ + "(a) Node-Edge Consistency" + ], + "image_footnote": [], + "bbox": [ + 161, + 89, + 336, + 224 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3a5c414e81235605c68b93b9bed1b65faf0beb7ac894e0401a41738429c8868f.jpg", + "image_caption": [ + "(b) Edge-Edge Consistency", + "Figure 2: Illustration of various consistency in SimMatchV2. The green and blue nodes are from the weakly and strongly augmented view of the unlabeled data, the red nodes are the other samples." + ], + "image_footnote": [], + "bbox": [ + 359, + 88, + 609, + 220 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/77973cd795f807da3d6f12147c88fb6ef683e0988013a0217b41324e101c206e.jpg", + "image_caption": [ + "(c) Edge-Node Consistency" + ], + "image_footnote": [], + "bbox": [ + 627, + 87, + 807, + 218 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "be pulled closer on embedding space but pushes apart all the other images away. BYOL [22] and SimSiam [12] omit the negatives in contrastive learning and leverage an asymmetric structure to avoid representation collapse. Recently, some works (e.g. CoMatch [34] and SimMatch [66]) have been proposed to jointly leverage both class-level and instance-level consistencies and greatly improve the SSL performance, especially in the case when only very limited labeled data is available. There are also many works [6, 5] exploring the two-stage framework (i.e. pre-training followed by fine-tuning with SSL algorithms like FixMatch), which shows the importance of the self-supervised pretrained to the SSL tasks.", + "bbox": [ + 75, + 276, + 473, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 510, + 210, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 542, + 279, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to [34, 66], we define a batch of labeled data as $\\mathcal{X} = \\{x_b : b \\in (1, \\dots, B)\\}$ , and a batch of unlabeled data as $\\mathcal{U} = \\{u_b : b \\in (1, \\dots, \\mu B)\\}$ , where $\\mu$ is the coefficient of unlabeled batch size. Then, a neural network encoder will extract the feature $h$ from the augmented images. We use $h_b^w$ and $h_b^s$ to denote the feature extracted from the weakly and strongly augmented images, respectively. Then, we use $W \\in \\mathbb{R}$ to represent a linear layer followed by a SoftMax function that transforms the feature to a class probability i.e. $p = \\mathrm{SoftMax}(W^T h)$ . Finally, we adopt the $g(\\cdot)$ to denote a two-layer non-linear projection head which maps the feature to a low-dimensional representation vector i.e. $z = g(h)$ . By following the common setting as in [9, 24, 66, 34], the representation $z$ will be normalized to a unit vector by default. Note that the reason we use $z$ as the representation instead of $h$ is that we need to store a large number of samples as in [24]. The dimension of $z$ is much smaller than $h$ which makes the training more efficient. For example, $h$ is 2048-D and $z$ is 128-D for ResNet-50.", + "bbox": [ + 75, + 574, + 468, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For labeled data, we perform standard supervised learning with the cross-entropy loss, which can be represented", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "by Eq. (1), where $y_{b}$ is the one-hot ground truth label,", + "bbox": [ + 498, + 276, + 856, + 292 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = - \\frac {1}{B} \\sum_ {b = 1} ^ {B} y _ {b} \\log p _ {b} ^ {w}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 296, + 890, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Consistency Regularization with Graph", + "text_level": 1, + "bbox": [ + 498, + 343, + 841, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SimMatchV2 considers the augmented view of a sample as a node in the graph, which consists of a node label $p$ (class prediction for unlabeled data and ground truth for labeled data) and the representation $z$ . Different nodes are connected by edges measured by the similarity of the representations. We use strong/weak node to name the node from strongly/weakly augmented views.", + "bbox": [ + 496, + 366, + 890, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a graph with $K$ nodes, we have the node labels $\\{p_k : k \\in (1, \\dots, K)\\}$ and the corresponding representations $\\{z_k : k \\in (1, \\dots, K)\\}$ . In our implementation, the graph nodes will be stored in a memory bank as in [24, 66, 34]. We can store both labeled and unlabeled nodes. The node label will be the ground truth for labeled data and $p$ for unlabeled data. Suppose we have an unlabeled weak node with the representation $z_b^w$ . We use $e_{b,i}^w$ to denote the edge between the $z_b^w$ and the $i^{th}$ node on the graph, which can be expressed by Eq. (2). $t$ is a temperature parameter that controls the sharpness of the similarity distribution. We also use $e_{b,i}^w$ to denote edges between $z_b^w$ and all the nodes on the graph (i.e. $e_{b,i}^w$ is a scalar and $e_b^w \\in R^{1 \\times K}$ ).", + "bbox": [ + 498, + 472, + 892, + 674 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ne _ {b, i} ^ {w} = \\frac {\\exp \\left(z _ {b} ^ {w} \\cdot z _ {i} / t\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(z _ {b} ^ {w} \\cdot z _ {k} / t\\right)} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 679, + 890, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Node-Node Consistency. (Weak Node vs. Strong Node.) The node-node consistency aims to maintain the consistency of the label between the strong node and weak node of the same instance. Such consistency has been proposed in many previous works [3, 2, 43]. In this work, we have renamed it and integrated it into our graph-based framework. Concretely, we follow the design in previous works; given unlabeled data, the node-node consistency can be expressed by Eq. (3), where $\\tau$ is the confidence threshold.", + "bbox": [ + 496, + 719, + 893, + 856 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {n n} = - \\frac {1}{\\mu B} \\sum_ {b = 1} ^ {\\mu B} \\mathbb {1} \\left(\\max p _ {b} ^ {w} > \\tau\\right) p _ {b} ^ {w} \\log p _ {b} ^ {s}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 861, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16434", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Node-Edge Consistency. (Weak Node vs. Strong Edge.) The purpose of the node-edge consistency is to construct an edge $e_b^s$ by using the representation $z_b^s$ from the strong node, then use the edge $e_b^s$ as the weight to aggregate the node label from the graph. We expect the aggregated node label can be consistent with the label $p_b^w$ from the weak node. Finally, the node-edge consistency will be constrained by the cross-entropy loss between $p_b^w$ and the aggregated label, which can be expressed by Eq. (4).", + "bbox": [ + 76, + 90, + 472, + 229 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {n e} = - \\frac {1}{\\mu B} \\sum_ {b = 1} ^ {\\mu B} p _ {b} ^ {w} \\log \\left(\\sum_ {i = 1} ^ {K} p _ {i} e _ {b, i} ^ {s}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 248, + 468, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that the aggregated label $\\sum_{i=1}^{K} p_i e_{b,i}^s$ does not include the node label $p_b^s$ and $p_b^w$ from itself. The gradient will only be applied on $e_{b,i}^s$ .", + "bbox": [ + 76, + 311, + 468, + 361 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Edge-Edge Consistency. (Weak Edge vs. Strong Edge.) The edge-edge consistency can be regarded as a simple extension of the node-node consistency. Basically, the node-node consistency aims to let the model predict invariant labels for the weak and strong nodes of the same instances. Similarly, the edges should also be invariant with respect to the strong and weak nodes of the same instance. Thus, the edge-edge consistency can be formulated as the cross-entropy loss between the similarity distribution from different augmented views, which can be written as Eq. (5).", + "bbox": [ + 75, + 363, + 472, + 515 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {e e} = - \\frac {1}{\\mu B} \\sum_ {b = 1} ^ {\\mu B} \\sum_ {i = 1} ^ {K} e _ {b, i} ^ {w} \\log e _ {b, i} ^ {s} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 535, + 468, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Edge-Node Consistency. (Weak Edge vs. Strong Node.) This is somewhat similar to reversed operation to the node-edge consistency. Concretely, node-edge consistency requires the aggregated label $\\sum_{i=1}^{K} p_i e_{b,i}^s$ from the strong node to be consistent with the label $p_b^w$ from the weak node. In contrast, edge-node consistency requires the label $p_b^s$ from the strong node to be consistent with the aggregated label $\\sum_{i=1}^{K} p_i e_{b,i}^w$ from the weak node. Given an unlabeled weak node which has $z_b^w$ and $p_b^w$ . Then, we concatenate the unlabeled node with the graph to get $Z = [z_b^w, z_1, \\dots, z_k]$ , $Y = [p_b^w, p_1, \\dots, p_k]$ . Next, we use the representation matrix $Z$ to obtain the pairwise edge matrix $A$ (a.k.a. affinity matrix) in the same way as Eq. (2). Note that the difference between $e_b$ and $A$ is that $e_b$ only calculates the edges between the unlabeled data with respect to the node in the graph ( $e_b^w \\in R^{1 \\times K}$ ) but $A$ also contains the edge information within the graphs. ( $A \\in R^{(1 + K) \\times (1 + K)}$ ). We also want to emphasize that we will make $A$ a hollow matrix (i.e. $A_{i,j} = 0$ if $i == j$ ) before normalizing it to avoid the self-loop in the graph. Here, we adopt the transductive", + "bbox": [ + 76, + 595, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "inference [69] for aggregation:", + "bbox": [ + 500, + 90, + 705, + 107 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} Y _ {\\phi} = \\alpha \\cdot A Y _ {\\phi - 1} + (1 - \\alpha) \\cdot Y _ {0} \\\\ = (\\alpha \\cdot A) ^ {\\phi - 1} Y _ {0} + (1 - \\alpha) \\sum_ {i = 0} ^ {\\phi - 1} (\\alpha \\cdot A) ^ {i} Y _ {0}, \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 112, + 890, + 174 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi$ is the number of iterations that the label has been propagated, $Y_{0}$ is the initial state $[p_b^w,p_1,\\dots,p_k]$ , and $\\alpha$ is a parameter in (0,1) controlling the weight of the aggregated label and the initial state $Y_{0}$ . The node label will be propagated multiple times until it fully converges. When $\\phi \\rightarrow \\infty$ , we have $(\\alpha \\cdot A)^{\\phi -1} = 0$ since $0 < \\alpha < 1$ . We can also make an approximation such that $\\lim_{\\phi \\to \\infty}\\sum_{i = 0}^{\\phi -1}(\\alpha \\cdot A)^{i} = (I - \\alpha \\cdot A)^{-1}$ , where $I$ is the identity matrix. Thus, the converged $Y_{\\infty}$ can be derived as:", + "bbox": [ + 498, + 180, + 893, + 318 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY _ {\\infty} = (1 - \\alpha) (I - \\alpha \\cdot A) ^ {- 1} Y _ {0} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 324, + 890, + 342 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the converged pseudo-label of the unlabeled data can be obtained from $Y_{\\infty}[0]$ (i.e. the first row vector of $Y_{\\infty}$ ). Note that if we take $\\alpha = 1$ , then $Y_{1}[0]$ is identical to $\\sum_{i=1}^{K} p_{i} e_{b,i}^{w}$ . Since $Y_{\\infty}[0]$ contains both information from the node label and aggregated label. Thus, we can directly use it to replace the original $p_{b}^{w}$ in Eq. (3). Moreover, we should notice an inverse operator in Eq. (7), which is computationally very expensive for a large matrix. Thus, to reduce the complexity, we will only take the Top-N nearest node from $Z$ and $Y$ instead of using the whole graph.", + "bbox": [ + 498, + 347, + 893, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature Normalization. As we have presented in Eq. (3), one of the training objectives is to minimize the predictions between the weakly and strongly augmented views. However, significant gaps exist between the feature norms of different augmented views. Also, from $p = \\operatorname{SoftMax}(W^T h)$ , we can observe that the norm of $h$ does not affect the prediction (i.e. $\\operatorname{argmax}(p)$ ). Thus, minimizing the gaps between the feature norm is unnecessary and brings additional complexity to the training. To address this issue, we simply apply a layer normalization [56] on the feature $h$ , which can be expressed by Eq. (8). Finally, the linear layer $W$ and non-linear projection head $g(\\cdot)$ will be applied on top of the $\\widehat{h}$ ,", + "bbox": [ + 498, + 501, + 893, + 699 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {h} = \\operatorname {L a y e r N o r m} (h). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 705, + 890, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the overall objective of SimMatchV2 can be expressed by Eq. (9), where $\\lambda_{nn},\\lambda_{ee}$ , and $\\lambda_{ne}$ are the balancing factors that control the weights of the different losses.", + "bbox": [ + 498, + 729, + 893, + 777 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {s} + \\lambda_ {n n} \\mathcal {L} _ {n n} + \\lambda_ {n e} \\mathcal {L} _ {n e} + \\lambda_ {e e} \\mathcal {L} _ {e e} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 790, + 890, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 821, + 633, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Classical SSL Benchmark", + "text_level": 1, + "bbox": [ + 500, + 847, + 735, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We will first conduct our experiments on the classic SSL benchmark which is widely adopted by most SSL papers.", + "bbox": [ + 498, + 869, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16435", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/400dc6641308bea5a7063c8a68940ebaa1a44593609fd44d9dc7ea30939413fd.jpg", + "table_caption": [ + "Table 1: Error rate (%) and Rank on the classical SSL benchmark. The error rate and standard deviation are reported based on three runs. The experimental settings in this table are all based on the USB [51] code base. All results of other methods are directly copied from the USB [51] benchmark paper. The results for fully-supervised training on STL-10 is not provided since it contains an unlabeled set with unknown labels." + ], + "table_footnote": [], + "table_body": "
DatasetCIFAR-10CIFAR-100SVHNSTL-10Friedman rankFinal rankMean error rate
# Label402504000400250010000402501000402501000---
Fully-Supervised Supervised4.57±0.0618.96±0.062.14±0.01----
77.18±1.3256.24±3.4116.10±0.3289.60±0.4358.33±1.4136.83±0.2182.68±1.9124.17±1.6512.19±0.2375.4±0.6655.07±1.8335.42±0.48---
II-Model [32]76.35±1.6948.73±1.0713.63±0.0787.67±0.7956.40±0.6936.73±0.0580.07±1.2213.46±0.616.90±0.2274.89±0.5752.20±2.1131.34±0.6414.171548.20
Pseudo-Labelling [33]75.95±1.8651.12±2.9115.32±0.3588.18±0.8955.37±0.4836.58±0.1275.98±5.3616.47±0.599.37±0.4274.02±0.4751.90±1.8730.77±0.0414.081448.42
Mean Teacher [46]72.42±2.1037.56±4.908.29±0.1079.96±0.5344.37±0.6031.39±0.1149.34±7.903.44±0.023.28±0.0772.90±0.8349.30±2.0927.92±1.6511.751240.01
VAT [41]78.58±2.7828.87±3.6210.90±0.1683.60±4.2146.20±0.8032.14±0.3184.12±3.183.38±0.122.87±0.1873.33±0.4757.78±1.4740.98±0.9612.921345.23
MixMatch [3]35.18±3.8713.00±0.806.55±0.0564.91±3.3439.29±0.1327.74±0.2727.77±5.434.69±0.463.85±0.2849.84±0.5832.05±1.1620.17±0.6711.001127.09
ReMixMatch [2]8.13±0.586.34±0.224.65±0.0941.60±1.4825.72±0.0720.04±0.1316.43±13.775.65±0.355.36±0.5827.87±3.8511.14±0.526.44±0.157.751014.95
UDA [5]10.01±3.345.23±0.084.36±0.0945.48±0.3727.51±0.2823.12±0.455.28±0.421.93±0.031.94±0.0240.09±4.0310.11±1.156.23±0.286.42815.12
FixMatch [3]12.66±4.494.95±0.104.26±0.0145.38±2.0727.71±0.4222.06±0.103.37±1.011.97±0.012.02±0.0338.19±4.768.64±0.845.82±0.065.29514.75
Dash [57]9.29±3.285.10±0.284.36±0.1047.49±1.0527.47±0.3821.89±0.165.26±2.022.01±0.012.08±0.0942.00±4.9410.50±1.376.30±0.496.83915.32
CoMatch [34]6.51±1.185.35±0.144.27±0.1253.41±2.3629.78±0.1122.11±0.228.20±3.522.16±0.042.01±0.0413.74±4.207.63±0.945.71±0.085.75613.41
CRMatch [18]13.62±2.624.61±0.173.65±0.0437.76±1.4524.13±0.1619.89±0.232.60±0.771.98±0.041.95±0.0333.73±1.1714.87±5.096.53±0.364.08213.78
FlexMatch [4]5.29±0.294.97±0.074.24±0.0640.73±1.4426.17±0.1821.75±0.155.42±2.838.74±3.327.90±0.3029.12±5.049.85±1.356.08±0.345.92714.19
AdaMatch [4]5.09±0.215.13±0.054.36±0.0537.08±1.3526.66±0.3321.99±0.156.14±4.352.13±0.042.02±0.0519.95±5.178.59±0.436.01±0.024.79312.18
SimMatch [66]5.38±0.015.36±0.084.41±0.0739.32±0.7226.21±0.3721.50±0.117.60±2.112.48±0.612.05±0.0516.98±4.248.27±0.405.74±0.315.25412.10
SimMatchV2 (Ours)4.90±0.165.04±0.094.33±0.1636.68±0.8626.66±0.3821.37±0.207.92±2.802.92±0.812.85±0.9115.85±2.627.54±0.815.65±0.264.00111.81
", + "bbox": [ + 80, + 128, + 890, + 285 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "There are four datasets will be used in this experiment.", + "bbox": [ + 76, + 297, + 437, + 311 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- CIFAR-10 [31] consists of 60K 32x32 color images in 10 classes, with 6k images per class. The training and test set contains 5K and 1k samples, respectively. We conduct three different experiments on 40, 250, and 4,000 randomly selected images from the training set as the labeled data and use the rest of the training set as the unlabeled data.", + "- CIFAR-100 [31] is a more complex version of CIFAR-10, which contains 100 classes with 600 images each. There are 500 training images and 100 testing images per class. The experimental setting are similar to CIFAR-10 above except using 400, 2,500, and 10K labeled data for training.", + "- SVHN [42] is a real-world image dataset that contains 32x32 cropped digits from house numbers in Google Street View images. It contains 73,257 training images, with 26,032 testing images and 531,131 extra images. We merge the training and extra set and randomly select 40, 250, and 1,000 from it as the label data to conduct the experiment.", + "- STL-10 [14] consists of 5K labeled images, 100K unlabeled images, and 8,000 test images. We will randomly select 40, 250, and 1,000 from the 5K images; the rest images will be combined with 100K as the unlabeled data." + ], + "bbox": [ + 75, + 315, + 467, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Our implementation is mainly based on the USB [51] code base for a fair comparison. Most of the hyper-parameters are inherited from [43]. Specifically, we adopt WRN-28-2 [60] for CIFAR-10 and SVHN, WRN-28-8 for CIFAR-100, and WRN-37-2 for STL-10. We use an SGD optimizer with Nesterov momentum [45] for all experiments. The weight decay will be set to $5e^{-4}$ for most experiments except CIFAR-100, for which we set it as $1e^{-3}$ . For the learning rate schedule, we use a cosine learning rate decay [39], which adjusts the learning rate to $0.03 \\cdot \\cos\\left(\\frac{7\\pi s}{16S}\\right)$ where $s$ is the current training step, and $S$ is the total number of training steps. The model will be optimized for $2^{20}$ steps. We use 64 labeled data and 448 unlabeled data for each batch. All experiments will be conducted on a single GPU. The weak and strong augmentation policies are also directly inherited from the USB codebase. Note that the final performance will be reported based on", + "bbox": [ + 75, + 643, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the exponential moving average of the model with a 0.999 decay rate. For SimMatchV2 specific hyper-parameters, we set $\\lambda_{nn} = 1, \\lambda_{ne} = 1, \\lambda_{ee} = 1, t = 0.1, \\alpha = 0.1, \\tau = 0.95, n = 8$ . ( $n$ is the Top-N nearest node for Eq. (7)) These hyper-parameters are kept unchanged for all experiments. We maintain two memory banks for storing both unlabeled nodes and labeled nodes. The unlabeled bank will be used for calculating the Eq. (4) and Eq. (5). The labeled bank will be used for computing the Eq. (7).", + "bbox": [ + 496, + 297, + 890, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Performance on Classical SSL Benchmark. The overall results are reported in Table 1. We compare our SimMatchV2 against 14 SSL algorithms. As in Table 1, it is very hard for an algorithm to get the best performance on all settings. Many algorithms already meet or exceed the supervised baseline in some settings (CIFAR-10 and SVHN); the performance difference would not be significant in this case. For a more rigorous and fair comparison, we follow the steps from USB [51], which adopts the Friedman rank [19, 20] as the final evaluation metric. We also report the mean error rate for all settings. As can be seen, our SimMatchV2 achieves the best performance on 4 out of 12 settings and also has the highest Friedman rank and lowest mean error rate. CRMatch [18] also achieves a very good performance that ranked second in the final rank. It has the best performance on 5 out of the 12 settings, but it is less stable when the number of labeled data is limited, and the mean error rate is much worse than SimMatchV2. AdaMatch [4] and SimMatch [66] has a more stable performance under various conditions and they have a very close mean error rate compared to SimMatchV2. However, these two methods did not achieve any best results in this experiment, leading to a worse Friedman rank. Overall, the highest Friedman rank and lowest mean error rate for SimMatchV2 should demonstrate superiority in terms of both performance and stability.", + "bbox": [ + 496, + 436, + 890, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Experiments on ImageNet", + "text_level": 1, + "bbox": [ + 500, + 844, + 735, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we evaluate our SimMatchV2 on the large-scaled ImageNet dataset [15], which contains 1,000 categories", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16436", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6935e140a10ef8af42e0475de3800f9882d35a22d9c9334fb3f63b40e0e1d831.jpg", + "table_caption": [ + "Table 2: Top-1 an Top-5 Accuracy with ResNet-50 on ImageNet with $1\\%$ and $10\\%$ settings. Note that the underline in the parameters - inference columns denotes the methods which require more parameters than standard ResNet-50 during the inference stage." + ], + "table_footnote": [], + "table_body": "
Pre-training AlgorithmSemi-supervised AlgorithmPre-training EpochsSemi-supervised EpochsParamters trainParamters inferenceTop-1 Label fractionTop-5 Label fraction
1%10%1%10%
NoneVAT+EntMin. [41, 21, 61]-~10025.6M25.6M--47.083.4
S4L-Rotation [61]-~20025.6M25.6M--53.483.8
UDA [55]-~53025.6M25.6M-68.8-88.8
FixMatch [43]-~30025.6M25.6M-71.5-89.1
FixMatch-EMAN [6]-~30025.6M25.6M-72.8-90.3
SsCL [64]-~80028.0M25.6M60.272.182.890.9
CoMatch [34]-~40028.0M25.6M66.073.686.491.6
SimMatch [66]-~40028.0M25.6M67.274.487.191.7
PCL [35]Fine-tune~200~2023.8M25.6M--75.385.6
SimCLR [9]~1000~6028.0M25.6M48.365.675.587.8
SimCLR V2 [10]~800~6032.2M29.8M57.968.482.589.2
BYOL [22]~1000~5035.1M25.6M53.268.878.489.0
SwAV [7]~800~2028.0M25.6M53.970.278.589.9
WCL [65]~800~6032.2M29.8M65.072.086.391.2
MoCo V2 [11]Fine-tune~800~2028.0M25.6M49.866.177.287.9
FixMatch-EMAN [6]~800~30028.0M25.6M61.473.982.191.0
PAWS [7]Fine-tune~300~5036.1M29.8M66.575.5--
NoneSimMatchV2 (Ours)-~10028.0M25.6M69.974.888.791.7
-~30028.0M25.6M71.976.290.092.2
", + "bbox": [ + 84, + 116, + 883, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with 1.2M training images and 50K validation images. Concretely, we follow previous works [34, 66, 1, 6] to take $1\\%$ and $10\\%$ samples from the training images as our labeled data and use the rest samples as the unlabeled data. To avoid the effects of different selections on the labeled data, we follow [9, 10, 6, 1, 65] to adopt a pre-defined labeled / unlabeled splits1 that has been adopted in most SSL papers.", + "bbox": [ + 75, + 357, + 468, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. For ImageNet experiments, most of our hyper-parameter and settings are directly inherited from [66]. Specifically, we use ResNet-50 [25] as our default backbone and use the SGD with Nesterov momentum as the optimizer. We set the weight decay to 0.0001 during the training. For the learning rate, we warm up the model for 5 epochs until it reaches 0.03, and it will be decayed to 0 with the cosine scheduler [39]. By following [6, 66], we use 64 labeled data and 320 unlabeled data in each batch; the data will be evenly distributed on 8 GPUs. For the hyper-parameter settings, we set $\\lambda_{nn} = 10$ , $\\lambda_{ne} = 10$ , $\\lambda_{ee} = 10$ , $t = 0.1$ , $\\alpha = 0.1$ , $\\tau = 0.7$ , $n = 128$ for $1\\%$ setting, and we use $\\lambda_{nn} = 5$ , $\\lambda_{ne} = 5$ for $10\\%$ since we found the performance is slightly better. For strong augmentation, we adopt a multi-crop strategy as in [68], which contains five crops with different sizes (224x224, 192x192, 160x160, 128x128, and 96x96); these crops will be augmented with the policy in [8]. For weak augmentation, we follow the same strategy in [66, 34].", + "bbox": [ + 75, + 464, + 470, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance on ImageNet. The performance has been shown in Table 2. With 100 epochs of training, SimMatchV2 achieves $70.4\\%$ and $74.8\\%$ Top-1 accuracy with $1\\%$ and $10\\%$ labeled data, which already surpass the performance of CoMatch [34], and SimMatch [66] with 400 epochs. With 300 epochs of training, SimMatchV2 achieves $72.1\\%$ and $76.2\\%$ Top-1 accuracy, which improves absolute", + "bbox": [ + 75, + 753, + 468, + 859 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "+5.2% and +0.7% over PAWS [1]. As we can observe that the improvement of 10% setting is much lesser than 1% setting. We doubt that this is because the performance of 76.2% is already very close to the fully supervised setting (76.5%), which should be the upper bound of the SSL tasks. Most importantly, our SimMatchV2 established a new state-of-the-art performance without any pre-training.", + "bbox": [ + 500, + 357, + 893, + 464 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/01b1c8f1098f182b936785798403fa904e25e997c9cd79eec79594aec04a512b.jpg", + "table_caption": [ + "Table 3: Runtime Comparison. MC denotes whether we use the multicrop strategy. The experiments are performed on $1\\%$ ImageNet with 100 epochs of training. $\\downarrow$ indicates lower is better and $\\uparrow$ means higher is better." + ], + "table_footnote": [], + "table_body": "
MethodMCGPU Memory ↓GPU days ↓1% Top-1 ↑
CoMatch [34]✗81G10.361.1
SimMatch [66]✗40G7.761.2
SimMatchV2✗42G7.863.7
PAWS [1]✓1048G22.663.8
SimMatchV2✓83G14.769.9
", + "bbox": [ + 501, + 515, + 893, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Efficiency Comparison. Next, We also compare the training efficiency for different methods. The experiments are conducted with 8 Nvidia GPUs except for PAWS [1], which requires at least 64 GPUs to run. We use mixed precision training for all methods. Note that CoMatch [34] and SimMatch [66] have their own labeled / unlabeled splits, which makes the comparison a bit unfair. In this experiment, we adopt their officially released codebase to reproduce the performance on the same splits.", + "bbox": [ + 496, + 597, + 890, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show the training statistics in Table 3. When we omit the multi-crops, the computational cost of SimMatchV2 is slightly higher compared to SimMatch [66] (i.e. +2G memory, +0.1 GPU days), and it is much more efficient than CoMatch [34]. We can also observe that SimMatchV2 has much better performance (+2.5%) than CoMatch [34] and SimMatch [66] even without the multi-crops. Moreover, when the multi-crops are included, the training time of SimMatchV2 is nearly doubled (7.8 days → 14.7 days), but it is still only 65% of PAWS [1] (14.7 days / 22.6 days). The extra training cost of PAWS [1] is from the support", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1https://www.tensorflow.org/datasets/catalog/Imagenet2012 Subset", + "bbox": [ + 76, + 875, + 450, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16437", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "set containing $960 \\times 7$ labeled data per batch. Most importantly, our SimMatchV2 achieves similar performance with PAWS (63.7% vs. 63.8%) [1] with only 35% training time (7.8 days / 22.6 days), and has an absolute improvement of +6.6% with 65% training time.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8ad309088d1147d06506d70866fd814ad718d5c6f590cbd26413d114599fdbeb.jpg", + "table_caption": [ + "Table 4: More experiments on different architectures. Note that we do not use any pre-training in this experiment." + ], + "table_footnote": [], + "table_body": "
MethodArchParamsEpochs1%10%
Semifommer [52]ViT-S + Conv40M600-75.5
ProbPseudo Mixup [5]ViT-Small22M100-70.9
SimMatchV2ViT-Small22M10063.775.9
ProbPseudo Mixup [5]ViT-Base86M100-73.5
SimMatchV2ViT-Base86M10065.376.8
ProbPseudo Mixup [5]ConvNext-T28M100-74.1
SimMatchV2ConvNext-T28M10071.377.8
ProbPseudo Mixup [5]ConvNext-S50M100-75.1
SimMatchV2ConvNext-S50M10072.378.4
", + "bbox": [ + 83, + 198, + 460, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Working with More Advanced Architectures. With the recent successes in model architecture design, many more powerful models have been proposed, especially the vision transformer and its variant [16, 48, 49, 38, 37]. Thus, we further apply our SimMatchV2 on ViT-Small [48], ViT-Base [16], ConvNext-T [38], and ConvNext-S [38] to show the generality on different architectures. In this experiment, we follow [5] to use 128 labeled data and 640 unlabeled per batch. We also adopt the AdamW [40] optimizer with the learning rate of $1.25e^{-4}$ and weight decay of 0.1. We present the results in Table 4. As we can see, SimMatchV2 with a better model can generally produce a better result. The exploration of cooperating with more advanced SSL pipeline [5] and stronger data augmentations (e.g. ProbPseudo Mixup [5]) will be left as a future work since it requires more hyper-parameter turnings and more sophisticated designs, which is not the focus of this paper.", + "bbox": [ + 75, + 321, + 468, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Experiments on Semi-iNat 2021", + "text_level": 1, + "bbox": [ + 76, + 585, + 354, + 601 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dea4dce821a89b7b0f3bb2d10cc225c6094d33394fcea580e0cde184d4adc6ee.jpg", + "table_caption": [ + "Table 5: Accuracy for real-world dataset - Semi-iNat 2021." + ], + "table_footnote": [], + "table_body": "
MethodSemi-iNat 2021
With Random InitWith MoCo Init
Top1Top5Top1Top5
Supervised19.0935.8534.9657.11
MixMatch [3]16.8930.83--
+CCSSL [58]19.6535.09--
FixMatch [43]21.4137.6540.3060.05
+CCSSL [58]31.2152.2541.2864.30
CoMatch [34]20.9438.9638.9461.85
+CCSSL [58]24.1243.2339.8563.68
SimMatchV2 (Ours)43.0563.8153.1373.40
", + "bbox": [ + 94, + 618, + 450, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we also test the performance of SimMatchV2 on Semi-iNat 2021 [44], which contains 810 different categories with 9,721 labeled data, 313,248 unlabeled data, and 4,050 validation data. Unlike the dataset we used in previous experiments, Semi-iNat 2021 is designed to expose some challenges encountered in a real-world scenario, which contains highly similar classes, class imbalance, and domain mismatch between the labeled and unlabeled data. Such a more realistic dataset should be a better benchmark to reflect the performance of the different methods.", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this experiment, most of the hyper-parameters and implementation details are the same as our ImageNet setting, except that we do not use the multi-crops strategy for fair comparison. The model will be optimized for $512 \\times 1024$ steps as in [58]. We also follow the setting in [58] to conduct two experiments to report the training from scratch performance and training from pre-trained model performance.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show the results in Table 5. Our SimMatchV2 achieves $43.05\\%$ Top-1 accuracy when training from scratch, which is $+11.84\\%$ better than the state-of-the-art method (FixMatch + CCSL). When the MoCo pre-trained model is adopted, our SimMatchV2 achieves a $53.13\\%$ Top-1 accuracy, which is $+10.08\\%$ better than the from-scratch performance, and $+11.85\\%$ better than FixMatch + CCSL. We believe this experiment will further demonstrate the superiority of SimMatchV2.", + "bbox": [ + 496, + 198, + 890, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 349, + 648, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we will empirically study our SimMatchV2 based on various conditions and show the effect of each component and hyper-parameter sensitivities of our methods. For all experiments in this section, we adopt the ResNet-50 as our backbone and train the model on ImageNet for 100 epochs on the $1\\%$ labeled data. The multicrops are not used by default unless we mentioned.", + "bbox": [ + 496, + 376, + 890, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training Dynamics. We first show the training dynamics of SimMatchV2 compared to SimMatch [66] and CoMatch [34]. We visualize the pseudo-labels accuracy, unlabeled data accuracy, and validation accuracy in Figure 3. As we can observe that the pseudo-labels from SimMatchV2 have a higher quality than SimMatch [66], especially in the late stage of the training. Both of these two methods generate better pseudo-labels than CoMatch [34]. We can also notice that SimMatchV2 consistently has better accuracy for the unlabeled data and validation data, which shows the advantages of our method.", + "bbox": [ + 496, + 483, + 890, + 648 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ddd3bc36aa2c1206639138b60a678d8ec0cc0211d44e35bb5937e8c6beb0d2b2.jpg", + "table_caption": [ + "Table 6: Effect of each component in SimMatchV2. N and E denote the node and edge, respectively. N-E means the node-edge consistency and vice versa. Feat norm means feature normalization." + ], + "table_footnote": [], + "table_body": "
N-NN-EE-EE-NFeat NormTop-1Improvement
✓52.9-
✓✓60.0+7.1
✓✓✓60.7+0.7
✓✓✓✓62.2+1.5
✓✓✓✓✓63.7+1.5
", + "bbox": [ + 503, + 696, + 887, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Each Components. In table 6, we show the effect of each design in SimMatchV2. Basically, if only node-node consistency is adopted, the setting is equivalent to the FixMatch [43] with distribution alignment [2], which is our baseline in the first row of the table; Next, we can see that the node-edge, edge-node, and edge-node consistencies give $+7.1\\%$ , $+0.7\\%$ , $+1.5\\%$ improvements, respectively. Finally, the feature normalization also boosts the performance", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16438", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9f24b8eb5de621e2fa8f4ee27bfaa75b8a1771ffa4e06f933ac23b96ea5fa336.jpg", + "image_caption": [ + "(a) Pseudo-Labels Accuracy" + ], + "image_footnote": [], + "bbox": [ + 138, + 89, + 367, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce6cc88b43487746d9e31ad5e91ec04deff66e3a651cb57ea5fd5aae7c4234ff.jpg", + "image_caption": [ + "(b) Unlabeled Data Accuracy", + "Figure 3: In this Figure, we compare the training dynamics of SimMatchV2 with SimMatch [66], and CoMatch [34] from 3 different perspectives. a) The accuracy of $p^w$ that pass the threshold $\\tau$ . b) The accuracy of $p^w$ regardless of the threshold. c) The accuracy on the validation data. All experiments are in 1% ImageNet setting with 100 epochs training." + ], + "image_footnote": [], + "bbox": [ + 370, + 90, + 596, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/acc214e8c318ffaec44db1c4de232ab6d7ca04e0016e4fffdbda44759864bb44.jpg", + "image_caption": [ + "(c) Validation Accuracy" + ], + "image_footnote": [], + "bbox": [ + 601, + 90, + 826, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "+1.5%. This experiment should demonstrate that these different designs are all very effective for SSL tasks.", + "bbox": [ + 76, + 294, + 470, + 325 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a94b22f782aede1e9014829522cd8dd598b44a8bdc3c646961695406ed2cc90d.jpg", + "table_caption": [ + "Table 7: Effect of memory bank size on the performance. $5\\mathrm{k}-163\\mathrm{K}$ are setting that we use unlabeled bank; labeled bank means we use all labeled bank (All labeled data 12K for $1\\%$ ImageNet)." + ], + "table_footnote": [], + "table_body": "
Bank Size5K10K40K82K163KLabeled Bank (12K)
Top-163.163.263.563.763.663.2
", + "bbox": [ + 78, + 371, + 467, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Node-Edge and Edge-Edge Consistency. We will study the node-edge and edge-edge consistency together since the memory bank size is their primary factor. The experimental results are shown in Table 7. We can observe that when unlabeled data is used, the performance will be slightly increased with a larger memory bank, and the best memory bank size is $82\\mathrm{K}$ . We can also notice that we get slightly lower performance with $12\\mathrm{K}$ labeled data. Therefore, we decide to use the unlabeled data for node-edge and edge-edge consistency.", + "bbox": [ + 75, + 414, + 468, + 566 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8cde9d25925282561de83ad0976462716614c8df0c6e48c212b2faf15e852b4f.jpg", + "table_caption": [ + "Table 8: Effect of Top-N to Edge-Node Consistency. (We use labeled bank for this experiment.)" + ], + "table_footnote": [], + "table_body": "
N8163264128256
Top-163.463.563.563.563.763.6
", + "bbox": [ + 78, + 599, + 465, + 636 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/12231734a35b1ea52bb8e132754653171e89b8e852a0c17c126450305dc35d22.jpg", + "table_caption": [ + "Table 9: Effect of labeled / unlabeled bank and $\\phi$ in Eq. (6) to Edge-Node Consistency." + ], + "table_footnote": [], + "table_body": "
Labeled Y∞Labeled Y1Unlabeled Y∞Unlabeled Y1
Top-163.763.362.462.2
", + "bbox": [ + 80, + 675, + 465, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Edge-Node Consistency. As we have mentioned, it is computationally very expensive to involve the whole memory bank for the edge-node consistency calculation. Thus, we only adopt the Top-N nearest neighbor to propagate the label. We show this ablation study in Table 8. The performance is very robust with the selection of the number of Top-N, the best results come from $N = 128$ , and the performance is only slightly decreased -0.3% in the worst case. Note that all experiments in Table 8 are performed with the labeled bank; we also have the ablation study to test the effect of the labeled / unlabeled bank and the $\\phi$ in Eq. (6). In Table 9, we can see that the performance with the labeled", + "bbox": [ + 75, + 719, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "bank is much better than the unlabeled bank. Also, the performance with the fully converged $Y_{\\infty}$ is $+0.4\\%$ better than $Y_{1}$ , which shows the necessity of our design.", + "bbox": [ + 498, + 294, + 890, + 339 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/68fcf41140650a82f0006db03fbb0f36bb4473d495b9f609b529aa5e307bfb69.jpg", + "table_caption": [ + "Table 10: Effect of Feature Normalization" + ], + "table_footnote": [], + "table_body": "
Top-1Norm Gap ↓CE Loss ↓
W/o Feat Norm62.21.832.75
W/ Feat Norm63.71.312.63
", + "bbox": [ + 535, + 358, + 854, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feature Normalization. Finally, we show the effect of feature normalization in Table 10. As we mentioned, feature normalization aims to reduce the norm gaps between the weak and strong augmented views. Thus, to validate this, we calculate the mean of the absolute value of $\\| h^s\\| -\\| h^w\\|$ across the whole training process. We name it \"norm gaps\" in Table 10. With the feature normalization, the norm gaps can be reduced from $1.83\\rightarrow 1.31$ . We also calculate the mean of the CE loss Eq. (3) (regardless of the threshold, i.e. $\\tau = 0$ ) across the training process.", + "bbox": [ + 496, + 409, + 890, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 575, + 619, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a new semi-supervised learning algorithm - SimMatchV2, which models consistency regularization from the graph perspective. SimMatchV2 provides four different types of consistency by leveraging the idea of message passing in graph theory and also proposes a simple feature normalization to reduce the gaps between the feature norm of different augmented views. The experiments on different benchmarks and various architectures demonstrate state-of-the-art performance. The limitation of this work is the missing exploration of cooperating with the more advanced SSL training pipelines and stronger data augmentations [5] since it needs more experiments to find the optimal setting. We plan to investigate the pre-training setting in our future work.", + "bbox": [ + 496, + 602, + 890, + 813 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 829, + 651, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the Australian Research Council under Project DP210101859 and the University of Sydney Research Accelerator (SOAR) Prize.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16439", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. arXiv preprint arXiv:2104.13963, 2021. 1, 6, 7", + "[2] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. arXiv preprint arXiv:1911.09785, 2019. 1, 2, 3, 5, 7", + "[3] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. arXiv preprint arXiv:1905.02249, 2019. 1, 2, 3, 5, 7", + "[4] David Berthelot, Rebecca Roelofs, Kihyuk Sohn, Nicholas Carlini, and Alex Kurakin. Adamatch: A unified approach to semi-supervised learning and domain adaptation. arXiv preprint arXiv:2106.04732, 2021. 5", + "[5] Zhaowei Cai, Avinash Ravichandran, Paolo Favaro, Manchen Wang, Davide Modolo, Rahul Bhotika, Zhuowen Tu, and Stefano Soatto. Semi-supervised vision transformers at scale. In NeurIPS, 2022. 3, 7, 8", + "[6] Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Charless Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 194-203, 2021. 3, 6", + "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. 2020. 2, 6", + "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. arXiv preprint arXiv:2104.14294, 2021. 6", + "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709, 2020. 2, 3, 6", + "[10] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey Hinton. Big self-supervised models are strong semi-supervised learners. arXiv preprint arXiv:2006.10029, 2020. 2, 6", + "[11] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2, 6", + "[12] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 2, 3", + "[13] Ching-Yao Chuang, Joshua Robinson, Lin Yen-Chen, Antonio Torralba, and Stefanie Jegelka. Debiased contrastive learning. In Advances in neural information processing systems, 2020. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. volume 15 of Proceedings of Machine Learning Research, pages 215-223, Fort Lauderdale, FL, USA, 11-13 Apr 2011. JMLR Workshop and Conference Proceedings. 5", + "[15] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPRO9, 2009. 1, 5", + "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 7", + "[17] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 1", + "[18] Yue Fan, Anna Kukleva, Dengxin Dai, and Bernt Schiele. Revisiting consistency regularization for semi-supervised learning. International Journal of Computer Vision, pages 1-18, 2022. 5", + "[19] Milton Friedman. The use of ranks to avoid the assumption of normality implicit in the analysis of variance. Journal of the american statistical association, 32(200):675-701, 1937. 5", + "[20] Milton Friedman. A comparison of alternative tests of significance for the problem of m rankings. The Annals of Mathematical Statistics, 11(1):86-92, 1940. 5", + "[21] Yves Grandvalet, Yoshua Bengio, et al. Semi-supervised learning by entropy minimization. CAP, 367:281-296, 2005. 6", + "[22] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. arXiv preprint arXiv:2006.07733, 2020. 2, 3, 6", + "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2", + "[24] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019. 2, 3", + "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. arXiv preprint arXiv:1512.03385, 2015. 6", + "[26] Qianjiang Hu, Xiao Wang, Wei Hu, and Guo-Jun Qi. Adco: Adversarial contrast for efficient learning of unsupervised representations from self-trained negative adversaries. arXiv preprint arXiv:2011.08435, 2020. 2", + "[27] Lang Huang, Shan You, Mingkai Zheng, Fei Wang, Chen Qian, and Toshihiko Yamasaki. Green hierarchical vision transformer for masked image modeling. Advances in Neural Information Processing Systems, 35:19997-20010, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16440", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Lang Huang, Shan You, Mingkai Zheng, Fei Wang, Chen Qian, and Toshihiko Yamasaki. Learning where to learn in cross-view self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14451-14460, 2022. 2", + "[29] Yannis Kalantidis, Mert Bulent Sariyildiz, Noe Pion, Philippe Weinzaepfel, and Diane Larlus. Hard negative mixing for contrastive learning. In Neural Information Processing Systems (NeurIPS), 2020. 2", + "[30] Soroush Abbasi Koohpayegani, Ajinkya Tejankar, and Hamed Pirsiavash. Mean shift for self-supervised learning. arXiv preprint arXiv:2105.07269, 2021. 2", + "[31] A. Krizhevsky. Learning multiple layers of features from tiny images. 2009. 5", + "[32] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 2, 5", + "[33] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 1, 5", + "[34] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regularization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9475-9484, 2021. 3, 5, 6, 7, 8", + "[35] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, 2021. 2, 6", + "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1", + "[37] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7", + "[38] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022. 7", + "[39] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016.5,6", + "[40] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7", + "[41] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979–1993, 2018. 5, 6", + "[42] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 5" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. arXiv preprint arXiv:2001.07685, 2020. 1, 2, 3, 5, 6, 7", + "[44] Jong-Chyi Su and Subhransu Maji. The semi-supervised inaturalist challenge at the fgvc8 workshop, 2021. 7", + "[45] Ilya Sutskever, James Martens, George Dahl, and Geoffrey Hinton. On the importance of initialization and momentum in deep learning. In International conference on machine learning, pages 1139-1147. PMLR, 2013. 5", + "[46] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. arXiv preprint arXiv:1703.01780, 2017. 2, 5", + "[47] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? arXiv preprint arXiv:2005.10243, 2020. 2", + "[48] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 7", + "[49] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 32-42, 2021. 7", + "[50] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. arXiv preprint arXiv:2005.10242, 2020. 2", + "[51] Yidong Wang, Hao Chen, Yue Fan, Wang Sun, Ran Tao, Wenxin Hou, Renjie Wang, Linyi Yang, Zhi Zhou, Lan-Zhe Guo, Heli Qi, Zhen Wu, Yu-Feng Li, Satoshi Nakamura, Wei Ye, Marios Savvides, Bhiksha Raj, Takahiro Shinozaki, Bernt Schiele, Jindong Wang, Xing Xie, and Yue Zhang. Usb: A unified semi-supervised learning benchmark for classification. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 5", + "[52] Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, and Yu-Gang Jiang. Semi-supervised vision transformers. In ECCV, 2022. 7", + "[53] Zhirong Wu, Yuanjun Xiong, X Yu Stella, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2", + "[54] Haoyu Xie, Changqi Wang, Mingkai Zheng, Minjing Dong, Shan You, Chong Fu, and Chang Xu. Boosting semisupervised semantic segmentation with probabilistic representations. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2938-2946, 2023. 2", + "[55] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information" + ], + "bbox": [ + 501, + 90, + 892, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16441", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Processing Systems, volume 33, pages 6256-6268. Curran Associates, Inc., 2020. 5, 6", + "[56] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and improving layer normalization. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 4", + "[57] Yi Xu, Lei Shang, Jinxing Ye, Qi Qian, Yu-Feng Li, Baigui Sun, Hao Li, and Rong Jin. Dash: Semi-supervised learning with dynamic thresholding. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 11525-11536. PMLR, 18-24 Jul 2021. 5", + "[58] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14421-14430, June 2022. 7", + "[59] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. In Advances in Neural Information Processing Systems, 2022. 1", + "[60] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 5", + "[61] Xiaohua Zhai, A. Oliver, A. Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1476-1485, 2019. 6", + "[62] Bowen Zhang, Yidong Wang, Wenxin Hou, HAO WU, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 18408-18419. Curran Associates, Inc., 2021. 5", + "[63] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 2", + "[64] Yuhang Zhang, Xiaopeng Zhang, Robert Qiu, Jie Li, Haohang Xu, Qi Tian, et al. Semi-supervised contrastive learning with similarity co-calibration. arXiv preprint arXiv:2105.07387, 2021. 6", + "[65] Mingkai Zheng, Fei Wang, Shan You, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Weakly supervised contrastive learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10042-10051, October 2021. 6", + "[66] Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14471-14481, June 2022. 2, 3, 5, 6, 7, 8", + "[67] Mingkai Zheng, Shan You, Fei Wang, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Ressl: Re" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "lational self-supervised learning with weak augmentation. arXiv preprint arXiv:2107.09282, 2021. 2", + "[68] Mingkai Zheng, Shan You, Fei Wang, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Relational self-supervised learning. arXiv preprint arXiv:2203.08717, 2022. 2, 6", + "[69] Dengyong Zhou, Olivier Bousquet, Thomas Lal, Jason Weston, and Bernhard Scholkopf. Learning with local and global consistency. In S. Thrun, L. Saul, and B. Scholkopf, editors, Advances in Neural Information Processing Systems, volume 16. MIT Press, 2003. 4" + ], + "bbox": [ + 501, + 92, + 890, + 246 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16442", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/SimMatchV2_ Semi-Supervised Learning with Graph Consistency/4de24035-8115-4285-8e43-089c94dd354f_model.json b/2023/SimMatchV2_ Semi-Supervised Learning with Graph Consistency/4de24035-8115-4285-8e43-089c94dd354f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d70f3c2b5d3c35f068d2bc8598359a104202b48c --- /dev/null +++ b/2023/SimMatchV2_ Semi-Supervised Learning with Graph Consistency/4de24035-8115-4285-8e43-089c94dd354f_model.json @@ -0,0 +1,2389 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.151, + 0.131, + 0.82, + 0.154 + ], + "angle": 0, + "content": "SimMatchV2: Semi-Supervised Learning with Graph Consistency" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.18, + 0.773, + 0.287 + ], + "angle": 0, + "content": "Mingkai Zheng\\(^{1}\\) Shan You\\(^{2*}\\) \nLang Huang\\(^{3}\\) Chen Luo\\(^{5}\\) Fei Wang\\(^{4}\\) Chen Qian\\(^{2}\\) Chang Xu\\(^{1}\\) \n\\(^{1}\\)School of Computer Science, Faculty of Engineering, The University of Sydney \n\\(^{2}\\)SenseTime Research \\({}^{3}\\) The University of Tokyo \n\\(^{4}\\)University of Science and Technology of China \n\\(^{5}\\)State Grid Anhui Electric Power Research Institute" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.474, + 0.718 + ], + "angle": 0, + "content": "Semi-Supervised image classification is one of the most fundamental problem in computer vision, which significantly reduces the need for human labor. In this paper, we introduce a new semi-supervised learning algorithm - SimMatchV2, which formulates various consistency regularizations between labeled and unlabeled data from the graph perspective. In SimMatchV2, we regard the augmented view of a sample as a node, which consists of a label and its corresponding representation. Different nodes are connected with the edges, which are measured by the similarity of the node representations. Inspired by the message passing and node classification in graph theory, we propose four types of consistencies, namely 1) node-node consistency, 2) node-edge consistency, 3) edge-edge consistency, and 4) edge-node consistency. We also uncover that a simple feature normalization can reduce the gaps of the feature norm between different augmented views, significantly improving the performance of SimMatchV2. Our SimMatchV2 has been validated on multiple semi-supervised learning benchmarks. Notably, with ResNet-50 as our backbone and 300 epochs of training, SimMatchV2 achieves \\(71.9\\%\\) and \\(76.2\\%\\) Top-1 Accuracy with \\(1\\%\\) and \\(10\\%\\) labeled examples on ImageNet, which significantly outperforms the previous methods and achieves state-of-the-art performance." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.746, + 0.21, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.772, + 0.47, + 0.879 + ], + "angle": 0, + "content": "In the past decades, deep learning has demonstrated its superior performance in various tasks [15, 36, 17, 59]. However, its outstanding performance is usually based on the fully-supervised setting, which requires a large amount of data and annotations. In a real-world scenario, high-quality data annotations are fairly hard to collect since it is generally expensive, time-consuming, and sometimes" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.324, + 0.889, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.54, + 0.895, + 0.604 + ], + "angle": 0, + "content": "Figure 1: The Top-1 Accuracy of ResNet-50 on ImagetNet with \\(1\\%\\) labeled data. Our proposed SimMatchV2 performs significantly better than previous methods with fewer training epochs. Most importantly, 100 epochs of SimMatchV2 training only takes 14.7 GPU days, but the performance is much higher than PAWS [1] with 300 epochs (68 GPU days)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.894, + 0.688 + ], + "angle": 0, + "content": "require expert knowledge ( e.g. medical images). Semi-supervised learning (SSL) has recently drawn much attention to the research community due to its efficiency in dealing with few labeled data and leveraging massive unlabeled data, significantly reducing the need for human labor." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.901 + ], + "angle": 0, + "content": "For the typical SSL algorithms, one standard paradigm is to train a semantic classifier on the labeled data, then use the classifier to make a prediction on the unlabeled data as the pseudo-label [33]. Besides, a class-level \"consistency regularization\" will be applied to enforce consistency between the class predictions of a model on different perturbations of the same input, which encourages the model to learn more robust and invariant features that generalize well to unseen data. Thus, the quality of pseudo-labels plays a crucial role in consistency regularization, and many works focus on generating stable and high-quality pseudo-labels. For example, MixMatch [3], ReMixMatch [2], and FixMatch [43] are three typical \"consistency regularization\" based methods that significantly improve the performance of the SSL" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.454, + 0.901 + ], + "angle": 0, + "content": "*Correspondence to: Shan You ." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16432" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "tasks. However, these methods highly depend on the classifier trained on the labeled data. A limited number of labels will result in a poor classifier that generates inaccurate pseudo-labels and leads to bad performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.47 + ], + "angle": 0, + "content": "With the recent success of unsupervised visual representation learning [9, 24, 10, 11, 22, 7, 12, 27, 28, 23], the pre-training and fine-tuning paradigm has gained popularity for the semi-supervised learning. In general, the model will be first trained on a large amount of unlabeled data with a proxy task (e.g. instance discrimination), which also imposes the consistency regularization on the instance level (i.e. forcing the model to give similar representation for the different perturbations of the same instance). Then, a classifier will be attached to the pre-trained model, and the whole model will be fine-tuned with the limited labeled data. Many works [29, 35, 26, 67, 68, 13, 50, 47, 30, 54] have been proposed to design a better proxy task to improve the quality of the representation. Apparently, the biggest advantage of this approach is that it does not depend on the supervised trained classifier but still yields high-quality representations for unlabeled data. Nevertheless, the drawback of this approach is also very obvious, since no labels were used during the pre-training, making the model hard to converge. Generally, the pre-training requires more than 800 epochs to converge, which is extremely time-consuming." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.471, + 0.471, + 0.848 + ], + "angle": 0, + "content": "Although many different types of consistency regularizations have been proposed, including class-level [3, 2, 43], instance-level [9, 24, 22], and those that consider both [66, 43], how to establish a complete and effective consistency regularizations remains unresolved. This paper introduces a new SSL algorithm - SimMatchV2, which provides a complete consistency regularization solution for various scenarios. The overall idea is inspired by the message passing and node classification in the graph theory. In SimMatchV2, we consider the augmented view of a sample as a node in the graph which consists of a node label and its corresponding representation. Different nodes are connected with the edges, which are measured by the similarity of the node representations. Thus, the information of different nodes can be efficiently propagated through the graph. To this end, we propose four types of consistency regularization, namely 1) node-node consistency, 2) node-edge consistency, 3) edge-edge consistency, and 4) edge-node consistency. Note that the 1) node-node consistency has been proposed in previous works, but our SimMatchV2 provides a complete and unified solution to design the consistencies. Moreover, we also propose a feature normalization technique to reduce the gaps between the feature norm between different augmented views and greatly improve the performance. Our contributions can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "- We propose SimMatchV2 - a novel semi-supervised algorithm that models various consistency regularizations from the graph perspectives and proposes four" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "types of consistencies to provide a complete solution for the regularization in SSL." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.127, + 0.892, + 0.173 + ], + "angle": 0, + "content": "- We show that a simple feature normalization reduces the gaps between the feature norm of different augmented views and greatly improves the performance." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.178, + 0.892, + 0.269 + ], + "angle": 0, + "content": "- SimMatchV2 sets a new state-of-the-art performance on various SSL benchmarks. Notably, with ResNet50 as our backbone and 300 epochs of training, SimMatchV2 achieves \\(71.9\\%\\) and \\(76.2\\%\\) Top-1 accuracy with \\(1\\%\\) and \\(10\\%\\) labeled data on ImageNet. This result is \\(+4.7\\%\\) and \\(+0.7\\%\\) better than prior arts." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.127, + 0.892, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.285, + 0.638, + 0.301 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Consistency Regularization with Pseudo-Label. Pseudo-label is a method that aims to generate labels for unlabeled data. As has been mentioned, it is often used in conjunction with consistency regularization, which constrains the model from making similar predictions on different perturbations of the same input. Most of these methods focus on generating high-quality pseudo-labels and exploring better perturbation strategies. For example, II-model [32] propose a temporal ensemble strategy for the pseudo-label to reduce the noise in the target. MeanTeacher [46] instead, maintain an exponential moving average (EMA) network to generate more reliable and stable pseudo-labels. Later, MixMatch [3] was proposed and greatly improved the SSL performance. Specifically, it adopts the averaged prediction of multiple strongly augmented views of the same image as the pseudo-label, then incorporates with the Mixup [63] augmentation that generates a convex combination of pairs of images and uses the corresponding mixed pseudo-label as the supervision target. ReMixMatch [2] improves MixMatch by using only a weakly augmented view to generate the pseudo-label and also introduces the distribution alignment to encourage the marginal distribution of predictions on unlabeled data to be consistent with the labeled data. FixMatch [43] abandons these complicated strategies and proposes a confidence-based threshold method to filter out the low-confidence pseudo-labels during the training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Self-Supervised Learning for SSL. Self-supervised learning has demonstrated its outstanding performance in learning visual representations without labels. It is intuitive and natural to leverage the proxy tasks from self-supervised learning to help with SSL tasks. As mentioned, the most straightforward solution is to pre-train the model in a self-supervised manner on a large amount of unlabeled data and then fine-tune the model with the limited labeled data. In this way, the design of the self-supervised task is particularly important. For example, instance discrimination [53] is one of the most popular proxy tasks, which aims to encourage different augmented views of the same image to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16433" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.163, + 0.09, + 0.338, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.227, + 0.313, + 0.239 + ], + "angle": 0, + "content": "(a) Node-Edge Consistency" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.089, + 0.61, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.428, + 0.227, + 0.545, + 0.239 + ], + "angle": 0, + "content": "(b) Edge-Edge Consistency" + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.088, + 0.808, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.227, + 0.776, + 0.239 + ], + "angle": 0, + "content": "(c) Edge-Node Consistency" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.242, + 0.894, + 0.269 + ], + "angle": 0, + "content": "Figure 2: Illustration of various consistency in SimMatchV2. The green and blue nodes are from the weakly and strongly augmented view of the unlabeled data, the red nodes are the other samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.474, + 0.473 + ], + "angle": 0, + "content": "be pulled closer on embedding space but pushes apart all the other images away. BYOL [22] and SimSiam [12] omit the negatives in contrastive learning and leverage an asymmetric structure to avoid representation collapse. Recently, some works (e.g. CoMatch [34] and SimMatch [66]) have been proposed to jointly leverage both class-level and instance-level consistencies and greatly improve the SSL performance, especially in the case when only very limited labeled data is available. There are also many works [6, 5] exploring the two-stage framework (i.e. pre-training followed by fine-tuning with SSL algorithms like FixMatch), which shows the importance of the self-supervised pretrained to the SSL tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.511, + 0.212, + 0.529 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.544, + 0.281, + 0.56 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.47, + 0.864 + ], + "angle": 0, + "content": "Similar to [34, 66], we define a batch of labeled data as \\(\\mathcal{X} = \\{x_b : b \\in (1, \\dots, B)\\}\\), and a batch of unlabeled data as \\(\\mathcal{U} = \\{u_b : b \\in (1, \\dots, \\mu B)\\}\\), where \\(\\mu\\) is the coefficient of unlabeled batch size. Then, a neural network encoder will extract the feature \\(h\\) from the augmented images. We use \\(h_b^w\\) and \\(h_b^s\\) to denote the feature extracted from the weakly and strongly augmented images, respectively. Then, we use \\(W \\in \\mathbb{R}\\) to represent a linear layer followed by a SoftMax function that transforms the feature to a class probability i.e. \\(p = \\mathrm{SoftMax}(W^T h)\\). Finally, we adopt the \\(g(\\cdot)\\) to denote a two-layer non-linear projection head which maps the feature to a low-dimensional representation vector i.e. \\(z = g(h)\\). By following the common setting as in [9, 24, 66, 34], the representation \\(z\\) will be normalized to a unit vector by default. Note that the reason we use \\(z\\) as the representation instead of \\(h\\) is that we need to store a large number of samples as in [24]. The dimension of \\(z\\) is much smaller than \\(h\\) which makes the training more efficient. For example, \\(h\\) is 2048-D and \\(z\\) is 128-D for ResNet-50." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "For labeled data, we perform standard supervised learning with the cross-entropy loss, which can be represented" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.277, + 0.857, + 0.293 + ], + "angle": 0, + "content": "by Eq. (1), where \\(y_{b}\\) is the one-hot ground truth label," + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.297, + 0.892, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = - \\frac {1}{B} \\sum_ {b = 1} ^ {B} y _ {b} \\log p _ {b} ^ {w}. \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.344, + 0.842, + 0.361 + ], + "angle": 0, + "content": "3.2. Consistency Regularization with Graph" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.473 + ], + "angle": 0, + "content": "SimMatchV2 considers the augmented view of a sample as a node in the graph, which consists of a node label \\( p \\) (class prediction for unlabeled data and ground truth for labeled data) and the representation \\( z \\). Different nodes are connected by edges measured by the similarity of the representations. We use strong/weak node to name the node from strongly/weakly augmented views." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.473, + 0.893, + 0.675 + ], + "angle": 0, + "content": "Given a graph with \\(K\\) nodes, we have the node labels \\(\\{p_k : k \\in (1, \\dots, K)\\}\\) and the corresponding representations \\(\\{z_k : k \\in (1, \\dots, K)\\}\\). In our implementation, the graph nodes will be stored in a memory bank as in [24, 66, 34]. We can store both labeled and unlabeled nodes. The node label will be the ground truth for labeled data and \\(p\\) for unlabeled data. Suppose we have an unlabeled weak node with the representation \\(z_b^w\\). We use \\(e_{b,i}^w\\) to denote the edge between the \\(z_b^w\\) and the \\(i^{th}\\) node on the graph, which can be expressed by Eq. (2). \\(t\\) is a temperature parameter that controls the sharpness of the similarity distribution. We also use \\(e_{b,i}^w\\) to denote edges between \\(z_b^w\\) and all the nodes on the graph (i.e. \\(e_{b,i}^w\\) is a scalar and \\(e_b^w \\in R^{1 \\times K}\\))." + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.68, + 0.892, + 0.716 + ], + "angle": 0, + "content": "\\[\ne _ {b, i} ^ {w} = \\frac {\\exp \\left(z _ {b} ^ {w} \\cdot z _ {i} / t\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(z _ {b} ^ {w} \\cdot z _ {k} / t\\right)} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.857 + ], + "angle": 0, + "content": "Node-Node Consistency. (Weak Node vs. Strong Node.) The node-node consistency aims to maintain the consistency of the label between the strong node and weak node of the same instance. Such consistency has been proposed in many previous works [3, 2, 43]. In this work, we have renamed it and integrated it into our graph-based framework. Concretely, we follow the design in previous works; given unlabeled data, the node-node consistency can be expressed by Eq. (3), where \\(\\tau\\) is the confidence threshold." + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.862, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {n n} = - \\frac {1}{\\mu B} \\sum_ {b = 1} ^ {\\mu B} \\mathbb {1} \\left(\\max p _ {b} ^ {w} > \\tau\\right) p _ {b} ^ {w} \\log p _ {b} ^ {s}. \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16434" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.23 + ], + "angle": 0, + "content": "Node-Edge Consistency. (Weak Node vs. Strong Edge.) The purpose of the node-edge consistency is to construct an edge \\( e_b^s \\) by using the representation \\( z_b^s \\) from the strong node, then use the edge \\( e_b^s \\) as the weight to aggregate the node label from the graph. We expect the aggregated node label can be consistent with the label \\( p_b^w \\) from the weak node. Finally, the node-edge consistency will be constrained by the cross-entropy loss between \\( p_b^w \\) and the aggregated label, which can be expressed by Eq. (4)." + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.249, + 0.47, + 0.29 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {n e} = - \\frac {1}{\\mu B} \\sum_ {b = 1} ^ {\\mu B} p _ {b} ^ {w} \\log \\left(\\sum_ {i = 1} ^ {K} p _ {i} e _ {b, i} ^ {s}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.312, + 0.47, + 0.362 + ], + "angle": 0, + "content": "Note that the aggregated label \\(\\sum_{i=1}^{K} p_i e_{b,i}^s\\) does not include the node label \\(p_b^s\\) and \\(p_b^w\\) from itself. The gradient will only be applied on \\(e_{b,i}^s\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.473, + 0.516 + ], + "angle": 0, + "content": "Edge-Edge Consistency. (Weak Edge vs. Strong Edge.) The edge-edge consistency can be regarded as a simple extension of the node-node consistency. Basically, the node-node consistency aims to let the model predict invariant labels for the weak and strong nodes of the same instances. Similarly, the edges should also be invariant with respect to the strong and weak nodes of the same instance. Thus, the edge-edge consistency can be formulated as the cross-entropy loss between the similarity distribution from different augmented views, which can be written as Eq. (5)." + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.536, + 0.47, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {e e} = - \\frac {1}{\\mu B} \\sum_ {b = 1} ^ {\\mu B} \\sum_ {i = 1} ^ {K} e _ {b, i} ^ {w} \\log e _ {b, i} ^ {s} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.597, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Edge-Node Consistency. (Weak Edge vs. Strong Node.) This is somewhat similar to reversed operation to the node-edge consistency. Concretely, node-edge consistency requires the aggregated label \\(\\sum_{i=1}^{K} p_i e_{b,i}^s\\) from the strong node to be consistent with the label \\(p_b^w\\) from the weak node. In contrast, edge-node consistency requires the label \\(p_b^s\\) from the strong node to be consistent with the aggregated label \\(\\sum_{i=1}^{K} p_i e_{b,i}^w\\) from the weak node. Given an unlabeled weak node which has \\(z_b^w\\) and \\(p_b^w\\). Then, we concatenate the unlabeled node with the graph to get \\(Z = [z_b^w, z_1, \\dots, z_k]\\), \\(Y = [p_b^w, p_1, \\dots, p_k]\\). Next, we use the representation matrix \\(Z\\) to obtain the pairwise edge matrix \\(A\\) (a.k.a. affinity matrix) in the same way as Eq. (2). Note that the difference between \\(e_b\\) and \\(A\\) is that \\(e_b\\) only calculates the edges between the unlabeled data with respect to the node in the graph (\\(e_b^w \\in R^{1 \\times K}\\)) but \\(A\\) also contains the edge information within the graphs. (\\(A \\in R^{(1 + K) \\times (1 + K)}\\)). We also want to emphasize that we will make \\(A\\) a hollow matrix (i.e. \\(A_{i,j} = 0\\) if \\(i == j\\)) before normalizing it to avoid the self-loop in the graph. Here, we adopt the transductive" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.707, + 0.108 + ], + "angle": 0, + "content": "inference [69] for aggregation:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.113, + 0.892, + 0.175 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} Y _ {\\phi} = \\alpha \\cdot A Y _ {\\phi - 1} + (1 - \\alpha) \\cdot Y _ {0} \\\\ = (\\alpha \\cdot A) ^ {\\phi - 1} Y _ {0} + (1 - \\alpha) \\sum_ {i = 0} ^ {\\phi - 1} (\\alpha \\cdot A) ^ {i} Y _ {0}, \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.181, + 0.895, + 0.319 + ], + "angle": 0, + "content": "where \\(\\phi\\) is the number of iterations that the label has been propagated, \\(Y_{0}\\) is the initial state \\([p_b^w,p_1,\\dots,p_k]\\), and \\(\\alpha\\) is a parameter in (0,1) controlling the weight of the aggregated label and the initial state \\(Y_{0}\\). The node label will be propagated multiple times until it fully converges. When \\(\\phi \\rightarrow \\infty\\), we have \\((\\alpha \\cdot A)^{\\phi -1} = 0\\) since \\(0 < \\alpha < 1\\). We can also make an approximation such that \\(\\lim_{\\phi \\to \\infty}\\sum_{i = 0}^{\\phi -1}(\\alpha \\cdot A)^{i} = (I - \\alpha \\cdot A)^{-1}\\), where \\(I\\) is the identity matrix. Thus, the converged \\(Y_{\\infty}\\) can be derived as:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.325, + 0.892, + 0.343 + ], + "angle": 0, + "content": "\\[\nY _ {\\infty} = (1 - \\alpha) (I - \\alpha \\cdot A) ^ {- 1} Y _ {0} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.348, + 0.895, + 0.502 + ], + "angle": 0, + "content": "Finally, the converged pseudo-label of the unlabeled data can be obtained from \\( Y_{\\infty}[0] \\) (i.e. the first row vector of \\( Y_{\\infty} \\)). Note that if we take \\( \\alpha = 1 \\), then \\( Y_{1}[0] \\) is identical to \\( \\sum_{i=1}^{K} p_{i} e_{b,i}^{w} \\). Since \\( Y_{\\infty}[0] \\) contains both information from the node label and aggregated label. Thus, we can directly use it to replace the original \\( p_{b}^{w} \\) in Eq. (3). Moreover, we should notice an inverse operator in Eq. (7), which is computationally very expensive for a large matrix. Thus, to reduce the complexity, we will only take the Top-N nearest node from \\( Z \\) and \\( Y \\) instead of using the whole graph." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.502, + 0.895, + 0.7 + ], + "angle": 0, + "content": "Feature Normalization. As we have presented in Eq. (3), one of the training objectives is to minimize the predictions between the weakly and strongly augmented views. However, significant gaps exist between the feature norms of different augmented views. Also, from \\( p = \\operatorname{SoftMax}(W^T h) \\), we can observe that the norm of \\( h \\) does not affect the prediction (i.e. \\( \\operatorname{argmax}(p) \\)). Thus, minimizing the gaps between the feature norm is unnecessary and brings additional complexity to the training. To address this issue, we simply apply a layer normalization [56] on the feature \\( h \\), which can be expressed by Eq. (8). Finally, the linear layer \\( W \\) and non-linear projection head \\( g(\\cdot) \\) will be applied on top of the \\( \\widehat{h} \\)," + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.706, + 0.892, + 0.725 + ], + "angle": 0, + "content": "\\[\n\\widehat {h} = \\operatorname {L a y e r N o r m} (h). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.731, + 0.895, + 0.778 + ], + "angle": 0, + "content": "Finally, the overall objective of SimMatchV2 can be expressed by Eq. (9), where \\(\\lambda_{nn},\\lambda_{ee}\\) , and \\(\\lambda_{ne}\\) are the balancing factors that control the weights of the different losses." + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.791, + 0.892, + 0.808 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {s} + \\lambda_ {n n} \\mathcal {L} _ {n n} + \\lambda_ {n e} \\mathcal {L} _ {n e} + \\lambda_ {e e} \\mathcal {L} _ {e e} \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.823, + 0.634, + 0.839 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.736, + 0.862 + ], + "angle": 0, + "content": "4.1. Classical SSL Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "We will first conduct our experiments on the classic SSL benchmark which is widely adopted by most SSL papers." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16435" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.091, + 0.892, + 0.128 + ], + "angle": 0, + "content": "Table 1: Error rate (%) and Rank on the classical SSL benchmark. The error rate and standard deviation are reported based on three runs. The experimental settings in this table are all based on the USB [51] code base. All results of other methods are directly copied from the USB [51] benchmark paper. The results for fully-supervised training on STL-10 is not provided since it contains an unlabeled set with unknown labels." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.13, + 0.892, + 0.286 + ], + "angle": 0, + "content": "
DatasetCIFAR-10CIFAR-100SVHNSTL-10Friedman rankFinal rankMean error rate
# Label402504000400250010000402501000402501000---
Fully-Supervised Supervised4.57±0.0618.96±0.062.14±0.01----
77.18±1.3256.24±3.4116.10±0.3289.60±0.4358.33±1.4136.83±0.2182.68±1.9124.17±1.6512.19±0.2375.4±0.6655.07±1.8335.42±0.48---
II-Model [32]76.35±1.6948.73±1.0713.63±0.0787.67±0.7956.40±0.6936.73±0.0580.07±1.2213.46±0.616.90±0.2274.89±0.5752.20±2.1131.34±0.6414.171548.20
Pseudo-Labelling [33]75.95±1.8651.12±2.9115.32±0.3588.18±0.8955.37±0.4836.58±0.1275.98±5.3616.47±0.599.37±0.4274.02±0.4751.90±1.8730.77±0.0414.081448.42
Mean Teacher [46]72.42±2.1037.56±4.908.29±0.1079.96±0.5344.37±0.6031.39±0.1149.34±7.903.44±0.023.28±0.0772.90±0.8349.30±2.0927.92±1.6511.751240.01
VAT [41]78.58±2.7828.87±3.6210.90±0.1683.60±4.2146.20±0.8032.14±0.3184.12±3.183.38±0.122.87±0.1873.33±0.4757.78±1.4740.98±0.9612.921345.23
MixMatch [3]35.18±3.8713.00±0.806.55±0.0564.91±3.3439.29±0.1327.74±0.2727.77±5.434.69±0.463.85±0.2849.84±0.5832.05±1.1620.17±0.6711.001127.09
ReMixMatch [2]8.13±0.586.34±0.224.65±0.0941.60±1.4825.72±0.0720.04±0.1316.43±13.775.65±0.355.36±0.5827.87±3.8511.14±0.526.44±0.157.751014.95
UDA [5]10.01±3.345.23±0.084.36±0.0945.48±0.3727.51±0.2823.12±0.455.28±0.421.93±0.031.94±0.0240.09±4.0310.11±1.156.23±0.286.42815.12
FixMatch [3]12.66±4.494.95±0.104.26±0.0145.38±2.0727.71±0.4222.06±0.103.37±1.011.97±0.012.02±0.0338.19±4.768.64±0.845.82±0.065.29514.75
Dash [57]9.29±3.285.10±0.284.36±0.1047.49±1.0527.47±0.3821.89±0.165.26±2.022.01±0.012.08±0.0942.00±4.9410.50±1.376.30±0.496.83915.32
CoMatch [34]6.51±1.185.35±0.144.27±0.1253.41±2.3629.78±0.1122.11±0.228.20±3.522.16±0.042.01±0.0413.74±4.207.63±0.945.71±0.085.75613.41
CRMatch [18]13.62±2.624.61±0.173.65±0.0437.76±1.4524.13±0.1619.89±0.232.60±0.771.98±0.041.95±0.0333.73±1.1714.87±5.096.53±0.364.08213.78
FlexMatch [4]5.29±0.294.97±0.074.24±0.0640.73±1.4426.17±0.1821.75±0.155.42±2.838.74±3.327.90±0.3029.12±5.049.85±1.356.08±0.345.92714.19
AdaMatch [4]5.09±0.215.13±0.054.36±0.0537.08±1.3526.66±0.3321.99±0.156.14±4.352.13±0.042.02±0.0519.95±5.178.59±0.436.01±0.024.79312.18
SimMatch [66]5.38±0.015.36±0.084.41±0.0739.32±0.7226.21±0.3721.50±0.117.60±2.112.48±0.612.05±0.0516.98±4.248.27±0.405.74±0.315.25412.10
SimMatchV2 (Ours)4.90±0.165.04±0.094.33±0.1636.68±0.8626.66±0.3821.37±0.207.92±2.802.92±0.812.85±0.9115.85±2.627.54±0.815.65±0.264.00111.81
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.298, + 0.438, + 0.313 + ], + "angle": 0, + "content": "There are four datasets will be used in this experiment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.316, + 0.468, + 0.406 + ], + "angle": 0, + "content": "- CIFAR-10 [31] consists of 60K 32x32 color images in 10 classes, with 6k images per class. The training and test set contains 5K and 1k samples, respectively. We conduct three different experiments on 40, 250, and 4,000 randomly selected images from the training set as the labeled data and use the rest of the training set as the unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.468, + 0.485 + ], + "angle": 0, + "content": "- CIFAR-100 [31] is a more complex version of CIFAR-10, which contains 100 classes with 600 images each. There are 500 training images and 100 testing images per class. The experimental setting are similar to CIFAR-10 above except using 400, 2,500, and 10K labeled data for training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.488, + 0.468, + 0.578 + ], + "angle": 0, + "content": "- SVHN [42] is a real-world image dataset that contains 32x32 cropped digits from house numbers in Google Street View images. It contains 73,257 training images, with 26,032 testing images and 531,131 extra images. We merge the training and extra set and randomly select 40, 250, and 1,000 from it as the label data to conduct the experiment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.468, + 0.641 + ], + "angle": 0, + "content": "- STL-10 [14] consists of 5K labeled images, 100K unlabeled images, and 8,000 test images. We will randomly select 40, 250, and 1,000 from the 5K images; the rest images will be combined with 100K as the unlabeled data." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.316, + 0.468, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. Our implementation is mainly based on the USB [51] code base for a fair comparison. Most of the hyper-parameters are inherited from [43]. Specifically, we adopt WRN-28-2 [60] for CIFAR-10 and SVHN, WRN-28-8 for CIFAR-100, and WRN-37-2 for STL-10. We use an SGD optimizer with Nesterov momentum [45] for all experiments. The weight decay will be set to \\( 5e^{-4} \\) for most experiments except CIFAR-100, for which we set it as \\( 1e^{-3} \\). For the learning rate schedule, we use a cosine learning rate decay [39], which adjusts the learning rate to \\( 0.03 \\cdot \\cos\\left(\\frac{7\\pi s}{16S}\\right) \\) where \\( s \\) is the current training step, and \\( S \\) is the total number of training steps. The model will be optimized for \\( 2^{20} \\) steps. We use 64 labeled data and 448 unlabeled data for each batch. All experiments will be conducted on a single GPU. The weak and strong augmentation policies are also directly inherited from the USB codebase. Note that the final performance will be reported based on" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.298, + 0.892, + 0.435 + ], + "angle": 0, + "content": "the exponential moving average of the model with a 0.999 decay rate. For SimMatchV2 specific hyper-parameters, we set \\(\\lambda_{nn} = 1, \\lambda_{ne} = 1, \\lambda_{ee} = 1, t = 0.1, \\alpha = 0.1, \\tau = 0.95, n = 8\\). (\\(n\\) is the Top-N nearest node for Eq. (7)) These hyper-parameters are kept unchanged for all experiments. We maintain two memory banks for storing both unlabeled nodes and labeled nodes. The unlabeled bank will be used for calculating the Eq. (4) and Eq. (5). The labeled bank will be used for computing the Eq. (7)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.83 + ], + "angle": 0, + "content": "Performance on Classical SSL Benchmark. The overall results are reported in Table 1. We compare our SimMatchV2 against 14 SSL algorithms. As in Table 1, it is very hard for an algorithm to get the best performance on all settings. Many algorithms already meet or exceed the supervised baseline in some settings (CIFAR-10 and SVHN); the performance difference would not be significant in this case. For a more rigorous and fair comparison, we follow the steps from USB [51], which adopts the Friedman rank [19, 20] as the final evaluation metric. We also report the mean error rate for all settings. As can be seen, our SimMatchV2 achieves the best performance on 4 out of 12 settings and also has the highest Friedman rank and lowest mean error rate. CRMatch [18] also achieves a very good performance that ranked second in the final rank. It has the best performance on 5 out of the 12 settings, but it is less stable when the number of labeled data is limited, and the mean error rate is much worse than SimMatchV2. AdaMatch [4] and SimMatch [66] has a more stable performance under various conditions and they have a very close mean error rate compared to SimMatchV2. However, these two methods did not achieve any best results in this experiment, leading to a worse Friedman rank. Overall, the highest Friedman rank and lowest mean error rate for SimMatchV2 should demonstrate superiority in terms of both performance and stability." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.736, + 0.861 + ], + "angle": 0, + "content": "4.2. Experiments on ImageNet" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Next, we evaluate our SimMatchV2 on the large-scaled ImageNet dataset [15], which contains 1,000 categories" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16436" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.894, + 0.117 + ], + "angle": 0, + "content": "Table 2: Top-1 an Top-5 Accuracy with ResNet-50 on ImageNet with \\(1\\%\\) and \\(10\\%\\) settings. Note that the underline in the parameters - inference columns denotes the methods which require more parameters than standard ResNet-50 during the inference stage." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.117, + 0.885, + 0.348 + ], + "angle": 0, + "content": "
Pre-training AlgorithmSemi-supervised AlgorithmPre-training EpochsSemi-supervised EpochsParamters trainParamters inferenceTop-1 Label fractionTop-5 Label fraction
1%10%1%10%
NoneVAT+EntMin. [41, 21, 61]-~10025.6M25.6M--47.083.4
S4L-Rotation [61]-~20025.6M25.6M--53.483.8
UDA [55]-~53025.6M25.6M-68.8-88.8
FixMatch [43]-~30025.6M25.6M-71.5-89.1
FixMatch-EMAN [6]-~30025.6M25.6M-72.8-90.3
SsCL [64]-~80028.0M25.6M60.272.182.890.9
CoMatch [34]-~40028.0M25.6M66.073.686.491.6
SimMatch [66]-~40028.0M25.6M67.274.487.191.7
PCL [35]Fine-tune~200~2023.8M25.6M--75.385.6
SimCLR [9]~1000~6028.0M25.6M48.365.675.587.8
SimCLR V2 [10]~800~6032.2M29.8M57.968.482.589.2
BYOL [22]~1000~5035.1M25.6M53.268.878.489.0
SwAV [7]~800~2028.0M25.6M53.970.278.589.9
WCL [65]~800~6032.2M29.8M65.072.086.391.2
MoCo V2 [11]Fine-tune~800~2028.0M25.6M49.866.177.287.9
FixMatch-EMAN [6]~800~30028.0M25.6M61.473.982.191.0
PAWS [7]Fine-tune~300~5036.1M29.8M66.575.5--
NoneSimMatchV2 (Ours)-~10028.0M25.6M69.974.888.791.7
-~30028.0M25.6M71.976.290.092.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.47, + 0.465 + ], + "angle": 0, + "content": "with 1.2M training images and 50K validation images. Concretely, we follow previous works [34, 66, 1, 6] to take \\(1\\%\\) and \\(10\\%\\) samples from the training images as our labeled data and use the rest samples as the unlabeled data. To avoid the effects of different selections on the labeled data, we follow [9, 10, 6, 1, 65] to adopt a pre-defined labeled / unlabeled splits1 that has been adopted in most SSL papers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.465, + 0.471, + 0.753 + ], + "angle": 0, + "content": "Implementation Details. For ImageNet experiments, most of our hyper-parameter and settings are directly inherited from [66]. Specifically, we use ResNet-50 [25] as our default backbone and use the SGD with Nesterov momentum as the optimizer. We set the weight decay to 0.0001 during the training. For the learning rate, we warm up the model for 5 epochs until it reaches 0.03, and it will be decayed to 0 with the cosine scheduler [39]. By following [6, 66], we use 64 labeled data and 320 unlabeled data in each batch; the data will be evenly distributed on 8 GPUs. For the hyper-parameter settings, we set \\(\\lambda_{nn} = 10\\), \\(\\lambda_{ne} = 10\\), \\(\\lambda_{ee} = 10\\), \\(t = 0.1\\), \\(\\alpha = 0.1\\), \\(\\tau = 0.7\\), \\(n = 128\\) for \\(1\\%\\) setting, and we use \\(\\lambda_{nn} = 5\\), \\(\\lambda_{ne} = 5\\) for \\(10\\%\\) since we found the performance is slightly better. For strong augmentation, we adopt a multi-crop strategy as in [68], which contains five crops with different sizes (224x224, 192x192, 160x160, 128x128, and 96x96); these crops will be augmented with the policy in [8]. For weak augmentation, we follow the same strategy in [66, 34]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.754, + 0.47, + 0.86 + ], + "angle": 0, + "content": "Performance on ImageNet. The performance has been shown in Table 2. With 100 epochs of training, SimMatchV2 achieves \\(70.4\\%\\) and \\(74.8\\%\\) Top-1 accuracy with \\(1\\%\\) and \\(10\\%\\) labeled data, which already surpass the performance of CoMatch [34], and SimMatch [66] with 400 epochs. With 300 epochs of training, SimMatchV2 achieves \\(72.1\\%\\) and \\(76.2\\%\\) Top-1 accuracy, which improves absolute" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.358, + 0.895, + 0.465 + ], + "angle": 0, + "content": "+5.2% and +0.7% over PAWS [1]. As we can observe that the improvement of 10% setting is much lesser than 1% setting. We doubt that this is because the performance of 76.2% is already very close to the fully supervised setting (76.5%), which should be the upper bound of the SSL tasks. Most importantly, our SimMatchV2 established a new state-of-the-art performance without any pre-training." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.477, + 0.894, + 0.515 + ], + "angle": 0, + "content": "Table 3: Runtime Comparison. MC denotes whether we use the multicrop strategy. The experiments are performed on \\(1\\%\\) ImageNet with 100 epochs of training. \\(\\downarrow\\) indicates lower is better and \\(\\uparrow\\) means higher is better." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.516, + 0.895, + 0.59 + ], + "angle": 0, + "content": "
MethodMCGPU Memory ↓GPU days ↓1% Top-1 ↑
CoMatch [34]✗81G10.361.1
SimMatch [66]✗40G7.761.2
SimMatchV2✗42G7.863.7
PAWS [1]✓1048G22.663.8
SimMatchV2✓83G14.769.9
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Efficiency Comparison. Next, We also compare the training efficiency for different methods. The experiments are conducted with 8 Nvidia GPUs except for PAWS [1], which requires at least 64 GPUs to run. We use mixed precision training for all methods. Note that CoMatch [34] and SimMatch [66] have their own labeled / unlabeled splits, which makes the comparison a bit unfair. In this experiment, we adopt their officially released codebase to reproduce the performance on the same splits." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We show the training statistics in Table 3. When we omit the multi-crops, the computational cost of SimMatchV2 is slightly higher compared to SimMatch [66] (i.e. +2G memory, +0.1 GPU days), and it is much more efficient than CoMatch [34]. We can also observe that SimMatchV2 has much better performance (+2.5%) than CoMatch [34] and SimMatch [66] even without the multi-crops. Moreover, when the multi-crops are included, the training time of SimMatchV2 is nearly doubled (7.8 days → 14.7 days), but it is still only 65% of PAWS [1] (14.7 days / 22.6 days). The extra training cost of PAWS [1] is from the support" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.451, + 0.901 + ], + "angle": 0, + "content": "1https://www.tensorflow.org/datasets/catalog/Imagenet2012 Subset" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16437" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "set containing \\(960 \\times 7\\) labeled data per batch. Most importantly, our SimMatchV2 achieves similar performance with PAWS (63.7% vs. 63.8%) [1] with only 35% training time (7.8 days / 22.6 days), and has an absolute improvement of +6.6% with 65% training time." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.173, + 0.47, + 0.198 + ], + "angle": 0, + "content": "Table 4: More experiments on different architectures. Note that we do not use any pre-training in this experiment." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.199, + 0.462, + 0.317 + ], + "angle": 0, + "content": "
MethodArchParamsEpochs1%10%
Semifommer [52]ViT-S + Conv40M600-75.5
ProbPseudo Mixup [5]ViT-Small22M100-70.9
SimMatchV2ViT-Small22M10063.775.9
ProbPseudo Mixup [5]ViT-Base86M100-73.5
SimMatchV2ViT-Base86M10065.376.8
ProbPseudo Mixup [5]ConvNext-T28M100-74.1
SimMatchV2ConvNext-T28M10071.377.8
ProbPseudo Mixup [5]ConvNext-S50M100-75.1
SimMatchV2ConvNext-S50M10072.378.4
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.47, + 0.581 + ], + "angle": 0, + "content": "Working with More Advanced Architectures. With the recent successes in model architecture design, many more powerful models have been proposed, especially the vision transformer and its variant [16, 48, 49, 38, 37]. Thus, we further apply our SimMatchV2 on ViT-Small [48], ViT-Base [16], ConvNext-T [38], and ConvNext-S [38] to show the generality on different architectures. In this experiment, we follow [5] to use 128 labeled data and 640 unlabeled per batch. We also adopt the AdamW [40] optimizer with the learning rate of \\(1.25e^{-4}\\) and weight decay of 0.1. We present the results in Table 4. As we can see, SimMatchV2 with a better model can generally produce a better result. The exploration of cooperating with more advanced SSL pipeline [5] and stronger data augmentations (e.g. ProbPseudo Mixup [5]) will be left as a future work since it requires more hyper-parameter turnings and more sophisticated designs, which is not the focus of this paper." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.587, + 0.356, + 0.602 + ], + "angle": 0, + "content": "4.3. Experiments on Semi-iNat 2021" + }, + { + "type": "table_caption", + "bbox": [ + 0.111, + 0.605, + 0.434, + 0.618 + ], + "angle": 0, + "content": "Table 5: Accuracy for real-world dataset - Semi-iNat 2021." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.619, + 0.451, + 0.741 + ], + "angle": 0, + "content": "
MethodSemi-iNat 2021
With Random InitWith MoCo Init
Top1Top5Top1Top5
Supervised19.0935.8534.9657.11
MixMatch [3]16.8930.83--
+CCSSL [58]19.6535.09--
FixMatch [43]21.4137.6540.3060.05
+CCSSL [58]31.2152.2541.2864.30
CoMatch [34]20.9438.9638.9461.85
+CCSSL [58]24.1243.2339.8563.68
SimMatchV2 (Ours)43.0563.8153.1373.40
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Finally, we also test the performance of SimMatchV2 on Semi-iNat 2021 [44], which contains 810 different categories with 9,721 labeled data, 313,248 unlabeled data, and 4,050 validation data. Unlike the dataset we used in previous experiments, Semi-iNat 2021 is designed to expose some challenges encountered in a real-world scenario, which contains highly similar classes, class imbalance, and domain mismatch between the labeled and unlabeled data. Such a more realistic dataset should be a better benchmark to reflect the performance of the different methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "In this experiment, most of the hyper-parameters and implementation details are the same as our ImageNet setting, except that we do not use the multi-crops strategy for fair comparison. The model will be optimized for \\(512 \\times 1024\\) steps as in [58]. We also follow the setting in [58] to conduct two experiments to report the training from scratch performance and training from pre-trained model performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.892, + 0.334 + ], + "angle": 0, + "content": "We show the results in Table 5. Our SimMatchV2 achieves \\(43.05\\%\\) Top-1 accuracy when training from scratch, which is \\(+11.84\\%\\) better than the state-of-the-art method (FixMatch + CCSL). When the MoCo pre-trained model is adopted, our SimMatchV2 achieves a \\(53.13\\%\\) Top-1 accuracy, which is \\(+10.08\\%\\) better than the from-scratch performance, and \\(+11.85\\%\\) better than FixMatch + CCSL. We believe this experiment will further demonstrate the superiority of SimMatchV2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.351, + 0.649, + 0.368 + ], + "angle": 0, + "content": "5. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.892, + 0.483 + ], + "angle": 0, + "content": "In this section, we will empirically study our SimMatchV2 based on various conditions and show the effect of each component and hyper-parameter sensitivities of our methods. For all experiments in this section, we adopt the ResNet-50 as our backbone and train the model on ImageNet for 100 epochs on the \\(1\\%\\) labeled data. The multicrops are not used by default unless we mentioned." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.892, + 0.65 + ], + "angle": 0, + "content": "Training Dynamics. We first show the training dynamics of SimMatchV2 compared to SimMatch [66] and CoMatch [34]. We visualize the pseudo-labels accuracy, unlabeled data accuracy, and validation accuracy in Figure 3. As we can observe that the pseudo-labels from SimMatchV2 have a higher quality than SimMatch [66], especially in the late stage of the training. Both of these two methods generate better pseudo-labels than CoMatch [34]. We can also notice that SimMatchV2 consistently has better accuracy for the unlabeled data and validation data, which shows the advantages of our method." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.659, + 0.892, + 0.696 + ], + "angle": 0, + "content": "Table 6: Effect of each component in SimMatchV2. N and E denote the node and edge, respectively. N-E means the node-edge consistency and vice versa. Feat norm means feature normalization." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.698, + 0.888, + 0.77 + ], + "angle": 0, + "content": "
N-NN-EE-EE-NFeat NormTop-1Improvement
✓52.9-
✓✓60.0+7.1
✓✓✓60.7+0.7
✓✓✓✓62.2+1.5
✓✓✓✓✓63.7+1.5
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effect of Each Components. In table 6, we show the effect of each design in SimMatchV2. Basically, if only node-node consistency is adopted, the setting is equivalent to the FixMatch [43] with distribution alignment [2], which is our baseline in the first row of the table; Next, we can see that the node-edge, edge-node, and edge-node consistencies give \\(+7.1\\%\\), \\(+0.7\\%\\), \\(+1.5\\%\\) improvements, respectively. Finally, the feature normalization also boosts the performance" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16438" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.14, + 0.09, + 0.368, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.234, + 0.331, + 0.247 + ], + "angle": 0, + "content": "(a) Pseudo-Labels Accuracy" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.091, + 0.597, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.233, + 0.562, + 0.246 + ], + "angle": 0, + "content": "(b) Unlabeled Data Accuracy" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.091, + 0.828, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.233, + 0.78, + 0.247 + ], + "angle": 0, + "content": "(c) Validation Accuracy" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.249, + 0.895, + 0.287 + ], + "angle": 0, + "content": "Figure 3: In this Figure, we compare the training dynamics of SimMatchV2 with SimMatch [66], and CoMatch [34] from 3 different perspectives. a) The accuracy of \\( p^w \\) that pass the threshold \\( \\tau \\). b) The accuracy of \\( p^w \\) regardless of the threshold. c) The accuracy on the validation data. All experiments are in 1% ImageNet setting with 100 epochs training." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.295, + 0.471, + 0.326 + ], + "angle": 0, + "content": "+1.5%. This experiment should demonstrate that these different designs are all very effective for SSL tasks." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.333, + 0.47, + 0.37 + ], + "angle": 0, + "content": "Table 7: Effect of memory bank size on the performance. \\(5\\mathrm{k}-163\\mathrm{K}\\) are setting that we use unlabeled bank; labeled bank means we use all labeled bank (All labeled data 12K for \\(1\\%\\) ImageNet)." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.372, + 0.468, + 0.408 + ], + "angle": 0, + "content": "
Bank Size5K10K40K82K163KLabeled Bank (12K)
Top-163.163.263.563.763.663.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.47, + 0.567 + ], + "angle": 0, + "content": "Node-Edge and Edge-Edge Consistency. We will study the node-edge and edge-edge consistency together since the memory bank size is their primary factor. The experimental results are shown in Table 7. We can observe that when unlabeled data is used, the performance will be slightly increased with a larger memory bank, and the best memory bank size is \\(82\\mathrm{K}\\). We can also notice that we get slightly lower performance with \\(12\\mathrm{K}\\) labeled data. Therefore, we decide to use the unlabeled data for node-edge and edge-edge consistency." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.574, + 0.47, + 0.6 + ], + "angle": 0, + "content": "Table 8: Effect of Top-N to Edge-Node Consistency. (We use labeled bank for this experiment.)" + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.601, + 0.466, + 0.637 + ], + "angle": 0, + "content": "
N8163264128256
Top-163.463.563.563.563.763.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.649, + 0.468, + 0.674 + ], + "angle": 0, + "content": "Table 9: Effect of labeled / unlabeled bank and \\(\\phi\\) in Eq. (6) to Edge-Node Consistency." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.676, + 0.466, + 0.713 + ], + "angle": 0, + "content": "
Labeled Y∞Labeled Y1Unlabeled Y∞Unlabeled Y1
Top-163.763.362.462.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Edge-Node Consistency. As we have mentioned, it is computationally very expensive to involve the whole memory bank for the edge-node consistency calculation. Thus, we only adopt the Top-N nearest neighbor to propagate the label. We show this ablation study in Table 8. The performance is very robust with the selection of the number of Top-N, the best results come from \\( N = 128 \\), and the performance is only slightly decreased -0.3% in the worst case. Note that all experiments in Table 8 are performed with the labeled bank; we also have the ablation study to test the effect of the labeled / unlabeled bank and the \\( \\phi \\) in Eq. (6). In Table 9, we can see that the performance with the labeled" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.295, + 0.892, + 0.34 + ], + "angle": 0, + "content": "bank is much better than the unlabeled bank. Also, the performance with the fully converged \\( Y_{\\infty} \\) is \\( +0.4\\% \\) better than \\( Y_{1} \\), which shows the necessity of our design." + }, + { + "type": "table_caption", + "bbox": [ + 0.577, + 0.345, + 0.816, + 0.358 + ], + "angle": 0, + "content": "Table 10: Effect of Feature Normalization" + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.359, + 0.855, + 0.406 + ], + "angle": 0, + "content": "
Top-1Norm Gap ↓CE Loss ↓
W/o Feat Norm62.21.832.75
W/ Feat Norm63.71.312.63
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.892, + 0.562 + ], + "angle": 0, + "content": "Feature Normalization. Finally, we show the effect of feature normalization in Table 10. As we mentioned, feature normalization aims to reduce the norm gaps between the weak and strong augmented views. Thus, to validate this, we calculate the mean of the absolute value of \\( \\| h^s\\| -\\| h^w\\| \\) across the whole training process. We name it \"norm gaps\" in Table 10. With the feature normalization, the norm gaps can be reduced from \\( 1.83\\rightarrow 1.31 \\). We also calculate the mean of the CE loss Eq. (3) (regardless of the threshold, i.e. \\( \\tau = 0 \\)) across the training process." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.577, + 0.62, + 0.593 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.892, + 0.814 + ], + "angle": 0, + "content": "In this work, we propose a new semi-supervised learning algorithm - SimMatchV2, which models consistency regularization from the graph perspective. SimMatchV2 provides four different types of consistency by leveraging the idea of message passing in graph theory and also proposes a simple feature normalization to reduce the gaps between the feature norm of different augmented views. The experiments on different benchmarks and various architectures demonstrate state-of-the-art performance. The limitation of this work is the missing exploration of cooperating with the more advanced SSL training pipelines and stronger data augmentations [5] since it needs more experiments to find the optimal setting. We plan to investigate the pre-training setting in our future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.652, + 0.847 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by the Australian Research Council under Project DP210101859 and the University of Sydney Research Accelerator (SOAR) Prize." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16439" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.186 + ], + "angle": 0, + "content": "[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. arXiv preprint arXiv:2104.13963, 2021. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. arXiv preprint arXiv:1911.09785, 2019. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.313 + ], + "angle": 0, + "content": "[3] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. arXiv preprint arXiv:1905.02249, 2019. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.315, + 0.471, + 0.372 + ], + "angle": 0, + "content": "[4] David Berthelot, Rebecca Roelofs, Kihyuk Sohn, Nicholas Carlini, and Alex Kurakin. Adamatch: A unified approach to semi-supervised learning and domain adaptation. arXiv preprint arXiv:2106.04732, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.47, + 0.429 + ], + "angle": 0, + "content": "[5] Zhaowei Cai, Avinash Ravichandran, Paolo Favaro, Manchen Wang, Davide Modolo, Rahul Bhotika, Zhuowen Tu, and Stefano Soatto. Semi-supervised vision transformers at scale. In NeurIPS, 2022. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.43, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[6] Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Charless Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 194-203, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.515, + 0.47, + 0.57 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.572, + 0.47, + 0.628 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. arXiv preprint arXiv:2104.14294, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.63, + 0.47, + 0.685 + ], + "angle": 0, + "content": "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.687, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[10] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey Hinton. Big self-supervised models are strong semi-supervised learners. arXiv preprint arXiv:2006.10029, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.745, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[11] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[12] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[13] Ching-Yao Chuang, Joshua Robinson, Lin Yen-Chen, Antonio Torralba, and Stefanie Jegelka. Debiased contrastive learning. In Advances in neural information processing systems, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[14] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. volume 15 of Proceedings of Machine Learning Research, pages 215-223, Fort Lauderdale, FL, USA, 11-13 Apr 2011. JMLR Workshop and Conference Proceedings. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[15] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPRO9, 2009. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[17] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[18] Yue Fan, Anna Kukleva, Dengxin Dai, and Bernt Schiele. Revisiting consistency regularization for semi-supervised learning. International Journal of Computer Vision, pages 1-18, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[19] Milton Friedman. The use of ranks to avoid the assumption of normality implicit in the analysis of variance. Journal of the american statistical association, 32(200):675-701, 1937. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.497 + ], + "angle": 0, + "content": "[20] Milton Friedman. A comparison of alternative tests of significance for the problem of m rankings. The Annals of Mathematical Statistics, 11(1):86-92, 1940. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[21] Yves Grandvalet, Yoshua Bengio, et al. Semi-supervised learning by entropy minimization. CAP, 367:281-296, 2005. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[22] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. arXiv preprint arXiv:2006.07733, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.622, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[24] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. arXiv preprint arXiv:1512.03385, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[26] Qianjiang Hu, Xiao Wang, Wei Hu, and Guo-Jun Qi. Adco: Adversarial contrast for efficient learning of unsupervised representations from self-trained negative adversaries. arXiv preprint arXiv:2011.08435, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Lang Huang, Shan You, Mingkai Zheng, Fei Wang, Chen Qian, and Toshihiko Yamasaki. Green hierarchical vision transformer for masked image modeling. Advances in Neural Information Processing Systems, 35:19997-20010, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16440" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[28] Lang Huang, Shan You, Mingkai Zheng, Fei Wang, Chen Qian, and Toshihiko Yamasaki. Learning where to learn in cross-view self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14451-14460, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[29] Yannis Kalantidis, Mert Bulent Sariyildiz, Noe Pion, Philippe Weinzaepfel, and Diane Larlus. Hard negative mixing for contrastive learning. In Neural Information Processing Systems (NeurIPS), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.472, + 0.26 + ], + "angle": 0, + "content": "[30] Soroush Abbasi Koohpayegani, Ajinkya Tejankar, and Hamed Pirsiavash. Mean shift for self-supervised learning. arXiv preprint arXiv:2105.07269, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.472, + 0.288 + ], + "angle": 0, + "content": "[31] A. Krizhevsky. Learning multiple layers of features from tiny images. 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.472, + 0.328 + ], + "angle": 0, + "content": "[32] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.472, + 0.385 + ], + "angle": 0, + "content": "[33] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.472, + 0.453 + ], + "angle": 0, + "content": "[34] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regularization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9475-9484, 2021. 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.472, + 0.51 + ], + "angle": 0, + "content": "[35] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.472, + 0.58 + ], + "angle": 0, + "content": "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.472, + 0.65 + ], + "angle": 0, + "content": "[37] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.472, + 0.718 + ], + "angle": 0, + "content": "[38] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.472, + 0.761 + ], + "angle": 0, + "content": "[39] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016.5,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.472, + 0.79 + ], + "angle": 0, + "content": "[40] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.472, + 0.858 + ], + "angle": 0, + "content": "[41] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979–1993, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[42] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[43] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. arXiv preprint arXiv:2001.07685, 2020. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.192 + ], + "angle": 0, + "content": "[44] Jong-Chyi Su and Subhransu Maji. The semi-supervised inaturalist challenge at the fgvc8 workshop, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.893, + 0.248 + ], + "angle": 0, + "content": "[45] Ilya Sutskever, James Martens, George Dahl, and Geoffrey Hinton. On the importance of initialization and momentum in deep learning. In International conference on machine learning, pages 1139-1147. PMLR, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.893, + 0.305 + ], + "angle": 0, + "content": "[46] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. arXiv preprint arXiv:1703.01780, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.893, + 0.362 + ], + "angle": 0, + "content": "[47] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? arXiv preprint arXiv:2005.10243, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.893, + 0.433 + ], + "angle": 0, + "content": "[48] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.893, + 0.503 + ], + "angle": 0, + "content": "[49] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 32-42, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.893, + 0.547 + ], + "angle": 0, + "content": "[50] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. arXiv preprint arXiv:2005.10242, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.893, + 0.671 + ], + "angle": 0, + "content": "[51] Yidong Wang, Hao Chen, Yue Fan, Wang Sun, Ran Tao, Wenxin Hou, Renjie Wang, Linyi Yang, Zhi Zhou, Lan-Zhe Guo, Heli Qi, Zhen Wu, Yu-Feng Li, Satoshi Nakamura, Wei Ye, Marios Savvides, Bhiksha Raj, Takahiro Shinozaki, Bernt Schiele, Jindong Wang, Xing Xie, and Yue Zhang. Usb: A unified semi-supervised learning benchmark for classification. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.673, + 0.893, + 0.716 + ], + "angle": 0, + "content": "[52] Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, and Yu-Gang Jiang. Semi-supervised vision transformers. In ECCV, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.893, + 0.773 + ], + "angle": 0, + "content": "[53] Zhirong Wu, Yuanjun Xiong, X Yu Stella, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.893, + 0.845 + ], + "angle": 0, + "content": "[54] Haoyu Xie, Changqi Wang, Mingkai Zheng, Minjing Dong, Shan You, Chong Fu, and Chang Xu. Boosting semisupervised semantic segmentation with probabilistic representations. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2938-2946, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[55] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16441" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Processing Systems, volume 33, pages 6256-6268. Curran Associates, Inc., 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[56] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and improving layer normalization. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[57] Yi Xu, Lei Shang, Jinxing Ye, Qi Qian, Yu-Feng Li, Baigui Sun, Hao Li, and Rong Jin. Dash: Semi-supervised learning with dynamic thresholding. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 11525-11536. PMLR, 18-24 Jul 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.387 + ], + "angle": 0, + "content": "[58] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14421-14430, June 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[59] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. In Advances in Neural Information Processing Systems, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.46 + ], + "angle": 0, + "content": "[60] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[61] Xiaohua Zhai, A. Oliver, A. Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1476-1485, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.52, + 0.469, + 0.628 + ], + "angle": 0, + "content": "[62] Bowen Zhang, Yidong Wang, Wenxin Hou, HAO WU, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 18408-18419. Curran Associates, Inc., 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[63] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[64] Yuhang Zhang, Xiaopeng Zhang, Robert Qiu, Jie Li, Haohang Xu, Qi Tian, et al. Semi-supervised contrastive learning with similarity co-calibration. arXiv preprint arXiv:2105.07387, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[65] Mingkai Zheng, Fei Wang, Shan You, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Weakly supervised contrastive learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10042-10051, October 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[66] Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14471-14481, June 2022. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[67] Mingkai Zheng, Shan You, Fei Wang, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Ressl: Re" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.891, + 0.12 + ], + "angle": 0, + "content": "lational self-supervised learning with weak augmentation. arXiv preprint arXiv:2107.09282, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[68] Mingkai Zheng, Shan You, Fei Wang, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Relational self-supervised learning. arXiv preprint arXiv:2203.08717, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[69] Dengyong Zhou, Olivier Bousquet, Thomas Lal, Jason Weston, and Bernhard Scholkopf. Learning with local and global consistency. In S. Thrun, L. Saul, and B. Scholkopf, editors, Advances in Neural Information Processing Systems, volume 16. MIT Press, 2003. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16442" + } + ] +] \ No newline at end of file diff --git a/2023/SimMatchV2_ Semi-Supervised Learning with Graph Consistency/4de24035-8115-4285-8e43-089c94dd354f_origin.pdf b/2023/SimMatchV2_ Semi-Supervised Learning with Graph Consistency/4de24035-8115-4285-8e43-089c94dd354f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9bcb9ef3650a2663cab980d4fb87016a3f18d2b4 --- /dev/null +++ b/2023/SimMatchV2_ Semi-Supervised Learning with Graph Consistency/4de24035-8115-4285-8e43-089c94dd354f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c48fdfccbc022c389d8b6eccd1cc6b560c4755cbc1c898af05a73ff09b5028e6 +size 443199 diff --git a/2023/SimMatchV2_ Semi-Supervised Learning with Graph Consistency/full.md b/2023/SimMatchV2_ Semi-Supervised Learning with Graph Consistency/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ac5d8e41689b12e529c554598244c7ea08fd2a2f --- /dev/null +++ b/2023/SimMatchV2_ Semi-Supervised Learning with Graph Consistency/full.md @@ -0,0 +1,327 @@ +# SimMatchV2: Semi-Supervised Learning with Graph Consistency + +Mingkai Zheng $^{1}$ Shan You $^{2*}$ +Lang Huang $^{3}$ Chen Luo $^{5}$ Fei Wang $^{4}$ Chen Qian $^{2}$ Chang Xu $^{1}$ $^{1}$ School of Computer Science, Faculty of Engineering, The University of Sydney + $^{2}$ SenseTime Research ${}^{3}$ The University of Tokyo + $^{4}$ University of Science and Technology of China + $^{5}$ State Grid Anhui Electric Power Research Institute + +# Abstract + +Semi-Supervised image classification is one of the most fundamental problem in computer vision, which significantly reduces the need for human labor. In this paper, we introduce a new semi-supervised learning algorithm - SimMatchV2, which formulates various consistency regularizations between labeled and unlabeled data from the graph perspective. In SimMatchV2, we regard the augmented view of a sample as a node, which consists of a label and its corresponding representation. Different nodes are connected with the edges, which are measured by the similarity of the node representations. Inspired by the message passing and node classification in graph theory, we propose four types of consistencies, namely 1) node-node consistency, 2) node-edge consistency, 3) edge-edge consistency, and 4) edge-node consistency. We also uncover that a simple feature normalization can reduce the gaps of the feature norm between different augmented views, significantly improving the performance of SimMatchV2. Our SimMatchV2 has been validated on multiple semi-supervised learning benchmarks. Notably, with ResNet-50 as our backbone and 300 epochs of training, SimMatchV2 achieves $71.9\%$ and $76.2\%$ Top-1 Accuracy with $1\%$ and $10\%$ labeled examples on ImageNet, which significantly outperforms the previous methods and achieves state-of-the-art performance. + +# 1. Introduction + +In the past decades, deep learning has demonstrated its superior performance in various tasks [15, 36, 17, 59]. However, its outstanding performance is usually based on the fully-supervised setting, which requires a large amount of data and annotations. In a real-world scenario, high-quality data annotations are fairly hard to collect since it is generally expensive, time-consuming, and sometimes + +![](images/453fecbdc9a8f91f96af2826d42b2b7def9719b7d570adfa980eabb118ec480b.jpg) +Figure 1: The Top-1 Accuracy of ResNet-50 on ImagetNet with $1\%$ labeled data. Our proposed SimMatchV2 performs significantly better than previous methods with fewer training epochs. Most importantly, 100 epochs of SimMatchV2 training only takes 14.7 GPU days, but the performance is much higher than PAWS [1] with 300 epochs (68 GPU days). + +require expert knowledge ( e.g. medical images). Semi-supervised learning (SSL) has recently drawn much attention to the research community due to its efficiency in dealing with few labeled data and leveraging massive unlabeled data, significantly reducing the need for human labor. + +For the typical SSL algorithms, one standard paradigm is to train a semantic classifier on the labeled data, then use the classifier to make a prediction on the unlabeled data as the pseudo-label [33]. Besides, a class-level "consistency regularization" will be applied to enforce consistency between the class predictions of a model on different perturbations of the same input, which encourages the model to learn more robust and invariant features that generalize well to unseen data. Thus, the quality of pseudo-labels plays a crucial role in consistency regularization, and many works focus on generating stable and high-quality pseudo-labels. For example, MixMatch [3], ReMixMatch [2], and FixMatch [43] are three typical "consistency regularization" based methods that significantly improve the performance of the SSL + +tasks. However, these methods highly depend on the classifier trained on the labeled data. A limited number of labels will result in a poor classifier that generates inaccurate pseudo-labels and leads to bad performance. + +With the recent success of unsupervised visual representation learning [9, 24, 10, 11, 22, 7, 12, 27, 28, 23], the pre-training and fine-tuning paradigm has gained popularity for the semi-supervised learning. In general, the model will be first trained on a large amount of unlabeled data with a proxy task (e.g. instance discrimination), which also imposes the consistency regularization on the instance level (i.e. forcing the model to give similar representation for the different perturbations of the same instance). Then, a classifier will be attached to the pre-trained model, and the whole model will be fine-tuned with the limited labeled data. Many works [29, 35, 26, 67, 68, 13, 50, 47, 30, 54] have been proposed to design a better proxy task to improve the quality of the representation. Apparently, the biggest advantage of this approach is that it does not depend on the supervised trained classifier but still yields high-quality representations for unlabeled data. Nevertheless, the drawback of this approach is also very obvious, since no labels were used during the pre-training, making the model hard to converge. Generally, the pre-training requires more than 800 epochs to converge, which is extremely time-consuming. + +Although many different types of consistency regularizations have been proposed, including class-level [3, 2, 43], instance-level [9, 24, 22], and those that consider both [66, 43], how to establish a complete and effective consistency regularizations remains unresolved. This paper introduces a new SSL algorithm - SimMatchV2, which provides a complete consistency regularization solution for various scenarios. The overall idea is inspired by the message passing and node classification in the graph theory. In SimMatchV2, we consider the augmented view of a sample as a node in the graph which consists of a node label and its corresponding representation. Different nodes are connected with the edges, which are measured by the similarity of the node representations. Thus, the information of different nodes can be efficiently propagated through the graph. To this end, we propose four types of consistency regularization, namely 1) node-node consistency, 2) node-edge consistency, 3) edge-edge consistency, and 4) edge-node consistency. Note that the 1) node-node consistency has been proposed in previous works, but our SimMatchV2 provides a complete and unified solution to design the consistencies. Moreover, we also propose a feature normalization technique to reduce the gaps between the feature norm between different augmented views and greatly improve the performance. Our contributions can be summarized as: + +- We propose SimMatchV2 - a novel semi-supervised algorithm that models various consistency regularizations from the graph perspectives and proposes four + +types of consistencies to provide a complete solution for the regularization in SSL. + +- We show that a simple feature normalization reduces the gaps between the feature norm of different augmented views and greatly improves the performance. +- SimMatchV2 sets a new state-of-the-art performance on various SSL benchmarks. Notably, with ResNet50 as our backbone and 300 epochs of training, SimMatchV2 achieves $71.9\%$ and $76.2\%$ Top-1 accuracy with $1\%$ and $10\%$ labeled data on ImageNet. This result is $+4.7\%$ and $+0.7\%$ better than prior arts. + +# 2. Related work + +Consistency Regularization with Pseudo-Label. Pseudo-label is a method that aims to generate labels for unlabeled data. As has been mentioned, it is often used in conjunction with consistency regularization, which constrains the model from making similar predictions on different perturbations of the same input. Most of these methods focus on generating high-quality pseudo-labels and exploring better perturbation strategies. For example, II-model [32] propose a temporal ensemble strategy for the pseudo-label to reduce the noise in the target. MeanTeacher [46] instead, maintain an exponential moving average (EMA) network to generate more reliable and stable pseudo-labels. Later, MixMatch [3] was proposed and greatly improved the SSL performance. Specifically, it adopts the averaged prediction of multiple strongly augmented views of the same image as the pseudo-label, then incorporates with the Mixup [63] augmentation that generates a convex combination of pairs of images and uses the corresponding mixed pseudo-label as the supervision target. ReMixMatch [2] improves MixMatch by using only a weakly augmented view to generate the pseudo-label and also introduces the distribution alignment to encourage the marginal distribution of predictions on unlabeled data to be consistent with the labeled data. FixMatch [43] abandons these complicated strategies and proposes a confidence-based threshold method to filter out the low-confidence pseudo-labels during the training. + +Self-Supervised Learning for SSL. Self-supervised learning has demonstrated its outstanding performance in learning visual representations without labels. It is intuitive and natural to leverage the proxy tasks from self-supervised learning to help with SSL tasks. As mentioned, the most straightforward solution is to pre-train the model in a self-supervised manner on a large amount of unlabeled data and then fine-tune the model with the limited labeled data. In this way, the design of the self-supervised task is particularly important. For example, instance discrimination [53] is one of the most popular proxy tasks, which aims to encourage different augmented views of the same image to + +![](images/7feb6baedfb99e70b404cc269aa65c21deb7b2be80231cd5f771684e470a29fc.jpg) +(a) Node-Edge Consistency + +![](images/3a5c414e81235605c68b93b9bed1b65faf0beb7ac894e0401a41738429c8868f.jpg) +(b) Edge-Edge Consistency +Figure 2: Illustration of various consistency in SimMatchV2. The green and blue nodes are from the weakly and strongly augmented view of the unlabeled data, the red nodes are the other samples. + +![](images/77973cd795f807da3d6f12147c88fb6ef683e0988013a0217b41324e101c206e.jpg) +(c) Edge-Node Consistency + +be pulled closer on embedding space but pushes apart all the other images away. BYOL [22] and SimSiam [12] omit the negatives in contrastive learning and leverage an asymmetric structure to avoid representation collapse. Recently, some works (e.g. CoMatch [34] and SimMatch [66]) have been proposed to jointly leverage both class-level and instance-level consistencies and greatly improve the SSL performance, especially in the case when only very limited labeled data is available. There are also many works [6, 5] exploring the two-stage framework (i.e. pre-training followed by fine-tuning with SSL algorithms like FixMatch), which shows the importance of the self-supervised pretrained to the SSL tasks. + +# 3. Methodology + +# 3.1. Problem Formulation + +Similar to [34, 66], we define a batch of labeled data as $\mathcal{X} = \{x_b : b \in (1, \dots, B)\}$ , and a batch of unlabeled data as $\mathcal{U} = \{u_b : b \in (1, \dots, \mu B)\}$ , where $\mu$ is the coefficient of unlabeled batch size. Then, a neural network encoder will extract the feature $h$ from the augmented images. We use $h_b^w$ and $h_b^s$ to denote the feature extracted from the weakly and strongly augmented images, respectively. Then, we use $W \in \mathbb{R}$ to represent a linear layer followed by a SoftMax function that transforms the feature to a class probability i.e. $p = \mathrm{SoftMax}(W^T h)$ . Finally, we adopt the $g(\cdot)$ to denote a two-layer non-linear projection head which maps the feature to a low-dimensional representation vector i.e. $z = g(h)$ . By following the common setting as in [9, 24, 66, 34], the representation $z$ will be normalized to a unit vector by default. Note that the reason we use $z$ as the representation instead of $h$ is that we need to store a large number of samples as in [24]. The dimension of $z$ is much smaller than $h$ which makes the training more efficient. For example, $h$ is 2048-D and $z$ is 128-D for ResNet-50. + +For labeled data, we perform standard supervised learning with the cross-entropy loss, which can be represented + +by Eq. (1), where $y_{b}$ is the one-hot ground truth label, + +$$ +\mathcal {L} _ {s} = - \frac {1}{B} \sum_ {b = 1} ^ {B} y _ {b} \log p _ {b} ^ {w}. \tag {1} +$$ + +# 3.2. Consistency Regularization with Graph + +SimMatchV2 considers the augmented view of a sample as a node in the graph, which consists of a node label $p$ (class prediction for unlabeled data and ground truth for labeled data) and the representation $z$ . Different nodes are connected by edges measured by the similarity of the representations. We use strong/weak node to name the node from strongly/weakly augmented views. + +Given a graph with $K$ nodes, we have the node labels $\{p_k : k \in (1, \dots, K)\}$ and the corresponding representations $\{z_k : k \in (1, \dots, K)\}$ . In our implementation, the graph nodes will be stored in a memory bank as in [24, 66, 34]. We can store both labeled and unlabeled nodes. The node label will be the ground truth for labeled data and $p$ for unlabeled data. Suppose we have an unlabeled weak node with the representation $z_b^w$ . We use $e_{b,i}^w$ to denote the edge between the $z_b^w$ and the $i^{th}$ node on the graph, which can be expressed by Eq. (2). $t$ is a temperature parameter that controls the sharpness of the similarity distribution. We also use $e_{b,i}^w$ to denote edges between $z_b^w$ and all the nodes on the graph (i.e. $e_{b,i}^w$ is a scalar and $e_b^w \in R^{1 \times K}$ ). + +$$ +e _ {b, i} ^ {w} = \frac {\exp \left(z _ {b} ^ {w} \cdot z _ {i} / t\right)}{\sum_ {k = 1} ^ {K} \exp \left(z _ {b} ^ {w} \cdot z _ {k} / t\right)} \tag {2} +$$ + +Node-Node Consistency. (Weak Node vs. Strong Node.) The node-node consistency aims to maintain the consistency of the label between the strong node and weak node of the same instance. Such consistency has been proposed in many previous works [3, 2, 43]. In this work, we have renamed it and integrated it into our graph-based framework. Concretely, we follow the design in previous works; given unlabeled data, the node-node consistency can be expressed by Eq. (3), where $\tau$ is the confidence threshold. + +$$ +\mathcal {L} _ {n n} = - \frac {1}{\mu B} \sum_ {b = 1} ^ {\mu B} \mathbb {1} \left(\max p _ {b} ^ {w} > \tau\right) p _ {b} ^ {w} \log p _ {b} ^ {s}. \tag {3} +$$ + +Node-Edge Consistency. (Weak Node vs. Strong Edge.) The purpose of the node-edge consistency is to construct an edge $e_b^s$ by using the representation $z_b^s$ from the strong node, then use the edge $e_b^s$ as the weight to aggregate the node label from the graph. We expect the aggregated node label can be consistent with the label $p_b^w$ from the weak node. Finally, the node-edge consistency will be constrained by the cross-entropy loss between $p_b^w$ and the aggregated label, which can be expressed by Eq. (4). + +$$ +\mathcal {L} _ {n e} = - \frac {1}{\mu B} \sum_ {b = 1} ^ {\mu B} p _ {b} ^ {w} \log \left(\sum_ {i = 1} ^ {K} p _ {i} e _ {b, i} ^ {s}\right) \tag {4} +$$ + +Note that the aggregated label $\sum_{i=1}^{K} p_i e_{b,i}^s$ does not include the node label $p_b^s$ and $p_b^w$ from itself. The gradient will only be applied on $e_{b,i}^s$ . + +Edge-Edge Consistency. (Weak Edge vs. Strong Edge.) The edge-edge consistency can be regarded as a simple extension of the node-node consistency. Basically, the node-node consistency aims to let the model predict invariant labels for the weak and strong nodes of the same instances. Similarly, the edges should also be invariant with respect to the strong and weak nodes of the same instance. Thus, the edge-edge consistency can be formulated as the cross-entropy loss between the similarity distribution from different augmented views, which can be written as Eq. (5). + +$$ +\mathcal {L} _ {e e} = - \frac {1}{\mu B} \sum_ {b = 1} ^ {\mu B} \sum_ {i = 1} ^ {K} e _ {b, i} ^ {w} \log e _ {b, i} ^ {s} \tag {5} +$$ + +Edge-Node Consistency. (Weak Edge vs. Strong Node.) This is somewhat similar to reversed operation to the node-edge consistency. Concretely, node-edge consistency requires the aggregated label $\sum_{i=1}^{K} p_i e_{b,i}^s$ from the strong node to be consistent with the label $p_b^w$ from the weak node. In contrast, edge-node consistency requires the label $p_b^s$ from the strong node to be consistent with the aggregated label $\sum_{i=1}^{K} p_i e_{b,i}^w$ from the weak node. Given an unlabeled weak node which has $z_b^w$ and $p_b^w$ . Then, we concatenate the unlabeled node with the graph to get $Z = [z_b^w, z_1, \dots, z_k]$ , $Y = [p_b^w, p_1, \dots, p_k]$ . Next, we use the representation matrix $Z$ to obtain the pairwise edge matrix $A$ (a.k.a. affinity matrix) in the same way as Eq. (2). Note that the difference between $e_b$ and $A$ is that $e_b$ only calculates the edges between the unlabeled data with respect to the node in the graph ( $e_b^w \in R^{1 \times K}$ ) but $A$ also contains the edge information within the graphs. ( $A \in R^{(1 + K) \times (1 + K)}$ ). We also want to emphasize that we will make $A$ a hollow matrix (i.e. $A_{i,j} = 0$ if $i == j$ ) before normalizing it to avoid the self-loop in the graph. Here, we adopt the transductive + +inference [69] for aggregation: + +$$ +\begin{array}{l} Y _ {\phi} = \alpha \cdot A Y _ {\phi - 1} + (1 - \alpha) \cdot Y _ {0} \\ = (\alpha \cdot A) ^ {\phi - 1} Y _ {0} + (1 - \alpha) \sum_ {i = 0} ^ {\phi - 1} (\alpha \cdot A) ^ {i} Y _ {0}, \tag {6} \\ \end{array} +$$ + +where $\phi$ is the number of iterations that the label has been propagated, $Y_{0}$ is the initial state $[p_b^w,p_1,\dots,p_k]$ , and $\alpha$ is a parameter in (0,1) controlling the weight of the aggregated label and the initial state $Y_{0}$ . The node label will be propagated multiple times until it fully converges. When $\phi \rightarrow \infty$ , we have $(\alpha \cdot A)^{\phi -1} = 0$ since $0 < \alpha < 1$ . We can also make an approximation such that $\lim_{\phi \to \infty}\sum_{i = 0}^{\phi -1}(\alpha \cdot A)^{i} = (I - \alpha \cdot A)^{-1}$ , where $I$ is the identity matrix. Thus, the converged $Y_{\infty}$ can be derived as: + +$$ +Y _ {\infty} = (1 - \alpha) (I - \alpha \cdot A) ^ {- 1} Y _ {0} \tag {7} +$$ + +Finally, the converged pseudo-label of the unlabeled data can be obtained from $Y_{\infty}[0]$ (i.e. the first row vector of $Y_{\infty}$ ). Note that if we take $\alpha = 1$ , then $Y_{1}[0]$ is identical to $\sum_{i=1}^{K} p_{i} e_{b,i}^{w}$ . Since $Y_{\infty}[0]$ contains both information from the node label and aggregated label. Thus, we can directly use it to replace the original $p_{b}^{w}$ in Eq. (3). Moreover, we should notice an inverse operator in Eq. (7), which is computationally very expensive for a large matrix. Thus, to reduce the complexity, we will only take the Top-N nearest node from $Z$ and $Y$ instead of using the whole graph. + +Feature Normalization. As we have presented in Eq. (3), one of the training objectives is to minimize the predictions between the weakly and strongly augmented views. However, significant gaps exist between the feature norms of different augmented views. Also, from $p = \operatorname{SoftMax}(W^T h)$ , we can observe that the norm of $h$ does not affect the prediction (i.e. $\operatorname{argmax}(p)$ ). Thus, minimizing the gaps between the feature norm is unnecessary and brings additional complexity to the training. To address this issue, we simply apply a layer normalization [56] on the feature $h$ , which can be expressed by Eq. (8). Finally, the linear layer $W$ and non-linear projection head $g(\cdot)$ will be applied on top of the $\widehat{h}$ , + +$$ +\widehat {h} = \operatorname {L a y e r N o r m} (h). \tag {8} +$$ + +Finally, the overall objective of SimMatchV2 can be expressed by Eq. (9), where $\lambda_{nn},\lambda_{ee}$ , and $\lambda_{ne}$ are the balancing factors that control the weights of the different losses. + +$$ +\mathcal {L} = \mathcal {L} _ {s} + \lambda_ {n n} \mathcal {L} _ {n n} + \lambda_ {n e} \mathcal {L} _ {n e} + \lambda_ {e e} \mathcal {L} _ {e e} \tag {9} +$$ + +# 4. Experiments + +# 4.1. Classical SSL Benchmark + +We will first conduct our experiments on the classic SSL benchmark which is widely adopted by most SSL papers. + +Table 1: Error rate (%) and Rank on the classical SSL benchmark. The error rate and standard deviation are reported based on three runs. The experimental settings in this table are all based on the USB [51] code base. All results of other methods are directly copied from the USB [51] benchmark paper. The results for fully-supervised training on STL-10 is not provided since it contains an unlabeled set with unknown labels. + +
DatasetCIFAR-10CIFAR-100SVHNSTL-10Friedman rankFinal rankMean error rate
# Label402504000400250010000402501000402501000---
Fully-Supervised Supervised4.57±0.0618.96±0.062.14±0.01----
77.18±1.3256.24±3.4116.10±0.3289.60±0.4358.33±1.4136.83±0.2182.68±1.9124.17±1.6512.19±0.2375.4±0.6655.07±1.8335.42±0.48---
II-Model [32]76.35±1.6948.73±1.0713.63±0.0787.67±0.7956.40±0.6936.73±0.0580.07±1.2213.46±0.616.90±0.2274.89±0.5752.20±2.1131.34±0.6414.171548.20
Pseudo-Labelling [33]75.95±1.8651.12±2.9115.32±0.3588.18±0.8955.37±0.4836.58±0.1275.98±5.3616.47±0.599.37±0.4274.02±0.4751.90±1.8730.77±0.0414.081448.42
Mean Teacher [46]72.42±2.1037.56±4.908.29±0.1079.96±0.5344.37±0.6031.39±0.1149.34±7.903.44±0.023.28±0.0772.90±0.8349.30±2.0927.92±1.6511.751240.01
VAT [41]78.58±2.7828.87±3.6210.90±0.1683.60±4.2146.20±0.8032.14±0.3184.12±3.183.38±0.122.87±0.1873.33±0.4757.78±1.4740.98±0.9612.921345.23
MixMatch [3]35.18±3.8713.00±0.806.55±0.0564.91±3.3439.29±0.1327.74±0.2727.77±5.434.69±0.463.85±0.2849.84±0.5832.05±1.1620.17±0.6711.001127.09
ReMixMatch [2]8.13±0.586.34±0.224.65±0.0941.60±1.4825.72±0.0720.04±0.1316.43±13.775.65±0.355.36±0.5827.87±3.8511.14±0.526.44±0.157.751014.95
UDA [5]10.01±3.345.23±0.084.36±0.0945.48±0.3727.51±0.2823.12±0.455.28±0.421.93±0.031.94±0.0240.09±4.0310.11±1.156.23±0.286.42815.12
FixMatch [3]12.66±4.494.95±0.104.26±0.0145.38±2.0727.71±0.4222.06±0.103.37±1.011.97±0.012.02±0.0338.19±4.768.64±0.845.82±0.065.29514.75
Dash [57]9.29±3.285.10±0.284.36±0.1047.49±1.0527.47±0.3821.89±0.165.26±2.022.01±0.012.08±0.0942.00±4.9410.50±1.376.30±0.496.83915.32
CoMatch [34]6.51±1.185.35±0.144.27±0.1253.41±2.3629.78±0.1122.11±0.228.20±3.522.16±0.042.01±0.0413.74±4.207.63±0.945.71±0.085.75613.41
CRMatch [18]13.62±2.624.61±0.173.65±0.0437.76±1.4524.13±0.1619.89±0.232.60±0.771.98±0.041.95±0.0333.73±1.1714.87±5.096.53±0.364.08213.78
FlexMatch [4]5.29±0.294.97±0.074.24±0.0640.73±1.4426.17±0.1821.75±0.155.42±2.838.74±3.327.90±0.3029.12±5.049.85±1.356.08±0.345.92714.19
AdaMatch [4]5.09±0.215.13±0.054.36±0.0537.08±1.3526.66±0.3321.99±0.156.14±4.352.13±0.042.02±0.0519.95±5.178.59±0.436.01±0.024.79312.18
SimMatch [66]5.38±0.015.36±0.084.41±0.0739.32±0.7226.21±0.3721.50±0.117.60±2.112.48±0.612.05±0.0516.98±4.248.27±0.405.74±0.315.25412.10
SimMatchV2 (Ours)4.90±0.165.04±0.094.33±0.1636.68±0.8626.66±0.3821.37±0.207.92±2.802.92±0.812.85±0.9115.85±2.627.54±0.815.65±0.264.00111.81
+ +There are four datasets will be used in this experiment. + +- CIFAR-10 [31] consists of 60K 32x32 color images in 10 classes, with 6k images per class. The training and test set contains 5K and 1k samples, respectively. We conduct three different experiments on 40, 250, and 4,000 randomly selected images from the training set as the labeled data and use the rest of the training set as the unlabeled data. +- CIFAR-100 [31] is a more complex version of CIFAR-10, which contains 100 classes with 600 images each. There are 500 training images and 100 testing images per class. The experimental setting are similar to CIFAR-10 above except using 400, 2,500, and 10K labeled data for training. +- SVHN [42] is a real-world image dataset that contains 32x32 cropped digits from house numbers in Google Street View images. It contains 73,257 training images, with 26,032 testing images and 531,131 extra images. We merge the training and extra set and randomly select 40, 250, and 1,000 from it as the label data to conduct the experiment. +- STL-10 [14] consists of 5K labeled images, 100K unlabeled images, and 8,000 test images. We will randomly select 40, 250, and 1,000 from the 5K images; the rest images will be combined with 100K as the unlabeled data. + +Implementation Details. Our implementation is mainly based on the USB [51] code base for a fair comparison. Most of the hyper-parameters are inherited from [43]. Specifically, we adopt WRN-28-2 [60] for CIFAR-10 and SVHN, WRN-28-8 for CIFAR-100, and WRN-37-2 for STL-10. We use an SGD optimizer with Nesterov momentum [45] for all experiments. The weight decay will be set to $5e^{-4}$ for most experiments except CIFAR-100, for which we set it as $1e^{-3}$ . For the learning rate schedule, we use a cosine learning rate decay [39], which adjusts the learning rate to $0.03 \cdot \cos\left(\frac{7\pi s}{16S}\right)$ where $s$ is the current training step, and $S$ is the total number of training steps. The model will be optimized for $2^{20}$ steps. We use 64 labeled data and 448 unlabeled data for each batch. All experiments will be conducted on a single GPU. The weak and strong augmentation policies are also directly inherited from the USB codebase. Note that the final performance will be reported based on + +the exponential moving average of the model with a 0.999 decay rate. For SimMatchV2 specific hyper-parameters, we set $\lambda_{nn} = 1, \lambda_{ne} = 1, \lambda_{ee} = 1, t = 0.1, \alpha = 0.1, \tau = 0.95, n = 8$ . ( $n$ is the Top-N nearest node for Eq. (7)) These hyper-parameters are kept unchanged for all experiments. We maintain two memory banks for storing both unlabeled nodes and labeled nodes. The unlabeled bank will be used for calculating the Eq. (4) and Eq. (5). The labeled bank will be used for computing the Eq. (7). + +Performance on Classical SSL Benchmark. The overall results are reported in Table 1. We compare our SimMatchV2 against 14 SSL algorithms. As in Table 1, it is very hard for an algorithm to get the best performance on all settings. Many algorithms already meet or exceed the supervised baseline in some settings (CIFAR-10 and SVHN); the performance difference would not be significant in this case. For a more rigorous and fair comparison, we follow the steps from USB [51], which adopts the Friedman rank [19, 20] as the final evaluation metric. We also report the mean error rate for all settings. As can be seen, our SimMatchV2 achieves the best performance on 4 out of 12 settings and also has the highest Friedman rank and lowest mean error rate. CRMatch [18] also achieves a very good performance that ranked second in the final rank. It has the best performance on 5 out of the 12 settings, but it is less stable when the number of labeled data is limited, and the mean error rate is much worse than SimMatchV2. AdaMatch [4] and SimMatch [66] has a more stable performance under various conditions and they have a very close mean error rate compared to SimMatchV2. However, these two methods did not achieve any best results in this experiment, leading to a worse Friedman rank. Overall, the highest Friedman rank and lowest mean error rate for SimMatchV2 should demonstrate superiority in terms of both performance and stability. + +# 4.2. Experiments on ImageNet + +Next, we evaluate our SimMatchV2 on the large-scaled ImageNet dataset [15], which contains 1,000 categories + +Table 2: Top-1 an Top-5 Accuracy with ResNet-50 on ImageNet with $1\%$ and $10\%$ settings. Note that the underline in the parameters - inference columns denotes the methods which require more parameters than standard ResNet-50 during the inference stage. + +
Pre-training AlgorithmSemi-supervised AlgorithmPre-training EpochsSemi-supervised EpochsParamters trainParamters inferenceTop-1 Label fractionTop-5 Label fraction
1%10%1%10%
NoneVAT+EntMin. [41, 21, 61]-~10025.6M25.6M--47.083.4
S4L-Rotation [61]-~20025.6M25.6M--53.483.8
UDA [55]-~53025.6M25.6M-68.8-88.8
FixMatch [43]-~30025.6M25.6M-71.5-89.1
FixMatch-EMAN [6]-~30025.6M25.6M-72.8-90.3
SsCL [64]-~80028.0M25.6M60.272.182.890.9
CoMatch [34]-~40028.0M25.6M66.073.686.491.6
SimMatch [66]-~40028.0M25.6M67.274.487.191.7
PCL [35]Fine-tune~200~2023.8M25.6M--75.385.6
SimCLR [9]~1000~6028.0M25.6M48.365.675.587.8
SimCLR V2 [10]~800~6032.2M29.8M57.968.482.589.2
BYOL [22]~1000~5035.1M25.6M53.268.878.489.0
SwAV [7]~800~2028.0M25.6M53.970.278.589.9
WCL [65]~800~6032.2M29.8M65.072.086.391.2
MoCo V2 [11]Fine-tune~800~2028.0M25.6M49.866.177.287.9
FixMatch-EMAN [6]~800~30028.0M25.6M61.473.982.191.0
PAWS [7]Fine-tune~300~5036.1M29.8M66.575.5--
NoneSimMatchV2 (Ours)-~10028.0M25.6M69.974.888.791.7
-~30028.0M25.6M71.976.290.092.2
+ +with 1.2M training images and 50K validation images. Concretely, we follow previous works [34, 66, 1, 6] to take $1\%$ and $10\%$ samples from the training images as our labeled data and use the rest samples as the unlabeled data. To avoid the effects of different selections on the labeled data, we follow [9, 10, 6, 1, 65] to adopt a pre-defined labeled / unlabeled splits1 that has been adopted in most SSL papers. + +Implementation Details. For ImageNet experiments, most of our hyper-parameter and settings are directly inherited from [66]. Specifically, we use ResNet-50 [25] as our default backbone and use the SGD with Nesterov momentum as the optimizer. We set the weight decay to 0.0001 during the training. For the learning rate, we warm up the model for 5 epochs until it reaches 0.03, and it will be decayed to 0 with the cosine scheduler [39]. By following [6, 66], we use 64 labeled data and 320 unlabeled data in each batch; the data will be evenly distributed on 8 GPUs. For the hyper-parameter settings, we set $\lambda_{nn} = 10$ , $\lambda_{ne} = 10$ , $\lambda_{ee} = 10$ , $t = 0.1$ , $\alpha = 0.1$ , $\tau = 0.7$ , $n = 128$ for $1\%$ setting, and we use $\lambda_{nn} = 5$ , $\lambda_{ne} = 5$ for $10\%$ since we found the performance is slightly better. For strong augmentation, we adopt a multi-crop strategy as in [68], which contains five crops with different sizes (224x224, 192x192, 160x160, 128x128, and 96x96); these crops will be augmented with the policy in [8]. For weak augmentation, we follow the same strategy in [66, 34]. + +Performance on ImageNet. The performance has been shown in Table 2. With 100 epochs of training, SimMatchV2 achieves $70.4\%$ and $74.8\%$ Top-1 accuracy with $1\%$ and $10\%$ labeled data, which already surpass the performance of CoMatch [34], and SimMatch [66] with 400 epochs. With 300 epochs of training, SimMatchV2 achieves $72.1\%$ and $76.2\%$ Top-1 accuracy, which improves absolute + ++5.2% and +0.7% over PAWS [1]. As we can observe that the improvement of 10% setting is much lesser than 1% setting. We doubt that this is because the performance of 76.2% is already very close to the fully supervised setting (76.5%), which should be the upper bound of the SSL tasks. Most importantly, our SimMatchV2 established a new state-of-the-art performance without any pre-training. + +Table 3: Runtime Comparison. MC denotes whether we use the multicrop strategy. The experiments are performed on $1\%$ ImageNet with 100 epochs of training. $\downarrow$ indicates lower is better and $\uparrow$ means higher is better. + +
MethodMCGPU Memory ↓GPU days ↓1% Top-1 ↑
CoMatch [34]✗81G10.361.1
SimMatch [66]✗40G7.761.2
SimMatchV2✗42G7.863.7
PAWS [1]✓1048G22.663.8
SimMatchV2✓83G14.769.9
+ +Efficiency Comparison. Next, We also compare the training efficiency for different methods. The experiments are conducted with 8 Nvidia GPUs except for PAWS [1], which requires at least 64 GPUs to run. We use mixed precision training for all methods. Note that CoMatch [34] and SimMatch [66] have their own labeled / unlabeled splits, which makes the comparison a bit unfair. In this experiment, we adopt their officially released codebase to reproduce the performance on the same splits. + +We show the training statistics in Table 3. When we omit the multi-crops, the computational cost of SimMatchV2 is slightly higher compared to SimMatch [66] (i.e. +2G memory, +0.1 GPU days), and it is much more efficient than CoMatch [34]. We can also observe that SimMatchV2 has much better performance (+2.5%) than CoMatch [34] and SimMatch [66] even without the multi-crops. Moreover, when the multi-crops are included, the training time of SimMatchV2 is nearly doubled (7.8 days → 14.7 days), but it is still only 65% of PAWS [1] (14.7 days / 22.6 days). The extra training cost of PAWS [1] is from the support + +set containing $960 \times 7$ labeled data per batch. Most importantly, our SimMatchV2 achieves similar performance with PAWS (63.7% vs. 63.8%) [1] with only 35% training time (7.8 days / 22.6 days), and has an absolute improvement of +6.6% with 65% training time. + +Table 4: More experiments on different architectures. Note that we do not use any pre-training in this experiment. + +
MethodArchParamsEpochs1%10%
Semifommer [52]ViT-S + Conv40M600-75.5
ProbPseudo Mixup [5]ViT-Small22M100-70.9
SimMatchV2ViT-Small22M10063.775.9
ProbPseudo Mixup [5]ViT-Base86M100-73.5
SimMatchV2ViT-Base86M10065.376.8
ProbPseudo Mixup [5]ConvNext-T28M100-74.1
SimMatchV2ConvNext-T28M10071.377.8
ProbPseudo Mixup [5]ConvNext-S50M100-75.1
SimMatchV2ConvNext-S50M10072.378.4
+ +Working with More Advanced Architectures. With the recent successes in model architecture design, many more powerful models have been proposed, especially the vision transformer and its variant [16, 48, 49, 38, 37]. Thus, we further apply our SimMatchV2 on ViT-Small [48], ViT-Base [16], ConvNext-T [38], and ConvNext-S [38] to show the generality on different architectures. In this experiment, we follow [5] to use 128 labeled data and 640 unlabeled per batch. We also adopt the AdamW [40] optimizer with the learning rate of $1.25e^{-4}$ and weight decay of 0.1. We present the results in Table 4. As we can see, SimMatchV2 with a better model can generally produce a better result. The exploration of cooperating with more advanced SSL pipeline [5] and stronger data augmentations (e.g. ProbPseudo Mixup [5]) will be left as a future work since it requires more hyper-parameter turnings and more sophisticated designs, which is not the focus of this paper. + +# 4.3. Experiments on Semi-iNat 2021 + +Table 5: Accuracy for real-world dataset - Semi-iNat 2021. + +
MethodSemi-iNat 2021
With Random InitWith MoCo Init
Top1Top5Top1Top5
Supervised19.0935.8534.9657.11
MixMatch [3]16.8930.83--
+CCSSL [58]19.6535.09--
FixMatch [43]21.4137.6540.3060.05
+CCSSL [58]31.2152.2541.2864.30
CoMatch [34]20.9438.9638.9461.85
+CCSSL [58]24.1243.2339.8563.68
SimMatchV2 (Ours)43.0563.8153.1373.40
+ +Finally, we also test the performance of SimMatchV2 on Semi-iNat 2021 [44], which contains 810 different categories with 9,721 labeled data, 313,248 unlabeled data, and 4,050 validation data. Unlike the dataset we used in previous experiments, Semi-iNat 2021 is designed to expose some challenges encountered in a real-world scenario, which contains highly similar classes, class imbalance, and domain mismatch between the labeled and unlabeled data. Such a more realistic dataset should be a better benchmark to reflect the performance of the different methods. + +In this experiment, most of the hyper-parameters and implementation details are the same as our ImageNet setting, except that we do not use the multi-crops strategy for fair comparison. The model will be optimized for $512 \times 1024$ steps as in [58]. We also follow the setting in [58] to conduct two experiments to report the training from scratch performance and training from pre-trained model performance. + +We show the results in Table 5. Our SimMatchV2 achieves $43.05\%$ Top-1 accuracy when training from scratch, which is $+11.84\%$ better than the state-of-the-art method (FixMatch + CCSL). When the MoCo pre-trained model is adopted, our SimMatchV2 achieves a $53.13\%$ Top-1 accuracy, which is $+10.08\%$ better than the from-scratch performance, and $+11.85\%$ better than FixMatch + CCSL. We believe this experiment will further demonstrate the superiority of SimMatchV2. + +# 5. Ablation study + +In this section, we will empirically study our SimMatchV2 based on various conditions and show the effect of each component and hyper-parameter sensitivities of our methods. For all experiments in this section, we adopt the ResNet-50 as our backbone and train the model on ImageNet for 100 epochs on the $1\%$ labeled data. The multicrops are not used by default unless we mentioned. + +Training Dynamics. We first show the training dynamics of SimMatchV2 compared to SimMatch [66] and CoMatch [34]. We visualize the pseudo-labels accuracy, unlabeled data accuracy, and validation accuracy in Figure 3. As we can observe that the pseudo-labels from SimMatchV2 have a higher quality than SimMatch [66], especially in the late stage of the training. Both of these two methods generate better pseudo-labels than CoMatch [34]. We can also notice that SimMatchV2 consistently has better accuracy for the unlabeled data and validation data, which shows the advantages of our method. + +Table 6: Effect of each component in SimMatchV2. N and E denote the node and edge, respectively. N-E means the node-edge consistency and vice versa. Feat norm means feature normalization. + +
N-NN-EE-EE-NFeat NormTop-1Improvement
✓52.9-
✓✓60.0+7.1
✓✓✓60.7+0.7
✓✓✓✓62.2+1.5
✓✓✓✓✓63.7+1.5
+ +Effect of Each Components. In table 6, we show the effect of each design in SimMatchV2. Basically, if only node-node consistency is adopted, the setting is equivalent to the FixMatch [43] with distribution alignment [2], which is our baseline in the first row of the table; Next, we can see that the node-edge, edge-node, and edge-node consistencies give $+7.1\%$ , $+0.7\%$ , $+1.5\%$ improvements, respectively. Finally, the feature normalization also boosts the performance + +![](images/9f24b8eb5de621e2fa8f4ee27bfaa75b8a1771ffa4e06f933ac23b96ea5fa336.jpg) +(a) Pseudo-Labels Accuracy + +![](images/ce6cc88b43487746d9e31ad5e91ec04deff66e3a651cb57ea5fd5aae7c4234ff.jpg) +(b) Unlabeled Data Accuracy +Figure 3: In this Figure, we compare the training dynamics of SimMatchV2 with SimMatch [66], and CoMatch [34] from 3 different perspectives. a) The accuracy of $p^w$ that pass the threshold $\tau$ . b) The accuracy of $p^w$ regardless of the threshold. c) The accuracy on the validation data. All experiments are in 1% ImageNet setting with 100 epochs training. + +![](images/acc214e8c318ffaec44db1c4de232ab6d7ca04e0016e4fffdbda44759864bb44.jpg) +(c) Validation Accuracy + ++1.5%. This experiment should demonstrate that these different designs are all very effective for SSL tasks. + +Table 7: Effect of memory bank size on the performance. $5\mathrm{k}-163\mathrm{K}$ are setting that we use unlabeled bank; labeled bank means we use all labeled bank (All labeled data 12K for $1\%$ ImageNet). + +
Bank Size5K10K40K82K163KLabeled Bank (12K)
Top-163.163.263.563.763.663.2
+ +Node-Edge and Edge-Edge Consistency. We will study the node-edge and edge-edge consistency together since the memory bank size is their primary factor. The experimental results are shown in Table 7. We can observe that when unlabeled data is used, the performance will be slightly increased with a larger memory bank, and the best memory bank size is $82\mathrm{K}$ . We can also notice that we get slightly lower performance with $12\mathrm{K}$ labeled data. Therefore, we decide to use the unlabeled data for node-edge and edge-edge consistency. + +Table 8: Effect of Top-N to Edge-Node Consistency. (We use labeled bank for this experiment.) + +
N8163264128256
Top-163.463.563.563.563.763.6
+ +Table 9: Effect of labeled / unlabeled bank and $\phi$ in Eq. (6) to Edge-Node Consistency. + +
Labeled Y∞Labeled Y1Unlabeled Y∞Unlabeled Y1
Top-163.763.362.462.2
+ +Edge-Node Consistency. As we have mentioned, it is computationally very expensive to involve the whole memory bank for the edge-node consistency calculation. Thus, we only adopt the Top-N nearest neighbor to propagate the label. We show this ablation study in Table 8. The performance is very robust with the selection of the number of Top-N, the best results come from $N = 128$ , and the performance is only slightly decreased -0.3% in the worst case. Note that all experiments in Table 8 are performed with the labeled bank; we also have the ablation study to test the effect of the labeled / unlabeled bank and the $\phi$ in Eq. (6). In Table 9, we can see that the performance with the labeled + +bank is much better than the unlabeled bank. Also, the performance with the fully converged $Y_{\infty}$ is $+0.4\%$ better than $Y_{1}$ , which shows the necessity of our design. + +Table 10: Effect of Feature Normalization + +
Top-1Norm Gap ↓CE Loss ↓
W/o Feat Norm62.21.832.75
W/ Feat Norm63.71.312.63
+ +Feature Normalization. Finally, we show the effect of feature normalization in Table 10. As we mentioned, feature normalization aims to reduce the norm gaps between the weak and strong augmented views. Thus, to validate this, we calculate the mean of the absolute value of $\| h^s\| -\| h^w\|$ across the whole training process. We name it "norm gaps" in Table 10. With the feature normalization, the norm gaps can be reduced from $1.83\rightarrow 1.31$ . We also calculate the mean of the CE loss Eq. (3) (regardless of the threshold, i.e. $\tau = 0$ ) across the training process. + +# 6. Conclusion + +In this work, we propose a new semi-supervised learning algorithm - SimMatchV2, which models consistency regularization from the graph perspective. SimMatchV2 provides four different types of consistency by leveraging the idea of message passing in graph theory and also proposes a simple feature normalization to reduce the gaps between the feature norm of different augmented views. The experiments on different benchmarks and various architectures demonstrate state-of-the-art performance. The limitation of this work is the missing exploration of cooperating with the more advanced SSL training pipelines and stronger data augmentations [5] since it needs more experiments to find the optimal setting. We plan to investigate the pre-training setting in our future work. + +# Acknowledgment + +This work was supported in part by the Australian Research Council under Project DP210101859 and the University of Sydney Research Accelerator (SOAR) Prize. + +# References + +[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. arXiv preprint arXiv:2104.13963, 2021. 1, 6, 7 +[2] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. arXiv preprint arXiv:1911.09785, 2019. 1, 2, 3, 5, 7 +[3] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. arXiv preprint arXiv:1905.02249, 2019. 1, 2, 3, 5, 7 +[4] David Berthelot, Rebecca Roelofs, Kihyuk Sohn, Nicholas Carlini, and Alex Kurakin. Adamatch: A unified approach to semi-supervised learning and domain adaptation. arXiv preprint arXiv:2106.04732, 2021. 5 +[5] Zhaowei Cai, Avinash Ravichandran, Paolo Favaro, Manchen Wang, Davide Modolo, Rahul Bhotika, Zhuowen Tu, and Stefano Soatto. Semi-supervised vision transformers at scale. In NeurIPS, 2022. 3, 7, 8 +[6] Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Charless Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 194-203, 2021. 3, 6 +[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. 2020. 2, 6 +[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. arXiv preprint arXiv:2104.14294, 2021. 6 +[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709, 2020. 2, 3, 6 +[10] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey Hinton. Big self-supervised models are strong semi-supervised learners. arXiv preprint arXiv:2006.10029, 2020. 2, 6 +[11] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2, 6 +[12] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 2, 3 +[13] Ching-Yao Chuang, Joshua Robinson, Lin Yen-Chen, Antonio Torralba, and Stefanie Jegelka. Debiased contrastive learning. In Advances in neural information processing systems, 2020. 2 + +[14] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. volume 15 of Proceedings of Machine Learning Research, pages 215-223, Fort Lauderdale, FL, USA, 11-13 Apr 2011. JMLR Workshop and Conference Proceedings. 5 +[15] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPRO9, 2009. 1, 5 +[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 7 +[17] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 1 +[18] Yue Fan, Anna Kukleva, Dengxin Dai, and Bernt Schiele. Revisiting consistency regularization for semi-supervised learning. International Journal of Computer Vision, pages 1-18, 2022. 5 +[19] Milton Friedman. The use of ranks to avoid the assumption of normality implicit in the analysis of variance. Journal of the american statistical association, 32(200):675-701, 1937. 5 +[20] Milton Friedman. A comparison of alternative tests of significance for the problem of m rankings. The Annals of Mathematical Statistics, 11(1):86-92, 1940. 5 +[21] Yves Grandvalet, Yoshua Bengio, et al. Semi-supervised learning by entropy minimization. CAP, 367:281-296, 2005. 6 +[22] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. arXiv preprint arXiv:2006.07733, 2020. 2, 3, 6 +[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2 +[24] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019. 2, 3 +[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. arXiv preprint arXiv:1512.03385, 2015. 6 +[26] Qianjiang Hu, Xiao Wang, Wei Hu, and Guo-Jun Qi. Adco: Adversarial contrast for efficient learning of unsupervised representations from self-trained negative adversaries. arXiv preprint arXiv:2011.08435, 2020. 2 +[27] Lang Huang, Shan You, Mingkai Zheng, Fei Wang, Chen Qian, and Toshihiko Yamasaki. Green hierarchical vision transformer for masked image modeling. Advances in Neural Information Processing Systems, 35:19997-20010, 2022. 2 + +[28] Lang Huang, Shan You, Mingkai Zheng, Fei Wang, Chen Qian, and Toshihiko Yamasaki. Learning where to learn in cross-view self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14451-14460, 2022. 2 +[29] Yannis Kalantidis, Mert Bulent Sariyildiz, Noe Pion, Philippe Weinzaepfel, and Diane Larlus. Hard negative mixing for contrastive learning. In Neural Information Processing Systems (NeurIPS), 2020. 2 +[30] Soroush Abbasi Koohpayegani, Ajinkya Tejankar, and Hamed Pirsiavash. Mean shift for self-supervised learning. arXiv preprint arXiv:2105.07269, 2021. 2 +[31] A. Krizhevsky. Learning multiple layers of features from tiny images. 2009. 5 +[32] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 2, 5 +[33] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 1, 5 +[34] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regularization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9475-9484, 2021. 3, 5, 6, 7, 8 +[35] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, 2021. 2, 6 +[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1 +[37] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7 +[38] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022. 7 +[39] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016.5,6 +[40] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7 +[41] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979–1993, 2018. 5, 6 +[42] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 5 + +[43] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. arXiv preprint arXiv:2001.07685, 2020. 1, 2, 3, 5, 6, 7 +[44] Jong-Chyi Su and Subhransu Maji. The semi-supervised inaturalist challenge at the fgvc8 workshop, 2021. 7 +[45] Ilya Sutskever, James Martens, George Dahl, and Geoffrey Hinton. On the importance of initialization and momentum in deep learning. In International conference on machine learning, pages 1139-1147. PMLR, 2013. 5 +[46] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. arXiv preprint arXiv:1703.01780, 2017. 2, 5 +[47] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? arXiv preprint arXiv:2005.10243, 2020. 2 +[48] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 7 +[49] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 32-42, 2021. 7 +[50] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. arXiv preprint arXiv:2005.10242, 2020. 2 +[51] Yidong Wang, Hao Chen, Yue Fan, Wang Sun, Ran Tao, Wenxin Hou, Renjie Wang, Linyi Yang, Zhi Zhou, Lan-Zhe Guo, Heli Qi, Zhen Wu, Yu-Feng Li, Satoshi Nakamura, Wei Ye, Marios Savvides, Bhiksha Raj, Takahiro Shinozaki, Bernt Schiele, Jindong Wang, Xing Xie, and Yue Zhang. Usb: A unified semi-supervised learning benchmark for classification. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 5 +[52] Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, and Yu-Gang Jiang. Semi-supervised vision transformers. In ECCV, 2022. 7 +[53] Zhirong Wu, Yuanjun Xiong, X Yu Stella, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2 +[54] Haoyu Xie, Changqi Wang, Mingkai Zheng, Minjing Dong, Shan You, Chong Fu, and Chang Xu. Boosting semisupervised semantic segmentation with probabilistic representations. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2938-2946, 2023. 2 +[55] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information + +Processing Systems, volume 33, pages 6256-6268. Curran Associates, Inc., 2020. 5, 6 +[56] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and improving layer normalization. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 4 +[57] Yi Xu, Lei Shang, Jinxing Ye, Qi Qian, Yu-Feng Li, Baigui Sun, Hao Li, and Rong Jin. Dash: Semi-supervised learning with dynamic thresholding. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 11525-11536. PMLR, 18-24 Jul 2021. 5 +[58] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14421-14430, June 2022. 7 +[59] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. In Advances in Neural Information Processing Systems, 2022. 1 +[60] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 5 +[61] Xiaohua Zhai, A. Oliver, A. Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1476-1485, 2019. 6 +[62] Bowen Zhang, Yidong Wang, Wenxin Hou, HAO WU, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 18408-18419. Curran Associates, Inc., 2021. 5 +[63] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 2 +[64] Yuhang Zhang, Xiaopeng Zhang, Robert Qiu, Jie Li, Haohang Xu, Qi Tian, et al. Semi-supervised contrastive learning with similarity co-calibration. arXiv preprint arXiv:2105.07387, 2021. 6 +[65] Mingkai Zheng, Fei Wang, Shan You, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Weakly supervised contrastive learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10042-10051, October 2021. 6 +[66] Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14471-14481, June 2022. 2, 3, 5, 6, 7, 8 +[67] Mingkai Zheng, Shan You, Fei Wang, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Ressl: Re + +lational self-supervised learning with weak augmentation. arXiv preprint arXiv:2107.09282, 2021. 2 +[68] Mingkai Zheng, Shan You, Fei Wang, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Relational self-supervised learning. arXiv preprint arXiv:2203.08717, 2022. 2, 6 +[69] Dengyong Zhou, Olivier Bousquet, Thomas Lal, Jason Weston, and Bernhard Scholkopf. Learning with local and global consistency. In S. Thrun, L. Saul, and B. Scholkopf, editors, Advances in Neural Information Processing Systems, volume 16. MIT Press, 2003. 4 \ No newline at end of file diff --git a/2023/SimMatchV2_ Semi-Supervised Learning with Graph Consistency/images.zip b/2023/SimMatchV2_ Semi-Supervised Learning with Graph Consistency/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..409592ac3d9a1ce4312b5efbaac3d9c0a3b60f95 --- /dev/null +++ b/2023/SimMatchV2_ Semi-Supervised Learning with Graph Consistency/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a8035715df3917af8e4ee343fb3a83786f37db04e048722549ce2dc1494e2cc +size 503820 diff --git a/2023/SimMatchV2_ Semi-Supervised Learning with Graph Consistency/layout.json b/2023/SimMatchV2_ Semi-Supervised Learning with Graph Consistency/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..09b329c1399290b76bff1c8a03280efd2de4daa4 --- /dev/null +++ b/2023/SimMatchV2_ Semi-Supervised Learning with Graph Consistency/layout.json @@ -0,0 +1,10763 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 92, + 103, + 501, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 103, + 501, + 121 + ], + "spans": [ + { + "bbox": [ + 92, + 103, + 501, + 121 + ], + "type": "text", + "content": "SimMatchV2: Semi-Supervised Learning with Graph Consistency" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "spans": [ + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "text", + "content": "Mingkai Zheng" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "text", + "content": " Shan You" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "text", + "content": " \nLang Huang" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "text", + "content": " Chen Luo" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "text", + "content": " Fei Wang" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "text", + "content": " Chen Qian" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "text", + "content": " Chang Xu" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "text", + "content": "School of Computer Science, Faculty of Engineering, The University of Sydney \n" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "text", + "content": "SenseTime Research " + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "text", + "content": " The University of Tokyo \n" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "text", + "content": "University of Science and Technology of China \n" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 132, + 142, + 473, + 227 + ], + "type": "text", + "content": "State Grid Anhui Electric Power Research Institute" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 280, + 290, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 290, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 290, + 568 + ], + "type": "text", + "content": "Semi-Supervised image classification is one of the most fundamental problem in computer vision, which significantly reduces the need for human labor. In this paper, we introduce a new semi-supervised learning algorithm - SimMatchV2, which formulates various consistency regularizations between labeled and unlabeled data from the graph perspective. In SimMatchV2, we regard the augmented view of a sample as a node, which consists of a label and its corresponding representation. Different nodes are connected with the edges, which are measured by the similarity of the node representations. Inspired by the message passing and node classification in graph theory, we propose four types of consistencies, namely 1) node-node consistency, 2) node-edge consistency, 3) edge-edge consistency, and 4) edge-node consistency. We also uncover that a simple feature normalization can reduce the gaps of the feature norm between different augmented views, significantly improving the performance of SimMatchV2. Our SimMatchV2 has been validated on multiple semi-supervised learning benchmarks. Notably, with ResNet-50 as our backbone and 300 epochs of training, SimMatchV2 achieves " + }, + { + "bbox": [ + 46, + 280, + 290, + 568 + ], + "type": "inline_equation", + "content": "71.9\\%" + }, + { + "bbox": [ + 46, + 280, + 290, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 280, + 290, + 568 + ], + "type": "inline_equation", + "content": "76.2\\%" + }, + { + "bbox": [ + 46, + 280, + 290, + 568 + ], + "type": "text", + "content": " Top-1 Accuracy with " + }, + { + "bbox": [ + 46, + 280, + 290, + 568 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 280, + 290, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 280, + 290, + 568 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 280, + 290, + 568 + ], + "type": "text", + "content": " labeled examples on ImageNet, which significantly outperforms the previous methods and achieves state-of-the-art performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 590, + 128, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 128, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 128, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 611, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 611, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 611, + 287, + 696 + ], + "type": "text", + "content": "In the past decades, deep learning has demonstrated its superior performance in various tasks [15, 36, 17, 59]. However, its outstanding performance is usually based on the fully-supervised setting, which requires a large amount of data and annotations. In a real-world scenario, high-quality data annotations are fairly hard to collect since it is generally expensive, time-consuming, and sometimes" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 256, + 544, + 423 + ], + "blocks": [ + { + "bbox": [ + 309, + 256, + 544, + 423 + ], + "lines": [ + { + "bbox": [ + 309, + 256, + 544, + 423 + ], + "spans": [ + { + "bbox": [ + 309, + 256, + 544, + 423 + ], + "type": "image", + "image_path": "453fecbdc9a8f91f96af2826d42b2b7def9719b7d570adfa980eabb118ec480b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 427, + 547, + 478 + ], + "lines": [ + { + "bbox": [ + 305, + 427, + 547, + 478 + ], + "spans": [ + { + "bbox": [ + 305, + 427, + 547, + 478 + ], + "type": "text", + "content": "Figure 1: The Top-1 Accuracy of ResNet-50 on ImagetNet with " + }, + { + "bbox": [ + 305, + 427, + 547, + 478 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 305, + 427, + 547, + 478 + ], + "type": "text", + "content": " labeled data. Our proposed SimMatchV2 performs significantly better than previous methods with fewer training epochs. Most importantly, 100 epochs of SimMatchV2 training only takes 14.7 GPU days, but the performance is much higher than PAWS [1] with 300 epochs (68 GPU days)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 485, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 544 + ], + "type": "text", + "content": "require expert knowledge ( e.g. medical images). Semi-supervised learning (SSL) has recently drawn much attention to the research community due to its efficiency in dealing with few labeled data and leveraging massive unlabeled data, significantly reducing the need for human labor." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": "For the typical SSL algorithms, one standard paradigm is to train a semantic classifier on the labeled data, then use the classifier to make a prediction on the unlabeled data as the pseudo-label [33]. Besides, a class-level \"consistency regularization\" will be applied to enforce consistency between the class predictions of a model on different perturbations of the same input, which encourages the model to learn more robust and invariant features that generalize well to unseen data. Thus, the quality of pseudo-labels plays a crucial role in consistency regularization, and many works focus on generating stable and high-quality pseudo-labels. For example, MixMatch [3], ReMixMatch [2], and FixMatch [43] are three typical \"consistency regularization\" based methods that significantly improve the performance of the SSL" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 277, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 277, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 277, + 713 + ], + "type": "text", + "content": "*Correspondence to: Shan You ." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16432" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "tasks. However, these methods highly depend on the classifier trained on the labeled data. A limited number of labels will result in a poor classifier that generates inaccurate pseudo-labels and leads to bad performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 372 + ], + "type": "text", + "content": "With the recent success of unsupervised visual representation learning [9, 24, 10, 11, 22, 7, 12, 27, 28, 23], the pre-training and fine-tuning paradigm has gained popularity for the semi-supervised learning. In general, the model will be first trained on a large amount of unlabeled data with a proxy task (e.g. instance discrimination), which also imposes the consistency regularization on the instance level (i.e. forcing the model to give similar representation for the different perturbations of the same instance). Then, a classifier will be attached to the pre-trained model, and the whole model will be fine-tuned with the limited labeled data. Many works [29, 35, 26, 67, 68, 13, 50, 47, 30, 54] have been proposed to design a better proxy task to improve the quality of the representation. Apparently, the biggest advantage of this approach is that it does not depend on the supervised trained classifier but still yields high-quality representations for unlabeled data. Nevertheless, the drawback of this approach is also very obvious, since no labels were used during the pre-training, making the model hard to converge. Generally, the pre-training requires more than 800 epochs to converge, which is extremely time-consuming." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 373, + 288, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 288, + 671 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 288, + 671 + ], + "type": "text", + "content": "Although many different types of consistency regularizations have been proposed, including class-level [3, 2, 43], instance-level [9, 24, 22], and those that consider both [66, 43], how to establish a complete and effective consistency regularizations remains unresolved. This paper introduces a new SSL algorithm - SimMatchV2, which provides a complete consistency regularization solution for various scenarios. The overall idea is inspired by the message passing and node classification in the graph theory. In SimMatchV2, we consider the augmented view of a sample as a node in the graph which consists of a node label and its corresponding representation. Different nodes are connected with the edges, which are measured by the similarity of the node representations. Thus, the information of different nodes can be efficiently propagated through the graph. To this end, we propose four types of consistency regularization, namely 1) node-node consistency, 2) node-edge consistency, 3) edge-edge consistency, and 4) edge-node consistency. Note that the 1) node-node consistency has been proposed in previous works, but our SimMatchV2 provides a complete and unified solution to design the consistencies. Moreover, we also propose a feature normalization technique to reduce the gaps between the feature norm between different augmented views and greatly improve the performance. Our contributions can be summarized as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 677, + 287, + 713 + ], + "type": "text", + "content": "- We propose SimMatchV2 - a novel semi-supervised algorithm that models various consistency regularizations from the graph perspectives and proposes four" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 324, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 545, + 96 + ], + "type": "text", + "content": "types of consistencies to provide a complete solution for the regularization in SSL." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 100, + 545, + 213 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 317, + 100, + 545, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 100, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 317, + 100, + 545, + 137 + ], + "type": "text", + "content": "- We show that a simple feature normalization reduces the gaps between the feature norm of different augmented views and greatly improves the performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 140, + 545, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 545, + 213 + ], + "type": "text", + "content": "- SimMatchV2 sets a new state-of-the-art performance on various SSL benchmarks. Notably, with ResNet50 as our backbone and 300 epochs of training, SimMatchV2 achieves " + }, + { + "bbox": [ + 317, + 140, + 545, + 213 + ], + "type": "inline_equation", + "content": "71.9\\%" + }, + { + "bbox": [ + 317, + 140, + 545, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 317, + 140, + 545, + 213 + ], + "type": "inline_equation", + "content": "76.2\\%" + }, + { + "bbox": [ + 317, + 140, + 545, + 213 + ], + "type": "text", + "content": " Top-1 accuracy with " + }, + { + "bbox": [ + 317, + 140, + 545, + 213 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 317, + 140, + 545, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 317, + 140, + 545, + 213 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 317, + 140, + 545, + 213 + ], + "type": "text", + "content": " labeled data on ImageNet. This result is " + }, + { + "bbox": [ + 317, + 140, + 545, + 213 + ], + "type": "inline_equation", + "content": "+4.7\\%" + }, + { + "bbox": [ + 317, + 140, + 545, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 317, + 140, + 545, + 213 + ], + "type": "inline_equation", + "content": "+0.7\\%" + }, + { + "bbox": [ + 317, + 140, + 545, + 213 + ], + "type": "text", + "content": " better than prior arts." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 225, + 390, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 225, + 390, + 238 + ], + "spans": [ + { + "bbox": [ + 306, + 225, + 390, + 238 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 246, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 545, + 569 + ], + "type": "text", + "content": "Consistency Regularization with Pseudo-Label. Pseudo-label is a method that aims to generate labels for unlabeled data. As has been mentioned, it is often used in conjunction with consistency regularization, which constrains the model from making similar predictions on different perturbations of the same input. Most of these methods focus on generating high-quality pseudo-labels and exploring better perturbation strategies. For example, II-model [32] propose a temporal ensemble strategy for the pseudo-label to reduce the noise in the target. MeanTeacher [46] instead, maintain an exponential moving average (EMA) network to generate more reliable and stable pseudo-labels. Later, MixMatch [3] was proposed and greatly improved the SSL performance. Specifically, it adopts the averaged prediction of multiple strongly augmented views of the same image as the pseudo-label, then incorporates with the Mixup [63] augmentation that generates a convex combination of pairs of images and uses the corresponding mixed pseudo-label as the supervision target. ReMixMatch [2] improves MixMatch by using only a weakly augmented view to generate the pseudo-label and also introduces the distribution alignment to encourage the marginal distribution of predictions on unlabeled data to be consistent with the labeled data. FixMatch [43] abandons these complicated strategies and proposes a confidence-based threshold method to filter out the low-confidence pseudo-labels during the training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Self-Supervised Learning for SSL. Self-supervised learning has demonstrated its outstanding performance in learning visual representations without labels. It is intuitive and natural to leverage the proxy tasks from self-supervised learning to help with SSL tasks. As mentioned, the most straightforward solution is to pre-train the model in a self-supervised manner on a large amount of unlabeled data and then fine-tune the model with the limited labeled data. In this way, the design of the self-supervised task is particularly important. For example, instance discrimination [53] is one of the most popular proxy tasks, which aims to encourage different augmented views of the same image to" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16433" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 99, + 71, + 206, + 178 + ], + "blocks": [ + { + "bbox": [ + 99, + 71, + 206, + 178 + ], + "lines": [ + { + "bbox": [ + 99, + 71, + 206, + 178 + ], + "spans": [ + { + "bbox": [ + 99, + 71, + 206, + 178 + ], + "type": "image", + "image_path": "7feb6baedfb99e70b404cc269aa65c21deb7b2be80231cd5f771684e470a29fc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 179, + 191, + 189 + ], + "lines": [ + { + "bbox": [ + 119, + 179, + 191, + 189 + ], + "spans": [ + { + "bbox": [ + 119, + 179, + 191, + 189 + ], + "type": "text", + "content": "(a) Node-Edge Consistency" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 70, + 373, + 175 + ], + "blocks": [ + { + "bbox": [ + 220, + 70, + 373, + 175 + ], + "lines": [ + { + "bbox": [ + 220, + 70, + 373, + 175 + ], + "spans": [ + { + "bbox": [ + 220, + 70, + 373, + 175 + ], + "type": "image", + "image_path": "3a5c414e81235605c68b93b9bed1b65faf0beb7ac894e0401a41738429c8868f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 179, + 333, + 189 + ], + "lines": [ + { + "bbox": [ + 261, + 179, + 333, + 189 + ], + "spans": [ + { + "bbox": [ + 261, + 179, + 333, + 189 + ], + "type": "text", + "content": "(b) Edge-Edge Consistency" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 191, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 46, + 191, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 191, + 547, + 213 + ], + "type": "text", + "content": "Figure 2: Illustration of various consistency in SimMatchV2. The green and blue nodes are from the weakly and strongly augmented view of the unlabeled data, the red nodes are the other samples." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 384, + 69, + 494, + 173 + ], + "blocks": [ + { + "bbox": [ + 384, + 69, + 494, + 173 + ], + "lines": [ + { + "bbox": [ + 384, + 69, + 494, + 173 + ], + "spans": [ + { + "bbox": [ + 384, + 69, + 494, + 173 + ], + "type": "image", + "image_path": "77973cd795f807da3d6f12147c88fb6ef683e0988013a0217b41324e101c206e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 179, + 474, + 189 + ], + "lines": [ + { + "bbox": [ + 403, + 179, + 474, + 189 + ], + "spans": [ + { + "bbox": [ + 403, + 179, + 474, + 189 + ], + "type": "text", + "content": "(c) Edge-Node Consistency" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 219, + 290, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 290, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 290, + 374 + ], + "type": "text", + "content": "be pulled closer on embedding space but pushes apart all the other images away. BYOL [22] and SimSiam [12] omit the negatives in contrastive learning and leverage an asymmetric structure to avoid representation collapse. Recently, some works (e.g. CoMatch [34] and SimMatch [66]) have been proposed to jointly leverage both class-level and instance-level consistencies and greatly improve the SSL performance, especially in the case when only very limited labeled data is available. There are also many works [6, 5] exploring the two-stage framework (i.e. pre-training followed by fine-tuning with SSL algorithms like FixMatch), which shows the importance of the self-supervised pretrained to the SSL tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 404, + 129, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 404, + 129, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 404, + 129, + 418 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 430, + 171, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 171, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 171, + 443 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "text", + "content": "Similar to [34, 66], we define a batch of labeled data as " + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{x_b : b \\in (1, \\dots, B)\\}" + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "text", + "content": ", and a batch of unlabeled data as " + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{U} = \\{u_b : b \\in (1, \\dots, \\mu B)\\}" + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "text", + "content": " is the coefficient of unlabeled batch size. Then, a neural network encoder will extract the feature " + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "text", + "content": " from the augmented images. We use " + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "inline_equation", + "content": "h_b^w" + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "inline_equation", + "content": "h_b^s" + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "text", + "content": " to denote the feature extracted from the weakly and strongly augmented images, respectively. Then, we use " + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "text", + "content": " to represent a linear layer followed by a SoftMax function that transforms the feature to a class probability i.e. " + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "inline_equation", + "content": "p = \\mathrm{SoftMax}(W^T h)" + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "text", + "content": ". Finally, we adopt the " + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "text", + "content": " to denote a two-layer non-linear projection head which maps the feature to a low-dimensional representation vector i.e. " + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "inline_equation", + "content": "z = g(h)" + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "text", + "content": ". By following the common setting as in [9, 24, 66, 34], the representation " + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "text", + "content": " will be normalized to a unit vector by default. Note that the reason we use " + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "text", + "content": " as the representation instead of " + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "text", + "content": " is that we need to store a large number of samples as in [24]. The dimension of " + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "text", + "content": " is much smaller than " + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "text", + "content": " which makes the training more efficient. For example, " + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "text", + "content": " is 2048-D and " + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 455, + 287, + 684 + ], + "type": "text", + "content": " is 128-D for ResNet-50." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "For labeled data, we perform standard supervised learning with the cross-entropy loss, which can be represented" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 219, + 524, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 219, + 524, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 219, + 524, + 232 + ], + "type": "text", + "content": "by Eq. (1), where " + }, + { + "bbox": [ + 305, + 219, + 524, + 232 + ], + "type": "inline_equation", + "content": "y_{b}" + }, + { + "bbox": [ + 305, + 219, + 524, + 232 + ], + "type": "text", + "content": " is the one-hot ground truth label," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 372, + 235, + 545, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 235, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 372, + 235, + 545, + 269 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = - \\frac {1}{B} \\sum_ {b = 1} ^ {B} y _ {b} \\log p _ {b} ^ {w}. \\tag {1}", + "image_path": "1a65561869ea7160b200e87968837af1b322f5cc70caf8cffd6845b0c8521bd0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 272, + 515, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 272, + 515, + 285 + ], + "spans": [ + { + "bbox": [ + 305, + 272, + 515, + 285 + ], + "type": "text", + "content": "3.2. Consistency Regularization with Graph" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 290, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 374 + ], + "type": "text", + "content": "SimMatchV2 considers the augmented view of a sample as a node in the graph, which consists of a node label " + }, + { + "bbox": [ + 304, + 290, + 545, + 374 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 290, + 545, + 374 + ], + "type": "text", + "content": " (class prediction for unlabeled data and ground truth for labeled data) and the representation " + }, + { + "bbox": [ + 304, + 290, + 545, + 374 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 290, + 545, + 374 + ], + "type": "text", + "content": ". Different nodes are connected by edges measured by the similarity of the representations. We use strong/weak node to name the node from strongly/weakly augmented views." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "spans": [ + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "text", + "content": "Given a graph with " + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "text", + "content": " nodes, we have the node labels " + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "inline_equation", + "content": "\\{p_k : k \\in (1, \\dots, K)\\}" + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "text", + "content": " and the corresponding representations " + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "inline_equation", + "content": "\\{z_k : k \\in (1, \\dots, K)\\}" + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "text", + "content": ". In our implementation, the graph nodes will be stored in a memory bank as in [24, 66, 34]. We can store both labeled and unlabeled nodes. The node label will be the ground truth for labeled data and " + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "text", + "content": " for unlabeled data. Suppose we have an unlabeled weak node with the representation " + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "inline_equation", + "content": "z_b^w" + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "inline_equation", + "content": "e_{b,i}^w" + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "text", + "content": " to denote the edge between the " + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "inline_equation", + "content": "z_b^w" + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "text", + "content": " node on the graph, which can be expressed by Eq. (2). " + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "text", + "content": " is a temperature parameter that controls the sharpness of the similarity distribution. We also use " + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "inline_equation", + "content": "e_{b,i}^w" + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "text", + "content": " to denote edges between " + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "inline_equation", + "content": "z_b^w" + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "text", + "content": " and all the nodes on the graph (i.e. " + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "inline_equation", + "content": "e_{b,i}^w" + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "text", + "content": " is a scalar and " + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "inline_equation", + "content": "e_b^w \\in R^{1 \\times K}" + }, + { + "bbox": [ + 305, + 374, + 546, + 534 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 365, + 538, + 545, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 538, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 365, + 538, + 545, + 567 + ], + "type": "interline_equation", + "content": "e _ {b, i} ^ {w} = \\frac {\\exp \\left(z _ {b} ^ {w} \\cdot z _ {i} / t\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(z _ {b} ^ {w} \\cdot z _ {k} / t\\right)} \\tag {2}", + "image_path": "ab4e3fcb26f28629a45c0ba37ec8ead18f78ef95baa6469226fe45311c56eb5c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 570, + 547, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 678 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 678 + ], + "type": "text", + "content": "Node-Node Consistency. (Weak Node vs. Strong Node.) The node-node consistency aims to maintain the consistency of the label between the strong node and weak node of the same instance. Such consistency has been proposed in many previous works [3, 2, 43]. In this work, we have renamed it and integrated it into our graph-based framework. Concretely, we follow the design in previous works; given unlabeled data, the node-node consistency can be expressed by Eq. (3), where " + }, + { + "bbox": [ + 304, + 570, + 547, + 678 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 570, + 547, + 678 + ], + "type": "text", + "content": " is the confidence threshold." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 334, + 682, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 682, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 334, + 682, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {n n} = - \\frac {1}{\\mu B} \\sum_ {b = 1} ^ {\\mu B} \\mathbb {1} \\left(\\max p _ {b} ^ {w} > \\tau\\right) p _ {b} ^ {w} \\log p _ {b} ^ {s}. \\tag {3}", + "image_path": "1a9b4511150735a7a01690624b657e73e830c31c2f6309e2996fd94720458ba6.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16434" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 182 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 182 + ], + "type": "text", + "content": "Node-Edge Consistency. (Weak Node vs. Strong Edge.) The purpose of the node-edge consistency is to construct an edge " + }, + { + "bbox": [ + 47, + 72, + 289, + 182 + ], + "type": "inline_equation", + "content": "e_b^s" + }, + { + "bbox": [ + 47, + 72, + 289, + 182 + ], + "type": "text", + "content": " by using the representation " + }, + { + "bbox": [ + 47, + 72, + 289, + 182 + ], + "type": "inline_equation", + "content": "z_b^s" + }, + { + "bbox": [ + 47, + 72, + 289, + 182 + ], + "type": "text", + "content": " from the strong node, then use the edge " + }, + { + "bbox": [ + 47, + 72, + 289, + 182 + ], + "type": "inline_equation", + "content": "e_b^s" + }, + { + "bbox": [ + 47, + 72, + 289, + 182 + ], + "type": "text", + "content": " as the weight to aggregate the node label from the graph. We expect the aggregated node label can be consistent with the label " + }, + { + "bbox": [ + 47, + 72, + 289, + 182 + ], + "type": "inline_equation", + "content": "p_b^w" + }, + { + "bbox": [ + 47, + 72, + 289, + 182 + ], + "type": "text", + "content": " from the weak node. Finally, the node-edge consistency will be constrained by the cross-entropy loss between " + }, + { + "bbox": [ + 47, + 72, + 289, + 182 + ], + "type": "inline_equation", + "content": "p_b^w" + }, + { + "bbox": [ + 47, + 72, + 289, + 182 + ], + "type": "text", + "content": " and the aggregated label, which can be expressed by Eq. (4)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 197, + 287, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 197, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 91, + 197, + 287, + 229 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {n e} = - \\frac {1}{\\mu B} \\sum_ {b = 1} ^ {\\mu B} p _ {b} ^ {w} \\log \\left(\\sum_ {i = 1} ^ {K} p _ {i} e _ {b, i} ^ {s}\\right) \\tag {4}", + "image_path": "8989a2b78a55648da5fa8cffa8f639e920693bbb2cd6f48f3126c347aa6830da.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 247, + 287, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 247, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 287, + 286 + ], + "type": "text", + "content": "Note that the aggregated label " + }, + { + "bbox": [ + 47, + 247, + 287, + 286 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{K} p_i e_{b,i}^s" + }, + { + "bbox": [ + 47, + 247, + 287, + 286 + ], + "type": "text", + "content": " does not include the node label " + }, + { + "bbox": [ + 47, + 247, + 287, + 286 + ], + "type": "inline_equation", + "content": "p_b^s" + }, + { + "bbox": [ + 47, + 247, + 287, + 286 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 247, + 287, + 286 + ], + "type": "inline_equation", + "content": "p_b^w" + }, + { + "bbox": [ + 47, + 247, + 287, + 286 + ], + "type": "text", + "content": " from itself. The gradient will only be applied on " + }, + { + "bbox": [ + 47, + 247, + 287, + 286 + ], + "type": "inline_equation", + "content": "e_{b,i}^s" + }, + { + "bbox": [ + 47, + 247, + 287, + 286 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 288, + 289, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 289, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 289, + 408 + ], + "type": "text", + "content": "Edge-Edge Consistency. (Weak Edge vs. Strong Edge.) The edge-edge consistency can be regarded as a simple extension of the node-node consistency. Basically, the node-node consistency aims to let the model predict invariant labels for the weak and strong nodes of the same instances. Similarly, the edges should also be invariant with respect to the strong and weak nodes of the same instance. Thus, the edge-edge consistency can be formulated as the cross-entropy loss between the similarity distribution from different augmented views, which can be written as Eq. (5)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 424, + 287, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 424, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 99, + 424, + 287, + 458 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {e e} = - \\frac {1}{\\mu B} \\sum_ {b = 1} ^ {\\mu B} \\sum_ {i = 1} ^ {K} e _ {b, i} ^ {w} \\log e _ {b, i} ^ {s} \\tag {5}", + "image_path": "d3a00a10dfaf7fcea497e9314dd2452b2b40912901dc2febf580ded24dab7670.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "content": "Edge-Node Consistency. (Weak Edge vs. Strong Node.) This is somewhat similar to reversed operation to the node-edge consistency. Concretely, node-edge consistency requires the aggregated label " + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{K} p_i e_{b,i}^s" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "content": " from the strong node to be consistent with the label " + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "inline_equation", + "content": "p_b^w" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "content": " from the weak node. In contrast, edge-node consistency requires the label " + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "inline_equation", + "content": "p_b^s" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "content": " from the strong node to be consistent with the aggregated label " + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{K} p_i e_{b,i}^w" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "content": " from the weak node. Given an unlabeled weak node which has " + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "inline_equation", + "content": "z_b^w" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "inline_equation", + "content": "p_b^w" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "content": ". Then, we concatenate the unlabeled node with the graph to get " + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "inline_equation", + "content": "Z = [z_b^w, z_1, \\dots, z_k]" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "inline_equation", + "content": "Y = [p_b^w, p_1, \\dots, p_k]" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "content": ". Next, we use the representation matrix " + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "content": " to obtain the pairwise edge matrix " + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "content": " (a.k.a. affinity matrix) in the same way as Eq. (2). Note that the difference between " + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "inline_equation", + "content": "e_b" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "content": " is that " + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "inline_equation", + "content": "e_b" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "content": " only calculates the edges between the unlabeled data with respect to the node in the graph (" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "inline_equation", + "content": "e_b^w \\in R^{1 \\times K}" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "content": ") but " + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "content": " also contains the edge information within the graphs. (" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "inline_equation", + "content": "A \\in R^{(1 + K) \\times (1 + K)}" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "content": "). We also want to emphasize that we will make " + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "content": " a hollow matrix (i.e. " + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "inline_equation", + "content": "A_{i,j} = 0" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "inline_equation", + "content": "i == j" + }, + { + "bbox": [ + 47, + 472, + 289, + 715 + ], + "type": "text", + "content": ") before normalizing it to avoid the self-loop in the graph. Here, we adopt the transductive" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 432, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 432, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 432, + 85 + ], + "type": "text", + "content": "inference [69] for aggregation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 331, + 89, + 545, + 138 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 89, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 331, + 89, + 545, + 138 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} Y _ {\\phi} = \\alpha \\cdot A Y _ {\\phi - 1} + (1 - \\alpha) \\cdot Y _ {0} \\\\ = (\\alpha \\cdot A) ^ {\\phi - 1} Y _ {0} + (1 - \\alpha) \\sum_ {i = 0} ^ {\\phi - 1} (\\alpha \\cdot A) ^ {i} Y _ {0}, \\tag {6} \\\\ \\end{array}", + "image_path": "3fab6f3dfda6c8da8a725e9b91c25980a1859d5e3688109837f0de97c6f92de6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "text", + "content": " is the number of iterations that the label has been propagated, " + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "inline_equation", + "content": "Y_{0}" + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "text", + "content": " is the initial state " + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "inline_equation", + "content": "[p_b^w,p_1,\\dots,p_k]" + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "text", + "content": " is a parameter in (0,1) controlling the weight of the aggregated label and the initial state " + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "inline_equation", + "content": "Y_{0}" + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "text", + "content": ". The node label will be propagated multiple times until it fully converges. When " + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "inline_equation", + "content": "\\phi \\rightarrow \\infty" + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "inline_equation", + "content": "(\\alpha \\cdot A)^{\\phi -1} = 0" + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "text", + "content": " since " + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "inline_equation", + "content": "0 < \\alpha < 1" + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "text", + "content": ". We can also make an approximation such that " + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "inline_equation", + "content": "\\lim_{\\phi \\to \\infty}\\sum_{i = 0}^{\\phi -1}(\\alpha \\cdot A)^{i} = (I - \\alpha \\cdot A)^{-1}" + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "text", + "content": " is the identity matrix. Thus, the converged " + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "inline_equation", + "content": "Y_{\\infty}" + }, + { + "bbox": [ + 305, + 143, + 547, + 252 + ], + "type": "text", + "content": " can be derived as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 361, + 257, + 545, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 257, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 361, + 257, + 545, + 271 + ], + "type": "interline_equation", + "content": "Y _ {\\infty} = (1 - \\alpha) (I - \\alpha \\cdot A) ^ {- 1} Y _ {0} \\tag {7}", + "image_path": "072118716f66a8e0d0e9770a1b38cbc424a295523ce5e92584361a9066d6831e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "type": "text", + "content": "Finally, the converged pseudo-label of the unlabeled data can be obtained from " + }, + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "type": "inline_equation", + "content": "Y_{\\infty}[0]" + }, + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "type": "text", + "content": " (i.e. the first row vector of " + }, + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "type": "inline_equation", + "content": "Y_{\\infty}" + }, + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "type": "text", + "content": "). Note that if we take " + }, + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "type": "inline_equation", + "content": "Y_{1}[0]" + }, + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "type": "text", + "content": " is identical to " + }, + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{K} p_{i} e_{b,i}^{w}" + }, + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "type": "inline_equation", + "content": "Y_{\\infty}[0]" + }, + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "type": "text", + "content": " contains both information from the node label and aggregated label. Thus, we can directly use it to replace the original " + }, + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "type": "inline_equation", + "content": "p_{b}^{w}" + }, + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "type": "text", + "content": " in Eq. (3). Moreover, we should notice an inverse operator in Eq. (7), which is computationally very expensive for a large matrix. Thus, to reduce the complexity, we will only take the Top-N nearest node from " + }, + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 305, + 275, + 547, + 397 + ], + "type": "text", + "content": " instead of using the whole graph." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 397, + 547, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 397, + 547, + 554 + ], + "spans": [ + { + "bbox": [ + 305, + 397, + 547, + 554 + ], + "type": "text", + "content": "Feature Normalization. As we have presented in Eq. (3), one of the training objectives is to minimize the predictions between the weakly and strongly augmented views. However, significant gaps exist between the feature norms of different augmented views. Also, from " + }, + { + "bbox": [ + 305, + 397, + 547, + 554 + ], + "type": "inline_equation", + "content": "p = \\operatorname{SoftMax}(W^T h)" + }, + { + "bbox": [ + 305, + 397, + 547, + 554 + ], + "type": "text", + "content": ", we can observe that the norm of " + }, + { + "bbox": [ + 305, + 397, + 547, + 554 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 305, + 397, + 547, + 554 + ], + "type": "text", + "content": " does not affect the prediction (i.e. " + }, + { + "bbox": [ + 305, + 397, + 547, + 554 + ], + "type": "inline_equation", + "content": "\\operatorname{argmax}(p)" + }, + { + "bbox": [ + 305, + 397, + 547, + 554 + ], + "type": "text", + "content": "). Thus, minimizing the gaps between the feature norm is unnecessary and brings additional complexity to the training. To address this issue, we simply apply a layer normalization [56] on the feature " + }, + { + "bbox": [ + 305, + 397, + 547, + 554 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 305, + 397, + 547, + 554 + ], + "type": "text", + "content": ", which can be expressed by Eq. (8). Finally, the linear layer " + }, + { + "bbox": [ + 305, + 397, + 547, + 554 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 305, + 397, + 547, + 554 + ], + "type": "text", + "content": " and non-linear projection head " + }, + { + "bbox": [ + 305, + 397, + 547, + 554 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 305, + 397, + 547, + 554 + ], + "type": "text", + "content": " will be applied on top of the " + }, + { + "bbox": [ + 305, + 397, + 547, + 554 + ], + "type": "inline_equation", + "content": "\\widehat{h}" + }, + { + "bbox": [ + 305, + 397, + 547, + 554 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 382, + 559, + 545, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 559, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 382, + 559, + 545, + 574 + ], + "type": "interline_equation", + "content": "\\widehat {h} = \\operatorname {L a y e r N o r m} (h). \\tag {8}", + "image_path": "e3b5ea81d5ec942b4512ede6c69c30bf6a3f0efa981e4d6c90c2b7ce654ffea6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 578, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 578, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 305, + 578, + 547, + 616 + ], + "type": "text", + "content": "Finally, the overall objective of SimMatchV2 can be expressed by Eq. (9), where " + }, + { + "bbox": [ + 305, + 578, + 547, + 616 + ], + "type": "inline_equation", + "content": "\\lambda_{nn},\\lambda_{ee}" + }, + { + "bbox": [ + 305, + 578, + 547, + 616 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 305, + 578, + 547, + 616 + ], + "type": "inline_equation", + "content": "\\lambda_{ne}" + }, + { + "bbox": [ + 305, + 578, + 547, + 616 + ], + "type": "text", + "content": " are the balancing factors that control the weights of the different losses." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 343, + 626, + 545, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 626, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 343, + 626, + 545, + 639 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {s} + \\lambda_ {n n} \\mathcal {L} _ {n n} + \\lambda_ {n e} \\mathcal {L} _ {n e} + \\lambda_ {e e} \\mathcal {L} _ {e e} \\tag {9}", + "image_path": "4d7315ed6e13b8d345e9e29469d78815c225a76d205103640a3cf17ed3508361.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 651, + 388, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 388, + 664 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 388, + 664 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 671, + 450, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 450, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 450, + 682 + ], + "type": "text", + "content": "4.1. Classical SSL Benchmark" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "We will first conduct our experiments on the classic SSL benchmark which is widely adopted by most SSL papers." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16435" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 102, + 545, + 226 + ], + "blocks": [ + { + "bbox": [ + 47, + 72, + 545, + 101 + ], + "lines": [ + { + "bbox": [ + 47, + 72, + 545, + 101 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 545, + 101 + ], + "type": "text", + "content": "Table 1: Error rate (%) and Rank on the classical SSL benchmark. The error rate and standard deviation are reported based on three runs. The experimental settings in this table are all based on the USB [51] code base. All results of other methods are directly copied from the USB [51] benchmark paper. The results for fully-supervised training on STL-10 is not provided since it contains an unlabeled set with unknown labels." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 102, + 545, + 226 + ], + "lines": [ + { + "bbox": [ + 49, + 102, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 49, + 102, + 545, + 226 + ], + "type": "table", + "html": "
DatasetCIFAR-10CIFAR-100SVHNSTL-10Friedman rankFinal rankMean error rate
# Label402504000400250010000402501000402501000---
Fully-Supervised Supervised4.57±0.0618.96±0.062.14±0.01----
77.18±1.3256.24±3.4116.10±0.3289.60±0.4358.33±1.4136.83±0.2182.68±1.9124.17±1.6512.19±0.2375.4±0.6655.07±1.8335.42±0.48---
II-Model [32]76.35±1.6948.73±1.0713.63±0.0787.67±0.7956.40±0.6936.73±0.0580.07±1.2213.46±0.616.90±0.2274.89±0.5752.20±2.1131.34±0.6414.171548.20
Pseudo-Labelling [33]75.95±1.8651.12±2.9115.32±0.3588.18±0.8955.37±0.4836.58±0.1275.98±5.3616.47±0.599.37±0.4274.02±0.4751.90±1.8730.77±0.0414.081448.42
Mean Teacher [46]72.42±2.1037.56±4.908.29±0.1079.96±0.5344.37±0.6031.39±0.1149.34±7.903.44±0.023.28±0.0772.90±0.8349.30±2.0927.92±1.6511.751240.01
VAT [41]78.58±2.7828.87±3.6210.90±0.1683.60±4.2146.20±0.8032.14±0.3184.12±3.183.38±0.122.87±0.1873.33±0.4757.78±1.4740.98±0.9612.921345.23
MixMatch [3]35.18±3.8713.00±0.806.55±0.0564.91±3.3439.29±0.1327.74±0.2727.77±5.434.69±0.463.85±0.2849.84±0.5832.05±1.1620.17±0.6711.001127.09
ReMixMatch [2]8.13±0.586.34±0.224.65±0.0941.60±1.4825.72±0.0720.04±0.1316.43±13.775.65±0.355.36±0.5827.87±3.8511.14±0.526.44±0.157.751014.95
UDA [5]10.01±3.345.23±0.084.36±0.0945.48±0.3727.51±0.2823.12±0.455.28±0.421.93±0.031.94±0.0240.09±4.0310.11±1.156.23±0.286.42815.12
FixMatch [3]12.66±4.494.95±0.104.26±0.0145.38±2.0727.71±0.4222.06±0.103.37±1.011.97±0.012.02±0.0338.19±4.768.64±0.845.82±0.065.29514.75
Dash [57]9.29±3.285.10±0.284.36±0.1047.49±1.0527.47±0.3821.89±0.165.26±2.022.01±0.012.08±0.0942.00±4.9410.50±1.376.30±0.496.83915.32
CoMatch [34]6.51±1.185.35±0.144.27±0.1253.41±2.3629.78±0.1122.11±0.228.20±3.522.16±0.042.01±0.0413.74±4.207.63±0.945.71±0.085.75613.41
CRMatch [18]13.62±2.624.61±0.173.65±0.0437.76±1.4524.13±0.1619.89±0.232.60±0.771.98±0.041.95±0.0333.73±1.1714.87±5.096.53±0.364.08213.78
FlexMatch [4]5.29±0.294.97±0.074.24±0.0640.73±1.4426.17±0.1821.75±0.155.42±2.838.74±3.327.90±0.3029.12±5.049.85±1.356.08±0.345.92714.19
AdaMatch [4]5.09±0.215.13±0.054.36±0.0537.08±1.3526.66±0.3321.99±0.156.14±4.352.13±0.042.02±0.0519.95±5.178.59±0.436.01±0.024.79312.18
SimMatch [66]5.38±0.015.36±0.084.41±0.0739.32±0.7226.21±0.3721.50±0.117.60±2.112.48±0.612.05±0.0516.98±4.248.27±0.405.74±0.315.25412.10
SimMatchV2 (Ours)4.90±0.165.04±0.094.33±0.1636.68±0.8626.66±0.3821.37±0.207.92±2.802.92±0.812.85±0.9115.85±2.627.54±0.815.65±0.264.00111.81
", + "image_path": "400dc6641308bea5a7063c8a68940ebaa1a44593609fd44d9dc7ea30939413fd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 236, + 268, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 268, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 268, + 247 + ], + "type": "text", + "content": "There are four datasets will be used in this experiment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 250, + 286, + 507 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 46, + 250, + 286, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 250, + 286, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 286, + 321 + ], + "type": "text", + "content": "- CIFAR-10 [31] consists of 60K 32x32 color images in 10 classes, with 6k images per class. The training and test set contains 5K and 1k samples, respectively. We conduct three different experiments on 40, 250, and 4,000 randomly selected images from the training set as the labeled data and use the rest of the training set as the unlabeled data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 324, + 286, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 286, + 384 + ], + "type": "text", + "content": "- CIFAR-100 [31] is a more complex version of CIFAR-10, which contains 100 classes with 600 images each. There are 500 training images and 100 testing images per class. The experimental setting are similar to CIFAR-10 above except using 400, 2,500, and 10K labeled data for training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 386, + 286, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 386, + 286, + 457 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 286, + 457 + ], + "type": "text", + "content": "- SVHN [42] is a real-world image dataset that contains 32x32 cropped digits from house numbers in Google Street View images. It contains 73,257 training images, with 26,032 testing images and 531,131 extra images. We merge the training and extra set and randomly select 40, 250, and 1,000 from it as the label data to conduct the experiment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 460, + 286, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 286, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 286, + 507 + ], + "type": "text", + "content": "- STL-10 [14] consists of 5K labeled images, 100K unlabeled images, and 8,000 test images. We will randomly select 40, 250, and 1,000 from the 5K images; the rest images will be combined with 100K as the unlabeled data." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "content": "Implementation Details. Our implementation is mainly based on the USB [51] code base for a fair comparison. Most of the hyper-parameters are inherited from [43]. Specifically, we adopt WRN-28-2 [60] for CIFAR-10 and SVHN, WRN-28-8 for CIFAR-100, and WRN-37-2 for STL-10. We use an SGD optimizer with Nesterov momentum [45] for all experiments. The weight decay will be set to " + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "inline_equation", + "content": "5e^{-4}" + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "content": " for most experiments except CIFAR-100, for which we set it as " + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "inline_equation", + "content": "1e^{-3}" + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "content": ". For the learning rate schedule, we use a cosine learning rate decay [39], which adjusts the learning rate to " + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "inline_equation", + "content": "0.03 \\cdot \\cos\\left(\\frac{7\\pi s}{16S}\\right)" + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "content": " is the current training step, and " + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "content": " is the total number of training steps. The model will be optimized for " + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "inline_equation", + "content": "2^{20}" + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "content": " steps. We use 64 labeled data and 448 unlabeled data for each batch. All experiments will be conducted on a single GPU. The weak and strong augmentation policies are also directly inherited from the USB codebase. Note that the final performance will be reported based on" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "text", + "content": "the exponential moving average of the model with a 0.999 decay rate. For SimMatchV2 specific hyper-parameters, we set " + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "inline_equation", + "content": "\\lambda_{nn} = 1, \\lambda_{ne} = 1, \\lambda_{ee} = 1, t = 0.1, \\alpha = 0.1, \\tau = 0.95, n = 8" + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "text", + "content": ". (" + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "text", + "content": " is the Top-N nearest node for Eq. (7)) These hyper-parameters are kept unchanged for all experiments. We maintain two memory banks for storing both unlabeled nodes and labeled nodes. The unlabeled bank will be used for calculating the Eq. (4) and Eq. (5). The labeled bank will be used for computing the Eq. (7)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 346, + 545, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 657 + ], + "type": "text", + "content": "Performance on Classical SSL Benchmark. The overall results are reported in Table 1. We compare our SimMatchV2 against 14 SSL algorithms. As in Table 1, it is very hard for an algorithm to get the best performance on all settings. Many algorithms already meet or exceed the supervised baseline in some settings (CIFAR-10 and SVHN); the performance difference would not be significant in this case. For a more rigorous and fair comparison, we follow the steps from USB [51], which adopts the Friedman rank [19, 20] as the final evaluation metric. We also report the mean error rate for all settings. As can be seen, our SimMatchV2 achieves the best performance on 4 out of 12 settings and also has the highest Friedman rank and lowest mean error rate. CRMatch [18] also achieves a very good performance that ranked second in the final rank. It has the best performance on 5 out of the 12 settings, but it is less stable when the number of labeled data is limited, and the mean error rate is much worse than SimMatchV2. AdaMatch [4] and SimMatch [66] has a more stable performance under various conditions and they have a very close mean error rate compared to SimMatchV2. However, these two methods did not achieve any best results in this experiment, leading to a worse Friedman rank. Overall, the highest Friedman rank and lowest mean error rate for SimMatchV2 should demonstrate superiority in terms of both performance and stability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 669, + 450, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 450, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 450, + 681 + ], + "type": "text", + "content": "4.2. Experiments on ImageNet" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Next, we evaluate our SimMatchV2 on the large-scaled ImageNet dataset [15], which contains 1,000 categories" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16436" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 92, + 541, + 275 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 547, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 547, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 547, + 92 + ], + "type": "text", + "content": "Table 2: Top-1 an Top-5 Accuracy with ResNet-50 on ImageNet with " + }, + { + "bbox": [ + 47, + 71, + 547, + 92 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 47, + 71, + 547, + 92 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 71, + 547, + 92 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 47, + 71, + 547, + 92 + ], + "type": "text", + "content": " settings. Note that the underline in the parameters - inference columns denotes the methods which require more parameters than standard ResNet-50 during the inference stage." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 92, + 541, + 275 + ], + "lines": [ + { + "bbox": [ + 52, + 92, + 541, + 275 + ], + "spans": [ + { + "bbox": [ + 52, + 92, + 541, + 275 + ], + "type": "table", + "html": "
Pre-training AlgorithmSemi-supervised AlgorithmPre-training EpochsSemi-supervised EpochsParamters trainParamters inferenceTop-1 Label fractionTop-5 Label fraction
1%10%1%10%
NoneVAT+EntMin. [41, 21, 61]-~10025.6M25.6M--47.083.4
S4L-Rotation [61]-~20025.6M25.6M--53.483.8
UDA [55]-~53025.6M25.6M-68.8-88.8
FixMatch [43]-~30025.6M25.6M-71.5-89.1
FixMatch-EMAN [6]-~30025.6M25.6M-72.8-90.3
SsCL [64]-~80028.0M25.6M60.272.182.890.9
CoMatch [34]-~40028.0M25.6M66.073.686.491.6
SimMatch [66]-~40028.0M25.6M67.274.487.191.7
PCL [35]Fine-tune~200~2023.8M25.6M--75.385.6
SimCLR [9]~1000~6028.0M25.6M48.365.675.587.8
SimCLR V2 [10]~800~6032.2M29.8M57.968.482.589.2
BYOL [22]~1000~5035.1M25.6M53.268.878.489.0
SwAV [7]~800~2028.0M25.6M53.970.278.589.9
WCL [65]~800~6032.2M29.8M65.072.086.391.2
MoCo V2 [11]Fine-tune~800~2028.0M25.6M49.866.177.287.9
FixMatch-EMAN [6]~800~30028.0M25.6M61.473.982.191.0
PAWS [7]Fine-tune~300~5036.1M29.8M66.575.5--
NoneSimMatchV2 (Ours)-~10028.0M25.6M69.974.888.791.7
-~30028.0M25.6M71.976.290.092.2
", + "image_path": "6935e140a10ef8af42e0475de3800f9882d35a22d9c9334fb3f63b40e0e1d831.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 283, + 287, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 368 + ], + "type": "text", + "content": "with 1.2M training images and 50K validation images. Concretely, we follow previous works [34, 66, 1, 6] to take " + }, + { + "bbox": [ + 46, + 283, + 287, + 368 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 283, + 287, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 283, + 287, + 368 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 283, + 287, + 368 + ], + "type": "text", + "content": " samples from the training images as our labeled data and use the rest samples as the unlabeled data. To avoid the effects of different selections on the labeled data, we follow [9, 10, 6, 1, 65] to adopt a pre-defined labeled / unlabeled splits1 that has been adopted in most SSL papers." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "spans": [ + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "text", + "content": "Implementation Details. For ImageNet experiments, most of our hyper-parameter and settings are directly inherited from [66]. Specifically, we use ResNet-50 [25] as our default backbone and use the SGD with Nesterov momentum as the optimizer. We set the weight decay to 0.0001 during the training. For the learning rate, we warm up the model for 5 epochs until it reaches 0.03, and it will be decayed to 0 with the cosine scheduler [39]. By following [6, 66], we use 64 labeled data and 320 unlabeled data in each batch; the data will be evenly distributed on 8 GPUs. For the hyper-parameter settings, we set " + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "inline_equation", + "content": "\\lambda_{nn} = 10" + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "inline_equation", + "content": "\\lambda_{ne} = 10" + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "inline_equation", + "content": "\\lambda_{ee} = 10" + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "inline_equation", + "content": "t = 0.1" + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "inline_equation", + "content": "\\tau = 0.7" + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "inline_equation", + "content": "n = 128" + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "text", + "content": " setting, and we use " + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "inline_equation", + "content": "\\lambda_{nn} = 5" + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "inline_equation", + "content": "\\lambda_{ne} = 5" + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 368, + 288, + 596 + ], + "type": "text", + "content": " since we found the performance is slightly better. For strong augmentation, we adopt a multi-crop strategy as in [68], which contains five crops with different sizes (224x224, 192x192, 160x160, 128x128, and 96x96); these crops will be augmented with the policy in [8]. For weak augmentation, we follow the same strategy in [66, 34]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 597, + 287, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 597, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 287, + 681 + ], + "type": "text", + "content": "Performance on ImageNet. The performance has been shown in Table 2. With 100 epochs of training, SimMatchV2 achieves " + }, + { + "bbox": [ + 46, + 597, + 287, + 681 + ], + "type": "inline_equation", + "content": "70.4\\%" + }, + { + "bbox": [ + 46, + 597, + 287, + 681 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 597, + 287, + 681 + ], + "type": "inline_equation", + "content": "74.8\\%" + }, + { + "bbox": [ + 46, + 597, + 287, + 681 + ], + "type": "text", + "content": " Top-1 accuracy with " + }, + { + "bbox": [ + 46, + 597, + 287, + 681 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 597, + 287, + 681 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 597, + 287, + 681 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 597, + 287, + 681 + ], + "type": "text", + "content": " labeled data, which already surpass the performance of CoMatch [34], and SimMatch [66] with 400 epochs. With 300 epochs of training, SimMatchV2 achieves " + }, + { + "bbox": [ + 46, + 597, + 287, + 681 + ], + "type": "inline_equation", + "content": "72.1\\%" + }, + { + "bbox": [ + 46, + 597, + 287, + 681 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 597, + 287, + 681 + ], + "type": "inline_equation", + "content": "76.2\\%" + }, + { + "bbox": [ + 46, + 597, + 287, + 681 + ], + "type": "text", + "content": " Top-1 accuracy, which improves absolute" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 283, + 547, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 283, + 547, + 368 + ], + "spans": [ + { + "bbox": [ + 306, + 283, + 547, + 368 + ], + "type": "text", + "content": "+5.2% and +0.7% over PAWS [1]. As we can observe that the improvement of 10% setting is much lesser than 1% setting. We doubt that this is because the performance of 76.2% is already very close to the fully supervised setting (76.5%), which should be the upper bound of the SSL tasks. Most importantly, our SimMatchV2 established a new state-of-the-art performance without any pre-training." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 408, + 547, + 467 + ], + "blocks": [ + { + "bbox": [ + 305, + 377, + 547, + 407 + ], + "lines": [ + { + "bbox": [ + 305, + 377, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 377, + 547, + 407 + ], + "type": "text", + "content": "Table 3: Runtime Comparison. MC denotes whether we use the multicrop strategy. The experiments are performed on " + }, + { + "bbox": [ + 305, + 377, + 547, + 407 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 305, + 377, + 547, + 407 + ], + "type": "text", + "content": " ImageNet with 100 epochs of training. " + }, + { + "bbox": [ + 305, + 377, + 547, + 407 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 305, + 377, + 547, + 407 + ], + "type": "text", + "content": " indicates lower is better and " + }, + { + "bbox": [ + 305, + 377, + 547, + 407 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 305, + 377, + 547, + 407 + ], + "type": "text", + "content": " means higher is better." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 408, + 547, + 467 + ], + "lines": [ + { + "bbox": [ + 307, + 408, + 547, + 467 + ], + "spans": [ + { + "bbox": [ + 307, + 408, + 547, + 467 + ], + "type": "table", + "html": "
MethodMCGPU Memory ↓GPU days ↓1% Top-1 ↑
CoMatch [34]✗81G10.361.1
SimMatch [66]✗40G7.761.2
SimMatchV2✗42G7.863.7
PAWS [1]✓1048G22.663.8
SimMatchV2✓83G14.769.9
", + "image_path": "01b1c8f1098f182b936785798403fa904e25e997c9cd79eec79594aec04a512b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "text", + "content": "Efficiency Comparison. Next, We also compare the training efficiency for different methods. The experiments are conducted with 8 Nvidia GPUs except for PAWS [1], which requires at least 64 GPUs to run. We use mixed precision training for all methods. Note that CoMatch [34] and SimMatch [66] have their own labeled / unlabeled splits, which makes the comparison a bit unfair. In this experiment, we adopt their officially released codebase to reproduce the performance on the same splits." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "We show the training statistics in Table 3. When we omit the multi-crops, the computational cost of SimMatchV2 is slightly higher compared to SimMatch [66] (i.e. +2G memory, +0.1 GPU days), and it is much more efficient than CoMatch [34]. We can also observe that SimMatchV2 has much better performance (+2.5%) than CoMatch [34] and SimMatch [66] even without the multi-crops. Moreover, when the multi-crops are included, the training time of SimMatchV2 is nearly doubled (7.8 days → 14.7 days), but it is still only 65% of PAWS [1] (14.7 days / 22.6 days). The extra training cost of PAWS [1] is from the support" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 276, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 276, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 276, + 713 + ], + "type": "text", + "content": "1https://www.tensorflow.org/datasets/catalog/Imagenet2012 Subset" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16437" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "set containing " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "960 \\times 7" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": " labeled data per batch. Most importantly, our SimMatchV2 achieves similar performance with PAWS (63.7% vs. 63.8%) [1] with only 35% training time (7.8 days / 22.6 days), and has an absolute improvement of +6.6% with 65% training time." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 51, + 157, + 282, + 251 + ], + "blocks": [ + { + "bbox": [ + 47, + 137, + 287, + 156 + ], + "lines": [ + { + "bbox": [ + 47, + 137, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 137, + 287, + 156 + ], + "type": "text", + "content": "Table 4: More experiments on different architectures. Note that we do not use any pre-training in this experiment." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 157, + 282, + 251 + ], + "lines": [ + { + "bbox": [ + 51, + 157, + 282, + 251 + ], + "spans": [ + { + "bbox": [ + 51, + 157, + 282, + 251 + ], + "type": "table", + "html": "
MethodArchParamsEpochs1%10%
Semifommer [52]ViT-S + Conv40M600-75.5
ProbPseudo Mixup [5]ViT-Small22M100-70.9
SimMatchV2ViT-Small22M10063.775.9
ProbPseudo Mixup [5]ViT-Base86M100-73.5
SimMatchV2ViT-Base86M10065.376.8
ProbPseudo Mixup [5]ConvNext-T28M100-74.1
SimMatchV2ConvNext-T28M10071.377.8
ProbPseudo Mixup [5]ConvNext-S50M100-75.1
SimMatchV2ConvNext-S50M10072.378.4
", + "image_path": "8ad309088d1147d06506d70866fd814ad718d5c6f590cbd26413d114599fdbeb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 255, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 460 + ], + "type": "text", + "content": "Working with More Advanced Architectures. With the recent successes in model architecture design, many more powerful models have been proposed, especially the vision transformer and its variant [16, 48, 49, 38, 37]. Thus, we further apply our SimMatchV2 on ViT-Small [48], ViT-Base [16], ConvNext-T [38], and ConvNext-S [38] to show the generality on different architectures. In this experiment, we follow [5] to use 128 labeled data and 640 unlabeled per batch. We also adopt the AdamW [40] optimizer with the learning rate of " + }, + { + "bbox": [ + 46, + 255, + 287, + 460 + ], + "type": "inline_equation", + "content": "1.25e^{-4}" + }, + { + "bbox": [ + 46, + 255, + 287, + 460 + ], + "type": "text", + "content": " and weight decay of 0.1. We present the results in Table 4. As we can see, SimMatchV2 with a better model can generally produce a better result. The exploration of cooperating with more advanced SSL pipeline [5] and stronger data augmentations (e.g. ProbPseudo Mixup [5]) will be left as a future work since it requires more hyper-parameter turnings and more sophisticated designs, which is not the focus of this paper." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 464, + 217, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 217, + 476 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 217, + 476 + ], + "type": "text", + "content": "4.3. Experiments on Semi-iNat 2021" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 58, + 490, + 276, + 586 + ], + "blocks": [ + { + "bbox": [ + 67, + 479, + 265, + 489 + ], + "lines": [ + { + "bbox": [ + 67, + 479, + 265, + 489 + ], + "spans": [ + { + "bbox": [ + 67, + 479, + 265, + 489 + ], + "type": "text", + "content": "Table 5: Accuracy for real-world dataset - Semi-iNat 2021." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 490, + 276, + 586 + ], + "lines": [ + { + "bbox": [ + 58, + 490, + 276, + 586 + ], + "spans": [ + { + "bbox": [ + 58, + 490, + 276, + 586 + ], + "type": "table", + "html": "
MethodSemi-iNat 2021
With Random InitWith MoCo Init
Top1Top5Top1Top5
Supervised19.0935.8534.9657.11
MixMatch [3]16.8930.83--
+CCSSL [58]19.6535.09--
FixMatch [43]21.4137.6540.3060.05
+CCSSL [58]31.2152.2541.2864.30
CoMatch [34]20.9438.9638.9461.85
+CCSSL [58]24.1243.2339.8563.68
SimMatchV2 (Ours)43.0563.8153.1373.40
", + "image_path": "dea4dce821a89b7b0f3bb2d10cc225c6094d33394fcea580e0cde184d4adc6ee.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Finally, we also test the performance of SimMatchV2 on Semi-iNat 2021 [44], which contains 810 different categories with 9,721 labeled data, 313,248 unlabeled data, and 4,050 validation data. Unlike the dataset we used in previous experiments, Semi-iNat 2021 is designed to expose some challenges encountered in a real-world scenario, which contains highly similar classes, class imbalance, and domain mismatch between the labeled and unlabeled data. Such a more realistic dataset should be a better benchmark to reflect the performance of the different methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "In this experiment, most of the hyper-parameters and implementation details are the same as our ImageNet setting, except that we do not use the multi-crops strategy for fair comparison. The model will be optimized for " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "512 \\times 1024" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " steps as in [58]. We also follow the setting in [58] to conduct two experiments to report the training from scratch performance and training from pre-trained model performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 157, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 545, + 264 + ], + "type": "text", + "content": "We show the results in Table 5. Our SimMatchV2 achieves " + }, + { + "bbox": [ + 304, + 157, + 545, + 264 + ], + "type": "inline_equation", + "content": "43.05\\%" + }, + { + "bbox": [ + 304, + 157, + 545, + 264 + ], + "type": "text", + "content": " Top-1 accuracy when training from scratch, which is " + }, + { + "bbox": [ + 304, + 157, + 545, + 264 + ], + "type": "inline_equation", + "content": "+11.84\\%" + }, + { + "bbox": [ + 304, + 157, + 545, + 264 + ], + "type": "text", + "content": " better than the state-of-the-art method (FixMatch + CCSL). When the MoCo pre-trained model is adopted, our SimMatchV2 achieves a " + }, + { + "bbox": [ + 304, + 157, + 545, + 264 + ], + "type": "inline_equation", + "content": "53.13\\%" + }, + { + "bbox": [ + 304, + 157, + 545, + 264 + ], + "type": "text", + "content": " Top-1 accuracy, which is " + }, + { + "bbox": [ + 304, + 157, + 545, + 264 + ], + "type": "inline_equation", + "content": "+10.08\\%" + }, + { + "bbox": [ + 304, + 157, + 545, + 264 + ], + "type": "text", + "content": " better than the from-scratch performance, and " + }, + { + "bbox": [ + 304, + 157, + 545, + 264 + ], + "type": "inline_equation", + "content": "+11.85\\%" + }, + { + "bbox": [ + 304, + 157, + 545, + 264 + ], + "type": "text", + "content": " better than FixMatch + CCSL. We believe this experiment will further demonstrate the superiority of SimMatchV2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 277, + 397, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 277, + 397, + 291 + ], + "spans": [ + { + "bbox": [ + 306, + 277, + 397, + 291 + ], + "type": "text", + "content": "5. Ablation study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 298, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 545, + 382 + ], + "type": "text", + "content": "In this section, we will empirically study our SimMatchV2 based on various conditions and show the effect of each component and hyper-parameter sensitivities of our methods. For all experiments in this section, we adopt the ResNet-50 as our backbone and train the model on ImageNet for 100 epochs on the " + }, + { + "bbox": [ + 304, + 298, + 545, + 382 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 298, + 545, + 382 + ], + "type": "text", + "content": " labeled data. The multicrops are not used by default unless we mentioned." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 383, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 545, + 514 + ], + "type": "text", + "content": "Training Dynamics. We first show the training dynamics of SimMatchV2 compared to SimMatch [66] and CoMatch [34]. We visualize the pseudo-labels accuracy, unlabeled data accuracy, and validation accuracy in Figure 3. As we can observe that the pseudo-labels from SimMatchV2 have a higher quality than SimMatch [66], especially in the late stage of the training. Both of these two methods generate better pseudo-labels than CoMatch [34]. We can also notice that SimMatchV2 consistently has better accuracy for the unlabeled data and validation data, which shows the advantages of our method." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 308, + 552, + 543, + 609 + ], + "blocks": [ + { + "bbox": [ + 306, + 521, + 545, + 551 + ], + "lines": [ + { + "bbox": [ + 306, + 521, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 521, + 545, + 551 + ], + "type": "text", + "content": "Table 6: Effect of each component in SimMatchV2. N and E denote the node and edge, respectively. N-E means the node-edge consistency and vice versa. Feat norm means feature normalization." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 552, + 543, + 609 + ], + "lines": [ + { + "bbox": [ + 308, + 552, + 543, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 552, + 543, + 609 + ], + "type": "table", + "html": "
N-NN-EE-EE-NFeat NormTop-1Improvement
✓52.9-
✓✓60.0+7.1
✓✓✓60.7+0.7
✓✓✓✓62.2+1.5
✓✓✓✓✓63.7+1.5
", + "image_path": "ddd3bc36aa2c1206639138b60a678d8ec0cc0211d44e35bb5937e8c6beb0d2b2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Effect of Each Components. In table 6, we show the effect of each design in SimMatchV2. Basically, if only node-node consistency is adopted, the setting is equivalent to the FixMatch [43] with distribution alignment [2], which is our baseline in the first row of the table; Next, we can see that the node-edge, edge-node, and edge-node consistencies give " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "+7.1\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "+0.7\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "+1.5\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " improvements, respectively. Finally, the feature normalization also boosts the performance" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16438" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 71, + 225, + 177 + ], + "blocks": [ + { + "bbox": [ + 85, + 71, + 225, + 177 + ], + "lines": [ + { + "bbox": [ + 85, + 71, + 225, + 177 + ], + "spans": [ + { + "bbox": [ + 85, + 71, + 225, + 177 + ], + "type": "image", + "image_path": "9f24b8eb5de621e2fa8f4ee27bfaa75b8a1771ffa4e06f933ac23b96ea5fa336.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 185, + 202, + 195 + ], + "lines": [ + { + "bbox": [ + 107, + 185, + 202, + 195 + ], + "spans": [ + { + "bbox": [ + 107, + 185, + 202, + 195 + ], + "type": "text", + "content": "(a) Pseudo-Labels Accuracy" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 227, + 72, + 365, + 177 + ], + "blocks": [ + { + "bbox": [ + 227, + 72, + 365, + 177 + ], + "lines": [ + { + "bbox": [ + 227, + 72, + 365, + 177 + ], + "spans": [ + { + "bbox": [ + 227, + 72, + 365, + 177 + ], + "type": "image", + "image_path": "ce6cc88b43487746d9e31ad5e91ec04deff66e3a651cb57ea5fd5aae7c4234ff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 184, + 343, + 194 + ], + "lines": [ + { + "bbox": [ + 247, + 184, + 343, + 194 + ], + "spans": [ + { + "bbox": [ + 247, + 184, + 343, + 194 + ], + "type": "text", + "content": "(b) Unlabeled Data Accuracy" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 197, + 547, + 227 + ], + "lines": [ + { + "bbox": [ + 46, + 197, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 197, + 547, + 227 + ], + "type": "text", + "content": "Figure 3: In this Figure, we compare the training dynamics of SimMatchV2 with SimMatch [66], and CoMatch [34] from 3 different perspectives. a) The accuracy of " + }, + { + "bbox": [ + 46, + 197, + 547, + 227 + ], + "type": "inline_equation", + "content": "p^w" + }, + { + "bbox": [ + 46, + 197, + 547, + 227 + ], + "type": "text", + "content": " that pass the threshold " + }, + { + "bbox": [ + 46, + 197, + 547, + 227 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 197, + 547, + 227 + ], + "type": "text", + "content": ". b) The accuracy of " + }, + { + "bbox": [ + 46, + 197, + 547, + 227 + ], + "type": "inline_equation", + "content": "p^w" + }, + { + "bbox": [ + 46, + 197, + 547, + 227 + ], + "type": "text", + "content": " regardless of the threshold. c) The accuracy on the validation data. All experiments are in 1% ImageNet setting with 100 epochs training." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 368, + 72, + 506, + 177 + ], + "blocks": [ + { + "bbox": [ + 368, + 72, + 506, + 177 + ], + "lines": [ + { + "bbox": [ + 368, + 72, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 368, + 72, + 506, + 177 + ], + "type": "image", + "image_path": "acc214e8c318ffaec44db1c4de232ab6d7ca04e0016e4fffdbda44759864bb44.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 184, + 477, + 195 + ], + "lines": [ + { + "bbox": [ + 397, + 184, + 477, + 195 + ], + "spans": [ + { + "bbox": [ + 397, + 184, + 477, + 195 + ], + "type": "text", + "content": "(c) Validation Accuracy" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 233, + 288, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 233, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 233, + 288, + 258 + ], + "type": "text", + "content": "+1.5%. This experiment should demonstrate that these different designs are all very effective for SSL tasks." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 48, + 294, + 286, + 323 + ], + "blocks": [ + { + "bbox": [ + 47, + 263, + 287, + 293 + ], + "lines": [ + { + "bbox": [ + 47, + 263, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 263, + 287, + 293 + ], + "type": "text", + "content": "Table 7: Effect of memory bank size on the performance. " + }, + { + "bbox": [ + 47, + 263, + 287, + 293 + ], + "type": "inline_equation", + "content": "5\\mathrm{k}-163\\mathrm{K}" + }, + { + "bbox": [ + 47, + 263, + 287, + 293 + ], + "type": "text", + "content": " are setting that we use unlabeled bank; labeled bank means we use all labeled bank (All labeled data 12K for " + }, + { + "bbox": [ + 47, + 263, + 287, + 293 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 47, + 263, + 287, + 293 + ], + "type": "text", + "content": " ImageNet)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 294, + 286, + 323 + ], + "lines": [ + { + "bbox": [ + 48, + 294, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 286, + 323 + ], + "type": "table", + "html": "
Bank Size5K10K40K82K163KLabeled Bank (12K)
Top-163.163.263.563.763.663.2
", + "image_path": "a94b22f782aede1e9014829522cd8dd598b44a8bdc3c646961695406ed2cc90d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 328, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 449 + ], + "type": "text", + "content": "Node-Edge and Edge-Edge Consistency. We will study the node-edge and edge-edge consistency together since the memory bank size is their primary factor. The experimental results are shown in Table 7. We can observe that when unlabeled data is used, the performance will be slightly increased with a larger memory bank, and the best memory bank size is " + }, + { + "bbox": [ + 46, + 328, + 287, + 449 + ], + "type": "inline_equation", + "content": "82\\mathrm{K}" + }, + { + "bbox": [ + 46, + 328, + 287, + 449 + ], + "type": "text", + "content": ". We can also notice that we get slightly lower performance with " + }, + { + "bbox": [ + 46, + 328, + 287, + 449 + ], + "type": "inline_equation", + "content": "12\\mathrm{K}" + }, + { + "bbox": [ + 46, + 328, + 287, + 449 + ], + "type": "text", + "content": " labeled data. Therefore, we decide to use the unlabeled data for node-edge and edge-edge consistency." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 48, + 475, + 285, + 504 + ], + "blocks": [ + { + "bbox": [ + 47, + 454, + 287, + 475 + ], + "lines": [ + { + "bbox": [ + 47, + 454, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 287, + 475 + ], + "type": "text", + "content": "Table 8: Effect of Top-N to Edge-Node Consistency. (We use labeled bank for this experiment.)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 475, + 285, + 504 + ], + "lines": [ + { + "bbox": [ + 48, + 475, + 285, + 504 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 285, + 504 + ], + "type": "table", + "html": "
N8163264128256
Top-163.463.563.563.563.763.6
", + "image_path": "8cde9d25925282561de83ad0976462716614c8df0c6e48c212b2faf15e852b4f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 49, + 535, + 285, + 564 + ], + "blocks": [ + { + "bbox": [ + 47, + 514, + 286, + 533 + ], + "lines": [ + { + "bbox": [ + 47, + 514, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 286, + 533 + ], + "type": "text", + "content": "Table 9: Effect of labeled / unlabeled bank and " + }, + { + "bbox": [ + 47, + 514, + 286, + 533 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 47, + 514, + 286, + 533 + ], + "type": "text", + "content": " in Eq. (6) to Edge-Node Consistency." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 535, + 285, + 564 + ], + "lines": [ + { + "bbox": [ + 49, + 535, + 285, + 564 + ], + "spans": [ + { + "bbox": [ + 49, + 535, + 285, + 564 + ], + "type": "table", + "html": "
Labeled Y∞Labeled Y1Unlabeled Y∞Unlabeled Y1
Top-163.763.362.462.2
", + "image_path": "12231734a35b1ea52bb8e132754653171e89b8e852a0c17c126450305dc35d22.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": "Edge-Node Consistency. As we have mentioned, it is computationally very expensive to involve the whole memory bank for the edge-node consistency calculation. Thus, we only adopt the Top-N nearest neighbor to propagate the label. We show this ablation study in Table 8. The performance is very robust with the selection of the number of Top-N, the best results come from " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "N = 128" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": ", and the performance is only slightly decreased -0.3% in the worst case. Note that all experiments in Table 8 are performed with the labeled bank; we also have the ablation study to test the effect of the labeled / unlabeled bank and the " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": " in Eq. (6). In Table 9, we can see that the performance with the labeled" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 233, + 545, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 233, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 305, + 233, + 545, + 269 + ], + "type": "text", + "content": "bank is much better than the unlabeled bank. Also, the performance with the fully converged " + }, + { + "bbox": [ + 305, + 233, + 545, + 269 + ], + "type": "inline_equation", + "content": "Y_{\\infty}" + }, + { + "bbox": [ + 305, + 233, + 545, + 269 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 305, + 233, + 545, + 269 + ], + "type": "inline_equation", + "content": "+0.4\\%" + }, + { + "bbox": [ + 305, + 233, + 545, + 269 + ], + "type": "text", + "content": " better than " + }, + { + "bbox": [ + 305, + 233, + 545, + 269 + ], + "type": "inline_equation", + "content": "Y_{1}" + }, + { + "bbox": [ + 305, + 233, + 545, + 269 + ], + "type": "text", + "content": ", which shows the necessity of our design." + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 328, + 284, + 523, + 321 + ], + "blocks": [ + { + "bbox": [ + 353, + 273, + 499, + 283 + ], + "lines": [ + { + "bbox": [ + 353, + 273, + 499, + 283 + ], + "spans": [ + { + "bbox": [ + 353, + 273, + 499, + 283 + ], + "type": "text", + "content": "Table 10: Effect of Feature Normalization" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 284, + 523, + 321 + ], + "lines": [ + { + "bbox": [ + 328, + 284, + 523, + 321 + ], + "spans": [ + { + "bbox": [ + 328, + 284, + 523, + 321 + ], + "type": "table", + "html": "
Top-1Norm Gap ↓CE Loss ↓
W/o Feat Norm62.21.832.75
W/ Feat Norm63.71.312.63
", + "image_path": "68fcf41140650a82f0006db03fbb0f36bb4473d495b9f609b529aa5e307bfb69.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 324, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 545, + 445 + ], + "type": "text", + "content": "Feature Normalization. Finally, we show the effect of feature normalization in Table 10. As we mentioned, feature normalization aims to reduce the norm gaps between the weak and strong augmented views. Thus, to validate this, we calculate the mean of the absolute value of " + }, + { + "bbox": [ + 304, + 324, + 545, + 445 + ], + "type": "inline_equation", + "content": "\\| h^s\\| -\\| h^w\\|" + }, + { + "bbox": [ + 304, + 324, + 545, + 445 + ], + "type": "text", + "content": " across the whole training process. We name it \"norm gaps\" in Table 10. With the feature normalization, the norm gaps can be reduced from " + }, + { + "bbox": [ + 304, + 324, + 545, + 445 + ], + "type": "inline_equation", + "content": "1.83\\rightarrow 1.31" + }, + { + "bbox": [ + 304, + 324, + 545, + 445 + ], + "type": "text", + "content": ". We also calculate the mean of the CE loss Eq. (3) (regardless of the threshold, i.e. " + }, + { + "bbox": [ + 304, + 324, + 545, + 445 + ], + "type": "inline_equation", + "content": "\\tau = 0" + }, + { + "bbox": [ + 304, + 324, + 545, + 445 + ], + "type": "text", + "content": ") across the training process." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 456, + 379, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 379, + 469 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 379, + 469 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 477, + 545, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 545, + 644 + ], + "type": "text", + "content": "In this work, we propose a new semi-supervised learning algorithm - SimMatchV2, which models consistency regularization from the graph perspective. SimMatchV2 provides four different types of consistency by leveraging the idea of message passing in graph theory and also proposes a simple feature normalization to reduce the gaps between the feature norm of different augmented views. The experiments on different benchmarks and various architectures demonstrate state-of-the-art performance. The limitation of this work is the missing exploration of cooperating with the more advanced SSL training pipelines and stronger data augmentations [5] since it needs more experiments to find the optimal setting. We plan to investigate the pre-training setting in our future work." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 657, + 399, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 399, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 399, + 670 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "This work was supported in part by the Australian Research Council under Project DP210101859 and the University of Sydney Research Accelerator (SOAR) Prize." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16439" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "type": "text", + "content": "[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. arXiv preprint arXiv:2104.13963, 2021. 1, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "text", + "content": "[2] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. arXiv preprint arXiv:1911.09785, 2019. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 247 + ], + "type": "text", + "content": "[3] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. arXiv preprint arXiv:1905.02249, 2019. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 249, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 249, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 249, + 288, + 294 + ], + "type": "text", + "content": "[4] David Berthelot, Rebecca Roelofs, Kihyuk Sohn, Nicholas Carlini, and Alex Kurakin. Adamatch: A unified approach to semi-supervised learning and domain adaptation. arXiv preprint arXiv:2106.04732, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 287, + 339 + ], + "type": "text", + "content": "[5] Zhaowei Cai, Avinash Ravichandran, Paolo Favaro, Manchen Wang, Davide Modolo, Rahul Bhotika, Zhuowen Tu, and Stefano Soatto. Semi-supervised vision transformers at scale. In NeurIPS, 2022. 3, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 340, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 287, + 407 + ], + "type": "text", + "content": "[6] Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Charless Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 194-203, 2021. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 407, + 287, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 287, + 451 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 287, + 451 + ], + "type": "text", + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. 2020. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 453, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 287, + 497 + ], + "type": "text", + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. arXiv preprint arXiv:2104.14294, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 498, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 287, + 542 + ], + "type": "text", + "content": "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709, 2020. 2, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "text", + "content": "[10] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey Hinton. Big self-supervised models are strong semi-supervised learners. arXiv preprint arXiv:2006.10029, 2020. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "type": "text", + "content": "[11] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[12] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[13] Ching-Yao Chuang, Joshua Robinson, Lin Yen-Chen, Antonio Torralba, and Stefanie Jegelka. Debiased contrastive learning. In Advances in neural information processing systems, 2020. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[14] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. volume 15 of Proceedings of Machine Learning Research, pages 215-223, Fort Lauderdale, FL, USA, 11-13 Apr 2011. JMLR Workshop and Conference Proceedings. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "text", + "content": "[15] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPRO9, 2009. 1, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 162, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 227 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "text", + "content": "[17] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 272, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 315 + ], + "type": "text", + "content": "[18] Yue Fan, Anna Kukleva, Dengxin Dai, and Bernt Schiele. Revisiting consistency regularization for semi-supervised learning. International Journal of Computer Vision, pages 1-18, 2022. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 316, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 358 + ], + "type": "text", + "content": "[19] Milton Friedman. The use of ranks to avoid the assumption of normality implicit in the analysis of variance. Journal of the american statistical association, 32(200):675-701, 1937. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 360, + 545, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 393 + ], + "type": "text", + "content": "[20] Milton Friedman. A comparison of alternative tests of significance for the problem of m rankings. The Annals of Mathematical Statistics, 11(1):86-92, 1940. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "type": "text", + "content": "[21] Yves Grandvalet, Yoshua Bengio, et al. Semi-supervised learning by entropy minimization. CAP, 367:281-296, 2005. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 426, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 491 + ], + "type": "text", + "content": "[22] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. arXiv preprint arXiv:2006.07733, 2020. 2, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 492, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 492, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 492, + 545, + 547 + ], + "type": "text", + "content": "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 548, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 591 + ], + "type": "text", + "content": "[24] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 592, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 624 + ], + "type": "text", + "content": "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. arXiv preprint arXiv:1512.03385, 2015. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[26] Qianjiang Hu, Xiao Wang, Wei Hu, and Guo-Jun Qi. Adco: Adversarial contrast for efficient learning of unsupervised representations from self-trained negative adversaries. arXiv preprint arXiv:2011.08435, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[27] Lang Huang, Shan You, Mingkai Zheng, Fei Wang, Chen Qian, and Toshihiko Yamasaki. Green hierarchical vision transformer for masked image modeling. Advances in Neural Information Processing Systems, 35:19997-20010, 2022. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16440" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[28] Lang Huang, Shan You, Mingkai Zheng, Fei Wang, Chen Qian, and Toshihiko Yamasaki. Learning where to learn in cross-view self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14451-14460, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 172 + ], + "type": "text", + "content": "[29] Yannis Kalantidis, Mert Bulent Sariyildiz, Noe Pion, Philippe Weinzaepfel, and Diane Larlus. Hard negative mixing for contrastive learning. In Neural Information Processing Systems (NeurIPS), 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 205 + ], + "type": "text", + "content": "[30] Soroush Abbasi Koohpayegani, Ajinkya Tejankar, and Hamed Pirsiavash. Mean shift for self-supervised learning. arXiv preprint arXiv:2105.07269, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 288, + 228 + ], + "type": "text", + "content": "[31] A. Krizhevsky. Learning multiple layers of features from tiny images. 2009. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 259 + ], + "type": "text", + "content": "[32] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 261, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 304 + ], + "type": "text", + "content": "[33] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 1, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 305, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 288, + 358 + ], + "type": "text", + "content": "[34] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regularization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9475-9484, 2021. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "type": "text", + "content": "[35] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, 2021. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "type": "text", + "content": "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 459, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 288, + 514 + ], + "type": "text", + "content": "[37] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "type": "text", + "content": "[38] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 288, + 602 + ], + "type": "text", + "content": "[39] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016.5,6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 288, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 625 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 625 + ], + "type": "text", + "content": "[40] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "type": "text", + "content": "[41] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979–1993, 2018. 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[42] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "type": "text", + "content": "[43] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. arXiv preprint arXiv:2001.07685, 2020. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 546, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 152 + ], + "type": "text", + "content": "[44] Jong-Chyi Su and Subhransu Maji. The semi-supervised inaturalist challenge at the fgvc8 workshop, 2021. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 546, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 546, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 546, + 196 + ], + "type": "text", + "content": "[45] Ilya Sutskever, James Martens, George Dahl, and Geoffrey Hinton. On the importance of initialization and momentum in deep learning. In International conference on machine learning, pages 1139-1147. PMLR, 2013. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 546, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 546, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 546, + 241 + ], + "type": "text", + "content": "[46] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. arXiv preprint arXiv:1703.01780, 2017. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 242, + 546, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 546, + 286 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 546, + 286 + ], + "type": "text", + "content": "[47] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? arXiv preprint arXiv:2005.10243, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 287, + 546, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 546, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 546, + 342 + ], + "type": "text", + "content": "[48] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 343, + 546, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 546, + 398 + ], + "type": "text", + "content": "[49] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 32-42, 2021. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 399, + 546, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 546, + 433 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 546, + 433 + ], + "type": "text", + "content": "[50] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. arXiv preprint arXiv:2005.10242, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 434, + 546, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 546, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 546, + 531 + ], + "type": "text", + "content": "[51] Yidong Wang, Hao Chen, Yue Fan, Wang Sun, Ran Tao, Wenxin Hou, Renjie Wang, Linyi Yang, Zhi Zhou, Lan-Zhe Guo, Heli Qi, Zhen Wu, Yu-Feng Li, Satoshi Nakamura, Wei Ye, Marios Savvides, Bhiksha Raj, Takahiro Shinozaki, Bernt Schiele, Jindong Wang, Xing Xie, and Yue Zhang. Usb: A unified semi-supervised learning benchmark for classification. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 533, + 546, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 546, + 567 + ], + "type": "text", + "content": "[52] Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, and Yu-Gang Jiang. Semi-supervised vision transformers. In ECCV, 2022. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 568, + 546, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 546, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 546, + 612 + ], + "type": "text", + "content": "[53] Zhirong Wu, Yuanjun Xiong, X Yu Stella, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 613, + 546, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 546, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 546, + 669 + ], + "type": "text", + "content": "[54] Haoyu Xie, Changqi Wang, Mingkai Zheng, Minjing Dong, Shan You, Chong Fu, and Chang Xu. Boosting semisupervised semantic segmentation with probabilistic representations. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2938-2946, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 669, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 546, + 714 + ], + "type": "text", + "content": "[55] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16441" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "text", + "content": "Processing Systems, volume 33, pages 6256-6268. Curran Associates, Inc., 2020. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "type": "text", + "content": "[56] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and improving layer normalization. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 239 + ], + "type": "text", + "content": "[57] Yi Xu, Lei Shang, Jinxing Ye, Qi Qian, Yu-Feng Li, Baigui Sun, Hao Li, and Rong Jin. Dash: Semi-supervised learning with dynamic thresholding. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 11525-11536. PMLR, 18-24 Jul 2021. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 306 + ], + "type": "text", + "content": "[58] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14421-14430, June 2022. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 308, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 342 + ], + "type": "text", + "content": "[59] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. In Advances in Neural Information Processing Systems, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 343, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 364 + ], + "type": "text", + "content": "[60] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "type": "text", + "content": "[61] Xiaohua Zhai, A. Oliver, A. Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1476-1485, 2019. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 497 + ], + "type": "text", + "content": "[62] Bowen Zhang, Yidong Wang, Wenxin Hou, HAO WU, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 18408-18419. Curran Associates, Inc., 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "text", + "content": "[63] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 534, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 577 + ], + "type": "text", + "content": "[64] Yuhang Zhang, Xiaopeng Zhang, Robert Qiu, Jie Li, Haohang Xu, Qi Tian, et al. Semi-supervised contrastive learning with similarity co-calibration. arXiv preprint arXiv:2105.07387, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "text", + "content": "[65] Mingkai Zheng, Fei Wang, Shan You, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Weakly supervised contrastive learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10042-10051, October 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[66] Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14471-14481, June 2022. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[67] Mingkai Zheng, Shan You, Fei Wang, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Ressl: Re" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 195 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "lational self-supervised learning with weak augmentation. arXiv preprint arXiv:2107.09282, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "text", + "content": "[68] Mingkai Zheng, Shan You, Fei Wang, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Relational self-supervised learning. arXiv preprint arXiv:2203.08717, 2022. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "text", + "content": "[69] Dengyong Zhou, Olivier Bousquet, Thomas Lal, Jason Weston, and Bernhard Scholkopf. Learning with local and global consistency. In S. Thrun, L. Saul, and B. Scholkopf, editors, Advances in Neural Information Processing Systems, volume 16. MIT Press, 2003. 4" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16442" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SimNP_ Learning Self-Similarity Priors Between Neural Points/54ca6d2a-b8a1-4a4e-a6f5-57eb0fddb4dd_content_list.json b/2023/SimNP_ Learning Self-Similarity Priors Between Neural Points/54ca6d2a-b8a1-4a4e-a6f5-57eb0fddb4dd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2becf29b5e65fdcc966e5b322f1652b9816ce507 --- /dev/null +++ b/2023/SimNP_ Learning Self-Similarity Priors Between Neural Points/54ca6d2a-b8a1-4a4e-a6f5-57eb0fddb4dd_content_list.json @@ -0,0 +1,1703 @@ +[ + { + "type": "text", + "text": "SimNP: Learning Self-Similarity Priors Between Neural Points", + "text_level": 1, + "bbox": [ + 166, + 130, + 803, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Christopher Wewer1", + "bbox": [ + 156, + 179, + 320, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Eddy $\\mathrm{Ilg}^2$", + "bbox": [ + 367, + 179, + 446, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bernt Schiele1,2", + "bbox": [ + 493, + 179, + 620, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jan Eric Lenssen1", + "bbox": [ + 666, + 180, + 808, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Max Planck Institute for Informatics, Saarland Informatics Campus, Germany $^{2}$ Saarland University, Saarland Informatics Campus, Germany", + "bbox": [ + 174, + 222, + 795, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{cwewer, jlenssen}@mpi-inf.mpg.de", + "bbox": [ + 339, + 262, + 630, + 277 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6ae94561a93fdad6372fa550a80ae09782a582fec02d56a82be7b638a045e63e.jpg", + "image_caption": [ + "a) Neural Point Representation" + ], + "image_footnote": [], + "bbox": [ + 81, + 303, + 367, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/906f0cc58cda70bd2c77410623bd608f83e22e95b70ee421d95a23d5c5c460c3.jpg", + "image_caption": [ + "PixelNeRF" + ], + "image_footnote": [], + "bbox": [ + 393, + 300, + 470, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/50a4fbe33bc150d6f501d2b2040a5a3d2bd49b9faf53e8a60b9fa2de4f142cb7.jpg", + "image_caption": [ + "VisionNeRF" + ], + "image_footnote": [], + "bbox": [ + 477, + 319, + 553, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b841b3a03b83ba82aaa2158df46baf78797f296174d388cfdcea767deea965bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 319, + 638, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d2bc123abd0238bf9879b8c04eb91f6a27a946eb81673f7cd1e1540d400a9a5.jpg", + "image_caption": [ + "b) Single-view reconstruction comparison.", + "Figure 1: a) We present SimNP, a renderable neural point radiance field that learns category-level self-similarities from data by connecting neural points to embeddings via optimized bipartite attention scores. b) The learned self-similarities can be used to transfer details from single- or few-view observations to unobserved, similar and symmetric parts of objects." + ], + "image_footnote": [], + "bbox": [ + 645, + 300, + 723, + 448 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ed172df9b9c52d7f7855c8d407b4ed81632708d29402dc18f60fbbaa4d04f4eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 300, + 807, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/10231a88a666945918f0114c81c85d4ce540834175d09f9f50a637462e306dcb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 300, + 883, + 439 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 539, + 313, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing neural field representations for 3D object reconstruction either (1) utilize object-level representations, but suffer from low-quality details due to conditioning on a global latent code, or (2) are able to perfectly reconstruct the observations, but fail to utilize object-level prior knowledge to infer unobserved regions. We present SimNP, a method to learn category-level self-similarities, which combines the advantages of both worlds by connecting neural point radiance fields with a category-level self-similarity representation. Our contribution is two-fold. (1) We design the first neural point representation on a category level by utilizing the concept of coherent point clouds. The resulting neural point radiance fields store a high level of detail for locally supported object regions. (2) We learn how information is shared between neural points in an unconstrained and unsupervised fashion, which allows to derive unobserved regions of an object during the reconstruction process from given observations. We show that SimNP is able to outperform previous methods in reconstructing symmetric unseen object regions, surpassing methods that build upon category-level or pixel-aligned radiance fields, while providing semantic correspondences between instances.", + "bbox": [ + 75, + 568, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 539, + 630, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The human visual system succeeds in deriving 3D representations of objects just from incomplete 2D observations. Key to this ability is that given observations are successfully complemented by previously learned information about the 3D world. Replicating this ability has been a longstanding goal in computer vision.", + "bbox": [ + 496, + 566, + 890, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since the task of reconstructing complete objects relies on generalization from a set of known examples, deep learning is an intuitive solution. The common approach is to use large amounts of data to train a category-level model [35, 41, 54, 29, 13, 26, 25, 22, 31, 9, 44, 57] and let the reconstruction process combine observations with the prior knowledge learned from data, which we refer to as the data prior. Notably, this introduces an inherent trade-off between contributions of the data prior and the observations.", + "bbox": [ + 496, + 657, + 892, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "On one extreme of the spectrum, NeRF-like methods [32, 55, 4] do not use a data prior at all. With a high number of observations and an optimization process, they are able to nearly perfectly reconstruct novel views of scenes and objects. However, this renders them incapable of deriving unseen regions. On the other extreme of the spectrum are methods that learn the full space of radiance or", + "bbox": [ + 496, + 795, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8841", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "signed distance functions belonging to a specific object category, such as SRN [41] and DeepSDF [35]. While these methods succeed in learning a complete representation of objects on an abstract level, they fail to represent individual details and the reconstruction process often performs retrieval from the learned data prior [42]. A similar behaviour has also been observed for generative models based on GANs [6] or diffusion [2], which yield visually impressive results but still diverge from given observations.", + "bbox": [ + 80, + 90, + 467, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key challenge lies in combining the strengths of the methods from both ends of the spectrum. While one can perform a highly detailed reconstruction of the visible regions, the object model should allow to reuse this information in unseen regions. To this end, it is important to know that most objects show many structured self-similarities, often arising from different types of symmetries, such as point/plane symmetries or more general variants. None of the current approaches try to explicitly learn such self-similarities to perform better inference.", + "bbox": [ + 80, + 227, + 467, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This is where SimNP comes in. We propose a better data prior vs. observation trade-off by combining the best of both worlds: (1) a category-level data prior encoding self-similarities on top of a (2) local representation with test-time optimization. Instead of learning the full space of radiance functions for a given category, we move to learning a data prior one level of abstraction higher, i.e., we learn how information can be shared between local object regions. This enables us to learn characteristic self-similarity patterns from training data, which are used to propagate information from visible to invisible parts during inference.", + "bbox": [ + 80, + 378, + 467, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As learning a representation for category-level self-similarities implies modeling relationships between local regions of objects, a key observation in this work is that neural point representations are especially well-suited to describe such relationships. Besides their capacity to capture high-frequency patterns, the underlying sparse point cloud allows for explicit formulations of similarities.", + "bbox": [ + 80, + 545, + 467, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the contributions of our work are:", + "bbox": [ + 98, + 650, + 403, + 662 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We present the first generalizable neural point radiance field for representation of objects categories.", + "2. We propose a simple but effective mechanism that learns general similarities between local object regions in an unconstrained and unsupervised fashion.", + "3. We show that our model improves upon state of the art in reconstructing unobserved regions from a single image and by outperforming existing two-view methods by a large margin. At the same time, it is more efficient in training and rendering." + ], + "bbox": [ + 91, + 672, + 465, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 80, + 845, + 215, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reconstruction from Observations Only While 3D reconstruction has traditionally been dominated by multi", + "bbox": [ + 80, + 871, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "stage pipelines [38, 39], NeRF-like approaches [32, 55, 4] revolutionized novel view synthesis by using a volumetric density representation in continuous 3D space. Although such approaches can achieve very accurate reconstructions, they only work for the visible regions and require a large number of input views.", + "bbox": [ + 503, + 90, + 888, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reconstruction with Data Priors To reduce the number of input views required for the reconstruction, many ways to leverage data priors have been proposed.", + "bbox": [ + 503, + 202, + 888, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pixel-Aligned methods such as PixelNeRF [54] leverage image-based rendering and use features that represent data priors obtained from CNNs [47, 45] or vision transformers [29] trained on large amounts of data. The feature for a 3D location in space is then derived from the associated local 2D image features. Methods specialized on multi-view input additionally utilize Multi-View Stereo (MVS) [7, 10, 23]. FE-NVS [17] uses an autoencoder-like architecture with a 2D and 3D U-Net. Voxel-Based approaches use an MLP taking in a local latent code [5, 30] to learn a data prior of geometry and appearance for a single voxel. Point-Based approaches [1, 37, 52] first obtain a point cloud from an RGB-D sensor or with MVS, and then initialize features for the points from CNNs similar as in image-based rendering. In contrast to the above, our method introduces a combined global and local representation. The local features of our approach can be optimized during test time and are propagated to unobserved regions via our learned category-specific attention.", + "bbox": [ + 503, + 247, + 888, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reconstruction with Object-Level Data Priors Early works for 3D reconstruction from single or few images leverage voxel grids [11, 16, 49]. However, while being able to reconstruct complete objects, for the latter it was shown that they do not actually perform reconstruction but image classification [42] and are therefore on the far end of data prior in the data prior vs. observation trade-off.", + "bbox": [ + 503, + 554, + 888, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some approaches model objects via 3D point clouds [13, 26, 25] but do not model surfaces and appearance. More recently, approaches that model a continuous representation of 3D space with MLPs were proposed. Scene Representation Networks [41] (SRN) model the scene as a continuous feature field in 3D space. Implicit representations of surfaces were introduced in DeepSDF [35] and Occupancy Networks [31, 9]. While these approaches can be extended to model the appearance on the surfaces [53, 34], recent approaches leverage neural radiance fields [22]. All of the mentioned models can be parameterized by a global latent code to model a distribution over objects as an object-level data prior. However, since prior and representation are global, the reconstructions usually lack details. In contrast, our method learns a global geometry prior with local appearance features enabling higher frequency details. We are", + "bbox": [ + 503, + 659, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8842", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c0214eb8a19189d53cb8d09089bb9d5c4e3bf77f03a800ec334743b321028648.jpg", + "image_caption": [ + "Figure 2: Overview of SimNP. Our method is a category-level, coherent neural point radiance field, where points are connected to embedding vectors $\\mathbf{E}$ via learnable attention scores $\\mathbf{A}$ . The representation can be rendered using ray marching and a neural renderer. (a) During training, all parameters $(\\square, \\square)$ are optimized using multi-view supervision. Networks, features $\\mathbf{S}$ , and scores are shared over the category $(\\square)$ , while embeddings are instance-specific $(\\square)$ . During inference, only embeddings $\\mathbf{E}$ $(\\square)$ are optimized from observations. In case of similar points $i, j$ (e.g., those shown in red), the network learned $a_{i,k} \\approx a_{j,k} \\forall k$ during training. Thus, supervision from one side means only one of points $i$ and $j$ needs to be visible to infer the value of embedding $k$ . (b) Given optimized embeddings, we can render the object from novel views." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 883, + 291 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the first to introduce a point-based neural radiance field representation on object level.", + "bbox": [ + 75, + 436, + 470, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Modeling Self-Similarity and Symmetry To our knowledge, learning general self-similarities of 3D structures is a novel concept and has not been explored before. However, many early works proposed to use or recover predefined symmetries to obtain improved reconstructions [43, 33, 14, 20, 15, 27, 3, 40, 36, 8], which can be seen as a specific instance of self-similarity. More recent work performs reconstruction of a single instance by jointly optimizing the position and orientation of a symmetry plane [21] or providing the symmetry as input [28]. Modeling plane or rotational symmetry is also common in object reconstruction in the wild [51, 50, 24, 46]. However, the constraints are only used for training and not for inference. In contrast to all of the above, SimNP learns arbitrary self-similarities from data without supervision.", + "bbox": [ + 75, + 484, + 468, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VisionNeRF [29] extends PixelNeRF [54] with a Vision Transformer (ViT) [12], which is able to globally propagate features among different rays. Therefore, one can argue that the transformer is able to learn category-level symmetries, but only implicitly, and only in 2D pixel space. In contrast, SimNP operates directly in 3D and learns local self-similarity relationships explicitly, which improves reconstructions of symmetric parts, as our results show.", + "bbox": [ + 75, + 710, + 468, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Self-Similarity Priors between Neural Points", + "text_level": 1, + "bbox": [ + 76, + 844, + 468, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present SimNP. An overview is shown in Figure 2. At the heart of our method is a neural point", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "representation that is comprised of a point cloud with attached feature vectors $\\mathcal{P} = (\\mathbf{P},\\mathbf{S},\\mathbf{F})$ with point positions $\\mathbf{P}\\in \\mathbb{R}^{N\\times 3}$ and two sets of point features $\\mathbf{S}\\in \\mathbb{R}^{N\\times D}$ and $\\mathbf{F}\\in \\mathbb{R}^{N\\times D}$ . The point features $\\mathbf{S}$ are shared across the whole category and encode a point identity, while features $\\mathbf{F}$ and positions $\\mathbf{P}$ are individual for each instance, encoding local density and radiance. Features $\\mathbf{F}$ are not explicitly stored but derived from embeddings $\\mathbf{E}$ (c.f. Sec. 3.1). Note that we require our point clouds to be coherent, meaning that over different instances of one category, a single point with index $i$ describes roughly the same part of the object, e.g., the area around the right rear mirror of a car.", + "bbox": [ + 496, + 436, + 893, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. The neural points are connected to a large set of embeddings $\\mathbf{E}$ via bipartite attention scores $\\mathbf{A}$ representing our category-level self-similarity, which is further detailed in Section 3.1. The neural point cloud can be volumetrically rendered from arbitrary views by using a network to decode color and density as described in Section 3.2. We employ the autodecoder framework [35], finding the optimal embeddings via optimization in training and inference, which is explained in Section 3.3. After assuming given point clouds for the introduction of our neural point radiance field, we cover the coherent point cloud prediction from single images independently in Section 3.4.", + "bbox": [ + 496, + 638, + 893, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Representing Category-Level Self-Similarity", + "text_level": 1, + "bbox": [ + 500, + 830, + 877, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We learn self-similarity by learning how information can be shared between coherent neural points. There are two characteristics why neural point clouds are a well-suited", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8843", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "representation for learning such explicit self-similarities: (1) They disentangle where the geometry is (represented by $\\mathbf{P}$ ) from how it looks (represented by $\\mathbf{F}$ ) and (2) they provide a discrete sampling of local parts that stays coherent across different instances due to this disentanglement. Since, with coherent point clouds, category-level self-similarities are invariant to variations in point location, similarities can be formulated on top of the neural point features $\\mathbf{F}$ , independent of individual instances of $\\mathbf{P}$ .", + "bbox": [ + 75, + 90, + 467, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, we store local density and radiance information in a larger number of embeddings $\\mathbf{E} \\in \\mathbb{R}^{M \\times D}$ and connect these to our neural points by", + "bbox": [ + 75, + 227, + 467, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} = \\operatorname {s o f t m a x} (\\mathbf {A}) \\cdot \\mathbf {E}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 285, + 467, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{A} \\in \\mathbb{R}^{N \\times M}$ are learnable attention scores between $N$ points and $M$ embeddings. The matrix $\\mathbf{A}$ is shared across the whole category and encodes the category-level self-similarity prior. It learns to connect two object points to the same embeddings, if they are similar and share information (see also Figure 2). Then, during reconstruction with single or few views, it is sufficient if an embedding receives gradients from only one of the connected points.", + "bbox": [ + 75, + 313, + 467, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Neural Point Rendering", + "text_level": 1, + "bbox": [ + 76, + 445, + 295, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rendering a neural point cloud $\\mathcal{P} = (\\mathbf{P},\\mathbf{S},\\mathbf{F})$ follows a ray casting approach with a single message passing step from neural points to ray samples, similar to that of Point-NeRF [52]. We cast rays through each pixel and sample points along the rays. An inherent advantage of point-based neural fields is that ray samples which are not in the vicinity of any neural points can be discarded before neural network application, which increases performance in contrast to dense approaches.", + "bbox": [ + 75, + 468, + 467, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The rendering network consists of two MLPs, a kernel MLP $K_{\\theta}$ describing local patches around each neural point, and a rendering MLP $F_{\\psi}$ producing density and radiance from aggregated features. Given a ray sample point $\\mathbf{x} \\in \\mathbb{R}^3$ , an intermediate feature vector $\\mathbf{h}$ is computed as", + "bbox": [ + 75, + 604, + 467, + 679 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} = \\frac {1}{W} \\sum_ {i \\in \\mathcal {N} (\\mathbf {x})} w (\\mathbf {x}, \\mathbf {p} _ {i}) \\cdot K _ {\\theta} (\\mathbf {x} - \\mathbf {p} _ {i}, \\mathbf {f} _ {i}, \\mathbf {s} _ {i}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 691, + 467, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{N}(\\mathbf{x})$ contains the indices of the $k$ -nearest neural points ( $k = 8$ in our case) of $\\mathbf{x}$ from $\\mathbf{P}$ within a radius $r$ . $w$ is a weight based on inverse point distance:", + "bbox": [ + 75, + 742, + 467, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw (\\mathbf {x}, \\mathbf {p} _ {i}) = \\frac {1}{\\left\\| \\mathbf {x} - \\mathbf {p} _ {i} \\right\\| _ {2}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 797, + 467, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $W$ is the sum of weights in the neighborhood. Then, the final radiance $\\mathbf{r}$ and density $\\sigma$ are obtained as", + "bbox": [ + 75, + 840, + 467, + 871 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n(\\mathbf {r}, \\sigma) = F _ {\\psi} (\\mathbf {h}, \\mathbf {d}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 885, + 467, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "based on features $\\mathbf{h}$ and view direction $\\mathbf{d} \\in \\mathbb{R}^3$ . For computing $\\mathcal{N}(\\mathbf{x})$ we created a custom PyTorch extension that implements an efficient, batch-wise voxel grid for ray sample $k$ -nn queries in CUDA. The module is based on the Point-NeRF code [52] and will be made publicly available with the rest of the code.", + "bbox": [ + 496, + 90, + 890, + 180 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that the rendering network is purely local, only receiving relative coordinates between ray samples and neural points. Thus, it is only able to learn a local surface model during training and no global, category-level information. The shared features $\\mathbf{S}$ are a key ingredient in this formulation. In our studies, we found that they help to train a high-quality category-level neural point renderer and that they represent a density/radiance template of the category. We refer to the supplemental material for further discussion.", + "bbox": [ + 496, + 183, + 890, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Training and Inference", + "text_level": 1, + "bbox": [ + 500, + 333, + 712, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For training and inference, we adopt the autodecoder framework [35, 5], thus, optimizing embeddings using gradient descent instead of predicting them using an encoder. Recent research has shown that approaches based on such test-time optimization have better capabilities when it comes to accurately representing the given observation [52].", + "bbox": [ + 496, + 357, + 890, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As input during training, we assume multi-view renderings of a large set of objects, including camera poses for each view. For simplicity, we omit camera poses from the following formulations. Let $\\mathbf{V} = f_{\\theta, \\psi}(\\mathbf{P}, \\mathbf{S}, \\mathbf{A}, \\mathbf{E})$ denote the function that renders an image $\\mathbf{V}$ from a neural point cloud with positions $\\mathbf{P}$ , embeddings $\\mathbf{E}$ , attention scores $\\mathbf{A}$ , and shared features $\\mathbf{S}$ from a given camera pose.", + "bbox": [ + 496, + 449, + 890, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Given a dataset $\\{(\\mathbf{P}_i, \\mathbf{I}_{i,j})\\}_{i = j = 1}^{K,W}$ of $K$ objects of one category with coherent point clouds $\\mathbf{P}_i$ , and $W$ renderings $\\{\\mathbf{I}_{i,j}\\}_{j = 1}^{W}$ from random views for each object $i$ , our training procedure jointly optimizes embeddings $\\mathbf{E}_i$ (omitted here for simplicity), attention scores $\\mathbf{A}$ , shared point features $\\mathbf{S}$ , and rendering network parameters $\\theta, \\psi$ :", + "bbox": [ + 496, + 579, + 890, + 672 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {A}}, \\hat {\\mathbf {S}}, \\hat {\\theta}, \\hat {\\psi} = \\underset {\\mathbf {A}, \\mathbf {S}, \\theta , \\psi} {\\arg \\min } \\sum_ {i, j} \\mathcal {L} \\left(f _ {\\theta , \\psi} \\left(\\mathbf {P} _ {i}, \\mathbf {S}, \\mathbf {A}, \\mathbf {E} _ {i}\\right), \\mathbf {I} _ {i, j}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 686, + 888, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}$ is chosen as Mean Squared Error (MSE). After training, we freeze point renderer parameters $(\\hat{\\theta},\\hat{\\psi})$ , shared features $\\hat{\\mathbf{S}}$ and category-level symmetries $\\hat{\\mathbf{A}}$ , which are used to guide gradients to the correct embeddings during test-time optimization.", + "bbox": [ + 496, + 734, + 890, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inference During inference, we assume that the camera pose is known. Given one or multiple views $\\{\\mathbf{I}_j\\}_{j=1}^W$ and the coherent point cloud $\\mathbf{P}$ , we can find embeddings $\\hat{\\mathbf{E}}$ by test-time optimization:", + "bbox": [ + 496, + 837, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8844", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0ef6299d438e900c485466324a21b1af2fb5fe4fbfad088716b12ea281bfee32.jpg", + "image_caption": [ + "Figure 3: Coherent point cloud prediction. An MLP with bottleneck is used to enforce the point cloud to be constructed as a low-rank deformation of a template (second to last layer output). During training, all trainable parameters $(\\square, \\square)$ are optimized using ground-truth point cloud supervision. During inference, the embedding $\\mathbf{z}$ $(\\square)$ can be optimized using different supervision signals: a ResNet predicting a point cloud from single image, mask, or depth." + ], + "image_footnote": [], + "bbox": [ + 78, + 90, + 890, + 220 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {E}} = \\underset {\\mathbf {E}} {\\arg \\min} \\sum_ {i} \\mathcal {L} \\left(f _ {\\hat {\\theta}, \\hat {\\psi}} (\\mathbf {P}, \\hat {\\mathbf {S}}, \\hat {\\mathbf {A}}, \\mathbf {E}), \\mathbf {I} _ {i}\\right). \\qquad (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 328, + 468, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Afterwards, the example can be rendered from other views.", + "bbox": [ + 76, + 366, + 468, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Coherent Point Clouds from Single Image", + "text_level": 1, + "bbox": [ + 76, + 388, + 434, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Until now, coherent point clouds were assumed as given for the sake of simplicity. Still, they also have to be obtained from a single image during test time and are a crucial part of our method. We separate the point cloud prediction from the rest of SimNP such that it can be tackled independently. Note that we do not optimize point positions based on rendering losses.", + "bbox": [ + 75, + 411, + 468, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The full architecture for point cloud prediction is shown in Figure 3. Just like for our rendering branch, we follow the autodecoder framework in order to allow for flexible supervision at test time. However, unlike the local neural point features for capturing fine details, we opt for a global representation of the point cloud in form of a single learnable latent code $\\mathbf{z} \\in \\mathbb{R}^l$ . During training, we optimize these embeddings jointly together with an MLP decoder $D_{\\phi}: \\mathbb{R}^l \\to \\mathbb{R}^{N \\times 3}$ predicting the neural point coordinates in canonical space, given by the normalized orientation and position of ShapeNet objects. The output of the second to last layer is a low-dimensional bottleneck in order to serve as a low-rank representation of a high-dimensional point template from $\\mathbb{R}^{N \\times 3}$ [18]. The global representation together with this low-rank regularization allows us to always obtain coherent point clouds irrespective of the final supervision signal.", + "bbox": [ + 75, + 517, + 470, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given point clouds $\\{\\mathcal{P}_i\\}_{i=1}^K$ of training examples, we train $D_{\\phi}$ and latent codes $\\{\\mathbf{z}_i\\}_{i=1}^K$ to minimize the 3D Chamfer Distance (CD):", + "bbox": [ + 75, + 773, + 472, + 819 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C D} = \\sum_ {i = 1} ^ {K} \\mathrm {C D} \\left(D _ {\\phi} \\left(\\mathbf {z} _ {i}\\right), \\mathcal {P} _ {i}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 824, + 468, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As one form of point cloud supervision during test time, we jointly train a ResNet18 [19] encoder with the image,", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "segmentation mask, and camera ray encodings [48] as input to predict the same point cloud after decoding with $D_{\\phi}$ , by minimizing the MSE. At test time, we optimize the latent codes while keeping encoder and decoder fixed. At first sight, it may seem counterintuitive to bring back an encoder into the autodecoder framework. However, by parameterizing the latent code directly, we gain flexibility regarding additional supervisory signals.", + "bbox": [ + 496, + 311, + 893, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Two optional sources of additional point cloud supervision at test time are the segmentation mask and a depth map, e.g., captured by an RGB-D camera. For the mask, we use a 2D Chamfer loss between its pixel coordinates and the 2D projection of the point cloud. Depth maps can be utilized by projecting the individual depth samples into the world frame and employing an asymmetric 3D Chamfer loss, i.e., minimizing the distance between each sample and its nearest point in the point cloud. Note that although occluded points from the perspective of the depth map are not explicitly supervised by this optional loss, they are still optimized implicitly because of our global latent representation.", + "bbox": [ + 496, + 431, + 893, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Besides leveraging these additional forms of data, test-time optimization of the point cloud is directly compatible with multi-view supervision. Given multiple views, the parameterized latent code $\\mathbf{z}$ fuses the point clouds predicted individually for each image by the encoder. The following experimental results demonstrate that our neural point representation can benefit significantly from flexible point cloud supervision at test time.", + "bbox": [ + 496, + 614, + 893, + 737 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments and Results", + "text_level": 1, + "bbox": [ + 500, + 753, + 735, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we describe experiments made with SimNP and present our results. The goal of the experiments is to provide evidence for the following statements: SimNP (1) improves on previous approaches in reconstructing unseen symmetric object parts (c.f. Section 4.2), (2) learns correct self-similarities that respect symmetries of the object category (c.f. Section 4.3), (3) provides a meaningful representation that can be used for interpolation", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8845", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ae63750fd7db16d1cc21b304f7e9081b2e2d81eba47c191230980a822694fcdf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Cat.Method1 Input View1 Input View (Sym.)2 Input Views
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
CarsSRN [41]22.220.8930.10722.050.8930.10624.820.9200.093
PixelNeRF [54]23.170.9050.11222.420.8950.12225.670.9350.083
CodeNeRF1 [22]23.800.910.106—1—1—125.710.930.099
FE-NVS2 [17]22.830.9060.099—2—2—224.640.9270.085
VisionNeRF [29]22.880.9080.08522.070.8980.091—3—3—3
Ours23.000.9110.08122.670.9100.08426.690.9480.055
Ours + GT PC23.690.9200.07723.310.9160.08027.290.9530.051
ChairsSRN [41]22.320.9040.09222.110.9010.09525.720.9350.072
PixelNeRF [54]23.720.9130.10123.150.9060.10926.210.9410.070
CodeNeRF1 [22]23.660.900.136—1—1—125.630.910.129
FE-NVS2 [17]23.210.9200.077—2—2—225.250.9400.065
VisionNeRF [29]24.480.9300.07723.880.9250.085—3—3—3
Ours23.460.9180.08123.030.9120.08827.360.9530.057
Ours + GT PC25.800.9400.07125.240.9340.07728.430.9610.053
", + "bbox": [ + 114, + 88, + 857, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: Novel view synthesis on ShapeNet cars and chairs. Our approach achieves state-of-the-art or competitive performance in single-view reconstruction and outperforms previous baselines by a large margin in the two-view setting. While PixelNeRF achieves a higher PSNR with single-view input, it can be seen that their results are blurry (c.f. Fig. 4a). Additionally to the standard setup, (Sym.) provides results on all target views that show mostly the opposite object side to the input view. We show results of our method using point cloud supervision (Ours + GT PC) at test time. These results are only to identify the gap induced by errors in point cloud prediction. The results shown in the rest of the paper use the Ours setup.", + "bbox": [ + 75, + 357, + 890, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(c.f. Section 4.4), (4) is a very efficient approach (c.f. Section 4.5), and (5) is compatible with test-time pose optimization (c.f. Section 4.6). We provide additional results in the appendix, such as more qualitative results, ablation studies, and an analysis of our point cloud prediction.", + "bbox": [ + 75, + 463, + 470, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 564, + 267, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments for category-specific novel view synthesis, given one or two input views. We make use of the ShapeNet dataset provided by SRN [41], which comprises 3514 cars and 6591 chairs split into training, validation, and test sets. Each training instance has been rendered from 50 random camera locations on a sphere around the object. The test sets have 704 and 1318 examples, respectively, all with the same 251 views from an Archimedean spiral and the same lighting as during training. As done by all baselines, view 64 and additionally view 104 are used as input for single- and two-view experiments. The image resolution is $128 \\times 128$ . We use PSNR, SSIM, and LPIPS [56] to compare our approach with SRN [41], PixelNeRF [54], FE-NVS [17], and VisionNeRF [29].", + "bbox": [ + 75, + 593, + 468, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use 512/4096 points and 512/128 32-dim. embeddings per object. The shared features are of dimensionality 64. After separate pretraining of the point cloud prediction, we train for $0.9 / 1.2\\mathrm{M}$ iterations. At test time, the instancespecific parameters are optimized for $10\\mathrm{k}$ iterations each. Please refer to the suppl. mat. for additional information.", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Single- and Two-View Reconstruction", + "text_level": 1, + "bbox": [ + 498, + 463, + 823, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 summarizes our quantitative results. We leverage three different test-time optimization setups. In our main setup (Ours), we use the ResNet and the mask for test time optimization of point clouds to ensure a fair comparison with previous approaches. This setup is used for all comparisons and qualitative results shown. Furthermore, we report results using ground-truth point clouds (Ours + GT PC) and provide results with additional ground-truth depth supervision in the supplement.", + "bbox": [ + 496, + 486, + 890, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In single-view reconstruction of cars, SimNP outperforms the state-of-the-art in SSIM and LPIPS. We attribute the small PSNR advantage of PixelNeRF to the blurriness of its reconstructions visible in Fig. 4a, which is confirmed by higher LPIPS. Also, we outperform PixelNeRF in inferring symmetric regions, as we show in columns 3-6 in Table 1 and in Figure 4b. SRN lacks details due to its global representation. PixelNeRF's reconstructions are overly-blurry in unseen regions, because pixel-aligned features are shared along the same rays of the input view without respecting the category-level self-similarities. VisionNeRF is not able to consistently transfer uncommon local color variations to symmetric object regions, which indicates a too strong", + "bbox": [ + 496, + 623, + 892, + 819 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{1}$ CodeNeRF [22] does not provide trained models. Note that the input views differ from the usual setup.", + "bbox": [ + 500, + 825, + 890, + 849 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2FE-NVS [17] does not have public code. Thus, shown numbers are taken from the paper/from authors, where available. SRN, PixelNeRF and VisionNeRF are evaluated using the PixelNeRF calc.metrics script.", + "bbox": [ + 500, + 849, + 890, + 887 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{3}$ VisionNeRF [29] not applicable for multiple input views.", + "bbox": [ + 517, + 887, + 826, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8846", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3cc050c691fc04980ca50010c96171c1d7f56792ae0c0fa8d1cf30516915fb95.jpg", + "image_caption": [ + "(a) Qualitative single-view reconstruction." + ], + "image_footnote": [], + "bbox": [ + 81, + 93, + 475, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec42be55d01d4a3b75a0611f326640cd852e0fcb1c37ed2e9a86ad09ecee9fd3.jpg", + "image_caption": [ + "(b) Metrics per view.", + "Figure 4: a) Our method enables more detailed reconstructions and can better transfer appearance information to symmetric regions compared to SRN, PixelNeRF, and VisionNeRF. b) We show metric comparisons for each target view in the 251-view SRN test spiral for single-view reconstruction of cars. As visible, our overall better performance can be attributed to views showing regions symmetric to the input view (green areas and example views 10, 84, 175). Also, the related object-level method SRN shows rather flat curves indicating a weak adaptation to observations. This is not the case for SimNP." + ], + "image_footnote": [], + "bbox": [ + 491, + 93, + 890, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "global prior. In contrast, as visible from Figure 4a and Figure 1, our approach correctly transfers local patterns by leveraging learned category-level self-similarities. As visible from Figure 4b and columns 3-6 in Table 1, it significantly outperforms all baselines for views on regions with symmetric counterparts in the input image.", + "bbox": [ + 75, + 580, + 468, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared to cars, the chairs category exhibits a larger variety in geometry but less texture details, leading to a smaller advantage through the self-similarity prior. Still, SimNP performs competitively compared to the baselines. Moreover, our strong results under the assumption of a given ground-truth point cloud (last rows in Table 1) indicate further potential of our method in combination with an improved point cloud prediction.", + "bbox": [ + 75, + 672, + 470, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the two-view setting, our approach surpasses previous methods by a large margin on both categories. Here, the benefit comes not only from modeling self-similarities but also from our local point-based representation. The results show that it is better suited for fusing multiple observations than pixel-aligned methods, while at the same time being able to fit details. Looking at the qualitative results in Fig. 5,", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we can observe detailed reconstructions. In Tab. 2, we compare different two view setup against each other. The same side setup, which does observe one side of the object with two views, already clearly improves the results.", + "bbox": [ + 496, + 580, + 892, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Learned Self-Similarities", + "text_level": 1, + "bbox": [ + 500, + 650, + 728, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As SimNP learns category-level self-similarities explicitly, we can directly inspect them to gain more insights. For visualizing the attention between neural points and embeddings, we proceed as follows. The rendering network computes weights between ray samples and neural points (see Eq. 3). By treating these weights just like RGB channels during ray marching, we obtain the influence of each neural point on each pixel. Finally, we multiply the learned category-level attention scores of a single embedding to the respective neural point influence, resulting in the self-similarity visualizations given in Figure 6. We describe the visualization procedure in detail in the supplemental material. Although we use as many embeddings as neural points (for cars) such that the attention could learn a one-to-one mapping, the representation learns to share information", + "bbox": [ + 496, + 672, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8847", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4f8dc54a567e14f3b72f4f983da2fcfd0cce4d4049f6394b2bbac0eb2d0c0f78.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Cat.View Idx.PSNR↑SSIM↑LPIPS↓
Carssingle view (64)23.000.9110.081
2 v. close (63,64)23.390.9160.078
2 v. same side (44,64)25.760.9380.064
2 v. opposite (64,104)26.690.9480.055
Chairssingle view (64)23.460.9180.081
2 v. close (63,64)24.240.9240.077
2 v. same side (44,64)26.500.9440.064
2 v. opposite (64,104)27.360.9530.057
", + "bbox": [ + 78, + 88, + 472, + 237 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2: Effect of input views. We evaluate different two view setups, comparing two very close views, 2 views from the same side of the object (approx. 90 degrees apart) and 2 views from opposite sides of the object. It can be seen that even if one side of the object is not seen (same side setup), the results improve strongly with 2 views.", + "bbox": [ + 76, + 246, + 470, + 338 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7054eeb93a0cb8380066b6cf641eeebf12c0ccb9ab315ab96635e769e2cc82e0.jpg", + "image_caption": [ + "Figure 5: Two-view reconstruction. SimNP learns a high-quality 3D object representation given only two input views." + ], + "image_footnote": [], + "bbox": [ + 81, + 347, + 460, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "between similar points. Further, we observe that the connections for all embeddings are symmetric with respect to a reflection plane, indicating that the method successfully learned a prior about the plane symmetry of each category.", + "bbox": [ + 75, + 756, + 468, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Meaningful Representation Space", + "text_level": 1, + "bbox": [ + 76, + 830, + 372, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 7 provides results for interpolating the point cloud latent code and/or the embeddings of two instances obtained by multi-view fitting. SimNP learns a meaningful, disentan-", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "RGB", + "bbox": [ + 529, + 92, + 557, + 103 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Attention Scores for 4 Embeddings", + "bbox": [ + 638, + 90, + 825, + 103 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5ed381c504ba0c5365fd8213255d3e202debec8aee012a6737fd473ad6ba919e.jpg", + "image_caption": [ + "Figure 6: Attention visualization. We render the influence of four different embeddings per category (one per column) on each ray, and visualize them on eight different examples. It becomes evident that embeddings specialize on self-similar local areas, e.g., wheels, or front lights. All embeddings show plane symmetric patterns, showing that our method recovers the symmetry of the category." + ], + "image_footnote": [], + "bbox": [ + 506, + 114, + 888, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gled representation space allowing for smooth shape and/or appearance transitions from one object to another. This is in contrast to pixel-aligned methods, which are not suited for interpolation. As an example, Fig. 8 shows a comparison with VisionNeRF, for which interpolation of the intermediate feature maps results in pixel space interpolation.", + "bbox": [ + 496, + 684, + 893, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Efficient Representation", + "text_level": 1, + "bbox": [ + 500, + 786, + 723, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SimNP renders a frame in 59ms, which is more than an order of magnitude faster than the pixel-aligned methods PixelNeRF (2.116s) and VisionNeRF (2.700s). We cannot fairly compare against FE-NVS, as no public code is available. The evaluation they provide in the paper is using amortized rendering, which is not comparable, as it only", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8848", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e84b33e4c40f11fbb435fdd7cc45fe4ccfc747846e90d21283f9dc319a2451dd.jpg", + "image_caption": [ + "Figure 7: Disentangled interpolation. Our point-based representation disentangles the basic shape from appearance. Point cloud latent code $\\mathbf{z}$ and embeddings $\\mathbf{E}$ can be interpolated independently resulting in a smooth shape and/or appearance transition between different objects." + ], + "image_footnote": [], + "bbox": [ + 81, + 99, + 460, + 391 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/dac93f30b97fd72df9e25c613a5b42d53f45afd434b5dbf5044661400cc0214c.jpg", + "image_caption": [ + "Figure 8: Interpolation comparison. Unlike VisionNeRF, which sticks to interpolation in pixel space for views similar to the input perspective, SimNP enables semantically meaningful interpolation like SRN but with more details." + ], + "image_footnote": [], + "bbox": [ + 84, + 503, + 467, + 569 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "works on many views simultaneously. However, we suspect that their method is also very fast. We train our method on a single A40 GPU for $\\approx 1.6$ days. In comparison, SRN and PixelNeRF take 6 days on an RTX 8000 or Titan RTX, respectively, FE-NVS 2.5 days on 30 V100, and VisionNeRF around 5 days on 16 A100 GPUs, according to the authors. Overall, our method is very efficient with resource-efficient training and fast rendering that shows potential to be applied to larger scale reconstruction in future work.", + "bbox": [ + 75, + 671, + 468, + 806 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.6. Pose Optimization", + "text_level": 1, + "bbox": [ + 76, + 816, + 253, + 832 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To alleviate the assumption of a known camera pose at test time, we investigate optimization of camera parameters. More precisely, for each test example, we initialize eight evenly distributed camera locations on the sphere around", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/16c4941e9dfbbd2c2214d8c208140d84c46f345575c75ee9218d630435d6243e.jpg", + "image_caption": [ + "Figure 9: Pose Optimization. SimNP is compatible with camera pose optimization. The figure compares single-view reconstruction results with and without given camera pose." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 874, + 236 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the object. The rotation matrix is always defined by the $z$ -axis pointing towards the sphere center and the $y$ -axis pointing upwards. At test time, we first initialize the point cloud latent using an encoder without ray encodings. Then, the pose in form of the camera location on the sphere is optimized by minimizing the Chamfer loss between the mask's pixel coordinates and the 2D projection of the point cloud. Lastly, we further finetune the point cloud with the Chamfer loss and optimize the embeddings for reconstruction of the input image. We select the pose that leads to the best reconstruction based on LPIPS against the single input image. Fig. 9 shows a comparison given vs. optimized camera pose. We achieve competitive performance (0.116 LPIPS) compared with CodeNeRF (0.114) without camera poses.", + "bbox": [ + 496, + 323, + 890, + 535 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 555, + 617, + 570 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We presented SimNP, the first object-level representation based on neural radiance fields, utilizing shared attention scores to learn category-specific self-similarities. Our method reaches state-of-the-art quality in single- and two-view reconstruction, while being highly efficient. Overall, it achieves a better data prior vs. observation trade-off. The improvements can be contributed to (1) leveraging a local neural point radiance field for the representation of object categories and (2) correctly propagating information between similar regions.", + "bbox": [ + 496, + 582, + 890, + 733 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations and Future Work The main limitation of SimNP is the assumption of a canonical space with ground-truth point clouds during training, which prohibits direct application on in-the-wild datasets. Therefore, an improved point cloud prediction in camera frame could be an exciting path for future work. Furthermore, we believe that it is promising to relax the point identities and to apply our self-similarity priors on a scene level to obtain large-scale, prior-driven reconstruction.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8849", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In European Conhference on Computer Vision (ECCV), 2020.", + "[2] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J. Mitra, and Paul Guerrero. RenderDiffusion: Image diffusion for 3D reconstruction, inpainting and generation. arXiv pre-print, 2022.", + "[3] François ARJ, Medioni GG, and Waupotitsch R. Mirror symmetry: 2-view stereo geometry. In Image and Vision Computing, 2003.", + "[4] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-NeRF: A multiscale representation for antiailiasing neural radiance fields. 2021.", + "[5] Rohan Chabra, Jan E. Lenssen, Eddy Ilg, Tanner Schmidt, Julian Straub, Steven Lovegrove, and Richard Newcombe. Deep local shapes: Learning local SDF priors for detailed 3D reconstruction. In European Conhference on Computer Vision (ECCV), 2020.", + "[6] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. Int. Conference on Computer Vision (ICCV), 2021.", + "[8] Xin Chen, Yuwei Li, Xi Luo, Tianjia Shao, Jingyi Yu, Kun Zhou, and Youyi Zheng. Autosweep: Recovering 3d editable objects from a single photograph. In ACM Trans. Gr., 2018.", + "[9] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019.", + "[10] Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (SRF): Learning view synthesis from sparse views of novel scenes. In Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2021.", + "[11] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3D-R2N2: A unified approach for single and multi-view 3d object reconstruction. In European Conhference on Computer Vision (ECCV), 2016.", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Int. Conference on Learning Representations (ICLR), 2021.", + "[13] Haoqiang Fan, Hao Su, and Leonidas Guibas. A point set generation network for 3d object reconstruction from a single image. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017." + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Roger Fawcett, Andrew Zisserman, and Michael Brady. Extracting structure from an affine view of a 3d point set with one or two bilateral symmetries. In Image and Vision Computing.", + "[15] Alexandre François, Gérard Medioni, and Roman Waupotitsch. Reconstructing mirror symmetric scenes from a single view using 2-view stereo geometry. In ICPR: Proceedings of the 16th International Conference on Pattern Recognition, 2002.", + "[16] Rohit Girdhar, David F. Fouhey, Mikel D. Rodriguez, and Abhinav Kumar Gupta. Learning a predictable and generative vector representation for objects. 2016.", + "[17] Pengsheng Guo, Miguel Angel Bautista, Alex Colburn, Liang Yang, Daniel Ulbricht, Joshua M. Susskind, and Qi Shan. Fast and explicit neural view synthesis. In Winter Conference on Applications of Computer Vision (WACV), 2022.", + "[18] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016.", + "[20] Du Q. Huynh. Affine reconstruction from monocular vision in the presence of a symmetry plane. In Int. Conference on Computer Vision (ICCV), 1999.", + "[21] Eldar Insafutdinov, Dylan Campbell, João F Henriques, and Andrea Vedaldi. SNeS: Learning probably symmetric neural surfaces from incomplete data. In European Conference on Computer Vision (ECCV), 2022.", + "[22] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Int. Conference on Computer Vision (ICCV), 2021.", + "[23] M. Johari, Y. Lepoittevin, and F. Fleuret. GeoNeRF: Generalizing nef with geometry priors. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[24] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In European Conhference on Computer Vision (ECCV), 2018.", + "[25] Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. In European Conhference on Computer Vision (ECCV), 2020.", + "[26] Andrey Kurenkov, Jingwei Ji, Animesh Garg, Viraj Mehta, JunYoung Gwak, Chris Choy, and Silvio Savarese. Deform-Net: Free-form deformation network for 3D shape reconstruction from a single image. 2018.", + "[27] Kevin Koser, Christopher Zach, and Marc Pollefeys. Dense 3d reconstruction of symmetric scenes from a single image. In Proceedings of the 33rd international conference on Pattern recognition, 2011.", + "[28] Xingyi Li, Chaoyi Hong, Yiran Wang, Zhiguo Cao, Ke Xian, and Guosheng Lin. Symmnerf: Learning to explore symmetry prior for single-view view synthesis. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 1726-1742, December 2022." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8850", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Kai-En Lin, Lin Yen-Chen, Wei-Sheng Lai, Tsung-Yi Lin, Yi-Chang Shih, and Ravi Ramamoorthi. Vision transformer for NeRF-based view synthesis from a single input image. In Winter Conference on Applications of Computer Vision (WACV), 2023.", + "[30] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In Int. Conference on Neural Information Processing Systems (NIPS), 2020.", + "[31] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019.", + "[32] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conhference on Computer Vision (ECCV), 2020.", + "[33] Andrew Zisserman Mukherjee, Dipti Prasad and Michael Brady. Shape from symmetry: Detecting and exploiting symmetry in affine images. In Philosophical Transactions: Physical Sciences and Engineering, 1995.", + "[34] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020.", + "[35] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019.", + "[36] Cody J. Phillips, Matthieu Lecce, and Kostas Daniilidis. Seeing glassware: from edge detection to pose estimation and shape recovery. In Robotics: Science and Systems, 2016.", + "[37] Ruslan Rakhimov, Andrei-Timotei Ardelean, Victor Lempitsky, and Evgeny Burnaev. NPBG++: Accelerating neural point-based graphics. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[38] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016.", + "[39] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016.", + "[40] Sudipta N. Sinha, Krishnan Ramnath, and Richard Szeliski. Detecting and reconstructing 3d mirror symmetric objects. In European Conhference on Computer Vision (ECCV), 2012.", + "[41] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3D-structure-aware neural scene representations. In Int. Conference on Neural Information Processing Systems (NIPS), 2019.", + "[42] Maxim Tatarchenko, Stephan R. Richter, Rene Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In Confer-" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ence on Computer Vision and Pattern Recognition (CVPR), 2019.", + "[43] Sebastian Thurn and Ben Wegbreit. Shape from symmetry. In Int. Conference on Computer Vision (ICCV), 2005.", + "[44] Garvita Tiwari, Dimitrije Antic, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Pose-ndf: Modeling human pose manifolds with neural distance fields. In European Conference on Computer Vision (ECCV), October 2022.", + "[45] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d scene representation and rendering. In Int. Conference on Computer Vision (ICCV), 2021.", + "[46] Shubham Tulsiani, Nilesh Kulkarni, and Abhinav Gupta. Implicit mesh reconstruction from unannotated image collections. In arXiv pre-print, 2020.", + "[47] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021.", + "[48] Daniel Watson, William Chan, Ricardo Martin Brulla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. In Int. Conference on Learning Representations (ICLR), 2023.", + "[49] Jiajun Wu, Chengkai Zhang, Tianfan Xue, William T. Freeman, and Joshua B. Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In Int. Conference on Neural Information Processing Systems (NIPS), 2016.", + "[50] Shangzhe Wu, Ameesh Makadia, Jiajun Wu, Noah Snavely, Richard Tucker, and Angjoo Kanazawa. De-rendering the world's revolutionary artefacts. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021.", + "[51] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020.", + "[52] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-NeRF: Point-based neural radiance fields. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[53] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In Int. Conference on Neural Information Processing Systems (NIPS), 2020.", + "[54] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021.", + "[55] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv pre-print, 2020.", + "[56] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Conference on Computer Vision and Pattern Recognition (CVPR), June 2018." + ], + "bbox": [ + 503, + 92, + 890, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8851", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 10 + }, + { + "type": "ref_text", + "text": "[57] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Toch: Spatio-temporal object-to-hand correspondence for motion refinement. In European Conference on Computer Vision (ECCV). Springer, October 2022.", + "bbox": [ + 78, + 90, + 470, + 148 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "8852", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/SimNP_ Learning Self-Similarity Priors Between Neural Points/54ca6d2a-b8a1-4a4e-a6f5-57eb0fddb4dd_model.json b/2023/SimNP_ Learning Self-Similarity Priors Between Neural Points/54ca6d2a-b8a1-4a4e-a6f5-57eb0fddb4dd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7bbe41e38cda1891db00c8a7e8f36d56673d949a --- /dev/null +++ b/2023/SimNP_ Learning Self-Similarity Priors Between Neural Points/54ca6d2a-b8a1-4a4e-a6f5-57eb0fddb4dd_model.json @@ -0,0 +1,2402 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.167, + 0.131, + 0.804, + 0.153 + ], + "angle": 0, + "content": "SimNP: Learning Self-Similarity Priors Between Neural Points" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.18, + 0.321, + 0.2 + ], + "angle": 0, + "content": "Christopher Wewer1" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.18, + 0.447, + 0.2 + ], + "angle": 0, + "content": "Eddy \\(\\mathrm{Ilg}^2\\)" + }, + { + "type": "text", + "bbox": [ + 0.495, + 0.18, + 0.621, + 0.198 + ], + "angle": 0, + "content": "Bernt Schiele1,2" + }, + { + "type": "text", + "bbox": [ + 0.668, + 0.181, + 0.81, + 0.198 + ], + "angle": 0, + "content": "Jan Eric Lenssen1" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.223, + 0.796, + 0.26 + ], + "angle": 0, + "content": "\\(^{1}\\)Max Planck Institute for Informatics, Saarland Informatics Campus, Germany \\(^{2}\\)Saarland University, Saarland Informatics Campus, Germany" + }, + { + "type": "text", + "bbox": [ + 0.34, + 0.263, + 0.631, + 0.278 + ], + "angle": 0, + "content": "{cwewer, jlenssen}@mpi-inf.mpg.de" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.304, + 0.368, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.453, + 0.333, + 0.468 + ], + "angle": 0, + "content": "a) Neural Point Representation" + }, + { + "type": "image", + "bbox": [ + 0.394, + 0.301, + 0.471, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.478, + 0.301, + 0.554, + 0.313 + ], + "angle": 0, + "content": "PixelNeRF" + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.32, + 0.555, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.302, + 0.637, + 0.313 + ], + "angle": 0, + "content": "VisionNeRF" + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.32, + 0.64, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.301, + 0.724, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.301, + 0.808, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.301, + 0.884, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.453, + 0.778, + 0.469 + ], + "angle": 0, + "content": "b) Single-view reconstruction comparison." + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.479, + 0.893, + 0.525 + ], + "angle": 0, + "content": "Figure 1: a) We present SimNP, a renderable neural point radiance field that learns category-level self-similarities from data by connecting neural points to embeddings via optimized bipartite attention scores. b) The learned self-similarities can be used to transfer details from single- or few-view observations to unobserved, similar and symmetric parts of objects." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.54, + 0.314, + 0.555 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Existing neural field representations for 3D object reconstruction either (1) utilize object-level representations, but suffer from low-quality details due to conditioning on a global latent code, or (2) are able to perfectly reconstruct the observations, but fail to utilize object-level prior knowledge to infer unobserved regions. We present SimNP, a method to learn category-level self-similarities, which combines the advantages of both worlds by connecting neural point radiance fields with a category-level self-similarity representation. Our contribution is two-fold. (1) We design the first neural point representation on a category level by utilizing the concept of coherent point clouds. The resulting neural point radiance fields store a high level of detail for locally supported object regions. (2) We learn how information is shared between neural points in an unconstrained and unsupervised fashion, which allows to derive unobserved regions of an object during the reconstruction process from given observations. We show that SimNP is able to outperform previous methods in reconstructing symmetric unseen object regions, surpassing methods that build upon category-level or pixel-aligned radiance fields, while providing semantic correspondences between instances." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.54, + 0.631, + 0.555 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.657 + ], + "angle": 0, + "content": "The human visual system succeeds in deriving 3D representations of objects just from incomplete 2D observations. Key to this ability is that given observations are successfully complemented by previously learned information about the 3D world. Replicating this ability has been a longstanding goal in computer vision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.893, + 0.794 + ], + "angle": 0, + "content": "Since the task of reconstructing complete objects relies on generalization from a set of known examples, deep learning is an intuitive solution. The common approach is to use large amounts of data to train a category-level model [35, 41, 54, 29, 13, 26, 25, 22, 31, 9, 44, 57] and let the reconstruction process combine observations with the prior knowledge learned from data, which we refer to as the data prior. Notably, this introduces an inherent trade-off between contributions of the data prior and the observations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.901 + ], + "angle": 0, + "content": "On one extreme of the spectrum, NeRF-like methods [32, 55, 4] do not use a data prior at all. With a high number of observations and an optimization process, they are able to nearly perfectly reconstruct novel views of scenes and objects. However, this renders them incapable of deriving unseen regions. On the other extreme of the spectrum are methods that learn the full space of radiance or" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8841" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.226 + ], + "angle": 0, + "content": "signed distance functions belonging to a specific object category, such as SRN [41] and DeepSDF [35]. While these methods succeed in learning a complete representation of objects on an abstract level, they fail to represent individual details and the reconstruction process often performs retrieval from the learned data prior [42]. A similar behaviour has also been observed for generative models based on GANs [6] or diffusion [2], which yield visually impressive results but still diverge from given observations." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.228, + 0.468, + 0.377 + ], + "angle": 0, + "content": "The key challenge lies in combining the strengths of the methods from both ends of the spectrum. While one can perform a highly detailed reconstruction of the visible regions, the object model should allow to reuse this information in unseen regions. To this end, it is important to know that most objects show many structured self-similarities, often arising from different types of symmetries, such as point/plane symmetries or more general variants. None of the current approaches try to explicitly learn such self-similarities to perform better inference." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.379, + 0.468, + 0.544 + ], + "angle": 0, + "content": "This is where SimNP comes in. We propose a better data prior vs. observation trade-off by combining the best of both worlds: (1) a category-level data prior encoding self-similarities on top of a (2) local representation with test-time optimization. Instead of learning the full space of radiance functions for a given category, we move to learning a data prior one level of abstraction higher, i.e., we learn how information can be shared between local object regions. This enables us to learn characteristic self-similarity patterns from training data, which are used to propagate information from visible to invisible parts during inference." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.546, + 0.468, + 0.649 + ], + "angle": 0, + "content": "As learning a representation for category-level self-similarities implies modeling relationships between local regions of objects, a key observation in this work is that neural point representations are especially well-suited to describe such relationships. Besides their capacity to capture high-frequency patterns, the underlying sparse point cloud allows for explicit formulations of similarities." + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.651, + 0.405, + 0.664 + ], + "angle": 0, + "content": "In summary, the contributions of our work are:" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.673, + 0.467, + 0.702 + ], + "angle": 0, + "content": "1. We present the first generalizable neural point radiance field for representation of objects categories." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.708, + 0.467, + 0.75 + ], + "angle": 0, + "content": "2. We propose a simple but effective mechanism that learns general similarities between local object regions in an unconstrained and unsupervised fashion." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.758, + 0.467, + 0.831 + ], + "angle": 0, + "content": "3. We show that our model improves upon state of the art in reconstructing unobserved regions from a single image and by outperforming existing two-view methods by a large margin. At the same time, it is more efficient in training and rendering." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.673, + 0.467, + 0.831 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.846, + 0.217, + 0.86 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.872, + 0.468, + 0.899 + ], + "angle": 0, + "content": "Reconstruction from Observations Only While 3D reconstruction has traditionally been dominated by multi" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.89, + 0.181 + ], + "angle": 0, + "content": "stage pipelines [38, 39], NeRF-like approaches [32, 55, 4] revolutionized novel view synthesis by using a volumetric density representation in continuous 3D space. Although such approaches can achieve very accurate reconstructions, they only work for the visible regions and require a large number of input views." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.203, + 0.89, + 0.246 + ], + "angle": 0, + "content": "Reconstruction with Data Priors To reduce the number of input views required for the reconstruction, many ways to leverage data priors have been proposed." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.248, + 0.89, + 0.533 + ], + "angle": 0, + "content": "Pixel-Aligned methods such as PixelNeRF [54] leverage image-based rendering and use features that represent data priors obtained from CNNs [47, 45] or vision transformers [29] trained on large amounts of data. The feature for a 3D location in space is then derived from the associated local 2D image features. Methods specialized on multi-view input additionally utilize Multi-View Stereo (MVS) [7, 10, 23]. FE-NVS [17] uses an autoencoder-like architecture with a 2D and 3D U-Net. Voxel-Based approaches use an MLP taking in a local latent code [5, 30] to learn a data prior of geometry and appearance for a single voxel. Point-Based approaches [1, 37, 52] first obtain a point cloud from an RGB-D sensor or with MVS, and then initialize features for the points from CNNs similar as in image-based rendering. In contrast to the above, our method introduces a combined global and local representation. The local features of our approach can be optimized during test time and are propagated to unobserved regions via our learned category-specific attention." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.555, + 0.89, + 0.658 + ], + "angle": 0, + "content": "Reconstruction with Object-Level Data Priors Early works for 3D reconstruction from single or few images leverage voxel grids [11, 16, 49]. However, while being able to reconstruct complete objects, for the latter it was shown that they do not actually perform reconstruction but image classification [42] and are therefore on the far end of data prior in the data prior vs. observation trade-off." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.66, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Some approaches model objects via 3D point clouds [13, 26, 25] but do not model surfaces and appearance. More recently, approaches that model a continuous representation of 3D space with MLPs were proposed. Scene Representation Networks [41] (SRN) model the scene as a continuous feature field in 3D space. Implicit representations of surfaces were introduced in DeepSDF [35] and Occupancy Networks [31, 9]. While these approaches can be extended to model the appearance on the surfaces [53, 34], recent approaches leverage neural radiance fields [22]. All of the mentioned models can be parameterized by a global latent code to model a distribution over objects as an object-level data prior. However, since prior and representation are global, the reconstructions usually lack details. In contrast, our method learns a global geometry prior with local appearance features enabling higher frequency details. We are" + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.956 + ], + "angle": 0, + "content": "8842" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.09, + 0.885, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.305, + 0.896, + 0.414 + ], + "angle": 0, + "content": "Figure 2: Overview of SimNP. Our method is a category-level, coherent neural point radiance field, where points are connected to embedding vectors \\(\\mathbf{E}\\) via learnable attention scores \\(\\mathbf{A}\\). The representation can be rendered using ray marching and a neural renderer. (a) During training, all parameters \\((\\square, \\square)\\) are optimized using multi-view supervision. Networks, features \\(\\mathbf{S}\\), and scores are shared over the category \\((\\square)\\), while embeddings are instance-specific \\((\\square)\\). During inference, only embeddings \\(\\mathbf{E}\\) \\((\\square)\\) are optimized from observations. In case of similar points \\(i, j\\) (e.g., those shown in red), the network learned \\(a_{i,k} \\approx a_{j,k} \\forall k\\) during training. Thus, supervision from one side means only one of points \\(i\\) and \\(j\\) needs to be visible to infer the value of embedding \\(k\\). (b) Given optimized embeddings, we can render the object from novel views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.438, + 0.471, + 0.469 + ], + "angle": 0, + "content": "the first to introduce a point-based neural radiance field representation on object level." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.47, + 0.712 + ], + "angle": 0, + "content": "Modeling Self-Similarity and Symmetry To our knowledge, learning general self-similarities of 3D structures is a novel concept and has not been explored before. However, many early works proposed to use or recover predefined symmetries to obtain improved reconstructions [43, 33, 14, 20, 15, 27, 3, 40, 36, 8], which can be seen as a specific instance of self-similarity. More recent work performs reconstruction of a single instance by jointly optimizing the position and orientation of a symmetry plane [21] or providing the symmetry as input [28]. Modeling plane or rotational symmetry is also common in object reconstruction in the wild [51, 50, 24, 46]. However, the constraints are only used for training and not for inference. In contrast to all of the above, SimNP learns arbitrary self-similarities from data without supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.47, + 0.834 + ], + "angle": 0, + "content": "VisionNeRF [29] extends PixelNeRF [54] with a Vision Transformer (ViT) [12], which is able to globally propagate features among different rays. Therefore, one can argue that the transformer is able to learn category-level symmetries, but only implicitly, and only in 2D pixel space. In contrast, SimNP operates directly in 3D and learns local self-similarity relationships explicitly, which improves reconstructions of symmetric parts, as our results show." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.47, + 0.863 + ], + "angle": 0, + "content": "3. Self-Similarity Priors between Neural Points" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this section, we present SimNP. An overview is shown in Figure 2. At the heart of our method is a neural point" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.438, + 0.895, + 0.621 + ], + "angle": 0, + "content": "representation that is comprised of a point cloud with attached feature vectors \\(\\mathcal{P} = (\\mathbf{P},\\mathbf{S},\\mathbf{F})\\) with point positions \\(\\mathbf{P}\\in \\mathbb{R}^{N\\times 3}\\) and two sets of point features \\(\\mathbf{S}\\in \\mathbb{R}^{N\\times D}\\) and \\(\\mathbf{F}\\in \\mathbb{R}^{N\\times D}\\). The point features \\(\\mathbf{S}\\) are shared across the whole category and encode a point identity, while features \\(\\mathbf{F}\\) and positions \\(\\mathbf{P}\\) are individual for each instance, encoding local density and radiance. Features \\(\\mathbf{F}\\) are not explicitly stored but derived from embeddings \\(\\mathbf{E}\\) (c.f. Sec. 3.1). Note that we require our point clouds to be coherent, meaning that over different instances of one category, a single point with index \\(i\\) describes roughly the same part of the object, e.g., the area around the right rear mirror of a car." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.895, + 0.822 + ], + "angle": 0, + "content": "Overview. The neural points are connected to a large set of embeddings \\(\\mathbf{E}\\) via bipartite attention scores \\(\\mathbf{A}\\) representing our category-level self-similarity, which is further detailed in Section 3.1. The neural point cloud can be volumetrically rendered from arbitrary views by using a network to decode color and density as described in Section 3.2. We employ the autodecoder framework [35], finding the optimal embeddings via optimization in training and inference, which is explained in Section 3.3. After assuming given point clouds for the introduction of our neural point radiance field, we cover the coherent point cloud prediction from single images independently in Section 3.4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.878, + 0.849 + ], + "angle": 0, + "content": "3.1. Representing Category-Level Self-Similarity" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We learn self-similarity by learning how information can be shared between coherent neural points. There are two characteristics why neural point clouds are a well-suited" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8843" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.226 + ], + "angle": 0, + "content": "representation for learning such explicit self-similarities: (1) They disentangle where the geometry is (represented by \\(\\mathbf{P}\\)) from how it looks (represented by \\(\\mathbf{F}\\)) and (2) they provide a discrete sampling of local parts that stays coherent across different instances due to this disentanglement. Since, with coherent point clouds, category-level self-similarities are invariant to variations in point location, similarities can be formulated on top of the neural point features \\(\\mathbf{F}\\), independent of individual instances of \\(\\mathbf{P}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.468, + 0.273 + ], + "angle": 0, + "content": "Formally, we store local density and radiance information in a larger number of embeddings \\(\\mathbf{E} \\in \\mathbb{R}^{M \\times D}\\) and connect these to our neural points by" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.286, + 0.468, + 0.302 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} = \\operatorname {s o f t m a x} (\\mathbf {A}) \\cdot \\mathbf {E}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.468, + 0.435 + ], + "angle": 0, + "content": "where \\(\\mathbf{A} \\in \\mathbb{R}^{N \\times M}\\) are learnable attention scores between \\(N\\) points and \\(M\\) embeddings. The matrix \\(\\mathbf{A}\\) is shared across the whole category and encodes the category-level self-similarity prior. It learns to connect two object points to the same embeddings, if they are similar and share information (see also Figure 2). Then, during reconstruction with single or few views, it is sufficient if an embedding receives gradients from only one of the connected points." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.446, + 0.296, + 0.461 + ], + "angle": 0, + "content": "3.2. Neural Point Rendering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.468, + 0.605 + ], + "angle": 0, + "content": "Rendering a neural point cloud \\(\\mathcal{P} = (\\mathbf{P},\\mathbf{S},\\mathbf{F})\\) follows a ray casting approach with a single message passing step from neural points to ray samples, similar to that of Point-NeRF [52]. We cast rays through each pixel and sample points along the rays. An inherent advantage of point-based neural fields is that ray samples which are not in the vicinity of any neural points can be discarded before neural network application, which increases performance in contrast to dense approaches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.468, + 0.68 + ], + "angle": 0, + "content": "The rendering network consists of two MLPs, a kernel MLP \\(K_{\\theta}\\) describing local patches around each neural point, and a rendering MLP \\(F_{\\psi}\\) producing density and radiance from aggregated features. Given a ray sample point \\(\\mathbf{x} \\in \\mathbb{R}^3\\), an intermediate feature vector \\(\\mathbf{h}\\) is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.692, + 0.468, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} = \\frac {1}{W} \\sum_ {i \\in \\mathcal {N} (\\mathbf {x})} w (\\mathbf {x}, \\mathbf {p} _ {i}) \\cdot K _ {\\theta} (\\mathbf {x} - \\mathbf {p} _ {i}, \\mathbf {f} _ {i}, \\mathbf {s} _ {i}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.743, + 0.468, + 0.788 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}(\\mathbf{x})\\) contains the indices of the \\(k\\)-nearest neural points (\\(k = 8\\) in our case) of \\(\\mathbf{x}\\) from \\(\\mathbf{P}\\) within a radius \\(r\\). \\(w\\) is a weight based on inverse point distance:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.799, + 0.468, + 0.832 + ], + "angle": 0, + "content": "\\[\nw (\\mathbf {x}, \\mathbf {p} _ {i}) = \\frac {1}{\\left\\| \\mathbf {x} - \\mathbf {p} _ {i} \\right\\| _ {2}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.842, + 0.468, + 0.872 + ], + "angle": 0, + "content": "and \\(W\\) is the sum of weights in the neighborhood. Then, the final radiance \\(\\mathbf{r}\\) and density \\(\\sigma\\) are obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.886, + 0.468, + 0.902 + ], + "angle": 0, + "content": "\\[\n(\\mathbf {r}, \\sigma) = F _ {\\psi} (\\mathbf {h}, \\mathbf {d}) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.891, + 0.181 + ], + "angle": 0, + "content": "based on features \\(\\mathbf{h}\\) and view direction \\(\\mathbf{d} \\in \\mathbb{R}^3\\). For computing \\(\\mathcal{N}(\\mathbf{x})\\) we created a custom PyTorch extension that implements an efficient, batch-wise voxel grid for ray sample \\(k\\)-nn queries in CUDA. The module is based on the Point-NeRF code [52] and will be made publicly available with the rest of the code." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.184, + 0.892, + 0.319 + ], + "angle": 0, + "content": "Note that the rendering network is purely local, only receiving relative coordinates between ray samples and neural points. Thus, it is only able to learn a local surface model during training and no global, category-level information. The shared features \\(\\mathbf{S}\\) are a key ingredient in this formulation. In our studies, we found that they help to train a high-quality category-level neural point renderer and that they represent a density/radiance template of the category. We refer to the supplemental material for further discussion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.334, + 0.713, + 0.349 + ], + "angle": 0, + "content": "3.3. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.892, + 0.449 + ], + "angle": 0, + "content": "For training and inference, we adopt the autodecoder framework [35, 5], thus, optimizing embeddings using gradient descent instead of predicting them using an encoder. Recent research has shown that approaches based on such test-time optimization have better capabilities when it comes to accurately representing the given observation [52]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.892, + 0.555 + ], + "angle": 0, + "content": "As input during training, we assume multi-view renderings of a large set of objects, including camera poses for each view. For simplicity, we omit camera poses from the following formulations. Let \\(\\mathbf{V} = f_{\\theta, \\psi}(\\mathbf{P}, \\mathbf{S}, \\mathbf{A}, \\mathbf{E})\\) denote the function that renders an image \\(\\mathbf{V}\\) from a neural point cloud with positions \\(\\mathbf{P}\\), embeddings \\(\\mathbf{E}\\), attention scores \\(\\mathbf{A}\\), and shared features \\(\\mathbf{S}\\) from a given camera pose." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.891, + 0.673 + ], + "angle": 0, + "content": "Training Given a dataset \\(\\{(\\mathbf{P}_i, \\mathbf{I}_{i,j})\\}_{i = j = 1}^{K,W}\\) of \\(K\\) objects of one category with coherent point clouds \\(\\mathbf{P}_i\\), and \\(W\\) renderings \\(\\{\\mathbf{I}_{i,j}\\}_{j = 1}^{W}\\) from random views for each object \\(i\\), our training procedure jointly optimizes embeddings \\(\\mathbf{E}_i\\) (omitted here for simplicity), attention scores \\(\\mathbf{A}\\), shared point features \\(\\mathbf{S}\\), and rendering network parameters \\(\\theta, \\psi\\):" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.687, + 0.89, + 0.733 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {A}}, \\hat {\\mathbf {S}}, \\hat {\\theta}, \\hat {\\psi} = \\underset {\\mathbf {A}, \\mathbf {S}, \\theta , \\psi} {\\arg \\min } \\sum_ {i, j} \\mathcal {L} \\left(f _ {\\theta , \\psi} \\left(\\mathbf {P} _ {i}, \\mathbf {S}, \\mathbf {A}, \\mathbf {E} _ {i}\\right), \\mathbf {I} _ {i, j}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.81 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}\\) is chosen as Mean Squared Error (MSE). After training, we freeze point renderer parameters \\((\\hat{\\theta},\\hat{\\psi})\\), shared features \\(\\hat{\\mathbf{S}}\\) and category-level symmetries \\(\\hat{\\mathbf{A}}\\), which are used to guide gradients to the correct embeddings during test-time optimization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.838, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Inference During inference, we assume that the camera pose is known. Given one or multiple views \\(\\{\\mathbf{I}_j\\}_{j=1}^W\\) and the coherent point cloud \\(\\mathbf{P}\\), we can find embeddings \\(\\hat{\\mathbf{E}}\\) by test-time optimization:" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.956 + ], + "angle": 0, + "content": "8844" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.092, + 0.891, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.231, + 0.897, + 0.294 + ], + "angle": 0, + "content": "Figure 3: Coherent point cloud prediction. An MLP with bottleneck is used to enforce the point cloud to be constructed as a low-rank deformation of a template (second to last layer output). During training, all trainable parameters \\((\\square, \\square)\\) are optimized using ground-truth point cloud supervision. During inference, the embedding \\(\\mathbf{z}\\) \\((\\square)\\) can be optimized using different supervision signals: a ResNet predicting a point cloud from single image, mask, or depth." + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.329, + 0.47, + 0.361 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {E}} = \\underset {\\mathbf {E}} {\\arg \\min} \\sum_ {i} \\mathcal {L} \\left(f _ {\\hat {\\theta}, \\hat {\\psi}} (\\mathbf {P}, \\hat {\\mathbf {S}}, \\hat {\\mathbf {A}}, \\mathbf {E}), \\mathbf {I} _ {i}\\right). \\qquad (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.367, + 0.47, + 0.382 + ], + "angle": 0, + "content": "Afterwards, the example can be rendered from other views." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.389, + 0.436, + 0.405 + ], + "angle": 0, + "content": "3.4. Coherent Point Clouds from Single Image" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.47, + 0.517 + ], + "angle": 0, + "content": "Until now, coherent point clouds were assumed as given for the sake of simplicity. Still, they also have to be obtained from a single image during test time and are a crucial part of our method. We separate the point cloud prediction from the rest of SimNP such that it can be tackled independently. Note that we do not optimize point positions based on rendering losses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.471, + 0.775 + ], + "angle": 0, + "content": "The full architecture for point cloud prediction is shown in Figure 3. Just like for our rendering branch, we follow the autodecoder framework in order to allow for flexible supervision at test time. However, unlike the local neural point features for capturing fine details, we opt for a global representation of the point cloud in form of a single learnable latent code \\(\\mathbf{z} \\in \\mathbb{R}^l\\). During training, we optimize these embeddings jointly together with an MLP decoder \\(D_{\\phi}: \\mathbb{R}^l \\to \\mathbb{R}^{N \\times 3}\\) predicting the neural point coordinates in canonical space, given by the normalized orientation and position of ShapeNet objects. The output of the second to last layer is a low-dimensional bottleneck in order to serve as a low-rank representation of a high-dimensional point template from \\(\\mathbb{R}^{N \\times 3}\\) [18]. The global representation together with this low-rank regularization allows us to always obtain coherent point clouds irrespective of the final supervision signal." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.774, + 0.473, + 0.82 + ], + "angle": 0, + "content": "Given point clouds \\(\\{\\mathcal{P}_i\\}_{i=1}^K\\) of training examples, we train \\(D_{\\phi}\\) and latent codes \\(\\{\\mathbf{z}_i\\}_{i=1}^K\\) to minimize the 3D Chamfer Distance (CD):" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.825, + 0.47, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C D} = \\sum_ {i = 1} ^ {K} \\mathrm {C D} \\left(D _ {\\phi} \\left(\\mathbf {z} _ {i}\\right), \\mathcal {P} _ {i}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "As one form of point cloud supervision during test time, we jointly train a ResNet18 [19] encoder with the image," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.895, + 0.433 + ], + "angle": 0, + "content": "segmentation mask, and camera ray encodings [48] as input to predict the same point cloud after decoding with \\( D_{\\phi} \\), by minimizing the MSE. At test time, we optimize the latent codes while keeping encoder and decoder fixed. At first sight, it may seem counterintuitive to bring back an encoder into the autodecoder framework. However, by parameterizing the latent code directly, we gain flexibility regarding additional supervisory signals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.895, + 0.617 + ], + "angle": 0, + "content": "Two optional sources of additional point cloud supervision at test time are the segmentation mask and a depth map, e.g., captured by an RGB-D camera. For the mask, we use a 2D Chamfer loss between its pixel coordinates and the 2D projection of the point cloud. Depth maps can be utilized by projecting the individual depth samples into the world frame and employing an asymmetric 3D Chamfer loss, i.e., minimizing the distance between each sample and its nearest point in the point cloud. Note that although occluded points from the perspective of the depth map are not explicitly supervised by this optional loss, they are still optimized implicitly because of our global latent representation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.616, + 0.895, + 0.738 + ], + "angle": 0, + "content": "Besides leveraging these additional forms of data, test-time optimization of the point cloud is directly compatible with multi-view supervision. Given multiple views, the parameterized latent code \\(\\mathbf{z}\\) fuses the point clouds predicted individually for each image by the encoder. The following experimental results demonstrate that our neural point representation can benefit significantly from flexible point cloud supervision at test time." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.754, + 0.736, + 0.772 + ], + "angle": 0, + "content": "4. Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this section, we describe experiments made with SimNP and present our results. The goal of the experiments is to provide evidence for the following statements: SimNP (1) improves on previous approaches in reconstructing unseen symmetric object parts (c.f. Section 4.2), (2) learns correct self-similarities that respect symmetries of the object category (c.f. Section 4.3), (3) provides a meaningful representation that can be used for interpolation" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8845" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.859, + 0.348 + ], + "angle": 0, + "content": "
Cat.Method1 Input View1 Input View (Sym.)2 Input Views
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
CarsSRN [41]22.220.8930.10722.050.8930.10624.820.9200.093
PixelNeRF [54]23.170.9050.11222.420.8950.12225.670.9350.083
CodeNeRF1 [22]23.800.910.106—1—1—125.710.930.099
FE-NVS2 [17]22.830.9060.099—2—2—224.640.9270.085
VisionNeRF [29]22.880.9080.08522.070.8980.091—3—3—3
Ours23.000.9110.08122.670.9100.08426.690.9480.055
Ours + GT PC23.690.9200.07723.310.9160.08027.290.9530.051
ChairsSRN [41]22.320.9040.09222.110.9010.09525.720.9350.072
PixelNeRF [54]23.720.9130.10123.150.9060.10926.210.9410.070
CodeNeRF1 [22]23.660.900.136—1—1—125.630.910.129
FE-NVS2 [17]23.210.9200.077—2—2—225.250.9400.065
VisionNeRF [29]24.480.9300.07723.880.9250.085—3—3—3
Ours23.460.9180.08123.030.9120.08827.360.9530.057
Ours + GT PC25.800.9400.07125.240.9340.07728.430.9610.053
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.358, + 0.892, + 0.451 + ], + "angle": 0, + "content": "Table 1: Novel view synthesis on ShapeNet cars and chairs. Our approach achieves state-of-the-art or competitive performance in single-view reconstruction and outperforms previous baselines by a large margin in the two-view setting. While PixelNeRF achieves a higher PSNR with single-view input, it can be seen that their results are blurry (c.f. Fig. 4a). Additionally to the standard setup, (Sym.) provides results on all target views that show mostly the opposite object side to the input view. We show results of our method using point cloud supervision (Ours + GT PC) at test time. These results are only to identify the gap induced by errors in point cloud prediction. The results shown in the rest of the paper use the Ours setup." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.464, + 0.471, + 0.541 + ], + "angle": 0, + "content": "(c.f. Section 4.4), (4) is a very efficient approach (c.f. Section 4.5), and (5) is compatible with test-time pose optimization (c.f. Section 4.6). We provide additional results in the appendix, such as more qualitative results, ablation studies, and an analysis of our point cloud prediction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.565, + 0.268, + 0.582 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.594, + 0.47, + 0.806 + ], + "angle": 0, + "content": "We conduct experiments for category-specific novel view synthesis, given one or two input views. We make use of the ShapeNet dataset provided by SRN [41], which comprises 3514 cars and 6591 chairs split into training, validation, and test sets. Each training instance has been rendered from 50 random camera locations on a sphere around the object. The test sets have 704 and 1318 examples, respectively, all with the same 251 views from an Archimedean spiral and the same lighting as during training. As done by all baselines, view 64 and additionally view 104 are used as input for single- and two-view experiments. The image resolution is \\(128 \\times 128\\). We use PSNR, SSIM, and LPIPS [56] to compare our approach with SRN [41], PixelNeRF [54], FE-NVS [17], and VisionNeRF [29]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We use 512/4096 points and 512/128 32-dim. embeddings per object. The shared features are of dimensionality 64. After separate pretraining of the point cloud prediction, we train for \\(0.9 / 1.2\\mathrm{M}\\) iterations. At test time, the instancespecific parameters are optimized for \\(10\\mathrm{k}\\) iterations each. Please refer to the suppl. mat. for additional information." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.464, + 0.825, + 0.48 + ], + "angle": 0, + "content": "4.2. Single- and Two-View Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.623 + ], + "angle": 0, + "content": "Table 1 summarizes our quantitative results. We leverage three different test-time optimization setups. In our main setup (Ours), we use the ResNet and the mask for test time optimization of point clouds to ensure a fair comparison with previous approaches. This setup is used for all comparisons and qualitative results shown. Furthermore, we report results using ground-truth point clouds (Ours + GT PC) and provide results with additional ground-truth depth supervision in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.893, + 0.82 + ], + "angle": 0, + "content": "In single-view reconstruction of cars, SimNP outperforms the state-of-the-art in SSIM and LPIPS. We attribute the small PSNR advantage of PixelNeRF to the blurriness of its reconstructions visible in Fig. 4a, which is confirmed by higher LPIPS. Also, we outperform PixelNeRF in inferring symmetric regions, as we show in columns 3-6 in Table 1 and in Figure 4b. SRN lacks details due to its global representation. PixelNeRF's reconstructions are overly-blurry in unseen regions, because pixel-aligned features are shared along the same rays of the input view without respecting the category-level self-similarities. VisionNeRF is not able to consistently transfer uncommon local color variations to symmetric object regions, which indicates a too strong" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.827, + 0.892, + 0.851 + ], + "angle": 0, + "content": "\\(^{1}\\)CodeNeRF [22] does not provide trained models. Note that the input views differ from the usual setup." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.851, + 0.892, + 0.888 + ], + "angle": 0, + "content": "2FE-NVS [17] does not have public code. Thus, shown numbers are taken from the paper/from authors, where available. SRN, PixelNeRF and VisionNeRF are evaluated using the PixelNeRF calc.metrics script." + }, + { + "type": "page_footnote", + "bbox": [ + 0.519, + 0.888, + 0.828, + 0.901 + ], + "angle": 0, + "content": "\\(^{3}\\)VisionNeRF [29] not applicable for multiple input views." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.827, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8846" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.094, + 0.476, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.459, + 0.414, + 0.473 + ], + "angle": 0, + "content": "(a) Qualitative single-view reconstruction." + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.094, + 0.891, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.46, + 0.756, + 0.473 + ], + "angle": 0, + "content": "(b) Metrics per view." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.486, + 0.895, + 0.562 + ], + "angle": 0, + "content": "Figure 4: a) Our method enables more detailed reconstructions and can better transfer appearance information to symmetric regions compared to SRN, PixelNeRF, and VisionNeRF. b) We show metric comparisons for each target view in the 251-view SRN test spiral for single-view reconstruction of cars. As visible, our overall better performance can be attributed to views showing regions symmetric to the input view (green areas and example views 10, 84, 175). Also, the related object-level method SRN shows rather flat curves indicating a weak adaptation to observations. This is not the case for SimNP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.673 + ], + "angle": 0, + "content": "global prior. In contrast, as visible from Figure 4a and Figure 1, our approach correctly transfers local patterns by leveraging learned category-level self-similarities. As visible from Figure 4b and columns 3-6 in Table 1, it significantly outperforms all baselines for views on regions with symmetric counterparts in the input image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.471, + 0.794 + ], + "angle": 0, + "content": "Compared to cars, the chairs category exhibits a larger variety in geometry but less texture details, leading to a smaller advantage through the self-similarity prior. Still, SimNP performs competitively compared to the baselines. Moreover, our strong results under the assumption of a given ground-truth point cloud (last rows in Table 1) indicate further potential of our method in combination with an improved point cloud prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In the two-view setting, our approach surpasses previous methods by a large margin on both categories. Here, the benefit comes not only from modeling self-similarities but also from our local point-based representation. The results show that it is better suited for fusing multiple observations than pixel-aligned methods, while at the same time being able to fit details. Looking at the qualitative results in Fig. 5," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.893, + 0.644 + ], + "angle": 0, + "content": "we can observe detailed reconstructions. In Tab. 2, we compare different two view setup against each other. The same side setup, which does observe one side of the object with two views, already clearly improves the results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.729, + 0.667 + ], + "angle": 0, + "content": "4.3. Learned Self-Similarities" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.895, + 0.902 + ], + "angle": 0, + "content": "As SimNP learns category-level self-similarities explicitly, we can directly inspect them to gain more insights. For visualizing the attention between neural points and embeddings, we proceed as follows. The rendering network computes weights between ray samples and neural points (see Eq. 3). By treating these weights just like RGB channels during ray marching, we obtain the influence of each neural point on each pixel. Finally, we multiply the learned category-level attention scores of a single embedding to the respective neural point influence, resulting in the self-similarity visualizations given in Figure 6. We describe the visualization procedure in detail in the supplemental material. Although we use as many embeddings as neural points (for cars) such that the attention could learn a one-to-one mapping, the representation learns to share information" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8847" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.473, + 0.238 + ], + "angle": 0, + "content": "
Cat.View Idx.PSNR↑SSIM↑LPIPS↓
Carssingle view (64)23.000.9110.081
2 v. close (63,64)23.390.9160.078
2 v. same side (44,64)25.760.9380.064
2 v. opposite (64,104)26.690.9480.055
Chairssingle view (64)23.460.9180.081
2 v. close (63,64)24.240.9240.077
2 v. same side (44,64)26.500.9440.064
2 v. opposite (64,104)27.360.9530.057
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.247, + 0.472, + 0.339 + ], + "angle": 0, + "content": "Table 2: Effect of input views. We evaluate different two view setups, comparing two very close views, 2 views from the same side of the object (approx. 90 degrees apart) and 2 views from opposite sides of the object. It can be seen that even if one side of the object is not seen (same side setup), the results improve strongly with 2 views." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.348, + 0.462, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.737 + ], + "angle": 0, + "content": "Figure 5: Two-view reconstruction. SimNP learns a high-quality 3D object representation given only two input views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.47, + 0.819 + ], + "angle": 0, + "content": "between similar points. Further, we observe that the connections for all embeddings are symmetric with respect to a reflection plane, indicating that the method successfully learned a prior about the plane symmetry of each category." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.373, + 0.849 + ], + "angle": 0, + "content": "4.4. Meaningful Representation Space" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Figure 7 provides results for interpolating the point cloud latent code and/or the embeddings of two instances obtained by multi-view fitting. SimNP learns a meaningful, disentan-" + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.093, + 0.558, + 0.104 + ], + "angle": 0, + "content": "RGB" + }, + { + "type": "text", + "bbox": [ + 0.64, + 0.092, + 0.826, + 0.104 + ], + "angle": 0, + "content": "Attention Scores for 4 Embeddings" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.116, + 0.89, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.558, + 0.895, + 0.667 + ], + "angle": 0, + "content": "Figure 6: Attention visualization. We render the influence of four different embeddings per category (one per column) on each ray, and visualize them on eight different examples. It becomes evident that embeddings specialize on self-similar local areas, e.g., wheels, or front lights. All embeddings show plane symmetric patterns, showing that our method recovers the symmetry of the category." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.895, + 0.779 + ], + "angle": 0, + "content": "gled representation space allowing for smooth shape and/or appearance transitions from one object to another. This is in contrast to pixel-aligned methods, which are not suited for interpolation. As an example, Fig. 8 shows a comparison with VisionNeRF, for which interpolation of the intermediate feature maps results in pixel space interpolation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.724, + 0.803 + ], + "angle": 0, + "content": "4.5. Efficient Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "SimNP renders a frame in 59ms, which is more than an order of magnitude faster than the pixel-aligned methods PixelNeRF (2.116s) and VisionNeRF (2.700s). We cannot fairly compare against FE-NVS, as no public code is available. The evaluation they provide in the paper is using amortized rendering, which is not comparable, as it only" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8848" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.101, + 0.462, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.407, + 0.471, + 0.483 + ], + "angle": 0, + "content": "Figure 7: Disentangled interpolation. Our point-based representation disentangles the basic shape from appearance. Point cloud latent code \\(\\mathbf{z}\\) and embeddings \\(\\mathbf{E}\\) can be interpolated independently resulting in a smooth shape and/or appearance transition between different objects." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.504, + 0.468, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.584, + 0.47, + 0.645 + ], + "angle": 0, + "content": "Figure 8: Interpolation comparison. Unlike VisionNeRF, which sticks to interpolation in pixel space for views similar to the input perspective, SimNP enables semantically meaningful interpolation like SRN but with more details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.47, + 0.807 + ], + "angle": 0, + "content": "works on many views simultaneously. However, we suspect that their method is also very fast. We train our method on a single A40 GPU for \\(\\approx 1.6\\) days. In comparison, SRN and PixelNeRF take 6 days on an RTX 8000 or Titan RTX, respectively, FE-NVS 2.5 days on 30 V100, and VisionNeRF around 5 days on 16 A100 GPUs, according to the authors. Overall, our method is very efficient with resource-efficient training and fast rendering that shows potential to be applied to larger scale reconstruction in future work." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.254, + 0.833 + ], + "angle": 0, + "content": "4.6. Pose Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To alleviate the assumption of a known camera pose at test time, we investigate optimization of camera parameters. More precisely, for each test example, we initialize eight evenly distributed camera locations on the sphere around" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.875, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.247, + 0.892, + 0.293 + ], + "angle": 0, + "content": "Figure 9: Pose Optimization. SimNP is compatible with camera pose optimization. The figure compares single-view reconstruction results with and without given camera pose." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.892, + 0.536 + ], + "angle": 0, + "content": "the object. The rotation matrix is always defined by the \\(z\\)-axis pointing towards the sphere center and the \\(y\\)-axis pointing upwards. At test time, we first initialize the point cloud latent using an encoder without ray encodings. Then, the pose in form of the camera location on the sphere is optimized by minimizing the Chamfer loss between the mask's pixel coordinates and the 2D projection of the point cloud. Lastly, we further finetune the point cloud with the Chamfer loss and optimize the embeddings for reconstruction of the input image. We select the pose that leads to the best reconstruction based on LPIPS against the single input image. Fig. 9 shows a comparison given vs. optimized camera pose. We achieve competitive performance (0.116 LPIPS) compared with CodeNeRF (0.114) without camera poses." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.556, + 0.619, + 0.571 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.734 + ], + "angle": 0, + "content": "We presented SimNP, the first object-level representation based on neural radiance fields, utilizing shared attention scores to learn category-specific self-similarities. Our method reaches state-of-the-art quality in single- and two-view reconstruction, while being highly efficient. Overall, it achieves a better data prior vs. observation trade-off. The improvements can be contributed to (1) leveraging a local neural point radiance field for the representation of object categories and (2) correctly propagating information between similar regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Limitations and Future Work The main limitation of SimNP is the assumption of a canonical space with ground-truth point clouds during training, which prohibits direct application on in-the-wild datasets. Therefore, an improved point cloud prediction in camera frame could be an exciting path for future work. Furthermore, we believe that it is promising to relax the point identities and to apply our self-similarity priors on a scene level to obtain large-scale, prior-driven reconstruction." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8849" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.169 + ], + "angle": 0, + "content": "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In European Conhference on Computer Vision (ECCV), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.468, + 0.227 + ], + "angle": 0, + "content": "[2] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J. Mitra, and Paul Guerrero. RenderDiffusion: Image diffusion for 3D reconstruction, inpainting and generation. arXiv pre-print, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.468, + 0.27 + ], + "angle": 0, + "content": "[3] François ARJ, Medioni GG, and Waupotitsch R. Mirror symmetry: 2-view stereo geometry. In Image and Vision Computing, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.468, + 0.325 + ], + "angle": 0, + "content": "[4] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-NeRF: A multiscale representation for antiailiasing neural radiance fields. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.468, + 0.395 + ], + "angle": 0, + "content": "[5] Rohan Chabra, Jan E. Lenssen, Eddy Ilg, Tanner Schmidt, Julian Straub, Steven Lovegrove, and Richard Newcombe. Deep local shapes: Learning local SDF priors for detailed 3D reconstruction. In European Conhference on Computer Vision (ECCV), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.468, + 0.479 + ], + "angle": 0, + "content": "[6] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.468, + 0.536 + ], + "angle": 0, + "content": "[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. Int. Conference on Computer Vision (ICCV), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.468, + 0.579 + ], + "angle": 0, + "content": "[8] Xin Chen, Yuwei Li, Xi Luo, Tianjia Shao, Jingyi Yu, Kun Zhou, and Youyi Zheng. Autosweep: Recovering 3d editable objects from a single photograph. In ACM Trans. Gr., 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.58, + 0.468, + 0.62 + ], + "angle": 0, + "content": "[9] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.468, + 0.689 + ], + "angle": 0, + "content": "[10] Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (SRF): Learning view synthesis from sparse views of novel scenes. In Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.468, + 0.747 + ], + "angle": 0, + "content": "[11] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3D-R2N2: A unified approach for single and multi-view 3d object reconstruction. In European Conhference on Computer Vision (ECCV), 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.468, + 0.844 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Int. Conference on Learning Representations (ICLR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[13] Haoqiang Fan, Hao Su, and Leonidas Guibas. A point set generation network for 3d object reconstruction from a single image. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[14] Roger Fawcett, Andrew Zisserman, and Michael Brady. Extracting structure from an affine view of a 3d point set with one or two bilateral symmetries. In Image and Vision Computing." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[15] Alexandre François, Gérard Medioni, and Roman Waupotitsch. Reconstructing mirror symmetric scenes from a single view using 2-view stereo geometry. In ICPR: Proceedings of the 16th International Conference on Pattern Recognition, 2002." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[16] Rohit Girdhar, David F. Fouhey, Mikel D. Rodriguez, and Abhinav Kumar Gupta. Learning a predictable and generative vector representation for objects. 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[17] Pengsheng Guo, Miguel Angel Bautista, Alex Colburn, Liang Yang, Daniel Ulbricht, Joshua M. Susskind, and Qi Shan. Fast and explicit neural view synthesis. In Winter Conference on Applications of Computer Vision (WACV), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[18] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.439 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[20] Du Q. Huynh. Affine reconstruction from monocular vision in the presence of a symmetry plane. In Int. Conference on Computer Vision (ICCV), 1999." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[21] Eldar Insafutdinov, Dylan Campbell, João F Henriques, and Andrea Vedaldi. SNeS: Learning probably symmetric neural surfaces from incomplete data. In European Conference on Computer Vision (ECCV), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[22] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Int. Conference on Computer Vision (ICCV), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[23] M. Johari, Y. Lepoittevin, and F. Fleuret. GeoNeRF: Generalizing nef with geometry priors. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[24] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In European Conhference on Computer Vision (ECCV), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[25] Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. In European Conhference on Computer Vision (ECCV), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[26] Andrey Kurenkov, Jingwei Ji, Animesh Garg, Viraj Mehta, JunYoung Gwak, Chris Choy, and Silvio Savarese. Deform-Net: Free-form deformation network for 3D shape reconstruction from a single image. 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[27] Kevin Koser, Christopher Zach, and Marc Pollefeys. Dense 3d reconstruction of symmetric scenes from a single image. In Proceedings of the 33rd international conference on Pattern recognition, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[28] Xingyi Li, Chaoyi Hong, Yiran Wang, Zhiguo Cao, Ke Xian, and Guosheng Lin. Symmnerf: Learning to explore symmetry prior for single-view view synthesis. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 1726-1742, December 2022." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8850" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[29] Kai-En Lin, Lin Yen-Chen, Wei-Sheng Lai, Tsung-Yi Lin, Yi-Chang Shih, and Ravi Ramamoorthi. Vision transformer for NeRF-based view synthesis from a single input image. In Winter Conference on Applications of Computer Vision (WACV), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[30] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In Int. Conference on Neural Information Processing Systems (NIPS), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.276 + ], + "angle": 0, + "content": "[31] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.47, + 0.346 + ], + "angle": 0, + "content": "[32] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conhference on Computer Vision (ECCV), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.47, + 0.403 + ], + "angle": 0, + "content": "[33] Andrew Zisserman Mukherjee, Dipti Prasad and Michael Brady. Shape from symmetry: Detecting and exploiting symmetry in affine images. In Philosophical Transactions: Physical Sciences and Engineering, 1995." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.47, + 0.473 + ], + "angle": 0, + "content": "[34] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.475, + 0.47, + 0.545 + ], + "angle": 0, + "content": "[35] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.47, + 0.588 + ], + "angle": 0, + "content": "[36] Cody J. Phillips, Matthieu Lecce, and Kostas Daniilidis. Seeing glassware: from edge detection to pose estimation and shape recovery. In Robotics: Science and Systems, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.47, + 0.644 + ], + "angle": 0, + "content": "[37] Ruslan Rakhimov, Andrei-Timotei Ardelean, Victor Lempitsky, and Evgeny Burnaev. NPBG++: Accelerating neural point-based graphics. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[38] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.47, + 0.744 + ], + "angle": 0, + "content": "[39] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[40] Sudipta N. Sinha, Krishnan Ramnath, and Richard Szeliski. Detecting and reconstructing 3d mirror symmetric objects. In European Conhference on Computer Vision (ECCV), 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.856 + ], + "angle": 0, + "content": "[41] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3D-structure-aware neural scene representations. In Int. Conference on Neural Information Processing Systems (NIPS), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[42] Maxim Tatarchenko, Stephan R. Richter, Rene Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In Confer-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ence on Computer Vision and Pattern Recognition (CVPR), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.149 + ], + "angle": 0, + "content": "[43] Sebastian Thurn and Ben Wegbreit. Shape from symmetry. In Int. Conference on Computer Vision (ICCV), 2005." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.151, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[44] Garvita Tiwari, Dimitrije Antic, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Pose-ndf: Modeling human pose manifolds with neural distance fields. In European Conference on Computer Vision (ECCV), October 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[45] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d scene representation and rendering. In Int. Conference on Computer Vision (ICCV), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.306 + ], + "angle": 0, + "content": "[46] Shubham Tulsiani, Nilesh Kulkarni, and Abhinav Gupta. Implicit mesh reconstruction from unannotated image collections. In arXiv pre-print, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.308, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[47] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.378, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[48] Daniel Watson, William Chan, Ricardo Martin Brulla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. In Int. Conference on Learning Representations (ICLR), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.435, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[49] Jiajun Wu, Chengkai Zhang, Tianfan Xue, William T. Freeman, and Joshua B. Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In Int. Conference on Neural Information Processing Systems (NIPS), 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[50] Shangzhe Wu, Ameesh Makadia, Jiajun Wu, Noah Snavely, Richard Tucker, and Angjoo Kanazawa. De-rendering the world's revolutionary artefacts. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[51] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[52] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-NeRF: Point-based neural radiance fields. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[53] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In Int. Conference on Neural Information Processing Systems (NIPS), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[54] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[55] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv pre-print, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[56] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Conference on Computer Vision and Pattern Recognition (CVPR), June 2018." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8851" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.472, + 0.15 + ], + "angle": 0, + "content": "[57] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Toch: Spatio-temporal object-to-hand correspondence for motion refinement. In European Conference on Computer Vision (ECCV). Springer, October 2022." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8852" + } + ] +] \ No newline at end of file diff --git a/2023/SimNP_ Learning Self-Similarity Priors Between Neural Points/54ca6d2a-b8a1-4a4e-a6f5-57eb0fddb4dd_origin.pdf b/2023/SimNP_ Learning Self-Similarity Priors Between Neural Points/54ca6d2a-b8a1-4a4e-a6f5-57eb0fddb4dd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..09770529396f2c2e7496d630a9c3c9be621f7b8b --- /dev/null +++ b/2023/SimNP_ Learning Self-Similarity Priors Between Neural Points/54ca6d2a-b8a1-4a4e-a6f5-57eb0fddb4dd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd7aacae0b9f0d7ce426db7a761ebbbdd86ff5cd4833b9026f3c246ab28f4d3c +size 2762103 diff --git a/2023/SimNP_ Learning Self-Similarity Priors Between Neural Points/full.md b/2023/SimNP_ Learning Self-Similarity Priors Between Neural Points/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f6693c3ba31dfaa8ddec41b986533a5ed2e19b29 --- /dev/null +++ b/2023/SimNP_ Learning Self-Similarity Priors Between Neural Points/full.md @@ -0,0 +1,329 @@ +# SimNP: Learning Self-Similarity Priors Between Neural Points + +Christopher Wewer1 + +Eddy $\mathrm{Ilg}^2$ + +Bernt Schiele1,2 + +Jan Eric Lenssen1 + +$^{1}$ Max Planck Institute for Informatics, Saarland Informatics Campus, Germany $^{2}$ Saarland University, Saarland Informatics Campus, Germany + +{cwewer, jlenssen}@mpi-inf.mpg.de + +![](images/6ae94561a93fdad6372fa550a80ae09782a582fec02d56a82be7b638a045e63e.jpg) +a) Neural Point Representation + +![](images/906f0cc58cda70bd2c77410623bd608f83e22e95b70ee421d95a23d5c5c460c3.jpg) +PixelNeRF + +![](images/50a4fbe33bc150d6f501d2b2040a5a3d2bd49b9faf53e8a60b9fa2de4f142cb7.jpg) +VisionNeRF + +![](images/b841b3a03b83ba82aaa2158df46baf78797f296174d388cfdcea767deea965bb.jpg) + +![](images/9d2bc123abd0238bf9879b8c04eb91f6a27a946eb81673f7cd1e1540d400a9a5.jpg) +b) Single-view reconstruction comparison. +Figure 1: a) We present SimNP, a renderable neural point radiance field that learns category-level self-similarities from data by connecting neural points to embeddings via optimized bipartite attention scores. b) The learned self-similarities can be used to transfer details from single- or few-view observations to unobserved, similar and symmetric parts of objects. + +![](images/ed172df9b9c52d7f7855c8d407b4ed81632708d29402dc18f60fbbaa4d04f4eb.jpg) + +![](images/10231a88a666945918f0114c81c85d4ce540834175d09f9f50a637462e306dcb.jpg) + +# Abstract + +Existing neural field representations for 3D object reconstruction either (1) utilize object-level representations, but suffer from low-quality details due to conditioning on a global latent code, or (2) are able to perfectly reconstruct the observations, but fail to utilize object-level prior knowledge to infer unobserved regions. We present SimNP, a method to learn category-level self-similarities, which combines the advantages of both worlds by connecting neural point radiance fields with a category-level self-similarity representation. Our contribution is two-fold. (1) We design the first neural point representation on a category level by utilizing the concept of coherent point clouds. The resulting neural point radiance fields store a high level of detail for locally supported object regions. (2) We learn how information is shared between neural points in an unconstrained and unsupervised fashion, which allows to derive unobserved regions of an object during the reconstruction process from given observations. We show that SimNP is able to outperform previous methods in reconstructing symmetric unseen object regions, surpassing methods that build upon category-level or pixel-aligned radiance fields, while providing semantic correspondences between instances. + +# 1. Introduction + +The human visual system succeeds in deriving 3D representations of objects just from incomplete 2D observations. Key to this ability is that given observations are successfully complemented by previously learned information about the 3D world. Replicating this ability has been a longstanding goal in computer vision. + +Since the task of reconstructing complete objects relies on generalization from a set of known examples, deep learning is an intuitive solution. The common approach is to use large amounts of data to train a category-level model [35, 41, 54, 29, 13, 26, 25, 22, 31, 9, 44, 57] and let the reconstruction process combine observations with the prior knowledge learned from data, which we refer to as the data prior. Notably, this introduces an inherent trade-off between contributions of the data prior and the observations. + +On one extreme of the spectrum, NeRF-like methods [32, 55, 4] do not use a data prior at all. With a high number of observations and an optimization process, they are able to nearly perfectly reconstruct novel views of scenes and objects. However, this renders them incapable of deriving unseen regions. On the other extreme of the spectrum are methods that learn the full space of radiance or + +signed distance functions belonging to a specific object category, such as SRN [41] and DeepSDF [35]. While these methods succeed in learning a complete representation of objects on an abstract level, they fail to represent individual details and the reconstruction process often performs retrieval from the learned data prior [42]. A similar behaviour has also been observed for generative models based on GANs [6] or diffusion [2], which yield visually impressive results but still diverge from given observations. + +The key challenge lies in combining the strengths of the methods from both ends of the spectrum. While one can perform a highly detailed reconstruction of the visible regions, the object model should allow to reuse this information in unseen regions. To this end, it is important to know that most objects show many structured self-similarities, often arising from different types of symmetries, such as point/plane symmetries or more general variants. None of the current approaches try to explicitly learn such self-similarities to perform better inference. + +This is where SimNP comes in. We propose a better data prior vs. observation trade-off by combining the best of both worlds: (1) a category-level data prior encoding self-similarities on top of a (2) local representation with test-time optimization. Instead of learning the full space of radiance functions for a given category, we move to learning a data prior one level of abstraction higher, i.e., we learn how information can be shared between local object regions. This enables us to learn characteristic self-similarity patterns from training data, which are used to propagate information from visible to invisible parts during inference. + +As learning a representation for category-level self-similarities implies modeling relationships between local regions of objects, a key observation in this work is that neural point representations are especially well-suited to describe such relationships. Besides their capacity to capture high-frequency patterns, the underlying sparse point cloud allows for explicit formulations of similarities. + +In summary, the contributions of our work are: + +1. We present the first generalizable neural point radiance field for representation of objects categories. +2. We propose a simple but effective mechanism that learns general similarities between local object regions in an unconstrained and unsupervised fashion. +3. We show that our model improves upon state of the art in reconstructing unobserved regions from a single image and by outperforming existing two-view methods by a large margin. At the same time, it is more efficient in training and rendering. + +# 2. Related Work + +Reconstruction from Observations Only While 3D reconstruction has traditionally been dominated by multi + +stage pipelines [38, 39], NeRF-like approaches [32, 55, 4] revolutionized novel view synthesis by using a volumetric density representation in continuous 3D space. Although such approaches can achieve very accurate reconstructions, they only work for the visible regions and require a large number of input views. + +Reconstruction with Data Priors To reduce the number of input views required for the reconstruction, many ways to leverage data priors have been proposed. + +Pixel-Aligned methods such as PixelNeRF [54] leverage image-based rendering and use features that represent data priors obtained from CNNs [47, 45] or vision transformers [29] trained on large amounts of data. The feature for a 3D location in space is then derived from the associated local 2D image features. Methods specialized on multi-view input additionally utilize Multi-View Stereo (MVS) [7, 10, 23]. FE-NVS [17] uses an autoencoder-like architecture with a 2D and 3D U-Net. Voxel-Based approaches use an MLP taking in a local latent code [5, 30] to learn a data prior of geometry and appearance for a single voxel. Point-Based approaches [1, 37, 52] first obtain a point cloud from an RGB-D sensor or with MVS, and then initialize features for the points from CNNs similar as in image-based rendering. In contrast to the above, our method introduces a combined global and local representation. The local features of our approach can be optimized during test time and are propagated to unobserved regions via our learned category-specific attention. + +Reconstruction with Object-Level Data Priors Early works for 3D reconstruction from single or few images leverage voxel grids [11, 16, 49]. However, while being able to reconstruct complete objects, for the latter it was shown that they do not actually perform reconstruction but image classification [42] and are therefore on the far end of data prior in the data prior vs. observation trade-off. + +Some approaches model objects via 3D point clouds [13, 26, 25] but do not model surfaces and appearance. More recently, approaches that model a continuous representation of 3D space with MLPs were proposed. Scene Representation Networks [41] (SRN) model the scene as a continuous feature field in 3D space. Implicit representations of surfaces were introduced in DeepSDF [35] and Occupancy Networks [31, 9]. While these approaches can be extended to model the appearance on the surfaces [53, 34], recent approaches leverage neural radiance fields [22]. All of the mentioned models can be parameterized by a global latent code to model a distribution over objects as an object-level data prior. However, since prior and representation are global, the reconstructions usually lack details. In contrast, our method learns a global geometry prior with local appearance features enabling higher frequency details. We are + +![](images/c0214eb8a19189d53cb8d09089bb9d5c4e3bf77f03a800ec334743b321028648.jpg) +Figure 2: Overview of SimNP. Our method is a category-level, coherent neural point radiance field, where points are connected to embedding vectors $\mathbf{E}$ via learnable attention scores $\mathbf{A}$ . The representation can be rendered using ray marching and a neural renderer. (a) During training, all parameters $(\square, \square)$ are optimized using multi-view supervision. Networks, features $\mathbf{S}$ , and scores are shared over the category $(\square)$ , while embeddings are instance-specific $(\square)$ . During inference, only embeddings $\mathbf{E}$ $(\square)$ are optimized from observations. In case of similar points $i, j$ (e.g., those shown in red), the network learned $a_{i,k} \approx a_{j,k} \forall k$ during training. Thus, supervision from one side means only one of points $i$ and $j$ needs to be visible to infer the value of embedding $k$ . (b) Given optimized embeddings, we can render the object from novel views. + +the first to introduce a point-based neural radiance field representation on object level. + +Modeling Self-Similarity and Symmetry To our knowledge, learning general self-similarities of 3D structures is a novel concept and has not been explored before. However, many early works proposed to use or recover predefined symmetries to obtain improved reconstructions [43, 33, 14, 20, 15, 27, 3, 40, 36, 8], which can be seen as a specific instance of self-similarity. More recent work performs reconstruction of a single instance by jointly optimizing the position and orientation of a symmetry plane [21] or providing the symmetry as input [28]. Modeling plane or rotational symmetry is also common in object reconstruction in the wild [51, 50, 24, 46]. However, the constraints are only used for training and not for inference. In contrast to all of the above, SimNP learns arbitrary self-similarities from data without supervision. + +VisionNeRF [29] extends PixelNeRF [54] with a Vision Transformer (ViT) [12], which is able to globally propagate features among different rays. Therefore, one can argue that the transformer is able to learn category-level symmetries, but only implicitly, and only in 2D pixel space. In contrast, SimNP operates directly in 3D and learns local self-similarity relationships explicitly, which improves reconstructions of symmetric parts, as our results show. + +# 3. Self-Similarity Priors between Neural Points + +In this section, we present SimNP. An overview is shown in Figure 2. At the heart of our method is a neural point + +representation that is comprised of a point cloud with attached feature vectors $\mathcal{P} = (\mathbf{P},\mathbf{S},\mathbf{F})$ with point positions $\mathbf{P}\in \mathbb{R}^{N\times 3}$ and two sets of point features $\mathbf{S}\in \mathbb{R}^{N\times D}$ and $\mathbf{F}\in \mathbb{R}^{N\times D}$ . The point features $\mathbf{S}$ are shared across the whole category and encode a point identity, while features $\mathbf{F}$ and positions $\mathbf{P}$ are individual for each instance, encoding local density and radiance. Features $\mathbf{F}$ are not explicitly stored but derived from embeddings $\mathbf{E}$ (c.f. Sec. 3.1). Note that we require our point clouds to be coherent, meaning that over different instances of one category, a single point with index $i$ describes roughly the same part of the object, e.g., the area around the right rear mirror of a car. + +Overview. The neural points are connected to a large set of embeddings $\mathbf{E}$ via bipartite attention scores $\mathbf{A}$ representing our category-level self-similarity, which is further detailed in Section 3.1. The neural point cloud can be volumetrically rendered from arbitrary views by using a network to decode color and density as described in Section 3.2. We employ the autodecoder framework [35], finding the optimal embeddings via optimization in training and inference, which is explained in Section 3.3. After assuming given point clouds for the introduction of our neural point radiance field, we cover the coherent point cloud prediction from single images independently in Section 3.4. + +# 3.1. Representing Category-Level Self-Similarity + +We learn self-similarity by learning how information can be shared between coherent neural points. There are two characteristics why neural point clouds are a well-suited + +representation for learning such explicit self-similarities: (1) They disentangle where the geometry is (represented by $\mathbf{P}$ ) from how it looks (represented by $\mathbf{F}$ ) and (2) they provide a discrete sampling of local parts that stays coherent across different instances due to this disentanglement. Since, with coherent point clouds, category-level self-similarities are invariant to variations in point location, similarities can be formulated on top of the neural point features $\mathbf{F}$ , independent of individual instances of $\mathbf{P}$ . + +Formally, we store local density and radiance information in a larger number of embeddings $\mathbf{E} \in \mathbb{R}^{M \times D}$ and connect these to our neural points by + +$$ +\mathbf {F} = \operatorname {s o f t m a x} (\mathbf {A}) \cdot \mathbf {E}, \tag {1} +$$ + +where $\mathbf{A} \in \mathbb{R}^{N \times M}$ are learnable attention scores between $N$ points and $M$ embeddings. The matrix $\mathbf{A}$ is shared across the whole category and encodes the category-level self-similarity prior. It learns to connect two object points to the same embeddings, if they are similar and share information (see also Figure 2). Then, during reconstruction with single or few views, it is sufficient if an embedding receives gradients from only one of the connected points. + +# 3.2. Neural Point Rendering + +Rendering a neural point cloud $\mathcal{P} = (\mathbf{P},\mathbf{S},\mathbf{F})$ follows a ray casting approach with a single message passing step from neural points to ray samples, similar to that of Point-NeRF [52]. We cast rays through each pixel and sample points along the rays. An inherent advantage of point-based neural fields is that ray samples which are not in the vicinity of any neural points can be discarded before neural network application, which increases performance in contrast to dense approaches. + +The rendering network consists of two MLPs, a kernel MLP $K_{\theta}$ describing local patches around each neural point, and a rendering MLP $F_{\psi}$ producing density and radiance from aggregated features. Given a ray sample point $\mathbf{x} \in \mathbb{R}^3$ , an intermediate feature vector $\mathbf{h}$ is computed as + +$$ +\mathbf {h} = \frac {1}{W} \sum_ {i \in \mathcal {N} (\mathbf {x})} w (\mathbf {x}, \mathbf {p} _ {i}) \cdot K _ {\theta} (\mathbf {x} - \mathbf {p} _ {i}, \mathbf {f} _ {i}, \mathbf {s} _ {i}), \tag {2} +$$ + +where $\mathcal{N}(\mathbf{x})$ contains the indices of the $k$ -nearest neural points ( $k = 8$ in our case) of $\mathbf{x}$ from $\mathbf{P}$ within a radius $r$ . $w$ is a weight based on inverse point distance: + +$$ +w (\mathbf {x}, \mathbf {p} _ {i}) = \frac {1}{\left\| \mathbf {x} - \mathbf {p} _ {i} \right\| _ {2}} \tag {3} +$$ + +and $W$ is the sum of weights in the neighborhood. Then, the final radiance $\mathbf{r}$ and density $\sigma$ are obtained as + +$$ +(\mathbf {r}, \sigma) = F _ {\psi} (\mathbf {h}, \mathbf {d}) \tag {4} +$$ + +based on features $\mathbf{h}$ and view direction $\mathbf{d} \in \mathbb{R}^3$ . For computing $\mathcal{N}(\mathbf{x})$ we created a custom PyTorch extension that implements an efficient, batch-wise voxel grid for ray sample $k$ -nn queries in CUDA. The module is based on the Point-NeRF code [52] and will be made publicly available with the rest of the code. + +Note that the rendering network is purely local, only receiving relative coordinates between ray samples and neural points. Thus, it is only able to learn a local surface model during training and no global, category-level information. The shared features $\mathbf{S}$ are a key ingredient in this formulation. In our studies, we found that they help to train a high-quality category-level neural point renderer and that they represent a density/radiance template of the category. We refer to the supplemental material for further discussion. + +# 3.3. Training and Inference + +For training and inference, we adopt the autodecoder framework [35, 5], thus, optimizing embeddings using gradient descent instead of predicting them using an encoder. Recent research has shown that approaches based on such test-time optimization have better capabilities when it comes to accurately representing the given observation [52]. + +As input during training, we assume multi-view renderings of a large set of objects, including camera poses for each view. For simplicity, we omit camera poses from the following formulations. Let $\mathbf{V} = f_{\theta, \psi}(\mathbf{P}, \mathbf{S}, \mathbf{A}, \mathbf{E})$ denote the function that renders an image $\mathbf{V}$ from a neural point cloud with positions $\mathbf{P}$ , embeddings $\mathbf{E}$ , attention scores $\mathbf{A}$ , and shared features $\mathbf{S}$ from a given camera pose. + +Training Given a dataset $\{(\mathbf{P}_i, \mathbf{I}_{i,j})\}_{i = j = 1}^{K,W}$ of $K$ objects of one category with coherent point clouds $\mathbf{P}_i$ , and $W$ renderings $\{\mathbf{I}_{i,j}\}_{j = 1}^{W}$ from random views for each object $i$ , our training procedure jointly optimizes embeddings $\mathbf{E}_i$ (omitted here for simplicity), attention scores $\mathbf{A}$ , shared point features $\mathbf{S}$ , and rendering network parameters $\theta, \psi$ : + +$$ +\hat {\mathbf {A}}, \hat {\mathbf {S}}, \hat {\theta}, \hat {\psi} = \underset {\mathbf {A}, \mathbf {S}, \theta , \psi} {\arg \min } \sum_ {i, j} \mathcal {L} \left(f _ {\theta , \psi} \left(\mathbf {P} _ {i}, \mathbf {S}, \mathbf {A}, \mathbf {E} _ {i}\right), \mathbf {I} _ {i, j}\right), \tag {5} +$$ + +where $\mathcal{L}$ is chosen as Mean Squared Error (MSE). After training, we freeze point renderer parameters $(\hat{\theta},\hat{\psi})$ , shared features $\hat{\mathbf{S}}$ and category-level symmetries $\hat{\mathbf{A}}$ , which are used to guide gradients to the correct embeddings during test-time optimization. + +Inference During inference, we assume that the camera pose is known. Given one or multiple views $\{\mathbf{I}_j\}_{j=1}^W$ and the coherent point cloud $\mathbf{P}$ , we can find embeddings $\hat{\mathbf{E}}$ by test-time optimization: + +![](images/0ef6299d438e900c485466324a21b1af2fb5fe4fbfad088716b12ea281bfee32.jpg) +Figure 3: Coherent point cloud prediction. An MLP with bottleneck is used to enforce the point cloud to be constructed as a low-rank deformation of a template (second to last layer output). During training, all trainable parameters $(\square, \square)$ are optimized using ground-truth point cloud supervision. During inference, the embedding $\mathbf{z}$ $(\square)$ can be optimized using different supervision signals: a ResNet predicting a point cloud from single image, mask, or depth. + +$$ +\hat {\mathbf {E}} = \underset {\mathbf {E}} {\arg \min} \sum_ {i} \mathcal {L} \left(f _ {\hat {\theta}, \hat {\psi}} (\mathbf {P}, \hat {\mathbf {S}}, \hat {\mathbf {A}}, \mathbf {E}), \mathbf {I} _ {i}\right). \qquad (6) +$$ + +Afterwards, the example can be rendered from other views. + +# 3.4. Coherent Point Clouds from Single Image + +Until now, coherent point clouds were assumed as given for the sake of simplicity. Still, they also have to be obtained from a single image during test time and are a crucial part of our method. We separate the point cloud prediction from the rest of SimNP such that it can be tackled independently. Note that we do not optimize point positions based on rendering losses. + +The full architecture for point cloud prediction is shown in Figure 3. Just like for our rendering branch, we follow the autodecoder framework in order to allow for flexible supervision at test time. However, unlike the local neural point features for capturing fine details, we opt for a global representation of the point cloud in form of a single learnable latent code $\mathbf{z} \in \mathbb{R}^l$ . During training, we optimize these embeddings jointly together with an MLP decoder $D_{\phi}: \mathbb{R}^l \to \mathbb{R}^{N \times 3}$ predicting the neural point coordinates in canonical space, given by the normalized orientation and position of ShapeNet objects. The output of the second to last layer is a low-dimensional bottleneck in order to serve as a low-rank representation of a high-dimensional point template from $\mathbb{R}^{N \times 3}$ [18]. The global representation together with this low-rank regularization allows us to always obtain coherent point clouds irrespective of the final supervision signal. + +Given point clouds $\{\mathcal{P}_i\}_{i=1}^K$ of training examples, we train $D_{\phi}$ and latent codes $\{\mathbf{z}_i\}_{i=1}^K$ to minimize the 3D Chamfer Distance (CD): + +$$ +\mathcal {L} _ {C D} = \sum_ {i = 1} ^ {K} \mathrm {C D} \left(D _ {\phi} \left(\mathbf {z} _ {i}\right), \mathcal {P} _ {i}\right). \tag {7} +$$ + +As one form of point cloud supervision during test time, we jointly train a ResNet18 [19] encoder with the image, + +segmentation mask, and camera ray encodings [48] as input to predict the same point cloud after decoding with $D_{\phi}$ , by minimizing the MSE. At test time, we optimize the latent codes while keeping encoder and decoder fixed. At first sight, it may seem counterintuitive to bring back an encoder into the autodecoder framework. However, by parameterizing the latent code directly, we gain flexibility regarding additional supervisory signals. + +Two optional sources of additional point cloud supervision at test time are the segmentation mask and a depth map, e.g., captured by an RGB-D camera. For the mask, we use a 2D Chamfer loss between its pixel coordinates and the 2D projection of the point cloud. Depth maps can be utilized by projecting the individual depth samples into the world frame and employing an asymmetric 3D Chamfer loss, i.e., minimizing the distance between each sample and its nearest point in the point cloud. Note that although occluded points from the perspective of the depth map are not explicitly supervised by this optional loss, they are still optimized implicitly because of our global latent representation. + +Besides leveraging these additional forms of data, test-time optimization of the point cloud is directly compatible with multi-view supervision. Given multiple views, the parameterized latent code $\mathbf{z}$ fuses the point clouds predicted individually for each image by the encoder. The following experimental results demonstrate that our neural point representation can benefit significantly from flexible point cloud supervision at test time. + +# 4. Experiments and Results + +In this section, we describe experiments made with SimNP and present our results. The goal of the experiments is to provide evidence for the following statements: SimNP (1) improves on previous approaches in reconstructing unseen symmetric object parts (c.f. Section 4.2), (2) learns correct self-similarities that respect symmetries of the object category (c.f. Section 4.3), (3) provides a meaningful representation that can be used for interpolation + +
Cat.Method1 Input View1 Input View (Sym.)2 Input Views
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
CarsSRN [41]22.220.8930.10722.050.8930.10624.820.9200.093
PixelNeRF [54]23.170.9050.11222.420.8950.12225.670.9350.083
CodeNeRF1 [22]23.800.910.106—1—1—125.710.930.099
FE-NVS2 [17]22.830.9060.099—2—2—224.640.9270.085
VisionNeRF [29]22.880.9080.08522.070.8980.091—3—3—3
Ours23.000.9110.08122.670.9100.08426.690.9480.055
Ours + GT PC23.690.9200.07723.310.9160.08027.290.9530.051
ChairsSRN [41]22.320.9040.09222.110.9010.09525.720.9350.072
PixelNeRF [54]23.720.9130.10123.150.9060.10926.210.9410.070
CodeNeRF1 [22]23.660.900.136—1—1—125.630.910.129
FE-NVS2 [17]23.210.9200.077—2—2—225.250.9400.065
VisionNeRF [29]24.480.9300.07723.880.9250.085—3—3—3
Ours23.460.9180.08123.030.9120.08827.360.9530.057
Ours + GT PC25.800.9400.07125.240.9340.07728.430.9610.053
+ +Table 1: Novel view synthesis on ShapeNet cars and chairs. Our approach achieves state-of-the-art or competitive performance in single-view reconstruction and outperforms previous baselines by a large margin in the two-view setting. While PixelNeRF achieves a higher PSNR with single-view input, it can be seen that their results are blurry (c.f. Fig. 4a). Additionally to the standard setup, (Sym.) provides results on all target views that show mostly the opposite object side to the input view. We show results of our method using point cloud supervision (Ours + GT PC) at test time. These results are only to identify the gap induced by errors in point cloud prediction. The results shown in the rest of the paper use the Ours setup. + +(c.f. Section 4.4), (4) is a very efficient approach (c.f. Section 4.5), and (5) is compatible with test-time pose optimization (c.f. Section 4.6). We provide additional results in the appendix, such as more qualitative results, ablation studies, and an analysis of our point cloud prediction. + +# 4.1. Experimental Setup + +We conduct experiments for category-specific novel view synthesis, given one or two input views. We make use of the ShapeNet dataset provided by SRN [41], which comprises 3514 cars and 6591 chairs split into training, validation, and test sets. Each training instance has been rendered from 50 random camera locations on a sphere around the object. The test sets have 704 and 1318 examples, respectively, all with the same 251 views from an Archimedean spiral and the same lighting as during training. As done by all baselines, view 64 and additionally view 104 are used as input for single- and two-view experiments. The image resolution is $128 \times 128$ . We use PSNR, SSIM, and LPIPS [56] to compare our approach with SRN [41], PixelNeRF [54], FE-NVS [17], and VisionNeRF [29]. + +We use 512/4096 points and 512/128 32-dim. embeddings per object. The shared features are of dimensionality 64. After separate pretraining of the point cloud prediction, we train for $0.9 / 1.2\mathrm{M}$ iterations. At test time, the instancespecific parameters are optimized for $10\mathrm{k}$ iterations each. Please refer to the suppl. mat. for additional information. + +# 4.2. Single- and Two-View Reconstruction + +Table 1 summarizes our quantitative results. We leverage three different test-time optimization setups. In our main setup (Ours), we use the ResNet and the mask for test time optimization of point clouds to ensure a fair comparison with previous approaches. This setup is used for all comparisons and qualitative results shown. Furthermore, we report results using ground-truth point clouds (Ours + GT PC) and provide results with additional ground-truth depth supervision in the supplement. + +In single-view reconstruction of cars, SimNP outperforms the state-of-the-art in SSIM and LPIPS. We attribute the small PSNR advantage of PixelNeRF to the blurriness of its reconstructions visible in Fig. 4a, which is confirmed by higher LPIPS. Also, we outperform PixelNeRF in inferring symmetric regions, as we show in columns 3-6 in Table 1 and in Figure 4b. SRN lacks details due to its global representation. PixelNeRF's reconstructions are overly-blurry in unseen regions, because pixel-aligned features are shared along the same rays of the input view without respecting the category-level self-similarities. VisionNeRF is not able to consistently transfer uncommon local color variations to symmetric object regions, which indicates a too strong + +![](images/3cc050c691fc04980ca50010c96171c1d7f56792ae0c0fa8d1cf30516915fb95.jpg) +(a) Qualitative single-view reconstruction. + +![](images/ec42be55d01d4a3b75a0611f326640cd852e0fcb1c37ed2e9a86ad09ecee9fd3.jpg) +(b) Metrics per view. +Figure 4: a) Our method enables more detailed reconstructions and can better transfer appearance information to symmetric regions compared to SRN, PixelNeRF, and VisionNeRF. b) We show metric comparisons for each target view in the 251-view SRN test spiral for single-view reconstruction of cars. As visible, our overall better performance can be attributed to views showing regions symmetric to the input view (green areas and example views 10, 84, 175). Also, the related object-level method SRN shows rather flat curves indicating a weak adaptation to observations. This is not the case for SimNP. + +global prior. In contrast, as visible from Figure 4a and Figure 1, our approach correctly transfers local patterns by leveraging learned category-level self-similarities. As visible from Figure 4b and columns 3-6 in Table 1, it significantly outperforms all baselines for views on regions with symmetric counterparts in the input image. + +Compared to cars, the chairs category exhibits a larger variety in geometry but less texture details, leading to a smaller advantage through the self-similarity prior. Still, SimNP performs competitively compared to the baselines. Moreover, our strong results under the assumption of a given ground-truth point cloud (last rows in Table 1) indicate further potential of our method in combination with an improved point cloud prediction. + +In the two-view setting, our approach surpasses previous methods by a large margin on both categories. Here, the benefit comes not only from modeling self-similarities but also from our local point-based representation. The results show that it is better suited for fusing multiple observations than pixel-aligned methods, while at the same time being able to fit details. Looking at the qualitative results in Fig. 5, + +we can observe detailed reconstructions. In Tab. 2, we compare different two view setup against each other. The same side setup, which does observe one side of the object with two views, already clearly improves the results. + +# 4.3. Learned Self-Similarities + +As SimNP learns category-level self-similarities explicitly, we can directly inspect them to gain more insights. For visualizing the attention between neural points and embeddings, we proceed as follows. The rendering network computes weights between ray samples and neural points (see Eq. 3). By treating these weights just like RGB channels during ray marching, we obtain the influence of each neural point on each pixel. Finally, we multiply the learned category-level attention scores of a single embedding to the respective neural point influence, resulting in the self-similarity visualizations given in Figure 6. We describe the visualization procedure in detail in the supplemental material. Although we use as many embeddings as neural points (for cars) such that the attention could learn a one-to-one mapping, the representation learns to share information + +
Cat.View Idx.PSNR↑SSIM↑LPIPS↓
Carssingle view (64)23.000.9110.081
2 v. close (63,64)23.390.9160.078
2 v. same side (44,64)25.760.9380.064
2 v. opposite (64,104)26.690.9480.055
Chairssingle view (64)23.460.9180.081
2 v. close (63,64)24.240.9240.077
2 v. same side (44,64)26.500.9440.064
2 v. opposite (64,104)27.360.9530.057
+ +Table 2: Effect of input views. We evaluate different two view setups, comparing two very close views, 2 views from the same side of the object (approx. 90 degrees apart) and 2 views from opposite sides of the object. It can be seen that even if one side of the object is not seen (same side setup), the results improve strongly with 2 views. + +![](images/7054eeb93a0cb8380066b6cf641eeebf12c0ccb9ab315ab96635e769e2cc82e0.jpg) +Figure 5: Two-view reconstruction. SimNP learns a high-quality 3D object representation given only two input views. + +between similar points. Further, we observe that the connections for all embeddings are symmetric with respect to a reflection plane, indicating that the method successfully learned a prior about the plane symmetry of each category. + +# 4.4. Meaningful Representation Space + +Figure 7 provides results for interpolating the point cloud latent code and/or the embeddings of two instances obtained by multi-view fitting. SimNP learns a meaningful, disentan- + +RGB + +Attention Scores for 4 Embeddings + +![](images/5ed381c504ba0c5365fd8213255d3e202debec8aee012a6737fd473ad6ba919e.jpg) +Figure 6: Attention visualization. We render the influence of four different embeddings per category (one per column) on each ray, and visualize them on eight different examples. It becomes evident that embeddings specialize on self-similar local areas, e.g., wheels, or front lights. All embeddings show plane symmetric patterns, showing that our method recovers the symmetry of the category. + +gled representation space allowing for smooth shape and/or appearance transitions from one object to another. This is in contrast to pixel-aligned methods, which are not suited for interpolation. As an example, Fig. 8 shows a comparison with VisionNeRF, for which interpolation of the intermediate feature maps results in pixel space interpolation. + +# 4.5. Efficient Representation + +SimNP renders a frame in 59ms, which is more than an order of magnitude faster than the pixel-aligned methods PixelNeRF (2.116s) and VisionNeRF (2.700s). We cannot fairly compare against FE-NVS, as no public code is available. The evaluation they provide in the paper is using amortized rendering, which is not comparable, as it only + +![](images/e84b33e4c40f11fbb435fdd7cc45fe4ccfc747846e90d21283f9dc319a2451dd.jpg) +Figure 7: Disentangled interpolation. Our point-based representation disentangles the basic shape from appearance. Point cloud latent code $\mathbf{z}$ and embeddings $\mathbf{E}$ can be interpolated independently resulting in a smooth shape and/or appearance transition between different objects. + +![](images/dac93f30b97fd72df9e25c613a5b42d53f45afd434b5dbf5044661400cc0214c.jpg) +Figure 8: Interpolation comparison. Unlike VisionNeRF, which sticks to interpolation in pixel space for views similar to the input perspective, SimNP enables semantically meaningful interpolation like SRN but with more details. + +works on many views simultaneously. However, we suspect that their method is also very fast. We train our method on a single A40 GPU for $\approx 1.6$ days. In comparison, SRN and PixelNeRF take 6 days on an RTX 8000 or Titan RTX, respectively, FE-NVS 2.5 days on 30 V100, and VisionNeRF around 5 days on 16 A100 GPUs, according to the authors. Overall, our method is very efficient with resource-efficient training and fast rendering that shows potential to be applied to larger scale reconstruction in future work. + +# 4.6. Pose Optimization + +To alleviate the assumption of a known camera pose at test time, we investigate optimization of camera parameters. More precisely, for each test example, we initialize eight evenly distributed camera locations on the sphere around + +![](images/16c4941e9dfbbd2c2214d8c208140d84c46f345575c75ee9218d630435d6243e.jpg) +Figure 9: Pose Optimization. SimNP is compatible with camera pose optimization. The figure compares single-view reconstruction results with and without given camera pose. + +the object. The rotation matrix is always defined by the $z$ -axis pointing towards the sphere center and the $y$ -axis pointing upwards. At test time, we first initialize the point cloud latent using an encoder without ray encodings. Then, the pose in form of the camera location on the sphere is optimized by minimizing the Chamfer loss between the mask's pixel coordinates and the 2D projection of the point cloud. Lastly, we further finetune the point cloud with the Chamfer loss and optimize the embeddings for reconstruction of the input image. We select the pose that leads to the best reconstruction based on LPIPS against the single input image. Fig. 9 shows a comparison given vs. optimized camera pose. We achieve competitive performance (0.116 LPIPS) compared with CodeNeRF (0.114) without camera poses. + +# 5. Conclusion + +We presented SimNP, the first object-level representation based on neural radiance fields, utilizing shared attention scores to learn category-specific self-similarities. Our method reaches state-of-the-art quality in single- and two-view reconstruction, while being highly efficient. Overall, it achieves a better data prior vs. observation trade-off. The improvements can be contributed to (1) leveraging a local neural point radiance field for the representation of object categories and (2) correctly propagating information between similar regions. + +Limitations and Future Work The main limitation of SimNP is the assumption of a canonical space with ground-truth point clouds during training, which prohibits direct application on in-the-wild datasets. Therefore, an improved point cloud prediction in camera frame could be an exciting path for future work. Furthermore, we believe that it is promising to relax the point identities and to apply our self-similarity priors on a scene level to obtain large-scale, prior-driven reconstruction. + +# References + +[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In European Conhference on Computer Vision (ECCV), 2020. +[2] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J. Mitra, and Paul Guerrero. RenderDiffusion: Image diffusion for 3D reconstruction, inpainting and generation. arXiv pre-print, 2022. +[3] François ARJ, Medioni GG, and Waupotitsch R. Mirror symmetry: 2-view stereo geometry. In Image and Vision Computing, 2003. +[4] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-NeRF: A multiscale representation for antiailiasing neural radiance fields. 2021. +[5] Rohan Chabra, Jan E. Lenssen, Eddy Ilg, Tanner Schmidt, Julian Straub, Steven Lovegrove, and Richard Newcombe. Deep local shapes: Learning local SDF priors for detailed 3D reconstruction. In European Conhference on Computer Vision (ECCV), 2020. +[6] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. Int. Conference on Computer Vision (ICCV), 2021. +[8] Xin Chen, Yuwei Li, Xi Luo, Tianjia Shao, Jingyi Yu, Kun Zhou, and Youyi Zheng. Autosweep: Recovering 3d editable objects from a single photograph. In ACM Trans. Gr., 2018. +[9] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. +[10] Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (SRF): Learning view synthesis from sparse views of novel scenes. In Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2021. +[11] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3D-R2N2: A unified approach for single and multi-view 3d object reconstruction. In European Conhference on Computer Vision (ECCV), 2016. +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Int. Conference on Learning Representations (ICLR), 2021. +[13] Haoqiang Fan, Hao Su, and Leonidas Guibas. A point set generation network for 3d object reconstruction from a single image. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. + +[14] Roger Fawcett, Andrew Zisserman, and Michael Brady. Extracting structure from an affine view of a 3d point set with one or two bilateral symmetries. In Image and Vision Computing. +[15] Alexandre François, Gérard Medioni, and Roman Waupotitsch. Reconstructing mirror symmetric scenes from a single view using 2-view stereo geometry. In ICPR: Proceedings of the 16th International Conference on Pattern Recognition, 2002. +[16] Rohit Girdhar, David F. Fouhey, Mikel D. Rodriguez, and Abhinav Kumar Gupta. Learning a predictable and generative vector representation for objects. 2016. +[17] Pengsheng Guo, Miguel Angel Bautista, Alex Colburn, Liang Yang, Daniel Ulbricht, Joshua M. Susskind, and Qi Shan. Fast and explicit neural view synthesis. In Winter Conference on Applications of Computer Vision (WACV), 2022. +[18] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. +[20] Du Q. Huynh. Affine reconstruction from monocular vision in the presence of a symmetry plane. In Int. Conference on Computer Vision (ICCV), 1999. +[21] Eldar Insafutdinov, Dylan Campbell, João F Henriques, and Andrea Vedaldi. SNeS: Learning probably symmetric neural surfaces from incomplete data. In European Conference on Computer Vision (ECCV), 2022. +[22] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Int. Conference on Computer Vision (ICCV), 2021. +[23] M. Johari, Y. Lepoittevin, and F. Fleuret. GeoNeRF: Generalizing nef with geometry priors. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[24] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In European Conhference on Computer Vision (ECCV), 2018. +[25] Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. In European Conhference on Computer Vision (ECCV), 2020. +[26] Andrey Kurenkov, Jingwei Ji, Animesh Garg, Viraj Mehta, JunYoung Gwak, Chris Choy, and Silvio Savarese. Deform-Net: Free-form deformation network for 3D shape reconstruction from a single image. 2018. +[27] Kevin Koser, Christopher Zach, and Marc Pollefeys. Dense 3d reconstruction of symmetric scenes from a single image. In Proceedings of the 33rd international conference on Pattern recognition, 2011. +[28] Xingyi Li, Chaoyi Hong, Yiran Wang, Zhiguo Cao, Ke Xian, and Guosheng Lin. Symmnerf: Learning to explore symmetry prior for single-view view synthesis. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 1726-1742, December 2022. + +[29] Kai-En Lin, Lin Yen-Chen, Wei-Sheng Lai, Tsung-Yi Lin, Yi-Chang Shih, and Ravi Ramamoorthi. Vision transformer for NeRF-based view synthesis from a single input image. In Winter Conference on Applications of Computer Vision (WACV), 2023. +[30] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In Int. Conference on Neural Information Processing Systems (NIPS), 2020. +[31] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. +[32] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conhference on Computer Vision (ECCV), 2020. +[33] Andrew Zisserman Mukherjee, Dipti Prasad and Michael Brady. Shape from symmetry: Detecting and exploiting symmetry in affine images. In Philosophical Transactions: Physical Sciences and Engineering, 1995. +[34] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. +[35] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. +[36] Cody J. Phillips, Matthieu Lecce, and Kostas Daniilidis. Seeing glassware: from edge detection to pose estimation and shape recovery. In Robotics: Science and Systems, 2016. +[37] Ruslan Rakhimov, Andrei-Timotei Ardelean, Victor Lempitsky, and Evgeny Burnaev. NPBG++: Accelerating neural point-based graphics. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[38] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. +[39] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. +[40] Sudipta N. Sinha, Krishnan Ramnath, and Richard Szeliski. Detecting and reconstructing 3d mirror symmetric objects. In European Conhference on Computer Vision (ECCV), 2012. +[41] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3D-structure-aware neural scene representations. In Int. Conference on Neural Information Processing Systems (NIPS), 2019. +[42] Maxim Tatarchenko, Stephan R. Richter, Rene Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In Confer- + +ence on Computer Vision and Pattern Recognition (CVPR), 2019. +[43] Sebastian Thurn and Ben Wegbreit. Shape from symmetry. In Int. Conference on Computer Vision (ICCV), 2005. +[44] Garvita Tiwari, Dimitrije Antic, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Pose-ndf: Modeling human pose manifolds with neural distance fields. In European Conference on Computer Vision (ECCV), October 2022. +[45] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d scene representation and rendering. In Int. Conference on Computer Vision (ICCV), 2021. +[46] Shubham Tulsiani, Nilesh Kulkarni, and Abhinav Gupta. Implicit mesh reconstruction from unannotated image collections. In arXiv pre-print, 2020. +[47] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. +[48] Daniel Watson, William Chan, Ricardo Martin Brulla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. In Int. Conference on Learning Representations (ICLR), 2023. +[49] Jiajun Wu, Chengkai Zhang, Tianfan Xue, William T. Freeman, and Joshua B. Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In Int. Conference on Neural Information Processing Systems (NIPS), 2016. +[50] Shangzhe Wu, Ameesh Makadia, Jiajun Wu, Noah Snavely, Richard Tucker, and Angjoo Kanazawa. De-rendering the world's revolutionary artefacts. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. +[51] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. +[52] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-NeRF: Point-based neural radiance fields. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[53] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In Int. Conference on Neural Information Processing Systems (NIPS), 2020. +[54] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. +[55] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv pre-print, 2020. +[56] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. + +[57] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Toch: Spatio-temporal object-to-hand correspondence for motion refinement. In European Conference on Computer Vision (ECCV). Springer, October 2022. \ No newline at end of file diff --git a/2023/SimNP_ Learning Self-Similarity Priors Between Neural Points/images.zip b/2023/SimNP_ Learning Self-Similarity Priors Between Neural Points/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..28dd3c2cf57dfbfc50f8c65f54afbe2a80b5293d --- /dev/null +++ b/2023/SimNP_ Learning Self-Similarity Priors Between Neural Points/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:527d2fdd92da458ed8e27a4635621d4005f499a2934411ca40f0c6c0303a605f +size 719835 diff --git a/2023/SimNP_ Learning Self-Similarity Priors Between Neural Points/layout.json b/2023/SimNP_ Learning Self-Similarity Priors Between Neural Points/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..de7e8cf055cbac9139755184f2578dae6e50d69e --- /dev/null +++ b/2023/SimNP_ Learning Self-Similarity Priors Between Neural Points/layout.json @@ -0,0 +1,9323 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 492, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 492, + 121 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 492, + 121 + ], + "type": "text", + "content": "SimNP: Learning Self-Similarity Priors Between Neural Points" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 142, + 196, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 142, + 196, + 158 + ], + "spans": [ + { + "bbox": [ + 96, + 142, + 196, + 158 + ], + "type": "text", + "content": "Christopher Wewer1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 225, + 142, + 273, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 142, + 273, + 158 + ], + "spans": [ + { + "bbox": [ + 225, + 142, + 273, + 158 + ], + "type": "text", + "content": "Eddy " + }, + { + "bbox": [ + 225, + 142, + 273, + 158 + ], + "type": "inline_equation", + "content": "\\mathrm{Ilg}^2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 142, + 380, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 142, + 380, + 156 + ], + "spans": [ + { + "bbox": [ + 302, + 142, + 380, + 156 + ], + "type": "text", + "content": "Bernt Schiele1,2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 408, + 143, + 495, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 143, + 495, + 156 + ], + "spans": [ + { + "bbox": [ + 408, + 143, + 495, + 156 + ], + "type": "text", + "content": "Jan Eric Lenssen1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 176, + 487, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 176, + 487, + 205 + ], + "spans": [ + { + "bbox": [ + 107, + 176, + 487, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 176, + 487, + 205 + ], + "type": "text", + "content": "Max Planck Institute for Informatics, Saarland Informatics Campus, Germany " + }, + { + "bbox": [ + 107, + 176, + 487, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 176, + 487, + 205 + ], + "type": "text", + "content": "Saarland University, Saarland Informatics Campus, Germany" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 208, + 208, + 386, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 208, + 386, + 220 + ], + "spans": [ + { + "bbox": [ + 208, + 208, + 386, + 220 + ], + "type": "text", + "content": "{cwewer, jlenssen}@mpi-inf.mpg.de" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 50, + 240, + 225, + 350 + ], + "blocks": [ + { + "bbox": [ + 50, + 240, + 225, + 350 + ], + "lines": [ + { + "bbox": [ + 50, + 240, + 225, + 350 + ], + "spans": [ + { + "bbox": [ + 50, + 240, + 225, + 350 + ], + "type": "image", + "image_path": "6ae94561a93fdad6372fa550a80ae09782a582fec02d56a82be7b638a045e63e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 358, + 203, + 370 + ], + "lines": [ + { + "bbox": [ + 77, + 358, + 203, + 370 + ], + "spans": [ + { + "bbox": [ + 77, + 358, + 203, + 370 + ], + "type": "text", + "content": "a) Neural Point Representation" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 241, + 238, + 288, + 349 + ], + "blocks": [ + { + "bbox": [ + 241, + 238, + 288, + 349 + ], + "lines": [ + { + "bbox": [ + 241, + 238, + 288, + 349 + ], + "spans": [ + { + "bbox": [ + 241, + 238, + 288, + 349 + ], + "type": "image", + "image_path": "906f0cc58cda70bd2c77410623bd608f83e22e95b70ee421d95a23d5c5c460c3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 292, + 238, + 339, + 247 + ], + "lines": [ + { + "bbox": [ + 292, + 238, + 339, + 247 + ], + "spans": [ + { + "bbox": [ + 292, + 238, + 339, + 247 + ], + "type": "text", + "content": "PixelNeRF" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 292, + 253, + 339, + 349 + ], + "blocks": [ + { + "bbox": [ + 292, + 253, + 339, + 349 + ], + "lines": [ + { + "bbox": [ + 292, + 253, + 339, + 349 + ], + "spans": [ + { + "bbox": [ + 292, + 253, + 339, + 349 + ], + "type": "image", + "image_path": "50a4fbe33bc150d6f501d2b2040a5a3d2bd49b9faf53e8a60b9fa2de4f142cb7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 239, + 389, + 247 + ], + "lines": [ + { + "bbox": [ + 345, + 239, + 389, + 247 + ], + "spans": [ + { + "bbox": [ + 345, + 239, + 389, + 247 + ], + "type": "text", + "content": "VisionNeRF" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 344, + 253, + 391, + 349 + ], + "blocks": [ + { + "bbox": [ + 344, + 253, + 391, + 349 + ], + "lines": [ + { + "bbox": [ + 344, + 253, + 391, + 349 + ], + "spans": [ + { + "bbox": [ + 344, + 253, + 391, + 349 + ], + "type": "image", + "image_path": "b841b3a03b83ba82aaa2158df46baf78797f296174d388cfdcea767deea965bb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 395, + 238, + 443, + 355 + ], + "blocks": [ + { + "bbox": [ + 395, + 238, + 443, + 355 + ], + "lines": [ + { + "bbox": [ + 395, + 238, + 443, + 355 + ], + "spans": [ + { + "bbox": [ + 395, + 238, + 443, + 355 + ], + "type": "image", + "image_path": "9d2bc123abd0238bf9879b8c04eb91f6a27a946eb81673f7cd1e1540d400a9a5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 358, + 476, + 371 + ], + "lines": [ + { + "bbox": [ + 304, + 358, + 476, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 476, + 371 + ], + "type": "text", + "content": "b) Single-view reconstruction comparison." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 379, + 546, + 415 + ], + "lines": [ + { + "bbox": [ + 45, + 379, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 45, + 379, + 546, + 415 + ], + "type": "text", + "content": "Figure 1: a) We present SimNP, a renderable neural point radiance field that learns category-level self-similarities from data by connecting neural points to embeddings via optimized bipartite attention scores. b) The learned self-similarities can be used to transfer details from single- or few-view observations to unobserved, similar and symmetric parts of objects." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 447, + 238, + 494, + 349 + ], + "blocks": [ + { + "bbox": [ + 447, + 238, + 494, + 349 + ], + "lines": [ + { + "bbox": [ + 447, + 238, + 494, + 349 + ], + "spans": [ + { + "bbox": [ + 447, + 238, + 494, + 349 + ], + "type": "image", + "image_path": "ed172df9b9c52d7f7855c8d407b4ed81632708d29402dc18f60fbbaa4d04f4eb.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 495, + 238, + 541, + 348 + ], + "blocks": [ + { + "bbox": [ + 495, + 238, + 541, + 348 + ], + "lines": [ + { + "bbox": [ + 495, + 238, + 541, + 348 + ], + "spans": [ + { + "bbox": [ + 495, + 238, + 541, + 348 + ], + "type": "image", + "image_path": "10231a88a666945918f0114c81c85d4ce540834175d09f9f50a637462e306dcb.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 143, + 427, + 192, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 427, + 192, + 439 + ], + "spans": [ + { + "bbox": [ + 143, + 427, + 192, + 439 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": "Existing neural field representations for 3D object reconstruction either (1) utilize object-level representations, but suffer from low-quality details due to conditioning on a global latent code, or (2) are able to perfectly reconstruct the observations, but fail to utilize object-level prior knowledge to infer unobserved regions. We present SimNP, a method to learn category-level self-similarities, which combines the advantages of both worlds by connecting neural point radiance fields with a category-level self-similarity representation. Our contribution is two-fold. (1) We design the first neural point representation on a category level by utilizing the concept of coherent point clouds. The resulting neural point radiance fields store a high level of detail for locally supported object regions. (2) We learn how information is shared between neural points in an unconstrained and unsupervised fashion, which allows to derive unobserved regions of an object during the reconstruction process from given observations. We show that SimNP is able to outperform previous methods in reconstructing symmetric unseen object regions, surpassing methods that build upon category-level or pixel-aligned radiance fields, while providing semantic correspondences between instances." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 427, + 386, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 427, + 386, + 439 + ], + "spans": [ + { + "bbox": [ + 307, + 427, + 386, + 439 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 449, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 520 + ], + "type": "text", + "content": "The human visual system succeeds in deriving 3D representations of objects just from incomplete 2D observations. Key to this ability is that given observations are successfully complemented by previously learned information about the 3D world. Replicating this ability has been a longstanding goal in computer vision." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 521, + 546, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 628 + ], + "type": "text", + "content": "Since the task of reconstructing complete objects relies on generalization from a set of known examples, deep learning is an intuitive solution. The common approach is to use large amounts of data to train a category-level model [35, 41, 54, 29, 13, 26, 25, 22, 31, 9, 44, 57] and let the reconstruction process combine observations with the prior knowledge learned from data, which we refer to as the data prior. Notably, this introduces an inherent trade-off between contributions of the data prior and the observations." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "On one extreme of the spectrum, NeRF-like methods [32, 55, 4] do not use a data prior at all. With a high number of observations and an optimization process, they are able to nearly perfectly reconstruct novel views of scenes and objects. However, this renders them incapable of deriving unseen regions. On the other extreme of the spectrum are methods that learn the full space of radiance or" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8841" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 178 + ], + "type": "text", + "content": "signed distance functions belonging to a specific object category, such as SRN [41] and DeepSDF [35]. While these methods succeed in learning a complete representation of objects on an abstract level, they fail to represent individual details and the reconstruction process often performs retrieval from the learned data prior [42]. A similar behaviour has also been observed for generative models based on GANs [6] or diffusion [2], which yield visually impressive results but still diverge from given observations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 180, + 286, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 180, + 286, + 298 + ], + "spans": [ + { + "bbox": [ + 49, + 180, + 286, + 298 + ], + "type": "text", + "content": "The key challenge lies in combining the strengths of the methods from both ends of the spectrum. While one can perform a highly detailed reconstruction of the visible regions, the object model should allow to reuse this information in unseen regions. To this end, it is important to know that most objects show many structured self-similarities, often arising from different types of symmetries, such as point/plane symmetries or more general variants. None of the current approaches try to explicitly learn such self-similarities to perform better inference." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 300, + 286, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 300, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 300, + 286, + 430 + ], + "type": "text", + "content": "This is where SimNP comes in. We propose a better data prior vs. observation trade-off by combining the best of both worlds: (1) a category-level data prior encoding self-similarities on top of a (2) local representation with test-time optimization. Instead of learning the full space of radiance functions for a given category, we move to learning a data prior one level of abstraction higher, i.e., we learn how information can be shared between local object regions. This enables us to learn characteristic self-similarity patterns from training data, which are used to propagate information from visible to invisible parts during inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 432, + 286, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 432, + 286, + 514 + ], + "spans": [ + { + "bbox": [ + 49, + 432, + 286, + 514 + ], + "type": "text", + "content": "As learning a representation for category-level self-similarities implies modeling relationships between local regions of objects, a key observation in this work is that neural point representations are especially well-suited to describe such relationships. Besides their capacity to capture high-frequency patterns, the underlying sparse point cloud allows for explicit formulations of similarities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 515, + 247, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 515, + 247, + 525 + ], + "spans": [ + { + "bbox": [ + 60, + 515, + 247, + 525 + ], + "type": "text", + "content": "In summary, the contributions of our work are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 533, + 285, + 658 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 57, + 533, + 285, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 285, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 285, + 555 + ], + "type": "text", + "content": "1. We present the first generalizable neural point radiance field for representation of objects categories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 560, + 285, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 560, + 285, + 594 + ], + "spans": [ + { + "bbox": [ + 56, + 560, + 285, + 594 + ], + "type": "text", + "content": "2. We propose a simple but effective mechanism that learns general similarities between local object regions in an unconstrained and unsupervised fashion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 600, + 285, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 285, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 285, + 658 + ], + "type": "text", + "content": "3. We show that our model improves upon state of the art in reconstructing unobserved regions from a single image and by outperforming existing two-view methods by a large margin. At the same time, it is more efficient in training and rendering." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 49, + 670, + 132, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 132, + 681 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 132, + 681 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 690, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 690, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 690, + 286, + 712 + ], + "type": "text", + "content": "Reconstruction from Observations Only While 3D reconstruction has traditionally been dominated by multi" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 308, + 72, + 544, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 143 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 143 + ], + "type": "text", + "content": "stage pipelines [38, 39], NeRF-like approaches [32, 55, 4] revolutionized novel view synthesis by using a volumetric density representation in continuous 3D space. Although such approaches can achieve very accurate reconstructions, they only work for the visible regions and require a large number of input views." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 308, + 160, + 544, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 160, + 544, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 160, + 544, + 194 + ], + "type": "text", + "content": "Reconstruction with Data Priors To reduce the number of input views required for the reconstruction, many ways to leverage data priors have been proposed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 196, + 544, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 196, + 544, + 422 + ], + "spans": [ + { + "bbox": [ + 308, + 196, + 544, + 422 + ], + "type": "text", + "content": "Pixel-Aligned methods such as PixelNeRF [54] leverage image-based rendering and use features that represent data priors obtained from CNNs [47, 45] or vision transformers [29] trained on large amounts of data. The feature for a 3D location in space is then derived from the associated local 2D image features. Methods specialized on multi-view input additionally utilize Multi-View Stereo (MVS) [7, 10, 23]. FE-NVS [17] uses an autoencoder-like architecture with a 2D and 3D U-Net. Voxel-Based approaches use an MLP taking in a local latent code [5, 30] to learn a data prior of geometry and appearance for a single voxel. Point-Based approaches [1, 37, 52] first obtain a point cloud from an RGB-D sensor or with MVS, and then initialize features for the points from CNNs similar as in image-based rendering. In contrast to the above, our method introduces a combined global and local representation. The local features of our approach can be optimized during test time and are propagated to unobserved regions via our learned category-specific attention." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 439, + 544, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 439, + 544, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 439, + 544, + 521 + ], + "type": "text", + "content": "Reconstruction with Object-Level Data Priors Early works for 3D reconstruction from single or few images leverage voxel grids [11, 16, 49]. However, while being able to reconstruct complete objects, for the latter it was shown that they do not actually perform reconstruction but image classification [42] and are therefore on the far end of data prior in the data prior vs. observation trade-off." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 522, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 544, + 712 + ], + "type": "text", + "content": "Some approaches model objects via 3D point clouds [13, 26, 25] but do not model surfaces and appearance. More recently, approaches that model a continuous representation of 3D space with MLPs were proposed. Scene Representation Networks [41] (SRN) model the scene as a continuous feature field in 3D space. Implicit representations of surfaces were introduced in DeepSDF [35] and Occupancy Networks [31, 9]. While these approaches can be extended to model the appearance on the surfaces [53, 34], recent approaches leverage neural radiance fields [22]. All of the mentioned models can be parameterized by a global latent code to model a distribution over objects as an object-level data prior. However, since prior and representation are global, the reconstructions usually lack details. In contrast, our method learns a global geometry prior with local appearance features enabling higher frequency details. We are" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "8842" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 541, + 231 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 541, + 231 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 541, + 231 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 541, + 231 + ], + "type": "image", + "image_path": "c0214eb8a19189d53cb8d09089bb9d5c4e3bf77f03a800ec334743b321028648.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "lines": [ + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "text", + "content": "Figure 2: Overview of SimNP. Our method is a category-level, coherent neural point radiance field, where points are connected to embedding vectors " + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "text", + "content": " via learnable attention scores " + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "text", + "content": ". The representation can be rendered using ray marching and a neural renderer. (a) During training, all parameters " + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "inline_equation", + "content": "(\\square, \\square)" + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "text", + "content": " are optimized using multi-view supervision. Networks, features " + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "text", + "content": ", and scores are shared over the category " + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "inline_equation", + "content": "(\\square)" + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "text", + "content": ", while embeddings are instance-specific " + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "inline_equation", + "content": "(\\square)" + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "text", + "content": ". During inference, only embeddings " + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "inline_equation", + "content": "(\\square)" + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "text", + "content": " are optimized from observations. In case of similar points " + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "inline_equation", + "content": "i, j" + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "text", + "content": " (e.g., those shown in red), the network learned " + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "inline_equation", + "content": "a_{i,k} \\approx a_{j,k} \\forall k" + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "text", + "content": " during training. Thus, supervision from one side means only one of points " + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "text", + "content": " needs to be visible to infer the value of embedding " + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 241, + 548, + 327 + ], + "type": "text", + "content": ". (b) Given optimized embeddings, we can render the object from novel views." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 346, + 288, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 288, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 288, + 371 + ], + "type": "text", + "content": "the first to introduce a point-based neural radiance field representation on object level." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 384, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 563 + ], + "type": "text", + "content": "Modeling Self-Similarity and Symmetry To our knowledge, learning general self-similarities of 3D structures is a novel concept and has not been explored before. However, many early works proposed to use or recover predefined symmetries to obtain improved reconstructions [43, 33, 14, 20, 15, 27, 3, 40, 36, 8], which can be seen as a specific instance of self-similarity. More recent work performs reconstruction of a single instance by jointly optimizing the position and orientation of a symmetry plane [21] or providing the symmetry as input [28]. Modeling plane or rotational symmetry is also common in object reconstruction in the wild [51, 50, 24, 46]. However, the constraints are only used for training and not for inference. In contrast to all of the above, SimNP learns arbitrary self-similarities from data without supervision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 563, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 660 + ], + "type": "text", + "content": "VisionNeRF [29] extends PixelNeRF [54] with a Vision Transformer (ViT) [12], which is able to globally propagate features among different rays. Therefore, one can argue that the transformer is able to learn category-level symmetries, but only implicitly, and only in 2D pixel space. In contrast, SimNP operates directly in 3D and learns local self-similarity relationships explicitly, which improves reconstructions of symmetric parts, as our results show." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 669, + 287, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 287, + 683 + ], + "type": "text", + "content": "3. Self-Similarity Priors between Neural Points" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "In this section, we present SimNP. An overview is shown in Figure 2. At the heart of our method is a neural point" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "text", + "content": "representation that is comprised of a point cloud with attached feature vectors " + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = (\\mathbf{P},\\mathbf{S},\\mathbf{F})" + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "text", + "content": " with point positions " + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{P}\\in \\mathbb{R}^{N\\times 3}" + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "text", + "content": " and two sets of point features " + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{S}\\in \\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{F}\\in \\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "text", + "content": ". The point features " + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "text", + "content": " are shared across the whole category and encode a point identity, while features " + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "text", + "content": " and positions " + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "text", + "content": " are individual for each instance, encoding local density and radiance. Features " + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "text", + "content": " are not explicitly stored but derived from embeddings " + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "text", + "content": " (c.f. Sec. 3.1). Note that we require our point clouds to be coherent, meaning that over different instances of one category, a single point with index " + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 346, + 547, + 491 + ], + "type": "text", + "content": " describes roughly the same part of the object, e.g., the area around the right rear mirror of a car." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 506, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 651 + ], + "type": "text", + "content": "Overview. The neural points are connected to a large set of embeddings " + }, + { + "bbox": [ + 304, + 506, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 304, + 506, + 547, + 651 + ], + "type": "text", + "content": " via bipartite attention scores " + }, + { + "bbox": [ + 304, + 506, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 506, + 547, + 651 + ], + "type": "text", + "content": " representing our category-level self-similarity, which is further detailed in Section 3.1. The neural point cloud can be volumetrically rendered from arbitrary views by using a network to decode color and density as described in Section 3.2. We employ the autodecoder framework [35], finding the optimal embeddings via optimization in training and inference, which is explained in Section 3.3. After assuming given point clouds for the introduction of our neural point radiance field, we cover the coherent point cloud prediction from single images independently in Section 3.4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 658, + 537, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 537, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 537, + 672 + ], + "type": "text", + "content": "3.1. Representing Category-Level Self-Similarity" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "We learn self-similarity by learning how information can be shared between coherent neural points. There are two characteristics why neural point clouds are a well-suited" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8843" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "text", + "content": "representation for learning such explicit self-similarities: (1) They disentangle where the geometry is (represented by " + }, + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "text", + "content": ") from how it looks (represented by " + }, + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "text", + "content": ") and (2) they provide a discrete sampling of local parts that stays coherent across different instances due to this disentanglement. Since, with coherent point clouds, category-level self-similarities are invariant to variations in point location, similarities can be formulated on top of the neural point features " + }, + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "text", + "content": ", independent of individual instances of " + }, + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 286, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 286, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 286, + 216 + ], + "type": "text", + "content": "Formally, we store local density and radiance information in a larger number of embeddings " + }, + { + "bbox": [ + 46, + 180, + 286, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{E} \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 46, + 180, + 286, + 216 + ], + "type": "text", + "content": " and connect these to our neural points by" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 122, + 226, + 286, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 226, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 122, + 226, + 286, + 239 + ], + "type": "interline_equation", + "content": "\\mathbf {F} = \\operatorname {s o f t m a x} (\\mathbf {A}) \\cdot \\mathbf {E}, \\tag {1}", + "image_path": "6f54caf7042ba027391b183bc4df7c6b284268fcf4acd055ccce756c0c7ecf6a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 248, + 286, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 286, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 286, + 344 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 248, + 286, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{N \\times M}" + }, + { + "bbox": [ + 46, + 248, + 286, + 344 + ], + "type": "text", + "content": " are learnable attention scores between " + }, + { + "bbox": [ + 46, + 248, + 286, + 344 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 248, + 286, + 344 + ], + "type": "text", + "content": " points and " + }, + { + "bbox": [ + 46, + 248, + 286, + 344 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 248, + 286, + 344 + ], + "type": "text", + "content": " embeddings. The matrix " + }, + { + "bbox": [ + 46, + 248, + 286, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 248, + 286, + 344 + ], + "type": "text", + "content": " is shared across the whole category and encodes the category-level self-similarity prior. It learns to connect two object points to the same embeddings, if they are similar and share information (see also Figure 2). Then, during reconstruction with single or few views, it is sufficient if an embedding receives gradients from only one of the connected points." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 353, + 181, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 353, + 181, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 353, + 181, + 365 + ], + "type": "text", + "content": "3.2. Neural Point Rendering" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 371, + 286, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 286, + 479 + ], + "type": "text", + "content": "Rendering a neural point cloud " + }, + { + "bbox": [ + 46, + 371, + 286, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = (\\mathbf{P},\\mathbf{S},\\mathbf{F})" + }, + { + "bbox": [ + 46, + 371, + 286, + 479 + ], + "type": "text", + "content": " follows a ray casting approach with a single message passing step from neural points to ray samples, similar to that of Point-NeRF [52]. We cast rays through each pixel and sample points along the rays. An inherent advantage of point-based neural fields is that ray samples which are not in the vicinity of any neural points can be discarded before neural network application, which increases performance in contrast to dense approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 479, + 286, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 286, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 286, + 538 + ], + "type": "text", + "content": "The rendering network consists of two MLPs, a kernel MLP " + }, + { + "bbox": [ + 46, + 479, + 286, + 538 + ], + "type": "inline_equation", + "content": "K_{\\theta}" + }, + { + "bbox": [ + 46, + 479, + 286, + 538 + ], + "type": "text", + "content": " describing local patches around each neural point, and a rendering MLP " + }, + { + "bbox": [ + 46, + 479, + 286, + 538 + ], + "type": "inline_equation", + "content": "F_{\\psi}" + }, + { + "bbox": [ + 46, + 479, + 286, + 538 + ], + "type": "text", + "content": " producing density and radiance from aggregated features. Given a ray sample point " + }, + { + "bbox": [ + 46, + 479, + 286, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 479, + 286, + 538 + ], + "type": "text", + "content": ", an intermediate feature vector " + }, + { + "bbox": [ + 46, + 479, + 286, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 46, + 479, + 286, + 538 + ], + "type": "text", + "content": " is computed as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 74, + 548, + 286, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 548, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 74, + 548, + 286, + 578 + ], + "type": "interline_equation", + "content": "\\mathbf {h} = \\frac {1}{W} \\sum_ {i \\in \\mathcal {N} (\\mathbf {x})} w (\\mathbf {x}, \\mathbf {p} _ {i}) \\cdot K _ {\\theta} (\\mathbf {x} - \\mathbf {p} _ {i}, \\mathbf {f} _ {i}, \\mathbf {s} _ {i}), \\tag {2}", + "image_path": "03d80ab18e124b8764d5ee639750aa95a3ba5d70a9e4e324ea9cac257b9015db.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 588, + 286, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 588, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 286, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 588, + 286, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 588, + 286, + 624 + ], + "type": "text", + "content": " contains the indices of the " + }, + { + "bbox": [ + 46, + 588, + 286, + 624 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 588, + 286, + 624 + ], + "type": "text", + "content": "-nearest neural points (" + }, + { + "bbox": [ + 46, + 588, + 286, + 624 + ], + "type": "inline_equation", + "content": "k = 8" + }, + { + "bbox": [ + 46, + 588, + 286, + 624 + ], + "type": "text", + "content": " in our case) of " + }, + { + "bbox": [ + 46, + 588, + 286, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 588, + 286, + 624 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 588, + 286, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 588, + 286, + 624 + ], + "type": "text", + "content": " within a radius " + }, + { + "bbox": [ + 46, + 588, + 286, + 624 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 588, + 286, + 624 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 588, + 286, + 624 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 588, + 286, + 624 + ], + "type": "text", + "content": " is a weight based on inverse point distance:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 118, + 632, + 286, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 632, + 286, + 658 + ], + "spans": [ + { + "bbox": [ + 118, + 632, + 286, + 658 + ], + "type": "interline_equation", + "content": "w (\\mathbf {x}, \\mathbf {p} _ {i}) = \\frac {1}{\\left\\| \\mathbf {x} - \\mathbf {p} _ {i} \\right\\| _ {2}} \\tag {3}", + "image_path": "4363e9803575c60799ae73453d5c40581c8b7887e1d34ff2e3e7d4fe6470c3c9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 666, + 286, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 286, + 690 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 286, + 690 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 46, + 666, + 286, + 690 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 666, + 286, + 690 + ], + "type": "text", + "content": " is the sum of weights in the neighborhood. Then, the final radiance " + }, + { + "bbox": [ + 46, + 666, + 286, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 666, + 286, + 690 + ], + "type": "text", + "content": " and density " + }, + { + "bbox": [ + 46, + 666, + 286, + 690 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 666, + 286, + 690 + ], + "type": "text", + "content": " are obtained as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 127, + 701, + 286, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 701, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 127, + 701, + 286, + 714 + ], + "type": "interline_equation", + "content": "(\\mathbf {r}, \\sigma) = F _ {\\psi} (\\mathbf {h}, \\mathbf {d}) \\tag {4}", + "image_path": "0feae81b9d6c00891337987e9986a23bd1707bd1467d925295df56779d20a0be.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "content": "based on features " + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "content": " and view direction " + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "inline_equation", + "content": "\\mathbf{d} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "content": ". For computing " + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "content": " we created a custom PyTorch extension that implements an efficient, batch-wise voxel grid for ray sample " + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "content": "-nn queries in CUDA. The module is based on the Point-NeRF code [52] and will be made publicly available with the rest of the code." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 145, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 145, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 545, + 252 + ], + "type": "text", + "content": "Note that the rendering network is purely local, only receiving relative coordinates between ray samples and neural points. Thus, it is only able to learn a local surface model during training and no global, category-level information. The shared features " + }, + { + "bbox": [ + 304, + 145, + 545, + 252 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 304, + 145, + 545, + 252 + ], + "type": "text", + "content": " are a key ingredient in this formulation. In our studies, we found that they help to train a high-quality category-level neural point renderer and that they represent a density/radiance template of the category. We refer to the supplemental material for further discussion." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 264, + 436, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 264, + 436, + 276 + ], + "spans": [ + { + "bbox": [ + 306, + 264, + 436, + 276 + ], + "type": "text", + "content": "3.3. Training and Inference" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 283, + 545, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 545, + 355 + ], + "type": "text", + "content": "For training and inference, we adopt the autodecoder framework [35, 5], thus, optimizing embeddings using gradient descent instead of predicting them using an encoder. Recent research has shown that approaches based on such test-time optimization have better capabilities when it comes to accurately representing the given observation [52]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "text", + "content": "As input during training, we assume multi-view renderings of a large set of objects, including camera poses for each view. For simplicity, we omit camera poses from the following formulations. Let " + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = f_{\\theta, \\psi}(\\mathbf{P}, \\mathbf{S}, \\mathbf{A}, \\mathbf{E})" + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "text", + "content": " denote the function that renders an image " + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "text", + "content": " from a neural point cloud with positions " + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "text", + "content": ", embeddings " + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "text", + "content": ", attention scores " + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "text", + "content": ", and shared features " + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 304, + 356, + 545, + 439 + ], + "type": "text", + "content": " from a given camera pose." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "text", + "content": "Training Given a dataset " + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{P}_i, \\mathbf{I}_{i,j})\\}_{i = j = 1}^{K,W}" + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "text", + "content": " objects of one category with coherent point clouds " + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i" + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "text", + "content": " renderings " + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_{i,j}\\}_{j = 1}^{W}" + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "text", + "content": " from random views for each object " + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "text", + "content": ", our training procedure jointly optimizes embeddings " + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_i" + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "text", + "content": " (omitted here for simplicity), attention scores " + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "text", + "content": ", shared point features " + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "text", + "content": ", and rendering network parameters " + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\theta, \\psi" + }, + { + "bbox": [ + 304, + 459, + 545, + 533 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 544, + 544, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 544, + 544, + 580 + ], + "spans": [ + { + "bbox": [ + 315, + 544, + 544, + 580 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {A}}, \\hat {\\mathbf {S}}, \\hat {\\theta}, \\hat {\\psi} = \\underset {\\mathbf {A}, \\mathbf {S}, \\theta , \\psi} {\\arg \\min } \\sum_ {i, j} \\mathcal {L} \\left(f _ {\\theta , \\psi} \\left(\\mathbf {P} _ {i}, \\mathbf {S}, \\mathbf {A}, \\mathbf {E} _ {i}\\right), \\mathbf {I} _ {i, j}\\right), \\tag {5}", + "image_path": "56a3c1dd1d4cfac3985c1112e62428688e2dac5d5b237caedd9d2b0e7b24b957.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "content": " is chosen as Mean Squared Error (MSE). After training, we freeze point renderer parameters " + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "inline_equation", + "content": "(\\hat{\\theta},\\hat{\\psi})" + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "content": ", shared features " + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}}" + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "content": " and category-level symmetries " + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}" + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "content": ", which are used to guide gradients to the correct embeddings during test-time optimization." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 663, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 663, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 663, + 545, + 713 + ], + "type": "text", + "content": "Inference During inference, we assume that the camera pose is known. Given one or multiple views " + }, + { + "bbox": [ + 304, + 663, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_j\\}_{j=1}^W" + }, + { + "bbox": [ + 304, + 663, + 545, + 713 + ], + "type": "text", + "content": " and the coherent point cloud " + }, + { + "bbox": [ + 304, + 663, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 663, + 545, + 713 + ], + "type": "text", + "content": ", we can find embeddings " + }, + { + "bbox": [ + 304, + 663, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{E}}" + }, + { + "bbox": [ + 304, + 663, + 545, + 713 + ], + "type": "text", + "content": " by test-time optimization:" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8844" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 72, + 545, + 175 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 545, + 175 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 545, + 175 + ], + "type": "image", + "image_path": "0ef6299d438e900c485466324a21b1af2fb5fe4fbfad088716b12ea281bfee32.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 182, + 548, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 548, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 548, + 232 + ], + "type": "text", + "content": "Figure 3: Coherent point cloud prediction. An MLP with bottleneck is used to enforce the point cloud to be constructed as a low-rank deformation of a template (second to last layer output). During training, all trainable parameters " + }, + { + "bbox": [ + 46, + 182, + 548, + 232 + ], + "type": "inline_equation", + "content": "(\\square, \\square)" + }, + { + "bbox": [ + 46, + 182, + 548, + 232 + ], + "type": "text", + "content": " are optimized using ground-truth point cloud supervision. During inference, the embedding " + }, + { + "bbox": [ + 46, + 182, + 548, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 182, + 548, + 232 + ], + "type": "inline_equation", + "content": "(\\square)" + }, + { + "bbox": [ + 46, + 182, + 548, + 232 + ], + "type": "text", + "content": " can be optimized using different supervision signals: a ResNet predicting a point cloud from single image, mask, or depth." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 260, + 287, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 260, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 77, + 260, + 287, + 285 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {E}} = \\underset {\\mathbf {E}} {\\arg \\min} \\sum_ {i} \\mathcal {L} \\left(f _ {\\hat {\\theta}, \\hat {\\psi}} (\\mathbf {P}, \\hat {\\mathbf {S}}, \\hat {\\mathbf {A}}, \\mathbf {E}), \\mathbf {I} _ {i}\\right). \\qquad (6)", + "image_path": "1746d92b8740658136c3f9d2972f5041f13125ca9df5a64fff9e3a04b9332149.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 290, + 287, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 290, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 47, + 290, + 287, + 302 + ], + "type": "text", + "content": "Afterwards, the example can be rendered from other views." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 308, + 266, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 266, + 320 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 266, + 320 + ], + "type": "text", + "content": "3.4. Coherent Point Clouds from Single Image" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 326, + 287, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 287, + 409 + ], + "type": "text", + "content": "Until now, coherent point clouds were assumed as given for the sake of simplicity. Still, they also have to be obtained from a single image during test time and are a crucial part of our method. We separate the point cloud prediction from the rest of SimNP such that it can be tackled independently. Note that we do not optimize point positions based on rendering losses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 410, + 288, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 288, + 613 + ], + "type": "text", + "content": "The full architecture for point cloud prediction is shown in Figure 3. Just like for our rendering branch, we follow the autodecoder framework in order to allow for flexible supervision at test time. However, unlike the local neural point features for capturing fine details, we opt for a global representation of the point cloud in form of a single learnable latent code " + }, + { + "bbox": [ + 46, + 410, + 288, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^l" + }, + { + "bbox": [ + 46, + 410, + 288, + 613 + ], + "type": "text", + "content": ". During training, we optimize these embeddings jointly together with an MLP decoder " + }, + { + "bbox": [ + 46, + 410, + 288, + 613 + ], + "type": "inline_equation", + "content": "D_{\\phi}: \\mathbb{R}^l \\to \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 46, + 410, + 288, + 613 + ], + "type": "text", + "content": " predicting the neural point coordinates in canonical space, given by the normalized orientation and position of ShapeNet objects. The output of the second to last layer is a low-dimensional bottleneck in order to serve as a low-rank representation of a high-dimensional point template from " + }, + { + "bbox": [ + 46, + 410, + 288, + 613 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 46, + 410, + 288, + 613 + ], + "type": "text", + "content": " [18]. The global representation together with this low-rank regularization allows us to always obtain coherent point clouds irrespective of the final supervision signal." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 613, + 289, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 289, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 289, + 649 + ], + "type": "text", + "content": "Given point clouds " + }, + { + "bbox": [ + 46, + 613, + 289, + 649 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{P}_i\\}_{i=1}^K" + }, + { + "bbox": [ + 46, + 613, + 289, + 649 + ], + "type": "text", + "content": " of training examples, we train " + }, + { + "bbox": [ + 46, + 613, + 289, + 649 + ], + "type": "inline_equation", + "content": "D_{\\phi}" + }, + { + "bbox": [ + 46, + 613, + 289, + 649 + ], + "type": "text", + "content": " and latent codes " + }, + { + "bbox": [ + 46, + 613, + 289, + 649 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_i\\}_{i=1}^K" + }, + { + "bbox": [ + 46, + 613, + 289, + 649 + ], + "type": "text", + "content": " to minimize the 3D Chamfer Distance (CD):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 653, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 653, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 287, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C D} = \\sum_ {i = 1} ^ {K} \\mathrm {C D} \\left(D _ {\\phi} \\left(\\mathbf {z} _ {i}\\right), \\mathcal {P} _ {i}\\right). \\tag {7}", + "image_path": "c7f855253537986bb455fd81ee161eb13a8747da005de9a487c896f3d4161f3a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "As one form of point cloud supervision during test time, we jointly train a ResNet18 [19] encoder with the image," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 247, + 547, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 547, + 342 + ], + "type": "text", + "content": "segmentation mask, and camera ray encodings [48] as input to predict the same point cloud after decoding with " + }, + { + "bbox": [ + 304, + 247, + 547, + 342 + ], + "type": "inline_equation", + "content": "D_{\\phi}" + }, + { + "bbox": [ + 304, + 247, + 547, + 342 + ], + "type": "text", + "content": ", by minimizing the MSE. At test time, we optimize the latent codes while keeping encoder and decoder fixed. At first sight, it may seem counterintuitive to bring back an encoder into the autodecoder framework. However, by parameterizing the latent code directly, we gain flexibility regarding additional supervisory signals." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 342, + 547, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 488 + ], + "type": "text", + "content": "Two optional sources of additional point cloud supervision at test time are the segmentation mask and a depth map, e.g., captured by an RGB-D camera. For the mask, we use a 2D Chamfer loss between its pixel coordinates and the 2D projection of the point cloud. Depth maps can be utilized by projecting the individual depth samples into the world frame and employing an asymmetric 3D Chamfer loss, i.e., minimizing the distance between each sample and its nearest point in the point cloud. Note that although occluded points from the perspective of the depth map are not explicitly supervised by this optional loss, they are still optimized implicitly because of our global latent representation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 487, + 547, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 547, + 584 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 547, + 584 + ], + "type": "text", + "content": "Besides leveraging these additional forms of data, test-time optimization of the point cloud is directly compatible with multi-view supervision. Given multiple views, the parameterized latent code " + }, + { + "bbox": [ + 304, + 487, + 547, + 584 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 487, + 547, + 584 + ], + "type": "text", + "content": " fuses the point clouds predicted individually for each image by the encoder. The following experimental results demonstrate that our neural point representation can benefit significantly from flexible point cloud supervision at test time." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 597, + 450, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 450, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 450, + 611 + ], + "type": "text", + "content": "4. Experiments and Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "In this section, we describe experiments made with SimNP and present our results. The goal of the experiments is to provide evidence for the following statements: SimNP (1) improves on previous approaches in reconstructing unseen symmetric object parts (c.f. Section 4.2), (2) learns correct self-similarities that respect symmetries of the object category (c.f. Section 4.3), (3) provides a meaningful representation that can be used for interpolation" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8845" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 525, + 275 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 525, + 275 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 525, + 275 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 525, + 275 + ], + "type": "table", + "html": "
Cat.Method1 Input View1 Input View (Sym.)2 Input Views
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
CarsSRN [41]22.220.8930.10722.050.8930.10624.820.9200.093
PixelNeRF [54]23.170.9050.11222.420.8950.12225.670.9350.083
CodeNeRF1 [22]23.800.910.106—1—1—125.710.930.099
FE-NVS2 [17]22.830.9060.099—2—2—224.640.9270.085
VisionNeRF [29]22.880.9080.08522.070.8980.091—3—3—3
Ours23.000.9110.08122.670.9100.08426.690.9480.055
Ours + GT PC23.690.9200.07723.310.9160.08027.290.9530.051
ChairsSRN [41]22.320.9040.09222.110.9010.09525.720.9350.072
PixelNeRF [54]23.720.9130.10123.150.9060.10926.210.9410.070
CodeNeRF1 [22]23.660.900.136—1—1—125.630.910.129
FE-NVS2 [17]23.210.9200.077—2—2—225.250.9400.065
VisionNeRF [29]24.480.9300.07723.880.9250.085—3—3—3
Ours23.460.9180.08123.030.9120.08827.360.9530.057
Ours + GT PC25.800.9400.07125.240.9340.07728.430.9610.053
", + "image_path": "ae63750fd7db16d1cc21b304f7e9081b2e2d81eba47c191230980a822694fcdf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 283, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 283, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 545, + 357 + ], + "type": "text", + "content": "Table 1: Novel view synthesis on ShapeNet cars and chairs. Our approach achieves state-of-the-art or competitive performance in single-view reconstruction and outperforms previous baselines by a large margin in the two-view setting. While PixelNeRF achieves a higher PSNR with single-view input, it can be seen that their results are blurry (c.f. Fig. 4a). Additionally to the standard setup, (Sym.) provides results on all target views that show mostly the opposite object side to the input view. We show results of our method using point cloud supervision (Ours + GT PC) at test time. These results are only to identify the gap induced by errors in point cloud prediction. The results shown in the rest of the paper use the Ours setup." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 367, + 288, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 367, + 288, + 428 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 288, + 428 + ], + "type": "text", + "content": "(c.f. Section 4.4), (4) is a very efficient approach (c.f. Section 4.5), and (5) is compatible with test-time pose optimization (c.f. Section 4.6). We provide additional results in the appendix, such as more qualitative results, ablation studies, and an analysis of our point cloud prediction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 447, + 164, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 447, + 164, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 447, + 164, + 460 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 470, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 470, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 287, + 638 + ], + "type": "text", + "content": "We conduct experiments for category-specific novel view synthesis, given one or two input views. We make use of the ShapeNet dataset provided by SRN [41], which comprises 3514 cars and 6591 chairs split into training, validation, and test sets. Each training instance has been rendered from 50 random camera locations on a sphere around the object. The test sets have 704 and 1318 examples, respectively, all with the same 251 views from an Archimedean spiral and the same lighting as during training. As done by all baselines, view 64 and additionally view 104 are used as input for single- and two-view experiments. The image resolution is " + }, + { + "bbox": [ + 46, + 470, + 287, + 638 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 470, + 287, + 638 + ], + "type": "text", + "content": ". We use PSNR, SSIM, and LPIPS [56] to compare our approach with SRN [41], PixelNeRF [54], FE-NVS [17], and VisionNeRF [29]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "We use 512/4096 points and 512/128 32-dim. embeddings per object. The shared features are of dimensionality 64. After separate pretraining of the point cloud prediction, we train for " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "0.9 / 1.2\\mathrm{M}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " iterations. At test time, the instancespecific parameters are optimized for " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "10\\mathrm{k}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " iterations each. Please refer to the suppl. mat. for additional information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 367, + 504, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 367, + 504, + 380 + ], + "spans": [ + { + "bbox": [ + 305, + 367, + 504, + 380 + ], + "type": "text", + "content": "4.2. Single- and Two-View Reconstruction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 385, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 493 + ], + "type": "text", + "content": "Table 1 summarizes our quantitative results. We leverage three different test-time optimization setups. In our main setup (Ours), we use the ResNet and the mask for test time optimization of point clouds to ensure a fair comparison with previous approaches. This setup is used for all comparisons and qualitative results shown. Furthermore, we report results using ground-truth point clouds (Ours + GT PC) and provide results with additional ground-truth depth supervision in the supplement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 494, + 546, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 546, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 546, + 649 + ], + "type": "text", + "content": "In single-view reconstruction of cars, SimNP outperforms the state-of-the-art in SSIM and LPIPS. We attribute the small PSNR advantage of PixelNeRF to the blurriness of its reconstructions visible in Fig. 4a, which is confirmed by higher LPIPS. Also, we outperform PixelNeRF in inferring symmetric regions, as we show in columns 3-6 in Table 1 and in Figure 4b. SRN lacks details due to its global representation. PixelNeRF's reconstructions are overly-blurry in unseen regions, because pixel-aligned features are shared along the same rays of the input view without respecting the category-level self-similarities. VisionNeRF is not able to consistently transfer uncommon local color variations to symmetric object regions, which indicates a too strong" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 654, + 545, + 673 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 654, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 306, + 654, + 545, + 673 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 306, + 654, + 545, + 673 + ], + "type": "text", + "content": "CodeNeRF [22] does not provide trained models. Note that the input views differ from the usual setup." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 673, + 545, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 673, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 306, + 673, + 545, + 703 + ], + "type": "text", + "content": "2FE-NVS [17] does not have public code. Thus, shown numbers are taken from the paper/from authors, where available. SRN, PixelNeRF and VisionNeRF are evaluated using the PixelNeRF calc.metrics script." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 703, + 506, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 703, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 703, + 506, + 713 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 317, + 703, + 506, + 713 + ], + "type": "text", + "content": "VisionNeRF [29] not applicable for multiple input views." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8846" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 74, + 291, + 357 + ], + "blocks": [ + { + "bbox": [ + 50, + 74, + 291, + 357 + ], + "lines": [ + { + "bbox": [ + 50, + 74, + 291, + 357 + ], + "spans": [ + { + "bbox": [ + 50, + 74, + 291, + 357 + ], + "type": "image", + "image_path": "3cc050c691fc04980ca50010c96171c1d7f56792ae0c0fa8d1cf30516915fb95.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 363, + 253, + 374 + ], + "lines": [ + { + "bbox": [ + 91, + 363, + 253, + 374 + ], + "spans": [ + { + "bbox": [ + 91, + 363, + 253, + 374 + ], + "type": "text", + "content": "(a) Qualitative single-view reconstruction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 301, + 74, + 545, + 361 + ], + "blocks": [ + { + "bbox": [ + 301, + 74, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 301, + 74, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 301, + 74, + 545, + 361 + ], + "type": "image", + "image_path": "ec42be55d01d4a3b75a0611f326640cd852e0fcb1c37ed2e9a86ad09ecee9fd3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 364, + 462, + 374 + ], + "lines": [ + { + "bbox": [ + 381, + 364, + 462, + 374 + ], + "spans": [ + { + "bbox": [ + 381, + 364, + 462, + 374 + ], + "type": "text", + "content": "(b) Metrics per view." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 384, + 547, + 445 + ], + "lines": [ + { + "bbox": [ + 46, + 384, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 547, + 445 + ], + "type": "text", + "content": "Figure 4: a) Our method enables more detailed reconstructions and can better transfer appearance information to symmetric regions compared to SRN, PixelNeRF, and VisionNeRF. b) We show metric comparisons for each target view in the 251-view SRN test spiral for single-view reconstruction of cars. As visible, our overall better performance can be attributed to views showing regions symmetric to the input view (green areas and example views 10, 84, 175). Also, the related object-level method SRN shows rather flat curves indicating a weak adaptation to observations. This is not the case for SimNP." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 460, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 533 + ], + "type": "text", + "content": "global prior. In contrast, as visible from Figure 4a and Figure 1, our approach correctly transfers local patterns by leveraging learned category-level self-similarities. As visible from Figure 4b and columns 3-6 in Table 1, it significantly outperforms all baselines for views on regions with symmetric counterparts in the input image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 628 + ], + "type": "text", + "content": "Compared to cars, the chairs category exhibits a larger variety in geometry but less texture details, leading to a smaller advantage through the self-similarity prior. Still, SimNP performs competitively compared to the baselines. Moreover, our strong results under the assumption of a given ground-truth point cloud (last rows in Table 1) indicate further potential of our method in combination with an improved point cloud prediction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "In the two-view setting, our approach surpasses previous methods by a large margin on both categories. Here, the benefit comes not only from modeling self-similarities but also from our local point-based representation. The results show that it is better suited for fusing multiple observations than pixel-aligned methods, while at the same time being able to fit details. Looking at the qualitative results in Fig. 5," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 460, + 546, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 546, + 510 + ], + "type": "text", + "content": "we can observe detailed reconstructions. In Tab. 2, we compare different two view setup against each other. The same side setup, which does observe one side of the object with two views, already clearly improves the results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 515, + 446, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 446, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 446, + 528 + ], + "type": "text", + "content": "4.3. Learned Self-Similarities" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 533, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 714 + ], + "type": "text", + "content": "As SimNP learns category-level self-similarities explicitly, we can directly inspect them to gain more insights. For visualizing the attention between neural points and embeddings, we proceed as follows. The rendering network computes weights between ray samples and neural points (see Eq. 3). By treating these weights just like RGB channels during ray marching, we obtain the influence of each neural point on each pixel. Finally, we multiply the learned category-level attention scores of a single embedding to the respective neural point influence, resulting in the self-similarity visualizations given in Figure 6. We describe the visualization procedure in detail in the supplemental material. Although we use as many embeddings as neural points (for cars) such that the attention could learn a one-to-one mapping, the representation learns to share information" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8847" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 289, + 188 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 289, + 188 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 289, + 188 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 289, + 188 + ], + "type": "table", + "html": "
Cat.View Idx.PSNR↑SSIM↑LPIPS↓
Carssingle view (64)23.000.9110.081
2 v. close (63,64)23.390.9160.078
2 v. same side (44,64)25.760.9380.064
2 v. opposite (64,104)26.690.9480.055
Chairssingle view (64)23.460.9180.081
2 v. close (63,64)24.240.9240.077
2 v. same side (44,64)26.500.9440.064
2 v. opposite (64,104)27.360.9530.057
", + "image_path": "4f8dc54a567e14f3b72f4f983da2fcfd0cce4d4049f6394b2bbac0eb2d0c0f78.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 195, + 288, + 268 + ], + "lines": [ + { + "bbox": [ + 47, + 195, + 288, + 268 + ], + "spans": [ + { + "bbox": [ + 47, + 195, + 288, + 268 + ], + "type": "text", + "content": "Table 2: Effect of input views. We evaluate different two view setups, comparing two very close views, 2 views from the same side of the object (approx. 90 degrees apart) and 2 views from opposite sides of the object. It can be seen that even if one side of the object is not seen (same side setup), the results improve strongly with 2 views." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 275, + 282, + 545 + ], + "blocks": [ + { + "bbox": [ + 50, + 275, + 282, + 545 + ], + "lines": [ + { + "bbox": [ + 50, + 275, + 282, + 545 + ], + "spans": [ + { + "bbox": [ + 50, + 275, + 282, + 545 + ], + "type": "image", + "image_path": "7054eeb93a0cb8380066b6cf641eeebf12c0ccb9ab315ab96635e769e2cc82e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 556, + 287, + 583 + ], + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 583 + ], + "type": "text", + "content": "Figure 5: Two-view reconstruction. SimNP learns a high-quality 3D object representation given only two input views." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 599, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 287, + 648 + ], + "type": "text", + "content": "between similar points. Further, we observe that the connections for all embeddings are symmetric with respect to a reflection plane, indicating that the method successfully learned a prior about the plane symmetry of each category." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 658, + 228, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 228, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 228, + 672 + ], + "type": "text", + "content": "4.4. Meaningful Representation Space" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Figure 7 provides results for interpolating the point cloud latent code and/or the embeddings of two instances obtained by multi-view fitting. SimNP learns a meaningful, disentan-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 324, + 73, + 341, + 82 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 73, + 341, + 82 + ], + "spans": [ + { + "bbox": [ + 324, + 73, + 341, + 82 + ], + "type": "text", + "content": "RGB" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 391, + 72, + 505, + 82 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 72, + 505, + 82 + ], + "spans": [ + { + "bbox": [ + 391, + 72, + 505, + 82 + ], + "type": "text", + "content": "Attention Scores for 4 Embeddings" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 91, + 544, + 430 + ], + "blocks": [ + { + "bbox": [ + 310, + 91, + 544, + 430 + ], + "lines": [ + { + "bbox": [ + 310, + 91, + 544, + 430 + ], + "spans": [ + { + "bbox": [ + 310, + 91, + 544, + 430 + ], + "type": "image", + "image_path": "5ed381c504ba0c5365fd8213255d3e202debec8aee012a6737fd473ad6ba919e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 441, + 547, + 528 + ], + "lines": [ + { + "bbox": [ + 304, + 441, + 547, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 547, + 528 + ], + "type": "text", + "content": "Figure 6: Attention visualization. We render the influence of four different embeddings per category (one per column) on each ray, and visualize them on eight different examples. It becomes evident that embeddings specialize on self-similar local areas, e.g., wheels, or front lights. All embeddings show plane symmetric patterns, showing that our method recovers the symmetry of the category." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 542, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 547, + 616 + ], + "type": "text", + "content": "gled representation space allowing for smooth shape and/or appearance transitions from one object to another. This is in contrast to pixel-aligned methods, which are not suited for interpolation. As an example, Fig. 8 shows a comparison with VisionNeRF, for which interpolation of the intermediate feature maps results in pixel space interpolation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 623, + 443, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 443, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 443, + 635 + ], + "type": "text", + "content": "4.5. Efficient Representation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "SimNP renders a frame in 59ms, which is more than an order of magnitude faster than the pixel-aligned methods PixelNeRF (2.116s) and VisionNeRF (2.700s). We cannot fairly compare against FE-NVS, as no public code is available. The evaluation they provide in the paper is using amortized rendering, which is not comparable, as it only" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8848" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 79, + 282, + 310 + ], + "blocks": [ + { + "bbox": [ + 50, + 79, + 282, + 310 + ], + "lines": [ + { + "bbox": [ + 50, + 79, + 282, + 310 + ], + "spans": [ + { + "bbox": [ + 50, + 79, + 282, + 310 + ], + "type": "image", + "image_path": "e84b33e4c40f11fbb435fdd7cc45fe4ccfc747846e90d21283f9dc319a2451dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 322, + 288, + 382 + ], + "lines": [ + { + "bbox": [ + 46, + 322, + 288, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 288, + 382 + ], + "type": "text", + "content": "Figure 7: Disentangled interpolation. Our point-based representation disentangles the basic shape from appearance. Point cloud latent code " + }, + { + "bbox": [ + 46, + 322, + 288, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 322, + 288, + 382 + ], + "type": "text", + "content": " and embeddings " + }, + { + "bbox": [ + 46, + 322, + 288, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 46, + 322, + 288, + 382 + ], + "type": "text", + "content": " can be interpolated independently resulting in a smooth shape and/or appearance transition between different objects." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 399, + 286, + 451 + ], + "blocks": [ + { + "bbox": [ + 52, + 399, + 286, + 451 + ], + "lines": [ + { + "bbox": [ + 52, + 399, + 286, + 451 + ], + "spans": [ + { + "bbox": [ + 52, + 399, + 286, + 451 + ], + "type": "image", + "image_path": "dac93f30b97fd72df9e25c613a5b42d53f45afd434b5dbf5044661400cc0214c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 462, + 287, + 510 + ], + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 510 + ], + "type": "text", + "content": "Figure 8: Interpolation comparison. Unlike VisionNeRF, which sticks to interpolation in pixel space for views similar to the input perspective, SimNP enables semantically meaningful interpolation like SRN but with more details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 532, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 639 + ], + "type": "text", + "content": "works on many views simultaneously. However, we suspect that their method is also very fast. We train our method on a single A40 GPU for " + }, + { + "bbox": [ + 46, + 532, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\approx 1.6" + }, + { + "bbox": [ + 46, + 532, + 287, + 639 + ], + "type": "text", + "content": " days. In comparison, SRN and PixelNeRF take 6 days on an RTX 8000 or Titan RTX, respectively, FE-NVS 2.5 days on 30 V100, and VisionNeRF around 5 days on 16 A100 GPUs, according to the authors. Overall, our method is very efficient with resource-efficient training and fast rendering that shows potential to be applied to larger scale reconstruction in future work." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 647, + 155, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 155, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 155, + 659 + ], + "type": "text", + "content": "4.6. Pose Optimization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "To alleviate the assumption of a known camera pose at test time, we investigate optimization of camera parameters. More precisely, for each test example, we initialize eight evenly distributed camera locations on the sphere around" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 535, + 187 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 535, + 187 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 535, + 187 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 535, + 187 + ], + "type": "image", + "image_path": "16c4941e9dfbbd2c2214d8c208140d84c46f345575c75ee9218d630435d6243e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 195, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 305, + 195, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 195, + 545, + 232 + ], + "type": "text", + "content": "Figure 9: Pose Optimization. SimNP is compatible with camera pose optimization. The figure compares single-view reconstruction results with and without given camera pose." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 256, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 545, + 424 + ], + "type": "text", + "content": "the object. The rotation matrix is always defined by the " + }, + { + "bbox": [ + 304, + 256, + 545, + 424 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 256, + 545, + 424 + ], + "type": "text", + "content": "-axis pointing towards the sphere center and the " + }, + { + "bbox": [ + 304, + 256, + 545, + 424 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 256, + 545, + 424 + ], + "type": "text", + "content": "-axis pointing upwards. At test time, we first initialize the point cloud latent using an encoder without ray encodings. Then, the pose in form of the camera location on the sphere is optimized by minimizing the Chamfer loss between the mask's pixel coordinates and the 2D projection of the point cloud. Lastly, we further finetune the point cloud with the Chamfer loss and optimize the embeddings for reconstruction of the input image. We select the pose that leads to the best reconstruction based on LPIPS against the single input image. Fig. 9 shows a comparison given vs. optimized camera pose. We achieve competitive performance (0.116 LPIPS) compared with CodeNeRF (0.114) without camera poses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 440, + 378, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 440, + 378, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 440, + 378, + 452 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 461, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 581 + ], + "type": "text", + "content": "We presented SimNP, the first object-level representation based on neural radiance fields, utilizing shared attention scores to learn category-specific self-similarities. Our method reaches state-of-the-art quality in single- and two-view reconstruction, while being highly efficient. Overall, it achieves a better data prior vs. observation trade-off. The improvements can be contributed to (1) leveraging a local neural point radiance field for the representation of object categories and (2) correctly propagating information between similar regions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Limitations and Future Work The main limitation of SimNP is the assumption of a canonical space with ground-truth point clouds during training, which prohibits direct application on in-the-wild datasets. Therefore, an improved point cloud prediction in camera frame could be an exciting path for future work. Furthermore, we believe that it is promising to relax the point identities and to apply our self-similarity priors on a scene level to obtain large-scale, prior-driven reconstruction." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8849" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 133 + ], + "type": "text", + "content": "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In European Conhference on Computer Vision (ECCV), 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 286, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 286, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 286, + 179 + ], + "type": "text", + "content": "[2] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J. Mitra, and Paul Guerrero. RenderDiffusion: Image diffusion for 3D reconstruction, inpainting and generation. arXiv pre-print, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 286, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 286, + 213 + ], + "type": "text", + "content": "[3] François ARJ, Medioni GG, and Waupotitsch R. Mirror symmetry: 2-view stereo geometry. In Image and Vision Computing, 2003." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 214, + 286, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 286, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 286, + 257 + ], + "type": "text", + "content": "[4] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-NeRF: A multiscale representation for antiailiasing neural radiance fields. 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 286, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 286, + 312 + ], + "type": "text", + "content": "[5] Rohan Chabra, Jan E. Lenssen, Eddy Ilg, Tanner Schmidt, Julian Straub, Steven Lovegrove, and Richard Newcombe. Deep local shapes: Learning local SDF priors for detailed 3D reconstruction. In European Conhference on Computer Vision (ECCV), 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 314, + 286, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 286, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 286, + 379 + ], + "type": "text", + "content": "[6] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 286, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 286, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 286, + 424 + ], + "type": "text", + "content": "[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. Int. Conference on Computer Vision (ICCV), 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 286, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 286, + 458 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 286, + 458 + ], + "type": "text", + "content": "[8] Xin Chen, Yuwei Li, Xi Luo, Tianjia Shao, Jingyi Yu, Kun Zhou, and Youyi Zheng. Autosweep: Recovering 3d editable objects from a single photograph. In ACM Trans. Gr., 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 459, + 286, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 459, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 459, + 286, + 491 + ], + "type": "text", + "content": "[9] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 492, + 286, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 286, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 286, + 545 + ], + "type": "text", + "content": "[10] Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (SRF): Learning view synthesis from sparse views of novel scenes. In Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 547, + 286, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 286, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 286, + 591 + ], + "type": "text", + "content": "[11] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3D-R2N2: A unified approach for single and multi-view 3d object reconstruction. In European Conhference on Computer Vision (ECCV), 2016." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 286, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 286, + 668 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Int. Conference on Learning Representations (ICLR), 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "text", + "content": "[13] Haoqiang Fan, Hao Su, and Leonidas Guibas. A point set generation network for 3d object reconstruction from a single image. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[14] Roger Fawcett, Andrew Zisserman, and Michael Brady. Extracting structure from an affine view of a 3d point set with one or two bilateral symmetries. In Image and Vision Computing." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[15] Alexandre François, Gérard Medioni, and Roman Waupotitsch. Reconstructing mirror symmetric scenes from a single view using 2-view stereo geometry. In ICPR: Proceedings of the 16th International Conference on Pattern Recognition, 2002." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 172, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 205 + ], + "type": "text", + "content": "[16] Rohit Girdhar, David F. Fouhey, Mikel D. Rodriguez, and Abhinav Kumar Gupta. Learning a predictable and generative vector representation for objects. 2016." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 249 + ], + "type": "text", + "content": "[17] Pengsheng Guo, Miguel Angel Bautista, Alex Colburn, Liang Yang, Daniel Ulbricht, Joshua M. Susskind, and Qi Shan. Fast and explicit neural view synthesis. In Winter Conference on Applications of Computer Vision (WACV), 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 250, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 545, + 304 + ], + "type": "text", + "content": "[18] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 304, + 545, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 347 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 349, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 545, + 381 + ], + "type": "text", + "content": "[20] Du Q. Huynh. Affine reconstruction from monocular vision in the presence of a symmetry plane. In Int. Conference on Computer Vision (ICCV), 1999." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 382, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 425 + ], + "type": "text", + "content": "[21] Eldar Insafutdinov, Dylan Campbell, João F Henriques, and Andrea Vedaldi. SNeS: Learning probably symmetric neural surfaces from incomplete data. In European Conference on Computer Vision (ECCV), 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "type": "text", + "content": "[22] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Int. Conference on Computer Vision (ICCV), 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "type": "text", + "content": "[23] M. Johari, Y. Lepoittevin, and F. Fleuret. GeoNeRF: Generalizing nef with geometry priors. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 493, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 545, + 536 + ], + "type": "text", + "content": "[24] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In European Conhference on Computer Vision (ECCV), 2018." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 537, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 569 + ], + "type": "text", + "content": "[25] Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. In European Conhference on Computer Vision (ECCV), 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 570, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 613 + ], + "type": "text", + "content": "[26] Andrey Kurenkov, Jingwei Ji, Animesh Garg, Viraj Mehta, JunYoung Gwak, Chris Choy, and Silvio Savarese. Deform-Net: Free-form deformation network for 3D shape reconstruction from a single image. 2018." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 614, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 657 + ], + "type": "text", + "content": "[27] Kevin Koser, Christopher Zach, and Marc Pollefeys. Dense 3d reconstruction of symmetric scenes from a single image. In Proceedings of the 33rd international conference on Pattern recognition, 2011." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[28] Xingyi Li, Chaoyi Hong, Yiran Wang, Zhiguo Cao, Ke Xian, and Guosheng Lin. Symmnerf: Learning to explore symmetry prior for single-view view synthesis. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 1726-1742, December 2022." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8850" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[29] Kai-En Lin, Lin Yen-Chen, Wei-Sheng Lai, Tsung-Yi Lin, Yi-Chang Shih, and Ravi Ramamoorthi. Vision transformer for NeRF-based view synthesis from a single input image. In Winter Conference on Applications of Computer Vision (WACV), 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "text", + "content": "[30] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In Int. Conference on Neural Information Processing Systems (NIPS), 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 218 + ], + "type": "text", + "content": "[31] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 274 + ], + "type": "text", + "content": "[32] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conhference on Computer Vision (ECCV), 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "type": "text", + "content": "[33] Andrew Zisserman Mukherjee, Dipti Prasad and Michael Brady. Shape from symmetry: Detecting and exploiting symmetry in affine images. In Philosophical Transactions: Physical Sciences and Engineering, 1995." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "type": "text", + "content": "[34] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 431 + ], + "type": "text", + "content": "[35] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 433, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 287, + 465 + ], + "type": "text", + "content": "[36] Cody J. Phillips, Matthieu Lecce, and Kostas Daniilidis. Seeing glassware: from edge detection to pose estimation and shape recovery. In Robotics: Science and Systems, 2016." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "type": "text", + "content": "[37] Ruslan Rakhimov, Andrei-Timotei Ardelean, Victor Lempitsky, and Evgeny Burnaev. NPBG++: Accelerating neural point-based graphics. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "type": "text", + "content": "[38] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "text", + "content": "[39] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 624 + ], + "type": "text", + "content": "[40] Sudipta N. Sinha, Krishnan Ramnath, and Richard Szeliski. Detecting and reconstructing 3d mirror symmetric objects. In European Conhference on Computer Vision (ECCV), 2012." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "text", + "content": "[41] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3D-structure-aware neural scene representations. In Int. Conference on Neural Information Processing Systems (NIPS), 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[42] Maxim Tatarchenko, Stephan R. Richter, Rene Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In Confer-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "ence on Computer Vision and Pattern Recognition (CVPR), 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 118 + ], + "type": "text", + "content": "[43] Sebastian Thurn and Ben Wegbreit. Shape from symmetry. In Int. Conference on Computer Vision (ICCV), 2005." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 119, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 119, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 119, + 545, + 173 + ], + "type": "text", + "content": "[44] Garvita Tiwari, Dimitrije Antic, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Pose-ndf: Modeling human pose manifolds with neural distance fields. In European Conference on Computer Vision (ECCV), October 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 207 + ], + "type": "text", + "content": "[45] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d scene representation and rendering. In Int. Conference on Computer Vision (ICCV), 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 209, + 545, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 242 + ], + "type": "text", + "content": "[46] Shubham Tulsiani, Nilesh Kulkarni, and Abhinav Gupta. Implicit mesh reconstruction from unannotated image collections. In arXiv pre-print, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 243, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 297 + ], + "type": "text", + "content": "[47] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 299, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 299, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 299, + 545, + 342 + ], + "type": "text", + "content": "[48] Daniel Watson, William Chan, Ricardo Martin Brulla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. In Int. Conference on Learning Representations (ICLR), 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 344, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 344, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 344, + 545, + 398 + ], + "type": "text", + "content": "[49] Jiajun Wu, Chengkai Zhang, Tianfan Xue, William T. Freeman, and Joshua B. Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In Int. Conference on Neural Information Processing Systems (NIPS), 2016." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "text", + "content": "[50] Shangzhe Wu, Ameesh Makadia, Jiajun Wu, Noah Snavely, Richard Tucker, and Angjoo Kanazawa. De-rendering the world's revolutionary artefacts. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "type": "text", + "content": "[51] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 490, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 533 + ], + "type": "text", + "content": "[52] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-NeRF: Point-based neural radiance fields. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 535, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 545, + 589 + ], + "type": "text", + "content": "[53] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In Int. Conference on Neural Information Processing Systems (NIPS), 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 590, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 633 + ], + "type": "text", + "content": "[54] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 635, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 668 + ], + "type": "text", + "content": "[55] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv pre-print, 2020." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 714 + ], + "type": "text", + "content": "[56] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Conference on Computer Vision and Pattern Recognition (CVPR), June 2018." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8851" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 118 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 118 + ], + "type": "text", + "content": "[57] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Toch: Spatio-temporal object-to-hand correspondence for motion refinement. In European Conference on Computer Vision (ECCV). Springer, October 2022." + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8852" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Similarity Min-Max_ Zero-Shot Day-Night Domain Adaptation/d2d1b76d-0c03-43ab-96b8-c3ce7c667abf_content_list.json b/2023/Similarity Min-Max_ Zero-Shot Day-Night Domain Adaptation/d2d1b76d-0c03-43ab-96b8-c3ce7c667abf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9f325543e49da887bc1f5cac25b878503659f111 --- /dev/null +++ b/2023/Similarity Min-Max_ Zero-Shot Day-Night Domain Adaptation/d2d1b76d-0c03-43ab-96b8-c3ce7c667abf_content_list.json @@ -0,0 +1,2161 @@ +[ + { + "type": "text", + "text": "Similarity Min-Max: Zero-Shot Day-Night Domain Adaptation", + "text_level": 1, + "bbox": [ + 165, + 130, + 803, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rundong Luo $^{1,2,3}$ Wenjing Wang $^{1}$ Wenhan Yang $^{4}$ Jiaying Liu $^{1*}$", + "bbox": [ + 222, + 179, + 751, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Wangxuan Institute of Computer Technology, Peking University", + "bbox": [ + 225, + 199, + 745, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of EECS, Peking University $^{3}$ School of CS, Peking University $^{4}$ Peng Cheng Laboratory", + "bbox": [ + 91, + 215, + 872, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Low-light conditions not only hamper human visual experience but also degrade the model's performance on downstream vision tasks. While existing works make remarkable progress on day-night domain adaptation, they rely heavily on domain knowledge derived from the task-specific nighttime dataset. This paper challenges a more complicated scenario with border applicability, i.e., zero-shot day-night domain adaptation, which eliminates reliance on any nighttime data. Unlike prior zero-shot adaptation approaches emphasizing either image-level translation or model-level adaptation, we propose a similarity min-max paradigm that considers them under a unified framework. On the image level, we darken images towards minimum feature similarity to enlarge the domain gap. Then on the model level, we maximize the feature similarity between the darkened images and their normal-light counterparts for better model adaptation. To the best of our knowledge, this work represents the pioneering effort in jointly optimizing both levels, resulting in a significant improvement of model generalizability. Extensive experiments demonstrate our method's effectiveness and broad applicability on various nighttime vision tasks, including classification, semantic segmentation, visual place recognition, and video action recognition. Our project page is available at https://red-fairy.github.io/ZeroShotDayNightDA-Webpage/.", + "bbox": [ + 75, + 301, + 473, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 726, + 209, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks are sensitive to insufficient illumination, and such deficiency has posed significant threats to safety-critical computer vision applications. Intuitively, insufficient illumination can be handled by low-light enhancement methods [23, 30, 34, 56, 60, 63], which aim at restoring low-light images to normal-light. However, enhancement models do not necessarily benefit downstream high-level vision tasks as they are optimized for human visual perception", + "bbox": [ + 75, + 752, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d3327f257d5b9e3bce31a1ab221a20fb298904db601444b49884b02a9cc843f3.jpg", + "image_caption": [ + "Figure 1. Left: Illustration of our similarity min-max framework for zero-shot day-night domain adaptation. Right: Our framework achieves state-of-the-art results on multiple downstream high-level vision tasks without seeing real nighttime images during training." + ], + "image_footnote": [], + "bbox": [ + 504, + 271, + 715, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b034ff2126266295ad5e8fbc5a4328efe916752eeb8561921729fa569208c479.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 282, + 883, + 407 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and neglect the need for machine vision.", + "bbox": [ + 500, + 516, + 767, + 531 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Much existing literature has focused on improving machine vision performance at night through domain adaptation. By aligning the distribution statistics between the nighttime and daytime datasets through image translation [2, 12, 45], self-supervised learning [52, 53], or multistage algorithms [10, 46, 47], these methods have greatly improved models' performance in nighttime environments. The primary assumption of domain adaptation is that the target domain data is readily available. Nevertheless, obtaining data from the task-specific target domain may be challenging in extreme practical application scenarios such as deep-space exploration and deep-sea analysis.", + "bbox": [ + 496, + 534, + 892, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To reduce the requirement on target domain data, zero-shot domain adaptation has emerged as a promising research direction, where adaptation is performed without accessing the target domain. Regarding day-night domain adaptation, the primary challenge is learning illumination-robust representations generalizable to both day and night modalities. To accomplish this goal under zero-shot constraints, Lengyel et al. [29] proposed a color invariant convolution for handling illumination changes. Cui et al. [8] designed a Reverse ISP pipeline and generated synthetic nighttime images with pseudo labels. However, image-level methods simply consider synthetic nighttime as pseudo", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8104", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "labeled data and overlook model-level feature extraction; model-level methods focus on adjusting model architecture but neglect image-level nighttime characteristics. Neither is effective enough capture the illumination-robust representations that could bridge the complex day-night domain gap.", + "bbox": [ + 75, + 90, + 467, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From this point of view, we devise a similarity min-max framework that involves two levels, as illustrated in Figure 1. On the image level, we generate a synthetic nighttime domain that shares minimum feature similarity with the daytime domain to enlarge the domain gap. On the model level, we learn illumination-robust representations by maximizing the feature similarity of images from the two domains for better model adaptation.", + "bbox": [ + 75, + 167, + 467, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Intuitive as it seems, solving this bi-level optimization problem is untrivial. Directly solving it may yield unsatisfactory results, e.g., meaningless images filled with zero values or identical features given all inputs. Therefore, we develop a stable training pipeline that can be considered a sequential operation on both the image and the model. Regarding the image, we propose an exposure-guided module to perform reliable and controllable nighttime image synthesis. Regarding the model, we align the representation of images from day and night domains through multi-task contrastive learning. Finally, our model achieves day-night adaptation without seeing real nighttime images.", + "bbox": [ + 75, + 287, + 467, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our framework can serve as a plug-and-play remedy to existing daytime models. To verify its effectiveness, we conduct extensive experiments on multiple high-level nighttime vision tasks, including classification, semantic segmentation, visual place recognition, and video action recognition. Results on various benchmarks demonstrate our superiority over the state-of-the-art.", + "bbox": [ + 75, + 469, + 467, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 96, + 574, + 397, + 588 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a similarity min-max framework for zero-shot day-night domain adaptation. Feature similarity between the original and darkened images is minimized by image-level translation and maximized by model-level adaptation. In this way, model's performance in nighttime is improved without accessing real nighttime images.", + "- We develop a stable training pipeline to solve this bi-level optimization problem. On the image level, we propose an exposure-guided module to perform reliable and controllable nighttime image synthesis. On the model level, we align the representation of images from day and night domains through multi-task contrastive learning.", + "- Our framework universally applies to various nighttime high-level vision tasks. Experiments on classification, semantic segmentation, visual place recognition, and video action recognition demonstrate the superiority of our method." + ], + "bbox": [ + 94, + 595, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 89, + 648, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Low-Light Enhancement. A straightforward approach to improve the model's performance in low light is brightening the test low-light images. Early non-learning practices exploit image processing tools such as histogram equalization [40] or image formation theories such as Retinex Theory [44]. Recent literature mainly takes advantage of the advance in deep learning. Trained on paired day-night data, some methods [33, 55, 56] simulate the image decomposition process of Retinex Theory. Others introduce adversarial learning [23] to support unpaired training. Zero-DCE [16, 30] designs a curve-based low-light enhancement model and trains in a zero-reference way. Advanced techniques, including frequency decomposition [24], feature pyramids [60, 63], and flow models [54] are also adopted in recent papers.", + "bbox": [ + 496, + 118, + 890, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Day-Night Domain Adaptation. Nighttime high-level vision has attracted increasing attention in recent years. Apart from pre-processing with enhancement models, day-night domain adaptation is also a viable solution. YOLO-in-the-dark [47] introduces the glue layer to mitigate the day-night domain gap. MAET [8] exploits image signal processing (ISP) for nighttime image generation and uses both synthetic and real nighttime images for training. HLA-face [52] proposes a joint high-low adaptation framework driven by self-supervised learning. Others [2, 37, 45, 46, 57] employ Generative Adversarial Network (GAN) to transfer labeled daytime data to nighttime.", + "bbox": [ + 496, + 349, + 890, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Zero-Shot Day-Night Domain Adaptation. Beyond Conventional adaptation, zero-shot approaches consider an even stricter condition where real nighttime images are inaccessible. For general tasks, existing methods either draw supports from extra task-irrelevant source and target domain data pairs [39, 51] or require underlying probability distribution of the target domain [20], which are inapplicable to our settings. For the day-night task, Lengyel et al. propose the Color Invariant Convolution (CICov) [29] to capture illumination-robust features. MAET [8] can be viewed as zero-shot when real nighttime images are discarded during finetuning. Besides, domain generalization methods [1, 6, 19, 26, 31, 38, 62] also apply to our settings since they do not know target domains, but they are too general to handle the complex day-night domain gap.", + "bbox": [ + 496, + 534, + 890, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite these advances, low-light enhancement concentrates on human vision and disregards downstream nighttime vision tasks. Conventional adaptation methods require task-specific nighttime datasets, which creates extra burdens on data collection and limits their generalizability to multiple tasks. Prior zero-shot adaptation methods fail to consider image-level and model-level jointly. In this paper, we propose a novel similarity min-max framework that could outperform existing methods by a large margin.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8105", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Similarity Min-Max Optimization", + "text_level": 1, + "bbox": [ + 76, + 89, + 385, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section introduces our approach for zero-shot daynight domain adaptation. We first explain our motivation, then introduce the overall framework and detailed designs.", + "bbox": [ + 76, + 114, + 467, + 161 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation", + "text_level": 1, + "bbox": [ + 76, + 169, + 197, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing methods, generally categorized into Operator-based and Darkening-based as shown in Figure 2, come across troubles in the day-night domain adaptation problem. Operator-based methods [29] rely on the manually designed operators at the model level to handle illumination variations, which are not adaptive to real complex scenarios. Darkening-based methods transfer labeled daytime data to nighttime by ISP [8] or GAN [2,28,45,46] only at the image level. However, the former is sensor-dependent and cannot generalize across devices and datasets, while the latter requires data from the task-specific nighttime domain and thus fails to generalize to our zero-shot setting.", + "bbox": [ + 75, + 191, + 467, + 373 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Intrinsically, the most critical issue of existing methods is their ignorance of the mutual effect between pixels and features. In our work, we make the first systematic investigation on this issue and propose a similarity min-max framework that thoroughly exploits the information from two sides. In detail, at the pixel (image) level, we minimize the feature similarity between original and darkened images by day-to-night translation. While at the feature (model) level, we maximize the feature similarity by representation alignment. This joint optimization leads to representations more robust to illumination changes.", + "bbox": [ + 75, + 373, + 467, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We formulate our framework as follows. Denote the feature extractor of the downstream model as $F(\\cdot)$ . Being robust to illumination requires the extracted feature of a daytime image $I$ and its nighttime version $D(I)$ to be similar, where $D(\\cdot)$ represents a darkening process. The limitation of existing darkening-based methods is that their $D$ does not consider the co-effect of $F$ . So we introduce additional constraints on $D$ : we require $D$ to minimize the similarity between the day feature $F(I)$ and the night feature $F(D(I))$ . This way, we guide the darkening process with high-level vision, forming a unified framework of $D$ and $F$ . At this point, we can integrate $D$ and $F$ as a min-max optimization problem:", + "bbox": [ + 75, + 539, + 467, + 734 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\theta_ {F}} \\min _ {\\theta_ {D}} \\quad \\operatorname {S i m} (F (I), F (D (I))), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 746, + 467, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta_{D}$ and $\\theta_{F}$ denote the parameters in $D$ and $F$ , and $\\mathrm{Sim}(\\cdot, \\cdot)$ measures the similarity between features.", + "bbox": [ + 76, + 779, + 467, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, trivial solutions exist in Eq. (1), such as $D$ generating entirely black images and $F$ extracting identical features for all inputs. We add regularizations to $D$ and $F$ accordingly to address this problem:", + "bbox": [ + 76, + 809, + 467, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\theta_ {F}} \\min _ {\\theta_ {D}} \\operatorname {S i m} (F (I), F (D (I))) + \\mathcal {R} _ {D} (\\theta_ {D}) - \\mathcal {R} _ {F} (\\theta_ {F}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 878, + 467, + 904 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3e038f359b6f9c02a61caa3907af7a7ffc45758d3278b4e6d432c3efc4bdb6a3.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 506, + 92, + 573, + 224 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bf9ec25602bbe837fdb4a3226ba6fc47352f5e75c1461704efe75688fb62a48e.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 578, + 92, + 700, + 229 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/43ce8ed74e5868059a3467aaa757985b8716b65ed79807bee41e0d11eac94bf1.jpg", + "image_caption": [ + "(c)", + "Figure 2. Comparison between different learning paradigms. D and N denote the daytime and nighttime domains, respectively. (a) Operator-based. (b) Darkening-based. (c) Our method." + ], + "image_footnote": [], + "bbox": [ + 705, + 92, + 883, + 231 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{R}_D$ and $\\mathcal{R}_F$ are intended to prevent model collapse.", + "bbox": [ + 500, + 323, + 888, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "How to design $\\mathcal{R}_D$ and $\\mathcal{R}_F$ properly is the key to solving Eq. (2). The following will introduce how we design $\\mathcal{R}_D$ and $\\mathcal{R}_F$ and build up the whole learning framework.", + "bbox": [ + 498, + 338, + 890, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Image-Level Similarity Minimization", + "text_level": 1, + "bbox": [ + 500, + 391, + 820, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section describes our design for the darkening module $D$ . We want $D$ to satisfy three properties:", + "bbox": [ + 498, + 414, + 888, + 445 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Stability. First and foremost, we need to prevent the similarity min-max optimization from collapsing, i.e., applying proper $\\mathcal{R}_D$ in Eq. (2).", + "- Generalization. $D$ should represent a generalized darkening process so the downstream model can learn useful knowledge from $D(I)$ to handle unseen nighttime scenes.", + "- Flexibility. We additionally expect flexible control over the degree of darkening, which could enable us to create diverse inputs beneficial for optimizing $F$ ." + ], + "bbox": [ + 517, + 452, + 888, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We design an exposure-guided pixel-wise mapping algorithm to satisfy the above properties. Unlike widely-used image-to-image darkening approaches [2, 28, 46] that rely heavily on real nighttime images, pixel-wise mapping adjusts images using a pre-selected function with learnable parameters. We empirically found that, by setting proper constraints on the mapping function, we can naturally avoid obtaining trivial solutions in the similarity min-max optimization (stability) and guarantee $D$ follows a typical low-light process (generalization). Finally, we add an exposure guidance mechanism for better flexibility. The detailed design will be illustrated as follows.", + "bbox": [ + 496, + 630, + 890, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Darkening Process. We first define a general function for tone mapping. Given an image $I \\in [0,1]^{C \\cdot H \\cdot W}$ , we use a non-linear mapping $f \\colon [0,1] \\to [0,1]$ and a pixel-wise adjustment map $\\mathcal{A} \\in [0,1]^{C \\cdot H \\cdot W}$ to process the image:", + "bbox": [ + 498, + 814, + 890, + 876 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD ^ {0} (I) = f (I, \\mathcal {A}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 883, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8106", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/953c17127184e661c998bea30ca2ad8e9c642ea338396f8b7fe8a23c11fc20a1.jpg", + "image_caption": [ + "(a) Image-Level Translation" + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 398, + 263 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/39e387222bde2848badf85a19314f557cee3180ca842609728545c05ee8bf139.jpg", + "image_caption": [ + "(b) Model-Level Adaptation", + "Figure 3. Our proposed similarity min-max framework for zero-shot day-night domain adaptation. (a) We first train a darkening module $D$ with a fixed feature extractor to generate synthesized nighttime images that share minimum similarity with their daytime counterparts. (b) After obtaining $D$ , we freeze its weights and maximize the day-night feature similarity to adapt the model to nighttime." + ], + "image_footnote": [], + "bbox": [ + 416, + 88, + 890, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Typically, $f$ should be monotonically increasing to preserve contrast and satisfy $f(1,\\alpha) = 1$ for all $\\alpha$ to avoid information loss (e.g., gamma correction). However, the latter constraint $f(1,\\alpha) = 1$ no longer holds for darkening. Therefore, we propose an auxiliary pixel-wise adjustment using a monotonic increasing function $g$ : $[0,1] \\to [0,1]$ parameterized by another adjustment map $\\mathcal{B} \\in [0,1]^{C\\cdot H\\cdot W}$ . Note that $g$ only serves as a complement and should be simple to avoid taking over the role of $f$ . The overall darkening process is formulated as:", + "bbox": [ + 75, + 382, + 468, + 532 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD (I) = g ^ {- 1} \\left(f \\left(g (I, \\mathcal {B}), \\mathcal {A}\\right), \\mathcal {B}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 555, + 468, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Both $\\mathcal{A}$ and $\\mathcal{B}$ are estimated by a mapping estimator conditioned on the input image $I$ .", + "bbox": [ + 76, + 594, + 468, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To guarantee $D$ represents a darkening process (i.e., $D(I) < I$ ), $f$ should additionally satisfy convexity. Specifically, we let $f$ be the iterative quadratic curve [16]: $f(x) = h^{(8)}(x), h(x,\\alpha) = \\alpha x^2 + (1 - \\alpha)x$ , and $g$ be the dividing operation: $g(x,\\beta) = x / \\beta$ in our implementation. Other kinds of curve forms are also considered and tested. Still, we empirically found that quadratic curves could bring slightly better results (results in Sec. 4.2).", + "bbox": [ + 76, + 630, + 468, + 751 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides, to enable flexible control over the exposure level, we feed an exposure map $E$ to the mapping estimator with $I$ , yielding the corresponding darkened image $D(I,E)$ . During training, the darkening module is encouraged to align the pixel value of $E$ and $D(I,E)$ . We use $D(I)$ and $D(I,E)$ interchangeably for simplicity.", + "bbox": [ + 76, + 755, + 468, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarity Minimization. The training objective of module $D$ involves two parts: similarity minimization and regularization. For the former, we directly reduce the distance", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "between features:", + "bbox": [ + 500, + 383, + 617, + 396 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D} ^ {\\text {s i m}} = \\frac {\\langle F (I) , F (D (I)) \\rangle}{\\left| \\left| F (I) \\right| \\right| _ {2} \\cdot \\left| \\left| F (D (I)) \\right| \\right| _ {2}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 402, + 890, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\langle \\cdot ,\\cdot \\rangle$ is the inner product between two vectors.", + "bbox": [ + 500, + 445, + 849, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The regularization term consists of four losses. Besides a color consistency loss $\\mathcal{L}_{col}$ [16] that corrects color deviations, three additional losses are proposed to regularize $D$ :", + "bbox": [ + 498, + 460, + 890, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Firstly, conditional exposure control is adopted to align the exposure map with the corresponding generated image:", + "bbox": [ + 498, + 506, + 890, + 536 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c - e x p} = \\sum_ {1 \\leq i \\leq H, 1 \\leq j \\leq W} | \\hat {D} _ {i, j} (I, E) - E _ {i, j} |, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 547, + 890, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{D}(I,E)$ is the channel-wise average of $D(I,E)$ . During training, each exposure map $E$ has identical entries uniformly sampled between $[0,0.5]$ .", + "bbox": [ + 498, + 587, + 888, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then we add constraints on $\\mathcal{A}$ . Intuitively, $\\mathcal{A}$ represents the degree of illumination reduction. Illumination usually varies slowly across a scene but encounters rapid variations from object to object. Following this property, we apply a loose total variance loss:", + "bbox": [ + 496, + 633, + 890, + 708 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l t v} (\\mathcal {A}) = \\sum_ {c \\in \\{R, G, B \\}} \\left(h \\left(\\left| \\nabla_ {x} \\mathcal {A} ^ {c} \\right|\\right) ^ {2} + h \\left(\\left| \\nabla_ {y} \\mathcal {A} ^ {c} \\right|\\right) ^ {2}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 715, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh (x) = \\max (\\alpha - | x - \\alpha |, 0), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 753, + 890, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\nabla_{x},\\nabla_{y}$ are gradient operations along the horizontal and vertical axis, respectively, and $\\alpha$ is a hyperparameter. Compared with the original total variance loss where $h$ is the identity function, our loose version allows the network to predict values of greater difference for adjacent pixels, which is common on objects' boundaries.", + "bbox": [ + 496, + 779, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we adopt $\\mathcal{L}_{flex}(\\mathcal{B}) = 1 - \\mathcal{B}$ to avoid model fitting to the exposure solely by $g$ .", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8107", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall training objective for $D$ is:", + "bbox": [ + 96, + 90, + 354, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D} = \\lambda_ {D} ^ {\\text {s i m}} \\mathcal {L} _ {D} ^ {\\text {s i m}} + \\mathcal {R} _ {D}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 116, + 467, + 133 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {D} = \\lambda_ {c - e x p} \\mathcal {L} _ {c - e x p} + \\lambda_ {c o l} \\mathcal {L} _ {c o l} + \\lambda_ {l t v} \\mathcal {L} _ {l t v} + \\lambda_ {f l e x} \\mathcal {L} _ {f l e x}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 137, + 470, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Model-Level Similarity Maximization", + "text_level": 1, + "bbox": [ + 76, + 179, + 401, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The darkening module $D$ grants us access to a synthetic nighttime domain. In this section, we exploit $D$ to learn illumination-robust representations.", + "bbox": [ + 76, + 202, + 468, + 246 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Contrastive learning [5, 17] is a self-supervised learning paradigm that contrasts positive and negative image pairs. However, images of the same class in classification or adjacent scenes in segmentation will form false negative pairs, thus hurting the model's performance. To alleviate these burdens, BYOL [15] proposes a non-negative variant that only aligns the feature between positive image pairs $\\{v, v^{+}\\}$ :", + "bbox": [ + 76, + 247, + 468, + 368 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {B Y O L}} (v, v ^ {+}) = 2 - \\frac {2 \\cdot \\langle z (q (F (v))) , q ^ {\\prime} \\left(F ^ {\\prime} (v ^ {+})\\right) \\rangle}{\\left| \\left| z (q (F (v))) \\right| \\right| _ {2} \\cdot \\left| \\left| q ^ {\\prime} \\left(F ^ {\\prime} (v ^ {+})\\right) \\right| \\right| _ {2}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 378, + 468, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $q, q'$ are projection heads, and $z$ is the prediction head. Both of them are MLPs with a single hidden layer. Note that $F'$ and $q'$ share the same architecture and weight initialization with $F$ and $q$ but receive no gradient and are updated by exponential moving average (EMA).", + "bbox": [ + 76, + 425, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarity Maximization. Motivated by BYOL, we maximize the feature similarity between synthetic nighttime and daytime domains by non-negative contrastive learning. Given a daytime image $I$ and an exposure map $E$ , we formulate the training objective as follows:", + "bbox": [ + 76, + 503, + 468, + 579 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {F} ^ {\\text {s i m}} = \\mathcal {L} _ {\\mathrm {B Y O L}} (I, D (I, E)) + \\mathcal {L} _ {\\mathrm {B Y O L}} (D (I, E), I). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 590, + 468, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that the measure of feature similarity is different between Eq. (5) and Eq. (12). Directly applying Eq. (5) to train $F$ brings poorer results due to potential feature degeneration. In comparison, the asymmetric projection head and stop gradient policies prevent the feature extractor $F$ from collapsing, i.e., working as the regularization $\\mathcal{R}_F$ in Eq. (2) together with the task loss (introduced below).", + "bbox": [ + 76, + 619, + 468, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, different from $E$ in Eq. (6), we use a compound exposure map $E'$ instead. $E'$ is first initialized with identical entries uniformly sampled between [0, 0.2] for simulating nighttime illumination. This range is the same for all downstream tasks, which does not introduce task-relevant prior. Then, we add pixel-wise noise $z_1$ and patch-wise noise $z_2$ to $E$ to simulate exposure discrepancy. Overall, $E'$ can be represented as:", + "bbox": [ + 76, + 724, + 468, + 845 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE ^ {\\prime} = \\mathcal {U} (0, 0. 2) + z _ {1} + z _ {2}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 857, + 468, + 875 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "See the supplementary for details on noise injection.", + "bbox": [ + 76, + 885, + 423, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/76228bff88e809d282988be7f1c83351000e58b3b658172c18577f02be06583d.jpg", + "image_caption": [ + "Figure 4. t-SNE [50] visualization of images' feature extracted by the original daytime model and our adapted model on CO-DaN [29]. Red, green, and blue dots represent the feature of daytime, synthesized nighttime, and real nighttime images, respectively. We only color the instances from the \"Car\" category for better visual quality. Additional visualization results are shown in the supplementary." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 885, + 248 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Besides $\\mathcal{L}_F^{sim}$ , we add task-specific supervision $\\mathcal{L}_{task}$ on both the original daytime and synthetic nighttime domain. The final training objective for $F$ is:", + "bbox": [ + 498, + 383, + 890, + 429 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {F} = \\lambda_ {F} ^ {\\text {s i m}} \\mathcal {L} _ {F} ^ {\\text {s i m}} + \\lambda_ {\\text {t a s k}} \\mathcal {L} _ {\\text {t a s k}}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 436, + 890, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Overall Training Pipeline", + "text_level": 1, + "bbox": [ + 500, + 469, + 732, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Having introduced the image-level similarity minimization (Section 3.2) and model-level similarity maximization (Section 3.3), this section discusses the overall pipeline, as shown in Figure 3.", + "bbox": [ + 498, + 493, + 890, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An intuitive idea is training $D$ and $F$ alternately like GAN [13, 64]. Nevertheless, balancing $D$ and $F$ increases the difficulty of parameter tuning and makes the optimization process unstable. We adopt a simple but effective two-step strategy to solve this problem: we first train $D$ and keep $F$ frozen, then train $F$ and keep $D$ frozen. Compared with the alternate strategy, our step-wise approach improves the performance on nighttime image classification (elaborate in Section 4.2) from $63.84\\%$ to $65.87\\%$ .", + "bbox": [ + 498, + 553, + 890, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We could also explain the merits of our min-max framework from the perspective of adversarial training [14, 35]. Module $D$ first produces the worst-case examples regarding feature similarity. Then, our model could learn the illumination-robust features by learning on these cases through similarity maximization. This technical commonality further justifies our motivation to build the similarity min-max framework.", + "bbox": [ + 496, + 689, + 890, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Across all downstream tasks, the feature extractor and task module are initialized by daytime pre-trained models. We first freeze the feature extractor and train the darkening module (image-level translation). Then, we keep the darkening module fixed and train the feature extractor and task module jointly (model-level adaptation).", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8108", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Empirical Justifications on Darkening Module", + "text_level": 1, + "bbox": [ + 76, + 90, + 467, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Simulating nighttime conditions without accessing real nighttime images is the key to our framework. Particularly, nighttime conditions bring semantic changes in addition to illumination changes, e.g., the dark environment with artificial lights on the second real nighttime image in Figure 4. However, an accurate simulation is extremely difficult since our prior knowledge is limited to \"low illumination\". Fortunately, unlike typical day-to-night image synthesis processes [41] which target the human visual experience, ours only care about the distribution of darkened images in the feature space. Leaving aside visual quality, we are pleased to find that the feature distribution of our synthesized nighttime domain is similar to that of the real nighttime domain as visualized in Figure 4(a). This observation demonstrates that our darkening process can characterize the night domain from the model-level perspective.", + "bbox": [ + 75, + 113, + 468, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Thanks to this property, the feature discrepancy between daytime and real nighttime domain is significantly reduced after model-level adaptation (red and blue dots in Figure 4). This discovery is consistent with the Maximum Mean Discrepancy (MMD) between the feature distribution of day and night modalities, which is 0.020 and 0.014 for the original and adapted models, respectively. We provide implementation details and additional empirical analysis using saliency maps in the supplementary.", + "bbox": [ + 75, + 356, + 468, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 506, + 209, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section provides the implementation details, benchmarking results, and ablation analysis of our method.", + "bbox": [ + 76, + 531, + 468, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 571, + 294, + 587 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our framework widely applies to various nighttime vision tasks. In the following, we evaluate our method with four representative tasks: image classification, semantic segmentation, visual place recognition, and video action recognition. Only daytime data are accessible for training and validation, while nighttime data are only used during evaluation. We benchmark our method with three categories of methods that require no dataset-specific target domain data: low-light enhancement, zero-shot day-night domain adaptation, and domain generalization. For low-light enhancement, enhancement models are trained on their original datasets. Then we adopt them as a pre-processing step to assist the daytime baseline. The results of our method is the average of three independent trails. Additional details are provided in the supplementary.", + "bbox": [ + 75, + 595, + 468, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Nighttime Image Classification", + "text_level": 1, + "bbox": [ + 76, + 830, + 349, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first consider one of the most fundamental vision tasks: image classification. CODaN [29] is a 10-class dataset containing a training set of 10000 daytime images", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9c213b5f9d16c22023119f0557a95e9763769afaad65594fd2363a27de2063c8.jpg", + "table_caption": [ + "Table 1. Top-1 classification accuracy on the CODaN nighttime test set [29]. $\\dagger$ denotes our re-implementation with both the original and synthesized image fed into the task module." + ], + "table_footnote": [], + "table_body": "
MethodTop-1 (%)
ResNet-18 [18]53.32
Low-Light Enhancement
EnlightenGAN [23]56.68
LEDNet [63]57.40
Zero-DCE++ [30]57.96
RUAS [33]58.36
SCI [34]58.68
URetinexNet [56]58.72
Domain Generalization
MixStyle [62]53.12
IRM [1]54.52
AdaBN [31]54.25
Zero-Shot Day-Night Domain Adaptation
MAET† [8]56.48
CICConv [29]60.32
Ours65.87
", + "bbox": [ + 511, + 142, + 877, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and a test set with 2500 daytime and nighttime images, respectively. We validate models on the daytime test set and evaluate them on the nighttime test set. The backbone is ResNet-18 [18].", + "bbox": [ + 498, + 460, + 890, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmarking results are shown in Table 1. Enhancement methods restore input low-light images from the human visual perspective while keeping the model untouched, resulting in limited performance gains. Domain generalization methods are designed for general tasks and perform poorly in unseen nighttime environments. MAET [8] relies on degrading transformation with sensor-specific parameters, which suffers from poor generalizability. CICConv [29] adopts learnable color invariant edge detectors, which are not robust to the complex illumination variation in real scenarios. In contrast, our method outperforms state-of-the-art methods by a large margin (60.32% v.s. 65.87%), demonstrating our unified framework could obtain features more robust to illumination shifts.", + "bbox": [ + 496, + 522, + 890, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablation Studies. We conduct ablation studies to justify our framework design in Table 2. Firstly, we study how to design the darkening module $D$ given $\\mathcal{L}_D^{sim}$ . The model-level adaptation stage (Section 3.3) remains the same for fair comparisons. Firstly, we replace our darkening module with heuristic image adjustment approaches, such as brightness adjustment (Brightness in PIL $^1$ ) and gamma correction ( $D(I) = I^{\\gamma}$ ). We implement these two approaches using a fixed darkening hyperparameter chosen after multi-", + "bbox": [ + 496, + 734, + 890, + 871 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1https://pillow.readthedocs.io/en/stable/reference/ImageEnhance.html", + "bbox": [ + 517, + 886, + 885, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8109", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d89df1c6dc0bd829142bdb9e31c5493f58d5452c9c734c15b09ba9cf58eb29d0.jpg", + "table_caption": [ + "Table 2. Ablation studies for module $D$ and similarity losses. We report the Top-1 accuracy on the CODaN [29] nighttime test set." + ], + "table_footnote": [], + "table_body": "
CategoryMethodTop-1 (%)
BaselineVanilla ResNet-1853.32
Module DBrightness adjustment57.96
HeuristicGamma correction63.96
Module DReciprocal curve62.60
LearnableGamma curve64.16
Similarityw/o LDSim and LFSim64.08
Lossw/o LDSim64.56
w/o LFSim64.88
Full version-65.87
", + "bbox": [ + 107, + 128, + 434, + 309 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ple trials and report the best score.", + "bbox": [ + 76, + 335, + 307, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we test other possible curve forms for $f$ . Both gamma curve $(f(x,\\alpha) = x^{\\frac{1}{\\alpha}}, \\alpha \\in (0,1])$ and reciprocal curve $(f(x,\\alpha) = \\frac{(1 - \\alpha)\\cdot x}{1 - \\alpha\\cdot x}, \\alpha \\in [0,1))$ bring slightly worse results than the iterative quadratic curve. Please refer to the supplementary for implementation details of these ablations and additional results on the segmentation task.", + "bbox": [ + 75, + 349, + 468, + 443 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we test our framework's performance when one or both of the similarity loss is absent. We find that either similarity loss alone can boost the model's nighttime performance while combining them achieves the best result.", + "bbox": [ + 75, + 443, + 468, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Nighttime Semantic Segmentation", + "text_level": 1, + "bbox": [ + 76, + 513, + 375, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we explore a more challenging nighttime vision task: semantic segmentation. We adopt RefineNet [32] with ResNet-101 backbone as the baseline. The daytime training dataset is Cityscapes [7], containing 2975 images for training and 500 images for validation, all with dense annotations. The nighttime testing datasets are Nighttime Driving [9] and Dark-Zurich [46]. These two datasets contain 50 coarsely annotated and 151 densely annotated nighttime street view images.", + "bbox": [ + 75, + 537, + 468, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We benchmark our method in Table 3. Low-light enhancement methods yield worse results than the baseline because they perform poorly on street scenes with complex light sources. Domain generalization methods fail to mitigate the huge day-night domain gap, leading to unsatisfactory results. Note that RobustNet [6] adopts DeepLabv3 [4] architecture, which is superior to RefineNet [32] adopted in our implementation. Among zero-shot adaptation methods, MAET [8] injects too much noise into images, leading to severe performance degradation. CICConv yields better results, but the improvement is limited. In comparison, our approach improves the mIoU scores to $44.9\\%$ on Nighttime Driving and $40.2\\%$ on Dark-Zurich.", + "bbox": [ + 75, + 672, + 468, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5 shows qualitative segmentation results on two nighttime datasets. Low-light enhancement methods per", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9733bec9aa9493474b9f24fd7371cf869586d9132f06465a538a3a8f783f5e2f.jpg", + "table_caption": [ + "Table 3. Semantic segmentation results on Nighttime Driving [9] and Dark-Zurich [46], reported as percentage mIoU scores." + ], + "table_footnote": [], + "table_body": "
MethodNighttime DrivingDark-Zurich
RefineNet [32]34.330.6
Low-Light Enhancement
EnlightenGAN [23]25.224.9
Zero-DCE++ [30]32.728.3
RUAS [33]25.123.4
SCI [34]28.625.7
URetinexNet [56]28.124.0
LEDNet [63]27.626.6
Domain Generalization
AdaBN [31]37.231.1
RobustNet [6]33.034.5
SAN-SAW [38]28.116.0
Zero-Shot Day-Night Domain Adaptation
MAET [8]28.126.4
CICConv [29]41.234.5
Ours44.940.2
", + "bbox": [ + 506, + 128, + 883, + 419 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/284ea32fd983ea25bd305b73798493689fd0c82a121f09211d99bae5d6126fe1.jpg", + "table_caption": [ + "Table 4. Visual place recognition results on Tokyo 24/7 [49]." + ], + "table_footnote": [], + "table_body": "
MethodmAP (%)
Zero-Shot Day-Night Domain Adaptation
EdgeMAC [42]75.9
U-Net jointly [21]79.8
GeM [43]85.0
CICConv-GeM [29]88.3
Ours-GeM90.4
Day-Night Domain Adaptation \n(night images are available for training)
U-Net jointly [21]86.5
EdgeMAC + CLAHE [21]90.5
EdgeMAC + U-Net jointly [21]90.0
", + "bbox": [ + 514, + 457, + 874, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "form poorly on nighttime street scenes. Our method better extracts information hidden by darkness and thus generates more accurate semantic maps.", + "bbox": [ + 498, + 685, + 890, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Visual Place Recognition at Night", + "text_level": 1, + "bbox": [ + 500, + 739, + 792, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Then we explore visual place recognition, which aims to retrieve images that illustrate the same scene of a query image from an image pool. Unlike classification and segmentation, place recognition methods are not end-to-end during inference. We extend our method based on GeM [43] with ResNet-101 backbone. In GeM, the network receives a tuple of images $\\{p,q,n_1,\\dots ,n_k\\}$ as input, in which the query $q$ only matches $p$ . The network is trained on a contrastive loss, similar to the model-level stage in our", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8110", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c34cccb7bbf33208d091438771560d3643b627f27d819ee8b04e018160bbd63.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 88, + 212, + 141 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1283e0a1259f724edc7c69db5ab16e4b4e33f511901c6088fcc7a7eec748a2ed.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 80, + 142, + 210, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f242153bf0ca9531624ffbfca6435a8bc251e2481f8f43c884ba17da0cb273b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 89, + 348, + 141 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4c7c62bf14bd881f5fd5aba9331243f0b27210fafadc7392ad4bd9c6bf7546ba.jpg", + "image_caption": [ + "(b) Ground truth" + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 346, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4fa47fd874dacefa77e2c26c48d5ccb82c010d853c2f9fdc7a1e1b14be2f6231.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 89, + 482, + 141 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce58d14809e4597c2ab1f07668d28c502f7245e0b1118fdd70ccbac0a213c414.jpg", + "image_caption": [ + "(c) RefineNet" + ], + "image_footnote": [], + "bbox": [ + 351, + 142, + 483, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dff8421f58b7d6ff244687f790104a06e1657ff977af59e6470b0005e719c810.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 89, + 617, + 141 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/26a89c7c67f4577da9f820c5edd3e2a1543a6f883fafb60adcf1fdf245f61615.jpg", + "image_caption": [ + "(d) MAET" + ], + "image_footnote": [], + "bbox": [ + 486, + 142, + 617, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0c15b0e42873f543d7a7b3edfaf98e404e2419f349ebc8d41f7cdaeb38856595.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 89, + 754, + 141 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f185793ec33a9ad215eb49cb9d824ee6b6ee4118aef5483023e8c60f98f62a7e.jpg", + "image_caption": [ + "(e) CICConv" + ], + "image_footnote": [], + "bbox": [ + 620, + 142, + 754, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/491b913ceaef18be9aa5a615cca5aa92bda6f20f1f235161785d0a4f3d5f5c29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 89, + 890, + 141 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/77a65d83066df2462a180e2af4dbfdae0fa824928f6d29cffefcd1f47074fc09.jpg", + "image_caption": [ + "(f) Ours" + ], + "image_footnote": [], + "bbox": [ + 756, + 142, + 890, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/287f8adf80db365e86267b91809eae6f51dd06d2588839fd7be381d2240e0af3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 210, + 212, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/158eb11862588069f2d9b43bcf528ef41b7d33a47fbca340dddf47d35ec725a2.jpg", + "image_caption": [ + "(g)SCI" + ], + "image_footnote": [], + "bbox": [ + 80, + 263, + 212, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ccf612991c94ec0cbdadc4098eb518c82ea6697c9b23dab6557df3339139d9fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 210, + 348, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/190baffd73ccf3e85e47ed02edbeb1438af65faa96aaa5c23771e9c02866bbe0.jpg", + "image_caption": [ + "Figure 5. Semantic segmentation results. For each group, the first row: Nighttime Driving [9], the second row: Dark-Zurich [46]." + ], + "image_footnote": [], + "bbox": [ + 215, + 263, + 348, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/82498635fb2d1d7504b70b1d1cc5ab9ae101cd69bbb50b18f76133f5e2595f66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 210, + 483, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9aead569bf4da46517ef802e185521cf3d353906ec99176561ae4f98473dd61a.jpg", + "image_caption": [ + "(h) URetinexNet" + ], + "image_footnote": [], + "bbox": [ + 351, + 263, + 483, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fe196eb1a464f355d116bbd80c08957b2c772bf93c2262c85b1a6567e5b2fb71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 210, + 619, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c7212c51eb4eed20fee3b2ce313c032b30cc72540e75624827306a19c414c436.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 263, + 619, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/32c5e15404eecadc62dcfd4dde9393d3c1c1d02f5f6460147c3692220ba97b6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 210, + 754, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e39f936847e30e759b6bafc371ca06ddb54f02d4245b005f1a01f4119d4fe61b.jpg", + "image_caption": [ + "(i) Zero-DCE++" + ], + "image_footnote": [], + "bbox": [ + 620, + 263, + 754, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/194e97d964d681aacadcd6ff2819c3abb38bc532c01afcb03471194739c24e28.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 210, + 890, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9d2705ec1807b155c053d78b81cc886453a28a63a71f155c1fc35443a50c9458.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 263, + 890, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1602b536bd4468fed7c3749932595e07db0103c673ac3b4090e2707c8013eab4.jpg", + "image_caption": [ + "(a) Query", + "Figure 6. Qualitative visual place recognition results. (a) A night query from the Tokyo 24/7 dataset [49]. (b) Image retrieved by GeM [43]. (c) Image retrieved by our method." + ], + "image_footnote": [], + "bbox": [ + 88, + 387, + 207, + 494 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f56a9621e95f59a95a6b0fbbc099f2a39a1a81108f5969bbb6e3fec5dc6d7379.jpg", + "image_caption": [ + "(b) GeM" + ], + "image_footnote": [], + "bbox": [ + 214, + 387, + 333, + 494 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/53dd682c1230ea9c589870472abbe52ee42eae5cb3facd77118c3e9217d31cb9.jpg", + "image_caption": [ + "(c) Ours" + ], + "image_footnote": [], + "bbox": [ + 341, + 387, + 457, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "framework. We retain the image-level stage and modify the model-level stage in our implementation. We first train the darkening module $D$ as usual. Then, we consider $D(p)$ as an additional matching for $p$ , i.e., an input tuple contains two positive samples (instead of one) and $k$ negative samples. We train our network on the Retrieval-SfM dataset [43] and evaluate it on the Tokyo 24/7 dataset [49], which contains city views in multiple illumination conditions and viewing directions.", + "bbox": [ + 75, + 595, + 467, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance is reported as mean Average Precision (mAP) in Table 4. Results of comparison methods are borrowed from [21] and [29]. Our method outperforms all zero-shot methods and is comparable to conventional domain adaptation methods. As shown in Figure 6, the baseline method gets fooled by the night's appearance, while our model finds the correct daytime image.", + "bbox": [ + 75, + 732, + 467, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Low-Light Video Action Recognition", + "text_level": 1, + "bbox": [ + 76, + 847, + 395, + 863 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although initially designed for images, our method also applies to video tasks. Here we consider an 11-class low-", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/05e4fb800ed13eff971fe8e1fd52915f7d1e2c40d365705b1c8c08d4fb2d04a9.jpg", + "table_caption": [ + "Table 5. Video action recognition results on ARID [58]." + ], + "table_footnote": [], + "table_body": "
MethodTop-1 (%)
I3D [3]47.02
Low-Light Video Enhancement
StableLLVE [59]45.08
SMOID [22]47.27
SGZ [61]46.42
Domain Generalization & Zero-Shot Day-Night Domain Adaptation
AdaBN [31]46.17
Ours51.52
", + "bbox": [ + 513, + 412, + 875, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "light video action recognition task. Normal light training data consists of 2.6k normal light video clips from HMDB51 [27], UCF101 [48], Kinetics-600 [25], and Moments in Time [36]. We evaluate our model on the official test split of the ARID dataset [58]. The action recognizer is I3D [3] based on 3D-ResNet [11].", + "bbox": [ + 496, + 625, + 890, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We extend our method to video as follows. When training the darkening module, we input frames extracted from video clips. $\\mathcal{A}$ and $\\mathcal{B}$ in Eq. (4) is estimated for every individual frame. We calculate $\\mathcal{L}_D^{sim}$ between video clips and other losses between frames. When generating low-light videos, frames are separately fed into the curve estimator while sharing the same exposure map $E^{\\prime}$ .", + "bbox": [ + 496, + 715, + 890, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We report the results as Top-1 accuracy. As shown in Table 5, video enhancement methods StableLLVE [59], SMOID [22], and SGZ [61] yield a limited performance gain. Meanwhile, our approach boosts models' performance by $4.38\\%$ , demonstrating our superiority on videos.", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8111", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 89, + 194, + 104 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose a novel approach for zero-shot day-night domain adaptation. Going beyond a simple focus on the image-level translation or model-level adaptation, we observe a complementary relationship between two aspects and build our framework upon the similarity min-max paradigm. Our proposed method can significantly boost the model's performance at nighttime without accessing the nighttime domain. Experiments on multiple datasets demonstrate the superiority and broad applicability of our approach.", + "bbox": [ + 75, + 116, + 472, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 76, + 286, + 243, + 303 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the Fundamental Research Funds for the Central Universities and the National Natural Science Foundation of China under Contract No.61772043 and a research achievement of Key Laboratory of Science, Technology and Standard in Press Industry (Key Laboratory of Intelligent Press Media Technology). This research work is also partially supported by the Basic and Frontier Research Project of PCL and the Major Key Project of PCL.", + "bbox": [ + 75, + 311, + 472, + 434 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 76, + 452, + 174, + 467 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv, 2019. 2, 6", + "[2] Vinicius F Arruda, Thiago M Paixão, Rodrigo F Berriel, Alberto F De Souza, Claudine Badue, Nicu Sebe, and Thiago Oliveira-Santos. Cross-domain car detection using unsupervised image-to-image translation: From day to night. In IJCNN, 2019. 1, 2, 3", + "[3] João Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, 2017. 8", + "[4] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv, 2017. 7", + "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 5", + "[6] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T. Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, 2021. 2, 7", + "[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 7", + "[8] Ziteng Cui, Guo-Jun Qi, Lin Gu, Shaodi You, Zenghui Zhang, and Tatsuya Harada. Multitask aet with orthogonal tangent regularity for dark object detection. In ICCV, 2021. 1, 2, 3, 6, 7" + ], + "bbox": [ + 86, + 479, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] Dengxin Dai and Luc Van Gool. Dark model adaptation: Semantic image segmentation from daytime to nighttime. In ITSC, 2018. 7, 8", + "[10] Xueqing Deng, Peng Wang, Xiaochen Lian, and Shawn Newsam. Nightlab: A dual-level architecture with hardness detection for segmentation at night. In CVPR, 2022. 1", + "[11] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 8", + "[12] Huan Gao, Jichang Guo, Guoli Wang, and Qian Zhang. Cross-domain correlation distillation for unsupervised domain adaptation in nighttime semantic segmentation. In CVPR, 2022. 1", + "[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 2020. 5", + "[14] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv, 2014. 5", + "[15] Jean-Bastien Grill, Florian Strub, Florent Altché, C. Tallec, Pierre H. Richemond, Elena Buchatskaya, C. Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, B. Piot, K. Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS, 2020. 5", + "[16] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In CVPR, 2020. 2, 4", + "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 5", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6", + "[19] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In CVPR, 2021. 2", + "[20] Masato Ishii, Takashi Takenouchi, and Masashi Sugiyama. Zero-shot domain adaptation based on attribute information. In ACML, 2019. 2", + "[21] Tomas Jenicek and Ondrej Chum. No fear of the dark: Image retrieval under varying illumination conditions. In ICCV, 2019. 7, 8", + "[22] Haiyang Jiang and Yinqiang Zheng. Learning to see moving objects in the dark. In ICCV, 2019. 8", + "[23] Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. *Enlightengan: Deep light enhancement without paired supervision. IEEE TIP*, 2021. 1, 2, 6, 7", + "[24] Yeying Jin, Wenhan Yang, and Robby T Tan. Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In ECCV, 2022. 2", + "[25] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola," + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8112", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Tim Green, Trevor Back, Paul Natev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv, 2017. 8", + "[26] Namyup Kim, Taeyoung Son, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Wedge: web-image assisted domain generalization for semantic segmentation. arXiv, 2021. 2", + "[27] Hildegard Kuehne, Hueihan Jhuang, Estibaliz Garrote, Tomaso A. Poggio, and Thomas Serre. HMDB: A large video database for human motion recognition. In ICCV, 2011. 8", + "[28] Hongjun Lee, Moonsoo Ra, and Whoi-Yul Kim. Nighttime data augmentation using gan for improving blind-spot detection. IEEE Access, 2020. 3", + "[29] Attila Lengyel, Sourav Garg, Michael Milford, and Jan C. van Gemert. Zero-shot day-night domain adaptation with a physics prior. In ICCV, 2021. 1, 2, 3, 5, 6, 7, 8", + "[30] Chongyi Li, Chunle Guo, and Chen Change Loy. Learning to enhance low-light image via zero-reference deep curve estimation. IEEE TPAMI, 2021. 1, 2, 6, 7", + "[31] Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting batch normalization for practical domain adaptation. In ICLRW, 2017. 2, 6, 7, 8", + "[32] Guosheng Lin, Anton Milan, Chunhua Shen, and Ian Reid. Refinenet: Multi-path refinement networks for high-resolution semantic segmentation. In CVPR, 2017. 7", + "[33] Risheng Liu, Long Ma, Jiaao Zhang, Xin Fan, and Zhongxuan Luo. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In CVPR, 2021. 2, 6, 7", + "[34] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In CVPR, 2022. 1, 6, 7", + "[35] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv, 2017. 5", + "[36] Mathew Monfort, Carl Vondrick, Aude Oliva, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa M. Brown, Quanfu Fan, and Dan Gutfreund. Moments in time dataset: One million videos for event understanding. IEEE TPAMI, 2020. 8", + "[37] Amitangshu Mukherjee, Ameya Joshi, Anuj Sharma, Chinmay Hegde, and Soumik Sarkar. Generative semantic domain adaptation for perception in autonomous driving. Journal of Big Data Analytics in Transportation, 2022. 2", + "[38] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 2, 7", + "[39] Kuan-Chuan Peng, Ziyan Wu, and Jan Ernst. Zero-shot deep domain adaptation. In ECCV, 2018. 2", + "[40] Stephen M Pizer, E Philip Amburn, John D Austin, Robert Cromartie, Ari Geselowitz, Trey Greer, Bart ter Haar Romy, John B Zimmerman, and Karel Zuiderveld. Adaptive histogram equalization and its variations. Computer Vision, Graphics, and Image Processing, 1987. 2", + "[41] Abhijith Punnappurath, Abdullah Abuolaim, Abdelrahman Abdelhamed, Alex Levinshtein, and Michael S Brown. Day-to-night image synthesis for training nighttime neural isps. In CVPR, 2022. 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Filip Radenovic, Giorgos Tolias, and Ondrej Chum. Deep shape matching. In ECCV, 2018. 7", + "[43] Filip Radenovic, Giorgos Tolias, and Ondrej Chum. Finetuning cnn image retrieval with no human annotation. IEEE TPAMI, 2019. 7, 8", + "[44] Zia-ur Rahman, Daniel J Jobson, and Glenn A Woodell. Retinex processing for automatic image enhancement. Journal of Electronic Imaging, 2004. 2", + "[45] Eduardo Romera, Luis M. Bergasa, Kailun Yang, Jose M. Alvarez, and Rafael Barea. Bridging the day and night domain gap for semantic segmentation. IEEE Intelligent Vehicles Symposium, 2019. 1, 2, 3", + "[46] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Guided curriculum model adaptation and uncertainty-aware evaluation for semantic nighttime image segmentation. In ICCV, 2019. 1, 2, 3, 7, 8", + "[47] Yukihiro Sasagawa and Hajime Nagahara. Yolo in the dark-domain adaptation method for merging multiple models. In ECCV, 2020. 1, 2", + "[48] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv, 2012. 8", + "[49] Akihiko Torii, Relja Arandjelovic, Josef Sivic, Masatoshi Okutomi, and Tomas Pajdla. 24/7 place recognition by view synthesis. In CVPR, 2015. 7, 8", + "[50] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 5", + "[51] Jinghua Wang and Jianmin Jiang. Conditional coupled generative adversarial networks for zero-shot domain adaptation. In ICCV, 2019. 2", + "[52] Wenjing Wang, Xinhao Wang, Wenhan Yang, and Jiaying Liu. Unsupervised face detection in the dark. IEEE TPAMI, 2022. 1, 2", + "[53] Wenjing Wang, Zhengbo Xu, Haofeng Huang, and Jiaying Liu. Self-aligned concave curve: Illumination enhancement for unsupervised adaptation. In ACM MM, 2022. 1", + "[54] Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, LapPui Chau, and Alex Kot. Low-light image enhancement with normalizing flow. In AAAI, 2022. 2", + "[55] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. In BMVC, 2017. 2", + "[56] Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Wenhan Yang, and Jianmin Jiang. Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. In CVPR, 2022. 1, 2, 6, 7", + "[57] Xinyi Wu, Zhenyao Wu, Hao Guo, Lili Ju, and Song Wang. Darnet: A one-stage domain adaptation network for unsupervised nighttime semantic segmentation. In CVPR, 2021. 2", + "[58] Yuecong Xu, Jianfei Yang, Haozhi Cao, Kezhi Mao, Jianxiong Yin, and Simon See. Arid: A new dataset for recognizing action in the dark. In Deep Learning for Human Activity Recognition, 2021. 8", + "[59] Fan Zhang, Yu Li, Shaodi You, and Ying Fu. Learning temporal consistency for low light video enhancement from single images. In CVPR, 2021. 8" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8113", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Yuzhi Zhao, Yongzhe Xu, Qiong Yan, Dingdong Yang, Xuehui Wang, and Lai-Man Po. D2hnet: Joint denoising and deblurring with hierarchical network for robust night image restoration. In ECCV, 2022. 1, 2", + "[61] Shen Zheng and Gaurav Gupta. Semantic-guided zero-shot learning for low-light image/video enhancement. In WACVW, 2022. 8", + "[62] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. In ICLR, 2021. 2, 6", + "[63] Shangchen Zhou, Chongyi Li, and Chen Change Loy. Lednet: Joint low-light enhancement and deblurring in the dark. In ECCV, 2022. 1, 2, 6, 7", + "[64] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. 5" + ], + "bbox": [ + 78, + 90, + 468, + 303 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8114", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Similarity Min-Max_ Zero-Shot Day-Night Domain Adaptation/d2d1b76d-0c03-43ab-96b8-c3ce7c667abf_model.json b/2023/Similarity Min-Max_ Zero-Shot Day-Night Domain Adaptation/d2d1b76d-0c03-43ab-96b8-c3ce7c667abf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b2295f821c6bdd13cca91b87c7277e8159932e0d --- /dev/null +++ b/2023/Similarity Min-Max_ Zero-Shot Day-Night Domain Adaptation/d2d1b76d-0c03-43ab-96b8-c3ce7c667abf_model.json @@ -0,0 +1,3016 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.166, + 0.131, + 0.804, + 0.154 + ], + "angle": 0, + "content": "Similarity Min-Max: Zero-Shot Day-Night Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.18, + 0.752, + 0.2 + ], + "angle": 0, + "content": "Rundong Luo\\(^{1,2,3}\\) Wenjing Wang\\(^{1}\\) Wenhan Yang\\(^{4}\\) Jiaying Liu\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.2, + 0.746, + 0.218 + ], + "angle": 0, + "content": "\\(^{1}\\) Wangxuan Institute of Computer Technology, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.217, + 0.873, + 0.235 + ], + "angle": 0, + "content": "\\(^{2}\\)School of EECS, Peking University \\(^{3}\\)School of CS, Peking University \\(^{4}\\)Peng Cheng Laboratory" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.695 + ], + "angle": 0, + "content": "Low-light conditions not only hamper human visual experience but also degrade the model's performance on downstream vision tasks. While existing works make remarkable progress on day-night domain adaptation, they rely heavily on domain knowledge derived from the task-specific nighttime dataset. This paper challenges a more complicated scenario with border applicability, i.e., zero-shot day-night domain adaptation, which eliminates reliance on any nighttime data. Unlike prior zero-shot adaptation approaches emphasizing either image-level translation or model-level adaptation, we propose a similarity min-max paradigm that considers them under a unified framework. On the image level, we darken images towards minimum feature similarity to enlarge the domain gap. Then on the model level, we maximize the feature similarity between the darkened images and their normal-light counterparts for better model adaptation. To the best of our knowledge, this work represents the pioneering effort in jointly optimizing both levels, resulting in a significant improvement of model generalizability. Extensive experiments demonstrate our method's effectiveness and broad applicability on various nighttime vision tasks, including classification, semantic segmentation, visual place recognition, and video action recognition. Our project page is available at https://red-fairy.github.io/ZeroShotDayNightDA-Webpage/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.727, + 0.21, + 0.743 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.753, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Deep neural networks are sensitive to insufficient illumination, and such deficiency has posed significant threats to safety-critical computer vision applications. Intuitively, insufficient illumination can be handled by low-light enhancement methods [23, 30, 34, 56, 60, 63], which aim at restoring low-light images to normal-light. However, enhancement models do not necessarily benefit downstream high-level vision tasks as they are optimized for human visual perception" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.272, + 0.716, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.724, + 0.283, + 0.884, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.427, + 0.893, + 0.484 + ], + "angle": 0, + "content": "Figure 1. Left: Illustration of our similarity min-max framework for zero-shot day-night domain adaptation. Right: Our framework achieves state-of-the-art results on multiple downstream high-level vision tasks without seeing real nighttime images during training." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.517, + 0.768, + 0.532 + ], + "angle": 0, + "content": "and neglect the need for machine vision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.893, + 0.717 + ], + "angle": 0, + "content": "Much existing literature has focused on improving machine vision performance at night through domain adaptation. By aligning the distribution statistics between the nighttime and daytime datasets through image translation [2, 12, 45], self-supervised learning [52, 53], or multistage algorithms [10, 46, 47], these methods have greatly improved models' performance in nighttime environments. The primary assumption of domain adaptation is that the target domain data is readily available. Nevertheless, obtaining data from the task-specific target domain may be challenging in extreme practical application scenarios such as deep-space exploration and deep-sea analysis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "To reduce the requirement on target domain data, zero-shot domain adaptation has emerged as a promising research direction, where adaptation is performed without accessing the target domain. Regarding day-night domain adaptation, the primary challenge is learning illumination-robust representations generalizable to both day and night modalities. To accomplish this goal under zero-shot constraints, Lengyel et al. [29] proposed a color invariant convolution for handling illumination changes. Cui et al. [8] designed a Reverse ISP pipeline and generated synthetic nighttime images with pseudo labels. However, image-level methods simply consider synthetic nighttime as pseudo" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8104" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.168 + ], + "angle": 0, + "content": "labeled data and overlook model-level feature extraction; model-level methods focus on adjusting model architecture but neglect image-level nighttime characteristics. Neither is effective enough capture the illumination-robust representations that could bridge the complex day-night domain gap." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.468, + 0.287 + ], + "angle": 0, + "content": "From this point of view, we devise a similarity min-max framework that involves two levels, as illustrated in Figure 1. On the image level, we generate a synthetic nighttime domain that shares minimum feature similarity with the daytime domain to enlarge the domain gap. On the model level, we learn illumination-robust representations by maximizing the feature similarity of images from the two domains for better model adaptation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.468, + 0.469 + ], + "angle": 0, + "content": "Intuitive as it seems, solving this bi-level optimization problem is untrivial. Directly solving it may yield unsatisfactory results, e.g., meaningless images filled with zero values or identical features given all inputs. Therefore, we develop a stable training pipeline that can be considered a sequential operation on both the image and the model. Regarding the image, we propose an exposure-guided module to perform reliable and controllable nighttime image synthesis. Regarding the model, we align the representation of images from day and night domains through multi-task contrastive learning. Finally, our model achieves day-night adaptation without seeing real nighttime images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.468, + 0.573 + ], + "angle": 0, + "content": "Our framework can serve as a plug-and-play remedy to existing daytime models. To verify its effectiveness, we conduct extensive experiments on multiple high-level nighttime vision tasks, including classification, semantic segmentation, visual place recognition, and video action recognition. Results on various benchmarks demonstrate our superiority over the state-of-the-art." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.575, + 0.398, + 0.589 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.597, + 0.468, + 0.703 + ], + "angle": 0, + "content": "- We propose a similarity min-max framework for zero-shot day-night domain adaptation. Feature similarity between the original and darkened images is minimized by image-level translation and maximized by model-level adaptation. In this way, model's performance in nighttime is improved without accessing real nighttime images." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.712, + 0.468, + 0.817 + ], + "angle": 0, + "content": "- We develop a stable training pipeline to solve this bi-level optimization problem. On the image level, we propose an exposure-guided module to perform reliable and controllable nighttime image synthesis. On the model level, we align the representation of images from day and night domains through multi-task contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.826, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- Our framework universally applies to various nighttime high-level vision tasks. Experiments on classification, semantic segmentation, visual place recognition, and video action recognition demonstrate the superiority of our method." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.597, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.65, + 0.107 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.119, + 0.892, + 0.346 + ], + "angle": 0, + "content": "Low-Light Enhancement. A straightforward approach to improve the model's performance in low light is brightening the test low-light images. Early non-learning practices exploit image processing tools such as histogram equalization [40] or image formation theories such as Retinex Theory [44]. Recent literature mainly takes advantage of the advance in deep learning. Trained on paired day-night data, some methods [33, 55, 56] simulate the image decomposition process of Retinex Theory. Others introduce adversarial learning [23] to support unpaired training. Zero-DCE [16, 30] designs a curve-based low-light enhancement model and trains in a zero-reference way. Advanced techniques, including frequency decomposition [24], feature pyramids [60, 63], and flow models [54] are also adopted in recent papers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.532 + ], + "angle": 0, + "content": "Day-Night Domain Adaptation. Nighttime high-level vision has attracted increasing attention in recent years. Apart from pre-processing with enhancement models, day-night domain adaptation is also a viable solution. YOLO-in-the-dark [47] introduces the glue layer to mitigate the day-night domain gap. MAET [8] exploits image signal processing (ISP) for nighttime image generation and uses both synthetic and real nighttime images for training. HLA-face [52] proposes a joint high-low adaptation framework driven by self-supervised learning. Others [2, 37, 45, 46, 57] employ Generative Adversarial Network (GAN) to transfer labeled daytime data to nighttime." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Zero-Shot Day-Night Domain Adaptation. Beyond Conventional adaptation, zero-shot approaches consider an even stricter condition where real nighttime images are inaccessible. For general tasks, existing methods either draw supports from extra task-irrelevant source and target domain data pairs [39, 51] or require underlying probability distribution of the target domain [20], which are inapplicable to our settings. For the day-night task, Lengyel et al. propose the Color Invariant Convolution (CICov) [29] to capture illumination-robust features. MAET [8] can be viewed as zero-shot when real nighttime images are discarded during finetuning. Besides, domain generalization methods [1, 6, 19, 26, 31, 38, 62] also apply to our settings since they do not know target domains, but they are too general to handle the complex day-night domain gap." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Despite these advances, low-light enhancement concentrates on human vision and disregards downstream nighttime vision tasks. Conventional adaptation methods require task-specific nighttime datasets, which creates extra burdens on data collection and limits their generalizability to multiple tasks. Prior zero-shot adaptation methods fail to consider image-level and model-level jointly. In this paper, we propose a novel similarity min-max framework that could outperform existing methods by a large margin." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8105" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.09, + 0.387, + 0.108 + ], + "angle": 0, + "content": "3. Similarity Min-Max Optimization" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.468, + 0.162 + ], + "angle": 0, + "content": "This section introduces our approach for zero-shot daynight domain adaptation. We first explain our motivation, then introduce the overall framework and detailed designs." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.17, + 0.199, + 0.184 + ], + "angle": 0, + "content": "3.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.193, + 0.468, + 0.374 + ], + "angle": 0, + "content": "Existing methods, generally categorized into Operator-based and Darkening-based as shown in Figure 2, come across troubles in the day-night domain adaptation problem. Operator-based methods [29] rely on the manually designed operators at the model level to handle illumination variations, which are not adaptive to real complex scenarios. Darkening-based methods transfer labeled daytime data to nighttime by ISP [8] or GAN [2,28,45,46] only at the image level. However, the former is sensor-dependent and cannot generalize across devices and datasets, while the latter requires data from the task-specific nighttime domain and thus fails to generalize to our zero-shot setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.374, + 0.468, + 0.54 + ], + "angle": 0, + "content": "Intrinsically, the most critical issue of existing methods is their ignorance of the mutual effect between pixels and features. In our work, we make the first systematic investigation on this issue and propose a similarity min-max framework that thoroughly exploits the information from two sides. In detail, at the pixel (image) level, we minimize the feature similarity between original and darkened images by day-to-night translation. While at the feature (model) level, we maximize the feature similarity by representation alignment. This joint optimization leads to representations more robust to illumination changes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.468, + 0.736 + ], + "angle": 0, + "content": "We formulate our framework as follows. Denote the feature extractor of the downstream model as \\( F(\\cdot) \\). Being robust to illumination requires the extracted feature of a daytime image \\( I \\) and its nighttime version \\( D(I) \\) to be similar, where \\( D(\\cdot) \\) represents a darkening process. The limitation of existing darkening-based methods is that their \\( D \\) does not consider the co-effect of \\( F \\). So we introduce additional constraints on \\( D \\): we require \\( D \\) to minimize the similarity between the day feature \\( F(I) \\) and the night feature \\( F(D(I)) \\). This way, we guide the darkening process with high-level vision, forming a unified framework of \\( D \\) and \\( F \\). At this point, we can integrate \\( D \\) and \\( F \\) as a min-max optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.747, + 0.468, + 0.771 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\theta_ {F}} \\min _ {\\theta_ {D}} \\quad \\operatorname {S i m} (F (I), F (D (I))), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.468, + 0.809 + ], + "angle": 0, + "content": "where \\(\\theta_{D}\\) and \\(\\theta_{F}\\) denote the parameters in \\(D\\) and \\(F\\), and \\(\\mathrm{Sim}(\\cdot, \\cdot)\\) measures the similarity between features." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.468, + 0.87 + ], + "angle": 0, + "content": "However, trivial solutions exist in Eq. (1), such as \\( D \\) generating entirely black images and \\( F \\) extracting identical features for all inputs. We add regularizations to \\( D \\) and \\( F \\) accordingly to address this problem:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.88, + 0.468, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\theta_ {F}} \\min _ {\\theta_ {D}} \\operatorname {S i m} (F (I), F (D (I))) + \\mathcal {R} _ {D} (\\theta_ {D}) - \\mathcal {R} _ {F} (\\theta_ {F}), \\tag {2}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.093, + 0.574, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.235, + 0.549, + 0.246 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.579, + 0.093, + 0.701, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.235, + 0.648, + 0.246 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.093, + 0.885, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.235, + 0.789, + 0.246 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.256, + 0.892, + 0.299 + ], + "angle": 0, + "content": "Figure 2. Comparison between different learning paradigms. D and N denote the daytime and nighttime domains, respectively. (a) Operator-based. (b) Darkening-based. (c) Our method." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.324, + 0.89, + 0.339 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}_D\\) and \\(\\mathcal{R}_F\\) are intended to prevent model collapse." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.339, + 0.891, + 0.383 + ], + "angle": 0, + "content": "How to design \\(\\mathcal{R}_D\\) and \\(\\mathcal{R}_F\\) properly is the key to solving Eq. (2). The following will introduce how we design \\(\\mathcal{R}_D\\) and \\(\\mathcal{R}_F\\) and build up the whole learning framework." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.392, + 0.821, + 0.408 + ], + "angle": 0, + "content": "3.2. Image-Level Similarity Minimization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.415, + 0.89, + 0.446 + ], + "angle": 0, + "content": "This section describes our design for the darkening module \\( D \\). We want \\( D \\) to satisfy three properties:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.453, + 0.89, + 0.5 + ], + "angle": 0, + "content": "- Stability. First and foremost, we need to prevent the similarity min-max optimization from collapsing, i.e., applying proper \\(\\mathcal{R}_D\\) in Eq. (2)." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.508, + 0.89, + 0.567 + ], + "angle": 0, + "content": "- Generalization. \\(D\\) should represent a generalized darkening process so the downstream model can learn useful knowledge from \\(D(I)\\) to handle unseen nighttime scenes." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.577, + 0.89, + 0.623 + ], + "angle": 0, + "content": "- Flexibility. We additionally expect flexible control over the degree of darkening, which could enable us to create diverse inputs beneficial for optimizing \\( F \\)." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.453, + 0.89, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.631, + 0.892, + 0.812 + ], + "angle": 0, + "content": "We design an exposure-guided pixel-wise mapping algorithm to satisfy the above properties. Unlike widely-used image-to-image darkening approaches [2, 28, 46] that rely heavily on real nighttime images, pixel-wise mapping adjusts images using a pre-selected function with learnable parameters. We empirically found that, by setting proper constraints on the mapping function, we can naturally avoid obtaining trivial solutions in the similarity min-max optimization (stability) and guarantee \\( D \\) follows a typical low-light process (generalization). Finally, we add an exposure guidance mechanism for better flexibility. The detailed design will be illustrated as follows." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.815, + 0.892, + 0.877 + ], + "angle": 0, + "content": "Darkening Process. We first define a general function for tone mapping. Given an image \\( I \\in [0,1]^{C \\cdot H \\cdot W} \\), we use a non-linear mapping \\( f \\colon [0,1] \\to [0,1] \\) and a pixel-wise adjustment map \\( \\mathcal{A} \\in [0,1]^{C \\cdot H \\cdot W} \\) to process the image:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.884, + 0.891, + 0.902 + ], + "angle": 0, + "content": "\\[\nD ^ {0} (I) = f (I, \\mathcal {A}). \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8106" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.4, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.281, + 0.328, + 0.297 + ], + "angle": 0, + "content": "(a) Image-Level Translation" + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.089, + 0.891, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.281, + 0.746, + 0.296 + ], + "angle": 0, + "content": "(b) Model-Level Adaptation" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.314, + 0.894, + 0.358 + ], + "angle": 0, + "content": "Figure 3. Our proposed similarity min-max framework for zero-shot day-night domain adaptation. (a) We first train a darkening module \\( D \\) with a fixed feature extractor to generate synthesized nighttime images that share minimum similarity with their daytime counterparts. (b) After obtaining \\( D \\), we freeze its weights and maximize the day-night feature similarity to adapt the model to nighttime." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.47, + 0.534 + ], + "angle": 0, + "content": "Typically, \\( f \\) should be monotonically increasing to preserve contrast and satisfy \\( f(1,\\alpha) = 1 \\) for all \\( \\alpha \\) to avoid information loss (e.g., gamma correction). However, the latter constraint \\( f(1,\\alpha) = 1 \\) no longer holds for darkening. Therefore, we propose an auxiliary pixel-wise adjustment using a monotonic increasing function \\( g \\): \\( [0,1] \\to [0,1] \\) parameterized by another adjustment map \\( \\mathcal{B} \\in [0,1]^{C\\cdot H\\cdot W} \\). Note that \\( g \\) only serves as a complement and should be simple to avoid taking over the role of \\( f \\). The overall darkening process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.556, + 0.469, + 0.575 + ], + "angle": 0, + "content": "\\[\nD (I) = g ^ {- 1} \\left(f \\left(g (I, \\mathcal {B}), \\mathcal {A}\\right), \\mathcal {B}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.595, + 0.469, + 0.626 + ], + "angle": 0, + "content": "Both \\(\\mathcal{A}\\) and \\(\\mathcal{B}\\) are estimated by a mapping estimator conditioned on the input image \\(I\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.631, + 0.469, + 0.752 + ], + "angle": 0, + "content": "To guarantee \\( D \\) represents a darkening process (i.e., \\( D(I) < I \\)), \\( f \\) should additionally satisfy convexity. Specifically, we let \\( f \\) be the iterative quadratic curve [16]: \\( f(x) = h^{(8)}(x), h(x,\\alpha) = \\alpha x^2 + (1 - \\alpha)x \\), and \\( g \\) be the dividing operation: \\( g(x,\\beta) = x / \\beta \\) in our implementation. Other kinds of curve forms are also considered and tested. Still, we empirically found that quadratic curves could bring slightly better results (results in Sec. 4.2)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.756, + 0.469, + 0.848 + ], + "angle": 0, + "content": "Besides, to enable flexible control over the exposure level, we feed an exposure map \\( E \\) to the mapping estimator with \\( I \\), yielding the corresponding darkened image \\( D(I,E) \\). During training, the darkening module is encouraged to align the pixel value of \\( E \\) and \\( D(I,E) \\). We use \\( D(I) \\) and \\( D(I,E) \\) interchangeably for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Similarity Minimization. The training objective of module \\( D \\) involves two parts: similarity minimization and regularization. For the former, we directly reduce the distance" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.384, + 0.619, + 0.397 + ], + "angle": 0, + "content": "between features:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.404, + 0.891, + 0.439 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D} ^ {\\text {s i m}} = \\frac {\\langle F (I) , F (D (I)) \\rangle}{\\left| \\left| F (I) \\right| \\right| _ {2} \\cdot \\left| \\left| F (D (I)) \\right| \\right| _ {2}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.446, + 0.85, + 0.46 + ], + "angle": 0, + "content": "where \\(\\langle \\cdot ,\\cdot \\rangle\\) is the inner product between two vectors." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.461, + 0.891, + 0.506 + ], + "angle": 0, + "content": "The regularization term consists of four losses. Besides a color consistency loss \\(\\mathcal{L}_{col}\\) [16] that corrects color deviations, three additional losses are proposed to regularize \\(D\\):" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.507, + 0.891, + 0.537 + ], + "angle": 0, + "content": "Firstly, conditional exposure control is adopted to align the exposure map with the corresponding generated image:" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.548, + 0.891, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c - e x p} = \\sum_ {1 \\leq i \\leq H, 1 \\leq j \\leq W} | \\hat {D} _ {i, j} (I, E) - E _ {i, j} |, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.588, + 0.89, + 0.634 + ], + "angle": 0, + "content": "where \\(\\hat{D}(I,E)\\) is the channel-wise average of \\(D(I,E)\\). During training, each exposure map \\(E\\) has identical entries uniformly sampled between \\([0,0.5]\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.635, + 0.891, + 0.709 + ], + "angle": 0, + "content": "Then we add constraints on \\(\\mathcal{A}\\). Intuitively, \\(\\mathcal{A}\\) represents the degree of illumination reduction. Illumination usually varies slowly across a scene but encounters rapid variations from object to object. Following this property, we apply a loose total variance loss:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.716, + 0.891, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l t v} (\\mathcal {A}) = \\sum_ {c \\in \\{R, G, B \\}} \\left(h \\left(\\left| \\nabla_ {x} \\mathcal {A} ^ {c} \\right|\\right) ^ {2} + h \\left(\\left| \\nabla_ {y} \\mathcal {A} ^ {c} \\right|\\right) ^ {2}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.754, + 0.891, + 0.771 + ], + "angle": 0, + "content": "\\[\nh (x) = \\max (\\alpha - | x - \\alpha |, 0), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.891, + 0.87 + ], + "angle": 0, + "content": "where \\(\\nabla_{x},\\nabla_{y}\\) are gradient operations along the horizontal and vertical axis, respectively, and \\(\\alpha\\) is a hyperparameter. Compared with the original total variance loss where \\(h\\) is the identity function, our loose version allows the network to predict values of greater difference for adjacent pixels, which is common on objects' boundaries." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.891, + 0.902 + ], + "angle": 0, + "content": "Finally, we adopt \\(\\mathcal{L}_{flex}(\\mathcal{B}) = 1 - \\mathcal{B}\\) to avoid model fitting to the exposure solely by \\(g\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8107" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.097, + 0.092, + 0.355, + 0.107 + ], + "angle": 0, + "content": "The overall training objective for \\(D\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.117, + 0.468, + 0.135 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D} = \\lambda_ {D} ^ {\\text {s i m}} \\mathcal {L} _ {D} ^ {\\text {s i m}} + \\mathcal {R} _ {D}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.138, + 0.472, + 0.168 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {D} = \\lambda_ {c - e x p} \\mathcal {L} _ {c - e x p} + \\lambda_ {c o l} \\mathcal {L} _ {c o l} + \\lambda_ {l t v} \\mathcal {L} _ {l t v} + \\lambda_ {f l e x} \\mathcal {L} _ {f l e x}. \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.18, + 0.403, + 0.196 + ], + "angle": 0, + "content": "3.3. Model-Level Similarity Maximization" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.203, + 0.469, + 0.247 + ], + "angle": 0, + "content": "The darkening module \\( D \\) grants us access to a synthetic nighttime domain. In this section, we exploit \\( D \\) to learn illumination-robust representations." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.248, + 0.469, + 0.369 + ], + "angle": 0, + "content": "Contrastive learning [5, 17] is a self-supervised learning paradigm that contrasts positive and negative image pairs. However, images of the same class in classification or adjacent scenes in segmentation will form false negative pairs, thus hurting the model's performance. To alleviate these burdens, BYOL [15] proposes a non-negative variant that only aligns the feature between positive image pairs \\(\\{v, v^{+}\\}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.379, + 0.469, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {B Y O L}} (v, v ^ {+}) = 2 - \\frac {2 \\cdot \\langle z (q (F (v))) , q ^ {\\prime} \\left(F ^ {\\prime} (v ^ {+})\\right) \\rangle}{\\left| \\left| z (q (F (v))) \\right| \\right| _ {2} \\cdot \\left| \\left| q ^ {\\prime} \\left(F ^ {\\prime} (v ^ {+})\\right) \\right| \\right| _ {2}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.426, + 0.469, + 0.501 + ], + "angle": 0, + "content": "where \\( q, q' \\) are projection heads, and \\( z \\) is the prediction head. Both of them are MLPs with a single hidden layer. Note that \\( F' \\) and \\( q' \\) share the same architecture and weight initialization with \\( F \\) and \\( q \\) but receive no gradient and are updated by exponential moving average (EMA)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.505, + 0.469, + 0.58 + ], + "angle": 0, + "content": "Similarity Maximization. Motivated by BYOL, we maximize the feature similarity between synthetic nighttime and daytime domains by non-negative contrastive learning. Given a daytime image \\(I\\) and an exposure map \\(E\\), we formulate the training objective as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.591, + 0.469, + 0.61 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {F} ^ {\\text {s i m}} = \\mathcal {L} _ {\\mathrm {B Y O L}} (I, D (I, E)) + \\mathcal {L} _ {\\mathrm {B Y O L}} (D (I, E), I). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.62, + 0.469, + 0.724 + ], + "angle": 0, + "content": "Note that the measure of feature similarity is different between Eq. (5) and Eq. (12). Directly applying Eq. (5) to train \\(F\\) brings poorer results due to potential feature degeneration. In comparison, the asymmetric projection head and stop gradient policies prevent the feature extractor \\(F\\) from collapsing, i.e., working as the regularization \\(\\mathcal{R}_F\\) in Eq. (2) together with the task loss (introduced below)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.726, + 0.469, + 0.846 + ], + "angle": 0, + "content": "Moreover, different from \\( E \\) in Eq. (6), we use a compound exposure map \\( E' \\) instead. \\( E' \\) is first initialized with identical entries uniformly sampled between [0, 0.2] for simulating nighttime illumination. This range is the same for all downstream tasks, which does not introduce task-relevant prior. Then, we add pixel-wise noise \\( z_1 \\) and patch-wise noise \\( z_2 \\) to \\( E \\) to simulate exposure discrepancy. Overall, \\( E' \\) can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.858, + 0.469, + 0.875 + ], + "angle": 0, + "content": "\\[\nE ^ {\\prime} = \\mathcal {U} (0, 0. 2) + z _ {1} + z _ {2}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.424, + 0.901 + ], + "angle": 0, + "content": "See the supplementary for details on noise injection." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.092, + 0.887, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.264, + 0.892, + 0.361 + ], + "angle": 0, + "content": "Figure 4. t-SNE [50] visualization of images' feature extracted by the original daytime model and our adapted model on CO-DaN [29]. Red, green, and blue dots represent the feature of daytime, synthesized nighttime, and real nighttime images, respectively. We only color the instances from the \"Car\" category for better visual quality. Additional visualization results are shown in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.385, + 0.892, + 0.43 + ], + "angle": 0, + "content": "Besides \\(\\mathcal{L}_F^{sim}\\), we add task-specific supervision \\(\\mathcal{L}_{task}\\) on both the original daytime and synthetic nighttime domain. The final training objective for \\(F\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.438, + 0.892, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {F} = \\lambda_ {F} ^ {\\text {s i m}} \\mathcal {L} _ {F} ^ {\\text {s i m}} + \\lambda_ {\\text {t a s k}} \\mathcal {L} _ {\\text {t a s k}}. \\tag {14}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.47, + 0.733, + 0.487 + ], + "angle": 0, + "content": "3.4. Overall Training Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.494, + 0.891, + 0.553 + ], + "angle": 0, + "content": "Having introduced the image-level similarity minimization (Section 3.2) and model-level similarity maximization (Section 3.3), this section discusses the overall pipeline, as shown in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.554, + 0.892, + 0.689 + ], + "angle": 0, + "content": "An intuitive idea is training \\( D \\) and \\( F \\) alternately like GAN [13, 64]. Nevertheless, balancing \\( D \\) and \\( F \\) increases the difficulty of parameter tuning and makes the optimization process unstable. We adopt a simple but effective two-step strategy to solve this problem: we first train \\( D \\) and keep \\( F \\) frozen, then train \\( F \\) and keep \\( D \\) frozen. Compared with the alternate strategy, our step-wise approach improves the performance on nighttime image classification (elaborate in Section 4.2) from \\( 63.84\\% \\) to \\( 65.87\\% \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.809 + ], + "angle": 0, + "content": "We could also explain the merits of our min-max framework from the perspective of adversarial training [14, 35]. Module \\(D\\) first produces the worst-case examples regarding feature similarity. Then, our model could learn the illumination-robust features by learning on these cases through similarity maximization. This technical commonality further justifies our motivation to build the similarity min-max framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Across all downstream tasks, the feature extractor and task module are initialized by daytime pre-trained models. We first freeze the feature extractor and train the darkening module (image-level translation). Then, we keep the darkening module fixed and train the feature extractor and task module jointly (model-level adaptation)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8108" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.468, + 0.108 + ], + "angle": 0, + "content": "3.5. Empirical Justifications on Darkening Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.355 + ], + "angle": 0, + "content": "Simulating nighttime conditions without accessing real nighttime images is the key to our framework. Particularly, nighttime conditions bring semantic changes in addition to illumination changes, e.g., the dark environment with artificial lights on the second real nighttime image in Figure 4. However, an accurate simulation is extremely difficult since our prior knowledge is limited to \"low illumination\". Fortunately, unlike typical day-to-night image synthesis processes [41] which target the human visual experience, ours only care about the distribution of darkened images in the feature space. Leaving aside visual quality, we are pleased to find that the feature distribution of our synthesized nighttime domain is similar to that of the real nighttime domain as visualized in Figure 4(a). This observation demonstrates that our darkening process can characterize the night domain from the model-level perspective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.47, + 0.493 + ], + "angle": 0, + "content": "Thanks to this property, the feature discrepancy between daytime and real nighttime domain is significantly reduced after model-level adaptation (red and blue dots in Figure 4). This discovery is consistent with the Maximum Mean Discrepancy (MMD) between the feature distribution of day and night modalities, which is 0.020 and 0.014 for the original and adapted models, respectively. We provide implementation details and additional empirical analysis using saliency maps in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.507, + 0.21, + 0.523 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.532, + 0.469, + 0.563 + ], + "angle": 0, + "content": "This section provides the implementation details, benchmarking results, and ablation analysis of our method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.572, + 0.295, + 0.588 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.822 + ], + "angle": 0, + "content": "Our framework widely applies to various nighttime vision tasks. In the following, we evaluate our method with four representative tasks: image classification, semantic segmentation, visual place recognition, and video action recognition. Only daytime data are accessible for training and validation, while nighttime data are only used during evaluation. We benchmark our method with three categories of methods that require no dataset-specific target domain data: low-light enhancement, zero-shot day-night domain adaptation, and domain generalization. For low-light enhancement, enhancement models are trained on their original datasets. Then we adopt them as a pre-processing step to assist the daytime baseline. The results of our method is the average of three independent trails. Additional details are provided in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.351, + 0.848 + ], + "angle": 0, + "content": "4.2. Nighttime Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We first consider one of the most fundamental vision tasks: image classification. CODaN [29] is a 10-class dataset containing a training set of 10000 daytime images" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.132 + ], + "angle": 0, + "content": "Table 1. Top-1 classification accuracy on the CODaN nighttime test set [29]. \\(\\dagger\\) denotes our re-implementation with both the original and synthesized image fed into the task module." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.143, + 0.878, + 0.433 + ], + "angle": 0, + "content": "
MethodTop-1 (%)
ResNet-18 [18]53.32
Low-Light Enhancement
EnlightenGAN [23]56.68
LEDNet [63]57.40
Zero-DCE++ [30]57.96
RUAS [33]58.36
SCI [34]58.68
URetinexNet [56]58.72
Domain Generalization
MixStyle [62]53.12
IRM [1]54.52
AdaBN [31]54.25
Zero-Shot Day-Night Domain Adaptation
MAET† [8]56.48
CICConv [29]60.32
Ours65.87
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.461, + 0.892, + 0.522 + ], + "angle": 0, + "content": "and a test set with 2500 daytime and nighttime images, respectively. We validate models on the daytime test set and evaluate them on the nighttime test set. The backbone is ResNet-18 [18]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Benchmarking results are shown in Table 1. Enhancement methods restore input low-light images from the human visual perspective while keeping the model untouched, resulting in limited performance gains. Domain generalization methods are designed for general tasks and perform poorly in unseen nighttime environments. MAET [8] relies on degrading transformation with sensor-specific parameters, which suffers from poor generalizability. CICConv [29] adopts learnable color invariant edge detectors, which are not robust to the complex illumination variation in real scenarios. In contrast, our method outperforms state-of-the-art methods by a large margin (60.32% v.s. 65.87%), demonstrating our unified framework could obtain features more robust to illumination shifts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.736, + 0.892, + 0.872 + ], + "angle": 0, + "content": "Ablation Studies. We conduct ablation studies to justify our framework design in Table 2. Firstly, we study how to design the darkening module \\(D\\) given \\(\\mathcal{L}_D^{sim}\\). The model-level adaptation stage (Section 3.3) remains the same for fair comparisons. Firstly, we replace our darkening module with heuristic image adjustment approaches, such as brightness adjustment (Brightness in PIL\\(^1\\)) and gamma correction (\\(D(I) = I^{\\gamma}\\)). We implement these two approaches using a fixed darkening hyperparameter chosen after multi-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.886, + 0.901 + ], + "angle": 0, + "content": "1https://pillow.readthedocs.io/en/stable/reference/ImageEnhance.html" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8109" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.118 + ], + "angle": 0, + "content": "Table 2. Ablation studies for module \\(D\\) and similarity losses. We report the Top-1 accuracy on the CODaN [29] nighttime test set." + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.129, + 0.436, + 0.31 + ], + "angle": 0, + "content": "
CategoryMethodTop-1 (%)
BaselineVanilla ResNet-1853.32
Module DBrightness adjustment57.96
HeuristicGamma correction63.96
Module DReciprocal curve62.60
LearnableGamma curve64.16
Similarityw/o LDSim and LFSim64.08
Lossw/o LDSim64.56
w/o LFSim64.88
Full version-65.87
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.336, + 0.308, + 0.35 + ], + "angle": 0, + "content": "ple trials and report the best score." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.469, + 0.444 + ], + "angle": 0, + "content": "Next, we test other possible curve forms for \\( f \\). Both gamma curve \\( (f(x,\\alpha) = x^{\\frac{1}{\\alpha}}, \\alpha \\in (0,1]) \\) and reciprocal curve \\( (f(x,\\alpha) = \\frac{(1 - \\alpha)\\cdot x}{1 - \\alpha\\cdot x}, \\alpha \\in [0,1)) \\) bring slightly worse results than the iterative quadratic curve. Please refer to the supplementary for implementation details of these ablations and additional results on the segmentation task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.469, + 0.504 + ], + "angle": 0, + "content": "Finally, we test our framework's performance when one or both of the similarity loss is absent. We find that either similarity loss alone can boost the model's nighttime performance while combining them achieves the best result." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.514, + 0.376, + 0.531 + ], + "angle": 0, + "content": "4.3. Nighttime Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.469, + 0.674 + ], + "angle": 0, + "content": "Next, we explore a more challenging nighttime vision task: semantic segmentation. We adopt RefineNet [32] with ResNet-101 backbone as the baseline. The daytime training dataset is Cityscapes [7], containing 2975 images for training and 500 images for validation, all with dense annotations. The nighttime testing datasets are Nighttime Driving [9] and Dark-Zurich [46]. These two datasets contain 50 coarsely annotated and 151 densely annotated nighttime street view images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.469, + 0.87 + ], + "angle": 0, + "content": "We benchmark our method in Table 3. Low-light enhancement methods yield worse results than the baseline because they perform poorly on street scenes with complex light sources. Domain generalization methods fail to mitigate the huge day-night domain gap, leading to unsatisfactory results. Note that RobustNet [6] adopts DeepLabv3 [4] architecture, which is superior to RefineNet [32] adopted in our implementation. Among zero-shot adaptation methods, MAET [8] injects too much noise into images, leading to severe performance degradation. CICConv yields better results, but the improvement is limited. In comparison, our approach improves the mIoU scores to \\(44.9\\%\\) on Nighttime Driving and \\(40.2\\%\\) on Dark-Zurich." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Figure 5 shows qualitative segmentation results on two nighttime datasets. Low-light enhancement methods per" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 3. Semantic segmentation results on Nighttime Driving [9] and Dark-Zurich [46], reported as percentage mIoU scores." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.129, + 0.885, + 0.42 + ], + "angle": 0, + "content": "
MethodNighttime DrivingDark-Zurich
RefineNet [32]34.330.6
Low-Light Enhancement
EnlightenGAN [23]25.224.9
Zero-DCE++ [30]32.728.3
RUAS [33]25.123.4
SCI [34]28.625.7
URetinexNet [56]28.124.0
LEDNet [63]27.626.6
Domain Generalization
AdaBN [31]37.231.1
RobustNet [6]33.034.5
SAN-SAW [38]28.116.0
Zero-Shot Day-Night Domain Adaptation
MAET [8]28.126.4
CICConv [29]41.234.5
Ours44.940.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.433, + 0.876, + 0.447 + ], + "angle": 0, + "content": "Table 4. Visual place recognition results on Tokyo 24/7 [49]." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.458, + 0.875, + 0.66 + ], + "angle": 0, + "content": "
MethodmAP (%)
Zero-Shot Day-Night Domain Adaptation
EdgeMAC [42]75.9
U-Net jointly [21]79.8
GeM [43]85.0
CICConv-GeM [29]88.3
Ours-GeM90.4
Day-Night Domain Adaptation \n(night images are available for training)
U-Net jointly [21]86.5
EdgeMAC + CLAHE [21]90.5
EdgeMAC + U-Net jointly [21]90.0
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.686, + 0.892, + 0.732 + ], + "angle": 0, + "content": "form poorly on nighttime street scenes. Our method better extracts information hidden by darkness and thus generates more accurate semantic maps." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.794, + 0.758 + ], + "angle": 0, + "content": "4.4. Visual Place Recognition at Night" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Then we explore visual place recognition, which aims to retrieve images that illustrate the same scene of a query image from an image pool. Unlike classification and segmentation, place recognition methods are not end-to-end during inference. We extend our method based on GeM [43] with ResNet-101 backbone. In GeM, the network receives a tuple of images \\(\\{p,q,n_1,\\dots ,n_k\\}\\) as input, in which the query \\(q\\) only matches \\(p\\). The network is trained on a contrastive loss, similar to the model-level stage in our" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8110" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.213, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.143, + 0.212, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.197, + 0.173, + 0.209 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.09, + 0.349, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.143, + 0.348, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.197, + 0.333, + 0.209 + ], + "angle": 0, + "content": "(b) Ground truth" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.09, + 0.483, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.143, + 0.484, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.378, + 0.197, + 0.461, + 0.209 + ], + "angle": 0, + "content": "(c) RefineNet" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.09, + 0.619, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.143, + 0.619, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.197, + 0.586, + 0.209 + ], + "angle": 0, + "content": "(d) MAET" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.09, + 0.755, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.143, + 0.755, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.197, + 0.723, + 0.209 + ], + "angle": 0, + "content": "(e) CICConv" + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.09, + 0.892, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.143, + 0.892, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.802, + 0.197, + 0.85, + 0.209 + ], + "angle": 0, + "content": "(f) Ours" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.212, + 0.213, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.265, + 0.213, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.323, + 0.225, + 0.335 + ], + "angle": 0, + "content": "(g)SCI" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.212, + 0.349, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.265, + 0.349, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.212, + 0.484, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.265, + 0.484, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.323, + 0.524, + 0.335 + ], + "angle": 0, + "content": "(h) URetinexNet" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.212, + 0.62, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.265, + 0.62, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.212, + 0.755, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.265, + 0.755, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.323, + 0.791, + 0.335 + ], + "angle": 0, + "content": "(i) Zero-DCE++" + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.212, + 0.892, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.265, + 0.892, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.35, + 0.865, + 0.364 + ], + "angle": 0, + "content": "Figure 5. Semantic segmentation results. For each group, the first row: Nighttime Driving [9], the second row: Dark-Zurich [46]." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.388, + 0.209, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.497, + 0.18, + 0.51 + ], + "angle": 0, + "content": "(a) Query" + }, + { + "type": "image", + "bbox": [ + 0.215, + 0.388, + 0.334, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.497, + 0.301, + 0.509 + ], + "angle": 0, + "content": "(b) GeM" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.388, + 0.459, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.497, + 0.425, + 0.51 + ], + "angle": 0, + "content": "(c) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.528, + 0.47, + 0.57 + ], + "angle": 0, + "content": "Figure 6. Qualitative visual place recognition results. (a) A night query from the Tokyo 24/7 dataset [49]. (b) Image retrieved by GeM [43]. (c) Image retrieved by our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.468, + 0.732 + ], + "angle": 0, + "content": "framework. We retain the image-level stage and modify the model-level stage in our implementation. We first train the darkening module \\( D \\) as usual. Then, we consider \\( D(p) \\) as an additional matching for \\( p \\), i.e., an input tuple contains two positive samples (instead of one) and \\( k \\) negative samples. We train our network on the Retrieval-SfM dataset [43] and evaluate it on the Tokyo 24/7 dataset [49], which contains city views in multiple illumination conditions and viewing directions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.468, + 0.839 + ], + "angle": 0, + "content": "Performance is reported as mean Average Precision (mAP) in Table 4. Results of comparison methods are borrowed from [21] and [29]. Our method outperforms all zero-shot methods and is comparable to conventional domain adaptation methods. As shown in Figure 6, the baseline method gets fooled by the night's appearance, while our model finds the correct daytime image." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.396, + 0.864 + ], + "angle": 0, + "content": "4.5. Low-Light Video Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Although initially designed for images, our method also applies to video tasks. Here we consider an 11-class low-" + }, + { + "type": "table_caption", + "bbox": [ + 0.529, + 0.388, + 0.862, + 0.403 + ], + "angle": 0, + "content": "Table 5. Video action recognition results on ARID [58]." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.414, + 0.877, + 0.595 + ], + "angle": 0, + "content": "
MethodTop-1 (%)
I3D [3]47.02
Low-Light Video Enhancement
StableLLVE [59]45.08
SMOID [22]47.27
SGZ [61]46.42
Domain Generalization & Zero-Shot Day-Night Domain Adaptation
AdaBN [31]46.17
Ours51.52
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.892, + 0.715 + ], + "angle": 0, + "content": "light video action recognition task. Normal light training data consists of 2.6k normal light video clips from HMDB51 [27], UCF101 [48], Kinetics-600 [25], and Moments in Time [36]. We evaluate our model on the official test split of the ARID dataset [58]. The action recognizer is I3D [3] based on 3D-ResNet [11]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.892, + 0.824 + ], + "angle": 0, + "content": "We extend our method to video as follows. When training the darkening module, we input frames extracted from video clips. \\(\\mathcal{A}\\) and \\(\\mathcal{B}\\) in Eq. (4) is estimated for every individual frame. We calculate \\(\\mathcal{L}_D^{sim}\\) between video clips and other losses between frames. When generating low-light videos, frames are separately fed into the curve estimator while sharing the same exposure map \\(E^{\\prime}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We report the results as Top-1 accuracy. As shown in Table 5, video enhancement methods StableLLVE [59], SMOID [22], and SGZ [61] yield a limited performance gain. Meanwhile, our approach boosts models' performance by \\(4.38\\%\\), demonstrating our superiority on videos." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8111" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.196, + 0.106 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.117, + 0.473, + 0.269 + ], + "angle": 0, + "content": "In this paper, we propose a novel approach for zero-shot day-night domain adaptation. Going beyond a simple focus on the image-level translation or model-level adaptation, we observe a complementary relationship between two aspects and build our framework upon the similarity min-max paradigm. Our proposed method can significantly boost the model's performance at nighttime without accessing the nighttime domain. Experiments on multiple datasets demonstrate the superiority and broad applicability of our approach." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.287, + 0.245, + 0.304 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.313, + 0.473, + 0.435 + ], + "angle": 0, + "content": "This work is supported by the Fundamental Research Funds for the Central Universities and the National Natural Science Foundation of China under Contract No.61772043 and a research achievement of Key Laboratory of Science, Technology and Standard in Press Industry (Key Laboratory of Intelligent Press Media Technology). This research work is also partially supported by the Basic and Frontier Research Project of PCL and the Major Key Project of PCL." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.453, + 0.175, + 0.468 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.47, + 0.507 + ], + "angle": 0, + "content": "[1] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.51, + 0.472, + 0.579 + ], + "angle": 0, + "content": "[2] Vinicius F Arruda, Thiago M Paixão, Rodrigo F Berriel, Alberto F De Souza, Claudine Badue, Nicu Sebe, and Thiago Oliveira-Santos. Cross-domain car detection using unsupervised image-to-image translation: From day to night. In IJCNN, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.582, + 0.47, + 0.622 + ], + "angle": 0, + "content": "[3] João Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.626, + 0.47, + 0.668 + ], + "angle": 0, + "content": "[4] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.671, + 0.469, + 0.712 + ], + "angle": 0, + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.715, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[6] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T. Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.773, + 0.469, + 0.841 + ], + "angle": 0, + "content": "[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.845, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[8] Ziteng Cui, Guo-Jun Qi, Lin Gu, Shaodi You, Zenghui Zhang, and Tatsuya Harada. Multitask aet with orthogonal tangent regularity for dark object detection. In ICCV, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.48, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[9] Dengxin Dai and Luc Van Gool. Dark model adaptation: Semantic image segmentation from daytime to nighttime. In ITSC, 2018. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[10] Xueqing Deng, Peng Wang, Xiaochen Lian, and Shawn Newsam. Nightlab: A dual-level architecture with hardness detection for segmentation at night. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[11] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.222, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[12] Huan Gao, Jichang Guo, Guoli Wang, and Qian Zhang. Cross-domain correlation distillation for unsupervised domain adaptation in nighttime semantic segmentation. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[14] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[15] Jean-Bastien Grill, Florian Strub, Florent Altché, C. Tallec, Pierre H. Richemond, Elena Buchatskaya, C. Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, B. Piot, K. Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.475, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[16] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In CVPR, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.532, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.575, + 0.892, + 0.614 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.617, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[19] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.66, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[20] Masato Ishii, Takashi Takenouchi, and Masashi Sugiyama. Zero-shot domain adaptation based on attribute information. In ACML, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.703, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[21] Tomas Jenicek and Ondrej Chum. No fear of the dark: Image retrieval under varying illumination conditions. In ICCV, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[22] Haiyang Jiang and Yinqiang Zheng. Learning to see moving objects in the dark. In ICCV, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[23] Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. *Enlightengan: Deep light enhancement without paired supervision. IEEE TIP*, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[24] Yeying Jin, Wenhan Yang, and Robby T Tan. Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[25] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8112" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "Tim Green, Trevor Back, Paul Natev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[26] Namyup Kim, Taeyoung Son, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Wedge: web-image assisted domain generalization for semantic segmentation. arXiv, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[27] Hildegard Kuehne, Hueihan Jhuang, Estibaliz Garrote, Tomaso A. Poggio, and Thomas Serre. HMDB: A large video database for human motion recognition. In ICCV, 2011. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[28] Hongjun Lee, Moonsoo Ra, and Whoi-Yul Kim. Nighttime data augmentation using gan for improving blind-spot detection. IEEE Access, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.316 + ], + "angle": 0, + "content": "[29] Attila Lengyel, Sourav Garg, Michael Milford, and Jan C. van Gemert. Zero-shot day-night domain adaptation with a physics prior. In ICCV, 2021. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.317, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[30] Chongyi Li, Chunle Guo, and Chen Change Loy. Learning to enhance low-light image via zero-reference deep curve estimation. IEEE TPAMI, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[31] Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting batch normalization for practical domain adaptation. In ICLRW, 2017. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.401, + 0.469, + 0.441 + ], + "angle": 0, + "content": "[32] Guosheng Lin, Anton Milan, Chunhua Shen, and Ian Reid. Refinenet: Multi-path refinement networks for high-resolution semantic segmentation. In CVPR, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[33] Risheng Liu, Long Ma, Jiaao Zhang, Xin Fan, and Zhongxuan Luo. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In CVPR, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[34] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In CVPR, 2022. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[35] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[36] Mathew Monfort, Carl Vondrick, Aude Oliva, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa M. Brown, Quanfu Fan, and Dan Gutfreund. Moments in time dataset: One million videos for event understanding. IEEE TPAMI, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.706 + ], + "angle": 0, + "content": "[37] Amitangshu Mukherjee, Ameya Joshi, Anuj Sharma, Chinmay Hegde, and Soumik Sarkar. Generative semantic domain adaptation for perception in autonomous driving. Journal of Big Data Analytics in Transportation, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[38] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[39] Kuan-Chuan Peng, Ziyan Wu, and Jan Ernst. Zero-shot deep domain adaptation. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.777, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[40] Stephen M Pizer, E Philip Amburn, John D Austin, Robert Cromartie, Ari Geselowitz, Trey Greer, Bart ter Haar Romy, John B Zimmerman, and Karel Zuiderveld. Adaptive histogram equalization and its variations. Computer Vision, Graphics, and Image Processing, 1987. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[41] Abhijith Punnappurath, Abdullah Abuolaim, Abdelrahman Abdelhamed, Alex Levinshtein, and Michael S Brown. Day-to-night image synthesis for training nighttime neural isps. In CVPR, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[42] Filip Radenovic, Giorgos Tolias, and Ondrej Chum. Deep shape matching. In ECCV, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[43] Filip Radenovic, Giorgos Tolias, and Ondrej Chum. Finetuning cnn image retrieval with no human annotation. IEEE TPAMI, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[44] Zia-ur Rahman, Daniel J Jobson, and Glenn A Woodell. Retinex processing for automatic image enhancement. Journal of Electronic Imaging, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[45] Eduardo Romera, Luis M. Bergasa, Kailun Yang, Jose M. Alvarez, and Rafael Barea. Bridging the day and night domain gap for semantic segmentation. IEEE Intelligent Vehicles Symposium, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[46] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Guided curriculum model adaptation and uncertainty-aware evaluation for semantic nighttime image segmentation. In ICCV, 2019. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[47] Yukihiro Sasagawa and Hajime Nagahara. Yolo in the dark-domain adaptation method for merging multiple models. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[48] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv, 2012. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[49] Akihiko Torii, Relja Arandjelovic, Josef Sivic, Masatoshi Okutomi, and Tomas Pajdla. 24/7 place recognition by view synthesis. In CVPR, 2015. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[50] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[51] Jinghua Wang and Jianmin Jiang. Conditional coupled generative adversarial networks for zero-shot domain adaptation. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[52] Wenjing Wang, Xinhao Wang, Wenhan Yang, and Jiaying Liu. Unsupervised face detection in the dark. IEEE TPAMI, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[53] Wenjing Wang, Zhengbo Xu, Haofeng Huang, and Jiaying Liu. Self-aligned concave curve: Illumination enhancement for unsupervised adaptation. In ACM MM, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.605, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[54] Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, LapPui Chau, and Alex Kot. Low-light image enhancement with normalizing flow. In AAAI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[55] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. In BMVC, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[56] Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Wenhan Yang, and Jianmin Jiang. Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. In CVPR, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.747, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[57] Xinyi Wu, Zhenyao Wu, Hao Guo, Lili Ju, and Song Wang. Darnet: A one-stage domain adaptation network for unsupervised nighttime semantic segmentation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[58] Yuecong Xu, Jianfei Yang, Haozhi Cao, Kezhi Mao, Jianxiong Yin, and Simon See. Arid: A new dataset for recognizing action in the dark. In Deep Learning for Human Activity Recognition, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[59] Fan Zhang, Yu Li, Shaodi You, and Ying Fu. Learning temporal consistency for low light video enhancement from single images. In CVPR, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8113" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[60] Yuzhi Zhao, Yongzhe Xu, Qiong Yan, Dingdong Yang, Xuehui Wang, and Lai-Man Po. D2hnet: Joint denoising and deblurring with hierarchical network for robust night image restoration. In ECCV, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[61] Shen Zheng and Gaurav Gupta. Semantic-guided zero-shot learning for low-light image/video enhancement. In WACVW, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[62] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. In ICLR, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.222, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[63] Shangchen Zhou, Chongyi Li, and Chen Change Loy. Lednet: Joint low-light enhancement and deblurring in the dark. In ECCV, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.469, + 0.304 + ], + "angle": 0, + "content": "[64] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8114" + } + ] +] \ No newline at end of file diff --git a/2023/Similarity Min-Max_ Zero-Shot Day-Night Domain Adaptation/d2d1b76d-0c03-43ab-96b8-c3ce7c667abf_origin.pdf b/2023/Similarity Min-Max_ Zero-Shot Day-Night Domain Adaptation/d2d1b76d-0c03-43ab-96b8-c3ce7c667abf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..087ecf577cd8698ffb92034e6e02c123fa7130ff --- /dev/null +++ b/2023/Similarity Min-Max_ Zero-Shot Day-Night Domain Adaptation/d2d1b76d-0c03-43ab-96b8-c3ce7c667abf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48c87c94c503c7e4e58661e5ee0e16f80303afbad51ab30dd42a79bf40a994e0 +size 3518817 diff --git a/2023/Similarity Min-Max_ Zero-Shot Day-Night Domain Adaptation/full.md b/2023/Similarity Min-Max_ Zero-Shot Day-Night Domain Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..737d71910f4cb16ad546fec2077ddaa7c338c673 --- /dev/null +++ b/2023/Similarity Min-Max_ Zero-Shot Day-Night Domain Adaptation/full.md @@ -0,0 +1,441 @@ +# Similarity Min-Max: Zero-Shot Day-Night Domain Adaptation + +Rundong Luo $^{1,2,3}$ Wenjing Wang $^{1}$ Wenhan Yang $^{4}$ Jiaying Liu $^{1*}$ + +$^{1}$ Wangxuan Institute of Computer Technology, Peking University + +$^{2}$ School of EECS, Peking University $^{3}$ School of CS, Peking University $^{4}$ Peng Cheng Laboratory + +# Abstract + +Low-light conditions not only hamper human visual experience but also degrade the model's performance on downstream vision tasks. While existing works make remarkable progress on day-night domain adaptation, they rely heavily on domain knowledge derived from the task-specific nighttime dataset. This paper challenges a more complicated scenario with border applicability, i.e., zero-shot day-night domain adaptation, which eliminates reliance on any nighttime data. Unlike prior zero-shot adaptation approaches emphasizing either image-level translation or model-level adaptation, we propose a similarity min-max paradigm that considers them under a unified framework. On the image level, we darken images towards minimum feature similarity to enlarge the domain gap. Then on the model level, we maximize the feature similarity between the darkened images and their normal-light counterparts for better model adaptation. To the best of our knowledge, this work represents the pioneering effort in jointly optimizing both levels, resulting in a significant improvement of model generalizability. Extensive experiments demonstrate our method's effectiveness and broad applicability on various nighttime vision tasks, including classification, semantic segmentation, visual place recognition, and video action recognition. Our project page is available at https://red-fairy.github.io/ZeroShotDayNightDA-Webpage/. + +# 1. Introduction + +Deep neural networks are sensitive to insufficient illumination, and such deficiency has posed significant threats to safety-critical computer vision applications. Intuitively, insufficient illumination can be handled by low-light enhancement methods [23, 30, 34, 56, 60, 63], which aim at restoring low-light images to normal-light. However, enhancement models do not necessarily benefit downstream high-level vision tasks as they are optimized for human visual perception + +![](images/d3327f257d5b9e3bce31a1ab221a20fb298904db601444b49884b02a9cc843f3.jpg) +Figure 1. Left: Illustration of our similarity min-max framework for zero-shot day-night domain adaptation. Right: Our framework achieves state-of-the-art results on multiple downstream high-level vision tasks without seeing real nighttime images during training. + +![](images/b034ff2126266295ad5e8fbc5a4328efe916752eeb8561921729fa569208c479.jpg) + +and neglect the need for machine vision. + +Much existing literature has focused on improving machine vision performance at night through domain adaptation. By aligning the distribution statistics between the nighttime and daytime datasets through image translation [2, 12, 45], self-supervised learning [52, 53], or multistage algorithms [10, 46, 47], these methods have greatly improved models' performance in nighttime environments. The primary assumption of domain adaptation is that the target domain data is readily available. Nevertheless, obtaining data from the task-specific target domain may be challenging in extreme practical application scenarios such as deep-space exploration and deep-sea analysis. + +To reduce the requirement on target domain data, zero-shot domain adaptation has emerged as a promising research direction, where adaptation is performed without accessing the target domain. Regarding day-night domain adaptation, the primary challenge is learning illumination-robust representations generalizable to both day and night modalities. To accomplish this goal under zero-shot constraints, Lengyel et al. [29] proposed a color invariant convolution for handling illumination changes. Cui et al. [8] designed a Reverse ISP pipeline and generated synthetic nighttime images with pseudo labels. However, image-level methods simply consider synthetic nighttime as pseudo + +labeled data and overlook model-level feature extraction; model-level methods focus on adjusting model architecture but neglect image-level nighttime characteristics. Neither is effective enough capture the illumination-robust representations that could bridge the complex day-night domain gap. + +From this point of view, we devise a similarity min-max framework that involves two levels, as illustrated in Figure 1. On the image level, we generate a synthetic nighttime domain that shares minimum feature similarity with the daytime domain to enlarge the domain gap. On the model level, we learn illumination-robust representations by maximizing the feature similarity of images from the two domains for better model adaptation. + +Intuitive as it seems, solving this bi-level optimization problem is untrivial. Directly solving it may yield unsatisfactory results, e.g., meaningless images filled with zero values or identical features given all inputs. Therefore, we develop a stable training pipeline that can be considered a sequential operation on both the image and the model. Regarding the image, we propose an exposure-guided module to perform reliable and controllable nighttime image synthesis. Regarding the model, we align the representation of images from day and night domains through multi-task contrastive learning. Finally, our model achieves day-night adaptation without seeing real nighttime images. + +Our framework can serve as a plug-and-play remedy to existing daytime models. To verify its effectiveness, we conduct extensive experiments on multiple high-level nighttime vision tasks, including classification, semantic segmentation, visual place recognition, and video action recognition. Results on various benchmarks demonstrate our superiority over the state-of-the-art. + +Our contributions are summarized as follows: + +- We propose a similarity min-max framework for zero-shot day-night domain adaptation. Feature similarity between the original and darkened images is minimized by image-level translation and maximized by model-level adaptation. In this way, model's performance in nighttime is improved without accessing real nighttime images. +- We develop a stable training pipeline to solve this bi-level optimization problem. On the image level, we propose an exposure-guided module to perform reliable and controllable nighttime image synthesis. On the model level, we align the representation of images from day and night domains through multi-task contrastive learning. +- Our framework universally applies to various nighttime high-level vision tasks. Experiments on classification, semantic segmentation, visual place recognition, and video action recognition demonstrate the superiority of our method. + +# 2. Related Works + +Low-Light Enhancement. A straightforward approach to improve the model's performance in low light is brightening the test low-light images. Early non-learning practices exploit image processing tools such as histogram equalization [40] or image formation theories such as Retinex Theory [44]. Recent literature mainly takes advantage of the advance in deep learning. Trained on paired day-night data, some methods [33, 55, 56] simulate the image decomposition process of Retinex Theory. Others introduce adversarial learning [23] to support unpaired training. Zero-DCE [16, 30] designs a curve-based low-light enhancement model and trains in a zero-reference way. Advanced techniques, including frequency decomposition [24], feature pyramids [60, 63], and flow models [54] are also adopted in recent papers. + +Day-Night Domain Adaptation. Nighttime high-level vision has attracted increasing attention in recent years. Apart from pre-processing with enhancement models, day-night domain adaptation is also a viable solution. YOLO-in-the-dark [47] introduces the glue layer to mitigate the day-night domain gap. MAET [8] exploits image signal processing (ISP) for nighttime image generation and uses both synthetic and real nighttime images for training. HLA-face [52] proposes a joint high-low adaptation framework driven by self-supervised learning. Others [2, 37, 45, 46, 57] employ Generative Adversarial Network (GAN) to transfer labeled daytime data to nighttime. + +Zero-Shot Day-Night Domain Adaptation. Beyond Conventional adaptation, zero-shot approaches consider an even stricter condition where real nighttime images are inaccessible. For general tasks, existing methods either draw supports from extra task-irrelevant source and target domain data pairs [39, 51] or require underlying probability distribution of the target domain [20], which are inapplicable to our settings. For the day-night task, Lengyel et al. propose the Color Invariant Convolution (CICov) [29] to capture illumination-robust features. MAET [8] can be viewed as zero-shot when real nighttime images are discarded during finetuning. Besides, domain generalization methods [1, 6, 19, 26, 31, 38, 62] also apply to our settings since they do not know target domains, but they are too general to handle the complex day-night domain gap. + +Despite these advances, low-light enhancement concentrates on human vision and disregards downstream nighttime vision tasks. Conventional adaptation methods require task-specific nighttime datasets, which creates extra burdens on data collection and limits their generalizability to multiple tasks. Prior zero-shot adaptation methods fail to consider image-level and model-level jointly. In this paper, we propose a novel similarity min-max framework that could outperform existing methods by a large margin. + +# 3. Similarity Min-Max Optimization + +This section introduces our approach for zero-shot daynight domain adaptation. We first explain our motivation, then introduce the overall framework and detailed designs. + +# 3.1. Motivation + +Existing methods, generally categorized into Operator-based and Darkening-based as shown in Figure 2, come across troubles in the day-night domain adaptation problem. Operator-based methods [29] rely on the manually designed operators at the model level to handle illumination variations, which are not adaptive to real complex scenarios. Darkening-based methods transfer labeled daytime data to nighttime by ISP [8] or GAN [2,28,45,46] only at the image level. However, the former is sensor-dependent and cannot generalize across devices and datasets, while the latter requires data from the task-specific nighttime domain and thus fails to generalize to our zero-shot setting. + +Intrinsically, the most critical issue of existing methods is their ignorance of the mutual effect between pixels and features. In our work, we make the first systematic investigation on this issue and propose a similarity min-max framework that thoroughly exploits the information from two sides. In detail, at the pixel (image) level, we minimize the feature similarity between original and darkened images by day-to-night translation. While at the feature (model) level, we maximize the feature similarity by representation alignment. This joint optimization leads to representations more robust to illumination changes. + +We formulate our framework as follows. Denote the feature extractor of the downstream model as $F(\cdot)$ . Being robust to illumination requires the extracted feature of a daytime image $I$ and its nighttime version $D(I)$ to be similar, where $D(\cdot)$ represents a darkening process. The limitation of existing darkening-based methods is that their $D$ does not consider the co-effect of $F$ . So we introduce additional constraints on $D$ : we require $D$ to minimize the similarity between the day feature $F(I)$ and the night feature $F(D(I))$ . This way, we guide the darkening process with high-level vision, forming a unified framework of $D$ and $F$ . At this point, we can integrate $D$ and $F$ as a min-max optimization problem: + +$$ +\max _ {\theta_ {F}} \min _ {\theta_ {D}} \quad \operatorname {S i m} (F (I), F (D (I))), \tag {1} +$$ + +where $\theta_{D}$ and $\theta_{F}$ denote the parameters in $D$ and $F$ , and $\mathrm{Sim}(\cdot, \cdot)$ measures the similarity between features. + +However, trivial solutions exist in Eq. (1), such as $D$ generating entirely black images and $F$ extracting identical features for all inputs. We add regularizations to $D$ and $F$ accordingly to address this problem: + +$$ +\max _ {\theta_ {F}} \min _ {\theta_ {D}} \operatorname {S i m} (F (I), F (D (I))) + \mathcal {R} _ {D} (\theta_ {D}) - \mathcal {R} _ {F} (\theta_ {F}), \tag {2} +$$ + +![](images/3e038f359b6f9c02a61caa3907af7a7ffc45758d3278b4e6d432c3efc4bdb6a3.jpg) +(a) + +![](images/bf9ec25602bbe837fdb4a3226ba6fc47352f5e75c1461704efe75688fb62a48e.jpg) +(b) + +![](images/43ce8ed74e5868059a3467aaa757985b8716b65ed79807bee41e0d11eac94bf1.jpg) +(c) +Figure 2. Comparison between different learning paradigms. D and N denote the daytime and nighttime domains, respectively. (a) Operator-based. (b) Darkening-based. (c) Our method. + +where $\mathcal{R}_D$ and $\mathcal{R}_F$ are intended to prevent model collapse. + +How to design $\mathcal{R}_D$ and $\mathcal{R}_F$ properly is the key to solving Eq. (2). The following will introduce how we design $\mathcal{R}_D$ and $\mathcal{R}_F$ and build up the whole learning framework. + +# 3.2. Image-Level Similarity Minimization + +This section describes our design for the darkening module $D$ . We want $D$ to satisfy three properties: + +- Stability. First and foremost, we need to prevent the similarity min-max optimization from collapsing, i.e., applying proper $\mathcal{R}_D$ in Eq. (2). +- Generalization. $D$ should represent a generalized darkening process so the downstream model can learn useful knowledge from $D(I)$ to handle unseen nighttime scenes. +- Flexibility. We additionally expect flexible control over the degree of darkening, which could enable us to create diverse inputs beneficial for optimizing $F$ . + +We design an exposure-guided pixel-wise mapping algorithm to satisfy the above properties. Unlike widely-used image-to-image darkening approaches [2, 28, 46] that rely heavily on real nighttime images, pixel-wise mapping adjusts images using a pre-selected function with learnable parameters. We empirically found that, by setting proper constraints on the mapping function, we can naturally avoid obtaining trivial solutions in the similarity min-max optimization (stability) and guarantee $D$ follows a typical low-light process (generalization). Finally, we add an exposure guidance mechanism for better flexibility. The detailed design will be illustrated as follows. + +Darkening Process. We first define a general function for tone mapping. Given an image $I \in [0,1]^{C \cdot H \cdot W}$ , we use a non-linear mapping $f \colon [0,1] \to [0,1]$ and a pixel-wise adjustment map $\mathcal{A} \in [0,1]^{C \cdot H \cdot W}$ to process the image: + +$$ +D ^ {0} (I) = f (I, \mathcal {A}). \tag {3} +$$ + +![](images/953c17127184e661c998bea30ca2ad8e9c642ea338396f8b7fe8a23c11fc20a1.jpg) +(a) Image-Level Translation + +![](images/39e387222bde2848badf85a19314f557cee3180ca842609728545c05ee8bf139.jpg) +(b) Model-Level Adaptation +Figure 3. Our proposed similarity min-max framework for zero-shot day-night domain adaptation. (a) We first train a darkening module $D$ with a fixed feature extractor to generate synthesized nighttime images that share minimum similarity with their daytime counterparts. (b) After obtaining $D$ , we freeze its weights and maximize the day-night feature similarity to adapt the model to nighttime. + +Typically, $f$ should be monotonically increasing to preserve contrast and satisfy $f(1,\alpha) = 1$ for all $\alpha$ to avoid information loss (e.g., gamma correction). However, the latter constraint $f(1,\alpha) = 1$ no longer holds for darkening. Therefore, we propose an auxiliary pixel-wise adjustment using a monotonic increasing function $g$ : $[0,1] \to [0,1]$ parameterized by another adjustment map $\mathcal{B} \in [0,1]^{C\cdot H\cdot W}$ . Note that $g$ only serves as a complement and should be simple to avoid taking over the role of $f$ . The overall darkening process is formulated as: + +$$ +D (I) = g ^ {- 1} \left(f \left(g (I, \mathcal {B}), \mathcal {A}\right), \mathcal {B}\right). \tag {4} +$$ + +Both $\mathcal{A}$ and $\mathcal{B}$ are estimated by a mapping estimator conditioned on the input image $I$ . + +To guarantee $D$ represents a darkening process (i.e., $D(I) < I$ ), $f$ should additionally satisfy convexity. Specifically, we let $f$ be the iterative quadratic curve [16]: $f(x) = h^{(8)}(x), h(x,\alpha) = \alpha x^2 + (1 - \alpha)x$ , and $g$ be the dividing operation: $g(x,\beta) = x / \beta$ in our implementation. Other kinds of curve forms are also considered and tested. Still, we empirically found that quadratic curves could bring slightly better results (results in Sec. 4.2). + +Besides, to enable flexible control over the exposure level, we feed an exposure map $E$ to the mapping estimator with $I$ , yielding the corresponding darkened image $D(I,E)$ . During training, the darkening module is encouraged to align the pixel value of $E$ and $D(I,E)$ . We use $D(I)$ and $D(I,E)$ interchangeably for simplicity. + +Similarity Minimization. The training objective of module $D$ involves two parts: similarity minimization and regularization. For the former, we directly reduce the distance + +between features: + +$$ +\mathcal {L} _ {D} ^ {\text {s i m}} = \frac {\langle F (I) , F (D (I)) \rangle}{\left| \left| F (I) \right| \right| _ {2} \cdot \left| \left| F (D (I)) \right| \right| _ {2}}, \tag {5} +$$ + +where $\langle \cdot ,\cdot \rangle$ is the inner product between two vectors. + +The regularization term consists of four losses. Besides a color consistency loss $\mathcal{L}_{col}$ [16] that corrects color deviations, three additional losses are proposed to regularize $D$ : + +Firstly, conditional exposure control is adopted to align the exposure map with the corresponding generated image: + +$$ +\mathcal {L} _ {c - e x p} = \sum_ {1 \leq i \leq H, 1 \leq j \leq W} | \hat {D} _ {i, j} (I, E) - E _ {i, j} |, \tag {6} +$$ + +where $\hat{D}(I,E)$ is the channel-wise average of $D(I,E)$ . During training, each exposure map $E$ has identical entries uniformly sampled between $[0,0.5]$ . + +Then we add constraints on $\mathcal{A}$ . Intuitively, $\mathcal{A}$ represents the degree of illumination reduction. Illumination usually varies slowly across a scene but encounters rapid variations from object to object. Following this property, we apply a loose total variance loss: + +$$ +\mathcal {L} _ {l t v} (\mathcal {A}) = \sum_ {c \in \{R, G, B \}} \left(h \left(\left| \nabla_ {x} \mathcal {A} ^ {c} \right|\right) ^ {2} + h \left(\left| \nabla_ {y} \mathcal {A} ^ {c} \right|\right) ^ {2}\right), \tag {7} +$$ + +$$ +h (x) = \max (\alpha - | x - \alpha |, 0), \tag {8} +$$ + +where $\nabla_{x},\nabla_{y}$ are gradient operations along the horizontal and vertical axis, respectively, and $\alpha$ is a hyperparameter. Compared with the original total variance loss where $h$ is the identity function, our loose version allows the network to predict values of greater difference for adjacent pixels, which is common on objects' boundaries. + +Finally, we adopt $\mathcal{L}_{flex}(\mathcal{B}) = 1 - \mathcal{B}$ to avoid model fitting to the exposure solely by $g$ . + +The overall training objective for $D$ is: + +$$ +\mathcal {L} _ {D} = \lambda_ {D} ^ {\text {s i m}} \mathcal {L} _ {D} ^ {\text {s i m}} + \mathcal {R} _ {D}, \tag {9} +$$ + +$$ +\mathcal {R} _ {D} = \lambda_ {c - e x p} \mathcal {L} _ {c - e x p} + \lambda_ {c o l} \mathcal {L} _ {c o l} + \lambda_ {l t v} \mathcal {L} _ {l t v} + \lambda_ {f l e x} \mathcal {L} _ {f l e x}. \tag {10} +$$ + +# 3.3. Model-Level Similarity Maximization + +The darkening module $D$ grants us access to a synthetic nighttime domain. In this section, we exploit $D$ to learn illumination-robust representations. + +Contrastive learning [5, 17] is a self-supervised learning paradigm that contrasts positive and negative image pairs. However, images of the same class in classification or adjacent scenes in segmentation will form false negative pairs, thus hurting the model's performance. To alleviate these burdens, BYOL [15] proposes a non-negative variant that only aligns the feature between positive image pairs $\{v, v^{+}\}$ : + +$$ +\mathcal {L} _ {\mathrm {B Y O L}} (v, v ^ {+}) = 2 - \frac {2 \cdot \langle z (q (F (v))) , q ^ {\prime} \left(F ^ {\prime} (v ^ {+})\right) \rangle}{\left| \left| z (q (F (v))) \right| \right| _ {2} \cdot \left| \left| q ^ {\prime} \left(F ^ {\prime} (v ^ {+})\right) \right| \right| _ {2}}, \tag {11} +$$ + +where $q, q'$ are projection heads, and $z$ is the prediction head. Both of them are MLPs with a single hidden layer. Note that $F'$ and $q'$ share the same architecture and weight initialization with $F$ and $q$ but receive no gradient and are updated by exponential moving average (EMA). + +Similarity Maximization. Motivated by BYOL, we maximize the feature similarity between synthetic nighttime and daytime domains by non-negative contrastive learning. Given a daytime image $I$ and an exposure map $E$ , we formulate the training objective as follows: + +$$ +\mathcal {L} _ {F} ^ {\text {s i m}} = \mathcal {L} _ {\mathrm {B Y O L}} (I, D (I, E)) + \mathcal {L} _ {\mathrm {B Y O L}} (D (I, E), I). \tag {12} +$$ + +Note that the measure of feature similarity is different between Eq. (5) and Eq. (12). Directly applying Eq. (5) to train $F$ brings poorer results due to potential feature degeneration. In comparison, the asymmetric projection head and stop gradient policies prevent the feature extractor $F$ from collapsing, i.e., working as the regularization $\mathcal{R}_F$ in Eq. (2) together with the task loss (introduced below). + +Moreover, different from $E$ in Eq. (6), we use a compound exposure map $E'$ instead. $E'$ is first initialized with identical entries uniformly sampled between [0, 0.2] for simulating nighttime illumination. This range is the same for all downstream tasks, which does not introduce task-relevant prior. Then, we add pixel-wise noise $z_1$ and patch-wise noise $z_2$ to $E$ to simulate exposure discrepancy. Overall, $E'$ can be represented as: + +$$ +E ^ {\prime} = \mathcal {U} (0, 0. 2) + z _ {1} + z _ {2}. \tag {13} +$$ + +See the supplementary for details on noise injection. + +![](images/76228bff88e809d282988be7f1c83351000e58b3b658172c18577f02be06583d.jpg) +Figure 4. t-SNE [50] visualization of images' feature extracted by the original daytime model and our adapted model on CO-DaN [29]. Red, green, and blue dots represent the feature of daytime, synthesized nighttime, and real nighttime images, respectively. We only color the instances from the "Car" category for better visual quality. Additional visualization results are shown in the supplementary. + +Besides $\mathcal{L}_F^{sim}$ , we add task-specific supervision $\mathcal{L}_{task}$ on both the original daytime and synthetic nighttime domain. The final training objective for $F$ is: + +$$ +\mathcal {L} _ {F} = \lambda_ {F} ^ {\text {s i m}} \mathcal {L} _ {F} ^ {\text {s i m}} + \lambda_ {\text {t a s k}} \mathcal {L} _ {\text {t a s k}}. \tag {14} +$$ + +# 3.4. Overall Training Pipeline + +Having introduced the image-level similarity minimization (Section 3.2) and model-level similarity maximization (Section 3.3), this section discusses the overall pipeline, as shown in Figure 3. + +An intuitive idea is training $D$ and $F$ alternately like GAN [13, 64]. Nevertheless, balancing $D$ and $F$ increases the difficulty of parameter tuning and makes the optimization process unstable. We adopt a simple but effective two-step strategy to solve this problem: we first train $D$ and keep $F$ frozen, then train $F$ and keep $D$ frozen. Compared with the alternate strategy, our step-wise approach improves the performance on nighttime image classification (elaborate in Section 4.2) from $63.84\%$ to $65.87\%$ . + +We could also explain the merits of our min-max framework from the perspective of adversarial training [14, 35]. Module $D$ first produces the worst-case examples regarding feature similarity. Then, our model could learn the illumination-robust features by learning on these cases through similarity maximization. This technical commonality further justifies our motivation to build the similarity min-max framework. + +Across all downstream tasks, the feature extractor and task module are initialized by daytime pre-trained models. We first freeze the feature extractor and train the darkening module (image-level translation). Then, we keep the darkening module fixed and train the feature extractor and task module jointly (model-level adaptation). + +# 3.5. Empirical Justifications on Darkening Module + +Simulating nighttime conditions without accessing real nighttime images is the key to our framework. Particularly, nighttime conditions bring semantic changes in addition to illumination changes, e.g., the dark environment with artificial lights on the second real nighttime image in Figure 4. However, an accurate simulation is extremely difficult since our prior knowledge is limited to "low illumination". Fortunately, unlike typical day-to-night image synthesis processes [41] which target the human visual experience, ours only care about the distribution of darkened images in the feature space. Leaving aside visual quality, we are pleased to find that the feature distribution of our synthesized nighttime domain is similar to that of the real nighttime domain as visualized in Figure 4(a). This observation demonstrates that our darkening process can characterize the night domain from the model-level perspective. + +Thanks to this property, the feature discrepancy between daytime and real nighttime domain is significantly reduced after model-level adaptation (red and blue dots in Figure 4). This discovery is consistent with the Maximum Mean Discrepancy (MMD) between the feature distribution of day and night modalities, which is 0.020 and 0.014 for the original and adapted models, respectively. We provide implementation details and additional empirical analysis using saliency maps in the supplementary. + +# 4. Experiments + +This section provides the implementation details, benchmarking results, and ablation analysis of our method. + +# 4.1. Implementation Details + +Our framework widely applies to various nighttime vision tasks. In the following, we evaluate our method with four representative tasks: image classification, semantic segmentation, visual place recognition, and video action recognition. Only daytime data are accessible for training and validation, while nighttime data are only used during evaluation. We benchmark our method with three categories of methods that require no dataset-specific target domain data: low-light enhancement, zero-shot day-night domain adaptation, and domain generalization. For low-light enhancement, enhancement models are trained on their original datasets. Then we adopt them as a pre-processing step to assist the daytime baseline. The results of our method is the average of three independent trails. Additional details are provided in the supplementary. + +# 4.2. Nighttime Image Classification + +We first consider one of the most fundamental vision tasks: image classification. CODaN [29] is a 10-class dataset containing a training set of 10000 daytime images + +Table 1. Top-1 classification accuracy on the CODaN nighttime test set [29]. $\dagger$ denotes our re-implementation with both the original and synthesized image fed into the task module. + +
MethodTop-1 (%)
ResNet-18 [18]53.32
Low-Light Enhancement
EnlightenGAN [23]56.68
LEDNet [63]57.40
Zero-DCE++ [30]57.96
RUAS [33]58.36
SCI [34]58.68
URetinexNet [56]58.72
Domain Generalization
MixStyle [62]53.12
IRM [1]54.52
AdaBN [31]54.25
Zero-Shot Day-Night Domain Adaptation
MAET† [8]56.48
CICConv [29]60.32
Ours65.87
+ +and a test set with 2500 daytime and nighttime images, respectively. We validate models on the daytime test set and evaluate them on the nighttime test set. The backbone is ResNet-18 [18]. + +Benchmarking results are shown in Table 1. Enhancement methods restore input low-light images from the human visual perspective while keeping the model untouched, resulting in limited performance gains. Domain generalization methods are designed for general tasks and perform poorly in unseen nighttime environments. MAET [8] relies on degrading transformation with sensor-specific parameters, which suffers from poor generalizability. CICConv [29] adopts learnable color invariant edge detectors, which are not robust to the complex illumination variation in real scenarios. In contrast, our method outperforms state-of-the-art methods by a large margin (60.32% v.s. 65.87%), demonstrating our unified framework could obtain features more robust to illumination shifts. + +Ablation Studies. We conduct ablation studies to justify our framework design in Table 2. Firstly, we study how to design the darkening module $D$ given $\mathcal{L}_D^{sim}$ . The model-level adaptation stage (Section 3.3) remains the same for fair comparisons. Firstly, we replace our darkening module with heuristic image adjustment approaches, such as brightness adjustment (Brightness in PIL $^1$ ) and gamma correction ( $D(I) = I^{\gamma}$ ). We implement these two approaches using a fixed darkening hyperparameter chosen after multi- + +Table 2. Ablation studies for module $D$ and similarity losses. We report the Top-1 accuracy on the CODaN [29] nighttime test set. + +
CategoryMethodTop-1 (%)
BaselineVanilla ResNet-1853.32
Module DBrightness adjustment57.96
HeuristicGamma correction63.96
Module DReciprocal curve62.60
LearnableGamma curve64.16
Similarityw/o LDSim and LFSim64.08
Lossw/o LDSim64.56
w/o LFSim64.88
Full version-65.87
+ +ple trials and report the best score. + +Next, we test other possible curve forms for $f$ . Both gamma curve $(f(x,\alpha) = x^{\frac{1}{\alpha}}, \alpha \in (0,1])$ and reciprocal curve $(f(x,\alpha) = \frac{(1 - \alpha)\cdot x}{1 - \alpha\cdot x}, \alpha \in [0,1))$ bring slightly worse results than the iterative quadratic curve. Please refer to the supplementary for implementation details of these ablations and additional results on the segmentation task. + +Finally, we test our framework's performance when one or both of the similarity loss is absent. We find that either similarity loss alone can boost the model's nighttime performance while combining them achieves the best result. + +# 4.3. Nighttime Semantic Segmentation + +Next, we explore a more challenging nighttime vision task: semantic segmentation. We adopt RefineNet [32] with ResNet-101 backbone as the baseline. The daytime training dataset is Cityscapes [7], containing 2975 images for training and 500 images for validation, all with dense annotations. The nighttime testing datasets are Nighttime Driving [9] and Dark-Zurich [46]. These two datasets contain 50 coarsely annotated and 151 densely annotated nighttime street view images. + +We benchmark our method in Table 3. Low-light enhancement methods yield worse results than the baseline because they perform poorly on street scenes with complex light sources. Domain generalization methods fail to mitigate the huge day-night domain gap, leading to unsatisfactory results. Note that RobustNet [6] adopts DeepLabv3 [4] architecture, which is superior to RefineNet [32] adopted in our implementation. Among zero-shot adaptation methods, MAET [8] injects too much noise into images, leading to severe performance degradation. CICConv yields better results, but the improvement is limited. In comparison, our approach improves the mIoU scores to $44.9\%$ on Nighttime Driving and $40.2\%$ on Dark-Zurich. + +Figure 5 shows qualitative segmentation results on two nighttime datasets. Low-light enhancement methods per + +Table 3. Semantic segmentation results on Nighttime Driving [9] and Dark-Zurich [46], reported as percentage mIoU scores. + +
MethodNighttime DrivingDark-Zurich
RefineNet [32]34.330.6
Low-Light Enhancement
EnlightenGAN [23]25.224.9
Zero-DCE++ [30]32.728.3
RUAS [33]25.123.4
SCI [34]28.625.7
URetinexNet [56]28.124.0
LEDNet [63]27.626.6
Domain Generalization
AdaBN [31]37.231.1
RobustNet [6]33.034.5
SAN-SAW [38]28.116.0
Zero-Shot Day-Night Domain Adaptation
MAET [8]28.126.4
CICConv [29]41.234.5
Ours44.940.2
+ +Table 4. Visual place recognition results on Tokyo 24/7 [49]. + +
MethodmAP (%)
Zero-Shot Day-Night Domain Adaptation
EdgeMAC [42]75.9
U-Net jointly [21]79.8
GeM [43]85.0
CICConv-GeM [29]88.3
Ours-GeM90.4
Day-Night Domain Adaptation +(night images are available for training)
U-Net jointly [21]86.5
EdgeMAC + CLAHE [21]90.5
EdgeMAC + U-Net jointly [21]90.0
+ +form poorly on nighttime street scenes. Our method better extracts information hidden by darkness and thus generates more accurate semantic maps. + +# 4.4. Visual Place Recognition at Night + +Then we explore visual place recognition, which aims to retrieve images that illustrate the same scene of a query image from an image pool. Unlike classification and segmentation, place recognition methods are not end-to-end during inference. We extend our method based on GeM [43] with ResNet-101 backbone. In GeM, the network receives a tuple of images $\{p,q,n_1,\dots ,n_k\}$ as input, in which the query $q$ only matches $p$ . The network is trained on a contrastive loss, similar to the model-level stage in our + +![](images/0c34cccb7bbf33208d091438771560d3643b627f27d819ee8b04e018160bbd63.jpg) + +![](images/1283e0a1259f724edc7c69db5ab16e4b4e33f511901c6088fcc7a7eec748a2ed.jpg) +(a) Input + +![](images/f242153bf0ca9531624ffbfca6435a8bc251e2481f8f43c884ba17da0cb273b0.jpg) + +![](images/4c7c62bf14bd881f5fd5aba9331243f0b27210fafadc7392ad4bd9c6bf7546ba.jpg) +(b) Ground truth + +![](images/4fa47fd874dacefa77e2c26c48d5ccb82c010d853c2f9fdc7a1e1b14be2f6231.jpg) + +![](images/ce58d14809e4597c2ab1f07668d28c502f7245e0b1118fdd70ccbac0a213c414.jpg) +(c) RefineNet + +![](images/dff8421f58b7d6ff244687f790104a06e1657ff977af59e6470b0005e719c810.jpg) + +![](images/26a89c7c67f4577da9f820c5edd3e2a1543a6f883fafb60adcf1fdf245f61615.jpg) +(d) MAET + +![](images/0c15b0e42873f543d7a7b3edfaf98e404e2419f349ebc8d41f7cdaeb38856595.jpg) + +![](images/f185793ec33a9ad215eb49cb9d824ee6b6ee4118aef5483023e8c60f98f62a7e.jpg) +(e) CICConv + +![](images/491b913ceaef18be9aa5a615cca5aa92bda6f20f1f235161785d0a4f3d5f5c29.jpg) + +![](images/77a65d83066df2462a180e2af4dbfdae0fa824928f6d29cffefcd1f47074fc09.jpg) +(f) Ours + +![](images/287f8adf80db365e86267b91809eae6f51dd06d2588839fd7be381d2240e0af3.jpg) + +![](images/158eb11862588069f2d9b43bcf528ef41b7d33a47fbca340dddf47d35ec725a2.jpg) +(g)SCI + +![](images/ccf612991c94ec0cbdadc4098eb518c82ea6697c9b23dab6557df3339139d9fa.jpg) + +![](images/190baffd73ccf3e85e47ed02edbeb1438af65faa96aaa5c23771e9c02866bbe0.jpg) +Figure 5. Semantic segmentation results. For each group, the first row: Nighttime Driving [9], the second row: Dark-Zurich [46]. + +![](images/82498635fb2d1d7504b70b1d1cc5ab9ae101cd69bbb50b18f76133f5e2595f66.jpg) + +![](images/9aead569bf4da46517ef802e185521cf3d353906ec99176561ae4f98473dd61a.jpg) +(h) URetinexNet + +![](images/fe196eb1a464f355d116bbd80c08957b2c772bf93c2262c85b1a6567e5b2fb71.jpg) + +![](images/c7212c51eb4eed20fee3b2ce313c032b30cc72540e75624827306a19c414c436.jpg) + +![](images/32c5e15404eecadc62dcfd4dde9393d3c1c1d02f5f6460147c3692220ba97b6a.jpg) + +![](images/e39f936847e30e759b6bafc371ca06ddb54f02d4245b005f1a01f4119d4fe61b.jpg) +(i) Zero-DCE++ + +![](images/194e97d964d681aacadcd6ff2819c3abb38bc532c01afcb03471194739c24e28.jpg) + +![](images/9d2705ec1807b155c053d78b81cc886453a28a63a71f155c1fc35443a50c9458.jpg) + +![](images/1602b536bd4468fed7c3749932595e07db0103c673ac3b4090e2707c8013eab4.jpg) +(a) Query +Figure 6. Qualitative visual place recognition results. (a) A night query from the Tokyo 24/7 dataset [49]. (b) Image retrieved by GeM [43]. (c) Image retrieved by our method. + +![](images/f56a9621e95f59a95a6b0fbbc099f2a39a1a81108f5969bbb6e3fec5dc6d7379.jpg) +(b) GeM + +![](images/53dd682c1230ea9c589870472abbe52ee42eae5cb3facd77118c3e9217d31cb9.jpg) +(c) Ours + +framework. We retain the image-level stage and modify the model-level stage in our implementation. We first train the darkening module $D$ as usual. Then, we consider $D(p)$ as an additional matching for $p$ , i.e., an input tuple contains two positive samples (instead of one) and $k$ negative samples. We train our network on the Retrieval-SfM dataset [43] and evaluate it on the Tokyo 24/7 dataset [49], which contains city views in multiple illumination conditions and viewing directions. + +Performance is reported as mean Average Precision (mAP) in Table 4. Results of comparison methods are borrowed from [21] and [29]. Our method outperforms all zero-shot methods and is comparable to conventional domain adaptation methods. As shown in Figure 6, the baseline method gets fooled by the night's appearance, while our model finds the correct daytime image. + +# 4.5. Low-Light Video Action Recognition + +Although initially designed for images, our method also applies to video tasks. Here we consider an 11-class low- + +Table 5. Video action recognition results on ARID [58]. + +
MethodTop-1 (%)
I3D [3]47.02
Low-Light Video Enhancement
StableLLVE [59]45.08
SMOID [22]47.27
SGZ [61]46.42
Domain Generalization & Zero-Shot Day-Night Domain Adaptation
AdaBN [31]46.17
Ours51.52
+ +light video action recognition task. Normal light training data consists of 2.6k normal light video clips from HMDB51 [27], UCF101 [48], Kinetics-600 [25], and Moments in Time [36]. We evaluate our model on the official test split of the ARID dataset [58]. The action recognizer is I3D [3] based on 3D-ResNet [11]. + +We extend our method to video as follows. When training the darkening module, we input frames extracted from video clips. $\mathcal{A}$ and $\mathcal{B}$ in Eq. (4) is estimated for every individual frame. We calculate $\mathcal{L}_D^{sim}$ between video clips and other losses between frames. When generating low-light videos, frames are separately fed into the curve estimator while sharing the same exposure map $E^{\prime}$ . + +We report the results as Top-1 accuracy. As shown in Table 5, video enhancement methods StableLLVE [59], SMOID [22], and SGZ [61] yield a limited performance gain. Meanwhile, our approach boosts models' performance by $4.38\%$ , demonstrating our superiority on videos. + +# 5. Conclusion + +In this paper, we propose a novel approach for zero-shot day-night domain adaptation. Going beyond a simple focus on the image-level translation or model-level adaptation, we observe a complementary relationship between two aspects and build our framework upon the similarity min-max paradigm. Our proposed method can significantly boost the model's performance at nighttime without accessing the nighttime domain. Experiments on multiple datasets demonstrate the superiority and broad applicability of our approach. + +# Acknowledgements + +This work is supported by the Fundamental Research Funds for the Central Universities and the National Natural Science Foundation of China under Contract No.61772043 and a research achievement of Key Laboratory of Science, Technology and Standard in Press Industry (Key Laboratory of Intelligent Press Media Technology). This research work is also partially supported by the Basic and Frontier Research Project of PCL and the Major Key Project of PCL. + +# References + +[1] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv, 2019. 2, 6 +[2] Vinicius F Arruda, Thiago M Paixão, Rodrigo F Berriel, Alberto F De Souza, Claudine Badue, Nicu Sebe, and Thiago Oliveira-Santos. Cross-domain car detection using unsupervised image-to-image translation: From day to night. In IJCNN, 2019. 1, 2, 3 +[3] João Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, 2017. 8 +[4] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv, 2017. 7 +[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 5 +[6] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T. Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, 2021. 2, 7 +[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 7 +[8] Ziteng Cui, Guo-Jun Qi, Lin Gu, Shaodi You, Zenghui Zhang, and Tatsuya Harada. Multitask aet with orthogonal tangent regularity for dark object detection. In ICCV, 2021. 1, 2, 3, 6, 7 + +[9] Dengxin Dai and Luc Van Gool. Dark model adaptation: Semantic image segmentation from daytime to nighttime. In ITSC, 2018. 7, 8 +[10] Xueqing Deng, Peng Wang, Xiaochen Lian, and Shawn Newsam. Nightlab: A dual-level architecture with hardness detection for segmentation at night. In CVPR, 2022. 1 +[11] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 8 +[12] Huan Gao, Jichang Guo, Guoli Wang, and Qian Zhang. Cross-domain correlation distillation for unsupervised domain adaptation in nighttime semantic segmentation. In CVPR, 2022. 1 +[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 2020. 5 +[14] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv, 2014. 5 +[15] Jean-Bastien Grill, Florian Strub, Florent Altché, C. Tallec, Pierre H. Richemond, Elena Buchatskaya, C. Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, B. Piot, K. Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS, 2020. 5 +[16] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In CVPR, 2020. 2, 4 +[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 5 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6 +[19] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In CVPR, 2021. 2 +[20] Masato Ishii, Takashi Takenouchi, and Masashi Sugiyama. Zero-shot domain adaptation based on attribute information. In ACML, 2019. 2 +[21] Tomas Jenicek and Ondrej Chum. No fear of the dark: Image retrieval under varying illumination conditions. In ICCV, 2019. 7, 8 +[22] Haiyang Jiang and Yinqiang Zheng. Learning to see moving objects in the dark. In ICCV, 2019. 8 +[23] Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. *Enlightengan: Deep light enhancement without paired supervision. IEEE TIP*, 2021. 1, 2, 6, 7 +[24] Yeying Jin, Wenhan Yang, and Robby T Tan. Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In ECCV, 2022. 2 +[25] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, + +Tim Green, Trevor Back, Paul Natev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv, 2017. 8 +[26] Namyup Kim, Taeyoung Son, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Wedge: web-image assisted domain generalization for semantic segmentation. arXiv, 2021. 2 +[27] Hildegard Kuehne, Hueihan Jhuang, Estibaliz Garrote, Tomaso A. Poggio, and Thomas Serre. HMDB: A large video database for human motion recognition. In ICCV, 2011. 8 +[28] Hongjun Lee, Moonsoo Ra, and Whoi-Yul Kim. Nighttime data augmentation using gan for improving blind-spot detection. IEEE Access, 2020. 3 +[29] Attila Lengyel, Sourav Garg, Michael Milford, and Jan C. van Gemert. Zero-shot day-night domain adaptation with a physics prior. In ICCV, 2021. 1, 2, 3, 5, 6, 7, 8 +[30] Chongyi Li, Chunle Guo, and Chen Change Loy. Learning to enhance low-light image via zero-reference deep curve estimation. IEEE TPAMI, 2021. 1, 2, 6, 7 +[31] Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting batch normalization for practical domain adaptation. In ICLRW, 2017. 2, 6, 7, 8 +[32] Guosheng Lin, Anton Milan, Chunhua Shen, and Ian Reid. Refinenet: Multi-path refinement networks for high-resolution semantic segmentation. In CVPR, 2017. 7 +[33] Risheng Liu, Long Ma, Jiaao Zhang, Xin Fan, and Zhongxuan Luo. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In CVPR, 2021. 2, 6, 7 +[34] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In CVPR, 2022. 1, 6, 7 +[35] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv, 2017. 5 +[36] Mathew Monfort, Carl Vondrick, Aude Oliva, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa M. Brown, Quanfu Fan, and Dan Gutfreund. Moments in time dataset: One million videos for event understanding. IEEE TPAMI, 2020. 8 +[37] Amitangshu Mukherjee, Ameya Joshi, Anuj Sharma, Chinmay Hegde, and Soumik Sarkar. Generative semantic domain adaptation for perception in autonomous driving. Journal of Big Data Analytics in Transportation, 2022. 2 +[38] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 2, 7 +[39] Kuan-Chuan Peng, Ziyan Wu, and Jan Ernst. Zero-shot deep domain adaptation. In ECCV, 2018. 2 +[40] Stephen M Pizer, E Philip Amburn, John D Austin, Robert Cromartie, Ari Geselowitz, Trey Greer, Bart ter Haar Romy, John B Zimmerman, and Karel Zuiderveld. Adaptive histogram equalization and its variations. Computer Vision, Graphics, and Image Processing, 1987. 2 +[41] Abhijith Punnappurath, Abdullah Abuolaim, Abdelrahman Abdelhamed, Alex Levinshtein, and Michael S Brown. Day-to-night image synthesis for training nighttime neural isps. In CVPR, 2022. 6 + +[42] Filip Radenovic, Giorgos Tolias, and Ondrej Chum. Deep shape matching. In ECCV, 2018. 7 +[43] Filip Radenovic, Giorgos Tolias, and Ondrej Chum. Finetuning cnn image retrieval with no human annotation. IEEE TPAMI, 2019. 7, 8 +[44] Zia-ur Rahman, Daniel J Jobson, and Glenn A Woodell. Retinex processing for automatic image enhancement. Journal of Electronic Imaging, 2004. 2 +[45] Eduardo Romera, Luis M. Bergasa, Kailun Yang, Jose M. Alvarez, and Rafael Barea. Bridging the day and night domain gap for semantic segmentation. IEEE Intelligent Vehicles Symposium, 2019. 1, 2, 3 +[46] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Guided curriculum model adaptation and uncertainty-aware evaluation for semantic nighttime image segmentation. In ICCV, 2019. 1, 2, 3, 7, 8 +[47] Yukihiro Sasagawa and Hajime Nagahara. Yolo in the dark-domain adaptation method for merging multiple models. In ECCV, 2020. 1, 2 +[48] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv, 2012. 8 +[49] Akihiko Torii, Relja Arandjelovic, Josef Sivic, Masatoshi Okutomi, and Tomas Pajdla. 24/7 place recognition by view synthesis. In CVPR, 2015. 7, 8 +[50] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 5 +[51] Jinghua Wang and Jianmin Jiang. Conditional coupled generative adversarial networks for zero-shot domain adaptation. In ICCV, 2019. 2 +[52] Wenjing Wang, Xinhao Wang, Wenhan Yang, and Jiaying Liu. Unsupervised face detection in the dark. IEEE TPAMI, 2022. 1, 2 +[53] Wenjing Wang, Zhengbo Xu, Haofeng Huang, and Jiaying Liu. Self-aligned concave curve: Illumination enhancement for unsupervised adaptation. In ACM MM, 2022. 1 +[54] Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, LapPui Chau, and Alex Kot. Low-light image enhancement with normalizing flow. In AAAI, 2022. 2 +[55] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. In BMVC, 2017. 2 +[56] Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Wenhan Yang, and Jianmin Jiang. Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. In CVPR, 2022. 1, 2, 6, 7 +[57] Xinyi Wu, Zhenyao Wu, Hao Guo, Lili Ju, and Song Wang. Darnet: A one-stage domain adaptation network for unsupervised nighttime semantic segmentation. In CVPR, 2021. 2 +[58] Yuecong Xu, Jianfei Yang, Haozhi Cao, Kezhi Mao, Jianxiong Yin, and Simon See. Arid: A new dataset for recognizing action in the dark. In Deep Learning for Human Activity Recognition, 2021. 8 +[59] Fan Zhang, Yu Li, Shaodi You, and Ying Fu. Learning temporal consistency for low light video enhancement from single images. In CVPR, 2021. 8 + +[60] Yuzhi Zhao, Yongzhe Xu, Qiong Yan, Dingdong Yang, Xuehui Wang, and Lai-Man Po. D2hnet: Joint denoising and deblurring with hierarchical network for robust night image restoration. In ECCV, 2022. 1, 2 +[61] Shen Zheng and Gaurav Gupta. Semantic-guided zero-shot learning for low-light image/video enhancement. In WACVW, 2022. 8 +[62] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. In ICLR, 2021. 2, 6 +[63] Shangchen Zhou, Chongyi Li, and Chen Change Loy. Lednet: Joint low-light enhancement and deblurring in the dark. In ECCV, 2022. 1, 2, 6, 7 +[64] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. 5 \ No newline at end of file diff --git a/2023/Similarity Min-Max_ Zero-Shot Day-Night Domain Adaptation/images.zip b/2023/Similarity Min-Max_ Zero-Shot Day-Night Domain Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..26f4f5c55a423f65ffa2bdb3072b03a3e7cc3658 --- /dev/null +++ b/2023/Similarity Min-Max_ Zero-Shot Day-Night Domain Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05cc4d685351d38cf7c938eb587249b38179ca731bd177913e597304440d2211 +size 579314 diff --git a/2023/Similarity Min-Max_ Zero-Shot Day-Night Domain Adaptation/layout.json b/2023/Similarity Min-Max_ Zero-Shot Day-Night Domain Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..07dfd67d5bb8b644515f1ab6eb333f9f3d0671f8 --- /dev/null +++ b/2023/Similarity Min-Max_ Zero-Shot Day-Night Domain Adaptation/layout.json @@ -0,0 +1,12419 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 101, + 103, + 492, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 103, + 492, + 121 + ], + "spans": [ + { + "bbox": [ + 101, + 103, + 492, + 121 + ], + "type": "text", + "content": "Similarity Min-Max: Zero-Shot Day-Night Domain Adaptation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 142, + 460, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 142, + 460, + 158 + ], + "spans": [ + { + "bbox": [ + 136, + 142, + 460, + 158 + ], + "type": "text", + "content": "Rundong Luo" + }, + { + "bbox": [ + 136, + 142, + 460, + 158 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 136, + 142, + 460, + 158 + ], + "type": "text", + "content": " Wenjing Wang" + }, + { + "bbox": [ + 136, + 142, + 460, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 142, + 460, + 158 + ], + "type": "text", + "content": " Wenhan Yang" + }, + { + "bbox": [ + 136, + 142, + 460, + 158 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 136, + 142, + 460, + 158 + ], + "type": "text", + "content": " Jiaying Liu" + }, + { + "bbox": [ + 136, + 142, + 460, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 158, + 456, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 158, + 456, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 158, + 456, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 158, + 456, + 172 + ], + "type": "text", + "content": " Wangxuan Institute of Computer Technology, Peking University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 171, + 534, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 171, + 534, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 171, + 534, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 56, + 171, + 534, + 186 + ], + "type": "text", + "content": "School of EECS, Peking University " + }, + { + "bbox": [ + 56, + 171, + 534, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 56, + 171, + 534, + 186 + ], + "type": "text", + "content": "School of CS, Peking University " + }, + { + "bbox": [ + 56, + 171, + 534, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 56, + 171, + 534, + 186 + ], + "type": "text", + "content": "Peng Cheng Laboratory" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 239, + 290, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 550 + ], + "type": "text", + "content": "Low-light conditions not only hamper human visual experience but also degrade the model's performance on downstream vision tasks. While existing works make remarkable progress on day-night domain adaptation, they rely heavily on domain knowledge derived from the task-specific nighttime dataset. This paper challenges a more complicated scenario with border applicability, i.e., zero-shot day-night domain adaptation, which eliminates reliance on any nighttime data. Unlike prior zero-shot adaptation approaches emphasizing either image-level translation or model-level adaptation, we propose a similarity min-max paradigm that considers them under a unified framework. On the image level, we darken images towards minimum feature similarity to enlarge the domain gap. Then on the model level, we maximize the feature similarity between the darkened images and their normal-light counterparts for better model adaptation. To the best of our knowledge, this work represents the pioneering effort in jointly optimizing both levels, resulting in a significant improvement of model generalizability. Extensive experiments demonstrate our method's effectiveness and broad applicability on various nighttime vision tasks, including classification, semantic segmentation, visual place recognition, and video action recognition. Our project page is available at https://red-fairy.github.io/ZeroShotDayNightDA-Webpage/." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 575, + 128, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 128, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 128, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 596, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 596, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 596, + 287, + 693 + ], + "type": "text", + "content": "Deep neural networks are sensitive to insufficient illumination, and such deficiency has posed significant threats to safety-critical computer vision applications. Intuitively, insufficient illumination can be handled by low-light enhancement methods [23, 30, 34, 56, 60, 63], which aim at restoring low-light images to normal-light. However, enhancement models do not necessarily benefit downstream high-level vision tasks as they are optimized for human visual perception" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 215, + 438, + 327 + ], + "blocks": [ + { + "bbox": [ + 309, + 215, + 438, + 327 + ], + "lines": [ + { + "bbox": [ + 309, + 215, + 438, + 327 + ], + "spans": [ + { + "bbox": [ + 309, + 215, + 438, + 327 + ], + "type": "image", + "image_path": "d3327f257d5b9e3bce31a1ab221a20fb298904db601444b49884b02a9cc843f3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 338, + 546, + 383 + ], + "lines": [ + { + "bbox": [ + 305, + 338, + 546, + 383 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 546, + 383 + ], + "type": "text", + "content": "Figure 1. Left: Illustration of our similarity min-max framework for zero-shot day-night domain adaptation. Right: Our framework achieves state-of-the-art results on multiple downstream high-level vision tasks without seeing real nighttime images during training." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 443, + 224, + 541, + 323 + ], + "blocks": [ + { + "bbox": [ + 443, + 224, + 541, + 323 + ], + "lines": [ + { + "bbox": [ + 443, + 224, + 541, + 323 + ], + "spans": [ + { + "bbox": [ + 443, + 224, + 541, + 323 + ], + "type": "image", + "image_path": "b034ff2126266295ad5e8fbc5a4328efe916752eeb8561921729fa569208c479.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 409, + 470, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 409, + 470, + 421 + ], + "spans": [ + { + "bbox": [ + 306, + 409, + 470, + 421 + ], + "type": "text", + "content": "and neglect the need for machine vision." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 423, + 546, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 546, + 567 + ], + "type": "text", + "content": "Much existing literature has focused on improving machine vision performance at night through domain adaptation. By aligning the distribution statistics between the nighttime and daytime datasets through image translation [2, 12, 45], self-supervised learning [52, 53], or multistage algorithms [10, 46, 47], these methods have greatly improved models' performance in nighttime environments. The primary assumption of domain adaptation is that the target domain data is readily available. Nevertheless, obtaining data from the task-specific target domain may be challenging in extreme practical application scenarios such as deep-space exploration and deep-sea analysis." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "To reduce the requirement on target domain data, zero-shot domain adaptation has emerged as a promising research direction, where adaptation is performed without accessing the target domain. Regarding day-night domain adaptation, the primary challenge is learning illumination-robust representations generalizable to both day and night modalities. To accomplish this goal under zero-shot constraints, Lengyel et al. [29] proposed a color invariant convolution for handling illumination changes. Cui et al. [8] designed a Reverse ISP pipeline and generated synthetic nighttime images with pseudo labels. However, image-level methods simply consider synthetic nighttime as pseudo" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8104" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 133 + ], + "type": "text", + "content": "labeled data and overlook model-level feature extraction; model-level methods focus on adjusting model architecture but neglect image-level nighttime characteristics. Neither is effective enough capture the illumination-robust representations that could bridge the complex day-night domain gap." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 286, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 227 + ], + "type": "text", + "content": "From this point of view, we devise a similarity min-max framework that involves two levels, as illustrated in Figure 1. On the image level, we generate a synthetic nighttime domain that shares minimum feature similarity with the daytime domain to enlarge the domain gap. On the model level, we learn illumination-robust representations by maximizing the feature similarity of images from the two domains for better model adaptation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 228, + 286, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 286, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 286, + 371 + ], + "type": "text", + "content": "Intuitive as it seems, solving this bi-level optimization problem is untrivial. Directly solving it may yield unsatisfactory results, e.g., meaningless images filled with zero values or identical features given all inputs. Therefore, we develop a stable training pipeline that can be considered a sequential operation on both the image and the model. Regarding the image, we propose an exposure-guided module to perform reliable and controllable nighttime image synthesis. Regarding the model, we align the representation of images from day and night domains through multi-task contrastive learning. Finally, our model achieves day-night adaptation without seeing real nighttime images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 372, + 286, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 286, + 453 + ], + "type": "text", + "content": "Our framework can serve as a plug-and-play remedy to existing daytime models. To verify its effectiveness, we conduct extensive experiments on multiple high-level nighttime vision tasks, including classification, semantic segmentation, visual place recognition, and video action recognition. Results on various benchmarks demonstrate our superiority over the state-of-the-art." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 455, + 243, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 455, + 243, + 466 + ], + "spans": [ + { + "bbox": [ + 59, + 455, + 243, + 466 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 472, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 58, + 472, + 286, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 472, + 286, + 556 + ], + "spans": [ + { + "bbox": [ + 58, + 472, + 286, + 556 + ], + "type": "text", + "content": "- We propose a similarity min-max framework for zero-shot day-night domain adaptation. Feature similarity between the original and darkened images is minimized by image-level translation and maximized by model-level adaptation. In this way, model's performance in nighttime is improved without accessing real nighttime images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 563, + 286, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 563, + 286, + 647 + ], + "spans": [ + { + "bbox": [ + 58, + 563, + 286, + 647 + ], + "type": "text", + "content": "- We develop a stable training pipeline to solve this bi-level optimization problem. On the image level, we propose an exposure-guided module to perform reliable and controllable nighttime image synthesis. On the model level, we align the representation of images from day and night domains through multi-task contrastive learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 654, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 654, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 654, + 286, + 713 + ], + "type": "text", + "content": "- Our framework universally applies to various nighttime high-level vision tasks. Experiments on classification, semantic segmentation, visual place recognition, and video action recognition demonstrate the superiority of our method." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 397, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 397, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 397, + 84 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 94, + 545, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 94, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 304, + 94, + 545, + 274 + ], + "type": "text", + "content": "Low-Light Enhancement. A straightforward approach to improve the model's performance in low light is brightening the test low-light images. Early non-learning practices exploit image processing tools such as histogram equalization [40] or image formation theories such as Retinex Theory [44]. Recent literature mainly takes advantage of the advance in deep learning. Trained on paired day-night data, some methods [33, 55, 56] simulate the image decomposition process of Retinex Theory. Others introduce adversarial learning [23] to support unpaired training. Zero-DCE [16, 30] designs a curve-based low-light enhancement model and trains in a zero-reference way. Advanced techniques, including frequency decomposition [24], feature pyramids [60, 63], and flow models [54] are also adopted in recent papers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 277, + 545, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 421 + ], + "type": "text", + "content": "Day-Night Domain Adaptation. Nighttime high-level vision has attracted increasing attention in recent years. Apart from pre-processing with enhancement models, day-night domain adaptation is also a viable solution. YOLO-in-the-dark [47] introduces the glue layer to mitigate the day-night domain gap. MAET [8] exploits image signal processing (ISP) for nighttime image generation and uses both synthetic and real nighttime images for training. HLA-face [52] proposes a joint high-low adaptation framework driven by self-supervised learning. Others [2, 37, 45, 46, 57] employ Generative Adversarial Network (GAN) to transfer labeled daytime data to nighttime." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 423, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 604 + ], + "type": "text", + "content": "Zero-Shot Day-Night Domain Adaptation. Beyond Conventional adaptation, zero-shot approaches consider an even stricter condition where real nighttime images are inaccessible. For general tasks, existing methods either draw supports from extra task-irrelevant source and target domain data pairs [39, 51] or require underlying probability distribution of the target domain [20], which are inapplicable to our settings. For the day-night task, Lengyel et al. propose the Color Invariant Convolution (CICov) [29] to capture illumination-robust features. MAET [8] can be viewed as zero-shot when real nighttime images are discarded during finetuning. Besides, domain generalization methods [1, 6, 19, 26, 31, 38, 62] also apply to our settings since they do not know target domains, but they are too general to handle the complex day-night domain gap." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Despite these advances, low-light enhancement concentrates on human vision and disregards downstream nighttime vision tasks. Conventional adaptation methods require task-specific nighttime datasets, which creates extra burdens on data collection and limits their generalizability to multiple tasks. Prior zero-shot adaptation methods fail to consider image-level and model-level jointly. In this paper, we propose a novel similarity min-max framework that could outperform existing methods by a large margin." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8105" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 236, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 236, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 236, + 85 + ], + "type": "text", + "content": "3. Similarity Min-Max Optimization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 286, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 286, + 128 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 286, + 128 + ], + "type": "text", + "content": "This section introduces our approach for zero-shot daynight domain adaptation. We first explain our motivation, then introduce the overall framework and detailed designs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 134, + 121, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 134, + 121, + 145 + ], + "spans": [ + { + "bbox": [ + 47, + 134, + 121, + 145 + ], + "type": "text", + "content": "3.1. Motivation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 152, + 286, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 152, + 286, + 296 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 286, + 296 + ], + "type": "text", + "content": "Existing methods, generally categorized into Operator-based and Darkening-based as shown in Figure 2, come across troubles in the day-night domain adaptation problem. Operator-based methods [29] rely on the manually designed operators at the model level to handle illumination variations, which are not adaptive to real complex scenarios. Darkening-based methods transfer labeled daytime data to nighttime by ISP [8] or GAN [2,28,45,46] only at the image level. However, the former is sensor-dependent and cannot generalize across devices and datasets, while the latter requires data from the task-specific nighttime domain and thus fails to generalize to our zero-shot setting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 296, + 286, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 296, + 286, + 427 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 286, + 427 + ], + "type": "text", + "content": "Intrinsically, the most critical issue of existing methods is their ignorance of the mutual effect between pixels and features. In our work, we make the first systematic investigation on this issue and propose a similarity min-max framework that thoroughly exploits the information from two sides. In detail, at the pixel (image) level, we minimize the feature similarity between original and darkened images by day-to-night translation. While at the feature (model) level, we maximize the feature similarity by representation alignment. This joint optimization leads to representations more robust to illumination changes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "text", + "content": "We formulate our framework as follows. Denote the feature extractor of the downstream model as " + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "text", + "content": ". Being robust to illumination requires the extracted feature of a daytime image " + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "text", + "content": " and its nighttime version " + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "inline_equation", + "content": "D(I)" + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "text", + "content": " to be similar, where " + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "inline_equation", + "content": "D(\\cdot)" + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "text", + "content": " represents a darkening process. The limitation of existing darkening-based methods is that their " + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "text", + "content": " does not consider the co-effect of " + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "text", + "content": ". So we introduce additional constraints on " + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "text", + "content": ": we require " + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "text", + "content": " to minimize the similarity between the day feature " + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "inline_equation", + "content": "F(I)" + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "text", + "content": " and the night feature " + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "inline_equation", + "content": "F(D(I))" + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "text", + "content": ". This way, we guide the darkening process with high-level vision, forming a unified framework of " + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "text", + "content": ". At this point, we can integrate " + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 427, + 286, + 582 + ], + "type": "text", + "content": " as a min-max optimization problem:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 591, + 286, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 591, + 286, + 610 + ], + "spans": [ + { + "bbox": [ + 96, + 591, + 286, + 610 + ], + "type": "interline_equation", + "content": "\\max _ {\\theta_ {F}} \\min _ {\\theta_ {D}} \\quad \\operatorname {S i m} (F (I), F (D (I))), \\tag {1}", + "image_path": "57afa9ba740c66ef6f8f5b80465032ea46df25adb79495fcb190edb71e3541c0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 617, + 286, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 286, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 617, + 286, + 640 + ], + "type": "inline_equation", + "content": "\\theta_{D}" + }, + { + "bbox": [ + 47, + 617, + 286, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 617, + 286, + 640 + ], + "type": "inline_equation", + "content": "\\theta_{F}" + }, + { + "bbox": [ + 47, + 617, + 286, + 640 + ], + "type": "text", + "content": " denote the parameters in " + }, + { + "bbox": [ + 47, + 617, + 286, + 640 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 617, + 286, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 617, + 286, + 640 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 617, + 286, + 640 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 617, + 286, + 640 + ], + "type": "inline_equation", + "content": "\\mathrm{Sim}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 47, + 617, + 286, + 640 + ], + "type": "text", + "content": " measures the similarity between features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 641, + 286, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 286, + 689 + ], + "type": "text", + "content": "However, trivial solutions exist in Eq. (1), such as " + }, + { + "bbox": [ + 47, + 641, + 286, + 689 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 641, + 286, + 689 + ], + "type": "text", + "content": " generating entirely black images and " + }, + { + "bbox": [ + 47, + 641, + 286, + 689 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 641, + 286, + 689 + ], + "type": "text", + "content": " extracting identical features for all inputs. We add regularizations to " + }, + { + "bbox": [ + 47, + 641, + 286, + 689 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 641, + 286, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 641, + 286, + 689 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 641, + 286, + 689 + ], + "type": "text", + "content": " accordingly to address this problem:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 52, + 696, + 286, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 696, + 286, + 716 + ], + "spans": [ + { + "bbox": [ + 52, + 696, + 286, + 716 + ], + "type": "interline_equation", + "content": "\\max _ {\\theta_ {F}} \\min _ {\\theta_ {D}} \\operatorname {S i m} (F (I), F (D (I))) + \\mathcal {R} _ {D} (\\theta_ {D}) - \\mathcal {R} _ {F} (\\theta_ {F}), \\tag {2}", + "image_path": "39155db9800514410a843875b3385c6dd81ec5486c62f56b711ffd8d5fbbf111.jpg" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 351, + 178 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 351, + 178 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 351, + 178 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 351, + 178 + ], + "type": "image", + "image_path": "3e038f359b6f9c02a61caa3907af7a7ffc45758d3278b4e6d432c3efc4bdb6a3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 186, + 335, + 194 + ], + "lines": [ + { + "bbox": [ + 325, + 186, + 335, + 194 + ], + "spans": [ + { + "bbox": [ + 325, + 186, + 335, + 194 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 354, + 73, + 429, + 182 + ], + "blocks": [ + { + "bbox": [ + 354, + 73, + 429, + 182 + ], + "lines": [ + { + "bbox": [ + 354, + 73, + 429, + 182 + ], + "spans": [ + { + "bbox": [ + 354, + 73, + 429, + 182 + ], + "type": "image", + "image_path": "bf9ec25602bbe837fdb4a3226ba6fc47352f5e75c1461704efe75688fb62a48e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 186, + 396, + 194 + ], + "lines": [ + { + "bbox": [ + 386, + 186, + 396, + 194 + ], + "spans": [ + { + "bbox": [ + 386, + 186, + 396, + 194 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 432, + 73, + 541, + 183 + ], + "blocks": [ + { + "bbox": [ + 432, + 73, + 541, + 183 + ], + "lines": [ + { + "bbox": [ + 432, + 73, + 541, + 183 + ], + "spans": [ + { + "bbox": [ + 432, + 73, + 541, + 183 + ], + "type": "image", + "image_path": "43ce8ed74e5868059a3467aaa757985b8716b65ed79807bee41e0d11eac94bf1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 186, + 482, + 194 + ], + "lines": [ + { + "bbox": [ + 471, + 186, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 471, + 186, + 482, + 194 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 202, + 545, + 236 + ], + "lines": [ + { + "bbox": [ + 305, + 202, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 305, + 202, + 545, + 236 + ], + "type": "text", + "content": "Figure 2. Comparison between different learning paradigms. D and N denote the daytime and nighttime domains, respectively. (a) Operator-based. (b) Darkening-based. (c) Our method." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 256, + 544, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 256, + 544, + 268 + ], + "spans": [ + { + "bbox": [ + 306, + 256, + 544, + 268 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 256, + 544, + 268 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_D" + }, + { + "bbox": [ + 306, + 256, + 544, + 268 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 256, + 544, + 268 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_F" + }, + { + "bbox": [ + 306, + 256, + 544, + 268 + ], + "type": "text", + "content": " are intended to prevent model collapse." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 268, + 545, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 268, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 305, + 268, + 545, + 303 + ], + "type": "text", + "content": "How to design " + }, + { + "bbox": [ + 305, + 268, + 545, + 303 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_D" + }, + { + "bbox": [ + 305, + 268, + 545, + 303 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 268, + 545, + 303 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_F" + }, + { + "bbox": [ + 305, + 268, + 545, + 303 + ], + "type": "text", + "content": " properly is the key to solving Eq. (2). The following will introduce how we design " + }, + { + "bbox": [ + 305, + 268, + 545, + 303 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_D" + }, + { + "bbox": [ + 305, + 268, + 545, + 303 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 268, + 545, + 303 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_F" + }, + { + "bbox": [ + 305, + 268, + 545, + 303 + ], + "type": "text", + "content": " and build up the whole learning framework." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 310, + 502, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 310, + 502, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 310, + 502, + 323 + ], + "type": "text", + "content": "3.2. Image-Level Similarity Minimization" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 328, + 544, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 328, + 544, + 353 + ], + "spans": [ + { + "bbox": [ + 305, + 328, + 544, + 353 + ], + "type": "text", + "content": "This section describes our design for the darkening module " + }, + { + "bbox": [ + 305, + 328, + 544, + 353 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 328, + 544, + 353 + ], + "type": "text", + "content": ". We want " + }, + { + "bbox": [ + 305, + 328, + 544, + 353 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 328, + 544, + 353 + ], + "type": "text", + "content": " to satisfy three properties:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 358, + 544, + 493 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 317, + 358, + 544, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 358, + 544, + 396 + ], + "spans": [ + { + "bbox": [ + 317, + 358, + 544, + 396 + ], + "type": "text", + "content": "- Stability. First and foremost, we need to prevent the similarity min-max optimization from collapsing, i.e., applying proper " + }, + { + "bbox": [ + 317, + 358, + 544, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_D" + }, + { + "bbox": [ + 317, + 358, + 544, + 396 + ], + "type": "text", + "content": " in Eq. (2)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 402, + 544, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 402, + 544, + 449 + ], + "spans": [ + { + "bbox": [ + 317, + 402, + 544, + 449 + ], + "type": "text", + "content": "- Generalization. " + }, + { + "bbox": [ + 317, + 402, + 544, + 449 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 317, + 402, + 544, + 449 + ], + "type": "text", + "content": " should represent a generalized darkening process so the downstream model can learn useful knowledge from " + }, + { + "bbox": [ + 317, + 402, + 544, + 449 + ], + "type": "inline_equation", + "content": "D(I)" + }, + { + "bbox": [ + 317, + 402, + 544, + 449 + ], + "type": "text", + "content": " to handle unseen nighttime scenes." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 456, + 544, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 544, + 493 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 544, + 493 + ], + "type": "text", + "content": "- Flexibility. We additionally expect flexible control over the degree of darkening, which could enable us to create diverse inputs beneficial for optimizing " + }, + { + "bbox": [ + 317, + 456, + 544, + 493 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 317, + 456, + 544, + 493 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 499, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 499, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 499, + 545, + 643 + ], + "type": "text", + "content": "We design an exposure-guided pixel-wise mapping algorithm to satisfy the above properties. Unlike widely-used image-to-image darkening approaches [2, 28, 46] that rely heavily on real nighttime images, pixel-wise mapping adjusts images using a pre-selected function with learnable parameters. We empirically found that, by setting proper constraints on the mapping function, we can naturally avoid obtaining trivial solutions in the similarity min-max optimization (stability) and guarantee " + }, + { + "bbox": [ + 304, + 499, + 545, + 643 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 499, + 545, + 643 + ], + "type": "text", + "content": " follows a typical low-light process (generalization). Finally, we add an exposure guidance mechanism for better flexibility. The detailed design will be illustrated as follows." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "spans": [ + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "type": "text", + "content": "Darkening Process. We first define a general function for tone mapping. Given an image " + }, + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "type": "inline_equation", + "content": "I \\in [0,1]^{C \\cdot H \\cdot W}" + }, + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "type": "text", + "content": ", we use a non-linear mapping " + }, + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "type": "inline_equation", + "content": "f \\colon [0,1] \\to [0,1]" + }, + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "type": "text", + "content": " and a pixel-wise adjustment map " + }, + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{A} \\in [0,1]^{C \\cdot H \\cdot W}" + }, + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "type": "text", + "content": " to process the image:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 388, + 700, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 700, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 388, + 700, + 545, + 714 + ], + "type": "interline_equation", + "content": "D ^ {0} (I) = f (I, \\mathcal {A}). \\tag {3}", + "image_path": "d4511b1ed8e2c252d391a3cfa1f16f856db96437253af7759b4e2d8e57c941cf.jpg" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8106" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 244, + 209 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 244, + 209 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 244, + 209 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 244, + 209 + ], + "type": "image", + "image_path": "953c17127184e661c998bea30ca2ad8e9c642ea338396f8b7fe8a23c11fc20a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 222, + 200, + 235 + ], + "lines": [ + { + "bbox": [ + 89, + 222, + 200, + 235 + ], + "spans": [ + { + "bbox": [ + 89, + 222, + 200, + 235 + ], + "type": "text", + "content": "(a) Image-Level Translation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 255, + 70, + 545, + 211 + ], + "blocks": [ + { + "bbox": [ + 255, + 70, + 545, + 211 + ], + "lines": [ + { + "bbox": [ + 255, + 70, + 545, + 211 + ], + "spans": [ + { + "bbox": [ + 255, + 70, + 545, + 211 + ], + "type": "image", + "image_path": "39e387222bde2848badf85a19314f557cee3180ca842609728545c05ee8bf139.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 222, + 456, + 234 + ], + "lines": [ + { + "bbox": [ + 345, + 222, + 456, + 234 + ], + "spans": [ + { + "bbox": [ + 345, + 222, + 456, + 234 + ], + "type": "text", + "content": "(b) Model-Level Adaptation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 248, + 547, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 248, + 547, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 547, + 283 + ], + "type": "text", + "content": "Figure 3. Our proposed similarity min-max framework for zero-shot day-night domain adaptation. (a) We first train a darkening module " + }, + { + "bbox": [ + 46, + 248, + 547, + 283 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 248, + 547, + 283 + ], + "type": "text", + "content": " with a fixed feature extractor to generate synthesized nighttime images that share minimum similarity with their daytime counterparts. (b) After obtaining " + }, + { + "bbox": [ + 46, + 248, + 547, + 283 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 248, + 547, + 283 + ], + "type": "text", + "content": ", we freeze its weights and maximize the day-night feature similarity to adapt the model to nighttime." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "text", + "content": "Typically, " + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "text", + "content": " should be monotonically increasing to preserve contrast and satisfy " + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "inline_equation", + "content": "f(1,\\alpha) = 1" + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "text", + "content": " to avoid information loss (e.g., gamma correction). However, the latter constraint " + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "inline_equation", + "content": "f(1,\\alpha) = 1" + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "text", + "content": " no longer holds for darkening. Therefore, we propose an auxiliary pixel-wise adjustment using a monotonic increasing function " + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "inline_equation", + "content": "[0,1] \\to [0,1]" + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "text", + "content": " parameterized by another adjustment map " + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{B} \\in [0,1]^{C\\cdot H\\cdot W}" + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "text", + "content": " only serves as a complement and should be simple to avoid taking over the role of " + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 303, + 287, + 422 + ], + "type": "text", + "content": ". The overall darkening process is formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 440, + 287, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 440, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 102, + 440, + 287, + 455 + ], + "type": "interline_equation", + "content": "D (I) = g ^ {- 1} \\left(f \\left(g (I, \\mathcal {B}), \\mathcal {A}\\right), \\mathcal {B}\\right). \\tag {4}", + "image_path": "7c4fd5bcc83de6838fb4c7bbc2c5cc9f4b3c326fcdef07c835a9adbb814b9c69.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 471, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 471, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 471, + 287, + 495 + ], + "type": "text", + "content": "Both " + }, + { + "bbox": [ + 47, + 471, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 47, + 471, + 287, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 471, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 47, + 471, + 287, + 495 + ], + "type": "text", + "content": " are estimated by a mapping estimator conditioned on the input image " + }, + { + "bbox": [ + 47, + 471, + 287, + 495 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 471, + 287, + 495 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 499, + 287, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 287, + 595 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 287, + 595 + ], + "type": "text", + "content": "To guarantee " + }, + { + "bbox": [ + 47, + 499, + 287, + 595 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 499, + 287, + 595 + ], + "type": "text", + "content": " represents a darkening process (i.e., " + }, + { + "bbox": [ + 47, + 499, + 287, + 595 + ], + "type": "inline_equation", + "content": "D(I) < I" + }, + { + "bbox": [ + 47, + 499, + 287, + 595 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 47, + 499, + 287, + 595 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 499, + 287, + 595 + ], + "type": "text", + "content": " should additionally satisfy convexity. Specifically, we let " + }, + { + "bbox": [ + 47, + 499, + 287, + 595 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 499, + 287, + 595 + ], + "type": "text", + "content": " be the iterative quadratic curve [16]: " + }, + { + "bbox": [ + 47, + 499, + 287, + 595 + ], + "type": "inline_equation", + "content": "f(x) = h^{(8)}(x), h(x,\\alpha) = \\alpha x^2 + (1 - \\alpha)x" + }, + { + "bbox": [ + 47, + 499, + 287, + 595 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 499, + 287, + 595 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 47, + 499, + 287, + 595 + ], + "type": "text", + "content": " be the dividing operation: " + }, + { + "bbox": [ + 47, + 499, + 287, + 595 + ], + "type": "inline_equation", + "content": "g(x,\\beta) = x / \\beta" + }, + { + "bbox": [ + 47, + 499, + 287, + 595 + ], + "type": "text", + "content": " in our implementation. Other kinds of curve forms are also considered and tested. Still, we empirically found that quadratic curves could bring slightly better results (results in Sec. 4.2)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 598, + 287, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 287, + 671 + ], + "type": "text", + "content": "Besides, to enable flexible control over the exposure level, we feed an exposure map " + }, + { + "bbox": [ + 47, + 598, + 287, + 671 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 47, + 598, + 287, + 671 + ], + "type": "text", + "content": " to the mapping estimator with " + }, + { + "bbox": [ + 47, + 598, + 287, + 671 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 598, + 287, + 671 + ], + "type": "text", + "content": ", yielding the corresponding darkened image " + }, + { + "bbox": [ + 47, + 598, + 287, + 671 + ], + "type": "inline_equation", + "content": "D(I,E)" + }, + { + "bbox": [ + 47, + 598, + 287, + 671 + ], + "type": "text", + "content": ". During training, the darkening module is encouraged to align the pixel value of " + }, + { + "bbox": [ + 47, + 598, + 287, + 671 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 47, + 598, + 287, + 671 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 598, + 287, + 671 + ], + "type": "inline_equation", + "content": "D(I,E)" + }, + { + "bbox": [ + 47, + 598, + 287, + 671 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 47, + 598, + 287, + 671 + ], + "type": "inline_equation", + "content": "D(I)" + }, + { + "bbox": [ + 47, + 598, + 287, + 671 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 598, + 287, + 671 + ], + "type": "inline_equation", + "content": "D(I,E)" + }, + { + "bbox": [ + 47, + 598, + 287, + 671 + ], + "type": "text", + "content": " interchangeably for simplicity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "Similarity Minimization. The training objective of module " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " involves two parts: similarity minimization and regularization. For the former, we directly reduce the distance" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 304, + 378, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 304, + 378, + 314 + ], + "spans": [ + { + "bbox": [ + 306, + 304, + 378, + 314 + ], + "type": "text", + "content": "between features:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 356, + 319, + 545, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 319, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 356, + 319, + 545, + 347 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D} ^ {\\text {s i m}} = \\frac {\\langle F (I) , F (D (I)) \\rangle}{\\left| \\left| F (I) \\right| \\right| _ {2} \\cdot \\left| \\left| F (D (I)) \\right| \\right| _ {2}}, \\tag {5}", + "image_path": "0f9031ad354956ae4e187450f132d7ef319daf94590bc0183b46fceaa309c8cc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 353, + 520, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 353, + 520, + 364 + ], + "spans": [ + { + "bbox": [ + 306, + 353, + 520, + 364 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 353, + 520, + 364 + ], + "type": "inline_equation", + "content": "\\langle \\cdot ,\\cdot \\rangle" + }, + { + "bbox": [ + 306, + 353, + 520, + 364 + ], + "type": "text", + "content": " is the inner product between two vectors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 365, + 545, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 545, + 400 + ], + "type": "text", + "content": "The regularization term consists of four losses. Besides a color consistency loss " + }, + { + "bbox": [ + 305, + 365, + 545, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{col}" + }, + { + "bbox": [ + 305, + 365, + 545, + 400 + ], + "type": "text", + "content": " [16] that corrects color deviations, three additional losses are proposed to regularize " + }, + { + "bbox": [ + 305, + 365, + 545, + 400 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 365, + 545, + 400 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 401, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 401, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 305, + 401, + 545, + 425 + ], + "type": "text", + "content": "Firstly, conditional exposure control is adopted to align the exposure map with the corresponding generated image:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 332, + 434, + 545, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 434, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 332, + 434, + 545, + 460 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c - e x p} = \\sum_ {1 \\leq i \\leq H, 1 \\leq j \\leq W} | \\hat {D} _ {i, j} (I, E) - E _ {i, j} |, \\tag {6}", + "image_path": "3052dec93eccddcf62902ace3ff80d77f55a8c6650d8fdb9cb4c7569ec2c9f76.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 465, + 544, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 465, + 544, + 502 + ], + "spans": [ + { + "bbox": [ + 305, + 465, + 544, + 502 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 465, + 544, + 502 + ], + "type": "inline_equation", + "content": "\\hat{D}(I,E)" + }, + { + "bbox": [ + 305, + 465, + 544, + 502 + ], + "type": "text", + "content": " is the channel-wise average of " + }, + { + "bbox": [ + 305, + 465, + 544, + 502 + ], + "type": "inline_equation", + "content": "D(I,E)" + }, + { + "bbox": [ + 305, + 465, + 544, + 502 + ], + "type": "text", + "content": ". During training, each exposure map " + }, + { + "bbox": [ + 305, + 465, + 544, + 502 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 305, + 465, + 544, + 502 + ], + "type": "text", + "content": " has identical entries uniformly sampled between " + }, + { + "bbox": [ + 305, + 465, + 544, + 502 + ], + "type": "inline_equation", + "content": "[0,0.5]" + }, + { + "bbox": [ + 305, + 465, + 544, + 502 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 502, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 545, + 561 + ], + "type": "text", + "content": "Then we add constraints on " + }, + { + "bbox": [ + 304, + 502, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 502, + 545, + 561 + ], + "type": "text", + "content": ". Intuitively, " + }, + { + "bbox": [ + 304, + 502, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 502, + 545, + 561 + ], + "type": "text", + "content": " represents the degree of illumination reduction. Illumination usually varies slowly across a scene but encounters rapid variations from object to object. Following this property, we apply a loose total variance loss:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 567, + 545, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 545, + 594 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l t v} (\\mathcal {A}) = \\sum_ {c \\in \\{R, G, B \\}} \\left(h \\left(\\left| \\nabla_ {x} \\mathcal {A} ^ {c} \\right|\\right) ^ {2} + h \\left(\\left| \\nabla_ {y} \\mathcal {A} ^ {c} \\right|\\right) ^ {2}\\right), \\tag {7}", + "image_path": "d769d9b44ea370e08f83ebd4c0f092751ed1ab659159932a1bff26a181d976ec.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 328, + 597, + 545, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 597, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 328, + 597, + 545, + 610 + ], + "type": "interline_equation", + "content": "h (x) = \\max (\\alpha - | x - \\alpha |, 0), \\tag {8}", + "image_path": "4f0df3e6e04c9a573a89b240fe4dc1d89bf330e8bfbd9040e0a40ca1e46e38d5.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\nabla_{x},\\nabla_{y}" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": " are gradient operations along the horizontal and vertical axis, respectively, and " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": " is a hyperparameter. Compared with the original total variance loss where " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": " is the identity function, our loose version allows the network to predict values of greater difference for adjacent pixels, which is common on objects' boundaries." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Finally, we adopt " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{flex}(\\mathcal{B}) = 1 - \\mathcal{B}" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " to avoid model fitting to the exposure solely by " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8107" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 59, + 72, + 217, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 72, + 217, + 84 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 217, + 84 + ], + "type": "text", + "content": "The overall training objective for " + }, + { + "bbox": [ + 59, + 72, + 217, + 84 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 59, + 72, + 217, + 84 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 286, + 106 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 92, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 92, + 286, + 106 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D} = \\lambda_ {D} ^ {\\text {s i m}} \\mathcal {L} _ {D} ^ {\\text {s i m}} + \\mathcal {R} _ {D}, \\tag {9}", + "image_path": "c0eb115ee2adbb4b760b4065c345b5c3b0c58af3036eff64ca800312d09543c0.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 109, + 288, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 109, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 48, + 109, + 288, + 133 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {D} = \\lambda_ {c - e x p} \\mathcal {L} _ {c - e x p} + \\lambda_ {c o l} \\mathcal {L} _ {c o l} + \\lambda_ {l t v} \\mathcal {L} _ {l t v} + \\lambda_ {f l e x} \\mathcal {L} _ {f l e x}. \\tag {10}", + "image_path": "d9b8c74a8a2bfc0beda4fb84e78f7c4b9803e9d94854117ed801b82710134396.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 142, + 246, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 142, + 246, + 155 + ], + "spans": [ + { + "bbox": [ + 47, + 142, + 246, + 155 + ], + "type": "text", + "content": "3.3. Model-Level Similarity Maximization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 160, + 287, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 160, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 47, + 160, + 287, + 195 + ], + "type": "text", + "content": "The darkening module " + }, + { + "bbox": [ + 47, + 160, + 287, + 195 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 160, + 287, + 195 + ], + "type": "text", + "content": " grants us access to a synthetic nighttime domain. In this section, we exploit " + }, + { + "bbox": [ + 47, + 160, + 287, + 195 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 160, + 287, + 195 + ], + "type": "text", + "content": " to learn illumination-robust representations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 196, + 287, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 196, + 287, + 292 + ], + "spans": [ + { + "bbox": [ + 47, + 196, + 287, + 292 + ], + "type": "text", + "content": "Contrastive learning [5, 17] is a self-supervised learning paradigm that contrasts positive and negative image pairs. However, images of the same class in classification or adjacent scenes in segmentation will form false negative pairs, thus hurting the model's performance. To alleviate these burdens, BYOL [15] proposes a non-negative variant that only aligns the feature between positive image pairs " + }, + { + "bbox": [ + 47, + 196, + 287, + 292 + ], + "type": "inline_equation", + "content": "\\{v, v^{+}\\}" + }, + { + "bbox": [ + 47, + 196, + 287, + 292 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 300, + 287, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 287, + 336 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {B Y O L}} (v, v ^ {+}) = 2 - \\frac {2 \\cdot \\langle z (q (F (v))) , q ^ {\\prime} \\left(F ^ {\\prime} (v ^ {+})\\right) \\rangle}{\\left| \\left| z (q (F (v))) \\right| \\right| _ {2} \\cdot \\left| \\left| q ^ {\\prime} \\left(F ^ {\\prime} (v ^ {+})\\right) \\right| \\right| _ {2}}, \\tag {11}", + "image_path": "46e07ba2700abcd6ed18f2eea86b04b765232fae1c077b8689c68fc9c1765b24.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "inline_equation", + "content": "q, q'" + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "text", + "content": " are projection heads, and " + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "text", + "content": " is the prediction head. Both of them are MLPs with a single hidden layer. Note that " + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "inline_equation", + "content": "F'" + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "inline_equation", + "content": "q'" + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "text", + "content": " share the same architecture and weight initialization with " + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 47, + 337, + 287, + 396 + ], + "type": "text", + "content": " but receive no gradient and are updated by exponential moving average (EMA)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 399, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 287, + 459 + ], + "type": "text", + "content": "Similarity Maximization. Motivated by BYOL, we maximize the feature similarity between synthetic nighttime and daytime domains by non-negative contrastive learning. Given a daytime image " + }, + { + "bbox": [ + 47, + 399, + 287, + 459 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 399, + 287, + 459 + ], + "type": "text", + "content": " and an exposure map " + }, + { + "bbox": [ + 47, + 399, + 287, + 459 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 47, + 399, + 287, + 459 + ], + "type": "text", + "content": ", we formulate the training objective as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 468, + 287, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 287, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {F} ^ {\\text {s i m}} = \\mathcal {L} _ {\\mathrm {B Y O L}} (I, D (I, E)) + \\mathcal {L} _ {\\mathrm {B Y O L}} (D (I, E), I). \\tag {12}", + "image_path": "65b4d292065ad2c44b96b108af88ca7675d55117b1aadcb49be0626a6858e7d3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 491, + 287, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 287, + 573 + ], + "type": "text", + "content": "Note that the measure of feature similarity is different between Eq. (5) and Eq. (12). Directly applying Eq. (5) to train " + }, + { + "bbox": [ + 47, + 491, + 287, + 573 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 491, + 287, + 573 + ], + "type": "text", + "content": " brings poorer results due to potential feature degeneration. In comparison, the asymmetric projection head and stop gradient policies prevent the feature extractor " + }, + { + "bbox": [ + 47, + 491, + 287, + 573 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 491, + 287, + 573 + ], + "type": "text", + "content": " from collapsing, i.e., working as the regularization " + }, + { + "bbox": [ + 47, + 491, + 287, + 573 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_F" + }, + { + "bbox": [ + 47, + 491, + 287, + 573 + ], + "type": "text", + "content": " in Eq. (2) together with the task loss (introduced below)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 574, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 287, + 670 + ], + "type": "text", + "content": "Moreover, different from " + }, + { + "bbox": [ + 47, + 574, + 287, + 670 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 47, + 574, + 287, + 670 + ], + "type": "text", + "content": " in Eq. (6), we use a compound exposure map " + }, + { + "bbox": [ + 47, + 574, + 287, + 670 + ], + "type": "inline_equation", + "content": "E'" + }, + { + "bbox": [ + 47, + 574, + 287, + 670 + ], + "type": "text", + "content": " instead. " + }, + { + "bbox": [ + 47, + 574, + 287, + 670 + ], + "type": "inline_equation", + "content": "E'" + }, + { + "bbox": [ + 47, + 574, + 287, + 670 + ], + "type": "text", + "content": " is first initialized with identical entries uniformly sampled between [0, 0.2] for simulating nighttime illumination. This range is the same for all downstream tasks, which does not introduce task-relevant prior. Then, we add pixel-wise noise " + }, + { + "bbox": [ + 47, + 574, + 287, + 670 + ], + "type": "inline_equation", + "content": "z_1" + }, + { + "bbox": [ + 47, + 574, + 287, + 670 + ], + "type": "text", + "content": " and patch-wise noise " + }, + { + "bbox": [ + 47, + 574, + 287, + 670 + ], + "type": "inline_equation", + "content": "z_2" + }, + { + "bbox": [ + 47, + 574, + 287, + 670 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 574, + 287, + 670 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 47, + 574, + 287, + 670 + ], + "type": "text", + "content": " to simulate exposure discrepancy. Overall, " + }, + { + "bbox": [ + 47, + 574, + 287, + 670 + ], + "type": "inline_equation", + "content": "E'" + }, + { + "bbox": [ + 47, + 574, + 287, + 670 + ], + "type": "text", + "content": " can be represented as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 112, + 679, + 287, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 679, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 112, + 679, + 287, + 693 + ], + "type": "interline_equation", + "content": "E ^ {\\prime} = \\mathcal {U} (0, 0. 2) + z _ {1} + z _ {2}. \\tag {13}", + "image_path": "d558899aa38e8abc98362782c7ac222ad5f828b9e418bd7ad6fd4ff2d50730cf.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 701, + 259, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 259, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 259, + 713 + ], + "type": "text", + "content": "See the supplementary for details on noise injection." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 542, + 197 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 542, + 197 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 542, + 197 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 542, + 197 + ], + "type": "image", + "image_path": "76228bff88e809d282988be7f1c83351000e58b3b658172c18577f02be06583d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 209, + 545, + 285 + ], + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 285 + ], + "type": "text", + "content": "Figure 4. t-SNE [50] visualization of images' feature extracted by the original daytime model and our adapted model on CO-DaN [29]. Red, green, and blue dots represent the feature of daytime, synthesized nighttime, and real nighttime images, respectively. We only color the instances from the \"Car\" category for better visual quality. Additional visualization results are shown in the supplementary." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 304, + 545, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 545, + 340 + ], + "type": "text", + "content": "Besides " + }, + { + "bbox": [ + 305, + 304, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_F^{sim}" + }, + { + "bbox": [ + 305, + 304, + 545, + 340 + ], + "type": "text", + "content": ", we add task-specific supervision " + }, + { + "bbox": [ + 305, + 304, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{task}" + }, + { + "bbox": [ + 305, + 304, + 545, + 340 + ], + "type": "text", + "content": " on both the original daytime and synthetic nighttime domain. The final training objective for " + }, + { + "bbox": [ + 305, + 304, + 545, + 340 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 304, + 545, + 340 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 360, + 346, + 545, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 346, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 360, + 346, + 545, + 360 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {F} = \\lambda_ {F} ^ {\\text {s i m}} \\mathcal {L} _ {F} ^ {\\text {s i m}} + \\lambda_ {\\text {t a s k}} \\mathcal {L} _ {\\text {t a s k}}. \\tag {14}", + "image_path": "d75ed5c84def83880e3175fc9786287dee3ba5ecda7f423e40d7d70db480ffc9.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 372, + 448, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 372, + 448, + 385 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 448, + 385 + ], + "type": "text", + "content": "3.4. Overall Training Pipeline" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 391, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 391, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 391, + 545, + 437 + ], + "type": "text", + "content": "Having introduced the image-level similarity minimization (Section 3.2) and model-level similarity maximization (Section 3.3), this section discusses the overall pipeline, as shown in Figure 3." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "text", + "content": "An intuitive idea is training " + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "text", + "content": " alternately like GAN [13, 64]. Nevertheless, balancing " + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "text", + "content": " increases the difficulty of parameter tuning and makes the optimization process unstable. We adopt a simple but effective two-step strategy to solve this problem: we first train " + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "text", + "content": " and keep " + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "text", + "content": " frozen, then train " + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "text", + "content": " and keep " + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "text", + "content": " frozen. Compared with the alternate strategy, our step-wise approach improves the performance on nighttime image classification (elaborate in Section 4.2) from " + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "inline_equation", + "content": "63.84\\%" + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "inline_equation", + "content": "65.87\\%" + }, + { + "bbox": [ + 305, + 438, + 545, + 545 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 546, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 640 + ], + "type": "text", + "content": "We could also explain the merits of our min-max framework from the perspective of adversarial training [14, 35]. Module " + }, + { + "bbox": [ + 304, + 546, + 545, + 640 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 546, + 545, + 640 + ], + "type": "text", + "content": " first produces the worst-case examples regarding feature similarity. Then, our model could learn the illumination-robust features by learning on these cases through similarity maximization. This technical commonality further justifies our motivation to build the similarity min-max framework." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Across all downstream tasks, the feature extractor and task module are initialized by daytime pre-trained models. We first freeze the feature extractor and train the darkening module (image-level translation). Then, we keep the darkening module fixed and train the feature extractor and task module jointly (model-level adaptation)." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8108" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 85 + ], + "type": "text", + "content": "3.5. Empirical Justifications on Darkening Module" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 281 + ], + "type": "text", + "content": "Simulating nighttime conditions without accessing real nighttime images is the key to our framework. Particularly, nighttime conditions bring semantic changes in addition to illumination changes, e.g., the dark environment with artificial lights on the second real nighttime image in Figure 4. However, an accurate simulation is extremely difficult since our prior knowledge is limited to \"low illumination\". Fortunately, unlike typical day-to-night image synthesis processes [41] which target the human visual experience, ours only care about the distribution of darkened images in the feature space. Leaving aside visual quality, we are pleased to find that the feature distribution of our synthesized nighttime domain is similar to that of the real nighttime domain as visualized in Figure 4(a). This observation demonstrates that our darkening process can characterize the night domain from the model-level perspective." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 282, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 390 + ], + "type": "text", + "content": "Thanks to this property, the feature discrepancy between daytime and real nighttime domain is significantly reduced after model-level adaptation (red and blue dots in Figure 4). This discovery is consistent with the Maximum Mean Discrepancy (MMD) between the feature distribution of day and night modalities, which is 0.020 and 0.014 for the original and adapted models, respectively. We provide implementation details and additional empirical analysis using saliency maps in the supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 401, + 128, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 401, + 128, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 401, + 128, + 414 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 421, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 421, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 421, + 287, + 445 + ], + "type": "text", + "content": "This section provides the implementation details, benchmarking results, and ablation analysis of our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 453, + 180, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 180, + 465 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 180, + 465 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 472, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 651 + ], + "type": "text", + "content": "Our framework widely applies to various nighttime vision tasks. In the following, we evaluate our method with four representative tasks: image classification, semantic segmentation, visual place recognition, and video action recognition. Only daytime data are accessible for training and validation, while nighttime data are only used during evaluation. We benchmark our method with three categories of methods that require no dataset-specific target domain data: low-light enhancement, zero-shot day-night domain adaptation, and domain generalization. For low-light enhancement, enhancement models are trained on their original datasets. Then we adopt them as a pre-processing step to assist the daytime baseline. The results of our method is the average of three independent trails. Additional details are provided in the supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 658, + 214, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 214, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 214, + 671 + ], + "type": "text", + "content": "4.2. Nighttime Image Classification" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "We first consider one of the most fundamental vision tasks: image classification. CODaN [29] is a 10-class dataset containing a training set of 10000 daytime images" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 313, + 113, + 537, + 342 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "type": "text", + "content": "Table 1. Top-1 classification accuracy on the CODaN nighttime test set [29]. " + }, + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "type": "text", + "content": " denotes our re-implementation with both the original and synthesized image fed into the task module." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 113, + 537, + 342 + ], + "lines": [ + { + "bbox": [ + 313, + 113, + 537, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 113, + 537, + 342 + ], + "type": "table", + "html": "
MethodTop-1 (%)
ResNet-18 [18]53.32
Low-Light Enhancement
EnlightenGAN [23]56.68
LEDNet [63]57.40
Zero-DCE++ [30]57.96
RUAS [33]58.36
SCI [34]58.68
URetinexNet [56]58.72
Domain Generalization
MixStyle [62]53.12
IRM [1]54.52
AdaBN [31]54.25
Zero-Shot Day-Night Domain Adaptation
MAET† [8]56.48
CICConv [29]60.32
Ours65.87
", + "image_path": "9c213b5f9d16c22023119f0557a95e9763769afaad65594fd2363a27de2063c8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 365, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 545, + 413 + ], + "type": "text", + "content": "and a test set with 2500 daytime and nighttime images, respectively. We validate models on the daytime test set and evaluate them on the nighttime test set. The backbone is ResNet-18 [18]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 414, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 581 + ], + "type": "text", + "content": "Benchmarking results are shown in Table 1. Enhancement methods restore input low-light images from the human visual perspective while keeping the model untouched, resulting in limited performance gains. Domain generalization methods are designed for general tasks and perform poorly in unseen nighttime environments. MAET [8] relies on degrading transformation with sensor-specific parameters, which suffers from poor generalizability. CICConv [29] adopts learnable color invariant edge detectors, which are not robust to the complex illumination variation in real scenarios. In contrast, our method outperforms state-of-the-art methods by a large margin (60.32% v.s. 65.87%), demonstrating our unified framework could obtain features more robust to illumination shifts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 545, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 690 + ], + "type": "text", + "content": "Ablation Studies. We conduct ablation studies to justify our framework design in Table 2. Firstly, we study how to design the darkening module " + }, + { + "bbox": [ + 304, + 582, + 545, + 690 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 582, + 545, + 690 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 304, + 582, + 545, + 690 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D^{sim}" + }, + { + "bbox": [ + 304, + 582, + 545, + 690 + ], + "type": "text", + "content": ". The model-level adaptation stage (Section 3.3) remains the same for fair comparisons. Firstly, we replace our darkening module with heuristic image adjustment approaches, such as brightness adjustment (Brightness in PIL" + }, + { + "bbox": [ + 304, + 582, + 545, + 690 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 304, + 582, + 545, + 690 + ], + "type": "text", + "content": ") and gamma correction (" + }, + { + "bbox": [ + 304, + 582, + 545, + 690 + ], + "type": "inline_equation", + "content": "D(I) = I^{\\gamma}" + }, + { + "bbox": [ + 304, + 582, + 545, + 690 + ], + "type": "text", + "content": "). We implement these two approaches using a fixed darkening hyperparameter chosen after multi-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 542, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 542, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 542, + 713 + ], + "type": "text", + "content": "1https://pillow.readthedocs.io/en/stable/reference/ImageEnhance.html" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8109" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 102, + 266, + 245 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "text", + "content": "Table 2. Ablation studies for module " + }, + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "text", + "content": " and similarity losses. We report the Top-1 accuracy on the CODaN [29] nighttime test set." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 102, + 266, + 245 + ], + "lines": [ + { + "bbox": [ + 66, + 102, + 266, + 245 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 266, + 245 + ], + "type": "table", + "html": "
CategoryMethodTop-1 (%)
BaselineVanilla ResNet-1853.32
Module DBrightness adjustment57.96
HeuristicGamma correction63.96
Module DReciprocal curve62.60
LearnableGamma curve64.16
Similarityw/o LDSim and LFSim64.08
Lossw/o LDSim64.56
w/o LFSim64.88
Full version-65.87
", + "image_path": "d89df1c6dc0bd829142bdb9e31c5493f58d5452c9c734c15b09ba9cf58eb29d0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 266, + 188, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 188, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 188, + 277 + ], + "type": "text", + "content": "ple trials and report the best score." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 277, + 287, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 351 + ], + "type": "text", + "content": "Next, we test other possible curve forms for " + }, + { + "bbox": [ + 46, + 277, + 287, + 351 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 277, + 287, + 351 + ], + "type": "text", + "content": ". Both gamma curve " + }, + { + "bbox": [ + 46, + 277, + 287, + 351 + ], + "type": "inline_equation", + "content": "(f(x,\\alpha) = x^{\\frac{1}{\\alpha}}, \\alpha \\in (0,1])" + }, + { + "bbox": [ + 46, + 277, + 287, + 351 + ], + "type": "text", + "content": " and reciprocal curve " + }, + { + "bbox": [ + 46, + 277, + 287, + 351 + ], + "type": "inline_equation", + "content": "(f(x,\\alpha) = \\frac{(1 - \\alpha)\\cdot x}{1 - \\alpha\\cdot x}, \\alpha \\in [0,1))" + }, + { + "bbox": [ + 46, + 277, + 287, + 351 + ], + "type": "text", + "content": " bring slightly worse results than the iterative quadratic curve. Please refer to the supplementary for implementation details of these ablations and additional results on the segmentation task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 351, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 287, + 399 + ], + "type": "text", + "content": "Finally, we test our framework's performance when one or both of the similarity loss is absent. We find that either similarity loss alone can boost the model's nighttime performance while combining them achieves the best result." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 407, + 230, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 230, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 230, + 420 + ], + "type": "text", + "content": "4.3. Nighttime Semantic Segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "type": "text", + "content": "Next, we explore a more challenging nighttime vision task: semantic segmentation. We adopt RefineNet [32] with ResNet-101 backbone as the baseline. The daytime training dataset is Cityscapes [7], containing 2975 images for training and 500 images for validation, all with dense annotations. The nighttime testing datasets are Nighttime Driving [9] and Dark-Zurich [46]. These two datasets contain 50 coarsely annotated and 151 densely annotated nighttime street view images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 533, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 689 + ], + "type": "text", + "content": "We benchmark our method in Table 3. Low-light enhancement methods yield worse results than the baseline because they perform poorly on street scenes with complex light sources. Domain generalization methods fail to mitigate the huge day-night domain gap, leading to unsatisfactory results. Note that RobustNet [6] adopts DeepLabv3 [4] architecture, which is superior to RefineNet [32] adopted in our implementation. Among zero-shot adaptation methods, MAET [8] injects too much noise into images, leading to severe performance degradation. CICConv yields better results, but the improvement is limited. In comparison, our approach improves the mIoU scores to " + }, + { + "bbox": [ + 46, + 533, + 287, + 689 + ], + "type": "inline_equation", + "content": "44.9\\%" + }, + { + "bbox": [ + 46, + 533, + 287, + 689 + ], + "type": "text", + "content": " on Nighttime Driving and " + }, + { + "bbox": [ + 46, + 533, + 287, + 689 + ], + "type": "inline_equation", + "content": "40.2\\%" + }, + { + "bbox": [ + 46, + 533, + 287, + 689 + ], + "type": "text", + "content": " on Dark-Zurich." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Figure 5 shows qualitative segmentation results on two nighttime datasets. Low-light enhancement methods per" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 310, + 102, + 541, + 332 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 3. Semantic segmentation results on Nighttime Driving [9] and Dark-Zurich [46], reported as percentage mIoU scores." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 102, + 541, + 332 + ], + "lines": [ + { + "bbox": [ + 310, + 102, + 541, + 332 + ], + "spans": [ + { + "bbox": [ + 310, + 102, + 541, + 332 + ], + "type": "table", + "html": "
MethodNighttime DrivingDark-Zurich
RefineNet [32]34.330.6
Low-Light Enhancement
EnlightenGAN [23]25.224.9
Zero-DCE++ [30]32.728.3
RUAS [33]25.123.4
SCI [34]28.625.7
URetinexNet [56]28.124.0
LEDNet [63]27.626.6
Domain Generalization
AdaBN [31]37.231.1
RobustNet [6]33.034.5
SAN-SAW [38]28.116.0
Zero-Shot Day-Night Domain Adaptation
MAET [8]28.126.4
CICConv [29]41.234.5
Ours44.940.2
", + "image_path": "9733bec9aa9493474b9f24fd7371cf869586d9132f06465a538a3a8f783f5e2f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 362, + 535, + 522 + ], + "blocks": [ + { + "bbox": [ + 315, + 342, + 536, + 354 + ], + "lines": [ + { + "bbox": [ + 315, + 342, + 536, + 354 + ], + "spans": [ + { + "bbox": [ + 315, + 342, + 536, + 354 + ], + "type": "text", + "content": "Table 4. Visual place recognition results on Tokyo 24/7 [49]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 362, + 535, + 522 + ], + "lines": [ + { + "bbox": [ + 315, + 362, + 535, + 522 + ], + "spans": [ + { + "bbox": [ + 315, + 362, + 535, + 522 + ], + "type": "table", + "html": "
MethodmAP (%)
Zero-Shot Day-Night Domain Adaptation
EdgeMAC [42]75.9
U-Net jointly [21]79.8
GeM [43]85.0
CICConv-GeM [29]88.3
Ours-GeM90.4
Day-Night Domain Adaptation \n(night images are available for training)
U-Net jointly [21]86.5
EdgeMAC + CLAHE [21]90.5
EdgeMAC + U-Net jointly [21]90.0
", + "image_path": "284ea32fd983ea25bd305b73798493689fd0c82a121f09211d99bae5d6126fe1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 543, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 543, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 305, + 543, + 545, + 579 + ], + "type": "text", + "content": "form poorly on nighttime street scenes. Our method better extracts information hidden by darkness and thus generates more accurate semantic maps." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 586, + 485, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 485, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 485, + 600 + ], + "type": "text", + "content": "4.4. Visual Place Recognition at Night" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Then we explore visual place recognition, which aims to retrieve images that illustrate the same scene of a query image from an image pool. Unlike classification and segmentation, place recognition methods are not end-to-end during inference. We extend our method based on GeM [43] with ResNet-101 backbone. In GeM, the network receives a tuple of images " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{p,q,n_1,\\dots ,n_k\\}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " as input, in which the query " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " only matches " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". The network is trained on a contrastive loss, similar to the model-level stage in our" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8110" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 130, + 112 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 130, + 112 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 130, + 112 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 130, + 112 + ], + "type": "image", + "image_path": "0c34cccb7bbf33208d091438771560d3643b627f27d819ee8b04e018160bbd63.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 113, + 129, + 154 + ], + "blocks": [ + { + "bbox": [ + 49, + 113, + 129, + 154 + ], + "lines": [ + { + "bbox": [ + 49, + 113, + 129, + 154 + ], + "spans": [ + { + "bbox": [ + 49, + 113, + 129, + 154 + ], + "type": "image", + "image_path": "1283e0a1259f724edc7c69db5ab16e4b4e33f511901c6088fcc7a7eec748a2ed.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 156, + 105, + 165 + ], + "lines": [ + { + "bbox": [ + 74, + 156, + 105, + 165 + ], + "spans": [ + { + "bbox": [ + 74, + 156, + 105, + 165 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 132, + 71, + 213, + 112 + ], + "blocks": [ + { + "bbox": [ + 132, + 71, + 213, + 112 + ], + "lines": [ + { + "bbox": [ + 132, + 71, + 213, + 112 + ], + "spans": [ + { + "bbox": [ + 132, + 71, + 213, + 112 + ], + "type": "image", + "image_path": "f242153bf0ca9531624ffbfca6435a8bc251e2481f8f43c884ba17da0cb273b0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 132, + 113, + 212, + 154 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 212, + 154 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 212, + 154 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 212, + 154 + ], + "type": "image", + "image_path": "4c7c62bf14bd881f5fd5aba9331243f0b27210fafadc7392ad4bd9c6bf7546ba.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 156, + 203, + 165 + ], + "lines": [ + { + "bbox": [ + 144, + 156, + 203, + 165 + ], + "spans": [ + { + "bbox": [ + 144, + 156, + 203, + 165 + ], + "type": "text", + "content": "(b) Ground truth" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 215, + 71, + 295, + 112 + ], + "blocks": [ + { + "bbox": [ + 215, + 71, + 295, + 112 + ], + "lines": [ + { + "bbox": [ + 215, + 71, + 295, + 112 + ], + "spans": [ + { + "bbox": [ + 215, + 71, + 295, + 112 + ], + "type": "image", + "image_path": "4fa47fd874dacefa77e2c26c48d5ccb82c010d853c2f9fdc7a1e1b14be2f6231.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 215, + 113, + 296, + 154 + ], + "blocks": [ + { + "bbox": [ + 215, + 113, + 296, + 154 + ], + "lines": [ + { + "bbox": [ + 215, + 113, + 296, + 154 + ], + "spans": [ + { + "bbox": [ + 215, + 113, + 296, + 154 + ], + "type": "image", + "image_path": "ce58d14809e4597c2ab1f07668d28c502f7245e0b1118fdd70ccbac0a213c414.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 156, + 282, + 165 + ], + "lines": [ + { + "bbox": [ + 231, + 156, + 282, + 165 + ], + "spans": [ + { + "bbox": [ + 231, + 156, + 282, + 165 + ], + "type": "text", + "content": "(c) RefineNet" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 298, + 71, + 378, + 112 + ], + "blocks": [ + { + "bbox": [ + 298, + 71, + 378, + 112 + ], + "lines": [ + { + "bbox": [ + 298, + 71, + 378, + 112 + ], + "spans": [ + { + "bbox": [ + 298, + 71, + 378, + 112 + ], + "type": "image", + "image_path": "dff8421f58b7d6ff244687f790104a06e1657ff977af59e6470b0005e719c810.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 298, + 113, + 378, + 154 + ], + "blocks": [ + { + "bbox": [ + 298, + 113, + 378, + 154 + ], + "lines": [ + { + "bbox": [ + 298, + 113, + 378, + 154 + ], + "spans": [ + { + "bbox": [ + 298, + 113, + 378, + 154 + ], + "type": "image", + "image_path": "26a89c7c67f4577da9f820c5edd3e2a1543a6f883fafb60adcf1fdf245f61615.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 156, + 358, + 165 + ], + "lines": [ + { + "bbox": [ + 320, + 156, + 358, + 165 + ], + "spans": [ + { + "bbox": [ + 320, + 156, + 358, + 165 + ], + "type": "text", + "content": "(d) MAET" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 380, + 71, + 462, + 112 + ], + "blocks": [ + { + "bbox": [ + 380, + 71, + 462, + 112 + ], + "lines": [ + { + "bbox": [ + 380, + 71, + 462, + 112 + ], + "spans": [ + { + "bbox": [ + 380, + 71, + 462, + 112 + ], + "type": "image", + "image_path": "0c15b0e42873f543d7a7b3edfaf98e404e2419f349ebc8d41f7cdaeb38856595.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 380, + 113, + 462, + 154 + ], + "blocks": [ + { + "bbox": [ + 380, + 113, + 462, + 154 + ], + "lines": [ + { + "bbox": [ + 380, + 113, + 462, + 154 + ], + "spans": [ + { + "bbox": [ + 380, + 113, + 462, + 154 + ], + "type": "image", + "image_path": "f185793ec33a9ad215eb49cb9d824ee6b6ee4118aef5483023e8c60f98f62a7e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 156, + 442, + 165 + ], + "lines": [ + { + "bbox": [ + 402, + 156, + 442, + 165 + ], + "spans": [ + { + "bbox": [ + 402, + 156, + 442, + 165 + ], + "type": "text", + "content": "(e) CICConv" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 463, + 71, + 545, + 112 + ], + "blocks": [ + { + "bbox": [ + 463, + 71, + 545, + 112 + ], + "lines": [ + { + "bbox": [ + 463, + 71, + 545, + 112 + ], + "spans": [ + { + "bbox": [ + 463, + 71, + 545, + 112 + ], + "type": "image", + "image_path": "491b913ceaef18be9aa5a615cca5aa92bda6f20f1f235161785d0a4f3d5f5c29.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 463, + 113, + 545, + 154 + ], + "blocks": [ + { + "bbox": [ + 463, + 113, + 545, + 154 + ], + "lines": [ + { + "bbox": [ + 463, + 113, + 545, + 154 + ], + "spans": [ + { + "bbox": [ + 463, + 113, + 545, + 154 + ], + "type": "image", + "image_path": "77a65d83066df2462a180e2af4dbfdae0fa824928f6d29cffefcd1f47074fc09.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 156, + 520, + 165 + ], + "lines": [ + { + "bbox": [ + 490, + 156, + 520, + 165 + ], + "spans": [ + { + "bbox": [ + 490, + 156, + 520, + 165 + ], + "type": "text", + "content": "(f) Ours" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 49, + 167, + 130, + 209 + ], + "blocks": [ + { + "bbox": [ + 49, + 167, + 130, + 209 + ], + "lines": [ + { + "bbox": [ + 49, + 167, + 130, + 209 + ], + "spans": [ + { + "bbox": [ + 49, + 167, + 130, + 209 + ], + "type": "image", + "image_path": "287f8adf80db365e86267b91809eae6f51dd06d2588839fd7be381d2240e0af3.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 49, + 209, + 130, + 251 + ], + "blocks": [ + { + "bbox": [ + 49, + 209, + 130, + 251 + ], + "lines": [ + { + "bbox": [ + 49, + 209, + 130, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 130, + 251 + ], + "type": "image", + "image_path": "158eb11862588069f2d9b43bcf528ef41b7d33a47fbca340dddf47d35ec725a2.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 255, + 137, + 265 + ], + "lines": [ + { + "bbox": [ + 110, + 255, + 137, + 265 + ], + "spans": [ + { + "bbox": [ + 110, + 255, + 137, + 265 + ], + "type": "text", + "content": "(g)SCI" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 132, + 167, + 213, + 209 + ], + "blocks": [ + { + "bbox": [ + 132, + 167, + 213, + 209 + ], + "lines": [ + { + "bbox": [ + 132, + 167, + 213, + 209 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 213, + 209 + ], + "type": "image", + "image_path": "ccf612991c94ec0cbdadc4098eb518c82ea6697c9b23dab6557df3339139d9fa.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 132, + 209, + 213, + 251 + ], + "blocks": [ + { + "bbox": [ + 132, + 209, + 213, + 251 + ], + "lines": [ + { + "bbox": [ + 132, + 209, + 213, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 213, + 251 + ], + "type": "image", + "image_path": "190baffd73ccf3e85e47ed02edbeb1438af65faa96aaa5c23771e9c02866bbe0.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 277, + 529, + 288 + ], + "lines": [ + { + "bbox": [ + 63, + 277, + 529, + 288 + ], + "spans": [ + { + "bbox": [ + 63, + 277, + 529, + 288 + ], + "type": "text", + "content": "Figure 5. Semantic segmentation results. For each group, the first row: Nighttime Driving [9], the second row: Dark-Zurich [46]." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 215, + 167, + 296, + 209 + ], + "blocks": [ + { + "bbox": [ + 215, + 167, + 296, + 209 + ], + "lines": [ + { + "bbox": [ + 215, + 167, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 215, + 167, + 296, + 209 + ], + "type": "image", + "image_path": "82498635fb2d1d7504b70b1d1cc5ab9ae101cd69bbb50b18f76133f5e2595f66.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 215, + 209, + 296, + 251 + ], + "blocks": [ + { + "bbox": [ + 215, + 209, + 296, + 251 + ], + "lines": [ + { + "bbox": [ + 215, + 209, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 215, + 209, + 296, + 251 + ], + "type": "image", + "image_path": "9aead569bf4da46517ef802e185521cf3d353906ec99176561ae4f98473dd61a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 255, + 320, + 265 + ], + "lines": [ + { + "bbox": [ + 260, + 255, + 320, + 265 + ], + "spans": [ + { + "bbox": [ + 260, + 255, + 320, + 265 + ], + "type": "text", + "content": "(h) URetinexNet" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 298, + 167, + 379, + 209 + ], + "blocks": [ + { + "bbox": [ + 298, + 167, + 379, + 209 + ], + "lines": [ + { + "bbox": [ + 298, + 167, + 379, + 209 + ], + "spans": [ + { + "bbox": [ + 298, + 167, + 379, + 209 + ], + "type": "image", + "image_path": "fe196eb1a464f355d116bbd80c08957b2c772bf93c2262c85b1a6567e5b2fb71.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 298, + 209, + 379, + 251 + ], + "blocks": [ + { + "bbox": [ + 298, + 209, + 379, + 251 + ], + "lines": [ + { + "bbox": [ + 298, + 209, + 379, + 251 + ], + "spans": [ + { + "bbox": [ + 298, + 209, + 379, + 251 + ], + "type": "image", + "image_path": "c7212c51eb4eed20fee3b2ce313c032b30cc72540e75624827306a19c414c436.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 380, + 167, + 462, + 209 + ], + "blocks": [ + { + "bbox": [ + 380, + 167, + 462, + 209 + ], + "lines": [ + { + "bbox": [ + 380, + 167, + 462, + 209 + ], + "spans": [ + { + "bbox": [ + 380, + 167, + 462, + 209 + ], + "type": "image", + "image_path": "32c5e15404eecadc62dcfd4dde9393d3c1c1d02f5f6460147c3692220ba97b6a.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 380, + 209, + 462, + 251 + ], + "blocks": [ + { + "bbox": [ + 380, + 209, + 462, + 251 + ], + "lines": [ + { + "bbox": [ + 380, + 209, + 462, + 251 + ], + "spans": [ + { + "bbox": [ + 380, + 209, + 462, + 251 + ], + "type": "image", + "image_path": "e39f936847e30e759b6bafc371ca06ddb54f02d4245b005f1a01f4119d4fe61b.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 255, + 484, + 265 + ], + "lines": [ + { + "bbox": [ + 425, + 255, + 484, + 265 + ], + "spans": [ + { + "bbox": [ + 425, + 255, + 484, + 265 + ], + "type": "text", + "content": "(i) Zero-DCE++" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 463, + 167, + 545, + 209 + ], + "blocks": [ + { + "bbox": [ + 463, + 167, + 545, + 209 + ], + "lines": [ + { + "bbox": [ + 463, + 167, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 463, + 167, + 545, + 209 + ], + "type": "image", + "image_path": "194e97d964d681aacadcd6ff2819c3abb38bc532c01afcb03471194739c24e28.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 463, + 209, + 545, + 251 + ], + "blocks": [ + { + "bbox": [ + 463, + 209, + 545, + 251 + ], + "lines": [ + { + "bbox": [ + 463, + 209, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 463, + 209, + 545, + 251 + ], + "type": "image", + "image_path": "9d2705ec1807b155c053d78b81cc886453a28a63a71f155c1fc35443a50c9458.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 54, + 307, + 127, + 392 + ], + "blocks": [ + { + "bbox": [ + 54, + 307, + 127, + 392 + ], + "lines": [ + { + "bbox": [ + 54, + 307, + 127, + 392 + ], + "spans": [ + { + "bbox": [ + 54, + 307, + 127, + 392 + ], + "type": "image", + "image_path": "1602b536bd4468fed7c3749932595e07db0103c673ac3b4090e2707c8013eab4.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 393, + 110, + 403 + ], + "lines": [ + { + "bbox": [ + 75, + 393, + 110, + 403 + ], + "spans": [ + { + "bbox": [ + 75, + 393, + 110, + 403 + ], + "type": "text", + "content": "(a) Query" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 418, + 287, + 451 + ], + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 451 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 451 + ], + "type": "text", + "content": "Figure 6. Qualitative visual place recognition results. (a) A night query from the Tokyo 24/7 dataset [49]. (b) Image retrieved by GeM [43]. (c) Image retrieved by our method." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 131, + 307, + 204, + 392 + ], + "blocks": [ + { + "bbox": [ + 131, + 307, + 204, + 392 + ], + "lines": [ + { + "bbox": [ + 131, + 307, + 204, + 392 + ], + "spans": [ + { + "bbox": [ + 131, + 307, + 204, + 392 + ], + "type": "image", + "image_path": "f56a9621e95f59a95a6b0fbbc099f2a39a1a81108f5969bbb6e3fec5dc6d7379.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 152, + 393, + 184, + 403 + ], + "lines": [ + { + "bbox": [ + 152, + 393, + 184, + 403 + ], + "spans": [ + { + "bbox": [ + 152, + 393, + 184, + 403 + ], + "type": "text", + "content": "(b) GeM" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 209, + 307, + 280, + 392 + ], + "blocks": [ + { + "bbox": [ + 209, + 307, + 280, + 392 + ], + "lines": [ + { + "bbox": [ + 209, + 307, + 280, + 392 + ], + "spans": [ + { + "bbox": [ + 209, + 307, + 280, + 392 + ], + "type": "image", + "image_path": "53dd682c1230ea9c589870472abbe52ee42eae5cb3facd77118c3e9217d31cb9.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 393, + 260, + 403 + ], + "lines": [ + { + "bbox": [ + 229, + 393, + 260, + 403 + ], + "spans": [ + { + "bbox": [ + 229, + 393, + 260, + 403 + ], + "type": "text", + "content": "(c) Ours" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "bbox": [ + 46, + 472, + 286, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 286, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 286, + 579 + ], + "type": "text", + "content": "framework. We retain the image-level stage and modify the model-level stage in our implementation. We first train the darkening module " + }, + { + "bbox": [ + 46, + 472, + 286, + 579 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 472, + 286, + 579 + ], + "type": "text", + "content": " as usual. Then, we consider " + }, + { + "bbox": [ + 46, + 472, + 286, + 579 + ], + "type": "inline_equation", + "content": "D(p)" + }, + { + "bbox": [ + 46, + 472, + 286, + 579 + ], + "type": "text", + "content": " as an additional matching for " + }, + { + "bbox": [ + 46, + 472, + 286, + 579 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 472, + 286, + 579 + ], + "type": "text", + "content": ", i.e., an input tuple contains two positive samples (instead of one) and " + }, + { + "bbox": [ + 46, + 472, + 286, + 579 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 472, + 286, + 579 + ], + "type": "text", + "content": " negative samples. We train our network on the Retrieval-SfM dataset [43] and evaluate it on the Tokyo 24/7 dataset [49], which contains city views in multiple illumination conditions and viewing directions." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 46, + 580, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 286, + 664 + ], + "type": "text", + "content": "Performance is reported as mean Average Precision (mAP) in Table 4. Results of comparison methods are borrowed from [21] and [29]. Our method outperforms all zero-shot methods and is comparable to conventional domain adaptation methods. As shown in Figure 6, the baseline method gets fooled by the night's appearance, while our model finds the correct daytime image." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 47, + 671, + 242, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 242, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 242, + 684 + ], + "type": "text", + "content": "4.5. Low-Light Video Action Recognition" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Although initially designed for images, our method also applies to video tasks. Here we consider an 11-class low-" + } + ] + } + ], + "index": 44 + }, + { + "type": "table", + "bbox": [ + 314, + 327, + 536, + 471 + ], + "blocks": [ + { + "bbox": [ + 323, + 307, + 527, + 319 + ], + "lines": [ + { + "bbox": [ + 323, + 307, + 527, + 319 + ], + "spans": [ + { + "bbox": [ + 323, + 307, + 527, + 319 + ], + "type": "text", + "content": "Table 5. Video action recognition results on ARID [58]." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 327, + 536, + 471 + ], + "lines": [ + { + "bbox": [ + 314, + 327, + 536, + 471 + ], + "spans": [ + { + "bbox": [ + 314, + 327, + 536, + 471 + ], + "type": "table", + "html": "
MethodTop-1 (%)
I3D [3]47.02
Low-Light Video Enhancement
StableLLVE [59]45.08
SMOID [22]47.27
SGZ [61]46.42
Domain Generalization & Zero-Shot Day-Night Domain Adaptation
AdaBN [31]46.17
Ours51.52
", + "image_path": "05e4fb800ed13eff971fe8e1fd52915f7d1e2c40d365705b1c8c08d4fb2d04a9.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "table_body" + } + ], + "index": 46 + }, + { + "bbox": [ + 304, + 495, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 566 + ], + "type": "text", + "content": "light video action recognition task. Normal light training data consists of 2.6k normal light video clips from HMDB51 [27], UCF101 [48], Kinetics-600 [25], and Moments in Time [36]. We evaluate our model on the official test split of the ARID dataset [58]. The action recognizer is I3D [3] based on 3D-ResNet [11]." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 567, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 652 + ], + "type": "text", + "content": "We extend our method to video as follows. When training the darkening module, we input frames extracted from video clips. " + }, + { + "bbox": [ + 304, + 567, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 567, + 545, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 567, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 567, + 545, + 652 + ], + "type": "text", + "content": " in Eq. (4) is estimated for every individual frame. We calculate " + }, + { + "bbox": [ + 304, + 567, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D^{sim}" + }, + { + "bbox": [ + 304, + 567, + 545, + 652 + ], + "type": "text", + "content": " between video clips and other losses between frames. When generating low-light videos, frames are separately fed into the curve estimator while sharing the same exposure map " + }, + { + "bbox": [ + 304, + 567, + 545, + 652 + ], + "type": "inline_equation", + "content": "E^{\\prime}" + }, + { + "bbox": [ + 304, + 567, + 545, + 652 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "We report the results as Top-1 accuracy. As shown in Table 5, video enhancement methods StableLLVE [59], SMOID [22], and SGZ [61] yield a limited performance gain. Meanwhile, our approach boosts models' performance by " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "4.38\\%" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", demonstrating our superiority on videos." + } + ] + } + ], + "index": 49 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8111" + } + ] + } + ], + "index": 50 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 119, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 119, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 119, + 83 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 92, + 289, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 92, + 289, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 92, + 289, + 213 + ], + "type": "text", + "content": "In this paper, we propose a novel approach for zero-shot day-night domain adaptation. Going beyond a simple focus on the image-level translation or model-level adaptation, we observe a complementary relationship between two aspects and build our framework upon the similarity min-max paradigm. Our proposed method can significantly boost the model's performance at nighttime without accessing the nighttime domain. Experiments on multiple datasets demonstrate the superiority and broad applicability of our approach." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 227, + 149, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 227, + 149, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 227, + 149, + 240 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 247, + 289, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 289, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 289, + 344 + ], + "type": "text", + "content": "This work is supported by the Fundamental Research Funds for the Central Universities and the National Natural Science Foundation of China under Contract No.61772043 and a research achievement of Key Laboratory of Science, Technology and Standard in Press Industry (Key Laboratory of Intelligent Press Media Technology). This research work is also partially supported by the Basic and Frontier Research Project of PCL and the Major Key Project of PCL." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 358, + 107, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 107, + 370 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 107, + 370 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 380, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 380, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 401 + ], + "type": "text", + "content": "[1] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv, 2019. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 403, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 458 + ], + "type": "text", + "content": "[2] Vinicius F Arruda, Thiago M Paixão, Rodrigo F Berriel, Alberto F De Souza, Claudine Badue, Nicu Sebe, and Thiago Oliveira-Santos. Cross-domain car detection using unsupervised image-to-image translation: From day to night. In IJCNN, 2019. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 460, + 287, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 460, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 53, + 460, + 287, + 492 + ], + "type": "text", + "content": "[3] João Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, 2017. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 495, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 495, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 53, + 495, + 287, + 529 + ], + "type": "text", + "content": "[4] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv, 2017. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 531, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 531, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 53, + 531, + 287, + 563 + ], + "type": "text", + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 566, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 566, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 53, + 566, + 287, + 609 + ], + "type": "text", + "content": "[6] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T. Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, 2021. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 612, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 612, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 53, + 612, + 287, + 666 + ], + "type": "text", + "content": "[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 669, + 287, + 712 + ], + "type": "text", + "content": "[8] Ziteng Cui, Guo-Jun Qi, Lin Gu, Shaodi You, Zenghui Zhang, and Tatsuya Harada. Multitask aet with orthogonal tangent regularity for dark object detection. In ICCV, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "text", + "content": "[9] Dengxin Dai and Luc Van Gool. Dark model adaptation: Semantic image segmentation from daytime to nighttime. In ITSC, 2018. 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "type": "text", + "content": "[10] Xueqing Deng, Peng Wang, Xiaochen Lian, and Shawn Newsam. Nightlab: A dual-level architecture with hardness detection for segmentation at night. In CVPR, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[11] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 217 + ], + "type": "text", + "content": "[12] Huan Gao, Jichang Guo, Guoli Wang, and Qian Zhang. Cross-domain correlation distillation for unsupervised domain adaptation in nighttime semantic segmentation. In CVPR, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "text", + "content": "[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 264, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 296 + ], + "type": "text", + "content": "[14] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv, 2014. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 298, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 373 + ], + "type": "text", + "content": "[15] Jean-Bastien Grill, Florian Strub, Florent Altché, C. Tallec, Pierre H. Richemond, Elena Buchatskaya, C. Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, B. Piot, K. Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS, 2020. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 376, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 419 + ], + "type": "text", + "content": "[16] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In CVPR, 2020. 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 421, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 421, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 421, + 545, + 453 + ], + "type": "text", + "content": "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 455, + 545, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 455, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 545, + 486 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 488, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 488, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 545, + 521 + ], + "type": "text", + "content": "[19] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In CVPR, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 522, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 545, + 555 + ], + "type": "text", + "content": "[20] Masato Ishii, Takashi Takenouchi, and Masashi Sugiyama. Zero-shot domain adaptation based on attribute information. In ACML, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 556, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 556, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 556, + 545, + 588 + ], + "type": "text", + "content": "[21] Tomas Jenicek and Ondrej Chum. No fear of the dark: Image retrieval under varying illumination conditions. In ICCV, 2019. 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 590, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 611 + ], + "type": "text", + "content": "[22] Haiyang Jiang and Yinqiang Zheng. Learning to see moving objects in the dark. In ICCV, 2019. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "type": "text", + "content": "[23] Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. *Enlightengan: Deep light enhancement without paired supervision. IEEE TIP*, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "text", + "content": "[24] Yeying Jin, Wenhan Yang, and Robby T Tan. Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In ECCV, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "text", + "content": "[25] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola," + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8112" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Tim Green, Trevor Back, Paul Natev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv, 2017. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "text", + "content": "[26] Namyup Kim, Taeyoung Son, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Wedge: web-image assisted domain generalization for semantic segmentation. arXiv, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "text", + "content": "[27] Hildegard Kuehne, Hueihan Jhuang, Estibaliz Garrote, Tomaso A. Poggio, and Thomas Serre. HMDB: A large video database for human motion recognition. In ICCV, 2011. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 216 + ], + "type": "text", + "content": "[28] Hongjun Lee, Moonsoo Ra, and Whoi-Yul Kim. Nighttime data augmentation using gan for improving blind-spot detection. IEEE Access, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 250 + ], + "type": "text", + "content": "[29] Attila Lengyel, Sourav Garg, Michael Milford, and Jan C. van Gemert. Zero-shot day-night domain adaptation with a physics prior. In ICCV, 2021. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 251, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 287, + 282 + ], + "type": "text", + "content": "[30] Chongyi Li, Chunle Guo, and Chen Change Loy. Learning to enhance low-light image via zero-reference deep curve estimation. IEEE TPAMI, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 284, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 287, + 316 + ], + "type": "text", + "content": "[31] Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting batch normalization for practical domain adaptation. In ICLRW, 2017. 2, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 317, + 287, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 317, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 317, + 287, + 349 + ], + "type": "text", + "content": "[32] Guosheng Lin, Anton Milan, Chunhua Shen, and Ian Reid. Refinenet: Multi-path refinement networks for high-resolution semantic segmentation. In CVPR, 2017. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "text", + "content": "[33] Risheng Liu, Long Ma, Jiaao Zhang, Xin Fan, and Zhongxuan Luo. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In CVPR, 2021. 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 393, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 426 + ], + "type": "text", + "content": "[34] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In CVPR, 2022. 1, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "type": "text", + "content": "[35] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv, 2017. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "text", + "content": "[36] Mathew Monfort, Carl Vondrick, Aude Oliva, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa M. Brown, Quanfu Fan, and Dan Gutfreund. Moments in time dataset: One million videos for event understanding. IEEE TPAMI, 2020. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 514, + 287, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 559 + ], + "type": "text", + "content": "[37] Amitangshu Mukherjee, Ameya Joshi, Anuj Sharma, Chinmay Hegde, and Soumik Sarkar. Generative semantic domain adaptation for perception in autonomous driving. Journal of Big Data Analytics in Transportation, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 559, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 591 + ], + "type": "text", + "content": "[38] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 2, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 592, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 613 + ], + "type": "text", + "content": "[39] Kuan-Chuan Peng, Ziyan Wu, and Jan Ernst. Zero-shot deep domain adaptation. In ECCV, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 615, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 615, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 615, + 287, + 669 + ], + "type": "text", + "content": "[40] Stephen M Pizer, E Philip Amburn, John D Austin, Robert Cromartie, Ari Geselowitz, Trey Greer, Bart ter Haar Romy, John B Zimmerman, and Karel Zuiderveld. Adaptive histogram equalization and its variations. Computer Vision, Graphics, and Image Processing, 1987. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[41] Abhijith Punnappurath, Abdullah Abuolaim, Abdelrahman Abdelhamed, Alex Levinshtein, and Michael S Brown. Day-to-night image synthesis for training nighttime neural isps. In CVPR, 2022. 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[42] Filip Radenovic, Giorgos Tolias, and Ondrej Chum. Deep shape matching. In ECCV, 2018. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 96, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 128 + ], + "type": "text", + "content": "[43] Filip Radenovic, Giorgos Tolias, and Ondrej Chum. Finetuning cnn image retrieval with no human annotation. IEEE TPAMI, 2019. 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "text", + "content": "[44] Zia-ur Rahman, Daniel J Jobson, and Glenn A Woodell. Retinex processing for automatic image enhancement. Journal of Electronic Imaging, 2004. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 163, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 207 + ], + "type": "text", + "content": "[45] Eduardo Romera, Luis M. Bergasa, Kailun Yang, Jose M. Alvarez, and Rafael Barea. Bridging the day and night domain gap for semantic segmentation. IEEE Intelligent Vehicles Symposium, 2019. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "text", + "content": "[46] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Guided curriculum model adaptation and uncertainty-aware evaluation for semantic nighttime image segmentation. In ICCV, 2019. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 252, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 285 + ], + "type": "text", + "content": "[47] Yukihiro Sasagawa and Hajime Nagahara. Yolo in the dark-domain adaptation method for merging multiple models. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "text", + "content": "[48] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv, 2012. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 320, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 354 + ], + "type": "text", + "content": "[49] Akihiko Torii, Relja Arandjelovic, Josef Sivic, Masatoshi Okutomi, and Tomas Pajdla. 24/7 place recognition by view synthesis. In CVPR, 2015. 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 354, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 376 + ], + "type": "text", + "content": "[50] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 377, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 545, + 409 + ], + "type": "text", + "content": "[51] Jinghua Wang and Jianmin Jiang. Conditional coupled generative adversarial networks for zero-shot domain adaptation. In ICCV, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "type": "text", + "content": "[52] Wenjing Wang, Xinhao Wang, Wenhan Yang, and Jiaying Liu. Unsupervised face detection in the dark. IEEE TPAMI, 2022. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 445, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 477 + ], + "type": "text", + "content": "[53] Wenjing Wang, Zhengbo Xu, Haofeng Huang, and Jiaying Liu. Self-aligned concave curve: Illumination enhancement for unsupervised adaptation. In ACM MM, 2022. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 479, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 511 + ], + "type": "text", + "content": "[54] Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, LapPui Chau, and Alex Kot. Low-light image enhancement with normalizing flow. In AAAI, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "type": "text", + "content": "[55] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. In BMVC, 2017. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 545, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 590 + ], + "type": "text", + "content": "[56] Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Wenhan Yang, and Jianmin Jiang. Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. In CVPR, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 591, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 591, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 591, + 545, + 633 + ], + "type": "text", + "content": "[57] Xinyi Wu, Zhenyao Wu, Hao Guo, Lili Ju, and Song Wang. Darnet: A one-stage domain adaptation network for unsupervised nighttime semantic segmentation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "text", + "content": "[58] Yuecong Xu, Jianfei Yang, Haozhi Cao, Kezhi Mao, Jianxiong Yin, and Simon See. Arid: A new dataset for recognizing action in the dark. In Deep Learning for Human Activity Recognition, 2021. 8" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[59] Fan Zhang, Yu Li, Shaodi You, and Ying Fu. Learning temporal consistency for low light video enhancement from single images. In CVPR, 2021. 8" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8113" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 240 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[60] Yuzhi Zhao, Yongzhe Xu, Qiong Yan, Dingdong Yang, Xuehui Wang, and Lai-Man Po. D2hnet: Joint denoising and deblurring with hierarchical network for robust night image restoration. In ECCV, 2022. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "text", + "content": "[61] Shen Zheng and Gaurav Gupta. Semantic-guided zero-shot learning for low-light image/video enhancement. In WACVW, 2022. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 287, + 174 + ], + "type": "text", + "content": "[62] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. In ICLR, 2021. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 175, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 287, + 206 + ], + "type": "text", + "content": "[63] Shangchen Zhou, Chongyi Li, and Chen Change Loy. Lednet: Joint low-light enhancement and deblurring in the dark. In ECCV, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 209, + 287, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 287, + 240 + ], + "type": "text", + "content": "[64] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. 5" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8114" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Simoun_ Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning/17dc049b-d1dc-4e17-921e-1317f662d966_content_list.json b/2023/Simoun_ Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning/17dc049b-d1dc-4e17-921e-1317f662d966_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..890e0252c6651a68863353089bc613081952ba0e --- /dev/null +++ b/2023/Simoun_ Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning/17dc049b-d1dc-4e17-921e-1317f662d966_content_list.json @@ -0,0 +1,1608 @@ +[ + { + "type": "text", + "text": "Simoun: Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning", + "text_level": 1, + "bbox": [ + 158, + 164, + 815, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yangru Huang $^{1}$ , Peixi Peng $^{2,4*}$ , Yifan Zhao $^{1}$ , Yunpeng Zhai $^{1}$ , Haoran Xu $^{3,4}$ , Yonghong Tian $^{1,2,4*}$", + "bbox": [ + 140, + 229, + 836, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science, Peking University", + "bbox": [ + 316, + 247, + 666, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Electronic and Computer Engineering, Peking University", + "bbox": [ + 243, + 263, + 733, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Intelligent Systems Engineering, Sun Yat-sen University $^{4}$ Peng Cheng Laboratory", + "bbox": [ + 151, + 277, + 828, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yrhuang@stu.pku.edu.cn, {pxpeng, zhaoyf, ypzhai, yhtian}@pku.edu.cn, xuhr9@mail2.sysu.edu.cn", + "bbox": [ + 125, + 297, + 852, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 259, + 342, + 330, + 356 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Efficient motion and appearance modeling are critical for vision-based Reinforcement Learning (RL). However, existing methods struggle to reconcile motion and appearance information within the state representations learned from a single observation encoder. To address the problem, we present Synergizing Interactive Motion-appearance Understanding (Simoun), a unified framework for vision-based RL. Given consecutive observation frames, Simoun deliberately and interactively learns both motion and appearance features through a dual-path network architecture. The learning process collaborates with a structural interactive module, which explores the latent motion-appearance structures from the two network paths to leverage their complementarity. To promote sample efficiency, we further design a consistency-guided curiosity module to encourage the exploration of under-learned observations. During training, the curiosity module provides intrinsic rewards according to the consistency of environmental temporal dynamics, which are deduced from both motion and appearance network paths. Experiments conducted on DeepMind control suite and CARLA automatic driving benchmarks demonstrate the effectiveness of Simoun, where it performs favorably against state-of-the-art methods.", + "bbox": [ + 114, + 372, + 473, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 117, + 715, + 235, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reinforcement learning (RL) from visual signals has achieved great success in recent years. Compared with learning from hand-crafted states, vision-based RL eliminates the arduous task of designing states with manual feature engineering. Therefore, it is beneficial for a variety of tasks such as video game playing [25, 40], robot manipulation [50], and autonomous navigation [4]. However, one of", + "bbox": [ + 116, + 739, + 473, + 838 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cb2dcbe537421bf96fdf6defd6d1f075d5a29e8ffebc48afd892c01c3662331e.jpg", + "image_caption": [ + "Figure 1. Illustration of different observation encoding schemes. (a) Encoding individual frames with a shared encoder. (b) Encoding stacked multiple frames. (c) Encoding individual frames and latent vector differences (latent flows). (d) Encoding motion and appearance independently and interactively." + ], + "image_footnote": [], + "bbox": [ + 511, + 340, + 846, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the major challenges of vision-based RL lies in its high dimensional observation, which is less interpretable and leads to low rewards and inefficient sampling [19, 20]. As a result, robust visual understanding is crucial to bridge the gap between vision and state-based RL in terms of performance and sample efficiency.", + "bbox": [ + 500, + 667, + 854, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To make observation comprehensible for agents, recent works have recognized the importance of learning high-quality visual features [20, 5, 29, 49, 42]. From the perspective of decision-making, two kinds of features are essential for vision-based RL: the motion features, which closely relate to the actions performed by the agents, and the appearance features, which describe the contemporary environmental states. Despite their importance, few works", + "bbox": [ + 500, + 752, + 854, + 863 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 138, + 45, + 207, + 80 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 259, + 42, + 781, + 56 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 336, + 56, + 702, + 70 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 308, + 70, + 732, + 84 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 133, + 849, + 250, + 862 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "176", + "bbox": [ + 486, + 888, + 509, + 898 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "have tried to model these two kinds of features explicitly. Most methods tend to encode single or stacked multiple frames via a convolutional encoder, as shown in Fig. 1(a) and Fig. 1(b). The former neglects motion feature modeling, while the latter tightly entangles the motion and appearance information into a single network, causing information bias and optimization difficulty. To alleviate these issues, Shang et al. [33] propose to encode only appearance features in single frames and impose motion dynamics by latent vector differences (termed latent flow, see Fig. 1(c)). In addition, several world models learned directly from input images are also developed [11, 31, 3]. Nevertheless, the motion features of these methods are still computed from spatial latent space, which may lead to sub-optimal representation learning in complicated environments.", + "bbox": [ + 114, + 128, + 473, + 334 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the limitations described above, we present Synergizing Interactive Motion-appearance Understanding (Simoun) for vision-based RL. The design principle of Simoun is to explicitly learn both motion and appearance features in the early encoding stage and then interactively fuse them at the late stage for accurate decision-making. As shown in Fig. 1(d), given consecutive environment frames, Simoun learns motion and appearance information by separate network paths. The motion path (colored in blue) explicitly models motion clues (such as the speed and direction of cars) from the residual frame of multiple neighboring input frames. The appearance path (colored in orange) models the environmental spatial structures and focuses on identifying patterns and objects (such as cars and traffic lights) from every single frame. Additionally, a structural interactive module further extracts latent motion-appearance structures reflected by the correlations of the dual-path features. It then modulates both paths with the computed structure masks. In this way, each path can take complementary information from the other during learning, and agents are able to better understand the spatial and temporal context of the environment. Finally, the latent vectors from the two paths are fused for decision-making.", + "bbox": [ + 114, + 338, + 473, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although the dual-path design of Simoun promotes observation interpretability, low sample efficiency still exists due to finite data and sparse rewards. To alleviate this issue, a consistency-guided curiosity module is further designed. The idea is to adapt the learning process by concentrating more on under-learned observations, which can be deduced from the consistency of the motion and appearance paths. Intuitively, both the motion path and the appearance path describe the same observations. Hence, the dynamic information inferred from the following two sources should remain consistent: 1) latent motion vectors learned directly from the motion path and 2) differences between the appearance path latent vectors over multiple neighboring frames. If the opposite were true, then it indicates premature motion-appearance understanding, thus more explo", + "bbox": [ + 114, + 657, + 473, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ration should be added. In this way, we build a strong correlation between reward discovery and state novelty. During training, the consistency-guided curiosity module provides intrinsic rewards in addition to extrinsic rewards from the environment, resulting in more efficient exploration.", + "bbox": [ + 498, + 128, + 856, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments on both CARLA and OpenAI DMControl environments show that the proposed method performs favorably against state-of-the-arts. Overall, the contributions of this paper are threefold:", + "bbox": [ + 498, + 198, + 856, + 252 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) We propose Simoun, a novel dual-path learning framework that explicitly and interactively learns both motion and appearance information from observations.", + "(2) We design a structural interactive module to fully explore the complementarity of the two paths in Simoun and thereby further enhance visual understanding.", + "(3) We devise a consistency-guided curiosity module to encourage the exploration of under-learned observations. The proposed module effectively increases sample efficiency by providing intrinsic rewards for the agents." + ], + "bbox": [ + 500, + 253, + 854, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 401, + 635, + 416 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-based Reinforcement Learning To improve the performance and sample efficiency of vision-based RL, existing works can be roughly divided into three groups: 1) designing auxiliary loss/learning tasks [20, 14, 36, 49, 23], 2) employing various data augmentation techniques [21, 41, 2, 17, 24], and 3) modeling environment dynamics [12, 11, 13, 7, 27]. However, most existing works utilize a single-path network with multi-frame inputs, in which the motion and appearance information is tightly entangled without explicit separation. One exception is Flare [33], which also utilizes a single-path network to encode each frame individually and models motion information explicitly by taking latent vector differences. Although Flare achieves improved performance, its motion information is still computed from single-frame appearance features, causing insufficient temporal information extraction.", + "bbox": [ + 498, + 424, + 854, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dual-path Networks for Visual Modeling There is a rich literature of works on visual modeling with dual-path networks. One of the earliest works is the two-stream CNNs for action recognition [35], which utilizes a spatial stream with single-frame input and a temporal stream taking multiframe optical flows. Thereafter, the concept of dual-path networks is heavily explored with various fusion strategies on different tasks [34, 6, 39, 43]. Our approach differs from existing ones in terms of task, architecture and learning mechanism. Compare with other visual tasks, vision-based RL needs to extract fine-grained motion details across different time steps, which poses a great challenge to RL models. To the best of our knowledge, Simoun is one of the first vision-based RL methods which aim to explicitly learn motion and appearance features by a dual-path architecture. Meanwhile, its structural interactive module is", + "bbox": [ + 498, + 643, + 856, + 862 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "177", + "bbox": [ + 486, + 888, + 509, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5df02ca57eb18b68f9e9bf127e3ec6b2a9ad426b9cc2caa043529ed1fba78b4d.jpg", + "image_caption": [ + "Figure 2. The proposed dual-path learning framework Simoun. It introduces two modules: 1) a structural interactive module that leverages the complementarity between the two network paths by extracting latent motion-appearance structures, and 2) a consistency-guided curiosity module that provides intrinsic rewards according to the dynamic consistency between the motion and appearance features." + ], + "image_footnote": [], + "bbox": [ + 135, + 128, + 836, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "also deliberately designed to extract motion and appearance clues residing in consecutive observation frames. By learning with RL-oriented objectives and curiosity-driven strategies, Simoun successfully imports the idea of dual-path networks into robust visual control.", + "bbox": [ + 114, + 441, + 473, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Intrinsic Reward Exploration One of the key elements of RL is the reward function, which aims to quantify the \"goodness\" of the agent's decisions. However, the problem is that designing dense, well-defined extrinsic reward functions is difficult and unscalable. One possible solution is to introduce an intrinsic reward function, which is calculated by the agents. Existing intrinsic reward exploration methods mainly focus on counting or predicting state novelty [1], prediction error [28], uncertainty [22], or environmental dynamics [32]. These methods are typically designed for the general state-based instead of vision-based RL. The works most relevant to us are CCFDM [26] and CCLF [37], which also formulate intrinsic rewards for visual-based RL. CCFDM utilizes forward dynamics and CCLF is based on a contrastive term. Differently, the consistency-guided curiosity module in Simoun exploits the dynamic consistency extracted from both motion and appearance features to achieve reliable intrinsic reward estimation.", + "bbox": [ + 114, + 510, + 473, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 116, + 770, + 240, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We start by formalizing the task of vision-based RL in Sec. 3.1 and then delineate Simoun in detail. The general idea of Simoun is to 1) explicitly model the motion/appearance dynamics (and their correlations) of the agent's operating ambiance and 2) capitalize on the learned", + "bbox": [ + 114, + 792, + 473, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dynamics to achieve efficient exploration. As shown in Fig. 2, our framework consists of two parts. First, given inputs in the form of consecutive frames, it interactively models the motion and appearance features using a dualpath architecture with a set of targeted objectives and a structural interactive module (Sec. 3.2). Then the decision-making strategy is learned with adaptive curiosity assignment steered by the proposed consistency-guided curiosity module (Sec. 3.3). The overall learning objective and inference process of Simoun are finally summarized in Sec. 3.4.", + "bbox": [ + 496, + 441, + 857, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem formulation", + "text_level": 1, + "bbox": [ + 500, + 585, + 679, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-based RL can be formulated as a Partially Observable Markov Decision Process (POMDP) $\\mathcal{M} = < \\mathcal{O},\\mathcal{A},\\mathcal{P},\\mathcal{R},\\gamma >$ where $\\mathcal{O}$ denotes the observation space containing pixel frames $\\mathbf{o}_t$ at different time step $t$ and $\\mathcal{A}$ denotes the action space. At each $t$ the agent chooses an action $a_{t}\\in \\mathcal{A}$ . $\\mathcal{P}(\\mathbf{o}_{t + 1}|\\mathbf{o}_t,a_t)$ is the observation transition, $\\mathcal{R}(\\mathbf{o}_t,a_t)$ is the reward function, and $\\gamma \\in [0,1)$ is the discount factor. The goal is to identify an optimal policy $\\pi$ that maximizes the expected cumulative reward based on the visual observations rather than the complete environment state:", + "bbox": [ + 496, + 599, + 856, + 748 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nJ (\\pi) = \\sum_ {t} \\mathbb {E} _ {\\left(\\mathbf {o} _ {t}, a _ {t}\\right) \\sim \\pi} \\left[ \\mathcal {R} \\left(\\mathbf {o} _ {t}, a _ {t}\\right) \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 750, + 854, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During training, $\\pi$ is used to interact with the environment, and the related data are stored in a replay buffer $\\mathcal{B}$ .", + "bbox": [ + 496, + 780, + 854, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Dual-path Interactive Modeling", + "text_level": 1, + "bbox": [ + 500, + 813, + 756, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In vision-based RL, motion and appearance clues are both vital for the agent to perform accurate decision", + "bbox": [ + 498, + 835, + 854, + 863 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "178", + "bbox": [ + 486, + 888, + 509, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "making. Simoun adopts a dual-path architecture to interactively learn motion and appearance features. Specifically, it consists of three components: the motion path, the appearance path, and the structural interactive module that serves as a switchboard between the motion and appearance structures learned from the two paths.", + "bbox": [ + 114, + 128, + 473, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion Path Motion information is crucial for the visual system to understand the dynamics of the surrounding environment. The motion path aims to extract the motion-related features (such as the velocity of moving objects) from the changes between consecutive frames. Given a tuple of three adjacent observations $[\\mathbf{o}_{t - 2},\\mathbf{o}_{t - 1},\\mathbf{o}_t]$ sampled from replay buffer $\\mathcal{B}$ , the input of the motion path is the residual of adjacent frames $[(0_{t - 1} - 0_{t - 2}),(0_t - 0_{t - 1})]$ , concatenated along channel dimensions. A convolutional encoder $\\mathcal{E}^m$ is then used to learn a lower-dimensional motion representation. Specifically, the encoder includes four convolution layers with $3\\times 3$ kernel size and ReLU nonlinearity. Denote the feature map of the last convolution layer as $\\mathbf{F}_t^m\\in \\mathbb{R}^{c\\times h\\times w}$ , a fully connected (FC) layer with layer normalization (LN) is used to reduce the dimension of $\\mathbf{F}_t^m$ to get the motion feature vector $\\mathbf{f}_t^m$ .", + "bbox": [ + 114, + 214, + 473, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further impel the motion path to catch abundant environment temporal structures, a motion-aware constraint is deliberately designed. Specifically, given $\\mathbf{f}_t^m$ and corresponding action $a_{t}$ at timestep $t$ , an action-conditioned two-FC-layer transition model $\\mathcal{G}$ is used to obtain a motion-action joint representation of the current timestep. A latent-space transition loss is then defined as:", + "bbox": [ + 114, + 436, + 473, + 531 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t r a n} = \\left\\| \\mathcal {G} \\left(\\mathbf {f} _ {t} ^ {m}, u _ {t}\\right) - \\mathbf {f} _ {t + 1} ^ {m} \\right\\| _ {2} ^ {2}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 534, + 472, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which encourages $\\mathbf{f}_t^m$ to model the motion trends over time by predicting the future temporal dynamics $\\mathbf{f}_{t + 1}^{m}$ of the next time step.", + "bbox": [ + 114, + 551, + 472, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Appearance Path The appearance path is designed to extract the visual appearance representations of the object and scene from individual observation frames. The input of the appearance path is a single frame $\\mathbf{o}_t$ from the replay buffer $\\mathcal{B}$ . Similar to the motion path, the appearance path also adopts a four-layer encoder $\\mathcal{E}^a$ with the same architecture (except for the number of input channels of the first layer) to get feature map $\\mathbf{F}_t^a\\in \\mathbb{R}^{c\\times h\\times w}$ , followed by FC and LN layers to get the appearance representation $\\mathbf{f}_t^a$ .", + "bbox": [ + 114, + 595, + 473, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To explicitly make appearance representations discriminative to different scenes, we take inspiration from CURL [20] and adopt an unsupervised contrastive loss between similar and dissimilar sample pairs. Given observation frame $\\mathbf{o}_i$ , we consider $\\tilde{\\mathbf{o}}_i$ (obtained from $\\mathbf{o}_i$ by data augmentation) as the positive sample, while samples coming from different observation frames are regarded as negatives. The contrastive loss is then defined as:", + "bbox": [ + 114, + 722, + 473, + 832 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8b95db8882fb3931dede5d8f3b89fbe4514957d962f6cbf4ae2300b8e67efd28.jpg", + "image_caption": [ + "Figure 3. Illustration of the proposed structural interactive module." + ], + "image_footnote": [], + "bbox": [ + 509, + 128, + 848, + 313 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c o n} = - \\log \\frac {\\exp \\left(\\mathbf {f} _ {i} ^ {a ^ {\\top}} \\tilde {\\mathbf {f}} _ {i} ^ {a}\\right)}{\\exp \\left(\\mathbf {f} _ {i} ^ {a ^ {\\top}} \\tilde {\\mathbf {f}} _ {i} ^ {a}\\right) + \\sum_ {j = 0} ^ {K - 1} \\exp \\left(\\mathbf {f} _ {i} ^ {a ^ {\\top}} \\tilde {\\mathbf {f}} _ {j} ^ {a}\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 369, + 854, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_t^a$ and $\\tilde{\\mathbf{f}}_t^a$ are appearance representations of $\\mathbf{o}_t$ and $\\tilde{\\mathbf{o}}_t$ , respectively. $j$ is the sample index in a training batch with $K$ samples. Note that the augmented data is only used for representation learning and is inaccessible during decision-making. This is to avoid unstable training, which is easily caused by the augmented data dominating the evaluation of the Q-target.", + "bbox": [ + 498, + 409, + 854, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Structural Interactive Module Instead of simply concatenating the features from motion and appearance paths as state representations, interactively communicating the learned motion-appearance structures between both paths can lead to a more robust visual understanding. Intuitively, as shown in Fig. 3(left), given consecutive observation frames, the motion structure can be revealed by inter-frame appearance variation. Meanwhile, the appearance structure can be obtained by spatial pixel correlation. Therefore, the structural interactive module embraces a single and interframe relation discovering mechanism for efficient motion-appearance structure mining and propagation.", + "bbox": [ + 498, + 508, + 854, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, as shown in Fig. 3(right), the inputs of the structural interactive module are motion feature map $\\mathbf{F}_t^m$ and appearance feature maps $\\mathbf{F}_{t - 2}^{a},\\mathbf{F}_{t - 1}^{a}$ and $\\mathbf{F}_t^a$ .To extract motion structures, an inter-frame attention mechanism is designed. In particular, $\\mathbf{F}_{t - 1}^{a}$ is treated as the query, and $\\mathbf{F}_{t - 2}^{a},\\mathbf{F}_{t}^{a}$ is used as two keys. A inter-frame attention map $\\mathbf{X}$ can be obtained as:", + "bbox": [ + 500, + 675, + 854, + 770 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} = \\sigma \\left(\\sigma \\left(\\hat {\\mathbf {F}} _ {t - 2} ^ {a} ^ {\\top} \\hat {\\mathbf {F}} _ {t - 1} ^ {a}\\right) + \\sigma \\left(\\hat {\\mathbf {F}} _ {t} ^ {a \\top} \\hat {\\mathbf {F}} _ {t - 1} ^ {a}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 772, + 854, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{F}}_t^a\\in \\mathbb{R}^{c\\times h\\times w}$ denotes a new feature map generated by feeding $\\mathbf{F}_t^a$ to a convolution layer for reducing the spatial complexity ( $\\hat{\\mathbf{F}}_{t - 1}^a$ and $\\hat{\\mathbf{F}}_{t - 2}^a$ can be obtained similarly). $\\sigma$ denotes the Softmax function. Combining the outputs of two initial Softmax functions might not yield values sum", + "bbox": [ + 498, + 792, + 854, + 863 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "We choose three here following the practice in [44, 47].", + "bbox": [ + 133, + 849, + 408, + 863 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "179", + "bbox": [ + 486, + 888, + 509, + 898 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/50e6f72ebcab46802578cdc37e638c109921863340f1079d7e2328bd703cc8ee.jpg", + "image_caption": [ + "Figure 4. The idea of consistency-guided curiosity module." + ], + "image_footnote": [], + "bbox": [ + 127, + 127, + 460, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ming to 1, so an additional Softmax is used to normalize the final attention scores. The soft indexing of $\\hat{\\mathbf{F}}_{t - 1}^{a}$ by $\\hat{\\mathbf{F}}_{t - 2}^{a}$ or $\\hat{\\mathbf{F}}_t^a$ is performed at the spatial dimensions, resulting in a soft attention map of dimensions $hw\\times hw$ . Then a spatial gating operation is applied on $\\mathbf{X}$ using weights calculated from both $\\mathbf{F}_t^m$ and $\\mathbf{F}_t^a$ to consolidate appearance structures and obtain two motion-appearance structure masks:", + "bbox": [ + 116, + 354, + 472, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} ^ {m} = \\mathbf {X} \\cdot \\hat {\\mathbf {F}} _ {t} ^ {m}, \\mathbf {X} ^ {a} = \\mathbf {X} \\cdot \\hat {\\mathbf {F}} _ {t} ^ {a}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 455, + 470, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{F}}_t^m$ is a feature map of $\\mathbf{F}_t^m$ obtained with another convolution layer. Finally, $\\mathbf{F}_t^m$ and $\\mathbf{F}_t^a$ are enhanced by modulating with their corresponding motion-appearance structure masks:", + "bbox": [ + 116, + 472, + 472, + 526 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {t} ^ {m} = \\mathbf {F} _ {t} ^ {m} + \\boldsymbol {\\beta} \\cdot \\mathbf {X} ^ {m}, \\mathbf {F} _ {t} ^ {a} = \\mathbf {F} _ {t} ^ {a} + \\boldsymbol {\\beta} \\cdot \\mathbf {X} ^ {a}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 530, + 470, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ is a learnable parameter initialized as zero [48], and the enhanced $\\mathbf{F}_t^m$ and $\\mathbf{F}_t^a$ instead of the original ones are further used to calculate $\\mathbf{f}_t^m$ and $\\mathbf{f}_t^a$ .", + "bbox": [ + 116, + 547, + 472, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After acquiring the structure-enhanced motion and appearance features, they are finally fused as $\\mathbf{f}_t = [\\mathbf{f}_t^m,\\mathbf{f}_t^a ]$ where $\\mathbf{f}_t$ is used as the final state representation. To better capture reward-related features from the fused state representation, a reward function $\\mathcal{R}$ is further introduced to predict a numerical reward value to each state-action pair. $\\mathcal{R}$ has a similar architecture with the transition model $\\mathcal{G}$ except the output dimension is set to one. A reward loss $\\mathcal{L}_{re}$ is then defined as the mean squared error between the predicted and actual reward:", + "bbox": [ + 116, + 590, + 473, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e} = \\left(\\mathcal {R} \\left(\\mathbf {f} _ {t}, a _ {t}\\right) - r _ {t + 1} ^ {e}\\right) ^ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 728, + 472, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $r_{t + 1}^{e}$ is the actual external reward value at the next time step, which is returned by the environment.", + "bbox": [ + 116, + 747, + 472, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Consistency-guided Curiosity Module", + "text_level": 1, + "bbox": [ + 116, + 782, + 413, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given state representation $\\mathbf{f}_t$ from the dual-path model, we adapt SAC [9, 10] as the base RL algorithm following", + "bbox": [ + 116, + 803, + 472, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "previous methods [47, 5], which aims to maximize the expected cumulative reward to find an optimal policy by approximating the action-value function $Q$ and a stochastic policy $\\pi$ based on a $\\alpha$ -discounted maximum entropy $\\mathcal{H}(\\cdot)$ :", + "bbox": [ + 500, + 128, + 854, + 184 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nJ (\\pi) = \\sum_ {t} \\mathbb {E} _ {\\left(\\mathbf {o} _ {t}, a _ {t}\\right) \\sim \\pi} \\left[ r \\left(\\mathbf {o} _ {t}, a _ {t}\\right) + \\alpha \\mathcal {H} \\left(\\pi \\left(\\cdot \\mid \\mathbf {o} _ {t}\\right)\\right) \\right]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 186, + 854, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The action-value function $Q$ are learned by minimizing the soft Bellman error:", + "bbox": [ + 500, + 215, + 854, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {Q} = \\mathbb {E} _ {\\left(\\mathbf {o} _ {t}, a _ {t}\\right)} \\left(Q \\left(\\mathbf {o} _ {t}, a _ {t}\\right) - \\left(r _ {t} + \\lambda V \\left(\\mathbf {o} _ {t + 1}\\right)\\right)\\right) ^ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 244, + 854, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the soft state value $V$ can be estimated by sampling an action under the current policy:", + "bbox": [ + 500, + 263, + 854, + 291 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV \\left(\\mathbf {o} _ {t + 1}\\right) = \\mathbb {E} _ {a _ {t + 1} \\sim \\pi} \\left[ \\bar {Q} \\left(\\mathbf {o} _ {t + 1}; a _ {t + 1}\\right) - \\alpha \\log \\pi \\left(a _ {t + 1} \\mid \\mathbf {o} _ {t + 1}\\right) \\right], \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 309, + 854, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\overline{Q}$ denotes the exponential moving average of the parameters of $Q$ . The policy is optimized by decreasing the difference between the exponential of the soft-Q function and the policy:", + "bbox": [ + 500, + 338, + 854, + 392 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\pi} = \\mathbb {E} _ {a _ {t} \\sim \\pi} \\left[ \\alpha \\log \\pi \\left(a _ {t} \\mid \\mathbf {o} _ {t}\\right) - Q \\left(\\mathbf {o} _ {t}, a _ {t}\\right) \\right]. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 395, + 854, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although SAC algorithm introduces the entropy term to encourage exploration, it still heavily relies on carefully engineered environmental extrinsic rewards and suffers from low sample efficiency. Instead of extrinsic rewards, intrinsic curiosity can be a powerful concept to endow an agent with an automated mechanism to continuously explore its environment in the absence of task information. During training, in addition to the extrinsic rewards $r^e$ obtained from the environment, the curiosity module also provides intrinsic rewards $r^i$ . The learning objective SAC in Eq. 8 is therefore extended as:", + "bbox": [ + 498, + 422, + 854, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nJ (\\pi) = \\sum_ {t} \\mathbb {E} _ {\\left(\\mathbf {o} _ {t}, a _ {t}\\right) \\sim \\pi} \\left[ r ^ {e} \\left(\\mathbf {o} _ {t}, a _ {t}\\right) + \\alpha \\mathcal {H} \\left(\\pi (\\cdot | \\mathbf {o} _ {t})\\right) + r ^ {i} \\left(\\mathbf {o} _ {t}, a _ {t}\\right) \\right]. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 571, + 864, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To get proper intrinsic rewards $r^i$ , we propose to leverage the consistency of motion and appearance features from the dual-path model. Intuitively, as shown in Fig. 4, the environmental temporal dynamics can be obtained from two sources: 1) directly learned from the motion path, and 2) through the variations of the spatial features learned from the spatial path, similar to the latent flow [33]. Optimally, these two sources should be consistent with each other. That is, the motion feature $\\mathbf{f}_t^m$ should be similar to the difference between the spatial features $\\mathbf{f}_{t - 2}^a$ , $\\mathbf{f}_{t - 1}^a$ and $\\mathbf{f}_t^a$ . The idea of the consistency-guided curiosity module is then to encourage the agent to explore when the temporal dynamics produced by the two sources are inconsistent. To this end, the intrinsic reward $r^i$ is defined as:", + "bbox": [ + 498, + 613, + 854, + 804 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr ^ {i} = C e ^ {- \\lambda t} d \\left(\\left\\| \\mathbf {f} _ {t} ^ {m} \\right\\|. \\left| \\mathbf {f} _ {t - 1} ^ {a} - \\mathbf {f} _ {t - 2} ^ {a} \\right| + \\left| \\mathbf {f} _ {t} ^ {a} - \\mathbf {f} _ {t - 1} ^ {a} \\right|\\right) \\frac {r _ {\\max} ^ {e}}{t _ {\\max} ^ {i}}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 805, + 854, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C$ is temperature weight, $\\lambda$ is an exponential decay weight, $d$ is the L2 distance function, $t$ is the environment", + "bbox": [ + 500, + 835, + 854, + 862 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2We use concatenation to fuse $\\mathbf{f}_t^m$ and $\\mathbf{f}_t^a$ for simplicity, advanced fusion mechanisms are also applicable.", + "bbox": [ + 116, + 838, + 472, + 862 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "180", + "bbox": [ + 488, + 888, + 509, + 898 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Inference procedure of Simoun" + ], + "code_body": "1: for each environment step $t$ do \n2: Collect observation frames $\\mathbf{o}_{t-2}$ , $\\mathbf{o}_{t-1}$ and $\\mathbf{o}_t$ . \n3: Extract motion and appearance feature maps: \n4: $\\mathbf{F}_t^m = \\mathcal{E}^m ([\\mathbf{o}_{t-1} - \\mathbf{o}_{t-2}), (\\mathbf{o}_t - \\mathbf{o}_{t-1})]$ , \n5: $\\mathbf{F}_j^a = \\mathcal{E}^a (\\mathbf{o}_j)$ , $j = t - 2, \\dots, t$ \n6: Structural Interactive Modeling: \n7: $\\mathbf{X} = \\sigma (\\sigma (\\hat{\\mathbf{F}}_{t-2}^a \\top \\hat{\\mathbf{F}}_{t-1}^a) + \\sigma (\\hat{\\mathbf{F}}_t^{a\\top}\\hat{\\mathbf{F}}_{t-1}^a))$ \n8: $\\mathbf{F}_t^m = \\mathbf{F}_t^m + \\beta \\cdot (\\mathbf{X} \\cdot \\hat{\\mathbf{F}}_t^m)$ \n9: $\\mathbf{F}_t^a = \\mathbf{F}_t^a + \\beta \\cdot (\\mathbf{X} \\cdot \\hat{\\mathbf{F}}_t^a)$ \n10: Extract low-dimensional features: \n11: $\\mathbf{f}_t^m = \\mathrm{LayerNorm}(\\mathrm{FC}(\\mathbf{F}_t^m))$ \n12: $\\mathbf{f}_t^a = \\mathrm{LayerNorm}(\\mathrm{FC}(\\mathbf{F}_t^a))$ \n13: Take action based on the fused features: \n14: $\\mathbf{f}_t = [\\mathbf{f}_t^m, \\mathbf{f}_t^a]$ \n15: $a_t \\sim \\pi (a_t | \\mathbf{f}_t)$ \n16: Environment state transition: \n17: $\\mathbf{o}_{t+1} \\sim \\mathcal{P}(\\mathbf{o}_{t+1} | \\mathbf{o}_t, a_t)$ \n18: end for", + "bbox": [ + 125, + 143, + 468, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "time step, $|\\cdot |$ is the vector element-wise absolute operation (used to avoid the cancellation of $\\mathbf{f}_{t - 1}^s$ ), $r_{max}^{e}$ and $r_{max}^{i}$ are the maximum extrinsic reward value and intrinsic reward value over the time step $[0,t]$ , which serve as normalization term to balance $r^i$ and $r^e$ . At the beginning of the training, $r_{max}^{e}$ and $r_{max}^{i}$ are assigned with an initial value and dynamically updated as the training goes on. By rewarding the agent to explore motion-appearance inconsistent states, the curiosity module encourages the agent to learn more efficiently, leading to better performance and fewer learning steps.", + "bbox": [ + 114, + 419, + 473, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Overall Training Objective", + "text_level": 1, + "bbox": [ + 116, + 578, + 338, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By capitalizing on interactive motion-appearance understanding, Simoun learns from vision to control in an end-to-end manner by optimizing the following training objective:", + "bbox": [ + 114, + 599, + 473, + 641 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\underbrace {\\mathcal {L} _ {t r a n}} _ {M o t i o n} + \\underbrace {\\mathcal {L} _ {c o n}} _ {A p p e a r a n c e} + \\underbrace {\\mathcal {L} _ {r e}} _ {S t a t e} + \\underbrace {\\mathcal {L} _ {Q} + \\mathcal {L} _ {\\pi}} _ {R L}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 657, + 472, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the objective jointly considers the learning of motion and appearance features, as well as getting efficient state representation for high-performance RL policy learning. The detailed inference procedure is presented in Alg. 1.", + "bbox": [ + 114, + 690, + 473, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. An Information Bias Analysis of Simoun", + "text_level": 1, + "bbox": [ + 116, + 757, + 444, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As we demonstrated in Fig. 1, existing observation encoding paradigms model motion and appearance information either heuristically without special design or using preliminary techniques (such as taking latent difference). In contrast, Simoun deliberately models both information and emphasizes their interactions. In this section, we analyze", + "bbox": [ + 114, + 780, + 473, + 863 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f7a4511aa94bf9946b7464cbb16926a6e095b96eb6b75af418a55486e34c5051.jpg", + "image_caption": [ + "Figure 5. Feature component analysis of different models. M-Path and A-path represent the motion and appearance paths of Simoun, respectively. Strong information bias can be observed from the baseline models and Simoun notably alleviates the bias and decreases meaningless residual units." + ], + "image_footnote": [], + "bbox": [ + 500, + 126, + 854, + 279 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the representations learned by Simoun and other observation encoding paradigms from an information bias perspective to gain an in-depth understanding of them.", + "bbox": [ + 498, + 369, + 854, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, we leverage a recent approach [18] that can quantify static and dynamic information learned by any spatial-temporal model. The approach estimates the amount of static vs. dynamic bias based on the mutual information between sampled input sequence pairs. It then calculates the percentage of units (channel dimensions) of the model feature layer that encodes several pre-defined information factors (static, dynamic, joint, and residual). The quantifying results of Simoun and other models are illustrated in Fig. 5. It is immediately evident that previous methods tend to encode the static factor more than the dynamic, indicating a strong bias toward appearance information. It is also clear that the two paths of Simoun learn corresponding static and dynamic factors as expected, with the motion path having more dynamic units than the appearance path. By fusing the two paths together, Simoun enjoys abundant dynamic and static information, meanwhile having minimal residual units that do not involve any dynamic or static factor. As will be shown in the experiment section, such an abundant and informative representation significantly improves the decision-making process.", + "bbox": [ + 498, + 411, + 854, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 714, + 619, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we explore how Simoun can improve vision-based RL in terms of sample efficiency and performance gains. Two benchmarks are used for evaluation: DeepMind Control Suite (DMControl) [38] for continuous control and CARLA [4] for autonomous driving.", + "bbox": [ + 498, + 738, + 854, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental Settings: Simoun is implemented on the basis of the SAC algorithm [10]. For DMControl, to avoid the potential effects of different hyperparameters, we follow the previous training setup of DrQ [44] and choose", + "bbox": [ + 498, + 806, + 854, + 863 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "181", + "bbox": [ + 488, + 888, + 509, + 898 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/df26181c36d88852960075c7c23a6d3c1c839defd084608aafa7ebb89e8be530.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
100K Step ScoresSACDreamerCURLDrQSVEASPRPlayVirtualMLRCCLFOurs
Cartpole, Swingup237±49326±27582±146759±92727±86799±42816±36806±48799±61851±14
Reacher, Easy239±183314±155538±233601±213811±115638±269785±142866±103738±99910±65
Cheetah, Run118±13235±137299±48344±67375±54467±36474±50482±38317±38518±57
Walker, Walk95±19277±12403±24612±164747±65398±165460±173643±114648±110869±45
Finger, Spin230±194341±70767±56901±104859±77868±143915±49907±58994±42872±90
Ball in cup, Catch85±130246±174769±43913±53915±71861±233929±31933±16914±20971±17
Average167.3289.8559.7688.3739.0719.0729.8772.8735.0831.8
500k Step Scores
Cartpole, Swingup330±73762±27841±45868±10865±10870±12865±11872±5869±9876±4
Reacher, Easy307±65793±164929±44942±71944±52925±79942±66957±41941±48979±44
Cheetah, Run85±51570±253518±28660±96682±65716±47719±51674±37588±22693±53
Walker, Walk71±52897±49902±43921±45919±24916±75928±30939±10936±23966±7
Finger, Spin346±95796±183926±45938±103924±93924±132963±40973±31974±6964±19
Ball in cup, Catch162±122879±87959±27963±9960±19963±8967±5964±14961±9970±12
Average216.8782.8845.8882.0882.3920.0897.3896.5878.2908.0
", + "bbox": [ + 120, + 126, + 854, + 308 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d27ae9b6e8da07bb8388bd7aed9f94604db9edcb865ad6d9df0354b8ad910fad.jpg", + "table_caption": [ + "Table 1. Comparison with state-of-the-art methods on DMControl benchmark. The best results are shown in bold." + ], + "table_footnote": [], + "table_body": "
MeasurementsSACCURLDrQFlareDeepMDPOurs
Episode reward ↑121 ±26.1134 ±15.1154 ±21.5132 ±24.7170 ±36.1281 ±30.4
Distance (m) ↑74 ±17.4128 ±32.595 ±27.290 ±14.6132 ±20.4207 ±15.5
Crash intensity ↓3930 ±80.33050 ±100.32419 ±72.32668 ±95.72136 ±69.31813 ±57.8
Average steer ↓17.52% ±0.021%16.60% ±0.025%15.79% ±0.018%11.48% ±0.022%10.22% ±0.015%14.84% ±0.012%
Average brake ↓1.81% ±0.013%2.94% ±0.021%1.70% ±0.039%1.52% ±0.014%1.65% ±0.007%2.14% ±0.018%
", + "bbox": [ + 122, + 338, + 849, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison with state-of-the-art methods on CARLA benchmark. $\\uparrow$ indicates larger is better and $\\downarrow$ means the opposite.", + "bbox": [ + 142, + 421, + 828, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "six commonly adopted tasks: Walker-walk, Finger-spin, Cartpole-swingup, Reacher-easy, Cheetah-run and Ball in cup-catch. To evaluate sample efficiency, the performance at 100k and 500k environment steps are reported during the training stage. For CARLA, we mostly follow the setup of DBC [47], where the goal is to travel as far as possible on Highway 8 of Town 4 in 1000 time steps without any collisions with 20 moving cars. For observation acquisition, we horizontally concatenate the images from three cameras on the roof of vehicles to get $84 \\times 252$ images. Random convolution [21] is adopted for data augmentation in Eq. 3. All experiments are trained across 5 random seeds to report the mean and standard deviation of the rewards. More details can be found in the Supplementary Material.", + "bbox": [ + 114, + 457, + 473, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Methods Compared: We extensively compare Simoun with a variety of methods including the SAC [10] baseline, explicit motion modeling approach (Flare [33]), auxiliary loss-based approaches (CURL [20], MLR [46]), data augmentation approaches (DrQ [44], SVEA [15]), dynamic modeling based approaches (Dreamer [11], SPR [30], PlayVirtual [45], DeepMDP [8]), and curiosity based approach (CCLF [37]).", + "bbox": [ + 114, + 648, + 473, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Performance Comparisons", + "text_level": 1, + "bbox": [ + 116, + 771, + 336, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on DMControl Table 1 presents the experimental results on the DMControl benchmark. It can be observed that Simoun achieves considerable performance gains compared to other state-of-the-art methods. In particular, at $100\\mathrm{k}$ steps, significant performance improvement can al", + "bbox": [ + 114, + 792, + 473, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ready be reached by Simoun, which indicates improved sample efficiency.", + "bbox": [ + 500, + 457, + 856, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on CARLA The results of the CARLA benchmark are reported in Table 2. It is clear that Simoun outperforms all other methods on the episode reward. Additionally, the average driving distance is farther than other methods by a large margin and the average collision intensity is also smaller. Although the driving smoothness of Simoun is slightly decreased due to increased steer and brake, this small cost has led to considerable overall reward gain to break through the current status quo.", + "bbox": [ + 498, + 484, + 856, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 619, + 638, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of the Dual-path Design To demonstrate the effectiveness of the dual-path design in Simoun, we compare it with three single-path methods (depicted in Fig. 1): individual frame encoding, stacked frames encoding, latent flow encoding, and our dual-path encoding. For a fair comparison, the motion and appearance losses of Simoun are adopted for all four models to eliminate the affection of loss difference, and the interactive and curiosity modules are also removed from the dual-path model. The results are shown in Fig. 6 (left). Several observations can be made: 1) The low performance of individual frame encoding (black line) indicates the importance of modeling motion information. 2) By considering the motion across frames, stacked frame encoding (green line) performs much better than individual frame encoding. 3) The latent flow encoding (blue line) improves over stacked frame encoding", + "bbox": [ + 498, + 643, + 857, + 864 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "182", + "bbox": [ + 488, + 888, + 509, + 898 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dae08905ec192cb390fd34f8dd63b02476edbfc943e7da8d50c213bf4b24f7ae.jpg", + "image_caption": [ + "Figure 6. Experiment results on CARLA benchmark. Left: Performance on different encoding paradigms. Middle: Ablation study of each component of Simoun. Right: Comparison with models having more parameters." + ], + "image_footnote": [], + "bbox": [ + 127, + 125, + 367, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/de0bff32cca3531b6dccaf71c9f90a2fd7330246cac5cd0d8595ea3c0e0c6b7c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 126, + 604, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2faa47d456470448a415fffdbb2e78cd8757516ec1271af109a5ceb4629d802b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 126, + 843, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/38e26d4d4f00d87a6dd7210bb57d54996e55d596a1de5a983744d555af1641d1.jpg", + "image_caption": [ + "Figure 7. Visualizations of motion-appearance structure masks." + ], + "image_footnote": [], + "bbox": [ + 124, + 309, + 460, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with explicit motion modeling. However, there is limited room for further improvement due to the preliminary technique used for motion extraction. 4) The proposed dualpath model (red line) remarkably outperforms the other three methods by modeling motion and appearance explicitly. Interestingly, the results in Fig. 6 (left) echo perfectly with the information bias degrees illustrated and discussed in Fig. 5 and Sec. 4, which indicates potentially deeper connections between motion-appearance information bias and the performance of decision-making.", + "bbox": [ + 114, + 462, + 472, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Each Components in Simoun To investigate the contribution of each component in Simoun, we first evaluate the performance of each individual path, then test the dual-path model by gradually adding the structural interactive module and the consistency-guided curiosity module. It can be found in Fig. 6 (middle) that the individual path gives relatively low performance when trained separately, with the motion path performing better than the appearance path. When adopting the dual-path model, both the structural interactive module and the consistency-guided curiosity module can further improve performance, which demonstrates their effectiveness. To better understand which visual clues does Simoun concentrate, we visualize the motion-appearance structure masks $(\\mathbf{X}^m$ and $\\mathbf{X}^a$ in Eq. 5) of the two paths. As can be observed from Fig. 7, the motion path tends to focus on the moving trajectory of other vehicles (positions where the vehicles passed by), while the appearance path focuses strongly on spatial positions where the other vehicles exist.", + "bbox": [ + 114, + 602, + 473, + 863 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/98f7364f51785f3792cc7fd6ef0a01c0f0501203727accabd61b31b666a9c0b9.jpg", + "image_caption": [ + "Figure 8. Evaluation of the generalization ability on four different environment domains on DMControl benchmark." + ], + "image_footnote": [], + "bbox": [ + 532, + 309, + 823, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Further Discussions", + "text_level": 1, + "bbox": [ + 500, + 496, + 671, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Does the performance gain of the dual-path model come from the increased network parameters? To answer this question, we compare the dual-path model with a double-channeled stacked frame model, which has nearly $2 \\times$ more parameters. From Fig. 6 (right) we can observe that increasing model parameters indeed improves performance. However, the dual-path model still outperforms the double-channeled stacked frames model using only half of its parameters. This proves the effectiveness of the dualpath model mainly comes from its motion-appearance modeling paradigm, rather than increased network capacity.", + "bbox": [ + 498, + 526, + 856, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Does Simoun improves domain generalization? To evaluate the domain generalization ability of Simoun, we select the \"catch\" task on the \"ball in cup\" scenario as the source domain and test Simoun on DMControl Generalization Benchmark [16] with four different environment domains (color easy, color hard, video easy, and video hard). Fig. 8 shows that Simoun performs on par with other methods on color-shifted domains. However, the performance is much better on video-shifted domains, where the background is also moving. This is attributed to the specialized modeling of motion information by Simoun, which drives the agent to pay more attention to reward-related motions rather than the irrelevant dynamic background.", + "bbox": [ + 498, + 684, + 856, + 863 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "183", + "bbox": [ + 488, + 888, + 509, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 117, + 128, + 225, + 141 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We have proposed Simoun, a unified framework for vision-based RL with a dual-path network for motion and appearance understanding. The design of Simoun demonstrates the effectiveness of motion-appearance structural interaction, and further shows the benefits of consistency-guided intrinsic curiosity. Empirical results suggest that the proposed method has advantages in terms of sample efficiency, performance gains, and generalization ability. By analyzing Simoun from an information bias perspective, we build a connection between motion-appearance information bias and vision-based RL performance. We hope this connection can further inspire more efficient model designs for vision-based RL tasks.", + "bbox": [ + 116, + 143, + 473, + 321 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 117, + 334, + 279, + 348 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The study was funded by the Key-Area Research and Development Program of Guangdong Province with contract No. 2020B0101380001, as well as the National Natural Science Foundation of China under contracts No. 62027804, No. 61825101, No. 62088102 and No. 62202010. Computing support was provided by Pengcheng Cloudbrain.", + "bbox": [ + 116, + 349, + 473, + 433 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 117, + 444, + 204, + 457 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Marc Bellemare, Sriram Srinivasan, Georg Ostrovski, Tom Schaul, David Saxton, and Remi Munos. Unifying count-based exploration and intrinsic motivation. In Advances in Neural Information Processing Systems, 2016. 4", + "[2] David Bertoin and Emmanuel Rachelson. Local feature swapping for generalization in reinforcement learning. In International Conference on Learning Representations, 2022. 3", + "[3] Fei Deng, Ingook Jang, and Sungjin Ahn. Dreamerpro: Reconstruction-free model-based reinforcement learning with prototypical representations. In International Conference on Machine Learning, 2022. 3", + "[4] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on Robot Learning, pages 1-16. PMLR, 2017. 2, 7", + "[5] Linxi Fan, Guanzhi Wang, De-An Huang, Zhiding Yu, Li Fei-Fei, Yuke Zhu, and Animashree Anandkumar. Secant: Self-expert cloning for zero-shot generalization of visual policies. In International Conference on Machine Learning, 2021. 2, 6", + "[6] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition. In Computer Vision and Pattern Recognition, 2016. 3", + "[7] Xiang Fu, Ge Yang, Pulkit Agrawal, and Tommi Jaakkola. Learning task informed abstractions. In International Conference on Machine Learning, 2021. 3", + "[8] Carles Gelada, Saurabh Kumar, Jacob Buckman, Ofir Nachum, and Marc G Bellemare. Deepmdp: Learning continuous latent space models for representation learning. In International Conference on Machine Learning, 2019. 8" + ], + "bbox": [ + 125, + 467, + 473, + 872 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. In International Conference on Machine Learning, 2018. 6", + "[10] Tuomas Haarnoja, Aurick Zhou, Kristian Hartikainen, George Tucker, Sehoon Ha, Jie Tan, Vikash Kumar, Henry Zhu, Abhishek Gupta, Pieter Abbeel, et al. Soft actor-critic algorithms and applications. arXiv preprint arXiv:1812.05905, 2018.6,7,8", + "[11] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603, 2019. 3, 8", + "[12] Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. In International Conference on Machine Learning, 2019. 3", + "[13] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. arXiv preprint arXiv:2010.02193, 2020. 3", + "[14] Nicklas Hansen, Rishabh Jangir, Yu Sun, Guillem Alenyà, Pieter Abbeel, Alexei A Efros, Lerrel Pinto, and Xiaolong Wang. Self-supervised policy adaptation during deployment. In nternational Conference on Learning Representations, 2021. 3", + "[15] Nicklas Hansen, Hao Su, and Xiaolong Wang. Stabilizing deep q-learning with convnets and vision transformers under data augmentation. In Advances in Neural Information Processing Systems, 2021. 8", + "[16] Nicklas Hansen and Xiaolong Wang. Generalization in reinforcement learning by soft data augmentation. In International Conference on Robotics and Automation, 2021. 9", + "[17] Yangru Huang, Peixi Peng, Yifan Zhao, Guangyao Chen, and Yonghong Tian. Spectrum random masking for generalization in image-based reinforcement learning. In Advances in Neural Information Processing Systems. 3", + "[18] Matthew Kowal, Mennatullah Siam, Md Amirul Islam, Neil DB Bruce, Richard P Wildes, and Konstantinos G Derpanis. A deeper dive into what deep spatiotemporal networks encode: Quantifying static vs. dynamic information. In Computer Vision and Pattern Recognition, 2022. 7", + "[19] Misha Laskin, Kimin Lee, Adam Stooke, Lerrel Pinto, Pieter Abbeel, and Aravind Srinivas. Reinforcement learning with augmented data. In Advances in Neural Information Processing Systems, 2020. 2", + "[20] Michael Laskin, Aravind Srinivas, and Pieter Abbeel. Curl: Contrastive unsupervised representations for reinforcement learning. In International Conference on Machine Learning, 2020. 2, 3, 5, 8", + "[21] Kimin Lee, Kibok Lee, Jinwoo Shin, and Honglak Lee. Network randomization: A simple technique for generalization in deep reinforcement learning. In International Conference on Machine Learning, 2020. 3, 8", + "[22] Kevin Li, Abhishek Gupta, Ashwin Reddy, Vitchyr H Pong, Aurick Zhou, Justin Yu, and Sergey Levine. Mural: Meta-learning uncertainty-aware rewards for outcome-driven reinforcement learning. In International Conference on Machine Learning, 2021. 4" + ], + "bbox": [ + 501, + 128, + 854, + 872 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "184", + "bbox": [ + 488, + 888, + 509, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Xiang Li, Jinghuan Shang, Srijan Das, and Michael S Ryoo. Does self-supervised learning really improve reinforcement learning from pixels? In Advances in Neural Information Processing Systems, 2022. 3", + "[24] Sicong Liu, Xi Sheryl Zhang, Yushuo Li, Yifan Zhang, and Jian Cheng. On the data-efficiency with contrastive image transformation in reinforcement learning. In International Conference on Learning Representations, 2023. 3", + "[25] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602, 2013. 2", + "[26] Thanh Nguyen, Tung M Luu, Thang Vu, and Chang D Yoo. Sample-efficient reinforcement learning representation learning with curiosity contrastive forward dynamics model. In International Conference on Intelligent Robots and Systems, 2021. 4", + "[27] Minting Pan, Xiangming Zhu, Yunbo Wang, and Xiaokang Yang. Iso-dream: Isolating and leveraging noncontrollable visual dynamics in world models. In Advances in Neural Information Processing Systems, 2022. 3", + "[28] Deepak Pathak, Pulkit Agrawal, Alexei A Efros, and Trevor Darrell. Curiosity-driven exploration by self-supervised prediction. In International Conference on Machine Learning, 2017. 4", + "[29] Roberta Raileanu, Maxwell Goldstein, Denis Yarats, Ilya Kostrikov, and Rob Fergus. Automatic data augmentation for generalization in reinforcement learning. In Advances in Neural Information Processing Systems, 2021. 2", + "[30] Max Schwarzer, Ankesh Anand, Rishab Goel, R Devon Hjelm, Aaron Courville, and Philip Bachman. Data-efficient reinforcement learning with self-predictive representations. arXiv preprint arXiv:2007.05929, 2020.8", + "[31] Ramanan Sekar, Oleh Rybkin, Kostas Daniilidis, Pieter Abbeel, Danijar Hafner, and Deepak Pathak. Planning to explore via self-supervised world models. In International Conference on Machine Learning, 2020. 3", + "[32] Younggyo Seo, Lili Chen, Jinwoo Shin, Honglak Lee, Pieter Abbeel, and Kimin Lee. State entropy maximization with random encoders for efficient exploration. In International Conference on Machine Learning, 2021. 4", + "[33] Wenling Shang, Xiaofei Wang, Aravind Srinivas, Aravind Rajeswaran, Yang Gao, Pieter Abbeel, and Misha Laskin. Reinforcement learning with latent flow. In Advances in Neural Information Processing Systems, 2021. 3, 6, 8", + "[34] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Computer Vision and Pattern Recognition, 2019. 3", + "[35] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In Advances in Neural Information Processing Systems, 2014. 3", + "[36] Adam Stooke, Kimin Lee, Pieter Abbeel, and Michael Laskin. Decoupling representation learning from reinforcement learning. In International Conference on Machine Learning, pages 9870-9879. PMLR, 2021. 3", + "[37] Chenyu Sun, Hangwei Qian, and Chunyan Miao. Cclf: A contrastive-curiosity-driven learning framework for" + ], + "bbox": [ + 117, + 128, + 473, + 864 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sample-efficient reinforcement learning. arXiv preprint arXiv:2205.00943, 2022. 4, 8", + "[38] Yuval Tassa, Yotam Doron, Alistair Muldal, Tom Erez, Yazhe Li, Diego de Las Casas, David Budden, Abbas Abdelmaleki, Josh Merel, Andrew Lefrancq, et al. Deepmind control suite. arXiv preprint arXiv:1801.00690, 2018. 7", + "[39] Anirudh Thatipelli, Sanath Narayan, Salman Khan, Rao Muhammad Anwer, Fahad Shahbaz Khan, and Bernard Ghanem. Spatio-temporal relation modeling for few-shot action recognition. In Computer Vision and Pattern Recognition, 2022. 3", + "[40] Oriol Vinyals, Timo Ewalds, Sergey Bartunov, Petko Georgiev, Alexander Sasha Vezhnevets, Michelle Yeo, Alireza Makhzani, Heinrich Kuttler, John Agapiou, Julian Schrittwieser, et al. Starcraft ii: A new challenge for reinforcement learning. arXiv preprint arXiv:1708.04782, 2017. 2", + "[41] Kaixin Wang, Bingyi Kang, Jie Shao, and Jiashi Feng. Improving generalization in reinforcement learning with mixture regularization. In Advances in Neural Information Processing Systems, 2020. 3", + "[42] Xudong Wang, Long Lian, and Stella X Yu. Unsupervised visual attention and invariance for reinforcement learning. In Computer Vision and Pattern Recognition, 2021. 2", + "[43] Wangmeng Xiang, Chao Li, Biao Wang, Xihan Wei, XianSheng Hua, and Lei Zhang. Spatiotemporal self-attention modeling with temporal patch shift for action recognition. In European Conference on Computer Vision, 2022. 3", + "[44] Denis Yarats, Ilya Kostrikov, and Rob Fergus. Image augmentation is all you need: Regularizing deep reinforcement learning from pixels. In International Conference on Learning Representations, 2020. 5, 7, 8", + "[45] Tao Yu, Cuiling Lan, Wenjun Zeng, Mingxiao Feng, Zhizheng Zhang, and Zhibo Chen. Playvirtual: Augmenting cycle-consistent virtual trajectories for reinforcement learning. In Advances in Neural Information Processing Systems, 2021. 8", + "[46] Tao Yu, Zhizheng Zhang, Cuiling Lan, Zhibo Chen, and Yan Lu. Mask-based latent reconstruction for reinforcement learning. In Advances in Neural Information Processing Systems. 8", + "[47] Amy Zhang, Rowan McAllister, Roberto Calandra, Yarin Gal, and Sergey Levine. Learning invariant representations for reinforcement learning without reconstruction. arXiv preprint arXiv:2006.10742, 2020. 5, 6, 8", + "[48] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In International Conference on Machine Learning, 2019. 6", + "[49] Jinhua Zhu, Yingce Xia, Lijun Wu, Jiajun Deng, Wengang Zhou, Tao Qin, Tie-Yan Liu, and Houqiang Li. Masked contrastive representation learning for reinforcement learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 3", + "[50] Yuke Zhu, Josiah Wong, Ajay Mandlekar, and Roberto Martín-Martín. robosuite: A modular simulation framework and benchmark for robot learning. arXiv preprint arXiv:2009.12293, 2020. 2" + ], + "bbox": [ + 503, + 130, + 854, + 858 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "185", + "bbox": [ + 488, + 888, + 509, + 898 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Simoun_ Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning/17dc049b-d1dc-4e17-921e-1317f662d966_model.json b/2023/Simoun_ Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning/17dc049b-d1dc-4e17-921e-1317f662d966_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4b7256363e37be821ddd8bf505d2fb588917d030 --- /dev/null +++ b/2023/Simoun_ Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning/17dc049b-d1dc-4e17-921e-1317f662d966_model.json @@ -0,0 +1,2189 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.14, + 0.046, + 0.209, + 0.082 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.261, + 0.044, + 0.782, + 0.058 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.338, + 0.058, + 0.704, + 0.071 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.31, + 0.071, + 0.733, + 0.085 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.159, + 0.165, + 0.816, + 0.206 + ], + "angle": 0, + "content": "Simoun: Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning" + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.23, + 0.838, + 0.248 + ], + "angle": 0, + "content": "Yangru Huang\\(^{1}\\), Peixi Peng\\(^{2,4*}\\), Yifan Zhao\\(^{1}\\), Yunpeng Zhai\\(^{1}\\), Haoran Xu\\(^{3,4}\\), Yonghong Tian\\(^{1,2,4*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.248, + 0.667, + 0.264 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.264, + 0.735, + 0.279 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Electronic and Computer Engineering, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.279, + 0.83, + 0.296 + ], + "angle": 0, + "content": "\\(^{3}\\)School of Intelligent Systems Engineering, Sun Yat-sen University \\(^{4}\\)Peng Cheng Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.298, + 0.854, + 0.312 + ], + "angle": 0, + "content": "yrhuang@stu.pku.edu.cn, {pxpeng, zhaoyf, ypzhai, yhtian}@pku.edu.cn, xuhr9@mail2.sysu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.26, + 0.343, + 0.331, + 0.357 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.373, + 0.474, + 0.688 + ], + "angle": 0, + "content": "Efficient motion and appearance modeling are critical for vision-based Reinforcement Learning (RL). However, existing methods struggle to reconcile motion and appearance information within the state representations learned from a single observation encoder. To address the problem, we present Synergizing Interactive Motion-appearance Understanding (Simoun), a unified framework for vision-based RL. Given consecutive observation frames, Simoun deliberately and interactively learns both motion and appearance features through a dual-path network architecture. The learning process collaborates with a structural interactive module, which explores the latent motion-appearance structures from the two network paths to leverage their complementarity. To promote sample efficiency, we further design a consistency-guided curiosity module to encourage the exploration of under-learned observations. During training, the curiosity module provides intrinsic rewards according to the consistency of environmental temporal dynamics, which are deduced from both motion and appearance network paths. Experiments conducted on DeepMind control suite and CARLA automatic driving benchmarks demonstrate the effectiveness of Simoun, where it performs favorably against state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.118, + 0.717, + 0.236, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.741, + 0.474, + 0.839 + ], + "angle": 0, + "content": "Reinforcement learning (RL) from visual signals has achieved great success in recent years. Compared with learning from hand-crafted states, vision-based RL eliminates the arduous task of designing states with manual feature engineering. Therefore, it is beneficial for a variety of tasks such as video game playing [25, 40], robot manipulation [50], and autonomous navigation [4]. However, one of" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.342, + 0.848, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.576, + 0.857, + 0.64 + ], + "angle": 0, + "content": "Figure 1. Illustration of different observation encoding schemes. (a) Encoding individual frames with a shared encoder. (b) Encoding stacked multiple frames. (c) Encoding individual frames and latent vector differences (latent flows). (d) Encoding motion and appearance independently and interactively." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.668, + 0.856, + 0.751 + ], + "angle": 0, + "content": "the major challenges of vision-based RL lies in its high dimensional observation, which is less interpretable and leads to low rewards and inefficient sampling [19, 20]. As a result, robust visual understanding is crucial to bridge the gap between vision and state-based RL in terms of performance and sample efficiency." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.753, + 0.856, + 0.864 + ], + "angle": 0, + "content": "To make observation comprehensible for agents, recent works have recognized the importance of learning high-quality visual features [20, 5, 29, 49, 42]. From the perspective of decision-making, two kinds of features are essential for vision-based RL: the motion features, which closely relate to the actions performed by the agents, and the appearance features, which describe the contemporary environmental states. Despite their importance, few works" + }, + { + "type": "page_footnote", + "bbox": [ + 0.134, + 0.851, + 0.251, + 0.863 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.89, + 0.511, + 0.9 + ], + "angle": 0, + "content": "176" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.116, + 0.13, + 0.474, + 0.335 + ], + "angle": 0, + "content": "have tried to model these two kinds of features explicitly. Most methods tend to encode single or stacked multiple frames via a convolutional encoder, as shown in Fig. 1(a) and Fig. 1(b). The former neglects motion feature modeling, while the latter tightly entangles the motion and appearance information into a single network, causing information bias and optimization difficulty. To alleviate these issues, Shang et al. [33] propose to encode only appearance features in single frames and impose motion dynamics by latent vector differences (termed latent flow, see Fig. 1(c)). In addition, several world models learned directly from input images are also developed [11, 31, 3]. Nevertheless, the motion features of these methods are still computed from spatial latent space, which may lead to sub-optimal representation learning in complicated environments." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.339, + 0.475, + 0.655 + ], + "angle": 0, + "content": "To overcome the limitations described above, we present Synergizing Interactive Motion-appearance Understanding (Simoun) for vision-based RL. The design principle of Simoun is to explicitly learn both motion and appearance features in the early encoding stage and then interactively fuse them at the late stage for accurate decision-making. As shown in Fig. 1(d), given consecutive environment frames, Simoun learns motion and appearance information by separate network paths. The motion path (colored in blue) explicitly models motion clues (such as the speed and direction of cars) from the residual frame of multiple neighboring input frames. The appearance path (colored in orange) models the environmental spatial structures and focuses on identifying patterns and objects (such as cars and traffic lights) from every single frame. Additionally, a structural interactive module further extracts latent motion-appearance structures reflected by the correlations of the dual-path features. It then modulates both paths with the computed structure masks. In this way, each path can take complementary information from the other during learning, and agents are able to better understand the spatial and temporal context of the environment. Finally, the latent vectors from the two paths are fused for decision-making." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.658, + 0.474, + 0.864 + ], + "angle": 0, + "content": "Although the dual-path design of Simoun promotes observation interpretability, low sample efficiency still exists due to finite data and sparse rewards. To alleviate this issue, a consistency-guided curiosity module is further designed. The idea is to adapt the learning process by concentrating more on under-learned observations, which can be deduced from the consistency of the motion and appearance paths. Intuitively, both the motion path and the appearance path describe the same observations. Hence, the dynamic information inferred from the following two sources should remain consistent: 1) latent motion vectors learned directly from the motion path and 2) differences between the appearance path latent vectors over multiple neighboring frames. If the opposite were true, then it indicates premature motion-appearance understanding, thus more explo" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.13, + 0.857, + 0.198 + ], + "angle": 0, + "content": "ration should be added. In this way, we build a strong correlation between reward discovery and state novelty. During training, the consistency-guided curiosity module provides intrinsic rewards in addition to extrinsic rewards from the environment, resulting in more efficient exploration." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.199, + 0.857, + 0.253 + ], + "angle": 0, + "content": "Experiments on both CARLA and OpenAI DMControl environments show that the proposed method performs favorably against state-of-the-arts. Overall, the contributions of this paper are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.254, + 0.855, + 0.293 + ], + "angle": 0, + "content": "(1) We propose Simoun, a novel dual-path learning framework that explicitly and interactively learns both motion and appearance information from observations." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.294, + 0.855, + 0.335 + ], + "angle": 0, + "content": "(2) We design a structural interactive module to fully explore the complementarity of the two paths in Simoun and thereby further enhance visual understanding." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.335, + 0.855, + 0.39 + ], + "angle": 0, + "content": "(3) We devise a consistency-guided curiosity module to encourage the exploration of under-learned observations. The proposed module effectively increases sample efficiency by providing intrinsic rewards for the agents." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.254, + 0.855, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.402, + 0.636, + 0.417 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.425, + 0.856, + 0.644 + ], + "angle": 0, + "content": "Vision-based Reinforcement Learning To improve the performance and sample efficiency of vision-based RL, existing works can be roughly divided into three groups: 1) designing auxiliary loss/learning tasks [20, 14, 36, 49, 23], 2) employing various data augmentation techniques [21, 41, 2, 17, 24], and 3) modeling environment dynamics [12, 11, 13, 7, 27]. However, most existing works utilize a single-path network with multi-frame inputs, in which the motion and appearance information is tightly entangled without explicit separation. One exception is Flare [33], which also utilizes a single-path network to encode each frame individually and models motion information explicitly by taking latent vector differences. Although Flare achieves improved performance, its motion information is still computed from single-frame appearance features, causing insufficient temporal information extraction." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.644, + 0.857, + 0.863 + ], + "angle": 0, + "content": "Dual-path Networks for Visual Modeling There is a rich literature of works on visual modeling with dual-path networks. One of the earliest works is the two-stream CNNs for action recognition [35], which utilizes a spatial stream with single-frame input and a temporal stream taking multiframe optical flows. Thereafter, the concept of dual-path networks is heavily explored with various fusion strategies on different tasks [34, 6, 39, 43]. Our approach differs from existing ones in terms of task, architecture and learning mechanism. Compare with other visual tasks, vision-based RL needs to extract fine-grained motion details across different time steps, which poses a great challenge to RL models. To the best of our knowledge, Simoun is one of the first vision-based RL methods which aim to explicitly learn motion and appearance features by a dual-path architecture. Meanwhile, its structural interactive module is" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.89, + 0.511, + 0.901 + ], + "angle": 0, + "content": "177" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.136, + 0.13, + 0.837, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.382, + 0.859, + 0.422 + ], + "angle": 0, + "content": "Figure 2. The proposed dual-path learning framework Simoun. It introduces two modules: 1) a structural interactive module that leverages the complementarity between the two network paths by extracting latent motion-appearance structures, and 2) a consistency-guided curiosity module that provides intrinsic rewards according to the dynamic consistency between the motion and appearance features." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.442, + 0.475, + 0.51 + ], + "angle": 0, + "content": "also deliberately designed to extract motion and appearance clues residing in consecutive observation frames. By learning with RL-oriented objectives and curiosity-driven strategies, Simoun successfully imports the idea of dual-path networks into robust visual control." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.511, + 0.474, + 0.758 + ], + "angle": 0, + "content": "Intrinsic Reward Exploration One of the key elements of RL is the reward function, which aims to quantify the \"goodness\" of the agent's decisions. However, the problem is that designing dense, well-defined extrinsic reward functions is difficult and unscalable. One possible solution is to introduce an intrinsic reward function, which is calculated by the agents. Existing intrinsic reward exploration methods mainly focus on counting or predicting state novelty [1], prediction error [28], uncertainty [22], or environmental dynamics [32]. These methods are typically designed for the general state-based instead of vision-based RL. The works most relevant to us are CCFDM [26] and CCLF [37], which also formulate intrinsic rewards for visual-based RL. CCFDM utilizes forward dynamics and CCLF is based on a contrastive term. Differently, the consistency-guided curiosity module in Simoun exploits the dynamic consistency extracted from both motion and appearance features to achieve reliable intrinsic reward estimation." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.771, + 0.241, + 0.788 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.794, + 0.475, + 0.864 + ], + "angle": 0, + "content": "We start by formalizing the task of vision-based RL in Sec. 3.1 and then delineate Simoun in detail. The general idea of Simoun is to 1) explicitly model the motion/appearance dynamics (and their correlations) of the agent's operating ambiance and 2) capitalize on the learned" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.858, + 0.581 + ], + "angle": 0, + "content": "dynamics to achieve efficient exploration. As shown in Fig. 2, our framework consists of two parts. First, given inputs in the form of consecutive frames, it interactively models the motion and appearance features using a dualpath architecture with a set of targeted objectives and a structural interactive module (Sec. 3.2). Then the decision-making strategy is learned with adaptive curiosity assignment steered by the proposed consistency-guided curiosity module (Sec. 3.3). The overall learning objective and inference process of Simoun are finally summarized in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.586, + 0.681, + 0.599 + ], + "angle": 0, + "content": "3.1. Problem formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.857, + 0.749 + ], + "angle": 0, + "content": "Vision-based RL can be formulated as a Partially Observable Markov Decision Process (POMDP) \\(\\mathcal{M} = < \\mathcal{O},\\mathcal{A},\\mathcal{P},\\mathcal{R},\\gamma >\\) where \\(\\mathcal{O}\\) denotes the observation space containing pixel frames \\(\\mathbf{o}_t\\) at different time step \\(t\\) and \\(\\mathcal{A}\\) denotes the action space. At each \\(t\\) the agent chooses an action \\(a_{t}\\in \\mathcal{A}\\). \\(\\mathcal{P}(\\mathbf{o}_{t + 1}|\\mathbf{o}_t,a_t)\\) is the observation transition, \\(\\mathcal{R}(\\mathbf{o}_t,a_t)\\) is the reward function, and \\(\\gamma \\in [0,1)\\) is the discount factor. The goal is to identify an optimal policy \\(\\pi\\) that maximizes the expected cumulative reward based on the visual observations rather than the complete environment state:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.751, + 0.855, + 0.779 + ], + "angle": 0, + "content": "\\[\nJ (\\pi) = \\sum_ {t} \\mathbb {E} _ {\\left(\\mathbf {o} _ {t}, a _ {t}\\right) \\sim \\pi} \\left[ \\mathcal {R} \\left(\\mathbf {o} _ {t}, a _ {t}\\right) \\right]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.855, + 0.809 + ], + "angle": 0, + "content": "During training, \\(\\pi\\) is used to interact with the environment, and the related data are stored in a replay buffer \\(\\mathcal{B}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.814, + 0.757, + 0.83 + ], + "angle": 0, + "content": "3.2. Dual-path Interactive Modeling" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.836, + 0.856, + 0.864 + ], + "angle": 0, + "content": "In vision-based RL, motion and appearance clues are both vital for the agent to perform accurate decision" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.89, + 0.511, + 0.9 + ], + "angle": 0, + "content": "178" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.116, + 0.13, + 0.474, + 0.212 + ], + "angle": 0, + "content": "making. Simoun adopts a dual-path architecture to interactively learn motion and appearance features. Specifically, it consists of three components: the motion path, the appearance path, and the structural interactive module that serves as a switchboard between the motion and appearance structures learned from the two paths." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.215, + 0.474, + 0.434 + ], + "angle": 0, + "content": "Motion Path Motion information is crucial for the visual system to understand the dynamics of the surrounding environment. The motion path aims to extract the motion-related features (such as the velocity of moving objects) from the changes between consecutive frames. Given a tuple of three adjacent observations \\([\\mathbf{o}_{t - 2},\\mathbf{o}_{t - 1},\\mathbf{o}_t]\\) sampled from replay buffer \\(\\mathcal{B}\\), the input of the motion path is the residual of adjacent frames \\([(0_{t - 1} - 0_{t - 2}),(0_t - 0_{t - 1})]\\), concatenated along channel dimensions. A convolutional encoder \\(\\mathcal{E}^m\\) is then used to learn a lower-dimensional motion representation. Specifically, the encoder includes four convolution layers with \\(3\\times 3\\) kernel size and ReLU nonlinearity. Denote the feature map of the last convolution layer as \\(\\mathbf{F}_t^m\\in \\mathbb{R}^{c\\times h\\times w}\\), a fully connected (FC) layer with layer normalization (LN) is used to reduce the dimension of \\(\\mathbf{F}_t^m\\) to get the motion feature vector \\(\\mathbf{f}_t^m\\)." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.437, + 0.474, + 0.532 + ], + "angle": 0, + "content": "To further impel the motion path to catch abundant environment temporal structures, a motion-aware constraint is deliberately designed. Specifically, given \\(\\mathbf{f}_t^m\\) and corresponding action \\(a_{t}\\) at timestep \\(t\\), an action-conditioned two-FC-layer transition model \\(\\mathcal{G}\\) is used to obtain a motion-action joint representation of the current timestep. A latent-space transition loss is then defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.535, + 0.473, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t r a n} = \\left\\| \\mathcal {G} \\left(\\mathbf {f} _ {t} ^ {m}, u _ {t}\\right) - \\mathbf {f} _ {t + 1} ^ {m} \\right\\| _ {2} ^ {2}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.553, + 0.473, + 0.595 + ], + "angle": 0, + "content": "which encourages \\(\\mathbf{f}_t^m\\) to model the motion trends over time by predicting the future temporal dynamics \\(\\mathbf{f}_{t + 1}^{m}\\) of the next time step." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.597, + 0.474, + 0.72 + ], + "angle": 0, + "content": "Appearance Path The appearance path is designed to extract the visual appearance representations of the object and scene from individual observation frames. The input of the appearance path is a single frame \\(\\mathbf{o}_t\\) from the replay buffer \\(\\mathcal{B}\\). Similar to the motion path, the appearance path also adopts a four-layer encoder \\(\\mathcal{E}^a\\) with the same architecture (except for the number of input channels of the first layer) to get feature map \\(\\mathbf{F}_t^a\\in \\mathbb{R}^{c\\times h\\times w}\\), followed by FC and LN layers to get the appearance representation \\(\\mathbf{f}_t^a\\)." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.723, + 0.474, + 0.833 + ], + "angle": 0, + "content": "To explicitly make appearance representations discriminative to different scenes, we take inspiration from CURL [20] and adopt an unsupervised contrastive loss between similar and dissimilar sample pairs. Given observation frame \\(\\mathbf{o}_i\\), we consider \\(\\tilde{\\mathbf{o}}_i\\) (obtained from \\(\\mathbf{o}_i\\) by data augmentation) as the positive sample, while samples coming from different observation frames are regarded as negatives. The contrastive loss is then defined as:" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.13, + 0.849, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.316, + 0.856, + 0.33 + ], + "angle": 0, + "content": "Figure 3. Illustration of the proposed structural interactive module." + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.37, + 0.856, + 0.408 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c o n} = - \\log \\frac {\\exp \\left(\\mathbf {f} _ {i} ^ {a ^ {\\top}} \\tilde {\\mathbf {f}} _ {i} ^ {a}\\right)}{\\exp \\left(\\mathbf {f} _ {i} ^ {a ^ {\\top}} \\tilde {\\mathbf {f}} _ {i} ^ {a}\\right) + \\sum_ {j = 0} ^ {K - 1} \\exp \\left(\\mathbf {f} _ {i} ^ {a ^ {\\top}} \\tilde {\\mathbf {f}} _ {j} ^ {a}\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.41, + 0.856, + 0.507 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_t^a\\) and \\(\\tilde{\\mathbf{f}}_t^a\\) are appearance representations of \\(\\mathbf{o}_t\\) and \\(\\tilde{\\mathbf{o}}_t\\), respectively. \\(j\\) is the sample index in a training batch with \\(K\\) samples. Note that the augmented data is only used for representation learning and is inaccessible during decision-making. This is to avoid unstable training, which is easily caused by the augmented data dominating the evaluation of the Q-target." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.509, + 0.856, + 0.674 + ], + "angle": 0, + "content": "Structural Interactive Module Instead of simply concatenating the features from motion and appearance paths as state representations, interactively communicating the learned motion-appearance structures between both paths can lead to a more robust visual understanding. Intuitively, as shown in Fig. 3(left), given consecutive observation frames, the motion structure can be revealed by inter-frame appearance variation. Meanwhile, the appearance structure can be obtained by spatial pixel correlation. Therefore, the structural interactive module embraces a single and interframe relation discovering mechanism for efficient motion-appearance structure mining and propagation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.676, + 0.856, + 0.771 + ], + "angle": 0, + "content": "Specifically, as shown in Fig. 3(right), the inputs of the structural interactive module are motion feature map \\(\\mathbf{F}_t^m\\) and appearance feature maps \\(\\mathbf{F}_{t - 2}^{a},\\mathbf{F}_{t - 1}^{a}\\) and \\(\\mathbf{F}_t^a\\) .To extract motion structures, an inter-frame attention mechanism is designed. In particular, \\(\\mathbf{F}_{t - 1}^{a}\\) is treated as the query, and \\(\\mathbf{F}_{t - 2}^{a},\\mathbf{F}_{t}^{a}\\) is used as two keys. A inter-frame attention map \\(\\mathbf{X}\\) can be obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.773, + 0.855, + 0.791 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} = \\sigma \\left(\\sigma \\left(\\hat {\\mathbf {F}} _ {t - 2} ^ {a} ^ {\\top} \\hat {\\mathbf {F}} _ {t - 1} ^ {a}\\right) + \\sigma \\left(\\hat {\\mathbf {F}} _ {t} ^ {a \\top} \\hat {\\mathbf {F}} _ {t - 1} ^ {a}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.793, + 0.856, + 0.864 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{F}}_t^a\\in \\mathbb{R}^{c\\times h\\times w}\\) denotes a new feature map generated by feeding \\(\\mathbf{F}_t^a\\) to a convolution layer for reducing the spatial complexity (\\(\\hat{\\mathbf{F}}_{t - 1}^a\\) and \\(\\hat{\\mathbf{F}}_{t - 2}^a\\) can be obtained similarly). \\(\\sigma\\) denotes the Softmax function. Combining the outputs of two initial Softmax functions might not yield values sum" + }, + { + "type": "page_footnote", + "bbox": [ + 0.134, + 0.85, + 0.41, + 0.864 + ], + "angle": 0, + "content": "We choose three here following the practice in [44, 47]." + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.89, + 0.511, + 0.9 + ], + "angle": 0, + "content": "179" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.128, + 0.462, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.321, + 0.454, + 0.334 + ], + "angle": 0, + "content": "Figure 4. The idea of consistency-guided curiosity module." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.355, + 0.473, + 0.452 + ], + "angle": 0, + "content": "ming to 1, so an additional Softmax is used to normalize the final attention scores. The soft indexing of \\(\\hat{\\mathbf{F}}_{t - 1}^{a}\\) by \\(\\hat{\\mathbf{F}}_{t - 2}^{a}\\) or \\(\\hat{\\mathbf{F}}_t^a\\) is performed at the spatial dimensions, resulting in a soft attention map of dimensions \\(hw\\times hw\\). Then a spatial gating operation is applied on \\(\\mathbf{X}\\) using weights calculated from both \\(\\mathbf{F}_t^m\\) and \\(\\mathbf{F}_t^a\\) to consolidate appearance structures and obtain two motion-appearance structure masks:" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.456, + 0.472, + 0.471 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} ^ {m} = \\mathbf {X} \\cdot \\hat {\\mathbf {F}} _ {t} ^ {m}, \\mathbf {X} ^ {a} = \\mathbf {X} \\cdot \\hat {\\mathbf {F}} _ {t} ^ {a}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.473, + 0.473, + 0.527 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{F}}_t^m\\) is a feature map of \\(\\mathbf{F}_t^m\\) obtained with another convolution layer. Finally, \\(\\mathbf{F}_t^m\\) and \\(\\mathbf{F}_t^a\\) are enhanced by modulating with their corresponding motion-appearance structure masks:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.531, + 0.472, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {t} ^ {m} = \\mathbf {F} _ {t} ^ {m} + \\boldsymbol {\\beta} \\cdot \\mathbf {X} ^ {m}, \\mathbf {F} _ {t} ^ {a} = \\mathbf {F} _ {t} ^ {a} + \\boldsymbol {\\beta} \\cdot \\mathbf {X} ^ {a}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.549, + 0.473, + 0.59 + ], + "angle": 0, + "content": "where \\(\\beta\\) is a learnable parameter initialized as zero [48], and the enhanced \\(\\mathbf{F}_t^m\\) and \\(\\mathbf{F}_t^a\\) instead of the original ones are further used to calculate \\(\\mathbf{f}_t^m\\) and \\(\\mathbf{f}_t^a\\)." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.591, + 0.474, + 0.726 + ], + "angle": 0, + "content": "After acquiring the structure-enhanced motion and appearance features, they are finally fused as \\(\\mathbf{f}_t = [\\mathbf{f}_t^m,\\mathbf{f}_t^a ]\\) where \\(\\mathbf{f}_t\\) is used as the final state representation. To better capture reward-related features from the fused state representation, a reward function \\(\\mathcal{R}\\) is further introduced to predict a numerical reward value to each state-action pair. \\(\\mathcal{R}\\) has a similar architecture with the transition model \\(\\mathcal{G}\\) except the output dimension is set to one. A reward loss \\(\\mathcal{L}_{re}\\) is then defined as the mean squared error between the predicted and actual reward:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.729, + 0.473, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e} = \\left(\\mathcal {R} \\left(\\mathbf {f} _ {t}, a _ {t}\\right) - r _ {t + 1} ^ {e}\\right) ^ {2}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.748, + 0.473, + 0.775 + ], + "angle": 0, + "content": "where \\(r_{t + 1}^{e}\\) is the actual external reward value at the next time step, which is returned by the environment." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.783, + 0.414, + 0.798 + ], + "angle": 0, + "content": "3.3. Consistency-guided Curiosity Module" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.804, + 0.473, + 0.832 + ], + "angle": 0, + "content": "Given state representation \\(\\mathbf{f}_t\\) from the dual-path model, we adapt SAC [9, 10] as the base RL algorithm following" + }, + { + "type": "page_footnote", + "bbox": [ + 0.117, + 0.839, + 0.473, + 0.863 + ], + "angle": 0, + "content": "2We use concatenation to fuse \\(\\mathbf{f}_t^m\\) and \\(\\mathbf{f}_t^a\\) for simplicity, advanced fusion mechanisms are also applicable." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.13, + 0.855, + 0.185 + ], + "angle": 0, + "content": "previous methods [47, 5], which aims to maximize the expected cumulative reward to find an optimal policy by approximating the action-value function \\( Q \\) and a stochastic policy \\( \\pi \\) based on a \\( \\alpha \\)-discounted maximum entropy \\( \\mathcal{H}(\\cdot) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.187, + 0.855, + 0.215 + ], + "angle": 0, + "content": "\\[\nJ (\\pi) = \\sum_ {t} \\mathbb {E} _ {\\left(\\mathbf {o} _ {t}, a _ {t}\\right) \\sim \\pi} \\left[ r \\left(\\mathbf {o} _ {t}, a _ {t}\\right) + \\alpha \\mathcal {H} \\left(\\pi \\left(\\cdot \\mid \\mathbf {o} _ {t}\\right)\\right) \\right]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.217, + 0.855, + 0.243 + ], + "angle": 0, + "content": "The action-value function \\(Q\\) are learned by minimizing the soft Bellman error:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.246, + 0.855, + 0.263 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {Q} = \\mathbb {E} _ {\\left(\\mathbf {o} _ {t}, a _ {t}\\right)} \\left(Q \\left(\\mathbf {o} _ {t}, a _ {t}\\right) - \\left(r _ {t} + \\lambda V \\left(\\mathbf {o} _ {t + 1}\\right)\\right)\\right) ^ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.265, + 0.855, + 0.292 + ], + "angle": 0, + "content": "and the soft state value \\(V\\) can be estimated by sampling an action under the current policy:" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.31, + 0.855, + 0.339 + ], + "angle": 0, + "content": "\\[\nV \\left(\\mathbf {o} _ {t + 1}\\right) = \\mathbb {E} _ {a _ {t + 1} \\sim \\pi} \\left[ \\bar {Q} \\left(\\mathbf {o} _ {t + 1}; a _ {t + 1}\\right) - \\alpha \\log \\pi \\left(a _ {t + 1} \\mid \\mathbf {o} _ {t + 1}\\right) \\right], \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.339, + 0.855, + 0.393 + ], + "angle": 0, + "content": "where \\(\\overline{Q}\\) denotes the exponential moving average of the parameters of \\(Q\\). The policy is optimized by decreasing the difference between the exponential of the soft-Q function and the policy:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.396, + 0.855, + 0.419 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\pi} = \\mathbb {E} _ {a _ {t} \\sim \\pi} \\left[ \\alpha \\log \\pi \\left(a _ {t} \\mid \\mathbf {o} _ {t}\\right) - Q \\left(\\mathbf {o} _ {t}, a _ {t}\\right) \\right]. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.423, + 0.856, + 0.572 + ], + "angle": 0, + "content": "Although SAC algorithm introduces the entropy term to encourage exploration, it still heavily relies on carefully engineered environmental extrinsic rewards and suffers from low sample efficiency. Instead of extrinsic rewards, intrinsic curiosity can be a powerful concept to endow an agent with an automated mechanism to continuously explore its environment in the absence of task information. During training, in addition to the extrinsic rewards \\( r^e \\) obtained from the environment, the curiosity module also provides intrinsic rewards \\( r^i \\). The learning objective SAC in Eq. 8 is therefore extended as:" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.573, + 0.866, + 0.612 + ], + "angle": 0, + "content": "\\[\nJ (\\pi) = \\sum_ {t} \\mathbb {E} _ {\\left(\\mathbf {o} _ {t}, a _ {t}\\right) \\sim \\pi} \\left[ r ^ {e} \\left(\\mathbf {o} _ {t}, a _ {t}\\right) + \\alpha \\mathcal {H} \\left(\\pi (\\cdot | \\mathbf {o} _ {t})\\right) + r ^ {i} \\left(\\mathbf {o} _ {t}, a _ {t}\\right) \\right]. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.614, + 0.856, + 0.805 + ], + "angle": 0, + "content": "To get proper intrinsic rewards \\( r^i \\), we propose to leverage the consistency of motion and appearance features from the dual-path model. Intuitively, as shown in Fig. 4, the environmental temporal dynamics can be obtained from two sources: 1) directly learned from the motion path, and 2) through the variations of the spatial features learned from the spatial path, similar to the latent flow [33]. Optimally, these two sources should be consistent with each other. That is, the motion feature \\( \\mathbf{f}_t^m \\) should be similar to the difference between the spatial features \\( \\mathbf{f}_{t - 2}^a \\), \\( \\mathbf{f}_{t - 1}^a \\) and \\( \\mathbf{f}_t^a \\). The idea of the consistency-guided curiosity module is then to encourage the agent to explore when the temporal dynamics produced by the two sources are inconsistent. To this end, the intrinsic reward \\( r^i \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.806, + 0.855, + 0.836 + ], + "angle": 0, + "content": "\\[\nr ^ {i} = C e ^ {- \\lambda t} d \\left(\\left\\| \\mathbf {f} _ {t} ^ {m} \\right\\|. \\left| \\mathbf {f} _ {t - 1} ^ {a} - \\mathbf {f} _ {t - 2} ^ {a} \\right| + \\left| \\mathbf {f} _ {t} ^ {a} - \\mathbf {f} _ {t - 1} ^ {a} \\right|\\right) \\frac {r _ {\\max} ^ {e}}{t _ {\\max} ^ {i}}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.836, + 0.855, + 0.863 + ], + "angle": 0, + "content": "where \\(C\\) is temperature weight, \\(\\lambda\\) is an exponential decay weight, \\(d\\) is the L2 distance function, \\(t\\) is the environment" + }, + { + "type": "page_number", + "bbox": [ + 0.489, + 0.89, + 0.511, + 0.9 + ], + "angle": 0, + "content": "180" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.12, + 0.129, + 0.385, + 0.143 + ], + "angle": 0, + "content": "Algorithm 1 Inference procedure of Simoun" + }, + { + "type": "algorithm", + "bbox": [ + 0.126, + 0.145, + 0.47, + 0.395 + ], + "angle": 0, + "content": "1: for each environment step \\(t\\) do \n2: Collect observation frames \\(\\mathbf{o}_{t-2}\\), \\(\\mathbf{o}_{t-1}\\) and \\(\\mathbf{o}_t\\). \n3: Extract motion and appearance feature maps: \n4: \\(\\mathbf{F}_t^m = \\mathcal{E}^m ([\\mathbf{o}_{t-1} - \\mathbf{o}_{t-2}), (\\mathbf{o}_t - \\mathbf{o}_{t-1})]\\), \n5: \\(\\mathbf{F}_j^a = \\mathcal{E}^a (\\mathbf{o}_j)\\), \\(j = t - 2, \\dots, t\\) \n6: Structural Interactive Modeling: \n7: \\(\\mathbf{X} = \\sigma (\\sigma (\\hat{\\mathbf{F}}_{t-2}^a \\top \\hat{\\mathbf{F}}_{t-1}^a) + \\sigma (\\hat{\\mathbf{F}}_t^{a\\top}\\hat{\\mathbf{F}}_{t-1}^a))\\) \n8: \\(\\mathbf{F}_t^m = \\mathbf{F}_t^m + \\beta \\cdot (\\mathbf{X} \\cdot \\hat{\\mathbf{F}}_t^m)\\) \n9: \\(\\mathbf{F}_t^a = \\mathbf{F}_t^a + \\beta \\cdot (\\mathbf{X} \\cdot \\hat{\\mathbf{F}}_t^a)\\) \n10: Extract low-dimensional features: \n11: \\(\\mathbf{f}_t^m = \\mathrm{LayerNorm}(\\mathrm{FC}(\\mathbf{F}_t^m))\\) \n12: \\(\\mathbf{f}_t^a = \\mathrm{LayerNorm}(\\mathrm{FC}(\\mathbf{F}_t^a))\\) \n13: Take action based on the fused features: \n14: \\(\\mathbf{f}_t = [\\mathbf{f}_t^m, \\mathbf{f}_t^a]\\) \n15: \\(a_t \\sim \\pi (a_t | \\mathbf{f}_t)\\) \n16: Environment state transition: \n17: \\(\\mathbf{o}_{t+1} \\sim \\mathcal{P}(\\mathbf{o}_{t+1} | \\mathbf{o}_t, a_t)\\) \n18: end for" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.42, + 0.474, + 0.571 + ], + "angle": 0, + "content": "time step, \\(|\\cdot |\\) is the vector element-wise absolute operation (used to avoid the cancellation of \\(\\mathbf{f}_{t - 1}^s\\)), \\(r_{max}^{e}\\) and \\(r_{max}^{i}\\) are the maximum extrinsic reward value and intrinsic reward value over the time step \\([0,t]\\), which serve as normalization term to balance \\(r^i\\) and \\(r^e\\). At the beginning of the training, \\(r_{max}^{e}\\) and \\(r_{max}^{i}\\) are assigned with an initial value and dynamically updated as the training goes on. By rewarding the agent to explore motion-appearance inconsistent states, the curiosity module encourages the agent to learn more efficiently, leading to better performance and fewer learning steps." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.579, + 0.339, + 0.594 + ], + "angle": 0, + "content": "3.4. Overall Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.6, + 0.474, + 0.642 + ], + "angle": 0, + "content": "By capitalizing on interactive motion-appearance understanding, Simoun learns from vision to control in an end-to-end manner by optimizing the following training objective:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.658, + 0.473, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\underbrace {\\mathcal {L} _ {t r a n}} _ {M o t i o n} + \\underbrace {\\mathcal {L} _ {c o n}} _ {A p p e a r a n c e} + \\underbrace {\\mathcal {L} _ {r e}} _ {S t a t e} + \\underbrace {\\mathcal {L} _ {Q} + \\mathcal {L} _ {\\pi}} _ {R L}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.691, + 0.474, + 0.746 + ], + "angle": 0, + "content": "where the objective jointly considers the learning of motion and appearance features, as well as getting efficient state representation for high-performance RL policy learning. The detailed inference procedure is presented in Alg. 1." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.758, + 0.445, + 0.773 + ], + "angle": 0, + "content": "4. An Information Bias Analysis of Simoun" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.781, + 0.474, + 0.864 + ], + "angle": 0, + "content": "As we demonstrated in Fig. 1, existing observation encoding paradigms model motion and appearance information either heuristically without special design or using preliminary techniques (such as taking latent difference). In contrast, Simoun deliberately models both information and emphasizes their interactions. In this section, we analyze" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.127, + 0.855, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.283, + 0.857, + 0.346 + ], + "angle": 0, + "content": "Figure 5. Feature component analysis of different models. M-Path and A-path represent the motion and appearance paths of Simoun, respectively. Strong information bias can be observed from the baseline models and Simoun notably alleviates the bias and decreases meaningless residual units." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.37, + 0.856, + 0.412 + ], + "angle": 0, + "content": "the representations learned by Simoun and other observation encoding paradigms from an information bias perspective to gain an in-depth understanding of them." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.412, + 0.856, + 0.701 + ], + "angle": 0, + "content": "Specifically, we leverage a recent approach [18] that can quantify static and dynamic information learned by any spatial-temporal model. The approach estimates the amount of static vs. dynamic bias based on the mutual information between sampled input sequence pairs. It then calculates the percentage of units (channel dimensions) of the model feature layer that encodes several pre-defined information factors (static, dynamic, joint, and residual). The quantifying results of Simoun and other models are illustrated in Fig. 5. It is immediately evident that previous methods tend to encode the static factor more than the dynamic, indicating a strong bias toward appearance information. It is also clear that the two paths of Simoun learn corresponding static and dynamic factors as expected, with the motion path having more dynamic units than the appearance path. By fusing the two paths together, Simoun enjoys abundant dynamic and static information, meanwhile having minimal residual units that do not involve any dynamic or static factor. As will be shown in the experiment section, such an abundant and informative representation significantly improves the decision-making process." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.715, + 0.62, + 0.73 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.739, + 0.856, + 0.807 + ], + "angle": 0, + "content": "In this section, we explore how Simoun can improve vision-based RL in terms of sample efficiency and performance gains. Two benchmarks are used for evaluation: DeepMind Control Suite (DMControl) [38] for continuous control and CARLA [4] for autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.808, + 0.856, + 0.864 + ], + "angle": 0, + "content": "Experimental Settings: Simoun is implemented on the basis of the SAC algorithm [10]. For DMControl, to avoid the potential effects of different hyperparameters, we follow the previous training setup of DrQ [44] and choose" + }, + { + "type": "page_number", + "bbox": [ + 0.489, + 0.89, + 0.51, + 0.9 + ], + "angle": 0, + "content": "181" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.122, + 0.127, + 0.855, + 0.309 + ], + "angle": 0, + "content": "
100K Step ScoresSACDreamerCURLDrQSVEASPRPlayVirtualMLRCCLFOurs
Cartpole, Swingup237±49326±27582±146759±92727±86799±42816±36806±48799±61851±14
Reacher, Easy239±183314±155538±233601±213811±115638±269785±142866±103738±99910±65
Cheetah, Run118±13235±137299±48344±67375±54467±36474±50482±38317±38518±57
Walker, Walk95±19277±12403±24612±164747±65398±165460±173643±114648±110869±45
Finger, Spin230±194341±70767±56901±104859±77868±143915±49907±58994±42872±90
Ball in cup, Catch85±130246±174769±43913±53915±71861±233929±31933±16914±20971±17
Average167.3289.8559.7688.3739.0719.0729.8772.8735.0831.8
500k Step Scores
Cartpole, Swingup330±73762±27841±45868±10865±10870±12865±11872±5869±9876±4
Reacher, Easy307±65793±164929±44942±71944±52925±79942±66957±41941±48979±44
Cheetah, Run85±51570±253518±28660±96682±65716±47719±51674±37588±22693±53
Walker, Walk71±52897±49902±43921±45919±24916±75928±30939±10936±23966±7
Finger, Spin346±95796±183926±45938±103924±93924±132963±40973±31974±6964±19
Ball in cup, Catch162±122879±87959±27963±9960±19963±8967±5964±14961±9970±12
Average216.8782.8845.8882.0882.3920.0897.3896.5878.2908.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.182, + 0.316, + 0.794, + 0.328 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-art methods on DMControl benchmark. The best results are shown in bold." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.339, + 0.851, + 0.414 + ], + "angle": 0, + "content": "
MeasurementsSACCURLDrQFlareDeepMDPOurs
Episode reward ↑121 ±26.1134 ±15.1154 ±21.5132 ±24.7170 ±36.1281 ±30.4
Distance (m) ↑74 ±17.4128 ±32.595 ±27.290 ±14.6132 ±20.4207 ±15.5
Crash intensity ↓3930 ±80.33050 ±100.32419 ±72.32668 ±95.72136 ±69.31813 ±57.8
Average steer ↓17.52% ±0.021%16.60% ±0.025%15.79% ±0.018%11.48% ±0.022%10.22% ±0.015%14.84% ±0.012%
Average brake ↓1.81% ±0.013%2.94% ±0.021%1.70% ±0.039%1.52% ±0.014%1.65% ±0.007%2.14% ±0.018%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.143, + 0.422, + 0.829, + 0.435 + ], + "angle": 0, + "content": "Table 2. Comparison with state-of-the-art methods on CARLA benchmark. \\( \\uparrow \\) indicates larger is better and \\( \\downarrow \\) means the opposite." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.458, + 0.474, + 0.649 + ], + "angle": 0, + "content": "six commonly adopted tasks: Walker-walk, Finger-spin, Cartpole-swingup, Reacher-easy, Cheetah-run and Ball in cup-catch. To evaluate sample efficiency, the performance at 100k and 500k environment steps are reported during the training stage. For CARLA, we mostly follow the setup of DBC [47], where the goal is to travel as far as possible on Highway 8 of Town 4 in 1000 time steps without any collisions with 20 moving cars. For observation acquisition, we horizontally concatenate the images from three cameras on the roof of vehicles to get \\(84 \\times 252\\) images. Random convolution [21] is adopted for data augmentation in Eq. 3. All experiments are trained across 5 random seeds to report the mean and standard deviation of the rewards. More details can be found in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.65, + 0.475, + 0.761 + ], + "angle": 0, + "content": "Methods Compared: We extensively compare Simoun with a variety of methods including the SAC [10] baseline, explicit motion modeling approach (Flare [33]), auxiliary loss-based approaches (CURL [20], MLR [46]), data augmentation approaches (DrQ [44], SVEA [15]), dynamic modeling based approaches (Dreamer [11], SPR [30], PlayVirtual [45], DeepMDP [8]), and curiosity based approach (CCLF [37])." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.772, + 0.338, + 0.787 + ], + "angle": 0, + "content": "5.1. Performance Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.794, + 0.475, + 0.864 + ], + "angle": 0, + "content": "Results on DMControl Table 1 presents the experimental results on the DMControl benchmark. It can be observed that Simoun achieves considerable performance gains compared to other state-of-the-art methods. In particular, at \\(100\\mathrm{k}\\) steps, significant performance improvement can al" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.458, + 0.857, + 0.485 + ], + "angle": 0, + "content": "ready be reached by Simoun, which indicates improved sample efficiency." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.486, + 0.857, + 0.61 + ], + "angle": 0, + "content": "Results on CARLA The results of the CARLA benchmark are reported in Table 2. It is clear that Simoun outperforms all other methods on the episode reward. Additionally, the average driving distance is farther than other methods by a large margin and the average collision intensity is also smaller. Although the driving smoothness of Simoun is slightly decreased due to increased steer and brake, this small cost has led to considerable overall reward gain to break through the current status quo." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.621, + 0.64, + 0.637 + ], + "angle": 0, + "content": "5.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.644, + 0.858, + 0.865 + ], + "angle": 0, + "content": "Effectiveness of the Dual-path Design To demonstrate the effectiveness of the dual-path design in Simoun, we compare it with three single-path methods (depicted in Fig. 1): individual frame encoding, stacked frames encoding, latent flow encoding, and our dual-path encoding. For a fair comparison, the motion and appearance losses of Simoun are adopted for all four models to eliminate the affection of loss difference, and the interactive and curiosity modules are also removed from the dual-path model. The results are shown in Fig. 6 (left). Several observations can be made: 1) The low performance of individual frame encoding (black line) indicates the importance of modeling motion information. 2) By considering the motion across frames, stacked frame encoding (green line) performs much better than individual frame encoding. 3) The latent flow encoding (blue line) improves over stacked frame encoding" + }, + { + "type": "page_number", + "bbox": [ + 0.489, + 0.89, + 0.511, + 0.9 + ], + "angle": 0, + "content": "182" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.126, + 0.368, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.127, + 0.605, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.127, + 0.844, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.265, + 0.856, + 0.292 + ], + "angle": 0, + "content": "Figure 6. Experiment results on CARLA benchmark. Left: Performance on different encoding paradigms. Middle: Ablation study of each component of Simoun. Right: Comparison with models having more parameters." + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.31, + 0.462, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.118, + 0.423, + 0.458, + 0.435 + ], + "angle": 0, + "content": "Figure 7. Visualizations of motion-appearance structure masks." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.463, + 0.473, + 0.601 + ], + "angle": 0, + "content": "with explicit motion modeling. However, there is limited room for further improvement due to the preliminary technique used for motion extraction. 4) The proposed dualpath model (red line) remarkably outperforms the other three methods by modeling motion and appearance explicitly. Interestingly, the results in Fig. 6 (left) echo perfectly with the information bias degrees illustrated and discussed in Fig. 5 and Sec. 4, which indicates potentially deeper connections between motion-appearance information bias and the performance of decision-making." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.603, + 0.474, + 0.864 + ], + "angle": 0, + "content": "Effectiveness of Each Components in Simoun To investigate the contribution of each component in Simoun, we first evaluate the performance of each individual path, then test the dual-path model by gradually adding the structural interactive module and the consistency-guided curiosity module. It can be found in Fig. 6 (middle) that the individual path gives relatively low performance when trained separately, with the motion path performing better than the appearance path. When adopting the dual-path model, both the structural interactive module and the consistency-guided curiosity module can further improve performance, which demonstrates their effectiveness. To better understand which visual clues does Simoun concentrate, we visualize the motion-appearance structure masks \\((\\mathbf{X}^m\\) and \\(\\mathbf{X}^a\\) in Eq. 5) of the two paths. As can be observed from Fig. 7, the motion path tends to focus on the moving trajectory of other vehicles (positions where the vehicles passed by), while the appearance path focuses strongly on spatial positions where the other vehicles exist." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.31, + 0.825, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.436, + 0.857, + 0.461 + ], + "angle": 0, + "content": "Figure 8. Evaluation of the generalization ability on four different environment domains on DMControl benchmark." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.497, + 0.673, + 0.511 + ], + "angle": 0, + "content": "5.3. Further Discussions" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.527, + 0.857, + 0.678 + ], + "angle": 0, + "content": "Does the performance gain of the dual-path model come from the increased network parameters? To answer this question, we compare the dual-path model with a double-channeled stacked frame model, which has nearly \\(2 \\times\\) more parameters. From Fig. 6 (right) we can observe that increasing model parameters indeed improves performance. However, the dual-path model still outperforms the double-channeled stacked frames model using only half of its parameters. This proves the effectiveness of the dualpath model mainly comes from its motion-appearance modeling paradigm, rather than increased network capacity." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.685, + 0.857, + 0.864 + ], + "angle": 0, + "content": "Does Simoun improves domain generalization? To evaluate the domain generalization ability of Simoun, we select the \"catch\" task on the \"ball in cup\" scenario as the source domain and test Simoun on DMControl Generalization Benchmark [16] with four different environment domains (color easy, color hard, video easy, and video hard). Fig. 8 shows that Simoun performs on par with other methods on color-shifted domains. However, the performance is much better on video-shifted domains, where the background is also moving. This is attributed to the specialized modeling of motion information by Simoun, which drives the agent to pay more attention to reward-related motions rather than the irrelevant dynamic background." + }, + { + "type": "page_number", + "bbox": [ + 0.489, + 0.89, + 0.511, + 0.9 + ], + "angle": 0, + "content": "183" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.118, + 0.129, + 0.226, + 0.142 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.145, + 0.474, + 0.323 + ], + "angle": 0, + "content": "We have proposed Simoun, a unified framework for vision-based RL with a dual-path network for motion and appearance understanding. The design of Simoun demonstrates the effectiveness of motion-appearance structural interaction, and further shows the benefits of consistency-guided intrinsic curiosity. Empirical results suggest that the proposed method has advantages in terms of sample efficiency, performance gains, and generalization ability. By analyzing Simoun from an information bias perspective, we build a connection between motion-appearance information bias and vision-based RL performance. We hope this connection can further inspire more efficient model designs for vision-based RL tasks." + }, + { + "type": "title", + "bbox": [ + 0.118, + 0.335, + 0.281, + 0.349 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.351, + 0.474, + 0.434 + ], + "angle": 0, + "content": "The study was funded by the Key-Area Research and Development Program of Guangdong Province with contract No. 2020B0101380001, as well as the National Natural Science Foundation of China under contracts No. 62027804, No. 61825101, No. 62088102 and No. 62202010. Computing support was provided by Pengcheng Cloudbrain." + }, + { + "type": "title", + "bbox": [ + 0.119, + 0.445, + 0.205, + 0.458 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.468, + 0.474, + 0.518 + ], + "angle": 0, + "content": "[1] Marc Bellemare, Sriram Srinivasan, Georg Ostrovski, Tom Schaul, David Saxton, and Remi Munos. Unifying count-based exploration and intrinsic motivation. In Advances in Neural Information Processing Systems, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.519, + 0.473, + 0.568 + ], + "angle": 0, + "content": "[2] David Bertoin and Emmanuel Rachelson. Local feature swapping for generalization in reinforcement learning. In International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.57, + 0.474, + 0.619 + ], + "angle": 0, + "content": "[3] Fei Deng, Ingook Jang, and Sungjin Ahn. Dreamerpro: Reconstruction-free model-based reinforcement learning with prototypical representations. In International Conference on Machine Learning, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.621, + 0.473, + 0.67 + ], + "angle": 0, + "content": "[4] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on Robot Learning, pages 1-16. PMLR, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.672, + 0.473, + 0.734 + ], + "angle": 0, + "content": "[5] Linxi Fan, Guanzhi Wang, De-An Huang, Zhiding Yu, Li Fei-Fei, Yuke Zhu, and Animashree Anandkumar. Secant: Self-expert cloning for zero-shot generalization of visual policies. In International Conference on Machine Learning, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.736, + 0.473, + 0.784 + ], + "angle": 0, + "content": "[6] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition. In Computer Vision and Pattern Recognition, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.786, + 0.473, + 0.824 + ], + "angle": 0, + "content": "[7] Xiang Fu, Ge Yang, Pulkit Agrawal, and Tommi Jaakkola. Learning task informed abstractions. In International Conference on Machine Learning, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.825, + 0.473, + 0.873 + ], + "angle": 0, + "content": "[8] Carles Gelada, Saurabh Kumar, Jacob Buckman, Ofir Nachum, and Marc G Bellemare. Deepmdp: Learning continuous latent space models for representation learning. In International Conference on Machine Learning, 2019. 8" + }, + { + "type": "list", + "bbox": [ + 0.126, + 0.468, + 0.474, + 0.873 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.13, + 0.855, + 0.18 + ], + "angle": 0, + "content": "[9] Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. In International Conference on Machine Learning, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.182, + 0.855, + 0.244 + ], + "angle": 0, + "content": "[10] Tuomas Haarnoja, Aurick Zhou, Kristian Hartikainen, George Tucker, Sehoon Ha, Jie Tan, Vikash Kumar, Henry Zhu, Abhishek Gupta, Pieter Abbeel, et al. Soft actor-critic algorithms and applications. arXiv preprint arXiv:1812.05905, 2018.6,7,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.855, + 0.295 + ], + "angle": 0, + "content": "[11] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603, 2019. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.297, + 0.855, + 0.348 + ], + "angle": 0, + "content": "[12] Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. In International Conference on Machine Learning, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.855, + 0.386 + ], + "angle": 0, + "content": "[13] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. arXiv preprint arXiv:2010.02193, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.387, + 0.855, + 0.449 + ], + "angle": 0, + "content": "[14] Nicklas Hansen, Rishabh Jangir, Yu Sun, Guillem Alenyà, Pieter Abbeel, Alexei A Efros, Lerrel Pinto, and Xiaolong Wang. Self-supervised policy adaptation during deployment. In nternational Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.451, + 0.855, + 0.5 + ], + "angle": 0, + "content": "[15] Nicklas Hansen, Hao Su, and Xiaolong Wang. Stabilizing deep q-learning with convnets and vision transformers under data augmentation. In Advances in Neural Information Processing Systems, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.502, + 0.855, + 0.54 + ], + "angle": 0, + "content": "[16] Nicklas Hansen and Xiaolong Wang. Generalization in reinforcement learning by soft data augmentation. In International Conference on Robotics and Automation, 2021. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.541, + 0.855, + 0.591 + ], + "angle": 0, + "content": "[17] Yangru Huang, Peixi Peng, Yifan Zhao, Guangyao Chen, and Yonghong Tian. Spectrum random masking for generalization in image-based reinforcement learning. In Advances in Neural Information Processing Systems. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.592, + 0.855, + 0.655 + ], + "angle": 0, + "content": "[18] Matthew Kowal, Mennatullah Siam, Md Amirul Islam, Neil DB Bruce, Richard P Wildes, and Konstantinos G Derpanis. A deeper dive into what deep spatiotemporal networks encode: Quantifying static vs. dynamic information. In Computer Vision and Pattern Recognition, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.657, + 0.855, + 0.706 + ], + "angle": 0, + "content": "[19] Misha Laskin, Kimin Lee, Adam Stooke, Lerrel Pinto, Pieter Abbeel, and Aravind Srinivas. Reinforcement learning with augmented data. In Advances in Neural Information Processing Systems, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.855, + 0.757 + ], + "angle": 0, + "content": "[20] Michael Laskin, Aravind Srinivas, and Pieter Abbeel. Curl: Contrastive unsupervised representations for reinforcement learning. In International Conference on Machine Learning, 2020. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.758, + 0.855, + 0.809 + ], + "angle": 0, + "content": "[21] Kimin Lee, Kibok Lee, Jinwoo Shin, and Honglak Lee. Network randomization: A simple technique for generalization in deep reinforcement learning. In International Conference on Machine Learning, 2020. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.81, + 0.855, + 0.873 + ], + "angle": 0, + "content": "[22] Kevin Li, Abhishek Gupta, Ashwin Reddy, Vitchyr H Pong, Aurick Zhou, Justin Yu, and Sergey Levine. Mural: Meta-learning uncertainty-aware rewards for outcome-driven reinforcement learning. In International Conference on Machine Learning, 2021. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.13, + 0.855, + 0.873 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.489, + 0.89, + 0.511, + 0.9 + ], + "angle": 0, + "content": "184" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.13, + 0.474, + 0.18 + ], + "angle": 0, + "content": "[23] Xiang Li, Jinghuan Shang, Srijan Das, and Michael S Ryoo. Does self-supervised learning really improve reinforcement learning from pixels? In Advances in Neural Information Processing Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.119, + 0.182, + 0.473, + 0.232 + ], + "angle": 0, + "content": "[24] Sicong Liu, Xi Sheryl Zhang, Yushuo Li, Yifan Zhang, and Jian Cheng. On the data-efficiency with contrastive image transformation in reinforcement learning. In International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.233, + 0.473, + 0.282 + ], + "angle": 0, + "content": "[25] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.283, + 0.473, + 0.345 + ], + "angle": 0, + "content": "[26] Thanh Nguyen, Tung M Luu, Thang Vu, and Chang D Yoo. Sample-efficient reinforcement learning representation learning with curiosity contrastive forward dynamics model. In International Conference on Intelligent Robots and Systems, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.347, + 0.473, + 0.396 + ], + "angle": 0, + "content": "[27] Minting Pan, Xiangming Zhu, Yunbo Wang, and Xiaokang Yang. Iso-dream: Isolating and leveraging noncontrollable visual dynamics in world models. In Advances in Neural Information Processing Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.398, + 0.473, + 0.446 + ], + "angle": 0, + "content": "[28] Deepak Pathak, Pulkit Agrawal, Alexei A Efros, and Trevor Darrell. Curiosity-driven exploration by self-supervised prediction. In International Conference on Machine Learning, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.448, + 0.473, + 0.497 + ], + "angle": 0, + "content": "[29] Roberta Raileanu, Maxwell Goldstein, Denis Yarats, Ilya Kostrikov, and Rob Fergus. Automatic data augmentation for generalization in reinforcement learning. In Advances in Neural Information Processing Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.499, + 0.473, + 0.548 + ], + "angle": 0, + "content": "[30] Max Schwarzer, Ankesh Anand, Rishab Goel, R Devon Hjelm, Aaron Courville, and Philip Bachman. Data-efficient reinforcement learning with self-predictive representations. arXiv preprint arXiv:2007.05929, 2020.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.55, + 0.473, + 0.599 + ], + "angle": 0, + "content": "[31] Ramanan Sekar, Oleh Rybkin, Kostas Daniilidis, Pieter Abbeel, Danijar Hafner, and Deepak Pathak. Planning to explore via self-supervised world models. In International Conference on Machine Learning, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.601, + 0.473, + 0.65 + ], + "angle": 0, + "content": "[32] Younggyo Seo, Lili Chen, Jinwoo Shin, Honglak Lee, Pieter Abbeel, and Kimin Lee. State entropy maximization with random encoders for efficient exploration. In International Conference on Machine Learning, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.652, + 0.473, + 0.7 + ], + "angle": 0, + "content": "[33] Wenling Shang, Xiaofei Wang, Aravind Srinivas, Aravind Rajeswaran, Yang Gao, Pieter Abbeel, and Misha Laskin. Reinforcement learning with latent flow. In Advances in Neural Information Processing Systems, 2021. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.702, + 0.473, + 0.751 + ], + "angle": 0, + "content": "[34] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Computer Vision and Pattern Recognition, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.753, + 0.473, + 0.79 + ], + "angle": 0, + "content": "[35] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In Advances in Neural Information Processing Systems, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.791, + 0.473, + 0.841 + ], + "angle": 0, + "content": "[36] Adam Stooke, Kimin Lee, Pieter Abbeel, and Michael Laskin. Decoupling representation learning from reinforcement learning. In International Conference on Machine Learning, pages 9870-9879. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.842, + 0.473, + 0.866 + ], + "angle": 0, + "content": "[37] Chenyu Sun, Hangwei Qian, and Chunyan Miao. Cclf: A contrastive-curiosity-driven learning framework for" + }, + { + "type": "list", + "bbox": [ + 0.119, + 0.13, + 0.474, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.131, + 0.855, + 0.155 + ], + "angle": 0, + "content": "sample-efficient reinforcement learning. arXiv preprint arXiv:2205.00943, 2022. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.157, + 0.855, + 0.207 + ], + "angle": 0, + "content": "[38] Yuval Tassa, Yotam Doron, Alistair Muldal, Tom Erez, Yazhe Li, Diego de Las Casas, David Budden, Abbas Abdelmaleki, Josh Merel, Andrew Lefrancq, et al. Deepmind control suite. arXiv preprint arXiv:1801.00690, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.209, + 0.855, + 0.271 + ], + "angle": 0, + "content": "[39] Anirudh Thatipelli, Sanath Narayan, Salman Khan, Rao Muhammad Anwer, Fahad Shahbaz Khan, and Bernard Ghanem. Spatio-temporal relation modeling for few-shot action recognition. In Computer Vision and Pattern Recognition, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.272, + 0.855, + 0.345 + ], + "angle": 0, + "content": "[40] Oriol Vinyals, Timo Ewalds, Sergey Bartunov, Petko Georgiev, Alexander Sasha Vezhnevets, Michelle Yeo, Alireza Makhzani, Heinrich Kuttler, John Agapiou, Julian Schrittwieser, et al. Starcraft ii: A new challenge for reinforcement learning. arXiv preprint arXiv:1708.04782, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.347, + 0.855, + 0.398 + ], + "angle": 0, + "content": "[41] Kaixin Wang, Bingyi Kang, Jie Shao, and Jiashi Feng. Improving generalization in reinforcement learning with mixture regularization. In Advances in Neural Information Processing Systems, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.855, + 0.437 + ], + "angle": 0, + "content": "[42] Xudong Wang, Long Lian, and Stella X Yu. Unsupervised visual attention and invariance for reinforcement learning. In Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.439, + 0.855, + 0.489 + ], + "angle": 0, + "content": "[43] Wangmeng Xiang, Chao Li, Biao Wang, Xihan Wei, XianSheng Hua, and Lei Zhang. Spatiotemporal self-attention modeling with temporal patch shift for action recognition. In European Conference on Computer Vision, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.855, + 0.54 + ], + "angle": 0, + "content": "[44] Denis Yarats, Ilya Kostrikov, and Rob Fergus. Image augmentation is all you need: Regularizing deep reinforcement learning from pixels. In International Conference on Learning Representations, 2020. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.541, + 0.855, + 0.603 + ], + "angle": 0, + "content": "[45] Tao Yu, Cuiling Lan, Wenjun Zeng, Mingxiao Feng, Zhizheng Zhang, and Zhibo Chen. Playvirtual: Augmenting cycle-consistent virtual trajectories for reinforcement learning. In Advances in Neural Information Processing Systems, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.605, + 0.855, + 0.654 + ], + "angle": 0, + "content": "[46] Tao Yu, Zhizheng Zhang, Cuiling Lan, Zhibo Chen, and Yan Lu. Mask-based latent reconstruction for reinforcement learning. In Advances in Neural Information Processing Systems. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.656, + 0.855, + 0.707 + ], + "angle": 0, + "content": "[47] Amy Zhang, Rowan McAllister, Roberto Calandra, Yarin Gal, and Sergey Levine. Learning invariant representations for reinforcement learning without reconstruction. arXiv preprint arXiv:2006.10742, 2020. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.708, + 0.855, + 0.746 + ], + "angle": 0, + "content": "[48] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In International Conference on Machine Learning, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.855, + 0.808 + ], + "angle": 0, + "content": "[49] Jinhua Zhu, Yingce Xia, Lijun Wu, Jiajun Deng, Wengang Zhou, Tao Qin, Tie-Yan Liu, and Houqiang Li. Masked contrastive representation learning for reinforcement learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.81, + 0.855, + 0.859 + ], + "angle": 0, + "content": "[50] Yuke Zhu, Josiah Wong, Ajay Mandlekar, and Roberto Martín-Martín. robosuite: A modular simulation framework and benchmark for robot learning. arXiv preprint arXiv:2009.12293, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.131, + 0.855, + 0.859 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.489, + 0.89, + 0.511, + 0.9 + ], + "angle": 0, + "content": "185" + } + ] +] \ No newline at end of file diff --git a/2023/Simoun_ Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning/17dc049b-d1dc-4e17-921e-1317f662d966_origin.pdf b/2023/Simoun_ Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning/17dc049b-d1dc-4e17-921e-1317f662d966_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..57340523aefa6eec91e28dd1869fd28f477401f7 --- /dev/null +++ b/2023/Simoun_ Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning/17dc049b-d1dc-4e17-921e-1317f662d966_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e489c0a73333662a6f4af051d8a2803b9efc0c38d92526d33a2a46c8b6d69c90 +size 4268432 diff --git a/2023/Simoun_ Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning/full.md b/2023/Simoun_ Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..211f26c56aae062def4c61e7332c089a0f0241be --- /dev/null +++ b/2023/Simoun_ Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning/full.md @@ -0,0 +1,336 @@ +# Simoun: Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning + +Yangru Huang $^{1}$ , Peixi Peng $^{2,4*}$ , Yifan Zhao $^{1}$ , Yunpeng Zhai $^{1}$ , Haoran Xu $^{3,4}$ , Yonghong Tian $^{1,2,4*}$ + +$^{1}$ School of Computer Science, Peking University + +$^{2}$ School of Electronic and Computer Engineering, Peking University + +$^{3}$ School of Intelligent Systems Engineering, Sun Yat-sen University $^{4}$ Peng Cheng Laboratory + +yrhuang@stu.pku.edu.cn, {pxpeng, zhaoyf, ypzhai, yhtian}@pku.edu.cn, xuhr9@mail2.sysu.edu.cn + +# Abstract + +Efficient motion and appearance modeling are critical for vision-based Reinforcement Learning (RL). However, existing methods struggle to reconcile motion and appearance information within the state representations learned from a single observation encoder. To address the problem, we present Synergizing Interactive Motion-appearance Understanding (Simoun), a unified framework for vision-based RL. Given consecutive observation frames, Simoun deliberately and interactively learns both motion and appearance features through a dual-path network architecture. The learning process collaborates with a structural interactive module, which explores the latent motion-appearance structures from the two network paths to leverage their complementarity. To promote sample efficiency, we further design a consistency-guided curiosity module to encourage the exploration of under-learned observations. During training, the curiosity module provides intrinsic rewards according to the consistency of environmental temporal dynamics, which are deduced from both motion and appearance network paths. Experiments conducted on DeepMind control suite and CARLA automatic driving benchmarks demonstrate the effectiveness of Simoun, where it performs favorably against state-of-the-art methods. + +# 1. Introduction + +Reinforcement learning (RL) from visual signals has achieved great success in recent years. Compared with learning from hand-crafted states, vision-based RL eliminates the arduous task of designing states with manual feature engineering. Therefore, it is beneficial for a variety of tasks such as video game playing [25, 40], robot manipulation [50], and autonomous navigation [4]. However, one of + +![](images/cb2dcbe537421bf96fdf6defd6d1f075d5a29e8ffebc48afd892c01c3662331e.jpg) +Figure 1. Illustration of different observation encoding schemes. (a) Encoding individual frames with a shared encoder. (b) Encoding stacked multiple frames. (c) Encoding individual frames and latent vector differences (latent flows). (d) Encoding motion and appearance independently and interactively. + +the major challenges of vision-based RL lies in its high dimensional observation, which is less interpretable and leads to low rewards and inefficient sampling [19, 20]. As a result, robust visual understanding is crucial to bridge the gap between vision and state-based RL in terms of performance and sample efficiency. + +To make observation comprehensible for agents, recent works have recognized the importance of learning high-quality visual features [20, 5, 29, 49, 42]. From the perspective of decision-making, two kinds of features are essential for vision-based RL: the motion features, which closely relate to the actions performed by the agents, and the appearance features, which describe the contemporary environmental states. Despite their importance, few works + +have tried to model these two kinds of features explicitly. Most methods tend to encode single or stacked multiple frames via a convolutional encoder, as shown in Fig. 1(a) and Fig. 1(b). The former neglects motion feature modeling, while the latter tightly entangles the motion and appearance information into a single network, causing information bias and optimization difficulty. To alleviate these issues, Shang et al. [33] propose to encode only appearance features in single frames and impose motion dynamics by latent vector differences (termed latent flow, see Fig. 1(c)). In addition, several world models learned directly from input images are also developed [11, 31, 3]. Nevertheless, the motion features of these methods are still computed from spatial latent space, which may lead to sub-optimal representation learning in complicated environments. + +To overcome the limitations described above, we present Synergizing Interactive Motion-appearance Understanding (Simoun) for vision-based RL. The design principle of Simoun is to explicitly learn both motion and appearance features in the early encoding stage and then interactively fuse them at the late stage for accurate decision-making. As shown in Fig. 1(d), given consecutive environment frames, Simoun learns motion and appearance information by separate network paths. The motion path (colored in blue) explicitly models motion clues (such as the speed and direction of cars) from the residual frame of multiple neighboring input frames. The appearance path (colored in orange) models the environmental spatial structures and focuses on identifying patterns and objects (such as cars and traffic lights) from every single frame. Additionally, a structural interactive module further extracts latent motion-appearance structures reflected by the correlations of the dual-path features. It then modulates both paths with the computed structure masks. In this way, each path can take complementary information from the other during learning, and agents are able to better understand the spatial and temporal context of the environment. Finally, the latent vectors from the two paths are fused for decision-making. + +Although the dual-path design of Simoun promotes observation interpretability, low sample efficiency still exists due to finite data and sparse rewards. To alleviate this issue, a consistency-guided curiosity module is further designed. The idea is to adapt the learning process by concentrating more on under-learned observations, which can be deduced from the consistency of the motion and appearance paths. Intuitively, both the motion path and the appearance path describe the same observations. Hence, the dynamic information inferred from the following two sources should remain consistent: 1) latent motion vectors learned directly from the motion path and 2) differences between the appearance path latent vectors over multiple neighboring frames. If the opposite were true, then it indicates premature motion-appearance understanding, thus more explo + +ration should be added. In this way, we build a strong correlation between reward discovery and state novelty. During training, the consistency-guided curiosity module provides intrinsic rewards in addition to extrinsic rewards from the environment, resulting in more efficient exploration. + +Experiments on both CARLA and OpenAI DMControl environments show that the proposed method performs favorably against state-of-the-arts. Overall, the contributions of this paper are threefold: + +(1) We propose Simoun, a novel dual-path learning framework that explicitly and interactively learns both motion and appearance information from observations. +(2) We design a structural interactive module to fully explore the complementarity of the two paths in Simoun and thereby further enhance visual understanding. +(3) We devise a consistency-guided curiosity module to encourage the exploration of under-learned observations. The proposed module effectively increases sample efficiency by providing intrinsic rewards for the agents. + +# 2. Related Works + +Vision-based Reinforcement Learning To improve the performance and sample efficiency of vision-based RL, existing works can be roughly divided into three groups: 1) designing auxiliary loss/learning tasks [20, 14, 36, 49, 23], 2) employing various data augmentation techniques [21, 41, 2, 17, 24], and 3) modeling environment dynamics [12, 11, 13, 7, 27]. However, most existing works utilize a single-path network with multi-frame inputs, in which the motion and appearance information is tightly entangled without explicit separation. One exception is Flare [33], which also utilizes a single-path network to encode each frame individually and models motion information explicitly by taking latent vector differences. Although Flare achieves improved performance, its motion information is still computed from single-frame appearance features, causing insufficient temporal information extraction. + +Dual-path Networks for Visual Modeling There is a rich literature of works on visual modeling with dual-path networks. One of the earliest works is the two-stream CNNs for action recognition [35], which utilizes a spatial stream with single-frame input and a temporal stream taking multiframe optical flows. Thereafter, the concept of dual-path networks is heavily explored with various fusion strategies on different tasks [34, 6, 39, 43]. Our approach differs from existing ones in terms of task, architecture and learning mechanism. Compare with other visual tasks, vision-based RL needs to extract fine-grained motion details across different time steps, which poses a great challenge to RL models. To the best of our knowledge, Simoun is one of the first vision-based RL methods which aim to explicitly learn motion and appearance features by a dual-path architecture. Meanwhile, its structural interactive module is + +![](images/5df02ca57eb18b68f9e9bf127e3ec6b2a9ad426b9cc2caa043529ed1fba78b4d.jpg) +Figure 2. The proposed dual-path learning framework Simoun. It introduces two modules: 1) a structural interactive module that leverages the complementarity between the two network paths by extracting latent motion-appearance structures, and 2) a consistency-guided curiosity module that provides intrinsic rewards according to the dynamic consistency between the motion and appearance features. + +also deliberately designed to extract motion and appearance clues residing in consecutive observation frames. By learning with RL-oriented objectives and curiosity-driven strategies, Simoun successfully imports the idea of dual-path networks into robust visual control. + +Intrinsic Reward Exploration One of the key elements of RL is the reward function, which aims to quantify the "goodness" of the agent's decisions. However, the problem is that designing dense, well-defined extrinsic reward functions is difficult and unscalable. One possible solution is to introduce an intrinsic reward function, which is calculated by the agents. Existing intrinsic reward exploration methods mainly focus on counting or predicting state novelty [1], prediction error [28], uncertainty [22], or environmental dynamics [32]. These methods are typically designed for the general state-based instead of vision-based RL. The works most relevant to us are CCFDM [26] and CCLF [37], which also formulate intrinsic rewards for visual-based RL. CCFDM utilizes forward dynamics and CCLF is based on a contrastive term. Differently, the consistency-guided curiosity module in Simoun exploits the dynamic consistency extracted from both motion and appearance features to achieve reliable intrinsic reward estimation. + +# 3. Methodology + +We start by formalizing the task of vision-based RL in Sec. 3.1 and then delineate Simoun in detail. The general idea of Simoun is to 1) explicitly model the motion/appearance dynamics (and their correlations) of the agent's operating ambiance and 2) capitalize on the learned + +dynamics to achieve efficient exploration. As shown in Fig. 2, our framework consists of two parts. First, given inputs in the form of consecutive frames, it interactively models the motion and appearance features using a dualpath architecture with a set of targeted objectives and a structural interactive module (Sec. 3.2). Then the decision-making strategy is learned with adaptive curiosity assignment steered by the proposed consistency-guided curiosity module (Sec. 3.3). The overall learning objective and inference process of Simoun are finally summarized in Sec. 3.4. + +# 3.1. Problem formulation + +Vision-based RL can be formulated as a Partially Observable Markov Decision Process (POMDP) $\mathcal{M} = < \mathcal{O},\mathcal{A},\mathcal{P},\mathcal{R},\gamma >$ where $\mathcal{O}$ denotes the observation space containing pixel frames $\mathbf{o}_t$ at different time step $t$ and $\mathcal{A}$ denotes the action space. At each $t$ the agent chooses an action $a_{t}\in \mathcal{A}$ . $\mathcal{P}(\mathbf{o}_{t + 1}|\mathbf{o}_t,a_t)$ is the observation transition, $\mathcal{R}(\mathbf{o}_t,a_t)$ is the reward function, and $\gamma \in [0,1)$ is the discount factor. The goal is to identify an optimal policy $\pi$ that maximizes the expected cumulative reward based on the visual observations rather than the complete environment state: + +$$ +J (\pi) = \sum_ {t} \mathbb {E} _ {\left(\mathbf {o} _ {t}, a _ {t}\right) \sim \pi} \left[ \mathcal {R} \left(\mathbf {o} _ {t}, a _ {t}\right) \right]. \tag {1} +$$ + +During training, $\pi$ is used to interact with the environment, and the related data are stored in a replay buffer $\mathcal{B}$ . + +# 3.2. Dual-path Interactive Modeling + +In vision-based RL, motion and appearance clues are both vital for the agent to perform accurate decision + +making. Simoun adopts a dual-path architecture to interactively learn motion and appearance features. Specifically, it consists of three components: the motion path, the appearance path, and the structural interactive module that serves as a switchboard between the motion and appearance structures learned from the two paths. + +Motion Path Motion information is crucial for the visual system to understand the dynamics of the surrounding environment. The motion path aims to extract the motion-related features (such as the velocity of moving objects) from the changes between consecutive frames. Given a tuple of three adjacent observations $[\mathbf{o}_{t - 2},\mathbf{o}_{t - 1},\mathbf{o}_t]$ sampled from replay buffer $\mathcal{B}$ , the input of the motion path is the residual of adjacent frames $[(0_{t - 1} - 0_{t - 2}),(0_t - 0_{t - 1})]$ , concatenated along channel dimensions. A convolutional encoder $\mathcal{E}^m$ is then used to learn a lower-dimensional motion representation. Specifically, the encoder includes four convolution layers with $3\times 3$ kernel size and ReLU nonlinearity. Denote the feature map of the last convolution layer as $\mathbf{F}_t^m\in \mathbb{R}^{c\times h\times w}$ , a fully connected (FC) layer with layer normalization (LN) is used to reduce the dimension of $\mathbf{F}_t^m$ to get the motion feature vector $\mathbf{f}_t^m$ . + +To further impel the motion path to catch abundant environment temporal structures, a motion-aware constraint is deliberately designed. Specifically, given $\mathbf{f}_t^m$ and corresponding action $a_{t}$ at timestep $t$ , an action-conditioned two-FC-layer transition model $\mathcal{G}$ is used to obtain a motion-action joint representation of the current timestep. A latent-space transition loss is then defined as: + +$$ +\mathcal {L} _ {t r a n} = \left\| \mathcal {G} \left(\mathbf {f} _ {t} ^ {m}, u _ {t}\right) - \mathbf {f} _ {t + 1} ^ {m} \right\| _ {2} ^ {2}. \tag {2} +$$ + +which encourages $\mathbf{f}_t^m$ to model the motion trends over time by predicting the future temporal dynamics $\mathbf{f}_{t + 1}^{m}$ of the next time step. + +Appearance Path The appearance path is designed to extract the visual appearance representations of the object and scene from individual observation frames. The input of the appearance path is a single frame $\mathbf{o}_t$ from the replay buffer $\mathcal{B}$ . Similar to the motion path, the appearance path also adopts a four-layer encoder $\mathcal{E}^a$ with the same architecture (except for the number of input channels of the first layer) to get feature map $\mathbf{F}_t^a\in \mathbb{R}^{c\times h\times w}$ , followed by FC and LN layers to get the appearance representation $\mathbf{f}_t^a$ . + +To explicitly make appearance representations discriminative to different scenes, we take inspiration from CURL [20] and adopt an unsupervised contrastive loss between similar and dissimilar sample pairs. Given observation frame $\mathbf{o}_i$ , we consider $\tilde{\mathbf{o}}_i$ (obtained from $\mathbf{o}_i$ by data augmentation) as the positive sample, while samples coming from different observation frames are regarded as negatives. The contrastive loss is then defined as: + +![](images/8b95db8882fb3931dede5d8f3b89fbe4514957d962f6cbf4ae2300b8e67efd28.jpg) +Figure 3. Illustration of the proposed structural interactive module. + +$$ +\mathcal {L} _ {c o n} = - \log \frac {\exp \left(\mathbf {f} _ {i} ^ {a ^ {\top}} \tilde {\mathbf {f}} _ {i} ^ {a}\right)}{\exp \left(\mathbf {f} _ {i} ^ {a ^ {\top}} \tilde {\mathbf {f}} _ {i} ^ {a}\right) + \sum_ {j = 0} ^ {K - 1} \exp \left(\mathbf {f} _ {i} ^ {a ^ {\top}} \tilde {\mathbf {f}} _ {j} ^ {a}\right)}, \tag {3} +$$ + +where $\mathbf{f}_t^a$ and $\tilde{\mathbf{f}}_t^a$ are appearance representations of $\mathbf{o}_t$ and $\tilde{\mathbf{o}}_t$ , respectively. $j$ is the sample index in a training batch with $K$ samples. Note that the augmented data is only used for representation learning and is inaccessible during decision-making. This is to avoid unstable training, which is easily caused by the augmented data dominating the evaluation of the Q-target. + +Structural Interactive Module Instead of simply concatenating the features from motion and appearance paths as state representations, interactively communicating the learned motion-appearance structures between both paths can lead to a more robust visual understanding. Intuitively, as shown in Fig. 3(left), given consecutive observation frames, the motion structure can be revealed by inter-frame appearance variation. Meanwhile, the appearance structure can be obtained by spatial pixel correlation. Therefore, the structural interactive module embraces a single and interframe relation discovering mechanism for efficient motion-appearance structure mining and propagation. + +Specifically, as shown in Fig. 3(right), the inputs of the structural interactive module are motion feature map $\mathbf{F}_t^m$ and appearance feature maps $\mathbf{F}_{t - 2}^{a},\mathbf{F}_{t - 1}^{a}$ and $\mathbf{F}_t^a$ .To extract motion structures, an inter-frame attention mechanism is designed. In particular, $\mathbf{F}_{t - 1}^{a}$ is treated as the query, and $\mathbf{F}_{t - 2}^{a},\mathbf{F}_{t}^{a}$ is used as two keys. A inter-frame attention map $\mathbf{X}$ can be obtained as: + +$$ +\mathbf {X} = \sigma \left(\sigma \left(\hat {\mathbf {F}} _ {t - 2} ^ {a} ^ {\top} \hat {\mathbf {F}} _ {t - 1} ^ {a}\right) + \sigma \left(\hat {\mathbf {F}} _ {t} ^ {a \top} \hat {\mathbf {F}} _ {t - 1} ^ {a}\right)\right), \tag {4} +$$ + +where $\hat{\mathbf{F}}_t^a\in \mathbb{R}^{c\times h\times w}$ denotes a new feature map generated by feeding $\mathbf{F}_t^a$ to a convolution layer for reducing the spatial complexity ( $\hat{\mathbf{F}}_{t - 1}^a$ and $\hat{\mathbf{F}}_{t - 2}^a$ can be obtained similarly). $\sigma$ denotes the Softmax function. Combining the outputs of two initial Softmax functions might not yield values sum + +![](images/50e6f72ebcab46802578cdc37e638c109921863340f1079d7e2328bd703cc8ee.jpg) +Figure 4. The idea of consistency-guided curiosity module. + +ming to 1, so an additional Softmax is used to normalize the final attention scores. The soft indexing of $\hat{\mathbf{F}}_{t - 1}^{a}$ by $\hat{\mathbf{F}}_{t - 2}^{a}$ or $\hat{\mathbf{F}}_t^a$ is performed at the spatial dimensions, resulting in a soft attention map of dimensions $hw\times hw$ . Then a spatial gating operation is applied on $\mathbf{X}$ using weights calculated from both $\mathbf{F}_t^m$ and $\mathbf{F}_t^a$ to consolidate appearance structures and obtain two motion-appearance structure masks: + +$$ +\mathbf {X} ^ {m} = \mathbf {X} \cdot \hat {\mathbf {F}} _ {t} ^ {m}, \mathbf {X} ^ {a} = \mathbf {X} \cdot \hat {\mathbf {F}} _ {t} ^ {a}, \tag {5} +$$ + +where $\hat{\mathbf{F}}_t^m$ is a feature map of $\mathbf{F}_t^m$ obtained with another convolution layer. Finally, $\mathbf{F}_t^m$ and $\mathbf{F}_t^a$ are enhanced by modulating with their corresponding motion-appearance structure masks: + +$$ +\mathbf {F} _ {t} ^ {m} = \mathbf {F} _ {t} ^ {m} + \boldsymbol {\beta} \cdot \mathbf {X} ^ {m}, \mathbf {F} _ {t} ^ {a} = \mathbf {F} _ {t} ^ {a} + \boldsymbol {\beta} \cdot \mathbf {X} ^ {a}, \tag {6} +$$ + +where $\beta$ is a learnable parameter initialized as zero [48], and the enhanced $\mathbf{F}_t^m$ and $\mathbf{F}_t^a$ instead of the original ones are further used to calculate $\mathbf{f}_t^m$ and $\mathbf{f}_t^a$ . + +After acquiring the structure-enhanced motion and appearance features, they are finally fused as $\mathbf{f}_t = [\mathbf{f}_t^m,\mathbf{f}_t^a ]$ where $\mathbf{f}_t$ is used as the final state representation. To better capture reward-related features from the fused state representation, a reward function $\mathcal{R}$ is further introduced to predict a numerical reward value to each state-action pair. $\mathcal{R}$ has a similar architecture with the transition model $\mathcal{G}$ except the output dimension is set to one. A reward loss $\mathcal{L}_{re}$ is then defined as the mean squared error between the predicted and actual reward: + +$$ +\mathcal {L} _ {r e} = \left(\mathcal {R} \left(\mathbf {f} _ {t}, a _ {t}\right) - r _ {t + 1} ^ {e}\right) ^ {2}, \tag {7} +$$ + +where $r_{t + 1}^{e}$ is the actual external reward value at the next time step, which is returned by the environment. + +# 3.3. Consistency-guided Curiosity Module + +Given state representation $\mathbf{f}_t$ from the dual-path model, we adapt SAC [9, 10] as the base RL algorithm following + +previous methods [47, 5], which aims to maximize the expected cumulative reward to find an optimal policy by approximating the action-value function $Q$ and a stochastic policy $\pi$ based on a $\alpha$ -discounted maximum entropy $\mathcal{H}(\cdot)$ : + +$$ +J (\pi) = \sum_ {t} \mathbb {E} _ {\left(\mathbf {o} _ {t}, a _ {t}\right) \sim \pi} \left[ r \left(\mathbf {o} _ {t}, a _ {t}\right) + \alpha \mathcal {H} \left(\pi \left(\cdot \mid \mathbf {o} _ {t}\right)\right) \right]. \tag {8} +$$ + +The action-value function $Q$ are learned by minimizing the soft Bellman error: + +$$ +\mathcal {L} _ {Q} = \mathbb {E} _ {\left(\mathbf {o} _ {t}, a _ {t}\right)} \left(Q \left(\mathbf {o} _ {t}, a _ {t}\right) - \left(r _ {t} + \lambda V \left(\mathbf {o} _ {t + 1}\right)\right)\right) ^ {2}, \tag {9} +$$ + +and the soft state value $V$ can be estimated by sampling an action under the current policy: + +$$ +V \left(\mathbf {o} _ {t + 1}\right) = \mathbb {E} _ {a _ {t + 1} \sim \pi} \left[ \bar {Q} \left(\mathbf {o} _ {t + 1}; a _ {t + 1}\right) - \alpha \log \pi \left(a _ {t + 1} \mid \mathbf {o} _ {t + 1}\right) \right], \tag {10} +$$ + +where $\overline{Q}$ denotes the exponential moving average of the parameters of $Q$ . The policy is optimized by decreasing the difference between the exponential of the soft-Q function and the policy: + +$$ +\mathcal {L} _ {\pi} = \mathbb {E} _ {a _ {t} \sim \pi} \left[ \alpha \log \pi \left(a _ {t} \mid \mathbf {o} _ {t}\right) - Q \left(\mathbf {o} _ {t}, a _ {t}\right) \right]. \tag {11} +$$ + +Although SAC algorithm introduces the entropy term to encourage exploration, it still heavily relies on carefully engineered environmental extrinsic rewards and suffers from low sample efficiency. Instead of extrinsic rewards, intrinsic curiosity can be a powerful concept to endow an agent with an automated mechanism to continuously explore its environment in the absence of task information. During training, in addition to the extrinsic rewards $r^e$ obtained from the environment, the curiosity module also provides intrinsic rewards $r^i$ . The learning objective SAC in Eq. 8 is therefore extended as: + +$$ +J (\pi) = \sum_ {t} \mathbb {E} _ {\left(\mathbf {o} _ {t}, a _ {t}\right) \sim \pi} \left[ r ^ {e} \left(\mathbf {o} _ {t}, a _ {t}\right) + \alpha \mathcal {H} \left(\pi (\cdot | \mathbf {o} _ {t})\right) + r ^ {i} \left(\mathbf {o} _ {t}, a _ {t}\right) \right]. \tag {12} +$$ + +To get proper intrinsic rewards $r^i$ , we propose to leverage the consistency of motion and appearance features from the dual-path model. Intuitively, as shown in Fig. 4, the environmental temporal dynamics can be obtained from two sources: 1) directly learned from the motion path, and 2) through the variations of the spatial features learned from the spatial path, similar to the latent flow [33]. Optimally, these two sources should be consistent with each other. That is, the motion feature $\mathbf{f}_t^m$ should be similar to the difference between the spatial features $\mathbf{f}_{t - 2}^a$ , $\mathbf{f}_{t - 1}^a$ and $\mathbf{f}_t^a$ . The idea of the consistency-guided curiosity module is then to encourage the agent to explore when the temporal dynamics produced by the two sources are inconsistent. To this end, the intrinsic reward $r^i$ is defined as: + +$$ +r ^ {i} = C e ^ {- \lambda t} d \left(\left\| \mathbf {f} _ {t} ^ {m} \right\|. \left| \mathbf {f} _ {t - 1} ^ {a} - \mathbf {f} _ {t - 2} ^ {a} \right| + \left| \mathbf {f} _ {t} ^ {a} - \mathbf {f} _ {t - 1} ^ {a} \right|\right) \frac {r _ {\max} ^ {e}}{t _ {\max} ^ {i}}, \tag {13} +$$ + +where $C$ is temperature weight, $\lambda$ is an exponential decay weight, $d$ is the L2 distance function, $t$ is the environment + +Algorithm 1 Inference procedure of Simoun +1: for each environment step $t$ do +2: Collect observation frames $\mathbf{o}_{t-2}$ , $\mathbf{o}_{t-1}$ and $\mathbf{o}_t$ . +3: Extract motion and appearance feature maps: +4: $\mathbf{F}_t^m = \mathcal{E}^m ([\mathbf{o}_{t-1} - \mathbf{o}_{t-2}), (\mathbf{o}_t - \mathbf{o}_{t-1})]$ , +5: $\mathbf{F}_j^a = \mathcal{E}^a (\mathbf{o}_j)$ , $j = t - 2, \dots, t$ +6: Structural Interactive Modeling: +7: $\mathbf{X} = \sigma (\sigma (\hat{\mathbf{F}}_{t-2}^a \top \hat{\mathbf{F}}_{t-1}^a) + \sigma (\hat{\mathbf{F}}_t^{a\top}\hat{\mathbf{F}}_{t-1}^a))$ +8: $\mathbf{F}_t^m = \mathbf{F}_t^m + \beta \cdot (\mathbf{X} \cdot \hat{\mathbf{F}}_t^m)$ +9: $\mathbf{F}_t^a = \mathbf{F}_t^a + \beta \cdot (\mathbf{X} \cdot \hat{\mathbf{F}}_t^a)$ +10: Extract low-dimensional features: +11: $\mathbf{f}_t^m = \mathrm{LayerNorm}(\mathrm{FC}(\mathbf{F}_t^m))$ +12: $\mathbf{f}_t^a = \mathrm{LayerNorm}(\mathrm{FC}(\mathbf{F}_t^a))$ +13: Take action based on the fused features: +14: $\mathbf{f}_t = [\mathbf{f}_t^m, \mathbf{f}_t^a]$ +15: $a_t \sim \pi (a_t | \mathbf{f}_t)$ +16: Environment state transition: +17: $\mathbf{o}_{t+1} \sim \mathcal{P}(\mathbf{o}_{t+1} | \mathbf{o}_t, a_t)$ +18: end for + +time step, $|\cdot |$ is the vector element-wise absolute operation (used to avoid the cancellation of $\mathbf{f}_{t - 1}^s$ ), $r_{max}^{e}$ and $r_{max}^{i}$ are the maximum extrinsic reward value and intrinsic reward value over the time step $[0,t]$ , which serve as normalization term to balance $r^i$ and $r^e$ . At the beginning of the training, $r_{max}^{e}$ and $r_{max}^{i}$ are assigned with an initial value and dynamically updated as the training goes on. By rewarding the agent to explore motion-appearance inconsistent states, the curiosity module encourages the agent to learn more efficiently, leading to better performance and fewer learning steps. + +# 3.4. Overall Training Objective + +By capitalizing on interactive motion-appearance understanding, Simoun learns from vision to control in an end-to-end manner by optimizing the following training objective: + +$$ +\mathcal {L} = \underbrace {\mathcal {L} _ {t r a n}} _ {M o t i o n} + \underbrace {\mathcal {L} _ {c o n}} _ {A p p e a r a n c e} + \underbrace {\mathcal {L} _ {r e}} _ {S t a t e} + \underbrace {\mathcal {L} _ {Q} + \mathcal {L} _ {\pi}} _ {R L}, \tag {14} +$$ + +where the objective jointly considers the learning of motion and appearance features, as well as getting efficient state representation for high-performance RL policy learning. The detailed inference procedure is presented in Alg. 1. + +# 4. An Information Bias Analysis of Simoun + +As we demonstrated in Fig. 1, existing observation encoding paradigms model motion and appearance information either heuristically without special design or using preliminary techniques (such as taking latent difference). In contrast, Simoun deliberately models both information and emphasizes their interactions. In this section, we analyze + +![](images/f7a4511aa94bf9946b7464cbb16926a6e095b96eb6b75af418a55486e34c5051.jpg) +Figure 5. Feature component analysis of different models. M-Path and A-path represent the motion and appearance paths of Simoun, respectively. Strong information bias can be observed from the baseline models and Simoun notably alleviates the bias and decreases meaningless residual units. + +the representations learned by Simoun and other observation encoding paradigms from an information bias perspective to gain an in-depth understanding of them. + +Specifically, we leverage a recent approach [18] that can quantify static and dynamic information learned by any spatial-temporal model. The approach estimates the amount of static vs. dynamic bias based on the mutual information between sampled input sequence pairs. It then calculates the percentage of units (channel dimensions) of the model feature layer that encodes several pre-defined information factors (static, dynamic, joint, and residual). The quantifying results of Simoun and other models are illustrated in Fig. 5. It is immediately evident that previous methods tend to encode the static factor more than the dynamic, indicating a strong bias toward appearance information. It is also clear that the two paths of Simoun learn corresponding static and dynamic factors as expected, with the motion path having more dynamic units than the appearance path. By fusing the two paths together, Simoun enjoys abundant dynamic and static information, meanwhile having minimal residual units that do not involve any dynamic or static factor. As will be shown in the experiment section, such an abundant and informative representation significantly improves the decision-making process. + +# 5. Experiments + +In this section, we explore how Simoun can improve vision-based RL in terms of sample efficiency and performance gains. Two benchmarks are used for evaluation: DeepMind Control Suite (DMControl) [38] for continuous control and CARLA [4] for autonomous driving. + +Experimental Settings: Simoun is implemented on the basis of the SAC algorithm [10]. For DMControl, to avoid the potential effects of different hyperparameters, we follow the previous training setup of DrQ [44] and choose + +
100K Step ScoresSACDreamerCURLDrQSVEASPRPlayVirtualMLRCCLFOurs
Cartpole, Swingup237±49326±27582±146759±92727±86799±42816±36806±48799±61851±14
Reacher, Easy239±183314±155538±233601±213811±115638±269785±142866±103738±99910±65
Cheetah, Run118±13235±137299±48344±67375±54467±36474±50482±38317±38518±57
Walker, Walk95±19277±12403±24612±164747±65398±165460±173643±114648±110869±45
Finger, Spin230±194341±70767±56901±104859±77868±143915±49907±58994±42872±90
Ball in cup, Catch85±130246±174769±43913±53915±71861±233929±31933±16914±20971±17
Average167.3289.8559.7688.3739.0719.0729.8772.8735.0831.8
500k Step Scores
Cartpole, Swingup330±73762±27841±45868±10865±10870±12865±11872±5869±9876±4
Reacher, Easy307±65793±164929±44942±71944±52925±79942±66957±41941±48979±44
Cheetah, Run85±51570±253518±28660±96682±65716±47719±51674±37588±22693±53
Walker, Walk71±52897±49902±43921±45919±24916±75928±30939±10936±23966±7
Finger, Spin346±95796±183926±45938±103924±93924±132963±40973±31974±6964±19
Ball in cup, Catch162±122879±87959±27963±9960±19963±8967±5964±14961±9970±12
Average216.8782.8845.8882.0882.3920.0897.3896.5878.2908.0
+ +Table 1. Comparison with state-of-the-art methods on DMControl benchmark. The best results are shown in bold. + +
MeasurementsSACCURLDrQFlareDeepMDPOurs
Episode reward ↑121 ±26.1134 ±15.1154 ±21.5132 ±24.7170 ±36.1281 ±30.4
Distance (m) ↑74 ±17.4128 ±32.595 ±27.290 ±14.6132 ±20.4207 ±15.5
Crash intensity ↓3930 ±80.33050 ±100.32419 ±72.32668 ±95.72136 ±69.31813 ±57.8
Average steer ↓17.52% ±0.021%16.60% ±0.025%15.79% ±0.018%11.48% ±0.022%10.22% ±0.015%14.84% ±0.012%
Average brake ↓1.81% ±0.013%2.94% ±0.021%1.70% ±0.039%1.52% ±0.014%1.65% ±0.007%2.14% ±0.018%
+ +Table 2. Comparison with state-of-the-art methods on CARLA benchmark. $\uparrow$ indicates larger is better and $\downarrow$ means the opposite. + +six commonly adopted tasks: Walker-walk, Finger-spin, Cartpole-swingup, Reacher-easy, Cheetah-run and Ball in cup-catch. To evaluate sample efficiency, the performance at 100k and 500k environment steps are reported during the training stage. For CARLA, we mostly follow the setup of DBC [47], where the goal is to travel as far as possible on Highway 8 of Town 4 in 1000 time steps without any collisions with 20 moving cars. For observation acquisition, we horizontally concatenate the images from three cameras on the roof of vehicles to get $84 \times 252$ images. Random convolution [21] is adopted for data augmentation in Eq. 3. All experiments are trained across 5 random seeds to report the mean and standard deviation of the rewards. More details can be found in the Supplementary Material. + +Methods Compared: We extensively compare Simoun with a variety of methods including the SAC [10] baseline, explicit motion modeling approach (Flare [33]), auxiliary loss-based approaches (CURL [20], MLR [46]), data augmentation approaches (DrQ [44], SVEA [15]), dynamic modeling based approaches (Dreamer [11], SPR [30], PlayVirtual [45], DeepMDP [8]), and curiosity based approach (CCLF [37]). + +# 5.1. Performance Comparisons + +Results on DMControl Table 1 presents the experimental results on the DMControl benchmark. It can be observed that Simoun achieves considerable performance gains compared to other state-of-the-art methods. In particular, at $100\mathrm{k}$ steps, significant performance improvement can al + +ready be reached by Simoun, which indicates improved sample efficiency. + +Results on CARLA The results of the CARLA benchmark are reported in Table 2. It is clear that Simoun outperforms all other methods on the episode reward. Additionally, the average driving distance is farther than other methods by a large margin and the average collision intensity is also smaller. Although the driving smoothness of Simoun is slightly decreased due to increased steer and brake, this small cost has led to considerable overall reward gain to break through the current status quo. + +# 5.2. Ablation Study + +Effectiveness of the Dual-path Design To demonstrate the effectiveness of the dual-path design in Simoun, we compare it with three single-path methods (depicted in Fig. 1): individual frame encoding, stacked frames encoding, latent flow encoding, and our dual-path encoding. For a fair comparison, the motion and appearance losses of Simoun are adopted for all four models to eliminate the affection of loss difference, and the interactive and curiosity modules are also removed from the dual-path model. The results are shown in Fig. 6 (left). Several observations can be made: 1) The low performance of individual frame encoding (black line) indicates the importance of modeling motion information. 2) By considering the motion across frames, stacked frame encoding (green line) performs much better than individual frame encoding. 3) The latent flow encoding (blue line) improves over stacked frame encoding + +![](images/dae08905ec192cb390fd34f8dd63b02476edbfc943e7da8d50c213bf4b24f7ae.jpg) +Figure 6. Experiment results on CARLA benchmark. Left: Performance on different encoding paradigms. Middle: Ablation study of each component of Simoun. Right: Comparison with models having more parameters. + +![](images/de0bff32cca3531b6dccaf71c9f90a2fd7330246cac5cd0d8595ea3c0e0c6b7c.jpg) + +![](images/2faa47d456470448a415fffdbb2e78cd8757516ec1271af109a5ceb4629d802b.jpg) + +![](images/38e26d4d4f00d87a6dd7210bb57d54996e55d596a1de5a983744d555af1641d1.jpg) +Figure 7. Visualizations of motion-appearance structure masks. + +with explicit motion modeling. However, there is limited room for further improvement due to the preliminary technique used for motion extraction. 4) The proposed dualpath model (red line) remarkably outperforms the other three methods by modeling motion and appearance explicitly. Interestingly, the results in Fig. 6 (left) echo perfectly with the information bias degrees illustrated and discussed in Fig. 5 and Sec. 4, which indicates potentially deeper connections between motion-appearance information bias and the performance of decision-making. + +Effectiveness of Each Components in Simoun To investigate the contribution of each component in Simoun, we first evaluate the performance of each individual path, then test the dual-path model by gradually adding the structural interactive module and the consistency-guided curiosity module. It can be found in Fig. 6 (middle) that the individual path gives relatively low performance when trained separately, with the motion path performing better than the appearance path. When adopting the dual-path model, both the structural interactive module and the consistency-guided curiosity module can further improve performance, which demonstrates their effectiveness. To better understand which visual clues does Simoun concentrate, we visualize the motion-appearance structure masks $(\mathbf{X}^m$ and $\mathbf{X}^a$ in Eq. 5) of the two paths. As can be observed from Fig. 7, the motion path tends to focus on the moving trajectory of other vehicles (positions where the vehicles passed by), while the appearance path focuses strongly on spatial positions where the other vehicles exist. + +![](images/98f7364f51785f3792cc7fd6ef0a01c0f0501203727accabd61b31b666a9c0b9.jpg) +Figure 8. Evaluation of the generalization ability on four different environment domains on DMControl benchmark. + +# 5.3. Further Discussions + +Does the performance gain of the dual-path model come from the increased network parameters? To answer this question, we compare the dual-path model with a double-channeled stacked frame model, which has nearly $2 \times$ more parameters. From Fig. 6 (right) we can observe that increasing model parameters indeed improves performance. However, the dual-path model still outperforms the double-channeled stacked frames model using only half of its parameters. This proves the effectiveness of the dualpath model mainly comes from its motion-appearance modeling paradigm, rather than increased network capacity. + +Does Simoun improves domain generalization? To evaluate the domain generalization ability of Simoun, we select the "catch" task on the "ball in cup" scenario as the source domain and test Simoun on DMControl Generalization Benchmark [16] with four different environment domains (color easy, color hard, video easy, and video hard). Fig. 8 shows that Simoun performs on par with other methods on color-shifted domains. However, the performance is much better on video-shifted domains, where the background is also moving. This is attributed to the specialized modeling of motion information by Simoun, which drives the agent to pay more attention to reward-related motions rather than the irrelevant dynamic background. + +# 6. Conclusion + +We have proposed Simoun, a unified framework for vision-based RL with a dual-path network for motion and appearance understanding. The design of Simoun demonstrates the effectiveness of motion-appearance structural interaction, and further shows the benefits of consistency-guided intrinsic curiosity. Empirical results suggest that the proposed method has advantages in terms of sample efficiency, performance gains, and generalization ability. By analyzing Simoun from an information bias perspective, we build a connection between motion-appearance information bias and vision-based RL performance. We hope this connection can further inspire more efficient model designs for vision-based RL tasks. + +# 7. Acknowledgement + +The study was funded by the Key-Area Research and Development Program of Guangdong Province with contract No. 2020B0101380001, as well as the National Natural Science Foundation of China under contracts No. 62027804, No. 61825101, No. 62088102 and No. 62202010. Computing support was provided by Pengcheng Cloudbrain. + +# References + +[1] Marc Bellemare, Sriram Srinivasan, Georg Ostrovski, Tom Schaul, David Saxton, and Remi Munos. Unifying count-based exploration and intrinsic motivation. In Advances in Neural Information Processing Systems, 2016. 4 +[2] David Bertoin and Emmanuel Rachelson. Local feature swapping for generalization in reinforcement learning. In International Conference on Learning Representations, 2022. 3 +[3] Fei Deng, Ingook Jang, and Sungjin Ahn. Dreamerpro: Reconstruction-free model-based reinforcement learning with prototypical representations. In International Conference on Machine Learning, 2022. 3 +[4] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on Robot Learning, pages 1-16. PMLR, 2017. 2, 7 +[5] Linxi Fan, Guanzhi Wang, De-An Huang, Zhiding Yu, Li Fei-Fei, Yuke Zhu, and Animashree Anandkumar. Secant: Self-expert cloning for zero-shot generalization of visual policies. In International Conference on Machine Learning, 2021. 2, 6 +[6] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition. In Computer Vision and Pattern Recognition, 2016. 3 +[7] Xiang Fu, Ge Yang, Pulkit Agrawal, and Tommi Jaakkola. Learning task informed abstractions. In International Conference on Machine Learning, 2021. 3 +[8] Carles Gelada, Saurabh Kumar, Jacob Buckman, Ofir Nachum, and Marc G Bellemare. Deepmdp: Learning continuous latent space models for representation learning. In International Conference on Machine Learning, 2019. 8 + +[9] Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. In International Conference on Machine Learning, 2018. 6 +[10] Tuomas Haarnoja, Aurick Zhou, Kristian Hartikainen, George Tucker, Sehoon Ha, Jie Tan, Vikash Kumar, Henry Zhu, Abhishek Gupta, Pieter Abbeel, et al. Soft actor-critic algorithms and applications. arXiv preprint arXiv:1812.05905, 2018.6,7,8 +[11] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603, 2019. 3, 8 +[12] Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. In International Conference on Machine Learning, 2019. 3 +[13] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. arXiv preprint arXiv:2010.02193, 2020. 3 +[14] Nicklas Hansen, Rishabh Jangir, Yu Sun, Guillem Alenyà, Pieter Abbeel, Alexei A Efros, Lerrel Pinto, and Xiaolong Wang. Self-supervised policy adaptation during deployment. In nternational Conference on Learning Representations, 2021. 3 +[15] Nicklas Hansen, Hao Su, and Xiaolong Wang. Stabilizing deep q-learning with convnets and vision transformers under data augmentation. In Advances in Neural Information Processing Systems, 2021. 8 +[16] Nicklas Hansen and Xiaolong Wang. Generalization in reinforcement learning by soft data augmentation. In International Conference on Robotics and Automation, 2021. 9 +[17] Yangru Huang, Peixi Peng, Yifan Zhao, Guangyao Chen, and Yonghong Tian. Spectrum random masking for generalization in image-based reinforcement learning. In Advances in Neural Information Processing Systems. 3 +[18] Matthew Kowal, Mennatullah Siam, Md Amirul Islam, Neil DB Bruce, Richard P Wildes, and Konstantinos G Derpanis. A deeper dive into what deep spatiotemporal networks encode: Quantifying static vs. dynamic information. In Computer Vision and Pattern Recognition, 2022. 7 +[19] Misha Laskin, Kimin Lee, Adam Stooke, Lerrel Pinto, Pieter Abbeel, and Aravind Srinivas. Reinforcement learning with augmented data. In Advances in Neural Information Processing Systems, 2020. 2 +[20] Michael Laskin, Aravind Srinivas, and Pieter Abbeel. Curl: Contrastive unsupervised representations for reinforcement learning. In International Conference on Machine Learning, 2020. 2, 3, 5, 8 +[21] Kimin Lee, Kibok Lee, Jinwoo Shin, and Honglak Lee. Network randomization: A simple technique for generalization in deep reinforcement learning. In International Conference on Machine Learning, 2020. 3, 8 +[22] Kevin Li, Abhishek Gupta, Ashwin Reddy, Vitchyr H Pong, Aurick Zhou, Justin Yu, and Sergey Levine. Mural: Meta-learning uncertainty-aware rewards for outcome-driven reinforcement learning. In International Conference on Machine Learning, 2021. 4 + +[23] Xiang Li, Jinghuan Shang, Srijan Das, and Michael S Ryoo. Does self-supervised learning really improve reinforcement learning from pixels? In Advances in Neural Information Processing Systems, 2022. 3 +[24] Sicong Liu, Xi Sheryl Zhang, Yushuo Li, Yifan Zhang, and Jian Cheng. On the data-efficiency with contrastive image transformation in reinforcement learning. In International Conference on Learning Representations, 2023. 3 +[25] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602, 2013. 2 +[26] Thanh Nguyen, Tung M Luu, Thang Vu, and Chang D Yoo. Sample-efficient reinforcement learning representation learning with curiosity contrastive forward dynamics model. In International Conference on Intelligent Robots and Systems, 2021. 4 +[27] Minting Pan, Xiangming Zhu, Yunbo Wang, and Xiaokang Yang. Iso-dream: Isolating and leveraging noncontrollable visual dynamics in world models. In Advances in Neural Information Processing Systems, 2022. 3 +[28] Deepak Pathak, Pulkit Agrawal, Alexei A Efros, and Trevor Darrell. Curiosity-driven exploration by self-supervised prediction. In International Conference on Machine Learning, 2017. 4 +[29] Roberta Raileanu, Maxwell Goldstein, Denis Yarats, Ilya Kostrikov, and Rob Fergus. Automatic data augmentation for generalization in reinforcement learning. In Advances in Neural Information Processing Systems, 2021. 2 +[30] Max Schwarzer, Ankesh Anand, Rishab Goel, R Devon Hjelm, Aaron Courville, and Philip Bachman. Data-efficient reinforcement learning with self-predictive representations. arXiv preprint arXiv:2007.05929, 2020.8 +[31] Ramanan Sekar, Oleh Rybkin, Kostas Daniilidis, Pieter Abbeel, Danijar Hafner, and Deepak Pathak. Planning to explore via self-supervised world models. In International Conference on Machine Learning, 2020. 3 +[32] Younggyo Seo, Lili Chen, Jinwoo Shin, Honglak Lee, Pieter Abbeel, and Kimin Lee. State entropy maximization with random encoders for efficient exploration. In International Conference on Machine Learning, 2021. 4 +[33] Wenling Shang, Xiaofei Wang, Aravind Srinivas, Aravind Rajeswaran, Yang Gao, Pieter Abbeel, and Misha Laskin. Reinforcement learning with latent flow. In Advances in Neural Information Processing Systems, 2021. 3, 6, 8 +[34] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Computer Vision and Pattern Recognition, 2019. 3 +[35] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In Advances in Neural Information Processing Systems, 2014. 3 +[36] Adam Stooke, Kimin Lee, Pieter Abbeel, and Michael Laskin. Decoupling representation learning from reinforcement learning. In International Conference on Machine Learning, pages 9870-9879. PMLR, 2021. 3 +[37] Chenyu Sun, Hangwei Qian, and Chunyan Miao. Cclf: A contrastive-curiosity-driven learning framework for + +sample-efficient reinforcement learning. arXiv preprint arXiv:2205.00943, 2022. 4, 8 +[38] Yuval Tassa, Yotam Doron, Alistair Muldal, Tom Erez, Yazhe Li, Diego de Las Casas, David Budden, Abbas Abdelmaleki, Josh Merel, Andrew Lefrancq, et al. Deepmind control suite. arXiv preprint arXiv:1801.00690, 2018. 7 +[39] Anirudh Thatipelli, Sanath Narayan, Salman Khan, Rao Muhammad Anwer, Fahad Shahbaz Khan, and Bernard Ghanem. Spatio-temporal relation modeling for few-shot action recognition. In Computer Vision and Pattern Recognition, 2022. 3 +[40] Oriol Vinyals, Timo Ewalds, Sergey Bartunov, Petko Georgiev, Alexander Sasha Vezhnevets, Michelle Yeo, Alireza Makhzani, Heinrich Kuttler, John Agapiou, Julian Schrittwieser, et al. Starcraft ii: A new challenge for reinforcement learning. arXiv preprint arXiv:1708.04782, 2017. 2 +[41] Kaixin Wang, Bingyi Kang, Jie Shao, and Jiashi Feng. Improving generalization in reinforcement learning with mixture regularization. In Advances in Neural Information Processing Systems, 2020. 3 +[42] Xudong Wang, Long Lian, and Stella X Yu. Unsupervised visual attention and invariance for reinforcement learning. In Computer Vision and Pattern Recognition, 2021. 2 +[43] Wangmeng Xiang, Chao Li, Biao Wang, Xihan Wei, XianSheng Hua, and Lei Zhang. Spatiotemporal self-attention modeling with temporal patch shift for action recognition. In European Conference on Computer Vision, 2022. 3 +[44] Denis Yarats, Ilya Kostrikov, and Rob Fergus. Image augmentation is all you need: Regularizing deep reinforcement learning from pixels. In International Conference on Learning Representations, 2020. 5, 7, 8 +[45] Tao Yu, Cuiling Lan, Wenjun Zeng, Mingxiao Feng, Zhizheng Zhang, and Zhibo Chen. Playvirtual: Augmenting cycle-consistent virtual trajectories for reinforcement learning. In Advances in Neural Information Processing Systems, 2021. 8 +[46] Tao Yu, Zhizheng Zhang, Cuiling Lan, Zhibo Chen, and Yan Lu. Mask-based latent reconstruction for reinforcement learning. In Advances in Neural Information Processing Systems. 8 +[47] Amy Zhang, Rowan McAllister, Roberto Calandra, Yarin Gal, and Sergey Levine. Learning invariant representations for reinforcement learning without reconstruction. arXiv preprint arXiv:2006.10742, 2020. 5, 6, 8 +[48] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In International Conference on Machine Learning, 2019. 6 +[49] Jinhua Zhu, Yingce Xia, Lijun Wu, Jiajun Deng, Wengang Zhou, Tao Qin, Tie-Yan Liu, and Houqiang Li. Masked contrastive representation learning for reinforcement learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 3 +[50] Yuke Zhu, Josiah Wong, Ajay Mandlekar, and Roberto Martín-Martín. robosuite: A modular simulation framework and benchmark for robot learning. arXiv preprint arXiv:2009.12293, 2020. 2 \ No newline at end of file diff --git a/2023/Simoun_ Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning/images.zip b/2023/Simoun_ Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c5aa4097a46ee2ab5696d6bb952b2c33cf80e4b3 --- /dev/null +++ b/2023/Simoun_ Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:104abe406db034f49a70b2e0030e7ebd5b139b34922906307d9e328f95170362 +size 572943 diff --git a/2023/Simoun_ Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning/layout.json b/2023/Simoun_ Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..70417877f2c9621a7b243aea6256abf1f3869929 --- /dev/null +++ b/2023/Simoun_ Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning/layout.json @@ -0,0 +1,9435 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 130, + 499, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 130, + 499, + 163 + ], + "spans": [ + { + "bbox": [ + 97, + 130, + 499, + 163 + ], + "type": "text", + "content": "Simoun: Synergizing Interactive Motion-appearance Understanding for Vision-based Reinforcement Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 182, + 512, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 182, + 512, + 196 + ], + "spans": [ + { + "bbox": [ + 86, + 182, + 512, + 196 + ], + "type": "text", + "content": "Yangru Huang" + }, + { + "bbox": [ + 86, + 182, + 512, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 182, + 512, + 196 + ], + "type": "text", + "content": ", Peixi Peng" + }, + { + "bbox": [ + 86, + 182, + 512, + 196 + ], + "type": "inline_equation", + "content": "^{2,4*}" + }, + { + "bbox": [ + 86, + 182, + 512, + 196 + ], + "type": "text", + "content": ", Yifan Zhao" + }, + { + "bbox": [ + 86, + 182, + 512, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 182, + 512, + 196 + ], + "type": "text", + "content": ", Yunpeng Zhai" + }, + { + "bbox": [ + 86, + 182, + 512, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 182, + 512, + 196 + ], + "type": "text", + "content": ", Haoran Xu" + }, + { + "bbox": [ + 86, + 182, + 512, + 196 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 86, + 182, + 512, + 196 + ], + "type": "text", + "content": ", Yonghong Tian" + }, + { + "bbox": [ + 86, + 182, + 512, + 196 + ], + "type": "inline_equation", + "content": "^{1,2,4*}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 194, + 196, + 408, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 196, + 408, + 209 + ], + "spans": [ + { + "bbox": [ + 194, + 196, + 408, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 194, + 196, + 408, + 209 + ], + "type": "text", + "content": "School of Computer Science, Peking University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 149, + 209, + 449, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 209, + 449, + 220 + ], + "spans": [ + { + "bbox": [ + 149, + 209, + 449, + 220 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 209, + 449, + 220 + ], + "type": "text", + "content": "School of Electronic and Computer Engineering, Peking University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 93, + 220, + 507, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 220, + 507, + 234 + ], + "spans": [ + { + "bbox": [ + 93, + 220, + 507, + 234 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 93, + 220, + 507, + 234 + ], + "type": "text", + "content": "School of Intelligent Systems Engineering, Sun Yat-sen University " + }, + { + "bbox": [ + 93, + 220, + 507, + 234 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 93, + 220, + 507, + 234 + ], + "type": "text", + "content": "Peng Cheng Laboratory" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 236, + 522, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 236, + 522, + 247 + ], + "spans": [ + { + "bbox": [ + 77, + 236, + 522, + 247 + ], + "type": "text", + "content": "yrhuang@stu.pku.edu.cn, {pxpeng, zhaoyf, ypzhai, yhtian}@pku.edu.cn, xuhr9@mail2.sysu.edu.cn" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 159, + 271, + 202, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 271, + 202, + 282 + ], + "spans": [ + { + "bbox": [ + 159, + 271, + 202, + 282 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 70, + 295, + 290, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 295, + 290, + 544 + ], + "spans": [ + { + "bbox": [ + 70, + 295, + 290, + 544 + ], + "type": "text", + "content": "Efficient motion and appearance modeling are critical for vision-based Reinforcement Learning (RL). However, existing methods struggle to reconcile motion and appearance information within the state representations learned from a single observation encoder. To address the problem, we present Synergizing Interactive Motion-appearance Understanding (Simoun), a unified framework for vision-based RL. Given consecutive observation frames, Simoun deliberately and interactively learns both motion and appearance features through a dual-path network architecture. The learning process collaborates with a structural interactive module, which explores the latent motion-appearance structures from the two network paths to leverage their complementarity. To promote sample efficiency, we further design a consistency-guided curiosity module to encourage the exploration of under-learned observations. During training, the curiosity module provides intrinsic rewards according to the consistency of environmental temporal dynamics, which are deduced from both motion and appearance network paths. Experiments conducted on DeepMind control suite and CARLA automatic driving benchmarks demonstrate the effectiveness of Simoun, where it performs favorably against state-of-the-art methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 72, + 567, + 144, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 567, + 144, + 578 + ], + "spans": [ + { + "bbox": [ + 72, + 567, + 144, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 71, + 586, + 290, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 586, + 290, + 664 + ], + "spans": [ + { + "bbox": [ + 71, + 586, + 290, + 664 + ], + "type": "text", + "content": "Reinforcement learning (RL) from visual signals has achieved great success in recent years. Compared with learning from hand-crafted states, vision-based RL eliminates the arduous task of designing states with manual feature engineering. Therefore, it is beneficial for a variety of tasks such as video game playing [25, 40], robot manipulation [50], and autonomous navigation [4]. However, one of" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 313, + 270, + 518, + 453 + ], + "blocks": [ + { + "bbox": [ + 313, + 270, + 518, + 453 + ], + "lines": [ + { + "bbox": [ + 313, + 270, + 518, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 518, + 453 + ], + "type": "image", + "image_path": "cb2dcbe537421bf96fdf6defd6d1f075d5a29e8ffebc48afd892c01c3662331e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 456, + 524, + 506 + ], + "lines": [ + { + "bbox": [ + 306, + 456, + 524, + 506 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 524, + 506 + ], + "type": "text", + "content": "Figure 1. Illustration of different observation encoding schemes. (a) Encoding individual frames with a shared encoder. (b) Encoding stacked multiple frames. (c) Encoding individual frames and latent vector differences (latent flows). (d) Encoding motion and appearance independently and interactively." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 529, + 523, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 523, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 523, + 594 + ], + "type": "text", + "content": "the major challenges of vision-based RL lies in its high dimensional observation, which is less interpretable and leads to low rewards and inefficient sampling [19, 20]. As a result, robust visual understanding is crucial to bridge the gap between vision and state-based RL in terms of performance and sample efficiency." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 596, + 523, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 596, + 523, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 596, + 523, + 684 + ], + "type": "text", + "content": "To make observation comprehensible for agents, recent works have recognized the importance of learning high-quality visual features [20, 5, 29, 49, 42]. From the perspective of decision-making, two kinds of features are essential for vision-based RL: the motion features, which closely relate to the actions performed by the agents, and the appearance features, which describe the contemporary environmental states. Despite their importance, few works" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "spans": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 159, + 34, + 478, + 45 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 34, + 478, + 45 + ], + "spans": [ + { + "bbox": [ + 159, + 34, + 478, + 45 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 206, + 45, + 430, + 56 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 45, + 430, + 56 + ], + "spans": [ + { + "bbox": [ + 206, + 45, + 430, + 56 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 189, + 56, + 448, + 67 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 56, + 448, + 67 + ], + "spans": [ + { + "bbox": [ + 189, + 56, + 448, + 67 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 673, + 153, + 683 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 673, + 153, + 683 + ], + "spans": [ + { + "bbox": [ + 82, + 673, + 153, + 683 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 298, + 704, + 312, + 712 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 704, + 312, + 712 + ], + "spans": [ + { + "bbox": [ + 298, + 704, + 312, + 712 + ], + "type": "text", + "content": "176" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 102, + 290, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 102, + 290, + 265 + ], + "spans": [ + { + "bbox": [ + 70, + 102, + 290, + 265 + ], + "type": "text", + "content": "have tried to model these two kinds of features explicitly. Most methods tend to encode single or stacked multiple frames via a convolutional encoder, as shown in Fig. 1(a) and Fig. 1(b). The former neglects motion feature modeling, while the latter tightly entangles the motion and appearance information into a single network, causing information bias and optimization difficulty. To alleviate these issues, Shang et al. [33] propose to encode only appearance features in single frames and impose motion dynamics by latent vector differences (termed latent flow, see Fig. 1(c)). In addition, several world models learned directly from input images are also developed [11, 31, 3]. Nevertheless, the motion features of these methods are still computed from spatial latent space, which may lead to sub-optimal representation learning in complicated environments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 268, + 290, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 268, + 290, + 518 + ], + "spans": [ + { + "bbox": [ + 70, + 268, + 290, + 518 + ], + "type": "text", + "content": "To overcome the limitations described above, we present Synergizing Interactive Motion-appearance Understanding (Simoun) for vision-based RL. The design principle of Simoun is to explicitly learn both motion and appearance features in the early encoding stage and then interactively fuse them at the late stage for accurate decision-making. As shown in Fig. 1(d), given consecutive environment frames, Simoun learns motion and appearance information by separate network paths. The motion path (colored in blue) explicitly models motion clues (such as the speed and direction of cars) from the residual frame of multiple neighboring input frames. The appearance path (colored in orange) models the environmental spatial structures and focuses on identifying patterns and objects (such as cars and traffic lights) from every single frame. Additionally, a structural interactive module further extracts latent motion-appearance structures reflected by the correlations of the dual-path features. It then modulates both paths with the computed structure masks. In this way, each path can take complementary information from the other during learning, and agents are able to better understand the spatial and temporal context of the environment. Finally, the latent vectors from the two paths are fused for decision-making." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 521, + 290, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 521, + 290, + 684 + ], + "spans": [ + { + "bbox": [ + 70, + 521, + 290, + 684 + ], + "type": "text", + "content": "Although the dual-path design of Simoun promotes observation interpretability, low sample efficiency still exists due to finite data and sparse rewards. To alleviate this issue, a consistency-guided curiosity module is further designed. The idea is to adapt the learning process by concentrating more on under-learned observations, which can be deduced from the consistency of the motion and appearance paths. Intuitively, both the motion path and the appearance path describe the same observations. Hence, the dynamic information inferred from the following two sources should remain consistent: 1) latent motion vectors learned directly from the motion path and 2) differences between the appearance path latent vectors over multiple neighboring frames. If the opposite were true, then it indicates premature motion-appearance understanding, thus more explo" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 102, + 524, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 102, + 524, + 156 + ], + "spans": [ + { + "bbox": [ + 305, + 102, + 524, + 156 + ], + "type": "text", + "content": "ration should be added. In this way, we build a strong correlation between reward discovery and state novelty. During training, the consistency-guided curiosity module provides intrinsic rewards in addition to extrinsic rewards from the environment, resulting in more efficient exploration." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 157, + 524, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 157, + 524, + 200 + ], + "spans": [ + { + "bbox": [ + 305, + 157, + 524, + 200 + ], + "type": "text", + "content": "Experiments on both CARLA and OpenAI DMControl environments show that the proposed method performs favorably against state-of-the-arts. Overall, the contributions of this paper are threefold:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 201, + 523, + 308 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 201, + 523, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 201, + 523, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 201, + 523, + 232 + ], + "type": "text", + "content": "(1) We propose Simoun, a novel dual-path learning framework that explicitly and interactively learns both motion and appearance information from observations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 232, + 523, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 232, + 523, + 265 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 523, + 265 + ], + "type": "text", + "content": "(2) We design a structural interactive module to fully explore the complementarity of the two paths in Simoun and thereby further enhance visual understanding." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 265, + 523, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 265, + 523, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 265, + 523, + 308 + ], + "type": "text", + "content": "(3) We devise a consistency-guided curiosity module to encourage the exploration of under-learned observations. The proposed module effectively increases sample efficiency by providing intrinsic rewards for the agents." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 318, + 389, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 318, + 389, + 330 + ], + "spans": [ + { + "bbox": [ + 306, + 318, + 389, + 330 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 336, + 523, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 336, + 523, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 336, + 523, + 510 + ], + "type": "text", + "content": "Vision-based Reinforcement Learning To improve the performance and sample efficiency of vision-based RL, existing works can be roughly divided into three groups: 1) designing auxiliary loss/learning tasks [20, 14, 36, 49, 23], 2) employing various data augmentation techniques [21, 41, 2, 17, 24], and 3) modeling environment dynamics [12, 11, 13, 7, 27]. However, most existing works utilize a single-path network with multi-frame inputs, in which the motion and appearance information is tightly entangled without explicit separation. One exception is Flare [33], which also utilizes a single-path network to encode each frame individually and models motion information explicitly by taking latent vector differences. Although Flare achieves improved performance, its motion information is still computed from single-frame appearance features, causing insufficient temporal information extraction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 510, + 524, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 510, + 524, + 683 + ], + "spans": [ + { + "bbox": [ + 305, + 510, + 524, + 683 + ], + "type": "text", + "content": "Dual-path Networks for Visual Modeling There is a rich literature of works on visual modeling with dual-path networks. One of the earliest works is the two-stream CNNs for action recognition [35], which utilizes a spatial stream with single-frame input and a temporal stream taking multiframe optical flows. Thereafter, the concept of dual-path networks is heavily explored with various fusion strategies on different tasks [34, 6, 39, 43]. Our approach differs from existing ones in terms of task, architecture and learning mechanism. Compare with other visual tasks, vision-based RL needs to extract fine-grained motion details across different time steps, which poses a great challenge to RL models. To the best of our knowledge, Simoun is one of the first vision-based RL methods which aim to explicitly learn motion and appearance features by a dual-path architecture. Meanwhile, its structural interactive module is" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 704, + 312, + 713 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 704, + 312, + 713 + ], + "spans": [ + { + "bbox": [ + 298, + 704, + 312, + 713 + ], + "type": "text", + "content": "177" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 102, + 512, + 300 + ], + "blocks": [ + { + "bbox": [ + 83, + 102, + 512, + 300 + ], + "lines": [ + { + "bbox": [ + 83, + 102, + 512, + 300 + ], + "spans": [ + { + "bbox": [ + 83, + 102, + 512, + 300 + ], + "type": "image", + "image_path": "5df02ca57eb18b68f9e9bf127e3ec6b2a9ad426b9cc2caa043529ed1fba78b4d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 302, + 525, + 334 + ], + "lines": [ + { + "bbox": [ + 70, + 302, + 525, + 334 + ], + "spans": [ + { + "bbox": [ + 70, + 302, + 525, + 334 + ], + "type": "text", + "content": "Figure 2. The proposed dual-path learning framework Simoun. It introduces two modules: 1) a structural interactive module that leverages the complementarity between the two network paths by extracting latent motion-appearance structures, and 2) a consistency-guided curiosity module that provides intrinsic rewards according to the dynamic consistency between the motion and appearance features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 350, + 290, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 350, + 290, + 403 + ], + "spans": [ + { + "bbox": [ + 70, + 350, + 290, + 403 + ], + "type": "text", + "content": "also deliberately designed to extract motion and appearance clues residing in consecutive observation frames. By learning with RL-oriented objectives and curiosity-driven strategies, Simoun successfully imports the idea of dual-path networks into robust visual control." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 404, + 290, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 404, + 290, + 600 + ], + "spans": [ + { + "bbox": [ + 70, + 404, + 290, + 600 + ], + "type": "text", + "content": "Intrinsic Reward Exploration One of the key elements of RL is the reward function, which aims to quantify the \"goodness\" of the agent's decisions. However, the problem is that designing dense, well-defined extrinsic reward functions is difficult and unscalable. One possible solution is to introduce an intrinsic reward function, which is calculated by the agents. Existing intrinsic reward exploration methods mainly focus on counting or predicting state novelty [1], prediction error [28], uncertainty [22], or environmental dynamics [32]. These methods are typically designed for the general state-based instead of vision-based RL. The works most relevant to us are CCFDM [26] and CCLF [37], which also formulate intrinsic rewards for visual-based RL. CCFDM utilizes forward dynamics and CCLF is based on a contrastive term. Differently, the consistency-guided curiosity module in Simoun exploits the dynamic consistency extracted from both motion and appearance features to achieve reliable intrinsic reward estimation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 610, + 147, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 610, + 147, + 624 + ], + "spans": [ + { + "bbox": [ + 71, + 610, + 147, + 624 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 628, + 290, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 628, + 290, + 684 + ], + "spans": [ + { + "bbox": [ + 70, + 628, + 290, + 684 + ], + "type": "text", + "content": "We start by formalizing the task of vision-based RL in Sec. 3.1 and then delineate Simoun in detail. The general idea of Simoun is to 1) explicitly model the motion/appearance dynamics (and their correlations) of the agent's operating ambiance and 2) capitalize on the learned" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 350, + 525, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 525, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 525, + 460 + ], + "type": "text", + "content": "dynamics to achieve efficient exploration. As shown in Fig. 2, our framework consists of two parts. First, given inputs in the form of consecutive frames, it interactively models the motion and appearance features using a dualpath architecture with a set of targeted objectives and a structural interactive module (Sec. 3.2). Then the decision-making strategy is learned with adaptive curiosity assignment steered by the proposed consistency-guided curiosity module (Sec. 3.3). The overall learning objective and inference process of Simoun are finally summarized in Sec. 3.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 464, + 416, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 416, + 474 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 416, + 474 + ], + "type": "text", + "content": "3.1. Problem formulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "text", + "content": "Vision-based RL can be formulated as a Partially Observable Markov Decision Process (POMDP) " + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = < \\mathcal{O},\\mathcal{A},\\mathcal{P},\\mathcal{R},\\gamma >" + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "text", + "content": " denotes the observation space containing pixel frames " + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_t" + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "text", + "content": " at different time step " + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "text", + "content": " denotes the action space. At each " + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "text", + "content": " the agent chooses an action " + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "inline_equation", + "content": "a_{t}\\in \\mathcal{A}" + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\mathbf{o}_{t + 1}|\\mathbf{o}_t,a_t)" + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "text", + "content": " is the observation transition, " + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\mathbf{o}_t,a_t)" + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "text", + "content": " is the reward function, and " + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "inline_equation", + "content": "\\gamma \\in [0,1)" + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "text", + "content": " is the discount factor. The goal is to identify an optimal policy " + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 475, + 524, + 593 + ], + "type": "text", + "content": " that maximizes the expected cumulative reward based on the visual observations rather than the complete environment state:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 350, + 594, + 523, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 594, + 523, + 616 + ], + "spans": [ + { + "bbox": [ + 350, + 594, + 523, + 616 + ], + "type": "interline_equation", + "content": "J (\\pi) = \\sum_ {t} \\mathbb {E} _ {\\left(\\mathbf {o} _ {t}, a _ {t}\\right) \\sim \\pi} \\left[ \\mathcal {R} \\left(\\mathbf {o} _ {t}, a _ {t}\\right) \\right]. \\tag {1}", + "image_path": "e4698a28236680c81adf795d41492fae1bd57ba2512911537e4beb9003f9098e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 618, + 523, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 523, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 523, + 640 + ], + "type": "text", + "content": "During training, " + }, + { + "bbox": [ + 304, + 618, + 523, + 640 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 618, + 523, + 640 + ], + "type": "text", + "content": " is used to interact with the environment, and the related data are stored in a replay buffer " + }, + { + "bbox": [ + 304, + 618, + 523, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 618, + 523, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 644, + 463, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 644, + 463, + 657 + ], + "spans": [ + { + "bbox": [ + 306, + 644, + 463, + 657 + ], + "type": "text", + "content": "3.2. Dual-path Interactive Modeling" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 662, + 523, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 662, + 523, + 684 + ], + "spans": [ + { + "bbox": [ + 305, + 662, + 523, + 684 + ], + "type": "text", + "content": "In vision-based RL, motion and appearance clues are both vital for the agent to perform accurate decision" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 704, + 312, + 712 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 704, + 312, + 712 + ], + "spans": [ + { + "bbox": [ + 298, + 704, + 312, + 712 + ], + "type": "text", + "content": "178" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 102, + 290, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 102, + 290, + 167 + ], + "spans": [ + { + "bbox": [ + 70, + 102, + 290, + 167 + ], + "type": "text", + "content": "making. Simoun adopts a dual-path architecture to interactively learn motion and appearance features. Specifically, it consists of three components: the motion path, the appearance path, and the structural interactive module that serves as a switchboard between the motion and appearance structures learned from the two paths." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 170, + 290, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 170, + 290, + 343 + ], + "spans": [ + { + "bbox": [ + 70, + 170, + 290, + 343 + ], + "type": "text", + "content": "Motion Path Motion information is crucial for the visual system to understand the dynamics of the surrounding environment. The motion path aims to extract the motion-related features (such as the velocity of moving objects) from the changes between consecutive frames. Given a tuple of three adjacent observations " + }, + { + "bbox": [ + 70, + 170, + 290, + 343 + ], + "type": "inline_equation", + "content": "[\\mathbf{o}_{t - 2},\\mathbf{o}_{t - 1},\\mathbf{o}_t]" + }, + { + "bbox": [ + 70, + 170, + 290, + 343 + ], + "type": "text", + "content": " sampled from replay buffer " + }, + { + "bbox": [ + 70, + 170, + 290, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 70, + 170, + 290, + 343 + ], + "type": "text", + "content": ", the input of the motion path is the residual of adjacent frames " + }, + { + "bbox": [ + 70, + 170, + 290, + 343 + ], + "type": "inline_equation", + "content": "[(0_{t - 1} - 0_{t - 2}),(0_t - 0_{t - 1})]" + }, + { + "bbox": [ + 70, + 170, + 290, + 343 + ], + "type": "text", + "content": ", concatenated along channel dimensions. A convolutional encoder " + }, + { + "bbox": [ + 70, + 170, + 290, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^m" + }, + { + "bbox": [ + 70, + 170, + 290, + 343 + ], + "type": "text", + "content": " is then used to learn a lower-dimensional motion representation. Specifically, the encoder includes four convolution layers with " + }, + { + "bbox": [ + 70, + 170, + 290, + 343 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 70, + 170, + 290, + 343 + ], + "type": "text", + "content": " kernel size and ReLU nonlinearity. Denote the feature map of the last convolution layer as " + }, + { + "bbox": [ + 70, + 170, + 290, + 343 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^m\\in \\mathbb{R}^{c\\times h\\times w}" + }, + { + "bbox": [ + 70, + 170, + 290, + 343 + ], + "type": "text", + "content": ", a fully connected (FC) layer with layer normalization (LN) is used to reduce the dimension of " + }, + { + "bbox": [ + 70, + 170, + 290, + 343 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^m" + }, + { + "bbox": [ + 70, + 170, + 290, + 343 + ], + "type": "text", + "content": " to get the motion feature vector " + }, + { + "bbox": [ + 70, + 170, + 290, + 343 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^m" + }, + { + "bbox": [ + 70, + 170, + 290, + 343 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 346, + 290, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 346, + 290, + 421 + ], + "spans": [ + { + "bbox": [ + 70, + 346, + 290, + 421 + ], + "type": "text", + "content": "To further impel the motion path to catch abundant environment temporal structures, a motion-aware constraint is deliberately designed. Specifically, given " + }, + { + "bbox": [ + 70, + 346, + 290, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^m" + }, + { + "bbox": [ + 70, + 346, + 290, + 421 + ], + "type": "text", + "content": " and corresponding action " + }, + { + "bbox": [ + 70, + 346, + 290, + 421 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 70, + 346, + 290, + 421 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 70, + 346, + 290, + 421 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 70, + 346, + 290, + 421 + ], + "type": "text", + "content": ", an action-conditioned two-FC-layer transition model " + }, + { + "bbox": [ + 70, + 346, + 290, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 70, + 346, + 290, + 421 + ], + "type": "text", + "content": " is used to obtain a motion-action joint representation of the current timestep. A latent-space transition loss is then defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 423, + 289, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 423, + 289, + 436 + ], + "spans": [ + { + "bbox": [ + 123, + 423, + 289, + 436 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t r a n} = \\left\\| \\mathcal {G} \\left(\\mathbf {f} _ {t} ^ {m}, u _ {t}\\right) - \\mathbf {f} _ {t + 1} ^ {m} \\right\\| _ {2} ^ {2}. \\tag {2}", + "image_path": "69414ccd601e7848e9c3502a026a2fbd91bbb32ced94952d627e160f31f6b683.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 437, + 289, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 437, + 289, + 471 + ], + "spans": [ + { + "bbox": [ + 70, + 437, + 289, + 471 + ], + "type": "text", + "content": "which encourages " + }, + { + "bbox": [ + 70, + 437, + 289, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^m" + }, + { + "bbox": [ + 70, + 437, + 289, + 471 + ], + "type": "text", + "content": " to model the motion trends over time by predicting the future temporal dynamics " + }, + { + "bbox": [ + 70, + 437, + 289, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{t + 1}^{m}" + }, + { + "bbox": [ + 70, + 437, + 289, + 471 + ], + "type": "text", + "content": " of the next time step." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 472, + 290, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 472, + 290, + 570 + ], + "spans": [ + { + "bbox": [ + 70, + 472, + 290, + 570 + ], + "type": "text", + "content": "Appearance Path The appearance path is designed to extract the visual appearance representations of the object and scene from individual observation frames. The input of the appearance path is a single frame " + }, + { + "bbox": [ + 70, + 472, + 290, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_t" + }, + { + "bbox": [ + 70, + 472, + 290, + 570 + ], + "type": "text", + "content": " from the replay buffer " + }, + { + "bbox": [ + 70, + 472, + 290, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 70, + 472, + 290, + 570 + ], + "type": "text", + "content": ". Similar to the motion path, the appearance path also adopts a four-layer encoder " + }, + { + "bbox": [ + 70, + 472, + 290, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^a" + }, + { + "bbox": [ + 70, + 472, + 290, + 570 + ], + "type": "text", + "content": " with the same architecture (except for the number of input channels of the first layer) to get feature map " + }, + { + "bbox": [ + 70, + 472, + 290, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^a\\in \\mathbb{R}^{c\\times h\\times w}" + }, + { + "bbox": [ + 70, + 472, + 290, + 570 + ], + "type": "text", + "content": ", followed by FC and LN layers to get the appearance representation " + }, + { + "bbox": [ + 70, + 472, + 290, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^a" + }, + { + "bbox": [ + 70, + 472, + 290, + 570 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 572, + 290, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 572, + 290, + 659 + ], + "spans": [ + { + "bbox": [ + 70, + 572, + 290, + 659 + ], + "type": "text", + "content": "To explicitly make appearance representations discriminative to different scenes, we take inspiration from CURL [20] and adopt an unsupervised contrastive loss between similar and dissimilar sample pairs. Given observation frame " + }, + { + "bbox": [ + 70, + 572, + 290, + 659 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_i" + }, + { + "bbox": [ + 70, + 572, + 290, + 659 + ], + "type": "text", + "content": ", we consider " + }, + { + "bbox": [ + 70, + 572, + 290, + 659 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{o}}_i" + }, + { + "bbox": [ + 70, + 572, + 290, + 659 + ], + "type": "text", + "content": " (obtained from " + }, + { + "bbox": [ + 70, + 572, + 290, + 659 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_i" + }, + { + "bbox": [ + 70, + 572, + 290, + 659 + ], + "type": "text", + "content": " by data augmentation) as the positive sample, while samples coming from different observation frames are regarded as negatives. The contrastive loss is then defined as:" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 312, + 102, + 519, + 248 + ], + "blocks": [ + { + "bbox": [ + 312, + 102, + 519, + 248 + ], + "lines": [ + { + "bbox": [ + 312, + 102, + 519, + 248 + ], + "spans": [ + { + "bbox": [ + 312, + 102, + 519, + 248 + ], + "type": "image", + "image_path": "8b95db8882fb3931dede5d8f3b89fbe4514957d962f6cbf4ae2300b8e67efd28.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 250, + 523, + 261 + ], + "lines": [ + { + "bbox": [ + 306, + 250, + 523, + 261 + ], + "spans": [ + { + "bbox": [ + 306, + 250, + 523, + 261 + ], + "type": "text", + "content": "Figure 3. Illustration of the proposed structural interactive module." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 293, + 523, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 293, + 523, + 323 + ], + "spans": [ + { + "bbox": [ + 317, + 293, + 523, + 323 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c o n} = - \\log \\frac {\\exp \\left(\\mathbf {f} _ {i} ^ {a ^ {\\top}} \\tilde {\\mathbf {f}} _ {i} ^ {a}\\right)}{\\exp \\left(\\mathbf {f} _ {i} ^ {a ^ {\\top}} \\tilde {\\mathbf {f}} _ {i} ^ {a}\\right) + \\sum_ {j = 0} ^ {K - 1} \\exp \\left(\\mathbf {f} _ {i} ^ {a ^ {\\top}} \\tilde {\\mathbf {f}} _ {j} ^ {a}\\right)}, \\tag {3}", + "image_path": "ad1aa55a9202432042c9106afe641e1a77a7aa3db423631cc6ffb14c03db2359.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 324, + 523, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 523, + 401 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 523, + 401 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 324, + 523, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^a" + }, + { + "bbox": [ + 305, + 324, + 523, + 401 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 324, + 523, + 401 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{f}}_t^a" + }, + { + "bbox": [ + 305, + 324, + 523, + 401 + ], + "type": "text", + "content": " are appearance representations of " + }, + { + "bbox": [ + 305, + 324, + 523, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_t" + }, + { + "bbox": [ + 305, + 324, + 523, + 401 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 324, + 523, + 401 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{o}}_t" + }, + { + "bbox": [ + 305, + 324, + 523, + 401 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 305, + 324, + 523, + 401 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 324, + 523, + 401 + ], + "type": "text", + "content": " is the sample index in a training batch with " + }, + { + "bbox": [ + 305, + 324, + 523, + 401 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 324, + 523, + 401 + ], + "type": "text", + "content": " samples. Note that the augmented data is only used for representation learning and is inaccessible during decision-making. This is to avoid unstable training, which is easily caused by the augmented data dominating the evaluation of the Q-target." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 403, + 523, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 403, + 523, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 403, + 523, + 533 + ], + "type": "text", + "content": "Structural Interactive Module Instead of simply concatenating the features from motion and appearance paths as state representations, interactively communicating the learned motion-appearance structures between both paths can lead to a more robust visual understanding. Intuitively, as shown in Fig. 3(left), given consecutive observation frames, the motion structure can be revealed by inter-frame appearance variation. Meanwhile, the appearance structure can be obtained by spatial pixel correlation. Therefore, the structural interactive module embraces a single and interframe relation discovering mechanism for efficient motion-appearance structure mining and propagation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 535, + 523, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 535, + 523, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 535, + 523, + 610 + ], + "type": "text", + "content": "Specifically, as shown in Fig. 3(right), the inputs of the structural interactive module are motion feature map " + }, + { + "bbox": [ + 306, + 535, + 523, + 610 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^m" + }, + { + "bbox": [ + 306, + 535, + 523, + 610 + ], + "type": "text", + "content": " and appearance feature maps " + }, + { + "bbox": [ + 306, + 535, + 523, + 610 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{t - 2}^{a},\\mathbf{F}_{t - 1}^{a}" + }, + { + "bbox": [ + 306, + 535, + 523, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 535, + 523, + 610 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^a" + }, + { + "bbox": [ + 306, + 535, + 523, + 610 + ], + "type": "text", + "content": " .To extract motion structures, an inter-frame attention mechanism is designed. In particular, " + }, + { + "bbox": [ + 306, + 535, + 523, + 610 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{t - 1}^{a}" + }, + { + "bbox": [ + 306, + 535, + 523, + 610 + ], + "type": "text", + "content": " is treated as the query, and " + }, + { + "bbox": [ + 306, + 535, + 523, + 610 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{t - 2}^{a},\\mathbf{F}_{t}^{a}" + }, + { + "bbox": [ + 306, + 535, + 523, + 610 + ], + "type": "text", + "content": " is used as two keys. A inter-frame attention map " + }, + { + "bbox": [ + 306, + 535, + 523, + 610 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 306, + 535, + 523, + 610 + ], + "type": "text", + "content": " can be obtained as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 337, + 612, + 523, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 612, + 523, + 626 + ], + "spans": [ + { + "bbox": [ + 337, + 612, + 523, + 626 + ], + "type": "interline_equation", + "content": "\\mathbf {X} = \\sigma \\left(\\sigma \\left(\\hat {\\mathbf {F}} _ {t - 2} ^ {a} ^ {\\top} \\hat {\\mathbf {F}} _ {t - 1} ^ {a}\\right) + \\sigma \\left(\\hat {\\mathbf {F}} _ {t} ^ {a \\top} \\hat {\\mathbf {F}} _ {t - 1} ^ {a}\\right)\\right), \\tag {4}", + "image_path": "e4a3dd3ef0db592fa83e701ce0b12977bd07d53af5780241e571737e7269fea1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 628, + 523, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 628, + 523, + 684 + ], + "spans": [ + { + "bbox": [ + 305, + 628, + 523, + 684 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 628, + 523, + 684 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_t^a\\in \\mathbb{R}^{c\\times h\\times w}" + }, + { + "bbox": [ + 305, + 628, + 523, + 684 + ], + "type": "text", + "content": " denotes a new feature map generated by feeding " + }, + { + "bbox": [ + 305, + 628, + 523, + 684 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^a" + }, + { + "bbox": [ + 305, + 628, + 523, + 684 + ], + "type": "text", + "content": " to a convolution layer for reducing the spatial complexity (" + }, + { + "bbox": [ + 305, + 628, + 523, + 684 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{t - 1}^a" + }, + { + "bbox": [ + 305, + 628, + 523, + 684 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 628, + 523, + 684 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{t - 2}^a" + }, + { + "bbox": [ + 305, + 628, + 523, + 684 + ], + "type": "text", + "content": " can be obtained similarly). " + }, + { + "bbox": [ + 305, + 628, + 523, + 684 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 305, + 628, + 523, + 684 + ], + "type": "text", + "content": " denotes the Softmax function. Combining the outputs of two initial Softmax functions might not yield values sum" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 82, + 673, + 250, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 673, + 250, + 684 + ], + "spans": [ + { + "bbox": [ + 82, + 673, + 250, + 684 + ], + "type": "text", + "content": "We choose three here following the practice in [44, 47]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 298, + 704, + 312, + 712 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 704, + 312, + 712 + ], + "spans": [ + { + "bbox": [ + 298, + 704, + 312, + 712 + ], + "type": "text", + "content": "179" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 101, + 282, + 248 + ], + "blocks": [ + { + "bbox": [ + 78, + 101, + 282, + 248 + ], + "lines": [ + { + "bbox": [ + 78, + 101, + 282, + 248 + ], + "spans": [ + { + "bbox": [ + 78, + 101, + 282, + 248 + ], + "type": "image", + "image_path": "50e6f72ebcab46802578cdc37e638c109921863340f1079d7e2328bd703cc8ee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 254, + 277, + 264 + ], + "lines": [ + { + "bbox": [ + 83, + 254, + 277, + 264 + ], + "spans": [ + { + "bbox": [ + 83, + 254, + 277, + 264 + ], + "type": "text", + "content": "Figure 4. The idea of consistency-guided curiosity module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 71, + 281, + 289, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 281, + 289, + 357 + ], + "spans": [ + { + "bbox": [ + 71, + 281, + 289, + 357 + ], + "type": "text", + "content": "ming to 1, so an additional Softmax is used to normalize the final attention scores. The soft indexing of " + }, + { + "bbox": [ + 71, + 281, + 289, + 357 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{t - 1}^{a}" + }, + { + "bbox": [ + 71, + 281, + 289, + 357 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 71, + 281, + 289, + 357 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{t - 2}^{a}" + }, + { + "bbox": [ + 71, + 281, + 289, + 357 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 71, + 281, + 289, + 357 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_t^a" + }, + { + "bbox": [ + 71, + 281, + 289, + 357 + ], + "type": "text", + "content": " is performed at the spatial dimensions, resulting in a soft attention map of dimensions " + }, + { + "bbox": [ + 71, + 281, + 289, + 357 + ], + "type": "inline_equation", + "content": "hw\\times hw" + }, + { + "bbox": [ + 71, + 281, + 289, + 357 + ], + "type": "text", + "content": ". Then a spatial gating operation is applied on " + }, + { + "bbox": [ + 71, + 281, + 289, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 71, + 281, + 289, + 357 + ], + "type": "text", + "content": " using weights calculated from both " + }, + { + "bbox": [ + 71, + 281, + 289, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^m" + }, + { + "bbox": [ + 71, + 281, + 289, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 71, + 281, + 289, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^a" + }, + { + "bbox": [ + 71, + 281, + 289, + 357 + ], + "type": "text", + "content": " to consolidate appearance structures and obtain two motion-appearance structure masks:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 361, + 288, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 361, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 123, + 361, + 288, + 373 + ], + "type": "interline_equation", + "content": "\\mathbf {X} ^ {m} = \\mathbf {X} \\cdot \\hat {\\mathbf {F}} _ {t} ^ {m}, \\mathbf {X} ^ {a} = \\mathbf {X} \\cdot \\hat {\\mathbf {F}} _ {t} ^ {a}, \\tag {5}", + "image_path": "6187d5cb078d9dfe92692bba449625facb1cc73439e52619cfaca09f9fab7562.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 374, + 289, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 374, + 289, + 417 + ], + "spans": [ + { + "bbox": [ + 71, + 374, + 289, + 417 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 71, + 374, + 289, + 417 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_t^m" + }, + { + "bbox": [ + 71, + 374, + 289, + 417 + ], + "type": "text", + "content": " is a feature map of " + }, + { + "bbox": [ + 71, + 374, + 289, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^m" + }, + { + "bbox": [ + 71, + 374, + 289, + 417 + ], + "type": "text", + "content": " obtained with another convolution layer. Finally, " + }, + { + "bbox": [ + 71, + 374, + 289, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^m" + }, + { + "bbox": [ + 71, + 374, + 289, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 71, + 374, + 289, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^a" + }, + { + "bbox": [ + 71, + 374, + 289, + 417 + ], + "type": "text", + "content": " are enhanced by modulating with their corresponding motion-appearance structure masks:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 102, + 420, + 288, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 420, + 288, + 432 + ], + "spans": [ + { + "bbox": [ + 102, + 420, + 288, + 432 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {t} ^ {m} = \\mathbf {F} _ {t} ^ {m} + \\boldsymbol {\\beta} \\cdot \\mathbf {X} ^ {m}, \\mathbf {F} _ {t} ^ {a} = \\mathbf {F} _ {t} ^ {a} + \\boldsymbol {\\beta} \\cdot \\mathbf {X} ^ {a}, \\tag {6}", + "image_path": "23803145428c03021efcbf889fcad9da6cfef3f249fc6d203377a2cbacecbbda.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 71, + 434, + 289, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 434, + 289, + 467 + ], + "spans": [ + { + "bbox": [ + 71, + 434, + 289, + 467 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 71, + 434, + 289, + 467 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 71, + 434, + 289, + 467 + ], + "type": "text", + "content": " is a learnable parameter initialized as zero [48], and the enhanced " + }, + { + "bbox": [ + 71, + 434, + 289, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^m" + }, + { + "bbox": [ + 71, + 434, + 289, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 71, + 434, + 289, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^a" + }, + { + "bbox": [ + 71, + 434, + 289, + 467 + ], + "type": "text", + "content": " instead of the original ones are further used to calculate " + }, + { + "bbox": [ + 71, + 434, + 289, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^m" + }, + { + "bbox": [ + 71, + 434, + 289, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 71, + 434, + 289, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^a" + }, + { + "bbox": [ + 71, + 434, + 289, + 467 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 71, + 468, + 290, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 468, + 290, + 574 + ], + "spans": [ + { + "bbox": [ + 71, + 468, + 290, + 574 + ], + "type": "text", + "content": "After acquiring the structure-enhanced motion and appearance features, they are finally fused as " + }, + { + "bbox": [ + 71, + 468, + 290, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t = [\\mathbf{f}_t^m,\\mathbf{f}_t^a ]" + }, + { + "bbox": [ + 71, + 468, + 290, + 574 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 71, + 468, + 290, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t" + }, + { + "bbox": [ + 71, + 468, + 290, + 574 + ], + "type": "text", + "content": " is used as the final state representation. To better capture reward-related features from the fused state representation, a reward function " + }, + { + "bbox": [ + 71, + 468, + 290, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 71, + 468, + 290, + 574 + ], + "type": "text", + "content": " is further introduced to predict a numerical reward value to each state-action pair. " + }, + { + "bbox": [ + 71, + 468, + 290, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 71, + 468, + 290, + 574 + ], + "type": "text", + "content": " has a similar architecture with the transition model " + }, + { + "bbox": [ + 71, + 468, + 290, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 71, + 468, + 290, + 574 + ], + "type": "text", + "content": " except the output dimension is set to one. A reward loss " + }, + { + "bbox": [ + 71, + 468, + 290, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{re}" + }, + { + "bbox": [ + 71, + 468, + 290, + 574 + ], + "type": "text", + "content": " is then defined as the mean squared error between the predicted and actual reward:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 129, + 577, + 289, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 577, + 289, + 590 + ], + "spans": [ + { + "bbox": [ + 129, + 577, + 289, + 590 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e} = \\left(\\mathcal {R} \\left(\\mathbf {f} _ {t}, a _ {t}\\right) - r _ {t + 1} ^ {e}\\right) ^ {2}, \\tag {7}", + "image_path": "6bc221b13bc86ab1191fedf0a4cfd7699f02dd0804831d15dd0ef12bf2bc4e43.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 71, + 592, + 289, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 592, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 71, + 592, + 289, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 71, + 592, + 289, + 613 + ], + "type": "inline_equation", + "content": "r_{t + 1}^{e}" + }, + { + "bbox": [ + 71, + 592, + 289, + 613 + ], + "type": "text", + "content": " is the actual external reward value at the next time step, which is returned by the environment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 71, + 620, + 253, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 620, + 253, + 632 + ], + "spans": [ + { + "bbox": [ + 71, + 620, + 253, + 632 + ], + "type": "text", + "content": "3.3. Consistency-guided Curiosity Module" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 71, + 636, + 289, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 636, + 289, + 658 + ], + "spans": [ + { + "bbox": [ + 71, + 636, + 289, + 658 + ], + "type": "text", + "content": "Given state representation " + }, + { + "bbox": [ + 71, + 636, + 289, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t" + }, + { + "bbox": [ + 71, + 636, + 289, + 658 + ], + "type": "text", + "content": " from the dual-path model, we adapt SAC [9, 10] as the base RL algorithm following" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 102, + 523, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 102, + 523, + 146 + ], + "spans": [ + { + "bbox": [ + 306, + 102, + 523, + 146 + ], + "type": "text", + "content": "previous methods [47, 5], which aims to maximize the expected cumulative reward to find an optimal policy by approximating the action-value function " + }, + { + "bbox": [ + 306, + 102, + 523, + 146 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 306, + 102, + 523, + 146 + ], + "type": "text", + "content": " and a stochastic policy " + }, + { + "bbox": [ + 306, + 102, + 523, + 146 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 306, + 102, + 523, + 146 + ], + "type": "text", + "content": " based on a " + }, + { + "bbox": [ + 306, + 102, + 523, + 146 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 306, + 102, + 523, + 146 + ], + "type": "text", + "content": "-discounted maximum entropy " + }, + { + "bbox": [ + 306, + 102, + 523, + 146 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\cdot)" + }, + { + "bbox": [ + 306, + 102, + 523, + 146 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 148, + 523, + 170 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 148, + 523, + 170 + ], + "spans": [ + { + "bbox": [ + 318, + 148, + 523, + 170 + ], + "type": "interline_equation", + "content": "J (\\pi) = \\sum_ {t} \\mathbb {E} _ {\\left(\\mathbf {o} _ {t}, a _ {t}\\right) \\sim \\pi} \\left[ r \\left(\\mathbf {o} _ {t}, a _ {t}\\right) + \\alpha \\mathcal {H} \\left(\\pi \\left(\\cdot \\mid \\mathbf {o} _ {t}\\right)\\right) \\right]. \\tag {8}", + "image_path": "6442d1edbe8a6f1e49dc180764f533cbee061ef9fc16477eb81d30956e8fa77e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 171, + 523, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 171, + 523, + 192 + ], + "spans": [ + { + "bbox": [ + 306, + 171, + 523, + 192 + ], + "type": "text", + "content": "The action-value function " + }, + { + "bbox": [ + 306, + 171, + 523, + 192 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 306, + 171, + 523, + 192 + ], + "type": "text", + "content": " are learned by minimizing the soft Bellman error:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 194, + 523, + 208 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 194, + 523, + 208 + ], + "spans": [ + { + "bbox": [ + 321, + 194, + 523, + 208 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {Q} = \\mathbb {E} _ {\\left(\\mathbf {o} _ {t}, a _ {t}\\right)} \\left(Q \\left(\\mathbf {o} _ {t}, a _ {t}\\right) - \\left(r _ {t} + \\lambda V \\left(\\mathbf {o} _ {t + 1}\\right)\\right)\\right) ^ {2}, \\tag {9}", + "image_path": "c8947ce13c38549213cf0d686dff0a8fa795c378ffbe0eb2a6137f474c842da4.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 209, + 523, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 209, + 523, + 231 + ], + "spans": [ + { + "bbox": [ + 306, + 209, + 523, + 231 + ], + "type": "text", + "content": "and the soft state value " + }, + { + "bbox": [ + 306, + 209, + 523, + 231 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 306, + 209, + 523, + 231 + ], + "type": "text", + "content": " can be estimated by sampling an action under the current policy:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 245, + 523, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 245, + 523, + 268 + ], + "spans": [ + { + "bbox": [ + 306, + 245, + 523, + 268 + ], + "type": "interline_equation", + "content": "V \\left(\\mathbf {o} _ {t + 1}\\right) = \\mathbb {E} _ {a _ {t + 1} \\sim \\pi} \\left[ \\bar {Q} \\left(\\mathbf {o} _ {t + 1}; a _ {t + 1}\\right) - \\alpha \\log \\pi \\left(a _ {t + 1} \\mid \\mathbf {o} _ {t + 1}\\right) \\right], \\tag {10}", + "image_path": "b9308aabb18611168a2acb00f84473dbf3010914663bc358a65335d1fd483f2b.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 268, + 523, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 268, + 523, + 311 + ], + "spans": [ + { + "bbox": [ + 306, + 268, + 523, + 311 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 268, + 523, + 311 + ], + "type": "inline_equation", + "content": "\\overline{Q}" + }, + { + "bbox": [ + 306, + 268, + 523, + 311 + ], + "type": "text", + "content": " denotes the exponential moving average of the parameters of " + }, + { + "bbox": [ + 306, + 268, + 523, + 311 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 306, + 268, + 523, + 311 + ], + "type": "text", + "content": ". The policy is optimized by decreasing the difference between the exponential of the soft-Q function and the policy:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 337, + 313, + 523, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 313, + 523, + 331 + ], + "spans": [ + { + "bbox": [ + 337, + 313, + 523, + 331 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\pi} = \\mathbb {E} _ {a _ {t} \\sim \\pi} \\left[ \\alpha \\log \\pi \\left(a _ {t} \\mid \\mathbf {o} _ {t}\\right) - Q \\left(\\mathbf {o} _ {t}, a _ {t}\\right) \\right]. \\tag {11}", + "image_path": "58cef388d7e18b13109081bd73d6acbf1c8698753c7e9af625b4eb2d1bd3d0b9.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 335, + 523, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 335, + 523, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 523, + 453 + ], + "type": "text", + "content": "Although SAC algorithm introduces the entropy term to encourage exploration, it still heavily relies on carefully engineered environmental extrinsic rewards and suffers from low sample efficiency. Instead of extrinsic rewards, intrinsic curiosity can be a powerful concept to endow an agent with an automated mechanism to continuously explore its environment in the absence of task information. During training, in addition to the extrinsic rewards " + }, + { + "bbox": [ + 305, + 335, + 523, + 453 + ], + "type": "inline_equation", + "content": "r^e" + }, + { + "bbox": [ + 305, + 335, + 523, + 453 + ], + "type": "text", + "content": " obtained from the environment, the curiosity module also provides intrinsic rewards " + }, + { + "bbox": [ + 305, + 335, + 523, + 453 + ], + "type": "inline_equation", + "content": "r^i" + }, + { + "bbox": [ + 305, + 335, + 523, + 453 + ], + "type": "text", + "content": ". The learning objective SAC in Eq. 8 is therefore extended as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 453, + 529, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 529, + 484 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 529, + 484 + ], + "type": "interline_equation", + "content": "J (\\pi) = \\sum_ {t} \\mathbb {E} _ {\\left(\\mathbf {o} _ {t}, a _ {t}\\right) \\sim \\pi} \\left[ r ^ {e} \\left(\\mathbf {o} _ {t}, a _ {t}\\right) + \\alpha \\mathcal {H} \\left(\\pi (\\cdot | \\mathbf {o} _ {t})\\right) + r ^ {i} \\left(\\mathbf {o} _ {t}, a _ {t}\\right) \\right]. \\tag {12}", + "image_path": "d0c1ddd3e48017c4512246c5a2b3e10febddfaf3bb470226aa9178a302a41b9b.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 486, + 523, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 486, + 523, + 637 + ], + "spans": [ + { + "bbox": [ + 305, + 486, + 523, + 637 + ], + "type": "text", + "content": "To get proper intrinsic rewards " + }, + { + "bbox": [ + 305, + 486, + 523, + 637 + ], + "type": "inline_equation", + "content": "r^i" + }, + { + "bbox": [ + 305, + 486, + 523, + 637 + ], + "type": "text", + "content": ", we propose to leverage the consistency of motion and appearance features from the dual-path model. Intuitively, as shown in Fig. 4, the environmental temporal dynamics can be obtained from two sources: 1) directly learned from the motion path, and 2) through the variations of the spatial features learned from the spatial path, similar to the latent flow [33]. Optimally, these two sources should be consistent with each other. That is, the motion feature " + }, + { + "bbox": [ + 305, + 486, + 523, + 637 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^m" + }, + { + "bbox": [ + 305, + 486, + 523, + 637 + ], + "type": "text", + "content": " should be similar to the difference between the spatial features " + }, + { + "bbox": [ + 305, + 486, + 523, + 637 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{t - 2}^a" + }, + { + "bbox": [ + 305, + 486, + 523, + 637 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 486, + 523, + 637 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{t - 1}^a" + }, + { + "bbox": [ + 305, + 486, + 523, + 637 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 486, + 523, + 637 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^a" + }, + { + "bbox": [ + 305, + 486, + 523, + 637 + ], + "type": "text", + "content": ". The idea of the consistency-guided curiosity module is then to encourage the agent to explore when the temporal dynamics produced by the two sources are inconsistent. To this end, the intrinsic reward " + }, + { + "bbox": [ + 305, + 486, + 523, + 637 + ], + "type": "inline_equation", + "content": "r^i" + }, + { + "bbox": [ + 305, + 486, + 523, + 637 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 310, + 638, + 523, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 638, + 523, + 662 + ], + "spans": [ + { + "bbox": [ + 310, + 638, + 523, + 662 + ], + "type": "interline_equation", + "content": "r ^ {i} = C e ^ {- \\lambda t} d \\left(\\left\\| \\mathbf {f} _ {t} ^ {m} \\right\\|. \\left| \\mathbf {f} _ {t - 1} ^ {a} - \\mathbf {f} _ {t - 2} ^ {a} \\right| + \\left| \\mathbf {f} _ {t} ^ {a} - \\mathbf {f} _ {t - 1} ^ {a} \\right|\\right) \\frac {r _ {\\max} ^ {e}}{t _ {\\max} ^ {i}}, \\tag {13}", + "image_path": "aa11634dc834a77cf9f423a5e277f80cbc7fc1fc9b15360dd3e5b27053499282.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 662, + 523, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 662, + 523, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 662, + 523, + 683 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 662, + 523, + 683 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 306, + 662, + 523, + 683 + ], + "type": "text", + "content": " is temperature weight, " + }, + { + "bbox": [ + 306, + 662, + 523, + 683 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 306, + 662, + 523, + 683 + ], + "type": "text", + "content": " is an exponential decay weight, " + }, + { + "bbox": [ + 306, + 662, + 523, + 683 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 306, + 662, + 523, + 683 + ], + "type": "text", + "content": " is the L2 distance function, " + }, + { + "bbox": [ + 306, + 662, + 523, + 683 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 306, + 662, + 523, + 683 + ], + "type": "text", + "content": " is the environment" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 71, + 664, + 289, + 683 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 664, + 289, + 683 + ], + "spans": [ + { + "bbox": [ + 71, + 664, + 289, + 683 + ], + "type": "text", + "content": "2We use concatenation to fuse " + }, + { + "bbox": [ + 71, + 664, + 289, + 683 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^m" + }, + { + "bbox": [ + 71, + 664, + 289, + 683 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 71, + 664, + 289, + 683 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^a" + }, + { + "bbox": [ + 71, + 664, + 289, + 683 + ], + "type": "text", + "content": " for simplicity, advanced fusion mechanisms are also applicable." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 299, + 704, + 312, + 712 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 704, + 312, + 712 + ], + "spans": [ + { + "bbox": [ + 299, + 704, + 312, + 712 + ], + "type": "text", + "content": "180" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 77, + 114, + 287, + 312 + ], + "blocks": [ + { + "bbox": [ + 73, + 102, + 235, + 113 + ], + "lines": [ + { + "bbox": [ + 73, + 102, + 235, + 113 + ], + "spans": [ + { + "bbox": [ + 73, + 102, + 235, + 113 + ], + "type": "text", + "content": "Algorithm 1 Inference procedure of Simoun" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "lines": [ + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "text", + "content": "1: for each environment step " + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "text", + "content": " do \n2: Collect observation frames " + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_{t-2}" + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_{t-1}" + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_t" + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "text", + "content": ". \n3: Extract motion and appearance feature maps: \n4: " + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^m = \\mathcal{E}^m ([\\mathbf{o}_{t-1} - \\mathbf{o}_{t-2}), (\\mathbf{o}_t - \\mathbf{o}_{t-1})]" + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "text", + "content": ", \n5: " + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_j^a = \\mathcal{E}^a (\\mathbf{o}_j)" + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "inline_equation", + "content": "j = t - 2, \\dots, t" + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "text", + "content": " \n6: Structural Interactive Modeling: \n7: " + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = \\sigma (\\sigma (\\hat{\\mathbf{F}}_{t-2}^a \\top \\hat{\\mathbf{F}}_{t-1}^a) + \\sigma (\\hat{\\mathbf{F}}_t^{a\\top}\\hat{\\mathbf{F}}_{t-1}^a))" + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^m = \\mathbf{F}_t^m + \\beta \\cdot (\\mathbf{X} \\cdot \\hat{\\mathbf{F}}_t^m)" + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "text", + "content": " \n9: " + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^a = \\mathbf{F}_t^a + \\beta \\cdot (\\mathbf{X} \\cdot \\hat{\\mathbf{F}}_t^a)" + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "text", + "content": " \n10: Extract low-dimensional features: \n11: " + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^m = \\mathrm{LayerNorm}(\\mathrm{FC}(\\mathbf{F}_t^m))" + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "text", + "content": " \n12: " + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^a = \\mathrm{LayerNorm}(\\mathrm{FC}(\\mathbf{F}_t^a))" + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "text", + "content": " \n13: Take action based on the fused features: \n14: " + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t = [\\mathbf{f}_t^m, \\mathbf{f}_t^a]" + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "text", + "content": " \n15: " + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "inline_equation", + "content": "a_t \\sim \\pi (a_t | \\mathbf{f}_t)" + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "text", + "content": " \n16: Environment state transition: \n17: " + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_{t+1} \\sim \\mathcal{P}(\\mathbf{o}_{t+1} | \\mathbf{o}_t, a_t)" + }, + { + "bbox": [ + 77, + 114, + 287, + 312 + ], + "type": "text", + "content": " \n18: end for" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "spans": [ + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "type": "text", + "content": "time step, " + }, + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "type": "inline_equation", + "content": "|\\cdot |" + }, + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "type": "text", + "content": " is the vector element-wise absolute operation (used to avoid the cancellation of " + }, + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{t - 1}^s" + }, + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "type": "inline_equation", + "content": "r_{max}^{e}" + }, + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "type": "inline_equation", + "content": "r_{max}^{i}" + }, + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "type": "text", + "content": " are the maximum extrinsic reward value and intrinsic reward value over the time step " + }, + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "type": "inline_equation", + "content": "[0,t]" + }, + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "type": "text", + "content": ", which serve as normalization term to balance " + }, + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "type": "inline_equation", + "content": "r^i" + }, + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "type": "inline_equation", + "content": "r^e" + }, + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "type": "text", + "content": ". At the beginning of the training, " + }, + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "type": "inline_equation", + "content": "r_{max}^{e}" + }, + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "type": "inline_equation", + "content": "r_{max}^{i}" + }, + { + "bbox": [ + 70, + 332, + 290, + 452 + ], + "type": "text", + "content": " are assigned with an initial value and dynamically updated as the training goes on. By rewarding the agent to explore motion-appearance inconsistent states, the curiosity module encourages the agent to learn more efficiently, leading to better performance and fewer learning steps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 458, + 207, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 458, + 207, + 470 + ], + "spans": [ + { + "bbox": [ + 71, + 458, + 207, + 470 + ], + "type": "text", + "content": "3.4. Overall Training Objective" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 475, + 290, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 475, + 290, + 508 + ], + "spans": [ + { + "bbox": [ + 70, + 475, + 290, + 508 + ], + "type": "text", + "content": "By capitalizing on interactive motion-appearance understanding, Simoun learns from vision to control in an end-to-end manner by optimizing the following training objective:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 521, + 289, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 521, + 289, + 545 + ], + "spans": [ + { + "bbox": [ + 88, + 521, + 289, + 545 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\underbrace {\\mathcal {L} _ {t r a n}} _ {M o t i o n} + \\underbrace {\\mathcal {L} _ {c o n}} _ {A p p e a r a n c e} + \\underbrace {\\mathcal {L} _ {r e}} _ {S t a t e} + \\underbrace {\\mathcal {L} _ {Q} + \\mathcal {L} _ {\\pi}} _ {R L}, \\tag {14}", + "image_path": "f761f2c500b64b6720ed5d3ec4a9c61085121b5716b65368ecb3bdf9c7abb472.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 547, + 290, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 547, + 290, + 590 + ], + "spans": [ + { + "bbox": [ + 70, + 547, + 290, + 590 + ], + "type": "text", + "content": "where the objective jointly considers the learning of motion and appearance features, as well as getting efficient state representation for high-performance RL policy learning. The detailed inference procedure is presented in Alg. 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 71, + 600, + 272, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 600, + 272, + 612 + ], + "spans": [ + { + "bbox": [ + 71, + 600, + 272, + 612 + ], + "type": "text", + "content": "4. An Information Bias Analysis of Simoun" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 618, + 290, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 618, + 290, + 684 + ], + "spans": [ + { + "bbox": [ + 70, + 618, + 290, + 684 + ], + "type": "text", + "content": "As we demonstrated in Fig. 1, existing observation encoding paradigms model motion and appearance information either heuristically without special design or using preliminary techniques (such as taking latent difference). In contrast, Simoun deliberately models both information and emphasizes their interactions. In this section, we analyze" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 100, + 523, + 221 + ], + "blocks": [ + { + "bbox": [ + 306, + 100, + 523, + 221 + ], + "lines": [ + { + "bbox": [ + 306, + 100, + 523, + 221 + ], + "spans": [ + { + "bbox": [ + 306, + 100, + 523, + 221 + ], + "type": "image", + "image_path": "f7a4511aa94bf9946b7464cbb16926a6e095b96eb6b75af418a55486e34c5051.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 224, + 524, + 274 + ], + "lines": [ + { + "bbox": [ + 305, + 224, + 524, + 274 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 524, + 274 + ], + "type": "text", + "content": "Figure 5. Feature component analysis of different models. M-Path and A-path represent the motion and appearance paths of Simoun, respectively. Strong information bias can be observed from the baseline models and Simoun notably alleviates the bias and decreases meaningless residual units." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 293, + 523, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 523, + 326 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 523, + 326 + ], + "type": "text", + "content": "the representations learned by Simoun and other observation encoding paradigms from an information bias perspective to gain an in-depth understanding of them." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 326, + 523, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 326, + 523, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 326, + 523, + 555 + ], + "type": "text", + "content": "Specifically, we leverage a recent approach [18] that can quantify static and dynamic information learned by any spatial-temporal model. The approach estimates the amount of static vs. dynamic bias based on the mutual information between sampled input sequence pairs. It then calculates the percentage of units (channel dimensions) of the model feature layer that encodes several pre-defined information factors (static, dynamic, joint, and residual). The quantifying results of Simoun and other models are illustrated in Fig. 5. It is immediately evident that previous methods tend to encode the static factor more than the dynamic, indicating a strong bias toward appearance information. It is also clear that the two paths of Simoun learn corresponding static and dynamic factors as expected, with the motion path having more dynamic units than the appearance path. By fusing the two paths together, Simoun enjoys abundant dynamic and static information, meanwhile having minimal residual units that do not involve any dynamic or static factor. As will be shown in the experiment section, such an abundant and informative representation significantly improves the decision-making process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 566, + 379, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 566, + 379, + 578 + ], + "spans": [ + { + "bbox": [ + 306, + 566, + 379, + 578 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 585, + 523, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 585, + 523, + 639 + ], + "spans": [ + { + "bbox": [ + 305, + 585, + 523, + 639 + ], + "type": "text", + "content": "In this section, we explore how Simoun can improve vision-based RL in terms of sample efficiency and performance gains. Two benchmarks are used for evaluation: DeepMind Control Suite (DMControl) [38] for continuous control and CARLA [4] for autonomous driving." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 639, + 523, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 639, + 523, + 684 + ], + "spans": [ + { + "bbox": [ + 305, + 639, + 523, + 684 + ], + "type": "text", + "content": "Experimental Settings: Simoun is implemented on the basis of the SAC algorithm [10]. For DMControl, to avoid the potential effects of different hyperparameters, we follow the previous training setup of DrQ [44] and choose" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 704, + 312, + 712 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 704, + 312, + 712 + ], + "spans": [ + { + "bbox": [ + 299, + 704, + 312, + 712 + ], + "type": "text", + "content": "181" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 100, + 523, + 244 + ], + "blocks": [ + { + "bbox": [ + 74, + 100, + 523, + 244 + ], + "lines": [ + { + "bbox": [ + 74, + 100, + 523, + 244 + ], + "spans": [ + { + "bbox": [ + 74, + 100, + 523, + 244 + ], + "type": "table", + "html": "
100K Step ScoresSACDreamerCURLDrQSVEASPRPlayVirtualMLRCCLFOurs
Cartpole, Swingup237±49326±27582±146759±92727±86799±42816±36806±48799±61851±14
Reacher, Easy239±183314±155538±233601±213811±115638±269785±142866±103738±99910±65
Cheetah, Run118±13235±137299±48344±67375±54467±36474±50482±38317±38518±57
Walker, Walk95±19277±12403±24612±164747±65398±165460±173643±114648±110869±45
Finger, Spin230±194341±70767±56901±104859±77868±143915±49907±58994±42872±90
Ball in cup, Catch85±130246±174769±43913±53915±71861±233929±31933±16914±20971±17
Average167.3289.8559.7688.3739.0719.0729.8772.8735.0831.8
500k Step Scores
Cartpole, Swingup330±73762±27841±45868±10865±10870±12865±11872±5869±9876±4
Reacher, Easy307±65793±164929±44942±71944±52925±79942±66957±41941±48979±44
Cheetah, Run85±51570±253518±28660±96682±65716±47719±51674±37588±22693±53
Walker, Walk71±52897±49902±43921±45919±24916±75928±30939±10936±23966±7
Finger, Spin346±95796±183926±45938±103924±93924±132963±40973±31974±6964±19
Ball in cup, Catch162±122879±87959±27963±9960±19963±8967±5964±14961±9970±12
Average216.8782.8845.8882.0882.3920.0897.3896.5878.2908.0
", + "image_path": "df26181c36d88852960075c7c23a6d3c1c839defd084608aafa7ebb89e8be530.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 75, + 268, + 520, + 327 + ], + "blocks": [ + { + "bbox": [ + 111, + 250, + 485, + 259 + ], + "lines": [ + { + "bbox": [ + 111, + 250, + 485, + 259 + ], + "spans": [ + { + "bbox": [ + 111, + 250, + 485, + 259 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-art methods on DMControl benchmark. The best results are shown in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 268, + 520, + 327 + ], + "lines": [ + { + "bbox": [ + 75, + 268, + 520, + 327 + ], + "spans": [ + { + "bbox": [ + 75, + 268, + 520, + 327 + ], + "type": "table", + "html": "
MeasurementsSACCURLDrQFlareDeepMDPOurs
Episode reward ↑121 ±26.1134 ±15.1154 ±21.5132 ±24.7170 ±36.1281 ±30.4
Distance (m) ↑74 ±17.4128 ±32.595 ±27.290 ±14.6132 ±20.4207 ±15.5
Crash intensity ↓3930 ±80.33050 ±100.32419 ±72.32668 ±95.72136 ±69.31813 ±57.8
Average steer ↓17.52% ±0.021%16.60% ±0.025%15.79% ±0.018%11.48% ±0.022%10.22% ±0.015%14.84% ±0.012%
Average brake ↓1.81% ±0.013%2.94% ±0.021%1.70% ±0.039%1.52% ±0.014%1.65% ±0.007%2.14% ±0.018%
", + "image_path": "d27ae9b6e8da07bb8388bd7aed9f94604db9edcb865ad6d9df0354b8ad910fad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 334, + 507, + 344 + ], + "lines": [ + { + "bbox": [ + 87, + 334, + 507, + 344 + ], + "spans": [ + { + "bbox": [ + 87, + 334, + 507, + 344 + ], + "type": "text", + "content": "Table 2. Comparison with state-of-the-art methods on CARLA benchmark. " + }, + { + "bbox": [ + 87, + 334, + 507, + 344 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 87, + 334, + 507, + 344 + ], + "type": "text", + "content": " indicates larger is better and " + }, + { + "bbox": [ + 87, + 334, + 507, + 344 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 87, + 334, + 507, + 344 + ], + "type": "text", + "content": " means the opposite." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 70, + 362, + 290, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 362, + 290, + 514 + ], + "spans": [ + { + "bbox": [ + 70, + 362, + 290, + 514 + ], + "type": "text", + "content": "six commonly adopted tasks: Walker-walk, Finger-spin, Cartpole-swingup, Reacher-easy, Cheetah-run and Ball in cup-catch. To evaluate sample efficiency, the performance at 100k and 500k environment steps are reported during the training stage. For CARLA, we mostly follow the setup of DBC [47], where the goal is to travel as far as possible on Highway 8 of Town 4 in 1000 time steps without any collisions with 20 moving cars. For observation acquisition, we horizontally concatenate the images from three cameras on the roof of vehicles to get " + }, + { + "bbox": [ + 70, + 362, + 290, + 514 + ], + "type": "inline_equation", + "content": "84 \\times 252" + }, + { + "bbox": [ + 70, + 362, + 290, + 514 + ], + "type": "text", + "content": " images. Random convolution [21] is adopted for data augmentation in Eq. 3. All experiments are trained across 5 random seeds to report the mean and standard deviation of the rewards. More details can be found in the Supplementary Material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 514, + 290, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 514, + 290, + 602 + ], + "spans": [ + { + "bbox": [ + 70, + 514, + 290, + 602 + ], + "type": "text", + "content": "Methods Compared: We extensively compare Simoun with a variety of methods including the SAC [10] baseline, explicit motion modeling approach (Flare [33]), auxiliary loss-based approaches (CURL [20], MLR [46]), data augmentation approaches (DrQ [44], SVEA [15]), dynamic modeling based approaches (Dreamer [11], SPR [30], PlayVirtual [45], DeepMDP [8]), and curiosity based approach (CCLF [37])." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 71, + 611, + 206, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 611, + 206, + 623 + ], + "spans": [ + { + "bbox": [ + 71, + 611, + 206, + 623 + ], + "type": "text", + "content": "5.1. Performance Comparisons" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 628, + 290, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 628, + 290, + 684 + ], + "spans": [ + { + "bbox": [ + 70, + 628, + 290, + 684 + ], + "type": "text", + "content": "Results on DMControl Table 1 presents the experimental results on the DMControl benchmark. It can be observed that Simoun achieves considerable performance gains compared to other state-of-the-art methods. In particular, at " + }, + { + "bbox": [ + 70, + 628, + 290, + 684 + ], + "type": "inline_equation", + "content": "100\\mathrm{k}" + }, + { + "bbox": [ + 70, + 628, + 290, + 684 + ], + "type": "text", + "content": " steps, significant performance improvement can al" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 362, + 524, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 362, + 524, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 362, + 524, + 384 + ], + "type": "text", + "content": "ready be reached by Simoun, which indicates improved sample efficiency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 384, + 524, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 524, + 483 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 524, + 483 + ], + "type": "text", + "content": "Results on CARLA The results of the CARLA benchmark are reported in Table 2. It is clear that Simoun outperforms all other methods on the episode reward. Additionally, the average driving distance is farther than other methods by a large margin and the average collision intensity is also smaller. Although the driving smoothness of Simoun is slightly decreased due to increased steer and brake, this small cost has led to considerable overall reward gain to break through the current status quo." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 491, + 391, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 391, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 391, + 504 + ], + "type": "text", + "content": "5.2. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 510, + 525, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 510, + 525, + 685 + ], + "spans": [ + { + "bbox": [ + 305, + 510, + 525, + 685 + ], + "type": "text", + "content": "Effectiveness of the Dual-path Design To demonstrate the effectiveness of the dual-path design in Simoun, we compare it with three single-path methods (depicted in Fig. 1): individual frame encoding, stacked frames encoding, latent flow encoding, and our dual-path encoding. For a fair comparison, the motion and appearance losses of Simoun are adopted for all four models to eliminate the affection of loss difference, and the interactive and curiosity modules are also removed from the dual-path model. The results are shown in Fig. 6 (left). Several observations can be made: 1) The low performance of individual frame encoding (black line) indicates the importance of modeling motion information. 2) By considering the motion across frames, stacked frame encoding (green line) performs much better than individual frame encoding. 3) The latent flow encoding (blue line) improves over stacked frame encoding" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 704, + 312, + 712 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 704, + 312, + 712 + ], + "spans": [ + { + "bbox": [ + 299, + 704, + 312, + 712 + ], + "type": "text", + "content": "182" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 99, + 225, + 207 + ], + "blocks": [ + { + "bbox": [ + 78, + 99, + 225, + 207 + ], + "lines": [ + { + "bbox": [ + 78, + 99, + 225, + 207 + ], + "spans": [ + { + "bbox": [ + 78, + 99, + 225, + 207 + ], + "type": "image", + "image_path": "dae08905ec192cb390fd34f8dd63b02476edbfc943e7da8d50c213bf4b24f7ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 209, + 523, + 231 + ], + "lines": [ + { + "bbox": [ + 71, + 209, + 523, + 231 + ], + "spans": [ + { + "bbox": [ + 71, + 209, + 523, + 231 + ], + "type": "text", + "content": "Figure 6. Experiment results on CARLA benchmark. Left: Performance on different encoding paradigms. Middle: Ablation study of each component of Simoun. Right: Comparison with models having more parameters." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 100, + 370, + 207 + ], + "blocks": [ + { + "bbox": [ + 225, + 100, + 370, + 207 + ], + "lines": [ + { + "bbox": [ + 225, + 100, + 370, + 207 + ], + "spans": [ + { + "bbox": [ + 225, + 100, + 370, + 207 + ], + "type": "image", + "image_path": "de0bff32cca3531b6dccaf71c9f90a2fd7330246cac5cd0d8595ea3c0e0c6b7c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 370, + 100, + 516, + 207 + ], + "blocks": [ + { + "bbox": [ + 370, + 100, + 516, + 207 + ], + "lines": [ + { + "bbox": [ + 370, + 100, + 516, + 207 + ], + "spans": [ + { + "bbox": [ + 370, + 100, + 516, + 207 + ], + "type": "image", + "image_path": "2faa47d456470448a415fffdbb2e78cd8757516ec1271af109a5ceb4629d802b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 76, + 245, + 282, + 332 + ], + "blocks": [ + { + "bbox": [ + 76, + 245, + 282, + 332 + ], + "lines": [ + { + "bbox": [ + 76, + 245, + 282, + 332 + ], + "spans": [ + { + "bbox": [ + 76, + 245, + 282, + 332 + ], + "type": "image", + "image_path": "38e26d4d4f00d87a6dd7210bb57d54996e55d596a1de5a983744d555af1641d1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 335, + 280, + 344 + ], + "lines": [ + { + "bbox": [ + 72, + 335, + 280, + 344 + ], + "spans": [ + { + "bbox": [ + 72, + 335, + 280, + 344 + ], + "type": "text", + "content": "Figure 7. Visualizations of motion-appearance structure masks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 366, + 289, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 366, + 289, + 475 + ], + "spans": [ + { + "bbox": [ + 70, + 366, + 289, + 475 + ], + "type": "text", + "content": "with explicit motion modeling. However, there is limited room for further improvement due to the preliminary technique used for motion extraction. 4) The proposed dualpath model (red line) remarkably outperforms the other three methods by modeling motion and appearance explicitly. Interestingly, the results in Fig. 6 (left) echo perfectly with the information bias degrees illustrated and discussed in Fig. 5 and Sec. 4, which indicates potentially deeper connections between motion-appearance information bias and the performance of decision-making." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 477, + 290, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 477, + 290, + 684 + ], + "spans": [ + { + "bbox": [ + 70, + 477, + 290, + 684 + ], + "type": "text", + "content": "Effectiveness of Each Components in Simoun To investigate the contribution of each component in Simoun, we first evaluate the performance of each individual path, then test the dual-path model by gradually adding the structural interactive module and the consistency-guided curiosity module. It can be found in Fig. 6 (middle) that the individual path gives relatively low performance when trained separately, with the motion path performing better than the appearance path. When adopting the dual-path model, both the structural interactive module and the consistency-guided curiosity module can further improve performance, which demonstrates their effectiveness. To better understand which visual clues does Simoun concentrate, we visualize the motion-appearance structure masks " + }, + { + "bbox": [ + 70, + 477, + 290, + 684 + ], + "type": "inline_equation", + "content": "(\\mathbf{X}^m" + }, + { + "bbox": [ + 70, + 477, + 290, + 684 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 70, + 477, + 290, + 684 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^a" + }, + { + "bbox": [ + 70, + 477, + 290, + 684 + ], + "type": "text", + "content": " in Eq. 5) of the two paths. As can be observed from Fig. 7, the motion path tends to focus on the moving trajectory of other vehicles (positions where the vehicles passed by), while the appearance path focuses strongly on spatial positions where the other vehicles exist." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 326, + 245, + 504, + 342 + ], + "blocks": [ + { + "bbox": [ + 326, + 245, + 504, + 342 + ], + "lines": [ + { + "bbox": [ + 326, + 245, + 504, + 342 + ], + "spans": [ + { + "bbox": [ + 326, + 245, + 504, + 342 + ], + "type": "image", + "image_path": "98f7364f51785f3792cc7fd6ef0a01c0f0501203727accabd61b31b666a9c0b9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 345, + 524, + 365 + ], + "lines": [ + { + "bbox": [ + 306, + 345, + 524, + 365 + ], + "spans": [ + { + "bbox": [ + 306, + 345, + 524, + 365 + ], + "type": "text", + "content": "Figure 8. Evaluation of the generalization ability on four different environment domains on DMControl benchmark." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 393, + 411, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 393, + 411, + 404 + ], + "spans": [ + { + "bbox": [ + 306, + 393, + 411, + 404 + ], + "type": "text", + "content": "5.3. Further Discussions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 417, + 524, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 417, + 524, + 536 + ], + "spans": [ + { + "bbox": [ + 305, + 417, + 524, + 536 + ], + "type": "text", + "content": "Does the performance gain of the dual-path model come from the increased network parameters? To answer this question, we compare the dual-path model with a double-channeled stacked frame model, which has nearly " + }, + { + "bbox": [ + 305, + 417, + 524, + 536 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 305, + 417, + 524, + 536 + ], + "type": "text", + "content": " more parameters. From Fig. 6 (right) we can observe that increasing model parameters indeed improves performance. However, the dual-path model still outperforms the double-channeled stacked frames model using only half of its parameters. This proves the effectiveness of the dualpath model mainly comes from its motion-appearance modeling paradigm, rather than increased network capacity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 542, + 524, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 542, + 524, + 684 + ], + "spans": [ + { + "bbox": [ + 305, + 542, + 524, + 684 + ], + "type": "text", + "content": "Does Simoun improves domain generalization? To evaluate the domain generalization ability of Simoun, we select the \"catch\" task on the \"ball in cup\" scenario as the source domain and test Simoun on DMControl Generalization Benchmark [16] with four different environment domains (color easy, color hard, video easy, and video hard). Fig. 8 shows that Simoun performs on par with other methods on color-shifted domains. However, the performance is much better on video-shifted domains, where the background is also moving. This is attributed to the specialized modeling of motion information by Simoun, which drives the agent to pay more attention to reward-related motions rather than the irrelevant dynamic background." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 704, + 312, + 712 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 704, + 312, + 712 + ], + "spans": [ + { + "bbox": [ + 299, + 704, + 312, + 712 + ], + "type": "text", + "content": "183" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 72, + 102, + 138, + 112 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 102, + 138, + 112 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 138, + 112 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 71, + 114, + 290, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 114, + 290, + 255 + ], + "spans": [ + { + "bbox": [ + 71, + 114, + 290, + 255 + ], + "type": "text", + "content": "We have proposed Simoun, a unified framework for vision-based RL with a dual-path network for motion and appearance understanding. The design of Simoun demonstrates the effectiveness of motion-appearance structural interaction, and further shows the benefits of consistency-guided intrinsic curiosity. Empirical results suggest that the proposed method has advantages in terms of sample efficiency, performance gains, and generalization ability. By analyzing Simoun from an information bias perspective, we build a connection between motion-appearance information bias and vision-based RL performance. We hope this connection can further inspire more efficient model designs for vision-based RL tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 72, + 265, + 171, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 265, + 171, + 276 + ], + "spans": [ + { + "bbox": [ + 72, + 265, + 171, + 276 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 277, + 290, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 277, + 290, + 343 + ], + "spans": [ + { + "bbox": [ + 71, + 277, + 290, + 343 + ], + "type": "text", + "content": "The study was funded by the Key-Area Research and Development Program of Guangdong Province with contract No. 2020B0101380001, as well as the National Natural Science Foundation of China under contracts No. 62027804, No. 61825101, No. 62088102 and No. 62202010. Computing support was provided by Pengcheng Cloudbrain." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 352, + 125, + 362 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 352, + 125, + 362 + ], + "spans": [ + { + "bbox": [ + 72, + 352, + 125, + 362 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 370, + 290, + 691 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 370, + 290, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 370, + 290, + 410 + ], + "spans": [ + { + "bbox": [ + 77, + 370, + 290, + 410 + ], + "type": "text", + "content": "[1] Marc Bellemare, Sriram Srinivasan, Georg Ostrovski, Tom Schaul, David Saxton, and Remi Munos. Unifying count-based exploration and intrinsic motivation. In Advances in Neural Information Processing Systems, 2016. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 411, + 289, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 411, + 289, + 449 + ], + "spans": [ + { + "bbox": [ + 77, + 411, + 289, + 449 + ], + "type": "text", + "content": "[2] David Bertoin and Emmanuel Rachelson. Local feature swapping for generalization in reinforcement learning. In International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 77, + 451, + 290, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 451, + 290, + 490 + ], + "spans": [ + { + "bbox": [ + 77, + 451, + 290, + 490 + ], + "type": "text", + "content": "[3] Fei Deng, Ingook Jang, and Sungjin Ahn. Dreamerpro: Reconstruction-free model-based reinforcement learning with prototypical representations. In International Conference on Machine Learning, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 491, + 289, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 491, + 289, + 530 + ], + "spans": [ + { + "bbox": [ + 77, + 491, + 289, + 530 + ], + "type": "text", + "content": "[4] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on Robot Learning, pages 1-16. PMLR, 2017. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 532, + 289, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 532, + 289, + 581 + ], + "spans": [ + { + "bbox": [ + 77, + 532, + 289, + 581 + ], + "type": "text", + "content": "[5] Linxi Fan, Guanzhi Wang, De-An Huang, Zhiding Yu, Li Fei-Fei, Yuke Zhu, and Animashree Anandkumar. Secant: Self-expert cloning for zero-shot generalization of visual policies. In International Conference on Machine Learning, 2021. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 77, + 582, + 289, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 582, + 289, + 620 + ], + "spans": [ + { + "bbox": [ + 77, + 582, + 289, + 620 + ], + "type": "text", + "content": "[6] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition. In Computer Vision and Pattern Recognition, 2016. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 77, + 622, + 289, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 622, + 289, + 652 + ], + "spans": [ + { + "bbox": [ + 77, + 622, + 289, + 652 + ], + "type": "text", + "content": "[7] Xiang Fu, Ge Yang, Pulkit Agrawal, and Tommi Jaakkola. Learning task informed abstractions. In International Conference on Machine Learning, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 653, + 289, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 653, + 289, + 691 + ], + "spans": [ + { + "bbox": [ + 77, + 653, + 289, + 691 + ], + "type": "text", + "content": "[8] Carles Gelada, Saurabh Kumar, Jacob Buckman, Ofir Nachum, and Marc G Bellemare. Deepmdp: Learning continuous latent space models for representation learning. In International Conference on Machine Learning, 2019. 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 102, + 523, + 691 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 311, + 102, + 523, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 102, + 523, + 142 + ], + "spans": [ + { + "bbox": [ + 311, + 102, + 523, + 142 + ], + "type": "text", + "content": "[9] Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. In International Conference on Machine Learning, 2018. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 144, + 523, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 144, + 523, + 193 + ], + "spans": [ + { + "bbox": [ + 307, + 144, + 523, + 193 + ], + "type": "text", + "content": "[10] Tuomas Haarnoja, Aurick Zhou, Kristian Hartikainen, George Tucker, Sehoon Ha, Jie Tan, Vikash Kumar, Henry Zhu, Abhishek Gupta, Pieter Abbeel, et al. Soft actor-critic algorithms and applications. arXiv preprint arXiv:1812.05905, 2018.6,7,8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 194, + 523, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 523, + 233 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 523, + 233 + ], + "type": "text", + "content": "[11] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603, 2019. 3, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 235, + 523, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 235, + 523, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 235, + 523, + 275 + ], + "type": "text", + "content": "[12] Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. In International Conference on Machine Learning, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 276, + 523, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 523, + 305 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 523, + 305 + ], + "type": "text", + "content": "[13] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. arXiv preprint arXiv:2010.02193, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 306, + 523, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 306, + 523, + 355 + ], + "spans": [ + { + "bbox": [ + 308, + 306, + 523, + 355 + ], + "type": "text", + "content": "[14] Nicklas Hansen, Rishabh Jangir, Yu Sun, Guillem Alenyà, Pieter Abbeel, Alexei A Efros, Lerrel Pinto, and Xiaolong Wang. Self-supervised policy adaptation during deployment. In nternational Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 357, + 523, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 357, + 523, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 357, + 523, + 396 + ], + "type": "text", + "content": "[15] Nicklas Hansen, Hao Su, and Xiaolong Wang. Stabilizing deep q-learning with convnets and vision transformers under data augmentation. In Advances in Neural Information Processing Systems, 2021. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 397, + 523, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 397, + 523, + 427 + ], + "spans": [ + { + "bbox": [ + 308, + 397, + 523, + 427 + ], + "type": "text", + "content": "[16] Nicklas Hansen and Xiaolong Wang. Generalization in reinforcement learning by soft data augmentation. In International Conference on Robotics and Automation, 2021. 9" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 428, + 523, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 428, + 523, + 468 + ], + "spans": [ + { + "bbox": [ + 308, + 428, + 523, + 468 + ], + "type": "text", + "content": "[17] Yangru Huang, Peixi Peng, Yifan Zhao, Guangyao Chen, and Yonghong Tian. Spectrum random masking for generalization in image-based reinforcement learning. In Advances in Neural Information Processing Systems. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 468, + 523, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 523, + 518 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 523, + 518 + ], + "type": "text", + "content": "[18] Matthew Kowal, Mennatullah Siam, Md Amirul Islam, Neil DB Bruce, Richard P Wildes, and Konstantinos G Derpanis. A deeper dive into what deep spatiotemporal networks encode: Quantifying static vs. dynamic information. In Computer Vision and Pattern Recognition, 2022. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 520, + 523, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 520, + 523, + 559 + ], + "spans": [ + { + "bbox": [ + 308, + 520, + 523, + 559 + ], + "type": "text", + "content": "[19] Misha Laskin, Kimin Lee, Adam Stooke, Lerrel Pinto, Pieter Abbeel, and Aravind Srinivas. Reinforcement learning with augmented data. In Advances in Neural Information Processing Systems, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 559, + 523, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 523, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 523, + 599 + ], + "type": "text", + "content": "[20] Michael Laskin, Aravind Srinivas, and Pieter Abbeel. Curl: Contrastive unsupervised representations for reinforcement learning. In International Conference on Machine Learning, 2020. 2, 3, 5, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 600, + 523, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 600, + 523, + 640 + ], + "spans": [ + { + "bbox": [ + 308, + 600, + 523, + 640 + ], + "type": "text", + "content": "[21] Kimin Lee, Kibok Lee, Jinwoo Shin, and Honglak Lee. Network randomization: A simple technique for generalization in deep reinforcement learning. In International Conference on Machine Learning, 2020. 3, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 641, + 523, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 641, + 523, + 691 + ], + "spans": [ + { + "bbox": [ + 308, + 641, + 523, + 691 + ], + "type": "text", + "content": "[22] Kevin Li, Abhishek Gupta, Ashwin Reddy, Vitchyr H Pong, Aurick Zhou, Justin Yu, and Sergey Levine. Mural: Meta-learning uncertainty-aware rewards for outcome-driven reinforcement learning. In International Conference on Machine Learning, 2021. 4" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 704, + 312, + 712 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 704, + 312, + 712 + ], + "spans": [ + { + "bbox": [ + 299, + 704, + 312, + 712 + ], + "type": "text", + "content": "184" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 72, + 102, + 290, + 685 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 73, + 102, + 290, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 102, + 290, + 142 + ], + "spans": [ + { + "bbox": [ + 73, + 102, + 290, + 142 + ], + "type": "text", + "content": "[23] Xiang Li, Jinghuan Shang, Srijan Das, and Michael S Ryoo. Does self-supervised learning really improve reinforcement learning from pixels? In Advances in Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 72, + 144, + 289, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 144, + 289, + 183 + ], + "spans": [ + { + "bbox": [ + 72, + 144, + 289, + 183 + ], + "type": "text", + "content": "[24] Sicong Liu, Xi Sheryl Zhang, Yushuo Li, Yifan Zhang, and Jian Cheng. On the data-efficiency with contrastive image transformation in reinforcement learning. In International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 73, + 184, + 289, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 184, + 289, + 223 + ], + "spans": [ + { + "bbox": [ + 73, + 184, + 289, + 223 + ], + "type": "text", + "content": "[25] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602, 2013. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 224, + 289, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 224, + 289, + 273 + ], + "spans": [ + { + "bbox": [ + 73, + 224, + 289, + 273 + ], + "type": "text", + "content": "[26] Thanh Nguyen, Tung M Luu, Thang Vu, and Chang D Yoo. Sample-efficient reinforcement learning representation learning with curiosity contrastive forward dynamics model. In International Conference on Intelligent Robots and Systems, 2021. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 73, + 274, + 289, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 274, + 289, + 313 + ], + "spans": [ + { + "bbox": [ + 73, + 274, + 289, + 313 + ], + "type": "text", + "content": "[27] Minting Pan, Xiangming Zhu, Yunbo Wang, and Xiaokang Yang. Iso-dream: Isolating and leveraging noncontrollable visual dynamics in world models. In Advances in Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 73, + 315, + 289, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 315, + 289, + 353 + ], + "spans": [ + { + "bbox": [ + 73, + 315, + 289, + 353 + ], + "type": "text", + "content": "[28] Deepak Pathak, Pulkit Agrawal, Alexei A Efros, and Trevor Darrell. Curiosity-driven exploration by self-supervised prediction. In International Conference on Machine Learning, 2017. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 73, + 354, + 289, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 354, + 289, + 393 + ], + "spans": [ + { + "bbox": [ + 73, + 354, + 289, + 393 + ], + "type": "text", + "content": "[29] Roberta Raileanu, Maxwell Goldstein, Denis Yarats, Ilya Kostrikov, and Rob Fergus. Automatic data augmentation for generalization in reinforcement learning. In Advances in Neural Information Processing Systems, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 73, + 395, + 289, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 395, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 73, + 395, + 289, + 434 + ], + "type": "text", + "content": "[30] Max Schwarzer, Ankesh Anand, Rishab Goel, R Devon Hjelm, Aaron Courville, and Philip Bachman. Data-efficient reinforcement learning with self-predictive representations. arXiv preprint arXiv:2007.05929, 2020.8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 73, + 435, + 289, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 435, + 289, + 474 + ], + "spans": [ + { + "bbox": [ + 73, + 435, + 289, + 474 + ], + "type": "text", + "content": "[31] Ramanan Sekar, Oleh Rybkin, Kostas Daniilidis, Pieter Abbeel, Danijar Hafner, and Deepak Pathak. Planning to explore via self-supervised world models. In International Conference on Machine Learning, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 73, + 475, + 289, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 475, + 289, + 514 + ], + "spans": [ + { + "bbox": [ + 73, + 475, + 289, + 514 + ], + "type": "text", + "content": "[32] Younggyo Seo, Lili Chen, Jinwoo Shin, Honglak Lee, Pieter Abbeel, and Kimin Lee. State entropy maximization with random encoders for efficient exploration. In International Conference on Machine Learning, 2021. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 73, + 516, + 289, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 516, + 289, + 554 + ], + "spans": [ + { + "bbox": [ + 73, + 516, + 289, + 554 + ], + "type": "text", + "content": "[33] Wenling Shang, Xiaofei Wang, Aravind Srinivas, Aravind Rajeswaran, Yang Gao, Pieter Abbeel, and Misha Laskin. Reinforcement learning with latent flow. In Advances in Neural Information Processing Systems, 2021. 3, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 73, + 555, + 289, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 555, + 289, + 594 + ], + "spans": [ + { + "bbox": [ + 73, + 555, + 289, + 594 + ], + "type": "text", + "content": "[34] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Computer Vision and Pattern Recognition, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 73, + 596, + 289, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 596, + 289, + 625 + ], + "spans": [ + { + "bbox": [ + 73, + 596, + 289, + 625 + ], + "type": "text", + "content": "[35] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In Advances in Neural Information Processing Systems, 2014. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 73, + 626, + 289, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 626, + 289, + 666 + ], + "spans": [ + { + "bbox": [ + 73, + 626, + 289, + 666 + ], + "type": "text", + "content": "[36] Adam Stooke, Kimin Lee, Pieter Abbeel, and Michael Laskin. Decoupling representation learning from reinforcement learning. In International Conference on Machine Learning, pages 9870-9879. PMLR, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 73, + 666, + 289, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 666, + 289, + 685 + ], + "spans": [ + { + "bbox": [ + 73, + 666, + 289, + 685 + ], + "type": "text", + "content": "[37] Chenyu Sun, Hangwei Qian, and Chunyan Miao. Cclf: A contrastive-curiosity-driven learning framework for" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 103, + 523, + 680 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 103, + 523, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 103, + 523, + 122 + ], + "spans": [ + { + "bbox": [ + 326, + 103, + 523, + 122 + ], + "type": "text", + "content": "sample-efficient reinforcement learning. arXiv preprint arXiv:2205.00943, 2022. 4, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 124, + 523, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 124, + 523, + 163 + ], + "spans": [ + { + "bbox": [ + 308, + 124, + 523, + 163 + ], + "type": "text", + "content": "[38] Yuval Tassa, Yotam Doron, Alistair Muldal, Tom Erez, Yazhe Li, Diego de Las Casas, David Budden, Abbas Abdelmaleki, Josh Merel, Andrew Lefrancq, et al. Deepmind control suite. arXiv preprint arXiv:1801.00690, 2018. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 165, + 523, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 165, + 523, + 214 + ], + "spans": [ + { + "bbox": [ + 308, + 165, + 523, + 214 + ], + "type": "text", + "content": "[39] Anirudh Thatipelli, Sanath Narayan, Salman Khan, Rao Muhammad Anwer, Fahad Shahbaz Khan, and Bernard Ghanem. Spatio-temporal relation modeling for few-shot action recognition. In Computer Vision and Pattern Recognition, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 215, + 523, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 215, + 523, + 273 + ], + "spans": [ + { + "bbox": [ + 308, + 215, + 523, + 273 + ], + "type": "text", + "content": "[40] Oriol Vinyals, Timo Ewalds, Sergey Bartunov, Petko Georgiev, Alexander Sasha Vezhnevets, Michelle Yeo, Alireza Makhzani, Heinrich Kuttler, John Agapiou, Julian Schrittwieser, et al. Starcraft ii: A new challenge for reinforcement learning. arXiv preprint arXiv:1708.04782, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 274, + 523, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 523, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 523, + 315 + ], + "type": "text", + "content": "[41] Kaixin Wang, Bingyi Kang, Jie Shao, and Jiashi Feng. Improving generalization in reinforcement learning with mixture regularization. In Advances in Neural Information Processing Systems, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 316, + 523, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 523, + 346 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 523, + 346 + ], + "type": "text", + "content": "[42] Xudong Wang, Long Lian, and Stella X Yu. Unsupervised visual attention and invariance for reinforcement learning. In Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 347, + 523, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 347, + 523, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 347, + 523, + 387 + ], + "type": "text", + "content": "[43] Wangmeng Xiang, Chao Li, Biao Wang, Xihan Wei, XianSheng Hua, and Lei Zhang. Spatiotemporal self-attention modeling with temporal patch shift for action recognition. In European Conference on Computer Vision, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 388, + 523, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 523, + 427 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 523, + 427 + ], + "type": "text", + "content": "[44] Denis Yarats, Ilya Kostrikov, and Rob Fergus. Image augmentation is all you need: Regularizing deep reinforcement learning from pixels. In International Conference on Learning Representations, 2020. 5, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 428, + 523, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 428, + 523, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 428, + 523, + 477 + ], + "type": "text", + "content": "[45] Tao Yu, Cuiling Lan, Wenjun Zeng, Mingxiao Feng, Zhizheng Zhang, and Zhibo Chen. Playvirtual: Augmenting cycle-consistent virtual trajectories for reinforcement learning. In Advances in Neural Information Processing Systems, 2021. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 479, + 523, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 523, + 517 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 523, + 517 + ], + "type": "text", + "content": "[46] Tao Yu, Zhizheng Zhang, Cuiling Lan, Zhibo Chen, and Yan Lu. Mask-based latent reconstruction for reinforcement learning. In Advances in Neural Information Processing Systems. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 519, + 523, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 519, + 523, + 559 + ], + "spans": [ + { + "bbox": [ + 308, + 519, + 523, + 559 + ], + "type": "text", + "content": "[47] Amy Zhang, Rowan McAllister, Roberto Calandra, Yarin Gal, and Sergey Levine. Learning invariant representations for reinforcement learning without reconstruction. arXiv preprint arXiv:2006.10742, 2020. 5, 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 560, + 523, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 560, + 523, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 560, + 523, + 590 + ], + "type": "text", + "content": "[48] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In International Conference on Machine Learning, 2019. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 591, + 523, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 523, + 639 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 523, + 639 + ], + "type": "text", + "content": "[49] Jinhua Zhu, Yingce Xia, Lijun Wu, Jiajun Deng, Wengang Zhou, Tao Qin, Tie-Yan Liu, and Houqiang Li. Masked contrastive representation learning for reinforcement learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 641, + 523, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 641, + 523, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 641, + 523, + 680 + ], + "type": "text", + "content": "[50] Yuke Zhu, Josiah Wong, Ajay Mandlekar, and Roberto Martín-Martín. robosuite: A modular simulation framework and benchmark for robot learning. arXiv preprint arXiv:2009.12293, 2020. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 704, + 312, + 712 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 704, + 312, + 712 + ], + "spans": [ + { + "bbox": [ + 299, + 704, + 312, + 712 + ], + "type": "text", + "content": "185" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Simple Baselines for Interactive Video Retrieval with Questions and Answers/149533a1-79f1-428b-885a-6c528f5bd0a8_content_list.json b/2023/Simple Baselines for Interactive Video Retrieval with Questions and Answers/149533a1-79f1-428b-885a-6c528f5bd0a8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6e8090bd666a60659dce84e0b0065af69c83e94a --- /dev/null +++ b/2023/Simple Baselines for Interactive Video Retrieval with Questions and Answers/149533a1-79f1-428b-885a-6c528f5bd0a8_content_list.json @@ -0,0 +1,1471 @@ +[ + { + "type": "text", + "text": "Simple Baselines for Interactive Video Retrieval with Questions and Answers", + "text_level": 1, + "bbox": [ + 96, + 130, + 872, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kaiqu Liang \nPrinceton University", + "text_level": 1, + "bbox": [ + 264, + 181, + 436, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "k12471@princeton.edu", + "bbox": [ + 264, + 219, + 442, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Samuel Albanie \nUniversity of Cambridge", + "text_level": 1, + "bbox": [ + 506, + 181, + 705, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sma71@cam.ac.uk", + "bbox": [ + 540, + 220, + 674, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 312, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To date, the majority of video retrieval systems have been optimized for a \"single-shot\" scenario in which the user submits a query in isolation, ignoring previous interactions with the system. Recently, there has been renewed interest in interactive systems to enhance retrieval, but existing approaches are complex and deliver limited gains in performance. In this work, we revisit this topic and propose several simple yet effective baselines for interactive video retrieval via question-answering. We employ a VideoQA model to simulate user interactions and show that this enables the productive study of the interactive retrieval task without access to ground truth dialogue data. Experiments on MSR-VTT, MSVD, and AVSD show that our framework using question-based interaction significantly improves the performance of text-based video retrieval systems. Code is available at https://github.com/kevinliang888/IVR-QA-baselines.", + "bbox": [ + 75, + 300, + 473, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 571, + 207, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given the surging popularity of video content, there is a pressing need to develop tools that enable users to search video collections accurately and efficiently. The text-video retrieval task aims to meet this need by ranking videos among a gallery according to how well they match a given text query. This task has seen tremendous progress in recent years, benefiting from large-scale pretraining datasets [47, 41], multimodal architectures [46, 38, 17] and robust optimisation strategies for cross-modal representation learning [45].", + "bbox": [ + 75, + 598, + 470, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Much of the work to date has focused on a \"single-shot\" scenario in which a user submits a single query in isolation. While this may be applicable in some scenarios, there are many in which it would be useful to allow refinements to the initial ranking (particularly when the pool of valid responses to the initial query is large). To a large degree, this focus single-shot systems can be explained by practicality: obtaining the data required to develop and evaluate interactive retrieval systems is extremely challenging. User studies that engage individuals in interactive sessions are difficult to scale, while production data from product deployments is", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b199ceb1d8159e2b33ffb88b74e490e23c03f69a21bb18078ce751b246283506.jpg", + "image_caption": [ + "Figure 1: Interactive video retrieval with questions and answers. We revisit interactive video retrieval with a simple framework that employs Video Question Answering (VideoQA) to simulate user responses, sidestepping the need to collect dialogue data for development and evaluation." + ], + "image_footnote": [], + "bbox": [ + 506, + 271, + 890, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "challenging to adapt to the development of new systems.", + "bbox": [ + 500, + 590, + 872, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although it has received less attention, the topic of interactive retrieval is far from new—it has been studied for more than two decades [53]. Following an initial query, various mechanisms have been offered to the user to enable them to refine results: relevance scores [53] (corresponding to highly relevant, irrelevant, etc.), relative attributes [29] (e.g. \"shiner than these\"), natural language descriptions [19, 57] (e.g. \"unlike the given result, the one I want has fur on the back\") and question-answer dialog [13, 43] (e.g. \"Q: are there any people in the shot?\" A: No\").", + "bbox": [ + 496, + 609, + 893, + 761 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Among dialogue-based approaches, the recent work of Madasu et al. [43] highlighted the potential of question-answering with free-form text questions to enhance video retrieval performance. At the heart of this approach is a question-generating system that probes the user for more details about their desired target in order to refine the retrieval ranking. In such a framework, user responses can be simulated by employing a question-answering model to synthesise plausible answers to the questions (Fig. 1). Our", + "bbox": [ + 496, + 763, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11091", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "work draws inspiration from ViReD [43], but differs in two key aspects. (1) The ViReD answer generator works by first captioning, then extracting the answer from the text of the caption. The disadvantage of this approach is that the answer indirectly depends on the visual content. Instead, we adopt a video question-answering model to simulate the answer, incorporating the information from the video directly with the question. (2) ViReD fine-tunes on AVSD with the human-collected dialogue while our approach does not require any fine-tuning. We show that our approach generalises well to multiple datasets, and obtains significantly better performance than ViReD.", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we make the following contributions: (1) We propose a simple, yet effective framework for interactive video retrieval based on video question answering, and explore the design space of question generators within this framework. (2) We develop three simple question generators that leverage heuristics and learned models to gather discriminative information beyond objects (scenes, actions, colour etc.) in the target video. (3) Through careful experiments, we demonstrate that our system generalizes well to several video datasets and obtains significant improvement over baseline without fine-tuning. These results, which can be interpreted as leveraging an approximate interactive oracle, highlight the potential of question-based interactive retrieval frameworks to substantially enhance video search.", + "bbox": [ + 75, + 272, + 472, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 497, + 215, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-video Retrieval A long-running theme of text-video retrieval research has been the development of effective fusion mechanisms for cross-modal learning [68, 25, 26, 67, 58]. Another theme of vision-language retrieval has sought to tackle challenges associated with hubness [52] in the embedding space [14, 5]. In the past few years, there has been increasing emphasis on the use of pre-trained models that enable retrieval systems to leverage large quantities of training data that is available for related tasks, often spanning multiple modalities [38, 46, 17, 11]. Recently, there has been a particular focus on building atop models that can leverage the most scalable forms of pretraining data (e.g. CLIP [51]), which have demonstrated compelling performance across a range of vision and language tasks. Luo et al. [40] proposed CLIP4CLIP which adapts CLIP for text-video retrieval through a combination of architectural changes and finetuning. Other approaches have included the use of a curriculum to learn from images and video [3], BERT-style architectures [56], and additional pretraining tasks [34]. Our approach is built upon BLIP [35], a unified framework for retrieval, question answering, and captioning. Differently from the works listed above, we focus on the interactive retrieval setting.", + "bbox": [ + 75, + 523, + 472, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Question Answering. The goal of visual question-answering systems is to answer questions posed in natural", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "language that concern visual content in the form of an image or video. A broad range of work in this space has sought to combine spatial image representations and sequential question representations [2, 4, 16, 39, 61, 62, 66]. To answer questions about video content, spatio-temporal video representations that encode motion and appearance have been studied in detail [15, 18, 21, 22, 23, 24, 31, 33, 70]. There has also been work that extends these ideas to open-ended visual question answering [20, 64]. Recent work on goal-oriented visual dialog [12, 13, 32, 10, 49] uses policy-based reinforcement learning [49, 12, 13] or optimizing information gain [32] to improve the performance. However, these works require intensive training on human-annotated dialog datasets and do not trivially generalise to video. In terms of VideoQA, a few recent works have explored zero-shot approaches [64, 69], where models are only trained on automatically mined video clips with short text descriptions. Our VideoQA model is based on BLIP [35] which is trained on manually annotated image-question-answer triplets.", + "bbox": [ + 496, + 90, + 893, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interactive cross-modal retrieval A number of interactive image retrieval systems have been developed, drawing inspiration from visual dialogue [60, 53, 29, 28, 27, 50, 65]. In such systems, users provide feedback to an agent to guide it toward the desired retrieval target. Two families of feedback have been studied in detail: relevance feedback and difference feedback. For the former [60, 53], users provide a relevance score for the current retrieval results. The system then re-ranks its initial results using the feedback provided by the user. For the latter approach [29, 28, 27, 50, 65], users communicate the difference between the target image and a reference image to the system with cues such as tags or descriptions. The system then \"whittles away\" irrelevant images, iteratively filtering to select the desired target. However, one drawback of these systems is that they place a heavy time cost on the user who must describe what is missing in the initial search results. For example, description-based methods [19, 57] require users to provide detailed sentence-level feedback and tag-based methods require users to provide a collection of attributes [27, 28, 29].", + "bbox": [ + 496, + 380, + 895, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these concerns, Cai et al [7] propose a partial query system in which potentially discriminative objects are suggested to the user with the goal of narrowing down the search results quickly. However, this system doesn't utilize other cues in the image such as object color, actions, scenes, background, etc., so it can require many interactions with the user to obtain good retrieval results. By using free-form questions to elicit information from the user, our system avoids this restriction to solely object-based queries.", + "bbox": [ + 496, + 685, + 893, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most related to our approach, the ViReD work of Madasu et al. [43] highlighted the potential of question-answering with free-form text questions to enhance video retrieval performance. However, rather than performing VideoQA directly, ViReD employs captioning followed by text-based", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11092", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "question-answering to answer questions about the video, bottlenecking answer quality. In addition, the ViReD question generator is complex and dependent on dialog history for training—our approach requires no fine-tuning and generalises effectively to multiple datasets.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 184, + 166, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Interactive video retrieval with Q&A", + "text_level": 1, + "bbox": [ + 76, + 209, + 395, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Components. Our interactive video retrieval system contains four components: a Text Encoder, a Visual Encoder, a Ranker, and a Q&A System. The Text Encoder and Visual Encoder embed partial queries and videos into a shared textual-visual space. The Ranker computes the similarity scores between text and video features which are used to rank the videos. To accelerate inference, we adopt the approach of [36] that first selects $K$ candidates based on the video-text feature similarity, and then reranks the selected candidates based on their pairwise Image-Text Matching (ITM) scores. Specifically, the model uses an ITM head (a linear layer) to predict the score indicating whether an image-text pair is positive (matched) or negative (unmatched) given their multimodal feature.", + "bbox": [ + 75, + 234, + 470, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Q&A System forms the central module in the framework (Fig. 3). It comprises a captioner adopted from the BLIP caption model, a question generator, and an answer generator that takes the form of a VideoQA (video question-answering) model that simulates user responses. We adopt the BLIP VideoQA model [35]. Instead of formulating VideoQA as a multi-answer classification task [9, 37], Li et al [35] interpret it as an answer generation task, enabling open-ended VideoQA. This open-ended VideoQA allows more freedom to generate and combine answers, lending flexibility to the system.", + "bbox": [ + 75, + 446, + 468, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Interaction workflow. The complete interaction process is demonstrated in Fig. 2. The user provides an initial text query to the system. The interaction loop then begins, comprising four main steps. (1) Generate input to the Q&A System: The current (partial) query together with candidate retrieved videos is passed to the Q&A system. To obtain the candidate videos, we first generate the text embedding corresponding to the partial query and video embeddings for all videos. The Ranker then computes similarity scores in embedding space (via cosine similarity) to rank the videos. We select as candidate videos those who fall within the top $K$ places of the ranking and feed them as input to the Q&A System. (2) Question generation: We explore several kinds of question generators (described in detail in Sec. 3.2 and Sec. 3.3). Different question generators require access to different inputs. For example, for the Auto-text-vid question generation approach (Sec. 3.3), we first generate captions for the candidate videos, then we generate questions based on both the captions and the partial query. (3) User answer sim", + "bbox": [ + 75, + 613, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ulation by VideoQA model: Given the question and the target videos, we use the VideoQA model to generate an answer. (4) New query generation: We generate a complete sentence by incorporating the answer and adding it to the original textual query. This completes one round of interaction.", + "bbox": [ + 496, + 90, + 890, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Combining responses after each round of interaction After each round of interaction, we obtain new information with the potential to improve retrieval performance. It is therefore worth considering how this information is best integrated. One approach is to treat each new piece of information as a separate query, then adopt a strategy for multi-query retrieval. In recent work, Wang et al [59] propose two post hoc inference methods, Similarity aggregation (SA) and Rank aggregation (RA), to extend single-query to multi-query video retrieval. However, initial experiments with these approaches did not prove promising. We believe this is because, while each new piece of information may provide complementary information, it may perform poorly as a query in isolation. However, SA and RA work best with individual queries of high quality. We, therefore, adopt a different approach: we aggregate pieces of information by concatenating them into one large query. In this way, the initial query can be iteratively refined by adding more complementary information. In detail, we concatenate multiple queries using a separation token [SEP]. This token is derived from BERT pretraining, in which it is used to separate sentences as follows: \"[CLS]\" sentence A [SEP] sentence B [SEP]\", suggesting its applicability for separating content in our application setting.", + "bbox": [ + 496, + 167, + 893, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Heuristic Question Generator", + "text_level": 1, + "bbox": [ + 500, + 527, + 764, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now turn the design of question generators used in our system, starting with a simple heuristic question generator.", + "bbox": [ + 496, + 551, + 890, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The problem of asking useful questions about visual content has received attention in the field of Visual Question Generation. Mostafazadeh et al [48] proposed a framework that generates questions that can potentially engage a human in starting a conversation. They observe that asking a question that can be answered simply by looking at the image would be of interest to the computer vision community, but those questions are neither natural nor engaging for a person to answer. Nevertheless, for the purpose of improving retrieval performance, we focus on concrete, discriminative questions (rather than abstract and engaging questions). In particular, to develop a heuristic question generator, we target goal-driven questions [30] that focus on extracting objects (\"what is the person throwing?\"), attributes (\"what kind of shirt is the person wearing?\"), color (\"what color is the frisbee?\"), material (\"what material is the frisbee?\"), etc.", + "bbox": [ + 496, + 583, + 890, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We are particularly interested in obtaining information that is highly discriminative and complementary to the original query. Ask & Confirm [7] argues that objects provide discriminative cues to distinguish a target from other candi", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11093", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a53c8d5942e64569028088623a22771cdfeea6ad84b2bb3ab5dcf04c00c7b112.jpg", + "image_caption": [ + "Figure 2: Model architecture of interactive video retrieval with Q&A (Questions and Answers). The user provides an initial textual query to the system. The interaction loop contains 4 main steps. Step 1: Generate input to the Q&A system. For Heuristic and Auto-text question generation approaches, the input to the Q&A system are just the queries. For Auto-text-vid question generation approach, the input contains both queries and the candidate videos. Step 2: Question generation. Step 3: User answer simulation by VQA model. Step 4: New query generation. This fulfills one round of the interaction." + ], + "image_footnote": [], + "bbox": [ + 81, + 92, + 869, + 272 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1fc26f07da1dfb6211fdff7e2c8ef6490422da5819b33f0fc88ffe88a73aadf2.jpg", + "image_caption": [ + "Figure 3: Model architecture of Q&A System. The interaction loop contains either two or three steps depending on which question generation approach is used. Step 1: Caption generation. Step 2: Question generation. Step 3: Answer generation. Auto-text-vid approach contains step 1, 2 and 3 while Heuristic and Auto-text question generation approaches contain only step 2 and 3." + ], + "image_footnote": [], + "bbox": [ + 84, + 388, + 464, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dates in the context of image retrieval. Similarly, we observe the importance of the objects in the video retrieval task. However, other cues such as color, human actions, and scenes are also important, and we, therefore, aim to construct a question generator that incorporates these cues.", + "bbox": [ + 75, + 714, + 470, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Heuristic Question Generator. Our first approach to question generation employs hand-crafted heuristics. Our strategy is to first obtain objects from the initial textual query (via pattern matching) and then ask questions about these objects. If an object is a living thing such as a person or animal, then we enquire about both the action and scene associated with it. If the object is a non-living thing such as a lamp or table,", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "then we ask only the scene. If no object is found in the initial textual query, we ask the user to identify an object in the scene. The follow-up questions then proceed as above.", + "bbox": [ + 496, + 387, + 890, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since living things are often associated with actions, if the object is a living thing, then we employ a template of the form \"who is doing what in what place\". If the object is a non-living thing, the template used is \"who is in what place\". Once the answer is obtained, it is appended to the question, then both are appended to the current query for the next round of video retrieval.", + "bbox": [ + 496, + 434, + 892, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One potential downside of this approach is the redundancy between questions and the query. For example, suppose the initial textual query is \"a man is singing\", then when we ask \"what is the man doing?\", the answer may add no new information. In practice, this is typically not the case. We find that the information of the initial query is often restricted to addressing some temporal window within the video.", + "bbox": [ + 496, + 541, + 893, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the question provided to the user can often address a larger portion of the video. For the example above, the man may walk through the street in the following frames after singing. The answer generator could encode this information to provide a complementary cue for retrieval.", + "bbox": [ + 496, + 648, + 893, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Parametric Question Generators", + "text_level": 1, + "bbox": [ + 500, + 739, + 785, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The heuristic question generator captures intuition about what makes for useful questions but lacks flexibility. We therefore explore question generation mechanisms that leverage a large language model that possesses the ability to ask flexible, open-ended questions. The first, Auto-text, asks questions based only on the text query. The second, Autotext-vid, asks questions based on both the text query and the top- $K$ ranked video candidates for the current query (here $K$ is a hyperparameter of the system).", + "bbox": [ + 496, + 763, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11094", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The foundation of both question generators is T-0++ [54], a large language model that outperforms or matches GPT-3 [6] on many NLP tasks. Since this language model has been trained on various forms of question answering and shown to be very effective at zero-shot task generalization, we directly apply this model to generate questions without finetuning.", + "bbox": [ + 76, + 90, + 472, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Auto-text. To generate questions from the query, we provide the following template to the language model: \"Suppose you are given the following video descriptions $Q_{i}$ , What question would you ask to help you unique identify the video?\" where $Q_{0}$ is the initial query and $Q_{i}$ is the concatenated query used to retrieve videos after the $i^{\\text{th}}$ round of interaction. The new query after each round incorporates the answer to the question through concatenation. Note that in terms of the Auto-text question generation approach, the question depends both on the original query and the new descriptions generated after each round of the interaction.", + "bbox": [ + 75, + 183, + 472, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Auto-text-vid. Differently from Auto-text, Auto-text-vid conditions its question generation on both the text query and the current top- $k$ ranked retrieved videos (selected via cosine similarity for the query corresponding to the current round of interaction). This is because, intuitively, we expect the top-ranked videos selected at each round to share some common information with the target video. The use of the top- $k$ retrieved videos as conditioning information was also considered by [43], who found it beneficial.", + "bbox": [ + 75, + 349, + 472, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To proceed, we first generate captions for each of the top-ranked videos using the BLIP captioner [35]. Then we use both these captions and the original text query to generate questions. The input sentence to the language model employs the following template: \"Suppose you are given the following video descriptions: $C_i$ . What question would you ask to help you unique identify the video described as follows: $Q_i$ ?\" where $C_i$ are the generated captions at the $i^{\\text{th}}$ round of interaction and $Q_i$ is the concatenated textual query used to retrieve videos after the $i^{\\text{th}}$ round of interaction. The questions generated are based on those common features, so the machine obtains more information about the target video through the interaction.", + "bbox": [ + 75, + 486, + 472, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark. Since the top-ranked videos change at each round, there is some variation in the questions generated by Auto-text-vid. This may provide complementary cues, but also carries the risk of additional noise (both through mistakes in the synthetic captions and through mismatches in content between the candidates and the target).", + "bbox": [ + 75, + 681, + 472, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Question Augmentation", + "text_level": 1, + "bbox": [ + 76, + 785, + 294, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further increase the information extracted from the user about the target video, we also introduce two further simple techniques to augment the questions: Ask Segment (AS) and Ask Object (AO).", + "bbox": [ + 75, + 809, + 470, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ask Segment (AS) augmentation asks the user questions about each half of the video independently. Videos evolve", + "bbox": [ + 75, + 869, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "over time and different objects and actions may take place in different frames. As a consequence, asking the same question about different halves of the video often yields very different answers.", + "bbox": [ + 496, + 90, + 893, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ask Object (AO) augmentation enquires about more objects in the scene. This approach is motivated by Ask & Confirm [7] in which Cai et al observe that objects are often highly discriminative. Rather than searching for discriminative objects as candidates through a complex RL policy [7], we simply ask the user which objects are in the scene.", + "bbox": [ + 496, + 152, + 893, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that these strategies can be combined to generate more discriminative questions for interactive video retrieval.", + "bbox": [ + 496, + 243, + 893, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 498, + 289, + 624, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first describe the datasets used in our experiments (Sec. 4.1). We then discuss implementation details (Sec. 4.2) and evaluation metrics (Sec. 4.3).", + "bbox": [ + 496, + 315, + 893, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 498, + 372, + 602, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MSR-VTT [63] is a video benchmark that is widely used for text-video retrieval. It contains 10,000 videos sourced from YouTube. Each video is annotated with 20 natural language descriptions. Following previous work, we report retrieval results on the 1K-A test set.", + "bbox": [ + 496, + 395, + 893, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MSVD [8] is a dataset initially collected for translation and paraphrase evaluation. It contains 1,970 videos, spanning a duration from 1 second up to 62 seconds. The train, validation, and test splits contain 1,200, 100, and 670 videos, respectively. Each video contains approximately 40 associated sentences in English.", + "bbox": [ + 496, + 470, + 893, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AVSD (Audio-Visual Scene aware Dialog dataset) [1] is a collection of videos associated with ground truth dialog data. Each video dialog consists of 10 rounds of human-generated questions and answers describing various details about the video content. The AVSD dataset consists of 7985 training samples and 1863 validation samples. We follow prior work [43] and split the validation samples into a new validation split of 863 samples and 1000 test samples. Note that our approach does not make use of the dialog data, since we simulate the user directly. We use only the caption as the initial textual query to retrieve the video.", + "bbox": [ + 496, + 563, + 893, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 739, + 718, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the visual encoder, we employ a vision transformer with ViT-B/16 which contains 12 layers and 12 heads. This model has a hidden dimensionality of 768, an MLP size of 3072, and an input patch size of $16 \\times 16$ . The text transformer used for text encoder is BERTBASE with 12 layers, 12 self-attention heads, and a hidden size of 768. The text transformer used for the image-grounded text encoder is the same as the text encoder except that it uses additional cross-attention layers to model vision-language interactions.", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11095", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4575ab25945bde5373aebff5e1198759d16dde79358ab7a8757f3f4751c5d37c.jpg", + "image_caption": [ + "a woman demonstrates how to cook chitlins", + "Q: what is the woman doing? \nA: washing meat", + "Q: what is the man doing? \nA: looking at camera", + "Q: where is the meat? \nA: in sink", + "Q: where is the mirror? \nA: on wall" + ], + "image_footnote": [ + "Round:0", + "Rank : 953", + "Round:1", + "Rank:5", + "Round:3", + "Rank:1" + ], + "bbox": [ + 91, + 95, + 223, + 176 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dbf2ed63ac1a610d91798f37c638c476abdc648f6e7dfdbd41607cc568062a25.jpg", + "image_caption": [ + "a girl and a man are talking to each other" + ], + "image_footnote": [ + "Round:0", + "Rank:421", + "Round:1", + "Rank:74", + "Round:4", + "Rank:2" + ], + "bbox": [ + 91, + 195, + 220, + 272 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/79a72c2d13ddf6673743adffd90279d3739a6e32d703fafd83820cef69ac46ef.jpg", + "image_caption": [ + "A video of someone talking about some girls", + "Q: what are the girls doing? A: looking at each other", + "Q: where is the silver earring? \nA: in left ear" + ], + "image_footnote": [ + "Round:0", + "Rank:353", + "Round:1", + "Rank:6", + "Round:4", + "Rank:1" + ], + "bbox": [ + 91, + 289, + 220, + 364 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8f4c2d3c350deb4211464c39e1c8bb71851fa4e9c62df3e13670cbb8d90a0cdf.jpg", + "image_caption": [ + "a man is wearing a cap", + "Q: what is the man doing? \nA: clapping", + "Q: what is the object? A: baseball glove", + "Figure 4: Qualitative results of the dialogue generated by our Q & A System on MSR-VTT. The questions are generated by the Heuristic Question Generator. The initial queries of the target videos are shown in orange boxes. The questions and answers generated through each interaction are shown in blue boxes. Both the initial retrieval rank and the rank after the interaction are demonstrated. In each case, we observe a boost in performance. Note, however, that the answers are not always correct (e.g. in the fourth row \"A: baseball glove\" when the sport does not appear to be baseball). However, since the answer is \"nearly correct\", it still provides a useful cue to the retrieval model." + ], + "image_footnote": [ + "Round:0", + "Rank : 39", + "Round:2", + "Rank : 11", + "Round:3", + "Rank:1" + ], + "bbox": [ + 88, + 388, + 220, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The architecture of the image-grounded text decoder is similar to the image-grounded text encoder but it replaces the bi-directional self-attention layers with causal self-attention layers. All videos are resized to $384 \\times 384$ as input.", + "bbox": [ + 75, + 606, + 468, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our interactive system, we adopt the BLIP retrieval model, the VQA model, and the captioner. The BLIP retrieval model is composed of an image encoder, a text encoder, and an image-grounded text encoder. The BLIP VQA model is composed of an image encoder, an image-grounded text encoder, and an image-grounded text decoder. The BLIP captioner contains an image encoder and an image-grounded text decoder. To perform zero-shot transfer video retrieval, we uniformly sample 8 frames per video and concatenate the frame features into a single sequence. To generate questions, we adopt $\\mathrm{T0 + + }$ [54], a powerful encoder-decoder language model trained on a collection of English-prompted datasets. This model attains strong zero-shot task generalization on English natural language prompts (outperforming GPT-3 on many tasks while being 16x smaller).", + "bbox": [ + 75, + 674, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Auto-text-vid question generation approach depends on the current top-k ranked retrieved videos. We experimented with different k (3, 5, 10, 15) and observed that when $\\mathrm{k} = 5$ , we obtained the best performance. This may be because setting k to be too large produces too much noise in the generated captions while setting k to be too small does not provide an opportunity to extract all potentially discriminative information from related videos.", + "bbox": [ + 496, + 606, + 893, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. In addition to the BLIP baseline, we also experimented with the interactive approach in [43] for fair comparisons. As public code is not available at the time of writing, we simulate their answer generation by first captioning the target video (by BLIP captioner) and then using the large language model $(\\mathrm{T0}++)$ to generate the answer based on the caption and question. On the AVSD dataset, we also experimented with the approach in [44, 43] using 10 rounds of manually annotated human dialogue history for each video. However, simply concatenating all 10 rounds of ground truth dialog with the initial text query got a slightly", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11096", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "worse performance than the baseline. Instead, we concatenated all the answers with the initial text query for retrieval.", + "bbox": [ + 75, + 90, + 470, + 122 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Evaluation", + "text_level": 1, + "bbox": [ + 76, + 131, + 197, + 145 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. We adopt evaluation metrics that are widely used in the standard single-query retrieval setting and report recall@K (R@K, higher the better) and median rank (MdR, lower the better) in our experiments. R@K calculates the percentage of test data for which the ground-truth video is found in the retrieved K videos. MdR metric captures the median rank of the retrieved ground truth videos.", + "bbox": [ + 75, + 155, + 468, + 260 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Time cost. We next assess the time cost of producing answers during the interaction. For this, we sample 50 videos from the MSR-VTT dataset. We aim to compare our answer generation approach using the VideoQA model with the approach in ViReD [43] (We call this answer generation approach $CAP + LM$ ). In addition, we also conduct a user study on answer generation. We invite 4 users to answer the auto-generated question (Auto-text) on the target videos and record the time they spent on each question. (Note that each user went through all 50 videos before the interaction.)", + "bbox": [ + 75, + 261, + 470, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 76, + 426, + 161, + 441 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ef6ef9045b8d3bd302f4b7487e73ab8751cb40d5ad64a1abe1541a62b3f72427.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodITAR1 ↑R5 ↑R10 ↑MdR ↓
FiT [3]×31.961.172.63
CLIP4Clip [40]×41.569.479.32
CenterCLIP [71]×48.473.882.02
X-CLIP [42]×49.375.884.82
BLIP [35]×43.365.674.72
BLIP + ViReD [43]✓47.072.381.42
BLIP + Auto-text w/o AO✓47.672.881.82
BLIP + Auto-text-vid w/o AO✓49.072.980.02
BLIP + Auto-text✓62.284.590.71
BLIP + Auto-text-vid✓57.682.288.71
BLIP + Heuristic✓67.989.594.91
", + "bbox": [ + 81, + 462, + 468, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we discuss experimental results. On MSR-VTT (Tab. 1), without AO (asking objects), we observe that both the Auto-text and Auto-text-vid question generation approaches outperform the BLIP baseline. We observe that Auto-text-vid, which makes use of the top-ranked retrieved videos at each round, performs better than Auto-text. With AO, there is a significant performance boost for both approaches. For instance, we observe that the Auto-text Q&A System outperforms the baseline by $+18.9\\%$ in recall@1.", + "bbox": [ + 75, + 763, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/31a9bcb22e1711571c6e2b02cf0197f269b00583013f9e06e956fa021f7c334e.jpg", + "table_caption": [ + "Table 1: Comparison to state-of-the-art methods for text-to-video retrieval on the 1k test split of MSRVTT dataset. ITA indicates whether it is an interactive retrieval system. Recall metrics are reported in percent." + ], + "table_footnote": [], + "table_body": "
MethodITAR@1↑R@5↑R@10↑MdR↓
FiT [3]×33.764.776.33
CLIP4Clip [40]×46.276.184.62
X-CLIP [42]×50.480.6--
CenterCLIP [71]×50.680.388.41
BLIP [35]×44.273.780.92
BLIP + ViReD [43]✓51.078.187.21
BLIP + Auto-text✓68.791.295.81
BLIP + Auto-text-vid✓66.188.893.71
BLIP + Heuristic✓74.293.497.91
", + "bbox": [ + 506, + 88, + 887, + 250 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a6220412ad5aeb9661021ce5b399705ca0c9f01316942ab5c90cd6a9281ef79c.jpg", + "table_caption": [ + "Table 2: Comparison to state-of-the-art methods for text-to-video retrieval on the test split of the MSVD dataset. ITA indicates whether it is an interactive retrieval system." + ], + "table_footnote": [], + "table_body": "
MethodITAR@1 ↑R@5 ↑R@10 ↑MdR ↓
FiT [3]×5.618.427.525
LSTM [44]✓4.213.522.1-
FiT + Human dialog [43]✓10.828.940.018
ViReD [43]✓24.949.060.86
BLIP [35]×27.350.061.25.5
BLIP + Human dialog [43]✓27.450.462.75
BLIP + ViReD [43]✓28.454.366.84
BLIP + Auto-text✓36.061.671.63
BLIP + Auto-text-vid✓36.561.071.13
BLIP + Heuristic✓39.569.177.82
", + "bbox": [ + 504, + 343, + 890, + 508 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/da41c4e4bcae48042343a1565c23d3366953bcdaf8b96191be3700b1ba9cf1c2.jpg", + "table_caption": [ + "Table 3: Comparison to state-of-the-art methods for text-to-video retrieval on the test split of AVSD dataset. ITA indicates whether it is an interactive retrieval system." + ], + "table_footnote": [], + "table_body": "
VideoQACAP+LMUser
Avg time (s)0.18.05.7
", + "bbox": [ + 562, + 604, + 828, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4: Comparison of the average time cost of different answer and question generation approaches. We show the results of the answer generation approach of VQA (ours) and CAP+LM (simulation of ViReD [43] answer generator) and the user study.", + "bbox": [ + 498, + 657, + 890, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Interestingly, we observe that Auto-text outperforms Autotext-vid when used in combination with AO. This suggests that simply asking the user directly for the objects that they are interested in (rather than inferring them from closely ranked videos) offers higher signals.", + "bbox": [ + 496, + 777, + 890, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also observe that the Heuristic question generator yields dramatic gains: the recall@1 is improved by $24.6\\%$ (from $43.3\\%$ to $67.9\\%$ ), $30\\%$ (from $44.2\\%$ to $74.2\\%$ ) and", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11097", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3225b0ef270f835d04b1226e6990e1c085dab9a82b8bf007c2e79ea478cfde7e.jpg", + "image_caption": [ + "Retrieval results vs Number of queries", + "Figure 5: Retrieval result vs Number of queries on MSRVTT. Performance of all different approaches improves as the number of rounds increases. Heuristic question generation approach achieves the most significant improvement." + ], + "image_footnote": [], + "bbox": [ + 98, + 104, + 450, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$12.2\\%$ (from $27.3\\%$ to $39.5\\%$ ) on MSR-VTT, MSVD and AVSD respectively. Using Auto-text question generator, the BLIP baseline recall@1 is improved by $18.9\\%$ (from $43.3\\%$ to $62.2\\%$ ), $24.5\\%$ (from $44.2\\%$ to $68.7\\%$ ), and $8.7\\%$ (from $27.3\\%$ to $36.0\\%$ ) on MSR-VTT, MSVD, and AVSD respectively. This indicates that generated questions elicit useful information from the (simulated) user that substantially improves retrieval performance. Compared to prior interactive approaches (ViReD, Human dialog) on AVSD, our approach does not use dialog training data for finetuning but nevertheless outperforms existing work. Interestingly, the heuristic approach performs best over all three different datasets.", + "bbox": [ + 75, + 416, + 468, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As seen in Fig. 4, a few rounds of interaction are sufficient to locate the large video. We demonstrate the cases where the initial query is incomplete, hindering performance. However, using the question-answering system, the search results quickly improve and we obtain significantly better retrieval performance. The generated questions are of high quality, aiming to confirm discriminative information from the video. In addition, the answers generated by VideoQA model are reasonable in the majority of cases (though not always perfect). They successfully capture different aspects of the information in the video (such as action, location, and object). Another useful property of the generated answers is that they are short phrases, simulating concise user answers. Number of queries. Additional rounds of interaction lead to additional new queries. We observe in Fig. 5 that the retrieval performance increases as the number of queries increases. This effect is clearest for Heuristic question generation approach. The performance of both Auto-text and Auto-text-vid increases faster than their variants without AO. This further confirms that asking users about objects in the", + "bbox": [ + 75, + 598, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "videos is an effective approach to improve retrieval performance at each round of the interaction. We observe that without AO, recall@1 of Auto-text-vid increases faster than that of Auto-text while the opposite result is observed with AO. This agrees with our argument \"simply asking the user directly for the objects that they are interested in (rather than inferring them from closely ranked videos) offers higher signal\". For all the approaches, Performance stabilizes as the number of queries grows beyond 6. Note that the maximum number of queries for heuristic question generation is fixed to 6 and using 6 rounds works best.", + "bbox": [ + 496, + 90, + 893, + 256 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computational cost. As shown in Tab. 4, Our approach using the VideoQA model is significantly more efficient than $\\mathrm{CAP + LM}$ . This is as expected because two steps (captioning and question-answering) are involved in $\\mathrm{CAP + LM}$ approach while VideoQA directly incorporates the video and question information to generate answers. In our naive implementation, $\\mathrm{CAP + LM}$ is even slower than the user response (though this could be optimised).", + "bbox": [ + 496, + 257, + 893, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 500, + 393, + 612, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. (1) $\\mathrm{T0 + + }$ [55] is good at question-answering tasks but might not be good at generating questions. Therefore fine-tuning the model for the question generation task could be helpful. (2) We adopt the BLIP model for both retrieval and VideoQA. In principle, there is the potential for \"leakage\" in which the VideoQA model encodes information that is specific to BLIP, but irrelevant to the video. In practice, different parameter checkpoints are used for different VideoQA and retrieval components, so we believe that such leakage is unlikely to occur in practice.", + "bbox": [ + 496, + 419, + 893, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Broader Impact. Improved video retrieval has a wide range of societally useful applications, spanning domains such as education, entertainment, and security. However, as with many technologies, it is subject to dual use. For instance, it could be used to facilitate unlawful surveillance.", + "bbox": [ + 496, + 571, + 893, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 662, + 617, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we proposed several simple baselines for interactive video retrieval based on the generation of questions and answers. Through experiments on standard benchmarks, we demonstrated the effectiveness of generated questions in combination with a VideoQA model to simulate user responses through the interaction process. In future work, we hope to explore the use of question-answering systems to simulate users in other retrieval tasks.", + "bbox": [ + 496, + 688, + 893, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. SA would like to acknowledge Z. Novak and N. Novak in enabling his contribution, and support from an Isaac Newton grant and an EPSRC HPC grant. KL would like to thank Bill Byrne, Olga Russakovsky, Yu Wu, Vishaal Udandarao, Allison Chen, Xindi Wu, and Zhiwei Deng for helpful discussions and feedback.", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11098", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Huda Alamri, Vincent Cartillier, Abhishek Das, Jue Wang, Anoop Cherian, Irfan Essa, Dhruv Batra, Tim K Marks, Chiori Hori, Peter Anderson, et al. Audio visual scene-aware dialog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7558-7567, 2019. 5", + "[2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018. 2", + "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 2, 7", + "[4] Hedi Ben-Younes, Rémi Cadene, Matthieu Cord, and Nicolas Thome. Mutan: Multimodal tucker fusion for visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2612-2620, 2017. 2", + "[5] Simion-Vlad Bogolin, Ioana Croitoru, Hailin Jin, Yang Liu, and Samuel Albanie. Cross modal retrieval with querybank normalisation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5194-5205, 2022. 2", + "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 5", + "[7] Guanyu Cai, Jun Zhang, Xinyang Jiang, Yifei Gong, Lianghua He, Fufu Yu, Pai Peng, Xiaowei Guo, Feiyue Huang, and Xing Sun. Ask&confirm: active detail enriching for cross-modal retrieval with partial query. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1835-1844, 2021. 2, 3, 5", + "[8] David Chen and William Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 190-200, Portland, Oregon, USA, June 2011. Association for Computational Linguistics. 5", + "[9] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 3", + "[10] Michael Cogswell, Jiasen Lu, Rishabh Jain, Stefan Lee, Devi Parikh, and Dhruv Batra. Dialog without dialog data: Learning visual dialog agents from vqa data. Advances in Neural Information Processing Systems, 33:19988-19999, 2020. 2", + "[11] Ioana Croitoru, Simion-Vlad Bogolin, Marius Leordeanu, Hailin Jin, Andrew Zisserman, Samuel Albanie, and Yang Liu. Teachtext: Crossmodal generalized distillation for text-video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11583-11593, 2021. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, José MF Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 326-335, 2017. 2", + "[13] Abhishek Das, Satwik Kottur, José MF Moura, Stefan Lee, and Dhruv Batra. Learning cooperative visual dialog agents with deep reinforcement learning. In Proceedings of the IEEE international conference on computer vision, pages 2951–2960, 2017. 1, 2", + "[14] Georgiana Dinu, Angeliki Lazaridou, and Marco Baroni. Improving zero-shot learning by mitigating the hubness problem. arXiv preprint arXiv:1412.6568, 2014. 2", + "[15] Chenyou Fan, Xiaofan Zhang, Shu Zhang, Wensheng Wang, Chi Zhang, and Heng Huang. Heterogeneous memory enhanced multimodal attention model for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1999–2007, 2019. 2", + "[16] Akira Fukui, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, and Marcus Rohrbach. Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv preprint arXiv:1606.01847, 2016. 2", + "[17] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision, pages 214-229. Springer, 2020. 1, 2", + "[18] Jiyang Gao, Runzhou Ge, Kan Chen, and Ram Nevatia. Motion-appearance co-memory networks for video question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6576–6585, 2018. 2", + "[19] Xiaoxiao Guo, Hui Wu, Yu Cheng, Steven Rennie, Gerald Tesauro, and Rogerio Feris. Dialog-based interactive image retrieval. Advances in neural information processing systems, 31, 2018. 1, 2", + "[20] Hexiang Hu, Wei-Lun Chao, and Fei Sha. Learning answer embeddings for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5428-5436, 2018. 2", + "[21] Deng Huang, Peihao Chen, Runhao Zeng, Qing Du, Mingkui Tan, and Chuang Gan. Location-aware graph convolutional networks for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11021-11028, 2020. 2", + "[22] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2758–2766, 2017. 2", + "[23] Jianwen Jiang, Ziqiang Chen, Haojie Lin, Xibin Zhao, and Yue Gao. Divide and conquer: Question-guided spatiotemporal contextual attention for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11101-11108, 2020. 2", + "[24] Pin Jiang and Yahong Han. Reasoning with heterogeneous graph alignment for video question answering. In Proceedings" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11099", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the AAAI Conference on Artificial Intelligence, volume 34, pages 11109-11116, 2020. 2", + "[25] Dotan Kaufman, Gil Levi, Tal Hassner, and Lior Wolf. Temporal tessellation: A unified approach for video analysis. In Proceedings of the IEEE International Conference on Computer Vision, pages 94-104, 2017. 2", + "[26] Ryan Kiros, Ruslan Salakhutdinov, and Richard S Zemel. Unifying visual-semantic embeddings with multimodal neural language models. arXiv preprint arXiv:1411.2539, 2014. 2", + "[27] Adriana Kovashka and Kristen Grauman. Attribute pivots for guiding relevance feedback in image search. In Proceedings of the IEEE International Conference on Computer Vision, pages 297-304, 2013. 2", + "[28] Adriana Kovashka and Kristen Grauman. Attributes for image retrieval. In Visual Attributes, pages 89-117. Springer, 2017. 2", + "[29] Adriana Kovashka, Devi Parikh, and Kristen Grauman. Whittlesearch: Image search with relative attribute feedback. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2973-2980. IEEE, 2012. 1, 2", + "[30] Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Information maximizing visual question generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2008–2018, 2019. 3", + "[31] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9972–9981, 2020. 2", + "[32] Sang-Woo Lee, Yu-Jung Heo, and Byoung-Tak Zhang. Answerer in questioner's mind: Information theoretic approach to goal-oriented visual dialog. Advances in neural information processing systems, 31, 2018. 2", + "[33] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 2", + "[34] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 2", + "[35] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 3, 5, 7", + "[36] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 3", + "[37] Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV, 2020. 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts. arXiv preprint arXiv:1907.13487, 2019. 1, 2", + "[39] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Hierarchical question-image co-attention for visual question answering. Advances in neural information processing systems, 29, 2016. 2", + "[40] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 2, 7", + "[41] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 1", + "[42] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 638–647, 2022. 7", + "[43] Avinash Madasu, Junier Oliva, and Gedas Bertasius. Learning to retrieve videos by asking questions. arXiv preprint arXiv:2205.05739, 2022. 1, 2, 5, 6, 7", + "[44] Sho Maeoki, Kohei Uehara, and Tatsuya Harada. Interactive video retrieval with dialog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 952-953, 2020. 6, 7", + "[45] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 1", + "[46] Antoine Miech, Ivan Laptev, and Josef Sivic. Learning a text-video embedding from incomplete and heterogeneous data. arXiv preprint arXiv:1804.02516, 2018. 1, 2", + "[47] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1", + "[48] Nasrin Mostafazadeh, Ishan Misra, Jacob Devlin, Margaret Mitchell, Xiaodong He, and Lucy Vanderwende. Generating natural questions about an image. arXiv preprint arXiv:1603.06059, 2016. 3", + "[49] Aishwarya Padmakumar and Raymond J Mooney. Dialog policy learning for joint clarification and active learning queries. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 13604-13612, 2021. 2", + "[50] Devi Parikh and Kristen Grauman. Relative attributes. In 2011 International Conference on Computer Vision, pages 503-510. IEEE, 2011. 2", + "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11100", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2", + "[52] Milos Radovanovic, Alexandros Nanopoulos, and Mirjana Ivanovic. Hubs in space: Popular nearest neighbors in high-dimensional data. Journal of Machine Learning Research, 11(sept):2487-2531, 2010. 2", + "[53] Yong Rui, Thomas S Huang, Michael Ortega, and Sharad Mehrotra. Relevance feedback: a power tool for interactive content-based image retrieval. IEEE Transactions on circuits and systems for video technology, 8(5):644-655, 1998. 1, 2", + "[54] Victor Sanh, Albert Webson, Colin Raffel, Stephen Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey, M Saiful Bari, Canwen Xu, Urmish Thakker, Shanya Sharma Sharma, Eliza Szczechla, Taewoon Kim, Gunjan Chhablani, Nihal Nayak, Debajyoti Datta, Jonathan Chang, Mike Tian-Jian Jiang, Han Wang, Matteo Manica, Sheng Shen, Zheng Xin Yong, Harshit Pandey, Rachel Bawden, Thomas Wang, Trishala Neeraj, Jos Rozen, Abheesht Sharma, Andrea Santilli, Thibault Fevy, Jason Alan Fries, Ryan Teehan, Teven Le Scao, Stella Biderman, Leo Gao, Thomas Wolf, and Alexander M Rush. Multitask prompted training enables zero-shot task generalization. In International Conference on Learning Representations, 2022. 5, 6", + "[55] Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. Multitask prompted training enables zero-shot task generalization. arXiv preprint arXiv:2110.08207, 2021. 8", + "[56] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7464-7473, 2019. 2", + "[57] Fuwen Tan, Paola Cascante-Bonilla, Xiaoxiao Guo, Hui Wu, Song Feng, and Vicente Ordonez. Drill-down: Interactive retrieval of complex scenes using natural language queries. Advances in neural information processing systems, 32, 2019. 1, 2", + "[58] Atousa Torabi, Niket Tandon, and Leonid Sigal. Learning language-visual embedding for movie understanding with natural-language. arXiv preprint arXiv:1609.08124, 2016. 2", + "[59] Zeyu Wang, Yu Wu, Karthik Narasimhan, and Olga Russakovsky. Multi-query video retrieval. arXiv preprint arXiv:2201.03639, 2022. 3", + "[60] Hong Wu, Hanqing Lu, and Songde Ma. Willhunter: interactive image retrieval with multilevel relevance. In Proceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004., volume 2, pages 1009-1012. IEEE, 2004. 2", + "[61] Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and textual question answering. In International conference on machine learning, pages 2397-2406. PMLR, 2016. 2", + "[62] Huijuan Xu and Kate Saenko. Ask, attend and answer: Exploring question-guided spatial attention for visual question" + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "answering. In European conference on computer vision, pages 451-466. Springer, 2016. 2", + "[63] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 5", + "[64] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1686-1697, 2021. 2", + "[65] Xinru Yang, Haozhi Qi, Mingyang Li, and Alexander Hauptmann. From a glance to\" gotcha\": Interactive facial image retrieval with progressive relevance feedback. arXiv preprint arXiv:2007.15683, 2020. 2", + "[66] Zichao Yang, Xiaodong He, Jianfeng Gao, Li Deng, and Alex Smola. Stacked attention networks for image question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 21-29, 2016. 2", + "[67] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 471-487, 2018. 2", + "[68] Youngjae Yu, Hyungjin Ko, Jongwook Choi, and Gunhee Kim. End-to-end concept word detection for video captioning, retrieval, and question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3165-3173, 2017. 2", + "[69] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16375-16387, 2022. 2", + "[70] Zheng-Jun Zha, Jiawei Liu, Tianhao Yang, and Yongdong Zhang. Spatiotemporal-textual co-attention network for video question answering. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 15(2s):1-18, 2019. 2", + "[71] Shuai Zhao, Linchao Zhu, Xiaohan Wang, and Yi Yang. Centerclip: Token clustering for efficient text-video retrieval. arXiv preprint arXiv:2205.00823, 2022. 7" + ], + "bbox": [ + 501, + 92, + 893, + 685 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11101", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Simple Baselines for Interactive Video Retrieval with Questions and Answers/149533a1-79f1-428b-885a-6c528f5bd0a8_model.json b/2023/Simple Baselines for Interactive Video Retrieval with Questions and Answers/149533a1-79f1-428b-885a-6c528f5bd0a8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..964e8417268b18d68ff29142f6a3a8174ab514ed --- /dev/null +++ b/2023/Simple Baselines for Interactive Video Retrieval with Questions and Answers/149533a1-79f1-428b-885a-6c528f5bd0a8_model.json @@ -0,0 +1,2609 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.131, + 0.873, + 0.153 + ], + "angle": 0, + "content": "Simple Baselines for Interactive Video Retrieval with Questions and Answers" + }, + { + "type": "title", + "bbox": [ + 0.265, + 0.182, + 0.437, + 0.217 + ], + "angle": 0, + "content": "Kaiqu Liang \nPrinceton University" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.22, + 0.443, + 0.234 + ], + "angle": 0, + "content": "k12471@princeton.edu" + }, + { + "type": "title", + "bbox": [ + 0.508, + 0.182, + 0.707, + 0.217 + ], + "angle": 0, + "content": "Samuel Albanie \nUniversity of Cambridge" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.221, + 0.675, + 0.233 + ], + "angle": 0, + "content": "sma71@cam.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.313, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.543 + ], + "angle": 0, + "content": "To date, the majority of video retrieval systems have been optimized for a \"single-shot\" scenario in which the user submits a query in isolation, ignoring previous interactions with the system. Recently, there has been renewed interest in interactive systems to enhance retrieval, but existing approaches are complex and deliver limited gains in performance. In this work, we revisit this topic and propose several simple yet effective baselines for interactive video retrieval via question-answering. We employ a VideoQA model to simulate user interactions and show that this enables the productive study of the interactive retrieval task without access to ground truth dialogue data. Experiments on MSR-VTT, MSVD, and AVSD show that our framework using question-based interaction significantly improves the performance of text-based video retrieval systems. Code is available at https://github.com/kevinliang888/IVR-QA-baselines." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.573, + 0.208, + 0.589 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.471, + 0.734 + ], + "angle": 0, + "content": "Given the surging popularity of video content, there is a pressing need to develop tools that enable users to search video collections accurately and efficiently. The text-video retrieval task aims to meet this need by ranking videos among a gallery according to how well they match a given text query. This task has seen tremendous progress in recent years, benefiting from large-scale pretraining datasets [47, 41], multimodal architectures [46, 38, 17] and robust optimisation strategies for cross-modal representation learning [45]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Much of the work to date has focused on a \"single-shot\" scenario in which a user submits a single query in isolation. While this may be applicable in some scenarios, there are many in which it would be useful to allow refinements to the initial ranking (particularly when the pool of valid responses to the initial query is large). To a large degree, this focus single-shot systems can be explained by practicality: obtaining the data required to develop and evaluate interactive retrieval systems is extremely challenging. User studies that engage individuals in interactive sessions are difficult to scale, while production data from product deployments is" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.272, + 0.891, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.482, + 0.895, + 0.558 + ], + "angle": 0, + "content": "Figure 1: Interactive video retrieval with questions and answers. We revisit interactive video retrieval with a simple framework that employs Video Question Answering (VideoQA) to simulate user responses, sidestepping the need to collect dialogue data for development and evaluation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.592, + 0.874, + 0.607 + ], + "angle": 0, + "content": "challenging to adapt to the development of new systems." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.895, + 0.762 + ], + "angle": 0, + "content": "Although it has received less attention, the topic of interactive retrieval is far from new—it has been studied for more than two decades [53]. Following an initial query, various mechanisms have been offered to the user to enable them to refine results: relevance scores [53] (corresponding to highly relevant, irrelevant, etc.), relative attributes [29] (e.g. \"shiner than these\"), natural language descriptions [19, 57] (e.g. \"unlike the given result, the one I want has fur on the back\") and question-answer dialog [13, 43] (e.g. \"Q: are there any people in the shot?\" A: No\")." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Among dialogue-based approaches, the recent work of Madasu et al. [43] highlighted the potential of question-answering with free-form text questions to enhance video retrieval performance. At the heart of this approach is a question-generating system that probes the user for more details about their desired target in order to refine the retrieval ranking. In such a framework, user responses can be simulated by employing a question-answering model to synthesise plausible answers to the questions (Fig. 1). Our" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11091" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.272 + ], + "angle": 0, + "content": "work draws inspiration from ViReD [43], but differs in two key aspects. (1) The ViReD answer generator works by first captioning, then extracting the answer from the text of the caption. The disadvantage of this approach is that the answer indirectly depends on the visual content. Instead, we adopt a video question-answering model to simulate the answer, incorporating the information from the video directly with the question. (2) ViReD fine-tunes on AVSD with the human-collected dialogue while our approach does not require any fine-tuning. We show that our approach generalises well to multiple datasets, and obtains significantly better performance than ViReD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.473, + 0.485 + ], + "angle": 0, + "content": "In this work, we make the following contributions: (1) We propose a simple, yet effective framework for interactive video retrieval based on video question answering, and explore the design space of question generators within this framework. (2) We develop three simple question generators that leverage heuristics and learned models to gather discriminative information beyond objects (scenes, actions, colour etc.) in the target video. (3) Through careful experiments, we demonstrate that our system generalizes well to several video datasets and obtains significant improvement over baseline without fine-tuning. These results, which can be interpreted as leveraging an approximate interactive oracle, highlight the potential of question-based interactive retrieval frameworks to substantially enhance video search." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.498, + 0.216, + 0.514 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.473, + 0.869 + ], + "angle": 0, + "content": "Text-to-video Retrieval A long-running theme of text-video retrieval research has been the development of effective fusion mechanisms for cross-modal learning [68, 25, 26, 67, 58]. Another theme of vision-language retrieval has sought to tackle challenges associated with hubness [52] in the embedding space [14, 5]. In the past few years, there has been increasing emphasis on the use of pre-trained models that enable retrieval systems to leverage large quantities of training data that is available for related tasks, often spanning multiple modalities [38, 46, 17, 11]. Recently, there has been a particular focus on building atop models that can leverage the most scalable forms of pretraining data (e.g. CLIP [51]), which have demonstrated compelling performance across a range of vision and language tasks. Luo et al. [40] proposed CLIP4CLIP which adapts CLIP for text-video retrieval through a combination of architectural changes and finetuning. Other approaches have included the use of a curriculum to learn from images and video [3], BERT-style architectures [56], and additional pretraining tasks [34]. Our approach is built upon BLIP [35], a unified framework for retrieval, question answering, and captioning. Differently from the works listed above, we focus on the interactive retrieval setting." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Visual Question Answering. The goal of visual question-answering systems is to answer questions posed in natural" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.379 + ], + "angle": 0, + "content": "language that concern visual content in the form of an image or video. A broad range of work in this space has sought to combine spatial image representations and sequential question representations [2, 4, 16, 39, 61, 62, 66]. To answer questions about video content, spatio-temporal video representations that encode motion and appearance have been studied in detail [15, 18, 21, 22, 23, 24, 31, 33, 70]. There has also been work that extends these ideas to open-ended visual question answering [20, 64]. Recent work on goal-oriented visual dialog [12, 13, 32, 10, 49] uses policy-based reinforcement learning [49, 12, 13] or optimizing information gain [32] to improve the performance. However, these works require intensive training on human-annotated dialog datasets and do not trivially generalise to video. In terms of VideoQA, a few recent works have explored zero-shot approaches [64, 69], where models are only trained on automatically mined video clips with short text descriptions. Our VideoQA model is based on BLIP [35] which is trained on manually annotated image-question-answer triplets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.897, + 0.683 + ], + "angle": 0, + "content": "Interactive cross-modal retrieval A number of interactive image retrieval systems have been developed, drawing inspiration from visual dialogue [60, 53, 29, 28, 27, 50, 65]. In such systems, users provide feedback to an agent to guide it toward the desired retrieval target. Two families of feedback have been studied in detail: relevance feedback and difference feedback. For the former [60, 53], users provide a relevance score for the current retrieval results. The system then re-ranks its initial results using the feedback provided by the user. For the latter approach [29, 28, 27, 50, 65], users communicate the difference between the target image and a reference image to the system with cues such as tags or descriptions. The system then \"whittles away\" irrelevant images, iteratively filtering to select the desired target. However, one drawback of these systems is that they place a heavy time cost on the user who must describe what is missing in the initial search results. For example, description-based methods [19, 57] require users to provide detailed sentence-level feedback and tag-based methods require users to provide a collection of attributes [27, 28, 29]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.895, + 0.822 + ], + "angle": 0, + "content": "To address these concerns, Cai et al [7] propose a partial query system in which potentially discriminative objects are suggested to the user with the goal of narrowing down the search results quickly. However, this system doesn't utilize other cues in the image such as object color, actions, scenes, background, etc., so it can require many interactions with the user to obtain good retrieval results. By using free-form questions to elicit information from the user, our system avoids this restriction to solely object-based queries." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Most related to our approach, the ViReD work of Madasu et al. [43] highlighted the potential of question-answering with free-form text questions to enhance video retrieval performance. However, rather than performing VideoQA directly, ViReD employs captioning followed by text-based" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11092" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.472, + 0.168 + ], + "angle": 0, + "content": "question-answering to answer questions about the video, bottlenecking answer quality. In addition, the ViReD question generator is complex and dependent on dialog history for training—our approach requires no fine-tuning and generalises effectively to multiple datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.185, + 0.168, + 0.2 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.21, + 0.396, + 0.226 + ], + "angle": 0, + "content": "3.1. Interactive video retrieval with Q&A" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.235, + 0.471, + 0.446 + ], + "angle": 0, + "content": "Components. Our interactive video retrieval system contains four components: a Text Encoder, a Visual Encoder, a Ranker, and a Q&A System. The Text Encoder and Visual Encoder embed partial queries and videos into a shared textual-visual space. The Ranker computes the similarity scores between text and video features which are used to rank the videos. To accelerate inference, we adopt the approach of [36] that first selects \\( K \\) candidates based on the video-text feature similarity, and then reranks the selected candidates based on their pairwise Image-Text Matching (ITM) scores. Specifically, the model uses an ITM head (a linear layer) to predict the score indicating whether an image-text pair is positive (matched) or negative (unmatched) given their multimodal feature." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.47, + 0.612 + ], + "angle": 0, + "content": "The Q&A System forms the central module in the framework (Fig. 3). It comprises a captioner adopted from the BLIP caption model, a question generator, and an answer generator that takes the form of a VideoQA (video question-answering) model that simulates user responses. We adopt the BLIP VideoQA model [35]. Instead of formulating VideoQA as a multi-answer classification task [9, 37], Li et al [35] interpret it as an answer generation task, enabling open-ended VideoQA. This open-ended VideoQA allows more freedom to generate and combine answers, lending flexibility to the system." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Interaction workflow. The complete interaction process is demonstrated in Fig. 2. The user provides an initial text query to the system. The interaction loop then begins, comprising four main steps. (1) Generate input to the Q&A System: The current (partial) query together with candidate retrieved videos is passed to the Q&A system. To obtain the candidate videos, we first generate the text embedding corresponding to the partial query and video embeddings for all videos. The Ranker then computes similarity scores in embedding space (via cosine similarity) to rank the videos. We select as candidate videos those who fall within the top \\( K \\) places of the ranking and feed them as input to the Q&A System. (2) Question generation: We explore several kinds of question generators (described in detail in Sec. 3.2 and Sec. 3.3). Different question generators require access to different inputs. For example, for the Auto-text-vid question generation approach (Sec. 3.3), we first generate captions for the candidate videos, then we generate questions based on both the captions and the partial query. (3) User answer sim" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "ulation by VideoQA model: Given the question and the target videos, we use the VideoQA model to generate an answer. (4) New query generation: We generate a complete sentence by incorporating the answer and adding it to the original textual query. This completes one round of interaction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.169, + 0.895, + 0.516 + ], + "angle": 0, + "content": "Combining responses after each round of interaction After each round of interaction, we obtain new information with the potential to improve retrieval performance. It is therefore worth considering how this information is best integrated. One approach is to treat each new piece of information as a separate query, then adopt a strategy for multi-query retrieval. In recent work, Wang et al [59] propose two post hoc inference methods, Similarity aggregation (SA) and Rank aggregation (RA), to extend single-query to multi-query video retrieval. However, initial experiments with these approaches did not prove promising. We believe this is because, while each new piece of information may provide complementary information, it may perform poorly as a query in isolation. However, SA and RA work best with individual queries of high quality. We, therefore, adopt a different approach: we aggregate pieces of information by concatenating them into one large query. In this way, the initial query can be iteratively refined by adding more complementary information. In detail, we concatenate multiple queries using a separation token [SEP]. This token is derived from BERT pretraining, in which it is used to separate sentences as follows: \"[CLS]\" sentence A [SEP] sentence B [SEP]\", suggesting its applicability for separating content in our application setting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.528, + 0.766, + 0.543 + ], + "angle": 0, + "content": "3.2. Heuristic Question Generator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.582 + ], + "angle": 0, + "content": "We now turn the design of question generators used in our system, starting with a simple heuristic question generator." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.839 + ], + "angle": 0, + "content": "The problem of asking useful questions about visual content has received attention in the field of Visual Question Generation. Mostafazadeh et al [48] proposed a framework that generates questions that can potentially engage a human in starting a conversation. They observe that asking a question that can be answered simply by looking at the image would be of interest to the computer vision community, but those questions are neither natural nor engaging for a person to answer. Nevertheless, for the purpose of improving retrieval performance, we focus on concrete, discriminative questions (rather than abstract and engaging questions). In particular, to develop a heuristic question generator, we target goal-driven questions [30] that focus on extracting objects (\"what is the person throwing?\"), attributes (\"what kind of shirt is the person wearing?\"), color (\"what color is the frisbee?\"), material (\"what material is the frisbee?\"), etc." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We are particularly interested in obtaining information that is highly discriminative and complementary to the original query. Ask & Confirm [7] argues that objects provide discriminative cues to distinguish a target from other candi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11093" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.093, + 0.87, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.285, + 0.896, + 0.364 + ], + "angle": 0, + "content": "Figure 2: Model architecture of interactive video retrieval with Q&A (Questions and Answers). The user provides an initial textual query to the system. The interaction loop contains 4 main steps. Step 1: Generate input to the Q&A system. For Heuristic and Auto-text question generation approaches, the input to the Q&A system are just the queries. For Auto-text-vid question generation approach, the input contains both queries and the candidate videos. Step 2: Question generation. Step 3: User answer simulation by VQA model. Step 4: New query generation. This fulfills one round of the interaction." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.389, + 0.465, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.574, + 0.472, + 0.682 + ], + "angle": 0, + "content": "Figure 3: Model architecture of Q&A System. The interaction loop contains either two or three steps depending on which question generation approach is used. Step 1: Caption generation. Step 2: Question generation. Step 3: Answer generation. Auto-text-vid approach contains step 1, 2 and 3 while Heuristic and Auto-text question generation approaches contain only step 2 and 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.715, + 0.471, + 0.792 + ], + "angle": 0, + "content": "dates in the context of image retrieval. Similarly, we observe the importance of the objects in the video retrieval task. However, other cues such as color, human actions, and scenes are also important, and we, therefore, aim to construct a question generator that incorporates these cues." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Heuristic Question Generator. Our first approach to question generation employs hand-crafted heuristics. Our strategy is to first obtain objects from the initial textual query (via pattern matching) and then ask questions about these objects. If an object is a living thing such as a person or animal, then we enquire about both the action and scene associated with it. If the object is a non-living thing such as a lamp or table," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.892, + 0.434 + ], + "angle": 0, + "content": "then we ask only the scene. If no object is found in the initial textual query, we ask the user to identify an object in the scene. The follow-up questions then proceed as above." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.435, + 0.893, + 0.54 + ], + "angle": 0, + "content": "Since living things are often associated with actions, if the object is a living thing, then we employ a template of the form \"who is doing what in what place\". If the object is a non-living thing, the template used is \"who is in what place\". Once the answer is obtained, it is appended to the question, then both are appended to the current query for the next round of video retrieval." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.894, + 0.648 + ], + "angle": 0, + "content": "One potential downside of this approach is the redundancy between questions and the query. For example, suppose the initial textual query is \"a man is singing\", then when we ask \"what is the man doing?\", the answer may add no new information. In practice, this is typically not the case. We find that the information of the initial query is often restricted to addressing some temporal window within the video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.894, + 0.726 + ], + "angle": 0, + "content": "However, the question provided to the user can often address a larger portion of the video. For the example above, the man may walk through the street in the following frames after singing. The answer generator could encode this information to provide a complementary cue for retrieval." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.787, + 0.756 + ], + "angle": 0, + "content": "3.3. Parametric Question Generators" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.902 + ], + "angle": 0, + "content": "The heuristic question generator captures intuition about what makes for useful questions but lacks flexibility. We therefore explore question generation mechanisms that leverage a large language model that possesses the ability to ask flexible, open-ended questions. The first, Auto-text, asks questions based only on the text query. The second, Autotext-vid, asks questions based on both the text query and the top-\\(K\\) ranked video candidates for the current query (here \\(K\\) is a hyperparameter of the system)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11094" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.183 + ], + "angle": 0, + "content": "The foundation of both question generators is T-0++ [54], a large language model that outperforms or matches GPT-3 [6] on many NLP tasks. Since this language model has been trained on various forms of question answering and shown to be very effective at zero-shot task generalization, we directly apply this model to generate questions without finetuning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.473, + 0.349 + ], + "angle": 0, + "content": "Auto-text. To generate questions from the query, we provide the following template to the language model: \"Suppose you are given the following video descriptions \\( Q_{i} \\), What question would you ask to help you unique identify the video?\" where \\( Q_{0} \\) is the initial query and \\( Q_{i} \\) is the concatenated query used to retrieve videos after the \\( i^{\\text{th}} \\) round of interaction. The new query after each round incorporates the answer to the question through concatenation. Note that in terms of the Auto-text question generation approach, the question depends both on the original query and the new descriptions generated after each round of the interaction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.473, + 0.485 + ], + "angle": 0, + "content": "Auto-text-vid. Differently from Auto-text, Auto-text-vid conditions its question generation on both the text query and the current top-\\(k\\) ranked retrieved videos (selected via cosine similarity for the query corresponding to the current round of interaction). This is because, intuitively, we expect the top-ranked videos selected at each round to share some common information with the target video. The use of the top-\\(k\\) retrieved videos as conditioning information was also considered by [43], who found it beneficial." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.473, + 0.682 + ], + "angle": 0, + "content": "To proceed, we first generate captions for each of the top-ranked videos using the BLIP captioner [35]. Then we use both these captions and the original text query to generate questions. The input sentence to the language model employs the following template: \"Suppose you are given the following video descriptions: \\( C_i \\). What question would you ask to help you unique identify the video described as follows: \\( Q_i \\)?\" where \\( C_i \\) are the generated captions at the \\( i^{\\text{th}} \\) round of interaction and \\( Q_i \\) is the concatenated textual query used to retrieve videos after the \\( i^{\\text{th}} \\) round of interaction. The questions generated are based on those common features, so the machine obtains more information about the target video through the interaction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.473, + 0.775 + ], + "angle": 0, + "content": "Remark. Since the top-ranked videos change at each round, there is some variation in the questions generated by Auto-text-vid. This may provide complementary cues, but also carries the risk of additional noise (both through mistakes in the synthetic captions and through mismatches in content between the candidates and the target)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.295, + 0.802 + ], + "angle": 0, + "content": "3.4. Question Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.472, + 0.87 + ], + "angle": 0, + "content": "To further increase the information extracted from the user about the target video, we also introduce two further simple techniques to augment the questions: Ask Segment (AS) and Ask Object (AO)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Ask Segment (AS) augmentation asks the user questions about each half of the video independently. Videos evolve" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.092, + 0.894, + 0.151 + ], + "angle": 0, + "content": "over time and different objects and actions may take place in different frames. As a consequence, asking the same question about different halves of the video often yields very different answers." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.153, + 0.895, + 0.243 + ], + "angle": 0, + "content": "Ask Object (AO) augmentation enquires about more objects in the scene. This approach is motivated by Ask & Confirm [7] in which Cai et al observe that objects are often highly discriminative. Rather than searching for discriminative objects as candidates through a complex RL policy [7], we simply ask the user which objects are in the scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.894, + 0.275 + ], + "angle": 0, + "content": "Note that these strategies can be combined to generate more discriminative questions for interactive video retrieval." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.29, + 0.625, + 0.307 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.316, + 0.894, + 0.362 + ], + "angle": 0, + "content": "In this section, we first describe the datasets used in our experiments (Sec. 4.1). We then discuss implementation details (Sec. 4.2) and evaluation metrics (Sec. 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.373, + 0.603, + 0.387 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.396, + 0.894, + 0.47 + ], + "angle": 0, + "content": "MSR-VTT [63] is a video benchmark that is widely used for text-video retrieval. It contains 10,000 videos sourced from YouTube. Each video is annotated with 20 natural language descriptions. Following previous work, we report retrieval results on the 1K-A test set." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.472, + 0.895, + 0.563 + ], + "angle": 0, + "content": "MSVD [8] is a dataset initially collected for translation and paraphrase evaluation. It contains 1,970 videos, spanning a duration from 1 second up to 62 seconds. The train, validation, and test splits contain 1,200, 100, and 670 videos, respectively. Each video contains approximately 40 associated sentences in English." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.564, + 0.895, + 0.73 + ], + "angle": 0, + "content": "AVSD (Audio-Visual Scene aware Dialog dataset) [1] is a collection of videos associated with ground truth dialog data. Each video dialog consists of 10 rounds of human-generated questions and answers describing various details about the video content. The AVSD dataset consists of 7985 training samples and 1863 validation samples. We follow prior work [43] and split the validation samples into a new validation split of 863 samples and 1000 test samples. Note that our approach does not make use of the dialog data, since we simulate the user directly. We use only the caption as the initial textual query to retrieve the video." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.741, + 0.719, + 0.757 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.765, + 0.895, + 0.901 + ], + "angle": 0, + "content": "For the visual encoder, we employ a vision transformer with ViT-B/16 which contains 12 layers and 12 heads. This model has a hidden dimensionality of 768, an MLP size of 3072, and an input patch size of \\(16 \\times 16\\). The text transformer used for text encoder is BERTBASE with 12 layers, 12 self-attention heads, and a hidden size of 768. The text transformer used for the image-grounded text encoder is the same as the text encoder except that it uses additional cross-attention layers to model vision-language interactions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11095" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.096, + 0.225, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.106, + 0.411, + 0.131 + ], + "angle": 0, + "content": "a woman demonstrates how to cook chitlins" + }, + { + "type": "image_footnote", + "bbox": [ + 0.318, + 0.147, + 0.376, + 0.158 + ], + "angle": 0, + "content": "Round:0" + }, + { + "type": "image_footnote", + "bbox": [ + 0.318, + 0.166, + 0.377, + 0.177 + ], + "angle": 0, + "content": "Rank : 953" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.196, + 0.221, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.203, + 0.431, + 0.227 + ], + "angle": 0, + "content": "a girl and a man are talking to each other" + }, + { + "type": "image_footnote", + "bbox": [ + 0.317, + 0.242, + 0.375, + 0.254 + ], + "angle": 0, + "content": "Round:0" + }, + { + "type": "image_footnote", + "bbox": [ + 0.317, + 0.262, + 0.377, + 0.273 + ], + "angle": 0, + "content": "Rank:421" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.29, + 0.221, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.3, + 0.429, + 0.325 + ], + "angle": 0, + "content": "A video of someone talking about some girls" + }, + { + "type": "image_footnote", + "bbox": [ + 0.315, + 0.343, + 0.374, + 0.354 + ], + "angle": 0, + "content": "Round:0" + }, + { + "type": "image_footnote", + "bbox": [ + 0.316, + 0.361, + 0.376, + 0.372 + ], + "angle": 0, + "content": "Rank:353" + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.39, + 0.222, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.407, + 0.402, + 0.419 + ], + "angle": 0, + "content": "a man is wearing a cap" + }, + { + "type": "image_footnote", + "bbox": [ + 0.315, + 0.442, + 0.374, + 0.454 + ], + "angle": 0, + "content": "Round:0" + }, + { + "type": "image_footnote", + "bbox": [ + 0.315, + 0.462, + 0.368, + 0.473 + ], + "angle": 0, + "content": "Rank : 39" + }, + { + "type": "image_caption", + "bbox": [ + 0.483, + 0.106, + 0.645, + 0.132 + ], + "angle": 0, + "content": "Q: what is the woman doing? \nA: washing meat" + }, + { + "type": "image_footnote", + "bbox": [ + 0.525, + 0.148, + 0.584, + 0.159 + ], + "angle": 0, + "content": "Round:1" + }, + { + "type": "image_footnote", + "bbox": [ + 0.526, + 0.168, + 0.575, + 0.179 + ], + "angle": 0, + "content": "Rank:5" + }, + { + "type": "image_caption", + "bbox": [ + 0.484, + 0.203, + 0.63, + 0.228 + ], + "angle": 0, + "content": "Q: what is the man doing? \nA: looking at camera" + }, + { + "type": "image_footnote", + "bbox": [ + 0.525, + 0.242, + 0.584, + 0.253 + ], + "angle": 0, + "content": "Round:1" + }, + { + "type": "image_footnote", + "bbox": [ + 0.525, + 0.261, + 0.583, + 0.272 + ], + "angle": 0, + "content": "Rank:74" + }, + { + "type": "image_caption", + "bbox": [ + 0.483, + 0.299, + 0.634, + 0.324 + ], + "angle": 0, + "content": "Q: what are the girls doing? A: looking at each other" + }, + { + "type": "image_footnote", + "bbox": [ + 0.523, + 0.342, + 0.582, + 0.353 + ], + "angle": 0, + "content": "Round:1" + }, + { + "type": "image_footnote", + "bbox": [ + 0.523, + 0.361, + 0.574, + 0.372 + ], + "angle": 0, + "content": "Rank:6" + }, + { + "type": "image_caption", + "bbox": [ + 0.481, + 0.401, + 0.626, + 0.426 + ], + "angle": 0, + "content": "Q: what is the man doing? \nA: clapping" + }, + { + "type": "image_footnote", + "bbox": [ + 0.522, + 0.442, + 0.582, + 0.453 + ], + "angle": 0, + "content": "Round:2" + }, + { + "type": "image_footnote", + "bbox": [ + 0.522, + 0.461, + 0.579, + 0.472 + ], + "angle": 0, + "content": "Rank : 11" + }, + { + "type": "image_caption", + "bbox": [ + 0.694, + 0.106, + 0.824, + 0.13 + ], + "angle": 0, + "content": "Q: where is the meat? \nA: in sink" + }, + { + "type": "image_footnote", + "bbox": [ + 0.736, + 0.148, + 0.794, + 0.159 + ], + "angle": 0, + "content": "Round:3" + }, + { + "type": "image_footnote", + "bbox": [ + 0.736, + 0.168, + 0.785, + 0.179 + ], + "angle": 0, + "content": "Rank:1" + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.203, + 0.828, + 0.226 + ], + "angle": 0, + "content": "Q: where is the mirror? \nA: on wall" + }, + { + "type": "image_footnote", + "bbox": [ + 0.736, + 0.242, + 0.794, + 0.253 + ], + "angle": 0, + "content": "Round:4" + }, + { + "type": "image_footnote", + "bbox": [ + 0.736, + 0.262, + 0.785, + 0.272 + ], + "angle": 0, + "content": "Rank:2" + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.299, + 0.861, + 0.323 + ], + "angle": 0, + "content": "Q: where is the silver earring? \nA: in left ear" + }, + { + "type": "image_footnote", + "bbox": [ + 0.734, + 0.342, + 0.793, + 0.353 + ], + "angle": 0, + "content": "Round:4" + }, + { + "type": "image_footnote", + "bbox": [ + 0.734, + 0.361, + 0.784, + 0.372 + ], + "angle": 0, + "content": "Rank:1" + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.4, + 0.816, + 0.425 + ], + "angle": 0, + "content": "Q: what is the object? A: baseball glove" + }, + { + "type": "image_footnote", + "bbox": [ + 0.733, + 0.442, + 0.792, + 0.453 + ], + "angle": 0, + "content": "Round:3" + }, + { + "type": "image_footnote", + "bbox": [ + 0.733, + 0.462, + 0.783, + 0.473 + ], + "angle": 0, + "content": "Rank:1" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.489, + 0.892, + 0.581 + ], + "angle": 0, + "content": "Figure 4: Qualitative results of the dialogue generated by our Q & A System on MSR-VTT. The questions are generated by the Heuristic Question Generator. The initial queries of the target videos are shown in orange boxes. The questions and answers generated through each interaction are shown in blue boxes. Both the initial retrieval rank and the rank after the interaction are demonstrated. In each case, we observe a boost in performance. Note, however, that the answers are not always correct (e.g. in the fourth row \"A: baseball glove\" when the sport does not appear to be baseball). However, since the answer is \"nearly correct\", it still provides a useful cue to the retrieval model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.607, + 0.47, + 0.669 + ], + "angle": 0, + "content": "The architecture of the image-grounded text decoder is similar to the image-grounded text encoder but it replaces the bi-directional self-attention layers with causal self-attention layers. All videos are resized to \\(384 \\times 384\\) as input." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In our interactive system, we adopt the BLIP retrieval model, the VQA model, and the captioner. The BLIP retrieval model is composed of an image encoder, a text encoder, and an image-grounded text encoder. The BLIP VQA model is composed of an image encoder, an image-grounded text encoder, and an image-grounded text decoder. The BLIP captioner contains an image encoder and an image-grounded text decoder. To perform zero-shot transfer video retrieval, we uniformly sample 8 frames per video and concatenate the frame features into a single sequence. To generate questions, we adopt \\( \\mathrm{T0 + + } \\) [54], a powerful encoder-decoder language model trained on a collection of English-prompted datasets. This model attains strong zero-shot task generalization on English natural language prompts (outperforming GPT-3 on many tasks while being 16x smaller)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.894, + 0.728 + ], + "angle": 0, + "content": "The Auto-text-vid question generation approach depends on the current top-k ranked retrieved videos. We experimented with different k (3, 5, 10, 15) and observed that when \\( \\mathrm{k} = 5 \\), we obtained the best performance. This may be because setting k to be too large produces too much noise in the generated captions while setting k to be too small does not provide an opportunity to extract all potentially discriminative information from related videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Baselines. In addition to the BLIP baseline, we also experimented with the interactive approach in [43] for fair comparisons. As public code is not available at the time of writing, we simulate their answer generation by first captioning the target video (by BLIP captioner) and then using the large language model \\((\\mathrm{T0}++)\\) to generate the answer based on the caption and question. On the AVSD dataset, we also experimented with the approach in [44, 43] using 10 rounds of manually annotated human dialogue history for each video. However, simply concatenating all 10 rounds of ground truth dialog with the initial text query got a slightly" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11096" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.472, + 0.123 + ], + "angle": 0, + "content": "worse performance than the baseline. Instead, we concatenated all the answers with the initial text query for retrieval." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.132, + 0.199, + 0.146 + ], + "angle": 0, + "content": "4.3. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.156, + 0.47, + 0.261 + ], + "angle": 0, + "content": "Metrics. We adopt evaluation metrics that are widely used in the standard single-query retrieval setting and report recall@K (R@K, higher the better) and median rank (MdR, lower the better) in our experiments. R@K calculates the percentage of test data for which the ground-truth video is found in the retrieved K videos. MdR metric captures the median rank of the retrieved ground truth videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.471, + 0.413 + ], + "angle": 0, + "content": "Time cost. We next assess the time cost of producing answers during the interaction. For this, we sample 50 videos from the MSR-VTT dataset. We aim to compare our answer generation approach using the VideoQA model with the approach in ViReD [43] (We call this answer generation approach \\(CAP + LM\\)). In addition, we also conduct a user study on answer generation. We invite 4 users to answer the auto-generated question (Auto-text) on the target videos and record the time they spent on each question. (Note that each user went through all 50 videos before the interaction.)" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.428, + 0.163, + 0.442 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "table", + "bbox": [ + 0.083, + 0.463, + 0.469, + 0.653 + ], + "angle": 0, + "content": "
MethodITAR1 ↑R5 ↑R10 ↑MdR ↓
FiT [3]×31.961.172.63
CLIP4Clip [40]×41.569.479.32
CenterCLIP [71]×48.473.882.02
X-CLIP [42]×49.375.884.82
BLIP [35]×43.365.674.72
BLIP + ViReD [43]✓47.072.381.42
BLIP + Auto-text w/o AO✓47.672.881.82
BLIP + Auto-text-vid w/o AO✓49.072.980.02
BLIP + Auto-text✓62.284.590.71
BLIP + Auto-text-vid✓57.682.288.71
BLIP + Heuristic✓67.989.594.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.672, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Table 1: Comparison to state-of-the-art methods for text-to-video retrieval on the 1k test split of MSRVTT dataset. ITA indicates whether it is an interactive retrieval system. Recall metrics are reported in percent." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In this section, we discuss experimental results. On MSR-VTT (Tab. 1), without AO (asking objects), we observe that both the Auto-text and Auto-text-vid question generation approaches outperform the BLIP baseline. We observe that Auto-text-vid, which makes use of the top-ranked retrieved videos at each round, performs better than Auto-text. With AO, there is a significant performance boost for both approaches. For instance, we observe that the Auto-text Q&A System outperforms the baseline by \\(+18.9\\%\\) in recall@1." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.089, + 0.888, + 0.25 + ], + "angle": 0, + "content": "
MethodITAR@1↑R@5↑R@10↑MdR↓
FiT [3]×33.764.776.33
CLIP4Clip [40]×46.276.184.62
X-CLIP [42]×50.480.6--
CenterCLIP [71]×50.680.388.41
BLIP [35]×44.273.780.92
BLIP + ViReD [43]✓51.078.187.21
BLIP + Auto-text✓68.791.295.81
BLIP + Auto-text-vid✓66.188.893.71
BLIP + Heuristic✓74.293.497.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.267, + 0.895, + 0.314 + ], + "angle": 0, + "content": "Table 2: Comparison to state-of-the-art methods for text-to-video retrieval on the test split of the MSVD dataset. ITA indicates whether it is an interactive retrieval system." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.344, + 0.892, + 0.51 + ], + "angle": 0, + "content": "
MethodITAR@1 ↑R@5 ↑R@10 ↑MdR ↓
FiT [3]×5.618.427.525
LSTM [44]✓4.213.522.1-
FiT + Human dialog [43]✓10.828.940.018
ViReD [43]✓24.949.060.86
BLIP [35]×27.350.061.25.5
BLIP + Human dialog [43]✓27.450.462.75
BLIP + ViReD [43]✓28.454.366.84
BLIP + Auto-text✓36.061.671.63
BLIP + Auto-text-vid✓36.561.071.13
BLIP + Heuristic✓39.569.177.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.53, + 0.894, + 0.575 + ], + "angle": 0, + "content": "Table 3: Comparison to state-of-the-art methods for text-to-video retrieval on the test split of AVSD dataset. ITA indicates whether it is an interactive retrieval system." + }, + { + "type": "table", + "bbox": [ + 0.563, + 0.605, + 0.83, + 0.643 + ], + "angle": 0, + "content": "
VideoQACAP+LMUser
Avg time (s)0.18.05.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.659, + 0.892, + 0.735 + ], + "angle": 0, + "content": "Table 4: Comparison of the average time cost of different answer and question generation approaches. We show the results of the answer generation approach of VQA (ours) and CAP+LM (simulation of ViReD [43] answer generator) and the user study." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Interestingly, we observe that Auto-text outperforms Autotext-vid when used in combination with AO. This suggests that simply asking the user directly for the objects that they are interested in (rather than inferring them from closely ranked videos) offers higher signals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We also observe that the Heuristic question generator yields dramatic gains: the recall@1 is improved by \\(24.6\\%\\) (from \\(43.3\\%\\) to \\(67.9\\%\\)), \\(30\\%\\) (from \\(44.2\\%\\) to \\(74.2\\%\\)) and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11097" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.093, + 0.42, + 0.105 + ], + "angle": 0, + "content": "Retrieval results vs Number of queries" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.106, + 0.451, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.33, + 0.472, + 0.391 + ], + "angle": 0, + "content": "Figure 5: Retrieval result vs Number of queries on MSRVTT. Performance of all different approaches improves as the number of rounds increases. Heuristic question generation approach achieves the most significant improvement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.47, + 0.599 + ], + "angle": 0, + "content": "\\(12.2\\%\\) (from \\(27.3\\%\\) to \\(39.5\\%\\)) on MSR-VTT, MSVD and AVSD respectively. Using Auto-text question generator, the BLIP baseline recall@1 is improved by \\(18.9\\%\\) (from \\(43.3\\%\\) to \\(62.2\\%\\)), \\(24.5\\%\\) (from \\(44.2\\%\\) to \\(68.7\\%\\)), and \\(8.7\\%\\) (from \\(27.3\\%\\) to \\(36.0\\%\\)) on MSR-VTT, MSVD, and AVSD respectively. This indicates that generated questions elicit useful information from the (simulated) user that substantially improves retrieval performance. Compared to prior interactive approaches (ViReD, Human dialog) on AVSD, our approach does not use dialog training data for finetuning but nevertheless outperforms existing work. Interestingly, the heuristic approach performs best over all three different datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.471, + 0.902 + ], + "angle": 0, + "content": "As seen in Fig. 4, a few rounds of interaction are sufficient to locate the large video. We demonstrate the cases where the initial query is incomplete, hindering performance. However, using the question-answering system, the search results quickly improve and we obtain significantly better retrieval performance. The generated questions are of high quality, aiming to confirm discriminative information from the video. In addition, the answers generated by VideoQA model are reasonable in the majority of cases (though not always perfect). They successfully capture different aspects of the information in the video (such as action, location, and object). Another useful property of the generated answers is that they are short phrases, simulating concise user answers. Number of queries. Additional rounds of interaction lead to additional new queries. We observe in Fig. 5 that the retrieval performance increases as the number of queries increases. This effect is clearest for Heuristic question generation approach. The performance of both Auto-text and Auto-text-vid increases faster than their variants without AO. This further confirms that asking users about objects in the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.257 + ], + "angle": 0, + "content": "videos is an effective approach to improve retrieval performance at each round of the interaction. We observe that without AO, recall@1 of Auto-text-vid increases faster than that of Auto-text while the opposite result is observed with AO. This agrees with our argument \"simply asking the user directly for the objects that they are interested in (rather than inferring them from closely ranked videos) offers higher signal\". For all the approaches, Performance stabilizes as the number of queries grows beyond 6. Note that the maximum number of queries for heuristic question generation is fixed to 6 and using 6 rounds works best." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.895, + 0.379 + ], + "angle": 0, + "content": "Computational cost. As shown in Tab. 4, Our approach using the VideoQA model is significantly more efficient than \\(\\mathrm{CAP + LM}\\). This is as expected because two steps (captioning and question-answering) are involved in \\(\\mathrm{CAP + LM}\\) approach while VideoQA directly incorporates the video and question information to generate answers. In our naive implementation, \\(\\mathrm{CAP + LM}\\) is even slower than the user response (though this could be optimised)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.395, + 0.613, + 0.41 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.42, + 0.895, + 0.571 + ], + "angle": 0, + "content": "Limitations. (1) \\(\\mathrm{T0 + + }\\) [55] is good at question-answering tasks but might not be good at generating questions. Therefore fine-tuning the model for the question generation task could be helpful. (2) We adopt the BLIP model for both retrieval and VideoQA. In principle, there is the potential for \"leakage\" in which the VideoQA model encodes information that is specific to BLIP, but irrelevant to the video. In practice, different parameter checkpoints are used for different VideoQA and retrieval components, so we believe that such leakage is unlikely to occur in practice." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.894, + 0.647 + ], + "angle": 0, + "content": "Broader Impact. Improved video retrieval has a wide range of societally useful applications, spanning domains such as education, entertainment, and security. However, as with many technologies, it is subject to dual use. For instance, it could be used to facilitate unlawful surveillance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.663, + 0.619, + 0.679 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.809 + ], + "angle": 0, + "content": "In this work, we proposed several simple baselines for interactive video retrieval based on the generation of questions and answers. Through experiments on standard benchmarks, we demonstrated the effectiveness of generated questions in combination with a VideoQA model to simulate user responses through the interaction process. In future work, we hope to explore the use of question-answering systems to simulate users in other retrieval tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. SA would like to acknowledge Z. Novak and N. Novak in enabling his contribution, and support from an Isaac Newton grant and an EPSRC HPC grant. KL would like to thank Bill Byrne, Olga Russakovsky, Yu Wu, Vishaal Udandarao, Allison Chen, Xindi Wu, and Zhiwei Deng for helpful discussions and feedback." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11098" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Huda Alamri, Vincent Cartillier, Abhishek Das, Jue Wang, Anoop Cherian, Irfan Essa, Dhruv Batra, Tim K Marks, Chiori Hori, Peter Anderson, et al. Audio visual scene-aware dialog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7558-7567, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.472, + 0.337 + ], + "angle": 0, + "content": "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.472, + 0.396 + ], + "angle": 0, + "content": "[4] Hedi Ben-Younes, Rémi Cadene, Matthieu Cord, and Nicolas Thome. Mutan: Multimodal tucker fusion for visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2612-2620, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.398, + 0.472, + 0.464 + ], + "angle": 0, + "content": "[5] Simion-Vlad Bogolin, Ioana Croitoru, Hailin Jin, Yang Liu, and Samuel Albanie. Cross modal retrieval with querybank normalisation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5194-5205, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.472, + 0.536 + ], + "angle": 0, + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.538, + 0.472, + 0.62 + ], + "angle": 0, + "content": "[7] Guanyu Cai, Jun Zhang, Xinyang Jiang, Yifei Gong, Lianghua He, Fufu Yu, Pai Peng, Xiaowei Guo, Feiyue Huang, and Xing Sun. Ask&confirm: active detail enriching for cross-modal retrieval with partial query. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1835-1844, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.622, + 0.472, + 0.704 + ], + "angle": 0, + "content": "[8] David Chen and William Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 190-200, Portland, Oregon, USA, June 2011. Association for Computational Linguistics. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.706, + 0.472, + 0.759 + ], + "angle": 0, + "content": "[9] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.472, + 0.817 + ], + "angle": 0, + "content": "[10] Michael Cogswell, Jiasen Lu, Rishabh Jain, Stefan Lee, Devi Parikh, and Dhruv Batra. Dialog without dialog data: Learning visual dialog agents from vqa data. Advances in Neural Information Processing Systems, 33:19988-19999, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.472, + 0.899 + ], + "angle": 0, + "content": "[11] Ioana Croitoru, Simion-Vlad Bogolin, Marius Leordeanu, Hailin Jin, Andrew Zisserman, Samuel Albanie, and Yang Liu. Teachtext: Crossmodal generalized distillation for text-video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11583-11593, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[12] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, José MF Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 326-335, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[13] Abhishek Das, Satwik Kottur, José MF Moura, Stefan Lee, and Dhruv Batra. Learning cooperative visual dialog agents with deep reinforcement learning. In Proceedings of the IEEE international conference on computer vision, pages 2951–2960, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.894, + 0.275 + ], + "angle": 0, + "content": "[14] Georgiana Dinu, Angeliki Lazaridou, and Marco Baroni. Improving zero-shot learning by mitigating the hubness problem. arXiv preprint arXiv:1412.6568, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.894, + 0.359 + ], + "angle": 0, + "content": "[15] Chenyou Fan, Xiaofan Zhang, Shu Zhang, Wensheng Wang, Chi Zhang, and Heng Huang. Heterogeneous memory enhanced multimodal attention model for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1999–2007, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.894, + 0.417 + ], + "angle": 0, + "content": "[16] Akira Fukui, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, and Marcus Rohrbach. Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv preprint arXiv:1606.01847, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.894, + 0.473 + ], + "angle": 0, + "content": "[17] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision, pages 214-229. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.894, + 0.545 + ], + "angle": 0, + "content": "[18] Jiyang Gao, Runzhou Ge, Kan Chen, and Ram Nevatia. Motion-appearance co-memory networks for video question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6576–6585, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.894, + 0.601 + ], + "angle": 0, + "content": "[19] Xiaoxiao Guo, Hui Wu, Yu Cheng, Steven Rennie, Gerald Tesauro, and Rogerio Feris. Dialog-based interactive image retrieval. Advances in neural information processing systems, 31, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.894, + 0.659 + ], + "angle": 0, + "content": "[20] Hexiang Hu, Wei-Lun Chao, and Fei Sha. Learning answer embeddings for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5428-5436, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.894, + 0.729 + ], + "angle": 0, + "content": "[21] Deng Huang, Peihao Chen, Runhao Zeng, Qing Du, Mingkui Tan, and Chuang Gan. Location-aware graph convolutional networks for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11021-11028, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.894, + 0.8 + ], + "angle": 0, + "content": "[22] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2758–2766, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[23] Jianwen Jiang, Ziqiang Chen, Haojie Lin, Xibin Zhao, and Yue Gao. Divide and conquer: Question-guided spatiotemporal contextual attention for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11101-11108, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[24] Pin Jiang and Yahong Han. Reasoning with heterogeneous graph alignment for video question answering. In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11099" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "of the AAAI Conference on Artificial Intelligence, volume 34, pages 11109-11116, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.178 + ], + "angle": 0, + "content": "[25] Dotan Kaufman, Gil Levi, Tal Hassner, and Lior Wolf. Temporal tessellation: A unified approach for video analysis. In Proceedings of the IEEE International Conference on Computer Vision, pages 94-104, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.471, + 0.221 + ], + "angle": 0, + "content": "[26] Ryan Kiros, Ruslan Salakhutdinov, and Richard S Zemel. Unifying visual-semantic embeddings with multimodal neural language models. arXiv preprint arXiv:1411.2539, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.471, + 0.277 + ], + "angle": 0, + "content": "[27] Adriana Kovashka and Kristen Grauman. Attribute pivots for guiding relevance feedback in image search. In Proceedings of the IEEE International Conference on Computer Vision, pages 297-304, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.279, + 0.471, + 0.318 + ], + "angle": 0, + "content": "[28] Adriana Kovashka and Kristen Grauman. Attributes for image retrieval. In Visual Attributes, pages 89-117. Springer, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.322, + 0.471, + 0.377 + ], + "angle": 0, + "content": "[29] Adriana Kovashka, Devi Parikh, and Kristen Grauman. Whittlesearch: Image search with relative attribute feedback. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2973-2980. IEEE, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.379, + 0.471, + 0.434 + ], + "angle": 0, + "content": "[30] Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Information maximizing visual question generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2008–2018, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.436, + 0.471, + 0.503 + ], + "angle": 0, + "content": "[31] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9972–9981, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.506, + 0.471, + 0.561 + ], + "angle": 0, + "content": "[32] Sang-Woo Lee, Yu-Jung Heo, and Byoung-Tak Zhang. Answerer in questioner's mind: Information theoretic approach to goal-oriented visual dialog. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.564, + 0.471, + 0.632 + ], + "angle": 0, + "content": "[33] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.634, + 0.471, + 0.702 + ], + "angle": 0, + "content": "[34] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.471, + 0.759 + ], + "angle": 0, + "content": "[35] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.471, + 0.83 + ], + "angle": 0, + "content": "[36] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.899 + ], + "angle": 0, + "content": "[37] Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[38] Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts. arXiv preprint arXiv:1907.13487, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[39] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Hierarchical question-image co-attention for visual question answering. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.893, + 0.261 + ], + "angle": 0, + "content": "[40] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.894, + 0.318 + ], + "angle": 0, + "content": "[41] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.893, + 0.388 + ], + "angle": 0, + "content": "[42] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 638–647, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.893, + 0.432 + ], + "angle": 0, + "content": "[43] Avinash Madasu, Junier Oliva, and Gedas Bertasius. Learning to retrieve videos by asking questions. arXiv preprint arXiv:2205.05739, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.893, + 0.489 + ], + "angle": 0, + "content": "[44] Sho Maeoki, Kohei Uehara, and Tatsuya Harada. Interactive video retrieval with dialog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 952-953, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.893, + 0.572 + ], + "angle": 0, + "content": "[45] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.893, + 0.617 + ], + "angle": 0, + "content": "[46] Antoine Miech, Ivan Laptev, and Josef Sivic. Learning a text-video embedding from incomplete and heterogeneous data. arXiv preprint arXiv:1804.02516, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.894, + 0.7 + ], + "angle": 0, + "content": "[47] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.893, + 0.757 + ], + "angle": 0, + "content": "[48] Nasrin Mostafazadeh, Ishan Misra, Jacob Devlin, Margaret Mitchell, Xiaodong He, and Lucy Vanderwende. Generating natural questions about an image. arXiv preprint arXiv:1603.06059, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.893, + 0.815 + ], + "angle": 0, + "content": "[49] Aishwarya Padmakumar and Raymond J Mooney. Dialog policy learning for joint clarification and active learning queries. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 13604-13612, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.893, + 0.857 + ], + "angle": 0, + "content": "[50] Devi Parikh and Kristen Grauman. Relative attributes. In 2011 International Conference on Computer Vision, pages 503-510. IEEE, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11100" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.191 + ], + "angle": 0, + "content": "[52] Milos Radovanovic, Alexandros Nanopoulos, and Mirjana Ivanovic. Hubs in space: Popular nearest neighbors in high-dimensional data. Journal of Machine Learning Research, 11(sept):2487-2531, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[53] Yong Rui, Thomas S Huang, Michael Ortega, and Sharad Mehrotra. Relevance feedback: a power tool for interactive content-based image retrieval. IEEE Transactions on circuits and systems for video technology, 8(5):644-655, 1998. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.472, + 0.444 + ], + "angle": 0, + "content": "[54] Victor Sanh, Albert Webson, Colin Raffel, Stephen Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey, M Saiful Bari, Canwen Xu, Urmish Thakker, Shanya Sharma Sharma, Eliza Szczechla, Taewoon Kim, Gunjan Chhablani, Nihal Nayak, Debajyoti Datta, Jonathan Chang, Mike Tian-Jian Jiang, Han Wang, Matteo Manica, Sheng Shen, Zheng Xin Yong, Harshit Pandey, Rachel Bawden, Thomas Wang, Trishala Neeraj, Jos Rozen, Abheesht Sharma, Andrea Santilli, Thibault Fevy, Jason Alan Fries, Ryan Teehan, Teven Le Scao, Stella Biderman, Leo Gao, Thomas Wolf, and Alexander M Rush. Multitask prompted training enables zero-shot task generalization. In International Conference on Learning Representations, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.471, + 0.514 + ], + "angle": 0, + "content": "[55] Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. Multitask prompted training enables zero-shot task generalization. arXiv preprint arXiv:2110.08207, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.471, + 0.586 + ], + "angle": 0, + "content": "[56] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7464-7473, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.471, + 0.655 + ], + "angle": 0, + "content": "[57] Fuwen Tan, Paola Cascante-Bonilla, Xiaoxiao Guo, Hui Wu, Song Feng, and Vicente Ordonez. Drill-down: Interactive retrieval of complex scenes using natural language queries. Advances in neural information processing systems, 32, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.471, + 0.7 + ], + "angle": 0, + "content": "[58] Atousa Torabi, Niket Tandon, and Leonid Sigal. Learning language-visual embedding for movie understanding with natural-language. arXiv preprint arXiv:1609.08124, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.471, + 0.743 + ], + "angle": 0, + "content": "[59] Zeyu Wang, Yu Wu, Karthik Narasimhan, and Olga Russakovsky. Multi-query video retrieval. arXiv preprint arXiv:2201.03639, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.471, + 0.813 + ], + "angle": 0, + "content": "[60] Hong Wu, Hanqing Lu, and Songde Ma. Willhunter: interactive image retrieval with multilevel relevance. In Proceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004., volume 2, pages 1009-1012. IEEE, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[61] Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and textual question answering. In International conference on machine learning, pages 2397-2406. PMLR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[62] Huijuan Xu and Kate Saenko. Ask, attend and answer: Exploring question-guided spatial attention for visual question" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "answering. In European conference on computer vision, pages 451-466. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[63] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.895, + 0.247 + ], + "angle": 0, + "content": "[64] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1686-1697, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.895, + 0.303 + ], + "angle": 0, + "content": "[65] Xinru Yang, Haozhi Qi, Mingyang Li, and Alexander Hauptmann. From a glance to\" gotcha\": Interactive facial image retrieval with progressive relevance feedback. arXiv preprint arXiv:2007.15683, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.895, + 0.361 + ], + "angle": 0, + "content": "[66] Zichao Yang, Xiaodong He, Jianfeng Gao, Li Deng, and Alex Smola. Stacked attention networks for image question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 21-29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.895, + 0.417 + ], + "angle": 0, + "content": "[67] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 471-487, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.895, + 0.488 + ], + "angle": 0, + "content": "[68] Youngjae Yu, Hyungjin Ko, Jongwook Choi, and Gunhee Kim. End-to-end concept word detection for video captioning, retrieval, and question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3165-3173, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.895, + 0.572 + ], + "angle": 0, + "content": "[69] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16375-16387, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.574, + 0.895, + 0.642 + ], + "angle": 0, + "content": "[70] Zheng-Jun Zha, Jiawei Liu, Tianhao Yang, and Yongdong Zhang. Spatiotemporal-textual co-attention network for video question answering. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 15(2s):1-18, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.643, + 0.895, + 0.686 + ], + "angle": 0, + "content": "[71] Shuai Zhao, Linchao Zhu, Xiaohan Wang, and Yi Yang. Centerclip: Token clustering for efficient text-video retrieval. arXiv preprint arXiv:2205.00823, 2022. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11101" + } + ] +] \ No newline at end of file diff --git a/2023/Simple Baselines for Interactive Video Retrieval with Questions and Answers/149533a1-79f1-428b-885a-6c528f5bd0a8_origin.pdf b/2023/Simple Baselines for Interactive Video Retrieval with Questions and Answers/149533a1-79f1-428b-885a-6c528f5bd0a8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a7daf556e9827d20d7870e8ef3d3f374c1bea6f2 --- /dev/null +++ b/2023/Simple Baselines for Interactive Video Retrieval with Questions and Answers/149533a1-79f1-428b-885a-6c528f5bd0a8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d90f640d3d9ab128017c6ff2a021633aaf2ad357a8a0e61c20e28b2b5ce6f322 +size 913182 diff --git a/2023/Simple Baselines for Interactive Video Retrieval with Questions and Answers/full.md b/2023/Simple Baselines for Interactive Video Retrieval with Questions and Answers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5c2e4ba6adb81bfa9523b75269af871d1423d0e1 --- /dev/null +++ b/2023/Simple Baselines for Interactive Video Retrieval with Questions and Answers/full.md @@ -0,0 +1,329 @@ +# Simple Baselines for Interactive Video Retrieval with Questions and Answers + +# Kaiqu Liang +Princeton University + +k12471@princeton.edu + +# Samuel Albanie +University of Cambridge + +sma71@cam.ac.uk + +# Abstract + +To date, the majority of video retrieval systems have been optimized for a "single-shot" scenario in which the user submits a query in isolation, ignoring previous interactions with the system. Recently, there has been renewed interest in interactive systems to enhance retrieval, but existing approaches are complex and deliver limited gains in performance. In this work, we revisit this topic and propose several simple yet effective baselines for interactive video retrieval via question-answering. We employ a VideoQA model to simulate user interactions and show that this enables the productive study of the interactive retrieval task without access to ground truth dialogue data. Experiments on MSR-VTT, MSVD, and AVSD show that our framework using question-based interaction significantly improves the performance of text-based video retrieval systems. Code is available at https://github.com/kevinliang888/IVR-QA-baselines. + +# 1. Introduction + +Given the surging popularity of video content, there is a pressing need to develop tools that enable users to search video collections accurately and efficiently. The text-video retrieval task aims to meet this need by ranking videos among a gallery according to how well they match a given text query. This task has seen tremendous progress in recent years, benefiting from large-scale pretraining datasets [47, 41], multimodal architectures [46, 38, 17] and robust optimisation strategies for cross-modal representation learning [45]. + +Much of the work to date has focused on a "single-shot" scenario in which a user submits a single query in isolation. While this may be applicable in some scenarios, there are many in which it would be useful to allow refinements to the initial ranking (particularly when the pool of valid responses to the initial query is large). To a large degree, this focus single-shot systems can be explained by practicality: obtaining the data required to develop and evaluate interactive retrieval systems is extremely challenging. User studies that engage individuals in interactive sessions are difficult to scale, while production data from product deployments is + +![](images/b199ceb1d8159e2b33ffb88b74e490e23c03f69a21bb18078ce751b246283506.jpg) +Figure 1: Interactive video retrieval with questions and answers. We revisit interactive video retrieval with a simple framework that employs Video Question Answering (VideoQA) to simulate user responses, sidestepping the need to collect dialogue data for development and evaluation. + +challenging to adapt to the development of new systems. + +Although it has received less attention, the topic of interactive retrieval is far from new—it has been studied for more than two decades [53]. Following an initial query, various mechanisms have been offered to the user to enable them to refine results: relevance scores [53] (corresponding to highly relevant, irrelevant, etc.), relative attributes [29] (e.g. "shiner than these"), natural language descriptions [19, 57] (e.g. "unlike the given result, the one I want has fur on the back") and question-answer dialog [13, 43] (e.g. "Q: are there any people in the shot?" A: No"). + +Among dialogue-based approaches, the recent work of Madasu et al. [43] highlighted the potential of question-answering with free-form text questions to enhance video retrieval performance. At the heart of this approach is a question-generating system that probes the user for more details about their desired target in order to refine the retrieval ranking. In such a framework, user responses can be simulated by employing a question-answering model to synthesise plausible answers to the questions (Fig. 1). Our + +work draws inspiration from ViReD [43], but differs in two key aspects. (1) The ViReD answer generator works by first captioning, then extracting the answer from the text of the caption. The disadvantage of this approach is that the answer indirectly depends on the visual content. Instead, we adopt a video question-answering model to simulate the answer, incorporating the information from the video directly with the question. (2) ViReD fine-tunes on AVSD with the human-collected dialogue while our approach does not require any fine-tuning. We show that our approach generalises well to multiple datasets, and obtains significantly better performance than ViReD. + +In this work, we make the following contributions: (1) We propose a simple, yet effective framework for interactive video retrieval based on video question answering, and explore the design space of question generators within this framework. (2) We develop three simple question generators that leverage heuristics and learned models to gather discriminative information beyond objects (scenes, actions, colour etc.) in the target video. (3) Through careful experiments, we demonstrate that our system generalizes well to several video datasets and obtains significant improvement over baseline without fine-tuning. These results, which can be interpreted as leveraging an approximate interactive oracle, highlight the potential of question-based interactive retrieval frameworks to substantially enhance video search. + +# 2. Related work + +Text-to-video Retrieval A long-running theme of text-video retrieval research has been the development of effective fusion mechanisms for cross-modal learning [68, 25, 26, 67, 58]. Another theme of vision-language retrieval has sought to tackle challenges associated with hubness [52] in the embedding space [14, 5]. In the past few years, there has been increasing emphasis on the use of pre-trained models that enable retrieval systems to leverage large quantities of training data that is available for related tasks, often spanning multiple modalities [38, 46, 17, 11]. Recently, there has been a particular focus on building atop models that can leverage the most scalable forms of pretraining data (e.g. CLIP [51]), which have demonstrated compelling performance across a range of vision and language tasks. Luo et al. [40] proposed CLIP4CLIP which adapts CLIP for text-video retrieval through a combination of architectural changes and finetuning. Other approaches have included the use of a curriculum to learn from images and video [3], BERT-style architectures [56], and additional pretraining tasks [34]. Our approach is built upon BLIP [35], a unified framework for retrieval, question answering, and captioning. Differently from the works listed above, we focus on the interactive retrieval setting. + +Visual Question Answering. The goal of visual question-answering systems is to answer questions posed in natural + +language that concern visual content in the form of an image or video. A broad range of work in this space has sought to combine spatial image representations and sequential question representations [2, 4, 16, 39, 61, 62, 66]. To answer questions about video content, spatio-temporal video representations that encode motion and appearance have been studied in detail [15, 18, 21, 22, 23, 24, 31, 33, 70]. There has also been work that extends these ideas to open-ended visual question answering [20, 64]. Recent work on goal-oriented visual dialog [12, 13, 32, 10, 49] uses policy-based reinforcement learning [49, 12, 13] or optimizing information gain [32] to improve the performance. However, these works require intensive training on human-annotated dialog datasets and do not trivially generalise to video. In terms of VideoQA, a few recent works have explored zero-shot approaches [64, 69], where models are only trained on automatically mined video clips with short text descriptions. Our VideoQA model is based on BLIP [35] which is trained on manually annotated image-question-answer triplets. + +Interactive cross-modal retrieval A number of interactive image retrieval systems have been developed, drawing inspiration from visual dialogue [60, 53, 29, 28, 27, 50, 65]. In such systems, users provide feedback to an agent to guide it toward the desired retrieval target. Two families of feedback have been studied in detail: relevance feedback and difference feedback. For the former [60, 53], users provide a relevance score for the current retrieval results. The system then re-ranks its initial results using the feedback provided by the user. For the latter approach [29, 28, 27, 50, 65], users communicate the difference between the target image and a reference image to the system with cues such as tags or descriptions. The system then "whittles away" irrelevant images, iteratively filtering to select the desired target. However, one drawback of these systems is that they place a heavy time cost on the user who must describe what is missing in the initial search results. For example, description-based methods [19, 57] require users to provide detailed sentence-level feedback and tag-based methods require users to provide a collection of attributes [27, 28, 29]. + +To address these concerns, Cai et al [7] propose a partial query system in which potentially discriminative objects are suggested to the user with the goal of narrowing down the search results quickly. However, this system doesn't utilize other cues in the image such as object color, actions, scenes, background, etc., so it can require many interactions with the user to obtain good retrieval results. By using free-form questions to elicit information from the user, our system avoids this restriction to solely object-based queries. + +Most related to our approach, the ViReD work of Madasu et al. [43] highlighted the potential of question-answering with free-form text questions to enhance video retrieval performance. However, rather than performing VideoQA directly, ViReD employs captioning followed by text-based + +question-answering to answer questions about the video, bottlenecking answer quality. In addition, the ViReD question generator is complex and dependent on dialog history for training—our approach requires no fine-tuning and generalises effectively to multiple datasets. + +# 3. Method + +# 3.1. Interactive video retrieval with Q&A + +Components. Our interactive video retrieval system contains four components: a Text Encoder, a Visual Encoder, a Ranker, and a Q&A System. The Text Encoder and Visual Encoder embed partial queries and videos into a shared textual-visual space. The Ranker computes the similarity scores between text and video features which are used to rank the videos. To accelerate inference, we adopt the approach of [36] that first selects $K$ candidates based on the video-text feature similarity, and then reranks the selected candidates based on their pairwise Image-Text Matching (ITM) scores. Specifically, the model uses an ITM head (a linear layer) to predict the score indicating whether an image-text pair is positive (matched) or negative (unmatched) given their multimodal feature. + +The Q&A System forms the central module in the framework (Fig. 3). It comprises a captioner adopted from the BLIP caption model, a question generator, and an answer generator that takes the form of a VideoQA (video question-answering) model that simulates user responses. We adopt the BLIP VideoQA model [35]. Instead of formulating VideoQA as a multi-answer classification task [9, 37], Li et al [35] interpret it as an answer generation task, enabling open-ended VideoQA. This open-ended VideoQA allows more freedom to generate and combine answers, lending flexibility to the system. + +Interaction workflow. The complete interaction process is demonstrated in Fig. 2. The user provides an initial text query to the system. The interaction loop then begins, comprising four main steps. (1) Generate input to the Q&A System: The current (partial) query together with candidate retrieved videos is passed to the Q&A system. To obtain the candidate videos, we first generate the text embedding corresponding to the partial query and video embeddings for all videos. The Ranker then computes similarity scores in embedding space (via cosine similarity) to rank the videos. We select as candidate videos those who fall within the top $K$ places of the ranking and feed them as input to the Q&A System. (2) Question generation: We explore several kinds of question generators (described in detail in Sec. 3.2 and Sec. 3.3). Different question generators require access to different inputs. For example, for the Auto-text-vid question generation approach (Sec. 3.3), we first generate captions for the candidate videos, then we generate questions based on both the captions and the partial query. (3) User answer sim + +ulation by VideoQA model: Given the question and the target videos, we use the VideoQA model to generate an answer. (4) New query generation: We generate a complete sentence by incorporating the answer and adding it to the original textual query. This completes one round of interaction. + +Combining responses after each round of interaction After each round of interaction, we obtain new information with the potential to improve retrieval performance. It is therefore worth considering how this information is best integrated. One approach is to treat each new piece of information as a separate query, then adopt a strategy for multi-query retrieval. In recent work, Wang et al [59] propose two post hoc inference methods, Similarity aggregation (SA) and Rank aggregation (RA), to extend single-query to multi-query video retrieval. However, initial experiments with these approaches did not prove promising. We believe this is because, while each new piece of information may provide complementary information, it may perform poorly as a query in isolation. However, SA and RA work best with individual queries of high quality. We, therefore, adopt a different approach: we aggregate pieces of information by concatenating them into one large query. In this way, the initial query can be iteratively refined by adding more complementary information. In detail, we concatenate multiple queries using a separation token [SEP]. This token is derived from BERT pretraining, in which it is used to separate sentences as follows: "[CLS]" sentence A [SEP] sentence B [SEP]", suggesting its applicability for separating content in our application setting. + +# 3.2. Heuristic Question Generator + +We now turn the design of question generators used in our system, starting with a simple heuristic question generator. + +The problem of asking useful questions about visual content has received attention in the field of Visual Question Generation. Mostafazadeh et al [48] proposed a framework that generates questions that can potentially engage a human in starting a conversation. They observe that asking a question that can be answered simply by looking at the image would be of interest to the computer vision community, but those questions are neither natural nor engaging for a person to answer. Nevertheless, for the purpose of improving retrieval performance, we focus on concrete, discriminative questions (rather than abstract and engaging questions). In particular, to develop a heuristic question generator, we target goal-driven questions [30] that focus on extracting objects ("what is the person throwing?"), attributes ("what kind of shirt is the person wearing?"), color ("what color is the frisbee?"), material ("what material is the frisbee?"), etc. + +We are particularly interested in obtaining information that is highly discriminative and complementary to the original query. Ask & Confirm [7] argues that objects provide discriminative cues to distinguish a target from other candi + +![](images/a53c8d5942e64569028088623a22771cdfeea6ad84b2bb3ab5dcf04c00c7b112.jpg) +Figure 2: Model architecture of interactive video retrieval with Q&A (Questions and Answers). The user provides an initial textual query to the system. The interaction loop contains 4 main steps. Step 1: Generate input to the Q&A system. For Heuristic and Auto-text question generation approaches, the input to the Q&A system are just the queries. For Auto-text-vid question generation approach, the input contains both queries and the candidate videos. Step 2: Question generation. Step 3: User answer simulation by VQA model. Step 4: New query generation. This fulfills one round of the interaction. + +![](images/1fc26f07da1dfb6211fdff7e2c8ef6490422da5819b33f0fc88ffe88a73aadf2.jpg) +Figure 3: Model architecture of Q&A System. The interaction loop contains either two or three steps depending on which question generation approach is used. Step 1: Caption generation. Step 2: Question generation. Step 3: Answer generation. Auto-text-vid approach contains step 1, 2 and 3 while Heuristic and Auto-text question generation approaches contain only step 2 and 3. + +dates in the context of image retrieval. Similarly, we observe the importance of the objects in the video retrieval task. However, other cues such as color, human actions, and scenes are also important, and we, therefore, aim to construct a question generator that incorporates these cues. + +Heuristic Question Generator. Our first approach to question generation employs hand-crafted heuristics. Our strategy is to first obtain objects from the initial textual query (via pattern matching) and then ask questions about these objects. If an object is a living thing such as a person or animal, then we enquire about both the action and scene associated with it. If the object is a non-living thing such as a lamp or table, + +then we ask only the scene. If no object is found in the initial textual query, we ask the user to identify an object in the scene. The follow-up questions then proceed as above. + +Since living things are often associated with actions, if the object is a living thing, then we employ a template of the form "who is doing what in what place". If the object is a non-living thing, the template used is "who is in what place". Once the answer is obtained, it is appended to the question, then both are appended to the current query for the next round of video retrieval. + +One potential downside of this approach is the redundancy between questions and the query. For example, suppose the initial textual query is "a man is singing", then when we ask "what is the man doing?", the answer may add no new information. In practice, this is typically not the case. We find that the information of the initial query is often restricted to addressing some temporal window within the video. + +However, the question provided to the user can often address a larger portion of the video. For the example above, the man may walk through the street in the following frames after singing. The answer generator could encode this information to provide a complementary cue for retrieval. + +# 3.3. Parametric Question Generators + +The heuristic question generator captures intuition about what makes for useful questions but lacks flexibility. We therefore explore question generation mechanisms that leverage a large language model that possesses the ability to ask flexible, open-ended questions. The first, Auto-text, asks questions based only on the text query. The second, Autotext-vid, asks questions based on both the text query and the top- $K$ ranked video candidates for the current query (here $K$ is a hyperparameter of the system). + +The foundation of both question generators is T-0++ [54], a large language model that outperforms or matches GPT-3 [6] on many NLP tasks. Since this language model has been trained on various forms of question answering and shown to be very effective at zero-shot task generalization, we directly apply this model to generate questions without finetuning. + +Auto-text. To generate questions from the query, we provide the following template to the language model: "Suppose you are given the following video descriptions $Q_{i}$ , What question would you ask to help you unique identify the video?" where $Q_{0}$ is the initial query and $Q_{i}$ is the concatenated query used to retrieve videos after the $i^{\text{th}}$ round of interaction. The new query after each round incorporates the answer to the question through concatenation. Note that in terms of the Auto-text question generation approach, the question depends both on the original query and the new descriptions generated after each round of the interaction. + +Auto-text-vid. Differently from Auto-text, Auto-text-vid conditions its question generation on both the text query and the current top- $k$ ranked retrieved videos (selected via cosine similarity for the query corresponding to the current round of interaction). This is because, intuitively, we expect the top-ranked videos selected at each round to share some common information with the target video. The use of the top- $k$ retrieved videos as conditioning information was also considered by [43], who found it beneficial. + +To proceed, we first generate captions for each of the top-ranked videos using the BLIP captioner [35]. Then we use both these captions and the original text query to generate questions. The input sentence to the language model employs the following template: "Suppose you are given the following video descriptions: $C_i$ . What question would you ask to help you unique identify the video described as follows: $Q_i$ ?" where $C_i$ are the generated captions at the $i^{\text{th}}$ round of interaction and $Q_i$ is the concatenated textual query used to retrieve videos after the $i^{\text{th}}$ round of interaction. The questions generated are based on those common features, so the machine obtains more information about the target video through the interaction. + +Remark. Since the top-ranked videos change at each round, there is some variation in the questions generated by Auto-text-vid. This may provide complementary cues, but also carries the risk of additional noise (both through mistakes in the synthetic captions and through mismatches in content between the candidates and the target). + +# 3.4. Question Augmentation + +To further increase the information extracted from the user about the target video, we also introduce two further simple techniques to augment the questions: Ask Segment (AS) and Ask Object (AO). + +Ask Segment (AS) augmentation asks the user questions about each half of the video independently. Videos evolve + +over time and different objects and actions may take place in different frames. As a consequence, asking the same question about different halves of the video often yields very different answers. + +Ask Object (AO) augmentation enquires about more objects in the scene. This approach is motivated by Ask & Confirm [7] in which Cai et al observe that objects are often highly discriminative. Rather than searching for discriminative objects as candidates through a complex RL policy [7], we simply ask the user which objects are in the scene. + +Note that these strategies can be combined to generate more discriminative questions for interactive video retrieval. + +# 4. Experiment + +In this section, we first describe the datasets used in our experiments (Sec. 4.1). We then discuss implementation details (Sec. 4.2) and evaluation metrics (Sec. 4.3). + +# 4.1. Datasets + +MSR-VTT [63] is a video benchmark that is widely used for text-video retrieval. It contains 10,000 videos sourced from YouTube. Each video is annotated with 20 natural language descriptions. Following previous work, we report retrieval results on the 1K-A test set. + +MSVD [8] is a dataset initially collected for translation and paraphrase evaluation. It contains 1,970 videos, spanning a duration from 1 second up to 62 seconds. The train, validation, and test splits contain 1,200, 100, and 670 videos, respectively. Each video contains approximately 40 associated sentences in English. + +AVSD (Audio-Visual Scene aware Dialog dataset) [1] is a collection of videos associated with ground truth dialog data. Each video dialog consists of 10 rounds of human-generated questions and answers describing various details about the video content. The AVSD dataset consists of 7985 training samples and 1863 validation samples. We follow prior work [43] and split the validation samples into a new validation split of 863 samples and 1000 test samples. Note that our approach does not make use of the dialog data, since we simulate the user directly. We use only the caption as the initial textual query to retrieve the video. + +# 4.2. Implementation Details + +For the visual encoder, we employ a vision transformer with ViT-B/16 which contains 12 layers and 12 heads. This model has a hidden dimensionality of 768, an MLP size of 3072, and an input patch size of $16 \times 16$ . The text transformer used for text encoder is BERTBASE with 12 layers, 12 self-attention heads, and a hidden size of 768. The text transformer used for the image-grounded text encoder is the same as the text encoder except that it uses additional cross-attention layers to model vision-language interactions. + +a woman demonstrates how to cook chitlins +Q: what is the woman doing? +A: washing meat +Q: what is the man doing? +A: looking at camera +Q: where is the meat? +A: in sink +Q: where is the mirror? +A: on wall +![](images/4575ab25945bde5373aebff5e1198759d16dde79358ab7a8757f3f4751c5d37c.jpg) +Round:0 +Rank : 953 +Round:1 +Rank:5 +Round:3 +Rank:1 + +a girl and a man are talking to each other +![](images/dbf2ed63ac1a610d91798f37c638c476abdc648f6e7dfdbd41607cc568062a25.jpg) +Round:0 +Rank:421 +Round:1 +Rank:74 +Round:4 +Rank:2 + +A video of someone talking about some girls +Q: what are the girls doing? A: looking at each other +Q: where is the silver earring? +A: in left ear +![](images/79a72c2d13ddf6673743adffd90279d3739a6e32d703fafd83820cef69ac46ef.jpg) +Round:0 +Rank:353 +Round:1 +Rank:6 +Round:4 +Rank:1 + +a man is wearing a cap +Q: what is the man doing? +A: clapping +Q: what is the object? A: baseball glove +Figure 4: Qualitative results of the dialogue generated by our Q & A System on MSR-VTT. The questions are generated by the Heuristic Question Generator. The initial queries of the target videos are shown in orange boxes. The questions and answers generated through each interaction are shown in blue boxes. Both the initial retrieval rank and the rank after the interaction are demonstrated. In each case, we observe a boost in performance. Note, however, that the answers are not always correct (e.g. in the fourth row "A: baseball glove" when the sport does not appear to be baseball). However, since the answer is "nearly correct", it still provides a useful cue to the retrieval model. +![](images/8f4c2d3c350deb4211464c39e1c8bb71851fa4e9c62df3e13670cbb8d90a0cdf.jpg) +Round:0 +Rank : 39 +Round:2 +Rank : 11 +Round:3 +Rank:1 + +The architecture of the image-grounded text decoder is similar to the image-grounded text encoder but it replaces the bi-directional self-attention layers with causal self-attention layers. All videos are resized to $384 \times 384$ as input. + +In our interactive system, we adopt the BLIP retrieval model, the VQA model, and the captioner. The BLIP retrieval model is composed of an image encoder, a text encoder, and an image-grounded text encoder. The BLIP VQA model is composed of an image encoder, an image-grounded text encoder, and an image-grounded text decoder. The BLIP captioner contains an image encoder and an image-grounded text decoder. To perform zero-shot transfer video retrieval, we uniformly sample 8 frames per video and concatenate the frame features into a single sequence. To generate questions, we adopt $\mathrm{T0 + + }$ [54], a powerful encoder-decoder language model trained on a collection of English-prompted datasets. This model attains strong zero-shot task generalization on English natural language prompts (outperforming GPT-3 on many tasks while being 16x smaller). + +The Auto-text-vid question generation approach depends on the current top-k ranked retrieved videos. We experimented with different k (3, 5, 10, 15) and observed that when $\mathrm{k} = 5$ , we obtained the best performance. This may be because setting k to be too large produces too much noise in the generated captions while setting k to be too small does not provide an opportunity to extract all potentially discriminative information from related videos. + +Baselines. In addition to the BLIP baseline, we also experimented with the interactive approach in [43] for fair comparisons. As public code is not available at the time of writing, we simulate their answer generation by first captioning the target video (by BLIP captioner) and then using the large language model $(\mathrm{T0}++)$ to generate the answer based on the caption and question. On the AVSD dataset, we also experimented with the approach in [44, 43] using 10 rounds of manually annotated human dialogue history for each video. However, simply concatenating all 10 rounds of ground truth dialog with the initial text query got a slightly + +worse performance than the baseline. Instead, we concatenated all the answers with the initial text query for retrieval. + +# 4.3. Evaluation + +Metrics. We adopt evaluation metrics that are widely used in the standard single-query retrieval setting and report recall@K (R@K, higher the better) and median rank (MdR, lower the better) in our experiments. R@K calculates the percentage of test data for which the ground-truth video is found in the retrieved K videos. MdR metric captures the median rank of the retrieved ground truth videos. + +Time cost. We next assess the time cost of producing answers during the interaction. For this, we sample 50 videos from the MSR-VTT dataset. We aim to compare our answer generation approach using the VideoQA model with the approach in ViReD [43] (We call this answer generation approach $CAP + LM$ ). In addition, we also conduct a user study on answer generation. We invite 4 users to answer the auto-generated question (Auto-text) on the target videos and record the time they spent on each question. (Note that each user went through all 50 videos before the interaction.) + +# 5. Results + +
MethodITAR1 ↑R5 ↑R10 ↑MdR ↓
FiT [3]×31.961.172.63
CLIP4Clip [40]×41.569.479.32
CenterCLIP [71]×48.473.882.02
X-CLIP [42]×49.375.884.82
BLIP [35]×43.365.674.72
BLIP + ViReD [43]✓47.072.381.42
BLIP + Auto-text w/o AO✓47.672.881.82
BLIP + Auto-text-vid w/o AO✓49.072.980.02
BLIP + Auto-text✓62.284.590.71
BLIP + Auto-text-vid✓57.682.288.71
BLIP + Heuristic✓67.989.594.91
+ +In this section, we discuss experimental results. On MSR-VTT (Tab. 1), without AO (asking objects), we observe that both the Auto-text and Auto-text-vid question generation approaches outperform the BLIP baseline. We observe that Auto-text-vid, which makes use of the top-ranked retrieved videos at each round, performs better than Auto-text. With AO, there is a significant performance boost for both approaches. For instance, we observe that the Auto-text Q&A System outperforms the baseline by $+18.9\%$ in recall@1. + +Table 1: Comparison to state-of-the-art methods for text-to-video retrieval on the 1k test split of MSRVTT dataset. ITA indicates whether it is an interactive retrieval system. Recall metrics are reported in percent. + +
MethodITAR@1↑R@5↑R@10↑MdR↓
FiT [3]×33.764.776.33
CLIP4Clip [40]×46.276.184.62
X-CLIP [42]×50.480.6--
CenterCLIP [71]×50.680.388.41
BLIP [35]×44.273.780.92
BLIP + ViReD [43]✓51.078.187.21
BLIP + Auto-text✓68.791.295.81
BLIP + Auto-text-vid✓66.188.893.71
BLIP + Heuristic✓74.293.497.91
+ +Table 2: Comparison to state-of-the-art methods for text-to-video retrieval on the test split of the MSVD dataset. ITA indicates whether it is an interactive retrieval system. + +
MethodITAR@1 ↑R@5 ↑R@10 ↑MdR ↓
FiT [3]×5.618.427.525
LSTM [44]✓4.213.522.1-
FiT + Human dialog [43]✓10.828.940.018
ViReD [43]✓24.949.060.86
BLIP [35]×27.350.061.25.5
BLIP + Human dialog [43]✓27.450.462.75
BLIP + ViReD [43]✓28.454.366.84
BLIP + Auto-text✓36.061.671.63
BLIP + Auto-text-vid✓36.561.071.13
BLIP + Heuristic✓39.569.177.82
+ +Table 3: Comparison to state-of-the-art methods for text-to-video retrieval on the test split of AVSD dataset. ITA indicates whether it is an interactive retrieval system. + +
VideoQACAP+LMUser
Avg time (s)0.18.05.7
+ +Table 4: Comparison of the average time cost of different answer and question generation approaches. We show the results of the answer generation approach of VQA (ours) and CAP+LM (simulation of ViReD [43] answer generator) and the user study. + +Interestingly, we observe that Auto-text outperforms Autotext-vid when used in combination with AO. This suggests that simply asking the user directly for the objects that they are interested in (rather than inferring them from closely ranked videos) offers higher signals. + +We also observe that the Heuristic question generator yields dramatic gains: the recall@1 is improved by $24.6\%$ (from $43.3\%$ to $67.9\%$ ), $30\%$ (from $44.2\%$ to $74.2\%$ ) and + +![](images/3225b0ef270f835d04b1226e6990e1c085dab9a82b8bf007c2e79ea478cfde7e.jpg) +Retrieval results vs Number of queries +Figure 5: Retrieval result vs Number of queries on MSRVTT. Performance of all different approaches improves as the number of rounds increases. Heuristic question generation approach achieves the most significant improvement. + +$12.2\%$ (from $27.3\%$ to $39.5\%$ ) on MSR-VTT, MSVD and AVSD respectively. Using Auto-text question generator, the BLIP baseline recall@1 is improved by $18.9\%$ (from $43.3\%$ to $62.2\%$ ), $24.5\%$ (from $44.2\%$ to $68.7\%$ ), and $8.7\%$ (from $27.3\%$ to $36.0\%$ ) on MSR-VTT, MSVD, and AVSD respectively. This indicates that generated questions elicit useful information from the (simulated) user that substantially improves retrieval performance. Compared to prior interactive approaches (ViReD, Human dialog) on AVSD, our approach does not use dialog training data for finetuning but nevertheless outperforms existing work. Interestingly, the heuristic approach performs best over all three different datasets. + +As seen in Fig. 4, a few rounds of interaction are sufficient to locate the large video. We demonstrate the cases where the initial query is incomplete, hindering performance. However, using the question-answering system, the search results quickly improve and we obtain significantly better retrieval performance. The generated questions are of high quality, aiming to confirm discriminative information from the video. In addition, the answers generated by VideoQA model are reasonable in the majority of cases (though not always perfect). They successfully capture different aspects of the information in the video (such as action, location, and object). Another useful property of the generated answers is that they are short phrases, simulating concise user answers. Number of queries. Additional rounds of interaction lead to additional new queries. We observe in Fig. 5 that the retrieval performance increases as the number of queries increases. This effect is clearest for Heuristic question generation approach. The performance of both Auto-text and Auto-text-vid increases faster than their variants without AO. This further confirms that asking users about objects in the + +videos is an effective approach to improve retrieval performance at each round of the interaction. We observe that without AO, recall@1 of Auto-text-vid increases faster than that of Auto-text while the opposite result is observed with AO. This agrees with our argument "simply asking the user directly for the objects that they are interested in (rather than inferring them from closely ranked videos) offers higher signal". For all the approaches, Performance stabilizes as the number of queries grows beyond 6. Note that the maximum number of queries for heuristic question generation is fixed to 6 and using 6 rounds works best. + +Computational cost. As shown in Tab. 4, Our approach using the VideoQA model is significantly more efficient than $\mathrm{CAP + LM}$ . This is as expected because two steps (captioning and question-answering) are involved in $\mathrm{CAP + LM}$ approach while VideoQA directly incorporates the video and question information to generate answers. In our naive implementation, $\mathrm{CAP + LM}$ is even slower than the user response (though this could be optimised). + +# 6. Discussion + +Limitations. (1) $\mathrm{T0 + + }$ [55] is good at question-answering tasks but might not be good at generating questions. Therefore fine-tuning the model for the question generation task could be helpful. (2) We adopt the BLIP model for both retrieval and VideoQA. In principle, there is the potential for "leakage" in which the VideoQA model encodes information that is specific to BLIP, but irrelevant to the video. In practice, different parameter checkpoints are used for different VideoQA and retrieval components, so we believe that such leakage is unlikely to occur in practice. + +Broader Impact. Improved video retrieval has a wide range of societally useful applications, spanning domains such as education, entertainment, and security. However, as with many technologies, it is subject to dual use. For instance, it could be used to facilitate unlawful surveillance. + +# 7. Conclusion + +In this work, we proposed several simple baselines for interactive video retrieval based on the generation of questions and answers. Through experiments on standard benchmarks, we demonstrated the effectiveness of generated questions in combination with a VideoQA model to simulate user responses through the interaction process. In future work, we hope to explore the use of question-answering systems to simulate users in other retrieval tasks. + +Acknowledgements. SA would like to acknowledge Z. Novak and N. Novak in enabling his contribution, and support from an Isaac Newton grant and an EPSRC HPC grant. KL would like to thank Bill Byrne, Olga Russakovsky, Yu Wu, Vishaal Udandarao, Allison Chen, Xindi Wu, and Zhiwei Deng for helpful discussions and feedback. + +# References + +[1] Huda Alamri, Vincent Cartillier, Abhishek Das, Jue Wang, Anoop Cherian, Irfan Essa, Dhruv Batra, Tim K Marks, Chiori Hori, Peter Anderson, et al. Audio visual scene-aware dialog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7558-7567, 2019. 5 +[2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018. 2 +[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 2, 7 +[4] Hedi Ben-Younes, Rémi Cadene, Matthieu Cord, and Nicolas Thome. Mutan: Multimodal tucker fusion for visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2612-2620, 2017. 2 +[5] Simion-Vlad Bogolin, Ioana Croitoru, Hailin Jin, Yang Liu, and Samuel Albanie. Cross modal retrieval with querybank normalisation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5194-5205, 2022. 2 +[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 5 +[7] Guanyu Cai, Jun Zhang, Xinyang Jiang, Yifei Gong, Lianghua He, Fufu Yu, Pai Peng, Xiaowei Guo, Feiyue Huang, and Xing Sun. Ask&confirm: active detail enriching for cross-modal retrieval with partial query. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1835-1844, 2021. 2, 3, 5 +[8] David Chen and William Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 190-200, Portland, Oregon, USA, June 2011. Association for Computational Linguistics. 5 +[9] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 3 +[10] Michael Cogswell, Jiasen Lu, Rishabh Jain, Stefan Lee, Devi Parikh, and Dhruv Batra. Dialog without dialog data: Learning visual dialog agents from vqa data. Advances in Neural Information Processing Systems, 33:19988-19999, 2020. 2 +[11] Ioana Croitoru, Simion-Vlad Bogolin, Marius Leordeanu, Hailin Jin, Andrew Zisserman, Samuel Albanie, and Yang Liu. Teachtext: Crossmodal generalized distillation for text-video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11583-11593, 2021. 2 + +[12] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, José MF Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 326-335, 2017. 2 +[13] Abhishek Das, Satwik Kottur, José MF Moura, Stefan Lee, and Dhruv Batra. Learning cooperative visual dialog agents with deep reinforcement learning. In Proceedings of the IEEE international conference on computer vision, pages 2951–2960, 2017. 1, 2 +[14] Georgiana Dinu, Angeliki Lazaridou, and Marco Baroni. Improving zero-shot learning by mitigating the hubness problem. arXiv preprint arXiv:1412.6568, 2014. 2 +[15] Chenyou Fan, Xiaofan Zhang, Shu Zhang, Wensheng Wang, Chi Zhang, and Heng Huang. Heterogeneous memory enhanced multimodal attention model for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1999–2007, 2019. 2 +[16] Akira Fukui, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, and Marcus Rohrbach. Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv preprint arXiv:1606.01847, 2016. 2 +[17] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision, pages 214-229. Springer, 2020. 1, 2 +[18] Jiyang Gao, Runzhou Ge, Kan Chen, and Ram Nevatia. Motion-appearance co-memory networks for video question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6576–6585, 2018. 2 +[19] Xiaoxiao Guo, Hui Wu, Yu Cheng, Steven Rennie, Gerald Tesauro, and Rogerio Feris. Dialog-based interactive image retrieval. Advances in neural information processing systems, 31, 2018. 1, 2 +[20] Hexiang Hu, Wei-Lun Chao, and Fei Sha. Learning answer embeddings for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5428-5436, 2018. 2 +[21] Deng Huang, Peihao Chen, Runhao Zeng, Qing Du, Mingkui Tan, and Chuang Gan. Location-aware graph convolutional networks for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11021-11028, 2020. 2 +[22] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2758–2766, 2017. 2 +[23] Jianwen Jiang, Ziqiang Chen, Haojie Lin, Xibin Zhao, and Yue Gao. Divide and conquer: Question-guided spatiotemporal contextual attention for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11101-11108, 2020. 2 +[24] Pin Jiang and Yahong Han. Reasoning with heterogeneous graph alignment for video question answering. In Proceedings + +of the AAAI Conference on Artificial Intelligence, volume 34, pages 11109-11116, 2020. 2 +[25] Dotan Kaufman, Gil Levi, Tal Hassner, and Lior Wolf. Temporal tessellation: A unified approach for video analysis. In Proceedings of the IEEE International Conference on Computer Vision, pages 94-104, 2017. 2 +[26] Ryan Kiros, Ruslan Salakhutdinov, and Richard S Zemel. Unifying visual-semantic embeddings with multimodal neural language models. arXiv preprint arXiv:1411.2539, 2014. 2 +[27] Adriana Kovashka and Kristen Grauman. Attribute pivots for guiding relevance feedback in image search. In Proceedings of the IEEE International Conference on Computer Vision, pages 297-304, 2013. 2 +[28] Adriana Kovashka and Kristen Grauman. Attributes for image retrieval. In Visual Attributes, pages 89-117. Springer, 2017. 2 +[29] Adriana Kovashka, Devi Parikh, and Kristen Grauman. Whittlesearch: Image search with relative attribute feedback. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2973-2980. IEEE, 2012. 1, 2 +[30] Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Information maximizing visual question generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2008–2018, 2019. 3 +[31] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9972–9981, 2020. 2 +[32] Sang-Woo Lee, Yu-Jung Heo, and Byoung-Tak Zhang. Answerer in questioner's mind: Information theoretic approach to goal-oriented visual dialog. Advances in neural information processing systems, 31, 2018. 2 +[33] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 2 +[34] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 2 +[35] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 3, 5, 7 +[36] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 3 +[37] Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV, 2020. 3 + +[38] Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts. arXiv preprint arXiv:1907.13487, 2019. 1, 2 +[39] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Hierarchical question-image co-attention for visual question answering. Advances in neural information processing systems, 29, 2016. 2 +[40] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 2, 7 +[41] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 1 +[42] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 638–647, 2022. 7 +[43] Avinash Madasu, Junier Oliva, and Gedas Bertasius. Learning to retrieve videos by asking questions. arXiv preprint arXiv:2205.05739, 2022. 1, 2, 5, 6, 7 +[44] Sho Maeoki, Kohei Uehara, and Tatsuya Harada. Interactive video retrieval with dialog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 952-953, 2020. 6, 7 +[45] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 1 +[46] Antoine Miech, Ivan Laptev, and Josef Sivic. Learning a text-video embedding from incomplete and heterogeneous data. arXiv preprint arXiv:1804.02516, 2018. 1, 2 +[47] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1 +[48] Nasrin Mostafazadeh, Ishan Misra, Jacob Devlin, Margaret Mitchell, Xiaodong He, and Lucy Vanderwende. Generating natural questions about an image. arXiv preprint arXiv:1603.06059, 2016. 3 +[49] Aishwarya Padmakumar and Raymond J Mooney. Dialog policy learning for joint clarification and active learning queries. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 13604-13612, 2021. 2 +[50] Devi Parikh and Kristen Grauman. Relative attributes. In 2011 International Conference on Computer Vision, pages 503-510. IEEE, 2011. 2 +[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning + +transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2 +[52] Milos Radovanovic, Alexandros Nanopoulos, and Mirjana Ivanovic. Hubs in space: Popular nearest neighbors in high-dimensional data. Journal of Machine Learning Research, 11(sept):2487-2531, 2010. 2 +[53] Yong Rui, Thomas S Huang, Michael Ortega, and Sharad Mehrotra. Relevance feedback: a power tool for interactive content-based image retrieval. IEEE Transactions on circuits and systems for video technology, 8(5):644-655, 1998. 1, 2 +[54] Victor Sanh, Albert Webson, Colin Raffel, Stephen Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey, M Saiful Bari, Canwen Xu, Urmish Thakker, Shanya Sharma Sharma, Eliza Szczechla, Taewoon Kim, Gunjan Chhablani, Nihal Nayak, Debajyoti Datta, Jonathan Chang, Mike Tian-Jian Jiang, Han Wang, Matteo Manica, Sheng Shen, Zheng Xin Yong, Harshit Pandey, Rachel Bawden, Thomas Wang, Trishala Neeraj, Jos Rozen, Abheesht Sharma, Andrea Santilli, Thibault Fevy, Jason Alan Fries, Ryan Teehan, Teven Le Scao, Stella Biderman, Leo Gao, Thomas Wolf, and Alexander M Rush. Multitask prompted training enables zero-shot task generalization. In International Conference on Learning Representations, 2022. 5, 6 +[55] Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. Multitask prompted training enables zero-shot task generalization. arXiv preprint arXiv:2110.08207, 2021. 8 +[56] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7464-7473, 2019. 2 +[57] Fuwen Tan, Paola Cascante-Bonilla, Xiaoxiao Guo, Hui Wu, Song Feng, and Vicente Ordonez. Drill-down: Interactive retrieval of complex scenes using natural language queries. Advances in neural information processing systems, 32, 2019. 1, 2 +[58] Atousa Torabi, Niket Tandon, and Leonid Sigal. Learning language-visual embedding for movie understanding with natural-language. arXiv preprint arXiv:1609.08124, 2016. 2 +[59] Zeyu Wang, Yu Wu, Karthik Narasimhan, and Olga Russakovsky. Multi-query video retrieval. arXiv preprint arXiv:2201.03639, 2022. 3 +[60] Hong Wu, Hanqing Lu, and Songde Ma. Willhunter: interactive image retrieval with multilevel relevance. In Proceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004., volume 2, pages 1009-1012. IEEE, 2004. 2 +[61] Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and textual question answering. In International conference on machine learning, pages 2397-2406. PMLR, 2016. 2 +[62] Huijuan Xu and Kate Saenko. Ask, attend and answer: Exploring question-guided spatial attention for visual question + +answering. In European conference on computer vision, pages 451-466. Springer, 2016. 2 +[63] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 5 +[64] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1686-1697, 2021. 2 +[65] Xinru Yang, Haozhi Qi, Mingyang Li, and Alexander Hauptmann. From a glance to" gotcha": Interactive facial image retrieval with progressive relevance feedback. arXiv preprint arXiv:2007.15683, 2020. 2 +[66] Zichao Yang, Xiaodong He, Jianfeng Gao, Li Deng, and Alex Smola. Stacked attention networks for image question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 21-29, 2016. 2 +[67] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 471-487, 2018. 2 +[68] Youngjae Yu, Hyungjin Ko, Jongwook Choi, and Gunhee Kim. End-to-end concept word detection for video captioning, retrieval, and question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3165-3173, 2017. 2 +[69] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16375-16387, 2022. 2 +[70] Zheng-Jun Zha, Jiawei Liu, Tianhao Yang, and Yongdong Zhang. Spatiotemporal-textual co-attention network for video question answering. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 15(2s):1-18, 2019. 2 +[71] Shuai Zhao, Linchao Zhu, Xiaohan Wang, and Yi Yang. Centerclip: Token clustering for efficient text-video retrieval. arXiv preprint arXiv:2205.00823, 2022. 7 \ No newline at end of file diff --git a/2023/Simple Baselines for Interactive Video Retrieval with Questions and Answers/images.zip b/2023/Simple Baselines for Interactive Video Retrieval with Questions and Answers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..66f784c7eda0a1bc5e97d5c8aef6df3f110145c0 --- /dev/null +++ b/2023/Simple Baselines for Interactive Video Retrieval with Questions and Answers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6edc40b612c449e66accce82c1513ce77cf300021bede19af7b4bfa298563fd +size 284740 diff --git a/2023/Simple Baselines for Interactive Video Retrieval with Questions and Answers/layout.json b/2023/Simple Baselines for Interactive Video Retrieval with Questions and Answers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ce1768a7f464c43ed16db38c7182ee6413446762 --- /dev/null +++ b/2023/Simple Baselines for Interactive Video Retrieval with Questions and Answers/layout.json @@ -0,0 +1,8806 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 534, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 534, + 121 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 534, + 121 + ], + "type": "text", + "content": "Simple Baselines for Interactive Video Retrieval with Questions and Answers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 144, + 267, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 144, + 267, + 171 + ], + "spans": [ + { + "bbox": [ + 162, + 144, + 267, + 171 + ], + "type": "text", + "content": "Kaiqu Liang \nPrinceton University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 162, + 174, + 271, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 174, + 271, + 185 + ], + "spans": [ + { + "bbox": [ + 162, + 174, + 271, + 185 + ], + "type": "text", + "content": "k12471@princeton.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 310, + 144, + 432, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 144, + 432, + 171 + ], + "spans": [ + { + "bbox": [ + 310, + 144, + 432, + 171 + ], + "type": "text", + "content": "Samuel Albanie \nUniversity of Cambridge" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 331, + 175, + 413, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 175, + 413, + 184 + ], + "spans": [ + { + "bbox": [ + 331, + 175, + 413, + 184 + ], + "type": "text", + "content": "sma71@cam.ac.uk" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 238, + 290, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 430 + ], + "type": "text", + "content": "To date, the majority of video retrieval systems have been optimized for a \"single-shot\" scenario in which the user submits a query in isolation, ignoring previous interactions with the system. Recently, there has been renewed interest in interactive systems to enhance retrieval, but existing approaches are complex and deliver limited gains in performance. In this work, we revisit this topic and propose several simple yet effective baselines for interactive video retrieval via question-answering. We employ a VideoQA model to simulate user interactions and show that this enables the productive study of the interactive retrieval task without access to ground truth dialogue data. Experiments on MSR-VTT, MSVD, and AVSD show that our framework using question-based interaction significantly improves the performance of text-based video retrieval systems. Code is available at https://github.com/kevinliang888/IVR-QA-baselines." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 453, + 127, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 127, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 127, + 466 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 474, + 288, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 288, + 581 + ], + "type": "text", + "content": "Given the surging popularity of video content, there is a pressing need to develop tools that enable users to search video collections accurately and efficiently. The text-video retrieval task aims to meet this need by ranking videos among a gallery according to how well they match a given text query. This task has seen tremendous progress in recent years, benefiting from large-scale pretraining datasets [47, 41], multimodal architectures [46, 38, 17] and robust optimisation strategies for cross-modal representation learning [45]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "Much of the work to date has focused on a \"single-shot\" scenario in which a user submits a single query in isolation. While this may be applicable in some scenarios, there are many in which it would be useful to allow refinements to the initial ranking (particularly when the pool of valid responses to the initial query is large). To a large degree, this focus single-shot systems can be explained by practicality: obtaining the data required to develop and evaluate interactive retrieval systems is extremely challenging. User studies that engage individuals in interactive sessions are difficult to scale, while production data from product deployments is" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 215, + 545, + 373 + ], + "blocks": [ + { + "bbox": [ + 310, + 215, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 310, + 215, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 310, + 215, + 545, + 373 + ], + "type": "image", + "image_path": "b199ceb1d8159e2b33ffb88b74e490e23c03f69a21bb18078ce751b246283506.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 381, + 547, + 441 + ], + "lines": [ + { + "bbox": [ + 305, + 381, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 305, + 381, + 547, + 441 + ], + "type": "text", + "content": "Figure 1: Interactive video retrieval with questions and answers. We revisit interactive video retrieval with a simple framework that employs Video Question Answering (VideoQA) to simulate user responses, sidestepping the need to collect dialogue data for development and evaluation." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 468, + 534, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 534, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 534, + 480 + ], + "type": "text", + "content": "challenging to adapt to the development of new systems." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 483, + 547, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 547, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 547, + 603 + ], + "type": "text", + "content": "Although it has received less attention, the topic of interactive retrieval is far from new—it has been studied for more than two decades [53]. Following an initial query, various mechanisms have been offered to the user to enable them to refine results: relevance scores [53] (corresponding to highly relevant, irrelevant, etc.), relative attributes [29] (e.g. \"shiner than these\"), natural language descriptions [19, 57] (e.g. \"unlike the given result, the one I want has fur on the back\") and question-answer dialog [13, 43] (e.g. \"Q: are there any people in the shot?\" A: No\")." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "Among dialogue-based approaches, the recent work of Madasu et al. [43] highlighted the potential of question-answering with free-form text questions to enhance video retrieval performance. At the heart of this approach is a question-generating system that probes the user for more details about their desired target in order to refine the retrieval ranking. In such a framework, user responses can be simulated by employing a question-answering model to synthesise plausible answers to the questions (Fig. 1). Our" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11091" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "work draws inspiration from ViReD [43], but differs in two key aspects. (1) The ViReD answer generator works by first captioning, then extracting the answer from the text of the caption. The disadvantage of this approach is that the answer indirectly depends on the visual content. Instead, we adopt a video question-answering model to simulate the answer, incorporating the information from the video directly with the question. (2) ViReD fine-tunes on AVSD with the human-collected dialogue while our approach does not require any fine-tuning. We show that our approach generalises well to multiple datasets, and obtains significantly better performance than ViReD." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "type": "text", + "content": "In this work, we make the following contributions: (1) We propose a simple, yet effective framework for interactive video retrieval based on video question answering, and explore the design space of question generators within this framework. (2) We develop three simple question generators that leverage heuristics and learned models to gather discriminative information beyond objects (scenes, actions, colour etc.) in the target video. (3) Through careful experiments, we demonstrate that our system generalizes well to several video datasets and obtains significant improvement over baseline without fine-tuning. These results, which can be interpreted as leveraging an approximate interactive oracle, highlight the potential of question-based interactive retrieval frameworks to substantially enhance video search." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 394, + 132, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 132, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 132, + 407 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 415, + 289, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 289, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 289, + 688 + ], + "type": "text", + "content": "Text-to-video Retrieval A long-running theme of text-video retrieval research has been the development of effective fusion mechanisms for cross-modal learning [68, 25, 26, 67, 58]. Another theme of vision-language retrieval has sought to tackle challenges associated with hubness [52] in the embedding space [14, 5]. In the past few years, there has been increasing emphasis on the use of pre-trained models that enable retrieval systems to leverage large quantities of training data that is available for related tasks, often spanning multiple modalities [38, 46, 17, 11]. Recently, there has been a particular focus on building atop models that can leverage the most scalable forms of pretraining data (e.g. CLIP [51]), which have demonstrated compelling performance across a range of vision and language tasks. Luo et al. [40] proposed CLIP4CLIP which adapts CLIP for text-video retrieval through a combination of architectural changes and finetuning. Other approaches have included the use of a curriculum to learn from images and video [3], BERT-style architectures [56], and additional pretraining tasks [34]. Our approach is built upon BLIP [35], a unified framework for retrieval, question answering, and captioning. Differently from the works listed above, we focus on the interactive retrieval setting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Visual Question Answering. The goal of visual question-answering systems is to answer questions posed in natural" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "type": "text", + "content": "language that concern visual content in the form of an image or video. A broad range of work in this space has sought to combine spatial image representations and sequential question representations [2, 4, 16, 39, 61, 62, 66]. To answer questions about video content, spatio-temporal video representations that encode motion and appearance have been studied in detail [15, 18, 21, 22, 23, 24, 31, 33, 70]. There has also been work that extends these ideas to open-ended visual question answering [20, 64]. Recent work on goal-oriented visual dialog [12, 13, 32, 10, 49] uses policy-based reinforcement learning [49, 12, 13] or optimizing information gain [32] to improve the performance. However, these works require intensive training on human-annotated dialog datasets and do not trivially generalise to video. In terms of VideoQA, a few recent works have explored zero-shot approaches [64, 69], where models are only trained on automatically mined video clips with short text descriptions. Our VideoQA model is based on BLIP [35] which is trained on manually annotated image-question-answer triplets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 301, + 548, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 548, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 548, + 540 + ], + "type": "text", + "content": "Interactive cross-modal retrieval A number of interactive image retrieval systems have been developed, drawing inspiration from visual dialogue [60, 53, 29, 28, 27, 50, 65]. In such systems, users provide feedback to an agent to guide it toward the desired retrieval target. Two families of feedback have been studied in detail: relevance feedback and difference feedback. For the former [60, 53], users provide a relevance score for the current retrieval results. The system then re-ranks its initial results using the feedback provided by the user. For the latter approach [29, 28, 27, 50, 65], users communicate the difference between the target image and a reference image to the system with cues such as tags or descriptions. The system then \"whittles away\" irrelevant images, iteratively filtering to select the desired target. However, one drawback of these systems is that they place a heavy time cost on the user who must describe what is missing in the initial search results. For example, description-based methods [19, 57] require users to provide detailed sentence-level feedback and tag-based methods require users to provide a collection of attributes [27, 28, 29]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 543, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 547, + 651 + ], + "type": "text", + "content": "To address these concerns, Cai et al [7] propose a partial query system in which potentially discriminative objects are suggested to the user with the goal of narrowing down the search results quickly. However, this system doesn't utilize other cues in the image such as object color, actions, scenes, background, etc., so it can require many interactions with the user to obtain good retrieval results. By using free-form questions to elicit information from the user, our system avoids this restriction to solely object-based queries." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "Most related to our approach, the ViReD work of Madasu et al. [43] highlighted the potential of question-answering with free-form text questions to enhance video retrieval performance. However, rather than performing VideoQA directly, ViReD employs captioning followed by text-based" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11092" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "question-answering to answer questions about the video, bottlenecking answer quality. In addition, the ViReD question generator is complex and dependent on dialog history for training—our approach requires no fine-tuning and generalises effectively to multiple datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 146, + 102, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 146, + 102, + 158 + ], + "spans": [ + { + "bbox": [ + 47, + 146, + 102, + 158 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 166, + 242, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 166, + 242, + 178 + ], + "spans": [ + { + "bbox": [ + 47, + 166, + 242, + 178 + ], + "type": "text", + "content": "3.1. Interactive video retrieval with Q&A" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 186, + 288, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 288, + 353 + ], + "type": "text", + "content": "Components. Our interactive video retrieval system contains four components: a Text Encoder, a Visual Encoder, a Ranker, and a Q&A System. The Text Encoder and Visual Encoder embed partial queries and videos into a shared textual-visual space. The Ranker computes the similarity scores between text and video features which are used to rank the videos. To accelerate inference, we adopt the approach of [36] that first selects " + }, + { + "bbox": [ + 46, + 186, + 288, + 353 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 186, + 288, + 353 + ], + "type": "text", + "content": " candidates based on the video-text feature similarity, and then reranks the selected candidates based on their pairwise Image-Text Matching (ITM) scores. Specifically, the model uses an ITM head (a linear layer) to predict the score indicating whether an image-text pair is positive (matched) or negative (unmatched) given their multimodal feature." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 354, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 484 + ], + "type": "text", + "content": "The Q&A System forms the central module in the framework (Fig. 3). It comprises a captioner adopted from the BLIP caption model, a question generator, and an answer generator that takes the form of a VideoQA (video question-answering) model that simulates user responses. We adopt the BLIP VideoQA model [35]. Instead of formulating VideoQA as a multi-answer classification task [9, 37], Li et al [35] interpret it as an answer generation task, enabling open-ended VideoQA. This open-ended VideoQA allows more freedom to generate and combine answers, lending flexibility to the system." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": "Interaction workflow. The complete interaction process is demonstrated in Fig. 2. The user provides an initial text query to the system. The interaction loop then begins, comprising four main steps. (1) Generate input to the Q&A System: The current (partial) query together with candidate retrieved videos is passed to the Q&A system. To obtain the candidate videos, we first generate the text embedding corresponding to the partial query and video embeddings for all videos. The Ranker then computes similarity scores in embedding space (via cosine similarity) to rank the videos. We select as candidate videos those who fall within the top " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": " places of the ranking and feed them as input to the Q&A System. (2) Question generation: We explore several kinds of question generators (described in detail in Sec. 3.2 and Sec. 3.3). Different question generators require access to different inputs. For example, for the Auto-text-vid question generation approach (Sec. 3.3), we first generate captions for the candidate videos, then we generate questions based on both the captions and the partial query. (3) User answer sim" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": "ulation by VideoQA model: Given the question and the target videos, we use the VideoQA model to generate an answer. (4) New query generation: We generate a complete sentence by incorporating the answer and adding it to the original textual query. This completes one round of interaction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 133, + 547, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 547, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 547, + 408 + ], + "type": "text", + "content": "Combining responses after each round of interaction After each round of interaction, we obtain new information with the potential to improve retrieval performance. It is therefore worth considering how this information is best integrated. One approach is to treat each new piece of information as a separate query, then adopt a strategy for multi-query retrieval. In recent work, Wang et al [59] propose two post hoc inference methods, Similarity aggregation (SA) and Rank aggregation (RA), to extend single-query to multi-query video retrieval. However, initial experiments with these approaches did not prove promising. We believe this is because, while each new piece of information may provide complementary information, it may perform poorly as a query in isolation. However, SA and RA work best with individual queries of high quality. We, therefore, adopt a different approach: we aggregate pieces of information by concatenating them into one large query. In this way, the initial query can be iteratively refined by adding more complementary information. In detail, we concatenate multiple queries using a separation token [SEP]. This token is derived from BERT pretraining, in which it is used to separate sentences as follows: \"[CLS]\" sentence A [SEP] sentence B [SEP]\", suggesting its applicability for separating content in our application setting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 418, + 468, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 468, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 468, + 430 + ], + "type": "text", + "content": "3.2. Heuristic Question Generator" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 437, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 460 + ], + "type": "text", + "content": "We now turn the design of question generators used in our system, starting with a simple heuristic question generator." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 462, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 664 + ], + "type": "text", + "content": "The problem of asking useful questions about visual content has received attention in the field of Visual Question Generation. Mostafazadeh et al [48] proposed a framework that generates questions that can potentially engage a human in starting a conversation. They observe that asking a question that can be answered simply by looking at the image would be of interest to the computer vision community, but those questions are neither natural nor engaging for a person to answer. Nevertheless, for the purpose of improving retrieval performance, we focus on concrete, discriminative questions (rather than abstract and engaging questions). In particular, to develop a heuristic question generator, we target goal-driven questions [30] that focus on extracting objects (\"what is the person throwing?\"), attributes (\"what kind of shirt is the person wearing?\"), color (\"what color is the frisbee?\"), material (\"what material is the frisbee?\"), etc." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "We are particularly interested in obtaining information that is highly discriminative and complementary to the original query. Ask & Confirm [7] argues that objects provide discriminative cues to distinguish a target from other candi" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11093" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 73, + 532, + 216 + ], + "blocks": [ + { + "bbox": [ + 50, + 73, + 532, + 216 + ], + "lines": [ + { + "bbox": [ + 50, + 73, + 532, + 216 + ], + "spans": [ + { + "bbox": [ + 50, + 73, + 532, + 216 + ], + "type": "image", + "image_path": "a53c8d5942e64569028088623a22771cdfeea6ad84b2bb3ab5dcf04c00c7b112.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 225, + 548, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 225, + 548, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 548, + 288 + ], + "type": "text", + "content": "Figure 2: Model architecture of interactive video retrieval with Q&A (Questions and Answers). The user provides an initial textual query to the system. The interaction loop contains 4 main steps. Step 1: Generate input to the Q&A system. For Heuristic and Auto-text question generation approaches, the input to the Q&A system are just the queries. For Auto-text-vid question generation approach, the input contains both queries and the candidate videos. Step 2: Question generation. Step 3: User answer simulation by VQA model. Step 4: New query generation. This fulfills one round of the interaction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 308, + 284, + 439 + ], + "blocks": [ + { + "bbox": [ + 52, + 308, + 284, + 439 + ], + "lines": [ + { + "bbox": [ + 52, + 308, + 284, + 439 + ], + "spans": [ + { + "bbox": [ + 52, + 308, + 284, + 439 + ], + "type": "image", + "image_path": "1fc26f07da1dfb6211fdff7e2c8ef6490422da5819b33f0fc88ffe88a73aadf2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 454, + 288, + 540 + ], + "lines": [ + { + "bbox": [ + 46, + 454, + 288, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 454, + 288, + 540 + ], + "type": "text", + "content": "Figure 3: Model architecture of Q&A System. The interaction loop contains either two or three steps depending on which question generation approach is used. Step 1: Caption generation. Step 2: Question generation. Step 3: Answer generation. Auto-text-vid approach contains step 1, 2 and 3 while Heuristic and Auto-text question generation approaches contain only step 2 and 3." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 566, + 288, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 566, + 288, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 288, + 627 + ], + "type": "text", + "content": "dates in the context of image retrieval. Similarly, we observe the importance of the objects in the video retrieval task. However, other cues such as color, human actions, and scenes are also important, and we, therefore, aim to construct a question generator that incorporates these cues." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "Heuristic Question Generator. Our first approach to question generation employs hand-crafted heuristics. Our strategy is to first obtain objects from the initial textual query (via pattern matching) and then ask questions about these objects. If an object is a living thing such as a person or animal, then we enquire about both the action and scene associated with it. If the object is a non-living thing such as a lamp or table," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 307, + 545, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 545, + 343 + ], + "type": "text", + "content": "then we ask only the scene. If no object is found in the initial textual query, we ask the user to identify an object in the scene. The follow-up questions then proceed as above." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 344, + 546, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 546, + 427 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 546, + 427 + ], + "type": "text", + "content": "Since living things are often associated with actions, if the object is a living thing, then we employ a template of the form \"who is doing what in what place\". If the object is a non-living thing, the template used is \"who is in what place\". Once the answer is obtained, it is appended to the question, then both are appended to the current query for the next round of video retrieval." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 429, + 547, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 547, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 547, + 513 + ], + "type": "text", + "content": "One potential downside of this approach is the redundancy between questions and the query. For example, suppose the initial textual query is \"a man is singing\", then when we ask \"what is the man doing?\", the answer may add no new information. In practice, this is typically not the case. We find that the information of the initial query is often restricted to addressing some temporal window within the video." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 514, + 547, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 547, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 547, + 574 + ], + "type": "text", + "content": "However, the question provided to the user can often address a larger portion of the video. For the example above, the man may walk through the street in the following frames after singing. The answer generator could encode this information to provide a complementary cue for retrieval." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 586, + 481, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 481, + 598 + ], + "type": "text", + "content": "3.3. Parametric Question Generators" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "The heuristic question generator captures intuition about what makes for useful questions but lacks flexibility. We therefore explore question generation mechanisms that leverage a large language model that possesses the ability to ask flexible, open-ended questions. The first, Auto-text, asks questions based only on the text query. The second, Autotext-vid, asks questions based on both the text query and the top-" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": " ranked video candidates for the current query (here " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": " is a hyperparameter of the system)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11094" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 144 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 144 + ], + "type": "text", + "content": "The foundation of both question generators is T-0++ [54], a large language model that outperforms or matches GPT-3 [6] on many NLP tasks. Since this language model has been trained on various forms of question answering and shown to be very effective at zero-shot task generalization, we directly apply this model to generate questions without finetuning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 145, + 289, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 289, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 289, + 276 + ], + "type": "text", + "content": "Auto-text. To generate questions from the query, we provide the following template to the language model: \"Suppose you are given the following video descriptions " + }, + { + "bbox": [ + 46, + 145, + 289, + 276 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 46, + 145, + 289, + 276 + ], + "type": "text", + "content": ", What question would you ask to help you unique identify the video?\" where " + }, + { + "bbox": [ + 46, + 145, + 289, + 276 + ], + "type": "inline_equation", + "content": "Q_{0}" + }, + { + "bbox": [ + 46, + 145, + 289, + 276 + ], + "type": "text", + "content": " is the initial query and " + }, + { + "bbox": [ + 46, + 145, + 289, + 276 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 46, + 145, + 289, + 276 + ], + "type": "text", + "content": " is the concatenated query used to retrieve videos after the " + }, + { + "bbox": [ + 46, + 145, + 289, + 276 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 46, + 145, + 289, + 276 + ], + "type": "text", + "content": " round of interaction. The new query after each round incorporates the answer to the question through concatenation. Note that in terms of the Auto-text question generation approach, the question depends both on the original query and the new descriptions generated after each round of the interaction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 277, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 289, + 384 + ], + "type": "text", + "content": "Auto-text-vid. Differently from Auto-text, Auto-text-vid conditions its question generation on both the text query and the current top-" + }, + { + "bbox": [ + 46, + 277, + 289, + 384 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 277, + 289, + 384 + ], + "type": "text", + "content": " ranked retrieved videos (selected via cosine similarity for the query corresponding to the current round of interaction). This is because, intuitively, we expect the top-ranked videos selected at each round to share some common information with the target video. The use of the top-" + }, + { + "bbox": [ + 46, + 277, + 289, + 384 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 277, + 289, + 384 + ], + "type": "text", + "content": " retrieved videos as conditioning information was also considered by [43], who found it beneficial." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 385, + 289, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 289, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 289, + 540 + ], + "type": "text", + "content": "To proceed, we first generate captions for each of the top-ranked videos using the BLIP captioner [35]. Then we use both these captions and the original text query to generate questions. The input sentence to the language model employs the following template: \"Suppose you are given the following video descriptions: " + }, + { + "bbox": [ + 46, + 385, + 289, + 540 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 46, + 385, + 289, + 540 + ], + "type": "text", + "content": ". What question would you ask to help you unique identify the video described as follows: " + }, + { + "bbox": [ + 46, + 385, + 289, + 540 + ], + "type": "inline_equation", + "content": "Q_i" + }, + { + "bbox": [ + 46, + 385, + 289, + 540 + ], + "type": "text", + "content": "?\" where " + }, + { + "bbox": [ + 46, + 385, + 289, + 540 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 46, + 385, + 289, + 540 + ], + "type": "text", + "content": " are the generated captions at the " + }, + { + "bbox": [ + 46, + 385, + 289, + 540 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 46, + 385, + 289, + 540 + ], + "type": "text", + "content": " round of interaction and " + }, + { + "bbox": [ + 46, + 385, + 289, + 540 + ], + "type": "inline_equation", + "content": "Q_i" + }, + { + "bbox": [ + 46, + 385, + 289, + 540 + ], + "type": "text", + "content": " is the concatenated textual query used to retrieve videos after the " + }, + { + "bbox": [ + 46, + 385, + 289, + 540 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 46, + 385, + 289, + 540 + ], + "type": "text", + "content": " round of interaction. The questions generated are based on those common features, so the machine obtains more information about the target video through the interaction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 540, + 289, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 289, + 613 + ], + "type": "text", + "content": "Remark. Since the top-ranked videos change at each round, there is some variation in the questions generated by Auto-text-vid. This may provide complementary cues, but also carries the risk of additional noise (both through mistakes in the synthetic captions and through mismatches in content between the candidates and the target)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 622, + 180, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 180, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 180, + 635 + ], + "type": "text", + "content": "3.4. Question Augmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 689 + ], + "type": "text", + "content": "To further increase the information extracted from the user about the target video, we also introduce two further simple techniques to augment the questions: Ask Segment (AS) and Ask Object (AO)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "content": "Ask Segment (AS) augmentation asks the user questions about each half of the video independently. Videos evolve" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 119 + ], + "type": "text", + "content": "over time and different objects and actions may take place in different frames. As a consequence, asking the same question about different halves of the video often yields very different answers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 121, + 547, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 547, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 547, + 192 + ], + "type": "text", + "content": "Ask Object (AO) augmentation enquires about more objects in the scene. This approach is motivated by Ask & Confirm [7] in which Cai et al observe that objects are often highly discriminative. Rather than searching for discriminative objects as candidates through a complex RL policy [7], we simply ask the user which objects are in the scene." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 193, + 547, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 547, + 217 + ], + "type": "text", + "content": "Note that these strategies can be combined to generate more discriminative questions for interactive video retrieval." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 229, + 382, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 229, + 382, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 229, + 382, + 243 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 250, + 547, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 250, + 547, + 286 + ], + "spans": [ + { + "bbox": [ + 304, + 250, + 547, + 286 + ], + "type": "text", + "content": "In this section, we first describe the datasets used in our experiments (Sec. 4.1). We then discuss implementation details (Sec. 4.2) and evaluation metrics (Sec. 4.3)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 295, + 369, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 295, + 369, + 306 + ], + "spans": [ + { + "bbox": [ + 305, + 295, + 369, + 306 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 313, + 547, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 313, + 547, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 547, + 372 + ], + "type": "text", + "content": "MSR-VTT [63] is a video benchmark that is widely used for text-video retrieval. It contains 10,000 videos sourced from YouTube. Each video is annotated with 20 natural language descriptions. Following previous work, we report retrieval results on the 1K-A test set." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 373, + 547, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 547, + 445 + ], + "type": "text", + "content": "MSVD [8] is a dataset initially collected for translation and paraphrase evaluation. It contains 1,970 videos, spanning a duration from 1 second up to 62 seconds. The train, validation, and test splits contain 1,200, 100, and 670 videos, respectively. Each video contains approximately 40 associated sentences in English." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 446, + 547, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 547, + 578 + ], + "type": "text", + "content": "AVSD (Audio-Visual Scene aware Dialog dataset) [1] is a collection of videos associated with ground truth dialog data. Each video dialog consists of 10 rounds of human-generated questions and answers describing various details about the video content. The AVSD dataset consists of 7985 training samples and 1863 validation samples. We follow prior work [43] and split the validation samples into a new validation split of 863 samples and 1000 test samples. Note that our approach does not make use of the dialog data, since we simulate the user directly. We use only the caption as the initial textual query to retrieve the video." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 586, + 440, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 586, + 440, + 599 + ], + "spans": [ + { + "bbox": [ + 305, + 586, + 440, + 599 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "For the visual encoder, we employ a vision transformer with ViT-B/16 which contains 12 layers and 12 heads. This model has a hidden dimensionality of 768, an MLP size of 3072, and an input patch size of " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": ". The text transformer used for text encoder is BERTBASE with 12 layers, 12 self-attention heads, and a hidden size of 768. The text transformer used for the image-grounded text encoder is the same as the text encoder except that it uses additional cross-attention layers to model vision-language interactions." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11095" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 76, + 137, + 140 + ], + "blocks": [ + { + "bbox": [ + 56, + 76, + 137, + 140 + ], + "lines": [ + { + "bbox": [ + 56, + 76, + 137, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 76, + 137, + 140 + ], + "type": "image", + "image_path": "4575ab25945bde5373aebff5e1198759d16dde79358ab7a8757f3f4751c5d37c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 83, + 251, + 103 + ], + "lines": [ + { + "bbox": [ + 168, + 83, + 251, + 103 + ], + "spans": [ + { + "bbox": [ + 168, + 83, + 251, + 103 + ], + "type": "text", + "content": "a woman demonstrates how to cook chitlins" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 194, + 116, + 230, + 125 + ], + "lines": [ + { + "bbox": [ + 194, + 116, + 230, + 125 + ], + "spans": [ + { + "bbox": [ + 194, + 116, + 230, + 125 + ], + "type": "text", + "content": "Round:0" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 194, + 131, + 230, + 140 + ], + "lines": [ + { + "bbox": [ + 194, + 131, + 230, + 140 + ], + "spans": [ + { + "bbox": [ + 194, + 131, + 230, + 140 + ], + "type": "text", + "content": "Rank : 953" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 295, + 83, + 394, + 104 + ], + "lines": [ + { + "bbox": [ + 295, + 83, + 394, + 104 + ], + "spans": [ + { + "bbox": [ + 295, + 83, + 394, + 104 + ], + "type": "text", + "content": "Q: what is the woman doing? \nA: washing meat" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 117, + 357, + 125 + ], + "lines": [ + { + "bbox": [ + 321, + 117, + 357, + 125 + ], + "spans": [ + { + "bbox": [ + 321, + 117, + 357, + 125 + ], + "type": "text", + "content": "Round:1" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 321, + 133, + 351, + 141 + ], + "lines": [ + { + "bbox": [ + 321, + 133, + 351, + 141 + ], + "spans": [ + { + "bbox": [ + 321, + 133, + 351, + 141 + ], + "type": "text", + "content": "Rank:5" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 296, + 160, + 385, + 180 + ], + "lines": [ + { + "bbox": [ + 296, + 160, + 385, + 180 + ], + "spans": [ + { + "bbox": [ + 296, + 160, + 385, + 180 + ], + "type": "text", + "content": "Q: what is the man doing? \nA: looking at camera" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 424, + 83, + 504, + 102 + ], + "lines": [ + { + "bbox": [ + 424, + 83, + 504, + 102 + ], + "spans": [ + { + "bbox": [ + 424, + 83, + 504, + 102 + ], + "type": "text", + "content": "Q: where is the meat? \nA: in sink" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 450, + 117, + 485, + 125 + ], + "lines": [ + { + "bbox": [ + 450, + 117, + 485, + 125 + ], + "spans": [ + { + "bbox": [ + 450, + 117, + 485, + 125 + ], + "type": "text", + "content": "Round:3" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 450, + 133, + 480, + 141 + ], + "lines": [ + { + "bbox": [ + 450, + 133, + 480, + 141 + ], + "spans": [ + { + "bbox": [ + 450, + 133, + 480, + 141 + ], + "type": "text", + "content": "Rank:1" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 425, + 160, + 506, + 178 + ], + "lines": [ + { + "bbox": [ + 425, + 160, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 425, + 160, + 506, + 178 + ], + "type": "text", + "content": "Q: where is the mirror? \nA: on wall" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 155, + 135, + 216 + ], + "blocks": [ + { + "bbox": [ + 56, + 155, + 135, + 216 + ], + "lines": [ + { + "bbox": [ + 56, + 155, + 135, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 155, + 135, + 216 + ], + "type": "image", + "image_path": "dbf2ed63ac1a610d91798f37c638c476abdc648f6e7dfdbd41607cc568062a25.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 160, + 263, + 179 + ], + "lines": [ + { + "bbox": [ + 169, + 160, + 263, + 179 + ], + "spans": [ + { + "bbox": [ + 169, + 160, + 263, + 179 + ], + "type": "text", + "content": "a girl and a man are talking to each other" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 194, + 191, + 229, + 201 + ], + "lines": [ + { + "bbox": [ + 194, + 191, + 229, + 201 + ], + "spans": [ + { + "bbox": [ + 194, + 191, + 229, + 201 + ], + "type": "text", + "content": "Round:0" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 194, + 207, + 230, + 216 + ], + "lines": [ + { + "bbox": [ + 194, + 207, + 230, + 216 + ], + "spans": [ + { + "bbox": [ + 194, + 207, + 230, + 216 + ], + "type": "text", + "content": "Rank:421" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 321, + 191, + 357, + 200 + ], + "lines": [ + { + "bbox": [ + 321, + 191, + 357, + 200 + ], + "spans": [ + { + "bbox": [ + 321, + 191, + 357, + 200 + ], + "type": "text", + "content": "Round:1" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 321, + 206, + 356, + 215 + ], + "lines": [ + { + "bbox": [ + 321, + 206, + 356, + 215 + ], + "spans": [ + { + "bbox": [ + 321, + 206, + 356, + 215 + ], + "type": "text", + "content": "Rank:74" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 450, + 191, + 485, + 200 + ], + "lines": [ + { + "bbox": [ + 450, + 191, + 485, + 200 + ], + "spans": [ + { + "bbox": [ + 450, + 191, + 485, + 200 + ], + "type": "text", + "content": "Round:4" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 450, + 207, + 480, + 215 + ], + "lines": [ + { + "bbox": [ + 450, + 207, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 450, + 207, + 480, + 215 + ], + "type": "text", + "content": "Rank:2" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 229, + 135, + 289 + ], + "blocks": [ + { + "bbox": [ + 56, + 229, + 135, + 289 + ], + "lines": [ + { + "bbox": [ + 56, + 229, + 135, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 135, + 289 + ], + "type": "image", + "image_path": "79a72c2d13ddf6673743adffd90279d3739a6e32d703fafd83820cef69ac46ef.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 237, + 262, + 257 + ], + "lines": [ + { + "bbox": [ + 168, + 237, + 262, + 257 + ], + "spans": [ + { + "bbox": [ + 168, + 237, + 262, + 257 + ], + "type": "text", + "content": "A video of someone talking about some girls" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 192, + 271, + 228, + 280 + ], + "lines": [ + { + "bbox": [ + 192, + 271, + 228, + 280 + ], + "spans": [ + { + "bbox": [ + 192, + 271, + 228, + 280 + ], + "type": "text", + "content": "Round:0" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 193, + 285, + 230, + 294 + ], + "lines": [ + { + "bbox": [ + 193, + 285, + 230, + 294 + ], + "spans": [ + { + "bbox": [ + 193, + 285, + 230, + 294 + ], + "type": "text", + "content": "Rank:353" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 295, + 236, + 388, + 256 + ], + "lines": [ + { + "bbox": [ + 295, + 236, + 388, + 256 + ], + "spans": [ + { + "bbox": [ + 295, + 236, + 388, + 256 + ], + "type": "text", + "content": "Q: what are the girls doing? A: looking at each other" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 270, + 356, + 279 + ], + "lines": [ + { + "bbox": [ + 320, + 270, + 356, + 279 + ], + "spans": [ + { + "bbox": [ + 320, + 270, + 356, + 279 + ], + "type": "text", + "content": "Round:1" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 320, + 285, + 351, + 294 + ], + "lines": [ + { + "bbox": [ + 320, + 285, + 351, + 294 + ], + "spans": [ + { + "bbox": [ + 320, + 285, + 351, + 294 + ], + "type": "text", + "content": "Rank:6" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 424, + 236, + 526, + 255 + ], + "lines": [ + { + "bbox": [ + 424, + 236, + 526, + 255 + ], + "spans": [ + { + "bbox": [ + 424, + 236, + 526, + 255 + ], + "type": "text", + "content": "Q: where is the silver earring? \nA: in left ear" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 449, + 270, + 485, + 279 + ], + "lines": [ + { + "bbox": [ + 449, + 270, + 485, + 279 + ], + "spans": [ + { + "bbox": [ + 449, + 270, + 485, + 279 + ], + "type": "text", + "content": "Round:4" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 449, + 285, + 479, + 294 + ], + "lines": [ + { + "bbox": [ + 449, + 285, + 479, + 294 + ], + "spans": [ + { + "bbox": [ + 449, + 285, + 479, + 294 + ], + "type": "text", + "content": "Rank:1" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 54, + 308, + 135, + 369 + ], + "blocks": [ + { + "bbox": [ + 54, + 308, + 135, + 369 + ], + "lines": [ + { + "bbox": [ + 54, + 308, + 135, + 369 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 135, + 369 + ], + "type": "image", + "image_path": "8f4c2d3c350deb4211464c39e1c8bb71851fa4e9c62df3e13670cbb8d90a0cdf.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 322, + 246, + 331 + ], + "lines": [ + { + "bbox": [ + 168, + 322, + 246, + 331 + ], + "spans": [ + { + "bbox": [ + 168, + 322, + 246, + 331 + ], + "type": "text", + "content": "a man is wearing a cap" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 192, + 350, + 228, + 359 + ], + "lines": [ + { + "bbox": [ + 192, + 350, + 228, + 359 + ], + "spans": [ + { + "bbox": [ + 192, + 350, + 228, + 359 + ], + "type": "text", + "content": "Round:0" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 192, + 365, + 225, + 374 + ], + "lines": [ + { + "bbox": [ + 192, + 365, + 225, + 374 + ], + "spans": [ + { + "bbox": [ + 192, + 365, + 225, + 374 + ], + "type": "text", + "content": "Rank : 39" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 294, + 317, + 383, + 337 + ], + "lines": [ + { + "bbox": [ + 294, + 317, + 383, + 337 + ], + "spans": [ + { + "bbox": [ + 294, + 317, + 383, + 337 + ], + "type": "text", + "content": "Q: what is the man doing? \nA: clapping" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 350, + 356, + 358 + ], + "lines": [ + { + "bbox": [ + 319, + 350, + 356, + 358 + ], + "spans": [ + { + "bbox": [ + 319, + 350, + 356, + 358 + ], + "type": "text", + "content": "Round:2" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 319, + 365, + 354, + 373 + ], + "lines": [ + { + "bbox": [ + 319, + 365, + 354, + 373 + ], + "spans": [ + { + "bbox": [ + 319, + 365, + 354, + 373 + ], + "type": "text", + "content": "Rank : 11" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 424, + 316, + 499, + 336 + ], + "lines": [ + { + "bbox": [ + 424, + 316, + 499, + 336 + ], + "spans": [ + { + "bbox": [ + 424, + 316, + 499, + 336 + ], + "type": "text", + "content": "Q: what is the object? A: baseball glove" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 448, + 350, + 484, + 358 + ], + "lines": [ + { + "bbox": [ + 448, + 350, + 484, + 358 + ], + "spans": [ + { + "bbox": [ + 448, + 350, + 484, + 358 + ], + "type": "text", + "content": "Round:3" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 448, + 365, + 479, + 374 + ], + "lines": [ + { + "bbox": [ + 448, + 365, + 479, + 374 + ], + "spans": [ + { + "bbox": [ + 448, + 365, + 479, + 374 + ], + "type": "text", + "content": "Rank:1" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 45, + 387, + 545, + 460 + ], + "lines": [ + { + "bbox": [ + 45, + 387, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 45, + 387, + 545, + 460 + ], + "type": "text", + "content": "Figure 4: Qualitative results of the dialogue generated by our Q & A System on MSR-VTT. The questions are generated by the Heuristic Question Generator. The initial queries of the target videos are shown in orange boxes. The questions and answers generated through each interaction are shown in blue boxes. Both the initial retrieval rank and the rank after the interaction are demonstrated. In each case, we observe a boost in performance. Note, however, that the answers are not always correct (e.g. in the fourth row \"A: baseball glove\" when the sport does not appear to be baseball). However, since the answer is \"nearly correct\", it still provides a useful cue to the retrieval model." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 480, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 480, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 287, + 529 + ], + "type": "text", + "content": "The architecture of the image-grounded text decoder is similar to the image-grounded text encoder but it replaces the bi-directional self-attention layers with causal self-attention layers. All videos are resized to " + }, + { + "bbox": [ + 46, + 480, + 287, + 529 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 46, + 480, + 287, + 529 + ], + "type": "text", + "content": " as input." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "text", + "content": "In our interactive system, we adopt the BLIP retrieval model, the VQA model, and the captioner. The BLIP retrieval model is composed of an image encoder, a text encoder, and an image-grounded text encoder. The BLIP VQA model is composed of an image encoder, an image-grounded text encoder, and an image-grounded text decoder. The BLIP captioner contains an image encoder and an image-grounded text decoder. To perform zero-shot transfer video retrieval, we uniformly sample 8 frames per video and concatenate the frame features into a single sequence. To generate questions, we adopt " + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{T0 + + }" + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "text", + "content": " [54], a powerful encoder-decoder language model trained on a collection of English-prompted datasets. This model attains strong zero-shot task generalization on English natural language prompts (outperforming GPT-3 on many tasks while being 16x smaller)." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "type": "text", + "content": "The Auto-text-vid question generation approach depends on the current top-k ranked retrieved videos. We experimented with different k (3, 5, 10, 15) and observed that when " + }, + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "type": "inline_equation", + "content": "\\mathrm{k} = 5" + }, + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "type": "text", + "content": ", we obtained the best performance. This may be because setting k to be too large produces too much noise in the generated captions while setting k to be too small does not provide an opportunity to extract all potentially discriminative information from related videos." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Baselines. In addition to the BLIP baseline, we also experimented with the interactive approach in [43] for fair comparisons. As public code is not available at the time of writing, we simulate their answer generation by first captioning the target video (by BLIP captioner) and then using the large language model " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "(\\mathrm{T0}++)" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " to generate the answer based on the caption and question. On the AVSD dataset, we also experimented with the approach in [44, 43] using 10 rounds of manually annotated human dialogue history for each video. However, simply concatenating all 10 rounds of ground truth dialog with the initial text query got a slightly" + } + ] + } + ], + "index": 44 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11096" + } + ] + } + ], + "index": 45 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 97 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 97 + ], + "type": "text", + "content": "worse performance than the baseline. Instead, we concatenated all the answers with the initial text query for retrieval." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 104, + 121, + 115 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 104, + 121, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 104, + 121, + 115 + ], + "type": "text", + "content": "4.3. Evaluation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 123, + 287, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 123, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 46, + 123, + 287, + 206 + ], + "type": "text", + "content": "Metrics. We adopt evaluation metrics that are widely used in the standard single-query retrieval setting and report recall@K (R@K, higher the better) and median rank (MdR, lower the better) in our experiments. R@K calculates the percentage of test data for which the ground-truth video is found in the retrieved K videos. MdR metric captures the median rank of the retrieved ground truth videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 207, + 288, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 288, + 327 + ], + "type": "text", + "content": "Time cost. We next assess the time cost of producing answers during the interaction. For this, we sample 50 videos from the MSR-VTT dataset. We aim to compare our answer generation approach using the VideoQA model with the approach in ViReD [43] (We call this answer generation approach " + }, + { + "bbox": [ + 46, + 207, + 288, + 327 + ], + "type": "inline_equation", + "content": "CAP + LM" + }, + { + "bbox": [ + 46, + 207, + 288, + 327 + ], + "type": "text", + "content": "). In addition, we also conduct a user study on answer generation. We invite 4 users to answer the auto-generated question (Auto-text) on the target videos and record the time they spent on each question. (Note that each user went through all 50 videos before the interaction.)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 338, + 99, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 99, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 99, + 350 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 50, + 366, + 287, + 517 + ], + "blocks": [ + { + "bbox": [ + 50, + 366, + 287, + 517 + ], + "lines": [ + { + "bbox": [ + 50, + 366, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 50, + 366, + 287, + 517 + ], + "type": "table", + "html": "
MethodITAR1 ↑R5 ↑R10 ↑MdR ↓
FiT [3]×31.961.172.63
CLIP4Clip [40]×41.569.479.32
CenterCLIP [71]×48.473.882.02
X-CLIP [42]×49.375.884.82
BLIP [35]×43.365.674.72
BLIP + ViReD [43]✓47.072.381.42
BLIP + Auto-text w/o AO✓47.672.881.82
BLIP + Auto-text-vid w/o AO✓49.072.980.02
BLIP + Auto-text✓62.284.590.71
BLIP + Auto-text-vid✓57.682.288.71
BLIP + Heuristic✓67.989.594.91
", + "image_path": "ef6ef9045b8d3bd302f4b7487e73ab8751cb40d5ad64a1abe1541a62b3f72427.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": "In this section, we discuss experimental results. On MSR-VTT (Tab. 1), without AO (asking objects), we observe that both the Auto-text and Auto-text-vid question generation approaches outperform the BLIP baseline. We observe that Auto-text-vid, which makes use of the top-ranked retrieved videos at each round, performs better than Auto-text. With AO, there is a significant performance boost for both approaches. For instance, we observe that the Auto-text Q&A System outperforms the baseline by " + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "inline_equation", + "content": "+18.9\\%" + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": " in recall@1." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 310, + 70, + 543, + 198 + ], + "blocks": [ + { + "bbox": [ + 46, + 532, + 287, + 581 + ], + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 581 + ], + "type": "text", + "content": "Table 1: Comparison to state-of-the-art methods for text-to-video retrieval on the 1k test split of MSRVTT dataset. ITA indicates whether it is an interactive retrieval system. Recall metrics are reported in percent." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 70, + 543, + 198 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 543, + 198 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 543, + 198 + ], + "type": "table", + "html": "
MethodITAR@1↑R@5↑R@10↑MdR↓
FiT [3]×33.764.776.33
CLIP4Clip [40]×46.276.184.62
X-CLIP [42]×50.480.6--
CenterCLIP [71]×50.680.388.41
BLIP [35]×44.273.780.92
BLIP + ViReD [43]✓51.078.187.21
BLIP + Auto-text✓68.791.295.81
BLIP + Auto-text-vid✓66.188.893.71
BLIP + Heuristic✓74.293.497.91
", + "image_path": "31a9bcb22e1711571c6e2b02cf0197f269b00583013f9e06e956fa021f7c334e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 272, + 545, + 403 + ], + "blocks": [ + { + "bbox": [ + 305, + 211, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 305, + 211, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 305, + 211, + 547, + 248 + ], + "type": "text", + "content": "Table 2: Comparison to state-of-the-art methods for text-to-video retrieval on the test split of the MSVD dataset. ITA indicates whether it is an interactive retrieval system." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 272, + 545, + 403 + ], + "lines": [ + { + "bbox": [ + 309, + 272, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 309, + 272, + 545, + 403 + ], + "type": "table", + "html": "
MethodITAR@1 ↑R@5 ↑R@10 ↑MdR ↓
FiT [3]×5.618.427.525
LSTM [44]✓4.213.522.1-
FiT + Human dialog [43]✓10.828.940.018
ViReD [43]✓24.949.060.86
BLIP [35]×27.350.061.25.5
BLIP + Human dialog [43]✓27.450.462.75
BLIP + ViReD [43]✓28.454.366.84
BLIP + Auto-text✓36.061.671.63
BLIP + Auto-text-vid✓36.561.071.13
BLIP + Heuristic✓39.569.177.82
", + "image_path": "a6220412ad5aeb9661021ce5b399705ca0c9f01316942ab5c90cd6a9281ef79c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 344, + 479, + 507, + 509 + ], + "blocks": [ + { + "bbox": [ + 305, + 419, + 547, + 455 + ], + "lines": [ + { + "bbox": [ + 305, + 419, + 547, + 455 + ], + "spans": [ + { + "bbox": [ + 305, + 419, + 547, + 455 + ], + "type": "text", + "content": "Table 3: Comparison to state-of-the-art methods for text-to-video retrieval on the test split of AVSD dataset. ITA indicates whether it is an interactive retrieval system." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 344, + 479, + 507, + 509 + ], + "lines": [ + { + "bbox": [ + 344, + 479, + 507, + 509 + ], + "spans": [ + { + "bbox": [ + 344, + 479, + 507, + 509 + ], + "type": "table", + "html": "
VideoQACAP+LMUser
Avg time (s)0.18.05.7
", + "image_path": "da41c4e4bcae48042343a1565c23d3366953bcdaf8b96191be3700b1ba9cf1c2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 521, + 545, + 582 + ], + "lines": [ + { + "bbox": [ + 305, + 521, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 305, + 521, + 545, + 582 + ], + "type": "text", + "content": "Table 4: Comparison of the average time cost of different answer and question generation approaches. We show the results of the answer generation approach of VQA (ours) and CAP+LM (simulation of ViReD [43] answer generator) and the user study." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 616, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 676 + ], + "type": "text", + "content": "Interestingly, we observe that Auto-text outperforms Autotext-vid when used in combination with AO. This suggests that simply asking the user directly for the objects that they are interested in (rather than inferring them from closely ranked videos) offers higher signals." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "We also observe that the Heuristic question generator yields dramatic gains: the recall@1 is improved by " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "24.6\\%" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " (from " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "43.3\\%" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "67.9\\%" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " (from " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "44.2\\%" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "74.2\\%" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ") and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11097" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 83, + 276, + 248 + ], + "blocks": [ + { + "bbox": [ + 97, + 73, + 257, + 83 + ], + "lines": [ + { + "bbox": [ + 97, + 73, + 257, + 83 + ], + "spans": [ + { + "bbox": [ + 97, + 73, + 257, + 83 + ], + "type": "text", + "content": "Retrieval results vs Number of queries" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 83, + 276, + 248 + ], + "lines": [ + { + "bbox": [ + 60, + 83, + 276, + 248 + ], + "spans": [ + { + "bbox": [ + 60, + 83, + 276, + 248 + ], + "type": "image", + "image_path": "3225b0ef270f835d04b1226e6990e1c085dab9a82b8bf007c2e79ea478cfde7e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 261, + 288, + 309 + ], + "lines": [ + { + "bbox": [ + 46, + 261, + 288, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 288, + 309 + ], + "type": "text", + "content": "Figure 5: Retrieval result vs Number of queries on MSRVTT. Performance of all different approaches improves as the number of rounds increases. Heuristic question generation approach achieves the most significant improvement." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "inline_equation", + "content": "12.2\\%" + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "text", + "content": " (from " + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "inline_equation", + "content": "27.3\\%" + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "inline_equation", + "content": "39.5\\%" + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "text", + "content": ") on MSR-VTT, MSVD and AVSD respectively. Using Auto-text question generator, the BLIP baseline recall@1 is improved by " + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "inline_equation", + "content": "18.9\\%" + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "text", + "content": " (from " + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "inline_equation", + "content": "43.3\\%" + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "inline_equation", + "content": "62.2\\%" + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "inline_equation", + "content": "24.5\\%" + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "text", + "content": " (from " + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "inline_equation", + "content": "44.2\\%" + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "inline_equation", + "content": "68.7\\%" + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "inline_equation", + "content": "8.7\\%" + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "text", + "content": " (from " + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "inline_equation", + "content": "27.3\\%" + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "inline_equation", + "content": "36.0\\%" + }, + { + "bbox": [ + 46, + 330, + 287, + 474 + ], + "type": "text", + "content": ") on MSR-VTT, MSVD, and AVSD respectively. This indicates that generated questions elicit useful information from the (simulated) user that substantially improves retrieval performance. Compared to prior interactive approaches (ViReD, Human dialog) on AVSD, our approach does not use dialog training data for finetuning but nevertheless outperforms existing work. Interestingly, the heuristic approach performs best over all three different datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "text", + "content": "As seen in Fig. 4, a few rounds of interaction are sufficient to locate the large video. We demonstrate the cases where the initial query is incomplete, hindering performance. However, using the question-answering system, the search results quickly improve and we obtain significantly better retrieval performance. The generated questions are of high quality, aiming to confirm discriminative information from the video. In addition, the answers generated by VideoQA model are reasonable in the majority of cases (though not always perfect). They successfully capture different aspects of the information in the video (such as action, location, and object). Another useful property of the generated answers is that they are short phrases, simulating concise user answers. Number of queries. Additional rounds of interaction lead to additional new queries. We observe in Fig. 5 that the retrieval performance increases as the number of queries increases. This effect is clearest for Heuristic question generation approach. The performance of both Auto-text and Auto-text-vid increases faster than their variants without AO. This further confirms that asking users about objects in the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 203 + ], + "type": "text", + "content": "videos is an effective approach to improve retrieval performance at each round of the interaction. We observe that without AO, recall@1 of Auto-text-vid increases faster than that of Auto-text while the opposite result is observed with AO. This agrees with our argument \"simply asking the user directly for the objects that they are interested in (rather than inferring them from closely ranked videos) offers higher signal\". For all the approaches, Performance stabilizes as the number of queries grows beyond 6. Note that the maximum number of queries for heuristic question generation is fixed to 6 and using 6 rounds works best." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 204, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 547, + 300 + ], + "type": "text", + "content": "Computational cost. As shown in Tab. 4, Our approach using the VideoQA model is significantly more efficient than " + }, + { + "bbox": [ + 304, + 204, + 547, + 300 + ], + "type": "inline_equation", + "content": "\\mathrm{CAP + LM}" + }, + { + "bbox": [ + 304, + 204, + 547, + 300 + ], + "type": "text", + "content": ". This is as expected because two steps (captioning and question-answering) are involved in " + }, + { + "bbox": [ + 304, + 204, + 547, + 300 + ], + "type": "inline_equation", + "content": "\\mathrm{CAP + LM}" + }, + { + "bbox": [ + 304, + 204, + 547, + 300 + ], + "type": "text", + "content": " approach while VideoQA directly incorporates the video and question information to generate answers. In our naive implementation, " + }, + { + "bbox": [ + 304, + 204, + 547, + 300 + ], + "type": "inline_equation", + "content": "\\mathrm{CAP + LM}" + }, + { + "bbox": [ + 304, + 204, + 547, + 300 + ], + "type": "text", + "content": " is even slower than the user response (though this could be optimised)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 312, + 375, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 312, + 375, + 324 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 375, + 324 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 332, + 547, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 332, + 547, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 547, + 452 + ], + "type": "text", + "content": "Limitations. (1) " + }, + { + "bbox": [ + 304, + 332, + 547, + 452 + ], + "type": "inline_equation", + "content": "\\mathrm{T0 + + }" + }, + { + "bbox": [ + 304, + 332, + 547, + 452 + ], + "type": "text", + "content": " [55] is good at question-answering tasks but might not be good at generating questions. Therefore fine-tuning the model for the question generation task could be helpful. (2) We adopt the BLIP model for both retrieval and VideoQA. In principle, there is the potential for \"leakage\" in which the VideoQA model encodes information that is specific to BLIP, but irrelevant to the video. In practice, different parameter checkpoints are used for different VideoQA and retrieval components, so we believe that such leakage is unlikely to occur in practice." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 453, + 547, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 547, + 512 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 547, + 512 + ], + "type": "text", + "content": "Broader Impact. Improved video retrieval has a wide range of societally useful applications, spanning domains such as education, entertainment, and security. However, as with many technologies, it is subject to dual use. For instance, it could be used to facilitate unlawful surveillance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 525, + 378, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 378, + 537 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 378, + 537 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 640 + ], + "type": "text", + "content": "In this work, we proposed several simple baselines for interactive video retrieval based on the generation of questions and answers. Through experiments on standard benchmarks, we demonstrated the effectiveness of generated questions in combination with a VideoQA model to simulate user responses through the interaction process. In future work, we hope to explore the use of question-answering systems to simulate users in other retrieval tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgements. SA would like to acknowledge Z. Novak and N. Novak in enabling his contribution, and support from an Isaac Newton grant and an EPSRC HPC grant. KL would like to thank Bill Byrne, Olga Russakovsky, Yu Wu, Vishaal Udandarao, Allison Chen, Xindi Wu, and Zhiwei Deng for helpful discussions and feedback." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11098" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Huda Alamri, Vincent Cartillier, Abhishek Das, Jue Wang, Anoop Cherian, Irfan Essa, Dhruv Batra, Tim K Marks, Chiori Hori, Peter Anderson, et al. Audio visual scene-aware dialog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7558-7567, 2019. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "type": "text", + "content": "[2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 288, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 266 + ], + "type": "text", + "content": "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 269, + 288, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 313 + ], + "type": "text", + "content": "[4] Hedi Ben-Younes, Rémi Cadene, Matthieu Cord, and Nicolas Thome. Mutan: Multimodal tucker fusion for visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2612-2620, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 315, + 288, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 315, + 288, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 315, + 288, + 367 + ], + "type": "text", + "content": "[5] Simion-Vlad Bogolin, Ioana Croitoru, Hailin Jin, Yang Liu, and Samuel Albanie. Cross modal retrieval with querybank normalisation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5194-5205, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 369, + 288, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 424 + ], + "type": "text", + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 426, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 288, + 491 + ], + "type": "text", + "content": "[7] Guanyu Cai, Jun Zhang, Xinyang Jiang, Yifei Gong, Lianghua He, Fufu Yu, Pai Peng, Xiaowei Guo, Feiyue Huang, and Xing Sun. Ask&confirm: active detail enriching for cross-modal retrieval with partial query. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1835-1844, 2021. 2, 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 492, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 492, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 492, + 288, + 557 + ], + "type": "text", + "content": "[8] David Chen and William Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 190-200, Portland, Oregon, USA, June 2011. Association for Computational Linguistics. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 559, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 559, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 559, + 288, + 601 + ], + "type": "text", + "content": "[9] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 603, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 288, + 647 + ], + "type": "text", + "content": "[10] Michael Cogswell, Jiasen Lu, Rishabh Jain, Stefan Lee, Devi Parikh, and Dhruv Batra. Dialog without dialog data: Learning visual dialog agents from vqa data. Advances in Neural Information Processing Systems, 33:19988-19999, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "text", + "content": "[11] Ioana Croitoru, Simion-Vlad Bogolin, Marius Leordeanu, Hailin Jin, Andrew Zisserman, Samuel Albanie, and Yang Liu. Teachtext: Crossmodal generalized distillation for text-video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11583-11593, 2021. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "text", + "content": "[12] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, José MF Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 326-335, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 183 + ], + "type": "text", + "content": "[13] Abhishek Das, Satwik Kottur, José MF Moura, Stefan Lee, and Dhruv Batra. Learning cooperative visual dialog agents with deep reinforcement learning. In Proceedings of the IEEE international conference on computer vision, pages 2951–2960, 2017. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 185, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 547, + 217 + ], + "type": "text", + "content": "[14] Georgiana Dinu, Angeliki Lazaridou, and Marco Baroni. Improving zero-shot learning by mitigating the hubness problem. arXiv preprint arXiv:1412.6568, 2014. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 220, + 547, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 547, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 547, + 284 + ], + "type": "text", + "content": "[15] Chenyou Fan, Xiaofan Zhang, Shu Zhang, Wensheng Wang, Chi Zhang, and Heng Huang. Heterogeneous memory enhanced multimodal attention model for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1999–2007, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 286, + 547, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 547, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 547, + 330 + ], + "type": "text", + "content": "[16] Akira Fukui, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, and Marcus Rohrbach. Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv preprint arXiv:1606.01847, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 332, + 547, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 547, + 374 + ], + "type": "text", + "content": "[17] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision, pages 214-229. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 376, + 547, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 547, + 431 + ], + "type": "text", + "content": "[18] Jiyang Gao, Runzhou Ge, Kan Chen, and Ram Nevatia. Motion-appearance co-memory networks for video question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6576–6585, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 434, + 547, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 547, + 475 + ], + "type": "text", + "content": "[19] Xiaoxiao Guo, Hui Wu, Yu Cheng, Steven Rennie, Gerald Tesauro, and Rogerio Feris. Dialog-based interactive image retrieval. Advances in neural information processing systems, 31, 2018. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 478, + 547, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 547, + 521 + ], + "type": "text", + "content": "[20] Hexiang Hu, Wei-Lun Chao, and Fei Sha. Learning answer embeddings for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5428-5436, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 523, + 547, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 547, + 577 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 547, + 577 + ], + "type": "text", + "content": "[21] Deng Huang, Peihao Chen, Runhao Zeng, Qing Du, Mingkui Tan, and Chuang Gan. Location-aware graph convolutional networks for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11021-11028, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 579, + 547, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 547, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 547, + 633 + ], + "type": "text", + "content": "[22] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2758–2766, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 635, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 547, + 689 + ], + "type": "text", + "content": "[23] Jianwen Jiang, Ziqiang Chen, Haojie Lin, Xibin Zhao, and Yue Gao. Divide and conquer: Question-guided spatiotemporal contextual attention for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11101-11108, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 692, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 547, + 714 + ], + "type": "text", + "content": "[24] Pin Jiang and Yahong Han. Reasoning with heterogeneous graph alignment for video question answering. In Proceedings" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11099" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "of the AAAI Conference on Artificial Intelligence, volume 34, pages 11109-11116, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "type": "text", + "content": "[25] Dotan Kaufman, Gil Levi, Tal Hassner, and Lior Wolf. Temporal tessellation: A unified approach for video analysis. In Proceedings of the IEEE International Conference on Computer Vision, pages 94-104, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 288, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 288, + 175 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 288, + 175 + ], + "type": "text", + "content": "[26] Ryan Kiros, Ruslan Salakhutdinov, and Richard S Zemel. Unifying visual-semantic embeddings with multimodal neural language models. arXiv preprint arXiv:1411.2539, 2014. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 288, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 219 + ], + "type": "text", + "content": "[27] Adriana Kovashka and Kristen Grauman. Attribute pivots for guiding relevance feedback in image search. In Proceedings of the IEEE International Conference on Computer Vision, pages 297-304, 2013. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 251 + ], + "type": "text", + "content": "[28] Adriana Kovashka and Kristen Grauman. Attributes for image retrieval. In Visual Attributes, pages 89-117. Springer, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 255, + 288, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 288, + 298 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 288, + 298 + ], + "type": "text", + "content": "[29] Adriana Kovashka, Devi Parikh, and Kristen Grauman. Whittlesearch: Image search with relative attribute feedback. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2973-2980. IEEE, 2012. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 300, + 288, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 300, + 288, + 343 + ], + "spans": [ + { + "bbox": [ + 48, + 300, + 288, + 343 + ], + "type": "text", + "content": "[30] Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Information maximizing visual question generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2008–2018, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 345, + 288, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 345, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 345, + 288, + 398 + ], + "type": "text", + "content": "[31] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9972–9981, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 400, + 288, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 400, + 288, + 444 + ], + "spans": [ + { + "bbox": [ + 48, + 400, + 288, + 444 + ], + "type": "text", + "content": "[32] Sang-Woo Lee, Yu-Jung Heo, and Byoung-Tak Zhang. Answerer in questioner's mind: Information theoretic approach to goal-oriented visual dialog. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 446, + 288, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 446, + 288, + 500 + ], + "spans": [ + { + "bbox": [ + 48, + 446, + 288, + 500 + ], + "type": "text", + "content": "[33] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 502, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 288, + 555 + ], + "type": "text", + "content": "[34] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "type": "text", + "content": "[35] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 3, 5, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 288, + 657 + ], + "type": "text", + "content": "[36] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[37] Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV, 2020. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "text", + "content": "[38] Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts. arXiv preprint arXiv:1907.13487, 2019. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "text", + "content": "[39] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Hierarchical question-image co-attention for visual question answering. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 546, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 546, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 546, + 206 + ], + "type": "text", + "content": "[40] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 209, + 547, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 547, + 251 + ], + "type": "text", + "content": "[41] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 254, + 546, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 546, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 546, + 307 + ], + "type": "text", + "content": "[42] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 638–647, 2022. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 309, + 546, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 546, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 546, + 342 + ], + "type": "text", + "content": "[43] Avinash Madasu, Junier Oliva, and Gedas Bertasius. Learning to retrieve videos by asking questions. arXiv preprint arXiv:2205.05739, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 343, + 546, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 546, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 546, + 387 + ], + "type": "text", + "content": "[44] Sho Maeoki, Kohei Uehara, and Tatsuya Harada. Interactive video retrieval with dialog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 952-953, 2020. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 388, + 546, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 546, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 546, + 453 + ], + "type": "text", + "content": "[45] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 455, + 546, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 546, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 546, + 488 + ], + "type": "text", + "content": "[46] Antoine Miech, Ivan Laptev, and Josef Sivic. Learning a text-video embedding from incomplete and heterogeneous data. arXiv preprint arXiv:1804.02516, 2018. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 490, + 547, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 547, + 554 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 547, + 554 + ], + "type": "text", + "content": "[47] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 556, + 546, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 546, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 546, + 599 + ], + "type": "text", + "content": "[48] Nasrin Mostafazadeh, Ishan Misra, Jacob Devlin, Margaret Mitchell, Xiaodong He, and Lucy Vanderwende. Generating natural questions about an image. arXiv preprint arXiv:1603.06059, 2016. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 601, + 546, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 546, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 546, + 645 + ], + "type": "text", + "content": "[49] Aishwarya Padmakumar and Raymond J Mooney. Dialog policy learning for joint clarification and active learning queries. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 13604-13612, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 647, + 546, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 546, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 546, + 678 + ], + "type": "text", + "content": "[50] Devi Parikh and Kristen Grauman. Relative attributes. In 2011 International Conference on Computer Vision, pages 503-510. IEEE, 2011. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "type": "text", + "content": "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11100" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 151 + ], + "type": "text", + "content": "[52] Milos Radovanovic, Alexandros Nanopoulos, and Mirjana Ivanovic. Hubs in space: Popular nearest neighbors in high-dimensional data. Journal of Machine Learning Research, 11(sept):2487-2531, 2010. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "type": "text", + "content": "[53] Yong Rui, Thomas S Huang, Michael Ortega, and Sharad Mehrotra. Relevance feedback: a power tool for interactive content-based image retrieval. IEEE Transactions on circuits and systems for video technology, 8(5):644-655, 1998. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 288, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 288, + 351 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 288, + 351 + ], + "type": "text", + "content": "[54] Victor Sanh, Albert Webson, Colin Raffel, Stephen Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey, M Saiful Bari, Canwen Xu, Urmish Thakker, Shanya Sharma Sharma, Eliza Szczechla, Taewoon Kim, Gunjan Chhablani, Nihal Nayak, Debajyoti Datta, Jonathan Chang, Mike Tian-Jian Jiang, Han Wang, Matteo Manica, Sheng Shen, Zheng Xin Yong, Harshit Pandey, Rachel Bawden, Thomas Wang, Trishala Neeraj, Jos Rozen, Abheesht Sharma, Andrea Santilli, Thibault Fevy, Jason Alan Fries, Ryan Teehan, Teven Le Scao, Stella Biderman, Leo Gao, Thomas Wolf, and Alexander M Rush. Multitask prompted training enables zero-shot task generalization. In International Conference on Learning Representations, 2022. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 353, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 288, + 407 + ], + "type": "text", + "content": "[55] Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. Multitask prompted training enables zero-shot task generalization. arXiv preprint arXiv:2110.08207, 2021. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 409, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 288, + 464 + ], + "type": "text", + "content": "[56] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7464-7473, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 465, + 288, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 288, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 288, + 518 + ], + "type": "text", + "content": "[57] Fuwen Tan, Paola Cascante-Bonilla, Xiaoxiao Guo, Hui Wu, Song Feng, and Vicente Ordonez. Drill-down: Interactive retrieval of complex scenes using natural language queries. Advances in neural information processing systems, 32, 2019. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 521, + 288, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 288, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 288, + 554 + ], + "type": "text", + "content": "[58] Atousa Torabi, Niket Tandon, and Leonid Sigal. Learning language-visual embedding for movie understanding with natural-language. arXiv preprint arXiv:1609.08124, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 555, + 288, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 288, + 588 + ], + "type": "text", + "content": "[59] Zeyu Wang, Yu Wu, Karthik Narasimhan, and Olga Russakovsky. Multi-query video retrieval. arXiv preprint arXiv:2201.03639, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 590, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 643 + ], + "type": "text", + "content": "[60] Hong Wu, Hanqing Lu, and Songde Ma. Willhunter: interactive image retrieval with multilevel relevance. In Proceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004., volume 2, pages 1009-1012. IEEE, 2004. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "text", + "content": "[61] Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and textual question answering. In International conference on machine learning, pages 2397-2406. PMLR, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[62] Huijuan Xu and Kate Saenko. Ask, attend and answer: Exploring question-guided spatial attention for visual question" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 543 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "answering. In European conference on computer vision, pages 451-466. Springer, 2016. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "text", + "content": "[63] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 141, + 547, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 195 + ], + "type": "text", + "content": "[64] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1686-1697, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 197, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 547, + 239 + ], + "type": "text", + "content": "[65] Xinru Yang, Haozhi Qi, Mingyang Li, and Alexander Hauptmann. From a glance to\" gotcha\": Interactive facial image retrieval with progressive relevance feedback. arXiv preprint arXiv:2007.15683, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 241, + 547, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 547, + 285 + ], + "type": "text", + "content": "[66] Zichao Yang, Xiaodong He, Jianfeng Gao, Li Deng, and Alex Smola. Stacked attention networks for image question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 21-29, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 286, + 547, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 547, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 547, + 330 + ], + "type": "text", + "content": "[67] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 471-487, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 331, + 547, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 547, + 386 + ], + "type": "text", + "content": "[68] Youngjae Yu, Hyungjin Ko, Jongwook Choi, and Gunhee Kim. End-to-end concept word detection for video captioning, retrieval, and question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3165-3173, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 387, + 547, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 547, + 453 + ], + "type": "text", + "content": "[69] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16375-16387, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 454, + 547, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 454, + 547, + 508 + ], + "spans": [ + { + "bbox": [ + 307, + 454, + 547, + 508 + ], + "type": "text", + "content": "[70] Zheng-Jun Zha, Jiawei Liu, Tianhao Yang, and Yongdong Zhang. Spatiotemporal-textual co-attention network for video question answering. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 15(2s):1-18, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 509, + 547, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 509, + 547, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 509, + 547, + 543 + ], + "type": "text", + "content": "[71] Shuai Zhao, Linchao Zhu, Xiaohan Wang, and Yi Yang. Centerclip: Token clustering for efficient text-video retrieval. arXiv preprint arXiv:2205.00823, 2022. 7" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11101" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss/081c8308-e453-453c-8d9f-1e2b9b04b642_content_list.json b/2023/Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss/081c8308-e453-453c-8d9f-1e2b9b04b642_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d75421ced20552252936f27da5518d01298d3bf5 --- /dev/null +++ b/2023/Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss/081c8308-e453-453c-8d9f-1e2b9b04b642_content_list.json @@ -0,0 +1,1818 @@ +[ + { + "type": "text", + "text": "Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss", + "text_level": 1, + "bbox": [ + 227, + 130, + 743, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SoonCheol Noh DongEon Jeong Jee-Hyong Lee*", + "bbox": [ + 253, + 203, + 722, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sungkyunkwan University", + "bbox": [ + 377, + 222, + 589, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Suwon, Korea", + "bbox": [ + 426, + 239, + 542, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yhnuhb27, wjdehddjs, john}@skku.edu", + "bbox": [ + 326, + 258, + 643, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning models need to detect out-of-distribution (OOD) data in the inference stage because they are trained to estimate the train distribution and infer the data sampled from the distribution. Many methods have been proposed, but they have some limitations, such as requiring additional data, input processing, or high computational cost. Moreover, most methods have hyperparameters to be set by users, which have a significant impact on the detection rate. We propose a simple and effective OOD detection method by combining the feature norm and the Mahalanobis distance obtained from classification models trained with the cosine-based softmax loss. Our method is practical because it does not use additional data for training, is about three times faster when inferencing than the methods using the input processing, and is easy to apply because it does not have any hyperparameters for OOD detection. We confirm that our method is superior to or at least comparable to state-of-the-art OOD detection methods through the experiments.", + "bbox": [ + 75, + 340, + 472, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 656, + 209, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning models are generally designed to target the closed world. Models are trained based on the assumption that the inputs of the inference stage will be sampled from the same distribution as the train distribution. This means that the model may not make correct predictions if inputs are sampled from other distributions. The model must be able to distinguish whether the input is from the train distribution or not. This problem has been established as the task of out-of-distribution (OOD) detection.", + "bbox": [ + 75, + 681, + 468, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Some researchers tried to detect OOD samples using confidence scores from classification models. Hendrycks and Gimpel simply used the confidence scores, but this approach was not very effective because of the overconfi", + "bbox": [ + 75, + 816, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dence of deep learning models [9]. Hendrycks, Mazeika, and Dietterich (2019) and Papadopoulos et al. tried to lower the confidence of OOD samples using auxiliary large OOD datasets [10, 21]. However, collecting a massive OOD dataset is highly expensive, and training classification models with such a dataset takes a long time. Liang, Li and Srikant (2018) simply adopted the input processing technique, which added a small perturbation to the input samples to increase the confidence of in-distribution (InD) samples, but they assumed OOD samples were given in advance, which is not practical [16].", + "bbox": [ + 500, + 310, + 892, + 476 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Some researchers defined measures to detect OOD samples based on input features learned by classification models. Lee et al. proposed a measure based on Mahalanobis distance, but they also assumed that OOD samples were given in advance as Liang et al. did [15, 16]. Sastry and Oore proposed a measure based on various correlations between features, but their approach was very time-consuming in evaluating the OOD measure [23]. Hsu et al. tried to obtain the OOD probability by modifying the network structure, which caused the degradation of classification performances [12].", + "bbox": [ + 496, + 477, + 892, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Additional popular technique is the input processing, which was adopted by most approaches [16, 15, 23]. The input processing is to generate $x'$ by adding a small perturbation to input data $x$ to make the boundary of in-distribution compact. It may be helpful to increase the detection accuracy under certain conditions. However, generating the perturbation takes almost twice of the inference time, so the total inference becomes three times longer. The OOD detection performance is very susceptible to the size of the perturbation, so it should be carefully determined.", + "bbox": [ + 496, + 643, + 893, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Some OOD detection approaches had various hyperparameters, including the perturbation size of input processing. The hyperparameters were usually tuned by using additional datasets to maximize the detection performance. Liang et al. [16] and Lee et al. [15] tuned them with given OOD samples, and Hsu et al. [12] and Sastry and Oore [23] used the validation samples. However, it is hard to find out", + "bbox": [ + 496, + 795, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 94, + 886, + 223, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16560", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hyperparameter values generally applicable to any distributions of OOD samples. When the values are found in a specific dataset, the model is easy to be biased to the dataset.", + "bbox": [ + 75, + 90, + 467, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper proposes a simple and effective approach that can detect OOD samples without hyperparameters, additional datasets, and input processing. We detect OOD samples by combining norm and Mahalanobis distances of features. We find that the models trained with the cosine-based softmax loss are more advantageous than the ones trained with the regular softmax loss for OOD detection. We mathematically conjecture that the feature norm from the models with the cosine-based softmax loss can be a measure for the probability of in-distribution, and confirm its effectiveness experimentally. We also develop a robust Mahalanobis distance-based measure for the model with the cosine-based softmax loss. We propose an OOD score function by combining the feature norm and the Mahalanobis distance without any hyperparmaeters.", + "bbox": [ + 75, + 136, + 467, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method do not need the input processing because it shows high accuracy in OOD detection without it. So, the additional computation for OOD detection by our method is very small. Feature norms and the Mahalanobis distances can be obtained with tiny additional calculations from classification models. Since the cosine-based softmax loss is generally applicable to any deep models, our OOD detection approach can be easily extended to various structures without a classification accuracy loss. Our method is also practical because it does not use any additional datasets. We confirm the OOD detection performance on various image datasets through experiments. Our method is compared with others using additional datasets, performing the input processing, or tuning hyperparameters to maximize performance. The experiment shows that our approach is superior to or at least comparable to state-of-the-art OOD detection methods.", + "bbox": [ + 75, + 362, + 467, + 617 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 632, + 218, + 647 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous OOD detection approaches were based on various models, such as classification models [9, 16, 15, 10, 21, 17, 12, 23], generative models [22, 19, 25, 4, 26, 3] and Bayesian models [6, 2, 18].", + "bbox": [ + 75, + 657, + 467, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "OOD detection is usually required as a pre-processing of the classification task, which is one of the crucial tasks of deep learning. Due to its effectiveness, many researchers tried simultaneously performing OOD detection and classification in one model. In this section, we cover OOD detection-related approaches based on classification models.", + "bbox": [ + 75, + 717, + 467, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Training with Large OOD Datasets", + "text_level": 1, + "bbox": [ + 76, + 832, + 382, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hendrycks et al. proposed a method to lower the maximum softmax probability (MSP) for OOD samples from classification models [10]. They improved OOD detection", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "performance by adding a loss term which was based on large OOD datasets. The added loss term was designed so that softmax values for OOD samples be closed to a uniform distribution. Papadopoulos et al. further improved OOD detection performance by adding another loss term [21]. The loss term prevented classification models from overconfidence. Liu et al. proposed an energy-based measure for models trained with large OOD datasets [17].", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These methods had the disadvantage of collecting extensive additional datasets to train classification models. The collected OOD dataset was used, assuming it represented all unknown OODs. However, it is virtually impossible to collect data that can represent all OOD samples in real-world situations. In addition, training a model with a large amount of OOD data takes much longer than regular training.", + "bbox": [ + 496, + 215, + 890, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Tuning Hyperparameters Using Known OOD Samples", + "text_level": 1, + "bbox": [ + 498, + 342, + 890, + 373 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some OOD detection approaches had hyperparameters that need to be tuned. Liang et al. [16] and Lee et al. [15] used specific OOD samples to tune hyperparameters under the assumption that distribution of OOD samples were given in advance. Liang et al. [16] tuned the temperature scaling factor and a perturbation size of input processing. Lee et al. [15] tuned weights for the feature ensemble and a perturbation size of input processing. However, these approaches are not feasible because it is difficult to know from which distribution OOD samples will occur in a real-world situation.", + "bbox": [ + 496, + 383, + 890, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Tuning Hyperparameters Using In-distribution Samples", + "text_level": 1, + "bbox": [ + 498, + 571, + 890, + 602 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, some OOD detection approaches without using OOD samples were proposed because it was unrealistic to use OOD samples. Hsu et al. modified the classifying layer to obtain the probability that an input was in-distribution [12]. However, the modified classifying layer made the classification accuracy unstable. They also proposed a strategy that the perturbation size of the input processing was set without OOD samples, but it was not valid to all datasets. Sastry and Oore [23] proposed measures for OOD detection based on the Gram matrices of features. They stored the maximum and minimum values of the elements of the Gram matrix for each class, layer, power, and index of the matrix. They detected OOD samples using the stored values. However, it required heavy computation and a longer time than other approaches. Those approaches used additional InD validation dataset for tuning hyperparameters: Hsu et al. [12] tuned the perturbation size of input processing, and Sastry and Oore [23] tuned normalization factors for each layer.", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16561", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.4. Using Input Processing", + "text_level": 1, + "bbox": [ + 76, + 90, + 292, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The input processing is an operation that generates $x'$ by adding a small perturbation to an input $x$ . Many previous approaches applied the input processing because this process effectively improved OOD detection performance. However, the perturbation size, epsilon, needed to be tuned appropriately. The epsilon has a significant impact on performance, so it should be chosen carefully.", + "bbox": [ + 75, + 113, + 467, + 219 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Liang et al. [16] and Lee et al. [15] chose the epsilon values using given specific OOD samples. The chosen epsilon may not be valid to other OOD samples because it was tuned with a distribution of given OOD samples. Hsu et al. [12] proposed a method to set the epsilon value without OOD samples. They tuned the epsilon using the InD validation dataset, but their method did not generally work for all cases. The detection performance decreased in some datasets.", + "bbox": [ + 75, + 219, + 467, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another disadvantage is that the input processing requires a lot of computation. We need one forward loop and one backward loop to add a perturbation to an input. We need one more forward loop to determine whether the perturbed input is OOD. Since the time taken for the forward loop and the backward loop is almost the same, the total inference time increases by about three times compared to the inference time without the input processing.", + "bbox": [ + 75, + 354, + 467, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.5. Using Self-supervised Learning", + "text_level": 1, + "bbox": [ + 76, + 483, + 354, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, a self-supervised learning technique that can learn a deep learning model without a label is attracting attention. In many studies in the field of OOD detection, self-supervised learning techniques are being tried to improve performance or reduce dependence on data labels. Some approaches build new data by applying transforms such as rotation to the image, and train models by giving the transform applied to the image as a label [7, 11, 27, 1]. Other approaches use contrastive learning techniques to train the model [27, 24]. However, in most cases, the OOD detection performance is not as high as when learning using labels. In addition, since the model is trained only for the purpose of OOD detection or representation learning, an additional method is needed to perform classification.", + "bbox": [ + 75, + 506, + 467, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 729, + 250, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A deep learning-based classification model can be decomposed into a feature extractor $f(\\cdot)$ and a classifier $g(\\cdot)$ which is usually a fully connected layer. When training with the standard softmax loss, a logit is defined as:", + "bbox": [ + 75, + 755, + 467, + 816 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {l o g i t} (x, c) = f (x) \\cdot w _ {c} + b _ {c}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 830, + 467, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and with the cosine-based softmax loss, it is defined as:", + "bbox": [ + 76, + 849, + 442, + 864 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {l o g i t} (x, c) = \\frac {f (x) \\cdot w _ {c}}{| f (x) | \\cdot | w _ {c} |}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 869, + 467, + 905 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/825b49fc37e084e1c1ab0f01121b7cec8973b23fd1c07afb0f86c1efd4c238d6.jpg", + "image_caption": [ + "(a) Standard softmax loss" + ], + "image_footnote": [], + "bbox": [ + 501, + 99, + 684, + 191 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6d7b815eb1d20eb07273025a433b8073fe90b6e0ea7619058c6bed45b3416c3c.jpg", + "image_caption": [ + "(b) Cosine-based softmax loss", + "Figure 1: The feature norm distributions of InD (red) and OOD (blue) data. (a) is from DenseNet3 trained with the standard softmax loss and (b) is from DenseNet3 trained with the cosine-based softmax loss." + ], + "image_footnote": [], + "bbox": [ + 696, + 99, + 870, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $w_{c}$ is the connection weights of $g(\\cdot)$ for a class $c$ . The cosine-based softmax loss is widely used in face recognition field [28, 5, 29].", + "bbox": [ + 496, + 323, + 890, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce two useful measures for OOD detection that can be simply extracted from models trained with cosine-based softmax loss: a feature norm-based measure and a Mahalanobis distance-based measure. We then propose an OOD score by combining these two measures. Let us describe each measure and the score function in detail.", + "bbox": [ + 496, + 368, + 890, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Cosine-based Softmax Loss and Feature Norm", + "text_level": 1, + "bbox": [ + 500, + 468, + 890, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When we train a classification model with the closed-world assumption, the logit value for a class $c$ is considered as being proportional to $P(y = c|x)$ . However, in the open world, the logit value should be interpreted as being proportional to $P(y = c|x, D_{in})$ where $D_{in}$ is the event that $x$ is in the train distribution. The probability can be decomposed using $P(y = c, D_{in}|x)$ and $P(D_{in}|x)$ as follows:", + "bbox": [ + 496, + 491, + 890, + 597 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {l o g i t} (x, c) \\propto P (y = c | x, D _ {i n}) = \\frac {P (y = c , D _ {i n} | x)}{P \\left(D _ {i n} | x\\right)}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 607, + 890, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "If we use the cosine-based softmax loss, Equations 2 and 3 can be combined as follows:", + "bbox": [ + 500, + 648, + 890, + 678 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {f (x) \\cdot w _ {c}}{| f (x) | \\cdot | w _ {c} |} \\propto \\frac {P (y = c , D _ {i n} | x)}{P \\left(D _ {i n} | x\\right)}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 686, + 890, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The numerator of the right-hand side includes a class-related term, $c$ , and the denominator, $P(D_{in}|x)$ , does not. So, we reformulate the left-hand side so that only the numerator has class-related terms as follows:", + "bbox": [ + 496, + 729, + 890, + 787 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {f (x) \\cdot w _ {c} / | w _ {c} |}{| f (x) |} \\propto \\frac {P (y = c , D _ {i n} | x)}{P \\left(D _ {i n} | x\\right)}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 796, + 890, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since $w_{c}$ is the connection weight to $g(\\cdot)$ , we can say it is class-related, and since $f(x)$ is the feature representation of $x$ , we can say it is classifier-independent. Based on this, we rigorously infer that both denominators can be proportional", + "bbox": [ + 496, + 840, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16562", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c0b5afe3ce518fab97a13dc7f710470d428577900820d4dcd6cdb05ec1f732a0.jpg", + "image_caption": [ + "(a) Distribution of $|f(x)|$" + ], + "image_footnote": [], + "bbox": [ + 80, + 99, + 259, + 202 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ac6384f326f9c55319542164ee2e17d973fc31da3561b548d0c6df0194472bb5.jpg", + "image_caption": [ + "(b) Mean of $|f(x)|$" + ], + "image_footnote": [], + "bbox": [ + 272, + 99, + 449, + 202 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to each, i.e., $|f(x)| \\propto P(D_{in}|x)$ , because the two terms in Equation 5 are proportional, and their denominators and numerators share class-related characteristics.", + "bbox": [ + 75, + 349, + 468, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To confirm this, we perform some OOD detection experiments. We try to detect SVHN using $|f(x)|$ from DenseNet3 model trained for CIFAR-10 with the cosine-based softmax loss. We compare this to the model using the standard softmax loss as shown in Figure 1. The feature norm distributions of OOD and InD by the standard loss model overlap much, but those of the cosine loss model have a very small overlap between SVHN and CIFAR-10, which says that the cosine loss model can be better to discriminate ODD than the standard model.", + "bbox": [ + 75, + 397, + 468, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To verify that the length of feature norm decrease as the samples get closer to OOD, we create mixup samples by mixing SVHN and CIFAR-10 images as shown in Equation 6 and observed the feature norm.", + "bbox": [ + 75, + 549, + 468, + 608 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {T e s t I m g} = \\lambda \\cdot \\operatorname {I n D} + (1 - \\lambda) \\cdot \\operatorname {O O D} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 623, + 468, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the mixup factor, $\\lambda$ decreases, that is, as the samples get closer to OOD, the feature norm also decreases as shown in Figure 2. The feature norm of an image from the cosine loss model tends to increase as the image gets closer to InD samples and decreases closer to OOD samples, which supports our conclusion that $|f(x)|$ from a cosine loss model is proportional to $P(D_{in}|x)$ . Also, it supports that the feature norm can be a measure for OOD detection. Additional observations, including CIFAR-100 and SVHN, are presented in the appendix.", + "bbox": [ + 75, + 652, + 468, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Mahalanobis Distance in Models Trained with Cosine-based Softmax Loss", + "text_level": 1, + "bbox": [ + 76, + 816, + 468, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We may assume that features of InD samples obtained from internal layers of classification models follow the Gaussian distribution. We adopt an OOD score defined by", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2913a3b03bd1ba529efe9fadfd27919720bf4feb1ab15c505f627377bd3b4b7d.jpg", + "image_caption": [ + "(a) 1st Block (SVHN vs. TINc)" + ], + "image_footnote": [], + "bbox": [ + 501, + 98, + 686, + 185 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1540a4e03e4294be16f53776a9a35fa65b03108b1140bc6176a75ff781ed0d41.jpg", + "image_caption": [ + "(b) 1st Block (SVHN vs. iSUN)" + ], + "image_footnote": [], + "bbox": [ + 696, + 98, + 879, + 185 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d1ba4001561006a0b2c01e633b46a4f109f7fb8f790596caa1d3f7efaaffafd8.jpg", + "image_caption": [ + "Figure 2: The distribution of feature norm according to the mixup factor (a), and the mean value of feature norm according to the mixup factor. It can be observed that as OOD samples are more mixed into a sample, the feature norm tends to decrease." + ], + "image_footnote": [], + "bbox": [ + 501, + 215, + 684, + 305 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/41c6e10e01943d5d85d3d56f8dc6107449ca80a3b495e64431d9bbb25d32b8da.jpg", + "image_caption": [ + "(c) 2nd Block (SVHN vs. TINc)", + "(d) 2nd Block (SVHN vs. iSUN)" + ], + "image_footnote": [], + "bbox": [ + 696, + 215, + 877, + 305 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a78bc1dab61379cc9f038bd0b47dc810ac1d233d091151ae70a4c6921a28f07f.jpg", + "image_caption": [ + "(e) 3rd Block (SVHN vs. TINc)" + ], + "image_footnote": [], + "bbox": [ + 501, + 335, + 684, + 425 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ea705c855111256c1c3523f9ed65914eef230856b758dc977ce8244b7ca4b046.jpg", + "image_caption": [ + "(f) 3rd Block (SVHN vs. iSUN)", + "Figure 3: Distributions of $MD$ scores from each block of DenseNet3 trained with SVHN. The OODs datasets (blue) is TinyImageNet-crop (left) and iSUN (right)." + ], + "image_footnote": [], + "bbox": [ + 696, + 335, + 877, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lee et al. [15] based on the Mahalanobis distance:", + "bbox": [ + 500, + 542, + 831, + 556 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} M D _ {B} (x) \\\\ = - \\max _ {c} - \\left(f _ {B} (x) - \\mu_ {c, B}\\right) ^ {\\top} \\Sigma_ {B} ^ {- 1} \\left(f _ {B} (x) - \\mu_ {c, B}\\right), \\quad (7) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 571, + 890, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu_{c,B}$ and $\\Sigma_B$ are the mean and the covariance of features from a block $B$ . This score is very discriminative for OOD detection as shown in Figure 3. Since the distributions of $MD$ scores of InD and OOD samples have small overlaps, we can detect OOD samples using the scores. However, $MD$ scores from the penultimate layer overlap if the model is trained with cosine-based softmax loss, as shown in Figure 2(e)-(f).", + "bbox": [ + 496, + 622, + 890, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The reason why $MD$ scores from penultimate layer is not discriminative is that the feature from the penultimate layer do not follow a Gaussian distribution. In the model trained with cosine-based softmax loss, the softmax probability with $\\sigma_c$ , the softmax value for a class $c$ , is calculated as follows:", + "bbox": [ + 496, + 742, + 890, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {c} (g (f (x))) = \\frac {\\exp \\left(w _ {c} \\cdot f (x) / \\left(\\left| w _ {c} \\right| \\cdot | f (x) |\\right)\\right)}{\\sum_ {c ^ {\\prime}} \\exp \\left(w _ {c ^ {\\prime}} \\cdot f (x) / \\left(\\left| w _ {c ^ {\\prime}} \\right| \\cdot | f (x) |\\right)\\right)}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 853, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16563", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/d8cc9fb07b80701d78a4657d634bc82c220a370270d3d88bbcadd9bf97a05b1e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ApproachAuxiliary Dataset# of Hyperp.Input Processing
OE [10]Large OOD dataset0×
ODIN [16]OOD samples2✓
MAHA [15]OOD samples4✓
G-ODIN [12]InD samples1✓
GRAM [23]InD samples100×
COD (Proposed)None0×
", + "bbox": [ + 251, + 88, + 718, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: OOD detection approach comparison. The number of hyperparameters is counted based on DenseNet3 with 100 layers.", + "bbox": [ + 75, + 207, + 893, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The classification model is trained so that the softmax probability reflects $P(y = c|f(x))$ . If the feature distribution from the penultimate layer is assumed to be the multivariate Gaussian distribution for each class, and the class prior is $P(y = c) = \\frac{\\beta_c}{\\sum_{c'} \\beta_{c'}}$ where $\\beta_c$ is unnormalized prior for class $c$ , the probability of class $c$ with $f(x)$ is as follows:", + "bbox": [ + 76, + 265, + 472, + 361 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P (y = c \\mid f (x)) \\\\ = \\frac {P (y = c) P (f (x) | y = c)}{\\sum_ {c ^ {\\prime}} P (y = c ^ {\\prime}) P (f (x) | y = c ^ {\\prime})} \\\\ = \\frac {\\exp \\left(\\mu_ {c} ^ {T} \\Sigma^ {- 1} f (x) - \\frac {1}{2} \\mu_ {c} ^ {T} \\Sigma^ {- 1} \\mu_ {c} + \\log \\beta_ {c}\\right)}{\\sum_ {c ^ {\\prime}} \\exp \\left(\\mu_ {c ^ {\\prime}} ^ {T} \\Sigma^ {- 1} f (x) - \\frac {1}{2} \\mu_ {c ^ {\\prime}} ^ {T} \\Sigma^ {- 1} \\mu_ {c ^ {\\prime}} + \\log \\beta_ {c ^ {\\prime}}\\right)}. \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 390, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "That is, if the feature from the penultimate layer follows a Gaussian distribution, $P(y = c|f(x))$ is based on linear expressions. Since a nonlinear expression cannot be transformed in a linear expression, the softmax probability from the cosine loss model (Equation 8) cannot be converted to a similar form of $P(y = c|f(x))$ (Equation 9), which means that the features from the penultimate layer is not expected to follow a Gaussian distribution.", + "bbox": [ + 75, + 510, + 468, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If Equation 8 is rewritten using a normalized feature $f^{\\prime}(x) = f(x) / |f(x)|$ , it is as follows:", + "bbox": [ + 76, + 632, + 468, + 664 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {c} (g (f (x))) = \\frac {\\exp \\left(w _ {c} ^ {\\top} f ^ {\\prime} (x) / \\left(\\left| w _ {c} ^ {\\top} \\right|\\right)\\right)}{\\sum_ {c ^ {\\prime}} \\exp \\left(w _ {c ^ {\\prime}} ^ {\\top} f ^ {\\prime} (x) / \\left(\\left| w _ {c ^ {\\prime}} ^ {\\top} \\right|\\right)\\right)}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 686, + 468, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Equation 10 is based on linear expressions of $f'(x)$ , from which we conjecture that $f'(x)$ may follow a Gaussian distribution. Since $MD$ scores are useful under the assumption that features follow a Gaussian distribution, we redefine the feature for $MD$ from the penultimate layer as $f'(x) = f(x) / |f(x)|$ instead of $f(x)$ . We re-depict the distribution of $MD$ in Figures 2-(e) and (f) using $f'(x)$ as shown in Figure 4. The overlap of the distributions is significantly reduced compared to $MD$ based on $f(x)$ , which means that $MD$ based on $f'(x)$ may be better than $f(x)$ for OOD detection.", + "bbox": [ + 75, + 734, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/069b255e7d1cf33cf84628fa72a425a26ef0f52c098a564b504635a0b50eb375.jpg", + "image_caption": [ + "(a) 3rd Block (SVHN vs TINc)" + ], + "image_footnote": [], + "bbox": [ + 501, + 272, + 686, + 359 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d6c43b08bec461609b8de4b457afbce9907805edb48808435420b343c1967626.jpg", + "image_caption": [ + "(b) 3rd Block (SVHN vs iSUN)", + "Figure 4: Distributions of $MD$ scores based on $f^{\\prime}(x)$ . The configuration is the same as Figure 3." + ], + "image_footnote": [], + "bbox": [ + 696, + 271, + 879, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate $MD$ scores for the penultimate layer with the redefined feature, $f'(x) = f(x) / |f(x)|$ and for other layers with $f_B(x)$ . Finally, we combine $MD$ scores from each block as Equation 11 by product:", + "bbox": [ + 498, + 460, + 892, + 521 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM D (x) = \\prod_ {B} \\left(M D _ {B} (x)\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 532, + 890, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Proposed OOD Score Function", + "text_level": 1, + "bbox": [ + 500, + 571, + 774, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We define a cosine-based OOD score function, $COD(x)$ , with the feature norm and the Mahalanobis distance-based measure, $MD$ . Since the feature norm tends to appear large for InD samples and $MD$ tends to be large for OOD samples, we define $COD(x)$ by dividing the feature norm by $MD$ as follows:", + "bbox": [ + 498, + 595, + 890, + 686 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC O D (x) = \\frac {| f (x) |}{M D (x)}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 695, + 890, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$COD(x)$ does not have any hyperparameter to be tuned. Also, no additional dataset is required to use $COD(x)$ , and it shows a high discriminative performance without the input processing. $COD(x)$ can be calculated with a very small computation in a forward loop for classification. We compare the previous approaches in the viewpoint of auxiliary datasets, the number of hyperparameters and the input processing in Table 1. Most of the previous methods used auxiliary datasets, and had several to one hundred hyperparameters, and needed the input processing to increase the performance, while our method does not need any of them.", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16564", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 90, + 200, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section shows the effectiveness of our approach through experiments with various image datasets.", + "bbox": [ + 76, + 114, + 467, + 146 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment setting", + "text_level": 1, + "bbox": [ + 76, + 155, + 261, + 171 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets: We use CIFAR-10/100 [14], and SVHN [20] images for InD datasets. We use the training dataset for classification and use the test datasets to evaluate OOD detection. We use TinyImageNet-crop (TINc), TinyImageNet-resize (TINr), LSUN-crop (LSUNc), LSUN-resize (LSUNr), and iSUN as OOD datasets. These datasets have been used in many studies related to OOD detection. Additionally, when CIFAR-10/100 are in-distribution, SVHN test dataset is also used as an OOD dataset, and when SVHN image is in-distribution, CIFAR-10/100 test datasets are also used as OOD datasets.", + "bbox": [ + 76, + 179, + 467, + 344 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines: We compare our approach to previous OOD detection methods: ODIN [16], MAHA [15], G-ODIN [12], and GRAM [23].", + "bbox": [ + 76, + 351, + 467, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Networks and Training Details: We train ResNet34 [8] and DenseNet3 [13] models with the same setting as the baseline approaches. We train ResNet34 for 200 epochs with a batch size of 128 and a weight decay of 0.0005. We train DenseNet3 with 100 layers for 300 epochs with a batch size of 64 and a weight decay of 0.0001. In both training, the optimizer is SGD with a momentum of 0.9, and the learning rate starts with 0.1 and decays by a factor 0.1 at $50\\%$ and $75\\%$ of the training progress.", + "bbox": [ + 76, + 402, + 467, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When we train models with the cosine-based softmax loss, logit values are between -1 and 1. For expanding the range of logits, we adopt the temperature scaling with a factor of $T$ as in Equation 13. The factor of $T$ needs to be set to 0.1 for SVHN, CIFAR-10 and 0.05 for CIFAR-100 so that the classification performance is maximized. The logit scaled by $T$ is defined as:", + "bbox": [ + 76, + 540, + 467, + 643 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {l o g i t} _ {T} (x) = \\operatorname {l o g i t} (x) / T. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 661, + 467, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics: We evaluate the detection performance using the two metrics that are most widely used for OOD detection. The first metric is the area under the receiver operating characteristic curve (AUROC). It is used to evaluate the average performance considering various thresholds. The second is TNR at TPR $p\\%$ . We evaluate with $p = 95$ as the metric used in the baseline approaches.", + "bbox": [ + 76, + 685, + 467, + 791 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 76, + 801, + 169, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 shows an overall comparison with DenseNet3. Our approach achieves superior performance in all the datasets. Compared to ODIN, our approach has higher performance in all the cases. Compared to G-ODIN, our approach has higher performance except one cases. Com", + "bbox": [ + 76, + 824, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pared to MAHA, our approach has higher performance except some cases.", + "bbox": [ + 498, + 90, + 888, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In some cases, GRAM shows a higher performance than the proposed method, but GRAM is very unstable depending on OOD datasets. For example, GRAM shows a particularly poor performance on LSUNc. It shows a TNR@95 score of 88.4 for CIFAR-10 and 65.6 for CIFAR-100. Our approach is more even performances across all OOD datasets, and achieves the highest average performances. Also, GRAM requires significantly heavy computation. When comparing the wall clock time in the same environment (GTX 2080 Ti GPU with Xeon Silver 4021 CPU * 2), the GRAM takes 0.35 seconds to calculate OOD score for one image, but our approach takes only about 0.018 seconds. Our approach is 19 times faster than GRAM because GRAM evaluates a lot of measures in the forward loop while our approach needs very small additional computation.", + "bbox": [ + 498, + 122, + 890, + 363 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The head of Table 2 shows what each approach needs to detect OOD samples. ODIN and MAHA use OOD samples and InD validation samples, and perform the input processing. G-ODIN does not use OOD samples but uses InD validation samples and performs the input processing. GRAM uses InD validation samples to tune hyperparameters. Our approach needs nothing except a training dataset. Our approach is the fastest method because the input processing is not required. Our method is superior to or at least comparable to state-of-the-art OOD detection methods even though it does not require any additional datasets. Our model also shows better performance with ResNet34. The results are presented in the appendix.", + "bbox": [ + 498, + 364, + 890, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 500, + 577, + 612, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Layer Selection for $MD$", + "text_level": 1, + "bbox": [ + 500, + 603, + 718, + 618 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our approach, $COD$ , uses $MD$ values from all the blocks of the model. However, it can be more effective if we can choose some layers which are helpful for OOD detection. Useful layers for OOD detection may depend on the model structure, the training dataset, and the OOD dataset. However, OOD samples are not available in advance.", + "bbox": [ + 498, + 627, + 890, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We will discuss how to select useful layers with synthetic OOD samples. We may generate synthetic OOD samples from the training dataset. For example, we may use some training samples rotated by 90 degrees as synthetic OOD samples. Then, we evaluate AUROC scores for each layer, and choose layers of which AUROC scores are greater than a threshold such as $60\\%$ of AUROC. In the case of CIFAR10, CIFAR100, and SVHN, it is confirmed that the AUROC score of each $MD$ was over $60\\%$ . Therefore, it can be a reasonable threshold to choose layers.", + "bbox": [ + 498, + 718, + 890, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We apply the layer selection to a more realistic dataset, ImageNet-30. In this case, $MD$ from the last layer, $MD$", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16565", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/89ca4be6844ecdec980500f5d336193eca03f85557c6f9c16c692b32a5776d0c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
OODAUROCTNR@95
Use a training datasetUse a training dataset
Use InD validation samples(ours)Use InD validation samples(ours)
Perform the input processingGRAMPerform the input processingGRAMCOD
Use OOD samplesG-ODINUse OOD samplesG-ODIN
ODINMAHAODINMAHA
InD: CIFAR-10
SVHN95.598.198.899.199.586.290.894.096.198.0
TINc97.695.398.799.399.487.084.293.496.797.9
TINr98.598.899.199.799.592.495.095.898.897.8
LSUNc93.680.298.397.599.970.648.291.588.499.6
LSUNr99.299.399.499.999.696.297.297.699.598.9
iSUN98.798.999.499.899.693.295.397.599.098.7
MEAN97.295.199.099.299.687.685.195.096.498.5
InD: CIFAR-100
SVHN93.897.295.997.397.070.682.577.089.382.4
TINc88.388.897.697.797.951.060.187.889.089.3
TINr85.297.498.699.098.642.686.693.395.793.1
LSUNc91.481.795.391.498.957.842.175.065.595.1
LSUNr85.598.098.799.398.641.291.493.897.293.3
iSUN84.597.498.499.098.537.487.092.595.992.5
MEAN88.193.497.497.398.350.175.086.688.891.0
InD: SVHN
CIFAR-1091.498.9-95.599.071.796.8-80.496.1
CIFAR-100----98.9----95.8
TINc----99.8----99.7
TINr95.199.9-99.799.884.199.9-99.199.5
LSUNc----99.8----99.4
LSUNr94.599.9-99.899.881.199.9-99.599.4
iSUN94.799.9-99.899.882.299.9-99.499.5
MEAN----99.6----98.5
", + "bbox": [ + 84, + 87, + 885, + 592 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/34258861fd43ae7f3c45230af57463d10280aee0f1356fdda421a0b429ee0643.jpg", + "table_caption": [ + "Table 2: OOD detection performance (%) with a DenseNet3 classification model. Our performance is averaged over 5 runs. ODIN, MAHA and GRAM are cited from the work by Sastry and Oore [23] and G-ODIN is cited from the work by Hsu et al. [12]. In the case of SVHN, all baselines did not present their performances. Dash symbols stand for not-available. The mean rows are the averages of the performances over all OOD datasets. For each case, the best results are in bold." + ], + "table_footnote": [], + "table_body": "
ResNet18\nInD: ImageNet-30Test AccOOD Dataset
DTDCUB-200CaltechDogsPlaces-365Food-101PetFlowers
CSI97.093.793.491.997.792.587.096.996.0
COD97.294.092.891.895.993.780.496.295.4
", + "bbox": [ + 127, + 689, + 843, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3: OOD detection performance comparison with CSI and COD with layer selection.", + "bbox": [ + 186, + 763, + 779, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "obtained from the re-defined feature of the last layer, shows an AUROC score of over $80\\%$ , and the remaining blocks show AUROC scores of about $50\\%$ . Based on this, we choose only the last layer for $COD$ , and perform OOD detection for ImageNet-30. ResNet18 is used for comparison", + "bbox": [ + 75, + 818, + 470, + 895 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with other approach. The baseline is a contrastive learning-based model, CSI [27], one of SOTA models. CSI is heavier to apply because it use a specific training method for OOD detection such as input transform and contrastive learning. On the other hand, $COD$ is simple and easy to apply be", + "bbox": [ + 498, + 818, + 890, + 895 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16566", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cause it does not require anything but generating classifiers in the training process. Nevertheless, as shown in Table 3, $COD$ shows comparable performance to CSI.", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Classification Accuracy", + "text_level": 1, + "bbox": [ + 76, + 150, + 294, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Some approaches for OOD detection suffered from the degradation of classification performance by change of model structures or loss functions [10, 21, 12]. In this case, two models should be stored aside, one for classification and the other for OOD detection. OOD detection models also need to have high classification performance for efficiency. Table 4 shows that models trained with the cosine-based softmax loss are accurate as much as the standard loss. Our approach can be applied to various model architectures because it only changes the calculation of logits. Also, there is no disadvantage to classification performance.", + "bbox": [ + 75, + 175, + 470, + 340 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a742f2a013b0f038272f347b9780a50ed772cce7e61ff24d301563d88e6e8df9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
modelInDStandardCosine
ResNet34CIFAR-1095.2595.34
CIFAR-10077.7077.78
SVHN96.4896.41
DenseNet3CIFAR-1095.1495.08
CIFAR-10076.4376.87
SVHN96.4296.33
", + "bbox": [ + 106, + 356, + 442, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4: Comparison of classification performance", + "bbox": [ + 104, + 479, + 439, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Is the Input Processing Helpful?", + "text_level": 1, + "bbox": [ + 76, + 540, + 359, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The input processing is helpful for performance improvement when the perturbation size is properly chosen. We analyze the performance improvement by the perturbation size of the input processing. We train DenseNet3 with CIFAR-100 and observe the AUROC improvements with different perturbations for SVHN, LSUNc, and iSUN datasets. We perform ODIN and MAHA with pre-trained DenseNet3 with CIFAR-100. Figure 5 shows the improvement by the input processing on the three method.", + "bbox": [ + 75, + 565, + 468, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Each dataset has a different pattern of improvement. For example, our method is improved for SVHN but degraded for iSUN when the epsilon is large. ODIN and MAHA also have similar pattern to our method. This means that it is challenging to find the perturbation size that generally increases the performance regardless of OOD datasets. Hsu et al. tried to find the perturbation size without using OOD datasets, but the performance decreased in some cases [12].", + "bbox": [ + 75, + 702, + 468, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The input processing is helpful where the distribution of OOD samples are known. However, in reality, OOD samples can be generated from any distribution. The input processing significantly increases the amount of computation. It is reasonable not to use input processing, which increases", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dc3048eef10c4b0a757c135fbe3c4743a0ef63cc441400a3ce133a54a33a3ffc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 87, + 854, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/121e7516251752cea139f330627e2857ccda030f12714d37d457f382f6ef0d48.jpg", + "image_caption": [ + "(a)ODIN" + ], + "image_footnote": [], + "bbox": [ + 500, + 229, + 854, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5c2e22dd6992980ce4b832c6cf3d92b9197357fa5ef62f1baa5728fa8e8eabec.jpg", + "image_caption": [ + "(b) MAHA", + "(c) Proposed Method", + "Figure 5: OOD detection performance (AUROC) increment by the perturbation size." + ], + "image_footnote": [], + "bbox": [ + 498, + 375, + 854, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "computation cost and has a potential to decrease performance.", + "bbox": [ + 498, + 617, + 888, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Fine-tuning with Pre-trained Model", + "text_level": 1, + "bbox": [ + 500, + 662, + 810, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We verify that our approach is also effective to fine-tuned models. We first train a model with the standard softmax, and then fine-tune the last block of the feature extractor and the classifier for only 20 epochs. In the fine-tuning process, all settings except for the learning rate are kept consistent with the previous configuration. The learning rate for fine-tuning starts with 0.5 and decays by a factor 0.1 at $50\\%$ and $75\\%$ of the training progress.", + "bbox": [ + 496, + 686, + 890, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We can notify that our approach still works for fine-tuned models. The OOD detection performances are comparable to or even higher than those of our original models, as shown in Table 5. The DenseNet3 trained with the standard softmax for SVHN shows a classification accuracy of $96.4\\%$ , but its fine-tuned model shows an accuracy of", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16567", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2b5d4fa8717eb08d949d4d2b85cc2b1eb78c7a469ab2095fa67b091b003ea076.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
InDOODAUROCTNR95
CIFAR-10 (Acc: 95.1 → 94.1)SVHN99.4598.35
TINc99.5399.61
TINr99.7399.23
LSUNc99.8299.96
LSUNr99.7399.34
iSUN99.7299.27
CIFAR-100 (Acc: 76.4 → 75.7)SVHN96.3979.48
TINc97.5388.87
TINr98.0389.37
LSUNc98.7296.81
LSUNr98.1690.91
iSUN97.9988.76
SVHN (Acc: 96.4 → 96.1)CIFAR-1099.1798.18
CIFAR-10099.2298.33
TINc99.94100.00
TINr99.8899.96
LSUNc99.94100.00
LSUNr99.91100.00
iSUN99.9199.99
", + "bbox": [ + 78, + 89, + 472, + 398 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 5: Classification and OOD detection performance of fine-tuned DenseNet3 models. All figures are in %, and are the average results of 5 experiments each.", + "bbox": [ + 75, + 409, + 470, + 455 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "$96.1\\%$ . Even though the models are fine-tuned with a limited number of epochs, the classification accuracy and the OOD detection performances remain almost same. The results with ResNet34 are also presented in the appendix.", + "bbox": [ + 75, + 506, + 468, + 568 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 595, + 194, + 612 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We proposed a novel OOD detection method that can be applied to deep neural network-based classifiers. It is simple and fast since it uses only the feature norm and the Mahalanobis distances. Our approach does not use additional data, does not have hyperparameters, and does not perform input processing. Our approach is easy to apply because it only needs to apply the cosine-based softmax loss.", + "bbox": [ + 75, + 627, + 468, + 733 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 76, + 763, + 243, + 781 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-00421, AI Graduate School Support Program (Sungkyunkwan University)), and also supported by the MSIT under the ICT Creative Consilience Program (IITP-2023-2020-0-01821) supervised by the IITP.", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 501, + 90, + 596, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Liron Bergman and Yedid Hoshen. Classification-based anomaly detection for general data. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 3", + "[2] Wenhu Chen, Yilin Shen, Hongxia Jin, and William Wang. A variational dirichlet framework for out-of-distribution detection. arXiv preprint arXiv:1811.07308, 2018. 2", + "[3] Sung-Ik Choi and Sae-Young Chung. Novelty detection via blurring. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 2", + "[4] Erik Daxberger and José Miguel Hernández-Lobato. Bayesian variational autoencoders for unsupervised out-of-distribution detection. arXiv preprint arXiv:1912.05651, 2019. 2", + "[5] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 4690-4699. Computer Vision Foundation / IEEE, 2019. 3", + "[6] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016, volume 48 of JMLR Workshop and Conference Proceedings, pages 1050-1059. JMLR.org, 2016. 2", + "[7] Izhak Golan and Ran El-Yaniv. Deep anomaly detection using geometric transformations. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada, pages 9781-9791, 2018. 3", + "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society, 2016. 6", + "[9] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. 1, 2", + "[10] Dan Hendrycks, Mantas Mazeika, and Thomas G. Dietterich. Deep anomaly detection with outlier exposure. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. 1, 2, 5, 8", + "[11] Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self-supervised learning can improve model robustness and uncertainty. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS" + ], + "bbox": [ + 501, + 114, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16568", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2019, December 8-14, 2019, Vancouver, BC, Canada, pages 15637-15648, 2019. 3", + "[12] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized ODIN: detecting out-of-distribution image without learning from out-of-distribution data. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 10948-10957. IEEE, 2020. 1, 2, 3, 5, 6, 7, 8", + "[13] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 2261-2269. IEEE Computer Society, 2017. 6", + "[14] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, 2009. 6", + "[15] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada, pages 7167-7177, 2018. 1, 2, 3, 4, 5, 6", + "[16] Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net, 2018. 1, 2, 3, 5, 6", + "[17] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In Advances in Neural Information Processing Systems, volume 33, pages 21464-21475. Curran Associates, Inc., 2020. 2", + "[18] Andrey Malinin and Mark J. F. Gales. Predictive uncertainty estimation via prior networks. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada, pages 7047-7058, 2018. 2", + "[19] Eric Nalisnick, Akihiro Matsukawa, Yee Whye Teh, and Balaji Lakshminarayanan. Detecting out-of-distribution inputs to deep generative models using a test for typicality. arXiv preprint arXiv:1906.02994, 5:5, 2019. 2", + "[20] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 6", + "[21] Aristotelis-Angelos Papadopoulos, Mohammad Reza Rajati, Nazim Shaikh, and Jiamian Wang. Outlier exposure with confidence control for out-of-distribution detection. Neurocomputing, 441:138-150, 2021. 1, 2, 8", + "[22] Jie Ren, Peter J. Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark A. DePristo, Joshua V. Dillon, and Balaji Lakshminarayanan. Likelihood ratios for out-of-distribution detection. In Advances in Neural Information Processing Sys-" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 14680-14691, 2019. 2", + "[23] Chandramouli Shama Sastry and Sageev Oore. Detecting out-of-distribution examples with gram matrices. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 8491-8501. PMLR, 2020. 1, 2, 5, 6, 7", + "[24] Vikash Sehwag, Mung Chiang, and Prateek Mittal. $\\{\\mathrm{SSD}\\}$ : A unified framework for self-supervised outlier detection. In International Conference on Learning Representations, 2021. 3", + "[25] Joan Serrà, David Álvarez, Vicenç Gómez, Olga Slizovskaia, José F. Núñez, and Jordi Luque. Input complexity and out-of-distribution detection with likelihood-based generative models. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 2", + "[26] Jiaming Song, Yang Song, and Stefano Ermon. Unsupervised out-of-distribution detection with batch normalization. arXiv preprint arXiv:1910.09115, 2019. 2", + "[27] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. In Advances in Neural Information Processing Systems, volume 33, pages 11839-11852. Curran Associates, Inc., 2020. 3, 7", + "[28] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 5265-5274. IEEE Computer Society, 2018. 3", + "[29] Xiao Zhang, Rui Zhao, Yu Qiao, Xiaogang Wang, and Hongsheng Li. Adacos: Adaptively scaling cosine logits for effectively learning deep face representations. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 10823-10832. Computer Vision Foundation / IEEE, 2019. 3" + ], + "bbox": [ + 501, + 92, + 890, + 641 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16569", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss/081c8308-e453-453c-8d9f-1e2b9b04b642_model.json b/2023/Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss/081c8308-e453-453c-8d9f-1e2b9b04b642_model.json new file mode 100644 index 0000000000000000000000000000000000000000..319b11532b28ff7d3e6be225b7c3e3334f530014 --- /dev/null +++ b/2023/Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss/081c8308-e453-453c-8d9f-1e2b9b04b642_model.json @@ -0,0 +1,2255 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.228, + 0.131, + 0.744, + 0.175 + ], + "angle": 0, + "content": "Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.204, + 0.723, + 0.222 + ], + "angle": 0, + "content": "SoonCheol Noh DongEon Jeong Jee-Hyong Lee*" + }, + { + "type": "text", + "bbox": [ + 0.379, + 0.223, + 0.59, + 0.24 + ], + "angle": 0, + "content": "Sungkyunkwan University" + }, + { + "type": "text", + "bbox": [ + 0.428, + 0.241, + 0.544, + 0.256 + ], + "angle": 0, + "content": "Suwon, Korea" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.26, + 0.644, + 0.275 + ], + "angle": 0, + "content": "{yhnuhb27, wjdehddjs, john}@skku.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.473, + 0.614 + ], + "angle": 0, + "content": "Deep learning models need to detect out-of-distribution (OOD) data in the inference stage because they are trained to estimate the train distribution and infer the data sampled from the distribution. Many methods have been proposed, but they have some limitations, such as requiring additional data, input processing, or high computational cost. Moreover, most methods have hyperparameters to be set by users, which have a significant impact on the detection rate. We propose a simple and effective OOD detection method by combining the feature norm and the Mahalanobis distance obtained from classification models trained with the cosine-based softmax loss. Our method is practical because it does not use additional data for training, is about three times faster when inferencing than the methods using the input processing, and is easy to apply because it does not have any hyperparameters for OOD detection. We confirm that our method is superior to or at least comparable to state-of-the-art OOD detection methods through the experiments." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.657, + 0.21, + 0.673 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.47, + 0.817 + ], + "angle": 0, + "content": "Deep learning models are generally designed to target the closed world. Models are trained based on the assumption that the inputs of the inference stage will be sampled from the same distribution as the train distribution. This means that the model may not make correct predictions if inputs are sampled from other distributions. The model must be able to distinguish whether the input is from the train distribution or not. This problem has been established as the task of out-of-distribution (OOD) detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.818, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Some researchers tried to detect OOD samples using confidence scores from classification models. Hendrycks and Gimpel simply used the confidence scores, but this approach was not very effective because of the overconfi" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.311, + 0.893, + 0.477 + ], + "angle": 0, + "content": "dence of deep learning models [9]. Hendrycks, Mazeika, and Dietterich (2019) and Papadopoulos et al. tried to lower the confidence of OOD samples using auxiliary large OOD datasets [10, 21]. However, collecting a massive OOD dataset is highly expensive, and training classification models with such a dataset takes a long time. Liang, Li and Srikant (2018) simply adopted the input processing technique, which added a small perturbation to the input samples to increase the confidence of in-distribution (InD) samples, but they assumed OOD samples were given in advance, which is not practical [16]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.893, + 0.644 + ], + "angle": 0, + "content": "Some researchers defined measures to detect OOD samples based on input features learned by classification models. Lee et al. proposed a measure based on Mahalanobis distance, but they also assumed that OOD samples were given in advance as Liang et al. did [15, 16]. Sastry and Oore proposed a measure based on various correlations between features, but their approach was very time-consuming in evaluating the OOD measure [23]. Hsu et al. tried to obtain the OOD probability by modifying the network structure, which caused the degradation of classification performances [12]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.795 + ], + "angle": 0, + "content": "Additional popular technique is the input processing, which was adopted by most approaches [16, 15, 23]. The input processing is to generate \\( x' \\) by adding a small perturbation to input data \\( x \\) to make the boundary of in-distribution compact. It may be helpful to increase the detection accuracy under certain conditions. However, generating the perturbation takes almost twice of the inference time, so the total inference becomes three times longer. The OOD detection performance is very susceptible to the size of the perturbation, so it should be carefully determined." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Some OOD detection approaches had various hyperparameters, including the perturbation size of input processing. The hyperparameters were usually tuned by using additional datasets to maximize the detection performance. Liang et al. [16] and Lee et al. [15] tuned them with given OOD samples, and Hsu et al. [12] and Sastry and Oore [23] used the validation samples. However, it is hard to find out" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.225, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "16560" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.137 + ], + "angle": 0, + "content": "hyperparameter values generally applicable to any distributions of OOD samples. When the values are found in a specific dataset, the model is easy to be biased to the dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.468, + 0.362 + ], + "angle": 0, + "content": "This paper proposes a simple and effective approach that can detect OOD samples without hyperparameters, additional datasets, and input processing. We detect OOD samples by combining norm and Mahalanobis distances of features. We find that the models trained with the cosine-based softmax loss are more advantageous than the ones trained with the regular softmax loss for OOD detection. We mathematically conjecture that the feature norm from the models with the cosine-based softmax loss can be a measure for the probability of in-distribution, and confirm its effectiveness experimentally. We also develop a robust Mahalanobis distance-based measure for the model with the cosine-based softmax loss. We propose an OOD score function by combining the feature norm and the Mahalanobis distance without any hyperparmaeters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.468, + 0.618 + ], + "angle": 0, + "content": "Our method do not need the input processing because it shows high accuracy in OOD detection without it. So, the additional computation for OOD detection by our method is very small. Feature norms and the Mahalanobis distances can be obtained with tiny additional calculations from classification models. Since the cosine-based softmax loss is generally applicable to any deep models, our OOD detection approach can be easily extended to various structures without a classification accuracy loss. Our method is also practical because it does not use any additional datasets. We confirm the OOD detection performance on various image datasets through experiments. Our method is compared with others using additional datasets, performing the input processing, or tuning hyperparameters to maximize performance. The experiment shows that our approach is superior to or at least comparable to state-of-the-art OOD detection methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.633, + 0.219, + 0.648 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.468, + 0.717 + ], + "angle": 0, + "content": "Previous OOD detection approaches were based on various models, such as classification models [9, 16, 15, 10, 21, 17, 12, 23], generative models [22, 19, 25, 4, 26, 3] and Bayesian models [6, 2, 18]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.468, + 0.822 + ], + "angle": 0, + "content": "OOD detection is usually required as a pre-processing of the classification task, which is one of the crucial tasks of deep learning. Due to its effectiveness, many researchers tried simultaneously performing OOD detection and classification in one model. In this section, we cover OOD detection-related approaches based on classification models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.383, + 0.849 + ], + "angle": 0, + "content": "2.1. Training with Large OOD Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Hendrycks et al. proposed a method to lower the maximum softmax probability (MSP) for OOD samples from classification models [10]. They improved OOD detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "performance by adding a loss term which was based on large OOD datasets. The added loss term was designed so that softmax values for OOD samples be closed to a uniform distribution. Papadopoulos et al. further improved OOD detection performance by adding another loss term [21]. The loss term prevented classification models from overconfidence. Liu et al. proposed an energy-based measure for models trained with large OOD datasets [17]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.217, + 0.892, + 0.322 + ], + "angle": 0, + "content": "These methods had the disadvantage of collecting extensive additional datasets to train classification models. The collected OOD dataset was used, assuming it represented all unknown OODs. However, it is virtually impossible to collect data that can represent all OOD samples in real-world situations. In addition, training a model with a large amount of OOD data takes much longer than regular training." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.343, + 0.892, + 0.374 + ], + "angle": 0, + "content": "2.2. Tuning Hyperparameters Using Known OOD Samples" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.55 + ], + "angle": 0, + "content": "Some OOD detection approaches had hyperparameters that need to be tuned. Liang et al. [16] and Lee et al. [15] used specific OOD samples to tune hyperparameters under the assumption that distribution of OOD samples were given in advance. Liang et al. [16] tuned the temperature scaling factor and a perturbation size of input processing. Lee et al. [15] tuned weights for the feature ensemble and a perturbation size of input processing. However, these approaches are not feasible because it is difficult to know from which distribution OOD samples will occur in a real-world situation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.572, + 0.892, + 0.603 + ], + "angle": 0, + "content": "2.3. Tuning Hyperparameters Using In-distribution Samples" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recently, some OOD detection approaches without using OOD samples were proposed because it was unrealistic to use OOD samples. Hsu et al. modified the classifying layer to obtain the probability that an input was in-distribution [12]. However, the modified classifying layer made the classification accuracy unstable. They also proposed a strategy that the perturbation size of the input processing was set without OOD samples, but it was not valid to all datasets. Sastry and Oore [23] proposed measures for OOD detection based on the Gram matrices of features. They stored the maximum and minimum values of the elements of the Gram matrix for each class, layer, power, and index of the matrix. They detected OOD samples using the stored values. However, it required heavy computation and a longer time than other approaches. Those approaches used additional InD validation dataset for tuning hyperparameters: Hsu et al. [12] tuned the perturbation size of input processing, and Sastry and Oore [23] tuned normalization factors for each layer." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16561" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.294, + 0.108 + ], + "angle": 0, + "content": "2.4. Using Input Processing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.468, + 0.22 + ], + "angle": 0, + "content": "The input processing is an operation that generates \\( x' \\) by adding a small perturbation to an input \\( x \\). Many previous approaches applied the input processing because this process effectively improved OOD detection performance. However, the perturbation size, epsilon, needed to be tuned appropriately. The epsilon has a significant impact on performance, so it should be chosen carefully." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.22, + 0.468, + 0.355 + ], + "angle": 0, + "content": "Liang et al. [16] and Lee et al. [15] chose the epsilon values using given specific OOD samples. The chosen epsilon may not be valid to other OOD samples because it was tuned with a distribution of given OOD samples. Hsu et al. [12] proposed a method to set the epsilon value without OOD samples. They tuned the epsilon using the InD validation dataset, but their method did not generally work for all cases. The detection performance decreased in some datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.468, + 0.477 + ], + "angle": 0, + "content": "Another disadvantage is that the input processing requires a lot of computation. We need one forward loop and one backward loop to add a perturbation to an input. We need one more forward loop to determine whether the perturbed input is OOD. Since the time taken for the forward loop and the backward loop is almost the same, the total inference time increases by about three times compared to the inference time without the input processing." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.484, + 0.355, + 0.501 + ], + "angle": 0, + "content": "2.5. Using Self-supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.468, + 0.718 + ], + "angle": 0, + "content": "Recently, a self-supervised learning technique that can learn a deep learning model without a label is attracting attention. In many studies in the field of OOD detection, self-supervised learning techniques are being tried to improve performance or reduce dependence on data labels. Some approaches build new data by applying transforms such as rotation to the image, and train models by giving the transform applied to the image as a label [7, 11, 27, 1]. Other approaches use contrastive learning techniques to train the model [27, 24]. However, in most cases, the OOD detection performance is not as high as when learning using labels. In addition, since the model is trained only for the purpose of OOD detection or representation learning, an additional method is needed to perform classification." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.731, + 0.251, + 0.748 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.756, + 0.468, + 0.817 + ], + "angle": 0, + "content": "A deep learning-based classification model can be decomposed into a feature extractor \\( f(\\cdot) \\) and a classifier \\( g(\\cdot) \\) which is usually a fully connected layer. When training with the standard softmax loss, a logit is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.831, + 0.468, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\operatorname {l o g i t} (x, c) = f (x) \\cdot w _ {c} + b _ {c}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.851, + 0.443, + 0.866 + ], + "angle": 0, + "content": "and with the cosine-based softmax loss, it is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.871, + 0.468, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\operatorname {l o g i t} (x, c) = \\frac {f (x) \\cdot w _ {c}}{| f (x) | \\cdot | w _ {c} |}, \\tag {2}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.1, + 0.685, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.198, + 0.676, + 0.211 + ], + "angle": 0, + "content": "(a) Standard softmax loss" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.1, + 0.871, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.198, + 0.883, + 0.211 + ], + "angle": 0, + "content": "(b) Cosine-based softmax loss" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.224, + 0.892, + 0.283 + ], + "angle": 0, + "content": "Figure 1: The feature norm distributions of InD (red) and OOD (blue) data. (a) is from DenseNet3 trained with the standard softmax loss and (b) is from DenseNet3 trained with the cosine-based softmax loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.892, + 0.368 + ], + "angle": 0, + "content": "where \\( w_{c} \\) is the connection weights of \\( g(\\cdot) \\) for a class \\( c \\). The cosine-based softmax loss is widely used in face recognition field [28, 5, 29]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.892, + 0.459 + ], + "angle": 0, + "content": "We introduce two useful measures for OOD detection that can be simply extracted from models trained with cosine-based softmax loss: a feature norm-based measure and a Mahalanobis distance-based measure. We then propose an OOD score by combining these two measures. Let us describe each measure and the score function in detail." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.469, + 0.892, + 0.484 + ], + "angle": 0, + "content": "3.1. Cosine-based Softmax Loss and Feature Norm" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.598 + ], + "angle": 0, + "content": "When we train a classification model with the closed-world assumption, the logit value for a class \\( c \\) is considered as being proportional to \\( P(y = c|x) \\). However, in the open world, the logit value should be interpreted as being proportional to \\( P(y = c|x, D_{in}) \\) where \\( D_{in} \\) is the event that \\( x \\) is in the train distribution. The probability can be decomposed using \\( P(y = c, D_{in}|x) \\) and \\( P(D_{in}|x) \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.608, + 0.892, + 0.642 + ], + "angle": 0, + "content": "\\[\n\\operatorname {l o g i t} (x, c) \\propto P (y = c | x, D _ {i n}) = \\frac {P (y = c , D _ {i n} | x)}{P \\left(D _ {i n} | x\\right)}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.65, + 0.892, + 0.679 + ], + "angle": 0, + "content": "If we use the cosine-based softmax loss, Equations 2 and 3 can be combined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.687, + 0.891, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\frac {f (x) \\cdot w _ {c}}{| f (x) | \\cdot | w _ {c} |} \\propto \\frac {P (y = c , D _ {i n} | x)}{P \\left(D _ {i n} | x\\right)}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.73, + 0.891, + 0.789 + ], + "angle": 0, + "content": "The numerator of the right-hand side includes a class-related term, \\( c \\), and the denominator, \\( P(D_{in}|x) \\), does not. So, we reformulate the left-hand side so that only the numerator has class-related terms as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.797, + 0.891, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\frac {f (x) \\cdot w _ {c} / | w _ {c} |}{| f (x) |} \\propto \\frac {P (y = c , D _ {i n} | x)}{P \\left(D _ {i n} | x\\right)}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Since \\( w_{c} \\) is the connection weight to \\( g(\\cdot) \\), we can say it is class-related, and since \\( f(x) \\) is the feature representation of \\( x \\), we can say it is classifier-independent. Based on this, we rigorously infer that both denominators can be proportional" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16562" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.101, + 0.26, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.209, + 0.252, + 0.223 + ], + "angle": 0, + "content": "(a) Distribution of \\( |f(x)| \\)" + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.1, + 0.45, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.209, + 0.426, + 0.223 + ], + "angle": 0, + "content": "(b) Mean of \\( |f(x)| \\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.235, + 0.47, + 0.309 + ], + "angle": 0, + "content": "Figure 2: The distribution of feature norm according to the mixup factor (a), and the mean value of feature norm according to the mixup factor. It can be observed that as OOD samples are more mixed into a sample, the feature norm tends to decrease." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.469, + 0.397 + ], + "angle": 0, + "content": "to each, i.e., \\( |f(x)| \\propto P(D_{in}|x) \\), because the two terms in Equation 5 are proportional, and their denominators and numerators share class-related characteristics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.469, + 0.548 + ], + "angle": 0, + "content": "To confirm this, we perform some OOD detection experiments. We try to detect SVHN using \\( |f(x)| \\) from DenseNet3 model trained for CIFAR-10 with the cosine-based softmax loss. We compare this to the model using the standard softmax loss as shown in Figure 1. The feature norm distributions of OOD and InD by the standard loss model overlap much, but those of the cosine loss model have a very small overlap between SVHN and CIFAR-10, which says that the cosine loss model can be better to discriminate ODD than the standard model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.469, + 0.609 + ], + "angle": 0, + "content": "To verify that the length of feature norm decrease as the samples get closer to OOD, we create mixup samples by mixing SVHN and CIFAR-10 images as shown in Equation 6 and observed the feature norm." + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.624, + 0.469, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\operatorname {T e s t I m g} = \\lambda \\cdot \\operatorname {I n D} + (1 - \\lambda) \\cdot \\operatorname {O O D} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.654, + 0.469, + 0.804 + ], + "angle": 0, + "content": "As the mixup factor, \\(\\lambda\\) decreases, that is, as the samples get closer to OOD, the feature norm also decreases as shown in Figure 2. The feature norm of an image from the cosine loss model tends to increase as the image gets closer to InD samples and decreases closer to OOD samples, which supports our conclusion that \\(|f(x)|\\) from a cosine loss model is proportional to \\(P(D_{in}|x)\\). Also, it supports that the feature norm can be a measure for OOD detection. Additional observations, including CIFAR-100 and SVHN, are presented in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.469, + 0.846 + ], + "angle": 0, + "content": "3.2. Mahalanobis Distance in Models Trained with Cosine-based Softmax Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We may assume that features of InD samples obtained from internal layers of classification models follow the Gaussian distribution. We adopt an OOD score defined by" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.099, + 0.687, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.192, + 0.693, + 0.205 + ], + "angle": 0, + "content": "(a) 1st Block (SVHN vs. TINc)" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.099, + 0.88, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.192, + 0.888, + 0.205 + ], + "angle": 0, + "content": "(b) 1st Block (SVHN vs. iSUN)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.217, + 0.685, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.312, + 0.695, + 0.325 + ], + "angle": 0, + "content": "(c) 2nd Block (SVHN vs. TINc)" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.217, + 0.879, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.312, + 0.889, + 0.325 + ], + "angle": 0, + "content": "(d) 2nd Block (SVHN vs. iSUN)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.337, + 0.685, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.432, + 0.694, + 0.445 + ], + "angle": 0, + "content": "(e) 3rd Block (SVHN vs. TINc)" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.337, + 0.879, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.432, + 0.888, + 0.445 + ], + "angle": 0, + "content": "(f) 3rd Block (SVHN vs. iSUN)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.459, + 0.892, + 0.504 + ], + "angle": 0, + "content": "Figure 3: Distributions of \\( MD \\) scores from each block of DenseNet3 trained with SVHN. The OODs datasets (blue) is TinyImageNet-crop (left) and iSUN (right)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.543, + 0.833, + 0.557 + ], + "angle": 0, + "content": "Lee et al. [15] based on the Mahalanobis distance:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.572, + 0.891, + 0.614 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} M D _ {B} (x) \\\\ = - \\max _ {c} - \\left(f _ {B} (x) - \\mu_ {c, B}\\right) ^ {\\top} \\Sigma_ {B} ^ {- 1} \\left(f _ {B} (x) - \\mu_ {c, B}\\right), \\quad (7) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.892, + 0.743 + ], + "angle": 0, + "content": "where \\(\\mu_{c,B}\\) and \\(\\Sigma_B\\) are the mean and the covariance of features from a block \\(B\\). This score is very discriminative for OOD detection as shown in Figure 3. Since the distributions of \\(MD\\) scores of InD and OOD samples have small overlaps, we can detect OOD samples using the scores. However, \\(MD\\) scores from the penultimate layer overlap if the model is trained with cosine-based softmax loss, as shown in Figure 2(e)-(f)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.743, + 0.892, + 0.832 + ], + "angle": 0, + "content": "The reason why \\( MD \\) scores from penultimate layer is not discriminative is that the feature from the penultimate layer do not follow a Gaussian distribution. In the model trained with cosine-based softmax loss, the softmax probability with \\( \\sigma_c \\), the softmax value for a class \\( c \\), is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.854, + 0.891, + 0.9 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {c} (g (f (x))) = \\frac {\\exp \\left(w _ {c} \\cdot f (x) / \\left(\\left| w _ {c} \\right| \\cdot | f (x) |\\right)\\right)}{\\sum_ {c ^ {\\prime}} \\exp \\left(w _ {c ^ {\\prime}} \\cdot f (x) / \\left(\\left| w _ {c ^ {\\prime}} \\right| \\cdot | f (x) |\\right)\\right)}. \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16563" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.252, + 0.089, + 0.72, + 0.198 + ], + "angle": 0, + "content": "
ApproachAuxiliary Dataset# of Hyperp.Input Processing
OE [10]Large OOD dataset0×
ODIN [16]OOD samples2✓
MAHA [15]OOD samples4✓
G-ODIN [12]InD samples1✓
GRAM [23]InD samples100×
COD (Proposed)None0×
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.208, + 0.895, + 0.24 + ], + "angle": 0, + "content": "Table 1: OOD detection approach comparison. The number of hyperparameters is counted based on DenseNet3 with 100 layers." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.266, + 0.473, + 0.362 + ], + "angle": 0, + "content": "The classification model is trained so that the softmax probability reflects \\( P(y = c|f(x)) \\). If the feature distribution from the penultimate layer is assumed to be the multivariate Gaussian distribution for each class, and the class prior is \\( P(y = c) = \\frac{\\beta_c}{\\sum_{c'} \\beta_{c'}} \\) where \\( \\beta_c \\) is unnormalized prior for class \\( c \\), the probability of class \\( c \\) with \\( f(x) \\) is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.391, + 0.47, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P (y = c \\mid f (x)) \\\\ = \\frac {P (y = c) P (f (x) | y = c)}{\\sum_ {c ^ {\\prime}} P (y = c ^ {\\prime}) P (f (x) | y = c ^ {\\prime})} \\\\ = \\frac {\\exp \\left(\\mu_ {c} ^ {T} \\Sigma^ {- 1} f (x) - \\frac {1}{2} \\mu_ {c} ^ {T} \\Sigma^ {- 1} \\mu_ {c} + \\log \\beta_ {c}\\right)}{\\sum_ {c ^ {\\prime}} \\exp \\left(\\mu_ {c ^ {\\prime}} ^ {T} \\Sigma^ {- 1} f (x) - \\frac {1}{2} \\mu_ {c ^ {\\prime}} ^ {T} \\Sigma^ {- 1} \\mu_ {c ^ {\\prime}} + \\log \\beta_ {c ^ {\\prime}}\\right)}. \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.47, + 0.631 + ], + "angle": 0, + "content": "That is, if the feature from the penultimate layer follows a Gaussian distribution, \\( P(y = c|f(x)) \\) is based on linear expressions. Since a nonlinear expression cannot be transformed in a linear expression, the softmax probability from the cosine loss model (Equation 8) cannot be converted to a similar form of \\( P(y = c|f(x)) \\) (Equation 9), which means that the features from the penultimate layer is not expected to follow a Gaussian distribution." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.633, + 0.47, + 0.665 + ], + "angle": 0, + "content": "If Equation 8 is rewritten using a normalized feature \\( f^{\\prime}(x) = f(x) / |f(x)| \\), it is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.687, + 0.47, + 0.724 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {c} (g (f (x))) = \\frac {\\exp \\left(w _ {c} ^ {\\top} f ^ {\\prime} (x) / \\left(\\left| w _ {c} ^ {\\top} \\right|\\right)\\right)}{\\sum_ {c ^ {\\prime}} \\exp \\left(w _ {c ^ {\\prime}} ^ {\\top} f ^ {\\prime} (x) / \\left(\\left| w _ {c ^ {\\prime}} ^ {\\top} \\right|\\right)\\right)}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Equation 10 is based on linear expressions of \\( f'(x) \\), from which we conjecture that \\( f'(x) \\) may follow a Gaussian distribution. Since \\( MD \\) scores are useful under the assumption that features follow a Gaussian distribution, we redefine the feature for \\( MD \\) from the penultimate layer as \\( f'(x) = f(x) / |f(x)| \\) instead of \\( f(x) \\). We re-depict the distribution of \\( MD \\) in Figures 2-(e) and (f) using \\( f'(x) \\) as shown in Figure 4. The overlap of the distributions is significantly reduced compared to \\( MD \\) based on \\( f(x) \\), which means that \\( MD \\) based on \\( f'(x) \\) may be better than \\( f(x) \\) for OOD detection." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.273, + 0.687, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.365, + 0.693, + 0.38 + ], + "angle": 0, + "content": "(a) 3rd Block (SVHN vs TINc)" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.272, + 0.88, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.365, + 0.888, + 0.38 + ], + "angle": 0, + "content": "(b) 3rd Block (SVHN vs iSUN)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.391, + 0.892, + 0.423 + ], + "angle": 0, + "content": "Figure 4: Distributions of \\( MD \\) scores based on \\( f^{\\prime}(x) \\). The configuration is the same as Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.461, + 0.893, + 0.522 + ], + "angle": 0, + "content": "We evaluate \\( MD \\) scores for the penultimate layer with the redefined feature, \\( f'(x) = f(x) / |f(x)| \\) and for other layers with \\( f_B(x) \\). Finally, we combine \\( MD \\) scores from each block as Equation 11 by product:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.533, + 0.892, + 0.565 + ], + "angle": 0, + "content": "\\[\nM D (x) = \\prod_ {B} \\left(M D _ {B} (x)\\right). \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.572, + 0.776, + 0.589 + ], + "angle": 0, + "content": "3.3. Proposed OOD Score Function" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.596, + 0.892, + 0.687 + ], + "angle": 0, + "content": "We define a cosine-based OOD score function, \\( COD(x) \\), with the feature norm and the Mahalanobis distance-based measure, \\( MD \\). Since the feature norm tends to appear large for InD samples and \\( MD \\) tends to be large for OOD samples, we define \\( COD(x) \\) by dividing the feature norm by \\( MD \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.696, + 0.892, + 0.731 + ], + "angle": 0, + "content": "\\[\nC O D (x) = \\frac {| f (x) |}{M D (x)}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "\\( COD(x) \\) does not have any hyperparameter to be tuned. Also, no additional dataset is required to use \\( COD(x) \\), and it shows a high discriminative performance without the input processing. \\( COD(x) \\) can be calculated with a very small computation in a forward loop for classification. We compare the previous approaches in the viewpoint of auxiliary datasets, the number of hyperparameters and the input processing in Table 1. Most of the previous methods used auxiliary datasets, and had several to one hundred hyperparameters, and needed the input processing to increase the performance, while our method does not need any of them." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16564" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.201, + 0.108 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.116, + 0.468, + 0.147 + ], + "angle": 0, + "content": "This section shows the effectiveness of our approach through experiments with various image datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.156, + 0.262, + 0.172 + ], + "angle": 0, + "content": "4.1. Experiment setting" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.18, + 0.468, + 0.345 + ], + "angle": 0, + "content": "Datasets: We use CIFAR-10/100 [14], and SVHN [20] images for InD datasets. We use the training dataset for classification and use the test datasets to evaluate OOD detection. We use TinyImageNet-crop (TINc), TinyImageNet-resize (TINr), LSUN-crop (LSUNc), LSUN-resize (LSUNr), and iSUN as OOD datasets. These datasets have been used in many studies related to OOD detection. Additionally, when CIFAR-10/100 are in-distribution, SVHN test dataset is also used as an OOD dataset, and when SVHN image is in-distribution, CIFAR-10/100 test datasets are also used as OOD datasets." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.352, + 0.468, + 0.398 + ], + "angle": 0, + "content": "Baselines: We compare our approach to previous OOD detection methods: ODIN [16], MAHA [15], G-ODIN [12], and GRAM [23]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.404, + 0.468, + 0.54 + ], + "angle": 0, + "content": "Networks and Training Details: We train ResNet34 [8] and DenseNet3 [13] models with the same setting as the baseline approaches. We train ResNet34 for 200 epochs with a batch size of 128 and a weight decay of 0.0005. We train DenseNet3 with 100 layers for 300 epochs with a batch size of 64 and a weight decay of 0.0001. In both training, the optimizer is SGD with a momentum of 0.9, and the learning rate starts with 0.1 and decays by a factor 0.1 at \\(50\\%\\) and \\(75\\%\\) of the training progress." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.541, + 0.468, + 0.645 + ], + "angle": 0, + "content": "When we train models with the cosine-based softmax loss, logit values are between -1 and 1. For expanding the range of logits, we adopt the temperature scaling with a factor of \\( T \\) as in Equation 13. The factor of \\( T \\) needs to be set to 0.1 for SVHN, CIFAR-10 and 0.05 for CIFAR-100 so that the classification performance is maximized. The logit scaled by \\( T \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.662, + 0.468, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\operatorname {l o g i t} _ {T} (x) = \\operatorname {l o g i t} (x) / T. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.686, + 0.468, + 0.792 + ], + "angle": 0, + "content": "Evaluation Metrics: We evaluate the detection performance using the two metrics that are most widely used for OOD detection. The first metric is the area under the receiver operating characteristic curve (AUROC). It is used to evaluate the average performance considering various thresholds. The second is TNR at TPR \\( p\\% \\). We evaluate with \\( p = 95 \\) as the metric used in the baseline approaches." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.802, + 0.17, + 0.817 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Table 2 shows an overall comparison with DenseNet3. Our approach achieves superior performance in all the datasets. Compared to ODIN, our approach has higher performance in all the cases. Compared to G-ODIN, our approach has higher performance except one cases. Com" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "pared to MAHA, our approach has higher performance except some cases." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.123, + 0.892, + 0.364 + ], + "angle": 0, + "content": "In some cases, GRAM shows a higher performance than the proposed method, but GRAM is very unstable depending on OOD datasets. For example, GRAM shows a particularly poor performance on LSUNc. It shows a TNR@95 score of 88.4 for CIFAR-10 and 65.6 for CIFAR-100. Our approach is more even performances across all OOD datasets, and achieves the highest average performances. Also, GRAM requires significantly heavy computation. When comparing the wall clock time in the same environment (GTX 2080 Ti GPU with Xeon Silver 4021 CPU * 2), the GRAM takes 0.35 seconds to calculate OOD score for one image, but our approach takes only about 0.018 seconds. Our approach is 19 times faster than GRAM because GRAM evaluates a lot of measures in the forward loop while our approach needs very small additional computation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.365, + 0.892, + 0.562 + ], + "angle": 0, + "content": "The head of Table 2 shows what each approach needs to detect OOD samples. ODIN and MAHA use OOD samples and InD validation samples, and perform the input processing. G-ODIN does not use OOD samples but uses InD validation samples and performs the input processing. GRAM uses InD validation samples to tune hyperparameters. Our approach needs nothing except a training dataset. Our approach is the fastest method because the input processing is not required. Our method is superior to or at least comparable to state-of-the-art OOD detection methods even though it does not require any additional datasets. Our model also shows better performance with ResNet34. The results are presented in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.578, + 0.614, + 0.593 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.719, + 0.619 + ], + "angle": 0, + "content": "5.1. Layer Selection for \\(MD\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.628, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Our approach, \\( COD \\), uses \\( MD \\) values from all the blocks of the model. However, it can be more effective if we can choose some layers which are helpful for OOD detection. Useful layers for OOD detection may depend on the model structure, the training dataset, and the OOD dataset. However, OOD samples are not available in advance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.719, + 0.892, + 0.87 + ], + "angle": 0, + "content": "We will discuss how to select useful layers with synthetic OOD samples. We may generate synthetic OOD samples from the training dataset. For example, we may use some training samples rotated by 90 degrees as synthetic OOD samples. Then, we evaluate AUROC scores for each layer, and choose layers of which AUROC scores are greater than a threshold such as \\(60\\%\\) of AUROC. In the case of CIFAR10, CIFAR100, and SVHN, it is confirmed that the AUROC score of each \\(MD\\) was over \\(60\\%\\). Therefore, it can be a reasonable threshold to choose layers." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We apply the layer selection to a more realistic dataset, ImageNet-30. In this case, \\(MD\\) from the last layer, \\(MD\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16565" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.088, + 0.887, + 0.593 + ], + "angle": 0, + "content": "
OODAUROCTNR@95
Use a training datasetUse a training dataset
Use InD validation samples(ours)Use InD validation samples(ours)
Perform the input processingGRAMPerform the input processingGRAMCOD
Use OOD samplesG-ODINUse OOD samplesG-ODIN
ODINMAHAODINMAHA
InD: CIFAR-10
SVHN95.598.198.899.199.586.290.894.096.198.0
TINc97.695.398.799.399.487.084.293.496.797.9
TINr98.598.899.199.799.592.495.095.898.897.8
LSUNc93.680.298.397.599.970.648.291.588.499.6
LSUNr99.299.399.499.999.696.297.297.699.598.9
iSUN98.798.999.499.899.693.295.397.599.098.7
MEAN97.295.199.099.299.687.685.195.096.498.5
InD: CIFAR-100
SVHN93.897.295.997.397.070.682.577.089.382.4
TINc88.388.897.697.797.951.060.187.889.089.3
TINr85.297.498.699.098.642.686.693.395.793.1
LSUNc91.481.795.391.498.957.842.175.065.595.1
LSUNr85.598.098.799.398.641.291.493.897.293.3
iSUN84.597.498.499.098.537.487.092.595.992.5
MEAN88.193.497.497.398.350.175.086.688.891.0
InD: SVHN
CIFAR-1091.498.9-95.599.071.796.8-80.496.1
CIFAR-100----98.9----95.8
TINc----99.8----99.7
TINr95.199.9-99.799.884.199.9-99.199.5
LSUNc----99.8----99.4
LSUNr94.599.9-99.899.881.199.9-99.599.4
iSUN94.799.9-99.899.882.299.9-99.499.5
MEAN----99.6----98.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.603, + 0.895, + 0.667 + ], + "angle": 0, + "content": "Table 2: OOD detection performance (%) with a DenseNet3 classification model. Our performance is averaged over 5 runs. ODIN, MAHA and GRAM are cited from the work by Sastry and Oore [23] and G-ODIN is cited from the work by Hsu et al. [12]. In the case of SVHN, all baselines did not present their performances. Dash symbols stand for not-available. The mean rows are the averages of the performances over all OOD datasets. For each case, the best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.128, + 0.69, + 0.844, + 0.753 + ], + "angle": 0, + "content": "
ResNet18\nInD: ImageNet-30Test AccOOD Dataset
DTDCUB-200CaltechDogsPlaces-365Food-101PetFlowers
CSI97.093.793.491.997.792.587.096.996.0
COD97.294.092.891.895.993.780.496.295.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.187, + 0.764, + 0.781, + 0.78 + ], + "angle": 0, + "content": "Table 3: OOD detection performance comparison with CSI and COD with layer selection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.819, + 0.471, + 0.896 + ], + "angle": 0, + "content": "obtained from the re-defined feature of the last layer, shows an AUROC score of over \\(80\\%\\), and the remaining blocks show AUROC scores of about \\(50\\%\\). Based on this, we choose only the last layer for \\(COD\\), and perform OOD detection for ImageNet-30. ResNet18 is used for comparison" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.819, + 0.892, + 0.896 + ], + "angle": 0, + "content": "with other approach. The baseline is a contrastive learning-based model, CSI [27], one of SOTA models. CSI is heavier to apply because it use a specific training method for OOD detection such as input transform and contrastive learning. On the other hand, \\( COD \\) is simple and easy to apply be" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16566" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "cause it does not require anything but generating classifiers in the training process. Nevertheless, as shown in Table 3, \\( COD \\) shows comparable performance to CSI." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.151, + 0.295, + 0.168 + ], + "angle": 0, + "content": "5.2. Classification Accuracy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.176, + 0.471, + 0.342 + ], + "angle": 0, + "content": "Some approaches for OOD detection suffered from the degradation of classification performance by change of model structures or loss functions [10, 21, 12]. In this case, two models should be stored aside, one for classification and the other for OOD detection. OOD detection models also need to have high classification performance for efficiency. Table 4 shows that models trained with the cosine-based softmax loss are accurate as much as the standard loss. Our approach can be applied to various model architectures because it only changes the calculation of logits. Also, there is no disadvantage to classification performance." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.357, + 0.443, + 0.472 + ], + "angle": 0, + "content": "
modelInDStandardCosine
ResNet34CIFAR-1095.2595.34
CIFAR-10077.7077.78
SVHN96.4896.41
DenseNet3CIFAR-1095.1495.08
CIFAR-10076.4376.87
SVHN96.4296.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.106, + 0.481, + 0.441, + 0.497 + ], + "angle": 0, + "content": "Table 4: Comparison of classification performance" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.541, + 0.36, + 0.558 + ], + "angle": 0, + "content": "5.3. Is the Input Processing Helpful?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.469, + 0.702 + ], + "angle": 0, + "content": "The input processing is helpful for performance improvement when the perturbation size is properly chosen. We analyze the performance improvement by the perturbation size of the input processing. We train DenseNet3 with CIFAR-100 and observe the AUROC improvements with different perturbations for SVHN, LSUNc, and iSUN datasets. We perform ODIN and MAHA with pre-trained DenseNet3 with CIFAR-100. Figure 5 shows the improvement by the input processing on the three method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Each dataset has a different pattern of improvement. For example, our method is improved for SVHN but degraded for iSUN when the epsilon is large. ODIN and MAHA also have similar pattern to our method. This means that it is challenging to find the perturbation size that generally increases the performance regardless of OOD datasets. Hsu et al. tried to find the perturbation size without using OOD datasets, but the performance decreased in some cases [12]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The input processing is helpful where the distribution of OOD samples are known. However, in reality, OOD samples can be generated from any distribution. The input processing significantly increases the amount of computation. It is reasonable not to use input processing, which increases" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.088, + 0.856, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.217, + 0.728, + 0.231 + ], + "angle": 0, + "content": "(a)ODIN" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.231, + 0.856, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.36, + 0.731, + 0.374 + ], + "angle": 0, + "content": "(b) MAHA" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.376, + 0.856, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.518, + 0.761, + 0.532 + ], + "angle": 0, + "content": "(c) Proposed Method" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.545, + 0.892, + 0.575 + ], + "angle": 0, + "content": "Figure 5: OOD detection performance (AUROC) increment by the perturbation size." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.618, + 0.89, + 0.647 + ], + "angle": 0, + "content": "computation cost and has a potential to decrease performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.663, + 0.812, + 0.679 + ], + "angle": 0, + "content": "5.4. Fine-tuning with Pre-trained Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.809 + ], + "angle": 0, + "content": "We verify that our approach is also effective to fine-tuned models. We first train a model with the standard softmax, and then fine-tune the last block of the feature extractor and the classifier for only 20 epochs. In the fine-tuning process, all settings except for the learning rate are kept consistent with the previous configuration. The learning rate for fine-tuning starts with 0.5 and decays by a factor 0.1 at \\(50\\%\\) and \\(75\\%\\) of the training progress." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We can notify that our approach still works for fine-tuned models. The OOD detection performances are comparable to or even higher than those of our original models, as shown in Table 5. The DenseNet3 trained with the standard softmax for SVHN shows a classification accuracy of \\(96.4\\%\\), but its fine-tuned model shows an accuracy of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16567" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.09, + 0.473, + 0.4 + ], + "angle": 0, + "content": "
InDOODAUROCTNR95
CIFAR-10 (Acc: 95.1 → 94.1)SVHN99.4598.35
TINc99.5399.61
TINr99.7399.23
LSUNc99.8299.96
LSUNr99.7399.34
iSUN99.7299.27
CIFAR-100 (Acc: 76.4 → 75.7)SVHN96.3979.48
TINc97.5388.87
TINr98.0389.37
LSUNc98.7296.81
LSUNr98.1690.91
iSUN97.9988.76
SVHN (Acc: 96.4 → 96.1)CIFAR-1099.1798.18
CIFAR-10099.2298.33
TINc99.94100.00
TINr99.8899.96
LSUNc99.94100.00
LSUNr99.91100.00
iSUN99.9199.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.41, + 0.472, + 0.456 + ], + "angle": 0, + "content": "Table 5: Classification and OOD detection performance of fine-tuned DenseNet3 models. All figures are in %, and are the average results of 5 experiments each." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.569 + ], + "angle": 0, + "content": "\\(96.1\\%\\). Even though the models are fine-tuned with a limited number of epochs, the classification accuracy and the OOD detection performances remain almost same. The results with ResNet34 are also presented in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.597, + 0.196, + 0.613 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.734 + ], + "angle": 0, + "content": "We proposed a novel OOD detection method that can be applied to deep neural network-based classifiers. It is simple and fast since it uses only the feature norm and the Mahalanobis distances. Our approach does not use additional data, does not have hyperparameters, and does not perform input processing. Our approach is easy to apply because it only needs to apply the cosine-based softmax loss." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.765, + 0.245, + 0.782 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-00421, AI Graduate School Support Program (Sungkyunkwan University)), and also supported by the MSIT under the ICT Creative Consilience Program (IITP-2023-2020-0-01821) supervised by the IITP." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.091, + 0.597, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.116, + 0.892, + 0.184 + ], + "angle": 0, + "content": "[1] Liron Bergman and Yedid Hoshen. Classification-based anomaly detection for general data. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.188, + 0.894, + 0.23 + ], + "angle": 0, + "content": "[2] Wenhu Chen, Yilin Shen, Hongxia Jin, and William Wang. A variational dirichlet framework for out-of-distribution detection. arXiv preprint arXiv:1811.07308, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.232, + 0.895, + 0.287 + ], + "angle": 0, + "content": "[3] Sung-Ik Choi and Sae-Young Chung. Novelty detection via blurring. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.289, + 0.895, + 0.343 + ], + "angle": 0, + "content": "[4] Erik Daxberger and José Miguel Hernández-Lobato. Bayesian variational autoencoders for unsupervised out-of-distribution detection. arXiv preprint arXiv:1912.05651, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.346, + 0.895, + 0.43 + ], + "angle": 0, + "content": "[5] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 4690-4699. Computer Vision Foundation / IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.433, + 0.895, + 0.528 + ], + "angle": 0, + "content": "[6] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016, volume 48 of JMLR Workshop and Conference Proceedings, pages 1050-1059. JMLR.org, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.531, + 0.895, + 0.613 + ], + "angle": 0, + "content": "[7] Izhak Golan and Ran El-Yaniv. Deep anomaly detection using geometric transformations. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada, pages 9781-9791, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.617, + 0.895, + 0.686 + ], + "angle": 0, + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.688, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[9] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.76, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[10] Dan Hendrycks, Mantas Mazeika, and Thomas G. Dietterich. Deep anomaly detection with outlier exposure. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.831, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[11] Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self-supervised learning can improve model robustness and uncertainty. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.116, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16568" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "2019, December 8-14, 2019, Vancouver, BC, Canada, pages 15637-15648, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[12] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized ODIN: detecting out-of-distribution image without learning from out-of-distribution data. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 10948-10957. IEEE, 2020. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[13] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 2261-2269. IEEE Computer Society, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.471, + 0.332 + ], + "angle": 0, + "content": "[14] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.471, + 0.432 + ], + "angle": 0, + "content": "[15] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada, pages 7167-7177, 2018. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.471, + 0.517 + ], + "angle": 0, + "content": "[16] Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net, 2018. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.471, + 0.573 + ], + "angle": 0, + "content": "[17] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In Advances in Neural Information Processing Systems, volume 33, pages 21464-21475. Curran Associates, Inc., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.471, + 0.658 + ], + "angle": 0, + "content": "[18] Andrey Malinin and Mark J. F. Gales. Predictive uncertainty estimation via prior networks. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada, pages 7047-7058, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.471, + 0.716 + ], + "angle": 0, + "content": "[19] Eric Nalisnick, Akihiro Matsukawa, Yee Whye Teh, and Balaji Lakshminarayanan. Detecting out-of-distribution inputs to deep generative models using a test for typicality. arXiv preprint arXiv:1906.02994, 5:5, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.471, + 0.786 + ], + "angle": 0, + "content": "[20] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[21] Aristotelis-Angelos Papadopoulos, Mohammad Reza Rajati, Nazim Shaikh, and Jiamian Wang. Outlier exposure with confidence control for out-of-distribution detection. Neurocomputing, 441:138-150, 2021. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[22] Jie Ren, Peter J. Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark A. DePristo, Joshua V. Dillon, and Balaji Lakshminarayanan. Likelihood ratios for out-of-distribution detection. In Advances in Neural Information Processing Sys-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "tems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 14680-14691, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[23] Chandramouli Shama Sastry and Sageev Oore. Detecting out-of-distribution examples with gram matrices. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 8491-8501. PMLR, 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[24] Vikash Sehwag, Mung Chiang, and Prateek Mittal. \\(\\{\\mathrm{SSD}\\}\\): A unified framework for self-supervised outlier detection. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[25] Joan Serrà, David Álvarez, Vicenç Gómez, Olga Slizovskaia, José F. Núñez, and Jordi Luque. Input complexity and out-of-distribution detection with likelihood-based generative models. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.361, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[26] Jiaming Song, Yang Song, and Stefano Ermon. Unsupervised out-of-distribution detection with batch normalization. arXiv preprint arXiv:1910.09115, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.403, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[27] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. In Advances in Neural Information Processing Systems, volume 33, pages 11839-11852. Curran Associates, Inc., 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.474, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[28] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 5265-5274. IEEE Computer Society, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.559, + 0.892, + 0.642 + ], + "angle": 0, + "content": "[29] Xiao Zhang, Rui Zhao, Yu Qiao, Xiaogang Wang, and Hongsheng Li. Adacos: Adaptively scaling cosine logits for effectively learning deep face representations. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 10823-10832. Computer Vision Foundation / IEEE, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16569" + } + ] +] \ No newline at end of file diff --git a/2023/Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss/081c8308-e453-453c-8d9f-1e2b9b04b642_origin.pdf b/2023/Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss/081c8308-e453-453c-8d9f-1e2b9b04b642_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f2809b5818fc8a6e9fcab5b005ff8c14ed2042b3 --- /dev/null +++ b/2023/Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss/081c8308-e453-453c-8d9f-1e2b9b04b642_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:745b6409661f24d7d58605543df4b2d87195237b2c12791724234fb99da02ce7 +size 790992 diff --git a/2023/Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss/full.md b/2023/Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss/full.md new file mode 100644 index 0000000000000000000000000000000000000000..31d862bb3fedfb0f22c8f741fdc77a3c6b838f35 --- /dev/null +++ b/2023/Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss/full.md @@ -0,0 +1,352 @@ +# Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss + +SoonCheol Noh DongEon Jeong Jee-Hyong Lee* + +Sungkyunkwan University + +Suwon, Korea + +{yhnuhb27, wjdehddjs, john}@skku.edu + +# Abstract + +Deep learning models need to detect out-of-distribution (OOD) data in the inference stage because they are trained to estimate the train distribution and infer the data sampled from the distribution. Many methods have been proposed, but they have some limitations, such as requiring additional data, input processing, or high computational cost. Moreover, most methods have hyperparameters to be set by users, which have a significant impact on the detection rate. We propose a simple and effective OOD detection method by combining the feature norm and the Mahalanobis distance obtained from classification models trained with the cosine-based softmax loss. Our method is practical because it does not use additional data for training, is about three times faster when inferencing than the methods using the input processing, and is easy to apply because it does not have any hyperparameters for OOD detection. We confirm that our method is superior to or at least comparable to state-of-the-art OOD detection methods through the experiments. + +# 1. Introduction + +Deep learning models are generally designed to target the closed world. Models are trained based on the assumption that the inputs of the inference stage will be sampled from the same distribution as the train distribution. This means that the model may not make correct predictions if inputs are sampled from other distributions. The model must be able to distinguish whether the input is from the train distribution or not. This problem has been established as the task of out-of-distribution (OOD) detection. + +Some researchers tried to detect OOD samples using confidence scores from classification models. Hendrycks and Gimpel simply used the confidence scores, but this approach was not very effective because of the overconfi + +dence of deep learning models [9]. Hendrycks, Mazeika, and Dietterich (2019) and Papadopoulos et al. tried to lower the confidence of OOD samples using auxiliary large OOD datasets [10, 21]. However, collecting a massive OOD dataset is highly expensive, and training classification models with such a dataset takes a long time. Liang, Li and Srikant (2018) simply adopted the input processing technique, which added a small perturbation to the input samples to increase the confidence of in-distribution (InD) samples, but they assumed OOD samples were given in advance, which is not practical [16]. + +Some researchers defined measures to detect OOD samples based on input features learned by classification models. Lee et al. proposed a measure based on Mahalanobis distance, but they also assumed that OOD samples were given in advance as Liang et al. did [15, 16]. Sastry and Oore proposed a measure based on various correlations between features, but their approach was very time-consuming in evaluating the OOD measure [23]. Hsu et al. tried to obtain the OOD probability by modifying the network structure, which caused the degradation of classification performances [12]. + +Additional popular technique is the input processing, which was adopted by most approaches [16, 15, 23]. The input processing is to generate $x'$ by adding a small perturbation to input data $x$ to make the boundary of in-distribution compact. It may be helpful to increase the detection accuracy under certain conditions. However, generating the perturbation takes almost twice of the inference time, so the total inference becomes three times longer. The OOD detection performance is very susceptible to the size of the perturbation, so it should be carefully determined. + +Some OOD detection approaches had various hyperparameters, including the perturbation size of input processing. The hyperparameters were usually tuned by using additional datasets to maximize the detection performance. Liang et al. [16] and Lee et al. [15] tuned them with given OOD samples, and Hsu et al. [12] and Sastry and Oore [23] used the validation samples. However, it is hard to find out + +hyperparameter values generally applicable to any distributions of OOD samples. When the values are found in a specific dataset, the model is easy to be biased to the dataset. + +This paper proposes a simple and effective approach that can detect OOD samples without hyperparameters, additional datasets, and input processing. We detect OOD samples by combining norm and Mahalanobis distances of features. We find that the models trained with the cosine-based softmax loss are more advantageous than the ones trained with the regular softmax loss for OOD detection. We mathematically conjecture that the feature norm from the models with the cosine-based softmax loss can be a measure for the probability of in-distribution, and confirm its effectiveness experimentally. We also develop a robust Mahalanobis distance-based measure for the model with the cosine-based softmax loss. We propose an OOD score function by combining the feature norm and the Mahalanobis distance without any hyperparmaeters. + +Our method do not need the input processing because it shows high accuracy in OOD detection without it. So, the additional computation for OOD detection by our method is very small. Feature norms and the Mahalanobis distances can be obtained with tiny additional calculations from classification models. Since the cosine-based softmax loss is generally applicable to any deep models, our OOD detection approach can be easily extended to various structures without a classification accuracy loss. Our method is also practical because it does not use any additional datasets. We confirm the OOD detection performance on various image datasets through experiments. Our method is compared with others using additional datasets, performing the input processing, or tuning hyperparameters to maximize performance. The experiment shows that our approach is superior to or at least comparable to state-of-the-art OOD detection methods. + +# 2. Related Work + +Previous OOD detection approaches were based on various models, such as classification models [9, 16, 15, 10, 21, 17, 12, 23], generative models [22, 19, 25, 4, 26, 3] and Bayesian models [6, 2, 18]. + +OOD detection is usually required as a pre-processing of the classification task, which is one of the crucial tasks of deep learning. Due to its effectiveness, many researchers tried simultaneously performing OOD detection and classification in one model. In this section, we cover OOD detection-related approaches based on classification models. + +# 2.1. Training with Large OOD Datasets + +Hendrycks et al. proposed a method to lower the maximum softmax probability (MSP) for OOD samples from classification models [10]. They improved OOD detection + +performance by adding a loss term which was based on large OOD datasets. The added loss term was designed so that softmax values for OOD samples be closed to a uniform distribution. Papadopoulos et al. further improved OOD detection performance by adding another loss term [21]. The loss term prevented classification models from overconfidence. Liu et al. proposed an energy-based measure for models trained with large OOD datasets [17]. + +These methods had the disadvantage of collecting extensive additional datasets to train classification models. The collected OOD dataset was used, assuming it represented all unknown OODs. However, it is virtually impossible to collect data that can represent all OOD samples in real-world situations. In addition, training a model with a large amount of OOD data takes much longer than regular training. + +# 2.2. Tuning Hyperparameters Using Known OOD Samples + +Some OOD detection approaches had hyperparameters that need to be tuned. Liang et al. [16] and Lee et al. [15] used specific OOD samples to tune hyperparameters under the assumption that distribution of OOD samples were given in advance. Liang et al. [16] tuned the temperature scaling factor and a perturbation size of input processing. Lee et al. [15] tuned weights for the feature ensemble and a perturbation size of input processing. However, these approaches are not feasible because it is difficult to know from which distribution OOD samples will occur in a real-world situation. + +# 2.3. Tuning Hyperparameters Using In-distribution Samples + +Recently, some OOD detection approaches without using OOD samples were proposed because it was unrealistic to use OOD samples. Hsu et al. modified the classifying layer to obtain the probability that an input was in-distribution [12]. However, the modified classifying layer made the classification accuracy unstable. They also proposed a strategy that the perturbation size of the input processing was set without OOD samples, but it was not valid to all datasets. Sastry and Oore [23] proposed measures for OOD detection based on the Gram matrices of features. They stored the maximum and minimum values of the elements of the Gram matrix for each class, layer, power, and index of the matrix. They detected OOD samples using the stored values. However, it required heavy computation and a longer time than other approaches. Those approaches used additional InD validation dataset for tuning hyperparameters: Hsu et al. [12] tuned the perturbation size of input processing, and Sastry and Oore [23] tuned normalization factors for each layer. + +# 2.4. Using Input Processing + +The input processing is an operation that generates $x'$ by adding a small perturbation to an input $x$ . Many previous approaches applied the input processing because this process effectively improved OOD detection performance. However, the perturbation size, epsilon, needed to be tuned appropriately. The epsilon has a significant impact on performance, so it should be chosen carefully. + +Liang et al. [16] and Lee et al. [15] chose the epsilon values using given specific OOD samples. The chosen epsilon may not be valid to other OOD samples because it was tuned with a distribution of given OOD samples. Hsu et al. [12] proposed a method to set the epsilon value without OOD samples. They tuned the epsilon using the InD validation dataset, but their method did not generally work for all cases. The detection performance decreased in some datasets. + +Another disadvantage is that the input processing requires a lot of computation. We need one forward loop and one backward loop to add a perturbation to an input. We need one more forward loop to determine whether the perturbed input is OOD. Since the time taken for the forward loop and the backward loop is almost the same, the total inference time increases by about three times compared to the inference time without the input processing. + +# 2.5. Using Self-supervised Learning + +Recently, a self-supervised learning technique that can learn a deep learning model without a label is attracting attention. In many studies in the field of OOD detection, self-supervised learning techniques are being tried to improve performance or reduce dependence on data labels. Some approaches build new data by applying transforms such as rotation to the image, and train models by giving the transform applied to the image as a label [7, 11, 27, 1]. Other approaches use contrastive learning techniques to train the model [27, 24]. However, in most cases, the OOD detection performance is not as high as when learning using labels. In addition, since the model is trained only for the purpose of OOD detection or representation learning, an additional method is needed to perform classification. + +# 3. Proposed Method + +A deep learning-based classification model can be decomposed into a feature extractor $f(\cdot)$ and a classifier $g(\cdot)$ which is usually a fully connected layer. When training with the standard softmax loss, a logit is defined as: + +$$ +\operatorname {l o g i t} (x, c) = f (x) \cdot w _ {c} + b _ {c}, \tag {1} +$$ + +and with the cosine-based softmax loss, it is defined as: + +$$ +\operatorname {l o g i t} (x, c) = \frac {f (x) \cdot w _ {c}}{| f (x) | \cdot | w _ {c} |}, \tag {2} +$$ + +![](images/825b49fc37e084e1c1ab0f01121b7cec8973b23fd1c07afb0f86c1efd4c238d6.jpg) +(a) Standard softmax loss + +![](images/6d7b815eb1d20eb07273025a433b8073fe90b6e0ea7619058c6bed45b3416c3c.jpg) +(b) Cosine-based softmax loss +Figure 1: The feature norm distributions of InD (red) and OOD (blue) data. (a) is from DenseNet3 trained with the standard softmax loss and (b) is from DenseNet3 trained with the cosine-based softmax loss. + +where $w_{c}$ is the connection weights of $g(\cdot)$ for a class $c$ . The cosine-based softmax loss is widely used in face recognition field [28, 5, 29]. + +We introduce two useful measures for OOD detection that can be simply extracted from models trained with cosine-based softmax loss: a feature norm-based measure and a Mahalanobis distance-based measure. We then propose an OOD score by combining these two measures. Let us describe each measure and the score function in detail. + +# 3.1. Cosine-based Softmax Loss and Feature Norm + +When we train a classification model with the closed-world assumption, the logit value for a class $c$ is considered as being proportional to $P(y = c|x)$ . However, in the open world, the logit value should be interpreted as being proportional to $P(y = c|x, D_{in})$ where $D_{in}$ is the event that $x$ is in the train distribution. The probability can be decomposed using $P(y = c, D_{in}|x)$ and $P(D_{in}|x)$ as follows: + +$$ +\operatorname {l o g i t} (x, c) \propto P (y = c | x, D _ {i n}) = \frac {P (y = c , D _ {i n} | x)}{P \left(D _ {i n} | x\right)}. \tag {3} +$$ + +If we use the cosine-based softmax loss, Equations 2 and 3 can be combined as follows: + +$$ +\frac {f (x) \cdot w _ {c}}{| f (x) | \cdot | w _ {c} |} \propto \frac {P (y = c , D _ {i n} | x)}{P \left(D _ {i n} | x\right)}. \tag {4} +$$ + +The numerator of the right-hand side includes a class-related term, $c$ , and the denominator, $P(D_{in}|x)$ , does not. So, we reformulate the left-hand side so that only the numerator has class-related terms as follows: + +$$ +\frac {f (x) \cdot w _ {c} / | w _ {c} |}{| f (x) |} \propto \frac {P (y = c , D _ {i n} | x)}{P \left(D _ {i n} | x\right)}. \tag {5} +$$ + +Since $w_{c}$ is the connection weight to $g(\cdot)$ , we can say it is class-related, and since $f(x)$ is the feature representation of $x$ , we can say it is classifier-independent. Based on this, we rigorously infer that both denominators can be proportional + +![](images/c0b5afe3ce518fab97a13dc7f710470d428577900820d4dcd6cdb05ec1f732a0.jpg) +(a) Distribution of $|f(x)|$ + +![](images/ac6384f326f9c55319542164ee2e17d973fc31da3561b548d0c6df0194472bb5.jpg) +(b) Mean of $|f(x)|$ + +to each, i.e., $|f(x)| \propto P(D_{in}|x)$ , because the two terms in Equation 5 are proportional, and their denominators and numerators share class-related characteristics. + +To confirm this, we perform some OOD detection experiments. We try to detect SVHN using $|f(x)|$ from DenseNet3 model trained for CIFAR-10 with the cosine-based softmax loss. We compare this to the model using the standard softmax loss as shown in Figure 1. The feature norm distributions of OOD and InD by the standard loss model overlap much, but those of the cosine loss model have a very small overlap between SVHN and CIFAR-10, which says that the cosine loss model can be better to discriminate ODD than the standard model. + +To verify that the length of feature norm decrease as the samples get closer to OOD, we create mixup samples by mixing SVHN and CIFAR-10 images as shown in Equation 6 and observed the feature norm. + +$$ +\operatorname {T e s t I m g} = \lambda \cdot \operatorname {I n D} + (1 - \lambda) \cdot \operatorname {O O D} \tag {6} +$$ + +As the mixup factor, $\lambda$ decreases, that is, as the samples get closer to OOD, the feature norm also decreases as shown in Figure 2. The feature norm of an image from the cosine loss model tends to increase as the image gets closer to InD samples and decreases closer to OOD samples, which supports our conclusion that $|f(x)|$ from a cosine loss model is proportional to $P(D_{in}|x)$ . Also, it supports that the feature norm can be a measure for OOD detection. Additional observations, including CIFAR-100 and SVHN, are presented in the appendix. + +# 3.2. Mahalanobis Distance in Models Trained with Cosine-based Softmax Loss + +We may assume that features of InD samples obtained from internal layers of classification models follow the Gaussian distribution. We adopt an OOD score defined by + +![](images/2913a3b03bd1ba529efe9fadfd27919720bf4feb1ab15c505f627377bd3b4b7d.jpg) +(a) 1st Block (SVHN vs. TINc) + +![](images/1540a4e03e4294be16f53776a9a35fa65b03108b1140bc6176a75ff781ed0d41.jpg) +(b) 1st Block (SVHN vs. iSUN) + +![](images/d1ba4001561006a0b2c01e633b46a4f109f7fb8f790596caa1d3f7efaaffafd8.jpg) +Figure 2: The distribution of feature norm according to the mixup factor (a), and the mean value of feature norm according to the mixup factor. It can be observed that as OOD samples are more mixed into a sample, the feature norm tends to decrease. + +![](images/41c6e10e01943d5d85d3d56f8dc6107449ca80a3b495e64431d9bbb25d32b8da.jpg) +(c) 2nd Block (SVHN vs. TINc) +(d) 2nd Block (SVHN vs. iSUN) + +![](images/a78bc1dab61379cc9f038bd0b47dc810ac1d233d091151ae70a4c6921a28f07f.jpg) +(e) 3rd Block (SVHN vs. TINc) + +![](images/ea705c855111256c1c3523f9ed65914eef230856b758dc977ce8244b7ca4b046.jpg) +(f) 3rd Block (SVHN vs. iSUN) +Figure 3: Distributions of $MD$ scores from each block of DenseNet3 trained with SVHN. The OODs datasets (blue) is TinyImageNet-crop (left) and iSUN (right). + +Lee et al. [15] based on the Mahalanobis distance: + +$$ +\begin{array}{l} M D _ {B} (x) \\ = - \max _ {c} - \left(f _ {B} (x) - \mu_ {c, B}\right) ^ {\top} \Sigma_ {B} ^ {- 1} \left(f _ {B} (x) - \mu_ {c, B}\right), \quad (7) \\ \end{array} +$$ + +where $\mu_{c,B}$ and $\Sigma_B$ are the mean and the covariance of features from a block $B$ . This score is very discriminative for OOD detection as shown in Figure 3. Since the distributions of $MD$ scores of InD and OOD samples have small overlaps, we can detect OOD samples using the scores. However, $MD$ scores from the penultimate layer overlap if the model is trained with cosine-based softmax loss, as shown in Figure 2(e)-(f). + +The reason why $MD$ scores from penultimate layer is not discriminative is that the feature from the penultimate layer do not follow a Gaussian distribution. In the model trained with cosine-based softmax loss, the softmax probability with $\sigma_c$ , the softmax value for a class $c$ , is calculated as follows: + +$$ +\sigma_ {c} (g (f (x))) = \frac {\exp \left(w _ {c} \cdot f (x) / \left(\left| w _ {c} \right| \cdot | f (x) |\right)\right)}{\sum_ {c ^ {\prime}} \exp \left(w _ {c ^ {\prime}} \cdot f (x) / \left(\left| w _ {c ^ {\prime}} \right| \cdot | f (x) |\right)\right)}. \tag {8} +$$ + +
ApproachAuxiliary Dataset# of Hyperp.Input Processing
OE [10]Large OOD dataset0×
ODIN [16]OOD samples2✓
MAHA [15]OOD samples4✓
G-ODIN [12]InD samples1✓
GRAM [23]InD samples100×
COD (Proposed)None0×
+ +Table 1: OOD detection approach comparison. The number of hyperparameters is counted based on DenseNet3 with 100 layers. + +The classification model is trained so that the softmax probability reflects $P(y = c|f(x))$ . If the feature distribution from the penultimate layer is assumed to be the multivariate Gaussian distribution for each class, and the class prior is $P(y = c) = \frac{\beta_c}{\sum_{c'} \beta_{c'}}$ where $\beta_c$ is unnormalized prior for class $c$ , the probability of class $c$ with $f(x)$ is as follows: + +$$ +\begin{array}{l} P (y = c \mid f (x)) \\ = \frac {P (y = c) P (f (x) | y = c)}{\sum_ {c ^ {\prime}} P (y = c ^ {\prime}) P (f (x) | y = c ^ {\prime})} \\ = \frac {\exp \left(\mu_ {c} ^ {T} \Sigma^ {- 1} f (x) - \frac {1}{2} \mu_ {c} ^ {T} \Sigma^ {- 1} \mu_ {c} + \log \beta_ {c}\right)}{\sum_ {c ^ {\prime}} \exp \left(\mu_ {c ^ {\prime}} ^ {T} \Sigma^ {- 1} f (x) - \frac {1}{2} \mu_ {c ^ {\prime}} ^ {T} \Sigma^ {- 1} \mu_ {c ^ {\prime}} + \log \beta_ {c ^ {\prime}}\right)}. \tag {9} \\ \end{array} +$$ + +That is, if the feature from the penultimate layer follows a Gaussian distribution, $P(y = c|f(x))$ is based on linear expressions. Since a nonlinear expression cannot be transformed in a linear expression, the softmax probability from the cosine loss model (Equation 8) cannot be converted to a similar form of $P(y = c|f(x))$ (Equation 9), which means that the features from the penultimate layer is not expected to follow a Gaussian distribution. + +If Equation 8 is rewritten using a normalized feature $f^{\prime}(x) = f(x) / |f(x)|$ , it is as follows: + +$$ +\sigma_ {c} (g (f (x))) = \frac {\exp \left(w _ {c} ^ {\top} f ^ {\prime} (x) / \left(\left| w _ {c} ^ {\top} \right|\right)\right)}{\sum_ {c ^ {\prime}} \exp \left(w _ {c ^ {\prime}} ^ {\top} f ^ {\prime} (x) / \left(\left| w _ {c ^ {\prime}} ^ {\top} \right|\right)\right)}. \tag {10} +$$ + +Equation 10 is based on linear expressions of $f'(x)$ , from which we conjecture that $f'(x)$ may follow a Gaussian distribution. Since $MD$ scores are useful under the assumption that features follow a Gaussian distribution, we redefine the feature for $MD$ from the penultimate layer as $f'(x) = f(x) / |f(x)|$ instead of $f(x)$ . We re-depict the distribution of $MD$ in Figures 2-(e) and (f) using $f'(x)$ as shown in Figure 4. The overlap of the distributions is significantly reduced compared to $MD$ based on $f(x)$ , which means that $MD$ based on $f'(x)$ may be better than $f(x)$ for OOD detection. + +![](images/069b255e7d1cf33cf84628fa72a425a26ef0f52c098a564b504635a0b50eb375.jpg) +(a) 3rd Block (SVHN vs TINc) + +![](images/d6c43b08bec461609b8de4b457afbce9907805edb48808435420b343c1967626.jpg) +(b) 3rd Block (SVHN vs iSUN) +Figure 4: Distributions of $MD$ scores based on $f^{\prime}(x)$ . The configuration is the same as Figure 3. + +We evaluate $MD$ scores for the penultimate layer with the redefined feature, $f'(x) = f(x) / |f(x)|$ and for other layers with $f_B(x)$ . Finally, we combine $MD$ scores from each block as Equation 11 by product: + +$$ +M D (x) = \prod_ {B} \left(M D _ {B} (x)\right). \tag {11} +$$ + +# 3.3. Proposed OOD Score Function + +We define a cosine-based OOD score function, $COD(x)$ , with the feature norm and the Mahalanobis distance-based measure, $MD$ . Since the feature norm tends to appear large for InD samples and $MD$ tends to be large for OOD samples, we define $COD(x)$ by dividing the feature norm by $MD$ as follows: + +$$ +C O D (x) = \frac {| f (x) |}{M D (x)}. \tag {12} +$$ + +$COD(x)$ does not have any hyperparameter to be tuned. Also, no additional dataset is required to use $COD(x)$ , and it shows a high discriminative performance without the input processing. $COD(x)$ can be calculated with a very small computation in a forward loop for classification. We compare the previous approaches in the viewpoint of auxiliary datasets, the number of hyperparameters and the input processing in Table 1. Most of the previous methods used auxiliary datasets, and had several to one hundred hyperparameters, and needed the input processing to increase the performance, while our method does not need any of them. + +# 4. Experiment + +This section shows the effectiveness of our approach through experiments with various image datasets. + +# 4.1. Experiment setting + +Datasets: We use CIFAR-10/100 [14], and SVHN [20] images for InD datasets. We use the training dataset for classification and use the test datasets to evaluate OOD detection. We use TinyImageNet-crop (TINc), TinyImageNet-resize (TINr), LSUN-crop (LSUNc), LSUN-resize (LSUNr), and iSUN as OOD datasets. These datasets have been used in many studies related to OOD detection. Additionally, when CIFAR-10/100 are in-distribution, SVHN test dataset is also used as an OOD dataset, and when SVHN image is in-distribution, CIFAR-10/100 test datasets are also used as OOD datasets. + +Baselines: We compare our approach to previous OOD detection methods: ODIN [16], MAHA [15], G-ODIN [12], and GRAM [23]. + +Networks and Training Details: We train ResNet34 [8] and DenseNet3 [13] models with the same setting as the baseline approaches. We train ResNet34 for 200 epochs with a batch size of 128 and a weight decay of 0.0005. We train DenseNet3 with 100 layers for 300 epochs with a batch size of 64 and a weight decay of 0.0001. In both training, the optimizer is SGD with a momentum of 0.9, and the learning rate starts with 0.1 and decays by a factor 0.1 at $50\%$ and $75\%$ of the training progress. + +When we train models with the cosine-based softmax loss, logit values are between -1 and 1. For expanding the range of logits, we adopt the temperature scaling with a factor of $T$ as in Equation 13. The factor of $T$ needs to be set to 0.1 for SVHN, CIFAR-10 and 0.05 for CIFAR-100 so that the classification performance is maximized. The logit scaled by $T$ is defined as: + +$$ +\operatorname {l o g i t} _ {T} (x) = \operatorname {l o g i t} (x) / T. \tag {13} +$$ + +Evaluation Metrics: We evaluate the detection performance using the two metrics that are most widely used for OOD detection. The first metric is the area under the receiver operating characteristic curve (AUROC). It is used to evaluate the average performance considering various thresholds. The second is TNR at TPR $p\%$ . We evaluate with $p = 95$ as the metric used in the baseline approaches. + +# 4.2. Results + +Table 2 shows an overall comparison with DenseNet3. Our approach achieves superior performance in all the datasets. Compared to ODIN, our approach has higher performance in all the cases. Compared to G-ODIN, our approach has higher performance except one cases. Com + +pared to MAHA, our approach has higher performance except some cases. + +In some cases, GRAM shows a higher performance than the proposed method, but GRAM is very unstable depending on OOD datasets. For example, GRAM shows a particularly poor performance on LSUNc. It shows a TNR@95 score of 88.4 for CIFAR-10 and 65.6 for CIFAR-100. Our approach is more even performances across all OOD datasets, and achieves the highest average performances. Also, GRAM requires significantly heavy computation. When comparing the wall clock time in the same environment (GTX 2080 Ti GPU with Xeon Silver 4021 CPU * 2), the GRAM takes 0.35 seconds to calculate OOD score for one image, but our approach takes only about 0.018 seconds. Our approach is 19 times faster than GRAM because GRAM evaluates a lot of measures in the forward loop while our approach needs very small additional computation. + +The head of Table 2 shows what each approach needs to detect OOD samples. ODIN and MAHA use OOD samples and InD validation samples, and perform the input processing. G-ODIN does not use OOD samples but uses InD validation samples and performs the input processing. GRAM uses InD validation samples to tune hyperparameters. Our approach needs nothing except a training dataset. Our approach is the fastest method because the input processing is not required. Our method is superior to or at least comparable to state-of-the-art OOD detection methods even though it does not require any additional datasets. Our model also shows better performance with ResNet34. The results are presented in the appendix. + +# 5. Discussion + +# 5.1. Layer Selection for $MD$ + +Our approach, $COD$ , uses $MD$ values from all the blocks of the model. However, it can be more effective if we can choose some layers which are helpful for OOD detection. Useful layers for OOD detection may depend on the model structure, the training dataset, and the OOD dataset. However, OOD samples are not available in advance. + +We will discuss how to select useful layers with synthetic OOD samples. We may generate synthetic OOD samples from the training dataset. For example, we may use some training samples rotated by 90 degrees as synthetic OOD samples. Then, we evaluate AUROC scores for each layer, and choose layers of which AUROC scores are greater than a threshold such as $60\%$ of AUROC. In the case of CIFAR10, CIFAR100, and SVHN, it is confirmed that the AUROC score of each $MD$ was over $60\%$ . Therefore, it can be a reasonable threshold to choose layers. + +We apply the layer selection to a more realistic dataset, ImageNet-30. In this case, $MD$ from the last layer, $MD$ + +
OODAUROCTNR@95
Use a training datasetUse a training dataset
Use InD validation samples(ours)Use InD validation samples(ours)
Perform the input processingGRAMPerform the input processingGRAMCOD
Use OOD samplesG-ODINUse OOD samplesG-ODIN
ODINMAHAODINMAHA
InD: CIFAR-10
SVHN95.598.198.899.199.586.290.894.096.198.0
TINc97.695.398.799.399.487.084.293.496.797.9
TINr98.598.899.199.799.592.495.095.898.897.8
LSUNc93.680.298.397.599.970.648.291.588.499.6
LSUNr99.299.399.499.999.696.297.297.699.598.9
iSUN98.798.999.499.899.693.295.397.599.098.7
MEAN97.295.199.099.299.687.685.195.096.498.5
InD: CIFAR-100
SVHN93.897.295.997.397.070.682.577.089.382.4
TINc88.388.897.697.797.951.060.187.889.089.3
TINr85.297.498.699.098.642.686.693.395.793.1
LSUNc91.481.795.391.498.957.842.175.065.595.1
LSUNr85.598.098.799.398.641.291.493.897.293.3
iSUN84.597.498.499.098.537.487.092.595.992.5
MEAN88.193.497.497.398.350.175.086.688.891.0
InD: SVHN
CIFAR-1091.498.9-95.599.071.796.8-80.496.1
CIFAR-100----98.9----95.8
TINc----99.8----99.7
TINr95.199.9-99.799.884.199.9-99.199.5
LSUNc----99.8----99.4
LSUNr94.599.9-99.899.881.199.9-99.599.4
iSUN94.799.9-99.899.882.299.9-99.499.5
MEAN----99.6----98.5
+ +Table 2: OOD detection performance (%) with a DenseNet3 classification model. Our performance is averaged over 5 runs. ODIN, MAHA and GRAM are cited from the work by Sastry and Oore [23] and G-ODIN is cited from the work by Hsu et al. [12]. In the case of SVHN, all baselines did not present their performances. Dash symbols stand for not-available. The mean rows are the averages of the performances over all OOD datasets. For each case, the best results are in bold. + +
ResNet18 +InD: ImageNet-30Test AccOOD Dataset
DTDCUB-200CaltechDogsPlaces-365Food-101PetFlowers
CSI97.093.793.491.997.792.587.096.996.0
COD97.294.092.891.895.993.780.496.295.4
+ +Table 3: OOD detection performance comparison with CSI and COD with layer selection. + +obtained from the re-defined feature of the last layer, shows an AUROC score of over $80\%$ , and the remaining blocks show AUROC scores of about $50\%$ . Based on this, we choose only the last layer for $COD$ , and perform OOD detection for ImageNet-30. ResNet18 is used for comparison + +with other approach. The baseline is a contrastive learning-based model, CSI [27], one of SOTA models. CSI is heavier to apply because it use a specific training method for OOD detection such as input transform and contrastive learning. On the other hand, $COD$ is simple and easy to apply be + +cause it does not require anything but generating classifiers in the training process. Nevertheless, as shown in Table 3, $COD$ shows comparable performance to CSI. + +# 5.2. Classification Accuracy + +Some approaches for OOD detection suffered from the degradation of classification performance by change of model structures or loss functions [10, 21, 12]. In this case, two models should be stored aside, one for classification and the other for OOD detection. OOD detection models also need to have high classification performance for efficiency. Table 4 shows that models trained with the cosine-based softmax loss are accurate as much as the standard loss. Our approach can be applied to various model architectures because it only changes the calculation of logits. Also, there is no disadvantage to classification performance. + +
modelInDStandardCosine
ResNet34CIFAR-1095.2595.34
CIFAR-10077.7077.78
SVHN96.4896.41
DenseNet3CIFAR-1095.1495.08
CIFAR-10076.4376.87
SVHN96.4296.33
+ +Table 4: Comparison of classification performance + +# 5.3. Is the Input Processing Helpful? + +The input processing is helpful for performance improvement when the perturbation size is properly chosen. We analyze the performance improvement by the perturbation size of the input processing. We train DenseNet3 with CIFAR-100 and observe the AUROC improvements with different perturbations for SVHN, LSUNc, and iSUN datasets. We perform ODIN and MAHA with pre-trained DenseNet3 with CIFAR-100. Figure 5 shows the improvement by the input processing on the three method. + +Each dataset has a different pattern of improvement. For example, our method is improved for SVHN but degraded for iSUN when the epsilon is large. ODIN and MAHA also have similar pattern to our method. This means that it is challenging to find the perturbation size that generally increases the performance regardless of OOD datasets. Hsu et al. tried to find the perturbation size without using OOD datasets, but the performance decreased in some cases [12]. + +The input processing is helpful where the distribution of OOD samples are known. However, in reality, OOD samples can be generated from any distribution. The input processing significantly increases the amount of computation. It is reasonable not to use input processing, which increases + +![](images/dc3048eef10c4b0a757c135fbe3c4743a0ef63cc441400a3ce133a54a33a3ffc.jpg) + +![](images/121e7516251752cea139f330627e2857ccda030f12714d37d457f382f6ef0d48.jpg) +(a)ODIN + +![](images/5c2e22dd6992980ce4b832c6cf3d92b9197357fa5ef62f1baa5728fa8e8eabec.jpg) +(b) MAHA +(c) Proposed Method +Figure 5: OOD detection performance (AUROC) increment by the perturbation size. + +computation cost and has a potential to decrease performance. + +# 5.4. Fine-tuning with Pre-trained Model + +We verify that our approach is also effective to fine-tuned models. We first train a model with the standard softmax, and then fine-tune the last block of the feature extractor and the classifier for only 20 epochs. In the fine-tuning process, all settings except for the learning rate are kept consistent with the previous configuration. The learning rate for fine-tuning starts with 0.5 and decays by a factor 0.1 at $50\%$ and $75\%$ of the training progress. + +We can notify that our approach still works for fine-tuned models. The OOD detection performances are comparable to or even higher than those of our original models, as shown in Table 5. The DenseNet3 trained with the standard softmax for SVHN shows a classification accuracy of $96.4\%$ , but its fine-tuned model shows an accuracy of + +
InDOODAUROCTNR95
CIFAR-10 (Acc: 95.1 → 94.1)SVHN99.4598.35
TINc99.5399.61
TINr99.7399.23
LSUNc99.8299.96
LSUNr99.7399.34
iSUN99.7299.27
CIFAR-100 (Acc: 76.4 → 75.7)SVHN96.3979.48
TINc97.5388.87
TINr98.0389.37
LSUNc98.7296.81
LSUNr98.1690.91
iSUN97.9988.76
SVHN (Acc: 96.4 → 96.1)CIFAR-1099.1798.18
CIFAR-10099.2298.33
TINc99.94100.00
TINr99.8899.96
LSUNc99.94100.00
LSUNr99.91100.00
iSUN99.9199.99
+ +Table 5: Classification and OOD detection performance of fine-tuned DenseNet3 models. All figures are in %, and are the average results of 5 experiments each. + +$96.1\%$ . Even though the models are fine-tuned with a limited number of epochs, the classification accuracy and the OOD detection performances remain almost same. The results with ResNet34 are also presented in the appendix. + +# 6. Conclusion + +We proposed a novel OOD detection method that can be applied to deep neural network-based classifiers. It is simple and fast since it uses only the feature norm and the Mahalanobis distances. Our approach does not use additional data, does not have hyperparameters, and does not perform input processing. Our approach is easy to apply because it only needs to apply the cosine-based softmax loss. + +# Acknowledgements + +This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-00421, AI Graduate School Support Program (Sungkyunkwan University)), and also supported by the MSIT under the ICT Creative Consilience Program (IITP-2023-2020-0-01821) supervised by the IITP. + +# References + +[1] Liron Bergman and Yedid Hoshen. Classification-based anomaly detection for general data. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 3 +[2] Wenhu Chen, Yilin Shen, Hongxia Jin, and William Wang. A variational dirichlet framework for out-of-distribution detection. arXiv preprint arXiv:1811.07308, 2018. 2 +[3] Sung-Ik Choi and Sae-Young Chung. Novelty detection via blurring. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 2 +[4] Erik Daxberger and José Miguel Hernández-Lobato. Bayesian variational autoencoders for unsupervised out-of-distribution detection. arXiv preprint arXiv:1912.05651, 2019. 2 +[5] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 4690-4699. Computer Vision Foundation / IEEE, 2019. 3 +[6] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016, volume 48 of JMLR Workshop and Conference Proceedings, pages 1050-1059. JMLR.org, 2016. 2 +[7] Izhak Golan and Ran El-Yaniv. Deep anomaly detection using geometric transformations. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada, pages 9781-9791, 2018. 3 +[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society, 2016. 6 +[9] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. 1, 2 +[10] Dan Hendrycks, Mantas Mazeika, and Thomas G. Dietterich. Deep anomaly detection with outlier exposure. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. 1, 2, 5, 8 +[11] Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self-supervised learning can improve model robustness and uncertainty. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS + +2019, December 8-14, 2019, Vancouver, BC, Canada, pages 15637-15648, 2019. 3 +[12] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized ODIN: detecting out-of-distribution image without learning from out-of-distribution data. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 10948-10957. IEEE, 2020. 1, 2, 3, 5, 6, 7, 8 +[13] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 2261-2269. IEEE Computer Society, 2017. 6 +[14] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, 2009. 6 +[15] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada, pages 7167-7177, 2018. 1, 2, 3, 4, 5, 6 +[16] Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net, 2018. 1, 2, 3, 5, 6 +[17] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In Advances in Neural Information Processing Systems, volume 33, pages 21464-21475. Curran Associates, Inc., 2020. 2 +[18] Andrey Malinin and Mark J. F. Gales. Predictive uncertainty estimation via prior networks. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada, pages 7047-7058, 2018. 2 +[19] Eric Nalisnick, Akihiro Matsukawa, Yee Whye Teh, and Balaji Lakshminarayanan. Detecting out-of-distribution inputs to deep generative models using a test for typicality. arXiv preprint arXiv:1906.02994, 5:5, 2019. 2 +[20] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 6 +[21] Aristotelis-Angelos Papadopoulos, Mohammad Reza Rajati, Nazim Shaikh, and Jiamian Wang. Outlier exposure with confidence control for out-of-distribution detection. Neurocomputing, 441:138-150, 2021. 1, 2, 8 +[22] Jie Ren, Peter J. Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark A. DePristo, Joshua V. Dillon, and Balaji Lakshminarayanan. Likelihood ratios for out-of-distribution detection. In Advances in Neural Information Processing Sys- + +tems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 14680-14691, 2019. 2 +[23] Chandramouli Shama Sastry and Sageev Oore. Detecting out-of-distribution examples with gram matrices. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 8491-8501. PMLR, 2020. 1, 2, 5, 6, 7 +[24] Vikash Sehwag, Mung Chiang, and Prateek Mittal. $\{\mathrm{SSD}\}$ : A unified framework for self-supervised outlier detection. In International Conference on Learning Representations, 2021. 3 +[25] Joan Serrà, David Álvarez, Vicenç Gómez, Olga Slizovskaia, José F. Núñez, and Jordi Luque. Input complexity and out-of-distribution detection with likelihood-based generative models. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 2 +[26] Jiaming Song, Yang Song, and Stefano Ermon. Unsupervised out-of-distribution detection with batch normalization. arXiv preprint arXiv:1910.09115, 2019. 2 +[27] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. In Advances in Neural Information Processing Systems, volume 33, pages 11839-11852. Curran Associates, Inc., 2020. 3, 7 +[28] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 5265-5274. IEEE Computer Society, 2018. 3 +[29] Xiao Zhang, Rui Zhao, Yu Qiao, Xiaogang Wang, and Hongsheng Li. Adacos: Adaptively scaling cosine logits for effectively learning deep face representations. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 10823-10832. Computer Vision Foundation / IEEE, 2019. 3 \ No newline at end of file diff --git a/2023/Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss/images.zip b/2023/Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..149a7aac6907f1ee90bc3417ee06c6bed637ec41 --- /dev/null +++ b/2023/Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d43115b40713eb20f4767271fdd9ac48d145756f35233fec353b2ebf349535b2 +size 597487 diff --git a/2023/Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss/layout.json b/2023/Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7d3047ed171c37a56a21934fcb496ce7537d3f08 --- /dev/null +++ b/2023/Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss/layout.json @@ -0,0 +1,8936 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 139, + 103, + 455, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 103, + 455, + 138 + ], + "spans": [ + { + "bbox": [ + 139, + 103, + 455, + 138 + ], + "type": "text", + "content": "Simple and Effective Out-of-Distribution Detection via Cosine-based Softmax Loss" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 155, + 161, + 442, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 161, + 442, + 175 + ], + "spans": [ + { + "bbox": [ + 155, + 161, + 442, + 175 + ], + "type": "text", + "content": "SoonCheol Noh DongEon Jeong Jee-Hyong Lee*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 231, + 176, + 361, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 176, + 361, + 190 + ], + "spans": [ + { + "bbox": [ + 231, + 176, + 361, + 190 + ], + "type": "text", + "content": "Sungkyunkwan University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 261, + 190, + 332, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 190, + 332, + 202 + ], + "spans": [ + { + "bbox": [ + 261, + 190, + 332, + 202 + ], + "type": "text", + "content": "Suwon, Korea" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 200, + 205, + 394, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 205, + 394, + 217 + ], + "spans": [ + { + "bbox": [ + 200, + 205, + 394, + 217 + ], + "type": "text", + "content": "{yhnuhb27, wjdehddjs, john}@skku.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 270, + 289, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 289, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 289, + 486 + ], + "type": "text", + "content": "Deep learning models need to detect out-of-distribution (OOD) data in the inference stage because they are trained to estimate the train distribution and infer the data sampled from the distribution. Many methods have been proposed, but they have some limitations, such as requiring additional data, input processing, or high computational cost. Moreover, most methods have hyperparameters to be set by users, which have a significant impact on the detection rate. We propose a simple and effective OOD detection method by combining the feature norm and the Mahalanobis distance obtained from classification models trained with the cosine-based softmax loss. Our method is practical because it does not use additional data for training, is about three times faster when inferencing than the methods using the input processing, and is easy to apply because it does not have any hyperparameters for OOD detection. We confirm that our method is superior to or at least comparable to state-of-the-art OOD detection methods through the experiments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 520, + 128, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 128, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 128, + 533 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "type": "text", + "content": "Deep learning models are generally designed to target the closed world. Models are trained based on the assumption that the inputs of the inference stage will be sampled from the same distribution as the train distribution. This means that the model may not make correct predictions if inputs are sampled from other distributions. The model must be able to distinguish whether the input is from the train distribution or not. This problem has been established as the task of out-of-distribution (OOD) detection." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 647, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 647, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 647, + 287, + 696 + ], + "type": "text", + "content": "Some researchers tried to detect OOD samples using confidence scores from classification models. Hendrycks and Gimpel simply used the confidence scores, but this approach was not very effective because of the overconfi" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 246, + 546, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 246, + 546, + 377 + ], + "spans": [ + { + "bbox": [ + 306, + 246, + 546, + 377 + ], + "type": "text", + "content": "dence of deep learning models [9]. Hendrycks, Mazeika, and Dietterich (2019) and Papadopoulos et al. tried to lower the confidence of OOD samples using auxiliary large OOD datasets [10, 21]. However, collecting a massive OOD dataset is highly expensive, and training classification models with such a dataset takes a long time. Liang, Li and Srikant (2018) simply adopted the input processing technique, which added a small perturbation to the input samples to increase the confidence of in-distribution (InD) samples, but they assumed OOD samples were given in advance, which is not practical [16]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 378, + 546, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 546, + 510 + ], + "type": "text", + "content": "Some researchers defined measures to detect OOD samples based on input features learned by classification models. Lee et al. proposed a measure based on Mahalanobis distance, but they also assumed that OOD samples were given in advance as Liang et al. did [15, 16]. Sastry and Oore proposed a measure based on various correlations between features, but their approach was very time-consuming in evaluating the OOD measure [23]. Hsu et al. tried to obtain the OOD probability by modifying the network structure, which caused the degradation of classification performances [12]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 510, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 629 + ], + "type": "text", + "content": "Additional popular technique is the input processing, which was adopted by most approaches [16, 15, 23]. The input processing is to generate " + }, + { + "bbox": [ + 304, + 510, + 547, + 629 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 304, + 510, + 547, + 629 + ], + "type": "text", + "content": " by adding a small perturbation to input data " + }, + { + "bbox": [ + 304, + 510, + 547, + 629 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 510, + 547, + 629 + ], + "type": "text", + "content": " to make the boundary of in-distribution compact. It may be helpful to increase the detection accuracy under certain conditions. However, generating the perturbation takes almost twice of the inference time, so the total inference becomes three times longer. The OOD detection performance is very susceptible to the size of the perturbation, so it should be carefully determined." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "type": "text", + "content": "Some OOD detection approaches had various hyperparameters, including the perturbation size of input processing. The hyperparameters were usually tuned by using additional datasets to maximize the detection performance. Liang et al. [16] and Lee et al. [15] tuned them with given OOD samples, and Hsu et al. [12] and Sastry and Oore [23] used the validation samples. However, it is hard to find out" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16560" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "type": "text", + "content": "hyperparameter values generally applicable to any distributions of OOD samples. When the values are found in a specific dataset, the model is easy to be biased to the dataset." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 286, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 286, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 286, + 286 + ], + "type": "text", + "content": "This paper proposes a simple and effective approach that can detect OOD samples without hyperparameters, additional datasets, and input processing. We detect OOD samples by combining norm and Mahalanobis distances of features. We find that the models trained with the cosine-based softmax loss are more advantageous than the ones trained with the regular softmax loss for OOD detection. We mathematically conjecture that the feature norm from the models with the cosine-based softmax loss can be a measure for the probability of in-distribution, and confirm its effectiveness experimentally. We also develop a robust Mahalanobis distance-based measure for the model with the cosine-based softmax loss. We propose an OOD score function by combining the feature norm and the Mahalanobis distance without any hyperparmaeters." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 287, + 286, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 286, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 286, + 489 + ], + "type": "text", + "content": "Our method do not need the input processing because it shows high accuracy in OOD detection without it. So, the additional computation for OOD detection by our method is very small. Feature norms and the Mahalanobis distances can be obtained with tiny additional calculations from classification models. Since the cosine-based softmax loss is generally applicable to any deep models, our OOD detection approach can be easily extended to various structures without a classification accuracy loss. Our method is also practical because it does not use any additional datasets. We confirm the OOD detection performance on various image datasets through experiments. Our method is compared with others using additional datasets, performing the input processing, or tuning hyperparameters to maximize performance. The experiment shows that our approach is superior to or at least comparable to state-of-the-art OOD detection methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 501, + 134, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 134, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 134, + 513 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 521, + 286, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 286, + 567 + ], + "type": "text", + "content": "Previous OOD detection approaches were based on various models, such as classification models [9, 16, 15, 10, 21, 17, 12, 23], generative models [22, 19, 25, 4, 26, 3] and Bayesian models [6, 2, 18]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 568, + 286, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 286, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 286, + 651 + ], + "type": "text", + "content": "OOD detection is usually required as a pre-processing of the classification task, which is one of the crucial tasks of deep learning. Due to its effectiveness, many researchers tried simultaneously performing OOD detection and classification in one model. In this section, we cover OOD detection-related approaches based on classification models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 659, + 234, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 234, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 234, + 672 + ], + "type": "text", + "content": "2.1. Training with Large OOD Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "Hendrycks et al. proposed a method to lower the maximum softmax probability (MSP) for OOD samples from classification models [10]. They improved OOD detection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "performance by adding a loss term which was based on large OOD datasets. The added loss term was designed so that softmax values for OOD samples be closed to a uniform distribution. Papadopoulos et al. further improved OOD detection performance by adding another loss term [21]. The loss term prevented classification models from overconfidence. Liu et al. proposed an energy-based measure for models trained with large OOD datasets [17]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 171, + 545, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 545, + 255 + ], + "type": "text", + "content": "These methods had the disadvantage of collecting extensive additional datasets to train classification models. The collected OOD dataset was used, assuming it represented all unknown OODs. However, it is virtually impossible to collect data that can represent all OOD samples in real-world situations. In addition, training a model with a large amount of OOD data takes much longer than regular training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 271, + 545, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 271, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 305, + 271, + 545, + 296 + ], + "type": "text", + "content": "2.2. Tuning Hyperparameters Using Known OOD Samples" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 304, + 545, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 435 + ], + "type": "text", + "content": "Some OOD detection approaches had hyperparameters that need to be tuned. Liang et al. [16] and Lee et al. [15] used specific OOD samples to tune hyperparameters under the assumption that distribution of OOD samples were given in advance. Liang et al. [16] tuned the temperature scaling factor and a perturbation size of input processing. Lee et al. [15] tuned weights for the feature ensemble and a perturbation size of input processing. However, these approaches are not feasible because it is difficult to know from which distribution OOD samples will occur in a real-world situation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 453, + 545, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 453, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 305, + 453, + 545, + 477 + ], + "type": "text", + "content": "2.3. Tuning Hyperparameters Using In-distribution Samples" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "Recently, some OOD detection approaches without using OOD samples were proposed because it was unrealistic to use OOD samples. Hsu et al. modified the classifying layer to obtain the probability that an input was in-distribution [12]. However, the modified classifying layer made the classification accuracy unstable. They also proposed a strategy that the perturbation size of the input processing was set without OOD samples, but it was not valid to all datasets. Sastry and Oore [23] proposed measures for OOD detection based on the Gram matrices of features. They stored the maximum and minimum values of the elements of the Gram matrix for each class, layer, power, and index of the matrix. They detected OOD samples using the stored values. However, it required heavy computation and a longer time than other approaches. Those approaches used additional InD validation dataset for tuning hyperparameters: Hsu et al. [12] tuned the perturbation size of input processing, and Sastry and Oore [23] tuned normalization factors for each layer." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16561" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 179, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 179, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 179, + 85 + ], + "type": "text", + "content": "2.4. Using Input Processing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 286, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 286, + 174 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 286, + 174 + ], + "type": "text", + "content": "The input processing is an operation that generates " + }, + { + "bbox": [ + 46, + 90, + 286, + 174 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 46, + 90, + 286, + 174 + ], + "type": "text", + "content": " by adding a small perturbation to an input " + }, + { + "bbox": [ + 46, + 90, + 286, + 174 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 90, + 286, + 174 + ], + "type": "text", + "content": ". Many previous approaches applied the input processing because this process effectively improved OOD detection performance. However, the perturbation size, epsilon, needed to be tuned appropriately. The epsilon has a significant impact on performance, so it should be chosen carefully." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 174, + 286, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 174, + 286, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 286, + 281 + ], + "type": "text", + "content": "Liang et al. [16] and Lee et al. [15] chose the epsilon values using given specific OOD samples. The chosen epsilon may not be valid to other OOD samples because it was tuned with a distribution of given OOD samples. Hsu et al. [12] proposed a method to set the epsilon value without OOD samples. They tuned the epsilon using the InD validation dataset, but their method did not generally work for all cases. The detection performance decreased in some datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 281, + 286, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 286, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 286, + 377 + ], + "type": "text", + "content": "Another disadvantage is that the input processing requires a lot of computation. We need one forward loop and one backward loop to add a perturbation to an input. We need one more forward loop to determine whether the perturbed input is OOD. Since the time taken for the forward loop and the backward loop is almost the same, the total inference time increases by about three times compared to the inference time without the input processing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 383, + 217, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 383, + 217, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 383, + 217, + 396 + ], + "type": "text", + "content": "2.5. Using Self-supervised Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 401, + 286, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 286, + 568 + ], + "type": "text", + "content": "Recently, a self-supervised learning technique that can learn a deep learning model without a label is attracting attention. In many studies in the field of OOD detection, self-supervised learning techniques are being tried to improve performance or reduce dependence on data labels. Some approaches build new data by applying transforms such as rotation to the image, and train models by giving the transform applied to the image as a label [7, 11, 27, 1]. Other approaches use contrastive learning techniques to train the model [27, 24]. However, in most cases, the OOD detection performance is not as high as when learning using labels. In addition, since the model is trained only for the purpose of OOD detection or representation learning, an additional method is needed to perform classification." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 578, + 153, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 153, + 592 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 153, + 592 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 598, + 286, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 598, + 286, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 598, + 286, + 647 + ], + "type": "text", + "content": "A deep learning-based classification model can be decomposed into a feature extractor " + }, + { + "bbox": [ + 46, + 598, + 286, + 647 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 46, + 598, + 286, + 647 + ], + "type": "text", + "content": " and a classifier " + }, + { + "bbox": [ + 46, + 598, + 286, + 647 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 46, + 598, + 286, + 647 + ], + "type": "text", + "content": " which is usually a fully connected layer. When training with the standard softmax loss, a logit is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 658, + 286, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 658, + 286, + 671 + ], + "spans": [ + { + "bbox": [ + 107, + 658, + 286, + 671 + ], + "type": "interline_equation", + "content": "\\operatorname {l o g i t} (x, c) = f (x) \\cdot w _ {c} + b _ {c}, \\tag {1}", + "image_path": "d1a9a063ef502122f5addad4f5b17f80e25c8f1399c1b0c1d354694ba58895f7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 673, + 271, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 271, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 271, + 685 + ], + "type": "text", + "content": "and with the cosine-based softmax loss, it is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 111, + 689, + 286, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 689, + 286, + 717 + ], + "spans": [ + { + "bbox": [ + 111, + 689, + 286, + 717 + ], + "type": "interline_equation", + "content": "\\operatorname {l o g i t} (x, c) = \\frac {f (x) \\cdot w _ {c}}{| f (x) | \\cdot | w _ {c} |}, \\tag {2}", + "image_path": "b20b646bbbdfa7314867beac4a8d98fff15cf07b834e27af01437027baff8fc5.jpg" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 79, + 419, + 152 + ], + "blocks": [ + { + "bbox": [ + 307, + 79, + 419, + 152 + ], + "lines": [ + { + "bbox": [ + 307, + 79, + 419, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 79, + 419, + 152 + ], + "type": "image", + "image_path": "825b49fc37e084e1c1ab0f01121b7cec8973b23fd1c07afb0f86c1efd4c238d6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 156, + 413, + 167 + ], + "lines": [ + { + "bbox": [ + 320, + 156, + 413, + 167 + ], + "spans": [ + { + "bbox": [ + 320, + 156, + 413, + 167 + ], + "type": "text", + "content": "(a) Standard softmax loss" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 426, + 79, + 533, + 152 + ], + "blocks": [ + { + "bbox": [ + 426, + 79, + 533, + 152 + ], + "lines": [ + { + "bbox": [ + 426, + 79, + 533, + 152 + ], + "spans": [ + { + "bbox": [ + 426, + 79, + 533, + 152 + ], + "type": "image", + "image_path": "6d7b815eb1d20eb07273025a433b8073fe90b6e0ea7619058c6bed45b3416c3c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 156, + 540, + 167 + ], + "lines": [ + { + "bbox": [ + 430, + 156, + 540, + 167 + ], + "spans": [ + { + "bbox": [ + 430, + 156, + 540, + 167 + ], + "type": "text", + "content": "(b) Cosine-based softmax loss" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 177, + 545, + 224 + ], + "lines": [ + { + "bbox": [ + 304, + 177, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 304, + 177, + 545, + 224 + ], + "type": "text", + "content": "Figure 1: The feature norm distributions of InD (red) and OOD (blue) data. (a) is from DenseNet3 trained with the standard softmax loss and (b) is from DenseNet3 trained with the cosine-based softmax loss." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 256, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 545, + 291 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 256, + 545, + 291 + ], + "type": "inline_equation", + "content": "w_{c}" + }, + { + "bbox": [ + 304, + 256, + 545, + 291 + ], + "type": "text", + "content": " is the connection weights of " + }, + { + "bbox": [ + 304, + 256, + 545, + 291 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 304, + 256, + 545, + 291 + ], + "type": "text", + "content": " for a class " + }, + { + "bbox": [ + 304, + 256, + 545, + 291 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 256, + 545, + 291 + ], + "type": "text", + "content": ". The cosine-based softmax loss is widely used in face recognition field [28, 5, 29]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 292, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 545, + 363 + ], + "type": "text", + "content": "We introduce two useful measures for OOD detection that can be simply extracted from models trained with cosine-based softmax loss: a feature norm-based measure and a Mahalanobis distance-based measure. We then propose an OOD score by combining these two measures. Let us describe each measure and the score function in detail." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 371, + 545, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 545, + 383 + ], + "type": "text", + "content": "3.1. Cosine-based Softmax Loss and Feature Norm" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 389, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 473 + ], + "type": "text", + "content": "When we train a classification model with the closed-world assumption, the logit value for a class " + }, + { + "bbox": [ + 304, + 389, + 545, + 473 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 389, + 545, + 473 + ], + "type": "text", + "content": " is considered as being proportional to " + }, + { + "bbox": [ + 304, + 389, + 545, + 473 + ], + "type": "inline_equation", + "content": "P(y = c|x)" + }, + { + "bbox": [ + 304, + 389, + 545, + 473 + ], + "type": "text", + "content": ". However, in the open world, the logit value should be interpreted as being proportional to " + }, + { + "bbox": [ + 304, + 389, + 545, + 473 + ], + "type": "inline_equation", + "content": "P(y = c|x, D_{in})" + }, + { + "bbox": [ + 304, + 389, + 545, + 473 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 389, + 545, + 473 + ], + "type": "inline_equation", + "content": "D_{in}" + }, + { + "bbox": [ + 304, + 389, + 545, + 473 + ], + "type": "text", + "content": " is the event that " + }, + { + "bbox": [ + 304, + 389, + 545, + 473 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 389, + 545, + 473 + ], + "type": "text", + "content": " is in the train distribution. The probability can be decomposed using " + }, + { + "bbox": [ + 304, + 389, + 545, + 473 + ], + "type": "inline_equation", + "content": "P(y = c, D_{in}|x)" + }, + { + "bbox": [ + 304, + 389, + 545, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 389, + 545, + 473 + ], + "type": "inline_equation", + "content": "P(D_{in}|x)" + }, + { + "bbox": [ + 304, + 389, + 545, + 473 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 481, + 545, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 545, + 508 + ], + "type": "interline_equation", + "content": "\\operatorname {l o g i t} (x, c) \\propto P (y = c | x, D _ {i n}) = \\frac {P (y = c , D _ {i n} | x)}{P \\left(D _ {i n} | x\\right)}. \\tag {3}", + "image_path": "179067ac2fffeba48206b2a3fa15fb2be30a8ba7c533db0fcaa64f3ef7d0485a.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 514, + 545, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 545, + 537 + ], + "type": "text", + "content": "If we use the cosine-based softmax loss, Equations 2 and 3 can be combined as follows:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 355, + 544, + 545, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 544, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 355, + 544, + 545, + 571 + ], + "type": "interline_equation", + "content": "\\frac {f (x) \\cdot w _ {c}}{| f (x) | \\cdot | w _ {c} |} \\propto \\frac {P (y = c , D _ {i n} | x)}{P \\left(D _ {i n} | x\\right)}. \\tag {4}", + "image_path": "7be01e75e1b1a004186e9429b53c8c54c19be9f72509c0bcfaae06c9d64d0350.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 578, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 624 + ], + "type": "text", + "content": "The numerator of the right-hand side includes a class-related term, " + }, + { + "bbox": [ + 304, + 578, + 545, + 624 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 578, + 545, + 624 + ], + "type": "text", + "content": ", and the denominator, " + }, + { + "bbox": [ + 304, + 578, + 545, + 624 + ], + "type": "inline_equation", + "content": "P(D_{in}|x)" + }, + { + "bbox": [ + 304, + 578, + 545, + 624 + ], + "type": "text", + "content": ", does not. So, we reformulate the left-hand side so that only the numerator has class-related terms as follows:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 350, + 631, + 545, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 631, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 350, + 631, + 545, + 658 + ], + "type": "interline_equation", + "content": "\\frac {f (x) \\cdot w _ {c} / | w _ {c} |}{| f (x) |} \\propto \\frac {P (y = c , D _ {i n} | x)}{P \\left(D _ {i n} | x\\right)}. \\tag {5}", + "image_path": "cd08f6c34b88822225ac034591d73b4b3a6301f663d9f73debd85b992b3f4a23.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "inline_equation", + "content": "w_{c}" + }, + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "text", + "content": " is the connection weight to " + }, + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "text", + "content": ", we can say it is class-related, and since " + }, + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "text", + "content": " is the feature representation of " + }, + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "text", + "content": ", we can say it is classifier-independent. Based on this, we rigorously infer that both denominators can be proportional" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16562" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 79, + 159, + 160 + ], + "blocks": [ + { + "bbox": [ + 49, + 79, + 159, + 160 + ], + "lines": [ + { + "bbox": [ + 49, + 79, + 159, + 160 + ], + "spans": [ + { + "bbox": [ + 49, + 79, + 159, + 160 + ], + "type": "image", + "image_path": "c0b5afe3ce518fab97a13dc7f710470d428577900820d4dcd6cdb05ec1f732a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 165, + 154, + 176 + ], + "lines": [ + { + "bbox": [ + 62, + 165, + 154, + 176 + ], + "spans": [ + { + "bbox": [ + 62, + 165, + 154, + 176 + ], + "type": "text", + "content": "(a) Distribution of " + }, + { + "bbox": [ + 62, + 165, + 154, + 176 + ], + "type": "inline_equation", + "content": "|f(x)|" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 167, + 79, + 275, + 160 + ], + "blocks": [ + { + "bbox": [ + 167, + 79, + 275, + 160 + ], + "lines": [ + { + "bbox": [ + 167, + 79, + 275, + 160 + ], + "spans": [ + { + "bbox": [ + 167, + 79, + 275, + 160 + ], + "type": "image", + "image_path": "ac6384f326f9c55319542164ee2e17d973fc31da3561b548d0c6df0194472bb5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 165, + 260, + 176 + ], + "lines": [ + { + "bbox": [ + 192, + 165, + 260, + 176 + ], + "spans": [ + { + "bbox": [ + 192, + 165, + 260, + 176 + ], + "type": "text", + "content": "(b) Mean of " + }, + { + "bbox": [ + 192, + 165, + 260, + 176 + ], + "type": "inline_equation", + "content": "|f(x)|" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 277, + 287, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 314 + ], + "type": "text", + "content": "to each, i.e., " + }, + { + "bbox": [ + 46, + 277, + 287, + 314 + ], + "type": "inline_equation", + "content": "|f(x)| \\propto P(D_{in}|x)" + }, + { + "bbox": [ + 46, + 277, + 287, + 314 + ], + "type": "text", + "content": ", because the two terms in Equation 5 are proportional, and their denominators and numerators share class-related characteristics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 315, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 287, + 434 + ], + "type": "text", + "content": "To confirm this, we perform some OOD detection experiments. We try to detect SVHN using " + }, + { + "bbox": [ + 46, + 315, + 287, + 434 + ], + "type": "inline_equation", + "content": "|f(x)|" + }, + { + "bbox": [ + 46, + 315, + 287, + 434 + ], + "type": "text", + "content": " from DenseNet3 model trained for CIFAR-10 with the cosine-based softmax loss. We compare this to the model using the standard softmax loss as shown in Figure 1. The feature norm distributions of OOD and InD by the standard loss model overlap much, but those of the cosine loss model have a very small overlap between SVHN and CIFAR-10, which says that the cosine loss model can be better to discriminate ODD than the standard model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 435, + 287, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 482 + ], + "type": "text", + "content": "To verify that the length of feature norm decrease as the samples get closer to OOD, we create mixup samples by mixing SVHN and CIFAR-10 images as shown in Equation 6 and observed the feature norm." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 89, + 494, + 287, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 494, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 89, + 494, + 287, + 507 + ], + "type": "interline_equation", + "content": "\\operatorname {T e s t I m g} = \\lambda \\cdot \\operatorname {I n D} + (1 - \\lambda) \\cdot \\operatorname {O O D} \\tag {6}", + "image_path": "9dbffd101cc6387c8f5655b3088893f66381441c1e9fb8c98a4e1efa4a9485e0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "text", + "content": "As the mixup factor, " + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "text", + "content": " decreases, that is, as the samples get closer to OOD, the feature norm also decreases as shown in Figure 2. The feature norm of an image from the cosine loss model tends to increase as the image gets closer to InD samples and decreases closer to OOD samples, which supports our conclusion that " + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "inline_equation", + "content": "|f(x)|" + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "text", + "content": " from a cosine loss model is proportional to " + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "inline_equation", + "content": "P(D_{in}|x)" + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "text", + "content": ". Also, it supports that the feature norm can be a measure for OOD detection. Additional observations, including CIFAR-100 and SVHN, are presented in the appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 647, + 287, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 287, + 670 + ], + "type": "text", + "content": "3.2. Mahalanobis Distance in Models Trained with Cosine-based Softmax Loss" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "We may assume that features of InD samples obtained from internal layers of classification models follow the Gaussian distribution. We adopt an OOD score defined by" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 78, + 420, + 147 + ], + "blocks": [ + { + "bbox": [ + 307, + 78, + 420, + 147 + ], + "lines": [ + { + "bbox": [ + 307, + 78, + 420, + 147 + ], + "spans": [ + { + "bbox": [ + 307, + 78, + 420, + 147 + ], + "type": "image", + "image_path": "2913a3b03bd1ba529efe9fadfd27919720bf4feb1ab15c505f627377bd3b4b7d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 152, + 424, + 162 + ], + "lines": [ + { + "bbox": [ + 310, + 152, + 424, + 162 + ], + "spans": [ + { + "bbox": [ + 310, + 152, + 424, + 162 + ], + "type": "text", + "content": "(a) 1st Block (SVHN vs. TINc)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 426, + 78, + 538, + 147 + ], + "blocks": [ + { + "bbox": [ + 426, + 78, + 538, + 147 + ], + "lines": [ + { + "bbox": [ + 426, + 78, + 538, + 147 + ], + "spans": [ + { + "bbox": [ + 426, + 78, + 538, + 147 + ], + "type": "image", + "image_path": "1540a4e03e4294be16f53776a9a35fa65b03108b1140bc6176a75ff781ed0d41.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 152, + 543, + 162 + ], + "lines": [ + { + "bbox": [ + 427, + 152, + 543, + 162 + ], + "spans": [ + { + "bbox": [ + 427, + 152, + 543, + 162 + ], + "type": "text", + "content": "(b) 1st Block (SVHN vs. iSUN)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 171, + 419, + 242 + ], + "blocks": [ + { + "bbox": [ + 46, + 186, + 287, + 244 + ], + "lines": [ + { + "bbox": [ + 46, + 186, + 287, + 244 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 287, + 244 + ], + "type": "text", + "content": "Figure 2: The distribution of feature norm according to the mixup factor (a), and the mean value of feature norm according to the mixup factor. It can be observed that as OOD samples are more mixed into a sample, the feature norm tends to decrease." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 171, + 419, + 242 + ], + "lines": [ + { + "bbox": [ + 307, + 171, + 419, + 242 + ], + "spans": [ + { + "bbox": [ + 307, + 171, + 419, + 242 + ], + "type": "image", + "image_path": "d1ba4001561006a0b2c01e633b46a4f109f7fb8f790596caa1d3f7efaaffafd8.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 426, + 171, + 537, + 242 + ], + "blocks": [ + { + "bbox": [ + 308, + 247, + 425, + 257 + ], + "lines": [ + { + "bbox": [ + 308, + 247, + 425, + 257 + ], + "spans": [ + { + "bbox": [ + 308, + 247, + 425, + 257 + ], + "type": "text", + "content": "(c) 2nd Block (SVHN vs. TINc)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 426, + 171, + 537, + 242 + ], + "lines": [ + { + "bbox": [ + 426, + 171, + 537, + 242 + ], + "spans": [ + { + "bbox": [ + 426, + 171, + 537, + 242 + ], + "type": "image", + "image_path": "41c6e10e01943d5d85d3d56f8dc6107449ca80a3b495e64431d9bbb25d32b8da.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 247, + 544, + 257 + ], + "lines": [ + { + "bbox": [ + 427, + 247, + 544, + 257 + ], + "spans": [ + { + "bbox": [ + 427, + 247, + 544, + 257 + ], + "type": "text", + "content": "(d) 2nd Block (SVHN vs. iSUN)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 307, + 266, + 419, + 337 + ], + "blocks": [ + { + "bbox": [ + 307, + 266, + 419, + 337 + ], + "lines": [ + { + "bbox": [ + 307, + 266, + 419, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 419, + 337 + ], + "type": "image", + "image_path": "a78bc1dab61379cc9f038bd0b47dc810ac1d233d091151ae70a4c6921a28f07f.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 342, + 424, + 352 + ], + "lines": [ + { + "bbox": [ + 308, + 342, + 424, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 424, + 352 + ], + "type": "text", + "content": "(e) 3rd Block (SVHN vs. TINc)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 426, + 266, + 537, + 337 + ], + "blocks": [ + { + "bbox": [ + 426, + 266, + 537, + 337 + ], + "lines": [ + { + "bbox": [ + 426, + 266, + 537, + 337 + ], + "spans": [ + { + "bbox": [ + 426, + 266, + 537, + 337 + ], + "type": "image", + "image_path": "ea705c855111256c1c3523f9ed65914eef230856b758dc977ce8244b7ca4b046.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 342, + 543, + 352 + ], + "lines": [ + { + "bbox": [ + 427, + 342, + 543, + 352 + ], + "spans": [ + { + "bbox": [ + 427, + 342, + 543, + 352 + ], + "type": "text", + "content": "(f) 3rd Block (SVHN vs. iSUN)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 363, + 545, + 399 + ], + "lines": [ + { + "bbox": [ + 305, + 363, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 363, + 545, + 399 + ], + "type": "text", + "content": "Figure 3: Distributions of " + }, + { + "bbox": [ + 305, + 363, + 545, + 399 + ], + "type": "inline_equation", + "content": "MD" + }, + { + "bbox": [ + 305, + 363, + 545, + 399 + ], + "type": "text", + "content": " scores from each block of DenseNet3 trained with SVHN. The OODs datasets (blue) is TinyImageNet-crop (left) and iSUN (right)." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 430, + 509, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 509, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 509, + 441 + ], + "type": "text", + "content": "Lee et al. [15] based on the Mahalanobis distance:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 453, + 545, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 453, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 317, + 453, + 545, + 486 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} M D _ {B} (x) \\\\ = - \\max _ {c} - \\left(f _ {B} (x) - \\mu_ {c, B}\\right) ^ {\\top} \\Sigma_ {B} ^ {- 1} \\left(f _ {B} (x) - \\mu_ {c, B}\\right), \\quad (7) \\\\ \\end{array}", + "image_path": "e41414ee7b55f846f83d8f01ff4e59025a141b551cd45b04bbe77ee9b294f5d4.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 493, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 545, + 588 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 493, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\mu_{c,B}" + }, + { + "bbox": [ + 304, + 493, + 545, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 493, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\Sigma_B" + }, + { + "bbox": [ + 304, + 493, + 545, + 588 + ], + "type": "text", + "content": " are the mean and the covariance of features from a block " + }, + { + "bbox": [ + 304, + 493, + 545, + 588 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 493, + 545, + 588 + ], + "type": "text", + "content": ". This score is very discriminative for OOD detection as shown in Figure 3. Since the distributions of " + }, + { + "bbox": [ + 304, + 493, + 545, + 588 + ], + "type": "inline_equation", + "content": "MD" + }, + { + "bbox": [ + 304, + 493, + 545, + 588 + ], + "type": "text", + "content": " scores of InD and OOD samples have small overlaps, we can detect OOD samples using the scores. However, " + }, + { + "bbox": [ + 304, + 493, + 545, + 588 + ], + "type": "inline_equation", + "content": "MD" + }, + { + "bbox": [ + 304, + 493, + 545, + 588 + ], + "type": "text", + "content": " scores from the penultimate layer overlap if the model is trained with cosine-based softmax loss, as shown in Figure 2(e)-(f)." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 588, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 545, + 658 + ], + "type": "text", + "content": "The reason why " + }, + { + "bbox": [ + 304, + 588, + 545, + 658 + ], + "type": "inline_equation", + "content": "MD" + }, + { + "bbox": [ + 304, + 588, + 545, + 658 + ], + "type": "text", + "content": " scores from penultimate layer is not discriminative is that the feature from the penultimate layer do not follow a Gaussian distribution. In the model trained with cosine-based softmax loss, the softmax probability with " + }, + { + "bbox": [ + 304, + 588, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\sigma_c" + }, + { + "bbox": [ + 304, + 588, + 545, + 658 + ], + "type": "text", + "content": ", the softmax value for a class " + }, + { + "bbox": [ + 304, + 588, + 545, + 658 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 588, + 545, + 658 + ], + "type": "text", + "content": ", is calculated as follows:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 314, + 676, + 545, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 676, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 314, + 676, + 545, + 712 + ], + "type": "interline_equation", + "content": "\\sigma_ {c} (g (f (x))) = \\frac {\\exp \\left(w _ {c} \\cdot f (x) / \\left(\\left| w _ {c} \\right| \\cdot | f (x) |\\right)\\right)}{\\sum_ {c ^ {\\prime}} \\exp \\left(w _ {c ^ {\\prime}} \\cdot f (x) / \\left(\\left| w _ {c ^ {\\prime}} \\right| \\cdot | f (x) |\\right)\\right)}. \\tag {8}", + "image_path": "94b918fbb5e6459700be5e17448f1ad978e69c98db25e3862a5e4e0c29313768.jpg" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16563" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 154, + 70, + 440, + 156 + ], + "blocks": [ + { + "bbox": [ + 154, + 70, + 440, + 156 + ], + "lines": [ + { + "bbox": [ + 154, + 70, + 440, + 156 + ], + "spans": [ + { + "bbox": [ + 154, + 70, + 440, + 156 + ], + "type": "table", + "html": "
ApproachAuxiliary Dataset# of Hyperp.Input Processing
OE [10]Large OOD dataset0×
ODIN [16]OOD samples2✓
MAHA [15]OOD samples4✓
G-ODIN [12]InD samples1✓
GRAM [23]InD samples100×
COD (Proposed)None0×
", + "image_path": "d8cc9fb07b80701d78a4657d634bc82c220a370270d3d88bbcadd9bf97a05b1e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 164, + 547, + 190 + ], + "lines": [ + { + "bbox": [ + 46, + 164, + 547, + 190 + ], + "spans": [ + { + "bbox": [ + 46, + 164, + 547, + 190 + ], + "type": "text", + "content": "Table 1: OOD detection approach comparison. The number of hyperparameters is counted based on DenseNet3 with 100 layers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 210, + 289, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 210, + 289, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 210, + 289, + 286 + ], + "type": "text", + "content": "The classification model is trained so that the softmax probability reflects " + }, + { + "bbox": [ + 47, + 210, + 289, + 286 + ], + "type": "inline_equation", + "content": "P(y = c|f(x))" + }, + { + "bbox": [ + 47, + 210, + 289, + 286 + ], + "type": "text", + "content": ". If the feature distribution from the penultimate layer is assumed to be the multivariate Gaussian distribution for each class, and the class prior is " + }, + { + "bbox": [ + 47, + 210, + 289, + 286 + ], + "type": "inline_equation", + "content": "P(y = c) = \\frac{\\beta_c}{\\sum_{c'} \\beta_{c'}}" + }, + { + "bbox": [ + 47, + 210, + 289, + 286 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 47, + 210, + 289, + 286 + ], + "type": "inline_equation", + "content": "\\beta_c" + }, + { + "bbox": [ + 47, + 210, + 289, + 286 + ], + "type": "text", + "content": " is unnormalized prior for class " + }, + { + "bbox": [ + 47, + 210, + 289, + 286 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 210, + 289, + 286 + ], + "type": "text", + "content": ", the probability of class " + }, + { + "bbox": [ + 47, + 210, + 289, + 286 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 210, + 289, + 286 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 210, + 289, + 286 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 47, + 210, + 289, + 286 + ], + "type": "text", + "content": " is as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 309, + 287, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 309, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 57, + 309, + 287, + 396 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P (y = c \\mid f (x)) \\\\ = \\frac {P (y = c) P (f (x) | y = c)}{\\sum_ {c ^ {\\prime}} P (y = c ^ {\\prime}) P (f (x) | y = c ^ {\\prime})} \\\\ = \\frac {\\exp \\left(\\mu_ {c} ^ {T} \\Sigma^ {- 1} f (x) - \\frac {1}{2} \\mu_ {c} ^ {T} \\Sigma^ {- 1} \\mu_ {c} + \\log \\beta_ {c}\\right)}{\\sum_ {c ^ {\\prime}} \\exp \\left(\\mu_ {c ^ {\\prime}} ^ {T} \\Sigma^ {- 1} f (x) - \\frac {1}{2} \\mu_ {c ^ {\\prime}} ^ {T} \\Sigma^ {- 1} \\mu_ {c ^ {\\prime}} + \\log \\beta_ {c ^ {\\prime}}\\right)}. \\tag {9} \\\\ \\end{array}", + "image_path": "802fe91f9c23c258891f0640319ef0fda65966351519861bfa783380850cba3f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 404, + 287, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 287, + 499 + ], + "type": "text", + "content": "That is, if the feature from the penultimate layer follows a Gaussian distribution, " + }, + { + "bbox": [ + 46, + 404, + 287, + 499 + ], + "type": "inline_equation", + "content": "P(y = c|f(x))" + }, + { + "bbox": [ + 46, + 404, + 287, + 499 + ], + "type": "text", + "content": " is based on linear expressions. Since a nonlinear expression cannot be transformed in a linear expression, the softmax probability from the cosine loss model (Equation 8) cannot be converted to a similar form of " + }, + { + "bbox": [ + 46, + 404, + 287, + 499 + ], + "type": "inline_equation", + "content": "P(y = c|f(x))" + }, + { + "bbox": [ + 46, + 404, + 287, + 499 + ], + "type": "text", + "content": " (Equation 9), which means that the features from the penultimate layer is not expected to follow a Gaussian distribution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 501, + 287, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 287, + 526 + ], + "type": "text", + "content": "If Equation 8 is rewritten using a normalized feature " + }, + { + "bbox": [ + 47, + 501, + 287, + 526 + ], + "type": "inline_equation", + "content": "f^{\\prime}(x) = f(x) / |f(x)|" + }, + { + "bbox": [ + 47, + 501, + 287, + 526 + ], + "type": "text", + "content": ", it is as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 65, + 544, + 287, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 544, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 65, + 544, + 287, + 573 + ], + "type": "interline_equation", + "content": "\\sigma_ {c} (g (f (x))) = \\frac {\\exp \\left(w _ {c} ^ {\\top} f ^ {\\prime} (x) / \\left(\\left| w _ {c} ^ {\\top} \\right|\\right)\\right)}{\\sum_ {c ^ {\\prime}} \\exp \\left(w _ {c ^ {\\prime}} ^ {\\top} f ^ {\\prime} (x) / \\left(\\left| w _ {c ^ {\\prime}} ^ {\\top} \\right|\\right)\\right)}. \\tag {10}", + "image_path": "7962e9b1043f855a87ef40c489ab6d35d3fca9679ec6bedaea42559ba13257d4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": "Equation 10 is based on linear expressions of " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "f'(x)" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": ", from which we conjecture that " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "f'(x)" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": " may follow a Gaussian distribution. Since " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "MD" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": " scores are useful under the assumption that features follow a Gaussian distribution, we redefine the feature for " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "MD" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": " from the penultimate layer as " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "f'(x) = f(x) / |f(x)|" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": ". We re-depict the distribution of " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "MD" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": " in Figures 2-(e) and (f) using " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "f'(x)" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": " as shown in Figure 4. The overlap of the distributions is significantly reduced compared to " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "MD" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": ", which means that " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "MD" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "f'(x)" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": " may be better than " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": " for OOD detection." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 216, + 420, + 285 + ], + "blocks": [ + { + "bbox": [ + 307, + 216, + 420, + 285 + ], + "lines": [ + { + "bbox": [ + 307, + 216, + 420, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 216, + 420, + 285 + ], + "type": "image", + "image_path": "069b255e7d1cf33cf84628fa72a425a26ef0f52c098a564b504635a0b50eb375.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 289, + 424, + 300 + ], + "lines": [ + { + "bbox": [ + 310, + 289, + 424, + 300 + ], + "spans": [ + { + "bbox": [ + 310, + 289, + 424, + 300 + ], + "type": "text", + "content": "(a) 3rd Block (SVHN vs TINc)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 426, + 215, + 538, + 285 + ], + "blocks": [ + { + "bbox": [ + 426, + 215, + 538, + 285 + ], + "lines": [ + { + "bbox": [ + 426, + 215, + 538, + 285 + ], + "spans": [ + { + "bbox": [ + 426, + 215, + 538, + 285 + ], + "type": "image", + "image_path": "d6c43b08bec461609b8de4b457afbce9907805edb48808435420b343c1967626.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 289, + 543, + 300 + ], + "lines": [ + { + "bbox": [ + 427, + 289, + 543, + 300 + ], + "spans": [ + { + "bbox": [ + 427, + 289, + 543, + 300 + ], + "type": "text", + "content": "(b) 3rd Block (SVHN vs iSUN)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 309, + 545, + 335 + ], + "lines": [ + { + "bbox": [ + 305, + 309, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 309, + 545, + 335 + ], + "type": "text", + "content": "Figure 4: Distributions of " + }, + { + "bbox": [ + 305, + 309, + 545, + 335 + ], + "type": "inline_equation", + "content": "MD" + }, + { + "bbox": [ + 305, + 309, + 545, + 335 + ], + "type": "text", + "content": " scores based on " + }, + { + "bbox": [ + 305, + 309, + 545, + 335 + ], + "type": "inline_equation", + "content": "f^{\\prime}(x)" + }, + { + "bbox": [ + 305, + 309, + 545, + 335 + ], + "type": "text", + "content": ". The configuration is the same as Figure 3." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 365, + 546, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 546, + 413 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 546, + 413 + ], + "type": "text", + "content": "We evaluate " + }, + { + "bbox": [ + 305, + 365, + 546, + 413 + ], + "type": "inline_equation", + "content": "MD" + }, + { + "bbox": [ + 305, + 365, + 546, + 413 + ], + "type": "text", + "content": " scores for the penultimate layer with the redefined feature, " + }, + { + "bbox": [ + 305, + 365, + 546, + 413 + ], + "type": "inline_equation", + "content": "f'(x) = f(x) / |f(x)|" + }, + { + "bbox": [ + 305, + 365, + 546, + 413 + ], + "type": "text", + "content": " and for other layers with " + }, + { + "bbox": [ + 305, + 365, + 546, + 413 + ], + "type": "inline_equation", + "content": "f_B(x)" + }, + { + "bbox": [ + 305, + 365, + 546, + 413 + ], + "type": "text", + "content": ". Finally, we combine " + }, + { + "bbox": [ + 305, + 365, + 546, + 413 + ], + "type": "inline_equation", + "content": "MD" + }, + { + "bbox": [ + 305, + 365, + 546, + 413 + ], + "type": "text", + "content": " scores from each block as Equation 11 by product:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 368, + 422, + 545, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 422, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 368, + 422, + 545, + 447 + ], + "type": "interline_equation", + "content": "M D (x) = \\prod_ {B} \\left(M D _ {B} (x)\\right). \\tag {11}", + "image_path": "bb650071af5c67e9dc00e4d951100e3e5414c56d68976d55df14fa69ff11862e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 453, + 474, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 474, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 474, + 466 + ], + "type": "text", + "content": "3.3. Proposed OOD Score Function" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 472, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 472, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 545, + 544 + ], + "type": "text", + "content": "We define a cosine-based OOD score function, " + }, + { + "bbox": [ + 305, + 472, + 545, + 544 + ], + "type": "inline_equation", + "content": "COD(x)" + }, + { + "bbox": [ + 305, + 472, + 545, + 544 + ], + "type": "text", + "content": ", with the feature norm and the Mahalanobis distance-based measure, " + }, + { + "bbox": [ + 305, + 472, + 545, + 544 + ], + "type": "inline_equation", + "content": "MD" + }, + { + "bbox": [ + 305, + 472, + 545, + 544 + ], + "type": "text", + "content": ". Since the feature norm tends to appear large for InD samples and " + }, + { + "bbox": [ + 305, + 472, + 545, + 544 + ], + "type": "inline_equation", + "content": "MD" + }, + { + "bbox": [ + 305, + 472, + 545, + 544 + ], + "type": "text", + "content": " tends to be large for OOD samples, we define " + }, + { + "bbox": [ + 305, + 472, + 545, + 544 + ], + "type": "inline_equation", + "content": "COD(x)" + }, + { + "bbox": [ + 305, + 472, + 545, + 544 + ], + "type": "text", + "content": " by dividing the feature norm by " + }, + { + "bbox": [ + 305, + 472, + 545, + 544 + ], + "type": "inline_equation", + "content": "MD" + }, + { + "bbox": [ + 305, + 472, + 545, + 544 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 379, + 551, + 545, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 551, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 379, + 551, + 545, + 578 + ], + "type": "interline_equation", + "content": "C O D (x) = \\frac {| f (x) |}{M D (x)}. \\tag {12}", + "image_path": "a3503a1992c9a3ce11cdd6d1d5a1e4aa8e383866237c28dd55476856607f9126.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "COD(x)" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " does not have any hyperparameter to be tuned. Also, no additional dataset is required to use " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "COD(x)" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": ", and it shows a high discriminative performance without the input processing. " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "COD(x)" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " can be calculated with a very small computation in a forward loop for classification. We compare the previous approaches in the viewpoint of auxiliary datasets, the number of hyperparameters and the input processing in Table 1. Most of the previous methods used auxiliary datasets, and had several to one hundred hyperparameters, and needed the input processing to increase the performance, while our method does not need any of them." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16564" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 123, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 123, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 123, + 85 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 286, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 286, + 116 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 286, + 116 + ], + "type": "text", + "content": "This section shows the effectiveness of our approach through experiments with various image datasets." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 123, + 160, + 136 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 123, + 160, + 136 + ], + "spans": [ + { + "bbox": [ + 47, + 123, + 160, + 136 + ], + "type": "text", + "content": "4.1. Experiment setting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 142, + 286, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 142, + 286, + 273 + ], + "spans": [ + { + "bbox": [ + 47, + 142, + 286, + 273 + ], + "type": "text", + "content": "Datasets: We use CIFAR-10/100 [14], and SVHN [20] images for InD datasets. We use the training dataset for classification and use the test datasets to evaluate OOD detection. We use TinyImageNet-crop (TINc), TinyImageNet-resize (TINr), LSUN-crop (LSUNc), LSUN-resize (LSUNr), and iSUN as OOD datasets. These datasets have been used in many studies related to OOD detection. Additionally, when CIFAR-10/100 are in-distribution, SVHN test dataset is also used as an OOD dataset, and when SVHN image is in-distribution, CIFAR-10/100 test datasets are also used as OOD datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 278, + 286, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 278, + 286, + 315 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 286, + 315 + ], + "type": "text", + "content": "Baselines: We compare our approach to previous OOD detection methods: ODIN [16], MAHA [15], G-ODIN [12], and GRAM [23]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 319, + 286, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 286, + 427 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 286, + 427 + ], + "type": "text", + "content": "Networks and Training Details: We train ResNet34 [8] and DenseNet3 [13] models with the same setting as the baseline approaches. We train ResNet34 for 200 epochs with a batch size of 128 and a weight decay of 0.0005. We train DenseNet3 with 100 layers for 300 epochs with a batch size of 64 and a weight decay of 0.0001. In both training, the optimizer is SGD with a momentum of 0.9, and the learning rate starts with 0.1 and decays by a factor 0.1 at " + }, + { + "bbox": [ + 47, + 319, + 286, + 427 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 47, + 319, + 286, + 427 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 319, + 286, + 427 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 47, + 319, + 286, + 427 + ], + "type": "text", + "content": " of the training progress." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 428, + 286, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 428, + 286, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 428, + 286, + 510 + ], + "type": "text", + "content": "When we train models with the cosine-based softmax loss, logit values are between -1 and 1. For expanding the range of logits, we adopt the temperature scaling with a factor of " + }, + { + "bbox": [ + 47, + 428, + 286, + 510 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 428, + 286, + 510 + ], + "type": "text", + "content": " as in Equation 13. The factor of " + }, + { + "bbox": [ + 47, + 428, + 286, + 510 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 428, + 286, + 510 + ], + "type": "text", + "content": " needs to be set to 0.1 for SVHN, CIFAR-10 and 0.05 for CIFAR-100 so that the classification performance is maximized. The logit scaled by " + }, + { + "bbox": [ + 47, + 428, + 286, + 510 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 428, + 286, + 510 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 524, + 286, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 524, + 286, + 537 + ], + "spans": [ + { + "bbox": [ + 116, + 524, + 286, + 537 + ], + "type": "interline_equation", + "content": "\\operatorname {l o g i t} _ {T} (x) = \\operatorname {l o g i t} (x) / T. \\tag {13}", + "image_path": "09216d59b87f1aa6c9d705bbab6e4522b75aa6ea185c313eee6421c074ab65bb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 543, + 286, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 543, + 286, + 627 + ], + "spans": [ + { + "bbox": [ + 47, + 543, + 286, + 627 + ], + "type": "text", + "content": "Evaluation Metrics: We evaluate the detection performance using the two metrics that are most widely used for OOD detection. The first metric is the area under the receiver operating characteristic curve (AUROC). It is used to evaluate the average performance considering various thresholds. The second is TNR at TPR " + }, + { + "bbox": [ + 47, + 543, + 286, + 627 + ], + "type": "inline_equation", + "content": "p\\%" + }, + { + "bbox": [ + 47, + 543, + 286, + 627 + ], + "type": "text", + "content": ". We evaluate with " + }, + { + "bbox": [ + 47, + 543, + 286, + 627 + ], + "type": "inline_equation", + "content": "p = 95" + }, + { + "bbox": [ + 47, + 543, + 286, + 627 + ], + "type": "text", + "content": " as the metric used in the baseline approaches." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 635, + 104, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 104, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 104, + 647 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 653, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 286, + 713 + ], + "type": "text", + "content": "Table 2 shows an overall comparison with DenseNet3. Our approach achieves superior performance in all the datasets. Compared to ODIN, our approach has higher performance in all the cases. Compared to G-ODIN, our approach has higher performance except one cases. Com" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "content": "pared to MAHA, our approach has higher performance except some cases." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 97, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 97, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 305, + 97, + 545, + 288 + ], + "type": "text", + "content": "In some cases, GRAM shows a higher performance than the proposed method, but GRAM is very unstable depending on OOD datasets. For example, GRAM shows a particularly poor performance on LSUNc. It shows a TNR@95 score of 88.4 for CIFAR-10 and 65.6 for CIFAR-100. Our approach is more even performances across all OOD datasets, and achieves the highest average performances. Also, GRAM requires significantly heavy computation. When comparing the wall clock time in the same environment (GTX 2080 Ti GPU with Xeon Silver 4021 CPU * 2), the GRAM takes 0.35 seconds to calculate OOD score for one image, but our approach takes only about 0.018 seconds. Our approach is 19 times faster than GRAM because GRAM evaluates a lot of measures in the forward loop while our approach needs very small additional computation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 289, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 545, + 445 + ], + "type": "text", + "content": "The head of Table 2 shows what each approach needs to detect OOD samples. ODIN and MAHA use OOD samples and InD validation samples, and perform the input processing. G-ODIN does not use OOD samples but uses InD validation samples and performs the input processing. GRAM uses InD validation samples to tune hyperparameters. Our approach needs nothing except a training dataset. Our approach is the fastest method because the input processing is not required. Our method is superior to or at least comparable to state-of-the-art OOD detection methods even though it does not require any additional datasets. Our model also shows better performance with ResNet34. The results are presented in the appendix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 457, + 375, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 457, + 375, + 469 + ], + "spans": [ + { + "bbox": [ + 306, + 457, + 375, + 469 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 478, + 440, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 440, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 440, + 490 + ], + "type": "text", + "content": "5.1. Layer Selection for " + }, + { + "bbox": [ + 306, + 478, + 440, + 490 + ], + "type": "inline_equation", + "content": "MD" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 497, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 497, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 305, + 497, + 545, + 568 + ], + "type": "text", + "content": "Our approach, " + }, + { + "bbox": [ + 305, + 497, + 545, + 568 + ], + "type": "inline_equation", + "content": "COD" + }, + { + "bbox": [ + 305, + 497, + 545, + 568 + ], + "type": "text", + "content": ", uses " + }, + { + "bbox": [ + 305, + 497, + 545, + 568 + ], + "type": "inline_equation", + "content": "MD" + }, + { + "bbox": [ + 305, + 497, + 545, + 568 + ], + "type": "text", + "content": " values from all the blocks of the model. However, it can be more effective if we can choose some layers which are helpful for OOD detection. Useful layers for OOD detection may depend on the model structure, the training dataset, and the OOD dataset. However, OOD samples are not available in advance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 569, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 569, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 305, + 569, + 545, + 689 + ], + "type": "text", + "content": "We will discuss how to select useful layers with synthetic OOD samples. We may generate synthetic OOD samples from the training dataset. For example, we may use some training samples rotated by 90 degrees as synthetic OOD samples. Then, we evaluate AUROC scores for each layer, and choose layers of which AUROC scores are greater than a threshold such as " + }, + { + "bbox": [ + 305, + 569, + 545, + 689 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 305, + 569, + 545, + 689 + ], + "type": "text", + "content": " of AUROC. In the case of CIFAR10, CIFAR100, and SVHN, it is confirmed that the AUROC score of each " + }, + { + "bbox": [ + 305, + 569, + 545, + 689 + ], + "type": "inline_equation", + "content": "MD" + }, + { + "bbox": [ + 305, + 569, + 545, + 689 + ], + "type": "text", + "content": " was over " + }, + { + "bbox": [ + 305, + 569, + 545, + 689 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 305, + 569, + 545, + 689 + ], + "type": "text", + "content": ". Therefore, it can be a reasonable threshold to choose layers." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We apply the layer selection to a more realistic dataset, ImageNet-30. In this case, " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "MD" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " from the last layer, " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "MD" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16565" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 69, + 542, + 469 + ], + "blocks": [ + { + "bbox": [ + 52, + 69, + 542, + 469 + ], + "lines": [ + { + "bbox": [ + 52, + 69, + 542, + 469 + ], + "spans": [ + { + "bbox": [ + 52, + 69, + 542, + 469 + ], + "type": "table", + "html": "
OODAUROCTNR@95
Use a training datasetUse a training dataset
Use InD validation samples(ours)Use InD validation samples(ours)
Perform the input processingGRAMPerform the input processingGRAMCOD
Use OOD samplesG-ODINUse OOD samplesG-ODIN
ODINMAHAODINMAHA
InD: CIFAR-10
SVHN95.598.198.899.199.586.290.894.096.198.0
TINc97.695.398.799.399.487.084.293.496.797.9
TINr98.598.899.199.799.592.495.095.898.897.8
LSUNc93.680.298.397.599.970.648.291.588.499.6
LSUNr99.299.399.499.999.696.297.297.699.598.9
iSUN98.798.999.499.899.693.295.397.599.098.7
MEAN97.295.199.099.299.687.685.195.096.498.5
InD: CIFAR-100
SVHN93.897.295.997.397.070.682.577.089.382.4
TINc88.388.897.697.797.951.060.187.889.089.3
TINr85.297.498.699.098.642.686.693.395.793.1
LSUNc91.481.795.391.498.957.842.175.065.595.1
LSUNr85.598.098.799.398.641.291.493.897.293.3
iSUN84.597.498.499.098.537.487.092.595.992.5
MEAN88.193.497.497.398.350.175.086.688.891.0
InD: SVHN
CIFAR-1091.498.9-95.599.071.796.8-80.496.1
CIFAR-100----98.9----95.8
TINc----99.8----99.7
TINr95.199.9-99.799.884.199.9-99.199.5
LSUNc----99.8----99.4
LSUNr94.599.9-99.899.881.199.9-99.599.4
iSUN94.799.9-99.899.882.299.9-99.499.5
MEAN----99.6----98.5
", + "image_path": "89ca4be6844ecdec980500f5d336193eca03f85557c6f9c16c692b32a5776d0c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 78, + 546, + 516, + 596 + ], + "blocks": [ + { + "bbox": [ + 46, + 477, + 547, + 528 + ], + "lines": [ + { + "bbox": [ + 46, + 477, + 547, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 477, + 547, + 528 + ], + "type": "text", + "content": "Table 2: OOD detection performance (%) with a DenseNet3 classification model. Our performance is averaged over 5 runs. ODIN, MAHA and GRAM are cited from the work by Sastry and Oore [23] and G-ODIN is cited from the work by Hsu et al. [12]. In the case of SVHN, all baselines did not present their performances. Dash symbols stand for not-available. The mean rows are the averages of the performances over all OOD datasets. For each case, the best results are in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 78, + 546, + 516, + 596 + ], + "lines": [ + { + "bbox": [ + 78, + 546, + 516, + 596 + ], + "spans": [ + { + "bbox": [ + 78, + 546, + 516, + 596 + ], + "type": "table", + "html": "
ResNet18\nInD: ImageNet-30Test AccOOD Dataset
DTDCUB-200CaltechDogsPlaces-365Food-101PetFlowers
CSI97.093.793.491.997.792.587.096.996.0
COD97.294.092.891.895.993.780.496.295.4
", + "image_path": "34258861fd43ae7f3c45230af57463d10280aee0f1356fdda421a0b429ee0643.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 605, + 477, + 617 + ], + "lines": [ + { + "bbox": [ + 114, + 605, + 477, + 617 + ], + "spans": [ + { + "bbox": [ + 114, + 605, + 477, + 617 + ], + "type": "text", + "content": "Table 3: OOD detection performance comparison with CSI and COD with layer selection." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 648, + 288, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 648, + 288, + 709 + ], + "spans": [ + { + "bbox": [ + 46, + 648, + 288, + 709 + ], + "type": "text", + "content": "obtained from the re-defined feature of the last layer, shows an AUROC score of over " + }, + { + "bbox": [ + 46, + 648, + 288, + 709 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 648, + 288, + 709 + ], + "type": "text", + "content": ", and the remaining blocks show AUROC scores of about " + }, + { + "bbox": [ + 46, + 648, + 288, + 709 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 648, + 288, + 709 + ], + "type": "text", + "content": ". Based on this, we choose only the last layer for " + }, + { + "bbox": [ + 46, + 648, + 288, + 709 + ], + "type": "inline_equation", + "content": "COD" + }, + { + "bbox": [ + 46, + 648, + 288, + 709 + ], + "type": "text", + "content": ", and perform OOD detection for ImageNet-30. ResNet18 is used for comparison" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 648, + 545, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 648, + 545, + 709 + ], + "spans": [ + { + "bbox": [ + 305, + 648, + 545, + 709 + ], + "type": "text", + "content": "with other approach. The baseline is a contrastive learning-based model, CSI [27], one of SOTA models. CSI is heavier to apply because it use a specific training method for OOD detection such as input transform and contrastive learning. On the other hand, " + }, + { + "bbox": [ + 305, + 648, + 545, + 709 + ], + "type": "inline_equation", + "content": "COD" + }, + { + "bbox": [ + 305, + 648, + 545, + 709 + ], + "type": "text", + "content": " is simple and easy to apply be" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16566" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "cause it does not require anything but generating classifiers in the training process. Nevertheless, as shown in Table 3, " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "COD" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": " shows comparable performance to CSI." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 119, + 180, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 119, + 180, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 119, + 180, + 133 + ], + "type": "text", + "content": "5.2. Classification Accuracy" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 139, + 288, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 139, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 139, + 288, + 270 + ], + "type": "text", + "content": "Some approaches for OOD detection suffered from the degradation of classification performance by change of model structures or loss functions [10, 21, 12]. In this case, two models should be stored aside, one for classification and the other for OOD detection. OOD detection models also need to have high classification performance for efficiency. Table 4 shows that models trained with the cosine-based softmax loss are accurate as much as the standard loss. Our approach can be applied to various model architectures because it only changes the calculation of logits. Also, there is no disadvantage to classification performance." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 65, + 282, + 271, + 373 + ], + "blocks": [ + { + "bbox": [ + 65, + 282, + 271, + 373 + ], + "lines": [ + { + "bbox": [ + 65, + 282, + 271, + 373 + ], + "spans": [ + { + "bbox": [ + 65, + 282, + 271, + 373 + ], + "type": "table", + "html": "
modelInDStandardCosine
ResNet34CIFAR-1095.2595.34
CIFAR-10077.7077.78
SVHN96.4896.41
DenseNet3CIFAR-1095.1495.08
CIFAR-10076.4376.87
SVHN96.4296.33
", + "image_path": "a742f2a013b0f038272f347b9780a50ed772cce7e61ff24d301563d88e6e8df9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 380, + 269, + 393 + ], + "lines": [ + { + "bbox": [ + 64, + 380, + 269, + 393 + ], + "spans": [ + { + "bbox": [ + 64, + 380, + 269, + 393 + ], + "type": "text", + "content": "Table 4: Comparison of classification performance" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 428, + 220, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 428, + 220, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 428, + 220, + 441 + ], + "type": "text", + "content": "5.3. Is the Input Processing Helpful?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 448, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 287, + 555 + ], + "type": "text", + "content": "The input processing is helpful for performance improvement when the perturbation size is properly chosen. We analyze the performance improvement by the perturbation size of the input processing. We train DenseNet3 with CIFAR-100 and observe the AUROC improvements with different perturbations for SVHN, LSUNc, and iSUN datasets. We perform ODIN and MAHA with pre-trained DenseNet3 with CIFAR-100. Figure 5 shows the improvement by the input processing on the three method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 556, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 652 + ], + "type": "text", + "content": "Each dataset has a different pattern of improvement. For example, our method is improved for SVHN but degraded for iSUN when the epsilon is large. ODIN and MAHA also have similar pattern to our method. This means that it is challenging to find the perturbation size that generally increases the performance regardless of OOD datasets. Hsu et al. tried to find the perturbation size without using OOD datasets, but the performance decreased in some cases [12]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "The input processing is helpful where the distribution of OOD samples are known. However, in reality, OOD samples can be generated from any distribution. The input processing significantly increases the amount of computation. It is reasonable not to use input processing, which increases" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 69, + 523, + 167 + ], + "blocks": [ + { + "bbox": [ + 306, + 69, + 523, + 167 + ], + "lines": [ + { + "bbox": [ + 306, + 69, + 523, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 69, + 523, + 167 + ], + "type": "image", + "image_path": "dc3048eef10c4b0a757c135fbe3c4743a0ef63cc441400a3ce133a54a33a3ffc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 182, + 523, + 281 + ], + "blocks": [ + { + "bbox": [ + 408, + 171, + 445, + 182 + ], + "lines": [ + { + "bbox": [ + 408, + 171, + 445, + 182 + ], + "spans": [ + { + "bbox": [ + 408, + 171, + 445, + 182 + ], + "type": "text", + "content": "(a)ODIN" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 182, + 523, + 281 + ], + "lines": [ + { + "bbox": [ + 306, + 182, + 523, + 281 + ], + "spans": [ + { + "bbox": [ + 306, + 182, + 523, + 281 + ], + "type": "image", + "image_path": "121e7516251752cea139f330627e2857ccda030f12714d37d457f382f6ef0d48.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 305, + 297, + 523, + 406 + ], + "blocks": [ + { + "bbox": [ + 405, + 285, + 447, + 296 + ], + "lines": [ + { + "bbox": [ + 405, + 285, + 447, + 296 + ], + "spans": [ + { + "bbox": [ + 405, + 285, + 447, + 296 + ], + "type": "text", + "content": "(b) MAHA" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 297, + 523, + 406 + ], + "lines": [ + { + "bbox": [ + 305, + 297, + 523, + 406 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 523, + 406 + ], + "type": "image", + "image_path": "5c2e22dd6992980ce4b832c6cf3d92b9197357fa5ef62f1baa5728fa8e8eabec.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 410, + 465, + 421 + ], + "lines": [ + { + "bbox": [ + 388, + 410, + 465, + 421 + ], + "spans": [ + { + "bbox": [ + 388, + 410, + 465, + 421 + ], + "type": "text", + "content": "(c) Proposed Method" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 431, + 545, + 455 + ], + "lines": [ + { + "bbox": [ + 305, + 431, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 305, + 431, + 545, + 455 + ], + "type": "text", + "content": "Figure 5: OOD detection performance (AUROC) increment by the perturbation size." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 489, + 544, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 489, + 544, + 512 + ], + "spans": [ + { + "bbox": [ + 305, + 489, + 544, + 512 + ], + "type": "text", + "content": "computation cost and has a potential to decrease performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 525, + 496, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 496, + 537 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 496, + 537 + ], + "type": "text", + "content": "5.4. Fine-tuning with Pre-trained Model" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 544, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 640 + ], + "type": "text", + "content": "We verify that our approach is also effective to fine-tuned models. We first train a model with the standard softmax, and then fine-tune the last block of the feature extractor and the classifier for only 20 epochs. In the fine-tuning process, all settings except for the learning rate are kept consistent with the previous configuration. The learning rate for fine-tuning starts with 0.5 and decays by a factor 0.1 at " + }, + { + "bbox": [ + 304, + 544, + 545, + 640 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 544, + 545, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 544, + 545, + 640 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 544, + 545, + 640 + ], + "type": "text", + "content": " of the training progress." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "We can notify that our approach still works for fine-tuned models. The OOD detection performances are comparable to or even higher than those of our original models, as shown in Table 5. The DenseNet3 trained with the standard softmax for SVHN shows a classification accuracy of " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "96.4\\%" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": ", but its fine-tuned model shows an accuracy of" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16567" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 289, + 316 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 289, + 316 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 289, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 289, + 316 + ], + "type": "table", + "html": "
InDOODAUROCTNR95
CIFAR-10 (Acc: 95.1 → 94.1)SVHN99.4598.35
TINc99.5399.61
TINr99.7399.23
LSUNc99.8299.96
LSUNr99.7399.34
iSUN99.7299.27
CIFAR-100 (Acc: 76.4 → 75.7)SVHN96.3979.48
TINc97.5388.87
TINr98.0389.37
LSUNc98.7296.81
LSUNr98.1690.91
iSUN97.9988.76
SVHN (Acc: 96.4 → 96.1)CIFAR-1099.1798.18
CIFAR-10099.2298.33
TINc99.94100.00
TINr99.8899.96
LSUNc99.94100.00
LSUNr99.91100.00
iSUN99.9199.99
", + "image_path": "2b5d4fa8717eb08d949d4d2b85cc2b1eb78c7a469ab2095fa67b091b003ea076.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 324, + 288, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 324, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 288, + 361 + ], + "type": "text", + "content": "Table 5: Classification and OOD detection performance of fine-tuned DenseNet3 models. All figures are in %, and are the average results of 5 experiments each." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 401, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 450 + ], + "type": "inline_equation", + "content": "96.1\\%" + }, + { + "bbox": [ + 46, + 401, + 287, + 450 + ], + "type": "text", + "content": ". Even though the models are fine-tuned with a limited number of epochs, the classification accuracy and the OOD detection performances remain almost same. The results with ResNet34 are also presented in the appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 472, + 119, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 119, + 485 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 119, + 485 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 497, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 581 + ], + "type": "text", + "content": "We proposed a novel OOD detection method that can be applied to deep neural network-based classifiers. It is simple and fast since it uses only the feature norm and the Mahalanobis distances. Our approach does not use additional data, does not have hyperparameters, and does not perform input processing. Our approach is easy to apply because it only needs to apply the cosine-based softmax loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 605, + 149, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 149, + 619 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 149, + 619 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-00421, AI Graduate School Support Program (Sungkyunkwan University)), and also supported by the MSIT under the ICT Creative Consilience Program (IITP-2023-2020-0-01821) supervised by the IITP." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 72, + 365, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 365, + 83 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 365, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 91, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 311, + 91, + 545, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 91, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 311, + 91, + 545, + 145 + ], + "type": "text", + "content": "[1] Liron Bergman and Yedid Hoshen. Classification-based anomaly detection for general data. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 311, + 148, + 547, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 148, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 311, + 148, + 547, + 182 + ], + "type": "text", + "content": "[2] Wenhu Chen, Yilin Shen, Hongxia Jin, and William Wang. A variational dirichlet framework for out-of-distribution detection. arXiv preprint arXiv:1811.07308, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 311, + 183, + 547, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 183, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 311, + 183, + 547, + 227 + ], + "type": "text", + "content": "[3] Sung-Ik Choi and Sae-Young Chung. Novelty detection via blurring. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 228, + 547, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 228, + 547, + 271 + ], + "spans": [ + { + "bbox": [ + 311, + 228, + 547, + 271 + ], + "type": "text", + "content": "[4] Erik Daxberger and José Miguel Hernández-Lobato. Bayesian variational autoencoders for unsupervised out-of-distribution detection. arXiv preprint arXiv:1912.05651, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 274, + 547, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 274, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 311, + 274, + 547, + 340 + ], + "type": "text", + "content": "[5] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 4690-4699. Computer Vision Foundation / IEEE, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 342, + 547, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 342, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 311, + 342, + 547, + 418 + ], + "type": "text", + "content": "[6] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016, volume 48 of JMLR Workshop and Conference Proceedings, pages 1050-1059. JMLR.org, 2016. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 420, + 547, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 420, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 311, + 420, + 547, + 485 + ], + "type": "text", + "content": "[7] Izhak Golan and Ran El-Yaniv. Deep anomaly detection using geometric transformations. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada, pages 9781-9791, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 488, + 547, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 488, + 547, + 543 + ], + "spans": [ + { + "bbox": [ + 311, + 488, + 547, + 543 + ], + "type": "text", + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society, 2016. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 544, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 544, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 311, + 544, + 545, + 600 + ], + "type": "text", + "content": "[9] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "type": "text", + "content": "[10] Dan Hendrycks, Mantas Mazeika, and Thomas G. Dietterich. Deep anomaly detection with outlier exposure. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. 1, 2, 5, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "type": "text", + "content": "[11] Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self-supervised learning can improve model robustness and uncertainty. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16568" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "2019, December 8-14, 2019, Vancouver, BC, Canada, pages 15637-15648, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 163 + ], + "type": "text", + "content": "[12] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized ODIN: detecting out-of-distribution image without learning from out-of-distribution data. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 10948-10957. IEEE, 2020. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 288, + 229 + ], + "type": "text", + "content": "[13] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 2261-2269. IEEE Computer Society, 2017. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 231, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 288, + 262 + ], + "type": "text", + "content": "[14] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, 2009. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 342 + ], + "type": "text", + "content": "[15] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada, pages 7167-7177, 2018. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 343, + 288, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 288, + 409 + ], + "type": "text", + "content": "[16] Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net, 2018. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 410, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 288, + 453 + ], + "type": "text", + "content": "[17] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In Advances in Neural Information Processing Systems, volume 33, pages 21464-21475. Curran Associates, Inc., 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 455, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 288, + 521 + ], + "type": "text", + "content": "[18] Andrey Malinin and Mark J. F. Gales. Predictive uncertainty estimation via prior networks. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada, pages 7047-7058, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 522, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 567 + ], + "type": "text", + "content": "[19] Eric Nalisnick, Akihiro Matsukawa, Yee Whye Teh, and Balaji Lakshminarayanan. Detecting out-of-distribution inputs to deep generative models using a test for typicality. arXiv preprint arXiv:1906.02994, 5:5, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 568, + 288, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 622 + ], + "type": "text", + "content": "[20] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 624, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 668 + ], + "type": "text", + "content": "[21] Aristotelis-Angelos Papadopoulos, Mohammad Reza Rajati, Nazim Shaikh, and Jiamian Wang. Outlier exposure with confidence control for out-of-distribution detection. Neurocomputing, 441:138-150, 2021. 1, 2, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "type": "text", + "content": "[22] Jie Ren, Peter J. Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark A. DePristo, Joshua V. Dillon, and Balaji Lakshminarayanan. Likelihood ratios for out-of-distribution detection. In Advances in Neural Information Processing Sys-" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 508 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "tems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 14680-14691, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 107, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 173 + ], + "type": "text", + "content": "[23] Chandramouli Shama Sastry and Sageev Oore. Detecting out-of-distribution examples with gram matrices. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 8491-8501. PMLR, 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "type": "text", + "content": "[24] Vikash Sehwag, Mung Chiang, and Prateek Mittal. " + }, + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{SSD}\\}" + }, + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "type": "text", + "content": ": A unified framework for self-supervised outlier detection. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 219, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 284 + ], + "type": "text", + "content": "[25] Joan Serrà, David Álvarez, Vicenç Gómez, Olga Slizovskaia, José F. Núñez, and Jordi Luque. Input complexity and out-of-distribution detection with likelihood-based generative models. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 285, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 545, + 318 + ], + "type": "text", + "content": "[26] Jiaming Song, Yang Song, and Stefano Ermon. Unsupervised out-of-distribution detection with batch normalization. arXiv preprint arXiv:1910.09115, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "type": "text", + "content": "[27] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. In Advances in Neural Information Processing Systems, volume 33, pages 11839-11852. Curran Associates, Inc., 2020. 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 375, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 375, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 375, + 545, + 441 + ], + "type": "text", + "content": "[28] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 5265-5274. IEEE Computer Society, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 442, + 545, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 442, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 308, + 442, + 545, + 508 + ], + "type": "text", + "content": "[29] Xiao Zhang, Rui Zhao, Yu Qiao, Xiaogang Wang, and Hongsheng Li. Adacos: Adaptively scaling cosine logits for effectively learning deep face representations. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 10823-10832. Computer Vision Foundation / IEEE, 2019. 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16569" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SimpleClick_ Interactive Image Segmentation with Simple Vision Transformers/2b87aeaa-0b31-4e94-84d5-d9bfaea7e328_content_list.json b/2023/SimpleClick_ Interactive Image Segmentation with Simple Vision Transformers/2b87aeaa-0b31-4e94-84d5-d9bfaea7e328_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0cdb265244fe24e569858ca96c3a639f0429fef0 --- /dev/null +++ b/2023/SimpleClick_ Interactive Image Segmentation with Simple Vision Transformers/2b87aeaa-0b31-4e94-84d5-d9bfaea7e328_content_list.json @@ -0,0 +1,1432 @@ +[ + { + "type": "text", + "text": "SimpleClick: Interactive Image Segmentation with Simple Vision Transformers", + "text_level": 1, + "bbox": [ + 84, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qin Liu, Zhenlin Xu, Gedas Bertasius, Marc Niethammer \nUniversity of North Carolina at Chapel Hill", + "bbox": [ + 256, + 180, + 712, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/uncbiag/SimpleClick", + "bbox": [ + 315, + 219, + 653, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 312, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Click-based interactive image segmentation aims at extracting objects with a limited user clicking. A hierarchical backbone is the de-facto architecture for current methods. Recently, the plain, non-hierarchical Vision Transformer (ViT) has emerged as a competitive backbone for dense prediction tasks. This design allows the original ViT to be a foundation model that can be finetuned for downstream tasks without redesigning a hierarchical backbone for pretraining. Although this design is simple and has been proven effective, it has not yet been explored for interactive image segmentation. To fill this gap, we propose SimpleClick, the first interactive segmentation method that leverages a plain backbone. Based on the plain backbone, we introduce a symmetric patch embedding layer that encodes clicks into the backbone with minor modifications to the backbone itself. With the plain backbone pretrained as a masked autoencoder (MAE), SimpleClick achieves state-of-the-art performance. Remarkably, our method achieves 4.15 NoC@90 on SBD, improving $21.8\\%$ over the previous best result. Extensive evaluation on medical images demonstrates the generalizability of our method. We provide a detailed computational analysis, highlighting the suitability of our method as a practical annotation tool.", + "bbox": [ + 75, + 301, + 473, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 678, + 209, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The goal of interactive image segmentation is to obtain high-quality pixel-level annotations with limited user interaction such as clicking. Interactive image segmentation approaches have been widely applied to annotate large-scale image datasets, which drive the success of deep models in various applications, including video understanding [5, 49], self-driving [7], and medical imaging [32, 41]. Much research has been devoted to explore interactive image segmentation with different interaction types, such as bounding boxes [47], polygons [1], clicks [43], scribbles [45], and their combinations [51]. Among them, the click-based approach is most common due to its simplicity and well-established training and evaluation protocols.", + "bbox": [ + 75, + 703, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0a4bb07328405001327f2a24e63f407a2e00e221145706dc4c0f640e4c9cdf4b.jpg", + "image_caption": [ + "NoC@90 on SBD dataset", + "Figure 1. Interactive segmentation results on SBD [20]. The metric \"NoC@90\" denotes the number of clicks required to obtain $90\\%$ IoU. The area of each bubble is proportional to the FLOPs of a model variant (Tab. 5). We show that plain ViTs outperform all hierarchical backbones for interactive image segmentation at a moderate computational cost." + ], + "image_footnote": [], + "bbox": [ + 503, + 287, + 888, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in click-based approaches mainly lie in two orthogonal directions: 1) the development of more effective backbone networks and 2) the exploration of more elaborate refinement modules built upon the backbone. For the former direction, different hierarchical backbones, including both ConvNets [30,43] and ViTs [10,33], have been developed for interactive segmentation. For the latter direction, various refinement modules, including local refinement [10,30] and click imitation [34], have been proposed to further boost segmentation performance. In this work, we delve into the former direction and focus on exploring a plain backbone for interactive segmentation.", + "bbox": [ + 496, + 627, + 893, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A hierarchical backbone is the predominant architecture for current interactive segmentation methods. This design is deeply rooted in ConvNets, represented by ResNet [22], and has been adopted by ViTs, represented by the Swin Transformer [35]. The motivation for a hierarchical backbone stems from the locality of convolution operations, leading", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22290", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to insufficient model receptive field size without the hierarchy. To increase the receptive field size, ConvNets have to progressively downsample feature maps to capture more global contextual information. Therefore, they often require a feature pyramid network such as FPN [28] to aggregate multi-scale representations for high-quality segmentation. However, this reasoning no longer applies for a plain ViT, in which global information can be captured from the first self-attention block. Because all feature maps in the ViT are of the same resolution, the motivation for an FPN-like feature pyramid also no longer remains. The above reasoning is supported by a recent finding that a plain ViT can serve as a strong backbone for object detection [26]. This finding indicates a general-purpose ViT backbone might be suitable for other tasks, which then can decouple pretraining from finetuning and transfer the benefits from readily available pretrained ViT models (e.g. MAE [21]) to these tasks. However, although this design is simple and has been proven effective, it has not yet been explored in interactive segmentation. In this work, we propose SimpleClick, the first plain-backbone method for interactive segmentation. The core of SimpleClick is a plain ViT backbone that maintains single-scale representations throughout. We only use the last feature map from the plain backbone to build a simple feature pyramid for segmentation, largely decoupling the general-purpose backbone from the segmentation-specific modules. To make SimpleClick more efficient, we use a light-weight MLP decoder to transform the simple feature pyramid into a segmentation (see Sec. 3 for details).", + "bbox": [ + 76, + 90, + 472, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We extensively evaluate our method on 10 public benchmarks, including both natural and medical images. With the plain backbone pretrained as a MAE [21], our method achieves 4.15 NoC@90 on SBD, which outperforms the previous best method by $21.8\\%$ without a complex FPN-like design and local refinement. We demonstrate the generalizability of our method by out-of-domain evaluation on medical images. We further analyze the computational efficiency of SimpleClick, highlighting its suitability as a practical annotation tool.", + "bbox": [ + 76, + 529, + 472, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are:", + "bbox": [ + 96, + 680, + 282, + 694 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose SimpleClick, the first plain-backbone method for interactive image segmentation.", + "- SimpleClick achieves state-of-the-art performance on natural images and shows strong generalizability on medical images.", + "- SimpleClick meets the computational efficiency requirement for a practical annotation tool, highlighting its readiness for real-world applications." + ], + "bbox": [ + 78, + 696, + 468, + 815 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 829, + 218, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interactive Image Segmentation Interactive image segmentation is a longstanding problem for which increasingly better solution approaches have been proposed. Early", + "bbox": [ + 76, + 854, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "works [6, 16, 18, 40] tackle this problem using graphs defined over image pixels. However, these methods only focus on low-level image features and therefore tend to have difficulty with complex objects.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thriving on large datasets, ConvNets [10, 30, 43, 47, 48] have evolved as the dominant architecture for high-quality interactive segmentation. ConvNet-based methods have explored various interaction types, such as bounding boxes [47], polygons [1], clicks [43], and scribbles [45]. Click-based approaches are the most common due to their simplicity and well-established training and evaluation protocols. Xu et al. [48] first proposed a click simulation strategy that has been adopted by follow-up work [10, 34, 43]. DEXTR [36] extracts a target object by specifying its four extreme points (left-most, right-most, top, and bottom pixels). FCA-Net [31] demonstrates the critical role of the first click for better segmentation. Recently, ViTs have been applied to interactive segmentation. FocalClick [10] uses SegFormer [46] as the backbone network and achieves state-of-the-art segmentation results with high computational efficiency. iSegFormer [33] uses a Swin Transformer [35] as the backbone network for interactive segmentation on medical images. Besides the contribution of backbones, some works are exploring elaborate refinement modules built upon the backbone. FocalClick [10] and FocusCut [30] propose similar local refinement modules for high-quality segmentation. PseudoClick [34] proposes a click-imitation mechanism by estimating the next click to further reduce human annotation cost. Our method differs from all previous click-based methods in its plain, non-hierarchical ViT backbone, enjoying the benefits from readily available pretrained ViT models (e.g. MAE [21]).", + "bbox": [ + 496, + 155, + 892, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Transformers for Non-Interactive Segmentation Recently, ViT-based approaches [17, 24, 44, 46, 50] have shown competitive performance on segmentation tasks compared to ConvNets. The original ViT [13] is a non-hierarchical architecture that only maintains single-scale feature maps throughout. SETR [52] and Segmenter [44] use the original ViT as the encoder for semantic segmentation. To allow for more efficient segmentation, the Swin Transformer [35] reintroduces a computational hierarchy into the original ViT architecture using shifted window attention, leading to a highly efficient hierarchical ViT backbone. SegFormer [46] designs hierarchical feature representations based on the original ViT using overlapped patch merging, combined with a light-weight MLP decoder for efficient segmentation. HRViT [17] integrates a high-resolution multi-branch architecture with ViTs to learn multi-scale representations. Recently, the original ViT has been reintroduced as a competitive backbone for semantic segmentation [8] and object detection [26], with the aid of MAE [21] pretraining and window attention. Inspired by this finding, we explore using a plain ViT as the backbone", + "bbox": [ + 496, + 583, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22291", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "network for interactive segmentation.", + "bbox": [ + 76, + 90, + 326, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 125, + 169, + 142 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is not to propose new modules but to adapt a plain-ViT backbone for interactive segmentation with minimal modifications. Fig. 2 shows the overview of our method. Sec. 3.1 briefly describes the plain segmentation backbone. Sec. 3.2 shows how to adapt the backbone for interactive segmentation. Sec. 3.3 introduces other modules of SimpleClick. Sec. 3.4 describes the training and inference details of our method.", + "bbox": [ + 75, + 152, + 472, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Plain Segmentation Backbone", + "text_level": 1, + "bbox": [ + 76, + 289, + 344, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use a plain ViT [13] as the segmentation backbone. Unlike previous hierarchical backbones, a plain ViT only maintains single-scale feature maps throughout. Given an image, the patch embedding layer of the plain ViT first divides the image into non-overlapping fixed-size patches (e.g. $16 \\times 16$ ). All patches are then flattened and linearly projected to a sequence of fixed-length tokens before feeding into self-attention blocks. These output tokens will be reshaped and upsampled to match the spatial size of the input image for segmentation. In this work, we consider three standard ViT backbones: ViT-B, ViT-L, and ViT-H (Tab. 1 shows the number of parameters of these backbones). For stable and efficient training, we use readily-available MAEs [21] as the pretrained weights for these backbones.", + "bbox": [ + 75, + 314, + 472, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Clicks Encoding for Segmentation Backbone", + "text_level": 1, + "bbox": [ + 76, + 541, + 455, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We adapt the above plain segmentation backbone for interactive segmentation by turning user interactions into a form of guidance learned by the network. Similar to previous click-based methods [34, 43], we use clicks as a form of user interaction and encode each click as a small disk on a two-channel masks [34, 43] (one channel for positive clicks while the other for negative ones; positive clicks should be put in the foreground, while the negative ones should be put in the background). We automatically simulate human clicks for efficient training and evaluation. The clicks simulation process is described in Sec.3.4. Note that for human evaluation, a human-in-the-loop will provide all the clicks.", + "bbox": [ + 75, + 566, + 468, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given human clicks or simulated clicks, we now introduce how to fuse them into the plain backbone. We first introduce a patch-embedding layer that is symmetric to the one in the backbone. Note that we also concatenate the previous segmentation to the clicks map as an additional channel for better performance. The two symmetric embedding layers operate on the image and the clicks map, respectively. The inputs are patchified, flattened, and projected to two vector sequences of the same dimension, followed by element-wise addition before feeding into the self-attention blocks.", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/f704b14aca42558be0a85823be26f4a4261db16d593eaa95549d8e8b9d6307a7.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Number of parameters of our models. The unit is a million. The percentage of parameters is shown in bracket. Most parameters are used by the ViT backbone." + ], + "table_body": "
Model↓ Module→ViT BackboneConv. NeckMLP Head
Ours-ViT-B (base)83.0 (89.3%)9.0 (9.7%)0.9 (1.0%)
Ours-ViT-L (large)290.8 (94.3%)16.5 (5.3%)1.1 (0.4%)
Ours-ViT-H (huge)604.0 (95.7%)25.8 (4.1%)1.3 (0.2%)
", + "bbox": [ + 506, + 89, + 885, + 146 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Other Modules", + "text_level": 1, + "bbox": [ + 500, + 213, + 653, + 228 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Simple Feature Pyramid For the hierarchical backbone, a feature pyramid is commonly produced by an FPN [28] to combine features from different stages. For the plain backbone, a feature pyramid can be generated in a much simpler way: by a set of parallel convolutional or deconvolutional layers using only the last feature map of the backbone. As shown in Fig. 2, given the input ViT feature map, a multiscale feature map can be produced by four convolutions with different strides. Though the effectiveness of this simple feature pyramid design is first demonstrated in ViTDet [26] for object detection, we show in this work the effectiveness of this simple feature pyramid design for interactive segmentation. We also propose several additional variants (Fig. 6) as part of an ablation study (Sec. 4.4).", + "bbox": [ + 496, + 237, + 893, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All-MLP Segmentation Head We implement a lightweight segmentation head using only MLP layers. It takes in the simple feature pyramid and produces a segmentation probability map of scale 1/4, followed by an upsampling operation to recover the original resolution. Note that this segmentation head avoids computationally demanding components and only accounts for up to $1\\%$ of the model parameters (Tab. 1). The key insight is that with a powerful pretrained backbone, a lightweight segmentation head is sufficient for interactive segmentation. The proposed all-MLP segmentation head works in three steps. First, each feature map from the simple feature pyramid goes through an MLP layer to transform it to an identical channel dimension (i.e. $C_2$ in Fig. 2). Second, all feature maps are upsampled to the same resolution (i.e. 1/4 in Fig. 2) for concatenation. Third, the concatenated features are fused by another MLP layer to produce a single-channel feature map, followed by a sigmoid function to obtain a segmentation probability map, which is then transformed to a binary segmentation given a predefined threshold (i.e. 0.5).", + "bbox": [ + 496, + 449, + 893, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.4. Training and Inference Settings", + "text_level": 1, + "bbox": [ + 500, + 763, + 779, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Backbone Pretraining Our backbone models are pretrained as MAEs [21] on ImageNet-1K [11]. In MAE pretraining, the ViT models reconstruct the randomly masked pixels of images while learning a universal representation. This simple self-supervised approach turns out to be an efficient and", + "bbox": [ + 498, + 787, + 893, + 864 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "This probability map may be miscalibrated and can be improved by calibration approaches [12].", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "footer", + "text": "22292", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8c33132e78719e53e82814db17c1001d00504536e4a869b37db8cd9742315476.jpg", + "image_caption": [ + "Figure 2. SimpleClick overview. Our method consists of three main modules: (a) a plain ViT backbone that maintains single-scale feature maps throughout; (b) a multi-scale simple feature pyramid that is generated from the last feature maps of the backbone by four parallel convolution or deconvolution layers; (c) a light-weight MLP decoder for segmentation. We also input the previous segmentation to improve the performance and to allow clicking from existing masks. User clicks are encoded as a two-channel disk map, combined with the previous segmentation as input. Boxes in blue are intermediate feature maps. The positional encoding is not shown for brevity." + ], + "image_footnote": [], + "bbox": [ + 187, + 95, + 781, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "scalable way to pretrain ViT models [21]. In this work, we do not perform pretraining ourselves. Instead, we simply use the readily available pretrained MAE weights from [21].", + "bbox": [ + 75, + 438, + 470, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Clicks Simulation and End-to-end Finetuning With the pretrained backbone, we finetune our model end-to-end on the interactive segmentation task. The finetuning pipeline can be briefly described as follows. First, we automatically simulate clicks based on the current segmentation and gold standard segmentation, without a human-in-the-loop providing the clicks. Specifically, we use a combination of random and iterative click simulation strategies, inspired by RITM [43]. The random click simulation strategy generates clicks in parallel, without considering the order of the clicks. The iterative click simulation strategy generates clicks iteratively, where the next click should be placed on the erroneous region of a prediction that was obtained using the previous clicks. This strategy is more similar to human clicking behavior. Second, we incorporate the segmentation from the previous interaction as an additional input for the backbone, further improving the segmentation quality. This also allows our method to refine from an existing segmentation, which is a desired feature for a practical annotation tool. We use the normalized focal loss [43] (NFL) to train all our models. Previous works [10, 43] show that NFL converges faster and achieves better performance than the widely used binary cross entropy loss for interactive segmentation tasks. Similar training pipelines have been proposed by RITM [43] and its follow-up works [9, 10, 34].", + "bbox": [ + 75, + 488, + 470, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human Evaluation and Automatic Evaluation There are two inference modes: automatic evaluation and human eval", + "bbox": [ + 76, + 869, + 470, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "uation. For automatic evaluation, clicks are automatically simulated based on the current segmentation and gold standard. For human evaluation, a human-in-the-loop provides all clicks based on their subjective evaluation of current segmentation results. We use automatic evaluation for quantitative analyses and human evaluation for a qualitative assessment of the interactive segmentation behavior.", + "bbox": [ + 496, + 438, + 893, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 558, + 633, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets We conducted experiments on 10 public datasets including 7 natural image datasets and 3 medical datasets. The details are as follows:", + "bbox": [ + 496, + 583, + 890, + 627 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- GrabCut [40]: 50 images (50 instances), each with clear foreground and background differences.", + "- Berkeley [37]: 96 images (100 instances); this dataset shares a small portion of images with GrabCut.", + "- DAVIS [39]: 50 videos; we only use the same 345 frames as used in [10, 30, 34, 43] for evaluation.", + "- Pascal VOC [14]: 1449 images (3427 instances) in the validation set. We only test on the validation set.", + "- SBD [20]: 8498 training images (20172 instances) and 2857 validation images (6671 instances). Following previous works [10, 30, 43], we train our model on the training set and evaluate on the validation set.", + "- COCO [29]+LVIS [19] (C+L): COCO contains 118K training images (1.2M instances); LVIS shares the same images with COCO but has much higher segmentation quality. We combine the two datasets for training.", + "- ssTEM [15]: two image stacks, each contains 20 medical images. We use the same stack that was used in [34]." + ], + "bbox": [ + 500, + 628, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22293", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/95a272e72935b50e99c599d489175aead5a44503e4efee2e376f3def16a5fb21.jpg", + "image_caption": [ + "Figure 3. Convergence analysis for models trained on either SBD [20] or COCO [29] + LVIS [19] (C+L). We report results on SBD [20] and Pascal VOC [14]. The metric is mean IoU given $k$ clicks ( $\\text{mIoU} @ k$ ). Our models in general require fewer clicks for a given accuracy level." + ], + "image_footnote": [], + "bbox": [ + 83, + 99, + 465, + 297 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/403668dc1826ba85844aa3b720a498bf1b1e2ccba3df7c6302fa3ec6c264c16e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 99, + 857, + 296 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e33e5d653278fcbac5a597e499b7054ab0fa52f629380660cbee6302b4ac2494.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneGrabCutBerkeleySBDDAVISPascal VOC
NoC85NoC90NoC85NoC90NoC85NoC90NoC85NoC90NoC85NoC90
\\( \\nmid \\)DIOS [48] \\( _{CVPR16} \\)FCN-6.04-8.65---12.586.88-
\\( \\nmid \\)FCA-Net [31] \\( _{CVPR20} \\)ResNet-101-2.08-3.92---7.572.69-
\\( \\nmid \\)LD [27] \\( _{CVPR18} \\)VGG-193.204.79--7.4110.785.059.57--
\\( \\nmid \\)BRS [23] \\( _{CVPR19} \\)DenseNet2.603.60-5.086.599.785.588.24--
\\( \\nmid \\)f-BRS [42] \\( _{CVPR20} \\)ResNet-1012.302.72-4.574.817.735.047.41--
\\( \\nmid \\)RITM [43] \\( _{Preprint21} \\)HRNet-181.762.041.873.223.395.434.946.712.513.03
\\( \\nmid \\)CDNet [9] \\( _{ICCV21} \\)ResNet-341.862.181.953.275.187.895.006.893.614.51
\\( \\nmid \\)PseudoClick [34] \\( _{ECCV22} \\)HRNet-181.682.041.853.233.385.404.816.572.342.74
\\( \\nmid \\)FocalClick [10] \\( _{CVPR22} \\)HRNet-18s1.862.06-3.144.306.524.926.48--
\\( \\nmid \\)FocalClick [10] \\( _{CVPR22} \\)SegF-B01.661.90-3.144.346.515.027.06--
\\( \\nmid \\)FocusCut [30] \\( _{CVPR22} \\)ResNet-501.601.78\\( 1.85^{\\dagger } \\)3.443.625.665.006.38--
\\( \\nmid \\)FocusCut [30] \\( _{CVPR22} \\)ResNet-1011.461.64\\( 1.81^{\\dagger } \\)3.013.405.314.856.22--
\\( \\nmid \\)OursViT-B1.401.541.442.463.285.244.105.482.382.81
\\( \\nmid \\)OursViT-L1.381.461.402.332.694.464.125.391.952.30
\\( \\nmid \\)OursViT-H1.321.441.362.092.514.154.205.341.882.20
\\( \\nmid \\)RITM [43] \\( _{Preprint21} \\)HRNet-321.461.561.432.103.595.714.115.342.192.57
\\( \\nmid \\)CDNet [9] \\( _{ICCV21} \\)ResNet-341.401.521.472.064.307.044.275.562.743.30
\\( \\nmid \\)PseudoClick [34] \\( _{ECCV22} \\)HRNet-321.361.501.402.083.465.543.795.111.942.25
\\( \\nmid \\)FocalClick [10] \\( _{CVPR22} \\)SegF-B01.401.661.592.274.566.864.045.492.973.52
\\( \\nmid \\)FocalClick [10] \\( _{CVPR22} \\)SegF-B31.441.501.551.923.535.593.614.902.462.88
\\( \\nmid \\)OursViT-B1.381.481.361.973.435.623.665.062.062.38
\\( \\nmid \\)OursViT-L1.321.401.341.892.954.893.264.811.721.96
\\( \\nmid \\)OursViT-H1.381.501.361.752.854.703.414.781.761.98
", + "bbox": [ + 76, + 352, + 895, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Comparison with previous results. We report results on five benchmarks: GrabCut [40], Berkeley [37], SBD [20], DAVIS [39], and Pascal VOC [14]. The best results are set in bold. $\\nsubseteq$ denotes a model trained on Pascal; $\\downarrow$ denotes a model trained on SBD; $\\sqcup$ denotes a model trained on COCO [29] + LVIS [19] (C+L); $\\dagger$ denotes a number reproduced by the released or retrained models. Our models achieve state-of-the-art performance on all benchmarks.", + "bbox": [ + 75, + 680, + 893, + 736 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- BraTS [4]: 369 magnetic resonance image (MRI) volumes; we test on the same 369 slices used in [34].", + "- OAIZIB [2]: 507 MRI volumes; we test on the same 150 slices (300 instances) as used in [33]." + ], + "bbox": [ + 76, + 760, + 467, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics Following previous works [30,42,43], we automatically simulate user clicks by comparing the current segmentation with the gold standard. In this simulation, the next click will be put at the center of the region with the largest error. We use the Number of Clicks (NoC)", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as the evaluation metric to calculate the number of clicks required to achieve a target Intersection over Union (IoU). We set two target IoUs: $85\\%$ and $90\\%$ , represented by $\\mathrm{NoC}\\% 85$ and $\\mathrm{NoC}\\% 90$ respectively. The maximum number of clicks for each instance is set to 20. We also use the average IoU given $k$ clicks ( $\\mathrm{mIoU} @ k$ ) as an evaluation metric to measure the segmentation quality given a fixed number of clicks.", + "bbox": [ + 496, + 760, + 890, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details We implement our models using Python and PyTorch [38]. We implement three models", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22294", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d708b42d1e3d20a3f9719d6baca1cb8d5099ca4dc24f23ccaf7dfb58edc3db65.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 95, + 336, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f272ee5755fba79b1d467fdc717a5c1fe2a925081f2a7f26b38a04addfe55e92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 95, + 619, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7d956fb6852536a52d07d67e800ed4ac655d3a1aa9baa93f30097890a0c1c673.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 95, + 897, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f0f6086146e3d93c579d3ae0f7cfa6ced33039a26070565b6a24aabcae8000dd.jpg", + "image_caption": [ + "Figure 4. Histogram analysis of IoU given a predefined number of clicks $k$ (IoU@k). We report analysis on SBD [20] and Pascal VOC [14] with models trained on SBD. Compared with the two baselines, our models achieve higher-quality segmentation with fewer failure cases." + ], + "image_footnote": [], + "bbox": [ + 76, + 239, + 338, + 378 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/82c52d1ea5aee0d2d8ef34005a7dac3ec029c5fdc6eeae2081e2e1c661e6f726.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 239, + 619, + 378 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/93f392859a70a3b67630aa0a53ab74e3522b235574105659a481aab2b9682f6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 239, + 898, + 378 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/156383fcea46ded5d4d404bcd1c98849c93f8ba081855f79ad8bcc76c295ffe0.jpg", + "image_caption": [ + "Figure 5. Convergence analysis for two medical image datasets: BraTS [4] and OAIZIB [2]. Models are trained on either SBD [20] or COCO [29] + LVIS [19] (denoted as $\\mathrm{C} + \\mathrm{L}$ ). The metric is mean IoU given $k$ clicks. Overall, our models require fewer clicks for a given accuracy level. The performance gain is more prominent for bigger models (e.g. ViT-H) or larger training sets (e.g. C+L)." + ], + "image_footnote": [], + "bbox": [ + 84, + 433, + 457, + 628 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b6940d6819f52f98d051c1c61d91f6ed4de83aef58d207e70fdc409e394d0563.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 433, + 854, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "based on three vanilla ViT models (i.e. ViT-B, ViT-L, and ViT-H). These backbone models are initialized with the MAE pretrained weights, and then are finetuned end-to-end with other modules. We train our models on either SBD or COCO+LVIS with 55 epochs; the initial learning rate is set to $5 \\times 10^{-5}$ and decreases to $5 \\times 10^{-6}$ after epoch 50. We set the batch size to 140 for ViT-Base, 72 for ViT-Large, and 32 for ViT-Huge to fit the models into GPU memory. All our models are trained on four NVIDIA RTX A6000 GPUs. We use the following data augmentation techniques: random resizing (scale range from 0.75 to 1.25), random flipping and rotation, random brightness contrast, and random cropping. Though the ViT backbone was pretrained on images of size", + "bbox": [ + 75, + 700, + 472, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$224 \\times 224$ , we finetune on $448 \\times 448$ with non-shifting window attention for better performance. We optimize using Adam with $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ .", + "bbox": [ + 496, + 700, + 890, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison with Previous Results", + "text_level": 1, + "bbox": [ + 498, + 756, + 802, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show in Tab. 2 the comparisons with previous state-of-the-art results. Our models achieve the best performance on all the five benchmarks. Remarkably, when trained on SBD training set, our ViT-H model achieves 4.15 NoC@90 on the SBD validation set, outperforming the previous best score by $21.8\\%$ . Since the SBD validation set contains the largest number of instances (6671 instances) among the five benchmarks, this improvement is convincing. When trained", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22295", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/98b09e19ef34e827c4e6a4417e80a123a6645cf5451a1977db5dca3696c1f170.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Out-of-domain evaluation on three medical image datasets: ssTEM [15], BraTS [4], and OAIZIB [2]. Our models generalize very well on the three datasets, without finetuning." + ], + "table_body": "
ModelssTEM\nmIoU@10BraTS\nmIoU@10 / 20OAIZIB\nmIoU@10 / 20
↓RITM-H18 [43]93.1587.05 / 90.4771.04 / 78.52
↓CDN-RN34 [9]66.7258.34 / 82.0738.07 / 61.17
↓RITM-H32 [43]94.1188.34 / 89.2575.27 / 75.18
↓CDN-RN34 [9]88.4680.24 / 86.6363.19 / 74.21
↓FC-SF-B0 [10]92.6286.02 / 90.7474.08 / 79.14
↓FC-SF-B3 [10]93.6188.62 / 90.5875.77 / 80.08
↓Ours-ViT-B93.7286.98 / 90.6776.05 / 79.61
↓Ours-ViT-L94.3488.43 / 90.8477.34 / 79.97
↓Ours-ViT-H94.0888.98 / 91.0077.50 / 80.10
", + "bbox": [ + 78, + 88, + 470, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on COCO+LVIS, our models also achieve state-of-the-art performance on all benchmarks. Fig. 7 shows several segmentation cases on DAVIS, including the worst case. Note that the DAVIS dataset requires high-quality segmentations because all its instances have a high-quality gold standard. Our models still achieve the state-of-the-art on DAVIS without using specific modules, such as a local refinement module [10], which is beneficial for high-quality segmentation. Fig. 3 shows that our method converges better than other methods with sufficient clicks, leading to fewer failure cases as shown in Fig. 4. We only report results on SBD and Pascal VOC, the top two largest datasets.", + "bbox": [ + 76, + 301, + 468, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Out-of-Domain Evaluation on Medical Images", + "text_level": 1, + "bbox": [ + 76, + 494, + 468, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further evaluate the generalizability of our models on three medical image datasets: ssTEM [15], BraTS [3], and OAIZIB [2]. Tab. 3 reports the evaluation results on these three datasets. Fig. 5 shows the convergence analysis on BraTS and OAIZIB. Overall, our models generalize well to medical images. We also find that the models trained on larger datasets (i.e. $\\mathrm{C + L}$ ) generalize better than the models trained on smaller datasets (i.e. SBD).", + "bbox": [ + 76, + 518, + 468, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Towards Practical Annotation Tool", + "text_level": 1, + "bbox": [ + 76, + 648, + 382, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tiny Backbone To allow for practical applications, especially on low-end devices with limited computational resources, we implement an extremely tiny backbone (i.e. ViT-xTiny) for SimpleClick. Compared with ViT-Base, ViT-xTiny decreases the embedding dimension from 768 to 160 and the number of attention blocks from 12 to 8. We end up with a SimpleClick-xTiny model, which is comparable with the tiny FocalClick models in terms of parameters. Comparison results in Tab. 4 show that our model outperforms FocalClick models, even though it is trained from scratch due to the lack of readily available pretrained weights.", + "bbox": [ + 76, + 672, + 468, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computational Analysis Tab. 5 shows a comparison of computational requirements with respect to model parameters, FLOPs, GPU memory consumption, and speed; the speed is measured by seconds per click (SPC). Fig. 1 shows", + "bbox": [ + 76, + 839, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c21f115adfcb14cb00059f1d17fa99b45b5ffc6cd78adc577f71a140d1d2d7c3.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Comparison results on SBD for tiny models. All models are trained on C+L with 230 epochs. Our SimpleClick-xTiny model outperforms FocalClick models without pretraining." + ], + "table_body": "
ModelBackbonePretrainedParams/MNoC85NoC90
FocalClickHRNet-18s-S1✓4.224.747.29
FocalClickSegFormer-B0-S1✓3.724.987.60
SimpleClickViT-xTiny✗3.724.717.09
", + "bbox": [ + 506, + 89, + 885, + 145 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3f3dde7d97a3c8d156d56a57cf88d42de1c186af5be9667e4861222df8b9c70b.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5. Computation comparison for model parameters, FLOPs, GPU memory consumption (measured by the maximum GPU memory managed by PyTorch's caching allocator), and speed (measured by seconds per click). Each method is denoted by its backbone. The small number in front of the model denotes the input size $(448\\times 448$ for our models by default). Even for our ViT-H model, the speed (132ms) and memory consumption (3.22G) are sufficient to meet the requirements of a practical annotation tool." + ], + "table_body": "
BackboneParams/MFLOPs/GMem/G↓SPC/ms
HR-18s 400 [43]4.2217.940.5054
HR-18 400 [43]10.0330.990.5256
HR-32 400 [43]30.9583.121.1286
Swin-B 400 [33]87.44138.211.4136
Swin-L 400 [33]195.90302.782.1444
SegF-B0 256 [10]3.723.420.1037
SegF-B3 256 [10]45.6624.750.3253
ResN-34 384 [9]23.47113.600.2534
ResN-50 384 [30]40.3678.820.85331
ResN-101 384 [30]59.35100.760.89355
Ours-ViT-xT 2243.722.630.1717
Ours-ViT-xT 4483.7210.520.2329
Ours-ViT-B 22496.4642.440.5134
Ours-ViT-B 44896.46169.780.8754
Ours-ViT-L 448322.18532.871.7286
Ours-ViT-H 448659.391401.933.22132
", + "bbox": [ + 501, + 200, + 890, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the interactive segmentation performance of methods in terms of FLOPs. In Fig. 1 and Tab. 5, each method is denoted by its backbone. For fair comparison, we evaluate all the methods on the same benchmark (i.e. GrabCut) and using the same computer (GPU: NVIDIA RTX A6000, CPU: Intel Silver×2). We only calculate the FLOPs in a single forward pass. For methods like FocusCut which require multiple forward passes for each click, the FLOPs may be much higher than reported. By default, our method takes images of size $448 \\times 448$ as the fixed input. Even for our ViT-H model, the speed (132ms) and memory consumption (3.22G) is sufficient to meet the requirements of a practical annotation tool.", + "bbox": [ + 496, + 561, + 890, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 770, + 653, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we ablate the backbone finetuning and feature pyramid design. Tab. 6 shows the ablation results. By default, we finetune the backbone along with other modules. As an ablation, we freeze the backbone during finetuning, leading to significantly worse performance. This ablation is explainable considering the ViT backbone takes most of the model parameters (Tab. 1). For the second ablation,", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22296", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/35802b1d10ba2be9ff5c3882c8a36bb273d9da7e497eb7cc5fbe7f5d9e7c158d.jpg", + "image_caption": [ + "(a) simple feature pyramid" + ], + "image_footnote": [], + "bbox": [ + 122, + 89, + 297, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/16cf33bbdfb30df78c8ac692f33c5127ac9495c2039b41c38ad9d994b593755d.jpg", + "image_caption": [ + "(b) single-scale variant" + ], + "image_footnote": [], + "bbox": [ + 305, + 90, + 478, + 156 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4522563d1266cff3666743c2bd79b0c55bd0132df7861db3d02715fa953d6ed5.jpg", + "image_caption": [ + "(c) parallel variant" + ], + "image_footnote": [], + "bbox": [ + 486, + 90, + 661, + 156 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/44b772ea34f42aabdba4d854168e21e494dc508808f4b8567fbb51b74c2ea160.jpg", + "image_caption": [ + "(d) partial variant" + ], + "image_footnote": [], + "bbox": [ + 669, + 90, + 846, + 156 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1dc18ac633ff811b441fc0705291c390879570e65fc8e1014049daeae3c9388c.jpg", + "image_caption": [ + "Figure 6. Simple feature pyramid and its variants: b) the single-scale variant ablates the multi-scale property; c) the parallel variant evenly extracts features from the backbone; d) the partial variant freezes the second half parameters of the backbone. Tab. 6 shows the comparison results.", + "(a) the worst case", + "Figure 7. Segmentation results on DAVIS [39]: (a) the worst case; (b) a challenging case; (c) five normal cases. The backbone model is ViT-L trained on COCO [29] + LVIS [19]. The segmentation probability maps are shown in blue; the segmentation maps are overlaid in red on the original images. The clicks are shown as green (positive click) or red (negative click) dots on the image. GT denotes ground truth." + ], + "image_footnote": [], + "bbox": [ + 83, + 233, + 312, + 468 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8bfdd83a2fd3634da377d7b52ef40491c63f15a354464fd503e24104a0295645.jpg", + "image_caption": [ + "(b) a challenging case" + ], + "image_footnote": [], + "bbox": [ + 313, + 233, + 542, + 468 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3ab6c7acf59ed0a008a8399380f51bd7d7d1b4e27840ece7b26b25125883198d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 234, + 658, + 468 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bcebc9d9442dfdb80f687f7228ba9aba5b368ee468147585e8350ffb35badee1.jpg", + "image_caption": [ + "(c) five normal cases" + ], + "image_footnote": [], + "bbox": [ + 658, + 234, + 772, + 468 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7b102bbd0da3fe420b8a57ea26bf5f3c3b44beaa5381e4943b6822c88a8e80b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 234, + 887, + 468 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/96a72af9dd43015beb554220c65c3fc6d3c71a080d7269a019d0124376b0925b.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6. Ablation study on backbone finetuning and feature pyramid (FP) design (Fig. 6). The metric is NoC@90. We have three findings in this ablation: 1) freezing the ViT backbone during finetuning significantly deteriorates the performance; 2) the multiscale property matters for the simple feature pyramid; 3) the last feature map from the backbone is sufficient to build an effective feature pyramid." + ], + "table_body": "
FP designfrozen ViTViT-BViT-L
SBDPascalSBDPascal
(a) simple FP✓11.486.939.755.59
(a) simple FP✗5.242.534.462.15
(b) single-scale✗6.562.805.532.48
(c) parallel✗7.213.096.262.79
(d) partial✗8.294.347.514.25
", + "bbox": [ + 83, + 549, + 464, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we compare the default simple feature pyramid design with three variants depicted in Fig. 6 (i.e. (b), (c), and (d)). First, we observe that the multi-scale representation matters for the feature pyramid. By ablating the multi-scale property in the simple feature pyramid, the performance drops considerably. We also notice that the last feature map from the backbone is strong enough to build the feature pyramid. The parallel feature pyramid generated by multi-stage feature maps", + "bbox": [ + 75, + 779, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from the backbone does not surpass the simple feature pyramid that only uses the last feature map of the backbone.", + "bbox": [ + 500, + 551, + 888, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations and Remarks", + "text_level": 1, + "bbox": [ + 500, + 609, + 738, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our best-performing model (ViT-H) is much larger than existing models, leading to concerns about an unfair comparison. Our method is not prompt-efficient as every click update requires recomputing image features. The recent advancements in interactive segmentation [25, 53] show an elegant solution to this issue. Besides, these methods use sparse vectors to represent clicks, which might be more efficient than dense masks. Other than these, our models may fail in some challenging scenarios such as objects with very thin and elongated shapes or cluttered occlusions((a) and (b) in Fig. 7). We leave the improvements for future work.", + "bbox": [ + 496, + 638, + 890, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We are entering an era of large-scale pretraining on multimodal foundation models, which is dramatically transforming the landscape of vision and language tasks. In this context, we hope SimpleClick will serve as a strong baseline for a new wave of high-performing interactive segmentation methods based on ViTs and large-scale pretraining.", + "bbox": [ + 496, + 809, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22297", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 76, + 90, + 204, + 104 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We proposed SimpleClick, a plain-backbone model for interactive image segmentation. Our method leveraged a general-purpose ViT backbone that can benefit from advancements in pretrained ViT models. With the readily-available MAE weights, SimpleClick achieved state-of-the-art performance on natural images and demonstrated strong generalizability to medical images. Our method is simple yet effective, highlighting its suitability as a strong baseline model and a practical annotation tool.", + "bbox": [ + 75, + 114, + 472, + 252 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 76, + 265, + 243, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank our reviewers and ACs for their invaluable comments and suggestions. Research reported in this publication was supported by the National Institutes of Health (NIH) under award number NIH 1R01AR072013. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.", + "bbox": [ + 75, + 290, + 468, + 381 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 393, + 174, + 410 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] David Acuna, Huan Ling, Amlan Kar, and Sanja Fidler. Efficient interactive annotation of segmentation datasets with polygon-rnn++. In CVPR, pages 859-868, 2018. 1, 2", + "[2] Felix Ambellan, Alexander Tack, Moritz Ehlke, and Stefan Zachow. Automated segmentation of knee bone and cartilage combining statistical shape knowledge and convolutional neural networks: Data from the osteoarthritis initiative. Medical image analysis, 52:109-118, 2019. 5, 6, 7", + "[3] Xue Bai and Guillermo Sapiro. A geodesic framework for fast interactive image and video segmentation and matting. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 7", + "[4] Ujjwal Baid, Satyam Ghodasara, Suyash Mohan, Michel Bilello, Evan Calabrese, Errol Colak, Keyvan Farahani, Jayashree Kalpathy-Cramer, Felipe C Kitamura, Sarthak Pati, et al. The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314, 2021. 5, 6, 7", + "[5] Gedas Bertasius and Lorenzo Torresani. Classifying, segmenting, and tracking object instances in video with mask propagation. In CVPR, pages 9739-9748, 2020. 1", + "[6] Yuri Y Boykov and M-P Jolly. Interactive graph cuts for optimal boundary & region segmentation of objects in ND images. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, volume 1, pages 105-112. IEEE, 2001. 2", + "[7] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 1", + "[8] Wuyang Chen, Xianzhi Du, Fan Yang, Lucas Beyer, Xiaohua Zhai, Tsung-Yi Lin, Huizhong Chen, Jing Li, Xiaodan Song, Zhangyang Wang, et al. A simple single-scale vision" + ], + "bbox": [ + 86, + 419, + 472, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "transformer for object localization and instance segmentation. arXiv preprint arXiv:2112.09747, 2021. 2", + "[9] Xi Chen, Zhiyan Zhao, Feiwu Yu, Yilei Zhang, and Manni Duan. Conditional diffusion for interactive segmentation. In ICCV, pages 7345-7354, 2021. 4, 5, 7", + "[10] Xi Chen, Zhiyan Zhao, Yilei Zhang, Manni Duan, Donglian Qi, and Hengshuang Zhao. FocalClick: Towards practical interactive image segmentation. In CVPR, pages 1300-1309, 2022. 1, 2, 4, 5, 7", + "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3", + "[12] Zhipeng Ding, Xu Han, Peirong Liu, and Marc Niethammer. Local temperature scaling for probability calibration. In ICCV, pages 6889-6899, 2021. 3", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3", + "[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 4, 5, 6", + "[15] Stephan Gerhard, Jan Funke, Julien Martel, Albert Cardona, and Richard Fetter. Segmented anisotropic sstem dataset of neural tissue. figshare, pages 0–0, 2013. 4, 7", + "[16] Leo Grady. Random walks for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 28(11):1768-1783, 2006. 2", + "[17] Jiaqi Gu, Hyoukjun Kwon, Dilin Wang, Wei Ye, Meng Li, Yu-Hsin Chen, Liangzhen Lai, Vikas Chandra, and David Z Pan. Multi-scale high-resolution vision transformer for semantic segmentation. In CVPR, pages 12094-12103, 2022. 2", + "[18] Varun Gulshan, Carsten Rother, Antonio Criminisi, Andrew Blake, and Andrew Zisserman. Geodesic star convexity for interactive image segmentation. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3129-3136. IEEE, 2010. 2", + "[19] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, pages 5356-5364, 2019. 4, 5, 6, 8", + "[20] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In 2011 international conference on computer vision, pages 991-998. IEEE, 2011. 1, 4, 5, 6", + "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 2, 3, 4", + "[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22298", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Won-Dong Jang and Chang-Su Kim. Interactive image segmentation via backpropagating refinement scheme. In CVPR, pages 5297-5306, 2019. 5", + "[24] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM computing surveys (CSUR), 54(10s):1-41, 2022. 2", + "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv:2304.02643, 2023. 8", + "[26] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. arXiv preprint arXiv:2203.16527, 2022. 2, 3", + "[27] Zhuwen Li, Qifeng Chen, and Vladlen Koltun. Interactive image segmentation with latent diversity. In CVPR, pages 577-585, 2018. 5", + "[28] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 2, 3", + "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 4, 5, 6, 8", + "[30] Zheng Lin, Zheng-Peng Duan, Zhao Zhang, Chun-Le Guo, and Ming-Ming Cheng. FocusCut: Diving into a focus view in interactive segmentation. In CVPR, pages 2637-2646, 2022. 1, 2, 4, 5, 7", + "[31] Zheng Lin, Zhao Zhang, Lin-Zhuo Chen, Ming-Ming Cheng, and Shao-Ping Lu. Interactive image segmentation with first click attention. In CVPR, pages 13339-13348, 2020. 2, 5", + "[32] Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Arnaud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen Awm Van Der Laak, Bram Van Ginneken, and Clara I Sánchez. A survey on deep learning in medical image analysis. Medical image analysis, 42:60-88, 2017. 1", + "[33] Qin Liu, Zhenlin Xu, Yining Jiao, and Marc Niethammer. iSegFormer: Interactive segmentation via transformers with application to 3d knee mr images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 464-474. Springer, 2022. 1, 2, 5, 7", + "[34] Qin Liu, Meng Zheng, Benjamin Planche, Srikrishna Karanam, Terrence Chen, Marc Niethammer, and Ziyan Wu. PseudoClick: Interactive image segmentation with click imitation. arXiv preprint arXiv:2207.05282, 2022. 1, 2, 3, 4, 5", + "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 1, 2", + "[36] Kevis-Kokitsi Maninis, Sergi Caelles, Jordi Pont-Tuset, and Luc Van Gool. Deep extreme cut: From extreme points to object segmentation. In CVPR, pages 616-625, 2018. 2", + "[37] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, volume 2, pages 416-423. IEEE, 2001. 4, 5", + "[38] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5", + "[39] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. 4, 5, 8", + "[40] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. \"grabcut\" interactive foreground extraction using iterated graph cuts. ACM transactions on graphics (TOG), 23(3):309-314, 2004. 2, 4, 5", + "[41] Dinggang Shen, Guorong Wu, and Heung-II Suk. Deep learning in medical image analysis. Annual review of biomedical engineering, 19:221, 2017. 1", + "[42] Konstantin Sofiuk, Ilia Petrov, Olga Barinova, and Anton Konushin. f-brs: Rethinking backpropagating refinement for interactive segmentation. In CVPR, pages 8623-8632, 2020. 5", + "[43] Konstantin Sofiuk, Ilia A Petrov, and Anton Konushin. Reviving iterative training with mask guidance for interactive segmentation. arXiv preprint arXiv:2102.06583, 2021. 1, 2, 3, 4, 5, 7", + "[44] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, pages 7262-7272, 2021. 2", + "[45] Jiajun Wu, Yibiao Zhao, Jun-Yan Zhu, Siwei Luo, and Zhuowen Tu. Milcut: A sweeping line multiple instance learning paradigm for interactive image segmentation. In CVPR, pages 256-263, 2014. 1, 2", + "[46] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. SegFormer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 2", + "[47] Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas Huang. Deep grabcut for object selection. arXiv preprint arXiv:1707.00243, 2017. 1, 2", + "[48] Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas S Huang. Deep interactive object selection. In CVPR, pages 373-381, 2016. 2, 5", + "[49] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 1", + "[50] Yuhui Yuan, Rao Fu, Lang Huang, Weihong Lin, Chao Zhang, Xilin Chen, and Jingdong Wang. Hrformer: High-resolution vision transformer for dense predict. Advances in Neural Information Processing Systems, 34:7281-7293, 2021. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22299", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Shiyin Zhang, Jun Hao Liew, Yunchao Wei, Shikui Wei, and Yao Zhao. Interactive object segmentation with inside-outside guidance. In CVPR, pages 12234-12244, 2020. 1", + "[52] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. 2", + "[53] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv:2304.06718, 2023. 8" + ], + "bbox": [ + 78, + 90, + 467, + 246 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22300", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/SimpleClick_ Interactive Image Segmentation with Simple Vision Transformers/2b87aeaa-0b31-4e94-84d5-d9bfaea7e328_model.json b/2023/SimpleClick_ Interactive Image Segmentation with Simple Vision Transformers/2b87aeaa-0b31-4e94-84d5-d9bfaea7e328_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f7b372840685317e4fd00ecb2af9bf5936e8937b --- /dev/null +++ b/2023/SimpleClick_ Interactive Image Segmentation with Simple Vision Transformers/2b87aeaa-0b31-4e94-84d5-d9bfaea7e328_model.json @@ -0,0 +1,2213 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.085, + 0.131, + 0.885, + 0.153 + ], + "angle": 0, + "content": "SimpleClick: Interactive Image Segmentation with Simple Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.181, + 0.713, + 0.217 + ], + "angle": 0, + "content": "Qin Liu, Zhenlin Xu, Gedas Bertasius, Marc Niethammer \nUniversity of North Carolina at Chapel Hill" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.22, + 0.654, + 0.234 + ], + "angle": 0, + "content": "https://github.com/uncbiag/SimpleClick" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.313, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.65 + ], + "angle": 0, + "content": "Click-based interactive image segmentation aims at extracting objects with a limited user clicking. A hierarchical backbone is the de-facto architecture for current methods. Recently, the plain, non-hierarchical Vision Transformer (ViT) has emerged as a competitive backbone for dense prediction tasks. This design allows the original ViT to be a foundation model that can be finetuned for downstream tasks without redesigning a hierarchical backbone for pretraining. Although this design is simple and has been proven effective, it has not yet been explored for interactive image segmentation. To fill this gap, we propose SimpleClick, the first interactive segmentation method that leverages a plain backbone. Based on the plain backbone, we introduce a symmetric patch embedding layer that encodes clicks into the backbone with minor modifications to the backbone itself. With the plain backbone pretrained as a masked autoencoder (MAE), SimpleClick achieves state-of-the-art performance. Remarkably, our method achieves 4.15 NoC@90 on SBD, improving \\(21.8\\%\\) over the previous best result. Extensive evaluation on medical images demonstrates the generalizability of our method. We provide a detailed computational analysis, highlighting the suitability of our method as a practical annotation tool." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.679, + 0.21, + 0.695 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The goal of interactive image segmentation is to obtain high-quality pixel-level annotations with limited user interaction such as clicking. Interactive image segmentation approaches have been widely applied to annotate large-scale image datasets, which drive the success of deep models in various applications, including video understanding [5, 49], self-driving [7], and medical imaging [32, 41]. Much research has been devoted to explore interactive image segmentation with different interaction types, such as bounding boxes [47], polygons [1], clicks [43], scribbles [45], and their combinations [51]. Among them, the click-based approach is most common due to its simplicity and well-established training and evaluation protocols." + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.269, + 0.696, + 0.285 + ], + "angle": 0, + "content": "NoC@90 on SBD dataset" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.289, + 0.889, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.516, + 0.895, + 0.6 + ], + "angle": 0, + "content": "Figure 1. Interactive segmentation results on SBD [20]. The metric \"NoC@90\" denotes the number of clicks required to obtain \\(90\\%\\) IoU. The area of each bubble is proportional to the FLOPs of a model variant (Tab. 5). We show that plain ViTs outperform all hierarchical backbones for interactive image segmentation at a moderate computational cost." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.895, + 0.809 + ], + "angle": 0, + "content": "Recent advances in click-based approaches mainly lie in two orthogonal directions: 1) the development of more effective backbone networks and 2) the exploration of more elaborate refinement modules built upon the backbone. For the former direction, different hierarchical backbones, including both ConvNets [30,43] and ViTs [10,33], have been developed for interactive segmentation. For the latter direction, various refinement modules, including local refinement [10,30] and click imitation [34], have been proposed to further boost segmentation performance. In this work, we delve into the former direction and focus on exploring a plain backbone for interactive segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.903 + ], + "angle": 0, + "content": "A hierarchical backbone is the predominant architecture for current interactive segmentation methods. This design is deeply rooted in ConvNets, represented by ResNet [22], and has been adopted by ViTs, represented by the Swin Transformer [35]. The motivation for a hierarchical backbone stems from the locality of convolution operations, leading" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22290" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.529 + ], + "angle": 0, + "content": "to insufficient model receptive field size without the hierarchy. To increase the receptive field size, ConvNets have to progressively downsample feature maps to capture more global contextual information. Therefore, they often require a feature pyramid network such as FPN [28] to aggregate multi-scale representations for high-quality segmentation. However, this reasoning no longer applies for a plain ViT, in which global information can be captured from the first self-attention block. Because all feature maps in the ViT are of the same resolution, the motivation for an FPN-like feature pyramid also no longer remains. The above reasoning is supported by a recent finding that a plain ViT can serve as a strong backbone for object detection [26]. This finding indicates a general-purpose ViT backbone might be suitable for other tasks, which then can decouple pretraining from finetuning and transfer the benefits from readily available pretrained ViT models (e.g. MAE [21]) to these tasks. However, although this design is simple and has been proven effective, it has not yet been explored in interactive segmentation. In this work, we propose SimpleClick, the first plain-backbone method for interactive segmentation. The core of SimpleClick is a plain ViT backbone that maintains single-scale representations throughout. We only use the last feature map from the plain backbone to build a simple feature pyramid for segmentation, largely decoupling the general-purpose backbone from the segmentation-specific modules. To make SimpleClick more efficient, we use a light-weight MLP decoder to transform the simple feature pyramid into a segmentation (see Sec. 3 for details)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.53, + 0.473, + 0.68 + ], + "angle": 0, + "content": "We extensively evaluate our method on 10 public benchmarks, including both natural and medical images. With the plain backbone pretrained as a MAE [21], our method achieves 4.15 NoC@90 on SBD, which outperforms the previous best method by \\(21.8\\%\\) without a complex FPN-like design and local refinement. We demonstrate the generalizability of our method by out-of-domain evaluation on medical images. We further analyze the computational efficiency of SimpleClick, highlighting its suitability as a practical annotation tool." + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.681, + 0.283, + 0.695 + ], + "angle": 0, + "content": "Our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.697, + 0.469, + 0.725 + ], + "angle": 0, + "content": "- We propose SimpleClick, the first plain-backbone method for interactive image segmentation." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.727, + 0.469, + 0.77 + ], + "angle": 0, + "content": "- SimpleClick achieves state-of-the-art performance on natural images and shows strong generalizability on medical images." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.772, + 0.469, + 0.816 + ], + "angle": 0, + "content": "- SimpleClick meets the computational efficiency requirement for a practical annotation tool, highlighting its readiness for real-world applications." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.697, + 0.469, + 0.816 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.83, + 0.22, + 0.846 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Interactive Image Segmentation Interactive image segmentation is a longstanding problem for which increasingly better solution approaches have been proposed. Early" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "works [6, 16, 18, 40] tackle this problem using graphs defined over image pixels. However, these methods only focus on low-level image features and therefore tend to have difficulty with complex objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.156, + 0.893, + 0.581 + ], + "angle": 0, + "content": "Thriving on large datasets, ConvNets [10, 30, 43, 47, 48] have evolved as the dominant architecture for high-quality interactive segmentation. ConvNet-based methods have explored various interaction types, such as bounding boxes [47], polygons [1], clicks [43], and scribbles [45]. Click-based approaches are the most common due to their simplicity and well-established training and evaluation protocols. Xu et al. [48] first proposed a click simulation strategy that has been adopted by follow-up work [10, 34, 43]. DEXTR [36] extracts a target object by specifying its four extreme points (left-most, right-most, top, and bottom pixels). FCA-Net [31] demonstrates the critical role of the first click for better segmentation. Recently, ViTs have been applied to interactive segmentation. FocalClick [10] uses SegFormer [46] as the backbone network and achieves state-of-the-art segmentation results with high computational efficiency. iSegFormer [33] uses a Swin Transformer [35] as the backbone network for interactive segmentation on medical images. Besides the contribution of backbones, some works are exploring elaborate refinement modules built upon the backbone. FocalClick [10] and FocusCut [30] propose similar local refinement modules for high-quality segmentation. PseudoClick [34] proposes a click-imitation mechanism by estimating the next click to further reduce human annotation cost. Our method differs from all previous click-based methods in its plain, non-hierarchical ViT backbone, enjoying the benefits from readily available pretrained ViT models (e.g. MAE [21])." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Vision Transformers for Non-Interactive Segmentation Recently, ViT-based approaches [17, 24, 44, 46, 50] have shown competitive performance on segmentation tasks compared to ConvNets. The original ViT [13] is a non-hierarchical architecture that only maintains single-scale feature maps throughout. SETR [52] and Segmenter [44] use the original ViT as the encoder for semantic segmentation. To allow for more efficient segmentation, the Swin Transformer [35] reintroduces a computational hierarchy into the original ViT architecture using shifted window attention, leading to a highly efficient hierarchical ViT backbone. SegFormer [46] designs hierarchical feature representations based on the original ViT using overlapped patch merging, combined with a light-weight MLP decoder for efficient segmentation. HRViT [17] integrates a high-resolution multi-branch architecture with ViTs to learn multi-scale representations. Recently, the original ViT has been reintroduced as a competitive backbone for semantic segmentation [8] and object detection [26], with the aid of MAE [21] pretraining and window attention. Inspired by this finding, we explore using a plain ViT as the backbone" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22291" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.327, + 0.108 + ], + "angle": 0, + "content": "network for interactive segmentation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.126, + 0.17, + 0.143 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.154, + 0.473, + 0.276 + ], + "angle": 0, + "content": "Our goal is not to propose new modules but to adapt a plain-ViT backbone for interactive segmentation with minimal modifications. Fig. 2 shows the overview of our method. Sec. 3.1 briefly describes the plain segmentation backbone. Sec. 3.2 shows how to adapt the backbone for interactive segmentation. Sec. 3.3 introduces other modules of SimpleClick. Sec. 3.4 describes the training and inference details of our method." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.29, + 0.345, + 0.307 + ], + "angle": 0, + "content": "3.1. Plain Segmentation Backbone" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.315, + 0.473, + 0.527 + ], + "angle": 0, + "content": "We use a plain ViT [13] as the segmentation backbone. Unlike previous hierarchical backbones, a plain ViT only maintains single-scale feature maps throughout. Given an image, the patch embedding layer of the plain ViT first divides the image into non-overlapping fixed-size patches (e.g. \\(16 \\times 16\\)). All patches are then flattened and linearly projected to a sequence of fixed-length tokens before feeding into self-attention blocks. These output tokens will be reshaped and upsampled to match the spatial size of the input image for segmentation. In this work, we consider three standard ViT backbones: ViT-B, ViT-L, and ViT-H (Tab. 1 shows the number of parameters of these backbones). For stable and efficient training, we use readily-available MAEs [21] as the pretrained weights for these backbones." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.542, + 0.456, + 0.559 + ], + "angle": 0, + "content": "3.2. Clicks Encoding for Segmentation Backbone" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.749 + ], + "angle": 0, + "content": "We adapt the above plain segmentation backbone for interactive segmentation by turning user interactions into a form of guidance learned by the network. Similar to previous click-based methods [34, 43], we use clicks as a form of user interaction and encode each click as a small disk on a two-channel masks [34, 43] (one channel for positive clicks while the other for negative ones; positive clicks should be put in the foreground, while the negative ones should be put in the background). We automatically simulate human clicks for efficient training and evaluation. The clicks simulation process is described in Sec.3.4. Note that for human evaluation, a human-in-the-loop will provide all the clicks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Given human clicks or simulated clicks, we now introduce how to fuse them into the plain backbone. We first introduce a patch-embedding layer that is symmetric to the one in the backbone. Note that we also concatenate the previous segmentation to the clicks map as an additional channel for better performance. The two symmetric embedding layers operate on the image and the clicks map, respectively. The inputs are patchified, flattened, and projected to two vector sequences of the same dimension, followed by element-wise addition before feeding into the self-attention blocks." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.09, + 0.887, + 0.147 + ], + "angle": 0, + "content": "
Model↓ Module→ViT BackboneConv. NeckMLP Head
Ours-ViT-B (base)83.0 (89.3%)9.0 (9.7%)0.9 (1.0%)
Ours-ViT-L (large)290.8 (94.3%)16.5 (5.3%)1.1 (0.4%)
Ours-ViT-H (huge)604.0 (95.7%)25.8 (4.1%)1.3 (0.2%)
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.147, + 0.892, + 0.188 + ], + "angle": 0, + "content": "Table 1. Number of parameters of our models. The unit is a million. The percentage of parameters is shown in bracket. Most parameters are used by the ViT backbone." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.214, + 0.655, + 0.229 + ], + "angle": 0, + "content": "3.3. Other Modules" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.238, + 0.895, + 0.449 + ], + "angle": 0, + "content": "Simple Feature Pyramid For the hierarchical backbone, a feature pyramid is commonly produced by an FPN [28] to combine features from different stages. For the plain backbone, a feature pyramid can be generated in a much simpler way: by a set of parallel convolutional or deconvolutional layers using only the last feature map of the backbone. As shown in Fig. 2, given the input ViT feature map, a multiscale feature map can be produced by four convolutions with different strides. Though the effectiveness of this simple feature pyramid design is first demonstrated in ViTDet [26] for object detection, we show in this work the effectiveness of this simple feature pyramid design for interactive segmentation. We also propose several additional variants (Fig. 6) as part of an ablation study (Sec. 4.4)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.895, + 0.753 + ], + "angle": 0, + "content": "All-MLP Segmentation Head We implement a lightweight segmentation head using only MLP layers. It takes in the simple feature pyramid and produces a segmentation probability map of scale 1/4, followed by an upsampling operation to recover the original resolution. Note that this segmentation head avoids computationally demanding components and only accounts for up to \\(1\\%\\) of the model parameters (Tab. 1). The key insight is that with a powerful pretrained backbone, a lightweight segmentation head is sufficient for interactive segmentation. The proposed all-MLP segmentation head works in three steps. First, each feature map from the simple feature pyramid goes through an MLP layer to transform it to an identical channel dimension (i.e. \\(C_2\\) in Fig. 2). Second, all feature maps are upsampled to the same resolution (i.e. 1/4 in Fig. 2) for concatenation. Third, the concatenated features are fused by another MLP layer to produce a single-channel feature map, followed by a sigmoid function to obtain a segmentation probability map, which is then transformed to a binary segmentation given a predefined threshold (i.e. 0.5)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.764, + 0.78, + 0.781 + ], + "angle": 0, + "content": "3.4. Training and Inference Settings" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.788, + 0.894, + 0.865 + ], + "angle": 0, + "content": "Backbone Pretraining Our backbone models are pretrained as MAEs [21] on ImageNet-1K [11]. In MAE pretraining, the ViT models reconstruct the randomly masked pixels of images while learning a universal representation. This simple self-supervised approach turns out to be an efficient and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This probability map may be miscalibrated and can be improved by calibration approaches [12]." + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22292" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.189, + 0.096, + 0.782, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.345, + 0.895, + 0.416 + ], + "angle": 0, + "content": "Figure 2. SimpleClick overview. Our method consists of three main modules: (a) a plain ViT backbone that maintains single-scale feature maps throughout; (b) a multi-scale simple feature pyramid that is generated from the last feature maps of the backbone by four parallel convolution or deconvolution layers; (c) a light-weight MLP decoder for segmentation. We also input the previous segmentation to improve the performance and to allow clicking from existing masks. User clicks are encoded as a two-channel disk map, combined with the previous segmentation as input. Boxes in blue are intermediate feature maps. The positional encoding is not shown for brevity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.471, + 0.485 + ], + "angle": 0, + "content": "scalable way to pretrain ViT models [21]. In this work, we do not perform pretraining ourselves. Instead, we simply use the readily available pretrained MAE weights from [21]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.471, + 0.866 + ], + "angle": 0, + "content": "Clicks Simulation and End-to-end Finetuning With the pretrained backbone, we finetune our model end-to-end on the interactive segmentation task. The finetuning pipeline can be briefly described as follows. First, we automatically simulate clicks based on the current segmentation and gold standard segmentation, without a human-in-the-loop providing the clicks. Specifically, we use a combination of random and iterative click simulation strategies, inspired by RITM [43]. The random click simulation strategy generates clicks in parallel, without considering the order of the clicks. The iterative click simulation strategy generates clicks iteratively, where the next click should be placed on the erroneous region of a prediction that was obtained using the previous clicks. This strategy is more similar to human clicking behavior. Second, we incorporate the segmentation from the previous interaction as an additional input for the backbone, further improving the segmentation quality. This also allows our method to refine from an existing segmentation, which is a desired feature for a practical annotation tool. We use the normalized focal loss [43] (NFL) to train all our models. Previous works [10, 43] show that NFL converges faster and achieves better performance than the widely used binary cross entropy loss for interactive segmentation tasks. Similar training pipelines have been proposed by RITM [43] and its follow-up works [9, 10, 34]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.9 + ], + "angle": 0, + "content": "Human Evaluation and Automatic Evaluation There are two inference modes: automatic evaluation and human eval" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.895, + 0.545 + ], + "angle": 0, + "content": "uation. For automatic evaluation, clicks are automatically simulated based on the current segmentation and gold standard. For human evaluation, a human-in-the-loop provides all clicks based on their subjective evaluation of current segmentation results. We use automatic evaluation for quantitative analyses and human evaluation for a qualitative assessment of the interactive segmentation behavior." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.559, + 0.634, + 0.575 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.628 + ], + "angle": 0, + "content": "Datasets We conducted experiments on 10 public datasets including 7 natural image datasets and 3 medical datasets. The details are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.629, + 0.891, + 0.659 + ], + "angle": 0, + "content": "- GrabCut [40]: 50 images (50 instances), each with clear foreground and background differences." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.66, + 0.891, + 0.689 + ], + "angle": 0, + "content": "- Berkeley [37]: 96 images (100 instances); this dataset shares a small portion of images with GrabCut." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.69, + 0.89, + 0.72 + ], + "angle": 0, + "content": "- DAVIS [39]: 50 videos; we only use the same 345 frames as used in [10, 30, 34, 43] for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.72, + 0.89, + 0.75 + ], + "angle": 0, + "content": "- Pascal VOC [14]: 1449 images (3427 instances) in the validation set. We only test on the validation set." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.75, + 0.891, + 0.809 + ], + "angle": 0, + "content": "- SBD [20]: 8498 training images (20172 instances) and 2857 validation images (6671 instances). Following previous works [10, 30, 43], we train our model on the training set and evaluate on the validation set." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.811, + 0.891, + 0.87 + ], + "angle": 0, + "content": "- COCO [29]+LVIS [19] (C+L): COCO contains 118K training images (1.2M instances); LVIS shares the same images with COCO but has much higher segmentation quality. We combine the two datasets for training." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "- ssTEM [15]: two image stacks, each contains 20 medical images. We use the same stack that was used in [34]." + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.629, + 0.891, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22293" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.101, + 0.466, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.1, + 0.858, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.301, + 0.895, + 0.342 + ], + "angle": 0, + "content": "Figure 3. Convergence analysis for models trained on either SBD [20] or COCO [29] + LVIS [19] (C+L). We report results on SBD [20] and Pascal VOC [14]. The metric is mean IoU given \\( k \\) clicks (\\( \\text{mIoU} @ k \\)). Our models in general require fewer clicks for a given accuracy level." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.353, + 0.896, + 0.68 + ], + "angle": 0, + "content": "
MethodBackboneGrabCutBerkeleySBDDAVISPascal VOC
NoC85NoC90NoC85NoC90NoC85NoC90NoC85NoC90NoC85NoC90
\\( \\nmid \\)DIOS [48] \\( _{CVPR16} \\)FCN-6.04-8.65---12.586.88-
\\( \\nmid \\)FCA-Net [31] \\( _{CVPR20} \\)ResNet-101-2.08-3.92---7.572.69-
\\( \\nmid \\)LD [27] \\( _{CVPR18} \\)VGG-193.204.79--7.4110.785.059.57--
\\( \\nmid \\)BRS [23] \\( _{CVPR19} \\)DenseNet2.603.60-5.086.599.785.588.24--
\\( \\nmid \\)f-BRS [42] \\( _{CVPR20} \\)ResNet-1012.302.72-4.574.817.735.047.41--
\\( \\nmid \\)RITM [43] \\( _{Preprint21} \\)HRNet-181.762.041.873.223.395.434.946.712.513.03
\\( \\nmid \\)CDNet [9] \\( _{ICCV21} \\)ResNet-341.862.181.953.275.187.895.006.893.614.51
\\( \\nmid \\)PseudoClick [34] \\( _{ECCV22} \\)HRNet-181.682.041.853.233.385.404.816.572.342.74
\\( \\nmid \\)FocalClick [10] \\( _{CVPR22} \\)HRNet-18s1.862.06-3.144.306.524.926.48--
\\( \\nmid \\)FocalClick [10] \\( _{CVPR22} \\)SegF-B01.661.90-3.144.346.515.027.06--
\\( \\nmid \\)FocusCut [30] \\( _{CVPR22} \\)ResNet-501.601.78\\( 1.85^{\\dagger } \\)3.443.625.665.006.38--
\\( \\nmid \\)FocusCut [30] \\( _{CVPR22} \\)ResNet-1011.461.64\\( 1.81^{\\dagger } \\)3.013.405.314.856.22--
\\( \\nmid \\)OursViT-B1.401.541.442.463.285.244.105.482.382.81
\\( \\nmid \\)OursViT-L1.381.461.402.332.694.464.125.391.952.30
\\( \\nmid \\)OursViT-H1.321.441.362.092.514.154.205.341.882.20
\\( \\nmid \\)RITM [43] \\( _{Preprint21} \\)HRNet-321.461.561.432.103.595.714.115.342.192.57
\\( \\nmid \\)CDNet [9] \\( _{ICCV21} \\)ResNet-341.401.521.472.064.307.044.275.562.743.30
\\( \\nmid \\)PseudoClick [34] \\( _{ECCV22} \\)HRNet-321.361.501.402.083.465.543.795.111.942.25
\\( \\nmid \\)FocalClick [10] \\( _{CVPR22} \\)SegF-B01.401.661.592.274.566.864.045.492.973.52
\\( \\nmid \\)FocalClick [10] \\( _{CVPR22} \\)SegF-B31.441.501.551.923.535.593.614.902.462.88
\\( \\nmid \\)OursViT-B1.381.481.361.973.435.623.665.062.062.38
\\( \\nmid \\)OursViT-L1.321.401.341.892.954.893.264.811.721.96
\\( \\nmid \\)OursViT-H1.381.501.361.752.854.703.414.781.761.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.681, + 0.895, + 0.737 + ], + "angle": 0, + "content": "Table 2. Comparison with previous results. We report results on five benchmarks: GrabCut [40], Berkeley [37], SBD [20], DAVIS [39], and Pascal VOC [14]. The best results are set in bold. \\(\\nsubseteq\\) denotes a model trained on Pascal; \\(\\downarrow\\) denotes a model trained on SBD; \\(\\sqcup\\) denotes a model trained on COCO [29] + LVIS [19] (C+L); \\(\\dagger\\) denotes a number reproduced by the released or retrained models. Our models achieve state-of-the-art performance on all benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.761, + 0.468, + 0.79 + ], + "angle": 0, + "content": "- BraTS [4]: 369 magnetic resonance image (MRI) volumes; we test on the same 369 slices used in [34]." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.792, + 0.468, + 0.821 + ], + "angle": 0, + "content": "- OAIZIB [2]: 507 MRI volumes; we test on the same 150 slices (300 instances) as used in [33]." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.761, + 0.468, + 0.821 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics Following previous works [30,42,43], we automatically simulate user clicks by comparing the current segmentation with the gold standard. In this simulation, the next click will be put at the center of the region with the largest error. We use the Number of Clicks (NoC)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.761, + 0.892, + 0.868 + ], + "angle": 0, + "content": "as the evaluation metric to calculate the number of clicks required to achieve a target Intersection over Union (IoU). We set two target IoUs: \\(85\\%\\) and \\(90\\%\\), represented by \\(\\mathrm{NoC}\\% 85\\) and \\(\\mathrm{NoC}\\% 90\\) respectively. The maximum number of clicks for each instance is set to 20. We also use the average IoU given \\(k\\) clicks (\\(\\mathrm{mIoU} @ k\\)) as an evaluation metric to measure the segmentation quality given a fixed number of clicks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details We implement our models using Python and PyTorch [38]. We implement three models" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22294" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.096, + 0.338, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.096, + 0.62, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.097, + 0.898, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.077, + 0.24, + 0.339, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.24, + 0.62, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.24, + 0.899, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.382, + 0.892, + 0.41 + ], + "angle": 0, + "content": "Figure 4. Histogram analysis of IoU given a predefined number of clicks \\( k \\) (IoU@k). We report analysis on SBD [20] and Pascal VOC [14] with models trained on SBD. Compared with the two baselines, our models achieve higher-quality segmentation with fewer failure cases." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.434, + 0.459, + 0.63 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.434, + 0.856, + 0.63 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.634, + 0.894, + 0.678 + ], + "angle": 0, + "content": "Figure 5. Convergence analysis for two medical image datasets: BraTS [4] and OAIZIB [2]. Models are trained on either SBD [20] or COCO [29] + LVIS [19] (denoted as \\(\\mathrm{C} + \\mathrm{L}\\)). The metric is mean IoU given \\(k\\) clicks. Overall, our models require fewer clicks for a given accuracy level. The performance gain is more prominent for bigger models (e.g. ViT-H) or larger training sets (e.g. C+L)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.473, + 0.899 + ], + "angle": 0, + "content": "based on three vanilla ViT models (i.e. ViT-B, ViT-L, and ViT-H). These backbone models are initialized with the MAE pretrained weights, and then are finetuned end-to-end with other modules. We train our models on either SBD or COCO+LVIS with 55 epochs; the initial learning rate is set to \\(5 \\times 10^{-5}\\) and decreases to \\(5 \\times 10^{-6}\\) after epoch 50. We set the batch size to 140 for ViT-Base, 72 for ViT-Large, and 32 for ViT-Huge to fit the models into GPU memory. All our models are trained on four NVIDIA RTX A6000 GPUs. We use the following data augmentation techniques: random resizing (scale range from 0.75 to 1.25), random flipping and rotation, random brightness contrast, and random cropping. Though the ViT backbone was pretrained on images of size" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.892, + 0.748 + ], + "angle": 0, + "content": "\\(224 \\times 224\\), we finetune on \\(448 \\times 448\\) with non-shifting window attention for better performance. We optimize using Adam with \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.757, + 0.803, + 0.773 + ], + "angle": 0, + "content": "4.1. Comparison with Previous Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We show in Tab. 2 the comparisons with previous state-of-the-art results. Our models achieve the best performance on all the five benchmarks. Remarkably, when trained on SBD training set, our ViT-H model achieves 4.15 NoC@90 on the SBD validation set, outperforming the previous best score by \\(21.8\\%\\). Since the SBD validation set contains the largest number of instances (6671 instances) among the five benchmarks, this improvement is convincing. When trained" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22295" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.472, + 0.236 + ], + "angle": 0, + "content": "
ModelssTEM\nmIoU@10BraTS\nmIoU@10 / 20OAIZIB\nmIoU@10 / 20
↓RITM-H18 [43]93.1587.05 / 90.4771.04 / 78.52
↓CDN-RN34 [9]66.7258.34 / 82.0738.07 / 61.17
↓RITM-H32 [43]94.1188.34 / 89.2575.27 / 75.18
↓CDN-RN34 [9]88.4680.24 / 86.6363.19 / 74.21
↓FC-SF-B0 [10]92.6286.02 / 90.7474.08 / 79.14
↓FC-SF-B3 [10]93.6188.62 / 90.5875.77 / 80.08
↓Ours-ViT-B93.7286.98 / 90.6776.05 / 79.61
↓Ours-ViT-L94.3488.43 / 90.8477.34 / 79.97
↓Ours-ViT-H94.0888.98 / 91.0077.50 / 80.10
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.078, + 0.237, + 0.47, + 0.278 + ], + "angle": 0, + "content": "Table 3. Out-of-domain evaluation on three medical image datasets: ssTEM [15], BraTS [4], and OAIZIB [2]. Our models generalize very well on the three datasets, without finetuning." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.303, + 0.47, + 0.485 + ], + "angle": 0, + "content": "on COCO+LVIS, our models also achieve state-of-the-art performance on all benchmarks. Fig. 7 shows several segmentation cases on DAVIS, including the worst case. Note that the DAVIS dataset requires high-quality segmentations because all its instances have a high-quality gold standard. Our models still achieve the state-of-the-art on DAVIS without using specific modules, such as a local refinement module [10], which is beneficial for high-quality segmentation. Fig. 3 shows that our method converges better than other methods with sufficient clicks, leading to fewer failure cases as shown in Fig. 4. We only report results on SBD and Pascal VOC, the top two largest datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.495, + 0.469, + 0.513 + ], + "angle": 0, + "content": "4.2. Out-of-Domain Evaluation on Medical Images" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.519, + 0.469, + 0.64 + ], + "angle": 0, + "content": "We further evaluate the generalizability of our models on three medical image datasets: ssTEM [15], BraTS [3], and OAIZIB [2]. Tab. 3 reports the evaluation results on these three datasets. Fig. 5 shows the convergence analysis on BraTS and OAIZIB. Overall, our models generalize well to medical images. We also find that the models trained on larger datasets (i.e. \\(\\mathrm{C + L}\\)) generalize better than the models trained on smaller datasets (i.e. SBD)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.65, + 0.383, + 0.667 + ], + "angle": 0, + "content": "4.3. Towards Practical Annotation Tool" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.674, + 0.469, + 0.84 + ], + "angle": 0, + "content": "Tiny Backbone To allow for practical applications, especially on low-end devices with limited computational resources, we implement an extremely tiny backbone (i.e. ViT-xTiny) for SimpleClick. Compared with ViT-Base, ViT-xTiny decreases the embedding dimension from 768 to 160 and the number of attention blocks from 12 to 8. We end up with a SimpleClick-xTiny model, which is comparable with the tiny FocalClick models in terms of parameters. Comparison results in Tab. 4 show that our model outperforms FocalClick models, even though it is trained from scratch due to the lack of readily available pretrained weights." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Computational Analysis Tab. 5 shows a comparison of computational requirements with respect to model parameters, FLOPs, GPU memory consumption, and speed; the speed is measured by seconds per click (SPC). Fig. 1 shows" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.09, + 0.887, + 0.146 + ], + "angle": 0, + "content": "
ModelBackbonePretrainedParams/MNoC85NoC90
FocalClickHRNet-18s-S1✓4.224.747.29
FocalClickSegFormer-B0-S1✓3.724.987.60
SimpleClickViT-xTiny✗3.724.717.09
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.147, + 0.892, + 0.188 + ], + "angle": 0, + "content": "Table 4. Comparison results on SBD for tiny models. All models are trained on C+L with 230 epochs. Our SimpleClick-xTiny model outperforms FocalClick models without pretraining." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.202, + 0.892, + 0.426 + ], + "angle": 0, + "content": "
BackboneParams/MFLOPs/GMem/G↓SPC/ms
HR-18s 400 [43]4.2217.940.5054
HR-18 400 [43]10.0330.990.5256
HR-32 400 [43]30.9583.121.1286
Swin-B 400 [33]87.44138.211.4136
Swin-L 400 [33]195.90302.782.1444
SegF-B0 256 [10]3.723.420.1037
SegF-B3 256 [10]45.6624.750.3253
ResN-34 384 [9]23.47113.600.2534
ResN-50 384 [30]40.3678.820.85331
ResN-101 384 [30]59.35100.760.89355
Ours-ViT-xT 2243.722.630.1717
Ours-ViT-xT 4483.7210.520.2329
Ours-ViT-B 22496.4642.440.5134
Ours-ViT-B 44896.46169.780.8754
Ours-ViT-L 448322.18532.871.7286
Ours-ViT-H 448659.391401.933.22132
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.427, + 0.892, + 0.537 + ], + "angle": 0, + "content": "Table 5. Computation comparison for model parameters, FLOPs, GPU memory consumption (measured by the maximum GPU memory managed by PyTorch's caching allocator), and speed (measured by seconds per click). Each method is denoted by its backbone. The small number in front of the model denotes the input size \\((448\\times 448\\) for our models by default). Even for our ViT-H model, the speed (132ms) and memory consumption (3.22G) are sufficient to meet the requirements of a practical annotation tool." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.759 + ], + "angle": 0, + "content": "the interactive segmentation performance of methods in terms of FLOPs. In Fig. 1 and Tab. 5, each method is denoted by its backbone. For fair comparison, we evaluate all the methods on the same benchmark (i.e. GrabCut) and using the same computer (GPU: NVIDIA RTX A6000, CPU: Intel Silver×2). We only calculate the FLOPs in a single forward pass. For methods like FocusCut which require multiple forward passes for each click, the FLOPs may be much higher than reported. By default, our method takes images of size \\(448 \\times 448\\) as the fixed input. Even for our ViT-H model, the speed (132ms) and memory consumption (3.22G) is sufficient to meet the requirements of a practical annotation tool." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.655, + 0.788 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we ablate the backbone finetuning and feature pyramid design. Tab. 6 shows the ablation results. By default, we finetune the backbone along with other modules. As an ablation, we freeze the backbone during finetuning, leading to significantly worse performance. This ablation is explainable considering the ViT backbone takes most of the model parameters (Tab. 1). For the second ablation," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22296" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.09, + 0.298, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.162, + 0.27, + 0.173 + ], + "angle": 0, + "content": "(a) simple feature pyramid" + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.091, + 0.48, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.162, + 0.447, + 0.173 + ], + "angle": 0, + "content": "(b) single-scale variant" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.091, + 0.663, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.163, + 0.616, + 0.173 + ], + "angle": 0, + "content": "(c) parallel variant" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.091, + 0.848, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.163, + 0.798, + 0.173 + ], + "angle": 0, + "content": "(d) partial variant" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.177, + 0.894, + 0.218 + ], + "angle": 0, + "content": "Figure 6. Simple feature pyramid and its variants: b) the single-scale variant ablates the multi-scale property; c) the parallel variant evenly extracts features from the backbone; d) the partial variant freezes the second half parameters of the backbone. Tab. 6 shows the comparison results." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.234, + 0.313, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.152, + 0.471, + 0.246, + 0.481 + ], + "angle": 0, + "content": "(a) the worst case" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.234, + 0.543, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.372, + 0.47, + 0.489, + 0.482 + ], + "angle": 0, + "content": "(b) a challenging case" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.235, + 0.659, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.235, + 0.774, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.778, + 0.235, + 0.888, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.47, + 0.772, + 0.481 + ], + "angle": 0, + "content": "(c) five normal cases" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.486, + 0.893, + 0.528 + ], + "angle": 0, + "content": "Figure 7. Segmentation results on DAVIS [39]: (a) the worst case; (b) a challenging case; (c) five normal cases. The backbone model is ViT-L trained on COCO [29] + LVIS [19]. The segmentation probability maps are shown in blue; the segmentation maps are overlaid in red on the original images. The clicks are shown as green (positive click) or red (negative click) dots on the image. GT denotes ground truth." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.55, + 0.465, + 0.649 + ], + "angle": 0, + "content": "
FP designfrozen ViTViT-BViT-L
SBDPascalSBDPascal
(a) simple FP✓11.486.939.755.59
(a) simple FP✗5.242.534.462.15
(b) single-scale✗6.562.805.532.48
(c) parallel✗7.213.096.262.79
(d) partial✗8.294.347.514.25
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.65, + 0.47, + 0.745 + ], + "angle": 0, + "content": "Table 6. Ablation study on backbone finetuning and feature pyramid (FP) design (Fig. 6). The metric is NoC@90. We have three findings in this ablation: 1) freezing the ViT backbone during finetuning significantly deteriorates the performance; 2) the multiscale property matters for the simple feature pyramid; 3) the last feature map from the backbone is sufficient to build an effective feature pyramid." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.902 + ], + "angle": 0, + "content": "we compare the default simple feature pyramid design with three variants depicted in Fig. 6 (i.e. (b), (c), and (d)). First, we observe that the multi-scale representation matters for the feature pyramid. By ablating the multi-scale property in the simple feature pyramid, the performance drops considerably. We also notice that the last feature map from the backbone is strong enough to build the feature pyramid. The parallel feature pyramid generated by multi-stage feature maps" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.553, + 0.89, + 0.583 + ], + "angle": 0, + "content": "from the backbone does not surpass the simple feature pyramid that only uses the last feature map of the backbone." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.61, + 0.739, + 0.626 + ], + "angle": 0, + "content": "5. Limitations and Remarks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.892, + 0.806 + ], + "angle": 0, + "content": "Our best-performing model (ViT-H) is much larger than existing models, leading to concerns about an unfair comparison. Our method is not prompt-efficient as every click update requires recomputing image features. The recent advancements in interactive segmentation [25, 53] show an elegant solution to this issue. Besides, these methods use sparse vectors to represent clicks, which might be more efficient than dense masks. Other than these, our models may fail in some challenging scenarios such as objects with very thin and elongated shapes or cluttered occlusions((a) and (b) in Fig. 7). We leave the improvements for future work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We are entering an era of large-scale pretraining on multimodal foundation models, which is dramatically transforming the landscape of vision and language tasks. In this context, we hope SimpleClick will serve as a strong baseline for a new wave of high-performing interactive segmentation methods based on ViTs and large-scale pretraining." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22297" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.205, + 0.106 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.473, + 0.253 + ], + "angle": 0, + "content": "We proposed SimpleClick, a plain-backbone model for interactive image segmentation. Our method leveraged a general-purpose ViT backbone that can benefit from advancements in pretrained ViT models. With the readily-available MAE weights, SimpleClick achieved state-of-the-art performance on natural images and demonstrated strong generalizability to medical images. Our method is simple yet effective, highlighting its suitability as a strong baseline model and a practical annotation tool." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.266, + 0.245, + 0.284 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.47, + 0.382 + ], + "angle": 0, + "content": "We thank our reviewers and ACs for their invaluable comments and suggestions. Research reported in this publication was supported by the National Institutes of Health (NIH) under award number NIH 1R01AR072013. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.395, + 0.175, + 0.411 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.42, + 0.47, + 0.463 + ], + "angle": 0, + "content": "[1] David Acuna, Huan Ling, Amlan Kar, and Sanja Fidler. Efficient interactive annotation of segmentation datasets with polygon-rnn++. In CVPR, pages 859-868, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.464, + 0.472, + 0.533 + ], + "angle": 0, + "content": "[2] Felix Ambellan, Alexander Tack, Moritz Ehlke, and Stefan Zachow. Automated segmentation of knee bone and cartilage combining statistical shape knowledge and convolutional neural networks: Data from the osteoarthritis initiative. Medical image analysis, 52:109-118, 2019. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.534, + 0.473, + 0.589 + ], + "angle": 0, + "content": "[3] Xue Bai and Guillermo Sapiro. A geodesic framework for fast interactive image and video segmentation and matting. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.591, + 0.47, + 0.674 + ], + "angle": 0, + "content": "[4] Ujjwal Baid, Satyam Ghodasara, Suyash Mohan, Michel Bilello, Evan Calabrese, Errol Colak, Keyvan Farahani, Jayashree Kalpathy-Cramer, Felipe C Kitamura, Sarthak Pati, et al. The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314, 2021. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.675, + 0.469, + 0.717 + ], + "angle": 0, + "content": "[5] Gedas Bertasius and Lorenzo Torresani. Classifying, segmenting, and tracking object instances in video with mask propagation. In CVPR, pages 9739-9748, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.718, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[6] Yuri Y Boykov and M-P Jolly. Interactive graph cuts for optimal boundary & region segmentation of objects in ND images. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, volume 1, pages 105-112. IEEE, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.788, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[7] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.858, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[8] Wuyang Chen, Xianzhi Du, Fan Yang, Lucas Beyer, Xiaohua Zhai, Tsung-Yi Lin, Huizhong Chen, Jing Li, Xiaodan Song, Zhangyang Wang, et al. A simple single-scale vision" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.42, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "transformer for object localization and instance segmentation. arXiv preprint arXiv:2112.09747, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.123, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[9] Xi Chen, Zhiyan Zhao, Feiwu Yu, Yilei Zhang, and Manni Duan. Conditional diffusion for interactive segmentation. In ICCV, pages 7345-7354, 2021. 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.166, + 0.892, + 0.222 + ], + "angle": 0, + "content": "[10] Xi Chen, Zhiyan Zhao, Yilei Zhang, Manni Duan, Donglian Qi, and Hengshuang Zhao. FocalClick: Towards practical interactive image segmentation. In CVPR, pages 1300-1309, 2022. 1, 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.224, + 0.892, + 0.28 + ], + "angle": 0, + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.282, + 0.892, + 0.324 + ], + "angle": 0, + "content": "[12] Zhipeng Ding, Xu Han, Peirong Liu, and Marc Niethammer. Local temperature scaling for probability calibration. In ICCV, pages 6889-6899, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.326, + 0.892, + 0.408 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.411, + 0.892, + 0.467 + ], + "angle": 0, + "content": "[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.469, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[15] Stephan Gerhard, Jan Funke, Julien Martel, Albert Cardona, and Richard Fetter. Segmented anisotropic sstem dataset of neural tissue. figshare, pages 0–0, 2013. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.513, + 0.892, + 0.554 + ], + "angle": 0, + "content": "[16] Leo Grady. Random walks for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 28(11):1768-1783, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.556, + 0.892, + 0.625 + ], + "angle": 0, + "content": "[17] Jiaqi Gu, Hyoukjun Kwon, Dilin Wang, Wei Ye, Meng Li, Yu-Hsin Chen, Liangzhen Lai, Vikas Chandra, and David Z Pan. Multi-scale high-resolution vision transformer for semantic segmentation. In CVPR, pages 12094-12103, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.628, + 0.892, + 0.697 + ], + "angle": 0, + "content": "[18] Varun Gulshan, Carsten Rother, Antonio Criminisi, Andrew Blake, and Andrew Zisserman. Geodesic star convexity for interactive image segmentation. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3129-3136. IEEE, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.7, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[19] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, pages 5356-5364, 2019. 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.743, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[20] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In 2011 international conference on computer vision, pages 991-998. IEEE, 2011. 1, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.801, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.858, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22298" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[23] Won-Dong Jang and Chang-Su Kim. Interactive image segmentation via backpropagating refinement scheme. In CVPR, pages 5297-5306, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[24] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM computing surveys (CSUR), 54(10s):1-41, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv:2304.02643, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.469, + 0.291 + ], + "angle": 0, + "content": "[26] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. arXiv preprint arXiv:2203.16527, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.293, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[27] Zhuwen Li, Qifeng Chen, and Vladlen Koltun. Interactive image segmentation with latent diversity. In CVPR, pages 577-585, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.334, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[28] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.391, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.449, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[30] Zheng Lin, Zheng-Peng Duan, Zhao Zhang, Chun-Le Guo, and Ming-Ming Cheng. FocusCut: Diving into a focus view in interactive segmentation. In CVPR, pages 2637-2646, 2022. 1, 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.505, + 0.469, + 0.547 + ], + "angle": 0, + "content": "[31] Zheng Lin, Zhao Zhang, Lin-Zhuo Chen, Ming-Ming Cheng, and Shao-Ping Lu. Interactive image segmentation with first click attention. In CVPR, pages 13339-13348, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.548, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[32] Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Arnaud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen Awm Van Der Laak, Bram Van Ginneken, and Clara I Sánchez. A survey on deep learning in medical image analysis. Medical image analysis, 42:60-88, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.632, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[33] Qin Liu, Zhenlin Xu, Yining Jiao, and Marc Niethammer. iSegFormer: Interactive segmentation via transformers with application to 3d knee mr images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 464-474. Springer, 2022. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.703, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[34] Qin Liu, Meng Zheng, Benjamin Planche, Srikrishna Karanam, Terrence Chen, Marc Niethammer, and Ziyan Wu. PseudoClick: Interactive image segmentation with click imitation. arXiv preprint arXiv:2207.05282, 2022. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.773, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.83, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[36] Kevis-Kokitsi Maninis, Sergi Caelles, Jordi Pont-Tuset, and Luc Van Gool. Deep extreme cut: From extreme points to object segmentation. In CVPR, pages 616-625, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[37] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, volume 2, pages 416-423. IEEE, 2001. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.151, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[38] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.292 + ], + "angle": 0, + "content": "[39] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.294, + 0.892, + 0.349 + ], + "angle": 0, + "content": "[40] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. \"grabcut\" interactive foreground extraction using iterated graph cuts. ACM transactions on graphics (TOG), 23(3):309-314, 2004. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.351, + 0.892, + 0.394 + ], + "angle": 0, + "content": "[41] Dinggang Shen, Guorong Wu, and Heung-II Suk. Deep learning in medical image analysis. Annual review of biomedical engineering, 19:221, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.396, + 0.892, + 0.45 + ], + "angle": 0, + "content": "[42] Konstantin Sofiuk, Ilia Petrov, Olga Barinova, and Anton Konushin. f-brs: Rethinking backpropagating refinement for interactive segmentation. In CVPR, pages 8623-8632, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.453, + 0.892, + 0.508 + ], + "angle": 0, + "content": "[43] Konstantin Sofiuk, Ilia A Petrov, and Anton Konushin. Reviving iterative training with mask guidance for interactive segmentation. arXiv preprint arXiv:2102.06583, 2021. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[44] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, pages 7262-7272, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.556, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[45] Jiajun Wu, Yibiao Zhao, Jun-Yan Zhu, Siwei Luo, and Zhuowen Tu. Milcut: A sweeping line multiple instance learning paradigm for interactive image segmentation. In CVPR, pages 256-263, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.613, + 0.892, + 0.682 + ], + "angle": 0, + "content": "[46] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. SegFormer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.685, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[47] Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas Huang. Deep grabcut for object selection. arXiv preprint arXiv:1707.00243, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.729, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[48] Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas S Huang. Deep interactive object selection. In CVPR, pages 373-381, 2016. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.773, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[49] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[50] Yuhui Yuan, Rao Fu, Lang Huang, Weihong Lin, Chao Zhang, Xilin Chen, and Jingdong Wang. Hrformer: High-resolution vision transformer for dense predict. Advances in Neural Information Processing Systems, 34:7281-7293, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22299" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[51] Shiyin Zhang, Jun Hao Liew, Yunchao Wei, Shikui Wei, and Yao Zhao. Interactive object segmentation with inside-outside guidance. In CVPR, pages 12234-12244, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.468, + 0.205 + ], + "angle": 0, + "content": "[52] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.468, + 0.247 + ], + "angle": 0, + "content": "[53] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv:2304.06718, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22300" + } + ] +] \ No newline at end of file diff --git a/2023/SimpleClick_ Interactive Image Segmentation with Simple Vision Transformers/2b87aeaa-0b31-4e94-84d5-d9bfaea7e328_origin.pdf b/2023/SimpleClick_ Interactive Image Segmentation with Simple Vision Transformers/2b87aeaa-0b31-4e94-84d5-d9bfaea7e328_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ee395d2fbc32faf210c3176699b39d0ff2fdf3d1 --- /dev/null +++ b/2023/SimpleClick_ Interactive Image Segmentation with Simple Vision Transformers/2b87aeaa-0b31-4e94-84d5-d9bfaea7e328_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71bff23b15e7cb57784422933accf422d3491131ab56ba764a95520071e9f3ef +size 1870407 diff --git a/2023/SimpleClick_ Interactive Image Segmentation with Simple Vision Transformers/full.md b/2023/SimpleClick_ Interactive Image Segmentation with Simple Vision Transformers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e63be0d43a23e18ed38e2abdb32720567be0be4c --- /dev/null +++ b/2023/SimpleClick_ Interactive Image Segmentation with Simple Vision Transformers/full.md @@ -0,0 +1,283 @@ +# SimpleClick: Interactive Image Segmentation with Simple Vision Transformers + +Qin Liu, Zhenlin Xu, Gedas Bertasius, Marc Niethammer +University of North Carolina at Chapel Hill + +https://github.com/uncbiag/SimpleClick + +# Abstract + +Click-based interactive image segmentation aims at extracting objects with a limited user clicking. A hierarchical backbone is the de-facto architecture for current methods. Recently, the plain, non-hierarchical Vision Transformer (ViT) has emerged as a competitive backbone for dense prediction tasks. This design allows the original ViT to be a foundation model that can be finetuned for downstream tasks without redesigning a hierarchical backbone for pretraining. Although this design is simple and has been proven effective, it has not yet been explored for interactive image segmentation. To fill this gap, we propose SimpleClick, the first interactive segmentation method that leverages a plain backbone. Based on the plain backbone, we introduce a symmetric patch embedding layer that encodes clicks into the backbone with minor modifications to the backbone itself. With the plain backbone pretrained as a masked autoencoder (MAE), SimpleClick achieves state-of-the-art performance. Remarkably, our method achieves 4.15 NoC@90 on SBD, improving $21.8\%$ over the previous best result. Extensive evaluation on medical images demonstrates the generalizability of our method. We provide a detailed computational analysis, highlighting the suitability of our method as a practical annotation tool. + +# 1. Introduction + +The goal of interactive image segmentation is to obtain high-quality pixel-level annotations with limited user interaction such as clicking. Interactive image segmentation approaches have been widely applied to annotate large-scale image datasets, which drive the success of deep models in various applications, including video understanding [5, 49], self-driving [7], and medical imaging [32, 41]. Much research has been devoted to explore interactive image segmentation with different interaction types, such as bounding boxes [47], polygons [1], clicks [43], scribbles [45], and their combinations [51]. Among them, the click-based approach is most common due to its simplicity and well-established training and evaluation protocols. + +![](images/0a4bb07328405001327f2a24e63f407a2e00e221145706dc4c0f640e4c9cdf4b.jpg) +NoC@90 on SBD dataset +Figure 1. Interactive segmentation results on SBD [20]. The metric "NoC@90" denotes the number of clicks required to obtain $90\%$ IoU. The area of each bubble is proportional to the FLOPs of a model variant (Tab. 5). We show that plain ViTs outperform all hierarchical backbones for interactive image segmentation at a moderate computational cost. + +Recent advances in click-based approaches mainly lie in two orthogonal directions: 1) the development of more effective backbone networks and 2) the exploration of more elaborate refinement modules built upon the backbone. For the former direction, different hierarchical backbones, including both ConvNets [30,43] and ViTs [10,33], have been developed for interactive segmentation. For the latter direction, various refinement modules, including local refinement [10,30] and click imitation [34], have been proposed to further boost segmentation performance. In this work, we delve into the former direction and focus on exploring a plain backbone for interactive segmentation. + +A hierarchical backbone is the predominant architecture for current interactive segmentation methods. This design is deeply rooted in ConvNets, represented by ResNet [22], and has been adopted by ViTs, represented by the Swin Transformer [35]. The motivation for a hierarchical backbone stems from the locality of convolution operations, leading + +to insufficient model receptive field size without the hierarchy. To increase the receptive field size, ConvNets have to progressively downsample feature maps to capture more global contextual information. Therefore, they often require a feature pyramid network such as FPN [28] to aggregate multi-scale representations for high-quality segmentation. However, this reasoning no longer applies for a plain ViT, in which global information can be captured from the first self-attention block. Because all feature maps in the ViT are of the same resolution, the motivation for an FPN-like feature pyramid also no longer remains. The above reasoning is supported by a recent finding that a plain ViT can serve as a strong backbone for object detection [26]. This finding indicates a general-purpose ViT backbone might be suitable for other tasks, which then can decouple pretraining from finetuning and transfer the benefits from readily available pretrained ViT models (e.g. MAE [21]) to these tasks. However, although this design is simple and has been proven effective, it has not yet been explored in interactive segmentation. In this work, we propose SimpleClick, the first plain-backbone method for interactive segmentation. The core of SimpleClick is a plain ViT backbone that maintains single-scale representations throughout. We only use the last feature map from the plain backbone to build a simple feature pyramid for segmentation, largely decoupling the general-purpose backbone from the segmentation-specific modules. To make SimpleClick more efficient, we use a light-weight MLP decoder to transform the simple feature pyramid into a segmentation (see Sec. 3 for details). + +We extensively evaluate our method on 10 public benchmarks, including both natural and medical images. With the plain backbone pretrained as a MAE [21], our method achieves 4.15 NoC@90 on SBD, which outperforms the previous best method by $21.8\%$ without a complex FPN-like design and local refinement. We demonstrate the generalizability of our method by out-of-domain evaluation on medical images. We further analyze the computational efficiency of SimpleClick, highlighting its suitability as a practical annotation tool. + +Our main contributions are: + +- We propose SimpleClick, the first plain-backbone method for interactive image segmentation. +- SimpleClick achieves state-of-the-art performance on natural images and shows strong generalizability on medical images. +- SimpleClick meets the computational efficiency requirement for a practical annotation tool, highlighting its readiness for real-world applications. + +# 2. Related Work + +Interactive Image Segmentation Interactive image segmentation is a longstanding problem for which increasingly better solution approaches have been proposed. Early + +works [6, 16, 18, 40] tackle this problem using graphs defined over image pixels. However, these methods only focus on low-level image features and therefore tend to have difficulty with complex objects. + +Thriving on large datasets, ConvNets [10, 30, 43, 47, 48] have evolved as the dominant architecture for high-quality interactive segmentation. ConvNet-based methods have explored various interaction types, such as bounding boxes [47], polygons [1], clicks [43], and scribbles [45]. Click-based approaches are the most common due to their simplicity and well-established training and evaluation protocols. Xu et al. [48] first proposed a click simulation strategy that has been adopted by follow-up work [10, 34, 43]. DEXTR [36] extracts a target object by specifying its four extreme points (left-most, right-most, top, and bottom pixels). FCA-Net [31] demonstrates the critical role of the first click for better segmentation. Recently, ViTs have been applied to interactive segmentation. FocalClick [10] uses SegFormer [46] as the backbone network and achieves state-of-the-art segmentation results with high computational efficiency. iSegFormer [33] uses a Swin Transformer [35] as the backbone network for interactive segmentation on medical images. Besides the contribution of backbones, some works are exploring elaborate refinement modules built upon the backbone. FocalClick [10] and FocusCut [30] propose similar local refinement modules for high-quality segmentation. PseudoClick [34] proposes a click-imitation mechanism by estimating the next click to further reduce human annotation cost. Our method differs from all previous click-based methods in its plain, non-hierarchical ViT backbone, enjoying the benefits from readily available pretrained ViT models (e.g. MAE [21]). + +Vision Transformers for Non-Interactive Segmentation Recently, ViT-based approaches [17, 24, 44, 46, 50] have shown competitive performance on segmentation tasks compared to ConvNets. The original ViT [13] is a non-hierarchical architecture that only maintains single-scale feature maps throughout. SETR [52] and Segmenter [44] use the original ViT as the encoder for semantic segmentation. To allow for more efficient segmentation, the Swin Transformer [35] reintroduces a computational hierarchy into the original ViT architecture using shifted window attention, leading to a highly efficient hierarchical ViT backbone. SegFormer [46] designs hierarchical feature representations based on the original ViT using overlapped patch merging, combined with a light-weight MLP decoder for efficient segmentation. HRViT [17] integrates a high-resolution multi-branch architecture with ViTs to learn multi-scale representations. Recently, the original ViT has been reintroduced as a competitive backbone for semantic segmentation [8] and object detection [26], with the aid of MAE [21] pretraining and window attention. Inspired by this finding, we explore using a plain ViT as the backbone + +network for interactive segmentation. + +# 3. Method + +Our goal is not to propose new modules but to adapt a plain-ViT backbone for interactive segmentation with minimal modifications. Fig. 2 shows the overview of our method. Sec. 3.1 briefly describes the plain segmentation backbone. Sec. 3.2 shows how to adapt the backbone for interactive segmentation. Sec. 3.3 introduces other modules of SimpleClick. Sec. 3.4 describes the training and inference details of our method. + +# 3.1. Plain Segmentation Backbone + +We use a plain ViT [13] as the segmentation backbone. Unlike previous hierarchical backbones, a plain ViT only maintains single-scale feature maps throughout. Given an image, the patch embedding layer of the plain ViT first divides the image into non-overlapping fixed-size patches (e.g. $16 \times 16$ ). All patches are then flattened and linearly projected to a sequence of fixed-length tokens before feeding into self-attention blocks. These output tokens will be reshaped and upsampled to match the spatial size of the input image for segmentation. In this work, we consider three standard ViT backbones: ViT-B, ViT-L, and ViT-H (Tab. 1 shows the number of parameters of these backbones). For stable and efficient training, we use readily-available MAEs [21] as the pretrained weights for these backbones. + +# 3.2. Clicks Encoding for Segmentation Backbone + +We adapt the above plain segmentation backbone for interactive segmentation by turning user interactions into a form of guidance learned by the network. Similar to previous click-based methods [34, 43], we use clicks as a form of user interaction and encode each click as a small disk on a two-channel masks [34, 43] (one channel for positive clicks while the other for negative ones; positive clicks should be put in the foreground, while the negative ones should be put in the background). We automatically simulate human clicks for efficient training and evaluation. The clicks simulation process is described in Sec.3.4. Note that for human evaluation, a human-in-the-loop will provide all the clicks. + +Given human clicks or simulated clicks, we now introduce how to fuse them into the plain backbone. We first introduce a patch-embedding layer that is symmetric to the one in the backbone. Note that we also concatenate the previous segmentation to the clicks map as an additional channel for better performance. The two symmetric embedding layers operate on the image and the clicks map, respectively. The inputs are patchified, flattened, and projected to two vector sequences of the same dimension, followed by element-wise addition before feeding into the self-attention blocks. + +
Model↓ Module→ViT BackboneConv. NeckMLP Head
Ours-ViT-B (base)83.0 (89.3%)9.0 (9.7%)0.9 (1.0%)
Ours-ViT-L (large)290.8 (94.3%)16.5 (5.3%)1.1 (0.4%)
Ours-ViT-H (huge)604.0 (95.7%)25.8 (4.1%)1.3 (0.2%)
+ +Table 1. Number of parameters of our models. The unit is a million. The percentage of parameters is shown in bracket. Most parameters are used by the ViT backbone. + +# 3.3. Other Modules + +Simple Feature Pyramid For the hierarchical backbone, a feature pyramid is commonly produced by an FPN [28] to combine features from different stages. For the plain backbone, a feature pyramid can be generated in a much simpler way: by a set of parallel convolutional or deconvolutional layers using only the last feature map of the backbone. As shown in Fig. 2, given the input ViT feature map, a multiscale feature map can be produced by four convolutions with different strides. Though the effectiveness of this simple feature pyramid design is first demonstrated in ViTDet [26] for object detection, we show in this work the effectiveness of this simple feature pyramid design for interactive segmentation. We also propose several additional variants (Fig. 6) as part of an ablation study (Sec. 4.4). + +All-MLP Segmentation Head We implement a lightweight segmentation head using only MLP layers. It takes in the simple feature pyramid and produces a segmentation probability map of scale 1/4, followed by an upsampling operation to recover the original resolution. Note that this segmentation head avoids computationally demanding components and only accounts for up to $1\%$ of the model parameters (Tab. 1). The key insight is that with a powerful pretrained backbone, a lightweight segmentation head is sufficient for interactive segmentation. The proposed all-MLP segmentation head works in three steps. First, each feature map from the simple feature pyramid goes through an MLP layer to transform it to an identical channel dimension (i.e. $C_2$ in Fig. 2). Second, all feature maps are upsampled to the same resolution (i.e. 1/4 in Fig. 2) for concatenation. Third, the concatenated features are fused by another MLP layer to produce a single-channel feature map, followed by a sigmoid function to obtain a segmentation probability map, which is then transformed to a binary segmentation given a predefined threshold (i.e. 0.5). + +# 3.4. Training and Inference Settings + +Backbone Pretraining Our backbone models are pretrained as MAEs [21] on ImageNet-1K [11]. In MAE pretraining, the ViT models reconstruct the randomly masked pixels of images while learning a universal representation. This simple self-supervised approach turns out to be an efficient and + +![](images/8c33132e78719e53e82814db17c1001d00504536e4a869b37db8cd9742315476.jpg) +Figure 2. SimpleClick overview. Our method consists of three main modules: (a) a plain ViT backbone that maintains single-scale feature maps throughout; (b) a multi-scale simple feature pyramid that is generated from the last feature maps of the backbone by four parallel convolution or deconvolution layers; (c) a light-weight MLP decoder for segmentation. We also input the previous segmentation to improve the performance and to allow clicking from existing masks. User clicks are encoded as a two-channel disk map, combined with the previous segmentation as input. Boxes in blue are intermediate feature maps. The positional encoding is not shown for brevity. + +scalable way to pretrain ViT models [21]. In this work, we do not perform pretraining ourselves. Instead, we simply use the readily available pretrained MAE weights from [21]. + +Clicks Simulation and End-to-end Finetuning With the pretrained backbone, we finetune our model end-to-end on the interactive segmentation task. The finetuning pipeline can be briefly described as follows. First, we automatically simulate clicks based on the current segmentation and gold standard segmentation, without a human-in-the-loop providing the clicks. Specifically, we use a combination of random and iterative click simulation strategies, inspired by RITM [43]. The random click simulation strategy generates clicks in parallel, without considering the order of the clicks. The iterative click simulation strategy generates clicks iteratively, where the next click should be placed on the erroneous region of a prediction that was obtained using the previous clicks. This strategy is more similar to human clicking behavior. Second, we incorporate the segmentation from the previous interaction as an additional input for the backbone, further improving the segmentation quality. This also allows our method to refine from an existing segmentation, which is a desired feature for a practical annotation tool. We use the normalized focal loss [43] (NFL) to train all our models. Previous works [10, 43] show that NFL converges faster and achieves better performance than the widely used binary cross entropy loss for interactive segmentation tasks. Similar training pipelines have been proposed by RITM [43] and its follow-up works [9, 10, 34]. + +Human Evaluation and Automatic Evaluation There are two inference modes: automatic evaluation and human eval + +uation. For automatic evaluation, clicks are automatically simulated based on the current segmentation and gold standard. For human evaluation, a human-in-the-loop provides all clicks based on their subjective evaluation of current segmentation results. We use automatic evaluation for quantitative analyses and human evaluation for a qualitative assessment of the interactive segmentation behavior. + +# 4. Experiments + +Datasets We conducted experiments on 10 public datasets including 7 natural image datasets and 3 medical datasets. The details are as follows: + +- GrabCut [40]: 50 images (50 instances), each with clear foreground and background differences. +- Berkeley [37]: 96 images (100 instances); this dataset shares a small portion of images with GrabCut. +- DAVIS [39]: 50 videos; we only use the same 345 frames as used in [10, 30, 34, 43] for evaluation. +- Pascal VOC [14]: 1449 images (3427 instances) in the validation set. We only test on the validation set. +- SBD [20]: 8498 training images (20172 instances) and 2857 validation images (6671 instances). Following previous works [10, 30, 43], we train our model on the training set and evaluate on the validation set. +- COCO [29]+LVIS [19] (C+L): COCO contains 118K training images (1.2M instances); LVIS shares the same images with COCO but has much higher segmentation quality. We combine the two datasets for training. +- ssTEM [15]: two image stacks, each contains 20 medical images. We use the same stack that was used in [34]. + +![](images/95a272e72935b50e99c599d489175aead5a44503e4efee2e376f3def16a5fb21.jpg) +Figure 3. Convergence analysis for models trained on either SBD [20] or COCO [29] + LVIS [19] (C+L). We report results on SBD [20] and Pascal VOC [14]. The metric is mean IoU given $k$ clicks ( $\text{mIoU} @ k$ ). Our models in general require fewer clicks for a given accuracy level. + +![](images/403668dc1826ba85844aa3b720a498bf1b1e2ccba3df7c6302fa3ec6c264c16e.jpg) + +
MethodBackboneGrabCutBerkeleySBDDAVISPascal VOC
NoC85NoC90NoC85NoC90NoC85NoC90NoC85NoC90NoC85NoC90
\( \nmid \)DIOS [48] \( _{CVPR16} \)FCN-6.04-8.65---12.586.88-
\( \nmid \)FCA-Net [31] \( _{CVPR20} \)ResNet-101-2.08-3.92---7.572.69-
\( \nmid \)LD [27] \( _{CVPR18} \)VGG-193.204.79--7.4110.785.059.57--
\( \nmid \)BRS [23] \( _{CVPR19} \)DenseNet2.603.60-5.086.599.785.588.24--
\( \nmid \)f-BRS [42] \( _{CVPR20} \)ResNet-1012.302.72-4.574.817.735.047.41--
\( \nmid \)RITM [43] \( _{Preprint21} \)HRNet-181.762.041.873.223.395.434.946.712.513.03
\( \nmid \)CDNet [9] \( _{ICCV21} \)ResNet-341.862.181.953.275.187.895.006.893.614.51
\( \nmid \)PseudoClick [34] \( _{ECCV22} \)HRNet-181.682.041.853.233.385.404.816.572.342.74
\( \nmid \)FocalClick [10] \( _{CVPR22} \)HRNet-18s1.862.06-3.144.306.524.926.48--
\( \nmid \)FocalClick [10] \( _{CVPR22} \)SegF-B01.661.90-3.144.346.515.027.06--
\( \nmid \)FocusCut [30] \( _{CVPR22} \)ResNet-501.601.78\( 1.85^{\dagger } \)3.443.625.665.006.38--
\( \nmid \)FocusCut [30] \( _{CVPR22} \)ResNet-1011.461.64\( 1.81^{\dagger } \)3.013.405.314.856.22--
\( \nmid \)OursViT-B1.401.541.442.463.285.244.105.482.382.81
\( \nmid \)OursViT-L1.381.461.402.332.694.464.125.391.952.30
\( \nmid \)OursViT-H1.321.441.362.092.514.154.205.341.882.20
\( \nmid \)RITM [43] \( _{Preprint21} \)HRNet-321.461.561.432.103.595.714.115.342.192.57
\( \nmid \)CDNet [9] \( _{ICCV21} \)ResNet-341.401.521.472.064.307.044.275.562.743.30
\( \nmid \)PseudoClick [34] \( _{ECCV22} \)HRNet-321.361.501.402.083.465.543.795.111.942.25
\( \nmid \)FocalClick [10] \( _{CVPR22} \)SegF-B01.401.661.592.274.566.864.045.492.973.52
\( \nmid \)FocalClick [10] \( _{CVPR22} \)SegF-B31.441.501.551.923.535.593.614.902.462.88
\( \nmid \)OursViT-B1.381.481.361.973.435.623.665.062.062.38
\( \nmid \)OursViT-L1.321.401.341.892.954.893.264.811.721.96
\( \nmid \)OursViT-H1.381.501.361.752.854.703.414.781.761.98
+ +Table 2. Comparison with previous results. We report results on five benchmarks: GrabCut [40], Berkeley [37], SBD [20], DAVIS [39], and Pascal VOC [14]. The best results are set in bold. $\nsubseteq$ denotes a model trained on Pascal; $\downarrow$ denotes a model trained on SBD; $\sqcup$ denotes a model trained on COCO [29] + LVIS [19] (C+L); $\dagger$ denotes a number reproduced by the released or retrained models. Our models achieve state-of-the-art performance on all benchmarks. + +- BraTS [4]: 369 magnetic resonance image (MRI) volumes; we test on the same 369 slices used in [34]. +- OAIZIB [2]: 507 MRI volumes; we test on the same 150 slices (300 instances) as used in [33]. + +Evaluation Metrics Following previous works [30,42,43], we automatically simulate user clicks by comparing the current segmentation with the gold standard. In this simulation, the next click will be put at the center of the region with the largest error. We use the Number of Clicks (NoC) + +as the evaluation metric to calculate the number of clicks required to achieve a target Intersection over Union (IoU). We set two target IoUs: $85\%$ and $90\%$ , represented by $\mathrm{NoC}\% 85$ and $\mathrm{NoC}\% 90$ respectively. The maximum number of clicks for each instance is set to 20. We also use the average IoU given $k$ clicks ( $\mathrm{mIoU} @ k$ ) as an evaluation metric to measure the segmentation quality given a fixed number of clicks. + +Implementation Details We implement our models using Python and PyTorch [38]. We implement three models + +![](images/d708b42d1e3d20a3f9719d6baca1cb8d5099ca4dc24f23ccaf7dfb58edc3db65.jpg) + +![](images/f272ee5755fba79b1d467fdc717a5c1fe2a925081f2a7f26b38a04addfe55e92.jpg) + +![](images/7d956fb6852536a52d07d67e800ed4ac655d3a1aa9baa93f30097890a0c1c673.jpg) + +![](images/f0f6086146e3d93c579d3ae0f7cfa6ced33039a26070565b6a24aabcae8000dd.jpg) +Figure 4. Histogram analysis of IoU given a predefined number of clicks $k$ (IoU@k). We report analysis on SBD [20] and Pascal VOC [14] with models trained on SBD. Compared with the two baselines, our models achieve higher-quality segmentation with fewer failure cases. + +![](images/82c52d1ea5aee0d2d8ef34005a7dac3ec029c5fdc6eeae2081e2e1c661e6f726.jpg) + +![](images/93f392859a70a3b67630aa0a53ab74e3522b235574105659a481aab2b9682f6b.jpg) + +![](images/156383fcea46ded5d4d404bcd1c98849c93f8ba081855f79ad8bcc76c295ffe0.jpg) +Figure 5. Convergence analysis for two medical image datasets: BraTS [4] and OAIZIB [2]. Models are trained on either SBD [20] or COCO [29] + LVIS [19] (denoted as $\mathrm{C} + \mathrm{L}$ ). The metric is mean IoU given $k$ clicks. Overall, our models require fewer clicks for a given accuracy level. The performance gain is more prominent for bigger models (e.g. ViT-H) or larger training sets (e.g. C+L). + +![](images/b6940d6819f52f98d051c1c61d91f6ed4de83aef58d207e70fdc409e394d0563.jpg) + +based on three vanilla ViT models (i.e. ViT-B, ViT-L, and ViT-H). These backbone models are initialized with the MAE pretrained weights, and then are finetuned end-to-end with other modules. We train our models on either SBD or COCO+LVIS with 55 epochs; the initial learning rate is set to $5 \times 10^{-5}$ and decreases to $5 \times 10^{-6}$ after epoch 50. We set the batch size to 140 for ViT-Base, 72 for ViT-Large, and 32 for ViT-Huge to fit the models into GPU memory. All our models are trained on four NVIDIA RTX A6000 GPUs. We use the following data augmentation techniques: random resizing (scale range from 0.75 to 1.25), random flipping and rotation, random brightness contrast, and random cropping. Though the ViT backbone was pretrained on images of size + +$224 \times 224$ , we finetune on $448 \times 448$ with non-shifting window attention for better performance. We optimize using Adam with $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ . + +# 4.1. Comparison with Previous Results + +We show in Tab. 2 the comparisons with previous state-of-the-art results. Our models achieve the best performance on all the five benchmarks. Remarkably, when trained on SBD training set, our ViT-H model achieves 4.15 NoC@90 on the SBD validation set, outperforming the previous best score by $21.8\%$ . Since the SBD validation set contains the largest number of instances (6671 instances) among the five benchmarks, this improvement is convincing. When trained + +
ModelssTEM +mIoU@10BraTS +mIoU@10 / 20OAIZIB +mIoU@10 / 20
↓RITM-H18 [43]93.1587.05 / 90.4771.04 / 78.52
↓CDN-RN34 [9]66.7258.34 / 82.0738.07 / 61.17
↓RITM-H32 [43]94.1188.34 / 89.2575.27 / 75.18
↓CDN-RN34 [9]88.4680.24 / 86.6363.19 / 74.21
↓FC-SF-B0 [10]92.6286.02 / 90.7474.08 / 79.14
↓FC-SF-B3 [10]93.6188.62 / 90.5875.77 / 80.08
↓Ours-ViT-B93.7286.98 / 90.6776.05 / 79.61
↓Ours-ViT-L94.3488.43 / 90.8477.34 / 79.97
↓Ours-ViT-H94.0888.98 / 91.0077.50 / 80.10
+ +Table 3. Out-of-domain evaluation on three medical image datasets: ssTEM [15], BraTS [4], and OAIZIB [2]. Our models generalize very well on the three datasets, without finetuning. + +on COCO+LVIS, our models also achieve state-of-the-art performance on all benchmarks. Fig. 7 shows several segmentation cases on DAVIS, including the worst case. Note that the DAVIS dataset requires high-quality segmentations because all its instances have a high-quality gold standard. Our models still achieve the state-of-the-art on DAVIS without using specific modules, such as a local refinement module [10], which is beneficial for high-quality segmentation. Fig. 3 shows that our method converges better than other methods with sufficient clicks, leading to fewer failure cases as shown in Fig. 4. We only report results on SBD and Pascal VOC, the top two largest datasets. + +# 4.2. Out-of-Domain Evaluation on Medical Images + +We further evaluate the generalizability of our models on three medical image datasets: ssTEM [15], BraTS [3], and OAIZIB [2]. Tab. 3 reports the evaluation results on these three datasets. Fig. 5 shows the convergence analysis on BraTS and OAIZIB. Overall, our models generalize well to medical images. We also find that the models trained on larger datasets (i.e. $\mathrm{C + L}$ ) generalize better than the models trained on smaller datasets (i.e. SBD). + +# 4.3. Towards Practical Annotation Tool + +Tiny Backbone To allow for practical applications, especially on low-end devices with limited computational resources, we implement an extremely tiny backbone (i.e. ViT-xTiny) for SimpleClick. Compared with ViT-Base, ViT-xTiny decreases the embedding dimension from 768 to 160 and the number of attention blocks from 12 to 8. We end up with a SimpleClick-xTiny model, which is comparable with the tiny FocalClick models in terms of parameters. Comparison results in Tab. 4 show that our model outperforms FocalClick models, even though it is trained from scratch due to the lack of readily available pretrained weights. + +Computational Analysis Tab. 5 shows a comparison of computational requirements with respect to model parameters, FLOPs, GPU memory consumption, and speed; the speed is measured by seconds per click (SPC). Fig. 1 shows + +
ModelBackbonePretrainedParams/MNoC85NoC90
FocalClickHRNet-18s-S1✓4.224.747.29
FocalClickSegFormer-B0-S1✓3.724.987.60
SimpleClickViT-xTiny✗3.724.717.09
+ +Table 4. Comparison results on SBD for tiny models. All models are trained on C+L with 230 epochs. Our SimpleClick-xTiny model outperforms FocalClick models without pretraining. + +
BackboneParams/MFLOPs/GMem/G↓SPC/ms
HR-18s 400 [43]4.2217.940.5054
HR-18 400 [43]10.0330.990.5256
HR-32 400 [43]30.9583.121.1286
Swin-B 400 [33]87.44138.211.4136
Swin-L 400 [33]195.90302.782.1444
SegF-B0 256 [10]3.723.420.1037
SegF-B3 256 [10]45.6624.750.3253
ResN-34 384 [9]23.47113.600.2534
ResN-50 384 [30]40.3678.820.85331
ResN-101 384 [30]59.35100.760.89355
Ours-ViT-xT 2243.722.630.1717
Ours-ViT-xT 4483.7210.520.2329
Ours-ViT-B 22496.4642.440.5134
Ours-ViT-B 44896.46169.780.8754
Ours-ViT-L 448322.18532.871.7286
Ours-ViT-H 448659.391401.933.22132
+ +Table 5. Computation comparison for model parameters, FLOPs, GPU memory consumption (measured by the maximum GPU memory managed by PyTorch's caching allocator), and speed (measured by seconds per click). Each method is denoted by its backbone. The small number in front of the model denotes the input size $(448\times 448$ for our models by default). Even for our ViT-H model, the speed (132ms) and memory consumption (3.22G) are sufficient to meet the requirements of a practical annotation tool. + +the interactive segmentation performance of methods in terms of FLOPs. In Fig. 1 and Tab. 5, each method is denoted by its backbone. For fair comparison, we evaluate all the methods on the same benchmark (i.e. GrabCut) and using the same computer (GPU: NVIDIA RTX A6000, CPU: Intel Silver×2). We only calculate the FLOPs in a single forward pass. For methods like FocusCut which require multiple forward passes for each click, the FLOPs may be much higher than reported. By default, our method takes images of size $448 \times 448$ as the fixed input. Even for our ViT-H model, the speed (132ms) and memory consumption (3.22G) is sufficient to meet the requirements of a practical annotation tool. + +# 4.4. Ablation Study + +In this section, we ablate the backbone finetuning and feature pyramid design. Tab. 6 shows the ablation results. By default, we finetune the backbone along with other modules. As an ablation, we freeze the backbone during finetuning, leading to significantly worse performance. This ablation is explainable considering the ViT backbone takes most of the model parameters (Tab. 1). For the second ablation, + +![](images/35802b1d10ba2be9ff5c3882c8a36bb273d9da7e497eb7cc5fbe7f5d9e7c158d.jpg) +(a) simple feature pyramid + +![](images/16cf33bbdfb30df78c8ac692f33c5127ac9495c2039b41c38ad9d994b593755d.jpg) +(b) single-scale variant + +![](images/4522563d1266cff3666743c2bd79b0c55bd0132df7861db3d02715fa953d6ed5.jpg) +(c) parallel variant + +![](images/44b772ea34f42aabdba4d854168e21e494dc508808f4b8567fbb51b74c2ea160.jpg) +(d) partial variant + +![](images/1dc18ac633ff811b441fc0705291c390879570e65fc8e1014049daeae3c9388c.jpg) +Figure 6. Simple feature pyramid and its variants: b) the single-scale variant ablates the multi-scale property; c) the parallel variant evenly extracts features from the backbone; d) the partial variant freezes the second half parameters of the backbone. Tab. 6 shows the comparison results. +(a) the worst case +Figure 7. Segmentation results on DAVIS [39]: (a) the worst case; (b) a challenging case; (c) five normal cases. The backbone model is ViT-L trained on COCO [29] + LVIS [19]. The segmentation probability maps are shown in blue; the segmentation maps are overlaid in red on the original images. The clicks are shown as green (positive click) or red (negative click) dots on the image. GT denotes ground truth. + +![](images/8bfdd83a2fd3634da377d7b52ef40491c63f15a354464fd503e24104a0295645.jpg) +(b) a challenging case + +![](images/3ab6c7acf59ed0a008a8399380f51bd7d7d1b4e27840ece7b26b25125883198d.jpg) + +![](images/bcebc9d9442dfdb80f687f7228ba9aba5b368ee468147585e8350ffb35badee1.jpg) +(c) five normal cases + +![](images/7b102bbd0da3fe420b8a57ea26bf5f3c3b44beaa5381e4943b6822c88a8e80b5.jpg) + +
FP designfrozen ViTViT-BViT-L
SBDPascalSBDPascal
(a) simple FP✓11.486.939.755.59
(a) simple FP✗5.242.534.462.15
(b) single-scale✗6.562.805.532.48
(c) parallel✗7.213.096.262.79
(d) partial✗8.294.347.514.25
+ +Table 6. Ablation study on backbone finetuning and feature pyramid (FP) design (Fig. 6). The metric is NoC@90. We have three findings in this ablation: 1) freezing the ViT backbone during finetuning significantly deteriorates the performance; 2) the multiscale property matters for the simple feature pyramid; 3) the last feature map from the backbone is sufficient to build an effective feature pyramid. + +we compare the default simple feature pyramid design with three variants depicted in Fig. 6 (i.e. (b), (c), and (d)). First, we observe that the multi-scale representation matters for the feature pyramid. By ablating the multi-scale property in the simple feature pyramid, the performance drops considerably. We also notice that the last feature map from the backbone is strong enough to build the feature pyramid. The parallel feature pyramid generated by multi-stage feature maps + +from the backbone does not surpass the simple feature pyramid that only uses the last feature map of the backbone. + +# 5. Limitations and Remarks + +Our best-performing model (ViT-H) is much larger than existing models, leading to concerns about an unfair comparison. Our method is not prompt-efficient as every click update requires recomputing image features. The recent advancements in interactive segmentation [25, 53] show an elegant solution to this issue. Besides, these methods use sparse vectors to represent clicks, which might be more efficient than dense masks. Other than these, our models may fail in some challenging scenarios such as objects with very thin and elongated shapes or cluttered occlusions((a) and (b) in Fig. 7). We leave the improvements for future work. + +We are entering an era of large-scale pretraining on multimodal foundation models, which is dramatically transforming the landscape of vision and language tasks. In this context, we hope SimpleClick will serve as a strong baseline for a new wave of high-performing interactive segmentation methods based on ViTs and large-scale pretraining. + +# 6. Conclusions + +We proposed SimpleClick, a plain-backbone model for interactive image segmentation. Our method leveraged a general-purpose ViT backbone that can benefit from advancements in pretrained ViT models. With the readily-available MAE weights, SimpleClick achieved state-of-the-art performance on natural images and demonstrated strong generalizability to medical images. Our method is simple yet effective, highlighting its suitability as a strong baseline model and a practical annotation tool. + +# Acknowledgements + +We thank our reviewers and ACs for their invaluable comments and suggestions. Research reported in this publication was supported by the National Institutes of Health (NIH) under award number NIH 1R01AR072013. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH. + +# References + +[1] David Acuna, Huan Ling, Amlan Kar, and Sanja Fidler. Efficient interactive annotation of segmentation datasets with polygon-rnn++. In CVPR, pages 859-868, 2018. 1, 2 +[2] Felix Ambellan, Alexander Tack, Moritz Ehlke, and Stefan Zachow. Automated segmentation of knee bone and cartilage combining statistical shape knowledge and convolutional neural networks: Data from the osteoarthritis initiative. Medical image analysis, 52:109-118, 2019. 5, 6, 7 +[3] Xue Bai and Guillermo Sapiro. A geodesic framework for fast interactive image and video segmentation and matting. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 7 +[4] Ujjwal Baid, Satyam Ghodasara, Suyash Mohan, Michel Bilello, Evan Calabrese, Errol Colak, Keyvan Farahani, Jayashree Kalpathy-Cramer, Felipe C Kitamura, Sarthak Pati, et al. The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314, 2021. 5, 6, 7 +[5] Gedas Bertasius and Lorenzo Torresani. Classifying, segmenting, and tracking object instances in video with mask propagation. In CVPR, pages 9739-9748, 2020. 1 +[6] Yuri Y Boykov and M-P Jolly. Interactive graph cuts for optimal boundary & region segmentation of objects in ND images. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, volume 1, pages 105-112. IEEE, 2001. 2 +[7] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 1 +[8] Wuyang Chen, Xianzhi Du, Fan Yang, Lucas Beyer, Xiaohua Zhai, Tsung-Yi Lin, Huizhong Chen, Jing Li, Xiaodan Song, Zhangyang Wang, et al. A simple single-scale vision + +transformer for object localization and instance segmentation. arXiv preprint arXiv:2112.09747, 2021. 2 +[9] Xi Chen, Zhiyan Zhao, Feiwu Yu, Yilei Zhang, and Manni Duan. Conditional diffusion for interactive segmentation. In ICCV, pages 7345-7354, 2021. 4, 5, 7 +[10] Xi Chen, Zhiyan Zhao, Yilei Zhang, Manni Duan, Donglian Qi, and Hengshuang Zhao. FocalClick: Towards practical interactive image segmentation. In CVPR, pages 1300-1309, 2022. 1, 2, 4, 5, 7 +[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3 +[12] Zhipeng Ding, Xu Han, Peirong Liu, and Marc Niethammer. Local temperature scaling for probability calibration. In ICCV, pages 6889-6899, 2021. 3 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3 +[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 4, 5, 6 +[15] Stephan Gerhard, Jan Funke, Julien Martel, Albert Cardona, and Richard Fetter. Segmented anisotropic sstem dataset of neural tissue. figshare, pages 0–0, 2013. 4, 7 +[16] Leo Grady. Random walks for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 28(11):1768-1783, 2006. 2 +[17] Jiaqi Gu, Hyoukjun Kwon, Dilin Wang, Wei Ye, Meng Li, Yu-Hsin Chen, Liangzhen Lai, Vikas Chandra, and David Z Pan. Multi-scale high-resolution vision transformer for semantic segmentation. In CVPR, pages 12094-12103, 2022. 2 +[18] Varun Gulshan, Carsten Rother, Antonio Criminisi, Andrew Blake, and Andrew Zisserman. Geodesic star convexity for interactive image segmentation. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3129-3136. IEEE, 2010. 2 +[19] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, pages 5356-5364, 2019. 4, 5, 6, 8 +[20] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In 2011 international conference on computer vision, pages 991-998. IEEE, 2011. 1, 4, 5, 6 +[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 2, 3, 4 +[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1 + +[23] Won-Dong Jang and Chang-Su Kim. Interactive image segmentation via backpropagating refinement scheme. In CVPR, pages 5297-5306, 2019. 5 +[24] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM computing surveys (CSUR), 54(10s):1-41, 2022. 2 +[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv:2304.02643, 2023. 8 +[26] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. arXiv preprint arXiv:2203.16527, 2022. 2, 3 +[27] Zhuwen Li, Qifeng Chen, and Vladlen Koltun. Interactive image segmentation with latent diversity. In CVPR, pages 577-585, 2018. 5 +[28] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 2, 3 +[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 4, 5, 6, 8 +[30] Zheng Lin, Zheng-Peng Duan, Zhao Zhang, Chun-Le Guo, and Ming-Ming Cheng. FocusCut: Diving into a focus view in interactive segmentation. In CVPR, pages 2637-2646, 2022. 1, 2, 4, 5, 7 +[31] Zheng Lin, Zhao Zhang, Lin-Zhuo Chen, Ming-Ming Cheng, and Shao-Ping Lu. Interactive image segmentation with first click attention. In CVPR, pages 13339-13348, 2020. 2, 5 +[32] Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Arnaud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen Awm Van Der Laak, Bram Van Ginneken, and Clara I Sánchez. A survey on deep learning in medical image analysis. Medical image analysis, 42:60-88, 2017. 1 +[33] Qin Liu, Zhenlin Xu, Yining Jiao, and Marc Niethammer. iSegFormer: Interactive segmentation via transformers with application to 3d knee mr images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 464-474. Springer, 2022. 1, 2, 5, 7 +[34] Qin Liu, Meng Zheng, Benjamin Planche, Srikrishna Karanam, Terrence Chen, Marc Niethammer, and Ziyan Wu. PseudoClick: Interactive image segmentation with click imitation. arXiv preprint arXiv:2207.05282, 2022. 1, 2, 3, 4, 5 +[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 1, 2 +[36] Kevis-Kokitsi Maninis, Sergi Caelles, Jordi Pont-Tuset, and Luc Van Gool. Deep extreme cut: From extreme points to object segmentation. In CVPR, pages 616-625, 2018. 2 +[37] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and + +its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, volume 2, pages 416-423. IEEE, 2001. 4, 5 +[38] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5 +[39] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. 4, 5, 8 +[40] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. "grabcut" interactive foreground extraction using iterated graph cuts. ACM transactions on graphics (TOG), 23(3):309-314, 2004. 2, 4, 5 +[41] Dinggang Shen, Guorong Wu, and Heung-II Suk. Deep learning in medical image analysis. Annual review of biomedical engineering, 19:221, 2017. 1 +[42] Konstantin Sofiuk, Ilia Petrov, Olga Barinova, and Anton Konushin. f-brs: Rethinking backpropagating refinement for interactive segmentation. In CVPR, pages 8623-8632, 2020. 5 +[43] Konstantin Sofiuk, Ilia A Petrov, and Anton Konushin. Reviving iterative training with mask guidance for interactive segmentation. arXiv preprint arXiv:2102.06583, 2021. 1, 2, 3, 4, 5, 7 +[44] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, pages 7262-7272, 2021. 2 +[45] Jiajun Wu, Yibiao Zhao, Jun-Yan Zhu, Siwei Luo, and Zhuowen Tu. Milcut: A sweeping line multiple instance learning paradigm for interactive image segmentation. In CVPR, pages 256-263, 2014. 1, 2 +[46] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. SegFormer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 2 +[47] Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas Huang. Deep grabcut for object selection. arXiv preprint arXiv:1707.00243, 2017. 1, 2 +[48] Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas S Huang. Deep interactive object selection. In CVPR, pages 373-381, 2016. 2, 5 +[49] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 1 +[50] Yuhui Yuan, Rao Fu, Lang Huang, Weihong Lin, Chao Zhang, Xilin Chen, and Jingdong Wang. Hrformer: High-resolution vision transformer for dense predict. Advances in Neural Information Processing Systems, 34:7281-7293, 2021. 2 + +[51] Shiyin Zhang, Jun Hao Liew, Yunchao Wei, Shikui Wei, and Yao Zhao. Interactive object segmentation with inside-outside guidance. In CVPR, pages 12234-12244, 2020. 1 +[52] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. 2 +[53] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv:2304.06718, 2023. 8 \ No newline at end of file diff --git a/2023/SimpleClick_ Interactive Image Segmentation with Simple Vision Transformers/images.zip b/2023/SimpleClick_ Interactive Image Segmentation with Simple Vision Transformers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..11f1fbea77ad9986c54aadaf2ae327d2ed9517e1 --- /dev/null +++ b/2023/SimpleClick_ Interactive Image Segmentation with Simple Vision Transformers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0a3bfbbc097878808f8633227b49a3831909d49de75a237507fc4e7bb215d75 +size 772351 diff --git a/2023/SimpleClick_ Interactive Image Segmentation with Simple Vision Transformers/layout.json b/2023/SimpleClick_ Interactive Image Segmentation with Simple Vision Transformers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..893091fa1fc3fa3f08db73fe35a91a78649aab6a --- /dev/null +++ b/2023/SimpleClick_ Interactive Image Segmentation with Simple Vision Transformers/layout.json @@ -0,0 +1,7482 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "type": "text", + "content": "SimpleClick: Interactive Image Segmentation with Simple Vision Transformers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 157, + 143, + 436, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 143, + 436, + 171 + ], + "spans": [ + { + "bbox": [ + 157, + 143, + 436, + 171 + ], + "type": "text", + "content": "Qin Liu, Zhenlin Xu, Gedas Bertasius, Marc Niethammer \nUniversity of North Carolina at Chapel Hill" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 193, + 174, + 400, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 174, + 400, + 185 + ], + "spans": [ + { + "bbox": [ + 193, + 174, + 400, + 185 + ], + "type": "text", + "content": "https://github.com/uncbiag/SimpleClick" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 239, + 290, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 514 + ], + "type": "text", + "content": "Click-based interactive image segmentation aims at extracting objects with a limited user clicking. A hierarchical backbone is the de-facto architecture for current methods. Recently, the plain, non-hierarchical Vision Transformer (ViT) has emerged as a competitive backbone for dense prediction tasks. This design allows the original ViT to be a foundation model that can be finetuned for downstream tasks without redesigning a hierarchical backbone for pretraining. Although this design is simple and has been proven effective, it has not yet been explored for interactive image segmentation. To fill this gap, we propose SimpleClick, the first interactive segmentation method that leverages a plain backbone. Based on the plain backbone, we introduce a symmetric patch embedding layer that encodes clicks into the backbone with minor modifications to the backbone itself. With the plain backbone pretrained as a masked autoencoder (MAE), SimpleClick achieves state-of-the-art performance. Remarkably, our method achieves 4.15 NoC@90 on SBD, improving " + }, + { + "bbox": [ + 46, + 239, + 290, + 514 + ], + "type": "inline_equation", + "content": "21.8\\%" + }, + { + "bbox": [ + 46, + 239, + 290, + 514 + ], + "type": "text", + "content": " over the previous best result. Extensive evaluation on medical images demonstrates the generalizability of our method. We provide a detailed computational analysis, highlighting the suitability of our method as a practical annotation tool." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 537, + 128, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 128, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 128, + 550 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 557, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 715 + ], + "type": "text", + "content": "The goal of interactive image segmentation is to obtain high-quality pixel-level annotations with limited user interaction such as clicking. Interactive image segmentation approaches have been widely applied to annotate large-scale image datasets, which drive the success of deep models in various applications, including video understanding [5, 49], self-driving [7], and medical imaging [32, 41]. Much research has been devoted to explore interactive image segmentation with different interaction types, such as bounding boxes [47], polygons [1], clicks [43], scribbles [45], and their combinations [51]. Among them, the click-based approach is most common due to its simplicity and well-established training and evaluation protocols." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 228, + 544, + 406 + ], + "blocks": [ + { + "bbox": [ + 309, + 213, + 425, + 225 + ], + "lines": [ + { + "bbox": [ + 309, + 213, + 425, + 225 + ], + "spans": [ + { + "bbox": [ + 309, + 213, + 425, + 225 + ], + "type": "text", + "content": "NoC@90 on SBD dataset" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 228, + 544, + 406 + ], + "lines": [ + { + "bbox": [ + 308, + 228, + 544, + 406 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 544, + 406 + ], + "type": "image", + "image_path": "0a4bb07328405001327f2a24e63f407a2e00e221145706dc4c0f640e4c9cdf4b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 408, + 547, + 475 + ], + "lines": [ + { + "bbox": [ + 304, + 408, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 547, + 475 + ], + "type": "text", + "content": "Figure 1. Interactive segmentation results on SBD [20]. The metric \"NoC@90\" denotes the number of clicks required to obtain " + }, + { + "bbox": [ + 304, + 408, + 547, + 475 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 408, + 547, + 475 + ], + "type": "text", + "content": " IoU. The area of each bubble is proportional to the FLOPs of a model variant (Tab. 5). We show that plain ViTs outperform all hierarchical backbones for interactive image segmentation at a moderate computational cost." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 497, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 547, + 640 + ], + "type": "text", + "content": "Recent advances in click-based approaches mainly lie in two orthogonal directions: 1) the development of more effective backbone networks and 2) the exploration of more elaborate refinement modules built upon the backbone. For the former direction, different hierarchical backbones, including both ConvNets [30,43] and ViTs [10,33], have been developed for interactive segmentation. For the latter direction, various refinement modules, including local refinement [10,30] and click imitation [34], have been proposed to further boost segmentation performance. In this work, we delve into the former direction and focus on exploring a plain backbone for interactive segmentation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "A hierarchical backbone is the predominant architecture for current interactive segmentation methods. This design is deeply rooted in ConvNets, represented by ResNet [22], and has been adopted by ViTs, represented by the Swin Transformer [35]. The motivation for a hierarchical backbone stems from the locality of convolution operations, leading" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22290" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 418 + ], + "type": "text", + "content": "to insufficient model receptive field size without the hierarchy. To increase the receptive field size, ConvNets have to progressively downsample feature maps to capture more global contextual information. Therefore, they often require a feature pyramid network such as FPN [28] to aggregate multi-scale representations for high-quality segmentation. However, this reasoning no longer applies for a plain ViT, in which global information can be captured from the first self-attention block. Because all feature maps in the ViT are of the same resolution, the motivation for an FPN-like feature pyramid also no longer remains. The above reasoning is supported by a recent finding that a plain ViT can serve as a strong backbone for object detection [26]. This finding indicates a general-purpose ViT backbone might be suitable for other tasks, which then can decouple pretraining from finetuning and transfer the benefits from readily available pretrained ViT models (e.g. MAE [21]) to these tasks. However, although this design is simple and has been proven effective, it has not yet been explored in interactive segmentation. In this work, we propose SimpleClick, the first plain-backbone method for interactive segmentation. The core of SimpleClick is a plain ViT backbone that maintains single-scale representations throughout. We only use the last feature map from the plain backbone to build a simple feature pyramid for segmentation, largely decoupling the general-purpose backbone from the segmentation-specific modules. To make SimpleClick more efficient, we use a light-weight MLP decoder to transform the simple feature pyramid into a segmentation (see Sec. 3 for details)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 419, + 289, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 289, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 289, + 538 + ], + "type": "text", + "content": "We extensively evaluate our method on 10 public benchmarks, including both natural and medical images. With the plain backbone pretrained as a MAE [21], our method achieves 4.15 NoC@90 on SBD, which outperforms the previous best method by " + }, + { + "bbox": [ + 47, + 419, + 289, + 538 + ], + "type": "inline_equation", + "content": "21.8\\%" + }, + { + "bbox": [ + 47, + 419, + 289, + 538 + ], + "type": "text", + "content": " without a complex FPN-like design and local refinement. We demonstrate the generalizability of our method by out-of-domain evaluation on medical images. We further analyze the computational efficiency of SimpleClick, highlighting its suitability as a practical annotation tool." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 539, + 173, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 539, + 173, + 550 + ], + "spans": [ + { + "bbox": [ + 59, + 539, + 173, + 550 + ], + "type": "text", + "content": "Our main contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 552, + 287, + 646 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 552, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 552, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 552, + 287, + 574 + ], + "type": "text", + "content": "- We propose SimpleClick, the first plain-backbone method for interactive image segmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 575, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 575, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 575, + 287, + 609 + ], + "type": "text", + "content": "- SimpleClick achieves state-of-the-art performance on natural images and shows strong generalizability on medical images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 611, + 287, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 287, + 646 + ], + "type": "text", + "content": "- SimpleClick meets the computational efficiency requirement for a practical annotation tool, highlighting its readiness for real-world applications." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 657, + 134, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 134, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 134, + 670 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": "Interactive Image Segmentation Interactive image segmentation is a longstanding problem for which increasingly better solution approaches have been proposed. Early" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "works [6, 16, 18, 40] tackle this problem using graphs defined over image pixels. However, these methods only focus on low-level image features and therefore tend to have difficulty with complex objects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 123, + 546, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 123, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 123, + 546, + 460 + ], + "type": "text", + "content": "Thriving on large datasets, ConvNets [10, 30, 43, 47, 48] have evolved as the dominant architecture for high-quality interactive segmentation. ConvNet-based methods have explored various interaction types, such as bounding boxes [47], polygons [1], clicks [43], and scribbles [45]. Click-based approaches are the most common due to their simplicity and well-established training and evaluation protocols. Xu et al. [48] first proposed a click simulation strategy that has been adopted by follow-up work [10, 34, 43]. DEXTR [36] extracts a target object by specifying its four extreme points (left-most, right-most, top, and bottom pixels). FCA-Net [31] demonstrates the critical role of the first click for better segmentation. Recently, ViTs have been applied to interactive segmentation. FocalClick [10] uses SegFormer [46] as the backbone network and achieves state-of-the-art segmentation results with high computational efficiency. iSegFormer [33] uses a Swin Transformer [35] as the backbone network for interactive segmentation on medical images. Besides the contribution of backbones, some works are exploring elaborate refinement modules built upon the backbone. FocalClick [10] and FocusCut [30] propose similar local refinement modules for high-quality segmentation. PseudoClick [34] proposes a click-imitation mechanism by estimating the next click to further reduce human annotation cost. Our method differs from all previous click-based methods in its plain, non-hierarchical ViT backbone, enjoying the benefits from readily available pretrained ViT models (e.g. MAE [21])." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "content": "Vision Transformers for Non-Interactive Segmentation Recently, ViT-based approaches [17, 24, 44, 46, 50] have shown competitive performance on segmentation tasks compared to ConvNets. The original ViT [13] is a non-hierarchical architecture that only maintains single-scale feature maps throughout. SETR [52] and Segmenter [44] use the original ViT as the encoder for semantic segmentation. To allow for more efficient segmentation, the Swin Transformer [35] reintroduces a computational hierarchy into the original ViT architecture using shifted window attention, leading to a highly efficient hierarchical ViT backbone. SegFormer [46] designs hierarchical feature representations based on the original ViT using overlapped patch merging, combined with a light-weight MLP decoder for efficient segmentation. HRViT [17] integrates a high-resolution multi-branch architecture with ViTs to learn multi-scale representations. Recently, the original ViT has been reintroduced as a competitive backbone for semantic segmentation [8] and object detection [26], with the aid of MAE [21] pretraining and window attention. Inspired by this finding, we explore using a plain ViT as the backbone" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22291" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 200, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 200, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 200, + 85 + ], + "type": "text", + "content": "network for interactive segmentation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 99, + 104, + 113 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 99, + 104, + 113 + ], + "spans": [ + { + "bbox": [ + 47, + 99, + 104, + 113 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 121, + 289, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 289, + 218 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 289, + 218 + ], + "type": "text", + "content": "Our goal is not to propose new modules but to adapt a plain-ViT backbone for interactive segmentation with minimal modifications. Fig. 2 shows the overview of our method. Sec. 3.1 briefly describes the plain segmentation backbone. Sec. 3.2 shows how to adapt the backbone for interactive segmentation. Sec. 3.3 introduces other modules of SimpleClick. Sec. 3.4 describes the training and inference details of our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 229, + 211, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 229, + 211, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 229, + 211, + 243 + ], + "type": "text", + "content": "3.1. Plain Segmentation Backbone" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 249, + 289, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 249, + 289, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 289, + 417 + ], + "type": "text", + "content": "We use a plain ViT [13] as the segmentation backbone. Unlike previous hierarchical backbones, a plain ViT only maintains single-scale feature maps throughout. Given an image, the patch embedding layer of the plain ViT first divides the image into non-overlapping fixed-size patches (e.g. " + }, + { + "bbox": [ + 46, + 249, + 289, + 417 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 46, + 249, + 289, + 417 + ], + "type": "text", + "content": "). All patches are then flattened and linearly projected to a sequence of fixed-length tokens before feeding into self-attention blocks. These output tokens will be reshaped and upsampled to match the spatial size of the input image for segmentation. In this work, we consider three standard ViT backbones: ViT-B, ViT-L, and ViT-H (Tab. 1 shows the number of parameters of these backbones). For stable and efficient training, we use readily-available MAEs [21] as the pretrained weights for these backbones." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 429, + 279, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 279, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 279, + 442 + ], + "type": "text", + "content": "3.2. Clicks Encoding for Segmentation Backbone" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 449, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 593 + ], + "type": "text", + "content": "We adapt the above plain segmentation backbone for interactive segmentation by turning user interactions into a form of guidance learned by the network. Similar to previous click-based methods [34, 43], we use clicks as a form of user interaction and encode each click as a small disk on a two-channel masks [34, 43] (one channel for positive clicks while the other for negative ones; positive clicks should be put in the foreground, while the negative ones should be put in the background). We automatically simulate human clicks for efficient training and evaluation. The clicks simulation process is described in Sec.3.4. Note that for human evaluation, a human-in-the-loop will provide all the clicks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Given human clicks or simulated clicks, we now introduce how to fuse them into the plain backbone. We first introduce a patch-embedding layer that is symmetric to the one in the backbone. Note that we also concatenate the previous segmentation to the clicks map as an additional channel for better performance. The two symmetric embedding layers operate on the image and the clicks map, respectively. The inputs are patchified, flattened, and projected to two vector sequences of the same dimension, followed by element-wise addition before feeding into the self-attention blocks." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 310, + 71, + 542, + 116 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 542, + 116 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 542, + 116 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 542, + 116 + ], + "type": "table", + "html": "
Model↓ Module→ViT BackboneConv. NeckMLP Head
Ours-ViT-B (base)83.0 (89.3%)9.0 (9.7%)0.9 (1.0%)
Ours-ViT-L (large)290.8 (94.3%)16.5 (5.3%)1.1 (0.4%)
Ours-ViT-H (huge)604.0 (95.7%)25.8 (4.1%)1.3 (0.2%)
", + "image_path": "f704b14aca42558be0a85823be26f4a4261db16d593eaa95549d8e8b9d6307a7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 116, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 305, + 116, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 305, + 116, + 545, + 148 + ], + "type": "text", + "content": "Table 1. Number of parameters of our models. The unit is a million. The percentage of parameters is shown in bracket. Most parameters are used by the ViT backbone." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 169, + 400, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 169, + 400, + 181 + ], + "spans": [ + { + "bbox": [ + 306, + 169, + 400, + 181 + ], + "type": "text", + "content": "3.3. Other Modules" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 188, + 547, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 188, + 547, + 355 + ], + "spans": [ + { + "bbox": [ + 304, + 188, + 547, + 355 + ], + "type": "text", + "content": "Simple Feature Pyramid For the hierarchical backbone, a feature pyramid is commonly produced by an FPN [28] to combine features from different stages. For the plain backbone, a feature pyramid can be generated in a much simpler way: by a set of parallel convolutional or deconvolutional layers using only the last feature map of the backbone. As shown in Fig. 2, given the input ViT feature map, a multiscale feature map can be produced by four convolutions with different strides. Though the effectiveness of this simple feature pyramid design is first demonstrated in ViTDet [26] for object detection, we show in this work the effectiveness of this simple feature pyramid design for interactive segmentation. We also propose several additional variants (Fig. 6) as part of an ablation study (Sec. 4.4)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 356, + 547, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 547, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 547, + 596 + ], + "type": "text", + "content": "All-MLP Segmentation Head We implement a lightweight segmentation head using only MLP layers. It takes in the simple feature pyramid and produces a segmentation probability map of scale 1/4, followed by an upsampling operation to recover the original resolution. Note that this segmentation head avoids computationally demanding components and only accounts for up to " + }, + { + "bbox": [ + 304, + 356, + 547, + 596 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 356, + 547, + 596 + ], + "type": "text", + "content": " of the model parameters (Tab. 1). The key insight is that with a powerful pretrained backbone, a lightweight segmentation head is sufficient for interactive segmentation. The proposed all-MLP segmentation head works in three steps. First, each feature map from the simple feature pyramid goes through an MLP layer to transform it to an identical channel dimension (i.e. " + }, + { + "bbox": [ + 304, + 356, + 547, + 596 + ], + "type": "inline_equation", + "content": "C_2" + }, + { + "bbox": [ + 304, + 356, + 547, + 596 + ], + "type": "text", + "content": " in Fig. 2). Second, all feature maps are upsampled to the same resolution (i.e. 1/4 in Fig. 2) for concatenation. Third, the concatenated features are fused by another MLP layer to produce a single-channel feature map, followed by a sigmoid function to obtain a segmentation probability map, which is then transformed to a binary segmentation given a predefined threshold (i.e. 0.5)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 605, + 477, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 605, + 477, + 618 + ], + "spans": [ + { + "bbox": [ + 306, + 605, + 477, + 618 + ], + "type": "text", + "content": "3.4. Training and Inference Settings" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 624, + 547, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 624, + 547, + 685 + ], + "spans": [ + { + "bbox": [ + 305, + 624, + 547, + 685 + ], + "type": "text", + "content": "Backbone Pretraining Our backbone models are pretrained as MAEs [21] on ImageNet-1K [11]. In MAE pretraining, the ViT models reconstruct the randomly masked pixels of images while learning a universal representation. This simple self-supervised approach turns out to be an efficient and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "This probability map may be miscalibrated and can be improved by calibration approaches [12]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22292" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 76, + 478, + 270 + ], + "blocks": [ + { + "bbox": [ + 115, + 76, + 478, + 270 + ], + "lines": [ + { + "bbox": [ + 115, + 76, + 478, + 270 + ], + "spans": [ + { + "bbox": [ + 115, + 76, + 478, + 270 + ], + "type": "image", + "image_path": "8c33132e78719e53e82814db17c1001d00504536e4a869b37db8cd9742315476.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 273, + 547, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 273, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 547, + 329 + ], + "type": "text", + "content": "Figure 2. SimpleClick overview. Our method consists of three main modules: (a) a plain ViT backbone that maintains single-scale feature maps throughout; (b) a multi-scale simple feature pyramid that is generated from the last feature maps of the backbone by four parallel convolution or deconvolution layers; (c) a light-weight MLP decoder for segmentation. We also input the previous segmentation to improve the performance and to allow clicking from existing masks. User clicks are encoded as a two-channel disk map, combined with the previous segmentation as input. Boxes in blue are intermediate feature maps. The positional encoding is not shown for brevity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 347, + 288, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 288, + 384 + ], + "type": "text", + "content": "scalable way to pretrain ViT models [21]. In this work, we do not perform pretraining ourselves. Instead, we simply use the readily available pretrained MAE weights from [21]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 387, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 288, + 685 + ], + "type": "text", + "content": "Clicks Simulation and End-to-end Finetuning With the pretrained backbone, we finetune our model end-to-end on the interactive segmentation task. The finetuning pipeline can be briefly described as follows. First, we automatically simulate clicks based on the current segmentation and gold standard segmentation, without a human-in-the-loop providing the clicks. Specifically, we use a combination of random and iterative click simulation strategies, inspired by RITM [43]. The random click simulation strategy generates clicks in parallel, without considering the order of the clicks. The iterative click simulation strategy generates clicks iteratively, where the next click should be placed on the erroneous region of a prediction that was obtained using the previous clicks. This strategy is more similar to human clicking behavior. Second, we incorporate the segmentation from the previous interaction as an additional input for the backbone, further improving the segmentation quality. This also allows our method to refine from an existing segmentation, which is a desired feature for a practical annotation tool. We use the normalized focal loss [43] (NFL) to train all our models. Previous works [10, 43] show that NFL converges faster and achieves better performance than the widely used binary cross entropy loss for interactive segmentation tasks. Similar training pipelines have been proposed by RITM [43] and its follow-up works [9, 10, 34]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 712 + ], + "type": "text", + "content": "Human Evaluation and Automatic Evaluation There are two inference modes: automatic evaluation and human eval" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 347, + 547, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 547, + 431 + ], + "type": "text", + "content": "uation. For automatic evaluation, clicks are automatically simulated based on the current segmentation and gold standard. For human evaluation, a human-in-the-loop provides all clicks based on their subjective evaluation of current segmentation results. We use automatic evaluation for quantitative analyses and human evaluation for a qualitative assessment of the interactive segmentation behavior." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 442, + 388, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 442, + 388, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 442, + 388, + 455 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 462, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 497 + ], + "type": "text", + "content": "Datasets We conducted experiments on 10 public datasets including 7 natural image datasets and 3 medical datasets. The details are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 498, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 306, + 498, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 545, + 521 + ], + "type": "text", + "content": "- GrabCut [40]: 50 images (50 instances), each with clear foreground and background differences." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 522, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 522, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 306, + 522, + 545, + 545 + ], + "type": "text", + "content": "- Berkeley [37]: 96 images (100 instances); this dataset shares a small portion of images with GrabCut." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 546, + 544, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 546, + 544, + 570 + ], + "spans": [ + { + "bbox": [ + 306, + 546, + 544, + 570 + ], + "type": "text", + "content": "- DAVIS [39]: 50 videos; we only use the same 345 frames as used in [10, 30, 34, 43] for evaluation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 570, + 544, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 570, + 544, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 570, + 544, + 594 + ], + "type": "text", + "content": "- Pascal VOC [14]: 1449 images (3427 instances) in the validation set. We only test on the validation set." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 594, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 594, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 306, + 594, + 545, + 640 + ], + "type": "text", + "content": "- SBD [20]: 8498 training images (20172 instances) and 2857 validation images (6671 instances). Following previous works [10, 30, 43], we train our model on the training set and evaluate on the validation set." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 642, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 642, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 306, + 642, + 545, + 689 + ], + "type": "text", + "content": "- COCO [29]+LVIS [19] (C+L): COCO contains 118K training images (1.2M instances); LVIS shares the same images with COCO but has much higher segmentation quality. We combine the two datasets for training." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "- ssTEM [15]: two image stacks, each contains 20 medical images. We use the same stack that was used in [34]." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22293" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 79, + 285, + 236 + ], + "blocks": [ + { + "bbox": [ + 51, + 79, + 285, + 236 + ], + "lines": [ + { + "bbox": [ + 51, + 79, + 285, + 236 + ], + "spans": [ + { + "bbox": [ + 51, + 79, + 285, + 236 + ], + "type": "image", + "image_path": "95a272e72935b50e99c599d489175aead5a44503e4efee2e376f3def16a5fb21.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 238, + 547, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 238, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 547, + 270 + ], + "type": "text", + "content": "Figure 3. Convergence analysis for models trained on either SBD [20] or COCO [29] + LVIS [19] (C+L). We report results on SBD [20] and Pascal VOC [14]. The metric is mean IoU given " + }, + { + "bbox": [ + 46, + 238, + 547, + 270 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 238, + 547, + 270 + ], + "type": "text", + "content": " clicks (" + }, + { + "bbox": [ + 46, + 238, + 547, + 270 + ], + "type": "inline_equation", + "content": "\\text{mIoU} @ k" + }, + { + "bbox": [ + 46, + 238, + 547, + 270 + ], + "type": "text", + "content": "). Our models in general require fewer clicks for a given accuracy level." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 294, + 79, + 525, + 235 + ], + "blocks": [ + { + "bbox": [ + 294, + 79, + 525, + 235 + ], + "lines": [ + { + "bbox": [ + 294, + 79, + 525, + 235 + ], + "spans": [ + { + "bbox": [ + 294, + 79, + 525, + 235 + ], + "type": "image", + "image_path": "403668dc1826ba85844aa3b720a498bf1b1e2ccba3df7c6302fa3ec6c264c16e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 47, + 279, + 548, + 538 + ], + "blocks": [ + { + "bbox": [ + 47, + 279, + 548, + 538 + ], + "lines": [ + { + "bbox": [ + 47, + 279, + 548, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 279, + 548, + 538 + ], + "type": "table", + "html": "
MethodBackboneGrabCutBerkeleySBDDAVISPascal VOC
NoC85NoC90NoC85NoC90NoC85NoC90NoC85NoC90NoC85NoC90
\\( \\nmid \\)DIOS [48] \\( _{CVPR16} \\)FCN-6.04-8.65---12.586.88-
\\( \\nmid \\)FCA-Net [31] \\( _{CVPR20} \\)ResNet-101-2.08-3.92---7.572.69-
\\( \\nmid \\)LD [27] \\( _{CVPR18} \\)VGG-193.204.79--7.4110.785.059.57--
\\( \\nmid \\)BRS [23] \\( _{CVPR19} \\)DenseNet2.603.60-5.086.599.785.588.24--
\\( \\nmid \\)f-BRS [42] \\( _{CVPR20} \\)ResNet-1012.302.72-4.574.817.735.047.41--
\\( \\nmid \\)RITM [43] \\( _{Preprint21} \\)HRNet-181.762.041.873.223.395.434.946.712.513.03
\\( \\nmid \\)CDNet [9] \\( _{ICCV21} \\)ResNet-341.862.181.953.275.187.895.006.893.614.51
\\( \\nmid \\)PseudoClick [34] \\( _{ECCV22} \\)HRNet-181.682.041.853.233.385.404.816.572.342.74
\\( \\nmid \\)FocalClick [10] \\( _{CVPR22} \\)HRNet-18s1.862.06-3.144.306.524.926.48--
\\( \\nmid \\)FocalClick [10] \\( _{CVPR22} \\)SegF-B01.661.90-3.144.346.515.027.06--
\\( \\nmid \\)FocusCut [30] \\( _{CVPR22} \\)ResNet-501.601.78\\( 1.85^{\\dagger } \\)3.443.625.665.006.38--
\\( \\nmid \\)FocusCut [30] \\( _{CVPR22} \\)ResNet-1011.461.64\\( 1.81^{\\dagger } \\)3.013.405.314.856.22--
\\( \\nmid \\)OursViT-B1.401.541.442.463.285.244.105.482.382.81
\\( \\nmid \\)OursViT-L1.381.461.402.332.694.464.125.391.952.30
\\( \\nmid \\)OursViT-H1.321.441.362.092.514.154.205.341.882.20
\\( \\nmid \\)RITM [43] \\( _{Preprint21} \\)HRNet-321.461.561.432.103.595.714.115.342.192.57
\\( \\nmid \\)CDNet [9] \\( _{ICCV21} \\)ResNet-341.401.521.472.064.307.044.275.562.743.30
\\( \\nmid \\)PseudoClick [34] \\( _{ECCV22} \\)HRNet-321.361.501.402.083.465.543.795.111.942.25
\\( \\nmid \\)FocalClick [10] \\( _{CVPR22} \\)SegF-B01.401.661.592.274.566.864.045.492.973.52
\\( \\nmid \\)FocalClick [10] \\( _{CVPR22} \\)SegF-B31.441.501.551.923.535.593.614.902.462.88
\\( \\nmid \\)OursViT-B1.381.481.361.973.435.623.665.062.062.38
\\( \\nmid \\)OursViT-L1.321.401.341.892.954.893.264.811.721.96
\\( \\nmid \\)OursViT-H1.381.501.361.752.854.703.414.781.761.98
", + "image_path": "e33e5d653278fcbac5a597e499b7054ab0fa52f629380660cbee6302b4ac2494.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 539, + 547, + 583 + ], + "lines": [ + { + "bbox": [ + 46, + 539, + 547, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 547, + 583 + ], + "type": "text", + "content": "Table 2. Comparison with previous results. We report results on five benchmarks: GrabCut [40], Berkeley [37], SBD [20], DAVIS [39], and Pascal VOC [14]. The best results are set in bold. " + }, + { + "bbox": [ + 46, + 539, + 547, + 583 + ], + "type": "inline_equation", + "content": "\\nsubseteq" + }, + { + "bbox": [ + 46, + 539, + 547, + 583 + ], + "type": "text", + "content": " denotes a model trained on Pascal; " + }, + { + "bbox": [ + 46, + 539, + 547, + 583 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 539, + 547, + 583 + ], + "type": "text", + "content": " denotes a model trained on SBD; " + }, + { + "bbox": [ + 46, + 539, + 547, + 583 + ], + "type": "inline_equation", + "content": "\\sqcup" + }, + { + "bbox": [ + 46, + 539, + 547, + 583 + ], + "type": "text", + "content": " denotes a model trained on COCO [29] + LVIS [19] (C+L); " + }, + { + "bbox": [ + 46, + 539, + 547, + 583 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 539, + 547, + 583 + ], + "type": "text", + "content": " denotes a number reproduced by the released or retrained models. Our models achieve state-of-the-art performance on all benchmarks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 602, + 286, + 650 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 602, + 286, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 602, + 286, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 602, + 286, + 625 + ], + "type": "text", + "content": "- BraTS [4]: 369 magnetic resonance image (MRI) volumes; we test on the same 369 slices used in [34]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 627, + 286, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 286, + 650 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 286, + 650 + ], + "type": "text", + "content": "- OAIZIB [2]: 507 MRI volumes; we test on the same 150 slices (300 instances) as used in [33]." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "Evaluation Metrics Following previous works [30,42,43], we automatically simulate user clicks by comparing the current segmentation with the gold standard. In this simulation, the next click will be put at the center of the region with the largest error. We use the Number of Clicks (NoC)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "text", + "content": "as the evaluation metric to calculate the number of clicks required to achieve a target Intersection over Union (IoU). We set two target IoUs: " + }, + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "text", + "content": ", represented by " + }, + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{NoC}\\% 85" + }, + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{NoC}\\% 90" + }, + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "text", + "content": " respectively. The maximum number of clicks for each instance is set to 20. We also use the average IoU given " + }, + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "text", + "content": " clicks (" + }, + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{mIoU} @ k" + }, + { + "bbox": [ + 304, + 602, + 545, + 687 + ], + "type": "text", + "content": ") as an evaluation metric to measure the segmentation quality given a fixed number of clicks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details We implement our models using Python and PyTorch [38]. We implement three models" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22294" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 76, + 206, + 186 + ], + "blocks": [ + { + "bbox": [ + 53, + 76, + 206, + 186 + ], + "lines": [ + { + "bbox": [ + 53, + 76, + 206, + 186 + ], + "spans": [ + { + "bbox": [ + 53, + 76, + 206, + 186 + ], + "type": "image", + "image_path": "d708b42d1e3d20a3f9719d6baca1cb8d5099ca4dc24f23ccaf7dfb58edc3db65.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 222, + 76, + 379, + 186 + ], + "blocks": [ + { + "bbox": [ + 222, + 76, + 379, + 186 + ], + "lines": [ + { + "bbox": [ + 222, + 76, + 379, + 186 + ], + "spans": [ + { + "bbox": [ + 222, + 76, + 379, + 186 + ], + "type": "image", + "image_path": "f272ee5755fba79b1d467fdc717a5c1fe2a925081f2a7f26b38a04addfe55e92.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 391, + 76, + 549, + 186 + ], + "blocks": [ + { + "bbox": [ + 391, + 76, + 549, + 186 + ], + "lines": [ + { + "bbox": [ + 391, + 76, + 549, + 186 + ], + "spans": [ + { + "bbox": [ + 391, + 76, + 549, + 186 + ], + "type": "image", + "image_path": "7d956fb6852536a52d07d67e800ed4ac655d3a1aa9baa93f30097890a0c1c673.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 47, + 190, + 207, + 300 + ], + "blocks": [ + { + "bbox": [ + 47, + 190, + 207, + 300 + ], + "lines": [ + { + "bbox": [ + 47, + 190, + 207, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 207, + 300 + ], + "type": "image", + "image_path": "f0f6086146e3d93c579d3ae0f7cfa6ced33039a26070565b6a24aabcae8000dd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 302, + 545, + 324 + ], + "lines": [ + { + "bbox": [ + 46, + 302, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 545, + 324 + ], + "type": "text", + "content": "Figure 4. Histogram analysis of IoU given a predefined number of clicks " + }, + { + "bbox": [ + 46, + 302, + 545, + 324 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 302, + 545, + 324 + ], + "type": "text", + "content": " (IoU@k). We report analysis on SBD [20] and Pascal VOC [14] with models trained on SBD. Compared with the two baselines, our models achieve higher-quality segmentation with fewer failure cases." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 223, + 190, + 379, + 300 + ], + "blocks": [ + { + "bbox": [ + 223, + 190, + 379, + 300 + ], + "lines": [ + { + "bbox": [ + 223, + 190, + 379, + 300 + ], + "spans": [ + { + "bbox": [ + 223, + 190, + 379, + 300 + ], + "type": "image", + "image_path": "82c52d1ea5aee0d2d8ef34005a7dac3ec029c5fdc6eeae2081e2e1c661e6f726.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 395, + 190, + 550, + 300 + ], + "blocks": [ + { + "bbox": [ + 395, + 190, + 550, + 300 + ], + "lines": [ + { + "bbox": [ + 395, + 190, + 550, + 300 + ], + "spans": [ + { + "bbox": [ + 395, + 190, + 550, + 300 + ], + "type": "image", + "image_path": "93f392859a70a3b67630aa0a53ab74e3522b235574105659a481aab2b9682f6b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 52, + 343, + 280, + 498 + ], + "blocks": [ + { + "bbox": [ + 52, + 343, + 280, + 498 + ], + "lines": [ + { + "bbox": [ + 52, + 343, + 280, + 498 + ], + "spans": [ + { + "bbox": [ + 52, + 343, + 280, + 498 + ], + "type": "image", + "image_path": "156383fcea46ded5d4d404bcd1c98849c93f8ba081855f79ad8bcc76c295ffe0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 502, + 547, + 536 + ], + "lines": [ + { + "bbox": [ + 45, + 502, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 45, + 502, + 547, + 536 + ], + "type": "text", + "content": "Figure 5. Convergence analysis for two medical image datasets: BraTS [4] and OAIZIB [2]. Models are trained on either SBD [20] or COCO [29] + LVIS [19] (denoted as " + }, + { + "bbox": [ + 45, + 502, + 547, + 536 + ], + "type": "inline_equation", + "content": "\\mathrm{C} + \\mathrm{L}" + }, + { + "bbox": [ + 45, + 502, + 547, + 536 + ], + "type": "text", + "content": "). The metric is mean IoU given " + }, + { + "bbox": [ + 45, + 502, + 547, + 536 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 45, + 502, + 547, + 536 + ], + "type": "text", + "content": " clicks. Overall, our models require fewer clicks for a given accuracy level. The performance gain is more prominent for bigger models (e.g. ViT-H) or larger training sets (e.g. C+L)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 294, + 343, + 523, + 498 + ], + "blocks": [ + { + "bbox": [ + 294, + 343, + 523, + 498 + ], + "lines": [ + { + "bbox": [ + 294, + 343, + 523, + 498 + ], + "spans": [ + { + "bbox": [ + 294, + 343, + 523, + 498 + ], + "type": "image", + "image_path": "b6940d6819f52f98d051c1c61d91f6ed4de83aef58d207e70fdc409e394d0563.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 555, + 289, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 289, + 712 + ], + "type": "text", + "content": "based on three vanilla ViT models (i.e. ViT-B, ViT-L, and ViT-H). These backbone models are initialized with the MAE pretrained weights, and then are finetuned end-to-end with other modules. We train our models on either SBD or COCO+LVIS with 55 epochs; the initial learning rate is set to " + }, + { + "bbox": [ + 46, + 555, + 289, + 712 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 555, + 289, + 712 + ], + "type": "text", + "content": " and decreases to " + }, + { + "bbox": [ + 46, + 555, + 289, + 712 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-6}" + }, + { + "bbox": [ + 46, + 555, + 289, + 712 + ], + "type": "text", + "content": " after epoch 50. We set the batch size to 140 for ViT-Base, 72 for ViT-Large, and 32 for ViT-Huge to fit the models into GPU memory. All our models are trained on four NVIDIA RTX A6000 GPUs. We use the following data augmentation techniques: random resizing (scale range from 0.75 to 1.25), random flipping and rotation, random brightness contrast, and random cropping. Though the ViT backbone was pretrained on images of size" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 555, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 592 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 555, + 545, + 592 + ], + "type": "text", + "content": ", we finetune on " + }, + { + "bbox": [ + 304, + 555, + 545, + 592 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 304, + 555, + 545, + 592 + ], + "type": "text", + "content": " with non-shifting window attention for better performance. We optimize using Adam with " + }, + { + "bbox": [ + 304, + 555, + 545, + 592 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 304, + 555, + 545, + 592 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 555, + 545, + 592 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 304, + 555, + 545, + 592 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 599, + 491, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 599, + 491, + 612 + ], + "spans": [ + { + "bbox": [ + 305, + 599, + 491, + 612 + ], + "type": "text", + "content": "4.1. Comparison with Previous Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "We show in Tab. 2 the comparisons with previous state-of-the-art results. Our models achieve the best performance on all the five benchmarks. Remarkably, when trained on SBD training set, our ViT-H model achieves 4.15 NoC@90 on the SBD validation set, outperforming the previous best score by " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "21.8\\%" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": ". Since the SBD validation set contains the largest number of instances (6671 instances) among the five benchmarks, this improvement is convincing. When trained" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22295" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 288, + 186 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 186 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 186 + ], + "type": "table", + "html": "
ModelssTEM\nmIoU@10BraTS\nmIoU@10 / 20OAIZIB\nmIoU@10 / 20
↓RITM-H18 [43]93.1587.05 / 90.4771.04 / 78.52
↓CDN-RN34 [9]66.7258.34 / 82.0738.07 / 61.17
↓RITM-H32 [43]94.1188.34 / 89.2575.27 / 75.18
↓CDN-RN34 [9]88.4680.24 / 86.6363.19 / 74.21
↓FC-SF-B0 [10]92.6286.02 / 90.7474.08 / 79.14
↓FC-SF-B3 [10]93.6188.62 / 90.5875.77 / 80.08
↓Ours-ViT-B93.7286.98 / 90.6776.05 / 79.61
↓Ours-ViT-L94.3488.43 / 90.8477.34 / 79.97
↓Ours-ViT-H94.0888.98 / 91.0077.50 / 80.10
", + "image_path": "98b09e19ef34e827c4e6a4417e80a123a6645cf5451a1977db5dca3696c1f170.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 187, + 287, + 220 + ], + "lines": [ + { + "bbox": [ + 47, + 187, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 47, + 187, + 287, + 220 + ], + "type": "text", + "content": "Table 3. Out-of-domain evaluation on three medical image datasets: ssTEM [15], BraTS [4], and OAIZIB [2]. Our models generalize very well on the three datasets, without finetuning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 239, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 287, + 384 + ], + "type": "text", + "content": "on COCO+LVIS, our models also achieve state-of-the-art performance on all benchmarks. Fig. 7 shows several segmentation cases on DAVIS, including the worst case. Note that the DAVIS dataset requires high-quality segmentations because all its instances have a high-quality gold standard. Our models still achieve the state-of-the-art on DAVIS without using specific modules, such as a local refinement module [10], which is beneficial for high-quality segmentation. Fig. 3 shows that our method converges better than other methods with sufficient clicks, leading to fewer failure cases as shown in Fig. 4. We only report results on SBD and Pascal VOC, the top two largest datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 392, + 287, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 287, + 406 + ], + "type": "text", + "content": "4.2. Out-of-Domain Evaluation on Medical Images" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 411, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 287, + 506 + ], + "type": "text", + "content": "We further evaluate the generalizability of our models on three medical image datasets: ssTEM [15], BraTS [3], and OAIZIB [2]. Tab. 3 reports the evaluation results on these three datasets. Fig. 5 shows the convergence analysis on BraTS and OAIZIB. Overall, our models generalize well to medical images. We also find that the models trained on larger datasets (i.e. " + }, + { + "bbox": [ + 47, + 411, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{C + L}" + }, + { + "bbox": [ + 47, + 411, + 287, + 506 + ], + "type": "text", + "content": ") generalize better than the models trained on smaller datasets (i.e. SBD)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 514, + 234, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 234, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 234, + 528 + ], + "type": "text", + "content": "4.3. Towards Practical Annotation Tool" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 533, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 287, + 665 + ], + "type": "text", + "content": "Tiny Backbone To allow for practical applications, especially on low-end devices with limited computational resources, we implement an extremely tiny backbone (i.e. ViT-xTiny) for SimpleClick. Compared with ViT-Base, ViT-xTiny decreases the embedding dimension from 768 to 160 and the number of attention blocks from 12 to 8. We end up with a SimpleClick-xTiny model, which is comparable with the tiny FocalClick models in terms of parameters. Comparison results in Tab. 4 show that our model outperforms FocalClick models, even though it is trained from scratch due to the lack of readily available pretrained weights." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": "Computational Analysis Tab. 5 shows a comparison of computational requirements with respect to model parameters, FLOPs, GPU memory consumption, and speed; the speed is measured by seconds per click (SPC). Fig. 1 shows" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 310, + 71, + 542, + 115 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 542, + 115 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 542, + 115 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 542, + 115 + ], + "type": "table", + "html": "
ModelBackbonePretrainedParams/MNoC85NoC90
FocalClickHRNet-18s-S1✓4.224.747.29
FocalClickSegFormer-B0-S1✓3.724.987.60
SimpleClickViT-xTiny✗3.724.717.09
", + "image_path": "c21f115adfcb14cb00059f1d17fa99b45b5ffc6cd78adc577f71a140d1d2d7c3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 116, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 306, + 116, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 306, + 116, + 545, + 148 + ], + "type": "text", + "content": "Table 4. Comparison results on SBD for tiny models. All models are trained on C+L with 230 epochs. Our SimpleClick-xTiny model outperforms FocalClick models without pretraining." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 159, + 545, + 337 + ], + "blocks": [ + { + "bbox": [ + 307, + 159, + 545, + 337 + ], + "lines": [ + { + "bbox": [ + 307, + 159, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 159, + 545, + 337 + ], + "type": "table", + "html": "
BackboneParams/MFLOPs/GMem/G↓SPC/ms
HR-18s 400 [43]4.2217.940.5054
HR-18 400 [43]10.0330.990.5256
HR-32 400 [43]30.9583.121.1286
Swin-B 400 [33]87.44138.211.4136
Swin-L 400 [33]195.90302.782.1444
SegF-B0 256 [10]3.723.420.1037
SegF-B3 256 [10]45.6624.750.3253
ResN-34 384 [9]23.47113.600.2534
ResN-50 384 [30]40.3678.820.85331
ResN-101 384 [30]59.35100.760.89355
Ours-ViT-xT 2243.722.630.1717
Ours-ViT-xT 4483.7210.520.2329
Ours-ViT-B 22496.4642.440.5134
Ours-ViT-B 44896.46169.780.8754
Ours-ViT-L 448322.18532.871.7286
Ours-ViT-H 448659.391401.933.22132
", + "image_path": "3f3dde7d97a3c8d156d56a57cf88d42de1c186af5be9667e4861222df8b9c70b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 338, + 545, + 425 + ], + "lines": [ + { + "bbox": [ + 306, + 338, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 545, + 425 + ], + "type": "text", + "content": "Table 5. Computation comparison for model parameters, FLOPs, GPU memory consumption (measured by the maximum GPU memory managed by PyTorch's caching allocator), and speed (measured by seconds per click). Each method is denoted by its backbone. The small number in front of the model denotes the input size " + }, + { + "bbox": [ + 306, + 338, + 545, + 425 + ], + "type": "inline_equation", + "content": "(448\\times 448" + }, + { + "bbox": [ + 306, + 338, + 545, + 425 + ], + "type": "text", + "content": " for our models by default). Even for our ViT-H model, the speed (132ms) and memory consumption (3.22G) are sufficient to meet the requirements of a practical annotation tool." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 445, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 601 + ], + "type": "text", + "content": "the interactive segmentation performance of methods in terms of FLOPs. In Fig. 1 and Tab. 5, each method is denoted by its backbone. For fair comparison, we evaluate all the methods on the same benchmark (i.e. GrabCut) and using the same computer (GPU: NVIDIA RTX A6000, CPU: Intel Silver×2). We only calculate the FLOPs in a single forward pass. For methods like FocusCut which require multiple forward passes for each click, the FLOPs may be much higher than reported. By default, our method takes images of size " + }, + { + "bbox": [ + 304, + 445, + 545, + 601 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 304, + 445, + 545, + 601 + ], + "type": "text", + "content": " as the fixed input. Even for our ViT-H model, the speed (132ms) and memory consumption (3.22G) is sufficient to meet the requirements of a practical annotation tool." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 610, + 400, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 400, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 400, + 624 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "In this section, we ablate the backbone finetuning and feature pyramid design. Tab. 6 shows the ablation results. By default, we finetune the backbone along with other modules. As an ablation, we freeze the backbone during finetuning, leading to significantly worse performance. This ablation is explainable considering the ViT backbone takes most of the model parameters (Tab. 1). For the second ablation," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22296" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 71, + 182, + 125 + ], + "blocks": [ + { + "bbox": [ + 75, + 71, + 182, + 125 + ], + "lines": [ + { + "bbox": [ + 75, + 71, + 182, + 125 + ], + "spans": [ + { + "bbox": [ + 75, + 71, + 182, + 125 + ], + "type": "image", + "image_path": "35802b1d10ba2be9ff5c3882c8a36bb273d9da7e497eb7cc5fbe7f5d9e7c158d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 128, + 165, + 137 + ], + "lines": [ + { + "bbox": [ + 89, + 128, + 165, + 137 + ], + "spans": [ + { + "bbox": [ + 89, + 128, + 165, + 137 + ], + "type": "text", + "content": "(a) simple feature pyramid" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 187, + 72, + 293, + 124 + ], + "blocks": [ + { + "bbox": [ + 187, + 72, + 293, + 124 + ], + "lines": [ + { + "bbox": [ + 187, + 72, + 293, + 124 + ], + "spans": [ + { + "bbox": [ + 187, + 72, + 293, + 124 + ], + "type": "image", + "image_path": "16cf33bbdfb30df78c8ac692f33c5127ac9495c2039b41c38ad9d994b593755d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 207, + 128, + 273, + 137 + ], + "lines": [ + { + "bbox": [ + 207, + 128, + 273, + 137 + ], + "spans": [ + { + "bbox": [ + 207, + 128, + 273, + 137 + ], + "type": "text", + "content": "(b) single-scale variant" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 298, + 72, + 405, + 124 + ], + "blocks": [ + { + "bbox": [ + 298, + 72, + 405, + 124 + ], + "lines": [ + { + "bbox": [ + 298, + 72, + 405, + 124 + ], + "spans": [ + { + "bbox": [ + 298, + 72, + 405, + 124 + ], + "type": "image", + "image_path": "4522563d1266cff3666743c2bd79b0c55bd0132df7861db3d02715fa953d6ed5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 129, + 376, + 137 + ], + "lines": [ + { + "bbox": [ + 324, + 129, + 376, + 137 + ], + "spans": [ + { + "bbox": [ + 324, + 129, + 376, + 137 + ], + "type": "text", + "content": "(c) parallel variant" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 410, + 72, + 518, + 124 + ], + "blocks": [ + { + "bbox": [ + 410, + 72, + 518, + 124 + ], + "lines": [ + { + "bbox": [ + 410, + 72, + 518, + 124 + ], + "spans": [ + { + "bbox": [ + 410, + 72, + 518, + 124 + ], + "type": "image", + "image_path": "44b772ea34f42aabdba4d854168e21e494dc508808f4b8567fbb51b74c2ea160.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 129, + 488, + 137 + ], + "lines": [ + { + "bbox": [ + 438, + 129, + 488, + 137 + ], + "spans": [ + { + "bbox": [ + 438, + 129, + 488, + 137 + ], + "type": "text", + "content": "(d) partial variant" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 51, + 185, + 191, + 371 + ], + "blocks": [ + { + "bbox": [ + 46, + 140, + 547, + 172 + ], + "lines": [ + { + "bbox": [ + 46, + 140, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 46, + 140, + 547, + 172 + ], + "type": "text", + "content": "Figure 6. Simple feature pyramid and its variants: b) the single-scale variant ablates the multi-scale property; c) the parallel variant evenly extracts features from the backbone; d) the partial variant freezes the second half parameters of the backbone. Tab. 6 shows the comparison results." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 185, + 191, + 371 + ], + "lines": [ + { + "bbox": [ + 51, + 185, + 191, + 371 + ], + "spans": [ + { + "bbox": [ + 51, + 185, + 191, + 371 + ], + "type": "image", + "image_path": "1dc18ac633ff811b441fc0705291c390879570e65fc8e1014049daeae3c9388c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 373, + 150, + 380 + ], + "lines": [ + { + "bbox": [ + 93, + 373, + 150, + 380 + ], + "spans": [ + { + "bbox": [ + 93, + 373, + 150, + 380 + ], + "type": "text", + "content": "(a) the worst case" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 384, + 546, + 418 + ], + "lines": [ + { + "bbox": [ + 46, + 384, + 546, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 546, + 418 + ], + "type": "text", + "content": "Figure 7. Segmentation results on DAVIS [39]: (a) the worst case; (b) a challenging case; (c) five normal cases. The backbone model is ViT-L trained on COCO [29] + LVIS [19]. The segmentation probability maps are shown in blue; the segmentation maps are overlaid in red on the original images. The clicks are shown as green (positive click) or red (negative click) dots on the image. GT denotes ground truth." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 192, + 185, + 332, + 371 + ], + "blocks": [ + { + "bbox": [ + 192, + 185, + 332, + 371 + ], + "lines": [ + { + "bbox": [ + 192, + 185, + 332, + 371 + ], + "spans": [ + { + "bbox": [ + 192, + 185, + 332, + 371 + ], + "type": "image", + "image_path": "8bfdd83a2fd3634da377d7b52ef40491c63f15a354464fd503e24104a0295645.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 372, + 299, + 381 + ], + "lines": [ + { + "bbox": [ + 227, + 372, + 299, + 381 + ], + "spans": [ + { + "bbox": [ + 227, + 372, + 299, + 381 + ], + "type": "text", + "content": "(b) a challenging case" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 332, + 186, + 403, + 371 + ], + "blocks": [ + { + "bbox": [ + 332, + 186, + 403, + 371 + ], + "lines": [ + { + "bbox": [ + 332, + 186, + 403, + 371 + ], + "spans": [ + { + "bbox": [ + 332, + 186, + 403, + 371 + ], + "type": "image", + "image_path": "3ab6c7acf59ed0a008a8399380f51bd7d7d1b4e27840ece7b26b25125883198d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 403, + 186, + 473, + 371 + ], + "blocks": [ + { + "bbox": [ + 403, + 186, + 473, + 371 + ], + "lines": [ + { + "bbox": [ + 403, + 186, + 473, + 371 + ], + "spans": [ + { + "bbox": [ + 403, + 186, + 473, + 371 + ], + "type": "image", + "image_path": "bcebc9d9442dfdb80f687f7228ba9aba5b368ee468147585e8350ffb35badee1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 372, + 472, + 380 + ], + "lines": [ + { + "bbox": [ + 405, + 372, + 472, + 380 + ], + "spans": [ + { + "bbox": [ + 405, + 372, + 472, + 380 + ], + "type": "text", + "content": "(c) five normal cases" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 476, + 186, + 543, + 371 + ], + "blocks": [ + { + "bbox": [ + 476, + 186, + 543, + 371 + ], + "lines": [ + { + "bbox": [ + 476, + 186, + 543, + 371 + ], + "spans": [ + { + "bbox": [ + 476, + 186, + 543, + 371 + ], + "type": "image", + "image_path": "7b102bbd0da3fe420b8a57ea26bf5f3c3b44beaa5381e4943b6822c88a8e80b5.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 51, + 435, + 284, + 514 + ], + "blocks": [ + { + "bbox": [ + 51, + 435, + 284, + 514 + ], + "lines": [ + { + "bbox": [ + 51, + 435, + 284, + 514 + ], + "spans": [ + { + "bbox": [ + 51, + 435, + 284, + 514 + ], + "type": "table", + "html": "
FP designfrozen ViTViT-BViT-L
SBDPascalSBDPascal
(a) simple FP✓11.486.939.755.59
(a) simple FP✗5.242.534.462.15
(b) single-scale✗6.562.805.532.48
(c) parallel✗7.213.096.262.79
(d) partial✗8.294.347.514.25
", + "image_path": "96a72af9dd43015beb554220c65c3fc6d3c71a080d7269a019d0124376b0925b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 514, + 287, + 590 + ], + "lines": [ + { + "bbox": [ + 46, + 514, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 287, + 590 + ], + "type": "text", + "content": "Table 6. Ablation study on backbone finetuning and feature pyramid (FP) design (Fig. 6). The metric is NoC@90. We have three findings in this ablation: 1) freezing the ViT backbone during finetuning significantly deteriorates the performance; 2) the multiscale property matters for the simple feature pyramid; 3) the last feature map from the backbone is sufficient to build an effective feature pyramid." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": "we compare the default simple feature pyramid design with three variants depicted in Fig. 6 (i.e. (b), (c), and (d)). First, we observe that the multi-scale representation matters for the feature pyramid. By ablating the multi-scale property in the simple feature pyramid, the performance drops considerably. We also notice that the last feature map from the backbone is strong enough to build the feature pyramid. The parallel feature pyramid generated by multi-stage feature maps" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 437, + 544, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 437, + 544, + 461 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 544, + 461 + ], + "type": "text", + "content": "from the backbone does not surpass the simple feature pyramid that only uses the last feature map of the backbone." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 483, + 452, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 452, + 495 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 452, + 495 + ], + "type": "text", + "content": "5. Limitations and Remarks" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "type": "text", + "content": "Our best-performing model (ViT-H) is much larger than existing models, leading to concerns about an unfair comparison. Our method is not prompt-efficient as every click update requires recomputing image features. The recent advancements in interactive segmentation [25, 53] show an elegant solution to this issue. Besides, these methods use sparse vectors to represent clicks, which might be more efficient than dense masks. Other than these, our models may fail in some challenging scenarios such as objects with very thin and elongated shapes or cluttered occlusions((a) and (b) in Fig. 7). We leave the improvements for future work." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "content": "We are entering an era of large-scale pretraining on multimodal foundation models, which is dramatically transforming the landscape of vision and language tasks. In this context, we hope SimpleClick will serve as a strong baseline for a new wave of high-performing interactive segmentation methods based on ViTs and large-scale pretraining." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22297" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 125, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 125, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 125, + 83 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 200 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 200 + ], + "type": "text", + "content": "We proposed SimpleClick, a plain-backbone model for interactive image segmentation. Our method leveraged a general-purpose ViT backbone that can benefit from advancements in pretrained ViT models. With the readily-available MAE weights, SimpleClick achieved state-of-the-art performance on natural images and demonstrated strong generalizability to medical images. Our method is simple yet effective, highlighting its suitability as a strong baseline model and a practical annotation tool." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 210, + 149, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 210, + 149, + 224 + ], + "spans": [ + { + "bbox": [ + 47, + 210, + 149, + 224 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 230, + 287, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 230, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 287, + 302 + ], + "type": "text", + "content": "We thank our reviewers and ACs for their invaluable comments and suggestions. Research reported in this publication was supported by the National Institutes of Health (NIH) under award number NIH 1R01AR072013. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 312, + 107, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 312, + 107, + 325 + ], + "spans": [ + { + "bbox": [ + 48, + 312, + 107, + 325 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 332, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 332, + 287, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 332, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 53, + 332, + 287, + 366 + ], + "type": "text", + "content": "[1] David Acuna, Huan Ling, Amlan Kar, and Sanja Fidler. Efficient interactive annotation of segmentation datasets with polygon-rnn++. In CVPR, pages 859-868, 2018. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 367, + 288, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 367, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 53, + 367, + 288, + 422 + ], + "type": "text", + "content": "[2] Felix Ambellan, Alexander Tack, Moritz Ehlke, and Stefan Zachow. Automated segmentation of knee bone and cartilage combining statistical shape knowledge and convolutional neural networks: Data from the osteoarthritis initiative. Medical image analysis, 52:109-118, 2019. 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 422, + 289, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 422, + 289, + 466 + ], + "spans": [ + { + "bbox": [ + 53, + 422, + 289, + 466 + ], + "type": "text", + "content": "[3] Xue Bai and Guillermo Sapiro. A geodesic framework for fast interactive image and video segmentation and matting. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 468, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 287, + 533 + ], + "type": "text", + "content": "[4] Ujjwal Baid, Satyam Ghodasara, Suyash Mohan, Michel Bilello, Evan Calabrese, Errol Colak, Keyvan Farahani, Jayashree Kalpathy-Cramer, Felipe C Kitamura, Sarthak Pati, et al. The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314, 2021. 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 534, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 534, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 53, + 534, + 287, + 567 + ], + "type": "text", + "content": "[5] Gedas Bertasius and Lorenzo Torresani. Classifying, segmenting, and tracking object instances in video with mask propagation. In CVPR, pages 9739-9748, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 568, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 568, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 53, + 568, + 287, + 622 + ], + "type": "text", + "content": "[6] Yuri Y Boykov and M-P Jolly. Interactive graph cuts for optimal boundary & region segmentation of objects in ND images. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, volume 1, pages 105-112. IEEE, 2001. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 53, + 624, + 287, + 678 + ], + "type": "text", + "content": "[7] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 679, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 679, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 679, + 287, + 714 + ], + "type": "text", + "content": "[8] Wuyang Chen, Xianzhi Du, Fan Yang, Lucas Beyer, Xiaohua Zhai, Tsung-Yi Lin, Huizhong Chen, Jing Li, Xiaodan Song, Zhangyang Wang, et al. A simple single-scale vision" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "transformer for object localization and instance segmentation. arXiv preprint arXiv:2112.09747, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 97, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 97, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 311, + 97, + 545, + 129 + ], + "type": "text", + "content": "[9] Xi Chen, Zhiyan Zhao, Feiwu Yu, Yilei Zhang, and Manni Duan. Conditional diffusion for interactive segmentation. In ICCV, pages 7345-7354, 2021. 4, 5, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 131, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 131, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 131, + 545, + 175 + ], + "type": "text", + "content": "[10] Xi Chen, Zhiyan Zhao, Yilei Zhang, Manni Duan, Donglian Qi, and Hengshuang Zhao. FocalClick: Towards practical interactive image segmentation. In CVPR, pages 1300-1309, 2022. 1, 2, 4, 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 177, + 545, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 177, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 307, + 177, + 545, + 221 + ], + "type": "text", + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 223, + 545, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 223, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 308, + 223, + 545, + 256 + ], + "type": "text", + "content": "[12] Zhipeng Ding, Xu Han, Peirong Liu, and Marc Niethammer. Local temperature scaling for probability calibration. In ICCV, pages 6889-6899, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 258, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 258, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 308, + 258, + 545, + 323 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 325, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 325, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 308, + 325, + 545, + 369 + ], + "type": "text", + "content": "[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 4, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 371, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 371, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 545, + 404 + ], + "type": "text", + "content": "[15] Stephan Gerhard, Jan Funke, Julien Martel, Albert Cardona, and Richard Fetter. Segmented anisotropic sstem dataset of neural tissue. figshare, pages 0–0, 2013. 4, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 406, + 545, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 406, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 308, + 406, + 545, + 438 + ], + "type": "text", + "content": "[16] Leo Grady. Random walks for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 28(11):1768-1783, 2006. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 440, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 440, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 308, + 440, + 545, + 495 + ], + "type": "text", + "content": "[17] Jiaqi Gu, Hyoukjun Kwon, Dilin Wang, Wei Ye, Meng Li, Yu-Hsin Chen, Liangzhen Lai, Vikas Chandra, and David Z Pan. Multi-scale high-resolution vision transformer for semantic segmentation. In CVPR, pages 12094-12103, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 497, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 497, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 308, + 497, + 545, + 552 + ], + "type": "text", + "content": "[18] Varun Gulshan, Carsten Rother, Antonio Criminisi, Andrew Blake, and Andrew Zisserman. Geodesic star convexity for interactive image segmentation. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3129-3136. IEEE, 2010. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 554, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 554, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 554, + 545, + 586 + ], + "type": "text", + "content": "[19] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, pages 5356-5364, 2019. 4, 5, 6, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 588, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 588, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 588, + 545, + 633 + ], + "type": "text", + "content": "[20] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In 2011 international conference on computer vision, pages 991-998. IEEE, 2011. 1, 4, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 634, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 634, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 308, + 634, + 545, + 677 + ], + "type": "text", + "content": "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 2, 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 679, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 679, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 679, + 545, + 714 + ], + "type": "text", + "content": "[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22298" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[23] Won-Dong Jang and Chang-Su Kim. Interactive image segmentation via backpropagating refinement scheme. In CVPR, pages 5297-5306, 2019. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "type": "text", + "content": "[24] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM computing surveys (CSUR), 54(10s):1-41, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 287, + 195 + ], + "type": "text", + "content": "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv:2304.02643, 2023. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 287, + 230 + ], + "type": "text", + "content": "[26] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. arXiv preprint arXiv:2203.16527, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 232, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 232, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 232, + 287, + 262 + ], + "type": "text", + "content": "[27] Zhuwen Li, Qifeng Chen, and Vladlen Koltun. Interactive image segmentation with latent diversity. In CVPR, pages 577-585, 2018. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 287, + 308 + ], + "type": "text", + "content": "[28] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 309, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 309, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 49, + 309, + 287, + 354 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 4, 5, 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 355, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 355, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 49, + 355, + 287, + 398 + ], + "type": "text", + "content": "[30] Zheng Lin, Zheng-Peng Duan, Zhao Zhang, Chun-Le Guo, and Ming-Ming Cheng. FocusCut: Diving into a focus view in interactive segmentation. In CVPR, pages 2637-2646, 2022. 1, 2, 4, 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 399, + 287, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 399, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 49, + 399, + 287, + 433 + ], + "type": "text", + "content": "[31] Zheng Lin, Zhao Zhang, Lin-Zhuo Chen, Ming-Ming Cheng, and Shao-Ping Lu. Interactive image segmentation with first click attention. In CVPR, pages 13339-13348, 2020. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 434, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 434, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 49, + 434, + 287, + 498 + ], + "type": "text", + "content": "[32] Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Arnaud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen Awm Van Der Laak, Bram Van Ginneken, and Clara I Sánchez. A survey on deep learning in medical image analysis. Medical image analysis, 42:60-88, 2017. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 500, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 500, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 500, + 287, + 555 + ], + "type": "text", + "content": "[33] Qin Liu, Zhenlin Xu, Yining Jiao, and Marc Niethammer. iSegFormer: Interactive segmentation via transformers with application to 3d knee mr images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 464-474. Springer, 2022. 1, 2, 5, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 556, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 556, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 49, + 556, + 287, + 609 + ], + "type": "text", + "content": "[34] Qin Liu, Meng Zheng, Benjamin Planche, Srikrishna Karanam, Terrence Chen, Marc Niethammer, and Ziyan Wu. PseudoClick: Interactive image segmentation with click imitation. arXiv preprint arXiv:2207.05282, 2022. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 612, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 612, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 49, + 612, + 287, + 656 + ], + "type": "text", + "content": "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 657, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 657, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 49, + 657, + 287, + 690 + ], + "type": "text", + "content": "[36] Kevis-Kokitsi Maninis, Sergi Caelles, Jordi Pont-Tuset, and Luc Van Gool. Deep extreme cut: From extreme points to object segmentation. In CVPR, pages 616-625, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 691, + 287, + 713 + ], + "type": "text", + "content": "[37] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "type": "text", + "content": "its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, volume 2, pages 416-423. IEEE, 2001. 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 119, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 119, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 308, + 119, + 545, + 184 + ], + "type": "text", + "content": "[38] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 186, + 545, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 231 + ], + "type": "text", + "content": "[39] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. 4, 5, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 232, + 545, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 276 + ], + "type": "text", + "content": "[40] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. \"grabcut\" interactive foreground extraction using iterated graph cuts. ACM transactions on graphics (TOG), 23(3):309-314, 2004. 2, 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 277, + 545, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 277, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 308, + 277, + 545, + 312 + ], + "type": "text", + "content": "[41] Dinggang Shen, Guorong Wu, and Heung-II Suk. Deep learning in medical image analysis. Annual review of biomedical engineering, 19:221, 2017. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 313, + 545, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 313, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 308, + 313, + 545, + 356 + ], + "type": "text", + "content": "[42] Konstantin Sofiuk, Ilia Petrov, Olga Barinova, and Anton Konushin. f-brs: Rethinking backpropagating refinement for interactive segmentation. In CVPR, pages 8623-8632, 2020. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 358, + 545, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 358, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 308, + 358, + 545, + 402 + ], + "type": "text", + "content": "[43] Konstantin Sofiuk, Ilia A Petrov, and Anton Konushin. Reviving iterative training with mask guidance for interactive segmentation. arXiv preprint arXiv:2102.06583, 2021. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 405, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 545, + 437 + ], + "type": "text", + "content": "[44] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, pages 7262-7272, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 440, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 440, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 308, + 440, + 545, + 483 + ], + "type": "text", + "content": "[45] Jiajun Wu, Yibiao Zhao, Jun-Yan Zhu, Siwei Luo, and Zhuowen Tu. Milcut: A sweeping line multiple instance learning paradigm for interactive image segmentation. In CVPR, pages 256-263, 2014. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 485, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 485, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 308, + 485, + 545, + 540 + ], + "type": "text", + "content": "[46] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. SegFormer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 542, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 542, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 308, + 542, + 545, + 575 + ], + "type": "text", + "content": "[47] Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas Huang. Deep grabcut for object selection. arXiv preprint arXiv:1707.00243, 2017. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 577, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 577, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 577, + 545, + 610 + ], + "type": "text", + "content": "[48] Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas S Huang. Deep interactive object selection. In CVPR, pages 373-381, 2016. 2, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 612, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 612, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 612, + 545, + 656 + ], + "type": "text", + "content": "[49] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[50] Yuhui Yuan, Rao Fu, Lang Huang, Weihong Lin, Chao Zhang, Xilin Chen, and Jingdong Wang. Hrformer: High-resolution vision transformer for dense predict. Advances in Neural Information Processing Systems, 34:7281-7293, 2021. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22299" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 195 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[51] Shiyin Zhang, Jun Hao Liew, Yunchao Wei, Shikui Wei, and Yao Zhao. Interactive object segmentation with inside-outside guidance. In CVPR, pages 12234-12244, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 286, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 286, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 286, + 162 + ], + "type": "text", + "content": "[52] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 286, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 286, + 195 + ], + "type": "text", + "content": "[53] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv:2304.06718, 2023. 8" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22300" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Simulating Fluids in Real-World Still Images/d51a02b7-7fc4-4f85-a456-fd173559877f_content_list.json b/2023/Simulating Fluids in Real-World Still Images/d51a02b7-7fc4-4f85-a456-fd173559877f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8b322f300256155e01887cec7127d01b86209546 --- /dev/null +++ b/2023/Simulating Fluids in Real-World Still Images/d51a02b7-7fc4-4f85-a456-fd173559877f_content_list.json @@ -0,0 +1,1673 @@ +[ + { + "type": "text", + "text": "Simulating Fluids in Real-World Still Images", + "text_level": 1, + "bbox": [ + 256, + 130, + 712, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Siming Fan $^{1*}$ , Jingtan Piao $^{1,3*}$ , Chen Qian $^{1}$ , Hongsheng Li $^{2,3,4\\boxtimes}$ , Kwan-Yee Lin $^{2,3\\boxtimes}$ , SenseTime Research, Shanghai AI Laboratory, The Chinese University of Hong Kong, CPII", + "bbox": [ + 107, + 178, + 861, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{fansiming,qianchen}@sensetime.com, 1155116308@link.cuhk.edu.hk,", + "bbox": [ + 200, + 218, + 764, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hsli@ee.cuhk.edu.hk, junyilin@cuhk.edu.hk", + "bbox": [ + 300, + 237, + 666, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/12b0fd2234020a29b22095c20374d547abc6e5befb82d30306aea0c165030ea6.jpg", + "image_caption": [ + "Figure 1. Given a still image and a coarse hint (mask and arrows) of motion as input (left), our model estimates fluid motion to generate an animating video (right). Holynski [11] is a state-of-the-art method. Please refer to the supplemental materials for video playing." + ], + "image_footnote": [], + "bbox": [ + 78, + 258, + 232, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/65a5a608f1cb7fc964b626b8426cd82553465cad58e5080876b0256b469bc9a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 232, + 260, + 893, + 422 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 476, + 313, + 492 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we tackle the problem of real-world fluid animation from a still image. The key of our system is a surface-based layered representation, where the scene is decoupled into a surface fluid layer and an impervious background layer with corresponding transparencies to characterize the composition of the two layers. The animated video can be produced by warping only the surface fluid layer according to the estimation of fluid motions and recombining it with the background. In addition, we introduce surface-only fluid simulation, a 2.5D fluid calculation, as a replacement for motion estimation. Specifically, we leverage triangular mesh based on a monocular depth estimator to represent fluid surface layer and simulate the motion with the inspiration of classic physics theory of hybrid Lagrangian-Eulerian method, along with a learnable network so as to adapt to complex real-world image textures. Extensive experiments indicate our method's competitive performance for common fluid scenes and better robustness and reasonability under complex transparent fluid scenarios. Moreover, as proposed surface-based layer representation and surface-only fluid simulation naturally disentangle the scene, interactive editing such as adding objects and texture replacing could be easily achieved with", + "bbox": [ + 73, + 508, + 473, + 857 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "realistic results. Code, and dataset are publicly available.", + "bbox": [ + 500, + 478, + 880, + 493 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 522, + 632, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given an image of a scene containing liquid such as streams and waterfalls, humans can imagine how the liquid in the still image would move. The problem of animating fluid from a single image has become a blooming topic recently studied by a series of work [6, 11, 27, 9, 21]. Their general pipeline can be summarized as: (1) estimating optical flow that indicates the movement of liquid and (2) synthesizing future frame image based on the motion estimation. Although these methods achieve impressive visual effects on simple fluid regions, how to handle complex contents with challenging transparency, collisions, and thin structures in real-world scenes is still a challenging and open problem.", + "bbox": [ + 496, + 547, + 892, + 743 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The key to the limitation of previous work is that the generation process of videos is modeled in a global manner. Specifically, the subsequent frames are generated based on warping existing textures on input images [3] or neural features [6]. Such operations regard the scene as a whole and warp all the contents in the scene together. For some complex transparent fluid cases, as shown in Figure 2(b), we can see through the water to observe the background rocks beneath the liquid. Textures on those static objects", + "bbox": [ + 496, + 743, + 893, + 878 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution and joint first authorship.", + "bbox": [ + 94, + 875, + 336, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding authors.", + "bbox": [ + 107, + 888, + 243, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Project page: https://slr-sfs.github.io/", + "bbox": [ + 529, + 888, + 810, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15922", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are also improperly deformed and result in unnatural animations. To decompose the influence of motion on fluids and surrounding static objects, we propose a two-layer representation for scenes with fluids, namely Surface-based Layered Representation (SLR), which models the motion of the surface fluid and the rest background separately. To achieve such decomposition from the still image during inference, we propose a temporal-based training scheme, which facilitates the network to learn the decomposition with multiple supervisions over spatial and temporal dimensions of running fluid videos.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another major difficulty in animating a single fluid image is predicting the motion flow of the fluid. There exist two types of methods. The first category is the interactive motion prediction methods. They require sparse hints (ie, labelling the velocity direction and relative amplitude) on random pixels of the liquid region as additional inputs. The feature clustering model [27] or convolutional network [11] is then used to generate a dense motion field. As the precision depends on the level of motion details a user can provide, such methods are time-costing in complex fluid scenes. The second category is automatic methods [6]. These works regard the task as a domain transformation problem, and use image-to-image translation techniques to predict the motion represented by 2-channel optical flows. Ambiguous velocity (e.g., a flat river surface can flow either to the left or to the right, as shown in Figure 2(a)) and tremendous changes of velocity in a period of time caused by the collisions with objects (as shown in Figure 2(c)) may pose great challenges to the above methods. To tackle the challenges, we introduce a simplified Navier-Stokes simulation into the motion estimation process, called Surface-only Fluid Simulation (SFS). Specifically, the initial motion prediction of the liquid region is firstly calculated through sparse user inputs and interpolation inside the whole region. In parallel, a 3D mesh is built based on a monocular depth estimator. Then, the N-S equation is solved on the mesh surfaces and based on the initial motion to obtain a simulated motion field in a short period of time. Finally, since the real thickness of the fluid is neglected during the simulation, a refinement step is expected to compensate for the motion offset at height. Thus, a DNN-based motion translator is applied subsequently to obtain the refined motion field that adapts naturally to the input image. With SFS, we can enforce precise controls on the movement of fluid.", + "bbox": [ + 75, + 261, + 472, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our work has three major contributions. (a) We propose a new and learnable representation, surface-based layered representation, which decomposes the fluid surface and the static objects in the scene to better synthesize the fluid animating videos from a single fluid image. (b) We design a surface-only fluid simulation to model the evolution of the image fluids with better visual effects. (c) Based on the proposed surface-based layered representation", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8f4a06c3ce854edc582cdb7bf66e242a650536f414d983e0696d568af4453a8b.jpg", + "image_caption": [ + "Figure 2. Motivation. (a) Given a still input image, the motion may be ambiguous. (b) Animating via previous SOTA method [11] causes background rocks beneath the liquid improperly deformed. (c) Motion prediction methods such as [11, 27] cannot well capture motion changes caused by the collisions with objects." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 893, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and surface-only fluid simulation, image editing with fluids is achieved with realistic and vivid results. In addition, We can easily extend our fluid animation system to downstream editing tasks, thanks to the SLR and SFS designs in our framework. For example, we can augment objects in the fluid region or background region, and simulate the interactions between the fluids and objects. We can also replace textures to different layers to create different scenes.", + "bbox": [ + 496, + 333, + 893, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 470, + 643, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Video Generation", + "text_level": 1, + "bbox": [ + 500, + 497, + 671, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In early studies, researchers focus on generating new video context by changing the attributes of recorded frames [32, 33]. Traditional methods of video motion generation with movable substances are based on stochastic process analysis [34] and color transfer [32]. Given a reference video indicating the temporal changes of a scene, texture movement and sequential images can be generated using a patch aligned method and energy-based optimizing [33]. Similar methods are applied to single image input [3]. However, obvious artifacts exist, and the smoothness of the video is not guaranteed. These drawbacks make the usage of such methods limited to typical types of scene samples.", + "bbox": [ + 496, + 521, + 890, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, since the development of deep neural networks in image generation field [15], plenty of innovative efforts have been made to improve the synthesis quality of images [7, 16, 8, 35, 25]. Techniques such as GANs [8], perceptual loss [17], partial convolution [25] and cycle-consistency [35] have been widely used as standard components. Built upon the success of image generation, several attempts have created the precedents of neural video generation. A few methods are developed to force the network to learn intermediate scene representation for predicting the future frames or interpolating key frames, once the input sequences and camera trajectory are given (note: as we focus on single image input, we skip the review of", + "bbox": [ + 496, + 704, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15923", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "such multi-frames input trend). Given a single image as input, latent generative methods like [24] proposes an autoregressive model to generate infinite pixels forming sequential outputs. InfinityGAN [23] disentangles global appearances, local structures and textures with a structure synthesizer and texture synthesizer to generate images with spatial size and level of details not attainable before. [31] tries to use Fourier transforms in time variance showing the phase variations to represent the flow and generate moving videos. [20] spends efforts in digging into the semantic meanings in images and generates videos correspondingly to achieve a natural result. [6] decouples motion and appearance through learning intermediate flow fields and color transfer map. However, these methods are either focusing on spatial-wise consistency [20], or paying attention to coarse-level time-variance scene changes [31, 6]. None of them discusses or can handle the particularity of fluid with fine detail motions.", + "bbox": [ + 75, + 90, + 472, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Fluid Simulation", + "text_level": 1, + "bbox": [ + 76, + 376, + 243, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Physical Fluid Simulation. Traditional methods originated from computer graphics to achieve accurate approximations on various types of fluids. Hybrid Lagrangian-Eulerian Methods [2, 36, 10] have successfully simulated the motion of incompressible fluids in a pre-defined 2D or 3D box environment, which solves pressure projection on Eulerian grid and advect on Lagrangian particles. To improve convergence ability and flexibility on complex shapes, material point method (MPM) [1] has been popular and more subsequent methods [18, 19] have been proposed to stabilize the convergence when iterating velocities. However, these methods are time-consuming in large scenes. To reduce the cost of simulating a large-range of fluid, methods have been proposed to simulate only a small range of fluid [4], or the surface of the fluid [5, 13] to produce a faster approximation by avoiding high computational cost in 3D volumetric solvers. Although these works achieve amazing simulation effects, they are still limited to synthetic scenarios due to scene-specific modelling.", + "bbox": [ + 75, + 400, + 472, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Animating Fluids with Data Prior. Given a still image as animation target and a set of real-world water video candidates as animation bank, [31, 29] propose to retrieve a video or a group of videos from the bank that has similar content with the target image as the references. Then they transfer motion and appearance of references to each region of interest on target image to form the water animation video. These methods open the door of real-world fluid animation with data prior in-the-wild. However, the seamless alignment of animated results with source input image can not be well guaranteed and requiring lots of manual efforts. With the development of deep learning, recent works [11, 27] propose to automatically predict each frame of motion and appearance via learning from reconstruction loss", + "bbox": [ + 75, + 688, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with real world videos for training. [11] first simplifies motion estimation part as a single frame motion prediction via motion Eulerian integration and then proposes a deep feature warping technique to narrow down the size of blank areas caused by warping. Finally, it adaptively blends forward and backward features through a learnable parameter $Z$ to generate animated fluid video. Based on [11], [27] proposes to regress fluid motion conditioned to user's sparse guidance and generate paired training data through motion speed clustering. The work further proposes to use multiscale representation to capture different fluid speed in different resolution. Although these methods take a step further to real-world scenes with impressive results, it is hard for them to handle complex context relations over fluids due to the single representation to the entire scene that regards different objects and textures as equal.", + "bbox": [ + 496, + 90, + 893, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike previous work, our system marries the advantages of both physical simulation and learning-based pipelines. For textures, to avoid interplay between fluids and surroundings, we force the network to decouple the scene into a new two-layer representation SLR, that includes a static background and time-variates surface fluid layers2. Meanwhile, the network can hallucinate reasonable and photo-realistic textures in vacated regions thanks to the large-scale data prior. For motion, to skirt scene-specific modelling while keeping physiC reasonability, we propose a three-step motion prediction. We first use sparse labels from the human interface to generate the initial rough motion and then calculate reasonable evolution in a short period by introducing a 2.5D surface simulation, SFS. Later, a smooth motion translator is used to refine the motion trend to better fit real-world examples with data prior. Moreover, as a consequence, we can generate natural fluid animations with the network training on plenty of real-world examples and flexibly edit the main liquid region and boundary condition to create various visual effects.", + "bbox": [ + 496, + 333, + 895, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 647, + 591, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a real-world image that contains liquid, our goal is to generate a video clip that animates vivid flowing of the fluid. To achieve this goal, a system is required to (1) represent complex context relations between fluids and surrounding environments; and (2) to estimate motion fields from single image input with reasonability and efficiency. We propose a novel system with two main designs that satisfy the above requirements. For texture, in contrast with previous methods that regard all elements in the scene as an entirety, our system learns a surface-based layered representation (SLR) (Sec. 3.1) that decomposes the scene into surface fluid layer and impervious background layer. Since the tex", + "bbox": [ + 496, + 672, + 893, + 854 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2This idea shares a similar spirit with [26] in separating a scene into both static and dynamic parts. Whereas, our SLR could handle transparent fluids, and only require a single image as input.", + "bbox": [ + 498, + 862, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15924", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ture characteristics of surface fluid and rest background are assigned to different layers, such a design could overcome the negative influence of motion flows between fluids and surroundings (e.g., improper texture warping, as shown in Figure 2(b)). As for motion estimation, the system provides a three-step motion estimation alternative by introducing a surface-only fluid simulation (SFS) into the motion field prediction (Sec. 3.2). The motion prediction steps combine advantage of both physical- and learning-based motion estimation pipelines, thus ensuring physical reasonability and meanwhile skirting the scene-specific modelling problem. Each component is detailed in following subsections.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Learning of Decomposition of Images", + "text_level": 1, + "bbox": [ + 76, + 281, + 401, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this subsection, we first introduce Surface-based Layered Representation (SLR). Then, we show how to obtain animated videos through SLR. Finally, we present how to learn SLR with a neural network (illustrated in Figure 3), and challenges we meet during optimizing the network since there is no oracle ground-truth supervision for fluid scene decomposition.", + "bbox": [ + 75, + 304, + 468, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Surface-based Layered Representation. Given a still image $I(T_0)$ as input, a series of RGB frames $\\mathbf{I} = \\{I(T_0),\\dots,I(T_n)\\}$ are outputs to synthesize time-space variations of the scene with running fluids. As we aim to handle complex context of fluid scene through disentangling fluid motion and still objects beneath/interacting with surface fluid, we propose to learn a novel intermediate representation whose output is an impervious background layer image $I_b$ , a temporal sequence of surface fluid layer images $\\mathbf{I}_f = \\{I_f(T_0),\\dots,I_f(T_n)\\}$ and transparency factors $\\alpha_{b},\\alpha_{f} = \\{\\alpha_{f}(T_{0}),\\dots,\\alpha_{f}(T_{n})\\}$ , where $\\alpha_{*}\\in [0,1]$ . Transparency factors indicate the contribution of colors on each layer to the final image at each time. In practice, we utilize a convolution network to output the simplification of the representation: $F_{\\theta}: (I(T_0),I(T_n),\\mathbf{M})\\to (I_b,I_f(T_0),I_f(T_n),\\alpha_b,\\alpha_f(T_0),\\alpha_f(T_n))$ , where the network only outputs the estimated first and last frames of $\\mathbf{I}_f$ as well as $\\alpha_{f}$ . $\\mathbf{M}$ is the estimated motion fields of $T_0$ to $T_{1}$ from the single input $I(T_0)$ . We will detail the synthesis of $\\mathbf{M}$ in Sec. 3.2.", + "bbox": [ + 75, + 410, + 468, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We expect the background to be identical across all times, while the fluid surface deforms with realistic motions. To this end, $F_{\\theta}$ is enforced to automatically map the original texture of non-liquid and static regions as well as the reasonable texture of under-liquid surface regions to the background layer (detailed in later subsection). As for fluid surface learning, recurrent estimation for all frames (RNN) is a possible alternative that can be tailored from video gen", + "bbox": [ + 75, + 710, + 468, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "eration [30]. However, such an intuitive design might cause frame distortion as time goes on due to the accumulated errors and hard convergence (also demonstrated in [11]). In contrast, We use the linear combination of the first and last frames to construct the images at intermediate frames. This choice not only transforms the problem into interpolation rather than extrapolation, which is easier to converge, but also is able to generate an endless cyclic video rather than a video of a fixed interval. In detail, we tailor the symmetric softmax splatting, which is a pixel-warping strategy utilized in [11, 28], to our framework. We extend the splatting operation from a single image feature splatting into two levels – fluid surface feature splatting and fluid alpha splatting as follows :", + "bbox": [ + 496, + 90, + 890, + 284 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {f} \\left(T _ {i}; \\hat {\\mathbf {x}} ^ {\\prime}\\right) = \\frac {\\sum_ {\\hat {\\mathbf {x}} \\in \\mathcal {X}} \\alpha \\left(T _ {i} ; \\hat {\\mathbf {x}}\\right) \\cdot D _ {f} \\left(T _ {j} ; \\hat {\\mathbf {x}}\\right) \\cdot \\exp \\left(Z _ {f} \\left(T _ {j}; \\hat {\\mathbf {x}}\\right)\\right)}{\\sum_ {\\hat {\\mathbf {x}} \\in \\mathcal {X}} \\alpha \\left(T _ {i}; \\hat {\\mathbf {x}}\\right) \\cdot \\exp \\left(Z _ {f} \\left(T _ {j}; \\hat {\\mathbf {x}}\\right)\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 289, + 890, + 321 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {f} ^ {\\prime} \\left(T _ {i}; \\hat {\\mathbf {x}} ^ {\\prime}\\right) = \\frac {\\sum_ {\\hat {\\mathbf {x}} \\in \\mathcal {X}} \\alpha \\left(T _ {i} ; \\hat {\\mathbf {x}}\\right) \\cdot \\alpha_ {f} ^ {\\prime} \\left(T _ {j} ; \\hat {\\mathbf {x}}\\right) \\cdot \\exp \\left(Z _ {f} \\left(T _ {j} ; \\hat {\\mathbf {x}}\\right)\\right)}{\\sum_ {\\hat {\\mathbf {x}} \\in \\mathcal {X}} \\alpha \\left(T _ {i} ; \\hat {\\mathbf {x}}\\right) \\cdot \\exp \\left(Z _ {f} \\left(T _ {j} ; \\hat {\\mathbf {x}}\\right)\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 324, + 890, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $j \\in \\{0, n\\}$ is the start index or last index of sequences, indicating forward splatting the frame 0 or backward splatting the frame $n$ , which means the value $j$ is:", + "bbox": [ + 496, + 364, + 890, + 410 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nj = \\left\\{\\begin{array}{l l}0,&\\text {i f} \\hat {x} ^ {\\prime} = M _ {0 \\rightarrow t} (\\hat {x}),\\\\n,&\\text {i f} \\hat {x} ^ {\\prime} = M _ {0 \\rightarrow t - N} (\\hat {x}).\\end{array}\\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 430, + 890, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$D_{f}(T_{i};\\hat{\\mathbf{x}}^{\\prime})$ is the fused fluid feature value of target frame at pixel $\\hat{\\mathbf{x}}$ , and $\\alpha_f^\\prime (T_i;\\hat{\\mathbf{x}}^\\prime)$ is the transparency value of target frame at pixel $\\hat{\\mathbf{x}}$ . $D_{f}(T_{i};\\hat{\\mathbf{x}})$ represents value of $D_{f}(T_{i})$ at pixel $\\hat{\\mathbf{x}}$ , $\\alpha_{f}(T_{i};\\hat{\\mathbf{x}})$ and $\\alpha_f^\\prime (T_i;\\hat{\\mathbf{x}})$ are similar. During inference, we set $D_{f}(T_{n}) = D_{f}(T_{0})$ and $\\alpha_{f}(T_{n}) = \\alpha_{f}(T_{0})$ since only single image is available. $\\mathcal{X}$ is the set of pixels satisfying $M_{0\\rightarrow t}(\\hat{\\mathbf{x}}) = \\hat{\\mathbf{x}}^{\\prime}$ or $M_{0\\rightarrow t - N}(\\hat{\\mathbf{x}}) = \\hat{\\mathbf{x}}^{\\prime}$ , $M_{0\\rightarrow t}$ is displacement map generated from $\\mathbf{M}$ through euler integration [11, 28], and $\\alpha (T_i;\\hat{\\mathbf{x}}^{\\prime})$ is a temporal coefficient to fuse features from forward warping and backward warping to ensure endless looping:", + "bbox": [ + 496, + 472, + 890, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha \\left(T _ {i}; \\hat {\\mathbf {x}} ^ {\\prime}\\right) = \\left\\{\\begin{array}{l l}\\frac {t}{n},&\\text {i f} \\hat {x} ^ {\\prime} = M _ {0 \\rightarrow t} (\\hat {x}),\\\\1 - \\frac {t}{n},&\\text {i f} \\hat {x} ^ {\\prime} = M _ {0 \\rightarrow t - n} (\\hat {x}).\\end{array}\\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 647, + 890, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then a fluid feature decoder is applied to $D_{f}(T_{i})$ to generate surface fluid image $I_{f}(T_{i})$ and a alpha refinement network is applied to $\\alpha_f'(T_i)$ to generate transparency alpha $\\alpha_{f}(T_{i})$ to avoid vacant alpha particles.", + "bbox": [ + 496, + 688, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Animating Fluids with SLR. Once we obtain the final surface fluid layer $I_{f}(T_{i})$ and transparency $\\alpha_{f}(T_{i})$ , the reconstructed final image $I(T_{i})$ can be acquired by adding the impervious background layer $I_{b}$ back to the surface fluid layer $I_{f}(T_{i})$ with $\\alpha_{b}$ . Here we use separate $\\alpha$ channels for background and fluid transparency predictions. This design helps to stabilize the training of decomposition. The core reason is that when a single fluid transparency $\\alpha_{f}$ is warped according to the motion flow, it leaves a blank region on the boundary, and makes the gradients hard to be", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "3The word \"impervious\" is not rigorous, as background layer is not only formed from static object but also non-surface fluid which contains air molecules in real environment. However, we assume the fluid is incompressible, and thus a little bit abuse the word \"impervious background\" to indicate the non-surface fluid regions.", + "bbox": [ + 75, + 838, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15925", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d3022add8b5dd7b327414a4ffeb8837e765a9c118d7b06f824d305d6927933d3.jpg", + "image_caption": [ + "Figure 3. Training and testing pipelines. The background layer and background alpha are predicted only at the first frame. Both fluid features and fluid alpha are predicted at the first and last frames, and then symmetricly splatted and blended using the softmax splatting at frame $T_{i}$ . A decoder is used to decode fluid features into images and refined alpha. For frame $T_{0}$ , motion is calculated following pipeline Figure 4, and for frame $T_{i}$ , motion is calculated through Euler integration. During inference, $T_{n}$ is replaced by $T_{0}$ to create a cyclic video." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 893, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "backpropagated. Although with the two-channel implementation, there still is a $\\alpha_{b}$ on that vacated regions to fascinate the network refining the texture. In formula we have", + "bbox": [ + 75, + 470, + 470, + 517 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI \\left(T _ {i}\\right) = \\frac {\\alpha_ {f} \\left(T _ {i}\\right) I _ {f} \\left(T _ {i}\\right) + \\alpha_ {b} I _ {b}}{\\alpha_ {f} \\left(T _ {i}\\right) + \\alpha_ {b}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 527, + 468, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For corresponding network architectures to achieve the above targets, as shown in Fig 3, the encoder is used to extract background texture and $\\alpha$ channel as well as surface fluid layers' features and its $\\alpha$ . The decoder is used to refine the fused fluid layer that combines the layers from the first and last frames to obtain the complete surface fluid layer, and the final animated image is reconstructed by compositing background and surface fluid layers.", + "bbox": [ + 75, + 568, + 468, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimizing the SLR. The training process is divided into three parts, in which, we separately train (1) the surface fluid branch containing the encoder that predicts the fluid feature layer from image input and the decoder that reconstructs the warped fluid layer, (2) the background branch that predicts the background layer from the image input, and (3) jointly train them together with $\\alpha$ learning and surface fluid/ background branches finetuning. For each training step, the paired training data $< (I(T_0), I(T_n)), I(T_i) >$ are randomly picked among 60-frames sequence without constraint on interval. We use dense optical flow estimation network [14] to extract reasonable movement of each video as pseudo motion ground-truth, and to ensure smoothness, we average the flow with window size of 30 frames. Due", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to space limits, we detail the training of each stage in the supplementary materials.", + "bbox": [ + 498, + 470, + 890, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Motion Field Estimation", + "text_level": 1, + "bbox": [ + 500, + 512, + 723, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the previous section, we have described the core necessities to learn an SLR but skipped the synthesis process of one of its input, motion fields $\\mathbf{M}$ . In this section, we discuss how to obtain $\\mathbf{M}$ .", + "bbox": [ + 496, + 537, + 890, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, the motion of each frame can be generated from a conventional optical flow estimator [14]. During inference, as the input is a single image, we expect to synthesize a reasonable motion map from the observed scene. We propose a motion estimation process that separates the motion prediction into three parts taking advantage of both physical-based and learning-based pipelines for fluid simulation. First, interactive labelling of the sparse velocity of several pixels on the image is required to guide the generated motion directions of $I(T_0)$ , along with a fusion with weights calculated based on Gaussian filtering of the distance maps from the labeled pixel to build a dense velocity initial map. Second, a surface-only fluid simulation is calculated based on initial motion and estimated scene depth to simulate a following convincing evolution of the liquid. Note that this step only calculates for the first frame. Last, a DNN-based pixel-to-pixel motion translator is utilized to refine the flow to create a more realistic and natural motion map. This translator could be considered as generating a compensation for a detailed motion map that is", + "bbox": [ + 496, + 598, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15926", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9f75a0df9b2dfe7966d6e6e2026c7ada309e5193716dddf1816587371302d9ed.jpg", + "image_caption": [ + "6uizisauus 6uizisauus pue uonnnsr (e)" + ], + "image_footnote": [], + "bbox": [ + 122, + 89, + 295, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ff40a85ade35dcae10f018149db4dca6fee956f16d88fd2f044c6914eb4909f0.jpg", + "image_caption": [ + "2. Label Mask and Velocity Hint" + ], + "image_footnote": [], + "bbox": [ + 295, + 89, + 468, + 161 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/38ba20ee3bd6a16b12cfe736e0ce1ee1d431ed8ed56b9d0b942b7409c9063224.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 124, + 172, + 236, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/641bdf13877f0b84f955c378646caa2c5eff228b0dbeb392703208ab91049648.jpg", + "image_caption": [ + "1. Input Image", + "4. Mono-Depth Estimation" + ], + "image_footnote": [], + "bbox": [ + 241, + 172, + 352, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/145265d27db8f6cafa89256a9819cbb45795e7a0b25c033a601a25f9857dcf07.jpg", + "image_caption": [ + "5. Build 3D Mesh" + ], + "image_footnote": [], + "bbox": [ + 356, + 172, + 468, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f279058ea667c6d55a7dc890ea243467c79ea322c4a23da6aadd0544f0f915bb.jpg", + "image_caption": [ + "3. Generate Dense Velocity", + "6. Mesh-Based Simulation" + ], + "image_footnote": [], + "bbox": [ + 124, + 224, + 236, + 271 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cb7bc957b4e56a7b68ff354cadf2ad58acddd720b29274166606f09d02790e17.jpg", + "image_caption": [ + "7. Refine Velocity" + ], + "image_footnote": [], + "bbox": [ + 241, + 224, + 352, + 272 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a9caf5cbf6b56b3d70b51435e3c8b1c685f4331ecff0be76fab467a68b5d7d2b.jpg", + "image_caption": [ + "8. Synthesize Result" + ], + "image_footnote": [], + "bbox": [ + 356, + 224, + 467, + 271 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6fa6e7a9dad637d9e9b745aa4f1cd8ceb32d87b1e7a806558a06a88820406200.jpg", + "image_caption": [ + "9. Editing AR objects" + ], + "image_footnote": [], + "bbox": [ + 125, + 281, + 236, + 324 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9e09676cde456c99c2180fea141999763d3ed4cf80bc754826a694ab22dc6ce8.jpg", + "image_caption": [ + "10. Add object in Fluid Layer" + ], + "image_footnote": [], + "bbox": [ + 241, + 281, + 351, + 324 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ed95a40ef165f10778aa5e4e356a7caa37a6c6e001319ebe3fd53364d296d4fe.jpg", + "image_caption": [ + "1. Add object in Background layer" + ], + "image_footnote": [], + "bbox": [ + 356, + 281, + 467, + 324 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a33d82210fdda115b7af5a5dceda58b9b557fa8a755528ec77f37bb3e1ab249e.jpg", + "image_caption": [ + "12. Mark Intersection Vertex \"Solid", + "Figure 4. Motion calculation pipeline. (a) is initialization of fluid region, rock region, motion, and Mesh. (a,b) combine to form our 2.5D simulation and animation pipeline, where (4-5) is not included if using 2D simulation method. (a,c) combine to form our editable simulation and synthesis pipeline, where the user can edit 3D objects and create corresponding motion." + ], + "image_footnote": [], + "bbox": [ + 124, + 333, + 236, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/885ccab6f9e04014b188dabc5b45106d9255102c1956cfe3b2e7e43344d5247e.jpg", + "image_caption": [ + "Mesh-Based Simulation" + ], + "image_footnote": [], + "bbox": [ + 241, + 333, + 352, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bf57bb705454dc24c6444a0c7c7f97f25e4ca55063a174d0a7cf498e5083ff37.jpg", + "image_caption": [ + "14. Edited Synthesized Result" + ], + "image_footnote": [], + "bbox": [ + 356, + 333, + 467, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "harmonic with fluid textures. The pipeline is illustrated in Figure 4.", + "bbox": [ + 75, + 482, + 468, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Initial Motion from Interactive Sparse Labeling. To obtain determined as well as editable motion direction, we choose an interactive alternative to generate initial flow for image $I(T_0)$ following [27]. With no more than 10 discrete notations on images, and a fixed segmentation of the liquid region, we use nearest-neighbors-averaging to assign all the pixels that are inside the liquid region an initial velocity.", + "bbox": [ + 75, + 513, + 468, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Surface-only Fluid Simulation. After the initial state is estimated, we tailor the traditional graphics pipeline with discretization and linear equation to simulate the follow-up evolution of the liquid. Intuitively, we can directly apply simulation on 2D grids that align with image pixels by initializing each pixel as an Eulerian grid and several Lagrangian particles near each pixel. However, since such perspective projection of the 3D scene may cause severe distortion when direct manipulating the motion in 2D, as shown in Figure 7, we thus turn to use a mono-depth estimator [22] to estimate the 3D surface of the liquid. After the 3D surface is constructed inside the region of the liquid, a mesh surface-based velocity is calculated, with the restriction of projected velocity:", + "bbox": [ + 75, + 619, + 468, + 830 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {d}{d t} \\left[ \\begin{array}{l} u \\\\ v \\end{array} \\right] = \\left[ \\begin{array}{c c} f x & 0 \\\\ 0 & f y \\end{array} \\right] \\left[ \\begin{array}{c c c} 1 / z & 0 & - x / z ^ {2} \\\\ 0 & 1 / z & - y / z ^ {2} \\end{array} \\right] \\left[ \\begin{array}{c} \\frac {d x}{d t} \\\\ \\frac {d y}{d t} \\\\ \\frac {d z}{d t} \\end{array} \\right] \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 829, + 468, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $fx, fy, cx, cy$ is camera intrinsic parameters, $u, v$ is", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8e2bc8d256c01b1d907d1362cbb75cec30df8e4b2a94a9c4a8074026b7e87162.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 88, + 638, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f4f3659c2596b56d5aca88aa5b0fad3dbd2deeb6e229609166a5399e73c8a775.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 89, + 761, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c7e56c13e2b0350c9f6544a503ea131c67bddcb7eec7fdf51eb2a3bfd7dc9ae0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 89, + 890, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2793bf549947b7c855f27ccaa441b1271f533b5e07813fb1f12ea4ce8e0f5286.jpg", + "image_caption": [ + "Figure 5. Examples for Different Fluid Scenes. (a) The surface thickness of the waterfall and river are gauzy. (b) The fountain cannot be considered a surface. (c) Fluid with vortex contains complex motion." + ], + "image_footnote": [], + "bbox": [ + 501, + 146, + 640, + 202 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a3397f45c2a8e229f0d7b5a17ec9c61dd74fef053035c366b54bb5ce64d3485f.jpg", + "image_caption": [ + "(b)Not a Surface" + ], + "image_footnote": [], + "bbox": [ + 676, + 145, + 725, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bb145d3a50eb212c96f65a3cf2513508f0b9399a40543016b34cc91f90dc5882.jpg", + "image_caption": [ + "(c)Complex Initialization" + ], + "image_footnote": [], + "bbox": [ + 764, + 145, + 890, + 199 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the velocity on the projected image and $x, y, z$ the 3D position in the scene. Since we have modelled the liquid as a surface, we replace traditional Laplacian operator with Laplacian-Beltrami operator on the mesh and use Mixed-Euler-Lagrange approach [36] implemented in Taichi [12] to simulate consequent motion of each vertex. Detailed formula derivation can be seen in the supplementary. The particle binding on the surface is randomly sampled according to the surface area. When a particle is moved outside the mesh surface, we use the nearest projection to force the position to fall onto the mesh. For the boundary of the mesh, we calculate the velocity of the boundary triangle, judging whether the velocity on the edge boundary is inward or outward. For outward edges, the particle is abandoned after moving out of the liquid region. For inward region, a particle source is built to maintain the overall number of particles to be roughly stable. With the texture of surface fluid layer binding to each vertex, we can warp fluid layer with a soft rasterizer to simulate flows. In this way, given arbitrary flow initial state, a stable velocity can be calculated to approximate real-world flow after a period of simulation.", + "bbox": [ + 496, + 266, + 890, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Smooth Motion Translator. With the help of above $2.5D$ fluid simulation, we can generate a time-variant velocity among the sequence. However, the simulated velocity is still a rough trend since it only works on fluid surface without considering thickness variation of fluids. Thus, the SFS can work well in cases like Figure 5(a). It fails in cases with distinct depth measurement differences, as shown in the Figure 5(b) due to the lack of height information during calculation. Also, It is fragile to cases with complex initial state as shown Figure 5(c) due to error-prone/time-consuming interactive initial motion labeling for such cases. Thus, we need extra scheme to compensate or rectify for detailed motion that is harmonic with the fluid textures. To this end, we use a convolution-based network to transfer the simulated motion into a more detailed and reasonable one, which is pixel-aligned with image textures. The network is trained as a traditional style-transfer task, with a L2 loss between output velocity and ground truth. An external patch discriminator is used to refine the pattern of the velocity distribution. All the settings are similar to the image translation task pipeline [15].", + "bbox": [ + 496, + 583, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15927", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 89, + 209, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we present experiments and detailed analysis to demonstrate the contributions of our method both quantitatively and qualitatively. For all the cases shown in this paper, please refer to our demo video and project page for better dynamic visual effects.", + "bbox": [ + 75, + 114, + 468, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Quality of Fluid Video Synthesis", + "text_level": 1, + "bbox": [ + 76, + 199, + 362, + 215 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. To learn SLR, we train our network on a series of videos containing a large region of fluid movement and still background [11]. For training of motion refinement network, we use the same settings in [27]. All networks are trained at a resolution of $256 \\times 256$ .", + "bbox": [ + 75, + 223, + 468, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets and Evaluation Metrics. In addition to validation set in [11], we collect extra 126 scenes to form a new test set, which we called Complex Liquid Animation in-the-Wild (CLAW) test set, for evaluating the performance on both dense fluids (such as rivers and oceans) and sparse or transparent fluids (such as streams or splashes). Note that, in this subsection, the motion for all the methods to be compared is guided by the optical flow of the whole video to distinguish the performance of the synthesizing techniques. We use PSNR to show the whole average error, SSIM to show the errors in the region with plenty of textures and LPIPS, a perceptual based loss, to show visual quality difference. For more details, please refer to supplementary.", + "bbox": [ + 75, + 297, + 468, + 494 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c3373a7e1e3c48845d7076140da2f7bcd532a304b0a17cc0ba5c16d21fa6dcea.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Quantitative comparison. (a) Results on Holynski's common val set [11], evaluating first to 60-th frames. \"Fluid Region\" means static region is replaced by input image during metric calculation for focusing on the statistic of fluid. This could be done by replacing the non-fluid region with the corresponding textures of input images. (b) Results on CLAW test set." + ], + "table_body": "
DatasetMethodsAll RegionFluid Region
LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑
HolysnikReproduced Holysnik0.079825.030.77870.065725.880.8007
CommonModified Holysnik(Baseline)0.079324.750.77580.065625.720.8000
Val SetOurs0.083425.140.77950.065726.100.8030
OurReproduced Holysnik0.206720.260.59550.202920.360.5961
CLAWModified Holysnik(Baseline)0.207819.970.59230.204120.100.5934
TestsetOurs0.204020.790.60800.197520.800.6077
", + "bbox": [ + 78, + 500, + 468, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparisons with baseline. Table 1 shows the results between our method and state-of-the-art method [11], which is a single layer learning-based system, a typical method that animates the scene in a global manner. We re-implement [11] in Pytorch (Reproduced Holynski), since the source code is not public available. For other comparison baselines, Modified Holynski is built upon Reproduced Holynski but with the modification of replacing convolution to partial convolution in the fluid decoder. The third model (Ours) is our SLR model4. Three observations can", + "bbox": [ + 75, + 667, + 468, + 820 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "be included from Table 1: (1) PSNR and SSIM metrics of \"Modified Holynski\" drop a little compared to \"Reproduced Holynski\" in both statistics on all- and fluid region. We infer that, although partial convolution could better capture context than conventional one to impart reasonable texture in vacated regions in theory, simply replacing partial convolution into single-layer framework cannot essentially improve the image quality due to the wrong context raised from improper warping of both background and fluid. While, with SLR, the partial convolution can strengthen its advantage on context capturing for more proper fluid context textures provided in surface fluid layer. (2) The LPIPS of Ours model drops a little compared to the Modified Holynski model in the \"All Region\" of Holynski common validation set, but has nearly the same LPIPS in the \"Fluid Region\". For the other two metrics, our model leads to an uptick. Since most regions of Holynski's validation set are static, the blending of the fluid and background in the non-fluid region could lead to slight performance drops in terms of LPIPS. However, it could be eased by replacing the non-fluid region with the corresponding textures of input images, thus we care more about the performance in the \"Fluid Region\" instead of the \"All Region\". Such a phenomenon tells us that the two-layer decomposition design in Ours model does not harm the reconstruction ability on final images. To better unfold the analysis, we further conduct the close-up evaluations on Holynski's val set. Concretely, we crop the fluid region by motion threshold and only evaluate these visible fluid background regions. The LPIPS, PSNR, SSIM of Holynski's method are 0.230, 19.01, 0.565, and Ours are 0.227, 19.64, 0.581. This setting leads to larger performance margins with the baseline, which demonstrate the advantage of our system. (3) When evaluating only in the fluid region, we can see model Ours reaches a comparable result with the \"Modified Holynski\" model in Holynski validation set [11], but improves significantly in CLAW test set. The major reason behind this observation is that when it comes to complex scenes as included in our CLAW test set, our proposed two-layer representation suppresses the influence between surrounding and surface fluids. Thus we can achieve better results than single-layer baselines that regard all elements in the scene as an entirety during warping.", + "bbox": [ + 496, + 90, + 890, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons with SOTA methods. In this paragraph, we provide comparisons with more state-of-the-art methods - Endo et.al [6], Tavi et.al [9], and Mahapatra et.al [27]. The first two methods are centered on animating general scenes, where Endo et.al [6] is designed for landscape, and Tavi et.al [9] focuses on animating periodic patterns in still images. These two methods could be applied to fluid scenarios. In contrast, Mahapatra et.al [27] is specialized for fluid scenarios, where a multi-scale representation is introduced to capture different fluid speeds in different resolutions. Concretely, as Endo's code is publicly available, we", + "bbox": [ + 496, + 734, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "4For Ours, we use (Ours (stage 1) in Table 1 in our supplementary material.) as surface fluid model initialization. Since our training contains three stages, and we train 100 epochs at the first stage to get the model (Ours (stage 1)), then decay learning rate and fine-tune 50 epochs for fluid at the last stage in practice, we also apply the same training strategy for the first two baselines to ensure fair comparisons.", + "bbox": [ + 75, + 827, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15928", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "compare our system with it both quantitatively and qualitatively. For other methods, we compare them qualitatively in Fig 6. Specifically, the result of Mahapatra is from official project page, and result of Tavi is from official app. Endo's method can only control motion with a latent code. Thus, we test ten different latent codes and select the best one to compare. We observe that these three methods suffer from animation distortions with different levels – (1) from the zoom-in perspective, all baselines improperly move the still stones with the fluids. Specifically, Tavi's results have significant ghosting artifacts, which might be caused by its one-fold periodic motion and alpha-blending designs. (2) From the full image perspective, Endo's method has difficulty in estimating fluid motion. The river's motion is incorrectly predicted to be from right to left, and the waterfall flows upwards to the right. The quantitative results (Tab 1) provide the demonstration from another aspect. This phenomenon is foreseeable, as Endo's framework considers motion in a global manner. In contrast, our two-layer method animates plausible fluid without affecting still stones. Please refer to the supplementary material for more comparisons and straightforward video effects.", + "bbox": [ + 76, + 90, + 472, + 422 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c3b82864853e11bcd14c71f95201b96a64b759c2eb71a1633b4fa0b416b33d6d.jpg", + "image_caption": [ + "Figure 6. Qualitatively Comparisons. (e-f) are tested in $288 \\times 512$ , others are tested in $\\mathrm{W} = 1280$ . LPIPS/PSNR/SSIM of the whole video are colored in orange." + ], + "image_footnote": [], + "bbox": [ + 76, + 426, + 276, + 604 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b95a418ae96c45f6b2e1bcc9516afda9dffd34d0b5d87ee641691ea4a8937fbb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 426, + 468, + 604 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/134f49c97552258fdbdf4c130f9e2df080fc70b66ac1ce656f8e68e677838a42.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsHolynski validation setCLAW test set
All RegionFluid RegionAll RegionFluid Region
LPIPS ↓PSNR ↑SSIM ↑LPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIM
Endo[6] HD0.35716.270.3510.12022.310.7410.35715.920.2960.34716.380.331
Ours HD0.14223.430.7150.11823.780.7280.19620.990.6110.19020.990.610
", + "bbox": [ + 78, + 657, + 468, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Study. We conduct a user study to compare visual effects between single layer baseline (Modified Holynski) and our two layers method(Ours in Table 1) on the full set of CLAW testset. For each scene, we animate 60 frames and output a two-second video. Animated videos are shown side by side without providing ground truth videos as reference to guide the participant focusing on the reasonability and photo-realistic. We ask the participants to select which one is better or neither is better, mainly according to three dimensions (photo-realistic, stereoscopic, and high-fidelity).", + "bbox": [ + 75, + 719, + 468, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Table 3, the participants can make a distinction between the two methods over most samples, leav", + "bbox": [ + 76, + 869, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/17eb2b988aa2970d3f65d355554e3032f162d2619919bd59b25f9cb49d04b567.jpg", + "image_caption": [ + "Figure 7. Different simulations. Orange arrows point out motion direction. All simulation methods use $75\\%$ weak incompression simulation instead of fully incompression for better visual effect. (row 1)Input image, sparse motion generated by sparse user hint required by each method. (row 2-4)ablation study of Mahapatra's method [27], 2D simulation and 2.5D SFS." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 699, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/595e45724f7d490db88b74f8e91a68fef9e6387900203677e022853619390eb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 88, + 893, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing $27.2\\%$ samples difficult to tell. Our method achieves better scores than the single-layer method [11] with $52.3\\%$ sample voting. This indicates the effectiveness of the proposed SLR in other aspects.", + "bbox": [ + 496, + 330, + 892, + 392 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fe1d90ac1d8386dfb75d8d6365a261781035ef415cde6cc480ef0c3247596e35.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons with Endo's method." + ], + "table_footnote": [], + "table_body": "
Single layerOursNeither
E-O-N125(20.5%)319(52.3%)165(27.2%)
", + "bbox": [ + 532, + 396, + 859, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. User study. The subjective scores are voted on all samples of our CLAW testset, which includes 122 videos with various real-world fluids scenarios(e.g., river, stream and waterfall) and complex context surrounding (e.g., semi-/transparent liquid, collisions and thin structures).", + "bbox": [ + 496, + 429, + 892, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Different Motion Prediction", + "text_level": 1, + "bbox": [ + 500, + 513, + 750, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 7 shows different motion prediction pipelines. Pipeline 1: An interactive sparse labelling of the motion and then extending all the velocities to the whole moving region, which results in a constant flow that remains large gap with realistic motion. We can see that the motion appears reasonable global with a network refinement. However, for local regions around static rocks, the flow of liquid is improperly stiff and has no interactions with the rock. While in real-world scenes, there should be vortex and sprays caused by collisions. Pipeline 2: Built upon pipeline 1, while before the motion refinement stage, we additionally initialize grids with fluid, rock or air masks as well as placing particles around each grid, then advocate particles by edited motion for once. Later, a NS equation with the incompressible fluid assumption is solved on 2d grid to obtain the motion of the next frame, which we call \"Simulated 2D Motion\" in the figure. Considering pipeline 2 is simulating on a 2D image plane, a refinement step (using a smooth motion translator) is applied to compensate for the motion offset at height and fine details. Pipeline 3: instead of simulating on a 2D image plane, the N-S equation is solved in $2.5D$ . Specifically, we regress the depth with a pretrained monocular depth estimator and consequently build a 3D mesh over the fluid. Then, we solve the N-S equation on the mesh surface. With", + "bbox": [ + 496, + 537, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15929", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3cb386aad973956f81bfac3dd0c14d7a0a3b0d27f5eb143130268d01dfbd8612.jpg", + "image_caption": [ + "Figure 8. Interactive Editing. (Bottom) Initial all moving region as 'Fluid' statue. (Top) Initial 'Solid' statue at overlap area of mesh and new object." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 467, + 203 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the help of simulation, we can see in the left scene of the figure that the effect of turbulence around the rock can be animated, and the trend maintains the refined output. From the right scene in the figure, we can observe that the absolute value of the velocity is unnatural in Pipeline 2, which regards all particles on an identical plane. However, since the liquid surface leans in the picture, there should be a foreshortening effect on the projection. In contrast, pipeline3 samples particles on the surface and such simulation enable near-plane motion to carry more speed.", + "bbox": [ + 75, + 258, + 467, + 410 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Quantitative Comparisons. Tab 4 shows the ablations of motion refinement in terms of endpoint error metric (lower is better). We evaluate all scenes in both two datasets with user input limited to five sparse hints. The results prove motion refinement is essential for achieving a better ability to represent GT motion. Refer to supplementary for quantitative metrics of the three motion prediction pipelines.", + "bbox": [ + 75, + 410, + 467, + 515 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/165f55a44214b741e76712de9bf6484922f803d601b2db33648536eccaad6aeb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsDataset
Holynski Val SetCLAW Test Set
DenseMotion0.2990.596
RefinedMotion0.2690.582
", + "bbox": [ + 130, + 521, + 416, + 569 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 4. Quantitative Motion Refinement Ablation.", + "bbox": [ + 109, + 570, + 429, + 580 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Byproducts", + "text_level": 1, + "bbox": [ + 76, + 588, + 197, + 604 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Interactive Editing. We can augment fluid scene while keeping realistic fluid animation. For example, we can place an imaginary stone in a river that is originally slack water, and the flow could be updated to the new one with the effect of a vortex around the stone, as can be viewed in Simulation w/ Fluid-Rock Interaction of Figure 8. To achieve this, we only need to calculate interface of inserted stone mesh and fluid surface mesh. Then, we classify the part of stone above/beneath the river surface according to the depth. Later, fluid motion is updated with boundary conditions during simulation. Refer to supplementary for details.", + "bbox": [ + 75, + 613, + 467, + 779 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Layer Replacement Editing. We can change texture attributes of both surface fluid and background by simply replacing either fluid layer or background layer with another reference texture. For example, as shown in Figure 9, the fluid layer can be replaced by/regulated with a starry sky video with an endless loop to form dreamlike fluid effects that might remind the user of a Chinese clas", + "bbox": [ + 75, + 794, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/788a76442a5b61cd572bdf34b31f41d0eba9e806f75dd7cfbc705355ac9b721d.jpg", + "image_caption": [ + "Figure 9. Layer Replacement Editing. First row shows our SLR output and new background. The second row shows results of our background layer replacement effects, fluid layer replacement effects, and replacement effects using motion instead of alpha Arrows specify the motion direction and speed." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 888, + 224 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "sical poetry—as if the Silver River fell from azure sky. For a downstream task like such texture attribute replacement, accurately segmentation of all fluid regions is necessary to reach plausible visual effects. However, it is hard to be accomplished by brute force segmentation with either motion prediction (Fluid Replacement Editing by Motion Threshold w/o SLR in the figure) or manually labelling. Since real-world fluid scenes contain complex context relations such as semi-transparency, different motion behaviour among fluid regions in one scene and so on. On the contrary, since our proposed surface-based layer representation (SLR) naturally disentangles scene into surface fluid and background, we can achieve such editing easily.", + "bbox": [ + 496, + 306, + 890, + 502 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 511, + 617, + 526 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We propose a system for fluid animation from a still image. Specifically, for fluid texture, a neural network is designed to decompose the still image into surface liquids and background, which form a novel two-layered representation, SLR, to represent fluid scene. Compared to previous work, SLR can help better simulating the transparent fluids as well as other fluids scenes with complex surrounding. For motion estimation, we introduce surface-only fluid simulation (SFS), which provides a wider range of editing abilities on the fluid image. The system ensures physical reasonability and meanwhile skirting the scene-specific modelling problem. However, as our method works in a semi-supervised manner, without supervision from ideal decomposition or graphic-based rendering, the outputs may have some artifacts on boundaries. We expect more research on better disentangling and representing fluid surface with backgrounds, as well as more precise motion prediction.", + "bbox": [ + 496, + 537, + 890, + 792 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements. This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK.", + "bbox": [ + 496, + 795, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15930", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Scott G Bardenhagen and Edward M Kober. The generalized interpolation material point method. CMES, 5:477-496, 2004. 3", + "[2] Robert Bridson. Fluid simulation for computer graphics. AK Peters/CRC Press, 2015. 3", + "[3] Yung-Yu Chuang, Dan B Goldman, Ke Colin Zheng, Brian Curless, David H Salesin, and Richard Szeliski. Animating pictures with stochastic motion textures. ACM TOG, 24:853-860, 2005. 1, 2", + "[4] Vincenzo Citro, Paolo Luchini, Filippo Giannetti, and Franco Auteri. Efficient stabilization and acceleration of numerical simulation of fluid flows by residual recombination. J. Comput. Phys., 344:234-246, 2017. 3", + "[5] Fang Da, David Hahn, Christopher Batty, Chris Wojtan, and Eitan Grinspun. Surface-only liquids. ACM TOG, 35:1-12, 2016. 3", + "[6] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating landscape: self-supervised learning of decoupled motion and appearance for single-image video synthesis. ACM TOG, 38:175:1-175:19, 2019. 1, 2, 3, 7", + "[7] L. A. Gatys, A. S. Ecker, and M. Bethge. Image style transfer using convolutional neural networks. In CVPR, 2016. 2", + "[8] I.J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial networks. NeurIPS, 3:2672-2680, 2014. 2", + "[9] Tavi Halperin, Hanit Hakim, Orestis Vantzos, Gershon Hochman, Netai Benaim, Lior Sassy, Michael Kupchik, Ofir Bibi, and Ohad Fried. Endless loops: detecting and animating periodic patterns in still images. ACM TOG, 40:1-12, 2021. 1, 7", + "[10] Francis H Harlow. The particle-in-cell computing method for fluid dynamics. Methods Comput. Phys., 3:319-343, 1964. 3", + "[11] Aleksander Holynski, Brian L Curless, Steven M Seitz, and Richard Szeliski. Animating pictures with eulerian motion fields. In CVPR, 2021. 1, 2, 3, 4, 7, 8", + "[12] Yuanming Hu, Tzu-Mao Li, Luke Anderson, Jonathan Ragan-Kelley, and Frédo Durand. Taichi: a language for high-performance computation on spatially sparse data structures. ACM TOG, 38:201, 2019. 6", + "[13] Libo Huang, Ziyin Qu, Xun Tan, Xinxin Zhang, Dominik L Michels, and Chenfanfu Jiang. Ships, splashes, and waves on a vast ocean. ACM TOG, 40:1-15, 2021. 3", + "[14] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, 2017. 5", + "[15] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, 2017. 2, 6", + "[16] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 2", + "[17] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 2" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] ByungMoon Kim, Yingjie Liu, Ignacio Llamas, and Jaroslaw R Rossignac. Flowfixer: Using bfecc for fluid simulation. Technical report, Georgia Institute of Technology, 2005. 3", + "[19] Theodore Kim, Nils Thürey, Doug James, and Markus Gross. Wavelet turbulence for fluid simulation. ACM TOG, 27:1-6, 2008. 3", + "[20] Pierre-Yves Laffont, Zhile Ren, Xiaofeng Tao, Chao Qian, and James Hays. Transient attributes for high-level understanding and editing of outdoor scenes. ACM TOG, 33:1-11, 2014. 3", + "[21] Thi-Ngoc-Hanh Le, Chih-Kuo Yeh, Ying-Chi Lin, and Tong-Yee Lee. Animating still natural images using warping. TOMM, 2022. 1", + "[22] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In CVPR, 2018. 6", + "[23] Chieh Hubert Lin, Hsin-Ying Lee, Yen-Chi Cheng, Sergey Tulyakov, and Ming-Hsuan Yang. Infinitygan: Towards infinite-resolution image synthesis. arXiv preprint, arXiv:2104.03963, 2021. 3", + "[24] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In ICCV, 2021. 3", + "[25] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In ECCV, 2018. 2", + "[26] Erika Lu, Forrester Cole, Tali Dekel, Andrew Zisserman, William T Freeman, and Michael Rubinstein. Omnimatte: associating objects and their effects in video. In CVPR, 2021. 3", + "[27] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. arXiv preprint, arXiv:2112.03051, 2021. 1, 2, 3, 6, 7, 8", + "[28] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In CVPR, 2020. 4", + "[29] Makoto Okabe, Yoshinori Dobashi, and Ken Anjyo. Animating pictures of water scenes using video retrieval. Vis. Comput., 34:347-358, 2018. 3", + "[30] Junting Pan, Chengyu Wang, Xu Jia, Jing Shao, Lu Sheng, Junjie Yan, and Xiaogang Wang. Video generation from single semantic label map. In CVPR, 2019. 4", + "[31] Ekta Prashnani, Maneli Noorkami, Daniel Vaquero, and Pradeep Sen. A phase-based approach for animating images using video examples. In Comput Graph Forum, 2017. 3", + "[32] Erik Reinhard, Michael Adhikhmin, Bruce Gooch, and Peter Shirley. Color transfer between images. Comput. Graph. Appl., 21:34-41, 2001. 2", + "[33] Arno Schödl, Richard Szeliski, David H Salesin, and Irfan Essa. Video textures. In SIGGRAPH, 2000. 2", + "[34] Meng Sun, Allan D. Jepson, and Eugene Fume. Video input driven animation (VIDA). In ICCV, 2003. 2", + "[35] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. 2", + "[36] Yongning Zhu and Robert Bridson. Animating sand as a fluid. ACM TOG, 24:965-972, 2005. 3, 6" + ], + "bbox": [ + 501, + 92, + 890, + 895 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15931", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Simulating Fluids in Real-World Still Images/d51a02b7-7fc4-4f85-a456-fd173559877f_model.json b/2023/Simulating Fluids in Real-World Still Images/d51a02b7-7fc4-4f85-a456-fd173559877f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4e5b173ca7b7bc53b02a99e278f3f2f856ba83e0 --- /dev/null +++ b/2023/Simulating Fluids in Real-World Still Images/d51a02b7-7fc4-4f85-a456-fd173559877f_model.json @@ -0,0 +1,2222 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.258, + 0.131, + 0.713, + 0.154 + ], + "angle": 0, + "content": "Simulating Fluids in Real-World Still Images" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.179, + 0.862, + 0.217 + ], + "angle": 0, + "content": "Siming Fan\\(^{1*}\\), Jingtan Piao\\(^{1,3*}\\), Chen Qian\\(^{1}\\), Hongsheng Li\\(^{2,3,4\\boxtimes}\\), Kwan-Yee Lin\\(^{2,3\\boxtimes}\\), SenseTime Research, Shanghai AI Laboratory, The Chinese University of Hong Kong, CPII" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.219, + 0.766, + 0.234 + ], + "angle": 0, + "content": "{fansiming,qianchen}@sensetime.com, 1155116308@link.cuhk.edu.hk," + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.238, + 0.668, + 0.251 + ], + "angle": 0, + "content": "hsli@ee.cuhk.edu.hk, junyilin@cuhk.edu.hk" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.259, + 0.233, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.261, + 0.895, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.43, + 0.893, + 0.46 + ], + "angle": 0, + "content": "Figure 1. Given a still image and a coarse hint (mask and arrows) of motion as input (left), our model estimates fluid motion to generate an animating video (right). Holynski [11] is a state-of-the-art method. Please refer to the supplemental materials for video playing." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.477, + 0.314, + 0.493 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.509, + 0.474, + 0.858 + ], + "angle": 0, + "content": "In this work, we tackle the problem of real-world fluid animation from a still image. The key of our system is a surface-based layered representation, where the scene is decoupled into a surface fluid layer and an impervious background layer with corresponding transparencies to characterize the composition of the two layers. The animated video can be produced by warping only the surface fluid layer according to the estimation of fluid motions and recombining it with the background. In addition, we introduce surface-only fluid simulation, a 2.5D fluid calculation, as a replacement for motion estimation. Specifically, we leverage triangular mesh based on a monocular depth estimator to represent fluid surface layer and simulate the motion with the inspiration of classic physics theory of hybrid Lagrangian-Eulerian method, along with a learnable network so as to adapt to complex real-world image textures. Extensive experiments indicate our method's competitive performance for common fluid scenes and better robustness and reasonability under complex transparent fluid scenarios. Moreover, as proposed surface-based layer representation and surface-only fluid simulation naturally disentangle the scene, interactive editing such as adding objects and texture replacing could be easily achieved with" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.479, + 0.882, + 0.494 + ], + "angle": 0, + "content": "realistic results. Code, and dataset are publicly available." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.523, + 0.633, + 0.538 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.893, + 0.744 + ], + "angle": 0, + "content": "Given an image of a scene containing liquid such as streams and waterfalls, humans can imagine how the liquid in the still image would move. The problem of animating fluid from a single image has become a blooming topic recently studied by a series of work [6, 11, 27, 9, 21]. Their general pipeline can be summarized as: (1) estimating optical flow that indicates the movement of liquid and (2) synthesizing future frame image based on the motion estimation. Although these methods achieve impressive visual effects on simple fluid regions, how to handle complex contents with challenging transparency, collisions, and thin structures in real-world scenes is still a challenging and open problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.744, + 0.895, + 0.88 + ], + "angle": 0, + "content": "The key to the limitation of previous work is that the generation process of videos is modeled in a global manner. Specifically, the subsequent frames are generated based on warping existing textures on input images [3] or neural features [6]. Such operations regard the scene as a whole and warp all the contents in the scene together. For some complex transparent fluid cases, as shown in Figure 2(b), we can see through the water to observe the background rocks beneath the liquid. Textures on those static objects" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.876, + 0.337, + 0.889 + ], + "angle": 0, + "content": "*Equal contribution and joint first authorship." + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.889, + 0.245, + 0.902 + ], + "angle": 0, + "content": "Corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.876, + 0.337, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.53, + 0.889, + 0.811, + 0.902 + ], + "angle": 0, + "content": "Project page: https://slr-sfs.github.io/" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15922" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "are also improperly deformed and result in unnatural animations. To decompose the influence of motion on fluids and surrounding static objects, we propose a two-layer representation for scenes with fluids, namely Surface-based Layered Representation (SLR), which models the motion of the surface fluid and the rest background separately. To achieve such decomposition from the still image during inference, we propose a temporal-based training scheme, which facilitates the network to learn the decomposition with multiple supervisions over spatial and temporal dimensions of running fluid videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.473, + 0.777 + ], + "angle": 0, + "content": "Another major difficulty in animating a single fluid image is predicting the motion flow of the fluid. There exist two types of methods. The first category is the interactive motion prediction methods. They require sparse hints (ie, labelling the velocity direction and relative amplitude) on random pixels of the liquid region as additional inputs. The feature clustering model [27] or convolutional network [11] is then used to generate a dense motion field. As the precision depends on the level of motion details a user can provide, such methods are time-costing in complex fluid scenes. The second category is automatic methods [6]. These works regard the task as a domain transformation problem, and use image-to-image translation techniques to predict the motion represented by 2-channel optical flows. Ambiguous velocity (e.g., a flat river surface can flow either to the left or to the right, as shown in Figure 2(a)) and tremendous changes of velocity in a period of time caused by the collisions with objects (as shown in Figure 2(c)) may pose great challenges to the above methods. To tackle the challenges, we introduce a simplified Navier-Stokes simulation into the motion estimation process, called Surface-only Fluid Simulation (SFS). Specifically, the initial motion prediction of the liquid region is firstly calculated through sparse user inputs and interpolation inside the whole region. In parallel, a 3D mesh is built based on a monocular depth estimator. Then, the N-S equation is solved on the mesh surfaces and based on the initial motion to obtain a simulated motion field in a short period of time. Finally, since the real thickness of the fluid is neglected during the simulation, a refinement step is expected to compensate for the motion offset at height. Thus, a DNN-based motion translator is applied subsequently to obtain the refined motion field that adapts naturally to the input image. With SFS, we can enforce precise controls on the movement of fluid." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.472, + 0.902 + ], + "angle": 0, + "content": "In summary, our work has three major contributions. (a) We propose a new and learnable representation, surface-based layered representation, which decomposes the fluid surface and the static objects in the scene to better synthesize the fluid animating videos from a single fluid image. (b) We design a surface-only fluid simulation to model the evolution of the image fluids with better visual effects. (c) Based on the proposed surface-based layered representation" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.895, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.251, + 0.895, + 0.322 + ], + "angle": 0, + "content": "Figure 2. Motivation. (a) Given a still input image, the motion may be ambiguous. (b) Animating via previous SOTA method [11] causes background rocks beneath the liquid improperly deformed. (c) Motion prediction methods such as [11, 27] cannot well capture motion changes caused by the collisions with objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.895, + 0.456 + ], + "angle": 0, + "content": "and surface-only fluid simulation, image editing with fluids is achieved with realistic and vivid results. In addition, We can easily extend our fluid animation system to downstream editing tasks, thanks to the SLR and SFS designs in our framework. For example, we can augment objects in the fluid region or background region, and simulate the interactions between the fluids and objects. We can also replace textures to different layers to create different scenes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.472, + 0.644, + 0.487 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.498, + 0.673, + 0.513 + ], + "angle": 0, + "content": "2.1. Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.704 + ], + "angle": 0, + "content": "In early studies, researchers focus on generating new video context by changing the attributes of recorded frames [32, 33]. Traditional methods of video motion generation with movable substances are based on stochastic process analysis [34] and color transfer [32]. Given a reference video indicating the temporal changes of a scene, texture movement and sequential images can be generated using a patch aligned method and energy-based optimizing [33]. Similar methods are applied to single image input [3]. However, obvious artifacts exist, and the smoothness of the video is not guaranteed. These drawbacks make the usage of such methods limited to typical types of scene samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.902 + ], + "angle": 0, + "content": "In recent years, since the development of deep neural networks in image generation field [15], plenty of innovative efforts have been made to improve the synthesis quality of images [7, 16, 8, 35, 25]. Techniques such as GANs [8], perceptual loss [17], partial convolution [25] and cycle-consistency [35] have been widely used as standard components. Built upon the success of image generation, several attempts have created the precedents of neural video generation. A few methods are developed to force the network to learn intermediate scene representation for predicting the future frames or interpolating key frames, once the input sequences and camera trajectory are given (note: as we focus on single image input, we skip the review of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15923" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.364 + ], + "angle": 0, + "content": "such multi-frames input trend). Given a single image as input, latent generative methods like [24] proposes an autoregressive model to generate infinite pixels forming sequential outputs. InfinityGAN [23] disentangles global appearances, local structures and textures with a structure synthesizer and texture synthesizer to generate images with spatial size and level of details not attainable before. [31] tries to use Fourier transforms in time variance showing the phase variations to represent the flow and generate moving videos. [20] spends efforts in digging into the semantic meanings in images and generates videos correspondingly to achieve a natural result. [6] decouples motion and appearance through learning intermediate flow fields and color transfer map. However, these methods are either focusing on spatial-wise consistency [20], or paying attention to coarse-level time-variance scene changes [31, 6]. None of them discusses or can handle the particularity of fluid with fine detail motions." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.377, + 0.245, + 0.392 + ], + "angle": 0, + "content": "2.2. Fluid Simulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.473, + 0.688 + ], + "angle": 0, + "content": "Physical Fluid Simulation. Traditional methods originated from computer graphics to achieve accurate approximations on various types of fluids. Hybrid Lagrangian-Eulerian Methods [2, 36, 10] have successfully simulated the motion of incompressible fluids in a pre-defined 2D or 3D box environment, which solves pressure projection on Eulerian grid and advect on Lagrangian particles. To improve convergence ability and flexibility on complex shapes, material point method (MPM) [1] has been popular and more subsequent methods [18, 19] have been proposed to stabilize the convergence when iterating velocities. However, these methods are time-consuming in large scenes. To reduce the cost of simulating a large-range of fluid, methods have been proposed to simulate only a small range of fluid [4], or the surface of the fluid [5, 13] to produce a faster approximation by avoiding high computational cost in 3D volumetric solvers. Although these works achieve amazing simulation effects, they are still limited to synthetic scenarios due to scene-specific modelling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Animating Fluids with Data Prior. Given a still image as animation target and a set of real-world water video candidates as animation bank, [31, 29] propose to retrieve a video or a group of videos from the bank that has similar content with the target image as the references. Then they transfer motion and appearance of references to each region of interest on target image to form the water animation video. These methods open the door of real-world fluid animation with data prior in-the-wild. However, the seamless alignment of animated results with source input image can not be well guaranteed and requiring lots of manual efforts. With the development of deep learning, recent works [11, 27] propose to automatically predict each frame of motion and appearance via learning from reconstruction loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.333 + ], + "angle": 0, + "content": "with real world videos for training. [11] first simplifies motion estimation part as a single frame motion prediction via motion Eulerian integration and then proposes a deep feature warping technique to narrow down the size of blank areas caused by warping. Finally, it adaptively blends forward and backward features through a learnable parameter \\( Z \\) to generate animated fluid video. Based on [11], [27] proposes to regress fluid motion conditioned to user's sparse guidance and generate paired training data through motion speed clustering. The work further proposes to use multiscale representation to capture different fluid speed in different resolution. Although these methods take a step further to real-world scenes with impressive results, it is hard for them to handle complex context relations over fluids due to the single representation to the entire scene that regards different objects and textures as equal." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.897, + 0.635 + ], + "angle": 0, + "content": "Unlike previous work, our system marries the advantages of both physical simulation and learning-based pipelines. For textures, to avoid interplay between fluids and surroundings, we force the network to decouple the scene into a new two-layer representation SLR, that includes a static background and time-variates surface fluid layers2. Meanwhile, the network can hallucinate reasonable and photo-realistic textures in vacated regions thanks to the large-scale data prior. For motion, to skirt scene-specific modelling while keeping physiC reasonability, we propose a three-step motion prediction. We first use sparse labels from the human interface to generate the initial rough motion and then calculate reasonable evolution in a short period by introducing a 2.5D surface simulation, SFS. Later, a smooth motion translator is used to refine the motion trend to better fit real-world examples with data prior. Moreover, as a consequence, we can generate natural fluid animations with the network training on plenty of real-world examples and flexibly edit the main liquid region and boundary condition to create various visual effects." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.648, + 0.593, + 0.664 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.895, + 0.856 + ], + "angle": 0, + "content": "Given a real-world image that contains liquid, our goal is to generate a video clip that animates vivid flowing of the fluid. To achieve this goal, a system is required to (1) represent complex context relations between fluids and surrounding environments; and (2) to estimate motion fields from single image input with reasonability and efficiency. We propose a novel system with two main designs that satisfy the above requirements. For texture, in contrast with previous methods that regard all elements in the scene as an entirety, our system learns a surface-based layered representation (SLR) (Sec. 3.1) that decomposes the scene into surface fluid layer and impervious background layer. Since the tex" + }, + { + "type": "page_footnote", + "bbox": [ + 0.499, + 0.863, + 0.895, + 0.903 + ], + "angle": 0, + "content": "2This idea shares a similar spirit with [26] in separating a scene into both static and dynamic parts. Whereas, our SLR could handle transparent fluids, and only require a single image as input." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "15924" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "ture characteristics of surface fluid and rest background are assigned to different layers, such a design could overcome the negative influence of motion flows between fluids and surroundings (e.g., improper texture warping, as shown in Figure 2(b)). As for motion estimation, the system provides a three-step motion estimation alternative by introducing a surface-only fluid simulation (SFS) into the motion field prediction (Sec. 3.2). The motion prediction steps combine advantage of both physical- and learning-based motion estimation pipelines, thus ensuring physical reasonability and meanwhile skirting the scene-specific modelling problem. Each component is detailed in following subsections." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.282, + 0.403, + 0.299 + ], + "angle": 0, + "content": "3.1. Learning of Decomposition of Images" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.47, + 0.41 + ], + "angle": 0, + "content": "In this subsection, we first introduce Surface-based Layered Representation (SLR). Then, we show how to obtain animated videos through SLR. Finally, we present how to learn SLR with a neural network (illustrated in Figure 3), and challenges we meet during optimizing the network since there is no oracle ground-truth supervision for fluid scene decomposition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.47, + 0.712 + ], + "angle": 0, + "content": "Surface-based Layered Representation. Given a still image \\( I(T_0) \\) as input, a series of RGB frames \\( \\mathbf{I} = \\{I(T_0),\\dots,I(T_n)\\} \\) are outputs to synthesize time-space variations of the scene with running fluids. As we aim to handle complex context of fluid scene through disentangling fluid motion and still objects beneath/interacting with surface fluid, we propose to learn a novel intermediate representation whose output is an impervious background layer image \\( I_b \\), a temporal sequence of surface fluid layer images \\( \\mathbf{I}_f = \\{I_f(T_0),\\dots,I_f(T_n)\\} \\) and transparency factors \\( \\alpha_{b},\\alpha_{f} = \\{\\alpha_{f}(T_{0}),\\dots,\\alpha_{f}(T_{n})\\} \\), where \\( \\alpha_{*}\\in [0,1] \\). Transparency factors indicate the contribution of colors on each layer to the final image at each time. In practice, we utilize a convolution network to output the simplification of the representation: \\( F_{\\theta}: (I(T_0),I(T_n),\\mathbf{M})\\to (I_b,I_f(T_0),I_f(T_n),\\alpha_b,\\alpha_f(T_0),\\alpha_f(T_n)) \\), where the network only outputs the estimated first and last frames of \\( \\mathbf{I}_f \\) as well as \\( \\alpha_{f} \\). \\( \\mathbf{M} \\) is the estimated motion fields of \\( T_0 \\) to \\( T_{1} \\) from the single input \\( I(T_0) \\). We will detail the synthesis of \\( \\mathbf{M} \\) in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.469, + 0.825 + ], + "angle": 0, + "content": "We expect the background to be identical across all times, while the fluid surface deforms with realistic motions. To this end, \\( F_{\\theta} \\) is enforced to automatically map the original texture of non-liquid and static regions as well as the reasonable texture of under-liquid surface regions to the background layer (detailed in later subsection). As for fluid surface learning, recurrent estimation for all frames (RNN) is a possible alternative that can be tailored from video gen" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.285 + ], + "angle": 0, + "content": "eration [30]. However, such an intuitive design might cause frame distortion as time goes on due to the accumulated errors and hard convergence (also demonstrated in [11]). In contrast, We use the linear combination of the first and last frames to construct the images at intermediate frames. This choice not only transforms the problem into interpolation rather than extrapolation, which is easier to converge, but also is able to generate an endless cyclic video rather than a video of a fixed interval. In detail, we tailor the symmetric softmax splatting, which is a pixel-warping strategy utilized in [11, 28], to our framework. We extend the splatting operation from a single image feature splatting into two levels – fluid surface feature splatting and fluid alpha splatting as follows :" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.29, + 0.892, + 0.323 + ], + "angle": 0, + "content": "\\[\nD _ {f} \\left(T _ {i}; \\hat {\\mathbf {x}} ^ {\\prime}\\right) = \\frac {\\sum_ {\\hat {\\mathbf {x}} \\in \\mathcal {X}} \\alpha \\left(T _ {i} ; \\hat {\\mathbf {x}}\\right) \\cdot D _ {f} \\left(T _ {j} ; \\hat {\\mathbf {x}}\\right) \\cdot \\exp \\left(Z _ {f} \\left(T _ {j}; \\hat {\\mathbf {x}}\\right)\\right)}{\\sum_ {\\hat {\\mathbf {x}} \\in \\mathcal {X}} \\alpha \\left(T _ {i}; \\hat {\\mathbf {x}}\\right) \\cdot \\exp \\left(Z _ {f} \\left(T _ {j}; \\hat {\\mathbf {x}}\\right)\\right)}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.325, + 0.892, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {f} ^ {\\prime} \\left(T _ {i}; \\hat {\\mathbf {x}} ^ {\\prime}\\right) = \\frac {\\sum_ {\\hat {\\mathbf {x}} \\in \\mathcal {X}} \\alpha \\left(T _ {i} ; \\hat {\\mathbf {x}}\\right) \\cdot \\alpha_ {f} ^ {\\prime} \\left(T _ {j} ; \\hat {\\mathbf {x}}\\right) \\cdot \\exp \\left(Z _ {f} \\left(T _ {j} ; \\hat {\\mathbf {x}}\\right)\\right)}{\\sum_ {\\hat {\\mathbf {x}} \\in \\mathcal {X}} \\alpha \\left(T _ {i} ; \\hat {\\mathbf {x}}\\right) \\cdot \\exp \\left(Z _ {f} \\left(T _ {j} ; \\hat {\\mathbf {x}}\\right)\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.892, + 0.411 + ], + "angle": 0, + "content": "where \\( j \\in \\{0, n\\} \\) is the start index or last index of sequences, indicating forward splatting the frame 0 or backward splatting the frame \\( n \\), which means the value \\( j \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.431, + 0.892, + 0.465 + ], + "angle": 0, + "content": "\\[\nj = \\left\\{\\begin{array}{l l}0,&\\text {i f} \\hat {x} ^ {\\prime} = M _ {0 \\rightarrow t} (\\hat {x}),\\\\n,&\\text {i f} \\hat {x} ^ {\\prime} = M _ {0 \\rightarrow t - N} (\\hat {x}).\\end{array}\\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.892, + 0.64 + ], + "angle": 0, + "content": "\\(D_{f}(T_{i};\\hat{\\mathbf{x}}^{\\prime})\\) is the fused fluid feature value of target frame at pixel \\(\\hat{\\mathbf{x}}\\), and \\(\\alpha_f^\\prime (T_i;\\hat{\\mathbf{x}}^\\prime)\\) is the transparency value of target frame at pixel \\(\\hat{\\mathbf{x}}\\). \\(D_{f}(T_{i};\\hat{\\mathbf{x}})\\) represents value of \\(D_{f}(T_{i})\\) at pixel \\(\\hat{\\mathbf{x}}\\), \\(\\alpha_{f}(T_{i};\\hat{\\mathbf{x}})\\) and \\(\\alpha_f^\\prime (T_i;\\hat{\\mathbf{x}})\\) are similar. During inference, we set \\(D_{f}(T_{n}) = D_{f}(T_{0})\\) and \\(\\alpha_{f}(T_{n}) = \\alpha_{f}(T_{0})\\) since only single image is available. \\(\\mathcal{X}\\) is the set of pixels satisfying \\(M_{0\\rightarrow t}(\\hat{\\mathbf{x}}) = \\hat{\\mathbf{x}}^{\\prime}\\) or \\(M_{0\\rightarrow t - N}(\\hat{\\mathbf{x}}) = \\hat{\\mathbf{x}}^{\\prime}\\), \\(M_{0\\rightarrow t}\\) is displacement map generated from \\(\\mathbf{M}\\) through euler integration [11, 28], and \\(\\alpha (T_i;\\hat{\\mathbf{x}}^{\\prime})\\) is a temporal coefficient to fuse features from forward warping and backward warping to ensure endless looping:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.648, + 0.892, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\alpha \\left(T _ {i}; \\hat {\\mathbf {x}} ^ {\\prime}\\right) = \\left\\{\\begin{array}{l l}\\frac {t}{n},&\\text {i f} \\hat {x} ^ {\\prime} = M _ {0 \\rightarrow t} (\\hat {x}),\\\\1 - \\frac {t}{n},&\\text {i f} \\hat {x} ^ {\\prime} = M _ {0 \\rightarrow t - n} (\\hat {x}).\\end{array}\\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Then a fluid feature decoder is applied to \\( D_{f}(T_{i}) \\) to generate surface fluid image \\( I_{f}(T_{i}) \\) and a alpha refinement network is applied to \\( \\alpha_f'(T_i) \\) to generate transparency alpha \\( \\alpha_{f}(T_{i}) \\) to avoid vacant alpha particles." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Animating Fluids with SLR. Once we obtain the final surface fluid layer \\( I_{f}(T_{i}) \\) and transparency \\( \\alpha_{f}(T_{i}) \\), the reconstructed final image \\( I(T_{i}) \\) can be acquired by adding the impervious background layer \\( I_{b} \\) back to the surface fluid layer \\( I_{f}(T_{i}) \\) with \\( \\alpha_{b} \\). Here we use separate \\( \\alpha \\) channels for background and fluid transparency predictions. This design helps to stabilize the training of decomposition. The core reason is that when a single fluid transparency \\( \\alpha_{f} \\) is warped according to the motion flow, it leaves a blank region on the boundary, and makes the gradients hard to be" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.839, + 0.47, + 0.901 + ], + "angle": 0, + "content": "3The word \"impervious\" is not rigorous, as background layer is not only formed from static object but also non-surface fluid which contains air molecules in real environment. However, we assume the fluid is incompressible, and thus a little bit abuse the word \"impervious background\" to indicate the non-surface fluid regions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15925" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.086, + 0.895, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.404, + 0.895, + 0.463 + ], + "angle": 0, + "content": "Figure 3. Training and testing pipelines. The background layer and background alpha are predicted only at the first frame. Both fluid features and fluid alpha are predicted at the first and last frames, and then symmetricly splatted and blended using the softmax splatting at frame \\( T_{i} \\). A decoder is used to decode fluid features into images and refined alpha. For frame \\( T_{0} \\), motion is calculated following pipeline Figure 4, and for frame \\( T_{i} \\), motion is calculated through Euler integration. During inference, \\( T_{n} \\) is replaced by \\( T_{0} \\) to create a cyclic video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.471, + 0.471, + 0.518 + ], + "angle": 0, + "content": "backpropagated. Although with the two-channel implementation, there still is a \\(\\alpha_{b}\\) on that vacated regions to fascinate the network refining the texture. In formula we have" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.528, + 0.47, + 0.562 + ], + "angle": 0, + "content": "\\[\nI \\left(T _ {i}\\right) = \\frac {\\alpha_ {f} \\left(T _ {i}\\right) I _ {f} \\left(T _ {i}\\right) + \\alpha_ {b} I _ {b}}{\\alpha_ {f} \\left(T _ {i}\\right) + \\alpha_ {b}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.689 + ], + "angle": 0, + "content": "For corresponding network architectures to achieve the above targets, as shown in Fig 3, the encoder is used to extract background texture and \\(\\alpha\\) channel as well as surface fluid layers' features and its \\(\\alpha\\). The decoder is used to refine the fused fluid layer that combines the layers from the first and last frames to obtain the complete surface fluid layer, and the final animated image is reconstructed by compositing background and surface fluid layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Optimizing the SLR. The training process is divided into three parts, in which, we separately train (1) the surface fluid branch containing the encoder that predicts the fluid feature layer from image input and the decoder that reconstructs the warped fluid layer, (2) the background branch that predicts the background layer from the image input, and (3) jointly train them together with \\(\\alpha\\) learning and surface fluid/ background branches finetuning. For each training step, the paired training data \\(< (I(T_0), I(T_n)), I(T_i) >\\) are randomly picked among 60-frames sequence without constraint on interval. We use dense optical flow estimation network [14] to extract reasonable movement of each video as pseudo motion ground-truth, and to ensure smoothness, we average the flow with window size of 30 frames. Due" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.471, + 0.892, + 0.503 + ], + "angle": 0, + "content": "to space limits, we detail the training of each stage in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.725, + 0.529 + ], + "angle": 0, + "content": "3.2. Motion Field Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.597 + ], + "angle": 0, + "content": "In the previous section, we have described the core necessities to learn an SLR but skipped the synthesis process of one of its input, motion fields \\(\\mathbf{M}\\). In this section, we discuss how to obtain \\(\\mathbf{M}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.893, + 0.901 + ], + "angle": 0, + "content": "During training, the motion of each frame can be generated from a conventional optical flow estimator [14]. During inference, as the input is a single image, we expect to synthesize a reasonable motion map from the observed scene. We propose a motion estimation process that separates the motion prediction into three parts taking advantage of both physical-based and learning-based pipelines for fluid simulation. First, interactive labelling of the sparse velocity of several pixels on the image is required to guide the generated motion directions of \\( I(T_0) \\), along with a fusion with weights calculated based on Gaussian filtering of the distance maps from the labeled pixel to build a dense velocity initial map. Second, a surface-only fluid simulation is calculated based on initial motion and estimated scene depth to simulate a following convincing evolution of the liquid. Note that this step only calculates for the first frame. Last, a DNN-based pixel-to-pixel motion translator is utilized to refine the flow to create a more realistic and natural motion map. This translator could be considered as generating a compensation for a detailed motion map that is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15926" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.096, + 0.115, + 0.377 + ], + "angle": 270, + "content": "6uizisauus 6uizisauus pue uonnnsr (e)" + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.09, + 0.296, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.162, + 0.245, + 0.171 + ], + "angle": 0, + "content": "1. Input Image" + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.09, + 0.469, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.163, + 0.453, + 0.171 + ], + "angle": 0, + "content": "2. Label Mask and Velocity Hint" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.173, + 0.237, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.174, + 0.353, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.174, + 0.469, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.218, + 0.227, + 0.225 + ], + "angle": 0, + "content": "3. Generate Dense Velocity" + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.218, + 0.34, + 0.224 + ], + "angle": 0, + "content": "4. Mono-Depth Estimation" + }, + { + "type": "image_caption", + "bbox": [ + 0.386, + 0.218, + 0.444, + 0.224 + ], + "angle": 0, + "content": "5. Build 3D Mesh" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.226, + 0.237, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.273, + 0.222, + 0.281 + ], + "angle": 0, + "content": "6. Mesh-Based Simulation" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.226, + 0.353, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.273, + 0.325, + 0.281 + ], + "angle": 0, + "content": "7. Refine Velocity" + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.226, + 0.468, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.273, + 0.447, + 0.28 + ], + "angle": 0, + "content": "8. Synthesize Result" + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.282, + 0.237, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.326, + 0.215, + 0.333 + ], + "angle": 0, + "content": "9. Editing AR objects" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.282, + 0.352, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.247, + 0.326, + 0.339, + 0.333 + ], + "angle": 0, + "content": "10. Add object in Fluid Layer" + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.282, + 0.468, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.326, + 0.466, + 0.333 + ], + "angle": 0, + "content": "1. Add object in Background layer" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.334, + 0.237, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.375, + 0.234, + 0.381 + ], + "angle": 0, + "content": "12. Mark Intersection Vertex \"Solid" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.334, + 0.353, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.376, + 0.345, + 0.382 + ], + "angle": 0, + "content": "Mesh-Based Simulation" + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.334, + 0.468, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.376, + 0.458, + 0.382 + ], + "angle": 0, + "content": "14. Edited Synthesized Result" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.392, + 0.47, + 0.476 + ], + "angle": 0, + "content": "Figure 4. Motion calculation pipeline. (a) is initialization of fluid region, rock region, motion, and Mesh. (a,b) combine to form our 2.5D simulation and animation pipeline, where (4-5) is not included if using 2D simulation method. (a,c) combine to form our editable simulation and synthesis pipeline, where the user can edit 3D objects and create corresponding motion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.469, + 0.514 + ], + "angle": 0, + "content": "harmonic with fluid textures. The pipeline is illustrated in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.514, + 0.469, + 0.62 + ], + "angle": 0, + "content": "Initial Motion from Interactive Sparse Labeling. To obtain determined as well as editable motion direction, we choose an interactive alternative to generate initial flow for image \\( I(T_0) \\) following [27]. With no more than 10 discrete notations on images, and a fixed segmentation of the liquid region, we use nearest-neighbors-averaging to assign all the pixels that are inside the liquid region an initial velocity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.469, + 0.831 + ], + "angle": 0, + "content": "Surface-only Fluid Simulation. After the initial state is estimated, we tailor the traditional graphics pipeline with discretization and linear equation to simulate the follow-up evolution of the liquid. Intuitively, we can directly apply simulation on 2D grids that align with image pixels by initializing each pixel as an Eulerian grid and several Lagrangian particles near each pixel. However, since such perspective projection of the 3D scene may cause severe distortion when direct manipulating the motion in 2D, as shown in Figure 7, we thus turn to use a mono-depth estimator [22] to estimate the 3D surface of the liquid. After the 3D surface is constructed inside the region of the liquid, a mesh surface-based velocity is calculated, with the restriction of projected velocity:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.83, + 0.469, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\frac {d}{d t} \\left[ \\begin{array}{l} u \\\\ v \\end{array} \\right] = \\left[ \\begin{array}{c c} f x & 0 \\\\ 0 & f y \\end{array} \\right] \\left[ \\begin{array}{c c c} 1 / z & 0 & - x / z ^ {2} \\\\ 0 & 1 / z & - y / z ^ {2} \\end{array} \\right] \\left[ \\begin{array}{c} \\frac {d x}{d t} \\\\ \\frac {d y}{d t} \\\\ \\frac {d z}{d t} \\end{array} \\right] \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Where \\( fx, fy, cx, cy \\) is camera intrinsic parameters, \\( u, v \\) is" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.64, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.09, + 0.763, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.09, + 0.892, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.147, + 0.641, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.146, + 0.726, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.146, + 0.892, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.197, + 0.726, + 0.203 + ], + "angle": 0, + "content": "(b)Not a Surface" + }, + { + "type": "image_caption", + "bbox": [ + 0.794, + 0.2, + 0.862, + 0.205 + ], + "angle": 0, + "content": "(c)Complex Initialization" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.207, + 0.892, + 0.263 + ], + "angle": 0, + "content": "Figure 5. Examples for Different Fluid Scenes. (a) The surface thickness of the waterfall and river are gauzy. (b) The fountain cannot be considered a surface. (c) Fluid with vortex contains complex motion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.267, + 0.892, + 0.584 + ], + "angle": 0, + "content": "the velocity on the projected image and \\( x, y, z \\) the 3D position in the scene. Since we have modelled the liquid as a surface, we replace traditional Laplacian operator with Laplacian-Beltrami operator on the mesh and use Mixed-Euler-Lagrange approach [36] implemented in Taichi [12] to simulate consequent motion of each vertex. Detailed formula derivation can be seen in the supplementary. The particle binding on the surface is randomly sampled according to the surface area. When a particle is moved outside the mesh surface, we use the nearest projection to force the position to fall onto the mesh. For the boundary of the mesh, we calculate the velocity of the boundary triangle, judging whether the velocity on the edge boundary is inward or outward. For outward edges, the particle is abandoned after moving out of the liquid region. For inward region, a particle source is built to maintain the overall number of particles to be roughly stable. With the texture of surface fluid layer binding to each vertex, we can warp fluid layer with a soft rasterizer to simulate flows. In this way, given arbitrary flow initial state, a stable velocity can be calculated to approximate real-world flow after a period of simulation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Smooth Motion Translator. With the help of above \\(2.5D\\) fluid simulation, we can generate a time-variant velocity among the sequence. However, the simulated velocity is still a rough trend since it only works on fluid surface without considering thickness variation of fluids. Thus, the SFS can work well in cases like Figure 5(a). It fails in cases with distinct depth measurement differences, as shown in the Figure 5(b) due to the lack of height information during calculation. Also, It is fragile to cases with complex initial state as shown Figure 5(c) due to error-prone/time-consuming interactive initial motion labeling for such cases. Thus, we need extra scheme to compensate or rectify for detailed motion that is harmonic with the fluid textures. To this end, we use a convolution-based network to transfer the simulated motion into a more detailed and reasonable one, which is pixel-aligned with image textures. The network is trained as a traditional style-transfer task, with a L2 loss between output velocity and ground truth. An external patch discriminator is used to refine the pattern of the velocity distribution. All the settings are similar to the image translation task pipeline [15]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15927" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.09, + 0.21, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.192 + ], + "angle": 0, + "content": "In this section, we present experiments and detailed analysis to demonstrate the contributions of our method both quantitatively and qualitatively. For all the cases shown in this paper, please refer to our demo video and project page for better dynamic visual effects." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.2, + 0.363, + 0.216 + ], + "angle": 0, + "content": "4.1. Quality of Fluid Video Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.224, + 0.469, + 0.299 + ], + "angle": 0, + "content": "Implementation Details. To learn SLR, we train our network on a series of videos containing a large region of fluid movement and still background [11]. For training of motion refinement network, we use the same settings in [27]. All networks are trained at a resolution of \\(256 \\times 256\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.299, + 0.47, + 0.496 + ], + "angle": 0, + "content": "Datasets and Evaluation Metrics. In addition to validation set in [11], we collect extra 126 scenes to form a new test set, which we called Complex Liquid Animation in-the-Wild (CLAW) test set, for evaluating the performance on both dense fluids (such as rivers and oceans) and sparse or transparent fluids (such as streams or splashes). Note that, in this subsection, the motion for all the methods to be compared is guided by the optical flow of the whole video to distinguish the performance of the synthesizing techniques. We use PSNR to show the whole average error, SSIM to show the errors in the region with plenty of textures and LPIPS, a perceptual based loss, to show visual quality difference. For more details, please refer to supplementary." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.5, + 0.469, + 0.58 + ], + "angle": 0, + "content": "
DatasetMethodsAll RegionFluid Region
LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑
HolysnikReproduced Holysnik0.079825.030.77870.065725.880.8007
CommonModified Holysnik(Baseline)0.079324.750.77580.065625.720.8000
Val SetOurs0.083425.140.77950.065726.100.8030
OurReproduced Holysnik0.206720.260.59550.202920.360.5961
CLAWModified Holysnik(Baseline)0.207819.970.59230.204120.100.5934
TestsetOurs0.204020.790.60800.197520.800.6077
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.58, + 0.469, + 0.661 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison. (a) Results on Holynski's common val set [11], evaluating first to 60-th frames. \"Fluid Region\" means static region is replaced by input image during metric calculation for focusing on the statistic of fluid. This could be done by replacing the non-fluid region with the corresponding textures of input images. (b) Results on CLAW test set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.668, + 0.47, + 0.821 + ], + "angle": 0, + "content": "Quantitative Comparisons with baseline. Table 1 shows the results between our method and state-of-the-art method [11], which is a single layer learning-based system, a typical method that animates the scene in a global manner. We re-implement [11] in Pytorch (Reproduced Holynski), since the source code is not public available. For other comparison baselines, Modified Holynski is built upon Reproduced Holynski but with the modification of replacing convolution to partial convolution in the fluid decoder. The third model (Ours) is our SLR model4. Three observations can" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.828, + 0.47, + 0.902 + ], + "angle": 0, + "content": "4For Ours, we use (Ours (stage 1) in Table 1 in our supplementary material.) as surface fluid model initialization. Since our training contains three stages, and we train 100 epochs at the first stage to get the model (Ours (stage 1)), then decay learning rate and fine-tune 50 epochs for fluid at the last stage in practice, we also apply the same training strategy for the first two baselines to ensure fair comparisons." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.729 + ], + "angle": 0, + "content": "be included from Table 1: (1) PSNR and SSIM metrics of \"Modified Holynski\" drop a little compared to \"Reproduced Holynski\" in both statistics on all- and fluid region. We infer that, although partial convolution could better capture context than conventional one to impart reasonable texture in vacated regions in theory, simply replacing partial convolution into single-layer framework cannot essentially improve the image quality due to the wrong context raised from improper warping of both background and fluid. While, with SLR, the partial convolution can strengthen its advantage on context capturing for more proper fluid context textures provided in surface fluid layer. (2) The LPIPS of Ours model drops a little compared to the Modified Holynski model in the \"All Region\" of Holynski common validation set, but has nearly the same LPIPS in the \"Fluid Region\". For the other two metrics, our model leads to an uptick. Since most regions of Holynski's validation set are static, the blending of the fluid and background in the non-fluid region could lead to slight performance drops in terms of LPIPS. However, it could be eased by replacing the non-fluid region with the corresponding textures of input images, thus we care more about the performance in the \"Fluid Region\" instead of the \"All Region\". Such a phenomenon tells us that the two-layer decomposition design in Ours model does not harm the reconstruction ability on final images. To better unfold the analysis, we further conduct the close-up evaluations on Holynski's val set. Concretely, we crop the fluid region by motion threshold and only evaluate these visible fluid background regions. The LPIPS, PSNR, SSIM of Holynski's method are 0.230, 19.01, 0.565, and Ours are 0.227, 19.64, 0.581. This setting leads to larger performance margins with the baseline, which demonstrate the advantage of our system. (3) When evaluating only in the fluid region, we can see model Ours reaches a comparable result with the \"Modified Holynski\" model in Holynski validation set [11], but improves significantly in CLAW test set. The major reason behind this observation is that when it comes to complex scenes as included in our CLAW test set, our proposed two-layer representation suppresses the influence between surrounding and surface fluids. Thus we can achieve better results than single-layer baselines that regard all elements in the scene as an entirety during warping." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Comparisons with SOTA methods. In this paragraph, we provide comparisons with more state-of-the-art methods - Endo et.al [6], Tavi et.al [9], and Mahapatra et.al [27]. The first two methods are centered on animating general scenes, where Endo et.al [6] is designed for landscape, and Tavi et.al [9] focuses on animating periodic patterns in still images. These two methods could be applied to fluid scenarios. In contrast, Mahapatra et.al [27] is specialized for fluid scenarios, where a multi-scale representation is introduced to capture different fluid speeds in different resolutions. Concretely, as Endo's code is publicly available, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15928" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.424 + ], + "angle": 0, + "content": "compare our system with it both quantitatively and qualitatively. For other methods, we compare them qualitatively in Fig 6. Specifically, the result of Mahapatra is from official project page, and result of Tavi is from official app. Endo's method can only control motion with a latent code. Thus, we test ten different latent codes and select the best one to compare. We observe that these three methods suffer from animation distortions with different levels – (1) from the zoom-in perspective, all baselines improperly move the still stones with the fluids. Specifically, Tavi's results have significant ghosting artifacts, which might be caused by its one-fold periodic motion and alpha-blending designs. (2) From the full image perspective, Endo's method has difficulty in estimating fluid motion. The river's motion is incorrectly predicted to be from right to left, and the waterfall flows upwards to the right. The quantitative results (Tab 1) provide the demonstration from another aspect. This phenomenon is foreseeable, as Endo's framework considers motion in a global manner. In contrast, our two-layer method animates plausible fluid without affecting still stones. Please refer to the supplementary material for more comparisons and straightforward video effects." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.427, + 0.277, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.427, + 0.47, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.609, + 0.47, + 0.651 + ], + "angle": 0, + "content": "Figure 6. Qualitatively Comparisons. (e-f) are tested in \\(288 \\times 512\\), others are tested in \\(\\mathrm{W} = 1280\\). LPIPS/PSNR/SSIM of the whole video are colored in orange." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.658, + 0.469, + 0.702 + ], + "angle": 0, + "content": "
MethodsHolynski validation setCLAW test set
All RegionFluid RegionAll RegionFluid Region
LPIPS ↓PSNR ↑SSIM ↑LPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIM
Endo[6] HD0.35716.270.3510.12022.310.7410.35715.920.2960.34716.380.331
Ours HD0.14223.430.7150.11823.780.7280.19620.990.6110.19020.990.610
" + }, + { + "type": "table_caption", + "bbox": [ + 0.099, + 0.701, + 0.449, + 0.713 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons with Endo's method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.871 + ], + "angle": 0, + "content": "User Study. We conduct a user study to compare visual effects between single layer baseline (Modified Holynski) and our two layers method(Ours in Table 1) on the full set of CLAW testset. For each scene, we animate 60 frames and output a two-second video. Animated videos are shown side by side without providing ground truth videos as reference to guide the participant focusing on the reasonability and photo-realistic. We ask the participants to select which one is better or neither is better, mainly according to three dimensions (photo-realistic, stereoscopic, and high-fidelity)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.9 + ], + "angle": 0, + "content": "As shown in Table 3, the participants can make a distinction between the two methods over most samples, leav" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.7, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.089, + 0.895, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.232, + 0.894, + 0.316 + ], + "angle": 0, + "content": "Figure 7. Different simulations. Orange arrows point out motion direction. All simulation methods use \\(75\\%\\) weak incompression simulation instead of fully incompression for better visual effect. (row 1)Input image, sparse motion generated by sparse user hint required by each method. (row 2-4)ablation study of Mahapatra's method [27], 2D simulation and 2.5D SFS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.893, + 0.393 + ], + "angle": 0, + "content": "ing \\(27.2\\%\\) samples difficult to tell. Our method achieves better scores than the single-layer method [11] with \\(52.3\\%\\) sample voting. This indicates the effectiveness of the proposed SLR in other aspects." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.397, + 0.86, + 0.428 + ], + "angle": 0, + "content": "
Single layerOursNeither
E-O-N125(20.5%)319(52.3%)165(27.2%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.43, + 0.893, + 0.498 + ], + "angle": 0, + "content": "Table 3. User study. The subjective scores are voted on all samples of our CLAW testset, which includes 122 videos with various real-world fluids scenarios(e.g., river, stream and waterfall) and complex context surrounding (e.g., semi-/transparent liquid, collisions and thin structures)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.515, + 0.75, + 0.53 + ], + "angle": 0, + "content": "4.2. Different Motion Prediction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Figure 7 shows different motion prediction pipelines. Pipeline 1: An interactive sparse labelling of the motion and then extending all the velocities to the whole moving region, which results in a constant flow that remains large gap with realistic motion. We can see that the motion appears reasonable global with a network refinement. However, for local regions around static rocks, the flow of liquid is improperly stiff and has no interactions with the rock. While in real-world scenes, there should be vortex and sprays caused by collisions. Pipeline 2: Built upon pipeline 1, while before the motion refinement stage, we additionally initialize grids with fluid, rock or air masks as well as placing particles around each grid, then advocate particles by edited motion for once. Later, a NS equation with the incompressible fluid assumption is solved on 2d grid to obtain the motion of the next frame, which we call \"Simulated 2D Motion\" in the figure. Considering pipeline 2 is simulating on a 2D image plane, a refinement step (using a smooth motion translator) is applied to compensate for the motion offset at height and fine details. Pipeline 3: instead of simulating on a 2D image plane, the N-S equation is solved in \\(2.5D\\). Specifically, we regress the depth with a pretrained monocular depth estimator and consequently build a 3D mesh over the fluid. Then, we solve the N-S equation on the mesh surface. With" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15929" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.468, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.206, + 0.47, + 0.247 + ], + "angle": 0, + "content": "Figure 8. Interactive Editing. (Bottom) Initial all moving region as 'Fluid' statue. (Top) Initial 'Solid' statue at overlap area of mesh and new object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.468, + 0.411 + ], + "angle": 0, + "content": "the help of simulation, we can see in the left scene of the figure that the effect of turbulence around the rock can be animated, and the trend maintains the refined output. From the right scene in the figure, we can observe that the absolute value of the velocity is unnatural in Pipeline 2, which regards all particles on an identical plane. However, since the liquid surface leans in the picture, there should be a foreshortening effect on the projection. In contrast, pipeline3 samples particles on the surface and such simulation enable near-plane motion to carry more speed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.468, + 0.516 + ], + "angle": 0, + "content": "Quantitative Comparisons. Tab 4 shows the ablations of motion refinement in terms of endpoint error metric (lower is better). We evaluate all scenes in both two datasets with user input limited to five sparse hints. The results prove motion refinement is essential for achieving a better ability to represent GT motion. Refer to supplementary for quantitative metrics of the three motion prediction pipelines." + }, + { + "type": "table", + "bbox": [ + 0.131, + 0.522, + 0.418, + 0.57 + ], + "angle": 0, + "content": "
MethodsDataset
Holynski Val SetCLAW Test Set
DenseMotion0.2990.596
RefinedMotion0.2690.582
" + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.571, + 0.43, + 0.582 + ], + "angle": 0, + "content": "Table 4. Quantitative Motion Refinement Ablation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.589, + 0.199, + 0.606 + ], + "angle": 0, + "content": "5. Byproducts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.468, + 0.78 + ], + "angle": 0, + "content": "Interactive Editing. We can augment fluid scene while keeping realistic fluid animation. For example, we can place an imaginary stone in a river that is originally slack water, and the flow could be updated to the new one with the effect of a vortex around the stone, as can be viewed in Simulation w/ Fluid-Rock Interaction of Figure 8. To achieve this, we only need to calculate interface of inserted stone mesh and fluid surface mesh. Then, we classify the part of stone above/beneath the river surface according to the depth. Later, fluid motion is updated with boundary conditions during simulation. Refer to supplementary for details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Layer Replacement Editing. We can change texture attributes of both surface fluid and background by simply replacing either fluid layer or background layer with another reference texture. For example, as shown in Figure 9, the fluid layer can be replaced by/regulated with a starry sky video with an endless loop to form dreamlike fluid effects that might remind the user of a Chinese clas" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.89, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.227, + 0.892, + 0.296 + ], + "angle": 0, + "content": "Figure 9. Layer Replacement Editing. First row shows our SLR output and new background. The second row shows results of our background layer replacement effects, fluid layer replacement effects, and replacement effects using motion instead of alpha Arrows specify the motion direction and speed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.892, + 0.503 + ], + "angle": 0, + "content": "sical poetry—as if the Silver River fell from azure sky. For a downstream task like such texture attribute replacement, accurately segmentation of all fluid regions is necessary to reach plausible visual effects. However, it is hard to be accomplished by brute force segmentation with either motion prediction (Fluid Replacement Editing by Motion Threshold w/o SLR in the figure) or manually labelling. Since real-world fluid scenes contain complex context relations such as semi-transparency, different motion behaviour among fluid regions in one scene and so on. On the contrary, since our proposed surface-based layer representation (SLR) naturally disentangles scene into surface fluid and background, we can achieve such editing easily." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.512, + 0.618, + 0.527 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.794 + ], + "angle": 0, + "content": "We propose a system for fluid animation from a still image. Specifically, for fluid texture, a neural network is designed to decompose the still image into surface liquids and background, which form a novel two-layered representation, SLR, to represent fluid scene. Compared to previous work, SLR can help better simulating the transparent fluids as well as other fluids scenes with complex surrounding. For motion estimation, we introduce surface-only fluid simulation (SFS), which provides a wider range of editing abilities on the fluid image. The system ensures physical reasonability and meanwhile skirting the scene-specific modelling problem. However, as our method works in a semi-supervised manner, without supervision from ideal decomposition or graphic-based rendering, the outputs may have some artifacts on boundaries. We expect more research on better disentangling and representing fluid surface with backgrounds, as well as more precise motion prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15930" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.155 + ], + "angle": 0, + "content": "[1] Scott G Bardenhagen and Edward M Kober. The generalized interpolation material point method. CMES, 5:477-496, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[2] Robert Bridson. Fluid simulation for computer graphics. AK Peters/CRC Press, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.47, + 0.24 + ], + "angle": 0, + "content": "[3] Yung-Yu Chuang, Dan B Goldman, Ke Colin Zheng, Brian Curless, David H Salesin, and Richard Szeliski. Animating pictures with stochastic motion textures. ACM TOG, 24:853-860, 2005. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.243, + 0.469, + 0.297 + ], + "angle": 0, + "content": "[4] Vincenzo Citro, Paolo Luchini, Filippo Giannetti, and Franco Auteri. Efficient stabilization and acceleration of numerical simulation of fluid flows by residual recombination. J. Comput. Phys., 344:234-246, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.469, + 0.338 + ], + "angle": 0, + "content": "[5] Fang Da, David Hahn, Christopher Batty, Chris Wojtan, and Eitan Grinspun. Surface-only liquids. ACM TOG, 35:1-12, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.469, + 0.394 + ], + "angle": 0, + "content": "[6] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating landscape: self-supervised learning of decoupled motion and appearance for single-image video synthesis. ACM TOG, 38:175:1-175:19, 2019. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.469, + 0.422 + ], + "angle": 0, + "content": "[7] L. A. Gatys, A. S. Ecker, and M. Bethge. Image style transfer using convolutional neural networks. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.469, + 0.465 + ], + "angle": 0, + "content": "[8] I.J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial networks. NeurIPS, 3:2672-2680, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.469, + 0.535 + ], + "angle": 0, + "content": "[9] Tavi Halperin, Hanit Hakim, Orestis Vantzos, Gershon Hochman, Netai Benaim, Lior Sassy, Michael Kupchik, Ofir Bibi, and Ohad Fried. Endless loops: detecting and animating periodic patterns in still images. ACM TOG, 40:1-12, 2021. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.537, + 0.469, + 0.576 + ], + "angle": 0, + "content": "[10] Francis H Harlow. The particle-in-cell computing method for fluid dynamics. Methods Comput. Phys., 3:319-343, 1964. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.579, + 0.469, + 0.619 + ], + "angle": 0, + "content": "[11] Aleksander Holynski, Brian L Curless, Steven M Seitz, and Richard Szeliski. Animating pictures with eulerian motion fields. In CVPR, 2021. 1, 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.469, + 0.675 + ], + "angle": 0, + "content": "[12] Yuanming Hu, Tzu-Mao Li, Luke Anderson, Jonathan Ragan-Kelley, and Frédo Durand. Taichi: a language for high-performance computation on spatially sparse data structures. ACM TOG, 38:201, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.717 + ], + "angle": 0, + "content": "[13] Libo Huang, Ziyin Qu, Xun Tan, Xinxin Zhang, Dominik L Michels, and Chenfanfu Jiang. Ships, splashes, and waves on a vast ocean. ACM TOG, 40:1-15, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[14] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[15] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[16] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[17] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[18] ByungMoon Kim, Yingjie Liu, Ignacio Llamas, and Jaroslaw R Rossignac. Flowfixer: Using bfecc for fluid simulation. Technical report, Georgia Institute of Technology, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[19] Theodore Kim, Nils Thürey, Doug James, and Markus Gross. Wavelet turbulence for fluid simulation. ACM TOG, 27:1-6, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.243 + ], + "angle": 0, + "content": "[20] Pierre-Yves Laffont, Zhile Ren, Xiaofeng Tao, Chao Qian, and James Hays. Transient attributes for high-level understanding and editing of outdoor scenes. ACM TOG, 33:1-11, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.892, + 0.285 + ], + "angle": 0, + "content": "[21] Thi-Ngoc-Hanh Le, Chih-Kuo Yeh, Ying-Chi Lin, and Tong-Yee Lee. Animating still natural images using warping. TOMM, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.287, + 0.892, + 0.326 + ], + "angle": 0, + "content": "[22] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.328, + 0.892, + 0.382 + ], + "angle": 0, + "content": "[23] Chieh Hubert Lin, Hsin-Ying Lee, Yen-Chi Cheng, Sergey Tulyakov, and Ming-Hsuan Yang. Infinitygan: Towards infinite-resolution image synthesis. arXiv preprint, arXiv:2104.03963, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.384, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[24] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.44, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[25] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.482, + 0.892, + 0.534 + ], + "angle": 0, + "content": "[26] Erika Lu, Forrester Cole, Tali Dekel, Andrew Zisserman, William T Freeman, and Michael Rubinstein. Omnimatte: associating objects and their effects in video. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.536, + 0.892, + 0.576 + ], + "angle": 0, + "content": "[27] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. arXiv preprint, arXiv:2112.03051, 2021. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.579, + 0.892, + 0.605 + ], + "angle": 0, + "content": "[28] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In CVPR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.607, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[29] Makoto Okabe, Yoshinori Dobashi, and Ken Anjyo. Animating pictures of water scenes using video retrieval. Vis. Comput., 34:347-358, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.648, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[30] Junting Pan, Chengyu Wang, Xu Jia, Jing Shao, Lu Sheng, Junjie Yan, and Xiaogang Wang. Video generation from single semantic label map. In CVPR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[31] Ekta Prashnani, Maneli Noorkami, Daniel Vaquero, and Pradeep Sen. A phase-based approach for animating images using video examples. In Comput Graph Forum, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[32] Erik Reinhard, Michael Adhikhmin, Bruce Gooch, and Peter Shirley. Color transfer between images. Comput. Graph. Appl., 21:34-41, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[33] Arno Schödl, Richard Szeliski, David H Salesin, and Irfan Essa. Video textures. In SIGGRAPH, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.801, + 0.892, + 0.827 + ], + "angle": 0, + "content": "[34] Meng Sun, Allan D. Jepson, and Eugene Fume. Video input driven animation (VIDA). In ICCV, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.828, + 0.892, + 0.868 + ], + "angle": 0, + "content": "[35] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.87, + 0.892, + 0.896 + ], + "angle": 0, + "content": "[36] Yongning Zhu and Robert Bridson. Animating sand as a fluid. ACM TOG, 24:965-972, 2005. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15931" + } + ] +] \ No newline at end of file diff --git a/2023/Simulating Fluids in Real-World Still Images/d51a02b7-7fc4-4f85-a456-fd173559877f_origin.pdf b/2023/Simulating Fluids in Real-World Still Images/d51a02b7-7fc4-4f85-a456-fd173559877f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ce43446fab892e9233ed12aea098a04f12e54d46 --- /dev/null +++ b/2023/Simulating Fluids in Real-World Still Images/d51a02b7-7fc4-4f85-a456-fd173559877f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:faeb44d6e04c6c64188a9ca7e022b6cd4fb3489edb5b94e9827c19bd7dad092c +size 9630123 diff --git a/2023/Simulating Fluids in Real-World Still Images/full.md b/2023/Simulating Fluids in Real-World Still Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6fd9e008c6c0146bebdc32dfa6560138d1e38b2c --- /dev/null +++ b/2023/Simulating Fluids in Real-World Still Images/full.md @@ -0,0 +1,309 @@ +# Simulating Fluids in Real-World Still Images + +Siming Fan $^{1*}$ , Jingtan Piao $^{1,3*}$ , Chen Qian $^{1}$ , Hongsheng Li $^{2,3,4\boxtimes}$ , Kwan-Yee Lin $^{2,3\boxtimes}$ , SenseTime Research, Shanghai AI Laboratory, The Chinese University of Hong Kong, CPII + +{fansiming,qianchen}@sensetime.com, 1155116308@link.cuhk.edu.hk, + +hsli@ee.cuhk.edu.hk, junyilin@cuhk.edu.hk + +![](images/12b0fd2234020a29b22095c20374d547abc6e5befb82d30306aea0c165030ea6.jpg) +Figure 1. Given a still image and a coarse hint (mask and arrows) of motion as input (left), our model estimates fluid motion to generate an animating video (right). Holynski [11] is a state-of-the-art method. Please refer to the supplemental materials for video playing. + +![](images/65a5a608f1cb7fc964b626b8426cd82553465cad58e5080876b0256b469bc9a7.jpg) + +# Abstract + +In this work, we tackle the problem of real-world fluid animation from a still image. The key of our system is a surface-based layered representation, where the scene is decoupled into a surface fluid layer and an impervious background layer with corresponding transparencies to characterize the composition of the two layers. The animated video can be produced by warping only the surface fluid layer according to the estimation of fluid motions and recombining it with the background. In addition, we introduce surface-only fluid simulation, a 2.5D fluid calculation, as a replacement for motion estimation. Specifically, we leverage triangular mesh based on a monocular depth estimator to represent fluid surface layer and simulate the motion with the inspiration of classic physics theory of hybrid Lagrangian-Eulerian method, along with a learnable network so as to adapt to complex real-world image textures. Extensive experiments indicate our method's competitive performance for common fluid scenes and better robustness and reasonability under complex transparent fluid scenarios. Moreover, as proposed surface-based layer representation and surface-only fluid simulation naturally disentangle the scene, interactive editing such as adding objects and texture replacing could be easily achieved with + +realistic results. Code, and dataset are publicly available. + +# 1. Introduction + +Given an image of a scene containing liquid such as streams and waterfalls, humans can imagine how the liquid in the still image would move. The problem of animating fluid from a single image has become a blooming topic recently studied by a series of work [6, 11, 27, 9, 21]. Their general pipeline can be summarized as: (1) estimating optical flow that indicates the movement of liquid and (2) synthesizing future frame image based on the motion estimation. Although these methods achieve impressive visual effects on simple fluid regions, how to handle complex contents with challenging transparency, collisions, and thin structures in real-world scenes is still a challenging and open problem. + +The key to the limitation of previous work is that the generation process of videos is modeled in a global manner. Specifically, the subsequent frames are generated based on warping existing textures on input images [3] or neural features [6]. Such operations regard the scene as a whole and warp all the contents in the scene together. For some complex transparent fluid cases, as shown in Figure 2(b), we can see through the water to observe the background rocks beneath the liquid. Textures on those static objects + +are also improperly deformed and result in unnatural animations. To decompose the influence of motion on fluids and surrounding static objects, we propose a two-layer representation for scenes with fluids, namely Surface-based Layered Representation (SLR), which models the motion of the surface fluid and the rest background separately. To achieve such decomposition from the still image during inference, we propose a temporal-based training scheme, which facilitates the network to learn the decomposition with multiple supervisions over spatial and temporal dimensions of running fluid videos. + +Another major difficulty in animating a single fluid image is predicting the motion flow of the fluid. There exist two types of methods. The first category is the interactive motion prediction methods. They require sparse hints (ie, labelling the velocity direction and relative amplitude) on random pixels of the liquid region as additional inputs. The feature clustering model [27] or convolutional network [11] is then used to generate a dense motion field. As the precision depends on the level of motion details a user can provide, such methods are time-costing in complex fluid scenes. The second category is automatic methods [6]. These works regard the task as a domain transformation problem, and use image-to-image translation techniques to predict the motion represented by 2-channel optical flows. Ambiguous velocity (e.g., a flat river surface can flow either to the left or to the right, as shown in Figure 2(a)) and tremendous changes of velocity in a period of time caused by the collisions with objects (as shown in Figure 2(c)) may pose great challenges to the above methods. To tackle the challenges, we introduce a simplified Navier-Stokes simulation into the motion estimation process, called Surface-only Fluid Simulation (SFS). Specifically, the initial motion prediction of the liquid region is firstly calculated through sparse user inputs and interpolation inside the whole region. In parallel, a 3D mesh is built based on a monocular depth estimator. Then, the N-S equation is solved on the mesh surfaces and based on the initial motion to obtain a simulated motion field in a short period of time. Finally, since the real thickness of the fluid is neglected during the simulation, a refinement step is expected to compensate for the motion offset at height. Thus, a DNN-based motion translator is applied subsequently to obtain the refined motion field that adapts naturally to the input image. With SFS, we can enforce precise controls on the movement of fluid. + +In summary, our work has three major contributions. (a) We propose a new and learnable representation, surface-based layered representation, which decomposes the fluid surface and the static objects in the scene to better synthesize the fluid animating videos from a single fluid image. (b) We design a surface-only fluid simulation to model the evolution of the image fluids with better visual effects. (c) Based on the proposed surface-based layered representation + +![](images/8f4a06c3ce854edc582cdb7bf66e242a650536f414d983e0696d568af4453a8b.jpg) +Figure 2. Motivation. (a) Given a still input image, the motion may be ambiguous. (b) Animating via previous SOTA method [11] causes background rocks beneath the liquid improperly deformed. (c) Motion prediction methods such as [11, 27] cannot well capture motion changes caused by the collisions with objects. + +and surface-only fluid simulation, image editing with fluids is achieved with realistic and vivid results. In addition, We can easily extend our fluid animation system to downstream editing tasks, thanks to the SLR and SFS designs in our framework. For example, we can augment objects in the fluid region or background region, and simulate the interactions between the fluids and objects. We can also replace textures to different layers to create different scenes. + +# 2. Related Work + +# 2.1. Video Generation + +In early studies, researchers focus on generating new video context by changing the attributes of recorded frames [32, 33]. Traditional methods of video motion generation with movable substances are based on stochastic process analysis [34] and color transfer [32]. Given a reference video indicating the temporal changes of a scene, texture movement and sequential images can be generated using a patch aligned method and energy-based optimizing [33]. Similar methods are applied to single image input [3]. However, obvious artifacts exist, and the smoothness of the video is not guaranteed. These drawbacks make the usage of such methods limited to typical types of scene samples. + +In recent years, since the development of deep neural networks in image generation field [15], plenty of innovative efforts have been made to improve the synthesis quality of images [7, 16, 8, 35, 25]. Techniques such as GANs [8], perceptual loss [17], partial convolution [25] and cycle-consistency [35] have been widely used as standard components. Built upon the success of image generation, several attempts have created the precedents of neural video generation. A few methods are developed to force the network to learn intermediate scene representation for predicting the future frames or interpolating key frames, once the input sequences and camera trajectory are given (note: as we focus on single image input, we skip the review of + +such multi-frames input trend). Given a single image as input, latent generative methods like [24] proposes an autoregressive model to generate infinite pixels forming sequential outputs. InfinityGAN [23] disentangles global appearances, local structures and textures with a structure synthesizer and texture synthesizer to generate images with spatial size and level of details not attainable before. [31] tries to use Fourier transforms in time variance showing the phase variations to represent the flow and generate moving videos. [20] spends efforts in digging into the semantic meanings in images and generates videos correspondingly to achieve a natural result. [6] decouples motion and appearance through learning intermediate flow fields and color transfer map. However, these methods are either focusing on spatial-wise consistency [20], or paying attention to coarse-level time-variance scene changes [31, 6]. None of them discusses or can handle the particularity of fluid with fine detail motions. + +# 2.2. Fluid Simulation + +Physical Fluid Simulation. Traditional methods originated from computer graphics to achieve accurate approximations on various types of fluids. Hybrid Lagrangian-Eulerian Methods [2, 36, 10] have successfully simulated the motion of incompressible fluids in a pre-defined 2D or 3D box environment, which solves pressure projection on Eulerian grid and advect on Lagrangian particles. To improve convergence ability and flexibility on complex shapes, material point method (MPM) [1] has been popular and more subsequent methods [18, 19] have been proposed to stabilize the convergence when iterating velocities. However, these methods are time-consuming in large scenes. To reduce the cost of simulating a large-range of fluid, methods have been proposed to simulate only a small range of fluid [4], or the surface of the fluid [5, 13] to produce a faster approximation by avoiding high computational cost in 3D volumetric solvers. Although these works achieve amazing simulation effects, they are still limited to synthetic scenarios due to scene-specific modelling. + +Animating Fluids with Data Prior. Given a still image as animation target and a set of real-world water video candidates as animation bank, [31, 29] propose to retrieve a video or a group of videos from the bank that has similar content with the target image as the references. Then they transfer motion and appearance of references to each region of interest on target image to form the water animation video. These methods open the door of real-world fluid animation with data prior in-the-wild. However, the seamless alignment of animated results with source input image can not be well guaranteed and requiring lots of manual efforts. With the development of deep learning, recent works [11, 27] propose to automatically predict each frame of motion and appearance via learning from reconstruction loss + +with real world videos for training. [11] first simplifies motion estimation part as a single frame motion prediction via motion Eulerian integration and then proposes a deep feature warping technique to narrow down the size of blank areas caused by warping. Finally, it adaptively blends forward and backward features through a learnable parameter $Z$ to generate animated fluid video. Based on [11], [27] proposes to regress fluid motion conditioned to user's sparse guidance and generate paired training data through motion speed clustering. The work further proposes to use multiscale representation to capture different fluid speed in different resolution. Although these methods take a step further to real-world scenes with impressive results, it is hard for them to handle complex context relations over fluids due to the single representation to the entire scene that regards different objects and textures as equal. + +Unlike previous work, our system marries the advantages of both physical simulation and learning-based pipelines. For textures, to avoid interplay between fluids and surroundings, we force the network to decouple the scene into a new two-layer representation SLR, that includes a static background and time-variates surface fluid layers2. Meanwhile, the network can hallucinate reasonable and photo-realistic textures in vacated regions thanks to the large-scale data prior. For motion, to skirt scene-specific modelling while keeping physiC reasonability, we propose a three-step motion prediction. We first use sparse labels from the human interface to generate the initial rough motion and then calculate reasonable evolution in a short period by introducing a 2.5D surface simulation, SFS. Later, a smooth motion translator is used to refine the motion trend to better fit real-world examples with data prior. Moreover, as a consequence, we can generate natural fluid animations with the network training on plenty of real-world examples and flexibly edit the main liquid region and boundary condition to create various visual effects. + +# 3. Method + +Given a real-world image that contains liquid, our goal is to generate a video clip that animates vivid flowing of the fluid. To achieve this goal, a system is required to (1) represent complex context relations between fluids and surrounding environments; and (2) to estimate motion fields from single image input with reasonability and efficiency. We propose a novel system with two main designs that satisfy the above requirements. For texture, in contrast with previous methods that regard all elements in the scene as an entirety, our system learns a surface-based layered representation (SLR) (Sec. 3.1) that decomposes the scene into surface fluid layer and impervious background layer. Since the tex + +ture characteristics of surface fluid and rest background are assigned to different layers, such a design could overcome the negative influence of motion flows between fluids and surroundings (e.g., improper texture warping, as shown in Figure 2(b)). As for motion estimation, the system provides a three-step motion estimation alternative by introducing a surface-only fluid simulation (SFS) into the motion field prediction (Sec. 3.2). The motion prediction steps combine advantage of both physical- and learning-based motion estimation pipelines, thus ensuring physical reasonability and meanwhile skirting the scene-specific modelling problem. Each component is detailed in following subsections. + +# 3.1. Learning of Decomposition of Images + +In this subsection, we first introduce Surface-based Layered Representation (SLR). Then, we show how to obtain animated videos through SLR. Finally, we present how to learn SLR with a neural network (illustrated in Figure 3), and challenges we meet during optimizing the network since there is no oracle ground-truth supervision for fluid scene decomposition. + +Surface-based Layered Representation. Given a still image $I(T_0)$ as input, a series of RGB frames $\mathbf{I} = \{I(T_0),\dots,I(T_n)\}$ are outputs to synthesize time-space variations of the scene with running fluids. As we aim to handle complex context of fluid scene through disentangling fluid motion and still objects beneath/interacting with surface fluid, we propose to learn a novel intermediate representation whose output is an impervious background layer image $I_b$ , a temporal sequence of surface fluid layer images $\mathbf{I}_f = \{I_f(T_0),\dots,I_f(T_n)\}$ and transparency factors $\alpha_{b},\alpha_{f} = \{\alpha_{f}(T_{0}),\dots,\alpha_{f}(T_{n})\}$ , where $\alpha_{*}\in [0,1]$ . Transparency factors indicate the contribution of colors on each layer to the final image at each time. In practice, we utilize a convolution network to output the simplification of the representation: $F_{\theta}: (I(T_0),I(T_n),\mathbf{M})\to (I_b,I_f(T_0),I_f(T_n),\alpha_b,\alpha_f(T_0),\alpha_f(T_n))$ , where the network only outputs the estimated first and last frames of $\mathbf{I}_f$ as well as $\alpha_{f}$ . $\mathbf{M}$ is the estimated motion fields of $T_0$ to $T_{1}$ from the single input $I(T_0)$ . We will detail the synthesis of $\mathbf{M}$ in Sec. 3.2. + +We expect the background to be identical across all times, while the fluid surface deforms with realistic motions. To this end, $F_{\theta}$ is enforced to automatically map the original texture of non-liquid and static regions as well as the reasonable texture of under-liquid surface regions to the background layer (detailed in later subsection). As for fluid surface learning, recurrent estimation for all frames (RNN) is a possible alternative that can be tailored from video gen + +eration [30]. However, such an intuitive design might cause frame distortion as time goes on due to the accumulated errors and hard convergence (also demonstrated in [11]). In contrast, We use the linear combination of the first and last frames to construct the images at intermediate frames. This choice not only transforms the problem into interpolation rather than extrapolation, which is easier to converge, but also is able to generate an endless cyclic video rather than a video of a fixed interval. In detail, we tailor the symmetric softmax splatting, which is a pixel-warping strategy utilized in [11, 28], to our framework. We extend the splatting operation from a single image feature splatting into two levels – fluid surface feature splatting and fluid alpha splatting as follows : + +$$ +D _ {f} \left(T _ {i}; \hat {\mathbf {x}} ^ {\prime}\right) = \frac {\sum_ {\hat {\mathbf {x}} \in \mathcal {X}} \alpha \left(T _ {i} ; \hat {\mathbf {x}}\right) \cdot D _ {f} \left(T _ {j} ; \hat {\mathbf {x}}\right) \cdot \exp \left(Z _ {f} \left(T _ {j}; \hat {\mathbf {x}}\right)\right)}{\sum_ {\hat {\mathbf {x}} \in \mathcal {X}} \alpha \left(T _ {i}; \hat {\mathbf {x}}\right) \cdot \exp \left(Z _ {f} \left(T _ {j}; \hat {\mathbf {x}}\right)\right)}, \tag {1} +$$ + +$$ +\alpha_ {f} ^ {\prime} \left(T _ {i}; \hat {\mathbf {x}} ^ {\prime}\right) = \frac {\sum_ {\hat {\mathbf {x}} \in \mathcal {X}} \alpha \left(T _ {i} ; \hat {\mathbf {x}}\right) \cdot \alpha_ {f} ^ {\prime} \left(T _ {j} ; \hat {\mathbf {x}}\right) \cdot \exp \left(Z _ {f} \left(T _ {j} ; \hat {\mathbf {x}}\right)\right)}{\sum_ {\hat {\mathbf {x}} \in \mathcal {X}} \alpha \left(T _ {i} ; \hat {\mathbf {x}}\right) \cdot \exp \left(Z _ {f} \left(T _ {j} ; \hat {\mathbf {x}}\right)\right)}, \tag {2} +$$ + +where $j \in \{0, n\}$ is the start index or last index of sequences, indicating forward splatting the frame 0 or backward splatting the frame $n$ , which means the value $j$ is: + +$$ +j = \left\{\begin{array}{l l}0,&\text {i f} \hat {x} ^ {\prime} = M _ {0 \rightarrow t} (\hat {x}),\\n,&\text {i f} \hat {x} ^ {\prime} = M _ {0 \rightarrow t - N} (\hat {x}).\end{array}\right. \tag {3} +$$ + +$D_{f}(T_{i};\hat{\mathbf{x}}^{\prime})$ is the fused fluid feature value of target frame at pixel $\hat{\mathbf{x}}$ , and $\alpha_f^\prime (T_i;\hat{\mathbf{x}}^\prime)$ is the transparency value of target frame at pixel $\hat{\mathbf{x}}$ . $D_{f}(T_{i};\hat{\mathbf{x}})$ represents value of $D_{f}(T_{i})$ at pixel $\hat{\mathbf{x}}$ , $\alpha_{f}(T_{i};\hat{\mathbf{x}})$ and $\alpha_f^\prime (T_i;\hat{\mathbf{x}})$ are similar. During inference, we set $D_{f}(T_{n}) = D_{f}(T_{0})$ and $\alpha_{f}(T_{n}) = \alpha_{f}(T_{0})$ since only single image is available. $\mathcal{X}$ is the set of pixels satisfying $M_{0\rightarrow t}(\hat{\mathbf{x}}) = \hat{\mathbf{x}}^{\prime}$ or $M_{0\rightarrow t - N}(\hat{\mathbf{x}}) = \hat{\mathbf{x}}^{\prime}$ , $M_{0\rightarrow t}$ is displacement map generated from $\mathbf{M}$ through euler integration [11, 28], and $\alpha (T_i;\hat{\mathbf{x}}^{\prime})$ is a temporal coefficient to fuse features from forward warping and backward warping to ensure endless looping: + +$$ +\alpha \left(T _ {i}; \hat {\mathbf {x}} ^ {\prime}\right) = \left\{\begin{array}{l l}\frac {t}{n},&\text {i f} \hat {x} ^ {\prime} = M _ {0 \rightarrow t} (\hat {x}),\\1 - \frac {t}{n},&\text {i f} \hat {x} ^ {\prime} = M _ {0 \rightarrow t - n} (\hat {x}).\end{array}\right. \tag {4} +$$ + +Then a fluid feature decoder is applied to $D_{f}(T_{i})$ to generate surface fluid image $I_{f}(T_{i})$ and a alpha refinement network is applied to $\alpha_f'(T_i)$ to generate transparency alpha $\alpha_{f}(T_{i})$ to avoid vacant alpha particles. + +Animating Fluids with SLR. Once we obtain the final surface fluid layer $I_{f}(T_{i})$ and transparency $\alpha_{f}(T_{i})$ , the reconstructed final image $I(T_{i})$ can be acquired by adding the impervious background layer $I_{b}$ back to the surface fluid layer $I_{f}(T_{i})$ with $\alpha_{b}$ . Here we use separate $\alpha$ channels for background and fluid transparency predictions. This design helps to stabilize the training of decomposition. The core reason is that when a single fluid transparency $\alpha_{f}$ is warped according to the motion flow, it leaves a blank region on the boundary, and makes the gradients hard to be + +![](images/d3022add8b5dd7b327414a4ffeb8837e765a9c118d7b06f824d305d6927933d3.jpg) +Figure 3. Training and testing pipelines. The background layer and background alpha are predicted only at the first frame. Both fluid features and fluid alpha are predicted at the first and last frames, and then symmetricly splatted and blended using the softmax splatting at frame $T_{i}$ . A decoder is used to decode fluid features into images and refined alpha. For frame $T_{0}$ , motion is calculated following pipeline Figure 4, and for frame $T_{i}$ , motion is calculated through Euler integration. During inference, $T_{n}$ is replaced by $T_{0}$ to create a cyclic video. + +backpropagated. Although with the two-channel implementation, there still is a $\alpha_{b}$ on that vacated regions to fascinate the network refining the texture. In formula we have + +$$ +I \left(T _ {i}\right) = \frac {\alpha_ {f} \left(T _ {i}\right) I _ {f} \left(T _ {i}\right) + \alpha_ {b} I _ {b}}{\alpha_ {f} \left(T _ {i}\right) + \alpha_ {b}} \tag {5} +$$ + +For corresponding network architectures to achieve the above targets, as shown in Fig 3, the encoder is used to extract background texture and $\alpha$ channel as well as surface fluid layers' features and its $\alpha$ . The decoder is used to refine the fused fluid layer that combines the layers from the first and last frames to obtain the complete surface fluid layer, and the final animated image is reconstructed by compositing background and surface fluid layers. + +Optimizing the SLR. The training process is divided into three parts, in which, we separately train (1) the surface fluid branch containing the encoder that predicts the fluid feature layer from image input and the decoder that reconstructs the warped fluid layer, (2) the background branch that predicts the background layer from the image input, and (3) jointly train them together with $\alpha$ learning and surface fluid/ background branches finetuning. For each training step, the paired training data $< (I(T_0), I(T_n)), I(T_i) >$ are randomly picked among 60-frames sequence without constraint on interval. We use dense optical flow estimation network [14] to extract reasonable movement of each video as pseudo motion ground-truth, and to ensure smoothness, we average the flow with window size of 30 frames. Due + +to space limits, we detail the training of each stage in the supplementary materials. + +# 3.2. Motion Field Estimation + +In the previous section, we have described the core necessities to learn an SLR but skipped the synthesis process of one of its input, motion fields $\mathbf{M}$ . In this section, we discuss how to obtain $\mathbf{M}$ . + +During training, the motion of each frame can be generated from a conventional optical flow estimator [14]. During inference, as the input is a single image, we expect to synthesize a reasonable motion map from the observed scene. We propose a motion estimation process that separates the motion prediction into three parts taking advantage of both physical-based and learning-based pipelines for fluid simulation. First, interactive labelling of the sparse velocity of several pixels on the image is required to guide the generated motion directions of $I(T_0)$ , along with a fusion with weights calculated based on Gaussian filtering of the distance maps from the labeled pixel to build a dense velocity initial map. Second, a surface-only fluid simulation is calculated based on initial motion and estimated scene depth to simulate a following convincing evolution of the liquid. Note that this step only calculates for the first frame. Last, a DNN-based pixel-to-pixel motion translator is utilized to refine the flow to create a more realistic and natural motion map. This translator could be considered as generating a compensation for a detailed motion map that is + +![](images/9f75a0df9b2dfe7966d6e6e2026c7ada309e5193716dddf1816587371302d9ed.jpg) +6uizisauus 6uizisauus pue uonnnsr (e) + +![](images/ff40a85ade35dcae10f018149db4dca6fee956f16d88fd2f044c6914eb4909f0.jpg) +2. Label Mask and Velocity Hint + +![](images/38ba20ee3bd6a16b12cfe736e0ce1ee1d431ed8ed56b9d0b942b7409c9063224.jpg) + +![](images/641bdf13877f0b84f955c378646caa2c5eff228b0dbeb392703208ab91049648.jpg) +1. Input Image +4. Mono-Depth Estimation + +![](images/145265d27db8f6cafa89256a9819cbb45795e7a0b25c033a601a25f9857dcf07.jpg) +5. Build 3D Mesh + +![](images/f279058ea667c6d55a7dc890ea243467c79ea322c4a23da6aadd0544f0f915bb.jpg) +3. Generate Dense Velocity +6. Mesh-Based Simulation + +![](images/cb7bc957b4e56a7b68ff354cadf2ad58acddd720b29274166606f09d02790e17.jpg) +7. Refine Velocity + +![](images/a9caf5cbf6b56b3d70b51435e3c8b1c685f4331ecff0be76fab467a68b5d7d2b.jpg) +8. Synthesize Result + +![](images/6fa6e7a9dad637d9e9b745aa4f1cd8ceb32d87b1e7a806558a06a88820406200.jpg) +9. Editing AR objects + +![](images/9e09676cde456c99c2180fea141999763d3ed4cf80bc754826a694ab22dc6ce8.jpg) +10. Add object in Fluid Layer + +![](images/ed95a40ef165f10778aa5e4e356a7caa37a6c6e001319ebe3fd53364d296d4fe.jpg) +1. Add object in Background layer + +![](images/a33d82210fdda115b7af5a5dceda58b9b557fa8a755528ec77f37bb3e1ab249e.jpg) +12. Mark Intersection Vertex "Solid +Figure 4. Motion calculation pipeline. (a) is initialization of fluid region, rock region, motion, and Mesh. (a,b) combine to form our 2.5D simulation and animation pipeline, where (4-5) is not included if using 2D simulation method. (a,c) combine to form our editable simulation and synthesis pipeline, where the user can edit 3D objects and create corresponding motion. + +![](images/885ccab6f9e04014b188dabc5b45106d9255102c1956cfe3b2e7e43344d5247e.jpg) +Mesh-Based Simulation + +![](images/bf57bb705454dc24c6444a0c7c7f97f25e4ca55063a174d0a7cf498e5083ff37.jpg) +14. Edited Synthesized Result + +harmonic with fluid textures. The pipeline is illustrated in Figure 4. + +Initial Motion from Interactive Sparse Labeling. To obtain determined as well as editable motion direction, we choose an interactive alternative to generate initial flow for image $I(T_0)$ following [27]. With no more than 10 discrete notations on images, and a fixed segmentation of the liquid region, we use nearest-neighbors-averaging to assign all the pixels that are inside the liquid region an initial velocity. + +Surface-only Fluid Simulation. After the initial state is estimated, we tailor the traditional graphics pipeline with discretization and linear equation to simulate the follow-up evolution of the liquid. Intuitively, we can directly apply simulation on 2D grids that align with image pixels by initializing each pixel as an Eulerian grid and several Lagrangian particles near each pixel. However, since such perspective projection of the 3D scene may cause severe distortion when direct manipulating the motion in 2D, as shown in Figure 7, we thus turn to use a mono-depth estimator [22] to estimate the 3D surface of the liquid. After the 3D surface is constructed inside the region of the liquid, a mesh surface-based velocity is calculated, with the restriction of projected velocity: + +$$ +\frac {d}{d t} \left[ \begin{array}{l} u \\ v \end{array} \right] = \left[ \begin{array}{c c} f x & 0 \\ 0 & f y \end{array} \right] \left[ \begin{array}{c c c} 1 / z & 0 & - x / z ^ {2} \\ 0 & 1 / z & - y / z ^ {2} \end{array} \right] \left[ \begin{array}{c} \frac {d x}{d t} \\ \frac {d y}{d t} \\ \frac {d z}{d t} \end{array} \right] \tag {6} +$$ + +Where $fx, fy, cx, cy$ is camera intrinsic parameters, $u, v$ is + +![](images/8e2bc8d256c01b1d907d1362cbb75cec30df8e4b2a94a9c4a8074026b7e87162.jpg) + +![](images/f4f3659c2596b56d5aca88aa5b0fad3dbd2deeb6e229609166a5399e73c8a775.jpg) + +![](images/c7e56c13e2b0350c9f6544a503ea131c67bddcb7eec7fdf51eb2a3bfd7dc9ae0.jpg) + +![](images/2793bf549947b7c855f27ccaa441b1271f533b5e07813fb1f12ea4ce8e0f5286.jpg) +Figure 5. Examples for Different Fluid Scenes. (a) The surface thickness of the waterfall and river are gauzy. (b) The fountain cannot be considered a surface. (c) Fluid with vortex contains complex motion. + +![](images/a3397f45c2a8e229f0d7b5a17ec9c61dd74fef053035c366b54bb5ce64d3485f.jpg) +(b)Not a Surface + +![](images/bb145d3a50eb212c96f65a3cf2513508f0b9399a40543016b34cc91f90dc5882.jpg) +(c)Complex Initialization + +the velocity on the projected image and $x, y, z$ the 3D position in the scene. Since we have modelled the liquid as a surface, we replace traditional Laplacian operator with Laplacian-Beltrami operator on the mesh and use Mixed-Euler-Lagrange approach [36] implemented in Taichi [12] to simulate consequent motion of each vertex. Detailed formula derivation can be seen in the supplementary. The particle binding on the surface is randomly sampled according to the surface area. When a particle is moved outside the mesh surface, we use the nearest projection to force the position to fall onto the mesh. For the boundary of the mesh, we calculate the velocity of the boundary triangle, judging whether the velocity on the edge boundary is inward or outward. For outward edges, the particle is abandoned after moving out of the liquid region. For inward region, a particle source is built to maintain the overall number of particles to be roughly stable. With the texture of surface fluid layer binding to each vertex, we can warp fluid layer with a soft rasterizer to simulate flows. In this way, given arbitrary flow initial state, a stable velocity can be calculated to approximate real-world flow after a period of simulation. + +Smooth Motion Translator. With the help of above $2.5D$ fluid simulation, we can generate a time-variant velocity among the sequence. However, the simulated velocity is still a rough trend since it only works on fluid surface without considering thickness variation of fluids. Thus, the SFS can work well in cases like Figure 5(a). It fails in cases with distinct depth measurement differences, as shown in the Figure 5(b) due to the lack of height information during calculation. Also, It is fragile to cases with complex initial state as shown Figure 5(c) due to error-prone/time-consuming interactive initial motion labeling for such cases. Thus, we need extra scheme to compensate or rectify for detailed motion that is harmonic with the fluid textures. To this end, we use a convolution-based network to transfer the simulated motion into a more detailed and reasonable one, which is pixel-aligned with image textures. The network is trained as a traditional style-transfer task, with a L2 loss between output velocity and ground truth. An external patch discriminator is used to refine the pattern of the velocity distribution. All the settings are similar to the image translation task pipeline [15]. + +# 4. Experiments + +In this section, we present experiments and detailed analysis to demonstrate the contributions of our method both quantitatively and qualitatively. For all the cases shown in this paper, please refer to our demo video and project page for better dynamic visual effects. + +# 4.1. Quality of Fluid Video Synthesis + +Implementation Details. To learn SLR, we train our network on a series of videos containing a large region of fluid movement and still background [11]. For training of motion refinement network, we use the same settings in [27]. All networks are trained at a resolution of $256 \times 256$ . + +Datasets and Evaluation Metrics. In addition to validation set in [11], we collect extra 126 scenes to form a new test set, which we called Complex Liquid Animation in-the-Wild (CLAW) test set, for evaluating the performance on both dense fluids (such as rivers and oceans) and sparse or transparent fluids (such as streams or splashes). Note that, in this subsection, the motion for all the methods to be compared is guided by the optical flow of the whole video to distinguish the performance of the synthesizing techniques. We use PSNR to show the whole average error, SSIM to show the errors in the region with plenty of textures and LPIPS, a perceptual based loss, to show visual quality difference. For more details, please refer to supplementary. + +
DatasetMethodsAll RegionFluid Region
LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑
HolysnikReproduced Holysnik0.079825.030.77870.065725.880.8007
CommonModified Holysnik(Baseline)0.079324.750.77580.065625.720.8000
Val SetOurs0.083425.140.77950.065726.100.8030
OurReproduced Holysnik0.206720.260.59550.202920.360.5961
CLAWModified Holysnik(Baseline)0.207819.970.59230.204120.100.5934
TestsetOurs0.204020.790.60800.197520.800.6077
+ +Table 1. Quantitative comparison. (a) Results on Holynski's common val set [11], evaluating first to 60-th frames. "Fluid Region" means static region is replaced by input image during metric calculation for focusing on the statistic of fluid. This could be done by replacing the non-fluid region with the corresponding textures of input images. (b) Results on CLAW test set. + +Quantitative Comparisons with baseline. Table 1 shows the results between our method and state-of-the-art method [11], which is a single layer learning-based system, a typical method that animates the scene in a global manner. We re-implement [11] in Pytorch (Reproduced Holynski), since the source code is not public available. For other comparison baselines, Modified Holynski is built upon Reproduced Holynski but with the modification of replacing convolution to partial convolution in the fluid decoder. The third model (Ours) is our SLR model4. Three observations can + +be included from Table 1: (1) PSNR and SSIM metrics of "Modified Holynski" drop a little compared to "Reproduced Holynski" in both statistics on all- and fluid region. We infer that, although partial convolution could better capture context than conventional one to impart reasonable texture in vacated regions in theory, simply replacing partial convolution into single-layer framework cannot essentially improve the image quality due to the wrong context raised from improper warping of both background and fluid. While, with SLR, the partial convolution can strengthen its advantage on context capturing for more proper fluid context textures provided in surface fluid layer. (2) The LPIPS of Ours model drops a little compared to the Modified Holynski model in the "All Region" of Holynski common validation set, but has nearly the same LPIPS in the "Fluid Region". For the other two metrics, our model leads to an uptick. Since most regions of Holynski's validation set are static, the blending of the fluid and background in the non-fluid region could lead to slight performance drops in terms of LPIPS. However, it could be eased by replacing the non-fluid region with the corresponding textures of input images, thus we care more about the performance in the "Fluid Region" instead of the "All Region". Such a phenomenon tells us that the two-layer decomposition design in Ours model does not harm the reconstruction ability on final images. To better unfold the analysis, we further conduct the close-up evaluations on Holynski's val set. Concretely, we crop the fluid region by motion threshold and only evaluate these visible fluid background regions. The LPIPS, PSNR, SSIM of Holynski's method are 0.230, 19.01, 0.565, and Ours are 0.227, 19.64, 0.581. This setting leads to larger performance margins with the baseline, which demonstrate the advantage of our system. (3) When evaluating only in the fluid region, we can see model Ours reaches a comparable result with the "Modified Holynski" model in Holynski validation set [11], but improves significantly in CLAW test set. The major reason behind this observation is that when it comes to complex scenes as included in our CLAW test set, our proposed two-layer representation suppresses the influence between surrounding and surface fluids. Thus we can achieve better results than single-layer baselines that regard all elements in the scene as an entirety during warping. + +Comparisons with SOTA methods. In this paragraph, we provide comparisons with more state-of-the-art methods - Endo et.al [6], Tavi et.al [9], and Mahapatra et.al [27]. The first two methods are centered on animating general scenes, where Endo et.al [6] is designed for landscape, and Tavi et.al [9] focuses on animating periodic patterns in still images. These two methods could be applied to fluid scenarios. In contrast, Mahapatra et.al [27] is specialized for fluid scenarios, where a multi-scale representation is introduced to capture different fluid speeds in different resolutions. Concretely, as Endo's code is publicly available, we + +compare our system with it both quantitatively and qualitatively. For other methods, we compare them qualitatively in Fig 6. Specifically, the result of Mahapatra is from official project page, and result of Tavi is from official app. Endo's method can only control motion with a latent code. Thus, we test ten different latent codes and select the best one to compare. We observe that these three methods suffer from animation distortions with different levels – (1) from the zoom-in perspective, all baselines improperly move the still stones with the fluids. Specifically, Tavi's results have significant ghosting artifacts, which might be caused by its one-fold periodic motion and alpha-blending designs. (2) From the full image perspective, Endo's method has difficulty in estimating fluid motion. The river's motion is incorrectly predicted to be from right to left, and the waterfall flows upwards to the right. The quantitative results (Tab 1) provide the demonstration from another aspect. This phenomenon is foreseeable, as Endo's framework considers motion in a global manner. In contrast, our two-layer method animates plausible fluid without affecting still stones. Please refer to the supplementary material for more comparisons and straightforward video effects. + +![](images/c3b82864853e11bcd14c71f95201b96a64b759c2eb71a1633b4fa0b416b33d6d.jpg) +Figure 6. Qualitatively Comparisons. (e-f) are tested in $288 \times 512$ , others are tested in $\mathrm{W} = 1280$ . LPIPS/PSNR/SSIM of the whole video are colored in orange. + +![](images/b95a418ae96c45f6b2e1bcc9516afda9dffd34d0b5d87ee641691ea4a8937fbb.jpg) + +
MethodsHolynski validation setCLAW test set
All RegionFluid RegionAll RegionFluid Region
LPIPS ↓PSNR ↑SSIM ↑LPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIM
Endo[6] HD0.35716.270.3510.12022.310.7410.35715.920.2960.34716.380.331
Ours HD0.14223.430.7150.11823.780.7280.19620.990.6110.19020.990.610
+ +User Study. We conduct a user study to compare visual effects between single layer baseline (Modified Holynski) and our two layers method(Ours in Table 1) on the full set of CLAW testset. For each scene, we animate 60 frames and output a two-second video. Animated videos are shown side by side without providing ground truth videos as reference to guide the participant focusing on the reasonability and photo-realistic. We ask the participants to select which one is better or neither is better, mainly according to three dimensions (photo-realistic, stereoscopic, and high-fidelity). + +As shown in Table 3, the participants can make a distinction between the two methods over most samples, leav + +![](images/17eb2b988aa2970d3f65d355554e3032f162d2619919bd59b25f9cb49d04b567.jpg) +Figure 7. Different simulations. Orange arrows point out motion direction. All simulation methods use $75\%$ weak incompression simulation instead of fully incompression for better visual effect. (row 1)Input image, sparse motion generated by sparse user hint required by each method. (row 2-4)ablation study of Mahapatra's method [27], 2D simulation and 2.5D SFS. + +![](images/595e45724f7d490db88b74f8e91a68fef9e6387900203677e022853619390eb4.jpg) + +ing $27.2\%$ samples difficult to tell. Our method achieves better scores than the single-layer method [11] with $52.3\%$ sample voting. This indicates the effectiveness of the proposed SLR in other aspects. + +Table 2. Quantitative comparisons with Endo's method. + +
Single layerOursNeither
E-O-N125(20.5%)319(52.3%)165(27.2%)
+ +Table 3. User study. The subjective scores are voted on all samples of our CLAW testset, which includes 122 videos with various real-world fluids scenarios(e.g., river, stream and waterfall) and complex context surrounding (e.g., semi-/transparent liquid, collisions and thin structures). + +# 4.2. Different Motion Prediction + +Figure 7 shows different motion prediction pipelines. Pipeline 1: An interactive sparse labelling of the motion and then extending all the velocities to the whole moving region, which results in a constant flow that remains large gap with realistic motion. We can see that the motion appears reasonable global with a network refinement. However, for local regions around static rocks, the flow of liquid is improperly stiff and has no interactions with the rock. While in real-world scenes, there should be vortex and sprays caused by collisions. Pipeline 2: Built upon pipeline 1, while before the motion refinement stage, we additionally initialize grids with fluid, rock or air masks as well as placing particles around each grid, then advocate particles by edited motion for once. Later, a NS equation with the incompressible fluid assumption is solved on 2d grid to obtain the motion of the next frame, which we call "Simulated 2D Motion" in the figure. Considering pipeline 2 is simulating on a 2D image plane, a refinement step (using a smooth motion translator) is applied to compensate for the motion offset at height and fine details. Pipeline 3: instead of simulating on a 2D image plane, the N-S equation is solved in $2.5D$ . Specifically, we regress the depth with a pretrained monocular depth estimator and consequently build a 3D mesh over the fluid. Then, we solve the N-S equation on the mesh surface. With + +![](images/3cb386aad973956f81bfac3dd0c14d7a0a3b0d27f5eb143130268d01dfbd8612.jpg) +Figure 8. Interactive Editing. (Bottom) Initial all moving region as 'Fluid' statue. (Top) Initial 'Solid' statue at overlap area of mesh and new object. + +the help of simulation, we can see in the left scene of the figure that the effect of turbulence around the rock can be animated, and the trend maintains the refined output. From the right scene in the figure, we can observe that the absolute value of the velocity is unnatural in Pipeline 2, which regards all particles on an identical plane. However, since the liquid surface leans in the picture, there should be a foreshortening effect on the projection. In contrast, pipeline3 samples particles on the surface and such simulation enable near-plane motion to carry more speed. + +Quantitative Comparisons. Tab 4 shows the ablations of motion refinement in terms of endpoint error metric (lower is better). We evaluate all scenes in both two datasets with user input limited to five sparse hints. The results prove motion refinement is essential for achieving a better ability to represent GT motion. Refer to supplementary for quantitative metrics of the three motion prediction pipelines. + +
MethodsDataset
Holynski Val SetCLAW Test Set
DenseMotion0.2990.596
RefinedMotion0.2690.582
+ +Table 4. Quantitative Motion Refinement Ablation. + +# 5. Byproducts + +Interactive Editing. We can augment fluid scene while keeping realistic fluid animation. For example, we can place an imaginary stone in a river that is originally slack water, and the flow could be updated to the new one with the effect of a vortex around the stone, as can be viewed in Simulation w/ Fluid-Rock Interaction of Figure 8. To achieve this, we only need to calculate interface of inserted stone mesh and fluid surface mesh. Then, we classify the part of stone above/beneath the river surface according to the depth. Later, fluid motion is updated with boundary conditions during simulation. Refer to supplementary for details. + +Layer Replacement Editing. We can change texture attributes of both surface fluid and background by simply replacing either fluid layer or background layer with another reference texture. For example, as shown in Figure 9, the fluid layer can be replaced by/regulated with a starry sky video with an endless loop to form dreamlike fluid effects that might remind the user of a Chinese clas + +![](images/788a76442a5b61cd572bdf34b31f41d0eba9e806f75dd7cfbc705355ac9b721d.jpg) +Figure 9. Layer Replacement Editing. First row shows our SLR output and new background. The second row shows results of our background layer replacement effects, fluid layer replacement effects, and replacement effects using motion instead of alpha Arrows specify the motion direction and speed. + +sical poetry—as if the Silver River fell from azure sky. For a downstream task like such texture attribute replacement, accurately segmentation of all fluid regions is necessary to reach plausible visual effects. However, it is hard to be accomplished by brute force segmentation with either motion prediction (Fluid Replacement Editing by Motion Threshold w/o SLR in the figure) or manually labelling. Since real-world fluid scenes contain complex context relations such as semi-transparency, different motion behaviour among fluid regions in one scene and so on. On the contrary, since our proposed surface-based layer representation (SLR) naturally disentangles scene into surface fluid and background, we can achieve such editing easily. + +# 6. Conclusion + +We propose a system for fluid animation from a still image. Specifically, for fluid texture, a neural network is designed to decompose the still image into surface liquids and background, which form a novel two-layered representation, SLR, to represent fluid scene. Compared to previous work, SLR can help better simulating the transparent fluids as well as other fluids scenes with complex surrounding. For motion estimation, we introduce surface-only fluid simulation (SFS), which provides a wider range of editing abilities on the fluid image. The system ensures physical reasonability and meanwhile skirting the scene-specific modelling problem. However, as our method works in a semi-supervised manner, without supervision from ideal decomposition or graphic-based rendering, the outputs may have some artifacts on boundaries. We expect more research on better disentangling and representing fluid surface with backgrounds, as well as more precise motion prediction. + +Acknowledgements. This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK. + +# References + +[1] Scott G Bardenhagen and Edward M Kober. The generalized interpolation material point method. CMES, 5:477-496, 2004. 3 +[2] Robert Bridson. Fluid simulation for computer graphics. AK Peters/CRC Press, 2015. 3 +[3] Yung-Yu Chuang, Dan B Goldman, Ke Colin Zheng, Brian Curless, David H Salesin, and Richard Szeliski. Animating pictures with stochastic motion textures. ACM TOG, 24:853-860, 2005. 1, 2 +[4] Vincenzo Citro, Paolo Luchini, Filippo Giannetti, and Franco Auteri. Efficient stabilization and acceleration of numerical simulation of fluid flows by residual recombination. J. Comput. Phys., 344:234-246, 2017. 3 +[5] Fang Da, David Hahn, Christopher Batty, Chris Wojtan, and Eitan Grinspun. Surface-only liquids. ACM TOG, 35:1-12, 2016. 3 +[6] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating landscape: self-supervised learning of decoupled motion and appearance for single-image video synthesis. ACM TOG, 38:175:1-175:19, 2019. 1, 2, 3, 7 +[7] L. A. Gatys, A. S. Ecker, and M. Bethge. Image style transfer using convolutional neural networks. In CVPR, 2016. 2 +[8] I.J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial networks. NeurIPS, 3:2672-2680, 2014. 2 +[9] Tavi Halperin, Hanit Hakim, Orestis Vantzos, Gershon Hochman, Netai Benaim, Lior Sassy, Michael Kupchik, Ofir Bibi, and Ohad Fried. Endless loops: detecting and animating periodic patterns in still images. ACM TOG, 40:1-12, 2021. 1, 7 +[10] Francis H Harlow. The particle-in-cell computing method for fluid dynamics. Methods Comput. Phys., 3:319-343, 1964. 3 +[11] Aleksander Holynski, Brian L Curless, Steven M Seitz, and Richard Szeliski. Animating pictures with eulerian motion fields. In CVPR, 2021. 1, 2, 3, 4, 7, 8 +[12] Yuanming Hu, Tzu-Mao Li, Luke Anderson, Jonathan Ragan-Kelley, and Frédo Durand. Taichi: a language for high-performance computation on spatially sparse data structures. ACM TOG, 38:201, 2019. 6 +[13] Libo Huang, Ziyin Qu, Xun Tan, Xinxin Zhang, Dominik L Michels, and Chenfanfu Jiang. Ships, splashes, and waves on a vast ocean. ACM TOG, 40:1-15, 2021. 3 +[14] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, 2017. 5 +[15] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, 2017. 2, 6 +[16] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 2 +[17] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 2 + +[18] ByungMoon Kim, Yingjie Liu, Ignacio Llamas, and Jaroslaw R Rossignac. Flowfixer: Using bfecc for fluid simulation. Technical report, Georgia Institute of Technology, 2005. 3 +[19] Theodore Kim, Nils Thürey, Doug James, and Markus Gross. Wavelet turbulence for fluid simulation. ACM TOG, 27:1-6, 2008. 3 +[20] Pierre-Yves Laffont, Zhile Ren, Xiaofeng Tao, Chao Qian, and James Hays. Transient attributes for high-level understanding and editing of outdoor scenes. ACM TOG, 33:1-11, 2014. 3 +[21] Thi-Ngoc-Hanh Le, Chih-Kuo Yeh, Ying-Chi Lin, and Tong-Yee Lee. Animating still natural images using warping. TOMM, 2022. 1 +[22] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In CVPR, 2018. 6 +[23] Chieh Hubert Lin, Hsin-Ying Lee, Yen-Chi Cheng, Sergey Tulyakov, and Ming-Hsuan Yang. Infinitygan: Towards infinite-resolution image synthesis. arXiv preprint, arXiv:2104.03963, 2021. 3 +[24] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In ICCV, 2021. 3 +[25] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In ECCV, 2018. 2 +[26] Erika Lu, Forrester Cole, Tali Dekel, Andrew Zisserman, William T Freeman, and Michael Rubinstein. Omnimatte: associating objects and their effects in video. In CVPR, 2021. 3 +[27] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. arXiv preprint, arXiv:2112.03051, 2021. 1, 2, 3, 6, 7, 8 +[28] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In CVPR, 2020. 4 +[29] Makoto Okabe, Yoshinori Dobashi, and Ken Anjyo. Animating pictures of water scenes using video retrieval. Vis. Comput., 34:347-358, 2018. 3 +[30] Junting Pan, Chengyu Wang, Xu Jia, Jing Shao, Lu Sheng, Junjie Yan, and Xiaogang Wang. Video generation from single semantic label map. In CVPR, 2019. 4 +[31] Ekta Prashnani, Maneli Noorkami, Daniel Vaquero, and Pradeep Sen. A phase-based approach for animating images using video examples. In Comput Graph Forum, 2017. 3 +[32] Erik Reinhard, Michael Adhikhmin, Bruce Gooch, and Peter Shirley. Color transfer between images. Comput. Graph. Appl., 21:34-41, 2001. 2 +[33] Arno Schödl, Richard Szeliski, David H Salesin, and Irfan Essa. Video textures. In SIGGRAPH, 2000. 2 +[34] Meng Sun, Allan D. Jepson, and Eugene Fume. Video input driven animation (VIDA). In ICCV, 2003. 2 +[35] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. 2 +[36] Yongning Zhu and Robert Bridson. Animating sand as a fluid. ACM TOG, 24:965-972, 2005. 3, 6 \ No newline at end of file diff --git a/2023/Simulating Fluids in Real-World Still Images/images.zip b/2023/Simulating Fluids in Real-World Still Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..44e3ea032b6519fe7fb35e3d61d6cd06ba7edc18 --- /dev/null +++ b/2023/Simulating Fluids in Real-World Still Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4082f557e765e94771b53aa17cd3dba8f3feebb5a5ac04776922c6b977afe2a2 +size 741125 diff --git a/2023/Simulating Fluids in Real-World Still Images/layout.json b/2023/Simulating Fluids in Real-World Still Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2e2856da4aea84e1f36de3cfc7baad3501e90c96 --- /dev/null +++ b/2023/Simulating Fluids in Real-World Still Images/layout.json @@ -0,0 +1,8613 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 157, + 103, + 436, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 103, + 436, + 121 + ], + "spans": [ + { + "bbox": [ + 157, + 103, + 436, + 121 + ], + "type": "text", + "content": "Simulating Fluids in Real-World Still Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 141, + 527, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 141, + 527, + 171 + ], + "spans": [ + { + "bbox": [ + 66, + 141, + 527, + 171 + ], + "type": "text", + "content": "Siming Fan" + }, + { + "bbox": [ + 66, + 141, + 527, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 66, + 141, + 527, + 171 + ], + "type": "text", + "content": ", Jingtan Piao" + }, + { + "bbox": [ + 66, + 141, + 527, + 171 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 66, + 141, + 527, + 171 + ], + "type": "text", + "content": ", Chen Qian" + }, + { + "bbox": [ + 66, + 141, + 527, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 141, + 527, + 171 + ], + "type": "text", + "content": ", Hongsheng Li" + }, + { + "bbox": [ + 66, + 141, + 527, + 171 + ], + "type": "inline_equation", + "content": "^{2,3,4\\boxtimes}" + }, + { + "bbox": [ + 66, + 141, + 527, + 171 + ], + "type": "text", + "content": ", Kwan-Yee Lin" + }, + { + "bbox": [ + 66, + 141, + 527, + 171 + ], + "type": "inline_equation", + "content": "^{2,3\\boxtimes}" + }, + { + "bbox": [ + 66, + 141, + 527, + 171 + ], + "type": "text", + "content": ", SenseTime Research, Shanghai AI Laboratory, The Chinese University of Hong Kong, CPII" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 173, + 468, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 173, + 468, + 185 + ], + "spans": [ + { + "bbox": [ + 123, + 173, + 468, + 185 + ], + "type": "text", + "content": "{fansiming,qianchen}@sensetime.com, 1155116308@link.cuhk.edu.hk," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 184, + 188, + 408, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 188, + 408, + 198 + ], + "spans": [ + { + "bbox": [ + 184, + 188, + 408, + 198 + ], + "type": "text", + "content": "hsli@ee.cuhk.edu.hk, junyilin@cuhk.edu.hk" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 48, + 205, + 142, + 331 + ], + "blocks": [ + { + "bbox": [ + 48, + 205, + 142, + 331 + ], + "lines": [ + { + "bbox": [ + 48, + 205, + 142, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 142, + 331 + ], + "type": "image", + "image_path": "12b0fd2234020a29b22095c20374d547abc6e5befb82d30306aea0c165030ea6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 340, + 546, + 364 + ], + "lines": [ + { + "bbox": [ + 46, + 340, + 546, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 546, + 364 + ], + "type": "text", + "content": "Figure 1. Given a still image and a coarse hint (mask and arrows) of motion as input (left), our model estimates fluid motion to generate an animating video (right). Holynski [11] is a state-of-the-art method. Please refer to the supplemental materials for video playing." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 142, + 206, + 547, + 335 + ], + "blocks": [ + { + "bbox": [ + 142, + 206, + 547, + 335 + ], + "lines": [ + { + "bbox": [ + 142, + 206, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 142, + 206, + 547, + 335 + ], + "type": "image", + "image_path": "65a5a608f1cb7fc964b626b8426cd82553465cad58e5080876b0256b469bc9a7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 377, + 192, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 377, + 192, + 390 + ], + "spans": [ + { + "bbox": [ + 143, + 377, + 192, + 390 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 403, + 290, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 403, + 290, + 679 + ], + "spans": [ + { + "bbox": [ + 45, + 403, + 290, + 679 + ], + "type": "text", + "content": "In this work, we tackle the problem of real-world fluid animation from a still image. The key of our system is a surface-based layered representation, where the scene is decoupled into a surface fluid layer and an impervious background layer with corresponding transparencies to characterize the composition of the two layers. The animated video can be produced by warping only the surface fluid layer according to the estimation of fluid motions and recombining it with the background. In addition, we introduce surface-only fluid simulation, a 2.5D fluid calculation, as a replacement for motion estimation. Specifically, we leverage triangular mesh based on a monocular depth estimator to represent fluid surface layer and simulate the motion with the inspiration of classic physics theory of hybrid Lagrangian-Eulerian method, along with a learnable network so as to adapt to complex real-world image textures. Extensive experiments indicate our method's competitive performance for common fluid scenes and better robustness and reasonability under complex transparent fluid scenarios. Moreover, as proposed surface-based layer representation and surface-only fluid simulation naturally disentangle the scene, interactive editing such as adding objects and texture replacing could be easily achieved with" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 379, + 539, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 379, + 539, + 391 + ], + "spans": [ + { + "bbox": [ + 306, + 379, + 539, + 391 + ], + "type": "text", + "content": "realistic results. Code, and dataset are publicly available." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 414, + 387, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 414, + 387, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 414, + 387, + 426 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 434, + 546, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 546, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 546, + 589 + ], + "type": "text", + "content": "Given an image of a scene containing liquid such as streams and waterfalls, humans can imagine how the liquid in the still image would move. The problem of animating fluid from a single image has become a blooming topic recently studied by a series of work [6, 11, 27, 9, 21]. Their general pipeline can be summarized as: (1) estimating optical flow that indicates the movement of liquid and (2) synthesizing future frame image based on the motion estimation. Although these methods achieve impressive visual effects on simple fluid regions, how to handle complex contents with challenging transparency, collisions, and thin structures in real-world scenes is still a challenging and open problem." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 589, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 589, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 589, + 547, + 696 + ], + "type": "text", + "content": "The key to the limitation of previous work is that the generation process of videos is modeled in a global manner. Specifically, the subsequent frames are generated based on warping existing textures on input images [3] or neural features [6]. Such operations regard the scene as a whole and warp all the contents in the scene together. For some complex transparent fluid cases, as shown in Figure 2(b), we can see through the water to observe the background rocks beneath the liquid. Textures on those static objects" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 206, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 206, + 704 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 206, + 704 + ], + "type": "text", + "content": "*Equal contribution and joint first authorship." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 66, + 704, + 149, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 704, + 149, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 704, + 149, + 714 + ], + "type": "text", + "content": "Corresponding authors." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 324, + 704, + 496, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 704, + 496, + 714 + ], + "spans": [ + { + "bbox": [ + 324, + 704, + 496, + 714 + ], + "type": "text", + "content": "Project page: https://slr-sfs.github.io/" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15922" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "are also improperly deformed and result in unnatural animations. To decompose the influence of motion on fluids and surrounding static objects, we propose a two-layer representation for scenes with fluids, namely Surface-based Layered Representation (SLR), which models the motion of the surface fluid and the rest background separately. To achieve such decomposition from the still image during inference, we propose a temporal-based training scheme, which facilitates the network to learn the decomposition with multiple supervisions over spatial and temporal dimensions of running fluid videos." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 207, + 289, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 289, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 289, + 615 + ], + "type": "text", + "content": "Another major difficulty in animating a single fluid image is predicting the motion flow of the fluid. There exist two types of methods. The first category is the interactive motion prediction methods. They require sparse hints (ie, labelling the velocity direction and relative amplitude) on random pixels of the liquid region as additional inputs. The feature clustering model [27] or convolutional network [11] is then used to generate a dense motion field. As the precision depends on the level of motion details a user can provide, such methods are time-costing in complex fluid scenes. The second category is automatic methods [6]. These works regard the task as a domain transformation problem, and use image-to-image translation techniques to predict the motion represented by 2-channel optical flows. Ambiguous velocity (e.g., a flat river surface can flow either to the left or to the right, as shown in Figure 2(a)) and tremendous changes of velocity in a period of time caused by the collisions with objects (as shown in Figure 2(c)) may pose great challenges to the above methods. To tackle the challenges, we introduce a simplified Navier-Stokes simulation into the motion estimation process, called Surface-only Fluid Simulation (SFS). Specifically, the initial motion prediction of the liquid region is firstly calculated through sparse user inputs and interpolation inside the whole region. In parallel, a 3D mesh is built based on a monocular depth estimator. Then, the N-S equation is solved on the mesh surfaces and based on the initial motion to obtain a simulated motion field in a short period of time. Finally, since the real thickness of the fluid is neglected during the simulation, a refinement step is expected to compensate for the motion offset at height. Thus, a DNN-based motion translator is applied subsequently to obtain the refined motion field that adapts naturally to the input image. With SFS, we can enforce precise controls on the movement of fluid." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "In summary, our work has three major contributions. (a) We propose a new and learnable representation, surface-based layered representation, which decomposes the fluid surface and the static objects in the scene to better synthesize the fluid animating videos from a single fluid image. (b) We design a surface-only fluid simulation to model the evolution of the image fluids with better visual effects. (c) Based on the proposed surface-based layered representation" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 547, + 194 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 547, + 194 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 547, + 194 + ], + "type": "image", + "image_path": "8f4a06c3ce854edc582cdb7bf66e242a650536f414d983e0696d568af4453a8b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 198, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 304, + 198, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 547, + 255 + ], + "type": "text", + "content": "Figure 2. Motivation. (a) Given a still input image, the motion may be ambiguous. (b) Animating via previous SOTA method [11] causes background rocks beneath the liquid improperly deformed. (c) Motion prediction methods such as [11, 27] cannot well capture motion changes caused by the collisions with objects." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 264, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 547, + 361 + ], + "type": "text", + "content": "and surface-only fluid simulation, image editing with fluids is achieved with realistic and vivid results. In addition, We can easily extend our fluid animation system to downstream editing tasks, thanks to the SLR and SFS designs in our framework. For example, we can augment objects in the fluid region or background region, and simulate the interactions between the fluids and objects. We can also replace textures to different layers to create different scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 373, + 394, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 394, + 385 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 394, + 385 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 394, + 411, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 411, + 406 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 411, + 406 + ], + "type": "text", + "content": "2.1. Video Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 413, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 557 + ], + "type": "text", + "content": "In early studies, researchers focus on generating new video context by changing the attributes of recorded frames [32, 33]. Traditional methods of video motion generation with movable substances are based on stochastic process analysis [34] and color transfer [32]. Given a reference video indicating the temporal changes of a scene, texture movement and sequential images can be generated using a patch aligned method and energy-based optimizing [33]. Similar methods are applied to single image input [3]. However, obvious artifacts exist, and the smoothness of the video is not guaranteed. These drawbacks make the usage of such methods limited to typical types of scene samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 714 + ], + "type": "text", + "content": "In recent years, since the development of deep neural networks in image generation field [15], plenty of innovative efforts have been made to improve the synthesis quality of images [7, 16, 8, 35, 25]. Techniques such as GANs [8], perceptual loss [17], partial convolution [25] and cycle-consistency [35] have been widely used as standard components. Built upon the success of image generation, several attempts have created the precedents of neural video generation. A few methods are developed to force the network to learn intermediate scene representation for predicting the future frames or interpolating key frames, once the input sequences and camera trajectory are given (note: as we focus on single image input, we skip the review of" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15923" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": "such multi-frames input trend). Given a single image as input, latent generative methods like [24] proposes an autoregressive model to generate infinite pixels forming sequential outputs. InfinityGAN [23] disentangles global appearances, local structures and textures with a structure synthesizer and texture synthesizer to generate images with spatial size and level of details not attainable before. [31] tries to use Fourier transforms in time variance showing the phase variations to represent the flow and generate moving videos. [20] spends efforts in digging into the semantic meanings in images and generates videos correspondingly to achieve a natural result. [6] decouples motion and appearance through learning intermediate flow fields and color transfer map. However, these methods are either focusing on spatial-wise consistency [20], or paying attention to coarse-level time-variance scene changes [31, 6]. None of them discusses or can handle the particularity of fluid with fine detail motions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 298, + 149, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 298, + 149, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 298, + 149, + 310 + ], + "type": "text", + "content": "2.2. Fluid Simulation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 317, + 289, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 289, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 289, + 544 + ], + "type": "text", + "content": "Physical Fluid Simulation. Traditional methods originated from computer graphics to achieve accurate approximations on various types of fluids. Hybrid Lagrangian-Eulerian Methods [2, 36, 10] have successfully simulated the motion of incompressible fluids in a pre-defined 2D or 3D box environment, which solves pressure projection on Eulerian grid and advect on Lagrangian particles. To improve convergence ability and flexibility on complex shapes, material point method (MPM) [1] has been popular and more subsequent methods [18, 19] have been proposed to stabilize the convergence when iterating velocities. However, these methods are time-consuming in large scenes. To reduce the cost of simulating a large-range of fluid, methods have been proposed to simulate only a small range of fluid [4], or the surface of the fluid [5, 13] to produce a faster approximation by avoiding high computational cost in 3D volumetric solvers. Although these works achieve amazing simulation effects, they are still limited to synthetic scenarios due to scene-specific modelling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 545, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 714 + ], + "type": "text", + "content": "Animating Fluids with Data Prior. Given a still image as animation target and a set of real-world water video candidates as animation bank, [31, 29] propose to retrieve a video or a group of videos from the bank that has similar content with the target image as the references. Then they transfer motion and appearance of references to each region of interest on target image to form the water animation video. These methods open the door of real-world fluid animation with data prior in-the-wild. However, the seamless alignment of animated results with source input image can not be well guaranteed and requiring lots of manual efforts. With the development of deep learning, recent works [11, 27] propose to automatically predict each frame of motion and appearance via learning from reconstruction loss" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "type": "text", + "content": "with real world videos for training. [11] first simplifies motion estimation part as a single frame motion prediction via motion Eulerian integration and then proposes a deep feature warping technique to narrow down the size of blank areas caused by warping. Finally, it adaptively blends forward and backward features through a learnable parameter " + }, + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "type": "text", + "content": " to generate animated fluid video. Based on [11], [27] proposes to regress fluid motion conditioned to user's sparse guidance and generate paired training data through motion speed clustering. The work further proposes to use multiscale representation to capture different fluid speed in different resolution. Although these methods take a step further to real-world scenes with impressive results, it is hard for them to handle complex context relations over fluids due to the single representation to the entire scene that regards different objects and textures as equal." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 264, + 548, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 548, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 548, + 502 + ], + "type": "text", + "content": "Unlike previous work, our system marries the advantages of both physical simulation and learning-based pipelines. For textures, to avoid interplay between fluids and surroundings, we force the network to decouple the scene into a new two-layer representation SLR, that includes a static background and time-variates surface fluid layers2. Meanwhile, the network can hallucinate reasonable and photo-realistic textures in vacated regions thanks to the large-scale data prior. For motion, to skirt scene-specific modelling while keeping physiC reasonability, we propose a three-step motion prediction. We first use sparse labels from the human interface to generate the initial rough motion and then calculate reasonable evolution in a short period by introducing a 2.5D surface simulation, SFS. Later, a smooth motion translator is used to refine the motion trend to better fit real-world examples with data prior. Moreover, as a consequence, we can generate natural fluid animations with the network training on plenty of real-world examples and flexibly edit the main liquid region and boundary condition to create various visual effects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 513, + 362, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 513, + 362, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 513, + 362, + 525 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "type": "text", + "content": "Given a real-world image that contains liquid, our goal is to generate a video clip that animates vivid flowing of the fluid. To achieve this goal, a system is required to (1) represent complex context relations between fluids and surrounding environments; and (2) to estimate motion fields from single image input with reasonability and efficiency. We propose a novel system with two main designs that satisfy the above requirements. For texture, in contrast with previous methods that regard all elements in the scene as an entirety, our system learns a surface-based layered representation (SLR) (Sec. 3.1) that decomposes the scene into surface fluid layer and impervious background layer. Since the tex" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 305, + 683, + 547, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 683, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 683, + 547, + 715 + ], + "type": "text", + "content": "2This idea shares a similar spirit with [26] in separating a scene into both static and dynamic parts. Whereas, our SLR could handle transparent fluids, and only require a single image as input." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15924" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "ture characteristics of surface fluid and rest background are assigned to different layers, such a design could overcome the negative influence of motion flows between fluids and surroundings (e.g., improper texture warping, as shown in Figure 2(b)). As for motion estimation, the system provides a three-step motion estimation alternative by introducing a surface-only fluid simulation (SFS) into the motion field prediction (Sec. 3.2). The motion prediction steps combine advantage of both physical- and learning-based motion estimation pipelines, thus ensuring physical reasonability and meanwhile skirting the scene-specific modelling problem. Each component is detailed in following subsections." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 223, + 246, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 223, + 246, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 246, + 236 + ], + "type": "text", + "content": "3.1. Learning of Decomposition of Images" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 241, + 287, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 287, + 324 + ], + "type": "text", + "content": "In this subsection, we first introduce Surface-based Layered Representation (SLR). Then, we show how to obtain animated videos through SLR. Finally, we present how to learn SLR with a neural network (illustrated in Figure 3), and challenges we meet during optimizing the network since there is no oracle ground-truth supervision for fluid scene decomposition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "text", + "content": "Surface-based Layered Representation. Given a still image " + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "inline_equation", + "content": "I(T_0)" + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "text", + "content": " as input, a series of RGB frames " + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = \\{I(T_0),\\dots,I(T_n)\\}" + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "text", + "content": " are outputs to synthesize time-space variations of the scene with running fluids. As we aim to handle complex context of fluid scene through disentangling fluid motion and still objects beneath/interacting with surface fluid, we propose to learn a novel intermediate representation whose output is an impervious background layer image " + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "inline_equation", + "content": "I_b" + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "text", + "content": ", a temporal sequence of surface fluid layer images " + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_f = \\{I_f(T_0),\\dots,I_f(T_n)\\}" + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "text", + "content": " and transparency factors " + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\alpha_{b},\\alpha_{f} = \\{\\alpha_{f}(T_{0}),\\dots,\\alpha_{f}(T_{n})\\}" + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\alpha_{*}\\in [0,1]" + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "text", + "content": ". Transparency factors indicate the contribution of colors on each layer to the final image at each time. In practice, we utilize a convolution network to output the simplification of the representation: " + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "inline_equation", + "content": "F_{\\theta}: (I(T_0),I(T_n),\\mathbf{M})\\to (I_b,I_f(T_0),I_f(T_n),\\alpha_b,\\alpha_f(T_0),\\alpha_f(T_n))" + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "text", + "content": ", where the network only outputs the estimated first and last frames of " + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_f" + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "text", + "content": " as well as " + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\alpha_{f}" + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "text", + "content": " is the estimated motion fields of " + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "inline_equation", + "content": "T_0" + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "inline_equation", + "content": "T_{1}" + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "text", + "content": " from the single input " + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "inline_equation", + "content": "I(T_0)" + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "text", + "content": ". We will detail the synthesis of " + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 325, + 287, + 563 + ], + "type": "text", + "content": " in Sec. 3.2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 563, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 653 + ], + "type": "text", + "content": "We expect the background to be identical across all times, while the fluid surface deforms with realistic motions. To this end, " + }, + { + "bbox": [ + 46, + 563, + 287, + 653 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 46, + 563, + 287, + 653 + ], + "type": "text", + "content": " is enforced to automatically map the original texture of non-liquid and static regions as well as the reasonable texture of under-liquid surface regions to the background layer (detailed in later subsection). As for fluid surface learning, recurrent estimation for all frames (RNN) is a possible alternative that can be tailored from video gen" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 225 + ], + "type": "text", + "content": "eration [30]. However, such an intuitive design might cause frame distortion as time goes on due to the accumulated errors and hard convergence (also demonstrated in [11]). In contrast, We use the linear combination of the first and last frames to construct the images at intermediate frames. This choice not only transforms the problem into interpolation rather than extrapolation, which is easier to converge, but also is able to generate an endless cyclic video rather than a video of a fixed interval. In detail, we tailor the symmetric softmax splatting, which is a pixel-warping strategy utilized in [11, 28], to our framework. We extend the splatting operation from a single image feature splatting into two levels – fluid surface feature splatting and fluid alpha splatting as follows :" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 312, + 229, + 545, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 229, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 312, + 229, + 545, + 255 + ], + "type": "interline_equation", + "content": "D _ {f} \\left(T _ {i}; \\hat {\\mathbf {x}} ^ {\\prime}\\right) = \\frac {\\sum_ {\\hat {\\mathbf {x}} \\in \\mathcal {X}} \\alpha \\left(T _ {i} ; \\hat {\\mathbf {x}}\\right) \\cdot D _ {f} \\left(T _ {j} ; \\hat {\\mathbf {x}}\\right) \\cdot \\exp \\left(Z _ {f} \\left(T _ {j}; \\hat {\\mathbf {x}}\\right)\\right)}{\\sum_ {\\hat {\\mathbf {x}} \\in \\mathcal {X}} \\alpha \\left(T _ {i}; \\hat {\\mathbf {x}}\\right) \\cdot \\exp \\left(Z _ {f} \\left(T _ {j}; \\hat {\\mathbf {x}}\\right)\\right)}, \\tag {1}", + "image_path": "e3b72dbf53d037fb95583b1576b4b56c266b2f1e9526530b7761e2466ea4d7ab.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 257, + 545, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 545, + 283 + ], + "type": "interline_equation", + "content": "\\alpha_ {f} ^ {\\prime} \\left(T _ {i}; \\hat {\\mathbf {x}} ^ {\\prime}\\right) = \\frac {\\sum_ {\\hat {\\mathbf {x}} \\in \\mathcal {X}} \\alpha \\left(T _ {i} ; \\hat {\\mathbf {x}}\\right) \\cdot \\alpha_ {f} ^ {\\prime} \\left(T _ {j} ; \\hat {\\mathbf {x}}\\right) \\cdot \\exp \\left(Z _ {f} \\left(T _ {j} ; \\hat {\\mathbf {x}}\\right)\\right)}{\\sum_ {\\hat {\\mathbf {x}} \\in \\mathcal {X}} \\alpha \\left(T _ {i} ; \\hat {\\mathbf {x}}\\right) \\cdot \\exp \\left(Z _ {f} \\left(T _ {j} ; \\hat {\\mathbf {x}}\\right)\\right)}, \\tag {2}", + "image_path": "cdf5eab4182f97e93c69490aae7655084952a7f22a12a073a13ee6cec7f2bf64.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 289, + 545, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 325 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 289, + 545, + 325 + ], + "type": "inline_equation", + "content": "j \\in \\{0, n\\}" + }, + { + "bbox": [ + 304, + 289, + 545, + 325 + ], + "type": "text", + "content": " is the start index or last index of sequences, indicating forward splatting the frame 0 or backward splatting the frame " + }, + { + "bbox": [ + 304, + 289, + 545, + 325 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 289, + 545, + 325 + ], + "type": "text", + "content": ", which means the value " + }, + { + "bbox": [ + 304, + 289, + 545, + 325 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 289, + 545, + 325 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 351, + 341, + 545, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 341, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 351, + 341, + 545, + 368 + ], + "type": "interline_equation", + "content": "j = \\left\\{\\begin{array}{l l}0,&\\text {i f} \\hat {x} ^ {\\prime} = M _ {0 \\rightarrow t} (\\hat {x}),\\\\n,&\\text {i f} \\hat {x} ^ {\\prime} = M _ {0 \\rightarrow t - N} (\\hat {x}).\\end{array}\\right. \\tag {3}", + "image_path": "669fa723b2a6d7beed6c181f7cf2d5db6e83760e7dab5294736fc2e2108bfee2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "inline_equation", + "content": "D_{f}(T_{i};\\hat{\\mathbf{x}}^{\\prime})" + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "content": " is the fused fluid feature value of target frame at pixel " + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\alpha_f^\\prime (T_i;\\hat{\\mathbf{x}}^\\prime)" + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "content": " is the transparency value of target frame at pixel " + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "inline_equation", + "content": "D_{f}(T_{i};\\hat{\\mathbf{x}})" + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "content": " represents value of " + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "inline_equation", + "content": "D_{f}(T_{i})" + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "content": " at pixel " + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\alpha_{f}(T_{i};\\hat{\\mathbf{x}})" + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\alpha_f^\\prime (T_i;\\hat{\\mathbf{x}})" + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "content": " are similar. During inference, we set " + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "inline_equation", + "content": "D_{f}(T_{n}) = D_{f}(T_{0})" + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\alpha_{f}(T_{n}) = \\alpha_{f}(T_{0})" + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "content": " since only single image is available. " + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "content": " is the set of pixels satisfying " + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "inline_equation", + "content": "M_{0\\rightarrow t}(\\hat{\\mathbf{x}}) = \\hat{\\mathbf{x}}^{\\prime}" + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "inline_equation", + "content": "M_{0\\rightarrow t - N}(\\hat{\\mathbf{x}}) = \\hat{\\mathbf{x}}^{\\prime}" + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "inline_equation", + "content": "M_{0\\rightarrow t}" + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "content": " is displacement map generated from " + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "content": " through euler integration [11, 28], and " + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\alpha (T_i;\\hat{\\mathbf{x}}^{\\prime})" + }, + { + "bbox": [ + 304, + 374, + 545, + 506 + ], + "type": "text", + "content": " is a temporal coefficient to fuse features from forward warping and backward warping to ensure endless looping:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 321, + 513, + 545, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 513, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 321, + 513, + 545, + 540 + ], + "type": "interline_equation", + "content": "\\alpha \\left(T _ {i}; \\hat {\\mathbf {x}} ^ {\\prime}\\right) = \\left\\{\\begin{array}{l l}\\frac {t}{n},&\\text {i f} \\hat {x} ^ {\\prime} = M _ {0 \\rightarrow t} (\\hat {x}),\\\\1 - \\frac {t}{n},&\\text {i f} \\hat {x} ^ {\\prime} = M _ {0 \\rightarrow t - n} (\\hat {x}).\\end{array}\\right. \\tag {4}", + "image_path": "aa1baf0fb3280a627d7139c3b010d69c80756e83e107393f10da808af8a23624.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 545, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 594 + ], + "type": "text", + "content": "Then a fluid feature decoder is applied to " + }, + { + "bbox": [ + 304, + 545, + 545, + 594 + ], + "type": "inline_equation", + "content": "D_{f}(T_{i})" + }, + { + "bbox": [ + 304, + 545, + 545, + 594 + ], + "type": "text", + "content": " to generate surface fluid image " + }, + { + "bbox": [ + 304, + 545, + 545, + 594 + ], + "type": "inline_equation", + "content": "I_{f}(T_{i})" + }, + { + "bbox": [ + 304, + 545, + 545, + 594 + ], + "type": "text", + "content": " and a alpha refinement network is applied to " + }, + { + "bbox": [ + 304, + 545, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\alpha_f'(T_i)" + }, + { + "bbox": [ + 304, + 545, + 545, + 594 + ], + "type": "text", + "content": " to generate transparency alpha " + }, + { + "bbox": [ + 304, + 545, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\alpha_{f}(T_{i})" + }, + { + "bbox": [ + 304, + 545, + 545, + 594 + ], + "type": "text", + "content": " to avoid vacant alpha particles." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Animating Fluids with SLR. Once we obtain the final surface fluid layer " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{f}(T_{i})" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " and transparency " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{f}(T_{i})" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ", the reconstructed final image " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "I(T_{i})" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " can be acquired by adding the impervious background layer " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{b}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " back to the surface fluid layer " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{f}(T_{i})" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{b}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ". Here we use separate " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " channels for background and fluid transparency predictions. This design helps to stabilize the training of decomposition. The core reason is that when a single fluid transparency " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{f}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " is warped according to the motion flow, it leaves a blank region on the boundary, and makes the gradients hard to be" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 664, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 664, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 664, + 287, + 713 + ], + "type": "text", + "content": "3The word \"impervious\" is not rigorous, as background layer is not only formed from static object but also non-surface fluid which contains air molecules in real environment. However, we assume the fluid is incompressible, and thus a little bit abuse the word \"impervious background\" to indicate the non-surface fluid regions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15925" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 547, + 318 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 547, + 318 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 547, + 318 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 547, + 318 + ], + "type": "image", + "image_path": "d3022add8b5dd7b327414a4ffeb8837e765a9c118d7b06f824d305d6927933d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "lines": [ + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": "Figure 3. Training and testing pipelines. The background layer and background alpha are predicted only at the first frame. Both fluid features and fluid alpha are predicted at the first and last frames, and then symmetricly splatted and blended using the softmax splatting at frame " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": ". A decoder is used to decode fluid features into images and refined alpha. For frame " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "T_{0}" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": ", motion is calculated following pipeline Figure 4, and for frame " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": ", motion is calculated through Euler integration. During inference, " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "T_{n}" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": " is replaced by " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "T_{0}" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": " to create a cyclic video." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 373, + 288, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 288, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 288, + 410 + ], + "type": "text", + "content": "backpropagated. Although with the two-channel implementation, there still is a " + }, + { + "bbox": [ + 46, + 373, + 288, + 410 + ], + "type": "inline_equation", + "content": "\\alpha_{b}" + }, + { + "bbox": [ + 46, + 373, + 288, + 410 + ], + "type": "text", + "content": " on that vacated regions to fascinate the network refining the texture. In formula we have" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 103, + 418, + 287, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 418, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 103, + 418, + 287, + 445 + ], + "type": "interline_equation", + "content": "I \\left(T _ {i}\\right) = \\frac {\\alpha_ {f} \\left(T _ {i}\\right) I _ {f} \\left(T _ {i}\\right) + \\alpha_ {b} I _ {b}}{\\alpha_ {f} \\left(T _ {i}\\right) + \\alpha_ {b}} \\tag {5}", + "image_path": "fe63ae3698a1a0b42193af1eca62456c061ff655425708a16d88286bb2911541.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 450, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 545 + ], + "type": "text", + "content": "For corresponding network architectures to achieve the above targets, as shown in Fig 3, the encoder is used to extract background texture and " + }, + { + "bbox": [ + 46, + 450, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 450, + 287, + 545 + ], + "type": "text", + "content": " channel as well as surface fluid layers' features and its " + }, + { + "bbox": [ + 46, + 450, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 450, + 287, + 545 + ], + "type": "text", + "content": ". The decoder is used to refine the fused fluid layer that combines the layers from the first and last frames to obtain the complete surface fluid layer, and the final animated image is reconstructed by compositing background and surface fluid layers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "Optimizing the SLR. The training process is divided into three parts, in which, we separately train (1) the surface fluid branch containing the encoder that predicts the fluid feature layer from image input and the decoder that reconstructs the warped fluid layer, (2) the background branch that predicts the background layer from the image input, and (3) jointly train them together with " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " learning and surface fluid/ background branches finetuning. For each training step, the paired training data " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "< (I(T_0), I(T_n)), I(T_i) >" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " are randomly picked among 60-frames sequence without constraint on interval. We use dense optical flow estimation network [14] to extract reasonable movement of each video as pseudo motion ground-truth, and to ensure smoothness, we average the flow with window size of 30 frames. Due" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 373, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 545, + 398 + ], + "type": "text", + "content": "to space limits, we detail the training of each stage in the supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 406, + 443, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 443, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 443, + 418 + ], + "type": "text", + "content": "3.2. Motion Field Estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 426, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 472 + ], + "type": "text", + "content": "In the previous section, we have described the core necessities to learn an SLR but skipped the synthesis process of one of its input, motion fields " + }, + { + "bbox": [ + 304, + 426, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 426, + 545, + 472 + ], + "type": "text", + "content": ". In this section, we discuss how to obtain " + }, + { + "bbox": [ + 304, + 426, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 426, + 545, + 472 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "content": "During training, the motion of each frame can be generated from a conventional optical flow estimator [14]. During inference, as the input is a single image, we expect to synthesize a reasonable motion map from the observed scene. We propose a motion estimation process that separates the motion prediction into three parts taking advantage of both physical-based and learning-based pipelines for fluid simulation. First, interactive labelling of the sparse velocity of several pixels on the image is required to guide the generated motion directions of " + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "inline_equation", + "content": "I(T_0)" + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "content": ", along with a fusion with weights calculated based on Gaussian filtering of the distance maps from the labeled pixel to build a dense velocity initial map. Second, a surface-only fluid simulation is calculated based on initial motion and estimated scene depth to simulate a following convincing evolution of the liquid. Note that this step only calculates for the first frame. Last, a DNN-based pixel-to-pixel motion translator is utilized to refine the flow to create a more realistic and natural motion map. This translator could be considered as generating a compensation for a detailed motion map that is" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15926" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 71, + 181, + 125 + ], + "blocks": [ + { + "bbox": [ + 53, + 76, + 70, + 298 + ], + "lines": [ + { + "bbox": [ + 53, + 76, + 70, + 298 + ], + "spans": [ + { + "bbox": [ + 53, + 76, + 70, + 298 + ], + "type": "text", + "content": "6uizisauus 6uizisauus pue uonnnsr (e)" + } + ] + } + ], + "index": 0, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 71, + 181, + 125 + ], + "lines": [ + { + "bbox": [ + 75, + 71, + 181, + 125 + ], + "spans": [ + { + "bbox": [ + 75, + 71, + 181, + 125 + ], + "type": "image", + "image_path": "9f75a0df9b2dfe7966d6e6e2026c7ada309e5193716dddf1816587371302d9ed.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 181, + 71, + 287, + 128 + ], + "blocks": [ + { + "bbox": [ + 181, + 71, + 287, + 128 + ], + "lines": [ + { + "bbox": [ + 181, + 71, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 181, + 71, + 287, + 128 + ], + "type": "image", + "image_path": "ff40a85ade35dcae10f018149db4dca6fee956f16d88fd2f044c6914eb4909f0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 129, + 277, + 135 + ], + "lines": [ + { + "bbox": [ + 192, + 129, + 277, + 135 + ], + "spans": [ + { + "bbox": [ + 192, + 129, + 277, + 135 + ], + "type": "text", + "content": "2. Label Mask and Velocity Hint" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 76, + 137, + 145, + 171 + ], + "blocks": [ + { + "bbox": [ + 76, + 137, + 145, + 171 + ], + "lines": [ + { + "bbox": [ + 76, + 137, + 145, + 171 + ], + "spans": [ + { + "bbox": [ + 76, + 137, + 145, + 171 + ], + "type": "image", + "image_path": "38ba20ee3bd6a16b12cfe736e0ce1ee1d431ed8ed56b9d0b942b7409c9063224.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 148, + 137, + 216, + 171 + ], + "blocks": [ + { + "bbox": [ + 110, + 128, + 149, + 135 + ], + "lines": [ + { + "bbox": [ + 110, + 128, + 149, + 135 + ], + "spans": [ + { + "bbox": [ + 110, + 128, + 149, + 135 + ], + "type": "text", + "content": "1. Input Image" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 148, + 137, + 216, + 171 + ], + "lines": [ + { + "bbox": [ + 148, + 137, + 216, + 171 + ], + "spans": [ + { + "bbox": [ + 148, + 137, + 216, + 171 + ], + "type": "image", + "image_path": "641bdf13877f0b84f955c378646caa2c5eff228b0dbeb392703208ab91049648.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 172, + 208, + 177 + ], + "lines": [ + { + "bbox": [ + 154, + 172, + 208, + 177 + ], + "spans": [ + { + "bbox": [ + 154, + 172, + 208, + 177 + ], + "type": "text", + "content": "4. Mono-Depth Estimation" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 218, + 137, + 287, + 171 + ], + "blocks": [ + { + "bbox": [ + 218, + 137, + 287, + 171 + ], + "lines": [ + { + "bbox": [ + 218, + 137, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 218, + 137, + 287, + 171 + ], + "type": "image", + "image_path": "145265d27db8f6cafa89256a9819cbb45795e7a0b25c033a601a25f9857dcf07.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 172, + 271, + 177 + ], + "lines": [ + { + "bbox": [ + 236, + 172, + 271, + 177 + ], + "spans": [ + { + "bbox": [ + 236, + 172, + 271, + 177 + ], + "type": "text", + "content": "5. Build 3D Mesh" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 76, + 178, + 145, + 215 + ], + "blocks": [ + { + "bbox": [ + 82, + 172, + 138, + 178 + ], + "lines": [ + { + "bbox": [ + 82, + 172, + 138, + 178 + ], + "spans": [ + { + "bbox": [ + 82, + 172, + 138, + 178 + ], + "type": "text", + "content": "3. Generate Dense Velocity" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 76, + 178, + 145, + 215 + ], + "lines": [ + { + "bbox": [ + 76, + 178, + 145, + 215 + ], + "spans": [ + { + "bbox": [ + 76, + 178, + 145, + 215 + ], + "type": "image", + "image_path": "f279058ea667c6d55a7dc890ea243467c79ea322c4a23da6aadd0544f0f915bb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 216, + 135, + 222 + ], + "lines": [ + { + "bbox": [ + 82, + 216, + 135, + 222 + ], + "spans": [ + { + "bbox": [ + 82, + 216, + 135, + 222 + ], + "type": "text", + "content": "6. Mesh-Based Simulation" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 148, + 178, + 216, + 216 + ], + "blocks": [ + { + "bbox": [ + 148, + 178, + 216, + 216 + ], + "lines": [ + { + "bbox": [ + 148, + 178, + 216, + 216 + ], + "spans": [ + { + "bbox": [ + 148, + 178, + 216, + 216 + ], + "type": "image", + "image_path": "cb7bc957b4e56a7b68ff354cadf2ad58acddd720b29274166606f09d02790e17.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 216, + 198, + 222 + ], + "lines": [ + { + "bbox": [ + 164, + 216, + 198, + 222 + ], + "spans": [ + { + "bbox": [ + 164, + 216, + 198, + 222 + ], + "type": "text", + "content": "7. Refine Velocity" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 218, + 178, + 286, + 215 + ], + "blocks": [ + { + "bbox": [ + 218, + 178, + 286, + 215 + ], + "lines": [ + { + "bbox": [ + 218, + 178, + 286, + 215 + ], + "spans": [ + { + "bbox": [ + 218, + 178, + 286, + 215 + ], + "type": "image", + "image_path": "a9caf5cbf6b56b3d70b51435e3c8b1c685f4331ecff0be76fab467a68b5d7d2b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 216, + 273, + 221 + ], + "lines": [ + { + "bbox": [ + 230, + 216, + 273, + 221 + ], + "spans": [ + { + "bbox": [ + 230, + 216, + 273, + 221 + ], + "type": "text", + "content": "8. Synthesize Result" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 77, + 223, + 145, + 257 + ], + "blocks": [ + { + "bbox": [ + 77, + 223, + 145, + 257 + ], + "lines": [ + { + "bbox": [ + 77, + 223, + 145, + 257 + ], + "spans": [ + { + "bbox": [ + 77, + 223, + 145, + 257 + ], + "type": "image", + "image_path": "6fa6e7a9dad637d9e9b745aa4f1cd8ceb32d87b1e7a806558a06a88820406200.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 258, + 131, + 263 + ], + "lines": [ + { + "bbox": [ + 89, + 258, + 131, + 263 + ], + "spans": [ + { + "bbox": [ + 89, + 258, + 131, + 263 + ], + "type": "text", + "content": "9. Editing AR objects" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 148, + 223, + 215, + 257 + ], + "blocks": [ + { + "bbox": [ + 148, + 223, + 215, + 257 + ], + "lines": [ + { + "bbox": [ + 148, + 223, + 215, + 257 + ], + "spans": [ + { + "bbox": [ + 148, + 223, + 215, + 257 + ], + "type": "image", + "image_path": "9e09676cde456c99c2180fea141999763d3ed4cf80bc754826a694ab22dc6ce8.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 258, + 207, + 263 + ], + "lines": [ + { + "bbox": [ + 151, + 258, + 207, + 263 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 207, + 263 + ], + "type": "text", + "content": "10. Add object in Fluid Layer" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 218, + 223, + 286, + 257 + ], + "blocks": [ + { + "bbox": [ + 218, + 223, + 286, + 257 + ], + "lines": [ + { + "bbox": [ + 218, + 223, + 286, + 257 + ], + "spans": [ + { + "bbox": [ + 218, + 223, + 286, + 257 + ], + "type": "image", + "image_path": "ed95a40ef165f10778aa5e4e356a7caa37a6c6e001319ebe3fd53364d296d4fe.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 258, + 285, + 263 + ], + "lines": [ + { + "bbox": [ + 217, + 258, + 285, + 263 + ], + "spans": [ + { + "bbox": [ + 217, + 258, + 285, + 263 + ], + "type": "text", + "content": "1. Add object in Background layer" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 76, + 264, + 145, + 297 + ], + "blocks": [ + { + "bbox": [ + 76, + 264, + 145, + 297 + ], + "lines": [ + { + "bbox": [ + 76, + 264, + 145, + 297 + ], + "spans": [ + { + "bbox": [ + 76, + 264, + 145, + 297 + ], + "type": "image", + "image_path": "a33d82210fdda115b7af5a5dceda58b9b557fa8a755528ec77f37bb3e1ab249e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 297, + 143, + 301 + ], + "lines": [ + { + "bbox": [ + 75, + 297, + 143, + 301 + ], + "spans": [ + { + "bbox": [ + 75, + 297, + 143, + 301 + ], + "type": "text", + "content": "12. Mark Intersection Vertex \"Solid" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 310, + 287, + 376 + ], + "lines": [ + { + "bbox": [ + 46, + 310, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 287, + 376 + ], + "type": "text", + "content": "Figure 4. Motion calculation pipeline. (a) is initialization of fluid region, rock region, motion, and Mesh. (a,b) combine to form our 2.5D simulation and animation pipeline, where (4-5) is not included if using 2D simulation method. (a,c) combine to form our editable simulation and synthesis pipeline, where the user can edit 3D objects and create corresponding motion." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 148, + 264, + 216, + 297 + ], + "blocks": [ + { + "bbox": [ + 148, + 264, + 216, + 297 + ], + "lines": [ + { + "bbox": [ + 148, + 264, + 216, + 297 + ], + "spans": [ + { + "bbox": [ + 148, + 264, + 216, + 297 + ], + "type": "image", + "image_path": "885ccab6f9e04014b188dabc5b45106d9255102c1956cfe3b2e7e43344d5247e.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 297, + 211, + 302 + ], + "lines": [ + { + "bbox": [ + 166, + 297, + 211, + 302 + ], + "spans": [ + { + "bbox": [ + 166, + 297, + 211, + 302 + ], + "type": "text", + "content": "Mesh-Based Simulation" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 218, + 264, + 286, + 297 + ], + "blocks": [ + { + "bbox": [ + 218, + 264, + 286, + 297 + ], + "lines": [ + { + "bbox": [ + 218, + 264, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 218, + 264, + 286, + 297 + ], + "type": "image", + "image_path": "bf57bb705454dc24c6444a0c7c7f97f25e4ca55063a174d0a7cf498e5083ff37.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 297, + 280, + 302 + ], + "lines": [ + { + "bbox": [ + 220, + 297, + 280, + 302 + ], + "spans": [ + { + "bbox": [ + 220, + 297, + 280, + 302 + ], + "type": "text", + "content": "14. Edited Synthesized Result" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 46, + 382, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 287, + 407 + ], + "type": "text", + "content": "harmonic with fluid textures. The pipeline is illustrated in Figure 4." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "type": "text", + "content": "Initial Motion from Interactive Sparse Labeling. To obtain determined as well as editable motion direction, we choose an interactive alternative to generate initial flow for image " + }, + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "type": "inline_equation", + "content": "I(T_0)" + }, + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "type": "text", + "content": " following [27]. With no more than 10 discrete notations on images, and a fixed segmentation of the liquid region, we use nearest-neighbors-averaging to assign all the pixels that are inside the liquid region an initial velocity." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 46, + 491, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 658 + ], + "type": "text", + "content": "Surface-only Fluid Simulation. After the initial state is estimated, we tailor the traditional graphics pipeline with discretization and linear equation to simulate the follow-up evolution of the liquid. Intuitively, we can directly apply simulation on 2D grids that align with image pixels by initializing each pixel as an Eulerian grid and several Lagrangian particles near each pixel. However, since such perspective projection of the 3D scene may cause severe distortion when direct manipulating the motion in 2D, as shown in Figure 7, we thus turn to use a mono-depth estimator [22] to estimate the 3D surface of the liquid. After the 3D surface is constructed inside the region of the liquid, a mesh surface-based velocity is calculated, with the restriction of projected velocity:" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 59, + 657, + 287, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 657, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 59, + 657, + 287, + 696 + ], + "type": "interline_equation", + "content": "\\frac {d}{d t} \\left[ \\begin{array}{l} u \\\\ v \\end{array} \\right] = \\left[ \\begin{array}{c c} f x & 0 \\\\ 0 & f y \\end{array} \\right] \\left[ \\begin{array}{c c c} 1 / z & 0 & - x / z ^ {2} \\\\ 0 & 1 / z & - y / z ^ {2} \\end{array} \\right] \\left[ \\begin{array}{c} \\frac {d x}{d t} \\\\ \\frac {d y}{d t} \\\\ \\frac {d z}{d t} \\end{array} \\right] \\tag {6}", + "image_path": "16328a946eb78206ea5adebc7704666ae8273c078ecb7142311fdb8e06b36101.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "fx, fy, cx, cy" + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": " is camera intrinsic parameters, " + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "u, v" + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 391, + 114 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 391, + 114 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 391, + 114 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 391, + 114 + ], + "type": "image", + "image_path": "8e2bc8d256c01b1d907d1362cbb75cec30df8e4b2a94a9c4a8074026b7e87162.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 392, + 71, + 466, + 114 + ], + "blocks": [ + { + "bbox": [ + 392, + 71, + 466, + 114 + ], + "lines": [ + { + "bbox": [ + 392, + 71, + 466, + 114 + ], + "spans": [ + { + "bbox": [ + 392, + 71, + 466, + 114 + ], + "type": "image", + "image_path": "f4f3659c2596b56d5aca88aa5b0fad3dbd2deeb6e229609166a5399e73c8a775.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 468, + 71, + 545, + 114 + ], + "blocks": [ + { + "bbox": [ + 468, + 71, + 545, + 114 + ], + "lines": [ + { + "bbox": [ + 468, + 71, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 468, + 71, + 545, + 114 + ], + "type": "image", + "image_path": "c7e56c13e2b0350c9f6544a503ea131c67bddcb7eec7fdf51eb2a3bfd7dc9ae0.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 307, + 116, + 392, + 160 + ], + "blocks": [ + { + "bbox": [ + 307, + 116, + 392, + 160 + ], + "lines": [ + { + "bbox": [ + 307, + 116, + 392, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 116, + 392, + 160 + ], + "type": "image", + "image_path": "2793bf549947b7c855f27ccaa441b1271f533b5e07813fb1f12ea4ce8e0f5286.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 163, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 305, + 163, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 305, + 163, + 545, + 208 + ], + "type": "text", + "content": "Figure 5. Examples for Different Fluid Scenes. (a) The surface thickness of the waterfall and river are gauzy. (b) The fountain cannot be considered a surface. (c) Fluid with vortex contains complex motion." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 414, + 115, + 444, + 148 + ], + "blocks": [ + { + "bbox": [ + 414, + 115, + 444, + 148 + ], + "lines": [ + { + "bbox": [ + 414, + 115, + 444, + 148 + ], + "spans": [ + { + "bbox": [ + 414, + 115, + 444, + 148 + ], + "type": "image", + "image_path": "a3397f45c2a8e229f0d7b5a17ec9c61dd74fef053035c366b54bb5ce64d3485f.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 156, + 444, + 160 + ], + "lines": [ + { + "bbox": [ + 414, + 156, + 444, + 160 + ], + "spans": [ + { + "bbox": [ + 414, + 156, + 444, + 160 + ], + "type": "text", + "content": "(b)Not a Surface" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 468, + 115, + 545, + 158 + ], + "blocks": [ + { + "bbox": [ + 468, + 115, + 545, + 158 + ], + "lines": [ + { + "bbox": [ + 468, + 115, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 468, + 115, + 545, + 158 + ], + "type": "image", + "image_path": "bb145d3a50eb212c96f65a3cf2513508f0b9399a40543016b34cc91f90dc5882.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 158, + 527, + 162 + ], + "lines": [ + { + "bbox": [ + 485, + 158, + 527, + 162 + ], + "spans": [ + { + "bbox": [ + 485, + 158, + 527, + 162 + ], + "type": "text", + "content": "(c)Complex Initialization" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "bbox": [ + 304, + 211, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 211, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 211, + 545, + 462 + ], + "type": "text", + "content": "the velocity on the projected image and " + }, + { + "bbox": [ + 304, + 211, + 545, + 462 + ], + "type": "inline_equation", + "content": "x, y, z" + }, + { + "bbox": [ + 304, + 211, + 545, + 462 + ], + "type": "text", + "content": " the 3D position in the scene. Since we have modelled the liquid as a surface, we replace traditional Laplacian operator with Laplacian-Beltrami operator on the mesh and use Mixed-Euler-Lagrange approach [36] implemented in Taichi [12] to simulate consequent motion of each vertex. Detailed formula derivation can be seen in the supplementary. The particle binding on the surface is randomly sampled according to the surface area. When a particle is moved outside the mesh surface, we use the nearest projection to force the position to fall onto the mesh. For the boundary of the mesh, we calculate the velocity of the boundary triangle, judging whether the velocity on the edge boundary is inward or outward. For outward edges, the particle is abandoned after moving out of the liquid region. For inward region, a particle source is built to maintain the overall number of particles to be roughly stable. With the texture of surface fluid layer binding to each vertex, we can warp fluid layer with a soft rasterizer to simulate flows. In this way, given arbitrary flow initial state, a stable velocity can be calculated to approximate real-world flow after a period of simulation." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "text", + "content": "Smooth Motion Translator. With the help of above " + }, + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "inline_equation", + "content": "2.5D" + }, + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "text", + "content": " fluid simulation, we can generate a time-variant velocity among the sequence. However, the simulated velocity is still a rough trend since it only works on fluid surface without considering thickness variation of fluids. Thus, the SFS can work well in cases like Figure 5(a). It fails in cases with distinct depth measurement differences, as shown in the Figure 5(b) due to the lack of height information during calculation. Also, It is fragile to cases with complex initial state as shown Figure 5(c) due to error-prone/time-consuming interactive initial motion labeling for such cases. Thus, we need extra scheme to compensate or rectify for detailed motion that is harmonic with the fluid textures. To this end, we use a convolution-based network to transfer the simulated motion into a more detailed and reasonable one, which is pixel-aligned with image textures. The network is trained as a traditional style-transfer task, with a L2 loss between output velocity and ground truth. An external patch discriminator is used to refine the pattern of the velocity distribution. All the settings are similar to the image translation task pipeline [15]." + } + ] + } + ], + "index": 45 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15927" + } + ] + } + ], + "index": 46 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 152 + ], + "type": "text", + "content": "In this section, we present experiments and detailed analysis to demonstrate the contributions of our method both quantitatively and qualitatively. For all the cases shown in this paper, please refer to our demo video and project page for better dynamic visual effects." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 158, + 222, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 158, + 222, + 171 + ], + "spans": [ + { + "bbox": [ + 47, + 158, + 222, + 171 + ], + "type": "text", + "content": "4.1. Quality of Fluid Video Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 177, + 287, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 177, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 287, + 236 + ], + "type": "text", + "content": "Implementation Details. To learn SLR, we train our network on a series of videos containing a large region of fluid movement and still background [11]. For training of motion refinement network, we use the same settings in [27]. All networks are trained at a resolution of " + }, + { + "bbox": [ + 46, + 177, + 287, + 236 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 177, + 287, + 236 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 236, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 287, + 392 + ], + "type": "text", + "content": "Datasets and Evaluation Metrics. In addition to validation set in [11], we collect extra 126 scenes to form a new test set, which we called Complex Liquid Animation in-the-Wild (CLAW) test set, for evaluating the performance on both dense fluids (such as rivers and oceans) and sparse or transparent fluids (such as streams or splashes). Note that, in this subsection, the motion for all the methods to be compared is guided by the optical flow of the whole video to distinguish the performance of the synthesizing techniques. We use PSNR to show the whole average error, SSIM to show the errors in the region with plenty of textures and LPIPS, a perceptual based loss, to show visual quality difference. For more details, please refer to supplementary." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 48, + 396, + 287, + 459 + ], + "blocks": [ + { + "bbox": [ + 48, + 396, + 287, + 459 + ], + "lines": [ + { + "bbox": [ + 48, + 396, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 396, + 287, + 459 + ], + "type": "table", + "html": "
DatasetMethodsAll RegionFluid Region
LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑
HolysnikReproduced Holysnik0.079825.030.77870.065725.880.8007
CommonModified Holysnik(Baseline)0.079324.750.77580.065625.720.8000
Val SetOurs0.083425.140.77950.065726.100.8030
OurReproduced Holysnik0.206720.260.59550.202920.360.5961
CLAWModified Holysnik(Baseline)0.207819.970.59230.204120.100.5934
TestsetOurs0.204020.790.60800.197520.800.6077
", + "image_path": "c3373a7e1e3c48845d7076140da2f7bcd532a304b0a17cc0ba5c16d21fa6dcea.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 459, + 287, + 523 + ], + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 523 + ], + "type": "text", + "content": "Table 1. Quantitative comparison. (a) Results on Holynski's common val set [11], evaluating first to 60-th frames. \"Fluid Region\" means static region is replaced by input image during metric calculation for focusing on the statistic of fluid. This could be done by replacing the non-fluid region with the corresponding textures of input images. (b) Results on CLAW test set." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 529, + 287, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 650 + ], + "type": "text", + "content": "Quantitative Comparisons with baseline. Table 1 shows the results between our method and state-of-the-art method [11], which is a single layer learning-based system, a typical method that animates the scene in a global manner. We re-implement [11] in Pytorch (Reproduced Holynski), since the source code is not public available. For other comparison baselines, Modified Holynski is built upon Reproduced Holynski but with the modification of replacing convolution to partial convolution in the fluid decoder. The third model (Ours) is our SLR model4. Three observations can" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 577 + ], + "type": "text", + "content": "be included from Table 1: (1) PSNR and SSIM metrics of \"Modified Holynski\" drop a little compared to \"Reproduced Holynski\" in both statistics on all- and fluid region. We infer that, although partial convolution could better capture context than conventional one to impart reasonable texture in vacated regions in theory, simply replacing partial convolution into single-layer framework cannot essentially improve the image quality due to the wrong context raised from improper warping of both background and fluid. While, with SLR, the partial convolution can strengthen its advantage on context capturing for more proper fluid context textures provided in surface fluid layer. (2) The LPIPS of Ours model drops a little compared to the Modified Holynski model in the \"All Region\" of Holynski common validation set, but has nearly the same LPIPS in the \"Fluid Region\". For the other two metrics, our model leads to an uptick. Since most regions of Holynski's validation set are static, the blending of the fluid and background in the non-fluid region could lead to slight performance drops in terms of LPIPS. However, it could be eased by replacing the non-fluid region with the corresponding textures of input images, thus we care more about the performance in the \"Fluid Region\" instead of the \"All Region\". Such a phenomenon tells us that the two-layer decomposition design in Ours model does not harm the reconstruction ability on final images. To better unfold the analysis, we further conduct the close-up evaluations on Holynski's val set. Concretely, we crop the fluid region by motion threshold and only evaluate these visible fluid background regions. The LPIPS, PSNR, SSIM of Holynski's method are 0.230, 19.01, 0.565, and Ours are 0.227, 19.64, 0.581. This setting leads to larger performance margins with the baseline, which demonstrate the advantage of our system. (3) When evaluating only in the fluid region, we can see model Ours reaches a comparable result with the \"Modified Holynski\" model in Holynski validation set [11], but improves significantly in CLAW test set. The major reason behind this observation is that when it comes to complex scenes as included in our CLAW test set, our proposed two-layer representation suppresses the influence between surrounding and surface fluids. Thus we can achieve better results than single-layer baselines that regard all elements in the scene as an entirety during warping." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 582, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 715 + ], + "type": "text", + "content": "Comparisons with SOTA methods. In this paragraph, we provide comparisons with more state-of-the-art methods - Endo et.al [6], Tavi et.al [9], and Mahapatra et.al [27]. The first two methods are centered on animating general scenes, where Endo et.al [6] is designed for landscape, and Tavi et.al [9] focuses on animating periodic patterns in still images. These two methods could be applied to fluid scenarios. In contrast, Mahapatra et.al [27] is specialized for fluid scenarios, where a multi-scale representation is introduced to capture different fluid speeds in different resolutions. Concretely, as Endo's code is publicly available, we" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 655, + 287, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 655, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 655, + 287, + 714 + ], + "type": "text", + "content": "4For Ours, we use (Ours (stage 1) in Table 1 in our supplementary material.) as surface fluid model initialization. Since our training contains three stages, and we train 100 epochs at the first stage to get the model (Ours (stage 1)), then decay learning rate and fine-tune 50 epochs for fluid at the last stage in practice, we also apply the same training strategy for the first two baselines to ensure fair comparisons." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15928" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": "compare our system with it both quantitatively and qualitatively. For other methods, we compare them qualitatively in Fig 6. Specifically, the result of Mahapatra is from official project page, and result of Tavi is from official app. Endo's method can only control motion with a latent code. Thus, we test ten different latent codes and select the best one to compare. We observe that these three methods suffer from animation distortions with different levels – (1) from the zoom-in perspective, all baselines improperly move the still stones with the fluids. Specifically, Tavi's results have significant ghosting artifacts, which might be caused by its one-fold periodic motion and alpha-blending designs. (2) From the full image perspective, Endo's method has difficulty in estimating fluid motion. The river's motion is incorrectly predicted to be from right to left, and the waterfall flows upwards to the right. The quantitative results (Tab 1) provide the demonstration from another aspect. This phenomenon is foreseeable, as Endo's framework considers motion in a global manner. In contrast, our two-layer method animates plausible fluid without affecting still stones. Please refer to the supplementary material for more comparisons and straightforward video effects." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 338, + 169, + 479 + ], + "blocks": [ + { + "bbox": [ + 47, + 338, + 169, + 479 + ], + "lines": [ + { + "bbox": [ + 47, + 338, + 169, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 169, + 479 + ], + "type": "image", + "image_path": "c3b82864853e11bcd14c71f95201b96a64b759c2eb71a1633b4fa0b416b33d6d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 482, + 287, + 515 + ], + "lines": [ + { + "bbox": [ + 47, + 482, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 482, + 287, + 515 + ], + "type": "text", + "content": "Figure 6. Qualitatively Comparisons. (e-f) are tested in " + }, + { + "bbox": [ + 47, + 482, + 287, + 515 + ], + "type": "inline_equation", + "content": "288 \\times 512" + }, + { + "bbox": [ + 47, + 482, + 287, + 515 + ], + "type": "text", + "content": ", others are tested in " + }, + { + "bbox": [ + 47, + 482, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\mathrm{W} = 1280" + }, + { + "bbox": [ + 47, + 482, + 287, + 515 + ], + "type": "text", + "content": ". LPIPS/PSNR/SSIM of the whole video are colored in orange." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 171, + 338, + 287, + 479 + ], + "blocks": [ + { + "bbox": [ + 171, + 338, + 287, + 479 + ], + "lines": [ + { + "bbox": [ + 171, + 338, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 171, + 338, + 287, + 479 + ], + "type": "image", + "image_path": "b95a418ae96c45f6b2e1bcc9516afda9dffd34d0b5d87ee641691ea4a8937fbb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 521, + 287, + 555 + ], + "blocks": [ + { + "bbox": [ + 48, + 521, + 287, + 555 + ], + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 555 + ], + "type": "table", + "html": "
MethodsHolynski validation setCLAW test set
All RegionFluid RegionAll RegionFluid Region
LPIPS ↓PSNR ↑SSIM ↑LPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIM
Endo[6] HD0.35716.270.3510.12022.310.7410.35715.920.2960.34716.380.331
Ours HD0.14223.430.7150.11823.780.7280.19620.990.6110.19020.990.610
", + "image_path": "134f49c97552258fdbdf4c130f9e2df080fc70b66ac1ce656f8e68e677838a42.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": "User Study. We conduct a user study to compare visual effects between single layer baseline (Modified Holynski) and our two layers method(Ours in Table 1) on the full set of CLAW testset. For each scene, we animate 60 frames and output a two-second video. Animated videos are shown side by side without providing ground truth videos as reference to guide the participant focusing on the reasonability and photo-realistic. We ask the participants to select which one is better or neither is better, mainly according to three dimensions (photo-realistic, stereoscopic, and high-fidelity)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "content": "As shown in Table 3, the participants can make a distinction between the two methods over most samples, leav" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 428, + 173 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 428, + 173 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 428, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 428, + 173 + ], + "type": "image", + "image_path": "17eb2b988aa2970d3f65d355554e3032f162d2619919bd59b25f9cb49d04b567.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 183, + 547, + 250 + ], + "lines": [ + { + "bbox": [ + 304, + 183, + 547, + 250 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 547, + 250 + ], + "type": "text", + "content": "Figure 7. Different simulations. Orange arrows point out motion direction. All simulation methods use " + }, + { + "bbox": [ + 304, + 183, + 547, + 250 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 183, + 547, + 250 + ], + "type": "text", + "content": " weak incompression simulation instead of fully incompression for better visual effect. (row 1)Input image, sparse motion generated by sparse user hint required by each method. (row 2-4)ablation study of Mahapatra's method [27], 2D simulation and 2.5D SFS." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 430, + 70, + 547, + 174 + ], + "blocks": [ + { + "bbox": [ + 430, + 70, + 547, + 174 + ], + "lines": [ + { + "bbox": [ + 430, + 70, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 430, + 70, + 547, + 174 + ], + "type": "image", + "image_path": "595e45724f7d490db88b74f8e91a68fef9e6387900203677e022853619390eb4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 262, + 546, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 546, + 311 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 546, + 311 + ], + "type": "text", + "content": "ing " + }, + { + "bbox": [ + 304, + 262, + 546, + 311 + ], + "type": "inline_equation", + "content": "27.2\\%" + }, + { + "bbox": [ + 304, + 262, + 546, + 311 + ], + "type": "text", + "content": " samples difficult to tell. Our method achieves better scores than the single-layer method [11] with " + }, + { + "bbox": [ + 304, + 262, + 546, + 311 + ], + "type": "inline_equation", + "content": "52.3\\%" + }, + { + "bbox": [ + 304, + 262, + 546, + 311 + ], + "type": "text", + "content": " sample voting. This indicates the effectiveness of the proposed SLR in other aspects." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 326, + 314, + 526, + 338 + ], + "blocks": [ + { + "bbox": [ + 60, + 555, + 274, + 564 + ], + "lines": [ + { + "bbox": [ + 60, + 555, + 274, + 564 + ], + "spans": [ + { + "bbox": [ + 60, + 555, + 274, + 564 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons with Endo's method." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 314, + 526, + 338 + ], + "lines": [ + { + "bbox": [ + 326, + 314, + 526, + 338 + ], + "spans": [ + { + "bbox": [ + 326, + 314, + 526, + 338 + ], + "type": "table", + "html": "
Single layerOursNeither
E-O-N125(20.5%)319(52.3%)165(27.2%)
", + "image_path": "fe1d90ac1d8386dfb75d8d6365a261781035ef415cde6cc480ef0c3247596e35.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 340, + 546, + 394 + ], + "lines": [ + { + "bbox": [ + 304, + 340, + 546, + 394 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 546, + 394 + ], + "type": "text", + "content": "Table 3. User study. The subjective scores are voted on all samples of our CLAW testset, which includes 122 videos with various real-world fluids scenarios(e.g., river, stream and waterfall) and complex context surrounding (e.g., semi-/transparent liquid, collisions and thin structures)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 407, + 459, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 459, + 419 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 459, + 419 + ], + "type": "text", + "content": "4.2. Different Motion Prediction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": "Figure 7 shows different motion prediction pipelines. Pipeline 1: An interactive sparse labelling of the motion and then extending all the velocities to the whole moving region, which results in a constant flow that remains large gap with realistic motion. We can see that the motion appears reasonable global with a network refinement. However, for local regions around static rocks, the flow of liquid is improperly stiff and has no interactions with the rock. While in real-world scenes, there should be vortex and sprays caused by collisions. Pipeline 2: Built upon pipeline 1, while before the motion refinement stage, we additionally initialize grids with fluid, rock or air masks as well as placing particles around each grid, then advocate particles by edited motion for once. Later, a NS equation with the incompressible fluid assumption is solved on 2d grid to obtain the motion of the next frame, which we call \"Simulated 2D Motion\" in the figure. Considering pipeline 2 is simulating on a 2D image plane, a refinement step (using a smooth motion translator) is applied to compensate for the motion offset at height and fine details. Pipeline 3: instead of simulating on a 2D image plane, the N-S equation is solved in " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "2.5D" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": ". Specifically, we regress the depth with a pretrained monocular depth estimator and consequently build a 3D mesh over the fluid. Then, we solve the N-S equation on the mesh surface. With" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15929" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 286, + 161 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 286, + 161 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 286, + 161 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 286, + 161 + ], + "type": "image", + "image_path": "3cb386aad973956f81bfac3dd0c14d7a0a3b0d27f5eb143130268d01dfbd8612.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 163, + 287, + 195 + ], + "lines": [ + { + "bbox": [ + 46, + 163, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 287, + 195 + ], + "type": "text", + "content": "Figure 8. Interactive Editing. (Bottom) Initial all moving region as 'Fluid' statue. (Top) Initial 'Solid' statue at overlap area of mesh and new object." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 286, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 286, + 325 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 286, + 325 + ], + "type": "text", + "content": "the help of simulation, we can see in the left scene of the figure that the effect of turbulence around the rock can be animated, and the trend maintains the refined output. From the right scene in the figure, we can observe that the absolute value of the velocity is unnatural in Pipeline 2, which regards all particles on an identical plane. However, since the liquid surface leans in the picture, there should be a foreshortening effect on the projection. In contrast, pipeline3 samples particles on the surface and such simulation enable near-plane motion to carry more speed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 325, + 286, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 286, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 286, + 408 + ], + "type": "text", + "content": "Quantitative Comparisons. Tab 4 shows the ablations of motion refinement in terms of endpoint error metric (lower is better). We evaluate all scenes in both two datasets with user input limited to five sparse hints. The results prove motion refinement is essential for achieving a better ability to represent GT motion. Refer to supplementary for quantitative metrics of the three motion prediction pipelines." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 80, + 413, + 255, + 451 + ], + "blocks": [ + { + "bbox": [ + 80, + 413, + 255, + 451 + ], + "lines": [ + { + "bbox": [ + 80, + 413, + 255, + 451 + ], + "spans": [ + { + "bbox": [ + 80, + 413, + 255, + 451 + ], + "type": "table", + "html": "
MethodsDataset
Holynski Val SetCLAW Test Set
DenseMotion0.2990.596
RefinedMotion0.2690.582
", + "image_path": "165f55a44214b741e76712de9bf6484922f803d601b2db33648536eccaad6aeb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 452, + 263, + 460 + ], + "lines": [ + { + "bbox": [ + 67, + 452, + 263, + 460 + ], + "spans": [ + { + "bbox": [ + 67, + 452, + 263, + 460 + ], + "type": "text", + "content": "Table 4. Quantitative Motion Refinement Ablation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 466, + 121, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 121, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 121, + 479 + ], + "type": "text", + "content": "5. Byproducts" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 286, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 286, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 286, + 617 + ], + "type": "text", + "content": "Interactive Editing. We can augment fluid scene while keeping realistic fluid animation. For example, we can place an imaginary stone in a river that is originally slack water, and the flow could be updated to the new one with the effect of a vortex around the stone, as can be viewed in Simulation w/ Fluid-Rock Interaction of Figure 8. To achieve this, we only need to calculate interface of inserted stone mesh and fluid surface mesh. Then, we classify the part of stone above/beneath the river surface according to the depth. Later, fluid motion is updated with boundary conditions during simulation. Refer to supplementary for details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 712 + ], + "type": "text", + "content": "Layer Replacement Editing. We can change texture attributes of both surface fluid and background by simply replacing either fluid layer or background layer with another reference texture. For example, as shown in Figure 9, the fluid layer can be replaced by/regulated with a starry sky video with an endless loop to form dreamlike fluid effects that might remind the user of a Chinese clas" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 544, + 178 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 544, + 178 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 544, + 178 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 544, + 178 + ], + "type": "image", + "image_path": "788a76442a5b61cd572bdf34b31f41d0eba9e806f75dd7cfbc705355ac9b721d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 179, + 545, + 234 + ], + "lines": [ + { + "bbox": [ + 304, + 179, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 304, + 179, + 545, + 234 + ], + "type": "text", + "content": "Figure 9. Layer Replacement Editing. First row shows our SLR output and new background. The second row shows results of our background layer replacement effects, fluid layer replacement effects, and replacement effects using motion instead of alpha Arrows specify the motion direction and speed." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 243, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 398 + ], + "type": "text", + "content": "sical poetry—as if the Silver River fell from azure sky. For a downstream task like such texture attribute replacement, accurately segmentation of all fluid regions is necessary to reach plausible visual effects. However, it is hard to be accomplished by brute force segmentation with either motion prediction (Fluid Replacement Editing by Motion Threshold w/o SLR in the figure) or manually labelling. Since real-world fluid scenes contain complex context relations such as semi-transparency, different motion behaviour among fluid regions in one scene and so on. On the contrary, since our proposed surface-based layer representation (SLR) naturally disentangles scene into surface fluid and background, we can achieve such editing easily." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 405, + 378, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 405, + 378, + 417 + ], + "spans": [ + { + "bbox": [ + 306, + 405, + 378, + 417 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 426, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 628 + ], + "type": "text", + "content": "We propose a system for fluid animation from a still image. Specifically, for fluid texture, a neural network is designed to decompose the still image into surface liquids and background, which form a novel two-layered representation, SLR, to represent fluid scene. Compared to previous work, SLR can help better simulating the transparent fluids as well as other fluids scenes with complex surrounding. For motion estimation, we introduce surface-only fluid simulation (SFS), which provides a wider range of editing abilities on the fluid image. The system ensures physical reasonability and meanwhile skirting the scene-specific modelling problem. However, as our method works in a semi-supervised manner, without supervision from ideal decomposition or graphic-based rendering, the outputs may have some artifacts on boundaries. We expect more research on better disentangling and representing fluid surface with backgrounds, as well as more precise motion prediction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements. This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15930" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "type": "text", + "content": "[1] Scott G Bardenhagen and Edward M Kober. The generalized interpolation material point method. CMES, 5:477-496, 2004. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 145 + ], + "type": "text", + "content": "[2] Robert Bridson. Fluid simulation for computer graphics. AK Peters/CRC Press, 2015. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 287, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 287, + 190 + ], + "type": "text", + "content": "[3] Yung-Yu Chuang, Dan B Goldman, Ke Colin Zheng, Brian Curless, David H Salesin, and Richard Szeliski. Animating pictures with stochastic motion textures. ACM TOG, 24:853-860, 2005. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 192, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 287, + 235 + ], + "type": "text", + "content": "[4] Vincenzo Citro, Paolo Luchini, Filippo Giannetti, and Franco Auteri. Efficient stabilization and acceleration of numerical simulation of fluid flows by residual recombination. J. Comput. Phys., 344:234-246, 2017. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 267 + ], + "type": "text", + "content": "[5] Fang Da, David Hahn, Christopher Batty, Chris Wojtan, and Eitan Grinspun. Surface-only liquids. ACM TOG, 35:1-12, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 270, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 287, + 312 + ], + "type": "text", + "content": "[6] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating landscape: self-supervised learning of decoupled motion and appearance for single-image video synthesis. ACM TOG, 38:175:1-175:19, 2019. 1, 2, 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 314, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 334 + ], + "type": "text", + "content": "[7] L. A. Gatys, A. S. Ecker, and M. Bethge. Image style transfer using convolutional neural networks. In CVPR, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 335, + 287, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 287, + 368 + ], + "type": "text", + "content": "[8] I.J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial networks. NeurIPS, 3:2672-2680, 2014. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 369, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 423 + ], + "type": "text", + "content": "[9] Tavi Halperin, Hanit Hakim, Orestis Vantzos, Gershon Hochman, Netai Benaim, Lior Sassy, Michael Kupchik, Ofir Bibi, and Ohad Fried. Endless loops: detecting and animating periodic patterns in still images. ACM TOG, 40:1-12, 2021. 1, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 425, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 425, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 48, + 425, + 287, + 456 + ], + "type": "text", + "content": "[10] Francis H Harlow. The particle-in-cell computing method for fluid dynamics. Methods Comput. Phys., 3:319-343, 1964. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 458, + 287, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 458, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 48, + 458, + 287, + 490 + ], + "type": "text", + "content": "[11] Aleksander Holynski, Brian L Curless, Steven M Seitz, and Richard Szeliski. Animating pictures with eulerian motion fields. In CVPR, 2021. 1, 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 491, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 534 + ], + "type": "text", + "content": "[12] Yuanming Hu, Tzu-Mao Li, Luke Anderson, Jonathan Ragan-Kelley, and Frédo Durand. Taichi: a language for high-performance computation on spatially sparse data structures. ACM TOG, 38:201, 2019. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 536, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 567 + ], + "type": "text", + "content": "[13] Libo Huang, Ziyin Qu, Xun Tan, Xinxin Zhang, Dominik L Michels, and Chenfanfu Jiang. Ships, splashes, and waves on a vast ocean. ACM TOG, 40:1-15, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "type": "text", + "content": "[14] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, 2017. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 614, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 645 + ], + "type": "text", + "content": "[15] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, 2017. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "text", + "content": "[16] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[17] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 709 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[18] ByungMoon Kim, Yingjie Liu, Ignacio Llamas, and Jaroslaw R Rossignac. Flowfixer: Using bfecc for fluid simulation. Technical report, Georgia Institute of Technology, 2005. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 118, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 148 + ], + "type": "text", + "content": "[19] Theodore Kim, Nils Thürey, Doug James, and Markus Gross. Wavelet turbulence for fluid simulation. ACM TOG, 27:1-6, 2008. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 150, + 545, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 192 + ], + "type": "text", + "content": "[20] Pierre-Yves Laffont, Zhile Ren, Xiaofeng Tao, Chao Qian, and James Hays. Transient attributes for high-level understanding and editing of outdoor scenes. ACM TOG, 33:1-11, 2014. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 194, + 545, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 225 + ], + "type": "text", + "content": "[21] Thi-Ngoc-Hanh Le, Chih-Kuo Yeh, Ying-Chi Lin, and Tong-Yee Lee. Animating still natural images using warping. TOMM, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 227, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 227, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 307, + 227, + 545, + 258 + ], + "type": "text", + "content": "[22] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In CVPR, 2018. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 259, + 545, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 259, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 307, + 259, + 545, + 302 + ], + "type": "text", + "content": "[23] Chieh Hubert Lin, Hsin-Ying Lee, Yen-Chi Cheng, Sergey Tulyakov, and Ming-Hsuan Yang. Infinitygan: Towards infinite-resolution image synthesis. arXiv preprint, arXiv:2104.03963, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 304, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 346 + ], + "type": "text", + "content": "[24] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In ICCV, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 348, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 348, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 348, + 545, + 380 + ], + "type": "text", + "content": "[25] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In ECCV, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 381, + 545, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 381, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 307, + 381, + 545, + 422 + ], + "type": "text", + "content": "[26] Erika Lu, Forrester Cole, Tali Dekel, Andrew Zisserman, William T Freeman, and Michael Rubinstein. Omnimatte: associating objects and their effects in video. In CVPR, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 424, + 545, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 424, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 307, + 424, + 545, + 456 + ], + "type": "text", + "content": "[27] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. arXiv preprint, arXiv:2112.03051, 2021. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 458, + 545, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 458, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 307, + 458, + 545, + 479 + ], + "type": "text", + "content": "[28] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In CVPR, 2020. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 480, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 480, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 480, + 545, + 511 + ], + "type": "text", + "content": "[29] Makoto Okabe, Yoshinori Dobashi, and Ken Anjyo. Animating pictures of water scenes using video retrieval. Vis. Comput., 34:347-358, 2018. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 513, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 513, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 513, + 545, + 544 + ], + "type": "text", + "content": "[30] Junting Pan, Chengyu Wang, Xu Jia, Jing Shao, Lu Sheng, Junjie Yan, and Xiaogang Wang. Video generation from single semantic label map. In CVPR, 2019. 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "text", + "content": "[31] Ekta Prashnani, Maneli Noorkami, Daniel Vaquero, and Pradeep Sen. A phase-based approach for animating images using video examples. In Comput Graph Forum, 2017. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 579, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 610 + ], + "type": "text", + "content": "[32] Erik Reinhard, Michael Adhikhmin, Bruce Gooch, and Peter Shirley. Color transfer between images. Comput. Graph. Appl., 21:34-41, 2001. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 612, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 545, + 632 + ], + "type": "text", + "content": "[33] Arno Schödl, Richard Szeliski, David H Salesin, and Irfan Essa. Video textures. In SIGGRAPH, 2000. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 634, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 634, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 307, + 634, + 545, + 654 + ], + "type": "text", + "content": "[34] Meng Sun, Allan D. Jepson, and Eugene Fume. Video input driven animation (VIDA). In ICCV, 2003. 2" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 655, + 545, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 655, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 307, + 655, + 545, + 687 + ], + "type": "text", + "content": "[35] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. 2" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 689, + 545, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 689, + 545, + 709 + ], + "spans": [ + { + "bbox": [ + 307, + 689, + 545, + 709 + ], + "type": "text", + "content": "[36] Yongning Zhu and Robert Bridson. Animating sand as a fluid. ACM TOG, 24:965-972, 2005. 3, 6" + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15931" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Single Depth-image 3D Reflection Symmetry and Shape Prediction/fcff96a5-45a1-4f7e-ab9f-0c309af710cd_content_list.json b/2023/Single Depth-image 3D Reflection Symmetry and Shape Prediction/fcff96a5-45a1-4f7e-ab9f-0c309af710cd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4cd085946456e42122182cfbae36ca77bdb75877 --- /dev/null +++ b/2023/Single Depth-image 3D Reflection Symmetry and Shape Prediction/fcff96a5-45a1-4f7e-ab9f-0c309af710cd_content_list.json @@ -0,0 +1,1728 @@ +[ + { + "type": "text", + "text": "Single Depth-image 3D Reflection Symmetry and Shape Prediction", + "text_level": 1, + "bbox": [ + 148, + 130, + 821, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhaoxuan Zhang, Bo Dong*, Tong Li, Felix Heide, Pieter Peers, Baocai Yin*, Xin Yang*", + "bbox": [ + 137, + 180, + 836, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 233, + 313, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present Iterative Symmetry Completion Network (ISCNet), a single depth-image shape completion method that exploits reflective symmetry cues to obtain more detailed shapes. The efficacy of single depth-image shape completion methods is often sensitive to the accuracy of the symmetry plane. ISCNet therefore jointly estimates the symmetry plane and shape completion iteratively; more complete shapes contribute to more robust symmetry plane estimates and vice versa. Furthermore, our shape completion method operates in the image domain, enabling more efficient high-resolution, detailed geometry reconstruction. We perform the shape completion from pairs of viewpoints, reflected across the symmetry plane, predicted by a reinforcement learning agent to improve robustness and to simultaneously explicitly leverage symmetry. We demonstrate the effectiveness of ISCNet on a variety of object categories on both synthetic and real-scanned datasets.", + "bbox": [ + 75, + 265, + 473, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 550, + 209, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Symmetry is an intrinsic geometric property present in the vast majority of man-made and natural objects [12, 17], and which has been exploited to improve the efficacy of applications such as shape matching [12], segmentation [13], object retrieval [8], and robotic grasping [34].", + "bbox": [ + 75, + 575, + 468, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we aim to leverage symmetry as an aid for the completion of shapes from partial observations [23, 24, 28]. Recently, Yao et al. [39] presented Front2Back, a framework for 3D shape reconstruction from a single RGB image. A key observation in Front2Back is that a (nearly) complete 3D model can be effectively described by a pair of 2.5D visible surface images taken from opposite views (i.e., front and back views). Based on this observation, Front2Back completes a shape by synthesizing a back view using a global 3D reflective symmetry plane from a given front view. However, the reconstruction quality is significantly affected by the input front view angle (Figure 1(d)). Moreover, Front2Back assumes the availability of a 'perfect' symmetry plane, which in practice is challenging to obtain from partial observations [21, 22, 45].", + "bbox": [ + 75, + 652, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9f65dacff26d8142651f49b7c010a15e32a60362be274588248a08729da8b788.jpg", + "image_caption": [ + "(a) Input Depth Map & Observed Surface" + ], + "image_footnote": [], + "bbox": [ + 517, + 233, + 630, + 311 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/62b67ac4a0495c76f7fcfbc44cd1dee505e2450af71025e1dfd9e61a6cc15eff.jpg", + "image_caption": [ + "(b) PoinTr [40]" + ], + "image_footnote": [], + "bbox": [ + 694, + 229, + 764, + 310 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/45e3b261f8434c20c6712b131cc05a4f62a92430b17c52552817f01a327efc1e.jpg", + "image_caption": [ + "(c) IF-Net [3]" + ], + "image_footnote": [], + "bbox": [ + 805, + 232, + 870, + 311 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ac5b3c58597abcef32d4ecf2fb59d32e7da9043fdd5ea829b28a3d5eebaebe0e.jpg", + "image_caption": [ + "(d) Front2Back [39]" + ], + "image_footnote": [], + "bbox": [ + 570, + 345, + 640, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e8e205a5f71b82ebf239db0d51f22344d918beed4737b561c41bb74594d22542.jpg", + "image_caption": [ + "(e) Ours" + ], + "image_footnote": [], + "bbox": [ + 687, + 345, + 758, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/948cade7807be767324c0d2e9ab6b6c72db87805a993e9096ca22fd00a671f82.jpg", + "image_caption": [ + "(f) Ground Truth", + "Figure 1. ISCNet more efficiently leverages symmetry cues for robust shape completion from a single depth image compared to prior methods." + ], + "image_footnote": [], + "bbox": [ + 805, + 345, + 870, + 422 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present ISCNet (Iterative Shape Completion Network), a single depth-image shape completion method that explicitly leverages symmetry cues. Our key insight is that symmetry detection and symmetry-based shape completion are complementary tasks that can provide constructive cues to the other task. Symmetry detection can help shape completion by providing object-centric information about missing geometrical features, and shape completion can provide more robust symmetry cues by filling in missing geometry. Departing from existing work [33] that assumes an ideal prediction of the symmetry plane, ISCNet iteratively refines the symmetry plane and the reconstructed shape jointly. Furthermore, unlike Front2Back [39], ISCNet does not rely on the initial input viewpoint to a-priori fix the synthetic back view camera position but instead leverages a trained reinforcement learning (RL) policy to select optimal pairs of reflection viewpoints at each iteration. We empirically found that predetermining the viewpoints is less effective than RL-based viewpoint selection which can more effectively deal with the dynamic shape of a partial point cloud during the refinement process.", + "bbox": [ + 496, + 489, + 892, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Specifically, ISCNet takes as input a single depth image and reprojects it to an incomplete point cloud. Instead of explicitly detecting a symmetry plane, we estimate the pose [30] that aligns the symmetry plane to the X-Y plane. Based on the current estimate of the point cloud and symmetry aligning pose, an RL agent determines a pair of $re$ -", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Co-corresponding authors.", + "bbox": [ + 94, + 886, + 243, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8896", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "flection viewpoints, reflected across the symmetry plane, for extracting a pair of partial depth and normal maps that are subsequently repaired using a novel 2D inpainting module. By directly working on depth image pairs, our method is independent of the size of the point cloud, and it can leverage geometry information from the symmetric counter view. However, because the virtual depth images are reprojections that differ from the input view, not all reprojected points are valid (e.g., the backside of points visible through a 'hole'). Therefore, we first identify reliable pixels before performing depth inpainting in the projected image domain. Finally, we reproject the inpainted maps back to 3D and update the point cloud. We progressively refine the point cloud by alternating between estimating reflection viewpoints and completing the point cloud for 20 iterations, updating the symmetry plane every 4th iteration, to balance efficiency versus efficacy. When updating the symmetry plane, we 'reset' the point cloud to its initial state to mitigate the impact of prior inaccurate symmetry plane estimations on the shape completion.", + "bbox": [ + 75, + 90, + 472, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We validate that ISCNet produces more accurate shape completions than existing methods (Figure 1) and confirm the contribution of each component in our model via a thorough ablation study. In summary, our contributions are: 1) a novel robust joint symmetry-detection / shape completion method; 2) a symmetric viewpoint-based 3D shape completion method for detailed high-resolution shape reconstruction; and 3) a reinforcement learning-based optimal reflection viewpoint selection solution for progressive shape completion.", + "bbox": [ + 75, + 393, + 472, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 558, + 218, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Symmetry Detection on 3D Shapes forms a key component in many downstream applications ranging from shape matching [12], to segmentation [13], to 3D object retrieval [8], to shape completion [23, 24, 27, 28] and to deformable object learning [36]. Symmetry detection of incomplete shape observations is significantly more challenging, and recent learning-based solutions have explored the use of multitask learning [21], sampling the distribution of plausible shapes [22], and building 3D cost volumes [45] to predict 3D symmetry planes. An alternative approach to predicting arbitrarily oriented 3D symmetry planes is to estimate a camera pose that aligns the local coordinate frame such that the X-Y plane acts as a canonical symmetry plane [30, 44]. Our symmetry detection module falls in this latter category. However, unlike prior work that a priori fixes the symmetry information as a prior, we jointly perform shape completion yielding a more robust and mutually reinforcing symmetry detection and shape completion.", + "bbox": [ + 75, + 583, + 472, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning-based 3D Shape Completion of Partial Point Cloud Observations have explored coarse-to-fine methods [15, 41], transformer-networks [37, 40], probabilis-", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tic modeling [16], and topology-aware approaches [26] to complete the 3D point cloud. However, these methods are memory intensive and their scalability (e.g., beyond $\\approx 16384$ points) is limited. In contrast, our approach does not impose such limitation on the number of 3D points as our method leverages 2D depth inpainting.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-image Shape Reconstruction aims to complete approximately half of the shape either directly from a 2D image [6, 7, 29, 31] or (indirectly) from a 2.5D input [25, 35, 39]. Front2Back [39] estimates, besides depth and surface normals, also symmetry information to aid in the surface completion task. However, Front2Back is limited to a single front-back camera pair. Our method leverages an RL agent to predict multiple virtual reflection viewpoints for shape completion, and it takes better advantage of the symmetry prior by inpainting symmetric image pairs. In addition, we iteratively refine the symmetry plane to obtain more accurate estimates of both shape and symmetry.", + "bbox": [ + 496, + 186, + 892, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Iterative-Symmetry Completion Network", + "text_level": 1, + "bbox": [ + 498, + 401, + 870, + 419 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our Iterative Symmetry Completion Network (ISCNet) takes a single-view depth map of a target object as input and iteratively completes the 3D mesh by alternating between refining the symmetry plane, the virtual reflection viewpoints, and the shape completion. To bootstrap the iterative process, we first project the input single-view depth map to an incomplete point cloud that is subsequently passed to a symmetry detection module and point cloud transposition operator. The symmetry detection module (subsection 3.1) provides an estimate of the object's symmetry plane that is subsequently leveraged by the point cloud transposition operator to align the point cloud symmetrically around the estimated symmetry plane. Next, a novel 3D shape completion module (subsection 3.2) progressively completes the 3D point cloud based on a pair of reflection viewpoints selected by an RL agent. For each reflection viewpoint pair, a depth and normal map pair is extracted from the partial point cloud. Due to the partial completion of the point cloud, the extracted depth and normal map pairs exhibit holes. To address this issue, we employ a specifically crafted inpainting module, which not only repairs these holes but also assigns a confidence score to each repaired pixel. Next, the inpainted depth and normal maps are merged into the point cloud guided by their associated confidence scores. Finally, to improve robustness, we iteratively refine the symmetry plane, the reflection viewpoints, and the shape completion (subsection 3.3). We limit the number of iterations to 20 and only update the symmetry plane and reset the point cloud every 4th iteration to balance robustness versus efficiency; Figure 2 visually summarizes ISCNet.", + "bbox": [ + 496, + 431, + 893, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8897", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a275c2232977bca901e91c8827aaa14f65d2274e8c0f0153c50909992d5ca5c6.jpg", + "image_caption": [ + "Figure 2. Summary of ISCNet: a partial shape extracted from a single depth-image is aligned such that the X-Y plane corresponds to the (detected) symmetry plane. Next, a pair of depth and normal maps are rendered from reflection viewpoints selected by an RL agent, and each symmetric pair of synthetic depth and normal maps is inpainted in the image-domain (see also Figure 3). These inpainted maps are then back-projected to a point cloud and merged. The symmetry plane is updated every 4th iteration, where the merged shape is then visualized from 3 selected views, $v_{8}^{V}$ , $v_{12}^{V}$ , and $v_{13}^{V}$ in this case, passing into the symmetry detection module separately, and the best-detected symmetry plane (with the highest confidence score) is used as input for the next iteration." + ], + "image_footnote": [], + "bbox": [ + 111, + 89, + 851, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. 3D Reflection Symmetry Detection", + "text_level": 1, + "bbox": [ + 76, + 375, + 377, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We estimate the symmetry plane of an object via a proxy task, namely, camera pose estimation, such that the symmetry plane aligns in local coordinates with the X-Y plane. Our Symmetry Detection Module follows the architecture of DenseFusion [30]. We pass a depth image to obtain a point cloud, pixel-wise features, and finally a pose estimation. Different from DenseFusion, we only use depth instead of RGB channels, and we do not require an object segmentation stage. Instead of predicting the symmetry parameters directly, we found it more robust to estimate a rotation matrix $R \\in SO(3)$ and a translation vector $t \\in \\mathbb{R}^3$ such that the symmetry plane's center point $p_s$ and normal direction $n_s$ are defined by:", + "bbox": [ + 75, + 398, + 473, + 597 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left( \\begin{array}{c} p _ {s} \\\\ n _ {s} \\end{array} \\right) = \\left( \\begin{array}{c} R ^ {- 1} \\cdot ([ 0, 0, 0 ] - t) \\\\ R ^ {- 1} \\cdot ([ 0, 0, 1 ] - t) \\end{array} \\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 606, + 468, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. 3D Shape Completion", + "text_level": 1, + "bbox": [ + 76, + 648, + 282, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a predicted symmetry plane, ISCNet progressively completes the 3D shape. Our shape completion module consists of three sequential sub-modules: viewpoint selection, 2D inpainting, and point cloud updating.", + "bbox": [ + 75, + 672, + 468, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Viewpoint Selection Module: ISCNet completes 3D shapes via 2D projected depth and normal maps. The exact choice of virtual viewpoints from which these depth and normals are generated is critical for robust shape completion, and it is sensitive to the irregular structures present in the partially completed point cloud. As the shape of a partial point cloud dynamically changes, it is impractical to pre-determine viewpoints; see our supplementary material for more details.", + "bbox": [ + 75, + 734, + 468, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We model the problem of selecting a viewpoint $v_{i}$ as a Markov decision process (MDP) because, similar to [38,", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "42], we are only concerned about the current state of a partial point cloud. To this end, we leverage an RL agent to select an optimal viewpoint trained with an Actor-Critic Proximal Policy Optimisation (PPO) algorithm [20]. The RL agent predicts an optimal viewpoint (i.e., action) based on the current 3D point cloud (i.e., state) to maximize an appropriately designed reward.", + "bbox": [ + 496, + 375, + 890, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, the continuous action space of the RL agent is defined by $\\theta$ and $\\phi$ angles (Figure 2), with corresponding viewpoint:", + "bbox": [ + 498, + 507, + 892, + 553 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx = a \\sin \\theta \\sin \\phi , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 606, + 890, + 621 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny = a \\cos \\theta , \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 625, + 890, + 640 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz = a \\sin \\theta \\cos \\phi , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 643, + 890, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta \\in [0^{\\circ},180^{\\circ}]$ and $\\phi \\in [-180^{\\circ},180^{\\circ}]$ ; $a$ is set at $1.5\\mathrm{m}$ . The view direction is always aimed at the object center.", + "bbox": [ + 498, + 723, + 890, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The prediction action reward is determined by four factors: accuracy $R^a$ , shape completeness $R^s$ , number of recovered effective points $R^p$ , and viewpoint diversity $R^v$ . $R^a$ and $R^s$ encourage an agent to choose a new viewpoint that benefits the prediction of missing parts from the original point cloud. $R^p$ and $R^v$ encourage the agent to increase the viewpoint diversity. Formally, the reward function for", + "bbox": [ + 496, + 794, + 892, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8898", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ce7e296968062e1675be6315403d26c904f2c9c3594821dd76f30e2f3d4c64f2.jpg", + "image_caption": [ + "Figure 3. Detailed visual summary of the viewpoint selection and the 2D inpainting modules." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 869, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "an action is defined as:", + "bbox": [ + 76, + 378, + 230, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR = w _ {1} R ^ {a} + w _ {2} R ^ {s} + w _ {3} R ^ {p} + w _ {4} R ^ {v}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 402, + 468, + 420 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR ^ {a} = \\frac {1}{\\mathcal {N} (\\hat {P})} \\sum_ {x \\in \\hat {P}} \\min _ {y \\in P _ {G T}} \\| x - y \\| _ {2} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 422, + 468, + 460 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR ^ {s} = S - S ^ {\\prime}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 464, + 468, + 481 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR ^ {p} = \\mathcal {N} \\left(\\left\\{d (x, P ^ {\\prime}) < 0. 0 0 2 | x \\in \\hat {P} \\right\\}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 484, + 468, + 502 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR ^ {v} = \\left(\\left| \\theta - \\theta^ {\\prime} \\right| + \\left| \\phi - \\phi^ {\\prime} \\right|\\right) / 1 8 0 \\cdot \\pi , \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 503, + 468, + 521 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS = \\frac {\\mathcal {N} \\left(\\left\\{d (y , P) < 0 . 0 2 \\mid y \\in P _ {G T} \\right\\}\\right)}{\\mathcal {N} \\left(P _ {G T}\\right)}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 523, + 468, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P, P'$ , and $\\hat{P}$ are the 3D points of the current point cloud, previous point cloud, and recovered point cloud by the action, respectively. Intuitively, $P = P' + \\hat{P}$ ; $P_{GT}$ is the ground truth reference point set; $d(x, P)$ denotes the shortest distance from a point $x$ to a point set $P$ ; $S$ and $S'$ are defined based on $P$ and $P'$ , respectively; $\\mathcal{N}(\\cdot)$ counts the number of points of a point set; $w_1, w_2, w_3, w_4$ are the weights for balancing the contribution of the corresponding reward, which are set to 25, 1, 1.5, and 2 in all experiments.", + "bbox": [ + 75, + 566, + 468, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The architecture of the RL agent is simple yet effective; it contains an encoder, decoder, critic and actor heads (Figure 3). We adopt a commonly used 3D point encoder [15, 18] which extracts features from the input point cloud for downstream processing by the actor and critic. The critic head is only used during training, and the actor head predicts a viewpoint based on the input 3D point cloud.", + "bbox": [ + 75, + 704, + 468, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The 2D Inpainting Module consists of three components: mask generation, normal map inpainting, and depth-map inpainting.", + "bbox": [ + 76, + 824, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The mask generator takes as input symmetric depth map pairs $D_{i}$ and $D_{i}^{\\prime}$ corresponding to symmetric views $v_{i}$ and", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$v_{i}^{\\prime}$ , and outputs two masks $M_{i}$ and $M_{i}^{\\prime}$ that mark the areas for each input map that need to be completed. To compute $M_{i}$ , we start from an initial completion mask $\\hat{M}_i$ obtained by marking all 'black' areas of a depth map $D_{i}$ . However, this initial mask misses two important corner cases. First, not all 'black' areas should be completed. For example, the space between the legs of a chair. Second, areas with a missing front surface covered by a backside surface also need to be filled in. To cover the first case, we utilize a silhouette map generator to predict silhouette maps $K_{i}$ of the target object to exclude erroneous 'holes'. The silhouette generation is modified from a standard U-Net architecture [19] by adding an extra input head and output head to address the pairwise inputs and outputs. Furthermore, we add pairwise skip connections between the two input and output heads for exchanging symmetry information. To provide additional depth cues, we concatenate to each input $D_{i}$ (and $D_{i}^{\\prime}$ ) the initial input depth map, $D^{I}$ , along the channel dimension. To address the second corner case, we compute a front-surface mask $H_{i}$ by comparing the angle between the camera view direction and the surface normal at each point; we set the front-surface mask to 1 if the angle is less than $90^{\\circ}$ , or 0 otherwise (i.e., back facing). The final mask $M_{i}$ is then obtained by combining $\\hat{M}_{i}$ , $K_{i}$ and $H_{i}$ as:", + "bbox": [ + 496, + 378, + 890, + 741 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i} = \\left(\\hat {M} _ {i} + H _ {i}\\right) \\odot K _ {i}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 752, + 890, + 770 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {M} _ {i} = \\Gamma (D _ {i}), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 773, + 890, + 790 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH _ {i} = \\Psi \\left(N _ {i}, v _ {i}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 794, + 890, + 810 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK _ {i}, K _ {i} ^ {\\prime} = \\mathscr {K} \\left(\\left[ D _ {i}, D ^ {I} \\right], \\left[ D _ {i} ^ {\\prime}, D ^ {I} \\right]\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 813, + 890, + 832 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Gamma (X), \\Psi (Z, y) = \\left\\{ \\begin{array}{l l} 1 & X = 0, \\langle Z, y \\rangle > 0 \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 835, + 890, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ is the element-wise multiplication; $\\mathcal{H}$ denotes the", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8899", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "silhouette map generator; $[\\cdot, \\cdot]$ indicates a channel-wise concatenation operation; $\\Gamma(X)$ and $\\Psi(Z, y)$ are two indicator functions, that only differ in the conditions for 1.", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The normal inpainting submodule $\\mathcal{N}$ takes as input a pair of partial normal maps $N_{i}$ and $N_{i}^{\\prime}$ corresponding to symmetric viewpoints $v_{i}$ and $v_{i}^{\\prime}$ as well as their corresponding masks $M_{i}$ and $M_{i}^{\\prime}$ , and outputs a pair of completed normal maps $\\hat{N}_i$ and $\\hat{N}_i^\\prime$ respectively:", + "bbox": [ + 75, + 136, + 467, + 213 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {N} _ {i}, \\hat {N} _ {i} ^ {\\prime} = \\mathcal {N} ([ N _ {i}, \\mathbf {1} - M _ {i} ], [ N _ {i} ^ {\\prime}, \\mathbf {1} - M _ {i} ^ {\\prime} ]). \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 222, + 467, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The normal inpainting submodule is based on PConv [14], a partial-convolution-layer-based UNet architecture that is robust to irregularly shaped holes [9, 14]. We modify PConv to accommodate the dual in and output by duplicating the input head and last layer of the decoder and adding pairwise skip connections. Note, PConv also applies a convolution with a unit fixed kernel on the input mask to perform mask-updating, and hence the target inpainting region should be marked 0 in the mask (i.e., the inverse of mask $M_{i}$ ).", + "bbox": [ + 75, + 250, + 467, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The depth-map inpainting submodule $\\mathcal{D}$ takes as input a pair of partial depth maps $D_{i}$ and $D_{i}^{\\prime}$ corresponding to symmetric viewpoints $v_{i}$ and $v_{i}^{\\prime}$ as well as their corresponding masks ( $M_{i}$ and $M_{i}^{\\prime}$ ) and inpainted normal maps ( $\\hat{N}_{i}$ and $\\hat{N}_{i}^{\\prime}$ ), and outputs a pair of completed depth maps $\\hat{D}_i$ and $\\hat{D}_i^\\prime$ and corresponding confidence maps $C_i$ and $C_i^\\prime$ :", + "bbox": [ + 76, + 387, + 467, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} (\\hat {D} _ {i}, C _ {i}), (\\hat {D} _ {i} ^ {\\prime}, C _ {i} ^ {\\prime}) = \\mathcal {D} \\left(\\left[ D _ {i}, \\hat {N} _ {i}, \\mathbf {1} - M _ {i} \\right], \\right. \\\\ \\left[ D _ {i} ^ {\\prime}, \\hat {N} _ {i} ^ {\\prime}, \\mathbf {1} - M _ {i} ^ {\\prime} \\right]). \\tag {17} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 489, + 468, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The confidence scores indicate the confidence of the inpainted pixel (i.e., the distance of the estimated depth-map to the ground truth map on missing regions). Our implementation of the depth inpainting submodule is also based on PConv [14] with similar modifications to support the additional input and output channels.", + "bbox": [ + 75, + 551, + 467, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Point Cloud Update: We merge the inpainted depth, normal, and confidence maps into the previous point cloud by reprojecting all points in each depth map and corresponding normals to a point cloud if their confidence score exceeds a threshold (set to 0.5 in our implementation).", + "bbox": [ + 75, + 642, + 467, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Iterative Refinement", + "text_level": 1, + "bbox": [ + 76, + 726, + 271, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To improve the robustness of the symmetry plane estimation and of the shape completion, we iteratively refine the output of ISCNet. A key observation is that we can select any suitable viewpoint to generate a new depth image to serve as input to update the symmetry plane. Because it is unlikely that a single viewpoint would be ideal for all cases, we project the updated point cloud from three pre-selected trial viewpoints to generate their respective depth maps. Next, the three depth maps are passed (separately) into the symmetry detection module, and the symmetry plane with", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the highest confidence score is subsequently used for the remainder of our pipeline. The set of pre-determined viewpoints $(v_{8}^{V}, v_{12}^{V}$ and $v_{13}^{V}$ from VRCNet [16]) are empirically chosen and kept constant for all experiments. The final updated point cloud is converted to a mesh surface using Poisson Surface Reconstruction [11]; see Figure 2.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To balance computational cost versus gain in accuracy, the number of iterations is set to 20 based on an experimental analysis on ShapeNet. We update the symmetry plane every 4th iteration such that the RL agent can leverage the progressively completed point cloud for improving the reflection viewpoint estimates. In addition, to avoid accumulating incorrectly completed shape-parts due to an inaccurate symmetry plane, we reset the point cloud to its initial state when re-estimating the symmetry plane.", + "bbox": [ + 496, + 181, + 890, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Loss Functions", + "text_level": 1, + "bbox": [ + 500, + 325, + 651, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ISCNet is trained in three stages. In the first stage, the symmetry detection component and 2D inpainting module are trained separately. In the second stage, all components except the RL agent of ISCNet are then jointly fine-tuned. Finally, we train the RL agent by using a PPO algorithm with the parameters of the other components fixed using the loss function from [20].", + "bbox": [ + 496, + 348, + 890, + 453 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first-stage symmetry detection loss is defined by:", + "bbox": [ + 500, + 454, + 867, + 469 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {S} = \\mathcal {C} \\left(\\mathcal {M} \\left(P, S _ {\\mathrm {E}}\\right), \\mathcal {M} \\left(P, S _ {\\mathrm {G}}\\right)\\right) + \\mathcal {B} \\left(\\Theta_ {t _ {1}} \\left(S _ {\\mathrm {E}} ^ {N}, S _ {\\mathrm {G}} ^ {N}\\right), 1\\right), \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 474, + 890, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{C}$ measures the Chamfer Distance (CD) between the reconstructed 3D points $P$ reflected (via a reflection function $\\mathcal{M}(\\cdot, \\cdot)$ ) over the estimated symmetry plane $S_{\\mathrm{E}}$ and reference symmetry plane $S_{\\mathrm{G}}$ . We further regularize this loss via a binary cross entropy (BCE) loss $(\\mathcal{B})$ measuring the acute angle similarity between $S_{\\mathrm{E}}^{N}$ and $S_{\\mathrm{G}}^{N}$ , the corresponding normalized symmetry plane normals of the estimated and reference, via an indicator function $\\Theta_{t_1}(X,Y)$ defined as:", + "bbox": [ + 496, + 508, + 890, + 642 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta_ {t _ {1}} (X, Y) = \\left\\{ \\begin{array}{l l} 1 & \\operatorname {a r c c o s} (| \\langle X, Y \\rangle |) < = t _ {1} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 648, + 890, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $t_1$ a slack parameter set to $5^{\\circ}$ .", + "bbox": [ + 500, + 696, + 732, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first-stage inpainting loss is defined as:", + "bbox": [ + 500, + 710, + 797, + 726 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {I} = \\mathcal {L} _ {\\mathrm {K}} + \\mathcal {L} _ {\\mathrm {N}} + \\mathcal {L} _ {\\mathrm {D}}, \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 733, + 890, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {K}} = \\mathcal {B} \\left(K _ {\\mathrm {E}}, K _ {\\mathrm {G}}\\right), \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 753, + 890, + 768 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {N} = \\left\\| N _ {E} ^ {M _ {\\mathrm {G}}} - N _ {G} ^ {M _ {\\mathrm {G}}} \\right\\| _ {1}, \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 772, + 890, + 791 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D}} = \\left\\| D _ {\\mathrm {E}} ^ {M _ {\\mathrm {G}}} - D _ {\\mathrm {G}} ^ {M _ {\\mathrm {G}}} \\right\\| _ {1} +\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 792, + 785, + 811 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} \\left(C _ {\\mathrm {E}} ^ {M _ {\\mathrm {G}}}, \\Phi_ {t _ {2}} \\left(D _ {\\mathrm {E}} ^ {M _ {\\mathrm {G}}}, D _ {\\mathrm {G}} ^ {M _ {\\mathrm {G}}}\\right)\\right), \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 814, + 890, + 834 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX ^ {M _ {\\mathrm {G}}} = X \\odot M _ {\\mathrm {G}}, \\tag {24}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 835, + 890, + 853 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi_ {t _ {2}} (X, Y) = \\left\\{ \\begin{array}{l l} 1 & \\| X - Y \\| _ {1} < = t _ {2} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {25}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 857, + 890, + 897 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8900", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2564c1ecca370653770670a2b8cfac8f1409171180e634d396fe58ff913f1fb3.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Quantitative evaluation of ISCNet compared to prior work on the synthetic ShapeNet dataset ( $\\odot$ point cloud methods, $\\diamond$ implicit function methods, and $\\dagger$ image-based methods) in terms of a mesh-to-mesh symmetric ${L}_{1}$ metric (MD) and the Chamfer Distance (CD) (lowest, 2nd lowest, 3rd lowest error color coding). In addition, we show the errors on three ablation variants of ISCNet (without symmetry plane alignment, without pair-wise inpainting, and without utilizing normal information), as well as the evolution of errors per iteration." + ], + "table_body": "
CD/MD (×102)PlaneCabinetCarChairLampSofaTableVesselAverage
PCN° [41]2.87/2.443.42/3.652.12/2.014.32/4.085.49/5.192.88/3.354.49/3.722.23/2.253.48/3.34
MSN° [15]3.29/2.642.57/2.201.95/1.693.60/2.914.58/4.313.61/3.123.62/3.042.08/1.893.16/2.72
PoinTr° [40]1.10/0.982.86/3.121.72/1.653.33/2.923.45/2.922.58/2.973.44/2.821.40/1.412.48/2.35
VRCNet° [16]2.61/1.922.77/2.201.92/1.613.43/2.613.59/2.873.41/2.733.89/3.121.92/1.572.94/2.33
SnowF° [37]1.08/0.832.83/3.021.73/1.663.08/2.813.53/2.782.46/2.513.51/2.861.52/1.302.47/2.22
VE-PCN° [32]8.80/5.396.72/5.952.91/2.377.87/6.5710.94/8.538.76/7.226.33/4.993.75/2.737.01/5.47
IFNet◇ [3]1.33/1.365.17/4.773.16/2.973.13/3.241.74/1.663.61/3.872.26/2.412.26/2.522.83/2.85
Front2Back_R5† [39]1.44/1.333.14/2.591.94/1.922.49/2.312.48/2.302.53/2.213.02/3.021.88/1.492.36/2.15
Front2Back_GT† [39]1.19/1.182.78/2.211.90/1.772.00/1.861.69/1.772.13/2.002.03/2.261.53/1.511.91/1.82
Ours† (Iteration 20)0.82/0.842.04/1.631.56/1.351.54/1.391.30/1.341.84/1.531.67/1.541.04/1.001.48/1.33
Unaligned ISCNet1.43/1.432.56/2.092.15/1.882.10/1.921.76/1.872.33/1.992.33/2.251.66/1.602.04/1.88
No-Pair ISCNet0.88/1.022.28/1.991.90/1.691.75/1.472.12/1.832.07/1.762.11/1.541.41/1.541.82/1.61
No-Normal ISCNet1.08/1.072.41/2.042.18/1.861.89/1.531.91/1.782.27/1.812.14/1.811.61/1.591.94/1.69
Iteration 40.89/0.832.23/1.941.74/1.521.67/1.541.43/1.422.03/1.791.70/1.581.15/1.061.61/1.46
Iteration 80.85/0.872.19/1.731.66/1.441.60/1.451.32/1.391.92/1.631.73/1.621.10/1.041.55/1.40
Iteration 120.84/0.872.16/1.721.60/1.381.57/1.411.28/1.371.88/1.551.70/1.581.06/1.041.51/1.36
Iteration 160.83/0.842.11/1.671.57/1.361.55/1.401.26/1.361.92/1.621.67/1.541.04/1.001.49/1.35
", + "bbox": [ + 99, + 88, + 872, + 315 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7dad58b13507594ceaaea4394748a0d9416b103cc5dca2abc92b835deb5bf183.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Quantitative evaluation of ISCNet compared to prior work on real-world ScanNet dataset. The notations and colors follow Table 1." + ], + "table_body": "
CD/MD (×102)CabinetChairLampSofaTableAverage
PCN° [41]9.76/10.446.27/7.609.60/10.456.02/7.6311.83/11.948.70/9.61
MSN° [15]5.71/5.175.86/5.405.74/4.774.95/4.016.43/6.865.74/5.24
PoinTr° [40]8.33/7.686.40/6.339.66/9.196.12/5.217.84/8.477.67/7.38
VRCNet° [16]5.05/3.985.87/5.565.95/5.034.45/3.646.00/6.605.47/4.96
SnowF° [37]8.66/7.896.47/6.438.99/8.875.63/4.917.22/8.757.39/7.37
VE-PCN° [32]9.26/6.926.38/5.986.12/5.485.64/4.708.06/7.737.09/6.16
IFNet◇ [3]12.62/7.095.38/6.608.95/8.205.05/4.215.92/7.137.59/6.64
Front2Back_GT† [39]11.22/8.876.86/7.557.81/6.646.92/6.576.85/7.737.93/7.47
Ours † (Iteration 20)4.88/3.893.67/3.994.62/4.163.27/2.864.53/5.164.19/4.01
", + "bbox": [ + 192, + 393, + 777, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{K}}$ , $\\mathcal{L}_{\\mathrm{D}}$ and $\\mathcal{L}_{\\mathrm{N}}$ measure the loss over: silhouette mask $K$ , depth $D$ , and normal $N$ between the estimated $(\\cdot_E)$ and reference $(\\cdot_G)$ components. $\\Phi_{t_2}$ is an indicator function on the $L_{1}$ distance between maps with a slack parameter $t_2$ set to 0.02, and $X^{M_G}$ is the masking operator with a reference mask over a component $X \\in \\{N, D, C\\}$ . At this stage, we do not yet explicitly enforce symmetry to allow the inpainting network to converge to a good starting point for the second stage.", + "bbox": [ + 75, + 556, + 468, + 693 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The second-stage loss is the sum of both first-stage losses plus an additional loss that explicitly enforces symmetry:", + "bbox": [ + 75, + 700, + 468, + 729 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\alpha \\mathcal {L} ^ {S} + \\beta \\mathcal {L} ^ {I} + \\gamma \\mathcal {L} ^ {J}, \\tag {26}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 768, + 468, + 787 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {J} = \\Sigma_ {i = 1} ^ {4} \\left(\\| D _ {i} ^ {M _ {\\mathrm {G}}} - \\mathcal {F} \\left(D _ {i} ^ {\\prime}\\right) \\odot M _ {\\mathrm {G}} \\| _ {1} + \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 791, + 406, + 809 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| N _ {i} ^ {M _ {\\mathrm {G}}} - \\mathcal {F} \\left(N _ {i} ^ {\\prime}\\right) \\odot M _ {\\mathrm {G}} \\right\\| _ {1}), \\tag {27}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 811, + 468, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha, \\beta$ , and $\\gamma$ are balancing weights, $\\mathcal{F}$ is a 2D horizontal mirror operator, and $X$ and $X'$ the respective components from symmetric viewpoints.", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Assessment", + "text_level": 1, + "bbox": [ + 500, + 556, + 620, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement ISCNet in PyTorch [10], and train and test ISCNet for eight categories (airplane, cabinet, car, chair, lamp, sofa, table, and vessel) from the ShapeNet dataset [2]. As in prior work [15, 41], we partition the dataset in 29,774 shapes for training and 1,200 for testing. We use the 26 viewpoints provided by VRCNet [16] to generate synthetic depth images and normal maps for training and 8 randomly generated viewpoints for testing. In addition, we also test ISCNet with 456 scanned objects from ScanNet [5], excluding the categories not present in ScanNet (i.e. airplanes, cars, and vessels). The corresponding reference shapes are provided by Scan2CAD [1]; our and all competing methods are only trained with ShapeNet.", + "bbox": [ + 496, + 582, + 890, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To train ISCNet, we first pre-train the four submodules separately for 50 epochs: symmetry detection using $\\mathcal{L}^{\\mathrm{S}}$ (Equation 18), mask generation using $\\mathcal{L}_{\\mathrm{K}}$ (Equation 21), 2D normal inpainting using $\\mathcal{L}_{\\mathrm{N}}$ (Equation 22), and 2D depth inpainting using $\\mathcal{L}_{\\mathrm{D}}$ (Equation 23). Finally, we refine the complete network for an additional 10 epochs using $\\mathcal{L}$ as loss and set the weight parameters $\\alpha$ , $\\beta$ , and $\\gamma$ to 0.01, 0.05, and 1, respectively (Equation 26). While the", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8901", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/baf130b169d97b08987abdc5ad61e14ed3031125166fb83d72d56ab672e3615e.jpg", + "image_caption": [ + "Figure 4. Visual comparison of shape completion quality of ISCNet versus prior work on the synthetic ShapeNet dataset; the observed areas marked in red (the second column)." + ], + "image_footnote": [], + "bbox": [ + 122, + 90, + 844, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1eaab14e4e00d79233956c7a3cd95612a8147d1355ac0e63a60f0877285f8630.jpg", + "image_caption": [ + "Figure 5. Visual comparison of shape completion quality of ISCNet versus prior work on real-world ScanNet dataset. Note that the sofa and table are non-symmetric objects." + ], + "image_footnote": [], + "bbox": [ + 130, + 470, + 836, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "training datasets are generated from 26 pre-defined viewpoints for the first training stage, the second stage (finetuning) uses randomly selected viewpoints instead of the viewpoints chosen by the RL agent. In the latter case, we generate the corresponding ground truth depth images and normal maps on the fly during the training process. As a consequence of this phased training strategy, the distribution of (RL selected) viewpoints on the testing dataset cover", + "bbox": [ + 75, + 768, + 472, + 891 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the whole sphere, thereby avoiding overfitting to the 26 pre-selected viewpoints (see the supplemental material for more details).", + "bbox": [ + 496, + 768, + 890, + 813 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We train all networks with the Adam optimizer with a learning rate of $2e^{-4}$ for pretraining and $4e^{-5}$ for finetuning on a dual NVIDIA Gforce RTX 3090 setup. On this setup, for 20 iterations, ISCNet takes 81 seconds to complete a shape from a single depth map, where the image", + "bbox": [ + 496, + 824, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8902", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rendering and point clouds merging take around 59 seconds.", + "bbox": [ + 76, + 90, + 467, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For validation, we measure the reconstruction quality by a mesh-to-mesh symmetric $L_{1}$ distance (MD) [4] between the reference and completed meshes. Further, to quantify the reconstruction accuracy of surface details, we follow Yao et al. [39] and use the Chamfer Distance (CD) on 100K Poisson disk sampled [43] surface points from the reconstructed and reference meshes. Finally, following Zhou et al. [45] we quantify symmetry plane accuracy by comparing the cosine distance between symmetry plane normals.", + "bbox": [ + 75, + 106, + 467, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 251, + 297, + 266 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D Shape Completion. We compare the performance of ISCNet to 8 competing methods subdivided in three categories: Point-cloud completion methods (PCN [41], MSN [15], PoinTr [40], VRCNet [16], ShowFlakeNet [37], VE-PCN [32]), implicit-function-based methods (IFNet [3]), and image-based completion methods (2 variants of Front2Back $[39]^{1}$ ). All competing methods are retrained on ShapeNet and tested with identical exemplars under the same conditions. The quantitative comparisons (using the CD and MD metrics) are summarized in Table 1 and Table 2 for ShapeNet and ScanNet, respectively. From these, we can see that ISCNet offers the best overall performance for the majority of categories on both synthetic and scanned data, followed by Front2Back [39]. SnowFlakeNet [37] performs better than ISCNet on the plane shapes with a marginal accuracy gain. Compared to Front2Back, we can see that ISCNet outperforms Front2Back with the ground truth symmetry information. For completeness, Table 3 compares ISCNet with Front2Back and IFNet for the remaining five classes from ShapeNet, without retraining or finetuning. Our approach still outperforms existing work except for the rifle class compared to IFNet.", + "bbox": [ + 75, + 273, + 468, + 606 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/250ed7ef5d91aa87f8db7b7db1dabc50b190b579f20156e1aebfb4e3bc1a7619.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Quantitative evaluation of ISCNet compared to Front2Back and IFNet for the remaining 5 classes of ShapeNet, i.e., unseen-5 labels." + ], + "table_body": "
(×102)BenchDisplaySpeakerRiflePhoneAverage (unseen-5)Average (total-13)
CDF2B2.373.193.401.723.132.762.24
IFNet1.813.736.351.373.343.323.02
Ours1.442.473.091.432.102.111.72
MDF2B2.192.973.171.592.972.582.11
IFNet1.773.656.241.353.283.263.01
Ours1.412.453.011.422.072.071.61
", + "bbox": [ + 76, + 618, + 493, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 4 and 5 provide additional qualitative comparisons. Visually, ISCNet provides the 'cleanest' shape completion for all eight tasks. In particular, in the plane completion task, ISCNet is better able to retain high-frequency surface details, and in the cabinet and lamp completion tasks", + "bbox": [ + 75, + 753, + 467, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ISCNet is the only method to fully complete the surface without holes. In the case of the chair and table completion tasks, ISCNet more robustly identifies areas to be completed without erroneously merging thin features. Although our method is designed for symmetric objects, we also obtain good reconstruction quality on non-symmetric objects (sofa and table in Figure 5).", + "bbox": [ + 500, + 90, + 890, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Symmetry Detection. To assess the accuracy of the symmetry plane estimation, we compare it against two state-of-the-art single-view based pose estimation and symmetry detection methods, RotationContinuity [44] and NeRD [45]. Figure 6a shows that ISCNet's second update of the symmetry plane offers a similar performance as NeRD. Furthermore, with increasing iterations, the accuracy of the estimated symmetry plane improves (i.e., compared to the first symmetry plane estimate2, the occurrence of symmetry plane estimates with an error less than $5^{\\circ}$ , improves from $32.86\\%$ to $56.96\\%$ ), demonstrating the effectiveness of the iterative refinement for symmetry plane estimation.", + "bbox": [ + 498, + 218, + 890, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 410, + 651, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Importance of Symmetry. We validate the effectiveness of symmetry cues for shape completion in two ablation experiments. As a first experiment (Table 1, Unaligned ISCNet), we replace the estimated transformation that aligns the symmetry plane with the X-Y plane by a random rotation matrix and a perturbed center (adding normal distributed noise with $\\sigma = 0.1$ ). Consequently, ISCNet cannot leverage any symmetry cues. In a second experiment (Table 1, No-Pair ISCNet), we process each of the reflection viewpoints separately. From Table 1, we can see that the unaligned ISCNet suffers a significant performance degradation; CD increases by 0.56 on average, and MD increases 0.55 on average. By aligning the shape to the symmetry plane, no-pair ISCNet is able to leverage symmetry implicitly (by learning to copy from the negative Z coordinate). However, compared to the full ISCNet, the no-pair ISCNet still sees a performance drop (CD and MD are increased by 0.34 and 0.28, respectively). We refer to the supplementary material for a qualitative comparison of excluding explicit leveraging of the symmetry cues.", + "bbox": [ + 498, + 434, + 890, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Role of Normal Inpainting. To demonstrate the importance of performing normal inpainting, we eliminate the inpainted normal as input from the depth inpainting submodule (Table 1, No-Normal ISCNet). Without normal cues, CD increases by 0.46 and MD by 0.36. We refer to the supplementary material for a qualitative visualization of the impact of omitting the normal inpainting step.", + "bbox": [ + 498, + 757, + 890, + 864 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "2The accuracy of first symmetry estimate is similar to that of DenseFusion [30], as our symmetry estimation is based this method.", + "bbox": [ + 500, + 875, + 888, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "1The information in [39] was insufficient to replicate the symmetry plane estimation module. Therefore, we include 2 variants, trained and tested with the ground truth symmetry plane (Front2Back_GT), and with the reference symmetry plane normal perturbed within $[-5^{\\circ}, 5^{\\circ}]$ (Front2Back_R5).", + "bbox": [ + 75, + 838, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8903", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Iterative Refinement. To demonstrate that both symmetry plane estimation as well as shape completion improve with additional iterations, we compare the MD and CD errors per iteration (Table 1) as well as the symmetry plane normal error progression per iteration (Figure 6b). In both cases, we can see that the error improves per iteration and the improvement diminishes per subsequent iteration (justifying a 20 iteration stopping criterion).", + "bbox": [ + 75, + 90, + 472, + 213 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ef08d0aeb46556f227056aa264db3866194e0fee906be0935fc8ec6080050fdc.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 81, + 227, + 269, + 352 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4919d2a3ea659882d9fd192d0e8746bdae4da6fb8d1a42af67424f6da78ca44b.jpg", + "image_caption": [ + "(b)", + "Figure 6. Impact of the number of iterations on (a) the accuracy of the symmetry plane estimation, and (b) the shape reconstruction CD/MD accuracy and average and median symmetry plane estimation errors. 'Our SP $n$ ' refers to ISCNet's $n$ -th predicted symmetry plane." + ], + "image_footnote": [], + "bbox": [ + 269, + 227, + 467, + 352 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations The main bottleneck in ISCNet is the image rendering and the 3D point cloud processing step in each iteration. Furthermore, ISCNet is only able to leverage a single global symmetry plane, and thus its performance is suboptimal when multiple symmetries are present.", + "bbox": [ + 75, + 454, + 468, + 530 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 545, + 194, + 560 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We present ISCNet, a single-depth image joint symmetry plane estimation and shape completion method that explicitly leverages symmetry cues. ISCNet alternates between estimating the symmetry plane and completing the shape to enhance the accuracy of both components. Our method is informed by two key insights: (1) shape completion and symmetry plane estimation are mutually beneficial tasks, and (2) by selecting symmetric virtual viewpoints, we can explicitly inform the inpainting module on symmetry relations. Our experiments and ablation studies demonstrate that ISCNet can yield more accurate shape completions from a single-depth image than competing methods.", + "bbox": [ + 75, + 571, + 468, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 76, + 768, + 235, + 786 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank the anonymous reviewers for the insightful and constructive comments. This work was supported in part by National Key Research and Development Program of China (2022ZD0210500, 2021ZD0111902), National Natural Science Foundation of China (61972067, U19B2039, U21B2038), the Distinguished Young Scholars Funding of Dalian (No. 2022RJ01). Pieter Peers was supported in part", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "by NSF grant IIS-1909028. Felix Heide was supported by an NSF CAREER Award (2047359), a Packard Foundation Fellowship, a Sloan Research Fellowship, a Sony Young Faculty Award, a Project X Innovation Award, and an Amazon Science Research Award.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 501, + 180, + 596, + 196 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 6", + "[2] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6", + "[3] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6970-6981, 2020. 1, 6, 8", + "[4] Paolo Cignoni, Claudio Rocchini, and Roberto Scopigno. Metro: measuring error on simplified surfaces. In Computer Graphics Forum, volume 17, pages 167-174, 1998. 8", + "[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6", + "[6] Shivam Duggal and Deepak Pathak. Topologically-aware deformation fields for single-view 3d reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1536-1546, 2022. 2", + "[7] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 605-613, 2017. 2", + "[8] Boqing Gong, Jianzhuang Liu, Xiaogang Wang, and Xiaou Tang. Learning semantic signatures for 3d object retrieval. IEEE Trans. Multimedia, 15(2):369-377, 2012. 1, 2", + "[9] Xiaoguang Han, Zhaoxuan Zhang, Dong Du, Mingdai Yang, Jingming Yu, Pan Pan, Xin Yang, Ligang Liu, Zixiang Xiong, and Shuguang Cui. Deep reinforcement learning of volume-guided progressive view inpainting for 3d point scene completion from a single depth image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 234-243, 2019. 5", + "[10] Sagar Imambi, Kolla Bhanu Prakash, and GR Kanagachidambaresan. Pytorch. In Programming with TensorFlow, pages 87-104. 2021. 6", + "[11] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006. 5", + "[12] Michael Kazhdan, Thomas Funkhouser, and Szymon Rusinkiewicz. Symmetry descriptors and 3d shape matching. In Proceedings of the 2004 Eurographics/ACM SIGGRAPH" + ], + "bbox": [ + 501, + 205, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8904", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "symposium on Geometry processing, pages 115-123, 2004. 1, 2", + "[13] Wai Ho Li and Lindsay Kleeman. Segmentation and modeling of visually symmetric objects by robot actions. The International Journal of Robotics Research, 30(9):1124-1142, 2011. 1, 2", + "[14] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Eur. Conf. Comput. Vis., pages 85-100, 2018. 5", + "[15] Minghua Liu, Lu Sheng, Sheng Yang, Jing Shao, and ShiMin Hu. Morphing and sampling network for dense point cloud completion. In AAAI, volume 34, pages 11596-11603, 2020. 2, 4, 6, 8", + "[16] Liang Pan, Xinyi Chen, Zhongang Cai, Junzhe Zhang, Haiyu Zhao, Shuai Yi, and Ziwei Liu. Variational relational point completion network. In Int. Conf. Comput. Vis., pages 8524-8533, 2021. 2, 5, 6, 8", + "[17] Joshua Podolak, Philip Shilane, Aleksey Golovinskiy, Szymon Rusinkiewicz, and Thomas Funkhouser. A planar-reflective symmetry transform for 3d shapes. In ACM SIGGRAPH 2006 Papers, pages 549-559. 2006. 1", + "[18] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 652-660, 2017. 4", + "[19] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241, 2015. 4", + "[20] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3, 5", + "[21] Yifei Shi, Junwen Huang, Hongjia Zhang, Xin Xu, Szymon Rusinkiewicz, and Kai Xu. Symmetrynet: learning to predict reflectional and rotational symmetries of 3d shapes from single-view rgb-d images. ACM Trans. Graph., 39(6):1-14, 2020. 1, 2", + "[22] Yifei Shi, Xin Xu, Junhua Xi, Xiaochang Hu, Dewen Hu, and Kai Xu. Learning to detect 3d symmetry from single-view rgb-d images with weak supervision. IEEE Trans. Pattern Anal. Mach. Intell., 2022. 1, 2", + "[23] Ivan Sipiran, Robert Gregor, and Tobias Schreck. Approximate symmetry detection in partial 3d meshes. In Computer Graphics Forum, volume 33, pages 131-140. Wiley Online Library, 2014. 1, 2", + "[24] Pablo Speciale, Martin R Oswald, Andrea Cohen, and Marc Pollefeys. A symmetry prior for convex variational 3d reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 313-328. Springer, 2016. 1, 2", + "[25] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B Tenenbaum, and William T Freeman. Pix3d: Dataset and methods for single-image 3d shape modeling. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2974-2983, 2018. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Junshu Tang, Zhijun Gong, Ran Yi, Yuan Xie, and Lizhuang Ma. Lake-net: Topology-aware point cloud completion by localizing aligned keypoints. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1726-1735, 2022. 2", + "[27] Demetri Terzopoulos, Andrew Witkin, and Michael Kass. Symmetry-seeking models and 3d object reconstruction. Int. J. Comput. Vis., 1(3):211-221, 1988. 2", + "[28] Sebastian Thrun and Ben Wegbreit. Shape from symmetry. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 2, pages 1824-1831. IEEE, 2005. 1, 2", + "[29] Shubham Tulsiani, Tinghui Zhou, Alexei A Efros, and Jitendra Malik. Multi-view supervision for single-view reconstruction via differentiable ray consistency. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2626-2634, 2017. 2", + "[30] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martin-Martin, Cewu Lu, Li Fei-Fei, and Silvio Savarese. Densefusion: 6d object pose estimation by iterative dense fusion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3343–3352, 2019. 1, 2, 3, 8", + "[31] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In *Eur. Conf. Comput. Vis.*, pages 52-67, 2018. 2", + "[32] Xiaogang Wang, Marcelo H Ang, and Gim Hee Lee. Voxel-based network for shape completion by leveraging edge generation. In Int. Conf. Comput. Vis., pages 13189-13198, 2021. 6, 8", + "[33] Xiaogang Wang, Marcelo H Ang Jr, and Gim Hee Lee. Cascaded refinement network for point cloud completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 790-799, 2020. 1", + "[34] Boyan Wei, Xianfeng Ye, Chengjiang Long, Zhenjun Du, Bangyu Li, Baocai Yin, and Xin Yang. Discriminative active learning for robotic grasping in cluttered scene. IEEE Robotics and Automation Letters, 8(3):1858-1865, 2023. 1", + "[35] Jiajun Wu, Yifan Wang, Tianfan Xue, Xingyuan Sun, Bill Freeman, and Josh Tenenbaum. Marrnet: 3d shape reconstruction via 2.5 d sketches. Adv. Neural Inform. Process. Syst., 2017. 2", + "[36] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-10, 2020. 2", + "[37] Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. Snowflakenet: Point cloud completion by snowflake point deconvolution with skip-transformer. In Int. Conf. Comput. Vis., pages 5499-5509, 2021. 2, 6, 8", + "[38] Xin Yang, Yuanbo Wang, Yaru Wang, Baocai Yin, Qiang Zhang, Xiaopeng Wei, and Hongbo Fu. Active object reconstruction using a guided view planner. In Proceedings of International Joint Conference on Artificial Intelligence, pages 4965-4971, 2018. 3", + "[39] Yuan Yao, Nico Schertler, Enrique Rosales, Helge Rhodin, Leonid Sigal, and Alla Sheffer. Front2back: Single view 3d" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8905", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "shape reconstruction via front to back prediction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 531-540, 2020. 1, 2, 6, 8", + "[40] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In Int. Conf. Comput. Vis., pages 12498-12507, 2021. 1, 2, 6, 8", + "[41] Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, and Martial Hebert. Pcn: Point completion network. In Int. Conf. 3D Vis., pages 728-737, 2018. 2, 6, 8", + "[42] Zhaoxuan Zhang, Xiaoguang Han, Bo Dong, Tong Li, Baocai Yin, and Xin Yang. Point cloud scene completion with joint color and semantic estimation from single rgb-d image. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3", + "[43] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3d: A modern library for 3d data processing. arXiv preprint arXiv:1801.09847, 2018. 8", + "[44] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5745-5753, 2019. 2, 8", + "[45] Yichao Zhou, Shichen Liu, and Yi Ma. Nerd: neural 3d reflection symmetry detector. In IEEE Conf. Comput. Vis. Pattern Recog., pages 15940-15949, 2021. 1, 2, 8" + ], + "bbox": [ + 78, + 90, + 468, + 446 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8906", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Single Depth-image 3D Reflection Symmetry and Shape Prediction/fcff96a5-45a1-4f7e-ab9f-0c309af710cd_model.json b/2023/Single Depth-image 3D Reflection Symmetry and Shape Prediction/fcff96a5-45a1-4f7e-ab9f-0c309af710cd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cba695f2af8616463ff7083f63a80b5e5372ba42 --- /dev/null +++ b/2023/Single Depth-image 3D Reflection Symmetry and Shape Prediction/fcff96a5-45a1-4f7e-ab9f-0c309af710cd_model.json @@ -0,0 +1,2290 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.131, + 0.823, + 0.154 + ], + "angle": 0, + "content": "Single Depth-image 3D Reflection Symmetry and Shape Prediction" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.181, + 0.838, + 0.201 + ], + "angle": 0, + "content": "Zhaoxuan Zhang, Bo Dong*, Tong Li, Felix Heide, Pieter Peers, Baocai Yin*, Xin Yang*" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.234, + 0.314, + 0.25 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.266, + 0.474, + 0.522 + ], + "angle": 0, + "content": "In this paper, we present Iterative Symmetry Completion Network (ISCNet), a single depth-image shape completion method that exploits reflective symmetry cues to obtain more detailed shapes. The efficacy of single depth-image shape completion methods is often sensitive to the accuracy of the symmetry plane. ISCNet therefore jointly estimates the symmetry plane and shape completion iteratively; more complete shapes contribute to more robust symmetry plane estimates and vice versa. Furthermore, our shape completion method operates in the image domain, enabling more efficient high-resolution, detailed geometry reconstruction. We perform the shape completion from pairs of viewpoints, reflected across the symmetry plane, predicted by a reinforcement learning agent to improve robustness and to simultaneously explicitly leverage symmetry. We demonstrate the effectiveness of ISCNet on a variety of object categories on both synthetic and real-scanned datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.551, + 0.21, + 0.567 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.47, + 0.652 + ], + "angle": 0, + "content": "Symmetry is an intrinsic geometric property present in the vast majority of man-made and natural objects [12, 17], and which has been exploited to improve the efficacy of applications such as shape matching [12], segmentation [13], object retrieval [8], and robotic grasping [34]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.471, + 0.879 + ], + "angle": 0, + "content": "In this paper, we aim to leverage symmetry as an aid for the completion of shapes from partial observations [23, 24, 28]. Recently, Yao et al. [39] presented Front2Back, a framework for 3D shape reconstruction from a single RGB image. A key observation in Front2Back is that a (nearly) complete 3D model can be effectively described by a pair of 2.5D visible surface images taken from opposite views (i.e., front and back views). Based on this observation, Front2Back completes a shape by synthesizing a back view using a global 3D reflective symmetry plane from a given front view. However, the reconstruction quality is significantly affected by the input front view angle (Figure 1(d)). Moreover, Front2Back assumes the availability of a 'perfect' symmetry plane, which in practice is challenging to obtain from partial observations [21, 22, 45]." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.234, + 0.631, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.315, + 0.657, + 0.341 + ], + "angle": 0, + "content": "(a) Input Depth Map & Observed Surface" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.231, + 0.766, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.321, + 0.757, + 0.337 + ], + "angle": 0, + "content": "(b) PoinTr [40]" + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.233, + 0.872, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.322, + 0.862, + 0.336 + ], + "angle": 0, + "content": "(c) IF-Net [3]" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.346, + 0.641, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.428, + 0.649, + 0.442 + ], + "angle": 0, + "content": "(d) Front2Back [39]" + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.346, + 0.759, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.428, + 0.736, + 0.441 + ], + "angle": 0, + "content": "(e) Ours" + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.346, + 0.872, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.428, + 0.871, + 0.441 + ], + "angle": 0, + "content": "(f) Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.443, + 0.895, + 0.485 + ], + "angle": 0, + "content": "Figure 1. ISCNet more efficiently leverages symmetry cues for robust shape completion from a single depth image compared to prior methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.893, + 0.809 + ], + "angle": 0, + "content": "We present ISCNet (Iterative Shape Completion Network), a single depth-image shape completion method that explicitly leverages symmetry cues. Our key insight is that symmetry detection and symmetry-based shape completion are complementary tasks that can provide constructive cues to the other task. Symmetry detection can help shape completion by providing object-centric information about missing geometrical features, and shape completion can provide more robust symmetry cues by filling in missing geometry. Departing from existing work [33] that assumes an ideal prediction of the symmetry plane, ISCNet iteratively refines the symmetry plane and the reconstructed shape jointly. Furthermore, unlike Front2Back [39], ISCNet does not rely on the initial input viewpoint to a-priori fix the synthetic back view camera position but instead leverages a trained reinforcement learning (RL) policy to select optimal pairs of reflection viewpoints at each iteration. We empirically found that predetermining the viewpoints is less effective than RL-based viewpoint selection which can more effectively deal with the dynamic shape of a partial point cloud during the refinement process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Specifically, ISCNet takes as input a single depth image and reprojects it to an incomplete point cloud. Instead of explicitly detecting a symmetry plane, we estimate the pose [30] that aligns the symmetry plane to the X-Y plane. Based on the current estimate of the point cloud and symmetry aligning pose, an RL agent determines a pair of \\( re \\)-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.245, + 0.9 + ], + "angle": 0, + "content": "*Co-corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8896" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.394 + ], + "angle": 0, + "content": "flection viewpoints, reflected across the symmetry plane, for extracting a pair of partial depth and normal maps that are subsequently repaired using a novel 2D inpainting module. By directly working on depth image pairs, our method is independent of the size of the point cloud, and it can leverage geometry information from the symmetric counter view. However, because the virtual depth images are reprojections that differ from the input view, not all reprojected points are valid (e.g., the backside of points visible through a 'hole'). Therefore, we first identify reliable pixels before performing depth inpainting in the projected image domain. Finally, we reproject the inpainted maps back to 3D and update the point cloud. We progressively refine the point cloud by alternating between estimating reflection viewpoints and completing the point cloud for 20 iterations, updating the symmetry plane every 4th iteration, to balance efficiency versus efficacy. When updating the symmetry plane, we 'reset' the point cloud to its initial state to mitigate the impact of prior inaccurate symmetry plane estimations on the shape completion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.473, + 0.545 + ], + "angle": 0, + "content": "We validate that ISCNet produces more accurate shape completions than existing methods (Figure 1) and confirm the contribution of each component in our model via a thorough ablation study. In summary, our contributions are: 1) a novel robust joint symmetry-detection / shape completion method; 2) a symmetric viewpoint-based 3D shape completion method for detailed high-resolution shape reconstruction; and 3) a reinforcement learning-based optimal reflection viewpoint selection solution for progressive shape completion." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.559, + 0.22, + 0.574 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.473, + 0.856 + ], + "angle": 0, + "content": "Symmetry Detection on 3D Shapes forms a key component in many downstream applications ranging from shape matching [12], to segmentation [13], to 3D object retrieval [8], to shape completion [23, 24, 27, 28] and to deformable object learning [36]. Symmetry detection of incomplete shape observations is significantly more challenging, and recent learning-based solutions have explored the use of multitask learning [21], sampling the distribution of plausible shapes [22], and building 3D cost volumes [45] to predict 3D symmetry planes. An alternative approach to predicting arbitrarily oriented 3D symmetry planes is to estimate a camera pose that aligns the local coordinate frame such that the X-Y plane acts as a canonical symmetry plane [30, 44]. Our symmetry detection module falls in this latter category. However, unlike prior work that a priori fixes the symmetry information as a prior, we jointly perform shape completion yielding a more robust and mutually reinforcing symmetry detection and shape completion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Learning-based 3D Shape Completion of Partial Point Cloud Observations have explored coarse-to-fine methods [15, 41], transformer-networks [37, 40], probabilis-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "tic modeling [16], and topology-aware approaches [26] to complete the 3D point cloud. However, these methods are memory intensive and their scalability (e.g., beyond \\(\\approx 16384\\) points) is limited. In contrast, our approach does not impose such limitation on the number of 3D points as our method leverages 2D depth inpainting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.188, + 0.893, + 0.37 + ], + "angle": 0, + "content": "Single-image Shape Reconstruction aims to complete approximately half of the shape either directly from a 2D image [6, 7, 29, 31] or (indirectly) from a 2.5D input [25, 35, 39]. Front2Back [39] estimates, besides depth and surface normals, also symmetry information to aid in the surface completion task. However, Front2Back is limited to a single front-back camera pair. Our method leverages an RL agent to predict multiple virtual reflection viewpoints for shape completion, and it takes better advantage of the symmetry prior by inpainting symmetric image pairs. In addition, we iteratively refine the symmetry plane to obtain more accurate estimates of both shape and symmetry." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.402, + 0.871, + 0.42 + ], + "angle": 0, + "content": "3. Iterative-Symmetry Completion Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.895, + 0.9 + ], + "angle": 0, + "content": "Our Iterative Symmetry Completion Network (ISCNet) takes a single-view depth map of a target object as input and iteratively completes the 3D mesh by alternating between refining the symmetry plane, the virtual reflection viewpoints, and the shape completion. To bootstrap the iterative process, we first project the input single-view depth map to an incomplete point cloud that is subsequently passed to a symmetry detection module and point cloud transposition operator. The symmetry detection module (subsection 3.1) provides an estimate of the object's symmetry plane that is subsequently leveraged by the point cloud transposition operator to align the point cloud symmetrically around the estimated symmetry plane. Next, a novel 3D shape completion module (subsection 3.2) progressively completes the 3D point cloud based on a pair of reflection viewpoints selected by an RL agent. For each reflection viewpoint pair, a depth and normal map pair is extracted from the partial point cloud. Due to the partial completion of the point cloud, the extracted depth and normal map pairs exhibit holes. To address this issue, we employ a specifically crafted inpainting module, which not only repairs these holes but also assigns a confidence score to each repaired pixel. Next, the inpainted depth and normal maps are merged into the point cloud guided by their associated confidence scores. Finally, to improve robustness, we iteratively refine the symmetry plane, the reflection viewpoints, and the shape completion (subsection 3.3). We limit the number of iterations to 20 and only update the symmetry plane and reset the point cloud every 4th iteration to balance robustness versus efficiency; Figure 2 visually summarizes ISCNet." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8897" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.112, + 0.09, + 0.852, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.268, + 0.895, + 0.353 + ], + "angle": 0, + "content": "Figure 2. Summary of ISCNet: a partial shape extracted from a single depth-image is aligned such that the X-Y plane corresponds to the (detected) symmetry plane. Next, a pair of depth and normal maps are rendered from reflection viewpoints selected by an RL agent, and each symmetric pair of synthetic depth and normal maps is inpainted in the image-domain (see also Figure 3). These inpainted maps are then back-projected to a point cloud and merged. The symmetry plane is updated every 4th iteration, where the merged shape is then visualized from 3 selected views, \\( v_{8}^{V} \\), \\( v_{12}^{V} \\), and \\( v_{13}^{V} \\) in this case, passing into the symmetry detection module separately, and the best-detected symmetry plane (with the highest confidence score) is used as input for the next iteration." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.375, + 0.379, + 0.391 + ], + "angle": 0, + "content": "3.1. 3D Reflection Symmetry Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.474, + 0.598 + ], + "angle": 0, + "content": "We estimate the symmetry plane of an object via a proxy task, namely, camera pose estimation, such that the symmetry plane aligns in local coordinates with the X-Y plane. Our Symmetry Detection Module follows the architecture of DenseFusion [30]. We pass a depth image to obtain a point cloud, pixel-wise features, and finally a pose estimation. Different from DenseFusion, we only use depth instead of RGB channels, and we do not require an object segmentation stage. Instead of predicting the symmetry parameters directly, we found it more robust to estimate a rotation matrix \\( R \\in SO(3) \\) and a translation vector \\( t \\in \\mathbb{R}^3 \\) such that the symmetry plane's center point \\( p_s \\) and normal direction \\( n_s \\) are defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.607, + 0.47, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\left( \\begin{array}{c} p _ {s} \\\\ n _ {s} \\end{array} \\right) = \\left( \\begin{array}{c} R ^ {- 1} \\cdot ([ 0, 0, 0 ] - t) \\\\ R ^ {- 1} \\cdot ([ 0, 0, 1 ] - t) \\end{array} \\right). \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.65, + 0.283, + 0.667 + ], + "angle": 0, + "content": "3.2. 3D Shape Completion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.469, + 0.735 + ], + "angle": 0, + "content": "Given a predicted symmetry plane, ISCNet progressively completes the 3D shape. Our shape completion module consists of three sequential sub-modules: viewpoint selection, 2D inpainting, and point cloud updating." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.869 + ], + "angle": 0, + "content": "Viewpoint Selection Module: ISCNet completes 3D shapes via 2D projected depth and normal maps. The exact choice of virtual viewpoints from which these depth and normals are generated is critical for robust shape completion, and it is sensitive to the irregular structures present in the partially completed point cloud. As the shape of a partial point cloud dynamically changes, it is impractical to pre-determine viewpoints; see our supplementary material for more details." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We model the problem of selecting a viewpoint \\( v_{i} \\) as a Markov decision process (MDP) because, similar to [38," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.376, + 0.892, + 0.483 + ], + "angle": 0, + "content": "42], we are only concerned about the current state of a partial point cloud. To this end, we leverage an RL agent to select an optimal viewpoint trained with an Actor-Critic Proximal Policy Optimisation (PPO) algorithm [20]. The RL agent predicts an optimal viewpoint (i.e., action) based on the current 3D point cloud (i.e., state) to maximize an appropriately designed reward." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.508, + 0.893, + 0.554 + ], + "angle": 0, + "content": "Specifically, the continuous action space of the RL agent is defined by \\(\\theta\\) and \\(\\phi\\) angles (Figure 2), with corresponding viewpoint:" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.607, + 0.891, + 0.622 + ], + "angle": 0, + "content": "\\[\nx = a \\sin \\theta \\sin \\phi , \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.626, + 0.891, + 0.641 + ], + "angle": 0, + "content": "\\[\ny = a \\cos \\theta , \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.645, + 0.891, + 0.66 + ], + "angle": 0, + "content": "\\[\nz = a \\sin \\theta \\cos \\phi , \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.724, + 0.892, + 0.768 + ], + "angle": 0, + "content": "where \\(\\theta \\in [0^{\\circ},180^{\\circ}]\\) and \\(\\phi \\in [-180^{\\circ},180^{\\circ}]\\); \\(a\\) is set at \\(1.5\\mathrm{m}\\). The view direction is always aimed at the object center." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.893, + 0.903 + ], + "angle": 0, + "content": "The prediction action reward is determined by four factors: accuracy \\( R^a \\), shape completeness \\( R^s \\), number of recovered effective points \\( R^p \\), and viewpoint diversity \\( R^v \\). \\( R^a \\) and \\( R^s \\) encourage an agent to choose a new viewpoint that benefits the prediction of missing parts from the original point cloud. \\( R^p \\) and \\( R^v \\) encourage the agent to increase the viewpoint diversity. Formally, the reward function for" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8898" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.09, + 0.87, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.358, + 0.761, + 0.373 + ], + "angle": 0, + "content": "Figure 3. Detailed visual summary of the viewpoint selection and the 2D inpainting modules." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.38, + 0.232, + 0.394 + ], + "angle": 0, + "content": "an action is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.404, + 0.469, + 0.421 + ], + "angle": 0, + "content": "\\[\nR = w _ {1} R ^ {a} + w _ {2} R ^ {s} + w _ {3} R ^ {p} + w _ {4} R ^ {v}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.424, + 0.469, + 0.462 + ], + "angle": 0, + "content": "\\[\nR ^ {a} = \\frac {1}{\\mathcal {N} (\\hat {P})} \\sum_ {x \\in \\hat {P}} \\min _ {y \\in P _ {G T}} \\| x - y \\| _ {2} ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.465, + 0.469, + 0.482 + ], + "angle": 0, + "content": "\\[\nR ^ {s} = S - S ^ {\\prime}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.485, + 0.469, + 0.503 + ], + "angle": 0, + "content": "\\[\nR ^ {p} = \\mathcal {N} \\left(\\left\\{d (x, P ^ {\\prime}) < 0. 0 0 2 | x \\in \\hat {P} \\right\\}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.505, + 0.469, + 0.522 + ], + "angle": 0, + "content": "\\[\nR ^ {v} = \\left(\\left| \\theta - \\theta^ {\\prime} \\right| + \\left| \\phi - \\phi^ {\\prime} \\right|\\right) / 1 8 0 \\cdot \\pi , \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.524, + 0.469, + 0.558 + ], + "angle": 0, + "content": "\\[\nS = \\frac {\\mathcal {N} \\left(\\left\\{d (y , P) < 0 . 0 2 \\mid y \\in P _ {G T} \\right\\}\\right)}{\\mathcal {N} \\left(P _ {G T}\\right)}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.469, + 0.704 + ], + "angle": 0, + "content": "where \\(P, P'\\), and \\(\\hat{P}\\) are the 3D points of the current point cloud, previous point cloud, and recovered point cloud by the action, respectively. Intuitively, \\(P = P' + \\hat{P}\\); \\(P_{GT}\\) is the ground truth reference point set; \\(d(x, P)\\) denotes the shortest distance from a point \\(x\\) to a point set \\(P\\); \\(S\\) and \\(S'\\) are defined based on \\(P\\) and \\(P'\\), respectively; \\(\\mathcal{N}(\\cdot)\\) counts the number of points of a point set; \\(w_1, w_2, w_3, w_4\\) are the weights for balancing the contribution of the corresponding reward, which are set to 25, 1, 1.5, and 2 in all experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.469, + 0.822 + ], + "angle": 0, + "content": "The architecture of the RL agent is simple yet effective; it contains an encoder, decoder, critic and actor heads (Figure 3). We adopt a commonly used 3D point encoder [15, 18] which extracts features from the input point cloud for downstream processing by the actor and critic. The critic head is only used during training, and the actor head predicts a viewpoint based on the input 3D point cloud." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.469, + 0.87 + ], + "angle": 0, + "content": "The 2D Inpainting Module consists of three components: mask generation, normal map inpainting, and depth-map inpainting." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The mask generator takes as input symmetric depth map pairs \\(D_{i}\\) and \\(D_{i}^{\\prime}\\) corresponding to symmetric views \\(v_{i}\\) and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.892, + 0.742 + ], + "angle": 0, + "content": "\\(v_{i}^{\\prime}\\), and outputs two masks \\(M_{i}\\) and \\(M_{i}^{\\prime}\\) that mark the areas for each input map that need to be completed. To compute \\(M_{i}\\), we start from an initial completion mask \\(\\hat{M}_i\\) obtained by marking all 'black' areas of a depth map \\(D_{i}\\). However, this initial mask misses two important corner cases. First, not all 'black' areas should be completed. For example, the space between the legs of a chair. Second, areas with a missing front surface covered by a backside surface also need to be filled in. To cover the first case, we utilize a silhouette map generator to predict silhouette maps \\(K_{i}\\) of the target object to exclude erroneous 'holes'. The silhouette generation is modified from a standard U-Net architecture [19] by adding an extra input head and output head to address the pairwise inputs and outputs. Furthermore, we add pairwise skip connections between the two input and output heads for exchanging symmetry information. To provide additional depth cues, we concatenate to each input \\(D_{i}\\) (and \\(D_{i}^{\\prime}\\)) the initial input depth map, \\(D^{I}\\), along the channel dimension. To address the second corner case, we compute a front-surface mask \\(H_{i}\\) by comparing the angle between the camera view direction and the surface normal at each point; we set the front-surface mask to 1 if the angle is less than \\(90^{\\circ}\\), or 0 otherwise (i.e., back facing). The final mask \\(M_{i}\\) is then obtained by combining \\(\\hat{M}_{i}\\), \\(K_{i}\\) and \\(H_{i}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.753, + 0.891, + 0.771 + ], + "angle": 0, + "content": "\\[\nM _ {i} = \\left(\\hat {M} _ {i} + H _ {i}\\right) \\odot K _ {i}, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.774, + 0.891, + 0.791 + ], + "angle": 0, + "content": "\\[\n\\hat {M} _ {i} = \\Gamma (D _ {i}), \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.795, + 0.891, + 0.811 + ], + "angle": 0, + "content": "\\[\nH _ {i} = \\Psi \\left(N _ {i}, v _ {i}\\right), \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.814, + 0.891, + 0.833 + ], + "angle": 0, + "content": "\\[\nK _ {i}, K _ {i} ^ {\\prime} = \\mathscr {K} \\left(\\left[ D _ {i}, D ^ {I} \\right], \\left[ D _ {i} ^ {\\prime}, D ^ {I} \\right]\\right), \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.836, + 0.891, + 0.875 + ], + "angle": 0, + "content": "\\[\n\\Gamma (X), \\Psi (Z, y) = \\left\\{ \\begin{array}{l l} 1 & X = 0, \\langle Z, y \\rangle > 0 \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.891, + 0.901 + ], + "angle": 0, + "content": "where \\(\\odot\\) is the element-wise multiplication; \\(\\mathcal{H}\\) denotes the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8899" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "silhouette map generator; \\([\\cdot, \\cdot]\\) indicates a channel-wise concatenation operation; \\(\\Gamma(X)\\) and \\(\\Psi(Z, y)\\) are two indicator functions, that only differ in the conditions for 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.468, + 0.214 + ], + "angle": 0, + "content": "The normal inpainting submodule \\(\\mathcal{N}\\) takes as input a pair of partial normal maps \\(N_{i}\\) and \\(N_{i}^{\\prime}\\) corresponding to symmetric viewpoints \\(v_{i}\\) and \\(v_{i}^{\\prime}\\) as well as their corresponding masks \\(M_{i}\\) and \\(M_{i}^{\\prime}\\), and outputs a pair of completed normal maps \\(\\hat{N}_i\\) and \\(\\hat{N}_i^\\prime\\) respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.223, + 0.468, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\hat {N} _ {i}, \\hat {N} _ {i} ^ {\\prime} = \\mathcal {N} ([ N _ {i}, \\mathbf {1} - M _ {i} ], [ N _ {i} ^ {\\prime}, \\mathbf {1} - M _ {i} ^ {\\prime} ]). \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.251, + 0.468, + 0.386 + ], + "angle": 0, + "content": "The normal inpainting submodule is based on PConv [14], a partial-convolution-layer-based UNet architecture that is robust to irregularly shaped holes [9, 14]. We modify PConv to accommodate the dual in and output by duplicating the input head and last layer of the decoder and adding pairwise skip connections. Note, PConv also applies a convolution with a unit fixed kernel on the input mask to perform mask-updating, and hence the target inpainting region should be marked 0 in the mask (i.e., the inverse of mask \\( M_{i} \\))." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.388, + 0.468, + 0.481 + ], + "angle": 0, + "content": "The depth-map inpainting submodule \\(\\mathcal{D}\\) takes as input a pair of partial depth maps \\(D_{i}\\) and \\(D_{i}^{\\prime}\\) corresponding to symmetric viewpoints \\(v_{i}\\) and \\(v_{i}^{\\prime}\\) as well as their corresponding masks (\\(M_{i}\\) and \\(M_{i}^{\\prime}\\)) and inpainted normal maps (\\(\\hat{N}_{i}\\) and \\(\\hat{N}_{i}^{\\prime}\\)), and outputs a pair of completed depth maps \\(\\hat{D}_i\\) and \\(\\hat{D}_i^\\prime\\) and corresponding confidence maps \\(C_i\\) and \\(C_i^\\prime\\):" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.49, + 0.469, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} (\\hat {D} _ {i}, C _ {i}), (\\hat {D} _ {i} ^ {\\prime}, C _ {i} ^ {\\prime}) = \\mathcal {D} \\left(\\left[ D _ {i}, \\hat {N} _ {i}, \\mathbf {1} - M _ {i} \\right], \\right. \\\\ \\left[ D _ {i} ^ {\\prime}, \\hat {N} _ {i} ^ {\\prime}, \\mathbf {1} - M _ {i} ^ {\\prime} \\right]). \\tag {17} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.468, + 0.642 + ], + "angle": 0, + "content": "The confidence scores indicate the confidence of the inpainted pixel (i.e., the distance of the estimated depth-map to the ground truth map on missing regions). Our implementation of the depth inpainting submodule is also based on PConv [14] with similar modifications to support the additional input and output channels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.468, + 0.718 + ], + "angle": 0, + "content": "Point Cloud Update: We merge the inpainted depth, normal, and confidence maps into the previous point cloud by reprojecting all points in each depth map and corresponding normals to a point cloud if their confidence score exceeds a threshold (set to 0.5 in our implementation)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.272, + 0.74 + ], + "angle": 0, + "content": "3.3. Iterative Refinement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To improve the robustness of the symmetry plane estimation and of the shape completion, we iteratively refine the output of ISCNet. A key observation is that we can select any suitable viewpoint to generate a new depth image to serve as input to update the symmetry plane. Because it is unlikely that a single viewpoint would be ideal for all cases, we project the updated point cloud from three pre-selected trial viewpoints to generate their respective depth maps. Next, the three depth maps are passed (separately) into the symmetry detection module, and the symmetry plane with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.891, + 0.182 + ], + "angle": 0, + "content": "the highest confidence score is subsequently used for the remainder of our pipeline. The set of pre-determined viewpoints \\((v_{8}^{V}, v_{12}^{V}\\) and \\(v_{13}^{V}\\) from VRCNet [16]) are empirically chosen and kept constant for all experiments. The final updated point cloud is converted to a mesh surface using Poisson Surface Reconstruction [11]; see Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.892, + 0.319 + ], + "angle": 0, + "content": "To balance computational cost versus gain in accuracy, the number of iterations is set to 20 based on an experimental analysis on ShapeNet. We update the symmetry plane every 4th iteration such that the RL agent can leverage the progressively completed point cloud for improving the reflection viewpoint estimates. In addition, to avoid accumulating incorrectly completed shape-parts due to an inaccurate symmetry plane, we reset the point cloud to its initial state when re-estimating the symmetry plane." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.327, + 0.652, + 0.34 + ], + "angle": 0, + "content": "3.4. Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.892, + 0.454 + ], + "angle": 0, + "content": "ISCNet is trained in three stages. In the first stage, the symmetry detection component and 2D inpainting module are trained separately. In the second stage, all components except the RL agent of ISCNet are then jointly fine-tuned. Finally, we train the RL agent by using a PPO algorithm with the parameters of the other components fixed using the loss function from [20]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.455, + 0.868, + 0.47 + ], + "angle": 0, + "content": "The first-stage symmetry detection loss is defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.476, + 0.891, + 0.509 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {S} = \\mathcal {C} \\left(\\mathcal {M} \\left(P, S _ {\\mathrm {E}}\\right), \\mathcal {M} \\left(P, S _ {\\mathrm {G}}\\right)\\right) + \\mathcal {B} \\left(\\Theta_ {t _ {1}} \\left(S _ {\\mathrm {E}} ^ {N}, S _ {\\mathrm {G}} ^ {N}\\right), 1\\right), \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.892, + 0.643 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}\\) measures the Chamfer Distance (CD) between the reconstructed 3D points \\(P\\) reflected (via a reflection function \\(\\mathcal{M}(\\cdot, \\cdot)\\)) over the estimated symmetry plane \\(S_{\\mathrm{E}}\\) and reference symmetry plane \\(S_{\\mathrm{G}}\\). We further regularize this loss via a binary cross entropy (BCE) loss \\((\\mathcal{B})\\) measuring the acute angle similarity between \\(S_{\\mathrm{E}}^{N}\\) and \\(S_{\\mathrm{G}}^{N}\\), the corresponding normalized symmetry plane normals of the estimated and reference, via an indicator function \\(\\Theta_{t_1}(X,Y)\\) defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.65, + 0.891, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\Theta_ {t _ {1}} (X, Y) = \\left\\{ \\begin{array}{l l} 1 & \\operatorname {a r c c o s} (| \\langle X, Y \\rangle |) < = t _ {1} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.697, + 0.733, + 0.711 + ], + "angle": 0, + "content": "with \\(t_1\\) a slack parameter set to \\(5^{\\circ}\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.712, + 0.799, + 0.727 + ], + "angle": 0, + "content": "The first-stage inpainting loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.734, + 0.891, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {I} = \\mathcal {L} _ {\\mathrm {K}} + \\mathcal {L} _ {\\mathrm {N}} + \\mathcal {L} _ {\\mathrm {D}}, \\tag {20}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.754, + 0.891, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {K}} = \\mathcal {B} \\left(K _ {\\mathrm {E}}, K _ {\\mathrm {G}}\\right), \\tag {21}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.773, + 0.891, + 0.792 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {N} = \\left\\| N _ {E} ^ {M _ {\\mathrm {G}}} - N _ {G} ^ {M _ {\\mathrm {G}}} \\right\\| _ {1}, \\tag {22}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.794, + 0.786, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D}} = \\left\\| D _ {\\mathrm {E}} ^ {M _ {\\mathrm {G}}} - D _ {\\mathrm {G}} ^ {M _ {\\mathrm {G}}} \\right\\| _ {1} +\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.815, + 0.891, + 0.835 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} \\left(C _ {\\mathrm {E}} ^ {M _ {\\mathrm {G}}}, \\Phi_ {t _ {2}} \\left(D _ {\\mathrm {E}} ^ {M _ {\\mathrm {G}}}, D _ {\\mathrm {G}} ^ {M _ {\\mathrm {G}}}\\right)\\right), \\tag {23}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.837, + 0.891, + 0.854 + ], + "angle": 0, + "content": "\\[\nX ^ {M _ {\\mathrm {G}}} = X \\odot M _ {\\mathrm {G}}, \\tag {24}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.858, + 0.891, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\Phi_ {t _ {2}} (X, Y) = \\left\\{ \\begin{array}{l l} 1 & \\| X - Y \\| _ {1} < = t _ {2} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {25}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8900" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.874, + 0.316 + ], + "angle": 0, + "content": "
CD/MD (×102)PlaneCabinetCarChairLampSofaTableVesselAverage
PCN° [41]2.87/2.443.42/3.652.12/2.014.32/4.085.49/5.192.88/3.354.49/3.722.23/2.253.48/3.34
MSN° [15]3.29/2.642.57/2.201.95/1.693.60/2.914.58/4.313.61/3.123.62/3.042.08/1.893.16/2.72
PoinTr° [40]1.10/0.982.86/3.121.72/1.653.33/2.923.45/2.922.58/2.973.44/2.821.40/1.412.48/2.35
VRCNet° [16]2.61/1.922.77/2.201.92/1.613.43/2.613.59/2.873.41/2.733.89/3.121.92/1.572.94/2.33
SnowF° [37]1.08/0.832.83/3.021.73/1.663.08/2.813.53/2.782.46/2.513.51/2.861.52/1.302.47/2.22
VE-PCN° [32]8.80/5.396.72/5.952.91/2.377.87/6.5710.94/8.538.76/7.226.33/4.993.75/2.737.01/5.47
IFNet◇ [3]1.33/1.365.17/4.773.16/2.973.13/3.241.74/1.663.61/3.872.26/2.412.26/2.522.83/2.85
Front2Back_R5† [39]1.44/1.333.14/2.591.94/1.922.49/2.312.48/2.302.53/2.213.02/3.021.88/1.492.36/2.15
Front2Back_GT† [39]1.19/1.182.78/2.211.90/1.772.00/1.861.69/1.772.13/2.002.03/2.261.53/1.511.91/1.82
Ours† (Iteration 20)0.82/0.842.04/1.631.56/1.351.54/1.391.30/1.341.84/1.531.67/1.541.04/1.001.48/1.33
Unaligned ISCNet1.43/1.432.56/2.092.15/1.882.10/1.921.76/1.872.33/1.992.33/2.251.66/1.602.04/1.88
No-Pair ISCNet0.88/1.022.28/1.991.90/1.691.75/1.472.12/1.832.07/1.762.11/1.541.41/1.541.82/1.61
No-Normal ISCNet1.08/1.072.41/2.042.18/1.861.89/1.531.91/1.782.27/1.812.14/1.811.61/1.591.94/1.69
Iteration 40.89/0.832.23/1.941.74/1.521.67/1.541.43/1.422.03/1.791.70/1.581.15/1.061.61/1.46
Iteration 80.85/0.872.19/1.731.66/1.441.60/1.451.32/1.391.92/1.631.73/1.621.10/1.041.55/1.40
Iteration 120.84/0.872.16/1.721.60/1.381.57/1.411.28/1.371.88/1.551.70/1.581.06/1.041.51/1.36
Iteration 160.83/0.842.11/1.671.57/1.361.55/1.401.26/1.361.92/1.621.67/1.541.04/1.001.49/1.35
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.316, + 0.894, + 0.384 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation of ISCNet compared to prior work on the synthetic ShapeNet dataset ( \\( \\odot \\) point cloud methods, \\( \\diamond \\) implicit function methods, and \\( \\dagger \\) image-based methods) in terms of a mesh-to-mesh symmetric \\( {L}_{1} \\) metric (MD) and the Chamfer Distance (CD) (lowest, 2nd lowest, 3rd lowest error color coding). In addition, we show the errors on three ablation variants of ISCNet (without symmetry plane alignment, without pair-wise inpainting, and without utilizing normal information), as well as the evolution of errors per iteration." + }, + { + "type": "table", + "bbox": [ + 0.194, + 0.394, + 0.778, + 0.52 + ], + "angle": 0, + "content": "
CD/MD (×102)CabinetChairLampSofaTableAverage
PCN° [41]9.76/10.446.27/7.609.60/10.456.02/7.6311.83/11.948.70/9.61
MSN° [15]5.71/5.175.86/5.405.74/4.774.95/4.016.43/6.865.74/5.24
PoinTr° [40]8.33/7.686.40/6.339.66/9.196.12/5.217.84/8.477.67/7.38
VRCNet° [16]5.05/3.985.87/5.565.95/5.034.45/3.646.00/6.605.47/4.96
SnowF° [37]8.66/7.896.47/6.438.99/8.875.63/4.917.22/8.757.39/7.37
VE-PCN° [32]9.26/6.926.38/5.986.12/5.485.64/4.708.06/7.737.09/6.16
IFNet◇ [3]12.62/7.095.38/6.608.95/8.205.05/4.215.92/7.137.59/6.64
Front2Back_GT† [39]11.22/8.876.86/7.557.81/6.646.92/6.576.85/7.737.93/7.47
Ours † (Iteration 20)4.88/3.893.67/3.994.62/4.163.27/2.864.53/5.164.19/4.01
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.52, + 0.892, + 0.532 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation of ISCNet compared to prior work on real-world ScanNet dataset. The notations and colors follow Table 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.47, + 0.694 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{K}}\\), \\(\\mathcal{L}_{\\mathrm{D}}\\) and \\(\\mathcal{L}_{\\mathrm{N}}\\) measure the loss over: silhouette mask \\(K\\), depth \\(D\\), and normal \\(N\\) between the estimated \\((\\cdot_E)\\) and reference \\((\\cdot_G)\\) components. \\(\\Phi_{t_2}\\) is an indicator function on the \\(L_{1}\\) distance between maps with a slack parameter \\(t_2\\) set to 0.02, and \\(X^{M_G}\\) is the masking operator with a reference mask over a component \\(X \\in \\{N, D, C\\}\\). At this stage, we do not yet explicitly enforce symmetry to allow the inpainting network to converge to a good starting point for the second stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.47, + 0.731 + ], + "angle": 0, + "content": "The second-stage loss is the sum of both first-stage losses plus an additional loss that explicitly enforces symmetry:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.77, + 0.469, + 0.788 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\alpha \\mathcal {L} ^ {S} + \\beta \\mathcal {L} ^ {I} + \\gamma \\mathcal {L} ^ {J}, \\tag {26}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.792, + 0.408, + 0.81 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {J} = \\Sigma_ {i = 1} ^ {4} \\left(\\| D _ {i} ^ {M _ {\\mathrm {G}}} - \\mathcal {F} \\left(D _ {i} ^ {\\prime}\\right) \\odot M _ {\\mathrm {G}} \\| _ {1} + \\right.\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.812, + 0.469, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\left\\| N _ {i} ^ {M _ {\\mathrm {G}}} - \\mathcal {F} \\left(N _ {i} ^ {\\prime}\\right) \\odot M _ {\\mathrm {G}} \\right\\| _ {1}), \\tag {27}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(\\alpha, \\beta\\), and \\(\\gamma\\) are balancing weights, \\(\\mathcal{F}\\) is a 2D horizontal mirror operator, and \\(X\\) and \\(X'\\) the respective components from symmetric viewpoints." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.557, + 0.621, + 0.572 + ], + "angle": 0, + "content": "4. Assessment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.779 + ], + "angle": 0, + "content": "We implement ISCNet in PyTorch [10], and train and test ISCNet for eight categories (airplane, cabinet, car, chair, lamp, sofa, table, and vessel) from the ShapeNet dataset [2]. As in prior work [15, 41], we partition the dataset in 29,774 shapes for training and 1,200 for testing. We use the 26 viewpoints provided by VRCNet [16] to generate synthetic depth images and normal maps for training and 8 randomly generated viewpoints for testing. In addition, we also test ISCNet with 456 scanned objects from ScanNet [5], excluding the categories not present in ScanNet (i.e. airplanes, cars, and vessels). The corresponding reference shapes are provided by Scan2CAD [1]; our and all competing methods are only trained with ShapeNet." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To train ISCNet, we first pre-train the four submodules separately for 50 epochs: symmetry detection using \\(\\mathcal{L}^{\\mathrm{S}}\\) (Equation 18), mask generation using \\(\\mathcal{L}_{\\mathrm{K}}\\) (Equation 21), 2D normal inpainting using \\(\\mathcal{L}_{\\mathrm{N}}\\) (Equation 22), and 2D depth inpainting using \\(\\mathcal{L}_{\\mathrm{D}}\\) (Equation 23). Finally, we refine the complete network for an additional 10 epochs using \\(\\mathcal{L}\\) as loss and set the weight parameters \\(\\alpha\\), \\(\\beta\\), and \\(\\gamma\\) to 0.01, 0.05, and 1, respectively (Equation 26). While the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8901" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.092, + 0.845, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.428, + 0.893, + 0.456 + ], + "angle": 0, + "content": "Figure 4. Visual comparison of shape completion quality of ISCNet versus prior work on the synthetic ShapeNet dataset; the observed areas marked in red (the second column)." + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.471, + 0.837, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.716, + 0.893, + 0.744 + ], + "angle": 0, + "content": "Figure 5. Visual comparison of shape completion quality of ISCNet versus prior work on real-world ScanNet dataset. Note that the sofa and table are non-symmetric objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.769, + 0.473, + 0.892 + ], + "angle": 0, + "content": "training datasets are generated from 26 pre-defined viewpoints for the first training stage, the second stage (finetuning) uses randomly selected viewpoints instead of the viewpoints chosen by the RL agent. In the latter case, we generate the corresponding ground truth depth images and normal maps on the fly during the training process. As a consequence of this phased training strategy, the distribution of (RL selected) viewpoints on the testing dataset cover" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.769, + 0.892, + 0.814 + ], + "angle": 0, + "content": "the whole sphere, thereby avoiding overfitting to the 26 pre-selected viewpoints (see the supplemental material for more details)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.893, + 0.903 + ], + "angle": 0, + "content": "We train all networks with the Adam optimizer with a learning rate of \\(2e^{-4}\\) for pretraining and \\(4e^{-5}\\) for finetuning on a dual NVIDIA Gforce RTX 3090 setup. On this setup, for 20 iterations, ISCNet takes 81 seconds to complete a shape from a single depth map, where the image" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8902" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.106 + ], + "angle": 0, + "content": "rendering and point clouds merging take around 59 seconds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.468, + 0.243 + ], + "angle": 0, + "content": "For validation, we measure the reconstruction quality by a mesh-to-mesh symmetric \\(L_{1}\\) distance (MD) [4] between the reference and completed meshes. Further, to quantify the reconstruction accuracy of surface details, we follow Yao et al. [39] and use the Chamfer Distance (CD) on 100K Poisson disk sampled [43] surface points from the reconstructed and reference meshes. Finally, following Zhou et al. [45] we quantify symmetry plane accuracy by comparing the cosine distance between symmetry plane normals." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.252, + 0.298, + 0.267 + ], + "angle": 0, + "content": "4.1.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.275, + 0.47, + 0.607 + ], + "angle": 0, + "content": "3D Shape Completion. We compare the performance of ISCNet to 8 competing methods subdivided in three categories: Point-cloud completion methods (PCN [41], MSN [15], PoinTr [40], VRCNet [16], ShowFlakeNet [37], VE-PCN [32]), implicit-function-based methods (IFNet [3]), and image-based completion methods (2 variants of Front2Back \\([39]^{1}\\)). All competing methods are retrained on ShapeNet and tested with identical exemplars under the same conditions. The quantitative comparisons (using the CD and MD metrics) are summarized in Table 1 and Table 2 for ShapeNet and ScanNet, respectively. From these, we can see that ISCNet offers the best overall performance for the majority of categories on both synthetic and scanned data, followed by Front2Back [39]. SnowFlakeNet [37] performs better than ISCNet on the plane shapes with a marginal accuracy gain. Compared to Front2Back, we can see that ISCNet outperforms Front2Back with the ground truth symmetry information. For completeness, Table 3 compares ISCNet with Front2Back and IFNet for the remaining five classes from ShapeNet, without retraining or finetuning. Our approach still outperforms existing work except for the rifle class compared to IFNet." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.619, + 0.495, + 0.702 + ], + "angle": 0, + "content": "
(×102)BenchDisplaySpeakerRiflePhoneAverage (unseen-5)Average (total-13)
CDF2B2.373.193.401.723.132.762.24
IFNet1.813.736.351.373.343.323.02
Ours1.442.473.091.432.102.111.72
MDF2B2.192.973.171.592.972.582.11
IFNet1.773.656.241.353.283.263.01
Ours1.412.453.011.422.072.071.61
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.7, + 0.468, + 0.741 + ], + "angle": 0, + "content": "Table 3. Quantitative evaluation of ISCNet compared to Front2Back and IFNet for the remaining 5 classes of ShapeNet, i.e., unseen-5 labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.755, + 0.468, + 0.83 + ], + "angle": 0, + "content": "Figure 4 and 5 provide additional qualitative comparisons. Visually, ISCNet provides the 'cleanest' shape completion for all eight tasks. In particular, in the plane completion task, ISCNet is better able to retain high-frequency surface details, and in the cabinet and lamp completion tasks" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "ISCNet is the only method to fully complete the surface without holes. In the case of the chair and table completion tasks, ISCNet more robustly identifies areas to be completed without erroneously merging thin features. Although our method is designed for symmetric objects, we also obtain good reconstruction quality on non-symmetric objects (sofa and table in Figure 5)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.219, + 0.892, + 0.4 + ], + "angle": 0, + "content": "Symmetry Detection. To assess the accuracy of the symmetry plane estimation, we compare it against two state-of-the-art single-view based pose estimation and symmetry detection methods, RotationContinuity [44] and NeRD [45]. Figure 6a shows that ISCNet's second update of the symmetry plane offers a similar performance as NeRD. Furthermore, with increasing iterations, the accuracy of the estimated symmetry plane improves (i.e., compared to the first symmetry plane estimate2, the occurrence of symmetry plane estimates with an error less than \\(5^{\\circ}\\), improves from \\(32.86\\%\\) to \\(56.96\\%\\)), demonstrating the effectiveness of the iterative refinement for symmetry plane estimation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.411, + 0.653, + 0.427 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.435, + 0.892, + 0.737 + ], + "angle": 0, + "content": "Importance of Symmetry. We validate the effectiveness of symmetry cues for shape completion in two ablation experiments. As a first experiment (Table 1, Unaligned ISCNet), we replace the estimated transformation that aligns the symmetry plane with the X-Y plane by a random rotation matrix and a perturbed center (adding normal distributed noise with \\(\\sigma = 0.1\\)). Consequently, ISCNet cannot leverage any symmetry cues. In a second experiment (Table 1, No-Pair ISCNet), we process each of the reflection viewpoints separately. From Table 1, we can see that the unaligned ISCNet suffers a significant performance degradation; CD increases by 0.56 on average, and MD increases 0.55 on average. By aligning the shape to the symmetry plane, no-pair ISCNet is able to leverage symmetry implicitly (by learning to copy from the negative Z coordinate). However, compared to the full ISCNet, the no-pair ISCNet still sees a performance drop (CD and MD are increased by 0.34 and 0.28, respectively). We refer to the supplementary material for a qualitative comparison of excluding explicit leveraging of the symmetry cues." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.758, + 0.892, + 0.865 + ], + "angle": 0, + "content": "Role of Normal Inpainting. To demonstrate the importance of performing normal inpainting, we eliminate the inpainted normal as input from the depth inpainting submodule (Table 1, No-Normal ISCNet). Without normal cues, CD increases by 0.46 and MD by 0.36. We refer to the supplementary material for a qualitative visualization of the impact of omitting the normal inpainting step." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.89, + 0.9 + ], + "angle": 0, + "content": "2The accuracy of first symmetry estimate is similar to that of DenseFusion [30], as our symmetry estimation is based this method." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.839, + 0.468, + 0.9 + ], + "angle": 0, + "content": "1The information in [39] was insufficient to replicate the symmetry plane estimation module. Therefore, we include 2 variants, trained and tested with the ground truth symmetry plane (Front2Back_GT), and with the reference symmetry plane normal perturbed within \\([-5^{\\circ}, 5^{\\circ}]\\) (Front2Back_R5)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8903" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.214 + ], + "angle": 0, + "content": "Impact of Iterative Refinement. To demonstrate that both symmetry plane estimation as well as shape completion improve with additional iterations, we compare the MD and CD errors per iteration (Table 1) as well as the symmetry plane normal error progression per iteration (Figure 6b). In both cases, we can see that the error improves per iteration and the improvement diminishes per subsequent iteration (justifying a 20 iteration stopping criterion)." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.228, + 0.27, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.354, + 0.197, + 0.364 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.228, + 0.468, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.354, + 0.375, + 0.364 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.37, + 0.47, + 0.44 + ], + "angle": 0, + "content": "Figure 6. Impact of the number of iterations on (a) the accuracy of the symmetry plane estimation, and (b) the shape reconstruction CD/MD accuracy and average and median symmetry plane estimation errors. 'Our SP \\( n \\)' refers to ISCNet's \\( n \\)-th predicted symmetry plane." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.47, + 0.531 + ], + "angle": 0, + "content": "Limitations The main bottleneck in ISCNet is the image rendering and the 3D point cloud processing step in each iteration. Furthermore, ISCNet is only able to leverage a single global symmetry plane, and thus its performance is suboptimal when multiple symmetries are present." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.546, + 0.195, + 0.561 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.47, + 0.754 + ], + "angle": 0, + "content": "We present ISCNet, a single-depth image joint symmetry plane estimation and shape completion method that explicitly leverages symmetry cues. ISCNet alternates between estimating the symmetry plane and completing the shape to enhance the accuracy of both components. Our method is informed by two key insights: (1) shape completion and symmetry plane estimation are mutually beneficial tasks, and (2) by selecting symmetric virtual viewpoints, we can explicitly inform the inpainting module on symmetry relations. Our experiments and ablation studies demonstrate that ISCNet can yield more accurate shape completions from a single-depth image than competing methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.77, + 0.236, + 0.787 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We thank the anonymous reviewers for the insightful and constructive comments. This work was supported in part by National Key Research and Development Program of China (2022ZD0210500, 2021ZD0111902), National Natural Science Foundation of China (61972067, U19B2039, U21B2038), the Distinguished Young Scholars Funding of Dalian (No. 2022RJ01). Pieter Peers was supported in part" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "by NSF grant IIS-1909028. Felix Heide was supported by an NSF CAREER Award (2047359), a Packard Foundation Fellowship, a Sloan Research Fellowship, a Sony Young Faculty Award, a Project X Innovation Award, and an Amazon Science Research Award." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.181, + 0.597, + 0.197 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.206, + 0.895, + 0.277 + ], + "angle": 0, + "content": "[1] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.278, + 0.895, + 0.346 + ], + "angle": 0, + "content": "[2] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.348, + 0.895, + 0.404 + ], + "angle": 0, + "content": "[3] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6970-6981, 2020. 1, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.405, + 0.895, + 0.447 + ], + "angle": 0, + "content": "[4] Paolo Cignoni, Claudio Rocchini, and Roberto Scopigno. Metro: measuring error on simplified surfaces. In Computer Graphics Forum, volume 17, pages 167-174, 1998. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.448, + 0.895, + 0.517 + ], + "angle": 0, + "content": "[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.518, + 0.895, + 0.572 + ], + "angle": 0, + "content": "[6] Shivam Duggal and Deepak Pathak. Topologically-aware deformation fields for single-view 3d reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1536-1546, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.575, + 0.895, + 0.63 + ], + "angle": 0, + "content": "[7] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 605-613, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.632, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[8] Boqing Gong, Jianzhuang Liu, Xiaogang Wang, and Xiaou Tang. Learning semantic signatures for 3d object retrieval. IEEE Trans. Multimedia, 15(2):369-377, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.675, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[9] Xiaoguang Han, Zhaoxuan Zhang, Dong Du, Mingdai Yang, Jingming Yu, Pan Pan, Xin Yang, Ligang Liu, Zixiang Xiong, and Shuguang Cui. Deep reinforcement learning of volume-guided progressive view inpainting for 3d point scene completion from a single depth image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 234-243, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.76, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[10] Sagar Imambi, Kolla Bhanu Prakash, and GR Kanagachidambaresan. Pytorch. In Programming with TensorFlow, pages 87-104. 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[11] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[12] Michael Kazhdan, Thomas Funkhouser, and Szymon Rusinkiewicz. Symmetry descriptors and 3d shape matching. In Proceedings of the 2004 Eurographics/ACM SIGGRAPH" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.206, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8904" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "symposium on Geometry processing, pages 115-123, 2004. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.175 + ], + "angle": 0, + "content": "[13] Wai Ho Li and Lindsay Kleeman. Segmentation and modeling of visually symmetric objects by robot actions. The International Journal of Robotics Research, 30(9):1124-1142, 2011. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[14] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Eur. Conf. Comput. Vis., pages 85-100, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.469, + 0.287 + ], + "angle": 0, + "content": "[15] Minghua Liu, Lu Sheng, Sheng Yang, Jing Shao, and ShiMin Hu. Morphing and sampling network for dense point cloud completion. In AAAI, volume 34, pages 11596-11603, 2020. 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.289, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[16] Liang Pan, Xinyi Chen, Zhongang Cai, Junzhe Zhang, Haiyu Zhao, Shuai Yi, and Ziwei Liu. Variational relational point completion network. In Int. Conf. Comput. Vis., pages 8524-8533, 2021. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.4 + ], + "angle": 0, + "content": "[17] Joshua Podolak, Philip Shilane, Aleksey Golovinskiy, Szymon Rusinkiewicz, and Thomas Funkhouser. A planar-reflective symmetry transform for 3d shapes. In ACM SIGGRAPH 2006 Papers, pages 549-559. 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.401, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[18] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 652-660, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.457, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[19] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.567 + ], + "angle": 0, + "content": "[20] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[21] Yifei Shi, Junwen Huang, Hongjia Zhang, Xin Xu, Szymon Rusinkiewicz, and Kai Xu. Symmetrynet: learning to predict reflectional and rotational symmetries of 3d shapes from single-view rgb-d images. ACM Trans. Graph., 39(6):1-14, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[22] Yifei Shi, Xin Xu, Junhua Xi, Xiaochang Hu, Dewen Hu, and Kai Xu. Learning to detect 3d symmetry from single-view rgb-d images with weak supervision. IEEE Trans. Pattern Anal. Mach. Intell., 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.693, + 0.469, + 0.748 + ], + "angle": 0, + "content": "[23] Ivan Sipiran, Robert Gregor, and Tobias Schreck. Approximate symmetry detection in partial 3d meshes. In Computer Graphics Forum, volume 33, pages 131-140. Wiley Online Library, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[24] Pablo Speciale, Martin R Oswald, Andrea Cohen, and Marc Pollefeys. A symmetry prior for convex variational 3d reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 313-328. Springer, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[25] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B Tenenbaum, and William T Freeman. Pix3d: Dataset and methods for single-image 3d shape modeling. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2974-2983, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[26] Junshu Tang, Zhijun Gong, Ran Yi, Yuan Xie, and Lizhuang Ma. Lake-net: Topology-aware point cloud completion by localizing aligned keypoints. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1726-1735, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[27] Demetri Terzopoulos, Andrew Witkin, and Michael Kass. Symmetry-seeking models and 3d object reconstruction. Int. J. Comput. Vis., 1(3):211-221, 1988. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[28] Sebastian Thrun and Ben Wegbreit. Shape from symmetry. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 2, pages 1824-1831. IEEE, 2005. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[29] Shubham Tulsiani, Tinghui Zhou, Alexei A Efros, and Jitendra Malik. Multi-view supervision for single-view reconstruction via differentiable ray consistency. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2626-2634, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[30] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martin-Martin, Cewu Lu, Li Fei-Fei, and Silvio Savarese. Densefusion: 6d object pose estimation by iterative dense fusion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3343–3352, 2019. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[31] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In *Eur. Conf. Comput. Vis.*, pages 52-67, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.435, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[32] Xiaogang Wang, Marcelo H Ang, and Gim Hee Lee. Voxel-based network for shape completion by leveraging edge generation. In Int. Conf. Comput. Vis., pages 13189-13198, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[33] Xiaogang Wang, Marcelo H Ang Jr, and Gim Hee Lee. Cascaded refinement network for point cloud completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 790-799, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[34] Boyan Wei, Xianfeng Ye, Chengjiang Long, Zhenjun Du, Bangyu Li, Baocai Yin, and Xin Yang. Discriminative active learning for robotic grasping in cluttered scene. IEEE Robotics and Automation Letters, 8(3):1858-1865, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.605, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[35] Jiajun Wu, Yifan Wang, Tianfan Xue, Xingyuan Sun, Bill Freeman, and Josh Tenenbaum. Marrnet: 3d shape reconstruction via 2.5 d sketches. Adv. Neural Inform. Process. Syst., 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[36] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-10, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[37] Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. Snowflakenet: Point cloud completion by snowflake point deconvolution with skip-transformer. In Int. Conf. Comput. Vis., pages 5499-5509, 2021. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[38] Xin Yang, Yuanbo Wang, Yaru Wang, Baocai Yin, Qiang Zhang, Xiaopeng Wei, and Hongbo Fu. Active object reconstruction using a guided view planner. In Proceedings of International Joint Conference on Artificial Intelligence, pages 4965-4971, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[39] Yuan Yao, Nico Schertler, Enrique Rosales, Helge Rhodin, Leonid Sigal, and Alla Sheffer. Front2back: Single view 3d" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8905" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "shape reconstruction via front to back prediction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 531-540, 2020. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[40] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In Int. Conf. Comput. Vis., pages 12498-12507, 2021. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.234 + ], + "angle": 0, + "content": "[41] Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, and Martial Hebert. Pcn: Point completion network. In Int. Conf. 3D Vis., pages 728-737, 2018. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.47, + 0.304 + ], + "angle": 0, + "content": "[42] Zhaoxuan Zhang, Xiaoguang Han, Bo Dong, Tong Li, Baocai Yin, and Xin Yang. Point cloud scene completion with joint color and semantic estimation from single rgb-d image. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.47, + 0.347 + ], + "angle": 0, + "content": "[43] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3d: A modern library for 3d data processing. arXiv preprint arXiv:1801.09847, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.47, + 0.403 + ], + "angle": 0, + "content": "[44] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5745-5753, 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.47, + 0.448 + ], + "angle": 0, + "content": "[45] Yichao Zhou, Shichen Liu, and Yi Ma. Nerd: neural 3d reflection symmetry detector. In IEEE Conf. Comput. Vis. Pattern Recog., pages 15940-15949, 2021. 1, 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "8906" + } + ] +] \ No newline at end of file diff --git a/2023/Single Depth-image 3D Reflection Symmetry and Shape Prediction/fcff96a5-45a1-4f7e-ab9f-0c309af710cd_origin.pdf b/2023/Single Depth-image 3D Reflection Symmetry and Shape Prediction/fcff96a5-45a1-4f7e-ab9f-0c309af710cd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c58408d17b76aefef0dda52fc908c45bae565062 --- /dev/null +++ b/2023/Single Depth-image 3D Reflection Symmetry and Shape Prediction/fcff96a5-45a1-4f7e-ab9f-0c309af710cd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a24f9e20edc172422cc60dd0bb774520facfefc759bfa9336ed8df2b683613f5 +size 3697245 diff --git a/2023/Single Depth-image 3D Reflection Symmetry and Shape Prediction/full.md b/2023/Single Depth-image 3D Reflection Symmetry and Shape Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..97f3cd9ce30ef279a059e5d9ee407236bae9d972 --- /dev/null +++ b/2023/Single Depth-image 3D Reflection Symmetry and Shape Prediction/full.md @@ -0,0 +1,372 @@ +# Single Depth-image 3D Reflection Symmetry and Shape Prediction + +Zhaoxuan Zhang, Bo Dong*, Tong Li, Felix Heide, Pieter Peers, Baocai Yin*, Xin Yang* + +# Abstract + +In this paper, we present Iterative Symmetry Completion Network (ISCNet), a single depth-image shape completion method that exploits reflective symmetry cues to obtain more detailed shapes. The efficacy of single depth-image shape completion methods is often sensitive to the accuracy of the symmetry plane. ISCNet therefore jointly estimates the symmetry plane and shape completion iteratively; more complete shapes contribute to more robust symmetry plane estimates and vice versa. Furthermore, our shape completion method operates in the image domain, enabling more efficient high-resolution, detailed geometry reconstruction. We perform the shape completion from pairs of viewpoints, reflected across the symmetry plane, predicted by a reinforcement learning agent to improve robustness and to simultaneously explicitly leverage symmetry. We demonstrate the effectiveness of ISCNet on a variety of object categories on both synthetic and real-scanned datasets. + +# 1. Introduction + +Symmetry is an intrinsic geometric property present in the vast majority of man-made and natural objects [12, 17], and which has been exploited to improve the efficacy of applications such as shape matching [12], segmentation [13], object retrieval [8], and robotic grasping [34]. + +In this paper, we aim to leverage symmetry as an aid for the completion of shapes from partial observations [23, 24, 28]. Recently, Yao et al. [39] presented Front2Back, a framework for 3D shape reconstruction from a single RGB image. A key observation in Front2Back is that a (nearly) complete 3D model can be effectively described by a pair of 2.5D visible surface images taken from opposite views (i.e., front and back views). Based on this observation, Front2Back completes a shape by synthesizing a back view using a global 3D reflective symmetry plane from a given front view. However, the reconstruction quality is significantly affected by the input front view angle (Figure 1(d)). Moreover, Front2Back assumes the availability of a 'perfect' symmetry plane, which in practice is challenging to obtain from partial observations [21, 22, 45]. + +![](images/9f65dacff26d8142651f49b7c010a15e32a60362be274588248a08729da8b788.jpg) +(a) Input Depth Map & Observed Surface + +![](images/62b67ac4a0495c76f7fcfbc44cd1dee505e2450af71025e1dfd9e61a6cc15eff.jpg) +(b) PoinTr [40] + +![](images/45e3b261f8434c20c6712b131cc05a4f62a92430b17c52552817f01a327efc1e.jpg) +(c) IF-Net [3] + +![](images/ac5b3c58597abcef32d4ecf2fb59d32e7da9043fdd5ea829b28a3d5eebaebe0e.jpg) +(d) Front2Back [39] + +![](images/e8e205a5f71b82ebf239db0d51f22344d918beed4737b561c41bb74594d22542.jpg) +(e) Ours + +![](images/948cade7807be767324c0d2e9ab6b6c72db87805a993e9096ca22fd00a671f82.jpg) +(f) Ground Truth +Figure 1. ISCNet more efficiently leverages symmetry cues for robust shape completion from a single depth image compared to prior methods. + +We present ISCNet (Iterative Shape Completion Network), a single depth-image shape completion method that explicitly leverages symmetry cues. Our key insight is that symmetry detection and symmetry-based shape completion are complementary tasks that can provide constructive cues to the other task. Symmetry detection can help shape completion by providing object-centric information about missing geometrical features, and shape completion can provide more robust symmetry cues by filling in missing geometry. Departing from existing work [33] that assumes an ideal prediction of the symmetry plane, ISCNet iteratively refines the symmetry plane and the reconstructed shape jointly. Furthermore, unlike Front2Back [39], ISCNet does not rely on the initial input viewpoint to a-priori fix the synthetic back view camera position but instead leverages a trained reinforcement learning (RL) policy to select optimal pairs of reflection viewpoints at each iteration. We empirically found that predetermining the viewpoints is less effective than RL-based viewpoint selection which can more effectively deal with the dynamic shape of a partial point cloud during the refinement process. + +Specifically, ISCNet takes as input a single depth image and reprojects it to an incomplete point cloud. Instead of explicitly detecting a symmetry plane, we estimate the pose [30] that aligns the symmetry plane to the X-Y plane. Based on the current estimate of the point cloud and symmetry aligning pose, an RL agent determines a pair of $re$ - + +flection viewpoints, reflected across the symmetry plane, for extracting a pair of partial depth and normal maps that are subsequently repaired using a novel 2D inpainting module. By directly working on depth image pairs, our method is independent of the size of the point cloud, and it can leverage geometry information from the symmetric counter view. However, because the virtual depth images are reprojections that differ from the input view, not all reprojected points are valid (e.g., the backside of points visible through a 'hole'). Therefore, we first identify reliable pixels before performing depth inpainting in the projected image domain. Finally, we reproject the inpainted maps back to 3D and update the point cloud. We progressively refine the point cloud by alternating between estimating reflection viewpoints and completing the point cloud for 20 iterations, updating the symmetry plane every 4th iteration, to balance efficiency versus efficacy. When updating the symmetry plane, we 'reset' the point cloud to its initial state to mitigate the impact of prior inaccurate symmetry plane estimations on the shape completion. + +We validate that ISCNet produces more accurate shape completions than existing methods (Figure 1) and confirm the contribution of each component in our model via a thorough ablation study. In summary, our contributions are: 1) a novel robust joint symmetry-detection / shape completion method; 2) a symmetric viewpoint-based 3D shape completion method for detailed high-resolution shape reconstruction; and 3) a reinforcement learning-based optimal reflection viewpoint selection solution for progressive shape completion. + +# 2. Related Work + +Symmetry Detection on 3D Shapes forms a key component in many downstream applications ranging from shape matching [12], to segmentation [13], to 3D object retrieval [8], to shape completion [23, 24, 27, 28] and to deformable object learning [36]. Symmetry detection of incomplete shape observations is significantly more challenging, and recent learning-based solutions have explored the use of multitask learning [21], sampling the distribution of plausible shapes [22], and building 3D cost volumes [45] to predict 3D symmetry planes. An alternative approach to predicting arbitrarily oriented 3D symmetry planes is to estimate a camera pose that aligns the local coordinate frame such that the X-Y plane acts as a canonical symmetry plane [30, 44]. Our symmetry detection module falls in this latter category. However, unlike prior work that a priori fixes the symmetry information as a prior, we jointly perform shape completion yielding a more robust and mutually reinforcing symmetry detection and shape completion. + +Learning-based 3D Shape Completion of Partial Point Cloud Observations have explored coarse-to-fine methods [15, 41], transformer-networks [37, 40], probabilis- + +tic modeling [16], and topology-aware approaches [26] to complete the 3D point cloud. However, these methods are memory intensive and their scalability (e.g., beyond $\approx 16384$ points) is limited. In contrast, our approach does not impose such limitation on the number of 3D points as our method leverages 2D depth inpainting. + +Single-image Shape Reconstruction aims to complete approximately half of the shape either directly from a 2D image [6, 7, 29, 31] or (indirectly) from a 2.5D input [25, 35, 39]. Front2Back [39] estimates, besides depth and surface normals, also symmetry information to aid in the surface completion task. However, Front2Back is limited to a single front-back camera pair. Our method leverages an RL agent to predict multiple virtual reflection viewpoints for shape completion, and it takes better advantage of the symmetry prior by inpainting symmetric image pairs. In addition, we iteratively refine the symmetry plane to obtain more accurate estimates of both shape and symmetry. + +# 3. Iterative-Symmetry Completion Network + +Our Iterative Symmetry Completion Network (ISCNet) takes a single-view depth map of a target object as input and iteratively completes the 3D mesh by alternating between refining the symmetry plane, the virtual reflection viewpoints, and the shape completion. To bootstrap the iterative process, we first project the input single-view depth map to an incomplete point cloud that is subsequently passed to a symmetry detection module and point cloud transposition operator. The symmetry detection module (subsection 3.1) provides an estimate of the object's symmetry plane that is subsequently leveraged by the point cloud transposition operator to align the point cloud symmetrically around the estimated symmetry plane. Next, a novel 3D shape completion module (subsection 3.2) progressively completes the 3D point cloud based on a pair of reflection viewpoints selected by an RL agent. For each reflection viewpoint pair, a depth and normal map pair is extracted from the partial point cloud. Due to the partial completion of the point cloud, the extracted depth and normal map pairs exhibit holes. To address this issue, we employ a specifically crafted inpainting module, which not only repairs these holes but also assigns a confidence score to each repaired pixel. Next, the inpainted depth and normal maps are merged into the point cloud guided by their associated confidence scores. Finally, to improve robustness, we iteratively refine the symmetry plane, the reflection viewpoints, and the shape completion (subsection 3.3). We limit the number of iterations to 20 and only update the symmetry plane and reset the point cloud every 4th iteration to balance robustness versus efficiency; Figure 2 visually summarizes ISCNet. + +![](images/a275c2232977bca901e91c8827aaa14f65d2274e8c0f0153c50909992d5ca5c6.jpg) +Figure 2. Summary of ISCNet: a partial shape extracted from a single depth-image is aligned such that the X-Y plane corresponds to the (detected) symmetry plane. Next, a pair of depth and normal maps are rendered from reflection viewpoints selected by an RL agent, and each symmetric pair of synthetic depth and normal maps is inpainted in the image-domain (see also Figure 3). These inpainted maps are then back-projected to a point cloud and merged. The symmetry plane is updated every 4th iteration, where the merged shape is then visualized from 3 selected views, $v_{8}^{V}$ , $v_{12}^{V}$ , and $v_{13}^{V}$ in this case, passing into the symmetry detection module separately, and the best-detected symmetry plane (with the highest confidence score) is used as input for the next iteration. + +# 3.1. 3D Reflection Symmetry Detection + +We estimate the symmetry plane of an object via a proxy task, namely, camera pose estimation, such that the symmetry plane aligns in local coordinates with the X-Y plane. Our Symmetry Detection Module follows the architecture of DenseFusion [30]. We pass a depth image to obtain a point cloud, pixel-wise features, and finally a pose estimation. Different from DenseFusion, we only use depth instead of RGB channels, and we do not require an object segmentation stage. Instead of predicting the symmetry parameters directly, we found it more robust to estimate a rotation matrix $R \in SO(3)$ and a translation vector $t \in \mathbb{R}^3$ such that the symmetry plane's center point $p_s$ and normal direction $n_s$ are defined by: + +$$ +\left( \begin{array}{c} p _ {s} \\ n _ {s} \end{array} \right) = \left( \begin{array}{c} R ^ {- 1} \cdot ([ 0, 0, 0 ] - t) \\ R ^ {- 1} \cdot ([ 0, 0, 1 ] - t) \end{array} \right). \tag {1} +$$ + +# 3.2. 3D Shape Completion + +Given a predicted symmetry plane, ISCNet progressively completes the 3D shape. Our shape completion module consists of three sequential sub-modules: viewpoint selection, 2D inpainting, and point cloud updating. + +Viewpoint Selection Module: ISCNet completes 3D shapes via 2D projected depth and normal maps. The exact choice of virtual viewpoints from which these depth and normals are generated is critical for robust shape completion, and it is sensitive to the irregular structures present in the partially completed point cloud. As the shape of a partial point cloud dynamically changes, it is impractical to pre-determine viewpoints; see our supplementary material for more details. + +We model the problem of selecting a viewpoint $v_{i}$ as a Markov decision process (MDP) because, similar to [38, + +42], we are only concerned about the current state of a partial point cloud. To this end, we leverage an RL agent to select an optimal viewpoint trained with an Actor-Critic Proximal Policy Optimisation (PPO) algorithm [20]. The RL agent predicts an optimal viewpoint (i.e., action) based on the current 3D point cloud (i.e., state) to maximize an appropriately designed reward. + +Specifically, the continuous action space of the RL agent is defined by $\theta$ and $\phi$ angles (Figure 2), with corresponding viewpoint: + +$$ +x = a \sin \theta \sin \phi , \tag {2} +$$ + +$$ +y = a \cos \theta , \tag {3} +$$ + +$$ +z = a \sin \theta \cos \phi , \tag {4} +$$ + +where $\theta \in [0^{\circ},180^{\circ}]$ and $\phi \in [-180^{\circ},180^{\circ}]$ ; $a$ is set at $1.5\mathrm{m}$ . The view direction is always aimed at the object center. + +The prediction action reward is determined by four factors: accuracy $R^a$ , shape completeness $R^s$ , number of recovered effective points $R^p$ , and viewpoint diversity $R^v$ . $R^a$ and $R^s$ encourage an agent to choose a new viewpoint that benefits the prediction of missing parts from the original point cloud. $R^p$ and $R^v$ encourage the agent to increase the viewpoint diversity. Formally, the reward function for + +![](images/ce7e296968062e1675be6315403d26c904f2c9c3594821dd76f30e2f3d4c64f2.jpg) +Figure 3. Detailed visual summary of the viewpoint selection and the 2D inpainting modules. + +an action is defined as: + +$$ +R = w _ {1} R ^ {a} + w _ {2} R ^ {s} + w _ {3} R ^ {p} + w _ {4} R ^ {v}, \tag {5} +$$ + +$$ +R ^ {a} = \frac {1}{\mathcal {N} (\hat {P})} \sum_ {x \in \hat {P}} \min _ {y \in P _ {G T}} \| x - y \| _ {2} ^ {2}, \tag {6} +$$ + +$$ +R ^ {s} = S - S ^ {\prime}, \tag {7} +$$ + +$$ +R ^ {p} = \mathcal {N} \left(\left\{d (x, P ^ {\prime}) < 0. 0 0 2 | x \in \hat {P} \right\}\right), \tag {8} +$$ + +$$ +R ^ {v} = \left(\left| \theta - \theta^ {\prime} \right| + \left| \phi - \phi^ {\prime} \right|\right) / 1 8 0 \cdot \pi , \tag {9} +$$ + +$$ +S = \frac {\mathcal {N} \left(\left\{d (y , P) < 0 . 0 2 \mid y \in P _ {G T} \right\}\right)}{\mathcal {N} \left(P _ {G T}\right)}, \tag {10} +$$ + +where $P, P'$ , and $\hat{P}$ are the 3D points of the current point cloud, previous point cloud, and recovered point cloud by the action, respectively. Intuitively, $P = P' + \hat{P}$ ; $P_{GT}$ is the ground truth reference point set; $d(x, P)$ denotes the shortest distance from a point $x$ to a point set $P$ ; $S$ and $S'$ are defined based on $P$ and $P'$ , respectively; $\mathcal{N}(\cdot)$ counts the number of points of a point set; $w_1, w_2, w_3, w_4$ are the weights for balancing the contribution of the corresponding reward, which are set to 25, 1, 1.5, and 2 in all experiments. + +The architecture of the RL agent is simple yet effective; it contains an encoder, decoder, critic and actor heads (Figure 3). We adopt a commonly used 3D point encoder [15, 18] which extracts features from the input point cloud for downstream processing by the actor and critic. The critic head is only used during training, and the actor head predicts a viewpoint based on the input 3D point cloud. + +The 2D Inpainting Module consists of three components: mask generation, normal map inpainting, and depth-map inpainting. + +The mask generator takes as input symmetric depth map pairs $D_{i}$ and $D_{i}^{\prime}$ corresponding to symmetric views $v_{i}$ and + +$v_{i}^{\prime}$ , and outputs two masks $M_{i}$ and $M_{i}^{\prime}$ that mark the areas for each input map that need to be completed. To compute $M_{i}$ , we start from an initial completion mask $\hat{M}_i$ obtained by marking all 'black' areas of a depth map $D_{i}$ . However, this initial mask misses two important corner cases. First, not all 'black' areas should be completed. For example, the space between the legs of a chair. Second, areas with a missing front surface covered by a backside surface also need to be filled in. To cover the first case, we utilize a silhouette map generator to predict silhouette maps $K_{i}$ of the target object to exclude erroneous 'holes'. The silhouette generation is modified from a standard U-Net architecture [19] by adding an extra input head and output head to address the pairwise inputs and outputs. Furthermore, we add pairwise skip connections between the two input and output heads for exchanging symmetry information. To provide additional depth cues, we concatenate to each input $D_{i}$ (and $D_{i}^{\prime}$ ) the initial input depth map, $D^{I}$ , along the channel dimension. To address the second corner case, we compute a front-surface mask $H_{i}$ by comparing the angle between the camera view direction and the surface normal at each point; we set the front-surface mask to 1 if the angle is less than $90^{\circ}$ , or 0 otherwise (i.e., back facing). The final mask $M_{i}$ is then obtained by combining $\hat{M}_{i}$ , $K_{i}$ and $H_{i}$ as: + +$$ +M _ {i} = \left(\hat {M} _ {i} + H _ {i}\right) \odot K _ {i}, \tag {11} +$$ + +$$ +\hat {M} _ {i} = \Gamma (D _ {i}), \tag {12} +$$ + +$$ +H _ {i} = \Psi \left(N _ {i}, v _ {i}\right), \tag {13} +$$ + +$$ +K _ {i}, K _ {i} ^ {\prime} = \mathscr {K} \left(\left[ D _ {i}, D ^ {I} \right], \left[ D _ {i} ^ {\prime}, D ^ {I} \right]\right), \tag {14} +$$ + +$$ +\Gamma (X), \Psi (Z, y) = \left\{ \begin{array}{l l} 1 & X = 0, \langle Z, y \rangle > 0 \\ 0 & \text {o t h e r w i s e} \end{array} , \right. \tag {15} +$$ + +where $\odot$ is the element-wise multiplication; $\mathcal{H}$ denotes the + +silhouette map generator; $[\cdot, \cdot]$ indicates a channel-wise concatenation operation; $\Gamma(X)$ and $\Psi(Z, y)$ are two indicator functions, that only differ in the conditions for 1. + +The normal inpainting submodule $\mathcal{N}$ takes as input a pair of partial normal maps $N_{i}$ and $N_{i}^{\prime}$ corresponding to symmetric viewpoints $v_{i}$ and $v_{i}^{\prime}$ as well as their corresponding masks $M_{i}$ and $M_{i}^{\prime}$ , and outputs a pair of completed normal maps $\hat{N}_i$ and $\hat{N}_i^\prime$ respectively: + +$$ +\hat {N} _ {i}, \hat {N} _ {i} ^ {\prime} = \mathcal {N} ([ N _ {i}, \mathbf {1} - M _ {i} ], [ N _ {i} ^ {\prime}, \mathbf {1} - M _ {i} ^ {\prime} ]). \tag {16} +$$ + +The normal inpainting submodule is based on PConv [14], a partial-convolution-layer-based UNet architecture that is robust to irregularly shaped holes [9, 14]. We modify PConv to accommodate the dual in and output by duplicating the input head and last layer of the decoder and adding pairwise skip connections. Note, PConv also applies a convolution with a unit fixed kernel on the input mask to perform mask-updating, and hence the target inpainting region should be marked 0 in the mask (i.e., the inverse of mask $M_{i}$ ). + +The depth-map inpainting submodule $\mathcal{D}$ takes as input a pair of partial depth maps $D_{i}$ and $D_{i}^{\prime}$ corresponding to symmetric viewpoints $v_{i}$ and $v_{i}^{\prime}$ as well as their corresponding masks ( $M_{i}$ and $M_{i}^{\prime}$ ) and inpainted normal maps ( $\hat{N}_{i}$ and $\hat{N}_{i}^{\prime}$ ), and outputs a pair of completed depth maps $\hat{D}_i$ and $\hat{D}_i^\prime$ and corresponding confidence maps $C_i$ and $C_i^\prime$ : + +$$ +\begin{array}{l} (\hat {D} _ {i}, C _ {i}), (\hat {D} _ {i} ^ {\prime}, C _ {i} ^ {\prime}) = \mathcal {D} \left(\left[ D _ {i}, \hat {N} _ {i}, \mathbf {1} - M _ {i} \right], \right. \\ \left[ D _ {i} ^ {\prime}, \hat {N} _ {i} ^ {\prime}, \mathbf {1} - M _ {i} ^ {\prime} \right]). \tag {17} \\ \end{array} +$$ + +The confidence scores indicate the confidence of the inpainted pixel (i.e., the distance of the estimated depth-map to the ground truth map on missing regions). Our implementation of the depth inpainting submodule is also based on PConv [14] with similar modifications to support the additional input and output channels. + +Point Cloud Update: We merge the inpainted depth, normal, and confidence maps into the previous point cloud by reprojecting all points in each depth map and corresponding normals to a point cloud if their confidence score exceeds a threshold (set to 0.5 in our implementation). + +# 3.3. Iterative Refinement + +To improve the robustness of the symmetry plane estimation and of the shape completion, we iteratively refine the output of ISCNet. A key observation is that we can select any suitable viewpoint to generate a new depth image to serve as input to update the symmetry plane. Because it is unlikely that a single viewpoint would be ideal for all cases, we project the updated point cloud from three pre-selected trial viewpoints to generate their respective depth maps. Next, the three depth maps are passed (separately) into the symmetry detection module, and the symmetry plane with + +the highest confidence score is subsequently used for the remainder of our pipeline. The set of pre-determined viewpoints $(v_{8}^{V}, v_{12}^{V}$ and $v_{13}^{V}$ from VRCNet [16]) are empirically chosen and kept constant for all experiments. The final updated point cloud is converted to a mesh surface using Poisson Surface Reconstruction [11]; see Figure 2. + +To balance computational cost versus gain in accuracy, the number of iterations is set to 20 based on an experimental analysis on ShapeNet. We update the symmetry plane every 4th iteration such that the RL agent can leverage the progressively completed point cloud for improving the reflection viewpoint estimates. In addition, to avoid accumulating incorrectly completed shape-parts due to an inaccurate symmetry plane, we reset the point cloud to its initial state when re-estimating the symmetry plane. + +# 3.4. Loss Functions + +ISCNet is trained in three stages. In the first stage, the symmetry detection component and 2D inpainting module are trained separately. In the second stage, all components except the RL agent of ISCNet are then jointly fine-tuned. Finally, we train the RL agent by using a PPO algorithm with the parameters of the other components fixed using the loss function from [20]. + +The first-stage symmetry detection loss is defined by: + +$$ +\mathcal {L} ^ {S} = \mathcal {C} \left(\mathcal {M} \left(P, S _ {\mathrm {E}}\right), \mathcal {M} \left(P, S _ {\mathrm {G}}\right)\right) + \mathcal {B} \left(\Theta_ {t _ {1}} \left(S _ {\mathrm {E}} ^ {N}, S _ {\mathrm {G}} ^ {N}\right), 1\right), \tag {18} +$$ + +where $\mathcal{C}$ measures the Chamfer Distance (CD) between the reconstructed 3D points $P$ reflected (via a reflection function $\mathcal{M}(\cdot, \cdot)$ ) over the estimated symmetry plane $S_{\mathrm{E}}$ and reference symmetry plane $S_{\mathrm{G}}$ . We further regularize this loss via a binary cross entropy (BCE) loss $(\mathcal{B})$ measuring the acute angle similarity between $S_{\mathrm{E}}^{N}$ and $S_{\mathrm{G}}^{N}$ , the corresponding normalized symmetry plane normals of the estimated and reference, via an indicator function $\Theta_{t_1}(X,Y)$ defined as: + +$$ +\Theta_ {t _ {1}} (X, Y) = \left\{ \begin{array}{l l} 1 & \operatorname {a r c c o s} (| \langle X, Y \rangle |) < = t _ {1} \\ 0 & \text {o t h e r w i s e} \end{array} , \right. \tag {19} +$$ + +with $t_1$ a slack parameter set to $5^{\circ}$ . + +The first-stage inpainting loss is defined as: + +$$ +\mathcal {L} ^ {I} = \mathcal {L} _ {\mathrm {K}} + \mathcal {L} _ {\mathrm {N}} + \mathcal {L} _ {\mathrm {D}}, \tag {20} +$$ + +$$ +\mathcal {L} _ {\mathrm {K}} = \mathcal {B} \left(K _ {\mathrm {E}}, K _ {\mathrm {G}}\right), \tag {21} +$$ + +$$ +\mathcal {L} _ {N} = \left\| N _ {E} ^ {M _ {\mathrm {G}}} - N _ {G} ^ {M _ {\mathrm {G}}} \right\| _ {1}, \tag {22} +$$ + +$$ +\mathcal {L} _ {\mathrm {D}} = \left\| D _ {\mathrm {E}} ^ {M _ {\mathrm {G}}} - D _ {\mathrm {G}} ^ {M _ {\mathrm {G}}} \right\| _ {1} + +$$ + +$$ +\mathcal {B} \left(C _ {\mathrm {E}} ^ {M _ {\mathrm {G}}}, \Phi_ {t _ {2}} \left(D _ {\mathrm {E}} ^ {M _ {\mathrm {G}}}, D _ {\mathrm {G}} ^ {M _ {\mathrm {G}}}\right)\right), \tag {23} +$$ + +$$ +X ^ {M _ {\mathrm {G}}} = X \odot M _ {\mathrm {G}}, \tag {24} +$$ + +$$ +\Phi_ {t _ {2}} (X, Y) = \left\{ \begin{array}{l l} 1 & \| X - Y \| _ {1} < = t _ {2} \\ 0 & \text {o t h e r w i s e} \end{array} , \right. \tag {25} +$$ + +
CD/MD (×102)PlaneCabinetCarChairLampSofaTableVesselAverage
PCN° [41]2.87/2.443.42/3.652.12/2.014.32/4.085.49/5.192.88/3.354.49/3.722.23/2.253.48/3.34
MSN° [15]3.29/2.642.57/2.201.95/1.693.60/2.914.58/4.313.61/3.123.62/3.042.08/1.893.16/2.72
PoinTr° [40]1.10/0.982.86/3.121.72/1.653.33/2.923.45/2.922.58/2.973.44/2.821.40/1.412.48/2.35
VRCNet° [16]2.61/1.922.77/2.201.92/1.613.43/2.613.59/2.873.41/2.733.89/3.121.92/1.572.94/2.33
SnowF° [37]1.08/0.832.83/3.021.73/1.663.08/2.813.53/2.782.46/2.513.51/2.861.52/1.302.47/2.22
VE-PCN° [32]8.80/5.396.72/5.952.91/2.377.87/6.5710.94/8.538.76/7.226.33/4.993.75/2.737.01/5.47
IFNet◇ [3]1.33/1.365.17/4.773.16/2.973.13/3.241.74/1.663.61/3.872.26/2.412.26/2.522.83/2.85
Front2Back_R5† [39]1.44/1.333.14/2.591.94/1.922.49/2.312.48/2.302.53/2.213.02/3.021.88/1.492.36/2.15
Front2Back_GT† [39]1.19/1.182.78/2.211.90/1.772.00/1.861.69/1.772.13/2.002.03/2.261.53/1.511.91/1.82
Ours† (Iteration 20)0.82/0.842.04/1.631.56/1.351.54/1.391.30/1.341.84/1.531.67/1.541.04/1.001.48/1.33
Unaligned ISCNet1.43/1.432.56/2.092.15/1.882.10/1.921.76/1.872.33/1.992.33/2.251.66/1.602.04/1.88
No-Pair ISCNet0.88/1.022.28/1.991.90/1.691.75/1.472.12/1.832.07/1.762.11/1.541.41/1.541.82/1.61
No-Normal ISCNet1.08/1.072.41/2.042.18/1.861.89/1.531.91/1.782.27/1.812.14/1.811.61/1.591.94/1.69
Iteration 40.89/0.832.23/1.941.74/1.521.67/1.541.43/1.422.03/1.791.70/1.581.15/1.061.61/1.46
Iteration 80.85/0.872.19/1.731.66/1.441.60/1.451.32/1.391.92/1.631.73/1.621.10/1.041.55/1.40
Iteration 120.84/0.872.16/1.721.60/1.381.57/1.411.28/1.371.88/1.551.70/1.581.06/1.041.51/1.36
Iteration 160.83/0.842.11/1.671.57/1.361.55/1.401.26/1.361.92/1.621.67/1.541.04/1.001.49/1.35
+ +Table 1. Quantitative evaluation of ISCNet compared to prior work on the synthetic ShapeNet dataset ( $\odot$ point cloud methods, $\diamond$ implicit function methods, and $\dagger$ image-based methods) in terms of a mesh-to-mesh symmetric ${L}_{1}$ metric (MD) and the Chamfer Distance (CD) (lowest, 2nd lowest, 3rd lowest error color coding). In addition, we show the errors on three ablation variants of ISCNet (without symmetry plane alignment, without pair-wise inpainting, and without utilizing normal information), as well as the evolution of errors per iteration. + +
CD/MD (×102)CabinetChairLampSofaTableAverage
PCN° [41]9.76/10.446.27/7.609.60/10.456.02/7.6311.83/11.948.70/9.61
MSN° [15]5.71/5.175.86/5.405.74/4.774.95/4.016.43/6.865.74/5.24
PoinTr° [40]8.33/7.686.40/6.339.66/9.196.12/5.217.84/8.477.67/7.38
VRCNet° [16]5.05/3.985.87/5.565.95/5.034.45/3.646.00/6.605.47/4.96
SnowF° [37]8.66/7.896.47/6.438.99/8.875.63/4.917.22/8.757.39/7.37
VE-PCN° [32]9.26/6.926.38/5.986.12/5.485.64/4.708.06/7.737.09/6.16
IFNet◇ [3]12.62/7.095.38/6.608.95/8.205.05/4.215.92/7.137.59/6.64
Front2Back_GT† [39]11.22/8.876.86/7.557.81/6.646.92/6.576.85/7.737.93/7.47
Ours † (Iteration 20)4.88/3.893.67/3.994.62/4.163.27/2.864.53/5.164.19/4.01
+ +Table 2. Quantitative evaluation of ISCNet compared to prior work on real-world ScanNet dataset. The notations and colors follow Table 1. + +where $\mathcal{L}_{\mathrm{K}}$ , $\mathcal{L}_{\mathrm{D}}$ and $\mathcal{L}_{\mathrm{N}}$ measure the loss over: silhouette mask $K$ , depth $D$ , and normal $N$ between the estimated $(\cdot_E)$ and reference $(\cdot_G)$ components. $\Phi_{t_2}$ is an indicator function on the $L_{1}$ distance between maps with a slack parameter $t_2$ set to 0.02, and $X^{M_G}$ is the masking operator with a reference mask over a component $X \in \{N, D, C\}$ . At this stage, we do not yet explicitly enforce symmetry to allow the inpainting network to converge to a good starting point for the second stage. + +The second-stage loss is the sum of both first-stage losses plus an additional loss that explicitly enforces symmetry: + +$$ +\mathcal {L} = \alpha \mathcal {L} ^ {S} + \beta \mathcal {L} ^ {I} + \gamma \mathcal {L} ^ {J}, \tag {26} +$$ + +$$ +\mathcal {L} ^ {J} = \Sigma_ {i = 1} ^ {4} \left(\| D _ {i} ^ {M _ {\mathrm {G}}} - \mathcal {F} \left(D _ {i} ^ {\prime}\right) \odot M _ {\mathrm {G}} \| _ {1} + \right. +$$ + +$$ +\left\| N _ {i} ^ {M _ {\mathrm {G}}} - \mathcal {F} \left(N _ {i} ^ {\prime}\right) \odot M _ {\mathrm {G}} \right\| _ {1}), \tag {27} +$$ + +where $\alpha, \beta$ , and $\gamma$ are balancing weights, $\mathcal{F}$ is a 2D horizontal mirror operator, and $X$ and $X'$ the respective components from symmetric viewpoints. + +# 4. Assessment + +We implement ISCNet in PyTorch [10], and train and test ISCNet for eight categories (airplane, cabinet, car, chair, lamp, sofa, table, and vessel) from the ShapeNet dataset [2]. As in prior work [15, 41], we partition the dataset in 29,774 shapes for training and 1,200 for testing. We use the 26 viewpoints provided by VRCNet [16] to generate synthetic depth images and normal maps for training and 8 randomly generated viewpoints for testing. In addition, we also test ISCNet with 456 scanned objects from ScanNet [5], excluding the categories not present in ScanNet (i.e. airplanes, cars, and vessels). The corresponding reference shapes are provided by Scan2CAD [1]; our and all competing methods are only trained with ShapeNet. + +To train ISCNet, we first pre-train the four submodules separately for 50 epochs: symmetry detection using $\mathcal{L}^{\mathrm{S}}$ (Equation 18), mask generation using $\mathcal{L}_{\mathrm{K}}$ (Equation 21), 2D normal inpainting using $\mathcal{L}_{\mathrm{N}}$ (Equation 22), and 2D depth inpainting using $\mathcal{L}_{\mathrm{D}}$ (Equation 23). Finally, we refine the complete network for an additional 10 epochs using $\mathcal{L}$ as loss and set the weight parameters $\alpha$ , $\beta$ , and $\gamma$ to 0.01, 0.05, and 1, respectively (Equation 26). While the + +![](images/baf130b169d97b08987abdc5ad61e14ed3031125166fb83d72d56ab672e3615e.jpg) +Figure 4. Visual comparison of shape completion quality of ISCNet versus prior work on the synthetic ShapeNet dataset; the observed areas marked in red (the second column). + +![](images/1eaab14e4e00d79233956c7a3cd95612a8147d1355ac0e63a60f0877285f8630.jpg) +Figure 5. Visual comparison of shape completion quality of ISCNet versus prior work on real-world ScanNet dataset. Note that the sofa and table are non-symmetric objects. + +training datasets are generated from 26 pre-defined viewpoints for the first training stage, the second stage (finetuning) uses randomly selected viewpoints instead of the viewpoints chosen by the RL agent. In the latter case, we generate the corresponding ground truth depth images and normal maps on the fly during the training process. As a consequence of this phased training strategy, the distribution of (RL selected) viewpoints on the testing dataset cover + +the whole sphere, thereby avoiding overfitting to the 26 pre-selected viewpoints (see the supplemental material for more details). + +We train all networks with the Adam optimizer with a learning rate of $2e^{-4}$ for pretraining and $4e^{-5}$ for finetuning on a dual NVIDIA Gforce RTX 3090 setup. On this setup, for 20 iterations, ISCNet takes 81 seconds to complete a shape from a single depth map, where the image + +rendering and point clouds merging take around 59 seconds. + +For validation, we measure the reconstruction quality by a mesh-to-mesh symmetric $L_{1}$ distance (MD) [4] between the reference and completed meshes. Further, to quantify the reconstruction accuracy of surface details, we follow Yao et al. [39] and use the Chamfer Distance (CD) on 100K Poisson disk sampled [43] surface points from the reconstructed and reference meshes. Finally, following Zhou et al. [45] we quantify symmetry plane accuracy by comparing the cosine distance between symmetry plane normals. + +# 4.1.Quantitative Evaluation + +3D Shape Completion. We compare the performance of ISCNet to 8 competing methods subdivided in three categories: Point-cloud completion methods (PCN [41], MSN [15], PoinTr [40], VRCNet [16], ShowFlakeNet [37], VE-PCN [32]), implicit-function-based methods (IFNet [3]), and image-based completion methods (2 variants of Front2Back $[39]^{1}$ ). All competing methods are retrained on ShapeNet and tested with identical exemplars under the same conditions. The quantitative comparisons (using the CD and MD metrics) are summarized in Table 1 and Table 2 for ShapeNet and ScanNet, respectively. From these, we can see that ISCNet offers the best overall performance for the majority of categories on both synthetic and scanned data, followed by Front2Back [39]. SnowFlakeNet [37] performs better than ISCNet on the plane shapes with a marginal accuracy gain. Compared to Front2Back, we can see that ISCNet outperforms Front2Back with the ground truth symmetry information. For completeness, Table 3 compares ISCNet with Front2Back and IFNet for the remaining five classes from ShapeNet, without retraining or finetuning. Our approach still outperforms existing work except for the rifle class compared to IFNet. + +
(×102)BenchDisplaySpeakerRiflePhoneAverage (unseen-5)Average (total-13)
CDF2B2.373.193.401.723.132.762.24
IFNet1.813.736.351.373.343.323.02
Ours1.442.473.091.432.102.111.72
MDF2B2.192.973.171.592.972.582.11
IFNet1.773.656.241.353.283.263.01
Ours1.412.453.011.422.072.071.61
+ +Table 3. Quantitative evaluation of ISCNet compared to Front2Back and IFNet for the remaining 5 classes of ShapeNet, i.e., unseen-5 labels. + +Figure 4 and 5 provide additional qualitative comparisons. Visually, ISCNet provides the 'cleanest' shape completion for all eight tasks. In particular, in the plane completion task, ISCNet is better able to retain high-frequency surface details, and in the cabinet and lamp completion tasks + +ISCNet is the only method to fully complete the surface without holes. In the case of the chair and table completion tasks, ISCNet more robustly identifies areas to be completed without erroneously merging thin features. Although our method is designed for symmetric objects, we also obtain good reconstruction quality on non-symmetric objects (sofa and table in Figure 5). + +Symmetry Detection. To assess the accuracy of the symmetry plane estimation, we compare it against two state-of-the-art single-view based pose estimation and symmetry detection methods, RotationContinuity [44] and NeRD [45]. Figure 6a shows that ISCNet's second update of the symmetry plane offers a similar performance as NeRD. Furthermore, with increasing iterations, the accuracy of the estimated symmetry plane improves (i.e., compared to the first symmetry plane estimate2, the occurrence of symmetry plane estimates with an error less than $5^{\circ}$ , improves from $32.86\%$ to $56.96\%$ ), demonstrating the effectiveness of the iterative refinement for symmetry plane estimation. + +# 4.2. Ablation Study + +Importance of Symmetry. We validate the effectiveness of symmetry cues for shape completion in two ablation experiments. As a first experiment (Table 1, Unaligned ISCNet), we replace the estimated transformation that aligns the symmetry plane with the X-Y plane by a random rotation matrix and a perturbed center (adding normal distributed noise with $\sigma = 0.1$ ). Consequently, ISCNet cannot leverage any symmetry cues. In a second experiment (Table 1, No-Pair ISCNet), we process each of the reflection viewpoints separately. From Table 1, we can see that the unaligned ISCNet suffers a significant performance degradation; CD increases by 0.56 on average, and MD increases 0.55 on average. By aligning the shape to the symmetry plane, no-pair ISCNet is able to leverage symmetry implicitly (by learning to copy from the negative Z coordinate). However, compared to the full ISCNet, the no-pair ISCNet still sees a performance drop (CD and MD are increased by 0.34 and 0.28, respectively). We refer to the supplementary material for a qualitative comparison of excluding explicit leveraging of the symmetry cues. + +Role of Normal Inpainting. To demonstrate the importance of performing normal inpainting, we eliminate the inpainted normal as input from the depth inpainting submodule (Table 1, No-Normal ISCNet). Without normal cues, CD increases by 0.46 and MD by 0.36. We refer to the supplementary material for a qualitative visualization of the impact of omitting the normal inpainting step. + +Impact of Iterative Refinement. To demonstrate that both symmetry plane estimation as well as shape completion improve with additional iterations, we compare the MD and CD errors per iteration (Table 1) as well as the symmetry plane normal error progression per iteration (Figure 6b). In both cases, we can see that the error improves per iteration and the improvement diminishes per subsequent iteration (justifying a 20 iteration stopping criterion). + +![](images/ef08d0aeb46556f227056aa264db3866194e0fee906be0935fc8ec6080050fdc.jpg) +(a) + +![](images/4919d2a3ea659882d9fd192d0e8746bdae4da6fb8d1a42af67424f6da78ca44b.jpg) +(b) +Figure 6. Impact of the number of iterations on (a) the accuracy of the symmetry plane estimation, and (b) the shape reconstruction CD/MD accuracy and average and median symmetry plane estimation errors. 'Our SP $n$ ' refers to ISCNet's $n$ -th predicted symmetry plane. + +Limitations The main bottleneck in ISCNet is the image rendering and the 3D point cloud processing step in each iteration. Furthermore, ISCNet is only able to leverage a single global symmetry plane, and thus its performance is suboptimal when multiple symmetries are present. + +# 5. Conclusion + +We present ISCNet, a single-depth image joint symmetry plane estimation and shape completion method that explicitly leverages symmetry cues. ISCNet alternates between estimating the symmetry plane and completing the shape to enhance the accuracy of both components. Our method is informed by two key insights: (1) shape completion and symmetry plane estimation are mutually beneficial tasks, and (2) by selecting symmetric virtual viewpoints, we can explicitly inform the inpainting module on symmetry relations. Our experiments and ablation studies demonstrate that ISCNet can yield more accurate shape completions from a single-depth image than competing methods. + +# Acknowledgments + +We thank the anonymous reviewers for the insightful and constructive comments. This work was supported in part by National Key Research and Development Program of China (2022ZD0210500, 2021ZD0111902), National Natural Science Foundation of China (61972067, U19B2039, U21B2038), the Distinguished Young Scholars Funding of Dalian (No. 2022RJ01). Pieter Peers was supported in part + +by NSF grant IIS-1909028. Felix Heide was supported by an NSF CAREER Award (2047359), a Packard Foundation Fellowship, a Sloan Research Fellowship, a Sony Young Faculty Award, a Project X Innovation Award, and an Amazon Science Research Award. + +# References + +[1] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 6 +[2] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6 +[3] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6970-6981, 2020. 1, 6, 8 +[4] Paolo Cignoni, Claudio Rocchini, and Roberto Scopigno. Metro: measuring error on simplified surfaces. In Computer Graphics Forum, volume 17, pages 167-174, 1998. 8 +[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6 +[6] Shivam Duggal and Deepak Pathak. Topologically-aware deformation fields for single-view 3d reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1536-1546, 2022. 2 +[7] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 605-613, 2017. 2 +[8] Boqing Gong, Jianzhuang Liu, Xiaogang Wang, and Xiaou Tang. Learning semantic signatures for 3d object retrieval. IEEE Trans. Multimedia, 15(2):369-377, 2012. 1, 2 +[9] Xiaoguang Han, Zhaoxuan Zhang, Dong Du, Mingdai Yang, Jingming Yu, Pan Pan, Xin Yang, Ligang Liu, Zixiang Xiong, and Shuguang Cui. Deep reinforcement learning of volume-guided progressive view inpainting for 3d point scene completion from a single depth image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 234-243, 2019. 5 +[10] Sagar Imambi, Kolla Bhanu Prakash, and GR Kanagachidambaresan. Pytorch. In Programming with TensorFlow, pages 87-104. 2021. 6 +[11] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006. 5 +[12] Michael Kazhdan, Thomas Funkhouser, and Szymon Rusinkiewicz. Symmetry descriptors and 3d shape matching. In Proceedings of the 2004 Eurographics/ACM SIGGRAPH + +symposium on Geometry processing, pages 115-123, 2004. 1, 2 +[13] Wai Ho Li and Lindsay Kleeman. Segmentation and modeling of visually symmetric objects by robot actions. The International Journal of Robotics Research, 30(9):1124-1142, 2011. 1, 2 +[14] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Eur. Conf. Comput. Vis., pages 85-100, 2018. 5 +[15] Minghua Liu, Lu Sheng, Sheng Yang, Jing Shao, and ShiMin Hu. Morphing and sampling network for dense point cloud completion. In AAAI, volume 34, pages 11596-11603, 2020. 2, 4, 6, 8 +[16] Liang Pan, Xinyi Chen, Zhongang Cai, Junzhe Zhang, Haiyu Zhao, Shuai Yi, and Ziwei Liu. Variational relational point completion network. In Int. Conf. Comput. Vis., pages 8524-8533, 2021. 2, 5, 6, 8 +[17] Joshua Podolak, Philip Shilane, Aleksey Golovinskiy, Szymon Rusinkiewicz, and Thomas Funkhouser. A planar-reflective symmetry transform for 3d shapes. In ACM SIGGRAPH 2006 Papers, pages 549-559. 2006. 1 +[18] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 652-660, 2017. 4 +[19] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241, 2015. 4 +[20] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3, 5 +[21] Yifei Shi, Junwen Huang, Hongjia Zhang, Xin Xu, Szymon Rusinkiewicz, and Kai Xu. Symmetrynet: learning to predict reflectional and rotational symmetries of 3d shapes from single-view rgb-d images. ACM Trans. Graph., 39(6):1-14, 2020. 1, 2 +[22] Yifei Shi, Xin Xu, Junhua Xi, Xiaochang Hu, Dewen Hu, and Kai Xu. Learning to detect 3d symmetry from single-view rgb-d images with weak supervision. IEEE Trans. Pattern Anal. Mach. Intell., 2022. 1, 2 +[23] Ivan Sipiran, Robert Gregor, and Tobias Schreck. Approximate symmetry detection in partial 3d meshes. In Computer Graphics Forum, volume 33, pages 131-140. Wiley Online Library, 2014. 1, 2 +[24] Pablo Speciale, Martin R Oswald, Andrea Cohen, and Marc Pollefeys. A symmetry prior for convex variational 3d reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 313-328. Springer, 2016. 1, 2 +[25] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B Tenenbaum, and William T Freeman. Pix3d: Dataset and methods for single-image 3d shape modeling. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2974-2983, 2018. 2 + +[26] Junshu Tang, Zhijun Gong, Ran Yi, Yuan Xie, and Lizhuang Ma. Lake-net: Topology-aware point cloud completion by localizing aligned keypoints. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1726-1735, 2022. 2 +[27] Demetri Terzopoulos, Andrew Witkin, and Michael Kass. Symmetry-seeking models and 3d object reconstruction. Int. J. Comput. Vis., 1(3):211-221, 1988. 2 +[28] Sebastian Thrun and Ben Wegbreit. Shape from symmetry. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 2, pages 1824-1831. IEEE, 2005. 1, 2 +[29] Shubham Tulsiani, Tinghui Zhou, Alexei A Efros, and Jitendra Malik. Multi-view supervision for single-view reconstruction via differentiable ray consistency. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2626-2634, 2017. 2 +[30] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martin-Martin, Cewu Lu, Li Fei-Fei, and Silvio Savarese. Densefusion: 6d object pose estimation by iterative dense fusion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3343–3352, 2019. 1, 2, 3, 8 +[31] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In *Eur. Conf. Comput. Vis.*, pages 52-67, 2018. 2 +[32] Xiaogang Wang, Marcelo H Ang, and Gim Hee Lee. Voxel-based network for shape completion by leveraging edge generation. In Int. Conf. Comput. Vis., pages 13189-13198, 2021. 6, 8 +[33] Xiaogang Wang, Marcelo H Ang Jr, and Gim Hee Lee. Cascaded refinement network for point cloud completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 790-799, 2020. 1 +[34] Boyan Wei, Xianfeng Ye, Chengjiang Long, Zhenjun Du, Bangyu Li, Baocai Yin, and Xin Yang. Discriminative active learning for robotic grasping in cluttered scene. IEEE Robotics and Automation Letters, 8(3):1858-1865, 2023. 1 +[35] Jiajun Wu, Yifan Wang, Tianfan Xue, Xingyuan Sun, Bill Freeman, and Josh Tenenbaum. Marrnet: 3d shape reconstruction via 2.5 d sketches. Adv. Neural Inform. Process. Syst., 2017. 2 +[36] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-10, 2020. 2 +[37] Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. Snowflakenet: Point cloud completion by snowflake point deconvolution with skip-transformer. In Int. Conf. Comput. Vis., pages 5499-5509, 2021. 2, 6, 8 +[38] Xin Yang, Yuanbo Wang, Yaru Wang, Baocai Yin, Qiang Zhang, Xiaopeng Wei, and Hongbo Fu. Active object reconstruction using a guided view planner. In Proceedings of International Joint Conference on Artificial Intelligence, pages 4965-4971, 2018. 3 +[39] Yuan Yao, Nico Schertler, Enrique Rosales, Helge Rhodin, Leonid Sigal, and Alla Sheffer. Front2back: Single view 3d + +shape reconstruction via front to back prediction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 531-540, 2020. 1, 2, 6, 8 +[40] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In Int. Conf. Comput. Vis., pages 12498-12507, 2021. 1, 2, 6, 8 +[41] Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, and Martial Hebert. Pcn: Point completion network. In Int. Conf. 3D Vis., pages 728-737, 2018. 2, 6, 8 +[42] Zhaoxuan Zhang, Xiaoguang Han, Bo Dong, Tong Li, Baocai Yin, and Xin Yang. Point cloud scene completion with joint color and semantic estimation from single rgb-d image. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3 +[43] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3d: A modern library for 3d data processing. arXiv preprint arXiv:1801.09847, 2018. 8 +[44] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5745-5753, 2019. 2, 8 +[45] Yichao Zhou, Shichen Liu, and Yi Ma. Nerd: neural 3d reflection symmetry detector. In IEEE Conf. Comput. Vis. Pattern Recog., pages 15940-15949, 2021. 1, 2, 8 \ No newline at end of file diff --git a/2023/Single Depth-image 3D Reflection Symmetry and Shape Prediction/images.zip b/2023/Single Depth-image 3D Reflection Symmetry and Shape Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1ace22c87601ad5ad44af8fe1fb9ae8e123b323a --- /dev/null +++ b/2023/Single Depth-image 3D Reflection Symmetry and Shape Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1eaacb3e49b87b44cf6a0887a26c1a78621ef9e2360a77f4e6bdcb15e3c94f23 +size 700271 diff --git a/2023/Single Depth-image 3D Reflection Symmetry and Shape Prediction/layout.json b/2023/Single Depth-image 3D Reflection Symmetry and Shape Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f2bc863fef5ea32507a9583290f5d0a9a9e7a583 --- /dev/null +++ b/2023/Single Depth-image 3D Reflection Symmetry and Shape Prediction/layout.json @@ -0,0 +1,9811 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 103, + 503, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 103, + 503, + 121 + ], + "spans": [ + { + "bbox": [ + 91, + 103, + 503, + 121 + ], + "type": "text", + "content": "Single Depth-image 3D Reflection Symmetry and Shape Prediction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 84, + 143, + 512, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 143, + 512, + 159 + ], + "spans": [ + { + "bbox": [ + 84, + 143, + 512, + 159 + ], + "type": "text", + "content": "Zhaoxuan Zhang, Bo Dong*, Tong Li, Felix Heide, Pieter Peers, Baocai Yin*, Xin Yang*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 185, + 192, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 185, + 192, + 198 + ], + "spans": [ + { + "bbox": [ + 143, + 185, + 192, + 198 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 210, + 290, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 210, + 290, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 290, + 413 + ], + "type": "text", + "content": "In this paper, we present Iterative Symmetry Completion Network (ISCNet), a single depth-image shape completion method that exploits reflective symmetry cues to obtain more detailed shapes. The efficacy of single depth-image shape completion methods is often sensitive to the accuracy of the symmetry plane. ISCNet therefore jointly estimates the symmetry plane and shape completion iteratively; more complete shapes contribute to more robust symmetry plane estimates and vice versa. Furthermore, our shape completion method operates in the image domain, enabling more efficient high-resolution, detailed geometry reconstruction. We perform the shape completion from pairs of viewpoints, reflected across the symmetry plane, predicted by a reinforcement learning agent to improve robustness and to simultaneously explicitly leverage symmetry. We demonstrate the effectiveness of ISCNet on a variety of object categories on both synthetic and real-scanned datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 436, + 128, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 128, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 128, + 449 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 456, + 287, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 516 + ], + "type": "text", + "content": "Symmetry is an intrinsic geometric property present in the vast majority of man-made and natural objects [12, 17], and which has been exploited to improve the efficacy of applications such as shape matching [12], segmentation [13], object retrieval [8], and robotic grasping [34]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 517, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 288, + 696 + ], + "type": "text", + "content": "In this paper, we aim to leverage symmetry as an aid for the completion of shapes from partial observations [23, 24, 28]. Recently, Yao et al. [39] presented Front2Back, a framework for 3D shape reconstruction from a single RGB image. A key observation in Front2Back is that a (nearly) complete 3D model can be effectively described by a pair of 2.5D visible surface images taken from opposite views (i.e., front and back views). Based on this observation, Front2Back completes a shape by synthesizing a back view using a global 3D reflective symmetry plane from a given front view. However, the reconstruction quality is significantly affected by the input front view angle (Figure 1(d)). Moreover, Front2Back assumes the availability of a 'perfect' symmetry plane, which in practice is challenging to obtain from partial observations [21, 22, 45]." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 185, + 386, + 247 + ], + "blocks": [ + { + "bbox": [ + 317, + 185, + 386, + 247 + ], + "lines": [ + { + "bbox": [ + 317, + 185, + 386, + 247 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 386, + 247 + ], + "type": "image", + "image_path": "9f65dacff26d8142651f49b7c010a15e32a60362be274588248a08729da8b788.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 249, + 402, + 270 + ], + "lines": [ + { + "bbox": [ + 317, + 249, + 402, + 270 + ], + "spans": [ + { + "bbox": [ + 317, + 249, + 402, + 270 + ], + "type": "text", + "content": "(a) Input Depth Map & Observed Surface" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 425, + 182, + 468, + 246 + ], + "blocks": [ + { + "bbox": [ + 425, + 182, + 468, + 246 + ], + "lines": [ + { + "bbox": [ + 425, + 182, + 468, + 246 + ], + "spans": [ + { + "bbox": [ + 425, + 182, + 468, + 246 + ], + "type": "image", + "image_path": "62b67ac4a0495c76f7fcfbc44cd1dee505e2450af71025e1dfd9e61a6cc15eff.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 254, + 463, + 266 + ], + "lines": [ + { + "bbox": [ + 406, + 254, + 463, + 266 + ], + "spans": [ + { + "bbox": [ + 406, + 254, + 463, + 266 + ], + "type": "text", + "content": "(b) PoinTr [40]" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 493, + 184, + 533, + 247 + ], + "blocks": [ + { + "bbox": [ + 493, + 184, + 533, + 247 + ], + "lines": [ + { + "bbox": [ + 493, + 184, + 533, + 247 + ], + "spans": [ + { + "bbox": [ + 493, + 184, + 533, + 247 + ], + "type": "image", + "image_path": "45e3b261f8434c20c6712b131cc05a4f62a92430b17c52552817f01a327efc1e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 255, + 527, + 266 + ], + "lines": [ + { + "bbox": [ + 476, + 255, + 527, + 266 + ], + "spans": [ + { + "bbox": [ + 476, + 255, + 527, + 266 + ], + "type": "text", + "content": "(c) IF-Net [3]" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 349, + 274, + 392, + 335 + ], + "blocks": [ + { + "bbox": [ + 349, + 274, + 392, + 335 + ], + "lines": [ + { + "bbox": [ + 349, + 274, + 392, + 335 + ], + "spans": [ + { + "bbox": [ + 349, + 274, + 392, + 335 + ], + "type": "image", + "image_path": "ac5b3c58597abcef32d4ecf2fb59d32e7da9043fdd5ea829b28a3d5eebaebe0e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 338, + 397, + 350 + ], + "lines": [ + { + "bbox": [ + 322, + 338, + 397, + 350 + ], + "spans": [ + { + "bbox": [ + 322, + 338, + 397, + 350 + ], + "type": "text", + "content": "(d) Front2Back [39]" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 421, + 274, + 464, + 335 + ], + "blocks": [ + { + "bbox": [ + 421, + 274, + 464, + 335 + ], + "lines": [ + { + "bbox": [ + 421, + 274, + 464, + 335 + ], + "spans": [ + { + "bbox": [ + 421, + 274, + 464, + 335 + ], + "type": "image", + "image_path": "e8e205a5f71b82ebf239db0d51f22344d918beed4737b561c41bb74594d22542.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 338, + 450, + 349 + ], + "lines": [ + { + "bbox": [ + 419, + 338, + 450, + 349 + ], + "spans": [ + { + "bbox": [ + 419, + 338, + 450, + 349 + ], + "type": "text", + "content": "(e) Ours" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 493, + 274, + 533, + 335 + ], + "blocks": [ + { + "bbox": [ + 493, + 274, + 533, + 335 + ], + "lines": [ + { + "bbox": [ + 493, + 274, + 533, + 335 + ], + "spans": [ + { + "bbox": [ + 493, + 274, + 533, + 335 + ], + "type": "image", + "image_path": "948cade7807be767324c0d2e9ab6b6c72db87805a993e9096ca22fd00a671f82.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 338, + 533, + 349 + ], + "lines": [ + { + "bbox": [ + 470, + 338, + 533, + 349 + ], + "spans": [ + { + "bbox": [ + 470, + 338, + 533, + 349 + ], + "type": "text", + "content": "(f) Ground Truth" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 350, + 547, + 384 + ], + "lines": [ + { + "bbox": [ + 305, + 350, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 350, + 547, + 384 + ], + "type": "text", + "content": "Figure 1. ISCNet more efficiently leverages symmetry cues for robust shape completion from a single depth image compared to prior methods." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 388, + 546, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 546, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 546, + 640 + ], + "type": "text", + "content": "We present ISCNet (Iterative Shape Completion Network), a single depth-image shape completion method that explicitly leverages symmetry cues. Our key insight is that symmetry detection and symmetry-based shape completion are complementary tasks that can provide constructive cues to the other task. Symmetry detection can help shape completion by providing object-centric information about missing geometrical features, and shape completion can provide more robust symmetry cues by filling in missing geometry. Departing from existing work [33] that assumes an ideal prediction of the symmetry plane, ISCNet iteratively refines the symmetry plane and the reconstructed shape jointly. Furthermore, unlike Front2Back [39], ISCNet does not rely on the initial input viewpoint to a-priori fix the synthetic back view camera position but instead leverages a trained reinforcement learning (RL) policy to select optimal pairs of reflection viewpoints at each iteration. We empirically found that predetermining the viewpoints is less effective than RL-based viewpoint selection which can more effectively deal with the dynamic shape of a partial point cloud during the refinement process." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "Specifically, ISCNet takes as input a single depth image and reprojects it to an incomplete point cloud. Instead of explicitly detecting a symmetry plane, we estimate the pose [30] that aligns the symmetry plane to the X-Y plane. Based on the current estimate of the point cloud and symmetry aligning pose, an RL agent determines a pair of " + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "inline_equation", + "content": "re" + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "-" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 149, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 149, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 149, + 712 + ], + "type": "text", + "content": "*Co-corresponding authors." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8896" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": "flection viewpoints, reflected across the symmetry plane, for extracting a pair of partial depth and normal maps that are subsequently repaired using a novel 2D inpainting module. By directly working on depth image pairs, our method is independent of the size of the point cloud, and it can leverage geometry information from the symmetric counter view. However, because the virtual depth images are reprojections that differ from the input view, not all reprojected points are valid (e.g., the backside of points visible through a 'hole'). Therefore, we first identify reliable pixels before performing depth inpainting in the projected image domain. Finally, we reproject the inpainted maps back to 3D and update the point cloud. We progressively refine the point cloud by alternating between estimating reflection viewpoints and completing the point cloud for 20 iterations, updating the symmetry plane every 4th iteration, to balance efficiency versus efficacy. When updating the symmetry plane, we 'reset' the point cloud to its initial state to mitigate the impact of prior inaccurate symmetry plane estimations on the shape completion." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 312, + 289, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 289, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 289, + 431 + ], + "type": "text", + "content": "We validate that ISCNet produces more accurate shape completions than existing methods (Figure 1) and confirm the contribution of each component in our model via a thorough ablation study. In summary, our contributions are: 1) a novel robust joint symmetry-detection / shape completion method; 2) a symmetric viewpoint-based 3D shape completion method for detailed high-resolution shape reconstruction; and 3) a reinforcement learning-based optimal reflection viewpoint selection solution for progressive shape completion." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 442, + 134, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 442, + 134, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 442, + 134, + 454 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 462, + 289, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 289, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 289, + 677 + ], + "type": "text", + "content": "Symmetry Detection on 3D Shapes forms a key component in many downstream applications ranging from shape matching [12], to segmentation [13], to 3D object retrieval [8], to shape completion [23, 24, 27, 28] and to deformable object learning [36]. Symmetry detection of incomplete shape observations is significantly more challenging, and recent learning-based solutions have explored the use of multitask learning [21], sampling the distribution of plausible shapes [22], and building 3D cost volumes [45] to predict 3D symmetry planes. An alternative approach to predicting arbitrarily oriented 3D symmetry planes is to estimate a camera pose that aligns the local coordinate frame such that the X-Y plane acts as a canonical symmetry plane [30, 44]. Our symmetry detection module falls in this latter category. However, unlike prior work that a priori fixes the symmetry information as a prior, we jointly perform shape completion yielding a more robust and mutually reinforcing symmetry detection and shape completion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "Learning-based 3D Shape Completion of Partial Point Cloud Observations have explored coarse-to-fine methods [15, 41], transformer-networks [37, 40], probabilis-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "tic modeling [16], and topology-aware approaches [26] to complete the 3D point cloud. However, these methods are memory intensive and their scalability (e.g., beyond " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\approx 16384" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " points) is limited. In contrast, our approach does not impose such limitation on the number of 3D points as our method leverages 2D depth inpainting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 148, + 546, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 148, + 546, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 546, + 293 + ], + "type": "text", + "content": "Single-image Shape Reconstruction aims to complete approximately half of the shape either directly from a 2D image [6, 7, 29, 31] or (indirectly) from a 2.5D input [25, 35, 39]. Front2Back [39] estimates, besides depth and surface normals, also symmetry information to aid in the surface completion task. However, Front2Back is limited to a single front-back camera pair. Our method leverages an RL agent to predict multiple virtual reflection viewpoints for shape completion, and it takes better advantage of the symmetry prior by inpainting symmetric image pairs. In addition, we iteratively refine the symmetry plane to obtain more accurate estimates of both shape and symmetry." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 318, + 533, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 318, + 533, + 332 + ], + "spans": [ + { + "bbox": [ + 305, + 318, + 533, + 332 + ], + "type": "text", + "content": "3. Iterative-Symmetry Completion Network" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 342, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 712 + ], + "type": "text", + "content": "Our Iterative Symmetry Completion Network (ISCNet) takes a single-view depth map of a target object as input and iteratively completes the 3D mesh by alternating between refining the symmetry plane, the virtual reflection viewpoints, and the shape completion. To bootstrap the iterative process, we first project the input single-view depth map to an incomplete point cloud that is subsequently passed to a symmetry detection module and point cloud transposition operator. The symmetry detection module (subsection 3.1) provides an estimate of the object's symmetry plane that is subsequently leveraged by the point cloud transposition operator to align the point cloud symmetrically around the estimated symmetry plane. Next, a novel 3D shape completion module (subsection 3.2) progressively completes the 3D point cloud based on a pair of reflection viewpoints selected by an RL agent. For each reflection viewpoint pair, a depth and normal map pair is extracted from the partial point cloud. Due to the partial completion of the point cloud, the extracted depth and normal map pairs exhibit holes. To address this issue, we employ a specifically crafted inpainting module, which not only repairs these holes but also assigns a confidence score to each repaired pixel. Next, the inpainted depth and normal maps are merged into the point cloud guided by their associated confidence scores. Finally, to improve robustness, we iteratively refine the symmetry plane, the reflection viewpoints, and the shape completion (subsection 3.3). We limit the number of iterations to 20 and only update the symmetry plane and reset the point cloud every 4th iteration to balance robustness versus efficiency; Figure 2 visually summarizes ISCNet." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8897" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 68, + 71, + 521, + 208 + ], + "blocks": [ + { + "bbox": [ + 68, + 71, + 521, + 208 + ], + "lines": [ + { + "bbox": [ + 68, + 71, + 521, + 208 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 521, + 208 + ], + "type": "image", + "image_path": "a275c2232977bca901e91c8827aaa14f65d2274e8c0f0153c50909992d5ca5c6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 212, + 547, + 279 + ], + "lines": [ + { + "bbox": [ + 46, + 212, + 547, + 279 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 547, + 279 + ], + "type": "text", + "content": "Figure 2. Summary of ISCNet: a partial shape extracted from a single depth-image is aligned such that the X-Y plane corresponds to the (detected) symmetry plane. Next, a pair of depth and normal maps are rendered from reflection viewpoints selected by an RL agent, and each symmetric pair of synthetic depth and normal maps is inpainted in the image-domain (see also Figure 3). These inpainted maps are then back-projected to a point cloud and merged. The symmetry plane is updated every 4th iteration, where the merged shape is then visualized from 3 selected views, " + }, + { + "bbox": [ + 46, + 212, + 547, + 279 + ], + "type": "inline_equation", + "content": "v_{8}^{V}" + }, + { + "bbox": [ + 46, + 212, + 547, + 279 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 212, + 547, + 279 + ], + "type": "inline_equation", + "content": "v_{12}^{V}" + }, + { + "bbox": [ + 46, + 212, + 547, + 279 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 212, + 547, + 279 + ], + "type": "inline_equation", + "content": "v_{13}^{V}" + }, + { + "bbox": [ + 46, + 212, + 547, + 279 + ], + "type": "text", + "content": " in this case, passing into the symmetry detection module separately, and the best-detected symmetry plane (with the highest confidence score) is used as input for the next iteration." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 297, + 231, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 231, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 231, + 309 + ], + "type": "text", + "content": "3.1. 3D Reflection Symmetry Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 290, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 290, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 290, + 473 + ], + "type": "text", + "content": "We estimate the symmetry plane of an object via a proxy task, namely, camera pose estimation, such that the symmetry plane aligns in local coordinates with the X-Y plane. Our Symmetry Detection Module follows the architecture of DenseFusion [30]. We pass a depth image to obtain a point cloud, pixel-wise features, and finally a pose estimation. Different from DenseFusion, we only use depth instead of RGB channels, and we do not require an object segmentation stage. Instead of predicting the symmetry parameters directly, we found it more robust to estimate a rotation matrix " + }, + { + "bbox": [ + 46, + 316, + 290, + 473 + ], + "type": "inline_equation", + "content": "R \\in SO(3)" + }, + { + "bbox": [ + 46, + 316, + 290, + 473 + ], + "type": "text", + "content": " and a translation vector " + }, + { + "bbox": [ + 46, + 316, + 290, + 473 + ], + "type": "inline_equation", + "content": "t \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 316, + 290, + 473 + ], + "type": "text", + "content": " such that the symmetry plane's center point " + }, + { + "bbox": [ + 46, + 316, + 290, + 473 + ], + "type": "inline_equation", + "content": "p_s" + }, + { + "bbox": [ + 46, + 316, + 290, + 473 + ], + "type": "text", + "content": " and normal direction " + }, + { + "bbox": [ + 46, + 316, + 290, + 473 + ], + "type": "inline_equation", + "content": "n_s" + }, + { + "bbox": [ + 46, + 316, + 290, + 473 + ], + "type": "text", + "content": " are defined by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 480, + 287, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 480, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 88, + 480, + 287, + 509 + ], + "type": "interline_equation", + "content": "\\left( \\begin{array}{c} p _ {s} \\\\ n _ {s} \\end{array} \\right) = \\left( \\begin{array}{c} R ^ {- 1} \\cdot ([ 0, 0, 0 ] - t) \\\\ R ^ {- 1} \\cdot ([ 0, 0, 1 ] - t) \\end{array} \\right). \\tag {1}", + "image_path": "55673eeeb824b86543422ae580d9565ef45d13b1fa1454ad51aa1a8aac9081d7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 514, + 173, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 173, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 173, + 528 + ], + "type": "text", + "content": "3.2. 3D Shape Completion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 582 + ], + "type": "text", + "content": "Given a predicted symmetry plane, ISCNet progressively completes the 3D shape. Our shape completion module consists of three sequential sub-modules: viewpoint selection, 2D inpainting, and point cloud updating." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 688 + ], + "type": "text", + "content": "Viewpoint Selection Module: ISCNet completes 3D shapes via 2D projected depth and normal maps. The exact choice of virtual viewpoints from which these depth and normals are generated is critical for robust shape completion, and it is sensitive to the irregular structures present in the partially completed point cloud. As the shape of a partial point cloud dynamically changes, it is impractical to pre-determine viewpoints; see our supplementary material for more details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "We model the problem of selecting a viewpoint " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " as a Markov decision process (MDP) because, similar to [38," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 297, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 545, + 382 + ], + "type": "text", + "content": "42], we are only concerned about the current state of a partial point cloud. To this end, we leverage an RL agent to select an optimal viewpoint trained with an Actor-Critic Proximal Policy Optimisation (PPO) algorithm [20]. The RL agent predicts an optimal viewpoint (i.e., action) based on the current 3D point cloud (i.e., state) to maximize an appropriately designed reward." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 402, + 546, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 402, + 546, + 438 + ], + "spans": [ + { + "bbox": [ + 305, + 402, + 546, + 438 + ], + "type": "text", + "content": "Specifically, the continuous action space of the RL agent is defined by " + }, + { + "bbox": [ + 305, + 402, + 546, + 438 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 402, + 546, + 438 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 402, + 546, + 438 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 305, + 402, + 546, + 438 + ], + "type": "text", + "content": " angles (Figure 2), with corresponding viewpoint:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 389, + 480, + 545, + 492 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 480, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 389, + 480, + 545, + 492 + ], + "type": "interline_equation", + "content": "x = a \\sin \\theta \\sin \\phi , \\tag {2}", + "image_path": "ef197c46f371f734d8cd53bfa109fc70f18b523cb70c1ab3f70580b98c241dba.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 390, + 495, + 545, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 495, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 390, + 495, + 545, + 507 + ], + "type": "interline_equation", + "content": "y = a \\cos \\theta , \\tag {3}", + "image_path": "60d626bfb0b918cb91401fbe810807335220495112a3877741cda525bc10ef53.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 390, + 510, + 545, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 510, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 390, + 510, + 545, + 522 + ], + "type": "interline_equation", + "content": "z = a \\sin \\theta \\cos \\phi , \\tag {4}", + "image_path": "41665f924aa70e02b7f82a00e348b89c65fedd96cd72239ff0d61fa29b008814.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 573, + 545, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 573, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 305, + 573, + 545, + 608 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 573, + 545, + 608 + ], + "type": "inline_equation", + "content": "\\theta \\in [0^{\\circ},180^{\\circ}]" + }, + { + "bbox": [ + 305, + 573, + 545, + 608 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 573, + 545, + 608 + ], + "type": "inline_equation", + "content": "\\phi \\in [-180^{\\circ},180^{\\circ}]" + }, + { + "bbox": [ + 305, + 573, + 545, + 608 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 305, + 573, + 545, + 608 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 305, + 573, + 545, + 608 + ], + "type": "text", + "content": " is set at " + }, + { + "bbox": [ + 305, + 573, + 545, + 608 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{m}" + }, + { + "bbox": [ + 305, + 573, + 545, + 608 + ], + "type": "text", + "content": ". The view direction is always aimed at the object center." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "text", + "content": "The prediction action reward is determined by four factors: accuracy " + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "inline_equation", + "content": "R^a" + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "text", + "content": ", shape completeness " + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "inline_equation", + "content": "R^s" + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "text", + "content": ", number of recovered effective points " + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "inline_equation", + "content": "R^p" + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "text", + "content": ", and viewpoint diversity " + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "inline_equation", + "content": "R^v" + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "inline_equation", + "content": "R^a" + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "inline_equation", + "content": "R^s" + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "text", + "content": " encourage an agent to choose a new viewpoint that benefits the prediction of missing parts from the original point cloud. " + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "inline_equation", + "content": "R^p" + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "inline_equation", + "content": "R^v" + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "text", + "content": " encourage the agent to increase the viewpoint diversity. Formally, the reward function for" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8898" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 532, + 279 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 532, + 279 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 532, + 279 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 532, + 279 + ], + "type": "image", + "image_path": "ce7e296968062e1675be6315403d26c904f2c9c3594821dd76f30e2f3d4c64f2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 283, + 465, + 295 + ], + "lines": [ + { + "bbox": [ + 126, + 283, + 465, + 295 + ], + "spans": [ + { + "bbox": [ + 126, + 283, + 465, + 295 + ], + "type": "text", + "content": "Figure 3. Detailed visual summary of the viewpoint selection and the 2D inpainting modules." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 300, + 141, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 141, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 141, + 312 + ], + "type": "text", + "content": "an action is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 319, + 287, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 319, + 287, + 333 + ], + "spans": [ + { + "bbox": [ + 89, + 319, + 287, + 333 + ], + "type": "interline_equation", + "content": "R = w _ {1} R ^ {a} + w _ {2} R ^ {s} + w _ {3} R ^ {p} + w _ {4} R ^ {v}, \\tag {5}", + "image_path": "d0ea0987297bebee7faf610b1c2c2322830b0149441c7fc007c6f4365d3b8c7b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 335, + 287, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 335, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 85, + 335, + 287, + 365 + ], + "type": "interline_equation", + "content": "R ^ {a} = \\frac {1}{\\mathcal {N} (\\hat {P})} \\sum_ {x \\in \\hat {P}} \\min _ {y \\in P _ {G T}} \\| x - y \\| _ {2} ^ {2}, \\tag {6}", + "image_path": "6c3b770b4a3e4b52444792a8551cfb06d863dd82cc26d5bbed81129fe3751072.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 368, + 287, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 368, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 86, + 368, + 287, + 381 + ], + "type": "interline_equation", + "content": "R ^ {s} = S - S ^ {\\prime}, \\tag {7}", + "image_path": "3e47ab0e9317efb7e318eb4ec4974565bf91a99726799d9420179e1253674547.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 384, + 287, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 384, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 86, + 384, + 287, + 398 + ], + "type": "interline_equation", + "content": "R ^ {p} = \\mathcal {N} \\left(\\left\\{d (x, P ^ {\\prime}) < 0. 0 0 2 | x \\in \\hat {P} \\right\\}\\right), \\tag {8}", + "image_path": "62d6ea26d391eb9787c93b1ee40b01bc77d00ffa3fcf37763ff533d77385b2e2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 86, + 399, + 287, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 399, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 86, + 399, + 287, + 413 + ], + "type": "interline_equation", + "content": "R ^ {v} = \\left(\\left| \\theta - \\theta^ {\\prime} \\right| + \\left| \\phi - \\phi^ {\\prime} \\right|\\right) / 1 8 0 \\cdot \\pi , \\tag {9}", + "image_path": "13ad2ec8058d09f81fb7407aa3e1ba892951cc1e6851a14e914652a983be3c17.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 91, + 415, + 287, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 415, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 91, + 415, + 287, + 441 + ], + "type": "interline_equation", + "content": "S = \\frac {\\mathcal {N} \\left(\\left\\{d (y , P) < 0 . 0 2 \\mid y \\in P _ {G T} \\right\\}\\right)}{\\mathcal {N} \\left(P _ {G T}\\right)}, \\tag {10}", + "image_path": "fab409442579d2bd206b9ef71cf0a42be662c73d0e7ea34b8362a21bdde42768.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "inline_equation", + "content": "P, P'" + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\hat{P}" + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": " are the 3D points of the current point cloud, previous point cloud, and recovered point cloud by the action, respectively. Intuitively, " + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "inline_equation", + "content": "P = P' + \\hat{P}" + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "inline_equation", + "content": "P_{GT}" + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": " is the ground truth reference point set; " + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "inline_equation", + "content": "d(x, P)" + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": " denotes the shortest distance from a point " + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": " to a point set " + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "inline_equation", + "content": "S'" + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": " are defined based on " + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "inline_equation", + "content": "P'" + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": ", respectively; " + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\cdot)" + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": " counts the number of points of a point set; " + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "inline_equation", + "content": "w_1, w_2, w_3, w_4" + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": " are the weights for balancing the contribution of the corresponding reward, which are set to 25, 1, 1.5, and 2 in all experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 558, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 651 + ], + "type": "text", + "content": "The architecture of the RL agent is simple yet effective; it contains an encoder, decoder, critic and actor heads (Figure 3). We adopt a commonly used 3D point encoder [15, 18] which extracts features from the input point cloud for downstream processing by the actor and critic. The critic head is only used during training, and the actor head predicts a viewpoint based on the input 3D point cloud." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 653, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 689 + ], + "type": "text", + "content": "The 2D Inpainting Module consists of three components: mask generation, normal map inpainting, and depth-map inpainting." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "The mask generator takes as input symmetric depth map pairs " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "D_{i}^{\\prime}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " corresponding to symmetric views " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "inline_equation", + "content": "v_{i}^{\\prime}" + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "text", + "content": ", and outputs two masks " + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "inline_equation", + "content": "M_{i}^{\\prime}" + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "text", + "content": " that mark the areas for each input map that need to be completed. To compute " + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "text", + "content": ", we start from an initial completion mask " + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "inline_equation", + "content": "\\hat{M}_i" + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "text", + "content": " obtained by marking all 'black' areas of a depth map " + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "text", + "content": ". However, this initial mask misses two important corner cases. First, not all 'black' areas should be completed. For example, the space between the legs of a chair. Second, areas with a missing front surface covered by a backside surface also need to be filled in. To cover the first case, we utilize a silhouette map generator to predict silhouette maps " + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "inline_equation", + "content": "K_{i}" + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "text", + "content": " of the target object to exclude erroneous 'holes'. The silhouette generation is modified from a standard U-Net architecture [19] by adding an extra input head and output head to address the pairwise inputs and outputs. Furthermore, we add pairwise skip connections between the two input and output heads for exchanging symmetry information. To provide additional depth cues, we concatenate to each input " + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "text", + "content": " (and " + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "inline_equation", + "content": "D_{i}^{\\prime}" + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "text", + "content": ") the initial input depth map, " + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "inline_equation", + "content": "D^{I}" + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "text", + "content": ", along the channel dimension. To address the second corner case, we compute a front-surface mask " + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "inline_equation", + "content": "H_{i}" + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "text", + "content": " by comparing the angle between the camera view direction and the surface normal at each point; we set the front-surface mask to 1 if the angle is less than " + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "text", + "content": ", or 0 otherwise (i.e., back facing). The final mask " + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "text", + "content": " is then obtained by combining " + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "inline_equation", + "content": "\\hat{M}_{i}" + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "inline_equation", + "content": "K_{i}" + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "inline_equation", + "content": "H_{i}" + }, + { + "bbox": [ + 304, + 300, + 545, + 587 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 375, + 596, + 545, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 596, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 375, + 596, + 545, + 610 + ], + "type": "interline_equation", + "content": "M _ {i} = \\left(\\hat {M} _ {i} + H _ {i}\\right) \\odot K _ {i}, \\tag {11}", + "image_path": "d88b609260a2b3607f1d373533de054ba1660f6e8b79a2678331c5605404ce6a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 376, + 613, + 545, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 613, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 376, + 613, + 545, + 626 + ], + "type": "interline_equation", + "content": "\\hat {M} _ {i} = \\Gamma (D _ {i}), \\tag {12}", + "image_path": "e9d19d46a6c600d7724bf587f3407c636438028039e9a908aa6de852f27d0579.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 376, + 629, + 545, + 642 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 629, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 376, + 629, + 545, + 642 + ], + "type": "interline_equation", + "content": "H _ {i} = \\Psi \\left(N _ {i}, v _ {i}\\right), \\tag {13}", + "image_path": "1e9619957f1d6f4fea575e12610e6fd5a2fab56ecc9d20c2d75c66e7a639f6f7.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 360, + 644, + 545, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 644, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 360, + 644, + 545, + 659 + ], + "type": "interline_equation", + "content": "K _ {i}, K _ {i} ^ {\\prime} = \\mathscr {K} \\left(\\left[ D _ {i}, D ^ {I} \\right], \\left[ D _ {i} ^ {\\prime}, D ^ {I} \\right]\\right), \\tag {14}", + "image_path": "c7e72dccea8a070611e0394de72b0c0311c29580e362d8db0a02a18bf0eaaa97.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 328, + 662, + 545, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 662, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 328, + 662, + 545, + 693 + ], + "type": "interline_equation", + "content": "\\Gamma (X), \\Psi (Z, y) = \\left\\{ \\begin{array}{l l} 1 & X = 0, \\langle Z, y \\rangle > 0 \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {15}", + "image_path": "b38ec7758495ab5cfd8f5ff091174e26e73c1421b7a287ce8cbf0b569e8f720a.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": " is the element-wise multiplication; " + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": " denotes the" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8899" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "silhouette map generator; " + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "[\\cdot, \\cdot]" + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": " indicates a channel-wise concatenation operation; " + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\Gamma(X)" + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\Psi(Z, y)" + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": " are two indicator functions, that only differ in the conditions for 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "type": "text", + "content": "The normal inpainting submodule " + }, + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "type": "text", + "content": " takes as input a pair of partial normal maps " + }, + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "type": "inline_equation", + "content": "N_{i}^{\\prime}" + }, + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "type": "text", + "content": " corresponding to symmetric viewpoints " + }, + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "type": "inline_equation", + "content": "v_{i}^{\\prime}" + }, + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "type": "text", + "content": " as well as their corresponding masks " + }, + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "type": "inline_equation", + "content": "M_{i}^{\\prime}" + }, + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "type": "text", + "content": ", and outputs a pair of completed normal maps " + }, + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "type": "inline_equation", + "content": "\\hat{N}_i" + }, + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "type": "inline_equation", + "content": "\\hat{N}_i^\\prime" + }, + { + "bbox": [ + 46, + 108, + 286, + 169 + ], + "type": "text", + "content": " respectively:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 176, + 286, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 176, + 286, + 192 + ], + "spans": [ + { + "bbox": [ + 70, + 176, + 286, + 192 + ], + "type": "interline_equation", + "content": "\\hat {N} _ {i}, \\hat {N} _ {i} ^ {\\prime} = \\mathcal {N} ([ N _ {i}, \\mathbf {1} - M _ {i} ], [ N _ {i} ^ {\\prime}, \\mathbf {1} - M _ {i} ^ {\\prime} ]). \\tag {16}", + "image_path": "724f77187a37b4734418c59ce19c618ff0d0c564fe72aa424e8533030f61158e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 198, + 286, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 286, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 286, + 305 + ], + "type": "text", + "content": "The normal inpainting submodule is based on PConv [14], a partial-convolution-layer-based UNet architecture that is robust to irregularly shaped holes [9, 14]. We modify PConv to accommodate the dual in and output by duplicating the input head and last layer of the decoder and adding pairwise skip connections. Note, PConv also applies a convolution with a unit fixed kernel on the input mask to perform mask-updating, and hence the target inpainting region should be marked 0 in the mask (i.e., the inverse of mask " + }, + { + "bbox": [ + 46, + 198, + 286, + 305 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 46, + 198, + 286, + 305 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "text", + "content": "The depth-map inpainting submodule " + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "text", + "content": " takes as input a pair of partial depth maps " + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "inline_equation", + "content": "D_{i}^{\\prime}" + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "text", + "content": " corresponding to symmetric viewpoints " + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "inline_equation", + "content": "v_{i}^{\\prime}" + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "text", + "content": " as well as their corresponding masks (" + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "inline_equation", + "content": "M_{i}^{\\prime}" + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "text", + "content": ") and inpainted normal maps (" + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "inline_equation", + "content": "\\hat{N}_{i}" + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "inline_equation", + "content": "\\hat{N}_{i}^{\\prime}" + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "text", + "content": "), and outputs a pair of completed depth maps " + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "inline_equation", + "content": "\\hat{D}_i" + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "inline_equation", + "content": "\\hat{D}_i^\\prime" + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "text", + "content": " and corresponding confidence maps " + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "inline_equation", + "content": "C_i^\\prime" + }, + { + "bbox": [ + 47, + 307, + 286, + 380 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 388, + 287, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 388, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 67, + 388, + 287, + 430 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} (\\hat {D} _ {i}, C _ {i}), (\\hat {D} _ {i} ^ {\\prime}, C _ {i} ^ {\\prime}) = \\mathcal {D} \\left(\\left[ D _ {i}, \\hat {N} _ {i}, \\mathbf {1} - M _ {i} \\right], \\right. \\\\ \\left[ D _ {i} ^ {\\prime}, \\hat {N} _ {i} ^ {\\prime}, \\mathbf {1} - M _ {i} ^ {\\prime} \\right]). \\tag {17} \\\\ \\end{array}", + "image_path": "7eb69b290e5121df4f9c1899e1bc06ecd7c9811b7db66f477e648d5e3bfa5732.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 437, + 286, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 286, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 286, + 508 + ], + "type": "text", + "content": "The confidence scores indicate the confidence of the inpainted pixel (i.e., the distance of the estimated depth-map to the ground truth map on missing regions). Our implementation of the depth inpainting submodule is also based on PConv [14] with similar modifications to support the additional input and output channels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 509, + 286, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 286, + 568 + ], + "type": "text", + "content": "Point Cloud Update: We merge the inpainted depth, normal, and confidence maps into the previous point cloud by reprojecting all points in each depth map and corresponding normals to a point cloud if their confidence score exceeds a threshold (set to 0.5 in our implementation)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 575, + 166, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 166, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 166, + 586 + ], + "type": "text", + "content": "3.3. Iterative Refinement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "To improve the robustness of the symmetry plane estimation and of the shape completion, we iteratively refine the output of ISCNet. A key observation is that we can select any suitable viewpoint to generate a new depth image to serve as input to update the symmetry plane. Because it is unlikely that a single viewpoint would be ideal for all cases, we project the updated point cloud from three pre-selected trial viewpoints to generate their respective depth maps. Next, the three depth maps are passed (separately) into the symmetry detection module, and the symmetry plane with" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "the highest confidence score is subsequently used for the remainder of our pipeline. The set of pre-determined viewpoints " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "(v_{8}^{V}, v_{12}^{V}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "v_{13}^{V}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " from VRCNet [16]) are empirically chosen and kept constant for all experiments. The final updated point cloud is converted to a mesh surface using Poisson Surface Reconstruction [11]; see Figure 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 144, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 252 + ], + "type": "text", + "content": "To balance computational cost versus gain in accuracy, the number of iterations is set to 20 based on an experimental analysis on ShapeNet. We update the symmetry plane every 4th iteration such that the RL agent can leverage the progressively completed point cloud for improving the reflection viewpoint estimates. In addition, to avoid accumulating incorrectly completed shape-parts due to an inaccurate symmetry plane, we reset the point cloud to its initial state when re-estimating the symmetry plane." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 258, + 399, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 258, + 399, + 269 + ], + "spans": [ + { + "bbox": [ + 306, + 258, + 399, + 269 + ], + "type": "text", + "content": "3.4. Loss Functions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 276, + 545, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 545, + 359 + ], + "type": "text", + "content": "ISCNet is trained in three stages. In the first stage, the symmetry detection component and 2D inpainting module are trained separately. In the second stage, all components except the RL agent of ISCNet are then jointly fine-tuned. Finally, we train the RL agent by using a PPO algorithm with the parameters of the other components fixed using the loss function from [20]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 360, + 531, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 531, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 531, + 372 + ], + "type": "text", + "content": "The first-stage symmetry detection loss is defined by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 376, + 545, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 376, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 312, + 376, + 545, + 403 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {S} = \\mathcal {C} \\left(\\mathcal {M} \\left(P, S _ {\\mathrm {E}}\\right), \\mathcal {M} \\left(P, S _ {\\mathrm {G}}\\right)\\right) + \\mathcal {B} \\left(\\Theta_ {t _ {1}} \\left(S _ {\\mathrm {E}} ^ {N}, S _ {\\mathrm {G}} ^ {N}\\right), 1\\right), \\tag {18}", + "image_path": "85122f3a6ad22c63ec8f73ab93dc59aba1ad2e13eac7a9c8f059072ec0e93810.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "type": "text", + "content": " measures the Chamfer Distance (CD) between the reconstructed 3D points " + }, + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "type": "text", + "content": " reflected (via a reflection function " + }, + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "type": "text", + "content": ") over the estimated symmetry plane " + }, + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{E}}" + }, + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "type": "text", + "content": " and reference symmetry plane " + }, + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{G}}" + }, + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "type": "text", + "content": ". We further regularize this loss via a binary cross entropy (BCE) loss " + }, + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "type": "inline_equation", + "content": "(\\mathcal{B})" + }, + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "type": "text", + "content": " measuring the acute angle similarity between " + }, + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{E}}^{N}" + }, + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{G}}^{N}" + }, + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "type": "text", + "content": ", the corresponding normalized symmetry plane normals of the estimated and reference, via an indicator function " + }, + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "type": "inline_equation", + "content": "\\Theta_{t_1}(X,Y)" + }, + { + "bbox": [ + 304, + 403, + 545, + 509 + ], + "type": "text", + "content": " defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 323, + 514, + 545, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 514, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 323, + 514, + 545, + 546 + ], + "type": "interline_equation", + "content": "\\Theta_ {t _ {1}} (X, Y) = \\left\\{ \\begin{array}{l l} 1 & \\operatorname {a r c c o s} (| \\langle X, Y \\rangle |) < = t _ {1} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {19}", + "image_path": "4379d8f95c8698327cadb46609c85afaadcf2683e08eb8209d9d8059acd7a65d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 552, + 448, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 448, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 448, + 563 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 306, + 552, + 448, + 563 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 306, + 552, + 448, + 563 + ], + "type": "text", + "content": " a slack parameter set to " + }, + { + "bbox": [ + 306, + 552, + 448, + 563 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 306, + 552, + 448, + 563 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 563, + 488, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 488, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 488, + 575 + ], + "type": "text", + "content": "The first-stage inpainting loss is defined as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 377, + 581, + 545, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 581, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 377, + 581, + 545, + 594 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {I} = \\mathcal {L} _ {\\mathrm {K}} + \\mathcal {L} _ {\\mathrm {N}} + \\mathcal {L} _ {\\mathrm {D}}, \\tag {20}", + "image_path": "02e403024a3725a0ab5afec58a28eca4bd0d16e99166366ce22cab715aa1ba9b.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 376, + 597, + 545, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 597, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 376, + 597, + 545, + 609 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {K}} = \\mathcal {B} \\left(K _ {\\mathrm {E}}, K _ {\\mathrm {G}}\\right), \\tag {21}", + "image_path": "7a68753445154efda7d8d75bdf10adaabffd6369332267fed4db4975a36448f7.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 376, + 612, + 545, + 627 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 612, + 545, + 627 + ], + "spans": [ + { + "bbox": [ + 376, + 612, + 545, + 627 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {N} = \\left\\| N _ {E} ^ {M _ {\\mathrm {G}}} - N _ {G} ^ {M _ {\\mathrm {G}}} \\right\\| _ {1}, \\tag {22}", + "image_path": "f2cb15b6f51e2017b99753890f5f749594c023ce0668a4ab03039cfac9ba85b6.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 376, + 628, + 481, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 628, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 376, + 628, + 481, + 643 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D}} = \\left\\| D _ {\\mathrm {E}} ^ {M _ {\\mathrm {G}}} - D _ {\\mathrm {G}} ^ {M _ {\\mathrm {G}}} \\right\\| _ {1} +", + "image_path": "1c789df37372a5d031095d3ed01b01700d6d79052891425bfbd2d6d3fdf0dccb.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 389, + 645, + 545, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 645, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 389, + 645, + 545, + 661 + ], + "type": "interline_equation", + "content": "\\mathcal {B} \\left(C _ {\\mathrm {E}} ^ {M _ {\\mathrm {G}}}, \\Phi_ {t _ {2}} \\left(D _ {\\mathrm {E}} ^ {M _ {\\mathrm {G}}}, D _ {\\mathrm {G}} ^ {M _ {\\mathrm {G}}}\\right)\\right), \\tag {23}", + "image_path": "9e45430e09059ef55167c2d5d96f8519c1fb20214022738ee6e43db236b7b03a.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 368, + 662, + 545, + 676 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 662, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 368, + 662, + 545, + 676 + ], + "type": "interline_equation", + "content": "X ^ {M _ {\\mathrm {G}}} = X \\odot M _ {\\mathrm {G}}, \\tag {24}", + "image_path": "0dd5e7701f2d0e43ab18440f78fb9cf3d08fccf6597762c65de22f9a1a6aff43.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 345, + 679, + 545, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 679, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 345, + 679, + 545, + 711 + ], + "type": "interline_equation", + "content": "\\Phi_ {t _ {2}} (X, Y) = \\left\\{ \\begin{array}{l l} 1 & \\| X - Y \\| _ {1} < = t _ {2} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {25}", + "image_path": "a1860d95ac15129a2cc52a7caf0c39c7a51ccea03b66f3971aa52b6c4289dbce.jpg" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8900" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 534, + 250 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 534, + 250 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 534, + 250 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 534, + 250 + ], + "type": "table", + "html": "
CD/MD (×102)PlaneCabinetCarChairLampSofaTableVesselAverage
PCN° [41]2.87/2.443.42/3.652.12/2.014.32/4.085.49/5.192.88/3.354.49/3.722.23/2.253.48/3.34
MSN° [15]3.29/2.642.57/2.201.95/1.693.60/2.914.58/4.313.61/3.123.62/3.042.08/1.893.16/2.72
PoinTr° [40]1.10/0.982.86/3.121.72/1.653.33/2.923.45/2.922.58/2.973.44/2.821.40/1.412.48/2.35
VRCNet° [16]2.61/1.922.77/2.201.92/1.613.43/2.613.59/2.873.41/2.733.89/3.121.92/1.572.94/2.33
SnowF° [37]1.08/0.832.83/3.021.73/1.663.08/2.813.53/2.782.46/2.513.51/2.861.52/1.302.47/2.22
VE-PCN° [32]8.80/5.396.72/5.952.91/2.377.87/6.5710.94/8.538.76/7.226.33/4.993.75/2.737.01/5.47
IFNet◇ [3]1.33/1.365.17/4.773.16/2.973.13/3.241.74/1.663.61/3.872.26/2.412.26/2.522.83/2.85
Front2Back_R5† [39]1.44/1.333.14/2.591.94/1.922.49/2.312.48/2.302.53/2.213.02/3.021.88/1.492.36/2.15
Front2Back_GT† [39]1.19/1.182.78/2.211.90/1.772.00/1.861.69/1.772.13/2.002.03/2.261.53/1.511.91/1.82
Ours† (Iteration 20)0.82/0.842.04/1.631.56/1.351.54/1.391.30/1.341.84/1.531.67/1.541.04/1.001.48/1.33
Unaligned ISCNet1.43/1.432.56/2.092.15/1.882.10/1.921.76/1.872.33/1.992.33/2.251.66/1.602.04/1.88
No-Pair ISCNet0.88/1.022.28/1.991.90/1.691.75/1.472.12/1.832.07/1.762.11/1.541.41/1.541.82/1.61
No-Normal ISCNet1.08/1.072.41/2.042.18/1.861.89/1.531.91/1.782.27/1.812.14/1.811.61/1.591.94/1.69
Iteration 40.89/0.832.23/1.941.74/1.521.67/1.541.43/1.422.03/1.791.70/1.581.15/1.061.61/1.46
Iteration 80.85/0.872.19/1.731.66/1.441.60/1.451.32/1.391.92/1.631.73/1.621.10/1.041.55/1.40
Iteration 120.84/0.872.16/1.721.60/1.381.57/1.411.28/1.371.88/1.551.70/1.581.06/1.041.51/1.36
Iteration 160.83/0.842.11/1.671.57/1.361.55/1.401.26/1.361.92/1.621.67/1.541.04/1.001.49/1.35
", + "image_path": "2564c1ecca370653770670a2b8cfac8f1409171180e634d396fe58ff913f1fb3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 250, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 250, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 547, + 304 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation of ISCNet compared to prior work on the synthetic ShapeNet dataset ( " + }, + { + "bbox": [ + 46, + 250, + 547, + 304 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 250, + 547, + 304 + ], + "type": "text", + "content": " point cloud methods, " + }, + { + "bbox": [ + 46, + 250, + 547, + 304 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 46, + 250, + 547, + 304 + ], + "type": "text", + "content": " implicit function methods, and " + }, + { + "bbox": [ + 46, + 250, + 547, + 304 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 250, + 547, + 304 + ], + "type": "text", + "content": " image-based methods) in terms of a mesh-to-mesh symmetric " + }, + { + "bbox": [ + 46, + 250, + 547, + 304 + ], + "type": "inline_equation", + "content": "{L}_{1}" + }, + { + "bbox": [ + 46, + 250, + 547, + 304 + ], + "type": "text", + "content": " metric (MD) and the Chamfer Distance (CD) (lowest, 2nd lowest, 3rd lowest error color coding). In addition, we show the errors on three ablation variants of ISCNet (without symmetry plane alignment, without pair-wise inpainting, and without utilizing normal information), as well as the evolution of errors per iteration." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 118, + 312, + 476, + 411 + ], + "blocks": [ + { + "bbox": [ + 118, + 312, + 476, + 411 + ], + "lines": [ + { + "bbox": [ + 118, + 312, + 476, + 411 + ], + "spans": [ + { + "bbox": [ + 118, + 312, + 476, + 411 + ], + "type": "table", + "html": "
CD/MD (×102)CabinetChairLampSofaTableAverage
PCN° [41]9.76/10.446.27/7.609.60/10.456.02/7.6311.83/11.948.70/9.61
MSN° [15]5.71/5.175.86/5.405.74/4.774.95/4.016.43/6.865.74/5.24
PoinTr° [40]8.33/7.686.40/6.339.66/9.196.12/5.217.84/8.477.67/7.38
VRCNet° [16]5.05/3.985.87/5.565.95/5.034.45/3.646.00/6.605.47/4.96
SnowF° [37]8.66/7.896.47/6.438.99/8.875.63/4.917.22/8.757.39/7.37
VE-PCN° [32]9.26/6.926.38/5.986.12/5.485.64/4.708.06/7.737.09/6.16
IFNet◇ [3]12.62/7.095.38/6.608.95/8.205.05/4.215.92/7.137.59/6.64
Front2Back_GT† [39]11.22/8.876.86/7.557.81/6.646.92/6.576.85/7.737.93/7.47
Ours † (Iteration 20)4.88/3.893.67/3.994.62/4.163.27/2.864.53/5.164.19/4.01
", + "image_path": "7dad58b13507594ceaaea4394748a0d9416b103cc5dca2abc92b835deb5bf183.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 411, + 545, + 421 + ], + "lines": [ + { + "bbox": [ + 47, + 411, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 545, + 421 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation of ISCNet compared to prior work on real-world ScanNet dataset. The notations and colors follow Table 1." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{K}}" + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{D}}" + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "text", + "content": " measure the loss over: silhouette mask " + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "text", + "content": ", depth " + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "text", + "content": ", and normal " + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "text", + "content": " between the estimated " + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "inline_equation", + "content": "(\\cdot_E)" + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "text", + "content": " and reference " + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "inline_equation", + "content": "(\\cdot_G)" + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "text", + "content": " components. " + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "inline_equation", + "content": "\\Phi_{t_2}" + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "text", + "content": " is an indicator function on the " + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "text", + "content": " distance between maps with a slack parameter " + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "text", + "content": " set to 0.02, and " + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "inline_equation", + "content": "X^{M_G}" + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "text", + "content": " is the masking operator with a reference mask over a component " + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "inline_equation", + "content": "X \\in \\{N, D, C\\}" + }, + { + "bbox": [ + 46, + 441, + 287, + 549 + ], + "type": "text", + "content": ". At this stage, we do not yet explicitly enforce symmetry to allow the inpainting network to converge to a good starting point for the second stage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 555, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 578 + ], + "type": "text", + "content": "The second-stage loss is the sum of both first-stage losses plus an additional loss that explicitly enforces symmetry:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 609, + 287, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 609, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 91, + 609, + 287, + 624 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\alpha \\mathcal {L} ^ {S} + \\beta \\mathcal {L} ^ {I} + \\gamma \\mathcal {L} ^ {J}, \\tag {26}", + "image_path": "b0a8705aa8ca8fe1d9641094ec217f0905fa93d128ad177905394e87dcc103c7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 86, + 627, + 249, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 627, + 249, + 641 + ], + "spans": [ + { + "bbox": [ + 86, + 627, + 249, + 641 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {J} = \\Sigma_ {i = 1} ^ {4} \\left(\\| D _ {i} ^ {M _ {\\mathrm {G}}} - \\mathcal {F} \\left(D _ {i} ^ {\\prime}\\right) \\odot M _ {\\mathrm {G}} \\| _ {1} + \\right.", + "image_path": "6d0cf7fc5c9a0537209a11f222a80f17f2fbef9b92347b18eb20c63d90143ddb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 643, + 287, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 643, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 107, + 643, + 287, + 658 + ], + "type": "interline_equation", + "content": "\\left\\| N _ {i} ^ {M _ {\\mathrm {G}}} - \\mathcal {F} \\left(N _ {i} ^ {\\prime}\\right) \\odot M _ {\\mathrm {G}} \\right\\| _ {1}), \\tag {27}", + "image_path": "883743019db83337a5a9931a0c4ec4ce1158671212a3810d29b7d5daa4fc33ce.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " are balancing weights, " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " is a 2D horizontal mirror operator, and " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "X'" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " the respective components from symmetric viewpoints." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 441, + 380, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 380, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 380, + 453 + ], + "type": "text", + "content": "4. Assessment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "text", + "content": "We implement ISCNet in PyTorch [10], and train and test ISCNet for eight categories (airplane, cabinet, car, chair, lamp, sofa, table, and vessel) from the ShapeNet dataset [2]. As in prior work [15, 41], we partition the dataset in 29,774 shapes for training and 1,200 for testing. We use the 26 viewpoints provided by VRCNet [16] to generate synthetic depth images and normal maps for training and 8 randomly generated viewpoints for testing. In addition, we also test ISCNet with 456 scanned objects from ScanNet [5], excluding the categories not present in ScanNet (i.e. airplanes, cars, and vessels). The corresponding reference shapes are provided by Scan2CAD [1]; our and all competing methods are only trained with ShapeNet." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "To train ISCNet, we first pre-train the four submodules separately for 50 epochs: symmetry detection using " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{S}}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " (Equation 18), mask generation using " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{K}}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " (Equation 21), 2D normal inpainting using " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{N}}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " (Equation 22), and 2D depth inpainting using " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{D}}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " (Equation 23). Finally, we refine the complete network for an additional 10 epochs using " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " as loss and set the weight parameters " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " to 0.01, 0.05, and 1, respectively (Equation 26). While the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8901" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 72, + 517, + 329 + ], + "blocks": [ + { + "bbox": [ + 75, + 72, + 517, + 329 + ], + "lines": [ + { + "bbox": [ + 75, + 72, + 517, + 329 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 517, + 329 + ], + "type": "image", + "image_path": "baf130b169d97b08987abdc5ad61e14ed3031125166fb83d72d56ab672e3615e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 338, + 546, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 338, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 546, + 361 + ], + "type": "text", + "content": "Figure 4. Visual comparison of shape completion quality of ISCNet versus prior work on the synthetic ShapeNet dataset; the observed areas marked in red (the second column)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 80, + 373, + 512, + 564 + ], + "blocks": [ + { + "bbox": [ + 80, + 373, + 512, + 564 + ], + "lines": [ + { + "bbox": [ + 80, + 373, + 512, + 564 + ], + "spans": [ + { + "bbox": [ + 80, + 373, + 512, + 564 + ], + "type": "image", + "image_path": "1eaab14e4e00d79233956c7a3cd95612a8147d1355ac0e63a60f0877285f8630.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 567, + 546, + 589 + ], + "lines": [ + { + "bbox": [ + 46, + 567, + 546, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 546, + 589 + ], + "type": "text", + "content": "Figure 5. Visual comparison of shape completion quality of ISCNet versus prior work on real-world ScanNet dataset. Note that the sofa and table are non-symmetric objects." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 609, + 289, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 289, + 706 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 289, + 706 + ], + "type": "text", + "content": "training datasets are generated from 26 pre-defined viewpoints for the first training stage, the second stage (finetuning) uses randomly selected viewpoints instead of the viewpoints chosen by the RL agent. In the latter case, we generate the corresponding ground truth depth images and normal maps on the fly during the training process. As a consequence of this phased training strategy, the distribution of (RL selected) viewpoints on the testing dataset cover" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 609, + 545, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 609, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 609, + 545, + 644 + ], + "type": "text", + "content": "the whole sphere, thereby avoiding overfitting to the 26 pre-selected viewpoints (see the supplemental material for more details)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 653, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 546, + 715 + ], + "type": "text", + "content": "We train all networks with the Adam optimizer with a learning rate of " + }, + { + "bbox": [ + 304, + 653, + 546, + 715 + ], + "type": "inline_equation", + "content": "2e^{-4}" + }, + { + "bbox": [ + 304, + 653, + 546, + 715 + ], + "type": "text", + "content": " for pretraining and " + }, + { + "bbox": [ + 304, + 653, + 546, + 715 + ], + "type": "inline_equation", + "content": "4e^{-5}" + }, + { + "bbox": [ + 304, + 653, + 546, + 715 + ], + "type": "text", + "content": " for finetuning on a dual NVIDIA Gforce RTX 3090 setup. On this setup, for 20 iterations, ISCNet takes 81 seconds to complete a shape from a single depth map, where the image" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8902" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 83 + ], + "type": "text", + "content": "rendering and point clouds merging take around 59 seconds." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 286, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 286, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 286, + 192 + ], + "type": "text", + "content": "For validation, we measure the reconstruction quality by a mesh-to-mesh symmetric " + }, + { + "bbox": [ + 46, + 84, + 286, + 192 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 84, + 286, + 192 + ], + "type": "text", + "content": " distance (MD) [4] between the reference and completed meshes. Further, to quantify the reconstruction accuracy of surface details, we follow Yao et al. [39] and use the Chamfer Distance (CD) on 100K Poisson disk sampled [43] surface points from the reconstructed and reference meshes. Finally, following Zhou et al. [45] we quantify symmetry plane accuracy by comparing the cosine distance between symmetry plane normals." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 199, + 182, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 199, + 182, + 211 + ], + "spans": [ + { + "bbox": [ + 47, + 199, + 182, + 211 + ], + "type": "text", + "content": "4.1.Quantitative Evaluation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 217, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 287, + 480 + ], + "type": "text", + "content": "3D Shape Completion. We compare the performance of ISCNet to 8 competing methods subdivided in three categories: Point-cloud completion methods (PCN [41], MSN [15], PoinTr [40], VRCNet [16], ShowFlakeNet [37], VE-PCN [32]), implicit-function-based methods (IFNet [3]), and image-based completion methods (2 variants of Front2Back " + }, + { + "bbox": [ + 46, + 217, + 287, + 480 + ], + "type": "inline_equation", + "content": "[39]^{1}" + }, + { + "bbox": [ + 46, + 217, + 287, + 480 + ], + "type": "text", + "content": "). All competing methods are retrained on ShapeNet and tested with identical exemplars under the same conditions. The quantitative comparisons (using the CD and MD metrics) are summarized in Table 1 and Table 2 for ShapeNet and ScanNet, respectively. From these, we can see that ISCNet offers the best overall performance for the majority of categories on both synthetic and scanned data, followed by Front2Back [39]. SnowFlakeNet [37] performs better than ISCNet on the plane shapes with a marginal accuracy gain. Compared to Front2Back, we can see that ISCNet outperforms Front2Back with the ground truth symmetry information. For completeness, Table 3 compares ISCNet with Front2Back and IFNet for the remaining five classes from ShapeNet, without retraining or finetuning. Our approach still outperforms existing work except for the rifle class compared to IFNet." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 47, + 490, + 302, + 555 + ], + "blocks": [ + { + "bbox": [ + 47, + 490, + 302, + 555 + ], + "lines": [ + { + "bbox": [ + 47, + 490, + 302, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 302, + 555 + ], + "type": "table", + "html": "
(×102)BenchDisplaySpeakerRiflePhoneAverage (unseen-5)Average (total-13)
CDF2B2.373.193.401.723.132.762.24
IFNet1.813.736.351.373.343.323.02
Ours1.442.473.091.432.102.111.72
MDF2B2.192.973.171.592.972.582.11
IFNet1.773.656.241.353.283.263.01
Ours1.412.453.011.422.072.071.61
", + "image_path": "250ed7ef5d91aa87f8db7b7db1dabc50b190b579f20156e1aebfb4e3bc1a7619.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 554, + 286, + 586 + ], + "lines": [ + { + "bbox": [ + 47, + 554, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 286, + 586 + ], + "type": "text", + "content": "Table 3. Quantitative evaluation of ISCNet compared to Front2Back and IFNet for the remaining 5 classes of ShapeNet, i.e., unseen-5 labels." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 597, + 286, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 597, + 286, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 286, + 657 + ], + "type": "text", + "content": "Figure 4 and 5 provide additional qualitative comparisons. Visually, ISCNet provides the 'cleanest' shape completion for all eight tasks. In particular, in the plane completion task, ISCNet is better able to retain high-frequency surface details, and in the cabinet and lamp completion tasks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "text", + "content": "ISCNet is the only method to fully complete the surface without holes. In the case of the chair and table completion tasks, ISCNet more robustly identifies areas to be completed without erroneously merging thin features. Although our method is designed for symmetric objects, we also obtain good reconstruction quality on non-symmetric objects (sofa and table in Figure 5)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 173, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 173, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 173, + 545, + 316 + ], + "type": "text", + "content": "Symmetry Detection. To assess the accuracy of the symmetry plane estimation, we compare it against two state-of-the-art single-view based pose estimation and symmetry detection methods, RotationContinuity [44] and NeRD [45]. Figure 6a shows that ISCNet's second update of the symmetry plane offers a similar performance as NeRD. Furthermore, with increasing iterations, the accuracy of the estimated symmetry plane improves (i.e., compared to the first symmetry plane estimate2, the occurrence of symmetry plane estimates with an error less than " + }, + { + "bbox": [ + 305, + 173, + 545, + 316 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 305, + 173, + 545, + 316 + ], + "type": "text", + "content": ", improves from " + }, + { + "bbox": [ + 305, + 173, + 545, + 316 + ], + "type": "inline_equation", + "content": "32.86\\%" + }, + { + "bbox": [ + 305, + 173, + 545, + 316 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 173, + 545, + 316 + ], + "type": "inline_equation", + "content": "56.96\\%" + }, + { + "bbox": [ + 305, + 173, + 545, + 316 + ], + "type": "text", + "content": "), demonstrating the effectiveness of the iterative refinement for symmetry plane estimation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 325, + 399, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 325, + 399, + 338 + ], + "spans": [ + { + "bbox": [ + 306, + 325, + 399, + 338 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 344, + 545, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 344, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 305, + 344, + 545, + 583 + ], + "type": "text", + "content": "Importance of Symmetry. We validate the effectiveness of symmetry cues for shape completion in two ablation experiments. As a first experiment (Table 1, Unaligned ISCNet), we replace the estimated transformation that aligns the symmetry plane with the X-Y plane by a random rotation matrix and a perturbed center (adding normal distributed noise with " + }, + { + "bbox": [ + 305, + 344, + 545, + 583 + ], + "type": "inline_equation", + "content": "\\sigma = 0.1" + }, + { + "bbox": [ + 305, + 344, + 545, + 583 + ], + "type": "text", + "content": "). Consequently, ISCNet cannot leverage any symmetry cues. In a second experiment (Table 1, No-Pair ISCNet), we process each of the reflection viewpoints separately. From Table 1, we can see that the unaligned ISCNet suffers a significant performance degradation; CD increases by 0.56 on average, and MD increases 0.55 on average. By aligning the shape to the symmetry plane, no-pair ISCNet is able to leverage symmetry implicitly (by learning to copy from the negative Z coordinate). However, compared to the full ISCNet, the no-pair ISCNet still sees a performance drop (CD and MD are increased by 0.34 and 0.28, respectively). We refer to the supplementary material for a qualitative comparison of excluding explicit leveraging of the symmetry cues." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 600, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 600, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 305, + 600, + 545, + 685 + ], + "type": "text", + "content": "Role of Normal Inpainting. To demonstrate the importance of performing normal inpainting, we eliminate the inpainted normal as input from the depth inpainting submodule (Table 1, No-Normal ISCNet). Without normal cues, CD increases by 0.46 and MD by 0.36. We refer to the supplementary material for a qualitative visualization of the impact of omitting the normal inpainting step." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 544, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 544, + 712 + ], + "type": "text", + "content": "2The accuracy of first symmetry estimate is similar to that of DenseFusion [30], as our symmetry estimation is based this method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 664, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 664, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 664, + 286, + 712 + ], + "type": "text", + "content": "1The information in [39] was insufficient to replicate the symmetry plane estimation module. Therefore, we include 2 variants, trained and tested with the ground truth symmetry plane (Front2Back_GT), and with the reference symmetry plane normal perturbed within " + }, + { + "bbox": [ + 46, + 664, + 286, + 712 + ], + "type": "inline_equation", + "content": "[-5^{\\circ}, 5^{\\circ}]" + }, + { + "bbox": [ + 46, + 664, + 286, + 712 + ], + "type": "text", + "content": " (Front2Back_R5)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8903" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "content": "Impact of Iterative Refinement. To demonstrate that both symmetry plane estimation as well as shape completion improve with additional iterations, we compare the MD and CD errors per iteration (Table 1) as well as the symmetry plane normal error progression per iteration (Figure 6b). In both cases, we can see that the error improves per iteration and the improvement diminishes per subsequent iteration (justifying a 20 iteration stopping criterion)." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 180, + 165, + 279 + ], + "blocks": [ + { + "bbox": [ + 50, + 180, + 165, + 279 + ], + "lines": [ + { + "bbox": [ + 50, + 180, + 165, + 279 + ], + "spans": [ + { + "bbox": [ + 50, + 180, + 165, + 279 + ], + "type": "image", + "image_path": "ef08d0aeb46556f227056aa264db3866194e0fee906be0935fc8ec6080050fdc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 280, + 120, + 288 + ], + "lines": [ + { + "bbox": [ + 111, + 280, + 120, + 288 + ], + "spans": [ + { + "bbox": [ + 111, + 280, + 120, + 288 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 165, + 180, + 286, + 279 + ], + "blocks": [ + { + "bbox": [ + 165, + 180, + 286, + 279 + ], + "lines": [ + { + "bbox": [ + 165, + 180, + 286, + 279 + ], + "spans": [ + { + "bbox": [ + 165, + 180, + 286, + 279 + ], + "type": "image", + "image_path": "4919d2a3ea659882d9fd192d0e8746bdae4da6fb8d1a42af67424f6da78ca44b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 280, + 229, + 288 + ], + "lines": [ + { + "bbox": [ + 220, + 280, + 229, + 288 + ], + "spans": [ + { + "bbox": [ + 220, + 280, + 229, + 288 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 293, + 287, + 348 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 348 + ], + "type": "text", + "content": "Figure 6. Impact of the number of iterations on (a) the accuracy of the symmetry plane estimation, and (b) the shape reconstruction CD/MD accuracy and average and median symmetry plane estimation errors. 'Our SP " + }, + { + "bbox": [ + 46, + 293, + 287, + 348 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 293, + 287, + 348 + ], + "type": "text", + "content": "' refers to ISCNet's " + }, + { + "bbox": [ + 46, + 293, + 287, + 348 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 293, + 287, + 348 + ], + "type": "text", + "content": "-th predicted symmetry plane." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 360, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 287, + 420 + ], + "type": "text", + "content": "Limitations The main bottleneck in ISCNet is the image rendering and the 3D point cloud processing step in each iteration. Furthermore, ISCNet is only able to leverage a single global symmetry plane, and thus its performance is suboptimal when multiple symmetries are present." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 432, + 119, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 432, + 119, + 444 + ], + "spans": [ + { + "bbox": [ + 47, + 432, + 119, + 444 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 453, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 597 + ], + "type": "text", + "content": "We present ISCNet, a single-depth image joint symmetry plane estimation and shape completion method that explicitly leverages symmetry cues. ISCNet alternates between estimating the symmetry plane and completing the shape to enhance the accuracy of both components. Our method is informed by two key insights: (1) shape completion and symmetry plane estimation are mutually beneficial tasks, and (2) by selecting symmetric virtual viewpoints, we can explicitly inform the inpainting module on symmetry relations. Our experiments and ablation studies demonstrate that ISCNet can yield more accurate shape completions from a single-depth image than competing methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 609, + 144, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 144, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 144, + 623 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "We thank the anonymous reviewers for the insightful and constructive comments. This work was supported in part by National Key Research and Development Program of China (2022ZD0210500, 2021ZD0111902), National Natural Science Foundation of China (61972067, U19B2039, U21B2038), the Distinguished Young Scholars Funding of Dalian (No. 2022RJ01). Pieter Peers was supported in part" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "by NSF grant IIS-1909028. Felix Heide was supported by an NSF CAREER Award (2047359), a Packard Foundation Fellowship, a Sloan Research Fellowship, a Sony Young Faculty Award, a Project X Innovation Award, and an Amazon Science Research Award." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 143, + 365, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 365, + 156 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 365, + 156 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 163, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 311, + 163, + 547, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 163, + 547, + 219 + ], + "spans": [ + { + "bbox": [ + 311, + 163, + 547, + 219 + ], + "type": "text", + "content": "[1] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 220, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 220, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 311, + 220, + 547, + 274 + ], + "type": "text", + "content": "[2] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 275, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 275, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 311, + 275, + 547, + 319 + ], + "type": "text", + "content": "[3] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6970-6981, 2020. 1, 6, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 320, + 547, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 320, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 311, + 320, + 547, + 354 + ], + "type": "text", + "content": "[4] Paolo Cignoni, Claudio Rocchini, and Roberto Scopigno. Metro: measuring error on simplified surfaces. In Computer Graphics Forum, volume 17, pages 167-174, 1998. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 354, + 547, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 354, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 311, + 354, + 547, + 409 + ], + "type": "text", + "content": "[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 410, + 547, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 410, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 311, + 410, + 547, + 453 + ], + "type": "text", + "content": "[6] Shivam Duggal and Deepak Pathak. Topologically-aware deformation fields for single-view 3d reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1536-1546, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 311, + 455, + 547, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 455, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 311, + 455, + 547, + 498 + ], + "type": "text", + "content": "[7] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 605-613, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 311, + 500, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 500, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 311, + 500, + 545, + 533 + ], + "type": "text", + "content": "[8] Boqing Gong, Jianzhuang Liu, Xiaogang Wang, and Xiaou Tang. Learning semantic signatures for 3d object retrieval. IEEE Trans. Multimedia, 15(2):369-377, 2012. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 311, + 534, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 534, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 311, + 534, + 545, + 600 + ], + "type": "text", + "content": "[9] Xiaoguang Han, Zhaoxuan Zhang, Dong Du, Mingdai Yang, Jingming Yu, Pan Pan, Xin Yang, Ligang Liu, Zixiang Xiong, and Shuguang Cui. Deep reinforcement learning of volume-guided progressive view inpainting for 3d point scene completion from a single depth image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 234-243, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "text", + "content": "[10] Sagar Imambi, Kolla Bhanu Prakash, and GR Kanagachidambaresan. Pytorch. In Programming with TensorFlow, pages 87-104. 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "text", + "content": "[11] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[12] Michael Kazhdan, Thomas Funkhouser, and Szymon Rusinkiewicz. Symmetry descriptors and 3d shape matching. In Proceedings of the 2004 Eurographics/ACM SIGGRAPH" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8904" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "symposium on Geometry processing, pages 115-123, 2004. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "type": "text", + "content": "[13] Wai Ho Li and Lindsay Kleeman. Segmentation and modeling of visually symmetric objects by robot actions. The International Journal of Robotics Research, 30(9):1124-1142, 2011. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 184 + ], + "type": "text", + "content": "[14] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Eur. Conf. Comput. Vis., pages 85-100, 2018. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 185, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 227 + ], + "type": "text", + "content": "[15] Minghua Liu, Lu Sheng, Sheng Yang, Jing Shao, and ShiMin Hu. Morphing and sampling network for dense point cloud completion. In AAAI, volume 34, pages 11596-11603, 2020. 2, 4, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "text", + "content": "[16] Liang Pan, Xinyi Chen, Zhongang Cai, Junzhe Zhang, Haiyu Zhao, Shuai Yi, and Ziwei Liu. Variational relational point completion network. In Int. Conf. Comput. Vis., pages 8524-8533, 2021. 2, 5, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 316 + ], + "type": "text", + "content": "[17] Joshua Podolak, Philip Shilane, Aleksey Golovinskiy, Szymon Rusinkiewicz, and Thomas Funkhouser. A planar-reflective symmetry transform for 3d shapes. In ACM SIGGRAPH 2006 Papers, pages 549-559. 2006. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 317, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 317, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 317, + 287, + 360 + ], + "type": "text", + "content": "[18] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 652-660, 2017. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 414 + ], + "type": "text", + "content": "[19] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241, 2015. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 449 + ], + "type": "text", + "content": "[20] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "text", + "content": "[21] Yifei Shi, Junwen Huang, Hongjia Zhang, Xin Xu, Szymon Rusinkiewicz, and Kai Xu. Symmetrynet: learning to predict reflectional and rotational symmetries of 3d shapes from single-view rgb-d images. ACM Trans. Graph., 39(6):1-14, 2020. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 504, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 547 + ], + "type": "text", + "content": "[22] Yifei Shi, Xin Xu, Junhua Xi, Xiaochang Hu, Dewen Hu, and Kai Xu. Learning to detect 3d symmetry from single-view rgb-d images with weak supervision. IEEE Trans. Pattern Anal. Mach. Intell., 2022. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 287, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 592 + ], + "type": "text", + "content": "[23] Ivan Sipiran, Robert Gregor, and Tobias Schreck. Approximate symmetry detection in partial 3d meshes. In Computer Graphics Forum, volume 33, pages 131-140. Wiley Online Library, 2014. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 593, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 287, + 657 + ], + "type": "text", + "content": "[24] Pablo Speciale, Martin R Oswald, Andrea Cohen, and Marc Pollefeys. A symmetry prior for convex variational 3d reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 313-328. Springer, 2016. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[25] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B Tenenbaum, and William T Freeman. Pix3d: Dataset and methods for single-image 3d shape modeling. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2974-2983, 2018. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[26] Junshu Tang, Zhijun Gong, Ran Yi, Yuan Xie, and Lizhuang Ma. Lake-net: Topology-aware point cloud completion by localizing aligned keypoints. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1726-1735, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 152 + ], + "type": "text", + "content": "[27] Demetri Terzopoulos, Andrew Witkin, and Michael Kass. Symmetry-seeking models and 3d object reconstruction. Int. J. Comput. Vis., 1(3):211-221, 1988. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[28] Sebastian Thrun and Ben Wegbreit. Shape from symmetry. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 2, pages 1824-1831. IEEE, 2005. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "type": "text", + "content": "[29] Shubham Tulsiani, Tinghui Zhou, Alexei A Efros, and Jitendra Malik. Multi-view supervision for single-view reconstruction via differentiable ray consistency. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2626-2634, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "type": "text", + "content": "[30] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martin-Martin, Cewu Lu, Li Fei-Fei, and Silvio Savarese. Densefusion: 6d object pose estimation by iterative dense fusion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3343–3352, 2019. 1, 2, 3, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "type": "text", + "content": "[31] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In *Eur. Conf. Comput. Vis.*, pages 52-67, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 344, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 344, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 344, + 545, + 387 + ], + "type": "text", + "content": "[32] Xiaogang Wang, Marcelo H Ang, and Gim Hee Lee. Voxel-based network for shape completion by leveraging edge generation. In Int. Conf. Comput. Vis., pages 13189-13198, 2021. 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "text", + "content": "[33] Xiaogang Wang, Marcelo H Ang Jr, and Gim Hee Lee. Cascaded refinement network for point cloud completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 790-799, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 434, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 477 + ], + "type": "text", + "content": "[34] Boyan Wei, Xianfeng Ye, Chengjiang Long, Zhenjun Du, Bangyu Li, Baocai Yin, and Xin Yang. Discriminative active learning for robotic grasping in cluttered scene. IEEE Robotics and Automation Letters, 8(3):1858-1865, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "type": "text", + "content": "[35] Jiajun Wu, Yifan Wang, Tianfan Xue, Xingyuan Sun, Bill Freeman, and Josh Tenenbaum. Marrnet: 3d shape reconstruction via 2.5 d sketches. Adv. Neural Inform. Process. Syst., 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "type": "text", + "content": "[36] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-10, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 580, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 545, + 633 + ], + "type": "text", + "content": "[37] Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. Snowflakenet: Point cloud completion by snowflake point deconvolution with skip-transformer. In Int. Conf. Comput. Vis., pages 5499-5509, 2021. 2, 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 635, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 690 + ], + "type": "text", + "content": "[38] Xin Yang, Yuanbo Wang, Yaru Wang, Baocai Yin, Qiang Zhang, Xiaopeng Wei, and Hongbo Fu. Active object reconstruction using a guided view planner. In Proceedings of International Joint Conference on Artificial Intelligence, pages 4965-4971, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[39] Yuan Yao, Nico Schertler, Enrique Rosales, Helge Rhodin, Leonid Sigal, and Alla Sheffer. Front2back: Single view 3d" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8905" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 354 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "shape reconstruction via front to back prediction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 531-540, 2020. 1, 2, 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[40] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In Int. Conf. Comput. Vis., pages 12498-12507, 2021. 1, 2, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "text", + "content": "[41] Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, and Martial Hebert. Pcn: Point completion network. In Int. Conf. 3D Vis., pages 728-737, 2018. 2, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 240 + ], + "type": "text", + "content": "[42] Zhaoxuan Zhang, Xiaoguang Han, Bo Dong, Tong Li, Baocai Yin, and Xin Yang. Point cloud scene completion with joint color and semantic estimation from single rgb-d image. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "type": "text", + "content": "[43] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3d: A modern library for 3d data processing. arXiv preprint arXiv:1801.09847, 2018. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "type": "text", + "content": "[44] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5745-5753, 2019. 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 319, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 354 + ], + "type": "text", + "content": "[45] Yichao Zhou, Shichen Liu, and Yi Ma. Nerd: neural 3d reflection symmetry detector. In IEEE Conf. Comput. Vis. Pattern Recog., pages 15940-15949, 2021. 1, 2, 8" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8906" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Single Image Deblurring with Row-dependent Blur Magnitude/05f13a7b-32f2-44f9-b585-5bd4c1beae6a_content_list.json b/2023/Single Image Deblurring with Row-dependent Blur Magnitude/05f13a7b-32f2-44f9-b585-5bd4c1beae6a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..680d1fe97bb81d5d2dd63538bf9b42540710d20f --- /dev/null +++ b/2023/Single Image Deblurring with Row-dependent Blur Magnitude/05f13a7b-32f2-44f9-b585-5bd4c1beae6a_content_list.json @@ -0,0 +1,1637 @@ +[ + { + "type": "text", + "text": "Single Image Deblurring with Row-dependent Blur Magnitude", + "text_level": 1, + "bbox": [ + 166, + 130, + 802, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiang Ji $^{1}$ Zhixiang Wang $^{1,2}$ Shin'ichi Satoh $^{2,1}$ Yinqiang Zheng $^{1\\dagger}$ $^{1}$ The University of Tokyo, Japan \n $^{2}$ National Institute of Informatics, Japan", + "bbox": [ + 186, + 179, + 781, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jixiang, wangzhixiang}@g.ecc.u-tokyo.ac.jp, satoh@nii.ac.jp, yqzheng@ai.u-tokyo.ac.jp", + "bbox": [ + 109, + 219, + 861, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image degradation often occurs during fast camera or object movements, regardless of the exposure modes: global shutter (GS) or rolling shutter (RS). Since these two exposure modes give rise to intrinsically different degradations, two restoration threads have been explored separately, i.e. motion deblurring of GS images and distortion correction of RS images, both of which are challenging restoration tasks, especially in the presence of a single input image. In this paper, we explore a novel in-between exposure mode, called global reset release (GRR) shutter, which produces GS-like blur but with row-dependent blur magnitude. We take advantage of this unique characteristic of GRR to explore the latent frames within a single image and restore a clear counterpart by only relying on these latent contexts. Specifically, we propose a residual spatially-compensated and spectrally-enhanced Transformer (RSS-T) block for row-dependent deblurring of a single GRR image. Its hierarchical positional encoding compensates global positional context of windows and enables order-awareness of the local pixel's position, along with a novel feed-forward network that simultaneously uses spatial and spectral information for gaining mixed global context. Extensive experimental results demonstrate that our method outperforms the state-of-the-art GS deblurring and RS correction methods on single GRR input.", + "bbox": [ + 76, + 301, + 473, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 680, + 209, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a fundamental research task of computer vision, image restoration has been explored for many years, aiming to recover a high-quality clean image from its corrupted counterpart by removing undesired degradation. Fast or even extreme motion, which is a major inducement to image degeneration and highly correlated to the camera's exposure mode: global shutter (GS) and rolling shutter (RS).", + "bbox": [ + 75, + 705, + 470, + 813 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In a GS image sensor, all pixels are reset simultaneously and immediately charged with exposure, then stored for sequential readout (Figure 1a). In the presence of fast and sudden motion, blurring effects will happen. [25, 39] depict", + "bbox": [ + 75, + 813, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "this motion-blurred output $\\mathbf{I}^B$ as an average of $N$ successive latent sharp frames $\\mathbf{I}_i^G$ :", + "bbox": [ + 498, + 268, + 890, + 301 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} ^ {B} = \\frac {1}{N} \\sum_ {i = 0} ^ {N - 1} \\mathbf {I} _ {i} ^ {G}, i = 0, 1, \\dots , N - 1.\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 308, + 830, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast, RS cameras expose image pixels row by row (including reset and readout), leading to an invariant time delay between consecutive scanlines (Figure 1b). Therefore, RS distortion effects, also known as jello effects, will appear if the camera is moving during the image acquisition, which is formulated in [23] as $[\\mathbf{I}^R]_i = [\\mathbf{I}_i^G]_i$ . $[\\mathbf{I}_i^G]_i$ is an operator that extracts the $i^{th}$ row from the $i^{th}$ latent GS frame $\\mathbf{I}_i^G$ and $N$ is also the height of RS image.", + "bbox": [ + 496, + 357, + 893, + 477 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although image deblurring itself is highly involved, it is believed that restoring clear images from single degraded captures of GS cameras is much more tractable than that of RS cameras as mentioned in [23]. Intuitively, motion blur removal process aims to disentangle target GS frame from redundant input without sophisticated displacement estimation or speculation of unknown context. On the contrary, RS correction needs to shift pixels of each scanline to one virtual GS canvas by building pixel-level correspondences. Moreover, the rectified output usually suffers from missing boundaries not directly captured during RS exposure. Methodologically, the coarse-to-fine network design has gained remarkable performance for single image deblurring [25, 30, 33, 45, 10, 46]. On the contrary, due to the ill-posed nature, single image RS correction relies heavily on strong prior assumptions, explicitly [35, 17, 32, 31] or implicitly [34, 58], which limits their applicability to real scenarios. So, recent research has moved onto multi-image RS correction to bypass the ill-posedness. But both classical [57, 1] and learning-based methods [23, 9] still cannot work well under complex dynamic scenes or drastic camera motions because of nontrivial pixel alignment.", + "bbox": [ + 496, + 478, + 895, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "It is largely overlooked in computer vision that there exists an in-between shutter mode, called global reset release (GRR) shutter, which can be easily enabled in some image sensors that originally work in GS mode (e.g., EV76C560 in EO-1312C Camera) or in RS mode (e.g., IMX178 in BFS-U3-63S4C camera). As shown in Figure 1c, GRR", + "bbox": [ + 496, + 810, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12269", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/251c803ab291b0c604ee2c1c78c148e794b9b3724baefc173a263118d1ff1b0f.jpg", + "image_caption": [ + "Figure 1: Different exposure modes and corresponding degrading processes induced by fast motions. (a) GS exposes all pixels simultaneously, and it causes the output with row-independent blur that can be interpreted as an average of the latent sharp frames. (b) RS captures pixels scanline by scanline, featured in higher frame rate, and lower cost than GS. However, RS sensors are prone to row-dependent distortion, which can be more challenging to correct than deblurring. (c) GRR begins exposure of all pixels simultaneously but ends row by row, causing blurring effects with row-dependent blur magnitude. A specialized deblurring task is required but without the need for RS distortion correction. GRR also has the advantage of reducing the flickering effect of artificial illuminants, such as green traffic lights." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 893, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "starts to expose all pixels simultaneously as GS does but reads out signals scanline by scanline like conventional RS does. Apparently, GRR leads to blurring effects. However, different from GS blur, the blur magnitude of GRR (the brightness level and Signal-Noise-Ratio as a consequence) is row-dependent. Very recently, Wang et al. [47] proposed to enable this global reset feature in RS sensors that turns RS correction into a more tractable deblurring problem and reported superior results of video deblurring in GRR mode. We note the cost to pay for this performance gain by switching from RS to GRR mode includes 1) the superiority of RS mode over GS mode in capture speed is sacrificed since GRR does not allow early scanning of the next frame before the end of the current frame; 2) scanning rows in the bottom are more likely to be saturated, and the standard automatic exposure mechanism should be adjusted accordingly. Therefore, in the presence of multiple consecutive frames (e.g. 8 frames are used in [47]), the authentic benefits of using GRR mode are slightly questionable.", + "bbox": [ + 75, + 516, + 472, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we deal with a more challenging scenario with a single input image, which is practically meaningful in avoiding well-known flaws of using consecutive frames, including 1) Larger buffer capacity is needed; 2) Sensitivity to video recording settings, like frame rate and deadtime between frames, leading to additional generalization issues;", + "bbox": [ + 75, + 810, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3) Temporal alignment with severe degradation is still challenging.", + "bbox": [ + 498, + 516, + 893, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By fully considering its unique exposure characteristic of GRR mode, we bring single GRR image deblurring into the community. Mathematically, row-dependent blur and brightness caused by varying exposure time of different scanlines could be formulated as:", + "bbox": [ + 496, + 547, + 893, + 622 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} ^ {R G} = \\mathbf {I} _ {0} ^ {G} + \\delta \\sum_ {i = 1} ^ {N - 1} \\left[ \\mathbf {I} _ {i} ^ {G} \\right] _ {i:},\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 632, + 782, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $[\\mathbf{I}_i^G ]_i$ is an operator that extracts image patch of $i^{th}$ row to the end from the $i^{th}$ latent GS frame $\\mathbf{I}_i^G$ and $\\delta$ denotes the ratio between readout time and the first scanline's exposure duration. As a result, directly applying existing GS deblurring algorithms on GRR deblurring may give rise to an adaptation issue. On the other hand, Transformers have shown significant performance gains on image restoration task [22, 50] by mitigating the shortcomings of CNNs (i.e., limited receptive field and content-independent kernel). Although shifted widow strategy [22, 50, 46, 8, 24] has largely reduced computational loads and made it possible to process high dimensional input, the introduced severe corruption to image spatial information has not been addressed, especially for spatially-sensitive tasks (e.g., GRR deblurring). Besides,", + "bbox": [ + 496, + 686, + 895, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12270", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "according to the spectral convolution theorem [14], updating a single value in the spectral domain globally affects all original data, which has been implemented for non-local receptive field and proven to be effective in capturing long-range context from frequency domain [4, 5, 55].", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to leverage the capability of self-attention and U-net [36] structure with multi-scale input/output to tackle the single GRR deblurring task. To this end, we propose shifted horizontal window-based Transformer block with spatial compensation and spectral enhancement, which consists of three main components: 1) Shifted Window-based Multi-head Self-Attention(SW-MSA) with horizontal partition strategy and hierarchical positional encoding. We update original squared window as rectangular one to better adapt row-dependent blur and brightness. Meanwhile, hierarchical positional encoding compensates global positional context of windows lost in window partition and enables order-awareness of local pixel's position within each window. 2) Spectrally-enhanced Feed-Forward Network (SeFFN) based on depth-wise convolutional layer. The spectral branch captures the discrepancy between blurry and sharp image pairs from the perspective of frequency domain, as a complementary part of self-attention. 3) Cross-scale Feature Fusion module based on Squeeze and Excitation block (CFFSE) [11], which enables us to actively emphasize or suppress the features from encoders with different scales. The main contributions of this paper are summarized as follows:", + "bbox": [ + 75, + 167, + 473, + 500 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an original Transformer block with spatial compensation and spectral enhancement to address single GRR deblurring by fully exploring the latent frames, which further validates the advantages of GRR exposure mode over its RS counterpart.", + "- Horizontal partition strategy and hierarchical positional encoding are used to compensate global positional context of windows and enable order-awareness of local pixel's position within each window. The spectrally-enhanced feed-forward network simultaneously uses spatial and spectral information for gaining mixed global context.", + "- To facilitate the development and evaluation of GRR deblurring, we take paired GRR/Sharp images to offer a new dataset captured under real scenes named GRR-real. Furthermore, by simulating data captured by three shutter modes in a comparable way, the benefits of using GRR for single image restoration are verified." + ], + "bbox": [ + 96, + 507, + 470, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 810, + 218, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deep Image Deblurring Deep learning methods have achieved significant success in multi [12, 26, 41, 54, 20, 41] or single [15, 18, 25, 27, 40, 45, 48, 51, 52, 44, 19, 29] image deblurring. Initially, researchers try to find the spatially", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "varying kernels of motion blur before estimating latent image and no exception for deep CNN methods [37, 43]. However, blur kernel is normally computed for all pixels leading to huge demands for memory and computation. Besides, kernel estimation process is overly sensitive to noise and saturation, which is not practical for real scenes. Later, DeepDeblur [25] firstly exploits kernel-free learning to directly construct the relation between blurry and sharp images in an end-to-end manner. And the coarse-to-fine strategy is also taken by following SRN [45] uses an encoder-decoder network with skip-connections for three scale levels to significantly reduce training difficulty and introduce obvious stability benefits. Adversarial training has also been extensively studied [15, 16]. Recently, Cho et al. [6] revisit the coarse-to-fine scheme and present a novel deblurring network (MIMO U-net) that can handle multi-scale blur efficiently. These methods have been proven to be effective and achieved remarkable performance in single image deblurring.", + "bbox": [ + 496, + 90, + 893, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rolling Shutter Correction RS correction methods with single image input could be divided into classical and learning based. Classical approaches heavily rely on strong prior assumptions, such as the scene is static, and the movement of the camera is limited to pure rotation or in-plane translation [35, 17, 32, 31]. Rengarajan et al. [35] propose to estimate motion by converting transformed curves to be straight based on the assumption that \"straight lines must remain straight\" and purely rotation of camera. [32] simplifies the real scene as Manhattan world to rectify the monocular RS image. These classical methods are barely applicable in real situations because of the restrictive prior assumptions. For learning-based methods, Rengarajan et al. [34] proposed a new CNN architecture based on long rectangular kernels to correct rolling shutter distortions by simply modeling the camera motion as translation+rotation polynomials. Zhuang et al. [58] extend [34] for learning to predict both the camera velocity and depth from a single RS image. The global shutter image is then recovered as a post-processing step. Recent research mainly focuses on multi-image RS correction [57, 1, 23, 9]. But those methods still cannot work under complex dynamic scenes or large camera motions.", + "bbox": [ + 496, + 382, + 895, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision Transformers As an alternative to CNN, the Transformer model has been adapted to numerous vision tasks such as image restoration [22, 50, 46], segmentation [49, 53], objection detection [3, 24]. The pioneering work of ViT [8] directly conducts self-attention on flattened patches by decomposing an image into a sequence of tokens and gets excellent results on image classification. To further overcome the quadratic complexity of original self-attention, Liu et al. [24] perform local windows with fixed size or window shift to help cross-window interaction. Liang et al. proposes an image restoration model, SwinIR [22] based on Swin Transformer [24]. Similarly, Uformer [46] imple", + "bbox": [ + 496, + 719, + 895, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12271", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6ff24b9af7a1cb0abb7690de950a025cd5e89d5ba223b9ef817388b8be218f30.jpg", + "image_caption": [], + "image_footnote": [ + "Element-wise multiplication" + ], + "bbox": [ + 117, + 88, + 851, + 297 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/66e13100157fffbddcaf4005c140bf24866040d93745662377a6a58b4e0f64a5.jpg", + "image_caption": [ + "Figure 2: Overview of our model architecture. The model is implemented in an encoder-decoder manner with multi-scale input and output. Each encoder/decoder (except for $\\mathsf{EB}_0$ and $\\mathsf{EB}_3$ ) consists of (a) specially designed residual spatially-compensated and spectrally-enhanced transformer block (RSS-T); (b) feature attention module (FAM)/convolution layer (Conv.). Cross-scale feature fusion based on squeeze and excitation block (CFF-SE) is explored to actively emphasize or suppress the features from encoders." + ], + "image_footnote": [], + "bbox": [ + 117, + 309, + 385, + 422 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/924e59d9634ca8d320f780f928c8435b1cce5ed6516d69f8586312bb8c14b2da.jpg", + "image_caption": [], + "image_footnote": [ + "Upsample", + "Downsample", + "$\\oplus$ Element-wise addition", + "$\\mathbb{C}$ Concatenation", + "In Out Input / Output projection" + ], + "bbox": [ + 398, + 309, + 651, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ments the Locally enhanced Window Transformer block in Unet structure and introduces a learnable multi-scale restoration modulator. All these methods have gained remarkable performance on image restoration tasks by mitigating the shortcomings of CNNs. But no one has addressed the corruption to image spatial information introduced by the widow partition strategy. And spectral feature, which is proven to capture long-range context effectively, has also been ignored in vision transformers.", + "bbox": [ + 75, + 537, + 472, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 695, + 168, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview The overall structure is in the manner of encoder-decoder with multi-scale inputs and outputs (Figure 2). Except for $\\mathbb{E}\\mathbb{B}_0$ and $\\mathbb{D}\\mathbb{B}_3$ , all encoder/decoder blocks contain a Feature Attention Module (FAM) or Convolution layer (Conv.) to refine aggregated features. Cross-scale feature fusion based on squeeze and excitation block (CFF-SE) is exploited to propagate information flow between encoders and decoders. To be specific, given a degraded image $\\mathbf{I} \\in \\mathbb{R}^{3 \\times H \\times W}$ , multi-scale input set $\\mathcal{I} = [\\mathbf{I}_0, \\mathbf{I}_1, \\mathbf{I}_2, \\mathbf{I}_3]$ , $\\mathbf{I}_k \\in \\mathbb{R}^{3 \\times \\frac{H}{2^k} \\times \\frac{W}{2^k}}$ ( $k = 0, 1, 2, 3$ ) is obtained by down-sampling operations. Then low-level feature $\\mathbf{X}_k \\in \\mathbb{R}^{2^k C \\times \\frac{H}{2^k} \\times \\frac{W}{2^k}}$ of", + "bbox": [ + 75, + 728, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "input $\\mathbf{I}_k$ is extracted by its own input projection module $In^k$ and will be fed into $\\mathbb{E}\\mathbb{B}_k$ . Similarly, the first three encoder blocks are followed by a down-sampling operation to guarantee the shape of feature map from $\\mathbb{E}\\mathbb{B}_{k - 1}$ is consistent with that of $\\mathbf{X}_k$ . For feature reconstruction, an up-sampling operation is inserted between each two decoders. And multi-scale output set $\\mathcal{O} = [\\hat{\\mathbf{I}}_0,\\hat{\\mathbf{I}}_1,\\hat{\\mathbf{I}}_2,\\hat{\\mathbf{I}}_3],\\hat{\\mathbf{I}}_l\\in \\mathbb{R}^{3\\times \\frac{H}{2^l}\\times \\frac{W}{2^l}}(l = 0,1,2,3)$ is generated by corresponding output projection module $Out^l$ and residual connection from input image $\\mathbf{I}_l$ . Figure 2 illustrates the detailed structure of our RSS-T block.", + "bbox": [ + 496, + 537, + 893, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. SW-MSA with Spatial Compensation", + "text_level": 1, + "bbox": [ + 500, + 696, + 821, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Two main challenges exist to directly extend the existing Transformer architecture to GRR deblurring. First, because of the characteristic of GRR exposure mode, generated spatial-variant blur and brightness are closely correlated to image rows, while normal Transformer blocks do not consider this. They divide the image into squared window [22, 50, 46], leading to large blur and brightness variance within each window, which makes the rectification process more challenging. Second, standard vision Transformer architecture cuts the computational cost due to the usage of self-attention conducted on non-overlapping windows, which enables high dimensional input processing.", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12272", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cd6bbc3504fe3fba787a40016c6536058cd29c2ef3fe42db9cadfc515f76a9dd.jpg", + "image_caption": [ + "Figure 3: Two main components of our proposed RSS-T. (a) The proposed SW-MSA with Spatial Compensation includes two core designs: horizontal window partition strategy and hierarchical position encoding comprising window positional encoding (WinPE) and pixel positional encoding (PixPE). (b) Se-FFN." + ], + "image_footnote": [], + "bbox": [ + 112, + 90, + 617, + 287 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a60eccb0ea8b4f41f4bf71a4e71a8d961c6f3e5f7947822ccf1c9782bf83bf4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 90, + 848, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "But during window partition, image spatial information has been corrupted. That absolute spatial location is crucial for spatially-sensitive tasks. To address the aforementioned issues, we propose SW-MSA with spatial compensation, as shown in Figure 3a, which has two core designs: horizontal window partition and hierarchical positional encoding.", + "bbox": [ + 75, + 368, + 470, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Horizontal Window Partition As discussed above, common squared windows without distinguishing blur and brightness variance along the row and column dimensions make rectification more challenging. We update the original squared window as rectangular to better adapt row-dependent blur and brightness. These long and narrow horizontal windows will mitigate large variances within each widow. Bhojanapalli et al. [2] has depicted a kind of low-rank bottleneck, i.e., while increasing the number of heads seemingly gives the model more expressive power, at the same time, the head size is reduced, which can decrease the expressive power. When input token number $n$ is larger than head size $d_h$ , it will create a low-rank bottleneck and lose its ability to represent arbitrary context vectors. Thus, by following this rule and considering computational cost, we set the horizontal window size as (2, 16), (4, 16), (2, 32), and (4, 32). For better comparison, we also provide a normal squared window size of (8, 8) and an extreme case with a window size of (2, full). The 'full' means the window width equals the feature map width. Given input feature $\\mathbf{X}_{in} \\in \\mathbb{R}^{B \\times H \\times W \\times C'}$ , the partition process is represented as:", + "bbox": [ + 75, + 464, + 473, + 781 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {w} = \\operatorname {W i n P a r t} \\left(\\mathbf {X} _ {i n}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 792, + 468, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{X}_w\\in \\mathbb{R}^{BN_w\\times H_w\\times W_w\\times C'}$ $H_{w},W_{w}$ denote height and width of windows and $N_{w} = H / H_{w}\\times W / W_{w}$ , which is the total number of divided windows.", + "bbox": [ + 76, + 818, + 468, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hierarchical Positional Encoding Window partition largely reduces computational cost but corrupts the image's", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "global positional information crucial to GRR deblurring. So, we propose hierarchical positional encoding: 1) Window Position Encoding (WinPE) compensates for the absolute global location lost in the partition process. 2) Pixel Position Encoding (PixPE) enables context of the local pixel's relative position within each window. Chu et al. [7] proposes a positional encoding generator to obtain Conditional Position Encodings (CPE), which can be efficiently implemented with a 2D depth convolution with kernel $k$ ( $k \\geq 3$ ) and $\\frac{k - 1}{2}$ zero paddings. They have proven zero paddings here are important to make the model aware of the absolute positions of each token. Li et al. [21] use a similar positional encoding strategy to capture temporal position for videos. Here, we extend CPE to the 3D situation and propose our WinPE:", + "bbox": [ + 496, + 368, + 893, + 580 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {W i n P E} \\left(\\mathcal {R} \\left(\\mathbf {X} _ {w}\\right)\\right) = \\operatorname {D W C o n v} \\left(\\mathcal {R} \\left(\\mathbf {X} _ {w}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 589, + 890, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{R}(\\cdot)$ is reshape operation and DWConv means 3D depthwise convolution with zero paddings. Each window's absolute position will be encoded by repeatedly applying depthwise convolution to the 3D feature space. Moreover, this encoding is dynamic and allows us to eliminate the constraint to the token's length when testing on different datasets, which perfectly matches the requirement of our window position encoding. The fixed size of the window will lead to the dynamic length of the window sequence once the image size changes. We also apply the relative position encoding [24, 38] for PixPE through a learnable parameter table $\\mathcal{B} \\in \\mathbb{R}^{(2H_w - 1) \\times (2W_w - 1)}$ . Overall, the formulation is:", + "bbox": [ + 496, + 614, + 893, + 796 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} = \\mathbf {X} _ {i n} + \\operatorname {M S A} \\left(\\ln \\left(\\mathbf {X} _ {e}\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 804, + 805, + 820 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {e} = \\mathbf {X} _ {w} + \\mathcal {R} (\\operatorname {D W C o n v} (\\mathcal {R} (\\mathbf {X} _ {w}))), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 823, + 890, + 840 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M S A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t M a x} \\left(\\mathbf {Q K} ^ {T} / \\sqrt {d _ {h}} + \\mathbf {B}\\right) \\mathbf {V},\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 844, + 848, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{Q}$ , $\\mathbf{K}$ and $\\mathbf{V}$ are projected query, key and value from input. $\\mathbf{B}$ is the relative position bias, whose values are taken", + "bbox": [ + 496, + 869, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12273", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "from $\\mathcal{B}$ . $\\mathrm{LN}(\\cdot)$ means layer norm and $d_h$ denotes head size.", + "bbox": [ + 76, + 90, + 468, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. Spectrally-enhanced FFN", + "text_level": 1, + "bbox": [ + 76, + 125, + 308, + 142 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Researchers have proven the spectral feature effectively captures long-range dependencies. The Fast Fourier Convolution (FFC) operator in [4] could efficiently implement a non-local receptive field. Inspired by this work, we devised our Se-FFN based on depth-wise convolutional layer. The spectral branch captures the discrepancy between blurry and sharp image pairs as a supplemental part of self-attention for effective long-range dependency modeling.", + "bbox": [ + 75, + 151, + 470, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Se-FFN includes two parts: spectral branch and spatial branch (Figure 3b). The spatial part first applies layer norm to the input feature followed by two convolutions with $1 \\times 1$ kernel and a $3 \\times 3$ depth-wise convolution to capture local information. The spectral branch is formulated as (a) convert the input feature map $\\mathbf{Y} \\in \\mathbb{R}^{B \\times H \\times W \\times C'}$ into the frequency domain to obtain $\\mathcal{F}(\\mathbf{Y})$ . (b) Concatenate the real part and imaginary part of $\\mathcal{F}(\\mathbf{Y})$ along channel dimension to get the spectral feature $\\tilde{\\mathbf{Y}} \\in \\mathbb{R}^{B \\times H \\times W/2 \\times 2C'}$ . (c) Exploit two $1 \\times 1$ convolutional layers with the LeakyReLU function in between to process the spectral feature. (d) Split the processed spectral feature as real and imaginary parts, then convert them back to spatial domain $\\mathbf{Y}_f \\in \\mathbb{R}^{B \\times H \\times W \\times C'}$ by inverse Fast Fourier Transform (iFFT). Finally, the output of Se-FFN is $\\mathbf{Z} = \\mathbf{Y} + \\mathbf{Y}_f + \\mathbf{Y}_s$ , where $\\mathbf{Y}_s$ is the output of spatial branch.", + "bbox": [ + 75, + 275, + 472, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Cross-scale Feature Fusion", + "text_level": 1, + "bbox": [ + 76, + 536, + 320, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In most conventional U-net structures, the corresponding encoder and decoder are related by a skip connection to allow information flow between the same scale level. Later in [6], authors present an Asymmetric Feature Fusion (AFF) module to enable cross-scale interaction but without any control. So, we present our CFF-SE block. As shown in Figure 2, before the convolutional block, the input firstly goes through a SE module to control the information flow from the respective scale level of all encoders, which enables us to actively emphasize or suppress the features. Thus, each decoder can selectively exploit multi-scale features, resulting in improved deblurring performance as will be demonstrated in the experimental part.", + "bbox": [ + 75, + 561, + 468, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 781, + 209, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Learning-based methods for deblurring or RS correction are usually trained on synthetic data, which has limitations of generalization. We, therefore, built an image acquisition system to collect a real dataset named GRR-real. The details about the system, training implementation, and more experimental results are offered in supplemental materials.", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison on GRR Deblurring", + "text_level": 1, + "bbox": [ + 500, + 90, + 790, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with Other Models Since single GRR deblurring is addressed for the first time, there is no specially devised algorithm. So, we compare our model with previous state-of-the-art approaches from:", + "bbox": [ + 498, + 117, + 893, + 179 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- GS image-based deblurring by exploiting coarse-to-fine strategy: basic U-net [36], SRN [45], MIMO-UNet[6] and Uformer[46].", + "- GS video-based deblurring by fusing neighboring temporal information: STR-CNN [12], DBN [42] and IFIRNN[26].", + "- RS video-based correction through predicting pixelwise displacement: JCD [56] and DSUR [23].1", + "- GRR video deblurring using a spatial-aware encoder with long-short term temporal information aggregator: NGS [47]. For fair comparisons, we not only provide the original setting with 8 frames as input $(\\mathrm{NGS}_8)$ but also update the model with a single input $(\\mathrm{NGS}_1)$ ." + ], + "bbox": [ + 517, + 189, + 893, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After retraining all models, Table 1 depicts quantitative results on the GRR-real test dataset. Overall, it demonstrates our specially devised algorithm significantly outperforms models from RS correction and GS deblurring, even though they take multi frames as input and exploit temporal information. As discussed in Section 1, RS correction usually resorts to nontrivial motion estimation and warping process. When directly used for GRR, they performed worse than deblurring models. Furthermore, we observe that, for deblurring models, those with video clips as input may degrade sharply, even are no match for some approaches of a single input. A key reason is that these models are struggling to align the content of different video frames. If failed, the redundant inputs would confuse learning process instead.", + "bbox": [ + 496, + 421, + 893, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although the tailored GRR video deblurring method $\\mathrm{NGS}_8$ [47] achieved higher metrics than ours (30.03/0.90 vs. 28.64/0.90), it heavily relies on temporal correlation of 8 frames and the performance decreased drastically when given single input $(\\mathrm{NGS}_1)$ . In contrast to RSS-T, the demand of $\\mathrm{NGS}_8$ for 8 consecutive frames in training or testing can be difficult to be satisfied, sometimes even impossible due to buffer restrictions. Besides, sensitivity to video capturing settings (e.g., frame rate and deadtime between frames) will lead to additional generalization issues. We also computed the complexity of all algorithms, which shows that our model is at medium level and outperforms SOTA RS correction or Transformer-based models. From qualitative", + "bbox": [ + 496, + 633, + 893, + 829 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Noting that single RS image correction methods either heavily rely on strong assumption [35, 17, 32, 31], which is apparently inapplicable to our natural scene dataset or require actual velocity or depth to train the model [34, 58] that are barely accessible in common cameras. So, we ignore them and only choose RS video correction methods to compare.", + "bbox": [ + 496, + 838, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12274", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6c07a62771f7d08e2a733bdcf45ea6ff87c24c889ccbb90d3f27042dc3a1eb3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 87, + 251, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/29abe3ac19cc63e5ff08bcd78ef5972a8fdcf5ad07d9dc6970ce8cb3859da507.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 88, + 406, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/20b2157bdc87a47c7cd87749cd790a5f7bf0d966f8f0a77750768a3e4783d90c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 88, + 562, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a4a7996a66ac7811165c1390bfdf59d4308b823d082138114dbb930c21604fbe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 88, + 717, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c3f4b062b1a6841df49321bf3609fd00d439a1ce77d2074476673903617a8423.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 88, + 874, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/865db4c06b59df0feede0bf35d7300e432c04b6052c71c99c6351e68d3b74cfb.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 93, + 193, + 250, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/422980c6580793f25e092ddd3499328134aacf6cbaefc02abf15057b5a463cdf.jpg", + "image_caption": [ + "U-net [36]" + ], + "image_footnote": [], + "bbox": [ + 251, + 193, + 406, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2513361c4ace0648e3623b5bed443956abd835b42ad3648a80c2b4ec45afc879.jpg", + "image_caption": [ + "SRN [45]" + ], + "image_footnote": [], + "bbox": [ + 406, + 193, + 562, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78b567243fa9597b7cd3a856cd4b05ff029977f7fa04ee5887b05f48fb4c7e80.jpg", + "image_caption": [ + "MIMO [6]" + ], + "image_footnote": [], + "bbox": [ + 562, + 193, + 717, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5b62b503d635ea90672aab6abd95d0326b91a5d968451e7ec5be2abe477f4a8.jpg", + "image_caption": [ + "MIMO+ [6]" + ], + "image_footnote": [], + "bbox": [ + 718, + 193, + 874, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ea96ddbaf44480c939644e3ca01ea9f34612c10e53f972d5bc05ae3192ddf076.jpg", + "image_caption": [ + "Uformer [46]", + "DSUR [23]", + "Figure 4: Qualitative results on the GRR-real dataset. We compare our method with state-of-the-art methods from RS correction, GS deblurring, and GRR video deblurring." + ], + "image_footnote": [], + "bbox": [ + 93, + 297, + 250, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/167ae600df2a4a4ab20cfbb7f443d24d05712ec1328b9f6e4a8f1573ec5777d4.jpg", + "image_caption": [ + "STR-CNN [12]", + "NGS1 [47]" + ], + "image_footnote": [], + "bbox": [ + 251, + 297, + 406, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ba9069f537e5479b8e23072ec87a3029b7e1ae742d9010c1534dd17ea6a7beef.jpg", + "image_caption": [ + "DBN [42]", + "NGS8 [47]" + ], + "image_footnote": [], + "bbox": [ + 406, + 297, + 562, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b1da72b802b6b386806df53af1d37082283917e12fb881276f37a3a66c7dc0f6.jpg", + "image_caption": [ + "IFIRNN [26]", + "RSS-T" + ], + "image_footnote": [], + "bbox": [ + 562, + 297, + 717, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5238d6ea6659bdb5d269380b0d8c5c8482c09df658a514805db104c79c20b2df.jpg", + "image_caption": [ + "JCD [56]", + "GT" + ], + "image_footnote": [], + "bbox": [ + 718, + 297, + 874, + 388 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/79274a1d5df776e15f56fa28cfa9fc86ccb10acb4ece038cb4477e6f8947105d.jpg", + "table_caption": [ + "Table 1: Quantitative comparison on GRR-real dataset. We compare our method with state-of-the-art motion deblurring, RS correction, and GRR video deblurring approaches. According to 'exposure mode/Input number', they could be divided as 'GS/Single', 'GS/Multi', 'RS/Multi', 'GRR/Multi', and 'GRR/single'. The performance is measured with mean PSNR/SSIM (↑). 'Full' denotes evaluation using full-size frames, while 'Top', 'Middle', and 'Bottom' represent using only a patch of images from the top, middle, and bottom areas." + ], + "table_footnote": [], + "table_body": "
MethodMode / InputEffectivenessEfficiency
TopMiddleBottomFullTime (s)Params (M)FLOPs (T)
Input-15.12 / 0.6721.83 / 0.7820.08 / 0.7617.61 / 0.74---
U-net [36]26.34 / 0.9223.83 / 0.8522.74 / 0.8323.57 / 0.870.00629.500.072
SRN [45]25.11 / 0.7824.34 / 0.7224.08 / 0.7223.84 / 0.740.014010.250.509
MIMO [6]GS / Single27.16 /0.9024.55 / 0.8224.04 / 0.7924.48 / 0.840.01236.810.315
MIMO+ [6]28.46 / 0.9126.31 / 0.8525.78 / 0.8326.12 / 0.860.024916.110.724
Uformer [46]25.04 / 0.9323.70 / 0.8622.20 / 0.8422.92 / 0.870.214720.630.257
STR-CNN [12]20.88 / 0.8323.28 / 0.7722.60 / 0.7521.39 / 0.780.01940.930.367
DBN [42]GS / Multi25.11 / 0.9025.90 / 0.8525.87 / 0.8224.97 / 0.850.022915.311.046
IFIRNN [26]27.19 / 0.9023.85 / 0.7823.21 / 0.7723.97 / 0.810.01351.640.581
JCD [56]RS / Multi27.64 / 0.9023.65 / 0.8019.76 / 0.7722.31 / 0.820.26258.670.326
DSUR [23]24.81 / 0.8723.87 / 0.7823.39 / 0.7623.35 / 0.800.30183.900.225
NGS8 [47]GRR / Multi31.71 / 0.9330.54 / 0.9029.12 / 0.8730.03 / 0.900.02338.671.266
NGS1 [47]GRR / Single26.61 / 0.8923.53 / 0.7822.83 / 0.7623.67 / 0.810.01874.560.083
RSS-T30.90 / 0.9328.60 / 0.8827.86 / 0.8628.64 / 0.900.147911.340.176
", + "bbox": [ + 102, + 547, + 862, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "results in Figure 4, although the images obtained by the existing networks exhibit mitigated distortions compared to the input, local details and structures were not sufficiently corrected, whereas our method produces sharper images that even visually outperform $\\mathrm{NGS}_8$ .", + "bbox": [ + 75, + 800, + 470, + 876 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Thirdparty Evaluation To further prove the practical value", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of our method, we evaluate it on another GRR dataset from [47]. As shown in Table 2, our method still outperforms other algorithms and is just slightly inferior to $\\mathrm{NGS}_8$ with as many as eight input frames (26.62/0.86 vs. 27.29/0.85).", + "bbox": [ + 498, + 800, + 890, + 861 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12275", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/754467b461aab5ae8977e3a20818dcce2433bcee8ce57ed1b567751d53548ed3.jpg", + "table_caption": [ + "Table 2: Comparison on another GRR dataset [47]." + ], + "table_footnote": [], + "table_body": "
MethodSET-II
FullTopMiddleBottom
Input17.82 / 0.7323.64 / 0.7721.45 / 0.7715.54 / 0.66
SRN [45]25.05 / 0.8124.32 / 0.7925.65 / 0.8127.02 / 0.83
STRCNN [12]22.59 / 0.8122.99 / 0.7923.46 / 0.8123.66 / 0.83
DBN [42]22.57 / 0.8123.24 / 0.8023.81 / 0.8123.24 / 0.82
IFIRNN[26]25.17 / 0.8224.77 / 0.8025.62 / 0.8126.94 / 0.84
ESTRNN [54]22.72 / 0.8323.42 / 0.8126.03 / 0.8322.86 / 0.83
DSUR [23]22.50 / 0.8022.49 / 0.7823.87 / 0.8123.38 / 0.83
JCD [56]25.33 / 0.8024.77 / 0.7825.71 / 0.8027.43 / 0.83
NGS8 [47]27.29 / 0.8526.96 / 0.8427.57 / 0.8528.35 / 0.86
NGS1 [47]25.19 / 0.7924.48 / 0.7725.53 / 0.7927.27 / 0.82
RSS-T26.62 / 0.8625.90 / 0.8426.82 / 0.8628.43 / 0.87
", + "bbox": [ + 81, + 112, + 465, + 301 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ec14846d8dd4fa17ebd46db42075cc57b70deb2a1c74492d62cf7d8abbe9988b.jpg", + "table_caption": [ + "Table 3: Architecture ablation on our GRR-real dataset." + ], + "table_footnote": [], + "table_body": "
PEFFNFusionPSNR / SSIM
WinPEPixPELeFNNSe-FNNAFFCFF-SE
v1✓✓25.69 / 0.85
v2✓✓✓26.50 / 0.86
v3✓✓✓26.10 / 0.86
v4✓✓✓✓26.83 / 0.87
v5✓✓✓✓27.24 / 0.87
RSS-T✓✓✓✓28.64 / 0.90
", + "bbox": [ + 81, + 345, + 465, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 498, + 228, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Model Architecture Ablation We conducted experiments to analyze the effectiveness of the core components of our RSS-T. The hierarchical position encoding consists of WinPE and PixPE. The feed-forward network is also implemented by LeFFN [46] to compare with our Se-FNN (v5 vs. RSS-T). As for the cross-scale fusion part, AFF [46] and our CFF-SE are used to explore the performance difference (v4 vs. v5). All the experimental results are listed in Table 3. From v1, v2, v3, and v4, when WinPE and PixPE are separately presented in baseline model v1, the performance goes up but the combination of them achieved the best metrics. Our Se-FNN and CFF-SE modules contributed to the further performance improvement of PSNR by $+1.81\\mathrm{dB}$ compared with the LeFFN and AFF.", + "bbox": [ + 75, + 530, + 470, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Window Size Ablation We introduced three types of window size (i.e., horizontal, squared, and full-width window) and clarified the advantages of the horizontal window for GRR deblurring in Section 3.1. Here, we present experimental results to support the claim. As Figure 5 shows, all horizontal window settings are superior to the standard squared one from quantitative and qualitative perspectives. The extreme case (2, full) is severely trapped in low-rank bottleneck [2] as the pixel number of this window is far more than the head size.", + "bbox": [ + 75, + 750, + 472, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/28cc50ef74d74788349acbfa0f3b6a480cb7e500e5b0d2ff2e6a5989b3cc1d23.jpg", + "image_caption": [ + "Figure 5: Window size ablation. Experimental results of all different window-size settings on our GRR-real dataset." + ], + "image_footnote": [], + "bbox": [ + 504, + 87, + 888, + 325 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e26407c8df00d9f090c6997474d827bf45a94f127e1aaa19fc34b0c497ecaa5.jpg", + "image_caption": [ + "Figure 6: Temporal alignment of three shutters. When synthesizing datasets, we strictly align the exposure of three modes." + ], + "image_footnote": [], + "bbox": [ + 506, + 383, + 885, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Comparison on Three Shutter Modes", + "text_level": 1, + "bbox": [ + 500, + 559, + 821, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Until now, a direct comparison of three shutter modes is still missing, although [47] has demonstrated that GRR has its superiority against RS in obtaining sharp images in the scenario of video restoration. Here, we evaluate them on single image Restoration. For a fair comparison, we synthesize strictly aligned images virtually captured by three modes (RS, GRR, and blur) for each scene by following the protocol of [28, 45, 6] (details in the supplemental material), and compare with SOTA methods for each mode.", + "bbox": [ + 496, + 582, + 890, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When compared with RS mode, Table 4 shows that GRR combined with our RSS-T model has gained significant advantages in obtaining single sharp image, even though DSUR and JCD take multiple frames as input. These results are consistent with the discussion in [47] that the GRR mode enables us to convert the RS correction problem into a more tractable one. Therefore, given an RS camera that supports GRR mode, we recommend switching to GRR mode for better restoration from a single input, if losing some capture speed of the original RS is permissible. Also, the exposure should be adjusted to avoid saturation of bottom rows.", + "bbox": [ + 496, + 718, + 892, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We additionally present comparisons with row-", + "bbox": [ + 519, + 885, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12276", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6883381720174d4adc04700af91a153cd65293181ff32ab9c84bc1c6ed6587c4.jpg", + "table_caption": [ + "Table 4: Quantitative comparison on three modes. The subscript of each method denotes the number of input frames. The reason we choose multi-frame RS correction is discussed in Section 4.1. MIMO+ and Uformer are trained on blur mode. DSUR and JCD are trained on the RS counterpart, while RSS-T is trained on GRR mode." + ], + "table_footnote": [], + "table_body": "
MethodFullTopMiddleBottom
DSUR2 [23]24.51 / 0.8128.42 / 0.8923.84 / 0.7824.21 / 0.77
JCD3 [56]23.52 / 0.7525.85 / 0.7923.87 / 0.7523.18 /0.70
RSS-T127.47 / 0.8927.40 / 0.9027.9 / 0.8928.10 / 0.89
MIMO+1[6]25.16 / 0.8626.75 / 0.8725.54 / 0.8525.58 / 0.85
Uformer1[46]24.28 / 0.8425.54 / 0.8624.26 / 0.8324.57 / 0.83
", + "bbox": [ + 81, + 186, + 467, + 281 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "independent deblurring issues under identical scenes. For fairness, we artificially elongate the exposure time of a GS camera to synthesize blurred images. As shown in Figure 6, the exposure time of GS is set to be the same as that of the full scanning period of RS and GRR. From Table 4, row-dependent blur from GRR is more tractable than blur generated by a GS camera. Because of the characteristics of GRR exposure, it implicitly encodes the temporal ordering of latent frames that mitigate the motion ambiguity [13] of blur images. And local details from the top sharper yet noisier rows further enhance the reconstruction performance. The practical implication on usage is that, given a GS camera in hand that supports GRR mode, we recommend switching to the GRR mode for single image deblurring with our proposed algorithm. Note that there is no additional cost to pay here since GRR mode will not slow down the original GS frame rate nor cause additional saturation since the longest exposure time of GRR is the same as the GS exposure time (Figure 6).", + "bbox": [ + 75, + 308, + 472, + 595 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 612, + 194, + 628 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we analyzed image degradations for conventional GS and RS cameras and further highlighted an in-between exposure mode, named GRR, arising from the global reset feature equipped with many RS sensors. Instead of RS distortion, GRR gives rise to a specialized blurring effect with row-dependent blur magnitude and brightness level. Based on the characteristics of GRR degradation, an original Transformer block with spatial compensation and spectral enhancement was devised, which has the advantage of being able to rectify the GRR degradation. The proposed hierarchical positional encoding well preserved the absolute and relative position context and the spectral branch of the feed-forward network enhanced the model's ability to capture long-range discrepancies. The experimental results also verified that our tailored model is effective.", + "bbox": [ + 75, + 638, + 470, + 864 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Potential Application As shown in Figure 7, GRR is mainly intended for industrial usage in controlled dark environment,", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2e0fb29440ee26553e76989229d118e2737cda86a3c3a8cbe48d801dc53db3f9.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 501, + 90, + 671, + 176 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/97902040cc3210cba77bc08f0dd8af75c1bfe8f967c575c607d4b066791e2982.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 653, + 90, + 807, + 185 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/31f922b570d546f42bc916c8b346add328bf96ea0c49e21224f9e827dd1c73af.jpg", + "image_caption": [ + "Figure 7: Application scenario of GRR mode. (a) Illustrated application scenario. (b) Synchronization between illumination and cameras." + ], + "image_footnote": [], + "bbox": [ + 810, + 97, + 890, + 186 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "such as parts inspection on a conveyor belt, together with strobe illuminant. In this scenario, GRR with synchronized illumination can capture distortion-free and blur-free images, with lower cost than using a GS sensor, and faster speed than using a RS sensor (Because, in this case, the exposure time of RS has to be elongated such that the first and last rows have sufficient overlap). Instead, without using strobe illumination, GRR boosted with our approach can still restore a sharp image. This will further reduce production and maintenance cost in practice.", + "bbox": [ + 496, + 287, + 893, + 439 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations When the illumination is bright enough, with appropriate exposure, GS can capture a sharp image without blur. The bottom parts of GRR can do the same, yet the top parts will be darker and thus noiser, or the top parts can be sharp, yet the the bottom parts will be blurry. In this scenario, GRR is clearly worse than GS, and our algorithm can help reduce the performance gap. On the other hand, RS will not have distortion and the image will be sharp and clean when dealing with completely static scenes. Yet GRR will have non-uniform brightness, although blur will not happen. Our algorithm can compensate this non-uniformity, yet GRR is worse than RS in this particular scenario.", + "bbox": [ + 496, + 443, + 893, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement This research was supported in part by JSPS KAKENHI Grant Numbers 22H00529, 20H05951, 22H03620, 22H05015, JST AIP Acceleration Research JP-MJCR22U4, the Value Exchange Engineering, a joint research project between Mercari, Inc. and RIISE.", + "bbox": [ + 496, + 628, + 893, + 705 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 719, + 596, + 734 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Cenek Albl, Zuzana Kukelova, Viktor Larsson, Michal Polic, Tomas Pajdla, and Konrad Schindler. From two rolling shutters to one global shutter. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2505-2513, 2020. 1, 3", + "[2] Srinadh Bhojanapalli, Chulhee Yun, Ankit Singh Rawat, Sashank Reddi, and Sanjiv Kumar. Low-rank bottleneck in multi-head attention models. In International Conference on Machine Learning, pages 864-873. PMLR, 2020. 5, 8", + "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End" + ], + "bbox": [ + 506, + 744, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12277", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3", + "[4] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolution. Advances in Neural Information Processing Systems, 33:4479-4488, 2020. 3, 6", + "[5] Lu Chi, Guiyu Tian, Yadong Mu, Lingxi Xie, and Qi Tian. Fast non-local neural networks with spectral residual learning. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2142-2151, 2019. 3", + "[6] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4641-4650, 2021. 3, 6, 7, 8, 9", + "[7] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Conditional positional encodings for vision transformers. arXiv preprint arXiv:2102.10882, 2021. 5", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3", + "[9] Bin Fan, Yuchao Dai, and Mingyi He. Sunet: symmetric undistortion network for rolling shutter correction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4541-4550, 2021. 1, 3", + "[10] Hongyun Gao, Xin Tao, Xiaoyong Shen, and Jiaya Jia. Dynamic scene deblurring with parameter selective sharing and nested skip connections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3848-3856, 2019. 1", + "[11] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 3", + "[12] Tae Hyun Kim, Kyoung Mu Lee, Bernhard Scholkopf, and Michael Hirsch. Online video deblurring via dynamic temporal blending network. In Proceedings of the IEEE International Conference on Computer Vision, pages 4038-4047, 2017. 3, 6, 7, 8", + "[13] Meiguang Jin, Givi Meishvili, and Paolo Favaro. Learning to extract a video sequence from a single motion-blurred image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6334-6342, 2018. 9", + "[14] Yitzhak Katznelson. An introduction to harmonic analysis. Cambridge University Press, 2004. 3", + "[15] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiri Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8183-8192, 2018. 3", + "[16] Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8878-8887, 2019. 3" + ], + "bbox": [ + 78, + 90, + 472, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Yizhen Lao and Omar Ait-Aider. A robust method for strong rolling shutter effects correction using lines with automatic feature selection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4795-4803, 2018. 1, 3, 6", + "[18] Junyong Lee, Hyeongseok Son, Jaesung Rim, Sunghyun Cho, and Seungyong Lee. Iterative filter adaptive network for single image defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2034-2042, 2021. 3", + "[19] Anat Levin. Blind motion deblurring using image statistics. Advances in Neural Information Processing Systems, 19, 2006. 3", + "[20] Dongxu Li, Chenchen Xu, Kaihao Zhang, Xin Yu, Yiran Zhong, Wenqi Ren, Hanna Suominen, and Hongdong Li. Arvo: Learning all-range volumetric correspondence for video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7721-7731, 2021. 3", + "[21] Kunchang Li, Yali Wang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatiotemporal representation learning. arXiv preprint arXiv:2201.04676, 2022. 5", + "[22] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 2, 3, 4", + "[23] Peidong Liu, Zhaopeng Cui, Viktor Larsson, and Marc Pollefeys. Deep shutter unrolling network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5941-5949, 2020. 1, 3, 6, 7, 8, 9", + "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 3, 5", + "[25] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 1, 3", + "[26] Seungjun Nah, Sanghyun Son, and Kyoung Mu Lee. Recurrent neural networks with intra-frame iterations for video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8102-8111, 2019. 3, 6, 7, 8", + "[27] Mehdi Noroozi, Paramanand Chandramouli, and Paolo Favaro. Motion deblurring in the wild. In German conference on pattern recognition, pages 65-77. Springer, 2017. 3", + "[28] Jihyong Oh and Munchurl Kim. Demfi: deep joint deblurring and multi-frame interpolation with flow-guided attentive correlation and recursive boosting. In European Conference on Computer Vision, pages 198-215. Springer, 2022. 8", + "[29] Chandramouli Paramanand and Ambasamudram N Rajagopalan. Non-uniform motion deblurring for bilayer scenes." + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12278", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1115-1122, 2013. 3", + "[30] Dongwon Park, Dong Un Kang, Jisoo Kim, and Se Young Chun. Multi-temporal recurrent neural networks for progressive non-uniform single image deblurring with incremental temporal training. In European Conference on Computer Vision, pages 327-343. Springer, 2020. 1", + "[31] Pulak Purkait and Christopher Zach. Minimal solvers for monocular rolling shutter compensation under ackermann motion. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 903–911. IEEE, 2018. 1, 3, 6", + "[32] Pulak Purkait, Christopher Zach, and Ales Leonardis. Rolling shutter correction in manhattan world. In Proceedings of the IEEE International Conference on Computer Vision, pages 882-890, 2017. 1, 3, 6", + "[33] Kuldeep Purohit and AN Rajagopalan. Region-adaptive dense network for efficient motion deblurring. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11882-11889, 2020. 1", + "[34] Vijay Rengarajan, Yogesh Balaji, and AN Rajagopalan. Unrolling the shutter: Cnn to correct motion distortions. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 2291-2299, 2017. 1, 3, 6", + "[35] Vijay Rengarajan, Ambasamudram N Rajagopalan, and Rangarajan Aravind. From bows to arrows: Rolling shutter rectification of urban scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2773-2781, 2016. 1, 3, 6", + "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3, 6, 7", + "[37] Christian J Schuler, Michael Hirsch, Stefan Harmeling, and Bernhard Schölkopf. Learning to deblur. IEEE transactions on pattern analysis and machine intelligence, 38(7):1439-1451, 2015. 3", + "[38] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. arXiv preprint arXiv:1803.02155, 2018. 5", + "[39] Wang Shen, Wenbo Bao, Guangtao Zhai, Li Chen, Xiongkuo Min, and Zhiyong Gao. Blurry video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5114-5123, 2020. 1", + "[40] Hyeongseok Son, Junyong Lee, Sunghyun Cho, and Seungyong Lee. Single image defocus deblurring using kernel-sharing parallel atrous convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2642-2650, 2021. 3", + "[41] Hyeongseok Son, Junyong Lee, Jonghyeop Lee, Sunghyun Cho, and Seungyong Lee. Recurrent video deblurring with blur-invariant motion estimation and pixel volumes. ACM Transactions on Graphics (TOG), 40(5):1-18, 2021. 3", + "[42] Shuochen Su, Maurizio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 1279-1288, 2017. 6, 7, 8", + "[43] Jian Sun, Wenfei Cao, Zongben Xu, and Jean Ponce. Learning a convolutional neural network for non-uniform motion blur removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 769-777, 2015. 3", + "[44] Yu-Wing Tai, Ping Tan, and Michael S Brown. Richardson-lucy deblurring for scenes under a projective motion path. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(8):1603-1618, 2010. 3", + "[45] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8174-8182, 2018. 1, 3, 6, 7, 8", + "[46] Zhendong Wang, Xiaodong Cun, Jianmin Bao, and Jianzhuang Liu. Uformer: A general u-shaped transformer for image restoration. arXiv preprint arXiv:2106.03106, 2021. 1, 2, 3, 4, 6, 7, 8, 9", + "[47] Zhixiang Wang, Xiang Ji, Jia-Bin Huang, Shin'ichi Satoh, Xiao Zhou, and Yinqiang Zheng. Neural global shutter: Learn to restore video from a rolling shutter camera with global reset feature. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17794-17803, 2022. 2, 6, 7, 8", + "[48] Patrick Wieschollek, Michael Hirsch, Bernhard Scholkopf, and Hendrik Lensch. Learning blind motion deblurring. In Proceedings of the IEEE international conference on computer vision, pages 231-240, 2017. 3", + "[49] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34, 2021. 3", + "[50] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. arXiv preprint arXiv:2111.09881, 2021. 2, 3, 4", + "[51] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 3", + "[52] Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Bjorn Stenger, Wei Liu, and Hongdong Li. Deblurring by realistic blurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2737-2746, 2020. 3", + "[53] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xi-ang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881–6890, 2021. 3", + "[54] Zhihang Zhong, Ye Gao, Yinqiang Zheng, and Bo Zheng. Efficient spatio-temporal recurrent neural network for video deblurring. In European Conference on Computer Vision, pages 191-207. Springer, 2020. 3, 8" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12279", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Zhisheng Zhong, Tiancheng Shen, Yibo Yang, Zhouchen Lin, and Chao Zhang. Joint sub-bands learning with clique structures for wavelet domain super-resolution. Advances in neural information processing systems, 31, 2018. 3", + "[56] Zhihang Zhong, Yinqiang Zheng, and Imari Sato. Towards rolling shutter correction and deblurring in dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9219-9228, 2021. 6, 7, 8, 9", + "[57] Bingbing Zhuang, Loong-Fah Cheong, and Gim Hee Lee. Rolling-shutter-aware differential sfm and image rectification. In Proceedings of the IEEE International Conference on Computer Vision, pages 948-956, 2017. 1, 3", + "[58] Bingbing Zhuang, Quoc-Huy Tran, Pan Ji, Loong-Fah Cheong, and Manmohan Chandraker. Learning structure-and-motion-aware rolling shutter correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4551-4560, 2019. 1, 3, 6" + ], + "bbox": [ + 78, + 90, + 470, + 345 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12280", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Single Image Deblurring with Row-dependent Blur Magnitude/05f13a7b-32f2-44f9-b585-5bd4c1beae6a_model.json b/2023/Single Image Deblurring with Row-dependent Blur Magnitude/05f13a7b-32f2-44f9-b585-5bd4c1beae6a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d09bba0876e58798956087aafa7c989ca144ee91 --- /dev/null +++ b/2023/Single Image Deblurring with Row-dependent Blur Magnitude/05f13a7b-32f2-44f9-b585-5bd4c1beae6a_model.json @@ -0,0 +1,2567 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.168, + 0.131, + 0.803, + 0.154 + ], + "angle": 0, + "content": "Single Image Deblurring with Row-dependent Blur Magnitude" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.18, + 0.782, + 0.218 + ], + "angle": 0, + "content": "Xiang Ji\\(^{1}\\) Zhixiang Wang\\(^{1,2}\\) Shin'ichi Satoh\\(^{2,1}\\) Yinqiang Zheng\\(^{1\\dagger}\\) \n\\(^{1}\\)The University of Tokyo, Japan \n\\(^{2}\\)National Institute of Informatics, Japan" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.22, + 0.862, + 0.235 + ], + "angle": 0, + "content": "{jixiang, wangzhixiang}@g.ecc.u-tokyo.ac.jp, satoh@nii.ac.jp, yqzheng@ai.u-tokyo.ac.jp" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.302, + 0.474, + 0.665 + ], + "angle": 0, + "content": "Image degradation often occurs during fast camera or object movements, regardless of the exposure modes: global shutter (GS) or rolling shutter (RS). Since these two exposure modes give rise to intrinsically different degradations, two restoration threads have been explored separately, i.e. motion deblurring of GS images and distortion correction of RS images, both of which are challenging restoration tasks, especially in the presence of a single input image. In this paper, we explore a novel in-between exposure mode, called global reset release (GRR) shutter, which produces GS-like blur but with row-dependent blur magnitude. We take advantage of this unique characteristic of GRR to explore the latent frames within a single image and restore a clear counterpart by only relying on these latent contexts. Specifically, we propose a residual spatially-compensated and spectrally-enhanced Transformer (RSS-T) block for row-dependent deblurring of a single GRR image. Its hierarchical positional encoding compensates global positional context of windows and enables order-awareness of the local pixel's position, along with a novel feed-forward network that simultaneously uses spatial and spectral information for gaining mixed global context. Extensive experimental results demonstrate that our method outperforms the state-of-the-art GS deblurring and RS correction methods on single GRR input." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.681, + 0.21, + 0.697 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.707, + 0.471, + 0.814 + ], + "angle": 0, + "content": "As a fundamental research task of computer vision, image restoration has been explored for many years, aiming to recover a high-quality clean image from its corrupted counterpart by removing undesired degradation. Fast or even extreme motion, which is a major inducement to image degeneration and highly correlated to the camera's exposure mode: global shutter (GS) and rolling shutter (RS)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.814, + 0.471, + 0.876 + ], + "angle": 0, + "content": "In a GS image sensor, all pixels are reset simultaneously and immediately charged with exposure, then stored for sequential readout (Figure 1a). In the presence of fast and sudden motion, blurring effects will happen. [25, 39] depict" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.269, + 0.892, + 0.302 + ], + "angle": 0, + "content": "this motion-blurred output \\(\\mathbf{I}^B\\) as an average of \\(N\\) successive latent sharp frames \\(\\mathbf{I}_i^G\\):" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.309, + 0.831, + 0.351 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} ^ {B} = \\frac {1}{N} \\sum_ {i = 0} ^ {N - 1} \\mathbf {I} _ {i} ^ {G}, i = 0, 1, \\dots , N - 1.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.895, + 0.478 + ], + "angle": 0, + "content": "In contrast, RS cameras expose image pixels row by row (including reset and readout), leading to an invariant time delay between consecutive scanlines (Figure 1b). Therefore, RS distortion effects, also known as jello effects, will appear if the camera is moving during the image acquisition, which is formulated in [23] as \\([\\mathbf{I}^R]_i = [\\mathbf{I}_i^G]_i\\). \\([\\mathbf{I}_i^G]_i\\) is an operator that extracts the \\(i^{th}\\) row from the \\(i^{th}\\) latent GS frame \\(\\mathbf{I}_i^G\\) and \\(N\\) is also the height of RS image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.896, + 0.81 + ], + "angle": 0, + "content": "Although image deblurring itself is highly involved, it is believed that restoring clear images from single degraded captures of GS cameras is much more tractable than that of RS cameras as mentioned in [23]. Intuitively, motion blur removal process aims to disentangle target GS frame from redundant input without sophisticated displacement estimation or speculation of unknown context. On the contrary, RS correction needs to shift pixels of each scanline to one virtual GS canvas by building pixel-level correspondences. Moreover, the rectified output usually suffers from missing boundaries not directly captured during RS exposure. Methodologically, the coarse-to-fine network design has gained remarkable performance for single image deblurring [25, 30, 33, 45, 10, 46]. On the contrary, due to the ill-posed nature, single image RS correction relies heavily on strong prior assumptions, explicitly [35, 17, 32, 31] or implicitly [34, 58], which limits their applicability to real scenarios. So, recent research has moved onto multi-image RS correction to bypass the ill-posedness. But both classical [57, 1] and learning-based methods [23, 9] still cannot work well under complex dynamic scenes or drastic camera motions because of nontrivial pixel alignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.897, + 0.902 + ], + "angle": 0, + "content": "It is largely overlooked in computer vision that there exists an in-between shutter mode, called global reset release (GRR) shutter, which can be easily enabled in some image sensors that originally work in GS mode (e.g., EV76C560 in EO-1312C Camera) or in RS mode (e.g., IMX178 in BFS-U3-63S4C camera). As shown in Figure 1c, GRR" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12269" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.895, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.385, + 0.897, + 0.492 + ], + "angle": 0, + "content": "Figure 1: Different exposure modes and corresponding degrading processes induced by fast motions. (a) GS exposes all pixels simultaneously, and it causes the output with row-independent blur that can be interpreted as an average of the latent sharp frames. (b) RS captures pixels scanline by scanline, featured in higher frame rate, and lower cost than GS. However, RS sensors are prone to row-dependent distortion, which can be more challenging to correct than deblurring. (c) GRR begins exposure of all pixels simultaneously but ends row by row, causing blurring effects with row-dependent blur magnitude. A specialized deblurring task is required but without the need for RS distortion correction. GRR also has the advantage of reducing the flickering effect of artificial illuminants, such as green traffic lights." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.473, + 0.804 + ], + "angle": 0, + "content": "starts to expose all pixels simultaneously as GS does but reads out signals scanline by scanline like conventional RS does. Apparently, GRR leads to blurring effects. However, different from GS blur, the blur magnitude of GRR (the brightness level and Signal-Noise-Ratio as a consequence) is row-dependent. Very recently, Wang et al. [47] proposed to enable this global reset feature in RS sensors that turns RS correction into a more tractable deblurring problem and reported superior results of video deblurring in GRR mode. We note the cost to pay for this performance gain by switching from RS to GRR mode includes 1) the superiority of RS mode over GS mode in capture speed is sacrificed since GRR does not allow early scanning of the next frame before the end of the current frame; 2) scanning rows in the bottom are more likely to be saturated, and the standard automatic exposure mechanism should be adjusted accordingly. Therefore, in the presence of multiple consecutive frames (e.g. 8 frames are used in [47]), the authentic benefits of using GRR mode are slightly questionable." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In this paper, we deal with a more challenging scenario with a single input image, which is practically meaningful in avoiding well-known flaws of using consecutive frames, including 1) Larger buffer capacity is needed; 2) Sensitivity to video recording settings, like frame rate and deadtime between frames, leading to additional generalization issues;" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.517, + 0.895, + 0.547 + ], + "angle": 0, + "content": "3) Temporal alignment with severe degradation is still challenging." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.895, + 0.623 + ], + "angle": 0, + "content": "By fully considering its unique exposure characteristic of GRR mode, we bring single GRR image deblurring into the community. Mathematically, row-dependent blur and brightness caused by varying exposure time of different scanlines could be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.633, + 0.784, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} ^ {R G} = \\mathbf {I} _ {0} ^ {G} + \\delta \\sum_ {i = 1} ^ {N - 1} \\left[ \\mathbf {I} _ {i} ^ {G} \\right] _ {i:},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.897, + 0.902 + ], + "angle": 0, + "content": "where \\([\\mathbf{I}_i^G ]_i\\) is an operator that extracts image patch of \\(i^{th}\\) row to the end from the \\(i^{th}\\) latent GS frame \\(\\mathbf{I}_i^G\\) and \\(\\delta\\) denotes the ratio between readout time and the first scanline's exposure duration. As a result, directly applying existing GS deblurring algorithms on GRR deblurring may give rise to an adaptation issue. On the other hand, Transformers have shown significant performance gains on image restoration task [22, 50] by mitigating the shortcomings of CNNs (i.e., limited receptive field and content-independent kernel). Although shifted widow strategy [22, 50, 46, 8, 24] has largely reduced computational loads and made it possible to process high dimensional input, the introduced severe corruption to image spatial information has not been addressed, especially for spatially-sensitive tasks (e.g., GRR deblurring). Besides," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12270" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "according to the spectral convolution theorem [14], updating a single value in the spectral domain globally affects all original data, which has been implemented for non-local receptive field and proven to be effective in capturing long-range context from frequency domain [4, 5, 55]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.474, + 0.501 + ], + "angle": 0, + "content": "We aim to leverage the capability of self-attention and U-net [36] structure with multi-scale input/output to tackle the single GRR deblurring task. To this end, we propose shifted horizontal window-based Transformer block with spatial compensation and spectral enhancement, which consists of three main components: 1) Shifted Window-based Multi-head Self-Attention(SW-MSA) with horizontal partition strategy and hierarchical positional encoding. We update original squared window as rectangular one to better adapt row-dependent blur and brightness. Meanwhile, hierarchical positional encoding compensates global positional context of windows lost in window partition and enables order-awareness of local pixel's position within each window. 2) Spectrally-enhanced Feed-Forward Network (SeFFN) based on depth-wise convolutional layer. The spectral branch captures the discrepancy between blurry and sharp image pairs from the perspective of frequency domain, as a complementary part of self-attention. 3) Cross-scale Feature Fusion module based on Squeeze and Excitation block (CFFSE) [11], which enables us to actively emphasize or suppress the features from encoders with different scales. The main contributions of this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.508, + 0.47, + 0.584 + ], + "angle": 0, + "content": "- We propose an original Transformer block with spatial compensation and spectral enhancement to address single GRR deblurring by fully exploring the latent frames, which further validates the advantages of GRR exposure mode over its RS counterpart." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.593, + 0.47, + 0.699 + ], + "angle": 0, + "content": "- Horizontal partition strategy and hierarchical positional encoding are used to compensate global positional context of windows and enable order-awareness of local pixel's position within each window. The spectrally-enhanced feed-forward network simultaneously uses spatial and spectral information for gaining mixed global context." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.708, + 0.471, + 0.8 + ], + "angle": 0, + "content": "- To facilitate the development and evaluation of GRR deblurring, we take paired GRR/Sharp images to offer a new dataset captured under real scenes named GRR-real. Furthermore, by simulating data captured by three shutter modes in a comparable way, the benefits of using GRR for single image restoration are verified." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.508, + 0.471, + 0.8 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.811, + 0.22, + 0.827 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Deep Image Deblurring Deep learning methods have achieved significant success in multi [12, 26, 41, 54, 20, 41] or single [15, 18, 25, 27, 40, 45, 48, 51, 52, 44, 19, 29] image deblurring. Initially, researchers try to find the spatially" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.38 + ], + "angle": 0, + "content": "varying kernels of motion blur before estimating latent image and no exception for deep CNN methods [37, 43]. However, blur kernel is normally computed for all pixels leading to huge demands for memory and computation. Besides, kernel estimation process is overly sensitive to noise and saturation, which is not practical for real scenes. Later, DeepDeblur [25] firstly exploits kernel-free learning to directly construct the relation between blurry and sharp images in an end-to-end manner. And the coarse-to-fine strategy is also taken by following SRN [45] uses an encoder-decoder network with skip-connections for three scale levels to significantly reduce training difficulty and introduce obvious stability benefits. Adversarial training has also been extensively studied [15, 16]. Recently, Cho et al. [6] revisit the coarse-to-fine scheme and present a novel deblurring network (MIMO U-net) that can handle multi-scale blur efficiently. These methods have been proven to be effective and achieved remarkable performance in single image deblurring." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.897, + 0.716 + ], + "angle": 0, + "content": "Rolling Shutter Correction RS correction methods with single image input could be divided into classical and learning based. Classical approaches heavily rely on strong prior assumptions, such as the scene is static, and the movement of the camera is limited to pure rotation or in-plane translation [35, 17, 32, 31]. Rengarajan et al. [35] propose to estimate motion by converting transformed curves to be straight based on the assumption that \"straight lines must remain straight\" and purely rotation of camera. [32] simplifies the real scene as Manhattan world to rectify the monocular RS image. These classical methods are barely applicable in real situations because of the restrictive prior assumptions. For learning-based methods, Rengarajan et al. [34] proposed a new CNN architecture based on long rectangular kernels to correct rolling shutter distortions by simply modeling the camera motion as translation+rotation polynomials. Zhuang et al. [58] extend [34] for learning to predict both the camera velocity and depth from a single RS image. The global shutter image is then recovered as a post-processing step. Recent research mainly focuses on multi-image RS correction [57, 1, 23, 9]. But those methods still cannot work under complex dynamic scenes or large camera motions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Vision Transformers As an alternative to CNN, the Transformer model has been adapted to numerous vision tasks such as image restoration [22, 50, 46], segmentation [49, 53], objection detection [3, 24]. The pioneering work of ViT [8] directly conducts self-attention on flattened patches by decomposing an image into a sequence of tokens and gets excellent results on image classification. To further overcome the quadratic complexity of original self-attention, Liu et al. [24] perform local windows with fixed size or window shift to help cross-window interaction. Liang et al. proposes an image restoration model, SwinIR [22] based on Swin Transformer [24]. Similarly, Uformer [46] imple" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12271" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.852, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.31, + 0.387, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.31, + 0.653, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.666, + 0.311, + 0.812, + 0.325 + ], + "angle": 0, + "content": "Element-wise multiplication" + }, + { + "type": "image_footnote", + "bbox": [ + 0.667, + 0.329, + 0.734, + 0.343 + ], + "angle": 0, + "content": "Upsample" + }, + { + "type": "image_footnote", + "bbox": [ + 0.667, + 0.348, + 0.747, + 0.361 + ], + "angle": 0, + "content": "Downsample" + }, + { + "type": "image_footnote", + "bbox": [ + 0.667, + 0.365, + 0.789, + 0.377 + ], + "angle": 0, + "content": "\\(\\oplus\\) Element-wise addition" + }, + { + "type": "image_footnote", + "bbox": [ + 0.667, + 0.383, + 0.753, + 0.394 + ], + "angle": 0, + "content": "\\(\\mathbb{C}\\) Concatenation" + }, + { + "type": "image_footnote", + "bbox": [ + 0.667, + 0.402, + 0.823, + 0.414 + ], + "angle": 0, + "content": "In Out Input / Output projection" + }, + { + "type": "list", + "bbox": [ + 0.666, + 0.311, + 0.823, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.435, + 0.895, + 0.511 + ], + "angle": 0, + "content": "Figure 2: Overview of our model architecture. The model is implemented in an encoder-decoder manner with multi-scale input and output. Each encoder/decoder (except for \\(\\mathsf{EB}_0\\) and \\(\\mathsf{EB}_3\\)) consists of (a) specially designed residual spatially-compensated and spectrally-enhanced transformer block (RSS-T); (b) feature attention module (FAM)/convolution layer (Conv.). Cross-scale feature fusion based on squeeze and excitation block (CFF-SE) is explored to actively emphasize or suppress the features from encoders." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.473, + 0.673 + ], + "angle": 0, + "content": "ments the Locally enhanced Window Transformer block in Unet structure and introduces a learnable multi-scale restoration modulator. All these methods have gained remarkable performance on image restoration tasks by mitigating the shortcomings of CNNs. But no one has addressed the corruption to image spatial information introduced by the widow partition strategy. And spectral feature, which is proven to capture long-range context effectively, has also been ignored in vision transformers." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.696, + 0.169, + 0.71 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.729, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Overview The overall structure is in the manner of encoder-decoder with multi-scale inputs and outputs (Figure 2). Except for \\(\\mathbb{E}\\mathbb{B}_0\\) and \\(\\mathbb{D}\\mathbb{B}_3\\), all encoder/decoder blocks contain a Feature Attention Module (FAM) or Convolution layer (Conv.) to refine aggregated features. Cross-scale feature fusion based on squeeze and excitation block (CFF-SE) is exploited to propagate information flow between encoders and decoders. To be specific, given a degraded image \\(\\mathbf{I} \\in \\mathbb{R}^{3 \\times H \\times W}\\), multi-scale input set \\(\\mathcal{I} = [\\mathbf{I}_0, \\mathbf{I}_1, \\mathbf{I}_2, \\mathbf{I}_3]\\), \\(\\mathbf{I}_k \\in \\mathbb{R}^{3 \\times \\frac{H}{2^k} \\times \\frac{W}{2^k}}\\) (\\(k = 0, 1, 2, 3\\)) is obtained by down-sampling operations. Then low-level feature \\(\\mathbf{X}_k \\in \\mathbb{R}^{2^k C \\times \\frac{H}{2^k} \\times \\frac{W}{2^k}}\\) of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.895, + 0.689 + ], + "angle": 0, + "content": "input \\(\\mathbf{I}_k\\) is extracted by its own input projection module \\(In^k\\) and will be fed into \\(\\mathbb{E}\\mathbb{B}_k\\). Similarly, the first three encoder blocks are followed by a down-sampling operation to guarantee the shape of feature map from \\(\\mathbb{E}\\mathbb{B}_{k - 1}\\) is consistent with that of \\(\\mathbf{X}_k\\). For feature reconstruction, an up-sampling operation is inserted between each two decoders. And multi-scale output set \\(\\mathcal{O} = [\\hat{\\mathbf{I}}_0,\\hat{\\mathbf{I}}_1,\\hat{\\mathbf{I}}_2,\\hat{\\mathbf{I}}_3],\\hat{\\mathbf{I}}_l\\in \\mathbb{R}^{3\\times \\frac{H}{2^l}\\times \\frac{W}{2^l}}(l = 0,1,2,3)\\) is generated by corresponding output projection module \\(Out^l\\) and residual connection from input image \\(\\mathbf{I}_l\\). Figure 2 illustrates the detailed structure of our RSS-T block." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.822, + 0.713 + ], + "angle": 0, + "content": "3.1. SW-MSA with Spatial Compensation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Two main challenges exist to directly extend the existing Transformer architecture to GRR deblurring. First, because of the characteristic of GRR exposure mode, generated spatial-variant blur and brightness are closely correlated to image rows, while normal Transformer blocks do not consider this. They divide the image into squared window [22, 50, 46], leading to large blur and brightness variance within each window, which makes the rectification process more challenging. Second, standard vision Transformer architecture cuts the computational cost due to the usage of self-attention conducted on non-overlapping windows, which enables high dimensional input processing." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12272" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.092, + 0.619, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.091, + 0.849, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.297, + 0.895, + 0.343 + ], + "angle": 0, + "content": "Figure 3: Two main components of our proposed RSS-T. (a) The proposed SW-MSA with Spatial Compensation includes two core designs: horizontal window partition strategy and hierarchical position encoding comprising window positional encoding (WinPE) and pixel positional encoding (PixPE). (b) Se-FFN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.471, + 0.461 + ], + "angle": 0, + "content": "But during window partition, image spatial information has been corrupted. That absolute spatial location is crucial for spatially-sensitive tasks. To address the aforementioned issues, we propose SW-MSA with spatial compensation, as shown in Figure 3a, which has two core designs: horizontal window partition and hierarchical positional encoding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.465, + 0.474, + 0.782 + ], + "angle": 0, + "content": "Horizontal Window Partition As discussed above, common squared windows without distinguishing blur and brightness variance along the row and column dimensions make rectification more challenging. We update the original squared window as rectangular to better adapt row-dependent blur and brightness. These long and narrow horizontal windows will mitigate large variances within each widow. Bhojanapalli et al. [2] has depicted a kind of low-rank bottleneck, i.e., while increasing the number of heads seemingly gives the model more expressive power, at the same time, the head size is reduced, which can decrease the expressive power. When input token number \\( n \\) is larger than head size \\( d_h \\), it will create a low-rank bottleneck and lose its ability to represent arbitrary context vectors. Thus, by following this rule and considering computational cost, we set the horizontal window size as (2, 16), (4, 16), (2, 32), and (4, 32). For better comparison, we also provide a normal squared window size of (8, 8) and an extreme case with a window size of (2, full). The 'full' means the window width equals the feature map width. Given input feature \\( \\mathbf{X}_{in} \\in \\mathbb{R}^{B \\times H \\times W \\times C'} \\), the partition process is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.794, + 0.47, + 0.81 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {w} = \\operatorname {W i n P a r t} \\left(\\mathbf {X} _ {i n}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.819, + 0.469, + 0.866 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}_w\\in \\mathbb{R}^{BN_w\\times H_w\\times W_w\\times C'}\\) \\(H_{w},W_{w}\\) denote height and width of windows and \\(N_{w} = H / H_{w}\\times W / W_{w}\\), which is the total number of divided windows." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Hierarchical Positional Encoding Window partition largely reduces computational cost but corrupts the image's" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.895, + 0.581 + ], + "angle": 0, + "content": "global positional information crucial to GRR deblurring. So, we propose hierarchical positional encoding: 1) Window Position Encoding (WinPE) compensates for the absolute global location lost in the partition process. 2) Pixel Position Encoding (PixPE) enables context of the local pixel's relative position within each window. Chu et al. [7] proposes a positional encoding generator to obtain Conditional Position Encodings (CPE), which can be efficiently implemented with a 2D depth convolution with kernel \\( k \\) (\\( k \\geq 3 \\)) and \\( \\frac{k - 1}{2} \\) zero paddings. They have proven zero paddings here are important to make the model aware of the absolute positions of each token. Li et al. [21] use a similar positional encoding strategy to capture temporal position for videos. Here, we extend CPE to the 3D situation and propose our WinPE:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.59, + 0.892, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\operatorname {W i n P E} \\left(\\mathcal {R} \\left(\\mathbf {X} _ {w}\\right)\\right) = \\operatorname {D W C o n v} \\left(\\mathcal {R} \\left(\\mathbf {X} _ {w}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.894, + 0.797 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}(\\cdot)\\) is reshape operation and DWConv means 3D depthwise convolution with zero paddings. Each window's absolute position will be encoded by repeatedly applying depthwise convolution to the 3D feature space. Moreover, this encoding is dynamic and allows us to eliminate the constraint to the token's length when testing on different datasets, which perfectly matches the requirement of our window position encoding. The fixed size of the window will lead to the dynamic length of the window sequence once the image size changes. We also apply the relative position encoding [24, 38] for PixPE through a learnable parameter table \\(\\mathcal{B} \\in \\mathbb{R}^{(2H_w - 1) \\times (2W_w - 1)}\\). Overall, the formulation is:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.805, + 0.806, + 0.821 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} = \\mathbf {X} _ {i n} + \\operatorname {M S A} \\left(\\ln \\left(\\mathbf {X} _ {e}\\right)\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.824, + 0.892, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {e} = \\mathbf {X} _ {w} + \\mathcal {R} (\\operatorname {D W C o n v} (\\mathcal {R} (\\mathbf {X} _ {w}))), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.845, + 0.849, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M S A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t M a x} \\left(\\mathbf {Q K} ^ {T} / \\sqrt {d _ {h}} + \\mathbf {B}\\right) \\mathbf {V},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.893, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q}\\), \\(\\mathbf{K}\\) and \\(\\mathbf{V}\\) are projected query, key and value from input. \\(\\mathbf{B}\\) is the relative position bias, whose values are taken" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12273" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.47, + 0.107 + ], + "angle": 0, + "content": "from \\(\\mathcal{B}\\). \\(\\mathrm{LN}(\\cdot)\\) means layer norm and \\(d_h\\) denotes head size." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.126, + 0.31, + 0.143 + ], + "angle": 0, + "content": "3.2. Spectrally-enhanced FFN" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.471, + 0.273 + ], + "angle": 0, + "content": "Researchers have proven the spectral feature effectively captures long-range dependencies. The Fast Fourier Convolution (FFC) operator in [4] could efficiently implement a non-local receptive field. Inspired by this work, we devised our Se-FFN based on depth-wise convolutional layer. The spectral branch captures the discrepancy between blurry and sharp image pairs as a supplemental part of self-attention for effective long-range dependency modeling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.473, + 0.518 + ], + "angle": 0, + "content": "Se-FFN includes two parts: spectral branch and spatial branch (Figure 3b). The spatial part first applies layer norm to the input feature followed by two convolutions with \\(1 \\times 1\\) kernel and a \\(3 \\times 3\\) depth-wise convolution to capture local information. The spectral branch is formulated as (a) convert the input feature map \\(\\mathbf{Y} \\in \\mathbb{R}^{B \\times H \\times W \\times C'}\\) into the frequency domain to obtain \\(\\mathcal{F}(\\mathbf{Y})\\). (b) Concatenate the real part and imaginary part of \\(\\mathcal{F}(\\mathbf{Y})\\) along channel dimension to get the spectral feature \\(\\tilde{\\mathbf{Y}} \\in \\mathbb{R}^{B \\times H \\times W/2 \\times 2C'}\\). (c) Exploit two \\(1 \\times 1\\) convolutional layers with the LeakyReLU function in between to process the spectral feature. (d) Split the processed spectral feature as real and imaginary parts, then convert them back to spatial domain \\(\\mathbf{Y}_f \\in \\mathbb{R}^{B \\times H \\times W \\times C'}\\) by inverse Fast Fourier Transform (iFFT). Finally, the output of Se-FFN is \\(\\mathbf{Z} = \\mathbf{Y} + \\mathbf{Y}_f + \\mathbf{Y}_s\\), where \\(\\mathbf{Y}_s\\) is the output of spatial branch." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.537, + 0.321, + 0.551 + ], + "angle": 0, + "content": "3.3. Cross-scale Feature Fusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.47, + 0.76 + ], + "angle": 0, + "content": "In most conventional U-net structures, the corresponding encoder and decoder are related by a skip connection to allow information flow between the same scale level. Later in [6], authors present an Asymmetric Feature Fusion (AFF) module to enable cross-scale interaction but without any control. So, we present our CFF-SE block. As shown in Figure 2, before the convolutional block, the input firstly goes through a SE module to control the information flow from the respective scale level of all encoders, which enables us to actively emphasize or suppress the features. Thus, each decoder can selectively exploit multi-scale features, resulting in improved deblurring performance as will be demonstrated in the experimental part." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.782, + 0.21, + 0.799 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Learning-based methods for deblurring or RS correction are usually trained on synthetic data, which has limitations of generalization. We, therefore, built an image acquisition system to collect a real dataset named GRR-real. The details about the system, training implementation, and more experimental results are offered in supplemental materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.791, + 0.108 + ], + "angle": 0, + "content": "4.1. Comparison on GRR Deblurring" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.118, + 0.894, + 0.18 + ], + "angle": 0, + "content": "Comparison with Other Models Since single GRR deblurring is addressed for the first time, there is no specially devised algorithm. So, we compare our model with previous state-of-the-art approaches from:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.19, + 0.892, + 0.235 + ], + "angle": 0, + "content": "- GS image-based deblurring by exploiting coarse-to-fine strategy: basic U-net [36], SRN [45], MIMO-UNet[6] and Uformer[46]." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.238, + 0.894, + 0.284 + ], + "angle": 0, + "content": "- GS video-based deblurring by fusing neighboring temporal information: STR-CNN [12], DBN [42] and IFIRNN[26]." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.295, + 0.894, + 0.326 + ], + "angle": 0, + "content": "- RS video-based correction through predicting pixelwise displacement: JCD [56] and DSUR [23].1" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.336, + 0.894, + 0.412 + ], + "angle": 0, + "content": "- GRR video deblurring using a spatial-aware encoder with long-short term temporal information aggregator: NGS [47]. For fair comparisons, we not only provide the original setting with 8 frames as input \\((\\mathrm{NGS}_8)\\) but also update the model with a single input \\((\\mathrm{NGS}_1)\\)." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.19, + 0.894, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.422, + 0.895, + 0.633 + ], + "angle": 0, + "content": "After retraining all models, Table 1 depicts quantitative results on the GRR-real test dataset. Overall, it demonstrates our specially devised algorithm significantly outperforms models from RS correction and GS deblurring, even though they take multi frames as input and exploit temporal information. As discussed in Section 1, RS correction usually resorts to nontrivial motion estimation and warping process. When directly used for GRR, they performed worse than deblurring models. Furthermore, we observe that, for deblurring models, those with video clips as input may degrade sharply, even are no match for some approaches of a single input. A key reason is that these models are struggling to align the content of different video frames. If failed, the redundant inputs would confuse learning process instead." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.634, + 0.895, + 0.83 + ], + "angle": 0, + "content": "Although the tailored GRR video deblurring method \\(\\mathrm{NGS}_8\\) [47] achieved higher metrics than ours (30.03/0.90 vs. 28.64/0.90), it heavily relies on temporal correlation of 8 frames and the performance decreased drastically when given single input \\((\\mathrm{NGS}_1)\\). In contrast to RSS-T, the demand of \\(\\mathrm{NGS}_8\\) for 8 consecutive frames in training or testing can be difficult to be satisfied, sometimes even impossible due to buffer restrictions. Besides, sensitivity to video capturing settings (e.g., frame rate and deadtime between frames) will lead to additional generalization issues. We also computed the complexity of all algorithms, which shows that our model is at medium level and outperforms SOTA RS correction or Transformer-based models. From qualitative" + }, + { + "type": "page_footnote", + "bbox": [ + 0.498, + 0.839, + 0.894, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)Noting that single RS image correction methods either heavily rely on strong assumption [35, 17, 32, 31], which is apparently inapplicable to our natural scene dataset or require actual velocity or depth to train the model [34, 58] that are barely accessible in common cameras. So, we ignore them and only choose RS video correction methods to compare." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12274" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.252, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.157, + 0.183, + 0.187, + 0.194 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.089, + 0.408, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.183, + 0.356, + 0.194 + ], + "angle": 0, + "content": "U-net [36]" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.089, + 0.563, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.458, + 0.183, + 0.51, + 0.194 + ], + "angle": 0, + "content": "SRN [45]" + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.089, + 0.718, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.183, + 0.668, + 0.194 + ], + "angle": 0, + "content": "MIMO [6]" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.089, + 0.875, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.183, + 0.828, + 0.194 + ], + "angle": 0, + "content": "MIMO+ [6]" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.194, + 0.251, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.287, + 0.208, + 0.298 + ], + "angle": 0, + "content": "Uformer [46]" + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.194, + 0.408, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.287, + 0.287, + 0.37, + 0.298 + ], + "angle": 0, + "content": "STR-CNN [12]" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.194, + 0.563, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.458, + 0.287, + 0.511, + 0.298 + ], + "angle": 0, + "content": "DBN [42]" + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.194, + 0.718, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.287, + 0.675, + 0.299 + ], + "angle": 0, + "content": "IFIRNN [26]" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.194, + 0.875, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.287, + 0.821, + 0.299 + ], + "angle": 0, + "content": "JCD [56]" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.299, + 0.251, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.391, + 0.203, + 0.403 + ], + "angle": 0, + "content": "DSUR [23]" + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.299, + 0.408, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.391, + 0.358, + 0.403 + ], + "angle": 0, + "content": "NGS1 [47]" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.299, + 0.563, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.391, + 0.514, + 0.403 + ], + "angle": 0, + "content": "NGS8 [47]" + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.299, + 0.718, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.622, + 0.391, + 0.659, + 0.402 + ], + "angle": 0, + "content": "RSS-T" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.299, + 0.875, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.391, + 0.806, + 0.402 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.416, + 0.892, + 0.447 + ], + "angle": 0, + "content": "Figure 4: Qualitative results on the GRR-real dataset. We compare our method with state-of-the-art methods from RS correction, GS deblurring, and GRR video deblurring." + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.461, + 0.895, + 0.537 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison on GRR-real dataset. We compare our method with state-of-the-art motion deblurring, RS correction, and GRR video deblurring approaches. According to 'exposure mode/Input number', they could be divided as 'GS/Single', 'GS/Multi', 'RS/Multi', 'GRR/Multi', and 'GRR/single'. The performance is measured with mean PSNR/SSIM (↑). 'Full' denotes evaluation using full-size frames, while 'Top', 'Middle', and 'Bottom' represent using only a patch of images from the top, middle, and bottom areas." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.548, + 0.864, + 0.778 + ], + "angle": 0, + "content": "
MethodMode / InputEffectivenessEfficiency
TopMiddleBottomFullTime (s)Params (M)FLOPs (T)
Input-15.12 / 0.6721.83 / 0.7820.08 / 0.7617.61 / 0.74---
U-net [36]26.34 / 0.9223.83 / 0.8522.74 / 0.8323.57 / 0.870.00629.500.072
SRN [45]25.11 / 0.7824.34 / 0.7224.08 / 0.7223.84 / 0.740.014010.250.509
MIMO [6]GS / Single27.16 /0.9024.55 / 0.8224.04 / 0.7924.48 / 0.840.01236.810.315
MIMO+ [6]28.46 / 0.9126.31 / 0.8525.78 / 0.8326.12 / 0.860.024916.110.724
Uformer [46]25.04 / 0.9323.70 / 0.8622.20 / 0.8422.92 / 0.870.214720.630.257
STR-CNN [12]20.88 / 0.8323.28 / 0.7722.60 / 0.7521.39 / 0.780.01940.930.367
DBN [42]GS / Multi25.11 / 0.9025.90 / 0.8525.87 / 0.8224.97 / 0.850.022915.311.046
IFIRNN [26]27.19 / 0.9023.85 / 0.7823.21 / 0.7723.97 / 0.810.01351.640.581
JCD [56]RS / Multi27.64 / 0.9023.65 / 0.8019.76 / 0.7722.31 / 0.820.26258.670.326
DSUR [23]24.81 / 0.8723.87 / 0.7823.39 / 0.7623.35 / 0.800.30183.900.225
NGS8 [47]GRR / Multi31.71 / 0.9330.54 / 0.9029.12 / 0.8730.03 / 0.900.02338.671.266
NGS1 [47]GRR / Single26.61 / 0.8923.53 / 0.7822.83 / 0.7623.67 / 0.810.01874.560.083
RSS-T30.90 / 0.9328.60 / 0.8827.86 / 0.8628.64 / 0.900.147911.340.176
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.801, + 0.471, + 0.877 + ], + "angle": 0, + "content": "results in Figure 4, although the images obtained by the existing networks exhibit mitigated distortions compared to the input, local details and structures were not sufficiently corrected, whereas our method produces sharper images that even visually outperform \\(\\mathrm{NGS}_8\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Thirdparty Evaluation To further prove the practical value" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.801, + 0.892, + 0.862 + ], + "angle": 0, + "content": "of our method, we evaluate it on another GRR dataset from [47]. As shown in Table 2, our method still outperforms other algorithms and is just slightly inferior to \\(\\mathrm{NGS}_8\\) with as many as eight input frames (26.62/0.86 vs. 27.29/0.85)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12275" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.095, + 0.09, + 0.451, + 0.105 + ], + "angle": 0, + "content": "Table 2: Comparison on another GRR dataset [47]." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.113, + 0.467, + 0.303 + ], + "angle": 0, + "content": "
MethodSET-II
FullTopMiddleBottom
Input17.82 / 0.7323.64 / 0.7721.45 / 0.7715.54 / 0.66
SRN [45]25.05 / 0.8124.32 / 0.7925.65 / 0.8127.02 / 0.83
STRCNN [12]22.59 / 0.8122.99 / 0.7923.46 / 0.8123.66 / 0.83
DBN [42]22.57 / 0.8123.24 / 0.8023.81 / 0.8123.24 / 0.82
IFIRNN[26]25.17 / 0.8224.77 / 0.8025.62 / 0.8126.94 / 0.84
ESTRNN [54]22.72 / 0.8323.42 / 0.8126.03 / 0.8322.86 / 0.83
DSUR [23]22.50 / 0.8022.49 / 0.7823.87 / 0.8123.38 / 0.83
JCD [56]25.33 / 0.8024.77 / 0.7825.71 / 0.8027.43 / 0.83
NGS8 [47]27.29 / 0.8526.96 / 0.8427.57 / 0.8528.35 / 0.86
NGS1 [47]25.19 / 0.7924.48 / 0.7725.53 / 0.7927.27 / 0.82
RSS-T26.62 / 0.8625.90 / 0.8426.82 / 0.8628.43 / 0.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.324, + 0.468, + 0.339 + ], + "angle": 0, + "content": "Table 3: Architecture ablation on our GRR-real dataset." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.347, + 0.466, + 0.466 + ], + "angle": 0, + "content": "
PEFFNFusionPSNR / SSIM
WinPEPixPELeFNNSe-FNNAFFCFF-SE
v1✓✓25.69 / 0.85
v2✓✓✓26.50 / 0.86
v3✓✓✓26.10 / 0.86
v4✓✓✓✓26.83 / 0.87
v5✓✓✓✓27.24 / 0.87
RSS-T✓✓✓✓28.64 / 0.90
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.499, + 0.229, + 0.515 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.471, + 0.742 + ], + "angle": 0, + "content": "Model Architecture Ablation We conducted experiments to analyze the effectiveness of the core components of our RSS-T. The hierarchical position encoding consists of WinPE and PixPE. The feed-forward network is also implemented by LeFFN [46] to compare with our Se-FNN (v5 vs. RSS-T). As for the cross-scale fusion part, AFF [46] and our CFF-SE are used to explore the performance difference (v4 vs. v5). All the experimental results are listed in Table 3. From v1, v2, v3, and v4, when WinPE and PixPE are separately presented in baseline model v1, the performance goes up but the combination of them achieved the best metrics. Our Se-FNN and CFF-SE modules contributed to the further performance improvement of PSNR by \\(+1.81\\mathrm{dB}\\) compared with the LeFFN and AFF." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Window Size Ablation We introduced three types of window size (i.e., horizontal, squared, and full-width window) and clarified the advantages of the horizontal window for GRR deblurring in Section 3.1. Here, we present experimental results to support the claim. As Figure 5 shows, all horizontal window settings are superior to the standard squared one from quantitative and qualitative perspectives. The extreme case (2, full) is severely trapped in low-rank bottleneck [2] as the pixel number of this window is far more than the head size." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.088, + 0.889, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.338, + 0.892, + 0.368 + ], + "angle": 0, + "content": "Figure 5: Window size ablation. Experimental results of all different window-size settings on our GRR-real dataset." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.384, + 0.886, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.488, + 0.892, + 0.532 + ], + "angle": 0, + "content": "Figure 6: Temporal alignment of three shutters. When synthesizing datasets, we strictly align the exposure of three modes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.822, + 0.576 + ], + "angle": 0, + "content": "4.3. Comparison on Three Shutter Modes" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Until now, a direct comparison of three shutter modes is still missing, although [47] has demonstrated that GRR has its superiority against RS in obtaining sharp images in the scenario of video restoration. Here, we evaluate them on single image Restoration. For a fair comparison, we synthesize strictly aligned images virtually captured by three modes (RS, GRR, and blur) for each scene by following the protocol of [28, 45, 6] (details in the supplemental material), and compare with SOTA methods for each mode." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.893, + 0.884 + ], + "angle": 0, + "content": "When compared with RS mode, Table 4 shows that GRR combined with our RSS-T model has gained significant advantages in obtaining single sharp image, even though DSUR and JCD take multiple frames as input. These results are consistent with the discussion in [47] that the GRR mode enables us to convert the RS correction problem into a more tractable one. Therefore, given an RS camera that supports GRR mode, we recommend switching to GRR mode for better restoration from a single input, if losing some capture speed of the original RS is permissible. Also, the exposure should be adjusted to avoid saturation of bottom rows." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We additionally present comparisons with row-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12276" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.472, + 0.18 + ], + "angle": 0, + "content": "Table 4: Quantitative comparison on three modes. The subscript of each method denotes the number of input frames. The reason we choose multi-frame RS correction is discussed in Section 4.1. MIMO+ and Uformer are trained on blur mode. DSUR and JCD are trained on the RS counterpart, while RSS-T is trained on GRR mode." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.188, + 0.468, + 0.282 + ], + "angle": 0, + "content": "
MethodFullTopMiddleBottom
DSUR2 [23]24.51 / 0.8128.42 / 0.8923.84 / 0.7824.21 / 0.77
JCD3 [56]23.52 / 0.7525.85 / 0.7923.87 / 0.7523.18 /0.70
RSS-T127.47 / 0.8927.40 / 0.9027.9 / 0.8928.10 / 0.89
MIMO+1[6]25.16 / 0.8626.75 / 0.8725.54 / 0.8525.58 / 0.85
Uformer1[46]24.28 / 0.8425.54 / 0.8624.26 / 0.8324.57 / 0.83
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.309, + 0.473, + 0.596 + ], + "angle": 0, + "content": "independent deblurring issues under identical scenes. For fairness, we artificially elongate the exposure time of a GS camera to synthesize blurred images. As shown in Figure 6, the exposure time of GS is set to be the same as that of the full scanning period of RS and GRR. From Table 4, row-dependent blur from GRR is more tractable than blur generated by a GS camera. Because of the characteristics of GRR exposure, it implicitly encodes the temporal ordering of latent frames that mitigate the motion ambiguity [13] of blur images. And local details from the top sharper yet noisier rows further enhance the reconstruction performance. The practical implication on usage is that, given a GS camera in hand that supports GRR mode, we recommend switching to the GRR mode for single image deblurring with our proposed algorithm. Note that there is no additional cost to pay here since GRR mode will not slow down the original GS frame rate nor cause additional saturation since the longest exposure time of GRR is the same as the GS exposure time (Figure 6)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.613, + 0.196, + 0.629 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.639, + 0.472, + 0.866 + ], + "angle": 0, + "content": "In this paper, we analyzed image degradations for conventional GS and RS cameras and further highlighted an in-between exposure mode, named GRR, arising from the global reset feature equipped with many RS sensors. Instead of RS distortion, GRR gives rise to a specialized blurring effect with row-dependent blur magnitude and brightness level. Based on the characteristics of GRR degradation, an original Transformer block with spatial compensation and spectral enhancement was devised, which has the advantage of being able to rectify the GRR degradation. The proposed hierarchical positional encoding well preserved the absolute and relative position context and the spectral branch of the feed-forward network enhanced the model's ability to capture long-range discrepancies. The experimental results also verified that our tailored model is effective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Potential Application As shown in Figure 7, GRR is mainly intended for industrial usage in controlled dark environment," + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.673, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.19, + 0.597, + 0.202 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.091, + 0.808, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.19, + 0.799, + 0.202 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.812, + 0.098, + 0.892, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.215, + 0.895, + 0.261 + ], + "angle": 0, + "content": "Figure 7: Application scenario of GRR mode. (a) Illustrated application scenario. (b) Synchronization between illumination and cameras." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.895, + 0.44 + ], + "angle": 0, + "content": "such as parts inspection on a conveyor belt, together with strobe illuminant. In this scenario, GRR with synchronized illumination can capture distortion-free and blur-free images, with lower cost than using a GS sensor, and faster speed than using a RS sensor (Because, in this case, the exposure time of RS has to be elongated such that the first and last rows have sufficient overlap). Instead, without using strobe illumination, GRR boosted with our approach can still restore a sharp image. This will further reduce production and maintenance cost in practice." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.895, + 0.626 + ], + "angle": 0, + "content": "Limitations When the illumination is bright enough, with appropriate exposure, GS can capture a sharp image without blur. The bottom parts of GRR can do the same, yet the top parts will be darker and thus noiser, or the top parts can be sharp, yet the the bottom parts will be blurry. In this scenario, GRR is clearly worse than GS, and our algorithm can help reduce the performance gap. On the other hand, RS will not have distortion and the image will be sharp and clean when dealing with completely static scenes. Yet GRR will have non-uniform brightness, although blur will not happen. Our algorithm can compensate this non-uniformity, yet GRR is worse than RS in this particular scenario." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.895, + 0.706 + ], + "angle": 0, + "content": "Acknowledgement This research was supported in part by JSPS KAKENHI Grant Numbers 22H00529, 20H05951, 22H03620, 22H05015, JST AIP Acceleration Research JP-MJCR22U4, the Value Exchange Engineering, a joint research project between Mercari, Inc. and RIISE." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.72, + 0.597, + 0.735 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.745, + 0.895, + 0.814 + ], + "angle": 0, + "content": "[1] Cenek Albl, Zuzana Kukelova, Viktor Larsson, Michal Polic, Tomas Pajdla, and Konrad Schindler. From two rolling shutters to one global shutter. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2505-2513, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.816, + 0.895, + 0.872 + ], + "angle": 0, + "content": "[2] Srinadh Bhojanapalli, Chulhee Yun, Ankit Singh Rawat, Sashank Reddi, and Sanjiv Kumar. Low-rank bottleneck in multi-head attention models. In International Conference on Machine Learning, pages 864-873. PMLR, 2020. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.873, + 0.895, + 0.902 + ], + "angle": 0, + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.745, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12277" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.133 + ], + "angle": 0, + "content": "to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.134, + 0.473, + 0.175 + ], + "angle": 0, + "content": "[4] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolution. Advances in Neural Information Processing Systems, 33:4479-4488, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.177, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[5] Lu Chi, Guiyu Tian, Yadong Mu, Lingxi Xie, and Qi Tian. Fast non-local neural networks with spectral residual learning. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2142-2151, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.233, + 0.472, + 0.3 + ], + "angle": 0, + "content": "[6] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4641-4650, 2021. 3, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.472, + 0.357 + ], + "angle": 0, + "content": "[7] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Conditional positional encodings for vision transformers. arXiv preprint arXiv:2102.10882, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.358, + 0.472, + 0.44 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.442, + 0.472, + 0.496 + ], + "angle": 0, + "content": "[9] Bin Fan, Yuchao Dai, and Mingyi He. Sunet: symmetric undistortion network for rolling shutter correction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4541-4550, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.497, + 0.472, + 0.566 + ], + "angle": 0, + "content": "[10] Hongyun Gao, Xin Tao, Xiaoyong Shen, and Jiaya Jia. Dynamic scene deblurring with parameter selective sharing and nested skip connections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3848-3856, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.567, + 0.472, + 0.608 + ], + "angle": 0, + "content": "[11] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.609, + 0.472, + 0.676 + ], + "angle": 0, + "content": "[12] Tae Hyun Kim, Kyoung Mu Lee, Bernhard Scholkopf, and Michael Hirsch. Online video deblurring via dynamic temporal blending network. In Proceedings of the IEEE International Conference on Computer Vision, pages 4038-4047, 2017. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.679, + 0.472, + 0.734 + ], + "angle": 0, + "content": "[13] Meiguang Jin, Givi Meishvili, and Paolo Favaro. Learning to extract a video sequence from a single motion-blurred image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6334-6342, 2018. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.735, + 0.472, + 0.762 + ], + "angle": 0, + "content": "[14] Yitzhak Katznelson. An introduction to harmonic analysis. Cambridge University Press, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.763, + 0.472, + 0.831 + ], + "angle": 0, + "content": "[15] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiri Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8183-8192, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.832, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[16] Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8878-8887, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.473, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[17] Yizhen Lao and Omar Ait-Aider. A robust method for strong rolling shutter effects correction using lines with automatic feature selection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4795-4803, 2018. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[18] Junyong Lee, Hyeongseok Son, Jaesung Rim, Sunghyun Cho, and Seungyong Lee. Iterative filter adaptive network for single image defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2034-2042, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.894, + 0.274 + ], + "angle": 0, + "content": "[19] Anat Levin. Blind motion deblurring using image statistics. Advances in Neural Information Processing Systems, 19, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.894, + 0.359 + ], + "angle": 0, + "content": "[20] Dongxu Li, Chenchen Xu, Kaihao Zhang, Xin Yu, Yiran Zhong, Wenqi Ren, Hanna Suominen, and Hongdong Li. Arvo: Learning all-range volumetric correspondence for video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7721-7731, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.894, + 0.417 + ], + "angle": 0, + "content": "[21] Kunchang Li, Yali Wang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatiotemporal representation learning. arXiv preprint arXiv:2201.04676, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.894, + 0.487 + ], + "angle": 0, + "content": "[22] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.894, + 0.545 + ], + "angle": 0, + "content": "[23] Peidong Liu, Zhaopeng Cui, Viktor Larsson, and Marc Pollefeys. Deep shutter unrolling network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5941-5949, 2020. 1, 3, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.894, + 0.616 + ], + "angle": 0, + "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.894, + 0.686 + ], + "angle": 0, + "content": "[25] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.894, + 0.757 + ], + "angle": 0, + "content": "[26] Seungjun Nah, Sanghyun Son, and Kyoung Mu Lee. Recurrent neural networks with intra-frame iterations for video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8102-8111, 2019. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.894, + 0.813 + ], + "angle": 0, + "content": "[27] Mehdi Noroozi, Paramanand Chandramouli, and Paolo Favaro. Motion deblurring in the wild. In German conference on pattern recognition, pages 65-77. Springer, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.894, + 0.872 + ], + "angle": 0, + "content": "[28] Jihyong Oh and Munchurl Kim. Demfi: deep joint deblurring and multi-frame interpolation with flow-guided attentive correlation and recursive boosting. In European Conference on Computer Vision, pages 198-215. Springer, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[29] Chandramouli Paramanand and Ambasamudram N Rajagopalan. Non-uniform motion deblurring for bilayer scenes." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12278" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1115-1122, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.471, + 0.191 + ], + "angle": 0, + "content": "[30] Dongwon Park, Dong Un Kang, Jisoo Kim, and Se Young Chun. Multi-temporal recurrent neural networks for progressive non-uniform single image deblurring with incremental temporal training. In European Conference on Computer Vision, pages 327-343. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.472, + 0.261 + ], + "angle": 0, + "content": "[31] Pulak Purkait and Christopher Zach. Minimal solvers for monocular rolling shutter compensation under ackermann motion. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 903–911. IEEE, 2018. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.471, + 0.318 + ], + "angle": 0, + "content": "[32] Pulak Purkait, Christopher Zach, and Ales Leonardis. Rolling shutter correction in manhattan world. In Proceedings of the IEEE International Conference on Computer Vision, pages 882-890, 2017. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.472, + 0.376 + ], + "angle": 0, + "content": "[33] Kuldeep Purohit and AN Rajagopalan. Region-adaptive dense network for efficient motion deblurring. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11882-11889, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.471, + 0.433 + ], + "angle": 0, + "content": "[34] Vijay Rengarajan, Yogesh Balaji, and AN Rajagopalan. Unrolling the shutter: Cnn to correct motion distortions. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 2291-2299, 2017. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.471, + 0.503 + ], + "angle": 0, + "content": "[35] Vijay Rengarajan, Ambasamudram N Rajagopalan, and Rangarajan Aravind. From bows to arrows: Rolling shutter rectification of urban scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2773-2781, 2016. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.471, + 0.573 + ], + "angle": 0, + "content": "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.471, + 0.63 + ], + "angle": 0, + "content": "[37] Christian J Schuler, Michael Hirsch, Stefan Harmeling, and Bernhard Schölkopf. Learning to deblur. IEEE transactions on pattern analysis and machine intelligence, 38(7):1439-1451, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.471, + 0.673 + ], + "angle": 0, + "content": "[38] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. arXiv preprint arXiv:1803.02155, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.471, + 0.731 + ], + "angle": 0, + "content": "[39] Wang Shen, Wenbo Bao, Guangtao Zhai, Li Chen, Xiongkuo Min, and Zhiyong Gao. Blurry video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5114-5123, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.471, + 0.801 + ], + "angle": 0, + "content": "[40] Hyeongseok Son, Junyong Lee, Sunghyun Cho, and Seungyong Lee. Single image defocus deblurring using kernel-sharing parallel atrous convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2642-2650, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[41] Hyeongseok Son, Junyong Lee, Jonghyeop Lee, Sunghyun Cho, and Seungyong Lee. Recurrent video deblurring with blur-invariant motion estimation and pixel volumes. ACM Transactions on Graphics (TOG), 40(5):1-18, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[42] Shuochen Su, Maurizio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.893, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 1279-1288, 2017. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[43] Jian Sun, Wenfei Cao, Zongben Xu, and Jean Ponce. Learning a convolutional neural network for non-uniform motion blur removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 769-777, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.234 + ], + "angle": 0, + "content": "[44] Yu-Wing Tai, Ping Tan, and Michael S Brown. Richardson-lucy deblurring for scenes under a projective motion path. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(8):1603-1618, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.893, + 0.292 + ], + "angle": 0, + "content": "[45] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8174-8182, 2018. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.893, + 0.347 + ], + "angle": 0, + "content": "[46] Zhendong Wang, Xiaodong Cun, Jianmin Bao, and Jianzhuang Liu. Uformer: A general u-shaped transformer for image restoration. arXiv preprint arXiv:2106.03106, 2021. 1, 2, 3, 4, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.894, + 0.433 + ], + "angle": 0, + "content": "[47] Zhixiang Wang, Xiang Ji, Jia-Bin Huang, Shin'ichi Satoh, Xiao Zhou, and Yinqiang Zheng. Neural global shutter: Learn to restore video from a rolling shutter camera with global reset feature. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17794-17803, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.435, + 0.894, + 0.49 + ], + "angle": 0, + "content": "[48] Patrick Wieschollek, Michael Hirsch, Bernhard Scholkopf, and Hendrik Lensch. Learning blind motion deblurring. In Proceedings of the IEEE international conference on computer vision, pages 231-240, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.894, + 0.559 + ], + "angle": 0, + "content": "[49] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.894, + 0.618 + ], + "angle": 0, + "content": "[50] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. arXiv preprint arXiv:2111.09881, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.893, + 0.689 + ], + "angle": 0, + "content": "[51] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.893, + 0.758 + ], + "angle": 0, + "content": "[52] Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Bjorn Stenger, Wei Liu, and Hongdong Li. Deblurring by realistic blurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2737-2746, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.894, + 0.845 + ], + "angle": 0, + "content": "[53] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xi-ang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881–6890, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[54] Zhihang Zhong, Ye Gao, Yinqiang Zheng, and Bo Zheng. Efficient spatio-temporal recurrent neural network for video deblurring. In European Conference on Computer Vision, pages 191-207. Springer, 2020. 3, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12279" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[55] Zhisheng Zhong, Tiancheng Shen, Yibo Yang, Zhouchen Lin, and Chao Zhang. Joint sub-bands learning with clique structures for wavelet domain super-resolution. Advances in neural information processing systems, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.472, + 0.217 + ], + "angle": 0, + "content": "[56] Zhihang Zhong, Yinqiang Zheng, and Imari Sato. Towards rolling shutter correction and deblurring in dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9219-9228, 2021. 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.472, + 0.275 + ], + "angle": 0, + "content": "[57] Bingbing Zhuang, Loong-Fah Cheong, and Gim Hee Lee. Rolling-shutter-aware differential sfm and image rectification. In Proceedings of the IEEE International Conference on Computer Vision, pages 948-956, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.472, + 0.346 + ], + "angle": 0, + "content": "[58] Bingbing Zhuang, Quoc-Huy Tran, Pan Ji, Loong-Fah Cheong, and Manmohan Chandraker. Learning structure-and-motion-aware rolling shutter correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4551-4560, 2019. 1, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "12280" + } + ] +] \ No newline at end of file diff --git a/2023/Single Image Deblurring with Row-dependent Blur Magnitude/05f13a7b-32f2-44f9-b585-5bd4c1beae6a_origin.pdf b/2023/Single Image Deblurring with Row-dependent Blur Magnitude/05f13a7b-32f2-44f9-b585-5bd4c1beae6a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cd7836ef70e12953cf3db59064b4a29ccd9e2585 --- /dev/null +++ b/2023/Single Image Deblurring with Row-dependent Blur Magnitude/05f13a7b-32f2-44f9-b585-5bd4c1beae6a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff43c61ec0b950c81cbcd3420b483f88fd8716283eb51a59bcce37c1acf2a404 +size 2428311 diff --git a/2023/Single Image Deblurring with Row-dependent Blur Magnitude/full.md b/2023/Single Image Deblurring with Row-dependent Blur Magnitude/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d2e36b63d0c182d248e1ff69bbf218cae6926b2a --- /dev/null +++ b/2023/Single Image Deblurring with Row-dependent Blur Magnitude/full.md @@ -0,0 +1,336 @@ +# Single Image Deblurring with Row-dependent Blur Magnitude + +Xiang Ji $^{1}$ Zhixiang Wang $^{1,2}$ Shin'ichi Satoh $^{2,1}$ Yinqiang Zheng $^{1\dagger}$ $^{1}$ The University of Tokyo, Japan + $^{2}$ National Institute of Informatics, Japan + +{jixiang, wangzhixiang}@g.ecc.u-tokyo.ac.jp, satoh@nii.ac.jp, yqzheng@ai.u-tokyo.ac.jp + +# Abstract + +Image degradation often occurs during fast camera or object movements, regardless of the exposure modes: global shutter (GS) or rolling shutter (RS). Since these two exposure modes give rise to intrinsically different degradations, two restoration threads have been explored separately, i.e. motion deblurring of GS images and distortion correction of RS images, both of which are challenging restoration tasks, especially in the presence of a single input image. In this paper, we explore a novel in-between exposure mode, called global reset release (GRR) shutter, which produces GS-like blur but with row-dependent blur magnitude. We take advantage of this unique characteristic of GRR to explore the latent frames within a single image and restore a clear counterpart by only relying on these latent contexts. Specifically, we propose a residual spatially-compensated and spectrally-enhanced Transformer (RSS-T) block for row-dependent deblurring of a single GRR image. Its hierarchical positional encoding compensates global positional context of windows and enables order-awareness of the local pixel's position, along with a novel feed-forward network that simultaneously uses spatial and spectral information for gaining mixed global context. Extensive experimental results demonstrate that our method outperforms the state-of-the-art GS deblurring and RS correction methods on single GRR input. + +# 1. Introduction + +As a fundamental research task of computer vision, image restoration has been explored for many years, aiming to recover a high-quality clean image from its corrupted counterpart by removing undesired degradation. Fast or even extreme motion, which is a major inducement to image degeneration and highly correlated to the camera's exposure mode: global shutter (GS) and rolling shutter (RS). + +In a GS image sensor, all pixels are reset simultaneously and immediately charged with exposure, then stored for sequential readout (Figure 1a). In the presence of fast and sudden motion, blurring effects will happen. [25, 39] depict + +this motion-blurred output $\mathbf{I}^B$ as an average of $N$ successive latent sharp frames $\mathbf{I}_i^G$ : + +$$ +\mathbf {I} ^ {B} = \frac {1}{N} \sum_ {i = 0} ^ {N - 1} \mathbf {I} _ {i} ^ {G}, i = 0, 1, \dots , N - 1. +$$ + +In contrast, RS cameras expose image pixels row by row (including reset and readout), leading to an invariant time delay between consecutive scanlines (Figure 1b). Therefore, RS distortion effects, also known as jello effects, will appear if the camera is moving during the image acquisition, which is formulated in [23] as $[\mathbf{I}^R]_i = [\mathbf{I}_i^G]_i$ . $[\mathbf{I}_i^G]_i$ is an operator that extracts the $i^{th}$ row from the $i^{th}$ latent GS frame $\mathbf{I}_i^G$ and $N$ is also the height of RS image. + +Although image deblurring itself is highly involved, it is believed that restoring clear images from single degraded captures of GS cameras is much more tractable than that of RS cameras as mentioned in [23]. Intuitively, motion blur removal process aims to disentangle target GS frame from redundant input without sophisticated displacement estimation or speculation of unknown context. On the contrary, RS correction needs to shift pixels of each scanline to one virtual GS canvas by building pixel-level correspondences. Moreover, the rectified output usually suffers from missing boundaries not directly captured during RS exposure. Methodologically, the coarse-to-fine network design has gained remarkable performance for single image deblurring [25, 30, 33, 45, 10, 46]. On the contrary, due to the ill-posed nature, single image RS correction relies heavily on strong prior assumptions, explicitly [35, 17, 32, 31] or implicitly [34, 58], which limits their applicability to real scenarios. So, recent research has moved onto multi-image RS correction to bypass the ill-posedness. But both classical [57, 1] and learning-based methods [23, 9] still cannot work well under complex dynamic scenes or drastic camera motions because of nontrivial pixel alignment. + +It is largely overlooked in computer vision that there exists an in-between shutter mode, called global reset release (GRR) shutter, which can be easily enabled in some image sensors that originally work in GS mode (e.g., EV76C560 in EO-1312C Camera) or in RS mode (e.g., IMX178 in BFS-U3-63S4C camera). As shown in Figure 1c, GRR + +![](images/251c803ab291b0c604ee2c1c78c148e794b9b3724baefc173a263118d1ff1b0f.jpg) +Figure 1: Different exposure modes and corresponding degrading processes induced by fast motions. (a) GS exposes all pixels simultaneously, and it causes the output with row-independent blur that can be interpreted as an average of the latent sharp frames. (b) RS captures pixels scanline by scanline, featured in higher frame rate, and lower cost than GS. However, RS sensors are prone to row-dependent distortion, which can be more challenging to correct than deblurring. (c) GRR begins exposure of all pixels simultaneously but ends row by row, causing blurring effects with row-dependent blur magnitude. A specialized deblurring task is required but without the need for RS distortion correction. GRR also has the advantage of reducing the flickering effect of artificial illuminants, such as green traffic lights. + +starts to expose all pixels simultaneously as GS does but reads out signals scanline by scanline like conventional RS does. Apparently, GRR leads to blurring effects. However, different from GS blur, the blur magnitude of GRR (the brightness level and Signal-Noise-Ratio as a consequence) is row-dependent. Very recently, Wang et al. [47] proposed to enable this global reset feature in RS sensors that turns RS correction into a more tractable deblurring problem and reported superior results of video deblurring in GRR mode. We note the cost to pay for this performance gain by switching from RS to GRR mode includes 1) the superiority of RS mode over GS mode in capture speed is sacrificed since GRR does not allow early scanning of the next frame before the end of the current frame; 2) scanning rows in the bottom are more likely to be saturated, and the standard automatic exposure mechanism should be adjusted accordingly. Therefore, in the presence of multiple consecutive frames (e.g. 8 frames are used in [47]), the authentic benefits of using GRR mode are slightly questionable. + +In this paper, we deal with a more challenging scenario with a single input image, which is practically meaningful in avoiding well-known flaws of using consecutive frames, including 1) Larger buffer capacity is needed; 2) Sensitivity to video recording settings, like frame rate and deadtime between frames, leading to additional generalization issues; + +3) Temporal alignment with severe degradation is still challenging. + +By fully considering its unique exposure characteristic of GRR mode, we bring single GRR image deblurring into the community. Mathematically, row-dependent blur and brightness caused by varying exposure time of different scanlines could be formulated as: + +$$ +\mathbf {I} ^ {R G} = \mathbf {I} _ {0} ^ {G} + \delta \sum_ {i = 1} ^ {N - 1} \left[ \mathbf {I} _ {i} ^ {G} \right] _ {i:}, +$$ + +where $[\mathbf{I}_i^G ]_i$ is an operator that extracts image patch of $i^{th}$ row to the end from the $i^{th}$ latent GS frame $\mathbf{I}_i^G$ and $\delta$ denotes the ratio between readout time and the first scanline's exposure duration. As a result, directly applying existing GS deblurring algorithms on GRR deblurring may give rise to an adaptation issue. On the other hand, Transformers have shown significant performance gains on image restoration task [22, 50] by mitigating the shortcomings of CNNs (i.e., limited receptive field and content-independent kernel). Although shifted widow strategy [22, 50, 46, 8, 24] has largely reduced computational loads and made it possible to process high dimensional input, the introduced severe corruption to image spatial information has not been addressed, especially for spatially-sensitive tasks (e.g., GRR deblurring). Besides, + +according to the spectral convolution theorem [14], updating a single value in the spectral domain globally affects all original data, which has been implemented for non-local receptive field and proven to be effective in capturing long-range context from frequency domain [4, 5, 55]. + +We aim to leverage the capability of self-attention and U-net [36] structure with multi-scale input/output to tackle the single GRR deblurring task. To this end, we propose shifted horizontal window-based Transformer block with spatial compensation and spectral enhancement, which consists of three main components: 1) Shifted Window-based Multi-head Self-Attention(SW-MSA) with horizontal partition strategy and hierarchical positional encoding. We update original squared window as rectangular one to better adapt row-dependent blur and brightness. Meanwhile, hierarchical positional encoding compensates global positional context of windows lost in window partition and enables order-awareness of local pixel's position within each window. 2) Spectrally-enhanced Feed-Forward Network (SeFFN) based on depth-wise convolutional layer. The spectral branch captures the discrepancy between blurry and sharp image pairs from the perspective of frequency domain, as a complementary part of self-attention. 3) Cross-scale Feature Fusion module based on Squeeze and Excitation block (CFFSE) [11], which enables us to actively emphasize or suppress the features from encoders with different scales. The main contributions of this paper are summarized as follows: + +- We propose an original Transformer block with spatial compensation and spectral enhancement to address single GRR deblurring by fully exploring the latent frames, which further validates the advantages of GRR exposure mode over its RS counterpart. +- Horizontal partition strategy and hierarchical positional encoding are used to compensate global positional context of windows and enable order-awareness of local pixel's position within each window. The spectrally-enhanced feed-forward network simultaneously uses spatial and spectral information for gaining mixed global context. +- To facilitate the development and evaluation of GRR deblurring, we take paired GRR/Sharp images to offer a new dataset captured under real scenes named GRR-real. Furthermore, by simulating data captured by three shutter modes in a comparable way, the benefits of using GRR for single image restoration are verified. + +# 2. Related Work + +Deep Image Deblurring Deep learning methods have achieved significant success in multi [12, 26, 41, 54, 20, 41] or single [15, 18, 25, 27, 40, 45, 48, 51, 52, 44, 19, 29] image deblurring. Initially, researchers try to find the spatially + +varying kernels of motion blur before estimating latent image and no exception for deep CNN methods [37, 43]. However, blur kernel is normally computed for all pixels leading to huge demands for memory and computation. Besides, kernel estimation process is overly sensitive to noise and saturation, which is not practical for real scenes. Later, DeepDeblur [25] firstly exploits kernel-free learning to directly construct the relation between blurry and sharp images in an end-to-end manner. And the coarse-to-fine strategy is also taken by following SRN [45] uses an encoder-decoder network with skip-connections for three scale levels to significantly reduce training difficulty and introduce obvious stability benefits. Adversarial training has also been extensively studied [15, 16]. Recently, Cho et al. [6] revisit the coarse-to-fine scheme and present a novel deblurring network (MIMO U-net) that can handle multi-scale blur efficiently. These methods have been proven to be effective and achieved remarkable performance in single image deblurring. + +Rolling Shutter Correction RS correction methods with single image input could be divided into classical and learning based. Classical approaches heavily rely on strong prior assumptions, such as the scene is static, and the movement of the camera is limited to pure rotation or in-plane translation [35, 17, 32, 31]. Rengarajan et al. [35] propose to estimate motion by converting transformed curves to be straight based on the assumption that "straight lines must remain straight" and purely rotation of camera. [32] simplifies the real scene as Manhattan world to rectify the monocular RS image. These classical methods are barely applicable in real situations because of the restrictive prior assumptions. For learning-based methods, Rengarajan et al. [34] proposed a new CNN architecture based on long rectangular kernels to correct rolling shutter distortions by simply modeling the camera motion as translation+rotation polynomials. Zhuang et al. [58] extend [34] for learning to predict both the camera velocity and depth from a single RS image. The global shutter image is then recovered as a post-processing step. Recent research mainly focuses on multi-image RS correction [57, 1, 23, 9]. But those methods still cannot work under complex dynamic scenes or large camera motions. + +Vision Transformers As an alternative to CNN, the Transformer model has been adapted to numerous vision tasks such as image restoration [22, 50, 46], segmentation [49, 53], objection detection [3, 24]. The pioneering work of ViT [8] directly conducts self-attention on flattened patches by decomposing an image into a sequence of tokens and gets excellent results on image classification. To further overcome the quadratic complexity of original self-attention, Liu et al. [24] perform local windows with fixed size or window shift to help cross-window interaction. Liang et al. proposes an image restoration model, SwinIR [22] based on Swin Transformer [24]. Similarly, Uformer [46] imple + +![](images/6ff24b9af7a1cb0abb7690de950a025cd5e89d5ba223b9ef817388b8be218f30.jpg) +Element-wise multiplication + +![](images/66e13100157fffbddcaf4005c140bf24866040d93745662377a6a58b4e0f64a5.jpg) +Figure 2: Overview of our model architecture. The model is implemented in an encoder-decoder manner with multi-scale input and output. Each encoder/decoder (except for $\mathsf{EB}_0$ and $\mathsf{EB}_3$ ) consists of (a) specially designed residual spatially-compensated and spectrally-enhanced transformer block (RSS-T); (b) feature attention module (FAM)/convolution layer (Conv.). Cross-scale feature fusion based on squeeze and excitation block (CFF-SE) is explored to actively emphasize or suppress the features from encoders. + +![](images/924e59d9634ca8d320f780f928c8435b1cce5ed6516d69f8586312bb8c14b2da.jpg) +Upsample +Downsample +$\oplus$ Element-wise addition +$\mathbb{C}$ Concatenation +In Out Input / Output projection + +ments the Locally enhanced Window Transformer block in Unet structure and introduces a learnable multi-scale restoration modulator. All these methods have gained remarkable performance on image restoration tasks by mitigating the shortcomings of CNNs. But no one has addressed the corruption to image spatial information introduced by the widow partition strategy. And spectral feature, which is proven to capture long-range context effectively, has also been ignored in vision transformers. + +# 3. Method + +Overview The overall structure is in the manner of encoder-decoder with multi-scale inputs and outputs (Figure 2). Except for $\mathbb{E}\mathbb{B}_0$ and $\mathbb{D}\mathbb{B}_3$ , all encoder/decoder blocks contain a Feature Attention Module (FAM) or Convolution layer (Conv.) to refine aggregated features. Cross-scale feature fusion based on squeeze and excitation block (CFF-SE) is exploited to propagate information flow between encoders and decoders. To be specific, given a degraded image $\mathbf{I} \in \mathbb{R}^{3 \times H \times W}$ , multi-scale input set $\mathcal{I} = [\mathbf{I}_0, \mathbf{I}_1, \mathbf{I}_2, \mathbf{I}_3]$ , $\mathbf{I}_k \in \mathbb{R}^{3 \times \frac{H}{2^k} \times \frac{W}{2^k}}$ ( $k = 0, 1, 2, 3$ ) is obtained by down-sampling operations. Then low-level feature $\mathbf{X}_k \in \mathbb{R}^{2^k C \times \frac{H}{2^k} \times \frac{W}{2^k}}$ of + +input $\mathbf{I}_k$ is extracted by its own input projection module $In^k$ and will be fed into $\mathbb{E}\mathbb{B}_k$ . Similarly, the first three encoder blocks are followed by a down-sampling operation to guarantee the shape of feature map from $\mathbb{E}\mathbb{B}_{k - 1}$ is consistent with that of $\mathbf{X}_k$ . For feature reconstruction, an up-sampling operation is inserted between each two decoders. And multi-scale output set $\mathcal{O} = [\hat{\mathbf{I}}_0,\hat{\mathbf{I}}_1,\hat{\mathbf{I}}_2,\hat{\mathbf{I}}_3],\hat{\mathbf{I}}_l\in \mathbb{R}^{3\times \frac{H}{2^l}\times \frac{W}{2^l}}(l = 0,1,2,3)$ is generated by corresponding output projection module $Out^l$ and residual connection from input image $\mathbf{I}_l$ . Figure 2 illustrates the detailed structure of our RSS-T block. + +# 3.1. SW-MSA with Spatial Compensation + +Two main challenges exist to directly extend the existing Transformer architecture to GRR deblurring. First, because of the characteristic of GRR exposure mode, generated spatial-variant blur and brightness are closely correlated to image rows, while normal Transformer blocks do not consider this. They divide the image into squared window [22, 50, 46], leading to large blur and brightness variance within each window, which makes the rectification process more challenging. Second, standard vision Transformer architecture cuts the computational cost due to the usage of self-attention conducted on non-overlapping windows, which enables high dimensional input processing. + +![](images/cd6bbc3504fe3fba787a40016c6536058cd29c2ef3fe42db9cadfc515f76a9dd.jpg) +Figure 3: Two main components of our proposed RSS-T. (a) The proposed SW-MSA with Spatial Compensation includes two core designs: horizontal window partition strategy and hierarchical position encoding comprising window positional encoding (WinPE) and pixel positional encoding (PixPE). (b) Se-FFN. + +![](images/a60eccb0ea8b4f41f4bf71a4e71a8d961c6f3e5f7947822ccf1c9782bf83bf4e.jpg) + +But during window partition, image spatial information has been corrupted. That absolute spatial location is crucial for spatially-sensitive tasks. To address the aforementioned issues, we propose SW-MSA with spatial compensation, as shown in Figure 3a, which has two core designs: horizontal window partition and hierarchical positional encoding. + +Horizontal Window Partition As discussed above, common squared windows without distinguishing blur and brightness variance along the row and column dimensions make rectification more challenging. We update the original squared window as rectangular to better adapt row-dependent blur and brightness. These long and narrow horizontal windows will mitigate large variances within each widow. Bhojanapalli et al. [2] has depicted a kind of low-rank bottleneck, i.e., while increasing the number of heads seemingly gives the model more expressive power, at the same time, the head size is reduced, which can decrease the expressive power. When input token number $n$ is larger than head size $d_h$ , it will create a low-rank bottleneck and lose its ability to represent arbitrary context vectors. Thus, by following this rule and considering computational cost, we set the horizontal window size as (2, 16), (4, 16), (2, 32), and (4, 32). For better comparison, we also provide a normal squared window size of (8, 8) and an extreme case with a window size of (2, full). The 'full' means the window width equals the feature map width. Given input feature $\mathbf{X}_{in} \in \mathbb{R}^{B \times H \times W \times C'}$ , the partition process is represented as: + +$$ +\mathbf {X} _ {w} = \operatorname {W i n P a r t} \left(\mathbf {X} _ {i n}\right), \tag {1} +$$ + +where $\mathbf{X}_w\in \mathbb{R}^{BN_w\times H_w\times W_w\times C'}$ $H_{w},W_{w}$ denote height and width of windows and $N_{w} = H / H_{w}\times W / W_{w}$ , which is the total number of divided windows. + +Hierarchical Positional Encoding Window partition largely reduces computational cost but corrupts the image's + +global positional information crucial to GRR deblurring. So, we propose hierarchical positional encoding: 1) Window Position Encoding (WinPE) compensates for the absolute global location lost in the partition process. 2) Pixel Position Encoding (PixPE) enables context of the local pixel's relative position within each window. Chu et al. [7] proposes a positional encoding generator to obtain Conditional Position Encodings (CPE), which can be efficiently implemented with a 2D depth convolution with kernel $k$ ( $k \geq 3$ ) and $\frac{k - 1}{2}$ zero paddings. They have proven zero paddings here are important to make the model aware of the absolute positions of each token. Li et al. [21] use a similar positional encoding strategy to capture temporal position for videos. Here, we extend CPE to the 3D situation and propose our WinPE: + +$$ +\operatorname {W i n P E} \left(\mathcal {R} \left(\mathbf {X} _ {w}\right)\right) = \operatorname {D W C o n v} \left(\mathcal {R} \left(\mathbf {X} _ {w}\right)\right), \tag {2} +$$ + +where $\mathcal{R}(\cdot)$ is reshape operation and DWConv means 3D depthwise convolution with zero paddings. Each window's absolute position will be encoded by repeatedly applying depthwise convolution to the 3D feature space. Moreover, this encoding is dynamic and allows us to eliminate the constraint to the token's length when testing on different datasets, which perfectly matches the requirement of our window position encoding. The fixed size of the window will lead to the dynamic length of the window sequence once the image size changes. We also apply the relative position encoding [24, 38] for PixPE through a learnable parameter table $\mathcal{B} \in \mathbb{R}^{(2H_w - 1) \times (2W_w - 1)}$ . Overall, the formulation is: + +$$ +\mathbf {Y} = \mathbf {X} _ {i n} + \operatorname {M S A} \left(\ln \left(\mathbf {X} _ {e}\right)\right), +$$ + +$$ +\mathbf {X} _ {e} = \mathbf {X} _ {w} + \mathcal {R} (\operatorname {D W C o n v} (\mathcal {R} (\mathbf {X} _ {w}))), \tag {3} +$$ + +$$ +\operatorname {M S A} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) = \operatorname {S o f t M a x} \left(\mathbf {Q K} ^ {T} / \sqrt {d _ {h}} + \mathbf {B}\right) \mathbf {V}, +$$ + +where $\mathbf{Q}$ , $\mathbf{K}$ and $\mathbf{V}$ are projected query, key and value from input. $\mathbf{B}$ is the relative position bias, whose values are taken + +from $\mathcal{B}$ . $\mathrm{LN}(\cdot)$ means layer norm and $d_h$ denotes head size. + +# 3.2. Spectrally-enhanced FFN + +Researchers have proven the spectral feature effectively captures long-range dependencies. The Fast Fourier Convolution (FFC) operator in [4] could efficiently implement a non-local receptive field. Inspired by this work, we devised our Se-FFN based on depth-wise convolutional layer. The spectral branch captures the discrepancy between blurry and sharp image pairs as a supplemental part of self-attention for effective long-range dependency modeling. + +Se-FFN includes two parts: spectral branch and spatial branch (Figure 3b). The spatial part first applies layer norm to the input feature followed by two convolutions with $1 \times 1$ kernel and a $3 \times 3$ depth-wise convolution to capture local information. The spectral branch is formulated as (a) convert the input feature map $\mathbf{Y} \in \mathbb{R}^{B \times H \times W \times C'}$ into the frequency domain to obtain $\mathcal{F}(\mathbf{Y})$ . (b) Concatenate the real part and imaginary part of $\mathcal{F}(\mathbf{Y})$ along channel dimension to get the spectral feature $\tilde{\mathbf{Y}} \in \mathbb{R}^{B \times H \times W/2 \times 2C'}$ . (c) Exploit two $1 \times 1$ convolutional layers with the LeakyReLU function in between to process the spectral feature. (d) Split the processed spectral feature as real and imaginary parts, then convert them back to spatial domain $\mathbf{Y}_f \in \mathbb{R}^{B \times H \times W \times C'}$ by inverse Fast Fourier Transform (iFFT). Finally, the output of Se-FFN is $\mathbf{Z} = \mathbf{Y} + \mathbf{Y}_f + \mathbf{Y}_s$ , where $\mathbf{Y}_s$ is the output of spatial branch. + +# 3.3. Cross-scale Feature Fusion + +In most conventional U-net structures, the corresponding encoder and decoder are related by a skip connection to allow information flow between the same scale level. Later in [6], authors present an Asymmetric Feature Fusion (AFF) module to enable cross-scale interaction but without any control. So, we present our CFF-SE block. As shown in Figure 2, before the convolutional block, the input firstly goes through a SE module to control the information flow from the respective scale level of all encoders, which enables us to actively emphasize or suppress the features. Thus, each decoder can selectively exploit multi-scale features, resulting in improved deblurring performance as will be demonstrated in the experimental part. + +# 4. Experiments + +Learning-based methods for deblurring or RS correction are usually trained on synthetic data, which has limitations of generalization. We, therefore, built an image acquisition system to collect a real dataset named GRR-real. The details about the system, training implementation, and more experimental results are offered in supplemental materials. + +# 4.1. Comparison on GRR Deblurring + +Comparison with Other Models Since single GRR deblurring is addressed for the first time, there is no specially devised algorithm. So, we compare our model with previous state-of-the-art approaches from: + +- GS image-based deblurring by exploiting coarse-to-fine strategy: basic U-net [36], SRN [45], MIMO-UNet[6] and Uformer[46]. +- GS video-based deblurring by fusing neighboring temporal information: STR-CNN [12], DBN [42] and IFIRNN[26]. +- RS video-based correction through predicting pixelwise displacement: JCD [56] and DSUR [23].1 +- GRR video deblurring using a spatial-aware encoder with long-short term temporal information aggregator: NGS [47]. For fair comparisons, we not only provide the original setting with 8 frames as input $(\mathrm{NGS}_8)$ but also update the model with a single input $(\mathrm{NGS}_1)$ . + +After retraining all models, Table 1 depicts quantitative results on the GRR-real test dataset. Overall, it demonstrates our specially devised algorithm significantly outperforms models from RS correction and GS deblurring, even though they take multi frames as input and exploit temporal information. As discussed in Section 1, RS correction usually resorts to nontrivial motion estimation and warping process. When directly used for GRR, they performed worse than deblurring models. Furthermore, we observe that, for deblurring models, those with video clips as input may degrade sharply, even are no match for some approaches of a single input. A key reason is that these models are struggling to align the content of different video frames. If failed, the redundant inputs would confuse learning process instead. + +Although the tailored GRR video deblurring method $\mathrm{NGS}_8$ [47] achieved higher metrics than ours (30.03/0.90 vs. 28.64/0.90), it heavily relies on temporal correlation of 8 frames and the performance decreased drastically when given single input $(\mathrm{NGS}_1)$ . In contrast to RSS-T, the demand of $\mathrm{NGS}_8$ for 8 consecutive frames in training or testing can be difficult to be satisfied, sometimes even impossible due to buffer restrictions. Besides, sensitivity to video capturing settings (e.g., frame rate and deadtime between frames) will lead to additional generalization issues. We also computed the complexity of all algorithms, which shows that our model is at medium level and outperforms SOTA RS correction or Transformer-based models. From qualitative + +![](images/6c07a62771f7d08e2a733bdcf45ea6ff87c24c889ccbb90d3f27042dc3a1eb3f.jpg) + +![](images/29abe3ac19cc63e5ff08bcd78ef5972a8fdcf5ad07d9dc6970ce8cb3859da507.jpg) + +![](images/20b2157bdc87a47c7cd87749cd790a5f7bf0d966f8f0a77750768a3e4783d90c.jpg) + +![](images/a4a7996a66ac7811165c1390bfdf59d4308b823d082138114dbb930c21604fbe.jpg) + +![](images/c3f4b062b1a6841df49321bf3609fd00d439a1ce77d2074476673903617a8423.jpg) + +![](images/865db4c06b59df0feede0bf35d7300e432c04b6052c71c99c6351e68d3b74cfb.jpg) +Input + +![](images/422980c6580793f25e092ddd3499328134aacf6cbaefc02abf15057b5a463cdf.jpg) +U-net [36] + +![](images/2513361c4ace0648e3623b5bed443956abd835b42ad3648a80c2b4ec45afc879.jpg) +SRN [45] + +![](images/78b567243fa9597b7cd3a856cd4b05ff029977f7fa04ee5887b05f48fb4c7e80.jpg) +MIMO [6] + +![](images/b5b62b503d635ea90672aab6abd95d0326b91a5d968451e7ec5be2abe477f4a8.jpg) +MIMO+ [6] + +![](images/ea96ddbaf44480c939644e3ca01ea9f34612c10e53f972d5bc05ae3192ddf076.jpg) +Uformer [46] +DSUR [23] +Figure 4: Qualitative results on the GRR-real dataset. We compare our method with state-of-the-art methods from RS correction, GS deblurring, and GRR video deblurring. + +![](images/167ae600df2a4a4ab20cfbb7f443d24d05712ec1328b9f6e4a8f1573ec5777d4.jpg) +STR-CNN [12] +NGS1 [47] + +![](images/ba9069f537e5479b8e23072ec87a3029b7e1ae742d9010c1534dd17ea6a7beef.jpg) +DBN [42] +NGS8 [47] + +![](images/b1da72b802b6b386806df53af1d37082283917e12fb881276f37a3a66c7dc0f6.jpg) +IFIRNN [26] +RSS-T + +![](images/5238d6ea6659bdb5d269380b0d8c5c8482c09df658a514805db104c79c20b2df.jpg) +JCD [56] +GT + +Table 1: Quantitative comparison on GRR-real dataset. We compare our method with state-of-the-art motion deblurring, RS correction, and GRR video deblurring approaches. According to 'exposure mode/Input number', they could be divided as 'GS/Single', 'GS/Multi', 'RS/Multi', 'GRR/Multi', and 'GRR/single'. The performance is measured with mean PSNR/SSIM (↑). 'Full' denotes evaluation using full-size frames, while 'Top', 'Middle', and 'Bottom' represent using only a patch of images from the top, middle, and bottom areas. + +
MethodMode / InputEffectivenessEfficiency
TopMiddleBottomFullTime (s)Params (M)FLOPs (T)
Input-15.12 / 0.6721.83 / 0.7820.08 / 0.7617.61 / 0.74---
U-net [36]26.34 / 0.9223.83 / 0.8522.74 / 0.8323.57 / 0.870.00629.500.072
SRN [45]25.11 / 0.7824.34 / 0.7224.08 / 0.7223.84 / 0.740.014010.250.509
MIMO [6]GS / Single27.16 /0.9024.55 / 0.8224.04 / 0.7924.48 / 0.840.01236.810.315
MIMO+ [6]28.46 / 0.9126.31 / 0.8525.78 / 0.8326.12 / 0.860.024916.110.724
Uformer [46]25.04 / 0.9323.70 / 0.8622.20 / 0.8422.92 / 0.870.214720.630.257
STR-CNN [12]20.88 / 0.8323.28 / 0.7722.60 / 0.7521.39 / 0.780.01940.930.367
DBN [42]GS / Multi25.11 / 0.9025.90 / 0.8525.87 / 0.8224.97 / 0.850.022915.311.046
IFIRNN [26]27.19 / 0.9023.85 / 0.7823.21 / 0.7723.97 / 0.810.01351.640.581
JCD [56]RS / Multi27.64 / 0.9023.65 / 0.8019.76 / 0.7722.31 / 0.820.26258.670.326
DSUR [23]24.81 / 0.8723.87 / 0.7823.39 / 0.7623.35 / 0.800.30183.900.225
NGS8 [47]GRR / Multi31.71 / 0.9330.54 / 0.9029.12 / 0.8730.03 / 0.900.02338.671.266
NGS1 [47]GRR / Single26.61 / 0.8923.53 / 0.7822.83 / 0.7623.67 / 0.810.01874.560.083
RSS-T30.90 / 0.9328.60 / 0.8827.86 / 0.8628.64 / 0.900.147911.340.176
+ +results in Figure 4, although the images obtained by the existing networks exhibit mitigated distortions compared to the input, local details and structures were not sufficiently corrected, whereas our method produces sharper images that even visually outperform $\mathrm{NGS}_8$ . + +Thirdparty Evaluation To further prove the practical value + +of our method, we evaluate it on another GRR dataset from [47]. As shown in Table 2, our method still outperforms other algorithms and is just slightly inferior to $\mathrm{NGS}_8$ with as many as eight input frames (26.62/0.86 vs. 27.29/0.85). + +Table 2: Comparison on another GRR dataset [47]. + +
MethodSET-II
FullTopMiddleBottom
Input17.82 / 0.7323.64 / 0.7721.45 / 0.7715.54 / 0.66
SRN [45]25.05 / 0.8124.32 / 0.7925.65 / 0.8127.02 / 0.83
STRCNN [12]22.59 / 0.8122.99 / 0.7923.46 / 0.8123.66 / 0.83
DBN [42]22.57 / 0.8123.24 / 0.8023.81 / 0.8123.24 / 0.82
IFIRNN[26]25.17 / 0.8224.77 / 0.8025.62 / 0.8126.94 / 0.84
ESTRNN [54]22.72 / 0.8323.42 / 0.8126.03 / 0.8322.86 / 0.83
DSUR [23]22.50 / 0.8022.49 / 0.7823.87 / 0.8123.38 / 0.83
JCD [56]25.33 / 0.8024.77 / 0.7825.71 / 0.8027.43 / 0.83
NGS8 [47]27.29 / 0.8526.96 / 0.8427.57 / 0.8528.35 / 0.86
NGS1 [47]25.19 / 0.7924.48 / 0.7725.53 / 0.7927.27 / 0.82
RSS-T26.62 / 0.8625.90 / 0.8426.82 / 0.8628.43 / 0.87
+ +Table 3: Architecture ablation on our GRR-real dataset. + +
PEFFNFusionPSNR / SSIM
WinPEPixPELeFNNSe-FNNAFFCFF-SE
v1✓✓25.69 / 0.85
v2✓✓✓26.50 / 0.86
v3✓✓✓26.10 / 0.86
v4✓✓✓✓26.83 / 0.87
v5✓✓✓✓27.24 / 0.87
RSS-T✓✓✓✓28.64 / 0.90
+ +# 4.2. Ablation Study + +Model Architecture Ablation We conducted experiments to analyze the effectiveness of the core components of our RSS-T. The hierarchical position encoding consists of WinPE and PixPE. The feed-forward network is also implemented by LeFFN [46] to compare with our Se-FNN (v5 vs. RSS-T). As for the cross-scale fusion part, AFF [46] and our CFF-SE are used to explore the performance difference (v4 vs. v5). All the experimental results are listed in Table 3. From v1, v2, v3, and v4, when WinPE and PixPE are separately presented in baseline model v1, the performance goes up but the combination of them achieved the best metrics. Our Se-FNN and CFF-SE modules contributed to the further performance improvement of PSNR by $+1.81\mathrm{dB}$ compared with the LeFFN and AFF. + +Window Size Ablation We introduced three types of window size (i.e., horizontal, squared, and full-width window) and clarified the advantages of the horizontal window for GRR deblurring in Section 3.1. Here, we present experimental results to support the claim. As Figure 5 shows, all horizontal window settings are superior to the standard squared one from quantitative and qualitative perspectives. The extreme case (2, full) is severely trapped in low-rank bottleneck [2] as the pixel number of this window is far more than the head size. + +![](images/28cc50ef74d74788349acbfa0f3b6a480cb7e500e5b0d2ff2e6a5989b3cc1d23.jpg) +Figure 5: Window size ablation. Experimental results of all different window-size settings on our GRR-real dataset. + +![](images/3e26407c8df00d9f090c6997474d827bf45a94f127e1aaa19fc34b0c497ecaa5.jpg) +Figure 6: Temporal alignment of three shutters. When synthesizing datasets, we strictly align the exposure of three modes. + +# 4.3. Comparison on Three Shutter Modes + +Until now, a direct comparison of three shutter modes is still missing, although [47] has demonstrated that GRR has its superiority against RS in obtaining sharp images in the scenario of video restoration. Here, we evaluate them on single image Restoration. For a fair comparison, we synthesize strictly aligned images virtually captured by three modes (RS, GRR, and blur) for each scene by following the protocol of [28, 45, 6] (details in the supplemental material), and compare with SOTA methods for each mode. + +When compared with RS mode, Table 4 shows that GRR combined with our RSS-T model has gained significant advantages in obtaining single sharp image, even though DSUR and JCD take multiple frames as input. These results are consistent with the discussion in [47] that the GRR mode enables us to convert the RS correction problem into a more tractable one. Therefore, given an RS camera that supports GRR mode, we recommend switching to GRR mode for better restoration from a single input, if losing some capture speed of the original RS is permissible. Also, the exposure should be adjusted to avoid saturation of bottom rows. + +We additionally present comparisons with row- + +Table 4: Quantitative comparison on three modes. The subscript of each method denotes the number of input frames. The reason we choose multi-frame RS correction is discussed in Section 4.1. MIMO+ and Uformer are trained on blur mode. DSUR and JCD are trained on the RS counterpart, while RSS-T is trained on GRR mode. + +
MethodFullTopMiddleBottom
DSUR2 [23]24.51 / 0.8128.42 / 0.8923.84 / 0.7824.21 / 0.77
JCD3 [56]23.52 / 0.7525.85 / 0.7923.87 / 0.7523.18 /0.70
RSS-T127.47 / 0.8927.40 / 0.9027.9 / 0.8928.10 / 0.89
MIMO+1[6]25.16 / 0.8626.75 / 0.8725.54 / 0.8525.58 / 0.85
Uformer1[46]24.28 / 0.8425.54 / 0.8624.26 / 0.8324.57 / 0.83
+ +independent deblurring issues under identical scenes. For fairness, we artificially elongate the exposure time of a GS camera to synthesize blurred images. As shown in Figure 6, the exposure time of GS is set to be the same as that of the full scanning period of RS and GRR. From Table 4, row-dependent blur from GRR is more tractable than blur generated by a GS camera. Because of the characteristics of GRR exposure, it implicitly encodes the temporal ordering of latent frames that mitigate the motion ambiguity [13] of blur images. And local details from the top sharper yet noisier rows further enhance the reconstruction performance. The practical implication on usage is that, given a GS camera in hand that supports GRR mode, we recommend switching to the GRR mode for single image deblurring with our proposed algorithm. Note that there is no additional cost to pay here since GRR mode will not slow down the original GS frame rate nor cause additional saturation since the longest exposure time of GRR is the same as the GS exposure time (Figure 6). + +# 5. Conclusion + +In this paper, we analyzed image degradations for conventional GS and RS cameras and further highlighted an in-between exposure mode, named GRR, arising from the global reset feature equipped with many RS sensors. Instead of RS distortion, GRR gives rise to a specialized blurring effect with row-dependent blur magnitude and brightness level. Based on the characteristics of GRR degradation, an original Transformer block with spatial compensation and spectral enhancement was devised, which has the advantage of being able to rectify the GRR degradation. The proposed hierarchical positional encoding well preserved the absolute and relative position context and the spectral branch of the feed-forward network enhanced the model's ability to capture long-range discrepancies. The experimental results also verified that our tailored model is effective. + +Potential Application As shown in Figure 7, GRR is mainly intended for industrial usage in controlled dark environment, + +![](images/2e0fb29440ee26553e76989229d118e2737cda86a3c3a8cbe48d801dc53db3f9.jpg) +(a) + +![](images/97902040cc3210cba77bc08f0dd8af75c1bfe8f967c575c607d4b066791e2982.jpg) +(b) + +![](images/31f922b570d546f42bc916c8b346add328bf96ea0c49e21224f9e827dd1c73af.jpg) +Figure 7: Application scenario of GRR mode. (a) Illustrated application scenario. (b) Synchronization between illumination and cameras. + +such as parts inspection on a conveyor belt, together with strobe illuminant. In this scenario, GRR with synchronized illumination can capture distortion-free and blur-free images, with lower cost than using a GS sensor, and faster speed than using a RS sensor (Because, in this case, the exposure time of RS has to be elongated such that the first and last rows have sufficient overlap). Instead, without using strobe illumination, GRR boosted with our approach can still restore a sharp image. This will further reduce production and maintenance cost in practice. + +Limitations When the illumination is bright enough, with appropriate exposure, GS can capture a sharp image without blur. The bottom parts of GRR can do the same, yet the top parts will be darker and thus noiser, or the top parts can be sharp, yet the the bottom parts will be blurry. In this scenario, GRR is clearly worse than GS, and our algorithm can help reduce the performance gap. On the other hand, RS will not have distortion and the image will be sharp and clean when dealing with completely static scenes. Yet GRR will have non-uniform brightness, although blur will not happen. Our algorithm can compensate this non-uniformity, yet GRR is worse than RS in this particular scenario. + +Acknowledgement This research was supported in part by JSPS KAKENHI Grant Numbers 22H00529, 20H05951, 22H03620, 22H05015, JST AIP Acceleration Research JP-MJCR22U4, the Value Exchange Engineering, a joint research project between Mercari, Inc. and RIISE. + +# References + +[1] Cenek Albl, Zuzana Kukelova, Viktor Larsson, Michal Polic, Tomas Pajdla, and Konrad Schindler. From two rolling shutters to one global shutter. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2505-2513, 2020. 1, 3 +[2] Srinadh Bhojanapalli, Chulhee Yun, Ankit Singh Rawat, Sashank Reddi, and Sanjiv Kumar. Low-rank bottleneck in multi-head attention models. In International Conference on Machine Learning, pages 864-873. PMLR, 2020. 5, 8 +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End + +to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3 +[4] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolution. Advances in Neural Information Processing Systems, 33:4479-4488, 2020. 3, 6 +[5] Lu Chi, Guiyu Tian, Yadong Mu, Lingxi Xie, and Qi Tian. Fast non-local neural networks with spectral residual learning. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2142-2151, 2019. 3 +[6] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4641-4650, 2021. 3, 6, 7, 8, 9 +[7] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Conditional positional encodings for vision transformers. arXiv preprint arXiv:2102.10882, 2021. 5 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3 +[9] Bin Fan, Yuchao Dai, and Mingyi He. Sunet: symmetric undistortion network for rolling shutter correction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4541-4550, 2021. 1, 3 +[10] Hongyun Gao, Xin Tao, Xiaoyong Shen, and Jiaya Jia. Dynamic scene deblurring with parameter selective sharing and nested skip connections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3848-3856, 2019. 1 +[11] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 3 +[12] Tae Hyun Kim, Kyoung Mu Lee, Bernhard Scholkopf, and Michael Hirsch. Online video deblurring via dynamic temporal blending network. In Proceedings of the IEEE International Conference on Computer Vision, pages 4038-4047, 2017. 3, 6, 7, 8 +[13] Meiguang Jin, Givi Meishvili, and Paolo Favaro. Learning to extract a video sequence from a single motion-blurred image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6334-6342, 2018. 9 +[14] Yitzhak Katznelson. An introduction to harmonic analysis. Cambridge University Press, 2004. 3 +[15] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiri Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8183-8192, 2018. 3 +[16] Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8878-8887, 2019. 3 + +[17] Yizhen Lao and Omar Ait-Aider. A robust method for strong rolling shutter effects correction using lines with automatic feature selection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4795-4803, 2018. 1, 3, 6 +[18] Junyong Lee, Hyeongseok Son, Jaesung Rim, Sunghyun Cho, and Seungyong Lee. Iterative filter adaptive network for single image defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2034-2042, 2021. 3 +[19] Anat Levin. Blind motion deblurring using image statistics. Advances in Neural Information Processing Systems, 19, 2006. 3 +[20] Dongxu Li, Chenchen Xu, Kaihao Zhang, Xin Yu, Yiran Zhong, Wenqi Ren, Hanna Suominen, and Hongdong Li. Arvo: Learning all-range volumetric correspondence for video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7721-7731, 2021. 3 +[21] Kunchang Li, Yali Wang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatiotemporal representation learning. arXiv preprint arXiv:2201.04676, 2022. 5 +[22] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 2, 3, 4 +[23] Peidong Liu, Zhaopeng Cui, Viktor Larsson, and Marc Pollefeys. Deep shutter unrolling network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5941-5949, 2020. 1, 3, 6, 7, 8, 9 +[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 3, 5 +[25] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 1, 3 +[26] Seungjun Nah, Sanghyun Son, and Kyoung Mu Lee. Recurrent neural networks with intra-frame iterations for video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8102-8111, 2019. 3, 6, 7, 8 +[27] Mehdi Noroozi, Paramanand Chandramouli, and Paolo Favaro. Motion deblurring in the wild. In German conference on pattern recognition, pages 65-77. Springer, 2017. 3 +[28] Jihyong Oh and Munchurl Kim. Demfi: deep joint deblurring and multi-frame interpolation with flow-guided attentive correlation and recursive boosting. In European Conference on Computer Vision, pages 198-215. Springer, 2022. 8 +[29] Chandramouli Paramanand and Ambasamudram N Rajagopalan. Non-uniform motion deblurring for bilayer scenes. + +In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1115-1122, 2013. 3 +[30] Dongwon Park, Dong Un Kang, Jisoo Kim, and Se Young Chun. Multi-temporal recurrent neural networks for progressive non-uniform single image deblurring with incremental temporal training. In European Conference on Computer Vision, pages 327-343. Springer, 2020. 1 +[31] Pulak Purkait and Christopher Zach. Minimal solvers for monocular rolling shutter compensation under ackermann motion. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 903–911. IEEE, 2018. 1, 3, 6 +[32] Pulak Purkait, Christopher Zach, and Ales Leonardis. Rolling shutter correction in manhattan world. In Proceedings of the IEEE International Conference on Computer Vision, pages 882-890, 2017. 1, 3, 6 +[33] Kuldeep Purohit and AN Rajagopalan. Region-adaptive dense network for efficient motion deblurring. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11882-11889, 2020. 1 +[34] Vijay Rengarajan, Yogesh Balaji, and AN Rajagopalan. Unrolling the shutter: Cnn to correct motion distortions. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 2291-2299, 2017. 1, 3, 6 +[35] Vijay Rengarajan, Ambasamudram N Rajagopalan, and Rangarajan Aravind. From bows to arrows: Rolling shutter rectification of urban scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2773-2781, 2016. 1, 3, 6 +[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3, 6, 7 +[37] Christian J Schuler, Michael Hirsch, Stefan Harmeling, and Bernhard Schölkopf. Learning to deblur. IEEE transactions on pattern analysis and machine intelligence, 38(7):1439-1451, 2015. 3 +[38] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. arXiv preprint arXiv:1803.02155, 2018. 5 +[39] Wang Shen, Wenbo Bao, Guangtao Zhai, Li Chen, Xiongkuo Min, and Zhiyong Gao. Blurry video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5114-5123, 2020. 1 +[40] Hyeongseok Son, Junyong Lee, Sunghyun Cho, and Seungyong Lee. Single image defocus deblurring using kernel-sharing parallel atrous convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2642-2650, 2021. 3 +[41] Hyeongseok Son, Junyong Lee, Jonghyeop Lee, Sunghyun Cho, and Seungyong Lee. Recurrent video deblurring with blur-invariant motion estimation and pixel volumes. ACM Transactions on Graphics (TOG), 40(5):1-18, 2021. 3 +[42] Shuochen Su, Maurizio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE + +Conference on Computer Vision and Pattern Recognition, pages 1279-1288, 2017. 6, 7, 8 +[43] Jian Sun, Wenfei Cao, Zongben Xu, and Jean Ponce. Learning a convolutional neural network for non-uniform motion blur removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 769-777, 2015. 3 +[44] Yu-Wing Tai, Ping Tan, and Michael S Brown. Richardson-lucy deblurring for scenes under a projective motion path. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(8):1603-1618, 2010. 3 +[45] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8174-8182, 2018. 1, 3, 6, 7, 8 +[46] Zhendong Wang, Xiaodong Cun, Jianmin Bao, and Jianzhuang Liu. Uformer: A general u-shaped transformer for image restoration. arXiv preprint arXiv:2106.03106, 2021. 1, 2, 3, 4, 6, 7, 8, 9 +[47] Zhixiang Wang, Xiang Ji, Jia-Bin Huang, Shin'ichi Satoh, Xiao Zhou, and Yinqiang Zheng. Neural global shutter: Learn to restore video from a rolling shutter camera with global reset feature. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17794-17803, 2022. 2, 6, 7, 8 +[48] Patrick Wieschollek, Michael Hirsch, Bernhard Scholkopf, and Hendrik Lensch. Learning blind motion deblurring. In Proceedings of the IEEE international conference on computer vision, pages 231-240, 2017. 3 +[49] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34, 2021. 3 +[50] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. arXiv preprint arXiv:2111.09881, 2021. 2, 3, 4 +[51] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 3 +[52] Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Bjorn Stenger, Wei Liu, and Hongdong Li. Deblurring by realistic blurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2737-2746, 2020. 3 +[53] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xi-ang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881–6890, 2021. 3 +[54] Zhihang Zhong, Ye Gao, Yinqiang Zheng, and Bo Zheng. Efficient spatio-temporal recurrent neural network for video deblurring. In European Conference on Computer Vision, pages 191-207. Springer, 2020. 3, 8 + +[55] Zhisheng Zhong, Tiancheng Shen, Yibo Yang, Zhouchen Lin, and Chao Zhang. Joint sub-bands learning with clique structures for wavelet domain super-resolution. Advances in neural information processing systems, 31, 2018. 3 +[56] Zhihang Zhong, Yinqiang Zheng, and Imari Sato. Towards rolling shutter correction and deblurring in dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9219-9228, 2021. 6, 7, 8, 9 +[57] Bingbing Zhuang, Loong-Fah Cheong, and Gim Hee Lee. Rolling-shutter-aware differential sfm and image rectification. In Proceedings of the IEEE International Conference on Computer Vision, pages 948-956, 2017. 1, 3 +[58] Bingbing Zhuang, Quoc-Huy Tran, Pan Ji, Loong-Fah Cheong, and Manmohan Chandraker. Learning structure-and-motion-aware rolling shutter correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4551-4560, 2019. 1, 3, 6 \ No newline at end of file diff --git a/2023/Single Image Deblurring with Row-dependent Blur Magnitude/images.zip b/2023/Single Image Deblurring with Row-dependent Blur Magnitude/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..74d1821aec27f30fd732def27f4c0b5ad96c1e17 --- /dev/null +++ b/2023/Single Image Deblurring with Row-dependent Blur Magnitude/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec3fcb89a1056694f1bebfbd00725b3ee85d825dc1ad0ee952e25d85ea092cdc +size 716590 diff --git a/2023/Single Image Deblurring with Row-dependent Blur Magnitude/layout.json b/2023/Single Image Deblurring with Row-dependent Blur Magnitude/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9b48821948efc0084d129c050c0d86274a52fcc8 --- /dev/null +++ b/2023/Single Image Deblurring with Row-dependent Blur Magnitude/layout.json @@ -0,0 +1,9417 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 491, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 491, + 121 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 491, + 121 + ], + "type": "text", + "content": "Single Image Deblurring with Row-dependent Blur Magnitude" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 142, + 478, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 142, + 478, + 172 + ], + "spans": [ + { + "bbox": [ + 114, + 142, + 478, + 172 + ], + "type": "text", + "content": "Xiang Ji" + }, + { + "bbox": [ + 114, + 142, + 478, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 114, + 142, + 478, + 172 + ], + "type": "text", + "content": " Zhixiang Wang" + }, + { + "bbox": [ + 114, + 142, + 478, + 172 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 114, + 142, + 478, + 172 + ], + "type": "text", + "content": " Shin'ichi Satoh" + }, + { + "bbox": [ + 114, + 142, + 478, + 172 + ], + "type": "inline_equation", + "content": "^{2,1}" + }, + { + "bbox": [ + 114, + 142, + 478, + 172 + ], + "type": "text", + "content": " Yinqiang Zheng" + }, + { + "bbox": [ + 114, + 142, + 478, + 172 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 114, + 142, + 478, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 114, + 142, + 478, + 172 + ], + "type": "text", + "content": "The University of Tokyo, Japan \n" + }, + { + "bbox": [ + 114, + 142, + 478, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 114, + 142, + 478, + 172 + ], + "type": "text", + "content": "National Institute of Informatics, Japan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 174, + 527, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 174, + 527, + 186 + ], + "spans": [ + { + "bbox": [ + 67, + 174, + 527, + 186 + ], + "type": "text", + "content": "{jixiang, wangzhixiang}@g.ecc.u-tokyo.ac.jp, satoh@nii.ac.jp, yqzheng@ai.u-tokyo.ac.jp" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 239, + 290, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 290, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 290, + 526 + ], + "type": "text", + "content": "Image degradation often occurs during fast camera or object movements, regardless of the exposure modes: global shutter (GS) or rolling shutter (RS). Since these two exposure modes give rise to intrinsically different degradations, two restoration threads have been explored separately, i.e. motion deblurring of GS images and distortion correction of RS images, both of which are challenging restoration tasks, especially in the presence of a single input image. In this paper, we explore a novel in-between exposure mode, called global reset release (GRR) shutter, which produces GS-like blur but with row-dependent blur magnitude. We take advantage of this unique characteristic of GRR to explore the latent frames within a single image and restore a clear counterpart by only relying on these latent contexts. Specifically, we propose a residual spatially-compensated and spectrally-enhanced Transformer (RSS-T) block for row-dependent deblurring of a single GRR image. Its hierarchical positional encoding compensates global positional context of windows and enables order-awareness of the local pixel's position, along with a novel feed-forward network that simultaneously uses spatial and spectral information for gaining mixed global context. Extensive experimental results demonstrate that our method outperforms the state-of-the-art GS deblurring and RS correction methods on single GRR input." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 559, + 288, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 288, + 644 + ], + "type": "text", + "content": "As a fundamental research task of computer vision, image restoration has been explored for many years, aiming to recover a high-quality clean image from its corrupted counterpart by removing undesired degradation. Fast or even extreme motion, which is a major inducement to image degeneration and highly correlated to the camera's exposure mode: global shutter (GS) and rolling shutter (RS)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 644, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 644, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 644, + 288, + 693 + ], + "type": "text", + "content": "In a GS image sensor, all pixels are reset simultaneously and immediately charged with exposure, then stored for sequential readout (Figure 1a). In the presence of fast and sudden motion, blurring effects will happen. [25, 39] depict" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 213, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 213, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 545, + 239 + ], + "type": "text", + "content": "this motion-blurred output " + }, + { + "bbox": [ + 305, + 213, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^B" + }, + { + "bbox": [ + 305, + 213, + 545, + 239 + ], + "type": "text", + "content": " as an average of " + }, + { + "bbox": [ + 305, + 213, + 545, + 239 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 213, + 545, + 239 + ], + "type": "text", + "content": " successive latent sharp frames " + }, + { + "bbox": [ + 305, + 213, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i^G" + }, + { + "bbox": [ + 305, + 213, + 545, + 239 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 342, + 244, + 508, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 244, + 508, + 277 + ], + "spans": [ + { + "bbox": [ + 342, + 244, + 508, + 277 + ], + "type": "interline_equation", + "content": "\\mathbf {I} ^ {B} = \\frac {1}{N} \\sum_ {i = 0} ^ {N - 1} \\mathbf {I} _ {i} ^ {G}, i = 0, 1, \\dots , N - 1.", + "image_path": "d468792830260834969188aa18a9e5d43fafb3526477cd162d86ea972114d9f6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 283, + 547, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 547, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 547, + 378 + ], + "type": "text", + "content": "In contrast, RS cameras expose image pixels row by row (including reset and readout), leading to an invariant time delay between consecutive scanlines (Figure 1b). Therefore, RS distortion effects, also known as jello effects, will appear if the camera is moving during the image acquisition, which is formulated in [23] as " + }, + { + "bbox": [ + 304, + 283, + 547, + 378 + ], + "type": "inline_equation", + "content": "[\\mathbf{I}^R]_i = [\\mathbf{I}_i^G]_i" + }, + { + "bbox": [ + 304, + 283, + 547, + 378 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 283, + 547, + 378 + ], + "type": "inline_equation", + "content": "[\\mathbf{I}_i^G]_i" + }, + { + "bbox": [ + 304, + 283, + 547, + 378 + ], + "type": "text", + "content": " is an operator that extracts the " + }, + { + "bbox": [ + 304, + 283, + 547, + 378 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 283, + 547, + 378 + ], + "type": "text", + "content": " row from the " + }, + { + "bbox": [ + 304, + 283, + 547, + 378 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 283, + 547, + 378 + ], + "type": "text", + "content": " latent GS frame " + }, + { + "bbox": [ + 304, + 283, + 547, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i^G" + }, + { + "bbox": [ + 304, + 283, + 547, + 378 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 283, + 547, + 378 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 283, + 547, + 378 + ], + "type": "text", + "content": " is also the height of RS image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 379, + 548, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 548, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 548, + 641 + ], + "type": "text", + "content": "Although image deblurring itself is highly involved, it is believed that restoring clear images from single degraded captures of GS cameras is much more tractable than that of RS cameras as mentioned in [23]. Intuitively, motion blur removal process aims to disentangle target GS frame from redundant input without sophisticated displacement estimation or speculation of unknown context. On the contrary, RS correction needs to shift pixels of each scanline to one virtual GS canvas by building pixel-level correspondences. Moreover, the rectified output usually suffers from missing boundaries not directly captured during RS exposure. Methodologically, the coarse-to-fine network design has gained remarkable performance for single image deblurring [25, 30, 33, 45, 10, 46]. On the contrary, due to the ill-posed nature, single image RS correction relies heavily on strong prior assumptions, explicitly [35, 17, 32, 31] or implicitly [34, 58], which limits their applicability to real scenarios. So, recent research has moved onto multi-image RS correction to bypass the ill-posedness. But both classical [57, 1] and learning-based methods [23, 9] still cannot work well under complex dynamic scenes or drastic camera motions because of nontrivial pixel alignment." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 548, + 714 + ], + "type": "text", + "content": "It is largely overlooked in computer vision that there exists an in-between shutter mode, called global reset release (GRR) shutter, which can be easily enabled in some image sensors that originally work in GS mode (e.g., EV76C560 in EO-1312C Camera) or in RS mode (e.g., IMX178 in BFS-U3-63S4C camera). As shown in Figure 1c, GRR" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12269" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 547, + 289 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 547, + 289 + ], + "type": "image", + "image_path": "251c803ab291b0c604ee2c1c78c148e794b9b3724baefc173a263118d1ff1b0f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 304, + 548, + 389 + ], + "lines": [ + { + "bbox": [ + 45, + 304, + 548, + 389 + ], + "spans": [ + { + "bbox": [ + 45, + 304, + 548, + 389 + ], + "type": "text", + "content": "Figure 1: Different exposure modes and corresponding degrading processes induced by fast motions. (a) GS exposes all pixels simultaneously, and it causes the output with row-independent blur that can be interpreted as an average of the latent sharp frames. (b) RS captures pixels scanline by scanline, featured in higher frame rate, and lower cost than GS. However, RS sensors are prone to row-dependent distortion, which can be more challenging to correct than deblurring. (c) GRR begins exposure of all pixels simultaneously but ends row by row, causing blurring effects with row-dependent blur magnitude. A specialized deblurring task is required but without the need for RS distortion correction. GRR also has the advantage of reducing the flickering effect of artificial illuminants, such as green traffic lights." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 409, + 289, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 289, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 289, + 636 + ], + "type": "text", + "content": "starts to expose all pixels simultaneously as GS does but reads out signals scanline by scanline like conventional RS does. Apparently, GRR leads to blurring effects. However, different from GS blur, the blur magnitude of GRR (the brightness level and Signal-Noise-Ratio as a consequence) is row-dependent. Very recently, Wang et al. [47] proposed to enable this global reset feature in RS sensors that turns RS correction into a more tractable deblurring problem and reported superior results of video deblurring in GRR mode. We note the cost to pay for this performance gain by switching from RS to GRR mode includes 1) the superiority of RS mode over GS mode in capture speed is sacrificed since GRR does not allow early scanning of the next frame before the end of the current frame; 2) scanning rows in the bottom are more likely to be saturated, and the standard automatic exposure mechanism should be adjusted accordingly. Therefore, in the presence of multiple consecutive frames (e.g. 8 frames are used in [47]), the authentic benefits of using GRR mode are slightly questionable." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "content": "In this paper, we deal with a more challenging scenario with a single input image, which is practically meaningful in avoiding well-known flaws of using consecutive frames, including 1) Larger buffer capacity is needed; 2) Sensitivity to video recording settings, like frame rate and deadtime between frames, leading to additional generalization issues;" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 409, + 547, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 409, + 547, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 409, + 547, + 433 + ], + "type": "text", + "content": "3) Temporal alignment with severe degradation is still challenging." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 434, + 547, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 547, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 547, + 493 + ], + "type": "text", + "content": "By fully considering its unique exposure characteristic of GRR mode, we bring single GRR image deblurring into the community. Mathematically, row-dependent blur and brightness caused by varying exposure time of different scanlines could be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 372, + 501, + 479, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 501, + 479, + 534 + ], + "spans": [ + { + "bbox": [ + 372, + 501, + 479, + 534 + ], + "type": "interline_equation", + "content": "\\mathbf {I} ^ {R G} = \\mathbf {I} _ {0} ^ {G} + \\delta \\sum_ {i = 1} ^ {N - 1} \\left[ \\mathbf {I} _ {i} ^ {G} \\right] _ {i:},", + "image_path": "7a0159952f9ae56a15a9f3dee87d5b021b7a9c8f135b74f10aed3658ae8e7ff9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 544, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 548, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 544, + 548, + 714 + ], + "type": "inline_equation", + "content": "[\\mathbf{I}_i^G ]_i" + }, + { + "bbox": [ + 304, + 544, + 548, + 714 + ], + "type": "text", + "content": " is an operator that extracts image patch of " + }, + { + "bbox": [ + 304, + 544, + 548, + 714 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 544, + 548, + 714 + ], + "type": "text", + "content": " row to the end from the " + }, + { + "bbox": [ + 304, + 544, + 548, + 714 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 544, + 548, + 714 + ], + "type": "text", + "content": " latent GS frame " + }, + { + "bbox": [ + 304, + 544, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i^G" + }, + { + "bbox": [ + 304, + 544, + 548, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 544, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 544, + 548, + 714 + ], + "type": "text", + "content": " denotes the ratio between readout time and the first scanline's exposure duration. As a result, directly applying existing GS deblurring algorithms on GRR deblurring may give rise to an adaptation issue. On the other hand, Transformers have shown significant performance gains on image restoration task [22, 50] by mitigating the shortcomings of CNNs (i.e., limited receptive field and content-independent kernel). Although shifted widow strategy [22, 50, 46, 8, 24] has largely reduced computational loads and made it possible to process high dimensional input, the introduced severe corruption to image spatial information has not been addressed, especially for spatially-sensitive tasks (e.g., GRR deblurring). Besides," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12270" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "according to the spectral convolution theorem [14], updating a single value in the spectral domain globally affects all original data, which has been implemented for non-local receptive field and proven to be effective in capturing long-range context from frequency domain [4, 5, 55]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 290, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 290, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 290, + 396 + ], + "type": "text", + "content": "We aim to leverage the capability of self-attention and U-net [36] structure with multi-scale input/output to tackle the single GRR deblurring task. To this end, we propose shifted horizontal window-based Transformer block with spatial compensation and spectral enhancement, which consists of three main components: 1) Shifted Window-based Multi-head Self-Attention(SW-MSA) with horizontal partition strategy and hierarchical positional encoding. We update original squared window as rectangular one to better adapt row-dependent blur and brightness. Meanwhile, hierarchical positional encoding compensates global positional context of windows lost in window partition and enables order-awareness of local pixel's position within each window. 2) Spectrally-enhanced Feed-Forward Network (SeFFN) based on depth-wise convolutional layer. The spectral branch captures the discrepancy between blurry and sharp image pairs from the perspective of frequency domain, as a complementary part of self-attention. 3) Cross-scale Feature Fusion module based on Squeeze and Excitation block (CFFSE) [11], which enables us to actively emphasize or suppress the features from encoders with different scales. The main contributions of this paper are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 402, + 288, + 633 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 59, + 402, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 402, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 59, + 402, + 287, + 462 + ], + "type": "text", + "content": "- We propose an original Transformer block with spatial compensation and spectral enhancement to address single GRR deblurring by fully exploring the latent frames, which further validates the advantages of GRR exposure mode over its RS counterpart." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 469, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 469, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 59, + 469, + 287, + 553 + ], + "type": "text", + "content": "- Horizontal partition strategy and hierarchical positional encoding are used to compensate global positional context of windows and enable order-awareness of local pixel's position within each window. The spectrally-enhanced feed-forward network simultaneously uses spatial and spectral information for gaining mixed global context." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 560, + 288, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 560, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 59, + 560, + 288, + 633 + ], + "type": "text", + "content": "- To facilitate the development and evaluation of GRR deblurring, we take paired GRR/Sharp images to offer a new dataset captured under real scenes named GRR-real. Furthermore, by simulating data captured by three shutter modes in a comparable way, the benefits of using GRR for single image restoration are verified." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 642, + 134, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 134, + 654 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 134, + 654 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Deep Image Deblurring Deep learning methods have achieved significant success in multi [12, 26, 41, 54, 20, 41] or single [15, 18, 25, 27, 40, 45, 48, 51, 52, 44, 19, 29] image deblurring. Initially, researchers try to find the spatially" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "type": "text", + "content": "varying kernels of motion blur before estimating latent image and no exception for deep CNN methods [37, 43]. However, blur kernel is normally computed for all pixels leading to huge demands for memory and computation. Besides, kernel estimation process is overly sensitive to noise and saturation, which is not practical for real scenes. Later, DeepDeblur [25] firstly exploits kernel-free learning to directly construct the relation between blurry and sharp images in an end-to-end manner. And the coarse-to-fine strategy is also taken by following SRN [45] uses an encoder-decoder network with skip-connections for three scale levels to significantly reduce training difficulty and introduce obvious stability benefits. Adversarial training has also been extensively studied [15, 16]. Recently, Cho et al. [6] revisit the coarse-to-fine scheme and present a novel deblurring network (MIMO U-net) that can handle multi-scale blur efficiently. These methods have been proven to be effective and achieved remarkable performance in single image deblurring." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 303, + 548, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 548, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 548, + 567 + ], + "type": "text", + "content": "Rolling Shutter Correction RS correction methods with single image input could be divided into classical and learning based. Classical approaches heavily rely on strong prior assumptions, such as the scene is static, and the movement of the camera is limited to pure rotation or in-plane translation [35, 17, 32, 31]. Rengarajan et al. [35] propose to estimate motion by converting transformed curves to be straight based on the assumption that \"straight lines must remain straight\" and purely rotation of camera. [32] simplifies the real scene as Manhattan world to rectify the monocular RS image. These classical methods are barely applicable in real situations because of the restrictive prior assumptions. For learning-based methods, Rengarajan et al. [34] proposed a new CNN architecture based on long rectangular kernels to correct rolling shutter distortions by simply modeling the camera motion as translation+rotation polynomials. Zhuang et al. [58] extend [34] for learning to predict both the camera velocity and depth from a single RS image. The global shutter image is then recovered as a post-processing step. Recent research mainly focuses on multi-image RS correction [57, 1, 23, 9]. But those methods still cannot work under complex dynamic scenes or large camera motions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": "Vision Transformers As an alternative to CNN, the Transformer model has been adapted to numerous vision tasks such as image restoration [22, 50, 46], segmentation [49, 53], objection detection [3, 24]. The pioneering work of ViT [8] directly conducts self-attention on flattened patches by decomposing an image into a sequence of tokens and gets excellent results on image classification. To further overcome the quadratic complexity of original self-attention, Liu et al. [24] perform local windows with fixed size or window shift to help cross-window interaction. Liang et al. proposes an image restoration model, SwinIR [22] based on Swin Transformer [24]. Similarly, Uformer [46] imple" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12271" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 521, + 236 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 521, + 236 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 521, + 236 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 521, + 236 + ], + "type": "image", + "image_path": "6ff24b9af7a1cb0abb7690de950a025cd5e89d5ba223b9ef817388b8be218f30.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 407, + 246, + 496, + 257 + ], + "lines": [ + { + "bbox": [ + 407, + 246, + 496, + 257 + ], + "spans": [ + { + "bbox": [ + 407, + 246, + 496, + 257 + ], + "type": "text", + "content": "Element-wise multiplication" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 72, + 245, + 236, + 335 + ], + "blocks": [ + { + "bbox": [ + 72, + 245, + 236, + 335 + ], + "lines": [ + { + "bbox": [ + 72, + 245, + 236, + 335 + ], + "spans": [ + { + "bbox": [ + 72, + 245, + 236, + 335 + ], + "type": "image", + "image_path": "66e13100157fffbddcaf4005c140bf24866040d93745662377a6a58b4e0f64a5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 344, + 547, + 404 + ], + "lines": [ + { + "bbox": [ + 46, + 344, + 547, + 404 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 547, + 404 + ], + "type": "text", + "content": "Figure 2: Overview of our model architecture. The model is implemented in an encoder-decoder manner with multi-scale input and output. Each encoder/decoder (except for " + }, + { + "bbox": [ + 46, + 344, + 547, + 404 + ], + "type": "inline_equation", + "content": "\\mathsf{EB}_0" + }, + { + "bbox": [ + 46, + 344, + 547, + 404 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 344, + 547, + 404 + ], + "type": "inline_equation", + "content": "\\mathsf{EB}_3" + }, + { + "bbox": [ + 46, + 344, + 547, + 404 + ], + "type": "text", + "content": ") consists of (a) specially designed residual spatially-compensated and spectrally-enhanced transformer block (RSS-T); (b) feature attention module (FAM)/convolution layer (Conv.). Cross-scale feature fusion based on squeeze and excitation block (CFF-SE) is explored to actively emphasize or suppress the features from encoders." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 244, + 245, + 399, + 335 + ], + "blocks": [ + { + "bbox": [ + 244, + 245, + 399, + 335 + ], + "lines": [ + { + "bbox": [ + 244, + 245, + 399, + 335 + ], + "spans": [ + { + "bbox": [ + 244, + 245, + 399, + 335 + ], + "type": "image", + "image_path": "924e59d9634ca8d320f780f928c8435b1cce5ed6516d69f8586312bb8c14b2da.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 260, + 449, + 271 + ], + "lines": [ + { + "bbox": [ + 408, + 260, + 449, + 271 + ], + "spans": [ + { + "bbox": [ + 408, + 260, + 449, + 271 + ], + "type": "text", + "content": "Upsample" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 408, + 275, + 457, + 285 + ], + "lines": [ + { + "bbox": [ + 408, + 275, + 457, + 285 + ], + "spans": [ + { + "bbox": [ + 408, + 275, + 457, + 285 + ], + "type": "text", + "content": "Downsample" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 408, + 289, + 482, + 298 + ], + "lines": [ + { + "bbox": [ + 408, + 289, + 482, + 298 + ], + "spans": [ + { + "bbox": [ + 408, + 289, + 482, + 298 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 408, + 289, + 482, + 298 + ], + "type": "text", + "content": " Element-wise addition" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 408, + 303, + 460, + 312 + ], + "lines": [ + { + "bbox": [ + 408, + 303, + 460, + 312 + ], + "spans": [ + { + "bbox": [ + 408, + 303, + 460, + 312 + ], + "type": "inline_equation", + "content": "\\mathbb{C}" + }, + { + "bbox": [ + 408, + 303, + 460, + 312 + ], + "type": "text", + "content": " Concatenation" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 408, + 318, + 503, + 327 + ], + "lines": [ + { + "bbox": [ + 408, + 318, + 503, + 327 + ], + "spans": [ + { + "bbox": [ + 408, + 318, + 503, + 327 + ], + "type": "text", + "content": "In Out Input / Output projection" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 426, + 289, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 289, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 289, + 533 + ], + "type": "text", + "content": "ments the Locally enhanced Window Transformer block in Unet structure and introduces a learnable multi-scale restoration modulator. All these methods have gained remarkable performance on image restoration tasks by mitigating the shortcomings of CNNs. But no one has addressed the corruption to image spatial information introduced by the widow partition strategy. And spectral feature, which is proven to capture long-range context effectively, has also been ignored in vision transformers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 551, + 103, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 103, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 103, + 562 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 577, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 577, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 577, + 289, + 715 + ], + "type": "text", + "content": "Overview The overall structure is in the manner of encoder-decoder with multi-scale inputs and outputs (Figure 2). Except for " + }, + { + "bbox": [ + 46, + 577, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{E}\\mathbb{B}_0" + }, + { + "bbox": [ + 46, + 577, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 577, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{D}\\mathbb{B}_3" + }, + { + "bbox": [ + 46, + 577, + 289, + 715 + ], + "type": "text", + "content": ", all encoder/decoder blocks contain a Feature Attention Module (FAM) or Convolution layer (Conv.) to refine aggregated features. Cross-scale feature fusion based on squeeze and excitation block (CFF-SE) is exploited to propagate information flow between encoders and decoders. To be specific, given a degraded image " + }, + { + "bbox": [ + 46, + 577, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 46, + 577, + 289, + 715 + ], + "type": "text", + "content": ", multi-scale input set " + }, + { + "bbox": [ + 46, + 577, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = [\\mathbf{I}_0, \\mathbf{I}_1, \\mathbf{I}_2, \\mathbf{I}_3]" + }, + { + "bbox": [ + 46, + 577, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 577, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_k \\in \\mathbb{R}^{3 \\times \\frac{H}{2^k} \\times \\frac{W}{2^k}}" + }, + { + "bbox": [ + 46, + 577, + 289, + 715 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 577, + 289, + 715 + ], + "type": "inline_equation", + "content": "k = 0, 1, 2, 3" + }, + { + "bbox": [ + 46, + 577, + 289, + 715 + ], + "type": "text", + "content": ") is obtained by down-sampling operations. Then low-level feature " + }, + { + "bbox": [ + 46, + 577, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_k \\in \\mathbb{R}^{2^k C \\times \\frac{H}{2^k} \\times \\frac{W}{2^k}}" + }, + { + "bbox": [ + 46, + 577, + 289, + 715 + ], + "type": "text", + "content": " of" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 426, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 545 + ], + "type": "text", + "content": "input " + }, + { + "bbox": [ + 304, + 426, + 547, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_k" + }, + { + "bbox": [ + 304, + 426, + 547, + 545 + ], + "type": "text", + "content": " is extracted by its own input projection module " + }, + { + "bbox": [ + 304, + 426, + 547, + 545 + ], + "type": "inline_equation", + "content": "In^k" + }, + { + "bbox": [ + 304, + 426, + 547, + 545 + ], + "type": "text", + "content": " and will be fed into " + }, + { + "bbox": [ + 304, + 426, + 547, + 545 + ], + "type": "inline_equation", + "content": "\\mathbb{E}\\mathbb{B}_k" + }, + { + "bbox": [ + 304, + 426, + 547, + 545 + ], + "type": "text", + "content": ". Similarly, the first three encoder blocks are followed by a down-sampling operation to guarantee the shape of feature map from " + }, + { + "bbox": [ + 304, + 426, + 547, + 545 + ], + "type": "inline_equation", + "content": "\\mathbb{E}\\mathbb{B}_{k - 1}" + }, + { + "bbox": [ + 304, + 426, + 547, + 545 + ], + "type": "text", + "content": " is consistent with that of " + }, + { + "bbox": [ + 304, + 426, + 547, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_k" + }, + { + "bbox": [ + 304, + 426, + 547, + 545 + ], + "type": "text", + "content": ". For feature reconstruction, an up-sampling operation is inserted between each two decoders. And multi-scale output set " + }, + { + "bbox": [ + 304, + 426, + 547, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{O} = [\\hat{\\mathbf{I}}_0,\\hat{\\mathbf{I}}_1,\\hat{\\mathbf{I}}_2,\\hat{\\mathbf{I}}_3],\\hat{\\mathbf{I}}_l\\in \\mathbb{R}^{3\\times \\frac{H}{2^l}\\times \\frac{W}{2^l}}(l = 0,1,2,3)" + }, + { + "bbox": [ + 304, + 426, + 547, + 545 + ], + "type": "text", + "content": " is generated by corresponding output projection module " + }, + { + "bbox": [ + 304, + 426, + 547, + 545 + ], + "type": "inline_equation", + "content": "Out^l" + }, + { + "bbox": [ + 304, + 426, + 547, + 545 + ], + "type": "text", + "content": " and residual connection from input image " + }, + { + "bbox": [ + 304, + 426, + 547, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_l" + }, + { + "bbox": [ + 304, + 426, + 547, + 545 + ], + "type": "text", + "content": ". Figure 2 illustrates the detailed structure of our RSS-T block." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 552, + 503, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 503, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 503, + 564 + ], + "type": "text", + "content": "3.1. SW-MSA with Spatial Compensation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "Two main challenges exist to directly extend the existing Transformer architecture to GRR deblurring. First, because of the characteristic of GRR exposure mode, generated spatial-variant blur and brightness are closely correlated to image rows, while normal Transformer blocks do not consider this. They divide the image into squared window [22, 50, 46], leading to large blur and brightness variance within each window, which makes the rectification process more challenging. Second, standard vision Transformer architecture cuts the computational cost due to the usage of self-attention conducted on non-overlapping windows, which enables high dimensional input processing." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12272" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 72, + 378, + 228 + ], + "blocks": [ + { + "bbox": [ + 69, + 72, + 378, + 228 + ], + "lines": [ + { + "bbox": [ + 69, + 72, + 378, + 228 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 378, + 228 + ], + "type": "image", + "image_path": "cd6bbc3504fe3fba787a40016c6536058cd29c2ef3fe42db9cadfc515f76a9dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 235, + 547, + 271 + ], + "lines": [ + { + "bbox": [ + 46, + 235, + 547, + 271 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 547, + 271 + ], + "type": "text", + "content": "Figure 3: Two main components of our proposed RSS-T. (a) The proposed SW-MSA with Spatial Compensation includes two core designs: horizontal window partition strategy and hierarchical position encoding comprising window positional encoding (WinPE) and pixel positional encoding (PixPE). (b) Se-FFN." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 381, + 72, + 519, + 227 + ], + "blocks": [ + { + "bbox": [ + 381, + 72, + 519, + 227 + ], + "lines": [ + { + "bbox": [ + 381, + 72, + 519, + 227 + ], + "spans": [ + { + "bbox": [ + 381, + 72, + 519, + 227 + ], + "type": "image", + "image_path": "a60eccb0ea8b4f41f4bf71a4e71a8d961c6f3e5f7947822ccf1c9782bf83bf4e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 292, + 288, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 288, + 365 + ], + "type": "text", + "content": "But during window partition, image spatial information has been corrupted. That absolute spatial location is crucial for spatially-sensitive tasks. To address the aforementioned issues, we propose SW-MSA with spatial compensation, as shown in Figure 3a, which has two core designs: horizontal window partition and hierarchical positional encoding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 368, + 290, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 368, + 290, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 368, + 290, + 619 + ], + "type": "text", + "content": "Horizontal Window Partition As discussed above, common squared windows without distinguishing blur and brightness variance along the row and column dimensions make rectification more challenging. We update the original squared window as rectangular to better adapt row-dependent blur and brightness. These long and narrow horizontal windows will mitigate large variances within each widow. Bhojanapalli et al. [2] has depicted a kind of low-rank bottleneck, i.e., while increasing the number of heads seemingly gives the model more expressive power, at the same time, the head size is reduced, which can decrease the expressive power. When input token number " + }, + { + "bbox": [ + 46, + 368, + 290, + 619 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 368, + 290, + 619 + ], + "type": "text", + "content": " is larger than head size " + }, + { + "bbox": [ + 46, + 368, + 290, + 619 + ], + "type": "inline_equation", + "content": "d_h" + }, + { + "bbox": [ + 46, + 368, + 290, + 619 + ], + "type": "text", + "content": ", it will create a low-rank bottleneck and lose its ability to represent arbitrary context vectors. Thus, by following this rule and considering computational cost, we set the horizontal window size as (2, 16), (4, 16), (2, 32), and (4, 32). For better comparison, we also provide a normal squared window size of (8, 8) and an extreme case with a window size of (2, full). The 'full' means the window width equals the feature map width. Given input feature " + }, + { + "bbox": [ + 46, + 368, + 290, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{in} \\in \\mathbb{R}^{B \\times H \\times W \\times C'}" + }, + { + "bbox": [ + 46, + 368, + 290, + 619 + ], + "type": "text", + "content": ", the partition process is represented as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 122, + 628, + 287, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 628, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 122, + 628, + 287, + 641 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {w} = \\operatorname {W i n P a r t} \\left(\\mathbf {X} _ {i n}\\right), \\tag {1}", + "image_path": "20e0fb8ae3665f855b76ce31e9b6263b7747582d19179ffc2a607e5d08bffadb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 648, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 648, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 648, + 287, + 685 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 648, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_w\\in \\mathbb{R}^{BN_w\\times H_w\\times W_w\\times C'}" + }, + { + "bbox": [ + 47, + 648, + 287, + 685 + ], + "type": "inline_equation", + "content": "H_{w},W_{w}" + }, + { + "bbox": [ + 47, + 648, + 287, + 685 + ], + "type": "text", + "content": " denote height and width of windows and " + }, + { + "bbox": [ + 47, + 648, + 287, + 685 + ], + "type": "inline_equation", + "content": "N_{w} = H / H_{w}\\times W / W_{w}" + }, + { + "bbox": [ + 47, + 648, + 287, + 685 + ], + "type": "text", + "content": ", which is the total number of divided windows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Hierarchical Positional Encoding Window partition largely reduces computational cost but corrupts the image's" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 292, + 547, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 547, + 460 + ], + "type": "text", + "content": "global positional information crucial to GRR deblurring. So, we propose hierarchical positional encoding: 1) Window Position Encoding (WinPE) compensates for the absolute global location lost in the partition process. 2) Pixel Position Encoding (PixPE) enables context of the local pixel's relative position within each window. Chu et al. [7] proposes a positional encoding generator to obtain Conditional Position Encodings (CPE), which can be efficiently implemented with a 2D depth convolution with kernel " + }, + { + "bbox": [ + 304, + 292, + 547, + 460 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 292, + 547, + 460 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 292, + 547, + 460 + ], + "type": "inline_equation", + "content": "k \\geq 3" + }, + { + "bbox": [ + 304, + 292, + 547, + 460 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 304, + 292, + 547, + 460 + ], + "type": "inline_equation", + "content": "\\frac{k - 1}{2}" + }, + { + "bbox": [ + 304, + 292, + 547, + 460 + ], + "type": "text", + "content": " zero paddings. They have proven zero paddings here are important to make the model aware of the absolute positions of each token. Li et al. [21] use a similar positional encoding strategy to capture temporal position for videos. Here, we extend CPE to the 3D situation and propose our WinPE:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 347, + 467, + 545, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 467, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 347, + 467, + 545, + 480 + ], + "type": "interline_equation", + "content": "\\operatorname {W i n P E} \\left(\\mathcal {R} \\left(\\mathbf {X} _ {w}\\right)\\right) = \\operatorname {D W C o n v} \\left(\\mathcal {R} \\left(\\mathbf {X} _ {w}\\right)\\right), \\tag {2}", + "image_path": "1c3cd245eee126b176f3832e42be3406c0e6674f7892e858ecf6e21891d807bc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 487, + 547, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 547, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 547, + 631 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 487, + 547, + 631 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\cdot)" + }, + { + "bbox": [ + 304, + 487, + 547, + 631 + ], + "type": "text", + "content": " is reshape operation and DWConv means 3D depthwise convolution with zero paddings. Each window's absolute position will be encoded by repeatedly applying depthwise convolution to the 3D feature space. Moreover, this encoding is dynamic and allows us to eliminate the constraint to the token's length when testing on different datasets, which perfectly matches the requirement of our window position encoding. The fixed size of the window will lead to the dynamic length of the window sequence once the image size changes. We also apply the relative position encoding [24, 38] for PixPE through a learnable parameter table " + }, + { + "bbox": [ + 304, + 487, + 547, + 631 + ], + "type": "inline_equation", + "content": "\\mathcal{B} \\in \\mathbb{R}^{(2H_w - 1) \\times (2W_w - 1)}" + }, + { + "bbox": [ + 304, + 487, + 547, + 631 + ], + "type": "text", + "content": ". Overall, the formulation is:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 376, + 637, + 493, + 650 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 637, + 493, + 650 + ], + "spans": [ + { + "bbox": [ + 376, + 637, + 493, + 650 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} = \\mathbf {X} _ {i n} + \\operatorname {M S A} \\left(\\ln \\left(\\mathbf {X} _ {e}\\right)\\right),", + "image_path": "2b6a23db59c73c23ed23189041d4f0a31109f54713fe1d097a21ddda8404b1d7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 372, + 652, + 545, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 652, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 372, + 652, + 545, + 666 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {e} = \\mathbf {X} _ {w} + \\mathcal {R} (\\operatorname {D W C o n v} (\\mathcal {R} (\\mathbf {X} _ {w}))), \\tag {3}", + "image_path": "6463eb1edbe397535fbab955b5ca9d087c88044cf5ce968f1815a87a8ef1ef7b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 669, + 519, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 669, + 519, + 683 + ], + "spans": [ + { + "bbox": [ + 321, + 669, + 519, + 683 + ], + "type": "interline_equation", + "content": "\\operatorname {M S A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t M a x} \\left(\\mathbf {Q K} ^ {T} / \\sqrt {d _ {h}} + \\mathbf {B}\\right) \\mathbf {V},", + "image_path": "e33275526fceb57b7fd797a81951958d7d5ddf87f96f271d0e325d88182af0bc.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 689, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 546, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 304, + 689, + 546, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 689, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 304, + 689, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 689, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 304, + 689, + 546, + 713 + ], + "type": "text", + "content": " are projected query, key and value from input. " + }, + { + "bbox": [ + 304, + 689, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 304, + 689, + 546, + 713 + ], + "type": "text", + "content": " is the relative position bias, whose values are taken" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "12273" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 84 + ], + "type": "text", + "content": "from " + }, + { + "bbox": [ + 47, + 72, + 287, + 84 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 47, + 72, + 287, + 84 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 72, + 287, + 84 + ], + "type": "inline_equation", + "content": "\\mathrm{LN}(\\cdot)" + }, + { + "bbox": [ + 47, + 72, + 287, + 84 + ], + "type": "text", + "content": " means layer norm and " + }, + { + "bbox": [ + 47, + 72, + 287, + 84 + ], + "type": "inline_equation", + "content": "d_h" + }, + { + "bbox": [ + 47, + 72, + 287, + 84 + ], + "type": "text", + "content": " denotes head size." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 99, + 189, + 113 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 99, + 189, + 113 + ], + "spans": [ + { + "bbox": [ + 47, + 99, + 189, + 113 + ], + "type": "text", + "content": "3.2. Spectrally-enhanced FFN" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 120, + 288, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 288, + 216 + ], + "type": "text", + "content": "Researchers have proven the spectral feature effectively captures long-range dependencies. The Fast Fourier Convolution (FFC) operator in [4] could efficiently implement a non-local receptive field. Inspired by this work, we devised our Se-FFN based on depth-wise convolutional layer. The spectral branch captures the discrepancy between blurry and sharp image pairs as a supplemental part of self-attention for effective long-range dependency modeling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "text", + "content": "Se-FFN includes two parts: spectral branch and spatial branch (Figure 3b). The spatial part first applies layer norm to the input feature followed by two convolutions with " + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "text", + "content": " kernel and a " + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "text", + "content": " depth-wise convolution to capture local information. The spectral branch is formulated as (a) convert the input feature map " + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} \\in \\mathbb{R}^{B \\times H \\times W \\times C'}" + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "text", + "content": " into the frequency domain to obtain " + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{Y})" + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "text", + "content": ". (b) Concatenate the real part and imaginary part of " + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{Y})" + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "text", + "content": " along channel dimension to get the spectral feature " + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Y}} \\in \\mathbb{R}^{B \\times H \\times W/2 \\times 2C'}" + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "text", + "content": ". (c) Exploit two " + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "text", + "content": " convolutional layers with the LeakyReLU function in between to process the spectral feature. (d) Split the processed spectral feature as real and imaginary parts, then convert them back to spatial domain " + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_f \\in \\mathbb{R}^{B \\times H \\times W \\times C'}" + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "text", + "content": " by inverse Fast Fourier Transform (iFFT). Finally, the output of Se-FFN is " + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} = \\mathbf{Y} + \\mathbf{Y}_f + \\mathbf{Y}_s" + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_s" + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "text", + "content": " is the output of spatial branch." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 425, + 196, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 425, + 196, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 425, + 196, + 436 + ], + "type": "text", + "content": "3.3. Cross-scale Feature Fusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 445, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 601 + ], + "type": "text", + "content": "In most conventional U-net structures, the corresponding encoder and decoder are related by a skip connection to allow information flow between the same scale level. Later in [6], authors present an Asymmetric Feature Fusion (AFF) module to enable cross-scale interaction but without any control. So, we present our CFF-SE block. As shown in Figure 2, before the convolutional block, the input firstly goes through a SE module to control the information flow from the respective scale level of all encoders, which enables us to actively emphasize or suppress the features. Thus, each decoder can selectively exploit multi-scale features, resulting in improved deblurring performance as will be demonstrated in the experimental part." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 619, + 128, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 619, + 128, + 632 + ], + "spans": [ + { + "bbox": [ + 47, + 619, + 128, + 632 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "Learning-based methods for deblurring or RS correction are usually trained on synthetic data, which has limitations of generalization. We, therefore, built an image acquisition system to collect a real dataset named GRR-real. The details about the system, training implementation, and more experimental results are offered in supplemental materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 484, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 484, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 484, + 85 + ], + "type": "text", + "content": "4.1. Comparison on GRR Deblurring" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 93, + 547, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 93, + 547, + 142 + ], + "spans": [ + { + "bbox": [ + 305, + 93, + 547, + 142 + ], + "type": "text", + "content": "Comparison with Other Models Since single GRR deblurring is addressed for the first time, there is no specially devised algorithm. So, we compare our model with previous state-of-the-art approaches from:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 150, + 547, + 326 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 317, + 150, + 545, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 150, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 150, + 545, + 186 + ], + "type": "text", + "content": "- GS image-based deblurring by exploiting coarse-to-fine strategy: basic U-net [36], SRN [45], MIMO-UNet[6] and Uformer[46]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 188, + 547, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 188, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 317, + 188, + 547, + 224 + ], + "type": "text", + "content": "- GS video-based deblurring by fusing neighboring temporal information: STR-CNN [12], DBN [42] and IFIRNN[26]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 233, + 547, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 233, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 317, + 233, + 547, + 258 + ], + "type": "text", + "content": "- RS video-based correction through predicting pixelwise displacement: JCD [56] and DSUR [23].1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 266, + 547, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 266, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 317, + 266, + 547, + 326 + ], + "type": "text", + "content": "- GRR video deblurring using a spatial-aware encoder with long-short term temporal information aggregator: NGS [47]. For fair comparisons, we not only provide the original setting with 8 frames as input " + }, + { + "bbox": [ + 317, + 266, + 547, + 326 + ], + "type": "inline_equation", + "content": "(\\mathrm{NGS}_8)" + }, + { + "bbox": [ + 317, + 266, + 547, + 326 + ], + "type": "text", + "content": " but also update the model with a single input " + }, + { + "bbox": [ + 317, + 266, + 547, + 326 + ], + "type": "inline_equation", + "content": "(\\mathrm{NGS}_1)" + }, + { + "bbox": [ + 317, + 266, + 547, + 326 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 334, + 547, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 547, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 547, + 501 + ], + "type": "text", + "content": "After retraining all models, Table 1 depicts quantitative results on the GRR-real test dataset. Overall, it demonstrates our specially devised algorithm significantly outperforms models from RS correction and GS deblurring, even though they take multi frames as input and exploit temporal information. As discussed in Section 1, RS correction usually resorts to nontrivial motion estimation and warping process. When directly used for GRR, they performed worse than deblurring models. Furthermore, we observe that, for deblurring models, those with video clips as input may degrade sharply, even are no match for some approaches of a single input. A key reason is that these models are struggling to align the content of different video frames. If failed, the redundant inputs would confuse learning process instead." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 502, + 547, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 547, + 657 + ], + "type": "text", + "content": "Although the tailored GRR video deblurring method " + }, + { + "bbox": [ + 304, + 502, + 547, + 657 + ], + "type": "inline_equation", + "content": "\\mathrm{NGS}_8" + }, + { + "bbox": [ + 304, + 502, + 547, + 657 + ], + "type": "text", + "content": " [47] achieved higher metrics than ours (30.03/0.90 vs. 28.64/0.90), it heavily relies on temporal correlation of 8 frames and the performance decreased drastically when given single input " + }, + { + "bbox": [ + 304, + 502, + 547, + 657 + ], + "type": "inline_equation", + "content": "(\\mathrm{NGS}_1)" + }, + { + "bbox": [ + 304, + 502, + 547, + 657 + ], + "type": "text", + "content": ". In contrast to RSS-T, the demand of " + }, + { + "bbox": [ + 304, + 502, + 547, + 657 + ], + "type": "inline_equation", + "content": "\\mathrm{NGS}_8" + }, + { + "bbox": [ + 304, + 502, + 547, + 657 + ], + "type": "text", + "content": " for 8 consecutive frames in training or testing can be difficult to be satisfied, sometimes even impossible due to buffer restrictions. Besides, sensitivity to video capturing settings (e.g., frame rate and deadtime between frames) will lead to additional generalization issues. We also computed the complexity of all algorithms, which shows that our model is at medium level and outperforms SOTA RS correction or Transformer-based models. From qualitative" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 304, + 664, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 664, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 547, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 304, + 664, + 547, + 713 + ], + "type": "text", + "content": "Noting that single RS image correction methods either heavily rely on strong assumption [35, 17, 32, 31], which is apparently inapplicable to our natural scene dataset or require actual velocity or depth to train the model [34, 58] that are barely accessible in common cameras. So, we ignore them and only choose RS video correction methods to compare." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12274" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 154, + 143 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 154, + 143 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 154, + 143 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 154, + 143 + ], + "type": "image", + "image_path": "6c07a62771f7d08e2a733bdcf45ea6ff87c24c889ccbb90d3f27042dc3a1eb3f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 154, + 70, + 249, + 143 + ], + "blocks": [ + { + "bbox": [ + 154, + 70, + 249, + 143 + ], + "lines": [ + { + "bbox": [ + 154, + 70, + 249, + 143 + ], + "spans": [ + { + "bbox": [ + 154, + 70, + 249, + 143 + ], + "type": "image", + "image_path": "29abe3ac19cc63e5ff08bcd78ef5972a8fdcf5ad07d9dc6970ce8cb3859da507.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 249, + 70, + 344, + 143 + ], + "blocks": [ + { + "bbox": [ + 249, + 70, + 344, + 143 + ], + "lines": [ + { + "bbox": [ + 249, + 70, + 344, + 143 + ], + "spans": [ + { + "bbox": [ + 249, + 70, + 344, + 143 + ], + "type": "image", + "image_path": "20b2157bdc87a47c7cd87749cd790a5f7bf0d966f8f0a77750768a3e4783d90c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 344, + 70, + 439, + 143 + ], + "blocks": [ + { + "bbox": [ + 344, + 70, + 439, + 143 + ], + "lines": [ + { + "bbox": [ + 344, + 70, + 439, + 143 + ], + "spans": [ + { + "bbox": [ + 344, + 70, + 439, + 143 + ], + "type": "image", + "image_path": "a4a7996a66ac7811165c1390bfdf59d4308b823d082138114dbb930c21604fbe.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 440, + 70, + 535, + 143 + ], + "blocks": [ + { + "bbox": [ + 440, + 70, + 535, + 143 + ], + "lines": [ + { + "bbox": [ + 440, + 70, + 535, + 143 + ], + "spans": [ + { + "bbox": [ + 440, + 70, + 535, + 143 + ], + "type": "image", + "image_path": "c3f4b062b1a6841df49321bf3609fd00d439a1ce77d2074476673903617a8423.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 57, + 153, + 153, + 225 + ], + "blocks": [ + { + "bbox": [ + 96, + 144, + 114, + 153 + ], + "lines": [ + { + "bbox": [ + 96, + 144, + 114, + 153 + ], + "spans": [ + { + "bbox": [ + 96, + 144, + 114, + 153 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 153, + 153, + 225 + ], + "lines": [ + { + "bbox": [ + 57, + 153, + 153, + 225 + ], + "spans": [ + { + "bbox": [ + 57, + 153, + 153, + 225 + ], + "type": "image", + "image_path": "865db4c06b59df0feede0bf35d7300e432c04b6052c71c99c6351e68d3b74cfb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 154, + 153, + 249, + 225 + ], + "blocks": [ + { + "bbox": [ + 184, + 144, + 217, + 153 + ], + "lines": [ + { + "bbox": [ + 184, + 144, + 217, + 153 + ], + "spans": [ + { + "bbox": [ + 184, + 144, + 217, + 153 + ], + "type": "text", + "content": "U-net [36]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 153, + 249, + 225 + ], + "lines": [ + { + "bbox": [ + 154, + 153, + 249, + 225 + ], + "spans": [ + { + "bbox": [ + 154, + 153, + 249, + 225 + ], + "type": "image", + "image_path": "422980c6580793f25e092ddd3499328134aacf6cbaefc02abf15057b5a463cdf.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 249, + 153, + 344, + 225 + ], + "blocks": [ + { + "bbox": [ + 280, + 144, + 312, + 153 + ], + "lines": [ + { + "bbox": [ + 280, + 144, + 312, + 153 + ], + "spans": [ + { + "bbox": [ + 280, + 144, + 312, + 153 + ], + "type": "text", + "content": "SRN [45]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 249, + 153, + 344, + 225 + ], + "lines": [ + { + "bbox": [ + 249, + 153, + 344, + 225 + ], + "spans": [ + { + "bbox": [ + 249, + 153, + 344, + 225 + ], + "type": "image", + "image_path": "2513361c4ace0648e3623b5bed443956abd835b42ad3648a80c2b4ec45afc879.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 344, + 153, + 439, + 225 + ], + "blocks": [ + { + "bbox": [ + 374, + 144, + 408, + 153 + ], + "lines": [ + { + "bbox": [ + 374, + 144, + 408, + 153 + ], + "spans": [ + { + "bbox": [ + 374, + 144, + 408, + 153 + ], + "type": "text", + "content": "MIMO [6]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 344, + 153, + 439, + 225 + ], + "lines": [ + { + "bbox": [ + 344, + 153, + 439, + 225 + ], + "spans": [ + { + "bbox": [ + 344, + 153, + 439, + 225 + ], + "type": "image", + "image_path": "78b567243fa9597b7cd3a856cd4b05ff029977f7fa04ee5887b05f48fb4c7e80.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 440, + 153, + 535, + 225 + ], + "blocks": [ + { + "bbox": [ + 468, + 144, + 506, + 153 + ], + "lines": [ + { + "bbox": [ + 468, + 144, + 506, + 153 + ], + "spans": [ + { + "bbox": [ + 468, + 144, + 506, + 153 + ], + "type": "text", + "content": "MIMO+ [6]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 440, + 153, + 535, + 225 + ], + "lines": [ + { + "bbox": [ + 440, + 153, + 535, + 225 + ], + "spans": [ + { + "bbox": [ + 440, + 153, + 535, + 225 + ], + "type": "image", + "image_path": "b5b62b503d635ea90672aab6abd95d0326b91a5d968451e7ec5be2abe477f4a8.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 57, + 236, + 153, + 308 + ], + "blocks": [ + { + "bbox": [ + 83, + 227, + 127, + 236 + ], + "lines": [ + { + "bbox": [ + 83, + 227, + 127, + 236 + ], + "spans": [ + { + "bbox": [ + 83, + 227, + 127, + 236 + ], + "type": "text", + "content": "Uformer [46]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 236, + 153, + 308 + ], + "lines": [ + { + "bbox": [ + 57, + 236, + 153, + 308 + ], + "spans": [ + { + "bbox": [ + 57, + 236, + 153, + 308 + ], + "type": "image", + "image_path": "ea96ddbaf44480c939644e3ca01ea9f34612c10e53f972d5bc05ae3192ddf076.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 309, + 124, + 319 + ], + "lines": [ + { + "bbox": [ + 86, + 309, + 124, + 319 + ], + "spans": [ + { + "bbox": [ + 86, + 309, + 124, + 319 + ], + "type": "text", + "content": "DSUR [23]" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 329, + 545, + 354 + ], + "lines": [ + { + "bbox": [ + 46, + 329, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 545, + 354 + ], + "type": "text", + "content": "Figure 4: Qualitative results on the GRR-real dataset. We compare our method with state-of-the-art methods from RS correction, GS deblurring, and GRR video deblurring." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 154, + 236, + 249, + 308 + ], + "blocks": [ + { + "bbox": [ + 175, + 227, + 226, + 236 + ], + "lines": [ + { + "bbox": [ + 175, + 227, + 226, + 236 + ], + "spans": [ + { + "bbox": [ + 175, + 227, + 226, + 236 + ], + "type": "text", + "content": "STR-CNN [12]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 236, + 249, + 308 + ], + "lines": [ + { + "bbox": [ + 154, + 236, + 249, + 308 + ], + "spans": [ + { + "bbox": [ + 154, + 236, + 249, + 308 + ], + "type": "image", + "image_path": "167ae600df2a4a4ab20cfbb7f443d24d05712ec1328b9f6e4a8f1573ec5777d4.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 309, + 219, + 319 + ], + "lines": [ + { + "bbox": [ + 182, + 309, + 219, + 319 + ], + "spans": [ + { + "bbox": [ + 182, + 309, + 219, + 319 + ], + "type": "text", + "content": "NGS1 [47]" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 249, + 236, + 344, + 308 + ], + "blocks": [ + { + "bbox": [ + 280, + 227, + 312, + 236 + ], + "lines": [ + { + "bbox": [ + 280, + 227, + 312, + 236 + ], + "spans": [ + { + "bbox": [ + 280, + 227, + 312, + 236 + ], + "type": "text", + "content": "DBN [42]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 249, + 236, + 344, + 308 + ], + "lines": [ + { + "bbox": [ + 249, + 236, + 344, + 308 + ], + "spans": [ + { + "bbox": [ + 249, + 236, + 344, + 308 + ], + "type": "image", + "image_path": "ba9069f537e5479b8e23072ec87a3029b7e1ae742d9010c1534dd17ea6a7beef.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 278, + 309, + 314, + 319 + ], + "lines": [ + { + "bbox": [ + 278, + 309, + 314, + 319 + ], + "spans": [ + { + "bbox": [ + 278, + 309, + 314, + 319 + ], + "type": "text", + "content": "NGS8 [47]" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 344, + 236, + 439, + 308 + ], + "blocks": [ + { + "bbox": [ + 370, + 227, + 413, + 236 + ], + "lines": [ + { + "bbox": [ + 370, + 227, + 413, + 236 + ], + "spans": [ + { + "bbox": [ + 370, + 227, + 413, + 236 + ], + "type": "text", + "content": "IFIRNN [26]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 344, + 236, + 439, + 308 + ], + "lines": [ + { + "bbox": [ + 344, + 236, + 439, + 308 + ], + "spans": [ + { + "bbox": [ + 344, + 236, + 439, + 308 + ], + "type": "image", + "image_path": "b1da72b802b6b386806df53af1d37082283917e12fb881276f37a3a66c7dc0f6.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 309, + 403, + 318 + ], + "lines": [ + { + "bbox": [ + 380, + 309, + 403, + 318 + ], + "spans": [ + { + "bbox": [ + 380, + 309, + 403, + 318 + ], + "type": "text", + "content": "RSS-T" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 440, + 236, + 535, + 308 + ], + "blocks": [ + { + "bbox": [ + 471, + 227, + 502, + 236 + ], + "lines": [ + { + "bbox": [ + 471, + 227, + 502, + 236 + ], + "spans": [ + { + "bbox": [ + 471, + 227, + 502, + 236 + ], + "type": "text", + "content": "JCD [56]" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 440, + 236, + 535, + 308 + ], + "lines": [ + { + "bbox": [ + 440, + 236, + 535, + 308 + ], + "spans": [ + { + "bbox": [ + 440, + 236, + 535, + 308 + ], + "type": "image", + "image_path": "5238d6ea6659bdb5d269380b0d8c5c8482c09df658a514805db104c79c20b2df.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 309, + 493, + 318 + ], + "lines": [ + { + "bbox": [ + 481, + 309, + 493, + 318 + ], + "spans": [ + { + "bbox": [ + 481, + 309, + 493, + 318 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "table", + "bbox": [ + 63, + 434, + 528, + 616 + ], + "blocks": [ + { + "bbox": [ + 45, + 365, + 547, + 425 + ], + "lines": [ + { + "bbox": [ + 45, + 365, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 45, + 365, + 547, + 425 + ], + "type": "text", + "content": "Table 1: Quantitative comparison on GRR-real dataset. We compare our method with state-of-the-art motion deblurring, RS correction, and GRR video deblurring approaches. According to 'exposure mode/Input number', they could be divided as 'GS/Single', 'GS/Multi', 'RS/Multi', 'GRR/Multi', and 'GRR/single'. The performance is measured with mean PSNR/SSIM (↑). 'Full' denotes evaluation using full-size frames, while 'Top', 'Middle', and 'Bottom' represent using only a patch of images from the top, middle, and bottom areas." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 434, + 528, + 616 + ], + "lines": [ + { + "bbox": [ + 63, + 434, + 528, + 616 + ], + "spans": [ + { + "bbox": [ + 63, + 434, + 528, + 616 + ], + "type": "table", + "html": "
MethodMode / InputEffectivenessEfficiency
TopMiddleBottomFullTime (s)Params (M)FLOPs (T)
Input-15.12 / 0.6721.83 / 0.7820.08 / 0.7617.61 / 0.74---
U-net [36]26.34 / 0.9223.83 / 0.8522.74 / 0.8323.57 / 0.870.00629.500.072
SRN [45]25.11 / 0.7824.34 / 0.7224.08 / 0.7223.84 / 0.740.014010.250.509
MIMO [6]GS / Single27.16 /0.9024.55 / 0.8224.04 / 0.7924.48 / 0.840.01236.810.315
MIMO+ [6]28.46 / 0.9126.31 / 0.8525.78 / 0.8326.12 / 0.860.024916.110.724
Uformer [46]25.04 / 0.9323.70 / 0.8622.20 / 0.8422.92 / 0.870.214720.630.257
STR-CNN [12]20.88 / 0.8323.28 / 0.7722.60 / 0.7521.39 / 0.780.01940.930.367
DBN [42]GS / Multi25.11 / 0.9025.90 / 0.8525.87 / 0.8224.97 / 0.850.022915.311.046
IFIRNN [26]27.19 / 0.9023.85 / 0.7823.21 / 0.7723.97 / 0.810.01351.640.581
JCD [56]RS / Multi27.64 / 0.9023.65 / 0.8019.76 / 0.7722.31 / 0.820.26258.670.326
DSUR [23]24.81 / 0.8723.87 / 0.7823.39 / 0.7623.35 / 0.800.30183.900.225
NGS8 [47]GRR / Multi31.71 / 0.9330.54 / 0.9029.12 / 0.8730.03 / 0.900.02338.671.266
NGS1 [47]GRR / Single26.61 / 0.8923.53 / 0.7822.83 / 0.7623.67 / 0.810.01874.560.083
RSS-T30.90 / 0.9328.60 / 0.8827.86 / 0.8628.64 / 0.900.147911.340.176
", + "image_path": "79274a1d5df776e15f56fa28cfa9fc86ccb10acb4ece038cb4477e6f8947105d.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "table_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 46, + 634, + 288, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 634, + 288, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 634, + 288, + 694 + ], + "type": "text", + "content": "results in Figure 4, although the images obtained by the existing networks exhibit mitigated distortions compared to the input, local details and structures were not sufficiently corrected, whereas our method produces sharper images that even visually outperform " + }, + { + "bbox": [ + 46, + 634, + 288, + 694 + ], + "type": "inline_equation", + "content": "\\mathrm{NGS}_8" + }, + { + "bbox": [ + 46, + 634, + 288, + 694 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "Thirdparty Evaluation To further prove the practical value" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 305, + 634, + 545, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 634, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 305, + 634, + 545, + 682 + ], + "type": "text", + "content": "of our method, we evaluate it on another GRR dataset from [47]. As shown in Table 2, our method still outperforms other algorithms and is just slightly inferior to " + }, + { + "bbox": [ + 305, + 634, + 545, + 682 + ], + "type": "inline_equation", + "content": "\\mathrm{NGS}_8" + }, + { + "bbox": [ + 305, + 634, + 545, + 682 + ], + "type": "text", + "content": " with as many as eight input frames (26.62/0.86 vs. 27.29/0.85)." + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "12275" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 89, + 285, + 239 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 276, + 83 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 276, + 83 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 276, + 83 + ], + "type": "text", + "content": "Table 2: Comparison on another GRR dataset [47]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 89, + 285, + 239 + ], + "lines": [ + { + "bbox": [ + 50, + 89, + 285, + 239 + ], + "spans": [ + { + "bbox": [ + 50, + 89, + 285, + 239 + ], + "type": "table", + "html": "
MethodSET-II
FullTopMiddleBottom
Input17.82 / 0.7323.64 / 0.7721.45 / 0.7715.54 / 0.66
SRN [45]25.05 / 0.8124.32 / 0.7925.65 / 0.8127.02 / 0.83
STRCNN [12]22.59 / 0.8122.99 / 0.7923.46 / 0.8123.66 / 0.83
DBN [42]22.57 / 0.8123.24 / 0.8023.81 / 0.8123.24 / 0.82
IFIRNN[26]25.17 / 0.8224.77 / 0.8025.62 / 0.8126.94 / 0.84
ESTRNN [54]22.72 / 0.8323.42 / 0.8126.03 / 0.8322.86 / 0.83
DSUR [23]22.50 / 0.8022.49 / 0.7823.87 / 0.8123.38 / 0.83
JCD [56]25.33 / 0.8024.77 / 0.7825.71 / 0.8027.43 / 0.83
NGS8 [47]27.29 / 0.8526.96 / 0.8427.57 / 0.8528.35 / 0.86
NGS1 [47]25.19 / 0.7924.48 / 0.7725.53 / 0.7927.27 / 0.82
RSS-T26.62 / 0.8625.90 / 0.8426.82 / 0.8628.43 / 0.87
", + "image_path": "754467b461aab5ae8977e3a20818dcce2433bcee8ce57ed1b567751d53548ed3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 274, + 285, + 369 + ], + "blocks": [ + { + "bbox": [ + 47, + 256, + 286, + 268 + ], + "lines": [ + { + "bbox": [ + 47, + 256, + 286, + 268 + ], + "spans": [ + { + "bbox": [ + 47, + 256, + 286, + 268 + ], + "type": "text", + "content": "Table 3: Architecture ablation on our GRR-real dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 274, + 285, + 369 + ], + "lines": [ + { + "bbox": [ + 50, + 274, + 285, + 369 + ], + "spans": [ + { + "bbox": [ + 50, + 274, + 285, + 369 + ], + "type": "table", + "html": "
PEFFNFusionPSNR / SSIM
WinPEPixPELeFNNSe-FNNAFFCFF-SE
v1✓✓25.69 / 0.85
v2✓✓✓26.50 / 0.86
v3✓✓✓26.10 / 0.86
v4✓✓✓✓26.83 / 0.87
v5✓✓✓✓27.24 / 0.87
RSS-T✓✓✓✓28.64 / 0.90
", + "image_path": "ec14846d8dd4fa17ebd46db42075cc57b70deb2a1c74492d62cf7d8abbe9988b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 395, + 140, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 395, + 140, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 395, + 140, + 407 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 420, + 288, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 288, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 288, + 587 + ], + "type": "text", + "content": "Model Architecture Ablation We conducted experiments to analyze the effectiveness of the core components of our RSS-T. The hierarchical position encoding consists of WinPE and PixPE. The feed-forward network is also implemented by LeFFN [46] to compare with our Se-FNN (v5 vs. RSS-T). As for the cross-scale fusion part, AFF [46] and our CFF-SE are used to explore the performance difference (v4 vs. v5). All the experimental results are listed in Table 3. From v1, v2, v3, and v4, when WinPE and PixPE are separately presented in baseline model v1, the performance goes up but the combination of them achieved the best metrics. Our Se-FNN and CFF-SE modules contributed to the further performance improvement of PSNR by " + }, + { + "bbox": [ + 46, + 420, + 288, + 587 + ], + "type": "inline_equation", + "content": "+1.81\\mathrm{dB}" + }, + { + "bbox": [ + 46, + 420, + 288, + 587 + ], + "type": "text", + "content": " compared with the LeFFN and AFF." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "text", + "content": "Window Size Ablation We introduced three types of window size (i.e., horizontal, squared, and full-width window) and clarified the advantages of the horizontal window for GRR deblurring in Section 3.1. Here, we present experimental results to support the claim. As Figure 5 shows, all horizontal window settings are superior to the standard squared one from quantitative and qualitative perspectives. The extreme case (2, full) is severely trapped in low-rank bottleneck [2] as the pixel number of this window is far more than the head size." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 69, + 544, + 258 + ], + "blocks": [ + { + "bbox": [ + 309, + 69, + 544, + 258 + ], + "lines": [ + { + "bbox": [ + 309, + 69, + 544, + 258 + ], + "spans": [ + { + "bbox": [ + 309, + 69, + 544, + 258 + ], + "type": "image", + "image_path": "28cc50ef74d74788349acbfa0f3b6a480cb7e500e5b0d2ff2e6a5989b3cc1d23.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 267, + 545, + 291 + ], + "lines": [ + { + "bbox": [ + 305, + 267, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 305, + 267, + 545, + 291 + ], + "type": "text", + "content": "Figure 5: Window size ablation. Experimental results of all different window-size settings on our GRR-real dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 304, + 542, + 377 + ], + "blocks": [ + { + "bbox": [ + 310, + 304, + 542, + 377 + ], + "lines": [ + { + "bbox": [ + 310, + 304, + 542, + 377 + ], + "spans": [ + { + "bbox": [ + 310, + 304, + 542, + 377 + ], + "type": "image", + "image_path": "3e26407c8df00d9f090c6997474d827bf45a94f127e1aaa19fc34b0c497ecaa5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 386, + 545, + 421 + ], + "lines": [ + { + "bbox": [ + 305, + 386, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 305, + 386, + 545, + 421 + ], + "type": "text", + "content": "Figure 6: Temporal alignment of three shutters. When synthesizing datasets, we strictly align the exposure of three modes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 443, + 503, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 503, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 503, + 456 + ], + "type": "text", + "content": "4.3. Comparison on Three Shutter Modes" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 461, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 568 + ], + "type": "text", + "content": "Until now, a direct comparison of three shutter modes is still missing, although [47] has demonstrated that GRR has its superiority against RS in obtaining sharp images in the scenario of video restoration. Here, we evaluate them on single image Restoration. For a fair comparison, we synthesize strictly aligned images virtually captured by three modes (RS, GRR, and blur) for each scene by following the protocol of [28, 45, 6] (details in the supplemental material), and compare with SOTA methods for each mode." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 569, + 546, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 546, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 546, + 700 + ], + "type": "text", + "content": "When compared with RS mode, Table 4 shows that GRR combined with our RSS-T model has gained significant advantages in obtaining single sharp image, even though DSUR and JCD take multiple frames as input. These results are consistent with the discussion in [47] that the GRR mode enables us to convert the RS correction problem into a more tractable one. Therefore, given an RS camera that supports GRR mode, we recommend switching to GRR mode for better restoration from a single input, if losing some capture speed of the original RS is permissible. Also, the exposure should be adjusted to avoid saturation of bottom rows." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 701, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 547, + 713 + ], + "type": "text", + "content": "We additionally present comparisons with row-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12276" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 148, + 286, + 223 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 288, + 142 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 288, + 142 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 288, + 142 + ], + "type": "text", + "content": "Table 4: Quantitative comparison on three modes. The subscript of each method denotes the number of input frames. The reason we choose multi-frame RS correction is discussed in Section 4.1. MIMO+ and Uformer are trained on blur mode. DSUR and JCD are trained on the RS counterpart, while RSS-T is trained on GRR mode." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 148, + 286, + 223 + ], + "lines": [ + { + "bbox": [ + 50, + 148, + 286, + 223 + ], + "spans": [ + { + "bbox": [ + 50, + 148, + 286, + 223 + ], + "type": "table", + "html": "
MethodFullTopMiddleBottom
DSUR2 [23]24.51 / 0.8128.42 / 0.8923.84 / 0.7824.21 / 0.77
JCD3 [56]23.52 / 0.7525.85 / 0.7923.87 / 0.7523.18 /0.70
RSS-T127.47 / 0.8927.40 / 0.9027.9 / 0.8928.10 / 0.89
MIMO+1[6]25.16 / 0.8626.75 / 0.8725.54 / 0.8525.58 / 0.85
Uformer1[46]24.28 / 0.8425.54 / 0.8624.26 / 0.8324.57 / 0.83
", + "image_path": "6883381720174d4adc04700af91a153cd65293181ff32ab9c84bc1c6ed6587c4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 244, + 289, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 244, + 289, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 289, + 472 + ], + "type": "text", + "content": "independent deblurring issues under identical scenes. For fairness, we artificially elongate the exposure time of a GS camera to synthesize blurred images. As shown in Figure 6, the exposure time of GS is set to be the same as that of the full scanning period of RS and GRR. From Table 4, row-dependent blur from GRR is more tractable than blur generated by a GS camera. Because of the characteristics of GRR exposure, it implicitly encodes the temporal ordering of latent frames that mitigate the motion ambiguity [13] of blur images. And local details from the top sharper yet noisier rows further enhance the reconstruction performance. The practical implication on usage is that, given a GS camera in hand that supports GRR mode, we recommend switching to the GRR mode for single image deblurring with our proposed algorithm. Note that there is no additional cost to pay here since GRR mode will not slow down the original GS frame rate nor cause additional saturation since the longest exposure time of GRR is the same as the GS exposure time (Figure 6)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 485, + 119, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 485, + 119, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 485, + 119, + 498 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 506, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 288, + 685 + ], + "type": "text", + "content": "In this paper, we analyzed image degradations for conventional GS and RS cameras and further highlighted an in-between exposure mode, named GRR, arising from the global reset feature equipped with many RS sensors. Instead of RS distortion, GRR gives rise to a specialized blurring effect with row-dependent blur magnitude and brightness level. Based on the characteristics of GRR degradation, an original Transformer block with spatial compensation and spectral enhancement was devised, which has the advantage of being able to rectify the GRR degradation. The proposed hierarchical positional encoding well preserved the absolute and relative position context and the spectral branch of the feed-forward network enhanced the model's ability to capture long-range discrepancies. The experimental results also verified that our tailored model is effective." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Potential Application As shown in Figure 7, GRR is mainly intended for industrial usage in controlled dark environment," + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 411, + 140 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 411, + 140 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 411, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 411, + 140 + ], + "type": "image", + "image_path": "2e0fb29440ee26553e76989229d118e2737cda86a3c3a8cbe48d801dc53db3f9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 150, + 365, + 159 + ], + "lines": [ + { + "bbox": [ + 354, + 150, + 365, + 159 + ], + "spans": [ + { + "bbox": [ + 354, + 150, + 365, + 159 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 400, + 72, + 494, + 147 + ], + "blocks": [ + { + "bbox": [ + 400, + 72, + 494, + 147 + ], + "lines": [ + { + "bbox": [ + 400, + 72, + 494, + 147 + ], + "spans": [ + { + "bbox": [ + 400, + 72, + 494, + 147 + ], + "type": "image", + "image_path": "97902040cc3210cba77bc08f0dd8af75c1bfe8f967c575c607d4b066791e2982.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 150, + 488, + 159 + ], + "lines": [ + { + "bbox": [ + 477, + 150, + 488, + 159 + ], + "spans": [ + { + "bbox": [ + 477, + 150, + 488, + 159 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 496, + 77, + 545, + 148 + ], + "blocks": [ + { + "bbox": [ + 496, + 77, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 496, + 77, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 496, + 77, + 545, + 148 + ], + "type": "image", + "image_path": "31f922b570d546f42bc916c8b346add328bf96ea0c49e21224f9e827dd1c73af.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 170, + 547, + 206 + ], + "lines": [ + { + "bbox": [ + 305, + 170, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 305, + 170, + 547, + 206 + ], + "type": "text", + "content": "Figure 7: Application scenario of GRR mode. (a) Illustrated application scenario. (b) Synchronization between illumination and cameras." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 228, + 547, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 348 + ], + "type": "text", + "content": "such as parts inspection on a conveyor belt, together with strobe illuminant. In this scenario, GRR with synchronized illumination can capture distortion-free and blur-free images, with lower cost than using a GS sensor, and faster speed than using a RS sensor (Because, in this case, the exposure time of RS has to be elongated such that the first and last rows have sufficient overlap). Instead, without using strobe illumination, GRR boosted with our approach can still restore a sharp image. This will further reduce production and maintenance cost in practice." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 351, + 547, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 547, + 495 + ], + "type": "text", + "content": "Limitations When the illumination is bright enough, with appropriate exposure, GS can capture a sharp image without blur. The bottom parts of GRR can do the same, yet the top parts will be darker and thus noiser, or the top parts can be sharp, yet the the bottom parts will be blurry. In this scenario, GRR is clearly worse than GS, and our algorithm can help reduce the performance gap. On the other hand, RS will not have distortion and the image will be sharp and clean when dealing with completely static scenes. Yet GRR will have non-uniform brightness, although blur will not happen. Our algorithm can compensate this non-uniformity, yet GRR is worse than RS in this particular scenario." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 498, + 547, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 559 + ], + "type": "text", + "content": "Acknowledgement This research was supported in part by JSPS KAKENHI Grant Numbers 22H00529, 20H05951, 22H03620, 22H05015, JST AIP Acceleration Research JP-MJCR22U4, the Value Exchange Engineering, a joint research project between Mercari, Inc. and RIISE." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 570, + 365, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 570, + 365, + 582 + ], + "spans": [ + { + "bbox": [ + 306, + 570, + 365, + 582 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 310, + 590, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 310, + 590, + 547, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 590, + 547, + 644 + ], + "spans": [ + { + "bbox": [ + 310, + 590, + 547, + 644 + ], + "type": "text", + "content": "[1] Cenek Albl, Zuzana Kukelova, Viktor Larsson, Michal Polic, Tomas Pajdla, and Konrad Schindler. From two rolling shutters to one global shutter. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2505-2513, 2020. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 646, + 547, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 646, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 311, + 646, + 547, + 690 + ], + "type": "text", + "content": "[2] Srinadh Bhojanapalli, Chulhee Yun, Ankit Singh Rawat, Sashank Reddi, and Sanjiv Kumar. Low-rank bottleneck in multi-head attention models. In International Conference on Machine Learning, pages 864-873. PMLR, 2020. 5, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 691, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 691, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 311, + 691, + 547, + 714 + ], + "type": "text", + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12277" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 289, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 105 + ], + "type": "text", + "content": "to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 106, + 289, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 106, + 289, + 138 + ], + "spans": [ + { + "bbox": [ + 53, + 106, + 289, + 138 + ], + "type": "text", + "content": "[4] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolution. Advances in Neural Information Processing Systems, 33:4479-4488, 2020. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 140, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 140, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 53, + 140, + 288, + 183 + ], + "type": "text", + "content": "[5] Lu Chi, Guiyu Tian, Yadong Mu, Lingxi Xie, and Qi Tian. Fast non-local neural networks with spectral residual learning. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2142-2151, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 184, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 184, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 184, + 288, + 237 + ], + "type": "text", + "content": "[6] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4641-4650, 2021. 3, 6, 7, 8, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "text", + "content": "[7] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Conditional positional encodings for vision transformers. arXiv preprint arXiv:2102.10882, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 283, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 283, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 53, + 283, + 288, + 348 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 350, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 350, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 350, + 288, + 392 + ], + "type": "text", + "content": "[9] Bin Fan, Yuchao Dai, and Mingyi He. Sunet: symmetric undistortion network for rolling shutter correction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4541-4550, 2021. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 393, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 288, + 448 + ], + "type": "text", + "content": "[10] Hongyun Gao, Xin Tao, Xiaoyong Shen, and Jiaya Jia. Dynamic scene deblurring with parameter selective sharing and nested skip connections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3848-3856, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 288, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 481 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 481 + ], + "type": "text", + "content": "[11] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 482, + 288, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 288, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 288, + 535 + ], + "type": "text", + "content": "[12] Tae Hyun Kim, Kyoung Mu Lee, Bernhard Scholkopf, and Michael Hirsch. Online video deblurring via dynamic temporal blending network. In Proceedings of the IEEE International Conference on Computer Vision, pages 4038-4047, 2017. 3, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 288, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 288, + 581 + ], + "type": "text", + "content": "[13] Meiguang Jin, Givi Meishvili, and Paolo Favaro. Learning to extract a video sequence from a single motion-blurred image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6334-6342, 2018. 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 582, + 288, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 288, + 603 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 288, + 603 + ], + "type": "text", + "content": "[14] Yitzhak Katznelson. An introduction to harmonic analysis. Cambridge University Press, 2004. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 604, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 604, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 604, + 288, + 658 + ], + "type": "text", + "content": "[15] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiri Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8183-8192, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[16] Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8878-8887, 2019. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "text", + "content": "[17] Yizhen Lao and Omar Ait-Aider. A robust method for strong rolling shutter effects correction using lines with automatic feature selection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4795-4803, 2018. 1, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "type": "text", + "content": "[18] Junyong Lee, Hyeongseok Son, Jaesung Rim, Sunghyun Cho, and Seungyong Lee. Iterative filter adaptive network for single image defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2034-2042, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 185, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 547, + 217 + ], + "type": "text", + "content": "[19] Anat Levin. Blind motion deblurring using image statistics. Advances in Neural Information Processing Systems, 19, 2006. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 219, + 547, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 547, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 547, + 284 + ], + "type": "text", + "content": "[20] Dongxu Li, Chenchen Xu, Kaihao Zhang, Xin Yu, Yiran Zhong, Wenqi Ren, Hanna Suominen, and Hongdong Li. Arvo: Learning all-range volumetric correspondence for video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7721-7731, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 286, + 547, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 547, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 547, + 330 + ], + "type": "text", + "content": "[21] Kunchang Li, Yali Wang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatiotemporal representation learning. arXiv preprint arXiv:2201.04676, 2022. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 332, + 547, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 547, + 385 + ], + "type": "text", + "content": "[22] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 388, + 547, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 547, + 431 + ], + "type": "text", + "content": "[23] Peidong Liu, Zhaopeng Cui, Viktor Larsson, and Marc Pollefeys. Deep shutter unrolling network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5941-5949, 2020. 1, 3, 6, 7, 8, 9" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 434, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 547, + 487 + ], + "type": "text", + "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 489, + 547, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 547, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 547, + 543 + ], + "type": "text", + "content": "[25] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 545, + 547, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 547, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 547, + 599 + ], + "type": "text", + "content": "[26] Seungjun Nah, Sanghyun Son, and Kyoung Mu Lee. Recurrent neural networks with intra-frame iterations for video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8102-8111, 2019. 3, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 601, + 547, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 547, + 643 + ], + "type": "text", + "content": "[27] Mehdi Noroozi, Paramanand Chandramouli, and Paolo Favaro. Motion deblurring in the wild. In German conference on pattern recognition, pages 65-77. Springer, 2017. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 646, + 547, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 547, + 690 + ], + "type": "text", + "content": "[28] Jihyong Oh and Munchurl Kim. Demfi: deep joint deblurring and multi-frame interpolation with flow-guided attentive correlation and recursive boosting. In European Conference on Computer Vision, pages 198-215. Springer, 2022. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 692, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 547, + 714 + ], + "type": "text", + "content": "[29] Chandramouli Paramanand and Ambasamudram N Rajagopalan. Non-uniform motion deblurring for bilayer scenes." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12278" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1115-1122, 2013. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "type": "text", + "content": "[30] Dongwon Park, Dong Un Kang, Jisoo Kim, and Se Young Chun. Multi-temporal recurrent neural networks for progressive non-uniform single image deblurring with incremental temporal training. In European Conference on Computer Vision, pages 327-343. Springer, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 206 + ], + "type": "text", + "content": "[31] Pulak Purkait and Christopher Zach. Minimal solvers for monocular rolling shutter compensation under ackermann motion. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 903–911. IEEE, 2018. 1, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 288, + 251 + ], + "type": "text", + "content": "[32] Pulak Purkait, Christopher Zach, and Ales Leonardis. Rolling shutter correction in manhattan world. In Proceedings of the IEEE International Conference on Computer Vision, pages 882-890, 2017. 1, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 288, + 297 + ], + "type": "text", + "content": "[33] Kuldeep Purohit and AN Rajagopalan. Region-adaptive dense network for efficient motion deblurring. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11882-11889, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 288, + 342 + ], + "type": "text", + "content": "[34] Vijay Rengarajan, Yogesh Balaji, and AN Rajagopalan. Unrolling the shutter: Cnn to correct motion distortions. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 2291-2299, 2017. 1, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 288, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 288, + 398 + ], + "type": "text", + "content": "[35] Vijay Rengarajan, Ambasamudram N Rajagopalan, and Rangarajan Aravind. From bows to arrows: Rolling shutter rectification of urban scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2773-2781, 2016. 1, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 453 + ], + "type": "text", + "content": "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 455, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 288, + 498 + ], + "type": "text", + "content": "[37] Christian J Schuler, Michael Hirsch, Stefan Harmeling, and Bernhard Schölkopf. Learning to deblur. IEEE transactions on pattern analysis and machine intelligence, 38(7):1439-1451, 2015. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 500, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 288, + 533 + ], + "type": "text", + "content": "[38] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. arXiv preprint arXiv:1803.02155, 2018. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "type": "text", + "content": "[39] Wang Shen, Wenbo Bao, Guangtao Zhai, Li Chen, Xiongkuo Min, and Zhiyong Gao. Blurry video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5114-5123, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 579, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 634 + ], + "type": "text", + "content": "[40] Hyeongseok Son, Junyong Lee, Sunghyun Cho, and Seungyong Lee. Single image defocus deblurring using kernel-sharing parallel atrous convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2642-2650, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "text", + "content": "[41] Hyeongseok Son, Junyong Lee, Jonghyeop Lee, Sunghyun Cho, and Seungyong Lee. Recurrent video deblurring with blur-invariant motion estimation and pixel volumes. ACM Transactions on Graphics (TOG), 40(5):1-18, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "text", + "content": "[42] Shuochen Su, Maurizio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 1279-1288, 2017. 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "text", + "content": "[43] Jian Sun, Wenfei Cao, Zongben Xu, and Jean Ponce. Learning a convolutional neural network for non-uniform motion blur removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 769-777, 2015. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 547, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 185 + ], + "type": "text", + "content": "[44] Yu-Wing Tai, Ping Tan, and Michael S Brown. Richardson-lucy deblurring for scenes under a projective motion path. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(8):1603-1618, 2010. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 546, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 546, + 231 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 546, + 231 + ], + "type": "text", + "content": "[45] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8174-8182, 2018. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 232, + 546, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 546, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 546, + 274 + ], + "type": "text", + "content": "[46] Zhendong Wang, Xiaodong Cun, Jianmin Bao, and Jianzhuang Liu. Uformer: A general u-shaped transformer for image restoration. arXiv preprint arXiv:2106.03106, 2021. 1, 2, 3, 4, 6, 7, 8, 9" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 276, + 547, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 547, + 342 + ], + "type": "text", + "content": "[47] Zhixiang Wang, Xiang Ji, Jia-Bin Huang, Shin'ichi Satoh, Xiao Zhou, and Yinqiang Zheng. Neural global shutter: Learn to restore video from a rolling shutter camera with global reset feature. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17794-17803, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 344, + 547, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 344, + 547, + 388 + ], + "spans": [ + { + "bbox": [ + 308, + 344, + 547, + 388 + ], + "type": "text", + "content": "[48] Patrick Wieschollek, Michael Hirsch, Bernhard Scholkopf, and Hendrik Lensch. Learning blind motion deblurring. In Proceedings of the IEEE international conference on computer vision, pages 231-240, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 388, + 547, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 547, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 547, + 442 + ], + "type": "text", + "content": "[49] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 445, + 547, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 547, + 489 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 547, + 489 + ], + "type": "text", + "content": "[50] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. arXiv preprint arXiv:2111.09881, 2021. 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 490, + 546, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 546, + 545 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 546, + 545 + ], + "type": "text", + "content": "[51] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 546, + 546, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 546, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 546, + 600 + ], + "type": "text", + "content": "[52] Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Bjorn Stenger, Wei Liu, and Hongdong Li. Deblurring by realistic blurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2737-2746, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 601, + 547, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 547, + 669 + ], + "type": "text", + "content": "[53] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xi-ang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881–6890, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 670, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 546, + 713 + ], + "type": "text", + "content": "[54] Zhihang Zhong, Ye Gao, Yinqiang Zheng, and Bo Zheng. Efficient spatio-temporal recurrent neural network for video deblurring. In European Conference on Computer Vision, pages 191-207. Springer, 2020. 3, 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12279" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 274 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[55] Zhisheng Zhong, Tiancheng Shen, Yibo Yang, Zhouchen Lin, and Chao Zhang. Joint sub-bands learning with clique structures for wavelet domain super-resolution. Advances in neural information processing systems, 31, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 171 + ], + "type": "text", + "content": "[56] Zhihang Zhong, Yinqiang Zheng, and Imari Sato. Towards rolling shutter correction and deblurring in dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9219-9228, 2021. 6, 7, 8, 9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 288, + 217 + ], + "type": "text", + "content": "[57] Bingbing Zhuang, Loong-Fah Cheong, and Gim Hee Lee. Rolling-shutter-aware differential sfm and image rectification. In Proceedings of the IEEE International Conference on Computer Vision, pages 948-956, 2017. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 288, + 274 + ], + "type": "text", + "content": "[58] Bingbing Zhuang, Quoc-Huy Tran, Pan Ji, Loong-Fah Cheong, and Manmohan Chandraker. Learning structure-and-motion-aware rolling shutter correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4551-4560, 2019. 1, 3, 6" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12280" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Single Image Defocus Deblurring via Implicit Neural Inverse Kernels/bee3d863-7fc6-4e0b-a339-87a4afbe8bb4_content_list.json b/2023/Single Image Defocus Deblurring via Implicit Neural Inverse Kernels/bee3d863-7fc6-4e0b-a339-87a4afbe8bb4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..51eeb8c43a33d3c6dae6212c8329c4287b21bf85 --- /dev/null +++ b/2023/Single Image Defocus Deblurring via Implicit Neural Inverse Kernels/bee3d863-7fc6-4e0b-a339-87a4afbe8bb4_content_list.json @@ -0,0 +1,1577 @@ +[ + { + "type": "text", + "text": "Single Image Defocus Deblurring via Implicit Neural Inverse Kernels", + "text_level": 1, + "bbox": [ + 135, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuhui Quan $^{1,2}$ Xin Yao $^{1,2}$ Hui Ji $^{3}$", + "bbox": [ + 292, + 179, + 676, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Engineering, South China University of Technology, Guangzhou 510006, China $^{2}$ Pazhou Lab, Guangzhou 510335, China", + "bbox": [ + 107, + 199, + 861, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Department of Mathematics, National University of Singapore, 119076, Singapore", + "bbox": [ + 205, + 234, + 767, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cseyhquan@scut.edu.cn, csxinyaoscut@gmail.com, matjh@nus.edu.sg*", + "bbox": [ + 205, + 253, + 767, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single image defocus deblurring (SIDD) is a challenging task due to the spatially-varying nature of defocus blur, characterized by per-pixel point spread functions (PSFs). Existing deep-learning-based methods for SIDD are limited by either over-fitting due to the lack of model constraints or under-parametrization that restricts their applicability to real-world images. To address the limitations, this paper proposes an interpretable approach that explicitly predicts inverse kernels with structural regularization. Motivated by the observation that defocus PSFs within an image often have similar shapes but different sizes, we represent the inverse kernels linearly over a multi-scale dictionary parameterized by implicit neural representations. We predict the corresponding representation coefficients via a duplex scale-recurrent neural network that jointly performs fine-to-coarse and coarse-to-fine estimations. Extensive experiments demonstrate that our approach achieves excellent performance using a lightweight model.", + "bbox": [ + 75, + 335, + 472, + 608 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 633, + 209, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When a camera captures an image, objects outside of the focal plane appear blurry. This effect is known as defocus blur and usually occurs when using a large camera aperture or capturing scenes with significantly varying depths. Removing defocus blur from a single image, known as SIDD, is desired for many practical applications; see e.g. [27, 14, 33, 60]. In general, defocus blur varies spatially. Each defocused pixel is a weighted average of its neighboring pixels in the latent all-in-focus image, with the weights determined by a spatially-varying unknown PSF.", + "bbox": [ + 75, + 659, + 468, + 810 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SIDD is closely related to defocus map estimation (DME) [3, 44, 7, 5, 51, 32, 17]. Existing DME methods typically model defocus PSFs using a Gaussian function", + "bbox": [ + 75, + 811, + 468, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "parameterized by its variance or a disc function parameterized by its radius. The resulting defocus map indicates the per-pixel blur amount in terms of the values of such model parameters, which can then be utilized for SIDD by applying a non-blind image deblurring (NID) algorithm using the corresponding PSFs; see e.g. [18]. However, DME remains a challenging task, and the estimated parameters of PSFs can be erroneous for many pixels. Moreover, Gaussian or disc-indicator functions are overly simplistic for real-world defocus PSFs. All these errors will be magnified in the NID process, resulting in noticeable artifacts.", + "bbox": [ + 500, + 305, + 892, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, many deep-learning-based methods have been proposed for SIDD; see e.g. [41, 25, 21, 47, 36, 40, 1, 56, 58, 23, 31]. Most train an end-to-end neural network (NN) that directly maps blurry images to their sharp correspondences. Although these methods provide better results than the DME-based two-step methods, their performance still has much room for improvement due to significant variations in the spatial distributions of defocus PSFs among different images. Many existing works (e.g. [21, 40, 56]) explicitly introduce spatially-varying processing blocks to better handle the spatial variance of defocus PSFs. However, those blocks lack constraints and may cause overfitting. A few studies (e.g. [36, 37]) represent defocus PSFs by some specific basis for geometric regularization and employ deep unrolling for an NN-based inversion process. Nevertheless, there is still room for improvement.", + "bbox": [ + 496, + 474, + 892, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper presents an end-to-end deep NN with high interpretability for SIDD, which explicitly predicts spatially-varying inverse kernels of defocus PSFs and performs deblurring using the predicted inverse kernels. The possible overfitting issue is tackled by providing a more accurate model with structural constraints/regularization for the inverse kernels. Specifically, the inverse kernels are modeled by a linear representation under a dictionary. To achieve a compact yet sufficiently general representation, we construct the dictionary with a multi-scale structure. This is motivated by the empirical observation (as seen in [47]) that defocus PSFs within an image tend to have the same shape", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work is supported by the Natural Science Foundation of Guangdong Province (Grant No. 2022A1515011755 and 2023A1515012841), and Singapore MOE AcRF Tier 1 Grant (WBS No. A-8000981-00-00).", + "bbox": [ + 75, + 862, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12600", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "but vary mainly in size. We show that if two defocus PSFs of different sizes have the same shape under an upsampling operation, so do the dictionary atoms of their inverse kernels. Thus, the dictionary atoms of a large size are defined as an upsampled version of the atoms of a small size.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although plain upsampling offers an economical method for generating larger dictionary atoms from smaller ones, the resulting atoms may have a limited frequency range. This can lead to poor approximations of certain inverse kernels. To address the limitation, we leverage implicit neural representation (INR) [46], a technique that uses coordinate-input NNs to represent geometric objects like 2D shapes, so as to efficiently parameterize multi-scale atoms with both sufficient coverage of high-frequency content and implicit regularization to alleviate overfitting.", + "bbox": [ + 75, + 167, + 468, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For the multi-scale representation coefficients, we separate them into scale-related and shape-related components, which are predicted by two scale-recurrent sub-NNs respectively. To better utilize information from different scales for coefficient prediction, we introduce a duplex scale-recurrent framework that performs both fine-to-coarse and coarse-to-fine estimations. Once the coefficients are predicted, the corresponding inverse kernels are applied to the input image for deblurring. The resulting pipeline is highly efficient and interpretable, leading to a lightweight yet effective model.", + "bbox": [ + 75, + 319, + 468, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To conclude, this paper proposes an end-to-end NN for SIDD, which exhibits a noticeable performance boost compared to existing ones, while maintaining low complexity. See below for our main contributions:", + "bbox": [ + 75, + 470, + 468, + 531 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A parametric inverse kernel prediction framework for SIDD, providing efficient and interpretable deep NNs.", + "- A multi-scale linear representation model for the inverse kernels of spatially-varying defocus blur, using an INR-based multi-scale dictionary.", + "- A duplex scale-recurrent framework for predicting the representation coefficients of inverse kernels." + ], + "bbox": [ + 96, + 539, + 467, + 651 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 667, + 225, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Existing Approaches for SIDD", + "text_level": 1, + "bbox": [ + 76, + 694, + 346, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SIDD can done by a two-step approach that combines DME (e.g. [44, 51, 17, 32]) and NID (e.g. [20]). However, such an approach has two inherent drawbacks. First, the errors in DME will be magnified in the following NID process. Second, simple Gaussian or disc functions do not fit real-world defocus PSFs well. While there has been steady progress in deep-learning-based methods for DME [59, 16] and for NID [38, 10, 33, 29, 30, 4, 34, 35], the inherent drawbacks of the two-step approach remain.", + "bbox": [ + 75, + 718, + 468, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep end-to-end learning has emerged as a promising approach for SIDD. The seminal work conducted by Abuolaim and Brown [2] trained an encoder-decoder CNN that", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "maps a defocused image to its in-focus correspondence, with significant performance gain over two-step methods. To better handle spatially-varying defocus blur, some studies introduced dynamic processing blocks, e.g., per-pixel filter prediction in Lee et al. [21], dynamic residual blocks in Ruan et al. [40], and attention in Zhang and Zhai [56].", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The use of dynamic processing blocks may increase the model complexity and the risk of overfitting. Regularization by auxiliary tasks is an effective approach for improving the generalization performance. In existing works, auxiliary tasks are often defined using DME ([25, 41, 58]), depth estimation [31], dual-view data [21, 1], or reblurring loss [23]. However, most of them require additional data sources.", + "bbox": [ + 496, + 181, + 890, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A more appealing approach is encoding physical priors of defocus blur into NNs. Quan et al. [36, 37] used a kernel mixture model to parameterize defocus PSFs and constructed a deep unrolling NN. This method can be viewed as the degradation learning [22, 49, 24, 11] that predicts the blurring operator or kernels for deblurring. Our approach differs in that it predicts inverse kernels, not the degradation with defocus PSFs. Son et al. [47] showed that inverse defocus kernels have an invariant shape across scales and emulated them in feature spaces via a series of kernel-sharing parallel Atrous convolutions. Their method inspired our work, and it can be viewed as representing inverse kernels for a feature tensor via a multi-scale dilated combination of a single atom and predicting spatially-invariant combination coefficients via channel attention. In comparison, our approach explicitly predicts inverse kernels on the input image, allowing better interpretability, and represents them using INR to improve effectiveness. Additionally, we predict spatially-varying combination coefficients to better handle spatially-similar (not invariant) defocus shapes.", + "bbox": [ + 496, + 287, + 892, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Related Techniques", + "text_level": 1, + "bbox": [ + 500, + 601, + 686, + 617 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep inverse kernels Several studies have included inverse kernels in deblurring NNs, in addition to [47]. Xu et al. [52] and Ren et al. [38] presented the CNNs that adopt tensor product for emulating separable inverse kernels to perform NID. Dong et al. [9, 10] and Pronina et al. [33] performed NID via Wiener inverse filtering in deep feature spaces. However, all these methods are designed for spatially-invariant blur. Ren et al. [55, 39] proposed a recurrent NN to mimic inverse kernels for removing spatially-varying blur of dynamic scenes, but not for SIDD.", + "bbox": [ + 496, + 625, + 890, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scale-recurrent NNs for deblurring Scale-recurrent NNs are commonly used for image deblurring, e.g. [36, 28, 48, 12, 54, 8, 15]. These methods typically employ coarse-to-fine estimation which is unidirectional and thus cannot fully leverage information from the subsequent scales for the prediction at the current scale. In contrast, our proposed NN utilizes a bi-directional structure via bi-LSTMs [13] for joint coarse-to-fine and fine-to-coarse estimation. This is", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12601", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "particularly useful for SIDD, as defocus blur effects show strong self-similarity over scales. While bi-LSTM has been used in [61] for multi-scale video deblurring, it is only used for bi-directional temporal processing and the multi-scale processing is still one-way. Also note that the multi-scale NN proposed in [6] for motion deblurring is not restrictive to unidirectional processing, but not in a recurrent manner.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "INR for image recovery INR has been applied to image recovery [42, 19, 50]. In comparison to these works, we use INR to parameterize inverse kernels, not images, for SIDD.", + "bbox": [ + 76, + 199, + 468, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. INR-Based Dictionary for Inverse Kernels", + "text_level": 1, + "bbox": [ + 76, + 256, + 452, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Inverse Kernels for Spatially-Varying Blur", + "text_level": 1, + "bbox": [ + 76, + 281, + 437, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider a simplistic uniform blurring model defined by $\\pmb{y} = \\pmb{k} * \\pmb{x}$ , where $*$ denotes convolution, $\\pmb{k}$ denotes the PSF, and $\\pmb{y}, \\pmb{x}$ denote the blurred image and its sharp correspondence, respectively. Let $\\mathcal{F}$ denote the discrete Fourier transform (DFT). According to the well-known convolution theorem, we have $\\mathcal{F}(\\pmb{y}) = \\mathcal{F}(\\pmb{x}) \\odot \\mathcal{F}(\\pmb{k})$ , where $\\odot$ denotes entry-wise product. Then, one can define a (pesudo) inverse kernel $\\pmb{k}^{\\dagger}$ by: for each $(\\omega_x, \\omega_y)$ ,", + "bbox": [ + 75, + 304, + 468, + 425 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} \\left(\\boldsymbol {k} ^ {\\dagger}\\right) \\left[ \\omega_ {x}, \\omega_ {y} \\right] = \\left(\\mathcal {F} (\\boldsymbol {k}) [ s ]\\right) ^ {- 1}, \\text {i f} | \\mathcal {F} (\\boldsymbol {k}) \\left[ \\omega_ {x}, \\omega_ {y} \\right] | \\neq 0; \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 431, + 468, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and 0 otherwise. Then, for an invertible PSF with only nonzero DFT coefficients, we have", + "bbox": [ + 75, + 463, + 468, + 492 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} (\\boldsymbol {x}) = \\mathcal {F} \\left(\\boldsymbol {k} ^ {\\dagger}\\right) \\odot \\mathcal {F} (\\boldsymbol {y}) \\Longrightarrow \\boldsymbol {x} = \\boldsymbol {k} ^ {\\dagger} * \\boldsymbol {y}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 500, + 467, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When the PSF $k$ is non-invertible, the result $k^{\\dagger} * y$ still provides a fair approximation to $x$ .", + "bbox": [ + 75, + 523, + 468, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One can extend (1) to handling spatially-varying blur. Let $\\mathbb{K}^{\\dagger} = \\{k_{i,j}^{\\dagger}\\}_{i,j}$ denote a set of inverse kernels. Then, the image $\\pmb{y}$ with spatially-varying blur can be recovered by a spatially-varying deconvolution process $\\mathcal{C}$ defined by", + "bbox": [ + 75, + 555, + 468, + 614 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} \\left(\\boldsymbol {y}; \\mathbb {K} ^ {\\dagger}\\right) [ i, j ] = \\sum_ {x, y} \\boldsymbol {k} _ {i, j} ^ {\\dagger} [ x, y ] \\boldsymbol {y} [ i - x, j - y ]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 622, + 467, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Both the shape and the size of $k_{i,j}^{\\dagger}$ depend on the PSFs determined by scene depths and camera settings. When scene depths have significant variations, the sizes of the defocus PSFs also vary significantly, leading to large variations in the size of $k_{i,j}^{\\dagger}$ and a large number of unknowns in $\\mathbb{K}^{\\dagger}$ . Without additional constraints on $\\mathbb{K}^{\\dagger}$ , overfitting can occur.", + "bbox": [ + 75, + 662, + 468, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multi-Scale Dictionary for Inverse Kernels", + "text_level": 1, + "bbox": [ + 76, + 763, + 437, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the overfitting caused by excessive degrees of freedom in $\\mathbb{K}^{\\dagger}$ , we parameterize the inverse kernels using a linear span over a dictionary with reduced number of atoms, resulting in inverse kernels that lie within a low-dimensional space. That is, we express an inverse kernel $\\pmb{k}^{\\dagger}$ as a linear combination of atoms from a dictionary $\\pmb{V} = [v_{1},\\dots ,v_{N}]$ :", + "bbox": [ + 75, + 786, + 468, + 877 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {k} ^ {\\dagger} = \\omega_ {1} \\boldsymbol {v} _ {1} + \\dots + \\omega_ {N} \\boldsymbol {v} _ {N}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 883, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\omega_{1},\\dots ,\\omega_{N}\\in \\mathbb{R}$ are the weights. Additionally, inverse kernels may have a large size. To accurately approximate them using as few atoms as possible, we introduce a multi-scale structure to the dictionary atoms, which is motivated by the observation that when two kernels have the same shape but different sizes, the larger one can be effectively approximated by upsampling the smaller one. Consequently, the dictionary atoms for the corresponding inverse kernels can also be related through such an upsampling process, as stated in the following proposition.", + "bbox": [ + 500, + 90, + 890, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proposition 1. Let $\\pmb{k}^{\\dagger}$ be an inverse kernel defined by (4). Let $\\mathcal{U}_s$ be a standard dyadic upsampling operator with a factor $s$ via expansion. Then", + "bbox": [ + 496, + 248, + 890, + 294 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathcal {U} _ {s} (\\boldsymbol {k})\\right) ^ {\\dagger} = \\omega_ {1} \\mathcal {U} _ {s} \\left(\\boldsymbol {v} _ {1}\\right) + \\dots + \\omega_ {N} \\mathcal {U} _ {s} \\left(\\boldsymbol {v} _ {N}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 303, + 888, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proof. See supplemental material for the proof.", + "bbox": [ + 500, + 330, + 813, + 345 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/84ccd284f25d6e8ba9319f23dd4670275be72eb35bc144e09f67ee8693263c78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 874, + 330, + 890, + 342 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proposition 1 implies that two PFSs with the same shape but different sizes can share the same dictionary for their inverse kernels up to a spatial scaling factor. For defocus PSFs within an image, they typically have similar shapes related to camera settings, as observed in [47]. However, their sizes will vary with the corresponding scene depths. Thus, we represent the inverse kernels of PSFs using different upsampled versions of a dictionary with small atoms. Such a multi-scale structure provides a physics-based implicit prior on the PSFs within the same image and leads to a more compact dictionary for representing inverse kernels of varying sizes. Specifically, we define a multi-scale dictionary composed by the sets (sub-dictionaries) $\\{V^r\\}_{r=1}^R$ :", + "bbox": [ + 496, + 354, + 890, + 551 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {V} ^ {r} = \\left[ \\boldsymbol {v} _ {1} ^ {r}, \\dots , \\boldsymbol {v} _ {N} ^ {r} \\right] \\subset \\mathbb {R} ^ {M _ {r} \\times M _ {r}}, \\boldsymbol {v} _ {n} ^ {r} = \\mathcal {U} _ {S _ {r}} (\\boldsymbol {v} _ {n}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 558, + 888, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\boldsymbol{v}_n^r$ denotes the atom at the $r$ -th scale, $S_r$ denotes an upsampling factor that determines the size parameter $M_r$ of the atom. As $r$ increases, $S_r$ becomes larger and $M_r$ gets smaller. The per-pixel inverse kernels are then expressed as", + "bbox": [ + 500, + 583, + 890, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {k} _ {i, j} ^ {\\dagger} = \\omega_ {1, i, j} \\boldsymbol {v} _ {1} ^ {r _ {i, j}} + \\dots + \\omega_ {N, i, j} \\boldsymbol {v} _ {N} ^ {r _ {i, j}} \\in \\mathbb {R} ^ {M _ {r _ {i, j}} \\times M _ {r _ {i, j}}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 650, + 888, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $r_{i,j}$ denotes the scale factor (index) related to $\\pmb{k}_{i,j}^{\\dagger}$ . Note that the atoms within the same sub-dictionary have a uniform size, while their sizes vary across different sub-dictionaries. That is, each sub-dictionary is utilized to represent inverse kernels of a specific size.", + "bbox": [ + 500, + 679, + 888, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Using the linear representation (7) and the fact that $\\pmb{v}_n^{r_{i,j}}*$ $\\pmb {y} = \\sum_r\\delta (r - r_{i,j})\\pmb{v}_n^r *\\pmb {y}$ with a Dirac delta function $\\delta$ , we can rewrite Eq. (3) as", + "bbox": [ + 500, + 755, + 888, + 801 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {C} (\\boldsymbol {y}) = \\sum_ {i, j} \\mathbf {1} _ {i, j} \\odot \\left(\\sum_ {n = 1} ^ {N} \\omega_ {n, i, j} \\boldsymbol {v} _ {n} ^ {r _ {i, j}} * \\boldsymbol {y}\\right) (8) \\\\ = \\sum_ {i, j} \\mathbf {1} _ {i, j} \\odot \\left(\\sum_ {n = 1} ^ {N} \\sum_ {r = 1} ^ {R} \\mu_ {r, i, j} \\omega_ {n, i, j} \\boldsymbol {v} _ {n} ^ {r} * \\boldsymbol {y}\\right), (9) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 809, + 888, + 895 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12602", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{1}_{i,j}$ denotes a mask with 1 for the $(i,j)$ -th entry being 1 and 0 otherwise, and $\\mu_{r,i,j} = \\delta (r - r_{i,j})$ . We can then perform spatially-varying inverse filtering by convolving each atom $\\pmb{v}_n^r$ in the dictionary with the input image $\\pmb{y}$ , and summing the resulting images by the spatially-varying weights $\\mu_{s,i,j}\\omega_{n,i,j}$ . The weights $\\mu_{s,i,j}$ and $\\omega_{n,i,j}$ control the size (adaptive to scene depth) and shape (adaptive to image) of the corresponding inverse kernel, respectively.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. INR Models for Dictionaries of Inverse Kernels", + "text_level": 1, + "bbox": [ + 76, + 219, + 468, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The dictionary $\\{V_r\\}_{r=1}^R$ used in (8) is generated by the base atoms $\\{v_j\\}_{j=1}^N$ of the smallest size. However, simply upsampling these base atoms cannot cover all high frequencies, i.e., certain high-pass filters of large sizes may not be well approximated by the linear span of these atoms obtained by plain upsampling. To address this limitation, we use INR to re-parameterize the multi-scale dictionary. INR provides a more expressive way to generate multi-scale atoms that cover more frequencies than plain upsampling. Additionally, INR provides implicit regularization to alleviate overfitting by preferring image structures over random noise, as shown in [42, 19, 50].", + "bbox": [ + 75, + 241, + 468, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we express $\\pmb{v}_n^r$ as", + "bbox": [ + 96, + 424, + 297, + 440 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {n} ^ {r} [ x, y ] = \\Phi_ {n} (x, y), [ x, y ] \\in [ 1, \\dots , M _ {r} ] \\times [ 1, \\dots , M _ {r} ],\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 446, + 465, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[x,y]$ denotes a spatial coordinate, and $\\Phi$ is an INR model implemented by a compact multi-layer perceptron (MLP) of a small size. The INR model maps spatial coordinates to a continuous function that implicitly represents the kernel atom. This function can be evaluated at any point in space to interpolate a kernel atom of an arbitrary size, forming a dictionary of inverse kernels with arbitrary scales.", + "bbox": [ + 75, + 469, + 468, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Implicit Neural Inverse Kernel Network", + "text_level": 1, + "bbox": [ + 76, + 587, + 434, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our proposed NN called INIKNet consists of recurrent deblurring stages over different scales for exploiting cross-scale contextual information. Each deblurring stage predicts and applies inverse kernels to deblur the input image. Concretely, each stage consists of two steps: (i) given an input image $\\pmb{y}$ , inferring two weight tensors $\\pmb{\\mu},\\pmb{\\omega}$ that contain the weights $\\{\\mu_{r,i,j}\\}_{r,i,j}$ and $\\{\\omega_{n,i,j}\\}_{n,i,j}$ respectively; and (ii) applying inverse filtering defined by (8) to deblurring $\\pmb{y}$ .", + "bbox": [ + 75, + 613, + 468, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Duplex Scale-Recurrent Framework", + "text_level": 1, + "bbox": [ + 76, + 741, + 393, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Current scale-recurrent deblurring NNs, e.g. [36, 28, 48, 12, 54, 8, 15], typically use an LSTM-based coarse-to-fine framework. The deblurred image from the current (coarser) scale is attached to the input of the next (finer) scale, resulting in unidirectional feature flow that only considers coarser scales when inferring at finer scales. This approach cannot fully leverage the additional information available at finer scales for deblurring on coarser scales, especially when the blur effects show strong similarity over space and scale.", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this limitation, we propose a duplex multiscale processing framework, which leverages bi-LSTM [13] as the sub-NNs to predict $\\mu$ and $\\omega$ . It improves predictions at coarser scales by incorporating additional information from finer scales through the feature flow in the fine-to-coarse pass. This refined information is then utilized to further improve predictions on finer scales via the feature flow through the coarse-to-fine pass. As a result, the prediction at the current scale benefits from the information extracted from both the previous and subsequent scales.", + "bbox": [ + 496, + 90, + 890, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 1, our INIKNet first downsamples the input image to multi-scale versions. For each scale, a U-Net shared across scales is first applied for extracting spatial features from the input. The extracted features are then fed to two bi-LSTM sub-NNs so as to exploit the dependencies among scales for feature refinement. The outputs of two bi-LSTMs sub-NNs are then used to predict $\\mu$ and $\\omega$ via an $1\\times 1$ convolutional layer, respectively. Afterward, the predicted weights are used by an Inverse Filtering Module (IFM) to perform deblurring using (8) to output the result.", + "bbox": [ + 496, + 242, + 890, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our INIKNet differs from existing coarse-to-fine NNs in that it does not concatenate the output image from the current scale to the input of the next scale, as this operation would not only make the feature flow unidirectional, but also prevent the current-scale processing from starting, until the deblurring is fully completed at the previous scale. By removing this operation, the U-Net and bi-LSTM can start running earlier, without waiting for the deblurred images from previous scales. Moreover, by cancelling the operation, the IFM only needs to be called at the original image scale during inference, even though it is still required during training to calculate the multi-scale loss.", + "bbox": [ + 496, + 393, + 892, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Details of Key Modules", + "text_level": 1, + "bbox": [ + 500, + 585, + 714, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Bi-LSTM sub-NNs Recall that the weights $\\mu$ and $\\omega$ correspond to two distinct properties of inverse kernels, namely size and shape, respectively. Their dependency on image scales also differs significantly. Therefore, rather than use a single bi-LSTM sub-NN to predict $\\mu$ and $\\omega$ jointly, we use two parallel bi-LSTM sub-NNs for separate predictions. Both sub-NNs share the same structure, consisting of convolutional LSTM cells [45] arranged in two paths for fine-to-coarse and coarse-to-fine estimations in opposite directions. Each LSTM cell takes a cell state and a hidden state as input and outputs updated states. The input cell state is defined as the features extracted by the U-Net, while the input hidden state for the next scale is set to the output one from the previous scale, forming a cross-scale connection.", + "bbox": [ + 496, + 609, + 890, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Weight predictors To convert the output features from the bi-LSTM sub-NNs to the weights $\\mu$ and $\\omega$ , two $1 \\times 1$ convolutional layers are applied, respectively. For predicting $\\mu$ , the Softmax function is used to impose non-negativity and $\\ell_1$ -normalization constraints. Such a soft relaxation of", + "bbox": [ + 496, + 825, + 892, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12603", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d18ae44616a483e33cf54e273e30b5a81a94276c9dd19a50eb96e85fca5e4260.jpg", + "image_caption": [ + "Figure 1: Architecture of proposed INIKNet for SIDD." + ], + "image_footnote": [], + "bbox": [ + 83, + 93, + 478, + 445 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a65389a81c005a72bc9106fb2b811f9eb8ee4eae426d77a2859443fe518586bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 92, + 885, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "weights $\\mu_{r,i,j}$ from 0,1 to [0,1] not only makes training easier, but also improves the representation accuracy of inverse kernels by including the use of additional atoms. This can be useful when the inverse kernel sizes fall between the gaps among the predefined non-consecutive atom sizes.", + "bbox": [ + 75, + 493, + 468, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "INR models of kernel atoms The MLPs used for INR have a structure similar to those in [46]. They comprise three fully-connected (FC) layers separated by a Sine activation function with a learnable frequency parameter that controls frequency content. The 2D coordinates are first mapped to a high-dimensional space and then projected to single values. One difference from [46] is that we first transform spatial coordinates to polar coordinates for slight improvement.", + "bbox": [ + 75, + 571, + 468, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "IFM for deblurring The IFM first calls the learned MLPs to generate all kernel atoms $\\{\\pmb{v}_n^r\\}_{n,r}$ and then calculates the feature tensors $\\pmb{f}[n,r] = \\pmb{v}_n^r*\\pmb{y}$ . Afterward, Eq. (8) is implemented by simple entry-wise product and summation operations on the tensors $\\pmb{f},\\pmb{\\mu},\\pmb{\\omega}$ .", + "bbox": [ + 75, + 696, + 468, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Training Loss", + "text_level": 1, + "bbox": [ + 76, + 785, + 220, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to the $\\ell_2$ loss $\\mathcal{L}_{\\mathrm{MSE}}$ defined by mean squared error (MSE), two loss functions often seen in image recovery are also used for training, including the frequency-domain loss $\\mathcal{L}_{\\mathrm{FD}}$ [6] defined as the $\\ell_1$ distance in the frequency domain, and the Learned Perceptual Image Patch Similarity (LPIPS) loss $\\mathcal{L}_{\\mathrm{LPIPS}}$ [57] predicted by a pre", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "trained NN. These three loss functions measure deblurring quality in the spatial, frequency and feature (perceptual) domains, respectively. The total loss is then given by", + "bbox": [ + 498, + 493, + 890, + 540 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {p} \\mathcal {L} _ {\\mathrm {M S E}} ^ {(p)} + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {F D}} ^ {(p)} + \\lambda_ {2} \\mathcal {L} _ {\\mathrm {L P I P S}} ^ {(p)}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 551, + 890, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p$ denotes the index of the $p$ th scale, and $\\lambda_1, \\lambda_2$ are empirically set to 0.1. Ground-truth images are downsampled accordingly for the supervision at different scales.", + "bbox": [ + 498, + 597, + 890, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 659, + 633, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and metrics Following [2, 41], we train models on two widely-used datasets, DPDD [2] and LFDOF [41], respectively. To assess their generalization performance, we also evaluate these models on two more datasets, RTF [7] and RealDOF [21]. Three quality metrics are used for quantitative evaluation: Peak Signal to Noise Ratio (PSNR, measured in dB), Structural SIMilarity index (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [57]. The model complexity is measured by both the number of parameters and the inference time on a $640 \\times 640$ image using an NVIDIA 2080Ti RTX GPU.", + "bbox": [ + 496, + 685, + 890, + 851 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details Throughout the experiments, we employ 8 scales and 10 INR-based atoms per scale: $R = 8$ and $N = 10$ . We use atoms of sizes $1 \\times 1, 3 \\times 3, \\dots, 15 \\times 15$ ,", + "bbox": [ + 498, + 854, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12604", + "bbox": [ + 480, + 944, + 517, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $1 \\times 1$ atom is fixed as a delta kernel for modeling in-focus regions. The frequency parameters of the sine activation functions in all INR models are initialized by random sampling from [2, 16]. We call Adam for training, with $4 \\times 10^{5}$ iterations and a batch size of 4. We start with an initial learning rate of $2 \\times 10^{-4}$ and halve it every $10^{5}$ iterations. Random flipping, rotation, and cropping (to $256 \\times 256$ pixels) are performed for training data augmentation.", + "bbox": [ + 75, + 90, + 470, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Performance Evaluation", + "text_level": 1, + "bbox": [ + 76, + 220, + 302, + 237 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared methods We select 9 NN-based SIDD methods for comparison, including the DPDNet [2], AIFNet [41], IFANet [21], KPAC [47], GKMNet [36], MDP [1], DRBNet [40], Restormer [53], and MPRNet [26] trained in [23] with a misalignment-robust scheme. The experimental results of the compared methods are reported either from their original papers or obtained by using their released pretrained models and codes. AIFNet is a model trained on LFDOF with the additional use of SYNDOF for its DME step. To compare its performance on DPDD, we retrain it using both DPDD and SYNDOF. DRBNet is a model trained on LFDOF and fine-tuned on DPDD. For a fair comparison, we retrain it on both datasets, respectively.", + "bbox": [ + 75, + 244, + 468, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on DPDD and LFDOF Table 1 presents the quantitative results on DPDD and LFDOF. Our INIKNet performs the best in PSNR on DPDD and in all three metrics on LFDOF, despite using a lightweight model whose parameter number is the second smallest and around $13.2\\%$ of that of Restormer, a general restoration model. Compared to the latest SIDD-specified MRPNet model [23], INIKNet achieves over $0.3\\mathrm{dB}$ PSNR improvement on DPDD while having only about $10\\%$ parameters to learn. Similar results are observed in the comparison with DRBNet, indicating that the performance gain of INIKNet comes from the design of NN structure rather than the increased model size. Though INIKNet is slightly larger than GKMNet, the smallest model, it achieves noticeably better performance. Compared to KPAC that uses dilated convolutions to mimic inverse kernels, INIKNet performs significantly better. In conclusion, the superiority and compactness of INIKNet demonstrate the effectiveness of its architecture design.", + "bbox": [ + 75, + 444, + 470, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generalization analysis on RealDOF and RTF The results of the DPDD/LFDOF-trained models evaluated on RealDOF and RTF are listed in Table 2. Our DPDD-trained INIKNet shows good generalization to different datasets and achieves the best performance on both RealDOF and RTF in all three metrics. The LFDOF-trained INIKNet is the top performer on RealDOF and achieves the best SSIM score on RTF, as well as the second-best scores in PSNR and LPIPS. Though AIFNet performs better in PSNR and LPIPS, it requires an additional dataset for DME and is much larger than INIKNet. In summary, INIKNet exhibits superior generalizability over other compared models.", + "bbox": [ + 75, + 719, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4e4f2879d70f81dfcd3d2ea2f8fa1c731d07365399f7d20d7fb6436496488b83.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelDPDDLFDOF#Params (106)Time (sec.)
PSNRSSIMLPIPSPSNRSSIMLPIPS
DPDNet24.3480.7470.277---32.251.430
AIFNet24.2130.7420.30929.6770.8840.20241.550.276
IFANet25.3660.7890.21729.7870.8720.15610.480.063
KPAC25.2210.7740.22628.9420.8570.1742.060.037
GKMNet25.4680.7890.21929.0810.8670.1711.410.129
MDP25.3470.7630.26828.0690.8340.18546.863.317
DRBNet25.4850.7920.25430.2530.8830.14711.690.085
MPRNet25.7300.7920.232---20.101.161
Restormer25.9800.8110.178---26.100.672
INIKNet26.0550.8030.18530.2930.8860.1321.980.228
", + "bbox": [ + 501, + 89, + 890, + 247 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/47a3d551c9f410106060ab10e255fecf72ae676dc94c160a41bf9ddd8930a6d2.jpg", + "table_caption": [ + "Table 1: Quantitative results on DPDD/LFDOF. Bold for best performers and underline for second-best performers." + ], + "table_footnote": [], + "table_body": "
ModelRealDOFRTF#Params (106)
PSNRSSIMLPIPSPSNRSSIMLPIPS
Trained on DPDDDPDNet22.8700.6700.42523.6080.5910.29632.25
AIFNet23.0930.6800.41324.0410.7580.28941.55
MDP23.5000.6810.44424.0120.7380.31246.86
KPAC23.9750.7620.33824.6180.7770.2362.06
IFANet24.7120.7480.30624.9240.8010.22710.48
GKMNet24.2540.7320.39224.9700.7890.2611.41
DRBNet24.8840.7510.37624.4630.7730.31111.69
MPRNet24.5410.7360.33924.5880.7880.30420.10
Restormer25.0910.7620.28524.2120.8210.22426.10
INIKNet25.2310.7650.28725.4500.8340.2151.98
Trained on LFDOFIFANet22.5040.6690.48326.4370.8380.23810.48
KPAC22.5500.6710.45725.9590.8030.2302.06
AIFNet22.6230.6670.46127.5520.8820.17641.55
GKMNet23.6090.7210.40826.9850.8560.2461.41
MDP22.7260.6800.45325.5800.8090.22846.86
DRBNet22.9100.6910.43726.7170.8530.20011.69
INIKNet23.8100.7240.35627.4010.8850.1791.98
", + "bbox": [ + 506, + 299, + 885, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2: Quantitative results on RealDOF/RTF. Bold for best performers and underline for second-best performers.", + "bbox": [ + 498, + 556, + 890, + 587 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative comparison We compare INIKNet with some SIDD-specified competitors on DPDD and LFDOF via visual inspection. See Fig. 2 for some deblurred images on four datasets, where complex textures such as those on wire fence and branches are successfully recovered by INIKNet. See also Fig. 3 for some results on the CUHK dataset [43]. While most methods fail under large blur, ours yields more visually satisfying results. These findings demonstrate that our INIKNet performs well in terms of visual quality.", + "bbox": [ + 496, + 603, + 890, + 741 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Ablation Studies and Analysis", + "text_level": 1, + "bbox": [ + 500, + 752, + 766, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablation studies We construct several variants of INIKNet and list their results on two datasets in Table 3.", + "bbox": [ + 498, + 776, + 890, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Non-MS Dict: We drop the multi-scale structure of the dictionary and instead use an individual INR model for each dictionary atom at each scale (thus yielding a larger model size). A noticeable performance degradation occurs. This is because the multi-scale structure of the INR-based dictionary imposes structural regularization to reduce overfitting.", + "bbox": [ + 496, + 809, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12605", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d738c76f004176285f2135664e6a221410f8981637ba70da43a0978e2c1dc83e.jpg", + "image_caption": [ + "Figure 2: Results on using models trained on DPDD (left part) and LFDOF (right part). Zoom-in for better views." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 529, + 496 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/259a5b098bb38f3343912588c4db9a3572482b1e03416ac5a055b6a1ff2dfd47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 85, + 880, + 496 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bd3368dbe2335aa8263df3c0540c6b5bf50a73f57ad5efee0d42d4cdba24be8b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelRealDOFRTF
PSNRSSIMLPIPSPSNRSSIMLPIPS
Non-MS Dict24.9160.7490.31025.1360.8090.254
Non-INR Dict25.0580.7540.30325.2370.8190.226
Unidirectional25.0990.7530.29425.3050.8240.227
Original INIKNet25.2310.7650.28725.4500.8340.215
", + "bbox": [ + 84, + 542, + 460, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3: Results of baseline models trained on DPDD.", + "bbox": [ + 91, + 636, + 452, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Non-INR Dict: When the INR-based dictionaries are replaced by the plain ones, where each base atom $v_{n}$ is defined as a learnable matrix and bi-linearly upsampled to form the multi-scale dictionary, the performance of INIKNet decreases. This is probably because the frequency content for the atoms generated by plain upsampling are limited, so do their generated inverse kernels.", + "bbox": [ + 75, + 667, + 468, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unidirectional: To simulate an unidirectional scale recurrent mechanism (e.g. [36]), we replace the bi-LSTM with double unidirectional LSTMs (for keeping the model size). The output from the last scale is attached to the input of the current scale for additional information. Such a scheme for coefficient prediction results in some performance degradation, highlighting the effectiveness of our proposed duplex scale-recurrent framework.", + "bbox": [ + 75, + 780, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization of coefficients and atoms The coefficient maps $\\mu$ and $\\omega$ are visualized in Fig. 4. We can see that the regions with larger blur amount tend to have larger coefficients $\\mu$ assigned to larger inverse kernels, and vice versa. In comparison, the coefficients $\\omega$ are more spatially similar (gray), consistent with the prior of spatially-similar shapes of inverse kernels. We use Fig. 5 to visualize some learned (non-)INR-based atoms in both the spatial and Fourier domain. The INR-based atoms within the same scale contain diverse patterns with varying frequencies and orientations, and meanwhile they have similar shapes across scales. Compared to the non-INR-based ones, the INR-based atoms encode richer spatial structures and frequency contents. Interestingly, the INR-based atoms exhibit strong isotropy, which is likely due to the isotropy of defocus PSFs (so do their inverse kernels). All these properties of the INR-based atoms enable our INIKNet to effectively represent and predict inverse kernels for spatially-varying defocus blur.", + "bbox": [ + 496, + 546, + 890, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 829, + 619, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This paper presented an interpretable end-to-end SIDD approach, which learns a duplex scale-recurrent NN to predict INR-driven dictionary-parameterized inverse kernels of", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12606", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f453722cad9142760713cb963b3cd48a9a288922863cd894392a109d48c3a1a5.jpg", + "image_caption": [ + "Figure 3: Results on CUHK dataset (without ground-truths) using models trained on DPD. Zoom-in for better views." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 890, + 492 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4ec66da483cb1d5e4b7c54ccb4f098f49ad562d4a69bc32559ee44057a86a917.jpg", + "image_caption": [ + "Figure 4: Visualization of $\\pmb{\\mu}$ maps and $\\omega$ maps." + ], + "image_footnote": [], + "bbox": [ + 84, + 529, + 888, + 664 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2c75e98d78c435e3ac290ac63b8e0a99078eccb3c236168faaf0ec7d518ab6ac.jpg", + "image_caption": [ + "Figure 5: Visualization of bilinearly upsampled (non-INR-based) atoms and INR-based atoms in two domains." + ], + "image_footnote": [], + "bbox": [ + 84, + 702, + 883, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "defocus blur and applies them to deblur images. Our proposed approach introduced effective structural regularization on inverse kernels for good generalization, which is", + "bbox": [ + 76, + 854, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "based on the shape similarity of defocus PSFs over an image. The resulting effective and light-weight model has demonstrated its advantages by extensive experiments.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12607", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abdullah Abuolaim, Mahmoud Afifi, and Michael S. Brown. Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1231-1239, 2022. 1, 2, 6", + "[2] Abdullah Abuolaim and Michael S. Brown. Defocus deblurring using dual-pixel data. In Proceedings of the European Conference on Computer Vision, pages 111-126. Springer, 2020. 2, 5, 6", + "[3] Yosuke Bando and Tomoyuki Nishita. Towards digital refocusing from a single photograph. In Proceedings of the Pacific Conference on Computer Graphics and Applications, pages 363-372. IEEE, 2007. 1", + "[4] Mingqin Chen, Yuhui Quan, Tongyao Pang, and Hui Ji. Non-blind image deconvolution via leveraging model uncertainty in an untrained deep neural network. International Journal of Computer Vision, 130(7):1770-1789, 2022. 2", + "[5] Sunghyun Cho and Seungyong Lee. Convergence analysis of map based blur kernel estimation. In Proceedings of the IEEE/International Conference on Computer Vision, pages 4808-4816, 2017. 1", + "[6] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4641-4650, 2021. 3, 5", + "[7] Laurent D'Andrés, Jordi Salvador, Axel Kochale, and Sabine Süssstrunk. Non-parametric blur map regression for depth of field extension. IEEE Transactions on Image Processing, 25(4):1660-1673, 2016. 1, 5", + "[8] Senyou Deng, Wenqi Ren, Yanyang Yan, Tao Wang, Fenglong Song, and Xiaochun Cao. Multi-scale separable network for ultra-high-definition video deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14030-14039, 2021. 2, 4", + "[9] Jiangxin Dong, Stefan Roth, and Bernt Schiele. Deep wiener deconvolution: Wiener meets deep learning for image deblurring. Advances in Neural Information Processing Systems, 33:1048-1059, 2020. 2", + "[10] Jiangxin Dong, Stefan Roth, and Bernt Schiele. Dwdn: deep wiener deconvolution network for non-blind image deblurring. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9960-9976, 2021. 2", + "[11] Zhenxuan Fang, Fangfang Wu, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Self-supervised non-uniform kernel estimation with flow-based motion prior for blind image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18105-18114, 2023. 2", + "[12] Hongyun Gao, Xin Tao, Xiaoyong Shen, and Jiaya Jia. Dynamic scene deblurring with parameter selective sharing and nested skip connections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3848-3856, 2019. 2, 4", + "[13] Alex Graves and Jürgen Schmidhuber. Framewise phoneme classification with bidirectional LSTM and other neural net-" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "work architectures. Neural networks, 18(5-6):602-610, 2005. 2, 4", + "[14] Fang Hua, Peter Johnson, Nadezhda Sazonova, Paulo Lopez-Meyer, and Stephanie Schuckers. Impact of out-of-focus blur on face recognition performance based on modular transfer function. In Proceedings of the International Conference on Biometrics, pages 85–90. IEEE, 2012. 1", + "[15] Bo Ji and Angela Yao. Multi-scale memory-based video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1919-1928, 2022. 2, 4", + "[16] Ali Karaali, Naomi Harte, and Claudio R. Jung. Deep multiscale feature learning for defocus blur estimation. IEEE Transactions on Image Processing, 31:1097-1106, 2022. 2", + "[17] Ali Karaali and Claudio Rosito Jung. Edge-based defocus blur estimation with adaptive scale selection. IEEE Transactions on Image Processing, 27(3):1126-1137, Mar. 2018. 1, 2", + "[18] Ali Karaali and Cláudio Rosito Jung. Svbr-net: A non-blind spatially varying defocus blur removal network. In Proceedings of the IEEE International Conference on Image Processing, pages 566-570. IEEE, 2022. 1", + "[19] Chaewon Kim, Jaeho Lee, and Jinwoo Shin. Zero-shot blind image denoising via implicit neural representations. arXiv preprint arXiv:2204.02405, 2022. 3, 4", + "[20] Dilip Krishnan and Rob Fergus. Fast image deconvolution using hyper-laplacian priors. In Advances in Neural Information Processing Systems, 2009. 2", + "[21] Junyong Lee, Hyeongseok Son, Jaesung Rim, Sunghyun Cho, and Seungyong Lee. Iterative filter adaptive network for single image defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2034-2042, 2021. 1, 2, 5, 6", + "[22] Dasong Li, Yi Zhang, Ka Chun Cheung, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. Learning degradation representations for image deblurring. In Proceedings of the European Conference on Computer Vision, pages 736-753. Springer, 2022. 2", + "[23] Yu Li, Dongwei Ren, Xinya Shu, and Wangmeng Zuo. Learning single image defocus deblurring with misaligned training pairs. Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 1, 2, 6", + "[24] Ziwei Luo, Haibin Huang, Lei Yu, Youwei Li, Haoqiang Fan, and Shuaicheng Liu. Deep constrained least squares for blind image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17642-17652, 2022. 2", + "[25] Haoyu Ma, Shaojun Liu, Qingmin Liao, Juncheng Zhang, and Jing-Hao Xue. Defocus image deblurring network with defocus map estimation as auxiliary task. IEEE Transactions on Image Processing, 31:216-226, 2021. 1, 2", + "[26] Armin Mehri, Parichehr B Ardakani, and Angel D Sappa. Mprnet: Multi-path residual network for lightweight image super resolution. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2704-2713, 2021. 6", + "[27] Sung-Jun Min, Kyeongbo Kong, and Suk-Ju Kang. Out-of-focus image deblurring for mobile display vision inspection." + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12608", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1", + "[28] Seungjun Nah, Tae Kim, and Kyoung Mu Lee. Deep multiscale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 257-265, 2017. 2, 4", + "[29] Yuesong Nan and Hui Ji. Deep learning for handling kernel/model uncertainty in image deconvolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2388-2397, 2020. 2", + "[30] Yuesong Nan, Yuhui Quan, and Hui Ji. Variational-EM-based deep learning for noise-blind image deblurring. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3626–3635, 2020. 2", + "[31] Saqib Nazir, Lorenzo Vaquero, Manuel Mucientes, Víctor M Brea, and Daniela Coltuc. Depth estimation and image restoration by deep learning from defocused images. arXiv preprint arXiv:2302.10730, 2023. 1, 2", + "[32] Jinsun Park, Yu-Wing Tai, Donghyeon Cho, and In So Kweon. A unified approach of multi-scale deep and handcrafted features for defocus estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1736-1745, 2017. 1, 2", + "[33] Valeriya Pronina, Filippos Kokkinos, Dmitry V Dylov, and Stamatios Lefkimmiatis. Microscopy image restoration with deep wiener-kolmogorov filters. In Proceedings of the European Conference on Computer Vision, pages 185-201. Springer, 2020. 1, 2", + "[34] Yuhui Quan, Zhuojie Chen, Huan Zheng, and Hui Ji. Learning deep non-blind image deconvolution without ground truths. In Proceedings of the European Conference on Computer Vision, pages 642-659, 2022. 2", + "[35] Yuhui Quan, Peikang Lin, Yong Xu, Yuesong Nan, and Hui Ji. Nonblind image deblurring via deep learning in complex field. IEEE Transactions on Neural Networks and Learning Systems, 33(10):5387-5400, 2022. 2", + "[36] Yuhui Quan, Hui Wu, Zicong, and Hui Ji. Gaussian kernel mixture network for single image defocus deblurring. In Advances in Neural Information Processing Systems, volume 34, 2021. 1, 2, 4, 6, 7", + "[37] Yuhui Quan, Hui Wu, Zicong, and Hui Ji. Neumann network with recursive kernels for single image defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5754-5763, June 2023. 1, 2", + "[38] Wenqi Ren, Jiawei Zhang, Lin Ma, Jinshan Pan, Xiaochun Cao, Wangmeng Zuo, Wei Liu, and Ming-Hsuan Yang. Deep non-blind deconvolution via generalized low-rank approximation. Advances in Neural Information Processing Systems, 31, 2018. 2", + "[39] Wenqi Ren, Jiawei Zhang, Jinshan Pan, Sifei Liu, Jimmy S Ren, Junping Du, Xiaochun Cao, and Ming-Hsuan Yang. Deblurring dynamic scenes via spatially varying recurrent neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):3974-3987, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Lingyan Ruan, Bin Chen, Jizhou Li, and Miuling Lam. Learning to deblur using light field generated and real defocus images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16304-16313, 2022. 1, 2, 6", + "[41] Lingyan Ruan, Bin Chen, Jizhou Li, and Miu-Ling Lam. Aifnet: All-in-focus image restoration network using a light field-based dataset. IEEE Transactions on Computational Imaging, 7:675-688, 2021. 1, 2, 5, 6", + "[42] Nimrod Shabtay, Eli Schwartz, and Raja Giryes. Pip: Positional-encoding image prior. arXiv preprint arXiv:2211.14298, 2022. 3, 4", + "[43] Jianping Shi, Li Xu, and Jiaya Jia. Discriminative blur detection features. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2965-2972, 2014. 6", + "[44] Jianping Shi, Li Xu, and Jiaya Jia. Just noticeable defocus blur detection and estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 657-665, 2015. 1, 2", + "[45] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in Neural Information Processing Systems, 28, 2015. 4", + "[46] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 2, 5", + "[47] Hyeongseok Son, Junyong Lee, Sunghyun Cho, and Seungyong Lee. Single image defocus deblurring using kernel-sharing parallel atrous convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2642-2650, 2021. 1, 2, 3, 6", + "[48] Xin Tao, Hongyun Gao, Yi Wang, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8174-8182, 2018. 2, 4", + "[49] Longguang Wang, Yingqian Wang, Xiaoyu Dong, Qingyu Xu, Jungang Yang, Wei An, and Yulan Guo. Unsupervised degradation representation learning for blind superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10581-10590, 2021. 2", + "[50] Shaowen Xie, Hao Zhu, Zhen Liu, Qi Zhang, You Zhou, Xun Cao, and Zhan Ma. Diner: Disorder-invariant implicit neural representation. arXiv preprint arXiv:2211.07871, 2022. 3, 4", + "[51] Guodong Xu, Yuhui Quan, and Hui Ji. Estimating defocus blur via rank of local patches. In Proceedings of the IEEE/International Conference on Computer Vision, pages 5371-5379, 2017. 1, 2", + "[52] Li Xu, Jimmy S Ren, Ce Liu, and Jiaya Jia. Deep convolutional neural network for image deconvolution. Advances in Neural Information Processing Systems, 27, 2014. 2", + "[53] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12609", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 6", + "[54] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5978-5986, 2019. 2, 4", + "[55] Jiawei Zhang, Jinshan Pan, Jimmy Ren, Yibing Song, Linchao Bao, Rynson WH Lau, and Ming-Hsuan Yang. Dynamic scene deblurring using spatially variant recurrent neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2521-2529, 2018. 2", + "[56] Jie Zhang and Wanming Zhai. Blind attention geometric restraint neural network for single image dynamic/defocus deblurring. IEEE Transactions on Neural Networks and Learning Systems, 2022. 1, 2", + "[57] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 5", + "[58] Wenda Zhao, Fei Wei, You He, and Hutchuan Lu. United defocus blur detection and deblurring via adversarial promoting learning. In Proceedings of the European Conference on Computer Vision, pages 569-586, 2022. 1, 2", + "[59] Wenda Zhao, Fan Zhao, Dong Wang, and Huchuan Lu. Defocus blur detection via multi-stream bottom-top-bottom network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42:1884-1897, 2020. 2", + "[60] Bolun Zheng, Quan Chen, Shanxin Yuan, Xiaofei Zhou, Hua Zhang, Jiyong Zhang, Chenggang Yan, and Gregory Slabaugh. Constrained predictive filters for single image bokeh rendering. IEEE Transactions on Computational Imaging, 8:346-357, 2022. 1", + "[61] Chao Zhu, Hang Dong, Jinshan Pan, Boyang Liang, Yuhao Huang, Lean Fu, and Fei Wang. Deep recurrent neural network with multi-scale bi-directional propagation for video deblurring. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 3598-3607, 2022. 3" + ], + "bbox": [ + 78, + 90, + 468, + 683 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12610", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Single Image Defocus Deblurring via Implicit Neural Inverse Kernels/bee3d863-7fc6-4e0b-a339-87a4afbe8bb4_model.json b/2023/Single Image Defocus Deblurring via Implicit Neural Inverse Kernels/bee3d863-7fc6-4e0b-a339-87a4afbe8bb4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c0b5627433e847513b1ebfe5bfc7bdf50539dcec --- /dev/null +++ b/2023/Single Image Defocus Deblurring via Implicit Neural Inverse Kernels/bee3d863-7fc6-4e0b-a339-87a4afbe8bb4_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.131, + 0.834, + 0.154 + ], + "angle": 0, + "content": "Single Image Defocus Deblurring via Implicit Neural Inverse Kernels" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.18, + 0.677, + 0.199 + ], + "angle": 0, + "content": "Yuhui Quan\\(^{1,2}\\) Xin Yao\\(^{1,2}\\) Hui Ji\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.2, + 0.862, + 0.233 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Engineering, South China University of Technology, Guangzhou 510006, China \\(^{2}\\) Pazhou Lab, Guangzhou 510335, China" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.235, + 0.768, + 0.253 + ], + "angle": 0, + "content": "3 Department of Mathematics, National University of Singapore, 119076, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.255, + 0.769, + 0.27 + ], + "angle": 0, + "content": "cseyhquan@scut.edu.cn, csxinyaoscut@gmail.com, matjh@nus.edu.sg*" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.473, + 0.609 + ], + "angle": 0, + "content": "Single image defocus deblurring (SIDD) is a challenging task due to the spatially-varying nature of defocus blur, characterized by per-pixel point spread functions (PSFs). Existing deep-learning-based methods for SIDD are limited by either over-fitting due to the lack of model constraints or under-parametrization that restricts their applicability to real-world images. To address the limitations, this paper proposes an interpretable approach that explicitly predicts inverse kernels with structural regularization. Motivated by the observation that defocus PSFs within an image often have similar shapes but different sizes, we represent the inverse kernels linearly over a multi-scale dictionary parameterized by implicit neural representations. We predict the corresponding representation coefficients via a duplex scale-recurrent neural network that jointly performs fine-to-coarse and coarse-to-fine estimations. Extensive experiments demonstrate that our approach achieves excellent performance using a lightweight model." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.635, + 0.21, + 0.65 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.811 + ], + "angle": 0, + "content": "When a camera captures an image, objects outside of the focal plane appear blurry. This effect is known as defocus blur and usually occurs when using a large camera aperture or capturing scenes with significantly varying depths. Removing defocus blur from a single image, known as SIDD, is desired for many practical applications; see e.g. [27, 14, 33, 60]. In general, defocus blur varies spatially. Each defocused pixel is a weighted average of its neighboring pixels in the latent all-in-focus image, with the weights determined by a spatially-varying unknown PSF." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.812, + 0.47, + 0.856 + ], + "angle": 0, + "content": "SIDD is closely related to defocus map estimation (DME) [3, 44, 7, 5, 51, 32, 17]. Existing DME methods typically model defocus PSFs using a Gaussian function" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.306, + 0.893, + 0.472 + ], + "angle": 0, + "content": "parameterized by its variance or a disc function parameterized by its radius. The resulting defocus map indicates the per-pixel blur amount in terms of the values of such model parameters, which can then be utilized for SIDD by applying a non-blind image deblurring (NID) algorithm using the corresponding PSFs; see e.g. [18]. However, DME remains a challenging task, and the estimated parameters of PSFs can be erroneous for many pixels. Moreover, Gaussian or disc-indicator functions are overly simplistic for real-world defocus PSFs. All these errors will be magnified in the NID process, resulting in noticeable artifacts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.893, + 0.717 + ], + "angle": 0, + "content": "Recently, many deep-learning-based methods have been proposed for SIDD; see e.g. [41, 25, 21, 47, 36, 40, 1, 56, 58, 23, 31]. Most train an end-to-end neural network (NN) that directly maps blurry images to their sharp correspondences. Although these methods provide better results than the DME-based two-step methods, their performance still has much room for improvement due to significant variations in the spatial distributions of defocus PSFs among different images. Many existing works (e.g. [21, 40, 56]) explicitly introduce spatially-varying processing blocks to better handle the spatial variance of defocus PSFs. However, those blocks lack constraints and may cause overfitting. A few studies (e.g. [36, 37]) represent defocus PSFs by some specific basis for geometric regularization and employ deep unrolling for an NN-based inversion process. Nevertheless, there is still room for improvement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "This paper presents an end-to-end deep NN with high interpretability for SIDD, which explicitly predicts spatially-varying inverse kernels of defocus PSFs and performs deblurring using the predicted inverse kernels. The possible overfitting issue is tackled by providing a more accurate model with structural constraints/regularization for the inverse kernels. Specifically, the inverse kernels are modeled by a linear representation under a dictionary. To achieve a compact yet sufficiently general representation, we construct the dictionary with a multi-scale structure. This is motivated by the empirical observation (as seen in [47]) that defocus PSFs within an image tend to have the same shape" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.47, + 0.901 + ], + "angle": 0, + "content": "*This work is supported by the Natural Science Foundation of Guangdong Province (Grant No. 2022A1515011755 and 2023A1515012841), and Singapore MOE AcRF Tier 1 Grant (WBS No. A-8000981-00-00)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12600" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "but vary mainly in size. We show that if two defocus PSFs of different sizes have the same shape under an upsampling operation, so do the dictionary atoms of their inverse kernels. Thus, the dictionary atoms of a large size are defined as an upsampled version of the atoms of a small size." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.47, + 0.319 + ], + "angle": 0, + "content": "Although plain upsampling offers an economical method for generating larger dictionary atoms from smaller ones, the resulting atoms may have a limited frequency range. This can lead to poor approximations of certain inverse kernels. To address the limitation, we leverage implicit neural representation (INR) [46], a technique that uses coordinate-input NNs to represent geometric objects like 2D shapes, so as to efficiently parameterize multi-scale atoms with both sufficient coverage of high-frequency content and implicit regularization to alleviate overfitting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.469, + 0.472 + ], + "angle": 0, + "content": "For the multi-scale representation coefficients, we separate them into scale-related and shape-related components, which are predicted by two scale-recurrent sub-NNs respectively. To better utilize information from different scales for coefficient prediction, we introduce a duplex scale-recurrent framework that performs both fine-to-coarse and coarse-to-fine estimations. Once the coefficients are predicted, the corresponding inverse kernels are applied to the input image for deblurring. The resulting pipeline is highly efficient and interpretable, leading to a lightweight yet effective model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.469, + 0.532 + ], + "angle": 0, + "content": "To conclude, this paper proposes an end-to-end NN for SIDD, which exhibits a noticeable performance boost compared to existing ones, while maintaining low complexity. See below for our main contributions:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.54, + 0.468, + 0.57 + ], + "angle": 0, + "content": "- A parametric inverse kernel prediction framework for SIDD, providing efficient and interpretable deep NNs." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.574, + 0.468, + 0.619 + ], + "angle": 0, + "content": "- A multi-scale linear representation model for the inverse kernels of spatially-varying defocus blur, using an INR-based multi-scale dictionary." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.623, + 0.468, + 0.652 + ], + "angle": 0, + "content": "- A duplex scale-recurrent framework for predicting the representation coefficients of inverse kernels." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.54, + 0.468, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.669, + 0.227, + 0.685 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.695, + 0.347, + 0.712 + ], + "angle": 0, + "content": "2.1. Existing Approaches for SIDD" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.469, + 0.854 + ], + "angle": 0, + "content": "SIDD can done by a two-step approach that combines DME (e.g. [44, 51, 17, 32]) and NID (e.g. [20]). However, such an approach has two inherent drawbacks. First, the errors in DME will be magnified in the following NID process. Second, simple Gaussian or disc functions do not fit real-world defocus PSFs well. While there has been steady progress in deep-learning-based methods for DME [59, 16] and for NID [38, 10, 33, 29, 30, 4, 34, 35], the inherent drawbacks of the two-step approach remain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Deep end-to-end learning has emerged as a promising approach for SIDD. The seminal work conducted by Abuolaim and Brown [2] trained an encoder-decoder CNN that" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "maps a defocused image to its in-focus correspondence, with significant performance gain over two-step methods. To better handle spatially-varying defocus blur, some studies introduced dynamic processing blocks, e.g., per-pixel filter prediction in Lee et al. [21], dynamic residual blocks in Ruan et al. [40], and attention in Zhang and Zhai [56]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.892, + 0.289 + ], + "angle": 0, + "content": "The use of dynamic processing blocks may increase the model complexity and the risk of overfitting. Regularization by auxiliary tasks is an effective approach for improving the generalization performance. In existing works, auxiliary tasks are often defined using DME ([25, 41, 58]), depth estimation [31], dual-view data [21, 1], or reblurring loss [23]. However, most of them require additional data sources." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.893, + 0.591 + ], + "angle": 0, + "content": "A more appealing approach is encoding physical priors of defocus blur into NNs. Quan et al. [36, 37] used a kernel mixture model to parameterize defocus PSFs and constructed a deep unrolling NN. This method can be viewed as the degradation learning [22, 49, 24, 11] that predicts the blurring operator or kernels for deblurring. Our approach differs in that it predicts inverse kernels, not the degradation with defocus PSFs. Son et al. [47] showed that inverse defocus kernels have an invariant shape across scales and emulated them in feature spaces via a series of kernel-sharing parallel Atrous convolutions. Their method inspired our work, and it can be viewed as representing inverse kernels for a feature tensor via a multi-scale dilated combination of a single atom and predicting spatially-invariant combination coefficients via channel attention. In comparison, our approach explicitly predicts inverse kernels on the input image, allowing better interpretability, and represents them using INR to improve effectiveness. Additionally, we predict spatially-varying combination coefficients to better handle spatially-similar (not invariant) defocus shapes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.602, + 0.687, + 0.618 + ], + "angle": 0, + "content": "2.2. Related Techniques" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.777 + ], + "angle": 0, + "content": "Deep inverse kernels Several studies have included inverse kernels in deblurring NNs, in addition to [47]. Xu et al. [52] and Ren et al. [38] presented the CNNs that adopt tensor product for emulating separable inverse kernels to perform NID. Dong et al. [9, 10] and Pronina et al. [33] performed NID via Wiener inverse filtering in deep feature spaces. However, all these methods are designed for spatially-invariant blur. Ren et al. [55, 39] proposed a recurrent NN to mimic inverse kernels for removing spatially-varying blur of dynamic scenes, but not for SIDD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Scale-recurrent NNs for deblurring Scale-recurrent NNs are commonly used for image deblurring, e.g. [36, 28, 48, 12, 54, 8, 15]. These methods typically employ coarse-to-fine estimation which is unidirectional and thus cannot fully leverage information from the subsequent scales for the prediction at the current scale. In contrast, our proposed NN utilizes a bi-directional structure via bi-LSTMs [13] for joint coarse-to-fine and fine-to-coarse estimation. This is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12601" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "particularly useful for SIDD, as defocus blur effects show strong self-similarity over scales. While bi-LSTM has been used in [61] for multi-scale video deblurring, it is only used for bi-directional temporal processing and the multi-scale processing is still one-way. Also note that the multi-scale NN proposed in [6] for motion deblurring is not restrictive to unidirectional processing, but not in a recurrent manner." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.2, + 0.469, + 0.245 + ], + "angle": 0, + "content": "INR for image recovery INR has been applied to image recovery [42, 19, 50]. In comparison to these works, we use INR to parameterize inverse kernels, not images, for SIDD." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.257, + 0.453, + 0.273 + ], + "angle": 0, + "content": "3. INR-Based Dictionary for Inverse Kernels" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.282, + 0.439, + 0.298 + ], + "angle": 0, + "content": "3.1. Inverse Kernels for Spatially-Varying Blur" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.469, + 0.426 + ], + "angle": 0, + "content": "Consider a simplistic uniform blurring model defined by \\( \\pmb{y} = \\pmb{k} * \\pmb{x} \\), where \\( * \\) denotes convolution, \\( \\pmb{k} \\) denotes the PSF, and \\( \\pmb{y}, \\pmb{x} \\) denote the blurred image and its sharp correspondence, respectively. Let \\( \\mathcal{F} \\) denote the discrete Fourier transform (DFT). According to the well-known convolution theorem, we have \\( \\mathcal{F}(\\pmb{y}) = \\mathcal{F}(\\pmb{x}) \\odot \\mathcal{F}(\\pmb{k}) \\), where \\( \\odot \\) denotes entry-wise product. Then, one can define a (pesudo) inverse kernel \\( \\pmb{k}^{\\dagger} \\) by: for each \\( (\\omega_x, \\omega_y) \\)," + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.433, + 0.469, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} \\left(\\boldsymbol {k} ^ {\\dagger}\\right) \\left[ \\omega_ {x}, \\omega_ {y} \\right] = \\left(\\mathcal {F} (\\boldsymbol {k}) [ s ]\\right) ^ {- 1}, \\text {i f} | \\mathcal {F} (\\boldsymbol {k}) \\left[ \\omega_ {x}, \\omega_ {y} \\right] | \\neq 0; \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.464, + 0.469, + 0.493 + ], + "angle": 0, + "content": "and 0 otherwise. Then, for an invertible PSF with only nonzero DFT coefficients, we have" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.501, + 0.468, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} (\\boldsymbol {x}) = \\mathcal {F} \\left(\\boldsymbol {k} ^ {\\dagger}\\right) \\odot \\mathcal {F} (\\boldsymbol {y}) \\Longrightarrow \\boldsymbol {x} = \\boldsymbol {k} ^ {\\dagger} * \\boldsymbol {y}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.469, + 0.555 + ], + "angle": 0, + "content": "When the PSF \\( k \\) is non-invertible, the result \\( k^{\\dagger} * y \\) still provides a fair approximation to \\( x \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.556, + 0.469, + 0.615 + ], + "angle": 0, + "content": "One can extend (1) to handling spatially-varying blur. Let \\(\\mathbb{K}^{\\dagger} = \\{k_{i,j}^{\\dagger}\\}_{i,j}\\) denote a set of inverse kernels. Then, the image \\(\\pmb{y}\\) with spatially-varying blur can be recovered by a spatially-varying deconvolution process \\(\\mathcal{C}\\) defined by" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.623, + 0.468, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} \\left(\\boldsymbol {y}; \\mathbb {K} ^ {\\dagger}\\right) [ i, j ] = \\sum_ {x, y} \\boldsymbol {k} _ {i, j} ^ {\\dagger} [ x, y ] \\boldsymbol {y} [ i - x, j - y ]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.664, + 0.469, + 0.756 + ], + "angle": 0, + "content": "Both the shape and the size of \\( k_{i,j}^{\\dagger} \\) depend on the PSFs determined by scene depths and camera settings. When scene depths have significant variations, the sizes of the defocus PSFs also vary significantly, leading to large variations in the size of \\( k_{i,j}^{\\dagger} \\) and a large number of unknowns in \\( \\mathbb{K}^{\\dagger} \\). Without additional constraints on \\( \\mathbb{K}^{\\dagger} \\), overfitting can occur." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.764, + 0.439, + 0.78 + ], + "angle": 0, + "content": "3.2. Multi-Scale Dictionary for Inverse Kernels" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.787, + 0.469, + 0.878 + ], + "angle": 0, + "content": "To address the overfitting caused by excessive degrees of freedom in \\(\\mathbb{K}^{\\dagger}\\), we parameterize the inverse kernels using a linear span over a dictionary with reduced number of atoms, resulting in inverse kernels that lie within a low-dimensional space. That is, we express an inverse kernel \\(\\pmb{k}^{\\dagger}\\) as a linear combination of atoms from a dictionary \\(\\pmb{V} = [v_{1},\\dots ,v_{N}]\\):" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.885, + 0.468, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {k} ^ {\\dagger} = \\omega_ {1} \\boldsymbol {v} _ {1} + \\dots + \\omega_ {N} \\boldsymbol {v} _ {N}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.243 + ], + "angle": 0, + "content": "where \\(\\omega_{1},\\dots ,\\omega_{N}\\in \\mathbb{R}\\) are the weights. Additionally, inverse kernels may have a large size. To accurately approximate them using as few atoms as possible, we introduce a multi-scale structure to the dictionary atoms, which is motivated by the observation that when two kernels have the same shape but different sizes, the larger one can be effectively approximated by upsampling the smaller one. Consequently, the dictionary atoms for the corresponding inverse kernels can also be related through such an upsampling process, as stated in the following proposition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.249, + 0.892, + 0.295 + ], + "angle": 0, + "content": "Proposition 1. Let \\( \\pmb{k}^{\\dagger} \\) be an inverse kernel defined by (4). Let \\( \\mathcal{U}_s \\) be a standard dyadic upsampling operator with a factor \\( s \\) via expansion. Then" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.304, + 0.89, + 0.323 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathcal {U} _ {s} (\\boldsymbol {k})\\right) ^ {\\dagger} = \\omega_ {1} \\mathcal {U} _ {s} \\left(\\boldsymbol {v} _ {1}\\right) + \\dots + \\omega_ {N} \\mathcal {U} _ {s} \\left(\\boldsymbol {v} _ {N}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.331, + 0.815, + 0.346 + ], + "angle": 0, + "content": "Proof. See supplemental material for the proof." + }, + { + "type": "image", + "bbox": [ + 0.875, + 0.331, + 0.892, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.552 + ], + "angle": 0, + "content": "Proposition 1 implies that two PFSs with the same shape but different sizes can share the same dictionary for their inverse kernels up to a spatial scaling factor. For defocus PSFs within an image, they typically have similar shapes related to camera settings, as observed in [47]. However, their sizes will vary with the corresponding scene depths. Thus, we represent the inverse kernels of PSFs using different upsampled versions of a dictionary with small atoms. Such a multi-scale structure provides a physics-based implicit prior on the PSFs within the same image and leads to a more compact dictionary for representing inverse kernels of varying sizes. Specifically, we define a multi-scale dictionary composed by the sets (sub-dictionaries) \\(\\{V^r\\}_{r=1}^R\\):" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.559, + 0.89, + 0.576 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {V} ^ {r} = \\left[ \\boldsymbol {v} _ {1} ^ {r}, \\dots , \\boldsymbol {v} _ {N} ^ {r} \\right] \\subset \\mathbb {R} ^ {M _ {r} \\times M _ {r}}, \\boldsymbol {v} _ {n} ^ {r} = \\mathcal {U} _ {S _ {r}} (\\boldsymbol {v} _ {n}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.584, + 0.892, + 0.644 + ], + "angle": 0, + "content": "where \\( \\boldsymbol{v}_n^r \\) denotes the atom at the \\( r \\)-th scale, \\( S_r \\) denotes an upsampling factor that determines the size parameter \\( M_r \\) of the atom. As \\( r \\) increases, \\( S_r \\) becomes larger and \\( M_r \\) gets smaller. The per-pixel inverse kernels are then expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.651, + 0.89, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {k} _ {i, j} ^ {\\dagger} = \\omega_ {1, i, j} \\boldsymbol {v} _ {1} ^ {r _ {i, j}} + \\dots + \\omega_ {N, i, j} \\boldsymbol {v} _ {N} ^ {r _ {i, j}} \\in \\mathbb {R} ^ {M _ {r _ {i, j}} \\times M _ {r _ {i, j}}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.68, + 0.89, + 0.755 + ], + "angle": 0, + "content": "where \\(r_{i,j}\\) denotes the scale factor (index) related to \\(\\pmb{k}_{i,j}^{\\dagger}\\). Note that the atoms within the same sub-dictionary have a uniform size, while their sizes vary across different sub-dictionaries. That is, each sub-dictionary is utilized to represent inverse kernels of a specific size." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.756, + 0.89, + 0.802 + ], + "angle": 0, + "content": "Using the linear representation (7) and the fact that \\(\\pmb{v}_n^{r_{i,j}}*\\) \\(\\pmb {y} = \\sum_r\\delta (r - r_{i,j})\\pmb{v}_n^r *\\pmb {y}\\) with a Dirac delta function \\(\\delta\\) , we can rewrite Eq. (3) as" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.81, + 0.89, + 0.896 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {C} (\\boldsymbol {y}) = \\sum_ {i, j} \\mathbf {1} _ {i, j} \\odot \\left(\\sum_ {n = 1} ^ {N} \\omega_ {n, i, j} \\boldsymbol {v} _ {n} ^ {r _ {i, j}} * \\boldsymbol {y}\\right) (8) \\\\ = \\sum_ {i, j} \\mathbf {1} _ {i, j} \\odot \\left(\\sum_ {n = 1} ^ {N} \\sum_ {r = 1} ^ {R} \\mu_ {r, i, j} \\omega_ {n, i, j} \\boldsymbol {v} _ {n} ^ {r} * \\boldsymbol {y}\\right), (9) \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12602" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "where \\(\\mathbf{1}_{i,j}\\) denotes a mask with 1 for the \\((i,j)\\)-th entry being 1 and 0 otherwise, and \\(\\mu_{r,i,j} = \\delta (r - r_{i,j})\\). We can then perform spatially-varying inverse filtering by convolving each atom \\(\\pmb{v}_n^r\\) in the dictionary with the input image \\(\\pmb{y}\\), and summing the resulting images by the spatially-varying weights \\(\\mu_{s,i,j}\\omega_{n,i,j}\\). The weights \\(\\mu_{s,i,j}\\) and \\(\\omega_{n,i,j}\\) control the size (adaptive to scene depth) and shape (adaptive to image) of the corresponding inverse kernel, respectively." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.22, + 0.47, + 0.236 + ], + "angle": 0, + "content": "3.3. INR Models for Dictionaries of Inverse Kernels" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.242, + 0.47, + 0.424 + ], + "angle": 0, + "content": "The dictionary \\(\\{V_r\\}_{r=1}^R\\) used in (8) is generated by the base atoms \\(\\{v_j\\}_{j=1}^N\\) of the smallest size. However, simply upsampling these base atoms cannot cover all high frequencies, i.e., certain high-pass filters of large sizes may not be well approximated by the linear span of these atoms obtained by plain upsampling. To address this limitation, we use INR to re-parameterize the multi-scale dictionary. INR provides a more expressive way to generate multi-scale atoms that cover more frequencies than plain upsampling. Additionally, INR provides implicit regularization to alleviate overfitting by preferring image structures over random noise, as shown in [42, 19, 50]." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.425, + 0.299, + 0.441 + ], + "angle": 0, + "content": "Specifically, we express \\(\\pmb{v}_n^r\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.447, + 0.466, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {n} ^ {r} [ x, y ] = \\Phi_ {n} (x, y), [ x, y ] \\in [ 1, \\dots , M _ {r} ] \\times [ 1, \\dots , M _ {r} ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.469, + 0.577 + ], + "angle": 0, + "content": "where \\([x,y]\\) denotes a spatial coordinate, and \\(\\Phi\\) is an INR model implemented by a compact multi-layer perceptron (MLP) of a small size. The INR model maps spatial coordinates to a continuous function that implicitly represents the kernel atom. This function can be evaluated at any point in space to interpolate a kernel atom of an arbitrary size, forming a dictionary of inverse kernels with arbitrary scales." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.588, + 0.436, + 0.605 + ], + "angle": 0, + "content": "4. Implicit Neural Inverse Kernel Network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.469, + 0.734 + ], + "angle": 0, + "content": "Our proposed NN called INIKNet consists of recurrent deblurring stages over different scales for exploiting cross-scale contextual information. Each deblurring stage predicts and applies inverse kernels to deblur the input image. Concretely, each stage consists of two steps: (i) given an input image \\(\\pmb{y}\\), inferring two weight tensors \\(\\pmb{\\mu},\\pmb{\\omega}\\) that contain the weights \\(\\{\\mu_{r,i,j}\\}_{r,i,j}\\) and \\(\\{\\omega_{n,i,j}\\}_{n,i,j}\\) respectively; and (ii) applying inverse filtering defined by (8) to deblurring \\(\\pmb{y}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.394, + 0.759 + ], + "angle": 0, + "content": "4.1. Duplex Scale-Recurrent Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Current scale-recurrent deblurring NNs, e.g. [36, 28, 48, 12, 54, 8, 15], typically use an LSTM-based coarse-to-fine framework. The deblurred image from the current (coarser) scale is attached to the input of the next (finer) scale, resulting in unidirectional feature flow that only considers coarser scales when inferring at finer scales. This approach cannot fully leverage the additional information available at finer scales for deblurring on coarser scales, especially when the blur effects show strong similarity over space and scale." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.242 + ], + "angle": 0, + "content": "To address this limitation, we propose a duplex multiscale processing framework, which leverages bi-LSTM [13] as the sub-NNs to predict \\(\\mu\\) and \\(\\omega\\). It improves predictions at coarser scales by incorporating additional information from finer scales through the feature flow in the fine-to-coarse pass. This refined information is then utilized to further improve predictions on finer scales via the feature flow through the coarse-to-fine pass. As a result, the prediction at the current scale benefits from the information extracted from both the previous and subsequent scales." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.892, + 0.394 + ], + "angle": 0, + "content": "As shown in Fig. 1, our INIKNet first downsamples the input image to multi-scale versions. For each scale, a U-Net shared across scales is first applied for extracting spatial features from the input. The extracted features are then fed to two bi-LSTM sub-NNs so as to exploit the dependencies among scales for feature refinement. The outputs of two bi-LSTMs sub-NNs are then used to predict \\(\\mu\\) and \\(\\omega\\) via an \\(1\\times 1\\) convolutional layer, respectively. Afterward, the predicted weights are used by an Inverse Filtering Module (IFM) to perform deblurring using (8) to output the result." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.395, + 0.893, + 0.576 + ], + "angle": 0, + "content": "Our INIKNet differs from existing coarse-to-fine NNs in that it does not concatenate the output image from the current scale to the input of the next scale, as this operation would not only make the feature flow unidirectional, but also prevent the current-scale processing from starting, until the deblurring is fully completed at the previous scale. By removing this operation, the U-Net and bi-LSTM can start running earlier, without waiting for the deblurred images from previous scales. Moreover, by cancelling the operation, the IFM only needs to be called at the original image scale during inference, even though it is still required during training to calculate the multi-scale loss." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.587, + 0.715, + 0.604 + ], + "angle": 0, + "content": "4.2. Details of Key Modules" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Bi-LSTM sub-NNs Recall that the weights \\(\\mu\\) and \\(\\omega\\) correspond to two distinct properties of inverse kernels, namely size and shape, respectively. Their dependency on image scales also differs significantly. Therefore, rather than use a single bi-LSTM sub-NN to predict \\(\\mu\\) and \\(\\omega\\) jointly, we use two parallel bi-LSTM sub-NNs for separate predictions. Both sub-NNs share the same structure, consisting of convolutional LSTM cells [45] arranged in two paths for fine-to-coarse and coarse-to-fine estimations in opposite directions. Each LSTM cell takes a cell state and a hidden state as input and outputs updated states. The input cell state is defined as the features extracted by the U-Net, while the input hidden state for the next scale is set to the output one from the previous scale, forming a cross-scale connection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.893, + 0.9 + ], + "angle": 0, + "content": "Weight predictors To convert the output features from the bi-LSTM sub-NNs to the weights \\(\\mu\\) and \\(\\omega\\), two \\(1 \\times 1\\) convolutional layers are applied, respectively. For predicting \\(\\mu\\), the Softmax function is used to impose non-negativity and \\(\\ell_1\\)-normalization constraints. Such a soft relaxation of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12603" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.094, + 0.48, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.093, + 0.887, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.45, + 0.666, + 0.468 + ], + "angle": 0, + "content": "Figure 1: Architecture of proposed INIKNet for SIDD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.47, + 0.571 + ], + "angle": 0, + "content": "weights \\(\\mu_{r,i,j}\\) from 0,1 to [0,1] not only makes training easier, but also improves the representation accuracy of inverse kernels by including the use of additional atoms. This can be useful when the inverse kernel sizes fall between the gaps among the predefined non-consecutive atom sizes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.573, + 0.47, + 0.695 + ], + "angle": 0, + "content": "INR models of kernel atoms The MLPs used for INR have a structure similar to those in [46]. They comprise three fully-connected (FC) layers separated by a Sine activation function with a learnable frequency parameter that controls frequency content. The 2D coordinates are first mapped to a high-dimensional space and then projected to single values. One difference from [46] is that we first transform spatial coordinates to polar coordinates for slight improvement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.47, + 0.775 + ], + "angle": 0, + "content": "IFM for deblurring The IFM first calls the learned MLPs to generate all kernel atoms \\(\\{\\pmb{v}_n^r\\}_{n,r}\\) and then calculates the feature tensors \\(\\pmb{f}[n,r] = \\pmb{v}_n^r*\\pmb{y}\\). Afterward, Eq. (8) is implemented by simple entry-wise product and summation operations on the tensors \\(\\pmb{f},\\pmb{\\mu},\\pmb{\\omega}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.221, + 0.803 + ], + "angle": 0, + "content": "4.3. Training Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In addition to the \\(\\ell_2\\) loss \\(\\mathcal{L}_{\\mathrm{MSE}}\\) defined by mean squared error (MSE), two loss functions often seen in image recovery are also used for training, including the frequency-domain loss \\(\\mathcal{L}_{\\mathrm{FD}}\\) [6] defined as the \\(\\ell_1\\) distance in the frequency domain, and the Learned Perceptual Image Patch Similarity (LPIPS) loss \\(\\mathcal{L}_{\\mathrm{LPIPS}}\\) [57] predicted by a pre" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.494, + 0.892, + 0.541 + ], + "angle": 0, + "content": "trained NN. These three loss functions measure deblurring quality in the spatial, frequency and feature (perceptual) domains, respectively. The total loss is then given by" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.552, + 0.892, + 0.586 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {p} \\mathcal {L} _ {\\mathrm {M S E}} ^ {(p)} + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {F D}} ^ {(p)} + \\lambda_ {2} \\mathcal {L} _ {\\mathrm {L P I P S}} ^ {(p)}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.598, + 0.892, + 0.644 + ], + "angle": 0, + "content": "where \\( p \\) denotes the index of the \\( p \\)th scale, and \\( \\lambda_1, \\lambda_2 \\) are empirically set to 0.1. Ground-truth images are downsampled accordingly for the supervision at different scales." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.66, + 0.634, + 0.677 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.852 + ], + "angle": 0, + "content": "Datasets and metrics Following [2, 41], we train models on two widely-used datasets, DPDD [2] and LFDOF [41], respectively. To assess their generalization performance, we also evaluate these models on two more datasets, RTF [7] and RealDOF [21]. Three quality metrics are used for quantitative evaluation: Peak Signal to Noise Ratio (PSNR, measured in dB), Structural SIMilarity index (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [57]. The model complexity is measured by both the number of parameters and the inference time on a \\(640 \\times 640\\) image using an NVIDIA 2080Ti RTX GPU." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Implementation details Throughout the experiments, we employ 8 scales and 10 INR-based atoms per scale: \\( R = 8 \\) and \\( N = 10 \\). We use atoms of sizes \\( 1 \\times 1, 3 \\times 3, \\dots, 15 \\times 15 \\)," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "12604" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.213 + ], + "angle": 0, + "content": "where the \\(1 \\times 1\\) atom is fixed as a delta kernel for modeling in-focus regions. The frequency parameters of the sine activation functions in all INR models are initialized by random sampling from [2, 16]. We call Adam for training, with \\(4 \\times 10^{5}\\) iterations and a batch size of 4. We start with an initial learning rate of \\(2 \\times 10^{-4}\\) and halve it every \\(10^{5}\\) iterations. Random flipping, rotation, and cropping (to \\(256 \\times 256\\) pixels) are performed for training data augmentation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.222, + 0.303, + 0.238 + ], + "angle": 0, + "content": "5.1. Performance Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.47, + 0.442 + ], + "angle": 0, + "content": "Compared methods We select 9 NN-based SIDD methods for comparison, including the DPDNet [2], AIFNet [41], IFANet [21], KPAC [47], GKMNet [36], MDP [1], DRBNet [40], Restormer [53], and MPRNet [26] trained in [23] with a misalignment-robust scheme. The experimental results of the compared methods are reported either from their original papers or obtained by using their released pretrained models and codes. AIFNet is a model trained on LFDOF with the additional use of SYNDOF for its DME step. To compare its performance on DPDD, we retrain it using both DPDD and SYNDOF. DRBNet is a model trained on LFDOF and fine-tuned on DPDD. For a fair comparison, we retrain it on both datasets, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.471, + 0.718 + ], + "angle": 0, + "content": "Evaluation on DPDD and LFDOF Table 1 presents the quantitative results on DPDD and LFDOF. Our INIKNet performs the best in PSNR on DPDD and in all three metrics on LFDOF, despite using a lightweight model whose parameter number is the second smallest and around \\(13.2\\%\\) of that of Restormer, a general restoration model. Compared to the latest SIDD-specified MRPNet model [23], INIKNet achieves over \\(0.3\\mathrm{dB}\\) PSNR improvement on DPDD while having only about \\(10\\%\\) parameters to learn. Similar results are observed in the comparison with DRBNet, indicating that the performance gain of INIKNet comes from the design of NN structure rather than the increased model size. Though INIKNet is slightly larger than GKMNet, the smallest model, it achieves noticeably better performance. Compared to KPAC that uses dilated convolutions to mimic inverse kernels, INIKNet performs significantly better. In conclusion, the superiority and compactness of INIKNet demonstrate the effectiveness of its architecture design." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Generalization analysis on RealDOF and RTF The results of the DPDD/LFDOF-trained models evaluated on RealDOF and RTF are listed in Table 2. Our DPDD-trained INIKNet shows good generalization to different datasets and achieves the best performance on both RealDOF and RTF in all three metrics. The LFDOF-trained INIKNet is the top performer on RealDOF and achieves the best SSIM score on RTF, as well as the second-best scores in PSNR and LPIPS. Though AIFNet performs better in PSNR and LPIPS, it requires an additional dataset for DME and is much larger than INIKNet. In summary, INIKNet exhibits superior generalizability over other compared models." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.09, + 0.892, + 0.248 + ], + "angle": 0, + "content": "
ModelDPDDLFDOF#Params (106)Time (sec.)
PSNRSSIMLPIPSPSNRSSIMLPIPS
DPDNet24.3480.7470.277---32.251.430
AIFNet24.2130.7420.30929.6770.8840.20241.550.276
IFANet25.3660.7890.21729.7870.8720.15610.480.063
KPAC25.2210.7740.22628.9420.8570.1742.060.037
GKMNet25.4680.7890.21929.0810.8670.1711.410.129
MDP25.3470.7630.26828.0690.8340.18546.863.317
DRBNet25.4850.7920.25430.2530.8830.14711.690.085
MPRNet25.7300.7920.232---20.101.161
Restormer25.9800.8110.178---26.100.672
INIKNet26.0550.8030.18530.2930.8860.1321.980.228
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.253, + 0.892, + 0.284 + ], + "angle": 0, + "content": "Table 1: Quantitative results on DPDD/LFDOF. Bold for best performers and underline for second-best performers." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.3, + 0.887, + 0.553 + ], + "angle": 0, + "content": "
ModelRealDOFRTF#Params (106)
PSNRSSIMLPIPSPSNRSSIMLPIPS
Trained on DPDDDPDNet22.8700.6700.42523.6080.5910.29632.25
AIFNet23.0930.6800.41324.0410.7580.28941.55
MDP23.5000.6810.44424.0120.7380.31246.86
KPAC23.9750.7620.33824.6180.7770.2362.06
IFANet24.7120.7480.30624.9240.8010.22710.48
GKMNet24.2540.7320.39224.9700.7890.2611.41
DRBNet24.8840.7510.37624.4630.7730.31111.69
MPRNet24.5410.7360.33924.5880.7880.30420.10
Restormer25.0910.7620.28524.2120.8210.22426.10
INIKNet25.2310.7650.28725.4500.8340.2151.98
Trained on LFDOFIFANet22.5040.6690.48326.4370.8380.23810.48
KPAC22.5500.6710.45725.9590.8030.2302.06
AIFNet22.6230.6670.46127.5520.8820.17641.55
GKMNet23.6090.7210.40826.9850.8560.2461.41
MDP22.7260.6800.45325.5800.8090.22846.86
DRBNet22.9100.6910.43726.7170.8530.20011.69
INIKNet23.8100.7240.35627.4010.8850.1791.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.557, + 0.892, + 0.588 + ], + "angle": 0, + "content": "Table 2: Quantitative results on RealDOF/RTF. Bold for best performers and underline for second-best performers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.892, + 0.742 + ], + "angle": 0, + "content": "Qualitative comparison We compare INIKNet with some SIDD-specified competitors on DPDD and LFDOF via visual inspection. See Fig. 2 for some deblurred images on four datasets, where complex textures such as those on wire fence and branches are successfully recovered by INIKNet. See also Fig. 3 for some results on the CUHK dataset [43]. While most methods fail under large blur, ours yields more visually satisfying results. These findings demonstrate that our INIKNet performs well in terms of visual quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.753, + 0.767, + 0.769 + ], + "angle": 0, + "content": "5.2. Ablation Studies and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.777, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Ablation studies We construct several variants of INIKNet and list their results on two datasets in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Non-MS Dict: We drop the multi-scale structure of the dictionary and instead use an individual INR model for each dictionary atom at each scale (thus yielding a larger model size). A noticeable performance degradation occurs. This is because the multi-scale structure of the INR-based dictionary imposes structural regularization to reduce overfitting." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12605" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.087, + 0.531, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.087, + 0.882, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.505, + 0.858, + 0.521 + ], + "angle": 0, + "content": "Figure 2: Results on using models trained on DPDD (left part) and LFDOF (right part). Zoom-in for better views." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.544, + 0.462, + 0.633 + ], + "angle": 0, + "content": "
ModelRealDOFRTF
PSNRSSIMLPIPSPSNRSSIMLPIPS
Non-MS Dict24.9160.7490.31025.1360.8090.254
Non-INR Dict25.0580.7540.30325.2370.8190.226
Unidirectional25.0990.7530.29425.3050.8240.227
Original INIKNet25.2310.7650.28725.4500.8340.215
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.637, + 0.454, + 0.651 + ], + "angle": 0, + "content": "Table 3: Results of baseline models trained on DPDD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.47, + 0.775 + ], + "angle": 0, + "content": "Non-INR Dict: When the INR-based dictionaries are replaced by the plain ones, where each base atom \\( v_{n} \\) is defined as a learnable matrix and bi-linearly upsampled to form the multi-scale dictionary, the performance of INIKNet decreases. This is probably because the frequency content for the atoms generated by plain upsampling are limited, so do their generated inverse kernels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Unidirectional: To simulate an unidirectional scale recurrent mechanism (e.g. [36]), we replace the bi-LSTM with double unidirectional LSTMs (for keeping the model size). The output from the last scale is attached to the input of the current scale for additional information. Such a scheme for coefficient prediction results in some performance degradation, highlighting the effectiveness of our proposed duplex scale-recurrent framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.892, + 0.819 + ], + "angle": 0, + "content": "Visualization of coefficients and atoms The coefficient maps \\(\\mu\\) and \\(\\omega\\) are visualized in Fig. 4. We can see that the regions with larger blur amount tend to have larger coefficients \\(\\mu\\) assigned to larger inverse kernels, and vice versa. In comparison, the coefficients \\(\\omega\\) are more spatially similar (gray), consistent with the prior of spatially-similar shapes of inverse kernels. We use Fig. 5 to visualize some learned (non-)INR-based atoms in both the spatial and Fourier domain. The INR-based atoms within the same scale contain diverse patterns with varying frequencies and orientations, and meanwhile they have similar shapes across scales. Compared to the non-INR-based ones, the INR-based atoms encode richer spatial structures and frequency contents. Interestingly, the INR-based atoms exhibit strong isotropy, which is likely due to the isotropy of defocus PSFs (so do their inverse kernels). All these properties of the INR-based atoms enable our INIKNet to effectively represent and predict inverse kernels for spatially-varying defocus blur." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.62, + 0.846 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "This paper presented an interpretable end-to-end SIDD approach, which learns a duplex scale-recurrent NN to predict INR-driven dictionary-parameterized inverse kernels of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12606" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.5, + 0.867, + 0.516 + ], + "angle": 0, + "content": "Figure 3: Results on CUHK dataset (without ground-truths) using models trained on DPD. Zoom-in for better views." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.53, + 0.889, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.674, + 0.642, + 0.689 + ], + "angle": 0, + "content": "Figure 4: Visualization of \\(\\pmb{\\mu}\\) maps and \\(\\omega\\) maps." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.703, + 0.885, + 0.807 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.813, + 0.846, + 0.829 + ], + "angle": 0, + "content": "Figure 5: Visualization of bilinearly upsampled (non-INR-based) atoms and INR-based atoms in two domains." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.855, + 0.471, + 0.901 + ], + "angle": 0, + "content": "defocus blur and applies them to deblur images. Our proposed approach introduced effective structural regularization on inverse kernels for good generalization, which is" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.855, + 0.892, + 0.901 + ], + "angle": 0, + "content": "based on the shape similarity of defocus PSFs over an image. The resulting effective and light-weight model has demonstrated its advantages by extensive experiments." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12607" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Abdullah Abuolaim, Mahmoud Afifi, and Michael S. Brown. Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1231-1239, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Abdullah Abuolaim and Michael S. Brown. Defocus deblurring using dual-pixel data. In Proceedings of the European Conference on Computer Vision, pages 111-126. Springer, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Yosuke Bando and Tomoyuki Nishita. Towards digital refocusing from a single photograph. In Proceedings of the Pacific Conference on Computer Graphics and Applications, pages 363-372. IEEE, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.354 + ], + "angle": 0, + "content": "[4] Mingqin Chen, Yuhui Quan, Tongyao Pang, and Hui Ji. Non-blind image deconvolution via leveraging model uncertainty in an untrained deep neural network. International Journal of Computer Vision, 130(7):1770-1789, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.471, + 0.409 + ], + "angle": 0, + "content": "[5] Sunghyun Cho and Seungyong Lee. Convergence analysis of map based blur kernel estimation. In Proceedings of the IEEE/International Conference on Computer Vision, pages 4808-4816, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.471, + 0.478 + ], + "angle": 0, + "content": "[6] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4641-4650, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[7] Laurent D'Andrés, Jordi Salvador, Axel Kochale, and Sabine Süssstrunk. Non-parametric blur map regression for depth of field extension. IEEE Transactions on Image Processing, 25(4):1660-1673, 2016. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.471, + 0.606 + ], + "angle": 0, + "content": "[8] Senyou Deng, Wenqi Ren, Yanyang Yan, Tao Wang, Fenglong Song, and Xiaochun Cao. Multi-scale separable network for ultra-high-definition video deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14030-14039, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.471, + 0.661 + ], + "angle": 0, + "content": "[9] Jiangxin Dong, Stefan Roth, and Bernt Schiele. Deep wiener deconvolution: Wiener meets deep learning for image deblurring. Advances in Neural Information Processing Systems, 33:1048-1059, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.663, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[10] Jiangxin Dong, Stefan Roth, and Bernt Schiele. Dwdn: deep wiener deconvolution network for non-blind image deblurring. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9960-9976, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.72, + 0.471, + 0.801 + ], + "angle": 0, + "content": "[11] Zhenxuan Fang, Fangfang Wu, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Self-supervised non-uniform kernel estimation with flow-based motion prior for blind image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18105-18114, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.803, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[12] Hongyun Gao, Xin Tao, Xiaoyong Shen, and Jiaya Jia. Dynamic scene deblurring with parameter selective sharing and nested skip connections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3848-3856, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[13] Alex Graves and Jürgen Schmidhuber. Framewise phoneme classification with bidirectional LSTM and other neural net-" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "work architectures. Neural networks, 18(5-6):602-610, 2005. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.894, + 0.189 + ], + "angle": 0, + "content": "[14] Fang Hua, Peter Johnson, Nadezhda Sazonova, Paulo Lopez-Meyer, and Stephanie Schuckers. Impact of out-of-focus blur on face recognition performance based on modular transfer function. In Proceedings of the International Conference on Biometrics, pages 85–90. IEEE, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.894, + 0.245 + ], + "angle": 0, + "content": "[15] Bo Ji and Angela Yao. Multi-scale memory-based video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1919-1928, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.894, + 0.287 + ], + "angle": 0, + "content": "[16] Ali Karaali, Naomi Harte, and Claudio R. Jung. Deep multiscale feature learning for defocus blur estimation. IEEE Transactions on Image Processing, 31:1097-1106, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.894, + 0.342 + ], + "angle": 0, + "content": "[17] Ali Karaali and Claudio Rosito Jung. Edge-based defocus blur estimation with adaptive scale selection. IEEE Transactions on Image Processing, 27(3):1126-1137, Mar. 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.343, + 0.894, + 0.399 + ], + "angle": 0, + "content": "[18] Ali Karaali and Cláudio Rosito Jung. Svbr-net: A non-blind spatially varying defocus blur removal network. In Proceedings of the IEEE International Conference on Image Processing, pages 566-570. IEEE, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.894, + 0.441 + ], + "angle": 0, + "content": "[19] Chaewon Kim, Jaeho Lee, and Jinwoo Shin. Zero-shot blind image denoising via implicit neural representations. arXiv preprint arXiv:2204.02405, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.894, + 0.483 + ], + "angle": 0, + "content": "[20] Dilip Krishnan and Rob Fergus. Fast image deconvolution using hyper-laplacian priors. In Advances in Neural Information Processing Systems, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.894, + 0.552 + ], + "angle": 0, + "content": "[21] Junyong Lee, Hyeongseok Son, Jaesung Rim, Sunghyun Cho, and Seungyong Lee. Iterative filter adaptive network for single image defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2034-2042, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.894, + 0.622 + ], + "angle": 0, + "content": "[22] Dasong Li, Yi Zhang, Ka Chun Cheung, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. Learning degradation representations for image deblurring. In Proceedings of the European Conference on Computer Vision, pages 736-753. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.894, + 0.678 + ], + "angle": 0, + "content": "[23] Yu Li, Dongwei Ren, Xinya Shu, and Wangmeng Zuo. Learning single image defocus deblurring with misaligned training pairs. Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.894, + 0.748 + ], + "angle": 0, + "content": "[24] Ziwei Luo, Haibin Huang, Lei Yu, Youwei Li, Haoqiang Fan, and Shuaicheng Liu. Deep constrained least squares for blind image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17642-17652, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.894, + 0.803 + ], + "angle": 0, + "content": "[25] Haoyu Ma, Shaojun Liu, Qingmin Liao, Juncheng Zhang, and Jing-Hao Xue. Defocus image deblurring network with defocus map estimation as auxiliary task. IEEE Transactions on Image Processing, 31:216-226, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.894, + 0.872 + ], + "angle": 0, + "content": "[26] Armin Mehri, Parichehr B Ardakani, and Angel D Sappa. Mprnet: Multi-path residual network for lightweight image super resolution. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2704-2713, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[27] Sung-Jun Min, Kyeongbo Kong, and Suk-Ju Kang. Out-of-focus image deblurring for mobile display vision inspection." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12608" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.123, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[28] Seungjun Nah, Tae Kim, and Kyoung Mu Lee. Deep multiscale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 257-265, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.195, + 0.471, + 0.25 + ], + "angle": 0, + "content": "[29] Yuesong Nan and Hui Ji. Deep learning for handling kernel/model uncertainty in image deconvolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2388-2397, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.253, + 0.47, + 0.308 + ], + "angle": 0, + "content": "[30] Yuesong Nan, Yuhui Quan, and Hui Ji. Variational-EM-based deep learning for noise-blind image deblurring. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3626–3635, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.311, + 0.47, + 0.367 + ], + "angle": 0, + "content": "[31] Saqib Nazir, Lorenzo Vaquero, Manuel Mucientes, Víctor M Brea, and Daniela Coltuc. Depth estimation and image restoration by deep learning from defocused images. arXiv preprint arXiv:2302.10730, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.369, + 0.469, + 0.438 + ], + "angle": 0, + "content": "[32] Jinsun Park, Yu-Wing Tai, Donghyeon Cho, and In So Kweon. A unified approach of multi-scale deep and handcrafted features for defocus estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1736-1745, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[33] Valeriya Pronina, Filippos Kokkinos, Dmitry V Dylov, and Stamatios Lefkimmiatis. Microscopy image restoration with deep wiener-kolmogorov filters. In Proceedings of the European Conference on Computer Vision, pages 185-201. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.513, + 0.469, + 0.569 + ], + "angle": 0, + "content": "[34] Yuhui Quan, Zhuojie Chen, Huan Zheng, and Hui Ji. Learning deep non-blind image deconvolution without ground truths. In Proceedings of the European Conference on Computer Vision, pages 642-659, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.572, + 0.469, + 0.626 + ], + "angle": 0, + "content": "[35] Yuhui Quan, Peikang Lin, Yong Xu, Yuesong Nan, and Hui Ji. Nonblind image deblurring via deep learning in complex field. IEEE Transactions on Neural Networks and Learning Systems, 33(10):5387-5400, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.63, + 0.469, + 0.684 + ], + "angle": 0, + "content": "[36] Yuhui Quan, Hui Wu, Zicong, and Hui Ji. Gaussian kernel mixture network for single image defocus deblurring. In Advances in Neural Information Processing Systems, volume 34, 2021. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[37] Yuhui Quan, Hui Wu, Zicong, and Hui Ji. Neumann network with recursive kernels for single image defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5754-5763, June 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[38] Wenqi Ren, Jiawei Zhang, Lin Ma, Jinshan Pan, Xiaochun Cao, Wangmeng Zuo, Wei Liu, and Ming-Hsuan Yang. Deep non-blind deconvolution via generalized low-rank approximation. Advances in Neural Information Processing Systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[39] Wenqi Ren, Jiawei Zhang, Jinshan Pan, Sifei Liu, Jimmy S Ren, Junping Du, Xiaochun Cao, and Ming-Hsuan Yang. Deblurring dynamic scenes via spatially varying recurrent neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):3974-3987, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[40] Lingyan Ruan, Bin Chen, Jizhou Li, and Miuling Lam. Learning to deblur using light field generated and real defocus images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16304-16313, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[41] Lingyan Ruan, Bin Chen, Jizhou Li, and Miu-Ling Lam. Aifnet: All-in-focus image restoration network using a light field-based dataset. IEEE Transactions on Computational Imaging, 7:675-688, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[42] Nimrod Shabtay, Eli Schwartz, and Raja Giryes. Pip: Positional-encoding image prior. arXiv preprint arXiv:2211.14298, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[43] Jianping Shi, Li Xu, and Jiaya Jia. Discriminative blur detection features. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2965-2972, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[44] Jianping Shi, Li Xu, and Jiaya Jia. Just noticeable defocus blur detection and estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 657-665, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[45] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in Neural Information Processing Systems, 28, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[46] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[47] Hyeongseok Son, Junyong Lee, Sunghyun Cho, and Seungyong Lee. Single image defocus deblurring using kernel-sharing parallel atrous convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2642-2650, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[48] Xin Tao, Hongyun Gao, Yi Wang, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8174-8182, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[49] Longguang Wang, Yingqian Wang, Xiaoyu Dong, Qingyu Xu, Jungang Yang, Wei An, and Yulan Guo. Unsupervised degradation representation learning for blind superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10581-10590, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[50] Shaowen Xie, Hao Zhu, Zhen Liu, Qi Zhang, You Zhou, Xun Cao, and Zhan Ma. Diner: Disorder-invariant implicit neural representation. arXiv preprint arXiv:2211.07871, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[51] Guodong Xu, Yuhui Quan, and Hui Ji. Estimating defocus blur via rank of local patches. In Proceedings of the IEEE/International Conference on Computer Vision, pages 5371-5379, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[52] Li Xu, Jimmy S Ren, Ce Liu, and Jiaya Jia. Deep convolutional neural network for image deconvolution. Advances in Neural Information Processing Systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[53] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12609" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.146 + ], + "angle": 0, + "content": "Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[54] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5978-5986, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.219, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[55] Jiawei Zhang, Jinshan Pan, Jimmy Ren, Yibing Song, Linchao Bao, Rynson WH Lau, and Ming-Hsuan Yang. Dynamic scene deblurring using spatially variant recurrent neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2521-2529, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.304, + 0.469, + 0.358 + ], + "angle": 0, + "content": "[56] Jie Zhang and Wanming Zhai. Blind attention geometric restraint neural network for single image dynamic/defocus deblurring. IEEE Transactions on Neural Networks and Learning Systems, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.361, + 0.469, + 0.429 + ], + "angle": 0, + "content": "[57] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.431, + 0.469, + 0.485 + ], + "angle": 0, + "content": "[58] Wenda Zhao, Fei Wei, You He, and Hutchuan Lu. United defocus blur detection and deblurring via adversarial promoting learning. In Proceedings of the European Conference on Computer Vision, pages 569-586, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.488, + 0.469, + 0.542 + ], + "angle": 0, + "content": "[59] Wenda Zhao, Fan Zhao, Dong Wang, and Huchuan Lu. Defocus blur detection via multi-stream bottom-top-bottom network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42:1884-1897, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.544, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[60] Bolun Zheng, Quan Chen, Shanxin Yuan, Xiaofei Zhou, Hua Zhang, Jiyong Zhang, Chenggang Yan, and Gregory Slabaugh. Constrained predictive filters for single image bokeh rendering. IEEE Transactions on Computational Imaging, 8:346-357, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.615, + 0.469, + 0.684 + ], + "angle": 0, + "content": "[61] Chao Zhu, Hang Dong, Jinshan Pan, Boyang Liang, Yuhao Huang, Lean Fu, and Fei Wang. Deep recurrent neural network with multi-scale bi-directional propagation for video deblurring. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 3598-3607, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12610" + } + ] +] \ No newline at end of file diff --git a/2023/Single Image Defocus Deblurring via Implicit Neural Inverse Kernels/bee3d863-7fc6-4e0b-a339-87a4afbe8bb4_origin.pdf b/2023/Single Image Defocus Deblurring via Implicit Neural Inverse Kernels/bee3d863-7fc6-4e0b-a339-87a4afbe8bb4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..05dafcfa696f54612bf40a17b74e96798b8118c7 --- /dev/null +++ b/2023/Single Image Defocus Deblurring via Implicit Neural Inverse Kernels/bee3d863-7fc6-4e0b-a339-87a4afbe8bb4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fb576ed3d95804c45c0f2f34a43ac496a8fd9e5d08bf118a5ae5b398a333847 +size 9056976 diff --git a/2023/Single Image Defocus Deblurring via Implicit Neural Inverse Kernels/full.md b/2023/Single Image Defocus Deblurring via Implicit Neural Inverse Kernels/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8d11a5267ff4240722953c21e180e2767705afe5 --- /dev/null +++ b/2023/Single Image Defocus Deblurring via Implicit Neural Inverse Kernels/full.md @@ -0,0 +1,319 @@ +# Single Image Defocus Deblurring via Implicit Neural Inverse Kernels + +Yuhui Quan $^{1,2}$ Xin Yao $^{1,2}$ Hui Ji $^{3}$ + +$^{1}$ School of Computer Science and Engineering, South China University of Technology, Guangzhou 510006, China $^{2}$ Pazhou Lab, Guangzhou 510335, China + +3 Department of Mathematics, National University of Singapore, 119076, Singapore + +cseyhquan@scut.edu.cn, csxinyaoscut@gmail.com, matjh@nus.edu.sg* + +# Abstract + +Single image defocus deblurring (SIDD) is a challenging task due to the spatially-varying nature of defocus blur, characterized by per-pixel point spread functions (PSFs). Existing deep-learning-based methods for SIDD are limited by either over-fitting due to the lack of model constraints or under-parametrization that restricts their applicability to real-world images. To address the limitations, this paper proposes an interpretable approach that explicitly predicts inverse kernels with structural regularization. Motivated by the observation that defocus PSFs within an image often have similar shapes but different sizes, we represent the inverse kernels linearly over a multi-scale dictionary parameterized by implicit neural representations. We predict the corresponding representation coefficients via a duplex scale-recurrent neural network that jointly performs fine-to-coarse and coarse-to-fine estimations. Extensive experiments demonstrate that our approach achieves excellent performance using a lightweight model. + +# 1. Introduction + +When a camera captures an image, objects outside of the focal plane appear blurry. This effect is known as defocus blur and usually occurs when using a large camera aperture or capturing scenes with significantly varying depths. Removing defocus blur from a single image, known as SIDD, is desired for many practical applications; see e.g. [27, 14, 33, 60]. In general, defocus blur varies spatially. Each defocused pixel is a weighted average of its neighboring pixels in the latent all-in-focus image, with the weights determined by a spatially-varying unknown PSF. + +SIDD is closely related to defocus map estimation (DME) [3, 44, 7, 5, 51, 32, 17]. Existing DME methods typically model defocus PSFs using a Gaussian function + +parameterized by its variance or a disc function parameterized by its radius. The resulting defocus map indicates the per-pixel blur amount in terms of the values of such model parameters, which can then be utilized for SIDD by applying a non-blind image deblurring (NID) algorithm using the corresponding PSFs; see e.g. [18]. However, DME remains a challenging task, and the estimated parameters of PSFs can be erroneous for many pixels. Moreover, Gaussian or disc-indicator functions are overly simplistic for real-world defocus PSFs. All these errors will be magnified in the NID process, resulting in noticeable artifacts. + +Recently, many deep-learning-based methods have been proposed for SIDD; see e.g. [41, 25, 21, 47, 36, 40, 1, 56, 58, 23, 31]. Most train an end-to-end neural network (NN) that directly maps blurry images to their sharp correspondences. Although these methods provide better results than the DME-based two-step methods, their performance still has much room for improvement due to significant variations in the spatial distributions of defocus PSFs among different images. Many existing works (e.g. [21, 40, 56]) explicitly introduce spatially-varying processing blocks to better handle the spatial variance of defocus PSFs. However, those blocks lack constraints and may cause overfitting. A few studies (e.g. [36, 37]) represent defocus PSFs by some specific basis for geometric regularization and employ deep unrolling for an NN-based inversion process. Nevertheless, there is still room for improvement. + +This paper presents an end-to-end deep NN with high interpretability for SIDD, which explicitly predicts spatially-varying inverse kernels of defocus PSFs and performs deblurring using the predicted inverse kernels. The possible overfitting issue is tackled by providing a more accurate model with structural constraints/regularization for the inverse kernels. Specifically, the inverse kernels are modeled by a linear representation under a dictionary. To achieve a compact yet sufficiently general representation, we construct the dictionary with a multi-scale structure. This is motivated by the empirical observation (as seen in [47]) that defocus PSFs within an image tend to have the same shape + +but vary mainly in size. We show that if two defocus PSFs of different sizes have the same shape under an upsampling operation, so do the dictionary atoms of their inverse kernels. Thus, the dictionary atoms of a large size are defined as an upsampled version of the atoms of a small size. + +Although plain upsampling offers an economical method for generating larger dictionary atoms from smaller ones, the resulting atoms may have a limited frequency range. This can lead to poor approximations of certain inverse kernels. To address the limitation, we leverage implicit neural representation (INR) [46], a technique that uses coordinate-input NNs to represent geometric objects like 2D shapes, so as to efficiently parameterize multi-scale atoms with both sufficient coverage of high-frequency content and implicit regularization to alleviate overfitting. + +For the multi-scale representation coefficients, we separate them into scale-related and shape-related components, which are predicted by two scale-recurrent sub-NNs respectively. To better utilize information from different scales for coefficient prediction, we introduce a duplex scale-recurrent framework that performs both fine-to-coarse and coarse-to-fine estimations. Once the coefficients are predicted, the corresponding inverse kernels are applied to the input image for deblurring. The resulting pipeline is highly efficient and interpretable, leading to a lightweight yet effective model. + +To conclude, this paper proposes an end-to-end NN for SIDD, which exhibits a noticeable performance boost compared to existing ones, while maintaining low complexity. See below for our main contributions: + +- A parametric inverse kernel prediction framework for SIDD, providing efficient and interpretable deep NNs. +- A multi-scale linear representation model for the inverse kernels of spatially-varying defocus blur, using an INR-based multi-scale dictionary. +- A duplex scale-recurrent framework for predicting the representation coefficients of inverse kernels. + +# 2. Related Works + +# 2.1. Existing Approaches for SIDD + +SIDD can done by a two-step approach that combines DME (e.g. [44, 51, 17, 32]) and NID (e.g. [20]). However, such an approach has two inherent drawbacks. First, the errors in DME will be magnified in the following NID process. Second, simple Gaussian or disc functions do not fit real-world defocus PSFs well. While there has been steady progress in deep-learning-based methods for DME [59, 16] and for NID [38, 10, 33, 29, 30, 4, 34, 35], the inherent drawbacks of the two-step approach remain. + +Deep end-to-end learning has emerged as a promising approach for SIDD. The seminal work conducted by Abuolaim and Brown [2] trained an encoder-decoder CNN that + +maps a defocused image to its in-focus correspondence, with significant performance gain over two-step methods. To better handle spatially-varying defocus blur, some studies introduced dynamic processing blocks, e.g., per-pixel filter prediction in Lee et al. [21], dynamic residual blocks in Ruan et al. [40], and attention in Zhang and Zhai [56]. + +The use of dynamic processing blocks may increase the model complexity and the risk of overfitting. Regularization by auxiliary tasks is an effective approach for improving the generalization performance. In existing works, auxiliary tasks are often defined using DME ([25, 41, 58]), depth estimation [31], dual-view data [21, 1], or reblurring loss [23]. However, most of them require additional data sources. + +A more appealing approach is encoding physical priors of defocus blur into NNs. Quan et al. [36, 37] used a kernel mixture model to parameterize defocus PSFs and constructed a deep unrolling NN. This method can be viewed as the degradation learning [22, 49, 24, 11] that predicts the blurring operator or kernels for deblurring. Our approach differs in that it predicts inverse kernels, not the degradation with defocus PSFs. Son et al. [47] showed that inverse defocus kernels have an invariant shape across scales and emulated them in feature spaces via a series of kernel-sharing parallel Atrous convolutions. Their method inspired our work, and it can be viewed as representing inverse kernels for a feature tensor via a multi-scale dilated combination of a single atom and predicting spatially-invariant combination coefficients via channel attention. In comparison, our approach explicitly predicts inverse kernels on the input image, allowing better interpretability, and represents them using INR to improve effectiveness. Additionally, we predict spatially-varying combination coefficients to better handle spatially-similar (not invariant) defocus shapes. + +# 2.2. Related Techniques + +Deep inverse kernels Several studies have included inverse kernels in deblurring NNs, in addition to [47]. Xu et al. [52] and Ren et al. [38] presented the CNNs that adopt tensor product for emulating separable inverse kernels to perform NID. Dong et al. [9, 10] and Pronina et al. [33] performed NID via Wiener inverse filtering in deep feature spaces. However, all these methods are designed for spatially-invariant blur. Ren et al. [55, 39] proposed a recurrent NN to mimic inverse kernels for removing spatially-varying blur of dynamic scenes, but not for SIDD. + +Scale-recurrent NNs for deblurring Scale-recurrent NNs are commonly used for image deblurring, e.g. [36, 28, 48, 12, 54, 8, 15]. These methods typically employ coarse-to-fine estimation which is unidirectional and thus cannot fully leverage information from the subsequent scales for the prediction at the current scale. In contrast, our proposed NN utilizes a bi-directional structure via bi-LSTMs [13] for joint coarse-to-fine and fine-to-coarse estimation. This is + +particularly useful for SIDD, as defocus blur effects show strong self-similarity over scales. While bi-LSTM has been used in [61] for multi-scale video deblurring, it is only used for bi-directional temporal processing and the multi-scale processing is still one-way. Also note that the multi-scale NN proposed in [6] for motion deblurring is not restrictive to unidirectional processing, but not in a recurrent manner. + +INR for image recovery INR has been applied to image recovery [42, 19, 50]. In comparison to these works, we use INR to parameterize inverse kernels, not images, for SIDD. + +# 3. INR-Based Dictionary for Inverse Kernels + +# 3.1. Inverse Kernels for Spatially-Varying Blur + +Consider a simplistic uniform blurring model defined by $\pmb{y} = \pmb{k} * \pmb{x}$ , where $*$ denotes convolution, $\pmb{k}$ denotes the PSF, and $\pmb{y}, \pmb{x}$ denote the blurred image and its sharp correspondence, respectively. Let $\mathcal{F}$ denote the discrete Fourier transform (DFT). According to the well-known convolution theorem, we have $\mathcal{F}(\pmb{y}) = \mathcal{F}(\pmb{x}) \odot \mathcal{F}(\pmb{k})$ , where $\odot$ denotes entry-wise product. Then, one can define a (pesudo) inverse kernel $\pmb{k}^{\dagger}$ by: for each $(\omega_x, \omega_y)$ , + +$$ +\mathcal {F} \left(\boldsymbol {k} ^ {\dagger}\right) \left[ \omega_ {x}, \omega_ {y} \right] = \left(\mathcal {F} (\boldsymbol {k}) [ s ]\right) ^ {- 1}, \text {i f} | \mathcal {F} (\boldsymbol {k}) \left[ \omega_ {x}, \omega_ {y} \right] | \neq 0; \tag {1} +$$ + +and 0 otherwise. Then, for an invertible PSF with only nonzero DFT coefficients, we have + +$$ +\mathcal {F} (\boldsymbol {x}) = \mathcal {F} \left(\boldsymbol {k} ^ {\dagger}\right) \odot \mathcal {F} (\boldsymbol {y}) \Longrightarrow \boldsymbol {x} = \boldsymbol {k} ^ {\dagger} * \boldsymbol {y}. \tag {2} +$$ + +When the PSF $k$ is non-invertible, the result $k^{\dagger} * y$ still provides a fair approximation to $x$ . + +One can extend (1) to handling spatially-varying blur. Let $\mathbb{K}^{\dagger} = \{k_{i,j}^{\dagger}\}_{i,j}$ denote a set of inverse kernels. Then, the image $\pmb{y}$ with spatially-varying blur can be recovered by a spatially-varying deconvolution process $\mathcal{C}$ defined by + +$$ +\mathcal {C} \left(\boldsymbol {y}; \mathbb {K} ^ {\dagger}\right) [ i, j ] = \sum_ {x, y} \boldsymbol {k} _ {i, j} ^ {\dagger} [ x, y ] \boldsymbol {y} [ i - x, j - y ]. \tag {3} +$$ + +Both the shape and the size of $k_{i,j}^{\dagger}$ depend on the PSFs determined by scene depths and camera settings. When scene depths have significant variations, the sizes of the defocus PSFs also vary significantly, leading to large variations in the size of $k_{i,j}^{\dagger}$ and a large number of unknowns in $\mathbb{K}^{\dagger}$ . Without additional constraints on $\mathbb{K}^{\dagger}$ , overfitting can occur. + +# 3.2. Multi-Scale Dictionary for Inverse Kernels + +To address the overfitting caused by excessive degrees of freedom in $\mathbb{K}^{\dagger}$ , we parameterize the inverse kernels using a linear span over a dictionary with reduced number of atoms, resulting in inverse kernels that lie within a low-dimensional space. That is, we express an inverse kernel $\pmb{k}^{\dagger}$ as a linear combination of atoms from a dictionary $\pmb{V} = [v_{1},\dots ,v_{N}]$ : + +$$ +\boldsymbol {k} ^ {\dagger} = \omega_ {1} \boldsymbol {v} _ {1} + \dots + \omega_ {N} \boldsymbol {v} _ {N}, \tag {4} +$$ + +where $\omega_{1},\dots ,\omega_{N}\in \mathbb{R}$ are the weights. Additionally, inverse kernels may have a large size. To accurately approximate them using as few atoms as possible, we introduce a multi-scale structure to the dictionary atoms, which is motivated by the observation that when two kernels have the same shape but different sizes, the larger one can be effectively approximated by upsampling the smaller one. Consequently, the dictionary atoms for the corresponding inverse kernels can also be related through such an upsampling process, as stated in the following proposition. + +Proposition 1. Let $\pmb{k}^{\dagger}$ be an inverse kernel defined by (4). Let $\mathcal{U}_s$ be a standard dyadic upsampling operator with a factor $s$ via expansion. Then + +$$ +\left(\mathcal {U} _ {s} (\boldsymbol {k})\right) ^ {\dagger} = \omega_ {1} \mathcal {U} _ {s} \left(\boldsymbol {v} _ {1}\right) + \dots + \omega_ {N} \mathcal {U} _ {s} \left(\boldsymbol {v} _ {N}\right). \tag {5} +$$ + +Proof. See supplemental material for the proof. + +![](images/84ccd284f25d6e8ba9319f23dd4670275be72eb35bc144e09f67ee8693263c78.jpg) + +Proposition 1 implies that two PFSs with the same shape but different sizes can share the same dictionary for their inverse kernels up to a spatial scaling factor. For defocus PSFs within an image, they typically have similar shapes related to camera settings, as observed in [47]. However, their sizes will vary with the corresponding scene depths. Thus, we represent the inverse kernels of PSFs using different upsampled versions of a dictionary with small atoms. Such a multi-scale structure provides a physics-based implicit prior on the PSFs within the same image and leads to a more compact dictionary for representing inverse kernels of varying sizes. Specifically, we define a multi-scale dictionary composed by the sets (sub-dictionaries) $\{V^r\}_{r=1}^R$ : + +$$ +\boldsymbol {V} ^ {r} = \left[ \boldsymbol {v} _ {1} ^ {r}, \dots , \boldsymbol {v} _ {N} ^ {r} \right] \subset \mathbb {R} ^ {M _ {r} \times M _ {r}}, \boldsymbol {v} _ {n} ^ {r} = \mathcal {U} _ {S _ {r}} (\boldsymbol {v} _ {n}), \tag {6} +$$ + +where $\boldsymbol{v}_n^r$ denotes the atom at the $r$ -th scale, $S_r$ denotes an upsampling factor that determines the size parameter $M_r$ of the atom. As $r$ increases, $S_r$ becomes larger and $M_r$ gets smaller. The per-pixel inverse kernels are then expressed as + +$$ +\boldsymbol {k} _ {i, j} ^ {\dagger} = \omega_ {1, i, j} \boldsymbol {v} _ {1} ^ {r _ {i, j}} + \dots + \omega_ {N, i, j} \boldsymbol {v} _ {N} ^ {r _ {i, j}} \in \mathbb {R} ^ {M _ {r _ {i, j}} \times M _ {r _ {i, j}}}, \tag {7} +$$ + +where $r_{i,j}$ denotes the scale factor (index) related to $\pmb{k}_{i,j}^{\dagger}$ . Note that the atoms within the same sub-dictionary have a uniform size, while their sizes vary across different sub-dictionaries. That is, each sub-dictionary is utilized to represent inverse kernels of a specific size. + +Using the linear representation (7) and the fact that $\pmb{v}_n^{r_{i,j}}*$ $\pmb {y} = \sum_r\delta (r - r_{i,j})\pmb{v}_n^r *\pmb {y}$ with a Dirac delta function $\delta$ , we can rewrite Eq. (3) as + +$$ +\begin{array}{l} \mathcal {C} (\boldsymbol {y}) = \sum_ {i, j} \mathbf {1} _ {i, j} \odot \left(\sum_ {n = 1} ^ {N} \omega_ {n, i, j} \boldsymbol {v} _ {n} ^ {r _ {i, j}} * \boldsymbol {y}\right) (8) \\ = \sum_ {i, j} \mathbf {1} _ {i, j} \odot \left(\sum_ {n = 1} ^ {N} \sum_ {r = 1} ^ {R} \mu_ {r, i, j} \omega_ {n, i, j} \boldsymbol {v} _ {n} ^ {r} * \boldsymbol {y}\right), (9) \\ \end{array} +$$ + +where $\mathbf{1}_{i,j}$ denotes a mask with 1 for the $(i,j)$ -th entry being 1 and 0 otherwise, and $\mu_{r,i,j} = \delta (r - r_{i,j})$ . We can then perform spatially-varying inverse filtering by convolving each atom $\pmb{v}_n^r$ in the dictionary with the input image $\pmb{y}$ , and summing the resulting images by the spatially-varying weights $\mu_{s,i,j}\omega_{n,i,j}$ . The weights $\mu_{s,i,j}$ and $\omega_{n,i,j}$ control the size (adaptive to scene depth) and shape (adaptive to image) of the corresponding inverse kernel, respectively. + +# 3.3. INR Models for Dictionaries of Inverse Kernels + +The dictionary $\{V_r\}_{r=1}^R$ used in (8) is generated by the base atoms $\{v_j\}_{j=1}^N$ of the smallest size. However, simply upsampling these base atoms cannot cover all high frequencies, i.e., certain high-pass filters of large sizes may not be well approximated by the linear span of these atoms obtained by plain upsampling. To address this limitation, we use INR to re-parameterize the multi-scale dictionary. INR provides a more expressive way to generate multi-scale atoms that cover more frequencies than plain upsampling. Additionally, INR provides implicit regularization to alleviate overfitting by preferring image structures over random noise, as shown in [42, 19, 50]. + +Specifically, we express $\pmb{v}_n^r$ as + +$$ +\boldsymbol {v} _ {n} ^ {r} [ x, y ] = \Phi_ {n} (x, y), [ x, y ] \in [ 1, \dots , M _ {r} ] \times [ 1, \dots , M _ {r} ], +$$ + +where $[x,y]$ denotes a spatial coordinate, and $\Phi$ is an INR model implemented by a compact multi-layer perceptron (MLP) of a small size. The INR model maps spatial coordinates to a continuous function that implicitly represents the kernel atom. This function can be evaluated at any point in space to interpolate a kernel atom of an arbitrary size, forming a dictionary of inverse kernels with arbitrary scales. + +# 4. Implicit Neural Inverse Kernel Network + +Our proposed NN called INIKNet consists of recurrent deblurring stages over different scales for exploiting cross-scale contextual information. Each deblurring stage predicts and applies inverse kernels to deblur the input image. Concretely, each stage consists of two steps: (i) given an input image $\pmb{y}$ , inferring two weight tensors $\pmb{\mu},\pmb{\omega}$ that contain the weights $\{\mu_{r,i,j}\}_{r,i,j}$ and $\{\omega_{n,i,j}\}_{n,i,j}$ respectively; and (ii) applying inverse filtering defined by (8) to deblurring $\pmb{y}$ . + +# 4.1. Duplex Scale-Recurrent Framework + +Current scale-recurrent deblurring NNs, e.g. [36, 28, 48, 12, 54, 8, 15], typically use an LSTM-based coarse-to-fine framework. The deblurred image from the current (coarser) scale is attached to the input of the next (finer) scale, resulting in unidirectional feature flow that only considers coarser scales when inferring at finer scales. This approach cannot fully leverage the additional information available at finer scales for deblurring on coarser scales, especially when the blur effects show strong similarity over space and scale. + +To address this limitation, we propose a duplex multiscale processing framework, which leverages bi-LSTM [13] as the sub-NNs to predict $\mu$ and $\omega$ . It improves predictions at coarser scales by incorporating additional information from finer scales through the feature flow in the fine-to-coarse pass. This refined information is then utilized to further improve predictions on finer scales via the feature flow through the coarse-to-fine pass. As a result, the prediction at the current scale benefits from the information extracted from both the previous and subsequent scales. + +As shown in Fig. 1, our INIKNet first downsamples the input image to multi-scale versions. For each scale, a U-Net shared across scales is first applied for extracting spatial features from the input. The extracted features are then fed to two bi-LSTM sub-NNs so as to exploit the dependencies among scales for feature refinement. The outputs of two bi-LSTMs sub-NNs are then used to predict $\mu$ and $\omega$ via an $1\times 1$ convolutional layer, respectively. Afterward, the predicted weights are used by an Inverse Filtering Module (IFM) to perform deblurring using (8) to output the result. + +Our INIKNet differs from existing coarse-to-fine NNs in that it does not concatenate the output image from the current scale to the input of the next scale, as this operation would not only make the feature flow unidirectional, but also prevent the current-scale processing from starting, until the deblurring is fully completed at the previous scale. By removing this operation, the U-Net and bi-LSTM can start running earlier, without waiting for the deblurred images from previous scales. Moreover, by cancelling the operation, the IFM only needs to be called at the original image scale during inference, even though it is still required during training to calculate the multi-scale loss. + +# 4.2. Details of Key Modules + +Bi-LSTM sub-NNs Recall that the weights $\mu$ and $\omega$ correspond to two distinct properties of inverse kernels, namely size and shape, respectively. Their dependency on image scales also differs significantly. Therefore, rather than use a single bi-LSTM sub-NN to predict $\mu$ and $\omega$ jointly, we use two parallel bi-LSTM sub-NNs for separate predictions. Both sub-NNs share the same structure, consisting of convolutional LSTM cells [45] arranged in two paths for fine-to-coarse and coarse-to-fine estimations in opposite directions. Each LSTM cell takes a cell state and a hidden state as input and outputs updated states. The input cell state is defined as the features extracted by the U-Net, while the input hidden state for the next scale is set to the output one from the previous scale, forming a cross-scale connection. + +Weight predictors To convert the output features from the bi-LSTM sub-NNs to the weights $\mu$ and $\omega$ , two $1 \times 1$ convolutional layers are applied, respectively. For predicting $\mu$ , the Softmax function is used to impose non-negativity and $\ell_1$ -normalization constraints. Such a soft relaxation of + +![](images/d18ae44616a483e33cf54e273e30b5a81a94276c9dd19a50eb96e85fca5e4260.jpg) +Figure 1: Architecture of proposed INIKNet for SIDD. + +![](images/a65389a81c005a72bc9106fb2b811f9eb8ee4eae426d77a2859443fe518586bc.jpg) + +weights $\mu_{r,i,j}$ from 0,1 to [0,1] not only makes training easier, but also improves the representation accuracy of inverse kernels by including the use of additional atoms. This can be useful when the inverse kernel sizes fall between the gaps among the predefined non-consecutive atom sizes. + +INR models of kernel atoms The MLPs used for INR have a structure similar to those in [46]. They comprise three fully-connected (FC) layers separated by a Sine activation function with a learnable frequency parameter that controls frequency content. The 2D coordinates are first mapped to a high-dimensional space and then projected to single values. One difference from [46] is that we first transform spatial coordinates to polar coordinates for slight improvement. + +IFM for deblurring The IFM first calls the learned MLPs to generate all kernel atoms $\{\pmb{v}_n^r\}_{n,r}$ and then calculates the feature tensors $\pmb{f}[n,r] = \pmb{v}_n^r*\pmb{y}$ . Afterward, Eq. (8) is implemented by simple entry-wise product and summation operations on the tensors $\pmb{f},\pmb{\mu},\pmb{\omega}$ . + +# 4.3. Training Loss + +In addition to the $\ell_2$ loss $\mathcal{L}_{\mathrm{MSE}}$ defined by mean squared error (MSE), two loss functions often seen in image recovery are also used for training, including the frequency-domain loss $\mathcal{L}_{\mathrm{FD}}$ [6] defined as the $\ell_1$ distance in the frequency domain, and the Learned Perceptual Image Patch Similarity (LPIPS) loss $\mathcal{L}_{\mathrm{LPIPS}}$ [57] predicted by a pre + +trained NN. These three loss functions measure deblurring quality in the spatial, frequency and feature (perceptual) domains, respectively. The total loss is then given by + +$$ +\mathcal {L} = \sum_ {p} \mathcal {L} _ {\mathrm {M S E}} ^ {(p)} + \lambda_ {1} \mathcal {L} _ {\mathrm {F D}} ^ {(p)} + \lambda_ {2} \mathcal {L} _ {\mathrm {L P I P S}} ^ {(p)}, \tag {10} +$$ + +where $p$ denotes the index of the $p$ th scale, and $\lambda_1, \lambda_2$ are empirically set to 0.1. Ground-truth images are downsampled accordingly for the supervision at different scales. + +# 5. Experiments + +Datasets and metrics Following [2, 41], we train models on two widely-used datasets, DPDD [2] and LFDOF [41], respectively. To assess their generalization performance, we also evaluate these models on two more datasets, RTF [7] and RealDOF [21]. Three quality metrics are used for quantitative evaluation: Peak Signal to Noise Ratio (PSNR, measured in dB), Structural SIMilarity index (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [57]. The model complexity is measured by both the number of parameters and the inference time on a $640 \times 640$ image using an NVIDIA 2080Ti RTX GPU. + +Implementation details Throughout the experiments, we employ 8 scales and 10 INR-based atoms per scale: $R = 8$ and $N = 10$ . We use atoms of sizes $1 \times 1, 3 \times 3, \dots, 15 \times 15$ , + +where the $1 \times 1$ atom is fixed as a delta kernel for modeling in-focus regions. The frequency parameters of the sine activation functions in all INR models are initialized by random sampling from [2, 16]. We call Adam for training, with $4 \times 10^{5}$ iterations and a batch size of 4. We start with an initial learning rate of $2 \times 10^{-4}$ and halve it every $10^{5}$ iterations. Random flipping, rotation, and cropping (to $256 \times 256$ pixels) are performed for training data augmentation. + +# 5.1. Performance Evaluation + +Compared methods We select 9 NN-based SIDD methods for comparison, including the DPDNet [2], AIFNet [41], IFANet [21], KPAC [47], GKMNet [36], MDP [1], DRBNet [40], Restormer [53], and MPRNet [26] trained in [23] with a misalignment-robust scheme. The experimental results of the compared methods are reported either from their original papers or obtained by using their released pretrained models and codes. AIFNet is a model trained on LFDOF with the additional use of SYNDOF for its DME step. To compare its performance on DPDD, we retrain it using both DPDD and SYNDOF. DRBNet is a model trained on LFDOF and fine-tuned on DPDD. For a fair comparison, we retrain it on both datasets, respectively. + +Evaluation on DPDD and LFDOF Table 1 presents the quantitative results on DPDD and LFDOF. Our INIKNet performs the best in PSNR on DPDD and in all three metrics on LFDOF, despite using a lightweight model whose parameter number is the second smallest and around $13.2\%$ of that of Restormer, a general restoration model. Compared to the latest SIDD-specified MRPNet model [23], INIKNet achieves over $0.3\mathrm{dB}$ PSNR improvement on DPDD while having only about $10\%$ parameters to learn. Similar results are observed in the comparison with DRBNet, indicating that the performance gain of INIKNet comes from the design of NN structure rather than the increased model size. Though INIKNet is slightly larger than GKMNet, the smallest model, it achieves noticeably better performance. Compared to KPAC that uses dilated convolutions to mimic inverse kernels, INIKNet performs significantly better. In conclusion, the superiority and compactness of INIKNet demonstrate the effectiveness of its architecture design. + +Generalization analysis on RealDOF and RTF The results of the DPDD/LFDOF-trained models evaluated on RealDOF and RTF are listed in Table 2. Our DPDD-trained INIKNet shows good generalization to different datasets and achieves the best performance on both RealDOF and RTF in all three metrics. The LFDOF-trained INIKNet is the top performer on RealDOF and achieves the best SSIM score on RTF, as well as the second-best scores in PSNR and LPIPS. Though AIFNet performs better in PSNR and LPIPS, it requires an additional dataset for DME and is much larger than INIKNet. In summary, INIKNet exhibits superior generalizability over other compared models. + +
ModelDPDDLFDOF#Params (106)Time (sec.)
PSNRSSIMLPIPSPSNRSSIMLPIPS
DPDNet24.3480.7470.277---32.251.430
AIFNet24.2130.7420.30929.6770.8840.20241.550.276
IFANet25.3660.7890.21729.7870.8720.15610.480.063
KPAC25.2210.7740.22628.9420.8570.1742.060.037
GKMNet25.4680.7890.21929.0810.8670.1711.410.129
MDP25.3470.7630.26828.0690.8340.18546.863.317
DRBNet25.4850.7920.25430.2530.8830.14711.690.085
MPRNet25.7300.7920.232---20.101.161
Restormer25.9800.8110.178---26.100.672
INIKNet26.0550.8030.18530.2930.8860.1321.980.228
+ +Table 1: Quantitative results on DPDD/LFDOF. Bold for best performers and underline for second-best performers. + +
ModelRealDOFRTF#Params (106)
PSNRSSIMLPIPSPSNRSSIMLPIPS
Trained on DPDDDPDNet22.8700.6700.42523.6080.5910.29632.25
AIFNet23.0930.6800.41324.0410.7580.28941.55
MDP23.5000.6810.44424.0120.7380.31246.86
KPAC23.9750.7620.33824.6180.7770.2362.06
IFANet24.7120.7480.30624.9240.8010.22710.48
GKMNet24.2540.7320.39224.9700.7890.2611.41
DRBNet24.8840.7510.37624.4630.7730.31111.69
MPRNet24.5410.7360.33924.5880.7880.30420.10
Restormer25.0910.7620.28524.2120.8210.22426.10
INIKNet25.2310.7650.28725.4500.8340.2151.98
Trained on LFDOFIFANet22.5040.6690.48326.4370.8380.23810.48
KPAC22.5500.6710.45725.9590.8030.2302.06
AIFNet22.6230.6670.46127.5520.8820.17641.55
GKMNet23.6090.7210.40826.9850.8560.2461.41
MDP22.7260.6800.45325.5800.8090.22846.86
DRBNet22.9100.6910.43726.7170.8530.20011.69
INIKNet23.8100.7240.35627.4010.8850.1791.98
+ +Table 2: Quantitative results on RealDOF/RTF. Bold for best performers and underline for second-best performers. + +Qualitative comparison We compare INIKNet with some SIDD-specified competitors on DPDD and LFDOF via visual inspection. See Fig. 2 for some deblurred images on four datasets, where complex textures such as those on wire fence and branches are successfully recovered by INIKNet. See also Fig. 3 for some results on the CUHK dataset [43]. While most methods fail under large blur, ours yields more visually satisfying results. These findings demonstrate that our INIKNet performs well in terms of visual quality. + +# 5.2. Ablation Studies and Analysis + +Ablation studies We construct several variants of INIKNet and list their results on two datasets in Table 3. + +Non-MS Dict: We drop the multi-scale structure of the dictionary and instead use an individual INR model for each dictionary atom at each scale (thus yielding a larger model size). A noticeable performance degradation occurs. This is because the multi-scale structure of the INR-based dictionary imposes structural regularization to reduce overfitting. + +![](images/d738c76f004176285f2135664e6a221410f8981637ba70da43a0978e2c1dc83e.jpg) +Figure 2: Results on using models trained on DPDD (left part) and LFDOF (right part). Zoom-in for better views. + +![](images/259a5b098bb38f3343912588c4db9a3572482b1e03416ac5a055b6a1ff2dfd47.jpg) + +
ModelRealDOFRTF
PSNRSSIMLPIPSPSNRSSIMLPIPS
Non-MS Dict24.9160.7490.31025.1360.8090.254
Non-INR Dict25.0580.7540.30325.2370.8190.226
Unidirectional25.0990.7530.29425.3050.8240.227
Original INIKNet25.2310.7650.28725.4500.8340.215
+ +Table 3: Results of baseline models trained on DPDD. + +Non-INR Dict: When the INR-based dictionaries are replaced by the plain ones, where each base atom $v_{n}$ is defined as a learnable matrix and bi-linearly upsampled to form the multi-scale dictionary, the performance of INIKNet decreases. This is probably because the frequency content for the atoms generated by plain upsampling are limited, so do their generated inverse kernels. + +Unidirectional: To simulate an unidirectional scale recurrent mechanism (e.g. [36]), we replace the bi-LSTM with double unidirectional LSTMs (for keeping the model size). The output from the last scale is attached to the input of the current scale for additional information. Such a scheme for coefficient prediction results in some performance degradation, highlighting the effectiveness of our proposed duplex scale-recurrent framework. + +Visualization of coefficients and atoms The coefficient maps $\mu$ and $\omega$ are visualized in Fig. 4. We can see that the regions with larger blur amount tend to have larger coefficients $\mu$ assigned to larger inverse kernels, and vice versa. In comparison, the coefficients $\omega$ are more spatially similar (gray), consistent with the prior of spatially-similar shapes of inverse kernels. We use Fig. 5 to visualize some learned (non-)INR-based atoms in both the spatial and Fourier domain. The INR-based atoms within the same scale contain diverse patterns with varying frequencies and orientations, and meanwhile they have similar shapes across scales. Compared to the non-INR-based ones, the INR-based atoms encode richer spatial structures and frequency contents. Interestingly, the INR-based atoms exhibit strong isotropy, which is likely due to the isotropy of defocus PSFs (so do their inverse kernels). All these properties of the INR-based atoms enable our INIKNet to effectively represent and predict inverse kernels for spatially-varying defocus blur. + +# 6. Conclusion + +This paper presented an interpretable end-to-end SIDD approach, which learns a duplex scale-recurrent NN to predict INR-driven dictionary-parameterized inverse kernels of + +![](images/f453722cad9142760713cb963b3cd48a9a288922863cd894392a109d48c3a1a5.jpg) +Figure 3: Results on CUHK dataset (without ground-truths) using models trained on DPD. Zoom-in for better views. + +![](images/4ec66da483cb1d5e4b7c54ccb4f098f49ad562d4a69bc32559ee44057a86a917.jpg) +Figure 4: Visualization of $\pmb{\mu}$ maps and $\omega$ maps. + +![](images/2c75e98d78c435e3ac290ac63b8e0a99078eccb3c236168faaf0ec7d518ab6ac.jpg) +Figure 5: Visualization of bilinearly upsampled (non-INR-based) atoms and INR-based atoms in two domains. + +defocus blur and applies them to deblur images. Our proposed approach introduced effective structural regularization on inverse kernels for good generalization, which is + +based on the shape similarity of defocus PSFs over an image. The resulting effective and light-weight model has demonstrated its advantages by extensive experiments. + +# References + +[1] Abdullah Abuolaim, Mahmoud Afifi, and Michael S. Brown. Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1231-1239, 2022. 1, 2, 6 +[2] Abdullah Abuolaim and Michael S. Brown. Defocus deblurring using dual-pixel data. In Proceedings of the European Conference on Computer Vision, pages 111-126. Springer, 2020. 2, 5, 6 +[3] Yosuke Bando and Tomoyuki Nishita. Towards digital refocusing from a single photograph. In Proceedings of the Pacific Conference on Computer Graphics and Applications, pages 363-372. IEEE, 2007. 1 +[4] Mingqin Chen, Yuhui Quan, Tongyao Pang, and Hui Ji. Non-blind image deconvolution via leveraging model uncertainty in an untrained deep neural network. International Journal of Computer Vision, 130(7):1770-1789, 2022. 2 +[5] Sunghyun Cho and Seungyong Lee. Convergence analysis of map based blur kernel estimation. In Proceedings of the IEEE/International Conference on Computer Vision, pages 4808-4816, 2017. 1 +[6] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4641-4650, 2021. 3, 5 +[7] Laurent D'Andrés, Jordi Salvador, Axel Kochale, and Sabine Süssstrunk. Non-parametric blur map regression for depth of field extension. IEEE Transactions on Image Processing, 25(4):1660-1673, 2016. 1, 5 +[8] Senyou Deng, Wenqi Ren, Yanyang Yan, Tao Wang, Fenglong Song, and Xiaochun Cao. Multi-scale separable network for ultra-high-definition video deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14030-14039, 2021. 2, 4 +[9] Jiangxin Dong, Stefan Roth, and Bernt Schiele. Deep wiener deconvolution: Wiener meets deep learning for image deblurring. Advances in Neural Information Processing Systems, 33:1048-1059, 2020. 2 +[10] Jiangxin Dong, Stefan Roth, and Bernt Schiele. Dwdn: deep wiener deconvolution network for non-blind image deblurring. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9960-9976, 2021. 2 +[11] Zhenxuan Fang, Fangfang Wu, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Self-supervised non-uniform kernel estimation with flow-based motion prior for blind image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18105-18114, 2023. 2 +[12] Hongyun Gao, Xin Tao, Xiaoyong Shen, and Jiaya Jia. Dynamic scene deblurring with parameter selective sharing and nested skip connections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3848-3856, 2019. 2, 4 +[13] Alex Graves and Jürgen Schmidhuber. Framewise phoneme classification with bidirectional LSTM and other neural net- + +work architectures. Neural networks, 18(5-6):602-610, 2005. 2, 4 +[14] Fang Hua, Peter Johnson, Nadezhda Sazonova, Paulo Lopez-Meyer, and Stephanie Schuckers. Impact of out-of-focus blur on face recognition performance based on modular transfer function. In Proceedings of the International Conference on Biometrics, pages 85–90. IEEE, 2012. 1 +[15] Bo Ji and Angela Yao. Multi-scale memory-based video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1919-1928, 2022. 2, 4 +[16] Ali Karaali, Naomi Harte, and Claudio R. Jung. Deep multiscale feature learning for defocus blur estimation. IEEE Transactions on Image Processing, 31:1097-1106, 2022. 2 +[17] Ali Karaali and Claudio Rosito Jung. Edge-based defocus blur estimation with adaptive scale selection. IEEE Transactions on Image Processing, 27(3):1126-1137, Mar. 2018. 1, 2 +[18] Ali Karaali and Cláudio Rosito Jung. Svbr-net: A non-blind spatially varying defocus blur removal network. In Proceedings of the IEEE International Conference on Image Processing, pages 566-570. IEEE, 2022. 1 +[19] Chaewon Kim, Jaeho Lee, and Jinwoo Shin. Zero-shot blind image denoising via implicit neural representations. arXiv preprint arXiv:2204.02405, 2022. 3, 4 +[20] Dilip Krishnan and Rob Fergus. Fast image deconvolution using hyper-laplacian priors. In Advances in Neural Information Processing Systems, 2009. 2 +[21] Junyong Lee, Hyeongseok Son, Jaesung Rim, Sunghyun Cho, and Seungyong Lee. Iterative filter adaptive network for single image defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2034-2042, 2021. 1, 2, 5, 6 +[22] Dasong Li, Yi Zhang, Ka Chun Cheung, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. Learning degradation representations for image deblurring. In Proceedings of the European Conference on Computer Vision, pages 736-753. Springer, 2022. 2 +[23] Yu Li, Dongwei Ren, Xinya Shu, and Wangmeng Zuo. Learning single image defocus deblurring with misaligned training pairs. Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 1, 2, 6 +[24] Ziwei Luo, Haibin Huang, Lei Yu, Youwei Li, Haoqiang Fan, and Shuaicheng Liu. Deep constrained least squares for blind image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17642-17652, 2022. 2 +[25] Haoyu Ma, Shaojun Liu, Qingmin Liao, Juncheng Zhang, and Jing-Hao Xue. Defocus image deblurring network with defocus map estimation as auxiliary task. IEEE Transactions on Image Processing, 31:216-226, 2021. 1, 2 +[26] Armin Mehri, Parichehr B Ardakani, and Angel D Sappa. Mprnet: Multi-path residual network for lightweight image super resolution. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2704-2713, 2021. 6 +[27] Sung-Jun Min, Kyeongbo Kong, and Suk-Ju Kang. Out-of-focus image deblurring for mobile display vision inspection. + +IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1 +[28] Seungjun Nah, Tae Kim, and Kyoung Mu Lee. Deep multiscale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 257-265, 2017. 2, 4 +[29] Yuesong Nan and Hui Ji. Deep learning for handling kernel/model uncertainty in image deconvolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2388-2397, 2020. 2 +[30] Yuesong Nan, Yuhui Quan, and Hui Ji. Variational-EM-based deep learning for noise-blind image deblurring. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3626–3635, 2020. 2 +[31] Saqib Nazir, Lorenzo Vaquero, Manuel Mucientes, Víctor M Brea, and Daniela Coltuc. Depth estimation and image restoration by deep learning from defocused images. arXiv preprint arXiv:2302.10730, 2023. 1, 2 +[32] Jinsun Park, Yu-Wing Tai, Donghyeon Cho, and In So Kweon. A unified approach of multi-scale deep and handcrafted features for defocus estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1736-1745, 2017. 1, 2 +[33] Valeriya Pronina, Filippos Kokkinos, Dmitry V Dylov, and Stamatios Lefkimmiatis. Microscopy image restoration with deep wiener-kolmogorov filters. In Proceedings of the European Conference on Computer Vision, pages 185-201. Springer, 2020. 1, 2 +[34] Yuhui Quan, Zhuojie Chen, Huan Zheng, and Hui Ji. Learning deep non-blind image deconvolution without ground truths. In Proceedings of the European Conference on Computer Vision, pages 642-659, 2022. 2 +[35] Yuhui Quan, Peikang Lin, Yong Xu, Yuesong Nan, and Hui Ji. Nonblind image deblurring via deep learning in complex field. IEEE Transactions on Neural Networks and Learning Systems, 33(10):5387-5400, 2022. 2 +[36] Yuhui Quan, Hui Wu, Zicong, and Hui Ji. Gaussian kernel mixture network for single image defocus deblurring. In Advances in Neural Information Processing Systems, volume 34, 2021. 1, 2, 4, 6, 7 +[37] Yuhui Quan, Hui Wu, Zicong, and Hui Ji. Neumann network with recursive kernels for single image defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5754-5763, June 2023. 1, 2 +[38] Wenqi Ren, Jiawei Zhang, Lin Ma, Jinshan Pan, Xiaochun Cao, Wangmeng Zuo, Wei Liu, and Ming-Hsuan Yang. Deep non-blind deconvolution via generalized low-rank approximation. Advances in Neural Information Processing Systems, 31, 2018. 2 +[39] Wenqi Ren, Jiawei Zhang, Jinshan Pan, Sifei Liu, Jimmy S Ren, Junping Du, Xiaochun Cao, and Ming-Hsuan Yang. Deblurring dynamic scenes via spatially varying recurrent neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):3974-3987, 2021. 2 + +[40] Lingyan Ruan, Bin Chen, Jizhou Li, and Miuling Lam. Learning to deblur using light field generated and real defocus images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16304-16313, 2022. 1, 2, 6 +[41] Lingyan Ruan, Bin Chen, Jizhou Li, and Miu-Ling Lam. Aifnet: All-in-focus image restoration network using a light field-based dataset. IEEE Transactions on Computational Imaging, 7:675-688, 2021. 1, 2, 5, 6 +[42] Nimrod Shabtay, Eli Schwartz, and Raja Giryes. Pip: Positional-encoding image prior. arXiv preprint arXiv:2211.14298, 2022. 3, 4 +[43] Jianping Shi, Li Xu, and Jiaya Jia. Discriminative blur detection features. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2965-2972, 2014. 6 +[44] Jianping Shi, Li Xu, and Jiaya Jia. Just noticeable defocus blur detection and estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 657-665, 2015. 1, 2 +[45] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in Neural Information Processing Systems, 28, 2015. 4 +[46] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 2, 5 +[47] Hyeongseok Son, Junyong Lee, Sunghyun Cho, and Seungyong Lee. Single image defocus deblurring using kernel-sharing parallel atrous convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2642-2650, 2021. 1, 2, 3, 6 +[48] Xin Tao, Hongyun Gao, Yi Wang, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8174-8182, 2018. 2, 4 +[49] Longguang Wang, Yingqian Wang, Xiaoyu Dong, Qingyu Xu, Jungang Yang, Wei An, and Yulan Guo. Unsupervised degradation representation learning for blind superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10581-10590, 2021. 2 +[50] Shaowen Xie, Hao Zhu, Zhen Liu, Qi Zhang, You Zhou, Xun Cao, and Zhan Ma. Diner: Disorder-invariant implicit neural representation. arXiv preprint arXiv:2211.07871, 2022. 3, 4 +[51] Guodong Xu, Yuhui Quan, and Hui Ji. Estimating defocus blur via rank of local patches. In Proceedings of the IEEE/International Conference on Computer Vision, pages 5371-5379, 2017. 1, 2 +[52] Li Xu, Jimmy S Ren, Ce Liu, and Jiaya Jia. Deep convolutional neural network for image deconvolution. Advances in Neural Information Processing Systems, 27, 2014. 2 +[53] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. + +Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 6 +[54] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5978-5986, 2019. 2, 4 +[55] Jiawei Zhang, Jinshan Pan, Jimmy Ren, Yibing Song, Linchao Bao, Rynson WH Lau, and Ming-Hsuan Yang. Dynamic scene deblurring using spatially variant recurrent neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2521-2529, 2018. 2 +[56] Jie Zhang and Wanming Zhai. Blind attention geometric restraint neural network for single image dynamic/defocus deblurring. IEEE Transactions on Neural Networks and Learning Systems, 2022. 1, 2 +[57] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 5 +[58] Wenda Zhao, Fei Wei, You He, and Hutchuan Lu. United defocus blur detection and deblurring via adversarial promoting learning. In Proceedings of the European Conference on Computer Vision, pages 569-586, 2022. 1, 2 +[59] Wenda Zhao, Fan Zhao, Dong Wang, and Huchuan Lu. Defocus blur detection via multi-stream bottom-top-bottom network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42:1884-1897, 2020. 2 +[60] Bolun Zheng, Quan Chen, Shanxin Yuan, Xiaofei Zhou, Hua Zhang, Jiyong Zhang, Chenggang Yan, and Gregory Slabaugh. Constrained predictive filters for single image bokeh rendering. IEEE Transactions on Computational Imaging, 8:346-357, 2022. 1 +[61] Chao Zhu, Hang Dong, Jinshan Pan, Boyang Liang, Yuhao Huang, Lean Fu, and Fei Wang. Deep recurrent neural network with multi-scale bi-directional propagation for video deblurring. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 3598-3607, 2022. 3 \ No newline at end of file diff --git a/2023/Single Image Defocus Deblurring via Implicit Neural Inverse Kernels/images.zip b/2023/Single Image Defocus Deblurring via Implicit Neural Inverse Kernels/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..df5b00bff61b7c924d586d6adaa7d607dc2776d9 --- /dev/null +++ b/2023/Single Image Defocus Deblurring via Implicit Neural Inverse Kernels/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46f5123aa178b1fde5b0c724eb423c8adf3a949428a98970dacae9875a22633f +size 1089838 diff --git a/2023/Single Image Defocus Deblurring via Implicit Neural Inverse Kernels/layout.json b/2023/Single Image Defocus Deblurring via Implicit Neural Inverse Kernels/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fba7ee271f5b0d419e31b5b4f5492192ce677480 --- /dev/null +++ b/2023/Single Image Defocus Deblurring via Implicit Neural Inverse Kernels/layout.json @@ -0,0 +1,8967 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "type": "text", + "content": "Single Image Defocus Deblurring via Implicit Neural Inverse Kernels" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 142, + 414, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 142, + 414, + 157 + ], + "spans": [ + { + "bbox": [ + 179, + 142, + 414, + 157 + ], + "type": "text", + "content": "Yuhui Quan" + }, + { + "bbox": [ + 179, + 142, + 414, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 179, + 142, + 414, + 157 + ], + "type": "text", + "content": " Xin Yao" + }, + { + "bbox": [ + 179, + 142, + 414, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 179, + 142, + 414, + 157 + ], + "type": "text", + "content": " Hui Ji" + }, + { + "bbox": [ + 179, + 142, + 414, + 157 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 158, + 527, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 158, + 527, + 184 + ], + "spans": [ + { + "bbox": [ + 66, + 158, + 527, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 158, + 527, + 184 + ], + "type": "text", + "content": "School of Computer Science and Engineering, South China University of Technology, Guangzhou 510006, China " + }, + { + "bbox": [ + 66, + 158, + 527, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 158, + 527, + 184 + ], + "type": "text", + "content": " Pazhou Lab, Guangzhou 510335, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 186, + 470, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 186, + 470, + 200 + ], + "spans": [ + { + "bbox": [ + 126, + 186, + 470, + 200 + ], + "type": "text", + "content": "3 Department of Mathematics, National University of Singapore, 119076, Singapore" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 201, + 470, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 201, + 470, + 213 + ], + "spans": [ + { + "bbox": [ + 126, + 201, + 470, + 213 + ], + "type": "text", + "content": "cseyhquan@scut.edu.cn, csxinyaoscut@gmail.com, matjh@nus.edu.sg*" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 266, + 289, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 289, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 289, + 482 + ], + "type": "text", + "content": "Single image defocus deblurring (SIDD) is a challenging task due to the spatially-varying nature of defocus blur, characterized by per-pixel point spread functions (PSFs). Existing deep-learning-based methods for SIDD are limited by either over-fitting due to the lack of model constraints or under-parametrization that restricts their applicability to real-world images. To address the limitations, this paper proposes an interpretable approach that explicitly predicts inverse kernels with structural regularization. Motivated by the observation that defocus PSFs within an image often have similar shapes but different sizes, we represent the inverse kernels linearly over a multi-scale dictionary parameterized by implicit neural representations. We predict the corresponding representation coefficients via a duplex scale-recurrent neural network that jointly performs fine-to-coarse and coarse-to-fine estimations. Extensive experiments demonstrate that our approach achieves excellent performance using a lightweight model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 502, + 128, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 128, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 128, + 514 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 522, + 287, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 642 + ], + "type": "text", + "content": "When a camera captures an image, objects outside of the focal plane appear blurry. This effect is known as defocus blur and usually occurs when using a large camera aperture or capturing scenes with significantly varying depths. Removing defocus blur from a single image, known as SIDD, is desired for many practical applications; see e.g. [27, 14, 33, 60]. In general, defocus blur varies spatially. Each defocused pixel is a weighted average of its neighboring pixels in the latent all-in-focus image, with the weights determined by a spatially-varying unknown PSF." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 643, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 643, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 643, + 287, + 677 + ], + "type": "text", + "content": "SIDD is closely related to defocus map estimation (DME) [3, 44, 7, 5, 51, 32, 17]. Existing DME methods typically model defocus PSFs using a Gaussian function" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 242, + 546, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 242, + 546, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 242, + 546, + 373 + ], + "type": "text", + "content": "parameterized by its variance or a disc function parameterized by its radius. The resulting defocus map indicates the per-pixel blur amount in terms of the values of such model parameters, which can then be utilized for SIDD by applying a non-blind image deblurring (NID) algorithm using the corresponding PSFs; see e.g. [18]. However, DME remains a challenging task, and the estimated parameters of PSFs can be erroneous for many pixels. Moreover, Gaussian or disc-indicator functions are overly simplistic for real-world defocus PSFs. All these errors will be magnified in the NID process, resulting in noticeable artifacts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "text", + "content": "Recently, many deep-learning-based methods have been proposed for SIDD; see e.g. [41, 25, 21, 47, 36, 40, 1, 56, 58, 23, 31]. Most train an end-to-end neural network (NN) that directly maps blurry images to their sharp correspondences. Although these methods provide better results than the DME-based two-step methods, their performance still has much room for improvement due to significant variations in the spatial distributions of defocus PSFs among different images. Many existing works (e.g. [21, 40, 56]) explicitly introduce spatially-varying processing blocks to better handle the spatial variance of defocus PSFs. However, those blocks lack constraints and may cause overfitting. A few studies (e.g. [36, 37]) represent defocus PSFs by some specific basis for geometric regularization and employ deep unrolling for an NN-based inversion process. Nevertheless, there is still room for improvement." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "This paper presents an end-to-end deep NN with high interpretability for SIDD, which explicitly predicts spatially-varying inverse kernels of defocus PSFs and performs deblurring using the predicted inverse kernels. The possible overfitting issue is tackled by providing a more accurate model with structural constraints/regularization for the inverse kernels. Specifically, the inverse kernels are modeled by a linear representation under a dictionary. To achieve a compact yet sufficiently general representation, we construct the dictionary with a multi-scale structure. This is motivated by the empirical observation (as seen in [47]) that defocus PSFs within an image tend to have the same shape" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "text", + "content": "*This work is supported by the Natural Science Foundation of Guangdong Province (Grant No. 2022A1515011755 and 2023A1515012841), and Singapore MOE AcRF Tier 1 Grant (WBS No. A-8000981-00-00)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12600" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "but vary mainly in size. We show that if two defocus PSFs of different sizes have the same shape under an upsampling operation, so do the dictionary atoms of their inverse kernels. Thus, the dictionary atoms of a large size are defined as an upsampled version of the atoms of a small size." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 287, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 287, + 252 + ], + "type": "text", + "content": "Although plain upsampling offers an economical method for generating larger dictionary atoms from smaller ones, the resulting atoms may have a limited frequency range. This can lead to poor approximations of certain inverse kernels. To address the limitation, we leverage implicit neural representation (INR) [46], a technique that uses coordinate-input NNs to represent geometric objects like 2D shapes, so as to efficiently parameterize multi-scale atoms with both sufficient coverage of high-frequency content and implicit regularization to alleviate overfitting." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 253, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 287, + 373 + ], + "type": "text", + "content": "For the multi-scale representation coefficients, we separate them into scale-related and shape-related components, which are predicted by two scale-recurrent sub-NNs respectively. To better utilize information from different scales for coefficient prediction, we introduce a duplex scale-recurrent framework that performs both fine-to-coarse and coarse-to-fine estimations. Once the coefficients are predicted, the corresponding inverse kernels are applied to the input image for deblurring. The resulting pipeline is highly efficient and interpretable, leading to a lightweight yet effective model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 373, + 287, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 421 + ], + "type": "text", + "content": "To conclude, this paper proposes an end-to-end NN for SIDD, which exhibits a noticeable performance boost compared to existing ones, while maintaining low complexity. See below for our main contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 427, + 286, + 516 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 59, + 427, + 286, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 427, + 286, + 451 + ], + "spans": [ + { + "bbox": [ + 59, + 427, + 286, + 451 + ], + "type": "text", + "content": "- A parametric inverse kernel prediction framework for SIDD, providing efficient and interpretable deep NNs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 454, + 286, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 454, + 286, + 490 + ], + "spans": [ + { + "bbox": [ + 59, + 454, + 286, + 490 + ], + "type": "text", + "content": "- A multi-scale linear representation model for the inverse kernels of spatially-varying defocus blur, using an INR-based multi-scale dictionary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 493, + 286, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 493, + 286, + 516 + ], + "spans": [ + { + "bbox": [ + 59, + 493, + 286, + 516 + ], + "type": "text", + "content": "- A duplex scale-recurrent framework for predicting the representation coefficients of inverse kernels." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 529, + 138, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 138, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 138, + 542 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 550, + 212, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 212, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 212, + 563 + ], + "type": "text", + "content": "2.1. Existing Approaches for SIDD" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 569, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 676 + ], + "type": "text", + "content": "SIDD can done by a two-step approach that combines DME (e.g. [44, 51, 17, 32]) and NID (e.g. [20]). However, such an approach has two inherent drawbacks. First, the errors in DME will be magnified in the following NID process. Second, simple Gaussian or disc functions do not fit real-world defocus PSFs well. While there has been steady progress in deep-learning-based methods for DME [59, 16] and for NID [38, 10, 33, 29, 30, 4, 34, 35], the inherent drawbacks of the two-step approach remain." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Deep end-to-end learning has emerged as a promising approach for SIDD. The seminal work conducted by Abuolaim and Brown [2] trained an encoder-decoder CNN that" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "maps a defocused image to its in-focus correspondence, with significant performance gain over two-step methods. To better handle spatially-varying defocus blur, some studies introduced dynamic processing blocks, e.g., per-pixel filter prediction in Lee et al. [21], dynamic residual blocks in Ruan et al. [40], and attention in Zhang and Zhai [56]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 228 + ], + "type": "text", + "content": "The use of dynamic processing blocks may increase the model complexity and the risk of overfitting. Regularization by auxiliary tasks is an effective approach for improving the generalization performance. In existing works, auxiliary tasks are often defined using DME ([25, 41, 58]), depth estimation [31], dual-view data [21, 1], or reblurring loss [23]. However, most of them require additional data sources." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 228, + 546, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 546, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 546, + 468 + ], + "type": "text", + "content": "A more appealing approach is encoding physical priors of defocus blur into NNs. Quan et al. [36, 37] used a kernel mixture model to parameterize defocus PSFs and constructed a deep unrolling NN. This method can be viewed as the degradation learning [22, 49, 24, 11] that predicts the blurring operator or kernels for deblurring. Our approach differs in that it predicts inverse kernels, not the degradation with defocus PSFs. Son et al. [47] showed that inverse defocus kernels have an invariant shape across scales and emulated them in feature spaces via a series of kernel-sharing parallel Atrous convolutions. Their method inspired our work, and it can be viewed as representing inverse kernels for a feature tensor via a multi-scale dilated combination of a single atom and predicting spatially-invariant combination coefficients via channel attention. In comparison, our approach explicitly predicts inverse kernels on the input image, allowing better interpretability, and represents them using INR to improve effectiveness. Additionally, we predict spatially-varying combination coefficients to better handle spatially-similar (not invariant) defocus shapes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 476, + 420, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 420, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 420, + 489 + ], + "type": "text", + "content": "2.2. Related Techniques" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 495, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 615 + ], + "type": "text", + "content": "Deep inverse kernels Several studies have included inverse kernels in deblurring NNs, in addition to [47]. Xu et al. [52] and Ren et al. [38] presented the CNNs that adopt tensor product for emulating separable inverse kernels to perform NID. Dong et al. [9, 10] and Pronina et al. [33] performed NID via Wiener inverse filtering in deep feature spaces. However, all these methods are designed for spatially-invariant blur. Ren et al. [55, 39] proposed a recurrent NN to mimic inverse kernels for removing spatially-varying blur of dynamic scenes, but not for SIDD." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Scale-recurrent NNs for deblurring Scale-recurrent NNs are commonly used for image deblurring, e.g. [36, 28, 48, 12, 54, 8, 15]. These methods typically employ coarse-to-fine estimation which is unidirectional and thus cannot fully leverage information from the subsequent scales for the prediction at the current scale. In contrast, our proposed NN utilizes a bi-directional structure via bi-LSTMs [13] for joint coarse-to-fine and fine-to-coarse estimation. This is" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12601" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "particularly useful for SIDD, as defocus blur effects show strong self-similarity over scales. While bi-LSTM has been used in [61] for multi-scale video deblurring, it is only used for bi-directional temporal processing and the multi-scale processing is still one-way. Also note that the multi-scale NN proposed in [6] for motion deblurring is not restrictive to unidirectional processing, but not in a recurrent manner." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 158, + 287, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 158, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 47, + 158, + 287, + 194 + ], + "type": "text", + "content": "INR for image recovery INR has been applied to image recovery [42, 19, 50]. In comparison to these works, we use INR to parameterize inverse kernels, not images, for SIDD." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 203, + 277, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 203, + 277, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 203, + 277, + 216 + ], + "type": "text", + "content": "3. INR-Based Dictionary for Inverse Kernels" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 223, + 268, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 223, + 268, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 268, + 236 + ], + "type": "text", + "content": "3.1. Inverse Kernels for Spatially-Varying Blur" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "type": "text", + "content": "Consider a simplistic uniform blurring model defined by " + }, + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\pmb{y} = \\pmb{k} * \\pmb{x}" + }, + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "type": "text", + "content": " denotes convolution, " + }, + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\pmb{k}" + }, + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "type": "text", + "content": " denotes the PSF, and " + }, + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\pmb{y}, \\pmb{x}" + }, + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "type": "text", + "content": " denote the blurred image and its sharp correspondence, respectively. Let " + }, + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "type": "text", + "content": " denote the discrete Fourier transform (DFT). According to the well-known convolution theorem, we have " + }, + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\pmb{y}) = \\mathcal{F}(\\pmb{x}) \\odot \\mathcal{F}(\\pmb{k})" + }, + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "type": "text", + "content": " denotes entry-wise product. Then, one can define a (pesudo) inverse kernel " + }, + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\pmb{k}^{\\dagger}" + }, + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "type": "text", + "content": " by: for each " + }, + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "type": "inline_equation", + "content": "(\\omega_x, \\omega_y)" + }, + { + "bbox": [ + 46, + 241, + 287, + 337 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 342, + 287, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 342, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 58, + 342, + 287, + 366 + ], + "type": "interline_equation", + "content": "\\mathcal {F} \\left(\\boldsymbol {k} ^ {\\dagger}\\right) \\left[ \\omega_ {x}, \\omega_ {y} \\right] = \\left(\\mathcal {F} (\\boldsymbol {k}) [ s ]\\right) ^ {- 1}, \\text {i f} | \\mathcal {F} (\\boldsymbol {k}) \\left[ \\omega_ {x}, \\omega_ {y} \\right] | \\neq 0; \\tag {1}", + "image_path": "fcf1b4d915009c37f7b9ae30866bb921a8233415221c72e7092aaf90e84ace21.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 367, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 367, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 287, + 390 + ], + "type": "text", + "content": "and 0 otherwise. Then, for an invertible PSF with only nonzero DFT coefficients, we have" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 84, + 396, + 286, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 396, + 286, + 409 + ], + "spans": [ + { + "bbox": [ + 84, + 396, + 286, + 409 + ], + "type": "interline_equation", + "content": "\\mathcal {F} (\\boldsymbol {x}) = \\mathcal {F} \\left(\\boldsymbol {k} ^ {\\dagger}\\right) \\odot \\mathcal {F} (\\boldsymbol {y}) \\Longrightarrow \\boldsymbol {x} = \\boldsymbol {k} ^ {\\dagger} * \\boldsymbol {y}. \\tag {2}", + "image_path": "45b5ca9949767e4e47b94912460a58a18aaba8fb89d633a4226800d6b7b4bd66.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 415, + 287, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 439 + ], + "type": "text", + "content": "When the PSF " + }, + { + "bbox": [ + 46, + 415, + 287, + 439 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 415, + 287, + 439 + ], + "type": "text", + "content": " is non-invertible, the result " + }, + { + "bbox": [ + 46, + 415, + 287, + 439 + ], + "type": "inline_equation", + "content": "k^{\\dagger} * y" + }, + { + "bbox": [ + 46, + 415, + 287, + 439 + ], + "type": "text", + "content": " still provides a fair approximation to " + }, + { + "bbox": [ + 46, + 415, + 287, + 439 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 415, + 287, + 439 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 440, + 287, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 440, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 287, + 487 + ], + "type": "text", + "content": "One can extend (1) to handling spatially-varying blur. Let " + }, + { + "bbox": [ + 46, + 440, + 287, + 487 + ], + "type": "inline_equation", + "content": "\\mathbb{K}^{\\dagger} = \\{k_{i,j}^{\\dagger}\\}_{i,j}" + }, + { + "bbox": [ + 46, + 440, + 287, + 487 + ], + "type": "text", + "content": " denote a set of inverse kernels. Then, the image " + }, + { + "bbox": [ + 46, + 440, + 287, + 487 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 46, + 440, + 287, + 487 + ], + "type": "text", + "content": " with spatially-varying blur can be recovered by a spatially-varying deconvolution process " + }, + { + "bbox": [ + 46, + 440, + 287, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 440, + 287, + 487 + ], + "type": "text", + "content": " defined by" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 75, + 493, + 286, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 493, + 286, + 517 + ], + "spans": [ + { + "bbox": [ + 75, + 493, + 286, + 517 + ], + "type": "interline_equation", + "content": "\\mathcal {C} \\left(\\boldsymbol {y}; \\mathbb {K} ^ {\\dagger}\\right) [ i, j ] = \\sum_ {x, y} \\boldsymbol {k} _ {i, j} ^ {\\dagger} [ x, y ] \\boldsymbol {y} [ i - x, j - y ]. \\tag {3}", + "image_path": "7688674f8cde2f236cc57c7c4166fdd26c4a585d99e80be1753e3bf2c36f7554.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 525, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 598 + ], + "type": "text", + "content": "Both the shape and the size of " + }, + { + "bbox": [ + 46, + 525, + 287, + 598 + ], + "type": "inline_equation", + "content": "k_{i,j}^{\\dagger}" + }, + { + "bbox": [ + 46, + 525, + 287, + 598 + ], + "type": "text", + "content": " depend on the PSFs determined by scene depths and camera settings. When scene depths have significant variations, the sizes of the defocus PSFs also vary significantly, leading to large variations in the size of " + }, + { + "bbox": [ + 46, + 525, + 287, + 598 + ], + "type": "inline_equation", + "content": "k_{i,j}^{\\dagger}" + }, + { + "bbox": [ + 46, + 525, + 287, + 598 + ], + "type": "text", + "content": " and a large number of unknowns in " + }, + { + "bbox": [ + 46, + 525, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\mathbb{K}^{\\dagger}" + }, + { + "bbox": [ + 46, + 525, + 287, + 598 + ], + "type": "text", + "content": ". Without additional constraints on " + }, + { + "bbox": [ + 46, + 525, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\mathbb{K}^{\\dagger}" + }, + { + "bbox": [ + 46, + 525, + 287, + 598 + ], + "type": "text", + "content": ", overfitting can occur." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 605, + 268, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 268, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 268, + 617 + ], + "type": "text", + "content": "3.2. Multi-Scale Dictionary for Inverse Kernels" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 623, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 623, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 623, + 287, + 695 + ], + "type": "text", + "content": "To address the overfitting caused by excessive degrees of freedom in " + }, + { + "bbox": [ + 46, + 623, + 287, + 695 + ], + "type": "inline_equation", + "content": "\\mathbb{K}^{\\dagger}" + }, + { + "bbox": [ + 46, + 623, + 287, + 695 + ], + "type": "text", + "content": ", we parameterize the inverse kernels using a linear span over a dictionary with reduced number of atoms, resulting in inverse kernels that lie within a low-dimensional space. That is, we express an inverse kernel " + }, + { + "bbox": [ + 46, + 623, + 287, + 695 + ], + "type": "inline_equation", + "content": "\\pmb{k}^{\\dagger}" + }, + { + "bbox": [ + 46, + 623, + 287, + 695 + ], + "type": "text", + "content": " as a linear combination of atoms from a dictionary " + }, + { + "bbox": [ + 46, + 623, + 287, + 695 + ], + "type": "inline_equation", + "content": "\\pmb{V} = [v_{1},\\dots ,v_{N}]" + }, + { + "bbox": [ + 46, + 623, + 287, + 695 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 111, + 700, + 286, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 700, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 111, + 700, + 286, + 713 + ], + "type": "interline_equation", + "content": "\\boldsymbol {k} ^ {\\dagger} = \\omega_ {1} \\boldsymbol {v} _ {1} + \\dots + \\omega_ {N} \\boldsymbol {v} _ {N}, \\tag {4}", + "image_path": "2ae70bd4b6950838c0e4857ac3124726aa12ead7c1da4f141f5add056366d36d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 192 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\omega_{1},\\dots ,\\omega_{N}\\in \\mathbb{R}" + }, + { + "bbox": [ + 306, + 72, + 545, + 192 + ], + "type": "text", + "content": " are the weights. Additionally, inverse kernels may have a large size. To accurately approximate them using as few atoms as possible, we introduce a multi-scale structure to the dictionary atoms, which is motivated by the observation that when two kernels have the same shape but different sizes, the larger one can be effectively approximated by upsampling the smaller one. Consequently, the dictionary atoms for the corresponding inverse kernels can also be related through such an upsampling process, as stated in the following proposition." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 197, + 545, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 197, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 304, + 197, + 545, + 233 + ], + "type": "text", + "content": "Proposition 1. Let " + }, + { + "bbox": [ + 304, + 197, + 545, + 233 + ], + "type": "inline_equation", + "content": "\\pmb{k}^{\\dagger}" + }, + { + "bbox": [ + 304, + 197, + 545, + 233 + ], + "type": "text", + "content": " be an inverse kernel defined by (4). Let " + }, + { + "bbox": [ + 304, + 197, + 545, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_s" + }, + { + "bbox": [ + 304, + 197, + 545, + 233 + ], + "type": "text", + "content": " be a standard dyadic upsampling operator with a factor " + }, + { + "bbox": [ + 304, + 197, + 545, + 233 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 197, + 545, + 233 + ], + "type": "text", + "content": " via expansion. Then" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 340, + 240, + 544, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 240, + 544, + 255 + ], + "spans": [ + { + "bbox": [ + 340, + 240, + 544, + 255 + ], + "type": "interline_equation", + "content": "\\left(\\mathcal {U} _ {s} (\\boldsymbol {k})\\right) ^ {\\dagger} = \\omega_ {1} \\mathcal {U} _ {s} \\left(\\boldsymbol {v} _ {1}\\right) + \\dots + \\omega_ {N} \\mathcal {U} _ {s} \\left(\\boldsymbol {v} _ {N}\\right). \\tag {5}", + "image_path": "63508435f3cbb04aa4dee2008c8d4049b335f22fce92cc95ef8bbca6e96f9fe0.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 262, + 498, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 498, + 274 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 498, + 274 + ], + "type": "text", + "content": "Proof. See supplemental material for the proof." + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 535, + 262, + 545, + 271 + ], + "blocks": [ + { + "bbox": [ + 535, + 262, + 545, + 271 + ], + "lines": [ + { + "bbox": [ + 535, + 262, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 535, + 262, + 545, + 271 + ], + "type": "image", + "image_path": "84ccd284f25d6e8ba9319f23dd4670275be72eb35bc144e09f67ee8693263c78.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "type": "text", + "content": "Proposition 1 implies that two PFSs with the same shape but different sizes can share the same dictionary for their inverse kernels up to a spatial scaling factor. For defocus PSFs within an image, they typically have similar shapes related to camera settings, as observed in [47]. However, their sizes will vary with the corresponding scene depths. Thus, we represent the inverse kernels of PSFs using different upsampled versions of a dictionary with small atoms. Such a multi-scale structure provides a physics-based implicit prior on the PSFs within the same image and leads to a more compact dictionary for representing inverse kernels of varying sizes. Specifically, we define a multi-scale dictionary composed by the sets (sub-dictionaries) " + }, + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\{V^r\\}_{r=1}^R" + }, + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 442, + 544, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 442, + 544, + 456 + ], + "spans": [ + { + "bbox": [ + 321, + 442, + 544, + 456 + ], + "type": "interline_equation", + "content": "\\boldsymbol {V} ^ {r} = \\left[ \\boldsymbol {v} _ {1} ^ {r}, \\dots , \\boldsymbol {v} _ {N} ^ {r} \\right] \\subset \\mathbb {R} ^ {M _ {r} \\times M _ {r}}, \\boldsymbol {v} _ {n} ^ {r} = \\mathcal {U} _ {S _ {r}} (\\boldsymbol {v} _ {n}), \\tag {6}", + "image_path": "156560f09298b43ab07a2f04256e0e134f08cb4bc4548f6b5d91bbcc70dea994.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 462, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 462, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 462, + 545, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 462, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}_n^r" + }, + { + "bbox": [ + 306, + 462, + 545, + 510 + ], + "type": "text", + "content": " denotes the atom at the " + }, + { + "bbox": [ + 306, + 462, + 545, + 510 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 306, + 462, + 545, + 510 + ], + "type": "text", + "content": "-th scale, " + }, + { + "bbox": [ + 306, + 462, + 545, + 510 + ], + "type": "inline_equation", + "content": "S_r" + }, + { + "bbox": [ + 306, + 462, + 545, + 510 + ], + "type": "text", + "content": " denotes an upsampling factor that determines the size parameter " + }, + { + "bbox": [ + 306, + 462, + 545, + 510 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 306, + 462, + 545, + 510 + ], + "type": "text", + "content": " of the atom. As " + }, + { + "bbox": [ + 306, + 462, + 545, + 510 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 306, + 462, + 545, + 510 + ], + "type": "text", + "content": " increases, " + }, + { + "bbox": [ + 306, + 462, + 545, + 510 + ], + "type": "inline_equation", + "content": "S_r" + }, + { + "bbox": [ + 306, + 462, + 545, + 510 + ], + "type": "text", + "content": " becomes larger and " + }, + { + "bbox": [ + 306, + 462, + 545, + 510 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 306, + 462, + 545, + 510 + ], + "type": "text", + "content": " gets smaller. The per-pixel inverse kernels are then expressed as" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 515, + 544, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 515, + 544, + 531 + ], + "spans": [ + { + "bbox": [ + 312, + 515, + 544, + 531 + ], + "type": "interline_equation", + "content": "\\boldsymbol {k} _ {i, j} ^ {\\dagger} = \\omega_ {1, i, j} \\boldsymbol {v} _ {1} ^ {r _ {i, j}} + \\dots + \\omega_ {N, i, j} \\boldsymbol {v} _ {N} ^ {r _ {i, j}} \\in \\mathbb {R} ^ {M _ {r _ {i, j}} \\times M _ {r _ {i, j}}}, \\tag {7}", + "image_path": "388390c67eadfbb5431be17fe255a50563f1bed4b9cd99276fc855c424aa049f.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 538, + 544, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 544, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 544, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 538, + 544, + 597 + ], + "type": "inline_equation", + "content": "r_{i,j}" + }, + { + "bbox": [ + 306, + 538, + 544, + 597 + ], + "type": "text", + "content": " denotes the scale factor (index) related to " + }, + { + "bbox": [ + 306, + 538, + 544, + 597 + ], + "type": "inline_equation", + "content": "\\pmb{k}_{i,j}^{\\dagger}" + }, + { + "bbox": [ + 306, + 538, + 544, + 597 + ], + "type": "text", + "content": ". Note that the atoms within the same sub-dictionary have a uniform size, while their sizes vary across different sub-dictionaries. That is, each sub-dictionary is utilized to represent inverse kernels of a specific size." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 598, + 544, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 544, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 544, + 635 + ], + "type": "text", + "content": "Using the linear representation (7) and the fact that " + }, + { + "bbox": [ + 306, + 598, + 544, + 635 + ], + "type": "inline_equation", + "content": "\\pmb{v}_n^{r_{i,j}}*" + }, + { + "bbox": [ + 306, + 598, + 544, + 635 + ], + "type": "inline_equation", + "content": "\\pmb {y} = \\sum_r\\delta (r - r_{i,j})\\pmb{v}_n^r *\\pmb {y}" + }, + { + "bbox": [ + 306, + 598, + 544, + 635 + ], + "type": "text", + "content": " with a Dirac delta function " + }, + { + "bbox": [ + 306, + 598, + 544, + 635 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 306, + 598, + 544, + 635 + ], + "type": "text", + "content": " , we can rewrite Eq. (3) as" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 323, + 641, + 544, + 709 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 641, + 544, + 709 + ], + "spans": [ + { + "bbox": [ + 323, + 641, + 544, + 709 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {C} (\\boldsymbol {y}) = \\sum_ {i, j} \\mathbf {1} _ {i, j} \\odot \\left(\\sum_ {n = 1} ^ {N} \\omega_ {n, i, j} \\boldsymbol {v} _ {n} ^ {r _ {i, j}} * \\boldsymbol {y}\\right) (8) \\\\ = \\sum_ {i, j} \\mathbf {1} _ {i, j} \\odot \\left(\\sum_ {n = 1} ^ {N} \\sum_ {r = 1} ^ {R} \\mu_ {r, i, j} \\omega_ {n, i, j} \\boldsymbol {v} _ {n} ^ {r} * \\boldsymbol {y}\\right), (9) \\\\ \\end{array}", + "image_path": "92d8d29c823363f1a5baef5134af1f10f5a86a13bfb56035c9ec5121828138ab.jpg" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12602" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_{i,j}" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " denotes a mask with 1 for the " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "-th entry being 1 and 0 otherwise, and " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\mu_{r,i,j} = \\delta (r - r_{i,j})" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": ". We can then perform spatially-varying inverse filtering by convolving each atom " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{v}_n^r" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " in the dictionary with the input image " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": ", and summing the resulting images by the spatially-varying weights " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\mu_{s,i,j}\\omega_{n,i,j}" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": ". The weights " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\mu_{s,i,j}" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\omega_{n,i,j}" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " control the size (adaptive to scene depth) and shape (adaptive to image) of the corresponding inverse kernel, respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 174, + 287, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 174, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 174, + 287, + 186 + ], + "type": "text", + "content": "3.3. INR Models for Dictionaries of Inverse Kernels" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 191, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 191, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 191, + 287, + 335 + ], + "type": "text", + "content": "The dictionary " + }, + { + "bbox": [ + 46, + 191, + 287, + 335 + ], + "type": "inline_equation", + "content": "\\{V_r\\}_{r=1}^R" + }, + { + "bbox": [ + 46, + 191, + 287, + 335 + ], + "type": "text", + "content": " used in (8) is generated by the base atoms " + }, + { + "bbox": [ + 46, + 191, + 287, + 335 + ], + "type": "inline_equation", + "content": "\\{v_j\\}_{j=1}^N" + }, + { + "bbox": [ + 46, + 191, + 287, + 335 + ], + "type": "text", + "content": " of the smallest size. However, simply upsampling these base atoms cannot cover all high frequencies, i.e., certain high-pass filters of large sizes may not be well approximated by the linear span of these atoms obtained by plain upsampling. To address this limitation, we use INR to re-parameterize the multi-scale dictionary. INR provides a more expressive way to generate multi-scale atoms that cover more frequencies than plain upsampling. Additionally, INR provides implicit regularization to alleviate overfitting by preferring image structures over random noise, as shown in [42, 19, 50]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 336, + 182, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 336, + 182, + 349 + ], + "spans": [ + { + "bbox": [ + 59, + 336, + 182, + 349 + ], + "type": "text", + "content": "Specifically, we express " + }, + { + "bbox": [ + 59, + 336, + 182, + 349 + ], + "type": "inline_equation", + "content": "\\pmb{v}_n^r" + }, + { + "bbox": [ + 59, + 336, + 182, + 349 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 354, + 285, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 285, + 368 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 285, + 368 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {n} ^ {r} [ x, y ] = \\Phi_ {n} (x, y), [ x, y ] \\in [ 1, \\dots , M _ {r} ] \\times [ 1, \\dots , M _ {r} ],", + "image_path": "3e6eee74960986f6ef1a99ccb8b7a1fdcdd885103306d9004f10ea44d2c2e015.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "inline_equation", + "content": "[x,y]" + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "text", + "content": " denotes a spatial coordinate, and " + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "text", + "content": " is an INR model implemented by a compact multi-layer perceptron (MLP) of a small size. The INR model maps spatial coordinates to a continuous function that implicitly represents the kernel atom. This function can be evaluated at any point in space to interpolate a kernel atom of an arbitrary size, forming a dictionary of inverse kernels with arbitrary scales." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 465, + 266, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 465, + 266, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 465, + 266, + 479 + ], + "type": "text", + "content": "4. Implicit Neural Inverse Kernel Network" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "type": "text", + "content": "Our proposed NN called INIKNet consists of recurrent deblurring stages over different scales for exploiting cross-scale contextual information. Each deblurring stage predicts and applies inverse kernels to deblur the input image. Concretely, each stage consists of two steps: (i) given an input image " + }, + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "type": "text", + "content": ", inferring two weight tensors " + }, + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu},\\pmb{\\omega}" + }, + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "type": "text", + "content": " that contain the weights " + }, + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\{\\mu_{r,i,j}\\}_{r,i,j}" + }, + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\{\\omega_{n,i,j}\\}_{n,i,j}" + }, + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "type": "text", + "content": " respectively; and (ii) applying inverse filtering defined by (8) to deblurring " + }, + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 587, + 241, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 241, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 241, + 601 + ], + "type": "text", + "content": "4.1. Duplex Scale-Recurrent Framework" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Current scale-recurrent deblurring NNs, e.g. [36, 28, 48, 12, 54, 8, 15], typically use an LSTM-based coarse-to-fine framework. The deblurred image from the current (coarser) scale is attached to the input of the next (finer) scale, resulting in unidirectional feature flow that only considers coarser scales when inferring at finer scales. This approach cannot fully leverage the additional information available at finer scales for deblurring on coarser scales, especially when the blur effects show strong similarity over space and scale." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": "To address this limitation, we propose a duplex multiscale processing framework, which leverages bi-LSTM [13] as the sub-NNs to predict " + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": ". It improves predictions at coarser scales by incorporating additional information from finer scales through the feature flow in the fine-to-coarse pass. This refined information is then utilized to further improve predictions on finer scales via the feature flow through the coarse-to-fine pass. As a result, the prediction at the current scale benefits from the information extracted from both the previous and subsequent scales." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 192, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 545, + 312 + ], + "type": "text", + "content": "As shown in Fig. 1, our INIKNet first downsamples the input image to multi-scale versions. For each scale, a U-Net shared across scales is first applied for extracting spatial features from the input. The extracted features are then fed to two bi-LSTM sub-NNs so as to exploit the dependencies among scales for feature refinement. The outputs of two bi-LSTMs sub-NNs are then used to predict " + }, + { + "bbox": [ + 304, + 192, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 192, + 545, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 192, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 192, + 545, + 312 + ], + "type": "text", + "content": " via an " + }, + { + "bbox": [ + 304, + 192, + 545, + 312 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 304, + 192, + 545, + 312 + ], + "type": "text", + "content": " convolutional layer, respectively. Afterward, the predicted weights are used by an Inverse Filtering Module (IFM) to perform deblurring using (8) to output the result." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 312, + 546, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 546, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 546, + 456 + ], + "type": "text", + "content": "Our INIKNet differs from existing coarse-to-fine NNs in that it does not concatenate the output image from the current scale to the input of the next scale, as this operation would not only make the feature flow unidirectional, but also prevent the current-scale processing from starting, until the deblurring is fully completed at the previous scale. By removing this operation, the U-Net and bi-LSTM can start running earlier, without waiting for the deblurred images from previous scales. Moreover, by cancelling the operation, the IFM only needs to be called at the original image scale during inference, even though it is still required during training to calculate the multi-scale loss." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 464, + 437, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 437, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 437, + 478 + ], + "type": "text", + "content": "4.2. Details of Key Modules" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "text", + "content": "Bi-LSTM sub-NNs Recall that the weights " + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "text", + "content": " correspond to two distinct properties of inverse kernels, namely size and shape, respectively. Their dependency on image scales also differs significantly. Therefore, rather than use a single bi-LSTM sub-NN to predict " + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 483, + 545, + 651 + ], + "type": "text", + "content": " jointly, we use two parallel bi-LSTM sub-NNs for separate predictions. Both sub-NNs share the same structure, consisting of convolutional LSTM cells [45] arranged in two paths for fine-to-coarse and coarse-to-fine estimations in opposite directions. Each LSTM cell takes a cell state and a hidden state as input and outputs updated states. The input cell state is defined as the features extracted by the U-Net, while the input hidden state for the next scale is set to the output one from the previous scale, forming a cross-scale connection." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 546, + 712 + ], + "type": "text", + "content": "Weight predictors To convert the output features from the bi-LSTM sub-NNs to the weights " + }, + { + "bbox": [ + 304, + 654, + 546, + 712 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 654, + 546, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 654, + 546, + 712 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 654, + 546, + 712 + ], + "type": "text", + "content": ", two " + }, + { + "bbox": [ + 304, + 654, + 546, + 712 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 654, + 546, + 712 + ], + "type": "text", + "content": " convolutional layers are applied, respectively. For predicting " + }, + { + "bbox": [ + 304, + 654, + 546, + 712 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 654, + 546, + 712 + ], + "type": "text", + "content": ", the Softmax function is used to impose non-negativity and " + }, + { + "bbox": [ + 304, + 654, + 546, + 712 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 304, + 654, + 546, + 712 + ], + "type": "text", + "content": "-normalization constraints. Such a soft relaxation of" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12603" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 74, + 293, + 353 + ], + "blocks": [ + { + "bbox": [ + 51, + 74, + 293, + 353 + ], + "lines": [ + { + "bbox": [ + 51, + 74, + 293, + 353 + ], + "spans": [ + { + "bbox": [ + 51, + 74, + 293, + 353 + ], + "type": "image", + "image_path": "d18ae44616a483e33cf54e273e30b5a81a94276c9dd19a50eb96e85fca5e4260.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 356, + 407, + 370 + ], + "lines": [ + { + "bbox": [ + 184, + 356, + 407, + 370 + ], + "spans": [ + { + "bbox": [ + 184, + 356, + 407, + 370 + ], + "type": "text", + "content": "Figure 1: Architecture of proposed INIKNet for SIDD." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 296, + 73, + 542, + 353 + ], + "blocks": [ + { + "bbox": [ + 296, + 73, + 542, + 353 + ], + "lines": [ + { + "bbox": [ + 296, + 73, + 542, + 353 + ], + "spans": [ + { + "bbox": [ + 296, + 73, + 542, + 353 + ], + "type": "image", + "image_path": "a65389a81c005a72bc9106fb2b811f9eb8ee4eae426d77a2859443fe518586bc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 391, + 287, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 287, + 452 + ], + "type": "text", + "content": "weights " + }, + { + "bbox": [ + 46, + 391, + 287, + 452 + ], + "type": "inline_equation", + "content": "\\mu_{r,i,j}" + }, + { + "bbox": [ + 46, + 391, + 287, + 452 + ], + "type": "text", + "content": " from 0,1 to [0,1] not only makes training easier, but also improves the representation accuracy of inverse kernels by including the use of additional atoms. This can be useful when the inverse kernel sizes fall between the gaps among the predefined non-consecutive atom sizes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 453, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 550 + ], + "type": "text", + "content": "INR models of kernel atoms The MLPs used for INR have a structure similar to those in [46]. They comprise three fully-connected (FC) layers separated by a Sine activation function with a learnable frequency parameter that controls frequency content. The 2D coordinates are first mapped to a high-dimensional space and then projected to single values. One difference from [46] is that we first transform spatial coordinates to polar coordinates for slight improvement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 552, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 613 + ], + "type": "text", + "content": "IFM for deblurring The IFM first calls the learned MLPs to generate all kernel atoms " + }, + { + "bbox": [ + 46, + 552, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\{\\pmb{v}_n^r\\}_{n,r}" + }, + { + "bbox": [ + 46, + 552, + 287, + 613 + ], + "type": "text", + "content": " and then calculates the feature tensors " + }, + { + "bbox": [ + 46, + 552, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\pmb{f}[n,r] = \\pmb{v}_n^r*\\pmb{y}" + }, + { + "bbox": [ + 46, + 552, + 287, + 613 + ], + "type": "text", + "content": ". Afterward, Eq. (8) is implemented by simple entry-wise product and summation operations on the tensors " + }, + { + "bbox": [ + 46, + 552, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\pmb{f},\\pmb{\\mu},\\pmb{\\omega}" + }, + { + "bbox": [ + 46, + 552, + 287, + 613 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 622, + 135, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 135, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 135, + 635 + ], + "type": "text", + "content": "4.3. Training Loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "In addition to the " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " loss " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " defined by mean squared error (MSE), two loss functions often seen in image recovery are also used for training, including the frequency-domain loss " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FD}}" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " [6] defined as the " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " distance in the frequency domain, and the Learned Perceptual Image Patch Similarity (LPIPS) loss " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LPIPS}}" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " [57] predicted by a pre" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 391, + 545, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 391, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 305, + 391, + 545, + 428 + ], + "type": "text", + "content": "trained NN. These three loss functions measure deblurring quality in the spatial, frequency and feature (perceptual) domains, respectively. The total loss is then given by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 348, + 437, + 545, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 437, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 348, + 437, + 545, + 464 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {p} \\mathcal {L} _ {\\mathrm {M S E}} ^ {(p)} + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {F D}} ^ {(p)} + \\lambda_ {2} \\mathcal {L} _ {\\mathrm {L P I P S}} ^ {(p)}, \\tag {10}", + "image_path": "7428e0cc71cce1d96dc25005c61a4a75d4e2f068daf05acf505adfa69785a4db.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 473, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 473, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 473, + 545, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 473, + 545, + 510 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 473, + 545, + 510 + ], + "type": "text", + "content": " denotes the index of the " + }, + { + "bbox": [ + 305, + 473, + 545, + 510 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 473, + 545, + 510 + ], + "type": "text", + "content": "th scale, and " + }, + { + "bbox": [ + 305, + 473, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2" + }, + { + "bbox": [ + 305, + 473, + 545, + 510 + ], + "type": "text", + "content": " are empirically set to 0.1. Ground-truth images are downsampled accordingly for the supervision at different scales." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 522, + 388, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 522, + 388, + 536 + ], + "spans": [ + { + "bbox": [ + 306, + 522, + 388, + 536 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 543, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 545, + 674 + ], + "type": "text", + "content": "Datasets and metrics Following [2, 41], we train models on two widely-used datasets, DPDD [2] and LFDOF [41], respectively. To assess their generalization performance, we also evaluate these models on two more datasets, RTF [7] and RealDOF [21]. Three quality metrics are used for quantitative evaluation: Peak Signal to Noise Ratio (PSNR, measured in dB), Structural SIMilarity index (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [57]. The model complexity is measured by both the number of parameters and the inference time on a " + }, + { + "bbox": [ + 304, + 543, + 545, + 674 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 304, + 543, + 545, + 674 + ], + "type": "text", + "content": " image using an NVIDIA 2080Ti RTX GPU." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "content": "Implementation details Throughout the experiments, we employ 8 scales and 10 INR-based atoms per scale: " + }, + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "inline_equation", + "content": "R = 8" + }, + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "inline_equation", + "content": "N = 10" + }, + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "content": ". We use atoms of sizes " + }, + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 1, 3 \\times 3, \\dots, 15 \\times 15" + }, + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "text", + "content": "12604" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "type": "text", + "content": " atom is fixed as a delta kernel for modeling in-focus regions. The frequency parameters of the sine activation functions in all INR models are initialized by random sampling from [2, 16]. We call Adam for training, with " + }, + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "type": "inline_equation", + "content": "4 \\times 10^{5}" + }, + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "type": "text", + "content": " iterations and a batch size of 4. We start with an initial learning rate of " + }, + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "type": "text", + "content": " and halve it every " + }, + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "type": "inline_equation", + "content": "10^{5}" + }, + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "type": "text", + "content": " iterations. Random flipping, rotation, and cropping (to " + }, + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "type": "text", + "content": " pixels) are performed for training data augmentation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 175, + 185, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 175, + 185, + 188 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 185, + 188 + ], + "type": "text", + "content": "5.1. Performance Evaluation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 194, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 287, + 350 + ], + "type": "text", + "content": "Compared methods We select 9 NN-based SIDD methods for comparison, including the DPDNet [2], AIFNet [41], IFANet [21], KPAC [47], GKMNet [36], MDP [1], DRBNet [40], Restormer [53], and MPRNet [26] trained in [23] with a misalignment-robust scheme. The experimental results of the compared methods are reported either from their original papers or obtained by using their released pretrained models and codes. AIFNet is a model trained on LFDOF with the additional use of SYNDOF for its DME step. To compare its performance on DPDD, we retrain it using both DPDD and SYNDOF. DRBNet is a model trained on LFDOF and fine-tuned on DPDD. For a fair comparison, we retrain it on both datasets, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 352, + 288, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 288, + 568 + ], + "type": "text", + "content": "Evaluation on DPDD and LFDOF Table 1 presents the quantitative results on DPDD and LFDOF. Our INIKNet performs the best in PSNR on DPDD and in all three metrics on LFDOF, despite using a lightweight model whose parameter number is the second smallest and around " + }, + { + "bbox": [ + 46, + 352, + 288, + 568 + ], + "type": "inline_equation", + "content": "13.2\\%" + }, + { + "bbox": [ + 46, + 352, + 288, + 568 + ], + "type": "text", + "content": " of that of Restormer, a general restoration model. Compared to the latest SIDD-specified MRPNet model [23], INIKNet achieves over " + }, + { + "bbox": [ + 46, + 352, + 288, + 568 + ], + "type": "inline_equation", + "content": "0.3\\mathrm{dB}" + }, + { + "bbox": [ + 46, + 352, + 288, + 568 + ], + "type": "text", + "content": " PSNR improvement on DPDD while having only about " + }, + { + "bbox": [ + 46, + 352, + 288, + 568 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 352, + 288, + 568 + ], + "type": "text", + "content": " parameters to learn. Similar results are observed in the comparison with DRBNet, indicating that the performance gain of INIKNet comes from the design of NN structure rather than the increased model size. Though INIKNet is slightly larger than GKMNet, the smallest model, it achieves noticeably better performance. Compared to KPAC that uses dilated convolutions to mimic inverse kernels, INIKNet performs significantly better. In conclusion, the superiority and compactness of INIKNet demonstrate the effectiveness of its architecture design." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": "Generalization analysis on RealDOF and RTF The results of the DPDD/LFDOF-trained models evaluated on RealDOF and RTF are listed in Table 2. Our DPDD-trained INIKNet shows good generalization to different datasets and achieves the best performance on both RealDOF and RTF in all three metrics. The LFDOF-trained INIKNet is the top performer on RealDOF and achieves the best SSIM score on RTF, as well as the second-best scores in PSNR and LPIPS. Though AIFNet performs better in PSNR and LPIPS, it requires an additional dataset for DME and is much larger than INIKNet. In summary, INIKNet exhibits superior generalizability over other compared models." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 71, + 545, + 196 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 545, + 196 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 545, + 196 + ], + "type": "table", + "html": "
ModelDPDDLFDOF#Params (106)Time (sec.)
PSNRSSIMLPIPSPSNRSSIMLPIPS
DPDNet24.3480.7470.277---32.251.430
AIFNet24.2130.7420.30929.6770.8840.20241.550.276
IFANet25.3660.7890.21729.7870.8720.15610.480.063
KPAC25.2210.7740.22628.9420.8570.1742.060.037
GKMNet25.4680.7890.21929.0810.8670.1711.410.129
MDP25.3470.7630.26828.0690.8340.18546.863.317
DRBNet25.4850.7920.25430.2530.8830.14711.690.085
MPRNet25.7300.7920.232---20.101.161
Restormer25.9800.8110.178---26.100.672
INIKNet26.0550.8030.18530.2930.8860.1321.980.228
", + "image_path": "4e4f2879d70f81dfcd3d2ea2f8fa1c731d07365399f7d20d7fb6436496488b83.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 310, + 237, + 542, + 437 + ], + "blocks": [ + { + "bbox": [ + 305, + 200, + 545, + 224 + ], + "lines": [ + { + "bbox": [ + 305, + 200, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 305, + 200, + 545, + 224 + ], + "type": "text", + "content": "Table 1: Quantitative results on DPDD/LFDOF. Bold for best performers and underline for second-best performers." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 237, + 542, + 437 + ], + "lines": [ + { + "bbox": [ + 310, + 237, + 542, + 437 + ], + "spans": [ + { + "bbox": [ + 310, + 237, + 542, + 437 + ], + "type": "table", + "html": "
ModelRealDOFRTF#Params (106)
PSNRSSIMLPIPSPSNRSSIMLPIPS
Trained on DPDDDPDNet22.8700.6700.42523.6080.5910.29632.25
AIFNet23.0930.6800.41324.0410.7580.28941.55
MDP23.5000.6810.44424.0120.7380.31246.86
KPAC23.9750.7620.33824.6180.7770.2362.06
IFANet24.7120.7480.30624.9240.8010.22710.48
GKMNet24.2540.7320.39224.9700.7890.2611.41
DRBNet24.8840.7510.37624.4630.7730.31111.69
MPRNet24.5410.7360.33924.5880.7880.30420.10
Restormer25.0910.7620.28524.2120.8210.22426.10
INIKNet25.2310.7650.28725.4500.8340.2151.98
Trained on LFDOFIFANet22.5040.6690.48326.4370.8380.23810.48
KPAC22.5500.6710.45725.9590.8030.2302.06
AIFNet22.6230.6670.46127.5520.8820.17641.55
GKMNet23.6090.7210.40826.9850.8560.2461.41
MDP22.7260.6800.45325.5800.8090.22846.86
DRBNet22.9100.6910.43726.7170.8530.20011.69
INIKNet23.8100.7240.35627.4010.8850.1791.98
", + "image_path": "47a3d551c9f410106060ab10e255fecf72ae676dc94c160a41bf9ddd8930a6d2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 441, + 545, + 465 + ], + "lines": [ + { + "bbox": [ + 305, + 441, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 545, + 465 + ], + "type": "text", + "content": "Table 2: Quantitative results on RealDOF/RTF. Bold for best performers and underline for second-best performers." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 478, + 545, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 478, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 304, + 478, + 545, + 587 + ], + "type": "text", + "content": "Qualitative comparison We compare INIKNet with some SIDD-specified competitors on DPDD and LFDOF via visual inspection. See Fig. 2 for some deblurred images on four datasets, where complex textures such as those on wire fence and branches are successfully recovered by INIKNet. See also Fig. 3 for some results on the CUHK dataset [43]. While most methods fail under large blur, ours yields more visually satisfying results. These findings demonstrate that our INIKNet performs well in terms of visual quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 596, + 469, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 596, + 469, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 596, + 469, + 609 + ], + "type": "text", + "content": "5.2. Ablation Studies and Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 615, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 615, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 305, + 615, + 545, + 639 + ], + "type": "text", + "content": "Ablation studies We construct several variants of INIKNet and list their results on two datasets in Table 3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": "Non-MS Dict: We drop the multi-scale structure of the dictionary and instead use an individual INR model for each dictionary atom at each scale (thus yielding a larger model size). A noticeable performance degradation occurs. This is because the multi-scale structure of the INR-based dictionary imposes structural regularization to reduce overfitting." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12605" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 324, + 393 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 324, + 393 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 324, + 393 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 324, + 393 + ], + "type": "image", + "image_path": "d738c76f004176285f2135664e6a221410f8981637ba70da43a0978e2c1dc83e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 399, + 525, + 412 + ], + "lines": [ + { + "bbox": [ + 67, + 399, + 525, + 412 + ], + "spans": [ + { + "bbox": [ + 67, + 399, + 525, + 412 + ], + "type": "text", + "content": "Figure 2: Results on using models trained on DPDD (left part) and LFDOF (right part). Zoom-in for better views." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 334, + 68, + 539, + 393 + ], + "blocks": [ + { + "bbox": [ + 334, + 68, + 539, + 393 + ], + "lines": [ + { + "bbox": [ + 334, + 68, + 539, + 393 + ], + "spans": [ + { + "bbox": [ + 334, + 68, + 539, + 393 + ], + "type": "image", + "image_path": "259a5b098bb38f3343912588c4db9a3572482b1e03416ac5a055b6a1ff2dfd47.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 52, + 430, + 282, + 501 + ], + "blocks": [ + { + "bbox": [ + 52, + 430, + 282, + 501 + ], + "lines": [ + { + "bbox": [ + 52, + 430, + 282, + 501 + ], + "spans": [ + { + "bbox": [ + 52, + 430, + 282, + 501 + ], + "type": "table", + "html": "
ModelRealDOFRTF
PSNRSSIMLPIPSPSNRSSIMLPIPS
Non-MS Dict24.9160.7490.31025.1360.8090.254
Non-INR Dict25.0580.7540.30325.2370.8190.226
Unidirectional25.0990.7530.29425.3050.8240.227
Original INIKNet25.2310.7650.28725.4500.8340.215
", + "image_path": "bd3368dbe2335aa8263df3c0540c6b5bf50a73f57ad5efee0d42d4cdba24be8b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 504, + 277, + 515 + ], + "lines": [ + { + "bbox": [ + 56, + 504, + 277, + 515 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 277, + 515 + ], + "type": "text", + "content": "Table 3: Results of baseline models trained on DPDD." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 529, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 613 + ], + "type": "text", + "content": "Non-INR Dict: When the INR-based dictionaries are replaced by the plain ones, where each base atom " + }, + { + "bbox": [ + 46, + 529, + 287, + 613 + ], + "type": "inline_equation", + "content": "v_{n}" + }, + { + "bbox": [ + 46, + 529, + 287, + 613 + ], + "type": "text", + "content": " is defined as a learnable matrix and bi-linearly upsampled to form the multi-scale dictionary, the performance of INIKNet decreases. This is probably because the frequency content for the atoms generated by plain upsampling are limited, so do their generated inverse kernels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "type": "text", + "content": "Unidirectional: To simulate an unidirectional scale recurrent mechanism (e.g. [36]), we replace the bi-LSTM with double unidirectional LSTMs (for keeping the model size). The output from the last scale is attached to the input of the current scale for additional information. Such a scheme for coefficient prediction results in some performance degradation, highlighting the effectiveness of our proposed duplex scale-recurrent framework." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 433, + 545, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 545, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 545, + 648 + ], + "type": "text", + "content": "Visualization of coefficients and atoms The coefficient maps " + }, + { + "bbox": [ + 304, + 433, + 545, + 648 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 433, + 545, + 648 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 433, + 545, + 648 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 433, + 545, + 648 + ], + "type": "text", + "content": " are visualized in Fig. 4. We can see that the regions with larger blur amount tend to have larger coefficients " + }, + { + "bbox": [ + 304, + 433, + 545, + 648 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 433, + 545, + 648 + ], + "type": "text", + "content": " assigned to larger inverse kernels, and vice versa. In comparison, the coefficients " + }, + { + "bbox": [ + 304, + 433, + 545, + 648 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 433, + 545, + 648 + ], + "type": "text", + "content": " are more spatially similar (gray), consistent with the prior of spatially-similar shapes of inverse kernels. We use Fig. 5 to visualize some learned (non-)INR-based atoms in both the spatial and Fourier domain. The INR-based atoms within the same scale contain diverse patterns with varying frequencies and orientations, and meanwhile they have similar shapes across scales. Compared to the non-INR-based ones, the INR-based atoms encode richer spatial structures and frequency contents. Interestingly, the INR-based atoms exhibit strong isotropy, which is likely due to the isotropy of defocus PSFs (so do their inverse kernels). All these properties of the INR-based atoms enable our INIKNet to effectively represent and predict inverse kernels for spatially-varying defocus blur." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 657, + 379, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 379, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 379, + 670 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "This paper presented an interpretable end-to-end SIDD approach, which learns a duplex scale-recurrent NN to predict INR-driven dictionary-parameterized inverse kernels of" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12606" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 545, + 390 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 390 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 390 + ], + "type": "image", + "image_path": "f453722cad9142760713cb963b3cd48a9a288922863cd894392a109d48c3a1a5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 396, + 530, + 408 + ], + "lines": [ + { + "bbox": [ + 61, + 396, + 530, + 408 + ], + "spans": [ + { + "bbox": [ + 61, + 396, + 530, + 408 + ], + "type": "text", + "content": "Figure 3: Results on CUHK dataset (without ground-truths) using models trained on DPD. Zoom-in for better views." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 419, + 544, + 526 + ], + "blocks": [ + { + "bbox": [ + 52, + 419, + 544, + 526 + ], + "lines": [ + { + "bbox": [ + 52, + 419, + 544, + 526 + ], + "spans": [ + { + "bbox": [ + 52, + 419, + 544, + 526 + ], + "type": "image", + "image_path": "4ec66da483cb1d5e4b7c54ccb4f098f49ad562d4a69bc32559ee44057a86a917.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 533, + 392, + 545 + ], + "lines": [ + { + "bbox": [ + 200, + 533, + 392, + 545 + ], + "spans": [ + { + "bbox": [ + 200, + 533, + 392, + 545 + ], + "type": "text", + "content": "Figure 4: Visualization of " + }, + { + "bbox": [ + 200, + 533, + 392, + 545 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 200, + 533, + 392, + 545 + ], + "type": "text", + "content": " maps and " + }, + { + "bbox": [ + 200, + 533, + 392, + 545 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 200, + 533, + 392, + 545 + ], + "type": "text", + "content": " maps." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 52, + 556, + 541, + 639 + ], + "blocks": [ + { + "bbox": [ + 52, + 556, + 541, + 639 + ], + "lines": [ + { + "bbox": [ + 52, + 556, + 541, + 639 + ], + "spans": [ + { + "bbox": [ + 52, + 556, + 541, + 639 + ], + "type": "image", + "image_path": "2c75e98d78c435e3ac290ac63b8e0a99078eccb3c236168faaf0ec7d518ab6ac.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 643, + 517, + 656 + ], + "lines": [ + { + "bbox": [ + 73, + 643, + 517, + 656 + ], + "spans": [ + { + "bbox": [ + 73, + 643, + 517, + 656 + ], + "type": "text", + "content": "Figure 5: Visualization of bilinearly upsampled (non-INR-based) atoms and INR-based atoms in two domains." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": "defocus blur and applies them to deblur images. Our proposed approach introduced effective structural regularization on inverse kernels for good generalization, which is" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "based on the shape similarity of defocus PSFs over an image. The resulting effective and light-weight model has demonstrated its advantages by extensive experiments." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12607" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Abdullah Abuolaim, Mahmoud Afifi, and Michael S. Brown. Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1231-1239, 2022. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 190 + ], + "type": "text", + "content": "[2] Abdullah Abuolaim and Michael S. Brown. Defocus deblurring using dual-pixel data. In Proceedings of the European Conference on Computer Vision, pages 111-126. Springer, 2020. 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "type": "text", + "content": "[3] Yosuke Bando and Tomoyuki Nishita. Towards digital refocusing from a single photograph. In Proceedings of the Pacific Conference on Computer Graphics and Applications, pages 363-372. IEEE, 2007. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "text", + "content": "[4] Mingqin Chen, Yuhui Quan, Tongyao Pang, and Hui Ji. Non-blind image deconvolution via leveraging model uncertainty in an untrained deep neural network. International Journal of Computer Vision, 130(7):1770-1789, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 281, + 288, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 288, + 323 + ], + "type": "text", + "content": "[5] Sunghyun Cho and Seungyong Lee. Convergence analysis of map based blur kernel estimation. In Proceedings of the IEEE/International Conference on Computer Vision, pages 4808-4816, 2017. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 288, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 378 + ], + "type": "text", + "content": "[6] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4641-4650, 2021. 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "type": "text", + "content": "[7] Laurent D'Andrés, Jordi Salvador, Axel Kochale, and Sabine Süssstrunk. Non-parametric blur map regression for depth of field extension. IEEE Transactions on Image Processing, 25(4):1660-1673, 2016. 1, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 288, + 479 + ], + "type": "text", + "content": "[8] Senyou Deng, Wenqi Ren, Yanyang Yan, Tao Wang, Fenglong Song, and Xiaochun Cao. Multi-scale separable network for ultra-high-definition video deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14030-14039, 2021. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 288, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 288, + 523 + ], + "type": "text", + "content": "[9] Jiangxin Dong, Stefan Roth, and Bernt Schiele. Deep wiener deconvolution: Wiener meets deep learning for image deblurring. Advances in Neural Information Processing Systems, 33:1048-1059, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "text", + "content": "[10] Jiangxin Dong, Stefan Roth, and Bernt Schiele. Dwdn: deep wiener deconvolution network for non-blind image deblurring. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9960-9976, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 288, + 634 + ], + "type": "text", + "content": "[11] Zhenxuan Fang, Fangfang Wu, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Self-supervised non-uniform kernel estimation with flow-based motion prior for blind image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18105-18114, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 690 + ], + "type": "text", + "content": "[12] Hongyun Gao, Xin Tao, Xiaoyong Shen, and Jiaya Jia. Dynamic scene deblurring with parameter selective sharing and nested skip connections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3848-3856, 2019. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "text", + "content": "[13] Alex Graves and Jürgen Schmidhuber. Framewise phoneme classification with bidirectional LSTM and other neural net-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "work architectures. Neural networks, 18(5-6):602-610, 2005. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 547, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 547, + 149 + ], + "type": "text", + "content": "[14] Fang Hua, Peter Johnson, Nadezhda Sazonova, Paulo Lopez-Meyer, and Stephanie Schuckers. Impact of out-of-focus blur on face recognition performance based on modular transfer function. In Proceedings of the International Conference on Biometrics, pages 85–90. IEEE, 2012. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 150, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 547, + 194 + ], + "type": "text", + "content": "[15] Bo Ji and Angela Yao. Multi-scale memory-based video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1919-1928, 2022. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 194, + 547, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 547, + 227 + ], + "type": "text", + "content": "[16] Ali Karaali, Naomi Harte, and Claudio R. Jung. Deep multiscale feature learning for defocus blur estimation. IEEE Transactions on Image Processing, 31:1097-1106, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 228, + 547, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 547, + 270 + ], + "type": "text", + "content": "[17] Ali Karaali and Claudio Rosito Jung. Edge-based defocus blur estimation with adaptive scale selection. IEEE Transactions on Image Processing, 27(3):1126-1137, Mar. 2018. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 271, + 547, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 271, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 271, + 547, + 316 + ], + "type": "text", + "content": "[18] Ali Karaali and Cláudio Rosito Jung. Svbr-net: A non-blind spatially varying defocus blur removal network. In Proceedings of the IEEE International Conference on Image Processing, pages 566-570. IEEE, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 316, + 547, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 547, + 349 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 547, + 349 + ], + "type": "text", + "content": "[19] Chaewon Kim, Jaeho Lee, and Jinwoo Shin. Zero-shot blind image denoising via implicit neural representations. arXiv preprint arXiv:2204.02405, 2022. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 350, + 547, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 547, + 382 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 547, + 382 + ], + "type": "text", + "content": "[20] Dilip Krishnan and Rob Fergus. Fast image deconvolution using hyper-laplacian priors. In Advances in Neural Information Processing Systems, 2009. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 383, + 547, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 547, + 437 + ], + "type": "text", + "content": "[21] Junyong Lee, Hyeongseok Son, Jaesung Rim, Sunghyun Cho, and Seungyong Lee. Iterative filter adaptive network for single image defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2034-2042, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 437, + 547, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 547, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 547, + 492 + ], + "type": "text", + "content": "[22] Dasong Li, Yi Zhang, Ka Chun Cheung, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. Learning degradation representations for image deblurring. In Proceedings of the European Conference on Computer Vision, pages 736-753. Springer, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 493, + 547, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 547, + 536 + ], + "type": "text", + "content": "[23] Yu Li, Dongwei Ren, Xinya Shu, and Wangmeng Zuo. Learning single image defocus deblurring with misaligned training pairs. Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 1, 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 537, + 547, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 547, + 592 + ], + "type": "text", + "content": "[24] Ziwei Luo, Haibin Huang, Lei Yu, Youwei Li, Haoqiang Fan, and Shuaicheng Liu. Deep constrained least squares for blind image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17642-17652, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 592, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 547, + 635 + ], + "type": "text", + "content": "[25] Haoyu Ma, Shaojun Liu, Qingmin Liao, Juncheng Zhang, and Jing-Hao Xue. Defocus image deblurring network with defocus map estimation as auxiliary task. IEEE Transactions on Image Processing, 31:216-226, 2021. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 636, + 547, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 547, + 690 + ], + "type": "text", + "content": "[26] Armin Mehri, Parichehr B Ardakani, and Angel D Sappa. Mprnet: Multi-path residual network for lightweight image super resolution. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2704-2713, 2021. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "type": "text", + "content": "[27] Sung-Jun Min, Kyeongbo Kong, and Suk-Ju Kang. Out-of-focus image deblurring for mobile display vision inspection." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12608" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 97, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 97, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 287, + 151 + ], + "type": "text", + "content": "[28] Seungjun Nah, Tae Kim, and Kyoung Mu Lee. Deep multiscale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 257-265, 2017. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 154, + 288, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 154, + 288, + 198 + ], + "spans": [ + { + "bbox": [ + 48, + 154, + 288, + 198 + ], + "type": "text", + "content": "[29] Yuesong Nan and Hui Ji. Deep learning for handling kernel/model uncertainty in image deconvolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2388-2397, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 200, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 200, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 200, + 287, + 243 + ], + "type": "text", + "content": "[30] Yuesong Nan, Yuhui Quan, and Hui Ji. Variational-EM-based deep learning for noise-blind image deblurring. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3626–3635, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 246, + 287, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 246, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 48, + 246, + 287, + 290 + ], + "type": "text", + "content": "[31] Saqib Nazir, Lorenzo Vaquero, Manuel Mucientes, Víctor M Brea, and Daniela Coltuc. Depth estimation and image restoration by deep learning from defocused images. arXiv preprint arXiv:2302.10730, 2023. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 292, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 292, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 292, + 287, + 346 + ], + "type": "text", + "content": "[32] Jinsun Park, Yu-Wing Tai, Donghyeon Cho, and In So Kweon. A unified approach of multi-scale deep and handcrafted features for defocus estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1736-1745, 2017. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "type": "text", + "content": "[33] Valeriya Pronina, Filippos Kokkinos, Dmitry V Dylov, and Stamatios Lefkimmiatis. Microscopy image restoration with deep wiener-kolmogorov filters. In Proceedings of the European Conference on Computer Vision, pages 185-201. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 406, + 287, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 406, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 48, + 406, + 287, + 450 + ], + "type": "text", + "content": "[34] Yuhui Quan, Zhuojie Chen, Huan Zheng, and Hui Ji. Learning deep non-blind image deconvolution without ground truths. In Proceedings of the European Conference on Computer Vision, pages 642-659, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 453, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 287, + 495 + ], + "type": "text", + "content": "[35] Yuhui Quan, Peikang Lin, Yong Xu, Yuesong Nan, and Hui Ji. Nonblind image deblurring via deep learning in complex field. IEEE Transactions on Neural Networks and Learning Systems, 33(10):5387-5400, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 498, + 287, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 541 + ], + "type": "text", + "content": "[36] Yuhui Quan, Hui Wu, Zicong, and Hui Ji. Gaussian kernel mixture network for single image defocus deblurring. In Advances in Neural Information Processing Systems, volume 34, 2021. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 597 + ], + "type": "text", + "content": "[37] Yuhui Quan, Hui Wu, Zicong, and Hui Ji. Neumann network with recursive kernels for single image defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5754-5763, June 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "type": "text", + "content": "[38] Wenqi Ren, Jiawei Zhang, Lin Ma, Jinshan Pan, Xiaochun Cao, Wangmeng Zuo, Wei Liu, and Ming-Hsuan Yang. Deep non-blind deconvolution via generalized low-rank approximation. Advances in Neural Information Processing Systems, 31, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[39] Wenqi Ren, Jiawei Zhang, Jinshan Pan, Sifei Liu, Jimmy S Ren, Junping Du, Xiaochun Cao, and Ming-Hsuan Yang. Deblurring dynamic scenes via spatially varying recurrent neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):3974-3987, 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[40] Lingyan Ruan, Bin Chen, Jizhou Li, and Miuling Lam. Learning to deblur using light field generated and real defocus images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16304-16313, 2022. 1, 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "text", + "content": "[41] Lingyan Ruan, Bin Chen, Jizhou Li, and Miu-Ling Lam. Aifnet: All-in-focus image restoration network using a light field-based dataset. IEEE Transactions on Computational Imaging, 7:675-688, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 206 + ], + "type": "text", + "content": "[42] Nimrod Shabtay, Eli Schwartz, and Raja Giryes. Pip: Positional-encoding image prior. arXiv preprint arXiv:2211.14298, 2022. 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 209, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 251 + ], + "type": "text", + "content": "[43] Jianping Shi, Li Xu, and Jiaya Jia. Discriminative blur detection features. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2965-2972, 2014. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "text", + "content": "[44] Jianping Shi, Li Xu, and Jiaya Jia. Just noticeable defocus blur detection and estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 657-665, 2015. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 299, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 352 + ], + "type": "text", + "content": "[45] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in Neural Information Processing Systems, 28, 2015. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "type": "text", + "content": "[46] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 399, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 453 + ], + "type": "text", + "content": "[47] Hyeongseok Son, Junyong Lee, Sunghyun Cho, and Seungyong Lee. Single image defocus deblurring using kernel-sharing parallel atrous convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2642-2650, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 456, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 509 + ], + "type": "text", + "content": "[48] Xin Tao, Hongyun Gao, Yi Wang, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8174-8182, 2018. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 511, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 576 + ], + "type": "text", + "content": "[49] Longguang Wang, Yingqian Wang, Xiaoyu Dong, Qingyu Xu, Jungang Yang, Wei An, and Yulan Guo. Unsupervised degradation representation learning for blind superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10581-10590, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 578, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 610 + ], + "type": "text", + "content": "[50] Shaowen Xie, Hao Zhu, Zhen Liu, Qi Zhang, You Zhou, Xun Cao, and Zhan Ma. Diner: Disorder-invariant implicit neural representation. arXiv preprint arXiv:2211.07871, 2022. 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "type": "text", + "content": "[51] Guodong Xu, Yuhui Quan, and Hui Ji. Estimating defocus blur via rank of local patches. In Proceedings of the IEEE/International Conference on Computer Vision, pages 5371-5379, 2017. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "text", + "content": "[52] Li Xu, Jimmy S Ren, Ce Liu, and Jiaya Jia. Deep convolutional neural network for image deconvolution. Advances in Neural Information Processing Systems, 27, 2014. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[53] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12609" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 541 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 115 + ], + "type": "text", + "content": "Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "type": "text", + "content": "[54] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5978-5986, 2019. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 173, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 173, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 173, + 287, + 239 + ], + "type": "text", + "content": "[55] Jiawei Zhang, Jinshan Pan, Jimmy Ren, Yibing Song, Linchao Bao, Rynson WH Lau, and Ming-Hsuan Yang. Dynamic scene deblurring using spatially variant recurrent neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2521-2529, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 240, + 287, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 240, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 49, + 240, + 287, + 283 + ], + "type": "text", + "content": "[56] Jie Zhang and Wanming Zhai. Blind attention geometric restraint neural network for single image dynamic/defocus deblurring. IEEE Transactions on Neural Networks and Learning Systems, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 285, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 285, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 49, + 285, + 287, + 339 + ], + "type": "text", + "content": "[57] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 341, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 341, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 49, + 341, + 287, + 384 + ], + "type": "text", + "content": "[58] Wenda Zhao, Fei Wei, You He, and Hutchuan Lu. United defocus blur detection and deblurring via adversarial promoting learning. In Proceedings of the European Conference on Computer Vision, pages 569-586, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 386, + 287, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 386, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 49, + 386, + 287, + 429 + ], + "type": "text", + "content": "[59] Wenda Zhao, Fan Zhao, Dong Wang, and Huchuan Lu. Defocus blur detection via multi-stream bottom-top-bottom network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42:1884-1897, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 430, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 430, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 49, + 430, + 287, + 485 + ], + "type": "text", + "content": "[60] Bolun Zheng, Quan Chen, Shanxin Yuan, Xiaofei Zhou, Hua Zhang, Jiyong Zhang, Chenggang Yan, and Gregory Slabaugh. Constrained predictive filters for single image bokeh rendering. IEEE Transactions on Computational Imaging, 8:346-357, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 487, + 287, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 487, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 49, + 487, + 287, + 541 + ], + "type": "text", + "content": "[61] Chao Zhu, Hang Dong, Jinshan Pan, Boyang Liang, Yuhao Huang, Lean Fu, and Fei Wang. Deep recurrent neural network with multi-scale bi-directional propagation for video deblurring. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 3598-3607, 2022. 3" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12610" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Single Image Reflection Separation via Component Synergy/6169238f-a2a5-4863-b901-a9125cd1dec8_content_list.json b/2023/Single Image Reflection Separation via Component Synergy/6169238f-a2a5-4863-b901-a9125cd1dec8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a1be9b2653379dbab41482db2abf746aa0734d03 --- /dev/null +++ b/2023/Single Image Reflection Separation via Component Synergy/6169238f-a2a5-4863-b901-a9125cd1dec8_content_list.json @@ -0,0 +1,1786 @@ +[ + { + "type": "text", + "text": "Single Image Reflection Separation via Component Synergy", + "text_level": 1, + "bbox": [ + 181, + 130, + 785, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qiming Hu Xiaojie Guo*", + "bbox": [ + 380, + 180, + 593, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "College of Intelligence and Computing, Tianjin University, Tianjin, China", + "bbox": [ + 192, + 199, + 776, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "huqiming@tju.edu.cn, xj.max.quo@gmail.com", + "bbox": [ + 300, + 219, + 666, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The reflection superposition phenomenon is complex and widely distributed in the real world, which derives various simplified linear and nonlinear formulations of the problem. In this paper, based on the investigation of the weaknesses of existing models, we propose a more general form of the superposition model by introducing a learnable residue term, which can effectively capture residual information during decomposition, guiding the separated layers to be complete. In order to fully capitalize on its advantages, we further design the network structure elaborately, including a novel dual-stream interaction mechanism and a powerful decomposition network with a semantic pyramid encoder. Extensive experiments and ablation studies are conducted to verify our superiority over state-of-the-art approaches on multiple real-world benchmark datasets. Our code is publicly available at https://github.com/mingcv/DSRNet.", + "bbox": [ + 75, + 300, + 473, + 541 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 569, + 209, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a typical layer superimposition scenario, pictures captured through glass-like surfaces may be blended with undesired reflection, which not only impairs the aesthetic value but also hinders the downstream tasks [31]. In the meantime, scenes of interest are possibly concealed in reflection [35]. Therefore, both the parts transmitted through a surface (transmission layer, $\\mathbf{T}$ ) and those reflected (reflection layer, $\\mathbf{R}$ ) are desired to be reconstructed from a superimposed image $\\mathbf{I}$ to fulfill the practical demands.", + "bbox": [ + 75, + 595, + 468, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Following a common assumption [17], $\\mathbf{I}$ is linearly composed by $\\mathbf{T}$ and $\\mathbf{R}$ expressed as $\\mathbf{I} = \\mathbf{T} + \\mathbf{R}$ , which has been popular due to its simplicity. However, in real-world scenarios, the reflection and transmission layers are likely to be weakened due to diffusion and other problems during the superimposition [35]. Therefore, several methods [34, 39] introduce scalars $\\alpha$ and $\\beta$ to the two components, respectively, obtaining $\\mathbf{I} = \\alpha \\mathbf{T} + \\beta \\mathbf{R}$ . However, the linear superimposition model is often violated due to over-exposure", + "bbox": [ + 75, + 731, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/56033888929df1fdf01f854eea064f3699add69cf7861a5af3cfee24ae60856a.jpg", + "image_caption": [ + "1" + ], + "image_footnote": [], + "bbox": [ + 504, + 266, + 578, + 324 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e54e280b6bb2ed473e36128b628896277f6b439bac2d5777531d79f7ea17d9b.jpg", + "image_caption": [ + "T (GT)" + ], + "image_footnote": [], + "bbox": [ + 581, + 268, + 655, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e195b4812d272af10aec6a4a3fc2d927c302ae4ab9f5f26ca0915392de0c8d3f.jpg", + "image_caption": [ + "R (GT)" + ], + "image_footnote": [], + "bbox": [ + 656, + 268, + 732, + 324 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/86ee0166c729d4e5266ac42b2144ea8e06c051c58ff5bd1e030717eb281cce07.jpg", + "image_caption": [ + "Crops" + ], + "image_footnote": [], + "bbox": [ + 733, + 268, + 810, + 324 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/802dece7a144651c19a192ab710a3618c747b3c26a3a06bc767c91f970f94b9d.jpg", + "image_caption": [ + "$\\widetilde{\\mathbf{R}} = \\mathbf{I} - \\mathbf{T}$" + ], + "image_footnote": [], + "bbox": [ + 813, + 268, + 888, + 324 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d7d00c836444e1eed4e6f51d155a1ce6162c62fe63603c379b392c702dd0457.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 342, + 656, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3940ba43f6c6ad3cb4fe7cd80e4a15d68daf11ef1c4aa786677c21703f8dff88.jpg", + "image_caption": [ + "$\\alpha \\mathbf{T} + \\mathbf{R}$" + ], + "image_footnote": [], + "bbox": [ + 581, + 401, + 656, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8be5773578d8e377a6b8052ab32d3c5a6aab1c6f2dbc64621e91d6b5821b2590.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 401, + 732, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0597247a6e7e50f04d11b6a84fe15ee250e611dc156ff89cc2159bc147603e1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 342, + 810, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e634e827a617c59fc25b85b37010c3f09ade9384bbfe0db24163efce5f7091e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 401, + 810, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a521327a5a64758d209357b7498ff74004e3701e64339178c6c229c5161ac21f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 820, + 342, + 880, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cb6574b1db09669c5542194ffd894a15586491ffd4373b947c6073c1c2ba07a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 821, + 407, + 879, + 453 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ea1e9750503db500715a77c1ae45a7b1f9f3115e6742e5706e62c39303c3e0d9.jpg", + "image_caption": [ + "(IBCLN, CVPR,2020)", + "WoT+R", + "(Dong et al., ICCV,2021)" + ], + "image_footnote": [], + "bbox": [ + 581, + 460, + 656, + 518 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df5d9a3aae5e2511b328004d2cb70fcf424892a60610677108f3ab6f57f82c6a.jpg", + "image_caption": [ + "Figure 1: A visual example drawn from the $\\mathrm{SIR}^2$ dataset. The first row displays a real input $\\mathbf{I}$ with ground-truth $\\mathbf{T}$ and $\\mathbf{R}$ . In rows 2-5, the separations based on different physical models are compared, where $\\hat{\\mathbf{T}}$ and $\\hat{\\mathbf{R}}$ represent the predictions. Our method is able to produce a \"Residue\" term, separating the nonlinearity from the layer reconstructions." + ], + "image_footnote": [], + "bbox": [ + 581, + 521, + 656, + 579 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/89194c8b48b6bbc71358e9e44f7a2e79e31dece73c35a930078eaf8b6cdce826.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 460, + 732, + 518 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/54eaa68eff54bf6efc27ca67174227a550c862193e5d01ec4eb8c58bad0347a7.jpg", + "image_caption": [ + "R" + ], + "image_footnote": [], + "bbox": [ + 658, + 522, + 732, + 579 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2a797903b894441eb6d467b600a5b4484d47f917f52e82295d490826236ecba7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 460, + 810, + 518 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3702c3be7fe95cd56c097304b2349284bb7fca4fad51233e003cb9af86696456.jpg", + "image_caption": [ + "Crops" + ], + "image_footnote": [], + "bbox": [ + 733, + 522, + 810, + 578 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/761dced01cc93bd9a6e25375956da6716dfefa58e796facb26a7726de2f98b17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 813, + 465, + 880, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d44f4ed51cf31336afabf11ce3e43995dd5f9350a395bdfa62645f84f5ff835.jpg", + "image_caption": [ + "Residue" + ], + "image_footnote": [], + "bbox": [ + 813, + 523, + 888, + 577 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and other phenomena [37]. Thus, an alpha-matting map $\\mathbf{W}$ is introduced to the model as $\\mathbf{I} = \\mathbf{W} \\circ \\mathbf{T} + \\bar{\\mathbf{W}} \\circ \\mathbf{R}$ with $\\bar{\\mathbf{W}} = \\mathbf{1} - \\mathbf{W}$ , which, however, increases the degree of freedom and makes this ill-posed problem much harder. In view of all these limitations, it is not a trivial task to represent different kinds of reflection scenarios with a single model. In this work, we take a step forward in advancing the solution and deliver a more general form of the superimposition procedure by introducing a residue term as follows:", + "bbox": [ + 496, + 715, + 893, + 852 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} = \\tilde {\\mathbf {T}} + \\tilde {\\mathbf {R}} = \\mathbf {T} + \\mathbf {R} + \\Phi (\\mathbf {T}, \\mathbf {R}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 859, + 890, + 877 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $\\mathbf{T}$ and $\\mathbf{R}$ denote the groundtruth scenes of the trans", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author. This work was supported by National Natural Science Foundation of China under Grant no. 62072327.", + "bbox": [ + 76, + 875, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13138", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mission and reflection layers, while their respective information contained in $\\mathbf{I}$ after superimposition and other degradations and finally reach the camera sensors is represented by $\\tilde{\\mathbf{T}}$ and $\\tilde{\\mathbf{R}}$ . $\\Phi(\\mathbf{T}, \\mathbf{R}) = \\mathbf{I} - \\mathbf{T} - \\mathbf{R}$ is the residue of the reconstruction, which can be caused by attenuation, overexposure, etc. $\\Phi(\\cdot, \\cdot)$ represents a group of functions that can be used to model the residue in specific situations. For example, $\\Phi(\\mathbf{T}, \\mathbf{R}) = 0$ in [23, 5, 17, 41, 36, 12], $\\Phi(\\mathbf{T}, \\mathbf{R}) = (\\alpha - 1)\\mathbf{T} + (\\beta - 1)\\mathbf{R}$ in [34, 39], $\\Phi(\\mathbf{T}, \\mathbf{R}) = -(\\bar{\\mathbf{W}} \\circ \\mathbf{T} + \\mathbf{W} \\circ \\mathbf{R})$ in [37, 42]. In an effort to rule these different cases, a learnable module is desired to be introduced to model the residue $\\Phi(\\mathbf{T}, \\mathbf{R})$ . To clarify it, we provide a visual example in Fig. 1, where the first row shows a superimposed image $\\mathbf{I}$ , its groundtruth transmission $\\mathbf{T}$ (GT) and reflection $\\mathbf{R}$ (GT) layers collected by SIR $^2$ dataset. According to the linear model, the guidance for reflection restoration is $\\tilde{\\mathbf{R}} = \\mathbf{I} - \\mathbf{T}$ with incomplete signals rather than $\\mathbf{R}$ (GT) that contains full information, which not only impairs the training of networks but also hinders the downstream usages. Therefore, as we point out in Eq. (1), a promising way out of this dilemma is to utilize ground-truth layers $\\mathbf{T}$ and $\\mathbf{R}$ as guiding signals and employ an extra residual term to handle the non-linearity (\"Residue\" in Fig. 1). As shown, the components that violate the linear assumption are separated from the layer predictions $\\hat{\\mathbf{T}}$ and $\\hat{\\mathbf{R}}$ , and thus more complete and precious layer separations than previous methods are obtained.", + "bbox": [ + 76, + 90, + 472, + 496 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further exploit the synergy between the components $\\mathbf{T}$ and $\\mathbf{R}$ , we deliberate upon the facilitation of intercomponent feature interaction. The efficiency of a dual-stream interactive network has been verified in SIRS problem by [12]. Building upon their analysis, we subsequently advance a more effective scheme. Inspired by the gated mechanisms [2, 32], we present a simple yet effective mutually-gated interaction diagram (MuGI) serving as a better feature interaction candidate, by means of which we design the MuGI block to build our decomposition network. Moreover, the hypercolumn [41], as a commonly used semantic information encoder in the SIRS task, tends to aggregate features in a lossy way. Therefore, we propose a dual-stream pyramid fusion network (DSFNet) to replace it, which hierarchically decomposes and fuses the multiscale semantic information leveraging our proposed MuGI blocks and dual-stream fusion blocks (DSF Block). The roughly decomposed features of layers are further delivered into the dual-stream fine-grained decomposition phase (DSDNet). The two sub-networks constitute the main branch of our Dual-stream Semantic-aware network with Residual correction, namely DSRNet.", + "bbox": [ + 76, + 500, + 472, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are as follows:", + "bbox": [ + 96, + 830, + 434, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We build a general form for SIRS by introducing a learnable residue term, which is more flexible to different scenarios and boosts the separation of layers;", + "bbox": [ + 94, + 854, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We exploit the synergy of features via mutually-gated interaction block within a dual-stream semantic-aware network, which facilitates information usage;", + "- Extensive experiments are conducted to demonstrate the effectiveness of our design. Overall, it achieves state-of-the-art performance against the alternatives on multiple real-world benchmarks for SIRS." + ], + "bbox": [ + 517, + 90, + 893, + 204 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 220, + 640, + 237 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Plenty of efforts have been devoted to handling the image reflection separation problem in the past decades. They either leverage multiple images to acquire more cues for the layer decomposition or exploit extra priors to manage the problem through a single image. In what follows, we organize the previous methods based on the images they require.", + "bbox": [ + 496, + 247, + 890, + 339 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multiple Image Reflection Separation. In general, the transmission component tends to be unpolarized, yet the reflection component varies when rotating the polarization filter mounted in front of a camera sensor. Inspired by this phenomenon, a variety of methods [28, 14, 13, 27, 21, 16] resort to the physical solution that separates different components through a sequence of images with varied polarization orientations. Besides the polarization cues, the layer separations can also be indicated by different focuses [6], stereo information [31], flash on/off [1, 15], relative motions [29, 8, 22, 38, 40, 25, 26] and scene consistency [9, 30, 10]. Although involving multiple images mitigates the ill-posedness of the separation problem, this pipeline requires professional devices (such as polarizers) and manual operations, which limits its application.", + "bbox": [ + 496, + 339, + 892, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single Image Reflection Separation. Compared with multiple-image solutions, single-image-based methods show more merits, being more flexible and requiring less manual operations. However, as there is no such thing as a free lunch, single-image schemes call for additional priors to cope with its ill-posedness nature, such as gradient-based constraints [19, 18, 17, 23] and manual annotations [17]. Specifically, Levin et al. [19, 18] impose sparse constraints on gradients to acquire the separations that have fewer edges. Levin and Weiss [17] demand several manual annotations of respective layers to obtain favorable decompositions, which involves additional human labor. Li and Brown [23] suppose that one layer is smoother than the other, thus applying unbalanced penalization on the gradients of the two layers to decompose them. Although these methods work fine in restricted scenarios, the complicated real-world conditions can considerably surpass their assumptions, leading to unsatisfactory results.", + "bbox": [ + 496, + 566, + 893, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The advent of deep learning technology mitigates the incongruity between the volume of data and modeling capacity. Certain underlying assumptions can be subtly embedded in the process of data synthesis, and subsequently", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13139", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ec01cc870caafaa657d0cea161c2e87a94fff8633972795cd349c32bb1846eae.jpg", + "image_caption": [ + "Figure 2: The architecture of our proposed DSRNet, which consists of two cascaded stages and a learnable residue module (LRM). In Stage 1, DSFNet aggregates and separates hierarchical semantic information with two interactive feature streams. The roughly separated features are further fed into the DSDNet for fine-grained decompositions in Stage 2, where the LRM takes decomposed features, separating out the components that violate the linear assumption." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 893, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "learned by deep models. Concretely, CEILNet [5] imposes the relative smoothness prior to the synthesis of reflection layers and combines them with transmission layers by addition. An edge-aware network is developed to capture the transmission components, which yet ignores high-level semantics that are likely to aid the SIRS task. Zhang et al. [41] hence introduce HyperColumn features [11] using a pre-trained VGG-19 network to acquire semantic awareness, besides the perceptual and adversarial losses. Moreover, the exclusivity loss is developed to penalize the intersected gradients. ERRNet [36] goes a step further by leveraging an additional set of misaligned pairs. However, it appears to overlook the estimation of the reflection layer, which is essentially an image component and potentially important to distinguish the transmission parts. Li et al. [24] thereby presents a two-stage network (RAGNet) to first estimate reflection components and predict transmission ones guided by them, whereas the reflection estimation is isolated from the transmission in this way. Further, Hu and Guo [12] come up with the YTMT strategy, which pays equal attention to the two components, and a dual-stream interactive network is developed to restore the two layers simultaneously. However, their linear assumption makes the predicted reflection components tend to be weak in many cases. Other than them, BDN [39] and IBCLN [20] make use of reflection models weighted by scalars, and iteratively estimate both the components. This scheme prevents the reflection from being too weak, but its restorations often struggle to be free from the transmission parts. Furthermore, Wen et al. [37] simulate the nonlinear superimposition phenomenon by predicting a three-channel alpha blending weight map with the adversarial guidance of col", + "bbox": [ + 75, + 411, + 472, + 897 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lected unpaired images. Dong et al. [3] develops an iterative network and estimates a probabilistic reflection confidence map in each iteration. However, it is not an easier task to estimate an alpha blending map than to determine each of the blended components, while the residual estimation only needs to retain the redundant information during decomposition, which is more practical than the former. Considering the drawbacks of the previous reflection models, we propose a general form of them with a learnable residue term, which is shown to be more effective and flexible.", + "bbox": [ + 496, + 411, + 893, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 594, + 635, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, our proposed DSRNet comprises two cascaded stages and a learnable residue module (LRM). In what follows, we first clarify the motivation of LRM and then introduce the MuGI block and the architecture of the DSRNet followed by its training loss functions.", + "bbox": [ + 496, + 621, + 893, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Residue of the Linear Combination", + "text_level": 1, + "bbox": [ + 498, + 709, + 807, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It often occurs that a superimposed image $\\mathbf{I}$ cannot be perfectly represented by the linear combination of $\\mathbf{T}$ and $\\mathbf{R}$ , leading to a residue term, as discussed in Section 1. Under its disturbance, the error of the additive reconstruction criterion $\\varepsilon = \\| \\mathbf{I} - (\\hat{\\mathbf{T}} +\\hat{\\mathbf{R}})\\|$ remains large even though the predicted layers can well reconstruct the ground truths. Further minimizing $\\varepsilon$ will put redundant information to either $\\hat{\\mathbf{T}}$ or $\\hat{\\mathbf{R}}$ , which induces the deviation of predictions from their ground truths instead. Moreover, there are many kinds of simplified physical models proposed by previous work with their own shortcomings. Naturally, a unified model is de", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13140", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sired to be put forward. Therefore, we introduce an extra residue term to offset the error in additive reconstruction, which, at the same time, unifies the different physical models. As depicted in Fig. 2, we leverage a learnable residue module (LRM) after Stage 2 to estimate the residual information during the decomposition by analyzing the features of $\\hat{\\mathbf{T}}$ and $\\hat{\\mathbf{R}}$ before the final layer. LRM consists of an interactive and a fusion part to collect and merge residual information from two branches. The dual-stream signals are concatenated before entering the fusion part. To constrain the space of residue, the $\\tanh(\\cdot)$ function is utilized as the final activation. With the participation of LRM, we define the following reconstruction loss with residual rectification $(\\mathrm{R}^3\\mathrm{Loss})$ :", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} := \\left\\| \\mathbf {I} - \\left(\\hat {\\mathbf {T}} + \\hat {\\mathbf {R}}\\right) - \\Phi (\\hat {\\mathbf {T}}, \\hat {\\mathbf {R}}) \\right\\| _ {1}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 311, + 468, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Phi$ denotes the LRM and $\\| \\cdot \\| _1$ means the $\\ell_1$ norm. Guided by this objective, the information beyond the additive model will flow to the residue term. Notably, this term can be totally discarded during the testing phase, avoiding extra computational costs.", + "bbox": [ + 75, + 339, + 468, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Mutually-gated Interactive Block", + "text_level": 1, + "bbox": [ + 76, + 424, + 369, + 440 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now that we employ the LRM to model the residual components during the additive reconstruction, the rest part can be seen as a linear model again as $\\tilde{\\mathbf{I}} = \\mathbf{I} - \\Phi (\\hat{\\mathbf{T}},\\hat{\\mathbf{R}}) = \\hat{\\mathbf{T}} +\\hat{\\mathbf{R}}$ . As displayed in Fig. 3 (a), we exploit the synergy between $\\hat{\\mathbf{T}}$ and $\\hat{\\mathbf{R}}$ by developing the Mutually-Gated Interactive mechanism, namely MuGI:", + "bbox": [ + 75, + 446, + 470, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\hat {\\mathbf {F}} _ {\\mathbf {T}} = \\mathcal {G} _ {1} \\left(\\mathbf {F} _ {\\mathbf {T}}\\right) \\circ \\mathcal {G} _ {2} \\left(\\mathbf {F} _ {\\mathbf {R}}\\right); \\\\ \\hat {\\mathbf {F}} _ {\\mathbf {R}} = \\mathcal {G} _ {1} \\left(\\mathbf {F} _ {\\mathbf {R}}\\right) \\circ \\mathcal {G} _ {2} \\left(\\mathbf {F} _ {\\mathbf {T}}\\right), \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 547, + 468, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{F_T},\\mathbf{F_R}\\in \\mathbb{R}^{H_1\\times W_1\\times C_1}$ denote intermediate features of streams for predicting $\\hat{\\mathbf{T}}$ and $\\hat{\\mathbf{R}}$ , respectively, and $\\hat{\\mathbf{F}}_{\\mathrm{T}},\\hat{\\mathbf{F}}_{\\mathrm{R}}\\in \\mathbb{R}^{H_2\\times W_2\\times C_2}$ represent the outputs of the mutual gate. $\\mathcal{G}_1$ and $\\mathcal{G}_2$ are functions selecting which part of features to engage the interaction. $\\circ$ indicates the elementwise multiplication. Here we use a simple implementation of MuGI to evaluate its capability of solving SIRS, which divides a feature map in half at the channel dimension, and $\\mathcal{G}_1(\\cdot)$ always selects the former half, while $\\mathcal{G}_2(\\cdot)$ chooses the latter. We have $H_{1} = H_{2}$ , $W_{1} = W_{2}$ and $C_2 = C_1 / 2$ in this way. Such a gated mechanism captures the mutual dependency of the two components in the feature space, say if either stream contains signals the other one desires, they tend to be split into the latter half (selected by $\\mathcal{G}_2(\\cdot)$ ) and transferred into the sibling stream by the gate.", + "bbox": [ + 75, + 597, + 468, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our simple implementation of MuGI makes it plug-and-play. Therefore, in general, most of the single-path blocks can be converted into a dual-stream one by parallel setting two of them and relating them with MuGI. In this paper, we follow a state-of-the-art block design presented by [2]", + "bbox": [ + 75, + 825, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f543ebc8190b575d8149e22d166c0f824467140a4ab24ad36f478e445e0d2c28.jpg", + "image_caption": [ + "(a). MuGI Mechanism" + ], + "image_footnote": [], + "bbox": [ + 506, + 107, + 880, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0d629e279f96afb3c67f50b88c529afdca7bd39dad85154d2ef385166dd5cf3e.jpg", + "image_caption": [ + "(b). MuGI Block", + "Figure 3: MuGI mechanism and MuGI Block." + ], + "image_footnote": [], + "bbox": [ + 504, + 196, + 885, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and convert its single-path NAF block into a dual-stream mutually-gated interactive block (MuGI block). Its diagram is depicted in Fig. 3 (b), in which the simple gates are replaced by MuGI gates. Specifically, the layer normalization makes dual-stream features comparable, $1 \\times 1$ convolution doubles the channel dimensions in order to conduct MuGI without information loss which reduces dimensions by half during the interaction, channel attention and the following $1 \\times 1$ convolution serve as fusion/reweighting roles.", + "bbox": [ + 498, + 409, + 890, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dual-stream Semantic-aware Network", + "text_level": 1, + "bbox": [ + 500, + 559, + 830, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The semantic information provided by pre-trained models is usually introduced to alleviate the impact of ill-posedness in SIRS. The HyperColumn [41, 36, 12], as a prevalent multi-scale semantic feature extractor in SIRS, aggregates semantics through interpolating the features extracted by a pre-trained model into the same scale as the input images. It then concatenates them together and employs a $1 \\times 1$ convolution to rapidly reduce the channel dimensions (typically from 1475 to 64 or 256), before feeding them into the decomposition networks. This strategy omits the inner relationship of multi-scale features and may discard informative signals. To overcome the drawbacks mentioned above, we present a dual-stream pyramid fusion network (DSFNet), which hierarchically aggregates the extracted features by jointly upsampling and interacting. As shown in Fig. 2, given the hierarchical features extracted by a pre-trained deep network (e.g., VGGNet-19), DSFNet gathers them in a bottom-up manner with MuGI blocks and dual-stream fusion blocks (DSF Block), the structure of which is shown in Fig. 2. \"Dual Conv\" means two parallel convolutional layers. The features extracted by the pre", + "bbox": [ + 496, + 583, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13141", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/00d7982576588fb0d41182a4bcc3e114d5965d3339ead231c560e79c9e97ef95.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsReal20 (20)Objects (200)Postcard (199)Wild (55)Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Zhang et al. [41]22.550.78822.680.87916.810.79721.520.83220.080.835
BDN [39]18.410.72622.720.85620.710.85922.360.83021.650.849
ERRNet [36]22.890.80324.870.89622.040.87624.250.85323.530.879
IBCLN [20]21.860.76224.870.89323.390.87524.710.88624.100.879
RAGNet [24]22.950.79326.150.90323.670.87925.530.88024.900.886
DMGN [7]20.710.77024.980.89922.920.87723.810.83523.800.877
Zheng et al. [43]20.170.75525.200.88023.260.90525.390.87824.190.885
YTMT [12]23.260.80624.870.89622.910.88425.480.89024.050.886
Ours24.230.82026.280.91424.560.90825.680.89625.400.905
Dong et al.†[3]23.340.81224.360.89823.720.90325.730.90224.210.897
Ours†23.910.81826.740.92024.830.91126.110.90625.750.910
", + "bbox": [ + 122, + 88, + 848, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: Quantitative results on four real-world benchmark datasets of methods. The best results are indicated in **bold.†** indicates extra training data that are involved as in [3].", + "bbox": [ + 75, + 308, + 893, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "trained network at each level first interact through a MuGI block, thereby features in the same scale are related and transformed into dual-stream features $(\\mathbf{F_T} = \\mathbf{F_R}$ for these blocks). After the interaction, the deepest features extracted by VGGNet are upscaled and then fused with shallower features through the DSF block, in which the dual-stream features come from the two adjacent scales are concatenated at channel dimension, fused by a 1x1 convolutional layer and then upscaled at each stream. The MuGI blocks are further employed to promote cross-scale interactions. Note that, shallow features extracted from RGB inputs are fused at the top of DSFNet, which preserves fine-grained details to restore both layers. The aggregated features provide rough separations of layers and are further refined by the second stage. The following stage (DSDNet) is constructed by the proposed MuGI blocks in a U-shaped manner. More details can be found in the supplementary materials.", + "bbox": [ + 75, + 366, + 472, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Loss Function", + "text_level": 1, + "bbox": [ + 76, + 630, + 222, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Besides the $\\mathbf{R}^3$ loss demonstrated in Section 3.1, we further introduce pixel, perceptual, and exclusion loss for pixel-wise and semantic fidelity as well as the gradient independence of layers, which are described as follows:", + "bbox": [ + 75, + 652, + 468, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pixel Loss. The pixel loss is used to constrain the consistency of the reconstructed layer and groundtruth in both the natural image domain and gradient domain, minimizing their errors as follows:", + "bbox": [ + 75, + 714, + 468, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {p i x} := \\| \\hat {\\mathbf {T}} - \\mathbf {T} \\| _ {2} ^ {2} + \\| \\hat {\\mathbf {R}} - \\mathbf {R} \\| _ {2} ^ {2} \\tag {4} \\\\ + \\alpha (\\| \\nabla \\hat {\\mathbf {T}} - \\nabla \\mathbf {T} \\| _ {1} + \\| \\nabla \\hat {\\mathbf {R}} - \\nabla \\mathbf {R} \\| _ {1}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 777, + 468, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\|_2$ denotes the $\\ell_2$ norm. $\\nabla$ stands for the gradient operator, which extracts gradients in both vertical and horizontal directions. We express one of them in the formulation for clarity and average them in practice. $\\alpha$ is set as 2 in our experiments.", + "bbox": [ + 75, + 824, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Perceptual Loss. The pixel-wise losses cannot assure the multi-scale consistency between predictions and their groundtruths, which may result in overwhelming punishment for the whole image due to small differences in local brightness. Therefore, the perceptual loss is further introduced as follows:", + "bbox": [ + 496, + 366, + 890, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p e r} := \\sum_ {i} \\omega_ {i} \\| \\phi_ {i} (\\hat {\\mathbf {T}}) - \\phi_ {i} (\\mathbf {T}) \\| _ {1}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 463, + 890, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\phi_i(\\cdot)$ designates the features drawn by layer $i\\in$ $\\{2,7,12,21,30\\}$ of a VGG-19 model. $\\omega_{i}$ s are combining weights of terms at different layers. We follow the setting of hyper-parameters in [36].", + "bbox": [ + 498, + 503, + 890, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Exclusion Loss. We introduce the exclusion loss to strengthen the gradient independence prior and reduce the structural coupling in estimated separations as below:", + "bbox": [ + 498, + 564, + 890, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {e x c} := \\frac {1}{N} \\sum_ {n = 0} ^ {N - 1} \\| \\Psi \\left(\\hat {\\mathbf {T}} ^ {\\downarrow n}, \\hat {\\mathbf {R}} ^ {\\downarrow n}\\right) \\| _ {2} ^ {2}, \\tag {6} \\\\ \\Psi (\\hat {\\mathbf {T}}, \\hat {\\mathbf {R}}) := \\tanh \\left(\\eta_ {1} | \\nabla \\hat {\\mathbf {T}} |\\right) \\circ \\tanh \\left(\\eta_ {2} | \\nabla \\hat {\\mathbf {R}} |\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 618, + 890, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{T}}^{\\downarrow n}$ and $\\hat{\\mathbf{R}}^{\\downarrow n}$ represent taking down-sampling by $2^{n}$ times ( $2^{N}$ at most) of $\\hat{\\mathbf{T}}$ and $\\hat{\\mathbf{R}}$ . $\\eta_{1}$ and $\\eta_{2}$ are normalization factors, which are identical to [41].", + "bbox": [ + 498, + 696, + 890, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gathering all the loss terms yields the final objective as:", + "bbox": [ + 517, + 744, + 885, + 758 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l l}} := \\mathcal {L} _ {\\text {p i x}} + \\beta_ {1} \\mathcal {L} _ {\\text {p e r}} + \\beta_ {2} \\mathcal {L} _ {\\text {e x c}} + \\beta_ {3} \\mathcal {L} _ {\\text {r e c}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 768, + 890, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta_{1} = 0.01$ , $\\beta_{2} = 1$ , and $\\beta_{3} = 0.2$ are set empirically.", + "bbox": [ + 498, + 794, + 890, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Validation", + "text_level": 1, + "bbox": [ + 498, + 821, + 730, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 847, + 717, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implemented in PyTorch, our models are optimized with the Adam optimizer on a single RTX 3090 GPU for 20", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13142", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b36c30d0e3a868cbcd511404139932fa80e9d12352846d23f6f5fef772b9d1fd.jpg", + "image_caption": [ + "Figure 4: Visual comparison of estimated transmission layers between state-of-the-arts and ours on real-world samples." + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 859, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "epochs at most to reach favorable overall performance. The learning rate is initialized as $10^{-4}$ and fixed during the training phase with a batch size of 1.", + "bbox": [ + 75, + 693, + 470, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. Our training dataset embraces both real and synthesized images. Following [5, 12], the training dataset are composed of 90 real pairs from [41] and 7,643 synthesized pairs from the PASCAL VOC dataset [4]. For synthesized data, transmission and reflection layers are weakened by coefficients $\\gamma_{1} \\in [0.8, 1.0]$ and $\\gamma_{2} \\in [0.4, 1.0]$ during blending them with the following model:", + "bbox": [ + 75, + 739, + 468, + 845 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {\\text {s y n}} = \\gamma_ {1} \\mathbf {T} _ {\\text {s y n}} + \\gamma_ {2} \\mathbf {R} _ {\\text {s y n}} - \\gamma_ {1} \\gamma_ {2} \\mathbf {T} _ {\\text {s y n}} \\circ \\mathbf {R} _ {\\text {s y n}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 857, + 468, + 875 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{T}_{syn},\\mathbf{R}_{syn}$ and $\\mathbf{I}_{syn}$ represent the transmission, re", + "bbox": [ + 76, + 885, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "flection, and superimposed layers during synthesis, respectively. This formulation is inspired by the \"screen\" blending mode in digital image processing, which always reserves lighter colors for the blending layers.", + "bbox": [ + 498, + 693, + 890, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Performance Evaluation", + "text_level": 1, + "bbox": [ + 498, + 768, + 723, + 784 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparison. A comparison is made between state-of-the-art methods as shown in Table 1, involving Zhang et al. [41], BDN [39], ERRNet [36], IBCLN [20], RAGNet [24], DMGN [7], Zheng et al. [43], YTMT [12] and ours on four real-world dataset, including Real20 [41] and three subsets of the $\\mathrm{SIR}^2$ Dataset [33]. Besides, to meet the configuration of [3], we train our DSRNet ad", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13143", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f931fc59e8419156fdeed9c2cc088dc884a3b229dba2bd65a9d19f630bbf4662.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 251, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bb297128c2955cc666671a00763ccd0c157f4150567ddc55778a802d132a1ed1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 253, + 88, + 406, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f51a50c0aabb19973dde4259174388174a86125dff3749c2d4bf7094f0a8258c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 88, + 560, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8958afe7f5e3ddee1b14c3bd566f43b163bee46a37f8f446ad7d85dcc12091d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 88, + 712, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7013d71ceeadfb10017ac1f6dc69ac322446ed27aef54f63b9db6d5396fddbcc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 88, + 867, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f5d2b60917c97df8636b27eeae956475412041c5e9060a62d158b6bac4077b57.jpg", + "image_caption": [ + "YTMT [12]" + ], + "image_footnote": [], + "bbox": [ + 98, + 186, + 251, + 265 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f5ef5a078f8e9aec5f972bdd91a453f99b1a7d68d49747621c5797ffa3af1e5.jpg", + "image_caption": [ + "Zhang et al. [41]", + "Dong et al. [3]" + ], + "image_footnote": [], + "bbox": [ + 251, + 186, + 406, + 263 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2606e1aad110ba01058e13f39177897efd92ed786618c8deb42074dfe0029ff0.jpg", + "image_caption": [ + "BDN [39]", + "Ours", + "Figure 5: Visual comparison of estimated reflections between state-of-the-arts and ours on a sample drawn from the $\\mathrm{SIR}^2$ dataset. Ours has a significant advantage over other alternatives by separating the residue component from the estimations." + ], + "image_footnote": [], + "bbox": [ + 406, + 186, + 560, + 265 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a16d8ec40f48c74b778c09db3304f953905e0e5f4858df6baa85dcef82b779d.jpg", + "image_caption": [ + "IBCLN [20]", + "Residue (ours)" + ], + "image_footnote": [], + "bbox": [ + 560, + 186, + 712, + 265 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14e79a74f5fe1a00dde28020204561efbc8c1e17ba44715add88a06d657abc98.jpg", + "image_caption": [ + "RAGNet [24]", + "GT" + ], + "image_footnote": [], + "bbox": [ + 714, + 186, + 867, + 265 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ditionally under their training settings, which includes 200 extra real pairs provided by the \"Nature\" dataset [20], and 13,700 synthesized image pairs provided by [41] instead.", + "bbox": [ + 75, + 354, + 468, + 400 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It turns out that our method shows its superiority over other competitors on all the testing datasets in both settings, gaining $0.50dB$ and $1.54dB$ in terms of average PSNR in the two settings, respectively. Given the four real-world datasets contains a variety of scenes, illumination conditions, and glass thickness, it is not a trivial task to achieve the best performance of all metrics on these datasets simultaneously. The experimental results demonstrate that our proposed SIRS scheme has significantly higher performance and stronger generalization ability, which explains our main contributions.", + "bbox": [ + 75, + 401, + 468, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative comparisons. To further explain our performance advantages, we provide visual comparisons of transmission layers in Fig. 4 against state-of-the-art methods, including Zhang et al. [41], BDN [39], ERRNet [36], IBCLN [20], RAGNet [24], YTMT [12], and Dong et al. [3]. As can be observed in Fig. 4, the method proposed by Zhang et al. fails to handle the case containing scattered reflection components in the second row. BDN has trouble dealing with the images with highlights like in the first row and even aggravates the reflections. The problem is likely to be caused by the linear reflection model, which lacks the ability to model specular highlights. For the results of ERRNet, which contains only a single branch to estimate transmission layers, it lacks the direct modeling of the reflections, therefore only removing parts of them. IBCLN also has trouble removing highlights and scattered reflections. The reflection layers in RAGNet are estimated without the participation of the transmission, hence showing inferior performance. YTMT can better cope with the highlights in the last row owing to its dual-stream design but is still limited by the linear assumption and cannot remove strong reflections. The method proposed by Dong et al. shows inferior", + "bbox": [ + 75, + 569, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "performance in both cases due to the difficulty of estimating the blending weight. Overall, our results are more visually favorable and contain fewer residual reflection components, which further reinforces our claims.", + "bbox": [ + 496, + 354, + 890, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further deliver a comparison of reflection layers in Fig. 5 to illustrate that our method can better reproduce the reflection scenes. It can be seen that the methods (Zhang et al., RAGNet, and YTMT) based on the additive model provide weak reflection maps. In the meantime, the methods based on the linear model with scalars (BDN and IB-CLN) and the one containing an alpha blending map can hardly avoid mixing the transmission components in reflection estimations. This phenomenon reflects the weakness of a fixed simplified model, requiring the reconstruction of more than two components. As shown by our results, introducing a residue term can significantly enhance the restoration quality of the reflection scene, and separates out the components beyond the additive relationship.", + "bbox": [ + 496, + 415, + 892, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 635, + 653, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To better analyze our improvements in the physical model and network structure, we carried out a detailed ablation study from the perspective of the reconstruction losses, feature interaction mechanisms, and feature encoders. We gather the results (PSNR and SSIM averaged on Real20 and SIR $^2$ datasets) in Table 2 and deliver a visual comparison in Fig. 6. In what follows we detail every point of them.", + "bbox": [ + 496, + 659, + 890, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reconstruction Loss. In view of the problems existing in the linear combination model, we use a learnable residue term to modify it, enhancing its flexibility so that it can cover more reflection scenes. In order to promote the learning of this residue term, we propose the reconstruction loss of residual correction. To illustrate the necessity of this setting, we perform an ablation study to compare the settings without reconstruction loss (w/o Recons. Loss) and using the linear reflection model (w/ Linear Recons.). It shows", + "bbox": [ + 496, + 765, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13144", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1333e793e07efcd6f663c828f2ef8ef784372781062c92fbf7fdbb884519706d.jpg", + "image_caption": [ + "w/o Recons. Loss" + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 300, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0b62487eafd26a254564f06105fcc1a8e27b87ea1c2c1084c6c3d8f186261603.jpg", + "image_caption": [ + "w/ Linear Recons." + ], + "image_footnote": [], + "bbox": [ + 300, + 88, + 480, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/99e2aa82a32ca4451814a27bb87f32eba06222be5f31d6aa2842c544155fa513.jpg", + "image_caption": [ + "w/o Feature Inter." + ], + "image_footnote": [], + "bbox": [ + 483, + 88, + 663, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/98be64bff4db26a49534933b374d2b2e5aa7900de78abaa956e3c32638f3b29f.jpg", + "image_caption": [ + "w/ YTMT Inter." + ], + "image_footnote": [], + "bbox": [ + 666, + 88, + 846, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eefe51869b7612f2bd7240a4d2b8f9b4fa5a3e6d29e456e10fac4a966ff92d06.jpg", + "image_caption": [ + "w/o Feature Enc." + ], + "image_footnote": [], + "bbox": [ + 119, + 196, + 300, + 285 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/465163df7e638b6f1e75711d10f4b5589651dcad877eda8e75e2d2fd2562da58.jpg", + "image_caption": [ + "w/ HyperColumn" + ], + "image_footnote": [], + "bbox": [ + 300, + 196, + 480, + 285 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd58a03ed6bfc8461cc78618a477a9a48f2bce6f3120555abd5b93b11b33a2a4.jpg", + "image_caption": [ + "Ours", + "Figure 6: The visual comparison of different settings involved in the ablation study." + ], + "image_footnote": [], + "bbox": [ + 483, + 196, + 663, + 285 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ffb652bc7a2d5885074497b446b50cd811500ae3c8b6d0d050f18e756dc6247.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 666, + 196, + 846, + 285 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/89c09bd080a8a411b794d8ed310ce0cf6d4d9c68f631ed9eb1d1594f37f015a9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsPSNRSSIM
w/o Recons. Loss24.840.897
w/ Linear Recons.22.210.881
w/o Feature Inter.24.810.895
w/ YTMT Inter.24.890.901
w/o Feature Enc.24.750.903
w/ HyperColumn23.990.894
Ours25.400.905
", + "bbox": [ + 143, + 356, + 405, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2: Ablation study on different configurations. The initial two rows compare the configuration devoid of any reconstruction loss and that integrating a linear reconstruction loss. Rows 3-4 offer a comparison between the settings lacking feature interaction and incorporating the YTMT feature interaction mechanism. Rows 5-6 shed light on configurations that exclude the feature encoder and utilize HyperColumn as the feature encoder. The last row shows the results of our full version.", + "bbox": [ + 75, + 500, + 468, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that the linear model is inferior to the setting without any reconstruction criterion because of the existence of $\\varepsilon$ . Meanwhile, without a reconstruction term, the model is likely to have trouble determining the portion of the components, excessively separating content to the reflection layer, resulting in textureless regions in the transmission layer as shown in Fig. 6.", + "bbox": [ + 75, + 655, + 468, + 761 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feature Interaction Mechanism. The feature interaction mechanism connects the two information streams in our proposed DSRNet, both exchanging and conditioning the dual-stream information, which facilitates the layer reconstruction. As shown in Table 2, the model performance appears to degrade after removing the feature interaction (w/o Feature Inter.), which tells us that the network can hardly achieve state-of-the-art performance via a two-branch network without feature interaction. Further, we include the", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "results of the YTMT feature interaction mechanism (w/ YTMT Inter.) by replacing MuGI with the YTMT mechanism. It can be seen that the model with the YTMT mechanism is only slightly better than the scheme without interaction, which demonstrates the superiority of the proposed MuGI mechanism.", + "bbox": [ + 496, + 358, + 890, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feature Encoder. In comparison to the hypercolumn framework introduced by the previous method (referred to as \"w/ HyperColumn\") and the approach that without using any feature encoder (\"w/o Feature Enc.\"), our method demonstrates noteworthy performance advantages. These advantages stem from two main factors. Firstly, our approach employs a hierarchical and gradual reduction of channel dimensions, followed by fusion operations. Secondly, DSFNet leverages a feature interaction mechanism, allowing high-level semantic features to interact before being fed into the subsequent stage.", + "bbox": [ + 496, + 450, + 892, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 632, + 617, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we modified the commonly-used linear model in the single image reflection separation task and proposed the reconstruction loss with residual correction. The proposed model is more flexible and effective compared with the previous methods. Meanwhile, we further improved the feature interaction mechanism in dual-stream networks and the usage of hierarchical semantic information, proposed the MuGI as a novel interaction paradigm, and a dual-stream semantic-aware network, namely DSR-Net. Extensive experiments revealed that our proposed method has achieved state-of-the-art performance on all real-world benchmark datasets and verified our contributions. In the future, further constraints are desired to be applied to the residue term to reduce its solution space, and more reflection models are hopefully to be covered by it to fit a wider range of reflection phenomena in the real world.", + "bbox": [ + 496, + 659, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13145", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Amit K. Agrawal, Ramesh Raskar, Shree K. Nayar, and Yuanzhen Li. Removing photography artifacts using gradient projection and flash-exposure sampling. TOG, 24(3):828-835, 2005. 2", + "[2] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In ECCV, pages 17-33, 2022. 2, 4", + "[3] Zheng Dong, Ke Xu, Yin Yang, Hujun Bao, Weiwei Xu, and Rynson W. H. Lau. Location-aware single image reflection removal. In ICCV, pages 4997-5006, 2021. 3, 5, 6, 7", + "[4] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes (VOC) challenge. *IJCV*, 88(2):303–338, 2010. 6", + "[5] Qingnan Fan, Jiaolong Yang, Gang Hua, Baoquan Chen, and David P. Wipf. A generic deep architecture for single image reflection removal and image smoothing. In ICCV, pages 3258-3267, 2017. 2, 3, 6", + "[6] Hany Farid and Edward H. Adelson. Separating reflections and lighting using independent components analysis. In CVPR, pages 1262-1267, 1999. 2", + "[7] Xin Feng, Wenjie Pei, Zihui Jia, Fanglin Chen, David Zhang, and Guangming Lu. Deep-masking generative network: A unified framework for background restoration from superimposed images. TIP, 30:4867-4882, 2021. 5, 6", + "[8] Kun Gai, Zhenwei Shi, and Changshui Zhang. Blind separation of superimposed moving images using image statistics. TPAMI, 34(1):19-32, 2012. 2", + "[9] Xiaojie Guo, Xiaochun Cao, and Yi Ma. Robust separation of reflection from multiple images. In CVPR, pages 2195-2202, 2014. 2", + "[10] Byeong-Ju Han and Jae-Young Sim. Reflection removal using low-rank matrix completion. In CVPR, pages 3872–3880, 2017. 2", + "[11] Bharath Hariharan, Pablo Andres Arbeláez, Ross B. Girshick, and Jitendra Malik. Hypercolumns for object segmentation and fine-grained localization. In CVPR, pages 447-456, 2015. 3", + "[12] Qiming Hu and Xiaojie Guo. *Trash or treasure? an interactive dual-stream strategy for single image reflection separation*. In *NeurIPS*, pages 24683–24694, 2021. 2, 3, 4, 5, 6, 7", + "[13] Naejin Kong, Yu-Wing Tai, and Joseph S. Shin. A physically-based approach to reflection separation: From physical modeling to constrained optimization. TPAMI, 36(2):209-221, 2014. 2", + "[14] Naejin Kong, Yu-Wing Tai, and Sung Yong Shin. High-quality reflection separation using polarized images. TIP, 20(12):3393-3405, 2011. 2", + "[15] Chenyang Lei and Qifeng Chen. Robust reflection removal with reflection-free flash-only cues. In CVPR, pages 14811-14820, 2021. 2", + "[16] Chenyang Lei, Xuhua Huang, Mengdi Zhang, Qiong Yan, Wenxiu Sun, and Qifeng Chen. Polarized reflection removal with perfect alignment in the wild. In CVPR, pages 1747-1755, 2020. 2" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Anat Levin and Yair Weiss. User assisted separation of reflections from a single image using a sparsity prior. TPAMI, 29(9):1647-1654, 2007. 1, 2", + "[18] Anat Levin, Assaf Zomet, and Yair Weiss. Learning to perceive transparency from the statistics of natural scenes. In NeurIPS, pages 1247-1254, 2002. 2", + "[19] Anat Levin, Assaf Zomet, and Yair Weiss. Separating reflections from a single image using local features. In CVPR, pages 306-313, 2004. 2", + "[20] Chao Li, Yixiao Yang, Kun He, Stephen Lin, and John E. Hopcroft. Single image reflection removal through cascaded refinement. In CVPR, pages 3562-3571, 2020. 3, 5, 6, 7", + "[21] Rui Li, Simeng Qiu, Guangming Zang, and Wolfgang Heidrich. Reflection separation via multi-bounce polarization state tracing. In ECCV, pages 781-796, 2020. 2", + "[22] Yu Li and Michael S. Brown. Exploiting reflection change for automatic reflection removal. In ICCV, pages 2432-2439, 2013. 2", + "[23] Yu Li and Michael S. Brown. Single image layer separation using relative smoothness. In CVPR, pages 2752-2759, 2014. 2", + "[24] Yu Li, Ming Liu, Yaling Yi, Qince Li, Dongwei Ren, and Wangmeng Zuo. Two-stage single image reflection removal with reflection-aware guidance. CoRR, abs/2012.00945, 2020. 3, 5, 6, 7", + "[25] Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Learning to see through obstructions. In CVPR, pages 14203-14212, 2020. 2", + "[26] Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Learning to see through obstructions with layered decomposition. TPAMI, 44(11):8387-8402, 2022. 2", + "[27] Youwei Lyu, Zhaopeng Cui, Si Li, Marc Pollefeys, and Boxin Shi. Reflection separation using a pair of unpolarized and polarized images. In NeurIPS, pages 14532-14542, 2019. 2", + "[28] Shree K. Nayar, Xi-Sheng Fang, and Terrance E. Boult. Separation of reflection components using color and polarization. IJCV, 21(3):163-186, 1997. 2", + "[29] Bernard Sarel and Michal Irani. Separating transparent layers of repetitive dynamic behaviors. In ICCV, pages 26-32, 2005. 2", + "[30] Christian Simon and In Kyu Park. Reflection removal for in-vehicle black box videos. In CVPR, pages 4231–4239, 2015. 2", + "[31] Sudipta N. Sinha, Johannes Kopf, Michael Goesele, Daniel Scharstein, and Richard Szeliski. Image-based rendering for scenes with reflections. TOG, 31(4):100:1-100:10, 2012. 1, 2", + "[32] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In CVPR, pages 5769-5780, 2022. 2", + "[33] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, and Alex C. Kot. Benchmarking single-image reflection removal algorithms. In ICCV, pages 3942-3950, 2017. 6" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13146", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, and Alex C Kot. Crrn: Multi-scale guided concurrent reflection removal network. In CVPR, pages 4777-4785, 2018. 1, 2", + "[35] Renjie Wan, Boxin Shi, Haoliang Li, Ling-Yu Duan, and Alex C. Kot. Reflection scene separation from a single image. In CVPR, pages 2395-2403, 2020. 1", + "[36] Kaixuan Wei, Jiaolong Yang, Ying Fu, David P. Wipf, and Hua Huang. Single image reflection removal exploiting misaligned training data and network enhancements. In CVPR, pages 8178-8187, 2019. 2, 3, 4, 5, 6, 7", + "[37] Qiang Wen, Yinjie Tan, Jing Qin, Wenxi Liu, Guoqiang Han, and Shengfeng He. Single image reflection removal beyond linearity. In CVPR, pages 3771-3779, 2019. 1, 2, 3", + "[38] Tianfan Xue, Michael Rubinstein, Ce Liu, and William T. Freeman. A computational approach for obstruction-free photography. TOG, 34(4):79:1-79:11, 2015. 2", + "[39] Jie Yang, Dong Gong, Lingqiao Liu, and Qinfeng Shi. Seeing deeply and bidirectionally: A deep learning approach for single image reflection removal. In ECCV, pages 675-691, 2018. 1, 2, 3, 5, 6, 7", + "[40] Jiaolong Yang, Hongdong Li, Yuchao Dai, and Robby T. Tan. Robust optical flow estimation of double-layer images under transparency or reflection. In CVPR, pages 1410-1419, 2016. 2", + "[41] Xuaner Cecilia Zhang, Ren Ng, and Qifeng Chen. Single image reflection separation with perceptual losses. In CVPR, pages 4786-4794, 2018. 2, 3, 4, 5, 6, 7", + "[42] Qian Zheng, Jinnan Chen, Zhan Lu, Boxin Shi, Xudong Jiang, Kim-Hui Yap, Ling-Yu Duan, and Alex C Kot. What does plate glass reveal about camera calibration? In CVPR, pages 3022-3032, 2020. 2", + "[43] Qian Zheng, Boxin Shi, Jinnan Chen, Xudong Jiang, LingYu Duan, and Alex C. Kot. Single image reflection removal with absorption effect. In CVPR, pages 13395-13404, 2021. 5, 6" + ], + "bbox": [ + 78, + 90, + 467, + 585 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13147", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Single Image Reflection Separation via Component Synergy/6169238f-a2a5-4863-b901-a9125cd1dec8_model.json b/2023/Single Image Reflection Separation via Component Synergy/6169238f-a2a5-4863-b901-a9125cd1dec8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fb4007412ce5e11299d341290871882f91ddfed9 --- /dev/null +++ b/2023/Single Image Reflection Separation via Component Synergy/6169238f-a2a5-4863-b901-a9125cd1dec8_model.json @@ -0,0 +1,2486 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.183, + 0.131, + 0.787, + 0.154 + ], + "angle": 0, + "content": "Single Image Reflection Separation via Component Synergy" + }, + { + "type": "text", + "bbox": [ + 0.382, + 0.181, + 0.594, + 0.199 + ], + "angle": 0, + "content": "Qiming Hu Xiaojie Guo*" + }, + { + "type": "text", + "bbox": [ + 0.193, + 0.2, + 0.777, + 0.218 + ], + "angle": 0, + "content": "College of Intelligence and Computing, Tianjin University, Tianjin, China" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.22, + 0.668, + 0.234 + ], + "angle": 0, + "content": "huqiming@tju.edu.cn, xj.max.quo@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.542 + ], + "angle": 0, + "content": "The reflection superposition phenomenon is complex and widely distributed in the real world, which derives various simplified linear and nonlinear formulations of the problem. In this paper, based on the investigation of the weaknesses of existing models, we propose a more general form of the superposition model by introducing a learnable residue term, which can effectively capture residual information during decomposition, guiding the separated layers to be complete. In order to fully capitalize on its advantages, we further design the network structure elaborately, including a novel dual-stream interaction mechanism and a powerful decomposition network with a semantic pyramid encoder. Extensive experiments and ablation studies are conducted to verify our superiority over state-of-the-art approaches on multiple real-world benchmark datasets. Our code is publicly available at https://github.com/mingcv/DSRNet." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.57, + 0.21, + 0.586 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.731 + ], + "angle": 0, + "content": "As a typical layer superimposition scenario, pictures captured through glass-like surfaces may be blended with undesired reflection, which not only impairs the aesthetic value but also hinders the downstream tasks [31]. In the meantime, scenes of interest are possibly concealed in reflection [35]. Therefore, both the parts transmitted through a surface (transmission layer, \\(\\mathbf{T}\\)) and those reflected (reflection layer, \\(\\mathbf{R}\\)) are desired to be reconstructed from a superimposed image \\(\\mathbf{I}\\) to fulfill the practical demands." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Following a common assumption [17], \\(\\mathbf{I}\\) is linearly composed by \\(\\mathbf{T}\\) and \\(\\mathbf{R}\\) expressed as \\(\\mathbf{I} = \\mathbf{T} + \\mathbf{R}\\), which has been popular due to its simplicity. However, in real-world scenarios, the reflection and transmission layers are likely to be weakened due to diffusion and other problems during the superimposition [35]. Therefore, several methods [34, 39] introduce scalars \\(\\alpha\\) and \\(\\beta\\) to the two components, respectively, obtaining \\(\\mathbf{I} = \\alpha \\mathbf{T} + \\beta \\mathbf{R}\\). However, the linear superimposition model is often violated due to over-exposure" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.267, + 0.58, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.328, + 0.547, + 0.34 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.269, + 0.656, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.594, + 0.328, + 0.642, + 0.341 + ], + "angle": 0, + "content": "T (GT)" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.269, + 0.733, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.328, + 0.718, + 0.342 + ], + "angle": 0, + "content": "R (GT)" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.269, + 0.812, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.328, + 0.794, + 0.342 + ], + "angle": 0, + "content": "Crops" + }, + { + "type": "image", + "bbox": [ + 0.815, + 0.269, + 0.889, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.817, + 0.327, + 0.886, + 0.341 + ], + "angle": 0, + "content": "\\(\\widetilde{\\mathbf{R}} = \\mathbf{I} - \\mathbf{T}\\)" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.343, + 0.657, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.415, + 0.569, + 0.427 + ], + "angle": 0, + "content": "\\(\\alpha \\mathbf{T} + \\mathbf{R}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.432, + 0.571, + 0.451 + ], + "angle": 0, + "content": "(IBCLN, CVPR,2020)" + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.402, + 0.657, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.402, + 0.733, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.343, + 0.812, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.402, + 0.812, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.821, + 0.343, + 0.881, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.408, + 0.88, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.473, + 0.574, + 0.487 + ], + "angle": 0, + "content": "WoT+R" + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.492, + 0.571, + 0.511 + ], + "angle": 0, + "content": "(Dong et al., ICCV,2021)" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.461, + 0.657, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.522, + 0.657, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.461, + 0.733, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.523, + 0.733, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.583, + 0.702, + 0.596 + ], + "angle": 0, + "content": "R" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.461, + 0.812, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.523, + 0.812, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.754, + 0.582, + 0.794, + 0.596 + ], + "angle": 0, + "content": "Crops" + }, + { + "type": "image", + "bbox": [ + 0.815, + 0.467, + 0.882, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.815, + 0.524, + 0.889, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.825, + 0.583, + 0.878, + 0.595 + ], + "angle": 0, + "content": "Residue" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.707 + ], + "angle": 0, + "content": "Figure 1: A visual example drawn from the \\(\\mathrm{SIR}^2\\) dataset. The first row displays a real input \\(\\mathbf{I}\\) with ground-truth \\(\\mathbf{T}\\) and \\(\\mathbf{R}\\). In rows 2-5, the separations based on different physical models are compared, where \\(\\hat{\\mathbf{T}}\\) and \\(\\hat{\\mathbf{R}}\\) represent the predictions. Our method is able to produce a \"Residue\" term, separating the nonlinearity from the layer reconstructions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.894, + 0.853 + ], + "angle": 0, + "content": "and other phenomena [37]. Thus, an alpha-matting map \\(\\mathbf{W}\\) is introduced to the model as \\(\\mathbf{I} = \\mathbf{W} \\circ \\mathbf{T} + \\bar{\\mathbf{W}} \\circ \\mathbf{R}\\) with \\(\\bar{\\mathbf{W}} = \\mathbf{1} - \\mathbf{W}\\), which, however, increases the degree of freedom and makes this ill-posed problem much harder. In view of all these limitations, it is not a trivial task to represent different kinds of reflection scenarios with a single model. In this work, we take a step forward in advancing the solution and deliver a more general form of the superimposition procedure by introducing a residue term as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.86, + 0.892, + 0.878 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} = \\tilde {\\mathbf {T}} + \\tilde {\\mathbf {R}} = \\mathbf {T} + \\mathbf {R} + \\Phi (\\mathbf {T}, \\mathbf {R}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{T}\\) and \\(\\mathbf{R}\\) denote the groundtruth scenes of the trans" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.469, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author. This work was supported by National Natural Science Foundation of China under Grant no. 62072327." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13138" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.497 + ], + "angle": 0, + "content": "mission and reflection layers, while their respective information contained in \\(\\mathbf{I}\\) after superimposition and other degradations and finally reach the camera sensors is represented by \\(\\tilde{\\mathbf{T}}\\) and \\(\\tilde{\\mathbf{R}}\\). \\(\\Phi(\\mathbf{T}, \\mathbf{R}) = \\mathbf{I} - \\mathbf{T} - \\mathbf{R}\\) is the residue of the reconstruction, which can be caused by attenuation, overexposure, etc. \\(\\Phi(\\cdot, \\cdot)\\) represents a group of functions that can be used to model the residue in specific situations. For example, \\(\\Phi(\\mathbf{T}, \\mathbf{R}) = 0\\) in [23, 5, 17, 41, 36, 12], \\(\\Phi(\\mathbf{T}, \\mathbf{R}) = (\\alpha - 1)\\mathbf{T} + (\\beta - 1)\\mathbf{R}\\) in [34, 39], \\(\\Phi(\\mathbf{T}, \\mathbf{R}) = -(\\bar{\\mathbf{W}} \\circ \\mathbf{T} + \\mathbf{W} \\circ \\mathbf{R})\\) in [37, 42]. In an effort to rule these different cases, a learnable module is desired to be introduced to model the residue \\(\\Phi(\\mathbf{T}, \\mathbf{R})\\). To clarify it, we provide a visual example in Fig. 1, where the first row shows a superimposed image \\(\\mathbf{I}\\), its groundtruth transmission \\(\\mathbf{T}\\) (GT) and reflection \\(\\mathbf{R}\\) (GT) layers collected by SIR\\(^2\\) dataset. According to the linear model, the guidance for reflection restoration is \\(\\tilde{\\mathbf{R}} = \\mathbf{I} - \\mathbf{T}\\) with incomplete signals rather than \\(\\mathbf{R}\\) (GT) that contains full information, which not only impairs the training of networks but also hinders the downstream usages. Therefore, as we point out in Eq. (1), a promising way out of this dilemma is to utilize ground-truth layers \\(\\mathbf{T}\\) and \\(\\mathbf{R}\\) as guiding signals and employ an extra residual term to handle the non-linearity (\"Residue\" in Fig. 1). As shown, the components that violate the linear assumption are separated from the layer predictions \\(\\hat{\\mathbf{T}}\\) and \\(\\hat{\\mathbf{R}}\\), and thus more complete and precious layer separations than previous methods are obtained." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.5, + 0.473, + 0.831 + ], + "angle": 0, + "content": "To further exploit the synergy between the components \\(\\mathbf{T}\\) and \\(\\mathbf{R}\\), we deliberate upon the facilitation of intercomponent feature interaction. The efficiency of a dual-stream interactive network has been verified in SIRS problem by [12]. Building upon their analysis, we subsequently advance a more effective scheme. Inspired by the gated mechanisms [2, 32], we present a simple yet effective mutually-gated interaction diagram (MuGI) serving as a better feature interaction candidate, by means of which we design the MuGI block to build our decomposition network. Moreover, the hypercolumn [41], as a commonly used semantic information encoder in the SIRS task, tends to aggregate features in a lossy way. Therefore, we propose a dual-stream pyramid fusion network (DSFNet) to replace it, which hierarchically decomposes and fuses the multiscale semantic information leveraging our proposed MuGI blocks and dual-stream fusion blocks (DSF Block). The roughly decomposed features of layers are further delivered into the dual-stream fine-grained decomposition phase (DSDNet). The two sub-networks constitute the main branch of our Dual-stream Semantic-aware network with Residual correction, namely DSRNet." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.832, + 0.435, + 0.846 + ], + "angle": 0, + "content": "In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "- We build a general form for SIRS by introducing a learnable residue term, which is more flexible to different scenarios and boosts the separation of layers;" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "- We exploit the synergy of features via mutually-gated interaction block within a dual-stream semantic-aware network, which facilitates information usage;" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.145, + 0.894, + 0.205 + ], + "angle": 0, + "content": "- Extensive experiments are conducted to demonstrate the effectiveness of our design. Overall, it achieves state-of-the-art performance against the alternatives on multiple real-world benchmarks for SIRS." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.092, + 0.894, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.222, + 0.642, + 0.238 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.248, + 0.892, + 0.34 + ], + "angle": 0, + "content": "Plenty of efforts have been devoted to handling the image reflection separation problem in the past decades. They either leverage multiple images to acquire more cues for the layer decomposition or exploit extra priors to manage the problem through a single image. In what follows, we organize the previous methods based on the images they require." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.893, + 0.567 + ], + "angle": 0, + "content": "Multiple Image Reflection Separation. In general, the transmission component tends to be unpolarized, yet the reflection component varies when rotating the polarization filter mounted in front of a camera sensor. Inspired by this phenomenon, a variety of methods [28, 14, 13, 27, 21, 16] resort to the physical solution that separates different components through a sequence of images with varied polarization orientations. Besides the polarization cues, the layer separations can also be indicated by different focuses [6], stereo information [31], flash on/off [1, 15], relative motions [29, 8, 22, 38, 40, 25, 26] and scene consistency [9, 30, 10]. Although involving multiple images mitigates the ill-posedness of the separation problem, this pipeline requires professional devices (such as polarizers) and manual operations, which limits its application." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.895, + 0.84 + ], + "angle": 0, + "content": "Single Image Reflection Separation. Compared with multiple-image solutions, single-image-based methods show more merits, being more flexible and requiring less manual operations. However, as there is no such thing as a free lunch, single-image schemes call for additional priors to cope with its ill-posedness nature, such as gradient-based constraints [19, 18, 17, 23] and manual annotations [17]. Specifically, Levin et al. [19, 18] impose sparse constraints on gradients to acquire the separations that have fewer edges. Levin and Weiss [17] demand several manual annotations of respective layers to obtain favorable decompositions, which involves additional human labor. Li and Brown [23] suppose that one layer is smoother than the other, thus applying unbalanced penalization on the gradients of the two layers to decompose them. Although these methods work fine in restricted scenarios, the complicated real-world conditions can considerably surpass their assumptions, leading to unsatisfactory results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.902 + ], + "angle": 0, + "content": "The advent of deep learning technology mitigates the incongruity between the volume of data and modeling capacity. Certain underlying assumptions can be subtly embedded in the process of data synthesis, and subsequently" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13139" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.895, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.325, + 0.895, + 0.388 + ], + "angle": 0, + "content": "Figure 2: The architecture of our proposed DSRNet, which consists of two cascaded stages and a learnable residue module (LRM). In Stage 1, DSFNet aggregates and separates hierarchical semantic information with two interactive feature streams. The roughly separated features are further fed into the DSDNet for fine-grained decompositions in Stage 2, where the LRM takes decomposed features, separating out the components that violate the linear assumption." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.473, + 0.898 + ], + "angle": 0, + "content": "learned by deep models. Concretely, CEILNet [5] imposes the relative smoothness prior to the synthesis of reflection layers and combines them with transmission layers by addition. An edge-aware network is developed to capture the transmission components, which yet ignores high-level semantics that are likely to aid the SIRS task. Zhang et al. [41] hence introduce HyperColumn features [11] using a pre-trained VGG-19 network to acquire semantic awareness, besides the perceptual and adversarial losses. Moreover, the exclusivity loss is developed to penalize the intersected gradients. ERRNet [36] goes a step further by leveraging an additional set of misaligned pairs. However, it appears to overlook the estimation of the reflection layer, which is essentially an image component and potentially important to distinguish the transmission parts. Li et al. [24] thereby presents a two-stage network (RAGNet) to first estimate reflection components and predict transmission ones guided by them, whereas the reflection estimation is isolated from the transmission in this way. Further, Hu and Guo [12] come up with the YTMT strategy, which pays equal attention to the two components, and a dual-stream interactive network is developed to restore the two layers simultaneously. However, their linear assumption makes the predicted reflection components tend to be weak in many cases. Other than them, BDN [39] and IBCLN [20] make use of reflection models weighted by scalars, and iteratively estimate both the components. This scheme prevents the reflection from being too weak, but its restorations often struggle to be free from the transmission parts. Furthermore, Wen et al. [37] simulate the nonlinear superimposition phenomenon by predicting a three-channel alpha blending weight map with the adversarial guidance of col" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.895, + 0.579 + ], + "angle": 0, + "content": "lected unpaired images. Dong et al. [3] develops an iterative network and estimates a probabilistic reflection confidence map in each iteration. However, it is not an easier task to estimate an alpha blending map than to determine each of the blended components, while the residual estimation only needs to retain the redundant information during decomposition, which is more practical than the former. Considering the drawbacks of the previous reflection models, we propose a general form of them with a learnable residue term, which is shown to be more effective and flexible." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.595, + 0.637, + 0.613 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.895, + 0.7 + ], + "angle": 0, + "content": "As shown in Fig. 2, our proposed DSRNet comprises two cascaded stages and a learnable residue module (LRM). In what follows, we first clarify the motivation of LRM and then introduce the MuGI block and the architecture of the DSRNet followed by its training loss functions." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.71, + 0.808, + 0.727 + ], + "angle": 0, + "content": "3.1. Residue of the Linear Combination" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "It often occurs that a superimposed image \\(\\mathbf{I}\\) cannot be perfectly represented by the linear combination of \\(\\mathbf{T}\\) and \\(\\mathbf{R}\\), leading to a residue term, as discussed in Section 1. Under its disturbance, the error of the additive reconstruction criterion \\(\\varepsilon = \\| \\mathbf{I} - (\\hat{\\mathbf{T}} +\\hat{\\mathbf{R}})\\|\\) remains large even though the predicted layers can well reconstruct the ground truths. Further minimizing \\(\\varepsilon\\) will put redundant information to either \\(\\hat{\\mathbf{T}}\\) or \\(\\hat{\\mathbf{R}}\\), which induces the deviation of predictions from their ground truths instead. Moreover, there are many kinds of simplified physical models proposed by previous work with their own shortcomings. Naturally, a unified model is de" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13140" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "sired to be put forward. Therefore, we introduce an extra residue term to offset the error in additive reconstruction, which, at the same time, unifies the different physical models. As depicted in Fig. 2, we leverage a learnable residue module (LRM) after Stage 2 to estimate the residual information during the decomposition by analyzing the features of \\(\\hat{\\mathbf{T}}\\) and \\(\\hat{\\mathbf{R}}\\) before the final layer. LRM consists of an interactive and a fusion part to collect and merge residual information from two branches. The dual-stream signals are concatenated before entering the fusion part. To constrain the space of residue, the \\(\\tanh(\\cdot)\\) function is utilized as the final activation. With the participation of LRM, we define the following reconstruction loss with residual rectification \\((\\mathrm{R}^3\\mathrm{Loss})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.313, + 0.47, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} := \\left\\| \\mathbf {I} - \\left(\\hat {\\mathbf {T}} + \\hat {\\mathbf {R}}\\right) - \\Phi (\\hat {\\mathbf {T}}, \\hat {\\mathbf {R}}) \\right\\| _ {1}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.47, + 0.416 + ], + "angle": 0, + "content": "where \\(\\Phi\\) denotes the LRM and \\(\\| \\cdot \\| _1\\) means the \\(\\ell_1\\) norm. Guided by this objective, the information beyond the additive model will flow to the residue term. Notably, this term can be totally discarded during the testing phase, avoiding extra computational costs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.425, + 0.37, + 0.441 + ], + "angle": 0, + "content": "3.2. Mutually-gated Interactive Block" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.471, + 0.539 + ], + "angle": 0, + "content": "Now that we employ the LRM to model the residual components during the additive reconstruction, the rest part can be seen as a linear model again as \\(\\tilde{\\mathbf{I}} = \\mathbf{I} - \\Phi (\\hat{\\mathbf{T}},\\hat{\\mathbf{R}}) = \\hat{\\mathbf{T}} +\\hat{\\mathbf{R}}\\). As displayed in Fig. 3 (a), we exploit the synergy between \\(\\hat{\\mathbf{T}}\\) and \\(\\hat{\\mathbf{R}}\\) by developing the Mutually-Gated Interactive mechanism, namely MuGI:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.548, + 0.47, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\hat {\\mathbf {F}} _ {\\mathbf {T}} = \\mathcal {G} _ {1} \\left(\\mathbf {F} _ {\\mathbf {T}}\\right) \\circ \\mathcal {G} _ {2} \\left(\\mathbf {F} _ {\\mathbf {R}}\\right); \\\\ \\hat {\\mathbf {F}} _ {\\mathbf {R}} = \\mathcal {G} _ {1} \\left(\\mathbf {F} _ {\\mathbf {R}}\\right) \\circ \\mathcal {G} _ {2} \\left(\\mathbf {F} _ {\\mathbf {T}}\\right), \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.825 + ], + "angle": 0, + "content": "where \\(\\mathbf{F_T},\\mathbf{F_R}\\in \\mathbb{R}^{H_1\\times W_1\\times C_1}\\) denote intermediate features of streams for predicting \\(\\hat{\\mathbf{T}}\\) and \\(\\hat{\\mathbf{R}}\\), respectively, and \\(\\hat{\\mathbf{F}}_{\\mathrm{T}},\\hat{\\mathbf{F}}_{\\mathrm{R}}\\in \\mathbb{R}^{H_2\\times W_2\\times C_2}\\) represent the outputs of the mutual gate. \\(\\mathcal{G}_1\\) and \\(\\mathcal{G}_2\\) are functions selecting which part of features to engage the interaction. \\(\\circ\\) indicates the elementwise multiplication. Here we use a simple implementation of MuGI to evaluate its capability of solving SIRS, which divides a feature map in half at the channel dimension, and \\(\\mathcal{G}_1(\\cdot)\\) always selects the former half, while \\(\\mathcal{G}_2(\\cdot)\\) chooses the latter. We have \\(H_{1} = H_{2}\\), \\(W_{1} = W_{2}\\) and \\(C_2 = C_1 / 2\\) in this way. Such a gated mechanism captures the mutual dependency of the two components in the feature space, say if either stream contains signals the other one desires, they tend to be split into the latter half (selected by \\(\\mathcal{G}_2(\\cdot)\\)) and transferred into the sibling stream by the gate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Our simple implementation of MuGI makes it plug-and-play. Therefore, in general, most of the single-path blocks can be converted into a dual-stream one by parallel setting two of them and relating them with MuGI. In this paper, we follow a state-of-the-art block design presented by [2]" + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.093, + 0.639, + 0.107 + ], + "angle": 0, + "content": "(a). MuGI Mechanism" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.108, + 0.882, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.181, + 0.608, + 0.194 + ], + "angle": 0, + "content": "(b). MuGI Block" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.197, + 0.887, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.366, + 0.848, + 0.381 + ], + "angle": 0, + "content": "Figure 3: MuGI mechanism and MuGI Block." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.41, + 0.892, + 0.547 + ], + "angle": 0, + "content": "and convert its single-path NAF block into a dual-stream mutually-gated interactive block (MuGI block). Its diagram is depicted in Fig. 3 (b), in which the simple gates are replaced by MuGI gates. Specifically, the layer normalization makes dual-stream features comparable, \\(1 \\times 1\\) convolution doubles the channel dimensions in order to conduct MuGI without information loss which reduces dimensions by half during the interaction, channel attention and the following \\(1 \\times 1\\) convolution serve as fusion/reweighting roles." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.831, + 0.575 + ], + "angle": 0, + "content": "3.3. Dual-stream Semantic-aware Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.903 + ], + "angle": 0, + "content": "The semantic information provided by pre-trained models is usually introduced to alleviate the impact of ill-posedness in SIRS. The HyperColumn [41, 36, 12], as a prevalent multi-scale semantic feature extractor in SIRS, aggregates semantics through interpolating the features extracted by a pre-trained model into the same scale as the input images. It then concatenates them together and employs a \\(1 \\times 1\\) convolution to rapidly reduce the channel dimensions (typically from 1475 to 64 or 256), before feeding them into the decomposition networks. This strategy omits the inner relationship of multi-scale features and may discard informative signals. To overcome the drawbacks mentioned above, we present a dual-stream pyramid fusion network (DSFNet), which hierarchically aggregates the extracted features by jointly upsampling and interacting. As shown in Fig. 2, given the hierarchical features extracted by a pre-trained deep network (e.g., VGGNet-19), DSFNet gathers them in a bottom-up manner with MuGI blocks and dual-stream fusion blocks (DSF Block), the structure of which is shown in Fig. 2. \"Dual Conv\" means two parallel convolutional layers. The features extracted by the pre" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13141" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.089, + 0.849, + 0.3 + ], + "angle": 0, + "content": "
MethodsReal20 (20)Objects (200)Postcard (199)Wild (55)Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Zhang et al. [41]22.550.78822.680.87916.810.79721.520.83220.080.835
BDN [39]18.410.72622.720.85620.710.85922.360.83021.650.849
ERRNet [36]22.890.80324.870.89622.040.87624.250.85323.530.879
IBCLN [20]21.860.76224.870.89323.390.87524.710.88624.100.879
RAGNet [24]22.950.79326.150.90323.670.87925.530.88024.900.886
DMGN [7]20.710.77024.980.89922.920.87723.810.83523.800.877
Zheng et al. [43]20.170.75525.200.88023.260.90525.390.87824.190.885
YTMT [12]23.260.80624.870.89622.910.88425.480.89024.050.886
Ours24.230.82026.280.91424.560.90825.680.89625.400.905
Dong et al.†[3]23.340.81224.360.89823.720.90325.730.90224.210.897
Ours†23.910.81826.740.92024.830.91126.110.90625.750.910
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.309, + 0.894, + 0.342 + ], + "angle": 0, + "content": "Table 1: Quantitative results on four real-world benchmark datasets of methods. The best results are indicated in **bold.†** indicates extra training data that are involved as in [3]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.473, + 0.624 + ], + "angle": 0, + "content": "trained network at each level first interact through a MuGI block, thereby features in the same scale are related and transformed into dual-stream features \\((\\mathbf{F_T} = \\mathbf{F_R}\\) for these blocks). After the interaction, the deepest features extracted by VGGNet are upscaled and then fused with shallower features through the DSF block, in which the dual-stream features come from the two adjacent scales are concatenated at channel dimension, fused by a 1x1 convolutional layer and then upscaled at each stream. The MuGI blocks are further employed to promote cross-scale interactions. Note that, shallow features extracted from RGB inputs are fused at the top of DSFNet, which preserves fine-grained details to restore both layers. The aggregated features provide rough separations of layers and are further refined by the second stage. The following stage (DSDNet) is constructed by the proposed MuGI blocks in a U-shaped manner. More details can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.631, + 0.223, + 0.646 + ], + "angle": 0, + "content": "3.4. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.654, + 0.47, + 0.715 + ], + "angle": 0, + "content": "Besides the \\(\\mathbf{R}^3\\) loss demonstrated in Section 3.1, we further introduce pixel, perceptual, and exclusion loss for pixel-wise and semantic fidelity as well as the gradient independence of layers, which are described as follows:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.715, + 0.47, + 0.775 + ], + "angle": 0, + "content": "Pixel Loss. The pixel loss is used to constrain the consistency of the reconstructed layer and groundtruth in both the natural image domain and gradient domain, minimizing their errors as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.779, + 0.47, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {p i x} := \\| \\hat {\\mathbf {T}} - \\mathbf {T} \\| _ {2} ^ {2} + \\| \\hat {\\mathbf {R}} - \\mathbf {R} \\| _ {2} ^ {2} \\tag {4} \\\\ + \\alpha (\\| \\nabla \\hat {\\mathbf {T}} - \\nabla \\mathbf {T} \\| _ {1} + \\| \\nabla \\hat {\\mathbf {R}} - \\nabla \\mathbf {R} \\| _ {1}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\(\\| \\cdot \\|_2\\) denotes the \\(\\ell_2\\) norm. \\(\\nabla\\) stands for the gradient operator, which extracts gradients in both vertical and horizontal directions. We express one of them in the formulation for clarity and average them in practice. \\(\\alpha\\) is set as 2 in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Perceptual Loss. The pixel-wise losses cannot assure the multi-scale consistency between predictions and their groundtruths, which may result in overwhelming punishment for the whole image due to small differences in local brightness. Therefore, the perceptual loss is further introduced as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.464, + 0.892, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p e r} := \\sum_ {i} \\omega_ {i} \\| \\phi_ {i} (\\hat {\\mathbf {T}}) - \\phi_ {i} (\\mathbf {T}) \\| _ {1}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.504, + 0.892, + 0.565 + ], + "angle": 0, + "content": "where \\(\\phi_i(\\cdot)\\) designates the features drawn by layer \\(i\\in\\) \\(\\{2,7,12,21,30\\}\\) of a VGG-19 model. \\(\\omega_{i}\\) s are combining weights of terms at different layers. We follow the setting of hyper-parameters in [36]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.565, + 0.892, + 0.611 + ], + "angle": 0, + "content": "Exclusion Loss. We introduce the exclusion loss to strengthen the gradient independence prior and reduce the structural coupling in estimated separations as below:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.619, + 0.892, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {e x c} := \\frac {1}{N} \\sum_ {n = 0} ^ {N - 1} \\| \\Psi \\left(\\hat {\\mathbf {T}} ^ {\\downarrow n}, \\hat {\\mathbf {R}} ^ {\\downarrow n}\\right) \\| _ {2} ^ {2}, \\tag {6} \\\\ \\Psi (\\hat {\\mathbf {T}}, \\hat {\\mathbf {R}}) := \\tanh \\left(\\eta_ {1} | \\nabla \\hat {\\mathbf {T}} |\\right) \\circ \\tanh \\left(\\eta_ {2} | \\nabla \\hat {\\mathbf {R}} |\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.698, + 0.892, + 0.745 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{T}}^{\\downarrow n}\\) and \\(\\hat{\\mathbf{R}}^{\\downarrow n}\\) represent taking down-sampling by \\(2^{n}\\) times (\\(2^{N}\\) at most) of \\(\\hat{\\mathbf{T}}\\) and \\(\\hat{\\mathbf{R}}\\). \\(\\eta_{1}\\) and \\(\\eta_{2}\\) are normalization factors, which are identical to [41]." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.745, + 0.887, + 0.76 + ], + "angle": 0, + "content": "Gathering all the loss terms yields the final objective as:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.77, + 0.892, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l l}} := \\mathcal {L} _ {\\text {p i x}} + \\beta_ {1} \\mathcal {L} _ {\\text {p e r}} + \\beta_ {2} \\mathcal {L} _ {\\text {e x c}} + \\beta_ {3} \\mathcal {L} _ {\\text {r e c}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.795, + 0.892, + 0.811 + ], + "angle": 0, + "content": "where \\(\\beta_{1} = 0.01\\), \\(\\beta_{2} = 1\\), and \\(\\beta_{3} = 0.2\\) are set empirically." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.822, + 0.731, + 0.839 + ], + "angle": 0, + "content": "4. Experimental Validation" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.718, + 0.864 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implemented in PyTorch, our models are optimized with the Adam optimizer on a single RTX 3090 GPU for 20" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13142" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.089, + 0.861, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.652, + 0.877, + 0.669 + ], + "angle": 0, + "content": "Figure 4: Visual comparison of estimated transmission layers between state-of-the-arts and ours on real-world samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.694, + 0.471, + 0.739 + ], + "angle": 0, + "content": "epochs at most to reach favorable overall performance. The learning rate is initialized as \\(10^{-4}\\) and fixed during the training phase with a batch size of 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.74, + 0.47, + 0.846 + ], + "angle": 0, + "content": "Dataset. Our training dataset embraces both real and synthesized images. Following [5, 12], the training dataset are composed of 90 real pairs from [41] and 7,643 synthesized pairs from the PASCAL VOC dataset [4]. For synthesized data, transmission and reflection layers are weakened by coefficients \\(\\gamma_{1} \\in [0.8, 1.0]\\) and \\(\\gamma_{2} \\in [0.4, 1.0]\\) during blending them with the following model:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.858, + 0.47, + 0.875 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {\\text {s y n}} = \\gamma_ {1} \\mathbf {T} _ {\\text {s y n}} + \\gamma_ {2} \\mathbf {R} _ {\\text {s y n}} - \\gamma_ {1} \\gamma_ {2} \\mathbf {T} _ {\\text {s y n}} \\circ \\mathbf {R} _ {\\text {s y n}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.471, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathbf{T}_{syn},\\mathbf{R}_{syn}\\) and \\(\\mathbf{I}_{syn}\\) represent the transmission, re" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.694, + 0.892, + 0.755 + ], + "angle": 0, + "content": "flection, and superimposed layers during synthesis, respectively. This formulation is inspired by the \"screen\" blending mode in digital image processing, which always reserves lighter colors for the blending layers." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.77, + 0.724, + 0.785 + ], + "angle": 0, + "content": "4.2. Performance Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Quantitative comparison. A comparison is made between state-of-the-art methods as shown in Table 1, involving Zhang et al. [41], BDN [39], ERRNet [36], IBCLN [20], RAGNet [24], DMGN [7], Zheng et al. [43], YTMT [12] and ours on four real-world dataset, including Real20 [41] and three subsets of the \\(\\mathrm{SIR}^2\\) Dataset [33]. Besides, to meet the configuration of [3], we train our DSRNet ad" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13143" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.252, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.173, + 0.193, + 0.185 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.089, + 0.407, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.173, + 0.381, + 0.186 + ], + "angle": 0, + "content": "Zhang et al. [41]" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.089, + 0.561, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.173, + 0.515, + 0.186 + ], + "angle": 0, + "content": "BDN [39]" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.089, + 0.714, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.173, + 0.674, + 0.186 + ], + "angle": 0, + "content": "IBCLN [20]" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.089, + 0.868, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.173, + 0.831, + 0.186 + ], + "angle": 0, + "content": "RAGNet [24]" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.187, + 0.252, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.271, + 0.212, + 0.285 + ], + "angle": 0, + "content": "YTMT [12]" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.187, + 0.407, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.284, + 0.271, + 0.374, + 0.285 + ], + "angle": 0, + "content": "Dong et al. [3]" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.187, + 0.561, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.272, + 0.499, + 0.284 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.187, + 0.714, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.271, + 0.682, + 0.284 + ], + "angle": 0, + "content": "Residue (ours)" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.187, + 0.868, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.272, + 0.802, + 0.283 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.297, + 0.892, + 0.328 + ], + "angle": 0, + "content": "Figure 5: Visual comparison of estimated reflections between state-of-the-arts and ours on a sample drawn from the \\(\\mathrm{SIR}^2\\) dataset. Ours has a significant advantage over other alternatives by separating the residue component from the estimations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.469, + 0.401 + ], + "angle": 0, + "content": "ditionally under their training settings, which includes 200 extra real pairs provided by the \"Nature\" dataset [20], and 13,700 synthesized image pairs provided by [41] instead." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.469, + 0.567 + ], + "angle": 0, + "content": "It turns out that our method shows its superiority over other competitors on all the testing datasets in both settings, gaining \\(0.50dB\\) and \\(1.54dB\\) in terms of average PSNR in the two settings, respectively. Given the four real-world datasets contains a variety of scenes, illumination conditions, and glass thickness, it is not a trivial task to achieve the best performance of all metrics on these datasets simultaneously. The experimental results demonstrate that our proposed SIRS scheme has significantly higher performance and stronger generalization ability, which explains our main contributions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Qualitative comparisons. To further explain our performance advantages, we provide visual comparisons of transmission layers in Fig. 4 against state-of-the-art methods, including Zhang et al. [41], BDN [39], ERRNet [36], IBCLN [20], RAGNet [24], YTMT [12], and Dong et al. [3]. As can be observed in Fig. 4, the method proposed by Zhang et al. fails to handle the case containing scattered reflection components in the second row. BDN has trouble dealing with the images with highlights like in the first row and even aggravates the reflections. The problem is likely to be caused by the linear reflection model, which lacks the ability to model specular highlights. For the results of ERRNet, which contains only a single branch to estimate transmission layers, it lacks the direct modeling of the reflections, therefore only removing parts of them. IBCLN also has trouble removing highlights and scattered reflections. The reflection layers in RAGNet are estimated without the participation of the transmission, hence showing inferior performance. YTMT can better cope with the highlights in the last row owing to its dual-stream design but is still limited by the linear assumption and cannot remove strong reflections. The method proposed by Dong et al. shows inferior" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.414 + ], + "angle": 0, + "content": "performance in both cases due to the difficulty of estimating the blending weight. Overall, our results are more visually favorable and contain fewer residual reflection components, which further reinforces our claims." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.893, + 0.627 + ], + "angle": 0, + "content": "We further deliver a comparison of reflection layers in Fig. 5 to illustrate that our method can better reproduce the reflection scenes. It can be seen that the methods (Zhang et al., RAGNet, and YTMT) based on the additive model provide weak reflection maps. In the meantime, the methods based on the linear model with scalars (BDN and IB-CLN) and the one containing an alpha blending map can hardly avoid mixing the transmission components in reflection estimations. This phenomenon reflects the weakness of a fixed simplified model, requiring the reconstruction of more than two components. As shown by our results, introducing a residue term can significantly enhance the restoration quality of the reflection scene, and separates out the components beyond the additive relationship." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.655, + 0.652 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.765 + ], + "angle": 0, + "content": "To better analyze our improvements in the physical model and network structure, we carried out a detailed ablation study from the perspective of the reconstruction losses, feature interaction mechanisms, and feature encoders. We gather the results (PSNR and SSIM averaged on Real20 and SIR\\(^2\\) datasets) in Table 2 and deliver a visual comparison in Fig. 6. In what follows we detail every point of them." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Reconstruction Loss. In view of the problems existing in the linear combination model, we use a learnable residue term to modify it, enhancing its flexibility so that it can cover more reflection scenes. In order to promote the learning of this residue term, we propose the reconstruction loss of residual correction. To illustrate the necessity of this setting, we perform an ablation study to compare the settings without reconstruction loss (w/o Recons. Loss) and using the linear reflection model (w/ Linear Recons.). It shows" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13144" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.089, + 0.301, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.157, + 0.183, + 0.264, + 0.195 + ], + "angle": 0, + "content": "w/o Recons. Loss" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.089, + 0.482, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.183, + 0.446, + 0.195 + ], + "angle": 0, + "content": "w/ Linear Recons." + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.089, + 0.665, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.183, + 0.627, + 0.195 + ], + "angle": 0, + "content": "w/o Feature Inter." + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.089, + 0.848, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.709, + 0.183, + 0.804, + 0.195 + ], + "angle": 0, + "content": "w/ YTMT Inter." + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.197, + 0.301, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.291, + 0.262, + 0.304 + ], + "angle": 0, + "content": "w/o Feature Enc." + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.197, + 0.482, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.291, + 0.445, + 0.305 + ], + "angle": 0, + "content": "w/ HyperColumn" + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.197, + 0.665, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.292, + 0.591, + 0.304 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.197, + 0.848, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.292, + 0.768, + 0.304 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.318, + 0.759, + 0.334 + ], + "angle": 0, + "content": "Figure 6: The visual comparison of different settings involved in the ablation study." + }, + { + "type": "table", + "bbox": [ + 0.144, + 0.357, + 0.406, + 0.489 + ], + "angle": 0, + "content": "
ModelsPSNRSSIM
w/o Recons. Loss24.840.897
w/ Linear Recons.22.210.881
w/o Feature Inter.24.810.895
w/ YTMT Inter.24.890.901
w/o Feature Enc.24.750.903
w/ HyperColumn23.990.894
Ours25.400.905
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.5, + 0.47, + 0.634 + ], + "angle": 0, + "content": "Table 2: Ablation study on different configurations. The initial two rows compare the configuration devoid of any reconstruction loss and that integrating a linear reconstruction loss. Rows 3-4 offer a comparison between the settings lacking feature interaction and incorporating the YTMT feature interaction mechanism. Rows 5-6 shed light on configurations that exclude the feature encoder and utilize HyperColumn as the feature encoder. The last row shows the results of our full version." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.469, + 0.762 + ], + "angle": 0, + "content": "that the linear model is inferior to the setting without any reconstruction criterion because of the existence of \\(\\varepsilon\\). Meanwhile, without a reconstruction term, the model is likely to have trouble determining the portion of the components, excessively separating content to the reflection layer, resulting in textureless regions in the transmission layer as shown in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Feature Interaction Mechanism. The feature interaction mechanism connects the two information streams in our proposed DSRNet, both exchanging and conditioning the dual-stream information, which facilitates the layer reconstruction. As shown in Table 2, the model performance appears to degrade after removing the feature interaction (w/o Feature Inter.), which tells us that the network can hardly achieve state-of-the-art performance via a two-branch network without feature interaction. Further, we include the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.892, + 0.448 + ], + "angle": 0, + "content": "results of the YTMT feature interaction mechanism (w/ YTMT Inter.) by replacing MuGI with the YTMT mechanism. It can be seen that the model with the YTMT mechanism is only slightly better than the scheme without interaction, which demonstrates the superiority of the proposed MuGI mechanism." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.451, + 0.893, + 0.618 + ], + "angle": 0, + "content": "Feature Encoder. In comparison to the hypercolumn framework introduced by the previous method (referred to as \"w/ HyperColumn\") and the approach that without using any feature encoder (\"w/o Feature Enc.\"), our method demonstrates noteworthy performance advantages. These advantages stem from two main factors. Firstly, our approach employs a hierarchical and gradual reduction of channel dimensions, followed by fusion operations. Secondly, DSFNet leverages a feature interaction mechanism, allowing high-level semantic features to interact before being fed into the subsequent stage." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.633, + 0.619, + 0.649 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In this paper, we modified the commonly-used linear model in the single image reflection separation task and proposed the reconstruction loss with residual correction. The proposed model is more flexible and effective compared with the previous methods. Meanwhile, we further improved the feature interaction mechanism in dual-stream networks and the usage of hierarchical semantic information, proposed the MuGI as a novel interaction paradigm, and a dual-stream semantic-aware network, namely DSR-Net. Extensive experiments revealed that our proposed method has achieved state-of-the-art performance on all real-world benchmark datasets and verified our contributions. In the future, further constraints are desired to be applied to the residue term to reduce its solution space, and more reflection models are hopefully to be covered by it to fit a wider range of reflection phenomena in the real world." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13145" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Amit K. Agrawal, Ramesh Raskar, Shree K. Nayar, and Yuanzhen Li. Removing photography artifacts using gradient projection and flash-exposure sampling. TOG, 24(3):828-835, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.47, + 0.211 + ], + "angle": 0, + "content": "[2] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In ECCV, pages 17-33, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.47, + 0.255 + ], + "angle": 0, + "content": "[3] Zheng Dong, Ke Xu, Yin Yang, Hujun Bao, Weiwei Xu, and Rynson W. H. Lau. Location-aware single image reflection removal. In ICCV, pages 4997-5006, 2021. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.468, + 0.31 + ], + "angle": 0, + "content": "[4] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes (VOC) challenge. *IJCV*, 88(2):303–338, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.468, + 0.367 + ], + "angle": 0, + "content": "[5] Qingnan Fan, Jiaolong Yang, Gang Hua, Baoquan Chen, and David P. Wipf. A generic deep architecture for single image reflection removal and image smoothing. In ICCV, pages 3258-3267, 2017. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.468, + 0.409 + ], + "angle": 0, + "content": "[6] Hany Farid and Edward H. Adelson. Separating reflections and lighting using independent components analysis. In CVPR, pages 1262-1267, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.468, + 0.465 + ], + "angle": 0, + "content": "[7] Xin Feng, Wenjie Pei, Zihui Jia, Fanglin Chen, David Zhang, and Guangming Lu. Deep-masking generative network: A unified framework for background restoration from superimposed images. TIP, 30:4867-4882, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.468, + 0.507 + ], + "angle": 0, + "content": "[8] Kun Gai, Zhenwei Shi, and Changshui Zhang. Blind separation of superimposed moving images using image statistics. TPAMI, 34(1):19-32, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.468, + 0.549 + ], + "angle": 0, + "content": "[9] Xiaojie Guo, Xiaochun Cao, and Yi Ma. Robust separation of reflection from multiple images. In CVPR, pages 2195-2202, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.551, + 0.468, + 0.592 + ], + "angle": 0, + "content": "[10] Byeong-Ju Han and Jae-Young Sim. Reflection removal using low-rank matrix completion. In CVPR, pages 3872–3880, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.593, + 0.468, + 0.647 + ], + "angle": 0, + "content": "[11] Bharath Hariharan, Pablo Andres Arbeláez, Ross B. Girshick, and Jitendra Malik. Hypercolumns for object segmentation and fine-grained localization. In CVPR, pages 447-456, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.468, + 0.702 + ], + "angle": 0, + "content": "[12] Qiming Hu and Xiaojie Guo. *Trash or treasure? an interactive dual-stream strategy for single image reflection separation*. In *NeurIPS*, pages 24683–24694, 2021. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.468, + 0.759 + ], + "angle": 0, + "content": "[13] Naejin Kong, Yu-Wing Tai, and Joseph S. Shin. A physically-based approach to reflection separation: From physical modeling to constrained optimization. TPAMI, 36(2):209-221, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.468, + 0.802 + ], + "angle": 0, + "content": "[14] Naejin Kong, Yu-Wing Tai, and Sung Yong Shin. High-quality reflection separation using polarized images. TIP, 20(12):3393-3405, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.468, + 0.844 + ], + "angle": 0, + "content": "[15] Chenyang Lei and Qifeng Chen. Robust reflection removal with reflection-free flash-only cues. In CVPR, pages 14811-14820, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[16] Chenyang Lei, Xuhua Huang, Mengdi Zhang, Qiong Yan, Wenxiu Sun, and Qifeng Chen. Polarized reflection removal with perfect alignment in the wild. In CVPR, pages 1747-1755, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[17] Anat Levin and Yair Weiss. User assisted separation of reflections from a single image using a sparsity prior. TPAMI, 29(9):1647-1654, 2007. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[18] Anat Levin, Assaf Zomet, and Yair Weiss. Learning to perceive transparency from the statistics of natural scenes. In NeurIPS, pages 1247-1254, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.18, + 0.892, + 0.221 + ], + "angle": 0, + "content": "[19] Anat Levin, Assaf Zomet, and Yair Weiss. Separating reflections from a single image using local features. In CVPR, pages 306-313, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.224, + 0.892, + 0.265 + ], + "angle": 0, + "content": "[20] Chao Li, Yixiao Yang, Kun He, Stephen Lin, and John E. Hopcroft. Single image reflection removal through cascaded refinement. In CVPR, pages 3562-3571, 2020. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.267, + 0.892, + 0.308 + ], + "angle": 0, + "content": "[21] Rui Li, Simeng Qiu, Guangming Zang, and Wolfgang Heidrich. Reflection separation via multi-bounce polarization state tracing. In ECCV, pages 781-796, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.311, + 0.892, + 0.352 + ], + "angle": 0, + "content": "[22] Yu Li and Michael S. Brown. Exploiting reflection change for automatic reflection removal. In ICCV, pages 2432-2439, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.354, + 0.892, + 0.395 + ], + "angle": 0, + "content": "[23] Yu Li and Michael S. Brown. Single image layer separation using relative smoothness. In CVPR, pages 2752-2759, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.398, + 0.892, + 0.452 + ], + "angle": 0, + "content": "[24] Yu Li, Ming Liu, Yaling Yi, Qince Li, Dongwei Ren, and Wangmeng Zuo. Two-stage single image reflection removal with reflection-aware guidance. CoRR, abs/2012.00945, 2020. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.497 + ], + "angle": 0, + "content": "[25] Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Learning to see through obstructions. In CVPR, pages 14203-14212, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.499, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[26] Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Learning to see through obstructions with layered decomposition. TPAMI, 44(11):8387-8402, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.556, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[27] Youwei Lyu, Zhaopeng Cui, Si Li, Marc Pollefeys, and Boxin Shi. Reflection separation using a pair of unpolarized and polarized images. In NeurIPS, pages 14532-14542, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.613, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[28] Shree K. Nayar, Xi-Sheng Fang, and Terrance E. Boult. Separation of reflection components using color and polarization. IJCV, 21(3):163-186, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.698 + ], + "angle": 0, + "content": "[29] Bernard Sarel and Michal Irani. Separating transparent layers of repetitive dynamic behaviors. In ICCV, pages 26-32, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.701, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[30] Christian Simon and In Kyu Park. Reflection removal for in-vehicle black box videos. In CVPR, pages 4231–4239, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[31] Sudipta N. Sinha, Johannes Kopf, Michael Goesele, Daniel Scharstein, and Richard Szeliski. Image-based rendering for scenes with reflections. TOG, 31(4):100:1-100:10, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[32] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In CVPR, pages 5769-5780, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[33] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, and Alex C. Kot. Benchmarking single-image reflection removal algorithms. In ICCV, pages 3942-3950, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13146" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[34] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, and Alex C Kot. Crrn: Multi-scale guided concurrent reflection removal network. In CVPR, pages 4777-4785, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.177 + ], + "angle": 0, + "content": "[35] Renjie Wan, Boxin Shi, Haoliang Li, Ling-Yu Duan, and Alex C. Kot. Reflection scene separation from a single image. In CVPR, pages 2395-2403, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[36] Kaixuan Wei, Jiaolong Yang, Ying Fu, David P. Wipf, and Hua Huang. Single image reflection removal exploiting misaligned training data and network enhancements. In CVPR, pages 8178-8187, 2019. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.468, + 0.276 + ], + "angle": 0, + "content": "[37] Qiang Wen, Yinjie Tan, Jing Qin, Wenxi Liu, Guoqiang Han, and Shengfeng He. Single image reflection removal beyond linearity. In CVPR, pages 3771-3779, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.278, + 0.468, + 0.319 + ], + "angle": 0, + "content": "[38] Tianfan Xue, Michael Rubinstein, Ce Liu, and William T. Freeman. A computational approach for obstruction-free photography. TOG, 34(4):79:1-79:11, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.321, + 0.468, + 0.375 + ], + "angle": 0, + "content": "[39] Jie Yang, Dong Gong, Lingqiao Liu, and Qinfeng Shi. Seeing deeply and bidirectionally: A deep learning approach for single image reflection removal. In ECCV, pages 675-691, 2018. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.377, + 0.468, + 0.431 + ], + "angle": 0, + "content": "[40] Jiaolong Yang, Hongdong Li, Yuchao Dai, and Robby T. Tan. Robust optical flow estimation of double-layer images under transparency or reflection. In CVPR, pages 1410-1419, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.434, + 0.468, + 0.475 + ], + "angle": 0, + "content": "[41] Xuaner Cecilia Zhang, Ren Ng, and Qifeng Chen. Single image reflection separation with perceptual losses. In CVPR, pages 4786-4794, 2018. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.477, + 0.468, + 0.532 + ], + "angle": 0, + "content": "[42] Qian Zheng, Jinnan Chen, Zhan Lu, Boxin Shi, Xudong Jiang, Kim-Hui Yap, Ling-Yu Duan, and Alex C Kot. What does plate glass reveal about camera calibration? In CVPR, pages 3022-3032, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.534, + 0.468, + 0.587 + ], + "angle": 0, + "content": "[43] Qian Zheng, Boxin Shi, Jinnan Chen, Xudong Jiang, LingYu Duan, and Alex C. Kot. Single image reflection removal with absorption effect. In CVPR, pages 13395-13404, 2021. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13147" + } + ] +] \ No newline at end of file diff --git a/2023/Single Image Reflection Separation via Component Synergy/6169238f-a2a5-4863-b901-a9125cd1dec8_origin.pdf b/2023/Single Image Reflection Separation via Component Synergy/6169238f-a2a5-4863-b901-a9125cd1dec8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4f9af8737065b5c66078f12d976f4142c8b55774 --- /dev/null +++ b/2023/Single Image Reflection Separation via Component Synergy/6169238f-a2a5-4863-b901-a9125cd1dec8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93fccec52792228a1f01f29630e19d3e2cf29953714d1bbd1b9a2542c468da28 +size 3006984 diff --git a/2023/Single Image Reflection Separation via Component Synergy/full.md b/2023/Single Image Reflection Separation via Component Synergy/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d97b046957dfb15815a8c8a1b6040b7531bf4765 --- /dev/null +++ b/2023/Single Image Reflection Separation via Component Synergy/full.md @@ -0,0 +1,350 @@ +# Single Image Reflection Separation via Component Synergy + +Qiming Hu Xiaojie Guo* + +College of Intelligence and Computing, Tianjin University, Tianjin, China + +huqiming@tju.edu.cn, xj.max.quo@gmail.com + +# Abstract + +The reflection superposition phenomenon is complex and widely distributed in the real world, which derives various simplified linear and nonlinear formulations of the problem. In this paper, based on the investigation of the weaknesses of existing models, we propose a more general form of the superposition model by introducing a learnable residue term, which can effectively capture residual information during decomposition, guiding the separated layers to be complete. In order to fully capitalize on its advantages, we further design the network structure elaborately, including a novel dual-stream interaction mechanism and a powerful decomposition network with a semantic pyramid encoder. Extensive experiments and ablation studies are conducted to verify our superiority over state-of-the-art approaches on multiple real-world benchmark datasets. Our code is publicly available at https://github.com/mingcv/DSRNet. + +# 1. Introduction + +As a typical layer superimposition scenario, pictures captured through glass-like surfaces may be blended with undesired reflection, which not only impairs the aesthetic value but also hinders the downstream tasks [31]. In the meantime, scenes of interest are possibly concealed in reflection [35]. Therefore, both the parts transmitted through a surface (transmission layer, $\mathbf{T}$ ) and those reflected (reflection layer, $\mathbf{R}$ ) are desired to be reconstructed from a superimposed image $\mathbf{I}$ to fulfill the practical demands. + +Following a common assumption [17], $\mathbf{I}$ is linearly composed by $\mathbf{T}$ and $\mathbf{R}$ expressed as $\mathbf{I} = \mathbf{T} + \mathbf{R}$ , which has been popular due to its simplicity. However, in real-world scenarios, the reflection and transmission layers are likely to be weakened due to diffusion and other problems during the superimposition [35]. Therefore, several methods [34, 39] introduce scalars $\alpha$ and $\beta$ to the two components, respectively, obtaining $\mathbf{I} = \alpha \mathbf{T} + \beta \mathbf{R}$ . However, the linear superimposition model is often violated due to over-exposure + +![](images/56033888929df1fdf01f854eea064f3699add69cf7861a5af3cfee24ae60856a.jpg) +1 + +![](images/5e54e280b6bb2ed473e36128b628896277f6b439bac2d5777531d79f7ea17d9b.jpg) +T (GT) + +![](images/e195b4812d272af10aec6a4a3fc2d927c302ae4ab9f5f26ca0915392de0c8d3f.jpg) +R (GT) + +![](images/86ee0166c729d4e5266ac42b2144ea8e06c051c58ff5bd1e030717eb281cce07.jpg) +Crops + +![](images/802dece7a144651c19a192ab710a3618c747b3c26a3a06bc767c91f970f94b9d.jpg) +$\widetilde{\mathbf{R}} = \mathbf{I} - \mathbf{T}$ + +![](images/1d7d00c836444e1eed4e6f51d155a1ce6162c62fe63603c379b392c702dd0457.jpg) + +![](images/3940ba43f6c6ad3cb4fe7cd80e4a15d68daf11ef1c4aa786677c21703f8dff88.jpg) +$\alpha \mathbf{T} + \mathbf{R}$ + +![](images/8be5773578d8e377a6b8052ab32d3c5a6aab1c6f2dbc64621e91d6b5821b2590.jpg) + +![](images/0597247a6e7e50f04d11b6a84fe15ee250e611dc156ff89cc2159bc147603e1c.jpg) + +![](images/e634e827a617c59fc25b85b37010c3f09ade9384bbfe0db24163efce5f7091e5.jpg) + +![](images/a521327a5a64758d209357b7498ff74004e3701e64339178c6c229c5161ac21f.jpg) + +![](images/cb6574b1db09669c5542194ffd894a15586491ffd4373b947c6073c1c2ba07a7.jpg) + +![](images/ea1e9750503db500715a77c1ae45a7b1f9f3115e6742e5706e62c39303c3e0d9.jpg) +(IBCLN, CVPR,2020) +WoT+R +(Dong et al., ICCV,2021) + +![](images/df5d9a3aae5e2511b328004d2cb70fcf424892a60610677108f3ab6f57f82c6a.jpg) +Figure 1: A visual example drawn from the $\mathrm{SIR}^2$ dataset. The first row displays a real input $\mathbf{I}$ with ground-truth $\mathbf{T}$ and $\mathbf{R}$ . In rows 2-5, the separations based on different physical models are compared, where $\hat{\mathbf{T}}$ and $\hat{\mathbf{R}}$ represent the predictions. Our method is able to produce a "Residue" term, separating the nonlinearity from the layer reconstructions. + +![](images/89194c8b48b6bbc71358e9e44f7a2e79e31dece73c35a930078eaf8b6cdce826.jpg) + +![](images/54eaa68eff54bf6efc27ca67174227a550c862193e5d01ec4eb8c58bad0347a7.jpg) +R + +![](images/2a797903b894441eb6d467b600a5b4484d47f917f52e82295d490826236ecba7.jpg) + +![](images/3702c3be7fe95cd56c097304b2349284bb7fca4fad51233e003cb9af86696456.jpg) +Crops + +![](images/761dced01cc93bd9a6e25375956da6716dfefa58e796facb26a7726de2f98b17.jpg) + +![](images/1d44f4ed51cf31336afabf11ce3e43995dd5f9350a395bdfa62645f84f5ff835.jpg) +Residue + +and other phenomena [37]. Thus, an alpha-matting map $\mathbf{W}$ is introduced to the model as $\mathbf{I} = \mathbf{W} \circ \mathbf{T} + \bar{\mathbf{W}} \circ \mathbf{R}$ with $\bar{\mathbf{W}} = \mathbf{1} - \mathbf{W}$ , which, however, increases the degree of freedom and makes this ill-posed problem much harder. In view of all these limitations, it is not a trivial task to represent different kinds of reflection scenarios with a single model. In this work, we take a step forward in advancing the solution and deliver a more general form of the superimposition procedure by introducing a residue term as follows: + +$$ +\mathbf {I} = \tilde {\mathbf {T}} + \tilde {\mathbf {R}} = \mathbf {T} + \mathbf {R} + \Phi (\mathbf {T}, \mathbf {R}), \tag {1} +$$ + +where $\mathbf{T}$ and $\mathbf{R}$ denote the groundtruth scenes of the trans + +mission and reflection layers, while their respective information contained in $\mathbf{I}$ after superimposition and other degradations and finally reach the camera sensors is represented by $\tilde{\mathbf{T}}$ and $\tilde{\mathbf{R}}$ . $\Phi(\mathbf{T}, \mathbf{R}) = \mathbf{I} - \mathbf{T} - \mathbf{R}$ is the residue of the reconstruction, which can be caused by attenuation, overexposure, etc. $\Phi(\cdot, \cdot)$ represents a group of functions that can be used to model the residue in specific situations. For example, $\Phi(\mathbf{T}, \mathbf{R}) = 0$ in [23, 5, 17, 41, 36, 12], $\Phi(\mathbf{T}, \mathbf{R}) = (\alpha - 1)\mathbf{T} + (\beta - 1)\mathbf{R}$ in [34, 39], $\Phi(\mathbf{T}, \mathbf{R}) = -(\bar{\mathbf{W}} \circ \mathbf{T} + \mathbf{W} \circ \mathbf{R})$ in [37, 42]. In an effort to rule these different cases, a learnable module is desired to be introduced to model the residue $\Phi(\mathbf{T}, \mathbf{R})$ . To clarify it, we provide a visual example in Fig. 1, where the first row shows a superimposed image $\mathbf{I}$ , its groundtruth transmission $\mathbf{T}$ (GT) and reflection $\mathbf{R}$ (GT) layers collected by SIR $^2$ dataset. According to the linear model, the guidance for reflection restoration is $\tilde{\mathbf{R}} = \mathbf{I} - \mathbf{T}$ with incomplete signals rather than $\mathbf{R}$ (GT) that contains full information, which not only impairs the training of networks but also hinders the downstream usages. Therefore, as we point out in Eq. (1), a promising way out of this dilemma is to utilize ground-truth layers $\mathbf{T}$ and $\mathbf{R}$ as guiding signals and employ an extra residual term to handle the non-linearity ("Residue" in Fig. 1). As shown, the components that violate the linear assumption are separated from the layer predictions $\hat{\mathbf{T}}$ and $\hat{\mathbf{R}}$ , and thus more complete and precious layer separations than previous methods are obtained. + +To further exploit the synergy between the components $\mathbf{T}$ and $\mathbf{R}$ , we deliberate upon the facilitation of intercomponent feature interaction. The efficiency of a dual-stream interactive network has been verified in SIRS problem by [12]. Building upon their analysis, we subsequently advance a more effective scheme. Inspired by the gated mechanisms [2, 32], we present a simple yet effective mutually-gated interaction diagram (MuGI) serving as a better feature interaction candidate, by means of which we design the MuGI block to build our decomposition network. Moreover, the hypercolumn [41], as a commonly used semantic information encoder in the SIRS task, tends to aggregate features in a lossy way. Therefore, we propose a dual-stream pyramid fusion network (DSFNet) to replace it, which hierarchically decomposes and fuses the multiscale semantic information leveraging our proposed MuGI blocks and dual-stream fusion blocks (DSF Block). The roughly decomposed features of layers are further delivered into the dual-stream fine-grained decomposition phase (DSDNet). The two sub-networks constitute the main branch of our Dual-stream Semantic-aware network with Residual correction, namely DSRNet. + +In summary, our main contributions are as follows: + +- We build a general form for SIRS by introducing a learnable residue term, which is more flexible to different scenarios and boosts the separation of layers; + +- We exploit the synergy of features via mutually-gated interaction block within a dual-stream semantic-aware network, which facilitates information usage; +- Extensive experiments are conducted to demonstrate the effectiveness of our design. Overall, it achieves state-of-the-art performance against the alternatives on multiple real-world benchmarks for SIRS. + +# 2. Related Work + +Plenty of efforts have been devoted to handling the image reflection separation problem in the past decades. They either leverage multiple images to acquire more cues for the layer decomposition or exploit extra priors to manage the problem through a single image. In what follows, we organize the previous methods based on the images they require. + +Multiple Image Reflection Separation. In general, the transmission component tends to be unpolarized, yet the reflection component varies when rotating the polarization filter mounted in front of a camera sensor. Inspired by this phenomenon, a variety of methods [28, 14, 13, 27, 21, 16] resort to the physical solution that separates different components through a sequence of images with varied polarization orientations. Besides the polarization cues, the layer separations can also be indicated by different focuses [6], stereo information [31], flash on/off [1, 15], relative motions [29, 8, 22, 38, 40, 25, 26] and scene consistency [9, 30, 10]. Although involving multiple images mitigates the ill-posedness of the separation problem, this pipeline requires professional devices (such as polarizers) and manual operations, which limits its application. + +Single Image Reflection Separation. Compared with multiple-image solutions, single-image-based methods show more merits, being more flexible and requiring less manual operations. However, as there is no such thing as a free lunch, single-image schemes call for additional priors to cope with its ill-posedness nature, such as gradient-based constraints [19, 18, 17, 23] and manual annotations [17]. Specifically, Levin et al. [19, 18] impose sparse constraints on gradients to acquire the separations that have fewer edges. Levin and Weiss [17] demand several manual annotations of respective layers to obtain favorable decompositions, which involves additional human labor. Li and Brown [23] suppose that one layer is smoother than the other, thus applying unbalanced penalization on the gradients of the two layers to decompose them. Although these methods work fine in restricted scenarios, the complicated real-world conditions can considerably surpass their assumptions, leading to unsatisfactory results. + +The advent of deep learning technology mitigates the incongruity between the volume of data and modeling capacity. Certain underlying assumptions can be subtly embedded in the process of data synthesis, and subsequently + +![](images/ec01cc870caafaa657d0cea161c2e87a94fff8633972795cd349c32bb1846eae.jpg) +Figure 2: The architecture of our proposed DSRNet, which consists of two cascaded stages and a learnable residue module (LRM). In Stage 1, DSFNet aggregates and separates hierarchical semantic information with two interactive feature streams. The roughly separated features are further fed into the DSDNet for fine-grained decompositions in Stage 2, where the LRM takes decomposed features, separating out the components that violate the linear assumption. + +learned by deep models. Concretely, CEILNet [5] imposes the relative smoothness prior to the synthesis of reflection layers and combines them with transmission layers by addition. An edge-aware network is developed to capture the transmission components, which yet ignores high-level semantics that are likely to aid the SIRS task. Zhang et al. [41] hence introduce HyperColumn features [11] using a pre-trained VGG-19 network to acquire semantic awareness, besides the perceptual and adversarial losses. Moreover, the exclusivity loss is developed to penalize the intersected gradients. ERRNet [36] goes a step further by leveraging an additional set of misaligned pairs. However, it appears to overlook the estimation of the reflection layer, which is essentially an image component and potentially important to distinguish the transmission parts. Li et al. [24] thereby presents a two-stage network (RAGNet) to first estimate reflection components and predict transmission ones guided by them, whereas the reflection estimation is isolated from the transmission in this way. Further, Hu and Guo [12] come up with the YTMT strategy, which pays equal attention to the two components, and a dual-stream interactive network is developed to restore the two layers simultaneously. However, their linear assumption makes the predicted reflection components tend to be weak in many cases. Other than them, BDN [39] and IBCLN [20] make use of reflection models weighted by scalars, and iteratively estimate both the components. This scheme prevents the reflection from being too weak, but its restorations often struggle to be free from the transmission parts. Furthermore, Wen et al. [37] simulate the nonlinear superimposition phenomenon by predicting a three-channel alpha blending weight map with the adversarial guidance of col + +lected unpaired images. Dong et al. [3] develops an iterative network and estimates a probabilistic reflection confidence map in each iteration. However, it is not an easier task to estimate an alpha blending map than to determine each of the blended components, while the residual estimation only needs to retain the redundant information during decomposition, which is more practical than the former. Considering the drawbacks of the previous reflection models, we propose a general form of them with a learnable residue term, which is shown to be more effective and flexible. + +# 3. Methodology + +As shown in Fig. 2, our proposed DSRNet comprises two cascaded stages and a learnable residue module (LRM). In what follows, we first clarify the motivation of LRM and then introduce the MuGI block and the architecture of the DSRNet followed by its training loss functions. + +# 3.1. Residue of the Linear Combination + +It often occurs that a superimposed image $\mathbf{I}$ cannot be perfectly represented by the linear combination of $\mathbf{T}$ and $\mathbf{R}$ , leading to a residue term, as discussed in Section 1. Under its disturbance, the error of the additive reconstruction criterion $\varepsilon = \| \mathbf{I} - (\hat{\mathbf{T}} +\hat{\mathbf{R}})\|$ remains large even though the predicted layers can well reconstruct the ground truths. Further minimizing $\varepsilon$ will put redundant information to either $\hat{\mathbf{T}}$ or $\hat{\mathbf{R}}$ , which induces the deviation of predictions from their ground truths instead. Moreover, there are many kinds of simplified physical models proposed by previous work with their own shortcomings. Naturally, a unified model is de + +sired to be put forward. Therefore, we introduce an extra residue term to offset the error in additive reconstruction, which, at the same time, unifies the different physical models. As depicted in Fig. 2, we leverage a learnable residue module (LRM) after Stage 2 to estimate the residual information during the decomposition by analyzing the features of $\hat{\mathbf{T}}$ and $\hat{\mathbf{R}}$ before the final layer. LRM consists of an interactive and a fusion part to collect and merge residual information from two branches. The dual-stream signals are concatenated before entering the fusion part. To constrain the space of residue, the $\tanh(\cdot)$ function is utilized as the final activation. With the participation of LRM, we define the following reconstruction loss with residual rectification $(\mathrm{R}^3\mathrm{Loss})$ : + +$$ +\mathcal {L} _ {r e c} := \left\| \mathbf {I} - \left(\hat {\mathbf {T}} + \hat {\mathbf {R}}\right) - \Phi (\hat {\mathbf {T}}, \hat {\mathbf {R}}) \right\| _ {1}, \tag {2} +$$ + +where $\Phi$ denotes the LRM and $\| \cdot \| _1$ means the $\ell_1$ norm. Guided by this objective, the information beyond the additive model will flow to the residue term. Notably, this term can be totally discarded during the testing phase, avoiding extra computational costs. + +# 3.2. Mutually-gated Interactive Block + +Now that we employ the LRM to model the residual components during the additive reconstruction, the rest part can be seen as a linear model again as $\tilde{\mathbf{I}} = \mathbf{I} - \Phi (\hat{\mathbf{T}},\hat{\mathbf{R}}) = \hat{\mathbf{T}} +\hat{\mathbf{R}}$ . As displayed in Fig. 3 (a), we exploit the synergy between $\hat{\mathbf{T}}$ and $\hat{\mathbf{R}}$ by developing the Mutually-Gated Interactive mechanism, namely MuGI: + +$$ +\left\{ \begin{array}{l} \hat {\mathbf {F}} _ {\mathbf {T}} = \mathcal {G} _ {1} \left(\mathbf {F} _ {\mathbf {T}}\right) \circ \mathcal {G} _ {2} \left(\mathbf {F} _ {\mathbf {R}}\right); \\ \hat {\mathbf {F}} _ {\mathbf {R}} = \mathcal {G} _ {1} \left(\mathbf {F} _ {\mathbf {R}}\right) \circ \mathcal {G} _ {2} \left(\mathbf {F} _ {\mathbf {T}}\right), \end{array} \right. \tag {3} +$$ + +where $\mathbf{F_T},\mathbf{F_R}\in \mathbb{R}^{H_1\times W_1\times C_1}$ denote intermediate features of streams for predicting $\hat{\mathbf{T}}$ and $\hat{\mathbf{R}}$ , respectively, and $\hat{\mathbf{F}}_{\mathrm{T}},\hat{\mathbf{F}}_{\mathrm{R}}\in \mathbb{R}^{H_2\times W_2\times C_2}$ represent the outputs of the mutual gate. $\mathcal{G}_1$ and $\mathcal{G}_2$ are functions selecting which part of features to engage the interaction. $\circ$ indicates the elementwise multiplication. Here we use a simple implementation of MuGI to evaluate its capability of solving SIRS, which divides a feature map in half at the channel dimension, and $\mathcal{G}_1(\cdot)$ always selects the former half, while $\mathcal{G}_2(\cdot)$ chooses the latter. We have $H_{1} = H_{2}$ , $W_{1} = W_{2}$ and $C_2 = C_1 / 2$ in this way. Such a gated mechanism captures the mutual dependency of the two components in the feature space, say if either stream contains signals the other one desires, they tend to be split into the latter half (selected by $\mathcal{G}_2(\cdot)$ ) and transferred into the sibling stream by the gate. + +Our simple implementation of MuGI makes it plug-and-play. Therefore, in general, most of the single-path blocks can be converted into a dual-stream one by parallel setting two of them and relating them with MuGI. In this paper, we follow a state-of-the-art block design presented by [2] + +![](images/f543ebc8190b575d8149e22d166c0f824467140a4ab24ad36f478e445e0d2c28.jpg) +(a). MuGI Mechanism + +![](images/0d629e279f96afb3c67f50b88c529afdca7bd39dad85154d2ef385166dd5cf3e.jpg) +(b). MuGI Block +Figure 3: MuGI mechanism and MuGI Block. + +and convert its single-path NAF block into a dual-stream mutually-gated interactive block (MuGI block). Its diagram is depicted in Fig. 3 (b), in which the simple gates are replaced by MuGI gates. Specifically, the layer normalization makes dual-stream features comparable, $1 \times 1$ convolution doubles the channel dimensions in order to conduct MuGI without information loss which reduces dimensions by half during the interaction, channel attention and the following $1 \times 1$ convolution serve as fusion/reweighting roles. + +# 3.3. Dual-stream Semantic-aware Network + +The semantic information provided by pre-trained models is usually introduced to alleviate the impact of ill-posedness in SIRS. The HyperColumn [41, 36, 12], as a prevalent multi-scale semantic feature extractor in SIRS, aggregates semantics through interpolating the features extracted by a pre-trained model into the same scale as the input images. It then concatenates them together and employs a $1 \times 1$ convolution to rapidly reduce the channel dimensions (typically from 1475 to 64 or 256), before feeding them into the decomposition networks. This strategy omits the inner relationship of multi-scale features and may discard informative signals. To overcome the drawbacks mentioned above, we present a dual-stream pyramid fusion network (DSFNet), which hierarchically aggregates the extracted features by jointly upsampling and interacting. As shown in Fig. 2, given the hierarchical features extracted by a pre-trained deep network (e.g., VGGNet-19), DSFNet gathers them in a bottom-up manner with MuGI blocks and dual-stream fusion blocks (DSF Block), the structure of which is shown in Fig. 2. "Dual Conv" means two parallel convolutional layers. The features extracted by the pre + +
MethodsReal20 (20)Objects (200)Postcard (199)Wild (55)Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Zhang et al. [41]22.550.78822.680.87916.810.79721.520.83220.080.835
BDN [39]18.410.72622.720.85620.710.85922.360.83021.650.849
ERRNet [36]22.890.80324.870.89622.040.87624.250.85323.530.879
IBCLN [20]21.860.76224.870.89323.390.87524.710.88624.100.879
RAGNet [24]22.950.79326.150.90323.670.87925.530.88024.900.886
DMGN [7]20.710.77024.980.89922.920.87723.810.83523.800.877
Zheng et al. [43]20.170.75525.200.88023.260.90525.390.87824.190.885
YTMT [12]23.260.80624.870.89622.910.88425.480.89024.050.886
Ours24.230.82026.280.91424.560.90825.680.89625.400.905
Dong et al.†[3]23.340.81224.360.89823.720.90325.730.90224.210.897
Ours†23.910.81826.740.92024.830.91126.110.90625.750.910
+ +Table 1: Quantitative results on four real-world benchmark datasets of methods. The best results are indicated in **bold.†** indicates extra training data that are involved as in [3]. + +trained network at each level first interact through a MuGI block, thereby features in the same scale are related and transformed into dual-stream features $(\mathbf{F_T} = \mathbf{F_R}$ for these blocks). After the interaction, the deepest features extracted by VGGNet are upscaled and then fused with shallower features through the DSF block, in which the dual-stream features come from the two adjacent scales are concatenated at channel dimension, fused by a 1x1 convolutional layer and then upscaled at each stream. The MuGI blocks are further employed to promote cross-scale interactions. Note that, shallow features extracted from RGB inputs are fused at the top of DSFNet, which preserves fine-grained details to restore both layers. The aggregated features provide rough separations of layers and are further refined by the second stage. The following stage (DSDNet) is constructed by the proposed MuGI blocks in a U-shaped manner. More details can be found in the supplementary materials. + +# 3.4. Loss Function + +Besides the $\mathbf{R}^3$ loss demonstrated in Section 3.1, we further introduce pixel, perceptual, and exclusion loss for pixel-wise and semantic fidelity as well as the gradient independence of layers, which are described as follows: + +Pixel Loss. The pixel loss is used to constrain the consistency of the reconstructed layer and groundtruth in both the natural image domain and gradient domain, minimizing their errors as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {p i x} := \| \hat {\mathbf {T}} - \mathbf {T} \| _ {2} ^ {2} + \| \hat {\mathbf {R}} - \mathbf {R} \| _ {2} ^ {2} \tag {4} \\ + \alpha (\| \nabla \hat {\mathbf {T}} - \nabla \mathbf {T} \| _ {1} + \| \nabla \hat {\mathbf {R}} - \nabla \mathbf {R} \| _ {1}), \\ \end{array} +$$ + +where $\| \cdot \|_2$ denotes the $\ell_2$ norm. $\nabla$ stands for the gradient operator, which extracts gradients in both vertical and horizontal directions. We express one of them in the formulation for clarity and average them in practice. $\alpha$ is set as 2 in our experiments. + +Perceptual Loss. The pixel-wise losses cannot assure the multi-scale consistency between predictions and their groundtruths, which may result in overwhelming punishment for the whole image due to small differences in local brightness. Therefore, the perceptual loss is further introduced as follows: + +$$ +\mathcal {L} _ {p e r} := \sum_ {i} \omega_ {i} \| \phi_ {i} (\hat {\mathbf {T}}) - \phi_ {i} (\mathbf {T}) \| _ {1}, \tag {5} +$$ + +where $\phi_i(\cdot)$ designates the features drawn by layer $i\in$ $\{2,7,12,21,30\}$ of a VGG-19 model. $\omega_{i}$ s are combining weights of terms at different layers. We follow the setting of hyper-parameters in [36]. + +Exclusion Loss. We introduce the exclusion loss to strengthen the gradient independence prior and reduce the structural coupling in estimated separations as below: + +$$ +\begin{array}{l} \mathcal {L} _ {e x c} := \frac {1}{N} \sum_ {n = 0} ^ {N - 1} \| \Psi \left(\hat {\mathbf {T}} ^ {\downarrow n}, \hat {\mathbf {R}} ^ {\downarrow n}\right) \| _ {2} ^ {2}, \tag {6} \\ \Psi (\hat {\mathbf {T}}, \hat {\mathbf {R}}) := \tanh \left(\eta_ {1} | \nabla \hat {\mathbf {T}} |\right) \circ \tanh \left(\eta_ {2} | \nabla \hat {\mathbf {R}} |\right), \\ \end{array} +$$ + +where $\hat{\mathbf{T}}^{\downarrow n}$ and $\hat{\mathbf{R}}^{\downarrow n}$ represent taking down-sampling by $2^{n}$ times ( $2^{N}$ at most) of $\hat{\mathbf{T}}$ and $\hat{\mathbf{R}}$ . $\eta_{1}$ and $\eta_{2}$ are normalization factors, which are identical to [41]. + +Gathering all the loss terms yields the final objective as: + +$$ +\mathcal {L} _ {\text {a l l}} := \mathcal {L} _ {\text {p i x}} + \beta_ {1} \mathcal {L} _ {\text {p e r}} + \beta_ {2} \mathcal {L} _ {\text {e x c}} + \beta_ {3} \mathcal {L} _ {\text {r e c}}, \tag {7} +$$ + +where $\beta_{1} = 0.01$ , $\beta_{2} = 1$ , and $\beta_{3} = 0.2$ are set empirically. + +# 4. Experimental Validation + +# 4.1. Implementation Details + +Implemented in PyTorch, our models are optimized with the Adam optimizer on a single RTX 3090 GPU for 20 + +![](images/b36c30d0e3a868cbcd511404139932fa80e9d12352846d23f6f5fef772b9d1fd.jpg) +Figure 4: Visual comparison of estimated transmission layers between state-of-the-arts and ours on real-world samples. + +epochs at most to reach favorable overall performance. The learning rate is initialized as $10^{-4}$ and fixed during the training phase with a batch size of 1. + +Dataset. Our training dataset embraces both real and synthesized images. Following [5, 12], the training dataset are composed of 90 real pairs from [41] and 7,643 synthesized pairs from the PASCAL VOC dataset [4]. For synthesized data, transmission and reflection layers are weakened by coefficients $\gamma_{1} \in [0.8, 1.0]$ and $\gamma_{2} \in [0.4, 1.0]$ during blending them with the following model: + +$$ +\mathbf {I} _ {\text {s y n}} = \gamma_ {1} \mathbf {T} _ {\text {s y n}} + \gamma_ {2} \mathbf {R} _ {\text {s y n}} - \gamma_ {1} \gamma_ {2} \mathbf {T} _ {\text {s y n}} \circ \mathbf {R} _ {\text {s y n}}, \tag {8} +$$ + +where $\mathbf{T}_{syn},\mathbf{R}_{syn}$ and $\mathbf{I}_{syn}$ represent the transmission, re + +flection, and superimposed layers during synthesis, respectively. This formulation is inspired by the "screen" blending mode in digital image processing, which always reserves lighter colors for the blending layers. + +# 4.2. Performance Evaluation + +Quantitative comparison. A comparison is made between state-of-the-art methods as shown in Table 1, involving Zhang et al. [41], BDN [39], ERRNet [36], IBCLN [20], RAGNet [24], DMGN [7], Zheng et al. [43], YTMT [12] and ours on four real-world dataset, including Real20 [41] and three subsets of the $\mathrm{SIR}^2$ Dataset [33]. Besides, to meet the configuration of [3], we train our DSRNet ad + +![](images/f931fc59e8419156fdeed9c2cc088dc884a3b229dba2bd65a9d19f630bbf4662.jpg) +Input + +![](images/bb297128c2955cc666671a00763ccd0c157f4150567ddc55778a802d132a1ed1.jpg) + +![](images/f51a50c0aabb19973dde4259174388174a86125dff3749c2d4bf7094f0a8258c.jpg) + +![](images/8958afe7f5e3ddee1b14c3bd566f43b163bee46a37f8f446ad7d85dcc12091d2.jpg) + +![](images/7013d71ceeadfb10017ac1f6dc69ac322446ed27aef54f63b9db6d5396fddbcc.jpg) + +![](images/f5d2b60917c97df8636b27eeae956475412041c5e9060a62d158b6bac4077b57.jpg) +YTMT [12] + +![](images/1f5ef5a078f8e9aec5f972bdd91a453f99b1a7d68d49747621c5797ffa3af1e5.jpg) +Zhang et al. [41] +Dong et al. [3] + +![](images/2606e1aad110ba01058e13f39177897efd92ed786618c8deb42074dfe0029ff0.jpg) +BDN [39] +Ours +Figure 5: Visual comparison of estimated reflections between state-of-the-arts and ours on a sample drawn from the $\mathrm{SIR}^2$ dataset. Ours has a significant advantage over other alternatives by separating the residue component from the estimations. + +![](images/7a16d8ec40f48c74b778c09db3304f953905e0e5f4858df6baa85dcef82b779d.jpg) +IBCLN [20] +Residue (ours) + +![](images/14e79a74f5fe1a00dde28020204561efbc8c1e17ba44715add88a06d657abc98.jpg) +RAGNet [24] +GT + +ditionally under their training settings, which includes 200 extra real pairs provided by the "Nature" dataset [20], and 13,700 synthesized image pairs provided by [41] instead. + +It turns out that our method shows its superiority over other competitors on all the testing datasets in both settings, gaining $0.50dB$ and $1.54dB$ in terms of average PSNR in the two settings, respectively. Given the four real-world datasets contains a variety of scenes, illumination conditions, and glass thickness, it is not a trivial task to achieve the best performance of all metrics on these datasets simultaneously. The experimental results demonstrate that our proposed SIRS scheme has significantly higher performance and stronger generalization ability, which explains our main contributions. + +Qualitative comparisons. To further explain our performance advantages, we provide visual comparisons of transmission layers in Fig. 4 against state-of-the-art methods, including Zhang et al. [41], BDN [39], ERRNet [36], IBCLN [20], RAGNet [24], YTMT [12], and Dong et al. [3]. As can be observed in Fig. 4, the method proposed by Zhang et al. fails to handle the case containing scattered reflection components in the second row. BDN has trouble dealing with the images with highlights like in the first row and even aggravates the reflections. The problem is likely to be caused by the linear reflection model, which lacks the ability to model specular highlights. For the results of ERRNet, which contains only a single branch to estimate transmission layers, it lacks the direct modeling of the reflections, therefore only removing parts of them. IBCLN also has trouble removing highlights and scattered reflections. The reflection layers in RAGNet are estimated without the participation of the transmission, hence showing inferior performance. YTMT can better cope with the highlights in the last row owing to its dual-stream design but is still limited by the linear assumption and cannot remove strong reflections. The method proposed by Dong et al. shows inferior + +performance in both cases due to the difficulty of estimating the blending weight. Overall, our results are more visually favorable and contain fewer residual reflection components, which further reinforces our claims. + +We further deliver a comparison of reflection layers in Fig. 5 to illustrate that our method can better reproduce the reflection scenes. It can be seen that the methods (Zhang et al., RAGNet, and YTMT) based on the additive model provide weak reflection maps. In the meantime, the methods based on the linear model with scalars (BDN and IB-CLN) and the one containing an alpha blending map can hardly avoid mixing the transmission components in reflection estimations. This phenomenon reflects the weakness of a fixed simplified model, requiring the reconstruction of more than two components. As shown by our results, introducing a residue term can significantly enhance the restoration quality of the reflection scene, and separates out the components beyond the additive relationship. + +# 4.3. Ablation Study + +To better analyze our improvements in the physical model and network structure, we carried out a detailed ablation study from the perspective of the reconstruction losses, feature interaction mechanisms, and feature encoders. We gather the results (PSNR and SSIM averaged on Real20 and SIR $^2$ datasets) in Table 2 and deliver a visual comparison in Fig. 6. In what follows we detail every point of them. + +Reconstruction Loss. In view of the problems existing in the linear combination model, we use a learnable residue term to modify it, enhancing its flexibility so that it can cover more reflection scenes. In order to promote the learning of this residue term, we propose the reconstruction loss of residual correction. To illustrate the necessity of this setting, we perform an ablation study to compare the settings without reconstruction loss (w/o Recons. Loss) and using the linear reflection model (w/ Linear Recons.). It shows + +![](images/1333e793e07efcd6f663c828f2ef8ef784372781062c92fbf7fdbb884519706d.jpg) +w/o Recons. Loss + +![](images/0b62487eafd26a254564f06105fcc1a8e27b87ea1c2c1084c6c3d8f186261603.jpg) +w/ Linear Recons. + +![](images/99e2aa82a32ca4451814a27bb87f32eba06222be5f31d6aa2842c544155fa513.jpg) +w/o Feature Inter. + +![](images/98be64bff4db26a49534933b374d2b2e5aa7900de78abaa956e3c32638f3b29f.jpg) +w/ YTMT Inter. + +![](images/eefe51869b7612f2bd7240a4d2b8f9b4fa5a3e6d29e456e10fac4a966ff92d06.jpg) +w/o Feature Enc. + +![](images/465163df7e638b6f1e75711d10f4b5589651dcad877eda8e75e2d2fd2562da58.jpg) +w/ HyperColumn + +![](images/cd58a03ed6bfc8461cc78618a477a9a48f2bce6f3120555abd5b93b11b33a2a4.jpg) +Ours +Figure 6: The visual comparison of different settings involved in the ablation study. + +![](images/0ffb652bc7a2d5885074497b446b50cd811500ae3c8b6d0d050f18e756dc6247.jpg) +GT + +
ModelsPSNRSSIM
w/o Recons. Loss24.840.897
w/ Linear Recons.22.210.881
w/o Feature Inter.24.810.895
w/ YTMT Inter.24.890.901
w/o Feature Enc.24.750.903
w/ HyperColumn23.990.894
Ours25.400.905
+ +Table 2: Ablation study on different configurations. The initial two rows compare the configuration devoid of any reconstruction loss and that integrating a linear reconstruction loss. Rows 3-4 offer a comparison between the settings lacking feature interaction and incorporating the YTMT feature interaction mechanism. Rows 5-6 shed light on configurations that exclude the feature encoder and utilize HyperColumn as the feature encoder. The last row shows the results of our full version. + +that the linear model is inferior to the setting without any reconstruction criterion because of the existence of $\varepsilon$ . Meanwhile, without a reconstruction term, the model is likely to have trouble determining the portion of the components, excessively separating content to the reflection layer, resulting in textureless regions in the transmission layer as shown in Fig. 6. + +Feature Interaction Mechanism. The feature interaction mechanism connects the two information streams in our proposed DSRNet, both exchanging and conditioning the dual-stream information, which facilitates the layer reconstruction. As shown in Table 2, the model performance appears to degrade after removing the feature interaction (w/o Feature Inter.), which tells us that the network can hardly achieve state-of-the-art performance via a two-branch network without feature interaction. Further, we include the + +results of the YTMT feature interaction mechanism (w/ YTMT Inter.) by replacing MuGI with the YTMT mechanism. It can be seen that the model with the YTMT mechanism is only slightly better than the scheme without interaction, which demonstrates the superiority of the proposed MuGI mechanism. + +Feature Encoder. In comparison to the hypercolumn framework introduced by the previous method (referred to as "w/ HyperColumn") and the approach that without using any feature encoder ("w/o Feature Enc."), our method demonstrates noteworthy performance advantages. These advantages stem from two main factors. Firstly, our approach employs a hierarchical and gradual reduction of channel dimensions, followed by fusion operations. Secondly, DSFNet leverages a feature interaction mechanism, allowing high-level semantic features to interact before being fed into the subsequent stage. + +# 5. Conclusion + +In this paper, we modified the commonly-used linear model in the single image reflection separation task and proposed the reconstruction loss with residual correction. The proposed model is more flexible and effective compared with the previous methods. Meanwhile, we further improved the feature interaction mechanism in dual-stream networks and the usage of hierarchical semantic information, proposed the MuGI as a novel interaction paradigm, and a dual-stream semantic-aware network, namely DSR-Net. Extensive experiments revealed that our proposed method has achieved state-of-the-art performance on all real-world benchmark datasets and verified our contributions. In the future, further constraints are desired to be applied to the residue term to reduce its solution space, and more reflection models are hopefully to be covered by it to fit a wider range of reflection phenomena in the real world. + +# References + +[1] Amit K. Agrawal, Ramesh Raskar, Shree K. Nayar, and Yuanzhen Li. Removing photography artifacts using gradient projection and flash-exposure sampling. TOG, 24(3):828-835, 2005. 2 +[2] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In ECCV, pages 17-33, 2022. 2, 4 +[3] Zheng Dong, Ke Xu, Yin Yang, Hujun Bao, Weiwei Xu, and Rynson W. H. Lau. Location-aware single image reflection removal. In ICCV, pages 4997-5006, 2021. 3, 5, 6, 7 +[4] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes (VOC) challenge. *IJCV*, 88(2):303–338, 2010. 6 +[5] Qingnan Fan, Jiaolong Yang, Gang Hua, Baoquan Chen, and David P. Wipf. A generic deep architecture for single image reflection removal and image smoothing. In ICCV, pages 3258-3267, 2017. 2, 3, 6 +[6] Hany Farid and Edward H. Adelson. Separating reflections and lighting using independent components analysis. In CVPR, pages 1262-1267, 1999. 2 +[7] Xin Feng, Wenjie Pei, Zihui Jia, Fanglin Chen, David Zhang, and Guangming Lu. Deep-masking generative network: A unified framework for background restoration from superimposed images. TIP, 30:4867-4882, 2021. 5, 6 +[8] Kun Gai, Zhenwei Shi, and Changshui Zhang. Blind separation of superimposed moving images using image statistics. TPAMI, 34(1):19-32, 2012. 2 +[9] Xiaojie Guo, Xiaochun Cao, and Yi Ma. Robust separation of reflection from multiple images. In CVPR, pages 2195-2202, 2014. 2 +[10] Byeong-Ju Han and Jae-Young Sim. Reflection removal using low-rank matrix completion. In CVPR, pages 3872–3880, 2017. 2 +[11] Bharath Hariharan, Pablo Andres Arbeláez, Ross B. Girshick, and Jitendra Malik. Hypercolumns for object segmentation and fine-grained localization. In CVPR, pages 447-456, 2015. 3 +[12] Qiming Hu and Xiaojie Guo. *Trash or treasure? an interactive dual-stream strategy for single image reflection separation*. In *NeurIPS*, pages 24683–24694, 2021. 2, 3, 4, 5, 6, 7 +[13] Naejin Kong, Yu-Wing Tai, and Joseph S. Shin. A physically-based approach to reflection separation: From physical modeling to constrained optimization. TPAMI, 36(2):209-221, 2014. 2 +[14] Naejin Kong, Yu-Wing Tai, and Sung Yong Shin. High-quality reflection separation using polarized images. TIP, 20(12):3393-3405, 2011. 2 +[15] Chenyang Lei and Qifeng Chen. Robust reflection removal with reflection-free flash-only cues. In CVPR, pages 14811-14820, 2021. 2 +[16] Chenyang Lei, Xuhua Huang, Mengdi Zhang, Qiong Yan, Wenxiu Sun, and Qifeng Chen. Polarized reflection removal with perfect alignment in the wild. In CVPR, pages 1747-1755, 2020. 2 + +[17] Anat Levin and Yair Weiss. User assisted separation of reflections from a single image using a sparsity prior. TPAMI, 29(9):1647-1654, 2007. 1, 2 +[18] Anat Levin, Assaf Zomet, and Yair Weiss. Learning to perceive transparency from the statistics of natural scenes. In NeurIPS, pages 1247-1254, 2002. 2 +[19] Anat Levin, Assaf Zomet, and Yair Weiss. Separating reflections from a single image using local features. In CVPR, pages 306-313, 2004. 2 +[20] Chao Li, Yixiao Yang, Kun He, Stephen Lin, and John E. Hopcroft. Single image reflection removal through cascaded refinement. In CVPR, pages 3562-3571, 2020. 3, 5, 6, 7 +[21] Rui Li, Simeng Qiu, Guangming Zang, and Wolfgang Heidrich. Reflection separation via multi-bounce polarization state tracing. In ECCV, pages 781-796, 2020. 2 +[22] Yu Li and Michael S. Brown. Exploiting reflection change for automatic reflection removal. In ICCV, pages 2432-2439, 2013. 2 +[23] Yu Li and Michael S. Brown. Single image layer separation using relative smoothness. In CVPR, pages 2752-2759, 2014. 2 +[24] Yu Li, Ming Liu, Yaling Yi, Qince Li, Dongwei Ren, and Wangmeng Zuo. Two-stage single image reflection removal with reflection-aware guidance. CoRR, abs/2012.00945, 2020. 3, 5, 6, 7 +[25] Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Learning to see through obstructions. In CVPR, pages 14203-14212, 2020. 2 +[26] Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Learning to see through obstructions with layered decomposition. TPAMI, 44(11):8387-8402, 2022. 2 +[27] Youwei Lyu, Zhaopeng Cui, Si Li, Marc Pollefeys, and Boxin Shi. Reflection separation using a pair of unpolarized and polarized images. In NeurIPS, pages 14532-14542, 2019. 2 +[28] Shree K. Nayar, Xi-Sheng Fang, and Terrance E. Boult. Separation of reflection components using color and polarization. IJCV, 21(3):163-186, 1997. 2 +[29] Bernard Sarel and Michal Irani. Separating transparent layers of repetitive dynamic behaviors. In ICCV, pages 26-32, 2005. 2 +[30] Christian Simon and In Kyu Park. Reflection removal for in-vehicle black box videos. In CVPR, pages 4231–4239, 2015. 2 +[31] Sudipta N. Sinha, Johannes Kopf, Michael Goesele, Daniel Scharstein, and Richard Szeliski. Image-based rendering for scenes with reflections. TOG, 31(4):100:1-100:10, 2012. 1, 2 +[32] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In CVPR, pages 5769-5780, 2022. 2 +[33] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, and Alex C. Kot. Benchmarking single-image reflection removal algorithms. In ICCV, pages 3942-3950, 2017. 6 + +[34] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, and Alex C Kot. Crrn: Multi-scale guided concurrent reflection removal network. In CVPR, pages 4777-4785, 2018. 1, 2 +[35] Renjie Wan, Boxin Shi, Haoliang Li, Ling-Yu Duan, and Alex C. Kot. Reflection scene separation from a single image. In CVPR, pages 2395-2403, 2020. 1 +[36] Kaixuan Wei, Jiaolong Yang, Ying Fu, David P. Wipf, and Hua Huang. Single image reflection removal exploiting misaligned training data and network enhancements. In CVPR, pages 8178-8187, 2019. 2, 3, 4, 5, 6, 7 +[37] Qiang Wen, Yinjie Tan, Jing Qin, Wenxi Liu, Guoqiang Han, and Shengfeng He. Single image reflection removal beyond linearity. In CVPR, pages 3771-3779, 2019. 1, 2, 3 +[38] Tianfan Xue, Michael Rubinstein, Ce Liu, and William T. Freeman. A computational approach for obstruction-free photography. TOG, 34(4):79:1-79:11, 2015. 2 +[39] Jie Yang, Dong Gong, Lingqiao Liu, and Qinfeng Shi. Seeing deeply and bidirectionally: A deep learning approach for single image reflection removal. In ECCV, pages 675-691, 2018. 1, 2, 3, 5, 6, 7 +[40] Jiaolong Yang, Hongdong Li, Yuchao Dai, and Robby T. Tan. Robust optical flow estimation of double-layer images under transparency or reflection. In CVPR, pages 1410-1419, 2016. 2 +[41] Xuaner Cecilia Zhang, Ren Ng, and Qifeng Chen. Single image reflection separation with perceptual losses. In CVPR, pages 4786-4794, 2018. 2, 3, 4, 5, 6, 7 +[42] Qian Zheng, Jinnan Chen, Zhan Lu, Boxin Shi, Xudong Jiang, Kim-Hui Yap, Ling-Yu Duan, and Alex C Kot. What does plate glass reveal about camera calibration? In CVPR, pages 3022-3032, 2020. 2 +[43] Qian Zheng, Boxin Shi, Jinnan Chen, Xudong Jiang, LingYu Duan, and Alex C. Kot. Single image reflection removal with absorption effect. In CVPR, pages 13395-13404, 2021. 5, 6 \ No newline at end of file diff --git a/2023/Single Image Reflection Separation via Component Synergy/images.zip b/2023/Single Image Reflection Separation via Component Synergy/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..37286a946774381460e0264dccd85ba93485b472 --- /dev/null +++ b/2023/Single Image Reflection Separation via Component Synergy/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:188bdf685be900f8f7f795a028eca01b0c391708ef851e0bfc2ba80b146ddb0c +size 760895 diff --git a/2023/Single Image Reflection Separation via Component Synergy/layout.json b/2023/Single Image Reflection Separation via Component Synergy/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1640a55cd170e77cefa4331213b7baec873645ae --- /dev/null +++ b/2023/Single Image Reflection Separation via Component Synergy/layout.json @@ -0,0 +1,10100 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 481, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 481, + 121 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 481, + 121 + ], + "type": "text", + "content": "Single Image Reflection Separation via Component Synergy" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 233, + 143, + 363, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 143, + 363, + 157 + ], + "spans": [ + { + "bbox": [ + 233, + 143, + 363, + 157 + ], + "type": "text", + "content": "Qiming Hu Xiaojie Guo*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 158, + 475, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 158, + 475, + 172 + ], + "spans": [ + { + "bbox": [ + 118, + 158, + 475, + 172 + ], + "type": "text", + "content": "College of Intelligence and Computing, Tianjin University, Tianjin, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 184, + 174, + 408, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 174, + 408, + 185 + ], + "spans": [ + { + "bbox": [ + 184, + 174, + 408, + 185 + ], + "type": "text", + "content": "huqiming@tju.edu.cn, xj.max.quo@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 238, + 290, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 429 + ], + "type": "text", + "content": "The reflection superposition phenomenon is complex and widely distributed in the real world, which derives various simplified linear and nonlinear formulations of the problem. In this paper, based on the investigation of the weaknesses of existing models, we propose a more general form of the superposition model by introducing a learnable residue term, which can effectively capture residual information during decomposition, guiding the separated layers to be complete. In order to fully capitalize on its advantages, we further design the network structure elaborately, including a novel dual-stream interaction mechanism and a powerful decomposition network with a semantic pyramid encoder. Extensive experiments and ablation studies are conducted to verify our superiority over state-of-the-art approaches on multiple real-world benchmark datasets. Our code is publicly available at https://github.com/mingcv/DSRNet." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 451, + 128, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 451, + 128, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 451, + 128, + 464 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 472, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 578 + ], + "type": "text", + "content": "As a typical layer superimposition scenario, pictures captured through glass-like surfaces may be blended with undesired reflection, which not only impairs the aesthetic value but also hinders the downstream tasks [31]. In the meantime, scenes of interest are possibly concealed in reflection [35]. Therefore, both the parts transmitted through a surface (transmission layer, " + }, + { + "bbox": [ + 46, + 472, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 46, + 472, + 287, + 578 + ], + "type": "text", + "content": ") and those reflected (reflection layer, " + }, + { + "bbox": [ + 46, + 472, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 46, + 472, + 287, + 578 + ], + "type": "text", + "content": ") are desired to be reconstructed from a superimposed image " + }, + { + "bbox": [ + 46, + 472, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 472, + 287, + 578 + ], + "type": "text", + "content": " to fulfill the practical demands." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 579, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 687 + ], + "type": "text", + "content": "Following a common assumption [17], " + }, + { + "bbox": [ + 46, + 579, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 579, + 287, + 687 + ], + "type": "text", + "content": " is linearly composed by " + }, + { + "bbox": [ + 46, + 579, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 46, + 579, + 287, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 579, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 46, + 579, + 287, + 687 + ], + "type": "text", + "content": " expressed as " + }, + { + "bbox": [ + 46, + 579, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = \\mathbf{T} + \\mathbf{R}" + }, + { + "bbox": [ + 46, + 579, + 287, + 687 + ], + "type": "text", + "content": ", which has been popular due to its simplicity. However, in real-world scenarios, the reflection and transmission layers are likely to be weakened due to diffusion and other problems during the superimposition [35]. Therefore, several methods [34, 39] introduce scalars " + }, + { + "bbox": [ + 46, + 579, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 579, + 287, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 579, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 579, + 287, + 687 + ], + "type": "text", + "content": " to the two components, respectively, obtaining " + }, + { + "bbox": [ + 46, + 579, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = \\alpha \\mathbf{T} + \\beta \\mathbf{R}" + }, + { + "bbox": [ + 46, + 579, + 287, + 687 + ], + "type": "text", + "content": ". However, the linear superimposition model is often violated due to over-exposure" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 211, + 354, + 257 + ], + "blocks": [ + { + "bbox": [ + 309, + 211, + 354, + 257 + ], + "lines": [ + { + "bbox": [ + 309, + 211, + 354, + 257 + ], + "spans": [ + { + "bbox": [ + 309, + 211, + 354, + 257 + ], + "type": "image", + "image_path": "56033888929df1fdf01f854eea064f3699add69cf7861a5af3cfee24ae60856a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 259, + 334, + 269 + ], + "lines": [ + { + "bbox": [ + 329, + 259, + 334, + 269 + ], + "spans": [ + { + "bbox": [ + 329, + 259, + 334, + 269 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 356, + 213, + 401, + 258 + ], + "blocks": [ + { + "bbox": [ + 356, + 213, + 401, + 258 + ], + "lines": [ + { + "bbox": [ + 356, + 213, + 401, + 258 + ], + "spans": [ + { + "bbox": [ + 356, + 213, + 401, + 258 + ], + "type": "image", + "image_path": "5e54e280b6bb2ed473e36128b628896277f6b439bac2d5777531d79f7ea17d9b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 363, + 259, + 392, + 270 + ], + "lines": [ + { + "bbox": [ + 363, + 259, + 392, + 270 + ], + "spans": [ + { + "bbox": [ + 363, + 259, + 392, + 270 + ], + "type": "text", + "content": "T (GT)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 402, + 213, + 448, + 257 + ], + "blocks": [ + { + "bbox": [ + 402, + 213, + 448, + 257 + ], + "lines": [ + { + "bbox": [ + 402, + 213, + 448, + 257 + ], + "spans": [ + { + "bbox": [ + 402, + 213, + 448, + 257 + ], + "type": "image", + "image_path": "e195b4812d272af10aec6a4a3fc2d927c302ae4ab9f5f26ca0915392de0c8d3f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 259, + 439, + 270 + ], + "lines": [ + { + "bbox": [ + 410, + 259, + 439, + 270 + ], + "spans": [ + { + "bbox": [ + 410, + 259, + 439, + 270 + ], + "type": "text", + "content": "R (GT)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 449, + 213, + 496, + 257 + ], + "blocks": [ + { + "bbox": [ + 449, + 213, + 496, + 257 + ], + "lines": [ + { + "bbox": [ + 449, + 213, + 496, + 257 + ], + "spans": [ + { + "bbox": [ + 449, + 213, + 496, + 257 + ], + "type": "image", + "image_path": "86ee0166c729d4e5266ac42b2144ea8e06c051c58ff5bd1e030717eb281cce07.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 259, + 485, + 270 + ], + "lines": [ + { + "bbox": [ + 460, + 259, + 485, + 270 + ], + "spans": [ + { + "bbox": [ + 460, + 259, + 485, + 270 + ], + "type": "text", + "content": "Crops" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 498, + 213, + 544, + 257 + ], + "blocks": [ + { + "bbox": [ + 498, + 213, + 544, + 257 + ], + "lines": [ + { + "bbox": [ + 498, + 213, + 544, + 257 + ], + "spans": [ + { + "bbox": [ + 498, + 213, + 544, + 257 + ], + "type": "image", + "image_path": "802dece7a144651c19a192ab710a3618c747b3c26a3a06bc767c91f970f94b9d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 258, + 542, + 270 + ], + "lines": [ + { + "bbox": [ + 500, + 258, + 542, + 270 + ], + "spans": [ + { + "bbox": [ + 500, + 258, + 542, + 270 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{R}} = \\mathbf{I} - \\mathbf{T}" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 310, + 271, + 402, + 316 + ], + "blocks": [ + { + "bbox": [ + 310, + 271, + 402, + 316 + ], + "lines": [ + { + "bbox": [ + 310, + 271, + 402, + 316 + ], + "spans": [ + { + "bbox": [ + 310, + 271, + 402, + 316 + ], + "type": "image", + "image_path": "1d7d00c836444e1eed4e6f51d155a1ce6162c62fe63603c379b392c702dd0457.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 356, + 318, + 402, + 363 + ], + "blocks": [ + { + "bbox": [ + 313, + 328, + 348, + 338 + ], + "lines": [ + { + "bbox": [ + 313, + 328, + 348, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 348, + 338 + ], + "type": "inline_equation", + "content": "\\alpha \\mathbf{T} + \\mathbf{R}" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 356, + 318, + 402, + 363 + ], + "lines": [ + { + "bbox": [ + 356, + 318, + 402, + 363 + ], + "spans": [ + { + "bbox": [ + 356, + 318, + 402, + 363 + ], + "type": "image", + "image_path": "3940ba43f6c6ad3cb4fe7cd80e4a15d68daf11ef1c4aa786677c21703f8dff88.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 403, + 318, + 448, + 363 + ], + "blocks": [ + { + "bbox": [ + 403, + 318, + 448, + 363 + ], + "lines": [ + { + "bbox": [ + 403, + 318, + 448, + 363 + ], + "spans": [ + { + "bbox": [ + 403, + 318, + 448, + 363 + ], + "type": "image", + "image_path": "8be5773578d8e377a6b8052ab32d3c5a6aab1c6f2dbc64621e91d6b5821b2590.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 450, + 271, + 496, + 317 + ], + "blocks": [ + { + "bbox": [ + 450, + 271, + 496, + 317 + ], + "lines": [ + { + "bbox": [ + 450, + 271, + 496, + 317 + ], + "spans": [ + { + "bbox": [ + 450, + 271, + 496, + 317 + ], + "type": "image", + "image_path": "0597247a6e7e50f04d11b6a84fe15ee250e611dc156ff89cc2159bc147603e1c.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 449, + 318, + 496, + 363 + ], + "blocks": [ + { + "bbox": [ + 449, + 318, + 496, + 363 + ], + "lines": [ + { + "bbox": [ + 449, + 318, + 496, + 363 + ], + "spans": [ + { + "bbox": [ + 449, + 318, + 496, + 363 + ], + "type": "image", + "image_path": "e634e827a617c59fc25b85b37010c3f09ade9384bbfe0db24163efce5f7091e5.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 502, + 271, + 539, + 312 + ], + "blocks": [ + { + "bbox": [ + 502, + 271, + 539, + 312 + ], + "lines": [ + { + "bbox": [ + 502, + 271, + 539, + 312 + ], + "spans": [ + { + "bbox": [ + 502, + 271, + 539, + 312 + ], + "type": "image", + "image_path": "a521327a5a64758d209357b7498ff74004e3701e64339178c6c229c5161ac21f.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 503, + 323, + 538, + 359 + ], + "blocks": [ + { + "bbox": [ + 503, + 323, + 538, + 359 + ], + "lines": [ + { + "bbox": [ + 503, + 323, + 538, + 359 + ], + "spans": [ + { + "bbox": [ + 503, + 323, + 538, + 359 + ], + "type": "image", + "image_path": "cb6574b1db09669c5542194ffd894a15586491ffd4373b947c6073c1c2ba07a7.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 356, + 365, + 402, + 411 + ], + "blocks": [ + { + "bbox": [ + 315, + 342, + 349, + 357 + ], + "lines": [ + { + "bbox": [ + 315, + 342, + 349, + 357 + ], + "spans": [ + { + "bbox": [ + 315, + 342, + 349, + 357 + ], + "type": "text", + "content": "(IBCLN, CVPR,2020)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 374, + 351, + 385 + ], + "lines": [ + { + "bbox": [ + 313, + 374, + 351, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 351, + 385 + ], + "type": "text", + "content": "WoT+R" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 314, + 389, + 349, + 404 + ], + "lines": [ + { + "bbox": [ + 314, + 389, + 349, + 404 + ], + "spans": [ + { + "bbox": [ + 314, + 389, + 349, + 404 + ], + "type": "text", + "content": "(Dong et al., ICCV,2021)" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 356, + 365, + 402, + 411 + ], + "lines": [ + { + "bbox": [ + 356, + 365, + 402, + 411 + ], + "spans": [ + { + "bbox": [ + 356, + 365, + 402, + 411 + ], + "type": "image", + "image_path": "ea1e9750503db500715a77c1ae45a7b1f9f3115e6742e5706e62c39303c3e0d9.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 356, + 413, + 402, + 459 + ], + "blocks": [ + { + "bbox": [ + 356, + 413, + 402, + 459 + ], + "lines": [ + { + "bbox": [ + 356, + 413, + 402, + 459 + ], + "spans": [ + { + "bbox": [ + 356, + 413, + 402, + 459 + ], + "type": "image", + "image_path": "df5d9a3aae5e2511b328004d2cb70fcf424892a60610677108f3ab6f57f82c6a.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 486, + 547, + 559 + ], + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 559 + ], + "type": "text", + "content": "Figure 1: A visual example drawn from the " + }, + { + "bbox": [ + 304, + 486, + 547, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{SIR}^2" + }, + { + "bbox": [ + 304, + 486, + 547, + 559 + ], + "type": "text", + "content": " dataset. The first row displays a real input " + }, + { + "bbox": [ + 304, + 486, + 547, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 304, + 486, + 547, + 559 + ], + "type": "text", + "content": " with ground-truth " + }, + { + "bbox": [ + 304, + 486, + 547, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 304, + 486, + 547, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 486, + 547, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 304, + 486, + 547, + 559 + ], + "type": "text", + "content": ". In rows 2-5, the separations based on different physical models are compared, where " + }, + { + "bbox": [ + 304, + 486, + 547, + 559 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}" + }, + { + "bbox": [ + 304, + 486, + 547, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 486, + 547, + 559 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{R}}" + }, + { + "bbox": [ + 304, + 486, + 547, + 559 + ], + "type": "text", + "content": " represent the predictions. Our method is able to produce a \"Residue\" term, separating the nonlinearity from the layer reconstructions." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 403, + 365, + 448, + 411 + ], + "blocks": [ + { + "bbox": [ + 403, + 365, + 448, + 411 + ], + "lines": [ + { + "bbox": [ + 403, + 365, + 448, + 411 + ], + "spans": [ + { + "bbox": [ + 403, + 365, + 448, + 411 + ], + "type": "image", + "image_path": "89194c8b48b6bbc71358e9e44f7a2e79e31dece73c35a930078eaf8b6cdce826.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 403, + 414, + 448, + 459 + ], + "blocks": [ + { + "bbox": [ + 403, + 414, + 448, + 459 + ], + "lines": [ + { + "bbox": [ + 403, + 414, + 448, + 459 + ], + "spans": [ + { + "bbox": [ + 403, + 414, + 448, + 459 + ], + "type": "image", + "image_path": "54eaa68eff54bf6efc27ca67174227a550c862193e5d01ec4eb8c58bad0347a7.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 461, + 429, + 472 + ], + "lines": [ + { + "bbox": [ + 422, + 461, + 429, + 472 + ], + "spans": [ + { + "bbox": [ + 422, + 461, + 429, + 472 + ], + "type": "text", + "content": "R" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 449, + 365, + 496, + 411 + ], + "blocks": [ + { + "bbox": [ + 449, + 365, + 496, + 411 + ], + "lines": [ + { + "bbox": [ + 449, + 365, + 496, + 411 + ], + "spans": [ + { + "bbox": [ + 449, + 365, + 496, + 411 + ], + "type": "image", + "image_path": "2a797903b894441eb6d467b600a5b4484d47f917f52e82295d490826236ecba7.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 449, + 414, + 496, + 458 + ], + "blocks": [ + { + "bbox": [ + 449, + 414, + 496, + 458 + ], + "lines": [ + { + "bbox": [ + 449, + 414, + 496, + 458 + ], + "spans": [ + { + "bbox": [ + 449, + 414, + 496, + 458 + ], + "type": "image", + "image_path": "3702c3be7fe95cd56c097304b2349284bb7fca4fad51233e003cb9af86696456.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 461, + 460, + 485, + 472 + ], + "lines": [ + { + "bbox": [ + 461, + 460, + 485, + 472 + ], + "spans": [ + { + "bbox": [ + 461, + 460, + 485, + 472 + ], + "type": "text", + "content": "Crops" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 498, + 369, + 539, + 407 + ], + "blocks": [ + { + "bbox": [ + 498, + 369, + 539, + 407 + ], + "lines": [ + { + "bbox": [ + 498, + 369, + 539, + 407 + ], + "spans": [ + { + "bbox": [ + 498, + 369, + 539, + 407 + ], + "type": "image", + "image_path": "761dced01cc93bd9a6e25375956da6716dfefa58e796facb26a7726de2f98b17.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 498, + 415, + 544, + 457 + ], + "blocks": [ + { + "bbox": [ + 498, + 415, + 544, + 457 + ], + "lines": [ + { + "bbox": [ + 498, + 415, + 544, + 457 + ], + "spans": [ + { + "bbox": [ + 498, + 415, + 544, + 457 + ], + "type": "image", + "image_path": "1d44f4ed51cf31336afabf11ce3e43995dd5f9350a395bdfa62645f84f5ff835.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 504, + 461, + 537, + 471 + ], + "lines": [ + { + "bbox": [ + 504, + 461, + 537, + 471 + ], + "spans": [ + { + "bbox": [ + 504, + 461, + 537, + 471 + ], + "type": "text", + "content": "Residue" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "bbox": [ + 304, + 567, + 547, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 547, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 547, + 675 + ], + "type": "text", + "content": "and other phenomena [37]. Thus, an alpha-matting map " + }, + { + "bbox": [ + 304, + 567, + 547, + 675 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 304, + 567, + 547, + 675 + ], + "type": "text", + "content": " is introduced to the model as " + }, + { + "bbox": [ + 304, + 567, + 547, + 675 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = \\mathbf{W} \\circ \\mathbf{T} + \\bar{\\mathbf{W}} \\circ \\mathbf{R}" + }, + { + "bbox": [ + 304, + 567, + 547, + 675 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 567, + 547, + 675 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{W}} = \\mathbf{1} - \\mathbf{W}" + }, + { + "bbox": [ + 304, + 567, + 547, + 675 + ], + "type": "text", + "content": ", which, however, increases the degree of freedom and makes this ill-posed problem much harder. In view of all these limitations, it is not a trivial task to represent different kinds of reflection scenarios with a single model. In this work, we take a step forward in advancing the solution and deliver a more general form of the superimposition procedure by introducing a residue term as follows:" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 358, + 681, + 545, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 681, + 545, + 695 + ], + "spans": [ + { + "bbox": [ + 358, + 681, + 545, + 695 + ], + "type": "interline_equation", + "content": "\\mathbf {I} = \\tilde {\\mathbf {T}} + \\tilde {\\mathbf {R}} = \\mathbf {T} + \\mathbf {R} + \\Phi (\\mathbf {T}, \\mathbf {R}), \\tag {1}", + "image_path": "830545ed7e62e17a503b640688ac476f280a45f694a0c33e7140b71fc8739131.jpg" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": " denote the groundtruth scenes of the trans" + } + ] + } + ], + "index": 46 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "text", + "content": "*Corresponding author. This work was supported by National Natural Science Foundation of China under Grant no. 62072327." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13138" + } + ] + } + ], + "index": 48 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "content": "mission and reflection layers, while their respective information contained in " + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "content": " after superimposition and other degradations and finally reach the camera sensors is represented by " + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{T}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{R}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "inline_equation", + "content": "\\Phi(\\mathbf{T}, \\mathbf{R}) = \\mathbf{I} - \\mathbf{T} - \\mathbf{R}" + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "content": " is the residue of the reconstruction, which can be caused by attenuation, overexposure, etc. " + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "inline_equation", + "content": "\\Phi(\\cdot, \\cdot)" + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "content": " represents a group of functions that can be used to model the residue in specific situations. For example, " + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "inline_equation", + "content": "\\Phi(\\mathbf{T}, \\mathbf{R}) = 0" + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "content": " in [23, 5, 17, 41, 36, 12], " + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "inline_equation", + "content": "\\Phi(\\mathbf{T}, \\mathbf{R}) = (\\alpha - 1)\\mathbf{T} + (\\beta - 1)\\mathbf{R}" + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "content": " in [34, 39], " + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "inline_equation", + "content": "\\Phi(\\mathbf{T}, \\mathbf{R}) = -(\\bar{\\mathbf{W}} \\circ \\mathbf{T} + \\mathbf{W} \\circ \\mathbf{R})" + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "content": " in [37, 42]. In an effort to rule these different cases, a learnable module is desired to be introduced to model the residue " + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "inline_equation", + "content": "\\Phi(\\mathbf{T}, \\mathbf{R})" + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "content": ". To clarify it, we provide a visual example in Fig. 1, where the first row shows a superimposed image " + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "content": ", its groundtruth transmission " + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "content": " (GT) and reflection " + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "content": " (GT) layers collected by SIR" + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "content": " dataset. According to the linear model, the guidance for reflection restoration is " + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{R}} = \\mathbf{I} - \\mathbf{T}" + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "content": " with incomplete signals rather than " + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "content": " (GT) that contains full information, which not only impairs the training of networks but also hinders the downstream usages. Therefore, as we point out in Eq. (1), a promising way out of this dilemma is to utilize ground-truth layers " + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "content": " as guiding signals and employ an extra residual term to handle the non-linearity (\"Residue\" in Fig. 1). As shown, the components that violate the linear assumption are separated from the layer predictions " + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{R}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 393 + ], + "type": "text", + "content": ", and thus more complete and precious layer separations than previous methods are obtained." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 396, + 289, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 289, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 289, + 658 + ], + "type": "text", + "content": "To further exploit the synergy between the components " + }, + { + "bbox": [ + 47, + 396, + 289, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 47, + 396, + 289, + 658 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 396, + 289, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 47, + 396, + 289, + 658 + ], + "type": "text", + "content": ", we deliberate upon the facilitation of intercomponent feature interaction. The efficiency of a dual-stream interactive network has been verified in SIRS problem by [12]. Building upon their analysis, we subsequently advance a more effective scheme. Inspired by the gated mechanisms [2, 32], we present a simple yet effective mutually-gated interaction diagram (MuGI) serving as a better feature interaction candidate, by means of which we design the MuGI block to build our decomposition network. Moreover, the hypercolumn [41], as a commonly used semantic information encoder in the SIRS task, tends to aggregate features in a lossy way. Therefore, we propose a dual-stream pyramid fusion network (DSFNet) to replace it, which hierarchically decomposes and fuses the multiscale semantic information leveraging our proposed MuGI blocks and dual-stream fusion blocks (DSF Block). The roughly decomposed features of layers are further delivered into the dual-stream fine-grained decomposition phase (DSDNet). The two sub-networks constitute the main branch of our Dual-stream Semantic-aware network with Residual correction, namely DSRNet." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 658, + 266, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 658, + 266, + 670 + ], + "spans": [ + { + "bbox": [ + 59, + 658, + 266, + 670 + ], + "type": "text", + "content": "In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 58, + 677, + 287, + 715 + ], + "type": "text", + "content": "- We build a general form for SIRS by introducing a learnable residue term, which is more flexible to different scenarios and boosts the separation of layers;" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 72, + 547, + 162 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 317, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 545, + 109 + ], + "type": "text", + "content": "- We exploit the synergy of features via mutually-gated interaction block within a dual-stream semantic-aware network, which facilitates information usage;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 114, + 547, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 114, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 114, + 547, + 162 + ], + "type": "text", + "content": "- Extensive experiments are conducted to demonstrate the effectiveness of our design. Overall, it achieves state-of-the-art performance against the alternatives on multiple real-world benchmarks for SIRS." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 175, + 392, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 175, + 392, + 188 + ], + "spans": [ + { + "bbox": [ + 306, + 175, + 392, + 188 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 196, + 545, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 196, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 196, + 545, + 269 + ], + "type": "text", + "content": "Plenty of efforts have been devoted to handling the image reflection separation problem in the past decades. They either leverage multiple images to acquire more cues for the layer decomposition or exploit extra priors to manage the problem through a single image. In what follows, we organize the previous methods based on the images they require." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 269, + 546, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 546, + 449 + ], + "type": "text", + "content": "Multiple Image Reflection Separation. In general, the transmission component tends to be unpolarized, yet the reflection component varies when rotating the polarization filter mounted in front of a camera sensor. Inspired by this phenomenon, a variety of methods [28, 14, 13, 27, 21, 16] resort to the physical solution that separates different components through a sequence of images with varied polarization orientations. Besides the polarization cues, the layer separations can also be indicated by different focuses [6], stereo information [31], flash on/off [1, 15], relative motions [29, 8, 22, 38, 40, 25, 26] and scene consistency [9, 30, 10]. Although involving multiple images mitigates the ill-posedness of the separation problem, this pipeline requires professional devices (such as polarizers) and manual operations, which limits its application." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 449, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 665 + ], + "type": "text", + "content": "Single Image Reflection Separation. Compared with multiple-image solutions, single-image-based methods show more merits, being more flexible and requiring less manual operations. However, as there is no such thing as a free lunch, single-image schemes call for additional priors to cope with its ill-posedness nature, such as gradient-based constraints [19, 18, 17, 23] and manual annotations [17]. Specifically, Levin et al. [19, 18] impose sparse constraints on gradients to acquire the separations that have fewer edges. Levin and Weiss [17] demand several manual annotations of respective layers to obtain favorable decompositions, which involves additional human labor. Li and Brown [23] suppose that one layer is smoother than the other, thus applying unbalanced penalization on the gradients of the two layers to decompose them. Although these methods work fine in restricted scenarios, the complicated real-world conditions can considerably surpass their assumptions, leading to unsatisfactory results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "The advent of deep learning technology mitigates the incongruity between the volume of data and modeling capacity. Certain underlying assumptions can be subtly embedded in the process of data synthesis, and subsequently" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13139" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 547, + 247 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 547, + 247 + ], + "type": "image", + "image_path": "ec01cc870caafaa657d0cea161c2e87a94fff8633972795cd349c32bb1846eae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 257, + 547, + 307 + ], + "lines": [ + { + "bbox": [ + 46, + 257, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 547, + 307 + ], + "type": "text", + "content": "Figure 2: The architecture of our proposed DSRNet, which consists of two cascaded stages and a learnable residue module (LRM). In Stage 1, DSFNet aggregates and separates hierarchical semantic information with two interactive feature streams. The roughly separated features are further fed into the DSDNet for fine-grained decompositions in Stage 2, where the LRM takes decomposed features, separating out the components that violate the linear assumption." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 326, + 289, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 289, + 711 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 289, + 711 + ], + "type": "text", + "content": "learned by deep models. Concretely, CEILNet [5] imposes the relative smoothness prior to the synthesis of reflection layers and combines them with transmission layers by addition. An edge-aware network is developed to capture the transmission components, which yet ignores high-level semantics that are likely to aid the SIRS task. Zhang et al. [41] hence introduce HyperColumn features [11] using a pre-trained VGG-19 network to acquire semantic awareness, besides the perceptual and adversarial losses. Moreover, the exclusivity loss is developed to penalize the intersected gradients. ERRNet [36] goes a step further by leveraging an additional set of misaligned pairs. However, it appears to overlook the estimation of the reflection layer, which is essentially an image component and potentially important to distinguish the transmission parts. Li et al. [24] thereby presents a two-stage network (RAGNet) to first estimate reflection components and predict transmission ones guided by them, whereas the reflection estimation is isolated from the transmission in this way. Further, Hu and Guo [12] come up with the YTMT strategy, which pays equal attention to the two components, and a dual-stream interactive network is developed to restore the two layers simultaneously. However, their linear assumption makes the predicted reflection components tend to be weak in many cases. Other than them, BDN [39] and IBCLN [20] make use of reflection models weighted by scalars, and iteratively estimate both the components. This scheme prevents the reflection from being too weak, but its restorations often struggle to be free from the transmission parts. Furthermore, Wen et al. [37] simulate the nonlinear superimposition phenomenon by predicting a three-channel alpha blending weight map with the adversarial guidance of col" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 326, + 547, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 547, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 547, + 458 + ], + "type": "text", + "content": "lected unpaired images. Dong et al. [3] develops an iterative network and estimates a probabilistic reflection confidence map in each iteration. However, it is not an easier task to estimate an alpha blending map than to determine each of the blended components, while the residual estimation only needs to retain the redundant information during decomposition, which is more practical than the former. Considering the drawbacks of the previous reflection models, we propose a general form of them with a learnable residue term, which is shown to be more effective and flexible." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 471, + 389, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 471, + 389, + 485 + ], + "spans": [ + { + "bbox": [ + 306, + 471, + 389, + 485 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 492, + 547, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 547, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 547, + 554 + ], + "type": "text", + "content": "As shown in Fig. 2, our proposed DSRNet comprises two cascaded stages and a learnable residue module (LRM). In what follows, we first clarify the motivation of LRM and then introduce the MuGI block and the architecture of the DSRNet followed by its training loss functions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 562, + 494, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 562, + 494, + 575 + ], + "spans": [ + { + "bbox": [ + 305, + 562, + 494, + 575 + ], + "type": "text", + "content": "3.1. Residue of the Linear Combination" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "It often occurs that a superimposed image " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " cannot be perfectly represented by the linear combination of " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": ", leading to a residue term, as discussed in Section 1. Under its disturbance, the error of the additive reconstruction criterion " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\varepsilon = \\| \\mathbf{I} - (\\hat{\\mathbf{T}} +\\hat{\\mathbf{R}})\\|" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " remains large even though the predicted layers can well reconstruct the ground truths. Further minimizing " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " will put redundant information to either " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{R}}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": ", which induces the deviation of predictions from their ground truths instead. Moreover, there are many kinds of simplified physical models proposed by previous work with their own shortcomings. Naturally, a unified model is de" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13140" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "sired to be put forward. Therefore, we introduce an extra residue term to offset the error in additive reconstruction, which, at the same time, unifies the different physical models. As depicted in Fig. 2, we leverage a learnable residue module (LRM) after Stage 2 to estimate the residual information during the decomposition by analyzing the features of " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{R}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " before the final layer. LRM consists of an interactive and a fusion part to collect and merge residual information from two branches. The dual-stream signals are concatenated before entering the fusion part. To constrain the space of residue, the " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\tanh(\\cdot)" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " function is utilized as the final activation. With the participation of LRM, we define the following reconstruction loss with residual rectification " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "(\\mathrm{R}^3\\mathrm{Loss})" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 247, + 287, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 247, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 91, + 247, + 287, + 262 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} := \\left\\| \\mathbf {I} - \\left(\\hat {\\mathbf {T}} + \\hat {\\mathbf {R}}\\right) - \\Phi (\\hat {\\mathbf {T}}, \\hat {\\mathbf {R}}) \\right\\| _ {1}, \\tag {2}", + "image_path": "4c7dd9bea8b9c40f7ece089a19bafbcd90b918472375a178b725f0f2c4e5c1a5.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 269, + 287, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 287, + 329 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 269, + 287, + 329 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 269, + 287, + 329 + ], + "type": "text", + "content": " denotes the LRM and " + }, + { + "bbox": [ + 46, + 269, + 287, + 329 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _1" + }, + { + "bbox": [ + 46, + 269, + 287, + 329 + ], + "type": "text", + "content": " means the " + }, + { + "bbox": [ + 46, + 269, + 287, + 329 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 46, + 269, + 287, + 329 + ], + "type": "text", + "content": " norm. Guided by this objective, the information beyond the additive model will flow to the residue term. Notably, this term can be totally discarded during the testing phase, avoiding extra computational costs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 336, + 226, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 226, + 349 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 226, + 349 + ], + "type": "text", + "content": "3.2. Mutually-gated Interactive Block" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 354, + 288, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 288, + 426 + ], + "type": "text", + "content": "Now that we employ the LRM to model the residual components during the additive reconstruction, the rest part can be seen as a linear model again as " + }, + { + "bbox": [ + 46, + 354, + 288, + 426 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{I}} = \\mathbf{I} - \\Phi (\\hat{\\mathbf{T}},\\hat{\\mathbf{R}}) = \\hat{\\mathbf{T}} +\\hat{\\mathbf{R}}" + }, + { + "bbox": [ + 46, + 354, + 288, + 426 + ], + "type": "text", + "content": ". As displayed in Fig. 3 (a), we exploit the synergy between " + }, + { + "bbox": [ + 46, + 354, + 288, + 426 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}" + }, + { + "bbox": [ + 46, + 354, + 288, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 354, + 288, + 426 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{R}}" + }, + { + "bbox": [ + 46, + 354, + 288, + 426 + ], + "type": "text", + "content": " by developing the Mutually-Gated Interactive mechanism, namely MuGI:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 434, + 287, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 434, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 113, + 434, + 287, + 466 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\hat {\\mathbf {F}} _ {\\mathbf {T}} = \\mathcal {G} _ {1} \\left(\\mathbf {F} _ {\\mathbf {T}}\\right) \\circ \\mathcal {G} _ {2} \\left(\\mathbf {F} _ {\\mathbf {R}}\\right); \\\\ \\hat {\\mathbf {F}} _ {\\mathbf {R}} = \\mathcal {G} _ {1} \\left(\\mathbf {F} _ {\\mathbf {R}}\\right) \\circ \\mathcal {G} _ {2} \\left(\\mathbf {F} _ {\\mathbf {T}}\\right), \\end{array} \\right. \\tag {3}", + "image_path": "37470e67da0343b6a65d5e95fda4380d05af2f866b0da23f0cc2fd97f34ccf7c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{F_T},\\mathbf{F_R}\\in \\mathbb{R}^{H_1\\times W_1\\times C_1}" + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "text", + "content": " denote intermediate features of streams for predicting " + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}" + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{R}}" + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "text", + "content": ", respectively, and " + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{\\mathrm{T}},\\hat{\\mathbf{F}}_{\\mathrm{R}}\\in \\mathbb{R}^{H_2\\times W_2\\times C_2}" + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "text", + "content": " represent the outputs of the mutual gate. " + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_1" + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_2" + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "text", + "content": " are functions selecting which part of features to engage the interaction. " + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "text", + "content": " indicates the elementwise multiplication. Here we use a simple implementation of MuGI to evaluate its capability of solving SIRS, which divides a feature map in half at the channel dimension, and " + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_1(\\cdot)" + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "text", + "content": " always selects the former half, while " + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_2(\\cdot)" + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "text", + "content": " chooses the latter. We have " + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "inline_equation", + "content": "H_{1} = H_{2}" + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "inline_equation", + "content": "W_{1} = W_{2}" + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "inline_equation", + "content": "C_2 = C_1 / 2" + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "text", + "content": " in this way. Such a gated mechanism captures the mutual dependency of the two components in the feature space, say if either stream contains signals the other one desires, they tend to be split into the latter half (selected by " + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_2(\\cdot)" + }, + { + "bbox": [ + 46, + 473, + 287, + 653 + ], + "type": "text", + "content": ") and transferred into the sibling stream by the gate." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": "Our simple implementation of MuGI makes it plug-and-play. Therefore, in general, most of the single-path blocks can be converted into a dual-stream one by parallel setting two of them and relating them with MuGI. In this paper, we follow a state-of-the-art block design presented by [2]" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 85, + 539, + 137 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 391, + 84 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 391, + 84 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 391, + 84 + ], + "type": "text", + "content": "(a). MuGI Mechanism" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 85, + 539, + 137 + ], + "lines": [ + { + "bbox": [ + 310, + 85, + 539, + 137 + ], + "spans": [ + { + "bbox": [ + 310, + 85, + 539, + 137 + ], + "type": "image", + "image_path": "f543ebc8190b575d8149e22d166c0f824467140a4ab24ad36f478e445e0d2c28.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 156, + 542, + 279 + ], + "blocks": [ + { + "bbox": [ + 311, + 143, + 372, + 153 + ], + "lines": [ + { + "bbox": [ + 311, + 143, + 372, + 153 + ], + "spans": [ + { + "bbox": [ + 311, + 143, + 372, + 153 + ], + "type": "text", + "content": "(b). MuGI Block" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 156, + 542, + 279 + ], + "lines": [ + { + "bbox": [ + 309, + 156, + 542, + 279 + ], + "spans": [ + { + "bbox": [ + 309, + 156, + 542, + 279 + ], + "type": "image", + "image_path": "0d629e279f96afb3c67f50b88c529afdca7bd39dad85154d2ef385166dd5cf3e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 289, + 518, + 301 + ], + "lines": [ + { + "bbox": [ + 331, + 289, + 518, + 301 + ], + "spans": [ + { + "bbox": [ + 331, + 289, + 518, + 301 + ], + "type": "text", + "content": "Figure 3: MuGI mechanism and MuGI Block." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 324, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 545, + 433 + ], + "type": "text", + "content": "and convert its single-path NAF block into a dual-stream mutually-gated interactive block (MuGI block). Its diagram is depicted in Fig. 3 (b), in which the simple gates are replaced by MuGI gates. Specifically, the layer normalization makes dual-stream features comparable, " + }, + { + "bbox": [ + 305, + 324, + 545, + 433 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 305, + 324, + 545, + 433 + ], + "type": "text", + "content": " convolution doubles the channel dimensions in order to conduct MuGI without information loss which reduces dimensions by half during the interaction, channel attention and the following " + }, + { + "bbox": [ + 305, + 324, + 545, + 433 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 305, + 324, + 545, + 433 + ], + "type": "text", + "content": " convolution serve as fusion/reweighting roles." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 443, + 508, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 508, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 508, + 455 + ], + "type": "text", + "content": "3.3. Dual-stream Semantic-aware Network" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 462, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 715 + ], + "type": "text", + "content": "The semantic information provided by pre-trained models is usually introduced to alleviate the impact of ill-posedness in SIRS. The HyperColumn [41, 36, 12], as a prevalent multi-scale semantic feature extractor in SIRS, aggregates semantics through interpolating the features extracted by a pre-trained model into the same scale as the input images. It then concatenates them together and employs a " + }, + { + "bbox": [ + 304, + 462, + 545, + 715 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 462, + 545, + 715 + ], + "type": "text", + "content": " convolution to rapidly reduce the channel dimensions (typically from 1475 to 64 or 256), before feeding them into the decomposition networks. This strategy omits the inner relationship of multi-scale features and may discard informative signals. To overcome the drawbacks mentioned above, we present a dual-stream pyramid fusion network (DSFNet), which hierarchically aggregates the extracted features by jointly upsampling and interacting. As shown in Fig. 2, given the hierarchical features extracted by a pre-trained deep network (e.g., VGGNet-19), DSFNet gathers them in a bottom-up manner with MuGI blocks and dual-stream fusion blocks (DSF Block), the structure of which is shown in Fig. 2. \"Dual Conv\" means two parallel convolutional layers. The features extracted by the pre" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13141" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 519, + 237 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 519, + 237 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 519, + 237 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 519, + 237 + ], + "type": "table", + "html": "
MethodsReal20 (20)Objects (200)Postcard (199)Wild (55)Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Zhang et al. [41]22.550.78822.680.87916.810.79721.520.83220.080.835
BDN [39]18.410.72622.720.85620.710.85922.360.83021.650.849
ERRNet [36]22.890.80324.870.89622.040.87624.250.85323.530.879
IBCLN [20]21.860.76224.870.89323.390.87524.710.88624.100.879
RAGNet [24]22.950.79326.150.90323.670.87925.530.88024.900.886
DMGN [7]20.710.77024.980.89922.920.87723.810.83523.800.877
Zheng et al. [43]20.170.75525.200.88023.260.90525.390.87824.190.885
YTMT [12]23.260.80624.870.89622.910.88425.480.89024.050.886
Ours24.230.82026.280.91424.560.90825.680.89625.400.905
Dong et al.†[3]23.340.81224.360.89823.720.90325.730.90224.210.897
Ours†23.910.81826.740.92024.830.91126.110.90625.750.910
", + "image_path": "00d7982576588fb0d41182a4bcc3e114d5965d3339ead231c560e79c9e97ef95.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 244, + 547, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 244, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 547, + 270 + ], + "type": "text", + "content": "Table 1: Quantitative results on four real-world benchmark datasets of methods. The best results are indicated in **bold.†** indicates extra training data that are involved as in [3]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 290, + 289, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 289, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 289, + 494 + ], + "type": "text", + "content": "trained network at each level first interact through a MuGI block, thereby features in the same scale are related and transformed into dual-stream features " + }, + { + "bbox": [ + 46, + 290, + 289, + 494 + ], + "type": "inline_equation", + "content": "(\\mathbf{F_T} = \\mathbf{F_R}" + }, + { + "bbox": [ + 46, + 290, + 289, + 494 + ], + "type": "text", + "content": " for these blocks). After the interaction, the deepest features extracted by VGGNet are upscaled and then fused with shallower features through the DSF block, in which the dual-stream features come from the two adjacent scales are concatenated at channel dimension, fused by a 1x1 convolutional layer and then upscaled at each stream. The MuGI blocks are further employed to promote cross-scale interactions. Note that, shallow features extracted from RGB inputs are fused at the top of DSFNet, which preserves fine-grained details to restore both layers. The aggregated features provide rough separations of layers and are further refined by the second stage. The following stage (DSDNet) is constructed by the proposed MuGI blocks in a U-shaped manner. More details can be found in the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 499, + 136, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 136, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 136, + 511 + ], + "type": "text", + "content": "3.4. Loss Function" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 517, + 287, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 566 + ], + "type": "text", + "content": "Besides the " + }, + { + "bbox": [ + 46, + 517, + 287, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^3" + }, + { + "bbox": [ + 46, + 517, + 287, + 566 + ], + "type": "text", + "content": " loss demonstrated in Section 3.1, we further introduce pixel, perceptual, and exclusion loss for pixel-wise and semantic fidelity as well as the gradient independence of layers, which are described as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 566, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 566, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 287, + 613 + ], + "type": "text", + "content": "Pixel Loss. The pixel loss is used to constrain the consistency of the reconstructed layer and groundtruth in both the natural image domain and gradient domain, minimizing their errors as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 616, + 287, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 616, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 77, + 616, + 287, + 649 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {p i x} := \\| \\hat {\\mathbf {T}} - \\mathbf {T} \\| _ {2} ^ {2} + \\| \\hat {\\mathbf {R}} - \\mathbf {R} \\| _ {2} ^ {2} \\tag {4} \\\\ + \\alpha (\\| \\nabla \\hat {\\mathbf {T}} - \\nabla \\mathbf {T} \\| _ {1} + \\| \\nabla \\hat {\\mathbf {R}} - \\nabla \\mathbf {R} \\| _ {1}), \\\\ \\end{array}", + "image_path": "3b501debff18bf55bd441823622ebd6212b1c3d26686c7c3ea65bca3afe1c613.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_2" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " norm. " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " stands for the gradient operator, which extracts gradients in both vertical and horizontal directions. We express one of them in the formulation for clarity and average them in practice. " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " is set as 2 in our experiments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 361 + ], + "type": "text", + "content": "Perceptual Loss. The pixel-wise losses cannot assure the multi-scale consistency between predictions and their groundtruths, which may result in overwhelming punishment for the whole image due to small differences in local brightness. Therefore, the perceptual loss is further introduced as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 355, + 367, + 545, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 367, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 355, + 367, + 545, + 392 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p e r} := \\sum_ {i} \\omega_ {i} \\| \\phi_ {i} (\\hat {\\mathbf {T}}) - \\phi_ {i} (\\mathbf {T}) \\| _ {1}, \\tag {5}", + "image_path": "19675301af2c5429917fb80a2719c2610bdc86af01ed407d12d7afb35ef5bf39.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 399, + 545, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 399, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 545, + 447 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 399, + 545, + 447 + ], + "type": "inline_equation", + "content": "\\phi_i(\\cdot)" + }, + { + "bbox": [ + 305, + 399, + 545, + 447 + ], + "type": "text", + "content": " designates the features drawn by layer " + }, + { + "bbox": [ + 305, + 399, + 545, + 447 + ], + "type": "inline_equation", + "content": "i\\in" + }, + { + "bbox": [ + 305, + 399, + 545, + 447 + ], + "type": "inline_equation", + "content": "\\{2,7,12,21,30\\}" + }, + { + "bbox": [ + 305, + 399, + 545, + 447 + ], + "type": "text", + "content": " of a VGG-19 model. " + }, + { + "bbox": [ + 305, + 399, + 545, + 447 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 305, + 399, + 545, + 447 + ], + "type": "text", + "content": " s are combining weights of terms at different layers. We follow the setting of hyper-parameters in [36]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 447, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 447, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 305, + 447, + 545, + 483 + ], + "type": "text", + "content": "Exclusion Loss. We introduce the exclusion loss to strengthen the gradient independence prior and reduce the structural coupling in estimated separations as below:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 323, + 490, + 545, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 490, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 323, + 490, + 545, + 546 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {e x c} := \\frac {1}{N} \\sum_ {n = 0} ^ {N - 1} \\| \\Psi \\left(\\hat {\\mathbf {T}} ^ {\\downarrow n}, \\hat {\\mathbf {R}} ^ {\\downarrow n}\\right) \\| _ {2} ^ {2}, \\tag {6} \\\\ \\Psi (\\hat {\\mathbf {T}}, \\hat {\\mathbf {R}}) := \\tanh \\left(\\eta_ {1} | \\nabla \\hat {\\mathbf {T}} |\\right) \\circ \\tanh \\left(\\eta_ {2} | \\nabla \\hat {\\mathbf {R}} |\\right), \\\\ \\end{array}", + "image_path": "022afbc8873e3de40a2543ad7471038c81733ee35084b6508a8ef30e08683c9b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 552, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 552, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 305, + 552, + 545, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 552, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}^{\\downarrow n}" + }, + { + "bbox": [ + 305, + 552, + 545, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 552, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{R}}^{\\downarrow n}" + }, + { + "bbox": [ + 305, + 552, + 545, + 590 + ], + "type": "text", + "content": " represent taking down-sampling by " + }, + { + "bbox": [ + 305, + 552, + 545, + 590 + ], + "type": "inline_equation", + "content": "2^{n}" + }, + { + "bbox": [ + 305, + 552, + 545, + 590 + ], + "type": "text", + "content": " times (" + }, + { + "bbox": [ + 305, + 552, + 545, + 590 + ], + "type": "inline_equation", + "content": "2^{N}" + }, + { + "bbox": [ + 305, + 552, + 545, + 590 + ], + "type": "text", + "content": " at most) of " + }, + { + "bbox": [ + 305, + 552, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}" + }, + { + "bbox": [ + 305, + 552, + 545, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 552, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{R}}" + }, + { + "bbox": [ + 305, + 552, + 545, + 590 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 552, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\eta_{1}" + }, + { + "bbox": [ + 305, + 552, + 545, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 552, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\eta_{2}" + }, + { + "bbox": [ + 305, + 552, + 545, + 590 + ], + "type": "text", + "content": " are normalization factors, which are identical to [41]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 590, + 542, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 542, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 542, + 601 + ], + "type": "text", + "content": "Gathering all the loss terms yields the final objective as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 335, + 609, + 545, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 609, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 335, + 609, + 545, + 623 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l l}} := \\mathcal {L} _ {\\text {p i x}} + \\beta_ {1} \\mathcal {L} _ {\\text {p e r}} + \\beta_ {2} \\mathcal {L} _ {\\text {e x c}} + \\beta_ {3} \\mathcal {L} _ {\\text {r e c}}, \\tag {7}", + "image_path": "52ad74781ddb725eaf2620f08e6dd90248d4329fcf698ec1a6e1bd8d22773eaf.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 629, + 545, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 629, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 305, + 629, + 545, + 642 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 629, + 545, + 642 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.01" + }, + { + "bbox": [ + 305, + 629, + 545, + 642 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 629, + 545, + 642 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 1" + }, + { + "bbox": [ + 305, + 629, + 545, + 642 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 629, + 545, + 642 + ], + "type": "inline_equation", + "content": "\\beta_{3} = 0.2" + }, + { + "bbox": [ + 305, + 629, + 545, + 642 + ], + "type": "text", + "content": " are set empirically." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 651, + 447, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 651, + 447, + 664 + ], + "spans": [ + { + "bbox": [ + 305, + 651, + 447, + 664 + ], + "type": "text", + "content": "4. Experimental Validation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 671, + 439, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 439, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 439, + 684 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Implemented in PyTorch, our models are optimized with the Adam optimizer on a single RTX 3090 GPU for 20" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "13142" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 526, + 507 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 526, + 507 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 526, + 507 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 526, + 507 + ], + "type": "image", + "image_path": "b36c30d0e3a868cbcd511404139932fa80e9d12352846d23f6f5fef772b9d1fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 516, + 536, + 529 + ], + "lines": [ + { + "bbox": [ + 56, + 516, + 536, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 516, + 536, + 529 + ], + "type": "text", + "content": "Figure 4: Visual comparison of estimated transmission layers between state-of-the-arts and ours on real-world samples." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 549, + 288, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 549, + 288, + 585 + ], + "spans": [ + { + "bbox": [ + 46, + 549, + 288, + 585 + ], + "type": "text", + "content": "epochs at most to reach favorable overall performance. The learning rate is initialized as " + }, + { + "bbox": [ + 46, + 549, + 288, + 585 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 549, + 288, + 585 + ], + "type": "text", + "content": " and fixed during the training phase with a batch size of 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 586, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 287, + 670 + ], + "type": "text", + "content": "Dataset. Our training dataset embraces both real and synthesized images. Following [5, 12], the training dataset are composed of 90 real pairs from [41] and 7,643 synthesized pairs from the PASCAL VOC dataset [4]. For synthesized data, transmission and reflection layers are weakened by coefficients " + }, + { + "bbox": [ + 46, + 586, + 287, + 670 + ], + "type": "inline_equation", + "content": "\\gamma_{1} \\in [0.8, 1.0]" + }, + { + "bbox": [ + 46, + 586, + 287, + 670 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 586, + 287, + 670 + ], + "type": "inline_equation", + "content": "\\gamma_{2} \\in [0.4, 1.0]" + }, + { + "bbox": [ + 46, + 586, + 287, + 670 + ], + "type": "text", + "content": " during blending them with the following model:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 679, + 287, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 679, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 70, + 679, + 287, + 693 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {\\text {s y n}} = \\gamma_ {1} \\mathbf {T} _ {\\text {s y n}} + \\gamma_ {2} \\mathbf {R} _ {\\text {s y n}} - \\gamma_ {1} \\gamma_ {2} \\mathbf {T} _ {\\text {s y n}} \\circ \\mathbf {R} _ {\\text {s y n}}, \\tag {8}", + "image_path": "dd0e7c4864f9d9a6ff9b9f9ff70cdc9090cb86798ff24a63593804045b628953.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 701, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 701, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{syn},\\mathbf{R}_{syn}" + }, + { + "bbox": [ + 47, + 701, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 701, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{syn}" + }, + { + "bbox": [ + 47, + 701, + 288, + 715 + ], + "type": "text", + "content": " represent the transmission, re" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 549, + 545, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 549, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 305, + 549, + 545, + 597 + ], + "type": "text", + "content": "flection, and superimposed layers during synthesis, respectively. This formulation is inspired by the \"screen\" blending mode in digital image processing, which always reserves lighter colors for the blending layers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 609, + 443, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 609, + 443, + 621 + ], + "spans": [ + { + "bbox": [ + 305, + 609, + 443, + 621 + ], + "type": "text", + "content": "4.2. Performance Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Quantitative comparison. A comparison is made between state-of-the-art methods as shown in Table 1, involving Zhang et al. [41], BDN [39], ERRNet [36], IBCLN [20], RAGNet [24], DMGN [7], Zheng et al. [43], YTMT [12] and ours on four real-world dataset, including Real20 [41] and three subsets of the " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SIR}^2" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " Dataset [33]. Besides, to meet the configuration of [3], we train our DSRNet ad" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "13143" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 154, + 133 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 154, + 133 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 154, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 154, + 133 + ], + "type": "image", + "image_path": "f931fc59e8419156fdeed9c2cc088dc884a3b229dba2bd65a9d19f630bbf4662.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 137, + 118, + 146 + ], + "lines": [ + { + "bbox": [ + 97, + 137, + 118, + 146 + ], + "spans": [ + { + "bbox": [ + 97, + 137, + 118, + 146 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 155, + 70, + 249, + 133 + ], + "blocks": [ + { + "bbox": [ + 155, + 70, + 249, + 133 + ], + "lines": [ + { + "bbox": [ + 155, + 70, + 249, + 133 + ], + "spans": [ + { + "bbox": [ + 155, + 70, + 249, + 133 + ], + "type": "image", + "image_path": "bb297128c2955cc666671a00763ccd0c157f4150567ddc55778a802d132a1ed1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 249, + 70, + 343, + 133 + ], + "blocks": [ + { + "bbox": [ + 249, + 70, + 343, + 133 + ], + "lines": [ + { + "bbox": [ + 249, + 70, + 343, + 133 + ], + "spans": [ + { + "bbox": [ + 249, + 70, + 343, + 133 + ], + "type": "image", + "image_path": "f51a50c0aabb19973dde4259174388174a86125dff3749c2d4bf7094f0a8258c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 343, + 70, + 436, + 133 + ], + "blocks": [ + { + "bbox": [ + 343, + 70, + 436, + 133 + ], + "lines": [ + { + "bbox": [ + 343, + 70, + 436, + 133 + ], + "spans": [ + { + "bbox": [ + 343, + 70, + 436, + 133 + ], + "type": "image", + "image_path": "8958afe7f5e3ddee1b14c3bd566f43b163bee46a37f8f446ad7d85dcc12091d2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 437, + 70, + 531, + 133 + ], + "blocks": [ + { + "bbox": [ + 437, + 70, + 531, + 133 + ], + "lines": [ + { + "bbox": [ + 437, + 70, + 531, + 133 + ], + "spans": [ + { + "bbox": [ + 437, + 70, + 531, + 133 + ], + "type": "image", + "image_path": "7013d71ceeadfb10017ac1f6dc69ac322446ed27aef54f63b9db6d5396fddbcc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 60, + 148, + 154, + 210 + ], + "blocks": [ + { + "bbox": [ + 60, + 148, + 154, + 210 + ], + "lines": [ + { + "bbox": [ + 60, + 148, + 154, + 210 + ], + "spans": [ + { + "bbox": [ + 60, + 148, + 154, + 210 + ], + "type": "image", + "image_path": "f5d2b60917c97df8636b27eeae956475412041c5e9060a62d158b6bac4077b57.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 214, + 129, + 225 + ], + "lines": [ + { + "bbox": [ + 85, + 214, + 129, + 225 + ], + "spans": [ + { + "bbox": [ + 85, + 214, + 129, + 225 + ], + "type": "text", + "content": "YTMT [12]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 154, + 148, + 249, + 209 + ], + "blocks": [ + { + "bbox": [ + 170, + 137, + 233, + 147 + ], + "lines": [ + { + "bbox": [ + 170, + 137, + 233, + 147 + ], + "spans": [ + { + "bbox": [ + 170, + 137, + 233, + 147 + ], + "type": "text", + "content": "Zhang et al. [41]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 148, + 249, + 209 + ], + "lines": [ + { + "bbox": [ + 154, + 148, + 249, + 209 + ], + "spans": [ + { + "bbox": [ + 154, + 148, + 249, + 209 + ], + "type": "image", + "image_path": "1f5ef5a078f8e9aec5f972bdd91a453f99b1a7d68d49747621c5797ffa3af1e5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 214, + 228, + 225 + ], + "lines": [ + { + "bbox": [ + 173, + 214, + 228, + 225 + ], + "spans": [ + { + "bbox": [ + 173, + 214, + 228, + 225 + ], + "type": "text", + "content": "Dong et al. [3]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 249, + 148, + 343, + 210 + ], + "blocks": [ + { + "bbox": [ + 277, + 137, + 315, + 147 + ], + "lines": [ + { + "bbox": [ + 277, + 137, + 315, + 147 + ], + "spans": [ + { + "bbox": [ + 277, + 137, + 315, + 147 + ], + "type": "text", + "content": "BDN [39]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 249, + 148, + 343, + 210 + ], + "lines": [ + { + "bbox": [ + 249, + 148, + 343, + 210 + ], + "spans": [ + { + "bbox": [ + 249, + 148, + 343, + 210 + ], + "type": "image", + "image_path": "2606e1aad110ba01058e13f39177897efd92ed786618c8deb42074dfe0029ff0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 215, + 305, + 224 + ], + "lines": [ + { + "bbox": [ + 286, + 215, + 305, + 224 + ], + "spans": [ + { + "bbox": [ + 286, + 215, + 305, + 224 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 235, + 545, + 259 + ], + "lines": [ + { + "bbox": [ + 46, + 235, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 545, + 259 + ], + "type": "text", + "content": "Figure 5: Visual comparison of estimated reflections between state-of-the-arts and ours on a sample drawn from the " + }, + { + "bbox": [ + 46, + 235, + 545, + 259 + ], + "type": "inline_equation", + "content": "\\mathrm{SIR}^2" + }, + { + "bbox": [ + 46, + 235, + 545, + 259 + ], + "type": "text", + "content": " dataset. Ours has a significant advantage over other alternatives by separating the residue component from the estimations." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 343, + 148, + 436, + 210 + ], + "blocks": [ + { + "bbox": [ + 367, + 137, + 412, + 147 + ], + "lines": [ + { + "bbox": [ + 367, + 137, + 412, + 147 + ], + "spans": [ + { + "bbox": [ + 367, + 137, + 412, + 147 + ], + "type": "text", + "content": "IBCLN [20]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 343, + 148, + 436, + 210 + ], + "lines": [ + { + "bbox": [ + 343, + 148, + 436, + 210 + ], + "spans": [ + { + "bbox": [ + 343, + 148, + 436, + 210 + ], + "type": "image", + "image_path": "7a16d8ec40f48c74b778c09db3304f953905e0e5f4858df6baa85dcef82b779d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 214, + 417, + 224 + ], + "lines": [ + { + "bbox": [ + 362, + 214, + 417, + 224 + ], + "spans": [ + { + "bbox": [ + 362, + 214, + 417, + 224 + ], + "type": "text", + "content": "Residue (ours)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 437, + 148, + 531, + 210 + ], + "blocks": [ + { + "bbox": [ + 459, + 137, + 508, + 147 + ], + "lines": [ + { + "bbox": [ + 459, + 137, + 508, + 147 + ], + "spans": [ + { + "bbox": [ + 459, + 137, + 508, + 147 + ], + "type": "text", + "content": "RAGNet [24]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 437, + 148, + 531, + 210 + ], + "lines": [ + { + "bbox": [ + 437, + 148, + 531, + 210 + ], + "spans": [ + { + "bbox": [ + 437, + 148, + 531, + 210 + ], + "type": "image", + "image_path": "14e79a74f5fe1a00dde28020204561efbc8c1e17ba44715add88a06d657abc98.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 215, + 490, + 224 + ], + "lines": [ + { + "bbox": [ + 477, + 215, + 490, + 224 + ], + "spans": [ + { + "bbox": [ + 477, + 215, + 490, + 224 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 281, + 287, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 317 + ], + "type": "text", + "content": "ditionally under their training settings, which includes 200 extra real pairs provided by the \"Nature\" dataset [20], and 13,700 synthesized image pairs provided by [41] instead." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 318, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 449 + ], + "type": "text", + "content": "It turns out that our method shows its superiority over other competitors on all the testing datasets in both settings, gaining " + }, + { + "bbox": [ + 46, + 318, + 287, + 449 + ], + "type": "inline_equation", + "content": "0.50dB" + }, + { + "bbox": [ + 46, + 318, + 287, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 318, + 287, + 449 + ], + "type": "inline_equation", + "content": "1.54dB" + }, + { + "bbox": [ + 46, + 318, + 287, + 449 + ], + "type": "text", + "content": " in terms of average PSNR in the two settings, respectively. Given the four real-world datasets contains a variety of scenes, illumination conditions, and glass thickness, it is not a trivial task to achieve the best performance of all metrics on these datasets simultaneously. The experimental results demonstrate that our proposed SIRS scheme has significantly higher performance and stronger generalization ability, which explains our main contributions." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 451, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 713 + ], + "type": "text", + "content": "Qualitative comparisons. To further explain our performance advantages, we provide visual comparisons of transmission layers in Fig. 4 against state-of-the-art methods, including Zhang et al. [41], BDN [39], ERRNet [36], IBCLN [20], RAGNet [24], YTMT [12], and Dong et al. [3]. As can be observed in Fig. 4, the method proposed by Zhang et al. fails to handle the case containing scattered reflection components in the second row. BDN has trouble dealing with the images with highlights like in the first row and even aggravates the reflections. The problem is likely to be caused by the linear reflection model, which lacks the ability to model specular highlights. For the results of ERRNet, which contains only a single branch to estimate transmission layers, it lacks the direct modeling of the reflections, therefore only removing parts of them. IBCLN also has trouble removing highlights and scattered reflections. The reflection layers in RAGNet are estimated without the participation of the transmission, hence showing inferior performance. YTMT can better cope with the highlights in the last row owing to its dual-stream design but is still limited by the linear assumption and cannot remove strong reflections. The method proposed by Dong et al. shows inferior" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 281, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 327 + ], + "type": "text", + "content": "performance in both cases due to the difficulty of estimating the blending weight. Overall, our results are more visually favorable and contain fewer residual reflection components, which further reinforces our claims." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 329, + 546, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 546, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 546, + 496 + ], + "type": "text", + "content": "We further deliver a comparison of reflection layers in Fig. 5 to illustrate that our method can better reproduce the reflection scenes. It can be seen that the methods (Zhang et al., RAGNet, and YTMT) based on the additive model provide weak reflection maps. In the meantime, the methods based on the linear model with scalars (BDN and IB-CLN) and the one containing an alpha blending map can hardly avoid mixing the transmission components in reflection estimations. This phenomenon reflects the weakness of a fixed simplified model, requiring the reconstruction of more than two components. As shown by our results, introducing a residue term can significantly enhance the restoration quality of the reflection scene, and separates out the components beyond the additive relationship." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 503, + 400, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 400, + 516 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 400, + 516 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 522, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 605 + ], + "type": "text", + "content": "To better analyze our improvements in the physical model and network structure, we carried out a detailed ablation study from the perspective of the reconstruction losses, feature interaction mechanisms, and feature encoders. We gather the results (PSNR and SSIM averaged on Real20 and SIR" + }, + { + "bbox": [ + 304, + 522, + 545, + 605 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 304, + 522, + 545, + 605 + ], + "type": "text", + "content": " datasets) in Table 2 and deliver a visual comparison in Fig. 6. In what follows we detail every point of them." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": "Reconstruction Loss. In view of the problems existing in the linear combination model, we use a learnable residue term to modify it, enhancing its flexibility so that it can cover more reflection scenes. In order to promote the learning of this residue term, we propose the reconstruction loss of residual correction. To illustrate the necessity of this setting, we perform an ablation study to compare the settings without reconstruction loss (w/o Recons. Loss) and using the linear reflection model (w/ Linear Recons.). It shows" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13144" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 184, + 140 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 184, + 140 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 184, + 140 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 184, + 140 + ], + "type": "image", + "image_path": "1333e793e07efcd6f663c828f2ef8ef784372781062c92fbf7fdbb884519706d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 144, + 161, + 154 + ], + "lines": [ + { + "bbox": [ + 96, + 144, + 161, + 154 + ], + "spans": [ + { + "bbox": [ + 96, + 144, + 161, + 154 + ], + "type": "text", + "content": "w/o Recons. Loss" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 70, + 294, + 140 + ], + "blocks": [ + { + "bbox": [ + 184, + 70, + 294, + 140 + ], + "lines": [ + { + "bbox": [ + 184, + 70, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 184, + 70, + 294, + 140 + ], + "type": "image", + "image_path": "0b62487eafd26a254564f06105fcc1a8e27b87ea1c2c1084c6c3d8f186261603.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 144, + 272, + 154 + ], + "lines": [ + { + "bbox": [ + 206, + 144, + 272, + 154 + ], + "spans": [ + { + "bbox": [ + 206, + 144, + 272, + 154 + ], + "type": "text", + "content": "w/ Linear Recons." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 296, + 70, + 406, + 140 + ], + "blocks": [ + { + "bbox": [ + 296, + 70, + 406, + 140 + ], + "lines": [ + { + "bbox": [ + 296, + 70, + 406, + 140 + ], + "spans": [ + { + "bbox": [ + 296, + 70, + 406, + 140 + ], + "type": "image", + "image_path": "99e2aa82a32ca4451814a27bb87f32eba06222be5f31d6aa2842c544155fa513.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 144, + 383, + 154 + ], + "lines": [ + { + "bbox": [ + 318, + 144, + 383, + 154 + ], + "spans": [ + { + "bbox": [ + 318, + 144, + 383, + 154 + ], + "type": "text", + "content": "w/o Feature Inter." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 408, + 70, + 518, + 140 + ], + "blocks": [ + { + "bbox": [ + 408, + 70, + 518, + 140 + ], + "lines": [ + { + "bbox": [ + 408, + 70, + 518, + 140 + ], + "spans": [ + { + "bbox": [ + 408, + 70, + 518, + 140 + ], + "type": "image", + "image_path": "98be64bff4db26a49534933b374d2b2e5aa7900de78abaa956e3c32638f3b29f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 144, + 492, + 154 + ], + "lines": [ + { + "bbox": [ + 433, + 144, + 492, + 154 + ], + "spans": [ + { + "bbox": [ + 433, + 144, + 492, + 154 + ], + "type": "text", + "content": "w/ YTMT Inter." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 73, + 156, + 184, + 226 + ], + "blocks": [ + { + "bbox": [ + 73, + 156, + 184, + 226 + ], + "lines": [ + { + "bbox": [ + 73, + 156, + 184, + 226 + ], + "spans": [ + { + "bbox": [ + 73, + 156, + 184, + 226 + ], + "type": "image", + "image_path": "eefe51869b7612f2bd7240a4d2b8f9b4fa5a3e6d29e456e10fac4a966ff92d06.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 230, + 160, + 240 + ], + "lines": [ + { + "bbox": [ + 97, + 230, + 160, + 240 + ], + "spans": [ + { + "bbox": [ + 97, + 230, + 160, + 240 + ], + "type": "text", + "content": "w/o Feature Enc." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 184, + 156, + 294, + 226 + ], + "blocks": [ + { + "bbox": [ + 184, + 156, + 294, + 226 + ], + "lines": [ + { + "bbox": [ + 184, + 156, + 294, + 226 + ], + "spans": [ + { + "bbox": [ + 184, + 156, + 294, + 226 + ], + "type": "image", + "image_path": "465163df7e638b6f1e75711d10f4b5589651dcad877eda8e75e2d2fd2562da58.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 230, + 272, + 241 + ], + "lines": [ + { + "bbox": [ + 208, + 230, + 272, + 241 + ], + "spans": [ + { + "bbox": [ + 208, + 230, + 272, + 241 + ], + "type": "text", + "content": "w/ HyperColumn" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 296, + 156, + 406, + 226 + ], + "blocks": [ + { + "bbox": [ + 296, + 156, + 406, + 226 + ], + "lines": [ + { + "bbox": [ + 296, + 156, + 406, + 226 + ], + "spans": [ + { + "bbox": [ + 296, + 156, + 406, + 226 + ], + "type": "image", + "image_path": "cd58a03ed6bfc8461cc78618a477a9a48f2bce6f3120555abd5b93b11b33a2a4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 231, + 361, + 240 + ], + "lines": [ + { + "bbox": [ + 342, + 231, + 361, + 240 + ], + "spans": [ + { + "bbox": [ + 342, + 231, + 361, + 240 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 128, + 251, + 464, + 264 + ], + "lines": [ + { + "bbox": [ + 128, + 251, + 464, + 264 + ], + "spans": [ + { + "bbox": [ + 128, + 251, + 464, + 264 + ], + "type": "text", + "content": "Figure 6: The visual comparison of different settings involved in the ablation study." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 408, + 156, + 518, + 226 + ], + "blocks": [ + { + "bbox": [ + 408, + 156, + 518, + 226 + ], + "lines": [ + { + "bbox": [ + 408, + 156, + 518, + 226 + ], + "spans": [ + { + "bbox": [ + 408, + 156, + 518, + 226 + ], + "type": "image", + "image_path": "0ffb652bc7a2d5885074497b446b50cd811500ae3c8b6d0d050f18e756dc6247.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 231, + 470, + 240 + ], + "lines": [ + { + "bbox": [ + 456, + 231, + 470, + 240 + ], + "spans": [ + { + "bbox": [ + 456, + 231, + 470, + 240 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 88, + 282, + 248, + 387 + ], + "blocks": [ + { + "bbox": [ + 88, + 282, + 248, + 387 + ], + "lines": [ + { + "bbox": [ + 88, + 282, + 248, + 387 + ], + "spans": [ + { + "bbox": [ + 88, + 282, + 248, + 387 + ], + "type": "table", + "html": "
ModelsPSNRSSIM
w/o Recons. Loss24.840.897
w/ Linear Recons.22.210.881
w/o Feature Inter.24.810.895
w/ YTMT Inter.24.890.901
w/o Feature Enc.24.750.903
w/ HyperColumn23.990.894
Ours25.400.905
", + "image_path": "89c09bd080a8a411b794d8ed310ce0cf6d4d9c68f631ed9eb1d1594f37f015a9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 396, + 287, + 502 + ], + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 502 + ], + "type": "text", + "content": "Table 2: Ablation study on different configurations. The initial two rows compare the configuration devoid of any reconstruction loss and that integrating a linear reconstruction loss. Rows 3-4 offer a comparison between the settings lacking feature interaction and incorporating the YTMT feature interaction mechanism. Rows 5-6 shed light on configurations that exclude the feature encoder and utilize HyperColumn as the feature encoder. The last row shows the results of our full version." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 519, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 603 + ], + "type": "text", + "content": "that the linear model is inferior to the setting without any reconstruction criterion because of the existence of " + }, + { + "bbox": [ + 46, + 519, + 287, + 603 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 46, + 519, + 287, + 603 + ], + "type": "text", + "content": ". Meanwhile, without a reconstruction term, the model is likely to have trouble determining the portion of the components, excessively separating content to the reflection layer, resulting in textureless regions in the transmission layer as shown in Fig. 6." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Feature Interaction Mechanism. The feature interaction mechanism connects the two information streams in our proposed DSRNet, both exchanging and conditioning the dual-stream information, which facilitates the layer reconstruction. As shown in Table 2, the model performance appears to degrade after removing the feature interaction (w/o Feature Inter.), which tells us that the network can hardly achieve state-of-the-art performance via a two-branch network without feature interaction. Further, we include the" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 284, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 545, + 354 + ], + "type": "text", + "content": "results of the YTMT feature interaction mechanism (w/ YTMT Inter.) by replacing MuGI with the YTMT mechanism. It can be seen that the model with the YTMT mechanism is only slightly better than the scheme without interaction, which demonstrates the superiority of the proposed MuGI mechanism." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 357, + 546, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 546, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 546, + 489 + ], + "type": "text", + "content": "Feature Encoder. In comparison to the hypercolumn framework introduced by the previous method (referred to as \"w/ HyperColumn\") and the approach that without using any feature encoder (\"w/o Feature Enc.\"), our method demonstrates noteworthy performance advantages. These advantages stem from two main factors. Firstly, our approach employs a hierarchical and gradual reduction of channel dimensions, followed by fusion operations. Secondly, DSFNet leverages a feature interaction mechanism, allowing high-level semantic features to interact before being fed into the subsequent stage." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 501, + 378, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 501, + 378, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 501, + 378, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": "In this paper, we modified the commonly-used linear model in the single image reflection separation task and proposed the reconstruction loss with residual correction. The proposed model is more flexible and effective compared with the previous methods. Meanwhile, we further improved the feature interaction mechanism in dual-stream networks and the usage of hierarchical semantic information, proposed the MuGI as a novel interaction paradigm, and a dual-stream semantic-aware network, namely DSR-Net. Extensive experiments revealed that our proposed method has achieved state-of-the-art performance on all real-world benchmark datasets and verified our contributions. In the future, further constraints are desired to be applied to the residue term to reduce its solution space, and more reflection models are hopefully to be covered by it to fit a wider range of reflection phenomena in the real world." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13145" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Amit K. Agrawal, Ramesh Raskar, Shree K. Nayar, and Yuanzhen Li. Removing photography artifacts using gradient projection and flash-exposure sampling. TOG, 24(3):828-835, 2005. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 287, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 287, + 167 + ], + "type": "text", + "content": "[2] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In ECCV, pages 17-33, 2022. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 287, + 201 + ], + "type": "text", + "content": "[3] Zheng Dong, Ke Xu, Yin Yang, Hujun Bao, Weiwei Xu, and Rynson W. H. Lau. Location-aware single image reflection removal. In ICCV, pages 4997-5006, 2021. 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 203, + 286, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 286, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 286, + 245 + ], + "type": "text", + "content": "[4] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes (VOC) challenge. *IJCV*, 88(2):303–338, 2010. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 286, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 286, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 286, + 290 + ], + "type": "text", + "content": "[5] Qingnan Fan, Jiaolong Yang, Gang Hua, Baoquan Chen, and David P. Wipf. A generic deep architecture for single image reflection removal and image smoothing. In ICCV, pages 3258-3267, 2017. 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 291, + 286, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 286, + 323 + ], + "type": "text", + "content": "[6] Hany Farid and Edward H. Adelson. Separating reflections and lighting using independent components analysis. In CVPR, pages 1262-1267, 1999. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 324, + 286, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 286, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 286, + 368 + ], + "type": "text", + "content": "[7] Xin Feng, Wenjie Pei, Zihui Jia, Fanglin Chen, David Zhang, and Guangming Lu. Deep-masking generative network: A unified framework for background restoration from superimposed images. TIP, 30:4867-4882, 2021. 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 286, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 286, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 286, + 401 + ], + "type": "text", + "content": "[8] Kun Gai, Zhenwei Shi, and Changshui Zhang. Blind separation of superimposed moving images using image statistics. TPAMI, 34(1):19-32, 2012. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 402, + 286, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 286, + 434 + ], + "type": "text", + "content": "[9] Xiaojie Guo, Xiaochun Cao, and Yi Ma. Robust separation of reflection from multiple images. In CVPR, pages 2195-2202, 2014. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 436, + 286, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 436, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 436, + 286, + 468 + ], + "type": "text", + "content": "[10] Byeong-Ju Han and Jae-Young Sim. Reflection removal using low-rank matrix completion. In CVPR, pages 3872–3880, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 469, + 286, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 469, + 286, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 469, + 286, + 512 + ], + "type": "text", + "content": "[11] Bharath Hariharan, Pablo Andres Arbeláez, Ross B. Girshick, and Jitendra Malik. Hypercolumns for object segmentation and fine-grained localization. In CVPR, pages 447-456, 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 514, + 286, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 286, + 555 + ], + "type": "text", + "content": "[12] Qiming Hu and Xiaojie Guo. *Trash or treasure? an interactive dual-stream strategy for single image reflection separation*. In *NeurIPS*, pages 24683–24694, 2021. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 558, + 286, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 286, + 601 + ], + "type": "text", + "content": "[13] Naejin Kong, Yu-Wing Tai, and Joseph S. Shin. A physically-based approach to reflection separation: From physical modeling to constrained optimization. TPAMI, 36(2):209-221, 2014. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 602, + 286, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 286, + 635 + ], + "type": "text", + "content": "[14] Naejin Kong, Yu-Wing Tai, and Sung Yong Shin. High-quality reflection separation using polarized images. TIP, 20(12):3393-3405, 2011. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 636, + 286, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 286, + 668 + ], + "type": "text", + "content": "[15] Chenyang Lei and Qifeng Chen. Robust reflection removal with reflection-free flash-only cues. In CVPR, pages 14811-14820, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[16] Chenyang Lei, Xuhua Huang, Mengdi Zhang, Qiong Yan, Wenxiu Sun, and Qifeng Chen. Polarized reflection removal with perfect alignment in the wild. In CVPR, pages 1747-1755, 2020. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[17] Anat Levin and Yair Weiss. User assisted separation of reflections from a single image using a sparsity prior. TPAMI, 29(9):1647-1654, 2007. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[18] Anat Levin, Assaf Zomet, and Yair Weiss. Learning to perceive transparency from the statistics of natural scenes. In NeurIPS, pages 1247-1254, 2002. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 142, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 142, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 142, + 545, + 175 + ], + "type": "text", + "content": "[19] Anat Levin, Assaf Zomet, and Yair Weiss. Separating reflections from a single image using local features. In CVPR, pages 306-313, 2004. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 177, + 545, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 177, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 177, + 545, + 209 + ], + "type": "text", + "content": "[20] Chao Li, Yixiao Yang, Kun He, Stephen Lin, and John E. Hopcroft. Single image reflection removal through cascaded refinement. In CVPR, pages 3562-3571, 2020. 3, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 211, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 211, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 211, + 545, + 243 + ], + "type": "text", + "content": "[21] Rui Li, Simeng Qiu, Guangming Zang, and Wolfgang Heidrich. Reflection separation via multi-bounce polarization state tracing. In ECCV, pages 781-796, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 246, + 545, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 246, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 307, + 246, + 545, + 278 + ], + "type": "text", + "content": "[22] Yu Li and Michael S. Brown. Exploiting reflection change for automatic reflection removal. In ICCV, pages 2432-2439, 2013. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 280, + 545, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 280, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 280, + 545, + 312 + ], + "type": "text", + "content": "[23] Yu Li and Michael S. Brown. Single image layer separation using relative smoothness. In CVPR, pages 2752-2759, 2014. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 315, + 545, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 315, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 315, + 545, + 357 + ], + "type": "text", + "content": "[24] Yu Li, Ming Liu, Yaling Yi, Qince Li, Dongwei Ren, and Wangmeng Zuo. Two-stage single image reflection removal with reflection-aware guidance. CoRR, abs/2012.00945, 2020. 3, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 360, + 545, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 393 + ], + "type": "text", + "content": "[25] Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Learning to see through obstructions. In CVPR, pages 14203-14212, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 395, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 395, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 395, + 545, + 437 + ], + "type": "text", + "content": "[26] Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Learning to see through obstructions with layered decomposition. TPAMI, 44(11):8387-8402, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 440, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 440, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 440, + 545, + 483 + ], + "type": "text", + "content": "[27] Youwei Lyu, Zhaopeng Cui, Si Li, Marc Pollefeys, and Boxin Shi. Reflection separation using a pair of unpolarized and polarized images. In NeurIPS, pages 14532-14542, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 485, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 485, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 485, + 545, + 518 + ], + "type": "text", + "content": "[28] Shree K. Nayar, Xi-Sheng Fang, and Terrance E. Boult. Separation of reflection components using color and polarization. IJCV, 21(3):163-186, 1997. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 520, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 552 + ], + "type": "text", + "content": "[29] Bernard Sarel and Michal Irani. Separating transparent layers of repetitive dynamic behaviors. In ICCV, pages 26-32, 2005. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 555, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 586 + ], + "type": "text", + "content": "[30] Christian Simon and In Kyu Park. Reflection removal for in-vehicle black box videos. In CVPR, pages 4231–4239, 2015. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "type": "text", + "content": "[31] Sudipta N. Sinha, Johannes Kopf, Michael Goesele, Daniel Scharstein, and Richard Szeliski. Image-based rendering for scenes with reflections. TOG, 31(4):100:1-100:10, 2012. 1, 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 635, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 677 + ], + "type": "text", + "content": "[32] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In CVPR, pages 5769-5780, 2022. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[33] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, and Alex C. Kot. Benchmarking single-image reflection removal algorithms. In ICCV, pages 3942-3950, 2017. 6" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13146" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 464 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[34] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, and Alex C Kot. Crrn: Multi-scale guided concurrent reflection removal network. In CVPR, pages 4777-4785, 2018. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "type": "text", + "content": "[35] Renjie Wan, Boxin Shi, Haoliang Li, Ling-Yu Duan, and Alex C. Kot. Reflection scene separation from a single image. In CVPR, pages 2395-2403, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 286, + 184 + ], + "type": "text", + "content": "[36] Kaixuan Wei, Jiaolong Yang, Ying Fu, David P. Wipf, and Hua Huang. Single image reflection removal exploiting misaligned training data and network enhancements. In CVPR, pages 8178-8187, 2019. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 286, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 286, + 218 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 286, + 218 + ], + "type": "text", + "content": "[37] Qiang Wen, Yinjie Tan, Jing Qin, Wenxi Liu, Guoqiang Han, and Shengfeng He. Single image reflection removal beyond linearity. In CVPR, pages 3771-3779, 2019. 1, 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 220, + 286, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 286, + 252 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 286, + 252 + ], + "type": "text", + "content": "[38] Tianfan Xue, Michael Rubinstein, Ce Liu, and William T. Freeman. A computational approach for obstruction-free photography. TOG, 34(4):79:1-79:11, 2015. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 254, + 286, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 254, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 254, + 286, + 297 + ], + "type": "text", + "content": "[39] Jie Yang, Dong Gong, Lingqiao Liu, and Qinfeng Shi. Seeing deeply and bidirectionally: A deep learning approach for single image reflection removal. In ECCV, pages 675-691, 2018. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 298, + 286, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 298, + 286, + 341 + ], + "spans": [ + { + "bbox": [ + 49, + 298, + 286, + 341 + ], + "type": "text", + "content": "[40] Jiaolong Yang, Hongdong Li, Yuchao Dai, and Robby T. Tan. Robust optical flow estimation of double-layer images under transparency or reflection. In CVPR, pages 1410-1419, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 343, + 286, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 343, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 49, + 343, + 286, + 376 + ], + "type": "text", + "content": "[41] Xuaner Cecilia Zhang, Ren Ng, and Qifeng Chen. Single image reflection separation with perceptual losses. In CVPR, pages 4786-4794, 2018. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 377, + 286, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 377, + 286, + 421 + ], + "spans": [ + { + "bbox": [ + 49, + 377, + 286, + 421 + ], + "type": "text", + "content": "[42] Qian Zheng, Jinnan Chen, Zhan Lu, Boxin Shi, Xudong Jiang, Kim-Hui Yap, Ling-Yu Duan, and Alex C Kot. What does plate glass reveal about camera calibration? In CVPR, pages 3022-3032, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 422, + 286, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 422, + 286, + 464 + ], + "spans": [ + { + "bbox": [ + 49, + 422, + 286, + 464 + ], + "type": "text", + "content": "[43] Qian Zheng, Boxin Shi, Jinnan Chen, Xudong Jiang, LingYu Duan, and Alex C. Kot. Single image reflection removal with absorption effect. In CVPR, pages 13395-13404, 2021. 5, 6" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13147" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Single-Stage Diffusion NeRF_ A Unified Approach to 3D Generation and Reconstruction/06f68dbb-5c23-4704-ae55-c667675f6d9f_content_list.json b/2023/Single-Stage Diffusion NeRF_ A Unified Approach to 3D Generation and Reconstruction/06f68dbb-5c23-4704-ae55-c667675f6d9f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c447beca3224c4ccb9553d0336fbfa7807d90661 --- /dev/null +++ b/2023/Single-Stage Diffusion NeRF_ A Unified Approach to 3D Generation and Reconstruction/06f68dbb-5c23-4704-ae55-c667675f6d9f_content_list.json @@ -0,0 +1,1670 @@ +[ + { + "type": "text", + "text": "Single-Stage Diffusion NeRF: A Unified Approach to 3D Generation and Reconstruction", + "text_level": 1, + "bbox": [ + 218, + 130, + 751, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hansheng Chen, $^{1,*}$ Jiatao Gu, $^{2}$ Anpei Chen, $^{3}$ Wei Tian, $^{1}$ Zhuowen Tu, $^{4}$ Lingjie Liu, $^{5}$ Hao Su $^{4}$", + "bbox": [ + 127, + 202, + 836, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tongji University $^{2}$ Apple $^{3}$ ETH Zürich $^{4}$ University of California, San Diego $^{5}$ University of Pennsylvania", + "bbox": [ + 205, + 224, + 761, + 263 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d1343f1e853acd282882509fbc88fdafe21842eaed4dbf19b49f503e7bd72eb.jpg", + "image_caption": [ + "Figure 1. During training, SSDNeRF jointly learns triplane features of individual scenes, a shared NeRF decoder, and a triplane diffusion prior. During testing, it can perform (a) unconditional generation, (b) single-view reconstruction, as well as multi-view reconstruction." + ], + "image_footnote": [], + "bbox": [ + 98, + 282, + 431, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f339408477b3f3c7291875e05a1e747a03ef1055091c536ce453bb7f528463d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 464, + 282, + 874, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 526, + 313, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D-aware image synthesis encompasses a variety of tasks, such as scene generation and novel view synthesis from images. Despite numerous task-specific methods, developing a comprehensive model remains challenging. In this paper, we present SSDNeRF, a unified approach that employs an expressive diffusion model to learn a generalizable prior of neural radiance fields (NeRF) from multi-view images of diverse objects. Previous studies have used two-stage approaches that rely on pretrained NeRFs as real data to train diffusion models. In contrast, we propose a new single-stage training paradigm with an end-to-end objective that jointly optimizes a NeRF auto-decoder and a latent diffusion model, enabling simultaneous 3D reconstruction and prior learning, even from sparsely available views. At test time, we can directly sample the diffusion prior for unconditional generation, or combine it with arbitrary observations of unseen objects for NeRF reconstruction. SSDNeRF demonstrates robust results comparable to or better than leading task-specific methods in unconditional generation and single/sparse-view 3D reconstruction. $^6$", + "bbox": [ + 73, + 560, + 473, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 526, + 632, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Synthesizing 3D visual contents has gained significant attention in computer vision and graphics, thanks to advances in neural rendering and generative models. Although numerous methods have emerged to handle individual tasks, such as single-/multi-view 3D reconstruction and 3D content generation, it remains a major challenge to develop a comprehensive framework that bridges the state of the art of multiple tasks. For instance, neural radiance fields (NeRF) [28] have shown impressive results in novel view synthesis by solving the inverse rendering problem via perscene fitting, which is suitable for dense-view inputs but difficult to generalize to sparse observations. In contrast, many sparse-view 3D reconstruction methods [55, 8, 25] rely on feed-forward image-to-3D encoders, but they are unable to handle ambiguity in the occluded region and generate crisp images. Regarding unconditional generation, 3D-aware generative adversarial networks (GAN) [31, 5, 16, 13] are partially limited in their usage of single-image discriminators, which cannot reason cross-view relationships to effectively learn from multi-view data.", + "bbox": [ + 496, + 551, + 892, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a unified approach to various 3D tasks (Fig. 1) by developing a holistic model that learns generalizable 3D priors from multi-view images. Inspired", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 47 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during a remote internship with UCSD. \n6Project page: https://lakonik.github.io/ssdnerf", + "bbox": [ + 93, + 875, + 436, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2416", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "by the success of 2D diffusion models [19, 47, 27, 38, 26], we present the Single-Stage Diffusion NeRF (SSDNeRF), which models the generative prior of scene latent codes with a 3D latent diffusion model (LDM).", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While similar LDMs have been applied in 2D and 3D generation in previous work [50, 38, 12, 2, 44, 29], they typically require two-stage training, where the first stage pretrains the variational auto-encoders (VAE) [23] or autodecoders [32] without diffusion models. In the case of diffusion NeRFs, however, we argue that two-stage training induces noisy patterns and artifacts in the latent code due to the uncertain nature of inverse rendering, particularly when training from sparse-view data, which prevents the diffusion model from learning a clean latent manifold effectively. To address this issue, we introduce a novel single-stage training paradigm that enables end-to-end learning of diffusion and NeRF weights ( $\\S 4.1$ ). This approach blends the generative and the rendering biases coherently for improved performance overall and allows for training on sparse-view data. Additionally, we show that the learned 3D priors of unconditional diffusion models can be exploited for flexible test-time scene sampling from arbitrary observations ( $\\S 4.2$ ).", + "bbox": [ + 75, + 152, + 470, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate SSDNeRF on multiple datasets of categorical single-object scenes, demonstrating strong performance overall. Our approach represents a significant step towards a unified framework for various 3D tasks.", + "bbox": [ + 75, + 426, + 468, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our main contributions are as follows:", + "bbox": [ + 96, + 488, + 447, + 502 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce SSDNeRF, a unified approach to all-round performance in unconditional 3D generation and image-based reconstruction;", + "- We propose a novel single-stage training paradigm that jointly learns NeRF reconstruction and diffusion model from multi-view images of a large number of objects. Notably, this enables training on as sparse as three views per scene, which is previously infeasible;", + "- A guidance-finetuning sampling scheme is developed to exploit the learned diffusion priors for 3D reconstruction from arbitrary number of views at test time." + ], + "bbox": [ + 96, + 508, + 467, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 691, + 218, + 707 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D GANs The generative adversarial framework [14] has been successfully adapted for 3D generation by integrating projection-based rendering into the generator. A variety of 3D representations have been explored previously, including point clouds, cuboids, spheres [24] and voxels [30] in early works, the more recent radiance fields [4, 41, 11, 42, 46] and feature fields [31, 16, 5] with volume renderer, and differentiable surface [13] with mesh renderer. The above methods are all trained with 2D image discriminators that are unable to reason cross-view relationships, making them heavily dependent on model bias for 3D consistency. As a result, multi-view data cannot be effectively exploited to", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "learn complex and diverse geometries. 3D GANs are mostly applied in unconditional generation. Although 3D completion from images is possible through GAN inversion [11], faithfulness is not guaranteed due to limited latent expressiveness, as shown in [29, 1].", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "View-Conditioned Regression and Generation Sparse-view 3D reconstruction can be tackled by regressing novel views from input images. Various architectures [8, 55, 25, 57] have been proposed to encode images into volume features, which can be projected to supervised target views through volume rendering. However, they cannot reason ambiguity and generate diverse and meaningful contents, which often leads to blurry results. In contrast, image-conditioned generative models are better at synthesizing distinct contents. 3DiM [53] proposes to generate novel views from a view-conditioned image diffusion model, but the model lacks 3D consistency bias. [58, 10, 17] distill priors of image-conditioned 2D diffusion models into NeRFs to enforce 3D constraints. These methods are parallel to our track as they model cross-view relationships in the image space, while our model is inherently 3D.", + "bbox": [ + 496, + 171, + 890, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Auto-Decoders and Diffusion NeRF NeRF's per-scene fitting scheme can be generalized to multi-scene fitting by sharing part of the parameters across all scenes, leaving the rest as individual scene codes [7]. Therefore, multi-scene NeRFs can be trained as auto-decoders [32], where the code bank and shared decoder weights are jointly learned. With proper architectures, scene codes can be treated as latents with Gaussian priors, allowing 3D completion and even generation [21, 45, 35]. However, like 3D GANs, the latents are not expressive enough for faithful reconstruction of detailed objects. [2, 12, 51] improve upon vanilla auto-decoders with latent diffusion priors. DiffRF [29] leverages the diffusion prior to perform 3D completion. These methods train the auto-decoders and diffusion models in two separate stages, which is subject to the limitations in § 3.2.", + "bbox": [ + 496, + 417, + 890, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 500, + 656, + 627, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. NeRF as an Auto-Decoder", + "text_level": 1, + "bbox": [ + 500, + 681, + 736, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a set of 2D images of a scene and their camera parameters, one can fit a scene model to reconstruct the light field in 3D space, expressed by a plenoptic function $y_{\\psi}(r)$ , where $r$ parameterizes the endpoint and direction of a ray in the world space, $\\psi$ denotes the scene model parameters, and $y \\in \\mathbb{R}_+^3$ represents the received light in RGB format. NeRF [28] represents the light field as integrated radiance along rays through the 3D volume. It models the scene geometry and appearance as functions of the position $p \\in \\mathbb{R}^3$ and view direction $d \\in \\mathbb{R}^3$ of a point in the world space, expressed as $\\rho_{\\psi}(p)$ and $c_{\\psi}(p,d)$ respectively, where $\\rho \\in \\mathbb{R}_{+}$ is the density output and $c \\in \\mathbb{R}_{+}^{3}$ is the RGB color output. Differentiable volume rendering is applied to compose the", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2417", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "received light $y$ from multiple point samples along a ray $r$ .", + "bbox": [ + 75, + 90, + 464, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NeRF can also generalize to multi-scene settings by sharing part of the model parameters across all scenes [7]. Given observations of multiples scenes $\\{y_{ij}^{\\mathrm{gt}},r_{ij}^{\\mathrm{gt}}\\}$ , where $y_{ij}^{\\mathrm{gt}},r_{ij}^{\\mathrm{gt}}$ is the $j$ -th pair of pixel RGB and ray of the $i$ -th scene, one can optimize the per-scene codes $\\{x_i\\}$ and shared parameters $\\psi$ by minimizing the L2 rendering loss:", + "bbox": [ + 75, + 106, + 467, + 196 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e n d}} \\left(\\left\\{x _ {i} \\right\\}, \\psi\\right) = \\mathbb {E} _ {i} \\left[ \\sum_ {j} \\frac {1}{2} \\left\\| y _ {i j} ^ {\\mathrm {g t}} - y _ {\\psi} \\left(x _ {i}, r _ {i j} ^ {\\mathrm {g t}}\\right) \\right\\| ^ {2} \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 203, + 468, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With this objective, the model is trained as an auto-decoder [32], where the scene codes $\\{x_{i}\\}$ can be interpreted as the latent codes, and the plenoptic function can be regarded as a decoder in the form of $p_{\\psi}(\\{y_j\\} |x,\\{r_j\\})\\coloneqq \\prod_j\\mathcal{N}(y_j|y_\\psi (x,r_j),I)$ , assuming independent Gaussians.", + "bbox": [ + 75, + 244, + 467, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Challenges in Bridging Generation and Reconstruction An auto-decoder with trained weights $\\psi$ can perform unconditional generation by decoding latent codes drawn from a Gaussian prior [35]. However, to ensure continuity in generation, a low-dimensional latent space and a complex decoder is required, which adds to the difficulty in optimizing the latent code to faithfully reconstruct any given views.", + "bbox": [ + 75, + 324, + 467, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Latent Diffusion Models", + "text_level": 1, + "bbox": [ + 76, + 438, + 302, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent diffusion models (LDM) learn a prior distribution $p_{\\phi}(x)$ in the latent space with parameters $\\phi$ , which enables the usage of more expressive latent representations, such as 2D grids for images [50, 38]. For neural field generation, previous work [2, 29, 12, 44] adopts a two-stage training scheme, where the auto-decoder is trained first to obtain the per-scene latent $x_{i}$ , which is then treated as real data to train the LDM. The LDM injects Gaussian perturbation $\\epsilon \\sim \\mathcal{N}(0, I)$ into the code $x_{i}$ , yielding a noisy code $x_{i}^{(t)} \\coloneqq \\alpha^{(t)} x_{i} + \\sigma^{(t)} \\epsilon$ at diffusion time step $t$ , under empirical noise schedule functions $\\alpha^{(t)}, \\sigma^{(t)}$ . A denoising network with trainable weights $\\phi$ is then tasked with removing the noise from $x_{i}^{(t)}$ to predict a denoised code $\\hat{x}_{i}$ . The network is typically trained with a simplified L2 denoising loss:", + "bbox": [ + 75, + 460, + 468, + 678 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d i f f}} (\\phi) = \\underset {i, t, \\epsilon} {\\mathbb {E}} \\left[ \\frac {1}{2} w ^ {(t)} \\left\\| \\hat {x} _ {\\phi} \\left(x _ {i} ^ {(t)}, t\\right) - x _ {i} \\right\\| ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 684, + 468, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t \\sim \\mathcal{U}(0,T)$ , $w^{(t)}$ is an empirical time dependent weighting function, and $\\hat{x}_{\\phi}(x_i^{(t)},t)$ formulates the time-conditioned denoising network.", + "bbox": [ + 75, + 727, + 468, + 775 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unconditional/Guided Sampling With trained weights $\\phi$ , one can sample from the diffusion prior $p_{\\psi}(x)$ using a variety of solvers (e.g., DDIM [47]) that recursively denoise $x^{(t)}$ , starting from random Gaussian noise $x^{(T)}$ , until reaching the denoised state $x^{(0)}$ . Moreover, the sampling process can be guided by the gradients of the rendering loss against known observations, allowing 3D reconstruction from images at test time [29].", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c311386e4d3a3a5f7956006093d8918813e36ffac7196bd2bfacf7dd1dcd8568.jpg", + "image_caption": [ + "Figure 2. Visualization of the scene code $x_{\\mathrm{XZ}}$ at XZ plane. Left column: Two-stage training without TV regularization induces noise and fails in 3-view reconstruction. Mid column: TV regularization imposes smoothing prior at the cost of textural details (top), yet still struggles to cope with sparse views (bottom). Right column: Our single-stage training encourages smooth yet detailed latents and allows for training with sparse views." + ], + "image_footnote": [], + "bbox": [ + 532, + 88, + 841, + 204 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Limitations of Two-Stage Training for 3D Tasks While LDMs with 2D image VAEs are typically trained in two stages [50, 38], training LDMs with NeRF auto-decoders poses an unprecedented challenge. An expressive latent code is underdetermined when obtained via rendering-based optimization, leading to noisy patterns that distract denoising networks (top-left of Fig. 2). Additionally, reconstructing NeRFs from sparse views without a learned prior is exceptionally difficult (bottom-left of Fig. 2), limiting training to dense-views settings.", + "bbox": [ + 496, + 330, + 890, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Proposed Method", + "text_level": 1, + "bbox": [ + 500, + 494, + 671, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To build a holistic model that unifies 3D generation and reconstruction, we propose SSDNeRF, a framework that conjoins the expressive triplane NeRF auto-decoder with a triplane latent diffusion model. Fig. 3 provides an overview of the model. In the following subsections, we elaborate on how training and testing are performed in detail.", + "bbox": [ + 496, + 520, + 890, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Single-Stage Diffusion NeRF Training", + "text_level": 1, + "bbox": [ + 498, + 619, + 826, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An auto-decoder can be regarded as a type of VAE that uses a lookup table encoder instead of the typical neural network encoder. As such, the training objective can be derived in a similar way as for VAEs. With NeRF decoder $p_{\\psi}(\\{y_j\\} |x,\\{r_j\\})$ and diffusion latent prior $p_{\\phi}(x)$ , the training objective is to minimize variational upper bound on the negative log-likelihood (NLL) of observed data $\\{y_{ij}^{\\mathrm{gt}}, r_{ij}^{\\mathrm{gt}}\\}$ [23, 36, 50]. In this paper, a simplified training loss is derived by ignoring the uncertainty (variance) in latent codes:", + "bbox": [ + 496, + 643, + 890, + 792 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\underbrace {\\underset {i} {\\mathbb {E}} \\left[ - \\log p _ {\\psi} \\left(\\left\\{y _ {i j} ^ {\\mathrm {g t}} \\right\\} \\mid x _ {i} , \\left\\{r _ {i j} ^ {\\mathrm {g t}} \\right\\}\\right) \\right]} _ {\\text {r e n d e r i n g l o s s} \\mathcal {L} _ {\\text {r e n d}}} + \\underbrace {\\underset {i} {\\mathbb {E}} \\left[ - \\log p _ {\\phi} \\left(x _ {i}\\right) \\right]} _ {\\text {p r i o r t e r m}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 801, + 890, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the scene codes $\\{x_{i}\\}$ , prior parameters $\\phi$ , and decoder parameters $\\psi$ are jointly optimized in a single training stage. This loss function consists of the rendering loss $\\mathcal{L}_{\\mathrm{rend}}$", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2418", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8e145baf531a412b27e037624d3792ff4f3700cd97c4869b6aefbcae9804a2c1.jpg", + "image_caption": [ + "Figure 3. An overview of SSDNeRF framework with a triplane NeRF representation. During training, we feed a batch of observations in the format of RGB values $y_{ij}^{\\mathrm{gt}}$ and rays $r_{ij}^{\\mathrm{gt}}$ . The corresponding scene code $x_{i}$ is randomly initialized and optimized by minimizing the rendering loss $\\mathcal{L}_{\\mathrm{rend}}$ and the diffusion loss $\\mathcal{L}_{\\mathrm{diff}}$ , and model parameters $\\phi, \\psi$ are also updated along the way." + ], + "image_footnote": [], + "bbox": [ + 135, + 87, + 841, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in Eq. (1) and a diffusion prior term in the form of NLL. Following [50, 54, 48], we replace the diffusion NLL with its approximate upper bound $\\mathcal{L}_{\\mathrm{diff}}$ in Eq. (2). This technique is also called score distillation in [33]. Adding empirical weighting factors, we finalize our training objective as:", + "bbox": [ + 75, + 328, + 470, + 404 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {\\text {r e n d}} \\mathcal {L} _ {\\text {r e n d}} (\\{x _ {i} \\}, \\psi) + \\lambda_ {\\text {d i f f}} \\mathcal {L} _ {\\text {d i f f}} (\\{x _ {i} \\}, \\phi). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 412, + 468, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Single-stage training constrains scene codes $\\{x_{i}\\}$ with both terms in the loss function, allowing the learned prior to complete the parts unseen to rendering. This is particularly beneficial to training on sparse-view data, where the expressive triplane codes are severely underdetermined.", + "bbox": [ + 75, + 438, + 468, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Balancing Rendering and Prior Weights The render-to-prior weight ratio $\\lambda_{\\mathrm{rend}} / \\lambda_{\\mathrm{diff}}$ is crucial to single-stage training. To make hyperparameters more generalizable to different settings, we design an empirical weighting mechanism, in which the diffusion loss is normalized by the exponential moving average (EMA) of the scene codes' Frobenius norms, expressed as $\\lambda_{\\mathrm{diff}} \\coloneqq c_{\\mathrm{diff}} / EMA\\left(\\|x_i\\|_F^2\\right)$ with a constant scale $c_{\\mathrm{diff}}$ , and the rendering weight is determined by the number of views available $N_{\\mathrm{v}}$ , expressed as $\\lambda_{\\mathrm{rend}} \\coloneqq c_{\\mathrm{rend}}(1 - e^{-0.1N_{\\mathrm{v}}}) / N_{\\mathrm{v}}$ with a constant scale $c_{\\mathrm{rend}}$ . Intuitively, $N_{\\mathrm{v}}$ -based weighting is a calibration to the ray independence assumption in the decoder $p_{\\psi}(\\{y_j\\} | x, \\{r_j\\}) \\coloneqq \\prod_j \\mathcal{N}(y_j | y_{\\psi}(x, r_j), I)$ , preventing the rendering loss from scaling linearly with the number of rays.", + "bbox": [ + 75, + 518, + 468, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Comparison to Two-Stage Generative Neural Fields Previous two-stage methods [2, 12, 29, 44] ignore the prior term $\\lambda_{\\mathrm{diff}}\\mathcal{L}_{\\mathrm{diff}}$ during the first stage of training the auto-decoders. This can be seen as setting the render-to-prior weight ratio $\\lambda_{\\mathrm{rend}} / \\lambda_{\\mathrm{diff}}$ to infinity, resulting in biased and noisy scene codes $x_{i}$ . Shue et al. [44] partially mitigate this issue by imposing total variation (TV) regularization on triplane scene codes to enforce a smoothing prior, which resembles the LDM constraints on the latent space (mid column of Fig. 2). Control3Diff [15] proposes to learn a conditional diffusion model on data generated by a 3D GAN pre", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "trained on single-view images. In contrast, our single-stage training aims to directly incorporate the diffusion prior to promote end-to-end coherence.", + "bbox": [ + 496, + 328, + 890, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Image-Guided Sampling and Finetuning", + "text_level": 1, + "bbox": [ + 498, + 380, + 846, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve generalizable test-time NeRF reconstruction that covers a wide spectrum from single-view to dense observations, we propose performing image-guided sampling and then finetuning the sampled codes considering both the diffusion prior and rendering likelihood.", + "bbox": [ + 496, + 402, + 890, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the reconstruction-guided sampling method by Ho et al. [20], we compute the approximated rendering gradients $g$ w.r.t. a noisy code $x^{(t)}$ , defined as:", + "bbox": [ + 496, + 479, + 890, + 525 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng \\leftarrow \\nabla_ {x ^ {(t)}} \\lambda_ {\\text {r e n d}} \\sum_ {j} \\frac {1}{2} \\left(\\frac {\\alpha^ {(t)}}{\\sigma^ {(t)}}\\right) ^ {2 \\omega} \\left\\| y _ {j} ^ {\\mathrm {g t}} - y _ {\\psi} \\left(\\hat {x} _ {\\phi} \\left(x ^ {(t)}, t\\right), r _ {j} ^ {\\mathrm {g t}}\\right) \\right\\| ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 529, + 890, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\left(\\alpha^{(t)} / \\sigma^{(t)}\\right)^{2\\omega}$ is an additional weighting factor based on signal-to-noise ratio (SNR), with hyperparameter $\\omega$ chosen to be 0.5 or 0.25 in our work. The guidance gradients $g$ are then combined with unconditional score prediction, expressed as a correction to the denoising output $\\hat{x}$ :", + "bbox": [ + 496, + 579, + 890, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} \\leftarrow \\hat {x} - \\lambda_ {\\mathrm {g d}} \\frac {\\sigma^ {(t) ^ {2}}}{\\alpha^ {(t)}} g \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 662, + 890, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with guidance scale $\\lambda_{\\mathrm{gd}}$ . We adopt the predictor-corrector sampler [49] to solve $x^{(0)}$ by alternating between a DDIM step [47] and multiple Langevin correction steps.", + "bbox": [ + 496, + 700, + 890, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We observe that the reconstruction guidance alone cannot strictly enforce rendering constraints towards faithful reconstruction. To address this issue, we reuse the training objective in Eq. (4) to finetune the sampled scene code $x$ , while freezing the diffusion and decoder parameters:", + "bbox": [ + 496, + 746, + 890, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {x} \\lambda_ {\\text {r e n d}} \\mathcal {L} _ {\\text {r e n d}} (x) + \\lambda_ {\\text {d i f f}} ^ {\\prime} \\mathcal {L} _ {\\text {d i f f}} (x), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 828, + 890, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{diff}}^{\\prime}$ is the test-time prior weight, which we find should be lower than the training weight $\\lambda_{\\mathrm{diff}}$ for best results, as the prior learned from the training dataset is less", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2419", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "reliable when transferred to a different testing dataset. We use Adam [22] to optimize the code $x$ for finetuning.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison to Previous NeRF Finetuning Approaches While finetuning with rendering loss is common in view-conditioned NeRF regression methods [8, 57], our finetuning approach differs in the use of diffusion prior loss on the 3D scene code, which significantly enhances generalization to novel views, as demonstrated in § 5.3.", + "bbox": [ + 76, + 127, + 468, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 227, + 294, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This subsection briefly describes some important technical details. More details can be found in the supplementary.", + "bbox": [ + 76, + 250, + 468, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prior Gradient Caching Triplane NeRF reconstruction requires at least hundreds of optimization iterations on each scene code $x_{i}$ . A problem with the single-stage training loss in Eq. (4) is that the diffusion loss $\\mathcal{L}_{\\mathrm{diff}}$ requires much longer time to evaluate than the native NeRF rendering loss $\\mathcal{L}_{\\mathrm{rend}}$ , reducing overall efficiency. To accelerate reconstruction in both training and test-time finetuning, we introduce a technique called prior gradient caching, which caches the back-propagated prior gradients $\\nabla_{x}\\lambda_{\\mathrm{diff}}\\mathcal{L}_{\\mathrm{diff}}$ for re-use in multiple Adam steps, while refreshing the rendering gradients $\\nabla_{x}\\lambda_{\\mathrm{rend}}\\mathcal{L}_{\\mathrm{rend}}$ in each of the steps, which allows for fewer diffusion passes than rendering. A training pseudocode is given in Algorithm 1.", + "bbox": [ + 76, + 286, + 468, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Denoising Parameterization and Weighting The denoising model $\\hat{x}_{\\phi}(x^{(t)}, t)$ is implemented as a U-Net [39] as in DDPM [19], with a total of 122M parameters. Its input and output are noisy and denoised triplane features, respectively, with channels of all three planes stacked together. For the prediction format, we adopt the $v$ -parameterization $\\hat{v}_{\\phi}(x^{(t)}, t)$ in [40], such that $\\hat{x} = \\alpha^{(t)} x^{(t)} - \\sigma^{(t)} \\hat{v}$ . Regarding the weighting function $w^{(t)}$ in the diffusion loss in Eq. (2), LSGM [50] employs two different mechanisms for optimizing latents $x_{i}$ and diffusion weights $\\phi$ , respectively, which we find unstable with NeRF auto-decoders. Instead, we observe that the SNR-based weighting $w^{(t)} = \\left(\\alpha^{(t)} / \\sigma^{(t)}\\right)^{2\\omega}$ used in Eq. (5) works well with our models.", + "bbox": [ + 76, + 488, + 468, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 700, + 209, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 726, + 179, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct experiments on the ShapeNet SRN [6, 45] and Amazon Berkeley Objects (ABO) Tables [9] datasets for benchmarking with previous work. The SRN dataset provides single-object scenes in two categories, i.e., Cars and Chairs, with a train/test split of 2458/703 for Cars and 4612/1317 for Chairs. Each train scene has 50 random views from a sphere and each test scene has 251 spiral views from the upper hemisphere. The ABO Tables dataset provides a train/test split of 1520/156 table scenes, where each scene has 91 views from the upper hemisphere. For", + "bbox": [ + 76, + 750, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Single-stage diffusion NeRF training" + ], + "code_body": "Input: $\\{y_{ij}^{\\mathrm{gt}},r_{ij}^{\\mathrm{gt}}\\}$ \n1 Initialize $\\{x_i\\} ,\\phi ,\\psi$ \n2 for $k_{\\mathrm{out}}\\coloneqq 1\\dots K_{\\mathrm{out}}$ do // outer loop of $K_{\\mathrm{out}}$ iterations \n3 Sample a batch of scenes $i\\in B_{\\mathrm{sc}}$ \n4 $g_{\\phi},g_{x}^{\\mathrm{diff}}\\gets \\nabla_{\\phi ,\\{x_{i}\\}_{B_{\\mathrm{sc}}}}\\lambda_{\\mathrm{diff}}\\mathcal{L}_{\\mathrm{diff}}$ // diffusion grad \n5 $\\phi \\leftarrow \\phi -Adam(g_{\\phi})$ \n6 for $k_{\\mathrm{in}}\\coloneqq 1\\dots K_{\\mathrm{in}}$ do // inner loop of $K_{\\mathrm{in}}$ iterations \n7 Sample a batch of rays $j\\in B_{\\mathrm{ray}}$ \n8 $g_{x}^{\\mathrm{rend}}\\gets \\nabla_{\\{x_i\\}_{B_{\\mathrm{sc}}}}\\lambda_{\\mathrm{rend}}\\mathcal{L}_{\\mathrm{rend}}$ // rendering grad \n9 $g_{x}\\gets g_{x}^{\\mathrm{rend}} + g_{x}^{\\mathrm{diff}}$ // add cached prior grad \n10 $\\{x_i\\}_{B_{\\mathrm{sc}}} \\leftarrow \\{x_i\\}_{B_{\\mathrm{sc}}} - Adam(g_x)$ \n11 if $k_{\\mathrm{in}} = K_{\\mathrm{in}}$ then // last inner iteration \n12 $\\begin{array}{rl} & g_{\\psi}\\gets \\nabla_{\\psi}\\lambda_{\\mathrm{rend}}\\mathcal{L}_{\\mathrm{rend}}\\\\ & \\psi \\leftarrow \\psi -Adam(g_{\\psi}) \\end{array}$ \n13", + "bbox": [ + 500, + 112, + 888, + 344 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1b3e61d3ae55170fd3b258bf85ccd1420bb185c650921f09a7bdd36450bb4858.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Unconditional generation results on SRN Cars and ABO Tables. † denotes results reported by Functa [12]. § denotes results reported by DiffRF [29]. * denotes results reproduced by us using the official public code with a bugfix. We show ±2σ intervals." + ], + "table_body": "
MethodTypeCarsTables
FID↓KID/10-3↓FID↓KID/10-3↓
Functa [12]LDM80.3---
π-GAN [4]GAN36.7†-41.67§13.82§
EG3D [5]GAN10.46*4.90*31.18§11.67§
DiffRF [29]LDM--27.0610.03
Ours (2-stage)LDM16.33±0.936.38±0.41--
Ours (1-stage)LDM11.08±1.113.47±0.2314.27±0.664.08±0.33
", + "bbox": [ + 509, + 369, + 887, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "both datasets, we use the provided renderings (resized to $128 \\times 128$ ) with ground truth poses for training and testing.", + "bbox": [ + 498, + 583, + 890, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Unconditional Generation", + "text_level": 1, + "bbox": [ + 500, + 623, + 736, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conduct evaluations for unconditional generation using the SRN Cars and ABO Tables dataset. The Cars dataset poses a challenge in generating sharp and intricate textures, whereas the Tables dataset comprises of diverse geometries with realistic materials. Models are trained on all images of the training set for 1M iterations.", + "bbox": [ + 498, + 646, + 890, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Protocol and Metrics For SRN Cars, following Functa [12], we sample 704 scenes from the diffusion model, and render each scene using the fixed 251 camera poses from the test set. For ABO Tables, following DiffRF [29], we sample 1000 scenes and render each scene with 10 random cameras. We adopt standard generation metrics including Fréchet Inception Distance (FID) [18] and Kernel Inception Distance (KID) [3]. The metrics' reference sets are all images in the test set for SRN Cars and all", + "bbox": [ + 496, + 741, + 890, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{7}$ https://github.com/nvlabs/eg3d/issues/67", + "bbox": [ + 514, + 886, + 841, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2420", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2edc5250fd0c2cbf1e22213f5c01dceeb54f2123df734d67bc9fd201d6e45a25.jpg", + "image_caption": [ + "(a) EG3D Tables" + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 339, + 223 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c402b9128283d431983729c792e80d3f6820e86608bc750ac552a63ab4539da9.jpg", + "image_caption": [ + "(b) DiffRF Tables" + ], + "image_footnote": [], + "bbox": [ + 361, + 88, + 611, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c784415f9ed0e3707e9c95fdd9dc45973dc1943923630e2974d84fbcedd6f1e4.jpg", + "image_caption": [ + "(c) SSDNeRF Tables (ours)" + ], + "image_footnote": [], + "bbox": [ + 633, + 88, + 890, + 223 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3722d95486aec137dd0b0e2ee8cc2a7dd306816752499b3d40b309e9eb27857f.jpg", + "image_caption": [ + "(d) EG3D Cars" + ], + "image_footnote": [], + "bbox": [ + 83, + 244, + 465, + 364 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/abd7f77808ff61a642e84969ed5f20c754a8b5922571cd827a8866b62685f627.jpg", + "image_caption": [ + "(e) SSDNeRF Cars (Ours)", + "Figure 4. Qualitative comparison between unconditional generative models trained on ABO Tables and SRN Cars." + ], + "image_footnote": [], + "bbox": [ + 503, + 244, + 885, + 364 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4e9d518a2993259e53992a6380787e2dcd7e5ce029a159483d51f2da5ec9cac2.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Single-view and two-view reconstruction results on SRN Cars and Chairs. For consistency with prior work, we use view #64 of the test scene as single-view input and view #64 and #104 as two-view input. † denotes results reported by 3DiM [53]. ‡ denotes results reported by VisionNeRF [25], § denotes results reported by PixelNeRF [21], * denotes results reproduced by us using the official code. - indicates results are unavailable." + ], + "table_body": "
MethodCars 1-viewCars 2-viewChairs 1-viewChairs 2-view
PSNR↑SSIM↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓FID↓
3DiM [53]21.010.57-8.99----17.050.53-6.57----
PixelNeRF [55]23.170.900.146‡59.24†25.660.94--23.720.910.128‡38.49†26.200.94--
SRN [45]22.25§0.89§0.129‡41.21†24.84§0.92§--22.89§0.89§0.104‡26.51†24.48§0.92§--
CodeNeRF [21]23.800.910.118*56.34*25.710.930.108*56.13*23.660.900.106*31.65*25.630.910.097*29.90*
VisionNeRF [25]22.880.910.08421.31†----24.480.930.07710.05†----
Ours (1-stage)23.520.910.07816.3926.490.940.05410.6624.350.930.06710.1326.940.950.05510.85
", + "bbox": [ + 84, + 414, + 888, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "images in the entire dataset for ABO Tables, respectively.", + "bbox": [ + 75, + 623, + 455, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison to the State of the Art As shown in Table 1, on SRN Cars, SSDNeRF (1-stage) outperforms EG3D in KID (a more suitable measure for small datasets) by a clear margin. Meanwhile, its FID is drastically better than Functa, which uses an LDM but with low dimensional latent codes. On ABO Tables, SSDNeRF shows significantly better performance than EG3D and DiffRF.", + "bbox": [ + 75, + 645, + 468, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Single- vs. Two-stage On SRN Cars, we compare the proposed single-stage training against two-stage training with tuned TV regularization using the same model architecture. The results in Table 1 indicate substantial advantage of single-stage training (KID/10 $^{-3}$ 3.47 vs. 6.38).", + "bbox": [ + 75, + 757, + 468, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results As shown in Fig. 4, SSDNeRF generates more regular geometries than the slightly skewed and distorted shapes by EG3D [5]. Compared to DiffRF [29], our method produces sharp details and reflective materials,", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "thanks to our more expressive model with latents of higher spatial resolution and view-dependent NeRF decoder.", + "bbox": [ + 496, + 623, + 892, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Sparse-View NeRF Reconstruction", + "text_level": 1, + "bbox": [ + 498, + 660, + 803, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section presents experiments on 3D reconstruction from sparse-view images of unseen objects in SRN Cars and Chairs test sets. The Cars dataset presents the challenge of recovering distinct textures, while the Chairs dataset requires accurate reconstruction of diverse shapes. Models are trained on all images of the training set for 80K iterations, as we find that longer schedule leads to decaying performance in reconstructing unseen objects. This behaviour is in accordance with the interpolation results in § 5.5.", + "bbox": [ + 496, + 684, + 892, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Protocol and Metrics We use the evaluation protocol and metrics in PixelNeRF [55]. Given input images sampled from each test scene, we obtain the triplane scene code via guidance-finetuning and evaluate novel view synthesis quality with respect to the unseen images.", + "bbox": [ + 496, + 825, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2421", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/49d6562de1a695052892f6a857e6a41805f3310a43157d3f04148967a61f6fb8.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of single-view reconstruction methods on unseen test objects in SRN Cars and Chairs." + ], + "image_footnote": [], + "bbox": [ + 78, + 85, + 486, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d4af214eaa886395617b5124a9cbb17b7208255d6465580ae0cc20b603d9aae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 85, + 893, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The image quality metrics include average peak signal-to-noise-ratio (PSNR), structural similarity (SSIM) [52], and Learned Perceptual Image Patch Similarity (LPIPS) [56]. In addition, we evaluate the FID between all synthesized images and ground truth images as in 3DiM [53].", + "bbox": [ + 75, + 383, + 468, + 459 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison to the State of the Art Table 2 compares SSDNeRF against previous approaches in single-view and two-view reconstruction settings. Overall, SSDNeRF reaches the best LPIPS of all tasks, indicating the best perceptual fidelity. In contrast, 3DiM generates high quality images (best FID) but with the lowest fidelity to the ground truth (lowest PSNR); CodeNeRF reports the best PSNR on single-view Cars, but its limited expressiveness leads to blurry outputs (Fig. 5) and less competitive LPIPS and FID; VisionNeRF achieves a balanced performance on all single-view metrics, but may struggle to generate textural details on the unseen side of cars (e.g., the other side of the ambulance in Fig. 5). Moreover, SSDNeRF exhibits a clear advantage in two-view reconstruction, achieving the best performance on all relevant metrics.", + "bbox": [ + 75, + 465, + 470, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Single- vs. Two-stage As demonstrated in Table 3, the model trained in a single stage (A0) outperforms the same architecture trained in two stages with TV regularization (A1) in all metrics of single-view reconstruction.", + "bbox": [ + 75, + 698, + 468, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Studies on Test-Time Finetuning As shown in Table 3, we evaluate the effectiveness of test-time finetuning and the contribution of the learned diffusion prior with two ablation experiments: (A2) removing the diffusion loss during finetuning and using only the rendering loss, and (A3) omitting the finetuning process entirely. The results indicate that finetuning with single-view rendering loss provides only marginal improvements over guided sampling (A2 vs. A3), while the learned diffusion prior significantly", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5eade70fbe34051b3c7f33d7cffe72bdfc360e2696f6478c67cdf039eb2eb663.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IDTrainingFinetuningPSNR↑SSIM↑LPIPS↓FID↓
A01-stageRend + Diff23.520.9130.07816.39
A12-stageRend + Diff22.830.9060.09020.97
A21-stageRend23.130.9070.08827.93
A31-stageNone23.070.9050.09230.95
", + "bbox": [ + 522, + 375, + 870, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation results on single-view reconstruction of SRN Cars.", + "bbox": [ + 500, + 460, + 890, + 486 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6be3c0ffb17e1dfb41f87f2dc35db23292d7a4fab94c65b9ab3a7c2097c00cd5.jpg", + "image_caption": [ + "Figure 6. LPIPS scores (lower is better) of novel view synthesis from sparse-to-dense inputs, evaluated on SRN Cars test set. The triplane baselines adopt mean initialization for better performance." + ], + "image_footnote": [], + "bbox": [ + 549, + 491, + 841, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "boosts the LPIPS and FID scores (A0 vs. A2), highlighting its importance in recovering sharp and distinct contents. Moreover, the qualitative results in Fig. 5 reveal that views with higher overlap to the input view benefit the most from finetuning, meeting our expectation that finetuning helps faithfully reconstruct the exact observations.", + "bbox": [ + 496, + 667, + 890, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sparse-to-Dense Reconstruction To validate that SSDNeRF seamlessly bridges sparse- and dense-view NeRF reconstruction, we evaluate its novel view synthesis performance with the number of input views varying from 1 to 32. We compare our model to the triplane NeRF baseline trained as an auto-decoder with optional TV regularization instead of diffusion prior. Meanwhile, we also evaluate CodeNeRF [21], an auto-decoder with 256-d latent codes. The results in Fig. 6 show that SSDNeRF excels in all settings,", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2422", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "especially in 1 to 4 views. In contrast, CodeNeRF is outperformed by vanilla triplane NeRF with more views.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Training SSDNeRF on Sparse-View Dataset", + "text_level": 1, + "bbox": [ + 76, + 128, + 449, + 145 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we train SSDNeRF on a sparse-view subset of the full SRN Cars training set, in which a fixed set of only three views are randomly picked from each scene. Note that a reasonable decline in performance compared to dense-view training is expected as the whole training dataset has been reduced to $6\\%$ of its original size.", + "bbox": [ + 75, + 152, + 468, + 243 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unconditional Generation We adopt a training trick that resets the triplane codes to their mean value halfway through training. This helps to prevent the model from getting stuck in a local minimum that overfits geometric artifacts. We also double the length of the training schedule accordingly. The model achieves a decent FID of $19.04 \\pm 1.10$ and a $\\mathrm{KID} / 10^{-3}$ of $8.28 \\pm 0.60$ . Results are visualized in Fig. 7.", + "bbox": [ + 75, + 248, + 468, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Single-View Reconstruction We adopt the same training strategy as in § 5.3. With our guidance-finetuning approach, the model achieves an LPIPS score of 0.106, even outperforming most of the previous methods in Table 2 that use the full training set.", + "bbox": [ + 75, + 358, + 468, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison to TV Regularization Fig. 8 (b) shows the RGB images and geometries represented by the scene latent codes learned from three views during training. By comparison, vanilla triplane auto-decoder with TV regularization (Fig. 8 (a)) often fails to reconstruct a scene from sparse views, leading to severe geometric artifacts. As a result, previously it has been infeasible to train two-stage models with expressive latents on sparse-view data.", + "bbox": [ + 75, + 438, + 468, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. NeRF Interpolation", + "text_level": 1, + "bbox": [ + 76, + 566, + 264, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Following DDIM [47], we can sample two initial values $x^{(T)} \\sim \\mathcal{N}(0, I)$ , interpolate them using spherical linear interpolation [43], and then use the deterministic solver to obtain interpolated samples. However, as noted by [34, 37], standard Gaussian diffusion models often result in nonsmooth interpolation. In SSDNeRF (with results shown in Fig. 9), we find that the model (a) trained with early stopping for sparse-view reconstruction produces reasonably smooth transitions between samples, while the model (b) trained with a longer schedule for unconditional generation produces distinct yet discontinuous samples. This suggests that early stopping preserves a smoother prior, leading to better generalization for sparse-view reconstruction.", + "bbox": [ + 75, + 590, + 468, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 800, + 194, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose SSDNeRF, which combines the diffusion model and NeRF representation through a novel single-stage training paradigm with an end-to-end justifiable loss. Notably, it overcomes the limitations in previous work where implicit neural fields must be obtained from", + "bbox": [ + 75, + 824, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3d9b02bc5936756dd254acea334242c101e7eee1d099efa483db6bdf6f7315ff.jpg", + "image_caption": [ + "Figure 7. Images generated by SSDNeRF trained on a 3-view subset of SRN Cars training set." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 883, + 146 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6528710554a00ad0a8d10d9183222b477c74e7c9556896a9090b92c9b4ff213d.jpg", + "image_caption": [ + "(a) Auto-decoder training w/ TV" + ], + "image_footnote": [], + "bbox": [ + 508, + 195, + 679, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f229c7790d811fd3eb42ec3ec9a2b8f1224523d1fbe345e9855e7ddf5ca49142.jpg", + "image_caption": [ + "(b) 1-stage joint training with diffusion" + ], + "image_footnote": [], + "bbox": [ + 705, + 205, + 883, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d7a5022e3f655c21a283a26ff6eeb65cfcd0ec2a845f2ff9dcab5c6b464ceb80.jpg", + "image_caption": [ + "Figure 8. Qualitative comparison between scene codes learned from 3 views by (a) triplane auto-decoder with TV regularization vs. (b) single-stage diffusion NeRF.", + "(a) Models trained with early stopping (80K iters) for sparse-view reconstruction" + ], + "image_footnote": [], + "bbox": [ + 501, + 342, + 888, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9948b75b0a03f87389db5a7e462e11c4c2badd23c3f280bb2db8b73ecc382d61.jpg", + "image_caption": [ + "(b) Models trained with long schedule (1M iters) for unconditional generation", + "Figure 9. Interpolation between the leftmost and rightmost samples using DDIM [47]." + ], + "image_footnote": [], + "bbox": [ + 501, + 416, + 887, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dense observations first, before training the diffusion models to learn their manifold. With strong performance on multiple benchmarks, SSDNeRF demonstrates a significant advancement towards a unified framework for general 3D content manipulation.", + "bbox": [ + 496, + 541, + 890, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work Currently, our method relies on ground truth camera parameters during both training and testing. Future work may explore transform-invariant models. Additionally, the diffusion prior can become discontinuous with prolonged training, which affects generalization. Although early stopping is temporarily used, a better network design or a larger training dataset may be able to address this problem fundamentally.", + "bbox": [ + 496, + 622, + 890, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements We thank Norman Müller for sharing the baseline results on ABO Tables. Hansheng Chen and Wei Tian acknowledge the funding by the National Natural Science Foundation of China (No. 52002285), the Shanghai Science and Technology Commission (No. 21ZR1467400), the original research project of Tongji University (No. 22120220593), the National Key R&D Program of China (No. 2021YFB2501104), and the Natural Science Foundation of Chongqing (No. 2023NSCQ-MSX4511).", + "bbox": [ + 496, + 750, + 892, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2423", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J. Mitra, and Paul Guerrero. RenderDiffusion: Image diffusion for 3D reconstruction, inpainting and generation. In CVPR, 2023. 2", + "[2] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, Afshin Dehghan, and Josh Susskind. Gaudi: A neural architect for immersive 3d scene generation. In NeurIPS, 2022. 2, 3, 4", + "[3] Mikołaj Binkowski, Danica J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In ICLR, 2018. 5", + "[4] Eric Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 2, 5", + "[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2, 5, 6", + "[6] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 5", + "[7] Anpei Chen, Zexiang Xu, Xinyue Wei, Siyu Tang, Hao Su, and Andreas Geiger. Factor fields: A unified framework for neural fields and beyond, 2023. 2, 3", + "[8] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, pages 14124-14133, 2021. 1, 2, 5", + "[9] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik. Abo: Dataset and benchmarks for real-world 3d object understanding. In CVPR, 2022. 5", + "[10] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In CVPR, 2023. 2", + "[11] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W. Taylor, and Joshua M. Susskind. Unconstrained scene generation with locally conditioned radiance fields. In ICCV, 2021. 2", + "[12] Emilien Dupont, Hyunjik Kim, S. M. Ali Eslami, Danilo Jimenez Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you can treat it like one. In ICML, 2022. 2, 3, 4, 5", + "[13] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In NeurIPS, 2022. 1, 2", + "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2", + "[15] Jiatao Gu, Qingzhe Gao, Shuangfei Zhai, Baoquan Chen, Lingjie Liu, and Josh Susskind. Learning controllable 3d diffusion models from single-view images, 2023. 4", + "[16] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In ICLR, 2022. 1, 2", + "[17] Jiatao Gu, Alex Trevithick, Kai-En Lin, Josh Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nerf-guided distillation from 3d-aware diffusion, 2023. 2", + "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 5", + "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 5", + "[20] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 4", + "[21] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In ICCV, pages 12949-12958, 2021. 2, 6, 7", + "[22] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5", + "[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 2, 3", + "[24] Yiyi Liao, Katja Schwarz, Lars Mescheder, and Andreas Geiger. Towards unsupervised learning of generative models for 3d controllable image synthesis. In CVPR, 2020. 2", + "[25] Kai-En Lin, Lin Yen-Chen, Wei-Sheng Lai, Tsung-Yi Lin, Yi-Chang Shih, and Ravi Ramamoorthi. Vision transformer for nef-based view synthesis from a single input image. In WACV, 2023. 1, 2, 6", + "[26] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11461-11471, 2022. 2", + "[27] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 2", + "[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2", + "[29] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In CVPR, 2023. 2, 3, 4, 5, 6", + "[30] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2424", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learning of 3d representations from natural images. In ICCV, 2019. 2", + "[31] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 1, 2", + "[32] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 2, 3", + "[33] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 4", + "[34] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In CVPR, 2022. 8", + "[35] Daniel Rebain, Mark Matthews, Kwang Moo Yi, Dmitry Lagun, and Andrea Tagliasacchi. Lolnerf: Learn from one look. In CVPR, 2022. 2, 3", + "[36] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In ICML, pages 1278-1286, 2014. 3", + "[37] Severi Rissanen, Markus Heinonen, and Arno Solin. Generative modelling with inverse heat dissipation. In ICLR, 2023. 8", + "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3", + "[39] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241, 2015. 5", + "[40] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In ICLR, 2022. 5", + "[41] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In NeurIPS, 2020. 2", + "[42] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. In NeurIPS, 2022. 2", + "[43] Ken Shoemake. Animating rotation with quaternion curves. In Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 245-254, 1985. 8", + "[44] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In CVPR, 2023. 2, 3, 4", + "[45] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In NeurIPS, 2019. 2, 5, 6", + "[46] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. In NeurIPS, 2022. 2", + "[47] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2, 3, 4, 8" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. Maximum likelihood training of score-based diffusion models. In NeurIPS, 2021. 4", + "[49] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 4", + "[50] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In NeurlPS, 2021. 2, 3, 4, 5", + "[51] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In CVPR, 2023. 2", + "[52] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 7", + "[53] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. In ICLR, 2023. 2, 6, 7", + "[54] Antoine Wehenkel and Gilles Louppe. Diffusion priors in variational autoencoders. In ICML Workshops, 2021. 4", + "[55] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 1, 2, 6", + "[56] Richard Zhang, Phillip Isola, Alexei Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7", + "[57] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In CVPR, 2022. 2, 5", + "[58] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In CVPR, 2023. 2" + ], + "bbox": [ + 501, + 92, + 890, + 602 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2425", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Single-Stage Diffusion NeRF_ A Unified Approach to 3D Generation and Reconstruction/06f68dbb-5c23-4704-ae55-c667675f6d9f_model.json b/2023/Single-Stage Diffusion NeRF_ A Unified Approach to 3D Generation and Reconstruction/06f68dbb-5c23-4704-ae55-c667675f6d9f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..aa0872722198f2aa4b208481381ea111dca0aa25 --- /dev/null +++ b/2023/Single-Stage Diffusion NeRF_ A Unified Approach to 3D Generation and Reconstruction/06f68dbb-5c23-4704-ae55-c667675f6d9f_model.json @@ -0,0 +1,2431 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.048 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.131, + 0.753, + 0.175 + ], + "angle": 0, + "content": "Single-Stage Diffusion NeRF: A Unified Approach to 3D Generation and Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.203, + 0.837, + 0.224 + ], + "angle": 0, + "content": "Hansheng Chen,\\(^{1,*}\\) Jiatao Gu,\\(^{2}\\) Anpei Chen,\\(^{3}\\) Wei Tian,\\(^{1}\\) Zhuowen Tu,\\(^{4}\\) Lingjie Liu,\\(^{5}\\) Hao Su\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.226, + 0.763, + 0.265 + ], + "angle": 0, + "content": "\\(^{1}\\)Tongji University \\(^{2}\\)Apple \\(^{3}\\)ETH Zürich \\(^{4}\\)University of California, San Diego \\(^{5}\\)University of Pennsylvania" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.284, + 0.432, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.283, + 0.875, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.478, + 0.893, + 0.507 + ], + "angle": 0, + "content": "Figure 1. During training, SSDNeRF jointly learns triplane features of individual scenes, a shared NeRF decoder, and a triplane diffusion prior. During testing, it can perform (a) unconditional generation, (b) single-view reconstruction, as well as multi-view reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.527, + 0.314, + 0.544 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.561, + 0.474, + 0.864 + ], + "angle": 0, + "content": "3D-aware image synthesis encompasses a variety of tasks, such as scene generation and novel view synthesis from images. Despite numerous task-specific methods, developing a comprehensive model remains challenging. In this paper, we present SSDNeRF, a unified approach that employs an expressive diffusion model to learn a generalizable prior of neural radiance fields (NeRF) from multi-view images of diverse objects. Previous studies have used two-stage approaches that rely on pretrained NeRFs as real data to train diffusion models. In contrast, we propose a new single-stage training paradigm with an end-to-end objective that jointly optimizes a NeRF auto-decoder and a latent diffusion model, enabling simultaneous 3D reconstruction and prior learning, even from sparsely available views. At test time, we can directly sample the diffusion prior for unconditional generation, or combine it with arbitrary observations of unseen objects for NeRF reconstruction. SSDNeRF demonstrates robust results comparable to or better than leading task-specific methods in unconditional generation and single/sparse-view 3D reconstruction.\\(^6\\)" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.527, + 0.633, + 0.543 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.893, + 0.854 + ], + "angle": 0, + "content": "Synthesizing 3D visual contents has gained significant attention in computer vision and graphics, thanks to advances in neural rendering and generative models. Although numerous methods have emerged to handle individual tasks, such as single-/multi-view 3D reconstruction and 3D content generation, it remains a major challenge to develop a comprehensive framework that bridges the state of the art of multiple tasks. For instance, neural radiance fields (NeRF) [28] have shown impressive results in novel view synthesis by solving the inverse rendering problem via perscene fitting, which is suitable for dense-view inputs but difficult to generalize to sparse observations. In contrast, many sparse-view 3D reconstruction methods [55, 8, 25] rely on feed-forward image-to-3D encoders, but they are unable to handle ambiguity in the occluded region and generate crisp images. Regarding unconditional generation, 3D-aware generative adversarial networks (GAN) [31, 5, 16, 13] are partially limited in their usage of single-image discriminators, which cannot reason cross-view relationships to effectively learn from multi-view data." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose a unified approach to various 3D tasks (Fig. 1) by developing a holistic model that learns generalizable 3D priors from multi-view images. Inspired" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.875, + 0.437, + 0.902 + ], + "angle": 0, + "content": "*Work done during a remote internship with UCSD. \n6Project page: https://lakonik.github.io/ssdnerf" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2416" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.152 + ], + "angle": 0, + "content": "by the success of 2D diffusion models [19, 47, 27, 38, 26], we present the Single-Stage Diffusion NeRF (SSDNeRF), which models the generative prior of scene latent codes with a 3D latent diffusion model (LDM)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.426 + ], + "angle": 0, + "content": "While similar LDMs have been applied in 2D and 3D generation in previous work [50, 38, 12, 2, 44, 29], they typically require two-stage training, where the first stage pretrains the variational auto-encoders (VAE) [23] or autodecoders [32] without diffusion models. In the case of diffusion NeRFs, however, we argue that two-stage training induces noisy patterns and artifacts in the latent code due to the uncertain nature of inverse rendering, particularly when training from sparse-view data, which prevents the diffusion model from learning a clean latent manifold effectively. To address this issue, we introduce a novel single-stage training paradigm that enables end-to-end learning of diffusion and NeRF weights (\\(\\S 4.1\\)). This approach blends the generative and the rendering biases coherently for improved performance overall and allows for training on sparse-view data. Additionally, we show that the learned 3D priors of unconditional diffusion models can be exploited for flexible test-time scene sampling from arbitrary observations (\\(\\S 4.2\\))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.47, + 0.487 + ], + "angle": 0, + "content": "We evaluate SSDNeRF on multiple datasets of categorical single-object scenes, demonstrating strong performance overall. Our approach represents a significant step towards a unified framework for various 3D tasks." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.489, + 0.449, + 0.503 + ], + "angle": 0, + "content": "To summarize, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.51, + 0.468, + 0.553 + ], + "angle": 0, + "content": "- We introduce SSDNeRF, a unified approach to all-round performance in unconditional 3D generation and image-based reconstruction;" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.555, + 0.468, + 0.629 + ], + "angle": 0, + "content": "- We propose a novel single-stage training paradigm that jointly learns NeRF reconstruction and diffusion model from multi-view images of a large number of objects. Notably, this enables training on as sparse as three views per scene, which is previously infeasible;" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.631, + 0.468, + 0.674 + ], + "angle": 0, + "content": "- A guidance-finetuning sampling scheme is developed to exploit the learned diffusion priors for 3D reconstruction from arbitrary number of views at test time." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.51, + 0.468, + 0.674 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.693, + 0.219, + 0.708 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "3D GANs The generative adversarial framework [14] has been successfully adapted for 3D generation by integrating projection-based rendering into the generator. A variety of 3D representations have been explored previously, including point clouds, cuboids, spheres [24] and voxels [30] in early works, the more recent radiance fields [4, 41, 11, 42, 46] and feature fields [31, 16, 5] with volume renderer, and differentiable surface [13] with mesh renderer. The above methods are all trained with 2D image discriminators that are unable to reason cross-view relationships, making them heavily dependent on model bias for 3D consistency. As a result, multi-view data cannot be effectively exploited to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "learn complex and diverse geometries. 3D GANs are mostly applied in unconditional generation. Although 3D completion from images is possible through GAN inversion [11], faithfulness is not guaranteed due to limited latent expressiveness, as shown in [29, 1]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.172, + 0.892, + 0.414 + ], + "angle": 0, + "content": "View-Conditioned Regression and Generation Sparse-view 3D reconstruction can be tackled by regressing novel views from input images. Various architectures [8, 55, 25, 57] have been proposed to encode images into volume features, which can be projected to supervised target views through volume rendering. However, they cannot reason ambiguity and generate diverse and meaningful contents, which often leads to blurry results. In contrast, image-conditioned generative models are better at synthesizing distinct contents. 3DiM [53] proposes to generate novel views from a view-conditioned image diffusion model, but the model lacks 3D consistency bias. [58, 10, 17] distill priors of image-conditioned 2D diffusion models into NeRFs to enforce 3D constraints. These methods are parallel to our track as they model cross-view relationships in the image space, while our model is inherently 3D." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.645 + ], + "angle": 0, + "content": "Auto-Decoders and Diffusion NeRF NeRF's per-scene fitting scheme can be generalized to multi-scene fitting by sharing part of the parameters across all scenes, leaving the rest as individual scene codes [7]. Therefore, multi-scene NeRFs can be trained as auto-decoders [32], where the code bank and shared decoder weights are jointly learned. With proper architectures, scene codes can be treated as latents with Gaussian priors, allowing 3D completion and even generation [21, 45, 35]. However, like 3D GANs, the latents are not expressive enough for faithful reconstruction of detailed objects. [2, 12, 51] improve upon vanilla auto-decoders with latent diffusion priors. DiffRF [29] leverages the diffusion prior to perform 3D completion. These methods train the auto-decoders and diffusion models in two separate stages, which is subject to the limitations in § 3.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.657, + 0.628, + 0.673 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.682, + 0.737, + 0.697 + ], + "angle": 0, + "content": "3.1. NeRF as an Auto-Decoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Given a set of 2D images of a scene and their camera parameters, one can fit a scene model to reconstruct the light field in 3D space, expressed by a plenoptic function \\( y_{\\psi}(r) \\), where \\( r \\) parameterizes the endpoint and direction of a ray in the world space, \\( \\psi \\) denotes the scene model parameters, and \\( y \\in \\mathbb{R}_+^3 \\) represents the received light in RGB format. NeRF [28] represents the light field as integrated radiance along rays through the 3D volume. It models the scene geometry and appearance as functions of the position \\( p \\in \\mathbb{R}^3 \\) and view direction \\( d \\in \\mathbb{R}^3 \\) of a point in the world space, expressed as \\( \\rho_{\\psi}(p) \\) and \\( c_{\\psi}(p,d) \\) respectively, where \\( \\rho \\in \\mathbb{R}_{+} \\) is the density output and \\( c \\in \\mathbb{R}_{+}^{3} \\) is the RGB color output. Differentiable volume rendering is applied to compose the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2417" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.465, + 0.107 + ], + "angle": 0, + "content": "received light \\(y\\) from multiple point samples along a ray \\(r\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.468, + 0.198 + ], + "angle": 0, + "content": "NeRF can also generalize to multi-scene settings by sharing part of the model parameters across all scenes [7]. Given observations of multiples scenes \\(\\{y_{ij}^{\\mathrm{gt}},r_{ij}^{\\mathrm{gt}}\\}\\), where \\(y_{ij}^{\\mathrm{gt}},r_{ij}^{\\mathrm{gt}}\\) is the \\(j\\)-th pair of pixel RGB and ray of the \\(i\\)-th scene, one can optimize the per-scene codes \\(\\{x_i\\}\\) and shared parameters \\(\\psi\\) by minimizing the L2 rendering loss:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.204, + 0.47, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e n d}} \\left(\\left\\{x _ {i} \\right\\}, \\psi\\right) = \\mathbb {E} _ {i} \\left[ \\sum_ {j} \\frac {1}{2} \\left\\| y _ {i j} ^ {\\mathrm {g t}} - y _ {\\psi} \\left(x _ {i}, r _ {i j} ^ {\\mathrm {g t}}\\right) \\right\\| ^ {2} \\right]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.245, + 0.468, + 0.322 + ], + "angle": 0, + "content": "With this objective, the model is trained as an auto-decoder [32], where the scene codes \\(\\{x_{i}\\}\\) can be interpreted as the latent codes, and the plenoptic function can be regarded as a decoder in the form of \\(p_{\\psi}(\\{y_j\\} |x,\\{r_j\\})\\coloneqq \\prod_j\\mathcal{N}(y_j|y_\\psi (x,r_j),I)\\), assuming independent Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.468, + 0.431 + ], + "angle": 0, + "content": "Challenges in Bridging Generation and Reconstruction An auto-decoder with trained weights \\(\\psi\\) can perform unconditional generation by decoding latent codes drawn from a Gaussian prior [35]. However, to ensure continuity in generation, a low-dimensional latent space and a complex decoder is required, which adds to the difficulty in optimizing the latent code to faithfully reconstruct any given views." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.439, + 0.303, + 0.453 + ], + "angle": 0, + "content": "3.2. Latent Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.47, + 0.679 + ], + "angle": 0, + "content": "Latent diffusion models (LDM) learn a prior distribution \\( p_{\\phi}(x) \\) in the latent space with parameters \\( \\phi \\), which enables the usage of more expressive latent representations, such as 2D grids for images [50, 38]. For neural field generation, previous work [2, 29, 12, 44] adopts a two-stage training scheme, where the auto-decoder is trained first to obtain the per-scene latent \\( x_{i} \\), which is then treated as real data to train the LDM. The LDM injects Gaussian perturbation \\( \\epsilon \\sim \\mathcal{N}(0, I) \\) into the code \\( x_{i} \\), yielding a noisy code \\( x_{i}^{(t)} \\coloneqq \\alpha^{(t)} x_{i} + \\sigma^{(t)} \\epsilon \\) at diffusion time step \\( t \\), under empirical noise schedule functions \\( \\alpha^{(t)}, \\sigma^{(t)} \\). A denoising network with trainable weights \\( \\phi \\) is then tasked with removing the noise from \\( x_{i}^{(t)} \\) to predict a denoised code \\( \\hat{x}_{i} \\). The network is typically trained with a simplified L2 denoising loss:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.685, + 0.47, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d i f f}} (\\phi) = \\underset {i, t, \\epsilon} {\\mathbb {E}} \\left[ \\frac {1}{2} w ^ {(t)} \\left\\| \\hat {x} _ {\\phi} \\left(x _ {i} ^ {(t)}, t\\right) - x _ {i} \\right\\| ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.728, + 0.47, + 0.776 + ], + "angle": 0, + "content": "where \\(t \\sim \\mathcal{U}(0,T)\\), \\(w^{(t)}\\) is an empirical time dependent weighting function, and \\(\\hat{x}_{\\phi}(x_i^{(t)},t)\\) formulates the time-conditioned denoising network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Unconditional/Guided Sampling With trained weights \\(\\phi\\), one can sample from the diffusion prior \\(p_{\\psi}(x)\\) using a variety of solvers (e.g., DDIM [47]) that recursively denoise \\(x^{(t)}\\), starting from random Gaussian noise \\(x^{(T)}\\), until reaching the denoised state \\(x^{(0)}\\). Moreover, the sampling process can be guided by the gradients of the rendering loss against known observations, allowing 3D reconstruction from images at test time [29]." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.089, + 0.843, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.212, + 0.895, + 0.31 + ], + "angle": 0, + "content": "Figure 2. Visualization of the scene code \\( x_{\\mathrm{XZ}} \\) at XZ plane. Left column: Two-stage training without TV regularization induces noise and fails in 3-view reconstruction. Mid column: TV regularization imposes smoothing prior at the cost of textural details (top), yet still struggles to cope with sparse views (bottom). Right column: Our single-stage training encourages smooth yet detailed latents and allows for training with sparse views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.892, + 0.484 + ], + "angle": 0, + "content": "Limitations of Two-Stage Training for 3D Tasks While LDMs with 2D image VAEs are typically trained in two stages [50, 38], training LDMs with NeRF auto-decoders poses an unprecedented challenge. An expressive latent code is underdetermined when obtained via rendering-based optimization, leading to noisy patterns that distract denoising networks (top-left of Fig. 2). Additionally, reconstructing NeRFs from sparse views without a learned prior is exceptionally difficult (bottom-left of Fig. 2), limiting training to dense-views settings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.496, + 0.673, + 0.513 + ], + "angle": 0, + "content": "4. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.612 + ], + "angle": 0, + "content": "To build a holistic model that unifies 3D generation and reconstruction, we propose SSDNeRF, a framework that conjoins the expressive triplane NeRF auto-decoder with a triplane latent diffusion model. Fig. 3 provides an overview of the model. In the following subsections, we elaborate on how training and testing are performed in detail." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.62, + 0.828, + 0.637 + ], + "angle": 0, + "content": "4.1. Single-Stage Diffusion NeRF Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.793 + ], + "angle": 0, + "content": "An auto-decoder can be regarded as a type of VAE that uses a lookup table encoder instead of the typical neural network encoder. As such, the training objective can be derived in a similar way as for VAEs. With NeRF decoder \\( p_{\\psi}(\\{y_j\\} |x,\\{r_j\\}) \\) and diffusion latent prior \\( p_{\\phi}(x) \\), the training objective is to minimize variational upper bound on the negative log-likelihood (NLL) of observed data \\( \\{y_{ij}^{\\mathrm{gt}}, r_{ij}^{\\mathrm{gt}}\\} \\) [23, 36, 50]. In this paper, a simplified training loss is derived by ignoring the uncertainty (variance) in latent codes:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.803, + 0.892, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\underbrace {\\underset {i} {\\mathbb {E}} \\left[ - \\log p _ {\\psi} \\left(\\left\\{y _ {i j} ^ {\\mathrm {g t}} \\right\\} \\mid x _ {i} , \\left\\{r _ {i j} ^ {\\mathrm {g t}} \\right\\}\\right) \\right]} _ {\\text {r e n d e r i n g l o s s} \\mathcal {L} _ {\\text {r e n d}}} + \\underbrace {\\underset {i} {\\mathbb {E}} \\left[ - \\log p _ {\\phi} \\left(x _ {i}\\right) \\right]} _ {\\text {p r i o r t e r m}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where the scene codes \\(\\{x_{i}\\}\\), prior parameters \\(\\phi\\), and decoder parameters \\(\\psi\\) are jointly optimized in a single training stage. This loss function consists of the rendering loss \\(\\mathcal{L}_{\\mathrm{rend}}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2418" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.136, + 0.088, + 0.843, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.261, + 0.895, + 0.306 + ], + "angle": 0, + "content": "Figure 3. An overview of SSDNeRF framework with a triplane NeRF representation. During training, we feed a batch of observations in the format of RGB values \\( y_{ij}^{\\mathrm{gt}} \\) and rays \\( r_{ij}^{\\mathrm{gt}} \\). The corresponding scene code \\( x_{i} \\) is randomly initialized and optimized by minimizing the rendering loss \\( \\mathcal{L}_{\\mathrm{rend}} \\) and the diffusion loss \\( \\mathcal{L}_{\\mathrm{diff}} \\), and model parameters \\( \\phi, \\psi \\) are also updated along the way." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.329, + 0.471, + 0.405 + ], + "angle": 0, + "content": "in Eq. (1) and a diffusion prior term in the form of NLL. Following [50, 54, 48], we replace the diffusion NLL with its approximate upper bound \\(\\mathcal{L}_{\\mathrm{diff}}\\) in Eq. (2). This technique is also called score distillation in [33]. Adding empirical weighting factors, we finalize our training objective as:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.413, + 0.47, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {\\text {r e n d}} \\mathcal {L} _ {\\text {r e n d}} (\\{x _ {i} \\}, \\psi) + \\lambda_ {\\text {d i f f}} \\mathcal {L} _ {\\text {d i f f}} (\\{x _ {i} \\}, \\phi). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.515 + ], + "angle": 0, + "content": "Single-stage training constrains scene codes \\(\\{x_{i}\\}\\) with both terms in the loss function, allowing the learned prior to complete the parts unseen to rendering. This is particularly beneficial to training on sparse-view data, where the expressive triplane codes are severely underdetermined." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.47, + 0.732 + ], + "angle": 0, + "content": "Balancing Rendering and Prior Weights The render-to-prior weight ratio \\(\\lambda_{\\mathrm{rend}} / \\lambda_{\\mathrm{diff}}\\) is crucial to single-stage training. To make hyperparameters more generalizable to different settings, we design an empirical weighting mechanism, in which the diffusion loss is normalized by the exponential moving average (EMA) of the scene codes' Frobenius norms, expressed as \\(\\lambda_{\\mathrm{diff}} \\coloneqq c_{\\mathrm{diff}} / EMA\\left(\\|x_i\\|_F^2\\right)\\) with a constant scale \\(c_{\\mathrm{diff}}\\), and the rendering weight is determined by the number of views available \\(N_{\\mathrm{v}}\\), expressed as \\(\\lambda_{\\mathrm{rend}} \\coloneqq c_{\\mathrm{rend}}(1 - e^{-0.1N_{\\mathrm{v}}}) / N_{\\mathrm{v}}\\) with a constant scale \\(c_{\\mathrm{rend}}\\). Intuitively, \\(N_{\\mathrm{v}}\\)-based weighting is a calibration to the ray independence assumption in the decoder \\(p_{\\psi}(\\{y_j\\} | x, \\{r_j\\}) \\coloneqq \\prod_j \\mathcal{N}(y_j | y_{\\psi}(x, r_j), I)\\), preventing the rendering loss from scaling linearly with the number of rays." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Comparison to Two-Stage Generative Neural Fields Previous two-stage methods [2, 12, 29, 44] ignore the prior term \\(\\lambda_{\\mathrm{diff}}\\mathcal{L}_{\\mathrm{diff}}\\) during the first stage of training the auto-decoders. This can be seen as setting the render-to-prior weight ratio \\(\\lambda_{\\mathrm{rend}} / \\lambda_{\\mathrm{diff}}\\) to infinity, resulting in biased and noisy scene codes \\(x_{i}\\). Shue et al. [44] partially mitigate this issue by imposing total variation (TV) regularization on triplane scene codes to enforce a smoothing prior, which resembles the LDM constraints on the latent space (mid column of Fig. 2). Control3Diff [15] proposes to learn a conditional diffusion model on data generated by a 3D GAN pre" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.329, + 0.892, + 0.374 + ], + "angle": 0, + "content": "trained on single-view images. In contrast, our single-stage training aims to directly incorporate the diffusion prior to promote end-to-end coherence." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.381, + 0.847, + 0.398 + ], + "angle": 0, + "content": "4.2. Image-Guided Sampling and Finetuning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.404, + 0.892, + 0.48 + ], + "angle": 0, + "content": "To achieve generalizable test-time NeRF reconstruction that covers a wide spectrum from single-view to dense observations, we propose performing image-guided sampling and then finetuning the sampled codes considering both the diffusion prior and rendering likelihood." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.892, + 0.526 + ], + "angle": 0, + "content": "Following the reconstruction-guided sampling method by Ho et al. [20], we compute the approximated rendering gradients \\( g \\) w.r.t. a noisy code \\( x^{(t)} \\), defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.502, + 0.53, + 0.892, + 0.58 + ], + "angle": 0, + "content": "\\[\ng \\leftarrow \\nabla_ {x ^ {(t)}} \\lambda_ {\\text {r e n d}} \\sum_ {j} \\frac {1}{2} \\left(\\frac {\\alpha^ {(t)}}{\\sigma^ {(t)}}\\right) ^ {2 \\omega} \\left\\| y _ {j} ^ {\\mathrm {g t}} - y _ {\\psi} \\left(\\hat {x} _ {\\phi} \\left(x ^ {(t)}, t\\right), r _ {j} ^ {\\mathrm {g t}}\\right) \\right\\| ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.658 + ], + "angle": 0, + "content": "where \\(\\left(\\alpha^{(t)} / \\sigma^{(t)}\\right)^{2\\omega}\\) is an additional weighting factor based on signal-to-noise ratio (SNR), with hyperparameter \\(\\omega\\) chosen to be 0.5 or 0.25 in our work. The guidance gradients \\(g\\) are then combined with unconditional score prediction, expressed as a correction to the denoising output \\(\\hat{x}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.663, + 0.892, + 0.698 + ], + "angle": 0, + "content": "\\[\n\\hat {x} \\leftarrow \\hat {x} - \\lambda_ {\\mathrm {g d}} \\frac {\\sigma^ {(t) ^ {2}}}{\\alpha^ {(t)}} g \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.892, + 0.747 + ], + "angle": 0, + "content": "with guidance scale \\(\\lambda_{\\mathrm{gd}}\\). We adopt the predictor-corrector sampler [49] to solve \\(x^{(0)}\\) by alternating between a DDIM step [47] and multiple Langevin correction steps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.892, + 0.823 + ], + "angle": 0, + "content": "We observe that the reconstruction guidance alone cannot strictly enforce rendering constraints towards faithful reconstruction. To address this issue, we reuse the training objective in Eq. (4) to finetune the sampled scene code \\( x \\), while freezing the diffusion and decoder parameters:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.829, + 0.892, + 0.851 + ], + "angle": 0, + "content": "\\[\n\\min _ {x} \\lambda_ {\\text {r e n d}} \\mathcal {L} _ {\\text {r e n d}} (x) + \\lambda_ {\\text {d i f f}} ^ {\\prime} \\mathcal {L} _ {\\text {d i f f}} (x), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{diff}}^{\\prime}\\) is the test-time prior weight, which we find should be lower than the training weight \\(\\lambda_{\\mathrm{diff}}\\) for best results, as the prior learned from the training dataset is less" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2419" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "reliable when transferred to a different testing dataset. We use Adam [22] to optimize the code \\( x \\) for finetuning." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.128, + 0.47, + 0.218 + ], + "angle": 0, + "content": "Comparison to Previous NeRF Finetuning Approaches While finetuning with rendering loss is common in view-conditioned NeRF regression methods [8, 57], our finetuning approach differs in the use of diffusion prior loss on the 3D scene code, which significantly enhances generalization to novel views, as demonstrated in § 5.3." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.228, + 0.295, + 0.244 + ], + "angle": 0, + "content": "4.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.251, + 0.469, + 0.282 + ], + "angle": 0, + "content": "This subsection briefly describes some important technical details. More details can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.287, + 0.47, + 0.484 + ], + "angle": 0, + "content": "Prior Gradient Caching Triplane NeRF reconstruction requires at least hundreds of optimization iterations on each scene code \\( x_{i} \\). A problem with the single-stage training loss in Eq. (4) is that the diffusion loss \\( \\mathcal{L}_{\\mathrm{diff}} \\) requires much longer time to evaluate than the native NeRF rendering loss \\( \\mathcal{L}_{\\mathrm{rend}} \\), reducing overall efficiency. To accelerate reconstruction in both training and test-time finetuning, we introduce a technique called prior gradient caching, which caches the back-propagated prior gradients \\( \\nabla_{x}\\lambda_{\\mathrm{diff}}\\mathcal{L}_{\\mathrm{diff}} \\) for re-use in multiple Adam steps, while refreshing the rendering gradients \\( \\nabla_{x}\\lambda_{\\mathrm{rend}}\\mathcal{L}_{\\mathrm{rend}} \\) in each of the steps, which allows for fewer diffusion passes than rendering. A training pseudocode is given in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.489, + 0.47, + 0.69 + ], + "angle": 0, + "content": "Denoising Parameterization and Weighting The denoising model \\(\\hat{x}_{\\phi}(x^{(t)}, t)\\) is implemented as a U-Net [39] as in DDPM [19], with a total of 122M parameters. Its input and output are noisy and denoised triplane features, respectively, with channels of all three planes stacked together. For the prediction format, we adopt the \\(v\\)-parameterization \\(\\hat{v}_{\\phi}(x^{(t)}, t)\\) in [40], such that \\(\\hat{x} = \\alpha^{(t)} x^{(t)} - \\sigma^{(t)} \\hat{v}\\). Regarding the weighting function \\(w^{(t)}\\) in the diffusion loss in Eq. (2), LSGM [50] employs two different mechanisms for optimizing latents \\(x_{i}\\) and diffusion weights \\(\\phi\\), respectively, which we find unstable with NeRF auto-decoders. Instead, we observe that the SNR-based weighting \\(w^{(t)} = \\left(\\alpha^{(t)} / \\sigma^{(t)}\\right)^{2\\omega}\\) used in Eq. (5) works well with our models." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.702, + 0.21, + 0.719 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.727, + 0.18, + 0.741 + ], + "angle": 0, + "content": "5.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We conduct experiments on the ShapeNet SRN [6, 45] and Amazon Berkeley Objects (ABO) Tables [9] datasets for benchmarking with previous work. The SRN dataset provides single-object scenes in two categories, i.e., Cars and Chairs, with a train/test split of 2458/703 for Cars and 4612/1317 for Chairs. Each train scene has 50 random views from a sphere and each test scene has 251 spiral views from the upper hemisphere. The ABO Tables dataset provides a train/test split of 1520/156 table scenes, where each scene has 91 views from the upper hemisphere. For" + }, + { + "type": "code_caption", + "bbox": [ + 0.51, + 0.095, + 0.85, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: Single-stage diffusion NeRF training" + }, + { + "type": "algorithm", + "bbox": [ + 0.501, + 0.113, + 0.89, + 0.345 + ], + "angle": 0, + "content": "Input: \\(\\{y_{ij}^{\\mathrm{gt}},r_{ij}^{\\mathrm{gt}}\\}\\) \n1 Initialize \\(\\{x_i\\} ,\\phi ,\\psi\\) \n2 for \\(k_{\\mathrm{out}}\\coloneqq 1\\dots K_{\\mathrm{out}}\\) do // outer loop of \\(K_{\\mathrm{out}}\\) iterations \n3 Sample a batch of scenes \\(i\\in B_{\\mathrm{sc}}\\) \n4 \\(g_{\\phi},g_{x}^{\\mathrm{diff}}\\gets \\nabla_{\\phi ,\\{x_{i}\\}_{B_{\\mathrm{sc}}}}\\lambda_{\\mathrm{diff}}\\mathcal{L}_{\\mathrm{diff}}\\) // diffusion grad \n5 \\(\\phi \\leftarrow \\phi -Adam(g_{\\phi})\\) \n6 for \\(k_{\\mathrm{in}}\\coloneqq 1\\dots K_{\\mathrm{in}}\\) do // inner loop of \\(K_{\\mathrm{in}}\\) iterations \n7 Sample a batch of rays \\(j\\in B_{\\mathrm{ray}}\\) \n8 \\(g_{x}^{\\mathrm{rend}}\\gets \\nabla_{\\{x_i\\}_{B_{\\mathrm{sc}}}}\\lambda_{\\mathrm{rend}}\\mathcal{L}_{\\mathrm{rend}}\\) // rendering grad \n9 \\(g_{x}\\gets g_{x}^{\\mathrm{rend}} + g_{x}^{\\mathrm{diff}}\\) // add cached prior grad \n10 \\(\\{x_i\\}_{B_{\\mathrm{sc}}} \\leftarrow \\{x_i\\}_{B_{\\mathrm{sc}}} - Adam(g_x)\\) \n11 if \\(k_{\\mathrm{in}} = K_{\\mathrm{in}}\\) then // last inner iteration \n12 \\(\\begin{array}{rl} & g_{\\psi}\\gets \\nabla_{\\psi}\\lambda_{\\mathrm{rend}}\\mathcal{L}_{\\mathrm{rend}}\\\\ & \\psi \\leftarrow \\psi -Adam(g_{\\psi}) \\end{array}\\) \n13" + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.37, + 0.888, + 0.499 + ], + "angle": 0, + "content": "
MethodTypeCarsTables
FID↓KID/10-3↓FID↓KID/10-3↓
Functa [12]LDM80.3---
π-GAN [4]GAN36.7†-41.67§13.82§
EG3D [5]GAN10.46*4.90*31.18§11.67§
DiffRF [29]LDM--27.0610.03
Ours (2-stage)LDM16.33±0.936.38±0.41--
Ours (1-stage)LDM11.08±1.113.47±0.2314.27±0.664.08±0.33
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.504, + 0.892, + 0.56 + ], + "angle": 0, + "content": "Table 1. Unconditional generation results on SRN Cars and ABO Tables. † denotes results reported by Functa [12]. § denotes results reported by DiffRF [29]. * denotes results reproduced by us using the official public code with a bugfix. We show ±2σ intervals." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.584, + 0.892, + 0.614 + ], + "angle": 0, + "content": "both datasets, we use the provided renderings (resized to \\(128 \\times 128\\)) with ground truth poses for training and testing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.624, + 0.738, + 0.638 + ], + "angle": 0, + "content": "5.2. Unconditional Generation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.647, + 0.892, + 0.737 + ], + "angle": 0, + "content": "In this section, we conduct evaluations for unconditional generation using the SRN Cars and ABO Tables dataset. The Cars dataset poses a challenge in generating sharp and intricate textures, whereas the Tables dataset comprises of diverse geometries with realistic materials. Models are trained on all images of the training set for 1M iterations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.742, + 0.892, + 0.879 + ], + "angle": 0, + "content": "Evaluation Protocol and Metrics For SRN Cars, following Functa [12], we sample 704 scenes from the diffusion model, and render each scene using the fixed 251 camera poses from the test set. For ABO Tables, following DiffRF [29], we sample 1000 scenes and render each scene with 10 random cameras. We adopt standard generation metrics including Fréchet Inception Distance (FID) [18] and Kernel Inception Distance (KID) [3]. The metrics' reference sets are all images in the test set for SRN Cars and all" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.842, + 0.9 + ], + "angle": 0, + "content": "\\(^{7}\\)https://github.com/nvlabs/eg3d/issues/67" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2420" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.341, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.225, + 0.255, + 0.237 + ], + "angle": 0, + "content": "(a) EG3D Tables" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.089, + 0.612, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.439, + 0.224, + 0.536, + 0.237 + ], + "angle": 0, + "content": "(b) DiffRF Tables" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.089, + 0.891, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.225, + 0.839, + 0.237 + ], + "angle": 0, + "content": "(c) SSDNeRF Tables (ours)" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.245, + 0.466, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.371, + 0.318, + 0.383 + ], + "angle": 0, + "content": "(d) EG3D Cars" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.245, + 0.886, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.37, + 0.766, + 0.383 + ], + "angle": 0, + "content": "(e) SSDNeRF Cars (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.389, + 0.822, + 0.404 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison between unconditional generative models trained on ABO Tables and SRN Cars." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.415, + 0.89, + 0.54 + ], + "angle": 0, + "content": "
MethodCars 1-viewCars 2-viewChairs 1-viewChairs 2-view
PSNR↑SSIM↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓FID↓
3DiM [53]21.010.57-8.99----17.050.53-6.57----
PixelNeRF [55]23.170.900.146‡59.24†25.660.94--23.720.910.128‡38.49†26.200.94--
SRN [45]22.25§0.89§0.129‡41.21†24.84§0.92§--22.89§0.89§0.104‡26.51†24.48§0.92§--
CodeNeRF [21]23.800.910.118*56.34*25.710.930.108*56.13*23.660.900.106*31.65*25.630.910.097*29.90*
VisionNeRF [25]22.880.910.08421.31†----24.480.930.07710.05†----
Ours (1-stage)23.520.910.07816.3926.490.940.05410.6624.350.930.06710.1326.940.950.05510.85
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.545, + 0.893, + 0.6 + ], + "angle": 0, + "content": "Table 2. Single-view and two-view reconstruction results on SRN Cars and Chairs. For consistency with prior work, we use view #64 of the test scene as single-view input and view #64 and #104 as two-view input. † denotes results reported by 3DiM [53]. ‡ denotes results reported by VisionNeRF [25], § denotes results reported by PixelNeRF [21], * denotes results reproduced by us using the official code. - indicates results are unavailable." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.457, + 0.64 + ], + "angle": 0, + "content": "images in the entire dataset for ABO Tables, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.646, + 0.47, + 0.751 + ], + "angle": 0, + "content": "Comparison to the State of the Art As shown in Table 1, on SRN Cars, SSDNeRF (1-stage) outperforms EG3D in KID (a more suitable measure for small datasets) by a clear margin. Meanwhile, its FID is drastically better than Functa, which uses an LDM but with low dimensional latent codes. On ABO Tables, SSDNeRF shows significantly better performance than EG3D and DiffRF." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.758, + 0.47, + 0.835 + ], + "angle": 0, + "content": "Single- vs. Two-stage On SRN Cars, we compare the proposed single-stage training against two-stage training with tuned TV regularization using the same model architecture. The results in Table 1 indicate substantial advantage of single-stage training (KID/10\\(^{-3}\\) 3.47 vs. 6.38)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Qualitative Results As shown in Fig. 4, SSDNeRF generates more regular geometries than the slightly skewed and distorted shapes by EG3D [5]. Compared to DiffRF [29], our method produces sharp details and reflective materials," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.893, + 0.654 + ], + "angle": 0, + "content": "thanks to our more expressive model with latents of higher spatial resolution and view-dependent NeRF decoder." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.661, + 0.804, + 0.678 + ], + "angle": 0, + "content": "5.3. Sparse-View NeRF Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.893, + 0.821 + ], + "angle": 0, + "content": "This section presents experiments on 3D reconstruction from sparse-view images of unseen objects in SRN Cars and Chairs test sets. The Cars dataset presents the challenge of recovering distinct textures, while the Chairs dataset requires accurate reconstruction of diverse shapes. Models are trained on all images of the training set for 80K iterations, as we find that longer schedule leads to decaying performance in reconstructing unseen objects. This behaviour is in accordance with the interpolation results in § 5.5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Evaluation Protocol and Metrics We use the evaluation protocol and metrics in PixelNeRF [55]. Given input images sampled from each test scene, we obtain the triplane scene code via guidance-finetuning and evaluate novel view synthesis quality with respect to the unseen images." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2421" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.087, + 0.487, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.087, + 0.895, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.345, + 0.838, + 0.36 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of single-view reconstruction methods on unseen test objects in SRN Cars and Chairs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.47, + 0.46 + ], + "angle": 0, + "content": "The image quality metrics include average peak signal-to-noise-ratio (PSNR), structural similarity (SSIM) [52], and Learned Perceptual Image Patch Similarity (LPIPS) [56]. In addition, we evaluate the FID between all synthesized images and ground truth images as in 3DiM [53]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.471, + 0.692 + ], + "angle": 0, + "content": "Comparison to the State of the Art Table 2 compares SSDNeRF against previous approaches in single-view and two-view reconstruction settings. Overall, SSDNeRF reaches the best LPIPS of all tasks, indicating the best perceptual fidelity. In contrast, 3DiM generates high quality images (best FID) but with the lowest fidelity to the ground truth (lowest PSNR); CodeNeRF reports the best PSNR on single-view Cars, but its limited expressiveness leads to blurry outputs (Fig. 5) and less competitive LPIPS and FID; VisionNeRF achieves a balanced performance on all single-view metrics, but may struggle to generate textural details on the unseen side of cars (e.g., the other side of the ambulance in Fig. 5). Moreover, SSDNeRF exhibits a clear advantage in two-view reconstruction, achieving the best performance on all relevant metrics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.47, + 0.759 + ], + "angle": 0, + "content": "Single- vs. Two-stage As demonstrated in Table 3, the model trained in a single stage (A0) outperforms the same architecture trained in two stages with TV regularization (A1) in all metrics of single-view reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Ablation Studies on Test-Time Finetuning As shown in Table 3, we evaluate the effectiveness of test-time finetuning and the contribution of the learned diffusion prior with two ablation experiments: (A2) removing the diffusion loss during finetuning and using only the rendering loss, and (A3) omitting the finetuning process entirely. The results indicate that finetuning with single-view rendering loss provides only marginal improvements over guided sampling (A2 vs. A3), while the learned diffusion prior significantly" + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.376, + 0.871, + 0.455 + ], + "angle": 0, + "content": "
IDTrainingFinetuningPSNR↑SSIM↑LPIPS↓FID↓
A01-stageRend + Diff23.520.9130.07816.39
A12-stageRend + Diff22.830.9060.09020.97
A21-stageRend23.130.9070.08827.93
A31-stageNone23.070.9050.09230.95
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.461, + 0.892, + 0.487 + ], + "angle": 0, + "content": "Table 3. Ablation results on single-view reconstruction of SRN Cars." + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.492, + 0.843, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.603, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Figure 6. LPIPS scores (lower is better) of novel view synthesis from sparse-to-dense inputs, evaluated on SRN Cars test set. The triplane baselines adopt mean initialization for better performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.892, + 0.76 + ], + "angle": 0, + "content": "boosts the LPIPS and FID scores (A0 vs. A2), highlighting its importance in recovering sharp and distinct contents. Moreover, the qualitative results in Fig. 5 reveal that views with higher overlap to the input view benefit the most from finetuning, meeting our expectation that finetuning helps faithfully reconstruct the exact observations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Sparse-to-Dense Reconstruction To validate that SSDNeRF seamlessly bridges sparse- and dense-view NeRF reconstruction, we evaluate its novel view synthesis performance with the number of input views varying from 1 to 32. We compare our model to the triplane NeRF baseline trained as an auto-decoder with optional TV regularization instead of diffusion prior. Meanwhile, we also evaluate CodeNeRF [21], an auto-decoder with 256-d latent codes. The results in Fig. 6 show that SSDNeRF excels in all settings," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2422" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "especially in 1 to 4 views. In contrast, CodeNeRF is outperformed by vanilla triplane NeRF with more views." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.13, + 0.45, + 0.146 + ], + "angle": 0, + "content": "5.4. Training SSDNeRF on Sparse-View Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.154, + 0.47, + 0.244 + ], + "angle": 0, + "content": "In this section, we train SSDNeRF on a sparse-view subset of the full SRN Cars training set, in which a fixed set of only three views are randomly picked from each scene. Note that a reasonable decline in performance compared to dense-view training is expected as the whole training dataset has been reduced to \\(6\\%\\) of its original size." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.249, + 0.469, + 0.354 + ], + "angle": 0, + "content": "Unconditional Generation We adopt a training trick that resets the triplane codes to their mean value halfway through training. This helps to prevent the model from getting stuck in a local minimum that overfits geometric artifacts. We also double the length of the training schedule accordingly. The model achieves a decent FID of \\(19.04 \\pm 1.10\\) and a \\(\\mathrm{KID} / 10^{-3}\\) of \\(8.28 \\pm 0.60\\). Results are visualized in Fig. 7." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.469, + 0.435 + ], + "angle": 0, + "content": "Single-View Reconstruction We adopt the same training strategy as in § 5.3. With our guidance-finetuning approach, the model achieves an LPIPS score of 0.106, even outperforming most of the previous methods in Table 2 that use the full training set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.56 + ], + "angle": 0, + "content": "Comparison to TV Regularization Fig. 8 (b) shows the RGB images and geometries represented by the scene latent codes learned from three views during training. By comparison, vanilla triplane auto-decoder with TV regularization (Fig. 8 (a)) often fails to reconstruct a scene from sparse views, leading to severe geometric artifacts. As a result, previously it has been infeasible to train two-stage models with expressive latents on sparse-view data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.568, + 0.265, + 0.584 + ], + "angle": 0, + "content": "5.5. NeRF Interpolation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.469, + 0.788 + ], + "angle": 0, + "content": "Following DDIM [47], we can sample two initial values \\( x^{(T)} \\sim \\mathcal{N}(0, I) \\), interpolate them using spherical linear interpolation [43], and then use the deterministic solver to obtain interpolated samples. However, as noted by [34, 37], standard Gaussian diffusion models often result in nonsmooth interpolation. In SSDNeRF (with results shown in Fig. 9), we find that the model (a) trained with early stopping for sparse-view reconstruction produces reasonably smooth transitions between samples, while the model (b) trained with a longer schedule for unconditional generation produces distinct yet discontinuous samples. This suggests that early stopping preserves a smoother prior, leading to better generalization for sparse-view reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.196, + 0.815 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.9 + ], + "angle": 0, + "content": "In this paper, we propose SSDNeRF, which combines the diffusion model and NeRF representation through a novel single-stage training paradigm with an end-to-end justifiable loss. Notably, it overcomes the limitations in previous work where implicit neural fields must be obtained from" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.092, + 0.885, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.153, + 0.892, + 0.181 + ], + "angle": 0, + "content": "Figure 7. Images generated by SSDNeRF trained on a 3-view subset of SRN Cars training set." + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.196, + 0.681, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.27, + 0.674, + 0.282 + ], + "angle": 0, + "content": "(a) Auto-decoder training w/ TV" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.206, + 0.884, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.703, + 0.27, + 0.891, + 0.282 + ], + "angle": 0, + "content": "(b) 1-stage joint training with diffusion" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.289, + 0.892, + 0.33 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison between scene codes learned from 3 views by (a) triplane auto-decoder with TV regularization vs. (b) single-stage diffusion NeRF." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.343, + 0.89, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.397, + 0.886, + 0.409 + ], + "angle": 0, + "content": "(a) Models trained with early stopping (80K iters) for sparse-view reconstruction" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.417, + 0.888, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.47, + 0.88, + 0.481 + ], + "angle": 0, + "content": "(b) Models trained with long schedule (1M iters) for unconditional generation" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.488, + 0.891, + 0.515 + ], + "angle": 0, + "content": "Figure 9. Interpolation between the leftmost and rightmost samples using DDIM [47]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.892, + 0.617 + ], + "angle": 0, + "content": "dense observations first, before training the diffusion models to learn their manifold. With strong performance on multiple benchmarks, SSDNeRF demonstrates a significant advancement towards a unified framework for general 3D content manipulation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.892, + 0.744 + ], + "angle": 0, + "content": "Limitations and Future Work Currently, our method relies on ground truth camera parameters during both training and testing. Future work may explore transform-invariant models. Additionally, the diffusion prior can become discontinuous with prolonged training, which affects generalization. Although early stopping is temporarily used, a better network design or a larger training dataset may be able to address this problem fundamentally." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements We thank Norman Müller for sharing the baseline results on ABO Tables. Hansheng Chen and Wei Tian acknowledge the funding by the National Natural Science Foundation of China (No. 52002285), the Shanghai Science and Technology Commission (No. 21ZR1467400), the original research project of Tongji University (No. 22120220593), the National Key R&D Program of China (No. 2021YFB2501104), and the Natural Science Foundation of Chongqing (No. 2023NSCQ-MSX4511)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2423" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J. Mitra, and Paul Guerrero. RenderDiffusion: Image diffusion for 3D reconstruction, inpainting and generation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, Afshin Dehghan, and Josh Susskind. Gaudi: A neural architect for immersive 3d scene generation. In NeurIPS, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.285 + ], + "angle": 0, + "content": "[3] Mikołaj Binkowski, Danica J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In ICLR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[4] Eric Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.47, + 0.415 + ], + "angle": 0, + "content": "[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.417, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[6] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.47, + 0.557 + ], + "angle": 0, + "content": "[7] Anpei Chen, Zexiang Xu, Xinyue Wei, Siyu Tang, Hao Su, and Andreas Geiger. Factor fields: A unified framework for neural fields and beyond, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.47, + 0.614 + ], + "angle": 0, + "content": "[8] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, pages 14124-14133, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[9] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik. Abo: Dataset and benchmarks for real-world 3d object understanding. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[10] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[11] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W. Taylor, and Joshua M. Susskind. Unconstrained scene generation with locally conditioned radiance fields. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[12] Emilien Dupont, Hyunjik Kim, S. M. Ali Eslami, Danilo Jimenez Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you can treat it like one. In ICML, 2022. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[13] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In NeurIPS, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[15] Jiatao Gu, Qingzhe Gao, Shuangfei Zhai, Baoquan Chen, Lingjie Liu, and Josh Susskind. Learning controllable 3d diffusion models from single-view images, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[16] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In ICLR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[17] Jiatao Gu, Alex Trevithick, Kai-En Lin, Josh Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nerf-guided distillation from 3d-aware diffusion, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[21] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In ICCV, pages 12949-12958, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[22] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[24] Yiyi Liao, Katja Schwarz, Lars Mescheder, and Andreas Geiger. Towards unsupervised learning of generative models for 3d controllable image synthesis. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[25] Kai-En Lin, Lin Yen-Chen, Wei-Sheng Lai, Tsung-Yi Lin, Yi-Chang Shih, and Ravi Ramamoorthi. Vision transformer for nef-based view synthesis from a single input image. In WACV, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.703 + ], + "angle": 0, + "content": "[26] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11461-11471, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[27] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[29] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In CVPR, 2023. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[30] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2424" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "learning of 3d representations from natural images. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.164 + ], + "angle": 0, + "content": "[31] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[32] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.223, + 0.469, + 0.263 + ], + "angle": 0, + "content": "[33] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.266, + 0.469, + 0.32 + ], + "angle": 0, + "content": "[34] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.324, + 0.469, + 0.365 + ], + "angle": 0, + "content": "[35] Daniel Rebain, Mark Matthews, Kwang Moo Yi, Dmitry Lagun, and Andrea Tagliasacchi. Lolnerf: Learn from one look. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.368, + 0.469, + 0.422 + ], + "angle": 0, + "content": "[36] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In ICML, pages 1278-1286, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.424, + 0.469, + 0.464 + ], + "angle": 0, + "content": "[37] Severi Rissanen, Markus Heinonen, and Arno Solin. Generative modelling with inverse heat dissipation. In ICLR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.468, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.567 + ], + "angle": 0, + "content": "[39] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.569, + 0.469, + 0.597 + ], + "angle": 0, + "content": "[40] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In ICLR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.469, + 0.64 + ], + "angle": 0, + "content": "[41] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.642, + 0.469, + 0.684 + ], + "angle": 0, + "content": "[42] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.469, + 0.728 + ], + "angle": 0, + "content": "[43] Ken Shoemake. Animating rotation with quaternion curves. In Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 245-254, 1985. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.729, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[44] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In CVPR, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.469, + 0.827 + ], + "angle": 0, + "content": "[45] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In NeurIPS, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[46] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[47] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2, 3, 4, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[48] Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. Maximum likelihood training of score-based diffusion models. In NeurIPS, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[49] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[50] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In NeurlPS, 2021. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[51] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[52] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[53] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. In ICLR, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[54] Antoine Wehenkel and Gilles Louppe. Diffusion priors in variational autoencoders. In ICML Workshops, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.435, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[55] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[56] Richard Zhang, Phillip Isola, Alexei Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.52, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[57] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In CVPR, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[58] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In CVPR, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.603 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2425" + } + ] +] \ No newline at end of file diff --git a/2023/Single-Stage Diffusion NeRF_ A Unified Approach to 3D Generation and Reconstruction/06f68dbb-5c23-4704-ae55-c667675f6d9f_origin.pdf b/2023/Single-Stage Diffusion NeRF_ A Unified Approach to 3D Generation and Reconstruction/06f68dbb-5c23-4704-ae55-c667675f6d9f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e10c067a5b012b7f15a4df241f8a0b730b37e41c --- /dev/null +++ b/2023/Single-Stage Diffusion NeRF_ A Unified Approach to 3D Generation and Reconstruction/06f68dbb-5c23-4704-ae55-c667675f6d9f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:defe51f6d8bb2cd780a2e0330317e3894b0e3aa153d8e71e7acfca9c956b0557 +size 9921923 diff --git a/2023/Single-Stage Diffusion NeRF_ A Unified Approach to 3D Generation and Reconstruction/full.md b/2023/Single-Stage Diffusion NeRF_ A Unified Approach to 3D Generation and Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..464e4d9f9f07717b04f0c4d4198a6c1f4d421178 --- /dev/null +++ b/2023/Single-Stage Diffusion NeRF_ A Unified Approach to 3D Generation and Reconstruction/full.md @@ -0,0 +1,349 @@ +# Single-Stage Diffusion NeRF: A Unified Approach to 3D Generation and Reconstruction + +Hansheng Chen, $^{1,*}$ Jiatao Gu, $^{2}$ Anpei Chen, $^{3}$ Wei Tian, $^{1}$ Zhuowen Tu, $^{4}$ Lingjie Liu, $^{5}$ Hao Su $^{4}$ + +$^{1}$ Tongji University $^{2}$ Apple $^{3}$ ETH Zürich $^{4}$ University of California, San Diego $^{5}$ University of Pennsylvania + +![](images/9d1343f1e853acd282882509fbc88fdafe21842eaed4dbf19b49f503e7bd72eb.jpg) +Figure 1. During training, SSDNeRF jointly learns triplane features of individual scenes, a shared NeRF decoder, and a triplane diffusion prior. During testing, it can perform (a) unconditional generation, (b) single-view reconstruction, as well as multi-view reconstruction. + +![](images/f339408477b3f3c7291875e05a1e747a03ef1055091c536ce453bb7f528463d0.jpg) + +# Abstract + +3D-aware image synthesis encompasses a variety of tasks, such as scene generation and novel view synthesis from images. Despite numerous task-specific methods, developing a comprehensive model remains challenging. In this paper, we present SSDNeRF, a unified approach that employs an expressive diffusion model to learn a generalizable prior of neural radiance fields (NeRF) from multi-view images of diverse objects. Previous studies have used two-stage approaches that rely on pretrained NeRFs as real data to train diffusion models. In contrast, we propose a new single-stage training paradigm with an end-to-end objective that jointly optimizes a NeRF auto-decoder and a latent diffusion model, enabling simultaneous 3D reconstruction and prior learning, even from sparsely available views. At test time, we can directly sample the diffusion prior for unconditional generation, or combine it with arbitrary observations of unseen objects for NeRF reconstruction. SSDNeRF demonstrates robust results comparable to or better than leading task-specific methods in unconditional generation and single/sparse-view 3D reconstruction. $^6$ + +# 1. Introduction + +Synthesizing 3D visual contents has gained significant attention in computer vision and graphics, thanks to advances in neural rendering and generative models. Although numerous methods have emerged to handle individual tasks, such as single-/multi-view 3D reconstruction and 3D content generation, it remains a major challenge to develop a comprehensive framework that bridges the state of the art of multiple tasks. For instance, neural radiance fields (NeRF) [28] have shown impressive results in novel view synthesis by solving the inverse rendering problem via perscene fitting, which is suitable for dense-view inputs but difficult to generalize to sparse observations. In contrast, many sparse-view 3D reconstruction methods [55, 8, 25] rely on feed-forward image-to-3D encoders, but they are unable to handle ambiguity in the occluded region and generate crisp images. Regarding unconditional generation, 3D-aware generative adversarial networks (GAN) [31, 5, 16, 13] are partially limited in their usage of single-image discriminators, which cannot reason cross-view relationships to effectively learn from multi-view data. + +In this paper, we propose a unified approach to various 3D tasks (Fig. 1) by developing a holistic model that learns generalizable 3D priors from multi-view images. Inspired + +by the success of 2D diffusion models [19, 47, 27, 38, 26], we present the Single-Stage Diffusion NeRF (SSDNeRF), which models the generative prior of scene latent codes with a 3D latent diffusion model (LDM). + +While similar LDMs have been applied in 2D and 3D generation in previous work [50, 38, 12, 2, 44, 29], they typically require two-stage training, where the first stage pretrains the variational auto-encoders (VAE) [23] or autodecoders [32] without diffusion models. In the case of diffusion NeRFs, however, we argue that two-stage training induces noisy patterns and artifacts in the latent code due to the uncertain nature of inverse rendering, particularly when training from sparse-view data, which prevents the diffusion model from learning a clean latent manifold effectively. To address this issue, we introduce a novel single-stage training paradigm that enables end-to-end learning of diffusion and NeRF weights ( $\S 4.1$ ). This approach blends the generative and the rendering biases coherently for improved performance overall and allows for training on sparse-view data. Additionally, we show that the learned 3D priors of unconditional diffusion models can be exploited for flexible test-time scene sampling from arbitrary observations ( $\S 4.2$ ). + +We evaluate SSDNeRF on multiple datasets of categorical single-object scenes, demonstrating strong performance overall. Our approach represents a significant step towards a unified framework for various 3D tasks. + +To summarize, our main contributions are as follows: + +- We introduce SSDNeRF, a unified approach to all-round performance in unconditional 3D generation and image-based reconstruction; +- We propose a novel single-stage training paradigm that jointly learns NeRF reconstruction and diffusion model from multi-view images of a large number of objects. Notably, this enables training on as sparse as three views per scene, which is previously infeasible; +- A guidance-finetuning sampling scheme is developed to exploit the learned diffusion priors for 3D reconstruction from arbitrary number of views at test time. + +# 2. Related Work + +3D GANs The generative adversarial framework [14] has been successfully adapted for 3D generation by integrating projection-based rendering into the generator. A variety of 3D representations have been explored previously, including point clouds, cuboids, spheres [24] and voxels [30] in early works, the more recent radiance fields [4, 41, 11, 42, 46] and feature fields [31, 16, 5] with volume renderer, and differentiable surface [13] with mesh renderer. The above methods are all trained with 2D image discriminators that are unable to reason cross-view relationships, making them heavily dependent on model bias for 3D consistency. As a result, multi-view data cannot be effectively exploited to + +learn complex and diverse geometries. 3D GANs are mostly applied in unconditional generation. Although 3D completion from images is possible through GAN inversion [11], faithfulness is not guaranteed due to limited latent expressiveness, as shown in [29, 1]. + +View-Conditioned Regression and Generation Sparse-view 3D reconstruction can be tackled by regressing novel views from input images. Various architectures [8, 55, 25, 57] have been proposed to encode images into volume features, which can be projected to supervised target views through volume rendering. However, they cannot reason ambiguity and generate diverse and meaningful contents, which often leads to blurry results. In contrast, image-conditioned generative models are better at synthesizing distinct contents. 3DiM [53] proposes to generate novel views from a view-conditioned image diffusion model, but the model lacks 3D consistency bias. [58, 10, 17] distill priors of image-conditioned 2D diffusion models into NeRFs to enforce 3D constraints. These methods are parallel to our track as they model cross-view relationships in the image space, while our model is inherently 3D. + +Auto-Decoders and Diffusion NeRF NeRF's per-scene fitting scheme can be generalized to multi-scene fitting by sharing part of the parameters across all scenes, leaving the rest as individual scene codes [7]. Therefore, multi-scene NeRFs can be trained as auto-decoders [32], where the code bank and shared decoder weights are jointly learned. With proper architectures, scene codes can be treated as latents with Gaussian priors, allowing 3D completion and even generation [21, 45, 35]. However, like 3D GANs, the latents are not expressive enough for faithful reconstruction of detailed objects. [2, 12, 51] improve upon vanilla auto-decoders with latent diffusion priors. DiffRF [29] leverages the diffusion prior to perform 3D completion. These methods train the auto-decoders and diffusion models in two separate stages, which is subject to the limitations in § 3.2. + +# 3. Background + +# 3.1. NeRF as an Auto-Decoder + +Given a set of 2D images of a scene and their camera parameters, one can fit a scene model to reconstruct the light field in 3D space, expressed by a plenoptic function $y_{\psi}(r)$ , where $r$ parameterizes the endpoint and direction of a ray in the world space, $\psi$ denotes the scene model parameters, and $y \in \mathbb{R}_+^3$ represents the received light in RGB format. NeRF [28] represents the light field as integrated radiance along rays through the 3D volume. It models the scene geometry and appearance as functions of the position $p \in \mathbb{R}^3$ and view direction $d \in \mathbb{R}^3$ of a point in the world space, expressed as $\rho_{\psi}(p)$ and $c_{\psi}(p,d)$ respectively, where $\rho \in \mathbb{R}_{+}$ is the density output and $c \in \mathbb{R}_{+}^{3}$ is the RGB color output. Differentiable volume rendering is applied to compose the + +received light $y$ from multiple point samples along a ray $r$ . + +NeRF can also generalize to multi-scene settings by sharing part of the model parameters across all scenes [7]. Given observations of multiples scenes $\{y_{ij}^{\mathrm{gt}},r_{ij}^{\mathrm{gt}}\}$ , where $y_{ij}^{\mathrm{gt}},r_{ij}^{\mathrm{gt}}$ is the $j$ -th pair of pixel RGB and ray of the $i$ -th scene, one can optimize the per-scene codes $\{x_i\}$ and shared parameters $\psi$ by minimizing the L2 rendering loss: + +$$ +\mathcal {L} _ {\text {r e n d}} \left(\left\{x _ {i} \right\}, \psi\right) = \mathbb {E} _ {i} \left[ \sum_ {j} \frac {1}{2} \left\| y _ {i j} ^ {\mathrm {g t}} - y _ {\psi} \left(x _ {i}, r _ {i j} ^ {\mathrm {g t}}\right) \right\| ^ {2} \right]. \tag {1} +$$ + +With this objective, the model is trained as an auto-decoder [32], where the scene codes $\{x_{i}\}$ can be interpreted as the latent codes, and the plenoptic function can be regarded as a decoder in the form of $p_{\psi}(\{y_j\} |x,\{r_j\})\coloneqq \prod_j\mathcal{N}(y_j|y_\psi (x,r_j),I)$ , assuming independent Gaussians. + +Challenges in Bridging Generation and Reconstruction An auto-decoder with trained weights $\psi$ can perform unconditional generation by decoding latent codes drawn from a Gaussian prior [35]. However, to ensure continuity in generation, a low-dimensional latent space and a complex decoder is required, which adds to the difficulty in optimizing the latent code to faithfully reconstruct any given views. + +# 3.2. Latent Diffusion Models + +Latent diffusion models (LDM) learn a prior distribution $p_{\phi}(x)$ in the latent space with parameters $\phi$ , which enables the usage of more expressive latent representations, such as 2D grids for images [50, 38]. For neural field generation, previous work [2, 29, 12, 44] adopts a two-stage training scheme, where the auto-decoder is trained first to obtain the per-scene latent $x_{i}$ , which is then treated as real data to train the LDM. The LDM injects Gaussian perturbation $\epsilon \sim \mathcal{N}(0, I)$ into the code $x_{i}$ , yielding a noisy code $x_{i}^{(t)} \coloneqq \alpha^{(t)} x_{i} + \sigma^{(t)} \epsilon$ at diffusion time step $t$ , under empirical noise schedule functions $\alpha^{(t)}, \sigma^{(t)}$ . A denoising network with trainable weights $\phi$ is then tasked with removing the noise from $x_{i}^{(t)}$ to predict a denoised code $\hat{x}_{i}$ . The network is typically trained with a simplified L2 denoising loss: + +$$ +\mathcal {L} _ {\mathrm {d i f f}} (\phi) = \underset {i, t, \epsilon} {\mathbb {E}} \left[ \frac {1}{2} w ^ {(t)} \left\| \hat {x} _ {\phi} \left(x _ {i} ^ {(t)}, t\right) - x _ {i} \right\| ^ {2} \right], \tag {2} +$$ + +where $t \sim \mathcal{U}(0,T)$ , $w^{(t)}$ is an empirical time dependent weighting function, and $\hat{x}_{\phi}(x_i^{(t)},t)$ formulates the time-conditioned denoising network. + +Unconditional/Guided Sampling With trained weights $\phi$ , one can sample from the diffusion prior $p_{\psi}(x)$ using a variety of solvers (e.g., DDIM [47]) that recursively denoise $x^{(t)}$ , starting from random Gaussian noise $x^{(T)}$ , until reaching the denoised state $x^{(0)}$ . Moreover, the sampling process can be guided by the gradients of the rendering loss against known observations, allowing 3D reconstruction from images at test time [29]. + +![](images/c311386e4d3a3a5f7956006093d8918813e36ffac7196bd2bfacf7dd1dcd8568.jpg) +Figure 2. Visualization of the scene code $x_{\mathrm{XZ}}$ at XZ plane. Left column: Two-stage training without TV regularization induces noise and fails in 3-view reconstruction. Mid column: TV regularization imposes smoothing prior at the cost of textural details (top), yet still struggles to cope with sparse views (bottom). Right column: Our single-stage training encourages smooth yet detailed latents and allows for training with sparse views. + +Limitations of Two-Stage Training for 3D Tasks While LDMs with 2D image VAEs are typically trained in two stages [50, 38], training LDMs with NeRF auto-decoders poses an unprecedented challenge. An expressive latent code is underdetermined when obtained via rendering-based optimization, leading to noisy patterns that distract denoising networks (top-left of Fig. 2). Additionally, reconstructing NeRFs from sparse views without a learned prior is exceptionally difficult (bottom-left of Fig. 2), limiting training to dense-views settings. + +# 4. Proposed Method + +To build a holistic model that unifies 3D generation and reconstruction, we propose SSDNeRF, a framework that conjoins the expressive triplane NeRF auto-decoder with a triplane latent diffusion model. Fig. 3 provides an overview of the model. In the following subsections, we elaborate on how training and testing are performed in detail. + +# 4.1. Single-Stage Diffusion NeRF Training + +An auto-decoder can be regarded as a type of VAE that uses a lookup table encoder instead of the typical neural network encoder. As such, the training objective can be derived in a similar way as for VAEs. With NeRF decoder $p_{\psi}(\{y_j\} |x,\{r_j\})$ and diffusion latent prior $p_{\phi}(x)$ , the training objective is to minimize variational upper bound on the negative log-likelihood (NLL) of observed data $\{y_{ij}^{\mathrm{gt}}, r_{ij}^{\mathrm{gt}}\}$ [23, 36, 50]. In this paper, a simplified training loss is derived by ignoring the uncertainty (variance) in latent codes: + +$$ +\mathcal {L} = \underbrace {\underset {i} {\mathbb {E}} \left[ - \log p _ {\psi} \left(\left\{y _ {i j} ^ {\mathrm {g t}} \right\} \mid x _ {i} , \left\{r _ {i j} ^ {\mathrm {g t}} \right\}\right) \right]} _ {\text {r e n d e r i n g l o s s} \mathcal {L} _ {\text {r e n d}}} + \underbrace {\underset {i} {\mathbb {E}} \left[ - \log p _ {\phi} \left(x _ {i}\right) \right]} _ {\text {p r i o r t e r m}}, \tag {3} +$$ + +where the scene codes $\{x_{i}\}$ , prior parameters $\phi$ , and decoder parameters $\psi$ are jointly optimized in a single training stage. This loss function consists of the rendering loss $\mathcal{L}_{\mathrm{rend}}$ + +![](images/8e145baf531a412b27e037624d3792ff4f3700cd97c4869b6aefbcae9804a2c1.jpg) +Figure 3. An overview of SSDNeRF framework with a triplane NeRF representation. During training, we feed a batch of observations in the format of RGB values $y_{ij}^{\mathrm{gt}}$ and rays $r_{ij}^{\mathrm{gt}}$ . The corresponding scene code $x_{i}$ is randomly initialized and optimized by minimizing the rendering loss $\mathcal{L}_{\mathrm{rend}}$ and the diffusion loss $\mathcal{L}_{\mathrm{diff}}$ , and model parameters $\phi, \psi$ are also updated along the way. + +in Eq. (1) and a diffusion prior term in the form of NLL. Following [50, 54, 48], we replace the diffusion NLL with its approximate upper bound $\mathcal{L}_{\mathrm{diff}}$ in Eq. (2). This technique is also called score distillation in [33]. Adding empirical weighting factors, we finalize our training objective as: + +$$ +\mathcal {L} = \lambda_ {\text {r e n d}} \mathcal {L} _ {\text {r e n d}} (\{x _ {i} \}, \psi) + \lambda_ {\text {d i f f}} \mathcal {L} _ {\text {d i f f}} (\{x _ {i} \}, \phi). \tag {4} +$$ + +Single-stage training constrains scene codes $\{x_{i}\}$ with both terms in the loss function, allowing the learned prior to complete the parts unseen to rendering. This is particularly beneficial to training on sparse-view data, where the expressive triplane codes are severely underdetermined. + +Balancing Rendering and Prior Weights The render-to-prior weight ratio $\lambda_{\mathrm{rend}} / \lambda_{\mathrm{diff}}$ is crucial to single-stage training. To make hyperparameters more generalizable to different settings, we design an empirical weighting mechanism, in which the diffusion loss is normalized by the exponential moving average (EMA) of the scene codes' Frobenius norms, expressed as $\lambda_{\mathrm{diff}} \coloneqq c_{\mathrm{diff}} / EMA\left(\|x_i\|_F^2\right)$ with a constant scale $c_{\mathrm{diff}}$ , and the rendering weight is determined by the number of views available $N_{\mathrm{v}}$ , expressed as $\lambda_{\mathrm{rend}} \coloneqq c_{\mathrm{rend}}(1 - e^{-0.1N_{\mathrm{v}}}) / N_{\mathrm{v}}$ with a constant scale $c_{\mathrm{rend}}$ . Intuitively, $N_{\mathrm{v}}$ -based weighting is a calibration to the ray independence assumption in the decoder $p_{\psi}(\{y_j\} | x, \{r_j\}) \coloneqq \prod_j \mathcal{N}(y_j | y_{\psi}(x, r_j), I)$ , preventing the rendering loss from scaling linearly with the number of rays. + +Comparison to Two-Stage Generative Neural Fields Previous two-stage methods [2, 12, 29, 44] ignore the prior term $\lambda_{\mathrm{diff}}\mathcal{L}_{\mathrm{diff}}$ during the first stage of training the auto-decoders. This can be seen as setting the render-to-prior weight ratio $\lambda_{\mathrm{rend}} / \lambda_{\mathrm{diff}}$ to infinity, resulting in biased and noisy scene codes $x_{i}$ . Shue et al. [44] partially mitigate this issue by imposing total variation (TV) regularization on triplane scene codes to enforce a smoothing prior, which resembles the LDM constraints on the latent space (mid column of Fig. 2). Control3Diff [15] proposes to learn a conditional diffusion model on data generated by a 3D GAN pre + +trained on single-view images. In contrast, our single-stage training aims to directly incorporate the diffusion prior to promote end-to-end coherence. + +# 4.2. Image-Guided Sampling and Finetuning + +To achieve generalizable test-time NeRF reconstruction that covers a wide spectrum from single-view to dense observations, we propose performing image-guided sampling and then finetuning the sampled codes considering both the diffusion prior and rendering likelihood. + +Following the reconstruction-guided sampling method by Ho et al. [20], we compute the approximated rendering gradients $g$ w.r.t. a noisy code $x^{(t)}$ , defined as: + +$$ +g \leftarrow \nabla_ {x ^ {(t)}} \lambda_ {\text {r e n d}} \sum_ {j} \frac {1}{2} \left(\frac {\alpha^ {(t)}}{\sigma^ {(t)}}\right) ^ {2 \omega} \left\| y _ {j} ^ {\mathrm {g t}} - y _ {\psi} \left(\hat {x} _ {\phi} \left(x ^ {(t)}, t\right), r _ {j} ^ {\mathrm {g t}}\right) \right\| ^ {2}, \tag {5} +$$ + +where $\left(\alpha^{(t)} / \sigma^{(t)}\right)^{2\omega}$ is an additional weighting factor based on signal-to-noise ratio (SNR), with hyperparameter $\omega$ chosen to be 0.5 or 0.25 in our work. The guidance gradients $g$ are then combined with unconditional score prediction, expressed as a correction to the denoising output $\hat{x}$ : + +$$ +\hat {x} \leftarrow \hat {x} - \lambda_ {\mathrm {g d}} \frac {\sigma^ {(t) ^ {2}}}{\alpha^ {(t)}} g \tag {6} +$$ + +with guidance scale $\lambda_{\mathrm{gd}}$ . We adopt the predictor-corrector sampler [49] to solve $x^{(0)}$ by alternating between a DDIM step [47] and multiple Langevin correction steps. + +We observe that the reconstruction guidance alone cannot strictly enforce rendering constraints towards faithful reconstruction. To address this issue, we reuse the training objective in Eq. (4) to finetune the sampled scene code $x$ , while freezing the diffusion and decoder parameters: + +$$ +\min _ {x} \lambda_ {\text {r e n d}} \mathcal {L} _ {\text {r e n d}} (x) + \lambda_ {\text {d i f f}} ^ {\prime} \mathcal {L} _ {\text {d i f f}} (x), \tag {7} +$$ + +where $\lambda_{\mathrm{diff}}^{\prime}$ is the test-time prior weight, which we find should be lower than the training weight $\lambda_{\mathrm{diff}}$ for best results, as the prior learned from the training dataset is less + +reliable when transferred to a different testing dataset. We use Adam [22] to optimize the code $x$ for finetuning. + +Comparison to Previous NeRF Finetuning Approaches While finetuning with rendering loss is common in view-conditioned NeRF regression methods [8, 57], our finetuning approach differs in the use of diffusion prior loss on the 3D scene code, which significantly enhances generalization to novel views, as demonstrated in § 5.3. + +# 4.3. Implementation Details + +This subsection briefly describes some important technical details. More details can be found in the supplementary. + +Prior Gradient Caching Triplane NeRF reconstruction requires at least hundreds of optimization iterations on each scene code $x_{i}$ . A problem with the single-stage training loss in Eq. (4) is that the diffusion loss $\mathcal{L}_{\mathrm{diff}}$ requires much longer time to evaluate than the native NeRF rendering loss $\mathcal{L}_{\mathrm{rend}}$ , reducing overall efficiency. To accelerate reconstruction in both training and test-time finetuning, we introduce a technique called prior gradient caching, which caches the back-propagated prior gradients $\nabla_{x}\lambda_{\mathrm{diff}}\mathcal{L}_{\mathrm{diff}}$ for re-use in multiple Adam steps, while refreshing the rendering gradients $\nabla_{x}\lambda_{\mathrm{rend}}\mathcal{L}_{\mathrm{rend}}$ in each of the steps, which allows for fewer diffusion passes than rendering. A training pseudocode is given in Algorithm 1. + +Denoising Parameterization and Weighting The denoising model $\hat{x}_{\phi}(x^{(t)}, t)$ is implemented as a U-Net [39] as in DDPM [19], with a total of 122M parameters. Its input and output are noisy and denoised triplane features, respectively, with channels of all three planes stacked together. For the prediction format, we adopt the $v$ -parameterization $\hat{v}_{\phi}(x^{(t)}, t)$ in [40], such that $\hat{x} = \alpha^{(t)} x^{(t)} - \sigma^{(t)} \hat{v}$ . Regarding the weighting function $w^{(t)}$ in the diffusion loss in Eq. (2), LSGM [50] employs two different mechanisms for optimizing latents $x_{i}$ and diffusion weights $\phi$ , respectively, which we find unstable with NeRF auto-decoders. Instead, we observe that the SNR-based weighting $w^{(t)} = \left(\alpha^{(t)} / \sigma^{(t)}\right)^{2\omega}$ used in Eq. (5) works well with our models. + +# 5. Experiments + +# 5.1. Datasets + +We conduct experiments on the ShapeNet SRN [6, 45] and Amazon Berkeley Objects (ABO) Tables [9] datasets for benchmarking with previous work. The SRN dataset provides single-object scenes in two categories, i.e., Cars and Chairs, with a train/test split of 2458/703 for Cars and 4612/1317 for Chairs. Each train scene has 50 random views from a sphere and each test scene has 251 spiral views from the upper hemisphere. The ABO Tables dataset provides a train/test split of 1520/156 table scenes, where each scene has 91 views from the upper hemisphere. For + +Algorithm 1: Single-stage diffusion NeRF training +Input: $\{y_{ij}^{\mathrm{gt}},r_{ij}^{\mathrm{gt}}\}$ +1 Initialize $\{x_i\} ,\phi ,\psi$ +2 for $k_{\mathrm{out}}\coloneqq 1\dots K_{\mathrm{out}}$ do // outer loop of $K_{\mathrm{out}}$ iterations +3 Sample a batch of scenes $i\in B_{\mathrm{sc}}$ +4 $g_{\phi},g_{x}^{\mathrm{diff}}\gets \nabla_{\phi ,\{x_{i}\}_{B_{\mathrm{sc}}}}\lambda_{\mathrm{diff}}\mathcal{L}_{\mathrm{diff}}$ // diffusion grad +5 $\phi \leftarrow \phi -Adam(g_{\phi})$ +6 for $k_{\mathrm{in}}\coloneqq 1\dots K_{\mathrm{in}}$ do // inner loop of $K_{\mathrm{in}}$ iterations +7 Sample a batch of rays $j\in B_{\mathrm{ray}}$ +8 $g_{x}^{\mathrm{rend}}\gets \nabla_{\{x_i\}_{B_{\mathrm{sc}}}}\lambda_{\mathrm{rend}}\mathcal{L}_{\mathrm{rend}}$ // rendering grad +9 $g_{x}\gets g_{x}^{\mathrm{rend}} + g_{x}^{\mathrm{diff}}$ // add cached prior grad +10 $\{x_i\}_{B_{\mathrm{sc}}} \leftarrow \{x_i\}_{B_{\mathrm{sc}}} - Adam(g_x)$ +11 if $k_{\mathrm{in}} = K_{\mathrm{in}}$ then // last inner iteration +12 $\begin{array}{rl} & g_{\psi}\gets \nabla_{\psi}\lambda_{\mathrm{rend}}\mathcal{L}_{\mathrm{rend}}\\ & \psi \leftarrow \psi -Adam(g_{\psi}) \end{array}$ +13 + +
MethodTypeCarsTables
FID↓KID/10-3↓FID↓KID/10-3↓
Functa [12]LDM80.3---
π-GAN [4]GAN36.7†-41.67§13.82§
EG3D [5]GAN10.46*4.90*31.18§11.67§
DiffRF [29]LDM--27.0610.03
Ours (2-stage)LDM16.33±0.936.38±0.41--
Ours (1-stage)LDM11.08±1.113.47±0.2314.27±0.664.08±0.33
+ +Table 1. Unconditional generation results on SRN Cars and ABO Tables. † denotes results reported by Functa [12]. § denotes results reported by DiffRF [29]. * denotes results reproduced by us using the official public code with a bugfix. We show ±2σ intervals. + +both datasets, we use the provided renderings (resized to $128 \times 128$ ) with ground truth poses for training and testing. + +# 5.2. Unconditional Generation + +In this section, we conduct evaluations for unconditional generation using the SRN Cars and ABO Tables dataset. The Cars dataset poses a challenge in generating sharp and intricate textures, whereas the Tables dataset comprises of diverse geometries with realistic materials. Models are trained on all images of the training set for 1M iterations. + +Evaluation Protocol and Metrics For SRN Cars, following Functa [12], we sample 704 scenes from the diffusion model, and render each scene using the fixed 251 camera poses from the test set. For ABO Tables, following DiffRF [29], we sample 1000 scenes and render each scene with 10 random cameras. We adopt standard generation metrics including Fréchet Inception Distance (FID) [18] and Kernel Inception Distance (KID) [3]. The metrics' reference sets are all images in the test set for SRN Cars and all + +![](images/2edc5250fd0c2cbf1e22213f5c01dceeb54f2123df734d67bc9fd201d6e45a25.jpg) +(a) EG3D Tables + +![](images/c402b9128283d431983729c792e80d3f6820e86608bc750ac552a63ab4539da9.jpg) +(b) DiffRF Tables + +![](images/c784415f9ed0e3707e9c95fdd9dc45973dc1943923630e2974d84fbcedd6f1e4.jpg) +(c) SSDNeRF Tables (ours) + +![](images/3722d95486aec137dd0b0e2ee8cc2a7dd306816752499b3d40b309e9eb27857f.jpg) +(d) EG3D Cars + +![](images/abd7f77808ff61a642e84969ed5f20c754a8b5922571cd827a8866b62685f627.jpg) +(e) SSDNeRF Cars (Ours) +Figure 4. Qualitative comparison between unconditional generative models trained on ABO Tables and SRN Cars. + +
MethodCars 1-viewCars 2-viewChairs 1-viewChairs 2-view
PSNR↑SSIM↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓FID↓
3DiM [53]21.010.57-8.99----17.050.53-6.57----
PixelNeRF [55]23.170.900.146‡59.24†25.660.94--23.720.910.128‡38.49†26.200.94--
SRN [45]22.25§0.89§0.129‡41.21†24.84§0.92§--22.89§0.89§0.104‡26.51†24.48§0.92§--
CodeNeRF [21]23.800.910.118*56.34*25.710.930.108*56.13*23.660.900.106*31.65*25.630.910.097*29.90*
VisionNeRF [25]22.880.910.08421.31†----24.480.930.07710.05†----
Ours (1-stage)23.520.910.07816.3926.490.940.05410.6624.350.930.06710.1326.940.950.05510.85
+ +Table 2. Single-view and two-view reconstruction results on SRN Cars and Chairs. For consistency with prior work, we use view #64 of the test scene as single-view input and view #64 and #104 as two-view input. † denotes results reported by 3DiM [53]. ‡ denotes results reported by VisionNeRF [25], § denotes results reported by PixelNeRF [21], * denotes results reproduced by us using the official code. - indicates results are unavailable. + +images in the entire dataset for ABO Tables, respectively. + +Comparison to the State of the Art As shown in Table 1, on SRN Cars, SSDNeRF (1-stage) outperforms EG3D in KID (a more suitable measure for small datasets) by a clear margin. Meanwhile, its FID is drastically better than Functa, which uses an LDM but with low dimensional latent codes. On ABO Tables, SSDNeRF shows significantly better performance than EG3D and DiffRF. + +Single- vs. Two-stage On SRN Cars, we compare the proposed single-stage training against two-stage training with tuned TV regularization using the same model architecture. The results in Table 1 indicate substantial advantage of single-stage training (KID/10 $^{-3}$ 3.47 vs. 6.38). + +Qualitative Results As shown in Fig. 4, SSDNeRF generates more regular geometries than the slightly skewed and distorted shapes by EG3D [5]. Compared to DiffRF [29], our method produces sharp details and reflective materials, + +thanks to our more expressive model with latents of higher spatial resolution and view-dependent NeRF decoder. + +# 5.3. Sparse-View NeRF Reconstruction + +This section presents experiments on 3D reconstruction from sparse-view images of unseen objects in SRN Cars and Chairs test sets. The Cars dataset presents the challenge of recovering distinct textures, while the Chairs dataset requires accurate reconstruction of diverse shapes. Models are trained on all images of the training set for 80K iterations, as we find that longer schedule leads to decaying performance in reconstructing unseen objects. This behaviour is in accordance with the interpolation results in § 5.5. + +Evaluation Protocol and Metrics We use the evaluation protocol and metrics in PixelNeRF [55]. Given input images sampled from each test scene, we obtain the triplane scene code via guidance-finetuning and evaluate novel view synthesis quality with respect to the unseen images. + +![](images/49d6562de1a695052892f6a857e6a41805f3310a43157d3f04148967a61f6fb8.jpg) +Figure 5. Qualitative comparison of single-view reconstruction methods on unseen test objects in SRN Cars and Chairs. + +![](images/9d4af214eaa886395617b5124a9cbb17b7208255d6465580ae0cc20b603d9aae.jpg) + +The image quality metrics include average peak signal-to-noise-ratio (PSNR), structural similarity (SSIM) [52], and Learned Perceptual Image Patch Similarity (LPIPS) [56]. In addition, we evaluate the FID between all synthesized images and ground truth images as in 3DiM [53]. + +Comparison to the State of the Art Table 2 compares SSDNeRF against previous approaches in single-view and two-view reconstruction settings. Overall, SSDNeRF reaches the best LPIPS of all tasks, indicating the best perceptual fidelity. In contrast, 3DiM generates high quality images (best FID) but with the lowest fidelity to the ground truth (lowest PSNR); CodeNeRF reports the best PSNR on single-view Cars, but its limited expressiveness leads to blurry outputs (Fig. 5) and less competitive LPIPS and FID; VisionNeRF achieves a balanced performance on all single-view metrics, but may struggle to generate textural details on the unseen side of cars (e.g., the other side of the ambulance in Fig. 5). Moreover, SSDNeRF exhibits a clear advantage in two-view reconstruction, achieving the best performance on all relevant metrics. + +Single- vs. Two-stage As demonstrated in Table 3, the model trained in a single stage (A0) outperforms the same architecture trained in two stages with TV regularization (A1) in all metrics of single-view reconstruction. + +Ablation Studies on Test-Time Finetuning As shown in Table 3, we evaluate the effectiveness of test-time finetuning and the contribution of the learned diffusion prior with two ablation experiments: (A2) removing the diffusion loss during finetuning and using only the rendering loss, and (A3) omitting the finetuning process entirely. The results indicate that finetuning with single-view rendering loss provides only marginal improvements over guided sampling (A2 vs. A3), while the learned diffusion prior significantly + +
IDTrainingFinetuningPSNR↑SSIM↑LPIPS↓FID↓
A01-stageRend + Diff23.520.9130.07816.39
A12-stageRend + Diff22.830.9060.09020.97
A21-stageRend23.130.9070.08827.93
A31-stageNone23.070.9050.09230.95
+ +Table 3. Ablation results on single-view reconstruction of SRN Cars. + +![](images/6be3c0ffb17e1dfb41f87f2dc35db23292d7a4fab94c65b9ab3a7c2097c00cd5.jpg) +Figure 6. LPIPS scores (lower is better) of novel view synthesis from sparse-to-dense inputs, evaluated on SRN Cars test set. The triplane baselines adopt mean initialization for better performance. + +boosts the LPIPS and FID scores (A0 vs. A2), highlighting its importance in recovering sharp and distinct contents. Moreover, the qualitative results in Fig. 5 reveal that views with higher overlap to the input view benefit the most from finetuning, meeting our expectation that finetuning helps faithfully reconstruct the exact observations. + +Sparse-to-Dense Reconstruction To validate that SSDNeRF seamlessly bridges sparse- and dense-view NeRF reconstruction, we evaluate its novel view synthesis performance with the number of input views varying from 1 to 32. We compare our model to the triplane NeRF baseline trained as an auto-decoder with optional TV regularization instead of diffusion prior. Meanwhile, we also evaluate CodeNeRF [21], an auto-decoder with 256-d latent codes. The results in Fig. 6 show that SSDNeRF excels in all settings, + +especially in 1 to 4 views. In contrast, CodeNeRF is outperformed by vanilla triplane NeRF with more views. + +# 5.4. Training SSDNeRF on Sparse-View Dataset + +In this section, we train SSDNeRF on a sparse-view subset of the full SRN Cars training set, in which a fixed set of only three views are randomly picked from each scene. Note that a reasonable decline in performance compared to dense-view training is expected as the whole training dataset has been reduced to $6\%$ of its original size. + +Unconditional Generation We adopt a training trick that resets the triplane codes to their mean value halfway through training. This helps to prevent the model from getting stuck in a local minimum that overfits geometric artifacts. We also double the length of the training schedule accordingly. The model achieves a decent FID of $19.04 \pm 1.10$ and a $\mathrm{KID} / 10^{-3}$ of $8.28 \pm 0.60$ . Results are visualized in Fig. 7. + +Single-View Reconstruction We adopt the same training strategy as in § 5.3. With our guidance-finetuning approach, the model achieves an LPIPS score of 0.106, even outperforming most of the previous methods in Table 2 that use the full training set. + +Comparison to TV Regularization Fig. 8 (b) shows the RGB images and geometries represented by the scene latent codes learned from three views during training. By comparison, vanilla triplane auto-decoder with TV regularization (Fig. 8 (a)) often fails to reconstruct a scene from sparse views, leading to severe geometric artifacts. As a result, previously it has been infeasible to train two-stage models with expressive latents on sparse-view data. + +# 5.5. NeRF Interpolation + +Following DDIM [47], we can sample two initial values $x^{(T)} \sim \mathcal{N}(0, I)$ , interpolate them using spherical linear interpolation [43], and then use the deterministic solver to obtain interpolated samples. However, as noted by [34, 37], standard Gaussian diffusion models often result in nonsmooth interpolation. In SSDNeRF (with results shown in Fig. 9), we find that the model (a) trained with early stopping for sparse-view reconstruction produces reasonably smooth transitions between samples, while the model (b) trained with a longer schedule for unconditional generation produces distinct yet discontinuous samples. This suggests that early stopping preserves a smoother prior, leading to better generalization for sparse-view reconstruction. + +# 6. Conclusion + +In this paper, we propose SSDNeRF, which combines the diffusion model and NeRF representation through a novel single-stage training paradigm with an end-to-end justifiable loss. Notably, it overcomes the limitations in previous work where implicit neural fields must be obtained from + +![](images/3d9b02bc5936756dd254acea334242c101e7eee1d099efa483db6bdf6f7315ff.jpg) +Figure 7. Images generated by SSDNeRF trained on a 3-view subset of SRN Cars training set. + +![](images/6528710554a00ad0a8d10d9183222b477c74e7c9556896a9090b92c9b4ff213d.jpg) +(a) Auto-decoder training w/ TV + +![](images/f229c7790d811fd3eb42ec3ec9a2b8f1224523d1fbe345e9855e7ddf5ca49142.jpg) +(b) 1-stage joint training with diffusion + +![](images/d7a5022e3f655c21a283a26ff6eeb65cfcd0ec2a845f2ff9dcab5c6b464ceb80.jpg) +Figure 8. Qualitative comparison between scene codes learned from 3 views by (a) triplane auto-decoder with TV regularization vs. (b) single-stage diffusion NeRF. +(a) Models trained with early stopping (80K iters) for sparse-view reconstruction + +![](images/9948b75b0a03f87389db5a7e462e11c4c2badd23c3f280bb2db8b73ecc382d61.jpg) +(b) Models trained with long schedule (1M iters) for unconditional generation +Figure 9. Interpolation between the leftmost and rightmost samples using DDIM [47]. + +dense observations first, before training the diffusion models to learn their manifold. With strong performance on multiple benchmarks, SSDNeRF demonstrates a significant advancement towards a unified framework for general 3D content manipulation. + +Limitations and Future Work Currently, our method relies on ground truth camera parameters during both training and testing. Future work may explore transform-invariant models. Additionally, the diffusion prior can become discontinuous with prolonged training, which affects generalization. Although early stopping is temporarily used, a better network design or a larger training dataset may be able to address this problem fundamentally. + +Acknowledgements We thank Norman Müller for sharing the baseline results on ABO Tables. Hansheng Chen and Wei Tian acknowledge the funding by the National Natural Science Foundation of China (No. 52002285), the Shanghai Science and Technology Commission (No. 21ZR1467400), the original research project of Tongji University (No. 22120220593), the National Key R&D Program of China (No. 2021YFB2501104), and the Natural Science Foundation of Chongqing (No. 2023NSCQ-MSX4511). + +# References + +[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J. Mitra, and Paul Guerrero. RenderDiffusion: Image diffusion for 3D reconstruction, inpainting and generation. In CVPR, 2023. 2 +[2] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, Afshin Dehghan, and Josh Susskind. Gaudi: A neural architect for immersive 3d scene generation. In NeurIPS, 2022. 2, 3, 4 +[3] Mikołaj Binkowski, Danica J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In ICLR, 2018. 5 +[4] Eric Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 2, 5 +[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2, 5, 6 +[6] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 5 +[7] Anpei Chen, Zexiang Xu, Xinyue Wei, Siyu Tang, Hao Su, and Andreas Geiger. Factor fields: A unified framework for neural fields and beyond, 2023. 2, 3 +[8] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, pages 14124-14133, 2021. 1, 2, 5 +[9] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik. Abo: Dataset and benchmarks for real-world 3d object understanding. In CVPR, 2022. 5 +[10] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In CVPR, 2023. 2 +[11] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W. Taylor, and Joshua M. Susskind. Unconstrained scene generation with locally conditioned radiance fields. In ICCV, 2021. 2 +[12] Emilien Dupont, Hyunjik Kim, S. M. Ali Eslami, Danilo Jimenez Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you can treat it like one. In ICML, 2022. 2, 3, 4, 5 +[13] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja + +Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In NeurIPS, 2022. 1, 2 +[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2 +[15] Jiatao Gu, Qingzhe Gao, Shuangfei Zhai, Baoquan Chen, Lingjie Liu, and Josh Susskind. Learning controllable 3d diffusion models from single-view images, 2023. 4 +[16] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In ICLR, 2022. 1, 2 +[17] Jiatao Gu, Alex Trevithick, Kai-En Lin, Josh Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nerf-guided distillation from 3d-aware diffusion, 2023. 2 +[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 5 +[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 5 +[20] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 4 +[21] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In ICCV, pages 12949-12958, 2021. 2, 6, 7 +[22] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5 +[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 2, 3 +[24] Yiyi Liao, Katja Schwarz, Lars Mescheder, and Andreas Geiger. Towards unsupervised learning of generative models for 3d controllable image synthesis. In CVPR, 2020. 2 +[25] Kai-En Lin, Lin Yen-Chen, Wei-Sheng Lai, Tsung-Yi Lin, Yi-Chang Shih, and Ravi Ramamoorthi. Vision transformer for nef-based view synthesis from a single input image. In WACV, 2023. 1, 2, 6 +[26] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11461-11471, 2022. 2 +[27] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 2 +[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2 +[29] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In CVPR, 2023. 2, 3, 4, 5, 6 +[30] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised + +learning of 3d representations from natural images. In ICCV, 2019. 2 +[31] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 1, 2 +[32] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 2, 3 +[33] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 4 +[34] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In CVPR, 2022. 8 +[35] Daniel Rebain, Mark Matthews, Kwang Moo Yi, Dmitry Lagun, and Andrea Tagliasacchi. Lolnerf: Learn from one look. In CVPR, 2022. 2, 3 +[36] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In ICML, pages 1278-1286, 2014. 3 +[37] Severi Rissanen, Markus Heinonen, and Arno Solin. Generative modelling with inverse heat dissipation. In ICLR, 2023. 8 +[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3 +[39] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241, 2015. 5 +[40] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In ICLR, 2022. 5 +[41] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In NeurIPS, 2020. 2 +[42] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. In NeurIPS, 2022. 2 +[43] Ken Shoemake. Animating rotation with quaternion curves. In Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 245-254, 1985. 8 +[44] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In CVPR, 2023. 2, 3, 4 +[45] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In NeurIPS, 2019. 2, 5, 6 +[46] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. In NeurIPS, 2022. 2 +[47] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2, 3, 4, 8 + +[48] Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. Maximum likelihood training of score-based diffusion models. In NeurIPS, 2021. 4 +[49] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 4 +[50] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In NeurlPS, 2021. 2, 3, 4, 5 +[51] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In CVPR, 2023. 2 +[52] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 7 +[53] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. In ICLR, 2023. 2, 6, 7 +[54] Antoine Wehenkel and Gilles Louppe. Diffusion priors in variational autoencoders. In ICML Workshops, 2021. 4 +[55] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 1, 2, 6 +[56] Richard Zhang, Phillip Isola, Alexei Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7 +[57] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In CVPR, 2022. 2, 5 +[58] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In CVPR, 2023. 2 \ No newline at end of file diff --git a/2023/Single-Stage Diffusion NeRF_ A Unified Approach to 3D Generation and Reconstruction/images.zip b/2023/Single-Stage Diffusion NeRF_ A Unified Approach to 3D Generation and Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e1386c681a2ec6a2097ad3328d74c705e0def82a --- /dev/null +++ b/2023/Single-Stage Diffusion NeRF_ A Unified Approach to 3D Generation and Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77fddc4a11761e70f237cf04c2accb6000f3c95147925e8eecc3eac29dceadd2 +size 608331 diff --git a/2023/Single-Stage Diffusion NeRF_ A Unified Approach to 3D Generation and Reconstruction/layout.json b/2023/Single-Stage Diffusion NeRF_ A Unified Approach to 3D Generation and Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5f88a38b469428482451b3c164e92aa4d8896d2e --- /dev/null +++ b/2023/Single-Stage Diffusion NeRF_ A Unified Approach to 3D Generation and Reconstruction/layout.json @@ -0,0 +1,10061 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 103, + 460, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 103, + 460, + 138 + ], + "spans": [ + { + "bbox": [ + 134, + 103, + 460, + 138 + ], + "type": "text", + "content": "Single-Stage Diffusion NeRF: A Unified Approach to 3D Generation and Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 160, + 512, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 160, + 512, + 177 + ], + "spans": [ + { + "bbox": [ + 78, + 160, + 512, + 177 + ], + "type": "text", + "content": "Hansheng Chen," + }, + { + "bbox": [ + 78, + 160, + 512, + 177 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 78, + 160, + 512, + 177 + ], + "type": "text", + "content": " Jiatao Gu," + }, + { + "bbox": [ + 78, + 160, + 512, + 177 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 160, + 512, + 177 + ], + "type": "text", + "content": " Anpei Chen," + }, + { + "bbox": [ + 78, + 160, + 512, + 177 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 78, + 160, + 512, + 177 + ], + "type": "text", + "content": " Wei Tian," + }, + { + "bbox": [ + 78, + 160, + 512, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 160, + 512, + 177 + ], + "type": "text", + "content": " Zhuowen Tu," + }, + { + "bbox": [ + 78, + 160, + 512, + 177 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 78, + 160, + 512, + 177 + ], + "type": "text", + "content": " Lingjie Liu," + }, + { + "bbox": [ + 78, + 160, + 512, + 177 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 78, + 160, + 512, + 177 + ], + "type": "text", + "content": " Hao Su" + }, + { + "bbox": [ + 78, + 160, + 512, + 177 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 178, + 466, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 178, + 466, + 209 + ], + "spans": [ + { + "bbox": [ + 126, + 178, + 466, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 126, + 178, + 466, + 209 + ], + "type": "text", + "content": "Tongji University " + }, + { + "bbox": [ + 126, + 178, + 466, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 178, + 466, + 209 + ], + "type": "text", + "content": "Apple " + }, + { + "bbox": [ + 126, + 178, + 466, + 209 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 126, + 178, + 466, + 209 + ], + "type": "text", + "content": "ETH Zürich " + }, + { + "bbox": [ + 126, + 178, + 466, + 209 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 126, + 178, + 466, + 209 + ], + "type": "text", + "content": "University of California, San Diego " + }, + { + "bbox": [ + 126, + 178, + 466, + 209 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 126, + 178, + 466, + 209 + ], + "type": "text", + "content": "University of Pennsylvania" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 60, + 224, + 264, + 376 + ], + "blocks": [ + { + "bbox": [ + 60, + 224, + 264, + 376 + ], + "lines": [ + { + "bbox": [ + 60, + 224, + 264, + 376 + ], + "spans": [ + { + "bbox": [ + 60, + 224, + 264, + 376 + ], + "type": "image", + "image_path": "9d1343f1e853acd282882509fbc88fdafe21842eaed4dbf19b49f503e7bd72eb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 378, + 546, + 401 + ], + "lines": [ + { + "bbox": [ + 46, + 378, + 546, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 546, + 401 + ], + "type": "text", + "content": "Figure 1. During training, SSDNeRF jointly learns triplane features of individual scenes, a shared NeRF decoder, and a triplane diffusion prior. During testing, it can perform (a) unconditional generation, (b) single-view reconstruction, as well as multi-view reconstruction." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 284, + 224, + 535, + 376 + ], + "blocks": [ + { + "bbox": [ + 284, + 224, + 535, + 376 + ], + "lines": [ + { + "bbox": [ + 284, + 224, + 535, + 376 + ], + "spans": [ + { + "bbox": [ + 284, + 224, + 535, + 376 + ], + "type": "image", + "image_path": "f339408477b3f3c7291875e05a1e747a03ef1055091c536ce453bb7f528463d0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 417, + 192, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 417, + 192, + 430 + ], + "spans": [ + { + "bbox": [ + 143, + 417, + 192, + 430 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 444, + 290, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 444, + 290, + 684 + ], + "spans": [ + { + "bbox": [ + 45, + 444, + 290, + 684 + ], + "type": "text", + "content": "3D-aware image synthesis encompasses a variety of tasks, such as scene generation and novel view synthesis from images. Despite numerous task-specific methods, developing a comprehensive model remains challenging. In this paper, we present SSDNeRF, a unified approach that employs an expressive diffusion model to learn a generalizable prior of neural radiance fields (NeRF) from multi-view images of diverse objects. Previous studies have used two-stage approaches that rely on pretrained NeRFs as real data to train diffusion models. In contrast, we propose a new single-stage training paradigm with an end-to-end objective that jointly optimizes a NeRF auto-decoder and a latent diffusion model, enabling simultaneous 3D reconstruction and prior learning, even from sparsely available views. At test time, we can directly sample the diffusion prior for unconditional generation, or combine it with arbitrary observations of unseen objects for NeRF reconstruction. SSDNeRF demonstrates robust results comparable to or better than leading task-specific methods in unconditional generation and single/sparse-view 3D reconstruction." + }, + { + "bbox": [ + 45, + 444, + 290, + 684 + ], + "type": "inline_equation", + "content": "^6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 417, + 387, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 417, + 387, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 417, + 387, + 430 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 546, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 676 + ], + "type": "text", + "content": "Synthesizing 3D visual contents has gained significant attention in computer vision and graphics, thanks to advances in neural rendering and generative models. Although numerous methods have emerged to handle individual tasks, such as single-/multi-view 3D reconstruction and 3D content generation, it remains a major challenge to develop a comprehensive framework that bridges the state of the art of multiple tasks. For instance, neural radiance fields (NeRF) [28] have shown impressive results in novel view synthesis by solving the inverse rendering problem via perscene fitting, which is suitable for dense-view inputs but difficult to generalize to sparse observations. In contrast, many sparse-view 3D reconstruction methods [55, 8, 25] rely on feed-forward image-to-3D encoders, but they are unable to handle ambiguity in the occluded region and generate crisp images. Regarding unconditional generation, 3D-aware generative adversarial networks (GAN) [31, 5, 16, 13] are partially limited in their usage of single-image discriminators, which cannot reason cross-view relationships to effectively learn from multi-view data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we propose a unified approach to various 3D tasks (Fig. 1) by developing a holistic model that learns generalizable 3D priors from multi-view images. Inspired" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 693, + 267, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 693, + 267, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 693, + 267, + 714 + ], + "type": "text", + "content": "*Work done during a remote internship with UCSD. \n6Project page: https://lakonik.github.io/ssdnerf" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2416" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "by the success of 2D diffusion models [19, 47, 27, 38, 26], we present the Single-Stage Diffusion NeRF (SSDNeRF), which models the generative prior of scene latent codes with a 3D latent diffusion model (LDM)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 337 + ], + "type": "text", + "content": "While similar LDMs have been applied in 2D and 3D generation in previous work [50, 38, 12, 2, 44, 29], they typically require two-stage training, where the first stage pretrains the variational auto-encoders (VAE) [23] or autodecoders [32] without diffusion models. In the case of diffusion NeRFs, however, we argue that two-stage training induces noisy patterns and artifacts in the latent code due to the uncertain nature of inverse rendering, particularly when training from sparse-view data, which prevents the diffusion model from learning a clean latent manifold effectively. To address this issue, we introduce a novel single-stage training paradigm that enables end-to-end learning of diffusion and NeRF weights (" + }, + { + "bbox": [ + 46, + 121, + 288, + 337 + ], + "type": "inline_equation", + "content": "\\S 4.1" + }, + { + "bbox": [ + 46, + 121, + 288, + 337 + ], + "type": "text", + "content": "). This approach blends the generative and the rendering biases coherently for improved performance overall and allows for training on sparse-view data. Additionally, we show that the learned 3D priors of unconditional diffusion models can be exploited for flexible test-time scene sampling from arbitrary observations (" + }, + { + "bbox": [ + 46, + 121, + 288, + 337 + ], + "type": "inline_equation", + "content": "\\S 4.2" + }, + { + "bbox": [ + 46, + 121, + 288, + 337 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 338, + 287, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 385 + ], + "type": "text", + "content": "We evaluate SSDNeRF on multiple datasets of categorical single-object scenes, demonstrating strong performance overall. Our approach represents a significant step towards a unified framework for various 3D tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 387, + 274, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 387, + 274, + 398 + ], + "spans": [ + { + "bbox": [ + 59, + 387, + 274, + 398 + ], + "type": "text", + "content": "To summarize, our main contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 403, + 286, + 533 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 59, + 403, + 286, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 403, + 286, + 437 + ], + "spans": [ + { + "bbox": [ + 59, + 403, + 286, + 437 + ], + "type": "text", + "content": "- We introduce SSDNeRF, a unified approach to all-round performance in unconditional 3D generation and image-based reconstruction;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 439, + 286, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 439, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 59, + 439, + 286, + 498 + ], + "type": "text", + "content": "- We propose a novel single-stage training paradigm that jointly learns NeRF reconstruction and diffusion model from multi-view images of a large number of objects. Notably, this enables training on as sparse as three views per scene, which is previously infeasible;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 499, + 286, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 499, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 59, + 499, + 286, + 533 + ], + "type": "text", + "content": "- A guidance-finetuning sampling scheme is developed to exploit the learned diffusion priors for 3D reconstruction from arbitrary number of views at test time." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 548, + 134, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 134, + 560 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 134, + 560 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "3D GANs The generative adversarial framework [14] has been successfully adapted for 3D generation by integrating projection-based rendering into the generator. A variety of 3D representations have been explored previously, including point clouds, cuboids, spheres [24] and voxels [30] in early works, the more recent radiance fields [4, 41, 11, 42, 46] and feature fields [31, 16, 5] with volume renderer, and differentiable surface [13] with mesh renderer. The above methods are all trained with 2D image discriminators that are unable to reason cross-view relationships, making them heavily dependent on model bias for 3D consistency. As a result, multi-view data cannot be effectively exploited to" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "learn complex and diverse geometries. 3D GANs are mostly applied in unconditional generation. Although 3D completion from images is possible through GAN inversion [11], faithfulness is not guaranteed due to limited latent expressiveness, as shown in [29, 1]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 136, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 136, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 136, + 545, + 327 + ], + "type": "text", + "content": "View-Conditioned Regression and Generation Sparse-view 3D reconstruction can be tackled by regressing novel views from input images. Various architectures [8, 55, 25, 57] have been proposed to encode images into volume features, which can be projected to supervised target views through volume rendering. However, they cannot reason ambiguity and generate diverse and meaningful contents, which often leads to blurry results. In contrast, image-conditioned generative models are better at synthesizing distinct contents. 3DiM [53] proposes to generate novel views from a view-conditioned image diffusion model, but the model lacks 3D consistency bias. [58, 10, 17] distill priors of image-conditioned 2D diffusion models into NeRFs to enforce 3D constraints. These methods are parallel to our track as they model cross-view relationships in the image space, while our model is inherently 3D." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 331, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 510 + ], + "type": "text", + "content": "Auto-Decoders and Diffusion NeRF NeRF's per-scene fitting scheme can be generalized to multi-scene fitting by sharing part of the parameters across all scenes, leaving the rest as individual scene codes [7]. Therefore, multi-scene NeRFs can be trained as auto-decoders [32], where the code bank and shared decoder weights are jointly learned. With proper architectures, scene codes can be treated as latents with Gaussian priors, allowing 3D completion and even generation [21, 45, 35]. However, like 3D GANs, the latents are not expressive enough for faithful reconstruction of detailed objects. [2, 12, 51] improve upon vanilla auto-decoders with latent diffusion priors. DiffRF [29] leverages the diffusion prior to perform 3D completion. These methods train the auto-decoders and diffusion models in two separate stages, which is subject to the limitations in § 3.2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 520, + 384, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 520, + 384, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 520, + 384, + 533 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 540, + 451, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 451, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 451, + 552 + ], + "type": "text", + "content": "3.1. NeRF as an Auto-Decoder" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Given a set of 2D images of a scene and their camera parameters, one can fit a scene model to reconstruct the light field in 3D space, expressed by a plenoptic function " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "y_{\\psi}(r)" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " parameterizes the endpoint and direction of a ray in the world space, " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " denotes the scene model parameters, and " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "y \\in \\mathbb{R}_+^3" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " represents the received light in RGB format. NeRF [28] represents the light field as integrated radiance along rays through the 3D volume. It models the scene geometry and appearance as functions of the position " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " and view direction " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "d \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " of a point in the world space, expressed as " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\rho_{\\psi}(p)" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "c_{\\psi}(p,d)" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " respectively, where " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\rho \\in \\mathbb{R}_{+}" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " is the density output and " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{R}_{+}^{3}" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " is the RGB color output. Differentiable volume rendering is applied to compose the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2417" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 284, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 284, + 84 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 284, + 84 + ], + "type": "text", + "content": "received light " + }, + { + "bbox": [ + 46, + 72, + 284, + 84 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 72, + 284, + 84 + ], + "type": "text", + "content": " from multiple point samples along a ray " + }, + { + "bbox": [ + 46, + 72, + 284, + 84 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 72, + 284, + 84 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 286, + 156 + ], + "type": "text", + "content": "NeRF can also generalize to multi-scene settings by sharing part of the model parameters across all scenes [7]. Given observations of multiples scenes " + }, + { + "bbox": [ + 46, + 84, + 286, + 156 + ], + "type": "inline_equation", + "content": "\\{y_{ij}^{\\mathrm{gt}},r_{ij}^{\\mathrm{gt}}\\}" + }, + { + "bbox": [ + 46, + 84, + 286, + 156 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 84, + 286, + 156 + ], + "type": "inline_equation", + "content": "y_{ij}^{\\mathrm{gt}},r_{ij}^{\\mathrm{gt}}" + }, + { + "bbox": [ + 46, + 84, + 286, + 156 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 84, + 286, + 156 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 84, + 286, + 156 + ], + "type": "text", + "content": "-th pair of pixel RGB and ray of the " + }, + { + "bbox": [ + 46, + 84, + 286, + 156 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 84, + 286, + 156 + ], + "type": "text", + "content": "-th scene, one can optimize the per-scene codes " + }, + { + "bbox": [ + 46, + 84, + 286, + 156 + ], + "type": "inline_equation", + "content": "\\{x_i\\}" + }, + { + "bbox": [ + 46, + 84, + 286, + 156 + ], + "type": "text", + "content": " and shared parameters " + }, + { + "bbox": [ + 46, + 84, + 286, + 156 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 84, + 286, + 156 + ], + "type": "text", + "content": " by minimizing the L2 rendering loss:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 161, + 287, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 161, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 58, + 161, + 287, + 192 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e n d}} \\left(\\left\\{x _ {i} \\right\\}, \\psi\\right) = \\mathbb {E} _ {i} \\left[ \\sum_ {j} \\frac {1}{2} \\left\\| y _ {i j} ^ {\\mathrm {g t}} - y _ {\\psi} \\left(x _ {i}, r _ {i j} ^ {\\mathrm {g t}}\\right) \\right\\| ^ {2} \\right]. \\tag {1}", + "image_path": "b4071b21e8bc13d914a9028d0e2cc06aebd07496444dd525276514b714dd809c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 194, + 286, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 286, + 255 + ], + "type": "text", + "content": "With this objective, the model is trained as an auto-decoder [32], where the scene codes " + }, + { + "bbox": [ + 46, + 194, + 286, + 255 + ], + "type": "inline_equation", + "content": "\\{x_{i}\\}" + }, + { + "bbox": [ + 46, + 194, + 286, + 255 + ], + "type": "text", + "content": " can be interpreted as the latent codes, and the plenoptic function can be regarded as a decoder in the form of " + }, + { + "bbox": [ + 46, + 194, + 286, + 255 + ], + "type": "inline_equation", + "content": "p_{\\psi}(\\{y_j\\} |x,\\{r_j\\})\\coloneqq \\prod_j\\mathcal{N}(y_j|y_\\psi (x,r_j),I)" + }, + { + "bbox": [ + 46, + 194, + 286, + 255 + ], + "type": "text", + "content": ", assuming independent Gaussians." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 257, + 286, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 286, + 341 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 286, + 341 + ], + "type": "text", + "content": "Challenges in Bridging Generation and Reconstruction An auto-decoder with trained weights " + }, + { + "bbox": [ + 46, + 257, + 286, + 341 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 257, + 286, + 341 + ], + "type": "text", + "content": " can perform unconditional generation by decoding latent codes drawn from a Gaussian prior [35]. However, to ensure continuity in generation, a low-dimensional latent space and a complex decoder is required, which adds to the difficulty in optimizing the latent code to faithfully reconstruct any given views." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 347, + 185, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 347, + 185, + 358 + ], + "spans": [ + { + "bbox": [ + 47, + 347, + 185, + 358 + ], + "type": "text", + "content": "3.2. Latent Diffusion Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "text", + "content": "Latent diffusion models (LDM) learn a prior distribution " + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "inline_equation", + "content": "p_{\\phi}(x)" + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "text", + "content": " in the latent space with parameters " + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "text", + "content": ", which enables the usage of more expressive latent representations, such as 2D grids for images [50, 38]. For neural field generation, previous work [2, 29, 12, 44] adopts a two-stage training scheme, where the auto-decoder is trained first to obtain the per-scene latent " + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "text", + "content": ", which is then treated as real data to train the LDM. The LDM injects Gaussian perturbation " + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "text", + "content": " into the code " + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "text", + "content": ", yielding a noisy code " + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "inline_equation", + "content": "x_{i}^{(t)} \\coloneqq \\alpha^{(t)} x_{i} + \\sigma^{(t)} \\epsilon" + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "text", + "content": " at diffusion time step " + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "text", + "content": ", under empirical noise schedule functions " + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\alpha^{(t)}, \\sigma^{(t)}" + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "text", + "content": ". A denoising network with trainable weights " + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "text", + "content": " is then tasked with removing the noise from " + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "inline_equation", + "content": "x_{i}^{(t)}" + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "text", + "content": " to predict a denoised code " + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\hat{x}_{i}" + }, + { + "bbox": [ + 46, + 365, + 287, + 537 + ], + "type": "text", + "content": ". The network is typically trained with a simplified L2 denoising loss:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 73, + 542, + 287, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 542, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 73, + 542, + 287, + 570 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d i f f}} (\\phi) = \\underset {i, t, \\epsilon} {\\mathbb {E}} \\left[ \\frac {1}{2} w ^ {(t)} \\left\\| \\hat {x} _ {\\phi} \\left(x _ {i} ^ {(t)}, t\\right) - x _ {i} \\right\\| ^ {2} \\right], \\tag {2}", + "image_path": "bbf2407292e975784b5dd99255e4bd0dc911c5307b94351e1e1026bfdd752f6f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 576, + 287, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 576, + 287, + 614 + ], + "spans": [ + { + "bbox": [ + 46, + 576, + 287, + 614 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 576, + 287, + 614 + ], + "type": "inline_equation", + "content": "t \\sim \\mathcal{U}(0,T)" + }, + { + "bbox": [ + 46, + 576, + 287, + 614 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 576, + 287, + 614 + ], + "type": "inline_equation", + "content": "w^{(t)}" + }, + { + "bbox": [ + 46, + 576, + 287, + 614 + ], + "type": "text", + "content": " is an empirical time dependent weighting function, and " + }, + { + "bbox": [ + 46, + 576, + 287, + 614 + ], + "type": "inline_equation", + "content": "\\hat{x}_{\\phi}(x_i^{(t)},t)" + }, + { + "bbox": [ + 46, + 576, + 287, + 614 + ], + "type": "text", + "content": " formulates the time-conditioned denoising network." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Unconditional/Guided Sampling With trained weights " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": ", one can sample from the diffusion prior " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "p_{\\psi}(x)" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " using a variety of solvers (e.g., DDIM [47]) that recursively denoise " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "x^{(t)}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": ", starting from random Gaussian noise " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "x^{(T)}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": ", until reaching the denoised state " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "x^{(0)}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": ". Moreover, the sampling process can be guided by the gradients of the rendering loss against known observations, allowing 3D reconstruction from images at test time [29]." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 326, + 70, + 515, + 162 + ], + "blocks": [ + { + "bbox": [ + 326, + 70, + 515, + 162 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 515, + 162 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 515, + 162 + ], + "type": "image", + "image_path": "c311386e4d3a3a5f7956006093d8918813e36ffac7196bd2bfacf7dd1dcd8568.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 167, + 547, + 245 + ], + "lines": [ + { + "bbox": [ + 304, + 167, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 547, + 245 + ], + "type": "text", + "content": "Figure 2. Visualization of the scene code " + }, + { + "bbox": [ + 304, + 167, + 547, + 245 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{XZ}}" + }, + { + "bbox": [ + 304, + 167, + 547, + 245 + ], + "type": "text", + "content": " at XZ plane. Left column: Two-stage training without TV regularization induces noise and fails in 3-view reconstruction. Mid column: TV regularization imposes smoothing prior at the cost of textural details (top), yet still struggles to cope with sparse views (bottom). Right column: Our single-stage training encourages smooth yet detailed latents and allows for training with sparse views." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 262, + 545, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 383 + ], + "type": "text", + "content": "Limitations of Two-Stage Training for 3D Tasks While LDMs with 2D image VAEs are typically trained in two stages [50, 38], training LDMs with NeRF auto-decoders poses an unprecedented challenge. An expressive latent code is underdetermined when obtained via rendering-based optimization, leading to noisy patterns that distract denoising networks (top-left of Fig. 2). Additionally, reconstructing NeRFs from sparse views without a learned prior is exceptionally difficult (bottom-left of Fig. 2), limiting training to dense-views settings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 392, + 411, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 411, + 406 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 411, + 406 + ], + "type": "text", + "content": "4. Proposed Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 412, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 484 + ], + "type": "text", + "content": "To build a holistic model that unifies 3D generation and reconstruction, we propose SSDNeRF, a framework that conjoins the expressive triplane NeRF auto-decoder with a triplane latent diffusion model. Fig. 3 provides an overview of the model. In the following subsections, we elaborate on how training and testing are performed in detail." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 491, + 506, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 491, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 305, + 491, + 506, + 504 + ], + "type": "text", + "content": "4.1. Single-Stage Diffusion NeRF Training" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "text", + "content": "An auto-decoder can be regarded as a type of VAE that uses a lookup table encoder instead of the typical neural network encoder. As such, the training objective can be derived in a similar way as for VAEs. With NeRF decoder " + }, + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "inline_equation", + "content": "p_{\\psi}(\\{y_j\\} |x,\\{r_j\\})" + }, + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "text", + "content": " and diffusion latent prior " + }, + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "inline_equation", + "content": "p_{\\phi}(x)" + }, + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "text", + "content": ", the training objective is to minimize variational upper bound on the negative log-likelihood (NLL) of observed data " + }, + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\{y_{ij}^{\\mathrm{gt}}, r_{ij}^{\\mathrm{gt}}\\}" + }, + { + "bbox": [ + 304, + 510, + 545, + 628 + ], + "type": "text", + "content": " [23, 36, 50]. In this paper, a simplified training loss is derived by ignoring the uncertainty (variance) in latent codes:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 310, + 635, + 545, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 635, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 310, + 635, + 545, + 670 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\underbrace {\\underset {i} {\\mathbb {E}} \\left[ - \\log p _ {\\psi} \\left(\\left\\{y _ {i j} ^ {\\mathrm {g t}} \\right\\} \\mid x _ {i} , \\left\\{r _ {i j} ^ {\\mathrm {g t}} \\right\\}\\right) \\right]} _ {\\text {r e n d e r i n g l o s s} \\mathcal {L} _ {\\text {r e n d}}} + \\underbrace {\\underset {i} {\\mathbb {E}} \\left[ - \\log p _ {\\phi} \\left(x _ {i}\\right) \\right]} _ {\\text {p r i o r t e r m}}, \\tag {3}", + "image_path": "62420eb966891ef83a8062e4f40661791e81fca314a6cdce10165ee5b839cc79.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "where the scene codes " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\{x_{i}\\}" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": ", prior parameters " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": ", and decoder parameters " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " are jointly optimized in a single training stage. This loss function consists of the rendering loss " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rend}}" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2418" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 69, + 515, + 202 + ], + "blocks": [ + { + "bbox": [ + 83, + 69, + 515, + 202 + ], + "lines": [ + { + "bbox": [ + 83, + 69, + 515, + 202 + ], + "spans": [ + { + "bbox": [ + 83, + 69, + 515, + 202 + ], + "type": "image", + "image_path": "8e145baf531a412b27e037624d3792ff4f3700cd97c4869b6aefbcae9804a2c1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 206, + 547, + 242 + ], + "lines": [ + { + "bbox": [ + 46, + 206, + 547, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 547, + 242 + ], + "type": "text", + "content": "Figure 3. An overview of SSDNeRF framework with a triplane NeRF representation. During training, we feed a batch of observations in the format of RGB values " + }, + { + "bbox": [ + 46, + 206, + 547, + 242 + ], + "type": "inline_equation", + "content": "y_{ij}^{\\mathrm{gt}}" + }, + { + "bbox": [ + 46, + 206, + 547, + 242 + ], + "type": "text", + "content": " and rays " + }, + { + "bbox": [ + 46, + 206, + 547, + 242 + ], + "type": "inline_equation", + "content": "r_{ij}^{\\mathrm{gt}}" + }, + { + "bbox": [ + 46, + 206, + 547, + 242 + ], + "type": "text", + "content": ". The corresponding scene code " + }, + { + "bbox": [ + 46, + 206, + 547, + 242 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 206, + 547, + 242 + ], + "type": "text", + "content": " is randomly initialized and optimized by minimizing the rendering loss " + }, + { + "bbox": [ + 46, + 206, + 547, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rend}}" + }, + { + "bbox": [ + 46, + 206, + 547, + 242 + ], + "type": "text", + "content": " and the diffusion loss " + }, + { + "bbox": [ + 46, + 206, + 547, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{diff}}" + }, + { + "bbox": [ + 46, + 206, + 547, + 242 + ], + "type": "text", + "content": ", and model parameters " + }, + { + "bbox": [ + 46, + 206, + 547, + 242 + ], + "type": "inline_equation", + "content": "\\phi, \\psi" + }, + { + "bbox": [ + 46, + 206, + 547, + 242 + ], + "type": "text", + "content": " are also updated along the way." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 260, + 288, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 260, + 288, + 320 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 288, + 320 + ], + "type": "text", + "content": "in Eq. (1) and a diffusion prior term in the form of NLL. Following [50, 54, 48], we replace the diffusion NLL with its approximate upper bound " + }, + { + "bbox": [ + 46, + 260, + 288, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{diff}}" + }, + { + "bbox": [ + 46, + 260, + 288, + 320 + ], + "type": "text", + "content": " in Eq. (2). This technique is also called score distillation in [33]. Adding empirical weighting factors, we finalize our training objective as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 327, + 287, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 327, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 63, + 327, + 287, + 340 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {\\text {r e n d}} \\mathcal {L} _ {\\text {r e n d}} (\\{x _ {i} \\}, \\psi) + \\lambda_ {\\text {d i f f}} \\mathcal {L} _ {\\text {d i f f}} (\\{x _ {i} \\}, \\phi). \\tag {4}", + "image_path": "b4f2497be30fd9f7e736fd7769fe8bf49296e55c22900690547c6a7a5ae0acbe.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 347, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 407 + ], + "type": "text", + "content": "Single-stage training constrains scene codes " + }, + { + "bbox": [ + 46, + 347, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\{x_{i}\\}" + }, + { + "bbox": [ + 46, + 347, + 287, + 407 + ], + "type": "text", + "content": " with both terms in the loss function, allowing the learned prior to complete the parts unseen to rendering. This is particularly beneficial to training on sparse-view data, where the expressive triplane codes are severely underdetermined." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 411, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 579 + ], + "type": "text", + "content": "Balancing Rendering and Prior Weights The render-to-prior weight ratio " + }, + { + "bbox": [ + 46, + 411, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{rend}} / \\lambda_{\\mathrm{diff}}" + }, + { + "bbox": [ + 46, + 411, + 287, + 579 + ], + "type": "text", + "content": " is crucial to single-stage training. To make hyperparameters more generalizable to different settings, we design an empirical weighting mechanism, in which the diffusion loss is normalized by the exponential moving average (EMA) of the scene codes' Frobenius norms, expressed as " + }, + { + "bbox": [ + 46, + 411, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{diff}} \\coloneqq c_{\\mathrm{diff}} / EMA\\left(\\|x_i\\|_F^2\\right)" + }, + { + "bbox": [ + 46, + 411, + 287, + 579 + ], + "type": "text", + "content": " with a constant scale " + }, + { + "bbox": [ + 46, + 411, + 287, + 579 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{diff}}" + }, + { + "bbox": [ + 46, + 411, + 287, + 579 + ], + "type": "text", + "content": ", and the rendering weight is determined by the number of views available " + }, + { + "bbox": [ + 46, + 411, + 287, + 579 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{v}}" + }, + { + "bbox": [ + 46, + 411, + 287, + 579 + ], + "type": "text", + "content": ", expressed as " + }, + { + "bbox": [ + 46, + 411, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{rend}} \\coloneqq c_{\\mathrm{rend}}(1 - e^{-0.1N_{\\mathrm{v}}}) / N_{\\mathrm{v}}" + }, + { + "bbox": [ + 46, + 411, + 287, + 579 + ], + "type": "text", + "content": " with a constant scale " + }, + { + "bbox": [ + 46, + 411, + 287, + 579 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{rend}}" + }, + { + "bbox": [ + 46, + 411, + 287, + 579 + ], + "type": "text", + "content": ". Intuitively, " + }, + { + "bbox": [ + 46, + 411, + 287, + 579 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{v}}" + }, + { + "bbox": [ + 46, + 411, + 287, + 579 + ], + "type": "text", + "content": "-based weighting is a calibration to the ray independence assumption in the decoder " + }, + { + "bbox": [ + 46, + 411, + 287, + 579 + ], + "type": "inline_equation", + "content": "p_{\\psi}(\\{y_j\\} | x, \\{r_j\\}) \\coloneqq \\prod_j \\mathcal{N}(y_j | y_{\\psi}(x, r_j), I)" + }, + { + "bbox": [ + 46, + 411, + 287, + 579 + ], + "type": "text", + "content": ", preventing the rendering loss from scaling linearly with the number of rays." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Comparison to Two-Stage Generative Neural Fields Previous two-stage methods [2, 12, 29, 44] ignore the prior term " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{diff}}\\mathcal{L}_{\\mathrm{diff}}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " during the first stage of training the auto-decoders. This can be seen as setting the render-to-prior weight ratio " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{rend}} / \\lambda_{\\mathrm{diff}}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " to infinity, resulting in biased and noisy scene codes " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": ". Shue et al. [44] partially mitigate this issue by imposing total variation (TV) regularization on triplane scene codes to enforce a smoothing prior, which resembles the LDM constraints on the latent space (mid column of Fig. 2). Control3Diff [15] proposes to learn a conditional diffusion model on data generated by a 3D GAN pre" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 260, + 545, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 260, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 304, + 260, + 545, + 296 + ], + "type": "text", + "content": "trained on single-view images. In contrast, our single-stage training aims to directly incorporate the diffusion prior to promote end-to-end coherence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 301, + 518, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 301, + 518, + 315 + ], + "spans": [ + { + "bbox": [ + 305, + 301, + 518, + 315 + ], + "type": "text", + "content": "4.2. Image-Guided Sampling and Finetuning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 319, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 380 + ], + "type": "text", + "content": "To achieve generalizable test-time NeRF reconstruction that covers a wide spectrum from single-view to dense observations, we propose performing image-guided sampling and then finetuning the sampled codes considering both the diffusion prior and rendering likelihood." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 380, + 545, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 416 + ], + "type": "text", + "content": "Following the reconstruction-guided sampling method by Ho et al. [20], we compute the approximated rendering gradients " + }, + { + "bbox": [ + 304, + 380, + 545, + 416 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 380, + 545, + 416 + ], + "type": "text", + "content": " w.r.t. a noisy code " + }, + { + "bbox": [ + 304, + 380, + 545, + 416 + ], + "type": "inline_equation", + "content": "x^{(t)}" + }, + { + "bbox": [ + 304, + 380, + 545, + 416 + ], + "type": "text", + "content": ", defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 419, + 545, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 419, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 419, + 545, + 459 + ], + "type": "interline_equation", + "content": "g \\leftarrow \\nabla_ {x ^ {(t)}} \\lambda_ {\\text {r e n d}} \\sum_ {j} \\frac {1}{2} \\left(\\frac {\\alpha^ {(t)}}{\\sigma^ {(t)}}\\right) ^ {2 \\omega} \\left\\| y _ {j} ^ {\\mathrm {g t}} - y _ {\\psi} \\left(\\hat {x} _ {\\phi} \\left(x ^ {(t)}, t\\right), r _ {j} ^ {\\mathrm {g t}}\\right) \\right\\| ^ {2}, \\tag {5}", + "image_path": "c1dce5eb07d8d1a475a41a47c5f03d61c3c3a5da666162aaf3553978ffd599f1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 459, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 459, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\left(\\alpha^{(t)} / \\sigma^{(t)}\\right)^{2\\omega}" + }, + { + "bbox": [ + 304, + 459, + 545, + 521 + ], + "type": "text", + "content": " is an additional weighting factor based on signal-to-noise ratio (SNR), with hyperparameter " + }, + { + "bbox": [ + 304, + 459, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 459, + 545, + 521 + ], + "type": "text", + "content": " chosen to be 0.5 or 0.25 in our work. The guidance gradients " + }, + { + "bbox": [ + 304, + 459, + 545, + 521 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 459, + 545, + 521 + ], + "type": "text", + "content": " are then combined with unconditional score prediction, expressed as a correction to the denoising output " + }, + { + "bbox": [ + 304, + 459, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 304, + 459, + 545, + 521 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 382, + 525, + 545, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 525, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 382, + 525, + 545, + 552 + ], + "type": "interline_equation", + "content": "\\hat {x} \\leftarrow \\hat {x} - \\lambda_ {\\mathrm {g d}} \\frac {\\sigma^ {(t) ^ {2}}}{\\alpha^ {(t)}} g \\tag {6}", + "image_path": "a3056c9e00823fe2b4923c1f4ad86b28fd0c77571b8a42a6ee343dc12c9d6271.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "text", + "content": "with guidance scale " + }, + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{gd}}" + }, + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "text", + "content": ". We adopt the predictor-corrector sampler [49] to solve " + }, + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "inline_equation", + "content": "x^{(0)}" + }, + { + "bbox": [ + 304, + 555, + 545, + 591 + ], + "type": "text", + "content": " by alternating between a DDIM step [47] and multiple Langevin correction steps." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 591, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 545, + 651 + ], + "type": "text", + "content": "We observe that the reconstruction guidance alone cannot strictly enforce rendering constraints towards faithful reconstruction. To address this issue, we reuse the training objective in Eq. (4) to finetune the sampled scene code " + }, + { + "bbox": [ + 304, + 591, + 545, + 651 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 591, + 545, + 651 + ], + "type": "text", + "content": ", while freezing the diffusion and decoder parameters:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 352, + 656, + 545, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 656, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 352, + 656, + 545, + 673 + ], + "type": "interline_equation", + "content": "\\min _ {x} \\lambda_ {\\text {r e n d}} \\mathcal {L} _ {\\text {r e n d}} (x) + \\lambda_ {\\text {d i f f}} ^ {\\prime} \\mathcal {L} _ {\\text {d i f f}} (x), \\tag {7}", + "image_path": "39d5d1ac5da959b7dd4409a0f3f9e3a20a3b8308fce3e1f9ac9c63f90ca4f299.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{diff}}^{\\prime}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " is the test-time prior weight, which we find should be lower than the training weight " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{diff}}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " for best results, as the prior learned from the training dataset is less" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2419" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "reliable when transferred to a different testing dataset. We use Adam [22] to optimize the code " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " for finetuning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 101, + 287, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 101, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 47, + 101, + 287, + 172 + ], + "type": "text", + "content": "Comparison to Previous NeRF Finetuning Approaches While finetuning with rendering loss is common in view-conditioned NeRF regression methods [8, 57], our finetuning approach differs in the use of diffusion prior loss on the 3D scene code, which significantly enhances generalization to novel views, as demonstrated in § 5.3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 180, + 180, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 180, + 180, + 193 + ], + "spans": [ + { + "bbox": [ + 47, + 180, + 180, + 193 + ], + "type": "text", + "content": "4.3. Implementation Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 198, + 287, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 198, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 287, + 223 + ], + "type": "text", + "content": "This subsection briefly describes some important technical details. More details can be found in the supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 227, + 287, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 227, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 227, + 287, + 383 + ], + "type": "text", + "content": "Prior Gradient Caching Triplane NeRF reconstruction requires at least hundreds of optimization iterations on each scene code " + }, + { + "bbox": [ + 47, + 227, + 287, + 383 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 47, + 227, + 287, + 383 + ], + "type": "text", + "content": ". A problem with the single-stage training loss in Eq. (4) is that the diffusion loss " + }, + { + "bbox": [ + 47, + 227, + 287, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{diff}}" + }, + { + "bbox": [ + 47, + 227, + 287, + 383 + ], + "type": "text", + "content": " requires much longer time to evaluate than the native NeRF rendering loss " + }, + { + "bbox": [ + 47, + 227, + 287, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rend}}" + }, + { + "bbox": [ + 47, + 227, + 287, + 383 + ], + "type": "text", + "content": ", reducing overall efficiency. To accelerate reconstruction in both training and test-time finetuning, we introduce a technique called prior gradient caching, which caches the back-propagated prior gradients " + }, + { + "bbox": [ + 47, + 227, + 287, + 383 + ], + "type": "inline_equation", + "content": "\\nabla_{x}\\lambda_{\\mathrm{diff}}\\mathcal{L}_{\\mathrm{diff}}" + }, + { + "bbox": [ + 47, + 227, + 287, + 383 + ], + "type": "text", + "content": " for re-use in multiple Adam steps, while refreshing the rendering gradients " + }, + { + "bbox": [ + 47, + 227, + 287, + 383 + ], + "type": "inline_equation", + "content": "\\nabla_{x}\\lambda_{\\mathrm{rend}}\\mathcal{L}_{\\mathrm{rend}}" + }, + { + "bbox": [ + 47, + 227, + 287, + 383 + ], + "type": "text", + "content": " in each of the steps, which allows for fewer diffusion passes than rendering. A training pseudocode is given in Algorithm 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 387, + 287, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 387, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 287, + 546 + ], + "type": "text", + "content": "Denoising Parameterization and Weighting The denoising model " + }, + { + "bbox": [ + 47, + 387, + 287, + 546 + ], + "type": "inline_equation", + "content": "\\hat{x}_{\\phi}(x^{(t)}, t)" + }, + { + "bbox": [ + 47, + 387, + 287, + 546 + ], + "type": "text", + "content": " is implemented as a U-Net [39] as in DDPM [19], with a total of 122M parameters. Its input and output are noisy and denoised triplane features, respectively, with channels of all three planes stacked together. For the prediction format, we adopt the " + }, + { + "bbox": [ + 47, + 387, + 287, + 546 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 47, + 387, + 287, + 546 + ], + "type": "text", + "content": "-parameterization " + }, + { + "bbox": [ + 47, + 387, + 287, + 546 + ], + "type": "inline_equation", + "content": "\\hat{v}_{\\phi}(x^{(t)}, t)" + }, + { + "bbox": [ + 47, + 387, + 287, + 546 + ], + "type": "text", + "content": " in [40], such that " + }, + { + "bbox": [ + 47, + 387, + 287, + 546 + ], + "type": "inline_equation", + "content": "\\hat{x} = \\alpha^{(t)} x^{(t)} - \\sigma^{(t)} \\hat{v}" + }, + { + "bbox": [ + 47, + 387, + 287, + 546 + ], + "type": "text", + "content": ". Regarding the weighting function " + }, + { + "bbox": [ + 47, + 387, + 287, + 546 + ], + "type": "inline_equation", + "content": "w^{(t)}" + }, + { + "bbox": [ + 47, + 387, + 287, + 546 + ], + "type": "text", + "content": " in the diffusion loss in Eq. (2), LSGM [50] employs two different mechanisms for optimizing latents " + }, + { + "bbox": [ + 47, + 387, + 287, + 546 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 47, + 387, + 287, + 546 + ], + "type": "text", + "content": " and diffusion weights " + }, + { + "bbox": [ + 47, + 387, + 287, + 546 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 47, + 387, + 287, + 546 + ], + "type": "text", + "content": ", respectively, which we find unstable with NeRF auto-decoders. Instead, we observe that the SNR-based weighting " + }, + { + "bbox": [ + 47, + 387, + 287, + 546 + ], + "type": "inline_equation", + "content": "w^{(t)} = \\left(\\alpha^{(t)} / \\sigma^{(t)}\\right)^{2\\omega}" + }, + { + "bbox": [ + 47, + 387, + 287, + 546 + ], + "type": "text", + "content": " used in Eq. (5) works well with our models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 555, + 128, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 128, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 128, + 569 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 575, + 110, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 110, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 110, + 586 + ], + "type": "text", + "content": "5.1. Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": "We conduct experiments on the ShapeNet SRN [6, 45] and Amazon Berkeley Objects (ABO) Tables [9] datasets for benchmarking with previous work. The SRN dataset provides single-object scenes in two categories, i.e., Cars and Chairs, with a train/test split of 2458/703 for Cars and 4612/1317 for Chairs. Each train scene has 50 random views from a sphere and each test scene has 251 spiral views from the upper hemisphere. The ABO Tables dataset provides a train/test split of 1520/156 table scenes, where each scene has 91 views from the upper hemisphere. For" + } + ] + } + ], + "index": 8 + }, + { + "type": "code", + "bbox": [ + 306, + 89, + 544, + 273 + ], + "blocks": [ + { + "bbox": [ + 312, + 75, + 520, + 87 + ], + "lines": [ + { + "bbox": [ + 312, + 75, + 520, + 87 + ], + "spans": [ + { + "bbox": [ + 312, + 75, + 520, + 87 + ], + "type": "text", + "content": "Algorithm 1: Single-stage diffusion NeRF training" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "lines": [ + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "spans": [ + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "inline_equation", + "content": "\\{y_{ij}^{\\mathrm{gt}},r_{ij}^{\\mathrm{gt}}\\}" + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "text", + "content": " \n1 Initialize " + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "inline_equation", + "content": "\\{x_i\\} ,\\phi ,\\psi" + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "text", + "content": " \n2 for " + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{out}}\\coloneqq 1\\dots K_{\\mathrm{out}}" + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "text", + "content": " do // outer loop of " + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{out}}" + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "text", + "content": " iterations \n3 Sample a batch of scenes " + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "inline_equation", + "content": "i\\in B_{\\mathrm{sc}}" + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "text", + "content": " \n4 " + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "inline_equation", + "content": "g_{\\phi},g_{x}^{\\mathrm{diff}}\\gets \\nabla_{\\phi ,\\{x_{i}\\}_{B_{\\mathrm{sc}}}}\\lambda_{\\mathrm{diff}}\\mathcal{L}_{\\mathrm{diff}}" + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "text", + "content": " // diffusion grad \n5 " + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "inline_equation", + "content": "\\phi \\leftarrow \\phi -Adam(g_{\\phi})" + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "text", + "content": " \n6 for " + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{in}}\\coloneqq 1\\dots K_{\\mathrm{in}}" + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "text", + "content": " do // inner loop of " + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{in}}" + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "text", + "content": " iterations \n7 Sample a batch of rays " + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "inline_equation", + "content": "j\\in B_{\\mathrm{ray}}" + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "text", + "content": " \n8 " + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "inline_equation", + "content": "g_{x}^{\\mathrm{rend}}\\gets \\nabla_{\\{x_i\\}_{B_{\\mathrm{sc}}}}\\lambda_{\\mathrm{rend}}\\mathcal{L}_{\\mathrm{rend}}" + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "text", + "content": " // rendering grad \n9 " + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "inline_equation", + "content": "g_{x}\\gets g_{x}^{\\mathrm{rend}} + g_{x}^{\\mathrm{diff}}" + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "text", + "content": " // add cached prior grad \n10 " + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "inline_equation", + "content": "\\{x_i\\}_{B_{\\mathrm{sc}}} \\leftarrow \\{x_i\\}_{B_{\\mathrm{sc}}} - Adam(g_x)" + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "text", + "content": " \n11 if " + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{in}} = K_{\\mathrm{in}}" + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "text", + "content": " then // last inner iteration \n12 " + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "inline_equation", + "content": "\\begin{array}{rl} & g_{\\psi}\\gets \\nabla_{\\psi}\\lambda_{\\mathrm{rend}}\\mathcal{L}_{\\mathrm{rend}}\\\\ & \\psi \\leftarrow \\psi -Adam(g_{\\psi}) \\end{array}" + }, + { + "bbox": [ + 306, + 89, + 544, + 273 + ], + "type": "text", + "content": " \n13" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_body" + } + ], + "index": 10, + "sub_type": "algorithm" + }, + { + "type": "table", + "bbox": [ + 312, + 293, + 543, + 395 + ], + "blocks": [ + { + "bbox": [ + 312, + 293, + 543, + 395 + ], + "lines": [ + { + "bbox": [ + 312, + 293, + 543, + 395 + ], + "spans": [ + { + "bbox": [ + 312, + 293, + 543, + 395 + ], + "type": "table", + "html": "
MethodTypeCarsTables
FID↓KID/10-3↓FID↓KID/10-3↓
Functa [12]LDM80.3---
π-GAN [4]GAN36.7†-41.67§13.82§
EG3D [5]GAN10.46*4.90*31.18§11.67§
DiffRF [29]LDM--27.0610.03
Ours (2-stage)LDM16.33±0.936.38±0.41--
Ours (1-stage)LDM11.08±1.113.47±0.2314.27±0.664.08±0.33
", + "image_path": "1b3e61d3ae55170fd3b258bf85ccd1420bb185c650921f09a7bdd36450bb4858.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 399, + 545, + 443 + ], + "lines": [ + { + "bbox": [ + 305, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 545, + 443 + ], + "type": "text", + "content": "Table 1. Unconditional generation results on SRN Cars and ABO Tables. † denotes results reported by Functa [12]. § denotes results reported by DiffRF [29]. * denotes results reproduced by us using the official public code with a bugfix. We show ±2σ intervals." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 462, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 462, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 305, + 462, + 545, + 486 + ], + "type": "text", + "content": "both datasets, we use the provided renderings (resized to " + }, + { + "bbox": [ + 305, + 462, + 545, + 486 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 305, + 462, + 545, + 486 + ], + "type": "text", + "content": ") with ground truth poses for training and testing." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 494, + 451, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 494, + 451, + 505 + ], + "spans": [ + { + "bbox": [ + 306, + 494, + 451, + 505 + ], + "type": "text", + "content": "5.2. Unconditional Generation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 512, + 545, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 512, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 305, + 512, + 545, + 583 + ], + "type": "text", + "content": "In this section, we conduct evaluations for unconditional generation using the SRN Cars and ABO Tables dataset. The Cars dataset poses a challenge in generating sharp and intricate textures, whereas the Tables dataset comprises of diverse geometries with realistic materials. Models are trained on all images of the training set for 1M iterations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 587, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 587, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 587, + 545, + 696 + ], + "type": "text", + "content": "Evaluation Protocol and Metrics For SRN Cars, following Functa [12], we sample 704 scenes from the diffusion model, and render each scene using the fixed 251 camera poses from the test set. For ABO Tables, following DiffRF [29], we sample 1000 scenes and render each scene with 10 random cameras. We adopt standard generation metrics including Fréchet Inception Distance (FID) [18] and Kernel Inception Distance (KID) [3]. The metrics' reference sets are all images in the test set for SRN Cars and all" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 515, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 515, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 515, + 712 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 315, + 702, + 515, + 712 + ], + "type": "text", + "content": "https://github.com/nvlabs/eg3d/issues/67" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2420" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 208, + 177 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 208, + 177 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 208, + 177 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 208, + 177 + ], + "type": "image", + "image_path": "2edc5250fd0c2cbf1e22213f5c01dceeb54f2123df734d67bc9fd201d6e45a25.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 178, + 156, + 187 + ], + "lines": [ + { + "bbox": [ + 99, + 178, + 156, + 187 + ], + "spans": [ + { + "bbox": [ + 99, + 178, + 156, + 187 + ], + "type": "text", + "content": "(a) EG3D Tables" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 221, + 70, + 374, + 175 + ], + "blocks": [ + { + "bbox": [ + 221, + 70, + 374, + 175 + ], + "lines": [ + { + "bbox": [ + 221, + 70, + 374, + 175 + ], + "spans": [ + { + "bbox": [ + 221, + 70, + 374, + 175 + ], + "type": "image", + "image_path": "c402b9128283d431983729c792e80d3f6820e86608bc750ac552a63ab4539da9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 177, + 328, + 187 + ], + "lines": [ + { + "bbox": [ + 268, + 177, + 328, + 187 + ], + "spans": [ + { + "bbox": [ + 268, + 177, + 328, + 187 + ], + "type": "text", + "content": "(b) DiffRF Tables" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 388, + 70, + 545, + 177 + ], + "blocks": [ + { + "bbox": [ + 388, + 70, + 545, + 177 + ], + "lines": [ + { + "bbox": [ + 388, + 70, + 545, + 177 + ], + "spans": [ + { + "bbox": [ + 388, + 70, + 545, + 177 + ], + "type": "image", + "image_path": "c784415f9ed0e3707e9c95fdd9dc45973dc1943923630e2974d84fbcedd6f1e4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 178, + 513, + 187 + ], + "lines": [ + { + "bbox": [ + 422, + 178, + 513, + 187 + ], + "spans": [ + { + "bbox": [ + 422, + 178, + 513, + 187 + ], + "type": "text", + "content": "(c) SSDNeRF Tables (ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 51, + 194, + 285, + 289 + ], + "blocks": [ + { + "bbox": [ + 51, + 194, + 285, + 289 + ], + "lines": [ + { + "bbox": [ + 51, + 194, + 285, + 289 + ], + "spans": [ + { + "bbox": [ + 51, + 194, + 285, + 289 + ], + "type": "image", + "image_path": "3722d95486aec137dd0b0e2ee8cc2a7dd306816752499b3d40b309e9eb27857f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 293, + 194, + 303 + ], + "lines": [ + { + "bbox": [ + 145, + 293, + 194, + 303 + ], + "spans": [ + { + "bbox": [ + 145, + 293, + 194, + 303 + ], + "type": "text", + "content": "(d) EG3D Cars" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 308, + 194, + 542, + 289 + ], + "blocks": [ + { + "bbox": [ + 308, + 194, + 542, + 289 + ], + "lines": [ + { + "bbox": [ + 308, + 194, + 542, + 289 + ], + "spans": [ + { + "bbox": [ + 308, + 194, + 542, + 289 + ], + "type": "image", + "image_path": "abd7f77808ff61a642e84969ed5f20c754a8b5922571cd827a8866b62685f627.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 293, + 468, + 303 + ], + "lines": [ + { + "bbox": [ + 383, + 293, + 468, + 303 + ], + "spans": [ + { + "bbox": [ + 383, + 293, + 468, + 303 + ], + "type": "text", + "content": "(e) SSDNeRF Cars (Ours)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 88, + 308, + 503, + 319 + ], + "lines": [ + { + "bbox": [ + 88, + 308, + 503, + 319 + ], + "spans": [ + { + "bbox": [ + 88, + 308, + 503, + 319 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison between unconditional generative models trained on ABO Tables and SRN Cars." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 52, + 328, + 544, + 427 + ], + "blocks": [ + { + "bbox": [ + 52, + 328, + 544, + 427 + ], + "lines": [ + { + "bbox": [ + 52, + 328, + 544, + 427 + ], + "spans": [ + { + "bbox": [ + 52, + 328, + 544, + 427 + ], + "type": "table", + "html": "
MethodCars 1-viewCars 2-viewChairs 1-viewChairs 2-view
PSNR↑SSIM↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓FID↓
3DiM [53]21.010.57-8.99----17.050.53-6.57----
PixelNeRF [55]23.170.900.146‡59.24†25.660.94--23.720.910.128‡38.49†26.200.94--
SRN [45]22.25§0.89§0.129‡41.21†24.84§0.92§--22.89§0.89§0.104‡26.51†24.48§0.92§--
CodeNeRF [21]23.800.910.118*56.34*25.710.930.108*56.13*23.660.900.106*31.65*25.630.910.097*29.90*
VisionNeRF [25]22.880.910.08421.31†----24.480.930.07710.05†----
Ours (1-stage)23.520.910.07816.3926.490.940.05410.6624.350.930.06710.1326.940.950.05510.85
", + "image_path": "4e9d518a2993259e53992a6380787e2dcd7e5ce029a159483d51f2da5ec9cac2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 431, + 546, + 475 + ], + "lines": [ + { + "bbox": [ + 46, + 431, + 546, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 546, + 475 + ], + "type": "text", + "content": "Table 2. Single-view and two-view reconstruction results on SRN Cars and Chairs. For consistency with prior work, we use view #64 of the test scene as single-view input and view #64 and #104 as two-view input. † denotes results reported by 3DiM [53]. ‡ denotes results reported by VisionNeRF [25], § denotes results reported by PixelNeRF [21], * denotes results reproduced by us using the official code. - indicates results are unavailable." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 494, + 279, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 279, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 279, + 506 + ], + "type": "text", + "content": "images in the entire dataset for ABO Tables, respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 511, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 511, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 511, + 287, + 594 + ], + "type": "text", + "content": "Comparison to the State of the Art As shown in Table 1, on SRN Cars, SSDNeRF (1-stage) outperforms EG3D in KID (a more suitable measure for small datasets) by a clear margin. Meanwhile, its FID is drastically better than Functa, which uses an LDM but with low dimensional latent codes. On ABO Tables, SSDNeRF shows significantly better performance than EG3D and DiffRF." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 600, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 600, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 600, + 287, + 661 + ], + "type": "text", + "content": "Single- vs. Two-stage On SRN Cars, we compare the proposed single-stage training against two-stage training with tuned TV regularization using the same model architecture. The results in Table 1 indicate substantial advantage of single-stage training (KID/10" + }, + { + "bbox": [ + 46, + 600, + 287, + 661 + ], + "type": "inline_equation", + "content": "^{-3}" + }, + { + "bbox": [ + 46, + 600, + 287, + 661 + ], + "type": "text", + "content": " 3.47 vs. 6.38)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Qualitative Results As shown in Fig. 4, SSDNeRF generates more regular geometries than the slightly skewed and distorted shapes by EG3D [5]. Compared to DiffRF [29], our method produces sharp details and reflective materials," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 494, + 546, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 546, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 546, + 517 + ], + "type": "text", + "content": "thanks to our more expressive model with latents of higher spatial resolution and view-dependent NeRF decoder." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 523, + 492, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 523, + 492, + 536 + ], + "spans": [ + { + "bbox": [ + 305, + 523, + 492, + 536 + ], + "type": "text", + "content": "5.3. Sparse-View NeRF Reconstruction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 542, + 546, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 546, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 546, + 650 + ], + "type": "text", + "content": "This section presents experiments on 3D reconstruction from sparse-view images of unseen objects in SRN Cars and Chairs test sets. The Cars dataset presents the challenge of recovering distinct textures, while the Chairs dataset requires accurate reconstruction of diverse shapes. Models are trained on all images of the training set for 80K iterations, as we find that longer schedule leads to decaying performance in reconstructing unseen objects. This behaviour is in accordance with the interpolation results in § 5.5." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "text", + "content": "Evaluation Protocol and Metrics We use the evaluation protocol and metrics in PixelNeRF [55]. Given input images sampled from each test scene, we obtain the triplane scene code via guidance-finetuning and evaluate novel view synthesis quality with respect to the unseen images." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2421" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 68, + 298, + 269 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 298, + 269 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 298, + 269 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 298, + 269 + ], + "type": "image", + "image_path": "49d6562de1a695052892f6a857e6a41805f3310a43157d3f04148967a61f6fb8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 273, + 512, + 285 + ], + "lines": [ + { + "bbox": [ + 79, + 273, + 512, + 285 + ], + "spans": [ + { + "bbox": [ + 79, + 273, + 512, + 285 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of single-view reconstruction methods on unseen test objects in SRN Cars and Chairs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 302, + 68, + 547, + 269 + ], + "blocks": [ + { + "bbox": [ + 302, + 68, + 547, + 269 + ], + "lines": [ + { + "bbox": [ + 302, + 68, + 547, + 269 + ], + "spans": [ + { + "bbox": [ + 302, + 68, + 547, + 269 + ], + "type": "image", + "image_path": "9d4af214eaa886395617b5124a9cbb17b7208255d6465580ae0cc20b603d9aae.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 304, + 287, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 364 + ], + "type": "text", + "content": "The image quality metrics include average peak signal-to-noise-ratio (PSNR), structural similarity (SSIM) [52], and Learned Perceptual Image Patch Similarity (LPIPS) [56]. In addition, we evaluate the FID between all synthesized images and ground truth images as in 3DiM [53]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 369, + 288, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 288, + 548 + ], + "type": "text", + "content": "Comparison to the State of the Art Table 2 compares SSDNeRF against previous approaches in single-view and two-view reconstruction settings. Overall, SSDNeRF reaches the best LPIPS of all tasks, indicating the best perceptual fidelity. In contrast, 3DiM generates high quality images (best FID) but with the lowest fidelity to the ground truth (lowest PSNR); CodeNeRF reports the best PSNR on single-view Cars, but its limited expressiveness leads to blurry outputs (Fig. 5) and less competitive LPIPS and FID; VisionNeRF achieves a balanced performance on all single-view metrics, but may struggle to generate textural details on the unseen side of cars (e.g., the other side of the ambulance in Fig. 5). Moreover, SSDNeRF exhibits a clear advantage in two-view reconstruction, achieving the best performance on all relevant metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 553, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 287, + 601 + ], + "type": "text", + "content": "Single- vs. Two-stage As demonstrated in Table 3, the model trained in a single stage (A0) outperforms the same architecture trained in two stages with TV regularization (A1) in all metrics of single-view reconstruction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Ablation Studies on Test-Time Finetuning As shown in Table 3, we evaluate the effectiveness of test-time finetuning and the contribution of the learned diffusion prior with two ablation experiments: (A2) removing the diffusion loss during finetuning and using only the rendering loss, and (A3) omitting the finetuning process entirely. The results indicate that finetuning with single-view rendering loss provides only marginal improvements over guided sampling (A2 vs. A3), while the learned diffusion prior significantly" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 320, + 297, + 533, + 360 + ], + "blocks": [ + { + "bbox": [ + 320, + 297, + 533, + 360 + ], + "lines": [ + { + "bbox": [ + 320, + 297, + 533, + 360 + ], + "spans": [ + { + "bbox": [ + 320, + 297, + 533, + 360 + ], + "type": "table", + "html": "
IDTrainingFinetuningPSNR↑SSIM↑LPIPS↓FID↓
A01-stageRend + Diff23.520.9130.07816.39
A12-stageRend + Diff22.830.9060.09020.97
A21-stageRend23.130.9070.08827.93
A31-stageNone23.070.9050.09230.95
", + "image_path": "5eade70fbe34051b3c7f33d7cffe72bdfc360e2696f6478c67cdf039eb2eb663.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 365, + 545, + 385 + ], + "lines": [ + { + "bbox": [ + 306, + 365, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 545, + 385 + ], + "type": "text", + "content": "Table 3. Ablation results on single-view reconstruction of SRN Cars." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 336, + 389, + 515, + 473 + ], + "blocks": [ + { + "bbox": [ + 336, + 389, + 515, + 473 + ], + "lines": [ + { + "bbox": [ + 336, + 389, + 515, + 473 + ], + "spans": [ + { + "bbox": [ + 336, + 389, + 515, + 473 + ], + "type": "image", + "image_path": "6be3c0ffb17e1dfb41f87f2dc35db23292d7a4fab94c65b9ab3a7c2097c00cd5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 477, + 545, + 510 + ], + "lines": [ + { + "bbox": [ + 305, + 477, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 477, + 545, + 510 + ], + "type": "text", + "content": "Figure 6. LPIPS scores (lower is better) of novel view synthesis from sparse-to-dense inputs, evaluated on SRN Cars test set. The triplane baselines adopt mean initialization for better performance." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 529, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 545, + 601 + ], + "type": "text", + "content": "boosts the LPIPS and FID scores (A0 vs. A2), highlighting its importance in recovering sharp and distinct contents. Moreover, the qualitative results in Fig. 5 reveal that views with higher overlap to the input view benefit the most from finetuning, meeting our expectation that finetuning helps faithfully reconstruct the exact observations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Sparse-to-Dense Reconstruction To validate that SSDNeRF seamlessly bridges sparse- and dense-view NeRF reconstruction, we evaluate its novel view synthesis performance with the number of input views varying from 1 to 32. We compare our model to the triplane NeRF baseline trained as an auto-decoder with optional TV regularization instead of diffusion prior. Meanwhile, we also evaluate CodeNeRF [21], an auto-decoder with 256-d latent codes. The results in Fig. 6 show that SSDNeRF excels in all settings," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2422" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "especially in 1 to 4 views. In contrast, CodeNeRF is outperformed by vanilla triplane NeRF with more views." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 102, + 275, + 115 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 102, + 275, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 275, + 115 + ], + "type": "text", + "content": "5.4. Training SSDNeRF on Sparse-View Dataset" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 121, + 287, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 287, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 287, + 193 + ], + "type": "text", + "content": "In this section, we train SSDNeRF on a sparse-view subset of the full SRN Cars training set, in which a fixed set of only three views are randomly picked from each scene. Note that a reasonable decline in performance compared to dense-view training is expected as the whole training dataset has been reduced to " + }, + { + "bbox": [ + 46, + 121, + 287, + 193 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 46, + 121, + 287, + 193 + ], + "type": "text", + "content": " of its original size." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 197, + 287, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 197, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 197, + 287, + 280 + ], + "type": "text", + "content": "Unconditional Generation We adopt a training trick that resets the triplane codes to their mean value halfway through training. This helps to prevent the model from getting stuck in a local minimum that overfits geometric artifacts. We also double the length of the training schedule accordingly. The model achieves a decent FID of " + }, + { + "bbox": [ + 46, + 197, + 287, + 280 + ], + "type": "inline_equation", + "content": "19.04 \\pm 1.10" + }, + { + "bbox": [ + 46, + 197, + 287, + 280 + ], + "type": "text", + "content": " and a " + }, + { + "bbox": [ + 46, + 197, + 287, + 280 + ], + "type": "inline_equation", + "content": "\\mathrm{KID} / 10^{-3}" + }, + { + "bbox": [ + 46, + 197, + 287, + 280 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 197, + 287, + 280 + ], + "type": "inline_equation", + "content": "8.28 \\pm 0.60" + }, + { + "bbox": [ + 46, + 197, + 287, + 280 + ], + "type": "text", + "content": ". Results are visualized in Fig. 7." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 284, + 287, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 287, + 344 + ], + "type": "text", + "content": "Single-View Reconstruction We adopt the same training strategy as in § 5.3. With our guidance-finetuning approach, the model achieves an LPIPS score of 0.106, even outperforming most of the previous methods in Table 2 that use the full training set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "type": "text", + "content": "Comparison to TV Regularization Fig. 8 (b) shows the RGB images and geometries represented by the scene latent codes learned from three views during training. By comparison, vanilla triplane auto-decoder with TV regularization (Fig. 8 (a)) often fails to reconstruct a scene from sparse views, leading to severe geometric artifacts. As a result, previously it has been infeasible to train two-stage models with expressive latents on sparse-view data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 449, + 162, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 162, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 162, + 462 + ], + "type": "text", + "content": "5.5. NeRF Interpolation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 468, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 624 + ], + "type": "text", + "content": "Following DDIM [47], we can sample two initial values " + }, + { + "bbox": [ + 46, + 468, + 287, + 624 + ], + "type": "inline_equation", + "content": "x^{(T)} \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 46, + 468, + 287, + 624 + ], + "type": "text", + "content": ", interpolate them using spherical linear interpolation [43], and then use the deterministic solver to obtain interpolated samples. However, as noted by [34, 37], standard Gaussian diffusion models often result in nonsmooth interpolation. In SSDNeRF (with results shown in Fig. 9), we find that the model (a) trained with early stopping for sparse-view reconstruction produces reasonably smooth transitions between samples, while the model (b) trained with a longer schedule for unconditional generation produces distinct yet discontinuous samples. This suggests that early stopping preserves a smoother prior, leading to better generalization for sparse-view reconstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 634, + 119, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 119, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 119, + 645 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": "In this paper, we propose SSDNeRF, which combines the diffusion model and NeRF representation through a novel single-stage training paradigm with an end-to-end justifiable loss. Notably, it overcomes the limitations in previous work where implicit neural fields must be obtained from" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 541, + 116 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 541, + 116 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 541, + 116 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 541, + 116 + ], + "type": "image", + "image_path": "3d9b02bc5936756dd254acea334242c101e7eee1d099efa483db6bdf6f7315ff.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 121, + 545, + 143 + ], + "lines": [ + { + "bbox": [ + 306, + 121, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 306, + 121, + 545, + 143 + ], + "type": "text", + "content": "Figure 7. Images generated by SSDNeRF trained on a 3-view subset of SRN Cars training set." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 311, + 155, + 416, + 210 + ], + "blocks": [ + { + "bbox": [ + 311, + 155, + 416, + 210 + ], + "lines": [ + { + "bbox": [ + 311, + 155, + 416, + 210 + ], + "spans": [ + { + "bbox": [ + 311, + 155, + 416, + 210 + ], + "type": "image", + "image_path": "6528710554a00ad0a8d10d9183222b477c74e7c9556896a9090b92c9b4ff213d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 213, + 412, + 223 + ], + "lines": [ + { + "bbox": [ + 316, + 213, + 412, + 223 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 412, + 223 + ], + "type": "text", + "content": "(a) Auto-decoder training w/ TV" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 432, + 163, + 541, + 210 + ], + "blocks": [ + { + "bbox": [ + 432, + 163, + 541, + 210 + ], + "lines": [ + { + "bbox": [ + 432, + 163, + 541, + 210 + ], + "spans": [ + { + "bbox": [ + 432, + 163, + 541, + 210 + ], + "type": "image", + "image_path": "f229c7790d811fd3eb42ec3ec9a2b8f1224523d1fbe345e9855e7ddf5ca49142.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 213, + 545, + 223 + ], + "lines": [ + { + "bbox": [ + 430, + 213, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 430, + 213, + 545, + 223 + ], + "type": "text", + "content": "(b) 1-stage joint training with diffusion" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 271, + 544, + 312 + ], + "blocks": [ + { + "bbox": [ + 305, + 228, + 545, + 261 + ], + "lines": [ + { + "bbox": [ + 305, + 228, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 545, + 261 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison between scene codes learned from 3 views by (a) triplane auto-decoder with TV regularization vs. (b) single-stage diffusion NeRF." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 271, + 544, + 312 + ], + "lines": [ + { + "bbox": [ + 307, + 271, + 544, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 271, + 544, + 312 + ], + "type": "image", + "image_path": "d7a5022e3f655c21a283a26ff6eeb65cfcd0ec2a845f2ff9dcab5c6b464ceb80.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 314, + 542, + 323 + ], + "lines": [ + { + "bbox": [ + 312, + 314, + 542, + 323 + ], + "spans": [ + { + "bbox": [ + 312, + 314, + 542, + 323 + ], + "type": "text", + "content": "(a) Models trained with early stopping (80K iters) for sparse-view reconstruction" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 307, + 330, + 543, + 370 + ], + "blocks": [ + { + "bbox": [ + 307, + 330, + 543, + 370 + ], + "lines": [ + { + "bbox": [ + 307, + 330, + 543, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 330, + 543, + 370 + ], + "type": "image", + "image_path": "9948b75b0a03f87389db5a7e462e11c4c2badd23c3f280bb2db8b73ecc382d61.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 372, + 538, + 380 + ], + "lines": [ + { + "bbox": [ + 315, + 372, + 538, + 380 + ], + "spans": [ + { + "bbox": [ + 315, + 372, + 538, + 380 + ], + "type": "text", + "content": "(b) Models trained with long schedule (1M iters) for unconditional generation" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 386, + 545, + 407 + ], + "lines": [ + { + "bbox": [ + 305, + 386, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 386, + 545, + 407 + ], + "type": "text", + "content": "Figure 9. Interpolation between the leftmost and rightmost samples using DDIM [47]." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 429, + 545, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 545, + 488 + ], + "type": "text", + "content": "dense observations first, before training the diffusion models to learn their manifold. With strong performance on multiple benchmarks, SSDNeRF demonstrates a significant advancement towards a unified framework for general 3D content manipulation." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 493, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 545, + 589 + ], + "type": "text", + "content": "Limitations and Future Work Currently, our method relies on ground truth camera parameters during both training and testing. Future work may explore transform-invariant models. Additionally, the diffusion prior can become discontinuous with prolonged training, which affects generalization. Although early stopping is temporarily used, a better network design or a larger training dataset may be able to address this problem fundamentally." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 594, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 712 + ], + "type": "text", + "content": "Acknowledgements We thank Norman Müller for sharing the baseline results on ABO Tables. Hansheng Chen and Wei Tian acknowledge the funding by the National Natural Science Foundation of China (No. 52002285), the Shanghai Science and Technology Commission (No. 21ZR1467400), the original research project of Tongji University (No. 22120220593), the National Key R&D Program of China (No. 2021YFB2501104), and the Natural Science Foundation of Chongqing (No. 2023NSCQ-MSX4511)." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2423" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J. Mitra, and Paul Guerrero. RenderDiffusion: Image diffusion for 3D reconstruction, inpainting and generation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "text", + "content": "[2] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, Afshin Dehghan, and Josh Susskind. Gaudi: A neural architect for immersive 3d scene generation. In NeurIPS, 2022. 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 287, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 225 + ], + "type": "text", + "content": "[3] Mikołaj Binkowski, Danica J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In ICLR, 2018. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 287, + 270 + ], + "type": "text", + "content": "[4] Eric Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 287, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 287, + 328 + ], + "type": "text", + "content": "[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 330, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 330, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 330, + 287, + 407 + ], + "type": "text", + "content": "[6] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 408, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 287, + 441 + ], + "type": "text", + "content": "[7] Anpei Chen, Zexiang Xu, Xinyue Wei, Siyu Tang, Hao Su, and Andreas Geiger. Factor fields: A unified framework for neural fields and beyond, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 287, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 287, + 486 + ], + "type": "text", + "content": "[8] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, pages 14124-14133, 2021. 1, 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 553 + ], + "type": "text", + "content": "[9] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik. Abo: Dataset and benchmarks for real-world 3d object understanding. In CVPR, 2022. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "text", + "content": "[10] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In CVPR, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "type": "text", + "content": "[11] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W. Taylor, and Joshua M. Susskind. Unconstrained scene generation with locally conditioned radiance fields. In ICCV, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "text", + "content": "[12] Emilien Dupont, Hyunjik Kim, S. M. Ali Eslami, Danilo Jimenez Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you can treat it like one. In ICML, 2022. 2, 3, 4, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[13] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In NeurIPS, 2022. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "text", + "content": "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 140, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 174 + ], + "type": "text", + "content": "[15] Jiatao Gu, Qingzhe Gao, Shuangfei Zhai, Baoquan Chen, Lingjie Liu, and Josh Susskind. Learning controllable 3d diffusion models from single-view images, 2023. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "text", + "content": "[16] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In ICLR, 2022. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "text", + "content": "[17] Jiatao Gu, Alex Trevithick, Kai-En Lin, Josh Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nerf-guided distillation from 3d-aware diffusion, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 253, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 297 + ], + "type": "text", + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 298, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 319 + ], + "type": "text", + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "type": "text", + "content": "[20] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 354, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 388 + ], + "type": "text", + "content": "[21] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In ICCV, pages 12949-12958, 2021. 2, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 388, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 409 + ], + "type": "text", + "content": "[22] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 411, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 432 + ], + "type": "text", + "content": "[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 434, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 466 + ], + "type": "text", + "content": "[24] Yiyi Liao, Katja Schwarz, Lars Mescheder, and Andreas Geiger. Towards unsupervised learning of generative models for 3d controllable image synthesis. In CVPR, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 468, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 511 + ], + "type": "text", + "content": "[25] Kai-En Lin, Lin Yen-Chen, Wei-Sheng Lai, Tsung-Yi Lin, Yi-Chang Shih, and Ravi Ramamoorthi. Vision transformer for nef-based view synthesis from a single input image. In WACV, 2023. 1, 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 512, + 545, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 556 + ], + "type": "text", + "content": "[26] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11461-11471, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "type": "text", + "content": "[27] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "text", + "content": "[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[29] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In CVPR, 2023. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[30] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2424" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "learning of 3d representations from natural images. In ICCV, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[31] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "type": "text", + "content": "[32] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 176, + 287, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 176, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 176, + 287, + 208 + ], + "type": "text", + "content": "[33] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 210, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 210, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 287, + 253 + ], + "type": "text", + "content": "[34] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In CVPR, 2022. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 256, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 256, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 48, + 256, + 287, + 289 + ], + "type": "text", + "content": "[35] Daniel Rebain, Mark Matthews, Kwang Moo Yi, Dmitry Lagun, and Andrea Tagliasacchi. Lolnerf: Learn from one look. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 291, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 291, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 48, + 291, + 287, + 334 + ], + "type": "text", + "content": "[36] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In ICML, pages 1278-1286, 2014. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 335, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 335, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 48, + 335, + 287, + 367 + ], + "type": "text", + "content": "[37] Severi Rissanen, Markus Heinonen, and Arno Solin. Generative modelling with inverse heat dissipation. In ICLR, 2023. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 370, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 370, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 370, + 287, + 403 + ], + "type": "text", + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 404, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 449 + ], + "type": "text", + "content": "[39] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241, 2015. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 450, + 287, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 450, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 48, + 450, + 287, + 472 + ], + "type": "text", + "content": "[40] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In ICLR, 2022. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 474, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 474, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 48, + 474, + 287, + 506 + ], + "type": "text", + "content": "[41] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 508, + 287, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 508, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 48, + 508, + 287, + 541 + ], + "type": "text", + "content": "[42] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 543, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 576 + ], + "type": "text", + "content": "[43] Ken Shoemake. Animating rotation with quaternion curves. In Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 245-254, 1985. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 577, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 577, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 577, + 287, + 609 + ], + "type": "text", + "content": "[44] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In CVPR, 2023. 2, 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 611, + 287, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 287, + 654 + ], + "type": "text", + "content": "[45] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In NeurIPS, 2019. 2, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "type": "text", + "content": "[46] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[47] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2, 3, 4, 8" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 477 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[48] Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. Maximum likelihood training of score-based diffusion models. In NeurIPS, 2021. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[49] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "text", + "content": "[50] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In NeurlPS, 2021. 2, 3, 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "type": "text", + "content": "[51] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In CVPR, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "text", + "content": "[52] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "text", + "content": "[53] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. In ICLR, 2023. 2, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 320, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 342 + ], + "type": "text", + "content": "[54] Antoine Wehenkel and Gilles Louppe. Diffusion priors in variational autoencoders. In ICML Workshops, 2021. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 344, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 344, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 344, + 545, + 375 + ], + "type": "text", + "content": "[55] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 1, 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 376, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 410 + ], + "type": "text", + "content": "[56] Richard Zhang, Phillip Isola, Alexei Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "type": "text", + "content": "[57] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In CVPR, 2022. 2, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 445, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 477 + ], + "type": "text", + "content": "[58] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In CVPR, 2023. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2425" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Size Does Matter_ Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network/2e5e4941-f171-4e8c-be33-35affe248d0e_content_list.json b/2023/Size Does Matter_ Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network/2e5e4941-f171-4e8c-be33-35affe248d0e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fd79a0b9ace5b4cbc101bf0597b74561d0a5adf8 --- /dev/null +++ b/2023/Size Does Matter_ Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network/2e5e4941-f171-4e8c-be33-35affe248d0e_content_list.json @@ -0,0 +1,1222 @@ +[ + { + "type": "text", + "text": "Size Does Matter: Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network", + "text_level": 1, + "bbox": [ + 148, + 130, + 821, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chieh-Yun Chen $^{1,2*}$ Yi-Chung Chen $^{1,3*}$ Hong-Han Shuai $^{2}$ Wen-Huang Cheng $^{3}$", + "bbox": [ + 143, + 202, + 825, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Stylins.ai $^{2}$ National Yang Ming Chiao Tung University $^{3}$ National Taiwan University", + "bbox": [ + 148, + 223, + 818, + 243 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/82f50f4453d85f38c84e709171b798272971efcaee9572fd1b7a2f92430550ae.jpg", + "image_caption": [ + "Figure 1. Multi-garment virtual try-on results with different sizes (tucked or untucked). $r$ is the ratio of torso length over shoulder width, which can be decided by customers in online shopping scenarios. Padding blue backgrounds for better visualizing clothing skeletons." + ], + "image_footnote": [], + "bbox": [ + 102, + 266, + 870, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 515, + 313, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Virtual try-on tasks aim at synthesizing realistic try-on results by trying target clothes on humans. Most previous works relied on the Thin Plate Spline or appearance flows to warp clothes to fit human body shapes. However, both approaches cannot handle complex warping, leading to over distortion or misalignment. Furthermore, there is a critical unaddressed challenge of adjusting clothing sizes for try-on. To tackle these issues, we propose a Clothing-Oriented Transformation Try-On Network (COTTON). COTTON leverages clothing structure with landmarks and segmentation to design a novel landmark-guided transformation for precisely deforming clothes, allowing for size adjustment during try-on. Additionally, to properly remove the clothing region from the human image without losing significant human characteristics, we propose a clothing elimination policy based on both transformed clothes and human segmentation. This method enables users to try on clothes tucked-in or untucked while retaining more human characteristics. Both qualitative and quantitative results show that COTTON outperforms the state-of-the-art high-resolution virtual try-on approaches. All the code is available at https://github.com/cotton6/COTTON-size-does-matter.", + "bbox": [ + 73, + 537, + 473, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 515, + 630, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image-based virtual try-on replaces clothing items on a person with the desired ones, creating realistic try-on results and lowering costs associated with on-model photos for the e-commerce industry1. Additionally, it enables customers to use virtual dressing rooms when shopping online, potentially enhancing the e-commerce experience and increasing conversion rates2. As online shopping becomes popular, the virtual try-on tasks have received more and more attention [2,4,5,9,13,17,23,28,29,33,35]. For example, to extend the virtual try-on resolution from low $(256\\times 192)$ to high $(1024\\times 768)$ , the most obvious difficulty is the misalignment issue. [5] proposed the alignment-aware segment normalization to remove the misleading information in the misaligned area. Additionally, [23] further performs appearance flow-based warping and segmentation map generation simultaneously to tackle misalignment.", + "bbox": [ + 496, + 542, + 892, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the considerable progress made in previous works, there are still some challenges that have not been adequately addressed, as outlined below. i) Handling complex warping without misalignment: General image", + "bbox": [ + 496, + 786, + 893, + 848 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 517, + 862, + 779, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://www.zmo.ai/aimodels/", + "bbox": [ + 519, + 875, + 746, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "2https://www.revery.ai", + "bbox": [ + 519, + 887, + 691, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7513", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "based virtual try-on frameworks [5, 16, 21, 28, 31] typically employ a clothing deformation module, such as the Thin Plate Spline (TPS) method, to align clothing images with the target human pose. However, research [6, 15, 23] has found that TPS may not effectively handle complex warping when different garment regions require different deformations. In attempts to improve warping results, they have replaced TPS with dense appearance flows. However, when the transformation between the garment and corresponding body parts is significant (as illustrated in case II in Fig. 4), the performance of both TPS (VITON-HD [5]) and flow-based (HR-VITON [23]) methods deteriorates drastically. Hence, existing methods still can't properly address the challenge of clothing deformation. ii) Adjusting clothing sizes: Previous works predicted the try-on segmentation conditioned on the clothing image by only considering the shape of the clothes without the scale information. Thus, given a clothing image, it is impossible for previous works to change the clothing size. This limitation severely restricts the practical application of virtual try-on because people often try on different sizes in the fitting room. iii) Appropriate clothing elimination policy: Previous works [5, 23, 28] rely on segmentation maps to eliminate the original clothing region and utilize the remaining area as a guide for trying on target clothes. However, this method can lead to either excessive or insufficient removal of the original image. For instance, when prior works try on upper clothes, the lower part of the upper clothes remains intact, which limits the length of the upper clothes, as shown in Fig. 6. The lower part of the upper clothing is forced to be tucked in, which reduces the model's ability to generate complete upper clothes, thereby reducing the practicality of virtual try-on. Additionally, previous works remove the entire arm of input human images, which eliminates essential human characteristics such as tattoos and arm width that should be preserved in the final output, as illustrated in case I of Fig. 5. Therefore, developing an appropriate clothing elimination strategy that preserves crucial information while removing the clothing region is a significant challenge.", + "bbox": [ + 76, + 90, + 472, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose a simple yet powerful approach called Clothing-Oriented Transformation Try-On Network (COTTON). Specifically, COTTON first exploits geometric information with Clothing Landmark Predictor and Clothing Segmentation Network to predict clothing landmarks and segmentation masks respectively. To overcome the first challenge posed by complex warping, we propose the Landmark-guided Transformation that first separates clothes into sub-parts via the segmentation mask and then uses clothing landmarks to estimate homography matrices that fit these sub-parts to the target human pose. To solve the second challenge, we introduce a clothing landmarks adjustment approach that allows users to change clothing sizes. Adjustment of landmarks will al", + "bbox": [ + 75, + 688, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ter the homography matrices and then cause clothing size changes, as demonstrated in Fig. 1, which remarkably enhances the practicality of virtual try-on. To address the third challenge, we employ the transformed clothing images to identify and remove the areas that would be covered by the clothes. This enables our proposed Clothing Elimination Policy to effectively eliminate the clothing region while preserving important details and offer flexibility in tucking the clothes or not, as shown in Fig. 6. Extensive experiments show that COTTON achieves superior results than state-of-the-arts both quantitatively and qualitatively. We summarize our contributions as follows:", + "bbox": [ + 496, + 90, + 893, + 272 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a Clothing-Oriented Transformation Try-On Network (COTTON), which improves clothing transformation quality and addresses complex warping misalignments by leveraging clothing geometry.", + "- We introduce adjustable clothing landmarks to provide clothing size information based only on images. To the best of our knowledge, this is the first work enabling 2D-based virtual try-on with different clothing sizes for approaching real-world try-on.", + "- To preserve critical information, we propose a Clothing Elimination Policy to properly remove clothing information while retaining valuable human characteristics, and offer flexibility in tucking the clothes or not.", + "- Extensive experiments on the Dress Code dataset [28] show that our model significantly outperforms SOTAs, e.g., at least $41.1\\%$ improvement in terms of FID." + ], + "bbox": [ + 517, + 282, + 890, + 553 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 565, + 650, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Clothing Deformation", + "text_level": 1, + "bbox": [ + 500, + 590, + 705, + 606 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Virtual try-on requires clothing deformation to align patterns and preserve clothing details. Thin Plate Spline (TPS) warping is widely used in virtual try-on tasks [5, 16, 21, 28, 31], but it has limited ability to model geometric changes and result in unnatural deformations. Thus, [2, 6, 11, 13, 15, 17, 23] conducted appearance flows as an alternative to TPS, but misalignment still has not be well addressed. To address misalignment, [5] improved TPS by eliminating irrelevant clothing texture information in misaligned regions. Meanwhile, [23] improved the deformation by designing pathways to jointly predict appearance flows and segmentation features. However, neither TPS nor appearance flow can deform partial regions of clothing without affecting other parts, e.g., warping the sleeves to present arm akimbo without affecting the torso part of the clothes, since they do not have the structural information of the clothes. To address this issue, [33] introduced normalized patches to learn spatial-agnostic clothing features but this approach may destroy the completeness of clothes", + "bbox": [ + 496, + 613, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7514", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4923e65b2fc850a5929ae0bb9a7352a5eb97006251701116aa2700d9a0632a96.jpg", + "image_caption": [ + "Figure 2. Overview of COTTON. Given a clothing image $C$ , COTTON uses Clothing Landmark Predictor and Clothing Segmentation Network to predict clothing landmarks $K_{C}$ and a clothing segmentation mask $M_{C}$ . $C$ is then warped by Landmark-guided Transformation with homography derived from human keypoints $K_{P}$ and $K_{C}$ to obtain the transformed images $C^{torso}$ and $C^{limb}$ . $r$ is used to adjust $K_{C}$ at this step for changing clothing size. The clothing elimination step uses $C^{torso}$ and $C^{limb}$ to mask human image $I$ and segmentation mask $S$ , resulting in clothing-agnostic image $I_{a}$ and segmentation $S_{a}$ . The variable $t$ controls whether clothing is tucked in or not. Finally, Outfit Generator aggregates the previous outputs to predict segmentation mask $\\hat{S}$ and synthesize the final try-on result $\\hat{I}$ ." + ], + "image_footnote": [], + "bbox": [ + 122, + 87, + 854, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "due to lack of semantic information. In contrast, our proposed COTTON separates clothes based on landmarks and segmentation masks to maintain clothing completeness and aligns clothes based on specific landmarks to address misalignment. Besides, [8] is a concurrent work. Both of our works improve garment warping by human-garment correspondence. While [8] relies on the computationally intensive DensePose method, our proposed $LT$ method is more efficient and enables clothing size adjustment.", + "bbox": [ + 75, + 412, + 470, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Virtual Try-on", + "text_level": 1, + "bbox": [ + 76, + 559, + 227, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Virtual try-on methods can be categorized into 3D-based methods [1,14,22,24,26,27,37] and 2D-based methods [5,6, 7,8,9,10,11,12,15,16,17,19,23,28,29,31,33,34,35]. Since 3D-based methods are time-consuming and real-world scenarios require fast inference speeds, we focus on 2D-based methods. [16] proposed a coarse-to-fine network to synthesize image-based virtual try-on results based on TPS warping. However, try-on results synthesized by [16,31] do not provide a clear silhouette of human body parts since they consider the human to be a general mask instead of considering the semantic information, e.g., legs, head, limbs, etc. To tackle this problem, [10,20] proposed the first semantic-guided virtual try-on network, which predicted the target semantic segmentation to provide human structural information for the network to synthesize photo-realistic body parts. Afterward, many try-on methods [5,6,7,9,23,29,35] are designed with an individual stage for predicting the target semantic segmentation before synthesizing virtual try-on results. However, the individual stage for predicting target semantic segmentation would lead to error accumulation as pointed out by [21]. In this paper, we propose an end-to", + "bbox": [ + 75, + 583, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "end Outfit Generator that updates semantic segmentation and try-on results simultaneously, and introduce a Clothing Elimination Policy that retains valuable human characteristics while removing original clothing information.", + "bbox": [ + 498, + 412, + 893, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 494, + 635, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a human image $I$ and a clothing image $C$ , the goal of virtual try-on is to obtain a synthetic image $\\hat{I}$ , which shows the human of $I$ wearing the clothing of $C$ . We design a Clothing-Oriented Transformation Try-On Network (COTTON) for generating high-resolution virtual try-on results. Fig. 2 provides an overview of the proposed COTTON. We use the Clothing Landmark Predictor and the Clothing Segmentation Network to respectively capture clothing landmarks and a clothing segmentation map of $C$ . Afterward, we separate the clothes into torso part $C^{torse}$ and limbs part $C^{limb}$ according to the clothing segmentation map. The separation of clothes enables COTTON to treat each part differently, and thus makes it adapt to handling complex warping well. To align the clothes with the target human pose, the clothing landmarks are adjusted with a ratio $r$ to fit human body proportions and used to calculate homography matrices. COTTON then warps the clothes with affine transformation. The warped clothes shed light on how to properly remove the clothing region of image $I$ without losing necessary information. We use a user-determined variable $t$ to control whether to eliminate the overlap region of clothes. In this case, we can generate tucked-in or untucked try-on results by choosing the value of $t$ . Finally, the Outfit Generator takes both human and clothing representation to generate the try-on result $\\hat{I}$ .", + "bbox": [ + 496, + 523, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7515", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Clothing Landmark Predictor (CLP)", + "text_level": 1, + "bbox": [ + 76, + 90, + 397, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One of the most challenging problems of virtual try-on is to transform a clothing image $C$ to fit the target human pose. To solve this challenge, several approaches have been proposed in previous works, including hyperparameter learning of TPS warping [5, 16, 28, 31], and flow-based deep deformation [6, 15, 23]. However, these approaches implicitly learn the underlying geometric information from the training dataset and only achieve mediocre results when handling complex warping. We argue that explicit geometric information plays the most critical role in virtual try-on and can be leveraged to further improve performance. For example, by thoroughly understanding the clothes' geometry, humans can easily capture the relation between clothes and humans, even if the clothes are with different textures and styles. Therefore, we propose the Clothing Landmark Predictor (CLP) to extract explicit geometric information.", + "bbox": [ + 75, + 114, + 472, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, to exploit the geometric information of clothes, our proposed CLP is trained to predict a set of landmarks $K_{C}$ , which correlate to different parts of human pose $K_{P}$ , where $K_{P}$ can be obtained from off-the-shelf human pose estimation methods, e.g., [3]. The ground truth of $K_{C}$ is labeled manually, and the number of landmarks varies with regard to the clothing types. For upper clothing, $|K_{C}| = 10$ , including neck, shoulders, elbows, wrists, and hips. These landmarks explicitly indicate the relation between clothes and human pose, and therefore can be used as a reference to generate transformation matrices. Given clothing image $C$ , the CLP is trained to output clothing joint heatmaps $\\hat{M}_{JH}$ and Part Affinity Field $\\hat{M}_{PAF}$ . Afterward, the predicted landmarks $K_{C}$ can be obtained from $\\hat{M}_{JH}$ and $\\hat{M}_{PAF}$ by using the Hungarian algorithm.", + "bbox": [ + 75, + 359, + 472, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Clothing Segmentation Network", + "text_level": 1, + "bbox": [ + 76, + 603, + 361, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each part of clothes has distinct properties that should be taken into consideration during transformation. For example, the sleeves of clothes usually suffer severe distortion during transformation while the torso part only undergoes a mild distortion. Another common noise in clothing try-on is the region around the neckline that must be concealed when the garment is worn. This region always introduces undesired variation into try-on results, and as such, should be eliminated. Though the Clothing Landmark Predictor provides the connection between the clothes and human poses, it is difficult to identify different regions with only a few landmarks. To find the explicit boundary of each part, pixel-level semantic segmentation is necessary. Therefore, we propose a Clothing Segmentation Network to predict the clothing segmentation mask $M_C$ , which separates the clothing into three sub-region, including torso part, sleeve part, and invisible part around the neckline. As such, each sub-region can be treated properly based on its characteristics.", + "bbox": [ + 75, + 628, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fd7d1e9ad3e08136529b6dd189d740d9fecb6a19a7dfca0885bc8d4dcbba3a23.jpg", + "image_caption": [ + "Figure 3. Clothing transformation steps. For the torso part, we use keypoints including shoulders and hips to calculate matrix $H_{c \\to p}^{torso}$ . For the limbs part, we first create bounding boxes around limbs and use them to calculate matrices $H_{c \\to p}^{limb,i}$ , $i \\in \\{1,2,3,4\\}$ ." + ], + "image_footnote": [], + "bbox": [ + 506, + 87, + 888, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Landmark-guided Transformation", + "text_level": 1, + "bbox": [ + 500, + 290, + 802, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, we introduce how to leverage the landmarks $K_{C}$ and segmentation mask $M_{C}$ to achieve high-fidelity clothing transformation. We first remove the undesired part around the neckline such as the green region of $M_{C}$ in Fig. 2, which is invisible in the final try-on result, via the clothing segmentation map. We then separate the remaining part into two parts, including the torso part and the limbs part. Let $K_{C}^{\\text{torso}}$ and $K_{P}^{\\text{torso}}$ denote the clothing and human keypoints belonging to the torso part, such as shoulders and hips. For aligning the torso part, we compute a homography matrix $H_{c\\rightarrow p}^{\\text{torso}} \\in R^{3\\times 3}$ that can project clothing landmarks to human keypoints, as follows:", + "bbox": [ + 496, + 313, + 890, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nk _ {p} = H _ {c \\rightarrow p} ^ {t o r s o} k _ {c}, \\quad \\forall k _ {c} \\in K _ {C} ^ {t o r s o}, \\forall k _ {p} \\in K _ {P} ^ {t o r s o}. (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 507, + 890, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Afterward, this perspective transformation is applied to get the transformed torso part $C^{torso}$ . On the other hand, as the limbs part usually suffers from severe distortion, it requires a more sophisticated procedure. Therefore, instead of directly warping the whole limbs part, we break the limbs part into four sub-limbs based on the clothing landmarks. Take a hoodie as an example in Fig. 3. The limbs part is divided into two upper arms and two forearms. We create bounding boxes for each sub-limb based on clothing landmarks and human keypoints to obtain the area of each sub-limb. The corners of bounding boxes are then used to compute 4 homography matrices, $H_{c\\rightarrow p}^{limb,i}, i\\in \\{1,2,3,4\\}$ , which project each sub-limb to the target human. Finally, the four sub-limbs are collected as the warped limbs part $C^{limb}$ .", + "bbox": [ + 496, + 537, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is worth noting that the predicted clothing landmarks $K_{C}$ only provide a clothes-human relation in the general case. However, since body proportion varies from person to person, the relative size of the same clothing on different persons may change. To generate a more realistic try-on result, we can adjust the clothing landmark $K_{C}$ to fit the body proportion of the target human. For instance, if two people have the same shoulder width but different torso lengths, the portion of their bodies covered by clothing should be different. In that case, the predicted landmarks are adjusted", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7516", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "accordingly to fit their body proportions. Let $r$ be the ratio of torso length over shoulder width, which can be derived from extra body information provided by customers in online shopping scenarios. Fig. 1 visualizes the results of clothing landmarks adjustment with different $r$ , which shows that the region covered by the clothes is varied with regards to different $r$ , and can therefore generate try-on results with different clothing sizes.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Clothing Elimination Policy", + "text_level": 1, + "bbox": [ + 76, + 224, + 326, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the goal of the virtual try-on task is to replace the clothes of a given person image $I$ with a target one, the information about the original clothes should be eliminated. One common approach is eliminating the clothing according to the human segmentation $S$ . However, [5] pointed out that simply removing the clothing from image $I$ leads to a performance drop for paired data training due to information leakage about the clothing shape, and therefore proposed a clothing-agnostic person representation to address this problem. Nevertheless, this approach may lead to valuable information loss, e.g., arm width in case I of Fig. 5, when whole arms are directly masked during inference.", + "bbox": [ + 75, + 250, + 468, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hence, we propose a Clothing Elimination Policy for the inference phase to preserve the important information while generating clothing-agnostic person representation. The regions that ought to be eliminated in image $I$ are i) the regions of original clothes and ii) the regions covered by target clothes $C$ when worn. Note that the latter is hard to measure with previous methods. Fortunately, since we have already obtained the reliable warped clothing image at the previous step, we can use it as a reference to find the region that would be covered after try-on. Specifically, we create a clothing-elimination mask $M_{elm}$ by union the original clothing region and the region covered by the transformed target clothes, which is then expanded using morphological dilation to cover the entire body. Moreover, when wearing tops, we make the clothing tucked or not in the try-on result by controlling $t \\in \\{0,1\\}$ to determine whether to eliminate the overlap region of bottoms. As such, we can effectively obtain clothing-agnostic person representation without overly masking the irrelevant regions, i.e.,", + "bbox": [ + 75, + 431, + 472, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM _ {e l m} = \\mathcal {M} (S ^ {t} \\cup (C ^ {t o r s e} \\cup C ^ {l i m b}) - t \\cdot (C ^ {t o r s e} \\cap S ^ {b})), (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 732, + 468, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{M}$ represents the morphological dilation; $S^t$ and $S^b$ denote top and bottom clothing segmentation respectively. Fig. 1 shows the results of different $t$ . After eliminating clothing information, we obtain clothing-agnostic images $I_a$ and segmentation maps $S_a$ . To bridge the performance gap between the training and inference phases, we randomly mask the whole arms of human images with a $50\\%$ probability during the training phase and apply our Clothing Elimination Policy only in the inference phase.", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Outfit Generator", + "text_level": 1, + "bbox": [ + 500, + 90, + 668, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the final stage, the Outfit Generator takes personal $(I_{a}, S_{a}$ and $K_{P})$ and clothing $(C^{torse}$ and $C^{limb})$ information as input and synthesizes the final try-on result. Since learning the shape and texture of clothing simultaneously is challenging for the generator, a segmentation network is commonly used to predict a segmentation mask as guidance to lead the try-on generator network [5, 20, 23]. However, they regarded segmentation prediction as an independent task and trained the segmentation network and try-on generator separately. In that case, the optimization of the two networks may be suboptimal. Moreover, the accumulated noise from the segmentation network may further deteriorate the performance of the final synthesis result [21].", + "bbox": [ + 496, + 114, + 893, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the problem, we propose a novel Outfit Generator by combining the segmentation network and try-on generator as one end-to-end training multi-task network. Specifically, $S_{a}$ , $K_{P}$ , $C^{torse}$ , $C^{limb}$ are first concatenated and down-sampled to a lower resolution as the input of the segmentation network to predict segmentation mask $\\hat{S}$ . Afterward, $\\hat{S}$ is up-sampled back to the original resolution and combined with $I_{a}$ as well as $C^{torse}$ and $C^{limb}$ as the input of the synthesis step. Finally, the generator integrates all the information and renders the synthesis result $\\hat{I}$ .", + "bbox": [ + 496, + 311, + 893, + 464 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I}, \\hat {S} = G (I _ {a}, S _ {a}, K _ {P}, C ^ {t o r s e}, C ^ {l i m b}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 470, + 890, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The objective function of the Outfit Generator includes the supervision of both the predicted segmentation mask and the try-on synthesis result. We apply focal loss $\\mathcal{L}_{\\text {focal }}$ [25] for segmentation. For try-on results, we use losses including reconstruction loss $\\mathcal{L}_{\\text {rec }}$ , perceptual loss $\\mathcal{L}_{\\text {VGG }}$ , and adversarial loss $\\mathcal{L}_{\\text {adv }}$ . The overall loss functions are as follows:", + "bbox": [ + 496, + 494, + 893, + 585 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {o v e r a l l}} = \\lambda_ {\\text {r e c}} \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {V G G} \\mathcal {L} _ {V G G} \\tag {4} \\\\ + \\lambda_ {f o c a l} \\mathcal {L} _ {f o c a l} + \\lambda_ {a d v} \\mathcal {L} _ {a d v}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 612, + 890, + 647 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c}} = \\| \\hat {I} - I \\| _ {1}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 651, + 890, + 667 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {V G G} = \\sum_ {k = 1} ^ {5} \\lambda_ {k} \\| \\phi_ {k} (\\hat {I}) - \\phi_ {k} (I) \\| _ {1}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 672, + 890, + 712 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f o c a l} = \\sum_ {i} - \\alpha_ {i} (1 - p _ {i}) ^ {\\gamma} \\log (p _ {i}), p _ {i} \\in \\hat {S}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 715, + 890, + 747 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a d v} = \\left\\| \\mathcal {D} (\\hat {I}) - \\mathbf {1} \\right\\| _ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 750, + 890, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\phi_{k}$ represents the $k^{th}$ layer output of VGG-19 network. $\\alpha_{i}$ and $p_i$ are respectively the importance weighting for the $i^{th}$ class and the probabilities of pixels belonging to the $i^{th}$ class. $\\gamma$ is the focusing parameter. $\\lambda_{rec},\\lambda_{VGG},\\lambda_{focal}$ and $\\lambda_{adv}$ denote the balance weights. $\\mathcal{D}$ is the multi-layer discriminators [32].", + "bbox": [ + 500, + 782, + 890, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3The comparison of two-stage and end-to-end training is in supplement.", + "bbox": [ + 514, + 886, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7517", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 90, + 209, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 125, + 266, + 142 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We created the Pure Cotton, a high-resolution $(1024\\times 768)$ outfit dataset, to mitigate the impact of noisy data on model performance. We strictly employed several rules to filter out unsuitable data, which are clearly outlined in the supplements, resulting in 16,428 frontal-view human models and in-shop clothing pairs, including 10,636 upper clothes and 5,792 bottom clothes. For training and testing, we divided the upper clothing set into 8,451 and 2,185 pairs, respectively, and the bottom clothing set into 4,626 and 1,166 pairs. To evaluate the model's robustness, we also experimented with the public Dress Code dataset [28].", + "bbox": [ + 75, + 159, + 468, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. To train Clothing Landmark Predictor (CLP) and Clothing Segmentation Network (CSN), we collected a small clothing dataset with 50 images for each clothing type, including shirts, turtleneck, hoodies, etc, and manually labeled landmarks and segmentation masks as ground truth. The architecture of CLP is based on the one in [3], and that of CSN is based on FCN-resnet50 [30]. Moreover, Outfit Generator is designed using an Unet and an encoder-decoder network. To ensure a fair comparison with previous studies, all figures are generated in the tucked-in setting unless stated otherwise. This setting is consistent with prior arts.", + "bbox": [ + 75, + 335, + 468, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our proposed method with 3 relevant state-of-the-art baselines, including HR-VITON [23], VITON-HD [5], and PASTA-GAN [33], by training all the models with our self-collected Pure Cotton dataset and the public Dress Code dataset [28]. Specifically, HR-VITON and VITON-HD both represent SOTA high-resolution virtual try-on methods with different warping modules. HR-VITON warps clothes with appearance flows and designs a feature fusion network for eliminating the mismatch issue, while VITON-HD warps clothes with TPS and proposes an alignment-aware segmentation normalization method to deal with the misalignment problem. Meanwhile, PASTA-GAN represents another warping method, a patch-routed disentanglement module, which is more related to our proposed clothing deformation method.", + "bbox": [ + 75, + 527, + 468, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We evaluate the model performance by three widely-used metrics for virtual try-on, including i) Structural Similarity (SSIM) [38], ii) Learned Perceptual Image Patch Similarity (LPIPS) [36], and iii) Fréchet Inception Distance (FID) [18]. SSIM measures the quality of three key features: luminance, contrast, and structure between the reconstruction results and the ground truth. LPIPS evaluates the perceptual similarity and FID is computed for evaluating the GAN performance.", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Qualitative Results", + "text_level": 1, + "bbox": [ + 500, + 90, + 684, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 4 and 5 show the visual comparisons in Pure Cotton and Dress Code datasets. Overall, our model achieves the most visually convincing high-resolution try-on results. The comparison analysis is discussed as follows.", + "bbox": [ + 498, + 114, + 890, + 174 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The previous warping methods, e.g., TPS (VITON-HD) and appearance flows (HR-VITON), did not consider clothing skeleton and sleeves segmentation, rendering them incapable of accurately synthesizing clothing patterns and features in respective positions, such as hit color mosaic (case I in Fig. 4) and logo in special positions (case II in Fig. 5). Additionally, these methods could not tackle regional clothing warping, such as arm akimbo in complex postures (case II in Fig. 4). In contrast, our model conducts the Clothing Landmark Predictor and Clothing Segmentation Network to obtain clothing structural information. Then, our proposed Landmark-guided Transformation helps to warp the clothes regionally. As such, our model could synthesize try-on results with clothing patterns and features located in the right position that closely follow the target clothing image, even in complex postures. On the other hand, PASTA-GAN conducts a regional clothing deformation method similar to our proposed method. However, there are three differences between PASTA-GAN and our model. Firstly, we split the clothes into 5 regions, i.e., torso, right/left upper arms, and right/left forearms, and kept the torso undivided. Secondly, we further consider clothing segmentation, including neckline and limb segmentation. Thirdly, PASTA-GAN deforms clothes from clothes on another human image, as demonstrated on the right-bottom side of the target clothes columns in Fig. 4. Due to these differences, PASTA-GAN would produce strange try-on results that highly follow the original clothing tied up and limb postures (the blue circle and the red arrow on Fig 4). In contrast, our model leverages clothing structural information to synthesize natural clothing wrinkles.", + "bbox": [ + 498, + 174, + 892, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additional Clothing Control. The elimination policy in our proposed approach allows for greater flexibility in virtual try-on, with options for both tucked-in and untucked clothing. As shown in Fig. 6, previous methods would limit the location of the upper clothes' lower boundary based on the original lower clothes' upper boundary. However, our approach allows for the clothing to be untucked based on the clothes' body length, providing additional cues for customers to judge the clothing style.", + "bbox": [ + 498, + 643, + 890, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Partial Clothing Resizing: Apart from demonstrating COTTON's clothing size adjustment in Fig. 1, we also present its ability to adjust partial clothing sizes in Fig. 7. In the real world, clothing designers would adjust different sizes only for specific parts, e.g., body or sleeve length, instead of the entire garment. COTTON can smoothly manipulate partial clothing sizes by adjusting the corresponding landmarks without impacting the rest of the clothing.", + "bbox": [ + 498, + 780, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7518", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/88bbb3a57198f79bae6a2e205f97d4c4374fe82e163f4dfdb9d6e2f06c1bbdec.jpg", + "image_caption": [ + "Figure 4. The visual comparison in the Pure Cotton dataset. The left two columns are the inputs and the rest present the results of baselines and our model. VITON-HD and HR-VITON did not consider sleeve segmentation, resulting in misaligned clothes with wrong pattern shifts (green dash line in case I) and broken warped clothes (case II). Moreover, without clothing skeletons, VITON-HD and HR-VITON failed to maintain pattern proportions (red and blue double-headed arrows in case I). PASTA-GAN also had issues in case I, where the partially tucked-in clothing shape from the original human image was directly copied to the try-on result, and in case II, where the try-on result had two additional limbs in front of the abdomen due to the human, representing target clothes, crossing their arms over the chest." + ], + "image_footnote": [], + "bbox": [ + 143, + 93, + 823, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fa331c3cfeed98f7dd1f04d04de7fbd70bfb267d1f5ea5a85523f4f82eda5226.jpg", + "image_caption": [ + "Figure 5. The visual comparison in the Dress Code dataset. In case I, our Clothing Elimination Policy well retains valuable human features, e.g., bulging muscles. For case II, our Landmark-guided Transformation aligns textures better based on the obtained clothing structure." + ], + "image_footnote": [], + "bbox": [ + 151, + 425, + 821, + 667 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dbbbb79aa58abbd171567925a642994356de1d8f9b2e4be60977705e2c5ee4ca.jpg", + "image_caption": [ + "Figure 6. Additional clothing control for tucked-in or untucked." + ], + "image_footnote": [], + "bbox": [ + 84, + 713, + 462, + 795 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Quantitative Results", + "text_level": 1, + "bbox": [ + 76, + 828, + 269, + 844 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 compares COTTON's performance with SOTAs in terms of SSIM, LPIPS, and FID. We evaluated SSIM and LPIPS on 2,000 random paired test sets from our Pure Cot", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ton dataset and 1,800 test sets from the Dress Code dataset. FID was evaluated on 2,000 and 1,800 unpaired test sets from Pure Cotton and Dress Code datasets, respectively. Human images wearing original clothing served as ground truth for FID calculation. The results show that COTTON consistently outperforms the baselines across all evaluation metrics on both datasets, with at least $18.6\\%$ improvement in LPIPS on Pure Cotton dataset and $41.1\\%$ in FID on Dress Code dataset. The results manifest that our method excels in generating detailed contents on Pure Cotton dataset and robust results on Dress Code dataset. Notably, COTTON shows lower performance with untucked clothes, as it ob", + "bbox": [ + 496, + 713, + 890, + 893 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7519", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e1a20a28c0d5bdb66fdadc91b140b5b2eff79e838992baa75ea409fbd88ebd25.jpg", + "image_caption": [ + "Sleeves length adjustment" + ], + "image_footnote": [], + "bbox": [ + 101, + 104, + 447, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/36e13e32b0199788795087a0358c02f610129c1afb7c457f32e5984e501c19ef.jpg", + "image_caption": [ + "Body length adjustment", + "Figure 7. Visualization of partial clothing adjustment." + ], + "image_footnote": [], + "bbox": [ + 101, + 301, + 447, + 452 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cad675f3d94dd95acc37bad00718b8605f45b4c17d8f72c26022c9965460a1be.jpg", + "table_caption": [ + "Table 1. Quantitative comparison on the test dataset." + ], + "table_footnote": [], + "table_body": "
MethodDatasetSSIM↑LPIPS↓FID↓
PASTA-GAN0.9180.121529.43
VITON-HD0.9440.056617.16
HR-VITONSelf-collected0.9480.038711.16
Ours (untucked)Pure Cotton0.9560.034910.96
Ours (tucked-in)0.9580.031510.17
PASTA-GAN0.8000.306633.92
VITON-HD0.9270.090622.53
HR-VITONPublic0.9270.075520.12
Ours (untucked)Dress Code0.9490.048212.58
Ours (tucked-in)0.9530.043811.86
", + "bbox": [ + 78, + 500, + 480, + 670 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9bb56ed274c4bb6b9553a656d6ef3a3458421b30fec4bb03b5e5d2efeb928d6a.jpg", + "table_caption": [ + "Table 2. The user study shows that COTTON is the most photorealistic and remains the most human and clothing characteristics." + ], + "table_footnote": [], + "table_body": "
MethodPASTA-GANVITON-HDHR-VITONCOTTON(Ours)
Photo-realistic9.08%7.56%29.82%53.54%
Try-on accuracy8.66%4.62%27.40%59.32%
", + "bbox": [ + 93, + 717, + 450, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "scures the upper portion of lower clothing, causing the try-on results to differ from the ground truth. Nevertheless, Fig. 6 demonstrates that COTTON can generate more visually appealing results when clothing is untucked.", + "bbox": [ + 75, + 823, + 467, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, we conduct a user study on the Pure Cotton", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dataset to further evaluate the visual quality by humans. The experiment involved 127 volunteers. We randomly sampled 30 image sets and divided them equally into two groups for evaluating: i) photo-realistic quality and ii) try-on accuracy. To evaluate the photo-realistic quality, volunteers would see four try-on results (without input information) synthesized by four different models, and choose the most photo-realistic one. Furthermore, for evaluating try-on accuracy, we provide a source human image, one target try-on clothing, and four try-on results synthesized by four different models. Based on the source human and the target clothes, volunteers should select the best try-on result. Table 2 shows that COTTON outperforms all other baselines in both photo-realistic quality and try-on accuracy and achieves even higher ratings for try-on accuracy. Volunteers reported that they mainly used realistic reproduced clothing patterns, including loose/fit torso and accurate sleeve length, with fewer artifacts as the criteria to determine the better model. Our proposed Clothing Landmark Predictor and Clothing Segmentation Network effectively retain clothing characteristics such as patterns and sleeve length, contributing significantly to COTTON's superior performance over other state-of-the-art models.", + "bbox": [ + 496, + 90, + 890, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 460, + 617, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents COTTON, a clothing-oriented transformation try-on network that synthesizes high-resolution virtual try-on results. By considering the clothing structure and leveraging it, COTTON improves the clothing transformation quality and allows for size adjustment during clothing try-on. Additionally, the proposed elimination policy enables COTTON to try on outfits untucked or not while preserving valuable human characteristics. The experiments demonstrate that COTTON outperforms the state-of-the-art virtual try-on works at $1024 \\times 768$ resolution, both qualitatively and quantitatively. In future work, we plan to improve the accessories segmentation and extend the system to try on more clothing categories, such as shoes, hats, and accessories. Overall, the proposed approach shows great promise for improving the online shopping experience and reducing the need for physical try-on.", + "bbox": [ + 496, + 488, + 890, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 751, + 658, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We would like to express our gratitude to Josephine, Wendy, Hans, Joyce, Leslie and May from the Epoch Foundation for their valuable support. This work was supported in part by the National Science and Technology Council of Taiwan under Grants NSTC-109-2221-E-009-114-MY3, NSTC-112-2221-E-A49-094-MY3, NSTC-109-2223-E-009-002-MY3, NSTC-112-2634-F-002-002-MBK and NSTC-112-2628-E-002-033-MY4.", + "bbox": [ + 496, + 779, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7520", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alakh Aggarwal, Jikai Wang, Steven Hogue, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. Layered-garment net: Generating multiple implicit garment layers from a single image. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2022. 3", + "[2] Shuai Bai, Huiling Zhou, Zhikang Li, Chang Zhou, and Hongxia Yang. Single stage virtual try-on via deformable attention flows. In European Conference on Computer Vision (ECCV), 2022. 1, 2", + "[3] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 4, 6", + "[4] Chieh-Yun Chen, Ling Lo, Pin-Jui Huang, Hong-Han Shuai, and Wen-Huang Cheng. Fashionmirror: Co-attention feature-remapping virtual try-on with sequential template poses. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1", + "[5] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. VITON-HD: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3, 4, 5, 6", + "[6] Ayush Chopra, Rishabh Jain, Mayur Hemani, and Balaji Krishnamurthy. Zflow: Gated appearance flow-based virtual try-on with 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3, 4", + "[7] Chien-Lung Chou, Chieh-Yun Chen, Chia-Wei Hsieh, Hong-Han Shuai, Jiaying Liu, and Wen-Huang Cheng. Template-free try-on image synthesis via semantic-guided optimization. IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 2021. 3", + "[8] Aiyu Cui, Sen He, Tao Xiang, and Antoine Toisoul. Learning garment densepose for robust warping in virtual try-on. In arXiv, 2023. 3", + "[9] Aiyu Cui, Daniel McKee, and Svetlana Lazebnik. Dressing in order: Recurrent person image generation for pose transfer, virtual try-on and outfit editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 3", + "[10] Haoye Dong, Xiaodan Liang, Xiaohui Shen, Bochao Wang, Hanjiang Lai, Jia Zhu, Zhiting Hu, and Jian Yin. Towards multi-pose guided virtual try-on network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 3", + "[11] Xin Dong, Fuwei Zhao, Zhenyu Xie, Xijin Zhang, Kang Du, Min Zheng, Xiang Long, Xiaodan Liang, and Jianchao Yang1. Dressing in the wild by watching dance videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3", + "[12] Ruili Feng, Cheng Ma, Chengji Shen, Xin Gao, Zhenjiang Liu, Xiaobo Li, Kairi Ou, Deli Zhao, and Zheng-Jun Zha. Weakly supervised high-fidelity clothing model generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yuying Ge, Yibing Song, Ruimao Zhang, Chongjian Ge, Wei Liu, and Ping Luo. Parser-free virtual try-on via distilling appearance flows. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2", + "[14] Oshri Halimi, Tuur Stuyck, Donglai Xiang, Timur Bagautdinov, He Wen, Ron Kimmel, Takaaki Shiratori, Chenglei Wu, Yaser Sheikh, and Fabian Prada. Pattern-based cloth registration and sparse-view animation. ACM Trans. Graph., 41(6), 2022. 3", + "[15] Xintong Han, Xiaojun Hu, Weilin Huang, and Matthew R. Scott. ClothFlow: a flow-based model for clothed person generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 3, 4", + "[16] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S. Davis. VITON: An image-based virtual try-on network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 3, 4", + "[17] Sen He, Yi-Zhe Song, and Tao Xiang. Style-based global appearance flow for virtual try-on. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3", + "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Proceedings of the International Conference on Neural Information Processing Systems (NIPS), 2017. 6", + "[19] Chia-Wei Hsieh, Chieh-Yun Chen, Chien-Lung Chou, Hong-Han Shuai, and Wen-Huang Cheng. Fit-me: Image-based virtual try-on with arbitrary poses. In IEEE International Conference on Image Processing (ICIP), 2019. 3", + "[20] Chia-Wei Hsieh, Chieh-Yun Chen, Chien-Lung Chou, Hong-Han Shuai, Jiaying Liu, and Wen-Huang Cheng. FashionOn: Semantic-guided image-based virtual try-on with detailed human and clothing information. In Proceedings of the $27^{th}$ ACM International Conference on Multimedia (ACM MM), 2019. 3, 5", + "[21] Thibaut Issenhuth, Jérémie Mary, and Clément Calauzènes. Do not mask what you do not need to mask: a parser-free virtual try-on. In European Conference on Computer Vision (ECCV), 2020. 2, 3, 5", + "[22] Maria Korosteleva and Sung-Hee Lee. Neuraltailor: Reconstructing sewing pattern structures from 3d point clouds of garments. ACM Trans. Graph., 41(4), 2022. 3", + "[23] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 4, 5, 6", + "[24] Ren Li, Benoit Guillard, Edoardo Remelli, and Pascal Fua. Dig: Draping implicit garment over the human body. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2022. 3", + "[25] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7521", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Xinqi Liu, Jituo Li, Guodong Lu, Dongliang Zhang, and Shihai Xing. Robust and automatic clothing reconstruction based on a single rgb image. Computers & Graphics, 2023. 3", + "[27] S. Majithia, S. N. Parameswaran, S. Babar, V. Garg, A. Srivastava, and A. Sharma. Robust 3d garment digitization from monocular 2d images for 3d virtual try-on systems. In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2022. 3", + "[28] Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, and Rita Cucchiara. Dress Code: High-Resolution Multi-Category Virtual Try-On. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 4, 6", + "[29] Assaf Neuberger, Eran Borenstein, Bar Hilleli, Eduard Oks, and Sharon Alpert. Image based virtual try-on network from unpaired data. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 3", + "[30] Evan Shelhamer, Jonathan Long, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 6", + "[31] Bochao Wang, Huabin Zheng, Xiaodan Liang, Yimin Chen, Liang Lin, and Meng Yang. Toward characteristic-preserving image-based virtual try-on network. In European Conference on Computer Vision (ECCV), 2018. 2, 3, 4", + "[32] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5", + "[33] Zhenyu Xie, Zaiyu Huang, Fuwei Zhao, Haoye Dong, Michael Kampffmeyer, and Xiaodan Liang. Towards scalable unpaired virtual try-on via patch-routed spatially-adaptive GAN. In Advances in Neural Information Processing Systems (NIPS), 2021. 1, 2, 3, 6", + "[34] Han Yang, Xinrui Yu, and Ziwei Liu. Full-range virtual try-on with recurrent tri-level transform. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[35] Han Yang, Ruimao Zhang, Xiaobao Guo, Wei Liu, Wang-meng Zuo, and Ping Luo. Towards photo-realistic virtual try-on by adaptively generating $\\leftrightarrow$ preserving image content. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 3", + "[36] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6", + "[37] Fuwei Zhao, Zhenyu Xie, Michael Kampffmeyer, Haoye Dong, Songfang Han, Tianxiang Zheng, Tao Zhang, and Xiaodan Liang. M3d-vton: A monocular-to-3d virtual try-on network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3", + "[38] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "to structural similarity. IEEE Transactions on Image Processing (TIP), 2004. 6", + "bbox": [ + 532, + 90, + 890, + 119 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7522", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Size Does Matter_ Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network/2e5e4941-f171-4e8c-be33-35affe248d0e_model.json b/2023/Size Does Matter_ Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network/2e5e4941-f171-4e8c-be33-35affe248d0e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..89e7c823f15a2f1d801dc42689ceda2d3d685a04 --- /dev/null +++ b/2023/Size Does Matter_ Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network/2e5e4941-f171-4e8c-be33-35affe248d0e_model.json @@ -0,0 +1,1716 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.131, + 0.823, + 0.177 + ], + "angle": 0, + "content": "Size Does Matter: Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.203, + 0.826, + 0.223 + ], + "angle": 0, + "content": "Chieh-Yun Chen\\(^{1,2*}\\) Yi-Chung Chen\\(^{1,3*}\\) Hong-Han Shuai\\(^{2}\\) Wen-Huang Cheng\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.224, + 0.82, + 0.244 + ], + "angle": 0, + "content": "\\(^{1}\\)Stylins.ai \\(^{2}\\)National Yang Ming Chiao Tung University \\(^{3}\\)National Taiwan University" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.267, + 0.871, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.474, + 0.893, + 0.503 + ], + "angle": 0, + "content": "Figure 1. Multi-garment virtual try-on results with different sizes (tucked or untucked). \\( r \\) is the ratio of torso length over shoulder width, which can be decided by customers in online shopping scenarios. Padding blue backgrounds for better visualizing clothing skeletons." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.516, + 0.314, + 0.531 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.538, + 0.474, + 0.885 + ], + "angle": 0, + "content": "Virtual try-on tasks aim at synthesizing realistic try-on results by trying target clothes on humans. Most previous works relied on the Thin Plate Spline or appearance flows to warp clothes to fit human body shapes. However, both approaches cannot handle complex warping, leading to over distortion or misalignment. Furthermore, there is a critical unaddressed challenge of adjusting clothing sizes for try-on. To tackle these issues, we propose a Clothing-Oriented Transformation Try-On Network (COTTON). COTTON leverages clothing structure with landmarks and segmentation to design a novel landmark-guided transformation for precisely deforming clothes, allowing for size adjustment during try-on. Additionally, to properly remove the clothing region from the human image without losing significant human characteristics, we propose a clothing elimination policy based on both transformed clothes and human segmentation. This method enables users to try on clothes tucked-in or untucked while retaining more human characteristics. Both qualitative and quantitative results show that COTTON outperforms the state-of-the-art high-resolution virtual try-on approaches. All the code is available at https://github.com/cotton6/COTTON-size-does-matter." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.516, + 0.631, + 0.531 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.893, + 0.785 + ], + "angle": 0, + "content": "Image-based virtual try-on replaces clothing items on a person with the desired ones, creating realistic try-on results and lowering costs associated with on-model photos for the e-commerce industry1. Additionally, it enables customers to use virtual dressing rooms when shopping online, potentially enhancing the e-commerce experience and increasing conversion rates2. As online shopping becomes popular, the virtual try-on tasks have received more and more attention [2,4,5,9,13,17,23,28,29,33,35]. For example, to extend the virtual try-on resolution from low \\((256\\times 192)\\) to high \\((1024\\times 768)\\), the most obvious difficulty is the misalignment issue. [5] proposed the alignment-aware segment normalization to remove the misleading information in the misaligned area. Additionally, [23] further performs appearance flow-based warping and segmentation map generation simultaneously to tackle misalignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.787, + 0.894, + 0.849 + ], + "angle": 0, + "content": "Despite the considerable progress made in previous works, there are still some challenges that have not been adequately addressed, as outlined below. i) Handling complex warping without misalignment: General image" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.863, + 0.781, + 0.876 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.52, + 0.876, + 0.747, + 0.888 + ], + "angle": 0, + "content": "1https://www.zmo.ai/aimodels/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.52, + 0.888, + 0.692, + 0.9 + ], + "angle": 0, + "content": "2https://www.revery.ai" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.863, + 0.781, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "7513" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.683 + ], + "angle": 0, + "content": "based virtual try-on frameworks [5, 16, 21, 28, 31] typically employ a clothing deformation module, such as the Thin Plate Spline (TPS) method, to align clothing images with the target human pose. However, research [6, 15, 23] has found that TPS may not effectively handle complex warping when different garment regions require different deformations. In attempts to improve warping results, they have replaced TPS with dense appearance flows. However, when the transformation between the garment and corresponding body parts is significant (as illustrated in case II in Fig. 4), the performance of both TPS (VITON-HD [5]) and flow-based (HR-VITON [23]) methods deteriorates drastically. Hence, existing methods still can't properly address the challenge of clothing deformation. ii) Adjusting clothing sizes: Previous works predicted the try-on segmentation conditioned on the clothing image by only considering the shape of the clothes without the scale information. Thus, given a clothing image, it is impossible for previous works to change the clothing size. This limitation severely restricts the practical application of virtual try-on because people often try on different sizes in the fitting room. iii) Appropriate clothing elimination policy: Previous works [5, 23, 28] rely on segmentation maps to eliminate the original clothing region and utilize the remaining area as a guide for trying on target clothes. However, this method can lead to either excessive or insufficient removal of the original image. For instance, when prior works try on upper clothes, the lower part of the upper clothes remains intact, which limits the length of the upper clothes, as shown in Fig. 6. The lower part of the upper clothing is forced to be tucked in, which reduces the model's ability to generate complete upper clothes, thereby reducing the practicality of virtual try-on. Additionally, previous works remove the entire arm of input human images, which eliminates essential human characteristics such as tattoos and arm width that should be preserved in the final output, as illustrated in case I of Fig. 5. Therefore, developing an appropriate clothing elimination strategy that preserves crucial information while removing the clothing region is a significant challenge." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To address these challenges, we propose a simple yet powerful approach called Clothing-Oriented Transformation Try-On Network (COTTON). Specifically, COTTON first exploits geometric information with Clothing Landmark Predictor and Clothing Segmentation Network to predict clothing landmarks and segmentation masks respectively. To overcome the first challenge posed by complex warping, we propose the Landmark-guided Transformation that first separates clothes into sub-parts via the segmentation mask and then uses clothing landmarks to estimate homography matrices that fit these sub-parts to the target human pose. To solve the second challenge, we introduce a clothing landmarks adjustment approach that allows users to change clothing sizes. Adjustment of landmarks will al" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.273 + ], + "angle": 0, + "content": "ter the homography matrices and then cause clothing size changes, as demonstrated in Fig. 1, which remarkably enhances the practicality of virtual try-on. To address the third challenge, we employ the transformed clothing images to identify and remove the areas that would be covered by the clothes. This enables our proposed Clothing Elimination Policy to effectively eliminate the clothing region while preserving important details and offer flexibility in tucking the clothes or not, as shown in Fig. 6. Extensive experiments show that COTTON achieves superior results than state-of-the-arts both quantitatively and qualitatively. We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.283, + 0.892, + 0.344 + ], + "angle": 0, + "content": "- We propose a Clothing-Oriented Transformation Try-On Network (COTTON), which improves clothing transformation quality and addresses complex warping misalignments by leveraging clothing geometry." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.353, + 0.892, + 0.429 + ], + "angle": 0, + "content": "- We introduce adjustable clothing landmarks to provide clothing size information based only on images. To the best of our knowledge, this is the first work enabling 2D-based virtual try-on with different clothing sizes for approaching real-world try-on." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.439, + 0.892, + 0.5 + ], + "angle": 0, + "content": "- To preserve critical information, we propose a Clothing Elimination Policy to properly remove clothing information while retaining valuable human characteristics, and offer flexibility in tucking the clothes or not." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.509, + 0.892, + 0.554 + ], + "angle": 0, + "content": "- Extensive experiments on the Dress Code dataset [28] show that our model significantly outperforms SOTAs, e.g., at least \\(41.1\\%\\) improvement in terms of FID." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.283, + 0.892, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.566, + 0.651, + 0.582 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.591, + 0.707, + 0.607 + ], + "angle": 0, + "content": "2.1. Clothing Deformation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Virtual try-on requires clothing deformation to align patterns and preserve clothing details. Thin Plate Spline (TPS) warping is widely used in virtual try-on tasks [5, 16, 21, 28, 31], but it has limited ability to model geometric changes and result in unnatural deformations. Thus, [2, 6, 11, 13, 15, 17, 23] conducted appearance flows as an alternative to TPS, but misalignment still has not be well addressed. To address misalignment, [5] improved TPS by eliminating irrelevant clothing texture information in misaligned regions. Meanwhile, [23] improved the deformation by designing pathways to jointly predict appearance flows and segmentation features. However, neither TPS nor appearance flow can deform partial regions of clothing without affecting other parts, e.g., warping the sleeves to present arm akimbo without affecting the torso part of the clothes, since they do not have the structural information of the clothes. To address this issue, [33] introduced normalized patches to learn spatial-agnostic clothing features but this approach may destroy the completeness of clothes" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7514" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.088, + 0.855, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.319, + 0.896, + 0.404 + ], + "angle": 0, + "content": "Figure 2. Overview of COTTON. Given a clothing image \\(C\\), COTTON uses Clothing Landmark Predictor and Clothing Segmentation Network to predict clothing landmarks \\(K_{C}\\) and a clothing segmentation mask \\(M_{C}\\). \\(C\\) is then warped by Landmark-guided Transformation with homography derived from human keypoints \\(K_{P}\\) and \\(K_{C}\\) to obtain the transformed images \\(C^{torso}\\) and \\(C^{limb}\\). \\(r\\) is used to adjust \\(K_{C}\\) at this step for changing clothing size. The clothing elimination step uses \\(C^{torso}\\) and \\(C^{limb}\\) to mask human image \\(I\\) and segmentation mask \\(S\\), resulting in clothing-agnostic image \\(I_{a}\\) and segmentation \\(S_{a}\\). The variable \\(t\\) controls whether clothing is tucked in or not. Finally, Outfit Generator aggregates the previous outputs to predict segmentation mask \\(\\hat{S}\\) and synthesize the final try-on result \\(\\hat{I}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.471, + 0.552 + ], + "angle": 0, + "content": "due to lack of semantic information. In contrast, our proposed COTTON separates clothes based on landmarks and segmentation masks to maintain clothing completeness and aligns clothes based on specific landmarks to address misalignment. Besides, [8] is a concurrent work. Both of our works improve garment warping by human-garment correspondence. While [8] relies on the computationally intensive DensePose method, our proposed \\( LT \\) method is more efficient and enables clothing size adjustment." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.56, + 0.228, + 0.577 + ], + "angle": 0, + "content": "2.2. Virtual Try-on" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Virtual try-on methods can be categorized into 3D-based methods [1,14,22,24,26,27,37] and 2D-based methods [5,6, 7,8,9,10,11,12,15,16,17,19,23,28,29,31,33,34,35]. Since 3D-based methods are time-consuming and real-world scenarios require fast inference speeds, we focus on 2D-based methods. [16] proposed a coarse-to-fine network to synthesize image-based virtual try-on results based on TPS warping. However, try-on results synthesized by [16,31] do not provide a clear silhouette of human body parts since they consider the human to be a general mask instead of considering the semantic information, e.g., legs, head, limbs, etc. To tackle this problem, [10,20] proposed the first semantic-guided virtual try-on network, which predicted the target semantic segmentation to provide human structural information for the network to synthesize photo-realistic body parts. Afterward, many try-on methods [5,6,7,9,23,29,35] are designed with an individual stage for predicting the target semantic segmentation before synthesizing virtual try-on results. However, the individual stage for predicting target semantic segmentation would lead to error accumulation as pointed out by [21]. In this paper, we propose an end-to" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.414, + 0.895, + 0.476 + ], + "angle": 0, + "content": "end Outfit Generator that updates semantic segmentation and try-on results simultaneously, and introduce a Clothing Elimination Policy that retains valuable human characteristics while removing original clothing information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.495, + 0.637, + 0.514 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Given a human image \\( I \\) and a clothing image \\( C \\), the goal of virtual try-on is to obtain a synthetic image \\( \\hat{I} \\), which shows the human of \\( I \\) wearing the clothing of \\( C \\). We design a Clothing-Oriented Transformation Try-On Network (COTTON) for generating high-resolution virtual try-on results. Fig. 2 provides an overview of the proposed COTTON. We use the Clothing Landmark Predictor and the Clothing Segmentation Network to respectively capture clothing landmarks and a clothing segmentation map of \\( C \\). Afterward, we separate the clothes into torso part \\( C^{torse} \\) and limbs part \\( C^{limb} \\) according to the clothing segmentation map. The separation of clothes enables COTTON to treat each part differently, and thus makes it adapt to handling complex warping well. To align the clothes with the target human pose, the clothing landmarks are adjusted with a ratio \\( r \\) to fit human body proportions and used to calculate homography matrices. COTTON then warps the clothes with affine transformation. The warped clothes shed light on how to properly remove the clothing region of image \\( I \\) without losing necessary information. We use a user-determined variable \\( t \\) to control whether to eliminate the overlap region of clothes. In this case, we can generate tucked-in or untucked try-on results by choosing the value of \\( t \\). Finally, the Outfit Generator takes both human and clothing representation to generate the try-on result \\( \\hat{I} \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "7515" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.398, + 0.108 + ], + "angle": 0, + "content": "3.1. Clothing Landmark Predictor (CLP)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.116, + 0.473, + 0.359 + ], + "angle": 0, + "content": "One of the most challenging problems of virtual try-on is to transform a clothing image \\( C \\) to fit the target human pose. To solve this challenge, several approaches have been proposed in previous works, including hyperparameter learning of TPS warping [5, 16, 28, 31], and flow-based deep deformation [6, 15, 23]. However, these approaches implicitly learn the underlying geometric information from the training dataset and only achieve mediocre results when handling complex warping. We argue that explicit geometric information plays the most critical role in virtual try-on and can be leveraged to further improve performance. For example, by thoroughly understanding the clothes' geometry, humans can easily capture the relation between clothes and humans, even if the clothes are with different textures and styles. Therefore, we propose the Clothing Landmark Predictor (CLP) to extract explicit geometric information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.361, + 0.473, + 0.59 + ], + "angle": 0, + "content": "Specifically, to exploit the geometric information of clothes, our proposed CLP is trained to predict a set of landmarks \\( K_{C} \\), which correlate to different parts of human pose \\( K_{P} \\), where \\( K_{P} \\) can be obtained from off-the-shelf human pose estimation methods, e.g., [3]. The ground truth of \\( K_{C} \\) is labeled manually, and the number of landmarks varies with regard to the clothing types. For upper clothing, \\( |K_{C}| = 10 \\), including neck, shoulders, elbows, wrists, and hips. These landmarks explicitly indicate the relation between clothes and human pose, and therefore can be used as a reference to generate transformation matrices. Given clothing image \\( C \\), the CLP is trained to output clothing joint heatmaps \\( \\hat{M}_{JH} \\) and Part Affinity Field \\( \\hat{M}_{PAF} \\). Afterward, the predicted landmarks \\( K_{C} \\) can be obtained from \\( \\hat{M}_{JH} \\) and \\( \\hat{M}_{PAF} \\) by using the Hungarian algorithm." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.604, + 0.362, + 0.62 + ], + "angle": 0, + "content": "3.2. Clothing Segmentation Network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Each part of clothes has distinct properties that should be taken into consideration during transformation. For example, the sleeves of clothes usually suffer severe distortion during transformation while the torso part only undergoes a mild distortion. Another common noise in clothing try-on is the region around the neckline that must be concealed when the garment is worn. This region always introduces undesired variation into try-on results, and as such, should be eliminated. Though the Clothing Landmark Predictor provides the connection between the clothes and human poses, it is difficult to identify different regions with only a few landmarks. To find the explicit boundary of each part, pixel-level semantic segmentation is necessary. Therefore, we propose a Clothing Segmentation Network to predict the clothing segmentation mask \\( M_C \\), which separates the clothing into three sub-region, including torso part, sleeve part, and invisible part around the neckline. As such, each sub-region can be treated properly based on its characteristics." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.088, + 0.889, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.223, + 0.892, + 0.281 + ], + "angle": 0, + "content": "Figure 3. Clothing transformation steps. For the torso part, we use keypoints including shoulders and hips to calculate matrix \\( H_{c \\to p}^{torso} \\). For the limbs part, we first create bounding boxes around limbs and use them to calculate matrices \\( H_{c \\to p}^{limb,i} \\), \\( i \\in \\{1,2,3,4\\} \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.291, + 0.803, + 0.307 + ], + "angle": 0, + "content": "3.3. Landmark-guided Transformation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.314, + 0.892, + 0.496 + ], + "angle": 0, + "content": "Here, we introduce how to leverage the landmarks \\( K_{C} \\) and segmentation mask \\( M_{C} \\) to achieve high-fidelity clothing transformation. We first remove the undesired part around the neckline such as the green region of \\( M_{C} \\) in Fig. 2, which is invisible in the final try-on result, via the clothing segmentation map. We then separate the remaining part into two parts, including the torso part and the limbs part. Let \\( K_{C}^{\\text{torso}} \\) and \\( K_{P}^{\\text{torso}} \\) denote the clothing and human keypoints belonging to the torso part, such as shoulders and hips. For aligning the torso part, we compute a homography matrix \\( H_{c\\rightarrow p}^{\\text{torso}} \\in R^{3\\times 3} \\) that can project clothing landmarks to human keypoints, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.508, + 0.892, + 0.528 + ], + "angle": 0, + "content": "\\[\nk _ {p} = H _ {c \\rightarrow p} ^ {t o r s o} k _ {c}, \\quad \\forall k _ {c} \\in K _ {C} ^ {t o r s o}, \\forall k _ {p} \\in K _ {P} ^ {t o r s o}. (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Afterward, this perspective transformation is applied to get the transformed torso part \\( C^{torso} \\). On the other hand, as the limbs part usually suffers from severe distortion, it requires a more sophisticated procedure. Therefore, instead of directly warping the whole limbs part, we break the limbs part into four sub-limbs based on the clothing landmarks. Take a hoodie as an example in Fig. 3. The limbs part is divided into two upper arms and two forearms. We create bounding boxes for each sub-limb based on clothing landmarks and human keypoints to obtain the area of each sub-limb. The corners of bounding boxes are then used to compute 4 homography matrices, \\( H_{c\\rightarrow p}^{limb,i}, i\\in \\{1,2,3,4\\} \\), which project each sub-limb to the target human. Finally, the four sub-limbs are collected as the warped limbs part \\( C^{limb} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "It is worth noting that the predicted clothing landmarks \\( K_{C} \\) only provide a clothes-human relation in the general case. However, since body proportion varies from person to person, the relative size of the same clothing on different persons may change. To generate a more realistic try-on result, we can adjust the clothing landmark \\( K_{C} \\) to fit the body proportion of the target human. For instance, if two people have the same shoulder width but different torso lengths, the portion of their bodies covered by clothing should be different. In that case, the predicted landmarks are adjusted" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7516" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "accordingly to fit their body proportions. Let \\( r \\) be the ratio of torso length over shoulder width, which can be derived from extra body information provided by customers in online shopping scenarios. Fig. 1 visualizes the results of clothing landmarks adjustment with different \\( r \\), which shows that the region covered by the clothes is varied with regards to different \\( r \\), and can therefore generate try-on results with different clothing sizes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.226, + 0.327, + 0.243 + ], + "angle": 0, + "content": "3.4. Clothing Elimination Policy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.25, + 0.47, + 0.432 + ], + "angle": 0, + "content": "Since the goal of the virtual try-on task is to replace the clothes of a given person image \\( I \\) with a target one, the information about the original clothes should be eliminated. One common approach is eliminating the clothing according to the human segmentation \\( S \\). However, [5] pointed out that simply removing the clothing from image \\( I \\) leads to a performance drop for paired data training due to information leakage about the clothing shape, and therefore proposed a clothing-agnostic person representation to address this problem. Nevertheless, this approach may lead to valuable information loss, e.g., arm width in case I of Fig. 5, when whole arms are directly masked during inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.473, + 0.721 + ], + "angle": 0, + "content": "Hence, we propose a Clothing Elimination Policy for the inference phase to preserve the important information while generating clothing-agnostic person representation. The regions that ought to be eliminated in image \\(I\\) are i) the regions of original clothes and ii) the regions covered by target clothes \\(C\\) when worn. Note that the latter is hard to measure with previous methods. Fortunately, since we have already obtained the reliable warped clothing image at the previous step, we can use it as a reference to find the region that would be covered after try-on. Specifically, we create a clothing-elimination mask \\(M_{elm}\\) by union the original clothing region and the region covered by the transformed target clothes, which is then expanded using morphological dilation to cover the entire body. Moreover, when wearing tops, we make the clothing tucked or not in the try-on result by controlling \\(t \\in \\{0,1\\}\\) to determine whether to eliminate the overlap region of bottoms. As such, we can effectively obtain clothing-agnostic person representation without overly masking the irrelevant regions, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.733, + 0.47, + 0.753 + ], + "angle": 0, + "content": "\\[\nM _ {e l m} = \\mathcal {M} (S ^ {t} \\cup (C ^ {t o r s e} \\cup C ^ {l i m b}) - t \\cdot (C ^ {t o r s e} \\cap S ^ {b})), (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}\\) represents the morphological dilation; \\(S^t\\) and \\(S^b\\) denote top and bottom clothing segmentation respectively. Fig. 1 shows the results of different \\(t\\). After eliminating clothing information, we obtain clothing-agnostic images \\(I_a\\) and segmentation maps \\(S_a\\). To bridge the performance gap between the training and inference phases, we randomly mask the whole arms of human images with a \\(50\\%\\) probability during the training phase and apply our Clothing Elimination Policy only in the inference phase." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.669, + 0.107 + ], + "angle": 0, + "content": "3.5. Outfit Generator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.312 + ], + "angle": 0, + "content": "In the final stage, the Outfit Generator takes personal \\((I_{a}, S_{a}\\) and \\(K_{P})\\) and clothing \\((C^{torse}\\) and \\(C^{limb})\\) information as input and synthesizes the final try-on result. Since learning the shape and texture of clothing simultaneously is challenging for the generator, a segmentation network is commonly used to predict a segmentation mask as guidance to lead the try-on generator network [5, 20, 23]. However, they regarded segmentation prediction as an independent task and trained the segmentation network and try-on generator separately. In that case, the optimization of the two networks may be suboptimal. Moreover, the accumulated noise from the segmentation network may further deteriorate the performance of the final synthesis result [21]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.895, + 0.465 + ], + "angle": 0, + "content": "To address the problem, we propose a novel Outfit Generator by combining the segmentation network and try-on generator as one end-to-end training multi-task network. Specifically, \\( S_{a} \\), \\( K_{P} \\), \\( C^{torse} \\), \\( C^{limb} \\) are first concatenated and down-sampled to a lower resolution as the input of the segmentation network to predict segmentation mask \\( \\hat{S} \\). Afterward, \\( \\hat{S} \\) is up-sampled back to the original resolution and combined with \\( I_{a} \\) as well as \\( C^{torse} \\) and \\( C^{limb} \\) as the input of the synthesis step. Finally, the generator integrates all the information and renders the synthesis result \\( \\hat{I} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.471, + 0.892, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\hat {I}, \\hat {S} = G (I _ {a}, S _ {a}, K _ {P}, C ^ {t o r s e}, C ^ {l i m b}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.496, + 0.895, + 0.587 + ], + "angle": 0, + "content": "The objective function of the Outfit Generator includes the supervision of both the predicted segmentation mask and the try-on synthesis result. We apply focal loss \\(\\mathcal{L}_{\\text {focal }}\\) [25] for segmentation. For try-on results, we use losses including reconstruction loss \\(\\mathcal{L}_{\\text {rec }}\\), perceptual loss \\(\\mathcal{L}_{\\text {VGG }}\\), and adversarial loss \\(\\mathcal{L}_{\\text {adv }}\\). The overall loss functions are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.613, + 0.892, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {o v e r a l l}} = \\lambda_ {\\text {r e c}} \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {V G G} \\mathcal {L} _ {V G G} \\tag {4} \\\\ + \\lambda_ {f o c a l} \\mathcal {L} _ {f o c a l} + \\lambda_ {a d v} \\mathcal {L} _ {a d v}, \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.652, + 0.892, + 0.669 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c}} = \\| \\hat {I} - I \\| _ {1}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.674, + 0.892, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {V G G} = \\sum_ {k = 1} ^ {5} \\lambda_ {k} \\| \\phi_ {k} (\\hat {I}) - \\phi_ {k} (I) \\| _ {1}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.717, + 0.892, + 0.748 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f o c a l} = \\sum_ {i} - \\alpha_ {i} (1 - p _ {i}) ^ {\\gamma} \\log (p _ {i}), p _ {i} \\in \\hat {S}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.751, + 0.892, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a d v} = \\left\\| \\mathcal {D} (\\hat {I}) - \\mathbf {1} \\right\\| _ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.784, + 0.892, + 0.876 + ], + "angle": 0, + "content": "where \\(\\phi_{k}\\) represents the \\(k^{th}\\) layer output of VGG-19 network. \\(\\alpha_{i}\\) and \\(p_i\\) are respectively the importance weighting for the \\(i^{th}\\) class and the probabilities of pixels belonging to the \\(i^{th}\\) class. \\(\\gamma\\) is the focusing parameter. \\(\\lambda_{rec},\\lambda_{VGG},\\lambda_{focal}\\) and \\(\\lambda_{adv}\\) denote the balance weights. \\(\\mathcal{D}\\) is the multi-layer discriminators [32]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.892, + 0.902 + ], + "angle": 0, + "content": "3The comparison of two-stage and end-to-end training is in supplement." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "7517" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.21, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.126, + 0.267, + 0.143 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.16, + 0.47, + 0.326 + ], + "angle": 0, + "content": "Dataset. We created the Pure Cotton, a high-resolution \\((1024\\times 768)\\) outfit dataset, to mitigate the impact of noisy data on model performance. We strictly employed several rules to filter out unsuitable data, which are clearly outlined in the supplements, resulting in 16,428 frontal-view human models and in-shop clothing pairs, including 10,636 upper clothes and 5,792 bottom clothes. For training and testing, we divided the upper clothing set into 8,451 and 2,185 pairs, respectively, and the bottom clothing set into 4,626 and 1,166 pairs. To evaluate the model's robustness, we also experimented with the public Dress Code dataset [28]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.47, + 0.518 + ], + "angle": 0, + "content": "Implementation Details. To train Clothing Landmark Predictor (CLP) and Clothing Segmentation Network (CSN), we collected a small clothing dataset with 50 images for each clothing type, including shirts, turtleneck, hoodies, etc, and manually labeled landmarks and segmentation masks as ground truth. The architecture of CLP is based on the one in [3], and that of CSN is based on FCN-resnet50 [30]. Moreover, Outfit Generator is designed using an Unet and an encoder-decoder network. To ensure a fair comparison with previous studies, all figures are generated in the tucked-in setting unless stated otherwise. This setting is consistent with prior arts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.528, + 0.47, + 0.755 + ], + "angle": 0, + "content": "Baselines. We compare our proposed method with 3 relevant state-of-the-art baselines, including HR-VITON [23], VITON-HD [5], and PASTA-GAN [33], by training all the models with our self-collected Pure Cotton dataset and the public Dress Code dataset [28]. Specifically, HR-VITON and VITON-HD both represent SOTA high-resolution virtual try-on methods with different warping modules. HR-VITON warps clothes with appearance flows and designs a feature fusion network for eliminating the mismatch issue, while VITON-HD warps clothes with TPS and proposes an alignment-aware segmentation normalization method to deal with the misalignment problem. Meanwhile, PASTA-GAN represents another warping method, a patch-routed disentanglement module, which is more related to our proposed clothing deformation method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Evaluation metrics. We evaluate the model performance by three widely-used metrics for virtual try-on, including i) Structural Similarity (SSIM) [38], ii) Learned Perceptual Image Patch Similarity (LPIPS) [36], and iii) Fréchet Inception Distance (FID) [18]. SSIM measures the quality of three key features: luminance, contrast, and structure between the reconstruction results and the ground truth. LPIPS evaluates the perceptual similarity and FID is computed for evaluating the GAN performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.685, + 0.108 + ], + "angle": 0, + "content": "4.2. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.115, + 0.892, + 0.175 + ], + "angle": 0, + "content": "Fig. 4 and 5 show the visual comparisons in Pure Cotton and Dress Code datasets. Overall, our model achieves the most visually convincing high-resolution try-on results. The comparison analysis is discussed as follows." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.175, + 0.893, + 0.643 + ], + "angle": 0, + "content": "The previous warping methods, e.g., TPS (VITON-HD) and appearance flows (HR-VITON), did not consider clothing skeleton and sleeves segmentation, rendering them incapable of accurately synthesizing clothing patterns and features in respective positions, such as hit color mosaic (case I in Fig. 4) and logo in special positions (case II in Fig. 5). Additionally, these methods could not tackle regional clothing warping, such as arm akimbo in complex postures (case II in Fig. 4). In contrast, our model conducts the Clothing Landmark Predictor and Clothing Segmentation Network to obtain clothing structural information. Then, our proposed Landmark-guided Transformation helps to warp the clothes regionally. As such, our model could synthesize try-on results with clothing patterns and features located in the right position that closely follow the target clothing image, even in complex postures. On the other hand, PASTA-GAN conducts a regional clothing deformation method similar to our proposed method. However, there are three differences between PASTA-GAN and our model. Firstly, we split the clothes into 5 regions, i.e., torso, right/left upper arms, and right/left forearms, and kept the torso undivided. Secondly, we further consider clothing segmentation, including neckline and limb segmentation. Thirdly, PASTA-GAN deforms clothes from clothes on another human image, as demonstrated on the right-bottom side of the target clothes columns in Fig. 4. Due to these differences, PASTA-GAN would produce strange try-on results that highly follow the original clothing tied up and limb postures (the blue circle and the red arrow on Fig 4). In contrast, our model leverages clothing structural information to synthesize natural clothing wrinkles." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.644, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Additional Clothing Control. The elimination policy in our proposed approach allows for greater flexibility in virtual try-on, with options for both tucked-in and untucked clothing. As shown in Fig. 6, previous methods would limit the location of the upper clothes' lower boundary based on the original lower clothes' upper boundary. However, our approach allows for the clothing to be untucked based on the clothes' body length, providing additional cues for customers to judge the clothing style." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Partial Clothing Resizing: Apart from demonstrating COTTON's clothing size adjustment in Fig. 1, we also present its ability to adjust partial clothing sizes in Fig. 7. In the real world, clothing designers would adjust different sizes only for specific parts, e.g., body or sleeve length, instead of the entire garment. COTTON can smoothly manipulate partial clothing sizes by adjusting the corresponding landmarks without impacting the rest of the clothing." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7518" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.144, + 0.094, + 0.825, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.337, + 0.892, + 0.421 + ], + "angle": 0, + "content": "Figure 4. The visual comparison in the Pure Cotton dataset. The left two columns are the inputs and the rest present the results of baselines and our model. VITON-HD and HR-VITON did not consider sleeve segmentation, resulting in misaligned clothes with wrong pattern shifts (green dash line in case I) and broken warped clothes (case II). Moreover, without clothing skeletons, VITON-HD and HR-VITON failed to maintain pattern proportions (red and blue double-headed arrows in case I). PASTA-GAN also had issues in case I, where the partially tucked-in clothing shape from the original human image was directly copied to the try-on result, and in case II, where the try-on result had two additional limbs in front of the abdomen due to the human, representing target clothes, crossing their arms over the chest." + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.426, + 0.822, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.672, + 0.89, + 0.701 + ], + "angle": 0, + "content": "Figure 5. The visual comparison in the Dress Code dataset. In case I, our Clothing Elimination Policy well retains valuable human features, e.g., bulging muscles. For case II, our Landmark-guided Transformation aligns textures better based on the obtained clothing structure." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.714, + 0.464, + 0.796 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.799, + 0.454, + 0.813 + ], + "angle": 0, + "content": "Figure 6. Additional clothing control for tucked-in or untucked." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.829, + 0.27, + 0.845 + ], + "angle": 0, + "content": "4.3. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Table 1 compares COTTON's performance with SOTAs in terms of SSIM, LPIPS, and FID. We evaluated SSIM and LPIPS on 2,000 random paired test sets from our Pure Cot" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.895 + ], + "angle": 0, + "content": "ton dataset and 1,800 test sets from the Dress Code dataset. FID was evaluated on 2,000 and 1,800 unpaired test sets from Pure Cotton and Dress Code datasets, respectively. Human images wearing original clothing served as ground truth for FID calculation. The results show that COTTON consistently outperforms the baselines across all evaluation metrics on both datasets, with at least \\(18.6\\%\\) improvement in LPIPS on Pure Cotton dataset and \\(41.1\\%\\) in FID on Dress Code dataset. The results manifest that our method excels in generating detailed contents on Pure Cotton dataset and robust results on Dress Code dataset. Notably, COTTON shows lower performance with untucked clothes, as it ob" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7519" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.094, + 0.237, + 0.105 + ], + "angle": 0, + "content": "Sleeves length adjustment" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.106, + 0.449, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.291, + 0.225, + 0.302 + ], + "angle": 0, + "content": "Body length adjustment" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.302, + 0.448, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.457, + 0.432, + 0.471 + ], + "angle": 0, + "content": "Figure 7. Visualization of partial clothing adjustment." + }, + { + "type": "table_caption", + "bbox": [ + 0.118, + 0.487, + 0.428, + 0.499 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on the test dataset." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.5, + 0.481, + 0.671 + ], + "angle": 0, + "content": "
MethodDatasetSSIM↑LPIPS↓FID↓
PASTA-GAN0.9180.121529.43
VITON-HD0.9440.056617.16
HR-VITONSelf-collected0.9480.038711.16
Ours (untucked)Pure Cotton0.9560.034910.96
Ours (tucked-in)0.9580.031510.17
PASTA-GAN0.8000.306633.92
VITON-HD0.9270.090622.53
HR-VITONPublic0.9270.075520.12
Ours (untucked)Dress Code0.9490.048212.58
Ours (tucked-in)0.9530.043811.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.69, + 0.468, + 0.718 + ], + "angle": 0, + "content": "Table 2. The user study shows that COTTON is the most photorealistic and remains the most human and clothing characteristics." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.718, + 0.451, + 0.815 + ], + "angle": 0, + "content": "
MethodPASTA-GANVITON-HDHR-VITONCOTTON(Ours)
Photo-realistic9.08%7.56%29.82%53.54%
Try-on accuracy8.66%4.62%27.40%59.32%
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.824, + 0.468, + 0.884 + ], + "angle": 0, + "content": "scures the upper portion of lower clothing, causing the try-on results to differ from the ground truth. Nevertheless, Fig. 6 demonstrates that COTTON can generate more visually appealing results when clothing is untucked." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Moreover, we conduct a user study on the Pure Cotton" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.439 + ], + "angle": 0, + "content": "dataset to further evaluate the visual quality by humans. The experiment involved 127 volunteers. We randomly sampled 30 image sets and divided them equally into two groups for evaluating: i) photo-realistic quality and ii) try-on accuracy. To evaluate the photo-realistic quality, volunteers would see four try-on results (without input information) synthesized by four different models, and choose the most photo-realistic one. Furthermore, for evaluating try-on accuracy, we provide a source human image, one target try-on clothing, and four try-on results synthesized by four different models. Based on the source human and the target clothes, volunteers should select the best try-on result. Table 2 shows that COTTON outperforms all other baselines in both photo-realistic quality and try-on accuracy and achieves even higher ratings for try-on accuracy. Volunteers reported that they mainly used realistic reproduced clothing patterns, including loose/fit torso and accurate sleeve length, with fewer artifacts as the criteria to determine the better model. Our proposed Clothing Landmark Predictor and Clothing Segmentation Network effectively retain clothing characteristics such as patterns and sleeve length, contributing significantly to COTTON's superior performance over other state-of-the-art models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.461, + 0.618, + 0.476 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.73 + ], + "angle": 0, + "content": "This paper presents COTTON, a clothing-oriented transformation try-on network that synthesizes high-resolution virtual try-on results. By considering the clothing structure and leveraging it, COTTON improves the clothing transformation quality and allows for size adjustment during clothing try-on. Additionally, the proposed elimination policy enables COTTON to try on outfits untucked or not while preserving valuable human characteristics. The experiments demonstrate that COTTON outperforms the state-of-the-art virtual try-on works at \\(1024 \\times 768\\) resolution, both qualitatively and quantitatively. In future work, we plan to improve the accessories segmentation and extend the system to try on more clothing categories, such as shoes, hats, and accessories. Overall, the proposed approach shows great promise for improving the online shopping experience and reducing the need for physical try-on." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.752, + 0.66, + 0.769 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We would like to express our gratitude to Josephine, Wendy, Hans, Joyce, Leslie and May from the Epoch Foundation for their valuable support. This work was supported in part by the National Science and Technology Council of Taiwan under Grants NSTC-109-2221-E-009-114-MY3, NSTC-112-2221-E-A49-094-MY3, NSTC-109-2223-E-009-002-MY3, NSTC-112-2634-F-002-002-MBK and NSTC-112-2628-E-002-033-MY4." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7520" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Alakh Aggarwal, Jikai Wang, Steven Hogue, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. Layered-garment net: Generating multiple implicit garment layers from a single image. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Shuai Bai, Huiling Zhou, Zhikang Li, Chang Zhou, and Hongxia Yang. Single stage virtual try-on via deformable attention flows. In European Conference on Computer Vision (ECCV), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.47, + 0.367 + ], + "angle": 0, + "content": "[4] Chieh-Yun Chen, Ling Lo, Pin-Jui Huang, Hong-Han Shuai, and Wen-Huang Cheng. Fashionmirror: Co-attention feature-remapping virtual try-on with sequential template poses. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.47, + 0.437 + ], + "angle": 0, + "content": "[5] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. VITON-HD: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.507 + ], + "angle": 0, + "content": "[6] Ayush Chopra, Rishabh Jain, Mayur Hemani, and Balaji Krishnamurthy. Zflow: Gated appearance flow-based virtual try-on with 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.47, + 0.577 + ], + "angle": 0, + "content": "[7] Chien-Lung Chou, Chieh-Yun Chen, Chia-Wei Hsieh, Hong-Han Shuai, Jiaying Liu, and Wen-Huang Cheng. Template-free try-on image synthesis via semantic-guided optimization. IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[8] Aiyu Cui, Sen He, Tao Xiang, and Antoine Toisoul. Learning garment densepose for robust warping in virtual try-on. In arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.471, + 0.69 + ], + "angle": 0, + "content": "[9] Aiyu Cui, Daniel McKee, and Svetlana Lazebnik. Dressing in order: Recurrent person image generation for pose transfer, virtual try-on and outfit editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[10] Haoye Dong, Xiaodan Liang, Xiaohui Shen, Bochao Wang, Hanjiang Lai, Jia Zhu, Zhiting Hu, and Jian Yin. Towards multi-pose guided virtual try-on network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[11] Xin Dong, Fuwei Zhao, Zhenyu Xie, Xijin Zhang, Kang Du, Min Zheng, Xiang Long, Xiaodan Liang, and Jianchao Yang1. Dressing in the wild by watching dance videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[12] Ruili Feng, Cheng Ma, Chengji Shen, Xin Gao, Zhenjiang Liu, Xiaobo Li, Kairi Ou, Deli Zhao, and Zheng-Jun Zha. Weakly supervised high-fidelity clothing model generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[13] Yuying Ge, Yibing Song, Ruimao Zhang, Chongjian Ge, Wei Liu, and Ping Luo. Parser-free virtual try-on via distilling appearance flows. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[14] Oshri Halimi, Tuur Stuyck, Donglai Xiang, Timur Bagautdinov, He Wen, Ron Kimmel, Takaaki Shiratori, Chenglei Wu, Yaser Sheikh, and Fabian Prada. Pattern-based cloth registration and sparse-view animation. ACM Trans. Graph., 41(6), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[15] Xintong Han, Xiaojun Hu, Weilin Huang, and Matthew R. Scott. ClothFlow: a flow-based model for clothed person generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[16] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S. Davis. VITON: An image-based virtual try-on network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[17] Sen He, Yi-Zhe Song, and Tao Xiang. Style-based global appearance flow for virtual try-on. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Proceedings of the International Conference on Neural Information Processing Systems (NIPS), 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[19] Chia-Wei Hsieh, Chieh-Yun Chen, Chien-Lung Chou, Hong-Han Shuai, and Wen-Huang Cheng. Fit-me: Image-based virtual try-on with arbitrary poses. In IEEE International Conference on Image Processing (ICIP), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[20] Chia-Wei Hsieh, Chieh-Yun Chen, Chien-Lung Chou, Hong-Han Shuai, Jiaying Liu, and Wen-Huang Cheng. FashionOn: Semantic-guided image-based virtual try-on with detailed human and clothing information. In Proceedings of the \\(27^{th}\\) ACM International Conference on Multimedia (ACM MM), 2019. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.617, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[21] Thibaut Issenhuth, Jérémie Mary, and Clément Calauzènes. Do not mask what you do not need to mask: a parser-free virtual try-on. In European Conference on Computer Vision (ECCV), 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.717 + ], + "angle": 0, + "content": "[22] Maria Korosteleva and Sung-Hee Lee. Neuraltailor: Reconstructing sewing pattern structures from 3d point clouds of garments. ACM Trans. Graph., 41(4), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[23] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[24] Ren Li, Benoit Guillard, Edoardo Remelli, and Pascal Fua. Dig: Draping implicit garment over the human body. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7521" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[26] Xinqi Liu, Jituo Li, Guodong Lu, Dongliang Zhang, and Shihai Xing. Robust and automatic clothing reconstruction based on a single rgb image. Computers & Graphics, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.472, + 0.217 + ], + "angle": 0, + "content": "[27] S. Majithia, S. N. Parameswaran, S. Babar, V. Garg, A. Srivastava, and A. Sharma. Robust 3d garment digitization from monocular 2d images for 3d virtual try-on systems. In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[28] Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, and Rita Cucchiara. Dress Code: High-Resolution Multi-Category Virtual Try-On. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.289, + 0.47, + 0.355 + ], + "angle": 0, + "content": "[29] Assaf Neuberger, Eran Borenstein, Bar Hilleli, Eduard Oks, and Sharon Alpert. Image based virtual try-on network from unpaired data. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.357, + 0.47, + 0.412 + ], + "angle": 0, + "content": "[30] Evan Shelhamer, Jonathan Long, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.47, + 0.468 + ], + "angle": 0, + "content": "[31] Bochao Wang, Huabin Zheng, Xiaodan Liang, Yimin Chen, Liang Lin, and Meng Yang. Toward characteristic-preserving image-based virtual try-on network. In European Conference on Computer Vision (ECCV), 2018. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.538 + ], + "angle": 0, + "content": "[32] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.47, + 0.607 + ], + "angle": 0, + "content": "[33] Zhenyu Xie, Zaiyu Huang, Fuwei Zhao, Haoye Dong, Michael Kampffmeyer, and Xiaodan Liang. Towards scalable unpaired virtual try-on via patch-routed spatially-adaptive GAN. In Advances in Neural Information Processing Systems (NIPS), 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.47, + 0.663 + ], + "angle": 0, + "content": "[34] Han Yang, Xinrui Yu, and Ziwei Liu. Full-range virtual try-on with recurrent tri-level transform. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.47, + 0.733 + ], + "angle": 0, + "content": "[35] Han Yang, Ruimao Zhang, Xiaobao Guo, Wei Liu, Wang-meng Zuo, and Ping Luo. Towards photo-realistic virtual try-on by adaptively generating \\(\\leftrightarrow\\) preserving image content. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[36] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[37] Fuwei Zhao, Zhenyu Xie, Michael Kampffmeyer, Haoye Dong, Songfang Han, Tianxiang Zheng, Tao Zhang, and Xiaodan Liang. M3d-vton: A monocular-to-3d virtual try-on network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[38] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "to structural similarity. IEEE Transactions on Image Processing (TIP), 2004. 6" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7522" + } + ] +] \ No newline at end of file diff --git a/2023/Size Does Matter_ Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network/2e5e4941-f171-4e8c-be33-35affe248d0e_origin.pdf b/2023/Size Does Matter_ Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network/2e5e4941-f171-4e8c-be33-35affe248d0e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..931d7b84a96c4ec4f54bd40c1264068e2ffebf45 --- /dev/null +++ b/2023/Size Does Matter_ Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network/2e5e4941-f171-4e8c-be33-35affe248d0e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:658d0793e263d55c5f4fb7cf763aa54a246db5f384b03ec7ba4b3aebde82b91c +size 1108466 diff --git a/2023/Size Does Matter_ Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network/full.md b/2023/Size Does Matter_ Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network/full.md new file mode 100644 index 0000000000000000000000000000000000000000..60cdd5eb3eeacab22433924ed6a3ea041ff5708f --- /dev/null +++ b/2023/Size Does Matter_ Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network/full.md @@ -0,0 +1,234 @@ +# Size Does Matter: Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network + +Chieh-Yun Chen $^{1,2*}$ Yi-Chung Chen $^{1,3*}$ Hong-Han Shuai $^{2}$ Wen-Huang Cheng $^{3}$ + +$^{1}$ Stylins.ai $^{2}$ National Yang Ming Chiao Tung University $^{3}$ National Taiwan University + +![](images/82f50f4453d85f38c84e709171b798272971efcaee9572fd1b7a2f92430550ae.jpg) +Figure 1. Multi-garment virtual try-on results with different sizes (tucked or untucked). $r$ is the ratio of torso length over shoulder width, which can be decided by customers in online shopping scenarios. Padding blue backgrounds for better visualizing clothing skeletons. + +# Abstract + +Virtual try-on tasks aim at synthesizing realistic try-on results by trying target clothes on humans. Most previous works relied on the Thin Plate Spline or appearance flows to warp clothes to fit human body shapes. However, both approaches cannot handle complex warping, leading to over distortion or misalignment. Furthermore, there is a critical unaddressed challenge of adjusting clothing sizes for try-on. To tackle these issues, we propose a Clothing-Oriented Transformation Try-On Network (COTTON). COTTON leverages clothing structure with landmarks and segmentation to design a novel landmark-guided transformation for precisely deforming clothes, allowing for size adjustment during try-on. Additionally, to properly remove the clothing region from the human image without losing significant human characteristics, we propose a clothing elimination policy based on both transformed clothes and human segmentation. This method enables users to try on clothes tucked-in or untucked while retaining more human characteristics. Both qualitative and quantitative results show that COTTON outperforms the state-of-the-art high-resolution virtual try-on approaches. All the code is available at https://github.com/cotton6/COTTON-size-does-matter. + +# 1. Introduction + +Image-based virtual try-on replaces clothing items on a person with the desired ones, creating realistic try-on results and lowering costs associated with on-model photos for the e-commerce industry1. Additionally, it enables customers to use virtual dressing rooms when shopping online, potentially enhancing the e-commerce experience and increasing conversion rates2. As online shopping becomes popular, the virtual try-on tasks have received more and more attention [2,4,5,9,13,17,23,28,29,33,35]. For example, to extend the virtual try-on resolution from low $(256\times 192)$ to high $(1024\times 768)$ , the most obvious difficulty is the misalignment issue. [5] proposed the alignment-aware segment normalization to remove the misleading information in the misaligned area. Additionally, [23] further performs appearance flow-based warping and segmentation map generation simultaneously to tackle misalignment. + +Despite the considerable progress made in previous works, there are still some challenges that have not been adequately addressed, as outlined below. i) Handling complex warping without misalignment: General image + +based virtual try-on frameworks [5, 16, 21, 28, 31] typically employ a clothing deformation module, such as the Thin Plate Spline (TPS) method, to align clothing images with the target human pose. However, research [6, 15, 23] has found that TPS may not effectively handle complex warping when different garment regions require different deformations. In attempts to improve warping results, they have replaced TPS with dense appearance flows. However, when the transformation between the garment and corresponding body parts is significant (as illustrated in case II in Fig. 4), the performance of both TPS (VITON-HD [5]) and flow-based (HR-VITON [23]) methods deteriorates drastically. Hence, existing methods still can't properly address the challenge of clothing deformation. ii) Adjusting clothing sizes: Previous works predicted the try-on segmentation conditioned on the clothing image by only considering the shape of the clothes without the scale information. Thus, given a clothing image, it is impossible for previous works to change the clothing size. This limitation severely restricts the practical application of virtual try-on because people often try on different sizes in the fitting room. iii) Appropriate clothing elimination policy: Previous works [5, 23, 28] rely on segmentation maps to eliminate the original clothing region and utilize the remaining area as a guide for trying on target clothes. However, this method can lead to either excessive or insufficient removal of the original image. For instance, when prior works try on upper clothes, the lower part of the upper clothes remains intact, which limits the length of the upper clothes, as shown in Fig. 6. The lower part of the upper clothing is forced to be tucked in, which reduces the model's ability to generate complete upper clothes, thereby reducing the practicality of virtual try-on. Additionally, previous works remove the entire arm of input human images, which eliminates essential human characteristics such as tattoos and arm width that should be preserved in the final output, as illustrated in case I of Fig. 5. Therefore, developing an appropriate clothing elimination strategy that preserves crucial information while removing the clothing region is a significant challenge. + +To address these challenges, we propose a simple yet powerful approach called Clothing-Oriented Transformation Try-On Network (COTTON). Specifically, COTTON first exploits geometric information with Clothing Landmark Predictor and Clothing Segmentation Network to predict clothing landmarks and segmentation masks respectively. To overcome the first challenge posed by complex warping, we propose the Landmark-guided Transformation that first separates clothes into sub-parts via the segmentation mask and then uses clothing landmarks to estimate homography matrices that fit these sub-parts to the target human pose. To solve the second challenge, we introduce a clothing landmarks adjustment approach that allows users to change clothing sizes. Adjustment of landmarks will al + +ter the homography matrices and then cause clothing size changes, as demonstrated in Fig. 1, which remarkably enhances the practicality of virtual try-on. To address the third challenge, we employ the transformed clothing images to identify and remove the areas that would be covered by the clothes. This enables our proposed Clothing Elimination Policy to effectively eliminate the clothing region while preserving important details and offer flexibility in tucking the clothes or not, as shown in Fig. 6. Extensive experiments show that COTTON achieves superior results than state-of-the-arts both quantitatively and qualitatively. We summarize our contributions as follows: + +- We propose a Clothing-Oriented Transformation Try-On Network (COTTON), which improves clothing transformation quality and addresses complex warping misalignments by leveraging clothing geometry. +- We introduce adjustable clothing landmarks to provide clothing size information based only on images. To the best of our knowledge, this is the first work enabling 2D-based virtual try-on with different clothing sizes for approaching real-world try-on. +- To preserve critical information, we propose a Clothing Elimination Policy to properly remove clothing information while retaining valuable human characteristics, and offer flexibility in tucking the clothes or not. +- Extensive experiments on the Dress Code dataset [28] show that our model significantly outperforms SOTAs, e.g., at least $41.1\%$ improvement in terms of FID. + +# 2. Related Works + +# 2.1. Clothing Deformation + +Virtual try-on requires clothing deformation to align patterns and preserve clothing details. Thin Plate Spline (TPS) warping is widely used in virtual try-on tasks [5, 16, 21, 28, 31], but it has limited ability to model geometric changes and result in unnatural deformations. Thus, [2, 6, 11, 13, 15, 17, 23] conducted appearance flows as an alternative to TPS, but misalignment still has not be well addressed. To address misalignment, [5] improved TPS by eliminating irrelevant clothing texture information in misaligned regions. Meanwhile, [23] improved the deformation by designing pathways to jointly predict appearance flows and segmentation features. However, neither TPS nor appearance flow can deform partial regions of clothing without affecting other parts, e.g., warping the sleeves to present arm akimbo without affecting the torso part of the clothes, since they do not have the structural information of the clothes. To address this issue, [33] introduced normalized patches to learn spatial-agnostic clothing features but this approach may destroy the completeness of clothes + +![](images/4923e65b2fc850a5929ae0bb9a7352a5eb97006251701116aa2700d9a0632a96.jpg) +Figure 2. Overview of COTTON. Given a clothing image $C$ , COTTON uses Clothing Landmark Predictor and Clothing Segmentation Network to predict clothing landmarks $K_{C}$ and a clothing segmentation mask $M_{C}$ . $C$ is then warped by Landmark-guided Transformation with homography derived from human keypoints $K_{P}$ and $K_{C}$ to obtain the transformed images $C^{torso}$ and $C^{limb}$ . $r$ is used to adjust $K_{C}$ at this step for changing clothing size. The clothing elimination step uses $C^{torso}$ and $C^{limb}$ to mask human image $I$ and segmentation mask $S$ , resulting in clothing-agnostic image $I_{a}$ and segmentation $S_{a}$ . The variable $t$ controls whether clothing is tucked in or not. Finally, Outfit Generator aggregates the previous outputs to predict segmentation mask $\hat{S}$ and synthesize the final try-on result $\hat{I}$ . + +due to lack of semantic information. In contrast, our proposed COTTON separates clothes based on landmarks and segmentation masks to maintain clothing completeness and aligns clothes based on specific landmarks to address misalignment. Besides, [8] is a concurrent work. Both of our works improve garment warping by human-garment correspondence. While [8] relies on the computationally intensive DensePose method, our proposed $LT$ method is more efficient and enables clothing size adjustment. + +# 2.2. Virtual Try-on + +Virtual try-on methods can be categorized into 3D-based methods [1,14,22,24,26,27,37] and 2D-based methods [5,6, 7,8,9,10,11,12,15,16,17,19,23,28,29,31,33,34,35]. Since 3D-based methods are time-consuming and real-world scenarios require fast inference speeds, we focus on 2D-based methods. [16] proposed a coarse-to-fine network to synthesize image-based virtual try-on results based on TPS warping. However, try-on results synthesized by [16,31] do not provide a clear silhouette of human body parts since they consider the human to be a general mask instead of considering the semantic information, e.g., legs, head, limbs, etc. To tackle this problem, [10,20] proposed the first semantic-guided virtual try-on network, which predicted the target semantic segmentation to provide human structural information for the network to synthesize photo-realistic body parts. Afterward, many try-on methods [5,6,7,9,23,29,35] are designed with an individual stage for predicting the target semantic segmentation before synthesizing virtual try-on results. However, the individual stage for predicting target semantic segmentation would lead to error accumulation as pointed out by [21]. In this paper, we propose an end-to + +end Outfit Generator that updates semantic segmentation and try-on results simultaneously, and introduce a Clothing Elimination Policy that retains valuable human characteristics while removing original clothing information. + +# 3. Methodology + +Given a human image $I$ and a clothing image $C$ , the goal of virtual try-on is to obtain a synthetic image $\hat{I}$ , which shows the human of $I$ wearing the clothing of $C$ . We design a Clothing-Oriented Transformation Try-On Network (COTTON) for generating high-resolution virtual try-on results. Fig. 2 provides an overview of the proposed COTTON. We use the Clothing Landmark Predictor and the Clothing Segmentation Network to respectively capture clothing landmarks and a clothing segmentation map of $C$ . Afterward, we separate the clothes into torso part $C^{torse}$ and limbs part $C^{limb}$ according to the clothing segmentation map. The separation of clothes enables COTTON to treat each part differently, and thus makes it adapt to handling complex warping well. To align the clothes with the target human pose, the clothing landmarks are adjusted with a ratio $r$ to fit human body proportions and used to calculate homography matrices. COTTON then warps the clothes with affine transformation. The warped clothes shed light on how to properly remove the clothing region of image $I$ without losing necessary information. We use a user-determined variable $t$ to control whether to eliminate the overlap region of clothes. In this case, we can generate tucked-in or untucked try-on results by choosing the value of $t$ . Finally, the Outfit Generator takes both human and clothing representation to generate the try-on result $\hat{I}$ . + +# 3.1. Clothing Landmark Predictor (CLP) + +One of the most challenging problems of virtual try-on is to transform a clothing image $C$ to fit the target human pose. To solve this challenge, several approaches have been proposed in previous works, including hyperparameter learning of TPS warping [5, 16, 28, 31], and flow-based deep deformation [6, 15, 23]. However, these approaches implicitly learn the underlying geometric information from the training dataset and only achieve mediocre results when handling complex warping. We argue that explicit geometric information plays the most critical role in virtual try-on and can be leveraged to further improve performance. For example, by thoroughly understanding the clothes' geometry, humans can easily capture the relation between clothes and humans, even if the clothes are with different textures and styles. Therefore, we propose the Clothing Landmark Predictor (CLP) to extract explicit geometric information. + +Specifically, to exploit the geometric information of clothes, our proposed CLP is trained to predict a set of landmarks $K_{C}$ , which correlate to different parts of human pose $K_{P}$ , where $K_{P}$ can be obtained from off-the-shelf human pose estimation methods, e.g., [3]. The ground truth of $K_{C}$ is labeled manually, and the number of landmarks varies with regard to the clothing types. For upper clothing, $|K_{C}| = 10$ , including neck, shoulders, elbows, wrists, and hips. These landmarks explicitly indicate the relation between clothes and human pose, and therefore can be used as a reference to generate transformation matrices. Given clothing image $C$ , the CLP is trained to output clothing joint heatmaps $\hat{M}_{JH}$ and Part Affinity Field $\hat{M}_{PAF}$ . Afterward, the predicted landmarks $K_{C}$ can be obtained from $\hat{M}_{JH}$ and $\hat{M}_{PAF}$ by using the Hungarian algorithm. + +# 3.2. Clothing Segmentation Network + +Each part of clothes has distinct properties that should be taken into consideration during transformation. For example, the sleeves of clothes usually suffer severe distortion during transformation while the torso part only undergoes a mild distortion. Another common noise in clothing try-on is the region around the neckline that must be concealed when the garment is worn. This region always introduces undesired variation into try-on results, and as such, should be eliminated. Though the Clothing Landmark Predictor provides the connection between the clothes and human poses, it is difficult to identify different regions with only a few landmarks. To find the explicit boundary of each part, pixel-level semantic segmentation is necessary. Therefore, we propose a Clothing Segmentation Network to predict the clothing segmentation mask $M_C$ , which separates the clothing into three sub-region, including torso part, sleeve part, and invisible part around the neckline. As such, each sub-region can be treated properly based on its characteristics. + +![](images/fd7d1e9ad3e08136529b6dd189d740d9fecb6a19a7dfca0885bc8d4dcbba3a23.jpg) +Figure 3. Clothing transformation steps. For the torso part, we use keypoints including shoulders and hips to calculate matrix $H_{c \to p}^{torso}$ . For the limbs part, we first create bounding boxes around limbs and use them to calculate matrices $H_{c \to p}^{limb,i}$ , $i \in \{1,2,3,4\}$ . + +# 3.3. Landmark-guided Transformation + +Here, we introduce how to leverage the landmarks $K_{C}$ and segmentation mask $M_{C}$ to achieve high-fidelity clothing transformation. We first remove the undesired part around the neckline such as the green region of $M_{C}$ in Fig. 2, which is invisible in the final try-on result, via the clothing segmentation map. We then separate the remaining part into two parts, including the torso part and the limbs part. Let $K_{C}^{\text{torso}}$ and $K_{P}^{\text{torso}}$ denote the clothing and human keypoints belonging to the torso part, such as shoulders and hips. For aligning the torso part, we compute a homography matrix $H_{c\rightarrow p}^{\text{torso}} \in R^{3\times 3}$ that can project clothing landmarks to human keypoints, as follows: + +$$ +k _ {p} = H _ {c \rightarrow p} ^ {t o r s o} k _ {c}, \quad \forall k _ {c} \in K _ {C} ^ {t o r s o}, \forall k _ {p} \in K _ {P} ^ {t o r s o}. (1) +$$ + +Afterward, this perspective transformation is applied to get the transformed torso part $C^{torso}$ . On the other hand, as the limbs part usually suffers from severe distortion, it requires a more sophisticated procedure. Therefore, instead of directly warping the whole limbs part, we break the limbs part into four sub-limbs based on the clothing landmarks. Take a hoodie as an example in Fig. 3. The limbs part is divided into two upper arms and two forearms. We create bounding boxes for each sub-limb based on clothing landmarks and human keypoints to obtain the area of each sub-limb. The corners of bounding boxes are then used to compute 4 homography matrices, $H_{c\rightarrow p}^{limb,i}, i\in \{1,2,3,4\}$ , which project each sub-limb to the target human. Finally, the four sub-limbs are collected as the warped limbs part $C^{limb}$ . + +It is worth noting that the predicted clothing landmarks $K_{C}$ only provide a clothes-human relation in the general case. However, since body proportion varies from person to person, the relative size of the same clothing on different persons may change. To generate a more realistic try-on result, we can adjust the clothing landmark $K_{C}$ to fit the body proportion of the target human. For instance, if two people have the same shoulder width but different torso lengths, the portion of their bodies covered by clothing should be different. In that case, the predicted landmarks are adjusted + +accordingly to fit their body proportions. Let $r$ be the ratio of torso length over shoulder width, which can be derived from extra body information provided by customers in online shopping scenarios. Fig. 1 visualizes the results of clothing landmarks adjustment with different $r$ , which shows that the region covered by the clothes is varied with regards to different $r$ , and can therefore generate try-on results with different clothing sizes. + +# 3.4. Clothing Elimination Policy + +Since the goal of the virtual try-on task is to replace the clothes of a given person image $I$ with a target one, the information about the original clothes should be eliminated. One common approach is eliminating the clothing according to the human segmentation $S$ . However, [5] pointed out that simply removing the clothing from image $I$ leads to a performance drop for paired data training due to information leakage about the clothing shape, and therefore proposed a clothing-agnostic person representation to address this problem. Nevertheless, this approach may lead to valuable information loss, e.g., arm width in case I of Fig. 5, when whole arms are directly masked during inference. + +Hence, we propose a Clothing Elimination Policy for the inference phase to preserve the important information while generating clothing-agnostic person representation. The regions that ought to be eliminated in image $I$ are i) the regions of original clothes and ii) the regions covered by target clothes $C$ when worn. Note that the latter is hard to measure with previous methods. Fortunately, since we have already obtained the reliable warped clothing image at the previous step, we can use it as a reference to find the region that would be covered after try-on. Specifically, we create a clothing-elimination mask $M_{elm}$ by union the original clothing region and the region covered by the transformed target clothes, which is then expanded using morphological dilation to cover the entire body. Moreover, when wearing tops, we make the clothing tucked or not in the try-on result by controlling $t \in \{0,1\}$ to determine whether to eliminate the overlap region of bottoms. As such, we can effectively obtain clothing-agnostic person representation without overly masking the irrelevant regions, i.e., + +$$ +M _ {e l m} = \mathcal {M} (S ^ {t} \cup (C ^ {t o r s e} \cup C ^ {l i m b}) - t \cdot (C ^ {t o r s e} \cap S ^ {b})), (2) +$$ + +where $\mathcal{M}$ represents the morphological dilation; $S^t$ and $S^b$ denote top and bottom clothing segmentation respectively. Fig. 1 shows the results of different $t$ . After eliminating clothing information, we obtain clothing-agnostic images $I_a$ and segmentation maps $S_a$ . To bridge the performance gap between the training and inference phases, we randomly mask the whole arms of human images with a $50\%$ probability during the training phase and apply our Clothing Elimination Policy only in the inference phase. + +# 3.5. Outfit Generator + +In the final stage, the Outfit Generator takes personal $(I_{a}, S_{a}$ and $K_{P})$ and clothing $(C^{torse}$ and $C^{limb})$ information as input and synthesizes the final try-on result. Since learning the shape and texture of clothing simultaneously is challenging for the generator, a segmentation network is commonly used to predict a segmentation mask as guidance to lead the try-on generator network [5, 20, 23]. However, they regarded segmentation prediction as an independent task and trained the segmentation network and try-on generator separately. In that case, the optimization of the two networks may be suboptimal. Moreover, the accumulated noise from the segmentation network may further deteriorate the performance of the final synthesis result [21]. + +To address the problem, we propose a novel Outfit Generator by combining the segmentation network and try-on generator as one end-to-end training multi-task network. Specifically, $S_{a}$ , $K_{P}$ , $C^{torse}$ , $C^{limb}$ are first concatenated and down-sampled to a lower resolution as the input of the segmentation network to predict segmentation mask $\hat{S}$ . Afterward, $\hat{S}$ is up-sampled back to the original resolution and combined with $I_{a}$ as well as $C^{torse}$ and $C^{limb}$ as the input of the synthesis step. Finally, the generator integrates all the information and renders the synthesis result $\hat{I}$ . + +$$ +\hat {I}, \hat {S} = G (I _ {a}, S _ {a}, K _ {P}, C ^ {t o r s e}, C ^ {l i m b}). \tag {3} +$$ + +The objective function of the Outfit Generator includes the supervision of both the predicted segmentation mask and the try-on synthesis result. We apply focal loss $\mathcal{L}_{\text {focal }}$ [25] for segmentation. For try-on results, we use losses including reconstruction loss $\mathcal{L}_{\text {rec }}$ , perceptual loss $\mathcal{L}_{\text {VGG }}$ , and adversarial loss $\mathcal{L}_{\text {adv }}$ . The overall loss functions are as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {o v e r a l l}} = \lambda_ {\text {r e c}} \mathcal {L} _ {\text {r e c}} + \lambda_ {V G G} \mathcal {L} _ {V G G} \tag {4} \\ + \lambda_ {f o c a l} \mathcal {L} _ {f o c a l} + \lambda_ {a d v} \mathcal {L} _ {a d v}, \\ \end{array} +$$ + +$$ +\mathcal {L} _ {\text {r e c}} = \| \hat {I} - I \| _ {1}, \tag {5} +$$ + +$$ +\mathcal {L} _ {V G G} = \sum_ {k = 1} ^ {5} \lambda_ {k} \| \phi_ {k} (\hat {I}) - \phi_ {k} (I) \| _ {1}, \tag {6} +$$ + +$$ +\mathcal {L} _ {f o c a l} = \sum_ {i} - \alpha_ {i} (1 - p _ {i}) ^ {\gamma} \log (p _ {i}), p _ {i} \in \hat {S}, \tag {7} +$$ + +$$ +\mathcal {L} _ {a d v} = \left\| \mathcal {D} (\hat {I}) - \mathbf {1} \right\| _ {2}, \tag {8} +$$ + +where $\phi_{k}$ represents the $k^{th}$ layer output of VGG-19 network. $\alpha_{i}$ and $p_i$ are respectively the importance weighting for the $i^{th}$ class and the probabilities of pixels belonging to the $i^{th}$ class. $\gamma$ is the focusing parameter. $\lambda_{rec},\lambda_{VGG},\lambda_{focal}$ and $\lambda_{adv}$ denote the balance weights. $\mathcal{D}$ is the multi-layer discriminators [32]. + +# 4. Experiments + +# 4.1. Experimental Setup + +Dataset. We created the Pure Cotton, a high-resolution $(1024\times 768)$ outfit dataset, to mitigate the impact of noisy data on model performance. We strictly employed several rules to filter out unsuitable data, which are clearly outlined in the supplements, resulting in 16,428 frontal-view human models and in-shop clothing pairs, including 10,636 upper clothes and 5,792 bottom clothes. For training and testing, we divided the upper clothing set into 8,451 and 2,185 pairs, respectively, and the bottom clothing set into 4,626 and 1,166 pairs. To evaluate the model's robustness, we also experimented with the public Dress Code dataset [28]. + +Implementation Details. To train Clothing Landmark Predictor (CLP) and Clothing Segmentation Network (CSN), we collected a small clothing dataset with 50 images for each clothing type, including shirts, turtleneck, hoodies, etc, and manually labeled landmarks and segmentation masks as ground truth. The architecture of CLP is based on the one in [3], and that of CSN is based on FCN-resnet50 [30]. Moreover, Outfit Generator is designed using an Unet and an encoder-decoder network. To ensure a fair comparison with previous studies, all figures are generated in the tucked-in setting unless stated otherwise. This setting is consistent with prior arts. + +Baselines. We compare our proposed method with 3 relevant state-of-the-art baselines, including HR-VITON [23], VITON-HD [5], and PASTA-GAN [33], by training all the models with our self-collected Pure Cotton dataset and the public Dress Code dataset [28]. Specifically, HR-VITON and VITON-HD both represent SOTA high-resolution virtual try-on methods with different warping modules. HR-VITON warps clothes with appearance flows and designs a feature fusion network for eliminating the mismatch issue, while VITON-HD warps clothes with TPS and proposes an alignment-aware segmentation normalization method to deal with the misalignment problem. Meanwhile, PASTA-GAN represents another warping method, a patch-routed disentanglement module, which is more related to our proposed clothing deformation method. + +Evaluation metrics. We evaluate the model performance by three widely-used metrics for virtual try-on, including i) Structural Similarity (SSIM) [38], ii) Learned Perceptual Image Patch Similarity (LPIPS) [36], and iii) Fréchet Inception Distance (FID) [18]. SSIM measures the quality of three key features: luminance, contrast, and structure between the reconstruction results and the ground truth. LPIPS evaluates the perceptual similarity and FID is computed for evaluating the GAN performance. + +# 4.2. Qualitative Results + +Fig. 4 and 5 show the visual comparisons in Pure Cotton and Dress Code datasets. Overall, our model achieves the most visually convincing high-resolution try-on results. The comparison analysis is discussed as follows. + +The previous warping methods, e.g., TPS (VITON-HD) and appearance flows (HR-VITON), did not consider clothing skeleton and sleeves segmentation, rendering them incapable of accurately synthesizing clothing patterns and features in respective positions, such as hit color mosaic (case I in Fig. 4) and logo in special positions (case II in Fig. 5). Additionally, these methods could not tackle regional clothing warping, such as arm akimbo in complex postures (case II in Fig. 4). In contrast, our model conducts the Clothing Landmark Predictor and Clothing Segmentation Network to obtain clothing structural information. Then, our proposed Landmark-guided Transformation helps to warp the clothes regionally. As such, our model could synthesize try-on results with clothing patterns and features located in the right position that closely follow the target clothing image, even in complex postures. On the other hand, PASTA-GAN conducts a regional clothing deformation method similar to our proposed method. However, there are three differences between PASTA-GAN and our model. Firstly, we split the clothes into 5 regions, i.e., torso, right/left upper arms, and right/left forearms, and kept the torso undivided. Secondly, we further consider clothing segmentation, including neckline and limb segmentation. Thirdly, PASTA-GAN deforms clothes from clothes on another human image, as demonstrated on the right-bottom side of the target clothes columns in Fig. 4. Due to these differences, PASTA-GAN would produce strange try-on results that highly follow the original clothing tied up and limb postures (the blue circle and the red arrow on Fig 4). In contrast, our model leverages clothing structural information to synthesize natural clothing wrinkles. + +Additional Clothing Control. The elimination policy in our proposed approach allows for greater flexibility in virtual try-on, with options for both tucked-in and untucked clothing. As shown in Fig. 6, previous methods would limit the location of the upper clothes' lower boundary based on the original lower clothes' upper boundary. However, our approach allows for the clothing to be untucked based on the clothes' body length, providing additional cues for customers to judge the clothing style. + +Partial Clothing Resizing: Apart from demonstrating COTTON's clothing size adjustment in Fig. 1, we also present its ability to adjust partial clothing sizes in Fig. 7. In the real world, clothing designers would adjust different sizes only for specific parts, e.g., body or sleeve length, instead of the entire garment. COTTON can smoothly manipulate partial clothing sizes by adjusting the corresponding landmarks without impacting the rest of the clothing. + +![](images/88bbb3a57198f79bae6a2e205f97d4c4374fe82e163f4dfdb9d6e2f06c1bbdec.jpg) +Figure 4. The visual comparison in the Pure Cotton dataset. The left two columns are the inputs and the rest present the results of baselines and our model. VITON-HD and HR-VITON did not consider sleeve segmentation, resulting in misaligned clothes with wrong pattern shifts (green dash line in case I) and broken warped clothes (case II). Moreover, without clothing skeletons, VITON-HD and HR-VITON failed to maintain pattern proportions (red and blue double-headed arrows in case I). PASTA-GAN also had issues in case I, where the partially tucked-in clothing shape from the original human image was directly copied to the try-on result, and in case II, where the try-on result had two additional limbs in front of the abdomen due to the human, representing target clothes, crossing their arms over the chest. + +![](images/fa331c3cfeed98f7dd1f04d04de7fbd70bfb267d1f5ea5a85523f4f82eda5226.jpg) +Figure 5. The visual comparison in the Dress Code dataset. In case I, our Clothing Elimination Policy well retains valuable human features, e.g., bulging muscles. For case II, our Landmark-guided Transformation aligns textures better based on the obtained clothing structure. + +![](images/dbbbb79aa58abbd171567925a642994356de1d8f9b2e4be60977705e2c5ee4ca.jpg) +Figure 6. Additional clothing control for tucked-in or untucked. + +# 4.3. Quantitative Results + +Table 1 compares COTTON's performance with SOTAs in terms of SSIM, LPIPS, and FID. We evaluated SSIM and LPIPS on 2,000 random paired test sets from our Pure Cot + +ton dataset and 1,800 test sets from the Dress Code dataset. FID was evaluated on 2,000 and 1,800 unpaired test sets from Pure Cotton and Dress Code datasets, respectively. Human images wearing original clothing served as ground truth for FID calculation. The results show that COTTON consistently outperforms the baselines across all evaluation metrics on both datasets, with at least $18.6\%$ improvement in LPIPS on Pure Cotton dataset and $41.1\%$ in FID on Dress Code dataset. The results manifest that our method excels in generating detailed contents on Pure Cotton dataset and robust results on Dress Code dataset. Notably, COTTON shows lower performance with untucked clothes, as it ob + +![](images/e1a20a28c0d5bdb66fdadc91b140b5b2eff79e838992baa75ea409fbd88ebd25.jpg) +Sleeves length adjustment + +![](images/36e13e32b0199788795087a0358c02f610129c1afb7c457f32e5984e501c19ef.jpg) +Body length adjustment +Figure 7. Visualization of partial clothing adjustment. + +Table 1. Quantitative comparison on the test dataset. + +
MethodDatasetSSIM↑LPIPS↓FID↓
PASTA-GAN0.9180.121529.43
VITON-HD0.9440.056617.16
HR-VITONSelf-collected0.9480.038711.16
Ours (untucked)Pure Cotton0.9560.034910.96
Ours (tucked-in)0.9580.031510.17
PASTA-GAN0.8000.306633.92
VITON-HD0.9270.090622.53
HR-VITONPublic0.9270.075520.12
Ours (untucked)Dress Code0.9490.048212.58
Ours (tucked-in)0.9530.043811.86
+ +Table 2. The user study shows that COTTON is the most photorealistic and remains the most human and clothing characteristics. + +
MethodPASTA-GANVITON-HDHR-VITONCOTTON(Ours)
Photo-realistic9.08%7.56%29.82%53.54%
Try-on accuracy8.66%4.62%27.40%59.32%
+ +scures the upper portion of lower clothing, causing the try-on results to differ from the ground truth. Nevertheless, Fig. 6 demonstrates that COTTON can generate more visually appealing results when clothing is untucked. + +Moreover, we conduct a user study on the Pure Cotton + +dataset to further evaluate the visual quality by humans. The experiment involved 127 volunteers. We randomly sampled 30 image sets and divided them equally into two groups for evaluating: i) photo-realistic quality and ii) try-on accuracy. To evaluate the photo-realistic quality, volunteers would see four try-on results (without input information) synthesized by four different models, and choose the most photo-realistic one. Furthermore, for evaluating try-on accuracy, we provide a source human image, one target try-on clothing, and four try-on results synthesized by four different models. Based on the source human and the target clothes, volunteers should select the best try-on result. Table 2 shows that COTTON outperforms all other baselines in both photo-realistic quality and try-on accuracy and achieves even higher ratings for try-on accuracy. Volunteers reported that they mainly used realistic reproduced clothing patterns, including loose/fit torso and accurate sleeve length, with fewer artifacts as the criteria to determine the better model. Our proposed Clothing Landmark Predictor and Clothing Segmentation Network effectively retain clothing characteristics such as patterns and sleeve length, contributing significantly to COTTON's superior performance over other state-of-the-art models. + +# 5. Conclusion + +This paper presents COTTON, a clothing-oriented transformation try-on network that synthesizes high-resolution virtual try-on results. By considering the clothing structure and leveraging it, COTTON improves the clothing transformation quality and allows for size adjustment during clothing try-on. Additionally, the proposed elimination policy enables COTTON to try on outfits untucked or not while preserving valuable human characteristics. The experiments demonstrate that COTTON outperforms the state-of-the-art virtual try-on works at $1024 \times 768$ resolution, both qualitatively and quantitatively. In future work, we plan to improve the accessories segmentation and extend the system to try on more clothing categories, such as shoes, hats, and accessories. Overall, the proposed approach shows great promise for improving the online shopping experience and reducing the need for physical try-on. + +# Acknowledgement + +We would like to express our gratitude to Josephine, Wendy, Hans, Joyce, Leslie and May from the Epoch Foundation for their valuable support. This work was supported in part by the National Science and Technology Council of Taiwan under Grants NSTC-109-2221-E-009-114-MY3, NSTC-112-2221-E-A49-094-MY3, NSTC-109-2223-E-009-002-MY3, NSTC-112-2634-F-002-002-MBK and NSTC-112-2628-E-002-033-MY4. + +# References + +[1] Alakh Aggarwal, Jikai Wang, Steven Hogue, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. Layered-garment net: Generating multiple implicit garment layers from a single image. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2022. 3 +[2] Shuai Bai, Huiling Zhou, Zhikang Li, Chang Zhou, and Hongxia Yang. Single stage virtual try-on via deformable attention flows. In European Conference on Computer Vision (ECCV), 2022. 1, 2 +[3] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 4, 6 +[4] Chieh-Yun Chen, Ling Lo, Pin-Jui Huang, Hong-Han Shuai, and Wen-Huang Cheng. Fashionmirror: Co-attention feature-remapping virtual try-on with sequential template poses. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1 +[5] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. VITON-HD: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3, 4, 5, 6 +[6] Ayush Chopra, Rishabh Jain, Mayur Hemani, and Balaji Krishnamurthy. Zflow: Gated appearance flow-based virtual try-on with 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3, 4 +[7] Chien-Lung Chou, Chieh-Yun Chen, Chia-Wei Hsieh, Hong-Han Shuai, Jiaying Liu, and Wen-Huang Cheng. Template-free try-on image synthesis via semantic-guided optimization. IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 2021. 3 +[8] Aiyu Cui, Sen He, Tao Xiang, and Antoine Toisoul. Learning garment densepose for robust warping in virtual try-on. In arXiv, 2023. 3 +[9] Aiyu Cui, Daniel McKee, and Svetlana Lazebnik. Dressing in order: Recurrent person image generation for pose transfer, virtual try-on and outfit editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 3 +[10] Haoye Dong, Xiaodan Liang, Xiaohui Shen, Bochao Wang, Hanjiang Lai, Jia Zhu, Zhiting Hu, and Jian Yin. Towards multi-pose guided virtual try-on network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 3 +[11] Xin Dong, Fuwei Zhao, Zhenyu Xie, Xijin Zhang, Kang Du, Min Zheng, Xiang Long, Xiaodan Liang, and Jianchao Yang1. Dressing in the wild by watching dance videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3 +[12] Ruili Feng, Cheng Ma, Chengji Shen, Xin Gao, Zhenjiang Liu, Xiaobo Li, Kairi Ou, Deli Zhao, and Zheng-Jun Zha. Weakly supervised high-fidelity clothing model generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 + +[13] Yuying Ge, Yibing Song, Ruimao Zhang, Chongjian Ge, Wei Liu, and Ping Luo. Parser-free virtual try-on via distilling appearance flows. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2 +[14] Oshri Halimi, Tuur Stuyck, Donglai Xiang, Timur Bagautdinov, He Wen, Ron Kimmel, Takaaki Shiratori, Chenglei Wu, Yaser Sheikh, and Fabian Prada. Pattern-based cloth registration and sparse-view animation. ACM Trans. Graph., 41(6), 2022. 3 +[15] Xintong Han, Xiaojun Hu, Weilin Huang, and Matthew R. Scott. ClothFlow: a flow-based model for clothed person generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 3, 4 +[16] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S. Davis. VITON: An image-based virtual try-on network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 3, 4 +[17] Sen He, Yi-Zhe Song, and Tao Xiang. Style-based global appearance flow for virtual try-on. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3 +[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Proceedings of the International Conference on Neural Information Processing Systems (NIPS), 2017. 6 +[19] Chia-Wei Hsieh, Chieh-Yun Chen, Chien-Lung Chou, Hong-Han Shuai, and Wen-Huang Cheng. Fit-me: Image-based virtual try-on with arbitrary poses. In IEEE International Conference on Image Processing (ICIP), 2019. 3 +[20] Chia-Wei Hsieh, Chieh-Yun Chen, Chien-Lung Chou, Hong-Han Shuai, Jiaying Liu, and Wen-Huang Cheng. FashionOn: Semantic-guided image-based virtual try-on with detailed human and clothing information. In Proceedings of the $27^{th}$ ACM International Conference on Multimedia (ACM MM), 2019. 3, 5 +[21] Thibaut Issenhuth, Jérémie Mary, and Clément Calauzènes. Do not mask what you do not need to mask: a parser-free virtual try-on. In European Conference on Computer Vision (ECCV), 2020. 2, 3, 5 +[22] Maria Korosteleva and Sung-Hee Lee. Neuraltailor: Reconstructing sewing pattern structures from 3d point clouds of garments. ACM Trans. Graph., 41(4), 2022. 3 +[23] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 4, 5, 6 +[24] Ren Li, Benoit Guillard, Edoardo Remelli, and Pascal Fua. Dig: Draping implicit garment over the human body. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2022. 3 +[25] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. 5 + +[26] Xinqi Liu, Jituo Li, Guodong Lu, Dongliang Zhang, and Shihai Xing. Robust and automatic clothing reconstruction based on a single rgb image. Computers & Graphics, 2023. 3 +[27] S. Majithia, S. N. Parameswaran, S. Babar, V. Garg, A. Srivastava, and A. Sharma. Robust 3d garment digitization from monocular 2d images for 3d virtual try-on systems. In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2022. 3 +[28] Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, and Rita Cucchiara. Dress Code: High-Resolution Multi-Category Virtual Try-On. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 4, 6 +[29] Assaf Neuberger, Eran Borenstein, Bar Hilleli, Eduard Oks, and Sharon Alpert. Image based virtual try-on network from unpaired data. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 3 +[30] Evan Shelhamer, Jonathan Long, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 6 +[31] Bochao Wang, Huabin Zheng, Xiaodan Liang, Yimin Chen, Liang Lin, and Meng Yang. Toward characteristic-preserving image-based virtual try-on network. In European Conference on Computer Vision (ECCV), 2018. 2, 3, 4 +[32] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5 +[33] Zhenyu Xie, Zaiyu Huang, Fuwei Zhao, Haoye Dong, Michael Kampffmeyer, and Xiaodan Liang. Towards scalable unpaired virtual try-on via patch-routed spatially-adaptive GAN. In Advances in Neural Information Processing Systems (NIPS), 2021. 1, 2, 3, 6 +[34] Han Yang, Xinrui Yu, and Ziwei Liu. Full-range virtual try-on with recurrent tri-level transform. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[35] Han Yang, Ruimao Zhang, Xiaobao Guo, Wei Liu, Wang-meng Zuo, and Ping Luo. Towards photo-realistic virtual try-on by adaptively generating $\leftrightarrow$ preserving image content. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 3 +[36] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6 +[37] Fuwei Zhao, Zhenyu Xie, Michael Kampffmeyer, Haoye Dong, Songfang Han, Tianxiang Zheng, Tao Zhang, and Xiaodan Liang. M3d-vton: A monocular-to-3d virtual try-on network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3 +[38] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility + +to structural similarity. IEEE Transactions on Image Processing (TIP), 2004. 6 \ No newline at end of file diff --git a/2023/Size Does Matter_ Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network/images.zip b/2023/Size Does Matter_ Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..74bd4fe84544a6d51b0ae0174571827aa94da6e2 --- /dev/null +++ b/2023/Size Does Matter_ Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffb1b4f0457f99d38d8f985f0dc8093465e986b1d1407024b1fa91a0f790649f +size 552981 diff --git a/2023/Size Does Matter_ Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network/layout.json b/2023/Size Does Matter_ Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6b339132291aba53ade3988ca649a89daf1d97a6 --- /dev/null +++ b/2023/Size Does Matter_ Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network/layout.json @@ -0,0 +1,7639 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 103, + 503, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 103, + 503, + 140 + ], + "spans": [ + { + "bbox": [ + 91, + 103, + 503, + 140 + ], + "type": "text", + "content": "Size Does Matter: Size-aware Virtual Try-on via Clothing-oriented Transformation Try-on Network" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 160, + 505, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 160, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 88, + 160, + 505, + 176 + ], + "type": "text", + "content": "Chieh-Yun Chen" + }, + { + "bbox": [ + 88, + 160, + 505, + 176 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 88, + 160, + 505, + 176 + ], + "type": "text", + "content": " Yi-Chung Chen" + }, + { + "bbox": [ + 88, + 160, + 505, + 176 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 88, + 160, + 505, + 176 + ], + "type": "text", + "content": " Hong-Han Shuai" + }, + { + "bbox": [ + 88, + 160, + 505, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 160, + 505, + 176 + ], + "type": "text", + "content": " Wen-Huang Cheng" + }, + { + "bbox": [ + 88, + 160, + 505, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 177, + 501, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 177, + 501, + 193 + ], + "spans": [ + { + "bbox": [ + 91, + 177, + 501, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 177, + 501, + 193 + ], + "type": "text", + "content": "Stylins.ai " + }, + { + "bbox": [ + 91, + 177, + 501, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 177, + 501, + 193 + ], + "type": "text", + "content": "National Yang Ming Chiao Tung University " + }, + { + "bbox": [ + 91, + 177, + 501, + 193 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 91, + 177, + 501, + 193 + ], + "type": "text", + "content": "National Taiwan University" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 63, + 211, + 533, + 369 + ], + "blocks": [ + { + "bbox": [ + 63, + 211, + 533, + 369 + ], + "lines": [ + { + "bbox": [ + 63, + 211, + 533, + 369 + ], + "spans": [ + { + "bbox": [ + 63, + 211, + 533, + 369 + ], + "type": "image", + "image_path": "82f50f4453d85f38c84e709171b798272971efcaee9572fd1b7a2f92430550ae.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 375, + 546, + 398 + ], + "lines": [ + { + "bbox": [ + 46, + 375, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 546, + 398 + ], + "type": "text", + "content": "Figure 1. Multi-garment virtual try-on results with different sizes (tucked or untucked). " + }, + { + "bbox": [ + 46, + 375, + 546, + 398 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 375, + 546, + 398 + ], + "type": "text", + "content": " is the ratio of torso length over shoulder width, which can be decided by customers in online shopping scenarios. Padding blue backgrounds for better visualizing clothing skeletons." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 408, + 192, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 408, + 192, + 420 + ], + "spans": [ + { + "bbox": [ + 143, + 408, + 192, + 420 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 426, + 290, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 426, + 290, + 700 + ], + "spans": [ + { + "bbox": [ + 45, + 426, + 290, + 700 + ], + "type": "text", + "content": "Virtual try-on tasks aim at synthesizing realistic try-on results by trying target clothes on humans. Most previous works relied on the Thin Plate Spline or appearance flows to warp clothes to fit human body shapes. However, both approaches cannot handle complex warping, leading to over distortion or misalignment. Furthermore, there is a critical unaddressed challenge of adjusting clothing sizes for try-on. To tackle these issues, we propose a Clothing-Oriented Transformation Try-On Network (COTTON). COTTON leverages clothing structure with landmarks and segmentation to design a novel landmark-guided transformation for precisely deforming clothes, allowing for size adjustment during try-on. Additionally, to properly remove the clothing region from the human image without losing significant human characteristics, we propose a clothing elimination policy based on both transformed clothes and human segmentation. This method enables users to try on clothes tucked-in or untucked while retaining more human characteristics. Both qualitative and quantitative results show that COTTON outperforms the state-of-the-art high-resolution virtual try-on approaches. All the code is available at https://github.com/cotton6/COTTON-size-does-matter." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 408, + 386, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 408, + 386, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 408, + 386, + 420 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 430, + 546, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 546, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 546, + 621 + ], + "type": "text", + "content": "Image-based virtual try-on replaces clothing items on a person with the desired ones, creating realistic try-on results and lowering costs associated with on-model photos for the e-commerce industry1. Additionally, it enables customers to use virtual dressing rooms when shopping online, potentially enhancing the e-commerce experience and increasing conversion rates2. As online shopping becomes popular, the virtual try-on tasks have received more and more attention [2,4,5,9,13,17,23,28,29,33,35]. For example, to extend the virtual try-on resolution from low " + }, + { + "bbox": [ + 304, + 430, + 546, + 621 + ], + "type": "inline_equation", + "content": "(256\\times 192)" + }, + { + "bbox": [ + 304, + 430, + 546, + 621 + ], + "type": "text", + "content": " to high " + }, + { + "bbox": [ + 304, + 430, + 546, + 621 + ], + "type": "inline_equation", + "content": "(1024\\times 768)" + }, + { + "bbox": [ + 304, + 430, + 546, + 621 + ], + "type": "text", + "content": ", the most obvious difficulty is the misalignment issue. [5] proposed the alignment-aware segment normalization to remove the misleading information in the misaligned area. Additionally, [23] further performs appearance flow-based warping and segmentation map generation simultaneously to tackle misalignment." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 623, + 547, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 623, + 547, + 672 + ], + "spans": [ + { + "bbox": [ + 304, + 623, + 547, + 672 + ], + "type": "text", + "content": "Despite the considerable progress made in previous works, there are still some challenges that have not been adequately addressed, as outlined below. i) Handling complex warping without misalignment: General image" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 317, + 683, + 477, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 683, + 477, + 693 + ], + "spans": [ + { + "bbox": [ + 317, + 683, + 477, + 693 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 693, + 457, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 693, + 457, + 703 + ], + "spans": [ + { + "bbox": [ + 318, + 693, + 457, + 703 + ], + "type": "text", + "content": "1https://www.zmo.ai/aimodels/" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 703, + 423, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 703, + 423, + 712 + ], + "spans": [ + { + "bbox": [ + 318, + 703, + 423, + 712 + ], + "type": "text", + "content": "2https://www.revery.ai" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "7513" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 540 + ], + "type": "text", + "content": "based virtual try-on frameworks [5, 16, 21, 28, 31] typically employ a clothing deformation module, such as the Thin Plate Spline (TPS) method, to align clothing images with the target human pose. However, research [6, 15, 23] has found that TPS may not effectively handle complex warping when different garment regions require different deformations. In attempts to improve warping results, they have replaced TPS with dense appearance flows. However, when the transformation between the garment and corresponding body parts is significant (as illustrated in case II in Fig. 4), the performance of both TPS (VITON-HD [5]) and flow-based (HR-VITON [23]) methods deteriorates drastically. Hence, existing methods still can't properly address the challenge of clothing deformation. ii) Adjusting clothing sizes: Previous works predicted the try-on segmentation conditioned on the clothing image by only considering the shape of the clothes without the scale information. Thus, given a clothing image, it is impossible for previous works to change the clothing size. This limitation severely restricts the practical application of virtual try-on because people often try on different sizes in the fitting room. iii) Appropriate clothing elimination policy: Previous works [5, 23, 28] rely on segmentation maps to eliminate the original clothing region and utilize the remaining area as a guide for trying on target clothes. However, this method can lead to either excessive or insufficient removal of the original image. For instance, when prior works try on upper clothes, the lower part of the upper clothes remains intact, which limits the length of the upper clothes, as shown in Fig. 6. The lower part of the upper clothing is forced to be tucked in, which reduces the model's ability to generate complete upper clothes, thereby reducing the practicality of virtual try-on. Additionally, previous works remove the entire arm of input human images, which eliminates essential human characteristics such as tattoos and arm width that should be preserved in the final output, as illustrated in case I of Fig. 5. Therefore, developing an appropriate clothing elimination strategy that preserves crucial information while removing the clothing region is a significant challenge." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": "To address these challenges, we propose a simple yet powerful approach called Clothing-Oriented Transformation Try-On Network (COTTON). Specifically, COTTON first exploits geometric information with Clothing Landmark Predictor and Clothing Segmentation Network to predict clothing landmarks and segmentation masks respectively. To overcome the first challenge posed by complex warping, we propose the Landmark-guided Transformation that first separates clothes into sub-parts via the segmentation mask and then uses clothing landmarks to estimate homography matrices that fit these sub-parts to the target human pose. To solve the second challenge, we introduce a clothing landmarks adjustment approach that allows users to change clothing sizes. Adjustment of landmarks will al" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "content": "ter the homography matrices and then cause clothing size changes, as demonstrated in Fig. 1, which remarkably enhances the practicality of virtual try-on. To address the third challenge, we employ the transformed clothing images to identify and remove the areas that would be covered by the clothes. This enables our proposed Clothing Elimination Policy to effectively eliminate the clothing region while preserving important details and offer flexibility in tucking the clothes or not, as shown in Fig. 6. Extensive experiments show that COTTON achieves superior results than state-of-the-arts both quantitatively and qualitatively. We summarize our contributions as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 317, + 224, + 545, + 438 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 317, + 224, + 545, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 224, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 317, + 224, + 545, + 272 + ], + "type": "text", + "content": "- We propose a Clothing-Oriented Transformation Try-On Network (COTTON), which improves clothing transformation quality and addresses complex warping misalignments by leveraging clothing geometry." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 279, + 545, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 279, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 317, + 279, + 545, + 339 + ], + "type": "text", + "content": "- We introduce adjustable clothing landmarks to provide clothing size information based only on images. To the best of our knowledge, this is the first work enabling 2D-based virtual try-on with different clothing sizes for approaching real-world try-on." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 347, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 347, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 317, + 347, + 545, + 396 + ], + "type": "text", + "content": "- To preserve critical information, we propose a Clothing Elimination Policy to properly remove clothing information while retaining valuable human characteristics, and offer flexibility in tucking the clothes or not." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 403, + 545, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 403, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 317, + 403, + 545, + 438 + ], + "type": "text", + "content": "- Extensive experiments on the Dress Code dataset [28] show that our model significantly outperforms SOTAs, e.g., at least " + }, + { + "bbox": [ + 317, + 403, + 545, + 438 + ], + "type": "inline_equation", + "content": "41.1\\%" + }, + { + "bbox": [ + 317, + 403, + 545, + 438 + ], + "type": "text", + "content": " improvement in terms of FID." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 448, + 398, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 448, + 398, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 448, + 398, + 460 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 468, + 432, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 432, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 432, + 480 + ], + "type": "text", + "content": "2.1. Clothing Deformation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "content": "Virtual try-on requires clothing deformation to align patterns and preserve clothing details. Thin Plate Spline (TPS) warping is widely used in virtual try-on tasks [5, 16, 21, 28, 31], but it has limited ability to model geometric changes and result in unnatural deformations. Thus, [2, 6, 11, 13, 15, 17, 23] conducted appearance flows as an alternative to TPS, but misalignment still has not be well addressed. To address misalignment, [5] improved TPS by eliminating irrelevant clothing texture information in misaligned regions. Meanwhile, [23] improved the deformation by designing pathways to jointly predict appearance flows and segmentation features. However, neither TPS nor appearance flow can deform partial regions of clothing without affecting other parts, e.g., warping the sleeves to present arm akimbo without affecting the torso part of the clothes, since they do not have the structural information of the clothes. To address this issue, [33] introduced normalized patches to learn spatial-agnostic clothing features but this approach may destroy the completeness of clothes" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7514" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 69, + 523, + 249 + ], + "blocks": [ + { + "bbox": [ + 75, + 69, + 523, + 249 + ], + "lines": [ + { + "bbox": [ + 75, + 69, + 523, + 249 + ], + "spans": [ + { + "bbox": [ + 75, + 69, + 523, + 249 + ], + "type": "image", + "image_path": "4923e65b2fc850a5929ae0bb9a7352a5eb97006251701116aa2700d9a0632a96.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "text", + "content": "Figure 2. Overview of COTTON. Given a clothing image " + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "text", + "content": ", COTTON uses Clothing Landmark Predictor and Clothing Segmentation Network to predict clothing landmarks " + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "inline_equation", + "content": "K_{C}" + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "text", + "content": " and a clothing segmentation mask " + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "inline_equation", + "content": "M_{C}" + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "text", + "content": " is then warped by Landmark-guided Transformation with homography derived from human keypoints " + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "inline_equation", + "content": "K_{P}" + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "inline_equation", + "content": "K_{C}" + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "text", + "content": " to obtain the transformed images " + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "inline_equation", + "content": "C^{torso}" + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "inline_equation", + "content": "C^{limb}" + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "text", + "content": " is used to adjust " + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "inline_equation", + "content": "K_{C}" + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "text", + "content": " at this step for changing clothing size. The clothing elimination step uses " + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "inline_equation", + "content": "C^{torso}" + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "inline_equation", + "content": "C^{limb}" + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "text", + "content": " to mask human image " + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "text", + "content": " and segmentation mask " + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "text", + "content": ", resulting in clothing-agnostic image " + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "inline_equation", + "content": "I_{a}" + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "text", + "content": " and segmentation " + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "inline_equation", + "content": "S_{a}" + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "text", + "content": ". The variable " + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "text", + "content": " controls whether clothing is tucked in or not. Finally, Outfit Generator aggregates the previous outputs to predict segmentation mask " + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "text", + "content": " and synthesize the final try-on result " + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 46, + 252, + 548, + 319 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 327, + 288, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 288, + 437 + ], + "type": "text", + "content": "due to lack of semantic information. In contrast, our proposed COTTON separates clothes based on landmarks and segmentation masks to maintain clothing completeness and aligns clothes based on specific landmarks to address misalignment. Besides, [8] is a concurrent work. Both of our works improve garment warping by human-garment correspondence. While [8] relies on the computationally intensive DensePose method, our proposed " + }, + { + "bbox": [ + 46, + 327, + 288, + 437 + ], + "type": "inline_equation", + "content": "LT" + }, + { + "bbox": [ + 46, + 327, + 288, + 437 + ], + "type": "text", + "content": " method is more efficient and enables clothing size adjustment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 443, + 139, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 139, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 139, + 456 + ], + "type": "text", + "content": "2.2. Virtual Try-on" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": "Virtual try-on methods can be categorized into 3D-based methods [1,14,22,24,26,27,37] and 2D-based methods [5,6, 7,8,9,10,11,12,15,16,17,19,23,28,29,31,33,34,35]. Since 3D-based methods are time-consuming and real-world scenarios require fast inference speeds, we focus on 2D-based methods. [16] proposed a coarse-to-fine network to synthesize image-based virtual try-on results based on TPS warping. However, try-on results synthesized by [16,31] do not provide a clear silhouette of human body parts since they consider the human to be a general mask instead of considering the semantic information, e.g., legs, head, limbs, etc. To tackle this problem, [10,20] proposed the first semantic-guided virtual try-on network, which predicted the target semantic segmentation to provide human structural information for the network to synthesize photo-realistic body parts. Afterward, many try-on methods [5,6,7,9,23,29,35] are designed with an individual stage for predicting the target semantic segmentation before synthesizing virtual try-on results. However, the individual stage for predicting target semantic segmentation would lead to error accumulation as pointed out by [21]. In this paper, we propose an end-to" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 327, + 547, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 327, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 547, + 376 + ], + "type": "text", + "content": "end Outfit Generator that updates semantic segmentation and try-on results simultaneously, and introduce a Clothing Elimination Policy that retains valuable human characteristics while removing original clothing information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 392, + 389, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 389, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 389, + 407 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": "Given a human image " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": " and a clothing image " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": ", the goal of virtual try-on is to obtain a synthetic image " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": ", which shows the human of " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": " wearing the clothing of " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": ". We design a Clothing-Oriented Transformation Try-On Network (COTTON) for generating high-resolution virtual try-on results. Fig. 2 provides an overview of the proposed COTTON. We use the Clothing Landmark Predictor and the Clothing Segmentation Network to respectively capture clothing landmarks and a clothing segmentation map of " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": ". Afterward, we separate the clothes into torso part " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "C^{torse}" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": " and limbs part " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "C^{limb}" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": " according to the clothing segmentation map. The separation of clothes enables COTTON to treat each part differently, and thus makes it adapt to handling complex warping well. To align the clothes with the target human pose, the clothing landmarks are adjusted with a ratio " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": " to fit human body proportions and used to calculate homography matrices. COTTON then warps the clothes with affine transformation. The warped clothes shed light on how to properly remove the clothing region of image " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": " without losing necessary information. We use a user-determined variable " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": " to control whether to eliminate the overlap region of clothes. In this case, we can generate tucked-in or untucked try-on results by choosing the value of " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": ". Finally, the Outfit Generator takes both human and clothing representation to generate the try-on result " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "7515" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 243, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 243, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 243, + 85 + ], + "type": "text", + "content": "3.1. Clothing Landmark Predictor (CLP)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 284 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 284 + ], + "type": "text", + "content": "One of the most challenging problems of virtual try-on is to transform a clothing image " + }, + { + "bbox": [ + 46, + 91, + 289, + 284 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 91, + 289, + 284 + ], + "type": "text", + "content": " to fit the target human pose. To solve this challenge, several approaches have been proposed in previous works, including hyperparameter learning of TPS warping [5, 16, 28, 31], and flow-based deep deformation [6, 15, 23]. However, these approaches implicitly learn the underlying geometric information from the training dataset and only achieve mediocre results when handling complex warping. We argue that explicit geometric information plays the most critical role in virtual try-on and can be leveraged to further improve performance. For example, by thoroughly understanding the clothes' geometry, humans can easily capture the relation between clothes and humans, even if the clothes are with different textures and styles. Therefore, we propose the Clothing Landmark Predictor (CLP) to extract explicit geometric information." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "text", + "content": "Specifically, to exploit the geometric information of clothes, our proposed CLP is trained to predict a set of landmarks " + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "inline_equation", + "content": "K_{C}" + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "text", + "content": ", which correlate to different parts of human pose " + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "inline_equation", + "content": "K_{P}" + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "inline_equation", + "content": "K_{P}" + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "text", + "content": " can be obtained from off-the-shelf human pose estimation methods, e.g., [3]. The ground truth of " + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "inline_equation", + "content": "K_{C}" + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "text", + "content": " is labeled manually, and the number of landmarks varies with regard to the clothing types. For upper clothing, " + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "inline_equation", + "content": "|K_{C}| = 10" + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "text", + "content": ", including neck, shoulders, elbows, wrists, and hips. These landmarks explicitly indicate the relation between clothes and human pose, and therefore can be used as a reference to generate transformation matrices. Given clothing image " + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "text", + "content": ", the CLP is trained to output clothing joint heatmaps " + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "inline_equation", + "content": "\\hat{M}_{JH}" + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "text", + "content": " and Part Affinity Field " + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "inline_equation", + "content": "\\hat{M}_{PAF}" + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "text", + "content": ". Afterward, the predicted landmarks " + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "inline_equation", + "content": "K_{C}" + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "text", + "content": " can be obtained from " + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "inline_equation", + "content": "\\hat{M}_{JH}" + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "inline_equation", + "content": "\\hat{M}_{PAF}" + }, + { + "bbox": [ + 46, + 285, + 289, + 467 + ], + "type": "text", + "content": " by using the Hungarian algorithm." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 478, + 221, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 221, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 221, + 491 + ], + "type": "text", + "content": "3.2. Clothing Segmentation Network" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": "Each part of clothes has distinct properties that should be taken into consideration during transformation. For example, the sleeves of clothes usually suffer severe distortion during transformation while the torso part only undergoes a mild distortion. Another common noise in clothing try-on is the region around the neckline that must be concealed when the garment is worn. This region always introduces undesired variation into try-on results, and as such, should be eliminated. Though the Clothing Landmark Predictor provides the connection between the clothes and human poses, it is difficult to identify different regions with only a few landmarks. To find the explicit boundary of each part, pixel-level semantic segmentation is necessary. Therefore, we propose a Clothing Segmentation Network to predict the clothing segmentation mask " + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "inline_equation", + "content": "M_C" + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": ", which separates the clothing into three sub-region, including torso part, sleeve part, and invisible part around the neckline. As such, each sub-region can be treated properly based on its characteristics." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 69, + 544, + 175 + ], + "blocks": [ + { + "bbox": [ + 310, + 69, + 544, + 175 + ], + "lines": [ + { + "bbox": [ + 310, + 69, + 544, + 175 + ], + "spans": [ + { + "bbox": [ + 310, + 69, + 544, + 175 + ], + "type": "image", + "image_path": "fd7d1e9ad3e08136529b6dd189d740d9fecb6a19a7dfca0885bc8d4dcbba3a23.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 176, + 545, + 222 + ], + "lines": [ + { + "bbox": [ + 304, + 176, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 304, + 176, + 545, + 222 + ], + "type": "text", + "content": "Figure 3. Clothing transformation steps. For the torso part, we use keypoints including shoulders and hips to calculate matrix " + }, + { + "bbox": [ + 304, + 176, + 545, + 222 + ], + "type": "inline_equation", + "content": "H_{c \\to p}^{torso}" + }, + { + "bbox": [ + 304, + 176, + 545, + 222 + ], + "type": "text", + "content": ". For the limbs part, we first create bounding boxes around limbs and use them to calculate matrices " + }, + { + "bbox": [ + 304, + 176, + 545, + 222 + ], + "type": "inline_equation", + "content": "H_{c \\to p}^{limb,i}" + }, + { + "bbox": [ + 304, + 176, + 545, + 222 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 176, + 545, + 222 + ], + "type": "inline_equation", + "content": "i \\in \\{1,2,3,4\\}" + }, + { + "bbox": [ + 304, + 176, + 545, + 222 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 230, + 491, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 230, + 491, + 243 + ], + "spans": [ + { + "bbox": [ + 306, + 230, + 491, + 243 + ], + "type": "text", + "content": "3.3. Landmark-guided Transformation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 248, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 248, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 248, + 545, + 392 + ], + "type": "text", + "content": "Here, we introduce how to leverage the landmarks " + }, + { + "bbox": [ + 304, + 248, + 545, + 392 + ], + "type": "inline_equation", + "content": "K_{C}" + }, + { + "bbox": [ + 304, + 248, + 545, + 392 + ], + "type": "text", + "content": " and segmentation mask " + }, + { + "bbox": [ + 304, + 248, + 545, + 392 + ], + "type": "inline_equation", + "content": "M_{C}" + }, + { + "bbox": [ + 304, + 248, + 545, + 392 + ], + "type": "text", + "content": " to achieve high-fidelity clothing transformation. We first remove the undesired part around the neckline such as the green region of " + }, + { + "bbox": [ + 304, + 248, + 545, + 392 + ], + "type": "inline_equation", + "content": "M_{C}" + }, + { + "bbox": [ + 304, + 248, + 545, + 392 + ], + "type": "text", + "content": " in Fig. 2, which is invisible in the final try-on result, via the clothing segmentation map. We then separate the remaining part into two parts, including the torso part and the limbs part. Let " + }, + { + "bbox": [ + 304, + 248, + 545, + 392 + ], + "type": "inline_equation", + "content": "K_{C}^{\\text{torso}}" + }, + { + "bbox": [ + 304, + 248, + 545, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 248, + 545, + 392 + ], + "type": "inline_equation", + "content": "K_{P}^{\\text{torso}}" + }, + { + "bbox": [ + 304, + 248, + 545, + 392 + ], + "type": "text", + "content": " denote the clothing and human keypoints belonging to the torso part, such as shoulders and hips. For aligning the torso part, we compute a homography matrix " + }, + { + "bbox": [ + 304, + 248, + 545, + 392 + ], + "type": "inline_equation", + "content": "H_{c\\rightarrow p}^{\\text{torso}} \\in R^{3\\times 3}" + }, + { + "bbox": [ + 304, + 248, + 545, + 392 + ], + "type": "text", + "content": " that can project clothing landmarks to human keypoints, as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 319, + 402, + 545, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 402, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 319, + 402, + 545, + 418 + ], + "type": "interline_equation", + "content": "k _ {p} = H _ {c \\rightarrow p} ^ {t o r s o} k _ {c}, \\quad \\forall k _ {c} \\in K _ {C} ^ {t o r s o}, \\forall k _ {p} \\in K _ {P} ^ {t o r s o}. (1)", + "image_path": "18b85d5bd1d2cfc65a872a35168a3b4bfa1d84e187aace3cc3b6c14f43e17bac.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "text", + "content": "Afterward, this perspective transformation is applied to get the transformed torso part " + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "inline_equation", + "content": "C^{torso}" + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "text", + "content": ". On the other hand, as the limbs part usually suffers from severe distortion, it requires a more sophisticated procedure. Therefore, instead of directly warping the whole limbs part, we break the limbs part into four sub-limbs based on the clothing landmarks. Take a hoodie as an example in Fig. 3. The limbs part is divided into two upper arms and two forearms. We create bounding boxes for each sub-limb based on clothing landmarks and human keypoints to obtain the area of each sub-limb. The corners of bounding boxes are then used to compute 4 homography matrices, " + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "inline_equation", + "content": "H_{c\\rightarrow p}^{limb,i}, i\\in \\{1,2,3,4\\}" + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "text", + "content": ", which project each sub-limb to the target human. Finally, the four sub-limbs are collected as the warped limbs part " + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "inline_equation", + "content": "C^{limb}" + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "It is worth noting that the predicted clothing landmarks " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "K_{C}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " only provide a clothes-human relation in the general case. However, since body proportion varies from person to person, the relative size of the same clothing on different persons may change. To generate a more realistic try-on result, we can adjust the clothing landmark " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "K_{C}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " to fit the body proportion of the target human. For instance, if two people have the same shoulder width but different torso lengths, the portion of their bodies covered by clothing should be different. In that case, the predicted landmarks are adjusted" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7516" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "accordingly to fit their body proportions. Let " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " be the ratio of torso length over shoulder width, which can be derived from extra body information provided by customers in online shopping scenarios. Fig. 1 visualizes the results of clothing landmarks adjustment with different " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": ", which shows that the region covered by the clothes is varied with regards to different " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": ", and can therefore generate try-on results with different clothing sizes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 178, + 200, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 178, + 200, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 200, + 192 + ], + "type": "text", + "content": "3.4. Clothing Elimination Policy" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 198, + 287, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 342 + ], + "type": "text", + "content": "Since the goal of the virtual try-on task is to replace the clothes of a given person image " + }, + { + "bbox": [ + 46, + 198, + 287, + 342 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 198, + 287, + 342 + ], + "type": "text", + "content": " with a target one, the information about the original clothes should be eliminated. One common approach is eliminating the clothing according to the human segmentation " + }, + { + "bbox": [ + 46, + 198, + 287, + 342 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 198, + 287, + 342 + ], + "type": "text", + "content": ". However, [5] pointed out that simply removing the clothing from image " + }, + { + "bbox": [ + 46, + 198, + 287, + 342 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 198, + 287, + 342 + ], + "type": "text", + "content": " leads to a performance drop for paired data training due to information leakage about the clothing shape, and therefore proposed a clothing-agnostic person representation to address this problem. Nevertheless, this approach may lead to valuable information loss, e.g., arm width in case I of Fig. 5, when whole arms are directly masked during inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 342, + 289, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 289, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 289, + 571 + ], + "type": "text", + "content": "Hence, we propose a Clothing Elimination Policy for the inference phase to preserve the important information while generating clothing-agnostic person representation. The regions that ought to be eliminated in image " + }, + { + "bbox": [ + 46, + 342, + 289, + 571 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 342, + 289, + 571 + ], + "type": "text", + "content": " are i) the regions of original clothes and ii) the regions covered by target clothes " + }, + { + "bbox": [ + 46, + 342, + 289, + 571 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 342, + 289, + 571 + ], + "type": "text", + "content": " when worn. Note that the latter is hard to measure with previous methods. Fortunately, since we have already obtained the reliable warped clothing image at the previous step, we can use it as a reference to find the region that would be covered after try-on. Specifically, we create a clothing-elimination mask " + }, + { + "bbox": [ + 46, + 342, + 289, + 571 + ], + "type": "inline_equation", + "content": "M_{elm}" + }, + { + "bbox": [ + 46, + 342, + 289, + 571 + ], + "type": "text", + "content": " by union the original clothing region and the region covered by the transformed target clothes, which is then expanded using morphological dilation to cover the entire body. Moreover, when wearing tops, we make the clothing tucked or not in the try-on result by controlling " + }, + { + "bbox": [ + 46, + 342, + 289, + 571 + ], + "type": "inline_equation", + "content": "t \\in \\{0,1\\}" + }, + { + "bbox": [ + 46, + 342, + 289, + 571 + ], + "type": "text", + "content": " to determine whether to eliminate the overlap region of bottoms. As such, we can effectively obtain clothing-agnostic person representation without overly masking the irrelevant regions, i.e.," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 580, + 287, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 580, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 52, + 580, + 287, + 596 + ], + "type": "interline_equation", + "content": "M _ {e l m} = \\mathcal {M} (S ^ {t} \\cup (C ^ {t o r s e} \\cup C ^ {l i m b}) - t \\cdot (C ^ {t o r s e} \\cap S ^ {b})), (2)", + "image_path": "d2e23a200f8edb9e75ec74393a33af84d1182510cac612d47f4d38ae1146e77b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": " represents the morphological dilation; " + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "inline_equation", + "content": "S^t" + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "inline_equation", + "content": "S^b" + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": " denote top and bottom clothing segmentation respectively. Fig. 1 shows the results of different " + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": ". After eliminating clothing information, we obtain clothing-agnostic images " + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "inline_equation", + "content": "I_a" + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": " and segmentation maps " + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "inline_equation", + "content": "S_a" + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": ". To bridge the performance gap between the training and inference phases, we randomly mask the whole arms of human images with a " + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": " probability during the training phase and apply our Clothing Elimination Policy only in the inference phase." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 409, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 409, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 409, + 84 + ], + "type": "text", + "content": "3.5. Outfit Generator" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 91, + 547, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 247 + ], + "type": "text", + "content": "In the final stage, the Outfit Generator takes personal " + }, + { + "bbox": [ + 304, + 91, + 547, + 247 + ], + "type": "inline_equation", + "content": "(I_{a}, S_{a}" + }, + { + "bbox": [ + 304, + 91, + 547, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 91, + 547, + 247 + ], + "type": "inline_equation", + "content": "K_{P})" + }, + { + "bbox": [ + 304, + 91, + 547, + 247 + ], + "type": "text", + "content": " and clothing " + }, + { + "bbox": [ + 304, + 91, + 547, + 247 + ], + "type": "inline_equation", + "content": "(C^{torse}" + }, + { + "bbox": [ + 304, + 91, + 547, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 91, + 547, + 247 + ], + "type": "inline_equation", + "content": "C^{limb})" + }, + { + "bbox": [ + 304, + 91, + 547, + 247 + ], + "type": "text", + "content": " information as input and synthesizes the final try-on result. Since learning the shape and texture of clothing simultaneously is challenging for the generator, a segmentation network is commonly used to predict a segmentation mask as guidance to lead the try-on generator network [5, 20, 23]. However, they regarded segmentation prediction as an independent task and trained the segmentation network and try-on generator separately. In that case, the optimization of the two networks may be suboptimal. Moreover, the accumulated noise from the segmentation network may further deteriorate the performance of the final synthesis result [21]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "text", + "content": "To address the problem, we propose a novel Outfit Generator by combining the segmentation network and try-on generator as one end-to-end training multi-task network. Specifically, " + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "inline_equation", + "content": "S_{a}" + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "inline_equation", + "content": "K_{P}" + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "inline_equation", + "content": "C^{torse}" + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "inline_equation", + "content": "C^{limb}" + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "text", + "content": " are first concatenated and down-sampled to a lower resolution as the input of the segmentation network to predict segmentation mask " + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "text", + "content": ". Afterward, " + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "text", + "content": " is up-sampled back to the original resolution and combined with " + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "inline_equation", + "content": "I_{a}" + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "text", + "content": " as well as " + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "inline_equation", + "content": "C^{torse}" + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "inline_equation", + "content": "C^{limb}" + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "text", + "content": " as the input of the synthesis step. Finally, the generator integrates all the information and renders the synthesis result " + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 304, + 247, + 547, + 368 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 348, + 373, + 545, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 373, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 348, + 373, + 545, + 388 + ], + "type": "interline_equation", + "content": "\\hat {I}, \\hat {S} = G (I _ {a}, S _ {a}, K _ {P}, C ^ {t o r s e}, C ^ {l i m b}). \\tag {3}", + "image_path": "93ba61d0167526ed1d37f4c20341a94719cde3bc04e44cb3ea5e0c363a338355.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 392, + 547, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 547, + 464 + ], + "type": "text", + "content": "The objective function of the Outfit Generator includes the supervision of both the predicted segmentation mask and the try-on synthesis result. We apply focal loss " + }, + { + "bbox": [ + 304, + 392, + 547, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {focal }}" + }, + { + "bbox": [ + 304, + 392, + 547, + 464 + ], + "type": "text", + "content": " [25] for segmentation. For try-on results, we use losses including reconstruction loss " + }, + { + "bbox": [ + 304, + 392, + 547, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {rec }}" + }, + { + "bbox": [ + 304, + 392, + 547, + 464 + ], + "type": "text", + "content": ", perceptual loss " + }, + { + "bbox": [ + 304, + 392, + 547, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {VGG }}" + }, + { + "bbox": [ + 304, + 392, + 547, + 464 + ], + "type": "text", + "content": ", and adversarial loss " + }, + { + "bbox": [ + 304, + 392, + 547, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {adv }}" + }, + { + "bbox": [ + 304, + 392, + 547, + 464 + ], + "type": "text", + "content": ". The overall loss functions are as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 331, + 485, + 545, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 485, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 331, + 485, + 545, + 513 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {o v e r a l l}} = \\lambda_ {\\text {r e c}} \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {V G G} \\mathcal {L} _ {V G G} \\tag {4} \\\\ + \\lambda_ {f o c a l} \\mathcal {L} _ {f o c a l} + \\lambda_ {a d v} \\mathcal {L} _ {a d v}, \\\\ \\end{array}", + "image_path": "7bfd33ac0992ad82d4632e30776fca6d79f2f9b216a994c0ef3fcefaa3ae6314.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 345, + 516, + 545, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 516, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 345, + 516, + 545, + 529 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c}} = \\| \\hat {I} - I \\| _ {1}, \\tag {5}", + "image_path": "06a9038e67953d0ad7510c499acd99949cf6e1c5b65cc940130fc8012de5dfc5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 338, + 533, + 545, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 533, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 338, + 533, + 545, + 564 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {V G G} = \\sum_ {k = 1} ^ {5} \\lambda_ {k} \\| \\phi_ {k} (\\hat {I}) - \\phi_ {k} (I) \\| _ {1}, \\tag {6}", + "image_path": "febc05d63587f9ab56fe0750a0de684b02e52295347e0df46ca6e622b80f105b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 338, + 567, + 545, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 567, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 338, + 567, + 545, + 592 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f o c a l} = \\sum_ {i} - \\alpha_ {i} (1 - p _ {i}) ^ {\\gamma} \\log (p _ {i}), p _ {i} \\in \\hat {S}, \\tag {7}", + "image_path": "0d1159ea62e753423f4426695d49938f808e7a50e76c09c7126ea1ee78487f67.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 344, + 594, + 545, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 594, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 344, + 594, + 545, + 609 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a d v} = \\left\\| \\mathcal {D} (\\hat {I}) - \\mathbf {1} \\right\\| _ {2}, \\tag {8}", + "image_path": "f24cc00ef5938f31a6df8570f71a4051d12b93dfdea0f02a2df6c41da5c3acbf.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\phi_{k}" + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "text", + "content": " layer output of VGG-19 network. " + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "text", + "content": " are respectively the importance weighting for the " + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "text", + "content": " class and the probabilities of pixels belonging to the " + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "text", + "content": " class. " + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "text", + "content": " is the focusing parameter. " + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\lambda_{rec},\\lambda_{VGG},\\lambda_{focal}" + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\lambda_{adv}" + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "text", + "content": " denote the balance weights. " + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 306, + 620, + 545, + 693 + ], + "type": "text", + "content": " is the multi-layer discriminators [32]." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 545, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 545, + 714 + ], + "type": "text", + "content": "3The comparison of two-stage and end-to-end training is in supplement." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "7517" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 128, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 128, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 128, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 99, + 163, + 113 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 99, + 163, + 113 + ], + "spans": [ + { + "bbox": [ + 47, + 99, + 163, + 113 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "text", + "content": "Dataset. We created the Pure Cotton, a high-resolution " + }, + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "inline_equation", + "content": "(1024\\times 768)" + }, + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "text", + "content": " outfit dataset, to mitigate the impact of noisy data on model performance. We strictly employed several rules to filter out unsuitable data, which are clearly outlined in the supplements, resulting in 16,428 frontal-view human models and in-shop clothing pairs, including 10,636 upper clothes and 5,792 bottom clothes. For training and testing, we divided the upper clothing set into 8,451 and 2,185 pairs, respectively, and the bottom clothing set into 4,626 and 1,166 pairs. To evaluate the model's robustness, we also experimented with the public Dress Code dataset [28]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 266, + 287, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 410 + ], + "type": "text", + "content": "Implementation Details. To train Clothing Landmark Predictor (CLP) and Clothing Segmentation Network (CSN), we collected a small clothing dataset with 50 images for each clothing type, including shirts, turtleneck, hoodies, etc, and manually labeled landmarks and segmentation masks as ground truth. The architecture of CLP is based on the one in [3], and that of CSN is based on FCN-resnet50 [30]. Moreover, Outfit Generator is designed using an Unet and an encoder-decoder network. To ensure a fair comparison with previous studies, all figures are generated in the tucked-in setting unless stated otherwise. This setting is consistent with prior arts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 418, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 597 + ], + "type": "text", + "content": "Baselines. We compare our proposed method with 3 relevant state-of-the-art baselines, including HR-VITON [23], VITON-HD [5], and PASTA-GAN [33], by training all the models with our self-collected Pure Cotton dataset and the public Dress Code dataset [28]. Specifically, HR-VITON and VITON-HD both represent SOTA high-resolution virtual try-on methods with different warping modules. HR-VITON warps clothes with appearance flows and designs a feature fusion network for eliminating the mismatch issue, while VITON-HD warps clothes with TPS and proposes an alignment-aware segmentation normalization method to deal with the misalignment problem. Meanwhile, PASTA-GAN represents another warping method, a patch-routed disentanglement module, which is more related to our proposed clothing deformation method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Evaluation metrics. We evaluate the model performance by three widely-used metrics for virtual try-on, including i) Structural Similarity (SSIM) [38], ii) Learned Perceptual Image Patch Similarity (LPIPS) [36], and iii) Fréchet Inception Distance (FID) [18]. SSIM measures the quality of three key features: luminance, contrast, and structure between the reconstruction results and the ground truth. LPIPS evaluates the perceptual similarity and FID is computed for evaluating the GAN performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 419, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 419, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 419, + 85 + ], + "type": "text", + "content": "4.2. Qualitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 91, + 545, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 545, + 138 + ], + "type": "text", + "content": "Fig. 4 and 5 show the visual comparisons in Pure Cotton and Dress Code datasets. Overall, our model achieves the most visually convincing high-resolution try-on results. The comparison analysis is discussed as follows." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 138, + 546, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 138, + 546, + 509 + ], + "spans": [ + { + "bbox": [ + 305, + 138, + 546, + 509 + ], + "type": "text", + "content": "The previous warping methods, e.g., TPS (VITON-HD) and appearance flows (HR-VITON), did not consider clothing skeleton and sleeves segmentation, rendering them incapable of accurately synthesizing clothing patterns and features in respective positions, such as hit color mosaic (case I in Fig. 4) and logo in special positions (case II in Fig. 5). Additionally, these methods could not tackle regional clothing warping, such as arm akimbo in complex postures (case II in Fig. 4). In contrast, our model conducts the Clothing Landmark Predictor and Clothing Segmentation Network to obtain clothing structural information. Then, our proposed Landmark-guided Transformation helps to warp the clothes regionally. As such, our model could synthesize try-on results with clothing patterns and features located in the right position that closely follow the target clothing image, even in complex postures. On the other hand, PASTA-GAN conducts a regional clothing deformation method similar to our proposed method. However, there are three differences between PASTA-GAN and our model. Firstly, we split the clothes into 5 regions, i.e., torso, right/left upper arms, and right/left forearms, and kept the torso undivided. Secondly, we further consider clothing segmentation, including neckline and limb segmentation. Thirdly, PASTA-GAN deforms clothes from clothes on another human image, as demonstrated on the right-bottom side of the target clothes columns in Fig. 4. Due to these differences, PASTA-GAN would produce strange try-on results that highly follow the original clothing tied up and limb postures (the blue circle and the red arrow on Fig 4). In contrast, our model leverages clothing structural information to synthesize natural clothing wrinkles." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 510, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 510, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 305, + 510, + 545, + 617 + ], + "type": "text", + "content": "Additional Clothing Control. The elimination policy in our proposed approach allows for greater flexibility in virtual try-on, with options for both tucked-in and untucked clothing. As shown in Fig. 6, previous methods would limit the location of the upper clothes' lower boundary based on the original lower clothes' upper boundary. However, our approach allows for the clothing to be untucked based on the clothes' body length, providing additional cues for customers to judge the clothing style." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 618, + 545, + 713 + ], + "type": "text", + "content": "Partial Clothing Resizing: Apart from demonstrating COTTON's clothing size adjustment in Fig. 1, we also present its ability to adjust partial clothing sizes in Fig. 7. In the real world, clothing designers would adjust different sizes only for specific parts, e.g., body or sleeve length, instead of the entire garment. COTTON can smoothly manipulate partial clothing sizes by adjusting the corresponding landmarks without impacting the rest of the clothing." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7518" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 74, + 504, + 262 + ], + "blocks": [ + { + "bbox": [ + 88, + 74, + 504, + 262 + ], + "lines": [ + { + "bbox": [ + 88, + 74, + 504, + 262 + ], + "spans": [ + { + "bbox": [ + 88, + 74, + 504, + 262 + ], + "type": "image", + "image_path": "88bbb3a57198f79bae6a2e205f97d4c4374fe82e163f4dfdb9d6e2f06c1bbdec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 266, + 545, + 333 + ], + "lines": [ + { + "bbox": [ + 46, + 266, + 545, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 545, + 333 + ], + "type": "text", + "content": "Figure 4. The visual comparison in the Pure Cotton dataset. The left two columns are the inputs and the rest present the results of baselines and our model. VITON-HD and HR-VITON did not consider sleeve segmentation, resulting in misaligned clothes with wrong pattern shifts (green dash line in case I) and broken warped clothes (case II). Moreover, without clothing skeletons, VITON-HD and HR-VITON failed to maintain pattern proportions (red and blue double-headed arrows in case I). PASTA-GAN also had issues in case I, where the partially tucked-in clothing shape from the original human image was directly copied to the try-on result, and in case II, where the try-on result had two additional limbs in front of the abdomen due to the human, representing target clothes, crossing their arms over the chest." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 93, + 337, + 503, + 529 + ], + "blocks": [ + { + "bbox": [ + 93, + 337, + 503, + 529 + ], + "lines": [ + { + "bbox": [ + 93, + 337, + 503, + 529 + ], + "spans": [ + { + "bbox": [ + 93, + 337, + 503, + 529 + ], + "type": "image", + "image_path": "fa331c3cfeed98f7dd1f04d04de7fbd70bfb267d1f5ea5a85523f4f82eda5226.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 532, + 544, + 555 + ], + "lines": [ + { + "bbox": [ + 46, + 532, + 544, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 544, + 555 + ], + "type": "text", + "content": "Figure 5. The visual comparison in the Dress Code dataset. In case I, our Clothing Elimination Policy well retains valuable human features, e.g., bulging muscles. For case II, our Landmark-guided Transformation aligns textures better based on the obtained clothing structure." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 52, + 565, + 283, + 630 + ], + "blocks": [ + { + "bbox": [ + 52, + 565, + 283, + 630 + ], + "lines": [ + { + "bbox": [ + 52, + 565, + 283, + 630 + ], + "spans": [ + { + "bbox": [ + 52, + 565, + 283, + 630 + ], + "type": "image", + "image_path": "dbbbb79aa58abbd171567925a642994356de1d8f9b2e4be60977705e2c5ee4ca.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 632, + 277, + 643 + ], + "lines": [ + { + "bbox": [ + 47, + 632, + 277, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 632, + 277, + 643 + ], + "type": "text", + "content": "Figure 6. Additional clothing control for tucked-in or untucked." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 656, + 165, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 656, + 165, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 656, + 165, + 669 + ], + "type": "text", + "content": "4.3. Quantitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "Table 1 compares COTTON's performance with SOTAs in terms of SSIM, LPIPS, and FID. We evaluated SSIM and LPIPS on 2,000 random paired test sets from our Pure Cot" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 565, + 545, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 545, + 708 + ], + "type": "text", + "content": "ton dataset and 1,800 test sets from the Dress Code dataset. FID was evaluated on 2,000 and 1,800 unpaired test sets from Pure Cotton and Dress Code datasets, respectively. Human images wearing original clothing served as ground truth for FID calculation. The results show that COTTON consistently outperforms the baselines across all evaluation metrics on both datasets, with at least " + }, + { + "bbox": [ + 304, + 565, + 545, + 708 + ], + "type": "inline_equation", + "content": "18.6\\%" + }, + { + "bbox": [ + 304, + 565, + 545, + 708 + ], + "type": "text", + "content": " improvement in LPIPS on Pure Cotton dataset and " + }, + { + "bbox": [ + 304, + 565, + 545, + 708 + ], + "type": "inline_equation", + "content": "41.1\\%" + }, + { + "bbox": [ + 304, + 565, + 545, + 708 + ], + "type": "text", + "content": " in FID on Dress Code dataset. The results manifest that our method excels in generating detailed contents on Pure Cotton dataset and robust results on Dress Code dataset. Notably, COTTON shows lower performance with untucked clothes, as it ob" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7519" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 83, + 274, + 228 + ], + "blocks": [ + { + "bbox": [ + 63, + 74, + 145, + 83 + ], + "lines": [ + { + "bbox": [ + 63, + 74, + 145, + 83 + ], + "spans": [ + { + "bbox": [ + 63, + 74, + 145, + 83 + ], + "type": "text", + "content": "Sleeves length adjustment" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 83, + 274, + 228 + ], + "lines": [ + { + "bbox": [ + 62, + 83, + 274, + 228 + ], + "spans": [ + { + "bbox": [ + 62, + 83, + 274, + 228 + ], + "type": "image", + "image_path": "e1a20a28c0d5bdb66fdadc91b140b5b2eff79e838992baa75ea409fbd88ebd25.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 239, + 274, + 358 + ], + "blocks": [ + { + "bbox": [ + 65, + 230, + 137, + 239 + ], + "lines": [ + { + "bbox": [ + 65, + 230, + 137, + 239 + ], + "spans": [ + { + "bbox": [ + 65, + 230, + 137, + 239 + ], + "type": "text", + "content": "Body length adjustment" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 239, + 274, + 358 + ], + "lines": [ + { + "bbox": [ + 62, + 239, + 274, + 358 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 274, + 358 + ], + "type": "image", + "image_path": "36e13e32b0199788795087a0358c02f610129c1afb7c457f32e5984e501c19ef.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 361, + 264, + 373 + ], + "lines": [ + { + "bbox": [ + 69, + 361, + 264, + 373 + ], + "spans": [ + { + "bbox": [ + 69, + 361, + 264, + 373 + ], + "type": "text", + "content": "Figure 7. Visualization of partial clothing adjustment." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 396, + 294, + 531 + ], + "blocks": [ + { + "bbox": [ + 72, + 385, + 261, + 395 + ], + "lines": [ + { + "bbox": [ + 72, + 385, + 261, + 395 + ], + "spans": [ + { + "bbox": [ + 72, + 385, + 261, + 395 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on the test dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 396, + 294, + 531 + ], + "lines": [ + { + "bbox": [ + 48, + 396, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 396, + 294, + 531 + ], + "type": "table", + "html": "
MethodDatasetSSIM↑LPIPS↓FID↓
PASTA-GAN0.9180.121529.43
VITON-HD0.9440.056617.16
HR-VITONSelf-collected0.9480.038711.16
Ours (untucked)Pure Cotton0.9560.034910.96
Ours (tucked-in)0.9580.031510.17
PASTA-GAN0.8000.306633.92
VITON-HD0.9270.090622.53
HR-VITONPublic0.9270.075520.12
Ours (untucked)Dress Code0.9490.048212.58
Ours (tucked-in)0.9530.043811.86
", + "image_path": "cad675f3d94dd95acc37bad00718b8605f45b4c17d8f72c26022c9965460a1be.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 57, + 568, + 276, + 645 + ], + "blocks": [ + { + "bbox": [ + 47, + 546, + 286, + 568 + ], + "lines": [ + { + "bbox": [ + 47, + 546, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 286, + 568 + ], + "type": "text", + "content": "Table 2. The user study shows that COTTON is the most photorealistic and remains the most human and clothing characteristics." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 568, + 276, + 645 + ], + "lines": [ + { + "bbox": [ + 57, + 568, + 276, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 568, + 276, + 645 + ], + "type": "table", + "html": "
MethodPASTA-GANVITON-HDHR-VITONCOTTON(Ours)
Photo-realistic9.08%7.56%29.82%53.54%
Try-on accuracy8.66%4.62%27.40%59.32%
", + "image_path": "9bb56ed274c4bb6b9553a656d6ef3a3458421b30fec4bb03b5e5d2efeb928d6a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 652, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 652, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 652, + 286, + 700 + ], + "type": "text", + "content": "scures the upper portion of lower clothing, causing the try-on results to differ from the ground truth. Nevertheless, Fig. 6 demonstrates that COTTON can generate more visually appealing results when clothing is untucked." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "Moreover, we conduct a user study on the Pure Cotton" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 347 + ], + "type": "text", + "content": "dataset to further evaluate the visual quality by humans. The experiment involved 127 volunteers. We randomly sampled 30 image sets and divided them equally into two groups for evaluating: i) photo-realistic quality and ii) try-on accuracy. To evaluate the photo-realistic quality, volunteers would see four try-on results (without input information) synthesized by four different models, and choose the most photo-realistic one. Furthermore, for evaluating try-on accuracy, we provide a source human image, one target try-on clothing, and four try-on results synthesized by four different models. Based on the source human and the target clothes, volunteers should select the best try-on result. Table 2 shows that COTTON outperforms all other baselines in both photo-realistic quality and try-on accuracy and achieves even higher ratings for try-on accuracy. Volunteers reported that they mainly used realistic reproduced clothing patterns, including loose/fit torso and accurate sleeve length, with fewer artifacts as the criteria to determine the better model. Our proposed Clothing Landmark Predictor and Clothing Segmentation Network effectively retain clothing characteristics such as patterns and sleeve length, contributing significantly to COTTON's superior performance over other state-of-the-art models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 365, + 378, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 365, + 378, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 378, + 376 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 387, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 578 + ], + "type": "text", + "content": "This paper presents COTTON, a clothing-oriented transformation try-on network that synthesizes high-resolution virtual try-on results. By considering the clothing structure and leveraging it, COTTON improves the clothing transformation quality and allows for size adjustment during clothing try-on. Additionally, the proposed elimination policy enables COTTON to try on outfits untucked or not while preserving valuable human characteristics. The experiments demonstrate that COTTON outperforms the state-of-the-art virtual try-on works at " + }, + { + "bbox": [ + 304, + 387, + 545, + 578 + ], + "type": "inline_equation", + "content": "1024 \\times 768" + }, + { + "bbox": [ + 304, + 387, + 545, + 578 + ], + "type": "text", + "content": " resolution, both qualitatively and quantitatively. In future work, we plan to improve the accessories segmentation and extend the system to try on more clothing categories, such as shoes, hats, and accessories. Overall, the proposed approach shows great promise for improving the online shopping experience and reducing the need for physical try-on." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 595, + 403, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 595, + 403, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 595, + 403, + 609 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "content": "We would like to express our gratitude to Josephine, Wendy, Hans, Joyce, Leslie and May from the Epoch Foundation for their valuable support. This work was supported in part by the National Science and Technology Council of Taiwan under Grants NSTC-109-2221-E-009-114-MY3, NSTC-112-2221-E-A49-094-MY3, NSTC-109-2223-E-009-002-MY3, NSTC-112-2634-F-002-002-MBK and NSTC-112-2628-E-002-033-MY4." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7520" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Alakh Aggarwal, Jikai Wang, Steven Hogue, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. Layered-garment net: Generating multiple implicit garment layers from a single image. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 190 + ], + "type": "text", + "content": "[2] Shuai Bai, Huiling Zhou, Zhikang Li, Chang Zhou, and Hongxia Yang. Single stage virtual try-on via deformable attention flows. In European Conference on Computer Vision (ECCV), 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "type": "text", + "content": "[3] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 290 + ], + "type": "text", + "content": "[4] Chieh-Yun Chen, Ling Lo, Pin-Jui Huang, Hong-Han Shuai, and Wen-Huang Cheng. Fashionmirror: Co-attention feature-remapping virtual try-on with sequential template poses. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 346 + ], + "type": "text", + "content": "[5] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. VITON-HD: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 401 + ], + "type": "text", + "content": "[6] Ayush Chopra, Rishabh Jain, Mayur Hemani, and Balaji Krishnamurthy. Zflow: Gated appearance flow-based virtual try-on with 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 456 + ], + "type": "text", + "content": "[7] Chien-Lung Chou, Chieh-Yun Chen, Chia-Wei Hsieh, Hong-Han Shuai, Jiaying Liu, and Wen-Huang Cheng. Template-free try-on image synthesis via semantic-guided optimization. IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 458, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 287, + 491 + ], + "type": "text", + "content": "[8] Aiyu Cui, Sen He, Tao Xiang, and Antoine Toisoul. Learning garment densepose for robust warping in virtual try-on. In arXiv, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 491, + 288, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 288, + 546 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 288, + 546 + ], + "type": "text", + "content": "[9] Aiyu Cui, Daniel McKee, and Svetlana Lazebnik. Dressing in order: Recurrent person image generation for pose transfer, virtual try-on and outfit editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 602 + ], + "type": "text", + "content": "[10] Haoye Dong, Xiaodan Liang, Xiaohui Shen, Bochao Wang, Hanjiang Lai, Jia Zhu, Zhiting Hu, and Jian Yin. Towards multi-pose guided virtual try-on network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 603, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 288, + 658 + ], + "type": "text", + "content": "[11] Xin Dong, Fuwei Zhao, Zhenyu Xie, Xijin Zhang, Kang Du, Min Zheng, Xiang Long, Xiaodan Liang, and Jianchao Yang1. Dressing in the wild by watching dance videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[12] Ruili Feng, Cheng Ma, Chengji Shen, Xin Gao, Zhenjiang Liu, Xiaobo Li, Kairi Ou, Deli Zhao, and Zheng-Jun Zha. Weakly supervised high-fidelity clothing model generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[13] Yuying Ge, Yibing Song, Ruimao Zhang, Chongjian Ge, Wei Liu, and Ping Luo. Parser-free virtual try-on via distilling appearance flows. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[14] Oshri Halimi, Tuur Stuyck, Donglai Xiang, Timur Bagautdinov, He Wen, Ron Kimmel, Takaaki Shiratori, Chenglei Wu, Yaser Sheikh, and Fabian Prada. Pattern-based cloth registration and sparse-view animation. ACM Trans. Graph., 41(6), 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 185, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 229 + ], + "type": "text", + "content": "[15] Xintong Han, Xiaojun Hu, Weilin Huang, and Matthew R. Scott. ClothFlow: a flow-based model for clothed person generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 231, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 274 + ], + "type": "text", + "content": "[16] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S. Davis. VITON: An image-based virtual try-on network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "text", + "content": "[17] Sen He, Yi-Zhe Song, and Tao Xiang. Style-based global appearance flow for virtual try-on. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 321, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 376 + ], + "type": "text", + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Proceedings of the International Conference on Neural Information Processing Systems (NIPS), 2017. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 376, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 421 + ], + "type": "text", + "content": "[19] Chia-Wei Hsieh, Chieh-Yun Chen, Chien-Lung Chou, Hong-Han Shuai, and Wen-Huang Cheng. Fit-me: Image-based virtual try-on with arbitrary poses. In IEEE International Conference on Image Processing (ICIP), 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 422, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 487 + ], + "type": "text", + "content": "[20] Chia-Wei Hsieh, Chieh-Yun Chen, Chien-Lung Chou, Hong-Han Shuai, Jiaying Liu, and Wen-Huang Cheng. FashionOn: Semantic-guided image-based virtual try-on with detailed human and clothing information. In Proceedings of the " + }, + { + "bbox": [ + 308, + 422, + 545, + 487 + ], + "type": "inline_equation", + "content": "27^{th}" + }, + { + "bbox": [ + 308, + 422, + 545, + 487 + ], + "type": "text", + "content": " ACM International Conference on Multimedia (ACM MM), 2019. 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 488, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 488, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 545, + 533 + ], + "type": "text", + "content": "[21] Thibaut Issenhuth, Jérémie Mary, and Clément Calauzènes. Do not mask what you do not need to mask: a parser-free virtual try-on. In European Conference on Computer Vision (ECCV), 2020. 2, 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "type": "text", + "content": "[22] Maria Korosteleva and Sung-Hee Lee. Neuraltailor: Reconstructing sewing pattern structures from 3d point clouds of garments. ACM Trans. Graph., 41(4), 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 568, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 622 + ], + "type": "text", + "content": "[23] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "text", + "content": "[24] Ren Li, Benoit Guillard, Edoardo Remelli, and Pascal Fua. Dig: Draping implicit garment over the human body. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7521" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[26] Xinqi Liu, Jituo Li, Guodong Lu, Dongliang Zhang, and Shihai Xing. Robust and automatic clothing reconstruction based on a single rgb image. Computers & Graphics, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "type": "text", + "content": "[27] S. Majithia, S. N. Parameswaran, S. Babar, V. Garg, A. Srivastava, and A. Sharma. Robust 3d garment digitization from monocular 2d images for 3d virtual try-on systems. In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 228 + ], + "type": "text", + "content": "[28] Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, and Rita Cucchiara. Dress Code: High-Resolution Multi-Category Virtual Try-On. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 281 + ], + "type": "text", + "content": "[29] Assaf Neuberger, Eran Borenstein, Bar Hilleli, Eduard Oks, and Sharon Alpert. Image based virtual try-on network from unpaired data. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 282, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 282, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 282, + 287, + 326 + ], + "type": "text", + "content": "[30] Evan Shelhamer, Jonathan Long, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "type": "text", + "content": "[31] Bochao Wang, Huabin Zheng, Xiaodan Liang, Yimin Chen, Liang Lin, and Meng Yang. Toward characteristic-preserving image-based virtual try-on network. In European Conference on Computer Vision (ECCV), 2018. 2, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "type": "text", + "content": "[32] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "type": "text", + "content": "[33] Zhenyu Xie, Zaiyu Huang, Fuwei Zhao, Haoye Dong, Michael Kampffmeyer, and Xiaodan Liang. Towards scalable unpaired virtual try-on via patch-routed spatially-adaptive GAN. In Advances in Neural Information Processing Systems (NIPS), 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "type": "text", + "content": "[34] Han Yang, Xinrui Yu, and Ziwei Liu. Full-range virtual try-on with recurrent tri-level transform. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 580 + ], + "type": "text", + "content": "[35] Han Yang, Ruimao Zhang, Xiaobao Guo, Wei Liu, Wang-meng Zuo, and Ping Luo. Towards photo-realistic virtual try-on by adaptively generating " + }, + { + "bbox": [ + 48, + 525, + 287, + 580 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 48, + 525, + 287, + 580 + ], + "type": "text", + "content": " preserving image content. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "text", + "content": "[36] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "type": "text", + "content": "[37] Fuwei Zhao, Zhenyu Xie, Michael Kampffmeyer, Haoye Dong, Songfang Han, Tianxiang Zheng, Tao Zhang, and Xiaodan Liang. M3d-vton: A monocular-to-3d virtual try-on network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[38] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "type": "text", + "content": "to structural similarity. IEEE Transactions on Image Processing (TIP), 2004. 6" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7522" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SkeleTR_ Towards Skeleton-based Action Recognition in the Wild/27665a84-18aa-44bc-b25b-e6429d7734c6_content_list.json b/2023/SkeleTR_ Towards Skeleton-based Action Recognition in the Wild/27665a84-18aa-44bc-b25b-e6429d7734c6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..84ac30549144a9f58d399ff3e6f56310d6934099 --- /dev/null +++ b/2023/SkeleTR_ Towards Skeleton-based Action Recognition in the Wild/27665a84-18aa-44bc-b25b-e6429d7734c6_content_list.json @@ -0,0 +1,1538 @@ +[ + { + "type": "text", + "text": "SkeleTR: Towards Skeleton-based Action Recognition in the Wild", + "text_level": 1, + "bbox": [ + 151, + 130, + 818, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haodong Duan $^{1*}$ Mingze Xu $^{2}$ Bing Shuai $^{2}$ Davide Modolo $^{2}$ Zhuowen Tu $^{2}$ Joseph Tighe $^{2}$ Alessandro Bergamo $^{2}$ The Chinese University of Hong Kong AWS AI Labs", + "bbox": [ + 166, + 179, + 803, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dhd.efz@gmail.com, {xumingze,bshuai,dmodolo,ztu,tighej,bergamo}@amazon.com", + "bbox": [ + 156, + 247, + 812, + 261 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 296, + 313, + 313 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present SkeleTR, a new framework for skeleton-based action recognition. In contrast to prior work, which focuses mainly on controlled environments, we target more general scenarios that typically involve a variable number of people and various forms of interaction between people. SkeleTR works with a two-stage paradigm. It first models the intra-person skeleton dynamics for each skeleton sequence with graph convolutions, and then uses stacked Transformer encoders to capture person interactions that are important for action recognition in general scenarios. To mitigate the negative impact of inaccurate skeleton associations, SkeleTR takes relative short skeleton sequences as input and increases the number of sequences. As a unified solution, SkeleTR can be directly applied to multiple skeleton-based action tasks, including video-level action classification, instance-level action detection, and group-level activity recognition. It also enables transfer learning and joint training across different action tasks and datasets, which result in performance improvement. When evaluated on various skeleton-based action recognition benchmarks, SkeleTR achieves the state-of-the-art performance.", + "bbox": [ + 75, + 328, + 473, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 654, + 209, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Skeleton-based action recognition has received increasing attention in recent years due to its robustness against background and illumination changes [64, 16]. Existing methods [36, 7, 8, 15] mostly use Graph Convolutional Networks (GCNs) as backbones and focus on modeling spatiotemporal information from long skeleton sequences separately. While observing promising results in simplified and controlled environments [46, 32], they are less effective in more realistic scenarios with a variable number of people performing diverse actions. First, considering that linking skeletons (e.g., using skeletons' confidence scores [64]) in a long time span may inevitably generate more association errors [12, 48], requiring long sequences as input can bring", + "bbox": [ + 75, + 679, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/abcaf131eac95a9937be84888aecc7bbdc0156d8545bb16e99b89b9c82682192.jpg", + "image_caption": [ + "Figure 1: SkeleTR is a general framework for skeleton action recognition which can handle different tasks, including video-level action classification, instance-level action detection, and group-level activity recognition. It adopts a two-stage paradigm: the intra-sequence and inter-sequence modeling, respectively." + ], + "image_footnote": [], + "bbox": [ + 506, + 295, + 890, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lots of noises. Second, prior works pay little attention to modeling the relations among a group of people, which plays an important role in recognizing both person interactions (e.g., talking to someone [22]) and group activities (e.g., pass/win-point in a volleyball game [24]).", + "bbox": [ + 496, + 536, + 893, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose SkeleTR for skeleton action recognition in real-world scenarios and handle multiple action tasks (i.e., video-level, instance-level, and group-level action recognition) in a general framework. SkeleTR samples a large number of short skeleton sequences instead of a few long ones (practice of prior works [64, 16]) from each video. Each short sequence directly corresponds to an atomic action of a single person, thus the collection of multiple short sequences in space and time can represent hierarchical actions (with a longer time span) of a single person or a video-level action involving multiple people. The new input paradigm can be applied to instance-level and video-level recognition of actions with different temporal granularities, and it has several unique advantages. First, association errors are less likely to happen in a shorter skeleton sequence [12, 48]. We show that the improved sequence consistency leads to better intra-sequence modeling. Second, this new format is more flexible and helps to create comprehensive yet compact inputs. With a finer temporal", + "bbox": [ + 496, + 613, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*The work was done during an Amazon internship.", + "bbox": [ + 94, + 886, + 367, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13634", + "bbox": [ + 480, + 924, + 519, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "granularity, the same input size budget can cover more human skeletons. It reduces redundant zero-paddings for a small number of people, and mitigates information loss during input sampling for a large number of people.", + "bbox": [ + 78, + 90, + 467, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SkeleTR adopts a two-stage paradigm as shown in Fig 1. Stage 1 performs intra-sequence modeling to capture the motion patterns of each individual skeleton sequence with a GCN backbone. The GCN backbone outputs a spatiotemporal feature for each sequence. Stage 2 concatenates all sequence features and feeds them into Transformer encoders to model inter-sequence relationships. At the end of stage 2, average pooling aggregates all features belonging to the same sequence into a single feature vector. By leveraging inter-sequence modeling, the output sequence feature incorporates the interactions between people and the global context of the video, which is helpful in both instance-level and video-level action recognition. To reduce the dimension of spatio-temporal skeleton features and keep the computations of Transformer tractable, we insert a novel Mix Pooling module before stage 2. It generates a compact representation with fine granularity for each sequence by applying multiple partial dimension reduction strategies in parallel.", + "bbox": [ + 78, + 152, + 467, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SkeleTR easily enables transfer learning and joint learning across different datasets and tasks. With a unified input format and model architecture, one can jointly train multiple tasks with different datasets, each with its own loss function and ground truth annotations. This allows for leveraging auxiliary datasets for pre-training or joint training of a target task with limited labeled data. In experiments, we find that joint training serves as a strong regularizer and improves recognition performance, especially in scenarios with limited training data.", + "bbox": [ + 78, + 424, + 467, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of SkeleTR, we conduct comprehensive study on three skeleton-based action recognition tasks: action classification [6, 65, 46], spatiotemporal action detection [22], and group activity recognition [24]. Experimental results show that SkeleTR works well with different GCN backbones and significantly outperforms prior works on all three tasks. For example, with ST-GCN++ [15] backbone, SkeleTR outperforms the state-of-the-art methods by $3.8\\%$ , $3.8\\%$ , $0.7\\%$ , and $1.1\\%$ on Poetics [64], AVA [22], Volleyball [24], and NTU-Inter [40]. With Poetics joint training, the AVA performance is further improved by $4\\%$ mAP. The great capability of SkeleTR leads to another noteworthy result: we first observe the strong complementarity of skeleton action recognition. Combining SkeleTR prediction and the RGB state-of-the-art [2, 39] with simple late fusion leads to great performance boost (2.3% Poetics Top-1, 3% AVA mAP).", + "bbox": [ + 78, + 575, + 467, + 829 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 78, + 844, + 223, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human Action Understanding. Action classification [6, 21] is the most fundamental task in human action under", + "bbox": [ + 78, + 869, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "standing, aiming at recognizing the video-level action in a short video clip. Another task of great interest is recognizing the instance/person-level actions in a video, either at the atomic [22, 28] or hierarchical [30] level. Meanwhile, other areas of interest include group activity recognition [24, 9], action localization [17, 25, 70], or action segmentation [27, 50]. Various modalities in videos can be utilized to solve the action understanding tasks, such as RGB appearance [58, 6, 19, 59], optical flow [49], audio waves [61], and human skeletons [64, 16]. Spatio-temporal representations learned end-to-end from RGB appearance yield good performance on standard benchmarks. However, they are data intensive and less robust to context changes. In this work, we propose a unified framework to learn video representations based on human skeletons, that can be easily applied to multiple action tasks.", + "bbox": [ + 503, + 90, + 888, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Skeleton-based Action Recognition. Human skeletons are robust against background and illumination changes [64, 47], and can be easily obtained with sensors [69] or pose estimators [4, 51]. In skeleton action recognition, the majority of research has focused on modeling the intra-person skeleton dynamics. With the rise of deep learning, manual feature engineering [23, 55, 57] has gradually been replaced by end-to-end representation learning with deep neural architectures (GCN [64], 2D-CNN [10], or 3D-CNN [16]). Among them, GCN-based approaches [64, 47, 36, 7] have been the most popular paradigm, which construct spatial-temporal graphs of human keypoint coordinates and perform spatial and temporal modeling. These methods model the intra-person skeleton dynamics for in-lab collected videos [46, 11], featuring small number of persons (1 or 2) per video and accurate person identities. For more challenging scenarios [6], the model performance decreases due to the noises produced in matching skeletons to sequences. Besides, GCN approaches primarily focus on predicting video-level actions, and do not benefit from context or interaction modeling when applied to instance-level action recognition. To this end, we propose SkeleTR, which expands existing GCN approaches and works universally well for diversified scenarios and recognition tasks.", + "bbox": [ + 503, + 337, + 888, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human Interaction Recognition. Human interaction recognition [43, 53, 67] is a sub-task of action recognition. It aims at identifying and categorizing activities between people. Early works [66, 26] built hand-crafted features for interaction modeling and adopted classic machine learning algorithms for activity classification. Subsequent works proposed to use two-stream RNNs to model the two-person interactions [41, 37, 56]. Recently, IGFormer [40] first proposed to use a two-stream Transformer for two-person interaction modeling and observed improved performance. However, most deep methods share the same limitations: 1. they are largely restricted to two-person scenarios and tend to suffer from exponentially growing inference", + "bbox": [ + 503, + 704, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13635", + "bbox": [ + 480, + 925, + 517, + 935 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c04aa9b76c8349ec0693112c454311d4fc298ff9cfb52ba2538b912c98b3ad5e.jpg", + "image_caption": [ + "Figure 2: The architecture of SkeLETR. SkeLETR takes a large set of short skeleton sequences as input and performs a two-stage modeling process. It first adopts a GCN backbone to model the intra-sequence skeleton dynamics. Then it processes the sequence features with stacked Transformer encoders, which helps to model person interactions and leverage the long-range context. Between the two stages, we adopt a Mix Pooling layer to compress the spatio-temporal features and keep computations tractable." + ], + "image_footnote": [], + "bbox": [ + 120, + 80, + 848, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "costs with the person number; 2. they fail to leverage the strong intra-sequence modeling capabilities of GCNs. To overcome these challenges, we introduce SkeleTR, which integrates intra-sequence and inter-sequence modeling with good scalability and small computation cost.", + "bbox": [ + 75, + 313, + 470, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Transformer. With the great success in NLP [54, 13], Transformers have been increasingly applied to diversified vision tasks [14, 34, 5, 74] as well as the RGB-based action understanding [35, 18, 1, 2, 63, 71], and have achieved promising results. For skeleton action recognition, there exist works [42, 68] that directly use Transformers for intra-sequence spatio-temporal modeling. However, without exploiting the locality, the Transformer-based motion modeling only leads to inferior recognition performance compared to GCN-based ones. In SkeleTR, to enjoy the best of both worlds, we adopt GCNs for intra-sequence motion modeling, and use Transformer encoders for intersequence relationship modeling among multiple persons' fine-grained skeleton features.", + "bbox": [ + 75, + 390, + 470, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SkeleTR", + "text_level": 1, + "bbox": [ + 76, + 619, + 174, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 647, + 187, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Both SkeleTR and prior GCN approaches directly take human joint coordinates as input. A person in a specific frame corresponds to a skeleton with $V$ joints (each has $C$ dimensions, $C$ is 2 or 3). Skeletons in $T$ different frames (1 skeleton per frame) can be linked in temporal order to form a skeleton sequence $\\mathbf{X}_i\\in \\mathbb{R}^{T\\times V\\times C}$ with length $T$ . A skeleton action recognizer takes $M$ skeleton sequences as input, and outputs a score vector $\\mathbf{S}\\in \\mathbb{R}^K$ for video-level action recognition, where $K$ is the number of action classes.", + "bbox": [ + 75, + 671, + 468, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spatial-temporal GCNs are widely used to model skeleton sequences for action classification, which focus on modeling intra-sequence skeleton dynamics. Most GCN methods [64, 47, 7] share a similar high-level design (Fig 2(a)), where a backbone with $\\mathrm{N}_1$ GCN blocks is used to extract features of each skeleton sequence. A GCN block com", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "prises a GCN layer and a TCN layer. The GCN layer performs spatial modeling by fusing joint features within the same skeleton. The TCN layer employs 1D Conv to capture the motion patterns of each joint. GCNs model long skeleton sequences that span the entire video in parallel, and aggregate multiple sequence features with average pooling to predict the video-level action. On existing skeleton action recognition benchmarks, GCNs exhibit superior intrasequence modeling capability than other architectures (e.g., RNN [73, 33] or Transformers [42, 68]).", + "bbox": [ + 496, + 313, + 893, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SkeleTR inherits the effective intra-sequence modeling of GCNs, while also enhancing the flexibility, generality, and inter-person modeling capability. SkeleTR takes $M$ short skeleton sequences as input (Sec. 3.2), and outputs one feature vector per sequence. It first adopts a GCN backbone to model the motion patterns of each skeleton sequence, then uses stacked Transformer encoders to capture the interactions and leverage long-range context (Sec. 3.3). Between two stages, we use a Mix Pooling layer to compress spatiotemporal sequence features into a compact multi-view representation to keep computation tractable. Fig 2 compares the design of SkeleTR and previous GCN methods.", + "bbox": [ + 496, + 465, + 893, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. The Input Construction of SkeleTR", + "text_level": 1, + "bbox": [ + 500, + 662, + 808, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For skeleton action classification in real-world scenarios, prior works [64, 47, 36] use a few long skeleton sequences as input, where skeletons at different timestamps are linked with rudimentary association methods. They use a naive confidence score based matching to link skeletons to sequences as long as the entire video (as shown in Fig 3). Specifically, they mostly keep $\\hat{M} = 2$ skeletons with the highest confidence scores per frame, and link skeletons based on confidence rank. In cases where it is applied to challenging videos involving multiple persons, this strategy can lead to noisy skeleton sequences and sub-optimal performance.", + "bbox": [ + 496, + 686, + 890, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, we propose to sample a large number of short skeleton sequences as input, and link skeletons at differ", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13636", + "bbox": [ + 480, + 925, + 519, + 936 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f8a33ba00e35f6bb53ce07affe21e3b7785cdd0c5d82e6a3df0dfdc252ce0fe9.jpg", + "image_caption": [ + "Figure 3: The difference of input formats. Prior works link skeletons at different timestamps based on the confidence score rank to form $\\hat{M} = 2$ long sequences, each covers the entire video. SkeleTR, instead, directly takes multiple shorter sequences as input. Each sequence only covers $1\\sim 2$ seconds, thus has better consistency than long sequences." + ], + "image_footnote": [], + "bbox": [ + 99, + 84, + 444, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ent timestamps into sequences based on their IOU similarity. Specifically, to sample sequences with length $T$ , we first uniformly select a set of key frames with a stride of $T / 2$ . For each skeleton in every key frame, we create a sequence by linking skeletons within a window of $T$ frames using IoU-based matching [3], resulting in a total of $M'$ sequences ( $M'$ is variable for different videos). Our new input format is compatible with a variable number of input sequences. By using different numbers of sequences ( $M$ ) for each training batch, SkeleTR can handle inputs with different $M$ , which helps to reduce the usages of zero-padding at test time and improves efficiency.", + "bbox": [ + 75, + 354, + 468, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Inter-sequence Modeling Stage", + "text_level": 1, + "bbox": [ + 73, + 542, + 346, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After processing short skeleton sequences with the GCN backbone, SkeleTR employs stacked Transformer encoders to model inter-sequence relationships. The process starts with reducing the feature dimension using a Mix Pooling layer, resulting in $L$ feature vectors for each sequence. Next, a sequence-level spatial-temporal embedding is combined with the feature vectors. The resulting sequence features $(M\\times L$ vectors in total) are concatenated and modeled using a standard Transformer with multiple encoder blocks [14]. This architecture is depicted in Fig 2(b). We present more details in the next paragraphs.", + "bbox": [ + 71, + 564, + 464, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mix Pooling. The cost of computing the attention matrices in a Transformer block grows quadratically with the input dimensionality, resulting in a computational cost of $\\Theta((M \\times T \\times V)^2)$ in our case, if no dimension reduction is performed ( $L = T \\times V$ ). Given a video with a typical configuration ( $T = 15$ , $V = 17$ ) and a Transformer consisting of only $\\mathrm{N}_2 = 2$ encoder blocks, the Transformer itself consumes 52.7 GFLOPs, which is impractical to compute. Another extreme is to remove the dimension $T$ and $V$ with global average pooling, which reduces the computational complexity to $\\Theta(M^2)$ but also results in loss of fine-grained", + "bbox": [ + 71, + 733, + 465, + 897 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "features for each skeleton sequence.", + "bbox": [ + 500, + 90, + 738, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Between two extremes $(L = T\\times V$ or $L = 1\\times 1)$ we adopt multi-stream partial dimension reductions to reduce the computation cost. We denote a partial dimension reduction stream as $A\\times B$ , indicating reducing temporal dimension $(T)$ to $A$ and spatial dimension $(V)$ to $B$ . Each stream provides a specific view of the spatio-temporal skeleton features. Multiple reduction streams can be combined in Mix Pooling to compress raw features into a comprehensive representation. Specifically, we adopt 3 partial dimension reduction streams: $1\\times 1$ , $T\\times 1$ , and $1\\times P$ ( $V$ joints aggregated to $P = 5$ parts: head, two arms, two legs) (Fig 2(c)). Mix Pooling concatenates the output of 3 streams and keeps $L = 1 + T + P$ features per sequence. The heterogeneous features are then processed with one single Transformer. Mix Pooling enables the cross-view feature interaction and helps keep computations tractable, yielding a complexity of $\\Theta(M^2(1 + T + P)^2)$ ( $\\sim 0.6$ GFLOPs / video).", + "bbox": [ + 498, + 107, + 890, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sequence-level Spatial Temporal Embedding. We further combine $L$ features per sequence with a sequence-level spatial-temporal embedding. We utilize a fully-connected layer to encode both the bounding box and the normalized timestamp (computed based on the entire video) of the center skeleton for each sequence, which then serves as the embedding. The embeddings are added with the Mix-Pooled features, and further processed by a layer normalization.", + "bbox": [ + 498, + 364, + 890, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformer-based Interaction Modeling. We use a standard Transformer with $\\mathrm{N}_2$ encoder blocks for modeling sequence features. Each block consists of a multi-head self-attention layer (MHSA) and a feed-forward network (FFN), and adopts the post-layernorm style. The multi-head design in the encoder blocks helps to capture various kinds of relationships among skeleton features, including but not limited to the inter-person interactions and intra-person long-range context. At last, we use intra-sequence average pooling to aggregate $L = (1 + T + P)$ features per sequence into a single one. The output sequence features, which incorporate both person interactions and a comprehensive context, are then used to predict the instance-level or video-level action. We present the detailed configuration of encoder blocks in the appendix.", + "bbox": [ + 498, + 486, + 890, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. SkeleTR as a General Framework", + "text_level": 1, + "bbox": [ + 500, + 724, + 795, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As a general framework for skeleton action recognition, SkeleTR can be applied to various skeleton-based action recognition tasks. It also enables and benefits from transfer learning and joint learning across different tasks.", + "bbox": [ + 498, + 748, + 890, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video-level Action Classification. The tasks of action classification [6, 65] and group activity recognition [24, 9] both take as input a video and predict the action at video-level. Note that the input videos may contain a variable number of people. In SkeleTR, we sample numerous short skeleton sequences that last $1 \\sim 2$ seconds. We then average the $M$", + "bbox": [ + 498, + 810, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13637", + "bbox": [ + 478, + 924, + 517, + 935 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e18071554911c9145d282bce8b7720c319570936826c1ff56d1f7d6887400996.jpg", + "image_caption": [ + "Figure 4: The curve of training / validation loss and validation mAP on AVA. Joint training with Posetics [65] significantly alleviates overfitting for AVA action detection. The jointly-trained model achieves much higher validation mAP, with the gap of training / validation losses largely shrinks." + ], + "image_footnote": [], + "bbox": [ + 76, + 89, + 467, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "output features, and use a fully-connected layer to predict the video action class. Thanks to the local motion modeling with GCN and interaction modeling with Transformers, SkeleTR achieves strong performance in both action classification and group activity recognition.", + "bbox": [ + 75, + 295, + 468, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instance-level Action Recognition. Spatio-temporal action detection [22] aims to recognize actions of each person at specific timestamps (instance-level recognition). For this task, we sample short skeleton sequences with length $T^{1}$ in a longer temporal window $(5 \\sim 10s)$ as input, and utilize the sequence-level features for instance-level action recognition. SkeleTR enables the modeling of both intra-sequence motion patterns and inter-sequence interactions for skeleton-based spatio-temporal action detection.", + "bbox": [ + 75, + 372, + 468, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transfer Learning and Joint Learning. SkeLETR enables to: 1. transfer the representation learned on one task to another new task; 2. jointly train different datasets and tasks, which in turn alleviate the overfitting issues on small datasets. Fig 4 shows the curves of training and validation losses, and validation mAP for AVA. We see that joint training serves as a strong regularizer and alleviates the severe overfitting. With joint training, the gap between training and validation losses largely shrinks, and the recognition performance on the validation set becomes much stronger.", + "bbox": [ + 75, + 508, + 468, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 676, + 207, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 702, + 179, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Kinetics-400 (K400) [6] contains videos with variable number of people. Some K400 actions cannot be distinguished by only using skeletons, thus we use a subset of K400, named **Posetics** [65], as the main benchmark for action classification. Posetics merges together classes featuring identical skeletons, resulting in 320 total classes (e.g., 'eating burger' and 'eating hotdog' are merged into 'eating something'). For K400 and Posetics, we use 2D human skeletons provided by [16] (extracted with Faster", + "bbox": [ + 75, + 724, + 468, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RCNN [44] (R50) and HRNet [51]). We provide K400 results in the appendix.", + "bbox": [ + 500, + 90, + 888, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NTU-Inter is a commonly used benchmark for two-person interaction recognition [40, 41]. It's a subset of the inlab collected NTURGB+D [32, 46], containing only mutual action classes (with two subjects). Consistent with NTURGB+D, NTU-Inter includes 4 data splits: NTU60-XSub, NTU60-XView, NTU120-XSub, NTU120-XSet. We use the official 3D skeletons for NTU-Inter experiments.", + "bbox": [ + 496, + 122, + 890, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AVA 2.2 [22] is a spatio-temporal action detection benchmark. It's a multi-label dataset with 80 action classes, falling into three major categories: person movement, object manipulation, and person interaction. Following the common practice [60], we use Faster-RCNN (ResNeXt101-FPN [31, 62]) for person detection and HRNet for pose estimation. The detector is pretrained on COCO and finetuned on AVA training bboxes. Following the benchmark guidelines, we evaluate 60 action classes. We report the overall mAP and the mAP of each major category.", + "bbox": [ + 496, + 228, + 892, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Volleyball [24] is a group activity recognition dataset with 8 volleyball group activities. Each clip has up to 13 persons, with the center frame annotated with GT person boxes. Following common practice [72, 52], we use the tracklets from [45] and HRNet for pose extraction, train models on training and validation sets, and report the test set results.", + "bbox": [ + 496, + 380, + 890, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 482, + 718, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architecture. Unless otherwise noted, we adopt ST-GCN++ [15] as the GCN backbone. Particularly, we consider two variants: SkeleTR-S (with fewer GCN blocks and smaller output dimension, details in appendix), and SkeleTR-L (original ST-GCN++.). The inter-sequence modeling stage consists of a Mix Pooling module and two Transformer encoder blocks. The feature dimension in each encoder blocks equals to the output channels of the GCN backbone (128 for SkeleTR-S and 256 for SkeleTR-L).", + "bbox": [ + 496, + 506, + 890, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. We train all models end-to-end with SGD and batch size 128. For all models, we set the base learning rate (LR) to 0.1 and adopt a multi-step LR scheduler with the decay factor 0.1. For Poetics and NTU-Inter, we train models for 100 epochs, and set milestones at the $70_{th}$ and $85_{th}$ epoch. For AVA and Volleyball, we train the models for 40 epochs, and set milestones at the $30_{th}$ and $35_{th}$ epoch. We use cross entropy loss $(\\mathbf{L}_{\\mathbf{CE}})$ for Poetics, NTU-Inter and Volleyball, and use binary cross entropy loss $(\\mathbf{L}_{\\mathbf{BCE}})$ for AVA (multi-label). We explicitly set a loss weight $\\lambda = 100$ in all AVA experiments due to the small magnitude of $\\mathbf{L}_{\\mathbf{BCE}}$ . In Poetics-AVA joint training, we concatenate samples in two datasets and optimize the model with a weighted combination of two losses: $\\mathbf{L} = \\mathbf{L}_{\\mathbf{CE}} + \\lambda \\mathbf{L}_{\\mathbf{BCE}}$ . For AVA, we sample short skeleton sequences in a small temporal window (5s / 10s) to form a training sample.", + "bbox": [ + 496, + 642, + 892, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference. We follow standard practices [15] to predict ac", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "${}^{1}T$ is restricted by annotation granularity. For example,AVA [22] annotates atomic actions at 1fps, for which $T$ should cover at most 2 seconds.", + "bbox": [ + 76, + 875, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13638", + "bbox": [ + 480, + 925, + 517, + 936 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9f0e1a2e62bb365246c46e5d7e729ebf8dcb135b5576c7da581c1f505e885247.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsPosetics Top-1Posetics Top-5
ST-TR [42]47.571.3
MS-G3D [36]52.675.8
CTR-GCN [7]54.176.7
ST-GCN++ [15]52.576.0
UNIK [65]52.575.7
PoseC3D [16]53.177.1
SkeleTR-S55.978.9
SkeleTR-L57.980.3
SkeleTR [C]57.680.1
", + "bbox": [ + 84, + 89, + 236, + 202 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3ef1279ff35c90915d69990681f905b3af2f5934aa31b71d3ba351c840fcfbbd.jpg", + "table_caption": [ + "(a) Action Classification: Poetics." + ], + "table_footnote": [], + "table_body": "
MethodsNTU60 XSubNTU60 XViewNTU120 XSubNTU120 XSet
ST-LSTM [33]83.087.363.066.6
LSTM-IRN [41]90.593.577.779.6
GeomNet [38]93.696.386.587.6
IGFormer [40]93.696.585.486.5
AS-GCN [29]87.192.077.879.3
CTR-GCN [7]91.694.383.284.4
SkeleTR-S94.697.187.488.0
SkeleTR-L94.797.587.888.3
SkeleTR [C]94.897.787.888.3
", + "bbox": [ + 246, + 89, + 478, + 202 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1fc1872797ddb842aecb2ac479c5cce79e57cc32c3f7359c997c45ed323f17d7.jpg", + "table_caption": [ + "(b) Action Recognition: NTU-Inter." + ], + "table_footnote": [], + "table_body": "
MethodsOverall mAPPerson MovementObject ManipulationPerson Interaction
ST-TR [42]17.337.38.418.8
MS-G3D [36]18.438.79.619.7
CTR-GCN [7]18.939.49.820.5
ST-GCN++ [15]18.639.19.720.0
UNIK [65]19.039.79.920.4
PoseC3D [16]18.339.39.319.1
SkeleTR-S22.843.812.326.9
SkeleTR-S*26.848.217.727.8
SkeleTR [C]*27.350.917.328.3
", + "bbox": [ + 490, + 89, + 754, + 202 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ae858a17cccd0531a041cd0b2a15971973ac6cee21d57dfbbe43ac4f38a9159a.jpg", + "table_caption": [ + "(c) Spatio-temporal Action Detection: AVA v2.2." + ], + "table_footnote": [], + "table_body": "
MethodsVolleyball Top-1
AT+ [20]92.3
POGARS [52]93.2
COMPOSER [72]93.7
ST-TR [42]93.3
CTR-GCN [7]93.6
ST-GCN++ [15]93.3
SkeleTR-S94.4
SkeleTR [C]94.4
", + "bbox": [ + 763, + 90, + 890, + 193 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion labels for trimmed videos. For AVA untrimmed videos, in inference we use a sliding window strategy to generate predictions, with a 1-second time step. We get an action prediction for a human proposal from each window containing the proposal, and then average all predictions associated with a human proposal to get the final result.", + "bbox": [ + 75, + 277, + 468, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 76, + 378, + 215, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare SkeleTR with state-of-the-art methods on three different skeleton action understanding tasks: action classification, spatio-temporal action detection, and group activity recognition. Unless otherwise noted, all methods use the same skeleton input as SkeleTR.", + "bbox": [ + 75, + 402, + 468, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Action Classification. Table 1a compares SkeleTR with state-of-the-art methods on Poetics. We use official training configurations when available [36, 16, 65]. For methods without official configurations [7, 15], we use the same hyper-parameter setting as SkeleTR. SkeleTR with all GCN backbones notably outperforms existing methods. Compared to CTR-GCN, SkeleTR-L achieves $3.6\\%$ higher Top-1 accuracy. Quantitative analysis shows that the improvement is more significant for multi-person scenarios: actions with $\\geq 5$ persons per frame have around $10\\%$ improvement in mean class Top-1, which is twice as much as the improvement for other actions. SkeleTR also achieves the highest accuracy on Poetics (with OpenPose skeletons) and K400 (with HRNet/OpenPose skeletons). Table 1b shows that on the more specific task of two-person interaction recognition, SkeleTR also outperforms existing interaction recognition methods [41, 38, 40] and skeleton action classification methods by a large margin.", + "bbox": [ + 75, + 478, + 468, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Spatio-temporal Action Detection. Skeleton-based spatio-temporal action detection is a new setting first proposed in this paper and we use AVA as the primary benchmark. For comparisons, we adapt state-of-the-art skeleton action classification algorithms by taking a single skeleton sequence (with 2 seconds length, centered at a person pro", + "bbox": [ + 75, + 750, + 468, + 840 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bc82804ac0441bc86d05e139458f5e381dc65630a8d988f63d2d72d9766e24ff.jpg", + "table_caption": [ + "(d) Group Activity Recognition: Volleyball.", + "Table 1: Comparing SkeleTR with previous state-of-the-art on versatile skeleton action recognition tasks. [C] denotes CTR-GCN backbone; * denotes joint training with Poetics. For Poetics and AVA, we follow the common practice [7, 65, 16] and compare the two-stream performance. For NTU-Inter and Volleyball, we only compare the joint modality performance as done in the prior works [52, 72]. In Table 1d, the method $\\mathrm{AT^{+}}$ uses groundtruth bboxes of all video frames for skeleton extraction." + ], + "table_footnote": [], + "table_body": "
MethodsPosetics Top-1Params (M)GFLOPsInference Speed (2080Ti, clip/second))
ST-TR [42]47.456.4426.9297
MS-G3D [36]52.632.8220.8699
CTR-GCN [7]54.101.365.36195
ST-GCN++ [15]52.461.315.44235
UNIK [65]52.523.2812.15156
PoseC3D [16]53.091.9014.6953
SkeleTR-S55.850.812.74336
SkeleTR-L57.903.827.30200
", + "bbox": [ + 506, + 276, + 890, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2: Efficiency statistics of models in Table 1a.", + "bbox": [ + 537, + 392, + 852, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "posal in the key frame) as input to predict the action for this person. We train all models with the same hyper-parameter setting as SkeleTR. Due to the lack of global context and interaction modeling, all existing methods perform poorly on the new task, while SkeleTR significantly outperforms them. Compared to the baseline ST-GCN++, SkeleTR-S improves the overall mAP by $4.1\\%$ . Among three major action categories, SkeleTR-S makes the largest improvement on person interaction $(7.0\\%)$ , thanks to its great inter-person modeling capability. When jointly trained with Poetics, SkeleTR-S enjoys another $4.1\\%$ overall mAP improvement. However, this time the improvement is basically for person movement and object manipulation, since AVA-style person interactions rarely appear in Kinetics videos.", + "bbox": [ + 496, + 410, + 890, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Group Activity Recognition. Most prior works on skeleton-based group activity recognition do not utilize the effective GCN architecture to model intra-person motion patterns. Instead, they directly model the relationship among skeletons with complicated transformer architectures. For a comprehensive comparison, we also adapt the state-of-the-art skeleton action recognition algorithms and benchmark them on the Volleyball dataset. Table 1d shows that, with good capability of both intra-sequence motion modeling and inter-sequence relationship modeling, SkeleTR outperforms all previous state-of-the-art on Volleyball, including the adapted GCN-based approaches.", + "bbox": [ + 496, + 625, + 890, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Computational Efficiency. SkeleTR is designed to be efficient and lightweight. Table 2 shows that SkeleTR-S outperforms previous state-of-the-art methods for skeleton-based action classification in terms of accuracy, while having the model with the smallest parameter size, GFLOPs, and fastest inference speed. In the appendix, we provide ad", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2Joint and Bone. Each stream takes a separate modality derived from skeleton coordinates as input. For joint stream, the input is the coordinates itself. For bone stream it is the difference of neighboring joints. 3Please see the appendix for these results.", + "bbox": [ + 75, + 849, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13639", + "bbox": [ + 480, + 925, + 517, + 936 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/23c777619f48d6e785d9db6863a2e70da3858bec9d46bbaf265a70f2a6a4877a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PoseticsAVA v2.2
MethodsTop-1Top-5Overall mAPPerson MovementObject ManipulationPerson Interaction
RGB74.4291.2931.3048.4423.6732.74
ST-GCN++52.4675.9718.6239.109.6819.98
RGB + ST-GCN++75.9891.5831.9552.2823.1433.13
SkeleTR57.9080.3326.8348.1917.7227.75
RGB + SkeleTR76.7092.0234.3454.7825.7734.93
", + "bbox": [ + 83, + 88, + 468, + 180 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9d6cfb891329ed3768f4cdcfdcf2e9bab68409720107bcd1fee2cfa771cd2de2.jpg", + "table_caption": [ + "Table 3: SkeleTR demonstrates great complementarity when fused with RGB recognizers. We fuse predictions by state-of-the-art RGB recognizers (TimeSformer [2] for Posetics, ACAR-R101 [39] for AVA) with skeleton predictions with $2:1$ ratio." + ], + "table_footnote": [], + "table_body": "
Interaction ModelingPosetics Top-1AVA Overall mAP
NA48.2117.56
Add Global51.0620.00
Concat Global50.0120.14
Non-Local48.4319.78
Transformer Encoders52.8120.92
", + "bbox": [ + 101, + 241, + 447, + 325 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/96ebd670c653a5bd5f784e495ca0248772fdb08b35010629e2e0cc3ecb3ca1bc.jpg", + "table_caption": [ + "Table 4: Transformer-based interaction modeling significantly outperforms multiple baselines." + ], + "table_footnote": [], + "table_body": "
Feature AggregationPosetics Top-1AVA Overall mAPHead Computation (GFLOPs)
1 × 152.8120.920.013
T × P52.7821.505.211
T × 153.4820.640.356
1 × P53.4021.350.080
1 × V52.1521.010.429
1 × 1 + T × 153.7521.000.392
1 × 1 + 1 × P53.0021.430.101
T × 1 + 1 × P53.1620.870.550
1 × 1 + T × 1 + 1 × P53.9021.650.594
", + "bbox": [ + 99, + 358, + 447, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5: Comparing different feature aggregation strategies.", + "bbox": [ + 81, + 484, + 462, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ditional evidence of SkeleTR's computational efficiency by comparing it to other methods designed for different tasks, such as IGFormer [40] and COMPOSER [72].", + "bbox": [ + 75, + 502, + 468, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Improved complementarity with RGB modality. To demonstrate the complementarity of SkeleTR upon RGB methods, we directly fuse the prediction of SkeleTR and the state-of-the-art RGB-based action recognizers. Table 3 shows that late fusion with SkeleTR leads to a significant boost (2.3% Posetics top-1, 3% AVA mAP) compared to the RGB state-of-the-art, which notably outperforms ST-GCN++ late fusion. On AVA, the improvement is extremely large for person movement (6.3% mAP), which emphasizes the importance of the skeleton in modeling human motions.", + "bbox": [ + 75, + 547, + 468, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation on Inter-Sequence Modeling", + "text_level": 1, + "bbox": [ + 76, + 710, + 403, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation experiments to validate the efficacy of our inter-sequence modeling stage. In experiments, we adopt the light ST-GCN++ as the backbone. We set $M$ to 20, $T$ to 30, and set the temporal window size to 5s for AVA. We report the recognition performance with a single joint modality on Posetics and AVA.", + "bbox": [ + 75, + 734, + 468, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first compare the Transformer-based modeling with other simple baselines for modeling global context or interaction. In all experiments, we pool both spatial and temporal dimensions and retain a single feature vector per sequence. We consider the following baselines: 1. NA: no", + "bbox": [ + 75, + 825, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f3a2e78b803c0ced29c4e133a3cbd36555db4d785728b9355fda555da41689ee.jpg", + "image_caption": [ + "Figure 5: Visualization of the weights in different attention heads in the last transformer encoder. $\\mathbf{P}_{ij}$ denotes the short skeleton sequence of person $j$ at the timestamp $i$ ." + ], + "image_footnote": [], + "bbox": [ + 511, + 83, + 875, + 215 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "interaction modeling; 2. Add Global: process the average feature with an MLP to get a Global feature, then add it back to each sequence feature; 3. Concat Global: concatenate the average feature with each sequence feature, then process with an MLP; 4. Non-Local: process all sequence features with a standard non-local [59] block, and apply an MLP to the output. Table 4 shows that, despite its low computation cost (0.013 GFLOPs / clip), the Transformer markedly enhances recognition performance relatively to the NA baseline (4.6% Posetics top-1, 3.4% AVA mAP), and considerably outperforms all alternatives.", + "bbox": [ + 496, + 261, + 890, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Both Non-Local and Transformer encoders adopt selfattention to model the inter-sequence relationship, while there is a significant difference in recognition performance between them. The reason is that the multi-head design in Transformers allows for modeling diverse relationships between individual sequences. Visualizing the attention weights of different attention heads, we find that some of them have clearly interpretable semantics. Fig 5 shows the attention weights of a 'handshake' action. Of the eight attention heads, the first head aggregates temporal features within a single person, while the fourth head models interactions between individuals. Specifically, the attention weights of the fourth head reveal that $P_{21}$ and $P_{22}$ attend most strongly to other skeleton features, which corresponds to the exact moment of the handshaking interaction.", + "bbox": [ + 496, + 426, + 892, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We then delve deeper into the Mix Pooling design, which helps to keep fine-grained sequence features and serves as a good trade-off between efficiency and performance. Table 5 compares Mix Pooling $(1 \\times 1 + T \\times 1 + 1 \\times P)$ with other feature aggregation strategies. Mix Pooling outperforms each individual pooling strategy or the combination of two strategies. It achieves the best performance on both datasets with a reasonable computation cost of 0.594 GFLOPs per clip.", + "bbox": [ + 496, + 654, + 890, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation on Input Construction", + "text_level": 1, + "bbox": [ + 500, + 786, + 779, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section we demonstrate the advantages of our new input format, both in recognition performance and computational efficiency. We report the recognition performance with a single joint modality.", + "bbox": [ + 496, + 809, + 890, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Large Number of Short Sequences. We use a large number of short skeleton sequences as input for skeleton action", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13640", + "bbox": [ + 480, + 926, + 517, + 936 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71cca6e19234087cb9f46a241a258b42637d1c1c0b68cf026b54710c66e060d5.jpg", + "image_caption": [ + "Figure 6: Fix the total number of skeletons and adjust the sequence number. Posetics Top-1 accuracy is visualized." + ], + "image_footnote": [], + "bbox": [ + 99, + 84, + 444, + 222 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9cfc97a2cb2f89330f2ad113f250d2c9444f3ec54e75c5c5c1547cb10d843d99.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsMTPosetics Top-1Avg GFLOPs for \nPosetics Video Clips
SkeleTR-S203053.902.430
SkeleTR-L[5, 20]3055.064.934
SkeleTR-L203054.956.653
SkeleTR-S401554.172.789
SkeleTR-L[10, 40]1555.353.532
SkeleTR-L401555.557.394
", + "bbox": [ + 99, + 257, + 447, + 356 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "understanding tasks. Besides being a unified input format for different tasks, the new format also benefits standard skeleton-based action classification under real-world scenarios. To investigate the effect of sequence granularity, we fix the total number of skeletons $M \\times T$ , and adjust the sequence number $M$ . We compare three models: ST-GCN++, SkeleTR-S, and SkeleTR-L (Fig 6). Interestingly, we find that all models obtain their highest performance when using a large number of short sequences as input, specifically, $40 \\times 15$ . Meanwhile, the improvement is especially large for SkeleTR: $2.1\\%$ top-1 for SkeleTR-S, $3\\%$ top-1 for SkeleTR-L (compared to the $2 \\times 300$ setting).", + "bbox": [ + 75, + 409, + 468, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Flexible Number of Sequences. SkeleTR is compatible with different numbers of sequences. To train and test SkeleTR with an adaptive squence number, we first set a dynamic range $[M_{min}, M_{max}]$ , and for each training batch, we randomly sample $\\tilde{M}$ from the range and use it as the sequence number for all videos in the batch. For each testing video, we use the actual sequence number $M'$ if $M_{min} \\leq M' \\leq M_{max}$ , and cap it to $M_{min}$ or $M_{max}$ otherwise. We present an ablation study using two input formats: $20 \\times 30$ and $40 \\times 15$ . For each format, we use a dynamic range ([5, 20], [10, 40]) of $M$ to train SkeleTR-L on Poetics, and report the results in Table 6. Compared to using a fixed $M$ , using an adaptive $M$ can achieve comparable recognition performance with a largely reduced inference cost (26% less GFLOPs for $T = 30$ , 52% less for $T = 15$ ).", + "bbox": [ + 75, + 590, + 470, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Transfer Learning and Joint Learning", + "text_level": 1, + "bbox": [ + 76, + 830, + 406, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SkeleTR is a general framework for versatile skeleton action recognition tasks, which enables transfer learning and joint learning across different tasks. To demonstrate this,", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dcb059b5903146c79ed56b80d4fad8633e471d33b3ad31c18f9320d59a2acd0a.jpg", + "table_caption": [ + "Table 6: Sequence number $M$ : fixed or adaptive? Utilizing an adaptive sequence number $M$ can save great amounts of computational resources and achieve comparable performance." + ], + "table_footnote": [], + "table_body": "
TrainingTemporal WindowOverall mAPPerson Movement mAPObject Manipulation mAPPerson Interaction mAP
From Scratch521.6542.3411.3725.66
Finetuning522.1243.0211.9125.76
Joint Training523.9244.2514.6526.06
From Scratch1021.3342.2411.2624.68
Finetuning1023.1144.0613.3225.85
Joint Training1025.6947.1316.5326.64
", + "bbox": [ + 524, + 87, + 870, + 189 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7: Transfer Learning and Joint Learning with SkeleTR-S. For a target task with limited training data, an auxiliary dataset can be leveraged to improve the performance and mitigate the overfitting, either by transfer learning or joint learning.", + "bbox": [ + 496, + 193, + 890, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we consider two tasks: Poetics action classification and AVA spatio-temporal action detection (with AVA being the smaller dataset). From-scratch training on AVA is easy to overfit (Fig 4). Pretraining on Poetics or joint training on AVA and Poetics both help to mitigate the overfitting, with joint training leading to the best recognition performance. Enlarging the temporal window size for AVA sampling, which leads to less diversified training samples, amplifies the gain of pretraining or joint training.", + "bbox": [ + 496, + 253, + 890, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Limitations", + "text_level": 1, + "bbox": [ + 500, + 404, + 625, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SkeleTR is currently trained with skeletons pre-extracted from RGB videos, and the adopted pose estimation solution is relatively heavy. To use SkeleTR in real-world apps with limited computation resource, additional efforts are required to switch to a light-weight pose estimation solution that can perform fast inference on CPUs or edge devices, and to retrain the models accordingly. Besides, the short sequence format proposed by SkeleTR strikes a good balance between mitigating skeleton matching noises and modeling intra-sequence dynamics in real-world scenarios. The format may not lead to extra benefits when the ground-truth skeleton sequences are available (details in appendix).", + "bbox": [ + 496, + 429, + 890, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 628, + 617, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented SkeleTR, a unified framework for skeleton action recognition in general scenarios. It extends existing GCN backbones and achieves state-of-the-art performance on multiple skeleton action understanding tasks, including video-level, instance-level, and group-level action recognition. Besides, SkeleTR is compatible with transfer learning and joint learning across different datasets and tasks, which further improves the capability of learned representations.", + "bbox": [ + 496, + 656, + 890, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Broader Impact. Using human skeletons for action recognition is more robust to environment changes and can help preserve user privacy. Real-world applications of skeleton action recognition include monitoring patient or elderly health, enabling autonomous driving, and facilitating human-computer interactions. However, there could be unintended usages and we advocate responsible usage complying with applicable laws and regulations.", + "bbox": [ + 496, + 779, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13641", + "bbox": [ + 480, + 925, + 517, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 3", + "[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 2, 3, 7", + "[3] Erik Bochinski, Volker Eiselein, and Thomas Sikora. High-speed tracking-by-detection without using image information. In AVSS, 2017. 4", + "[4] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. TPAMI, 2019. 2", + "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3", + "[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. 2, 4, 5", + "[7] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In ICCV, 2021. 1, 2, 3, 6", + "[8] Hyung-gun Chi, Myoung Hoon Ha, Seunggeun Chi, Sang Wan Lee, Qixing Huang, and Karthik Ramani. Infogcn: Representation learning for human skeleton-based action recognition. In CVPR, 2022. 1", + "[9] Wongun Choi, Khuram Shahid, and Silvio Savarese. What are they doing?: Collective activity classification using spatio-temporal relationship among people. In ICCVW, 2009. 2, 4", + "[10] Vasileios Choutas, Philippe Weinzaepfel, Jérôme Revaud, and Cordelia Schmid. Potion: Pose motion representation for action recognition. In CVPR, 2018. 2", + "[11] Liu Chunhui, Hu Yueyu, Li Yanghao, Song Sijie, and Liu Jiaying. Pku-mmd: A large scale benchmark for continuous multi-modal human action understanding. arXiv:1703.07475, 2017. 2", + "[12] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv:2003.09003, 2020. 1", + "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv:1810.04805, 2018.3", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv:2010.11929, 2020.3,4", + "[15] Haodong Duan, Jiaqi Wang, Kai Chen, and Dahua Lin. Pyskl: Towards good practices for skeleton action recognition. ACMM, 2022. 1, 2, 5, 6" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Haodong Duan, Yue Zhao, Kai Chen, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. In CVPR, 2022. 1, 2, 5, 6", + "[17] Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 2", + "[18] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. 3", + "[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 2", + "[20] Kirill Gavrilyuk, Ryan Sanford, Mehrsan Javan, and Cees GM Snoek. Actor-transformers for group activity recognition. In CVPR, 2020. 6", + "[21] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In ICCV, 2017. 2", + "[22] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, 2018. 1, 2, 5", + "[23] Mohamed E Hussein, Marwan Torki, Mohammad A Gowayyed, and Motaz El-Saban. Human action recognition using a temporal hierarchy of covariance descriptors on 3d joint locations. In IJCAI, 2013. 2", + "[24] Mostafa S Ibrahim, Srikanth Muralidharan, Zhiwei Deng, Arash Vahdat, and Greg Mori. A hierarchical deep temporal model for group activity recognition. In CVPR, 2016. 1, 2, 4, 5", + "[25] Haroon Idrees, Amir R Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The thumos challenge on action recognition for videos \"in the wild\". CVIU, 2017. 2", + "[26] Yanli Ji, Guo Ye, and Hong Cheng. Interactive body part contrast mining for human interaction recognition. In 2014 IEEE international conference on multimedia and expo workshops (ICMEW), pages 1-6. IEEE, 2014. 2", + "[27] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal-directed human activities. In CVPR, 2014. 2", + "[28] Ang Li, Meghanana Thotakuri, David A Ross, João Carreira, Alexander Vostrikov, and Andrew Zisserman. Theava-kinetics localized human actions video dataset. arXiv:2005.00214, 2020. 2", + "[29] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In CVPR, 2019. 6", + "[30] Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang. Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In ICCV, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13642", + "bbox": [ + 480, + 926, + 519, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 5", + "[32] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+d 120: A large-scale benchmark for 3d human activity understanding. TPAMI, 2020. 1, 5", + "[33] Jun Liu, Amir Shahroudy, Dong Xu, and Gang Wang. Spatio-temporal LSTM with trust gates for 3d human action recognition. In ECCV, 2016. 3, 6", + "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3", + "[35] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, 2022. 3", + "[36] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In CVPR, 2020. 1, 2, 3, 6", + "[37] Qianhui Men, Edmond SL Ho, Hubert PH Shum, and Howard Leung. A two-stream recurrent network for skeleton-based human interaction recognition. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 2771-2778. IEEE, 2021. 2", + "[38] Xuan Son Nguyen. Geomnet: A neural network based on riemannian geometries of spd matrix space and cholesky space for 3d skeleton-based interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13379-13389, 2021. 6", + "[39] Junting Pan, Siyu Chen, Mike Zheng Shou, Yu Liu, Jing Shao, and Hongsheng Li. Actor-context-actor relation network for spatio-temporal action localization. In CVPR, 2021. 2, 7", + "[40] Yunsheng Pang, Qiuhong Ke, Hossein Rahmani, James Bailey, and Jun Liu. Igformer: Interaction graph transformer for skeleton-based human interaction recognition. In ECCV, 2022. 2, 5, 6, 7", + "[41] Mauricio Perez, Jun Liu, and Alex C Kot. Interaction relational network for mutual action recognition. IEEE Transactions on Multimedia, 24:366-376, 2021. 2, 5, 6", + "[42] Chiara Plizzari, Marco Cannici, and Matteo Matteucci. Skeleton-based action recognition via spatial and temporal transformer networks. CVIU, 2021. 3, 6", + "[43] Michalis Raptis and Leonid Sigal. Poselet key-framing: A model for human activity recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2650-2657, 2013. 2", + "[44] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 2015. 5", + "[45] Kohei Sendo and Norimichi Ukita. Heatmapping of people involved in group activities. In MVA, 2019. 5", + "[46] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+d: A large scale dataset for 3d human activity analysis. In CVPR, 2016. 1, 2, 5" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In CVPR, 2019. 2, 3", + "[48] Bing Shuai, Alessandro Bergamo, Uta Buechler, Andrew Berneshawi, Alyssa Boden, and Joseph Tighe. Large scale real-world multi-person tracking. ECCV, 2022. 1", + "[49] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. NeurIPS, 2014. 2", + "[50] Sebastian Stein and Stephen J McKenna. Combining embedded accelerometers with computer vision for recognizing food preparation activities. In UbiComp, 2013. 2", + "[51] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In CVPR, 2019. 2, 5", + "[52] Haritha Thilakarathne, Aiden Nibali, Zhen He, and Stuart Morgan. Pose is all you need: The pose only group activity recognition system (pogars). Machine Vision and Applications, 2022. 5, 6", + "[53] Arash Vahdat, Bo Gao, Mani Ranjbar, and Greg Mori. A discriminative key pose sequence model for recognizing human interactions. In 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops), pages 1729-1736. IEEE, 2011. 2", + "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3", + "[55] Raviteja Vemulapalli, Felipe Arrate, and Rama Chellappa. Human action recognition by representing 3d skeletons as points in a lie group. In CVPR, 2014. 2", + "[56] Hongsong Wang and Liang Wang. Modeling temporal dynamics and spatial configurations of actions using two-stream recurrent neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 499-508, 2017. 2", + "[57] Jiang Wang, Zicheng Liu, Ying Wu, and Junsong Yuan. Mining actionlet ensemble for action recognition with depth cameras. In CVPR. IEEE, 2012. 2", + "[58] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In ECCV, 2016. 2", + "[59] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, 2018. 2, 7", + "[60] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In CVPR, 2019. 5", + "[61] Fanyi Xiao, Yong Jae Lee, Kristen Grauman, Jitendra Malik, and Christoph Feichtenhofer. Audiovisual slowfast networks for video recognition. arXiv:2001.08740, 2020. 2", + "[62] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017. 5", + "[63] Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Xia, Zhuowen Tu, and Stefano Soatto. Long short-term transformer for online action detection. NeurIPS, 2021. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13643", + "bbox": [ + 480, + 926, + 517, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 1, 2, 3", + "[65] Di Yang, Yaohui Wang, Antitza Dantcheva, Lorenzo Garattoni, Gianpiero Francesca, and François Bremond. Unik: A unified framework for real-world skeleton-based action recognition. ArXiv, 2021. 2, 4, 5, 6", + "[66] Kiwon Yun, Jean Honorio, Debaleena Chattopadhyay, Tamara L Berg, and Dimitris Samaras. Two-person interaction detection using body-pose features and multiple instance learning. In 2012 IEEE computer society conference on computer vision and pattern recognition workshops, pages 28–35. IEEE, 2012. 2", + "[67] Yimeng Zhang, Xiaoming Liu, Ming-Ching Chang, Weina Ge, and Tsuhan Chen. Spatio-temporal phrases for activity recognition. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part III 12, pages 707-721. Springer, 2012. 2", + "[68] Yuhan Zhang, Bo Wu, Wen Li, Lixin Duan, and Chuang Gan. Stst: Spatial-temporal specialized transformer for skeleton-based action recognition. In ACMMM, 2021. 3", + "[69] Zhengyou Zhang. Microsoft Kinect sensor and its effect. IEEE multimedia, 2012. 2", + "[70] Hang Zhao, Antonio Torralba, Lorenzo Torresani, and Zhicheng Yan. Hacs: Human action clips and segments dataset for recognition and temporal localization. In ICCV, 2019. 2", + "[71] Jiaojiao Zhao, Yanyi Zhang, Xinyu Li, Hao Chen, Bing Shuai, Mingze Xu, Chunhui Liu, Kaustav Kundu, Yuanjun Xiong, Davide Modolo, et al. Tuber: Tubelet transformer for video action detection. In CVPR, 2022. 3", + "[72] Honglu Zhou, Asim Kadav, Aviv Shamsian, Shijie Geng, Farley Lai, Long Zhao, Ting Liu, Mubbasir Kapadia, and Hans Peter Graf.Composer: Compositional learning of group activity in videos.ECCV, 2022.5,6,7", + "[73] Wentao Zhu, Cuiling Lan, Junliang Xing, Wenjun Zeng, Yanghao Li, Li Shen, and Xiaohui Xie. Co-occurrence feature learning for skeleton based action recognition using regularized deep LSTM networks. In AAAI, 2016. 3", + "[74] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv:2010.04159, 2020. 3" + ], + "bbox": [ + 78, + 90, + 467, + 699 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13644", + "bbox": [ + 480, + 926, + 517, + 936 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/SkeleTR_ Towards Skeleton-based Action Recognition in the Wild/27665a84-18aa-44bc-b25b-e6429d7734c6_model.json b/2023/SkeleTR_ Towards Skeleton-based Action Recognition in the Wild/27665a84-18aa-44bc-b25b-e6429d7734c6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..92d3733197f8082ac3cb751b983243755e27ecc5 --- /dev/null +++ b/2023/SkeleTR_ Towards Skeleton-based Action Recognition in the Wild/27665a84-18aa-44bc-b25b-e6429d7734c6_model.json @@ -0,0 +1,2356 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.153, + 0.131, + 0.819, + 0.154 + ], + "angle": 0, + "content": "SkeleTR: Towards Skeleton-based Action Recognition in the Wild" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.18, + 0.804, + 0.243 + ], + "angle": 0, + "content": "Haodong Duan\\(^{1*}\\) Mingze Xu\\(^{2}\\) Bing Shuai\\(^{2}\\) Davide Modolo\\(^{2}\\) Zhuowen Tu\\(^{2}\\) Joseph Tighe\\(^{2}\\) Alessandro Bergamo\\(^{2}\\) The Chinese University of Hong Kong AWS AI Labs" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.248, + 0.813, + 0.262 + ], + "angle": 0, + "content": "dhd.efz@gmail.com, {xumingze,bshuai,dmodolo,ztu,tighej,bergamo}@amazon.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.297, + 0.314, + 0.314 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.329, + 0.474, + 0.648 + ], + "angle": 0, + "content": "We present SkeleTR, a new framework for skeleton-based action recognition. In contrast to prior work, which focuses mainly on controlled environments, we target more general scenarios that typically involve a variable number of people and various forms of interaction between people. SkeleTR works with a two-stage paradigm. It first models the intra-person skeleton dynamics for each skeleton sequence with graph convolutions, and then uses stacked Transformer encoders to capture person interactions that are important for action recognition in general scenarios. To mitigate the negative impact of inaccurate skeleton associations, SkeleTR takes relative short skeleton sequences as input and increases the number of sequences. As a unified solution, SkeleTR can be directly applied to multiple skeleton-based action tasks, including video-level action classification, instance-level action detection, and group-level activity recognition. It also enables transfer learning and joint training across different action tasks and datasets, which result in performance improvement. When evaluated on various skeleton-based action recognition benchmarks, SkeleTR achieves the state-of-the-art performance." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.655, + 0.21, + 0.671 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Skeleton-based action recognition has received increasing attention in recent years due to its robustness against background and illumination changes [64, 16]. Existing methods [36, 7, 8, 15] mostly use Graph Convolutional Networks (GCNs) as backbones and focus on modeling spatiotemporal information from long skeleton sequences separately. While observing promising results in simplified and controlled environments [46, 32], they are less effective in more realistic scenarios with a variable number of people performing diverse actions. First, considering that linking skeletons (e.g., using skeletons' confidence scores [64]) in a long time span may inevitably generate more association errors [12, 48], requiring long sequences as input can bring" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.296, + 0.891, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.464, + 0.893, + 0.534 + ], + "angle": 0, + "content": "Figure 1: SkeleTR is a general framework for skeleton action recognition which can handle different tasks, including video-level action classification, instance-level action detection, and group-level activity recognition. It adopts a two-stage paradigm: the intra-sequence and inter-sequence modeling, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.894, + 0.613 + ], + "angle": 0, + "content": "lots of noises. Second, prior works pay little attention to modeling the relations among a group of people, which plays an important role in recognizing both person interactions (e.g., talking to someone [22]) and group activities (e.g., pass/win-point in a volleyball game [24])." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose SkeleTR for skeleton action recognition in real-world scenarios and handle multiple action tasks (i.e., video-level, instance-level, and group-level action recognition) in a general framework. SkeleTR samples a large number of short skeleton sequences instead of a few long ones (practice of prior works [64, 16]) from each video. Each short sequence directly corresponds to an atomic action of a single person, thus the collection of multiple short sequences in space and time can represent hierarchical actions (with a longer time span) of a single person or a video-level action involving multiple people. The new input paradigm can be applied to instance-level and video-level recognition of actions with different temporal granularities, and it has several unique advantages. First, association errors are less likely to happen in a shorter skeleton sequence [12, 48]. We show that the improved sequence consistency leads to better intra-sequence modeling. Second, this new format is more flexible and helps to create comprehensive yet compact inputs. With a finer temporal" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.368, + 0.902 + ], + "angle": 0, + "content": "*The work was done during an Amazon internship." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "13634" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.152 + ], + "angle": 0, + "content": "granularity, the same input size budget can cover more human skeletons. It reduces redundant zero-paddings for a small number of people, and mitigates information loss during input sampling for a large number of people." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.153, + 0.468, + 0.424 + ], + "angle": 0, + "content": "SkeleTR adopts a two-stage paradigm as shown in Fig 1. Stage 1 performs intra-sequence modeling to capture the motion patterns of each individual skeleton sequence with a GCN backbone. The GCN backbone outputs a spatiotemporal feature for each sequence. Stage 2 concatenates all sequence features and feeds them into Transformer encoders to model inter-sequence relationships. At the end of stage 2, average pooling aggregates all features belonging to the same sequence into a single feature vector. By leveraging inter-sequence modeling, the output sequence feature incorporates the interactions between people and the global context of the video, which is helpful in both instance-level and video-level action recognition. To reduce the dimension of spatio-temporal skeleton features and keep the computations of Transformer tractable, we insert a novel Mix Pooling module before stage 2. It generates a compact representation with fine granularity for each sequence by applying multiple partial dimension reduction strategies in parallel." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.425, + 0.468, + 0.574 + ], + "angle": 0, + "content": "SkeleTR easily enables transfer learning and joint learning across different datasets and tasks. With a unified input format and model architecture, one can jointly train multiple tasks with different datasets, each with its own loss function and ground truth annotations. This allows for leveraging auxiliary datasets for pre-training or joint training of a target task with limited labeled data. In experiments, we find that joint training serves as a strong regularizer and improves recognition performance, especially in scenarios with limited training data." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.576, + 0.468, + 0.83 + ], + "angle": 0, + "content": "To evaluate the effectiveness of SkeleTR, we conduct comprehensive study on three skeleton-based action recognition tasks: action classification [6, 65, 46], spatiotemporal action detection [22], and group activity recognition [24]. Experimental results show that SkeleTR works well with different GCN backbones and significantly outperforms prior works on all three tasks. For example, with ST-GCN++ [15] backbone, SkeleTR outperforms the state-of-the-art methods by \\(3.8\\%\\), \\(3.8\\%\\), \\(0.7\\%\\), and \\(1.1\\%\\) on Poetics [64], AVA [22], Volleyball [24], and NTU-Inter [40]. With Poetics joint training, the AVA performance is further improved by \\(4\\%\\) mAP. The great capability of SkeleTR leads to another noteworthy result: we first observe the strong complementarity of skeleton action recognition. Combining SkeleTR prediction and the RGB state-of-the-art [2, 39] with simple late fusion leads to great performance boost (2.3% Poetics Top-1, 3% AVA mAP)." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.845, + 0.225, + 0.86 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.871, + 0.468, + 0.899 + ], + "angle": 0, + "content": "Human Action Understanding. Action classification [6, 21] is the most fundamental task in human action under" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.89, + 0.332 + ], + "angle": 0, + "content": "standing, aiming at recognizing the video-level action in a short video clip. Another task of great interest is recognizing the instance/person-level actions in a video, either at the atomic [22, 28] or hierarchical [30] level. Meanwhile, other areas of interest include group activity recognition [24, 9], action localization [17, 25, 70], or action segmentation [27, 50]. Various modalities in videos can be utilized to solve the action understanding tasks, such as RGB appearance [58, 6, 19, 59], optical flow [49], audio waves [61], and human skeletons [64, 16]. Spatio-temporal representations learned end-to-end from RGB appearance yield good performance on standard benchmarks. However, they are data intensive and less robust to context changes. In this work, we propose a unified framework to learn video representations based on human skeletons, that can be easily applied to multiple action tasks." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.338, + 0.89, + 0.699 + ], + "angle": 0, + "content": "Skeleton-based Action Recognition. Human skeletons are robust against background and illumination changes [64, 47], and can be easily obtained with sensors [69] or pose estimators [4, 51]. In skeleton action recognition, the majority of research has focused on modeling the intra-person skeleton dynamics. With the rise of deep learning, manual feature engineering [23, 55, 57] has gradually been replaced by end-to-end representation learning with deep neural architectures (GCN [64], 2D-CNN [10], or 3D-CNN [16]). Among them, GCN-based approaches [64, 47, 36, 7] have been the most popular paradigm, which construct spatial-temporal graphs of human keypoint coordinates and perform spatial and temporal modeling. These methods model the intra-person skeleton dynamics for in-lab collected videos [46, 11], featuring small number of persons (1 or 2) per video and accurate person identities. For more challenging scenarios [6], the model performance decreases due to the noises produced in matching skeletons to sequences. Besides, GCN approaches primarily focus on predicting video-level actions, and do not benefit from context or interaction modeling when applied to instance-level action recognition. To this end, we propose SkeleTR, which expands existing GCN approaches and works universally well for diversified scenarios and recognition tasks." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.705, + 0.89, + 0.899 + ], + "angle": 0, + "content": "Human Interaction Recognition. Human interaction recognition [43, 53, 67] is a sub-task of action recognition. It aims at identifying and categorizing activities between people. Early works [66, 26] built hand-crafted features for interaction modeling and adopted classic machine learning algorithms for activity classification. Subsequent works proposed to use two-stream RNNs to model the two-person interactions [41, 37, 56]. Recently, IGFormer [40] first proposed to use a two-stream Transformer for two-person interaction modeling and observed improved performance. However, most deep methods share the same limitations: 1. they are largely restricted to two-person scenarios and tend to suffer from exponentially growing inference" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.926, + 0.518, + 0.936 + ], + "angle": 0, + "content": "13635" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.082, + 0.849, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.253, + 0.895, + 0.31 + ], + "angle": 0, + "content": "Figure 2: The architecture of SkeLETR. SkeLETR takes a large set of short skeleton sequences as input and performs a two-stage modeling process. It first adopts a GCN backbone to model the intra-sequence skeleton dynamics. Then it processes the sequence features with stacked Transformer encoders, which helps to model person interactions and leverage the long-range context. Between the two stages, we adopt a Mix Pooling layer to compress the spatio-temporal features and keep computations tractable." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.471, + 0.39 + ], + "angle": 0, + "content": "costs with the person number; 2. they fail to leverage the strong intra-sequence modeling capabilities of GCNs. To overcome these challenges, we introduce SkeleTR, which integrates intra-sequence and inter-sequence modeling with good scalability and small computation cost." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.391, + 0.471, + 0.603 + ], + "angle": 0, + "content": "Visual Transformer. With the great success in NLP [54, 13], Transformers have been increasingly applied to diversified vision tasks [14, 34, 5, 74] as well as the RGB-based action understanding [35, 18, 1, 2, 63, 71], and have achieved promising results. For skeleton action recognition, there exist works [42, 68] that directly use Transformers for intra-sequence spatio-temporal modeling. However, without exploiting the locality, the Transformer-based motion modeling only leads to inferior recognition performance compared to GCN-based ones. In SkeleTR, to enjoy the best of both worlds, we adopt GCNs for intra-sequence motion modeling, and use Transformer encoders for intersequence relationship modeling among multiple persons' fine-grained skeleton features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.621, + 0.175, + 0.637 + ], + "angle": 0, + "content": "3. SkeleTR" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.648, + 0.188, + 0.663 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Both SkeleTR and prior GCN approaches directly take human joint coordinates as input. A person in a specific frame corresponds to a skeleton with \\(V\\) joints (each has \\(C\\) dimensions, \\(C\\) is 2 or 3). Skeletons in \\(T\\) different frames (1 skeleton per frame) can be linked in temporal order to form a skeleton sequence \\(\\mathbf{X}_i\\in \\mathbb{R}^{T\\times V\\times C}\\) with length \\(T\\). A skeleton action recognizer takes \\(M\\) skeleton sequences as input, and outputs a score vector \\(\\mathbf{S}\\in \\mathbb{R}^K\\) for video-level action recognition, where \\(K\\) is the number of action classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Spatial-temporal GCNs are widely used to model skeleton sequences for action classification, which focus on modeling intra-sequence skeleton dynamics. Most GCN methods [64, 47, 7] share a similar high-level design (Fig 2(a)), where a backbone with \\(\\mathrm{N}_1\\) GCN blocks is used to extract features of each skeleton sequence. A GCN block com" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.314, + 0.895, + 0.466 + ], + "angle": 0, + "content": "prises a GCN layer and a TCN layer. The GCN layer performs spatial modeling by fusing joint features within the same skeleton. The TCN layer employs 1D Conv to capture the motion patterns of each joint. GCNs model long skeleton sequences that span the entire video in parallel, and aggregate multiple sequence features with average pooling to predict the video-level action. On existing skeleton action recognition benchmarks, GCNs exhibit superior intrasequence modeling capability than other architectures (e.g., RNN [73, 33] or Transformers [42, 68])." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.895, + 0.649 + ], + "angle": 0, + "content": "SkeleTR inherits the effective intra-sequence modeling of GCNs, while also enhancing the flexibility, generality, and inter-person modeling capability. SkeleTR takes \\(M\\) short skeleton sequences as input (Sec. 3.2), and outputs one feature vector per sequence. It first adopts a GCN backbone to model the motion patterns of each skeleton sequence, then uses stacked Transformer encoders to capture the interactions and leverage long-range context (Sec. 3.3). Between two stages, we use a Mix Pooling layer to compress spatiotemporal sequence features into a compact multi-view representation to keep computation tractable. Fig 2 compares the design of SkeleTR and previous GCN methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.663, + 0.81, + 0.68 + ], + "angle": 0, + "content": "3.2. The Input Construction of SkeleTR" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.868 + ], + "angle": 0, + "content": "For skeleton action classification in real-world scenarios, prior works [64, 47, 36] use a few long skeleton sequences as input, where skeletons at different timestamps are linked with rudimentary association methods. They use a naive confidence score based matching to link skeletons to sequences as long as the entire video (as shown in Fig 3). Specifically, they mostly keep \\(\\hat{M} = 2\\) skeletons with the highest confidence scores per frame, and link skeletons based on confidence rank. In cases where it is applied to challenging videos involving multiple persons, this strategy can lead to noisy skeleton sequences and sub-optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In contrast, we propose to sample a large number of short skeleton sequences as input, and link skeletons at differ" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.52, + 0.938 + ], + "angle": 0, + "content": "13636" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.085, + 0.446, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.074, + 0.27, + 0.467, + 0.353 + ], + "angle": 0, + "content": "Figure 3: The difference of input formats. Prior works link skeletons at different timestamps based on the confidence score rank to form \\(\\hat{M} = 2\\) long sequences, each covers the entire video. SkeleTR, instead, directly takes multiple shorter sequences as input. Each sequence only covers \\(1\\sim 2\\) seconds, thus has better consistency than long sequences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.47, + 0.537 + ], + "angle": 0, + "content": "ent timestamps into sequences based on their IOU similarity. Specifically, to sample sequences with length \\( T \\), we first uniformly select a set of key frames with a stride of \\( T / 2 \\). For each skeleton in every key frame, we create a sequence by linking skeletons within a window of \\( T \\) frames using IoU-based matching [3], resulting in a total of \\( M' \\) sequences (\\( M' \\) is variable for different videos). Our new input format is compatible with a variable number of input sequences. By using different numbers of sequences (\\( M \\)) for each training batch, SkeleTR can handle inputs with different \\( M \\), which helps to reduce the usages of zero-padding at test time and improves efficiency." + }, + { + "type": "title", + "bbox": [ + 0.074, + 0.544, + 0.347, + 0.56 + ], + "angle": 0, + "content": "3.3. Inter-sequence Modeling Stage" + }, + { + "type": "text", + "bbox": [ + 0.073, + 0.565, + 0.465, + 0.731 + ], + "angle": 0, + "content": "After processing short skeleton sequences with the GCN backbone, SkeleTR employs stacked Transformer encoders to model inter-sequence relationships. The process starts with reducing the feature dimension using a Mix Pooling layer, resulting in \\(L\\) feature vectors for each sequence. Next, a sequence-level spatial-temporal embedding is combined with the feature vectors. The resulting sequence features \\((M\\times L\\) vectors in total) are concatenated and modeled using a standard Transformer with multiple encoder blocks [14]. This architecture is depicted in Fig 2(b). We present more details in the next paragraphs." + }, + { + "type": "text", + "bbox": [ + 0.073, + 0.734, + 0.466, + 0.898 + ], + "angle": 0, + "content": "Mix Pooling. The cost of computing the attention matrices in a Transformer block grows quadratically with the input dimensionality, resulting in a computational cost of \\(\\Theta((M \\times T \\times V)^2)\\) in our case, if no dimension reduction is performed (\\(L = T \\times V\\)). Given a video with a typical configuration (\\(T = 15\\), \\(V = 17\\)) and a Transformer consisting of only \\(\\mathrm{N}_2 = 2\\) encoder blocks, the Transformer itself consumes 52.7 GFLOPs, which is impractical to compute. Another extreme is to remove the dimension \\(T\\) and \\(V\\) with global average pooling, which reduces the computational complexity to \\(\\Theta(M^2)\\) but also results in loss of fine-grained" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.74, + 0.106 + ], + "angle": 0, + "content": "features for each skeleton sequence." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.108, + 0.892, + 0.364 + ], + "angle": 0, + "content": "Between two extremes \\((L = T\\times V\\) or \\(L = 1\\times 1)\\) we adopt multi-stream partial dimension reductions to reduce the computation cost. We denote a partial dimension reduction stream as \\(A\\times B\\), indicating reducing temporal dimension \\((T)\\) to \\(A\\) and spatial dimension \\((V)\\) to \\(B\\). Each stream provides a specific view of the spatio-temporal skeleton features. Multiple reduction streams can be combined in Mix Pooling to compress raw features into a comprehensive representation. Specifically, we adopt 3 partial dimension reduction streams: \\(1\\times 1\\), \\(T\\times 1\\), and \\(1\\times P\\) (\\(V\\) joints aggregated to \\(P = 5\\) parts: head, two arms, two legs) (Fig 2(c)). Mix Pooling concatenates the output of 3 streams and keeps \\(L = 1 + T + P\\) features per sequence. The heterogeneous features are then processed with one single Transformer. Mix Pooling enables the cross-view feature interaction and helps keep computations tractable, yielding a complexity of \\(\\Theta(M^2(1 + T + P)^2)\\) (\\(\\sim 0.6\\) GFLOPs / video)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.365, + 0.892, + 0.486 + ], + "angle": 0, + "content": "Sequence-level Spatial Temporal Embedding. We further combine \\( L \\) features per sequence with a sequence-level spatial-temporal embedding. We utilize a fully-connected layer to encode both the bounding box and the normalized timestamp (computed based on the entire video) of the center skeleton for each sequence, which then serves as the embedding. The embeddings are added with the Mix-Pooled features, and further processed by a layer normalization." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.487, + 0.892, + 0.713 + ], + "angle": 0, + "content": "Transformer-based Interaction Modeling. We use a standard Transformer with \\(\\mathrm{N}_2\\) encoder blocks for modeling sequence features. Each block consists of a multi-head self-attention layer (MHSA) and a feed-forward network (FFN), and adopts the post-layernorm style. The multi-head design in the encoder blocks helps to capture various kinds of relationships among skeleton features, including but not limited to the inter-person interactions and intra-person long-range context. At last, we use intra-sequence average pooling to aggregate \\(L = (1 + T + P)\\) features per sequence into a single one. The output sequence features, which incorporate both person interactions and a comprehensive context, are then used to predict the instance-level or video-level action. We present the detailed configuration of encoder blocks in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.797, + 0.74 + ], + "angle": 0, + "content": "3.4. SkeleTR as a General Framework" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.749, + 0.891, + 0.809 + ], + "angle": 0, + "content": "As a general framework for skeleton action recognition, SkeleTR can be applied to various skeleton-based action recognition tasks. It also enables and benefits from transfer learning and joint learning across different tasks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Video-level Action Classification. The tasks of action classification [6, 65] and group activity recognition [24, 9] both take as input a video and predict the action at video-level. Note that the input videos may contain a variable number of people. In SkeleTR, we sample numerous short skeleton sequences that last \\(1 \\sim 2\\) seconds. We then average the \\(M\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.518, + 0.936 + ], + "angle": 0, + "content": "13637" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.09, + 0.468, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.22, + 0.47, + 0.29 + ], + "angle": 0, + "content": "Figure 4: The curve of training / validation loss and validation mAP on AVA. Joint training with Posetics [65] significantly alleviates overfitting for AVA action detection. The jointly-trained model achieves much higher validation mAP, with the gap of training / validation losses largely shrinks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.47, + 0.372 + ], + "angle": 0, + "content": "output features, and use a fully-connected layer to predict the video action class. Thanks to the local motion modeling with GCN and interaction modeling with Transformers, SkeleTR achieves strong performance in both action classification and group activity recognition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.373, + 0.469, + 0.508 + ], + "angle": 0, + "content": "Instance-level Action Recognition. Spatio-temporal action detection [22] aims to recognize actions of each person at specific timestamps (instance-level recognition). For this task, we sample short skeleton sequences with length \\( T^{1} \\) in a longer temporal window \\( (5 \\sim 10s) \\) as input, and utilize the sequence-level features for instance-level action recognition. SkeleTR enables the modeling of both intra-sequence motion patterns and inter-sequence interactions for skeleton-based spatio-temporal action detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.47, + 0.661 + ], + "angle": 0, + "content": "Transfer Learning and Joint Learning. SkeLETR enables to: 1. transfer the representation learned on one task to another new task; 2. jointly train different datasets and tasks, which in turn alleviate the overfitting issues on small datasets. Fig 4 shows the curves of training and validation losses, and validation mAP for AVA. We see that joint training serves as a strong regularizer and alleviates the severe overfitting. With joint training, the gap between training and validation losses largely shrinks, and the recognition performance on the validation set becomes much stronger." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.677, + 0.208, + 0.695 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.703, + 0.18, + 0.717 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.726, + 0.47, + 0.864 + ], + "angle": 0, + "content": "Kinetics-400 (K400) [6] contains videos with variable number of people. Some K400 actions cannot be distinguished by only using skeletons, thus we use a subset of K400, named **Posetics** [65], as the main benchmark for action classification. Posetics merges together classes featuring identical skeletons, resulting in 320 total classes (e.g., 'eating burger' and 'eating hotdog' are merged into 'eating something'). For K400 and Posetics, we use 2D human skeletons provided by [16] (extracted with Faster" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.091, + 0.89, + 0.122 + ], + "angle": 0, + "content": "RCNN [44] (R50) and HRNet [51]). We provide K400 results in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.123, + 0.892, + 0.228 + ], + "angle": 0, + "content": "NTU-Inter is a commonly used benchmark for two-person interaction recognition [40, 41]. It's a subset of the inlab collected NTURGB+D [32, 46], containing only mutual action classes (with two subjects). Consistent with NTURGB+D, NTU-Inter includes 4 data splits: NTU60-XSub, NTU60-XView, NTU120-XSub, NTU120-XSet. We use the official 3D skeletons for NTU-Inter experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.229, + 0.893, + 0.38 + ], + "angle": 0, + "content": "AVA 2.2 [22] is a spatio-temporal action detection benchmark. It's a multi-label dataset with 80 action classes, falling into three major categories: person movement, object manipulation, and person interaction. Following the common practice [60], we use Faster-RCNN (ResNeXt101-FPN [31, 62]) for person detection and HRNet for pose estimation. The detector is pretrained on COCO and finetuned on AVA training bboxes. Following the benchmark guidelines, we evaluate 60 action classes. We report the overall mAP and the mAP of each major category." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.472 + ], + "angle": 0, + "content": "Volleyball [24] is a group activity recognition dataset with 8 volleyball group activities. Each clip has up to 13 persons, with the center frame annotated with GT person boxes. Following common practice [72, 52], we use the tracklets from [45] and HRNet for pose extraction, train models on training and validation sets, and report the test set results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.483, + 0.719, + 0.499 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.643 + ], + "angle": 0, + "content": "Architecture. Unless otherwise noted, we adopt ST-GCN++ [15] as the GCN backbone. Particularly, we consider two variants: SkeleTR-S (with fewer GCN blocks and smaller output dimension, details in appendix), and SkeleTR-L (original ST-GCN++.). The inter-sequence modeling stage consists of a Mix Pooling module and two Transformer encoder blocks. The feature dimension in each encoder blocks equals to the output channels of the GCN backbone (128 for SkeleTR-S and 256 for SkeleTR-L)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.893, + 0.885 + ], + "angle": 0, + "content": "Training. We train all models end-to-end with SGD and batch size 128. For all models, we set the base learning rate (LR) to 0.1 and adopt a multi-step LR scheduler with the decay factor 0.1. For Poetics and NTU-Inter, we train models for 100 epochs, and set milestones at the \\(70_{th}\\) and \\(85_{th}\\) epoch. For AVA and Volleyball, we train the models for 40 epochs, and set milestones at the \\(30_{th}\\) and \\(35_{th}\\) epoch. We use cross entropy loss \\((\\mathbf{L}_{\\mathbf{CE}})\\) for Poetics, NTU-Inter and Volleyball, and use binary cross entropy loss \\((\\mathbf{L}_{\\mathbf{BCE}})\\) for AVA (multi-label). We explicitly set a loss weight \\(\\lambda = 100\\) in all AVA experiments due to the small magnitude of \\(\\mathbf{L}_{\\mathbf{BCE}}\\). In Poetics-AVA joint training, we concatenate samples in two datasets and optimize the model with a weighted combination of two losses: \\(\\mathbf{L} = \\mathbf{L}_{\\mathbf{CE}} + \\lambda \\mathbf{L}_{\\mathbf{BCE}}\\). For AVA, we sample short skeleton sequences in a small temporal window (5s / 10s) to form a training sample." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Inference. We follow standard practices [15] to predict ac" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.9 + ], + "angle": 0, + "content": "\\( {}^{1}T \\) is restricted by annotation granularity. For example,AVA [22] annotates atomic actions at 1fps, for which \\( T \\) should cover at most 2 seconds." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "13638" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.09, + 0.238, + 0.203 + ], + "angle": 0, + "content": "
MethodsPosetics Top-1Posetics Top-5
ST-TR [42]47.571.3
MS-G3D [36]52.675.8
CTR-GCN [7]54.176.7
ST-GCN++ [15]52.576.0
UNIK [65]52.575.7
PoseC3D [16]53.177.1
SkeleTR-S55.978.9
SkeleTR-L57.980.3
SkeleTR [C]57.680.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.083, + 0.205, + 0.235, + 0.215 + ], + "angle": 0, + "content": "(a) Action Classification: Poetics." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.09, + 0.48, + 0.203 + ], + "angle": 0, + "content": "
MethodsNTU60 XSubNTU60 XViewNTU120 XSubNTU120 XSet
ST-LSTM [33]83.087.363.066.6
LSTM-IRN [41]90.593.577.779.6
GeomNet [38]93.696.386.587.6
IGFormer [40]93.696.585.486.5
AS-GCN [29]87.192.077.879.3
CTR-GCN [7]91.694.383.284.4
SkeleTR-S94.697.187.488.0
SkeleTR-L94.797.587.888.3
SkeleTR [C]94.897.787.888.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.279, + 0.205, + 0.438, + 0.216 + ], + "angle": 0, + "content": "(b) Action Recognition: NTU-Inter." + }, + { + "type": "table", + "bbox": [ + 0.491, + 0.09, + 0.755, + 0.203 + ], + "angle": 0, + "content": "
MethodsOverall mAPPerson MovementObject ManipulationPerson Interaction
ST-TR [42]17.337.38.418.8
MS-G3D [36]18.438.79.619.7
CTR-GCN [7]18.939.49.820.5
ST-GCN++ [15]18.639.19.720.0
UNIK [65]19.039.79.920.4
PoseC3D [16]18.339.39.319.1
SkeleTR-S22.843.812.326.9
SkeleTR-S*26.848.217.727.8
SkeleTR [C]*27.350.917.328.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.205, + 0.726, + 0.216 + ], + "angle": 0, + "content": "(c) Spatio-temporal Action Detection: AVA v2.2." + }, + { + "type": "table", + "bbox": [ + 0.764, + 0.091, + 0.892, + 0.194 + ], + "angle": 0, + "content": "
MethodsVolleyball Top-1
AT+ [20]92.3
POGARS [52]93.2
COMPOSER [72]93.7
ST-TR [42]93.3
CTR-GCN [7]93.6
ST-GCN++ [15]93.3
SkeleTR-S94.4
SkeleTR [C]94.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.761, + 0.196, + 0.891, + 0.215 + ], + "angle": 0, + "content": "(d) Group Activity Recognition: Volleyball." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.219, + 0.893, + 0.274 + ], + "angle": 0, + "content": "Table 1: Comparing SkeleTR with previous state-of-the-art on versatile skeleton action recognition tasks. [C] denotes CTR-GCN backbone; * denotes joint training with Poetics. For Poetics and AVA, we follow the common practice [7, 65, 16] and compare the two-stream performance. For NTU-Inter and Volleyball, we only compare the joint modality performance as done in the prior works [52, 72]. In Table 1d, the method \\(\\mathrm{AT^{+}}\\) uses groundtruth bboxes of all video frames for skeleton extraction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.279, + 0.469, + 0.37 + ], + "angle": 0, + "content": "tion labels for trimmed videos. For AVA untrimmed videos, in inference we use a sliding window strategy to generate predictions, with a 1-second time step. We get an action prediction for a human proposal from each window containing the proposal, and then average all predictions associated with a human proposal to get the final result." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.38, + 0.216, + 0.394 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.469, + 0.478 + ], + "angle": 0, + "content": "We compare SkeleTR with state-of-the-art methods on three different skeleton action understanding tasks: action classification, spatio-temporal action detection, and group activity recognition. Unless otherwise noted, all methods use the same skeleton input as SkeleTR." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.469, + 0.75 + ], + "angle": 0, + "content": "Action Classification. Table 1a compares SkeleTR with state-of-the-art methods on Poetics. We use official training configurations when available [36, 16, 65]. For methods without official configurations [7, 15], we use the same hyper-parameter setting as SkeleTR. SkeleTR with all GCN backbones notably outperforms existing methods. Compared to CTR-GCN, SkeleTR-L achieves \\(3.6\\%\\) higher Top-1 accuracy. Quantitative analysis shows that the improvement is more significant for multi-person scenarios: actions with \\(\\geq 5\\) persons per frame have around \\(10\\%\\) improvement in mean class Top-1, which is twice as much as the improvement for other actions. SkeleTR also achieves the highest accuracy on Poetics (with OpenPose skeletons) and K400 (with HRNet/OpenPose skeletons). Table 1b shows that on the more specific task of two-person interaction recognition, SkeleTR also outperforms existing interaction recognition methods [41, 38, 40] and skeleton action classification methods by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.751, + 0.469, + 0.842 + ], + "angle": 0, + "content": "Spatio-temporal Action Detection. Skeleton-based spatio-temporal action detection is a new setting first proposed in this paper and we use AVA as the primary benchmark. For comparisons, we adapt state-of-the-art skeleton action classification algorithms by taking a single skeleton sequence (with 2 seconds length, centered at a person pro" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.277, + 0.892, + 0.391 + ], + "angle": 0, + "content": "
MethodsPosetics Top-1Params (M)GFLOPsInference Speed (2080Ti, clip/second))
ST-TR [42]47.456.4426.9297
MS-G3D [36]52.632.8220.8699
CTR-GCN [7]54.101.365.36195
ST-GCN++ [15]52.461.315.44235
UNIK [65]52.523.2812.15156
PoseC3D [16]53.091.9014.6953
SkeleTR-S55.850.812.74336
SkeleTR-L57.903.827.30200
" + }, + { + "type": "table_caption", + "bbox": [ + 0.538, + 0.393, + 0.854, + 0.406 + ], + "angle": 0, + "content": "Table 2: Efficiency statistics of models in Table 1a." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.892, + 0.623 + ], + "angle": 0, + "content": "posal in the key frame) as input to predict the action for this person. We train all models with the same hyper-parameter setting as SkeleTR. Due to the lack of global context and interaction modeling, all existing methods perform poorly on the new task, while SkeleTR significantly outperforms them. Compared to the baseline ST-GCN++, SkeleTR-S improves the overall mAP by \\(4.1\\%\\). Among three major action categories, SkeleTR-S makes the largest improvement on person interaction \\((7.0\\%)\\), thanks to its great inter-person modeling capability. When jointly trained with Poetics, SkeleTR-S enjoys another \\(4.1\\%\\) overall mAP improvement. However, this time the improvement is basically for person movement and object manipulation, since AVA-style person interactions rarely appear in Kinetics videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Group Activity Recognition. Most prior works on skeleton-based group activity recognition do not utilize the effective GCN architecture to model intra-person motion patterns. Instead, they directly model the relationship among skeletons with complicated transformer architectures. For a comprehensive comparison, we also adapt the state-of-the-art skeleton action recognition algorithms and benchmark them on the Volleyball dataset. Table 1d shows that, with good capability of both intra-sequence motion modeling and inter-sequence relationship modeling, SkeleTR outperforms all previous state-of-the-art on Volleyball, including the adapted GCN-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Computational Efficiency. SkeleTR is designed to be efficient and lightweight. Table 2 shows that SkeleTR-S outperforms previous state-of-the-art methods for skeleton-based action classification in terms of accuracy, while having the model with the smallest parameter size, GFLOPs, and fastest inference speed. In the appendix, we provide ad" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.85, + 0.469, + 0.901 + ], + "angle": 0, + "content": "2Joint and Bone. Each stream takes a separate modality derived from skeleton coordinates as input. For joint stream, the input is the coordinates itself. For bone stream it is the difference of neighboring joints. 3Please see the appendix for these results." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "13639" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.47, + 0.181 + ], + "angle": 0, + "content": "
PoseticsAVA v2.2
MethodsTop-1Top-5Overall mAPPerson MovementObject ManipulationPerson Interaction
RGB74.4291.2931.3048.4423.6732.74
ST-GCN++52.4675.9718.6239.109.6819.98
RGB + ST-GCN++75.9891.5831.9552.2823.1433.13
SkeleTR57.9080.3326.8348.1917.7227.75
RGB + SkeleTR76.7092.0234.3454.7825.7734.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.183, + 0.47, + 0.239 + ], + "angle": 0, + "content": "Table 3: SkeleTR demonstrates great complementarity when fused with RGB recognizers. We fuse predictions by state-of-the-art RGB recognizers (TimeSformer [2] for Posetics, ACAR-R101 [39] for AVA) with skeleton predictions with \\(2:1\\) ratio." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.242, + 0.449, + 0.326 + ], + "angle": 0, + "content": "
Interaction ModelingPosetics Top-1AVA Overall mAP
NA48.2117.56
Add Global51.0620.00
Concat Global50.0120.14
Non-Local48.4319.78
Transformer Encoders52.8120.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.328, + 0.468, + 0.356 + ], + "angle": 0, + "content": "Table 4: Transformer-based interaction modeling significantly outperforms multiple baselines." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.359, + 0.449, + 0.482 + ], + "angle": 0, + "content": "
Feature AggregationPosetics Top-1AVA Overall mAPHead Computation (GFLOPs)
1 × 152.8120.920.013
T × P52.7821.505.211
T × 153.4820.640.356
1 × P53.4021.350.080
1 × V52.1521.010.429
1 × 1 + T × 153.7521.000.392
1 × 1 + 1 × P53.0021.430.101
T × 1 + 1 × P53.1620.870.550
1 × 1 + T × 1 + 1 × P53.9021.650.594
" + }, + { + "type": "table_caption", + "bbox": [ + 0.082, + 0.486, + 0.464, + 0.501 + ], + "angle": 0, + "content": "Table 5: Comparing different feature aggregation strategies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.469, + 0.548 + ], + "angle": 0, + "content": "ditional evidence of SkeleTR's computational efficiency by comparing it to other methods designed for different tasks, such as IGFormer [40] and COMPOSER [72]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.47, + 0.701 + ], + "angle": 0, + "content": "Improved complementarity with RGB modality. To demonstrate the complementarity of SkeleTR upon RGB methods, we directly fuse the prediction of SkeleTR and the state-of-the-art RGB-based action recognizers. Table 3 shows that late fusion with SkeleTR leads to a significant boost (2.3% Posetics top-1, 3% AVA mAP) compared to the RGB state-of-the-art, which notably outperforms ST-GCN++ late fusion. On AVA, the improvement is extremely large for person movement (6.3% mAP), which emphasizes the importance of the skeleton in modeling human motions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.405, + 0.728 + ], + "angle": 0, + "content": "4.4. Ablation on Inter-Sequence Modeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.825 + ], + "angle": 0, + "content": "We conduct ablation experiments to validate the efficacy of our inter-sequence modeling stage. In experiments, we adopt the light ST-GCN++ as the backbone. We set \\(M\\) to 20, \\(T\\) to 30, and set the temporal window size to 5s for AVA. We report the recognition performance with a single joint modality on Posetics and AVA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.902 + ], + "angle": 0, + "content": "We first compare the Transformer-based modeling with other simple baselines for modeling global context or interaction. In all experiments, we pool both spatial and temporal dimensions and retain a single feature vector per sequence. We consider the following baselines: 1. NA: no" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.084, + 0.877, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.217, + 0.892, + 0.259 + ], + "angle": 0, + "content": "Figure 5: Visualization of the weights in different attention heads in the last transformer encoder. \\(\\mathbf{P}_{ij}\\) denotes the short skeleton sequence of person \\(j\\) at the timestamp \\(i\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.892, + 0.427 + ], + "angle": 0, + "content": "interaction modeling; 2. Add Global: process the average feature with an MLP to get a Global feature, then add it back to each sequence feature; 3. Concat Global: concatenate the average feature with each sequence feature, then process with an MLP; 4. Non-Local: process all sequence features with a standard non-local [59] block, and apply an MLP to the output. Table 4 shows that, despite its low computation cost (0.013 GFLOPs / clip), the Transformer markedly enhances recognition performance relatively to the NA baseline (4.6% Posetics top-1, 3.4% AVA mAP), and considerably outperforms all alternatives." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.893, + 0.653 + ], + "angle": 0, + "content": "Both Non-Local and Transformer encoders adopt selfattention to model the inter-sequence relationship, while there is a significant difference in recognition performance between them. The reason is that the multi-head design in Transformers allows for modeling diverse relationships between individual sequences. Visualizing the attention weights of different attention heads, we find that some of them have clearly interpretable semantics. Fig 5 shows the attention weights of a 'handshake' action. Of the eight attention heads, the first head aggregates temporal features within a single person, while the fourth head models interactions between individuals. Specifically, the attention weights of the fourth head reveal that \\( P_{21} \\) and \\( P_{22} \\) attend most strongly to other skeleton features, which corresponds to the exact moment of the handshaking interaction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.892, + 0.777 + ], + "angle": 0, + "content": "We then delve deeper into the Mix Pooling design, which helps to keep fine-grained sequence features and serves as a good trade-off between efficiency and performance. Table 5 compares Mix Pooling \\((1 \\times 1 + T \\times 1 + 1 \\times P)\\) with other feature aggregation strategies. Mix Pooling outperforms each individual pooling strategy or the combination of two strategies. It achieves the best performance on both datasets with a reasonable computation cost of 0.594 GFLOPs per clip." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.78, + 0.803 + ], + "angle": 0, + "content": "4.5. Ablation on Input Construction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.891, + 0.87 + ], + "angle": 0, + "content": "In this section we demonstrate the advantages of our new input format, both in recognition performance and computational efficiency. We report the recognition performance with a single joint modality." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Large Number of Short Sequences. We use a large number of short skeleton sequences as input for skeleton action" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "13640" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.085, + 0.445, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.225, + 0.47, + 0.253 + ], + "angle": 0, + "content": "Figure 6: Fix the total number of skeletons and adjust the sequence number. Posetics Top-1 accuracy is visualized." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.258, + 0.449, + 0.357 + ], + "angle": 0, + "content": "
MethodsMTPosetics Top-1Avg GFLOPs for \nPosetics Video Clips
SkeleTR-S203053.902.430
SkeleTR-L[5, 20]3055.064.934
SkeleTR-L203054.956.653
SkeleTR-S401554.172.789
SkeleTR-L[10, 40]1555.353.532
SkeleTR-L401555.557.394
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.361, + 0.47, + 0.403 + ], + "angle": 0, + "content": "Table 6: Sequence number \\( M \\) : fixed or adaptive? Utilizing an adaptive sequence number \\( M \\) can save great amounts of computational resources and achieve comparable performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.47, + 0.59 + ], + "angle": 0, + "content": "understanding tasks. Besides being a unified input format for different tasks, the new format also benefits standard skeleton-based action classification under real-world scenarios. To investigate the effect of sequence granularity, we fix the total number of skeletons \\( M \\times T \\), and adjust the sequence number \\( M \\). We compare three models: ST-GCN++, SkeleTR-S, and SkeleTR-L (Fig 6). Interestingly, we find that all models obtain their highest performance when using a large number of short sequences as input, specifically, \\( 40 \\times 15 \\). Meanwhile, the improvement is especially large for SkeleTR: \\( 2.1\\% \\) top-1 for SkeleTR-S, \\( 3\\% \\) top-1 for SkeleTR-L (compared to the \\( 2 \\times 300 \\) setting)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.471, + 0.82 + ], + "angle": 0, + "content": "Flexible Number of Sequences. SkeleTR is compatible with different numbers of sequences. To train and test SkeleTR with an adaptive squence number, we first set a dynamic range \\([M_{min}, M_{max}]\\), and for each training batch, we randomly sample \\(\\tilde{M}\\) from the range and use it as the sequence number for all videos in the batch. For each testing video, we use the actual sequence number \\(M'\\) if \\(M_{min} \\leq M' \\leq M_{max}\\), and cap it to \\(M_{min}\\) or \\(M_{max}\\) otherwise. We present an ablation study using two input formats: \\(20 \\times 30\\) and \\(40 \\times 15\\). For each format, we use a dynamic range ([5, 20], [10, 40]) of \\(M\\) to train SkeleTR-L on Poetics, and report the results in Table 6. Compared to using a fixed \\(M\\), using an adaptive \\(M\\) can achieve comparable recognition performance with a largely reduced inference cost (26% less GFLOPs for \\(T = 30\\), 52% less for \\(T = 15\\))." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.408, + 0.848 + ], + "angle": 0, + "content": "4.6. Transfer Learning and Joint Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "SkeleTR is a general framework for versatile skeleton action recognition tasks, which enables transfer learning and joint learning across different tasks. To demonstrate this," + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.088, + 0.872, + 0.19 + ], + "angle": 0, + "content": "
TrainingTemporal WindowOverall mAPPerson Movement mAPObject Manipulation mAPPerson Interaction mAP
From Scratch521.6542.3411.3725.66
Finetuning522.1243.0211.9125.76
Joint Training523.9244.2514.6526.06
From Scratch1021.3342.2411.2624.68
Finetuning1023.1144.0613.3225.85
Joint Training1025.6947.1316.5326.64
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.194, + 0.892, + 0.25 + ], + "angle": 0, + "content": "Table 7: Transfer Learning and Joint Learning with SkeleTR-S. For a target task with limited training data, an auxiliary dataset can be leveraged to improve the performance and mitigate the overfitting, either by transfer learning or joint learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.255, + 0.892, + 0.391 + ], + "angle": 0, + "content": "we consider two tasks: Poetics action classification and AVA spatio-temporal action detection (with AVA being the smaller dataset). From-scratch training on AVA is easy to overfit (Fig 4). Pretraining on Poetics or joint training on AVA and Poetics both help to mitigate the overfitting, with joint training leading to the best recognition performance. Enlarging the temporal window size for AVA sampling, which leads to less diversified training samples, amplifies the gain of pretraining or joint training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.405, + 0.627, + 0.42 + ], + "angle": 0, + "content": "4.7. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.612 + ], + "angle": 0, + "content": "SkeleTR is currently trained with skeletons pre-extracted from RGB videos, and the adopted pose estimation solution is relatively heavy. To use SkeleTR in real-world apps with limited computation resource, additional efforts are required to switch to a light-weight pose estimation solution that can perform fast inference on CPUs or edge devices, and to retrain the models accordingly. Besides, the short sequence format proposed by SkeleTR strikes a good balance between mitigating skeleton matching noises and modeling intra-sequence dynamics in real-world scenarios. The format may not lead to extra benefits when the ground-truth skeleton sequences are available (details in appendix)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.63, + 0.619, + 0.646 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.779 + ], + "angle": 0, + "content": "We presented SkeleTR, a unified framework for skeleton action recognition in general scenarios. It extends existing GCN backbones and achieves state-of-the-art performance on multiple skeleton action understanding tasks, including video-level, instance-level, and group-level action recognition. Besides, SkeleTR is compatible with transfer learning and joint learning across different datasets and tasks, which further improves the capability of learned representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Broader Impact. Using human skeletons for action recognition is more robust to environment changes and can help preserve user privacy. Real-world applications of skeleton action recognition include monitoring patient or elderly health, enabling autonomous driving, and facilitating human-computer interactions. However, there could be unintended usages and we advocate responsible usage complying with applicable laws and regulations." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.518, + 0.938 + ], + "angle": 0, + "content": "13641" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.468, + 0.2 + ], + "angle": 0, + "content": "[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.202, + 0.468, + 0.243 + ], + "angle": 0, + "content": "[3] Erik Bochinski, Volker Eiselein, and Thomas Sikora. High-speed tracking-by-detection without using image information. In AVSS, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.246, + 0.468, + 0.287 + ], + "angle": 0, + "content": "[4] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. TPAMI, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.289, + 0.468, + 0.33 + ], + "angle": 0, + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.332, + 0.468, + 0.372 + ], + "angle": 0, + "content": "[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.375, + 0.468, + 0.43 + ], + "angle": 0, + "content": "[7] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In ICCV, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.433, + 0.468, + 0.487 + ], + "angle": 0, + "content": "[8] Hyung-gun Chi, Myoung Hoon Ha, Seunggeun Chi, Sang Wan Lee, Qixing Huang, and Karthik Ramani. Infogcn: Representation learning for human skeleton-based action recognition. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.489, + 0.468, + 0.543 + ], + "angle": 0, + "content": "[9] Wongun Choi, Khuram Shahid, and Silvio Savarese. What are they doing?: Collective activity classification using spatio-temporal relationship among people. In ICCVW, 2009. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.468, + 0.587 + ], + "angle": 0, + "content": "[10] Vasileios Choutas, Philippe Weinzaepfel, Jérôme Revaud, and Cordelia Schmid. Potion: Pose motion representation for action recognition. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.468, + 0.644 + ], + "angle": 0, + "content": "[11] Liu Chunhui, Hu Yueyu, Li Yanghao, Song Sijie, and Liu Jiaying. Pku-mmd: A large scale benchmark for continuous multi-modal human action understanding. arXiv:1703.07475, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.468, + 0.715 + ], + "angle": 0, + "content": "[12] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv:2003.09003, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.468, + 0.771 + ], + "angle": 0, + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv:1810.04805, 2018.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.468, + 0.856 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv:2010.11929, 2020.3,4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[15] Haodong Duan, Jiaqi Wang, Kai Chen, and Dahua Lin. Pyskl: Towards good practices for skeleton action recognition. ACMM, 2022. 1, 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[16] Haodong Duan, Yue Zhao, Kai Chen, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. In CVPR, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[17] Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[18] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.265, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[20] Kirill Gavrilyuk, Ryan Sanford, Mehrsan Javan, and Cees GM Snoek. Actor-transformers for group activity recognition. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.308, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[21] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[22] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, 2018. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[23] Mohamed E Hussein, Marwan Torki, Mohammad A Gowayyed, and Motaz El-Saban. Human action recognition using a temporal hierarchy of covariance descriptors on 3d joint locations. In IJCAI, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[24] Mostafa S Ibrahim, Srikanth Muralidharan, Zhiwei Deng, Arash Vahdat, and Greg Mori. A hierarchical deep temporal model for group activity recognition. In CVPR, 2016. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[25] Haroon Idrees, Amir R Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The thumos challenge on action recognition for videos \"in the wild\". CVIU, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[26] Yanli Ji, Guo Ye, and Hong Cheng. Interactive body part contrast mining for human interaction recognition. In 2014 IEEE international conference on multimedia and expo workshops (ICMEW), pages 1-6. IEEE, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[27] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal-directed human activities. In CVPR, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[28] Ang Li, Meghanana Thotakuri, David A Ross, João Carreira, Alexander Vostrikov, and Andrew Zisserman. Theava-kinetics localized human actions video dataset. arXiv:2005.00214, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[29] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In CVPR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[30] Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang. Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In ICCV, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.52, + 0.938 + ], + "angle": 0, + "content": "13642" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[31] Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[32] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+d 120: A large-scale benchmark for 3d human activity understanding. TPAMI, 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[33] Jun Liu, Amir Shahroudy, Dong Xu, and Gang Wang. Spatio-temporal LSTM with trust gates for 3d human action recognition. In ECCV, 2016. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[35] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[36] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In CVPR, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.461 + ], + "angle": 0, + "content": "[37] Qianhui Men, Edmond SL Ho, Hubert PH Shum, and Howard Leung. A two-stream recurrent network for skeleton-based human interaction recognition. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 2771-2778. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.469, + 0.532 + ], + "angle": 0, + "content": "[38] Xuan Son Nguyen. Geomnet: A neural network based on riemannian geometries of spd matrix space and cholesky space for 3d skeleton-based interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13379-13389, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.469, + 0.586 + ], + "angle": 0, + "content": "[39] Junting Pan, Siyu Chen, Mike Zheng Shou, Yu Liu, Jing Shao, and Hongsheng Li. Actor-context-actor relation network for spatio-temporal action localization. In CVPR, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[40] Yunsheng Pang, Qiuhong Ke, Hossein Rahmani, James Bailey, and Jun Liu. Igformer: Interaction graph transformer for skeleton-based human interaction recognition. In ECCV, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[41] Mauricio Perez, Jun Liu, and Alex C Kot. Interaction relational network for mutual action recognition. IEEE Transactions on Multimedia, 24:366-376, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[42] Chiara Plizzari, Marco Cannici, and Matteo Matteucci. Skeleton-based action recognition via spatial and temporal transformer networks. CVIU, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[43] Michalis Raptis and Leonid Sigal. Poselet key-framing: A model for human activity recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2650-2657, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[44] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[45] Kohei Sendo and Norimichi Ukita. Heatmapping of people involved in group activities. In MVA, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[46] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+d: A large scale dataset for 3d human activity analysis. In CVPR, 2016. 1, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[47] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In CVPR, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[48] Bing Shuai, Alessandro Bergamo, Uta Buechler, Andrew Berneshawi, Alyssa Boden, and Joseph Tighe. Large scale real-world multi-person tracking. ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[49] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. NeurIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[50] Sebastian Stein and Stephen J McKenna. Combining embedded accelerometers with computer vision for recognizing food preparation activities. In UbiComp, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[51] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In CVPR, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[52] Haritha Thilakarathne, Aiden Nibali, Zhen He, and Stuart Morgan. Pose is all you need: The pose only group activity recognition system (pogars). Machine Vision and Applications, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[53] Arash Vahdat, Bo Gao, Mani Ranjbar, and Greg Mori. A discriminative key pose sequence model for recognizing human interactions. In 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops), pages 1729-1736. IEEE, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[55] Raviteja Vemulapalli, Felipe Arrate, and Rama Chellappa. Human action recognition by representing 3d skeletons as points in a lie group. In CVPR, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[56] Hongsong Wang and Liang Wang. Modeling temporal dynamics and spatial configurations of actions using two-stream recurrent neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 499-508, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[57] Jiang Wang, Zicheng Liu, Ying Wu, and Junsong Yuan. Mining actionlet ensemble for action recognition with depth cameras. In CVPR. IEEE, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[58] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[59] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, 2018. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[60] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[61] Fanyi Xiao, Yong Jae Lee, Kristen Grauman, Jitendra Malik, and Christoph Feichtenhofer. Audiovisual slowfast networks for video recognition. arXiv:2001.08740, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[62] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[63] Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Xia, Zhuowen Tu, and Stefano Soatto. Long short-term transformer for online action detection. NeurIPS, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "13643" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[64] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[65] Di Yang, Yaohui Wang, Antitza Dantcheva, Lorenzo Garattoni, Gianpiero Francesca, and François Bremond. Unik: A unified framework for real-world skeleton-based action recognition. ArXiv, 2021. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.274 + ], + "angle": 0, + "content": "[66] Kiwon Yun, Jean Honorio, Debaleena Chattopadhyay, Tamara L Berg, and Dimitris Samaras. Two-person interaction detection using body-pose features and multiple instance learning. In 2012 IEEE computer society conference on computer vision and pattern recognition workshops, pages 28–35. IEEE, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.468, + 0.359 + ], + "angle": 0, + "content": "[67] Yimeng Zhang, Xiaoming Liu, Ming-Ching Chang, Weina Ge, and Tsuhan Chen. Spatio-temporal phrases for activity recognition. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part III 12, pages 707-721. Springer, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.361, + 0.468, + 0.402 + ], + "angle": 0, + "content": "[68] Yuhan Zhang, Bo Wu, Wen Li, Lixin Duan, and Chuang Gan. Stst: Spatial-temporal specialized transformer for skeleton-based action recognition. In ACMMM, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.404, + 0.468, + 0.431 + ], + "angle": 0, + "content": "[69] Zhengyou Zhang. Microsoft Kinect sensor and its effect. IEEE multimedia, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.433, + 0.468, + 0.486 + ], + "angle": 0, + "content": "[70] Hang Zhao, Antonio Torralba, Lorenzo Torresani, and Zhicheng Yan. Hacs: Human action clips and segments dataset for recognition and temporal localization. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.489, + 0.468, + 0.543 + ], + "angle": 0, + "content": "[71] Jiaojiao Zhao, Yanyi Zhang, Xinyu Li, Hao Chen, Bing Shuai, Mingze Xu, Chunhui Liu, Kaustav Kundu, Yuanjun Xiong, Davide Modolo, et al. Tuber: Tubelet transformer for video action detection. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.546, + 0.468, + 0.6 + ], + "angle": 0, + "content": "[72] Honglu Zhou, Asim Kadav, Aviv Shamsian, Shijie Geng, Farley Lai, Long Zhao, Ting Liu, Mubbasir Kapadia, and Hans Peter Graf.Composer: Compositional learning of group activity in videos.ECCV, 2022.5,6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.602, + 0.468, + 0.657 + ], + "angle": 0, + "content": "[73] Wentao Zhu, Cuiling Lan, Junliang Xing, Wenjun Zeng, Yanghao Li, Li Shen, and Xiaohui Xie. Co-occurrence feature learning for skeleton based action recognition using regularized deep LSTM networks. In AAAI, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.66, + 0.468, + 0.7 + ], + "angle": 0, + "content": "[74] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv:2010.04159, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.7 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "13644" + } + ] +] \ No newline at end of file diff --git a/2023/SkeleTR_ Towards Skeleton-based Action Recognition in the Wild/27665a84-18aa-44bc-b25b-e6429d7734c6_origin.pdf b/2023/SkeleTR_ Towards Skeleton-based Action Recognition in the Wild/27665a84-18aa-44bc-b25b-e6429d7734c6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6215408ce263504e9036cc5e441733945b1d30e7 --- /dev/null +++ b/2023/SkeleTR_ Towards Skeleton-based Action Recognition in the Wild/27665a84-18aa-44bc-b25b-e6429d7734c6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35cf611b1782a631f2f160fbc7f51146f69e60dbd79e2fac0ef36fcb43a40934 +size 1523759 diff --git a/2023/SkeleTR_ Towards Skeleton-based Action Recognition in the Wild/full.md b/2023/SkeleTR_ Towards Skeleton-based Action Recognition in the Wild/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ba0060fc77442357836c3f178ddfe7d91224db4e --- /dev/null +++ b/2023/SkeleTR_ Towards Skeleton-based Action Recognition in the Wild/full.md @@ -0,0 +1,306 @@ +# SkeleTR: Towards Skeleton-based Action Recognition in the Wild + +Haodong Duan $^{1*}$ Mingze Xu $^{2}$ Bing Shuai $^{2}$ Davide Modolo $^{2}$ Zhuowen Tu $^{2}$ Joseph Tighe $^{2}$ Alessandro Bergamo $^{2}$ The Chinese University of Hong Kong AWS AI Labs + +dhd.efz@gmail.com, {xumingze,bshuai,dmodolo,ztu,tighej,bergamo}@amazon.com + +# Abstract + +We present SkeleTR, a new framework for skeleton-based action recognition. In contrast to prior work, which focuses mainly on controlled environments, we target more general scenarios that typically involve a variable number of people and various forms of interaction between people. SkeleTR works with a two-stage paradigm. It first models the intra-person skeleton dynamics for each skeleton sequence with graph convolutions, and then uses stacked Transformer encoders to capture person interactions that are important for action recognition in general scenarios. To mitigate the negative impact of inaccurate skeleton associations, SkeleTR takes relative short skeleton sequences as input and increases the number of sequences. As a unified solution, SkeleTR can be directly applied to multiple skeleton-based action tasks, including video-level action classification, instance-level action detection, and group-level activity recognition. It also enables transfer learning and joint training across different action tasks and datasets, which result in performance improvement. When evaluated on various skeleton-based action recognition benchmarks, SkeleTR achieves the state-of-the-art performance. + +# 1. Introduction + +Skeleton-based action recognition has received increasing attention in recent years due to its robustness against background and illumination changes [64, 16]. Existing methods [36, 7, 8, 15] mostly use Graph Convolutional Networks (GCNs) as backbones and focus on modeling spatiotemporal information from long skeleton sequences separately. While observing promising results in simplified and controlled environments [46, 32], they are less effective in more realistic scenarios with a variable number of people performing diverse actions. First, considering that linking skeletons (e.g., using skeletons' confidence scores [64]) in a long time span may inevitably generate more association errors [12, 48], requiring long sequences as input can bring + +![](images/abcaf131eac95a9937be84888aecc7bbdc0156d8545bb16e99b89b9c82682192.jpg) +Figure 1: SkeleTR is a general framework for skeleton action recognition which can handle different tasks, including video-level action classification, instance-level action detection, and group-level activity recognition. It adopts a two-stage paradigm: the intra-sequence and inter-sequence modeling, respectively. + +lots of noises. Second, prior works pay little attention to modeling the relations among a group of people, which plays an important role in recognizing both person interactions (e.g., talking to someone [22]) and group activities (e.g., pass/win-point in a volleyball game [24]). + +In this paper, we propose SkeleTR for skeleton action recognition in real-world scenarios and handle multiple action tasks (i.e., video-level, instance-level, and group-level action recognition) in a general framework. SkeleTR samples a large number of short skeleton sequences instead of a few long ones (practice of prior works [64, 16]) from each video. Each short sequence directly corresponds to an atomic action of a single person, thus the collection of multiple short sequences in space and time can represent hierarchical actions (with a longer time span) of a single person or a video-level action involving multiple people. The new input paradigm can be applied to instance-level and video-level recognition of actions with different temporal granularities, and it has several unique advantages. First, association errors are less likely to happen in a shorter skeleton sequence [12, 48]. We show that the improved sequence consistency leads to better intra-sequence modeling. Second, this new format is more flexible and helps to create comprehensive yet compact inputs. With a finer temporal + +granularity, the same input size budget can cover more human skeletons. It reduces redundant zero-paddings for a small number of people, and mitigates information loss during input sampling for a large number of people. + +SkeleTR adopts a two-stage paradigm as shown in Fig 1. Stage 1 performs intra-sequence modeling to capture the motion patterns of each individual skeleton sequence with a GCN backbone. The GCN backbone outputs a spatiotemporal feature for each sequence. Stage 2 concatenates all sequence features and feeds them into Transformer encoders to model inter-sequence relationships. At the end of stage 2, average pooling aggregates all features belonging to the same sequence into a single feature vector. By leveraging inter-sequence modeling, the output sequence feature incorporates the interactions between people and the global context of the video, which is helpful in both instance-level and video-level action recognition. To reduce the dimension of spatio-temporal skeleton features and keep the computations of Transformer tractable, we insert a novel Mix Pooling module before stage 2. It generates a compact representation with fine granularity for each sequence by applying multiple partial dimension reduction strategies in parallel. + +SkeleTR easily enables transfer learning and joint learning across different datasets and tasks. With a unified input format and model architecture, one can jointly train multiple tasks with different datasets, each with its own loss function and ground truth annotations. This allows for leveraging auxiliary datasets for pre-training or joint training of a target task with limited labeled data. In experiments, we find that joint training serves as a strong regularizer and improves recognition performance, especially in scenarios with limited training data. + +To evaluate the effectiveness of SkeleTR, we conduct comprehensive study on three skeleton-based action recognition tasks: action classification [6, 65, 46], spatiotemporal action detection [22], and group activity recognition [24]. Experimental results show that SkeleTR works well with different GCN backbones and significantly outperforms prior works on all three tasks. For example, with ST-GCN++ [15] backbone, SkeleTR outperforms the state-of-the-art methods by $3.8\%$ , $3.8\%$ , $0.7\%$ , and $1.1\%$ on Poetics [64], AVA [22], Volleyball [24], and NTU-Inter [40]. With Poetics joint training, the AVA performance is further improved by $4\%$ mAP. The great capability of SkeleTR leads to another noteworthy result: we first observe the strong complementarity of skeleton action recognition. Combining SkeleTR prediction and the RGB state-of-the-art [2, 39] with simple late fusion leads to great performance boost (2.3% Poetics Top-1, 3% AVA mAP). + +# 2. Related Works + +Human Action Understanding. Action classification [6, 21] is the most fundamental task in human action under + +standing, aiming at recognizing the video-level action in a short video clip. Another task of great interest is recognizing the instance/person-level actions in a video, either at the atomic [22, 28] or hierarchical [30] level. Meanwhile, other areas of interest include group activity recognition [24, 9], action localization [17, 25, 70], or action segmentation [27, 50]. Various modalities in videos can be utilized to solve the action understanding tasks, such as RGB appearance [58, 6, 19, 59], optical flow [49], audio waves [61], and human skeletons [64, 16]. Spatio-temporal representations learned end-to-end from RGB appearance yield good performance on standard benchmarks. However, they are data intensive and less robust to context changes. In this work, we propose a unified framework to learn video representations based on human skeletons, that can be easily applied to multiple action tasks. + +Skeleton-based Action Recognition. Human skeletons are robust against background and illumination changes [64, 47], and can be easily obtained with sensors [69] or pose estimators [4, 51]. In skeleton action recognition, the majority of research has focused on modeling the intra-person skeleton dynamics. With the rise of deep learning, manual feature engineering [23, 55, 57] has gradually been replaced by end-to-end representation learning with deep neural architectures (GCN [64], 2D-CNN [10], or 3D-CNN [16]). Among them, GCN-based approaches [64, 47, 36, 7] have been the most popular paradigm, which construct spatial-temporal graphs of human keypoint coordinates and perform spatial and temporal modeling. These methods model the intra-person skeleton dynamics for in-lab collected videos [46, 11], featuring small number of persons (1 or 2) per video and accurate person identities. For more challenging scenarios [6], the model performance decreases due to the noises produced in matching skeletons to sequences. Besides, GCN approaches primarily focus on predicting video-level actions, and do not benefit from context or interaction modeling when applied to instance-level action recognition. To this end, we propose SkeleTR, which expands existing GCN approaches and works universally well for diversified scenarios and recognition tasks. + +Human Interaction Recognition. Human interaction recognition [43, 53, 67] is a sub-task of action recognition. It aims at identifying and categorizing activities between people. Early works [66, 26] built hand-crafted features for interaction modeling and adopted classic machine learning algorithms for activity classification. Subsequent works proposed to use two-stream RNNs to model the two-person interactions [41, 37, 56]. Recently, IGFormer [40] first proposed to use a two-stream Transformer for two-person interaction modeling and observed improved performance. However, most deep methods share the same limitations: 1. they are largely restricted to two-person scenarios and tend to suffer from exponentially growing inference + +![](images/c04aa9b76c8349ec0693112c454311d4fc298ff9cfb52ba2538b912c98b3ad5e.jpg) +Figure 2: The architecture of SkeLETR. SkeLETR takes a large set of short skeleton sequences as input and performs a two-stage modeling process. It first adopts a GCN backbone to model the intra-sequence skeleton dynamics. Then it processes the sequence features with stacked Transformer encoders, which helps to model person interactions and leverage the long-range context. Between the two stages, we adopt a Mix Pooling layer to compress the spatio-temporal features and keep computations tractable. + +costs with the person number; 2. they fail to leverage the strong intra-sequence modeling capabilities of GCNs. To overcome these challenges, we introduce SkeleTR, which integrates intra-sequence and inter-sequence modeling with good scalability and small computation cost. + +Visual Transformer. With the great success in NLP [54, 13], Transformers have been increasingly applied to diversified vision tasks [14, 34, 5, 74] as well as the RGB-based action understanding [35, 18, 1, 2, 63, 71], and have achieved promising results. For skeleton action recognition, there exist works [42, 68] that directly use Transformers for intra-sequence spatio-temporal modeling. However, without exploiting the locality, the Transformer-based motion modeling only leads to inferior recognition performance compared to GCN-based ones. In SkeleTR, to enjoy the best of both worlds, we adopt GCNs for intra-sequence motion modeling, and use Transformer encoders for intersequence relationship modeling among multiple persons' fine-grained skeleton features. + +# 3. SkeleTR + +# 3.1. Overview + +Both SkeleTR and prior GCN approaches directly take human joint coordinates as input. A person in a specific frame corresponds to a skeleton with $V$ joints (each has $C$ dimensions, $C$ is 2 or 3). Skeletons in $T$ different frames (1 skeleton per frame) can be linked in temporal order to form a skeleton sequence $\mathbf{X}_i\in \mathbb{R}^{T\times V\times C}$ with length $T$ . A skeleton action recognizer takes $M$ skeleton sequences as input, and outputs a score vector $\mathbf{S}\in \mathbb{R}^K$ for video-level action recognition, where $K$ is the number of action classes. + +Spatial-temporal GCNs are widely used to model skeleton sequences for action classification, which focus on modeling intra-sequence skeleton dynamics. Most GCN methods [64, 47, 7] share a similar high-level design (Fig 2(a)), where a backbone with $\mathrm{N}_1$ GCN blocks is used to extract features of each skeleton sequence. A GCN block com + +prises a GCN layer and a TCN layer. The GCN layer performs spatial modeling by fusing joint features within the same skeleton. The TCN layer employs 1D Conv to capture the motion patterns of each joint. GCNs model long skeleton sequences that span the entire video in parallel, and aggregate multiple sequence features with average pooling to predict the video-level action. On existing skeleton action recognition benchmarks, GCNs exhibit superior intrasequence modeling capability than other architectures (e.g., RNN [73, 33] or Transformers [42, 68]). + +SkeleTR inherits the effective intra-sequence modeling of GCNs, while also enhancing the flexibility, generality, and inter-person modeling capability. SkeleTR takes $M$ short skeleton sequences as input (Sec. 3.2), and outputs one feature vector per sequence. It first adopts a GCN backbone to model the motion patterns of each skeleton sequence, then uses stacked Transformer encoders to capture the interactions and leverage long-range context (Sec. 3.3). Between two stages, we use a Mix Pooling layer to compress spatiotemporal sequence features into a compact multi-view representation to keep computation tractable. Fig 2 compares the design of SkeleTR and previous GCN methods. + +# 3.2. The Input Construction of SkeleTR + +For skeleton action classification in real-world scenarios, prior works [64, 47, 36] use a few long skeleton sequences as input, where skeletons at different timestamps are linked with rudimentary association methods. They use a naive confidence score based matching to link skeletons to sequences as long as the entire video (as shown in Fig 3). Specifically, they mostly keep $\hat{M} = 2$ skeletons with the highest confidence scores per frame, and link skeletons based on confidence rank. In cases where it is applied to challenging videos involving multiple persons, this strategy can lead to noisy skeleton sequences and sub-optimal performance. + +In contrast, we propose to sample a large number of short skeleton sequences as input, and link skeletons at differ + +![](images/f8a33ba00e35f6bb53ce07affe21e3b7785cdd0c5d82e6a3df0dfdc252ce0fe9.jpg) +Figure 3: The difference of input formats. Prior works link skeletons at different timestamps based on the confidence score rank to form $\hat{M} = 2$ long sequences, each covers the entire video. SkeleTR, instead, directly takes multiple shorter sequences as input. Each sequence only covers $1\sim 2$ seconds, thus has better consistency than long sequences. + +ent timestamps into sequences based on their IOU similarity. Specifically, to sample sequences with length $T$ , we first uniformly select a set of key frames with a stride of $T / 2$ . For each skeleton in every key frame, we create a sequence by linking skeletons within a window of $T$ frames using IoU-based matching [3], resulting in a total of $M'$ sequences ( $M'$ is variable for different videos). Our new input format is compatible with a variable number of input sequences. By using different numbers of sequences ( $M$ ) for each training batch, SkeleTR can handle inputs with different $M$ , which helps to reduce the usages of zero-padding at test time and improves efficiency. + +# 3.3. Inter-sequence Modeling Stage + +After processing short skeleton sequences with the GCN backbone, SkeleTR employs stacked Transformer encoders to model inter-sequence relationships. The process starts with reducing the feature dimension using a Mix Pooling layer, resulting in $L$ feature vectors for each sequence. Next, a sequence-level spatial-temporal embedding is combined with the feature vectors. The resulting sequence features $(M\times L$ vectors in total) are concatenated and modeled using a standard Transformer with multiple encoder blocks [14]. This architecture is depicted in Fig 2(b). We present more details in the next paragraphs. + +Mix Pooling. The cost of computing the attention matrices in a Transformer block grows quadratically with the input dimensionality, resulting in a computational cost of $\Theta((M \times T \times V)^2)$ in our case, if no dimension reduction is performed ( $L = T \times V$ ). Given a video with a typical configuration ( $T = 15$ , $V = 17$ ) and a Transformer consisting of only $\mathrm{N}_2 = 2$ encoder blocks, the Transformer itself consumes 52.7 GFLOPs, which is impractical to compute. Another extreme is to remove the dimension $T$ and $V$ with global average pooling, which reduces the computational complexity to $\Theta(M^2)$ but also results in loss of fine-grained + +features for each skeleton sequence. + +Between two extremes $(L = T\times V$ or $L = 1\times 1)$ we adopt multi-stream partial dimension reductions to reduce the computation cost. We denote a partial dimension reduction stream as $A\times B$ , indicating reducing temporal dimension $(T)$ to $A$ and spatial dimension $(V)$ to $B$ . Each stream provides a specific view of the spatio-temporal skeleton features. Multiple reduction streams can be combined in Mix Pooling to compress raw features into a comprehensive representation. Specifically, we adopt 3 partial dimension reduction streams: $1\times 1$ , $T\times 1$ , and $1\times P$ ( $V$ joints aggregated to $P = 5$ parts: head, two arms, two legs) (Fig 2(c)). Mix Pooling concatenates the output of 3 streams and keeps $L = 1 + T + P$ features per sequence. The heterogeneous features are then processed with one single Transformer. Mix Pooling enables the cross-view feature interaction and helps keep computations tractable, yielding a complexity of $\Theta(M^2(1 + T + P)^2)$ ( $\sim 0.6$ GFLOPs / video). + +Sequence-level Spatial Temporal Embedding. We further combine $L$ features per sequence with a sequence-level spatial-temporal embedding. We utilize a fully-connected layer to encode both the bounding box and the normalized timestamp (computed based on the entire video) of the center skeleton for each sequence, which then serves as the embedding. The embeddings are added with the Mix-Pooled features, and further processed by a layer normalization. + +Transformer-based Interaction Modeling. We use a standard Transformer with $\mathrm{N}_2$ encoder blocks for modeling sequence features. Each block consists of a multi-head self-attention layer (MHSA) and a feed-forward network (FFN), and adopts the post-layernorm style. The multi-head design in the encoder blocks helps to capture various kinds of relationships among skeleton features, including but not limited to the inter-person interactions and intra-person long-range context. At last, we use intra-sequence average pooling to aggregate $L = (1 + T + P)$ features per sequence into a single one. The output sequence features, which incorporate both person interactions and a comprehensive context, are then used to predict the instance-level or video-level action. We present the detailed configuration of encoder blocks in the appendix. + +# 3.4. SkeleTR as a General Framework + +As a general framework for skeleton action recognition, SkeleTR can be applied to various skeleton-based action recognition tasks. It also enables and benefits from transfer learning and joint learning across different tasks. + +Video-level Action Classification. The tasks of action classification [6, 65] and group activity recognition [24, 9] both take as input a video and predict the action at video-level. Note that the input videos may contain a variable number of people. In SkeleTR, we sample numerous short skeleton sequences that last $1 \sim 2$ seconds. We then average the $M$ + +![](images/e18071554911c9145d282bce8b7720c319570936826c1ff56d1f7d6887400996.jpg) +Figure 4: The curve of training / validation loss and validation mAP on AVA. Joint training with Posetics [65] significantly alleviates overfitting for AVA action detection. The jointly-trained model achieves much higher validation mAP, with the gap of training / validation losses largely shrinks. + +output features, and use a fully-connected layer to predict the video action class. Thanks to the local motion modeling with GCN and interaction modeling with Transformers, SkeleTR achieves strong performance in both action classification and group activity recognition. + +Instance-level Action Recognition. Spatio-temporal action detection [22] aims to recognize actions of each person at specific timestamps (instance-level recognition). For this task, we sample short skeleton sequences with length $T^{1}$ in a longer temporal window $(5 \sim 10s)$ as input, and utilize the sequence-level features for instance-level action recognition. SkeleTR enables the modeling of both intra-sequence motion patterns and inter-sequence interactions for skeleton-based spatio-temporal action detection. + +Transfer Learning and Joint Learning. SkeLETR enables to: 1. transfer the representation learned on one task to another new task; 2. jointly train different datasets and tasks, which in turn alleviate the overfitting issues on small datasets. Fig 4 shows the curves of training and validation losses, and validation mAP for AVA. We see that joint training serves as a strong regularizer and alleviates the severe overfitting. With joint training, the gap between training and validation losses largely shrinks, and the recognition performance on the validation set becomes much stronger. + +# 4. Experiments + +# 4.1. Datasets + +Kinetics-400 (K400) [6] contains videos with variable number of people. Some K400 actions cannot be distinguished by only using skeletons, thus we use a subset of K400, named **Posetics** [65], as the main benchmark for action classification. Posetics merges together classes featuring identical skeletons, resulting in 320 total classes (e.g., 'eating burger' and 'eating hotdog' are merged into 'eating something'). For K400 and Posetics, we use 2D human skeletons provided by [16] (extracted with Faster + +RCNN [44] (R50) and HRNet [51]). We provide K400 results in the appendix. + +NTU-Inter is a commonly used benchmark for two-person interaction recognition [40, 41]. It's a subset of the inlab collected NTURGB+D [32, 46], containing only mutual action classes (with two subjects). Consistent with NTURGB+D, NTU-Inter includes 4 data splits: NTU60-XSub, NTU60-XView, NTU120-XSub, NTU120-XSet. We use the official 3D skeletons for NTU-Inter experiments. + +AVA 2.2 [22] is a spatio-temporal action detection benchmark. It's a multi-label dataset with 80 action classes, falling into three major categories: person movement, object manipulation, and person interaction. Following the common practice [60], we use Faster-RCNN (ResNeXt101-FPN [31, 62]) for person detection and HRNet for pose estimation. The detector is pretrained on COCO and finetuned on AVA training bboxes. Following the benchmark guidelines, we evaluate 60 action classes. We report the overall mAP and the mAP of each major category. + +Volleyball [24] is a group activity recognition dataset with 8 volleyball group activities. Each clip has up to 13 persons, with the center frame annotated with GT person boxes. Following common practice [72, 52], we use the tracklets from [45] and HRNet for pose extraction, train models on training and validation sets, and report the test set results. + +# 4.2. Implementation Details + +Architecture. Unless otherwise noted, we adopt ST-GCN++ [15] as the GCN backbone. Particularly, we consider two variants: SkeleTR-S (with fewer GCN blocks and smaller output dimension, details in appendix), and SkeleTR-L (original ST-GCN++.). The inter-sequence modeling stage consists of a Mix Pooling module and two Transformer encoder blocks. The feature dimension in each encoder blocks equals to the output channels of the GCN backbone (128 for SkeleTR-S and 256 for SkeleTR-L). + +Training. We train all models end-to-end with SGD and batch size 128. For all models, we set the base learning rate (LR) to 0.1 and adopt a multi-step LR scheduler with the decay factor 0.1. For Poetics and NTU-Inter, we train models for 100 epochs, and set milestones at the $70_{th}$ and $85_{th}$ epoch. For AVA and Volleyball, we train the models for 40 epochs, and set milestones at the $30_{th}$ and $35_{th}$ epoch. We use cross entropy loss $(\mathbf{L}_{\mathbf{CE}})$ for Poetics, NTU-Inter and Volleyball, and use binary cross entropy loss $(\mathbf{L}_{\mathbf{BCE}})$ for AVA (multi-label). We explicitly set a loss weight $\lambda = 100$ in all AVA experiments due to the small magnitude of $\mathbf{L}_{\mathbf{BCE}}$ . In Poetics-AVA joint training, we concatenate samples in two datasets and optimize the model with a weighted combination of two losses: $\mathbf{L} = \mathbf{L}_{\mathbf{CE}} + \lambda \mathbf{L}_{\mathbf{BCE}}$ . For AVA, we sample short skeleton sequences in a small temporal window (5s / 10s) to form a training sample. + +Inference. We follow standard practices [15] to predict ac + +
MethodsPosetics Top-1Posetics Top-5
ST-TR [42]47.571.3
MS-G3D [36]52.675.8
CTR-GCN [7]54.176.7
ST-GCN++ [15]52.576.0
UNIK [65]52.575.7
PoseC3D [16]53.177.1
SkeleTR-S55.978.9
SkeleTR-L57.980.3
SkeleTR [C]57.680.1
+ +(a) Action Classification: Poetics. + +
MethodsNTU60 XSubNTU60 XViewNTU120 XSubNTU120 XSet
ST-LSTM [33]83.087.363.066.6
LSTM-IRN [41]90.593.577.779.6
GeomNet [38]93.696.386.587.6
IGFormer [40]93.696.585.486.5
AS-GCN [29]87.192.077.879.3
CTR-GCN [7]91.694.383.284.4
SkeleTR-S94.697.187.488.0
SkeleTR-L94.797.587.888.3
SkeleTR [C]94.897.787.888.3
+ +(b) Action Recognition: NTU-Inter. + +
MethodsOverall mAPPerson MovementObject ManipulationPerson Interaction
ST-TR [42]17.337.38.418.8
MS-G3D [36]18.438.79.619.7
CTR-GCN [7]18.939.49.820.5
ST-GCN++ [15]18.639.19.720.0
UNIK [65]19.039.79.920.4
PoseC3D [16]18.339.39.319.1
SkeleTR-S22.843.812.326.9
SkeleTR-S*26.848.217.727.8
SkeleTR [C]*27.350.917.328.3
+ +(c) Spatio-temporal Action Detection: AVA v2.2. + +
MethodsVolleyball Top-1
AT+ [20]92.3
POGARS [52]93.2
COMPOSER [72]93.7
ST-TR [42]93.3
CTR-GCN [7]93.6
ST-GCN++ [15]93.3
SkeleTR-S94.4
SkeleTR [C]94.4
+ +tion labels for trimmed videos. For AVA untrimmed videos, in inference we use a sliding window strategy to generate predictions, with a 1-second time step. We get an action prediction for a human proposal from each window containing the proposal, and then average all predictions associated with a human proposal to get the final result. + +# 4.3. Main Results + +We compare SkeleTR with state-of-the-art methods on three different skeleton action understanding tasks: action classification, spatio-temporal action detection, and group activity recognition. Unless otherwise noted, all methods use the same skeleton input as SkeleTR. + +Action Classification. Table 1a compares SkeleTR with state-of-the-art methods on Poetics. We use official training configurations when available [36, 16, 65]. For methods without official configurations [7, 15], we use the same hyper-parameter setting as SkeleTR. SkeleTR with all GCN backbones notably outperforms existing methods. Compared to CTR-GCN, SkeleTR-L achieves $3.6\%$ higher Top-1 accuracy. Quantitative analysis shows that the improvement is more significant for multi-person scenarios: actions with $\geq 5$ persons per frame have around $10\%$ improvement in mean class Top-1, which is twice as much as the improvement for other actions. SkeleTR also achieves the highest accuracy on Poetics (with OpenPose skeletons) and K400 (with HRNet/OpenPose skeletons). Table 1b shows that on the more specific task of two-person interaction recognition, SkeleTR also outperforms existing interaction recognition methods [41, 38, 40] and skeleton action classification methods by a large margin. + +Spatio-temporal Action Detection. Skeleton-based spatio-temporal action detection is a new setting first proposed in this paper and we use AVA as the primary benchmark. For comparisons, we adapt state-of-the-art skeleton action classification algorithms by taking a single skeleton sequence (with 2 seconds length, centered at a person pro + +(d) Group Activity Recognition: Volleyball. +Table 1: Comparing SkeleTR with previous state-of-the-art on versatile skeleton action recognition tasks. [C] denotes CTR-GCN backbone; * denotes joint training with Poetics. For Poetics and AVA, we follow the common practice [7, 65, 16] and compare the two-stream performance. For NTU-Inter and Volleyball, we only compare the joint modality performance as done in the prior works [52, 72]. In Table 1d, the method $\mathrm{AT^{+}}$ uses groundtruth bboxes of all video frames for skeleton extraction. + +
MethodsPosetics Top-1Params (M)GFLOPsInference Speed (2080Ti, clip/second))
ST-TR [42]47.456.4426.9297
MS-G3D [36]52.632.8220.8699
CTR-GCN [7]54.101.365.36195
ST-GCN++ [15]52.461.315.44235
UNIK [65]52.523.2812.15156
PoseC3D [16]53.091.9014.6953
SkeleTR-S55.850.812.74336
SkeleTR-L57.903.827.30200
+ +Table 2: Efficiency statistics of models in Table 1a. + +posal in the key frame) as input to predict the action for this person. We train all models with the same hyper-parameter setting as SkeleTR. Due to the lack of global context and interaction modeling, all existing methods perform poorly on the new task, while SkeleTR significantly outperforms them. Compared to the baseline ST-GCN++, SkeleTR-S improves the overall mAP by $4.1\%$ . Among three major action categories, SkeleTR-S makes the largest improvement on person interaction $(7.0\%)$ , thanks to its great inter-person modeling capability. When jointly trained with Poetics, SkeleTR-S enjoys another $4.1\%$ overall mAP improvement. However, this time the improvement is basically for person movement and object manipulation, since AVA-style person interactions rarely appear in Kinetics videos. + +Group Activity Recognition. Most prior works on skeleton-based group activity recognition do not utilize the effective GCN architecture to model intra-person motion patterns. Instead, they directly model the relationship among skeletons with complicated transformer architectures. For a comprehensive comparison, we also adapt the state-of-the-art skeleton action recognition algorithms and benchmark them on the Volleyball dataset. Table 1d shows that, with good capability of both intra-sequence motion modeling and inter-sequence relationship modeling, SkeleTR outperforms all previous state-of-the-art on Volleyball, including the adapted GCN-based approaches. + +Computational Efficiency. SkeleTR is designed to be efficient and lightweight. Table 2 shows that SkeleTR-S outperforms previous state-of-the-art methods for skeleton-based action classification in terms of accuracy, while having the model with the smallest parameter size, GFLOPs, and fastest inference speed. In the appendix, we provide ad + +
PoseticsAVA v2.2
MethodsTop-1Top-5Overall mAPPerson MovementObject ManipulationPerson Interaction
RGB74.4291.2931.3048.4423.6732.74
ST-GCN++52.4675.9718.6239.109.6819.98
RGB + ST-GCN++75.9891.5831.9552.2823.1433.13
SkeleTR57.9080.3326.8348.1917.7227.75
RGB + SkeleTR76.7092.0234.3454.7825.7734.93
+ +Table 3: SkeleTR demonstrates great complementarity when fused with RGB recognizers. We fuse predictions by state-of-the-art RGB recognizers (TimeSformer [2] for Posetics, ACAR-R101 [39] for AVA) with skeleton predictions with $2:1$ ratio. + +
Interaction ModelingPosetics Top-1AVA Overall mAP
NA48.2117.56
Add Global51.0620.00
Concat Global50.0120.14
Non-Local48.4319.78
Transformer Encoders52.8120.92
+ +Table 4: Transformer-based interaction modeling significantly outperforms multiple baselines. + +
Feature AggregationPosetics Top-1AVA Overall mAPHead Computation (GFLOPs)
1 × 152.8120.920.013
T × P52.7821.505.211
T × 153.4820.640.356
1 × P53.4021.350.080
1 × V52.1521.010.429
1 × 1 + T × 153.7521.000.392
1 × 1 + 1 × P53.0021.430.101
T × 1 + 1 × P53.1620.870.550
1 × 1 + T × 1 + 1 × P53.9021.650.594
+ +Table 5: Comparing different feature aggregation strategies. + +ditional evidence of SkeleTR's computational efficiency by comparing it to other methods designed for different tasks, such as IGFormer [40] and COMPOSER [72]. + +Improved complementarity with RGB modality. To demonstrate the complementarity of SkeleTR upon RGB methods, we directly fuse the prediction of SkeleTR and the state-of-the-art RGB-based action recognizers. Table 3 shows that late fusion with SkeleTR leads to a significant boost (2.3% Posetics top-1, 3% AVA mAP) compared to the RGB state-of-the-art, which notably outperforms ST-GCN++ late fusion. On AVA, the improvement is extremely large for person movement (6.3% mAP), which emphasizes the importance of the skeleton in modeling human motions. + +# 4.4. Ablation on Inter-Sequence Modeling + +We conduct ablation experiments to validate the efficacy of our inter-sequence modeling stage. In experiments, we adopt the light ST-GCN++ as the backbone. We set $M$ to 20, $T$ to 30, and set the temporal window size to 5s for AVA. We report the recognition performance with a single joint modality on Posetics and AVA. + +We first compare the Transformer-based modeling with other simple baselines for modeling global context or interaction. In all experiments, we pool both spatial and temporal dimensions and retain a single feature vector per sequence. We consider the following baselines: 1. NA: no + +![](images/f3a2e78b803c0ced29c4e133a3cbd36555db4d785728b9355fda555da41689ee.jpg) +Figure 5: Visualization of the weights in different attention heads in the last transformer encoder. $\mathbf{P}_{ij}$ denotes the short skeleton sequence of person $j$ at the timestamp $i$ . + +interaction modeling; 2. Add Global: process the average feature with an MLP to get a Global feature, then add it back to each sequence feature; 3. Concat Global: concatenate the average feature with each sequence feature, then process with an MLP; 4. Non-Local: process all sequence features with a standard non-local [59] block, and apply an MLP to the output. Table 4 shows that, despite its low computation cost (0.013 GFLOPs / clip), the Transformer markedly enhances recognition performance relatively to the NA baseline (4.6% Posetics top-1, 3.4% AVA mAP), and considerably outperforms all alternatives. + +Both Non-Local and Transformer encoders adopt selfattention to model the inter-sequence relationship, while there is a significant difference in recognition performance between them. The reason is that the multi-head design in Transformers allows for modeling diverse relationships between individual sequences. Visualizing the attention weights of different attention heads, we find that some of them have clearly interpretable semantics. Fig 5 shows the attention weights of a 'handshake' action. Of the eight attention heads, the first head aggregates temporal features within a single person, while the fourth head models interactions between individuals. Specifically, the attention weights of the fourth head reveal that $P_{21}$ and $P_{22}$ attend most strongly to other skeleton features, which corresponds to the exact moment of the handshaking interaction. + +We then delve deeper into the Mix Pooling design, which helps to keep fine-grained sequence features and serves as a good trade-off between efficiency and performance. Table 5 compares Mix Pooling $(1 \times 1 + T \times 1 + 1 \times P)$ with other feature aggregation strategies. Mix Pooling outperforms each individual pooling strategy or the combination of two strategies. It achieves the best performance on both datasets with a reasonable computation cost of 0.594 GFLOPs per clip. + +# 4.5. Ablation on Input Construction + +In this section we demonstrate the advantages of our new input format, both in recognition performance and computational efficiency. We report the recognition performance with a single joint modality. + +Large Number of Short Sequences. We use a large number of short skeleton sequences as input for skeleton action + +![](images/71cca6e19234087cb9f46a241a258b42637d1c1c0b68cf026b54710c66e060d5.jpg) +Figure 6: Fix the total number of skeletons and adjust the sequence number. Posetics Top-1 accuracy is visualized. + +
MethodsMTPosetics Top-1Avg GFLOPs for +Posetics Video Clips
SkeleTR-S203053.902.430
SkeleTR-L[5, 20]3055.064.934
SkeleTR-L203054.956.653
SkeleTR-S401554.172.789
SkeleTR-L[10, 40]1555.353.532
SkeleTR-L401555.557.394
+ +understanding tasks. Besides being a unified input format for different tasks, the new format also benefits standard skeleton-based action classification under real-world scenarios. To investigate the effect of sequence granularity, we fix the total number of skeletons $M \times T$ , and adjust the sequence number $M$ . We compare three models: ST-GCN++, SkeleTR-S, and SkeleTR-L (Fig 6). Interestingly, we find that all models obtain their highest performance when using a large number of short sequences as input, specifically, $40 \times 15$ . Meanwhile, the improvement is especially large for SkeleTR: $2.1\%$ top-1 for SkeleTR-S, $3\%$ top-1 for SkeleTR-L (compared to the $2 \times 300$ setting). + +Flexible Number of Sequences. SkeleTR is compatible with different numbers of sequences. To train and test SkeleTR with an adaptive squence number, we first set a dynamic range $[M_{min}, M_{max}]$ , and for each training batch, we randomly sample $\tilde{M}$ from the range and use it as the sequence number for all videos in the batch. For each testing video, we use the actual sequence number $M'$ if $M_{min} \leq M' \leq M_{max}$ , and cap it to $M_{min}$ or $M_{max}$ otherwise. We present an ablation study using two input formats: $20 \times 30$ and $40 \times 15$ . For each format, we use a dynamic range ([5, 20], [10, 40]) of $M$ to train SkeleTR-L on Poetics, and report the results in Table 6. Compared to using a fixed $M$ , using an adaptive $M$ can achieve comparable recognition performance with a largely reduced inference cost (26% less GFLOPs for $T = 30$ , 52% less for $T = 15$ ). + +# 4.6. Transfer Learning and Joint Learning + +SkeleTR is a general framework for versatile skeleton action recognition tasks, which enables transfer learning and joint learning across different tasks. To demonstrate this, + +Table 6: Sequence number $M$ : fixed or adaptive? Utilizing an adaptive sequence number $M$ can save great amounts of computational resources and achieve comparable performance. + +
TrainingTemporal WindowOverall mAPPerson Movement mAPObject Manipulation mAPPerson Interaction mAP
From Scratch521.6542.3411.3725.66
Finetuning522.1243.0211.9125.76
Joint Training523.9244.2514.6526.06
From Scratch1021.3342.2411.2624.68
Finetuning1023.1144.0613.3225.85
Joint Training1025.6947.1316.5326.64
+ +Table 7: Transfer Learning and Joint Learning with SkeleTR-S. For a target task with limited training data, an auxiliary dataset can be leveraged to improve the performance and mitigate the overfitting, either by transfer learning or joint learning. + +we consider two tasks: Poetics action classification and AVA spatio-temporal action detection (with AVA being the smaller dataset). From-scratch training on AVA is easy to overfit (Fig 4). Pretraining on Poetics or joint training on AVA and Poetics both help to mitigate the overfitting, with joint training leading to the best recognition performance. Enlarging the temporal window size for AVA sampling, which leads to less diversified training samples, amplifies the gain of pretraining or joint training. + +# 4.7. Limitations + +SkeleTR is currently trained with skeletons pre-extracted from RGB videos, and the adopted pose estimation solution is relatively heavy. To use SkeleTR in real-world apps with limited computation resource, additional efforts are required to switch to a light-weight pose estimation solution that can perform fast inference on CPUs or edge devices, and to retrain the models accordingly. Besides, the short sequence format proposed by SkeleTR strikes a good balance between mitigating skeleton matching noises and modeling intra-sequence dynamics in real-world scenarios. The format may not lead to extra benefits when the ground-truth skeleton sequences are available (details in appendix). + +# 5. Conclusion + +We presented SkeleTR, a unified framework for skeleton action recognition in general scenarios. It extends existing GCN backbones and achieves state-of-the-art performance on multiple skeleton action understanding tasks, including video-level, instance-level, and group-level action recognition. Besides, SkeleTR is compatible with transfer learning and joint learning across different datasets and tasks, which further improves the capability of learned representations. + +Broader Impact. Using human skeletons for action recognition is more robust to environment changes and can help preserve user privacy. Real-world applications of skeleton action recognition include monitoring patient or elderly health, enabling autonomous driving, and facilitating human-computer interactions. However, there could be unintended usages and we advocate responsible usage complying with applicable laws and regulations. + +# References + +[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 3 +[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 2, 3, 7 +[3] Erik Bochinski, Volker Eiselein, and Thomas Sikora. High-speed tracking-by-detection without using image information. In AVSS, 2017. 4 +[4] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. TPAMI, 2019. 2 +[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3 +[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. 2, 4, 5 +[7] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In ICCV, 2021. 1, 2, 3, 6 +[8] Hyung-gun Chi, Myoung Hoon Ha, Seunggeun Chi, Sang Wan Lee, Qixing Huang, and Karthik Ramani. Infogcn: Representation learning for human skeleton-based action recognition. In CVPR, 2022. 1 +[9] Wongun Choi, Khuram Shahid, and Silvio Savarese. What are they doing?: Collective activity classification using spatio-temporal relationship among people. In ICCVW, 2009. 2, 4 +[10] Vasileios Choutas, Philippe Weinzaepfel, Jérôme Revaud, and Cordelia Schmid. Potion: Pose motion representation for action recognition. In CVPR, 2018. 2 +[11] Liu Chunhui, Hu Yueyu, Li Yanghao, Song Sijie, and Liu Jiaying. Pku-mmd: A large scale benchmark for continuous multi-modal human action understanding. arXiv:1703.07475, 2017. 2 +[12] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv:2003.09003, 2020. 1 +[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv:1810.04805, 2018.3 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv:2010.11929, 2020.3,4 +[15] Haodong Duan, Jiaqi Wang, Kai Chen, and Dahua Lin. Pyskl: Towards good practices for skeleton action recognition. ACMM, 2022. 1, 2, 5, 6 + +[16] Haodong Duan, Yue Zhao, Kai Chen, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. In CVPR, 2022. 1, 2, 5, 6 +[17] Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 2 +[18] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. 3 +[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 2 +[20] Kirill Gavrilyuk, Ryan Sanford, Mehrsan Javan, and Cees GM Snoek. Actor-transformers for group activity recognition. In CVPR, 2020. 6 +[21] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The" something something" video database for learning and evaluating visual common sense. In ICCV, 2017. 2 +[22] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, 2018. 1, 2, 5 +[23] Mohamed E Hussein, Marwan Torki, Mohammad A Gowayyed, and Motaz El-Saban. Human action recognition using a temporal hierarchy of covariance descriptors on 3d joint locations. In IJCAI, 2013. 2 +[24] Mostafa S Ibrahim, Srikanth Muralidharan, Zhiwei Deng, Arash Vahdat, and Greg Mori. A hierarchical deep temporal model for group activity recognition. In CVPR, 2016. 1, 2, 4, 5 +[25] Haroon Idrees, Amir R Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The thumos challenge on action recognition for videos "in the wild". CVIU, 2017. 2 +[26] Yanli Ji, Guo Ye, and Hong Cheng. Interactive body part contrast mining for human interaction recognition. In 2014 IEEE international conference on multimedia and expo workshops (ICMEW), pages 1-6. IEEE, 2014. 2 +[27] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal-directed human activities. In CVPR, 2014. 2 +[28] Ang Li, Meghanana Thotakuri, David A Ross, João Carreira, Alexander Vostrikov, and Andrew Zisserman. Theava-kinetics localized human actions video dataset. arXiv:2005.00214, 2020. 2 +[29] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In CVPR, 2019. 6 +[30] Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang. Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In ICCV, 2021. 2 + +[31] Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 5 +[32] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+d 120: A large-scale benchmark for 3d human activity understanding. TPAMI, 2020. 1, 5 +[33] Jun Liu, Amir Shahroudy, Dong Xu, and Gang Wang. Spatio-temporal LSTM with trust gates for 3d human action recognition. In ECCV, 2016. 3, 6 +[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3 +[35] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, 2022. 3 +[36] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In CVPR, 2020. 1, 2, 3, 6 +[37] Qianhui Men, Edmond SL Ho, Hubert PH Shum, and Howard Leung. A two-stream recurrent network for skeleton-based human interaction recognition. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 2771-2778. IEEE, 2021. 2 +[38] Xuan Son Nguyen. Geomnet: A neural network based on riemannian geometries of spd matrix space and cholesky space for 3d skeleton-based interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13379-13389, 2021. 6 +[39] Junting Pan, Siyu Chen, Mike Zheng Shou, Yu Liu, Jing Shao, and Hongsheng Li. Actor-context-actor relation network for spatio-temporal action localization. In CVPR, 2021. 2, 7 +[40] Yunsheng Pang, Qiuhong Ke, Hossein Rahmani, James Bailey, and Jun Liu. Igformer: Interaction graph transformer for skeleton-based human interaction recognition. In ECCV, 2022. 2, 5, 6, 7 +[41] Mauricio Perez, Jun Liu, and Alex C Kot. Interaction relational network for mutual action recognition. IEEE Transactions on Multimedia, 24:366-376, 2021. 2, 5, 6 +[42] Chiara Plizzari, Marco Cannici, and Matteo Matteucci. Skeleton-based action recognition via spatial and temporal transformer networks. CVIU, 2021. 3, 6 +[43] Michalis Raptis and Leonid Sigal. Poselet key-framing: A model for human activity recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2650-2657, 2013. 2 +[44] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 2015. 5 +[45] Kohei Sendo and Norimichi Ukita. Heatmapping of people involved in group activities. In MVA, 2019. 5 +[46] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+d: A large scale dataset for 3d human activity analysis. In CVPR, 2016. 1, 2, 5 + +[47] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In CVPR, 2019. 2, 3 +[48] Bing Shuai, Alessandro Bergamo, Uta Buechler, Andrew Berneshawi, Alyssa Boden, and Joseph Tighe. Large scale real-world multi-person tracking. ECCV, 2022. 1 +[49] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. NeurIPS, 2014. 2 +[50] Sebastian Stein and Stephen J McKenna. Combining embedded accelerometers with computer vision for recognizing food preparation activities. In UbiComp, 2013. 2 +[51] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In CVPR, 2019. 2, 5 +[52] Haritha Thilakarathne, Aiden Nibali, Zhen He, and Stuart Morgan. Pose is all you need: The pose only group activity recognition system (pogars). Machine Vision and Applications, 2022. 5, 6 +[53] Arash Vahdat, Bo Gao, Mani Ranjbar, and Greg Mori. A discriminative key pose sequence model for recognizing human interactions. In 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops), pages 1729-1736. IEEE, 2011. 2 +[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3 +[55] Raviteja Vemulapalli, Felipe Arrate, and Rama Chellappa. Human action recognition by representing 3d skeletons as points in a lie group. In CVPR, 2014. 2 +[56] Hongsong Wang and Liang Wang. Modeling temporal dynamics and spatial configurations of actions using two-stream recurrent neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 499-508, 2017. 2 +[57] Jiang Wang, Zicheng Liu, Ying Wu, and Junsong Yuan. Mining actionlet ensemble for action recognition with depth cameras. In CVPR. IEEE, 2012. 2 +[58] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In ECCV, 2016. 2 +[59] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, 2018. 2, 7 +[60] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In CVPR, 2019. 5 +[61] Fanyi Xiao, Yong Jae Lee, Kristen Grauman, Jitendra Malik, and Christoph Feichtenhofer. Audiovisual slowfast networks for video recognition. arXiv:2001.08740, 2020. 2 +[62] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017. 5 +[63] Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Xia, Zhuowen Tu, and Stefano Soatto. Long short-term transformer for online action detection. NeurIPS, 2021. 3 + +[64] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 1, 2, 3 +[65] Di Yang, Yaohui Wang, Antitza Dantcheva, Lorenzo Garattoni, Gianpiero Francesca, and François Bremond. Unik: A unified framework for real-world skeleton-based action recognition. ArXiv, 2021. 2, 4, 5, 6 +[66] Kiwon Yun, Jean Honorio, Debaleena Chattopadhyay, Tamara L Berg, and Dimitris Samaras. Two-person interaction detection using body-pose features and multiple instance learning. In 2012 IEEE computer society conference on computer vision and pattern recognition workshops, pages 28–35. IEEE, 2012. 2 +[67] Yimeng Zhang, Xiaoming Liu, Ming-Ching Chang, Weina Ge, and Tsuhan Chen. Spatio-temporal phrases for activity recognition. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part III 12, pages 707-721. Springer, 2012. 2 +[68] Yuhan Zhang, Bo Wu, Wen Li, Lixin Duan, and Chuang Gan. Stst: Spatial-temporal specialized transformer for skeleton-based action recognition. In ACMMM, 2021. 3 +[69] Zhengyou Zhang. Microsoft Kinect sensor and its effect. IEEE multimedia, 2012. 2 +[70] Hang Zhao, Antonio Torralba, Lorenzo Torresani, and Zhicheng Yan. Hacs: Human action clips and segments dataset for recognition and temporal localization. In ICCV, 2019. 2 +[71] Jiaojiao Zhao, Yanyi Zhang, Xinyu Li, Hao Chen, Bing Shuai, Mingze Xu, Chunhui Liu, Kaustav Kundu, Yuanjun Xiong, Davide Modolo, et al. Tuber: Tubelet transformer for video action detection. In CVPR, 2022. 3 +[72] Honglu Zhou, Asim Kadav, Aviv Shamsian, Shijie Geng, Farley Lai, Long Zhao, Ting Liu, Mubbasir Kapadia, and Hans Peter Graf.Composer: Compositional learning of group activity in videos.ECCV, 2022.5,6,7 +[73] Wentao Zhu, Cuiling Lan, Junliang Xing, Wenjun Zeng, Yanghao Li, Li Shen, and Xiaohui Xie. Co-occurrence feature learning for skeleton based action recognition using regularized deep LSTM networks. In AAAI, 2016. 3 +[74] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv:2010.04159, 2020. 3 \ No newline at end of file diff --git a/2023/SkeleTR_ Towards Skeleton-based Action Recognition in the Wild/images.zip b/2023/SkeleTR_ Towards Skeleton-based Action Recognition in the Wild/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..724865193cabd5500477f61e274ac535190343b5 --- /dev/null +++ b/2023/SkeleTR_ Towards Skeleton-based Action Recognition in the Wild/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57e63c25aab7775cd49ce16b8c12d8ac20429560d99740cda51ecf8cd68f90da +size 461757 diff --git a/2023/SkeleTR_ Towards Skeleton-based Action Recognition in the Wild/layout.json b/2023/SkeleTR_ Towards Skeleton-based Action Recognition in the Wild/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ae55592c0550029121957980c4c7593062b3301e --- /dev/null +++ b/2023/SkeleTR_ Towards Skeleton-based Action Recognition in the Wild/layout.json @@ -0,0 +1,9470 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 93, + 103, + 501, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 103, + 501, + 121 + ], + "spans": [ + { + "bbox": [ + 93, + 103, + 501, + 121 + ], + "type": "text", + "content": "SkeleTR: Towards Skeleton-based Action Recognition in the Wild" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 142, + 492, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 142, + 492, + 192 + ], + "spans": [ + { + "bbox": [ + 102, + 142, + 492, + 192 + ], + "type": "text", + "content": "Haodong Duan" + }, + { + "bbox": [ + 102, + 142, + 492, + 192 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 102, + 142, + 492, + 192 + ], + "type": "text", + "content": " Mingze Xu" + }, + { + "bbox": [ + 102, + 142, + 492, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 142, + 492, + 192 + ], + "type": "text", + "content": " Bing Shuai" + }, + { + "bbox": [ + 102, + 142, + 492, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 142, + 492, + 192 + ], + "type": "text", + "content": " Davide Modolo" + }, + { + "bbox": [ + 102, + 142, + 492, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 142, + 492, + 192 + ], + "type": "text", + "content": " Zhuowen Tu" + }, + { + "bbox": [ + 102, + 142, + 492, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 142, + 492, + 192 + ], + "type": "text", + "content": " Joseph Tighe" + }, + { + "bbox": [ + 102, + 142, + 492, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 142, + 492, + 192 + ], + "type": "text", + "content": " Alessandro Bergamo" + }, + { + "bbox": [ + 102, + 142, + 492, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 142, + 492, + 192 + ], + "type": "text", + "content": " The Chinese University of Hong Kong AWS AI Labs" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 96, + 196, + 497, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 196, + 497, + 207 + ], + "spans": [ + { + "bbox": [ + 96, + 196, + 497, + 207 + ], + "type": "text", + "content": "dhd.efz@gmail.com, {xumingze,bshuai,dmodolo,ztu,tighej,bergamo}@amazon.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 235, + 192, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 235, + 192, + 248 + ], + "spans": [ + { + "bbox": [ + 143, + 235, + 192, + 248 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 260, + 290, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 260, + 290, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 290, + 513 + ], + "type": "text", + "content": "We present SkeleTR, a new framework for skeleton-based action recognition. In contrast to prior work, which focuses mainly on controlled environments, we target more general scenarios that typically involve a variable number of people and various forms of interaction between people. SkeleTR works with a two-stage paradigm. It first models the intra-person skeleton dynamics for each skeleton sequence with graph convolutions, and then uses stacked Transformer encoders to capture person interactions that are important for action recognition in general scenarios. To mitigate the negative impact of inaccurate skeleton associations, SkeleTR takes relative short skeleton sequences as input and increases the number of sequences. As a unified solution, SkeleTR can be directly applied to multiple skeleton-based action tasks, including video-level action classification, instance-level action detection, and group-level activity recognition. It also enables transfer learning and joint training across different action tasks and datasets, which result in performance improvement. When evaluated on various skeleton-based action recognition benchmarks, SkeleTR achieves the state-of-the-art performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 518, + 128, + 531 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 128, + 531 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 128, + 531 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 538, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 287, + 696 + ], + "type": "text", + "content": "Skeleton-based action recognition has received increasing attention in recent years due to its robustness against background and illumination changes [64, 16]. Existing methods [36, 7, 8, 15] mostly use Graph Convolutional Networks (GCNs) as backbones and focus on modeling spatiotemporal information from long skeleton sequences separately. While observing promising results in simplified and controlled environments [46, 32], they are less effective in more realistic scenarios with a variable number of people performing diverse actions. First, considering that linking skeletons (e.g., using skeletons' confidence scores [64]) in a long time span may inevitably generate more association errors [12, 48], requiring long sequences as input can bring" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 234, + 545, + 365 + ], + "blocks": [ + { + "bbox": [ + 310, + 234, + 545, + 365 + ], + "lines": [ + { + "bbox": [ + 310, + 234, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 310, + 234, + 545, + 365 + ], + "type": "image", + "image_path": "abcaf131eac95a9937be84888aecc7bbdc0156d8545bb16e99b89b9c82682192.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 367, + 546, + 422 + ], + "lines": [ + { + "bbox": [ + 304, + 367, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 546, + 422 + ], + "type": "text", + "content": "Figure 1: SkeleTR is a general framework for skeleton action recognition which can handle different tasks, including video-level action classification, instance-level action detection, and group-level activity recognition. It adopts a two-stage paradigm: the intra-sequence and inter-sequence modeling, respectively." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 425, + 547, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 547, + 485 + ], + "type": "text", + "content": "lots of noises. Second, prior works pay little attention to modeling the relations among a group of people, which plays an important role in recognizing both person interactions (e.g., talking to someone [22]) and group activities (e.g., pass/win-point in a volleyball game [24])." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "type": "text", + "content": "In this paper, we propose SkeleTR for skeleton action recognition in real-world scenarios and handle multiple action tasks (i.e., video-level, instance-level, and group-level action recognition) in a general framework. SkeleTR samples a large number of short skeleton sequences instead of a few long ones (practice of prior works [64, 16]) from each video. Each short sequence directly corresponds to an atomic action of a single person, thus the collection of multiple short sequences in space and time can represent hierarchical actions (with a longer time span) of a single person or a video-level action involving multiple people. The new input paradigm can be applied to instance-level and video-level recognition of actions with different temporal granularities, and it has several unique advantages. First, association errors are less likely to happen in a shorter skeleton sequence [12, 48]. We show that the improved sequence consistency leads to better intra-sequence modeling. Second, this new format is more flexible and helps to create comprehensive yet compact inputs. With a finer temporal" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 225, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 225, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 225, + 714 + ], + "type": "text", + "content": "*The work was done during an Amazon internship." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "type": "text", + "content": "13634" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 120 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 120 + ], + "type": "text", + "content": "granularity, the same input size budget can cover more human skeletons. It reduces redundant zero-paddings for a small number of people, and mitigates information loss during input sampling for a large number of people." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 121, + 286, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 121, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 121, + 286, + 335 + ], + "type": "text", + "content": "SkeleTR adopts a two-stage paradigm as shown in Fig 1. Stage 1 performs intra-sequence modeling to capture the motion patterns of each individual skeleton sequence with a GCN backbone. The GCN backbone outputs a spatiotemporal feature for each sequence. Stage 2 concatenates all sequence features and feeds them into Transformer encoders to model inter-sequence relationships. At the end of stage 2, average pooling aggregates all features belonging to the same sequence into a single feature vector. By leveraging inter-sequence modeling, the output sequence feature incorporates the interactions between people and the global context of the video, which is helpful in both instance-level and video-level action recognition. To reduce the dimension of spatio-temporal skeleton features and keep the computations of Transformer tractable, we insert a novel Mix Pooling module before stage 2. It generates a compact representation with fine granularity for each sequence by applying multiple partial dimension reduction strategies in parallel." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 336, + 286, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 336, + 286, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 336, + 286, + 454 + ], + "type": "text", + "content": "SkeleTR easily enables transfer learning and joint learning across different datasets and tasks. With a unified input format and model architecture, one can jointly train multiple tasks with different datasets, each with its own loss function and ground truth annotations. This allows for leveraging auxiliary datasets for pre-training or joint training of a target task with limited labeled data. In experiments, we find that joint training serves as a strong regularizer and improves recognition performance, especially in scenarios with limited training data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 456, + 286, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 286, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 286, + 657 + ], + "type": "text", + "content": "To evaluate the effectiveness of SkeleTR, we conduct comprehensive study on three skeleton-based action recognition tasks: action classification [6, 65, 46], spatiotemporal action detection [22], and group activity recognition [24]. Experimental results show that SkeleTR works well with different GCN backbones and significantly outperforms prior works on all three tasks. For example, with ST-GCN++ [15] backbone, SkeleTR outperforms the state-of-the-art methods by " + }, + { + "bbox": [ + 48, + 456, + 286, + 657 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 48, + 456, + 286, + 657 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 48, + 456, + 286, + 657 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 48, + 456, + 286, + 657 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 48, + 456, + 286, + 657 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 48, + 456, + 286, + 657 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 48, + 456, + 286, + 657 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 48, + 456, + 286, + 657 + ], + "type": "text", + "content": " on Poetics [64], AVA [22], Volleyball [24], and NTU-Inter [40]. With Poetics joint training, the AVA performance is further improved by " + }, + { + "bbox": [ + 48, + 456, + 286, + 657 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 48, + 456, + 286, + 657 + ], + "type": "text", + "content": " mAP. The great capability of SkeleTR leads to another noteworthy result: we first observe the strong complementarity of skeleton action recognition. Combining SkeleTR prediction and the RGB state-of-the-art [2, 39] with simple late fusion leads to great performance boost (2.3% Poetics Top-1, 3% AVA mAP)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 669, + 137, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 137, + 681 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 137, + 681 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 689, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 689, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 689, + 286, + 712 + ], + "type": "text", + "content": "Human Action Understanding. Action classification [6, 21] is the most fundamental task in human action under" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 72, + 544, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 262 + ], + "type": "text", + "content": "standing, aiming at recognizing the video-level action in a short video clip. Another task of great interest is recognizing the instance/person-level actions in a video, either at the atomic [22, 28] or hierarchical [30] level. Meanwhile, other areas of interest include group activity recognition [24, 9], action localization [17, 25, 70], or action segmentation [27, 50]. Various modalities in videos can be utilized to solve the action understanding tasks, such as RGB appearance [58, 6, 19, 59], optical flow [49], audio waves [61], and human skeletons [64, 16]. Spatio-temporal representations learned end-to-end from RGB appearance yield good performance on standard benchmarks. However, they are data intensive and less robust to context changes. In this work, we propose a unified framework to learn video representations based on human skeletons, that can be easily applied to multiple action tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 267, + 544, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 267, + 544, + 553 + ], + "spans": [ + { + "bbox": [ + 308, + 267, + 544, + 553 + ], + "type": "text", + "content": "Skeleton-based Action Recognition. Human skeletons are robust against background and illumination changes [64, 47], and can be easily obtained with sensors [69] or pose estimators [4, 51]. In skeleton action recognition, the majority of research has focused on modeling the intra-person skeleton dynamics. With the rise of deep learning, manual feature engineering [23, 55, 57] has gradually been replaced by end-to-end representation learning with deep neural architectures (GCN [64], 2D-CNN [10], or 3D-CNN [16]). Among them, GCN-based approaches [64, 47, 36, 7] have been the most popular paradigm, which construct spatial-temporal graphs of human keypoint coordinates and perform spatial and temporal modeling. These methods model the intra-person skeleton dynamics for in-lab collected videos [46, 11], featuring small number of persons (1 or 2) per video and accurate person identities. For more challenging scenarios [6], the model performance decreases due to the noises produced in matching skeletons to sequences. Besides, GCN approaches primarily focus on predicting video-level actions, and do not benefit from context or interaction modeling when applied to instance-level action recognition. To this end, we propose SkeleTR, which expands existing GCN approaches and works universally well for diversified scenarios and recognition tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 558, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 558, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 558, + 544, + 712 + ], + "type": "text", + "content": "Human Interaction Recognition. Human interaction recognition [43, 53, 67] is a sub-task of action recognition. It aims at identifying and categorizing activities between people. Early works [66, 26] built hand-crafted features for interaction modeling and adopted classic machine learning algorithms for activity classification. Subsequent works proposed to use two-stream RNNs to model the two-person interactions [41, 37, 56]. Recently, IGFormer [40] first proposed to use a two-stream Transformer for two-person interaction modeling and observed improved performance. However, most deep methods share the same limitations: 1. they are largely restricted to two-person scenarios and tend to suffer from exponentially growing inference" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 317, + 741 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 317, + 741 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 317, + 741 + ], + "type": "text", + "content": "13635" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 64, + 519, + 199 + ], + "blocks": [ + { + "bbox": [ + 74, + 64, + 519, + 199 + ], + "lines": [ + { + "bbox": [ + 74, + 64, + 519, + 199 + ], + "spans": [ + { + "bbox": [ + 74, + 64, + 519, + 199 + ], + "type": "image", + "image_path": "c04aa9b76c8349ec0693112c454311d4fc298ff9cfb52ba2538b912c98b3ad5e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 200, + 547, + 245 + ], + "lines": [ + { + "bbox": [ + 46, + 200, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 200, + 547, + 245 + ], + "type": "text", + "content": "Figure 2: The architecture of SkeLETR. SkeLETR takes a large set of short skeleton sequences as input and performs a two-stage modeling process. It first adopts a GCN backbone to model the intra-sequence skeleton dynamics. Then it processes the sequence features with stacked Transformer encoders, which helps to model person interactions and leverage the long-range context. Between the two stages, we adopt a Mix Pooling layer to compress the spatio-temporal features and keep computations tractable." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 248, + 288, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 288, + 308 + ], + "type": "text", + "content": "costs with the person number; 2. they fail to leverage the strong intra-sequence modeling capabilities of GCNs. To overcome these challenges, we introduce SkeleTR, which integrates intra-sequence and inter-sequence modeling with good scalability and small computation cost." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 309, + 288, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 309, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 288, + 477 + ], + "type": "text", + "content": "Visual Transformer. With the great success in NLP [54, 13], Transformers have been increasingly applied to diversified vision tasks [14, 34, 5, 74] as well as the RGB-based action understanding [35, 18, 1, 2, 63, 71], and have achieved promising results. For skeleton action recognition, there exist works [42, 68] that directly use Transformers for intra-sequence spatio-temporal modeling. However, without exploiting the locality, the Transformer-based motion modeling only leads to inferior recognition performance compared to GCN-based ones. In SkeleTR, to enjoy the best of both worlds, we adopt GCNs for intra-sequence motion modeling, and use Transformer encoders for intersequence relationship modeling among multiple persons' fine-grained skeleton features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 491, + 107, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 107, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 107, + 504 + ], + "type": "text", + "content": "3. SkeleTR" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 513, + 115, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 115, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 115, + 525 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "text", + "content": "Both SkeleTR and prior GCN approaches directly take human joint coordinates as input. A person in a specific frame corresponds to a skeleton with " + }, + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "text", + "content": " joints (each has " + }, + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "text", + "content": " dimensions, " + }, + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "text", + "content": " is 2 or 3). Skeletons in " + }, + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "text", + "content": " different frames (1 skeleton per frame) can be linked in temporal order to form a skeleton sequence " + }, + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i\\in \\mathbb{R}^{T\\times V\\times C}" + }, + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "text", + "content": " with length " + }, + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "text", + "content": ". A skeleton action recognizer takes " + }, + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "text", + "content": " skeleton sequences as input, and outputs a score vector " + }, + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{S}\\in \\mathbb{R}^K" + }, + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "text", + "content": " for video-level action recognition, where " + }, + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "text", + "content": " is the number of action classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Spatial-temporal GCNs are widely used to model skeleton sequences for action classification, which focus on modeling intra-sequence skeleton dynamics. Most GCN methods [64, 47, 7] share a similar high-level design (Fig 2(a)), where a backbone with " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{N}_1" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " GCN blocks is used to extract features of each skeleton sequence. A GCN block com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 248, + 547, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 248, + 547, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 248, + 547, + 369 + ], + "type": "text", + "content": "prises a GCN layer and a TCN layer. The GCN layer performs spatial modeling by fusing joint features within the same skeleton. The TCN layer employs 1D Conv to capture the motion patterns of each joint. GCNs model long skeleton sequences that span the entire video in parallel, and aggregate multiple sequence features with average pooling to predict the video-level action. On existing skeleton action recognition benchmarks, GCNs exhibit superior intrasequence modeling capability than other architectures (e.g., RNN [73, 33] or Transformers [42, 68])." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 369, + 547, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 547, + 514 + ], + "type": "text", + "content": "SkeleTR inherits the effective intra-sequence modeling of GCNs, while also enhancing the flexibility, generality, and inter-person modeling capability. SkeleTR takes " + }, + { + "bbox": [ + 304, + 369, + 547, + 514 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 369, + 547, + 514 + ], + "type": "text", + "content": " short skeleton sequences as input (Sec. 3.2), and outputs one feature vector per sequence. It first adopts a GCN backbone to model the motion patterns of each skeleton sequence, then uses stacked Transformer encoders to capture the interactions and leverage long-range context (Sec. 3.3). Between two stages, we use a Mix Pooling layer to compress spatiotemporal sequence features into a compact multi-view representation to keep computation tractable. Fig 2 compares the design of SkeleTR and previous GCN methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 525, + 495, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 495, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 495, + 538 + ], + "type": "text", + "content": "3.2. The Input Construction of SkeleTR" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 544, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 687 + ], + "type": "text", + "content": "For skeleton action classification in real-world scenarios, prior works [64, 47, 36] use a few long skeleton sequences as input, where skeletons at different timestamps are linked with rudimentary association methods. They use a naive confidence score based matching to link skeletons to sequences as long as the entire video (as shown in Fig 3). Specifically, they mostly keep " + }, + { + "bbox": [ + 304, + 544, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\hat{M} = 2" + }, + { + "bbox": [ + 304, + 544, + 545, + 687 + ], + "type": "text", + "content": " skeletons with the highest confidence scores per frame, and link skeletons based on confidence rank. In cases where it is applied to challenging videos involving multiple persons, this strategy can lead to noisy skeleton sequences and sub-optimal performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "In contrast, we propose to sample a large number of short skeleton sequences as input, and link skeletons at differ" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "type": "text", + "content": "13636" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 67, + 272, + 209 + ], + "blocks": [ + { + "bbox": [ + 61, + 67, + 272, + 209 + ], + "lines": [ + { + "bbox": [ + 61, + 67, + 272, + 209 + ], + "spans": [ + { + "bbox": [ + 61, + 67, + 272, + 209 + ], + "type": "image", + "image_path": "f8a33ba00e35f6bb53ce07affe21e3b7785cdd0c5d82e6a3df0dfdc252ce0fe9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 213, + 285, + 279 + ], + "lines": [ + { + "bbox": [ + 45, + 213, + 285, + 279 + ], + "spans": [ + { + "bbox": [ + 45, + 213, + 285, + 279 + ], + "type": "text", + "content": "Figure 3: The difference of input formats. Prior works link skeletons at different timestamps based on the confidence score rank to form " + }, + { + "bbox": [ + 45, + 213, + 285, + 279 + ], + "type": "inline_equation", + "content": "\\hat{M} = 2" + }, + { + "bbox": [ + 45, + 213, + 285, + 279 + ], + "type": "text", + "content": " long sequences, each covers the entire video. SkeleTR, instead, directly takes multiple shorter sequences as input. Each sequence only covers " + }, + { + "bbox": [ + 45, + 213, + 285, + 279 + ], + "type": "inline_equation", + "content": "1\\sim 2" + }, + { + "bbox": [ + 45, + 213, + 285, + 279 + ], + "type": "text", + "content": " seconds, thus has better consistency than long sequences." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "text", + "content": "ent timestamps into sequences based on their IOU similarity. Specifically, to sample sequences with length " + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "text", + "content": ", we first uniformly select a set of key frames with a stride of " + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "inline_equation", + "content": "T / 2" + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "text", + "content": ". For each skeleton in every key frame, we create a sequence by linking skeletons within a window of " + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "text", + "content": " frames using IoU-based matching [3], resulting in a total of " + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "inline_equation", + "content": "M'" + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "text", + "content": " sequences (" + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "inline_equation", + "content": "M'" + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "text", + "content": " is variable for different videos). Our new input format is compatible with a variable number of input sequences. By using different numbers of sequences (" + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "text", + "content": ") for each training batch, SkeleTR can handle inputs with different " + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "text", + "content": ", which helps to reduce the usages of zero-padding at test time and improves efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 430, + 212, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 430, + 212, + 443 + ], + "spans": [ + { + "bbox": [ + 45, + 430, + 212, + 443 + ], + "type": "text", + "content": "3.3. Inter-sequence Modeling Stage" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 44, + 447, + 284, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 44, + 447, + 284, + 578 + ], + "spans": [ + { + "bbox": [ + 44, + 447, + 284, + 578 + ], + "type": "text", + "content": "After processing short skeleton sequences with the GCN backbone, SkeleTR employs stacked Transformer encoders to model inter-sequence relationships. The process starts with reducing the feature dimension using a Mix Pooling layer, resulting in " + }, + { + "bbox": [ + 44, + 447, + 284, + 578 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 44, + 447, + 284, + 578 + ], + "type": "text", + "content": " feature vectors for each sequence. Next, a sequence-level spatial-temporal embedding is combined with the feature vectors. The resulting sequence features " + }, + { + "bbox": [ + 44, + 447, + 284, + 578 + ], + "type": "inline_equation", + "content": "(M\\times L" + }, + { + "bbox": [ + 44, + 447, + 284, + 578 + ], + "type": "text", + "content": " vectors in total) are concatenated and modeled using a standard Transformer with multiple encoder blocks [14]. This architecture is depicted in Fig 2(b). We present more details in the next paragraphs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 44, + 581, + 285, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 44, + 581, + 285, + 711 + ], + "spans": [ + { + "bbox": [ + 44, + 581, + 285, + 711 + ], + "type": "text", + "content": "Mix Pooling. The cost of computing the attention matrices in a Transformer block grows quadratically with the input dimensionality, resulting in a computational cost of " + }, + { + "bbox": [ + 44, + 581, + 285, + 711 + ], + "type": "inline_equation", + "content": "\\Theta((M \\times T \\times V)^2)" + }, + { + "bbox": [ + 44, + 581, + 285, + 711 + ], + "type": "text", + "content": " in our case, if no dimension reduction is performed (" + }, + { + "bbox": [ + 44, + 581, + 285, + 711 + ], + "type": "inline_equation", + "content": "L = T \\times V" + }, + { + "bbox": [ + 44, + 581, + 285, + 711 + ], + "type": "text", + "content": "). Given a video with a typical configuration (" + }, + { + "bbox": [ + 44, + 581, + 285, + 711 + ], + "type": "inline_equation", + "content": "T = 15" + }, + { + "bbox": [ + 44, + 581, + 285, + 711 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 44, + 581, + 285, + 711 + ], + "type": "inline_equation", + "content": "V = 17" + }, + { + "bbox": [ + 44, + 581, + 285, + 711 + ], + "type": "text", + "content": ") and a Transformer consisting of only " + }, + { + "bbox": [ + 44, + 581, + 285, + 711 + ], + "type": "inline_equation", + "content": "\\mathrm{N}_2 = 2" + }, + { + "bbox": [ + 44, + 581, + 285, + 711 + ], + "type": "text", + "content": " encoder blocks, the Transformer itself consumes 52.7 GFLOPs, which is impractical to compute. Another extreme is to remove the dimension " + }, + { + "bbox": [ + 44, + 581, + 285, + 711 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 44, + 581, + 285, + 711 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 44, + 581, + 285, + 711 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 44, + 581, + 285, + 711 + ], + "type": "text", + "content": " with global average pooling, which reduces the computational complexity to " + }, + { + "bbox": [ + 44, + 581, + 285, + 711 + ], + "type": "inline_equation", + "content": "\\Theta(M^2)" + }, + { + "bbox": [ + 44, + 581, + 285, + 711 + ], + "type": "text", + "content": " but also results in loss of fine-grained" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 452, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 452, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 452, + 83 + ], + "type": "text", + "content": "features for each skeleton sequence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "text", + "content": "Between two extremes " + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "inline_equation", + "content": "(L = T\\times V" + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "inline_equation", + "content": "L = 1\\times 1)" + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "text", + "content": " we adopt multi-stream partial dimension reductions to reduce the computation cost. We denote a partial dimension reduction stream as " + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "inline_equation", + "content": "A\\times B" + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "text", + "content": ", indicating reducing temporal dimension " + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "inline_equation", + "content": "(T)" + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "text", + "content": " and spatial dimension " + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "inline_equation", + "content": "(V)" + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "text", + "content": ". Each stream provides a specific view of the spatio-temporal skeleton features. Multiple reduction streams can be combined in Mix Pooling to compress raw features into a comprehensive representation. Specifically, we adopt 3 partial dimension reduction streams: " + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "inline_equation", + "content": "T\\times 1" + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "inline_equation", + "content": "1\\times P" + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "text", + "content": " joints aggregated to " + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "inline_equation", + "content": "P = 5" + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "text", + "content": " parts: head, two arms, two legs) (Fig 2(c)). Mix Pooling concatenates the output of 3 streams and keeps " + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "inline_equation", + "content": "L = 1 + T + P" + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "text", + "content": " features per sequence. The heterogeneous features are then processed with one single Transformer. Mix Pooling enables the cross-view feature interaction and helps keep computations tractable, yielding a complexity of " + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "inline_equation", + "content": "\\Theta(M^2(1 + T + P)^2)" + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "inline_equation", + "content": "\\sim 0.6" + }, + { + "bbox": [ + 305, + 85, + 545, + 288 + ], + "type": "text", + "content": " GFLOPs / video)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "text", + "content": "Sequence-level Spatial Temporal Embedding. We further combine " + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "text", + "content": " features per sequence with a sequence-level spatial-temporal embedding. We utilize a fully-connected layer to encode both the bounding box and the normalized timestamp (computed based on the entire video) of the center skeleton for each sequence, which then serves as the embedding. The embeddings are added with the Mix-Pooled features, and further processed by a layer normalization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 385, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 385, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 305, + 385, + 545, + 564 + ], + "type": "text", + "content": "Transformer-based Interaction Modeling. We use a standard Transformer with " + }, + { + "bbox": [ + 305, + 385, + 545, + 564 + ], + "type": "inline_equation", + "content": "\\mathrm{N}_2" + }, + { + "bbox": [ + 305, + 385, + 545, + 564 + ], + "type": "text", + "content": " encoder blocks for modeling sequence features. Each block consists of a multi-head self-attention layer (MHSA) and a feed-forward network (FFN), and adopts the post-layernorm style. The multi-head design in the encoder blocks helps to capture various kinds of relationships among skeleton features, including but not limited to the inter-person interactions and intra-person long-range context. At last, we use intra-sequence average pooling to aggregate " + }, + { + "bbox": [ + 305, + 385, + 545, + 564 + ], + "type": "inline_equation", + "content": "L = (1 + T + P)" + }, + { + "bbox": [ + 305, + 385, + 545, + 564 + ], + "type": "text", + "content": " features per sequence into a single one. The output sequence features, which incorporate both person interactions and a comprehensive context, are then used to predict the instance-level or video-level action. We present the detailed configuration of encoder blocks in the appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 574, + 487, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 487, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 487, + 586 + ], + "type": "text", + "content": "3.4. SkeleTR as a General Framework" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 593, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 593, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 305, + 593, + 545, + 640 + ], + "type": "text", + "content": "As a general framework for skeleton action recognition, SkeleTR can be applied to various skeleton-based action recognition tasks. It also enables and benefits from transfer learning and joint learning across different tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 642, + 545, + 713 + ], + "type": "text", + "content": "Video-level Action Classification. The tasks of action classification [6, 65] and group activity recognition [24, 9] both take as input a video and predict the action at video-level. Note that the input videos may contain a variable number of people. In SkeleTR, we sample numerous short skeleton sequences that last " + }, + { + "bbox": [ + 305, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "1 \\sim 2" + }, + { + "bbox": [ + 305, + 642, + 545, + 713 + ], + "type": "text", + "content": " seconds. We then average the " + }, + { + "bbox": [ + 305, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "M" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 741 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 741 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 741 + ], + "type": "text", + "content": "13637" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 71, + 286, + 171 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 286, + 171 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 286, + 171 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 286, + 171 + ], + "type": "image", + "image_path": "e18071554911c9145d282bce8b7720c319570936826c1ff56d1f7d6887400996.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 174, + 287, + 229 + ], + "lines": [ + { + "bbox": [ + 46, + 174, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 287, + 229 + ], + "type": "text", + "content": "Figure 4: The curve of training / validation loss and validation mAP on AVA. Joint training with Posetics [65] significantly alleviates overfitting for AVA action detection. The jointly-trained model achieves much higher validation mAP, with the gap of training / validation losses largely shrinks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 234, + 287, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 287, + 294 + ], + "type": "text", + "content": "output features, and use a fully-connected layer to predict the video action class. Thanks to the local motion modeling with GCN and interaction modeling with Transformers, SkeleTR achieves strong performance in both action classification and group activity recognition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 295, + 287, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 295, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 287, + 402 + ], + "type": "text", + "content": "Instance-level Action Recognition. Spatio-temporal action detection [22] aims to recognize actions of each person at specific timestamps (instance-level recognition). For this task, we sample short skeleton sequences with length " + }, + { + "bbox": [ + 46, + 295, + 287, + 402 + ], + "type": "inline_equation", + "content": "T^{1}" + }, + { + "bbox": [ + 46, + 295, + 287, + 402 + ], + "type": "text", + "content": " in a longer temporal window " + }, + { + "bbox": [ + 46, + 295, + 287, + 402 + ], + "type": "inline_equation", + "content": "(5 \\sim 10s)" + }, + { + "bbox": [ + 46, + 295, + 287, + 402 + ], + "type": "text", + "content": " as input, and utilize the sequence-level features for instance-level action recognition. SkeleTR enables the modeling of both intra-sequence motion patterns and inter-sequence interactions for skeleton-based spatio-temporal action detection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 403, + 287, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 523 + ], + "type": "text", + "content": "Transfer Learning and Joint Learning. SkeLETR enables to: 1. transfer the representation learned on one task to another new task; 2. jointly train different datasets and tasks, which in turn alleviate the overfitting issues on small datasets. Fig 4 shows the curves of training and validation losses, and validation mAP for AVA. We see that joint training serves as a strong regularizer and alleviates the severe overfitting. With joint training, the gap between training and validation losses largely shrinks, and the recognition performance on the validation set becomes much stronger." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 536, + 127, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 127, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 127, + 550 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 556, + 110, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 556, + 110, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 556, + 110, + 567 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 574, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 287, + 684 + ], + "type": "text", + "content": "Kinetics-400 (K400) [6] contains videos with variable number of people. Some K400 actions cannot be distinguished by only using skeletons, thus we use a subset of K400, named **Posetics** [65], as the main benchmark for action classification. Posetics merges together classes featuring identical skeletons, resulting in 320 total classes (e.g., 'eating burger' and 'eating hotdog' are merged into 'eating something'). For K400 and Posetics, we use 2D human skeletons provided by [16] (extracted with Faster" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 544, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 544, + 96 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 544, + 96 + ], + "type": "text", + "content": "RCNN [44] (R50) and HRNet [51]). We provide K400 results in the appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 97, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 97, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 97, + 545, + 180 + ], + "type": "text", + "content": "NTU-Inter is a commonly used benchmark for two-person interaction recognition [40, 41]. It's a subset of the inlab collected NTURGB+D [32, 46], containing only mutual action classes (with two subjects). Consistent with NTURGB+D, NTU-Inter includes 4 data splits: NTU60-XSub, NTU60-XView, NTU120-XSub, NTU120-XSet. We use the official 3D skeletons for NTU-Inter experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 181, + 546, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 181, + 546, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 546, + 300 + ], + "type": "text", + "content": "AVA 2.2 [22] is a spatio-temporal action detection benchmark. It's a multi-label dataset with 80 action classes, falling into three major categories: person movement, object manipulation, and person interaction. Following the common practice [60], we use Faster-RCNN (ResNeXt101-FPN [31, 62]) for person detection and HRNet for pose estimation. The detector is pretrained on COCO and finetuned on AVA training bboxes. Following the benchmark guidelines, we evaluate 60 action classes. We report the overall mAP and the mAP of each major category." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 301, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 545, + 373 + ], + "type": "text", + "content": "Volleyball [24] is a group activity recognition dataset with 8 volleyball group activities. Each clip has up to 13 persons, with the center frame annotated with GT person boxes. Following common practice [72, 52], we use the tracklets from [45] and HRNet for pose extraction, train models on training and validation sets, and report the test set results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 382, + 440, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 382, + 440, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 382, + 440, + 395 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 401, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 509 + ], + "type": "text", + "content": "Architecture. Unless otherwise noted, we adopt ST-GCN++ [15] as the GCN backbone. Particularly, we consider two variants: SkeleTR-S (with fewer GCN blocks and smaller output dimension, details in appendix), and SkeleTR-L (original ST-GCN++.). The inter-sequence modeling stage consists of a Mix Pooling module and two Transformer encoder blocks. The feature dimension in each encoder blocks equals to the output channels of the GCN backbone (128 for SkeleTR-S and 256 for SkeleTR-L)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "type": "text", + "content": "Training. We train all models end-to-end with SGD and batch size 128. For all models, we set the base learning rate (LR) to 0.1 and adopt a multi-step LR scheduler with the decay factor 0.1. For Poetics and NTU-Inter, we train models for 100 epochs, and set milestones at the " + }, + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "type": "inline_equation", + "content": "70_{th}" + }, + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "type": "inline_equation", + "content": "85_{th}" + }, + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "type": "text", + "content": " epoch. For AVA and Volleyball, we train the models for 40 epochs, and set milestones at the " + }, + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "type": "inline_equation", + "content": "30_{th}" + }, + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "type": "inline_equation", + "content": "35_{th}" + }, + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "type": "text", + "content": " epoch. We use cross entropy loss " + }, + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "type": "inline_equation", + "content": "(\\mathbf{L}_{\\mathbf{CE}})" + }, + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "type": "text", + "content": " for Poetics, NTU-Inter and Volleyball, and use binary cross entropy loss " + }, + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "type": "inline_equation", + "content": "(\\mathbf{L}_{\\mathbf{BCE}})" + }, + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "type": "text", + "content": " for AVA (multi-label). We explicitly set a loss weight " + }, + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "type": "inline_equation", + "content": "\\lambda = 100" + }, + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "type": "text", + "content": " in all AVA experiments due to the small magnitude of " + }, + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathbf{BCE}}" + }, + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "type": "text", + "content": ". In Poetics-AVA joint training, we concatenate samples in two datasets and optimize the model with a weighted combination of two losses: " + }, + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{L} = \\mathbf{L}_{\\mathbf{CE}} + \\lambda \\mathbf{L}_{\\mathbf{BCE}}" + }, + { + "bbox": [ + 304, + 509, + 546, + 700 + ], + "type": "text", + "content": ". For AVA, we sample short skeleton sequences in a small temporal window (5s / 10s) to form a training sample." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "Inference. We follow standard practices [15] to predict ac" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "inline_equation", + "content": "{}^{1}T" + }, + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "text", + "content": " is restricted by annotation granularity. For example,AVA [22] annotates atomic actions at 1fps, for which " + }, + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "text", + "content": " should cover at most 2 seconds." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "text", + "content": "13638" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 71, + 145, + 160 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 145, + 160 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 145, + 160 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 145, + 160 + ], + "type": "table", + "html": "
MethodsPosetics Top-1Posetics Top-5
ST-TR [42]47.571.3
MS-G3D [36]52.675.8
CTR-GCN [7]54.176.7
ST-GCN++ [15]52.576.0
UNIK [65]52.575.7
PoseC3D [16]53.177.1
SkeleTR-S55.978.9
SkeleTR-L57.980.3
SkeleTR [C]57.680.1
", + "image_path": "9f0e1a2e62bb365246c46e5d7e729ebf8dcb135b5576c7da581c1f505e885247.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 151, + 71, + 293, + 160 + ], + "blocks": [ + { + "bbox": [ + 50, + 162, + 143, + 170 + ], + "lines": [ + { + "bbox": [ + 50, + 162, + 143, + 170 + ], + "spans": [ + { + "bbox": [ + 50, + 162, + 143, + 170 + ], + "type": "text", + "content": "(a) Action Classification: Poetics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 71, + 293, + 160 + ], + "lines": [ + { + "bbox": [ + 151, + 71, + 293, + 160 + ], + "spans": [ + { + "bbox": [ + 151, + 71, + 293, + 160 + ], + "type": "table", + "html": "
MethodsNTU60 XSubNTU60 XViewNTU120 XSubNTU120 XSet
ST-LSTM [33]83.087.363.066.6
LSTM-IRN [41]90.593.577.779.6
GeomNet [38]93.696.386.587.6
IGFormer [40]93.696.585.486.5
AS-GCN [29]87.192.077.879.3
CTR-GCN [7]91.694.383.284.4
SkeleTR-S94.697.187.488.0
SkeleTR-L94.797.587.888.3
SkeleTR [C]94.897.787.888.3
", + "image_path": "3ef1279ff35c90915d69990681f905b3af2f5934aa31b71d3ba351c840fcfbbd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 300, + 71, + 462, + 160 + ], + "blocks": [ + { + "bbox": [ + 170, + 162, + 268, + 171 + ], + "lines": [ + { + "bbox": [ + 170, + 162, + 268, + 171 + ], + "spans": [ + { + "bbox": [ + 170, + 162, + 268, + 171 + ], + "type": "text", + "content": "(b) Action Recognition: NTU-Inter." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 300, + 71, + 462, + 160 + ], + "lines": [ + { + "bbox": [ + 300, + 71, + 462, + 160 + ], + "spans": [ + { + "bbox": [ + 300, + 71, + 462, + 160 + ], + "type": "table", + "html": "
MethodsOverall mAPPerson MovementObject ManipulationPerson Interaction
ST-TR [42]17.337.38.418.8
MS-G3D [36]18.438.79.619.7
CTR-GCN [7]18.939.49.820.5
ST-GCN++ [15]18.639.19.720.0
UNIK [65]19.039.79.920.4
PoseC3D [16]18.339.39.319.1
SkeleTR-S22.843.812.326.9
SkeleTR-S*26.848.217.727.8
SkeleTR [C]*27.350.917.328.3
", + "image_path": "1fc1872797ddb842aecb2ac479c5cce79e57cc32c3f7359c997c45ed323f17d7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 467, + 72, + 545, + 153 + ], + "blocks": [ + { + "bbox": [ + 313, + 162, + 444, + 171 + ], + "lines": [ + { + "bbox": [ + 313, + 162, + 444, + 171 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 444, + 171 + ], + "type": "text", + "content": "(c) Spatio-temporal Action Detection: AVA v2.2." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 467, + 72, + 545, + 153 + ], + "lines": [ + { + "bbox": [ + 467, + 72, + 545, + 153 + ], + "spans": [ + { + "bbox": [ + 467, + 72, + 545, + 153 + ], + "type": "table", + "html": "
MethodsVolleyball Top-1
AT+ [20]92.3
POGARS [52]93.2
COMPOSER [72]93.7
ST-TR [42]93.3
CTR-GCN [7]93.6
ST-GCN++ [15]93.3
SkeleTR-S94.4
SkeleTR [C]94.4
", + "image_path": "ae858a17cccd0531a041cd0b2a15971973ac6cee21d57dfbbe43ac4f38a9159a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 220, + 287, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 287, + 293 + ], + "type": "text", + "content": "tion labels for trimmed videos. For AVA untrimmed videos, in inference we use a sliding window strategy to generate predictions, with a 1-second time step. We get an action prediction for a human proposal from each window containing the proposal, and then average all predictions associated with a human proposal to get the final result." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 300, + 132, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 132, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 132, + 312 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 319, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 378 + ], + "type": "text", + "content": "We compare SkeleTR with state-of-the-art methods on three different skeleton action understanding tasks: action classification, spatio-temporal action detection, and group activity recognition. Unless otherwise noted, all methods use the same skeleton input as SkeleTR." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 379, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 594 + ], + "type": "text", + "content": "Action Classification. Table 1a compares SkeleTR with state-of-the-art methods on Poetics. We use official training configurations when available [36, 16, 65]. For methods without official configurations [7, 15], we use the same hyper-parameter setting as SkeleTR. SkeleTR with all GCN backbones notably outperforms existing methods. Compared to CTR-GCN, SkeleTR-L achieves " + }, + { + "bbox": [ + 46, + 379, + 287, + 594 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 46, + 379, + 287, + 594 + ], + "type": "text", + "content": " higher Top-1 accuracy. Quantitative analysis shows that the improvement is more significant for multi-person scenarios: actions with " + }, + { + "bbox": [ + 46, + 379, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\geq 5" + }, + { + "bbox": [ + 46, + 379, + 287, + 594 + ], + "type": "text", + "content": " persons per frame have around " + }, + { + "bbox": [ + 46, + 379, + 287, + 594 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 379, + 287, + 594 + ], + "type": "text", + "content": " improvement in mean class Top-1, which is twice as much as the improvement for other actions. SkeleTR also achieves the highest accuracy on Poetics (with OpenPose skeletons) and K400 (with HRNet/OpenPose skeletons). Table 1b shows that on the more specific task of two-person interaction recognition, SkeleTR also outperforms existing interaction recognition methods [41, 38, 40] and skeleton action classification methods by a large margin." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 594, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 666 + ], + "type": "text", + "content": "Spatio-temporal Action Detection. Skeleton-based spatio-temporal action detection is a new setting first proposed in this paper and we use AVA as the primary benchmark. For comparisons, we adapt state-of-the-art skeleton action classification algorithms by taking a single skeleton sequence (with 2 seconds length, centered at a person pro" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 310, + 219, + 545, + 309 + ], + "blocks": [ + { + "bbox": [ + 465, + 155, + 545, + 170 + ], + "lines": [ + { + "bbox": [ + 465, + 155, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 465, + 155, + 545, + 170 + ], + "type": "text", + "content": "(d) Group Activity Recognition: Volleyball." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 46, + 173, + 546, + 217 + ], + "lines": [ + { + "bbox": [ + 46, + 173, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 173, + 546, + 217 + ], + "type": "text", + "content": "Table 1: Comparing SkeleTR with previous state-of-the-art on versatile skeleton action recognition tasks. [C] denotes CTR-GCN backbone; * denotes joint training with Poetics. For Poetics and AVA, we follow the common practice [7, 65, 16] and compare the two-stream performance. For NTU-Inter and Volleyball, we only compare the joint modality performance as done in the prior works [52, 72]. In Table 1d, the method " + }, + { + "bbox": [ + 46, + 173, + 546, + 217 + ], + "type": "inline_equation", + "content": "\\mathrm{AT^{+}}" + }, + { + "bbox": [ + 46, + 173, + 546, + 217 + ], + "type": "text", + "content": " uses groundtruth bboxes of all video frames for skeleton extraction." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 219, + 545, + 309 + ], + "lines": [ + { + "bbox": [ + 310, + 219, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 310, + 219, + 545, + 309 + ], + "type": "table", + "html": "
MethodsPosetics Top-1Params (M)GFLOPsInference Speed (2080Ti, clip/second))
ST-TR [42]47.456.4426.9297
MS-G3D [36]52.632.8220.8699
CTR-GCN [7]54.101.365.36195
ST-GCN++ [15]52.461.315.44235
UNIK [65]52.523.2812.15156
PoseC3D [16]53.091.9014.6953
SkeleTR-S55.850.812.74336
SkeleTR-L57.903.827.30200
", + "image_path": "bc82804ac0441bc86d05e139458f5e381dc65630a8d988f63d2d72d9766e24ff.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 329, + 311, + 522, + 321 + ], + "lines": [ + { + "bbox": [ + 329, + 311, + 522, + 321 + ], + "spans": [ + { + "bbox": [ + 329, + 311, + 522, + 321 + ], + "type": "text", + "content": "Table 2: Efficiency statistics of models in Table 1a." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 325, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 545, + 493 + ], + "type": "text", + "content": "posal in the key frame) as input to predict the action for this person. We train all models with the same hyper-parameter setting as SkeleTR. Due to the lack of global context and interaction modeling, all existing methods perform poorly on the new task, while SkeleTR significantly outperforms them. Compared to the baseline ST-GCN++, SkeleTR-S improves the overall mAP by " + }, + { + "bbox": [ + 304, + 325, + 545, + 493 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 304, + 325, + 545, + 493 + ], + "type": "text", + "content": ". Among three major action categories, SkeleTR-S makes the largest improvement on person interaction " + }, + { + "bbox": [ + 304, + 325, + 545, + 493 + ], + "type": "inline_equation", + "content": "(7.0\\%)" + }, + { + "bbox": [ + 304, + 325, + 545, + 493 + ], + "type": "text", + "content": ", thanks to its great inter-person modeling capability. When jointly trained with Poetics, SkeleTR-S enjoys another " + }, + { + "bbox": [ + 304, + 325, + 545, + 493 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 304, + 325, + 545, + 493 + ], + "type": "text", + "content": " overall mAP improvement. However, this time the improvement is basically for person movement and object manipulation, since AVA-style person interactions rarely appear in Kinetics videos." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 495, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 639 + ], + "type": "text", + "content": "Group Activity Recognition. Most prior works on skeleton-based group activity recognition do not utilize the effective GCN architecture to model intra-person motion patterns. Instead, they directly model the relationship among skeletons with complicated transformer architectures. For a comprehensive comparison, we also adapt the state-of-the-art skeleton action recognition algorithms and benchmark them on the Volleyball dataset. Table 1d shows that, with good capability of both intra-sequence motion modeling and inter-sequence relationship modeling, SkeleTR outperforms all previous state-of-the-art on Volleyball, including the adapted GCN-based approaches." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Computational Efficiency. SkeleTR is designed to be efficient and lightweight. Table 2 shows that SkeleTR-S outperforms previous state-of-the-art methods for skeleton-based action classification in terms of accuracy, while having the model with the smallest parameter size, GFLOPs, and fastest inference speed. In the appendix, we provide ad" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "type": "text", + "content": "2Joint and Bone. Each stream takes a separate modality derived from skeleton coordinates as input. For joint stream, the input is the coordinates itself. For bone stream it is the difference of neighboring joints. 3Please see the appendix for these results." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "text", + "content": "13639" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 287, + 143 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 287, + 143 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 287, + 143 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 287, + 143 + ], + "type": "table", + "html": "
PoseticsAVA v2.2
MethodsTop-1Top-5Overall mAPPerson MovementObject ManipulationPerson Interaction
RGB74.4291.2931.3048.4423.6732.74
ST-GCN++52.4675.9718.6239.109.6819.98
RGB + ST-GCN++75.9891.5831.9552.2823.1433.13
SkeleTR57.9080.3326.8348.1917.7227.75
RGB + SkeleTR76.7092.0234.3454.7825.7734.93
", + "image_path": "23c777619f48d6e785d9db6863a2e70da3858bec9d46bbaf265a70f2a6a4877a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 62, + 191, + 274, + 258 + ], + "blocks": [ + { + "bbox": [ + 47, + 144, + 287, + 189 + ], + "lines": [ + { + "bbox": [ + 47, + 144, + 287, + 189 + ], + "spans": [ + { + "bbox": [ + 47, + 144, + 287, + 189 + ], + "type": "text", + "content": "Table 3: SkeleTR demonstrates great complementarity when fused with RGB recognizers. We fuse predictions by state-of-the-art RGB recognizers (TimeSformer [2] for Posetics, ACAR-R101 [39] for AVA) with skeleton predictions with " + }, + { + "bbox": [ + 47, + 144, + 287, + 189 + ], + "type": "inline_equation", + "content": "2:1" + }, + { + "bbox": [ + 47, + 144, + 287, + 189 + ], + "type": "text", + "content": " ratio." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 191, + 274, + 258 + ], + "lines": [ + { + "bbox": [ + 62, + 191, + 274, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 274, + 258 + ], + "type": "table", + "html": "
Interaction ModelingPosetics Top-1AVA Overall mAP
NA48.2117.56
Add Global51.0620.00
Concat Global50.0120.14
Non-Local48.4319.78
Transformer Encoders52.8120.92
", + "image_path": "9d6cfb891329ed3768f4cdcfdcf2e9bab68409720107bcd1fee2cfa771cd2de2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 61, + 284, + 274, + 381 + ], + "blocks": [ + { + "bbox": [ + 47, + 259, + 286, + 281 + ], + "lines": [ + { + "bbox": [ + 47, + 259, + 286, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 286, + 281 + ], + "type": "text", + "content": "Table 4: Transformer-based interaction modeling significantly outperforms multiple baselines." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 284, + 274, + 381 + ], + "lines": [ + { + "bbox": [ + 61, + 284, + 274, + 381 + ], + "spans": [ + { + "bbox": [ + 61, + 284, + 274, + 381 + ], + "type": "table", + "html": "
Feature AggregationPosetics Top-1AVA Overall mAPHead Computation (GFLOPs)
1 × 152.8120.920.013
T × P52.7821.505.211
T × 153.4820.640.356
1 × P53.4021.350.080
1 × V52.1521.010.429
1 × 1 + T × 153.7521.000.392
1 × 1 + 1 × P53.0021.430.101
T × 1 + 1 × P53.1620.870.550
1 × 1 + T × 1 + 1 × P53.9021.650.594
", + "image_path": "96ebd670c653a5bd5f784e495ca0248772fdb08b35010629e2e0cc3ecb3ca1bc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 50, + 384, + 283, + 396 + ], + "lines": [ + { + "bbox": [ + 50, + 384, + 283, + 396 + ], + "spans": [ + { + "bbox": [ + 50, + 384, + 283, + 396 + ], + "type": "text", + "content": "Table 5: Comparing different feature aggregation strategies." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 398, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 287, + 434 + ], + "type": "text", + "content": "ditional evidence of SkeleTR's computational efficiency by comparing it to other methods designed for different tasks, such as IGFormer [40] and COMPOSER [72]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 434, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 555 + ], + "type": "text", + "content": "Improved complementarity with RGB modality. To demonstrate the complementarity of SkeleTR upon RGB methods, we directly fuse the prediction of SkeleTR and the state-of-the-art RGB-based action recognizers. Table 3 shows that late fusion with SkeleTR leads to a significant boost (2.3% Posetics top-1, 3% AVA mAP) compared to the RGB state-of-the-art, which notably outperforms ST-GCN++ late fusion. On AVA, the improvement is extremely large for person movement (6.3% mAP), which emphasizes the importance of the skeleton in modeling human motions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 563, + 247, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 247, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 247, + 576 + ], + "type": "text", + "content": "4.4. Ablation on Inter-Sequence Modeling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "text", + "content": "We conduct ablation experiments to validate the efficacy of our inter-sequence modeling stage. In experiments, we adopt the light ST-GCN++ as the backbone. We set " + }, + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "text", + "content": " to 20, " + }, + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "text", + "content": " to 30, and set the temporal window size to 5s for AVA. We report the recognition performance with a single joint modality on Posetics and AVA." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": "We first compare the Transformer-based modeling with other simple baselines for modeling global context or interaction. In all experiments, we pool both spatial and temporal dimensions and retain a single feature vector per sequence. We consider the following baselines: 1. NA: no" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 313, + 66, + 536, + 171 + ], + "blocks": [ + { + "bbox": [ + 313, + 66, + 536, + 171 + ], + "lines": [ + { + "bbox": [ + 313, + 66, + 536, + 171 + ], + "spans": [ + { + "bbox": [ + 313, + 66, + 536, + 171 + ], + "type": "image", + "image_path": "f3a2e78b803c0ced29c4e133a3cbd36555db4d785728b9355fda555da41689ee.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 171, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 305, + 171, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 305, + 171, + 545, + 205 + ], + "type": "text", + "content": "Figure 5: Visualization of the weights in different attention heads in the last transformer encoder. " + }, + { + "bbox": [ + 305, + 171, + 545, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{ij}" + }, + { + "bbox": [ + 305, + 171, + 545, + 205 + ], + "type": "text", + "content": " denotes the short skeleton sequence of person " + }, + { + "bbox": [ + 305, + 171, + 545, + 205 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 171, + 545, + 205 + ], + "type": "text", + "content": " at the timestamp " + }, + { + "bbox": [ + 305, + 171, + 545, + 205 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 171, + 545, + 205 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 207, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 545, + 338 + ], + "type": "text", + "content": "interaction modeling; 2. Add Global: process the average feature with an MLP to get a Global feature, then add it back to each sequence feature; 3. Concat Global: concatenate the average feature with each sequence feature, then process with an MLP; 4. Non-Local: process all sequence features with a standard non-local [59] block, and apply an MLP to the output. Table 4 shows that, despite its low computation cost (0.013 GFLOPs / clip), the Transformer markedly enhances recognition performance relatively to the NA baseline (4.6% Posetics top-1, 3.4% AVA mAP), and considerably outperforms all alternatives." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 338, + 546, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 546, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 546, + 517 + ], + "type": "text", + "content": "Both Non-Local and Transformer encoders adopt selfattention to model the inter-sequence relationship, while there is a significant difference in recognition performance between them. The reason is that the multi-head design in Transformers allows for modeling diverse relationships between individual sequences. Visualizing the attention weights of different attention heads, we find that some of them have clearly interpretable semantics. Fig 5 shows the attention weights of a 'handshake' action. Of the eight attention heads, the first head aggregates temporal features within a single person, while the fourth head models interactions between individuals. Specifically, the attention weights of the fourth head reveal that " + }, + { + "bbox": [ + 304, + 338, + 546, + 517 + ], + "type": "inline_equation", + "content": "P_{21}" + }, + { + "bbox": [ + 304, + 338, + 546, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 338, + 546, + 517 + ], + "type": "inline_equation", + "content": "P_{22}" + }, + { + "bbox": [ + 304, + 338, + 546, + 517 + ], + "type": "text", + "content": " attend most strongly to other skeleton features, which corresponds to the exact moment of the handshaking interaction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 518, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 615 + ], + "type": "text", + "content": "We then delve deeper into the Mix Pooling design, which helps to keep fine-grained sequence features and serves as a good trade-off between efficiency and performance. Table 5 compares Mix Pooling " + }, + { + "bbox": [ + 304, + 518, + 545, + 615 + ], + "type": "inline_equation", + "content": "(1 \\times 1 + T \\times 1 + 1 \\times P)" + }, + { + "bbox": [ + 304, + 518, + 545, + 615 + ], + "type": "text", + "content": " with other feature aggregation strategies. Mix Pooling outperforms each individual pooling strategy or the combination of two strategies. It achieves the best performance on both datasets with a reasonable computation cost of 0.594 GFLOPs per clip." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 623, + 477, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 477, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 477, + 635 + ], + "type": "text", + "content": "4.5. Ablation on Input Construction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": "In this section we demonstrate the advantages of our new input format, both in recognition performance and computational efficiency. We report the recognition performance with a single joint modality." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Large Number of Short Sequences. We use a large number of short skeleton sequences as input for skeleton action" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "text", + "content": "13640" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 67, + 272, + 176 + ], + "blocks": [ + { + "bbox": [ + 61, + 67, + 272, + 176 + ], + "lines": [ + { + "bbox": [ + 61, + 67, + 272, + 176 + ], + "spans": [ + { + "bbox": [ + 61, + 67, + 272, + 176 + ], + "type": "image", + "image_path": "71cca6e19234087cb9f46a241a258b42637d1c1c0b68cf026b54710c66e060d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 178, + 287, + 200 + ], + "lines": [ + { + "bbox": [ + 47, + 178, + 287, + 200 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 287, + 200 + ], + "type": "text", + "content": "Figure 6: Fix the total number of skeletons and adjust the sequence number. Posetics Top-1 accuracy is visualized." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 204, + 274, + 282 + ], + "blocks": [ + { + "bbox": [ + 61, + 204, + 274, + 282 + ], + "lines": [ + { + "bbox": [ + 61, + 204, + 274, + 282 + ], + "spans": [ + { + "bbox": [ + 61, + 204, + 274, + 282 + ], + "type": "table", + "html": "
MethodsMTPosetics Top-1Avg GFLOPs for \nPosetics Video Clips
SkeleTR-S203053.902.430
SkeleTR-L[5, 20]3055.064.934
SkeleTR-L203054.956.653
SkeleTR-S401554.172.789
SkeleTR-L[10, 40]1555.353.532
SkeleTR-L401555.557.394
", + "image_path": "9cfc97a2cb2f89330f2ad113f250d2c9444f3ec54e75c5c5c1547cb10d843d99.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 324, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 467 + ], + "type": "text", + "content": "understanding tasks. Besides being a unified input format for different tasks, the new format also benefits standard skeleton-based action classification under real-world scenarios. To investigate the effect of sequence granularity, we fix the total number of skeletons " + }, + { + "bbox": [ + 46, + 324, + 287, + 467 + ], + "type": "inline_equation", + "content": "M \\times T" + }, + { + "bbox": [ + 46, + 324, + 287, + 467 + ], + "type": "text", + "content": ", and adjust the sequence number " + }, + { + "bbox": [ + 46, + 324, + 287, + 467 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 324, + 287, + 467 + ], + "type": "text", + "content": ". We compare three models: ST-GCN++, SkeleTR-S, and SkeleTR-L (Fig 6). Interestingly, we find that all models obtain their highest performance when using a large number of short sequences as input, specifically, " + }, + { + "bbox": [ + 46, + 324, + 287, + 467 + ], + "type": "inline_equation", + "content": "40 \\times 15" + }, + { + "bbox": [ + 46, + 324, + 287, + 467 + ], + "type": "text", + "content": ". Meanwhile, the improvement is especially large for SkeleTR: " + }, + { + "bbox": [ + 46, + 324, + 287, + 467 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 46, + 324, + 287, + 467 + ], + "type": "text", + "content": " top-1 for SkeleTR-S, " + }, + { + "bbox": [ + 46, + 324, + 287, + 467 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 46, + 324, + 287, + 467 + ], + "type": "text", + "content": " top-1 for SkeleTR-L (compared to the " + }, + { + "bbox": [ + 46, + 324, + 287, + 467 + ], + "type": "inline_equation", + "content": "2 \\times 300" + }, + { + "bbox": [ + 46, + 324, + 287, + 467 + ], + "type": "text", + "content": " setting)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "text", + "content": "Flexible Number of Sequences. SkeleTR is compatible with different numbers of sequences. To train and test SkeleTR with an adaptive squence number, we first set a dynamic range " + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "inline_equation", + "content": "[M_{min}, M_{max}]" + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "text", + "content": ", and for each training batch, we randomly sample " + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "inline_equation", + "content": "\\tilde{M}" + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "text", + "content": " from the range and use it as the sequence number for all videos in the batch. For each testing video, we use the actual sequence number " + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "inline_equation", + "content": "M'" + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "inline_equation", + "content": "M_{min} \\leq M' \\leq M_{max}" + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "text", + "content": ", and cap it to " + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "inline_equation", + "content": "M_{min}" + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "inline_equation", + "content": "M_{max}" + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "text", + "content": " otherwise. We present an ablation study using two input formats: " + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "inline_equation", + "content": "20 \\times 30" + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "inline_equation", + "content": "40 \\times 15" + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "text", + "content": ". For each format, we use a dynamic range ([5, 20], [10, 40]) of " + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "text", + "content": " to train SkeleTR-L on Poetics, and report the results in Table 6. Compared to using a fixed " + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "text", + "content": ", using an adaptive " + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "text", + "content": " can achieve comparable recognition performance with a largely reduced inference cost (26% less GFLOPs for " + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "inline_equation", + "content": "T = 30" + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "text", + "content": ", 52% less for " + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "inline_equation", + "content": "T = 15" + }, + { + "bbox": [ + 46, + 468, + 288, + 649 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 658, + 249, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 249, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 249, + 671 + ], + "type": "text", + "content": "4.6. Transfer Learning and Joint Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "SkeleTR is a general framework for versatile skeleton action recognition tasks, which enables transfer learning and joint learning across different tasks. To demonstrate this," + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 321, + 69, + 533, + 150 + ], + "blocks": [ + { + "bbox": [ + 47, + 285, + 287, + 319 + ], + "lines": [ + { + "bbox": [ + 47, + 285, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 287, + 319 + ], + "type": "text", + "content": "Table 6: Sequence number " + }, + { + "bbox": [ + 47, + 285, + 287, + 319 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 285, + 287, + 319 + ], + "type": "text", + "content": " : fixed or adaptive? Utilizing an adaptive sequence number " + }, + { + "bbox": [ + 47, + 285, + 287, + 319 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 285, + 287, + 319 + ], + "type": "text", + "content": " can save great amounts of computational resources and achieve comparable performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 69, + 533, + 150 + ], + "lines": [ + { + "bbox": [ + 321, + 69, + 533, + 150 + ], + "spans": [ + { + "bbox": [ + 321, + 69, + 533, + 150 + ], + "type": "table", + "html": "
TrainingTemporal WindowOverall mAPPerson Movement mAPObject Manipulation mAPPerson Interaction mAP
From Scratch521.6542.3411.3725.66
Finetuning522.1243.0211.9125.76
Joint Training523.9244.2514.6526.06
From Scratch1021.3342.2411.2624.68
Finetuning1023.1144.0613.3225.85
Joint Training1025.6947.1316.5326.64
", + "image_path": "dcb059b5903146c79ed56b80d4fad8633e471d33b3ad31c18f9320d59a2acd0a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 153, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 304, + 153, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 304, + 153, + 545, + 198 + ], + "type": "text", + "content": "Table 7: Transfer Learning and Joint Learning with SkeleTR-S. For a target task with limited training data, an auxiliary dataset can be leveraged to improve the performance and mitigate the overfitting, either by transfer learning or joint learning." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 201, + 545, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 545, + 309 + ], + "type": "text", + "content": "we consider two tasks: Poetics action classification and AVA spatio-temporal action detection (with AVA being the smaller dataset). From-scratch training on AVA is easy to overfit (Fig 4). Pretraining on Poetics or joint training on AVA and Poetics both help to mitigate the overfitting, with joint training leading to the best recognition performance. Enlarging the temporal window size for AVA sampling, which leads to less diversified training samples, amplifies the gain of pretraining or joint training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 320, + 383, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 320, + 383, + 332 + ], + "spans": [ + { + "bbox": [ + 306, + 320, + 383, + 332 + ], + "type": "text", + "content": "4.7. Limitations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 340, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 484 + ], + "type": "text", + "content": "SkeleTR is currently trained with skeletons pre-extracted from RGB videos, and the adopted pose estimation solution is relatively heavy. To use SkeleTR in real-world apps with limited computation resource, additional efforts are required to switch to a light-weight pose estimation solution that can perform fast inference on CPUs or edge devices, and to retrain the models accordingly. Besides, the short sequence format proposed by SkeleTR strikes a good balance between mitigating skeleton matching noises and modeling intra-sequence dynamics in real-world scenarios. The format may not lead to extra benefits when the ground-truth skeleton sequences are available (details in appendix)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 498, + 378, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 378, + 511 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 378, + 511 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 616 + ], + "type": "text", + "content": "We presented SkeleTR, a unified framework for skeleton action recognition in general scenarios. It extends existing GCN backbones and achieves state-of-the-art performance on multiple skeleton action understanding tasks, including video-level, instance-level, and group-level action recognition. Besides, SkeleTR is compatible with transfer learning and joint learning across different datasets and tasks, which further improves the capability of learned representations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": "Broader Impact. Using human skeletons for action recognition is more robust to environment changes and can help preserve user privacy. Real-world applications of skeleton action recognition include monitoring patient or elderly health, enabling autonomous driving, and facilitating human-computer interactions. However, there could be unintended usages and we advocate responsible usage complying with applicable laws and regulations." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "text", + "content": "13641" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "type": "text", + "content": "[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 2, 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 286, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 286, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 286, + 192 + ], + "type": "text", + "content": "[3] Erik Bochinski, Volker Eiselein, and Thomas Sikora. High-speed tracking-by-detection without using image information. In AVSS, 2017. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 194, + 286, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 286, + 227 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 286, + 227 + ], + "type": "text", + "content": "[4] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. TPAMI, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 228, + 286, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 286, + 261 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 286, + 261 + ], + "type": "text", + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 262, + 286, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 286, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 286, + 294 + ], + "type": "text", + "content": "[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. 2, 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 297, + 286, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 286, + 340 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 286, + 340 + ], + "type": "text", + "content": "[7] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In ICCV, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 342, + 286, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 286, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 286, + 385 + ], + "type": "text", + "content": "[8] Hyung-gun Chi, Myoung Hoon Ha, Seunggeun Chi, Sang Wan Lee, Qixing Huang, and Karthik Ramani. Infogcn: Representation learning for human skeleton-based action recognition. In CVPR, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 387, + 286, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 387, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 387, + 286, + 430 + ], + "type": "text", + "content": "[9] Wongun Choi, Khuram Shahid, and Silvio Savarese. What are they doing?: Collective activity classification using spatio-temporal relationship among people. In ICCVW, 2009. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 433, + 286, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 286, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 286, + 464 + ], + "type": "text", + "content": "[10] Vasileios Choutas, Philippe Weinzaepfel, Jérôme Revaud, and Cordelia Schmid. Potion: Pose motion representation for action recognition. In CVPR, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 467, + 286, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 286, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 286, + 510 + ], + "type": "text", + "content": "[11] Liu Chunhui, Hu Yueyu, Li Yanghao, Song Sijie, and Liu Jiaying. Pku-mmd: A large scale benchmark for continuous multi-modal human action understanding. arXiv:1703.07475, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 512, + 286, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 286, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 286, + 566 + ], + "type": "text", + "content": "[12] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv:2003.09003, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 568, + 286, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 286, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 286, + 610 + ], + "type": "text", + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv:1810.04805, 2018.3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 286, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 677 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv:2010.11929, 2020.3,4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "type": "text", + "content": "[15] Haodong Duan, Jiaqi Wang, Kai Chen, and Dahua Lin. Pyskl: Towards good practices for skeleton action recognition. ACMM, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Haodong Duan, Yue Zhao, Kai Chen, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. In CVPR, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[17] Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[18] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "text", + "content": "[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "text", + "content": "[20] Kirill Gavrilyuk, Ryan Sanford, Mehrsan Javan, and Cees GM Snoek. Actor-transformers for group activity recognition. In CVPR, 2020. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 243, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 308 + ], + "type": "text", + "content": "[21] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In ICCV, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 310, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 364 + ], + "type": "text", + "content": "[22] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, 2018. 1, 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "text", + "content": "[23] Mohamed E Hussein, Marwan Torki, Mohammad A Gowayyed, and Motaz El-Saban. Human action recognition using a temporal hierarchy of covariance descriptors on 3d joint locations. In IJCAI, 2013. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 411, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 453 + ], + "type": "text", + "content": "[24] Mostafa S Ibrahim, Srikanth Muralidharan, Zhiwei Deng, Arash Vahdat, and Greg Mori. A hierarchical deep temporal model for group activity recognition. In CVPR, 2016. 1, 2, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "type": "text", + "content": "[25] Haroon Idrees, Amir R Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The thumos challenge on action recognition for videos \"in the wild\". CVIU, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "type": "text", + "content": "[26] Yanli Ji, Guo Ye, and Hong Cheng. Interactive body part contrast mining for human interaction recognition. In 2014 IEEE international conference on multimedia and expo workshops (ICMEW), pages 1-6. IEEE, 2014. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "text", + "content": "[27] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal-directed human activities. In CVPR, 2014. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 580, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 545, + 622 + ], + "type": "text", + "content": "[28] Ang Li, Meghanana Thotakuri, David A Ross, João Carreira, Alexander Vostrikov, and Andrew Zisserman. Theava-kinetics localized human actions video dataset. arXiv:2005.00214, 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "type": "text", + "content": "[29] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In CVPR, 2019. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[30] Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang. Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In ICCV, 2021. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "type": "text", + "content": "13642" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[31] Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[32] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+d 120: A large-scale benchmark for 3d human activity understanding. TPAMI, 2020. 1, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "text", + "content": "[33] Jun Liu, Amir Shahroudy, Dong Xu, and Gang Wang. Spatio-temporal LSTM with trust gates for 3d human action recognition. In ECCV, 2016. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 228 + ], + "type": "text", + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "type": "text", + "content": "[35] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "text", + "content": "[36] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In CVPR, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 309, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 365 + ], + "type": "text", + "content": "[37] Qianhui Men, Edmond SL Ho, Hubert PH Shum, and Howard Leung. A two-stream recurrent network for skeleton-based human interaction recognition. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 2771-2778. IEEE, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 365, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 421 + ], + "type": "text", + "content": "[38] Xuan Son Nguyen. Geomnet: A neural network based on riemannian geometries of spd matrix space and cholesky space for 3d skeleton-based interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13379-13389, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "text", + "content": "[39] Junting Pan, Siyu Chen, Mike Zheng Shou, Yu Liu, Jing Shao, and Hongsheng Li. Actor-context-actor relation network for spatio-temporal action localization. In CVPR, 2021. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 466, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 509 + ], + "type": "text", + "content": "[40] Yunsheng Pang, Qiuhong Ke, Hossein Rahmani, James Bailey, and Jun Liu. Igformer: Interaction graph transformer for skeleton-based human interaction recognition. In ECCV, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 543 + ], + "type": "text", + "content": "[41] Mauricio Perez, Jun Liu, and Alex C Kot. Interaction relational network for mutual action recognition. IEEE Transactions on Multimedia, 24:366-376, 2021. 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "type": "text", + "content": "[42] Chiara Plizzari, Marco Cannici, and Matteo Matteucci. Skeleton-based action recognition via spatial and temporal transformer networks. CVIU, 2021. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "text", + "content": "[43] Michalis Raptis and Leonid Sigal. Poselet key-framing: A model for human activity recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2650-2657, 2013. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 656 + ], + "type": "text", + "content": "[44] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 2015. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 679 + ], + "type": "text", + "content": "[45] Kohei Sendo and Norimichi Ukita. Heatmapping of people involved in group activities. In MVA, 2019. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[46] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+d: A large scale dataset for 3d human activity analysis. In CVPR, 2016. 1, 2, 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[47] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In CVPR, 2019. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[48] Bing Shuai, Alessandro Bergamo, Uta Buechler, Andrew Berneshawi, Alyssa Boden, and Joseph Tighe. Large scale real-world multi-person tracking. ECCV, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[49] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. NeurIPS, 2014. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "text", + "content": "[50] Sebastian Stein and Stephen J McKenna. Combining embedded accelerometers with computer vision for recognizing food preparation activities. In UbiComp, 2013. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 209, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 240 + ], + "type": "text", + "content": "[51] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In CVPR, 2019. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "text", + "content": "[52] Haritha Thilakarathne, Aiden Nibali, Zhen He, and Stuart Morgan. Pose is all you need: The pose only group activity recognition system (pogars). Machine Vision and Applications, 2022. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 287, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 341 + ], + "type": "text", + "content": "[53] Arash Vahdat, Bo Gao, Mani Ranjbar, and Greg Mori. A discriminative key pose sequence model for recognizing human interactions. In 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops), pages 1729-1736. IEEE, 2011. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "type": "text", + "content": "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "type": "text", + "content": "[55] Raviteja Vemulapalli, Felipe Arrate, and Rama Chellappa. Human action recognition by representing 3d skeletons as points in a lie group. In CVPR, 2014. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "type": "text", + "content": "[56] Hongsong Wang and Liang Wang. Modeling temporal dynamics and spatial configurations of actions using two-stream recurrent neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 499-508, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 466, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 498 + ], + "type": "text", + "content": "[57] Jiang Wang, Zicheng Liu, Ying Wu, and Junsong Yuan. Mining actionlet ensemble for action recognition with depth cameras. In CVPR. IEEE, 2012. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 500, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 543 + ], + "type": "text", + "content": "[58] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In ECCV, 2016. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 544, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 567 + ], + "type": "text", + "content": "[59] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, 2018. 2, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 568, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 610 + ], + "type": "text", + "content": "[60] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In CVPR, 2019. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "type": "text", + "content": "[61] Fanyi Xiao, Yong Jae Lee, Kristen Grauman, Jitendra Malik, and Christoph Feichtenhofer. Audiovisual slowfast networks for video recognition. arXiv:2001.08740, 2020. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "text", + "content": "[62] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017. 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[63] Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Xia, Zhuowen Tu, and Stefano Soatto. Long short-term transformer for online action detection. NeurIPS, 2021. 3" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "text", + "content": "13643" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 554 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[64] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 1, 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "type": "text", + "content": "[65] Di Yang, Yaohui Wang, Antitza Dantcheva, Lorenzo Garattoni, Gianpiero Francesca, and François Bremond. Unik: A unified framework for real-world skeleton-based action recognition. ArXiv, 2021. 2, 4, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 217 + ], + "type": "text", + "content": "[66] Kiwon Yun, Jean Honorio, Debaleena Chattopadhyay, Tamara L Berg, and Dimitris Samaras. Two-person interaction detection using body-pose features and multiple instance learning. In 2012 IEEE computer society conference on computer vision and pattern recognition workshops, pages 28–35. IEEE, 2012. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 286, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 286, + 284 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 286, + 284 + ], + "type": "text", + "content": "[67] Yimeng Zhang, Xiaoming Liu, Ming-Ching Chang, Weina Ge, and Tsuhan Chen. Spatio-temporal phrases for activity recognition. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part III 12, pages 707-721. Springer, 2012. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 285, + 286, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 285, + 286, + 318 + ], + "spans": [ + { + "bbox": [ + 49, + 285, + 286, + 318 + ], + "type": "text", + "content": "[68] Yuhan Zhang, Bo Wu, Wen Li, Lixin Duan, and Chuang Gan. Stst: Spatial-temporal specialized transformer for skeleton-based action recognition. In ACMMM, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 319, + 286, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 319, + 286, + 341 + ], + "spans": [ + { + "bbox": [ + 49, + 319, + 286, + 341 + ], + "type": "text", + "content": "[69] Zhengyou Zhang. Microsoft Kinect sensor and its effect. IEEE multimedia, 2012. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 342, + 286, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 342, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 49, + 342, + 286, + 384 + ], + "type": "text", + "content": "[70] Hang Zhao, Antonio Torralba, Lorenzo Torresani, and Zhicheng Yan. Hacs: Human action clips and segments dataset for recognition and temporal localization. In ICCV, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 387, + 286, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 387, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 387, + 286, + 430 + ], + "type": "text", + "content": "[71] Jiaojiao Zhao, Yanyi Zhang, Xinyu Li, Hao Chen, Bing Shuai, Mingze Xu, Chunhui Liu, Kaustav Kundu, Yuanjun Xiong, Davide Modolo, et al. Tuber: Tubelet transformer for video action detection. In CVPR, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 432, + 286, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 432, + 286, + 475 + ], + "spans": [ + { + "bbox": [ + 49, + 432, + 286, + 475 + ], + "type": "text", + "content": "[72] Honglu Zhou, Asim Kadav, Aviv Shamsian, Shijie Geng, Farley Lai, Long Zhao, Ting Liu, Mubbasir Kapadia, and Hans Peter Graf.Composer: Compositional learning of group activity in videos.ECCV, 2022.5,6,7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 476, + 286, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 476, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 49, + 476, + 286, + 520 + ], + "type": "text", + "content": "[73] Wentao Zhu, Cuiling Lan, Junliang Xing, Wenjun Zeng, Yanghao Li, Li Shen, and Xiaohui Xie. Co-occurrence feature learning for skeleton based action recognition using regularized deep LSTM networks. In AAAI, 2016. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 522, + 286, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 522, + 286, + 554 + ], + "spans": [ + { + "bbox": [ + 49, + 522, + 286, + 554 + ], + "type": "text", + "content": "[74] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv:2010.04159, 2020. 3" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "text", + "content": "13644" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SkeletonMAE_ Graph-based Masked Autoencoder for Skeleton Sequence Pre-training/dcc19c04-02ae-4f8d-8f44-a86cdd91e8a4_content_list.json b/2023/SkeletonMAE_ Graph-based Masked Autoencoder for Skeleton Sequence Pre-training/dcc19c04-02ae-4f8d-8f44-a86cdd91e8a4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f3c2fdc63e3f97bd2af8074dc96a525d9e425a4b --- /dev/null +++ b/2023/SkeletonMAE_ Graph-based Masked Autoencoder for Skeleton Sequence Pre-training/dcc19c04-02ae-4f8d-8f44-a86cdd91e8a4_content_list.json @@ -0,0 +1,1746 @@ +[ + { + "type": "text", + "text": "SkeletonMAE: Graph-based Masked Autoencoder for Skeleton Sequence Pre-training", + "text_level": 1, + "bbox": [ + 105, + 150, + 863, + 195 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hong Yan $^{1*}$ Yang Liu $^{1*†}$ Yushen Wei $^{1}$ Zhen Li $^{2}$ Guanbin Li $^{1}$ Liang Lin $^{1}$ $^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China \n $^{2}$ The Chinese University of Hong Kong, Shenzhen, China", + "bbox": [ + 117, + 219, + 853, + 271 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yanh36,weiysh8}@mail2.sysu.edu.cn，{liuy856,liguanbin}@mail.sysu.edu.cn, lizhen@cuhk.edu.cn，linliang@ieee.org", + "bbox": [ + 154, + 273, + 811, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 226, + 336, + 307, + 351 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Skeleton sequence representation learning has shown great advantages for action recognition due to its promising ability to model human joints and topology. However, the current methods usually require sufficient labeled data for training computationally expensive models. Moreover, these methods ignore how to utilize the fine-grained dependencies among different skeleton joints to pre-train an efficient skeleton sequence learning model that can generalize well across different datasets. In this paper, we propose an efficient skeleton sequence learning framework, named Skeleton Sequence Learning (SSL). To comprehensively capture the human pose and obtain discriminative skeleton sequence representation, we build an asymmetric graph-based encoder-decoder pre-training architecture named SkeletonMAE, which embeds skeleton joint sequence into graph convolutional network and reconstructs the masked skeleton joints and edges based on the prior human topology knowledge. Then, the pre-trained SkeletonMAE encoder is integrated with the Spatial-Temporal Representation Learning (STRL) module to build the SSL framework. Extensive experimental results show that our SSL generalizes well across different datasets and outperforms the state-of-the-art self-supervised skeleton-based methods on FineGym, Diving48, NTU 60 and NTU 120 datasets. Moreover, we obtain comparable performance to some fully supervised methods. The code is available at https://github.com/HongYan1123/SkeletonMAE.", + "bbox": [ + 63, + 366, + 473, + 751 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 65, + 777, + 200, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human action recognition has attracted increasing attention in video understanding [83, 7, 45, 60, 39], due to its wide applications [2, 14, 61, 51, 52, 86, 44] in human-", + "bbox": [ + 65, + 801, + 468, + 844 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4fb8b97be0a14dbcf1dba2639876c35f5b1f977c3ad77152bd645c9d1b44440d.jpg", + "image_caption": [ + "Figure 1: Traditional MAE usually uses random masking strategy to reconstruct skeleton, which tends to ignore action-sensitive skeleton regions. Differently, our proposed SkeletonMAE reconstructs the masked skeleton joints and edges based on the prior human topology knowledge, to obtain a comprehensive perception of the action." + ], + "image_footnote": [], + "bbox": [ + 512, + 334, + 890, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "computer interaction, intelligent surveillance security, virtual reality, etc. In terms of visual perception [24], even without appearance information, humans can identify action categories by only observing the motion of joints. Unlike RGB videos [3, 14, 13], the skeleton sequences only contain the coordinate information of the key joints of the human body [84], which is high-level, light-weighted, and robust against complex backgrounds and various conditions including viewpoint, scale, and movement speed [11, 69]. Additionally, with the development of human pose estimation algorithms [8, 1, 85], the localization method of human joints (i.e., key points) has made great progress and it is feasible to obtain accurate skeleton sequences. At present, the existing 2D pose estimation method is more accurate and more robust than the 3D pose estimation methods [11]. In Figure 1 (a), we visualize 2D poses estimated with HRNet [67] for two action classes on FineGym dataset [59]. It", + "bbox": [ + 497, + 635, + 905, + 878 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 119, + 36, + 189, + 73 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 247, + 35, + 796, + 76 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Both authors contributed equally to this work.", + "bbox": [ + 89, + 852, + 347, + 865 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 92, + 865, + 220, + 876 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5606", + "bbox": [ + 468, + 891, + 502, + 903 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "can be seen that the 2D poses can accurately capture human skeletons and motion details.", + "bbox": [ + 63, + 115, + 468, + 142 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the promising ability to model multiple granularities and large variations in human motion, the skeleton sequence is more suitable to distinguish similar actions with subtle differences than the RGB data. To capture discriminative spatial-temporal motion patterns, most of the existing skeleton-based action recognition methods [11, 83, 5] are fully supervised and usually require large amounts of labeled data for training elaborate models, which is time-consuming and labor-intensive. To mitigate the problem of limited labeled training data, self-supervised skeleton-based action recognition methods [31, 18, 65] have attracted increasing attention recently. Some contrastive learning methods [31, 18] adopted data augmentation to generate pairs of positive and negative samples, but they rely heavily on the number of contrastive pairs. With the popularity of the encoder-decoder [64, 48], some methods [91, 65] reconstructed the masked skeleton sequence by link reconstruction to encourage the topological closeness following the paradigm of graph encoder-decoder. However, these methods are usually good at link prediction and node clustering but are unsatisfactory in node and graph classifications. For accurate action recognition, the fine-grained dependencies among different skeleton joints (i.e., graph classifications) are essential. Therefore, previous self-supervised learning-based methods tend to ignore the fine-grained dependencies among different skeleton joints, which restricts the generalization of self-supervised skeleton representation. As shown in Figure 1 (b)-(c), the arm joints and edges are essential to discriminate between these two similar actions. Different from the randomly masking strategy of MAE [20], our masking strategy is action-sensitive and reconstructs specific limbs or body parts that dominate the given action class. Our SkeletonMAE utilizes prior human topology knowledge to guide the reconstruction of the masked skeleton joints and edges to achieve a comprehensive perception of the joints, topology, and action.", + "bbox": [ + 63, + 148, + 472, + 661 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned challenges, we propose an efficient skeleton sequence representation learning framework, named Skeleton Sequence Learning (SSL). To fully discover the fine-grained dependencies among different skeleton joints, we build a novel asymmetric graph-based encoder-decoder pre-training architecture named SkeletonMAE that embeds skeleton joint sequences in Graph Convolutional Network (GCN). The SkeletonMAE aims to reconstruct the masked human skeleton joints and edges based on prior human topology knowledge so that it can infer the underlying topology of the joints and obtain a comprehensive perception of human action. To learn discriminative spatial-temporal skeleton representation, the pre-trained SkeletonMAE encoder is integrated with the Spatial-Temporal Representation Learning (STRL)", + "bbox": [ + 63, + 663, + 468, + 878 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "module to learn spatial-temporal dependencies. Finally, the SSL is fine-tuned on action recognition datasets. Extensive experimental results on FineGym, Diving48, NTU 60 and NTU 120 show that our SSL generalizes well across different datasets and outperforms the state-of-the-art methods significantly. Our contributions are summarized as follows:", + "bbox": [ + 497, + 115, + 904, + 201 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To comprehensively capture human pose and obtain discriminative skeleton sequence representation, we propose a graph-based encoder-decoder pre-training architecture named SkeletonMAE, that embeds skeleton joint sequence into GCN and utilize the prior human topology knowledge to guide the reconstruction of the underlying masked joints and topology.", + "- To learn comprehensive spatial-temporal dependencies for skeleton sequence, we propose an efficient skeleton sequence learning framework, named Skeleton Sequence Learning (SSL), which integrates the pre-trained SkeletonMAE encoder with the Spatial-Temporal Representation Learning (STRL) module.", + "- Extensive experimental results on FineGym, Diving48, NTU 60 and NTU 120 datasets show that our SSL outperforms the state-of-the-art self-supervised skeleton-based action recognition methods and achieves comparable performance compared with the state-of-the-art fully supervised methods." + ], + "bbox": [ + 517, + 214, + 900, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 500, + 645, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Action Recognition. One of the most challenging tasks for action recognition is to distinguish similar actions from subtle differences [40, 41, 42]. Recently, some challenging action recognition datasets like FineGym [59], Diving48 [34], NTU RGB+D 60 [58] and NTU RGB+D 120 [37] are proposed. These datasets contain a large number of challenging actions that require discriminative and fine-grained action representation learning. For example, in FineGym [59], an action is divided into action units, subactions, or phases, and the model is required to distinguish between \"split leap with 1 turn\" and \"switch leap with 1 turn\". The higher inter-class similarity and a new level of granularity in the fine-grained setting make it a challenging task, which makes coarse-grained backbones and methods [14, 2, 71, 79] struggle to overcome. To tackle the more challenging fine-grained action recognition task, most of the existing works [50, 33] are fully supervised and consider fine-grained actions as distinct categories and supervise the model to learn action semantics. However, collecting and labeling these fine-grained actions is time-consuming and labor-intensive, which limits the generalization of a well-trained model to different datasets. To utilize unlabeled data, we propose a graph-based encoder-decoder pre-training architecture named SkeletonMAE.", + "bbox": [ + 497, + 535, + 904, + 878 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5607", + "bbox": [ + 468, + 891, + 500, + 903 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Skeleton-based Action Recognition. Due to the promising ability to model multiple granularities and large variations in human motion, the skeleton data is more suitable for the aforementioned action recognition task than the RGB data [4]. Early skeleton-based action recognition methods are usually handcrafted, exploiting the geometric relationship of skeleton joints [47, 73, 76, 74], which greatly limits the feature representation of skeletons. Benefiting from the advantages of deep learning, some methods [95, 62, 63] utilized RNNs as the basic model, Du et al. [10] presented a pioneering work based on hierarchical RNNs. But RNNs easily suffer from vanishing gradients [21]. Inspired by the booming Graph Convolutional Networks (GCN) [28], Yan et al. [83] proposed a spatial-temporal graph convolutional network to learn the spatial and temporal pattern from skeleton data. However, their manually defined topology is arduous to model the relations among joints in underlying topology. Chen et al. [5] proposed a channel-wise topology graph convolution, which models channel-wise topology with a refinement method. Duan et al. [11] proposed a PoseConv3D model that relies on a 3D heatmap volume instead of a graph sequence as the base representation of human skeletons. Different from previous methods that required large amounts of labeled data for training elaborate models, we utilize unlabeled skeleton sequences to pretrain a graph-based encoder-decoder named SkeletonMAE to comprehensively capture human pose and obtain discriminative skeleton sequence representation.", + "bbox": [ + 63, + 115, + 472, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-supervised Learning for Skeleton Sequence. To learn more effective representation for unlabeled skeleton data, self-supervised learning [43] has achieved inspiring progress recently. For contrastive learning approaches, ASCAL [55] and SkeletonCLR [31] applied momentum encoders for contrastive learning with single-stream skeleton sequences. AimCLR [18] used an extreme data augmentation strategy to add additional hard contrastive pairs. Most contrastive learning methods adopt data augmentation to generate positive and negative pairs, but they rely heavily on the number of contrastive pairs. For generative learning approaches, LongT GAN [91] proposed the encoder-decoder to reconstruct masked input sequence skeletons. Based on the LongT GAN, P&C [65] strengthened the encoder and weakened the decoder for feature representation. Wu et al. [80] proposed a spatial-temporal masked autoencoder framework for self-supervised 3D skeleton-based action recognition. Colorization [87] used three pairs of encoder-decoder frameworks to learn spatial-temporal features from skeleton point clouds. Due to the limitation of the reconstruction criterion, previous generative methods usually fail to fully discover the fine-grained spatial-temporal dependencies among different skeleton joints. In our SkeletonMAE, we utilize the prior human topology", + "bbox": [ + 63, + 522, + 472, + 865 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "knowledge to infer the skeleton sequence and obtain a comprehensive perception of the action.", + "bbox": [ + 497, + 115, + 904, + 146 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 499, + 156, + 638, + 173 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the details of Skeleton Sequence Learning (SSL), which contains two parts: 1) pretraining SkeletonMAE and 2) fine-tuning on downstream datasets based on the pre-trained SkeletonMAE.", + "bbox": [ + 497, + 180, + 904, + 238 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Pre-training SkeletonMAE", + "text_level": 1, + "bbox": [ + 499, + 245, + 752, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce graph-based asymmetric encoder-decoder pre-training architecture SkeletonMAE, to learn human skeleton sequence representations without supervision. Since Graph Isomorphism Network (GIN) [82] provides a better inductive bias, it is more suitable for learning more generalized self-supervised representation [22]. Therefore, we adopt GIN as the backbone of SkeletonMAE. Besides, we evaluate different backbones of SkeletonMAE in Table 4, including GIN [82], GCN [28], and GAT [72].", + "bbox": [ + 497, + 268, + 904, + 397 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 SkeletonMAE Structure", + "text_level": 1, + "bbox": [ + 499, + 414, + 729, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the effective representation learning by masked autoencoder (MAE) in NLP [9], image recognition [20], and video recognition [70], we focus on the human skeleton sequence and build an asymmetric graph-based encoder-decoder pre-training architecture named SkeletonMAE that embeds skeleton sequence and its prior topology knowledge in GIN. The SkeletonMAE is implemented following the paradigm of graph generative self-supervised learning.", + "bbox": [ + 497, + 435, + 902, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We follow the joint label of the Kinetics Skeleton dataset [60]. Specifically, as Figure 2(d) shown, we divide all $N = 17$ joints into $R = 6$ local regions according to the natural parts of the body: $\\nu_{0},\\dots,\\nu_{5}$ . Notably, compared to the randomly masking strategy from MAE [20] to select skeleton joints, our masking strategy is action-sensitive and reconstructs specific limbs or body parts that dominate the given action class. Then, we mask these skeleton regions and make the SkeletonMAE reconstruct the masked joint features and their edges based on the adjacent joints. By reconstructing the masked skeleton joints and edges, the SkeletonMAE can infer the underlying topology of joints and obtain a comprehensive perception of the action.", + "bbox": [ + 497, + 549, + 904, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 2, the SkeletonMAE is an asymmetric encoder-decoder architecture, which includes an encoder and a decoder. The encoder consists of $L_{D}$ GIN layers that map the input 2D skeleton data to hidden features. The decoder, which consists of only one GIN layer, reconstructs the hidden features under the supervision of the reconstruction criterion. According to the prior human skeleton knowledge that the human skeleton can be represented as a graph with joints as vertices and limbs as edges, we formulate the human skeleton as the following graph structure.", + "bbox": [ + 497, + 735, + 904, + 878 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5608", + "bbox": [ + 468, + 891, + 502, + 903 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/35375e2e6c09ba63eb670e2bc3dbc5dc2c87d494400546cf2143b510478c75eb.jpg", + "image_caption": [ + "Figure 2: The details of our skeleton sequence pre-training architecture SkeletonMAE. (a) We build a GIN-based asymmetric encoder-decoder structure, to reconstruct joint features to enhance action representation ability. (b) The GIN-based encoder structure contains $L_{D}$ GIN neural network layers, to learn the joint representation spatially. (c) The decoder consists of one GIN layer, which uses the hidden features from the encoder as the input and reconstructs the original input joint features. (d) Partition the joints in the skeleton sequence according to the natural structure of the human body. 5 joints $\\{\\mathcal{V}_0:\\mathrm{Head}\\}$ , 4 joints $\\{\\mathcal{V}_1:\\mathrm{Torso}\\}$ , 2 joints $\\{\\mathcal{V}_2:\\mathrm{Left arm},\\mathcal{V}_3:\\mathrm{Right arm},\\mathcal{V}_4:\\mathrm{Left leg},\\mathcal{V}_5:\\mathrm{Right leg}\\}$ ." + ], + "image_footnote": [], + "bbox": [ + 67, + 112, + 904, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The skeleton sequence of two-dimensional coordinates of $N$ human skeleton joints and $T$ frames is pre-processed in the following way. Specifically, we embed all skeleton joints and their topology into a structure $\\mathcal{G}$ , the skeleton structure and the joint feature are fused to obtain a joint sequence matrix $\\mathbf{S} \\in \\mathbb{R}^{N \\times T \\times 2}$ . And then the $\\mathbf{S}$ is linearly transformed to $\\mathbf{S} \\in \\mathbb{R}^{N \\times T \\times D}$ with learnable parameters. We empirically set $\\mathrm{T}$ and $\\mathrm{D}$ to 64. For each skeleton frame $\\mathbf{X} \\in \\mathbb{R}^{N \\times D}$ from $\\mathbf{S}$ , let $\\mathcal{G} = (\\mathcal{V}, \\mathbf{A}, \\mathbf{X})$ denote a skeleton, where $\\mathcal{V} = \\{v_1, v_2, \\dots, v_N\\}$ is the node set that contains all skeleton joints, $N = |\\mathcal{V}|$ is the number of joints. The number of joints is $N = 17$ . $\\mathbf{A} \\in \\{0, 1\\}^{N \\times N}$ is an adjacency matrix, where $\\mathbf{A}_{i,j} = 1$ if joints $i$ and $j$ are physically connected, otherwise 0. The feature of $v_i$ is represented as $\\mathbf{x}_i \\in \\mathbb{R}^{1 \\times D}$ . And $G_E$ , $G_D$ denote the GIN encoder and the GIN decoder, respectively.", + "bbox": [ + 63, + 391, + 472, + 621 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2 Skeleton Joints Masking and Reconstruction", + "text_level": 1, + "bbox": [ + 63, + 638, + 440, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the prior human skeleton topology $\\mathbf{A}$ is embedded (Figure 2) and we specify the aggregation of joints in Section 3.1.1. Inspired by the GraphMAE [22] that randomly reconstructs the masked graph nodes, our SkeletonMAE reconstructs the masked skeleton feature $\\mathbf{X}$ based on the prior skeleton topology, rather than reconstructing graph structure $\\mathbf{A}$ [68, 17] or reconstructing both structure $\\mathbf{A}$ and features $\\mathbf{X}$ [57, 53].", + "bbox": [ + 63, + 662, + 468, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To mask skeleton joint features, we randomly select one or more joint sets from $\\mathcal{V} = \\{\\mathcal{V}_0,\\dots,\\mathcal{V}_5\\}$ , which consists of a subset $\\overline{\\mathcal{V}}\\subseteq \\mathcal{V}$ for masking. For the human skeleton sequence, each joint communicates with some of its adjacent joints to represent the specific action class. Therefore, it is not feasible to mask all joint sets for all action classes. Then, each of their features is masked with a learnable mask", + "bbox": [ + 63, + 777, + 468, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "token vector $[\\mathbf{MASK}] = \\mathbf{x}_{[\\mathbf{M}]} \\in \\mathbb{R}^D$ . Thus, the masked joint feature $\\overline{\\mathbf{x}}_i$ for $\\mathbf{v}_i \\in \\overline{\\mathcal{V}}$ in the masked feature matrix $\\overline{\\mathbf{X}}$ can be defined as $\\overline{\\mathbf{x}}_i = \\mathbf{x}_{[\\mathbf{M}]}$ if $\\mathbf{v}_i \\in \\overline{\\mathcal{V}}$ , otherwise $\\overline{\\mathbf{x}}_i = \\mathbf{x}_i$ . We set $\\overline{\\mathbf{X}} \\in \\mathbb{R}^{N \\times D}$ as the input joint feature matrix of the SkeletonMAE, and each joint feature in $\\overline{\\mathbf{X}}$ can be defined as $\\overline{\\mathbf{x}}_i = \\{\\mathbf{x}_{[\\mathbf{M}]}, \\mathbf{x}_i\\}$ , $i = 1, 2, \\dots, N$ . Therefore, the masked skeleton sequence can be formulated as $\\overline{\\mathcal{G}} = (\\mathcal{V}, \\mathbf{A}, \\overline{\\mathbf{X}})$ and the objective of SkeletonMAE is to reconstruct the masked skeleton features in $\\overline{\\mathcal{V}}$ given the partially observed joint features $\\overline{\\mathbf{X}}$ with the input adjacency matrix $\\mathbf{A}$ . The process of SkeletonMAE reconstruction is formalized as:", + "bbox": [ + 497, + 390, + 905, + 552 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\int \\mathbf {H} = G _ {E} (\\mathbf {A}, \\overline {{\\mathbf {X}}}), \\quad \\mathbf {H} \\in \\mathbb {R} ^ {N \\times D _ {h}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 574, + 557, + 902, + 573 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathbf {Y} = G _ {D} (\\mathbf {A}, \\mathbf {H}), \\quad \\mathbf {Y} \\in \\mathbb {R} ^ {N \\times D} \\right., \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 573, + 902, + 587 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{H}$ and $\\mathbf{Y}$ denote the encoder output and the decoder output, respectively. The objective of SkeletonMAE can be formalized as minimizing the divergence between $\\mathbf{X}$ and $\\mathbf{Y}$ .", + "bbox": [ + 497, + 596, + 905, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.3 Reconstruction Criterion", + "text_level": 1, + "bbox": [ + 499, + 655, + 737, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The common reconstruction criterion for masked autoencoders is a mean squared error (MSE) in image and video tasks. However, for skeleton sequence, the multidimensional and continuous nature of joint features makes MSE hard to achieve promising feature reconstruction because the MSE is sensitive to dimensionality and vector norms of features [15]. Inspired by the observation [16] that the $l_{2}$ -normalization in the cosine error maps vectors to a unit hyper-sphere and substantially improves the training stability, we utilize the cosine error as the reconstruction.", + "bbox": [ + 497, + 676, + 905, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To make the reconstruction criterion focus on harder ones among imbalanced easy-and-hard samples [22], we use the Re-weighted Cosine Error (RCE) for SkeletonMAE. The RCE is based on the intuition that we can down-weigh", + "bbox": [ + 497, + 819, + 904, + 878 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5609", + "bbox": [ + 468, + 891, + 502, + 903 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "easy samples' contribution in training by scaling the cosine error with a power of $\\beta \\geq 1$ . For predictions with high confidence, their corresponding cosine errors are usually smaller than 1 and decay faster to zero when the scaling factor $\\beta > 1$ . Formally, given the original feature $\\mathbf{X} \\in \\mathbb{R}^{N \\times D}$ and the reconstructed output $\\mathbf{Y} \\in \\mathbb{R}^{N \\times D}$ , the RCE is defined as:", + "bbox": [ + 63, + 114, + 468, + 213 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {R C E}} = \\sum_ {\\mathbf {v} _ {i} \\in \\overline {{\\mathcal {V}}}} \\left(\\frac {1}{| \\overline {{\\mathcal {V}}} |} - \\frac {\\mathbf {x} _ {i} ^ {\\mathrm {T}} \\cdot \\mathbf {z} _ {i}}{| \\overline {{\\mathcal {V}}} | \\times \\| \\mathbf {x} _ {i} \\| \\times \\| \\mathbf {z} _ {i} \\|}\\right) ^ {\\beta}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 115, + 220, + 468, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which represents the average of the similarity gap between the reconstructed feature and the input feature over all masked joints. And $\\beta$ is set to 2 in our work.", + "bbox": [ + 63, + 267, + 468, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By training the SkeletonMAE to reconstruct the skeleton sequence, the pre-trained SkeletonMAE can comprehensively perceive the human skeleton structure and obtain discriminative action representation. After pre-training, the SkeletonMAE can be elegantly embedded into the Skeleton Sequence Learning (SSL) framework for fine-tuning.", + "bbox": [ + 63, + 309, + 467, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Fine-tuning for Skeleton Action Recognition", + "text_level": 1, + "bbox": [ + 65, + 403, + 448, + 418 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the SkeletonMAE's generalization ability for skeleton action recognition, we build a complete skeleton action recognition model named Skeleton Sequence Learning (SSL), based on the pre-trained SkeletonMAE. To capture multiple-person interaction, we integrate two pre-trained SkeletonMAE encoders to build the Spatial-Temporal Representation Learning (STRL) module, as shown in Figure 3(b)-(c). The entire SSL consists of an $M$ -layer STRL model and a classifier. The SSL model is finally fine-tuned on skeleton action recognition datasets with cross-entropy loss to recognize actions.", + "bbox": [ + 63, + 425, + 468, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.1 Spatial-Temporal Representation Learning", + "text_level": 1, + "bbox": [ + 63, + 598, + 428, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The STRL contains two pre-trained SkeletonMAE encoders for Spatial Modeling (SM). The input of SM is skeleton sequence S. The output of SM is connected with the input by $1\\times 1$ convolution for residual connection (Figure 3 (b)).", + "bbox": [ + 63, + 619, + 467, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 3 (c), the input skeleton sequence $\\mathbf{S} \\in \\mathbb{R}^{N \\times T \\times D}$ is firstly added with the learnable temporal position embedding PE to obtain the skeleton sequence feature $\\mathbf{H}_t^{(l)} \\in \\mathbb{R}^{P \\times N \\times D^{(l)}}$ . To model multiple human skeleton interactions, we obtain two individual features ( $P = 2$ ) for two persons $\\mathbf{H}_{t,0}^{(l)} \\in \\mathbb{R}^{N \\times D^{(l)}}$ and $\\mathbf{H}_{t,1}^{(l)} \\in \\mathbb{R}^{N \\times D^{(l)}}$ from $\\mathbf{H}_t^{(l)}$ . Here, we take the joint feature of the 0-th person as an example, the operation of the 1-th person is implemented similarly. We send the joint representation $\\mathbf{H}_{t,0}^{(l)}$ and prior knowledge of the joint $\\widetilde{\\mathbf{A}}$ into the SM module,", + "bbox": [ + 63, + 678, + 468, + 832 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S M} \\left(\\mathbf {H} _ {t, 0} ^ {(l)}\\right) = \\operatorname {R e p e a t} \\left(\\operatorname {S P} \\left(G _ {E} \\left(\\widetilde {\\mathbf {A}}, \\mathbf {H} _ {t, 0} ^ {(l)}\\right)\\right); N\\right) \\oplus \\mathbf {H} _ {t, 0} ^ {(l)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 80, + 840, + 467, + 878 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fb87c7575a775872921782e56c839404148f88d67827dcabea1af4f08c84bb38.jpg", + "image_caption": [ + "Figure 3: The pipeline of Skeleton Sequence Learning (SSL). (a) During pre-training, we build an encoder-decoder module named SkeletonMAE that embeds skeleton joints and its prior topology knowledge into GIN and reconstructs the underlying masked joints and topology. (b) The SM consists of the pre-trained SkeletonMAE encoder. (c) We integrate SM structures to build the $M$ -layer Spatial-Temporal Representation Learning (STRL) model and then conduct end-to-end fine-tuning." + ], + "image_footnote": [], + "bbox": [ + 504, + 112, + 771, + 336 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ec9ea7da87ce471a951ce204826e2fea2f11ec2f1df86543b0532710f1a7f507.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 114, + 905, + 336 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $G_{E}$ is the SkeletonMAE encoder, $\\mathrm{SP}(\\cdot)$ denotes the sum-pooling, $\\mathrm{Repeat}(\\cdot;N)$ means repeating the single joint into $N$ joints representations after sum-pooling and then connect it with the $\\mathbf{H}_{t,0}^{(l)}$ residual to get the global joint representation $\\mathrm{SM}(\\mathbf{H}_{t,0}^{(l)})$ . In this way, the SM module can obtain global information through a single joint representation, and constrain some joint features through all joint representations. Similarly, $\\mathrm{SM}(\\mathbf{H}_{t,1}^{(l)})$ is obtained in the same way. As shown in Figure 3(c), we get the joint features $\\mathrm{SM}(\\mathbf{H}_t^{(l)})$ that contains the action interaction bewteen the 0-th person and the 1-th person. According to the update rules of graph convolution [28], we can obtain $\\mathbf{H}_t^{(l + 1)}$ from $\\mathbf{H}_t^{(l)}$ in a multi-layer GCN. For more details, please refer to the Supplementary in Section D. The final skeleton sequence representation is defined as follows:", + "bbox": [ + 497, + 517, + 902, + 745 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {t} ^ {(l + 1)} = \\sigma (\\mathbf {S M} (\\mathbf {H} _ {t} ^ {(l)}) \\mathbf {W} ^ {(l)}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 763, + 902, + 788 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{W}^{(l)}$ denotes the trainable weight matrix in the $l^{th}$ layer, $\\sigma(\\cdot)$ denotes the ReLU activation function. Then, we adopt the widely-used multi-scale temporal pooling [5, 38] to get the final output. Finally, a classifier consisting of MLP and softmax predicts the action class.", + "bbox": [ + 497, + 804, + 904, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5610", + "bbox": [ + 468, + 891, + 500, + 903 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 65, + 114, + 200, + 130 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All experiments are conducted with a single modality (2D pose) and evaluated on the corresponding train/test sets.", + "bbox": [ + 65, + 137, + 467, + 166 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 65, + 172, + 169, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our SSL on four benchmark datasets FineGym [59], Diving48 [34], NTU RGB+D 60 [58] and NTU RGB+D 120 [37] in the mainstream skeleton action recognition task. For all datasets except FineGym, we follow the pre-processing protocol provided by [11] to obtain the skeleton sequence from the 2D pose estimator. The pre-processing adopts a Top-Down approach for pose extraction, where the detector is Faster-RCNN [56] with ResNet50 backbone and the pose estimator is HRNet [67] pre-trained on COCO-keypoint [36]. To make a fair comparison, we added pixel noise to the joint during training, making the original joint confidence rate unreliable, thus we do not use the originally fixed threshold.", + "bbox": [ + 63, + 194, + 467, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FineGym is a large-scale fine-grained action recognition dataset for gymnastic videos, which contains 29K videos of 99 fine-grained gymnastics action classes, which requires action recognition methods to distinguish different sub-actions in the same video. In particular, it provides temporal annotations at both action and sub-action levels with a three-level semantic hierarchy. We follow the method [11] to extract the skeleton data from the 2D pose estimator.", + "bbox": [ + 63, + 388, + 467, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "NTU RGB+D 60 and 120. NTU RGB+D is a large-scale skeleton-based action recognition dataset, where NTU 60 contains 56,880 skeleton sequences and 60 action classes. NTU 120 has 114,480 skeleton sequences and 120 action categories. The action samples are captured from 155 different camera viewpoints. The subjects in NTU 120 are in a wide range of age distribution (from 10 to 57) and from different cultural backgrounds (15 countries), which brings very realistic variation to the quality of actions. The NTU 60 and 120 datasets have a large amount of variation in subjects, views, and backgrounds.", + "bbox": [ + 63, + 511, + 467, + 668 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Diving48 is a challenging fine-grained dataset that focuses on complex and competitive sports content analysis. It is formed of over 18k video clips from competitive dive championships that are distributed over 48 fine-grained classes with minimal biases. The difficulties of the dataset lie in that actions are similar and differ in body parts and their combinations which require the model to capture details and motion in body parts and combine them to perform classification. We report the accuracy on the official train/test split.", + "bbox": [ + 63, + 676, + 467, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 65, + 812, + 287, + 826 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this paper, our SkeletonMAE is optimized by the Adaptive Moment Estimation (Adam) with the initial learning rate as $1.5e^{-4}$ and the PReLU is the activation function.", + "bbox": [ + 63, + 834, + 467, + 876 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/657e307161cec89a5369e3540df788ba787a073073f110b32297f9feea62f2fe.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodModalityMean Acc. (%)
Fully Supervised
I3D [2]RGB64.4
ST-S3D [81]RGB72.9
TSN [77]RGB+Flow79.8
TRNms [93]RGB+Flow80.2
TSM [35]RGB+Flow81.2
GST-50 [46]RGB84.6
MTN [30]RGB88.5
LT-S3D [81]RGB88.9
TQN [90]RGB+Text90.6
PYSKL [11]Skeleton93.2
PYSKL [11]RGB+Skeleton+Limb95.6
Unsupervised Pre-train
SaL [49]RGB42.7
TCC [12]RGB45.6
GTA [19]RGB49.5
CARL [3]RGB60.4
SSL (ours)Skeleton91.8
", + "bbox": [ + 505, + 112, + 895, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: The comparison with the state-of-the-art unsupervised pre-train and supervised methods on FineGym.", + "bbox": [ + 499, + 374, + 900, + 403 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The batch size is 1024 and the training epoch is 50. At the fine-tuning stage, we use the Stochastic Gradient Descent (SGD) with momentum (0.9) and adopt the warm-up strategy for the first 5 epochs. The total fine-tuning epochs are 110. The learning rate is initialized to 0.1 and is divided by 10 at the 90 epoch and the 100 epoch. And we employ 0.1 for label smoothing. We use a large batch size of 128 to facilitate training our attention mechanism and enhancing the model's perception for all human action classes. Both our pre-training and fine-tuning models are implemented by PyTorch [54], and our SSL is trained on a single NVIDIA GeForce RTX 2080Ti GPU. For more details of implementation, please refer to the Supplementary in Section A.", + "bbox": [ + 497, + 429, + 900, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pre-training and Fine-tuning Setting. For each dataset, the SkeletonMAE encoder is pre-trained with unlabeled data from the training set. Then, we load the learned parameter weights to fine-tune the SSL model.", + "bbox": [ + 497, + 619, + 900, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. To make a fair comparison, we follow previous methods [11, 3, 18] and report the Mean Top-1 accuracy $(\\%)$ on FineGym dataset and Top-1 accuracy $(\\%)$ on Diving48, NTU 60, and NTU 120 datasets.", + "bbox": [ + 499, + 682, + 900, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Downstream Evaluation", + "text_level": 1, + "bbox": [ + 499, + 749, + 726, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a fair comparison, we compare SSL with other models without pre-training on large-scale action datasets, e.g., Kinetics [27] or Sports1M [26]. The comparison results on FineGym, NTU 60 & 120, and Diving48 datasets are shown in Table 1, Table 2, and Table 3, respectively.", + "bbox": [ + 497, + 771, + 900, + 842 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on FineGym Dataset. In Table 1, our SSL with skeleton input outperforms most of the fully supervised", + "bbox": [ + 499, + 847, + 900, + 876 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5611", + "bbox": [ + 468, + 891, + 500, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cedfa05fb94bb2bcc24d5a9dc473424b2b851d11aaa06fbe9dcf5f951acc6bf0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneSupervisionJoint Number2D SkeletonNTU 60NTU 120
X-sub (%)X-view (%)X-sub (%)X-set (%)
ST-GCN [83]GCNFully Supervised25X81.588.3--
AS-GCN [32]GCNFully Supervised25X86.894.2--
2s-AGCN [60]GCNFully Supervised25X88.595.182.984.9
Shift-GCN [6]GCNFully Supervised25X90.796.585.987.6
MS-G3D [45]GCNFully Supervised25X91.596.286.988.4
CTR-GCN [5]GCNFully Supervised25X92.496.888.990.6
PYSKL [11]CNNFully Supervised17✓93.796.686.089.6
SkeletonCLR [31]ST-GCNUnsupervised Pre-train25X82.288.973.675.3
CrossCLR [31]ST-GCNUnsupervised Pre-train25X86.292.580.580.4
Wu et al. [80]STTFormerUnsupervised Pre-train25X86.692.976.879.1
AimCLR [18]ST-GCNUnsupervised Pre-train25X86.992.880.180.9
3s-PSTL [94]ST-GCNUnsupervised Pre-train25X87.193.981.382.6
Colorization [87]DGCNNUnsupervised Pre-train25X88.094.9--
SSL(ours)STRLUnsupervised Pre-train17✓92.8(↑ 4.8)96.5(↑ 1.6)84.8(↑ 3.5)85.7(↑ 3.1)
", + "bbox": [ + 68, + 112, + 902, + 349 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/09a29f0e00cc88705fa87307c3285afc739853624a2728fad9a6b921678e2c8e.jpg", + "table_caption": [ + "Table 2: The comparison with state-of-the-art unsupervised pre-train and supervised methods on NTU 60 and NTU 120 datasets. $\\underline{\\mathbf{\\Pi}}_{-}^{\\prime}$ means the method with the second-best performance under unsupervised pre-training." + ], + "table_footnote": [], + "table_body": "
MethodPre-trainGFLOPsAcc.(%)
Fully Supervised
TSN [34]ImageNet-16.8
TRN [25]ImageNet-22.8
P3D [46]ImageNet-32.4
C3D [46]ImageNet-34.5
CorrNet [46]-74.835.5
CorrNet-R101 [75]ImageNet187.338.2
MG-TEA-ResNet50 [92]ImageNet-39.5
GSM [66]ImageNet107.440.3
TSM-R50 [29]ImageNet153.841.6
TMF [78]ImageNet-42.2
Unsupervised Pre-train
RESOUND-C3D [34]K400-16.4
Jenni et al. [23]K400-29.9
SSL (ours)Diving4842.834.1
", + "bbox": [ + 67, + 408, + 468, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3: The comparison with the unsupervised pre-train and supervised methods on the Diving48 dataset.", + "bbox": [ + 63, + 631, + 468, + 661 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods and achieves the best performance among unsupervised pre-train methods with RGB input. For the same input modality, our performance is lower than the fully supervised method PYSKL[11] (with the skeleton as input) by about $1.4\\%$ , because the PYSKL adopts stacks of visual heatmaps of skeleton joints as input while we only use human skeleton coordinates. This validates the promising discriminative ability of our skeleton sequence representation.", + "bbox": [ + 63, + 686, + 468, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on NTU 60 and NTU 120 Datasets. In Table 2, for NTU 60 X-sub and NTU 60 X-view, compared with unsupervised pre-train methods, our SSL outperforms the current state-of-the-art method Colorization [87] by $4.8\\%$ and $1.6\\%$ , respectively. Our SSL is also competitive compared", + "bbox": [ + 63, + 805, + 468, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with fully supervised methods, outperforming the first six fully supervised methods on NTU 60 X-sub. For NTU 120 X-sub and NTU 120 X-set, our SSL outperforms the previous best-unsupervised pre-train method 3s-PSTL [94] by $3.5\\%$ and $3.1\\%$ , respectively. Our SSL is superior compared with some fully supervised methods on NTU 120 X-sub and NTU 120 X-set. These results show that our SSL can learn discriminative skeleton representation from large-scale action recognition datasets due to the promising generalization ability of our SkeletonMAE.", + "bbox": [ + 497, + 410, + 904, + 553 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on Diving48 Dataset. Our SSL with skeleton input outperforms some fully supervised methods. Although our SSL is not pre-trained on additional large-scale pretraining action datasets in Table 3, it still achieves competitive performance among unsupervised pre-train methods. This validates that our pre-training model SkeletonMAE can learn discriminative skeleton sequence representation.", + "bbox": [ + 497, + 560, + 902, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results on FineGym and Diving48 validate that our SkeletonMAE has a promising ability to enhance the feature representation of skeleton sequence by comprehensively perceiving the underlying topology of actions, and the SSL can learn discriminative action representation.", + "bbox": [ + 497, + 661, + 902, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 499, + 745, + 670, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we analyze the contributions of essential components and hyper-parameters of our model. Note that unless otherwise specified, all experiments are verified on the FineGym dataset with masking body sub-region 3.", + "bbox": [ + 497, + 769, + 904, + 828 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Whether to load pre-trained model or not. To explore the effectiveness of loading the pre-trained SkeletonMAE encoder, we find that the accuracy is 86.3 without load", + "bbox": [ + 497, + 833, + 904, + 878 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5612", + "bbox": [ + 468, + 891, + 502, + 903 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b06c6de66b65b037a9309186e6d837b69cb7e8e88d69c935ae255b77921b0800.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LDMean Acc.
189.6
290.7
391.2
490.9
", + "bbox": [ + 90, + 112, + 206, + 182 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5e829114466724f13d860dd047a18f48d4dfdab9389fc4f842409003447d6777.jpg", + "table_caption": [ + "(a)" + ], + "table_footnote": [], + "table_body": "
MethodMean Acc.
GraphCL [89]86.5
JOAO [88]88.7
Ours(SkeletonMAE)91.2
", + "bbox": [ + 231, + 122, + 458, + 178 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/788823b018c4e153357a2f7c5812de40f31698964bfaa9faf8354a92e29da8b7.jpg", + "table_caption": [ + "(b)" + ], + "table_footnote": [], + "table_body": "
#Masked Body Part012345
GAT [72]86.888.188.989.589.490.0
GCN [28]87.688.989.390.689.590.5
GIN [82]88.689.590.291.290.391.2
", + "bbox": [ + 73, + 203, + 458, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing the pre-trained SkeletonMAE encoder (randomly initialized weights). As Table 5(a) shows, loading the pretrained model is always better than not loading it. This validates that our SkeletonMAE can learn more comprehensive and generalized representations for unlabeled fine-grained actions by reconstructing the skeleton joint features.", + "bbox": [ + 63, + 370, + 467, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GIN layers in SkeletonMAE. Table 4(a) shows the performance of using different GIN layers in the SkeletonMAE encoder. The performance is the best when $L_{D} = 3$ .", + "bbox": [ + 63, + 463, + 468, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with contrastive learning methods. To verify the superior ability of our SkeletonMAE when conducting skeleton sequence pre-training, we compare our SkeletonMAE with different contrastive learning methods GraphCL and JOAO. As shown in Table 4(b), our SkeletonMAE achieves the best performance. Besides, we visually compare the action representations of SkeletonMAE and GraphCL by PCA, as shown in Figure 4(a) and Figure 4(b). Compared to GraphCL, the skeleton representation of our SkeletonMAE appears to have a larger inter-class variance and smaller intra-class variance. This validates that our SkeletonMAE can comprehensively capture the human pose and obtain discriminative skeleton sequence representation. We observe similar patterns in all other classes but visualize only five categories for simplicity.", + "bbox": [ + 63, + 513, + 468, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Backbones and masked body parts in SkeletonMAE. As shown in Table 4(c), we show the accuracy of our SSL with different SkeletonMAE backbones and different masked body parts in SkeletonMAE. It can be seen that GIN is always better than both GAT and GCN under the same masked body part. This is because that GIN provides a better inductive bias for graph-level applications. Thus, it is more suitable for learning more generalized skeleton representations. Additionally, we can see that masking body sub-regions 3 and 5 are both optimal across all backbones,", + "bbox": [ + 63, + 733, + 468, + 876 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3ef71e46cc99393e7d444e7f52199e30fc1208963dabbf502bb6c9ef8253e405.jpg", + "table_caption": [ + "(c)", + "Table 4: (a) Mean accuracy of using the different number of GIN layers in SkeletonMAE encoder. (b) Comparison results with the contrastive learning method as the pre-trained encoder. (c) The results of using different backbones in SkeletonMAE under each masked body part are compared." + ], + "table_footnote": [], + "table_body": "
M1234
Mean Acc.89.1(↑2.8)90.6(↑4.3)91.2(↑4.9)91.0(↑4.7)
", + "bbox": [ + 500, + 112, + 904, + 143 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/89e3c714cbcc815314b4a3dcf39460b323e6a5877906b726a8b335aa65022df9.jpg", + "table_caption": [ + "(a)" + ], + "table_footnote": [], + "table_body": "
#Masked Joints Number591215
Ratio of Mask Joints Accuracy of SSL30%50%70%90%
89.790.389.990.1
Masked Body PartHigh91.8 (V3,V5)91.2 (V0,V3,V5)91.0 (V1,V2,V3, V4,V5)90.8 (V0,V1,V2, V3,V5)
Low91.1 (V2,V4)90.1 (V0,V1)91.0 (V1,V2,V3, V4,V5)90.2 (V0,V1,V3, V4,V5)
", + "bbox": [ + 500, + 168, + 907, + 290 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(b)", + "bbox": [ + 690, + 290, + 710, + 300 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5: (a) Results of four SSL variants. $\\uparrow$ represents the accuracy improvement relative to the random initialization of SkeletonMAE in SSL. (b) The comparison of our body part based masked and the random masked strategies. $\\nu_{0}-\\nu_{5}$ : Head, Torso, Left arm, Right arm, Left leg, Right leg.", + "bbox": [ + 497, + 308, + 902, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "which demonstrates the importance of reconstruction of human limbs in action recognition.", + "bbox": [ + 497, + 410, + 900, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Variants of SSL. To evaluate whether our pre-trained SkeletonMAE is effective across different skeleton action recognition models, we set the different number of STRL layers $(M = 1,2,3,4)$ to obtain four variants of the SSL. As shown in Table 5(a), all SSL variants outperform the random initialization of SkeletonMAE in SSL, which validates our body part masking strategy indeed improves the discriminative ability of skeleton feature by learning action-sensitive visual concepts. Additionally, three-layer STRL is the best due to the good compromise between efficiency and computational cost. Moreover, it also validates that our SkeletonMAE generalizes well across different models.", + "bbox": [ + 497, + 447, + 902, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Skeleton Masked Strategy. In Table 5(b), our masked body part strategy is fairly compared with the random masked strategy under the same masked joint conditions. Our method is better than the random mask method across all settings. As mentioned in Section 3.1.1, our masking strategy is action-sensitive and reconstructs specific limbs or body parts that dominates the given action class and is suitable for real-world skeleton-based action recognition. From Table 5, we can see that SSL effectively learns meaningful representations that encode the spatial relationships between joints, enabling the capture of crucial information about the structure and configuration of human movement. This notable achievement can be attributed to the integration of prior knowledge about the human body through pretraining the SkeletonMAE. We report results on the ntu 60 dataset in Supplementary Section E.", + "bbox": [ + 497, + 625, + 902, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Transferability of the SkeletonMAE across datasets.", + "bbox": [ + 500, + 862, + 900, + 876 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5613", + "bbox": [ + 468, + 891, + 500, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1bd449f4e6909eafdc3815d68bdde47a97bcb875710a756a5e4db1e849cc13de.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 72, + 114, + 200, + 192 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/255c89d302dfcd18f7cc9283a17a426442beeea21573ccfc133c175861011d0e.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 201, + 114, + 329, + 192 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b0f61c7d2d6aa89463196bc696b5d714b3f682d1efda655f9efa874ac41b16ed.jpg", + "image_caption": [ + "(c)", + "Figure 4: (a) and (b) are 2d-PCA of SkeletonMAE and GraphCL as pre-trained encoder representations. We randomly select five action classes for 2d-PCA visualization, (c) Confusion matrix for fine-grained action recognition." + ], + "image_footnote": [], + "bbox": [ + 334, + 112, + 463, + 192 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As shown in Figure 5(a), we pre-train SkeletonMAE on the FineGym dataset and then fine-tune it on NTU 60 X-sub and NTU 120 X-sub datasets. Compared with the method that uses the same dataset for pre-training and fine-tuning, our SkeletonMAE achieves better performance across all masked strategies when conducting dataset transfer. This shows that the SkeletonMAE can learn generalized skeleton representation and effectively transfer the strong representation ability to other datasets.", + "bbox": [ + 63, + 286, + 468, + 414 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.5. Visualization Analysis", + "text_level": 1, + "bbox": [ + 65, + 427, + 275, + 442 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Figure 6 shows the reconstruction process of the skeleton sequence by SkeletonMAE. From the same frame, the difference between the reconstructed skeleton sequence and the original skeleton sequence is slight, but overall the human body structure is reserved. This shows that the SkeletonMAE has good spatial representation learning ability. Moreover, SSL effectively captures the temporal evolution and distinguishing characteristics of actions by capturing the relationships between consecutive joint positions and poses. This highlights the positive impact of the multi-scale temporal dependence incorporated within the STRL module. Figure 4(c) shows that our SSL works well for fine-grained action recognition tasks on the FineGym dataset. More visualization results are in Supplementary Section C.", + "bbox": [ + 63, + 451, + 468, + 650 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 65, + 666, + 186, + 681 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose an efficient skeleton sequence learning framework SSL, to learn discriminative skeleton-based action representation. To comprehensively capture the human pose and obtain skeleton sequence representation, we propose a graph-based encoder-decoder pretraining architecture, SkeletonMAE, that embeds skeleton joint sequence into GCN and utilize the prior human topology knowledge to guide the reconstruction of the underlying masked joints and topology. Extensive experimental results show that our SSL achieves SOTA performance on four benchmark skeleton-based action recognition datasets. In our future work, we will build a multi-level feature refinement module to identify ambiguous skeleton actions.", + "bbox": [ + 63, + 692, + 467, + 876 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/54dd7b564b7cc918b149d8eeec8ece73b720a305fd5500f6b811f23c1e2e00ad.jpg", + "image_caption": [ + "(a) Masked body part" + ], + "image_footnote": [], + "bbox": [ + 561, + 112, + 842, + 228 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c03a1682afb7420414d6f69e2ce248a81872f4a2f53ddb88d23e843002a2d942.jpg", + "image_caption": [ + "(b) Masked body part", + "Figure 5: (a) The accuracies with mask body part of 0-5 on the NTU 60 X-sub dataset, and (b) the accuracies on the NTU 120 X-sub dataset. \\* means SkeletonMAE encoder pre-trained on the FineGym dataset." + ], + "image_footnote": [], + "bbox": [ + 561, + 257, + 842, + 372 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/dea9acba65f830331750a0eee144a9338b3fc3bd9fce2d4313fe411e01c39792.jpg", + "image_caption": [ + "(a) Original joint" + ], + "image_footnote": [], + "bbox": [ + 502, + 461, + 899, + 509 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9f5ff0b48515ec9109f88be3e6505e8309f047ece657364b22483d5e70580491.jpg", + "image_caption": [ + "(b) Masked joint" + ], + "image_footnote": [], + "bbox": [ + 502, + 524, + 897, + 564 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/70ebc3ed8c947882234cde932c6c1488f1e55a263922d0030cee929599c22391.jpg", + "image_caption": [ + "(c) Reconstructed joint", + "Figure 6: Visualization results for skeleton sequence of \"aerial walkover forward\" on FineGym. (a) The input skeleton sequence. (b) Masked skeleton sequence (masked parts are 3 and 5). (c) Reconstructed skeleton sequence." + ], + "image_footnote": [], + "bbox": [ + 502, + 579, + 897, + 618 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 722, + 663, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported in part by the National Key R&D Program of China under Grant No.2021ZD0111601, in part by the National Natural Science Foundation of China under Grant No.62002395 and No.61976250, in part by the Guangdong Basic and Applied Basic Research Foundation under Grant No.2023A1515011530, No.2021A1515012311 and No.2020B1515020048, and in part by the Guangzhou Science and Technology Planning Project under Grant No. 2023A04J2030.", + "bbox": [ + 497, + 748, + 902, + 876 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5614", + "bbox": [ + 468, + 891, + 500, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 65, + 112, + 164, + 128 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 1", + "[2] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1, 2, 6", + "[3] Minghao Chen, Fangyun Wei, Chong Li, and Deng Cai. Frame-wise action representations for long videos via sequence contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13801-13810, 2022. 1, 6", + "[4] Tailin Chen, Desen Zhou, Jian Wang, Shidong Wang, Yu Guan, Xuming He, and Errui Ding. Learning multi-granular spatio-temporal graph network for skeleton-based action recognition. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4334-4342, 2021. 3", + "[5] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13359-13368, 2021. 2, 3, 5, 7", + "[6] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 183-192, 2020. 7", + "[7] Sangwoo Cho, Muhammad Maqbool, Fei Liu, and Hassan Foroosh. Self-attention network for skeleton-based human action recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 635–644, 2020. 1", + "[8] Xiao Chu, Wei Yang, Wanli Ouyang, Cheng Ma, Alan L Yuille, and Xiaogang Wang. Multi-context attention for human pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1831-1840, 2017. 1", + "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3", + "[10] Yong Du, Wei Wang, and Liang Wang. Hierarchical recurrent neural network for skeleton based action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1110-1118, 2015. 3", + "[11] Haodong Duan, Yue Zhao, Kai Chen, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2969-2978, 2022. 1, 2, 3, 6, 7", + "[12] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. Temporal cycle-consistency learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1801-1810, 2019. 6" + ], + "bbox": [ + 67, + 137, + 468, + 875 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yin Fan, Xiangju Lu, Dian Li, and Yuanliu Liu. Video-based emotion recognition using cnn-rnn and c3d hybrid networks. In Proceedings of the 18th ACM international conference on multimodal interaction, pages 445-450, 2016. 1", + "[14] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1, 2", + "[15] Jerome H Friedman. On bias, variance, 0/1—loss, and the curse-of-dimensionality. Data mining and knowledge discovery, 1(1):55-77, 1997. 4", + "[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 4", + "[17] Aditya Grover and Jure Leskovec. node2vec: Scalable feature learning for networks. In Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining, pages 855-864, 2016. 4", + "[18] Tianyu Guo, Hong Liu, Zhan Chen, Mengyuan Liu, Tao Wang, and Runwei Ding. Contrastive learning from extremely augmented skeleton sequences for self-supervised action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 762-770, 2022. 2, 3, 6, 7", + "[19] Isma Hadji, Konstantinos G Derpanis, and Allan D Jepson. Representation learning via global temporal alignment and cycle-consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11068-11077, 2021. 6", + "[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 3", + "[21] Sepp Hochreiter, Yoshua Bengio, Paolo Frasconi, Jürgen Schmidhuber, et al. Gradient flow in recurrent nets: the difficulty of learning long-term dependencies, 2001. 3", + "[22] Zhenyu Hou, Xiao Liu, Yukuo Cen, Yuxiao Dong, Hongxia Yang, Chunjie Wang, and Jie Tang. Graphmae: Self-supervised masked graph autoencoders. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 594-604, 2022. 3, 4", + "[23] Simon Jenni and Hailin Jin. Time-equivariant contrastive video representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9970-9980, 2021. 7", + "[24] Gunnar Johansson. Visual perception of biological motion and a model for its analysis. Perception & psychophysics, 14(2):201-211, 1973. 1", + "[25] Gagan Kanojia, Sudhakar Kumawat, and Shanmuganathan Raman. Attentive spatio-temporal representation learning for diving classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 7" + ], + "bbox": [ + 502, + 116, + 902, + 875 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5615", + "bbox": [ + 468, + 891, + 500, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1725–1732, 2014. 6", + "[27] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6", + "[28] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 3, 5, 8", + "[29] Heeseung Kwon, Manjin Kim, Suha Kwak, and Minsu Cho. Learning self-similarity in space and time as generalized motion for video action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13065-13075, 2021. 7", + "[30] Mei Chee Leong, Hui Li Tan, Haosong Zhang, Liyuan Li, Feng Lin, and Joo Hwee Lim. Joint learning on the hierarchy representation for fine-grained human action recognition. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1059-1063. IEEE, 2021. 6", + "[31] Linguo Li, Minsi Wang, Bingbing Ni, Hang Wang, Jiancheng Yang, and Wenjun Zhang. 3d human action representation learning via cross-view consistency pursuit. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4741-4750, 2021. 2, 3, 7", + "[32] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3595–3603, 2019. 7", + "[33] Tianjiao Li, Lin Geng Foo, Qiuhong Ke, Hossein Rahmani, Anran Wang, Jinghua Wang, and Jun Liu. Dynamic spatio-temporal specialization learning for fine-grained action recognition. In European Conference on Computer Vision, pages 386-403. Springer, 2022. 2", + "[34] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018. 2, 6, 7", + "[35] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7083-7093, 2019. 6", + "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6", + "[37] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10):2684-2701, 2019. 2, 6" + ], + "bbox": [ + 67, + 116, + 468, + 875 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Mengyuan Liu, Hong Liu, and Chen Chen. Enhanced skeleton visualization for view invariant human action recognition. Pattern Recognition, 68:346-362, 2017. 5", + "[39] Yang Liu, Guanbin Li, and Liang Lin. Cross-modal causal relational reasoning for event-level visual question answering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1", + "[40] Yang Liu, Zhaoyang Lu, Jing Li, and Tao Yang. Hierarchically learned view-invariant representations for cross-view action recognition. IEEE Transactions on Circuits and Systems for Video Technology, 29(8):2416-2430, 2018. 2", + "[41] Yang Liu, Zhaoyang Lu, Jing Li, Tao Yang, and Chao Yao. Deep image-to-video adaptation and fusion networks for action recognition. IEEE Transactions on Image Processing, 29:3168-3182, 2019. 2", + "[42] Yang Liu, Keze Wang, Guanbin Li, and Liang Lin. Semantics-aware adaptive knowledge distillation for sensor-to-vision action recognition. IEEE Transactions on Image Processing, 30:5573-5588, 2021. 2", + "[43] Yang Liu, Keze Wang, Lingbo Liu, Haoyuan Lan, and Liang Lin. Tcgl: Temporal contrastive graph for self-supervised video representation learning. IEEE Transactions on Image Processing, 31:1978-1993, 2022. 3", + "[44] Yang Liu, Yu-Shen Wei, Hong Yan, Guan-Bin Li, and Liang Lin. Causal reasoning meets visual representation learning: A prospective study. Machine Intelligence Research, 19(6):485-511, 2022. 1", + "[45] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 143-152, 2020. 1, 7", + "[46] Chenxu Luo and Alan L Yuille. Grouped spatial-temporal aggregation for efficient action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5512-5521, 2019. 6, 7", + "[47] Fengjun Lv and Ramakant Nevatia. Recognition and segmentation of 3-d human action using hmm and multi-class adaboost. In European conference on computer vision, pages 359-372. Springer, 2006. 3", + "[48] Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2891-2900, 2017. 2", + "[49] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European conference on computer vision, pages 527-544. Springer, 2016. 6", + "[50] Jonathan Munro and Dima Damen. Multi-modal domain adaptation for fine-grained action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 122-132, 2020. 2", + "[51] Ashish S Nikam and Aarti G Ambekar. Sign language recognition using image based hand gesture recognition techniques. In 2016 online international conference on green engineering and technologies (IC-GET), pages 1-5. IEEE, 2016. 1" + ], + "bbox": [ + 502, + 116, + 900, + 875 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5616", + "bbox": [ + 468, + 891, + 500, + 902 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Cosmas Ifeanyi Nwakanma, Fabliha Bushra Islam, Mareska Pratiwi Maharani, Dong-Seong Kim, and JaeMin Lee. Iot-based vibration sensor data collection and emergency detection classification using long short term memory (lstm). In 2021 International Conference on Artificial Intelligence in Information and Communication (ICAIIC), pages 273-278. IEEE, 2021. 1", + "[53] Jiwong Park, Minsk Lee, Hyung Jin Chang, Kyuewang Lee, and Jin Young Choi. Symmetric graph convolutional autoencoder for unsupervised graph representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6519-6528, 2019. 4", + "[54] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 6", + "[55] Haocong Rao, Shihao Xu, Xiping Hu, Jun Cheng, and Bin Hu. Augmented skeleton based contrastive action learning with momentum lstm for unsupervised action recognition. Information Sciences, 569:90-109, 2021. 3", + "[56] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 6", + "[57] Amin Salehi and Hasan Davulcu. Graph attention autoencoders. arXiv preprint arXiv:1905.10715, 2019. 4", + "[58] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016. 2, 6", + "[59] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2616-2625, 2020. 1, 2, 6", + "[60] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12026-12035, 2019. 1, 3, 7", + "[61] Lei Shi, Yifan Zhang, Jing Hu, Jian Cheng, and Hanqing Lu. Gesture recognition using spatiotemporal deformable convolutional representation. In 2019 IEEE International Conference on Image Processing (ICIP), pages 1900-1904. IEEE, 2019. 1", + "[62] Sijie Song, Cuiling Lan, Junliang Xing, Wenjun Zeng, and Jiaying Liu. An end-to-end spatio-temporal attention model for human action recognition from skeleton data. In Proceedings of the AAAI conference on artificial intelligence, volume 31, 2017. 3", + "[63] Sijie Song, Cuiling Lan, Junliang Xing, Wenjun Zeng, and Jiaying Liu. Spatio-temporal attention-based LSTM networks for 3d action recognition and detection. IEEE Transactions on image processing, 27(7):3459-3471, 2018. 3", + "[64] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In International conference on machine learning, pages 843-852. PMLR, 2015. 2" + ], + "bbox": [ + 67, + 116, + 468, + 876 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Kun Su, Xiulong Liu, and Eli Shlizerman. Predict & cluster: Unsupervised skeleton based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9631-9640, 2020. 2, 3", + "[66] Swathikiran Sudhakaran, Sergio Escalera, and Oswald Lanz. Gate-shift networks for video action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1102-1111, 2020. 7", + "[67] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5693-5703, 2019. 1, 6", + "[68] Jian Tang, Meng Qu, Mingzhe Wang, Ming Zhang, Jun Yan, and Qiao zhu Mei. Line: Large-scale information network embedding. In Proceedings of the 24th international conference on world wide web, pages 1067-1077, 2015. 4", + "[69] Fida Mohammad Thoker, Hazel Doughty, and Cees GM Snoek. Skeleton-contrastive 3d action representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1655–1663, 2021. 1", + "[70] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022. 3", + "[71] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 2", + "[72] Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, Yoshua Bengio, et al. Graph attention networks. stat, 1050(20):10-48550, 2017. 3, 8", + "[73] Raviteja Vemulapalli, Felipe Arrate, and Rama Chellappa. Human action recognition by representing 3d skeletons as points in a lie group. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 588-595, 2014. 3", + "[74] Raviteja Vemulapalli and Rama Chellapa. Rolling rotations for recognizing human actions from 3d skeletal data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4471-4479, 2016. 3", + "[75] Heng Wang, Du Tran, Lorenzo Torresani, and Matt Feiszli. Video modeling with correlation networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 352-361, 2020. 7", + "[76] Jiang Wang, Zicheng Liu, Ying Wu, and Junsong Yuan. Mining actionlet ensemble for action recognition with depth cameras. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 1290-1297. IEEE, 2012. 3", + "[77] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 6", + "[78] Yanze Wang and Junyong Ye. Tmf: Temporal motion and fusion for action recognition. Computer Vision and Image Understanding, 213:103304, 2021. 7" + ], + "bbox": [ + 502, + 116, + 902, + 876 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "5617", + "bbox": [ + 468, + 891, + 500, + 902 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[79] Zhengwei Wang, Qi She, and Aljosa Smolic. Action-net: Multipath excitation for action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13214-13223, 2021. 2", + "[80] Wenhan Wu, Yilei Hua, Shiqian Wu, Chen Chen, Aidong Lu, et al. Skeletonmae: Spatial-temporal masked autoencoders for self-supervised skeleton action recognition. arXiv preprint arXiv:2209.02399, 2022. 3, 7", + "[81] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In Proceedings of the European conference on computer vision (ECCV), pages 305–321, 2018. 6", + "[82] Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826, 2018. 3, 8", + "[83] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Thirty-second AAAI conference on artificial intelligence, 2018. 1, 2, 3, 7", + "[84] Jie Yang, Chaoqun Wang, Zhen Li, Junle Wang, and Ruimao Zhang. Semantic human parsing via scalable semantic transfer over multiple label domains. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19424-19433, 2023. 1", + "[85] Jie Yang, Ailing Zeng, Shilong Liu, Feng Li, Ruimao Zhang, and Lei Zhang. Explicit box detection unifies end-to-end multi-person pose estimation. In The Eleventh International Conference on Learning Representations, 2023. 1", + "[86] LI Yang, Jin Huang, TIAN Feng, WANG Hong-An, and DAI Guo-Zhong. Gesture interaction in virtual reality. Virtual Reality & Intelligent Hardware, 1(1):84-112, 2019. 1", + "[87] Siyuan Yang, Jun Liu, Shijian Lu, Meng Hwa Er, and Alex C Kot. Skeleton cloud colorization for unsupervised 3d action representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13423-13433, 2021. 3, 7", + "[88] Yuning You, Tianlong Chen, Yang Shen, and Zhangyang Wang. Graph contrastive learning automated. In International Conference on Machine Learning, pages 12121-12132. PMLR, 2021. 8", + "[89] Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph contrastive learning with augmentations. Advances in neural information processing systems, 33:5812-5823, 2020. 8", + "[90] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Temporal query networks for fine-grained video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4486-4496, 2021. 6", + "[91] Nenggan Zheng, Jun Wen, Risheng Liu, Liangqu Long, Jianhua Dai, and Zhefeng Gong. Unsupervised representation learning with long-term dynamics for skeleton based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 2, 3", + "[92] Yuan Zhi, Zhan Tong, Limin Wang, and Gangshan Wu. Mgsampler: An explainable sampling strategy for video ac" + ], + "bbox": [ + 67, + 116, + 468, + 876 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1513-1522, 2021. 7", + "[93] Bolei Zhou, Alex Andonian, Aude Oliva, and Antonio Torralba. Temporal relational reasoning in videos. In Proceedings of the European conference on computer vision (ECCV), pages 803-818, 2018. 6", + "[94] Yujie Zhou, Haodong Duan, Anyi Rao, Bing Su, and Jiaqi Wang. Self-supervised action representation learning from partial spatio-temporal skeleton sequences. arXiv preprint arXiv:2302.09018, 2023. 7", + "[95] Wentao Zhu, Cuiling Lan, Junliang Xing, Wenjun Zeng, Yanghao Li, Li Shen, and Xiaohui Xie. Co-occurrence feature learning for skeleton based action recognition using regularized deep LSTM networks. In Proceedings of the AAAI conference on artificial intelligence, volume 30, 2016. 3" + ], + "bbox": [ + 502, + 116, + 900, + 328 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "5618", + "bbox": [ + 468, + 891, + 500, + 902 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2023/SkeletonMAE_ Graph-based Masked Autoencoder for Skeleton Sequence Pre-training/dcc19c04-02ae-4f8d-8f44-a86cdd91e8a4_model.json b/2023/SkeletonMAE_ Graph-based Masked Autoencoder for Skeleton Sequence Pre-training/dcc19c04-02ae-4f8d-8f44-a86cdd91e8a4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..aa5201bc72aa7fbdb66fd6fe9e2eb3b949506030 --- /dev/null +++ b/2023/SkeletonMAE_ Graph-based Masked Autoencoder for Skeleton Sequence Pre-training/dcc19c04-02ae-4f8d-8f44-a86cdd91e8a4_model.json @@ -0,0 +1,2866 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.12, + 0.038, + 0.191, + 0.074 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.248, + 0.036, + 0.798, + 0.077 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.152, + 0.864, + 0.196 + ], + "angle": 0, + "content": "SkeletonMAE: Graph-based Masked Autoencoder for Skeleton Sequence Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.22, + 0.854, + 0.272 + ], + "angle": 0, + "content": "Hong Yan\\(^{1*}\\) Yang Liu\\(^{1*†}\\) Yushen Wei\\(^{1}\\) Zhen Li\\(^{2}\\) Guanbin Li\\(^{1}\\) Liang Lin\\(^{1}\\) \n\\(^{1}\\)School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China \n\\(^{2}\\)The Chinese University of Hong Kong, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.274, + 0.813, + 0.305 + ], + "angle": 0, + "content": "{yanh36,weiysh8}@mail2.sysu.edu.cn，{liuy856,liguanbin}@mail.sysu.edu.cn, lizhen@cuhk.edu.cn，linliang@ieee.org" + }, + { + "type": "title", + "bbox": [ + 0.228, + 0.337, + 0.308, + 0.352 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.368, + 0.474, + 0.752 + ], + "angle": 0, + "content": "Skeleton sequence representation learning has shown great advantages for action recognition due to its promising ability to model human joints and topology. However, the current methods usually require sufficient labeled data for training computationally expensive models. Moreover, these methods ignore how to utilize the fine-grained dependencies among different skeleton joints to pre-train an efficient skeleton sequence learning model that can generalize well across different datasets. In this paper, we propose an efficient skeleton sequence learning framework, named Skeleton Sequence Learning (SSL). To comprehensively capture the human pose and obtain discriminative skeleton sequence representation, we build an asymmetric graph-based encoder-decoder pre-training architecture named SkeletonMAE, which embeds skeleton joint sequence into graph convolutional network and reconstructs the masked skeleton joints and edges based on the prior human topology knowledge. Then, the pre-trained SkeletonMAE encoder is integrated with the Spatial-Temporal Representation Learning (STRL) module to build the SSL framework. Extensive experimental results show that our SSL generalizes well across different datasets and outperforms the state-of-the-art self-supervised skeleton-based methods on FineGym, Diving48, NTU 60 and NTU 120 datasets. Moreover, we obtain comparable performance to some fully supervised methods. The code is available at https://github.com/HongYan1123/SkeletonMAE." + }, + { + "type": "title", + "bbox": [ + 0.067, + 0.778, + 0.201, + 0.793 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.066, + 0.802, + 0.469, + 0.845 + ], + "angle": 0, + "content": "Human action recognition has attracted increasing attention in video understanding [83, 7, 45, 60, 39], due to its wide applications [2, 14, 61, 51, 52, 86, 44] in human-" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.336, + 0.892, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.541, + 0.905, + 0.627 + ], + "angle": 0, + "content": "Figure 1: Traditional MAE usually uses random masking strategy to reconstruct skeleton, which tends to ignore action-sensitive skeleton regions. Differently, our proposed SkeletonMAE reconstructs the masked skeleton joints and edges based on the prior human topology knowledge, to obtain a comprehensive perception of the action." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.907, + 0.879 + ], + "angle": 0, + "content": "computer interaction, intelligent surveillance security, virtual reality, etc. In terms of visual perception [24], even without appearance information, humans can identify action categories by only observing the motion of joints. Unlike RGB videos [3, 14, 13], the skeleton sequences only contain the coordinate information of the key joints of the human body [84], which is high-level, light-weighted, and robust against complex backgrounds and various conditions including viewpoint, scale, and movement speed [11, 69]. Additionally, with the development of human pose estimation algorithms [8, 1, 85], the localization method of human joints (i.e., key points) has made great progress and it is feasible to obtain accurate skeleton sequences. At present, the existing 2D pose estimation method is more accurate and more robust than the 3D pose estimation methods [11]. In Figure 1 (a), we visualize 2D poses estimated with HRNet [67] for two action classes on FineGym dataset [59]. It" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.853, + 0.348, + 0.866 + ], + "angle": 0, + "content": "* Both authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.866, + 0.221, + 0.877 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.853, + 0.348, + 0.877 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.503, + 0.904 + ], + "angle": 0, + "content": "5606" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.065, + 0.116, + 0.469, + 0.143 + ], + "angle": 0, + "content": "can be seen that the 2D poses can accurately capture human skeletons and motion details." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.149, + 0.473, + 0.662 + ], + "angle": 0, + "content": "Due to the promising ability to model multiple granularities and large variations in human motion, the skeleton sequence is more suitable to distinguish similar actions with subtle differences than the RGB data. To capture discriminative spatial-temporal motion patterns, most of the existing skeleton-based action recognition methods [11, 83, 5] are fully supervised and usually require large amounts of labeled data for training elaborate models, which is time-consuming and labor-intensive. To mitigate the problem of limited labeled training data, self-supervised skeleton-based action recognition methods [31, 18, 65] have attracted increasing attention recently. Some contrastive learning methods [31, 18] adopted data augmentation to generate pairs of positive and negative samples, but they rely heavily on the number of contrastive pairs. With the popularity of the encoder-decoder [64, 48], some methods [91, 65] reconstructed the masked skeleton sequence by link reconstruction to encourage the topological closeness following the paradigm of graph encoder-decoder. However, these methods are usually good at link prediction and node clustering but are unsatisfactory in node and graph classifications. For accurate action recognition, the fine-grained dependencies among different skeleton joints (i.e., graph classifications) are essential. Therefore, previous self-supervised learning-based methods tend to ignore the fine-grained dependencies among different skeleton joints, which restricts the generalization of self-supervised skeleton representation. As shown in Figure 1 (b)-(c), the arm joints and edges are essential to discriminate between these two similar actions. Different from the randomly masking strategy of MAE [20], our masking strategy is action-sensitive and reconstructs specific limbs or body parts that dominate the given action class. Our SkeletonMAE utilizes prior human topology knowledge to guide the reconstruction of the masked skeleton joints and edges to achieve a comprehensive perception of the joints, topology, and action." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.665, + 0.47, + 0.879 + ], + "angle": 0, + "content": "To address the aforementioned challenges, we propose an efficient skeleton sequence representation learning framework, named Skeleton Sequence Learning (SSL). To fully discover the fine-grained dependencies among different skeleton joints, we build a novel asymmetric graph-based encoder-decoder pre-training architecture named SkeletonMAE that embeds skeleton joint sequences in Graph Convolutional Network (GCN). The SkeletonMAE aims to reconstruct the masked human skeleton joints and edges based on prior human topology knowledge so that it can infer the underlying topology of the joints and obtain a comprehensive perception of human action. To learn discriminative spatial-temporal skeleton representation, the pre-trained SkeletonMAE encoder is integrated with the Spatial-Temporal Representation Learning (STRL)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.116, + 0.905, + 0.203 + ], + "angle": 0, + "content": "module to learn spatial-temporal dependencies. Finally, the SSL is fine-tuned on action recognition datasets. Extensive experimental results on FineGym, Diving48, NTU 60 and NTU 120 show that our SSL generalizes well across different datasets and outperforms the state-of-the-art methods significantly. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.215, + 0.902, + 0.314 + ], + "angle": 0, + "content": "- To comprehensively capture human pose and obtain discriminative skeleton sequence representation, we propose a graph-based encoder-decoder pre-training architecture named SkeletonMAE, that embeds skeleton joint sequence into GCN and utilize the prior human topology knowledge to guide the reconstruction of the underlying masked joints and topology." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.315, + 0.902, + 0.399 + ], + "angle": 0, + "content": "- To learn comprehensive spatial-temporal dependencies for skeleton sequence, we propose an efficient skeleton sequence learning framework, named Skeleton Sequence Learning (SSL), which integrates the pre-trained SkeletonMAE encoder with the Spatial-Temporal Representation Learning (STRL) module." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.4, + 0.902, + 0.484 + ], + "angle": 0, + "content": "- Extensive experimental results on FineGym, Diving48, NTU 60 and NTU 120 datasets show that our SSL outperforms the state-of-the-art self-supervised skeleton-based action recognition methods and achieves comparable performance compared with the state-of-the-art fully supervised methods." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.215, + 0.902, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.501, + 0.646, + 0.517 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.536, + 0.905, + 0.879 + ], + "angle": 0, + "content": "Action Recognition. One of the most challenging tasks for action recognition is to distinguish similar actions from subtle differences [40, 41, 42]. Recently, some challenging action recognition datasets like FineGym [59], Diving48 [34], NTU RGB+D 60 [58] and NTU RGB+D 120 [37] are proposed. These datasets contain a large number of challenging actions that require discriminative and fine-grained action representation learning. For example, in FineGym [59], an action is divided into action units, subactions, or phases, and the model is required to distinguish between \"split leap with 1 turn\" and \"switch leap with 1 turn\". The higher inter-class similarity and a new level of granularity in the fine-grained setting make it a challenging task, which makes coarse-grained backbones and methods [14, 2, 71, 79] struggle to overcome. To tackle the more challenging fine-grained action recognition task, most of the existing works [50, 33] are fully supervised and consider fine-grained actions as distinct categories and supervise the model to learn action semantics. However, collecting and labeling these fine-grained actions is time-consuming and labor-intensive, which limits the generalization of a well-trained model to different datasets. To utilize unlabeled data, we propose a graph-based encoder-decoder pre-training architecture named SkeletonMAE." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.502, + 0.904 + ], + "angle": 0, + "content": "5607" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.065, + 0.116, + 0.473, + 0.514 + ], + "angle": 0, + "content": "Skeleton-based Action Recognition. Due to the promising ability to model multiple granularities and large variations in human motion, the skeleton data is more suitable for the aforementioned action recognition task than the RGB data [4]. Early skeleton-based action recognition methods are usually handcrafted, exploiting the geometric relationship of skeleton joints [47, 73, 76, 74], which greatly limits the feature representation of skeletons. Benefiting from the advantages of deep learning, some methods [95, 62, 63] utilized RNNs as the basic model, Du et al. [10] presented a pioneering work based on hierarchical RNNs. But RNNs easily suffer from vanishing gradients [21]. Inspired by the booming Graph Convolutional Networks (GCN) [28], Yan et al. [83] proposed a spatial-temporal graph convolutional network to learn the spatial and temporal pattern from skeleton data. However, their manually defined topology is arduous to model the relations among joints in underlying topology. Chen et al. [5] proposed a channel-wise topology graph convolution, which models channel-wise topology with a refinement method. Duan et al. [11] proposed a PoseConv3D model that relies on a 3D heatmap volume instead of a graph sequence as the base representation of human skeletons. Different from previous methods that required large amounts of labeled data for training elaborate models, we utilize unlabeled skeleton sequences to pretrain a graph-based encoder-decoder named SkeletonMAE to comprehensively capture human pose and obtain discriminative skeleton sequence representation." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.523, + 0.473, + 0.866 + ], + "angle": 0, + "content": "Self-supervised Learning for Skeleton Sequence. To learn more effective representation for unlabeled skeleton data, self-supervised learning [43] has achieved inspiring progress recently. For contrastive learning approaches, ASCAL [55] and SkeletonCLR [31] applied momentum encoders for contrastive learning with single-stream skeleton sequences. AimCLR [18] used an extreme data augmentation strategy to add additional hard contrastive pairs. Most contrastive learning methods adopt data augmentation to generate positive and negative pairs, but they rely heavily on the number of contrastive pairs. For generative learning approaches, LongT GAN [91] proposed the encoder-decoder to reconstruct masked input sequence skeletons. Based on the LongT GAN, P&C [65] strengthened the encoder and weakened the decoder for feature representation. Wu et al. [80] proposed a spatial-temporal masked autoencoder framework for self-supervised 3D skeleton-based action recognition. Colorization [87] used three pairs of encoder-decoder frameworks to learn spatial-temporal features from skeleton point clouds. Due to the limitation of the reconstruction criterion, previous generative methods usually fail to fully discover the fine-grained spatial-temporal dependencies among different skeleton joints. In our SkeletonMAE, we utilize the prior human topology" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.116, + 0.905, + 0.147 + ], + "angle": 0, + "content": "knowledge to infer the skeleton sequence and obtain a comprehensive perception of the action." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.157, + 0.64, + 0.174 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.181, + 0.905, + 0.239 + ], + "angle": 0, + "content": "In this section, we introduce the details of Skeleton Sequence Learning (SSL), which contains two parts: 1) pretraining SkeletonMAE and 2) fine-tuning on downstream datasets based on the pre-trained SkeletonMAE." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.247, + 0.753, + 0.263 + ], + "angle": 0, + "content": "3.1. Pre-training SkeletonMAE" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.269, + 0.905, + 0.398 + ], + "angle": 0, + "content": "In this section, we introduce graph-based asymmetric encoder-decoder pre-training architecture SkeletonMAE, to learn human skeleton sequence representations without supervision. Since Graph Isomorphism Network (GIN) [82] provides a better inductive bias, it is more suitable for learning more generalized self-supervised representation [22]. Therefore, we adopt GIN as the backbone of SkeletonMAE. Besides, we evaluate different backbones of SkeletonMAE in Table 4, including GIN [82], GCN [28], and GAT [72]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.415, + 0.731, + 0.429 + ], + "angle": 0, + "content": "3.1.1 SkeletonMAE Structure" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.437, + 0.904, + 0.55 + ], + "angle": 0, + "content": "Inspired by the effective representation learning by masked autoencoder (MAE) in NLP [9], image recognition [20], and video recognition [70], we focus on the human skeleton sequence and build an asymmetric graph-based encoder-decoder pre-training architecture named SkeletonMAE that embeds skeleton sequence and its prior topology knowledge in GIN. The SkeletonMAE is implemented following the paradigm of graph generative self-supervised learning." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.551, + 0.905, + 0.735 + ], + "angle": 0, + "content": "We follow the joint label of the Kinetics Skeleton dataset [60]. Specifically, as Figure 2(d) shown, we divide all \\(N = 17\\) joints into \\(R = 6\\) local regions according to the natural parts of the body: \\(\\nu_{0},\\dots,\\nu_{5}\\). Notably, compared to the randomly masking strategy from MAE [20] to select skeleton joints, our masking strategy is action-sensitive and reconstructs specific limbs or body parts that dominate the given action class. Then, we mask these skeleton regions and make the SkeletonMAE reconstruct the masked joint features and their edges based on the adjacent joints. By reconstructing the masked skeleton joints and edges, the SkeletonMAE can infer the underlying topology of joints and obtain a comprehensive perception of the action." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.736, + 0.905, + 0.879 + ], + "angle": 0, + "content": "As shown in Figure 2, the SkeletonMAE is an asymmetric encoder-decoder architecture, which includes an encoder and a decoder. The encoder consists of \\( L_{D} \\) GIN layers that map the input 2D skeleton data to hidden features. The decoder, which consists of only one GIN layer, reconstructs the hidden features under the supervision of the reconstruction criterion. According to the prior human skeleton knowledge that the human skeleton can be represented as a graph with joints as vertices and limbs as edges, we formulate the human skeleton as the following graph structure." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.503, + 0.904 + ], + "angle": 0, + "content": "5608" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.068, + 0.114, + 0.905, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.063, + 0.281, + 0.908, + 0.37 + ], + "angle": 0, + "content": "Figure 2: The details of our skeleton sequence pre-training architecture SkeletonMAE. (a) We build a GIN-based asymmetric encoder-decoder structure, to reconstruct joint features to enhance action representation ability. (b) The GIN-based encoder structure contains \\(L_{D}\\) GIN neural network layers, to learn the joint representation spatially. (c) The decoder consists of one GIN layer, which uses the hidden features from the encoder as the input and reconstructs the original input joint features. (d) Partition the joints in the skeleton sequence according to the natural structure of the human body. 5 joints \\(\\{\\mathcal{V}_0:\\mathrm{Head}\\}\\), 4 joints \\(\\{\\mathcal{V}_1:\\mathrm{Torso}\\}\\), 2 joints \\(\\{\\mathcal{V}_2:\\mathrm{Left arm},\\mathcal{V}_3:\\mathrm{Right arm},\\mathcal{V}_4:\\mathrm{Left leg},\\mathcal{V}_5:\\mathrm{Right leg}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.392, + 0.473, + 0.622 + ], + "angle": 0, + "content": "The skeleton sequence of two-dimensional coordinates of \\(N\\) human skeleton joints and \\(T\\) frames is pre-processed in the following way. Specifically, we embed all skeleton joints and their topology into a structure \\(\\mathcal{G}\\), the skeleton structure and the joint feature are fused to obtain a joint sequence matrix \\(\\mathbf{S} \\in \\mathbb{R}^{N \\times T \\times 2}\\). And then the \\(\\mathbf{S}\\) is linearly transformed to \\(\\mathbf{S} \\in \\mathbb{R}^{N \\times T \\times D}\\) with learnable parameters. We empirically set \\(\\mathrm{T}\\) and \\(\\mathrm{D}\\) to 64. For each skeleton frame \\(\\mathbf{X} \\in \\mathbb{R}^{N \\times D}\\) from \\(\\mathbf{S}\\), let \\(\\mathcal{G} = (\\mathcal{V}, \\mathbf{A}, \\mathbf{X})\\) denote a skeleton, where \\(\\mathcal{V} = \\{v_1, v_2, \\dots, v_N\\}\\) is the node set that contains all skeleton joints, \\(N = |\\mathcal{V}|\\) is the number of joints. The number of joints is \\(N = 17\\). \\(\\mathbf{A} \\in \\{0, 1\\}^{N \\times N}\\) is an adjacency matrix, where \\(\\mathbf{A}_{i,j} = 1\\) if joints \\(i\\) and \\(j\\) are physically connected, otherwise 0. The feature of \\(v_i\\) is represented as \\(\\mathbf{x}_i \\in \\mathbb{R}^{1 \\times D}\\). And \\(G_E\\), \\(G_D\\) denote the GIN encoder and the GIN decoder, respectively." + }, + { + "type": "title", + "bbox": [ + 0.065, + 0.64, + 0.442, + 0.655 + ], + "angle": 0, + "content": "3.1.2 Skeleton Joints Masking and Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.064, + 0.663, + 0.47, + 0.777 + ], + "angle": 0, + "content": "Since the prior human skeleton topology \\(\\mathbf{A}\\) is embedded (Figure 2) and we specify the aggregation of joints in Section 3.1.1. Inspired by the GraphMAE [22] that randomly reconstructs the masked graph nodes, our SkeletonMAE reconstructs the masked skeleton feature \\(\\mathbf{X}\\) based on the prior skeleton topology, rather than reconstructing graph structure \\(\\mathbf{A}\\) [68, 17] or reconstructing both structure \\(\\mathbf{A}\\) and features \\(\\mathbf{X}\\) [57, 53]." + }, + { + "type": "text", + "bbox": [ + 0.064, + 0.778, + 0.47, + 0.879 + ], + "angle": 0, + "content": "To mask skeleton joint features, we randomly select one or more joint sets from \\(\\mathcal{V} = \\{\\mathcal{V}_0,\\dots,\\mathcal{V}_5\\}\\), which consists of a subset \\(\\overline{\\mathcal{V}}\\subseteq \\mathcal{V}\\) for masking. For the human skeleton sequence, each joint communicates with some of its adjacent joints to represent the specific action class. Therefore, it is not feasible to mask all joint sets for all action classes. Then, each of their features is masked with a learnable mask" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.391, + 0.907, + 0.553 + ], + "angle": 0, + "content": "token vector \\([\\mathbf{MASK}] = \\mathbf{x}_{[\\mathbf{M}]} \\in \\mathbb{R}^D\\). Thus, the masked joint feature \\(\\overline{\\mathbf{x}}_i\\) for \\(\\mathbf{v}_i \\in \\overline{\\mathcal{V}}\\) in the masked feature matrix \\(\\overline{\\mathbf{X}}\\) can be defined as \\(\\overline{\\mathbf{x}}_i = \\mathbf{x}_{[\\mathbf{M}]}\\) if \\(\\mathbf{v}_i \\in \\overline{\\mathcal{V}}\\), otherwise \\(\\overline{\\mathbf{x}}_i = \\mathbf{x}_i\\). We set \\(\\overline{\\mathbf{X}} \\in \\mathbb{R}^{N \\times D}\\) as the input joint feature matrix of the SkeletonMAE, and each joint feature in \\(\\overline{\\mathbf{X}}\\) can be defined as \\(\\overline{\\mathbf{x}}_i = \\{\\mathbf{x}_{[\\mathbf{M}]}, \\mathbf{x}_i\\}\\), \\(i = 1, 2, \\dots, N\\). Therefore, the masked skeleton sequence can be formulated as \\(\\overline{\\mathcal{G}} = (\\mathcal{V}, \\mathbf{A}, \\overline{\\mathbf{X}})\\) and the objective of SkeletonMAE is to reconstruct the masked skeleton features in \\(\\overline{\\mathcal{V}}\\) given the partially observed joint features \\(\\overline{\\mathbf{X}}\\) with the input adjacency matrix \\(\\mathbf{A}\\). The process of SkeletonMAE reconstruction is formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.558, + 0.903, + 0.574 + ], + "angle": 0, + "content": "\\[\n\\int \\mathbf {H} = G _ {E} (\\mathbf {A}, \\overline {{\\mathbf {X}}}), \\quad \\mathbf {H} \\in \\mathbb {R} ^ {N \\times D _ {h}} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.574, + 0.903, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathbf {Y} = G _ {D} (\\mathbf {A}, \\mathbf {H}), \\quad \\mathbf {Y} \\in \\mathbb {R} ^ {N \\times D} \\right., \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.906, + 0.641 + ], + "angle": 0, + "content": "where \\(\\mathbf{H}\\) and \\(\\mathbf{Y}\\) denote the encoder output and the decoder output, respectively. The objective of SkeletonMAE can be formalized as minimizing the divergence between \\(\\mathbf{X}\\) and \\(\\mathbf{Y}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.656, + 0.738, + 0.67 + ], + "angle": 0, + "content": "3.1.3 Reconstruction Criterion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.678, + 0.906, + 0.82 + ], + "angle": 0, + "content": "The common reconstruction criterion for masked autoencoders is a mean squared error (MSE) in image and video tasks. However, for skeleton sequence, the multidimensional and continuous nature of joint features makes MSE hard to achieve promising feature reconstruction because the MSE is sensitive to dimensionality and vector norms of features [15]. Inspired by the observation [16] that the \\(l_{2}\\)-normalization in the cosine error maps vectors to a unit hyper-sphere and substantially improves the training stability, we utilize the cosine error as the reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.82, + 0.905, + 0.879 + ], + "angle": 0, + "content": "To make the reconstruction criterion focus on harder ones among imbalanced easy-and-hard samples [22], we use the Re-weighted Cosine Error (RCE) for SkeletonMAE. The RCE is based on the intuition that we can down-weigh" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.503, + 0.904 + ], + "angle": 0, + "content": "5609" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.065, + 0.115, + 0.469, + 0.214 + ], + "angle": 0, + "content": "easy samples' contribution in training by scaling the cosine error with a power of \\(\\beta \\geq 1\\). For predictions with high confidence, their corresponding cosine errors are usually smaller than 1 and decay faster to zero when the scaling factor \\(\\beta > 1\\). Formally, given the original feature \\(\\mathbf{X} \\in \\mathbb{R}^{N \\times D}\\) and the reconstructed output \\(\\mathbf{Y} \\in \\mathbb{R}^{N \\times D}\\), the RCE is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.221, + 0.469, + 0.26 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {R C E}} = \\sum_ {\\mathbf {v} _ {i} \\in \\overline {{\\mathcal {V}}}} \\left(\\frac {1}{| \\overline {{\\mathcal {V}}} |} - \\frac {\\mathbf {x} _ {i} ^ {\\mathrm {T}} \\cdot \\mathbf {z} _ {i}}{| \\overline {{\\mathcal {V}}} | \\times \\| \\mathbf {x} _ {i} \\| \\times \\| \\mathbf {z} _ {i} \\|}\\right) ^ {\\beta}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.268, + 0.469, + 0.31 + ], + "angle": 0, + "content": "which represents the average of the similarity gap between the reconstructed feature and the input feature over all masked joints. And \\(\\beta\\) is set to 2 in our work." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.311, + 0.468, + 0.397 + ], + "angle": 0, + "content": "By training the SkeletonMAE to reconstruct the skeleton sequence, the pre-trained SkeletonMAE can comprehensively perceive the human skeleton structure and obtain discriminative action representation. After pre-training, the SkeletonMAE can be elegantly embedded into the Skeleton Sequence Learning (SSL) framework for fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.066, + 0.404, + 0.45, + 0.419 + ], + "angle": 0, + "content": "3.2. Fine-tuning for Skeleton Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.426, + 0.469, + 0.581 + ], + "angle": 0, + "content": "To evaluate the SkeletonMAE's generalization ability for skeleton action recognition, we build a complete skeleton action recognition model named Skeleton Sequence Learning (SSL), based on the pre-trained SkeletonMAE. To capture multiple-person interaction, we integrate two pre-trained SkeletonMAE encoders to build the Spatial-Temporal Representation Learning (STRL) module, as shown in Figure 3(b)-(c). The entire SSL consists of an \\(M\\)-layer STRL model and a classifier. The SSL model is finally fine-tuned on skeleton action recognition datasets with cross-entropy loss to recognize actions." + }, + { + "type": "title", + "bbox": [ + 0.065, + 0.599, + 0.43, + 0.615 + ], + "angle": 0, + "content": "3.2.1 Spatial-Temporal Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.621, + 0.468, + 0.678 + ], + "angle": 0, + "content": "The STRL contains two pre-trained SkeletonMAE encoders for Spatial Modeling (SM). The input of SM is skeleton sequence S. The output of SM is connected with the input by \\(1\\times 1\\) convolution for residual connection (Figure 3 (b))." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.679, + 0.469, + 0.833 + ], + "angle": 0, + "content": "As shown in Figure 3 (c), the input skeleton sequence \\(\\mathbf{S} \\in \\mathbb{R}^{N \\times T \\times D}\\) is firstly added with the learnable temporal position embedding PE to obtain the skeleton sequence feature \\(\\mathbf{H}_t^{(l)} \\in \\mathbb{R}^{P \\times N \\times D^{(l)}}\\). To model multiple human skeleton interactions, we obtain two individual features (\\(P = 2\\)) for two persons \\(\\mathbf{H}_{t,0}^{(l)} \\in \\mathbb{R}^{N \\times D^{(l)}}\\) and \\(\\mathbf{H}_{t,1}^{(l)} \\in \\mathbb{R}^{N \\times D^{(l)}}\\) from \\(\\mathbf{H}_t^{(l)}\\). Here, we take the joint feature of the 0-th person as an example, the operation of the 1-th person is implemented similarly. We send the joint representation \\(\\mathbf{H}_{t,0}^{(l)}\\) and prior knowledge of the joint \\(\\widetilde{\\mathbf{A}}\\) into the SM module," + }, + { + "type": "equation", + "bbox": [ + 0.082, + 0.841, + 0.468, + 0.879 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S M} \\left(\\mathbf {H} _ {t, 0} ^ {(l)}\\right) = \\operatorname {R e p e a t} \\left(\\operatorname {S P} \\left(G _ {E} \\left(\\widetilde {\\mathbf {A}}, \\mathbf {H} _ {t, 0} ^ {(l)}\\right)\\right); N\\right) \\oplus \\mathbf {H} _ {t, 0} ^ {(l)}, \\tag {3}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.114, + 0.772, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.115, + 0.906, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.36, + 0.903, + 0.488 + ], + "angle": 0, + "content": "Figure 3: The pipeline of Skeleton Sequence Learning (SSL). (a) During pre-training, we build an encoder-decoder module named SkeletonMAE that embeds skeleton joints and its prior topology knowledge into GIN and reconstructs the underlying masked joints and topology. (b) The SM consists of the pre-trained SkeletonMAE encoder. (c) We integrate SM structures to build the \\(M\\)-layer Spatial-Temporal Representation Learning (STRL) model and then conduct end-to-end fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.903, + 0.747 + ], + "angle": 0, + "content": "where \\(G_{E}\\) is the SkeletonMAE encoder, \\(\\mathrm{SP}(\\cdot)\\) denotes the sum-pooling, \\(\\mathrm{Repeat}(\\cdot;N)\\) means repeating the single joint into \\(N\\) joints representations after sum-pooling and then connect it with the \\(\\mathbf{H}_{t,0}^{(l)}\\) residual to get the global joint representation \\(\\mathrm{SM}(\\mathbf{H}_{t,0}^{(l)})\\). In this way, the SM module can obtain global information through a single joint representation, and constrain some joint features through all joint representations. Similarly, \\(\\mathrm{SM}(\\mathbf{H}_{t,1}^{(l)})\\) is obtained in the same way. As shown in Figure 3(c), we get the joint features \\(\\mathrm{SM}(\\mathbf{H}_t^{(l)})\\) that contains the action interaction bewteen the 0-th person and the 1-th person. According to the update rules of graph convolution [28], we can obtain \\(\\mathbf{H}_t^{(l + 1)}\\) from \\(\\mathbf{H}_t^{(l)}\\) in a multi-layer GCN. For more details, please refer to the Supplementary in Section D. The final skeleton sequence representation is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.764, + 0.903, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {t} ^ {(l + 1)} = \\sigma (\\mathbf {S M} (\\mathbf {H} _ {t} ^ {(l)}) \\mathbf {W} ^ {(l)}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.805, + 0.905, + 0.879 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}^{(l)}\\) denotes the trainable weight matrix in the \\(l^{th}\\) layer, \\(\\sigma(\\cdot)\\) denotes the ReLU activation function. Then, we adopt the widely-used multi-scale temporal pooling [5, 38] to get the final output. Finally, a classifier consisting of MLP and softmax predicts the action class." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.502, + 0.904 + ], + "angle": 0, + "content": "5610" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.066, + 0.115, + 0.201, + 0.131 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.066, + 0.138, + 0.468, + 0.167 + ], + "angle": 0, + "content": "All experiments are conducted with a single modality (2D pose) and evaluated on the corresponding train/test sets." + }, + { + "type": "title", + "bbox": [ + 0.066, + 0.173, + 0.17, + 0.187 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.195, + 0.468, + 0.38 + ], + "angle": 0, + "content": "We evaluate our SSL on four benchmark datasets FineGym [59], Diving48 [34], NTU RGB+D 60 [58] and NTU RGB+D 120 [37] in the mainstream skeleton action recognition task. For all datasets except FineGym, we follow the pre-processing protocol provided by [11] to obtain the skeleton sequence from the 2D pose estimator. The pre-processing adopts a Top-Down approach for pose extraction, where the detector is Faster-RCNN [56] with ResNet50 backbone and the pose estimator is HRNet [67] pre-trained on COCO-keypoint [36]. To make a fair comparison, we added pixel noise to the joint during training, making the original joint confidence rate unreliable, thus we do not use the originally fixed threshold." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.389, + 0.468, + 0.503 + ], + "angle": 0, + "content": "FineGym is a large-scale fine-grained action recognition dataset for gymnastic videos, which contains 29K videos of 99 fine-grained gymnastics action classes, which requires action recognition methods to distinguish different sub-actions in the same video. In particular, it provides temporal annotations at both action and sub-action levels with a three-level semantic hierarchy. We follow the method [11] to extract the skeleton data from the 2D pose estimator." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.512, + 0.468, + 0.669 + ], + "angle": 0, + "content": "NTU RGB+D 60 and 120. NTU RGB+D is a large-scale skeleton-based action recognition dataset, where NTU 60 contains 56,880 skeleton sequences and 60 action classes. NTU 120 has 114,480 skeleton sequences and 120 action categories. The action samples are captured from 155 different camera viewpoints. The subjects in NTU 120 are in a wide range of age distribution (from 10 to 57) and from different cultural backgrounds (15 countries), which brings very realistic variation to the quality of actions. The NTU 60 and 120 datasets have a large amount of variation in subjects, views, and backgrounds." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.678, + 0.468, + 0.806 + ], + "angle": 0, + "content": "Diving48 is a challenging fine-grained dataset that focuses on complex and competitive sports content analysis. It is formed of over 18k video clips from competitive dive championships that are distributed over 48 fine-grained classes with minimal biases. The difficulties of the dataset lie in that actions are similar and differ in body parts and their combinations which require the model to capture details and motion in body parts and combine them to perform classification. We report the accuracy on the official train/test split." + }, + { + "type": "title", + "bbox": [ + 0.066, + 0.813, + 0.288, + 0.827 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.835, + 0.468, + 0.877 + ], + "angle": 0, + "content": "In this paper, our SkeletonMAE is optimized by the Adaptive Moment Estimation (Adam) with the initial learning rate as \\(1.5e^{-4}\\) and the PReLU is the activation function." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.113, + 0.897, + 0.365 + ], + "angle": 0, + "content": "
MethodModalityMean Acc. (%)
Fully Supervised
I3D [2]RGB64.4
ST-S3D [81]RGB72.9
TSN [77]RGB+Flow79.8
TRNms [93]RGB+Flow80.2
TSM [35]RGB+Flow81.2
GST-50 [46]RGB84.6
MTN [30]RGB88.5
LT-S3D [81]RGB88.9
TQN [90]RGB+Text90.6
PYSKL [11]Skeleton93.2
PYSKL [11]RGB+Skeleton+Limb95.6
Unsupervised Pre-train
SaL [49]RGB42.7
TCC [12]RGB45.6
GTA [19]RGB49.5
CARL [3]RGB60.4
SSL (ours)Skeleton91.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.375, + 0.902, + 0.404 + ], + "angle": 0, + "content": "Table 1: The comparison with the state-of-the-art unsupervised pre-train and supervised methods on FineGym." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.43, + 0.902, + 0.614 + ], + "angle": 0, + "content": "The batch size is 1024 and the training epoch is 50. At the fine-tuning stage, we use the Stochastic Gradient Descent (SGD) with momentum (0.9) and adopt the warm-up strategy for the first 5 epochs. The total fine-tuning epochs are 110. The learning rate is initialized to 0.1 and is divided by 10 at the 90 epoch and the 100 epoch. And we employ 0.1 for label smoothing. We use a large batch size of 128 to facilitate training our attention mechanism and enhancing the model's perception for all human action classes. Both our pre-training and fine-tuning models are implemented by PyTorch [54], and our SSL is trained on a single NVIDIA GeForce RTX 2080Ti GPU. For more details of implementation, please refer to the Supplementary in Section A." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.62, + 0.902, + 0.678 + ], + "angle": 0, + "content": "Pre-training and Fine-tuning Setting. For each dataset, the SkeletonMAE encoder is pre-trained with unlabeled data from the training set. Then, we load the learned parameter weights to fine-tune the SSL model." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.683, + 0.902, + 0.74 + ], + "angle": 0, + "content": "Evaluation Metrics. To make a fair comparison, we follow previous methods [11, 3, 18] and report the Mean Top-1 accuracy \\((\\%)\\) on FineGym dataset and Top-1 accuracy \\((\\%)\\) on Diving48, NTU 60, and NTU 120 datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.75, + 0.727, + 0.763 + ], + "angle": 0, + "content": "4.3. Downstream Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.772, + 0.902, + 0.843 + ], + "angle": 0, + "content": "For a fair comparison, we compare SSL with other models without pre-training on large-scale action datasets, e.g., Kinetics [27] or Sports1M [26]. The comparison results on FineGym, NTU 60 & 120, and Diving48 datasets are shown in Table 1, Table 2, and Table 3, respectively." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.848, + 0.902, + 0.877 + ], + "angle": 0, + "content": "Results on FineGym Dataset. In Table 1, our SSL with skeleton input outperforms most of the fully supervised" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.501, + 0.903 + ], + "angle": 0, + "content": "5611" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.069, + 0.113, + 0.904, + 0.35 + ], + "angle": 0, + "content": "
MethodBackboneSupervisionJoint Number2D SkeletonNTU 60NTU 120
X-sub (%)X-view (%)X-sub (%)X-set (%)
ST-GCN [83]GCNFully Supervised25X81.588.3--
AS-GCN [32]GCNFully Supervised25X86.894.2--
2s-AGCN [60]GCNFully Supervised25X88.595.182.984.9
Shift-GCN [6]GCNFully Supervised25X90.796.585.987.6
MS-G3D [45]GCNFully Supervised25X91.596.286.988.4
CTR-GCN [5]GCNFully Supervised25X92.496.888.990.6
PYSKL [11]CNNFully Supervised17✓93.796.686.089.6
SkeletonCLR [31]ST-GCNUnsupervised Pre-train25X82.288.973.675.3
CrossCLR [31]ST-GCNUnsupervised Pre-train25X86.292.580.580.4
Wu et al. [80]STTFormerUnsupervised Pre-train25X86.692.976.879.1
AimCLR [18]ST-GCNUnsupervised Pre-train25X86.992.880.180.9
3s-PSTL [94]ST-GCNUnsupervised Pre-train25X87.193.981.382.6
Colorization [87]DGCNNUnsupervised Pre-train25X88.094.9--
SSL(ours)STRLUnsupervised Pre-train17✓92.8(↑ 4.8)96.5(↑ 1.6)84.8(↑ 3.5)85.7(↑ 3.1)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.065, + 0.358, + 0.905, + 0.389 + ], + "angle": 0, + "content": "Table 2: The comparison with state-of-the-art unsupervised pre-train and supervised methods on NTU 60 and NTU 120 datasets. \\(\\underline{\\mathbf{\\Pi}}_{-}^{\\prime}\\) means the method with the second-best performance under unsupervised pre-training." + }, + { + "type": "table", + "bbox": [ + 0.068, + 0.409, + 0.47, + 0.624 + ], + "angle": 0, + "content": "
MethodPre-trainGFLOPsAcc.(%)
Fully Supervised
TSN [34]ImageNet-16.8
TRN [25]ImageNet-22.8
P3D [46]ImageNet-32.4
C3D [46]ImageNet-34.5
CorrNet [46]-74.835.5
CorrNet-R101 [75]ImageNet187.338.2
MG-TEA-ResNet50 [92]ImageNet-39.5
GSM [66]ImageNet107.440.3
TSM-R50 [29]ImageNet153.841.6
TMF [78]ImageNet-42.2
Unsupervised Pre-train
RESOUND-C3D [34]K400-16.4
Jenni et al. [23]K400-29.9
SSL (ours)Diving4842.834.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.065, + 0.632, + 0.47, + 0.662 + ], + "angle": 0, + "content": "Table 3: The comparison with the unsupervised pre-train and supervised methods on the Diving48 dataset." + }, + { + "type": "text", + "bbox": [ + 0.064, + 0.687, + 0.469, + 0.801 + ], + "angle": 0, + "content": "methods and achieves the best performance among unsupervised pre-train methods with RGB input. For the same input modality, our performance is lower than the fully supervised method PYSKL[11] (with the skeleton as input) by about \\(1.4\\%\\), because the PYSKL adopts stacks of visual heatmaps of skeleton joints as input while we only use human skeleton coordinates. This validates the promising discriminative ability of our skeleton sequence representation." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.806, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Results on NTU 60 and NTU 120 Datasets. In Table 2, for NTU 60 X-sub and NTU 60 X-view, compared with unsupervised pre-train methods, our SSL outperforms the current state-of-the-art method Colorization [87] by \\(4.8\\%\\) and \\(1.6\\%\\), respectively. Our SSL is also competitive compared" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.905, + 0.554 + ], + "angle": 0, + "content": "with fully supervised methods, outperforming the first six fully supervised methods on NTU 60 X-sub. For NTU 120 X-sub and NTU 120 X-set, our SSL outperforms the previous best-unsupervised pre-train method 3s-PSTL [94] by \\(3.5\\%\\) and \\(3.1\\%\\), respectively. Our SSL is superior compared with some fully supervised methods on NTU 120 X-sub and NTU 120 X-set. These results show that our SSL can learn discriminative skeleton representation from large-scale action recognition datasets due to the promising generalization ability of our SkeletonMAE." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.904, + 0.661 + ], + "angle": 0, + "content": "Results on Diving48 Dataset. Our SSL with skeleton input outperforms some fully supervised methods. Although our SSL is not pre-trained on additional large-scale pretraining action datasets in Table 3, it still achieves competitive performance among unsupervised pre-train methods. This validates that our pre-training model SkeletonMAE can learn discriminative skeleton sequence representation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.662, + 0.904, + 0.735 + ], + "angle": 0, + "content": "The results on FineGym and Diving48 validate that our SkeletonMAE has a promising ability to enhance the feature representation of skeleton sequence by comprehensively perceiving the underlying topology of actions, and the SSL can learn discriminative action representation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.747, + 0.671, + 0.761 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.77, + 0.905, + 0.829 + ], + "angle": 0, + "content": "In this section, we analyze the contributions of essential components and hyper-parameters of our model. Note that unless otherwise specified, all experiments are verified on the FineGym dataset with masking body sub-region 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.834, + 0.905, + 0.879 + ], + "angle": 0, + "content": "Whether to load pre-trained model or not. To explore the effectiveness of loading the pre-trained SkeletonMAE encoder, we find that the accuracy is 86.3 without load" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.503, + 0.904 + ], + "angle": 0, + "content": "5612" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.113, + 0.208, + 0.183 + ], + "angle": 0, + "content": "
LDMean Acc.
189.6
290.7
391.2
490.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.14, + 0.183, + 0.158, + 0.192 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.123, + 0.46, + 0.179 + ], + "angle": 0, + "content": "
MethodMean Acc.
GraphCL [89]86.5
JOAO [88]88.7
Ours(SkeletonMAE)91.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.338, + 0.179, + 0.357, + 0.19 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "table", + "bbox": [ + 0.075, + 0.204, + 0.46, + 0.26 + ], + "angle": 0, + "content": "
#Masked Body Part012345
GAT [72]86.888.188.989.589.490.0
GCN [28]87.688.989.390.689.590.5
GIN [82]88.689.590.291.290.391.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.258, + 0.26, + 0.278, + 0.27 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "table_caption", + "bbox": [ + 0.065, + 0.271, + 0.469, + 0.341 + ], + "angle": 0, + "content": "Table 4: (a) Mean accuracy of using the different number of GIN layers in SkeletonMAE encoder. (b) Comparison results with the contrastive learning method as the pre-trained encoder. (c) The results of using different backbones in SkeletonMAE under each masked body part are compared." + }, + { + "type": "text", + "bbox": [ + 0.064, + 0.371, + 0.468, + 0.456 + ], + "angle": 0, + "content": "ing the pre-trained SkeletonMAE encoder (randomly initialized weights). As Table 5(a) shows, loading the pretrained model is always better than not loading it. This validates that our SkeletonMAE can learn more comprehensive and generalized representations for unlabeled fine-grained actions by reconstructing the skeleton joint features." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.464, + 0.469, + 0.507 + ], + "angle": 0, + "content": "GIN layers in SkeletonMAE. Table 4(a) shows the performance of using different GIN layers in the SkeletonMAE encoder. The performance is the best when \\( L_{D} = 3 \\)." + }, + { + "type": "text", + "bbox": [ + 0.064, + 0.514, + 0.469, + 0.728 + ], + "angle": 0, + "content": "Comparison with contrastive learning methods. To verify the superior ability of our SkeletonMAE when conducting skeleton sequence pre-training, we compare our SkeletonMAE with different contrastive learning methods GraphCL and JOAO. As shown in Table 4(b), our SkeletonMAE achieves the best performance. Besides, we visually compare the action representations of SkeletonMAE and GraphCL by PCA, as shown in Figure 4(a) and Figure 4(b). Compared to GraphCL, the skeleton representation of our SkeletonMAE appears to have a larger inter-class variance and smaller intra-class variance. This validates that our SkeletonMAE can comprehensively capture the human pose and obtain discriminative skeleton sequence representation. We observe similar patterns in all other classes but visualize only five categories for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.064, + 0.735, + 0.469, + 0.877 + ], + "angle": 0, + "content": "Backbones and masked body parts in SkeletonMAE. As shown in Table 4(c), we show the accuracy of our SSL with different SkeletonMAE backbones and different masked body parts in SkeletonMAE. It can be seen that GIN is always better than both GAT and GCN under the same masked body part. This is because that GIN provides a better inductive bias for graph-level applications. Thus, it is more suitable for learning more generalized skeleton representations. Additionally, we can see that masking body sub-regions 3 and 5 are both optimal across all backbones," + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.113, + 0.905, + 0.144 + ], + "angle": 0, + "content": "
M1234
Mean Acc.89.1(↑2.8)90.6(↑4.3)91.2(↑4.9)91.0(↑4.7)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.692, + 0.145, + 0.712, + 0.157 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.169, + 0.908, + 0.291 + ], + "angle": 0, + "content": "
#Masked Joints Number591215
Ratio of Mask Joints Accuracy of SSL30%50%70%90%
89.790.389.990.1
Masked Body PartHigh91.8 (V3,V5)91.2 (V0,V3,V5)91.0 (V1,V2,V3, V4,V5)90.8 (V0,V1,V2, V3,V5)
Low91.1 (V2,V4)90.1 (V0,V1)91.0 (V1,V2,V3, V4,V5)90.2 (V0,V1,V3, V4,V5)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.692, + 0.291, + 0.712, + 0.301 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.309, + 0.904, + 0.381 + ], + "angle": 0, + "content": "Table 5: (a) Results of four SSL variants. \\(\\uparrow\\) represents the accuracy improvement relative to the random initialization of SkeletonMAE in SSL. (b) The comparison of our body part based masked and the random masked strategies. \\(\\nu_{0}-\\nu_{5}\\): Head, Torso, Left arm, Right arm, Left leg, Right leg." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.411, + 0.902, + 0.44 + ], + "angle": 0, + "content": "which demonstrates the importance of reconstruction of human limbs in action recognition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.904, + 0.619 + ], + "angle": 0, + "content": "Variants of SSL. To evaluate whether our pre-trained SkeletonMAE is effective across different skeleton action recognition models, we set the different number of STRL layers \\((M = 1,2,3,4)\\) to obtain four variants of the SSL. As shown in Table 5(a), all SSL variants outperform the random initialization of SkeletonMAE in SSL, which validates our body part masking strategy indeed improves the discriminative ability of skeleton feature by learning action-sensitive visual concepts. Additionally, three-layer STRL is the best due to the good compromise between efficiency and computational cost. Moreover, it also validates that our SkeletonMAE generalizes well across different models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.904, + 0.854 + ], + "angle": 0, + "content": "Skeleton Masked Strategy. In Table 5(b), our masked body part strategy is fairly compared with the random masked strategy under the same masked joint conditions. Our method is better than the random mask method across all settings. As mentioned in Section 3.1.1, our masking strategy is action-sensitive and reconstructs specific limbs or body parts that dominates the given action class and is suitable for real-world skeleton-based action recognition. From Table 5, we can see that SSL effectively learns meaningful representations that encode the spatial relationships between joints, enabling the capture of crucial information about the structure and configuration of human movement. This notable achievement can be attributed to the integration of prior knowledge about the human body through pretraining the SkeletonMAE. We report results on the ntu 60 dataset in Supplementary Section E." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.863, + 0.902, + 0.877 + ], + "angle": 0, + "content": "Transferability of the SkeletonMAE across datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.501, + 0.903 + ], + "angle": 0, + "content": "5613" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.073, + 0.115, + 0.2, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.195, + 0.15, + 0.206 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.203, + 0.115, + 0.331, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.195, + 0.276, + 0.207 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.113, + 0.465, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.195, + 0.401, + 0.207 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.065, + 0.214, + 0.468, + 0.271 + ], + "angle": 0, + "content": "Figure 4: (a) and (b) are 2d-PCA of SkeletonMAE and GraphCL as pre-trained encoder representations. We randomly select five action classes for 2d-PCA visualization, (c) Confusion matrix for fine-grained action recognition." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.287, + 0.469, + 0.415 + ], + "angle": 0, + "content": "As shown in Figure 5(a), we pre-train SkeletonMAE on the FineGym dataset and then fine-tune it on NTU 60 X-sub and NTU 120 X-sub datasets. Compared with the method that uses the same dataset for pre-training and fine-tuning, our SkeletonMAE achieves better performance across all masked strategies when conducting dataset transfer. This shows that the SkeletonMAE can learn generalized skeleton representation and effectively transfer the strong representation ability to other datasets." + }, + { + "type": "title", + "bbox": [ + 0.066, + 0.428, + 0.277, + 0.443 + ], + "angle": 0, + "content": "4.5. Visualization Analysis" + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.452, + 0.469, + 0.651 + ], + "angle": 0, + "content": "Figure 6 shows the reconstruction process of the skeleton sequence by SkeletonMAE. From the same frame, the difference between the reconstructed skeleton sequence and the original skeleton sequence is slight, but overall the human body structure is reserved. This shows that the SkeletonMAE has good spatial representation learning ability. Moreover, SSL effectively captures the temporal evolution and distinguishing characteristics of actions by capturing the relationships between consecutive joint positions and poses. This highlights the positive impact of the multi-scale temporal dependence incorporated within the STRL module. Figure 4(c) shows that our SSL works well for fine-grained action recognition tasks on the FineGym dataset. More visualization results are in Supplementary Section C." + }, + { + "type": "title", + "bbox": [ + 0.066, + 0.667, + 0.187, + 0.682 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.693, + 0.468, + 0.877 + ], + "angle": 0, + "content": "In this paper, we propose an efficient skeleton sequence learning framework SSL, to learn discriminative skeleton-based action representation. To comprehensively capture the human pose and obtain skeleton sequence representation, we propose a graph-based encoder-decoder pretraining architecture, SkeletonMAE, that embeds skeleton joint sequence into GCN and utilize the prior human topology knowledge to guide the reconstruction of the underlying masked joints and topology. Extensive experimental results show that our SSL achieves SOTA performance on four benchmark skeleton-based action recognition datasets. In our future work, we will build a multi-level feature refinement module to identify ambiguous skeleton actions." + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.114, + 0.843, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.23, + 0.807, + 0.242 + ], + "angle": 0, + "content": "(a) Masked body part" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.258, + 0.843, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.373, + 0.806, + 0.386 + ], + "angle": 0, + "content": "(b) Masked body part" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.388, + 0.903, + 0.445 + ], + "angle": 0, + "content": "Figure 5: (a) The accuracies with mask body part of 0-5 on the NTU 60 X-sub dataset, and (b) the accuracies on the NTU 120 X-sub dataset. \\* means SkeletonMAE encoder pre-trained on the FineGym dataset." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.463, + 0.9, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.511, + 0.739, + 0.52 + ], + "angle": 0, + "content": "(a) Original joint" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.525, + 0.899, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.565, + 0.741, + 0.574 + ], + "angle": 0, + "content": "(b) Masked joint" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.58, + 0.899, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.62, + 0.747, + 0.63 + ], + "angle": 0, + "content": "(c) Reconstructed joint" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.638, + 0.903, + 0.695 + ], + "angle": 0, + "content": "Figure 6: Visualization results for skeleton sequence of \"aerial walkover forward\" on FineGym. (a) The input skeleton sequence. (b) Masked skeleton sequence (masked parts are 3 and 5). (c) Reconstructed skeleton sequence." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.723, + 0.664, + 0.74 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.903, + 0.877 + ], + "angle": 0, + "content": "This work is supported in part by the National Key R&D Program of China under Grant No.2021ZD0111601, in part by the National Natural Science Foundation of China under Grant No.62002395 and No.61976250, in part by the Guangdong Basic and Applied Basic Research Foundation under Grant No.2023A1515011530, No.2021A1515012311 and No.2020B1515020048, and in part by the Guangzhou Science and Technology Planning Project under Grant No. 2023A04J2030." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.502, + 0.903 + ], + "angle": 0, + "content": "5614" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.067, + 0.114, + 0.165, + 0.129 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.075, + 0.138, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[1] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.191, + 0.469, + 0.243 + ], + "angle": 0, + "content": "[2] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.245, + 0.468, + 0.309 + ], + "angle": 0, + "content": "[3] Minghao Chen, Fangyun Wei, Chong Li, and Deng Cai. Frame-wise action representations for long videos via sequence contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13801-13810, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.311, + 0.468, + 0.375 + ], + "angle": 0, + "content": "[4] Tailin Chen, Desen Zhou, Jian Wang, Shidong Wang, Yu Guan, Xuming He, and Errui Ding. Learning multi-granular spatio-temporal graph network for skeleton-based action recognition. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4334-4342, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.376, + 0.468, + 0.442 + ], + "angle": 0, + "content": "[5] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13359-13368, 2021. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.443, + 0.469, + 0.507 + ], + "angle": 0, + "content": "[6] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 183-192, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.508, + 0.468, + 0.572 + ], + "angle": 0, + "content": "[7] Sangwoo Cho, Muhammad Maqbool, Fei Liu, and Hassan Foroosh. Self-attention network for skeleton-based human action recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 635–644, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.574, + 0.469, + 0.638 + ], + "angle": 0, + "content": "[8] Xiao Chu, Wei Yang, Wanli Ouyang, Cheng Ma, Alan L Yuille, and Xiaogang Wang. Multi-context attention for human pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1831-1840, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.64, + 0.468, + 0.692 + ], + "angle": 0, + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.068, + 0.693, + 0.468, + 0.745 + ], + "angle": 0, + "content": "[10] Yong Du, Wei Wang, and Liang Wang. Hierarchical recurrent neural network for skeleton based action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1110-1118, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.068, + 0.746, + 0.468, + 0.81 + ], + "angle": 0, + "content": "[11] Haodong Duan, Yue Zhao, Kai Chen, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2969-2978, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.068, + 0.812, + 0.468, + 0.876 + ], + "angle": 0, + "content": "[12] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. Temporal cycle-consistency learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1801-1810, 2019. 6" + }, + { + "type": "list", + "bbox": [ + 0.068, + 0.138, + 0.469, + 0.876 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.117, + 0.903, + 0.169 + ], + "angle": 0, + "content": "[13] Yin Fan, Xiangju Lu, Dian Li, and Yuanliu Liu. Video-based emotion recognition using cnn-rnn and c3d hybrid networks. In Proceedings of the 18th ACM international conference on multimodal interaction, pages 445-450, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.17, + 0.903, + 0.223 + ], + "angle": 0, + "content": "[14] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.224, + 0.903, + 0.263 + ], + "angle": 0, + "content": "[15] Jerome H Friedman. On bias, variance, 0/1—loss, and the curse-of-dimensionality. Data mining and knowledge discovery, 1(1):55-77, 1997. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.265, + 0.903, + 0.343 + ], + "angle": 0, + "content": "[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.903, + 0.397 + ], + "angle": 0, + "content": "[17] Aditya Grover and Jure Leskovec. node2vec: Scalable feature learning for networks. In Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining, pages 855-864, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.398, + 0.903, + 0.475 + ], + "angle": 0, + "content": "[18] Tianyu Guo, Hong Liu, Zhan Chen, Mengyuan Liu, Tao Wang, and Runwei Ding. Contrastive learning from extremely augmented skeleton sequences for self-supervised action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 762-770, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.903, + 0.542 + ], + "angle": 0, + "content": "[19] Isma Hadji, Konstantinos G Derpanis, and Allan D Jepson. Representation learning via global temporal alignment and cycle-consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11068-11077, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.543, + 0.903, + 0.609 + ], + "angle": 0, + "content": "[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.611, + 0.903, + 0.65 + ], + "angle": 0, + "content": "[21] Sepp Hochreiter, Yoshua Bengio, Paolo Frasconi, Jürgen Schmidhuber, et al. Gradient flow in recurrent nets: the difficulty of learning long-term dependencies, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.904, + 0.717 + ], + "angle": 0, + "content": "[22] Zhenyu Hou, Xiao Liu, Yukuo Cen, Yuxiao Dong, Hongxia Yang, Chunjie Wang, and Jie Tang. Graphmae: Self-supervised masked graph autoencoders. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 594-604, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.903, + 0.77 + ], + "angle": 0, + "content": "[23] Simon Jenni and Hailin Jin. Time-equivariant contrastive video representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9970-9980, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.772, + 0.903, + 0.81 + ], + "angle": 0, + "content": "[24] Gunnar Johansson. Visual perception of biological motion and a model for its analysis. Perception & psychophysics, 14(2):201-211, 1973. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.812, + 0.903, + 0.876 + ], + "angle": 0, + "content": "[25] Gagan Kanojia, Sudhakar Kumawat, and Shanmuganathan Raman. Attentive spatio-temporal representation learning for diving classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.117, + 0.904, + 0.876 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.501, + 0.903 + ], + "angle": 0, + "content": "5615" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.068, + 0.117, + 0.47, + 0.182 + ], + "angle": 0, + "content": "[26] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1725–1732, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.068, + 0.184, + 0.468, + 0.247 + ], + "angle": 0, + "content": "[27] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.25, + 0.468, + 0.288 + ], + "angle": 0, + "content": "[28] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.291, + 0.468, + 0.355 + ], + "angle": 0, + "content": "[29] Heeseung Kwon, Manjin Kim, Suha Kwak, and Minsu Cho. Learning self-similarity in space and time as generalized motion for video action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13065-13075, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.358, + 0.468, + 0.422 + ], + "angle": 0, + "content": "[30] Mei Chee Leong, Hui Li Tan, Haosong Zhang, Liyuan Li, Feng Lin, and Joo Hwee Lim. Joint learning on the hierarchy representation for fine-grained human action recognition. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1059-1063. IEEE, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.424, + 0.468, + 0.5 + ], + "angle": 0, + "content": "[31] Linguo Li, Minsi Wang, Bingbing Ni, Hang Wang, Jiancheng Yang, and Wenjun Zhang. 3d human action representation learning via cross-view consistency pursuit. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4741-4750, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.504, + 0.468, + 0.569 + ], + "angle": 0, + "content": "[32] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3595–3603, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.571, + 0.468, + 0.636 + ], + "angle": 0, + "content": "[33] Tianjiao Li, Lin Geng Foo, Qiuhong Ke, Hossein Rahmani, Anran Wang, Jinghua Wang, and Jun Liu. Dynamic spatio-temporal specialization learning for fine-grained action recognition. In European Conference on Computer Vision, pages 386-403. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.638, + 0.468, + 0.689 + ], + "angle": 0, + "content": "[34] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.691, + 0.468, + 0.743 + ], + "angle": 0, + "content": "[35] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7083-7093, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.746, + 0.468, + 0.809 + ], + "angle": 0, + "content": "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.812, + 0.468, + 0.876 + ], + "angle": 0, + "content": "[37] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10):2684-2701, 2019. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.068, + 0.117, + 0.47, + 0.876 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.117, + 0.902, + 0.155 + ], + "angle": 0, + "content": "[38] Mengyuan Liu, Hong Liu, and Chen Chen. Enhanced skeleton visualization for view invariant human action recognition. Pattern Recognition, 68:346-362, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.158, + 0.902, + 0.209 + ], + "angle": 0, + "content": "[39] Yang Liu, Guanbin Li, and Liang Lin. Cross-modal causal relational reasoning for event-level visual question answering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.212, + 0.902, + 0.262 + ], + "angle": 0, + "content": "[40] Yang Liu, Zhaoyang Lu, Jing Li, and Tao Yang. Hierarchically learned view-invariant representations for cross-view action recognition. IEEE Transactions on Circuits and Systems for Video Technology, 29(8):2416-2430, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.265, + 0.902, + 0.315 + ], + "angle": 0, + "content": "[41] Yang Liu, Zhaoyang Lu, Jing Li, Tao Yang, and Chao Yao. Deep image-to-video adaptation and fusion networks for action recognition. IEEE Transactions on Image Processing, 29:3168-3182, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.318, + 0.902, + 0.369 + ], + "angle": 0, + "content": "[42] Yang Liu, Keze Wang, Guanbin Li, and Liang Lin. Semantics-aware adaptive knowledge distillation for sensor-to-vision action recognition. IEEE Transactions on Image Processing, 30:5573-5588, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.902, + 0.422 + ], + "angle": 0, + "content": "[43] Yang Liu, Keze Wang, Lingbo Liu, Haoyuan Lan, and Liang Lin. Tcgl: Temporal contrastive graph for self-supervised video representation learning. IEEE Transactions on Image Processing, 31:1978-1993, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.425, + 0.902, + 0.475 + ], + "angle": 0, + "content": "[44] Yang Liu, Yu-Shen Wei, Hong Yan, Guan-Bin Li, and Liang Lin. Causal reasoning meets visual representation learning: A prospective study. Machine Intelligence Research, 19(6):485-511, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.478, + 0.902, + 0.543 + ], + "angle": 0, + "content": "[45] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 143-152, 2020. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.545, + 0.902, + 0.596 + ], + "angle": 0, + "content": "[46] Chenxu Luo and Alan L Yuille. Grouped spatial-temporal aggregation for efficient action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5512-5521, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.599, + 0.902, + 0.65 + ], + "angle": 0, + "content": "[47] Fengjun Lv and Ramakant Nevatia. Recognition and segmentation of 3-d human action using hmm and multi-class adaboost. In European conference on computer vision, pages 359-372. Springer, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.652, + 0.902, + 0.703 + ], + "angle": 0, + "content": "[48] Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2891-2900, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.705, + 0.902, + 0.757 + ], + "angle": 0, + "content": "[49] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European conference on computer vision, pages 527-544. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.759, + 0.902, + 0.811 + ], + "angle": 0, + "content": "[50] Jonathan Munro and Dima Damen. Multi-modal domain adaptation for fine-grained action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 122-132, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.813, + 0.902, + 0.876 + ], + "angle": 0, + "content": "[51] Ashish S Nikam and Aarti G Ambekar. Sign language recognition using image based hand gesture recognition techniques. In 2016 online international conference on green engineering and technologies (IC-GET), pages 1-5. IEEE, 2016. 1" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.117, + 0.902, + 0.876 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.502, + 0.903 + ], + "angle": 0, + "content": "5616" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.068, + 0.117, + 0.469, + 0.208 + ], + "angle": 0, + "content": "[52] Cosmas Ifeanyi Nwakanma, Fabliha Bushra Islam, Mareska Pratiwi Maharani, Dong-Seong Kim, and JaeMin Lee. Iot-based vibration sensor data collection and emergency detection classification using long short term memory (lstm). In 2021 International Conference on Artificial Intelligence in Information and Communication (ICAIIC), pages 273-278. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.068, + 0.209, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[53] Jiwong Park, Minsk Lee, Hyung Jin Chang, Kyuewang Lee, and Jin Young Choi. Symmetric graph convolutional autoencoder for unsupervised graph representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6519-6528, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.274, + 0.468, + 0.326 + ], + "angle": 0, + "content": "[54] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.327, + 0.468, + 0.378 + ], + "angle": 0, + "content": "[55] Haocong Rao, Shihao Xu, Xiping Hu, Jun Cheng, and Bin Hu. Augmented skeleton based contrastive action learning with momentum lstm for unsupervised action recognition. Information Sciences, 569:90-109, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.379, + 0.468, + 0.431 + ], + "angle": 0, + "content": "[56] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.432, + 0.468, + 0.457 + ], + "angle": 0, + "content": "[57] Amin Salehi and Hasan Davulcu. Graph attention autoencoders. arXiv preprint arXiv:1905.10715, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.458, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[58] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.511, + 0.468, + 0.575 + ], + "angle": 0, + "content": "[59] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2616-2625, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.577, + 0.469, + 0.641 + ], + "angle": 0, + "content": "[60] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12026-12035, 2019. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.642, + 0.468, + 0.706 + ], + "angle": 0, + "content": "[61] Lei Shi, Yifan Zhang, Jing Hu, Jian Cheng, and Hanqing Lu. Gesture recognition using spatiotemporal deformable convolutional representation. In 2019 IEEE International Conference on Image Processing (ICIP), pages 1900-1904. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.707, + 0.468, + 0.771 + ], + "angle": 0, + "content": "[62] Sijie Song, Cuiling Lan, Junliang Xing, Wenjun Zeng, and Jiaying Liu. An end-to-end spatio-temporal attention model for human action recognition from skeleton data. In Proceedings of the AAAI conference on artificial intelligence, volume 31, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.772, + 0.468, + 0.825 + ], + "angle": 0, + "content": "[63] Sijie Song, Cuiling Lan, Junliang Xing, Wenjun Zeng, and Jiaying Liu. Spatio-temporal attention-based LSTM networks for 3d action recognition and detection. IEEE Transactions on image processing, 27(7):3459-3471, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.826, + 0.468, + 0.877 + ], + "angle": 0, + "content": "[64] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In International conference on machine learning, pages 843-852. PMLR, 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.068, + 0.117, + 0.469, + 0.877 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.117, + 0.903, + 0.169 + ], + "angle": 0, + "content": "[65] Kun Su, Xiulong Liu, and Eli Shlizerman. Predict & cluster: Unsupervised skeleton based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9631-9640, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.17, + 0.903, + 0.222 + ], + "angle": 0, + "content": "[66] Swathikiran Sudhakaran, Sergio Escalera, and Oswald Lanz. Gate-shift networks for video action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1102-1111, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.223, + 0.903, + 0.286 + ], + "angle": 0, + "content": "[67] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5693-5703, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.903, + 0.34 + ], + "angle": 0, + "content": "[68] Jian Tang, Meng Qu, Mingzhe Wang, Ming Zhang, Jun Yan, and Qiao zhu Mei. Line: Large-scale information network embedding. In Proceedings of the 24th international conference on world wide web, pages 1067-1077, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.341, + 0.903, + 0.392 + ], + "angle": 0, + "content": "[69] Fida Mohammad Thoker, Hazel Doughty, and Cees GM Snoek. Skeleton-contrastive 3d action representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1655–1663, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.393, + 0.903, + 0.444 + ], + "angle": 0, + "content": "[70] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.445, + 0.903, + 0.51 + ], + "angle": 0, + "content": "[71] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.903, + 0.55 + ], + "angle": 0, + "content": "[72] Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, Yoshua Bengio, et al. Graph attention networks. stat, 1050(20):10-48550, 2017. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.55, + 0.903, + 0.613 + ], + "angle": 0, + "content": "[73] Raviteja Vemulapalli, Felipe Arrate, and Rama Chellappa. Human action recognition by representing 3d skeletons as points in a lie group. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 588-595, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.615, + 0.903, + 0.668 + ], + "angle": 0, + "content": "[74] Raviteja Vemulapalli and Rama Chellapa. Rolling rotations for recognizing human actions from 3d skeletal data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4471-4479, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.668, + 0.904, + 0.72 + ], + "angle": 0, + "content": "[75] Heng Wang, Du Tran, Lorenzo Torresani, and Matt Feiszli. Video modeling with correlation networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 352-361, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.903, + 0.772 + ], + "angle": 0, + "content": "[76] Jiang Wang, Zicheng Liu, Ying Wu, and Junsong Yuan. Mining actionlet ensemble for action recognition with depth cameras. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 1290-1297. IEEE, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.773, + 0.903, + 0.837 + ], + "angle": 0, + "content": "[77] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.838, + 0.903, + 0.877 + ], + "angle": 0, + "content": "[78] Yanze Wang and Junyong Ye. Tmf: Temporal motion and fusion for action recognition. Computer Vision and Image Understanding, 213:103304, 2021. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.117, + 0.904, + 0.877 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.501, + 0.903 + ], + "angle": 0, + "content": "5617" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.068, + 0.117, + 0.469, + 0.169 + ], + "angle": 0, + "content": "[79] Zhengwei Wang, Qi She, and Aljosa Smolic. Action-net: Multipath excitation for action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13214-13223, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.068, + 0.17, + 0.469, + 0.222 + ], + "angle": 0, + "content": "[80] Wenhan Wu, Yilei Hua, Shiqian Wu, Chen Chen, Aidong Lu, et al. Skeletonmae: Spatial-temporal masked autoencoders for self-supervised skeleton action recognition. arXiv preprint arXiv:2209.02399, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.224, + 0.468, + 0.289 + ], + "angle": 0, + "content": "[81] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In Proceedings of the European conference on computer vision (ECCV), pages 305–321, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.29, + 0.468, + 0.329 + ], + "angle": 0, + "content": "[82] Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826, 2018. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.331, + 0.468, + 0.382 + ], + "angle": 0, + "content": "[83] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Thirty-second AAAI conference on artificial intelligence, 2018. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.384, + 0.469, + 0.449 + ], + "angle": 0, + "content": "[84] Jie Yang, Chaoqun Wang, Zhen Li, Junle Wang, and Ruimao Zhang. Semantic human parsing via scalable semantic transfer over multiple label domains. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19424-19433, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.451, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[85] Jie Yang, Ailing Zeng, Shilong Liu, Feng Li, Ruimao Zhang, and Lei Zhang. Explicit box detection unifies end-to-end multi-person pose estimation. In The Eleventh International Conference on Learning Representations, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.504, + 0.469, + 0.543 + ], + "angle": 0, + "content": "[86] LI Yang, Jin Huang, TIAN Feng, WANG Hong-An, and DAI Guo-Zhong. Gesture interaction in virtual reality. Virtual Reality & Intelligent Hardware, 1(1):84-112, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.545, + 0.468, + 0.609 + ], + "angle": 0, + "content": "[87] Siyuan Yang, Jun Liu, Shijian Lu, Meng Hwa Er, and Alex C Kot. Skeleton cloud colorization for unsupervised 3d action representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13423-13433, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.611, + 0.468, + 0.662 + ], + "angle": 0, + "content": "[88] Yuning You, Tianlong Chen, Yang Shen, and Zhangyang Wang. Graph contrastive learning automated. In International Conference on Machine Learning, pages 12121-12132. PMLR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.664, + 0.468, + 0.716 + ], + "angle": 0, + "content": "[89] Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph contrastive learning with augmentations. Advances in neural information processing systems, 33:5812-5823, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.718, + 0.468, + 0.782 + ], + "angle": 0, + "content": "[90] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Temporal query networks for fine-grained video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4486-4496, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.785, + 0.468, + 0.849 + ], + "angle": 0, + "content": "[91] Nenggan Zheng, Jun Wen, Risheng Liu, Liangqu Long, Jianhua Dai, and Zhefeng Gong. Unsupervised representation learning with long-term dynamics for skeleton based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.851, + 0.468, + 0.877 + ], + "angle": 0, + "content": "[92] Yuan Zhi, Zhan Tong, Limin Wang, and Gangshan Wu. Mgsampler: An explainable sampling strategy for video ac" + }, + { + "type": "list", + "bbox": [ + 0.068, + 0.117, + 0.469, + 0.877 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.536, + 0.117, + 0.902, + 0.155 + ], + "angle": 0, + "content": "tion recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1513-1522, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.157, + 0.902, + 0.209 + ], + "angle": 0, + "content": "[93] Bolei Zhou, Alex Andonian, Aude Oliva, and Antonio Torralba. Temporal relational reasoning in videos. In Proceedings of the European conference on computer vision (ECCV), pages 803-818, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.211, + 0.902, + 0.261 + ], + "angle": 0, + "content": "[94] Yujie Zhou, Haodong Duan, Anyi Rao, Bing Su, and Jiaqi Wang. Self-supervised action representation learning from partial spatio-temporal skeleton sequences. arXiv preprint arXiv:2302.09018, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.902, + 0.329 + ], + "angle": 0, + "content": "[95] Wentao Zhu, Cuiling Lan, Junliang Xing, Wenjun Zeng, Yanghao Li, Li Shen, and Xiaohui Xie. Co-occurrence feature learning for skeleton based action recognition using regularized deep LSTM networks. In Proceedings of the AAAI conference on artificial intelligence, volume 30, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.117, + 0.902, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.501, + 0.903 + ], + "angle": 0, + "content": "5618" + } + ] +] \ No newline at end of file diff --git a/2023/SkeletonMAE_ Graph-based Masked Autoencoder for Skeleton Sequence Pre-training/dcc19c04-02ae-4f8d-8f44-a86cdd91e8a4_origin.pdf b/2023/SkeletonMAE_ Graph-based Masked Autoencoder for Skeleton Sequence Pre-training/dcc19c04-02ae-4f8d-8f44-a86cdd91e8a4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2f30d92f4ce366aa43814c0ee19c6e178a6240c6 --- /dev/null +++ b/2023/SkeletonMAE_ Graph-based Masked Autoencoder for Skeleton Sequence Pre-training/dcc19c04-02ae-4f8d-8f44-a86cdd91e8a4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67ee55ebc79cbfe9703e6de1e74bea6b10b5787b125c1daf00ae2651e1b212dd +size 642631 diff --git a/2023/SkeletonMAE_ Graph-based Masked Autoencoder for Skeleton Sequence Pre-training/full.md b/2023/SkeletonMAE_ Graph-based Masked Autoencoder for Skeleton Sequence Pre-training/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a01cc2f6781151a0dd9bf75720ed6bf0dec395b6 --- /dev/null +++ b/2023/SkeletonMAE_ Graph-based Masked Autoencoder for Skeleton Sequence Pre-training/full.md @@ -0,0 +1,367 @@ +# SkeletonMAE: Graph-based Masked Autoencoder for Skeleton Sequence Pre-training + +Hong Yan $^{1*}$ Yang Liu $^{1*†}$ Yushen Wei $^{1}$ Zhen Li $^{2}$ Guanbin Li $^{1}$ Liang Lin $^{1}$ $^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China + $^{2}$ The Chinese University of Hong Kong, Shenzhen, China + +{yanh36,weiysh8}@mail2.sysu.edu.cn，{liuy856,liguanbin}@mail.sysu.edu.cn, lizhen@cuhk.edu.cn，linliang@ieee.org + +# Abstract + +Skeleton sequence representation learning has shown great advantages for action recognition due to its promising ability to model human joints and topology. However, the current methods usually require sufficient labeled data for training computationally expensive models. Moreover, these methods ignore how to utilize the fine-grained dependencies among different skeleton joints to pre-train an efficient skeleton sequence learning model that can generalize well across different datasets. In this paper, we propose an efficient skeleton sequence learning framework, named Skeleton Sequence Learning (SSL). To comprehensively capture the human pose and obtain discriminative skeleton sequence representation, we build an asymmetric graph-based encoder-decoder pre-training architecture named SkeletonMAE, which embeds skeleton joint sequence into graph convolutional network and reconstructs the masked skeleton joints and edges based on the prior human topology knowledge. Then, the pre-trained SkeletonMAE encoder is integrated with the Spatial-Temporal Representation Learning (STRL) module to build the SSL framework. Extensive experimental results show that our SSL generalizes well across different datasets and outperforms the state-of-the-art self-supervised skeleton-based methods on FineGym, Diving48, NTU 60 and NTU 120 datasets. Moreover, we obtain comparable performance to some fully supervised methods. The code is available at https://github.com/HongYan1123/SkeletonMAE. + +# 1. Introduction + +Human action recognition has attracted increasing attention in video understanding [83, 7, 45, 60, 39], due to its wide applications [2, 14, 61, 51, 52, 86, 44] in human- + +![](images/4fb8b97be0a14dbcf1dba2639876c35f5b1f977c3ad77152bd645c9d1b44440d.jpg) +Figure 1: Traditional MAE usually uses random masking strategy to reconstruct skeleton, which tends to ignore action-sensitive skeleton regions. Differently, our proposed SkeletonMAE reconstructs the masked skeleton joints and edges based on the prior human topology knowledge, to obtain a comprehensive perception of the action. + +computer interaction, intelligent surveillance security, virtual reality, etc. In terms of visual perception [24], even without appearance information, humans can identify action categories by only observing the motion of joints. Unlike RGB videos [3, 14, 13], the skeleton sequences only contain the coordinate information of the key joints of the human body [84], which is high-level, light-weighted, and robust against complex backgrounds and various conditions including viewpoint, scale, and movement speed [11, 69]. Additionally, with the development of human pose estimation algorithms [8, 1, 85], the localization method of human joints (i.e., key points) has made great progress and it is feasible to obtain accurate skeleton sequences. At present, the existing 2D pose estimation method is more accurate and more robust than the 3D pose estimation methods [11]. In Figure 1 (a), we visualize 2D poses estimated with HRNet [67] for two action classes on FineGym dataset [59]. It + +can be seen that the 2D poses can accurately capture human skeletons and motion details. + +Due to the promising ability to model multiple granularities and large variations in human motion, the skeleton sequence is more suitable to distinguish similar actions with subtle differences than the RGB data. To capture discriminative spatial-temporal motion patterns, most of the existing skeleton-based action recognition methods [11, 83, 5] are fully supervised and usually require large amounts of labeled data for training elaborate models, which is time-consuming and labor-intensive. To mitigate the problem of limited labeled training data, self-supervised skeleton-based action recognition methods [31, 18, 65] have attracted increasing attention recently. Some contrastive learning methods [31, 18] adopted data augmentation to generate pairs of positive and negative samples, but they rely heavily on the number of contrastive pairs. With the popularity of the encoder-decoder [64, 48], some methods [91, 65] reconstructed the masked skeleton sequence by link reconstruction to encourage the topological closeness following the paradigm of graph encoder-decoder. However, these methods are usually good at link prediction and node clustering but are unsatisfactory in node and graph classifications. For accurate action recognition, the fine-grained dependencies among different skeleton joints (i.e., graph classifications) are essential. Therefore, previous self-supervised learning-based methods tend to ignore the fine-grained dependencies among different skeleton joints, which restricts the generalization of self-supervised skeleton representation. As shown in Figure 1 (b)-(c), the arm joints and edges are essential to discriminate between these two similar actions. Different from the randomly masking strategy of MAE [20], our masking strategy is action-sensitive and reconstructs specific limbs or body parts that dominate the given action class. Our SkeletonMAE utilizes prior human topology knowledge to guide the reconstruction of the masked skeleton joints and edges to achieve a comprehensive perception of the joints, topology, and action. + +To address the aforementioned challenges, we propose an efficient skeleton sequence representation learning framework, named Skeleton Sequence Learning (SSL). To fully discover the fine-grained dependencies among different skeleton joints, we build a novel asymmetric graph-based encoder-decoder pre-training architecture named SkeletonMAE that embeds skeleton joint sequences in Graph Convolutional Network (GCN). The SkeletonMAE aims to reconstruct the masked human skeleton joints and edges based on prior human topology knowledge so that it can infer the underlying topology of the joints and obtain a comprehensive perception of human action. To learn discriminative spatial-temporal skeleton representation, the pre-trained SkeletonMAE encoder is integrated with the Spatial-Temporal Representation Learning (STRL) + +module to learn spatial-temporal dependencies. Finally, the SSL is fine-tuned on action recognition datasets. Extensive experimental results on FineGym, Diving48, NTU 60 and NTU 120 show that our SSL generalizes well across different datasets and outperforms the state-of-the-art methods significantly. Our contributions are summarized as follows: + +- To comprehensively capture human pose and obtain discriminative skeleton sequence representation, we propose a graph-based encoder-decoder pre-training architecture named SkeletonMAE, that embeds skeleton joint sequence into GCN and utilize the prior human topology knowledge to guide the reconstruction of the underlying masked joints and topology. +- To learn comprehensive spatial-temporal dependencies for skeleton sequence, we propose an efficient skeleton sequence learning framework, named Skeleton Sequence Learning (SSL), which integrates the pre-trained SkeletonMAE encoder with the Spatial-Temporal Representation Learning (STRL) module. +- Extensive experimental results on FineGym, Diving48, NTU 60 and NTU 120 datasets show that our SSL outperforms the state-of-the-art self-supervised skeleton-based action recognition methods and achieves comparable performance compared with the state-of-the-art fully supervised methods. + +# 2. Related Work + +Action Recognition. One of the most challenging tasks for action recognition is to distinguish similar actions from subtle differences [40, 41, 42]. Recently, some challenging action recognition datasets like FineGym [59], Diving48 [34], NTU RGB+D 60 [58] and NTU RGB+D 120 [37] are proposed. These datasets contain a large number of challenging actions that require discriminative and fine-grained action representation learning. For example, in FineGym [59], an action is divided into action units, subactions, or phases, and the model is required to distinguish between "split leap with 1 turn" and "switch leap with 1 turn". The higher inter-class similarity and a new level of granularity in the fine-grained setting make it a challenging task, which makes coarse-grained backbones and methods [14, 2, 71, 79] struggle to overcome. To tackle the more challenging fine-grained action recognition task, most of the existing works [50, 33] are fully supervised and consider fine-grained actions as distinct categories and supervise the model to learn action semantics. However, collecting and labeling these fine-grained actions is time-consuming and labor-intensive, which limits the generalization of a well-trained model to different datasets. To utilize unlabeled data, we propose a graph-based encoder-decoder pre-training architecture named SkeletonMAE. + +Skeleton-based Action Recognition. Due to the promising ability to model multiple granularities and large variations in human motion, the skeleton data is more suitable for the aforementioned action recognition task than the RGB data [4]. Early skeleton-based action recognition methods are usually handcrafted, exploiting the geometric relationship of skeleton joints [47, 73, 76, 74], which greatly limits the feature representation of skeletons. Benefiting from the advantages of deep learning, some methods [95, 62, 63] utilized RNNs as the basic model, Du et al. [10] presented a pioneering work based on hierarchical RNNs. But RNNs easily suffer from vanishing gradients [21]. Inspired by the booming Graph Convolutional Networks (GCN) [28], Yan et al. [83] proposed a spatial-temporal graph convolutional network to learn the spatial and temporal pattern from skeleton data. However, their manually defined topology is arduous to model the relations among joints in underlying topology. Chen et al. [5] proposed a channel-wise topology graph convolution, which models channel-wise topology with a refinement method. Duan et al. [11] proposed a PoseConv3D model that relies on a 3D heatmap volume instead of a graph sequence as the base representation of human skeletons. Different from previous methods that required large amounts of labeled data for training elaborate models, we utilize unlabeled skeleton sequences to pretrain a graph-based encoder-decoder named SkeletonMAE to comprehensively capture human pose and obtain discriminative skeleton sequence representation. + +Self-supervised Learning for Skeleton Sequence. To learn more effective representation for unlabeled skeleton data, self-supervised learning [43] has achieved inspiring progress recently. For contrastive learning approaches, ASCAL [55] and SkeletonCLR [31] applied momentum encoders for contrastive learning with single-stream skeleton sequences. AimCLR [18] used an extreme data augmentation strategy to add additional hard contrastive pairs. Most contrastive learning methods adopt data augmentation to generate positive and negative pairs, but they rely heavily on the number of contrastive pairs. For generative learning approaches, LongT GAN [91] proposed the encoder-decoder to reconstruct masked input sequence skeletons. Based on the LongT GAN, P&C [65] strengthened the encoder and weakened the decoder for feature representation. Wu et al. [80] proposed a spatial-temporal masked autoencoder framework for self-supervised 3D skeleton-based action recognition. Colorization [87] used three pairs of encoder-decoder frameworks to learn spatial-temporal features from skeleton point clouds. Due to the limitation of the reconstruction criterion, previous generative methods usually fail to fully discover the fine-grained spatial-temporal dependencies among different skeleton joints. In our SkeletonMAE, we utilize the prior human topology + +knowledge to infer the skeleton sequence and obtain a comprehensive perception of the action. + +# 3. Methodology + +In this section, we introduce the details of Skeleton Sequence Learning (SSL), which contains two parts: 1) pretraining SkeletonMAE and 2) fine-tuning on downstream datasets based on the pre-trained SkeletonMAE. + +# 3.1. Pre-training SkeletonMAE + +In this section, we introduce graph-based asymmetric encoder-decoder pre-training architecture SkeletonMAE, to learn human skeleton sequence representations without supervision. Since Graph Isomorphism Network (GIN) [82] provides a better inductive bias, it is more suitable for learning more generalized self-supervised representation [22]. Therefore, we adopt GIN as the backbone of SkeletonMAE. Besides, we evaluate different backbones of SkeletonMAE in Table 4, including GIN [82], GCN [28], and GAT [72]. + +# 3.1.1 SkeletonMAE Structure + +Inspired by the effective representation learning by masked autoencoder (MAE) in NLP [9], image recognition [20], and video recognition [70], we focus on the human skeleton sequence and build an asymmetric graph-based encoder-decoder pre-training architecture named SkeletonMAE that embeds skeleton sequence and its prior topology knowledge in GIN. The SkeletonMAE is implemented following the paradigm of graph generative self-supervised learning. + +We follow the joint label of the Kinetics Skeleton dataset [60]. Specifically, as Figure 2(d) shown, we divide all $N = 17$ joints into $R = 6$ local regions according to the natural parts of the body: $\nu_{0},\dots,\nu_{5}$ . Notably, compared to the randomly masking strategy from MAE [20] to select skeleton joints, our masking strategy is action-sensitive and reconstructs specific limbs or body parts that dominate the given action class. Then, we mask these skeleton regions and make the SkeletonMAE reconstruct the masked joint features and their edges based on the adjacent joints. By reconstructing the masked skeleton joints and edges, the SkeletonMAE can infer the underlying topology of joints and obtain a comprehensive perception of the action. + +As shown in Figure 2, the SkeletonMAE is an asymmetric encoder-decoder architecture, which includes an encoder and a decoder. The encoder consists of $L_{D}$ GIN layers that map the input 2D skeleton data to hidden features. The decoder, which consists of only one GIN layer, reconstructs the hidden features under the supervision of the reconstruction criterion. According to the prior human skeleton knowledge that the human skeleton can be represented as a graph with joints as vertices and limbs as edges, we formulate the human skeleton as the following graph structure. + +![](images/35375e2e6c09ba63eb670e2bc3dbc5dc2c87d494400546cf2143b510478c75eb.jpg) +Figure 2: The details of our skeleton sequence pre-training architecture SkeletonMAE. (a) We build a GIN-based asymmetric encoder-decoder structure, to reconstruct joint features to enhance action representation ability. (b) The GIN-based encoder structure contains $L_{D}$ GIN neural network layers, to learn the joint representation spatially. (c) The decoder consists of one GIN layer, which uses the hidden features from the encoder as the input and reconstructs the original input joint features. (d) Partition the joints in the skeleton sequence according to the natural structure of the human body. 5 joints $\{\mathcal{V}_0:\mathrm{Head}\}$ , 4 joints $\{\mathcal{V}_1:\mathrm{Torso}\}$ , 2 joints $\{\mathcal{V}_2:\mathrm{Left arm},\mathcal{V}_3:\mathrm{Right arm},\mathcal{V}_4:\mathrm{Left leg},\mathcal{V}_5:\mathrm{Right leg}\}$ . + +The skeleton sequence of two-dimensional coordinates of $N$ human skeleton joints and $T$ frames is pre-processed in the following way. Specifically, we embed all skeleton joints and their topology into a structure $\mathcal{G}$ , the skeleton structure and the joint feature are fused to obtain a joint sequence matrix $\mathbf{S} \in \mathbb{R}^{N \times T \times 2}$ . And then the $\mathbf{S}$ is linearly transformed to $\mathbf{S} \in \mathbb{R}^{N \times T \times D}$ with learnable parameters. We empirically set $\mathrm{T}$ and $\mathrm{D}$ to 64. For each skeleton frame $\mathbf{X} \in \mathbb{R}^{N \times D}$ from $\mathbf{S}$ , let $\mathcal{G} = (\mathcal{V}, \mathbf{A}, \mathbf{X})$ denote a skeleton, where $\mathcal{V} = \{v_1, v_2, \dots, v_N\}$ is the node set that contains all skeleton joints, $N = |\mathcal{V}|$ is the number of joints. The number of joints is $N = 17$ . $\mathbf{A} \in \{0, 1\}^{N \times N}$ is an adjacency matrix, where $\mathbf{A}_{i,j} = 1$ if joints $i$ and $j$ are physically connected, otherwise 0. The feature of $v_i$ is represented as $\mathbf{x}_i \in \mathbb{R}^{1 \times D}$ . And $G_E$ , $G_D$ denote the GIN encoder and the GIN decoder, respectively. + +# 3.1.2 Skeleton Joints Masking and Reconstruction + +Since the prior human skeleton topology $\mathbf{A}$ is embedded (Figure 2) and we specify the aggregation of joints in Section 3.1.1. Inspired by the GraphMAE [22] that randomly reconstructs the masked graph nodes, our SkeletonMAE reconstructs the masked skeleton feature $\mathbf{X}$ based on the prior skeleton topology, rather than reconstructing graph structure $\mathbf{A}$ [68, 17] or reconstructing both structure $\mathbf{A}$ and features $\mathbf{X}$ [57, 53]. + +To mask skeleton joint features, we randomly select one or more joint sets from $\mathcal{V} = \{\mathcal{V}_0,\dots,\mathcal{V}_5\}$ , which consists of a subset $\overline{\mathcal{V}}\subseteq \mathcal{V}$ for masking. For the human skeleton sequence, each joint communicates with some of its adjacent joints to represent the specific action class. Therefore, it is not feasible to mask all joint sets for all action classes. Then, each of their features is masked with a learnable mask + +token vector $[\mathbf{MASK}] = \mathbf{x}_{[\mathbf{M}]} \in \mathbb{R}^D$ . Thus, the masked joint feature $\overline{\mathbf{x}}_i$ for $\mathbf{v}_i \in \overline{\mathcal{V}}$ in the masked feature matrix $\overline{\mathbf{X}}$ can be defined as $\overline{\mathbf{x}}_i = \mathbf{x}_{[\mathbf{M}]}$ if $\mathbf{v}_i \in \overline{\mathcal{V}}$ , otherwise $\overline{\mathbf{x}}_i = \mathbf{x}_i$ . We set $\overline{\mathbf{X}} \in \mathbb{R}^{N \times D}$ as the input joint feature matrix of the SkeletonMAE, and each joint feature in $\overline{\mathbf{X}}$ can be defined as $\overline{\mathbf{x}}_i = \{\mathbf{x}_{[\mathbf{M}]}, \mathbf{x}_i\}$ , $i = 1, 2, \dots, N$ . Therefore, the masked skeleton sequence can be formulated as $\overline{\mathcal{G}} = (\mathcal{V}, \mathbf{A}, \overline{\mathbf{X}})$ and the objective of SkeletonMAE is to reconstruct the masked skeleton features in $\overline{\mathcal{V}}$ given the partially observed joint features $\overline{\mathbf{X}}$ with the input adjacency matrix $\mathbf{A}$ . The process of SkeletonMAE reconstruction is formalized as: + +$$ +\int \mathbf {H} = G _ {E} (\mathbf {A}, \overline {{\mathbf {X}}}), \quad \mathbf {H} \in \mathbb {R} ^ {N \times D _ {h}} \tag {1} +$$ + +$$ +\left\{\mathbf {Y} = G _ {D} (\mathbf {A}, \mathbf {H}), \quad \mathbf {Y} \in \mathbb {R} ^ {N \times D} \right., \tag {1} +$$ + +where $\mathbf{H}$ and $\mathbf{Y}$ denote the encoder output and the decoder output, respectively. The objective of SkeletonMAE can be formalized as minimizing the divergence between $\mathbf{X}$ and $\mathbf{Y}$ . + +# 3.1.3 Reconstruction Criterion + +The common reconstruction criterion for masked autoencoders is a mean squared error (MSE) in image and video tasks. However, for skeleton sequence, the multidimensional and continuous nature of joint features makes MSE hard to achieve promising feature reconstruction because the MSE is sensitive to dimensionality and vector norms of features [15]. Inspired by the observation [16] that the $l_{2}$ -normalization in the cosine error maps vectors to a unit hyper-sphere and substantially improves the training stability, we utilize the cosine error as the reconstruction. + +To make the reconstruction criterion focus on harder ones among imbalanced easy-and-hard samples [22], we use the Re-weighted Cosine Error (RCE) for SkeletonMAE. The RCE is based on the intuition that we can down-weigh + +easy samples' contribution in training by scaling the cosine error with a power of $\beta \geq 1$ . For predictions with high confidence, their corresponding cosine errors are usually smaller than 1 and decay faster to zero when the scaling factor $\beta > 1$ . Formally, given the original feature $\mathbf{X} \in \mathbb{R}^{N \times D}$ and the reconstructed output $\mathbf{Y} \in \mathbb{R}^{N \times D}$ , the RCE is defined as: + +$$ +\mathcal {L} _ {\mathrm {R C E}} = \sum_ {\mathbf {v} _ {i} \in \overline {{\mathcal {V}}}} \left(\frac {1}{| \overline {{\mathcal {V}}} |} - \frac {\mathbf {x} _ {i} ^ {\mathrm {T}} \cdot \mathbf {z} _ {i}}{| \overline {{\mathcal {V}}} | \times \| \mathbf {x} _ {i} \| \times \| \mathbf {z} _ {i} \|}\right) ^ {\beta}, \tag {2} +$$ + +which represents the average of the similarity gap between the reconstructed feature and the input feature over all masked joints. And $\beta$ is set to 2 in our work. + +By training the SkeletonMAE to reconstruct the skeleton sequence, the pre-trained SkeletonMAE can comprehensively perceive the human skeleton structure and obtain discriminative action representation. After pre-training, the SkeletonMAE can be elegantly embedded into the Skeleton Sequence Learning (SSL) framework for fine-tuning. + +# 3.2. Fine-tuning for Skeleton Action Recognition + +To evaluate the SkeletonMAE's generalization ability for skeleton action recognition, we build a complete skeleton action recognition model named Skeleton Sequence Learning (SSL), based on the pre-trained SkeletonMAE. To capture multiple-person interaction, we integrate two pre-trained SkeletonMAE encoders to build the Spatial-Temporal Representation Learning (STRL) module, as shown in Figure 3(b)-(c). The entire SSL consists of an $M$ -layer STRL model and a classifier. The SSL model is finally fine-tuned on skeleton action recognition datasets with cross-entropy loss to recognize actions. + +# 3.2.1 Spatial-Temporal Representation Learning + +The STRL contains two pre-trained SkeletonMAE encoders for Spatial Modeling (SM). The input of SM is skeleton sequence S. The output of SM is connected with the input by $1\times 1$ convolution for residual connection (Figure 3 (b)). + +As shown in Figure 3 (c), the input skeleton sequence $\mathbf{S} \in \mathbb{R}^{N \times T \times D}$ is firstly added with the learnable temporal position embedding PE to obtain the skeleton sequence feature $\mathbf{H}_t^{(l)} \in \mathbb{R}^{P \times N \times D^{(l)}}$ . To model multiple human skeleton interactions, we obtain two individual features ( $P = 2$ ) for two persons $\mathbf{H}_{t,0}^{(l)} \in \mathbb{R}^{N \times D^{(l)}}$ and $\mathbf{H}_{t,1}^{(l)} \in \mathbb{R}^{N \times D^{(l)}}$ from $\mathbf{H}_t^{(l)}$ . Here, we take the joint feature of the 0-th person as an example, the operation of the 1-th person is implemented similarly. We send the joint representation $\mathbf{H}_{t,0}^{(l)}$ and prior knowledge of the joint $\widetilde{\mathbf{A}}$ into the SM module, + +$$ +\operatorname {S M} \left(\mathbf {H} _ {t, 0} ^ {(l)}\right) = \operatorname {R e p e a t} \left(\operatorname {S P} \left(G _ {E} \left(\widetilde {\mathbf {A}}, \mathbf {H} _ {t, 0} ^ {(l)}\right)\right); N\right) \oplus \mathbf {H} _ {t, 0} ^ {(l)}, \tag {3} +$$ + +![](images/fb87c7575a775872921782e56c839404148f88d67827dcabea1af4f08c84bb38.jpg) +Figure 3: The pipeline of Skeleton Sequence Learning (SSL). (a) During pre-training, we build an encoder-decoder module named SkeletonMAE that embeds skeleton joints and its prior topology knowledge into GIN and reconstructs the underlying masked joints and topology. (b) The SM consists of the pre-trained SkeletonMAE encoder. (c) We integrate SM structures to build the $M$ -layer Spatial-Temporal Representation Learning (STRL) model and then conduct end-to-end fine-tuning. + +![](images/ec9ea7da87ce471a951ce204826e2fea2f11ec2f1df86543b0532710f1a7f507.jpg) + +where $G_{E}$ is the SkeletonMAE encoder, $\mathrm{SP}(\cdot)$ denotes the sum-pooling, $\mathrm{Repeat}(\cdot;N)$ means repeating the single joint into $N$ joints representations after sum-pooling and then connect it with the $\mathbf{H}_{t,0}^{(l)}$ residual to get the global joint representation $\mathrm{SM}(\mathbf{H}_{t,0}^{(l)})$ . In this way, the SM module can obtain global information through a single joint representation, and constrain some joint features through all joint representations. Similarly, $\mathrm{SM}(\mathbf{H}_{t,1}^{(l)})$ is obtained in the same way. As shown in Figure 3(c), we get the joint features $\mathrm{SM}(\mathbf{H}_t^{(l)})$ that contains the action interaction bewteen the 0-th person and the 1-th person. According to the update rules of graph convolution [28], we can obtain $\mathbf{H}_t^{(l + 1)}$ from $\mathbf{H}_t^{(l)}$ in a multi-layer GCN. For more details, please refer to the Supplementary in Section D. The final skeleton sequence representation is defined as follows: + +$$ +\mathbf {H} _ {t} ^ {(l + 1)} = \sigma (\mathbf {S M} (\mathbf {H} _ {t} ^ {(l)}) \mathbf {W} ^ {(l)}). \tag {4} +$$ + +where $\mathbf{W}^{(l)}$ denotes the trainable weight matrix in the $l^{th}$ layer, $\sigma(\cdot)$ denotes the ReLU activation function. Then, we adopt the widely-used multi-scale temporal pooling [5, 38] to get the final output. Finally, a classifier consisting of MLP and softmax predicts the action class. + +# 4. Experiments + +All experiments are conducted with a single modality (2D pose) and evaluated on the corresponding train/test sets. + +# 4.1. Datasets + +We evaluate our SSL on four benchmark datasets FineGym [59], Diving48 [34], NTU RGB+D 60 [58] and NTU RGB+D 120 [37] in the mainstream skeleton action recognition task. For all datasets except FineGym, we follow the pre-processing protocol provided by [11] to obtain the skeleton sequence from the 2D pose estimator. The pre-processing adopts a Top-Down approach for pose extraction, where the detector is Faster-RCNN [56] with ResNet50 backbone and the pose estimator is HRNet [67] pre-trained on COCO-keypoint [36]. To make a fair comparison, we added pixel noise to the joint during training, making the original joint confidence rate unreliable, thus we do not use the originally fixed threshold. + +FineGym is a large-scale fine-grained action recognition dataset for gymnastic videos, which contains 29K videos of 99 fine-grained gymnastics action classes, which requires action recognition methods to distinguish different sub-actions in the same video. In particular, it provides temporal annotations at both action and sub-action levels with a three-level semantic hierarchy. We follow the method [11] to extract the skeleton data from the 2D pose estimator. + +NTU RGB+D 60 and 120. NTU RGB+D is a large-scale skeleton-based action recognition dataset, where NTU 60 contains 56,880 skeleton sequences and 60 action classes. NTU 120 has 114,480 skeleton sequences and 120 action categories. The action samples are captured from 155 different camera viewpoints. The subjects in NTU 120 are in a wide range of age distribution (from 10 to 57) and from different cultural backgrounds (15 countries), which brings very realistic variation to the quality of actions. The NTU 60 and 120 datasets have a large amount of variation in subjects, views, and backgrounds. + +Diving48 is a challenging fine-grained dataset that focuses on complex and competitive sports content analysis. It is formed of over 18k video clips from competitive dive championships that are distributed over 48 fine-grained classes with minimal biases. The difficulties of the dataset lie in that actions are similar and differ in body parts and their combinations which require the model to capture details and motion in body parts and combine them to perform classification. We report the accuracy on the official train/test split. + +# 4.2. Implementation Details + +In this paper, our SkeletonMAE is optimized by the Adaptive Moment Estimation (Adam) with the initial learning rate as $1.5e^{-4}$ and the PReLU is the activation function. + +
MethodModalityMean Acc. (%)
Fully Supervised
I3D [2]RGB64.4
ST-S3D [81]RGB72.9
TSN [77]RGB+Flow79.8
TRNms [93]RGB+Flow80.2
TSM [35]RGB+Flow81.2
GST-50 [46]RGB84.6
MTN [30]RGB88.5
LT-S3D [81]RGB88.9
TQN [90]RGB+Text90.6
PYSKL [11]Skeleton93.2
PYSKL [11]RGB+Skeleton+Limb95.6
Unsupervised Pre-train
SaL [49]RGB42.7
TCC [12]RGB45.6
GTA [19]RGB49.5
CARL [3]RGB60.4
SSL (ours)Skeleton91.8
+ +Table 1: The comparison with the state-of-the-art unsupervised pre-train and supervised methods on FineGym. + +The batch size is 1024 and the training epoch is 50. At the fine-tuning stage, we use the Stochastic Gradient Descent (SGD) with momentum (0.9) and adopt the warm-up strategy for the first 5 epochs. The total fine-tuning epochs are 110. The learning rate is initialized to 0.1 and is divided by 10 at the 90 epoch and the 100 epoch. And we employ 0.1 for label smoothing. We use a large batch size of 128 to facilitate training our attention mechanism and enhancing the model's perception for all human action classes. Both our pre-training and fine-tuning models are implemented by PyTorch [54], and our SSL is trained on a single NVIDIA GeForce RTX 2080Ti GPU. For more details of implementation, please refer to the Supplementary in Section A. + +Pre-training and Fine-tuning Setting. For each dataset, the SkeletonMAE encoder is pre-trained with unlabeled data from the training set. Then, we load the learned parameter weights to fine-tune the SSL model. + +Evaluation Metrics. To make a fair comparison, we follow previous methods [11, 3, 18] and report the Mean Top-1 accuracy $(\%)$ on FineGym dataset and Top-1 accuracy $(\%)$ on Diving48, NTU 60, and NTU 120 datasets. + +# 4.3. Downstream Evaluation + +For a fair comparison, we compare SSL with other models without pre-training on large-scale action datasets, e.g., Kinetics [27] or Sports1M [26]. The comparison results on FineGym, NTU 60 & 120, and Diving48 datasets are shown in Table 1, Table 2, and Table 3, respectively. + +Results on FineGym Dataset. In Table 1, our SSL with skeleton input outperforms most of the fully supervised + +
MethodBackboneSupervisionJoint Number2D SkeletonNTU 60NTU 120
X-sub (%)X-view (%)X-sub (%)X-set (%)
ST-GCN [83]GCNFully Supervised25X81.588.3--
AS-GCN [32]GCNFully Supervised25X86.894.2--
2s-AGCN [60]GCNFully Supervised25X88.595.182.984.9
Shift-GCN [6]GCNFully Supervised25X90.796.585.987.6
MS-G3D [45]GCNFully Supervised25X91.596.286.988.4
CTR-GCN [5]GCNFully Supervised25X92.496.888.990.6
PYSKL [11]CNNFully Supervised17✓93.796.686.089.6
SkeletonCLR [31]ST-GCNUnsupervised Pre-train25X82.288.973.675.3
CrossCLR [31]ST-GCNUnsupervised Pre-train25X86.292.580.580.4
Wu et al. [80]STTFormerUnsupervised Pre-train25X86.692.976.879.1
AimCLR [18]ST-GCNUnsupervised Pre-train25X86.992.880.180.9
3s-PSTL [94]ST-GCNUnsupervised Pre-train25X87.193.981.382.6
Colorization [87]DGCNNUnsupervised Pre-train25X88.094.9--
SSL(ours)STRLUnsupervised Pre-train17✓92.8(↑ 4.8)96.5(↑ 1.6)84.8(↑ 3.5)85.7(↑ 3.1)
+ +Table 2: The comparison with state-of-the-art unsupervised pre-train and supervised methods on NTU 60 and NTU 120 datasets. $\underline{\mathbf{\Pi}}_{-}^{\prime}$ means the method with the second-best performance under unsupervised pre-training. + +
MethodPre-trainGFLOPsAcc.(%)
Fully Supervised
TSN [34]ImageNet-16.8
TRN [25]ImageNet-22.8
P3D [46]ImageNet-32.4
C3D [46]ImageNet-34.5
CorrNet [46]-74.835.5
CorrNet-R101 [75]ImageNet187.338.2
MG-TEA-ResNet50 [92]ImageNet-39.5
GSM [66]ImageNet107.440.3
TSM-R50 [29]ImageNet153.841.6
TMF [78]ImageNet-42.2
Unsupervised Pre-train
RESOUND-C3D [34]K400-16.4
Jenni et al. [23]K400-29.9
SSL (ours)Diving4842.834.1
+ +Table 3: The comparison with the unsupervised pre-train and supervised methods on the Diving48 dataset. + +methods and achieves the best performance among unsupervised pre-train methods with RGB input. For the same input modality, our performance is lower than the fully supervised method PYSKL[11] (with the skeleton as input) by about $1.4\%$ , because the PYSKL adopts stacks of visual heatmaps of skeleton joints as input while we only use human skeleton coordinates. This validates the promising discriminative ability of our skeleton sequence representation. + +Results on NTU 60 and NTU 120 Datasets. In Table 2, for NTU 60 X-sub and NTU 60 X-view, compared with unsupervised pre-train methods, our SSL outperforms the current state-of-the-art method Colorization [87] by $4.8\%$ and $1.6\%$ , respectively. Our SSL is also competitive compared + +with fully supervised methods, outperforming the first six fully supervised methods on NTU 60 X-sub. For NTU 120 X-sub and NTU 120 X-set, our SSL outperforms the previous best-unsupervised pre-train method 3s-PSTL [94] by $3.5\%$ and $3.1\%$ , respectively. Our SSL is superior compared with some fully supervised methods on NTU 120 X-sub and NTU 120 X-set. These results show that our SSL can learn discriminative skeleton representation from large-scale action recognition datasets due to the promising generalization ability of our SkeletonMAE. + +Results on Diving48 Dataset. Our SSL with skeleton input outperforms some fully supervised methods. Although our SSL is not pre-trained on additional large-scale pretraining action datasets in Table 3, it still achieves competitive performance among unsupervised pre-train methods. This validates that our pre-training model SkeletonMAE can learn discriminative skeleton sequence representation. + +The results on FineGym and Diving48 validate that our SkeletonMAE has a promising ability to enhance the feature representation of skeleton sequence by comprehensively perceiving the underlying topology of actions, and the SSL can learn discriminative action representation. + +# 4.4. Ablation Studies + +In this section, we analyze the contributions of essential components and hyper-parameters of our model. Note that unless otherwise specified, all experiments are verified on the FineGym dataset with masking body sub-region 3. + +Whether to load pre-trained model or not. To explore the effectiveness of loading the pre-trained SkeletonMAE encoder, we find that the accuracy is 86.3 without load + +
LDMean Acc.
189.6
290.7
391.2
490.9
+ +(a) + +
MethodMean Acc.
GraphCL [89]86.5
JOAO [88]88.7
Ours(SkeletonMAE)91.2
+ +(b) + +
#Masked Body Part012345
GAT [72]86.888.188.989.589.490.0
GCN [28]87.688.989.390.689.590.5
GIN [82]88.689.590.291.290.391.2
+ +ing the pre-trained SkeletonMAE encoder (randomly initialized weights). As Table 5(a) shows, loading the pretrained model is always better than not loading it. This validates that our SkeletonMAE can learn more comprehensive and generalized representations for unlabeled fine-grained actions by reconstructing the skeleton joint features. + +GIN layers in SkeletonMAE. Table 4(a) shows the performance of using different GIN layers in the SkeletonMAE encoder. The performance is the best when $L_{D} = 3$ . + +Comparison with contrastive learning methods. To verify the superior ability of our SkeletonMAE when conducting skeleton sequence pre-training, we compare our SkeletonMAE with different contrastive learning methods GraphCL and JOAO. As shown in Table 4(b), our SkeletonMAE achieves the best performance. Besides, we visually compare the action representations of SkeletonMAE and GraphCL by PCA, as shown in Figure 4(a) and Figure 4(b). Compared to GraphCL, the skeleton representation of our SkeletonMAE appears to have a larger inter-class variance and smaller intra-class variance. This validates that our SkeletonMAE can comprehensively capture the human pose and obtain discriminative skeleton sequence representation. We observe similar patterns in all other classes but visualize only five categories for simplicity. + +Backbones and masked body parts in SkeletonMAE. As shown in Table 4(c), we show the accuracy of our SSL with different SkeletonMAE backbones and different masked body parts in SkeletonMAE. It can be seen that GIN is always better than both GAT and GCN under the same masked body part. This is because that GIN provides a better inductive bias for graph-level applications. Thus, it is more suitable for learning more generalized skeleton representations. Additionally, we can see that masking body sub-regions 3 and 5 are both optimal across all backbones, + +(c) +Table 4: (a) Mean accuracy of using the different number of GIN layers in SkeletonMAE encoder. (b) Comparison results with the contrastive learning method as the pre-trained encoder. (c) The results of using different backbones in SkeletonMAE under each masked body part are compared. + +
M1234
Mean Acc.89.1(↑2.8)90.6(↑4.3)91.2(↑4.9)91.0(↑4.7)
+ +(a) + +
#Masked Joints Number591215
Ratio of Mask Joints Accuracy of SSL30%50%70%90%
89.790.389.990.1
Masked Body PartHigh91.8 (V3,V5)91.2 (V0,V3,V5)91.0 (V1,V2,V3, V4,V5)90.8 (V0,V1,V2, V3,V5)
Low91.1 (V2,V4)90.1 (V0,V1)91.0 (V1,V2,V3, V4,V5)90.2 (V0,V1,V3, V4,V5)
+ +(b) + +Table 5: (a) Results of four SSL variants. $\uparrow$ represents the accuracy improvement relative to the random initialization of SkeletonMAE in SSL. (b) The comparison of our body part based masked and the random masked strategies. $\nu_{0}-\nu_{5}$ : Head, Torso, Left arm, Right arm, Left leg, Right leg. + +which demonstrates the importance of reconstruction of human limbs in action recognition. + +Variants of SSL. To evaluate whether our pre-trained SkeletonMAE is effective across different skeleton action recognition models, we set the different number of STRL layers $(M = 1,2,3,4)$ to obtain four variants of the SSL. As shown in Table 5(a), all SSL variants outperform the random initialization of SkeletonMAE in SSL, which validates our body part masking strategy indeed improves the discriminative ability of skeleton feature by learning action-sensitive visual concepts. Additionally, three-layer STRL is the best due to the good compromise between efficiency and computational cost. Moreover, it also validates that our SkeletonMAE generalizes well across different models. + +Skeleton Masked Strategy. In Table 5(b), our masked body part strategy is fairly compared with the random masked strategy under the same masked joint conditions. Our method is better than the random mask method across all settings. As mentioned in Section 3.1.1, our masking strategy is action-sensitive and reconstructs specific limbs or body parts that dominates the given action class and is suitable for real-world skeleton-based action recognition. From Table 5, we can see that SSL effectively learns meaningful representations that encode the spatial relationships between joints, enabling the capture of crucial information about the structure and configuration of human movement. This notable achievement can be attributed to the integration of prior knowledge about the human body through pretraining the SkeletonMAE. We report results on the ntu 60 dataset in Supplementary Section E. + +Transferability of the SkeletonMAE across datasets. + +![](images/1bd449f4e6909eafdc3815d68bdde47a97bcb875710a756a5e4db1e849cc13de.jpg) +(a) + +![](images/255c89d302dfcd18f7cc9283a17a426442beeea21573ccfc133c175861011d0e.jpg) +(b) + +![](images/b0f61c7d2d6aa89463196bc696b5d714b3f682d1efda655f9efa874ac41b16ed.jpg) +(c) +Figure 4: (a) and (b) are 2d-PCA of SkeletonMAE and GraphCL as pre-trained encoder representations. We randomly select five action classes for 2d-PCA visualization, (c) Confusion matrix for fine-grained action recognition. + +As shown in Figure 5(a), we pre-train SkeletonMAE on the FineGym dataset and then fine-tune it on NTU 60 X-sub and NTU 120 X-sub datasets. Compared with the method that uses the same dataset for pre-training and fine-tuning, our SkeletonMAE achieves better performance across all masked strategies when conducting dataset transfer. This shows that the SkeletonMAE can learn generalized skeleton representation and effectively transfer the strong representation ability to other datasets. + +# 4.5. Visualization Analysis + +Figure 6 shows the reconstruction process of the skeleton sequence by SkeletonMAE. From the same frame, the difference between the reconstructed skeleton sequence and the original skeleton sequence is slight, but overall the human body structure is reserved. This shows that the SkeletonMAE has good spatial representation learning ability. Moreover, SSL effectively captures the temporal evolution and distinguishing characteristics of actions by capturing the relationships between consecutive joint positions and poses. This highlights the positive impact of the multi-scale temporal dependence incorporated within the STRL module. Figure 4(c) shows that our SSL works well for fine-grained action recognition tasks on the FineGym dataset. More visualization results are in Supplementary Section C. + +# 5. Conclusion + +In this paper, we propose an efficient skeleton sequence learning framework SSL, to learn discriminative skeleton-based action representation. To comprehensively capture the human pose and obtain skeleton sequence representation, we propose a graph-based encoder-decoder pretraining architecture, SkeletonMAE, that embeds skeleton joint sequence into GCN and utilize the prior human topology knowledge to guide the reconstruction of the underlying masked joints and topology. Extensive experimental results show that our SSL achieves SOTA performance on four benchmark skeleton-based action recognition datasets. In our future work, we will build a multi-level feature refinement module to identify ambiguous skeleton actions. + +![](images/54dd7b564b7cc918b149d8eeec8ece73b720a305fd5500f6b811f23c1e2e00ad.jpg) +(a) Masked body part + +![](images/c03a1682afb7420414d6f69e2ce248a81872f4a2f53ddb88d23e843002a2d942.jpg) +(b) Masked body part +Figure 5: (a) The accuracies with mask body part of 0-5 on the NTU 60 X-sub dataset, and (b) the accuracies on the NTU 120 X-sub dataset. \* means SkeletonMAE encoder pre-trained on the FineGym dataset. + +![](images/dea9acba65f830331750a0eee144a9338b3fc3bd9fce2d4313fe411e01c39792.jpg) +(a) Original joint + +![](images/9f5ff0b48515ec9109f88be3e6505e8309f047ece657364b22483d5e70580491.jpg) +(b) Masked joint + +![](images/70ebc3ed8c947882234cde932c6c1488f1e55a263922d0030cee929599c22391.jpg) +(c) Reconstructed joint +Figure 6: Visualization results for skeleton sequence of "aerial walkover forward" on FineGym. (a) The input skeleton sequence. (b) Masked skeleton sequence (masked parts are 3 and 5). (c) Reconstructed skeleton sequence. + +# Acknowledgement + +This work is supported in part by the National Key R&D Program of China under Grant No.2021ZD0111601, in part by the National Natural Science Foundation of China under Grant No.62002395 and No.61976250, in part by the Guangdong Basic and Applied Basic Research Foundation under Grant No.2023A1515011530, No.2021A1515012311 and No.2020B1515020048, and in part by the Guangzhou Science and Technology Planning Project under Grant No. 2023A04J2030. + +# References + +[1] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 1 +[2] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1, 2, 6 +[3] Minghao Chen, Fangyun Wei, Chong Li, and Deng Cai. Frame-wise action representations for long videos via sequence contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13801-13810, 2022. 1, 6 +[4] Tailin Chen, Desen Zhou, Jian Wang, Shidong Wang, Yu Guan, Xuming He, and Errui Ding. Learning multi-granular spatio-temporal graph network for skeleton-based action recognition. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4334-4342, 2021. 3 +[5] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13359-13368, 2021. 2, 3, 5, 7 +[6] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 183-192, 2020. 7 +[7] Sangwoo Cho, Muhammad Maqbool, Fei Liu, and Hassan Foroosh. Self-attention network for skeleton-based human action recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 635–644, 2020. 1 +[8] Xiao Chu, Wei Yang, Wanli Ouyang, Cheng Ma, Alan L Yuille, and Xiaogang Wang. Multi-context attention for human pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1831-1840, 2017. 1 +[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3 +[10] Yong Du, Wei Wang, and Liang Wang. Hierarchical recurrent neural network for skeleton based action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1110-1118, 2015. 3 +[11] Haodong Duan, Yue Zhao, Kai Chen, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2969-2978, 2022. 1, 2, 3, 6, 7 +[12] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. Temporal cycle-consistency learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1801-1810, 2019. 6 + +[13] Yin Fan, Xiangju Lu, Dian Li, and Yuanliu Liu. Video-based emotion recognition using cnn-rnn and c3d hybrid networks. In Proceedings of the 18th ACM international conference on multimodal interaction, pages 445-450, 2016. 1 +[14] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1, 2 +[15] Jerome H Friedman. On bias, variance, 0/1—loss, and the curse-of-dimensionality. Data mining and knowledge discovery, 1(1):55-77, 1997. 4 +[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 4 +[17] Aditya Grover and Jure Leskovec. node2vec: Scalable feature learning for networks. In Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining, pages 855-864, 2016. 4 +[18] Tianyu Guo, Hong Liu, Zhan Chen, Mengyuan Liu, Tao Wang, and Runwei Ding. Contrastive learning from extremely augmented skeleton sequences for self-supervised action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 762-770, 2022. 2, 3, 6, 7 +[19] Isma Hadji, Konstantinos G Derpanis, and Allan D Jepson. Representation learning via global temporal alignment and cycle-consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11068-11077, 2021. 6 +[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 3 +[21] Sepp Hochreiter, Yoshua Bengio, Paolo Frasconi, Jürgen Schmidhuber, et al. Gradient flow in recurrent nets: the difficulty of learning long-term dependencies, 2001. 3 +[22] Zhenyu Hou, Xiao Liu, Yukuo Cen, Yuxiao Dong, Hongxia Yang, Chunjie Wang, and Jie Tang. Graphmae: Self-supervised masked graph autoencoders. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 594-604, 2022. 3, 4 +[23] Simon Jenni and Hailin Jin. Time-equivariant contrastive video representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9970-9980, 2021. 7 +[24] Gunnar Johansson. Visual perception of biological motion and a model for its analysis. Perception & psychophysics, 14(2):201-211, 1973. 1 +[25] Gagan Kanojia, Sudhakar Kumawat, and Shanmuganathan Raman. Attentive spatio-temporal representation learning for diving classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 7 + +[26] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1725–1732, 2014. 6 +[27] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6 +[28] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 3, 5, 8 +[29] Heeseung Kwon, Manjin Kim, Suha Kwak, and Minsu Cho. Learning self-similarity in space and time as generalized motion for video action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13065-13075, 2021. 7 +[30] Mei Chee Leong, Hui Li Tan, Haosong Zhang, Liyuan Li, Feng Lin, and Joo Hwee Lim. Joint learning on the hierarchy representation for fine-grained human action recognition. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1059-1063. IEEE, 2021. 6 +[31] Linguo Li, Minsi Wang, Bingbing Ni, Hang Wang, Jiancheng Yang, and Wenjun Zhang. 3d human action representation learning via cross-view consistency pursuit. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4741-4750, 2021. 2, 3, 7 +[32] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3595–3603, 2019. 7 +[33] Tianjiao Li, Lin Geng Foo, Qiuhong Ke, Hossein Rahmani, Anran Wang, Jinghua Wang, and Jun Liu. Dynamic spatio-temporal specialization learning for fine-grained action recognition. In European Conference on Computer Vision, pages 386-403. Springer, 2022. 2 +[34] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018. 2, 6, 7 +[35] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7083-7093, 2019. 6 +[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6 +[37] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10):2684-2701, 2019. 2, 6 + +[38] Mengyuan Liu, Hong Liu, and Chen Chen. Enhanced skeleton visualization for view invariant human action recognition. Pattern Recognition, 68:346-362, 2017. 5 +[39] Yang Liu, Guanbin Li, and Liang Lin. Cross-modal causal relational reasoning for event-level visual question answering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1 +[40] Yang Liu, Zhaoyang Lu, Jing Li, and Tao Yang. Hierarchically learned view-invariant representations for cross-view action recognition. IEEE Transactions on Circuits and Systems for Video Technology, 29(8):2416-2430, 2018. 2 +[41] Yang Liu, Zhaoyang Lu, Jing Li, Tao Yang, and Chao Yao. Deep image-to-video adaptation and fusion networks for action recognition. IEEE Transactions on Image Processing, 29:3168-3182, 2019. 2 +[42] Yang Liu, Keze Wang, Guanbin Li, and Liang Lin. Semantics-aware adaptive knowledge distillation for sensor-to-vision action recognition. IEEE Transactions on Image Processing, 30:5573-5588, 2021. 2 +[43] Yang Liu, Keze Wang, Lingbo Liu, Haoyuan Lan, and Liang Lin. Tcgl: Temporal contrastive graph for self-supervised video representation learning. IEEE Transactions on Image Processing, 31:1978-1993, 2022. 3 +[44] Yang Liu, Yu-Shen Wei, Hong Yan, Guan-Bin Li, and Liang Lin. Causal reasoning meets visual representation learning: A prospective study. Machine Intelligence Research, 19(6):485-511, 2022. 1 +[45] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 143-152, 2020. 1, 7 +[46] Chenxu Luo and Alan L Yuille. Grouped spatial-temporal aggregation for efficient action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5512-5521, 2019. 6, 7 +[47] Fengjun Lv and Ramakant Nevatia. Recognition and segmentation of 3-d human action using hmm and multi-class adaboost. In European conference on computer vision, pages 359-372. Springer, 2006. 3 +[48] Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2891-2900, 2017. 2 +[49] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European conference on computer vision, pages 527-544. Springer, 2016. 6 +[50] Jonathan Munro and Dima Damen. Multi-modal domain adaptation for fine-grained action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 122-132, 2020. 2 +[51] Ashish S Nikam and Aarti G Ambekar. Sign language recognition using image based hand gesture recognition techniques. In 2016 online international conference on green engineering and technologies (IC-GET), pages 1-5. IEEE, 2016. 1 + +[52] Cosmas Ifeanyi Nwakanma, Fabliha Bushra Islam, Mareska Pratiwi Maharani, Dong-Seong Kim, and JaeMin Lee. Iot-based vibration sensor data collection and emergency detection classification using long short term memory (lstm). In 2021 International Conference on Artificial Intelligence in Information and Communication (ICAIIC), pages 273-278. IEEE, 2021. 1 +[53] Jiwong Park, Minsk Lee, Hyung Jin Chang, Kyuewang Lee, and Jin Young Choi. Symmetric graph convolutional autoencoder for unsupervised graph representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6519-6528, 2019. 4 +[54] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 6 +[55] Haocong Rao, Shihao Xu, Xiping Hu, Jun Cheng, and Bin Hu. Augmented skeleton based contrastive action learning with momentum lstm for unsupervised action recognition. Information Sciences, 569:90-109, 2021. 3 +[56] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 6 +[57] Amin Salehi and Hasan Davulcu. Graph attention autoencoders. arXiv preprint arXiv:1905.10715, 2019. 4 +[58] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016. 2, 6 +[59] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2616-2625, 2020. 1, 2, 6 +[60] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12026-12035, 2019. 1, 3, 7 +[61] Lei Shi, Yifan Zhang, Jing Hu, Jian Cheng, and Hanqing Lu. Gesture recognition using spatiotemporal deformable convolutional representation. In 2019 IEEE International Conference on Image Processing (ICIP), pages 1900-1904. IEEE, 2019. 1 +[62] Sijie Song, Cuiling Lan, Junliang Xing, Wenjun Zeng, and Jiaying Liu. An end-to-end spatio-temporal attention model for human action recognition from skeleton data. In Proceedings of the AAAI conference on artificial intelligence, volume 31, 2017. 3 +[63] Sijie Song, Cuiling Lan, Junliang Xing, Wenjun Zeng, and Jiaying Liu. Spatio-temporal attention-based LSTM networks for 3d action recognition and detection. IEEE Transactions on image processing, 27(7):3459-3471, 2018. 3 +[64] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In International conference on machine learning, pages 843-852. PMLR, 2015. 2 + +[65] Kun Su, Xiulong Liu, and Eli Shlizerman. Predict & cluster: Unsupervised skeleton based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9631-9640, 2020. 2, 3 +[66] Swathikiran Sudhakaran, Sergio Escalera, and Oswald Lanz. Gate-shift networks for video action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1102-1111, 2020. 7 +[67] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5693-5703, 2019. 1, 6 +[68] Jian Tang, Meng Qu, Mingzhe Wang, Ming Zhang, Jun Yan, and Qiao zhu Mei. Line: Large-scale information network embedding. In Proceedings of the 24th international conference on world wide web, pages 1067-1077, 2015. 4 +[69] Fida Mohammad Thoker, Hazel Doughty, and Cees GM Snoek. Skeleton-contrastive 3d action representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1655–1663, 2021. 1 +[70] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022. 3 +[71] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 2 +[72] Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, Yoshua Bengio, et al. Graph attention networks. stat, 1050(20):10-48550, 2017. 3, 8 +[73] Raviteja Vemulapalli, Felipe Arrate, and Rama Chellappa. Human action recognition by representing 3d skeletons as points in a lie group. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 588-595, 2014. 3 +[74] Raviteja Vemulapalli and Rama Chellapa. Rolling rotations for recognizing human actions from 3d skeletal data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4471-4479, 2016. 3 +[75] Heng Wang, Du Tran, Lorenzo Torresani, and Matt Feiszli. Video modeling with correlation networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 352-361, 2020. 7 +[76] Jiang Wang, Zicheng Liu, Ying Wu, and Junsong Yuan. Mining actionlet ensemble for action recognition with depth cameras. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 1290-1297. IEEE, 2012. 3 +[77] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 6 +[78] Yanze Wang and Junyong Ye. Tmf: Temporal motion and fusion for action recognition. Computer Vision and Image Understanding, 213:103304, 2021. 7 + +[79] Zhengwei Wang, Qi She, and Aljosa Smolic. Action-net: Multipath excitation for action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13214-13223, 2021. 2 +[80] Wenhan Wu, Yilei Hua, Shiqian Wu, Chen Chen, Aidong Lu, et al. Skeletonmae: Spatial-temporal masked autoencoders for self-supervised skeleton action recognition. arXiv preprint arXiv:2209.02399, 2022. 3, 7 +[81] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In Proceedings of the European conference on computer vision (ECCV), pages 305–321, 2018. 6 +[82] Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826, 2018. 3, 8 +[83] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Thirty-second AAAI conference on artificial intelligence, 2018. 1, 2, 3, 7 +[84] Jie Yang, Chaoqun Wang, Zhen Li, Junle Wang, and Ruimao Zhang. Semantic human parsing via scalable semantic transfer over multiple label domains. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19424-19433, 2023. 1 +[85] Jie Yang, Ailing Zeng, Shilong Liu, Feng Li, Ruimao Zhang, and Lei Zhang. Explicit box detection unifies end-to-end multi-person pose estimation. In The Eleventh International Conference on Learning Representations, 2023. 1 +[86] LI Yang, Jin Huang, TIAN Feng, WANG Hong-An, and DAI Guo-Zhong. Gesture interaction in virtual reality. Virtual Reality & Intelligent Hardware, 1(1):84-112, 2019. 1 +[87] Siyuan Yang, Jun Liu, Shijian Lu, Meng Hwa Er, and Alex C Kot. Skeleton cloud colorization for unsupervised 3d action representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13423-13433, 2021. 3, 7 +[88] Yuning You, Tianlong Chen, Yang Shen, and Zhangyang Wang. Graph contrastive learning automated. In International Conference on Machine Learning, pages 12121-12132. PMLR, 2021. 8 +[89] Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph contrastive learning with augmentations. Advances in neural information processing systems, 33:5812-5823, 2020. 8 +[90] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Temporal query networks for fine-grained video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4486-4496, 2021. 6 +[91] Nenggan Zheng, Jun Wen, Risheng Liu, Liangqu Long, Jianhua Dai, and Zhefeng Gong. Unsupervised representation learning with long-term dynamics for skeleton based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 2, 3 +[92] Yuan Zhi, Zhan Tong, Limin Wang, and Gangshan Wu. Mgsampler: An explainable sampling strategy for video ac + +tion recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1513-1522, 2021. 7 +[93] Bolei Zhou, Alex Andonian, Aude Oliva, and Antonio Torralba. Temporal relational reasoning in videos. In Proceedings of the European conference on computer vision (ECCV), pages 803-818, 2018. 6 +[94] Yujie Zhou, Haodong Duan, Anyi Rao, Bing Su, and Jiaqi Wang. Self-supervised action representation learning from partial spatio-temporal skeleton sequences. arXiv preprint arXiv:2302.09018, 2023. 7 +[95] Wentao Zhu, Cuiling Lan, Junliang Xing, Wenjun Zeng, Yanghao Li, Li Shen, and Xiaohui Xie. Co-occurrence feature learning for skeleton based action recognition using regularized deep LSTM networks. In Proceedings of the AAAI conference on artificial intelligence, volume 30, 2016. 3 \ No newline at end of file diff --git a/2023/SkeletonMAE_ Graph-based Masked Autoencoder for Skeleton Sequence Pre-training/images.zip b/2023/SkeletonMAE_ Graph-based Masked Autoencoder for Skeleton Sequence Pre-training/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f8752d159080794d0aae1aad919e120f09286f72 --- /dev/null +++ b/2023/SkeletonMAE_ Graph-based Masked Autoencoder for Skeleton Sequence Pre-training/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:446a0897e75bff2abe345d3a4df8d8ed44e276c8b0ff655e573c6da1ebce04ff +size 547706 diff --git a/2023/SkeletonMAE_ Graph-based Masked Autoencoder for Skeleton Sequence Pre-training/layout.json b/2023/SkeletonMAE_ Graph-based Masked Autoencoder for Skeleton Sequence Pre-training/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0e54df43988fe0fe9f79dce3f9b896f7a0498491 --- /dev/null +++ b/2023/SkeletonMAE_ Graph-based Masked Autoencoder for Skeleton Sequence Pre-training/layout.json @@ -0,0 +1,10842 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 127, + 514, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 127, + 514, + 165 + ], + "spans": [ + { + "bbox": [ + 63, + 127, + 514, + 165 + ], + "type": "text", + "content": "SkeletonMAE: Graph-based Masked Autoencoder for Skeleton Sequence Pre-training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 185, + 508, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 185, + 508, + 229 + ], + "spans": [ + { + "bbox": [ + 70, + 185, + 508, + 229 + ], + "type": "text", + "content": "Hong Yan" + }, + { + "bbox": [ + 70, + 185, + 508, + 229 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 70, + 185, + 508, + 229 + ], + "type": "text", + "content": " Yang Liu" + }, + { + "bbox": [ + 70, + 185, + 508, + 229 + ], + "type": "inline_equation", + "content": "^{1*†}" + }, + { + "bbox": [ + 70, + 185, + 508, + 229 + ], + "type": "text", + "content": " Yushen Wei" + }, + { + "bbox": [ + 70, + 185, + 508, + 229 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 185, + 508, + 229 + ], + "type": "text", + "content": " Zhen Li" + }, + { + "bbox": [ + 70, + 185, + 508, + 229 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 185, + 508, + 229 + ], + "type": "text", + "content": " Guanbin Li" + }, + { + "bbox": [ + 70, + 185, + 508, + 229 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 185, + 508, + 229 + ], + "type": "text", + "content": " Liang Lin" + }, + { + "bbox": [ + 70, + 185, + 508, + 229 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 185, + 508, + 229 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 185, + 508, + 229 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China \n" + }, + { + "bbox": [ + 70, + 185, + 508, + 229 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 185, + 508, + 229 + ], + "type": "text", + "content": "The Chinese University of Hong Kong, Shenzhen, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 92, + 230, + 483, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 230, + 483, + 256 + ], + "spans": [ + { + "bbox": [ + 92, + 230, + 483, + 256 + ], + "type": "text", + "content": "{yanh36,weiysh8}@mail2.sysu.edu.cn，{liuy856,liguanbin}@mail.sysu.edu.cn, lizhen@cuhk.edu.cn，linliang@ieee.org" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 283, + 183, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 283, + 183, + 296 + ], + "spans": [ + { + "bbox": [ + 135, + 283, + 183, + 296 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 38, + 309, + 282, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 309, + 282, + 633 + ], + "spans": [ + { + "bbox": [ + 38, + 309, + 282, + 633 + ], + "type": "text", + "content": "Skeleton sequence representation learning has shown great advantages for action recognition due to its promising ability to model human joints and topology. However, the current methods usually require sufficient labeled data for training computationally expensive models. Moreover, these methods ignore how to utilize the fine-grained dependencies among different skeleton joints to pre-train an efficient skeleton sequence learning model that can generalize well across different datasets. In this paper, we propose an efficient skeleton sequence learning framework, named Skeleton Sequence Learning (SSL). To comprehensively capture the human pose and obtain discriminative skeleton sequence representation, we build an asymmetric graph-based encoder-decoder pre-training architecture named SkeletonMAE, which embeds skeleton joint sequence into graph convolutional network and reconstructs the masked skeleton joints and edges based on the prior human topology knowledge. Then, the pre-trained SkeletonMAE encoder is integrated with the Spatial-Temporal Representation Learning (STRL) module to build the SSL framework. Extensive experimental results show that our SSL generalizes well across different datasets and outperforms the state-of-the-art self-supervised skeleton-based methods on FineGym, Diving48, NTU 60 and NTU 120 datasets. Moreover, we obtain comparable performance to some fully supervised methods. The code is available at https://github.com/HongYan1123/SkeletonMAE." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 39, + 655, + 119, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 655, + 119, + 667 + ], + "spans": [ + { + "bbox": [ + 39, + 655, + 119, + 667 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 39, + 675, + 279, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 675, + 279, + 711 + ], + "spans": [ + { + "bbox": [ + 39, + 675, + 279, + 711 + ], + "type": "text", + "content": "Human action recognition has attracted increasing attention in video understanding [83, 7, 45, 60, 39], due to its wide applications [2, 14, 61, 51, 52, 86, 44] in human-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 305, + 282, + 530, + 450 + ], + "blocks": [ + { + "bbox": [ + 305, + 282, + 530, + 450 + ], + "lines": [ + { + "bbox": [ + 305, + 282, + 530, + 450 + ], + "spans": [ + { + "bbox": [ + 305, + 282, + 530, + 450 + ], + "type": "image", + "image_path": "4fb8b97be0a14dbcf1dba2639876c35f5b1f977c3ad77152bd645c9d1b44440d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 455, + 538, + 527 + ], + "lines": [ + { + "bbox": [ + 296, + 455, + 538, + 527 + ], + "spans": [ + { + "bbox": [ + 296, + 455, + 538, + 527 + ], + "type": "text", + "content": "Figure 1: Traditional MAE usually uses random masking strategy to reconstruct skeleton, which tends to ignore action-sensitive skeleton regions. Differently, our proposed SkeletonMAE reconstructs the masked skeleton joints and edges based on the prior human topology knowledge, to obtain a comprehensive perception of the action." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 535, + 539, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 535, + 539, + 740 + ], + "spans": [ + { + "bbox": [ + 296, + 535, + 539, + 740 + ], + "type": "text", + "content": "computer interaction, intelligent surveillance security, virtual reality, etc. In terms of visual perception [24], even without appearance information, humans can identify action categories by only observing the motion of joints. Unlike RGB videos [3, 14, 13], the skeleton sequences only contain the coordinate information of the key joints of the human body [84], which is high-level, light-weighted, and robust against complex backgrounds and various conditions including viewpoint, scale, and movement speed [11, 69]. Additionally, with the development of human pose estimation algorithms [8, 1, 85], the localization method of human joints (i.e., key points) has made great progress and it is feasible to obtain accurate skeleton sequences. At present, the existing 2D pose estimation method is more accurate and more robust than the 3D pose estimation methods [11]. In Figure 1 (a), we visualize 2D poses estimated with HRNet [67] for two action classes on FineGym dataset [59]. It" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 71, + 31, + 113, + 62 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 31, + 113, + 62 + ], + "spans": [ + { + "bbox": [ + 71, + 31, + 113, + 62 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 147, + 30, + 474, + 64 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 30, + 474, + 64 + ], + "spans": [ + { + "bbox": [ + 147, + 30, + 474, + 64 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 718, + 207, + 729 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 718, + 207, + 729 + ], + "spans": [ + { + "bbox": [ + 53, + 718, + 207, + 729 + ], + "type": "text", + "content": "* Both authors contributed equally to this work." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 729, + 131, + 738 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 729, + 131, + 738 + ], + "spans": [ + { + "bbox": [ + 55, + 729, + 131, + 738 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 279, + 751, + 299, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 299, + 761 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 299, + 761 + ], + "type": "text", + "content": "5606" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 38, + 97, + 279, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 97, + 279, + 120 + ], + "spans": [ + { + "bbox": [ + 38, + 97, + 279, + 120 + ], + "type": "text", + "content": "can be seen that the 2D poses can accurately capture human skeletons and motion details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 38, + 125, + 281, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 125, + 281, + 557 + ], + "spans": [ + { + "bbox": [ + 38, + 125, + 281, + 557 + ], + "type": "text", + "content": "Due to the promising ability to model multiple granularities and large variations in human motion, the skeleton sequence is more suitable to distinguish similar actions with subtle differences than the RGB data. To capture discriminative spatial-temporal motion patterns, most of the existing skeleton-based action recognition methods [11, 83, 5] are fully supervised and usually require large amounts of labeled data for training elaborate models, which is time-consuming and labor-intensive. To mitigate the problem of limited labeled training data, self-supervised skeleton-based action recognition methods [31, 18, 65] have attracted increasing attention recently. Some contrastive learning methods [31, 18] adopted data augmentation to generate pairs of positive and negative samples, but they rely heavily on the number of contrastive pairs. With the popularity of the encoder-decoder [64, 48], some methods [91, 65] reconstructed the masked skeleton sequence by link reconstruction to encourage the topological closeness following the paradigm of graph encoder-decoder. However, these methods are usually good at link prediction and node clustering but are unsatisfactory in node and graph classifications. For accurate action recognition, the fine-grained dependencies among different skeleton joints (i.e., graph classifications) are essential. Therefore, previous self-supervised learning-based methods tend to ignore the fine-grained dependencies among different skeleton joints, which restricts the generalization of self-supervised skeleton representation. As shown in Figure 1 (b)-(c), the arm joints and edges are essential to discriminate between these two similar actions. Different from the randomly masking strategy of MAE [20], our masking strategy is action-sensitive and reconstructs specific limbs or body parts that dominate the given action class. Our SkeletonMAE utilizes prior human topology knowledge to guide the reconstruction of the masked skeleton joints and edges to achieve a comprehensive perception of the joints, topology, and action." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 38, + 559, + 279, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 559, + 279, + 740 + ], + "spans": [ + { + "bbox": [ + 38, + 559, + 279, + 740 + ], + "type": "text", + "content": "To address the aforementioned challenges, we propose an efficient skeleton sequence representation learning framework, named Skeleton Sequence Learning (SSL). To fully discover the fine-grained dependencies among different skeleton joints, we build a novel asymmetric graph-based encoder-decoder pre-training architecture named SkeletonMAE that embeds skeleton joint sequences in Graph Convolutional Network (GCN). The SkeletonMAE aims to reconstruct the masked human skeleton joints and edges based on prior human topology knowledge so that it can infer the underlying topology of the joints and obtain a comprehensive perception of human action. To learn discriminative spatial-temporal skeleton representation, the pre-trained SkeletonMAE encoder is integrated with the Spatial-Temporal Representation Learning (STRL)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 296, + 97, + 538, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 97, + 538, + 170 + ], + "spans": [ + { + "bbox": [ + 296, + 97, + 538, + 170 + ], + "type": "text", + "content": "module to learn spatial-temporal dependencies. Finally, the SSL is fine-tuned on action recognition datasets. Extensive experimental results on FineGym, Diving48, NTU 60 and NTU 120 show that our SSL generalizes well across different datasets and outperforms the state-of-the-art methods significantly. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 308, + 181, + 536, + 407 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 308, + 181, + 536, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 181, + 536, + 264 + ], + "spans": [ + { + "bbox": [ + 308, + 181, + 536, + 264 + ], + "type": "text", + "content": "- To comprehensively capture human pose and obtain discriminative skeleton sequence representation, we propose a graph-based encoder-decoder pre-training architecture named SkeletonMAE, that embeds skeleton joint sequence into GCN and utilize the prior human topology knowledge to guide the reconstruction of the underlying masked joints and topology." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 265, + 536, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 265, + 536, + 335 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 536, + 335 + ], + "type": "text", + "content": "- To learn comprehensive spatial-temporal dependencies for skeleton sequence, we propose an efficient skeleton sequence learning framework, named Skeleton Sequence Learning (SSL), which integrates the pre-trained SkeletonMAE encoder with the Spatial-Temporal Representation Learning (STRL) module." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 309, + 336, + 536, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 336, + 536, + 407 + ], + "spans": [ + { + "bbox": [ + 309, + 336, + 536, + 407 + ], + "type": "text", + "content": "- Extensive experimental results on FineGym, Diving48, NTU 60 and NTU 120 datasets show that our SSL outperforms the state-of-the-art self-supervised skeleton-based action recognition methods and achieves comparable performance compared with the state-of-the-art fully supervised methods." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 298, + 421, + 384, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 421, + 384, + 435 + ], + "spans": [ + { + "bbox": [ + 298, + 421, + 384, + 435 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 451, + 538, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 451, + 538, + 740 + ], + "spans": [ + { + "bbox": [ + 296, + 451, + 538, + 740 + ], + "type": "text", + "content": "Action Recognition. One of the most challenging tasks for action recognition is to distinguish similar actions from subtle differences [40, 41, 42]. Recently, some challenging action recognition datasets like FineGym [59], Diving48 [34], NTU RGB+D 60 [58] and NTU RGB+D 120 [37] are proposed. These datasets contain a large number of challenging actions that require discriminative and fine-grained action representation learning. For example, in FineGym [59], an action is divided into action units, subactions, or phases, and the model is required to distinguish between \"split leap with 1 turn\" and \"switch leap with 1 turn\". The higher inter-class similarity and a new level of granularity in the fine-grained setting make it a challenging task, which makes coarse-grained backbones and methods [14, 2, 71, 79] struggle to overcome. To tackle the more challenging fine-grained action recognition task, most of the existing works [50, 33] are fully supervised and consider fine-grained actions as distinct categories and supervise the model to learn action semantics. However, collecting and labeling these fine-grained actions is time-consuming and labor-intensive, which limits the generalization of a well-trained model to different datasets. To utilize unlabeled data, we propose a graph-based encoder-decoder pre-training architecture named SkeletonMAE." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 751, + 298, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 298, + 761 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 298, + 761 + ], + "type": "text", + "content": "5607" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 38, + 97, + 281, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 97, + 281, + 432 + ], + "spans": [ + { + "bbox": [ + 38, + 97, + 281, + 432 + ], + "type": "text", + "content": "Skeleton-based Action Recognition. Due to the promising ability to model multiple granularities and large variations in human motion, the skeleton data is more suitable for the aforementioned action recognition task than the RGB data [4]. Early skeleton-based action recognition methods are usually handcrafted, exploiting the geometric relationship of skeleton joints [47, 73, 76, 74], which greatly limits the feature representation of skeletons. Benefiting from the advantages of deep learning, some methods [95, 62, 63] utilized RNNs as the basic model, Du et al. [10] presented a pioneering work based on hierarchical RNNs. But RNNs easily suffer from vanishing gradients [21]. Inspired by the booming Graph Convolutional Networks (GCN) [28], Yan et al. [83] proposed a spatial-temporal graph convolutional network to learn the spatial and temporal pattern from skeleton data. However, their manually defined topology is arduous to model the relations among joints in underlying topology. Chen et al. [5] proposed a channel-wise topology graph convolution, which models channel-wise topology with a refinement method. Duan et al. [11] proposed a PoseConv3D model that relies on a 3D heatmap volume instead of a graph sequence as the base representation of human skeletons. Different from previous methods that required large amounts of labeled data for training elaborate models, we utilize unlabeled skeleton sequences to pretrain a graph-based encoder-decoder named SkeletonMAE to comprehensively capture human pose and obtain discriminative skeleton sequence representation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 38, + 440, + 281, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 440, + 281, + 729 + ], + "spans": [ + { + "bbox": [ + 38, + 440, + 281, + 729 + ], + "type": "text", + "content": "Self-supervised Learning for Skeleton Sequence. To learn more effective representation for unlabeled skeleton data, self-supervised learning [43] has achieved inspiring progress recently. For contrastive learning approaches, ASCAL [55] and SkeletonCLR [31] applied momentum encoders for contrastive learning with single-stream skeleton sequences. AimCLR [18] used an extreme data augmentation strategy to add additional hard contrastive pairs. Most contrastive learning methods adopt data augmentation to generate positive and negative pairs, but they rely heavily on the number of contrastive pairs. For generative learning approaches, LongT GAN [91] proposed the encoder-decoder to reconstruct masked input sequence skeletons. Based on the LongT GAN, P&C [65] strengthened the encoder and weakened the decoder for feature representation. Wu et al. [80] proposed a spatial-temporal masked autoencoder framework for self-supervised 3D skeleton-based action recognition. Colorization [87] used three pairs of encoder-decoder frameworks to learn spatial-temporal features from skeleton point clouds. Due to the limitation of the reconstruction criterion, previous generative methods usually fail to fully discover the fine-grained spatial-temporal dependencies among different skeleton joints. In our SkeletonMAE, we utilize the prior human topology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 296, + 97, + 538, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 97, + 538, + 123 + ], + "spans": [ + { + "bbox": [ + 296, + 97, + 538, + 123 + ], + "type": "text", + "content": "knowledge to infer the skeleton sequence and obtain a comprehensive perception of the action." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 297, + 132, + 380, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 132, + 380, + 146 + ], + "spans": [ + { + "bbox": [ + 297, + 132, + 380, + 146 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 296, + 152, + 538, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 152, + 538, + 201 + ], + "spans": [ + { + "bbox": [ + 296, + 152, + 538, + 201 + ], + "type": "text", + "content": "In this section, we introduce the details of Skeleton Sequence Learning (SSL), which contains two parts: 1) pretraining SkeletonMAE and 2) fine-tuning on downstream datasets based on the pre-trained SkeletonMAE." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 297, + 207, + 448, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 207, + 448, + 221 + ], + "spans": [ + { + "bbox": [ + 297, + 207, + 448, + 221 + ], + "type": "text", + "content": "3.1. Pre-training SkeletonMAE" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 296, + 226, + 538, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 226, + 538, + 335 + ], + "spans": [ + { + "bbox": [ + 296, + 226, + 538, + 335 + ], + "type": "text", + "content": "In this section, we introduce graph-based asymmetric encoder-decoder pre-training architecture SkeletonMAE, to learn human skeleton sequence representations without supervision. Since Graph Isomorphism Network (GIN) [82] provides a better inductive bias, it is more suitable for learning more generalized self-supervised representation [22]. Therefore, we adopt GIN as the backbone of SkeletonMAE. Besides, we evaluate different backbones of SkeletonMAE in Table 4, including GIN [82], GCN [28], and GAT [72]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 297, + 349, + 434, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 349, + 434, + 361 + ], + "spans": [ + { + "bbox": [ + 297, + 349, + 434, + 361 + ], + "type": "text", + "content": "3.1.1 SkeletonMAE Structure" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 367, + 537, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 367, + 537, + 463 + ], + "spans": [ + { + "bbox": [ + 296, + 367, + 537, + 463 + ], + "type": "text", + "content": "Inspired by the effective representation learning by masked autoencoder (MAE) in NLP [9], image recognition [20], and video recognition [70], we focus on the human skeleton sequence and build an asymmetric graph-based encoder-decoder pre-training architecture named SkeletonMAE that embeds skeleton sequence and its prior topology knowledge in GIN. The SkeletonMAE is implemented following the paradigm of graph generative self-supervised learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 463, + 538, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 463, + 538, + 618 + ], + "spans": [ + { + "bbox": [ + 296, + 463, + 538, + 618 + ], + "type": "text", + "content": "We follow the joint label of the Kinetics Skeleton dataset [60]. Specifically, as Figure 2(d) shown, we divide all " + }, + { + "bbox": [ + 296, + 463, + 538, + 618 + ], + "type": "inline_equation", + "content": "N = 17" + }, + { + "bbox": [ + 296, + 463, + 538, + 618 + ], + "type": "text", + "content": " joints into " + }, + { + "bbox": [ + 296, + 463, + 538, + 618 + ], + "type": "inline_equation", + "content": "R = 6" + }, + { + "bbox": [ + 296, + 463, + 538, + 618 + ], + "type": "text", + "content": " local regions according to the natural parts of the body: " + }, + { + "bbox": [ + 296, + 463, + 538, + 618 + ], + "type": "inline_equation", + "content": "\\nu_{0},\\dots,\\nu_{5}" + }, + { + "bbox": [ + 296, + 463, + 538, + 618 + ], + "type": "text", + "content": ". Notably, compared to the randomly masking strategy from MAE [20] to select skeleton joints, our masking strategy is action-sensitive and reconstructs specific limbs or body parts that dominate the given action class. Then, we mask these skeleton regions and make the SkeletonMAE reconstruct the masked joint features and their edges based on the adjacent joints. By reconstructing the masked skeleton joints and edges, the SkeletonMAE can infer the underlying topology of joints and obtain a comprehensive perception of the action." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 619, + 538, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 619, + 538, + 740 + ], + "spans": [ + { + "bbox": [ + 296, + 619, + 538, + 740 + ], + "type": "text", + "content": "As shown in Figure 2, the SkeletonMAE is an asymmetric encoder-decoder architecture, which includes an encoder and a decoder. The encoder consists of " + }, + { + "bbox": [ + 296, + 619, + 538, + 740 + ], + "type": "inline_equation", + "content": "L_{D}" + }, + { + "bbox": [ + 296, + 619, + 538, + 740 + ], + "type": "text", + "content": " GIN layers that map the input 2D skeleton data to hidden features. The decoder, which consists of only one GIN layer, reconstructs the hidden features under the supervision of the reconstruction criterion. According to the prior human skeleton knowledge that the human skeleton can be represented as a graph with joints as vertices and limbs as edges, we formulate the human skeleton as the following graph structure." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 751, + 299, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 299, + 761 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 299, + 761 + ], + "type": "text", + "content": "5608" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 40, + 95, + 538, + 228 + ], + "blocks": [ + { + "bbox": [ + 40, + 95, + 538, + 228 + ], + "lines": [ + { + "bbox": [ + 40, + 95, + 538, + 228 + ], + "spans": [ + { + "bbox": [ + 40, + 95, + 538, + 228 + ], + "type": "image", + "image_path": "35375e2e6c09ba63eb670e2bc3dbc5dc2c87d494400546cf2143b510478c75eb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 37, + 236, + 540, + 311 + ], + "lines": [ + { + "bbox": [ + 37, + 236, + 540, + 311 + ], + "spans": [ + { + "bbox": [ + 37, + 236, + 540, + 311 + ], + "type": "text", + "content": "Figure 2: The details of our skeleton sequence pre-training architecture SkeletonMAE. (a) We build a GIN-based asymmetric encoder-decoder structure, to reconstruct joint features to enhance action representation ability. (b) The GIN-based encoder structure contains " + }, + { + "bbox": [ + 37, + 236, + 540, + 311 + ], + "type": "inline_equation", + "content": "L_{D}" + }, + { + "bbox": [ + 37, + 236, + 540, + 311 + ], + "type": "text", + "content": " GIN neural network layers, to learn the joint representation spatially. (c) The decoder consists of one GIN layer, which uses the hidden features from the encoder as the input and reconstructs the original input joint features. (d) Partition the joints in the skeleton sequence according to the natural structure of the human body. 5 joints " + }, + { + "bbox": [ + 37, + 236, + 540, + 311 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{V}_0:\\mathrm{Head}\\}" + }, + { + "bbox": [ + 37, + 236, + 540, + 311 + ], + "type": "text", + "content": ", 4 joints " + }, + { + "bbox": [ + 37, + 236, + 540, + 311 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{V}_1:\\mathrm{Torso}\\}" + }, + { + "bbox": [ + 37, + 236, + 540, + 311 + ], + "type": "text", + "content": ", 2 joints " + }, + { + "bbox": [ + 37, + 236, + 540, + 311 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{V}_2:\\mathrm{Left arm},\\mathcal{V}_3:\\mathrm{Right arm},\\mathcal{V}_4:\\mathrm{Left leg},\\mathcal{V}_5:\\mathrm{Right leg}\\}" + }, + { + "bbox": [ + 37, + 236, + 540, + 311 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "spans": [ + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": "The skeleton sequence of two-dimensional coordinates of " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": " human skeleton joints and " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": " frames is pre-processed in the following way. Specifically, we embed all skeleton joints and their topology into a structure " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": ", the skeleton structure and the joint feature are fused to obtain a joint sequence matrix " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{N \\times T \\times 2}" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": ". And then the " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": " is linearly transformed to " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{N \\times T \\times D}" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": " with learnable parameters. We empirically set " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "\\mathrm{T}" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "\\mathrm{D}" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": " to 64. For each skeleton frame " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = (\\mathcal{V}, \\mathbf{A}, \\mathbf{X})" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": " denote a skeleton, where " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{v_1, v_2, \\dots, v_N\\}" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": " is the node set that contains all skeleton joints, " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "N = |\\mathcal{V}|" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": " is the number of joints. The number of joints is " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "N = 17" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\{0, 1\\}^{N \\times N}" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": " is an adjacency matrix, where " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{i,j} = 1" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": " if joints " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": " are physically connected, otherwise 0. The feature of " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": " is represented as " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i \\in \\mathbb{R}^{1 \\times D}" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": ". And " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "G_E" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "inline_equation", + "content": "G_D" + }, + { + "bbox": [ + 38, + 330, + 281, + 523 + ], + "type": "text", + "content": " denote the GIN encoder and the GIN decoder, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 38, + 538, + 262, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 538, + 262, + 551 + ], + "spans": [ + { + "bbox": [ + 38, + 538, + 262, + 551 + ], + "type": "text", + "content": "3.1.2 Skeleton Joints Masking and Reconstruction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 38, + 558, + 279, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 558, + 279, + 654 + ], + "spans": [ + { + "bbox": [ + 38, + 558, + 279, + 654 + ], + "type": "text", + "content": "Since the prior human skeleton topology " + }, + { + "bbox": [ + 38, + 558, + 279, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 38, + 558, + 279, + 654 + ], + "type": "text", + "content": " is embedded (Figure 2) and we specify the aggregation of joints in Section 3.1.1. Inspired by the GraphMAE [22] that randomly reconstructs the masked graph nodes, our SkeletonMAE reconstructs the masked skeleton feature " + }, + { + "bbox": [ + 38, + 558, + 279, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 38, + 558, + 279, + 654 + ], + "type": "text", + "content": " based on the prior skeleton topology, rather than reconstructing graph structure " + }, + { + "bbox": [ + 38, + 558, + 279, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 38, + 558, + 279, + 654 + ], + "type": "text", + "content": " [68, 17] or reconstructing both structure " + }, + { + "bbox": [ + 38, + 558, + 279, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 38, + 558, + 279, + 654 + ], + "type": "text", + "content": " and features " + }, + { + "bbox": [ + 38, + 558, + 279, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 38, + 558, + 279, + 654 + ], + "type": "text", + "content": " [57, 53]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 38, + 655, + 279, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 655, + 279, + 740 + ], + "spans": [ + { + "bbox": [ + 38, + 655, + 279, + 740 + ], + "type": "text", + "content": "To mask skeleton joint features, we randomly select one or more joint sets from " + }, + { + "bbox": [ + 38, + 655, + 279, + 740 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{\\mathcal{V}_0,\\dots,\\mathcal{V}_5\\}" + }, + { + "bbox": [ + 38, + 655, + 279, + 740 + ], + "type": "text", + "content": ", which consists of a subset " + }, + { + "bbox": [ + 38, + 655, + 279, + 740 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{V}}\\subseteq \\mathcal{V}" + }, + { + "bbox": [ + 38, + 655, + 279, + 740 + ], + "type": "text", + "content": " for masking. For the human skeleton sequence, each joint communicates with some of its adjacent joints to represent the specific action class. Therefore, it is not feasible to mask all joint sets for all action classes. Then, each of their features is masked with a learnable mask" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "spans": [ + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "text", + "content": "token vector " + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "inline_equation", + "content": "[\\mathbf{MASK}] = \\mathbf{x}_{[\\mathbf{M}]} \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "text", + "content": ". Thus, the masked joint feature " + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{x}}_i" + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i \\in \\overline{\\mathcal{V}}" + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "text", + "content": " in the masked feature matrix " + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{X}}" + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "text", + "content": " can be defined as " + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{x}}_i = \\mathbf{x}_{[\\mathbf{M}]}" + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i \\in \\overline{\\mathcal{V}}" + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "text", + "content": ", otherwise " + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{x}}_i = \\mathbf{x}_i" + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{X}} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "text", + "content": " as the input joint feature matrix of the SkeletonMAE, and each joint feature in " + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{X}}" + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "text", + "content": " can be defined as " + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{x}}_i = \\{\\mathbf{x}_{[\\mathbf{M}]}, \\mathbf{x}_i\\}" + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "inline_equation", + "content": "i = 1, 2, \\dots, N" + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "text", + "content": ". Therefore, the masked skeleton sequence can be formulated as " + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{G}} = (\\mathcal{V}, \\mathbf{A}, \\overline{\\mathbf{X}})" + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "text", + "content": " and the objective of SkeletonMAE is to reconstruct the masked skeleton features in " + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{V}}" + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "text", + "content": " given the partially observed joint features " + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{X}}" + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "text", + "content": " with the input adjacency matrix " + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 296, + 329, + 539, + 465 + ], + "type": "text", + "content": ". The process of SkeletonMAE reconstruction is formalized as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 342, + 469, + 537, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 469, + 537, + 483 + ], + "spans": [ + { + "bbox": [ + 342, + 469, + 537, + 483 + ], + "type": "interline_equation", + "content": "\\int \\mathbf {H} = G _ {E} (\\mathbf {A}, \\overline {{\\mathbf {X}}}), \\quad \\mathbf {H} \\in \\mathbb {R} ^ {N \\times D _ {h}} \\tag {1}", + "image_path": "5006811803017f870ce591cf94887a484c0733097319943636d664d8c86eb683.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 344, + 483, + 537, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 483, + 537, + 495 + ], + "spans": [ + { + "bbox": [ + 344, + 483, + 537, + 495 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathbf {Y} = G _ {D} (\\mathbf {A}, \\mathbf {H}), \\quad \\mathbf {Y} \\in \\mathbb {R} ^ {N \\times D} \\right., \\tag {1}", + "image_path": "a852ab0d093e823025a17af9b10377156c8f9e6eb0092199246c3b5e0348ef8f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 502, + 539, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 502, + 539, + 539 + ], + "spans": [ + { + "bbox": [ + 296, + 502, + 539, + 539 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 502, + 539, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 296, + 502, + 539, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 502, + 539, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 296, + 502, + 539, + 539 + ], + "type": "text", + "content": " denote the encoder output and the decoder output, respectively. The objective of SkeletonMAE can be formalized as minimizing the divergence between " + }, + { + "bbox": [ + 296, + 502, + 539, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 296, + 502, + 539, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 502, + 539, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 296, + 502, + 539, + 539 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 297, + 552, + 439, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 552, + 439, + 564 + ], + "spans": [ + { + "bbox": [ + 297, + 552, + 439, + 564 + ], + "type": "text", + "content": "3.1.3 Reconstruction Criterion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 570, + 539, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 570, + 539, + 690 + ], + "spans": [ + { + "bbox": [ + 296, + 570, + 539, + 690 + ], + "type": "text", + "content": "The common reconstruction criterion for masked autoencoders is a mean squared error (MSE) in image and video tasks. However, for skeleton sequence, the multidimensional and continuous nature of joint features makes MSE hard to achieve promising feature reconstruction because the MSE is sensitive to dimensionality and vector norms of features [15]. Inspired by the observation [16] that the " + }, + { + "bbox": [ + 296, + 570, + 539, + 690 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 296, + 570, + 539, + 690 + ], + "type": "text", + "content": "-normalization in the cosine error maps vectors to a unit hyper-sphere and substantially improves the training stability, we utilize the cosine error as the reconstruction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 690, + 538, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 690, + 538, + 740 + ], + "spans": [ + { + "bbox": [ + 296, + 690, + 538, + 740 + ], + "type": "text", + "content": "To make the reconstruction criterion focus on harder ones among imbalanced easy-and-hard samples [22], we use the Re-weighted Cosine Error (RCE) for SkeletonMAE. The RCE is based on the intuition that we can down-weigh" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 751, + 299, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 299, + 761 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 299, + 761 + ], + "type": "text", + "content": "5609" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 38, + 96, + 279, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 96, + 279, + 180 + ], + "spans": [ + { + "bbox": [ + 38, + 96, + 279, + 180 + ], + "type": "text", + "content": "easy samples' contribution in training by scaling the cosine error with a power of " + }, + { + "bbox": [ + 38, + 96, + 279, + 180 + ], + "type": "inline_equation", + "content": "\\beta \\geq 1" + }, + { + "bbox": [ + 38, + 96, + 279, + 180 + ], + "type": "text", + "content": ". For predictions with high confidence, their corresponding cosine errors are usually smaller than 1 and decay faster to zero when the scaling factor " + }, + { + "bbox": [ + 38, + 96, + 279, + 180 + ], + "type": "inline_equation", + "content": "\\beta > 1" + }, + { + "bbox": [ + 38, + 96, + 279, + 180 + ], + "type": "text", + "content": ". Formally, given the original feature " + }, + { + "bbox": [ + 38, + 96, + 279, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 38, + 96, + 279, + 180 + ], + "type": "text", + "content": " and the reconstructed output " + }, + { + "bbox": [ + 38, + 96, + 279, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 38, + 96, + 279, + 180 + ], + "type": "text", + "content": ", the RCE is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 186, + 279, + 218 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 186, + 279, + 218 + ], + "spans": [ + { + "bbox": [ + 69, + 186, + 279, + 218 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {R C E}} = \\sum_ {\\mathbf {v} _ {i} \\in \\overline {{\\mathcal {V}}}} \\left(\\frac {1}{| \\overline {{\\mathcal {V}}} |} - \\frac {\\mathbf {x} _ {i} ^ {\\mathrm {T}} \\cdot \\mathbf {z} _ {i}}{| \\overline {{\\mathcal {V}}} | \\times \\| \\mathbf {x} _ {i} \\| \\times \\| \\mathbf {z} _ {i} \\|}\\right) ^ {\\beta}, \\tag {2}", + "image_path": "3eaf7bcafb9569e9e18f1046a31295d1edba311d5e17c4bd914d89e14638ad60.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 38, + 225, + 279, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 225, + 279, + 261 + ], + "spans": [ + { + "bbox": [ + 38, + 225, + 279, + 261 + ], + "type": "text", + "content": "which represents the average of the similarity gap between the reconstructed feature and the input feature over all masked joints. And " + }, + { + "bbox": [ + 38, + 225, + 279, + 261 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 38, + 225, + 279, + 261 + ], + "type": "text", + "content": " is set to 2 in our work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 38, + 261, + 278, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 261, + 278, + 334 + ], + "spans": [ + { + "bbox": [ + 38, + 261, + 278, + 334 + ], + "type": "text", + "content": "By training the SkeletonMAE to reconstruct the skeleton sequence, the pre-trained SkeletonMAE can comprehensively perceive the human skeleton structure and obtain discriminative action representation. After pre-training, the SkeletonMAE can be elegantly embedded into the Skeleton Sequence Learning (SSL) framework for fine-tuning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 39, + 340, + 267, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 340, + 267, + 352 + ], + "spans": [ + { + "bbox": [ + 39, + 340, + 267, + 352 + ], + "type": "text", + "content": "3.2. Fine-tuning for Skeleton Action Recognition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 38, + 358, + 279, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 358, + 279, + 489 + ], + "spans": [ + { + "bbox": [ + 38, + 358, + 279, + 489 + ], + "type": "text", + "content": "To evaluate the SkeletonMAE's generalization ability for skeleton action recognition, we build a complete skeleton action recognition model named Skeleton Sequence Learning (SSL), based on the pre-trained SkeletonMAE. To capture multiple-person interaction, we integrate two pre-trained SkeletonMAE encoders to build the Spatial-Temporal Representation Learning (STRL) module, as shown in Figure 3(b)-(c). The entire SSL consists of an " + }, + { + "bbox": [ + 38, + 358, + 279, + 489 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 38, + 358, + 279, + 489 + ], + "type": "text", + "content": "-layer STRL model and a classifier. The SSL model is finally fine-tuned on skeleton action recognition datasets with cross-entropy loss to recognize actions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 38, + 504, + 255, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 504, + 255, + 517 + ], + "spans": [ + { + "bbox": [ + 38, + 504, + 255, + 517 + ], + "type": "text", + "content": "3.2.1 Spatial-Temporal Representation Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 38, + 522, + 278, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 522, + 278, + 570 + ], + "spans": [ + { + "bbox": [ + 38, + 522, + 278, + 570 + ], + "type": "text", + "content": "The STRL contains two pre-trained SkeletonMAE encoders for Spatial Modeling (SM). The input of SM is skeleton sequence S. The output of SM is connected with the input by " + }, + { + "bbox": [ + 38, + 522, + 278, + 570 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 38, + 522, + 278, + 570 + ], + "type": "text", + "content": " convolution for residual connection (Figure 3 (b))." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 38, + 571, + 279, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 571, + 279, + 701 + ], + "spans": [ + { + "bbox": [ + 38, + 571, + 279, + 701 + ], + "type": "text", + "content": "As shown in Figure 3 (c), the input skeleton sequence " + }, + { + "bbox": [ + 38, + 571, + 279, + 701 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{N \\times T \\times D}" + }, + { + "bbox": [ + 38, + 571, + 279, + 701 + ], + "type": "text", + "content": " is firstly added with the learnable temporal position embedding PE to obtain the skeleton sequence feature " + }, + { + "bbox": [ + 38, + 571, + 279, + 701 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_t^{(l)} \\in \\mathbb{R}^{P \\times N \\times D^{(l)}}" + }, + { + "bbox": [ + 38, + 571, + 279, + 701 + ], + "type": "text", + "content": ". To model multiple human skeleton interactions, we obtain two individual features (" + }, + { + "bbox": [ + 38, + 571, + 279, + 701 + ], + "type": "inline_equation", + "content": "P = 2" + }, + { + "bbox": [ + 38, + 571, + 279, + 701 + ], + "type": "text", + "content": ") for two persons " + }, + { + "bbox": [ + 38, + 571, + 279, + 701 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{t,0}^{(l)} \\in \\mathbb{R}^{N \\times D^{(l)}}" + }, + { + "bbox": [ + 38, + 571, + 279, + 701 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 38, + 571, + 279, + 701 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{t,1}^{(l)} \\in \\mathbb{R}^{N \\times D^{(l)}}" + }, + { + "bbox": [ + 38, + 571, + 279, + 701 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 38, + 571, + 279, + 701 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_t^{(l)}" + }, + { + "bbox": [ + 38, + 571, + 279, + 701 + ], + "type": "text", + "content": ". Here, we take the joint feature of the 0-th person as an example, the operation of the 1-th person is implemented similarly. We send the joint representation " + }, + { + "bbox": [ + 38, + 571, + 279, + 701 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{t,0}^{(l)}" + }, + { + "bbox": [ + 38, + 571, + 279, + 701 + ], + "type": "text", + "content": " and prior knowledge of the joint " + }, + { + "bbox": [ + 38, + 571, + 279, + 701 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{A}}" + }, + { + "bbox": [ + 38, + 571, + 279, + 701 + ], + "type": "text", + "content": " into the SM module," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 708, + 278, + 740 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 708, + 278, + 740 + ], + "spans": [ + { + "bbox": [ + 48, + 708, + 278, + 740 + ], + "type": "interline_equation", + "content": "\\operatorname {S M} \\left(\\mathbf {H} _ {t, 0} ^ {(l)}\\right) = \\operatorname {R e p e a t} \\left(\\operatorname {S P} \\left(G _ {E} \\left(\\widetilde {\\mathbf {A}}, \\mathbf {H} _ {t, 0} ^ {(l)}\\right)\\right); N\\right) \\oplus \\mathbf {H} _ {t, 0} ^ {(l)}, \\tag {3}", + "image_path": "98611cb979bfa4e55e5f41fd4ec05572067d09083453f9dffed9f51591e90f52.jpg" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 300, + 95, + 459, + 283 + ], + "blocks": [ + { + "bbox": [ + 300, + 95, + 459, + 283 + ], + "lines": [ + { + "bbox": [ + 300, + 95, + 459, + 283 + ], + "spans": [ + { + "bbox": [ + 300, + 95, + 459, + 283 + ], + "type": "image", + "image_path": "fb87c7575a775872921782e56c839404148f88d67827dcabea1af4f08c84bb38.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 303, + 537, + 410 + ], + "lines": [ + { + "bbox": [ + 296, + 303, + 537, + 410 + ], + "spans": [ + { + "bbox": [ + 296, + 303, + 537, + 410 + ], + "type": "text", + "content": "Figure 3: The pipeline of Skeleton Sequence Learning (SSL). (a) During pre-training, we build an encoder-decoder module named SkeletonMAE that embeds skeleton joints and its prior topology knowledge into GIN and reconstructs the underlying masked joints and topology. (b) The SM consists of the pre-trained SkeletonMAE encoder. (c) We integrate SM structures to build the " + }, + { + "bbox": [ + 296, + 303, + 537, + 410 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 296, + 303, + 537, + 410 + ], + "type": "text", + "content": "-layer Spatial-Temporal Representation Learning (STRL) model and then conduct end-to-end fine-tuning." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 461, + 96, + 539, + 283 + ], + "blocks": [ + { + "bbox": [ + 461, + 96, + 539, + 283 + ], + "lines": [ + { + "bbox": [ + 461, + 96, + 539, + 283 + ], + "spans": [ + { + "bbox": [ + 461, + 96, + 539, + 283 + ], + "type": "image", + "image_path": "ec9ea7da87ce471a951ce204826e2fea2f11ec2f1df86543b0532710f1a7f507.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "spans": [ + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "inline_equation", + "content": "G_{E}" + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "text", + "content": " is the SkeletonMAE encoder, " + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{SP}(\\cdot)" + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "text", + "content": " denotes the sum-pooling, " + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{Repeat}(\\cdot;N)" + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "text", + "content": " means repeating the single joint into " + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "text", + "content": " joints representations after sum-pooling and then connect it with the " + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{t,0}^{(l)}" + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "text", + "content": " residual to get the global joint representation " + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{SM}(\\mathbf{H}_{t,0}^{(l)})" + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "text", + "content": ". In this way, the SM module can obtain global information through a single joint representation, and constrain some joint features through all joint representations. Similarly, " + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{SM}(\\mathbf{H}_{t,1}^{(l)})" + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "text", + "content": " is obtained in the same way. As shown in Figure 3(c), we get the joint features " + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{SM}(\\mathbf{H}_t^{(l)})" + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "text", + "content": " that contains the action interaction bewteen the 0-th person and the 1-th person. According to the update rules of graph convolution [28], we can obtain " + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_t^{(l + 1)}" + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_t^{(l)}" + }, + { + "bbox": [ + 296, + 436, + 537, + 628 + ], + "type": "text", + "content": " in a multi-layer GCN. For more details, please refer to the Supplementary in Section D. The final skeleton sequence representation is defined as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 351, + 643, + 537, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 643, + 537, + 664 + ], + "spans": [ + { + "bbox": [ + 351, + 643, + 537, + 664 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {t} ^ {(l + 1)} = \\sigma (\\mathbf {S M} (\\mathbf {H} _ {t} ^ {(l)}) \\mathbf {W} ^ {(l)}). \\tag {4}", + "image_path": "7a6af3b63795df9e241aacb9a15540ef28ccadfc3f2d132c243424a6fa7e4869.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 677, + 538, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 677, + 538, + 740 + ], + "spans": [ + { + "bbox": [ + 296, + 677, + 538, + 740 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 677, + 538, + 740 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{(l)}" + }, + { + "bbox": [ + 296, + 677, + 538, + 740 + ], + "type": "text", + "content": " denotes the trainable weight matrix in the " + }, + { + "bbox": [ + 296, + 677, + 538, + 740 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 296, + 677, + 538, + 740 + ], + "type": "text", + "content": " layer, " + }, + { + "bbox": [ + 296, + 677, + 538, + 740 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 296, + 677, + 538, + 740 + ], + "type": "text", + "content": " denotes the ReLU activation function. Then, we adopt the widely-used multi-scale temporal pooling [5, 38] to get the final output. Finally, a classifier consisting of MLP and softmax predicts the action class." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 751, + 298, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 298, + 761 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 298, + 761 + ], + "type": "text", + "content": "5610" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 39, + 96, + 119, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 96, + 119, + 110 + ], + "spans": [ + { + "bbox": [ + 39, + 96, + 119, + 110 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 39, + 116, + 278, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 116, + 278, + 140 + ], + "spans": [ + { + "bbox": [ + 39, + 116, + 278, + 140 + ], + "type": "text", + "content": "All experiments are conducted with a single modality (2D pose) and evaluated on the corresponding train/test sets." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 39, + 145, + 101, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 145, + 101, + 157 + ], + "spans": [ + { + "bbox": [ + 39, + 145, + 101, + 157 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 38, + 164, + 278, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 164, + 278, + 319 + ], + "spans": [ + { + "bbox": [ + 38, + 164, + 278, + 319 + ], + "type": "text", + "content": "We evaluate our SSL on four benchmark datasets FineGym [59], Diving48 [34], NTU RGB+D 60 [58] and NTU RGB+D 120 [37] in the mainstream skeleton action recognition task. For all datasets except FineGym, we follow the pre-processing protocol provided by [11] to obtain the skeleton sequence from the 2D pose estimator. The pre-processing adopts a Top-Down approach for pose extraction, where the detector is Faster-RCNN [56] with ResNet50 backbone and the pose estimator is HRNet [67] pre-trained on COCO-keypoint [36]. To make a fair comparison, we added pixel noise to the joint during training, making the original joint confidence rate unreliable, thus we do not use the originally fixed threshold." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 38, + 327, + 278, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 327, + 278, + 423 + ], + "spans": [ + { + "bbox": [ + 38, + 327, + 278, + 423 + ], + "type": "text", + "content": "FineGym is a large-scale fine-grained action recognition dataset for gymnastic videos, which contains 29K videos of 99 fine-grained gymnastics action classes, which requires action recognition methods to distinguish different sub-actions in the same video. In particular, it provides temporal annotations at both action and sub-action levels with a three-level semantic hierarchy. We follow the method [11] to extract the skeleton data from the 2D pose estimator." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 38, + 431, + 278, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 431, + 278, + 563 + ], + "spans": [ + { + "bbox": [ + 38, + 431, + 278, + 563 + ], + "type": "text", + "content": "NTU RGB+D 60 and 120. NTU RGB+D is a large-scale skeleton-based action recognition dataset, where NTU 60 contains 56,880 skeleton sequences and 60 action classes. NTU 120 has 114,480 skeleton sequences and 120 action categories. The action samples are captured from 155 different camera viewpoints. The subjects in NTU 120 are in a wide range of age distribution (from 10 to 57) and from different cultural backgrounds (15 countries), which brings very realistic variation to the quality of actions. The NTU 60 and 120 datasets have a large amount of variation in subjects, views, and backgrounds." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 38, + 570, + 278, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 570, + 278, + 678 + ], + "spans": [ + { + "bbox": [ + 38, + 570, + 278, + 678 + ], + "type": "text", + "content": "Diving48 is a challenging fine-grained dataset that focuses on complex and competitive sports content analysis. It is formed of over 18k video clips from competitive dive championships that are distributed over 48 fine-grained classes with minimal biases. The difficulties of the dataset lie in that actions are similar and differ in body parts and their combinations which require the model to capture details and motion in body parts and combine them to perform classification. We report the accuracy on the official train/test split." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 39, + 684, + 171, + 696 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 684, + 171, + 696 + ], + "spans": [ + { + "bbox": [ + 39, + 684, + 171, + 696 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 38, + 703, + 278, + 738 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 703, + 278, + 738 + ], + "spans": [ + { + "bbox": [ + 38, + 703, + 278, + 738 + ], + "type": "text", + "content": "In this paper, our SkeletonMAE is optimized by the Adaptive Moment Estimation (Adam) with the initial learning rate as " + }, + { + "bbox": [ + 38, + 703, + 278, + 738 + ], + "type": "inline_equation", + "content": "1.5e^{-4}" + }, + { + "bbox": [ + 38, + 703, + 278, + 738 + ], + "type": "text", + "content": " and the PReLU is the activation function." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 301, + 95, + 533, + 307 + ], + "blocks": [ + { + "bbox": [ + 301, + 95, + 533, + 307 + ], + "lines": [ + { + "bbox": [ + 301, + 95, + 533, + 307 + ], + "spans": [ + { + "bbox": [ + 301, + 95, + 533, + 307 + ], + "type": "table", + "html": "
MethodModalityMean Acc. (%)
Fully Supervised
I3D [2]RGB64.4
ST-S3D [81]RGB72.9
TSN [77]RGB+Flow79.8
TRNms [93]RGB+Flow80.2
TSM [35]RGB+Flow81.2
GST-50 [46]RGB84.6
MTN [30]RGB88.5
LT-S3D [81]RGB88.9
TQN [90]RGB+Text90.6
PYSKL [11]Skeleton93.2
PYSKL [11]RGB+Skeleton+Limb95.6
Unsupervised Pre-train
SaL [49]RGB42.7
TCC [12]RGB45.6
GTA [19]RGB49.5
CARL [3]RGB60.4
SSL (ours)Skeleton91.8
", + "image_path": "657e307161cec89a5369e3540df788ba787a073073f110b32297f9feea62f2fe.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 297, + 315, + 536, + 340 + ], + "lines": [ + { + "bbox": [ + 297, + 315, + 536, + 340 + ], + "spans": [ + { + "bbox": [ + 297, + 315, + 536, + 340 + ], + "type": "text", + "content": "Table 1: The comparison with the state-of-the-art unsupervised pre-train and supervised methods on FineGym." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 296, + 362, + 536, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 362, + 536, + 516 + ], + "spans": [ + { + "bbox": [ + 296, + 362, + 536, + 516 + ], + "type": "text", + "content": "The batch size is 1024 and the training epoch is 50. At the fine-tuning stage, we use the Stochastic Gradient Descent (SGD) with momentum (0.9) and adopt the warm-up strategy for the first 5 epochs. The total fine-tuning epochs are 110. The learning rate is initialized to 0.1 and is divided by 10 at the 90 epoch and the 100 epoch. And we employ 0.1 for label smoothing. We use a large batch size of 128 to facilitate training our attention mechanism and enhancing the model's perception for all human action classes. Both our pre-training and fine-tuning models are implemented by PyTorch [54], and our SSL is trained on a single NVIDIA GeForce RTX 2080Ti GPU. For more details of implementation, please refer to the Supplementary in Section A." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 522, + 536, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 522, + 536, + 570 + ], + "spans": [ + { + "bbox": [ + 296, + 522, + 536, + 570 + ], + "type": "text", + "content": "Pre-training and Fine-tuning Setting. For each dataset, the SkeletonMAE encoder is pre-trained with unlabeled data from the training set. Then, we load the learned parameter weights to fine-tune the SSL model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 297, + 575, + 536, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 575, + 536, + 623 + ], + "spans": [ + { + "bbox": [ + 297, + 575, + 536, + 623 + ], + "type": "text", + "content": "Evaluation Metrics. To make a fair comparison, we follow previous methods [11, 3, 18] and report the Mean Top-1 accuracy " + }, + { + "bbox": [ + 297, + 575, + 536, + 623 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 297, + 575, + 536, + 623 + ], + "type": "text", + "content": " on FineGym dataset and Top-1 accuracy " + }, + { + "bbox": [ + 297, + 575, + 536, + 623 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 297, + 575, + 536, + 623 + ], + "type": "text", + "content": " on Diving48, NTU 60, and NTU 120 datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 297, + 631, + 432, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 631, + 432, + 642 + ], + "spans": [ + { + "bbox": [ + 297, + 631, + 432, + 642 + ], + "type": "text", + "content": "4.3. Downstream Evaluation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 650, + 536, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 650, + 536, + 709 + ], + "spans": [ + { + "bbox": [ + 296, + 650, + 536, + 709 + ], + "type": "text", + "content": "For a fair comparison, we compare SSL with other models without pre-training on large-scale action datasets, e.g., Kinetics [27] or Sports1M [26]. The comparison results on FineGym, NTU 60 & 120, and Diving48 datasets are shown in Table 1, Table 2, and Table 3, respectively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 297, + 714, + 536, + 738 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 714, + 536, + 738 + ], + "spans": [ + { + "bbox": [ + 297, + 714, + 536, + 738 + ], + "type": "text", + "content": "Results on FineGym Dataset. In Table 1, our SSL with skeleton input outperforms most of the fully supervised" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "text", + "content": "5611" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 41, + 95, + 537, + 294 + ], + "blocks": [ + { + "bbox": [ + 41, + 95, + 537, + 294 + ], + "lines": [ + { + "bbox": [ + 41, + 95, + 537, + 294 + ], + "spans": [ + { + "bbox": [ + 41, + 95, + 537, + 294 + ], + "type": "table", + "html": "
MethodBackboneSupervisionJoint Number2D SkeletonNTU 60NTU 120
X-sub (%)X-view (%)X-sub (%)X-set (%)
ST-GCN [83]GCNFully Supervised25X81.588.3--
AS-GCN [32]GCNFully Supervised25X86.894.2--
2s-AGCN [60]GCNFully Supervised25X88.595.182.984.9
Shift-GCN [6]GCNFully Supervised25X90.796.585.987.6
MS-G3D [45]GCNFully Supervised25X91.596.286.988.4
CTR-GCN [5]GCNFully Supervised25X92.496.888.990.6
PYSKL [11]CNNFully Supervised17✓93.796.686.089.6
SkeletonCLR [31]ST-GCNUnsupervised Pre-train25X82.288.973.675.3
CrossCLR [31]ST-GCNUnsupervised Pre-train25X86.292.580.580.4
Wu et al. [80]STTFormerUnsupervised Pre-train25X86.692.976.879.1
AimCLR [18]ST-GCNUnsupervised Pre-train25X86.992.880.180.9
3s-PSTL [94]ST-GCNUnsupervised Pre-train25X87.193.981.382.6
Colorization [87]DGCNNUnsupervised Pre-train25X88.094.9--
SSL(ours)STRLUnsupervised Pre-train17✓92.8(↑ 4.8)96.5(↑ 1.6)84.8(↑ 3.5)85.7(↑ 3.1)
", + "image_path": "cedfa05fb94bb2bcc24d5a9dc473424b2b851d11aaa06fbe9dcf5f951acc6bf0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 40, + 344, + 279, + 525 + ], + "blocks": [ + { + "bbox": [ + 38, + 301, + 538, + 327 + ], + "lines": [ + { + "bbox": [ + 38, + 301, + 538, + 327 + ], + "spans": [ + { + "bbox": [ + 38, + 301, + 538, + 327 + ], + "type": "text", + "content": "Table 2: The comparison with state-of-the-art unsupervised pre-train and supervised methods on NTU 60 and NTU 120 datasets. " + }, + { + "bbox": [ + 38, + 301, + 538, + 327 + ], + "type": "inline_equation", + "content": "\\underline{\\mathbf{\\Pi}}_{-}^{\\prime}" + }, + { + "bbox": [ + 38, + 301, + 538, + 327 + ], + "type": "text", + "content": " means the method with the second-best performance under unsupervised pre-training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 40, + 344, + 279, + 525 + ], + "lines": [ + { + "bbox": [ + 40, + 344, + 279, + 525 + ], + "spans": [ + { + "bbox": [ + 40, + 344, + 279, + 525 + ], + "type": "table", + "html": "
MethodPre-trainGFLOPsAcc.(%)
Fully Supervised
TSN [34]ImageNet-16.8
TRN [25]ImageNet-22.8
P3D [46]ImageNet-32.4
C3D [46]ImageNet-34.5
CorrNet [46]-74.835.5
CorrNet-R101 [75]ImageNet187.338.2
MG-TEA-ResNet50 [92]ImageNet-39.5
GSM [66]ImageNet107.440.3
TSM-R50 [29]ImageNet153.841.6
TMF [78]ImageNet-42.2
Unsupervised Pre-train
RESOUND-C3D [34]K400-16.4
Jenni et al. [23]K400-29.9
SSL (ours)Diving4842.834.1
", + "image_path": "09a29f0e00cc88705fa87307c3285afc739853624a2728fad9a6b921678e2c8e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 38, + 532, + 279, + 557 + ], + "lines": [ + { + "bbox": [ + 38, + 532, + 279, + 557 + ], + "spans": [ + { + "bbox": [ + 38, + 532, + 279, + 557 + ], + "type": "text", + "content": "Table 3: The comparison with the unsupervised pre-train and supervised methods on the Diving48 dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 38, + 578, + 279, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 578, + 279, + 674 + ], + "spans": [ + { + "bbox": [ + 38, + 578, + 279, + 674 + ], + "type": "text", + "content": "methods and achieves the best performance among unsupervised pre-train methods with RGB input. For the same input modality, our performance is lower than the fully supervised method PYSKL[11] (with the skeleton as input) by about " + }, + { + "bbox": [ + 38, + 578, + 279, + 674 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 38, + 578, + 279, + 674 + ], + "type": "text", + "content": ", because the PYSKL adopts stacks of visual heatmaps of skeleton joints as input while we only use human skeleton coordinates. This validates the promising discriminative ability of our skeleton sequence representation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 38, + 678, + 279, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 678, + 279, + 740 + ], + "spans": [ + { + "bbox": [ + 38, + 678, + 279, + 740 + ], + "type": "text", + "content": "Results on NTU 60 and NTU 120 Datasets. In Table 2, for NTU 60 X-sub and NTU 60 X-view, compared with unsupervised pre-train methods, our SSL outperforms the current state-of-the-art method Colorization [87] by " + }, + { + "bbox": [ + 38, + 678, + 279, + 740 + ], + "type": "inline_equation", + "content": "4.8\\%" + }, + { + "bbox": [ + 38, + 678, + 279, + 740 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 38, + 678, + 279, + 740 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 38, + 678, + 279, + 740 + ], + "type": "text", + "content": ", respectively. Our SSL is also competitive compared" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 296, + 346, + 538, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 346, + 538, + 466 + ], + "spans": [ + { + "bbox": [ + 296, + 346, + 538, + 466 + ], + "type": "text", + "content": "with fully supervised methods, outperforming the first six fully supervised methods on NTU 60 X-sub. For NTU 120 X-sub and NTU 120 X-set, our SSL outperforms the previous best-unsupervised pre-train method 3s-PSTL [94] by " + }, + { + "bbox": [ + 296, + 346, + 538, + 466 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 296, + 346, + 538, + 466 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 346, + 538, + 466 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 296, + 346, + 538, + 466 + ], + "type": "text", + "content": ", respectively. Our SSL is superior compared with some fully supervised methods on NTU 120 X-sub and NTU 120 X-set. These results show that our SSL can learn discriminative skeleton representation from large-scale action recognition datasets due to the promising generalization ability of our SkeletonMAE." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 472, + 537, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 472, + 537, + 556 + ], + "spans": [ + { + "bbox": [ + 296, + 472, + 537, + 556 + ], + "type": "text", + "content": "Results on Diving48 Dataset. Our SSL with skeleton input outperforms some fully supervised methods. Although our SSL is not pre-trained on additional large-scale pretraining action datasets in Table 3, it still achieves competitive performance among unsupervised pre-train methods. This validates that our pre-training model SkeletonMAE can learn discriminative skeleton sequence representation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 557, + 537, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 557, + 537, + 618 + ], + "spans": [ + { + "bbox": [ + 296, + 557, + 537, + 618 + ], + "type": "text", + "content": "The results on FineGym and Diving48 validate that our SkeletonMAE has a promising ability to enhance the feature representation of skeleton sequence by comprehensively perceiving the underlying topology of actions, and the SSL can learn discriminative action representation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 297, + 628, + 399, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 628, + 399, + 640 + ], + "spans": [ + { + "bbox": [ + 297, + 628, + 399, + 640 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 648, + 538, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 648, + 538, + 698 + ], + "spans": [ + { + "bbox": [ + 296, + 648, + 538, + 698 + ], + "type": "text", + "content": "In this section, we analyze the contributions of essential components and hyper-parameters of our model. Note that unless otherwise specified, all experiments are verified on the FineGym dataset with masking body sub-region 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 702, + 538, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 702, + 538, + 740 + ], + "spans": [ + { + "bbox": [ + 296, + 702, + 538, + 740 + ], + "type": "text", + "content": "Whether to load pre-trained model or not. To explore the effectiveness of loading the pre-trained SkeletonMAE encoder, we find that the accuracy is 86.3 without load" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 751, + 299, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 299, + 761 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 299, + 761 + ], + "type": "text", + "content": "5612" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 54, + 95, + 123, + 154 + ], + "blocks": [ + { + "bbox": [ + 54, + 95, + 123, + 154 + ], + "lines": [ + { + "bbox": [ + 54, + 95, + 123, + 154 + ], + "spans": [ + { + "bbox": [ + 54, + 95, + 123, + 154 + ], + "type": "table", + "html": "
LDMean Acc.
189.6
290.7
391.2
490.9
", + "image_path": "b06c6de66b65b037a9309186e6d837b69cb7e8e88d69c935ae255b77921b0800.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 138, + 103, + 273, + 150 + ], + "blocks": [ + { + "bbox": [ + 83, + 154, + 94, + 161 + ], + "lines": [ + { + "bbox": [ + 83, + 154, + 94, + 161 + ], + "spans": [ + { + "bbox": [ + 83, + 154, + 94, + 161 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 103, + 273, + 150 + ], + "lines": [ + { + "bbox": [ + 138, + 103, + 273, + 150 + ], + "spans": [ + { + "bbox": [ + 138, + 103, + 273, + 150 + ], + "type": "table", + "html": "
MethodMean Acc.
GraphCL [89]86.5
JOAO [88]88.7
Ours(SkeletonMAE)91.2
", + "image_path": "5e829114466724f13d860dd047a18f48d4dfdab9389fc4f842409003447d6777.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 44, + 171, + 273, + 218 + ], + "blocks": [ + { + "bbox": [ + 201, + 150, + 212, + 159 + ], + "lines": [ + { + "bbox": [ + 201, + 150, + 212, + 159 + ], + "spans": [ + { + "bbox": [ + 201, + 150, + 212, + 159 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 44, + 171, + 273, + 218 + ], + "lines": [ + { + "bbox": [ + 44, + 171, + 273, + 218 + ], + "spans": [ + { + "bbox": [ + 44, + 171, + 273, + 218 + ], + "type": "table", + "html": "
#Masked Body Part012345
GAT [72]86.888.188.989.589.490.0
GCN [28]87.688.989.390.689.590.5
GIN [82]88.689.590.291.290.391.2
", + "image_path": "788823b018c4e153357a2f7c5812de40f31698964bfaa9faf8354a92e29da8b7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 38, + 312, + 278, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 312, + 278, + 383 + ], + "spans": [ + { + "bbox": [ + 38, + 312, + 278, + 383 + ], + "type": "text", + "content": "ing the pre-trained SkeletonMAE encoder (randomly initialized weights). As Table 5(a) shows, loading the pretrained model is always better than not loading it. This validates that our SkeletonMAE can learn more comprehensive and generalized representations for unlabeled fine-grained actions by reconstructing the skeleton joint features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 38, + 390, + 279, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 390, + 279, + 426 + ], + "spans": [ + { + "bbox": [ + 38, + 390, + 279, + 426 + ], + "type": "text", + "content": "GIN layers in SkeletonMAE. Table 4(a) shows the performance of using different GIN layers in the SkeletonMAE encoder. The performance is the best when " + }, + { + "bbox": [ + 38, + 390, + 279, + 426 + ], + "type": "inline_equation", + "content": "L_{D} = 3" + }, + { + "bbox": [ + 38, + 390, + 279, + 426 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 38, + 432, + 279, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 432, + 279, + 612 + ], + "spans": [ + { + "bbox": [ + 38, + 432, + 279, + 612 + ], + "type": "text", + "content": "Comparison with contrastive learning methods. To verify the superior ability of our SkeletonMAE when conducting skeleton sequence pre-training, we compare our SkeletonMAE with different contrastive learning methods GraphCL and JOAO. As shown in Table 4(b), our SkeletonMAE achieves the best performance. Besides, we visually compare the action representations of SkeletonMAE and GraphCL by PCA, as shown in Figure 4(a) and Figure 4(b). Compared to GraphCL, the skeleton representation of our SkeletonMAE appears to have a larger inter-class variance and smaller intra-class variance. This validates that our SkeletonMAE can comprehensively capture the human pose and obtain discriminative skeleton sequence representation. We observe similar patterns in all other classes but visualize only five categories for simplicity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 38, + 618, + 279, + 738 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 618, + 279, + 738 + ], + "spans": [ + { + "bbox": [ + 38, + 618, + 279, + 738 + ], + "type": "text", + "content": "Backbones and masked body parts in SkeletonMAE. As shown in Table 4(c), we show the accuracy of our SSL with different SkeletonMAE backbones and different masked body parts in SkeletonMAE. It can be seen that GIN is always better than both GAT and GCN under the same masked body part. This is because that GIN provides a better inductive bias for graph-level applications. Thus, it is more suitable for learning more generalized skeleton representations. Additionally, we can see that masking body sub-regions 3 and 5 are both optimal across all backbones," + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 298, + 95, + 538, + 121 + ], + "blocks": [ + { + "bbox": [ + 153, + 218, + 165, + 227 + ], + "lines": [ + { + "bbox": [ + 153, + 218, + 165, + 227 + ], + "spans": [ + { + "bbox": [ + 153, + 218, + 165, + 227 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 38, + 228, + 279, + 287 + ], + "lines": [ + { + "bbox": [ + 38, + 228, + 279, + 287 + ], + "spans": [ + { + "bbox": [ + 38, + 228, + 279, + 287 + ], + "type": "text", + "content": "Table 4: (a) Mean accuracy of using the different number of GIN layers in SkeletonMAE encoder. (b) Comparison results with the contrastive learning method as the pre-trained encoder. (c) The results of using different backbones in SkeletonMAE under each masked body part are compared." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 298, + 95, + 538, + 121 + ], + "lines": [ + { + "bbox": [ + 298, + 95, + 538, + 121 + ], + "spans": [ + { + "bbox": [ + 298, + 95, + 538, + 121 + ], + "type": "table", + "html": "
M1234
Mean Acc.89.1(↑2.8)90.6(↑4.3)91.2(↑4.9)91.0(↑4.7)
", + "image_path": "3ef71e46cc99393e7d444e7f52199e30fc1208963dabbf502bb6c9ef8253e405.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 298, + 142, + 540, + 245 + ], + "blocks": [ + { + "bbox": [ + 411, + 122, + 423, + 132 + ], + "lines": [ + { + "bbox": [ + 411, + 122, + 423, + 132 + ], + "spans": [ + { + "bbox": [ + 411, + 122, + 423, + 132 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 298, + 142, + 540, + 245 + ], + "lines": [ + { + "bbox": [ + 298, + 142, + 540, + 245 + ], + "spans": [ + { + "bbox": [ + 298, + 142, + 540, + 245 + ], + "type": "table", + "html": "
#Masked Joints Number591215
Ratio of Mask Joints Accuracy of SSL30%50%70%90%
89.790.389.990.1
Masked Body PartHigh91.8 (V3,V5)91.2 (V0,V3,V5)91.0 (V1,V2,V3, V4,V5)90.8 (V0,V1,V2, V3,V5)
Low91.1 (V2,V4)90.1 (V0,V1)91.0 (V1,V2,V3, V4,V5)90.2 (V0,V1,V3, V4,V5)
", + "image_path": "89e3c714cbcc815314b4a3dcf39460b323e6a5877906b726a8b335aa65022df9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 411, + 245, + 423, + 253 + ], + "lines": [ + { + "bbox": [ + 411, + 245, + 423, + 253 + ], + "spans": [ + { + "bbox": [ + 411, + 245, + 423, + 253 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 296, + 260, + 537, + 320 + ], + "lines": [ + { + "bbox": [ + 296, + 260, + 537, + 320 + ], + "spans": [ + { + "bbox": [ + 296, + 260, + 537, + 320 + ], + "type": "text", + "content": "Table 5: (a) Results of four SSL variants. " + }, + { + "bbox": [ + 296, + 260, + 537, + 320 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 296, + 260, + 537, + 320 + ], + "type": "text", + "content": " represents the accuracy improvement relative to the random initialization of SkeletonMAE in SSL. (b) The comparison of our body part based masked and the random masked strategies. " + }, + { + "bbox": [ + 296, + 260, + 537, + 320 + ], + "type": "inline_equation", + "content": "\\nu_{0}-\\nu_{5}" + }, + { + "bbox": [ + 296, + 260, + 537, + 320 + ], + "type": "text", + "content": ": Head, Torso, Left arm, Right arm, Left leg, Right leg." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 296, + 346, + 536, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 346, + 536, + 370 + ], + "spans": [ + { + "bbox": [ + 296, + 346, + 536, + 370 + ], + "type": "text", + "content": "which demonstrates the importance of reconstruction of human limbs in action recognition." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 377, + 537, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 377, + 537, + 521 + ], + "spans": [ + { + "bbox": [ + 296, + 377, + 537, + 521 + ], + "type": "text", + "content": "Variants of SSL. To evaluate whether our pre-trained SkeletonMAE is effective across different skeleton action recognition models, we set the different number of STRL layers " + }, + { + "bbox": [ + 296, + 377, + 537, + 521 + ], + "type": "inline_equation", + "content": "(M = 1,2,3,4)" + }, + { + "bbox": [ + 296, + 377, + 537, + 521 + ], + "type": "text", + "content": " to obtain four variants of the SSL. As shown in Table 5(a), all SSL variants outperform the random initialization of SkeletonMAE in SSL, which validates our body part masking strategy indeed improves the discriminative ability of skeleton feature by learning action-sensitive visual concepts. Additionally, three-layer STRL is the best due to the good compromise between efficiency and computational cost. Moreover, it also validates that our SkeletonMAE generalizes well across different models." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 527, + 537, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 527, + 537, + 719 + ], + "spans": [ + { + "bbox": [ + 296, + 527, + 537, + 719 + ], + "type": "text", + "content": "Skeleton Masked Strategy. In Table 5(b), our masked body part strategy is fairly compared with the random masked strategy under the same masked joint conditions. Our method is better than the random mask method across all settings. As mentioned in Section 3.1.1, our masking strategy is action-sensitive and reconstructs specific limbs or body parts that dominates the given action class and is suitable for real-world skeleton-based action recognition. From Table 5, we can see that SSL effectively learns meaningful representations that encode the spatial relationships between joints, enabling the capture of crucial information about the structure and configuration of human movement. This notable achievement can be attributed to the integration of prior knowledge about the human body through pretraining the SkeletonMAE. We report results on the ntu 60 dataset in Supplementary Section E." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 298, + 726, + 536, + 738 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 726, + 536, + 738 + ], + "spans": [ + { + "bbox": [ + 298, + 726, + 536, + 738 + ], + "type": "text", + "content": "Transferability of the SkeletonMAE across datasets." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "text", + "content": "5613" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 43, + 96, + 119, + 162 + ], + "blocks": [ + { + "bbox": [ + 43, + 96, + 119, + 162 + ], + "lines": [ + { + "bbox": [ + 43, + 96, + 119, + 162 + ], + "spans": [ + { + "bbox": [ + 43, + 96, + 119, + 162 + ], + "type": "image", + "image_path": "1bd449f4e6909eafdc3815d68bdde47a97bcb875710a756a5e4db1e849cc13de.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 164, + 89, + 173 + ], + "lines": [ + { + "bbox": [ + 78, + 164, + 89, + 173 + ], + "spans": [ + { + "bbox": [ + 78, + 164, + 89, + 173 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 120, + 96, + 196, + 162 + ], + "blocks": [ + { + "bbox": [ + 120, + 96, + 196, + 162 + ], + "lines": [ + { + "bbox": [ + 120, + 96, + 196, + 162 + ], + "spans": [ + { + "bbox": [ + 120, + 96, + 196, + 162 + ], + "type": "image", + "image_path": "255c89d302dfcd18f7cc9283a17a426442beeea21573ccfc133c175861011d0e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 152, + 164, + 164, + 174 + ], + "lines": [ + { + "bbox": [ + 152, + 164, + 164, + 174 + ], + "spans": [ + { + "bbox": [ + 152, + 164, + 164, + 174 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 199, + 95, + 276, + 162 + ], + "blocks": [ + { + "bbox": [ + 199, + 95, + 276, + 162 + ], + "lines": [ + { + "bbox": [ + 199, + 95, + 276, + 162 + ], + "spans": [ + { + "bbox": [ + 199, + 95, + 276, + 162 + ], + "type": "image", + "image_path": "b0f61c7d2d6aa89463196bc696b5d714b3f682d1efda655f9efa874ac41b16ed.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 164, + 238, + 174 + ], + "lines": [ + { + "bbox": [ + 227, + 164, + 238, + 174 + ], + "spans": [ + { + "bbox": [ + 227, + 164, + 238, + 174 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 38, + 180, + 278, + 228 + ], + "lines": [ + { + "bbox": [ + 38, + 180, + 278, + 228 + ], + "spans": [ + { + "bbox": [ + 38, + 180, + 278, + 228 + ], + "type": "text", + "content": "Figure 4: (a) and (b) are 2d-PCA of SkeletonMAE and GraphCL as pre-trained encoder representations. We randomly select five action classes for 2d-PCA visualization, (c) Confusion matrix for fine-grained action recognition." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 38, + 241, + 279, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 241, + 279, + 349 + ], + "spans": [ + { + "bbox": [ + 38, + 241, + 279, + 349 + ], + "type": "text", + "content": "As shown in Figure 5(a), we pre-train SkeletonMAE on the FineGym dataset and then fine-tune it on NTU 60 X-sub and NTU 120 X-sub datasets. Compared with the method that uses the same dataset for pre-training and fine-tuning, our SkeletonMAE achieves better performance across all masked strategies when conducting dataset transfer. This shows that the SkeletonMAE can learn generalized skeleton representation and effectively transfer the strong representation ability to other datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 39, + 360, + 164, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 360, + 164, + 373 + ], + "spans": [ + { + "bbox": [ + 39, + 360, + 164, + 373 + ], + "type": "text", + "content": "4.5. Visualization Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 38, + 380, + 279, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 380, + 279, + 548 + ], + "spans": [ + { + "bbox": [ + 38, + 380, + 279, + 548 + ], + "type": "text", + "content": "Figure 6 shows the reconstruction process of the skeleton sequence by SkeletonMAE. From the same frame, the difference between the reconstructed skeleton sequence and the original skeleton sequence is slight, but overall the human body structure is reserved. This shows that the SkeletonMAE has good spatial representation learning ability. Moreover, SSL effectively captures the temporal evolution and distinguishing characteristics of actions by capturing the relationships between consecutive joint positions and poses. This highlights the positive impact of the multi-scale temporal dependence incorporated within the STRL module. Figure 4(c) shows that our SSL works well for fine-grained action recognition tasks on the FineGym dataset. More visualization results are in Supplementary Section C." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 39, + 561, + 111, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 561, + 111, + 574 + ], + "spans": [ + { + "bbox": [ + 39, + 561, + 111, + 574 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 38, + 583, + 278, + 738 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 583, + 278, + 738 + ], + "spans": [ + { + "bbox": [ + 38, + 583, + 278, + 738 + ], + "type": "text", + "content": "In this paper, we propose an efficient skeleton sequence learning framework SSL, to learn discriminative skeleton-based action representation. To comprehensively capture the human pose and obtain skeleton sequence representation, we propose a graph-based encoder-decoder pretraining architecture, SkeletonMAE, that embeds skeleton joint sequence into GCN and utilize the prior human topology knowledge to guide the reconstruction of the underlying masked joints and topology. Extensive experimental results show that our SSL achieves SOTA performance on four benchmark skeleton-based action recognition datasets. In our future work, we will build a multi-level feature refinement module to identify ambiguous skeleton actions." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 334, + 95, + 501, + 192 + ], + "blocks": [ + { + "bbox": [ + 334, + 95, + 501, + 192 + ], + "lines": [ + { + "bbox": [ + 334, + 95, + 501, + 192 + ], + "spans": [ + { + "bbox": [ + 334, + 95, + 501, + 192 + ], + "type": "image", + "image_path": "54dd7b564b7cc918b149d8eeec8ece73b720a305fd5500f6b811f23c1e2e00ad.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 193, + 480, + 203 + ], + "lines": [ + { + "bbox": [ + 381, + 193, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 381, + 193, + 480, + 203 + ], + "type": "text", + "content": "(a) Masked body part" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 334, + 217, + 501, + 314 + ], + "blocks": [ + { + "bbox": [ + 334, + 217, + 501, + 314 + ], + "lines": [ + { + "bbox": [ + 334, + 217, + 501, + 314 + ], + "spans": [ + { + "bbox": [ + 334, + 217, + 501, + 314 + ], + "type": "image", + "image_path": "c03a1682afb7420414d6f69e2ce248a81872f4a2f53ddb88d23e843002a2d942.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 314, + 479, + 325 + ], + "lines": [ + { + "bbox": [ + 381, + 314, + 479, + 325 + ], + "spans": [ + { + "bbox": [ + 381, + 314, + 479, + 325 + ], + "type": "text", + "content": "(b) Masked body part" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 296, + 326, + 537, + 374 + ], + "lines": [ + { + "bbox": [ + 296, + 326, + 537, + 374 + ], + "spans": [ + { + "bbox": [ + 296, + 326, + 537, + 374 + ], + "type": "text", + "content": "Figure 5: (a) The accuracies with mask body part of 0-5 on the NTU 60 X-sub dataset, and (b) the accuracies on the NTU 120 X-sub dataset. \\* means SkeletonMAE encoder pre-trained on the FineGym dataset." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 299, + 389, + 535, + 429 + ], + "blocks": [ + { + "bbox": [ + 299, + 389, + 535, + 429 + ], + "lines": [ + { + "bbox": [ + 299, + 389, + 535, + 429 + ], + "spans": [ + { + "bbox": [ + 299, + 389, + 535, + 429 + ], + "type": "image", + "image_path": "dea9acba65f830331750a0eee144a9338b3fc3bd9fce2d4313fe411e01c39792.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 430, + 439, + 437 + ], + "lines": [ + { + "bbox": [ + 395, + 430, + 439, + 437 + ], + "spans": [ + { + "bbox": [ + 395, + 430, + 439, + 437 + ], + "type": "text", + "content": "(a) Original joint" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 299, + 442, + 534, + 475 + ], + "blocks": [ + { + "bbox": [ + 299, + 442, + 534, + 475 + ], + "lines": [ + { + "bbox": [ + 299, + 442, + 534, + 475 + ], + "spans": [ + { + "bbox": [ + 299, + 442, + 534, + 475 + ], + "type": "image", + "image_path": "9f5ff0b48515ec9109f88be3e6505e8309f047ece657364b22483d5e70580491.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 475, + 440, + 483 + ], + "lines": [ + { + "bbox": [ + 394, + 475, + 440, + 483 + ], + "spans": [ + { + "bbox": [ + 394, + 475, + 440, + 483 + ], + "type": "text", + "content": "(b) Masked joint" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 299, + 488, + 534, + 521 + ], + "blocks": [ + { + "bbox": [ + 299, + 488, + 534, + 521 + ], + "lines": [ + { + "bbox": [ + 299, + 488, + 534, + 521 + ], + "spans": [ + { + "bbox": [ + 299, + 488, + 534, + 521 + ], + "type": "image", + "image_path": "70ebc3ed8c947882234cde932c6c1488f1e55a263922d0030cee929599c22391.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 522, + 444, + 530 + ], + "lines": [ + { + "bbox": [ + 381, + 522, + 444, + 530 + ], + "spans": [ + { + "bbox": [ + 381, + 522, + 444, + 530 + ], + "type": "text", + "content": "(c) Reconstructed joint" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 296, + 537, + 537, + 585 + ], + "lines": [ + { + "bbox": [ + 296, + 537, + 537, + 585 + ], + "spans": [ + { + "bbox": [ + 296, + 537, + 537, + 585 + ], + "type": "text", + "content": "Figure 6: Visualization results for skeleton sequence of \"aerial walkover forward\" on FineGym. (a) The input skeleton sequence. (b) Masked skeleton sequence (masked parts are 3 and 5). (c) Reconstructed skeleton sequence." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 298, + 608, + 395, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 608, + 395, + 623 + ], + "spans": [ + { + "bbox": [ + 298, + 608, + 395, + 623 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "spans": [ + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "text", + "content": "This work is supported in part by the National Key R&D Program of China under Grant No.2021ZD0111601, in part by the National Natural Science Foundation of China under Grant No.62002395 and No.61976250, in part by the Guangdong Basic and Applied Basic Research Foundation under Grant No.2023A1515011530, No.2021A1515012311 and No.2020B1515020048, and in part by the Guangzhou Science and Technology Planning Project under Grant No. 2023A04J2030." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "text", + "content": "5614" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 39, + 95, + 98, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 95, + 98, + 108 + ], + "spans": [ + { + "bbox": [ + 39, + 95, + 98, + 108 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 40, + 116, + 279, + 737 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 44, + 116, + 279, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 44, + 116, + 279, + 159 + ], + "spans": [ + { + "bbox": [ + 44, + 116, + 279, + 159 + ], + "type": "text", + "content": "[1] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 160, + 279, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 160, + 279, + 204 + ], + "spans": [ + { + "bbox": [ + 45, + 160, + 279, + 204 + ], + "type": "text", + "content": "[2] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 206, + 278, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 206, + 278, + 260 + ], + "spans": [ + { + "bbox": [ + 45, + 206, + 278, + 260 + ], + "type": "text", + "content": "[3] Minghao Chen, Fangyun Wei, Chong Li, and Deng Cai. Frame-wise action representations for long videos via sequence contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13801-13810, 2022. 1, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 261, + 278, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 261, + 278, + 315 + ], + "spans": [ + { + "bbox": [ + 45, + 261, + 278, + 315 + ], + "type": "text", + "content": "[4] Tailin Chen, Desen Zhou, Jian Wang, Shidong Wang, Yu Guan, Xuming He, and Errui Ding. Learning multi-granular spatio-temporal graph network for skeleton-based action recognition. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4334-4342, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 316, + 278, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 316, + 278, + 372 + ], + "spans": [ + { + "bbox": [ + 45, + 316, + 278, + 372 + ], + "type": "text", + "content": "[5] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13359-13368, 2021. 2, 3, 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 373, + 279, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 373, + 279, + 426 + ], + "spans": [ + { + "bbox": [ + 45, + 373, + 279, + 426 + ], + "type": "text", + "content": "[6] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 183-192, 2020. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 427, + 278, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 427, + 278, + 481 + ], + "spans": [ + { + "bbox": [ + 45, + 427, + 278, + 481 + ], + "type": "text", + "content": "[7] Sangwoo Cho, Muhammad Maqbool, Fei Liu, and Hassan Foroosh. Self-attention network for skeleton-based human action recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 635–644, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 483, + 279, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 483, + 279, + 537 + ], + "spans": [ + { + "bbox": [ + 45, + 483, + 279, + 537 + ], + "type": "text", + "content": "[8] Xiao Chu, Wei Yang, Wanli Ouyang, Cheng Ma, Alan L Yuille, and Xiaogang Wang. Multi-context attention for human pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1831-1840, 2017. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 538, + 278, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 538, + 278, + 582 + ], + "spans": [ + { + "bbox": [ + 45, + 538, + 278, + 582 + ], + "type": "text", + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 40, + 583, + 278, + 627 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 583, + 278, + 627 + ], + "spans": [ + { + "bbox": [ + 40, + 583, + 278, + 627 + ], + "type": "text", + "content": "[10] Yong Du, Wei Wang, and Liang Wang. Hierarchical recurrent neural network for skeleton based action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1110-1118, 2015. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 40, + 628, + 278, + 682 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 628, + 278, + 682 + ], + "spans": [ + { + "bbox": [ + 40, + 628, + 278, + 682 + ], + "type": "text", + "content": "[11] Haodong Duan, Yue Zhao, Kai Chen, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2969-2978, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 40, + 683, + 278, + 737 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 683, + 278, + 737 + ], + "spans": [ + { + "bbox": [ + 40, + 683, + 278, + 737 + ], + "type": "text", + "content": "[12] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. Temporal cycle-consistency learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1801-1810, 2019. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 98, + 537, + 737 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 299, + 98, + 537, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 98, + 537, + 142 + ], + "spans": [ + { + "bbox": [ + 299, + 98, + 537, + 142 + ], + "type": "text", + "content": "[13] Yin Fan, Xiangju Lu, Dian Li, and Yuanliu Liu. Video-based emotion recognition using cnn-rnn and c3d hybrid networks. In Proceedings of the 18th ACM international conference on multimodal interaction, pages 445-450, 2016. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 143, + 537, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 143, + 537, + 187 + ], + "spans": [ + { + "bbox": [ + 299, + 143, + 537, + 187 + ], + "type": "text", + "content": "[14] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 188, + 537, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 188, + 537, + 221 + ], + "spans": [ + { + "bbox": [ + 299, + 188, + 537, + 221 + ], + "type": "text", + "content": "[15] Jerome H Friedman. On bias, variance, 0/1—loss, and the curse-of-dimensionality. Data mining and knowledge discovery, 1(1):55-77, 1997. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 223, + 537, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 223, + 537, + 288 + ], + "spans": [ + { + "bbox": [ + 299, + 223, + 537, + 288 + ], + "type": "text", + "content": "[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 289, + 537, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 289, + 537, + 334 + ], + "spans": [ + { + "bbox": [ + 299, + 289, + 537, + 334 + ], + "type": "text", + "content": "[17] Aditya Grover and Jure Leskovec. node2vec: Scalable feature learning for networks. In Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining, pages 855-864, 2016. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 335, + 537, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 335, + 537, + 399 + ], + "spans": [ + { + "bbox": [ + 299, + 335, + 537, + 399 + ], + "type": "text", + "content": "[18] Tianyu Guo, Hong Liu, Zhan Chen, Mengyuan Liu, Tao Wang, and Runwei Ding. Contrastive learning from extremely augmented skeleton sequences for self-supervised action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 762-770, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 401, + 537, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 401, + 537, + 456 + ], + "spans": [ + { + "bbox": [ + 299, + 401, + 537, + 456 + ], + "type": "text", + "content": "[19] Isma Hadji, Konstantinos G Derpanis, and Allan D Jepson. Representation learning via global temporal alignment and cycle-consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11068-11077, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 457, + 537, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 457, + 537, + 512 + ], + "spans": [ + { + "bbox": [ + 299, + 457, + 537, + 512 + ], + "type": "text", + "content": "[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 514, + 537, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 514, + 537, + 547 + ], + "spans": [ + { + "bbox": [ + 299, + 514, + 537, + 547 + ], + "type": "text", + "content": "[21] Sepp Hochreiter, Yoshua Bengio, Paolo Frasconi, Jürgen Schmidhuber, et al. Gradient flow in recurrent nets: the difficulty of learning long-term dependencies, 2001. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 548, + 537, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 548, + 537, + 603 + ], + "spans": [ + { + "bbox": [ + 299, + 548, + 537, + 603 + ], + "type": "text", + "content": "[22] Zhenyu Hou, Xiao Liu, Yukuo Cen, Yuxiao Dong, Hongxia Yang, Chunjie Wang, and Jie Tang. Graphmae: Self-supervised masked graph autoencoders. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 594-604, 2022. 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 604, + 537, + 648 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 604, + 537, + 648 + ], + "spans": [ + { + "bbox": [ + 299, + 604, + 537, + 648 + ], + "type": "text", + "content": "[23] Simon Jenni and Hailin Jin. Time-equivariant contrastive video representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9970-9980, 2021. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 650, + 537, + 682 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 650, + 537, + 682 + ], + "spans": [ + { + "bbox": [ + 299, + 650, + 537, + 682 + ], + "type": "text", + "content": "[24] Gunnar Johansson. Visual perception of biological motion and a model for its analysis. Perception & psychophysics, 14(2):201-211, 1973. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 683, + 537, + 737 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 683, + 537, + 737 + ], + "spans": [ + { + "bbox": [ + 299, + 683, + 537, + 737 + ], + "type": "text", + "content": "[25] Gagan Kanojia, Sudhakar Kumawat, and Shanmuganathan Raman. Attentive spatio-temporal representation learning for diving classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "text", + "content": "5615" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 40, + 98, + 279, + 737 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 40, + 98, + 279, + 153 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 98, + 279, + 153 + ], + "spans": [ + { + "bbox": [ + 40, + 98, + 279, + 153 + ], + "type": "text", + "content": "[26] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1725–1732, 2014. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 40, + 154, + 278, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 154, + 278, + 207 + ], + "spans": [ + { + "bbox": [ + 40, + 154, + 278, + 207 + ], + "type": "text", + "content": "[27] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 41, + 210, + 278, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 210, + 278, + 242 + ], + "spans": [ + { + "bbox": [ + 41, + 210, + 278, + 242 + ], + "type": "text", + "content": "[28] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 3, 5, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 41, + 245, + 278, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 245, + 278, + 298 + ], + "spans": [ + { + "bbox": [ + 41, + 245, + 278, + 298 + ], + "type": "text", + "content": "[29] Heeseung Kwon, Manjin Kim, Suha Kwak, and Minsu Cho. Learning self-similarity in space and time as generalized motion for video action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13065-13075, 2021. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 41, + 301, + 278, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 301, + 278, + 355 + ], + "spans": [ + { + "bbox": [ + 41, + 301, + 278, + 355 + ], + "type": "text", + "content": "[30] Mei Chee Leong, Hui Li Tan, Haosong Zhang, Liyuan Li, Feng Lin, and Joo Hwee Lim. Joint learning on the hierarchy representation for fine-grained human action recognition. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1059-1063. IEEE, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 41, + 357, + 278, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 357, + 278, + 421 + ], + "spans": [ + { + "bbox": [ + 41, + 357, + 278, + 421 + ], + "type": "text", + "content": "[31] Linguo Li, Minsi Wang, Bingbing Ni, Hang Wang, Jiancheng Yang, and Wenjun Zhang. 3d human action representation learning via cross-view consistency pursuit. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4741-4750, 2021. 2, 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 41, + 424, + 278, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 424, + 278, + 479 + ], + "spans": [ + { + "bbox": [ + 41, + 424, + 278, + 479 + ], + "type": "text", + "content": "[32] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3595–3603, 2019. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 41, + 480, + 278, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 480, + 278, + 535 + ], + "spans": [ + { + "bbox": [ + 41, + 480, + 278, + 535 + ], + "type": "text", + "content": "[33] Tianjiao Li, Lin Geng Foo, Qiuhong Ke, Hossein Rahmani, Anran Wang, Jinghua Wang, and Jun Liu. Dynamic spatio-temporal specialization learning for fine-grained action recognition. In European Conference on Computer Vision, pages 386-403. Springer, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 41, + 537, + 278, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 537, + 278, + 580 + ], + "spans": [ + { + "bbox": [ + 41, + 537, + 278, + 580 + ], + "type": "text", + "content": "[34] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018. 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 41, + 581, + 278, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 581, + 278, + 625 + ], + "spans": [ + { + "bbox": [ + 41, + 581, + 278, + 625 + ], + "type": "text", + "content": "[35] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7083-7093, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 41, + 628, + 278, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 628, + 278, + 681 + ], + "spans": [ + { + "bbox": [ + 41, + 628, + 278, + 681 + ], + "type": "text", + "content": "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 41, + 683, + 278, + 737 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 683, + 278, + 737 + ], + "spans": [ + { + "bbox": [ + 41, + 683, + 278, + 737 + ], + "type": "text", + "content": "[37] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10):2684-2701, 2019. 2, 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 98, + 536, + 737 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 299, + 98, + 536, + 130 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 98, + 536, + 130 + ], + "spans": [ + { + "bbox": [ + 299, + 98, + 536, + 130 + ], + "type": "text", + "content": "[38] Mengyuan Liu, Hong Liu, and Chen Chen. Enhanced skeleton visualization for view invariant human action recognition. Pattern Recognition, 68:346-362, 2017. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 299, + 133, + 536, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 133, + 536, + 175 + ], + "spans": [ + { + "bbox": [ + 299, + 133, + 536, + 175 + ], + "type": "text", + "content": "[39] Yang Liu, Guanbin Li, and Liang Lin. Cross-modal causal relational reasoning for event-level visual question answering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 178, + 536, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 178, + 536, + 220 + ], + "spans": [ + { + "bbox": [ + 299, + 178, + 536, + 220 + ], + "type": "text", + "content": "[40] Yang Liu, Zhaoyang Lu, Jing Li, and Tao Yang. Hierarchically learned view-invariant representations for cross-view action recognition. IEEE Transactions on Circuits and Systems for Video Technology, 29(8):2416-2430, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 223, + 536, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 223, + 536, + 265 + ], + "spans": [ + { + "bbox": [ + 299, + 223, + 536, + 265 + ], + "type": "text", + "content": "[41] Yang Liu, Zhaoyang Lu, Jing Li, Tao Yang, and Chao Yao. Deep image-to-video adaptation and fusion networks for action recognition. IEEE Transactions on Image Processing, 29:3168-3182, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 267, + 536, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 267, + 536, + 310 + ], + "spans": [ + { + "bbox": [ + 299, + 267, + 536, + 310 + ], + "type": "text", + "content": "[42] Yang Liu, Keze Wang, Guanbin Li, and Liang Lin. Semantics-aware adaptive knowledge distillation for sensor-to-vision action recognition. IEEE Transactions on Image Processing, 30:5573-5588, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 313, + 536, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 313, + 536, + 355 + ], + "spans": [ + { + "bbox": [ + 299, + 313, + 536, + 355 + ], + "type": "text", + "content": "[43] Yang Liu, Keze Wang, Lingbo Liu, Haoyuan Lan, and Liang Lin. Tcgl: Temporal contrastive graph for self-supervised video representation learning. IEEE Transactions on Image Processing, 31:1978-1993, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 357, + 536, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 357, + 536, + 399 + ], + "spans": [ + { + "bbox": [ + 299, + 357, + 536, + 399 + ], + "type": "text", + "content": "[44] Yang Liu, Yu-Shen Wei, Hong Yan, Guan-Bin Li, and Liang Lin. Causal reasoning meets visual representation learning: A prospective study. Machine Intelligence Research, 19(6):485-511, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 402, + 536, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 402, + 536, + 457 + ], + "spans": [ + { + "bbox": [ + 299, + 402, + 536, + 457 + ], + "type": "text", + "content": "[45] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 143-152, 2020. 1, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 458, + 536, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 458, + 536, + 501 + ], + "spans": [ + { + "bbox": [ + 299, + 458, + 536, + 501 + ], + "type": "text", + "content": "[46] Chenxu Luo and Alan L Yuille. Grouped spatial-temporal aggregation for efficient action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5512-5521, 2019. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 504, + 536, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 504, + 536, + 547 + ], + "spans": [ + { + "bbox": [ + 299, + 504, + 536, + 547 + ], + "type": "text", + "content": "[47] Fengjun Lv and Ramakant Nevatia. Recognition and segmentation of 3-d human action using hmm and multi-class adaboost. In European conference on computer vision, pages 359-372. Springer, 2006. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 548, + 536, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 548, + 536, + 591 + ], + "spans": [ + { + "bbox": [ + 299, + 548, + 536, + 591 + ], + "type": "text", + "content": "[48] Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2891-2900, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 593, + 536, + 637 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 593, + 536, + 637 + ], + "spans": [ + { + "bbox": [ + 299, + 593, + 536, + 637 + ], + "type": "text", + "content": "[49] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European conference on computer vision, pages 527-544. Springer, 2016. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 639, + 536, + 682 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 639, + 536, + 682 + ], + "spans": [ + { + "bbox": [ + 299, + 639, + 536, + 682 + ], + "type": "text", + "content": "[50] Jonathan Munro and Dima Damen. Multi-modal domain adaptation for fine-grained action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 122-132, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 684, + 536, + 737 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 684, + 536, + 737 + ], + "spans": [ + { + "bbox": [ + 299, + 684, + 536, + 737 + ], + "type": "text", + "content": "[51] Ashish S Nikam and Aarti G Ambekar. Sign language recognition using image based hand gesture recognition techniques. In 2016 online international conference on green engineering and technologies (IC-GET), pages 1-5. IEEE, 2016. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "text", + "content": "5616" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 40, + 98, + 279, + 738 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 40, + 98, + 279, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 98, + 279, + 175 + ], + "spans": [ + { + "bbox": [ + 40, + 98, + 279, + 175 + ], + "type": "text", + "content": "[52] Cosmas Ifeanyi Nwakanma, Fabliha Bushra Islam, Mareska Pratiwi Maharani, Dong-Seong Kim, and JaeMin Lee. Iot-based vibration sensor data collection and emergency detection classification using long short term memory (lstm). In 2021 International Conference on Artificial Intelligence in Information and Communication (ICAIIC), pages 273-278. IEEE, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 40, + 175, + 279, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 175, + 279, + 229 + ], + "spans": [ + { + "bbox": [ + 40, + 175, + 279, + 229 + ], + "type": "text", + "content": "[53] Jiwong Park, Minsk Lee, Hyung Jin Chang, Kyuewang Lee, and Jin Young Choi. Symmetric graph convolutional autoencoder for unsupervised graph representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6519-6528, 2019. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 41, + 230, + 278, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 230, + 278, + 274 + ], + "spans": [ + { + "bbox": [ + 41, + 230, + 278, + 274 + ], + "type": "text", + "content": "[54] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 41, + 275, + 278, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 275, + 278, + 318 + ], + "spans": [ + { + "bbox": [ + 41, + 275, + 278, + 318 + ], + "type": "text", + "content": "[55] Haocong Rao, Shihao Xu, Xiping Hu, Jun Cheng, and Bin Hu. Augmented skeleton based contrastive action learning with momentum lstm for unsupervised action recognition. Information Sciences, 569:90-109, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 41, + 319, + 278, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 319, + 278, + 362 + ], + "spans": [ + { + "bbox": [ + 41, + 319, + 278, + 362 + ], + "type": "text", + "content": "[56] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 41, + 363, + 278, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 363, + 278, + 384 + ], + "spans": [ + { + "bbox": [ + 41, + 363, + 278, + 384 + ], + "type": "text", + "content": "[57] Amin Salehi and Hasan Davulcu. Graph attention autoencoders. arXiv preprint arXiv:1905.10715, 2019. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 41, + 385, + 279, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 385, + 279, + 429 + ], + "spans": [ + { + "bbox": [ + 41, + 385, + 279, + 429 + ], + "type": "text", + "content": "[58] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 41, + 430, + 278, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 430, + 278, + 484 + ], + "spans": [ + { + "bbox": [ + 41, + 430, + 278, + 484 + ], + "type": "text", + "content": "[59] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2616-2625, 2020. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 41, + 485, + 279, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 485, + 279, + 539 + ], + "spans": [ + { + "bbox": [ + 41, + 485, + 279, + 539 + ], + "type": "text", + "content": "[60] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12026-12035, 2019. 1, 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 41, + 540, + 278, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 540, + 278, + 594 + ], + "spans": [ + { + "bbox": [ + 41, + 540, + 278, + 594 + ], + "type": "text", + "content": "[61] Lei Shi, Yifan Zhang, Jing Hu, Jian Cheng, and Hanqing Lu. Gesture recognition using spatiotemporal deformable convolutional representation. In 2019 IEEE International Conference on Image Processing (ICIP), pages 1900-1904. IEEE, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 41, + 595, + 278, + 649 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 595, + 278, + 649 + ], + "spans": [ + { + "bbox": [ + 41, + 595, + 278, + 649 + ], + "type": "text", + "content": "[62] Sijie Song, Cuiling Lan, Junliang Xing, Wenjun Zeng, and Jiaying Liu. An end-to-end spatio-temporal attention model for human action recognition from skeleton data. In Proceedings of the AAAI conference on artificial intelligence, volume 31, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 41, + 650, + 278, + 694 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 650, + 278, + 694 + ], + "spans": [ + { + "bbox": [ + 41, + 650, + 278, + 694 + ], + "type": "text", + "content": "[63] Sijie Song, Cuiling Lan, Junliang Xing, Wenjun Zeng, and Jiaying Liu. Spatio-temporal attention-based LSTM networks for 3d action recognition and detection. IEEE Transactions on image processing, 27(7):3459-3471, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 41, + 695, + 278, + 738 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 695, + 278, + 738 + ], + "spans": [ + { + "bbox": [ + 41, + 695, + 278, + 738 + ], + "type": "text", + "content": "[64] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In International conference on machine learning, pages 843-852. PMLR, 2015. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 98, + 537, + 738 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 299, + 98, + 537, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 98, + 537, + 142 + ], + "spans": [ + { + "bbox": [ + 299, + 98, + 537, + 142 + ], + "type": "text", + "content": "[65] Kun Su, Xiulong Liu, and Eli Shlizerman. Predict & cluster: Unsupervised skeleton based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9631-9640, 2020. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 143, + 537, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 143, + 537, + 186 + ], + "spans": [ + { + "bbox": [ + 299, + 143, + 537, + 186 + ], + "type": "text", + "content": "[66] Swathikiran Sudhakaran, Sergio Escalera, and Oswald Lanz. Gate-shift networks for video action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1102-1111, 2020. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 187, + 537, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 187, + 537, + 240 + ], + "spans": [ + { + "bbox": [ + 299, + 187, + 537, + 240 + ], + "type": "text", + "content": "[67] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5693-5703, 2019. 1, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 242, + 537, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 242, + 537, + 286 + ], + "spans": [ + { + "bbox": [ + 299, + 242, + 537, + 286 + ], + "type": "text", + "content": "[68] Jian Tang, Meng Qu, Mingzhe Wang, Ming Zhang, Jun Yan, and Qiao zhu Mei. Line: Large-scale information network embedding. In Proceedings of the 24th international conference on world wide web, pages 1067-1077, 2015. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 287, + 537, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 287, + 537, + 330 + ], + "spans": [ + { + "bbox": [ + 299, + 287, + 537, + 330 + ], + "type": "text", + "content": "[69] Fida Mohammad Thoker, Hazel Doughty, and Cees GM Snoek. Skeleton-contrastive 3d action representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1655–1663, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 330, + 537, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 330, + 537, + 373 + ], + "spans": [ + { + "bbox": [ + 299, + 330, + 537, + 373 + ], + "type": "text", + "content": "[70] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 374, + 537, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 374, + 537, + 429 + ], + "spans": [ + { + "bbox": [ + 299, + 374, + 537, + 429 + ], + "type": "text", + "content": "[71] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 430, + 537, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 430, + 537, + 463 + ], + "spans": [ + { + "bbox": [ + 299, + 430, + 537, + 463 + ], + "type": "text", + "content": "[72] Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, Yoshua Bengio, et al. Graph attention networks. stat, 1050(20):10-48550, 2017. 3, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 463, + 537, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 463, + 537, + 516 + ], + "spans": [ + { + "bbox": [ + 299, + 463, + 537, + 516 + ], + "type": "text", + "content": "[73] Raviteja Vemulapalli, Felipe Arrate, and Rama Chellappa. Human action recognition by representing 3d skeletons as points in a lie group. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 588-595, 2014. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 517, + 537, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 517, + 537, + 562 + ], + "spans": [ + { + "bbox": [ + 299, + 517, + 537, + 562 + ], + "type": "text", + "content": "[74] Raviteja Vemulapalli and Rama Chellapa. Rolling rotations for recognizing human actions from 3d skeletal data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4471-4479, 2016. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 562, + 537, + 606 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 562, + 537, + 606 + ], + "spans": [ + { + "bbox": [ + 299, + 562, + 537, + 606 + ], + "type": "text", + "content": "[75] Heng Wang, Du Tran, Lorenzo Torresani, and Matt Feiszli. Video modeling with correlation networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 352-361, 2020. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 607, + 537, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 607, + 537, + 650 + ], + "spans": [ + { + "bbox": [ + 299, + 607, + 537, + 650 + ], + "type": "text", + "content": "[76] Jiang Wang, Zicheng Liu, Ying Wu, and Junsong Yuan. Mining actionlet ensemble for action recognition with depth cameras. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 1290-1297. IEEE, 2012. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 650, + 537, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 650, + 537, + 704 + ], + "spans": [ + { + "bbox": [ + 299, + 650, + 537, + 704 + ], + "type": "text", + "content": "[77] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 705, + 537, + 738 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 705, + 537, + 738 + ], + "spans": [ + { + "bbox": [ + 299, + 705, + 537, + 738 + ], + "type": "text", + "content": "[78] Yanze Wang and Junyong Ye. Tmf: Temporal motion and fusion for action recognition. Computer Vision and Image Understanding, 213:103304, 2021. 7" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "text", + "content": "5617" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 40, + 98, + 279, + 738 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 40, + 98, + 279, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 98, + 279, + 142 + ], + "spans": [ + { + "bbox": [ + 40, + 98, + 279, + 142 + ], + "type": "text", + "content": "[79] Zhengwei Wang, Qi She, and Aljosa Smolic. Action-net: Multipath excitation for action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13214-13223, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 40, + 143, + 279, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 143, + 279, + 186 + ], + "spans": [ + { + "bbox": [ + 40, + 143, + 279, + 186 + ], + "type": "text", + "content": "[80] Wenhan Wu, Yilei Hua, Shiqian Wu, Chen Chen, Aidong Lu, et al. Skeletonmae: Spatial-temporal masked autoencoders for self-supervised skeleton action recognition. arXiv preprint arXiv:2209.02399, 2022. 3, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 41, + 188, + 278, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 188, + 278, + 243 + ], + "spans": [ + { + "bbox": [ + 41, + 188, + 278, + 243 + ], + "type": "text", + "content": "[81] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In Proceedings of the European conference on computer vision (ECCV), pages 305–321, 2018. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 41, + 244, + 278, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 244, + 278, + 277 + ], + "spans": [ + { + "bbox": [ + 41, + 244, + 278, + 277 + ], + "type": "text", + "content": "[82] Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826, 2018. 3, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 41, + 278, + 278, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 278, + 278, + 321 + ], + "spans": [ + { + "bbox": [ + 41, + 278, + 278, + 321 + ], + "type": "text", + "content": "[83] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Thirty-second AAAI conference on artificial intelligence, 2018. 1, 2, 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 41, + 323, + 279, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 323, + 279, + 378 + ], + "spans": [ + { + "bbox": [ + 41, + 323, + 279, + 378 + ], + "type": "text", + "content": "[84] Jie Yang, Chaoqun Wang, Zhen Li, Junle Wang, and Ruimao Zhang. Semantic human parsing via scalable semantic transfer over multiple label domains. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19424-19433, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 41, + 379, + 279, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 379, + 279, + 422 + ], + "spans": [ + { + "bbox": [ + 41, + 379, + 279, + 422 + ], + "type": "text", + "content": "[85] Jie Yang, Ailing Zeng, Shilong Liu, Feng Li, Ruimao Zhang, and Lei Zhang. Explicit box detection unifies end-to-end multi-person pose estimation. In The Eleventh International Conference on Learning Representations, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 41, + 424, + 279, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 424, + 279, + 457 + ], + "spans": [ + { + "bbox": [ + 41, + 424, + 279, + 457 + ], + "type": "text", + "content": "[86] LI Yang, Jin Huang, TIAN Feng, WANG Hong-An, and DAI Guo-Zhong. Gesture interaction in virtual reality. Virtual Reality & Intelligent Hardware, 1(1):84-112, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 41, + 458, + 278, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 458, + 278, + 512 + ], + "spans": [ + { + "bbox": [ + 41, + 458, + 278, + 512 + ], + "type": "text", + "content": "[87] Siyuan Yang, Jun Liu, Shijian Lu, Meng Hwa Er, and Alex C Kot. Skeleton cloud colorization for unsupervised 3d action representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13423-13433, 2021. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 41, + 514, + 278, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 514, + 278, + 557 + ], + "spans": [ + { + "bbox": [ + 41, + 514, + 278, + 557 + ], + "type": "text", + "content": "[88] Yuning You, Tianlong Chen, Yang Shen, and Zhangyang Wang. Graph contrastive learning automated. In International Conference on Machine Learning, pages 12121-12132. PMLR, 2021. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 41, + 559, + 278, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 559, + 278, + 602 + ], + "spans": [ + { + "bbox": [ + 41, + 559, + 278, + 602 + ], + "type": "text", + "content": "[89] Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph contrastive learning with augmentations. Advances in neural information processing systems, 33:5812-5823, 2020. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 41, + 604, + 278, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 604, + 278, + 658 + ], + "spans": [ + { + "bbox": [ + 41, + 604, + 278, + 658 + ], + "type": "text", + "content": "[90] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Temporal query networks for fine-grained video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4486-4496, 2021. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 41, + 660, + 278, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 660, + 278, + 714 + ], + "spans": [ + { + "bbox": [ + 41, + 660, + 278, + 714 + ], + "type": "text", + "content": "[91] Nenggan Zheng, Jun Wen, Risheng Liu, Liangqu Long, Jianhua Dai, and Zhefeng Gong. Unsupervised representation learning with long-term dynamics for skeleton based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 41, + 716, + 278, + 738 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 716, + 278, + 738 + ], + "spans": [ + { + "bbox": [ + 41, + 716, + 278, + 738 + ], + "type": "text", + "content": "[92] Yuan Zhi, Zhan Tong, Limin Wang, and Gangshan Wu. Mgsampler: An explainable sampling strategy for video ac" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 98, + 536, + 277 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 318, + 98, + 536, + 130 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 98, + 536, + 130 + ], + "spans": [ + { + "bbox": [ + 318, + 98, + 536, + 130 + ], + "type": "text", + "content": "tion recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1513-1522, 2021. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 132, + 536, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 132, + 536, + 175 + ], + "spans": [ + { + "bbox": [ + 299, + 132, + 536, + 175 + ], + "type": "text", + "content": "[93] Bolei Zhou, Alex Andonian, Aude Oliva, and Antonio Torralba. Temporal relational reasoning in videos. In Proceedings of the European conference on computer vision (ECCV), pages 803-818, 2018. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 177, + 536, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 177, + 536, + 219 + ], + "spans": [ + { + "bbox": [ + 299, + 177, + 536, + 219 + ], + "type": "text", + "content": "[94] Yujie Zhou, Haodong Duan, Anyi Rao, Bing Su, and Jiaqi Wang. Self-supervised action representation learning from partial spatio-temporal skeleton sequences. arXiv preprint arXiv:2302.09018, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 221, + 536, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 221, + 536, + 277 + ], + "spans": [ + { + "bbox": [ + 299, + 221, + 536, + 277 + ], + "type": "text", + "content": "[95] Wentao Zhu, Cuiling Lan, Junliang Xing, Wenjun Zeng, Yanghao Li, Li Shen, and Xiaohui Xie. Co-occurrence feature learning for skeleton based action recognition using regularized deep LSTM networks. In Proceedings of the AAAI conference on artificial intelligence, volume 30, 2016. 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "text", + "content": "5618" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation/6a47e5bb-0a8f-4500-ae70-d8f17ed82ef9_content_list.json b/2023/Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation/6a47e5bb-0a8f-4500-ae70-d8f17ed82ef9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..faac206b4fc957df5c6559625466041b767a0b9d --- /dev/null +++ b/2023/Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation/6a47e5bb-0a8f-4500-ae70-d8f17ed82ef9_content_list.json @@ -0,0 +1,1823 @@ +[ + { + "type": "text", + "text": "Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation", + "text_level": 1, + "bbox": [ + 94, + 130, + 875, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zijie $\\mathbf{W}\\mathbf{u}^{1*}$ Yaonan Wang $^{1}$ Mingtao Feng $^{2\\dagger}$ He Xie $^{1}$ Ajmal Mian $^{3}$", + "bbox": [ + 192, + 179, + 774, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Hunan University $^{2}$ Xidian University $^{3}$ The University of Western Australia", + "bbox": [ + 217, + 199, + 751, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 313, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion probabilistic models have achieved remarkable success in text guided image generation. However, generating 3D shapes is still challenging due to the lack of sufficient data containing 3D models along with their descriptions. Moreover, text based descriptions of 3D shapes are inherently ambiguous and lack details. In this paper, we propose a sketch and text guided probabilistic diffusion model for colored point cloud generation that conditions the denoising process jointly with a hand drawn sketch of the object and its textual description. We incrementally diffuse the point coordinates and color values in a joint diffusion process to reach a Gaussian distribution. Colored point cloud generation thus amounts to learning the reverse diffusion process, conditioned by the sketch and text, to iteratively recover the desired shape and color. Specifically, to learn effective sketch-text embedding, our model adaptively aggregates the joint embedding of text prompt and the sketch based on a capsule attention network. Our model uses staged diffusion to generate the shape and then assign colors to different parts conditioned on the appearance prompt while preserving precise shapes from the first stage. This gives our model the flexibility to extend to multiple tasks, such as appearance re-editing and part segmentation. Experimental results demonstrate that our model outperforms recent state-of-the-art in point cloud generation.", + "bbox": [ + 75, + 268, + 473, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 652, + 209, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Denoising Diffusion Probabilistic Models (DDPM) [19, 45] can generate novel images and videos, from their text descriptions. Recent diffusion models [38, 33, 39, 42] contain billions of parameters and are trained on millions of text-image [13] pairs sourced from the Internet, including existing datasets like MS-COCO [25]. The availability of such large corpora of text-image pair datasets is, arguably, the major driving force behind the success of research in image and video generation. However, such datasets of text-shape pairs are almost nonexistent causing a major bottleneck for advancing research in this direction.", + "bbox": [ + 75, + 669, + 470, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D shape generation has a wide range of applications including data augmentation, virtual/augmented reality [12],", + "bbox": [ + 76, + 835, + 468, + 868 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8ace3db44bcce4d77d5dca71fe737efd5c35089b91d5ed53ee903096b54be033.jpg", + "image_caption": [ + "Figure 1. Illustration of sketch and text guided 3D point cloud generation. The text and sketch complement each other." + ], + "image_footnote": [], + "bbox": [ + 500, + 250, + 895, + 361 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "manufacturing, and reverse engineering. Pioneering works in diffusion based 3D shape generation include Luo et al. [26] and Point Voxel Diffusion (PVD) [66]. Luo et al. [26] extend the DDPMs to 3D shape generation conditioned on shape latent that follows a prior distribution parameterized via normalized flows [9]. Their probabilistic generative model treats points as particles that, under heat, are diffused to a noise distribution. It then exploits the reverse diffusion process to learn the point distribution to be able to generate plausible point clouds. PVD [66] combines the denoising diffusion model with the point-voxel 3D shape representation. It performs the denoising steps by optimizing a variational lower bound to the conditional likelihood function to produce 3D shapes or multiple 3D completions [14] from a partial observation.", + "bbox": [ + 496, + 387, + 892, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The problem of text guided 3D model or point cloud generation is slightly different from that of images in the sense that there is a higher emphasis, at least implicitly, on geometry rather than appearance which puts yet another constraint on the training datasets. Current methods [26, 66] have mostly used the ShapeNet [3] dataset for text-shape learning combined with the ModelNet [55] datasets for 3D representation learning and validation. However, these datasets are relatively small and lack detailed textual descriptions. To work around the data paucity, CLIP-forge [42] extends the CLIP model [37] for 3D shape generation by aligning 3D shapes with text in the latent space. Although these text guided methods maintain diversity in the generated shapes, their common limitation is that the generated shapes are often not the desired ones, which is understandable given the shape ambiguity in the text.", + "bbox": [ + 496, + 613, + 893, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, to reduce the shape ambiguity, we propose a sketch and text guided probabilistic diffusion (STPD) model for colored point cloud generation that conditions the", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work performed during visit at the University of Western Australia", + "bbox": [ + 94, + 875, + 457, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 217, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8929", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "denoising process with sketch inputs, in addition to textual descriptions. Sketches can be hand drawn and give much more geometric details compared to text. We argue that, combined with text, sketches are a viable option for conditioning the denoising of 3D shapes given their direct relevance to the problem at hand and their history of successful use for 3D object reconstruction [65]. However, cross-modality inputs (sketch and text) [11] always suffer from large cross-modality discrepancies and intra-modality variations bringing in additional challenges of cross-modality fusion [28] and data sparsity. Especially, sketches are inherently sparse with only a very small proportion of informative pixels, and contain much less information than natural images.", + "bbox": [ + 75, + 90, + 472, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed STPD model can produce the geometry and appearance simultaneously and unambiguously from a sketch-text input, combining ease of access and completeness of description. It takes the sketch as the main shape description and shape information present in the text to complement the sketch input. Hence, shape information is extracted from the sketch as well as the text when there is shape information in the text. Appearance is extracted from the text input to give color to the generated point clouds as shown in Fig. 1. To summarize, our contributions are:", + "bbox": [ + 75, + 301, + 472, + 454 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose a staged probabilistic diffusion model for better control over the geometry and appearance to generate colored point clouds. Our model conditions the denoising process jointly with sketch and text inputs to generate 3D shapes. It can also be used for part-segmentation and appearance re-editing.", + "2. We propose an attention capsule based feature extraction module for encoding sketches, which are inherently sparse. Our method robustly gives more attention to the entities of the useful pixels and ignores the large number of meaningless capsules corresponding to the blank pixels in the sketch.", + "3. We present a sketch-text fusion network that efficiently abstracts the shape and color information from both sketch and language descriptions to guide the reverse 3D shape diffusion process." + ], + "bbox": [ + 88, + 457, + 468, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments on the ShapeNet dataset [3] and comparison to existing diffusion probabilistic model based generation as well as some classical shape reconstruction methods show that our model achieves state-of-the-art performance for colored point cloud generation. We also show the representation learning ability of our model by conducting 3D object classification experiments on the ModelNet40 [55] dataset and the application of our model to part segmentation on the ShapeNet-Parts dataset [3].", + "bbox": [ + 75, + 707, + 468, + 843 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 852, + 218, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We first survey diffusion models for generating images. Next, we discuss 3D shape reconstruction methods fol", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lowed by classical 3D shape generation methods. Finally, we discuss diffusion models for 3D shape generation.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-based 2D Image Generation", + "text_level": 1, + "bbox": [ + 500, + 128, + 790, + 143 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative Adversarial Network (GAN): Early methods for text based image generation were based on GANs. Xu et al. [59] used attention mechanism for fine-grained text to image generation. TediGAN [56] uses GAN inversion based on the pretrained StyleGAN [1] for text to image generation. Zhang et al.[63][64] used multiple GAN architectures that stacked generators and discriminators to progressively generate multi-scale images. GANs are known to be difficult to train and have recently been surpassed by diffusion models [8].", + "bbox": [ + 496, + 148, + 890, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Probabilistic Diffusion Models: Sohl-Dickstein et al. [45] proposed diffusion probabilistic models that are both tractable and flexible. Inspired by non-equilibrium thermodynamics, the diffusion process uses Markov chain [6] to convert a simple Gaussian distribution to a target data distribution. Keeping the steps of the chain small enough, each step can be analytically evaluated [16], and hence, the full chain can also be evaluated. At each step of the forward pass, a small perturbation is added to the data until it reaches the Gaussian distribution. Since the perturbation at each step is small, a model can be learned to estimate and remove the perturbation. It was not until 2020 that Ho et al. [19] showed the real power of diffusion models with high quality image generation by training on a weighted variational bound and denoising score matching with Langevin dynamics. Dhariwal and Nichol [8] showed the superiority of diffusion models over GANs. However, directly optimizing in pixel space comes with a high computational cost. To reduce the computational requirements, [39] performs diffusion in the latent feature representation space. Large models such as Imagen [41] and DALL-E2 [38], which uses the CLIP model [37], can generate diverse photo-realistic images. The success of diffusion models is largely attributed to fundamental research as well as the availability of huge datasets in the image domain.", + "bbox": [ + 496, + 299, + 892, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2.3D Reconstruction from Images", + "text_level": 1, + "bbox": [ + 500, + 684, + 781, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reconstructing 3D shapes from images has been studied for decades. Single-view reconstruction techniques are popular, however, recovering 3D shapes from a single image is an ill-posed problem, as it requires a strong prior shape knowledge. Fan et al. [10] proposed to directly regress point clouds from a single image. Wang et al. [51] deformed the ellipsoid mesh to form the 3D shape to learn the shape prior. Sketch2Mesh [17] employed the silhouettes matching match between a given single sketch and the learned model to refine the 3D shape. TMNet [32] uses triangular meshes and images to supervise the reconstruction of 3D shapes.", + "bbox": [ + 496, + 704, + 890, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multiview images contain more geometric information useful for 3D reconstruction. $\\mathrm{Pix2Vox + + [58]}$ fuse the mul", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8930", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tiple coarse 3D volumes, which are generated from each individual view, to form the well-refined 3D shapes. Wang et al. [49] include transformers to recover 3D information from multiviews. Reconstruction based methods usually demand exact image-to-shape match and strong prior shape knowledge, whereas generative methods offer a balance between diversity and alignment with user specifications.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Sketch, Text and 3D Shape Generation", + "text_level": 1, + "bbox": [ + 76, + 203, + 410, + 219 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sketch to 3D Shape: Sketches have been used as a sparse representation of natural images and 3D shapes. Sketches are quite illustrative, despite their abstract nature and simplicity. Sketch based 3D generation [31] can be coarsely categorized into evocative [50] and constructive [29] methods, which depend on retrieval from model collections or free-form modeling. Deep models learn the pattern of the sketch for 3D model generation. For example, Zhang et al. [65] proposed a view-aware technique to gain priors from existing 3D models using a single freehand sketch. DeepSketchHair [44] proposed a learning based system for intuitive 3D styling from 2D sketches. These methods have less dependence on a shape prior to the generation phase, but the tremendous requirement of shape-related data during training for pattern learning.", + "bbox": [ + 75, + 223, + 468, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text to 3D Shape: Text to shape generation is a relatively new direction. Examples include Text2shape [4] and CLIP-Forge [42]. Text2shape [4] generates colored 3D shapes using joint embedding of text and 3D shapes. CLIP-Forge [42] generates 3D shapes from the pre-trained text-to-image CLIP [37], addressing the lack of text-to-shape priors data. Although previous studies have considered sketch and text as flexible methods for 3D shape generation, they only refine the model in an implicit pattern that takes the shape priors library into optimization while ignoring the intrinsic information mining at the sketch and text level.", + "bbox": [ + 75, + 450, + 468, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Diffusion Models for 3D Shape Generation", + "text_level": 1, + "bbox": [ + 76, + 623, + 441, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent methods [26, 66, 62, 33] generate 3D shapes with denoising diffusion probabilistic models. DPM [26] introduced the flow-based model for point cloud embedding that conditions the diffusion model on geometry representation. PVD [66] trains a diffusion model on the point-voxel representation for 3D shapes. It performs the variational lower bound optimizing on unconditional shape generation and conditioned shape completion to produce complete 3D shapes. These methods show great promise in extending the diffusion theory to 3D shape generation and achieving good performance. However, their generations are relatively uncontrolled and lack certainty for our generation problem which needs to follow alignment with user instructions more closely. LION [62] proposed a text-based 3D generation model by combining the CLIP-forge to align the text-to-3D embedding. DreamFusion [33] is a generative method for text based 3D rendering, which replaces CLIP", + "bbox": [ + 75, + 643, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with a loss derived from the distillation of 2D diffusion models. Point-E [30] starts from a language prompt to generate the single view image with fine-tuned GLIDE model and then uses the generated image for image-to-shape diffusion. Nonetheless, these text-based 3D diffusion models rely on feature pre-alignment from text to image and usually fail when there are huge domain gaps. Our method focuses on information mining for 3D diffusion by exploring the 3D structure from sketch and text to facilitate efficient controlled generation of high quality 3D shapes.", + "bbox": [ + 496, + 90, + 890, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Approach", + "text_level": 1, + "bbox": [ + 498, + 247, + 689, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a sketch $\\mathbb{S}$ and text description $\\mathbb{T}$ of an object, the goal is to generate its 3D geometry and appearance. To achieve this, the proposed STPD model consists of three modules: sparse feature extraction network for sketch embedding (Section 3.1), sketch-text feature fusion (Section 3.2), and staged diffusion for shape and color generation (Section 3.3). Figure 2 shows an overview of our model.", + "bbox": [ + 496, + 265, + 890, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Sparse Feature Extraction of a Sketch", + "text_level": 1, + "bbox": [ + 498, + 375, + 833, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature extraction from a sketch faces three major challenges. Firstly, sketches are sparse and consist of very few informative pixels. Secondly, visual models are generally trained on large amounts of natural images and their attention mechanisms exploit the information in dense pixels to focus on the more relevant parts. This becomes challenging when the image (i.e. sketch) contains copious amounts of futile background pixels and the training dataset is also small. Finally, the relationship between different object parts is ambiguous due to self-occlusions in a single sketch. Below, as shown in Fig. 3, we present our capsule network using dynamic routing [40] with attention to effectively learning the sparse features from a sketch image.", + "bbox": [ + 496, + 392, + 890, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Primary Capsule Layer: Our sketch extraction starts from a pixel-wise CNN embedding $S_{N,D_{input},W,H} = f_{cnn}(\\mathbb{S})$ , where $N$ is the core number of caps, $D$ is the dimension of each caps, and $W, H$ are the width, height of the sketch. Based on these embeddings, we perform a $1 \\times 1$ Conv of filter size $D$ to share an affine transformation on each capsule channel and then turn it to a primary capsule $u_{N,D,w,h}^{i,0}$ .", + "bbox": [ + 496, + 588, + 890, + 696 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nu _ {(N, D, w, h)} ^ {i, 0} = \\operatorname {C o n v} \\left(S _ {(N, D, w, h)}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 696, + 890, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $i$ denotes the index of capsules in the current layer, $w,h$ is the size of current caps layers and Conv is a simple network comprising a $1\\times 1$ convolution layer with stride 1, followed by batch normalization, ReLU, and max pooling.", + "bbox": [ + 496, + 719, + 890, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic Routing with Attention: Dynamic routing is applied to strengthen the structural relationship inference between different sketch parts. This facilitates robust feature learning from sketches while keeping the network depth in control and the number of parameters low. We add an attention module to force the capsules to focus on the useful sketch parts and ignore the background. Each $s^j$ can be acquired by the sum of previous layer capsules,", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8931", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5a0c311ac4c4a3709e16936908d2ae079366bedf45a12ecf5fbf347517ceb780.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 88, + 509, + 190 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b69df35504669f84f471324036294e38160235d1c721427f3f08b97643fbd17e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 88, + 890, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/92a7b16a08ae7dcd5baec19c08e0d3b540a7fecf2d65cf623c400209f8298e12.jpg", + "image_caption": [ + "Figure 2. Overview of STPD framework. The sketch feature is first extracted with two level attention modules for sparse embedding and then flows with multi-head attention fusion to text. With the embedded features, we proposed a staged denoising probabilistic diffusion model for both the geometry and appearance of 3D shapes." + ], + "image_footnote": [], + "bbox": [ + 78, + 191, + 509, + 316 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a7e3f0c1be04271896399af966cc41409cf0cbc173c657b0f253cb92a191c8b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 186, + 879, + 316 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4664f824102e2ab4ec668d47ff938e0ea2453127d21e816f675e51e706a79ba3.jpg", + "image_caption": [ + "Figure 3. Proposed sketch feature extractor. Pixel-wise attention and instance-wise attention are implemented by a CNN and attention routing, respectively. The two level attentions are designed to extract the sparse information from hand drawn sketches." + ], + "image_footnote": [], + "bbox": [ + 78, + 376, + 473, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {(N, D, w, h)} ^ {j} = \\sum_ {i = 1, 2, \\dots} c _ {w, h} ^ {i j} a _ {(N, D, 1, 1)} ^ {i j} u _ {(N, D, w, h)} ^ {i, j - 1}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 547, + 468, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c^{ij}$ are the dynamic routing parameters computed by softmax layers and $a^{ij}$ is the attention weight obtained by $1 \\times 1$ Conv on lower level primary capsules and hence associated with each instance of the original sketch.", + "bbox": [ + 76, + 580, + 468, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to the $l$ times dynamic routing, the capsules are squashed [40], to cluster the informative parts, and then reshaped to a one dimension vector $S_{(1,D_f)}$ (where $D_{f} = N\\times D$ ) from the last $s^j$ . This $S$ forms the sketch embedding used to condition the reverse diffusion process.", + "bbox": [ + 76, + 643, + 468, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Sketch-Text Feature Fusion Embedding", + "text_level": 1, + "bbox": [ + 76, + 728, + 424, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We encode the input language expression as a sequence of textual embeddings with BERT [7]. Applying a linear projection to the BERT embedding, we get embedding $T$ of the same dimension as the sketch embedding $S$ . Both embeddings are fused with two multi-head attention modules to guide the reverse diffusion process.", + "bbox": [ + 75, + 750, + 468, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The attention-ed embedding is formulated as a set of cascaded multi-attention modules. The first attention takes the sketch feature $S$ as query to extract the relevant intermediate semantic information $I$ from the text embedding. The", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "second attention performs text-guided self-attention for the sketch feature to produce the conditional embedding that contains geometry/color information based on $I$ . Architectures for geometry and color diffusion are the same but have different parameters. The first multi-head attention (MH Attention 1 in Fig. 2) is defined as:", + "bbox": [ + 496, + 378, + 892, + 470 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {(1, D _ {f})} = \\operatorname {A t t e n} _ {1} \\left(Q _ {1} = S, K _ {1} = V _ {1} = T\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 477, + 890, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q_{1}, K_{1}, V_{1}$ are the Query, Key, Value and $I$ is an intermediate attention-ed feature passed to the second multi-head attention module (MH Attention 2) defined as:", + "bbox": [ + 496, + 500, + 890, + 546 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {p \\left(1, D _ {f}\\right)} = A t t e n _ {2} \\left(Q _ {2} = K _ {2} = I + S, V _ {2} = S\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 550, + 890, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q_{2}, K_{2}, V_{2}$ are the Query, Key, Value and $p = g$ for geometric diffusion and $p = a$ for appearance diffusion. Specifically, we take the sum of intermediate features $I$ with the sketch feature $S$ as the input of the query and key in the second multi-head attention. Then, the text prompt is introduced as the vector input to support the appearance denoising using the color information from the attention module. Note that, $C_{g}$ and $C_{a}$ are from the same cascaded attentions but focus on different descriptions (geometry or appearance) in each diffusion process. Appearance condition also requires information from sketch (shape) because different object parts can have different colors.", + "bbox": [ + 496, + 571, + 892, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Staged Colored Point Cloud Generation", + "text_level": 1, + "bbox": [ + 498, + 757, + 849, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Markov's chain: Given a sample $\\mathbf{x}_0 \\sim q(\\mathbf{x}_0)$ , noise is systematically added to $\\mathbf{x}_0$ in the forward process [19] step-by-step following the Markov's chain assumption until the Gaussian distribution is reached:", + "bbox": [ + 496, + 773, + 890, + 833 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {0: T}\\right) = q \\left(\\mathbf {x} _ {0}\\right) \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 835, + 890, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 878, + 836, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8932", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{N}(\\mu, \\sigma^2)$ denotes a Gaussian distribution. Here, $\\mathbf{x}_0$ can represent either coordinates or colors from 3D shapes. We train geometry and appearance diffusion separately to generate colored point clouds. The reverse process, $p_\\theta(\\mathbf{x}_t \\mid \\mathbf{x}_{t+1})$ , starts from a standard Gaussian prior and ends with the desired $\\mathbf{x}_0$ :", + "bbox": [ + 75, + 94, + 468, + 184 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {x} _ {0: T}\\right) = p \\left(\\mathbf {x} _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 181, + 468, + 219 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) = \\mathcal {N} \\left(\\mu_ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 223, + 398, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This cross-entropy minimization is designed such that it forces the reverse process to follow the pattern of noise addition of the forward process, and eventually, recover $\\mathbf{x}_0$ from the standard Gaussian noise distribution. Hence, the loss function can be derived as:", + "bbox": [ + 75, + 242, + 468, + 315 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t} = \\mathbb {E} _ {\\mathbf {x} _ {0}, e} [ C \\| \\tilde {\\mu} _ {t} (\\mathbf {x} _ {t}, \\mathbf {x} _ {0}) - \\mu_ {t} (\\mathbf {x} _ {t}, t) \\| ^ {2} ], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 316, + 468, + 334 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C$ is a constant.", + "bbox": [ + 76, + 335, + 227, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Geometry Diffusion: Denoting the colored point cloud as $\\mathbf{x}_0\\in \\mathbb{R}^{N\\times 6} = \\{\\mathbf{g}_0\\in \\mathbb{R}^{N\\times 3},\\mathbf{a}_0\\in \\mathbb{R}^{N\\times 3}\\}$ , consisting $N$ points with geometry $\\mathbf{g}_0$ and appearance $\\mathbf{a}_0$ coordinates. We first perform geometry diffusion to generate the 3D shape. STPD is trained by minimizing the cross-entropy loss between two diffusion chains with respect to the geometric conditional feature $C_g$ , and learning of model parameters $\\theta$ :", + "bbox": [ + 75, + 354, + 468, + 474 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {g _ {0} \\sim q (g _ {0}), g _ {1: T} \\sim q (g _ {1: T})} \\left[ \\sum_ {t = 1} ^ {T} \\log p _ {\\theta} \\left(\\mathbf {g} _ {t - 1} \\mid \\mathbf {g} _ {t}, C _ {g}\\right) \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 472, + 468, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C_g$ is derived from sketch-text embedding when the text also contains shape description, or only from sketch when there is no text. We add a switch to adjust the feature map selection for different scenarios. The joint posterior $q(g_{1:T})$ can be factorized by the assumption of Markov's chain $\\prod_{t=1}^{T} q(\\mathbf{g}_{t-1}|\\mathbf{g}_t,\\mathbf{g}_0)$ , where each is analytically tractable by a re-parameterized Gaussian distribution:", + "bbox": [ + 75, + 510, + 468, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {\\mu}} _ {t} = \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1})}{1 - \\bar {\\alpha} _ {t}} \\mathbf {x} _ {t} + \\frac {\\sqrt {\\bar {\\alpha} _ {t - 1}} \\beta_ {t}}{1 - \\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 626, + 468, + 657 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {g} _ {t - 1} \\mid \\mathbf {g} _ {t}, \\mathbf {g} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {g} _ {t - 1}; \\tilde {\\boldsymbol {\\mu}} _ {t}, \\tilde {\\boldsymbol {\\beta}} _ {t} \\mathbf {I}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 662, + 468, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{\\beta}_t$ is the known variance defined in the forward process, $\\alpha_{t} = 1 - \\beta_{t}$ , and $\\tilde{\\alpha} = \\prod_{s = 1}^{t}\\alpha_{s}$ . From Eq. 7 and this property, our diffusion model learns to match each $q\\left(\\mathbf{g}_{t - 1}\\mid \\mathbf{g}_t,\\mathbf{g}_0\\right)$ and $p_{\\theta}\\left(\\mathbf{g}_{t - 1}\\mid \\mathbf{g}_t\\right)$ by estimating the output noise $\\epsilon_{\\theta}(\\mathbf{g}_t,C_g,t)$ to match noise $\\epsilon$ :", + "bbox": [ + 75, + 694, + 468, + 770 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\mathbf {g} _ {t}, C _ {g}, t\\right) \\right\\| ^ {2}, \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 772, + 468, + 791 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The noise prediction network $\\epsilon_{\\theta}$ progressively generates the 3D shape using the reverse chain", + "bbox": [ + 75, + 797, + 468, + 827 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {g} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (g _ {t}, C _ {g}, t)\\right) + \\sqrt {\\beta_ {t}} \\epsilon .\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 829, + 449, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Appearance Diffusion: Appearance diffusion is performed while freezing the geometry parameters. The optimization", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "function, in this case, can be simply modified from Eq. 8 to learn a conditional generative model for appearance:", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {a _ {0} \\sim q (a _ {0}), a _ {1: T} \\sim q (a _ {1: T})} [ \\sum_ {t = 1} ^ {T} \\log p _ {\\theta} (\\mathbf {a} _ {t - 1} | \\mathbf {a} _ {t}, p _ {\\theta} (\\mathbf {g} _ {0}), C _ {a}) ]\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 121, + 895, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training loss for appearance diffusion is minimized and conditioned on appearance features $C_a$ and $p_{\\theta}(\\mathbf{g}_0)$ :", + "bbox": [ + 498, + 162, + 890, + 191 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\mathbf {a} _ {t}, p _ {\\theta} \\left(\\mathbf {g} _ {0}\\right), C _ {a}, t\\right) \\right\\| ^ {2}, \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 195, + 890, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Colored Point Cloud Generation: We can now formally define our joint generative model for colored point cloud generation as", + "bbox": [ + 498, + 217, + 890, + 262 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta_ {2}, \\theta_ {1}, \\phi} (\\mathbf {a} _ {0}, \\mathbf {g} _ {0}, C _ {g}, C _ {a}) = p _ {\\theta_ {2}} (\\mathbf {a} _ {0} | \\mathbf {g} _ {0}, C _ {a}) p _ {\\theta_ {1}} (\\mathbf {g} _ {0} | C _ {g}) p _ {\\phi} (C _ {g}, C _ {a}),\n$$\n", + "text_format": "latex", + "bbox": [ + 498, + 268, + 905, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\{\\mathbf{a}_0,\\mathbf{g}_0\\}$ denote the colored point cloud $\\mathbf{x}_0$ . $p_{\\phi}(C_g,C_a)$ refers to the sketch and textual embedding, and $p_{\\theta_2}(\\mathbf{a}_0),p_{\\phi}(\\mathbf{g}_0)$ are the appearance and geometry denoising processes respectively. Note that the appearance diffusion is conditioned on the geometry diffusion as well. We also tried extended diffusion to denoise from a 6D Gaussian rather than treating the shape and appearance separately. However, integrating the unrelated features causes undesirable interference in the diffusion process and consequently, STPD does not learn to produce point clouds with fine details.", + "bbox": [ + 498, + 292, + 890, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 450, + 632, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 472, + 689, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets: We use the chair, table, aeroplane, car categories of ShapeNet dataset [3], randomly split it into training-test sets at $80 - 20\\%$ ratio and render a single-image from each 3D object. We apply Canny edge detection [24] on the rendered ShapeNet images to acquire their single-channel edge images. We manually remove the background so that they resemble clean hand drawn sketches. For colored point clouds, we use the chair, table categories as only they have text descriptions [4]. To generate ground-truth point clouds, each 3D shape is randomly sampled at 2048 points and then translated to zero mean and normalized in the range $[-1,1]^3$ so that the Gaussian distribution assumption is satisfied during the forward pass.", + "bbox": [ + 498, + 492, + 890, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics: Previous works [26] used Chamfer Distance (CD) and Earth Mover's Distance (EMD) to evaluate the reconstruction quality of point clouds. Based on these, existing works report Coverage (COV) and Minimum matching distance (MMD). Although these metrics have their limitations, they still serve as a good reference to measure and compare the performance of generative methods. Specifically, COV evaluates the certainty of conditioned generation whereas MMD calculates the average distance of the point clouds in the reference set and their closest neighbors in the generated set. Another metric used in numerous works to evaluate the generation quality and diversity is 1-nearest neighbor accuracy (1-NNA) which directly quantifies distribution similarity between two sets. However, since", + "bbox": [ + 498, + 689, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8933", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b71d2e6c5f1cecb2f13fe6c0778c7a93efc94cc1e575c5923eb0b31dcc6936a1.jpg", + "image_caption": [ + "Figure 4. Visual comparison of the generated objects. 1) Inputs include text and free-hand sketch. 2) Shape methods with no color. 3) Colored shape generation methods. 4) Our method generates both geometry and appearance using sketch and text. The appearance can be re-edited by changing the language prompt while preserving geometry." + ], + "image_footnote": [], + "bbox": [ + 78, + 85, + 890, + 205 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7da6e7cd62bbf2213604f0c6a59540222a14d4717fd11a338d02040a999905fd.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Comparison of 3D shape generation for the chair, table categories of ShapeNet [3]. CD is multiplied by $10^{3}$ and EMD by $10^{1}$ . Our method generates the best shape, with or without color." + ], + "table_body": "
CategoriesInputMethodsMMD (↓) CD EMDCOV (%,↑) CD EMD
Reconstruction Shape (N×3)SketchPSGN [10]8.1747.253
Pix2Mesh [51]9.7717.347
TMNet [32]9.188.11
Sketch2Mesh [17]3.0182.786
SketchesPix2Vox++ [57]5.9765.760
EvoIT [49]5.7185.307
Generative Shape (N×3)TextText2Shape [5]9.8736.348
CLIP-forge [43]4.3673.818
SketchMixNF [34]7.6896.224
DPF-Net [22]7.5316.176
Sketch2Model [65]3.1762.967
×DPM [26]3.7641.986
Sketch,textSTPD (ours)1.6351.043
Generative Shape&Color (N×6)TextText2Shape10.347.476
Point-E (text) [30]2.5232.265
×DPM-extended4.7653.256
sketchPoint-E (sketch)3.0923.117
Sketch,textSTPD (ours)1.3461.017
", + "bbox": [ + 78, + 253, + 472, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "our aim is to generate 3D shapes faithful to the provided sketch, we stick to the first four metrics. Where other methods do not provide color, we compare shapes only.", + "bbox": [ + 75, + 602, + 468, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines: We compare with classical as well as diffusion based methods. In the former category, we select the major sketch [65], image [34] [22], and text [43] [5] based 3D object generation methods. We also compare to reconstruction based methods including single-view [10, 51, 32, 17] and multiviews [58, 49]. In the latter category, we select 3D diffusion (DPM) [26], Point-E [30], and Point Voxel Diffusion (PVD) [66]. Point-E (sketch) results are obtained after fine-tuning the Point-E image-to-shape diffusion model on our sketch-shape data for a fair comparison. We take ten generations for all methods and use the average error from ground truth for evaluation.", + "bbox": [ + 75, + 648, + 470, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 76, + 837, + 449, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 compares the point cloud generation quality of our method with existing state-of-the-art. For a fair comparison, we re-train all methods with our sketch-shape data of", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the chair, table categories, or their text descriptions when the method takes text only as input. Multiview methods are given sketches from 3 viewpoints. Only our method performs (single view) sketch and text guided 3D object generation. Since DPM [26] is unconditional, we generate many shapes and pick the best matches. For shape only (upper half of Table 1), the proposed STPD outperforms all methods, achieving $45.8\\%$ lower CD-MMD error from the nearest reconstruction method Sketch2Mesh [17] and $48.5\\%$ lower CD-MMD error from the nearest generative method Sketch2Model [65]. The multiview methods do not perform well since they are unable to establish correspondences between the sketches. Text guided generative methods have more diversity but struggle with alignment to user specifications given the inherent shape ambiguity in the text. Sketch guided methods perform better in general.", + "bbox": [ + 496, + 256, + 892, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 also compares colored point cloud (shape and color) generation. For this, we extend DPM to a 6D Gaussian and also compare with Text2Shape [4] and Point-E [30]. The proposed STPD again outperforms all methods, achieving $46.7\\%$ CD-MMD error from the nearest competitor Point-E [30]. In Fig. 4, we compare our results visually to the existing methods using hand-drawn sketches from SketchX [23]. Our method also offers to reedit the color (appearance) once the geometry is finalized (last column). Table 2 reports comparative results on the chair, table, aeroplane, car categories of ShapeNet. We choose these categories because PVD [66] provides comparisons on the chair, aeroplane, car categories and the table category is used in our previous experiment. From Table 2, we can see that STPD achieves the best performance on both evaluation metrics on all four object categories. SoftFlow [21], PointFlow [60], DPF-Net [22], DPM [26], Point-E [30], PVD [66] achieve an average Chamfer Distance of 2.464, 2.429, 2.168, 2.509, 2.182, 2.539 and an average EMD of 3.105, 3.191, 2.851, 2.553, 2.490, 2.661 respectively. Our STPD achieves 1.172, 1.261 average CD, EMD errors which are $46.3\\%$ , $49.3\\%$ lower than the nearest competitor Point-E [30].", + "bbox": [ + 496, + 503, + 892, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Human Evaluations: We also conducted human evaluations on 100 objects generated by our method and other methods. Table 3 shows the results where 30 people rate", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8934", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f21cab7196327195786279a418842650dd1f5a1469b549f052b782a74bd74671.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodschairtableaeroplanecar
CDEMDCDEMDCDEMDCDEMD
SoftFlow [21]2.7863.2954.8155.1390.4041.1981.8502.789
PointFlow [60]2.7073.6494.8045.0820.4031.1801.8032.851
DPF-Net [22]2.7633.3203.9864.6610.5281.1051.3962.318
DPM [26]3.2012.7184.6014.2980.4321.0271.8042.167
Point-E [30]2.9012.6673.8094.2900.4121.0051.6091.997
PVD [66]3.2112.9394.7314.5270.4421.0301.7742.146
STPD (ours)1.2091.1172.1171.8970.3110.8971.0491.136
", + "bbox": [ + 80, + 89, + 467, + 215 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cdc468c1889d3205187df328c9720a9f4dff01994c53f4dc955564b146dc4976.jpg", + "table_caption": [ + "Table 2. Comparison with baselines on 4 categories of the ShapeNet dataset [3]. CD is multiplied by $10^{3}$ and EMD by $10^{1}$ ." + ], + "table_footnote": [], + "table_body": "
Human rate (↑)SoftFlowPointFlowDPF-NetDPMPoint-EPVDSTPD
chair1.81.71.92.13.93.54.8
table2.11.93.53.03.42.54.6
aeroplane3.43.32.23.13.73.04.0
car2.52.43.43.53.22.94.3
", + "bbox": [ + 80, + 253, + 467, + 319 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/94d0f38edc4a9e651e530a6325687daa049674dea793e778c2080ac43bee9c14.jpg", + "table_caption": [ + "Table 3. Human evaluations: 100 generated objects rated on a scale of 1 to 5 by 30 people." + ], + "table_footnote": [], + "table_body": "
Input Modality/Model for ConditioningCD (×103)EMD (×101)
SketchText
×BERT28.32520.457
C1×1×10.9658.658
C1×1BERT9.6518.459
ResNet50BERT6.4095.480
ResNet50 + Atten.RoutingBERT1.9351.587
C1×1 + Atten.Routing×1.6891.348
C1×1 + Atten.Routing (concatenate)BERT3.7484.127
C1×1 + Atten.Routing (our fusion)BERT1.3901.101
", + "bbox": [ + 93, + 354, + 454, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the generations on a scale of 1 to 5 (higher is better). STPD outperforms the baseline methods on all individual object categories and achieves an average rating of 4.4, compared to average rating of 3.3 by the nearest competitor Point-E.", + "bbox": [ + 75, + 537, + 467, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 608, + 228, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform ablation study on the sketch and text modules to verify their contributions. Table 4 shows our results using CD and EMD metrics. Text conditioning, using BERT embeddings, alone achieves the worst performance. $C_{1 \\times 1}$ is a simple visual encoder comprising $1 \\times 1$ convolution followed by batch normalization and ReLU. Sketch conditioning, using simple $C_{1 \\times 1}$ embeddings, reduces the errors significantly since a sketch can describe shaping much better than text. Using sketch and text conditioning with $C_{1 \\times 1}$ with BERT embeddings, the errors do not reduce much. However, when a more sophisticated sketch encoding is performed with ResNet50 [18] and combined with BERT, the CD and EMD reduce notably. To account for the sparsity of sketches, when we add capsule network with attention routing on top of ResNet50, we see a dramatic improvement in both metrics. Our simple visual encoder $C_{1 \\times 1}$ with attention routing further improves the performance even without using text conditioning. As seen in", + "bbox": [ + 75, + 628, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d13430e58c9ff9b7355ee005b97a8fc89f00642c8f4a138d0368a0120c60165e.jpg", + "table_caption": [ + "Table 4. Ablation study on visual & textual modules. $C_{1 \\times 1}$ is a simple visual encoder comprising $1 \\times 1$ convolution followed by batch normalization and ReLU." + ], + "table_footnote": [], + "table_body": "
ModelModelNet10ModelNet40ScanObjectNN
PointFlow [60]93.786.878.2
PDGN [20]94.287.378.1
3D-GAN [54]91.083.376.7
AtlasNet [15]91.986.676.9
ShapeGF [2]90.284.675.8
DPM [26]94.287.679.7
STPD94.888.184.3
", + "bbox": [ + 509, + 89, + 880, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Object classification accuracy $(\\%)$", + "bbox": [ + 563, + 205, + 826, + 217 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cda21b06b74008462a46dcc22296ce71e3f4576bcad40c9b17f82d414e0f1daf.jpg", + "image_caption": [ + "Figure 5. Visualization of attention score maps for a sketch with respect to capsules. The $8 \\times 16$ size denotes 8 capsules of 16-dimension each. Capsule attention maps are learned from attention-routing. Dark parts denote high attention." + ], + "image_footnote": [], + "bbox": [ + 501, + 227, + 898, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the last row of the table, the best performance is achieved when we use sketch conditioning with $C_{1 \\times 1} +$ attention routing embeddings combined with text embeddings from BERT [7] using the proposed fusion method. The second last row of the table shows that a simple concatenation of the two embeddings does not work as well as our proposed fusion mechanism.", + "bbox": [ + 498, + 412, + 890, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Representation Learning Evaluation", + "text_level": 1, + "bbox": [ + 500, + 526, + 815, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the representation learned by our sketch embedding used to condition the diffusion model. We follow the work of MVCNN [46], ScanObjectNN [48] and convert rendered images from different viewpoints of the ModelNet dataset [55] to sketches and conduct object classification experiments. We train our STPD on the generated sketches and then take the conditioned embedded features to learn a linear SVM classifier for classification based on a single sketch embedding. The essence of the conditional diffusion model is that the conditioning should guide the diffusion to retrieve the 3D shape from Gaussian noise. This means that our conditional network should be able to represent specific object categories to aid classification. The results in Table 5 demonstrate that our architecture can extract discriminative features, that are also useful for accurate classification.", + "bbox": [ + 496, + 544, + 890, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Effectiveness of Capsule Network", + "text_level": 1, + "bbox": [ + 500, + 777, + 792, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 5 visualizes the generated attention scores from a typical sketch. More examples are given in the supplementary material along with quantitative results using Instance Scores for Sketch (ISS i.e. without attention routing) and Instance Score on Capsule (ISC i.e. with attention routing) metrics. ISS is the percentage of useful pixels divided by the total pixels of the sketch. For ISC, we compute the aver", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8935", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c73ff89596e7aed9bf3f5347d276b016b30dee73976a0ca6ff744e471db1410.jpg", + "image_caption": [ + "Figure 6. Visualization of part segmentation results for 4 categories of ShapeNet-Parts." + ], + "image_footnote": [], + "bbox": [ + 78, + 85, + 864, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "age of the highest attention score for each of the 16 capsule dimensions (column wise) and convert it to a percentage.", + "bbox": [ + 75, + 279, + 468, + 309 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Before the attention capsule, the sketches have ISS of $1.07\\%$ on average. After our capsule-attention module, ISC becomes $58.68\\%$ on average for better diffusion. This demonstrates the effectiveness of the Capsule Attention Network which attends to sparse data in sketches during diffusion training. Each capsule $(1\\times 16)$ denotes a learned feature, e.g. the table, from a sketch. It can be observed that given the distribution of the useful pixel in the sketch, only a few attention score values are dark i.e., attentioned. Caps number 2 and a few dimensions of Caps number 5 are given high weight to contribute to the final visual feature, while the other Caps that represent the white background of the sketch are ignored. The attention-based capsule network enables our method to effectively locate the referred objects in the sparse sketch input.", + "bbox": [ + 75, + 310, + 473, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Extending STPD to Part Segmentation Task", + "text_level": 1, + "bbox": [ + 76, + 547, + 468, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further show the feature learning ability of our STPD model, we use it for the task of part segmentation in point clouds. For this, we use the ShapeNet-Parts [3] dataset, focusing on the four object categories of Table 2.", + "bbox": [ + 75, + 574, + 468, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The part segmentation task can be extended from our colored point cloud generation model by assigning different colors to each part. We consider only two parts for the table and three parts for remaining three categories e.g. for aeroplane, the (body, wing, tail) are assigned (blue, green, red) colors respectively.", + "bbox": [ + 75, + 635, + 468, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our STPD, each coordinate is generated by our sketch-text guided geometric diffusion model. However, in the part segmentation task, the shape $\\mathbf{g}_0$ is given as input to the model. Here, we directly apply the text feature as the condition $C_a$ of appearance diffusion. For instance, we provide a generic text sentence \"An blue aeroplane with green wings and red tail\" for all aeroplanes.", + "bbox": [ + 75, + 726, + 468, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With the given input geometry and appearance description, we can formally define our joint generative model for point cloud segmentation as", + "bbox": [ + 75, + 832, + 468, + 878 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta , \\phi} \\left(\\mathbf {a} _ {0}, \\mathbf {g} _ {0}, C _ {a}\\right) = p _ {\\theta} \\left(\\mathbf {a} _ {0} \\mid \\mathbf {g} _ {0}, C _ {a}\\right) p _ {\\text {i n p u t}} \\left(\\mathbf {g} _ {0}\\right) p _ {\\phi} \\left(C _ {a}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 886, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/13fc5712042b2d701e97b398423bfe97ddd1d753096563c6146bacefa597cf36.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelmIoUcmIoU1chairtableaeroplanecar
PointNet [35]82.183.589.680.683.474.9
PointNet++ [36]83.384.690.882.682.477.3
DGCNN [52]83.884.990.682.684.077.8
Point-BERT [61]84.785.691.081.584.379.8
PointMLP [27]84.684.890.384.383.580.5
KPConv [47]85.185.991.183.684.681.1
P2P [53]85.085.991.683.784.380.4
STPD85.986.792.484.185.681.3
", + "bbox": [ + 514, + 276, + 879, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Comparative results of part segmentation on four categories of ShapeNet-Parts dataset.", + "bbox": [ + 500, + 405, + 890, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where the combination $\\{\\mathbf{a}_0, \\mathbf{g}_0\\}$ denotes the desired segmentation of the 3D point cloud. Note that, $\\mathbf{a}_0$ is generated as colors which are then used as segmentation labels. We apply K-means clustering for final refinement of the colors.", + "bbox": [ + 498, + 438, + 893, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results for 3D Shape Segmentation: Fig. 6 shows some visualizations of our results and Table 6 compares our method to some existing baselines. The proposed STPD achieves the highest mIoU compared to existing methods.", + "bbox": [ + 498, + 506, + 893, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 580, + 619, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed a novel sketch and text guided probabilistic diffusion model for 3D object generation. Our model takes a hand drawn sketch and text as inputs to condition the reverse diffusion process. With the strong generative ability and effective guidance by the sketch-text fused features, our model is able to achieve high alignment with user specifications. Our visual encoder is designed to focus on the informative sparse parts of sketches. The sketch-text fusion part gathers informative object context from visual and textual descriptions to improve the shape and appearance of the generated 3D objects. Experimental results demonstrate that our model achieves state-of-art performance in 3D object generation and shows promising applications for 3D object classification and part segmentation.", + "bbox": [ + 496, + 604, + 893, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 830, + 679, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Professor Ajmal Mian is the recipient of an Australian Research Council Future Fellowship Award (project number FT210100268) funded by the Australian Government.", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8936", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] A.H. Bermano, R. Gal, Y. Alaluf, R. Mokady, Y. Nitzan, O. Tov, O. Patashnik, and D. Cohen-Or. State-of-the-art in the architecture, methods and applications of stylegan. Computer Graphics Forum, 41(2):591-611, 2022. 2", + "[2] Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snively, and Bharath Hariharan. Learning gradient fields for shape generation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 7", + "[3] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 1, 2, 5, 6, 7, 8", + "[4] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian conference on computer vision, pages 100-116. Springer, 2018. 3, 5, 6", + "[5] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian conference on computer vision, pages 100-116. Springer, 2018. 6", + "[6] Shin-I Cheng, Yu-Jie Chen, Wei-Chen Chiu, Hung-Yu Tseng, and Hsin-Ying Lee. Adaptively-realistic image generation from stroke and sketch with diffusion model. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4054-4062, January 2023. 2", + "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 4, 7", + "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2", + "[9] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. arXiv preprint arXiv:1605.08803, 2016. 1", + "[10] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 2, 6", + "[11] Mingtao Feng, Syed Zulqarnain Gilani, Yaonan Wang, Liang Zhang, and Ajmal Mian. Relation graph network for 3d object detection in point clouds. IEEE Transactions on Image Processing, 30:92-107, 2020. 2", + "[12] Mingtao Feng, Haoran Hou, Liang Zhang, Yulan Guo, Hongshan Yu, Yaonan Wang, and Ajmal Mian. Exploring hierarchical spatial layout cues for 3d point cloud based scene graph prediction. IEEE Transactions on Multimedia, 2023. 1", + "[13] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph net-" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "work for object grounding in point cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3722-3731, 2021. 1", + "[14] Zhiheng Fu, Siyu Hong, Mengyi Liu, Hamid Laga, Mohammed Bennamoun, Farid Boussaid, and Yulan Guo. Multi-stage information diffusion for joint depth and surface normal estimation. Pattern Recognition, 141:109660, 2023. 1", + "[15] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 216-224, 2018. 7", + "[16] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10696-10706, June 2022. 2", + "[17] Benoit Guillard, Edoardo Remelli, Pierre Yvernay, and Pascal Fua. Sketch2mesh: Reconstructing and editing 3d shapes from sketches. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13023-13032, October 2021. 2, 6", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7", + "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1, 2, 4", + "[20] Le Hui, Rui Xu, Jin Xie, Jianjun Qian, and Jian Yang. Progressive point cloud deconvolution generation network. In European Conference on Computer Vision, pages 397-413. Springer, 2020. 7", + "[21] Hyeongju Kim, Hyeonseung Lee, Woo Hyun Kang, Joun Yeop Lee, and Nam Soo Kim. Softflow: Probabilistic framework for normalizing flow on manifolds. Advances in Neural Information Processing Systems, 33:16388-16397, 2020. 6, 7", + "[22] Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. In European Conference on Computer Vision, pages 694-710. Springer, 2020. 6, 7", + "[23] Yi Li, Timothy M Hospedales, Yi-Zhe Song, and Shaogang Gong. Fine-grained sketch-based image retrieval by matching deformable part models. 2014. 6", + "[24] Yibo Li and Bailun Liu. Improved edge detection algorithm for canny operator. In 2022 IEEE 10th Joint International Information Technology and Artificial Intelligence Conference (ITAIC), volume 10, pages 1-5. IEEE, 2022. 5", + "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1" + ], + "bbox": [ + 503, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8937", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2837-2845, 2021. 1, 3, 5, 6, 7", + "[27] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022. 8", + "[28] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Spectrum-guided multi-granularity referring video object segmentation. arXiv preprint arXiv:2307.13537, 2023. 2", + "[29] Yongwei Miao, Feixia Hu, Xudong Zhang, Jiazhou Chen, and Renato Pajarola. Symmsketch: Creating symmetric 3d free-form shapes from 2d sketches. Computational Visual Media, 1(1):3-16, 2015. 3", + "[30] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 3, 6, 7", + "[31] Luke Olsen, Faramarz F Samavati, Mario Costa Sousa, and Joaquim A Jorge. Sketch-based modeling: A survey. Computers & Graphics, 33(1):85-103, 2009. 3", + "[32] Junyi Pan, Xiaoguang Han, Weikai Chen, Jiapeng Tang, and Kui Jia. Deep mesh reconstruction from single rgb images via topology modification networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9964-9973, 2019. 2, 6", + "[33] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 3", + "[34] Janis Postels, Mengya Liu, Riccardo Spezialetti, Luc Van Gool, and Federico Tombari. Go with the flows: Mixtures of normalizing flows for point cloud generation and reconstruction. In 2021 International Conference on 3D Vision (3DV), pages 1249-1258. IEEE, 2021. 6", + "[35] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 8", + "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 8", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3", + "[38] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8821-8831. PMLR, 18-24 Jul 2021. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 1, 2", + "[40] Sara Sabour, Nicholas Frosst, and Geoffrey E Hinton. Dynamic routing between capsules. Advances in neural information processing systems, 30, 2017. 3, 4", + "[41] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems, volume 35, pages 36479-36494. Curran Associates, Inc., 2022. 2", + "[42] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 1, 3", + "[43] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 6", + "[44] Yuefan Shen, Changgeng Zhang, Hongbo Fu, Kun Zhou, and Youyi Zheng. Deepsketchhair: Deep sketch-based 3d hair modeling. IEEE Transactions on Visualization and Computer Graphics, 27(7):3250-3263, 2021. 3", + "[45] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 1, 2", + "[46] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In Proceedings of the IEEE international conference on computer vision, pages 945-953, 2015. 7", + "[47] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 8", + "[48] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In International Conference on Computer Vision (ICCV), 2019. 7", + "[49] Dan Wang, Xinrui Cui, Xun Chen, Zhengxia Zou, Tianyang Shi, Septimiu Salcudean, Z. Jane Wang, and Rabab Ward. Multi-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF International Conference on Com" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8938", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "puter Vision (ICCV), pages 5722-5731, October 2021. 3, 6", + "[50] Fang Wang, Le Kang, and Yi Li. Sketch-based 3d shape retrieval using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1875-1883, 2015. 3", + "[51] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 2, 6", + "[52] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog), 38(5):1-12, 2019. 8", + "[53] Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. P2p: Tuning pre-trained image models for point cloud analysis with point-to-pixel prompting. arXiv preprint arXiv:2208.02812, 2022. 8", + "[54] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 7", + "[55] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 1, 2, 7", + "[56] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2256-2265, June 2021. 2", + "[57] Haozhe Xie, Hongxun Yao, Xiaoshuai Sun, Shangchen Zhou, and Shengping Zhang. Pix2vox: Context-aware 3d reconstruction from single and multi-view images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2690-2698, 2019. 6", + "[58] Haozhe Xie, Hongxun Yao, Shengping Zhang, Shangchen Zhou, and Wenxiu Sun. Pix2vox++: Multi-scale context-aware 3d object reconstruction from single and multiple images. International Journal of Computer Vision, 128(12):2919-2935, 2020. 2, 6", + "[59] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Finegrained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 2", + "[60] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4541-4550, 2019. 6, 7", + "[61] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 8", + "[62] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022. 3", + "[63] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 2", + "[64] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan++: Realistic image synthesis with stacked generative adversarial networks. IEEE transactions on pattern analysis and machine intelligence, 41(8):1947-1962, 2018. 2", + "[65] Song-Hai Zhang, Yuan-Chen Guo, and Qing-Wen Gu. Sketch2model: View-aware 3d modeling from single free-hand sketches. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6012-6021, 2021. 2, 3, 6", + "[66] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826–5835, 2021. 1, 3, 6, 7" + ], + "bbox": [ + 503, + 92, + 892, + 472 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8939", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation/6a47e5bb-0a8f-4500-ae70-d8f17ed82ef9_model.json b/2023/Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation/6a47e5bb-0a8f-4500-ae70-d8f17ed82ef9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b3823c94ffdb4478b7d055c7542ddc0b553b36c3 --- /dev/null +++ b/2023/Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation/6a47e5bb-0a8f-4500-ae70-d8f17ed82ef9_model.json @@ -0,0 +1,2576 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.131, + 0.877, + 0.152 + ], + "angle": 0, + "content": "Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation" + }, + { + "type": "text", + "bbox": [ + 0.193, + 0.18, + 0.775, + 0.2 + ], + "angle": 0, + "content": "Zijie \\(\\mathbf{W}\\mathbf{u}^{1*}\\) Yaonan Wang\\(^{1}\\) Mingtao Feng\\(^{2\\dagger}\\) He Xie\\(^{1}\\) Ajmal Mian\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.2, + 0.753, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)Hunan University \\(^{2}\\)Xidian University \\(^{3}\\)The University of Western Australia" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.314, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.269, + 0.474, + 0.649 + ], + "angle": 0, + "content": "Diffusion probabilistic models have achieved remarkable success in text guided image generation. However, generating 3D shapes is still challenging due to the lack of sufficient data containing 3D models along with their descriptions. Moreover, text based descriptions of 3D shapes are inherently ambiguous and lack details. In this paper, we propose a sketch and text guided probabilistic diffusion model for colored point cloud generation that conditions the denoising process jointly with a hand drawn sketch of the object and its textual description. We incrementally diffuse the point coordinates and color values in a joint diffusion process to reach a Gaussian distribution. Colored point cloud generation thus amounts to learning the reverse diffusion process, conditioned by the sketch and text, to iteratively recover the desired shape and color. Specifically, to learn effective sketch-text embedding, our model adaptively aggregates the joint embedding of text prompt and the sketch based on a capsule attention network. Our model uses staged diffusion to generate the shape and then assign colors to different parts conditioned on the appearance prompt while preserving precise shapes from the first stage. This gives our model the flexibility to extend to multiple tasks, such as appearance re-editing and part segmentation. Experimental results demonstrate that our model outperforms recent state-of-the-art in point cloud generation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.653, + 0.21, + 0.668 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.471, + 0.837 + ], + "angle": 0, + "content": "Denoising Diffusion Probabilistic Models (DDPM) [19, 45] can generate novel images and videos, from their text descriptions. Recent diffusion models [38, 33, 39, 42] contain billions of parameters and are trained on millions of text-image [13] pairs sourced from the Internet, including existing datasets like MS-COCO [25]. The availability of such large corpora of text-image pair datasets is, arguably, the major driving force behind the success of research in image and video generation. However, such datasets of text-shape pairs are almost nonexistent causing a major bottleneck for advancing research in this direction." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.837, + 0.47, + 0.869 + ], + "angle": 0, + "content": "3D shape generation has a wide range of applications including data augmentation, virtual/augmented reality [12]," + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.25, + 0.897, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.36, + 0.893, + 0.387 + ], + "angle": 0, + "content": "Figure 1. Illustration of sketch and text guided 3D point cloud generation. The text and sketch complement each other." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.893, + 0.614 + ], + "angle": 0, + "content": "manufacturing, and reverse engineering. Pioneering works in diffusion based 3D shape generation include Luo et al. [26] and Point Voxel Diffusion (PVD) [66]. Luo et al. [26] extend the DDPMs to 3D shape generation conditioned on shape latent that follows a prior distribution parameterized via normalized flows [9]. Their probabilistic generative model treats points as particles that, under heat, are diffused to a noise distribution. It then exploits the reverse diffusion process to learn the point distribution to be able to generate plausible point clouds. PVD [66] combines the denoising diffusion model with the point-voxel 3D shape representation. It performs the denoising steps by optimizing a variational lower bound to the conditional likelihood function to produce 3D shapes or multiple 3D completions [14] from a partial observation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.856 + ], + "angle": 0, + "content": "The problem of text guided 3D model or point cloud generation is slightly different from that of images in the sense that there is a higher emphasis, at least implicitly, on geometry rather than appearance which puts yet another constraint on the training datasets. Current methods [26, 66] have mostly used the ShapeNet [3] dataset for text-shape learning combined with the ModelNet [55] datasets for 3D representation learning and validation. However, these datasets are relatively small and lack detailed textual descriptions. To work around the data paucity, CLIP-forge [42] extends the CLIP model [37] for 3D shape generation by aligning 3D shapes with text in the latent space. Although these text guided methods maintain diversity in the generated shapes, their common limitation is that the generated shapes are often not the desired ones, which is understandable given the shape ambiguity in the text." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this paper, to reduce the shape ambiguity, we propose a sketch and text guided probabilistic diffusion (STPD) model for colored point cloud generation that conditions the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.875, + 0.458, + 0.888 + ], + "angle": 0, + "content": "*Work performed during visit at the University of Western Australia" + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.218, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.875, + 0.458, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8929" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.302 + ], + "angle": 0, + "content": "denoising process with sketch inputs, in addition to textual descriptions. Sketches can be hand drawn and give much more geometric details compared to text. We argue that, combined with text, sketches are a viable option for conditioning the denoising of 3D shapes given their direct relevance to the problem at hand and their history of successful use for 3D object reconstruction [65]. However, cross-modality inputs (sketch and text) [11] always suffer from large cross-modality discrepancies and intra-modality variations bringing in additional challenges of cross-modality fusion [28] and data sparsity. Especially, sketches are inherently sparse with only a very small proportion of informative pixels, and contain much less information than natural images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.473, + 0.455 + ], + "angle": 0, + "content": "The proposed STPD model can produce the geometry and appearance simultaneously and unambiguously from a sketch-text input, combining ease of access and completeness of description. It takes the sketch as the main shape description and shape information present in the text to complement the sketch input. Hence, shape information is extracted from the sketch as well as the text when there is shape information in the text. Appearance is extracted from the text input to give color to the generated point clouds as shown in Fig. 1. To summarize, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.458, + 0.47, + 0.549 + ], + "angle": 0, + "content": "1. We propose a staged probabilistic diffusion model for better control over the geometry and appearance to generate colored point clouds. Our model conditions the denoising process jointly with sketch and text inputs to generate 3D shapes. It can also be used for part-segmentation and appearance re-editing." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.551, + 0.469, + 0.642 + ], + "angle": 0, + "content": "2. We propose an attention capsule based feature extraction module for encoding sketches, which are inherently sparse. Our method robustly gives more attention to the entities of the useful pixels and ignores the large number of meaningless capsules corresponding to the blank pixels in the sketch." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.643, + 0.469, + 0.703 + ], + "angle": 0, + "content": "3. We present a sketch-text fusion network that efficiently abstracts the shape and color information from both sketch and language descriptions to guide the reverse 3D shape diffusion process." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.458, + 0.47, + 0.703 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.708, + 0.47, + 0.844 + ], + "angle": 0, + "content": "Extensive experiments on the ShapeNet dataset [3] and comparison to existing diffusion probabilistic model based generation as well as some classical shape reconstruction methods show that our model achieves state-of-the-art performance for colored point cloud generation. We also show the representation learning ability of our model by conducting 3D object classification experiments on the ModelNet40 [55] dataset and the application of our model to part segmentation on the ShapeNet-Parts dataset [3]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.853, + 0.22, + 0.868 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We first survey diffusion models for generating images. Next, we discuss 3D shape reconstruction methods fol" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "lowed by classical 3D shape generation methods. Finally, we discuss diffusion models for 3D shape generation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.129, + 0.791, + 0.144 + ], + "angle": 0, + "content": "2.1. Text-based 2D Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.149, + 0.892, + 0.299 + ], + "angle": 0, + "content": "Generative Adversarial Network (GAN): Early methods for text based image generation were based on GANs. Xu et al. [59] used attention mechanism for fine-grained text to image generation. TediGAN [56] uses GAN inversion based on the pretrained StyleGAN [1] for text to image generation. Zhang et al.[63][64] used multiple GAN architectures that stacked generators and discriminators to progressively generate multi-scale images. GANs are known to be difficult to train and have recently been surpassed by diffusion models [8]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.893, + 0.679 + ], + "angle": 0, + "content": "Probabilistic Diffusion Models: Sohl-Dickstein et al. [45] proposed diffusion probabilistic models that are both tractable and flexible. Inspired by non-equilibrium thermodynamics, the diffusion process uses Markov chain [6] to convert a simple Gaussian distribution to a target data distribution. Keeping the steps of the chain small enough, each step can be analytically evaluated [16], and hence, the full chain can also be evaluated. At each step of the forward pass, a small perturbation is added to the data until it reaches the Gaussian distribution. Since the perturbation at each step is small, a model can be learned to estimate and remove the perturbation. It was not until 2020 that Ho et al. [19] showed the real power of diffusion models with high quality image generation by training on a weighted variational bound and denoising score matching with Langevin dynamics. Dhariwal and Nichol [8] showed the superiority of diffusion models over GANs. However, directly optimizing in pixel space comes with a high computational cost. To reduce the computational requirements, [39] performs diffusion in the latent feature representation space. Large models such as Imagen [41] and DALL-E2 [38], which uses the CLIP model [37], can generate diverse photo-realistic images. The success of diffusion models is largely attributed to fundamental research as well as the availability of huge datasets in the image domain." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.685, + 0.782, + 0.7 + ], + "angle": 0, + "content": "2.2.3D Reconstruction from Images" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Reconstructing 3D shapes from images has been studied for decades. Single-view reconstruction techniques are popular, however, recovering 3D shapes from a single image is an ill-posed problem, as it requires a strong prior shape knowledge. Fan et al. [10] proposed to directly regress point clouds from a single image. Wang et al. [51] deformed the ellipsoid mesh to form the 3D shape to learn the shape prior. Sketch2Mesh [17] employed the silhouettes matching match between a given single sketch and the learned model to refine the 3D shape. TMNet [32] uses triangular meshes and images to supervise the reconstruction of 3D shapes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Multiview images contain more geometric information useful for 3D reconstruction. \\(\\mathrm{Pix2Vox + + [58]}\\) fuse the mul" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8930" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "tiple coarse 3D volumes, which are generated from each individual view, to form the well-refined 3D shapes. Wang et al. [49] include transformers to recover 3D information from multiviews. Reconstruction based methods usually demand exact image-to-shape match and strong prior shape knowledge, whereas generative methods offer a balance between diversity and alignment with user specifications." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.204, + 0.411, + 0.22 + ], + "angle": 0, + "content": "2.3. Sketch, Text and 3D Shape Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.224, + 0.47, + 0.451 + ], + "angle": 0, + "content": "Sketch to 3D Shape: Sketches have been used as a sparse representation of natural images and 3D shapes. Sketches are quite illustrative, despite their abstract nature and simplicity. Sketch based 3D generation [31] can be coarsely categorized into evocative [50] and constructive [29] methods, which depend on retrieval from model collections or free-form modeling. Deep models learn the pattern of the sketch for 3D model generation. For example, Zhang et al. [65] proposed a view-aware technique to gain priors from existing 3D models using a single freehand sketch. DeepSketchHair [44] proposed a learning based system for intuitive 3D styling from 2D sketches. These methods have less dependence on a shape prior to the generation phase, but the tremendous requirement of shape-related data during training for pattern learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.47, + 0.618 + ], + "angle": 0, + "content": "Text to 3D Shape: Text to shape generation is a relatively new direction. Examples include Text2shape [4] and CLIP-Forge [42]. Text2shape [4] generates colored 3D shapes using joint embedding of text and 3D shapes. CLIP-Forge [42] generates 3D shapes from the pre-trained text-to-image CLIP [37], addressing the lack of text-to-shape priors data. Although previous studies have considered sketch and text as flexible methods for 3D shape generation, they only refine the model in an implicit pattern that takes the shape priors library into optimization while ignoring the intrinsic information mining at the sketch and text level." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.624, + 0.442, + 0.641 + ], + "angle": 0, + "content": "2.4. Diffusion Models for 3D Shape Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Recent methods [26, 66, 62, 33] generate 3D shapes with denoising diffusion probabilistic models. DPM [26] introduced the flow-based model for point cloud embedding that conditions the diffusion model on geometry representation. PVD [66] trains a diffusion model on the point-voxel representation for 3D shapes. It performs the variational lower bound optimizing on unconditional shape generation and conditioned shape completion to produce complete 3D shapes. These methods show great promise in extending the diffusion theory to 3D shape generation and achieving good performance. However, their generations are relatively uncontrolled and lack certainty for our generation problem which needs to follow alignment with user instructions more closely. LION [62] proposed a text-based 3D generation model by combining the CLIP-forge to align the text-to-3D embedding. DreamFusion [33] is a generative method for text based 3D rendering, which replaces CLIP" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.243 + ], + "angle": 0, + "content": "with a loss derived from the distillation of 2D diffusion models. Point-E [30] starts from a language prompt to generate the single view image with fine-tuned GLIDE model and then uses the generated image for image-to-shape diffusion. Nonetheless, these text-based 3D diffusion models rely on feature pre-alignment from text to image and usually fail when there are huge domain gaps. Our method focuses on information mining for 3D diffusion by exploring the 3D structure from sketch and text to facilitate efficient controlled generation of high quality 3D shapes." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.248, + 0.691, + 0.265 + ], + "angle": 0, + "content": "3. Proposed Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.892, + 0.372 + ], + "angle": 0, + "content": "Given a sketch \\(\\mathbb{S}\\) and text description \\(\\mathbb{T}\\) of an object, the goal is to generate its 3D geometry and appearance. To achieve this, the proposed STPD model consists of three modules: sparse feature extraction network for sketch embedding (Section 3.1), sketch-text feature fusion (Section 3.2), and staged diffusion for shape and color generation (Section 3.3). Figure 2 shows an overview of our model." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.376, + 0.834, + 0.391 + ], + "angle": 0, + "content": "3.1. Sparse Feature Extraction of a Sketch" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.892, + 0.589 + ], + "angle": 0, + "content": "Feature extraction from a sketch faces three major challenges. Firstly, sketches are sparse and consist of very few informative pixels. Secondly, visual models are generally trained on large amounts of natural images and their attention mechanisms exploit the information in dense pixels to focus on the more relevant parts. This becomes challenging when the image (i.e. sketch) contains copious amounts of futile background pixels and the training dataset is also small. Finally, the relationship between different object parts is ambiguous due to self-occlusions in a single sketch. Below, as shown in Fig. 3, we present our capsule network using dynamic routing [40] with attention to effectively learning the sparse features from a sketch image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.892, + 0.697 + ], + "angle": 0, + "content": "Primary Capsule Layer: Our sketch extraction starts from a pixel-wise CNN embedding \\( S_{N,D_{input},W,H} = f_{cnn}(\\mathbb{S}) \\), where \\( N \\) is the core number of caps, \\( D \\) is the dimension of each caps, and \\( W, H \\) are the width, height of the sketch. Based on these embeddings, we perform a \\( 1 \\times 1 \\) Conv of filter size \\( D \\) to share an affine transformation on each capsule channel and then turn it to a primary capsule \\( u_{N,D,w,h}^{i,0} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.697, + 0.891, + 0.717 + ], + "angle": 0, + "content": "\\[\nu _ {(N, D, w, h)} ^ {i, 0} = \\operatorname {C o n v} \\left(S _ {(N, D, w, h)}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.781 + ], + "angle": 0, + "content": "Here, \\(i\\) denotes the index of capsules in the current layer, \\(w,h\\) is the size of current caps layers and Conv is a simple network comprising a \\(1\\times 1\\) convolution layer with stride 1, followed by batch normalization, ReLU, and max pooling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Dynamic Routing with Attention: Dynamic routing is applied to strengthen the structural relationship inference between different sketch parts. This facilitates robust feature learning from sketches while keeping the network depth in control and the number of parameters low. We add an attention module to force the capsules to focus on the useful sketch parts and ignore the background. Each \\( s^j \\) can be acquired by the sum of previous layer capsules," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8931" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.51, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.089, + 0.891, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.192, + 0.51, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.187, + 0.88, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.323, + 0.895, + 0.368 + ], + "angle": 0, + "content": "Figure 2. Overview of STPD framework. The sketch feature is first extracted with two level attention modules for sparse embedding and then flows with multi-head attention fusion to text. With the embedded features, we proposed a staged denoising probabilistic diffusion model for both the geometry and appearance of 3D shapes." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.377, + 0.475, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.48, + 0.471, + 0.536 + ], + "angle": 0, + "content": "Figure 3. Proposed sketch feature extractor. Pixel-wise attention and instance-wise attention are implemented by a CNN and attention routing, respectively. The two level attentions are designed to extract the sparse information from hand drawn sketches." + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.548, + 0.47, + 0.58 + ], + "angle": 0, + "content": "\\[\ns _ {(N, D, w, h)} ^ {j} = \\sum_ {i = 1, 2, \\dots} c _ {w, h} ^ {i j} a _ {(N, D, 1, 1)} ^ {i j} u _ {(N, D, w, h)} ^ {i, j - 1}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.582, + 0.47, + 0.643 + ], + "angle": 0, + "content": "where \\( c^{ij} \\) are the dynamic routing parameters computed by softmax layers and \\( a^{ij} \\) is the attention weight obtained by \\( 1 \\times 1 \\) Conv on lower level primary capsules and hence associated with each instance of the original sketch." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.644, + 0.47, + 0.722 + ], + "angle": 0, + "content": "Similar to the \\(l\\) times dynamic routing, the capsules are squashed [40], to cluster the informative parts, and then reshaped to a one dimension vector \\(S_{(1,D_f)}\\) (where \\(D_{f} = N\\times D\\)) from the last \\(s^j\\). This \\(S\\) forms the sketch embedding used to condition the reverse diffusion process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.729, + 0.426, + 0.745 + ], + "angle": 0, + "content": "3.2. Sketch-Text Feature Fusion Embedding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.84 + ], + "angle": 0, + "content": "We encode the input language expression as a sequence of textual embeddings with BERT [7]. Applying a linear projection to the BERT embedding, we get embedding \\( T \\) of the same dimension as the sketch embedding \\( S \\). Both embeddings are fused with two multi-head attention modules to guide the reverse diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The attention-ed embedding is formulated as a set of cascaded multi-attention modules. The first attention takes the sketch feature \\( S \\) as query to extract the relevant intermediate semantic information \\( I \\) from the text embedding. The" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.893, + 0.472 + ], + "angle": 0, + "content": "second attention performs text-guided self-attention for the sketch feature to produce the conditional embedding that contains geometry/color information based on \\( I \\). Architectures for geometry and color diffusion are the same but have different parameters. The first multi-head attention (MH Attention 1 in Fig. 2) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.478, + 0.892, + 0.496 + ], + "angle": 0, + "content": "\\[\nI _ {(1, D _ {f})} = \\operatorname {A t t e n} _ {1} \\left(Q _ {1} = S, K _ {1} = V _ {1} = T\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.892, + 0.547 + ], + "angle": 0, + "content": "where \\(Q_{1}, K_{1}, V_{1}\\) are the Query, Key, Value and \\(I\\) is an intermediate attention-ed feature passed to the second multi-head attention module (MH Attention 2) defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.551, + 0.892, + 0.568 + ], + "angle": 0, + "content": "\\[\nC _ {p \\left(1, D _ {f}\\right)} = A t t e n _ {2} \\left(Q _ {2} = K _ {2} = I + S, V _ {2} = S\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.893, + 0.753 + ], + "angle": 0, + "content": "where \\( Q_{2}, K_{2}, V_{2} \\) are the Query, Key, Value and \\( p = g \\) for geometric diffusion and \\( p = a \\) for appearance diffusion. Specifically, we take the sum of intermediate features \\( I \\) with the sketch feature \\( S \\) as the input of the query and key in the second multi-head attention. Then, the text prompt is introduced as the vector input to support the appearance denoising using the color information from the attention module. Note that, \\( C_{g} \\) and \\( C_{a} \\) are from the same cascaded attentions but focus on different descriptions (geometry or appearance) in each diffusion process. Appearance condition also requires information from sketch (shape) because different object parts can have different colors." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.758, + 0.85, + 0.774 + ], + "angle": 0, + "content": "3.3. Staged Colored Point Cloud Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.775, + 0.892, + 0.834 + ], + "angle": 0, + "content": "Markov's chain: Given a sample \\(\\mathbf{x}_0 \\sim q(\\mathbf{x}_0)\\), noise is systematically added to \\(\\mathbf{x}_0\\) in the forward process [19] step-by-step following the Markov's chain assumption until the Gaussian distribution is reached:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.836, + 0.892, + 0.877 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {0: T}\\right) = q \\left(\\mathbf {x} _ {0}\\right) \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.88, + 0.838, + 0.906 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right),\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8932" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.095, + 0.47, + 0.185 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}(\\mu, \\sigma^2)\\) denotes a Gaussian distribution. Here, \\(\\mathbf{x}_0\\) can represent either coordinates or colors from 3D shapes. We train geometry and appearance diffusion separately to generate colored point clouds. The reverse process, \\(p_\\theta(\\mathbf{x}_t \\mid \\mathbf{x}_{t+1})\\), starts from a standard Gaussian prior and ends with the desired \\(\\mathbf{x}_0\\):" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.183, + 0.469, + 0.22 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {x} _ {0: T}\\right) = p \\left(\\mathbf {x} _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.224, + 0.4, + 0.241 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) = \\mathcal {N} \\left(\\mu_ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.469, + 0.316 + ], + "angle": 0, + "content": "This cross-entropy minimization is designed such that it forces the reverse process to follow the pattern of noise addition of the forward process, and eventually, recover \\(\\mathbf{x}_0\\) from the standard Gaussian noise distribution. Hence, the loss function can be derived as:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.317, + 0.469, + 0.335 + ], + "angle": 0, + "content": "\\[\nL _ {t} = \\mathbb {E} _ {\\mathbf {x} _ {0}, e} [ C \\| \\tilde {\\mu} _ {t} (\\mathbf {x} _ {t}, \\mathbf {x} _ {0}) - \\mu_ {t} (\\mathbf {x} _ {t}, t) \\| ^ {2} ], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.337, + 0.228, + 0.351 + ], + "angle": 0, + "content": "where \\(C\\) is a constant." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.469, + 0.475 + ], + "angle": 0, + "content": "Geometry Diffusion: Denoting the colored point cloud as \\(\\mathbf{x}_0\\in \\mathbb{R}^{N\\times 6} = \\{\\mathbf{g}_0\\in \\mathbb{R}^{N\\times 3},\\mathbf{a}_0\\in \\mathbb{R}^{N\\times 3}\\}\\), consisting \\(N\\) points with geometry \\(\\mathbf{g}_0\\) and appearance \\(\\mathbf{a}_0\\) coordinates. We first perform geometry diffusion to generate the 3D shape. STPD is trained by minimizing the cross-entropy loss between two diffusion chains with respect to the geometric conditional feature \\(C_g\\), and learning of model parameters \\(\\theta\\):" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.473, + 0.469, + 0.511 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {g _ {0} \\sim q (g _ {0}), g _ {1: T} \\sim q (g _ {1: T})} \\left[ \\sum_ {t = 1} ^ {T} \\log p _ {\\theta} \\left(\\mathbf {g} _ {t - 1} \\mid \\mathbf {g} _ {t}, C _ {g}\\right) \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.469, + 0.618 + ], + "angle": 0, + "content": "where \\( C_g \\) is derived from sketch-text embedding when the text also contains shape description, or only from sketch when there is no text. We add a switch to adjust the feature map selection for different scenarios. The joint posterior \\( q(g_{1:T}) \\) can be factorized by the assumption of Markov's chain \\( \\prod_{t=1}^{T} q(\\mathbf{g}_{t-1}|\\mathbf{g}_t,\\mathbf{g}_0) \\), where each is analytically tractable by a re-parameterized Gaussian distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.627, + 0.469, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {\\mu}} _ {t} = \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1})}{1 - \\bar {\\alpha} _ {t}} \\mathbf {x} _ {t} + \\frac {\\sqrt {\\bar {\\alpha} _ {t - 1}} \\beta_ {t}}{1 - \\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.664, + 0.469, + 0.689 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {g} _ {t - 1} \\mid \\mathbf {g} _ {t}, \\mathbf {g} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {g} _ {t - 1}; \\tilde {\\boldsymbol {\\mu}} _ {t}, \\tilde {\\boldsymbol {\\beta}} _ {t} \\mathbf {I}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.695, + 0.469, + 0.771 + ], + "angle": 0, + "content": "where \\(\\tilde{\\beta}_t\\) is the known variance defined in the forward process, \\(\\alpha_{t} = 1 - \\beta_{t}\\), and \\(\\tilde{\\alpha} = \\prod_{s = 1}^{t}\\alpha_{s}\\). From Eq. 7 and this property, our diffusion model learns to match each \\(q\\left(\\mathbf{g}_{t - 1}\\mid \\mathbf{g}_t,\\mathbf{g}_0\\right)\\) and \\(p_{\\theta}\\left(\\mathbf{g}_{t - 1}\\mid \\mathbf{g}_t\\right)\\) by estimating the output noise \\(\\epsilon_{\\theta}(\\mathbf{g}_t,C_g,t)\\) to match noise \\(\\epsilon\\):" + }, + { + "type": "equation", + "bbox": [ + 0.159, + 0.773, + 0.469, + 0.792 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\mathbf {g} _ {t}, C _ {g}, t\\right) \\right\\| ^ {2}, \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.799, + 0.469, + 0.828 + ], + "angle": 0, + "content": "The noise prediction network \\(\\epsilon_{\\theta}\\) progressively generates the 3D shape using the reverse chain" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.83, + 0.45, + 0.865 + ], + "angle": 0, + "content": "\\[\ng _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {g} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (g _ {t}, C _ {g}, t)\\right) + \\sqrt {\\beta_ {t}} \\epsilon .\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Appearance Diffusion: Appearance diffusion is performed while freezing the geometry parameters. The optimization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "function, in this case, can be simply modified from Eq. 8 to learn a conditional generative model for appearance:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.122, + 0.897, + 0.162 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {a _ {0} \\sim q (a _ {0}), a _ {1: T} \\sim q (a _ {1: T})} [ \\sum_ {t = 1} ^ {T} \\log p _ {\\theta} (\\mathbf {a} _ {t - 1} | \\mathbf {a} _ {t}, p _ {\\theta} (\\mathbf {g} _ {0}), C _ {a}) ]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.163, + 0.892, + 0.193 + ], + "angle": 0, + "content": "The training loss for appearance diffusion is minimized and conditioned on appearance features \\( C_a \\) and \\( p_{\\theta}(\\mathbf{g}_0) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.196, + 0.891, + 0.214 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\mathbf {a} _ {t}, p _ {\\theta} \\left(\\mathbf {g} _ {0}\\right), C _ {a}, t\\right) \\right\\| ^ {2}, \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.218, + 0.892, + 0.263 + ], + "angle": 0, + "content": "Colored Point Cloud Generation: We can now formally define our joint generative model for colored point cloud generation as" + }, + { + "type": "equation", + "bbox": [ + 0.499, + 0.27, + 0.906, + 0.287 + ], + "angle": 0, + "content": "\\[\np _ {\\theta_ {2}, \\theta_ {1}, \\phi} (\\mathbf {a} _ {0}, \\mathbf {g} _ {0}, C _ {g}, C _ {a}) = p _ {\\theta_ {2}} (\\mathbf {a} _ {0} | \\mathbf {g} _ {0}, C _ {a}) p _ {\\theta_ {1}} (\\mathbf {g} _ {0} | C _ {g}) p _ {\\phi} (C _ {g}, C _ {a}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.294, + 0.892, + 0.445 + ], + "angle": 0, + "content": "where \\(\\{\\mathbf{a}_0,\\mathbf{g}_0\\}\\) denote the colored point cloud \\(\\mathbf{x}_0\\). \\(p_{\\phi}(C_g,C_a)\\) refers to the sketch and textual embedding, and \\(p_{\\theta_2}(\\mathbf{a}_0),p_{\\phi}(\\mathbf{g}_0)\\) are the appearance and geometry denoising processes respectively. Note that the appearance diffusion is conditioned on the geometry diffusion as well. We also tried extended diffusion to denoise from a 6D Gaussian rather than treating the shape and appearance separately. However, integrating the unrelated features causes undesirable interference in the diffusion process and consequently, STPD does not learn to produce point clouds with fine details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.452, + 0.633, + 0.468 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.473, + 0.691, + 0.489 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.493, + 0.892, + 0.689 + ], + "angle": 0, + "content": "Datasets: We use the chair, table, aeroplane, car categories of ShapeNet dataset [3], randomly split it into training-test sets at \\(80 - 20\\%\\) ratio and render a single-image from each 3D object. We apply Canny edge detection [24] on the rendered ShapeNet images to acquire their single-channel edge images. We manually remove the background so that they resemble clean hand drawn sketches. For colored point clouds, we use the chair, table categories as only they have text descriptions [4]. To generate ground-truth point clouds, each 3D shape is randomly sampled at 2048 points and then translated to zero mean and normalized in the range \\([-1,1]^3\\) so that the Gaussian distribution assumption is satisfied during the forward pass." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics: Previous works [26] used Chamfer Distance (CD) and Earth Mover's Distance (EMD) to evaluate the reconstruction quality of point clouds. Based on these, existing works report Coverage (COV) and Minimum matching distance (MMD). Although these metrics have their limitations, they still serve as a good reference to measure and compare the performance of generative methods. Specifically, COV evaluates the certainty of conditioned generation whereas MMD calculates the average distance of the point clouds in the reference set and their closest neighbors in the generated set. Another metric used in numerous works to evaluate the generation quality and diversity is 1-nearest neighbor accuracy (1-NNA) which directly quantifies distribution similarity between two sets. However, since" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "8933" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.086, + 0.891, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.208, + 0.894, + 0.251 + ], + "angle": 0, + "content": "Figure 4. Visual comparison of the generated objects. 1) Inputs include text and free-hand sketch. 2) Shape methods with no color. 3) Colored shape generation methods. 4) Our method generates both geometry and appearance using sketch and text. The appearance can be re-edited by changing the language prompt while preserving geometry." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.255, + 0.473, + 0.55 + ], + "angle": 0, + "content": "
CategoriesInputMethodsMMD (↓) CD EMDCOV (%,↑) CD EMD
Reconstruction Shape (N×3)SketchPSGN [10]8.1747.253
Pix2Mesh [51]9.7717.347
TMNet [32]9.188.11
Sketch2Mesh [17]3.0182.786
SketchesPix2Vox++ [57]5.9765.760
EvoIT [49]5.7185.307
Generative Shape (N×3)TextText2Shape [5]9.8736.348
CLIP-forge [43]4.3673.818
SketchMixNF [34]7.6896.224
DPF-Net [22]7.5316.176
Sketch2Model [65]3.1762.967
×DPM [26]3.7641.986
Sketch,textSTPD (ours)1.6351.043
Generative Shape&Color (N×6)TextText2Shape10.347.476
Point-E (text) [30]2.5232.265
×DPM-extended4.7653.256
sketchPoint-E (sketch)3.0923.117
Sketch,textSTPD (ours)1.3461.017
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.551, + 0.47, + 0.593 + ], + "angle": 0, + "content": "Table 1. Comparison of 3D shape generation for the chair, table categories of ShapeNet [3]. CD is multiplied by \\(10^{3}\\) and EMD by \\(10^{1}\\). Our method generates the best shape, with or without color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.603, + 0.47, + 0.649 + ], + "angle": 0, + "content": "our aim is to generate 3D shapes faithful to the provided sketch, we stick to the first four metrics. Where other methods do not provide color, we compare shapes only." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.65, + 0.471, + 0.831 + ], + "angle": 0, + "content": "Baselines: We compare with classical as well as diffusion based methods. In the former category, we select the major sketch [65], image [34] [22], and text [43] [5] based 3D object generation methods. We also compare to reconstruction based methods including single-view [10, 51, 32, 17] and multiviews [58, 49]. In the latter category, we select 3D diffusion (DPM) [26], Point-E [30], and Point Voxel Diffusion (PVD) [66]. Point-E (sketch) results are obtained after fine-tuning the Point-E image-to-shape diffusion model on our sketch-shape data for a fair comparison. We take ten generations for all methods and use the average error from ground truth for evaluation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.838, + 0.45, + 0.854 + ], + "angle": 0, + "content": "4.2. Comparisons with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Table 1 compares the point cloud generation quality of our method with existing state-of-the-art. For a fair comparison, we re-train all methods with our sketch-shape data of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.257, + 0.893, + 0.499 + ], + "angle": 0, + "content": "the chair, table categories, or their text descriptions when the method takes text only as input. Multiview methods are given sketches from 3 viewpoints. Only our method performs (single view) sketch and text guided 3D object generation. Since DPM [26] is unconditional, we generate many shapes and pick the best matches. For shape only (upper half of Table 1), the proposed STPD outperforms all methods, achieving \\(45.8\\%\\) lower CD-MMD error from the nearest reconstruction method Sketch2Mesh [17] and \\(48.5\\%\\) lower CD-MMD error from the nearest generative method Sketch2Model [65]. The multiview methods do not perform well since they are unable to establish correspondences between the sketches. Text guided generative methods have more diversity but struggle with alignment to user specifications given the inherent shape ambiguity in the text. Sketch guided methods perform better in general." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.893, + 0.851 + ], + "angle": 0, + "content": "Table 1 also compares colored point cloud (shape and color) generation. For this, we extend DPM to a 6D Gaussian and also compare with Text2Shape [4] and Point-E [30]. The proposed STPD again outperforms all methods, achieving \\(46.7\\%\\) CD-MMD error from the nearest competitor Point-E [30]. In Fig. 4, we compare our results visually to the existing methods using hand-drawn sketches from SketchX [23]. Our method also offers to reedit the color (appearance) once the geometry is finalized (last column). Table 2 reports comparative results on the chair, table, aeroplane, car categories of ShapeNet. We choose these categories because PVD [66] provides comparisons on the chair, aeroplane, car categories and the table category is used in our previous experiment. From Table 2, we can see that STPD achieves the best performance on both evaluation metrics on all four object categories. SoftFlow [21], PointFlow [60], DPF-Net [22], DPM [26], Point-E [30], PVD [66] achieve an average Chamfer Distance of 2.464, 2.429, 2.168, 2.509, 2.182, 2.539 and an average EMD of 3.105, 3.191, 2.851, 2.553, 2.490, 2.661 respectively. Our STPD achieves 1.172, 1.261 average CD, EMD errors which are \\(46.3\\%\\), \\(49.3\\%\\) lower than the nearest competitor Point-E [30]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Human Evaluations: We also conducted human evaluations on 100 objects generated by our method and other methods. Table 3 shows the results where 30 people rate" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8934" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.09, + 0.468, + 0.216 + ], + "angle": 0, + "content": "
Methodschairtableaeroplanecar
CDEMDCDEMDCDEMDCDEMD
SoftFlow [21]2.7863.2954.8155.1390.4041.1981.8502.789
PointFlow [60]2.7073.6494.8045.0820.4031.1801.8032.851
DPF-Net [22]2.7633.3203.9864.6610.5281.1051.3962.318
DPM [26]3.2012.7184.6014.2980.4321.0271.8042.167
Point-E [30]2.9012.6673.8094.2900.4121.0051.6091.997
PVD [66]3.2112.9394.7314.5270.4421.0301.7742.146
STPD (ours)1.2091.1172.1171.8970.3110.8971.0491.136
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.217, + 0.468, + 0.245 + ], + "angle": 0, + "content": "Table 2. Comparison with baselines on 4 categories of the ShapeNet dataset [3]. CD is multiplied by \\(10^{3}\\) and EMD by \\(10^{1}\\)." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.255, + 0.468, + 0.32 + ], + "angle": 0, + "content": "
Human rate (↑)SoftFlowPointFlowDPF-NetDPMPoint-EPVDSTPD
chair1.81.71.92.13.93.54.8
table2.11.93.53.03.42.54.6
aeroplane3.43.32.23.13.73.04.0
car2.52.43.43.53.22.94.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.32, + 0.468, + 0.346 + ], + "angle": 0, + "content": "Table 3. Human evaluations: 100 generated objects rated on a scale of 1 to 5 by 30 people." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.356, + 0.455, + 0.493 + ], + "angle": 0, + "content": "
Input Modality/Model for ConditioningCD (×103)EMD (×101)
SketchText
×BERT28.32520.457
C1×1×10.9658.658
C1×1BERT9.6518.459
ResNet50BERT6.4095.480
ResNet50 + Atten.RoutingBERT1.9351.587
C1×1 + Atten.Routing×1.6891.348
C1×1 + Atten.Routing (concatenate)BERT3.7484.127
C1×1 + Atten.Routing (our fusion)BERT1.3901.101
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.492, + 0.468, + 0.532 + ], + "angle": 0, + "content": "Table 4. Ablation study on visual & textual modules. \\( C_{1 \\times 1} \\) is a simple visual encoder comprising \\( 1 \\times 1 \\) convolution followed by batch normalization and ReLU." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.468, + 0.599 + ], + "angle": 0, + "content": "the generations on a scale of 1 to 5 (higher is better). STPD outperforms the baseline methods on all individual object categories and achieves an average rating of 4.4, compared to average rating of 3.3 by the nearest competitor Point-E." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.609, + 0.229, + 0.625 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We perform ablation study on the sketch and text modules to verify their contributions. Table 4 shows our results using CD and EMD metrics. Text conditioning, using BERT embeddings, alone achieves the worst performance. \\(C_{1 \\times 1}\\) is a simple visual encoder comprising \\(1 \\times 1\\) convolution followed by batch normalization and ReLU. Sketch conditioning, using simple \\(C_{1 \\times 1}\\) embeddings, reduces the errors significantly since a sketch can describe shaping much better than text. Using sketch and text conditioning with \\(C_{1 \\times 1}\\) with BERT embeddings, the errors do not reduce much. However, when a more sophisticated sketch encoding is performed with ResNet50 [18] and combined with BERT, the CD and EMD reduce notably. To account for the sparsity of sketches, when we add capsule network with attention routing on top of ResNet50, we see a dramatic improvement in both metrics. Our simple visual encoder \\(C_{1 \\times 1}\\) with attention routing further improves the performance even without using text conditioning. As seen in" + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.09, + 0.882, + 0.206 + ], + "angle": 0, + "content": "
ModelModelNet10ModelNet40ScanObjectNN
PointFlow [60]93.786.878.2
PDGN [20]94.287.378.1
3D-GAN [54]91.083.376.7
AtlasNet [15]91.986.676.9
ShapeGF [2]90.284.675.8
DPM [26]94.287.679.7
STPD94.888.184.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.565, + 0.206, + 0.827, + 0.218 + ], + "angle": 0, + "content": "Table 5. Object classification accuracy \\((\\%)\\)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.228, + 0.9, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.347, + 0.892, + 0.402 + ], + "angle": 0, + "content": "Figure 5. Visualization of attention score maps for a sketch with respect to capsules. The \\(8 \\times 16\\) size denotes 8 capsules of 16-dimension each. Capsule attention maps are learned from attention-routing. Dark parts denote high attention." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.414, + 0.892, + 0.518 + ], + "angle": 0, + "content": "the last row of the table, the best performance is achieved when we use sketch conditioning with \\( C_{1 \\times 1} + \\) attention routing embeddings combined with text embeddings from BERT [7] using the proposed fusion method. The second last row of the table shows that a simple concatenation of the two embeddings does not work as well as our proposed fusion mechanism." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.527, + 0.816, + 0.543 + ], + "angle": 0, + "content": "4.4. Representation Learning Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.892, + 0.77 + ], + "angle": 0, + "content": "We evaluate the representation learned by our sketch embedding used to condition the diffusion model. We follow the work of MVCNN [46], ScanObjectNN [48] and convert rendered images from different viewpoints of the ModelNet dataset [55] to sketches and conduct object classification experiments. We train our STPD on the generated sketches and then take the conditioned embedded features to learn a linear SVM classifier for classification based on a single sketch embedding. The essence of the conditional diffusion model is that the conditioning should guide the diffusion to retrieve the 3D shape from Gaussian noise. This means that our conditional network should be able to represent specific object categories to aid classification. The results in Table 5 demonstrate that our architecture can extract discriminative features, that are also useful for accurate classification." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.779, + 0.794, + 0.794 + ], + "angle": 0, + "content": "4.5. Effectiveness of Capsule Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Fig. 5 visualizes the generated attention scores from a typical sketch. More examples are given in the supplementary material along with quantitative results using Instance Scores for Sketch (ISS i.e. without attention routing) and Instance Score on Capsule (ISC i.e. with attention routing) metrics. ISS is the percentage of useful pixels divided by the total pixels of the sketch. For ISC, we compute the aver" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "8935" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.087, + 0.865, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.255, + 0.746, + 0.27 + ], + "angle": 0, + "content": "Figure 6. Visualization of part segmentation results for 4 categories of ShapeNet-Parts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.28, + 0.47, + 0.31 + ], + "angle": 0, + "content": "age of the highest attention score for each of the 16 capsule dimensions (column wise) and convert it to a percentage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.474, + 0.538 + ], + "angle": 0, + "content": "Before the attention capsule, the sketches have ISS of \\(1.07\\%\\) on average. After our capsule-attention module, ISC becomes \\(58.68\\%\\) on average for better diffusion. This demonstrates the effectiveness of the Capsule Attention Network which attends to sparse data in sketches during diffusion training. Each capsule \\((1\\times 16)\\) denotes a learned feature, e.g. the table, from a sketch. It can be observed that given the distribution of the useful pixel in the sketch, only a few attention score values are dark i.e., attentioned. Caps number 2 and a few dimensions of Caps number 5 are given high weight to contribute to the final visual feature, while the other Caps that represent the white background of the sketch are ignored. The attention-based capsule network enables our method to effectively locate the referred objects in the sparse sketch input." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.549, + 0.47, + 0.566 + ], + "angle": 0, + "content": "5. Extending STPD to Part Segmentation Task" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.47, + 0.636 + ], + "angle": 0, + "content": "To further show the feature learning ability of our STPD model, we use it for the task of part segmentation in point clouds. For this, we use the ShapeNet-Parts [3] dataset, focusing on the four object categories of Table 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.636, + 0.47, + 0.727 + ], + "angle": 0, + "content": "The part segmentation task can be extended from our colored point cloud generation model by assigning different colors to each part. We consider only two parts for the table and three parts for remaining three categories e.g. for aeroplane, the (body, wing, tail) are assigned (blue, green, red) colors respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.47, + 0.832 + ], + "angle": 0, + "content": "In our STPD, each coordinate is generated by our sketch-text guided geometric diffusion model. However, in the part segmentation task, the shape \\(\\mathbf{g}_0\\) is given as input to the model. Here, we directly apply the text feature as the condition \\(C_a\\) of appearance diffusion. For instance, we provide a generic text sentence \"An blue aeroplane with green wings and red tail\" for all aeroplanes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.833, + 0.47, + 0.879 + ], + "angle": 0, + "content": "With the given input geometry and appearance description, we can formally define our joint generative model for point cloud segmentation as" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.887, + 0.47, + 0.902 + ], + "angle": 0, + "content": "\\[\np _ {\\theta , \\phi} \\left(\\mathbf {a} _ {0}, \\mathbf {g} _ {0}, C _ {a}\\right) = p _ {\\theta} \\left(\\mathbf {a} _ {0} \\mid \\mathbf {g} _ {0}, C _ {a}\\right) p _ {\\text {i n p u t}} \\left(\\mathbf {g} _ {0}\\right) p _ {\\phi} \\left(C _ {a}\\right), \\tag {13}\n\\]" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.277, + 0.88, + 0.405 + ], + "angle": 0, + "content": "
ModelmIoUcmIoU1chairtableaeroplanecar
PointNet [35]82.183.589.680.683.474.9
PointNet++ [36]83.384.690.882.682.477.3
DGCNN [52]83.884.990.682.684.077.8
Point-BERT [61]84.785.691.081.584.379.8
PointMLP [27]84.684.890.384.383.580.5
KPConv [47]85.185.991.183.684.681.1
P2P [53]85.085.991.683.784.380.4
STPD85.986.792.484.185.681.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.406, + 0.892, + 0.433 + ], + "angle": 0, + "content": "Table 6. Comparative results of part segmentation on four categories of ShapeNet-Parts dataset." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.439, + 0.894, + 0.502 + ], + "angle": 0, + "content": "where the combination \\(\\{\\mathbf{a}_0, \\mathbf{g}_0\\}\\) denotes the desired segmentation of the 3D point cloud. Note that, \\(\\mathbf{a}_0\\) is generated as colors which are then used as segmentation labels. We apply K-means clustering for final refinement of the colors." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.507, + 0.894, + 0.569 + ], + "angle": 0, + "content": "Results for 3D Shape Segmentation: Fig. 6 shows some visualizations of our results and Table 6 compares our method to some existing baselines. The proposed STPD achieves the highest mIoU compared to existing methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.581, + 0.62, + 0.597 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.895, + 0.819 + ], + "angle": 0, + "content": "We proposed a novel sketch and text guided probabilistic diffusion model for 3D object generation. Our model takes a hand drawn sketch and text as inputs to condition the reverse diffusion process. With the strong generative ability and effective guidance by the sketch-text fused features, our model is able to achieve high alignment with user specifications. Our visual encoder is designed to focus on the informative sparse parts of sketches. The sketch-text fusion part gathers informative object context from visual and textual descriptions to improve the shape and appearance of the generated 3D objects. Experimental results demonstrate that our model achieves state-of-art performance in 3D object generation and shows promising applications for 3D object classification and part segmentation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.681, + 0.848 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Professor Ajmal Mian is the recipient of an Australian Research Council Future Fellowship Award (project number FT210100268) funded by the Australian Government." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8936" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] A.H. Bermano, R. Gal, Y. Alaluf, R. Mokady, Y. Nitzan, O. Tov, O. Patashnik, and D. Cohen-Or. State-of-the-art in the architecture, methods and applications of stylegan. Computer Graphics Forum, 41(2):591-611, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snively, and Bharath Hariharan. Learning gradient fields for shape generation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.31 + ], + "angle": 0, + "content": "[3] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.382 + ], + "angle": 0, + "content": "[4] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian conference on computer vision, pages 100-116. Springer, 2018. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.471, + 0.451 + ], + "angle": 0, + "content": "[5] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian conference on computer vision, pages 100-116. Springer, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[6] Shin-I Cheng, Yu-Jie Chen, Wei-Chen Chiu, Hung-Yu Tseng, and Hsin-Ying Lee. Adaptively-realistic image generation from stroke and sketch with diffusion model. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4054-4062, January 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.471, + 0.59 + ], + "angle": 0, + "content": "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.592, + 0.471, + 0.634 + ], + "angle": 0, + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.471, + 0.675 + ], + "angle": 0, + "content": "[9] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. arXiv preprint arXiv:1605.08803, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.47, + 0.733 + ], + "angle": 0, + "content": "[10] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[11] Mingtao Feng, Syed Zulqarnain Gilani, Yaonan Wang, Liang Zhang, and Ajmal Mian. Relation graph network for 3d object detection in point clouds. IEEE Transactions on Image Processing, 30:92-107, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[12] Mingtao Feng, Haoran Hou, Liang Zhang, Yulan Guo, Hongshan Yu, Yaonan Wang, and Ajmal Mian. Exploring hierarchical spatial layout cues for 3d point cloud based scene graph prediction. IEEE Transactions on Multimedia, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[13] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph net-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "work for object grounding in point cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3722-3731, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.137, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[14] Zhiheng Fu, Siyu Hong, Mengyi Liu, Hamid Laga, Mohammed Bennamoun, Farid Boussaid, and Yulan Guo. Multi-stage information diffusion for joint depth and surface normal estimation. Pattern Recognition, 141:109660, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.895, + 0.276 + ], + "angle": 0, + "content": "[15] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 216-224, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[16] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10696-10706, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[17] Benoit Guillard, Edoardo Remelli, Pierre Yvernay, and Pascal Fua. Sketch2mesh: Reconstructing and editing 3d shapes from sketches. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13023-13032, October 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[20] Le Hui, Rui Xu, Jin Xie, Jianjun Qian, and Jian Yang. Progressive point cloud deconvolution generation network. In European Conference on Computer Vision, pages 397-413. Springer, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[21] Hyeongju Kim, Hyeonseung Lee, Woo Hyun Kang, Joun Yeop Lee, and Nam Soo Kim. Softflow: Probabilistic framework for normalizing flow on manifolds. Advances in Neural Information Processing Systems, 33:16388-16397, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[22] Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. In European Conference on Computer Vision, pages 694-710. Springer, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[23] Yi Li, Timothy M Hospedales, Yi-Zhe Song, and Shaogang Gong. Fine-grained sketch-based image retrieval by matching deformable part models. 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[24] Yibo Li and Bailun Liu. Improved edge detection algorithm for canny operator. In 2022 IEEE 10th Joint International Information Technology and Artificial Intelligence Conference (ITAIC), volume 10, pages 1-5. IEEE, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8937" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[26] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2837-2845, 2021. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[27] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.205, + 0.469, + 0.259 + ], + "angle": 0, + "content": "[28] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Spectrum-guided multi-granularity referring video object segmentation. arXiv preprint arXiv:2307.13537, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.261, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[29] Yongwei Miao, Feixia Hu, Xudong Zhang, Jiazhou Chen, and Renato Pajarola. Symmsketch: Creating symmetric 3d free-form shapes from 2d sketches. Computational Visual Media, 1(1):3-16, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.317, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[30] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.373, + 0.469, + 0.412 + ], + "angle": 0, + "content": "[31] Luke Olsen, Faramarz F Samavati, Mario Costa Sousa, and Joaquim A Jorge. Sketch-based modeling: A survey. Computers & Graphics, 33(1):85-103, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.414, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[32] Junyi Pan, Xiaoguang Han, Weikai Chen, Jiapeng Tang, and Kui Jia. Deep mesh reconstruction from single rgb images via topology modification networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9964-9973, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.485, + 0.469, + 0.524 + ], + "angle": 0, + "content": "[33] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.526, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[34] Janis Postels, Mengya Liu, Riccardo Spezialetti, Luc Van Gool, and Federico Tombari. Go with the flows: Mixtures of normalizing flows for point cloud generation and reconstruction. In 2021 International Conference on 3D Vision (3DV), pages 1249-1258. IEEE, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.596, + 0.469, + 0.662 + ], + "angle": 0, + "content": "[35] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.665, + 0.469, + 0.719 + ], + "angle": 0, + "content": "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.722, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.805, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[38] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8821-8831. PMLR, 18-24 Jul 2021. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[40] Sara Sabour, Nicholas Frosst, and Geoffrey E Hinton. Dynamic routing between capsules. Advances in neural information processing systems, 30, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.208, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[41] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems, volume 35, pages 36479-36494. Curran Associates, Inc., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.415 + ], + "angle": 0, + "content": "[42] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[43] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[44] Yuefan Shen, Changgeng Zhang, Hongbo Fu, Kun Zhou, and Youyi Zheng. Deepsketchhair: Deep sketch-based 3d hair modeling. IEEE Transactions on Visualization and Computer Graphics, 27(7):3250-3263, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[45] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[46] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In Proceedings of the IEEE international conference on computer vision, pages 945-953, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[47] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[48] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In International Conference on Computer Vision (ICCV), 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[49] Dan Wang, Xinrui Cui, Xun Chen, Zhengxia Zou, Tianyang Shi, Septimiu Salcudean, Z. Jane Wang, and Rabab Ward. Multi-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF International Conference on Com" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8938" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "puter Vision (ICCV), pages 5722-5731, October 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[50] Fang Wang, Le Kang, and Yi Li. Sketch-based 3d shape retrieval using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1875-1883, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[51] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[52] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog), 38(5):1-12, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.361 + ], + "angle": 0, + "content": "[53] Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. P2p: Tuning pre-trained image models for point cloud analysis with point-to-pixel prompting. arXiv preprint arXiv:2208.02812, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[54] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.469, + 0.504 + ], + "angle": 0, + "content": "[55] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.506, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[56] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2256-2265, June 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.646 + ], + "angle": 0, + "content": "[57] Haozhe Xie, Hongxun Yao, Xiaoshuai Sun, Shangchen Zhou, and Shengping Zhang. Pix2vox: Context-aware 3d reconstruction from single and multi-view images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2690-2698, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.648, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[58] Haozhe Xie, Hongxun Yao, Shengping Zhang, Shangchen Zhou, and Wenxiu Sun. Pix2vox++: Multi-scale context-aware 3d object reconstruction from single and multiple images. International Journal of Computer Vision, 128(12):2919-2935, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[59] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Finegrained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[60] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4541-4550, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[61] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.893, + 0.134 + ], + "angle": 0, + "content": "transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.893, + 0.19 + ], + "angle": 0, + "content": "[62] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.893, + 0.274 + ], + "angle": 0, + "content": "[63] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.893, + 0.345 + ], + "angle": 0, + "content": "[64] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan++: Realistic image synthesis with stacked generative adversarial networks. IEEE transactions on pattern analysis and machine intelligence, 41(8):1947-1962, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.348, + 0.893, + 0.415 + ], + "angle": 0, + "content": "[65] Song-Hai Zhang, Yuan-Chen Guo, and Qing-Wen Gu. Sketch2model: View-aware 3d modeling from single free-hand sketches. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6012-6021, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.893, + 0.473 + ], + "angle": 0, + "content": "[66] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826–5835, 2021. 1, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.893, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8939" + } + ] +] \ No newline at end of file diff --git a/2023/Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation/6a47e5bb-0a8f-4500-ae70-d8f17ed82ef9_origin.pdf b/2023/Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation/6a47e5bb-0a8f-4500-ae70-d8f17ed82ef9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0d294aabb09afe8f436130d4f3b1209eb1d893ca --- /dev/null +++ b/2023/Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation/6a47e5bb-0a8f-4500-ae70-d8f17ed82ef9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa94a82af7ec421c247d00759ce1941636db59868353c4066271d63ab2665e40 +size 3180991 diff --git a/2023/Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation/full.md b/2023/Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6b242915af8a03c88e9d77feacfd287734339dd0 --- /dev/null +++ b/2023/Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation/full.md @@ -0,0 +1,382 @@ +# Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation + +Zijie $\mathbf{W}\mathbf{u}^{1*}$ Yaonan Wang $^{1}$ Mingtao Feng $^{2\dagger}$ He Xie $^{1}$ Ajmal Mian $^{3}$ + +$^{1}$ Hunan University $^{2}$ Xidian University $^{3}$ The University of Western Australia + +# Abstract + +Diffusion probabilistic models have achieved remarkable success in text guided image generation. However, generating 3D shapes is still challenging due to the lack of sufficient data containing 3D models along with their descriptions. Moreover, text based descriptions of 3D shapes are inherently ambiguous and lack details. In this paper, we propose a sketch and text guided probabilistic diffusion model for colored point cloud generation that conditions the denoising process jointly with a hand drawn sketch of the object and its textual description. We incrementally diffuse the point coordinates and color values in a joint diffusion process to reach a Gaussian distribution. Colored point cloud generation thus amounts to learning the reverse diffusion process, conditioned by the sketch and text, to iteratively recover the desired shape and color. Specifically, to learn effective sketch-text embedding, our model adaptively aggregates the joint embedding of text prompt and the sketch based on a capsule attention network. Our model uses staged diffusion to generate the shape and then assign colors to different parts conditioned on the appearance prompt while preserving precise shapes from the first stage. This gives our model the flexibility to extend to multiple tasks, such as appearance re-editing and part segmentation. Experimental results demonstrate that our model outperforms recent state-of-the-art in point cloud generation. + +# 1. Introduction + +Denoising Diffusion Probabilistic Models (DDPM) [19, 45] can generate novel images and videos, from their text descriptions. Recent diffusion models [38, 33, 39, 42] contain billions of parameters and are trained on millions of text-image [13] pairs sourced from the Internet, including existing datasets like MS-COCO [25]. The availability of such large corpora of text-image pair datasets is, arguably, the major driving force behind the success of research in image and video generation. However, such datasets of text-shape pairs are almost nonexistent causing a major bottleneck for advancing research in this direction. + +3D shape generation has a wide range of applications including data augmentation, virtual/augmented reality [12], + +![](images/8ace3db44bcce4d77d5dca71fe737efd5c35089b91d5ed53ee903096b54be033.jpg) +Figure 1. Illustration of sketch and text guided 3D point cloud generation. The text and sketch complement each other. + +manufacturing, and reverse engineering. Pioneering works in diffusion based 3D shape generation include Luo et al. [26] and Point Voxel Diffusion (PVD) [66]. Luo et al. [26] extend the DDPMs to 3D shape generation conditioned on shape latent that follows a prior distribution parameterized via normalized flows [9]. Their probabilistic generative model treats points as particles that, under heat, are diffused to a noise distribution. It then exploits the reverse diffusion process to learn the point distribution to be able to generate plausible point clouds. PVD [66] combines the denoising diffusion model with the point-voxel 3D shape representation. It performs the denoising steps by optimizing a variational lower bound to the conditional likelihood function to produce 3D shapes or multiple 3D completions [14] from a partial observation. + +The problem of text guided 3D model or point cloud generation is slightly different from that of images in the sense that there is a higher emphasis, at least implicitly, on geometry rather than appearance which puts yet another constraint on the training datasets. Current methods [26, 66] have mostly used the ShapeNet [3] dataset for text-shape learning combined with the ModelNet [55] datasets for 3D representation learning and validation. However, these datasets are relatively small and lack detailed textual descriptions. To work around the data paucity, CLIP-forge [42] extends the CLIP model [37] for 3D shape generation by aligning 3D shapes with text in the latent space. Although these text guided methods maintain diversity in the generated shapes, their common limitation is that the generated shapes are often not the desired ones, which is understandable given the shape ambiguity in the text. + +In this paper, to reduce the shape ambiguity, we propose a sketch and text guided probabilistic diffusion (STPD) model for colored point cloud generation that conditions the + +denoising process with sketch inputs, in addition to textual descriptions. Sketches can be hand drawn and give much more geometric details compared to text. We argue that, combined with text, sketches are a viable option for conditioning the denoising of 3D shapes given their direct relevance to the problem at hand and their history of successful use for 3D object reconstruction [65]. However, cross-modality inputs (sketch and text) [11] always suffer from large cross-modality discrepancies and intra-modality variations bringing in additional challenges of cross-modality fusion [28] and data sparsity. Especially, sketches are inherently sparse with only a very small proportion of informative pixels, and contain much less information than natural images. + +The proposed STPD model can produce the geometry and appearance simultaneously and unambiguously from a sketch-text input, combining ease of access and completeness of description. It takes the sketch as the main shape description and shape information present in the text to complement the sketch input. Hence, shape information is extracted from the sketch as well as the text when there is shape information in the text. Appearance is extracted from the text input to give color to the generated point clouds as shown in Fig. 1. To summarize, our contributions are: + +1. We propose a staged probabilistic diffusion model for better control over the geometry and appearance to generate colored point clouds. Our model conditions the denoising process jointly with sketch and text inputs to generate 3D shapes. It can also be used for part-segmentation and appearance re-editing. +2. We propose an attention capsule based feature extraction module for encoding sketches, which are inherently sparse. Our method robustly gives more attention to the entities of the useful pixels and ignores the large number of meaningless capsules corresponding to the blank pixels in the sketch. +3. We present a sketch-text fusion network that efficiently abstracts the shape and color information from both sketch and language descriptions to guide the reverse 3D shape diffusion process. + +Extensive experiments on the ShapeNet dataset [3] and comparison to existing diffusion probabilistic model based generation as well as some classical shape reconstruction methods show that our model achieves state-of-the-art performance for colored point cloud generation. We also show the representation learning ability of our model by conducting 3D object classification experiments on the ModelNet40 [55] dataset and the application of our model to part segmentation on the ShapeNet-Parts dataset [3]. + +# 2. Related Work + +We first survey diffusion models for generating images. Next, we discuss 3D shape reconstruction methods fol + +lowed by classical 3D shape generation methods. Finally, we discuss diffusion models for 3D shape generation. + +# 2.1. Text-based 2D Image Generation + +Generative Adversarial Network (GAN): Early methods for text based image generation were based on GANs. Xu et al. [59] used attention mechanism for fine-grained text to image generation. TediGAN [56] uses GAN inversion based on the pretrained StyleGAN [1] for text to image generation. Zhang et al.[63][64] used multiple GAN architectures that stacked generators and discriminators to progressively generate multi-scale images. GANs are known to be difficult to train and have recently been surpassed by diffusion models [8]. + +Probabilistic Diffusion Models: Sohl-Dickstein et al. [45] proposed diffusion probabilistic models that are both tractable and flexible. Inspired by non-equilibrium thermodynamics, the diffusion process uses Markov chain [6] to convert a simple Gaussian distribution to a target data distribution. Keeping the steps of the chain small enough, each step can be analytically evaluated [16], and hence, the full chain can also be evaluated. At each step of the forward pass, a small perturbation is added to the data until it reaches the Gaussian distribution. Since the perturbation at each step is small, a model can be learned to estimate and remove the perturbation. It was not until 2020 that Ho et al. [19] showed the real power of diffusion models with high quality image generation by training on a weighted variational bound and denoising score matching with Langevin dynamics. Dhariwal and Nichol [8] showed the superiority of diffusion models over GANs. However, directly optimizing in pixel space comes with a high computational cost. To reduce the computational requirements, [39] performs diffusion in the latent feature representation space. Large models such as Imagen [41] and DALL-E2 [38], which uses the CLIP model [37], can generate diverse photo-realistic images. The success of diffusion models is largely attributed to fundamental research as well as the availability of huge datasets in the image domain. + +# 2.2.3D Reconstruction from Images + +Reconstructing 3D shapes from images has been studied for decades. Single-view reconstruction techniques are popular, however, recovering 3D shapes from a single image is an ill-posed problem, as it requires a strong prior shape knowledge. Fan et al. [10] proposed to directly regress point clouds from a single image. Wang et al. [51] deformed the ellipsoid mesh to form the 3D shape to learn the shape prior. Sketch2Mesh [17] employed the silhouettes matching match between a given single sketch and the learned model to refine the 3D shape. TMNet [32] uses triangular meshes and images to supervise the reconstruction of 3D shapes. + +Multiview images contain more geometric information useful for 3D reconstruction. $\mathrm{Pix2Vox + + [58]}$ fuse the mul + +tiple coarse 3D volumes, which are generated from each individual view, to form the well-refined 3D shapes. Wang et al. [49] include transformers to recover 3D information from multiviews. Reconstruction based methods usually demand exact image-to-shape match and strong prior shape knowledge, whereas generative methods offer a balance between diversity and alignment with user specifications. + +# 2.3. Sketch, Text and 3D Shape Generation + +Sketch to 3D Shape: Sketches have been used as a sparse representation of natural images and 3D shapes. Sketches are quite illustrative, despite their abstract nature and simplicity. Sketch based 3D generation [31] can be coarsely categorized into evocative [50] and constructive [29] methods, which depend on retrieval from model collections or free-form modeling. Deep models learn the pattern of the sketch for 3D model generation. For example, Zhang et al. [65] proposed a view-aware technique to gain priors from existing 3D models using a single freehand sketch. DeepSketchHair [44] proposed a learning based system for intuitive 3D styling from 2D sketches. These methods have less dependence on a shape prior to the generation phase, but the tremendous requirement of shape-related data during training for pattern learning. + +Text to 3D Shape: Text to shape generation is a relatively new direction. Examples include Text2shape [4] and CLIP-Forge [42]. Text2shape [4] generates colored 3D shapes using joint embedding of text and 3D shapes. CLIP-Forge [42] generates 3D shapes from the pre-trained text-to-image CLIP [37], addressing the lack of text-to-shape priors data. Although previous studies have considered sketch and text as flexible methods for 3D shape generation, they only refine the model in an implicit pattern that takes the shape priors library into optimization while ignoring the intrinsic information mining at the sketch and text level. + +# 2.4. Diffusion Models for 3D Shape Generation + +Recent methods [26, 66, 62, 33] generate 3D shapes with denoising diffusion probabilistic models. DPM [26] introduced the flow-based model for point cloud embedding that conditions the diffusion model on geometry representation. PVD [66] trains a diffusion model on the point-voxel representation for 3D shapes. It performs the variational lower bound optimizing on unconditional shape generation and conditioned shape completion to produce complete 3D shapes. These methods show great promise in extending the diffusion theory to 3D shape generation and achieving good performance. However, their generations are relatively uncontrolled and lack certainty for our generation problem which needs to follow alignment with user instructions more closely. LION [62] proposed a text-based 3D generation model by combining the CLIP-forge to align the text-to-3D embedding. DreamFusion [33] is a generative method for text based 3D rendering, which replaces CLIP + +with a loss derived from the distillation of 2D diffusion models. Point-E [30] starts from a language prompt to generate the single view image with fine-tuned GLIDE model and then uses the generated image for image-to-shape diffusion. Nonetheless, these text-based 3D diffusion models rely on feature pre-alignment from text to image and usually fail when there are huge domain gaps. Our method focuses on information mining for 3D diffusion by exploring the 3D structure from sketch and text to facilitate efficient controlled generation of high quality 3D shapes. + +# 3. Proposed Approach + +Given a sketch $\mathbb{S}$ and text description $\mathbb{T}$ of an object, the goal is to generate its 3D geometry and appearance. To achieve this, the proposed STPD model consists of three modules: sparse feature extraction network for sketch embedding (Section 3.1), sketch-text feature fusion (Section 3.2), and staged diffusion for shape and color generation (Section 3.3). Figure 2 shows an overview of our model. + +# 3.1. Sparse Feature Extraction of a Sketch + +Feature extraction from a sketch faces three major challenges. Firstly, sketches are sparse and consist of very few informative pixels. Secondly, visual models are generally trained on large amounts of natural images and their attention mechanisms exploit the information in dense pixels to focus on the more relevant parts. This becomes challenging when the image (i.e. sketch) contains copious amounts of futile background pixels and the training dataset is also small. Finally, the relationship between different object parts is ambiguous due to self-occlusions in a single sketch. Below, as shown in Fig. 3, we present our capsule network using dynamic routing [40] with attention to effectively learning the sparse features from a sketch image. + +Primary Capsule Layer: Our sketch extraction starts from a pixel-wise CNN embedding $S_{N,D_{input},W,H} = f_{cnn}(\mathbb{S})$ , where $N$ is the core number of caps, $D$ is the dimension of each caps, and $W, H$ are the width, height of the sketch. Based on these embeddings, we perform a $1 \times 1$ Conv of filter size $D$ to share an affine transformation on each capsule channel and then turn it to a primary capsule $u_{N,D,w,h}^{i,0}$ . + +$$ +u _ {(N, D, w, h)} ^ {i, 0} = \operatorname {C o n v} \left(S _ {(N, D, w, h)}\right). \tag {1} +$$ + +Here, $i$ denotes the index of capsules in the current layer, $w,h$ is the size of current caps layers and Conv is a simple network comprising a $1\times 1$ convolution layer with stride 1, followed by batch normalization, ReLU, and max pooling. + +Dynamic Routing with Attention: Dynamic routing is applied to strengthen the structural relationship inference between different sketch parts. This facilitates robust feature learning from sketches while keeping the network depth in control and the number of parameters low. We add an attention module to force the capsules to focus on the useful sketch parts and ignore the background. Each $s^j$ can be acquired by the sum of previous layer capsules, + +![](images/5a0c311ac4c4a3709e16936908d2ae079366bedf45a12ecf5fbf347517ceb780.jpg) + +![](images/b69df35504669f84f471324036294e38160235d1c721427f3f08b97643fbd17e.jpg) + +![](images/92a7b16a08ae7dcd5baec19c08e0d3b540a7fecf2d65cf623c400209f8298e12.jpg) +Figure 2. Overview of STPD framework. The sketch feature is first extracted with two level attention modules for sparse embedding and then flows with multi-head attention fusion to text. With the embedded features, we proposed a staged denoising probabilistic diffusion model for both the geometry and appearance of 3D shapes. + +![](images/a7e3f0c1be04271896399af966cc41409cf0cbc173c657b0f253cb92a191c8b5.jpg) + +![](images/4664f824102e2ab4ec668d47ff938e0ea2453127d21e816f675e51e706a79ba3.jpg) +Figure 3. Proposed sketch feature extractor. Pixel-wise attention and instance-wise attention are implemented by a CNN and attention routing, respectively. The two level attentions are designed to extract the sparse information from hand drawn sketches. + +$$ +s _ {(N, D, w, h)} ^ {j} = \sum_ {i = 1, 2, \dots} c _ {w, h} ^ {i j} a _ {(N, D, 1, 1)} ^ {i j} u _ {(N, D, w, h)} ^ {i, j - 1}, \tag {2} +$$ + +where $c^{ij}$ are the dynamic routing parameters computed by softmax layers and $a^{ij}$ is the attention weight obtained by $1 \times 1$ Conv on lower level primary capsules and hence associated with each instance of the original sketch. + +Similar to the $l$ times dynamic routing, the capsules are squashed [40], to cluster the informative parts, and then reshaped to a one dimension vector $S_{(1,D_f)}$ (where $D_{f} = N\times D$ ) from the last $s^j$ . This $S$ forms the sketch embedding used to condition the reverse diffusion process. + +# 3.2. Sketch-Text Feature Fusion Embedding + +We encode the input language expression as a sequence of textual embeddings with BERT [7]. Applying a linear projection to the BERT embedding, we get embedding $T$ of the same dimension as the sketch embedding $S$ . Both embeddings are fused with two multi-head attention modules to guide the reverse diffusion process. + +The attention-ed embedding is formulated as a set of cascaded multi-attention modules. The first attention takes the sketch feature $S$ as query to extract the relevant intermediate semantic information $I$ from the text embedding. The + +second attention performs text-guided self-attention for the sketch feature to produce the conditional embedding that contains geometry/color information based on $I$ . Architectures for geometry and color diffusion are the same but have different parameters. The first multi-head attention (MH Attention 1 in Fig. 2) is defined as: + +$$ +I _ {(1, D _ {f})} = \operatorname {A t t e n} _ {1} \left(Q _ {1} = S, K _ {1} = V _ {1} = T\right), \tag {3} +$$ + +where $Q_{1}, K_{1}, V_{1}$ are the Query, Key, Value and $I$ is an intermediate attention-ed feature passed to the second multi-head attention module (MH Attention 2) defined as: + +$$ +C _ {p \left(1, D _ {f}\right)} = A t t e n _ {2} \left(Q _ {2} = K _ {2} = I + S, V _ {2} = S\right), \tag {4} +$$ + +where $Q_{2}, K_{2}, V_{2}$ are the Query, Key, Value and $p = g$ for geometric diffusion and $p = a$ for appearance diffusion. Specifically, we take the sum of intermediate features $I$ with the sketch feature $S$ as the input of the query and key in the second multi-head attention. Then, the text prompt is introduced as the vector input to support the appearance denoising using the color information from the attention module. Note that, $C_{g}$ and $C_{a}$ are from the same cascaded attentions but focus on different descriptions (geometry or appearance) in each diffusion process. Appearance condition also requires information from sketch (shape) because different object parts can have different colors. + +# 3.3. Staged Colored Point Cloud Generation + +Markov's chain: Given a sample $\mathbf{x}_0 \sim q(\mathbf{x}_0)$ , noise is systematically added to $\mathbf{x}_0$ in the forward process [19] step-by-step following the Markov's chain assumption until the Gaussian distribution is reached: + +$$ +q \left(\mathbf {x} _ {0: T}\right) = q \left(\mathbf {x} _ {0}\right) \prod_ {t = 1} ^ {T} q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {t - 1}\right), \tag {5} +$$ + +$$ +q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {t - 1}\right) = \mathcal {N} \left(\sqrt {1 - \beta_ {t}} \mathbf {x} _ {t - 1}, \beta_ {t} \mathbf {I}\right), +$$ + +where $\mathcal{N}(\mu, \sigma^2)$ denotes a Gaussian distribution. Here, $\mathbf{x}_0$ can represent either coordinates or colors from 3D shapes. We train geometry and appearance diffusion separately to generate colored point clouds. The reverse process, $p_\theta(\mathbf{x}_t \mid \mathbf{x}_{t+1})$ , starts from a standard Gaussian prior and ends with the desired $\mathbf{x}_0$ : + +$$ +p _ {\theta} \left(\mathbf {x} _ {0: T}\right) = p \left(\mathbf {x} _ {T}\right) \prod_ {t = 1} ^ {T} p _ {\theta} \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}\right), \tag {6} +$$ + +$$ +p _ {\theta} \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}\right) = \mathcal {N} \left(\mu_ {\theta} \left(\mathbf {x} _ {t}, t\right), \sigma_ {t} ^ {2} \mathbf {I}\right). +$$ + +This cross-entropy minimization is designed such that it forces the reverse process to follow the pattern of noise addition of the forward process, and eventually, recover $\mathbf{x}_0$ from the standard Gaussian noise distribution. Hence, the loss function can be derived as: + +$$ +L _ {t} = \mathbb {E} _ {\mathbf {x} _ {0}, e} [ C \| \tilde {\mu} _ {t} (\mathbf {x} _ {t}, \mathbf {x} _ {0}) - \mu_ {t} (\mathbf {x} _ {t}, t) \| ^ {2} ], \tag {7} +$$ + +where $C$ is a constant. + +Geometry Diffusion: Denoting the colored point cloud as $\mathbf{x}_0\in \mathbb{R}^{N\times 6} = \{\mathbf{g}_0\in \mathbb{R}^{N\times 3},\mathbf{a}_0\in \mathbb{R}^{N\times 3}\}$ , consisting $N$ points with geometry $\mathbf{g}_0$ and appearance $\mathbf{a}_0$ coordinates. We first perform geometry diffusion to generate the 3D shape. STPD is trained by minimizing the cross-entropy loss between two diffusion chains with respect to the geometric conditional feature $C_g$ , and learning of model parameters $\theta$ : + +$$ +\min _ {\theta} \mathbb {E} _ {g _ {0} \sim q (g _ {0}), g _ {1: T} \sim q (g _ {1: T})} \left[ \sum_ {t = 1} ^ {T} \log p _ {\theta} \left(\mathbf {g} _ {t - 1} \mid \mathbf {g} _ {t}, C _ {g}\right) \right], \tag {8} +$$ + +where $C_g$ is derived from sketch-text embedding when the text also contains shape description, or only from sketch when there is no text. We add a switch to adjust the feature map selection for different scenarios. The joint posterior $q(g_{1:T})$ can be factorized by the assumption of Markov's chain $\prod_{t=1}^{T} q(\mathbf{g}_{t-1}|\mathbf{g}_t,\mathbf{g}_0)$ , where each is analytically tractable by a re-parameterized Gaussian distribution: + +$$ +\tilde {\boldsymbol {\mu}} _ {t} = \frac {\sqrt {\alpha_ {t}} (1 - \bar {\alpha} _ {t - 1})}{1 - \bar {\alpha} _ {t}} \mathbf {x} _ {t} + \frac {\sqrt {\bar {\alpha} _ {t - 1}} \beta_ {t}}{1 - \bar {\alpha} _ {t}} \mathbf {x} _ {0}, \tag {9} +$$ + +$$ +q \left(\mathbf {g} _ {t - 1} \mid \mathbf {g} _ {t}, \mathbf {g} _ {0}\right) = \mathcal {N} \left(\mathbf {g} _ {t - 1}; \tilde {\boldsymbol {\mu}} _ {t}, \tilde {\boldsymbol {\beta}} _ {t} \mathbf {I}\right), \tag {10} +$$ + +where $\tilde{\beta}_t$ is the known variance defined in the forward process, $\alpha_{t} = 1 - \beta_{t}$ , and $\tilde{\alpha} = \prod_{s = 1}^{t}\alpha_{s}$ . From Eq. 7 and this property, our diffusion model learns to match each $q\left(\mathbf{g}_{t - 1}\mid \mathbf{g}_t,\mathbf{g}_0\right)$ and $p_{\theta}\left(\mathbf{g}_{t - 1}\mid \mathbf{g}_t\right)$ by estimating the output noise $\epsilon_{\theta}(\mathbf{g}_t,C_g,t)$ to match noise $\epsilon$ : + +$$ +\left\| \epsilon - \epsilon_ {\theta} \left(\mathbf {g} _ {t}, C _ {g}, t\right) \right\| ^ {2}, \epsilon \sim \mathcal {N} (0, \mathbf {I}), \tag {11} +$$ + +The noise prediction network $\epsilon_{\theta}$ progressively generates the 3D shape using the reverse chain + +$$ +g _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(\mathbf {g} _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\theta} (g _ {t}, C _ {g}, t)\right) + \sqrt {\beta_ {t}} \epsilon . +$$ + +Appearance Diffusion: Appearance diffusion is performed while freezing the geometry parameters. The optimization + +function, in this case, can be simply modified from Eq. 8 to learn a conditional generative model for appearance: + +$$ +\min _ {\theta} \mathbb {E} _ {a _ {0} \sim q (a _ {0}), a _ {1: T} \sim q (a _ {1: T})} [ \sum_ {t = 1} ^ {T} \log p _ {\theta} (\mathbf {a} _ {t - 1} | \mathbf {a} _ {t}, p _ {\theta} (\mathbf {g} _ {0}), C _ {a}) ] +$$ + +The training loss for appearance diffusion is minimized and conditioned on appearance features $C_a$ and $p_{\theta}(\mathbf{g}_0)$ : + +$$ +\left\| \epsilon - \epsilon_ {\theta} \left(\mathbf {a} _ {t}, p _ {\theta} \left(\mathbf {g} _ {0}\right), C _ {a}, t\right) \right\| ^ {2}, \epsilon \sim \mathcal {N} (0, \mathbf {I}). \tag {12} +$$ + +Colored Point Cloud Generation: We can now formally define our joint generative model for colored point cloud generation as + +$$ +p _ {\theta_ {2}, \theta_ {1}, \phi} (\mathbf {a} _ {0}, \mathbf {g} _ {0}, C _ {g}, C _ {a}) = p _ {\theta_ {2}} (\mathbf {a} _ {0} | \mathbf {g} _ {0}, C _ {a}) p _ {\theta_ {1}} (\mathbf {g} _ {0} | C _ {g}) p _ {\phi} (C _ {g}, C _ {a}), +$$ + +where $\{\mathbf{a}_0,\mathbf{g}_0\}$ denote the colored point cloud $\mathbf{x}_0$ . $p_{\phi}(C_g,C_a)$ refers to the sketch and textual embedding, and $p_{\theta_2}(\mathbf{a}_0),p_{\phi}(\mathbf{g}_0)$ are the appearance and geometry denoising processes respectively. Note that the appearance diffusion is conditioned on the geometry diffusion as well. We also tried extended diffusion to denoise from a 6D Gaussian rather than treating the shape and appearance separately. However, integrating the unrelated features causes undesirable interference in the diffusion process and consequently, STPD does not learn to produce point clouds with fine details. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets: We use the chair, table, aeroplane, car categories of ShapeNet dataset [3], randomly split it into training-test sets at $80 - 20\%$ ratio and render a single-image from each 3D object. We apply Canny edge detection [24] on the rendered ShapeNet images to acquire their single-channel edge images. We manually remove the background so that they resemble clean hand drawn sketches. For colored point clouds, we use the chair, table categories as only they have text descriptions [4]. To generate ground-truth point clouds, each 3D shape is randomly sampled at 2048 points and then translated to zero mean and normalized in the range $[-1,1]^3$ so that the Gaussian distribution assumption is satisfied during the forward pass. + +Evaluation Metrics: Previous works [26] used Chamfer Distance (CD) and Earth Mover's Distance (EMD) to evaluate the reconstruction quality of point clouds. Based on these, existing works report Coverage (COV) and Minimum matching distance (MMD). Although these metrics have their limitations, they still serve as a good reference to measure and compare the performance of generative methods. Specifically, COV evaluates the certainty of conditioned generation whereas MMD calculates the average distance of the point clouds in the reference set and their closest neighbors in the generated set. Another metric used in numerous works to evaluate the generation quality and diversity is 1-nearest neighbor accuracy (1-NNA) which directly quantifies distribution similarity between two sets. However, since + +![](images/b71d2e6c5f1cecb2f13fe6c0778c7a93efc94cc1e575c5923eb0b31dcc6936a1.jpg) +Figure 4. Visual comparison of the generated objects. 1) Inputs include text and free-hand sketch. 2) Shape methods with no color. 3) Colored shape generation methods. 4) Our method generates both geometry and appearance using sketch and text. The appearance can be re-edited by changing the language prompt while preserving geometry. + +
CategoriesInputMethodsMMD (↓) CD EMDCOV (%,↑) CD EMD
Reconstruction Shape (N×3)SketchPSGN [10]8.1747.253
Pix2Mesh [51]9.7717.347
TMNet [32]9.188.11
Sketch2Mesh [17]3.0182.786
SketchesPix2Vox++ [57]5.9765.760
EvoIT [49]5.7185.307
Generative Shape (N×3)TextText2Shape [5]9.8736.348
CLIP-forge [43]4.3673.818
SketchMixNF [34]7.6896.224
DPF-Net [22]7.5316.176
Sketch2Model [65]3.1762.967
×DPM [26]3.7641.986
Sketch,textSTPD (ours)1.6351.043
Generative Shape&Color (N×6)TextText2Shape10.347.476
Point-E (text) [30]2.5232.265
×DPM-extended4.7653.256
sketchPoint-E (sketch)3.0923.117
Sketch,textSTPD (ours)1.3461.017
+ +Table 1. Comparison of 3D shape generation for the chair, table categories of ShapeNet [3]. CD is multiplied by $10^{3}$ and EMD by $10^{1}$ . Our method generates the best shape, with or without color. + +our aim is to generate 3D shapes faithful to the provided sketch, we stick to the first four metrics. Where other methods do not provide color, we compare shapes only. + +Baselines: We compare with classical as well as diffusion based methods. In the former category, we select the major sketch [65], image [34] [22], and text [43] [5] based 3D object generation methods. We also compare to reconstruction based methods including single-view [10, 51, 32, 17] and multiviews [58, 49]. In the latter category, we select 3D diffusion (DPM) [26], Point-E [30], and Point Voxel Diffusion (PVD) [66]. Point-E (sketch) results are obtained after fine-tuning the Point-E image-to-shape diffusion model on our sketch-shape data for a fair comparison. We take ten generations for all methods and use the average error from ground truth for evaluation. + +# 4.2. Comparisons with State-of-the-art Methods + +Table 1 compares the point cloud generation quality of our method with existing state-of-the-art. For a fair comparison, we re-train all methods with our sketch-shape data of + +the chair, table categories, or their text descriptions when the method takes text only as input. Multiview methods are given sketches from 3 viewpoints. Only our method performs (single view) sketch and text guided 3D object generation. Since DPM [26] is unconditional, we generate many shapes and pick the best matches. For shape only (upper half of Table 1), the proposed STPD outperforms all methods, achieving $45.8\%$ lower CD-MMD error from the nearest reconstruction method Sketch2Mesh [17] and $48.5\%$ lower CD-MMD error from the nearest generative method Sketch2Model [65]. The multiview methods do not perform well since they are unable to establish correspondences between the sketches. Text guided generative methods have more diversity but struggle with alignment to user specifications given the inherent shape ambiguity in the text. Sketch guided methods perform better in general. + +Table 1 also compares colored point cloud (shape and color) generation. For this, we extend DPM to a 6D Gaussian and also compare with Text2Shape [4] and Point-E [30]. The proposed STPD again outperforms all methods, achieving $46.7\%$ CD-MMD error from the nearest competitor Point-E [30]. In Fig. 4, we compare our results visually to the existing methods using hand-drawn sketches from SketchX [23]. Our method also offers to reedit the color (appearance) once the geometry is finalized (last column). Table 2 reports comparative results on the chair, table, aeroplane, car categories of ShapeNet. We choose these categories because PVD [66] provides comparisons on the chair, aeroplane, car categories and the table category is used in our previous experiment. From Table 2, we can see that STPD achieves the best performance on both evaluation metrics on all four object categories. SoftFlow [21], PointFlow [60], DPF-Net [22], DPM [26], Point-E [30], PVD [66] achieve an average Chamfer Distance of 2.464, 2.429, 2.168, 2.509, 2.182, 2.539 and an average EMD of 3.105, 3.191, 2.851, 2.553, 2.490, 2.661 respectively. Our STPD achieves 1.172, 1.261 average CD, EMD errors which are $46.3\%$ , $49.3\%$ lower than the nearest competitor Point-E [30]. + +Human Evaluations: We also conducted human evaluations on 100 objects generated by our method and other methods. Table 3 shows the results where 30 people rate + +
Methodschairtableaeroplanecar
CDEMDCDEMDCDEMDCDEMD
SoftFlow [21]2.7863.2954.8155.1390.4041.1981.8502.789
PointFlow [60]2.7073.6494.8045.0820.4031.1801.8032.851
DPF-Net [22]2.7633.3203.9864.6610.5281.1051.3962.318
DPM [26]3.2012.7184.6014.2980.4321.0271.8042.167
Point-E [30]2.9012.6673.8094.2900.4121.0051.6091.997
PVD [66]3.2112.9394.7314.5270.4421.0301.7742.146
STPD (ours)1.2091.1172.1171.8970.3110.8971.0491.136
+ +Table 2. Comparison with baselines on 4 categories of the ShapeNet dataset [3]. CD is multiplied by $10^{3}$ and EMD by $10^{1}$ . + +
Human rate (↑)SoftFlowPointFlowDPF-NetDPMPoint-EPVDSTPD
chair1.81.71.92.13.93.54.8
table2.11.93.53.03.42.54.6
aeroplane3.43.32.23.13.73.04.0
car2.52.43.43.53.22.94.3
+ +Table 3. Human evaluations: 100 generated objects rated on a scale of 1 to 5 by 30 people. + +
Input Modality/Model for ConditioningCD (×103)EMD (×101)
SketchText
×BERT28.32520.457
C1×1×10.9658.658
C1×1BERT9.6518.459
ResNet50BERT6.4095.480
ResNet50 + Atten.RoutingBERT1.9351.587
C1×1 + Atten.Routing×1.6891.348
C1×1 + Atten.Routing (concatenate)BERT3.7484.127
C1×1 + Atten.Routing (our fusion)BERT1.3901.101
+ +the generations on a scale of 1 to 5 (higher is better). STPD outperforms the baseline methods on all individual object categories and achieves an average rating of 4.4, compared to average rating of 3.3 by the nearest competitor Point-E. + +# 4.3. Ablation Study + +We perform ablation study on the sketch and text modules to verify their contributions. Table 4 shows our results using CD and EMD metrics. Text conditioning, using BERT embeddings, alone achieves the worst performance. $C_{1 \times 1}$ is a simple visual encoder comprising $1 \times 1$ convolution followed by batch normalization and ReLU. Sketch conditioning, using simple $C_{1 \times 1}$ embeddings, reduces the errors significantly since a sketch can describe shaping much better than text. Using sketch and text conditioning with $C_{1 \times 1}$ with BERT embeddings, the errors do not reduce much. However, when a more sophisticated sketch encoding is performed with ResNet50 [18] and combined with BERT, the CD and EMD reduce notably. To account for the sparsity of sketches, when we add capsule network with attention routing on top of ResNet50, we see a dramatic improvement in both metrics. Our simple visual encoder $C_{1 \times 1}$ with attention routing further improves the performance even without using text conditioning. As seen in + +Table 4. Ablation study on visual & textual modules. $C_{1 \times 1}$ is a simple visual encoder comprising $1 \times 1$ convolution followed by batch normalization and ReLU. + +
ModelModelNet10ModelNet40ScanObjectNN
PointFlow [60]93.786.878.2
PDGN [20]94.287.378.1
3D-GAN [54]91.083.376.7
AtlasNet [15]91.986.676.9
ShapeGF [2]90.284.675.8
DPM [26]94.287.679.7
STPD94.888.184.3
+ +Table 5. Object classification accuracy $(\%)$ + +![](images/cda21b06b74008462a46dcc22296ce71e3f4576bcad40c9b17f82d414e0f1daf.jpg) +Figure 5. Visualization of attention score maps for a sketch with respect to capsules. The $8 \times 16$ size denotes 8 capsules of 16-dimension each. Capsule attention maps are learned from attention-routing. Dark parts denote high attention. + +the last row of the table, the best performance is achieved when we use sketch conditioning with $C_{1 \times 1} +$ attention routing embeddings combined with text embeddings from BERT [7] using the proposed fusion method. The second last row of the table shows that a simple concatenation of the two embeddings does not work as well as our proposed fusion mechanism. + +# 4.4. Representation Learning Evaluation + +We evaluate the representation learned by our sketch embedding used to condition the diffusion model. We follow the work of MVCNN [46], ScanObjectNN [48] and convert rendered images from different viewpoints of the ModelNet dataset [55] to sketches and conduct object classification experiments. We train our STPD on the generated sketches and then take the conditioned embedded features to learn a linear SVM classifier for classification based on a single sketch embedding. The essence of the conditional diffusion model is that the conditioning should guide the diffusion to retrieve the 3D shape from Gaussian noise. This means that our conditional network should be able to represent specific object categories to aid classification. The results in Table 5 demonstrate that our architecture can extract discriminative features, that are also useful for accurate classification. + +# 4.5. Effectiveness of Capsule Network + +Fig. 5 visualizes the generated attention scores from a typical sketch. More examples are given in the supplementary material along with quantitative results using Instance Scores for Sketch (ISS i.e. without attention routing) and Instance Score on Capsule (ISC i.e. with attention routing) metrics. ISS is the percentage of useful pixels divided by the total pixels of the sketch. For ISC, we compute the aver + +![](images/9c73ff89596e7aed9bf3f5347d276b016b30dee73976a0ca6ff744e471db1410.jpg) +Figure 6. Visualization of part segmentation results for 4 categories of ShapeNet-Parts. + +age of the highest attention score for each of the 16 capsule dimensions (column wise) and convert it to a percentage. + +Before the attention capsule, the sketches have ISS of $1.07\%$ on average. After our capsule-attention module, ISC becomes $58.68\%$ on average for better diffusion. This demonstrates the effectiveness of the Capsule Attention Network which attends to sparse data in sketches during diffusion training. Each capsule $(1\times 16)$ denotes a learned feature, e.g. the table, from a sketch. It can be observed that given the distribution of the useful pixel in the sketch, only a few attention score values are dark i.e., attentioned. Caps number 2 and a few dimensions of Caps number 5 are given high weight to contribute to the final visual feature, while the other Caps that represent the white background of the sketch are ignored. The attention-based capsule network enables our method to effectively locate the referred objects in the sparse sketch input. + +# 5. Extending STPD to Part Segmentation Task + +To further show the feature learning ability of our STPD model, we use it for the task of part segmentation in point clouds. For this, we use the ShapeNet-Parts [3] dataset, focusing on the four object categories of Table 2. + +The part segmentation task can be extended from our colored point cloud generation model by assigning different colors to each part. We consider only two parts for the table and three parts for remaining three categories e.g. for aeroplane, the (body, wing, tail) are assigned (blue, green, red) colors respectively. + +In our STPD, each coordinate is generated by our sketch-text guided geometric diffusion model. However, in the part segmentation task, the shape $\mathbf{g}_0$ is given as input to the model. Here, we directly apply the text feature as the condition $C_a$ of appearance diffusion. For instance, we provide a generic text sentence "An blue aeroplane with green wings and red tail" for all aeroplanes. + +With the given input geometry and appearance description, we can formally define our joint generative model for point cloud segmentation as + +$$ +p _ {\theta , \phi} \left(\mathbf {a} _ {0}, \mathbf {g} _ {0}, C _ {a}\right) = p _ {\theta} \left(\mathbf {a} _ {0} \mid \mathbf {g} _ {0}, C _ {a}\right) p _ {\text {i n p u t}} \left(\mathbf {g} _ {0}\right) p _ {\phi} \left(C _ {a}\right), \tag {13} +$$ + +
ModelmIoUcmIoU1chairtableaeroplanecar
PointNet [35]82.183.589.680.683.474.9
PointNet++ [36]83.384.690.882.682.477.3
DGCNN [52]83.884.990.682.684.077.8
Point-BERT [61]84.785.691.081.584.379.8
PointMLP [27]84.684.890.384.383.580.5
KPConv [47]85.185.991.183.684.681.1
P2P [53]85.085.991.683.784.380.4
STPD85.986.792.484.185.681.3
+ +Table 6. Comparative results of part segmentation on four categories of ShapeNet-Parts dataset. + +where the combination $\{\mathbf{a}_0, \mathbf{g}_0\}$ denotes the desired segmentation of the 3D point cloud. Note that, $\mathbf{a}_0$ is generated as colors which are then used as segmentation labels. We apply K-means clustering for final refinement of the colors. + +Results for 3D Shape Segmentation: Fig. 6 shows some visualizations of our results and Table 6 compares our method to some existing baselines. The proposed STPD achieves the highest mIoU compared to existing methods. + +# 6. Conclusion + +We proposed a novel sketch and text guided probabilistic diffusion model for 3D object generation. Our model takes a hand drawn sketch and text as inputs to condition the reverse diffusion process. With the strong generative ability and effective guidance by the sketch-text fused features, our model is able to achieve high alignment with user specifications. Our visual encoder is designed to focus on the informative sparse parts of sketches. The sketch-text fusion part gathers informative object context from visual and textual descriptions to improve the shape and appearance of the generated 3D objects. Experimental results demonstrate that our model achieves state-of-art performance in 3D object generation and shows promising applications for 3D object classification and part segmentation. + +# 7. Acknowledgement + +Professor Ajmal Mian is the recipient of an Australian Research Council Future Fellowship Award (project number FT210100268) funded by the Australian Government. + +# References + +[1] A.H. Bermano, R. Gal, Y. Alaluf, R. Mokady, Y. Nitzan, O. Tov, O. Patashnik, and D. Cohen-Or. State-of-the-art in the architecture, methods and applications of stylegan. Computer Graphics Forum, 41(2):591-611, 2022. 2 +[2] Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snively, and Bharath Hariharan. Learning gradient fields for shape generation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 7 +[3] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 1, 2, 5, 6, 7, 8 +[4] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian conference on computer vision, pages 100-116. Springer, 2018. 3, 5, 6 +[5] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian conference on computer vision, pages 100-116. Springer, 2018. 6 +[6] Shin-I Cheng, Yu-Jie Chen, Wei-Chen Chiu, Hung-Yu Tseng, and Hsin-Ying Lee. Adaptively-realistic image generation from stroke and sketch with diffusion model. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4054-4062, January 2023. 2 +[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 4, 7 +[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2 +[9] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. arXiv preprint arXiv:1605.08803, 2016. 1 +[10] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 2, 6 +[11] Mingtao Feng, Syed Zulqarnain Gilani, Yaonan Wang, Liang Zhang, and Ajmal Mian. Relation graph network for 3d object detection in point clouds. IEEE Transactions on Image Processing, 30:92-107, 2020. 2 +[12] Mingtao Feng, Haoran Hou, Liang Zhang, Yulan Guo, Hongshan Yu, Yaonan Wang, and Ajmal Mian. Exploring hierarchical spatial layout cues for 3d point cloud based scene graph prediction. IEEE Transactions on Multimedia, 2023. 1 +[13] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph net- + +work for object grounding in point cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3722-3731, 2021. 1 +[14] Zhiheng Fu, Siyu Hong, Mengyi Liu, Hamid Laga, Mohammed Bennamoun, Farid Boussaid, and Yulan Guo. Multi-stage information diffusion for joint depth and surface normal estimation. Pattern Recognition, 141:109660, 2023. 1 +[15] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 216-224, 2018. 7 +[16] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10696-10706, June 2022. 2 +[17] Benoit Guillard, Edoardo Remelli, Pierre Yvernay, and Pascal Fua. Sketch2mesh: Reconstructing and editing 3d shapes from sketches. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13023-13032, October 2021. 2, 6 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7 +[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1, 2, 4 +[20] Le Hui, Rui Xu, Jin Xie, Jianjun Qian, and Jian Yang. Progressive point cloud deconvolution generation network. In European Conference on Computer Vision, pages 397-413. Springer, 2020. 7 +[21] Hyeongju Kim, Hyeonseung Lee, Woo Hyun Kang, Joun Yeop Lee, and Nam Soo Kim. Softflow: Probabilistic framework for normalizing flow on manifolds. Advances in Neural Information Processing Systems, 33:16388-16397, 2020. 6, 7 +[22] Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. In European Conference on Computer Vision, pages 694-710. Springer, 2020. 6, 7 +[23] Yi Li, Timothy M Hospedales, Yi-Zhe Song, and Shaogang Gong. Fine-grained sketch-based image retrieval by matching deformable part models. 2014. 6 +[24] Yibo Li and Bailun Liu. Improved edge detection algorithm for canny operator. In 2022 IEEE 10th Joint International Information Technology and Artificial Intelligence Conference (ITAIC), volume 10, pages 1-5. IEEE, 2022. 5 +[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1 + +[26] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2837-2845, 2021. 1, 3, 5, 6, 7 +[27] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022. 8 +[28] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Spectrum-guided multi-granularity referring video object segmentation. arXiv preprint arXiv:2307.13537, 2023. 2 +[29] Yongwei Miao, Feixia Hu, Xudong Zhang, Jiazhou Chen, and Renato Pajarola. Symmsketch: Creating symmetric 3d free-form shapes from 2d sketches. Computational Visual Media, 1(1):3-16, 2015. 3 +[30] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 3, 6, 7 +[31] Luke Olsen, Faramarz F Samavati, Mario Costa Sousa, and Joaquim A Jorge. Sketch-based modeling: A survey. Computers & Graphics, 33(1):85-103, 2009. 3 +[32] Junyi Pan, Xiaoguang Han, Weikai Chen, Jiapeng Tang, and Kui Jia. Deep mesh reconstruction from single rgb images via topology modification networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9964-9973, 2019. 2, 6 +[33] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 3 +[34] Janis Postels, Mengya Liu, Riccardo Spezialetti, Luc Van Gool, and Federico Tombari. Go with the flows: Mixtures of normalizing flows for point cloud generation and reconstruction. In 2021 International Conference on 3D Vision (3DV), pages 1249-1258. IEEE, 2021. 6 +[35] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 8 +[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 8 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3 +[38] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8821-8831. PMLR, 18-24 Jul 2021. 1, 2 + +[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 1, 2 +[40] Sara Sabour, Nicholas Frosst, and Geoffrey E Hinton. Dynamic routing between capsules. Advances in neural information processing systems, 30, 2017. 3, 4 +[41] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems, volume 35, pages 36479-36494. Curran Associates, Inc., 2022. 2 +[42] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 1, 3 +[43] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 6 +[44] Yuefan Shen, Changgeng Zhang, Hongbo Fu, Kun Zhou, and Youyi Zheng. Deepsketchhair: Deep sketch-based 3d hair modeling. IEEE Transactions on Visualization and Computer Graphics, 27(7):3250-3263, 2021. 3 +[45] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 1, 2 +[46] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In Proceedings of the IEEE international conference on computer vision, pages 945-953, 2015. 7 +[47] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 8 +[48] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In International Conference on Computer Vision (ICCV), 2019. 7 +[49] Dan Wang, Xinrui Cui, Xun Chen, Zhengxia Zou, Tianyang Shi, Septimiu Salcudean, Z. Jane Wang, and Rabab Ward. Multi-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF International Conference on Com + +puter Vision (ICCV), pages 5722-5731, October 2021. 3, 6 +[50] Fang Wang, Le Kang, and Yi Li. Sketch-based 3d shape retrieval using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1875-1883, 2015. 3 +[51] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 2, 6 +[52] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog), 38(5):1-12, 2019. 8 +[53] Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. P2p: Tuning pre-trained image models for point cloud analysis with point-to-pixel prompting. arXiv preprint arXiv:2208.02812, 2022. 8 +[54] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 7 +[55] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 1, 2, 7 +[56] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2256-2265, June 2021. 2 +[57] Haozhe Xie, Hongxun Yao, Xiaoshuai Sun, Shangchen Zhou, and Shengping Zhang. Pix2vox: Context-aware 3d reconstruction from single and multi-view images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2690-2698, 2019. 6 +[58] Haozhe Xie, Hongxun Yao, Shengping Zhang, Shangchen Zhou, and Wenxiu Sun. Pix2vox++: Multi-scale context-aware 3d object reconstruction from single and multiple images. International Journal of Computer Vision, 128(12):2919-2935, 2020. 2, 6 +[59] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Finegrained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 2 +[60] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4541-4550, 2019. 6, 7 +[61] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud + +transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 8 +[62] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022. 3 +[63] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 2 +[64] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan++: Realistic image synthesis with stacked generative adversarial networks. IEEE transactions on pattern analysis and machine intelligence, 41(8):1947-1962, 2018. 2 +[65] Song-Hai Zhang, Yuan-Chen Guo, and Qing-Wen Gu. Sketch2model: View-aware 3d modeling from single free-hand sketches. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6012-6021, 2021. 2, 3, 6 +[66] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826–5835, 2021. 1, 3, 6, 7 \ No newline at end of file diff --git a/2023/Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation/images.zip b/2023/Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1afe0964d81b82e96a3f2027539584e6eb102b4f --- /dev/null +++ b/2023/Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c4d68b5b87f790d892a89291bc9d9fb136cbeaaf93ba5cf88105aacfadda9ca2 +size 647605 diff --git a/2023/Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation/layout.json b/2023/Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..74b9eeb34ec1faee545febdc88e24a06c20b73cc --- /dev/null +++ b/2023/Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation/layout.json @@ -0,0 +1,9851 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 103, + 536, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 536, + 120 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 536, + 120 + ], + "type": "text", + "content": "Sketch and Text Guided Diffusion Model for Colored Point Cloud Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 142, + 474, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 142, + 474, + 158 + ], + "spans": [ + { + "bbox": [ + 118, + 142, + 474, + 158 + ], + "type": "text", + "content": "Zijie " + }, + { + "bbox": [ + 118, + 142, + 474, + 158 + ], + "type": "inline_equation", + "content": "\\mathbf{W}\\mathbf{u}^{1*}" + }, + { + "bbox": [ + 118, + 142, + 474, + 158 + ], + "type": "text", + "content": " Yaonan Wang" + }, + { + "bbox": [ + 118, + 142, + 474, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 142, + 474, + 158 + ], + "type": "text", + "content": " Mingtao Feng" + }, + { + "bbox": [ + 118, + 142, + 474, + 158 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 118, + 142, + 474, + 158 + ], + "type": "text", + "content": " He Xie" + }, + { + "bbox": [ + 118, + 142, + 474, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 142, + 474, + 158 + ], + "type": "text", + "content": " Ajmal Mian" + }, + { + "bbox": [ + 118, + 142, + 474, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 158, + 460, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 158, + 460, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 158, + 460, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 158, + 460, + 171 + ], + "type": "text", + "content": "Hunan University " + }, + { + "bbox": [ + 133, + 158, + 460, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 158, + 460, + 171 + ], + "type": "text", + "content": "Xidian University " + }, + { + "bbox": [ + 133, + 158, + 460, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 158, + 460, + 171 + ], + "type": "text", + "content": "The University of Western Australia" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 213, + 290, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 213, + 290, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 290, + 514 + ], + "type": "text", + "content": "Diffusion probabilistic models have achieved remarkable success in text guided image generation. However, generating 3D shapes is still challenging due to the lack of sufficient data containing 3D models along with their descriptions. Moreover, text based descriptions of 3D shapes are inherently ambiguous and lack details. In this paper, we propose a sketch and text guided probabilistic diffusion model for colored point cloud generation that conditions the denoising process jointly with a hand drawn sketch of the object and its textual description. We incrementally diffuse the point coordinates and color values in a joint diffusion process to reach a Gaussian distribution. Colored point cloud generation thus amounts to learning the reverse diffusion process, conditioned by the sketch and text, to iteratively recover the desired shape and color. Specifically, to learn effective sketch-text embedding, our model adaptively aggregates the joint embedding of text prompt and the sketch based on a capsule attention network. Our model uses staged diffusion to generate the shape and then assign colors to different parts conditioned on the appearance prompt while preserving precise shapes from the first stage. This gives our model the flexibility to extend to multiple tasks, such as appearance re-editing and part segmentation. Experimental results demonstrate that our model outperforms recent state-of-the-art in point cloud generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 517, + 128, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 128, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 128, + 529 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 530, + 288, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 288, + 662 + ], + "type": "text", + "content": "Denoising Diffusion Probabilistic Models (DDPM) [19, 45] can generate novel images and videos, from their text descriptions. Recent diffusion models [38, 33, 39, 42] contain billions of parameters and are trained on millions of text-image [13] pairs sourced from the Internet, including existing datasets like MS-COCO [25]. The availability of such large corpora of text-image pair datasets is, arguably, the major driving force behind the success of research in image and video generation. However, such datasets of text-shape pairs are almost nonexistent causing a major bottleneck for advancing research in this direction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 662, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 662, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 47, + 662, + 287, + 688 + ], + "type": "text", + "content": "3D shape generation has a wide range of applications including data augmentation, virtual/augmented reality [12]," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 198, + 548, + 286 + ], + "blocks": [ + { + "bbox": [ + 306, + 198, + 548, + 286 + ], + "lines": [ + { + "bbox": [ + 306, + 198, + 548, + 286 + ], + "spans": [ + { + "bbox": [ + 306, + 198, + 548, + 286 + ], + "type": "image", + "image_path": "8ace3db44bcce4d77d5dca71fe737efd5c35089b91d5ed53ee903096b54be033.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 285, + 546, + 306 + ], + "lines": [ + { + "bbox": [ + 304, + 285, + 546, + 306 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 546, + 306 + ], + "type": "text", + "content": "Figure 1. Illustration of sketch and text guided 3D point cloud generation. The text and sketch complement each other." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 307, + 546, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 546, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 546, + 486 + ], + "type": "text", + "content": "manufacturing, and reverse engineering. Pioneering works in diffusion based 3D shape generation include Luo et al. [26] and Point Voxel Diffusion (PVD) [66]. Luo et al. [26] extend the DDPMs to 3D shape generation conditioned on shape latent that follows a prior distribution parameterized via normalized flows [9]. Their probabilistic generative model treats points as particles that, under heat, are diffused to a noise distribution. It then exploits the reverse diffusion process to learn the point distribution to be able to generate plausible point clouds. PVD [66] combines the denoising diffusion model with the point-voxel 3D shape representation. It performs the denoising steps by optimizing a variational lower bound to the conditional likelihood function to produce 3D shapes or multiple 3D completions [14] from a partial observation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 486, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 677 + ], + "type": "text", + "content": "The problem of text guided 3D model or point cloud generation is slightly different from that of images in the sense that there is a higher emphasis, at least implicitly, on geometry rather than appearance which puts yet another constraint on the training datasets. Current methods [26, 66] have mostly used the ShapeNet [3] dataset for text-shape learning combined with the ModelNet [55] datasets for 3D representation learning and validation. However, these datasets are relatively small and lack detailed textual descriptions. To work around the data paucity, CLIP-forge [42] extends the CLIP model [37] for 3D shape generation by aligning 3D shapes with text in the latent space. Although these text guided methods maintain diversity in the generated shapes, their common limitation is that the generated shapes are often not the desired ones, which is understandable given the shape ambiguity in the text." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "In this paper, to reduce the shape ambiguity, we propose a sketch and text guided probabilistic diffusion (STPD) model for colored point cloud generation that conditions the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 280, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 280, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 280, + 703 + ], + "type": "text", + "content": "*Work performed during visit at the University of Western Australia" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 59, + 703, + 133, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 133, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 133, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8929" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": "denoising process with sketch inputs, in addition to textual descriptions. Sketches can be hand drawn and give much more geometric details compared to text. We argue that, combined with text, sketches are a viable option for conditioning the denoising of 3D shapes given their direct relevance to the problem at hand and their history of successful use for 3D object reconstruction [65]. However, cross-modality inputs (sketch and text) [11] always suffer from large cross-modality discrepancies and intra-modality variations bringing in additional challenges of cross-modality fusion [28] and data sparsity. Especially, sketches are inherently sparse with only a very small proportion of informative pixels, and contain much less information than natural images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 239, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 289, + 360 + ], + "type": "text", + "content": "The proposed STPD model can produce the geometry and appearance simultaneously and unambiguously from a sketch-text input, combining ease of access and completeness of description. It takes the sketch as the main shape description and shape information present in the text to complement the sketch input. Hence, shape information is extracted from the sketch as well as the text when there is shape information in the text. Appearance is extracted from the text input to give color to the generated point clouds as shown in Fig. 1. To summarize, our contributions are:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 362, + 287, + 556 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 362, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 287, + 434 + ], + "type": "text", + "content": "1. We propose a staged probabilistic diffusion model for better control over the geometry and appearance to generate colored point clouds. Our model conditions the denoising process jointly with sketch and text inputs to generate 3D shapes. It can also be used for part-segmentation and appearance re-editing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 436, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 436, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 54, + 436, + 287, + 508 + ], + "type": "text", + "content": "2. We propose an attention capsule based feature extraction module for encoding sketches, which are inherently sparse. Our method robustly gives more attention to the entities of the useful pixels and ignores the large number of meaningless capsules corresponding to the blank pixels in the sketch." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 509, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 509, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 54, + 509, + 287, + 556 + ], + "type": "text", + "content": "3. We present a sketch-text fusion network that efficiently abstracts the shape and color information from both sketch and language descriptions to guide the reverse 3D shape diffusion process." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 560, + 287, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 560, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 560, + 287, + 668 + ], + "type": "text", + "content": "Extensive experiments on the ShapeNet dataset [3] and comparison to existing diffusion probabilistic model based generation as well as some classical shape reconstruction methods show that our model achieves state-of-the-art performance for colored point cloud generation. We also show the representation learning ability of our model by conducting 3D object classification experiments on the ModelNet40 [55] dataset and the application of our model to part segmentation on the ShapeNet-Parts dataset [3]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 675, + 134, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 675, + 134, + 687 + ], + "spans": [ + { + "bbox": [ + 47, + 675, + 134, + 687 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "We first survey diffusion models for generating images. Next, we discuss 3D shape reconstruction methods fol" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "lowed by classical 3D shape generation methods. Finally, we discuss diffusion models for 3D shape generation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 102, + 484, + 114 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 102, + 484, + 114 + ], + "spans": [ + { + "bbox": [ + 306, + 102, + 484, + 114 + ], + "type": "text", + "content": "2.1. Text-based 2D Image Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 118, + 545, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 118, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 118, + 545, + 236 + ], + "type": "text", + "content": "Generative Adversarial Network (GAN): Early methods for text based image generation were based on GANs. Xu et al. [59] used attention mechanism for fine-grained text to image generation. TediGAN [56] uses GAN inversion based on the pretrained StyleGAN [1] for text to image generation. Zhang et al.[63][64] used multiple GAN architectures that stacked generators and discriminators to progressively generate multi-scale images. GANs are known to be difficult to train and have recently been surpassed by diffusion models [8]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 237, + 546, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 546, + 537 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 546, + 537 + ], + "type": "text", + "content": "Probabilistic Diffusion Models: Sohl-Dickstein et al. [45] proposed diffusion probabilistic models that are both tractable and flexible. Inspired by non-equilibrium thermodynamics, the diffusion process uses Markov chain [6] to convert a simple Gaussian distribution to a target data distribution. Keeping the steps of the chain small enough, each step can be analytically evaluated [16], and hence, the full chain can also be evaluated. At each step of the forward pass, a small perturbation is added to the data until it reaches the Gaussian distribution. Since the perturbation at each step is small, a model can be learned to estimate and remove the perturbation. It was not until 2020 that Ho et al. [19] showed the real power of diffusion models with high quality image generation by training on a weighted variational bound and denoising score matching with Langevin dynamics. Dhariwal and Nichol [8] showed the superiority of diffusion models over GANs. However, directly optimizing in pixel space comes with a high computational cost. To reduce the computational requirements, [39] performs diffusion in the latent feature representation space. Large models such as Imagen [41] and DALL-E2 [38], which uses the CLIP model [37], can generate diverse photo-realistic images. The success of diffusion models is largely attributed to fundamental research as well as the availability of huge datasets in the image domain." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 542, + 478, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 542, + 478, + 554 + ], + "spans": [ + { + "bbox": [ + 306, + 542, + 478, + 554 + ], + "type": "text", + "content": "2.2.3D Reconstruction from Images" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "content": "Reconstructing 3D shapes from images has been studied for decades. Single-view reconstruction techniques are popular, however, recovering 3D shapes from a single image is an ill-posed problem, as it requires a strong prior shape knowledge. Fan et al. [10] proposed to directly regress point clouds from a single image. Wang et al. [51] deformed the ellipsoid mesh to form the 3D shape to learn the shape prior. Sketch2Mesh [17] employed the silhouettes matching match between a given single sketch and the learned model to refine the 3D shape. TMNet [32] uses triangular meshes and images to supervise the reconstruction of 3D shapes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Multiview images contain more geometric information useful for 3D reconstruction. " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Pix2Vox + + [58]}" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " fuse the mul" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8930" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "tiple coarse 3D volumes, which are generated from each individual view, to form the well-refined 3D shapes. Wang et al. [49] include transformers to recover 3D information from multiviews. Reconstruction based methods usually demand exact image-to-shape match and strong prior shape knowledge, whereas generative methods offer a balance between diversity and alignment with user specifications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 161, + 251, + 174 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 161, + 251, + 174 + ], + "spans": [ + { + "bbox": [ + 47, + 161, + 251, + 174 + ], + "type": "text", + "content": "2.3. Sketch, Text and 3D Shape Generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 177, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 177, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 287, + 357 + ], + "type": "text", + "content": "Sketch to 3D Shape: Sketches have been used as a sparse representation of natural images and 3D shapes. Sketches are quite illustrative, despite their abstract nature and simplicity. Sketch based 3D generation [31] can be coarsely categorized into evocative [50] and constructive [29] methods, which depend on retrieval from model collections or free-form modeling. Deep models learn the pattern of the sketch for 3D model generation. For example, Zhang et al. [65] proposed a view-aware technique to gain priors from existing 3D models using a single freehand sketch. DeepSketchHair [44] proposed a learning based system for intuitive 3D styling from 2D sketches. These methods have less dependence on a shape prior to the generation phase, but the tremendous requirement of shape-related data during training for pattern learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 357, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 287, + 489 + ], + "type": "text", + "content": "Text to 3D Shape: Text to shape generation is a relatively new direction. Examples include Text2shape [4] and CLIP-Forge [42]. Text2shape [4] generates colored 3D shapes using joint embedding of text and 3D shapes. CLIP-Forge [42] generates 3D shapes from the pre-trained text-to-image CLIP [37], addressing the lack of text-to-shape priors data. Although previous studies have considered sketch and text as flexible methods for 3D shape generation, they only refine the model in an implicit pattern that takes the shape priors library into optimization while ignoring the intrinsic information mining at the sketch and text level." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 494, + 270, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 494, + 270, + 507 + ], + "spans": [ + { + "bbox": [ + 47, + 494, + 270, + 507 + ], + "type": "text", + "content": "2.4. Diffusion Models for 3D Shape Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "text", + "content": "Recent methods [26, 66, 62, 33] generate 3D shapes with denoising diffusion probabilistic models. DPM [26] introduced the flow-based model for point cloud embedding that conditions the diffusion model on geometry representation. PVD [66] trains a diffusion model on the point-voxel representation for 3D shapes. It performs the variational lower bound optimizing on unconditional shape generation and conditioned shape completion to produce complete 3D shapes. These methods show great promise in extending the diffusion theory to 3D shape generation and achieving good performance. However, their generations are relatively uncontrolled and lack certainty for our generation problem which needs to follow alignment with user instructions more closely. LION [62] proposed a text-based 3D generation model by combining the CLIP-forge to align the text-to-3D embedding. DreamFusion [33] is a generative method for text based 3D rendering, which replaces CLIP" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": "with a loss derived from the distillation of 2D diffusion models. Point-E [30] starts from a language prompt to generate the single view image with fine-tuned GLIDE model and then uses the generated image for image-to-shape diffusion. Nonetheless, these text-based 3D diffusion models rely on feature pre-alignment from text to image and usually fail when there are huge domain gaps. Our method focuses on information mining for 3D diffusion by exploring the 3D structure from sketch and text to facilitate efficient controlled generation of high quality 3D shapes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 196, + 422, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 196, + 422, + 209 + ], + "spans": [ + { + "bbox": [ + 305, + 196, + 422, + 209 + ], + "type": "text", + "content": "3. Proposed Approach" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 210, + 545, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 545, + 294 + ], + "type": "text", + "content": "Given a sketch " + }, + { + "bbox": [ + 304, + 210, + 545, + 294 + ], + "type": "inline_equation", + "content": "\\mathbb{S}" + }, + { + "bbox": [ + 304, + 210, + 545, + 294 + ], + "type": "text", + "content": " and text description " + }, + { + "bbox": [ + 304, + 210, + 545, + 294 + ], + "type": "inline_equation", + "content": "\\mathbb{T}" + }, + { + "bbox": [ + 304, + 210, + 545, + 294 + ], + "type": "text", + "content": " of an object, the goal is to generate its 3D geometry and appearance. To achieve this, the proposed STPD model consists of three modules: sparse feature extraction network for sketch embedding (Section 3.1), sketch-text feature fusion (Section 3.2), and staged diffusion for shape and color generation (Section 3.3). Figure 2 shows an overview of our model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 297, + 510, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 297, + 510, + 309 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 510, + 309 + ], + "type": "text", + "content": "3.1. Sparse Feature Extraction of a Sketch" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 311, + 545, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 545, + 466 + ], + "type": "text", + "content": "Feature extraction from a sketch faces three major challenges. Firstly, sketches are sparse and consist of very few informative pixels. Secondly, visual models are generally trained on large amounts of natural images and their attention mechanisms exploit the information in dense pixels to focus on the more relevant parts. This becomes challenging when the image (i.e. sketch) contains copious amounts of futile background pixels and the training dataset is also small. Finally, the relationship between different object parts is ambiguous due to self-occlusions in a single sketch. Below, as shown in Fig. 3, we present our capsule network using dynamic routing [40] with attention to effectively learning the sparse features from a sketch image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 466, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 545, + 552 + ], + "type": "text", + "content": "Primary Capsule Layer: Our sketch extraction starts from a pixel-wise CNN embedding " + }, + { + "bbox": [ + 304, + 466, + 545, + 552 + ], + "type": "inline_equation", + "content": "S_{N,D_{input},W,H} = f_{cnn}(\\mathbb{S})" + }, + { + "bbox": [ + 304, + 466, + 545, + 552 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 466, + 545, + 552 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 466, + 545, + 552 + ], + "type": "text", + "content": " is the core number of caps, " + }, + { + "bbox": [ + 304, + 466, + 545, + 552 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 466, + 545, + 552 + ], + "type": "text", + "content": " is the dimension of each caps, and " + }, + { + "bbox": [ + 304, + 466, + 545, + 552 + ], + "type": "inline_equation", + "content": "W, H" + }, + { + "bbox": [ + 304, + 466, + 545, + 552 + ], + "type": "text", + "content": " are the width, height of the sketch. Based on these embeddings, we perform a " + }, + { + "bbox": [ + 304, + 466, + 545, + 552 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 466, + 545, + 552 + ], + "type": "text", + "content": " Conv of filter size " + }, + { + "bbox": [ + 304, + 466, + 545, + 552 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 466, + 545, + 552 + ], + "type": "text", + "content": " to share an affine transformation on each capsule channel and then turn it to a primary capsule " + }, + { + "bbox": [ + 304, + 466, + 545, + 552 + ], + "type": "inline_equation", + "content": "u_{N,D,w,h}^{i,0}" + }, + { + "bbox": [ + 304, + 466, + 545, + 552 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 358, + 552, + 545, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 552, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 358, + 552, + 545, + 567 + ], + "type": "interline_equation", + "content": "u _ {(N, D, w, h)} ^ {i, 0} = \\operatorname {C o n v} \\left(S _ {(N, D, w, h)}\\right). \\tag {1}", + "image_path": "bcef1b4b012d9c5350bba88160e0422d1d7f1cd69881f76f5ab7e02e458cd0bb.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 545, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 618 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 618 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 304, + 570, + 545, + 618 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 570, + 545, + 618 + ], + "type": "text", + "content": " denotes the index of capsules in the current layer, " + }, + { + "bbox": [ + 304, + 570, + 545, + 618 + ], + "type": "inline_equation", + "content": "w,h" + }, + { + "bbox": [ + 304, + 570, + 545, + 618 + ], + "type": "text", + "content": " is the size of current caps layers and Conv is a simple network comprising a " + }, + { + "bbox": [ + 304, + 570, + 545, + 618 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 304, + 570, + 545, + 618 + ], + "type": "text", + "content": " convolution layer with stride 1, followed by batch normalization, ReLU, and max pooling." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Dynamic Routing with Attention: Dynamic routing is applied to strengthen the structural relationship inference between different sketch parts. This facilitates robust feature learning from sketches while keeping the network depth in control and the number of parameters low. We add an attention module to force the capsules to focus on the useful sketch parts and ignore the background. Each " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "s^j" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": " can be acquired by the sum of previous layer capsules," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8931" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 312, + 151 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 312, + 151 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 312, + 151 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 312, + 151 + ], + "type": "image", + "image_path": "5a0c311ac4c4a3709e16936908d2ae079366bedf45a12ecf5fbf347517ceb780.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 312, + 70, + 545, + 148 + ], + "blocks": [ + { + "bbox": [ + 312, + 70, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 545, + 148 + ], + "type": "image", + "image_path": "b69df35504669f84f471324036294e38160235d1c721427f3f08b97643fbd17e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 152, + 312, + 251 + ], + "blocks": [ + { + "bbox": [ + 48, + 152, + 312, + 251 + ], + "lines": [ + { + "bbox": [ + 48, + 152, + 312, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 312, + 251 + ], + "type": "image", + "image_path": "92a7b16a08ae7dcd5baec19c08e0d3b540a7fecf2d65cf623c400209f8298e12.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 255, + 547, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 547, + 291 + ], + "type": "text", + "content": "Figure 2. Overview of STPD framework. The sketch feature is first extracted with two level attention modules for sparse embedding and then flows with multi-head attention fusion to text. With the embedded features, we proposed a staged denoising probabilistic diffusion model for both the geometry and appearance of 3D shapes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 148, + 538, + 251 + ], + "blocks": [ + { + "bbox": [ + 312, + 148, + 538, + 251 + ], + "lines": [ + { + "bbox": [ + 312, + 148, + 538, + 251 + ], + "spans": [ + { + "bbox": [ + 312, + 148, + 538, + 251 + ], + "type": "image", + "image_path": "a7e3f0c1be04271896399af966cc41409cf0cbc173c657b0f253cb92a191c8b5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 298, + 290, + 376 + ], + "blocks": [ + { + "bbox": [ + 48, + 298, + 290, + 376 + ], + "lines": [ + { + "bbox": [ + 48, + 298, + 290, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 290, + 376 + ], + "type": "image", + "image_path": "4664f824102e2ab4ec668d47ff938e0ea2453127d21e816f675e51e706a79ba3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 380, + 288, + 424 + ], + "lines": [ + { + "bbox": [ + 46, + 380, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 288, + 424 + ], + "type": "text", + "content": "Figure 3. Proposed sketch feature extractor. Pixel-wise attention and instance-wise attention are implemented by a CNN and attention routing, respectively. The two level attentions are designed to extract the sparse information from hand drawn sketches." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 434, + 287, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 434, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 64, + 434, + 287, + 459 + ], + "type": "interline_equation", + "content": "s _ {(N, D, w, h)} ^ {j} = \\sum_ {i = 1, 2, \\dots} c _ {w, h} ^ {i j} a _ {(N, D, 1, 1)} ^ {i j} u _ {(N, D, w, h)} ^ {i, j - 1}, \\tag {2}", + "image_path": "9538b199b111ccf2ad7e4d95cec7f20144c93e289bf2070f364ae3743c2afa12.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 460, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 287, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 460, + 287, + 509 + ], + "type": "inline_equation", + "content": "c^{ij}" + }, + { + "bbox": [ + 47, + 460, + 287, + 509 + ], + "type": "text", + "content": " are the dynamic routing parameters computed by softmax layers and " + }, + { + "bbox": [ + 47, + 460, + 287, + 509 + ], + "type": "inline_equation", + "content": "a^{ij}" + }, + { + "bbox": [ + 47, + 460, + 287, + 509 + ], + "type": "text", + "content": " is the attention weight obtained by " + }, + { + "bbox": [ + 47, + 460, + 287, + 509 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 47, + 460, + 287, + 509 + ], + "type": "text", + "content": " Conv on lower level primary capsules and hence associated with each instance of the original sketch." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 510, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 287, + 571 + ], + "type": "text", + "content": "Similar to the " + }, + { + "bbox": [ + 47, + 510, + 287, + 571 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 510, + 287, + 571 + ], + "type": "text", + "content": " times dynamic routing, the capsules are squashed [40], to cluster the informative parts, and then reshaped to a one dimension vector " + }, + { + "bbox": [ + 47, + 510, + 287, + 571 + ], + "type": "inline_equation", + "content": "S_{(1,D_f)}" + }, + { + "bbox": [ + 47, + 510, + 287, + 571 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 47, + 510, + 287, + 571 + ], + "type": "inline_equation", + "content": "D_{f} = N\\times D" + }, + { + "bbox": [ + 47, + 510, + 287, + 571 + ], + "type": "text", + "content": ") from the last " + }, + { + "bbox": [ + 47, + 510, + 287, + 571 + ], + "type": "inline_equation", + "content": "s^j" + }, + { + "bbox": [ + 47, + 510, + 287, + 571 + ], + "type": "text", + "content": ". This " + }, + { + "bbox": [ + 47, + 510, + 287, + 571 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 510, + 287, + 571 + ], + "type": "text", + "content": " forms the sketch embedding used to condition the reverse diffusion process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 577, + 260, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 577, + 260, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 577, + 260, + 590 + ], + "type": "text", + "content": "3.2. Sketch-Text Feature Fusion Embedding" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "text", + "content": "We encode the input language expression as a sequence of textual embeddings with BERT [7]. Applying a linear projection to the BERT embedding, we get embedding " + }, + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "text", + "content": " of the same dimension as the sketch embedding " + }, + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "text", + "content": ". Both embeddings are fused with two multi-head attention modules to guide the reverse diffusion process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "The attention-ed embedding is formulated as a set of cascaded multi-attention modules. The first attention takes the sketch feature " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " as query to extract the relevant intermediate semantic information " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " from the text embedding. The" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 300, + 546, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 546, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 546, + 373 + ], + "type": "text", + "content": "second attention performs text-guided self-attention for the sketch feature to produce the conditional embedding that contains geometry/color information based on " + }, + { + "bbox": [ + 304, + 300, + 546, + 373 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 300, + 546, + 373 + ], + "type": "text", + "content": ". Architectures for geometry and color diffusion are the same but have different parameters. The first multi-head attention (MH Attention 1 in Fig. 2) is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 336, + 378, + 545, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 378, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 336, + 378, + 545, + 392 + ], + "type": "interline_equation", + "content": "I _ {(1, D _ {f})} = \\operatorname {A t t e n} _ {1} \\left(Q _ {1} = S, K _ {1} = V _ {1} = T\\right), \\tag {3}", + "image_path": "74c6423e4ad4d3c34982853b11d6e77ab196f5ca2583c947d0849ba73c4a843e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 396, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 396, + 545, + 433 + ], + "type": "inline_equation", + "content": "Q_{1}, K_{1}, V_{1}" + }, + { + "bbox": [ + 304, + 396, + 545, + 433 + ], + "type": "text", + "content": " are the Query, Key, Value and " + }, + { + "bbox": [ + 304, + 396, + 545, + 433 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 396, + 545, + 433 + ], + "type": "text", + "content": " is an intermediate attention-ed feature passed to the second multi-head attention module (MH Attention 2) defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 331, + 436, + 545, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 436, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 331, + 436, + 545, + 449 + ], + "type": "interline_equation", + "content": "C _ {p \\left(1, D _ {f}\\right)} = A t t e n _ {2} \\left(Q _ {2} = K _ {2} = I + S, V _ {2} = S\\right), \\tag {4}", + "image_path": "ec2d27de53c97ee5bf31acc6c7689a21343835aad55ff46e7e9965bb8998e6e1.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 453, + 546, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 546, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 546, + 596 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 453, + 546, + 596 + ], + "type": "inline_equation", + "content": "Q_{2}, K_{2}, V_{2}" + }, + { + "bbox": [ + 304, + 453, + 546, + 596 + ], + "type": "text", + "content": " are the Query, Key, Value and " + }, + { + "bbox": [ + 304, + 453, + 546, + 596 + ], + "type": "inline_equation", + "content": "p = g" + }, + { + "bbox": [ + 304, + 453, + 546, + 596 + ], + "type": "text", + "content": " for geometric diffusion and " + }, + { + "bbox": [ + 304, + 453, + 546, + 596 + ], + "type": "inline_equation", + "content": "p = a" + }, + { + "bbox": [ + 304, + 453, + 546, + 596 + ], + "type": "text", + "content": " for appearance diffusion. Specifically, we take the sum of intermediate features " + }, + { + "bbox": [ + 304, + 453, + 546, + 596 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 453, + 546, + 596 + ], + "type": "text", + "content": " with the sketch feature " + }, + { + "bbox": [ + 304, + 453, + 546, + 596 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 453, + 546, + 596 + ], + "type": "text", + "content": " as the input of the query and key in the second multi-head attention. Then, the text prompt is introduced as the vector input to support the appearance denoising using the color information from the attention module. Note that, " + }, + { + "bbox": [ + 304, + 453, + 546, + 596 + ], + "type": "inline_equation", + "content": "C_{g}" + }, + { + "bbox": [ + 304, + 453, + 546, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 453, + 546, + 596 + ], + "type": "inline_equation", + "content": "C_{a}" + }, + { + "bbox": [ + 304, + 453, + 546, + 596 + ], + "type": "text", + "content": " are from the same cascaded attentions but focus on different descriptions (geometry or appearance) in each diffusion process. Appearance condition also requires information from sketch (shape) because different object parts can have different colors." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 600, + 520, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 600, + 520, + 613 + ], + "spans": [ + { + "bbox": [ + 305, + 600, + 520, + 613 + ], + "type": "text", + "content": "3.3. Staged Colored Point Cloud Generation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 613, + 545, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 613, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 613, + 545, + 660 + ], + "type": "text", + "content": "Markov's chain: Given a sample " + }, + { + "bbox": [ + 304, + 613, + 545, + 660 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 \\sim q(\\mathbf{x}_0)" + }, + { + "bbox": [ + 304, + 613, + 545, + 660 + ], + "type": "text", + "content": ", noise is systematically added to " + }, + { + "bbox": [ + 304, + 613, + 545, + 660 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 304, + 613, + 545, + 660 + ], + "type": "text", + "content": " in the forward process [19] step-by-step following the Markov's chain assumption until the Gaussian distribution is reached:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 351, + 662, + 545, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 662, + 545, + 694 + ], + "spans": [ + { + "bbox": [ + 351, + 662, + 545, + 694 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {0: T}\\right) = q \\left(\\mathbf {x} _ {0}\\right) \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right), \\tag {5}", + "image_path": "6e8b324bc522922f1b14c30da37c72f179b8903e2a29a6bb3fc8eef1c8c082a3.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 339, + 696, + 512, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 696, + 512, + 717 + ], + "spans": [ + { + "bbox": [ + 339, + 696, + 512, + 717 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right),", + "image_path": "e2d4a5c4f78f9042c77d0baacacda4470f03f29a9a43f77c64d29d2e136377a6.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8932" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 75, + 287, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 75, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 46, + 75, + 287, + 146 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 75, + 287, + 146 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu, \\sigma^2)" + }, + { + "bbox": [ + 46, + 75, + 287, + 146 + ], + "type": "text", + "content": " denotes a Gaussian distribution. Here, " + }, + { + "bbox": [ + 46, + 75, + 287, + 146 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 75, + 287, + 146 + ], + "type": "text", + "content": " can represent either coordinates or colors from 3D shapes. We train geometry and appearance diffusion separately to generate colored point clouds. The reverse process, " + }, + { + "bbox": [ + 46, + 75, + 287, + 146 + ], + "type": "inline_equation", + "content": "p_\\theta(\\mathbf{x}_t \\mid \\mathbf{x}_{t+1})" + }, + { + "bbox": [ + 46, + 75, + 287, + 146 + ], + "type": "text", + "content": ", starts from a standard Gaussian prior and ends with the desired " + }, + { + "bbox": [ + 46, + 75, + 287, + 146 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 75, + 287, + 146 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 86, + 144, + 287, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 144, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 86, + 144, + 287, + 174 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {x} _ {0: T}\\right) = p \\left(\\mathbf {x} _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right), \\tag {6}", + "image_path": "fed0160ba6957e970e1c982bec50e1dfbe9702c4cff4f6d0398102ce3e87c69d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 88, + 177, + 244, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 177, + 244, + 190 + ], + "spans": [ + { + "bbox": [ + 88, + 177, + 244, + 190 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) = \\mathcal {N} \\left(\\mu_ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right).", + "image_path": "dcfa459229e682bad288af8d9ad5db9b8aabb24ac0fb7efb23cf5ef07cbbcd31.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 192, + 287, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 287, + 250 + ], + "type": "text", + "content": "This cross-entropy minimization is designed such that it forces the reverse process to follow the pattern of noise addition of the forward process, and eventually, recover " + }, + { + "bbox": [ + 46, + 192, + 287, + 250 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 192, + 287, + 250 + ], + "type": "text", + "content": " from the standard Gaussian noise distribution. Hence, the loss function can be derived as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 251, + 287, + 265 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 251, + 287, + 265 + ], + "spans": [ + { + "bbox": [ + 82, + 251, + 287, + 265 + ], + "type": "interline_equation", + "content": "L _ {t} = \\mathbb {E} _ {\\mathbf {x} _ {0}, e} [ C \\| \\tilde {\\mu} _ {t} (\\mathbf {x} _ {t}, \\mathbf {x} _ {0}) - \\mu_ {t} (\\mathbf {x} _ {t}, t) \\| ^ {2} ], \\tag {7}", + "image_path": "81dc2ecb5e3a08b204ce3366f08a5f6fefe2420b9f2c9e1a9bacab21fa4ecce0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 266, + 139, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 139, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 139, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 266, + 139, + 277 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 266, + 139, + 277 + ], + "type": "text", + "content": " is a constant." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "content": "Geometry Diffusion: Denoting the colored point cloud as " + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0\\in \\mathbb{R}^{N\\times 6} = \\{\\mathbf{g}_0\\in \\mathbb{R}^{N\\times 3},\\mathbf{a}_0\\in \\mathbb{R}^{N\\times 3}\\}" + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "content": ", consisting " + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "content": " points with geometry " + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_0" + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "content": " and appearance " + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_0" + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "content": " coordinates. We first perform geometry diffusion to generate the 3D shape. STPD is trained by minimizing the cross-entropy loss between two diffusion chains with respect to the geometric conditional feature " + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "inline_equation", + "content": "C_g" + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "content": ", and learning of model parameters " + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 374, + 287, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 287, + 404 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {g _ {0} \\sim q (g _ {0}), g _ {1: T} \\sim q (g _ {1: T})} \\left[ \\sum_ {t = 1} ^ {T} \\log p _ {\\theta} \\left(\\mathbf {g} _ {t - 1} \\mid \\mathbf {g} _ {t}, C _ {g}\\right) \\right], \\tag {8}", + "image_path": "d2d7a62e7038d0f3dd4c454233ec7451ce9bd218e2346305568d90aef27bdadd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 404, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 287, + 489 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 404, + 287, + 489 + ], + "type": "inline_equation", + "content": "C_g" + }, + { + "bbox": [ + 46, + 404, + 287, + 489 + ], + "type": "text", + "content": " is derived from sketch-text embedding when the text also contains shape description, or only from sketch when there is no text. We add a switch to adjust the feature map selection for different scenarios. The joint posterior " + }, + { + "bbox": [ + 46, + 404, + 287, + 489 + ], + "type": "inline_equation", + "content": "q(g_{1:T})" + }, + { + "bbox": [ + 46, + 404, + 287, + 489 + ], + "type": "text", + "content": " can be factorized by the assumption of Markov's chain " + }, + { + "bbox": [ + 46, + 404, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\prod_{t=1}^{T} q(\\mathbf{g}_{t-1}|\\mathbf{g}_t,\\mathbf{g}_0)" + }, + { + "bbox": [ + 46, + 404, + 287, + 489 + ], + "type": "text", + "content": ", where each is analytically tractable by a re-parameterized Gaussian distribution:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 496, + 287, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 496, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 77, + 496, + 287, + 521 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {\\mu}} _ {t} = \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1})}{1 - \\bar {\\alpha} _ {t}} \\mathbf {x} _ {t} + \\frac {\\sqrt {\\bar {\\alpha} _ {t - 1}} \\beta_ {t}}{1 - \\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}, \\tag {9}", + "image_path": "a4eefdc2cf1baa952a2a21f39ac7402bdc80130d5bb923653e8a5ac7334594fd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 85, + 525, + 287, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 525, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 85, + 525, + 287, + 545 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {g} _ {t - 1} \\mid \\mathbf {g} _ {t}, \\mathbf {g} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {g} _ {t - 1}; \\tilde {\\boldsymbol {\\mu}} _ {t}, \\tilde {\\boldsymbol {\\beta}} _ {t} \\mathbf {I}\\right), \\tag {10}", + "image_path": "9f1ba7c5f572af82fe0d67b0b4d7537864e9591e2b2f3b845d68b278a473379a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\tilde{\\beta}_t" + }, + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "text", + "content": " is the known variance defined in the forward process, " + }, + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\tilde{\\alpha} = \\prod_{s = 1}^{t}\\alpha_{s}" + }, + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "text", + "content": ". From Eq. 7 and this property, our diffusion model learns to match each " + }, + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "inline_equation", + "content": "q\\left(\\mathbf{g}_{t - 1}\\mid \\mathbf{g}_t,\\mathbf{g}_0\\right)" + }, + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "inline_equation", + "content": "p_{\\theta}\\left(\\mathbf{g}_{t - 1}\\mid \\mathbf{g}_t\\right)" + }, + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "text", + "content": " by estimating the output noise " + }, + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{g}_t,C_g,t)" + }, + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "text", + "content": " to match noise " + }, + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 97, + 612, + 287, + 627 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 612, + 287, + 627 + ], + "spans": [ + { + "bbox": [ + 97, + 612, + 287, + 627 + ], + "type": "interline_equation", + "content": "\\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\mathbf {g} _ {t}, C _ {g}, t\\right) \\right\\| ^ {2}, \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}), \\tag {11}", + "image_path": "818552c89ba6d4ac085af2b28e51b44d33e2d2bd9726b1116a3bc4235e40236b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 632, + 287, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 632, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 46, + 632, + 287, + 655 + ], + "type": "text", + "content": "The noise prediction network " + }, + { + "bbox": [ + 46, + 632, + 287, + 655 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 632, + 287, + 655 + ], + "type": "text", + "content": " progressively generates the 3D shape using the reverse chain" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 58, + 657, + 275, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 657, + 275, + 685 + ], + "spans": [ + { + "bbox": [ + 58, + 657, + 275, + 685 + ], + "type": "interline_equation", + "content": "g _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {g} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (g _ {t}, C _ {g}, t)\\right) + \\sqrt {\\beta_ {t}} \\epsilon .", + "image_path": "7867b76b6a4dfa48613fa5ebca1a9964bfdc8437efde5b0926d6522ab03c91de.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Appearance Diffusion: Appearance diffusion is performed while freezing the geometry parameters. The optimization" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "function, in this case, can be simply modified from Eq. 8 to learn a conditional generative model for appearance:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 96, + 548, + 128 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 96, + 548, + 128 + ], + "spans": [ + { + "bbox": [ + 306, + 96, + 548, + 128 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {a _ {0} \\sim q (a _ {0}), a _ {1: T} \\sim q (a _ {1: T})} [ \\sum_ {t = 1} ^ {T} \\log p _ {\\theta} (\\mathbf {a} _ {t - 1} | \\mathbf {a} _ {t}, p _ {\\theta} (\\mathbf {g} _ {0}), C _ {a}) ]", + "image_path": "80f4e3b5664cd1cafbfcf896b6286ea9c71f2473cd208ba6961fa4909acceec8.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 129, + 545, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 129, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 305, + 129, + 545, + 152 + ], + "type": "text", + "content": "The training loss for appearance diffusion is minimized and conditioned on appearance features " + }, + { + "bbox": [ + 305, + 129, + 545, + 152 + ], + "type": "inline_equation", + "content": "C_a" + }, + { + "bbox": [ + 305, + 129, + 545, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 129, + 545, + 152 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{g}_0)" + }, + { + "bbox": [ + 305, + 129, + 545, + 152 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 339, + 155, + 545, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 155, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 339, + 155, + 545, + 169 + ], + "type": "interline_equation", + "content": "\\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\mathbf {a} _ {t}, p _ {\\theta} \\left(\\mathbf {g} _ {0}\\right), C _ {a}, t\\right) \\right\\| ^ {2}, \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}). \\tag {12}", + "image_path": "6c9c6e2c5835f07dd78f622b01f4ada9b6576335f082bed5430dd8efc2945b11.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 172, + 545, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 172, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 305, + 172, + 545, + 208 + ], + "type": "text", + "content": "Colored Point Cloud Generation: We can now formally define our joint generative model for colored point cloud generation as" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 213, + 554, + 227 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 213, + 554, + 227 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 554, + 227 + ], + "type": "interline_equation", + "content": "p _ {\\theta_ {2}, \\theta_ {1}, \\phi} (\\mathbf {a} _ {0}, \\mathbf {g} _ {0}, C _ {g}, C _ {a}) = p _ {\\theta_ {2}} (\\mathbf {a} _ {0} | \\mathbf {g} _ {0}, C _ {a}) p _ {\\theta_ {1}} (\\mathbf {g} _ {0} | C _ {g}) p _ {\\phi} (C _ {g}, C _ {a}),", + "image_path": "80fe91e63a984bf219b2e819485ecbf33b08b6f183be1f961174a78af6aa5ee9.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 232, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 232, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 305, + 232, + 545, + 352 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 232, + 545, + 352 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{a}_0,\\mathbf{g}_0\\}" + }, + { + "bbox": [ + 305, + 232, + 545, + 352 + ], + "type": "text", + "content": " denote the colored point cloud " + }, + { + "bbox": [ + 305, + 232, + 545, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 305, + 232, + 545, + 352 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 232, + 545, + 352 + ], + "type": "inline_equation", + "content": "p_{\\phi}(C_g,C_a)" + }, + { + "bbox": [ + 305, + 232, + 545, + 352 + ], + "type": "text", + "content": " refers to the sketch and textual embedding, and " + }, + { + "bbox": [ + 305, + 232, + 545, + 352 + ], + "type": "inline_equation", + "content": "p_{\\theta_2}(\\mathbf{a}_0),p_{\\phi}(\\mathbf{g}_0)" + }, + { + "bbox": [ + 305, + 232, + 545, + 352 + ], + "type": "text", + "content": " are the appearance and geometry denoising processes respectively. Note that the appearance diffusion is conditioned on the geometry diffusion as well. We also tried extended diffusion to denoise from a 6D Gaussian rather than treating the shape and appearance separately. However, integrating the unrelated features causes undesirable interference in the diffusion process and consequently, STPD does not learn to produce point clouds with fine details." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 357, + 387, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 357, + 387, + 370 + ], + "spans": [ + { + "bbox": [ + 306, + 357, + 387, + 370 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 374, + 422, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 374, + 422, + 387 + ], + "spans": [ + { + "bbox": [ + 306, + 374, + 422, + 387 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 390, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 390, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 305, + 390, + 545, + 545 + ], + "type": "text", + "content": "Datasets: We use the chair, table, aeroplane, car categories of ShapeNet dataset [3], randomly split it into training-test sets at " + }, + { + "bbox": [ + 305, + 390, + 545, + 545 + ], + "type": "inline_equation", + "content": "80 - 20\\%" + }, + { + "bbox": [ + 305, + 390, + 545, + 545 + ], + "type": "text", + "content": " ratio and render a single-image from each 3D object. We apply Canny edge detection [24] on the rendered ShapeNet images to acquire their single-channel edge images. We manually remove the background so that they resemble clean hand drawn sketches. For colored point clouds, we use the chair, table categories as only they have text descriptions [4]. To generate ground-truth point clouds, each 3D shape is randomly sampled at 2048 points and then translated to zero mean and normalized in the range " + }, + { + "bbox": [ + 305, + 390, + 545, + 545 + ], + "type": "inline_equation", + "content": "[-1,1]^3" + }, + { + "bbox": [ + 305, + 390, + 545, + 545 + ], + "type": "text", + "content": " so that the Gaussian distribution assumption is satisfied during the forward pass." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": "Evaluation Metrics: Previous works [26] used Chamfer Distance (CD) and Earth Mover's Distance (EMD) to evaluate the reconstruction quality of point clouds. Based on these, existing works report Coverage (COV) and Minimum matching distance (MMD). Although these metrics have their limitations, they still serve as a good reference to measure and compare the performance of generative methods. Specifically, COV evaluates the certainty of conditioned generation whereas MMD calculates the average distance of the point clouds in the reference set and their closest neighbors in the generated set. Another metric used in numerous works to evaluate the generation quality and diversity is 1-nearest neighbor accuracy (1-NNA) which directly quantifies distribution similarity between two sets. However, since" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8933" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 68, + 545, + 163 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 545, + 163 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 545, + 163 + ], + "type": "image", + "image_path": "b71d2e6c5f1cecb2f13fe6c0778c7a93efc94cc1e575c5923eb0b31dcc6936a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 164, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 46, + 164, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 164, + 547, + 198 + ], + "type": "text", + "content": "Figure 4. Visual comparison of the generated objects. 1) Inputs include text and free-hand sketch. 2) Shape methods with no color. 3) Colored shape generation methods. 4) Our method generates both geometry and appearance using sketch and text. The appearance can be re-edited by changing the language prompt while preserving geometry." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 201, + 289, + 435 + ], + "blocks": [ + { + "bbox": [ + 48, + 201, + 289, + 435 + ], + "lines": [ + { + "bbox": [ + 48, + 201, + 289, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 201, + 289, + 435 + ], + "type": "table", + "html": "
CategoriesInputMethodsMMD (↓) CD EMDCOV (%,↑) CD EMD
Reconstruction Shape (N×3)SketchPSGN [10]8.1747.253
Pix2Mesh [51]9.7717.347
TMNet [32]9.188.11
Sketch2Mesh [17]3.0182.786
SketchesPix2Vox++ [57]5.9765.760
EvoIT [49]5.7185.307
Generative Shape (N×3)TextText2Shape [5]9.8736.348
CLIP-forge [43]4.3673.818
SketchMixNF [34]7.6896.224
DPF-Net [22]7.5316.176
Sketch2Model [65]3.1762.967
×DPM [26]3.7641.986
Sketch,textSTPD (ours)1.6351.043
Generative Shape&Color (N×6)TextText2Shape10.347.476
Point-E (text) [30]2.5232.265
×DPM-extended4.7653.256
sketchPoint-E (sketch)3.0923.117
Sketch,textSTPD (ours)1.3461.017
", + "image_path": "7da6e7cd62bbf2213604f0c6a59540222a14d4717fd11a338d02040a999905fd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 436, + 287, + 469 + ], + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 469 + ], + "type": "text", + "content": "Table 1. Comparison of 3D shape generation for the chair, table categories of ShapeNet [3]. CD is multiplied by " + }, + { + "bbox": [ + 46, + 436, + 287, + 469 + ], + "type": "inline_equation", + "content": "10^{3}" + }, + { + "bbox": [ + 46, + 436, + 287, + 469 + ], + "type": "text", + "content": " and EMD by " + }, + { + "bbox": [ + 46, + 436, + 287, + 469 + ], + "type": "inline_equation", + "content": "10^{1}" + }, + { + "bbox": [ + 46, + 436, + 287, + 469 + ], + "type": "text", + "content": ". Our method generates the best shape, with or without color." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 477, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 477, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 477, + 287, + 514 + ], + "type": "text", + "content": "our aim is to generate 3D shapes faithful to the provided sketch, we stick to the first four metrics. Where other methods do not provide color, we compare shapes only." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 514, + 288, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 288, + 658 + ], + "type": "text", + "content": "Baselines: We compare with classical as well as diffusion based methods. In the former category, we select the major sketch [65], image [34] [22], and text [43] [5] based 3D object generation methods. We also compare to reconstruction based methods including single-view [10, 51, 32, 17] and multiviews [58, 49]. In the latter category, we select 3D diffusion (DPM) [26], Point-E [30], and Point Voxel Diffusion (PVD) [66]. Point-E (sketch) results are obtained after fine-tuning the Point-E image-to-shape diffusion model on our sketch-shape data for a fair comparison. We take ten generations for all methods and use the average error from ground truth for evaluation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 663, + 275, + 676 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 663, + 275, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 663, + 275, + 676 + ], + "type": "text", + "content": "4.2. Comparisons with State-of-the-art Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Table 1 compares the point cloud generation quality of our method with existing state-of-the-art. For a fair comparison, we re-train all methods with our sketch-shape data of" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 203, + 546, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 203, + 546, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 203, + 546, + 395 + ], + "type": "text", + "content": "the chair, table categories, or their text descriptions when the method takes text only as input. Multiview methods are given sketches from 3 viewpoints. Only our method performs (single view) sketch and text guided 3D object generation. Since DPM [26] is unconditional, we generate many shapes and pick the best matches. For shape only (upper half of Table 1), the proposed STPD outperforms all methods, achieving " + }, + { + "bbox": [ + 304, + 203, + 546, + 395 + ], + "type": "inline_equation", + "content": "45.8\\%" + }, + { + "bbox": [ + 304, + 203, + 546, + 395 + ], + "type": "text", + "content": " lower CD-MMD error from the nearest reconstruction method Sketch2Mesh [17] and " + }, + { + "bbox": [ + 304, + 203, + 546, + 395 + ], + "type": "inline_equation", + "content": "48.5\\%" + }, + { + "bbox": [ + 304, + 203, + 546, + 395 + ], + "type": "text", + "content": " lower CD-MMD error from the nearest generative method Sketch2Model [65]. The multiview methods do not perform well since they are unable to establish correspondences between the sketches. Text guided generative methods have more diversity but struggle with alignment to user specifications given the inherent shape ambiguity in the text. Sketch guided methods perform better in general." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 399, + 546, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 546, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 546, + 673 + ], + "type": "text", + "content": "Table 1 also compares colored point cloud (shape and color) generation. For this, we extend DPM to a 6D Gaussian and also compare with Text2Shape [4] and Point-E [30]. The proposed STPD again outperforms all methods, achieving " + }, + { + "bbox": [ + 304, + 399, + 546, + 673 + ], + "type": "inline_equation", + "content": "46.7\\%" + }, + { + "bbox": [ + 304, + 399, + 546, + 673 + ], + "type": "text", + "content": " CD-MMD error from the nearest competitor Point-E [30]. In Fig. 4, we compare our results visually to the existing methods using hand-drawn sketches from SketchX [23]. Our method also offers to reedit the color (appearance) once the geometry is finalized (last column). Table 2 reports comparative results on the chair, table, aeroplane, car categories of ShapeNet. We choose these categories because PVD [66] provides comparisons on the chair, aeroplane, car categories and the table category is used in our previous experiment. From Table 2, we can see that STPD achieves the best performance on both evaluation metrics on all four object categories. SoftFlow [21], PointFlow [60], DPF-Net [22], DPM [26], Point-E [30], PVD [66] achieve an average Chamfer Distance of 2.464, 2.429, 2.168, 2.509, 2.182, 2.539 and an average EMD of 3.105, 3.191, 2.851, 2.553, 2.490, 2.661 respectively. Our STPD achieves 1.172, 1.261 average CD, EMD errors which are " + }, + { + "bbox": [ + 304, + 399, + 546, + 673 + ], + "type": "inline_equation", + "content": "46.3\\%" + }, + { + "bbox": [ + 304, + 399, + 546, + 673 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 399, + 546, + 673 + ], + "type": "inline_equation", + "content": "49.3\\%" + }, + { + "bbox": [ + 304, + 399, + 546, + 673 + ], + "type": "text", + "content": " lower than the nearest competitor Point-E [30]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "Human Evaluations: We also conducted human evaluations on 100 objects generated by our method and other methods. Table 3 shows the results where 30 people rate" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8934" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 71, + 286, + 171 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 286, + 171 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 286, + 171 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 286, + 171 + ], + "type": "table", + "html": "
Methodschairtableaeroplanecar
CDEMDCDEMDCDEMDCDEMD
SoftFlow [21]2.7863.2954.8155.1390.4041.1981.8502.789
PointFlow [60]2.7073.6494.8045.0820.4031.1801.8032.851
DPF-Net [22]2.7633.3203.9864.6610.5281.1051.3962.318
DPM [26]3.2012.7184.6014.2980.4321.0271.8042.167
Point-E [30]2.9012.6673.8094.2900.4121.0051.6091.997
PVD [66]3.2112.9394.7314.5270.4421.0301.7742.146
STPD (ours)1.2091.1172.1171.8970.3110.8971.0491.136
", + "image_path": "f21cab7196327195786279a418842650dd1f5a1469b549f052b782a74bd74671.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 201, + 286, + 253 + ], + "blocks": [ + { + "bbox": [ + 47, + 171, + 286, + 194 + ], + "lines": [ + { + "bbox": [ + 47, + 171, + 286, + 194 + ], + "spans": [ + { + "bbox": [ + 47, + 171, + 286, + 194 + ], + "type": "text", + "content": "Table 2. Comparison with baselines on 4 categories of the ShapeNet dataset [3]. CD is multiplied by " + }, + { + "bbox": [ + 47, + 171, + 286, + 194 + ], + "type": "inline_equation", + "content": "10^{3}" + }, + { + "bbox": [ + 47, + 171, + 286, + 194 + ], + "type": "text", + "content": " and EMD by " + }, + { + "bbox": [ + 47, + 171, + 286, + 194 + ], + "type": "inline_equation", + "content": "10^{1}" + }, + { + "bbox": [ + 47, + 171, + 286, + 194 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 201, + 286, + 253 + ], + "lines": [ + { + "bbox": [ + 49, + 201, + 286, + 253 + ], + "spans": [ + { + "bbox": [ + 49, + 201, + 286, + 253 + ], + "type": "table", + "html": "
Human rate (↑)SoftFlowPointFlowDPF-NetDPMPoint-EPVDSTPD
chair1.81.71.92.13.93.54.8
table2.11.93.53.03.42.54.6
aeroplane3.43.32.23.13.73.04.0
car2.52.43.43.53.22.94.3
", + "image_path": "cdc468c1889d3205187df328c9720a9f4dff01994c53f4dc955564b146dc4976.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 281, + 278, + 390 + ], + "blocks": [ + { + "bbox": [ + 47, + 253, + 286, + 274 + ], + "lines": [ + { + "bbox": [ + 47, + 253, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 286, + 274 + ], + "type": "text", + "content": "Table 3. Human evaluations: 100 generated objects rated on a scale of 1 to 5 by 30 people." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 281, + 278, + 390 + ], + "lines": [ + { + "bbox": [ + 57, + 281, + 278, + 390 + ], + "spans": [ + { + "bbox": [ + 57, + 281, + 278, + 390 + ], + "type": "table", + "html": "
Input Modality/Model for ConditioningCD (×103)EMD (×101)
SketchText
×BERT28.32520.457
C1×1×10.9658.658
C1×1BERT9.6518.459
ResNet50BERT6.4095.480
ResNet50 + Atten.RoutingBERT1.9351.587
C1×1 + Atten.Routing×1.6891.348
C1×1 + Atten.Routing (concatenate)BERT3.7484.127
C1×1 + Atten.Routing (our fusion)BERT1.3901.101
", + "image_path": "94d0f38edc4a9e651e530a6325687daa049674dea793e778c2080ac43bee9c14.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 286, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 286, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 286, + 474 + ], + "type": "text", + "content": "the generations on a scale of 1 to 5 (higher is better). STPD outperforms the baseline methods on all individual object categories and achieves an average rating of 4.4, compared to average rating of 3.3 by the nearest competitor Point-E." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 482, + 140, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 482, + 140, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 482, + 140, + 495 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": "We perform ablation study on the sketch and text modules to verify their contributions. Table 4 shows our results using CD and EMD metrics. Text conditioning, using BERT embeddings, alone achieves the worst performance. " + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "inline_equation", + "content": "C_{1 \\times 1}" + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": " is a simple visual encoder comprising " + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": " convolution followed by batch normalization and ReLU. Sketch conditioning, using simple " + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "inline_equation", + "content": "C_{1 \\times 1}" + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": " embeddings, reduces the errors significantly since a sketch can describe shaping much better than text. Using sketch and text conditioning with " + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "inline_equation", + "content": "C_{1 \\times 1}" + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": " with BERT embeddings, the errors do not reduce much. However, when a more sophisticated sketch encoding is performed with ResNet50 [18] and combined with BERT, the CD and EMD reduce notably. To account for the sparsity of sketches, when we add capsule network with attention routing on top of ResNet50, we see a dramatic improvement in both metrics. Our simple visual encoder " + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "inline_equation", + "content": "C_{1 \\times 1}" + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": " with attention routing further improves the performance even without using text conditioning. As seen in" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 312, + 71, + 539, + 163 + ], + "blocks": [ + { + "bbox": [ + 47, + 389, + 286, + 421 + ], + "lines": [ + { + "bbox": [ + 47, + 389, + 286, + 421 + ], + "spans": [ + { + "bbox": [ + 47, + 389, + 286, + 421 + ], + "type": "text", + "content": "Table 4. Ablation study on visual & textual modules. " + }, + { + "bbox": [ + 47, + 389, + 286, + 421 + ], + "type": "inline_equation", + "content": "C_{1 \\times 1}" + }, + { + "bbox": [ + 47, + 389, + 286, + 421 + ], + "type": "text", + "content": " is a simple visual encoder comprising " + }, + { + "bbox": [ + 47, + 389, + 286, + 421 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 47, + 389, + 286, + 421 + ], + "type": "text", + "content": " convolution followed by batch normalization and ReLU." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 71, + 539, + 163 + ], + "lines": [ + { + "bbox": [ + 312, + 71, + 539, + 163 + ], + "spans": [ + { + "bbox": [ + 312, + 71, + 539, + 163 + ], + "type": "table", + "html": "
ModelModelNet10ModelNet40ScanObjectNN
PointFlow [60]93.786.878.2
PDGN [20]94.287.378.1
3D-GAN [54]91.083.376.7
AtlasNet [15]91.986.676.9
ShapeGF [2]90.284.675.8
DPM [26]94.287.679.7
STPD94.888.184.3
", + "image_path": "d13430e58c9ff9b7355ee005b97a8fc89f00642c8f4a138d0368a0120c60165e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 345, + 163, + 506, + 172 + ], + "lines": [ + { + "bbox": [ + 345, + 163, + 506, + 172 + ], + "spans": [ + { + "bbox": [ + 345, + 163, + 506, + 172 + ], + "type": "text", + "content": "Table 5. Object classification accuracy " + }, + { + "bbox": [ + 345, + 163, + 506, + 172 + ], + "type": "inline_equation", + "content": "(\\%)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 307, + 180, + 550, + 274 + ], + "blocks": [ + { + "bbox": [ + 307, + 180, + 550, + 274 + ], + "lines": [ + { + "bbox": [ + 307, + 180, + 550, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 180, + 550, + 274 + ], + "type": "image", + "image_path": "cda21b06b74008462a46dcc22296ce71e3f4576bcad40c9b17f82d414e0f1daf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 274, + 545, + 318 + ], + "lines": [ + { + "bbox": [ + 305, + 274, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 305, + 274, + 545, + 318 + ], + "type": "text", + "content": "Figure 5. Visualization of attention score maps for a sketch with respect to capsules. The " + }, + { + "bbox": [ + 305, + 274, + 545, + 318 + ], + "type": "inline_equation", + "content": "8 \\times 16" + }, + { + "bbox": [ + 305, + 274, + 545, + 318 + ], + "type": "text", + "content": " size denotes 8 capsules of 16-dimension each. Capsule attention maps are learned from attention-routing. Dark parts denote high attention." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 327, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 327, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 545, + 410 + ], + "type": "text", + "content": "the last row of the table, the best performance is achieved when we use sketch conditioning with " + }, + { + "bbox": [ + 305, + 327, + 545, + 410 + ], + "type": "inline_equation", + "content": "C_{1 \\times 1} +" + }, + { + "bbox": [ + 305, + 327, + 545, + 410 + ], + "type": "text", + "content": " attention routing embeddings combined with text embeddings from BERT [7] using the proposed fusion method. The second last row of the table shows that a simple concatenation of the two embeddings does not work as well as our proposed fusion mechanism." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 417, + 499, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 417, + 499, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 417, + 499, + 430 + ], + "type": "text", + "content": "4.4. Representation Learning Evaluation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 431, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 545, + 609 + ], + "type": "text", + "content": "We evaluate the representation learned by our sketch embedding used to condition the diffusion model. We follow the work of MVCNN [46], ScanObjectNN [48] and convert rendered images from different viewpoints of the ModelNet dataset [55] to sketches and conduct object classification experiments. We train our STPD on the generated sketches and then take the conditioned embedded features to learn a linear SVM classifier for classification based on a single sketch embedding. The essence of the conditional diffusion model is that the conditioning should guide the diffusion to retrieve the 3D shape from Gaussian noise. This means that our conditional network should be able to represent specific object categories to aid classification. The results in Table 5 demonstrate that our architecture can extract discriminative features, that are also useful for accurate classification." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 616, + 485, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 485, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 485, + 628 + ], + "type": "text", + "content": "4.5. Effectiveness of Capsule Network" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Fig. 5 visualizes the generated attention scores from a typical sketch. More examples are given in the supplementary material along with quantitative results using Instance Scores for Sketch (ISS i.e. without attention routing) and Instance Score on Capsule (ISC i.e. with attention routing) metrics. ISS is the percentage of useful pixels divided by the total pixels of the sketch. For ISC, we compute the aver" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8935" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 68, + 529, + 201 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 529, + 201 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 529, + 201 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 529, + 201 + ], + "type": "image", + "image_path": "9c73ff89596e7aed9bf3f5347d276b016b30dee73976a0ca6ff744e471db1410.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 201, + 456, + 213 + ], + "lines": [ + { + "bbox": [ + 137, + 201, + 456, + 213 + ], + "spans": [ + { + "bbox": [ + 137, + 201, + 456, + 213 + ], + "type": "text", + "content": "Figure 6. Visualization of part segmentation results for 4 categories of ShapeNet-Parts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 221, + 287, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 221, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 287, + 245 + ], + "type": "text", + "content": "age of the highest attention score for each of the 16 capsule dimensions (column wise) and convert it to a percentage." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 246, + 290, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 290, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 290, + 426 + ], + "type": "text", + "content": "Before the attention capsule, the sketches have ISS of " + }, + { + "bbox": [ + 46, + 246, + 290, + 426 + ], + "type": "inline_equation", + "content": "1.07\\%" + }, + { + "bbox": [ + 46, + 246, + 290, + 426 + ], + "type": "text", + "content": " on average. After our capsule-attention module, ISC becomes " + }, + { + "bbox": [ + 46, + 246, + 290, + 426 + ], + "type": "inline_equation", + "content": "58.68\\%" + }, + { + "bbox": [ + 46, + 246, + 290, + 426 + ], + "type": "text", + "content": " on average for better diffusion. This demonstrates the effectiveness of the Capsule Attention Network which attends to sparse data in sketches during diffusion training. Each capsule " + }, + { + "bbox": [ + 46, + 246, + 290, + 426 + ], + "type": "inline_equation", + "content": "(1\\times 16)" + }, + { + "bbox": [ + 46, + 246, + 290, + 426 + ], + "type": "text", + "content": " denotes a learned feature, e.g. the table, from a sketch. It can be observed that given the distribution of the useful pixel in the sketch, only a few attention score values are dark i.e., attentioned. Caps number 2 and a few dimensions of Caps number 5 are given high weight to contribute to the final visual feature, while the other Caps that represent the white background of the sketch are ignored. The attention-based capsule network enables our method to effectively locate the referred objects in the sparse sketch input." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 434, + 287, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 287, + 448 + ], + "type": "text", + "content": "5. Extending STPD to Part Segmentation Task" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 455, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 287, + 503 + ], + "type": "text", + "content": "To further show the feature learning ability of our STPD model, we use it for the task of part segmentation in point clouds. For this, we use the ShapeNet-Parts [3] dataset, focusing on the four object categories of Table 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 503, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 503, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 287, + 575 + ], + "type": "text", + "content": "The part segmentation task can be extended from our colored point cloud generation model by assigning different colors to each part. We consider only two parts for the table and three parts for remaining three categories e.g. for aeroplane, the (body, wing, tail) are assigned (blue, green, red) colors respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 575, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 287, + 658 + ], + "type": "text", + "content": "In our STPD, each coordinate is generated by our sketch-text guided geometric diffusion model. However, in the part segmentation task, the shape " + }, + { + "bbox": [ + 46, + 575, + 287, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_0" + }, + { + "bbox": [ + 46, + 575, + 287, + 658 + ], + "type": "text", + "content": " is given as input to the model. Here, we directly apply the text feature as the condition " + }, + { + "bbox": [ + 46, + 575, + 287, + 658 + ], + "type": "inline_equation", + "content": "C_a" + }, + { + "bbox": [ + 46, + 575, + 287, + 658 + ], + "type": "text", + "content": " of appearance diffusion. For instance, we provide a generic text sentence \"An blue aeroplane with green wings and red tail\" for all aeroplanes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 659, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 659, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 659, + 287, + 696 + ], + "type": "text", + "content": "With the given input geometry and appearance description, we can formally define our joint generative model for point cloud segmentation as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 702, + 287, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 702, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 702, + 287, + 714 + ], + "type": "interline_equation", + "content": "p _ {\\theta , \\phi} \\left(\\mathbf {a} _ {0}, \\mathbf {g} _ {0}, C _ {a}\\right) = p _ {\\theta} \\left(\\mathbf {a} _ {0} \\mid \\mathbf {g} _ {0}, C _ {a}\\right) p _ {\\text {i n p u t}} \\left(\\mathbf {g} _ {0}\\right) p _ {\\phi} \\left(C _ {a}\\right), \\tag {13}", + "image_path": "b41ffb1386bde795163a883e3dc7f83394302030366ab8c9889f9488f601025c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 219, + 538, + 320 + ], + "blocks": [ + { + "bbox": [ + 315, + 219, + 538, + 320 + ], + "lines": [ + { + "bbox": [ + 315, + 219, + 538, + 320 + ], + "spans": [ + { + "bbox": [ + 315, + 219, + 538, + 320 + ], + "type": "table", + "html": "
ModelmIoUcmIoU1chairtableaeroplanecar
PointNet [35]82.183.589.680.683.474.9
PointNet++ [36]83.384.690.882.682.477.3
DGCNN [52]83.884.990.682.684.077.8
Point-BERT [61]84.785.691.081.584.379.8
PointMLP [27]84.684.890.384.383.580.5
KPConv [47]85.185.991.183.684.681.1
P2P [53]85.085.991.683.784.380.4
STPD85.986.792.484.185.681.3
", + "image_path": "13fc5712042b2d701e97b398423bfe97ddd1d753096563c6146bacefa597cf36.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 321, + 545, + 342 + ], + "lines": [ + { + "bbox": [ + 306, + 321, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 321, + 545, + 342 + ], + "type": "text", + "content": "Table 6. Comparative results of part segmentation on four categories of ShapeNet-Parts dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 347, + 547, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 347, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 305, + 347, + 547, + 397 + ], + "type": "text", + "content": "where the combination " + }, + { + "bbox": [ + 305, + 347, + 547, + 397 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{a}_0, \\mathbf{g}_0\\}" + }, + { + "bbox": [ + 305, + 347, + 547, + 397 + ], + "type": "text", + "content": " denotes the desired segmentation of the 3D point cloud. Note that, " + }, + { + "bbox": [ + 305, + 347, + 547, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_0" + }, + { + "bbox": [ + 305, + 347, + 547, + 397 + ], + "type": "text", + "content": " is generated as colors which are then used as segmentation labels. We apply K-means clustering for final refinement of the colors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 401, + 547, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 401, + 547, + 450 + ], + "spans": [ + { + "bbox": [ + 305, + 401, + 547, + 450 + ], + "type": "text", + "content": "Results for 3D Shape Segmentation: Fig. 6 shows some visualizations of our results and Table 6 compares our method to some existing baselines. The proposed STPD achieves the highest mIoU compared to existing methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 460, + 379, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 379, + 472 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 379, + 472 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 479, + 547, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 547, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 547, + 648 + ], + "type": "text", + "content": "We proposed a novel sketch and text guided probabilistic diffusion model for 3D object generation. Our model takes a hand drawn sketch and text as inputs to condition the reverse diffusion process. With the strong generative ability and effective guidance by the sketch-text fused features, our model is able to achieve high alignment with user specifications. Our visual encoder is designed to focus on the informative sparse parts of sketches. The sketch-text fusion part gathers informative object context from visual and textual descriptions to improve the shape and appearance of the generated 3D objects. Experimental results demonstrate that our model achieves state-of-art performance in 3D object generation and shows promising applications for 3D object classification and part segmentation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 658, + 416, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 416, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 416, + 671 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "Professor Ajmal Mian is the recipient of an Australian Research Council Future Fellowship Award (project number FT210100268) funded by the Australian Government." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8936" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] A.H. Bermano, R. Gal, Y. Alaluf, R. Mokady, Y. Nitzan, O. Tov, O. Patashnik, and D. Cohen-Or. State-of-the-art in the architecture, methods and applications of stylegan. Computer Graphics Forum, 41(2):591-611, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 190 + ], + "type": "text", + "content": "[2] Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snively, and Bharath Hariharan. Learning gradient fields for shape generation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 245 + ], + "type": "text", + "content": "[3] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 302 + ], + "type": "text", + "content": "[4] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian conference on computer vision, pages 100-116. Springer, 2018. 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 303, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 288, + 357 + ], + "type": "text", + "content": "[5] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian conference on computer vision, pages 100-116. Springer, 2018. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 423 + ], + "type": "text", + "content": "[6] Shin-I Cheng, Yu-Jie Chen, Wei-Chen Chiu, Hung-Yu Tseng, and Hsin-Ying Lee. Adaptively-realistic image generation from stroke and sketch with diffusion model. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4054-4062, January 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 424, + 288, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 467 + ], + "type": "text", + "content": "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 4, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 468, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 288, + 502 + ], + "type": "text", + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 502, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 288, + 534 + ], + "type": "text", + "content": "[9] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. arXiv preprint arXiv:1605.08803, 2016. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "type": "text", + "content": "[10] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "type": "text", + "content": "[11] Mingtao Feng, Syed Zulqarnain Gilani, Yaonan Wang, Liang Zhang, and Ajmal Mian. Relation graph network for 3d object detection in point clouds. IEEE Transactions on Image Processing, 30:92-107, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 678 + ], + "type": "text", + "content": "[12] Mingtao Feng, Haoran Hou, Liang Zhang, Yulan Guo, Hongshan Yu, Yaonan Wang, and Ajmal Mian. Exploring hierarchical spatial layout cues for 3d point cloud based scene graph prediction. IEEE Transactions on Multimedia, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[13] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph net-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "work for object grounding in point cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3722-3731, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 108, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 545, + 162 + ], + "type": "text", + "content": "[14] Zhiheng Fu, Siyu Hong, Mengyi Liu, Hamid Laga, Mohammed Bennamoun, Farid Boussaid, and Yulan Guo. Multi-stage information diffusion for joint depth and surface normal estimation. Pattern Recognition, 141:109660, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 547, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 547, + 218 + ], + "type": "text", + "content": "[15] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 216-224, 2018. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 220, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 284 + ], + "type": "text", + "content": "[16] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10696-10706, June 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "type": "text", + "content": "[17] Benoit Guillard, Edoardo Remelli, Pierre Yvernay, and Pascal Fua. Sketch2mesh: Reconstructing and editing 3d shapes from sketches. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13023-13032, October 2021. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 342, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 387 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 388, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 421 + ], + "type": "text", + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1, 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "type": "text", + "content": "[20] Le Hui, Rui Xu, Jin Xie, Jianjun Qian, and Jian Yang. Progressive point cloud deconvolution generation network. In European Conference on Computer Vision, pages 397-413. Springer, 2020. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 468, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 521 + ], + "type": "text", + "content": "[21] Hyeongju Kim, Hyeonseung Lee, Woo Hyun Kang, Joun Yeop Lee, and Nam Soo Kim. Softflow: Probabilistic framework for normalizing flow on manifolds. Advances in Neural Information Processing Systems, 33:16388-16397, 2020. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "type": "text", + "content": "[22] Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. In European Conference on Computer Vision, pages 694-710. Springer, 2020. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 568, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 601 + ], + "type": "text", + "content": "[23] Yi Li, Timothy M Hospedales, Yi-Zhe Song, and Shaogang Gong. Fine-grained sketch-based image retrieval by matching deformable part models. 2014. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 602, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 646 + ], + "type": "text", + "content": "[24] Yibo Li and Bailun Liu. Improved edge detection algorithm for canny operator. In 2022 IEEE 10th Joint International Information Technology and Artificial Intelligence Conference (ITAIC), volume 10, pages 1-5. IEEE, 2022. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 647, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 714 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8937" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[26] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2837-2845, 2021. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "type": "text", + "content": "[27] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 162, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 162, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 162, + 287, + 205 + ], + "type": "text", + "content": "[28] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Spectrum-guided multi-granularity referring video object segmentation. arXiv preprint arXiv:2307.13537, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 206, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 206, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 49, + 206, + 287, + 248 + ], + "type": "text", + "content": "[29] Yongwei Miao, Feixia Hu, Xudong Zhang, Jiazhou Chen, and Renato Pajarola. Symmsketch: Creating symmetric 3d free-form shapes from 2d sketches. Computational Visual Media, 1(1):3-16, 2015. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 251, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 251, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 49, + 251, + 287, + 293 + ], + "type": "text", + "content": "[30] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 295, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 295, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 49, + 295, + 287, + 326 + ], + "type": "text", + "content": "[31] Luke Olsen, Faramarz F Samavati, Mario Costa Sousa, and Joaquim A Jorge. Sketch-based modeling: A survey. Computers & Graphics, 33(1):85-103, 2009. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 327, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 327, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 49, + 327, + 287, + 381 + ], + "type": "text", + "content": "[32] Junyi Pan, Xiaoguang Han, Weikai Chen, Jiapeng Tang, and Kui Jia. Deep mesh reconstruction from single rgb images via topology modification networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9964-9973, 2019. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 384, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 384, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 49, + 384, + 287, + 415 + ], + "type": "text", + "content": "[33] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 416, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 416, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 49, + 416, + 287, + 470 + ], + "type": "text", + "content": "[34] Janis Postels, Mengya Liu, Riccardo Spezialetti, Luc Van Gool, and Federico Tombari. Go with the flows: Mixtures of normalizing flows for point cloud generation and reconstruction. In 2021 International Conference on 3D Vision (3DV), pages 1249-1258. IEEE, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 472, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 472, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 49, + 472, + 287, + 524 + ], + "type": "text", + "content": "[35] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 526, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 526, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 49, + 526, + 287, + 569 + ], + "type": "text", + "content": "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 571, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 571, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 49, + 571, + 287, + 635 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 637, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 637, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 637, + 287, + 712 + ], + "type": "text", + "content": "[38] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8821-8831. PMLR, 18-24 Jul 2021. 1, 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 308, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 73, + 545, + 128 + ], + "type": "text", + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 130, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 162 + ], + "type": "text", + "content": "[40] Sara Sabour, Nicholas Frosst, and Geoffrey E Hinton. Dynamic routing between capsules. Advances in neural information processing systems, 30, 2017. 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 164, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 164, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 545, + 261 + ], + "type": "text", + "content": "[41] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems, volume 35, pages 36479-36494. Curran Associates, Inc., 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 264, + 545, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 328 + ], + "type": "text", + "content": "[42] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 331, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 396 + ], + "type": "text", + "content": "[43] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "type": "text", + "content": "[44] Yuefan Shen, Changgeng Zhang, Hongbo Fu, Kun Zhou, and Youyi Zheng. Deepsketchhair: Deep sketch-based 3d hair modeling. IEEE Transactions on Visualization and Computer Graphics, 27(7):3250-3263, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "type": "text", + "content": "[45] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 501, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 553 + ], + "type": "text", + "content": "[46] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In Proceedings of the IEEE international conference on computer vision, pages 945-953, 2015. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 557, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 611 + ], + "type": "text", + "content": "[47] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "type": "text", + "content": "[48] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In International Conference on Computer Vision (ICCV), 2019. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[49] Dan Wang, Xinrui Cui, Xun Chen, Zhengxia Zou, Tianyang Shi, Septimiu Salcudean, Z. Jane Wang, and Rabab Ward. Multi-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF International Conference on Com" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8938" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "text", + "content": "puter Vision (ICCV), pages 5722-5731, October 2021. 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[50] Fang Wang, Le Kang, and Yi Li. Sketch-based 3d shape retrieval using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1875-1883, 2015. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "type": "text", + "content": "[51] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "text", + "content": "[52] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog), 38(5):1-12, 2019. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "type": "text", + "content": "[53] Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. P2p: Tuning pre-trained image models for point cloud analysis with point-to-pixel prompting. arXiv preprint arXiv:2208.02812, 2022. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 288, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 287, + 342 + ], + "type": "text", + "content": "[54] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 344, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 287, + 399 + ], + "type": "text", + "content": "[55] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 1, 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 400, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 400, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 400, + 287, + 453 + ], + "type": "text", + "content": "[56] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2256-2265, June 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 511 + ], + "type": "text", + "content": "[57] Haozhe Xie, Hongxun Yao, Xiaoshuai Sun, Shangchen Zhou, and Shengping Zhang. Pix2vox: Context-aware 3d reconstruction from single and multi-view images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2690-2698, 2019. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 513, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 513, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 513, + 287, + 567 + ], + "type": "text", + "content": "[58] Haozhe Xie, Hongxun Yao, Shengping Zhang, Shangchen Zhou, and Wenxiu Sun. Pix2vox++: Multi-scale context-aware 3d object reconstruction from single and multiple images. International Journal of Computer Vision, 128(12):2919-2935, 2020. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 633 + ], + "type": "text", + "content": "[59] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Finegrained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[60] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4541-4550, 2019. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[61] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 546, + 374 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 327, + 73, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 546, + 106 + ], + "type": "text", + "content": "transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 107, + 546, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 546, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 546, + 150 + ], + "type": "text", + "content": "[62] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 152, + 546, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 546, + 217 + ], + "type": "text", + "content": "[63] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 219, + 546, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 546, + 273 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 546, + 273 + ], + "type": "text", + "content": "[64] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan++: Realistic image synthesis with stacked generative adversarial networks. IEEE transactions on pattern analysis and machine intelligence, 41(8):1947-1962, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 275, + 546, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 275, + 546, + 328 + ], + "spans": [ + { + "bbox": [ + 308, + 275, + 546, + 328 + ], + "type": "text", + "content": "[65] Song-Hai Zhang, Yuan-Chen Guo, and Qing-Wen Gu. Sketch2model: View-aware 3d modeling from single free-hand sketches. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6012-6021, 2021. 2, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 331, + 546, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 546, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 546, + 374 + ], + "type": "text", + "content": "[66] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826–5835, 2021. 1, 3, 6, 7" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8939" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Skill Transformer_ A Monolithic Policy for Mobile Manipulation/e273608a-9944-4287-91dc-881cf0220c1e_content_list.json b/2023/Skill Transformer_ A Monolithic Policy for Mobile Manipulation/e273608a-9944-4287-91dc-881cf0220c1e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..21df0ab4b1bd8ad8f143b4dea7c9cf80765954c5 --- /dev/null +++ b/2023/Skill Transformer_ A Monolithic Policy for Mobile Manipulation/e273608a-9944-4287-91dc-881cf0220c1e_content_list.json @@ -0,0 +1,1600 @@ +[ + { + "type": "text", + "text": "Skill Transformer: A Monolithic Policy for Mobile Manipulation", + "text_level": 1, + "bbox": [ + 156, + 130, + 810, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoyu Huang \nGeorgia Tech", + "bbox": [ + 120, + 181, + 238, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dhruv Batra", + "bbox": [ + 326, + 181, + 454, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meta AI (FAIR), Georgia Tech", + "bbox": [ + 279, + 199, + 527, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Akshara Rai", + "bbox": [ + 581, + 181, + 683, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meta AI (FAIR)", + "bbox": [ + 566, + 199, + 697, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Andrew Szot", + "bbox": [ + 738, + 181, + 846, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Georgia Tech", + "bbox": [ + 738, + 199, + 846, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present Skill Transformer, an approach for solving long-horizon robotic tasks by combining conditional sequence modeling and skill modularity. Conditioned on ego-centric and proprioceptive observations of a robot, Skill Transformer is trained end-to-end to predict both a high-level skill (e.g., navigation, picking, placing), and a whole-body low-level action (e.g., base and arm motion), using a transformer architecture and demonstration trajectories that solve the full task. It retains the composability and modularity of the overall task through a skill predictor module while reasoning about low-level actions and avoiding hand-off errors, common in modular approaches. We test Skill Transformer on an embodied rearrangement benchmark and find it performs robust task planning and low-level control in new scenarios, achieving a $2.5x$ higher success rate than baselines in hard rearrangement problems.", + "bbox": [ + 75, + 300, + 473, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 587, + 209, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A long-standing goal of embodied AI is to build agents that can seamlessly operate in a home to accomplish tasks like preparing a meal, tidying up, or loading the dishwasher. To successfully carry out such tasks, an agent must be able to perform a wide range of diverse skills, like navigation and object manipulation, as well as high-level decision-making, all while operating with limited egocentric inputs. Our particular focus is the problem of embodied object rearrangement [41]. An example in Figure 1 shows that the Fetch mobile manipulator must rearrange a bowl from an initial position to a desired goal position (both circled in orange) in an unfamiliar environment without access to any pre-existing maps, 3D models, or precise object locations. Successfully completing this task requires executing both low-level motor commands for accurate arm and base movements, as well as high-level decision-making, such as determining whether a drawer needs to be opened before accessing a specific object.", + "bbox": [ + 75, + 612, + 470, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Long-horizon tasks like object rearrangement require", + "bbox": [ + 94, + 885, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "many low-level control steps and naturally break down into distinct phases of skills such as navigation, picking, placing, opening, and closing. Prior works [6, 33, 37] show that it is possible to end-to-end learn a single policy that can perform such diverse tasks and behaviors. Such a monolithic policy directly maps observations to low-level actions, reasoning both about which skill to execute and how to execute it. However, scaling end-to-end learning to long-horizon, multi-phase tasks with thousands of low-level steps and egocentric visual observations remains a challenging research question. Other works decompose the task into individual skills and then separately sequence them to solve the full task [40, 16]. While this decomposition makes the problem tractable, this modularity comes at a cost. Sequencing skills suffers from hand-off errors between skills, where one skill ends in a situation where the next skill is no longer feasible. Moreover, it does not address how to learn the skill sequencing or how to adapt the planned sequence to unknown disturbances.", + "bbox": [ + 496, + 268, + 893, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present Skill Transformer, an end-to-end trainable transformer policy for multi-skill tasks. Skill Transformer predicts both low-level actions and high-level skills, conditioned on a history of high-dimensional egocentric camera observations. By combining conditional sequence modeling with a skill prediction module, Skill Transformer can model long sequences of both high and low-level actions without breaking the problem into disconnected submodules. In Figure 1, Skill Transformer learns a policy that can perform diverse motor skills, as well as high-level reasoning over which motor skills to execute. We train Skill Transformer autoregressively with trajectories that solve the full rearrangement task. Skill Transformer is then tested in unseen settings that require a variable number of high-level and low-level actions, overcoming unknown perturbations.", + "bbox": [ + 496, + 561, + 893, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Skill Transformer achieves state-of-the-art performance in rearrangement tasks requiring variable-length high-level planning and delicate low-level actions, such as rearranging an object that can be in an open or closed receptacle, in the Habitat environment [41]. It achieves $1.45\\mathrm{x}$ higher success rates in the full task and up to $2.5\\mathrm{x}$ higher in especially challenging situations over modular and end-to-end baselines", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10852", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c585f2d9a95a0b53c9fa21e88d8ca29e6b43f55bbc2ececd28d2d27456965367.jpg", + "image_caption": [ + "Figure 1: A Fetch robot is placed in an unseen house and must rearrange an object from a specified start to a desired goal location from egocentric visual observations. The robot infers the drawer is closed and opens the drawer to access the object. Next, it picks the object, navigates to the goal, and places the object. The Skill Transformer is a monolithic \"pixels-to-actions\" policy for such long-horizon mobile manipulation problems which infers the active skill and the low-level continuous control." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 870, + 315 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and is more robust to environment disturbances, such as a drawer snapping shut after opening it, despite never seeing such disturbances during training. Furthermore, we demonstrate the importance of the design decisions in Skill Transformer and its ability to scale with larger training dataset sizes. The code is available at https://bit.ly/3qH2QQK.", + "bbox": [ + 75, + 402, + 470, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 510, + 218, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object Rearrangement: Object rearrangement can be formulated as a task and motion planning [15] problem where an agent must move objects to a desired configuration, potentially under some constraints. We refer to [4] for a comprehensive survey. Prior works have also used end-to-end RL training for learning rearrangement [44, 32, 14, 10], but unlike our work, they rely on abstract interactions with the environment, no physics, or simplified action spaces. Other works focus on tabletop rearrangement [28, 21, 12] while our work focuses on longer horizon mobile manipulation. Although mobile manipulation has witnessed significant achievements in robotics learning [29, 43, 26, 30, 19], due to the vastly increased complexity, in rearrangement tasks, a majority of previous works [6, 40] explicitly separate the base and arm movements to simplify the task. A few works [16, 47, 45] have shown the superiority of mobile manipulation over static manipulation. In the Habitat Rearrangement Challenge [40, 41], methods that decompose the rearrangement task into skills are the most successful [16, 45] but they require a fixed high-level plan and cannot dynamically select skills based on observations during execution.", + "bbox": [ + 75, + 537, + 468, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformers for Control: Prior work uses a transformer policy architecture to approach RL as a sequence", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "modeling problem [8, 49, 20, 7]. Our work also uses a transformer-based policy, but with a modified architecture that predicts high-level skills along with low-level actions. Other works demonstrate that transformer-based policies are competent multi-task learners [23, 48, 38, 6, 33]. Likewise, we examine how a transformer-based policy can learn multiple skills and compose them to solve long-horizon rearrangement tasks. [27, 5] further show that transformers can encode underlying dynamics of the environment that benefits learning for individual skills.", + "bbox": [ + 496, + 402, + 893, + 554 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Skill Sequencing: Other works investigate how to sequence together skills to accomplish a long horizon task. By utilizing skills, agents can reason over longer time horizons by operating in the lifted action space of skills rather than over low-level actions [39]. Skills may be either learned from demonstrations [13, 2, 17], from trial and error learning [36, 11], or hand defined [25]. This work demonstrates that a single transformer policy can learn multiple skills. However, when sequencing together skills, there may be \"hand-off\" issues when skills fail to transition smoothly. Prior works tackle this issue by learning to match the initiation and termination sets of skills [24, 9] or learning new skills to handle the chaining [3, 22]. Decision Diffuser [1] learns a generative model to combine skills, but requires specifying the skill combination order. We overcome these issues by distilling all skills into a single policy and training this monolithic policy end-to-end to solve the overall task while maintaining skill modularity through a skill-prediction module.", + "bbox": [ + 496, + 613, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10853", + "bbox": [ + 480, + 924, + 517, + 936 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Task", + "text_level": 1, + "bbox": [ + 76, + 89, + 142, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We focus on the problem of embodied object rearrangement, where a robot is spawned in an unseen house and must rearrange an object from a specified start position to a desired goal position entirely from on-board sensing and without any pre-built maps, object 3D models, or other privileged information. We follow the setup of the 2022 Habitat Rearrangement Challenge [41]. Specifically, a simulated Fetch robot [34] senses the world through a $256 \\times 256$ head-mounted RGB-D camera, robot joint positions, whether the robot is holding an object or not, and base egomotion giving the relative position of the robot since the start of the episode. The 3D position of the object to rearrange, along with the 3D position of where to rearrange the object to is also provided to the agent.", + "bbox": [ + 75, + 114, + 468, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The agent interacts with the world via a mobile base, 7DoF arm, and suction grip attached to the arm. The episode is successful if the agent calls an episode termination action when the object is within $15\\mathrm{cm}$ of its target. The agent fails if it excessively collides with the environment or fails to rearrange the object within the prescribed low-level time steps. See Appendix B for more task details.", + "bbox": [ + 75, + 327, + 468, + 433 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluate our method in the Rearrange-Hard Task, which presents challenges in both low-level actions and high-level reasoning. Unlike the Habitat Rearrangement tasks [16, 40] that prior works focus on, this task involves objects and goals that may spawn in closed receptacles. Therefore, the agent may need to perform intermediate actions, such as opening a drawer or fridge to access the target object or goal. Furthermore, the correct sequence of skills required to solve the task differs across episodes and is unavailable to the agent. As a result, the agent must infer the correct behaviors from egocentric visual observations and proprioception. This more natural setting better reflects real-world rearrangement scenarios, thus making it the focus of this work.", + "bbox": [ + 75, + 434, + 470, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From the dataset provided in [41], we create a harder version of the full Rearrange track dataset for evaluation. In this harder version, more episodes have objects and goals contained in closed receptacles. Specifically, the updated episode distribution in evaluation for the rearrange-hard task is divided into three splits:", + "bbox": [ + 75, + 646, + 468, + 736 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $25\\%$ easy scenes with objects initially accessible.", + "- $50\\%$ hard scenes with objects in closed receptacles.", + "- $25\\%$ very hard scenes with both objects and goals in closed receptacles." + ], + "bbox": [ + 76, + 737, + 467, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The original rearrange-hard dataset is used for training.", + "bbox": [ + 76, + 797, + 441, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 829, + 166, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose Skill Transformer, an approach for learning long-horizon, complex tasks that require sequencing multiple distinct skills. In Section 4.1, we describe the com", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ponents of Skill Transformer, which include a Skill Inference module for reasoning over skill selections and an Action Inference module to predict actions for controlling the robot, both modeled using a Transformer architecture and conditioned on past and current observations. Next, in Section 4.2, we detail how we train Skill Transformer in an autoregressive fashion from demonstrations of the full task.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Skill Transformer Architecture", + "text_level": 1, + "bbox": [ + 498, + 209, + 774, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Skill Transformer breaks the long-horizon task of object rearrangement into first inferring skill selections using a Skill Inference module and then predicting low-level control using an Action Inference module, given this skill choice. Both modules are implemented as causal transformer neural networks [31], conditioned on past and current observations of the robot.", + "bbox": [ + 496, + 233, + 890, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Input: Both modules in Skill Transformer take as input a sequence of tokens that encode robot observations. We refer to the maximum length of the input sequence as the context length $C$ . Let $o_{t}$ denote the $256 \\times 256 \\times 1$ depth visual observations, and $p_{t}$ denote the proprioceptive inputs of the joint angles and holding state at time $t$ . The visual observation $o_{t}$ is first encoded with a Resnet-18 encoder [18] to a 256-dimension visual embedding. Likewise, non-visual inputs are encoded with a multi-layer perceptron network into a 256-dimension sensor embedding. The visual and non-visual continuous embeddings are concatenated to form a 512-dimension input token $x_{t}$ . We also add a per-time step positional embedding to each input token $x_{t}$ to enable modeling time across the sequence. $x_{t}$ forms the input token that encodes the current observations, and Skill Transformer takes as input a history of such tokens up to context length $C$ . Note that we do not use RGB input as prior work [40] found it unnecessary for geometric object rearrangement, and training with only depth input is faster.", + "bbox": [ + 496, + 339, + 892, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Skill Inference: The first module of Skill Transformer is the Skill Inference module which predicts the current skill to be executed. The Skill Inference module takes as input a sequence of the input embeddings $\\{x_{t - C},\\ldots ,x_t\\}$ through a causal transformer network [42] with a 512 hidden dimension, to produce a one-hot skill vector $c_{t}$ corresponding to the predicted skill that should be active at the current time step $t$ . Using the causal mask, label $c_{i}$ only has access to the previous tokens $\\{t:t < i\\}$ in the input sequence. Then, we embed $c_{i}$ into a latent skill embedding $\\{z_{t - C},\\dots ,z_t\\}$ .", + "bbox": [ + 496, + 627, + 890, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Action Inference: Next, the Action Inference module outputs a distribution over robot low-level actions. The Action Inference module takes as input the input embedding $x_{t}$ , an embedding of the previous time step action $a_{t-1}$ , and the skill embedding $z_{t}$ . Each of these inputs are represented as distinct tokens in the input sequence of $\\{x_{t-C}, a_{t-C-1}, z_{t-C}, \\ldots, x_{t}, a_{t-1}, z_{t}\\}$ . This sequence is then fed into a transformer with the same architecture as the", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10854", + "bbox": [ + 480, + 925, + 519, + 936 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5c95c374819662e68a55ffd084578f8f7ef856cbd80ea6191a348ef3244adf24.jpg", + "image_caption": [ + "Figure 2: The Skill Transformer architecture. Skill Transformer learns to map egocentric depth images from the Fetch robot's head camera and proprioceptive state to actions controlling the robot, including the base movement, arm movement, and gripper state. First, the raw visual observations are processed by a visual encoder and concatenated with the proprioceptive state. Next, the Skill Inference module, implemented as a causal transformer sequence model, infers the skill ID that is executing for each time step. Finally, the Action Inference module predicts the low-level action to execute on the robot from the inferred skill, observation, and previous action information. The model is trained through end-to-end imitation learning." + ], + "image_footnote": [], + "bbox": [ + 119, + 89, + 851, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Skill Inference module. By taking as input the skill embeddings $z_{t}$ , the Action Inference module learns consistency in executing individual skills and the transitions between them. As we will show in Section 5.5, this also allows training Skill Transformer with datasets that only successfully complete a portion of the skills.", + "bbox": [ + 75, + 416, + 470, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Action Inference module then outputs a sequence of predicted next actions $\\{a_{t - C},\\dots a_t\\}$ which control the robot. These actions determine the whole body control actions composed of three components: arm joint positions, base velocities, and gripper state. The module first outputs a 12-dimension output vector. The first seven components are used as the mean of a normal distribution to sample the desired delta joint positions, which will then be actuated via a PD controller. The next two dimensions are used as the mean of a normal distribution to sample the desired base velocities, scaled by the maximum linear and angular velocities of the robot base. The remaining three outputs are passed through a softmax layer to generate probabilities for the gripper action, which indicate opening, closing, or maintaining the previous gripper state.", + "bbox": [ + 75, + 523, + 468, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unless otherwise specified, we use 2 transformer layers and a context length $C = 30$ for all architectures, giving 2M parameters for the transformer network in the Skill Inference module. The Action Inference module is also trained with the same architecture and a context length $C$ of 30 time steps, translating to a sequence length of 90 tokens. In total, between the observation encoders, Skill Inference, and Action Inference modules, the Skill Transformer policy network consists of 16M parameters.", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Skill Transformer Training", + "text_level": 1, + "bbox": [ + 498, + 416, + 746, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we describe how Skill Transformer is trained end-to-end to learn long-horizon rearrangement tasks from imitation learning. We train the policy using supervised learning on demonstration trajectories produced by a data-generating policy. We assume the data-generating policy demonstrates a composition of skills to solve the task. Such compositional approaches are the state-of-the-art for the Habitat Rearrangement task [40, 16].", + "bbox": [ + 496, + 438, + 890, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we assume the data-generating policy demonstrates one of $N$ skills, and each time step $t$ is labeled with the skill $\\mathrm{ID}y_{t}^{\\mathrm{skill}}$ that is currently executing. In our problem setting, these skills are: pick, place, navigate, open drawer, and open fridge. We also record the policy's actions at time step $t$ that consists of arm action $y_{t}^{\\mathrm{a}}$ , base velocity $y_{t}^{\\mathrm{b}}$ , and gripper state $y_{t}^{\\mathrm{g}}$ . In addition, we acquire from the simulator a one-dimensional indicator label $y_{t}^{\\mathrm{pred}}$ defined as $y_{t}^{\\mathrm{pred}} = 0$ if the object does not start in a closed receptacle and $y_{t}^{\\mathrm{pred}} = i$ if $i^{\\mathrm{th}}$ receptacle contains the object, at the time step $t$ .", + "bbox": [ + 496, + 560, + 892, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.1 Training Loss", + "text_level": 1, + "bbox": [ + 500, + 748, + 651, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We separately supervise the Skill Inference and Action Inference modules. We supervise the Skill Inference module over the context length $C$ using loss function:", + "bbox": [ + 496, + 772, + 890, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s k i l l}} = \\sum_ {t = 1} ^ {C} \\mathcal {L} _ {\\mathrm {F L}} \\left(c _ {t}, y _ {t} ^ {\\text {s k i l l}}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 816, + 890, + 856 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{FL}}$ represents the focal loss. The focal loss counteracts the imbalance of skill execution frequencies. For example, the navigation skill executes more frequently than", + "bbox": [ + 496, + 854, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10855", + "bbox": [ + 480, + 924, + 517, + 936 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the other skills as it is needed multiple times within a single episode, and takes more low-level time steps to execute than manipulation skills.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to train the Action Inference module in parallel with the Skill Inference module, we use teacher-forcing, and input the ground truth skill labels during training, instead of the inferred distributions that are used at test time. We supervise the Action Inference module over the context length $C$ with loss function:", + "bbox": [ + 75, + 137, + 467, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a c t}} = \\sum_ {t = 1} ^ {C} \\left(\\mathcal {L} _ {\\mathrm {M S E}} \\left(a _ {t} ^ {\\mathrm {b}}, y _ {t} ^ {\\mathrm {b}}\\right) + \\mathcal {L} _ {\\mathrm {M S E}} \\left(a _ {t} ^ {\\mathrm {a}}, y _ {t} ^ {\\mathrm {a}}\\right) + \\mathcal {L} _ {\\mathrm {F L}} \\left(a _ {t} ^ {\\mathrm {g}}, y _ {t} ^ {\\mathrm {g}}\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 227, + 467, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{MSE}}$ represents the mean-squared error loss.", + "bbox": [ + 76, + 266, + 419, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, we introduce auxiliary heads and losses for the Skill Inference module. The auxiliary head reconstructs oracle information unavailable to the policy. Specifically, assuming there exist $M$ receptacles that may contain the target object, the auxiliary head outputs an $M + 1$ dimensional distribution $v_{t}$ . We then supervise $v_{t}$ with $y_{t}^{\\mathrm{pred}}$ over the context length $C$ via the loss function:", + "bbox": [ + 75, + 281, + 467, + 385 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a u x}} = \\sum_ {t = 1} ^ {C} \\mathcal {L} _ {\\mathrm {C E}} \\left(v _ {t}, y _ {t} ^ {\\text {p r e d}}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 386, + 467, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{CE}}$ represents the Cross-Entropy loss.", + "bbox": [ + 76, + 426, + 377, + 440 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We incorporate this auxiliary loss because the data-generating policy used to generate the demonstrations has a low success rate in correctly opening the receptacles. This low success rate can hinder learning because Skill Transformer needs to reason if it failed due to not opening the receptacle at all or opening an incorrect receptacle. The auxiliary loss incentivizes Skill Transformer to identify the key features that lead to the decision to open specific receptacles and to avoid wrong ones. This auxiliary objective with privileged information is only used for training and not used at test time.", + "bbox": [ + 75, + 441, + 467, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final loss is the unweighted sum of individual loss terms: $\\mathcal{L} = \\mathcal{L}_{\\mathrm{skill}} + \\mathcal{L}_{\\mathrm{action}} + \\mathcal{L}_{\\mathrm{aux}}$ . Note that in training, we calculate the loss over all time steps, whereas at test time, we only output action $a_{t}$ for the current time step $t$ to the environment. While training, we also train the visual encoder, in contrast to previous works using transformers in embodied AI that only demonstrate performance with pre-trained visual encoders [8, 6]. Since Skill Transformer is an end-to-end trainable policy, it can also be trained with offline-RL using the rewards from the dataset or with online-RL by interacting with the simulator. We leave exploring Skill Transformer with RL training for future work.", + "bbox": [ + 75, + 608, + 467, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2 Training Scheme", + "text_level": 1, + "bbox": [ + 76, + 814, + 248, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use a similar scheme as [8], where we input a sequence of 30 timesteps and compute the loss for the skill and action distributions for all timesteps in the sequence using autoregressive masking. Skill Transformer uses a learning warm-", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "up phase where the learning rate is linearly increased at the start of training. After the warm-up phase, Skill Transformer uses a cosine learning rate scheduler which decays to 0 at the end of training. The entire dataset was iterated through approximately 6 times in total during the complete training process. We report the performance of the checkpoint with the highest success rate on seen scenes. We provide further details about Skill Transformer in Appendix A.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 224, + 630, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we describe a series of experiments used to validate our proposed approach. First, we introduce the experiment setup and state-of-the-art baselines for comparison. Then, we quantify the performance of Skill Transformer in the rearrange-hard task and compare it to the baselines. Furthermore, the robustness of the methods to disturbances at evaluation time are evaluated by adversarial perturbations, like closing drawers after the agent has opened them. Finally, we perform an ablation study to confirm the key design choices involved in Skill Transformer.", + "bbox": [ + 496, + 250, + 890, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 500, + 409, + 674, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct all experiments in the Habitat 2.0 [40] simulation environment and the rearrangement tasks as described in Section 3. For training, we use the train dataset split from the challenge consisting of 50,000 object configurations across 63 scenes. Each episode consists of starting object positions and a target location for the object to be rearranged. The robot positions are randomly generated in each episode. We first train individual skills (pick, place, navigate, open, close) that are used by a data generating policy to generate 10,000 demonstrations for training Skill Transformer. To train skills, we use the Multi-skill Mobile Manipulation (M3) [16] method with an oracle task planner with privileged information about the environment. This approach trains skills with RL and sequences them with an oracle planner. Further details of the data generating policy and the demonstration collected are detailed in Appendix C.1.", + "bbox": [ + 496, + 431, + 890, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Baselines", + "text_level": 1, + "bbox": [ + 500, + 696, + 607, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare Skill Transformer (ST) to the following baselines in the rearrange-hard task. The parameter count of each baseline policy is adjusted to match the 16M parameters in Skill Transformer.", + "bbox": [ + 496, + 719, + 890, + 777 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Monolithic RL (Mono): Train a monolithic neural network directly mapping sensor inputs to actions with end-to-end RL using DD-PPO [46], a distributed form of PPO [35]. The monolithic RL policy is trained for 100M timesteps, reflecting a wall time of 21h.", + "- Decision Transformer (DT) [8]: Train the monolithic policy from the offline dataset using a decision transformer architecture. The DT policy is trained with the" + ], + "bbox": [ + 500, + 780, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10856", + "bbox": [ + 480, + 926, + 517, + 936 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "same learning rate scheme and number of epochs as Skill Transformer. At test time, we condition the DT policy with the highest possible reward in the dataset.", + "bbox": [ + 89, + 90, + 467, + 135 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Decision Transformer with Auxiliary Skill Loss (DT (Skill)): Train the Decision Transformer with an auxiliary loss of skill prediction. This baseline incorporates an auxiliary head that predicts the current skill during training and is trained with focal loss like Skill Transformer.", + "bbox": [ + 76, + 136, + 467, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that Decision Transformer requires the environment reward at evaluation time to compute the reward-to-go input. Skill Transformer and the other baselines do not require this assumption. We also compare to the following hierarchical baselines which first decompose the task into skills and then sequence these skills to solve the tasks.", + "bbox": [ + 76, + 212, + 467, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Task-Planning + Skills RL (TP-SRL) [40]: Individually train pick, place, navigate, and open skills using RL and then sequence the skills together with a fixed task planner. Each skill predicts when it should end and the next skill should begin by a skill termination action. TP-SRL trains all manipulation skills with a fixed base.", + "bbox": [ + 76, + 303, + 467, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Multi-skill Mobile Manipulation (M3) [16]: M3 is a modular approach similar to TP-SRL, but instead trains skills as mobile manipulation skills with RL. All these skills are trained to be robust to diverse starting states. Like TP-SRL, it uses a fixed task planner to sequence the skills.", + "bbox": [ + 76, + 393, + 467, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Behavior Cloning Modular skills (BC-Modular): Individually train transformer-based policies for each of the skills using behavior cloning with the same training dataset as Skill Transformer. Then sequence together with the same fixed task planner as the other modular baselines. This baseline helps measure the gap in performance from training policies with behavior cloning as opposed to the RL-trained policies, M3 and TP-SRL.", + "bbox": [ + 76, + 470, + 467, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, these modular baselines are designed to only execute a fixed pre-defined sequence of skills (navigate, pick, navigate, place). They cannot adjust their skill sequence based on environment observations, i.e. the task plan is inaccurate when the robot needs to open a receptacle to pick or place an object.", + "bbox": [ + 76, + 590, + 467, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address this, we additionally include a version of the modular baselines, M3 (Oracle) and BC-Modular (Oracle), that extracts the correct task plan directly from the simulator, which we refer to as an oracle plan. In contrast, the proposed method, Skill Transformer, infers which skill to execute from egocentric observations without using this privileged oracle plan. Note that neither modular baseline re-executes skills if they fail, while Skill Transformer can re-plan. For further details about baselines and all hyperparameters, see Appendix C.2", + "bbox": [ + 76, + 683, + 467, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Rearrange-Hard", + "text_level": 1, + "bbox": [ + 76, + 845, + 241, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, we show the performance of Skill Transformer on the rearrange-hard task described in Section 3. In this task,", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the agent must rearrange a single object from the start to the goal, and the object or the goal may be located in a closed receptacle. Therefore, methods need high-level reasoning to decide if the receptacle should first be opened or not.", + "bbox": [ + 498, + 90, + 890, + 150 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the success rate of all methods on both \"Train\" (Seen) and \"Eval\" (Unseen) scenes and object configurations. The success rate is measured as the average number of episodes where the agent placed the target object within $15\\mathrm{cm}$ of the goal within 5000 steps and without excessively colliding with the environment. Specifically, the episode will terminate with failure if the agent accumulates more than $100\\mathrm{kN}$ of force throughout the episode or $10\\mathrm{kN}$ of instantaneous force due to collisions with the scene or objects. We also report the success rate on each of the three dataset splits described in Section 3. Our evaluation dataset splits have 100 episodes in the easy split, 200 in the hard split (100 for objects starting in closed receptacles and 100 for goals in closed receptacles), and 100 in the very hard split.", + "bbox": [ + 498, + 152, + 890, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 shows that Skill Transformer achieves higher success rates in both the \"Train\" and \"Eval\" settings than all of the baselines that do not have oracle planning. The Mono baseline cannot find any success, even after training for 100M steps. This demonstrates the difficulty of learning the rearrange-hard task from a reward signal alone.", + "bbox": [ + 498, + 381, + 890, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The modular TP-SRL, M3, and BC-Modular methods rely on a fixed task plan to execute the separately learned skill policies. As a result, the state-of-the-art M3 method achieves a significantly lower success rate than Skill Transformer in the overall result as well as the hard and very hard splits because it is unable to adapt the skill sequence to different situations. On the other hand, Skill Transformer does not rely on a fixed ordering of skills and can adapt its high-level task-solving to new situations, achieving a $35\\%$ increase in success rate across all episodes.", + "bbox": [ + 498, + 473, + 890, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Notice that even with an oracle plan, the success rate of M3 (Oracle) on the overall result for unseen scenes is only $3\\%$ higher than that of Skill Transformer. More importantly, Skill Transformer outperforms M3 (Oracle) in both hard and very hard settings, even without oracle task planning information. We find that Skill Transformer consistently achieves higher success rates in episodes with hidden objects or goals, while M3 (Oracle) experiences a substantial decrement as the difficulty increases. This difference is likely due to the low success rate in M3 (Oracle)'s opening skill, hindering its ability to complete the skill in one attempt. In contrast, as we will discuss quantitatively later in Section 5.4, Skill Transformer is capable of detecting and retrying failed skills, eventually leading to better performance in retrieving and placing the object.", + "bbox": [ + 498, + 627, + 890, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Further observe that BC-Modular exhibits the lowest performance among modular methods, with an overall success rate of only $5\\%$ even when employing oracle task", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10857", + "bbox": [ + 480, + 925, + 517, + 936 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/172f9ffce372c303b3b19667ad66d11865dfd649fe8dce75e538805d1a7d4de8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTrain (Seen)Eval (Unseen)
All EpisodesEasyHardVery HardAll EpisodesEasyHardVery Hard
BC-Modular (Oracle)5.67±0.1220.33±1.700.83±0.240.00±0.005.00±0.3515.33±1.701.67±1.650.33±0.47
M3 (Oracle)22.92±0.7259.33±1.7013.00±1.416.33±0.4722.58±1.1858.67±1.8912.67±2.366.33±0.47
BC-Modular5.08±0.4220.33±1.700.00±0.000.00±0.003.83±0.4215.33±1.700.00±0.000.00±0.00
TP-SRL7.25±0.8229.00±3.270.00±0.000.00±0.006.67±0.8526.67±3.400.00±0.000.00±0.00
M314.00±0.3555.67±0.940.16±0.470.00±0.0013.25±1.0853.00±4.320.00±0.000.00±0.00
Mono0.00±0.000.00±0.000.00±0.000.00±0.000.00±0.000.00±0.000.00±0.000.00±0.00
DT7.08±0.6319.33±0.473.83±0.622.00±0.946.50±0.4115.33±1.253.83±0.623.00±1.63
DT (Skill)8.17±0.6723.33±2.493.00±2.163.33±0.947.67±0.6618.67±2.624.83±1.702.33±1.25
ST (Ours)21.08±1.3644.00±0.8215.67±1.709.00±2.1619.17±1.0537.00±0.8215.83±0.858.00±3.27
", + "bbox": [ + 102, + 88, + 864, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1: Success rates on the rearrangement task. \"Eval\" results are from test episodes in unseen scenes. In \"Easy\", objects and goals start on accessible locations. In \"Hard\", objects start in closed receptacles. In \"Very Hard\" objects and goals start in closed receptacles. All results are averages across 100 episodes except for the \"Hard\" split which is an average across 200 episodes. Grayed methods at the top use the oracle skill order. We report the mean and standard deviation across 3 random training seeds. For hierarchical methods, we retrain all the skills for each random seed.", + "bbox": [ + 75, + 255, + 893, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "planning. This indicates that policies trained via Behavior Cloning, which uses only one-tenth the experience of those trained with RL, encounter greater difficulty in executing intricate mobile manipulation skills. Conversely, Skill Transformer, featuring the multi-skill Action Inference module, mitigates this limitation and achieves up to 4x performance improvements across all episodes.", + "bbox": [ + 75, + 358, + 470, + 464 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DT and DT (Skill) can also achieve some success in hard and very hard splits. Nonetheless, the success rates are significantly lower than Skill Transformer, especially in the very hard split. We observe that DT-based methods lack the capability to execute skills consistently and often mix up different skill behaviors. Also, notice that DT (Skill) with an auxiliary skill loss does not show a significant improvement in performance compared to DT without the skill loss. These observations highlight the superiority of directly conditioning on skills rather than as auxiliary rewards in learning tasks with clearly distinguished skills, such as rearrangement.", + "bbox": [ + 75, + 465, + 470, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Robustness to Perturbations", + "text_level": 1, + "bbox": [ + 76, + 662, + 331, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct a perturbation test to evaluate the replanning ability of Skill Transformer and baselines in the rearrange-hard task. The test perturbs the environment by closing the target receptacle after the agent had opened it for the first time. To succeed, the agent must detect the drawer has been closed and adjust the current plan to reopen the receptacle and retrieve the object. We compare Skill Transformer's skill prediction and M3's task planner against ground truth skill labels derived from the privileged states of the simulation environment (detailed in Appendix C.3). The test is performed on 400 episodes that involved only objects in the closed drawer. This test compares against the highest-performing modular baseline with access to oracle task plans, M3 (Oracle), and the highest", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/07b647e2f17b36c424d08f50fb8d81870e8ac20456e4de08cdec2279c74edd56.jpg", + "table_caption": [ + "performing end-to-end baseline, DT (Skill)." + ], + "table_footnote": [], + "table_body": "
MethodPicking Target ObjectPredicted Skill Accuracy
No PerturbPerturbNo PerturbPerturb
M3 (Oracle)24.000.5044.0010.00
DT (Skill)10.008.5039.0033.99
ST (Ours)26.7519.5060.0047.00
", + "bbox": [ + 504, + 383, + 883, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2: Success rates for picking up the target objects and accuracy of Skill Inference module and M3's oracle-plant task planner under both perturbed and unperturbed environment, measured by ground truth skill labels.", + "bbox": [ + 496, + 474, + 890, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The findings in Table 2 demonstrate that Skill Transformer outperforms the M3 (Oracle) baseline both in the success rate of picking up the object and skill selection accuracy. Specifically, when assessing the success rates of object retrieval, M3 (Oracle) performs comparably to Skill Transformer in the absence of perturbations, but its performance significantly deteriorates to nearly $0\\%$ with perturbations in the environment. In contrast, Skill Transformer exhibits adaptability to changes in the state of the receptacle and the ability to re-open the receptacle, achieving a success rate for picking the object of $19.5\\%$ in the presence of perturbations. Note that Skill Transformer is not trained with any perturbations, and this behavior emerges from the design choices made in Skill Transformer.", + "bbox": [ + 496, + 537, + 892, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Similarly, the accuracy of the predicted skills demonstrates that Skill Transformer outperforms M3 (Oracle) in high-level reasoning, both in the presence and absence of perturbations. Consistent with object retrieval, Skill Transformer demonstrates a notably higher accuracy, indicating a higher level of adaptability to perturbations, compared to M3 (Oracle), which suffers from the non-reactive task planner that does not correctly respond to perturbations.", + "bbox": [ + 496, + 750, + 893, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further compare Skill Transformer to DT (Skill), another end-to-end method that has the potential to adapt to", + "bbox": [ + 498, + 869, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10858", + "bbox": [ + 480, + 925, + 519, + 936 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c6895b0eee1cb630d8d19d7d154750c8f35e1365f468b1cdd789c3e2eef7cdda.jpg", + "image_caption": [ + "(a) Architecture" + ], + "image_footnote": [], + "bbox": [ + 138, + 89, + 367, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/06efc4a75aa8c59bbff841d3140185eebb0e4072be0b3d7b8cf936a60bcbeef7.jpg", + "image_caption": [ + "(b) Context Length", + "Figure 3: Success rates for ablation and analysis experiments across different policy architectures (Figure 3a), context lengths (Figure 3b), and training dataset sizes (Figure 3c) including performance on \"Eval\" and \"Train\" on the easy split setting." + ], + "image_footnote": [], + "bbox": [ + 370, + 90, + 596, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/48ab2462c259ae4ee067ab4ddd7dcadb75f6b6af004da18e67a8ee46514e278e.jpg", + "image_caption": [ + "(c) Dataset Size" + ], + "image_footnote": [], + "bbox": [ + 601, + 90, + 826, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "perturbations. DT (Skill) has a substantially lower success rate in picking the target object compared to Skill Transformer, with less than half the success rate in both perturbed and unperturbed environments. In contrast to object retrieval, DT (Skill) displays only slightly lower accuracy in predicting correct skill labels than Skill Transformer. Consistent with the qualitative findings in Section 5.3, these observations suggest that although DT (Skill) demonstrates relatively good high-level reasoning ability, its primary challenge remains in learning and effectively leveraging distinct skills. On the other hand, we find that Skill Transformer exhibits the combined advantages of adaptability to environment disturbances and proficiency in completing individual skills.", + "bbox": [ + 75, + 300, + 470, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Ablations and Analysis", + "text_level": 1, + "bbox": [ + 76, + 518, + 290, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We run ablations to address two key questions. Firstly, how necessary are the design choices of Skill Transformer, such as the transformer backbone, the Skill Inference module, and the choice of context length? Secondly, how do the dataset characteristics, such as its size and completeness, impact the performance? We perform all of the ablation experiments on the \"Easy\" split of the task.", + "bbox": [ + 75, + 542, + 468, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5.1 Policy Architecture Ablation", + "text_level": 1, + "bbox": [ + 76, + 665, + 331, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figure 3a, we evaluate the effectiveness of the transformer and the Skill Inference module for mobile manipulation tasks. By replacing the Transformer backbone with an LSTM backbone, we observe a significant decline in performance. We find that the policy struggles to effectively execute individual skills, and we posit that the LSTM architecture is not as data-efficient for learning multiple skills. Our results also indicate a similarly low success rate when the Skill Inference module is removed. Qualitatively, we discovered that without the Skill Inference module, the policy is not able to differentiate between individual skills, and frequently executes inconsistent actions.", + "bbox": [ + 75, + 688, + 468, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Context length refers to the maximum amount of historical information Skill Transformer takes as input. We vary", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the context length from 15 to 60 timesteps and evaluate its impact on performance. Shown in Figure 3b, a shorter context length of 15 timesteps performed slightly worse, while performance decreased as the context length increased from 30 to 60 timesteps. We posit that a shorter context length may not be able to fully capture the partially observable nature of the task, while a longer context length may lead to overfitting to the training dataset. Given the random robot start location during test time, the actual trajectory at test time will differ from that in the training dataset. The longer sequence exacerbates the discrepancy between the training dataset and the test time trajectory, resulting in an increased out-of-distribution error. A moderate context length may aid capturing short-term partially observed information and mitigating overfitting to the demonstration trajectories. The trend is similar to findings from previous works [49, 8].", + "bbox": [ + 496, + 300, + 893, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5.2 Dataset Size Analysis", + "text_level": 1, + "bbox": [ + 500, + 560, + 702, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As previously demonstrated [6], the size of the demonstration dataset is a vital aspect that impacts the efficacy of data-driven approaches such as Skill Transformer. As shown in Figure 3c, we trained Skill Transformer on subsets of the full demonstration dataset, ranging from 100 to 10k episodes. Our results showed that performance improved significantly when moving from 100 to 5k episodes, but the improvement is less pronounced when going from 5k to 10k demonstrations.", + "bbox": [ + 496, + 584, + 890, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5.3 Incomplete Demonstrations", + "text_level": 1, + "bbox": [ + 500, + 732, + 746, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In long-horizon tasks, obtaining sufficient demonstrations of the full trajectory can be challenging. We find that Skill Transformer is also able to work effectively when trained with datasets containing only subsets of the full task. To demonstrate this, we train Skill Transformer on only two subtasks of the task: navigate-pick and navigate-place, and it achieves a success rate of $36.5\\%$ in \"Train\" and $27.8\\%$ in \"Eval\" split, compared to $44\\%$ and $37\\%$ using complete trajectories. This ablation shows that, in contrast to DT, which uses reward-conditioning across an entire episode,", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10859", + "bbox": [ + 480, + 926, + 517, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Skill Transformer can also work with partial and incomplete demonstrations.", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 138, + 194, + 152 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This paper presents Skill Transformer, a novel end-to-end method for embodied object rearrangement tasks. Skill Transformer breaks the problem into predicting which skills to use and the low-level control conditioned on that skill, utilizing a causal transformer-based policy architecture. The proposed architecture outperforms baselines in challenging rearrangement problems requiring diverse skills and high-level reasoning over which skills to select.", + "bbox": [ + 75, + 164, + 468, + 285 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In future work, we will incorporate a more complex task specification, such as images of the desired state or natural language instructions. With the current geometric task specification, it is challenging to identify the target object after its container receptacle is manipulated. Specifically, because the object shifts inside a drawer when it is opened, its starting location no longer provides an indication of which objects is the target object. This limitation affects the success rate of retrieving objects in the challenging rearrange-hard task.", + "bbox": [ + 75, + 286, + 468, + 435 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 76, + 454, + 254, + 470 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The Georgia Tech effort was supported in part by NSF, ONR YIP, and ARO PECASE. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the U.S. Government, or any sponsor.", + "bbox": [ + 75, + 479, + 468, + 571 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 76, + 588, + 171, + 603 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anurag Ajay, Yilun Du, Abhi Gupta, Joshua Tenenbaum, Tommi Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision-making? arXiv preprint arXiv:2211.15657, 2022.", + "[2] Anurag Ajay, Aviral Kumar, Pulkit Agrawal, Sergey Levine, and Ofir Nachum. Opal: Offline primitive discovery for accelerating offline reinforcement learning. arXiv preprint arXiv:2010.13611, 2020.", + "[3] Akhil Bagaria and George Konidaris. Option discovery using deep skill chaining. In International Conference on Learning Representations, 2019.", + "[4] Dhruv Batra, Angel X Chang, Sonia Chernova, Andrew J Davison, Jia Deng, Vladlen Koltun, Sergey Levine, Jitendra Malik, Igor Mordatch, Roozbeh Mottaghi, et al. Rearrangement: A challenge for embodied ai. arXiv preprint arXiv:2011.01975, 2020.", + "[5] Rogerio Bonatti, Sai Vemprala, Shuang Ma, Felipe Frujeri, Shuhang Chen, and Ashish Kapoor. Pact: Perception-action causal transformer for autoregressive robotics pre-training. arXiv preprint arXiv:2209.11133, 2022." + ], + "bbox": [ + 84, + 613, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[6] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022.", + "[7] Micah Carroll, Jessy Lin, Orr Paradise, Raluca Georgescu, Mingfei Sun, David Bignell, Stephanie Milani, Katja Hofmann, Matthew Hausknecht, Anca Dragan, et al. Towards flexible inference in sequential decision problems via bidirectional transformers. arXiv preprint arXiv:2204.13326, 2022.", + "[8] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021.", + "[9] Alexander Clegg, Wenhao Yu, Jie Tan, C Karen Liu, and Greg Turk. Learning to dress: Synthesizing human dressing motion via deep reinforcement learning. ACM Transactions on Graphics (TOG), 37(6):1-10, 2018.", + "[10] Kiana Ehsani, Winson Han, Alvaro Herrasti, Eli VanderBilt, Luca Weihs, Eric Kolve, Aniruddha Kembhavi, and Roozbeh Mottaghi. Manipulathor: A framework for visual object manipulation. In CVPR, 2021.", + "[11] Benjamin Eysenbach, Abhishek Gupta, Julian Ibarz, and Sergey Levine. Diversity is all you need: Learning skills without a reward function. arXiv preprint arXiv:1802.06070, 2018.", + "[12] Pete Florence, Corey Lynch, Andy Zeng, Oscar A Ramirez, Ayzaan Wahid, Laura Downs, Adrian Wong, Johnny Lee, Igor Mordatch, and Jonathan Tompson. Implicit behavioral cloning. In Conference on Robot Learning, pages 158-168. PMLR, 2022.", + "[13] Roy Fox, Sanjay Krishnan, Ion Stoica, and Ken Goldberg. Multi-level discovery of deep options. arXiv preprint arXiv:1703.08294, 2017.", + "[14] Chuang Gan, Jeremy Schwartz, Seth Alter, Damian Mrowca, Martin Schrimpf, James Traer, Julian De Freitas, Jonas Kubilius, Abhishek Bhandwaldar, Nick Haber, et al. Threadworld: A platform for interactive multi-modal physical simulation. NeurIPS Datasets and Benchmarks Track, 2021.", + "[15] Caelan Reed Garrett, Rohan Chitnis, Rachel Holladay, Beomjoon Kim, Tom Silver, Leslie Pack Kaelbling, and Tomás Lozano-Pérez. Integrated task and motion planning. Annual review of control, robotics, and autonomous systems, 4:265-293, 2021.", + "[16] Jiayuan Gu, Devendra Singh Chaplot, Hao Su, and Jitendra Malik. Multi-skill mobile manipulation for object rearrangement. arXiv preprint arXiv:2209.02778, 2022.", + "[17] Kourosh Hakhamaneshi, Ruihan Zhao, Albert Zhan, Pieter Abbeel, and Michael Laskin. Hierarchical few-shot imitation with skill transition models. arXiv preprint arXiv:2107.08981, 2021.", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, 2016.", + "[19] Daniel Honerkamp, Tim Welschehold, and Abhinav Valada. Learning kinematic feasibility for mobile manipulation" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10860", + "bbox": [ + 480, + 926, + 517, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "through deep reinforcement learning. IEEE Robotics and Automation Letters, 6(4):6289-6296, 2021.", + "[20] Michael Janner, Qiyang Li, and Sergey Levine. Reinforcement learning as one big sequence modeling problem. In ICML 2021 Workshop on Unsupervised Reinforcement Learning, 2021.", + "[21] Dmitry Kalashnikov, Alex Irpan, Peter Pastor, Julian Ibarz, Alexander Herzog, Eric Jang, Deirdre Quillen, Ethan Holly, Mrinal Kalakrishnan, Vincent Vanhoucke, and Sergey Levine. Scalable deep reinforcement learning for vision-based robotic manipulation. In 2nd Annual Conference on Robot Learning, CoRL 2018, Zürich, Switzerland, 29-31 October 2018, Proceedings, volume 87 of Proceedings of Machine Learning Research, pages 651-673. PMLR, 2018.", + "[22] George Konidaris and Andrew Barto. Skill discovery in continuous reinforcement learning domains using skill chaining. Advances in neural information processing systems, 22, 2009.", + "[23] Kuang-Huei Lee, Ofir Nachum, Mengjiao Yang, Lisa Lee, Daniel Freeman, Winnie Xu, Sergio Guadarrama, Ian Fischer, Eric Jang, Henryk Michalewski, et al. Multi-game decision transformers. arXiv preprint arXiv:2205.15241, 2022.", + "[24] Youngwoon Lee, Joseph J Lim, Anima Anandkumar, and Yuke Zhu. Adversarial skill chaining for long-horizon robot manipulation via terminal state regularization. arXiv preprint arXiv:2111.07999, 2021.", + "[25] Youngwoon Lee, Shao-Hua Sun, Sriram Somasundaram, Edward S Hu, and Joseph J Lim. Composing complex skills by learning transition policies. In International Conference on Learning Representations, 2018.", + "[26] Chengshu Li, Fei Xia, Roberto Martin-Martin, and Silvio Savarese. Hrl4in: Hierarchical reinforcement learning for interactive navigation with mobile manipulators. In Conference on Robot Learning, pages 603-616. PMLR, 2020.", + "[27] Anthony Liang, Ishika Singh, Karl Pertsch, and Jesse Thomason. Transformer adapters for robot learning. In CoRL 2022 Workshop on Pre-training Robot Learning, 2022.", + "[28] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. arXiv preprint arXiv:2209.07753, 2022.", + "[29] Maria Vittoria Minniti, Farbod Farshidian, Ruben Grandia, and Marco Hutter. Whole-body mpc for a dynamically stable mobile manipulator. IEEE Robotics and Automation Letters, 4(4):3687-3694, 2019.", + "[30] Mayank Mittal, David Hoeller, Farbod Farshidian, Marco Hutter, and Animesh Garg. Articulated object interaction in unknown scenes with whole-body mobile manipulation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1647-1654. IEEE, 2022.", + "[31] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018.", + "[32] Ram Ramrakhya, Eric Undersander, Dhruv Batra, and Abhishek Das. Habitat-web: Learning embodied object-search strategies from human demonstrations at scale. In Proceed-" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5173-5183, 2022.", + "[33] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv preprint arXiv:2205.06175, 2022.", + "[34] Fetch robotics. Fetch. http://fetchrobotics.com/, 2020.", + "[35] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.", + "[36] Archit Sharma, Shixiang Gu, Sergey Levine, Vikash Kumar, and Karol Hausman. Dynamics-aware unsupervised discovery of skills. arXiv preprint arXiv:1907.01657, 2019.", + "[37] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In Conference on Robot Learning, pages 894–906. PMLR, 2022.", + "[38] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. arXiv preprint arXiv:2209.05451, 2022.", + "[39] Richard S Sutton, Doina Precup, and Satinder Singh. Between mdps and semi-mdps: A framework for temporal abstraction in reinforcement learning. Artificial intelligence, 112(1-2):181-211, 1999.", + "[40] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, et al. Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems, 34, 2021.", + "[41] Andrew Szot, Karmesh Yadav, Alex Clegg, Vincent-Pierre Berges, Aaron Gokaslan, Angel Chang, Manolis Savva, Zsolt Kira, and Dhruv Batra. Habitat rearrangement challenge 2022. https://aihabitat.org/challenge/2022_rearrange, 2022.", + "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.", + "[43] Cong Wang, Qifeng Zhang, Qiyan Tian, Shuo Li, Xiaohui Wang, David Lane, Yvan Petillot, and Sen Wang. Learning mobile manipulation through deep reinforcement learning. Sensors, 20(3):939, 2020.", + "[44] Luca Weihs, Matt Deitke, Aniruddha Kembhavi, and Roozbeh Mottaghi. Visual room rearrangement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5922-5931, 2021.", + "[45] Erik Wijmans, Irfan Essa, and Dhruv Batra. Ver: Scaling on-policy rl leads to the emergence of navigation in embodied rearrangement. In Advances in Neural Information Processing Systems, 2022.", + "[46] Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. arXiv preprint arXiv:1911.00357, 2019." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10861", + "bbox": [ + 480, + 926, + 517, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Josiah Wong, Albert Tung, Andrey Kurenkov, Ajay Mandlekar, Li Fei-Fei, Silvio Savarese, and Roberto Martin-Martin. Error-aware imitation learning from teleoperation data for mobile manipulation. In Conference on Robot Learning, pages 1367–1378. PMLR, 2022.", + "[48] Mengdi Xu, Yikang Shen, Shun Zhang, Yuchen Lu, Ding Zhao, Joshua Tenenbaum, and Chuang Gan. Prompting decision transformer for few-shot policy generalization. In International Conference on Machine Learning, pages 24631-24645. PMLR, 2022.", + "[49] Qinqing Zheng, Amy Zhang, and Aditya Grover. Online decision transformer. arXiv preprint arXiv:2202.05607, 2022." + ], + "bbox": [ + 78, + 90, + 468, + 260 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10862", + "bbox": [ + 480, + 926, + 517, + 936 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Skill Transformer_ A Monolithic Policy for Mobile Manipulation/e273608a-9944-4287-91dc-881cf0220c1e_model.json b/2023/Skill Transformer_ A Monolithic Policy for Mobile Manipulation/e273608a-9944-4287-91dc-881cf0220c1e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..52d012014017c7118f5e8d1c87fee4cc48533dc9 --- /dev/null +++ b/2023/Skill Transformer_ A Monolithic Policy for Mobile Manipulation/e273608a-9944-4287-91dc-881cf0220c1e_model.json @@ -0,0 +1,2191 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.158, + 0.131, + 0.812, + 0.153 + ], + "angle": 0, + "content": "Skill Transformer: A Monolithic Policy for Mobile Manipulation" + }, + { + "type": "text", + "bbox": [ + 0.122, + 0.182, + 0.24, + 0.22 + ], + "angle": 0, + "content": "Xiaoyu Huang \nGeorgia Tech" + }, + { + "type": "text", + "bbox": [ + 0.328, + 0.182, + 0.455, + 0.198 + ], + "angle": 0, + "content": "Dhruv Batra" + }, + { + "type": "text", + "bbox": [ + 0.281, + 0.2, + 0.528, + 0.218 + ], + "angle": 0, + "content": "Meta AI (FAIR), Georgia Tech" + }, + { + "type": "text", + "bbox": [ + 0.582, + 0.182, + 0.684, + 0.198 + ], + "angle": 0, + "content": "Akshara Rai" + }, + { + "type": "text", + "bbox": [ + 0.568, + 0.2, + 0.698, + 0.216 + ], + "angle": 0, + "content": "Meta AI (FAIR)" + }, + { + "type": "text", + "bbox": [ + 0.74, + 0.182, + 0.847, + 0.198 + ], + "angle": 0, + "content": "Andrew Szot" + }, + { + "type": "text", + "bbox": [ + 0.739, + 0.2, + 0.848, + 0.217 + ], + "angle": 0, + "content": "Georgia Tech" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.543 + ], + "angle": 0, + "content": "We present Skill Transformer, an approach for solving long-horizon robotic tasks by combining conditional sequence modeling and skill modularity. Conditioned on ego-centric and proprioceptive observations of a robot, Skill Transformer is trained end-to-end to predict both a high-level skill (e.g., navigation, picking, placing), and a whole-body low-level action (e.g., base and arm motion), using a transformer architecture and demonstration trajectories that solve the full task. It retains the composability and modularity of the overall task through a skill predictor module while reasoning about low-level actions and avoiding hand-off errors, common in modular approaches. We test Skill Transformer on an embodied rearrangement benchmark and find it performs robust task planning and low-level control in new scenarios, achieving a \\(2.5x\\) higher success rate than baselines in hard rearrangement problems." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.588, + 0.21, + 0.604 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.471, + 0.885 + ], + "angle": 0, + "content": "A long-standing goal of embodied AI is to build agents that can seamlessly operate in a home to accomplish tasks like preparing a meal, tidying up, or loading the dishwasher. To successfully carry out such tasks, an agent must be able to perform a wide range of diverse skills, like navigation and object manipulation, as well as high-level decision-making, all while operating with limited egocentric inputs. Our particular focus is the problem of embodied object rearrangement [41]. An example in Figure 1 shows that the Fetch mobile manipulator must rearrange a bowl from an initial position to a desired goal position (both circled in orange) in an unfamiliar environment without access to any pre-existing maps, 3D models, or precise object locations. Successfully completing this task requires executing both low-level motor commands for accurate arm and base movements, as well as high-level decision-making, such as determining whether a drawer needs to be opened before accessing a specific object." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Long-horizon tasks like object rearrangement require" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.895, + 0.557 + ], + "angle": 0, + "content": "many low-level control steps and naturally break down into distinct phases of skills such as navigation, picking, placing, opening, and closing. Prior works [6, 33, 37] show that it is possible to end-to-end learn a single policy that can perform such diverse tasks and behaviors. Such a monolithic policy directly maps observations to low-level actions, reasoning both about which skill to execute and how to execute it. However, scaling end-to-end learning to long-horizon, multi-phase tasks with thousands of low-level steps and egocentric visual observations remains a challenging research question. Other works decompose the task into individual skills and then separately sequence them to solve the full task [40, 16]. While this decomposition makes the problem tractable, this modularity comes at a cost. Sequencing skills suffers from hand-off errors between skills, where one skill ends in a situation where the next skill is no longer feasible. Moreover, it does not address how to learn the skill sequencing or how to adapt the planned sequence to unknown disturbances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.895, + 0.791 + ], + "angle": 0, + "content": "We present Skill Transformer, an end-to-end trainable transformer policy for multi-skill tasks. Skill Transformer predicts both low-level actions and high-level skills, conditioned on a history of high-dimensional egocentric camera observations. By combining conditional sequence modeling with a skill prediction module, Skill Transformer can model long sequences of both high and low-level actions without breaking the problem into disconnected submodules. In Figure 1, Skill Transformer learns a policy that can perform diverse motor skills, as well as high-level reasoning over which motor skills to execute. We train Skill Transformer autoregressively with trajectories that solve the full rearrangement task. Skill Transformer is then tested in unseen settings that require a variable number of high-level and low-level actions, overcoming unknown perturbations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Skill Transformer achieves state-of-the-art performance in rearrangement tasks requiring variable-length high-level planning and delicate low-level actions, such as rearranging an object that can be in an open or closed receptacle, in the Habitat environment [41]. It achieves \\(1.45\\mathrm{x}\\) higher success rates in the full task and up to \\(2.5\\mathrm{x}\\) higher in especially challenging situations over modular and end-to-end baselines" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10852" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.09, + 0.872, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.329, + 0.896, + 0.391 + ], + "angle": 0, + "content": "Figure 1: A Fetch robot is placed in an unseen house and must rearrange an object from a specified start to a desired goal location from egocentric visual observations. The robot infers the drawer is closed and opens the drawer to access the object. Next, it picks the object, navigates to the goal, and places the object. The Skill Transformer is a monolithic \"pixels-to-actions\" policy for such long-horizon mobile manipulation problems which infers the active skill and the low-level continuous control." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.471, + 0.496 + ], + "angle": 0, + "content": "and is more robust to environment disturbances, such as a drawer snapping shut after opening it, despite never seeing such disturbances during training. Furthermore, we demonstrate the importance of the design decisions in Skill Transformer and its ability to scale with larger training dataset sizes. The code is available at https://bit.ly/3qH2QQK." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.511, + 0.22, + 0.527 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.869 + ], + "angle": 0, + "content": "Object Rearrangement: Object rearrangement can be formulated as a task and motion planning [15] problem where an agent must move objects to a desired configuration, potentially under some constraints. We refer to [4] for a comprehensive survey. Prior works have also used end-to-end RL training for learning rearrangement [44, 32, 14, 10], but unlike our work, they rely on abstract interactions with the environment, no physics, or simplified action spaces. Other works focus on tabletop rearrangement [28, 21, 12] while our work focuses on longer horizon mobile manipulation. Although mobile manipulation has witnessed significant achievements in robotics learning [29, 43, 26, 30, 19], due to the vastly increased complexity, in rearrangement tasks, a majority of previous works [6, 40] explicitly separate the base and arm movements to simplify the task. A few works [16, 47, 45] have shown the superiority of mobile manipulation over static manipulation. In the Habitat Rearrangement Challenge [40, 41], methods that decompose the rearrangement task into skills are the most successful [16, 45] but they require a fixed high-level plan and cannot dynamically select skills based on observations during execution." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Transformers for Control: Prior work uses a transformer policy architecture to approach RL as a sequence" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.404, + 0.895, + 0.555 + ], + "angle": 0, + "content": "modeling problem [8, 49, 20, 7]. Our work also uses a transformer-based policy, but with a modified architecture that predicts high-level skills along with low-level actions. Other works demonstrate that transformer-based policies are competent multi-task learners [23, 48, 38, 6, 33]. Likewise, we examine how a transformer-based policy can learn multiple skills and compose them to solve long-horizon rearrangement tasks. [27, 5] further show that transformers can encode underlying dynamics of the environment that benefits learning for individual skills." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Skill Sequencing: Other works investigate how to sequence together skills to accomplish a long horizon task. By utilizing skills, agents can reason over longer time horizons by operating in the lifted action space of skills rather than over low-level actions [39]. Skills may be either learned from demonstrations [13, 2, 17], from trial and error learning [36, 11], or hand defined [25]. This work demonstrates that a single transformer policy can learn multiple skills. However, when sequencing together skills, there may be \"hand-off\" issues when skills fail to transition smoothly. Prior works tackle this issue by learning to match the initiation and termination sets of skills [24, 9] or learning new skills to handle the chaining [3, 22]. Decision Diffuser [1] learns a generative model to combine skills, but requires specifying the skill combination order. We overcome these issues by distilling all skills into a single policy and training this monolithic policy end-to-end to solve the overall task while maintaining skill modularity through a skill-prediction module." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "10853" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.143, + 0.105 + ], + "angle": 0, + "content": "3. Task" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.116, + 0.47, + 0.327 + ], + "angle": 0, + "content": "We focus on the problem of embodied object rearrangement, where a robot is spawned in an unseen house and must rearrange an object from a specified start position to a desired goal position entirely from on-board sensing and without any pre-built maps, object 3D models, or other privileged information. We follow the setup of the 2022 Habitat Rearrangement Challenge [41]. Specifically, a simulated Fetch robot [34] senses the world through a \\(256 \\times 256\\) head-mounted RGB-D camera, robot joint positions, whether the robot is holding an object or not, and base egomotion giving the relative position of the robot since the start of the episode. The 3D position of the object to rearrange, along with the 3D position of where to rearrange the object to is also provided to the agent." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.328, + 0.47, + 0.434 + ], + "angle": 0, + "content": "The agent interacts with the world via a mobile base, 7DoF arm, and suction grip attached to the arm. The episode is successful if the agent calls an episode termination action when the object is within \\(15\\mathrm{cm}\\) of its target. The agent fails if it excessively collides with the environment or fails to rearrange the object within the prescribed low-level time steps. See Appendix B for more task details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.435, + 0.471, + 0.646 + ], + "angle": 0, + "content": "We evaluate our method in the Rearrange-Hard Task, which presents challenges in both low-level actions and high-level reasoning. Unlike the Habitat Rearrangement tasks [16, 40] that prior works focus on, this task involves objects and goals that may spawn in closed receptacles. Therefore, the agent may need to perform intermediate actions, such as opening a drawer or fridge to access the target object or goal. Furthermore, the correct sequence of skills required to solve the task differs across episodes and is unavailable to the agent. As a result, the agent must infer the correct behaviors from egocentric visual observations and proprioception. This more natural setting better reflects real-world rearrangement scenarios, thus making it the focus of this work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.647, + 0.47, + 0.737 + ], + "angle": 0, + "content": "From the dataset provided in [41], we create a harder version of the full Rearrange track dataset for evaluation. In this harder version, more episodes have objects and goals contained in closed receptacles. Specifically, the updated episode distribution in evaluation for the rearrange-hard task is divided into three splits:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.738, + 0.418, + 0.752 + ], + "angle": 0, + "content": "- \\(25\\%\\) easy scenes with objects initially accessible." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.753, + 0.433, + 0.767 + ], + "angle": 0, + "content": "- \\(50\\%\\) hard scenes with objects in closed receptacles." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.768, + 0.468, + 0.797 + ], + "angle": 0, + "content": "- \\(25\\%\\) very hard scenes with both objects and goals in closed receptacles." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.738, + 0.468, + 0.797 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.799, + 0.442, + 0.814 + ], + "angle": 0, + "content": "The original rearrange-hard dataset is used for training." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.83, + 0.168, + 0.845 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We propose Skill Transformer, an approach for learning long-horizon, complex tasks that require sequencing multiple distinct skills. In Section 4.1, we describe the com" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "ponents of Skill Transformer, which include a Skill Inference module for reasoning over skill selections and an Action Inference module to predict actions for controlling the robot, both modeled using a Transformer architecture and conditioned on past and current observations. Next, in Section 4.2, we detail how we train Skill Transformer in an autoregressive fashion from demonstrations of the full task." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.21, + 0.776, + 0.225 + ], + "angle": 0, + "content": "4.1. Skill Transformer Architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.892, + 0.339 + ], + "angle": 0, + "content": "Skill Transformer breaks the long-horizon task of object rearrangement into first inferring skill selections using a Skill Inference module and then predicting low-level control using an Action Inference module, given this skill choice. Both modules are implemented as causal transformer neural networks [31], conditioned on past and current observations of the robot." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.893, + 0.627 + ], + "angle": 0, + "content": "Input: Both modules in Skill Transformer take as input a sequence of tokens that encode robot observations. We refer to the maximum length of the input sequence as the context length \\( C \\). Let \\( o_{t} \\) denote the \\( 256 \\times 256 \\times 1 \\) depth visual observations, and \\( p_{t} \\) denote the proprioceptive inputs of the joint angles and holding state at time \\( t \\). The visual observation \\( o_{t} \\) is first encoded with a Resnet-18 encoder [18] to a 256-dimension visual embedding. Likewise, non-visual inputs are encoded with a multi-layer perceptron network into a 256-dimension sensor embedding. The visual and non-visual continuous embeddings are concatenated to form a 512-dimension input token \\( x_{t} \\). We also add a per-time step positional embedding to each input token \\( x_{t} \\) to enable modeling time across the sequence. \\( x_{t} \\) forms the input token that encodes the current observations, and Skill Transformer takes as input a history of such tokens up to context length \\( C \\). Note that we do not use RGB input as prior work [40] found it unnecessary for geometric object rearrangement, and training with only depth input is faster." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Skill Inference: The first module of Skill Transformer is the Skill Inference module which predicts the current skill to be executed. The Skill Inference module takes as input a sequence of the input embeddings \\(\\{x_{t - C},\\ldots ,x_t\\}\\) through a causal transformer network [42] with a 512 hidden dimension, to produce a one-hot skill vector \\(c_{t}\\) corresponding to the predicted skill that should be active at the current time step \\(t\\). Using the causal mask, label \\(c_{i}\\) only has access to the previous tokens \\(\\{t:t < i\\}\\) in the input sequence. Then, we embed \\(c_{i}\\) into a latent skill embedding \\(\\{z_{t - C},\\dots ,z_t\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Action Inference: Next, the Action Inference module outputs a distribution over robot low-level actions. The Action Inference module takes as input the input embedding \\( x_{t} \\), an embedding of the previous time step action \\( a_{t-1} \\), and the skill embedding \\( z_{t} \\). Each of these inputs are represented as distinct tokens in the input sequence of \\( \\{x_{t-C}, a_{t-C-1}, z_{t-C}, \\ldots, x_{t}, a_{t-1}, z_{t}\\} \\). This sequence is then fed into a transformer with the same architecture as the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.52, + 0.938 + ], + "angle": 0, + "content": "10854" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.09, + 0.852, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.312, + 0.895, + 0.404 + ], + "angle": 0, + "content": "Figure 2: The Skill Transformer architecture. Skill Transformer learns to map egocentric depth images from the Fetch robot's head camera and proprioceptive state to actions controlling the robot, including the base movement, arm movement, and gripper state. First, the raw visual observations are processed by a visual encoder and concatenated with the proprioceptive state. Next, the Skill Inference module, implemented as a causal transformer sequence model, infers the skill ID that is executing for each time step. Finally, the Action Inference module predicts the low-level action to execute on the robot from the inferred skill, observation, and previous action information. The model is trained through end-to-end imitation learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.471, + 0.509 + ], + "angle": 0, + "content": "Skill Inference module. By taking as input the skill embeddings \\( z_{t} \\), the Action Inference module learns consistency in executing individual skills and the transitions between them. As we will show in Section 5.5, this also allows training Skill Transformer with datasets that only successfully complete a portion of the skills." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.47, + 0.751 + ], + "angle": 0, + "content": "The Action Inference module then outputs a sequence of predicted next actions \\(\\{a_{t - C},\\dots a_t\\}\\) which control the robot. These actions determine the whole body control actions composed of three components: arm joint positions, base velocities, and gripper state. The module first outputs a 12-dimension output vector. The first seven components are used as the mean of a normal distribution to sample the desired delta joint positions, which will then be actuated via a PD controller. The next two dimensions are used as the mean of a normal distribution to sample the desired base velocities, scaled by the maximum linear and angular velocities of the robot base. The remaining three outputs are passed through a softmax layer to generate probabilities for the gripper action, which indicate opening, closing, or maintaining the previous gripper state." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Unless otherwise specified, we use 2 transformer layers and a context length \\( C = 30 \\) for all architectures, giving 2M parameters for the transformer network in the Skill Inference module. The Action Inference module is also trained with the same architecture and a context length \\( C \\) of 30 time steps, translating to a sequence length of 90 tokens. In total, between the observation encoders, Skill Inference, and Action Inference modules, the Skill Transformer policy network consists of 16M parameters." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.417, + 0.747, + 0.433 + ], + "angle": 0, + "content": "4.2. Skill Transformer Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.892, + 0.56 + ], + "angle": 0, + "content": "In this section, we describe how Skill Transformer is trained end-to-end to learn long-horizon rearrangement tasks from imitation learning. We train the policy using supervised learning on demonstration trajectories produced by a data-generating policy. We assume the data-generating policy demonstrates a composition of skills to solve the task. Such compositional approaches are the state-of-the-art for the Habitat Rearrangement task [40, 16]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.893, + 0.732 + ], + "angle": 0, + "content": "Specifically, we assume the data-generating policy demonstrates one of \\(N\\) skills, and each time step \\(t\\) is labeled with the skill \\(\\mathrm{ID}y_{t}^{\\mathrm{skill}}\\) that is currently executing. In our problem setting, these skills are: pick, place, navigate, open drawer, and open fridge. We also record the policy's actions at time step \\(t\\) that consists of arm action \\(y_{t}^{\\mathrm{a}}\\), base velocity \\(y_{t}^{\\mathrm{b}}\\), and gripper state \\(y_{t}^{\\mathrm{g}}\\). In addition, we acquire from the simulator a one-dimensional indicator label \\(y_{t}^{\\mathrm{pred}}\\) defined as \\(y_{t}^{\\mathrm{pred}} = 0\\) if the object does not start in a closed receptacle and \\(y_{t}^{\\mathrm{pred}} = i\\) if \\(i^{\\mathrm{th}}\\) receptacle contains the object, at the time step \\(t\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.749, + 0.652, + 0.765 + ], + "angle": 0, + "content": "4.2.1 Training Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.773, + 0.892, + 0.817 + ], + "angle": 0, + "content": "We separately supervise the Skill Inference and Action Inference modules. We supervise the Skill Inference module over the context length \\( C \\) using loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.818, + 0.892, + 0.857 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s k i l l}} = \\sum_ {t = 1} ^ {C} \\mathcal {L} _ {\\mathrm {F L}} \\left(c _ {t}, y _ {t} ^ {\\text {s k i l l}}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{FL}}\\) represents the focal loss. The focal loss counteracts the imbalance of skill execution frequencies. For example, the navigation skill executes more frequently than" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "10855" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "the other skills as it is needed multiple times within a single episode, and takes more low-level time steps to execute than manipulation skills." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.468, + 0.228 + ], + "angle": 0, + "content": "In order to train the Action Inference module in parallel with the Skill Inference module, we use teacher-forcing, and input the ground truth skill labels during training, instead of the inferred distributions that are used at test time. We supervise the Action Inference module over the context length \\( C \\) with loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.228, + 0.468, + 0.267 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a c t}} = \\sum_ {t = 1} ^ {C} \\left(\\mathcal {L} _ {\\mathrm {M S E}} \\left(a _ {t} ^ {\\mathrm {b}}, y _ {t} ^ {\\mathrm {b}}\\right) + \\mathcal {L} _ {\\mathrm {M S E}} \\left(a _ {t} ^ {\\mathrm {a}}, y _ {t} ^ {\\mathrm {a}}\\right) + \\mathcal {L} _ {\\mathrm {F L}} \\left(a _ {t} ^ {\\mathrm {g}}, y _ {t} ^ {\\mathrm {g}}\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.267, + 0.42, + 0.281 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{MSE}}\\) represents the mean-squared error loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.282, + 0.468, + 0.386 + ], + "angle": 0, + "content": "In addition, we introduce auxiliary heads and losses for the Skill Inference module. The auxiliary head reconstructs oracle information unavailable to the policy. Specifically, assuming there exist \\(M\\) receptacles that may contain the target object, the auxiliary head outputs an \\(M + 1\\) dimensional distribution \\(v_{t}\\). We then supervise \\(v_{t}\\) with \\(y_{t}^{\\mathrm{pred}}\\) over the context length \\(C\\) via the loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.387, + 0.468, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a u x}} = \\sum_ {t = 1} ^ {C} \\mathcal {L} _ {\\mathrm {C E}} \\left(v _ {t}, y _ {t} ^ {\\text {p r e d}}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.427, + 0.378, + 0.441 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{CE}}\\) represents the Cross-Entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.468, + 0.607 + ], + "angle": 0, + "content": "We incorporate this auxiliary loss because the data-generating policy used to generate the demonstrations has a low success rate in correctly opening the receptacles. This low success rate can hinder learning because Skill Transformer needs to reason if it failed due to not opening the receptacle at all or opening an incorrect receptacle. The auxiliary loss incentivizes Skill Transformer to identify the key features that lead to the decision to open specific receptacles and to avoid wrong ones. This auxiliary objective with privileged information is only used for training and not used at test time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.468, + 0.79 + ], + "angle": 0, + "content": "The final loss is the unweighted sum of individual loss terms: \\(\\mathcal{L} = \\mathcal{L}_{\\mathrm{skill}} + \\mathcal{L}_{\\mathrm{action}} + \\mathcal{L}_{\\mathrm{aux}}\\). Note that in training, we calculate the loss over all time steps, whereas at test time, we only output action \\(a_{t}\\) for the current time step \\(t\\) to the environment. While training, we also train the visual encoder, in contrast to previous works using transformers in embodied AI that only demonstrate performance with pre-trained visual encoders [8, 6]. Since Skill Transformer is an end-to-end trainable policy, it can also be trained with offline-RL using the rewards from the dataset or with online-RL by interacting with the simulator. We leave exploring Skill Transformer with RL training for future work." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.249, + 0.83 + ], + "angle": 0, + "content": "4.2.2 Training Scheme" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We use a similar scheme as [8], where we input a sequence of 30 timesteps and compute the loss for the skill and action distributions for all timesteps in the sequence using autoregressive masking. Skill Transformer uses a learning warm-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "up phase where the learning rate is linearly increased at the start of training. After the warm-up phase, Skill Transformer uses a cosine learning rate scheduler which decays to 0 at the end of training. The entire dataset was iterated through approximately 6 times in total during the complete training process. We report the performance of the checkpoint with the highest success rate on seen scenes. We provide further details about Skill Transformer in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.225, + 0.632, + 0.241 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.25, + 0.892, + 0.401 + ], + "angle": 0, + "content": "In this section, we describe a series of experiments used to validate our proposed approach. First, we introduce the experiment setup and state-of-the-art baselines for comparison. Then, we quantify the performance of Skill Transformer in the rearrange-hard task and compare it to the baselines. Furthermore, the robustness of the methods to disturbances at evaluation time are evaluated by adversarial perturbations, like closing drawers after the agent has opened them. Finally, we perform an ablation study to confirm the key design choices involved in Skill Transformer." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.41, + 0.676, + 0.425 + ], + "angle": 0, + "content": "5.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.892, + 0.689 + ], + "angle": 0, + "content": "We conduct all experiments in the Habitat 2.0 [40] simulation environment and the rearrangement tasks as described in Section 3. For training, we use the train dataset split from the challenge consisting of 50,000 object configurations across 63 scenes. Each episode consists of starting object positions and a target location for the object to be rearranged. The robot positions are randomly generated in each episode. We first train individual skills (pick, place, navigate, open, close) that are used by a data generating policy to generate 10,000 demonstrations for training Skill Transformer. To train skills, we use the Multi-skill Mobile Manipulation (M3) [16] method with an oracle task planner with privileged information about the environment. This approach trains skills with RL and sequences them with an oracle planner. Further details of the data generating policy and the demonstration collected are detailed in Appendix C.1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.608, + 0.71 + ], + "angle": 0, + "content": "5.2. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.891, + 0.779 + ], + "angle": 0, + "content": "We compare Skill Transformer (ST) to the following baselines in the rearrange-hard task. The parameter count of each baseline policy is adjusted to match the 16M parameters in Skill Transformer." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.781, + 0.892, + 0.855 + ], + "angle": 0, + "content": "- Monolithic RL (Mono): Train a monolithic neural network directly mapping sensor inputs to actions with end-to-end RL using DD-PPO [46], a distributed form of PPO [35]. The monolithic RL policy is trained for 100M timesteps, reflecting a wall time of 21h." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "- Decision Transformer (DT) [8]: Train the monolithic policy from the offline dataset using a decision transformer architecture. The DT policy is trained with the" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "10856" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.468, + 0.136 + ], + "angle": 0, + "content": "same learning rate scheme and number of epochs as Skill Transformer. At test time, we condition the DT policy with the highest possible reward in the dataset." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.137, + 0.468, + 0.211 + ], + "angle": 0, + "content": "- Decision Transformer with Auxiliary Skill Loss (DT (Skill)): Train the Decision Transformer with an auxiliary loss of skill prediction. This baseline incorporates an auxiliary head that predicts the current skill during training and is trained with focal loss like Skill Transformer." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.213, + 0.468, + 0.303 + ], + "angle": 0, + "content": "Note that Decision Transformer requires the environment reward at evaluation time to compute the reward-to-go input. Skill Transformer and the other baselines do not require this assumption. We also compare to the following hierarchical baselines which first decompose the task into skills and then sequence these skills to solve the tasks." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.304, + 0.468, + 0.394 + ], + "angle": 0, + "content": "- Task-Planning + Skills RL (TP-SRL) [40]: Individually train pick, place, navigate, and open skills using RL and then sequence the skills together with a fixed task planner. Each skill predicts when it should end and the next skill should begin by a skill termination action. TP-SRL trains all manipulation skills with a fixed base." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.395, + 0.468, + 0.47 + ], + "angle": 0, + "content": "- Multi-skill Mobile Manipulation (M3) [16]: M3 is a modular approach similar to TP-SRL, but instead trains skills as mobile manipulation skills with RL. All these skills are trained to be robust to diverse starting states. Like TP-SRL, it uses a fixed task planner to sequence the skills." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.471, + 0.468, + 0.59 + ], + "angle": 0, + "content": "- Behavior Cloning Modular skills (BC-Modular): Individually train transformer-based policies for each of the skills using behavior cloning with the same training dataset as Skill Transformer. Then sequence together with the same fixed task planner as the other modular baselines. This baseline helps measure the gap in performance from training policies with behavior cloning as opposed to the RL-trained policies, M3 and TP-SRL." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.592, + 0.468, + 0.682 + ], + "angle": 0, + "content": "However, these modular baselines are designed to only execute a fixed pre-defined sequence of skills (navigate, pick, navigate, place). They cannot adjust their skill sequence based on environment observations, i.e. the task plan is inaccurate when the robot needs to open a receptacle to pick or place an object." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.684, + 0.468, + 0.835 + ], + "angle": 0, + "content": "To address this, we additionally include a version of the modular baselines, M3 (Oracle) and BC-Modular (Oracle), that extracts the correct task plan directly from the simulator, which we refer to as an oracle plan. In contrast, the proposed method, Skill Transformer, infers which skill to execute from egocentric observations without using this privileged oracle plan. Note that neither modular baseline re-executes skills if they fail, while Skill Transformer can re-plan. For further details about baselines and all hyperparameters, see Appendix C.2" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.847, + 0.242, + 0.862 + ], + "angle": 0, + "content": "5.3. Rearrange-Hard" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "First, we show the performance of Skill Transformer on the rearrange-hard task described in Section 3. In this task," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "the agent must rearrange a single object from the start to the goal, and the object or the goal may be located in a closed receptacle. Therefore, methods need high-level reasoning to decide if the receptacle should first be opened or not." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.154, + 0.892, + 0.38 + ], + "angle": 0, + "content": "We evaluate the success rate of all methods on both \"Train\" (Seen) and \"Eval\" (Unseen) scenes and object configurations. The success rate is measured as the average number of episodes where the agent placed the target object within \\(15\\mathrm{cm}\\) of the goal within 5000 steps and without excessively colliding with the environment. Specifically, the episode will terminate with failure if the agent accumulates more than \\(100\\mathrm{kN}\\) of force throughout the episode or \\(10\\mathrm{kN}\\) of instantaneous force due to collisions with the scene or objects. We also report the success rate on each of the three dataset splits described in Section 3. Our evaluation dataset splits have 100 episodes in the easy split, 200 in the hard split (100 for objects starting in closed receptacles and 100 for goals in closed receptacles), and 100 in the very hard split." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.382, + 0.892, + 0.473 + ], + "angle": 0, + "content": "Table 1 shows that Skill Transformer achieves higher success rates in both the \"Train\" and \"Eval\" settings than all of the baselines that do not have oracle planning. The Mono baseline cannot find any success, even after training for 100M steps. This demonstrates the difficulty of learning the rearrange-hard task from a reward signal alone." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.474, + 0.892, + 0.626 + ], + "angle": 0, + "content": "The modular TP-SRL, M3, and BC-Modular methods rely on a fixed task plan to execute the separately learned skill policies. As a result, the state-of-the-art M3 method achieves a significantly lower success rate than Skill Transformer in the overall result as well as the hard and very hard splits because it is unable to adapt the skill sequence to different situations. On the other hand, Skill Transformer does not rely on a fixed ordering of skills and can adapt its high-level task-solving to new situations, achieving a \\(35\\%\\) increase in success rate across all episodes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.628, + 0.892, + 0.853 + ], + "angle": 0, + "content": "Notice that even with an oracle plan, the success rate of M3 (Oracle) on the overall result for unseen scenes is only \\(3\\%\\) higher than that of Skill Transformer. More importantly, Skill Transformer outperforms M3 (Oracle) in both hard and very hard settings, even without oracle task planning information. We find that Skill Transformer consistently achieves higher success rates in episodes with hidden objects or goals, while M3 (Oracle) experiences a substantial decrement as the difficulty increases. This difference is likely due to the low success rate in M3 (Oracle)'s opening skill, hindering its ability to complete the skill in one attempt. In contrast, as we will discuss quantitatively later in Section 5.4, Skill Transformer is capable of detecting and retrying failed skills, eventually leading to better performance in retrieving and placing the object." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Further observe that BC-Modular exhibits the lowest performance among modular methods, with an overall success rate of only \\(5\\%\\) even when employing oracle task" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "10857" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.865, + 0.245 + ], + "angle": 0, + "content": "
MethodTrain (Seen)Eval (Unseen)
All EpisodesEasyHardVery HardAll EpisodesEasyHardVery Hard
BC-Modular (Oracle)5.67±0.1220.33±1.700.83±0.240.00±0.005.00±0.3515.33±1.701.67±1.650.33±0.47
M3 (Oracle)22.92±0.7259.33±1.7013.00±1.416.33±0.4722.58±1.1858.67±1.8912.67±2.366.33±0.47
BC-Modular5.08±0.4220.33±1.700.00±0.000.00±0.003.83±0.4215.33±1.700.00±0.000.00±0.00
TP-SRL7.25±0.8229.00±3.270.00±0.000.00±0.006.67±0.8526.67±3.400.00±0.000.00±0.00
M314.00±0.3555.67±0.940.16±0.470.00±0.0013.25±1.0853.00±4.320.00±0.000.00±0.00
Mono0.00±0.000.00±0.000.00±0.000.00±0.000.00±0.000.00±0.000.00±0.000.00±0.00
DT7.08±0.6319.33±0.473.83±0.622.00±0.946.50±0.4115.33±1.253.83±0.623.00±1.63
DT (Skill)8.17±0.6723.33±2.493.00±2.163.33±0.947.67±0.6618.67±2.624.83±1.702.33±1.25
ST (Ours)21.08±1.3644.00±0.8215.67±1.709.00±2.1619.17±1.0537.00±0.8215.83±0.858.00±3.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.256, + 0.895, + 0.334 + ], + "angle": 0, + "content": "Table 1: Success rates on the rearrangement task. \"Eval\" results are from test episodes in unseen scenes. In \"Easy\", objects and goals start on accessible locations. In \"Hard\", objects start in closed receptacles. In \"Very Hard\" objects and goals start in closed receptacles. All results are averages across 100 episodes except for the \"Hard\" split which is an average across 200 episodes. Grayed methods at the top use the oracle skill order. We report the mean and standard deviation across 3 random training seeds. For hierarchical methods, we retrain all the skills for each random seed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.471, + 0.465 + ], + "angle": 0, + "content": "planning. This indicates that policies trained via Behavior Cloning, which uses only one-tenth the experience of those trained with RL, encounter greater difficulty in executing intricate mobile manipulation skills. Conversely, Skill Transformer, featuring the multi-skill Action Inference module, mitigates this limitation and achieves up to 4x performance improvements across all episodes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.471, + 0.648 + ], + "angle": 0, + "content": "DT and DT (Skill) can also achieve some success in hard and very hard splits. Nonetheless, the success rates are significantly lower than Skill Transformer, especially in the very hard split. We observe that DT-based methods lack the capability to execute skills consistently and often mix up different skill behaviors. Also, notice that DT (Skill) with an auxiliary skill loss does not show a significant improvement in performance compared to DT without the skill loss. These observations highlight the superiority of directly conditioning on skills rather than as auxiliary rewards in learning tasks with clearly distinguished skills, such as rearrangement." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.664, + 0.333, + 0.68 + ], + "angle": 0, + "content": "5.4. Robustness to Perturbations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We conduct a perturbation test to evaluate the replanning ability of Skill Transformer and baselines in the rearrange-hard task. The test perturbs the environment by closing the target receptacle after the agent had opened it for the first time. To succeed, the agent must detect the drawer has been closed and adjust the current plan to reopen the receptacle and retrieve the object. We compare Skill Transformer's skill prediction and M3's task planner against ground truth skill labels derived from the privileged states of the simulation environment (detailed in Appendix C.3). The test is performed on 400 episodes that involved only objects in the closed drawer. This test compares against the highest-performing modular baseline with access to oracle task plans, M3 (Oracle), and the highest" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.359, + 0.791, + 0.375 + ], + "angle": 0, + "content": "performing end-to-end baseline, DT (Skill)." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.385, + 0.885, + 0.466 + ], + "angle": 0, + "content": "
MethodPicking Target ObjectPredicted Skill Accuracy
No PerturbPerturbNo PerturbPerturb
M3 (Oracle)24.000.5044.0010.00
DT (Skill)10.008.5039.0033.99
ST (Ours)26.7519.5060.0047.00
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.892, + 0.536 + ], + "angle": 0, + "content": "Table 2: Success rates for picking up the target objects and accuracy of Skill Inference module and M3's oracle-plant task planner under both perturbed and unperturbed environment, measured by ground truth skill labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.893, + 0.749 + ], + "angle": 0, + "content": "The findings in Table 2 demonstrate that Skill Transformer outperforms the M3 (Oracle) baseline both in the success rate of picking up the object and skill selection accuracy. Specifically, when assessing the success rates of object retrieval, M3 (Oracle) performs comparably to Skill Transformer in the absence of perturbations, but its performance significantly deteriorates to nearly \\(0\\%\\) with perturbations in the environment. In contrast, Skill Transformer exhibits adaptability to changes in the state of the receptacle and the ability to re-open the receptacle, achieving a success rate for picking the object of \\(19.5\\%\\) in the presence of perturbations. Note that Skill Transformer is not trained with any perturbations, and this behavior emerges from the design choices made in Skill Transformer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.894, + 0.871 + ], + "angle": 0, + "content": "Similarly, the accuracy of the predicted skills demonstrates that Skill Transformer outperforms M3 (Oracle) in high-level reasoning, both in the presence and absence of perturbations. Consistent with object retrieval, Skill Transformer demonstrates a notably higher accuracy, indicating a higher level of adaptability to perturbations, compared to M3 (Oracle), which suffers from the non-reactive task planner that does not correctly respond to perturbations." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We further compare Skill Transformer to DT (Skill), another end-to-end method that has the potential to adapt to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.52, + 0.938 + ], + "angle": 0, + "content": "10858" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.09, + 0.368, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.225, + 0.302, + 0.237 + ], + "angle": 0, + "content": "(a) Architecture" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.091, + 0.597, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.426, + 0.224, + 0.542, + 0.238 + ], + "angle": 0, + "content": "(b) Context Length" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.091, + 0.828, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.225, + 0.762, + 0.237 + ], + "angle": 0, + "content": "(c) Dataset Size" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.244, + 0.892, + 0.276 + ], + "angle": 0, + "content": "Figure 3: Success rates for ablation and analysis experiments across different policy architectures (Figure 3a), context lengths (Figure 3b), and training dataset sizes (Figure 3c) including performance on \"Eval\" and \"Train\" on the easy split setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.471, + 0.513 + ], + "angle": 0, + "content": "perturbations. DT (Skill) has a substantially lower success rate in picking the target object compared to Skill Transformer, with less than half the success rate in both perturbed and unperturbed environments. In contrast to object retrieval, DT (Skill) displays only slightly lower accuracy in predicting correct skill labels than Skill Transformer. Consistent with the qualitative findings in Section 5.3, these observations suggest that although DT (Skill) demonstrates relatively good high-level reasoning ability, its primary challenge remains in learning and effectively leveraging distinct skills. On the other hand, we find that Skill Transformer exhibits the combined advantages of adaptability to environment disturbances and proficiency in completing individual skills." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.52, + 0.292, + 0.536 + ], + "angle": 0, + "content": "5.5. Ablations and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.47, + 0.65 + ], + "angle": 0, + "content": "We run ablations to address two key questions. Firstly, how necessary are the design choices of Skill Transformer, such as the transformer backbone, the Skill Inference module, and the choice of context length? Secondly, how do the dataset characteristics, such as its size and completeness, impact the performance? We perform all of the ablation experiments on the \"Easy\" split of the task." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.332, + 0.681 + ], + "angle": 0, + "content": "5.5.1 Policy Architecture Ablation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.87 + ], + "angle": 0, + "content": "In Figure 3a, we evaluate the effectiveness of the transformer and the Skill Inference module for mobile manipulation tasks. By replacing the Transformer backbone with an LSTM backbone, we observe a significant decline in performance. We find that the policy struggles to effectively execute individual skills, and we posit that the LSTM architecture is not as data-efficient for learning multiple skills. Our results also indicate a similarly low success rate when the Skill Inference module is removed. Qualitatively, we discovered that without the Skill Inference module, the policy is not able to differentiate between individual skills, and frequently executes inconsistent actions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Context length refers to the maximum amount of historical information Skill Transformer takes as input. We vary" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.301, + 0.895, + 0.543 + ], + "angle": 0, + "content": "the context length from 15 to 60 timesteps and evaluate its impact on performance. Shown in Figure 3b, a shorter context length of 15 timesteps performed slightly worse, while performance decreased as the context length increased from 30 to 60 timesteps. We posit that a shorter context length may not be able to fully capture the partially observable nature of the task, while a longer context length may lead to overfitting to the training dataset. Given the random robot start location during test time, the actual trajectory at test time will differ from that in the training dataset. The longer sequence exacerbates the discrepancy between the training dataset and the test time trajectory, resulting in an increased out-of-distribution error. A moderate context length may aid capturing short-term partially observed information and mitigating overfitting to the demonstration trajectories. The trend is similar to findings from previous works [49, 8]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.561, + 0.704, + 0.577 + ], + "angle": 0, + "content": "5.5.2 Dataset Size Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.719 + ], + "angle": 0, + "content": "As previously demonstrated [6], the size of the demonstration dataset is a vital aspect that impacts the efficacy of data-driven approaches such as Skill Transformer. As shown in Figure 3c, we trained Skill Transformer on subsets of the full demonstration dataset, ranging from 100 to 10k episodes. Our results showed that performance improved significantly when moving from 100 to 5k episodes, but the improvement is less pronounced when going from 5k to 10k demonstrations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.733, + 0.747, + 0.747 + ], + "angle": 0, + "content": "5.5.3 Incomplete Demonstrations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In long-horizon tasks, obtaining sufficient demonstrations of the full trajectory can be challenging. We find that Skill Transformer is also able to work effectively when trained with datasets containing only subsets of the full task. To demonstrate this, we train Skill Transformer on only two subtasks of the task: navigate-pick and navigate-place, and it achieves a success rate of \\(36.5\\%\\) in \"Train\" and \\(27.8\\%\\) in \"Eval\" split, compared to \\(44\\%\\) and \\(37\\%\\) using complete trajectories. This ablation shows that, in contrast to DT, which uses reward-conditioning across an entire episode," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "10859" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "Skill Transformer can also work with partial and incomplete demonstrations." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.139, + 0.196, + 0.154 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.165, + 0.469, + 0.286 + ], + "angle": 0, + "content": "This paper presents Skill Transformer, a novel end-to-end method for embodied object rearrangement tasks. Skill Transformer breaks the problem into predicting which skills to use and the low-level control conditioned on that skill, utilizing a causal transformer-based policy architecture. The proposed architecture outperforms baselines in challenging rearrangement problems requiring diverse skills and high-level reasoning over which skills to select." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.287, + 0.469, + 0.436 + ], + "angle": 0, + "content": "In future work, we will incorporate a more complex task specification, such as images of the desired state or natural language instructions. With the current geometric task specification, it is challenging to identify the target object after its container receptacle is manipulated. Specifically, because the object shifts inside a drawer when it is opened, its starting location no longer provides an indication of which objects is the target object. This limitation affects the success rate of retrieving objects in the challenging rearrange-hard task." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.455, + 0.255, + 0.471 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.469, + 0.572 + ], + "angle": 0, + "content": "The Georgia Tech effort was supported in part by NSF, ONR YIP, and ARO PECASE. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the U.S. Government, or any sponsor." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.589, + 0.173, + 0.604 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.614, + 0.47, + 0.67 + ], + "angle": 0, + "content": "[1] Anurag Ajay, Yilun Du, Abhi Gupta, Joshua Tenenbaum, Tommi Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision-making? arXiv preprint arXiv:2211.15657, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.672, + 0.47, + 0.727 + ], + "angle": 0, + "content": "[2] Anurag Ajay, Aviral Kumar, Pulkit Agrawal, Sergey Levine, and Ofir Nachum. Opal: Offline primitive discovery for accelerating offline reinforcement learning. arXiv preprint arXiv:2010.13611, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.73, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[3] Akhil Bagaria and George Konidaris. Option discovery using deep skill chaining. In International Conference on Learning Representations, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.774, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[4] Dhruv Batra, Angel X Chang, Sonia Chernova, Andrew J Davison, Jia Deng, Vladlen Koltun, Sergey Levine, Jitendra Malik, Igor Mordatch, Roozbeh Mottaghi, et al. Rearrangement: A challenge for embodied ai. arXiv preprint arXiv:2011.01975, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.845, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[5] Rogerio Bonatti, Sai Vemprala, Shuang Ma, Felipe Frujeri, Shuhang Chen, and Ashish Kapoor. Pact: Perception-action causal transformer for autoregressive robotics pre-training. arXiv preprint arXiv:2209.11133, 2022." + }, + { + "type": "list", + "bbox": [ + 0.085, + 0.614, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[6] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.164, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[7] Micah Carroll, Jessy Lin, Orr Paradise, Raluca Georgescu, Mingfei Sun, David Bignell, Stephanie Milani, Katja Hofmann, Matthew Hausknecht, Anca Dragan, et al. Towards flexible inference in sequential decision problems via bidirectional transformers. arXiv preprint arXiv:2204.13326, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.248, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[8] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.319, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[9] Alexander Clegg, Wenhao Yu, Jie Tan, C Karen Liu, and Greg Turk. Learning to dress: Synthesizing human dressing motion via deep reinforcement learning. ACM Transactions on Graphics (TOG), 37(6):1-10, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[10] Kiana Ehsani, Winson Han, Alvaro Herrasti, Eli VanderBilt, Luca Weihs, Eric Kolve, Aniruddha Kembhavi, and Roozbeh Mottaghi. Manipulathor: A framework for visual object manipulation. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.486 + ], + "angle": 0, + "content": "[11] Benjamin Eysenbach, Abhishek Gupta, Julian Ibarz, and Sergey Levine. Diversity is all you need: Learning skills without a reward function. arXiv preprint arXiv:1802.06070, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.557 + ], + "angle": 0, + "content": "[12] Pete Florence, Corey Lynch, Andy Zeng, Oscar A Ramirez, Ayzaan Wahid, Laura Downs, Adrian Wong, Johnny Lee, Igor Mordatch, and Jonathan Tompson. Implicit behavioral cloning. In Conference on Robot Learning, pages 158-168. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.56, + 0.892, + 0.6 + ], + "angle": 0, + "content": "[13] Roy Fox, Sanjay Krishnan, Ion Stoica, and Ken Goldberg. Multi-level discovery of deep options. arXiv preprint arXiv:1703.08294, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.603, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[14] Chuang Gan, Jeremy Schwartz, Seth Alter, Damian Mrowca, Martin Schrimpf, James Traer, Julian De Freitas, Jonas Kubilius, Abhishek Bhandwaldar, Nick Haber, et al. Threadworld: A platform for interactive multi-modal physical simulation. NeurIPS Datasets and Benchmarks Track, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[15] Caelan Reed Garrett, Rohan Chitnis, Rachel Holladay, Beomjoon Kim, Tom Silver, Leslie Pack Kaelbling, and Tomás Lozano-Pérez. Integrated task and motion planning. Annual review of control, robotics, and autonomous systems, 4:265-293, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[16] Jiayuan Gu, Devendra Singh Chaplot, Hao Su, and Jitendra Malik. Multi-skill mobile manipulation for object rearrangement. arXiv preprint arXiv:2209.02778, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.841 + ], + "angle": 0, + "content": "[17] Kourosh Hakhamaneshi, Ruihan Zhao, Albert Zhan, Pieter Abbeel, and Michael Laskin. Hierarchical few-shot imitation with skill transition models. arXiv preprint arXiv:2107.08981, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.844, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[19] Daniel Honerkamp, Tim Welschehold, and Abhinav Valada. Learning kinematic feasibility for mobile manipulation" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "10860" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.093, + 0.47, + 0.12 + ], + "angle": 0, + "content": "through deep reinforcement learning. IEEE Robotics and Automation Letters, 6(4):6289-6296, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[20] Michael Janner, Qiyang Li, and Sergey Levine. Reinforcement learning as one big sequence modeling problem. In ICML 2021 Workshop on Unsupervised Reinforcement Learning, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[21] Dmitry Kalashnikov, Alex Irpan, Peter Pastor, Julian Ibarz, Alexander Herzog, Eric Jang, Deirdre Quillen, Ethan Holly, Mrinal Kalakrishnan, Vincent Vanhoucke, and Sergey Levine. Scalable deep reinforcement learning for vision-based robotic manipulation. In 2nd Annual Conference on Robot Learning, CoRL 2018, Zürich, Switzerland, 29-31 October 2018, Proceedings, volume 87 of Proceedings of Machine Learning Research, pages 651-673. PMLR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.469, + 0.345 + ], + "angle": 0, + "content": "[22] George Konidaris and Andrew Barto. Skill discovery in continuous reinforcement learning domains using skill chaining. Advances in neural information processing systems, 22, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[23] Kuang-Huei Lee, Ofir Nachum, Mengjiao Yang, Lisa Lee, Daniel Freeman, Winnie Xu, Sergio Guadarrama, Ian Fischer, Eric Jang, Henryk Michalewski, et al. Multi-game decision transformers. arXiv preprint arXiv:2205.15241, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.469, + 0.459 + ], + "angle": 0, + "content": "[24] Youngwoon Lee, Joseph J Lim, Anima Anandkumar, and Yuke Zhu. Adversarial skill chaining for long-horizon robot manipulation via terminal state regularization. arXiv preprint arXiv:2111.07999, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[25] Youngwoon Lee, Shao-Hua Sun, Sriram Somasundaram, Edward S Hu, and Joseph J Lim. Composing complex skills by learning transition policies. In International Conference on Learning Representations, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[26] Chengshu Li, Fei Xia, Roberto Martin-Martin, and Silvio Savarese. Hrl4in: Hierarchical reinforcement learning for interactive navigation with mobile manipulators. In Conference on Robot Learning, pages 603-616. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.629 + ], + "angle": 0, + "content": "[27] Anthony Liang, Ishika Singh, Karl Pertsch, and Jesse Thomason. Transformer adapters for robot learning. In CoRL 2022 Workshop on Pre-training Robot Learning, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[28] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. arXiv preprint arXiv:2209.07753, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[29] Maria Vittoria Minniti, Farbod Farshidian, Ruben Grandia, and Marco Hutter. Whole-body mpc for a dynamically stable mobile manipulator. IEEE Robotics and Automation Letters, 4(4):3687-3694, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[30] Mayank Mittal, David Hoeller, Farbod Farshidian, Marco Hutter, and Animesh Garg. Articulated object interaction in unknown scenes with whole-body mobile manipulation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1647-1654. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[31] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[32] Ram Ramrakhya, Eric Undersander, Dhruv Batra, and Abhishek Das. Habitat-web: Learning embodied object-search strategies from human demonstrations at scale. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5173-5183, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[33] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv preprint arXiv:2205.06175, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.892, + 0.222 + ], + "angle": 0, + "content": "[34] Fetch robotics. Fetch. http://fetchrobotics.com/, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.224, + 0.892, + 0.266 + ], + "angle": 0, + "content": "[35] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.268, + 0.892, + 0.31 + ], + "angle": 0, + "content": "[36] Archit Sharma, Shixiang Gu, Sergey Levine, Vikash Kumar, and Karol Hausman. Dynamics-aware unsupervised discovery of skills. arXiv preprint arXiv:1907.01657, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.312, + 0.892, + 0.354 + ], + "angle": 0, + "content": "[37] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In Conference on Robot Learning, pages 894–906. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.355, + 0.892, + 0.397 + ], + "angle": 0, + "content": "[38] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. arXiv preprint arXiv:2209.05451, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.399, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[39] Richard S Sutton, Doina Precup, and Satinder Singh. Between mdps and semi-mdps: A framework for temporal abstraction in reinforcement learning. Artificial intelligence, 112(1-2):181-211, 1999." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.457, + 0.892, + 0.54 + ], + "angle": 0, + "content": "[40] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, et al. Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems, 34, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.543, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[41] Andrew Szot, Karmesh Yadav, Alex Clegg, Vincent-Pierre Berges, Aaron Gokaslan, Angel Chang, Manolis Savva, Zsolt Kira, and Dhruv Batra. Habitat rearrangement challenge 2022. https://aihabitat.org/challenge/2022_rearrange, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.614, + 0.892, + 0.67 + ], + "angle": 0, + "content": "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[43] Cong Wang, Qifeng Zhang, Qiyan Tian, Shuo Li, Xiaohui Wang, David Lane, Yvan Petillot, and Sen Wang. Learning mobile manipulation through deep reinforcement learning. Sensors, 20(3):939, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.73, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[44] Luca Weihs, Matt Deitke, Aniruddha Kembhavi, and Roozbeh Mottaghi. Visual room rearrangement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5922-5931, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[45] Erik Wijmans, Irfan Essa, and Dhruv Batra. Ver: Scaling on-policy rl leads to the emergence of navigation in embodied rearrangement. In Advances in Neural Information Processing Systems, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[46] Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. arXiv preprint arXiv:1911.00357, 2019." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.518, + 0.938 + ], + "angle": 0, + "content": "10861" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[47] Josiah Wong, Albert Tung, Andrey Kurenkov, Ajay Mandlekar, Li Fei-Fei, Silvio Savarese, and Roberto Martin-Martin. Error-aware imitation learning from teleoperation data for mobile manipulation. In Conference on Robot Learning, pages 1367–1378. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[48] Mengdi Xu, Yikang Shen, Shun Zhang, Yuchen Lu, Ding Zhao, Joshua Tenenbaum, and Chuang Gan. Prompting decision transformer for few-shot policy generalization. In International Conference on Machine Learning, pages 24631-24645. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[49] Qinqing Zheng, Amy Zhang, and Aditya Grover. Online decision transformer. arXiv preprint arXiv:2202.05607, 2022." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "10862" + } + ] +] \ No newline at end of file diff --git a/2023/Skill Transformer_ A Monolithic Policy for Mobile Manipulation/e273608a-9944-4287-91dc-881cf0220c1e_origin.pdf b/2023/Skill Transformer_ A Monolithic Policy for Mobile Manipulation/e273608a-9944-4287-91dc-881cf0220c1e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bf3c9b4971aef4ae886b783303f8574305ef3267 --- /dev/null +++ b/2023/Skill Transformer_ A Monolithic Policy for Mobile Manipulation/e273608a-9944-4287-91dc-881cf0220c1e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac733f028ef0955c3a899e398f923303dc78d27343735e362cb37659d9d63ddc +size 1888496 diff --git a/2023/Skill Transformer_ A Monolithic Policy for Mobile Manipulation/full.md b/2023/Skill Transformer_ A Monolithic Policy for Mobile Manipulation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2d2a35420acba395c63dca68b0ca82dc25e3270e --- /dev/null +++ b/2023/Skill Transformer_ A Monolithic Policy for Mobile Manipulation/full.md @@ -0,0 +1,305 @@ +# Skill Transformer: A Monolithic Policy for Mobile Manipulation + +Xiaoyu Huang +Georgia Tech + +Dhruv Batra + +Meta AI (FAIR), Georgia Tech + +Akshara Rai + +Meta AI (FAIR) + +Andrew Szot + +Georgia Tech + +# Abstract + +We present Skill Transformer, an approach for solving long-horizon robotic tasks by combining conditional sequence modeling and skill modularity. Conditioned on ego-centric and proprioceptive observations of a robot, Skill Transformer is trained end-to-end to predict both a high-level skill (e.g., navigation, picking, placing), and a whole-body low-level action (e.g., base and arm motion), using a transformer architecture and demonstration trajectories that solve the full task. It retains the composability and modularity of the overall task through a skill predictor module while reasoning about low-level actions and avoiding hand-off errors, common in modular approaches. We test Skill Transformer on an embodied rearrangement benchmark and find it performs robust task planning and low-level control in new scenarios, achieving a $2.5x$ higher success rate than baselines in hard rearrangement problems. + +# 1. Introduction + +A long-standing goal of embodied AI is to build agents that can seamlessly operate in a home to accomplish tasks like preparing a meal, tidying up, or loading the dishwasher. To successfully carry out such tasks, an agent must be able to perform a wide range of diverse skills, like navigation and object manipulation, as well as high-level decision-making, all while operating with limited egocentric inputs. Our particular focus is the problem of embodied object rearrangement [41]. An example in Figure 1 shows that the Fetch mobile manipulator must rearrange a bowl from an initial position to a desired goal position (both circled in orange) in an unfamiliar environment without access to any pre-existing maps, 3D models, or precise object locations. Successfully completing this task requires executing both low-level motor commands for accurate arm and base movements, as well as high-level decision-making, such as determining whether a drawer needs to be opened before accessing a specific object. + +Long-horizon tasks like object rearrangement require + +many low-level control steps and naturally break down into distinct phases of skills such as navigation, picking, placing, opening, and closing. Prior works [6, 33, 37] show that it is possible to end-to-end learn a single policy that can perform such diverse tasks and behaviors. Such a monolithic policy directly maps observations to low-level actions, reasoning both about which skill to execute and how to execute it. However, scaling end-to-end learning to long-horizon, multi-phase tasks with thousands of low-level steps and egocentric visual observations remains a challenging research question. Other works decompose the task into individual skills and then separately sequence them to solve the full task [40, 16]. While this decomposition makes the problem tractable, this modularity comes at a cost. Sequencing skills suffers from hand-off errors between skills, where one skill ends in a situation where the next skill is no longer feasible. Moreover, it does not address how to learn the skill sequencing or how to adapt the planned sequence to unknown disturbances. + +We present Skill Transformer, an end-to-end trainable transformer policy for multi-skill tasks. Skill Transformer predicts both low-level actions and high-level skills, conditioned on a history of high-dimensional egocentric camera observations. By combining conditional sequence modeling with a skill prediction module, Skill Transformer can model long sequences of both high and low-level actions without breaking the problem into disconnected submodules. In Figure 1, Skill Transformer learns a policy that can perform diverse motor skills, as well as high-level reasoning over which motor skills to execute. We train Skill Transformer autoregressively with trajectories that solve the full rearrangement task. Skill Transformer is then tested in unseen settings that require a variable number of high-level and low-level actions, overcoming unknown perturbations. + +Skill Transformer achieves state-of-the-art performance in rearrangement tasks requiring variable-length high-level planning and delicate low-level actions, such as rearranging an object that can be in an open or closed receptacle, in the Habitat environment [41]. It achieves $1.45\mathrm{x}$ higher success rates in the full task and up to $2.5\mathrm{x}$ higher in especially challenging situations over modular and end-to-end baselines + +![](images/c585f2d9a95a0b53c9fa21e88d8ca29e6b43f55bbc2ececd28d2d27456965367.jpg) +Figure 1: A Fetch robot is placed in an unseen house and must rearrange an object from a specified start to a desired goal location from egocentric visual observations. The robot infers the drawer is closed and opens the drawer to access the object. Next, it picks the object, navigates to the goal, and places the object. The Skill Transformer is a monolithic "pixels-to-actions" policy for such long-horizon mobile manipulation problems which infers the active skill and the low-level continuous control. + +and is more robust to environment disturbances, such as a drawer snapping shut after opening it, despite never seeing such disturbances during training. Furthermore, we demonstrate the importance of the design decisions in Skill Transformer and its ability to scale with larger training dataset sizes. The code is available at https://bit.ly/3qH2QQK. + +# 2. Related Work + +Object Rearrangement: Object rearrangement can be formulated as a task and motion planning [15] problem where an agent must move objects to a desired configuration, potentially under some constraints. We refer to [4] for a comprehensive survey. Prior works have also used end-to-end RL training for learning rearrangement [44, 32, 14, 10], but unlike our work, they rely on abstract interactions with the environment, no physics, or simplified action spaces. Other works focus on tabletop rearrangement [28, 21, 12] while our work focuses on longer horizon mobile manipulation. Although mobile manipulation has witnessed significant achievements in robotics learning [29, 43, 26, 30, 19], due to the vastly increased complexity, in rearrangement tasks, a majority of previous works [6, 40] explicitly separate the base and arm movements to simplify the task. A few works [16, 47, 45] have shown the superiority of mobile manipulation over static manipulation. In the Habitat Rearrangement Challenge [40, 41], methods that decompose the rearrangement task into skills are the most successful [16, 45] but they require a fixed high-level plan and cannot dynamically select skills based on observations during execution. + +Transformers for Control: Prior work uses a transformer policy architecture to approach RL as a sequence + +modeling problem [8, 49, 20, 7]. Our work also uses a transformer-based policy, but with a modified architecture that predicts high-level skills along with low-level actions. Other works demonstrate that transformer-based policies are competent multi-task learners [23, 48, 38, 6, 33]. Likewise, we examine how a transformer-based policy can learn multiple skills and compose them to solve long-horizon rearrangement tasks. [27, 5] further show that transformers can encode underlying dynamics of the environment that benefits learning for individual skills. + +Skill Sequencing: Other works investigate how to sequence together skills to accomplish a long horizon task. By utilizing skills, agents can reason over longer time horizons by operating in the lifted action space of skills rather than over low-level actions [39]. Skills may be either learned from demonstrations [13, 2, 17], from trial and error learning [36, 11], or hand defined [25]. This work demonstrates that a single transformer policy can learn multiple skills. However, when sequencing together skills, there may be "hand-off" issues when skills fail to transition smoothly. Prior works tackle this issue by learning to match the initiation and termination sets of skills [24, 9] or learning new skills to handle the chaining [3, 22]. Decision Diffuser [1] learns a generative model to combine skills, but requires specifying the skill combination order. We overcome these issues by distilling all skills into a single policy and training this monolithic policy end-to-end to solve the overall task while maintaining skill modularity through a skill-prediction module. + +# 3. Task + +We focus on the problem of embodied object rearrangement, where a robot is spawned in an unseen house and must rearrange an object from a specified start position to a desired goal position entirely from on-board sensing and without any pre-built maps, object 3D models, or other privileged information. We follow the setup of the 2022 Habitat Rearrangement Challenge [41]. Specifically, a simulated Fetch robot [34] senses the world through a $256 \times 256$ head-mounted RGB-D camera, robot joint positions, whether the robot is holding an object or not, and base egomotion giving the relative position of the robot since the start of the episode. The 3D position of the object to rearrange, along with the 3D position of where to rearrange the object to is also provided to the agent. + +The agent interacts with the world via a mobile base, 7DoF arm, and suction grip attached to the arm. The episode is successful if the agent calls an episode termination action when the object is within $15\mathrm{cm}$ of its target. The agent fails if it excessively collides with the environment or fails to rearrange the object within the prescribed low-level time steps. See Appendix B for more task details. + +We evaluate our method in the Rearrange-Hard Task, which presents challenges in both low-level actions and high-level reasoning. Unlike the Habitat Rearrangement tasks [16, 40] that prior works focus on, this task involves objects and goals that may spawn in closed receptacles. Therefore, the agent may need to perform intermediate actions, such as opening a drawer or fridge to access the target object or goal. Furthermore, the correct sequence of skills required to solve the task differs across episodes and is unavailable to the agent. As a result, the agent must infer the correct behaviors from egocentric visual observations and proprioception. This more natural setting better reflects real-world rearrangement scenarios, thus making it the focus of this work. + +From the dataset provided in [41], we create a harder version of the full Rearrange track dataset for evaluation. In this harder version, more episodes have objects and goals contained in closed receptacles. Specifically, the updated episode distribution in evaluation for the rearrange-hard task is divided into three splits: + +- $25\%$ easy scenes with objects initially accessible. +- $50\%$ hard scenes with objects in closed receptacles. +- $25\%$ very hard scenes with both objects and goals in closed receptacles. + +The original rearrange-hard dataset is used for training. + +# 4. Method + +We propose Skill Transformer, an approach for learning long-horizon, complex tasks that require sequencing multiple distinct skills. In Section 4.1, we describe the com + +ponents of Skill Transformer, which include a Skill Inference module for reasoning over skill selections and an Action Inference module to predict actions for controlling the robot, both modeled using a Transformer architecture and conditioned on past and current observations. Next, in Section 4.2, we detail how we train Skill Transformer in an autoregressive fashion from demonstrations of the full task. + +# 4.1. Skill Transformer Architecture + +Skill Transformer breaks the long-horizon task of object rearrangement into first inferring skill selections using a Skill Inference module and then predicting low-level control using an Action Inference module, given this skill choice. Both modules are implemented as causal transformer neural networks [31], conditioned on past and current observations of the robot. + +Input: Both modules in Skill Transformer take as input a sequence of tokens that encode robot observations. We refer to the maximum length of the input sequence as the context length $C$ . Let $o_{t}$ denote the $256 \times 256 \times 1$ depth visual observations, and $p_{t}$ denote the proprioceptive inputs of the joint angles and holding state at time $t$ . The visual observation $o_{t}$ is first encoded with a Resnet-18 encoder [18] to a 256-dimension visual embedding. Likewise, non-visual inputs are encoded with a multi-layer perceptron network into a 256-dimension sensor embedding. The visual and non-visual continuous embeddings are concatenated to form a 512-dimension input token $x_{t}$ . We also add a per-time step positional embedding to each input token $x_{t}$ to enable modeling time across the sequence. $x_{t}$ forms the input token that encodes the current observations, and Skill Transformer takes as input a history of such tokens up to context length $C$ . Note that we do not use RGB input as prior work [40] found it unnecessary for geometric object rearrangement, and training with only depth input is faster. + +Skill Inference: The first module of Skill Transformer is the Skill Inference module which predicts the current skill to be executed. The Skill Inference module takes as input a sequence of the input embeddings $\{x_{t - C},\ldots ,x_t\}$ through a causal transformer network [42] with a 512 hidden dimension, to produce a one-hot skill vector $c_{t}$ corresponding to the predicted skill that should be active at the current time step $t$ . Using the causal mask, label $c_{i}$ only has access to the previous tokens $\{t:t < i\}$ in the input sequence. Then, we embed $c_{i}$ into a latent skill embedding $\{z_{t - C},\dots ,z_t\}$ . + +Action Inference: Next, the Action Inference module outputs a distribution over robot low-level actions. The Action Inference module takes as input the input embedding $x_{t}$ , an embedding of the previous time step action $a_{t-1}$ , and the skill embedding $z_{t}$ . Each of these inputs are represented as distinct tokens in the input sequence of $\{x_{t-C}, a_{t-C-1}, z_{t-C}, \ldots, x_{t}, a_{t-1}, z_{t}\}$ . This sequence is then fed into a transformer with the same architecture as the + +![](images/5c95c374819662e68a55ffd084578f8f7ef856cbd80ea6191a348ef3244adf24.jpg) +Figure 2: The Skill Transformer architecture. Skill Transformer learns to map egocentric depth images from the Fetch robot's head camera and proprioceptive state to actions controlling the robot, including the base movement, arm movement, and gripper state. First, the raw visual observations are processed by a visual encoder and concatenated with the proprioceptive state. Next, the Skill Inference module, implemented as a causal transformer sequence model, infers the skill ID that is executing for each time step. Finally, the Action Inference module predicts the low-level action to execute on the robot from the inferred skill, observation, and previous action information. The model is trained through end-to-end imitation learning. + +Skill Inference module. By taking as input the skill embeddings $z_{t}$ , the Action Inference module learns consistency in executing individual skills and the transitions between them. As we will show in Section 5.5, this also allows training Skill Transformer with datasets that only successfully complete a portion of the skills. + +The Action Inference module then outputs a sequence of predicted next actions $\{a_{t - C},\dots a_t\}$ which control the robot. These actions determine the whole body control actions composed of three components: arm joint positions, base velocities, and gripper state. The module first outputs a 12-dimension output vector. The first seven components are used as the mean of a normal distribution to sample the desired delta joint positions, which will then be actuated via a PD controller. The next two dimensions are used as the mean of a normal distribution to sample the desired base velocities, scaled by the maximum linear and angular velocities of the robot base. The remaining three outputs are passed through a softmax layer to generate probabilities for the gripper action, which indicate opening, closing, or maintaining the previous gripper state. + +Unless otherwise specified, we use 2 transformer layers and a context length $C = 30$ for all architectures, giving 2M parameters for the transformer network in the Skill Inference module. The Action Inference module is also trained with the same architecture and a context length $C$ of 30 time steps, translating to a sequence length of 90 tokens. In total, between the observation encoders, Skill Inference, and Action Inference modules, the Skill Transformer policy network consists of 16M parameters. + +# 4.2. Skill Transformer Training + +In this section, we describe how Skill Transformer is trained end-to-end to learn long-horizon rearrangement tasks from imitation learning. We train the policy using supervised learning on demonstration trajectories produced by a data-generating policy. We assume the data-generating policy demonstrates a composition of skills to solve the task. Such compositional approaches are the state-of-the-art for the Habitat Rearrangement task [40, 16]. + +Specifically, we assume the data-generating policy demonstrates one of $N$ skills, and each time step $t$ is labeled with the skill $\mathrm{ID}y_{t}^{\mathrm{skill}}$ that is currently executing. In our problem setting, these skills are: pick, place, navigate, open drawer, and open fridge. We also record the policy's actions at time step $t$ that consists of arm action $y_{t}^{\mathrm{a}}$ , base velocity $y_{t}^{\mathrm{b}}$ , and gripper state $y_{t}^{\mathrm{g}}$ . In addition, we acquire from the simulator a one-dimensional indicator label $y_{t}^{\mathrm{pred}}$ defined as $y_{t}^{\mathrm{pred}} = 0$ if the object does not start in a closed receptacle and $y_{t}^{\mathrm{pred}} = i$ if $i^{\mathrm{th}}$ receptacle contains the object, at the time step $t$ . + +# 4.2.1 Training Loss + +We separately supervise the Skill Inference and Action Inference modules. We supervise the Skill Inference module over the context length $C$ using loss function: + +$$ +\mathcal {L} _ {\text {s k i l l}} = \sum_ {t = 1} ^ {C} \mathcal {L} _ {\mathrm {F L}} \left(c _ {t}, y _ {t} ^ {\text {s k i l l}}\right) \tag {1} +$$ + +where $\mathcal{L}_{\mathrm{FL}}$ represents the focal loss. The focal loss counteracts the imbalance of skill execution frequencies. For example, the navigation skill executes more frequently than + +the other skills as it is needed multiple times within a single episode, and takes more low-level time steps to execute than manipulation skills. + +In order to train the Action Inference module in parallel with the Skill Inference module, we use teacher-forcing, and input the ground truth skill labels during training, instead of the inferred distributions that are used at test time. We supervise the Action Inference module over the context length $C$ with loss function: + +$$ +\mathcal {L} _ {\mathrm {a c t}} = \sum_ {t = 1} ^ {C} \left(\mathcal {L} _ {\mathrm {M S E}} \left(a _ {t} ^ {\mathrm {b}}, y _ {t} ^ {\mathrm {b}}\right) + \mathcal {L} _ {\mathrm {M S E}} \left(a _ {t} ^ {\mathrm {a}}, y _ {t} ^ {\mathrm {a}}\right) + \mathcal {L} _ {\mathrm {F L}} \left(a _ {t} ^ {\mathrm {g}}, y _ {t} ^ {\mathrm {g}}\right)\right) \tag {2} +$$ + +where $\mathcal{L}_{\mathrm{MSE}}$ represents the mean-squared error loss. + +In addition, we introduce auxiliary heads and losses for the Skill Inference module. The auxiliary head reconstructs oracle information unavailable to the policy. Specifically, assuming there exist $M$ receptacles that may contain the target object, the auxiliary head outputs an $M + 1$ dimensional distribution $v_{t}$ . We then supervise $v_{t}$ with $y_{t}^{\mathrm{pred}}$ over the context length $C$ via the loss function: + +$$ +\mathcal {L} _ {\text {a u x}} = \sum_ {t = 1} ^ {C} \mathcal {L} _ {\mathrm {C E}} \left(v _ {t}, y _ {t} ^ {\text {p r e d}}\right) \tag {3} +$$ + +where $\mathcal{L}_{\mathrm{CE}}$ represents the Cross-Entropy loss. + +We incorporate this auxiliary loss because the data-generating policy used to generate the demonstrations has a low success rate in correctly opening the receptacles. This low success rate can hinder learning because Skill Transformer needs to reason if it failed due to not opening the receptacle at all or opening an incorrect receptacle. The auxiliary loss incentivizes Skill Transformer to identify the key features that lead to the decision to open specific receptacles and to avoid wrong ones. This auxiliary objective with privileged information is only used for training and not used at test time. + +The final loss is the unweighted sum of individual loss terms: $\mathcal{L} = \mathcal{L}_{\mathrm{skill}} + \mathcal{L}_{\mathrm{action}} + \mathcal{L}_{\mathrm{aux}}$ . Note that in training, we calculate the loss over all time steps, whereas at test time, we only output action $a_{t}$ for the current time step $t$ to the environment. While training, we also train the visual encoder, in contrast to previous works using transformers in embodied AI that only demonstrate performance with pre-trained visual encoders [8, 6]. Since Skill Transformer is an end-to-end trainable policy, it can also be trained with offline-RL using the rewards from the dataset or with online-RL by interacting with the simulator. We leave exploring Skill Transformer with RL training for future work. + +# 4.2.2 Training Scheme + +We use a similar scheme as [8], where we input a sequence of 30 timesteps and compute the loss for the skill and action distributions for all timesteps in the sequence using autoregressive masking. Skill Transformer uses a learning warm- + +up phase where the learning rate is linearly increased at the start of training. After the warm-up phase, Skill Transformer uses a cosine learning rate scheduler which decays to 0 at the end of training. The entire dataset was iterated through approximately 6 times in total during the complete training process. We report the performance of the checkpoint with the highest success rate on seen scenes. We provide further details about Skill Transformer in Appendix A. + +# 5. Experiments + +In this section, we describe a series of experiments used to validate our proposed approach. First, we introduce the experiment setup and state-of-the-art baselines for comparison. Then, we quantify the performance of Skill Transformer in the rearrange-hard task and compare it to the baselines. Furthermore, the robustness of the methods to disturbances at evaluation time are evaluated by adversarial perturbations, like closing drawers after the agent has opened them. Finally, we perform an ablation study to confirm the key design choices involved in Skill Transformer. + +# 5.1. Experiment Setup + +We conduct all experiments in the Habitat 2.0 [40] simulation environment and the rearrangement tasks as described in Section 3. For training, we use the train dataset split from the challenge consisting of 50,000 object configurations across 63 scenes. Each episode consists of starting object positions and a target location for the object to be rearranged. The robot positions are randomly generated in each episode. We first train individual skills (pick, place, navigate, open, close) that are used by a data generating policy to generate 10,000 demonstrations for training Skill Transformer. To train skills, we use the Multi-skill Mobile Manipulation (M3) [16] method with an oracle task planner with privileged information about the environment. This approach trains skills with RL and sequences them with an oracle planner. Further details of the data generating policy and the demonstration collected are detailed in Appendix C.1. + +# 5.2. Baselines + +We compare Skill Transformer (ST) to the following baselines in the rearrange-hard task. The parameter count of each baseline policy is adjusted to match the 16M parameters in Skill Transformer. + +- Monolithic RL (Mono): Train a monolithic neural network directly mapping sensor inputs to actions with end-to-end RL using DD-PPO [46], a distributed form of PPO [35]. The monolithic RL policy is trained for 100M timesteps, reflecting a wall time of 21h. +- Decision Transformer (DT) [8]: Train the monolithic policy from the offline dataset using a decision transformer architecture. The DT policy is trained with the + +same learning rate scheme and number of epochs as Skill Transformer. At test time, we condition the DT policy with the highest possible reward in the dataset. + +- Decision Transformer with Auxiliary Skill Loss (DT (Skill)): Train the Decision Transformer with an auxiliary loss of skill prediction. This baseline incorporates an auxiliary head that predicts the current skill during training and is trained with focal loss like Skill Transformer. + +Note that Decision Transformer requires the environment reward at evaluation time to compute the reward-to-go input. Skill Transformer and the other baselines do not require this assumption. We also compare to the following hierarchical baselines which first decompose the task into skills and then sequence these skills to solve the tasks. + +- Task-Planning + Skills RL (TP-SRL) [40]: Individually train pick, place, navigate, and open skills using RL and then sequence the skills together with a fixed task planner. Each skill predicts when it should end and the next skill should begin by a skill termination action. TP-SRL trains all manipulation skills with a fixed base. + +- Multi-skill Mobile Manipulation (M3) [16]: M3 is a modular approach similar to TP-SRL, but instead trains skills as mobile manipulation skills with RL. All these skills are trained to be robust to diverse starting states. Like TP-SRL, it uses a fixed task planner to sequence the skills. + +- Behavior Cloning Modular skills (BC-Modular): Individually train transformer-based policies for each of the skills using behavior cloning with the same training dataset as Skill Transformer. Then sequence together with the same fixed task planner as the other modular baselines. This baseline helps measure the gap in performance from training policies with behavior cloning as opposed to the RL-trained policies, M3 and TP-SRL. + +However, these modular baselines are designed to only execute a fixed pre-defined sequence of skills (navigate, pick, navigate, place). They cannot adjust their skill sequence based on environment observations, i.e. the task plan is inaccurate when the robot needs to open a receptacle to pick or place an object. + +To address this, we additionally include a version of the modular baselines, M3 (Oracle) and BC-Modular (Oracle), that extracts the correct task plan directly from the simulator, which we refer to as an oracle plan. In contrast, the proposed method, Skill Transformer, infers which skill to execute from egocentric observations without using this privileged oracle plan. Note that neither modular baseline re-executes skills if they fail, while Skill Transformer can re-plan. For further details about baselines and all hyperparameters, see Appendix C.2 + +# 5.3. Rearrange-Hard + +First, we show the performance of Skill Transformer on the rearrange-hard task described in Section 3. In this task, + +the agent must rearrange a single object from the start to the goal, and the object or the goal may be located in a closed receptacle. Therefore, methods need high-level reasoning to decide if the receptacle should first be opened or not. + +We evaluate the success rate of all methods on both "Train" (Seen) and "Eval" (Unseen) scenes and object configurations. The success rate is measured as the average number of episodes where the agent placed the target object within $15\mathrm{cm}$ of the goal within 5000 steps and without excessively colliding with the environment. Specifically, the episode will terminate with failure if the agent accumulates more than $100\mathrm{kN}$ of force throughout the episode or $10\mathrm{kN}$ of instantaneous force due to collisions with the scene or objects. We also report the success rate on each of the three dataset splits described in Section 3. Our evaluation dataset splits have 100 episodes in the easy split, 200 in the hard split (100 for objects starting in closed receptacles and 100 for goals in closed receptacles), and 100 in the very hard split. + +Table 1 shows that Skill Transformer achieves higher success rates in both the "Train" and "Eval" settings than all of the baselines that do not have oracle planning. The Mono baseline cannot find any success, even after training for 100M steps. This demonstrates the difficulty of learning the rearrange-hard task from a reward signal alone. + +The modular TP-SRL, M3, and BC-Modular methods rely on a fixed task plan to execute the separately learned skill policies. As a result, the state-of-the-art M3 method achieves a significantly lower success rate than Skill Transformer in the overall result as well as the hard and very hard splits because it is unable to adapt the skill sequence to different situations. On the other hand, Skill Transformer does not rely on a fixed ordering of skills and can adapt its high-level task-solving to new situations, achieving a $35\%$ increase in success rate across all episodes. + +Notice that even with an oracle plan, the success rate of M3 (Oracle) on the overall result for unseen scenes is only $3\%$ higher than that of Skill Transformer. More importantly, Skill Transformer outperforms M3 (Oracle) in both hard and very hard settings, even without oracle task planning information. We find that Skill Transformer consistently achieves higher success rates in episodes with hidden objects or goals, while M3 (Oracle) experiences a substantial decrement as the difficulty increases. This difference is likely due to the low success rate in M3 (Oracle)'s opening skill, hindering its ability to complete the skill in one attempt. In contrast, as we will discuss quantitatively later in Section 5.4, Skill Transformer is capable of detecting and retrying failed skills, eventually leading to better performance in retrieving and placing the object. + +Further observe that BC-Modular exhibits the lowest performance among modular methods, with an overall success rate of only $5\%$ even when employing oracle task + +
MethodTrain (Seen)Eval (Unseen)
All EpisodesEasyHardVery HardAll EpisodesEasyHardVery Hard
BC-Modular (Oracle)5.67±0.1220.33±1.700.83±0.240.00±0.005.00±0.3515.33±1.701.67±1.650.33±0.47
M3 (Oracle)22.92±0.7259.33±1.7013.00±1.416.33±0.4722.58±1.1858.67±1.8912.67±2.366.33±0.47
BC-Modular5.08±0.4220.33±1.700.00±0.000.00±0.003.83±0.4215.33±1.700.00±0.000.00±0.00
TP-SRL7.25±0.8229.00±3.270.00±0.000.00±0.006.67±0.8526.67±3.400.00±0.000.00±0.00
M314.00±0.3555.67±0.940.16±0.470.00±0.0013.25±1.0853.00±4.320.00±0.000.00±0.00
Mono0.00±0.000.00±0.000.00±0.000.00±0.000.00±0.000.00±0.000.00±0.000.00±0.00
DT7.08±0.6319.33±0.473.83±0.622.00±0.946.50±0.4115.33±1.253.83±0.623.00±1.63
DT (Skill)8.17±0.6723.33±2.493.00±2.163.33±0.947.67±0.6618.67±2.624.83±1.702.33±1.25
ST (Ours)21.08±1.3644.00±0.8215.67±1.709.00±2.1619.17±1.0537.00±0.8215.83±0.858.00±3.27
+ +Table 1: Success rates on the rearrangement task. "Eval" results are from test episodes in unseen scenes. In "Easy", objects and goals start on accessible locations. In "Hard", objects start in closed receptacles. In "Very Hard" objects and goals start in closed receptacles. All results are averages across 100 episodes except for the "Hard" split which is an average across 200 episodes. Grayed methods at the top use the oracle skill order. We report the mean and standard deviation across 3 random training seeds. For hierarchical methods, we retrain all the skills for each random seed. + +planning. This indicates that policies trained via Behavior Cloning, which uses only one-tenth the experience of those trained with RL, encounter greater difficulty in executing intricate mobile manipulation skills. Conversely, Skill Transformer, featuring the multi-skill Action Inference module, mitigates this limitation and achieves up to 4x performance improvements across all episodes. + +DT and DT (Skill) can also achieve some success in hard and very hard splits. Nonetheless, the success rates are significantly lower than Skill Transformer, especially in the very hard split. We observe that DT-based methods lack the capability to execute skills consistently and often mix up different skill behaviors. Also, notice that DT (Skill) with an auxiliary skill loss does not show a significant improvement in performance compared to DT without the skill loss. These observations highlight the superiority of directly conditioning on skills rather than as auxiliary rewards in learning tasks with clearly distinguished skills, such as rearrangement. + +# 5.4. Robustness to Perturbations + +We conduct a perturbation test to evaluate the replanning ability of Skill Transformer and baselines in the rearrange-hard task. The test perturbs the environment by closing the target receptacle after the agent had opened it for the first time. To succeed, the agent must detect the drawer has been closed and adjust the current plan to reopen the receptacle and retrieve the object. We compare Skill Transformer's skill prediction and M3's task planner against ground truth skill labels derived from the privileged states of the simulation environment (detailed in Appendix C.3). The test is performed on 400 episodes that involved only objects in the closed drawer. This test compares against the highest-performing modular baseline with access to oracle task plans, M3 (Oracle), and the highest + +performing end-to-end baseline, DT (Skill). + +
MethodPicking Target ObjectPredicted Skill Accuracy
No PerturbPerturbNo PerturbPerturb
M3 (Oracle)24.000.5044.0010.00
DT (Skill)10.008.5039.0033.99
ST (Ours)26.7519.5060.0047.00
+ +Table 2: Success rates for picking up the target objects and accuracy of Skill Inference module and M3's oracle-plant task planner under both perturbed and unperturbed environment, measured by ground truth skill labels. + +The findings in Table 2 demonstrate that Skill Transformer outperforms the M3 (Oracle) baseline both in the success rate of picking up the object and skill selection accuracy. Specifically, when assessing the success rates of object retrieval, M3 (Oracle) performs comparably to Skill Transformer in the absence of perturbations, but its performance significantly deteriorates to nearly $0\%$ with perturbations in the environment. In contrast, Skill Transformer exhibits adaptability to changes in the state of the receptacle and the ability to re-open the receptacle, achieving a success rate for picking the object of $19.5\%$ in the presence of perturbations. Note that Skill Transformer is not trained with any perturbations, and this behavior emerges from the design choices made in Skill Transformer. + +Similarly, the accuracy of the predicted skills demonstrates that Skill Transformer outperforms M3 (Oracle) in high-level reasoning, both in the presence and absence of perturbations. Consistent with object retrieval, Skill Transformer demonstrates a notably higher accuracy, indicating a higher level of adaptability to perturbations, compared to M3 (Oracle), which suffers from the non-reactive task planner that does not correctly respond to perturbations. + +We further compare Skill Transformer to DT (Skill), another end-to-end method that has the potential to adapt to + +![](images/c6895b0eee1cb630d8d19d7d154750c8f35e1365f468b1cdd789c3e2eef7cdda.jpg) +(a) Architecture + +![](images/06efc4a75aa8c59bbff841d3140185eebb0e4072be0b3d7b8cf936a60bcbeef7.jpg) +(b) Context Length +Figure 3: Success rates for ablation and analysis experiments across different policy architectures (Figure 3a), context lengths (Figure 3b), and training dataset sizes (Figure 3c) including performance on "Eval" and "Train" on the easy split setting. + +![](images/48ab2462c259ae4ee067ab4ddd7dcadb75f6b6af004da18e67a8ee46514e278e.jpg) +(c) Dataset Size + +perturbations. DT (Skill) has a substantially lower success rate in picking the target object compared to Skill Transformer, with less than half the success rate in both perturbed and unperturbed environments. In contrast to object retrieval, DT (Skill) displays only slightly lower accuracy in predicting correct skill labels than Skill Transformer. Consistent with the qualitative findings in Section 5.3, these observations suggest that although DT (Skill) demonstrates relatively good high-level reasoning ability, its primary challenge remains in learning and effectively leveraging distinct skills. On the other hand, we find that Skill Transformer exhibits the combined advantages of adaptability to environment disturbances and proficiency in completing individual skills. + +# 5.5. Ablations and Analysis + +We run ablations to address two key questions. Firstly, how necessary are the design choices of Skill Transformer, such as the transformer backbone, the Skill Inference module, and the choice of context length? Secondly, how do the dataset characteristics, such as its size and completeness, impact the performance? We perform all of the ablation experiments on the "Easy" split of the task. + +# 5.5.1 Policy Architecture Ablation + +In Figure 3a, we evaluate the effectiveness of the transformer and the Skill Inference module for mobile manipulation tasks. By replacing the Transformer backbone with an LSTM backbone, we observe a significant decline in performance. We find that the policy struggles to effectively execute individual skills, and we posit that the LSTM architecture is not as data-efficient for learning multiple skills. Our results also indicate a similarly low success rate when the Skill Inference module is removed. Qualitatively, we discovered that without the Skill Inference module, the policy is not able to differentiate between individual skills, and frequently executes inconsistent actions. + +Context length refers to the maximum amount of historical information Skill Transformer takes as input. We vary + +the context length from 15 to 60 timesteps and evaluate its impact on performance. Shown in Figure 3b, a shorter context length of 15 timesteps performed slightly worse, while performance decreased as the context length increased from 30 to 60 timesteps. We posit that a shorter context length may not be able to fully capture the partially observable nature of the task, while a longer context length may lead to overfitting to the training dataset. Given the random robot start location during test time, the actual trajectory at test time will differ from that in the training dataset. The longer sequence exacerbates the discrepancy between the training dataset and the test time trajectory, resulting in an increased out-of-distribution error. A moderate context length may aid capturing short-term partially observed information and mitigating overfitting to the demonstration trajectories. The trend is similar to findings from previous works [49, 8]. + +# 5.5.2 Dataset Size Analysis + +As previously demonstrated [6], the size of the demonstration dataset is a vital aspect that impacts the efficacy of data-driven approaches such as Skill Transformer. As shown in Figure 3c, we trained Skill Transformer on subsets of the full demonstration dataset, ranging from 100 to 10k episodes. Our results showed that performance improved significantly when moving from 100 to 5k episodes, but the improvement is less pronounced when going from 5k to 10k demonstrations. + +# 5.5.3 Incomplete Demonstrations + +In long-horizon tasks, obtaining sufficient demonstrations of the full trajectory can be challenging. We find that Skill Transformer is also able to work effectively when trained with datasets containing only subsets of the full task. To demonstrate this, we train Skill Transformer on only two subtasks of the task: navigate-pick and navigate-place, and it achieves a success rate of $36.5\%$ in "Train" and $27.8\%$ in "Eval" split, compared to $44\%$ and $37\%$ using complete trajectories. This ablation shows that, in contrast to DT, which uses reward-conditioning across an entire episode, + +Skill Transformer can also work with partial and incomplete demonstrations. + +# 6. Conclusion + +This paper presents Skill Transformer, a novel end-to-end method for embodied object rearrangement tasks. Skill Transformer breaks the problem into predicting which skills to use and the low-level control conditioned on that skill, utilizing a causal transformer-based policy architecture. The proposed architecture outperforms baselines in challenging rearrangement problems requiring diverse skills and high-level reasoning over which skills to select. + +In future work, we will incorporate a more complex task specification, such as images of the desired state or natural language instructions. With the current geometric task specification, it is challenging to identify the target object after its container receptacle is manipulated. Specifically, because the object shifts inside a drawer when it is opened, its starting location no longer provides an indication of which objects is the target object. This limitation affects the success rate of retrieving objects in the challenging rearrange-hard task. + +# 7. Acknowledgments + +The Georgia Tech effort was supported in part by NSF, ONR YIP, and ARO PECASE. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the U.S. Government, or any sponsor. + +# References + +[1] Anurag Ajay, Yilun Du, Abhi Gupta, Joshua Tenenbaum, Tommi Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision-making? arXiv preprint arXiv:2211.15657, 2022. +[2] Anurag Ajay, Aviral Kumar, Pulkit Agrawal, Sergey Levine, and Ofir Nachum. Opal: Offline primitive discovery for accelerating offline reinforcement learning. arXiv preprint arXiv:2010.13611, 2020. +[3] Akhil Bagaria and George Konidaris. Option discovery using deep skill chaining. In International Conference on Learning Representations, 2019. +[4] Dhruv Batra, Angel X Chang, Sonia Chernova, Andrew J Davison, Jia Deng, Vladlen Koltun, Sergey Levine, Jitendra Malik, Igor Mordatch, Roozbeh Mottaghi, et al. Rearrangement: A challenge for embodied ai. arXiv preprint arXiv:2011.01975, 2020. +[5] Rogerio Bonatti, Sai Vemprala, Shuang Ma, Felipe Frujeri, Shuhang Chen, and Ashish Kapoor. Pact: Perception-action causal transformer for autoregressive robotics pre-training. arXiv preprint arXiv:2209.11133, 2022. + +[6] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. +[7] Micah Carroll, Jessy Lin, Orr Paradise, Raluca Georgescu, Mingfei Sun, David Bignell, Stephanie Milani, Katja Hofmann, Matthew Hausknecht, Anca Dragan, et al. Towards flexible inference in sequential decision problems via bidirectional transformers. arXiv preprint arXiv:2204.13326, 2022. +[8] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021. +[9] Alexander Clegg, Wenhao Yu, Jie Tan, C Karen Liu, and Greg Turk. Learning to dress: Synthesizing human dressing motion via deep reinforcement learning. ACM Transactions on Graphics (TOG), 37(6):1-10, 2018. +[10] Kiana Ehsani, Winson Han, Alvaro Herrasti, Eli VanderBilt, Luca Weihs, Eric Kolve, Aniruddha Kembhavi, and Roozbeh Mottaghi. Manipulathor: A framework for visual object manipulation. In CVPR, 2021. +[11] Benjamin Eysenbach, Abhishek Gupta, Julian Ibarz, and Sergey Levine. Diversity is all you need: Learning skills without a reward function. arXiv preprint arXiv:1802.06070, 2018. +[12] Pete Florence, Corey Lynch, Andy Zeng, Oscar A Ramirez, Ayzaan Wahid, Laura Downs, Adrian Wong, Johnny Lee, Igor Mordatch, and Jonathan Tompson. Implicit behavioral cloning. In Conference on Robot Learning, pages 158-168. PMLR, 2022. +[13] Roy Fox, Sanjay Krishnan, Ion Stoica, and Ken Goldberg. Multi-level discovery of deep options. arXiv preprint arXiv:1703.08294, 2017. +[14] Chuang Gan, Jeremy Schwartz, Seth Alter, Damian Mrowca, Martin Schrimpf, James Traer, Julian De Freitas, Jonas Kubilius, Abhishek Bhandwaldar, Nick Haber, et al. Threadworld: A platform for interactive multi-modal physical simulation. NeurIPS Datasets and Benchmarks Track, 2021. +[15] Caelan Reed Garrett, Rohan Chitnis, Rachel Holladay, Beomjoon Kim, Tom Silver, Leslie Pack Kaelbling, and Tomás Lozano-Pérez. Integrated task and motion planning. Annual review of control, robotics, and autonomous systems, 4:265-293, 2021. +[16] Jiayuan Gu, Devendra Singh Chaplot, Hao Su, and Jitendra Malik. Multi-skill mobile manipulation for object rearrangement. arXiv preprint arXiv:2209.02778, 2022. +[17] Kourosh Hakhamaneshi, Ruihan Zhao, Albert Zhan, Pieter Abbeel, and Michael Laskin. Hierarchical few-shot imitation with skill transition models. arXiv preprint arXiv:2107.08981, 2021. +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, 2016. +[19] Daniel Honerkamp, Tim Welschehold, and Abhinav Valada. Learning kinematic feasibility for mobile manipulation + +through deep reinforcement learning. IEEE Robotics and Automation Letters, 6(4):6289-6296, 2021. +[20] Michael Janner, Qiyang Li, and Sergey Levine. Reinforcement learning as one big sequence modeling problem. In ICML 2021 Workshop on Unsupervised Reinforcement Learning, 2021. +[21] Dmitry Kalashnikov, Alex Irpan, Peter Pastor, Julian Ibarz, Alexander Herzog, Eric Jang, Deirdre Quillen, Ethan Holly, Mrinal Kalakrishnan, Vincent Vanhoucke, and Sergey Levine. Scalable deep reinforcement learning for vision-based robotic manipulation. In 2nd Annual Conference on Robot Learning, CoRL 2018, Zürich, Switzerland, 29-31 October 2018, Proceedings, volume 87 of Proceedings of Machine Learning Research, pages 651-673. PMLR, 2018. +[22] George Konidaris and Andrew Barto. Skill discovery in continuous reinforcement learning domains using skill chaining. Advances in neural information processing systems, 22, 2009. +[23] Kuang-Huei Lee, Ofir Nachum, Mengjiao Yang, Lisa Lee, Daniel Freeman, Winnie Xu, Sergio Guadarrama, Ian Fischer, Eric Jang, Henryk Michalewski, et al. Multi-game decision transformers. arXiv preprint arXiv:2205.15241, 2022. +[24] Youngwoon Lee, Joseph J Lim, Anima Anandkumar, and Yuke Zhu. Adversarial skill chaining for long-horizon robot manipulation via terminal state regularization. arXiv preprint arXiv:2111.07999, 2021. +[25] Youngwoon Lee, Shao-Hua Sun, Sriram Somasundaram, Edward S Hu, and Joseph J Lim. Composing complex skills by learning transition policies. In International Conference on Learning Representations, 2018. +[26] Chengshu Li, Fei Xia, Roberto Martin-Martin, and Silvio Savarese. Hrl4in: Hierarchical reinforcement learning for interactive navigation with mobile manipulators. In Conference on Robot Learning, pages 603-616. PMLR, 2020. +[27] Anthony Liang, Ishika Singh, Karl Pertsch, and Jesse Thomason. Transformer adapters for robot learning. In CoRL 2022 Workshop on Pre-training Robot Learning, 2022. +[28] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. arXiv preprint arXiv:2209.07753, 2022. +[29] Maria Vittoria Minniti, Farbod Farshidian, Ruben Grandia, and Marco Hutter. Whole-body mpc for a dynamically stable mobile manipulator. IEEE Robotics and Automation Letters, 4(4):3687-3694, 2019. +[30] Mayank Mittal, David Hoeller, Farbod Farshidian, Marco Hutter, and Animesh Garg. Articulated object interaction in unknown scenes with whole-body mobile manipulation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1647-1654. IEEE, 2022. +[31] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. +[32] Ram Ramrakhya, Eric Undersander, Dhruv Batra, and Abhishek Das. Habitat-web: Learning embodied object-search strategies from human demonstrations at scale. In Proceed- + +ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5173-5183, 2022. +[33] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv preprint arXiv:2205.06175, 2022. +[34] Fetch robotics. Fetch. http://fetchrobotics.com/, 2020. +[35] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. +[36] Archit Sharma, Shixiang Gu, Sergey Levine, Vikash Kumar, and Karol Hausman. Dynamics-aware unsupervised discovery of skills. arXiv preprint arXiv:1907.01657, 2019. +[37] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In Conference on Robot Learning, pages 894–906. PMLR, 2022. +[38] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. arXiv preprint arXiv:2209.05451, 2022. +[39] Richard S Sutton, Doina Precup, and Satinder Singh. Between mdps and semi-mdps: A framework for temporal abstraction in reinforcement learning. Artificial intelligence, 112(1-2):181-211, 1999. +[40] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, et al. Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems, 34, 2021. +[41] Andrew Szot, Karmesh Yadav, Alex Clegg, Vincent-Pierre Berges, Aaron Gokaslan, Angel Chang, Manolis Savva, Zsolt Kira, and Dhruv Batra. Habitat rearrangement challenge 2022. https://aihabitat.org/challenge/2022_rearrange, 2022. +[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. +[43] Cong Wang, Qifeng Zhang, Qiyan Tian, Shuo Li, Xiaohui Wang, David Lane, Yvan Petillot, and Sen Wang. Learning mobile manipulation through deep reinforcement learning. Sensors, 20(3):939, 2020. +[44] Luca Weihs, Matt Deitke, Aniruddha Kembhavi, and Roozbeh Mottaghi. Visual room rearrangement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5922-5931, 2021. +[45] Erik Wijmans, Irfan Essa, and Dhruv Batra. Ver: Scaling on-policy rl leads to the emergence of navigation in embodied rearrangement. In Advances in Neural Information Processing Systems, 2022. +[46] Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. arXiv preprint arXiv:1911.00357, 2019. + +[47] Josiah Wong, Albert Tung, Andrey Kurenkov, Ajay Mandlekar, Li Fei-Fei, Silvio Savarese, and Roberto Martin-Martin. Error-aware imitation learning from teleoperation data for mobile manipulation. In Conference on Robot Learning, pages 1367–1378. PMLR, 2022. +[48] Mengdi Xu, Yikang Shen, Shun Zhang, Yuchen Lu, Ding Zhao, Joshua Tenenbaum, and Chuang Gan. Prompting decision transformer for few-shot policy generalization. In International Conference on Machine Learning, pages 24631-24645. PMLR, 2022. +[49] Qinqing Zheng, Amy Zhang, and Aditya Grover. Online decision transformer. arXiv preprint arXiv:2202.05607, 2022. \ No newline at end of file diff --git a/2023/Skill Transformer_ A Monolithic Policy for Mobile Manipulation/images.zip b/2023/Skill Transformer_ A Monolithic Policy for Mobile Manipulation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d5943f3ecdf40f50628db8c36067b440b643c48c --- /dev/null +++ b/2023/Skill Transformer_ A Monolithic Policy for Mobile Manipulation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a964f2593791877731c6067a8864c6cc2659df0f41d7eddfc8d1add457f1cd1 +size 294248 diff --git a/2023/Skill Transformer_ A Monolithic Policy for Mobile Manipulation/layout.json b/2023/Skill Transformer_ A Monolithic Policy for Mobile Manipulation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..253640eb2b4a1db8027f8d27e19fba4544a555e0 --- /dev/null +++ b/2023/Skill Transformer_ A Monolithic Policy for Mobile Manipulation/layout.json @@ -0,0 +1,8004 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 103, + 496, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 103, + 496, + 121 + ], + "spans": [ + { + "bbox": [ + 96, + 103, + 496, + 121 + ], + "type": "text", + "content": "Skill Transformer: A Monolithic Policy for Mobile Manipulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 144, + 146, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 144, + 146, + 174 + ], + "spans": [ + { + "bbox": [ + 74, + 144, + 146, + 174 + ], + "type": "text", + "content": "Xiaoyu Huang \nGeorgia Tech" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 200, + 144, + 278, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 144, + 278, + 156 + ], + "spans": [ + { + "bbox": [ + 200, + 144, + 278, + 156 + ], + "type": "text", + "content": "Dhruv Batra" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 171, + 158, + 323, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 158, + 323, + 172 + ], + "spans": [ + { + "bbox": [ + 171, + 158, + 323, + 172 + ], + "type": "text", + "content": "Meta AI (FAIR), Georgia Tech" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 356, + 144, + 418, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 144, + 418, + 156 + ], + "spans": [ + { + "bbox": [ + 356, + 144, + 418, + 156 + ], + "type": "text", + "content": "Akshara Rai" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 347, + 158, + 427, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 158, + 427, + 171 + ], + "spans": [ + { + "bbox": [ + 347, + 158, + 427, + 171 + ], + "type": "text", + "content": "Meta AI (FAIR)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 452, + 144, + 518, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 452, + 144, + 518, + 156 + ], + "spans": [ + { + "bbox": [ + 452, + 144, + 518, + 156 + ], + "type": "text", + "content": "Andrew Szot" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 452, + 158, + 518, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 452, + 158, + 518, + 171 + ], + "spans": [ + { + "bbox": [ + 452, + 158, + 518, + 171 + ], + "type": "text", + "content": "Georgia Tech" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 238, + 290, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 430 + ], + "type": "text", + "content": "We present Skill Transformer, an approach for solving long-horizon robotic tasks by combining conditional sequence modeling and skill modularity. Conditioned on ego-centric and proprioceptive observations of a robot, Skill Transformer is trained end-to-end to predict both a high-level skill (e.g., navigation, picking, placing), and a whole-body low-level action (e.g., base and arm motion), using a transformer architecture and demonstration trajectories that solve the full task. It retains the composability and modularity of the overall task through a skill predictor module while reasoning about low-level actions and avoiding hand-off errors, common in modular approaches. We test Skill Transformer on an embodied rearrangement benchmark and find it performs robust task planning and low-level control in new scenarios, achieving a " + }, + { + "bbox": [ + 46, + 238, + 290, + 430 + ], + "type": "inline_equation", + "content": "2.5x" + }, + { + "bbox": [ + 46, + 238, + 290, + 430 + ], + "type": "text", + "content": " higher success rate than baselines in hard rearrangement problems." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 465, + 128, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 465, + 128, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 465, + 128, + 478 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 485, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 288, + 700 + ], + "type": "text", + "content": "A long-standing goal of embodied AI is to build agents that can seamlessly operate in a home to accomplish tasks like preparing a meal, tidying up, or loading the dishwasher. To successfully carry out such tasks, an agent must be able to perform a wide range of diverse skills, like navigation and object manipulation, as well as high-level decision-making, all while operating with limited egocentric inputs. Our particular focus is the problem of embodied object rearrangement [41]. An example in Figure 1 shows that the Fetch mobile manipulator must rearrange a bowl from an initial position to a desired goal position (both circled in orange) in an unfamiliar environment without access to any pre-existing maps, 3D models, or precise object locations. Successfully completing this task requires executing both low-level motor commands for accurate arm and base movements, as well as high-level decision-making, such as determining whether a drawer needs to be opened before accessing a specific object." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 58, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 701, + 288, + 714 + ], + "type": "text", + "content": "Long-horizon tasks like object rearrangement require" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 213, + 547, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 441 + ], + "type": "text", + "content": "many low-level control steps and naturally break down into distinct phases of skills such as navigation, picking, placing, opening, and closing. Prior works [6, 33, 37] show that it is possible to end-to-end learn a single policy that can perform such diverse tasks and behaviors. Such a monolithic policy directly maps observations to low-level actions, reasoning both about which skill to execute and how to execute it. However, scaling end-to-end learning to long-horizon, multi-phase tasks with thousands of low-level steps and egocentric visual observations remains a challenging research question. Other works decompose the task into individual skills and then separately sequence them to solve the full task [40, 16]. While this decomposition makes the problem tractable, this modularity comes at a cost. Sequencing skills suffers from hand-off errors between skills, where one skill ends in a situation where the next skill is no longer feasible. Moreover, it does not address how to learn the skill sequencing or how to adapt the planned sequence to unknown disturbances." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 445, + 547, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 547, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 547, + 626 + ], + "type": "text", + "content": "We present Skill Transformer, an end-to-end trainable transformer policy for multi-skill tasks. Skill Transformer predicts both low-level actions and high-level skills, conditioned on a history of high-dimensional egocentric camera observations. By combining conditional sequence modeling with a skill prediction module, Skill Transformer can model long sequences of both high and low-level actions without breaking the problem into disconnected submodules. In Figure 1, Skill Transformer learns a policy that can perform diverse motor skills, as well as high-level reasoning over which motor skills to execute. We train Skill Transformer autoregressively with trajectories that solve the full rearrangement task. Skill Transformer is then tested in unseen settings that require a variable number of high-level and low-level actions, overcoming unknown perturbations." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Skill Transformer achieves state-of-the-art performance in rearrangement tasks requiring variable-length high-level planning and delicate low-level actions, such as rearranging an object that can be in an open or closed receptacle, in the Habitat environment [41]. It achieves " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "1.45\\mathrm{x}" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " higher success rates in the full task and up to " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "2.5\\mathrm{x}" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " higher in especially challenging situations over modular and end-to-end baselines" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10852" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 533, + 250 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 533, + 250 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 533, + 250 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 533, + 250 + ], + "type": "image", + "image_path": "c585f2d9a95a0b53c9fa21e88d8ca29e6b43f55bbc2ececd28d2d27456965367.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 260, + 548, + 309 + ], + "lines": [ + { + "bbox": [ + 46, + 260, + 548, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 548, + 309 + ], + "type": "text", + "content": "Figure 1: A Fetch robot is placed in an unseen house and must rearrange an object from a specified start to a desired goal location from egocentric visual observations. The robot infers the drawer is closed and opens the drawer to access the object. Next, it picks the object, navigates to the goal, and places the object. The Skill Transformer is a monolithic \"pixels-to-actions\" policy for such long-horizon mobile manipulation problems which infers the active skill and the low-level continuous control." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 319, + 288, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 288, + 392 + ], + "type": "text", + "content": "and is more robust to environment disturbances, such as a drawer snapping shut after opening it, despite never seeing such disturbances during training. Furthermore, we demonstrate the importance of the design decisions in Skill Transformer and its ability to scale with larger training dataset sizes. The code is available at https://bit.ly/3qH2QQK." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 404, + 134, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 404, + 134, + 417 + ], + "spans": [ + { + "bbox": [ + 47, + 404, + 134, + 417 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 688 + ], + "type": "text", + "content": "Object Rearrangement: Object rearrangement can be formulated as a task and motion planning [15] problem where an agent must move objects to a desired configuration, potentially under some constraints. We refer to [4] for a comprehensive survey. Prior works have also used end-to-end RL training for learning rearrangement [44, 32, 14, 10], but unlike our work, they rely on abstract interactions with the environment, no physics, or simplified action spaces. Other works focus on tabletop rearrangement [28, 21, 12] while our work focuses on longer horizon mobile manipulation. Although mobile manipulation has witnessed significant achievements in robotics learning [29, 43, 26, 30, 19], due to the vastly increased complexity, in rearrangement tasks, a majority of previous works [6, 40] explicitly separate the base and arm movements to simplify the task. A few works [16, 47, 45] have shown the superiority of mobile manipulation over static manipulation. In the Habitat Rearrangement Challenge [40, 41], methods that decompose the rearrangement task into skills are the most successful [16, 45] but they require a fixed high-level plan and cannot dynamically select skills based on observations during execution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Transformers for Control: Prior work uses a transformer policy architecture to approach RL as a sequence" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 319, + 547, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 547, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 547, + 439 + ], + "type": "text", + "content": "modeling problem [8, 49, 20, 7]. Our work also uses a transformer-based policy, but with a modified architecture that predicts high-level skills along with low-level actions. Other works demonstrate that transformer-based policies are competent multi-task learners [23, 48, 38, 6, 33]. Likewise, we examine how a transformer-based policy can learn multiple skills and compose them to solve long-horizon rearrangement tasks. [27, 5] further show that transformers can encode underlying dynamics of the environment that benefits learning for individual skills." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "content": "Skill Sequencing: Other works investigate how to sequence together skills to accomplish a long horizon task. By utilizing skills, agents can reason over longer time horizons by operating in the lifted action space of skills rather than over low-level actions [39]. Skills may be either learned from demonstrations [13, 2, 17], from trial and error learning [36, 11], or hand defined [25]. This work demonstrates that a single transformer policy can learn multiple skills. However, when sequencing together skills, there may be \"hand-off\" issues when skills fail to transition smoothly. Prior works tackle this issue by learning to match the initiation and termination sets of skills [24, 9] or learning new skills to handle the chaining [3, 22]. Decision Diffuser [1] learns a generative model to combine skills, but requires specifying the skill combination order. We overcome these issues by distilling all skills into a single policy and training this monolithic policy end-to-end to solve the overall task while maintaining skill modularity through a skill-prediction module." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "text", + "content": "10853" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 87, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 87, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 87, + 83 + ], + "type": "text", + "content": "3. Task" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 258 + ], + "type": "text", + "content": "We focus on the problem of embodied object rearrangement, where a robot is spawned in an unseen house and must rearrange an object from a specified start position to a desired goal position entirely from on-board sensing and without any pre-built maps, object 3D models, or other privileged information. We follow the setup of the 2022 Habitat Rearrangement Challenge [41]. Specifically, a simulated Fetch robot [34] senses the world through a " + }, + { + "bbox": [ + 46, + 91, + 287, + 258 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 91, + 287, + 258 + ], + "type": "text", + "content": " head-mounted RGB-D camera, robot joint positions, whether the robot is holding an object or not, and base egomotion giving the relative position of the robot since the start of the episode. The 3D position of the object to rearrange, along with the 3D position of where to rearrange the object to is also provided to the agent." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 259, + 287, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 259, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 287, + 343 + ], + "type": "text", + "content": "The agent interacts with the world via a mobile base, 7DoF arm, and suction grip attached to the arm. The episode is successful if the agent calls an episode termination action when the object is within " + }, + { + "bbox": [ + 46, + 259, + 287, + 343 + ], + "type": "inline_equation", + "content": "15\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 259, + 287, + 343 + ], + "type": "text", + "content": " of its target. The agent fails if it excessively collides with the environment or fails to rearrange the object within the prescribed low-level time steps. See Appendix B for more task details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 344, + 288, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 344, + 288, + 511 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 288, + 511 + ], + "type": "text", + "content": "We evaluate our method in the Rearrange-Hard Task, which presents challenges in both low-level actions and high-level reasoning. Unlike the Habitat Rearrangement tasks [16, 40] that prior works focus on, this task involves objects and goals that may spawn in closed receptacles. Therefore, the agent may need to perform intermediate actions, such as opening a drawer or fridge to access the target object or goal. Furthermore, the correct sequence of skills required to solve the task differs across episodes and is unavailable to the agent. As a result, the agent must infer the correct behaviors from egocentric visual observations and proprioception. This more natural setting better reflects real-world rearrangement scenarios, thus making it the focus of this work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 512, + 287, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 512, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 512, + 287, + 583 + ], + "type": "text", + "content": "From the dataset provided in [41], we create a harder version of the full Rearrange track dataset for evaluation. In this harder version, more episodes have objects and goals contained in closed receptacles. Specifically, the updated episode distribution in evaluation for the rearrange-hard task is divided into three splits:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 584, + 286, + 631 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 584, + 255, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 584, + 255, + 595 + ], + "spans": [ + { + "bbox": [ + 47, + 584, + 255, + 595 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 47, + 584, + 255, + 595 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 47, + 584, + 255, + 595 + ], + "type": "text", + "content": " easy scenes with objects initially accessible." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 596, + 264, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 264, + 607 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 264, + 607 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 47, + 596, + 264, + 607 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 47, + 596, + 264, + 607 + ], + "type": "text", + "content": " hard scenes with objects in closed receptacles." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 608, + 286, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 608, + 286, + 631 + ], + "spans": [ + { + "bbox": [ + 47, + 608, + 286, + 631 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 47, + 608, + 286, + 631 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 47, + 608, + 286, + 631 + ], + "type": "text", + "content": " very hard scenes with both objects and goals in closed receptacles." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 632, + 270, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 632, + 270, + 644 + ], + "spans": [ + { + "bbox": [ + 47, + 632, + 270, + 644 + ], + "type": "text", + "content": "The original rearrange-hard dataset is used for training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 657, + 102, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 102, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 102, + 669 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "We propose Skill Transformer, an approach for learning long-horizon, complex tasks that require sequencing multiple distinct skills. In Section 4.1, we describe the com" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "ponents of Skill Transformer, which include a Skill Inference module for reasoning over skill selections and an Action Inference module to predict actions for controlling the robot, both modeled using a Transformer architecture and conditioned on past and current observations. Next, in Section 4.2, we detail how we train Skill Transformer in an autoregressive fashion from demonstrations of the full task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 166, + 474, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 166, + 474, + 178 + ], + "spans": [ + { + "bbox": [ + 305, + 166, + 474, + 178 + ], + "type": "text", + "content": "4.1. Skill Transformer Architecture" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 185, + 545, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 545, + 268 + ], + "type": "text", + "content": "Skill Transformer breaks the long-horizon task of object rearrangement into first inferring skill selections using a Skill Inference module and then predicting low-level control using an Action Inference module, given this skill choice. Both modules are implemented as causal transformer neural networks [31], conditioned on past and current observations of the robot." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "text", + "content": "Input: Both modules in Skill Transformer take as input a sequence of tokens that encode robot observations. We refer to the maximum length of the input sequence as the context length " + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "inline_equation", + "content": "o_{t}" + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 1" + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "text", + "content": " depth visual observations, and " + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "inline_equation", + "content": "p_{t}" + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "text", + "content": " denote the proprioceptive inputs of the joint angles and holding state at time " + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "text", + "content": ". The visual observation " + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "inline_equation", + "content": "o_{t}" + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "text", + "content": " is first encoded with a Resnet-18 encoder [18] to a 256-dimension visual embedding. Likewise, non-visual inputs are encoded with a multi-layer perceptron network into a 256-dimension sensor embedding. The visual and non-visual continuous embeddings are concatenated to form a 512-dimension input token " + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "text", + "content": ". We also add a per-time step positional embedding to each input token " + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "text", + "content": " to enable modeling time across the sequence. " + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "text", + "content": " forms the input token that encodes the current observations, and Skill Transformer takes as input a history of such tokens up to context length " + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 269, + 546, + 496 + ], + "type": "text", + "content": ". Note that we do not use RGB input as prior work [40] found it unnecessary for geometric object rearrangement, and training with only depth input is faster." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "content": "Skill Inference: The first module of Skill Transformer is the Skill Inference module which predicts the current skill to be executed. The Skill Inference module takes as input a sequence of the input embeddings " + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\{x_{t - C},\\ldots ,x_t\\}" + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "content": " through a causal transformer network [42] with a 512 hidden dimension, to produce a one-hot skill vector " + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "content": " corresponding to the predicted skill that should be active at the current time step " + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "content": ". Using the causal mask, label " + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "content": " only has access to the previous tokens " + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\{t:t < i\\}" + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "content": " in the input sequence. Then, we embed " + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "content": " into a latent skill embedding " + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\{z_{t - C},\\dots ,z_t\\}" + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Action Inference: Next, the Action Inference module outputs a distribution over robot low-level actions. The Action Inference module takes as input the input embedding " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": ", an embedding of the previous time step action " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "a_{t-1}" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": ", and the skill embedding " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": ". Each of these inputs are represented as distinct tokens in the input sequence of " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{x_{t-C}, a_{t-C-1}, z_{t-C}, \\ldots, x_{t}, a_{t-1}, z_{t}\\}" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": ". This sequence is then fed into a transformer with the same architecture as the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "type": "text", + "content": "10854" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 71, + 521, + 242 + ], + "blocks": [ + { + "bbox": [ + 73, + 71, + 521, + 242 + ], + "lines": [ + { + "bbox": [ + 73, + 71, + 521, + 242 + ], + "spans": [ + { + "bbox": [ + 73, + 71, + 521, + 242 + ], + "type": "image", + "image_path": "5c95c374819662e68a55ffd084578f8f7ef856cbd80ea6191a348ef3244adf24.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 247, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 547, + 319 + ], + "type": "text", + "content": "Figure 2: The Skill Transformer architecture. Skill Transformer learns to map egocentric depth images from the Fetch robot's head camera and proprioceptive state to actions controlling the robot, including the base movement, arm movement, and gripper state. First, the raw visual observations are processed by a visual encoder and concatenated with the proprioceptive state. Next, the Skill Inference module, implemented as a causal transformer sequence model, infers the skill ID that is executing for each time step. Finally, the Action Inference module predicts the low-level action to execute on the robot from the inferred skill, observation, and previous action information. The model is trained through end-to-end imitation learning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 330, + 288, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 288, + 403 + ], + "type": "text", + "content": "Skill Inference module. By taking as input the skill embeddings " + }, + { + "bbox": [ + 46, + 330, + 288, + 403 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 330, + 288, + 403 + ], + "type": "text", + "content": ", the Action Inference module learns consistency in executing individual skills and the transitions between them. As we will show in Section 5.5, this also allows training Skill Transformer with datasets that only successfully complete a portion of the skills." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 415, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 594 + ], + "type": "text", + "content": "The Action Inference module then outputs a sequence of predicted next actions " + }, + { + "bbox": [ + 46, + 415, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\{a_{t - C},\\dots a_t\\}" + }, + { + "bbox": [ + 46, + 415, + 287, + 594 + ], + "type": "text", + "content": " which control the robot. These actions determine the whole body control actions composed of three components: arm joint positions, base velocities, and gripper state. The module first outputs a 12-dimension output vector. The first seven components are used as the mean of a normal distribution to sample the desired delta joint positions, which will then be actuated via a PD controller. The next two dimensions are used as the mean of a normal distribution to sample the desired base velocities, scaled by the maximum linear and angular velocities of the robot base. The remaining three outputs are passed through a softmax layer to generate probabilities for the gripper action, which indicate opening, closing, or maintaining the previous gripper state." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Unless otherwise specified, we use 2 transformer layers and a context length " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "C = 30" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " for all architectures, giving 2M parameters for the transformer network in the Skill Inference module. The Action Inference module is also trained with the same architecture and a context length " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " of 30 time steps, translating to a sequence length of 90 tokens. In total, between the observation encoders, Skill Inference, and Action Inference modules, the Skill Transformer policy network consists of 16M parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 330, + 457, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 330, + 457, + 342 + ], + "spans": [ + { + "bbox": [ + 305, + 330, + 457, + 342 + ], + "type": "text", + "content": "4.2. Skill Transformer Training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 347, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 545, + 443 + ], + "type": "text", + "content": "In this section, we describe how Skill Transformer is trained end-to-end to learn long-horizon rearrangement tasks from imitation learning. We train the policy using supervised learning on demonstration trajectories produced by a data-generating policy. We assume the data-generating policy demonstrates a composition of skills to solve the task. Such compositional approaches are the state-of-the-art for the Habitat Rearrangement task [40, 16]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "text", + "content": "Specifically, we assume the data-generating policy demonstrates one of " + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "text", + "content": " skills, and each time step " + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "text", + "content": " is labeled with the skill " + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}y_{t}^{\\mathrm{skill}}" + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "text", + "content": " that is currently executing. In our problem setting, these skills are: pick, place, navigate, open drawer, and open fridge. We also record the policy's actions at time step " + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "text", + "content": " that consists of arm action " + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "inline_equation", + "content": "y_{t}^{\\mathrm{a}}" + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "text", + "content": ", base velocity " + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "inline_equation", + "content": "y_{t}^{\\mathrm{b}}" + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "text", + "content": ", and gripper state " + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "inline_equation", + "content": "y_{t}^{\\mathrm{g}}" + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "text", + "content": ". In addition, we acquire from the simulator a one-dimensional indicator label " + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "inline_equation", + "content": "y_{t}^{\\mathrm{pred}}" + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "text", + "content": " defined as " + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "inline_equation", + "content": "y_{t}^{\\mathrm{pred}} = 0" + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "text", + "content": " if the object does not start in a closed receptacle and " + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "inline_equation", + "content": "y_{t}^{\\mathrm{pred}} = i" + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "text", + "content": " receptacle contains the object, at the time step " + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 444, + 546, + 579 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 593, + 399, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 593, + 399, + 605 + ], + "spans": [ + { + "bbox": [ + 306, + 593, + 399, + 605 + ], + "type": "text", + "content": "4.2.1 Training Loss" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 612, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 612, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 612, + 545, + 647 + ], + "type": "text", + "content": "We separately supervise the Skill Inference and Action Inference modules. We supervise the Skill Inference module over the context length " + }, + { + "bbox": [ + 304, + 612, + 545, + 647 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 612, + 545, + 647 + ], + "type": "text", + "content": " using loss function:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 377, + 647, + 545, + 678 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 647, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 377, + 647, + 545, + 678 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s k i l l}} = \\sum_ {t = 1} ^ {C} \\mathcal {L} _ {\\mathrm {F L}} \\left(c _ {t}, y _ {t} ^ {\\text {s k i l l}}\\right) \\tag {1}", + "image_path": "b5b344cba39a803964eb32d8c72ed6b40acad90b2900dc9c52dfb71700fb910f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FL}}" + }, + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "type": "text", + "content": " represents the focal loss. The focal loss counteracts the imbalance of skill execution frequencies. For example, the navigation skill executes more frequently than" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "text", + "content": "10855" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "the other skills as it is needed multiple times within a single episode, and takes more low-level time steps to execute than manipulation skills." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 286, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 286, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 286, + 180 + ], + "type": "text", + "content": "In order to train the Action Inference module in parallel with the Skill Inference module, we use teacher-forcing, and input the ground truth skill labels during training, instead of the inferred distributions that are used at test time. We supervise the Action Inference module over the context length " + }, + { + "bbox": [ + 46, + 109, + 286, + 180 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 109, + 286, + 180 + ], + "type": "text", + "content": " with loss function:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 180, + 286, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 286, + 211 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 286, + 211 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a c t}} = \\sum_ {t = 1} ^ {C} \\left(\\mathcal {L} _ {\\mathrm {M S E}} \\left(a _ {t} ^ {\\mathrm {b}}, y _ {t} ^ {\\mathrm {b}}\\right) + \\mathcal {L} _ {\\mathrm {M S E}} \\left(a _ {t} ^ {\\mathrm {a}}, y _ {t} ^ {\\mathrm {a}}\\right) + \\mathcal {L} _ {\\mathrm {F L}} \\left(a _ {t} ^ {\\mathrm {g}}, y _ {t} ^ {\\mathrm {g}}\\right)\\right) \\tag {2}", + "image_path": "56492925c25cd8d64d0106425a63b9ff351b55be37c93cb6ef9d289dd2ac4049.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 211, + 257, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 211, + 257, + 222 + ], + "spans": [ + { + "bbox": [ + 47, + 211, + 257, + 222 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 211, + 257, + 222 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 47, + 211, + 257, + 222 + ], + "type": "text", + "content": " represents the mean-squared error loss." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 223, + 286, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 223, + 286, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 286, + 305 + ], + "type": "text", + "content": "In addition, we introduce auxiliary heads and losses for the Skill Inference module. The auxiliary head reconstructs oracle information unavailable to the policy. Specifically, assuming there exist " + }, + { + "bbox": [ + 46, + 223, + 286, + 305 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 223, + 286, + 305 + ], + "type": "text", + "content": " receptacles that may contain the target object, the auxiliary head outputs an " + }, + { + "bbox": [ + 46, + 223, + 286, + 305 + ], + "type": "inline_equation", + "content": "M + 1" + }, + { + "bbox": [ + 46, + 223, + 286, + 305 + ], + "type": "text", + "content": " dimensional distribution " + }, + { + "bbox": [ + 46, + 223, + 286, + 305 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 46, + 223, + 286, + 305 + ], + "type": "text", + "content": ". We then supervise " + }, + { + "bbox": [ + 46, + 223, + 286, + 305 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 46, + 223, + 286, + 305 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 223, + 286, + 305 + ], + "type": "inline_equation", + "content": "y_{t}^{\\mathrm{pred}}" + }, + { + "bbox": [ + 46, + 223, + 286, + 305 + ], + "type": "text", + "content": " over the context length " + }, + { + "bbox": [ + 46, + 223, + 286, + 305 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 223, + 286, + 305 + ], + "type": "text", + "content": " via the loss function:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 120, + 306, + 286, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 306, + 286, + 338 + ], + "spans": [ + { + "bbox": [ + 120, + 306, + 286, + 338 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a u x}} = \\sum_ {t = 1} ^ {C} \\mathcal {L} _ {\\mathrm {C E}} \\left(v _ {t}, y _ {t} ^ {\\text {p r e d}}\\right) \\tag {3}", + "image_path": "b8297b6689a13b441c80bd14ab327435ea4f1036df353d31080c1e261dcaf278.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 338, + 231, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 231, + 349 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 231, + 349 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 338, + 231, + 349 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 47, + 338, + 231, + 349 + ], + "type": "text", + "content": " represents the Cross-Entropy loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 350, + 286, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 286, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 286, + 480 + ], + "type": "text", + "content": "We incorporate this auxiliary loss because the data-generating policy used to generate the demonstrations has a low success rate in correctly opening the receptacles. This low success rate can hinder learning because Skill Transformer needs to reason if it failed due to not opening the receptacle at all or opening an incorrect receptacle. The auxiliary loss incentivizes Skill Transformer to identify the key features that lead to the decision to open specific receptacles and to avoid wrong ones. This auxiliary objective with privileged information is only used for training and not used at test time." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 482, + 286, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 482, + 286, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 286, + 625 + ], + "type": "text", + "content": "The final loss is the unweighted sum of individual loss terms: " + }, + { + "bbox": [ + 46, + 482, + 286, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{\\mathrm{skill}} + \\mathcal{L}_{\\mathrm{action}} + \\mathcal{L}_{\\mathrm{aux}}" + }, + { + "bbox": [ + 46, + 482, + 286, + 625 + ], + "type": "text", + "content": ". Note that in training, we calculate the loss over all time steps, whereas at test time, we only output action " + }, + { + "bbox": [ + 46, + 482, + 286, + 625 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 46, + 482, + 286, + 625 + ], + "type": "text", + "content": " for the current time step " + }, + { + "bbox": [ + 46, + 482, + 286, + 625 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 482, + 286, + 625 + ], + "type": "text", + "content": " to the environment. While training, we also train the visual encoder, in contrast to previous works using transformers in embodied AI that only demonstrate performance with pre-trained visual encoders [8, 6]. Since Skill Transformer is an end-to-end trainable policy, it can also be trained with offline-RL using the rewards from the dataset or with online-RL by interacting with the simulator. We leave exploring Skill Transformer with RL training for future work." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 645, + 152, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 152, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 152, + 657 + ], + "type": "text", + "content": "4.2.2 Training Scheme" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "We use a similar scheme as [8], where we input a sequence of 30 timesteps and compute the loss for the skill and action distributions for all timesteps in the sequence using autoregressive masking. Skill Transformer uses a learning warm-" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "up phase where the learning rate is linearly increased at the start of training. After the warm-up phase, Skill Transformer uses a cosine learning rate scheduler which decays to 0 at the end of training. The entire dataset was iterated through approximately 6 times in total during the complete training process. We report the performance of the checkpoint with the highest success rate on seen scenes. We provide further details about Skill Transformer in Appendix A." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 178, + 386, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 178, + 386, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 386, + 190 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 198, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 317 + ], + "type": "text", + "content": "In this section, we describe a series of experiments used to validate our proposed approach. First, we introduce the experiment setup and state-of-the-art baselines for comparison. Then, we quantify the performance of Skill Transformer in the rearrange-hard task and compare it to the baselines. Furthermore, the robustness of the methods to disturbances at evaluation time are evaluated by adversarial perturbations, like closing drawers after the agent has opened them. Finally, we perform an ablation study to confirm the key design choices involved in Skill Transformer." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 324, + 413, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 324, + 413, + 336 + ], + "spans": [ + { + "bbox": [ + 306, + 324, + 413, + 336 + ], + "type": "text", + "content": "5.1. Experiment Setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 342, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 545 + ], + "type": "text", + "content": "We conduct all experiments in the Habitat 2.0 [40] simulation environment and the rearrangement tasks as described in Section 3. For training, we use the train dataset split from the challenge consisting of 50,000 object configurations across 63 scenes. Each episode consists of starting object positions and a target location for the object to be rearranged. The robot positions are randomly generated in each episode. We first train individual skills (pick, place, navigate, open, close) that are used by a data generating policy to generate 10,000 demonstrations for training Skill Transformer. To train skills, we use the Multi-skill Mobile Manipulation (M3) [16] method with an oracle task planner with privileged information about the environment. This approach trains skills with RL and sequences them with an oracle planner. Further details of the data generating policy and the demonstration collected are detailed in Appendix C.1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 552, + 372, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 372, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 372, + 562 + ], + "type": "text", + "content": "5.2. Baselines" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 616 + ], + "type": "text", + "content": "We compare Skill Transformer (ST) to the following baselines in the rearrange-hard task. The parameter count of each baseline policy is adjusted to match the 16M parameters in Skill Transformer." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 618, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 306, + 618, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 618, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 306, + 618, + 545, + 677 + ], + "type": "text", + "content": "- Monolithic RL (Mono): Train a monolithic neural network directly mapping sensor inputs to actions with end-to-end RL using DD-PPO [46], a distributed form of PPO [35]. The monolithic RL policy is trained for 100M timesteps, reflecting a wall time of 21h." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": "- Decision Transformer (DT) [8]: Train the monolithic policy from the offline dataset using a decision transformer architecture. The DT policy is trained with the" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "text", + "content": "10856" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 286, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 286, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 286, + 107 + ], + "type": "text", + "content": "same learning rate scheme and number of epochs as Skill Transformer. At test time, we condition the DT policy with the highest possible reward in the dataset." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 108, + 286, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 108, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 108, + 286, + 167 + ], + "type": "text", + "content": "- Decision Transformer with Auxiliary Skill Loss (DT (Skill)): Train the Decision Transformer with an auxiliary loss of skill prediction. This baseline incorporates an auxiliary head that predicts the current skill during training and is trained with focal loss like Skill Transformer." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 168, + 286, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 168, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 286, + 239 + ], + "type": "text", + "content": "Note that Decision Transformer requires the environment reward at evaluation time to compute the reward-to-go input. Skill Transformer and the other baselines do not require this assumption. We also compare to the following hierarchical baselines which first decompose the task into skills and then sequence these skills to solve the tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 240, + 286, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 240, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 286, + 312 + ], + "type": "text", + "content": "- Task-Planning + Skills RL (TP-SRL) [40]: Individually train pick, place, navigate, and open skills using RL and then sequence the skills together with a fixed task planner. Each skill predicts when it should end and the next skill should begin by a skill termination action. TP-SRL trains all manipulation skills with a fixed base." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 312, + 286, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 286, + 372 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 286, + 372 + ], + "type": "text", + "content": "- Multi-skill Mobile Manipulation (M3) [16]: M3 is a modular approach similar to TP-SRL, but instead trains skills as mobile manipulation skills with RL. All these skills are trained to be robust to diverse starting states. Like TP-SRL, it uses a fixed task planner to sequence the skills." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 373, + 286, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 286, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 286, + 467 + ], + "type": "text", + "content": "- Behavior Cloning Modular skills (BC-Modular): Individually train transformer-based policies for each of the skills using behavior cloning with the same training dataset as Skill Transformer. Then sequence together with the same fixed task planner as the other modular baselines. This baseline helps measure the gap in performance from training policies with behavior cloning as opposed to the RL-trained policies, M3 and TP-SRL." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 468, + 286, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 286, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 286, + 540 + ], + "type": "text", + "content": "However, these modular baselines are designed to only execute a fixed pre-defined sequence of skills (navigate, pick, navigate, place). They cannot adjust their skill sequence based on environment observations, i.e. the task plan is inaccurate when the robot needs to open a receptacle to pick or place an object." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 541, + 286, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 541, + 286, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 541, + 286, + 661 + ], + "type": "text", + "content": "To address this, we additionally include a version of the modular baselines, M3 (Oracle) and BC-Modular (Oracle), that extracts the correct task plan directly from the simulator, which we refer to as an oracle plan. In contrast, the proposed method, Skill Transformer, infers which skill to execute from egocentric observations without using this privileged oracle plan. Note that neither modular baseline re-executes skills if they fail, while Skill Transformer can re-plan. For further details about baselines and all hyperparameters, see Appendix C.2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 670, + 148, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 148, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 148, + 682 + ], + "type": "text", + "content": "5.3. Rearrange-Hard" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "First, we show the performance of Skill Transformer on the rearrange-hard task described in Section 3. In this task," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": "the agent must rearrange a single object from the start to the goal, and the object or the goal may be located in a closed receptacle. Therefore, methods need high-level reasoning to decide if the receptacle should first be opened or not." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 121, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 121, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 121, + 545, + 300 + ], + "type": "text", + "content": "We evaluate the success rate of all methods on both \"Train\" (Seen) and \"Eval\" (Unseen) scenes and object configurations. The success rate is measured as the average number of episodes where the agent placed the target object within " + }, + { + "bbox": [ + 305, + 121, + 545, + 300 + ], + "type": "inline_equation", + "content": "15\\mathrm{cm}" + }, + { + "bbox": [ + 305, + 121, + 545, + 300 + ], + "type": "text", + "content": " of the goal within 5000 steps and without excessively colliding with the environment. Specifically, the episode will terminate with failure if the agent accumulates more than " + }, + { + "bbox": [ + 305, + 121, + 545, + 300 + ], + "type": "inline_equation", + "content": "100\\mathrm{kN}" + }, + { + "bbox": [ + 305, + 121, + 545, + 300 + ], + "type": "text", + "content": " of force throughout the episode or " + }, + { + "bbox": [ + 305, + 121, + 545, + 300 + ], + "type": "inline_equation", + "content": "10\\mathrm{kN}" + }, + { + "bbox": [ + 305, + 121, + 545, + 300 + ], + "type": "text", + "content": " of instantaneous force due to collisions with the scene or objects. We also report the success rate on each of the three dataset splits described in Section 3. Our evaluation dataset splits have 100 episodes in the easy split, 200 in the hard split (100 for objects starting in closed receptacles and 100 for goals in closed receptacles), and 100 in the very hard split." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 302, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 302, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 302, + 545, + 374 + ], + "type": "text", + "content": "Table 1 shows that Skill Transformer achieves higher success rates in both the \"Train\" and \"Eval\" settings than all of the baselines that do not have oracle planning. The Mono baseline cannot find any success, even after training for 100M steps. This demonstrates the difficulty of learning the rearrange-hard task from a reward signal alone." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 375, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 375, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 305, + 375, + 545, + 495 + ], + "type": "text", + "content": "The modular TP-SRL, M3, and BC-Modular methods rely on a fixed task plan to execute the separately learned skill policies. As a result, the state-of-the-art M3 method achieves a significantly lower success rate than Skill Transformer in the overall result as well as the hard and very hard splits because it is unable to adapt the skill sequence to different situations. On the other hand, Skill Transformer does not rely on a fixed ordering of skills and can adapt its high-level task-solving to new situations, achieving a " + }, + { + "bbox": [ + 305, + 375, + 545, + 495 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 305, + 375, + 545, + 495 + ], + "type": "text", + "content": " increase in success rate across all episodes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 497, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 497, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 305, + 497, + 545, + 675 + ], + "type": "text", + "content": "Notice that even with an oracle plan, the success rate of M3 (Oracle) on the overall result for unseen scenes is only " + }, + { + "bbox": [ + 305, + 497, + 545, + 675 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 305, + 497, + 545, + 675 + ], + "type": "text", + "content": " higher than that of Skill Transformer. More importantly, Skill Transformer outperforms M3 (Oracle) in both hard and very hard settings, even without oracle task planning information. We find that Skill Transformer consistently achieves higher success rates in episodes with hidden objects or goals, while M3 (Oracle) experiences a substantial decrement as the difficulty increases. This difference is likely due to the low success rate in M3 (Oracle)'s opening skill, hindering its ability to complete the skill in one attempt. In contrast, as we will discuss quantitatively later in Section 5.4, Skill Transformer is capable of detecting and retrying failed skills, eventually leading to better performance in retrieving and placing the object." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Further observe that BC-Modular exhibits the lowest performance among modular methods, with an overall success rate of only " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " even when employing oracle task" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "text", + "content": "10857" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 529, + 194 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 529, + 194 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 529, + 194 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 529, + 194 + ], + "type": "table", + "html": "
MethodTrain (Seen)Eval (Unseen)
All EpisodesEasyHardVery HardAll EpisodesEasyHardVery Hard
BC-Modular (Oracle)5.67±0.1220.33±1.700.83±0.240.00±0.005.00±0.3515.33±1.701.67±1.650.33±0.47
M3 (Oracle)22.92±0.7259.33±1.7013.00±1.416.33±0.4722.58±1.1858.67±1.8912.67±2.366.33±0.47
BC-Modular5.08±0.4220.33±1.700.00±0.000.00±0.003.83±0.4215.33±1.700.00±0.000.00±0.00
TP-SRL7.25±0.8229.00±3.270.00±0.000.00±0.006.67±0.8526.67±3.400.00±0.000.00±0.00
M314.00±0.3555.67±0.940.16±0.470.00±0.0013.25±1.0853.00±4.320.00±0.000.00±0.00
Mono0.00±0.000.00±0.000.00±0.000.00±0.000.00±0.000.00±0.000.00±0.000.00±0.00
DT7.08±0.6319.33±0.473.83±0.622.00±0.946.50±0.4115.33±1.253.83±0.623.00±1.63
DT (Skill)8.17±0.6723.33±2.493.00±2.163.33±0.947.67±0.6618.67±2.624.83±1.702.33±1.25
ST (Ours)21.08±1.3644.00±0.8215.67±1.709.00±2.1619.17±1.0537.00±0.8215.83±0.858.00±3.27
", + "image_path": "172f9ffce372c303b3b19667ad66d11865dfd649fe8dce75e538805d1a7d4de8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 202, + 547, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 547, + 264 + ], + "type": "text", + "content": "Table 1: Success rates on the rearrangement task. \"Eval\" results are from test episodes in unseen scenes. In \"Easy\", objects and goals start on accessible locations. In \"Hard\", objects start in closed receptacles. In \"Very Hard\" objects and goals start in closed receptacles. All results are averages across 100 episodes except for the \"Hard\" split which is an average across 200 episodes. Grayed methods at the top use the oracle skill order. We report the mean and standard deviation across 3 random training seeds. For hierarchical methods, we retrain all the skills for each random seed." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 284, + 288, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 288, + 368 + ], + "type": "text", + "content": "planning. This indicates that policies trained via Behavior Cloning, which uses only one-tenth the experience of those trained with RL, encounter greater difficulty in executing intricate mobile manipulation skills. Conversely, Skill Transformer, featuring the multi-skill Action Inference module, mitigates this limitation and achieves up to 4x performance improvements across all episodes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 369, + 288, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 288, + 513 + ], + "type": "text", + "content": "DT and DT (Skill) can also achieve some success in hard and very hard splits. Nonetheless, the success rates are significantly lower than Skill Transformer, especially in the very hard split. We observe that DT-based methods lack the capability to execute skills consistently and often mix up different skill behaviors. Also, notice that DT (Skill) with an auxiliary skill loss does not show a significant improvement in performance compared to DT without the skill loss. These observations highlight the superiority of directly conditioning on skills rather than as auxiliary rewards in learning tasks with clearly distinguished skills, such as rearrangement." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 525, + 203, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 203, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 203, + 538 + ], + "type": "text", + "content": "5.4. Robustness to Perturbations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "We conduct a perturbation test to evaluate the replanning ability of Skill Transformer and baselines in the rearrange-hard task. The test perturbs the environment by closing the target receptacle after the agent had opened it for the first time. To succeed, the agent must detect the drawer has been closed and adjust the current plan to reopen the receptacle and retrieve the object. We compare Skill Transformer's skill prediction and M3's task planner against ground truth skill labels derived from the privileged states of the simulation environment (detailed in Appendix C.3). The test is performed on 400 episodes that involved only objects in the closed drawer. This test compares against the highest-performing modular baseline with access to oracle task plans, M3 (Oracle), and the highest" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 304, + 541, + 369 + ], + "blocks": [ + { + "bbox": [ + 306, + 284, + 484, + 297 + ], + "lines": [ + { + "bbox": [ + 306, + 284, + 484, + 297 + ], + "spans": [ + { + "bbox": [ + 306, + 284, + 484, + 297 + ], + "type": "text", + "content": "performing end-to-end baseline, DT (Skill)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 304, + 541, + 369 + ], + "lines": [ + { + "bbox": [ + 309, + 304, + 541, + 369 + ], + "spans": [ + { + "bbox": [ + 309, + 304, + 541, + 369 + ], + "type": "table", + "html": "
MethodPicking Target ObjectPredicted Skill Accuracy
No PerturbPerturbNo PerturbPerturb
M3 (Oracle)24.000.5044.0010.00
DT (Skill)10.008.5039.0033.99
ST (Ours)26.7519.5060.0047.00
", + "image_path": "07b647e2f17b36c424d08f50fb8d81870e8ac20456e4de08cdec2279c74edd56.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 376, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 424 + ], + "type": "text", + "content": "Table 2: Success rates for picking up the target objects and accuracy of Skill Inference module and M3's oracle-plant task planner under both perturbed and unperturbed environment, measured by ground truth skill labels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 426, + 546, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 546, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 546, + 593 + ], + "type": "text", + "content": "The findings in Table 2 demonstrate that Skill Transformer outperforms the M3 (Oracle) baseline both in the success rate of picking up the object and skill selection accuracy. Specifically, when assessing the success rates of object retrieval, M3 (Oracle) performs comparably to Skill Transformer in the absence of perturbations, but its performance significantly deteriorates to nearly " + }, + { + "bbox": [ + 304, + 426, + 546, + 593 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 304, + 426, + 546, + 593 + ], + "type": "text", + "content": " with perturbations in the environment. In contrast, Skill Transformer exhibits adaptability to changes in the state of the receptacle and the ability to re-open the receptacle, achieving a success rate for picking the object of " + }, + { + "bbox": [ + 304, + 426, + 546, + 593 + ], + "type": "inline_equation", + "content": "19.5\\%" + }, + { + "bbox": [ + 304, + 426, + 546, + 593 + ], + "type": "text", + "content": " in the presence of perturbations. Note that Skill Transformer is not trained with any perturbations, and this behavior emerges from the design choices made in Skill Transformer." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 594, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 689 + ], + "type": "text", + "content": "Similarly, the accuracy of the predicted skills demonstrates that Skill Transformer outperforms M3 (Oracle) in high-level reasoning, both in the presence and absence of perturbations. Consistent with object retrieval, Skill Transformer demonstrates a notably higher accuracy, indicating a higher level of adaptability to perturbations, compared to M3 (Oracle), which suffers from the non-reactive task planner that does not correctly respond to perturbations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "content": "We further compare Skill Transformer to DT (Skill), another end-to-end method that has the potential to adapt to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "type": "text", + "content": "10858" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 71, + 225, + 175 + ], + "blocks": [ + { + "bbox": [ + 85, + 71, + 225, + 175 + ], + "lines": [ + { + "bbox": [ + 85, + 71, + 225, + 175 + ], + "spans": [ + { + "bbox": [ + 85, + 71, + 225, + 175 + ], + "type": "image", + "image_path": "c6895b0eee1cb630d8d19d7d154750c8f35e1365f468b1cdd789c3e2eef7cdda.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 125, + 178, + 184, + 187 + ], + "lines": [ + { + "bbox": [ + 125, + 178, + 184, + 187 + ], + "spans": [ + { + "bbox": [ + 125, + 178, + 184, + 187 + ], + "type": "text", + "content": "(a) Architecture" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 227, + 72, + 365, + 174 + ], + "blocks": [ + { + "bbox": [ + 227, + 72, + 365, + 174 + ], + "lines": [ + { + "bbox": [ + 227, + 72, + 365, + 174 + ], + "spans": [ + { + "bbox": [ + 227, + 72, + 365, + 174 + ], + "type": "image", + "image_path": "06efc4a75aa8c59bbff841d3140185eebb0e4072be0b3d7b8cf936a60bcbeef7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 177, + 331, + 188 + ], + "lines": [ + { + "bbox": [ + 260, + 177, + 331, + 188 + ], + "spans": [ + { + "bbox": [ + 260, + 177, + 331, + 188 + ], + "type": "text", + "content": "(b) Context Length" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 193, + 545, + 218 + ], + "lines": [ + { + "bbox": [ + 46, + 193, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 545, + 218 + ], + "type": "text", + "content": "Figure 3: Success rates for ablation and analysis experiments across different policy architectures (Figure 3a), context lengths (Figure 3b), and training dataset sizes (Figure 3c) including performance on \"Eval\" and \"Train\" on the easy split setting." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 368, + 72, + 506, + 174 + ], + "blocks": [ + { + "bbox": [ + 368, + 72, + 506, + 174 + ], + "lines": [ + { + "bbox": [ + 368, + 72, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 368, + 72, + 506, + 174 + ], + "type": "image", + "image_path": "48ab2462c259ae4ee067ab4ddd7dcadb75f6b6af004da18e67a8ee46514e278e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 178, + 466, + 187 + ], + "lines": [ + { + "bbox": [ + 408, + 178, + 466, + 187 + ], + "spans": [ + { + "bbox": [ + 408, + 178, + 466, + 187 + ], + "type": "text", + "content": "(c) Dataset Size" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 238, + 288, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 288, + 406 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 288, + 406 + ], + "type": "text", + "content": "perturbations. DT (Skill) has a substantially lower success rate in picking the target object compared to Skill Transformer, with less than half the success rate in both perturbed and unperturbed environments. In contrast to object retrieval, DT (Skill) displays only slightly lower accuracy in predicting correct skill labels than Skill Transformer. Consistent with the qualitative findings in Section 5.3, these observations suggest that although DT (Skill) demonstrates relatively good high-level reasoning ability, its primary challenge remains in learning and effectively leveraging distinct skills. On the other hand, we find that Skill Transformer exhibits the combined advantages of adaptability to environment disturbances and proficiency in completing individual skills." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 411, + 178, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 178, + 424 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 178, + 424 + ], + "type": "text", + "content": "5.5. Ablations and Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 430, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 514 + ], + "type": "text", + "content": "We run ablations to address two key questions. Firstly, how necessary are the design choices of Skill Transformer, such as the transformer backbone, the Skill Inference module, and the choice of context length? Secondly, how do the dataset characteristics, such as its size and completeness, impact the performance? We perform all of the ablation experiments on the \"Easy\" split of the task." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 527, + 203, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 203, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 203, + 539 + ], + "type": "text", + "content": "5.5.1 Policy Architecture Ablation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "content": "In Figure 3a, we evaluate the effectiveness of the transformer and the Skill Inference module for mobile manipulation tasks. By replacing the Transformer backbone with an LSTM backbone, we observe a significant decline in performance. We find that the policy struggles to effectively execute individual skills, and we posit that the LSTM architecture is not as data-efficient for learning multiple skills. Our results also indicate a similarly low success rate when the Skill Inference module is removed. Qualitatively, we discovered that without the Skill Inference module, the policy is not able to differentiate between individual skills, and frequently executes inconsistent actions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Context length refers to the maximum amount of historical information Skill Transformer takes as input. We vary" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 238, + 547, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 238, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 238, + 547, + 430 + ], + "type": "text", + "content": "the context length from 15 to 60 timesteps and evaluate its impact on performance. Shown in Figure 3b, a shorter context length of 15 timesteps performed slightly worse, while performance decreased as the context length increased from 30 to 60 timesteps. We posit that a shorter context length may not be able to fully capture the partially observable nature of the task, while a longer context length may lead to overfitting to the training dataset. Given the random robot start location during test time, the actual trajectory at test time will differ from that in the training dataset. The longer sequence exacerbates the discrepancy between the training dataset and the test time trajectory, resulting in an increased out-of-distribution error. A moderate context length may aid capturing short-term partially observed information and mitigating overfitting to the demonstration trajectories. The trend is similar to findings from previous works [49, 8]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 444, + 430, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 444, + 430, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 430, + 456 + ], + "type": "text", + "content": "5.5.2 Dataset Size Analysis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 463, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 569 + ], + "type": "text", + "content": "As previously demonstrated [6], the size of the demonstration dataset is a vital aspect that impacts the efficacy of data-driven approaches such as Skill Transformer. As shown in Figure 3c, we trained Skill Transformer on subsets of the full demonstration dataset, ranging from 100 to 10k episodes. Our results showed that performance improved significantly when moving from 100 to 5k episodes, but the improvement is less pronounced when going from 5k to 10k demonstrations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 580, + 457, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 580, + 457, + 591 + ], + "spans": [ + { + "bbox": [ + 306, + 580, + 457, + 591 + ], + "type": "text", + "content": "5.5.3 Incomplete Demonstrations" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "In long-horizon tasks, obtaining sufficient demonstrations of the full trajectory can be challenging. We find that Skill Transformer is also able to work effectively when trained with datasets containing only subsets of the full task. To demonstrate this, we train Skill Transformer on only two subtasks of the task: navigate-pick and navigate-place, and it achieves a success rate of " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "36.5\\%" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " in \"Train\" and " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "27.8\\%" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " in \"Eval\" split, compared to " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "44\\%" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "37\\%" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " using complete trajectories. This ablation shows that, in contrast to DT, which uses reward-conditioning across an entire episode," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "text", + "content": "10859" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "Skill Transformer can also work with partial and incomplete demonstrations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 110, + 119, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 110, + 119, + 121 + ], + "spans": [ + { + "bbox": [ + 47, + 110, + 119, + 121 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 130, + 287, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 130, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 130, + 287, + 226 + ], + "type": "text", + "content": "This paper presents Skill Transformer, a novel end-to-end method for embodied object rearrangement tasks. Skill Transformer breaks the problem into predicting which skills to use and the low-level control conditioned on that skill, utilizing a causal transformer-based policy architecture. The proposed architecture outperforms baselines in challenging rearrangement problems requiring diverse skills and high-level reasoning over which skills to select." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 227, + 287, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 227, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 287, + 345 + ], + "type": "text", + "content": "In future work, we will incorporate a more complex task specification, such as images of the desired state or natural language instructions. With the current geometric task specification, it is challenging to identify the target object after its container receptacle is manipulated. Specifically, because the object shifts inside a drawer when it is opened, its starting location no longer provides an indication of which objects is the target object. This limitation affects the success rate of retrieving objects in the challenging rearrange-hard task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 360, + 156, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 156, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 156, + 373 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 380, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 453 + ], + "type": "text", + "content": "The Georgia Tech effort was supported in part by NSF, ONR YIP, and ARO PECASE. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the U.S. Government, or any sponsor." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 466, + 105, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 105, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 105, + 478 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 52, + 486, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 52, + 486, + 287, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 486, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 52, + 486, + 287, + 530 + ], + "type": "text", + "content": "[1] Anurag Ajay, Yilun Du, Abhi Gupta, Joshua Tenenbaum, Tommi Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision-making? arXiv preprint arXiv:2211.15657, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 532, + 287, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 532, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 52, + 532, + 287, + 575 + ], + "type": "text", + "content": "[2] Anurag Ajay, Aviral Kumar, Pulkit Agrawal, Sergey Levine, and Ofir Nachum. Opal: Offline primitive discovery for accelerating offline reinforcement learning. arXiv preprint arXiv:2010.13611, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 52, + 578, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 578, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 52, + 578, + 287, + 610 + ], + "type": "text", + "content": "[3] Akhil Bagaria and George Konidaris. Option discovery using deep skill chaining. In International Conference on Learning Representations, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 52, + 613, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 613, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 52, + 613, + 287, + 666 + ], + "type": "text", + "content": "[4] Dhruv Batra, Angel X Chang, Sonia Chernova, Andrew J Davison, Jia Deng, Vladlen Koltun, Sergey Levine, Jitendra Malik, Igor Mordatch, Roozbeh Mottaghi, et al. Rearrangement: A challenge for embodied ai. arXiv preprint arXiv:2011.01975, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 52, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 52, + 669, + 287, + 712 + ], + "type": "text", + "content": "[5] Rogerio Bonatti, Sai Vemprala, Shuang Ma, Felipe Frujeri, Shuhang Chen, and Ashish Kapoor. Pact: Perception-action causal transformer for autoregressive robotics pre-training. arXiv preprint arXiv:2209.11133, 2022." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 311, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 311, + 73, + 545, + 127 + ], + "type": "text", + "content": "[6] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 129, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 129, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 311, + 129, + 545, + 193 + ], + "type": "text", + "content": "[7] Micah Carroll, Jessy Lin, Orr Paradise, Raluca Georgescu, Mingfei Sun, David Bignell, Stephanie Milani, Katja Hofmann, Matthew Hausknecht, Anca Dragan, et al. Towards flexible inference in sequential decision problems via bidirectional transformers. arXiv preprint arXiv:2204.13326, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 196, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 196, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 311, + 196, + 545, + 251 + ], + "type": "text", + "content": "[8] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 252, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 252, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 311, + 252, + 545, + 296 + ], + "type": "text", + "content": "[9] Alexander Clegg, Wenhao Yu, Jie Tan, C Karen Liu, and Greg Turk. Learning to dress: Synthesizing human dressing motion via deep reinforcement learning. ACM Transactions on Graphics (TOG), 37(6):1-10, 2018." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "type": "text", + "content": "[10] Kiana Ehsani, Winson Han, Alvaro Herrasti, Eli VanderBilt, Luca Weihs, Eric Kolve, Aniruddha Kembhavi, and Roozbeh Mottaghi. Manipulathor: A framework for visual object manipulation. In CVPR, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 342, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 384 + ], + "type": "text", + "content": "[11] Benjamin Eysenbach, Abhishek Gupta, Julian Ibarz, and Sergey Levine. Diversity is all you need: Learning skills without a reward function. arXiv preprint arXiv:1802.06070, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 388, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 441 + ], + "type": "text", + "content": "[12] Pete Florence, Corey Lynch, Andy Zeng, Oscar A Ramirez, Ayzaan Wahid, Laura Downs, Adrian Wong, Johnny Lee, Igor Mordatch, and Jonathan Tompson. Implicit behavioral cloning. In Conference on Robot Learning, pages 158-168. PMLR, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 443, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 443, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 443, + 545, + 475 + ], + "type": "text", + "content": "[13] Roy Fox, Sanjay Krishnan, Ion Stoica, and Ken Goldberg. Multi-level discovery of deep options. arXiv preprint arXiv:1703.08294, 2017." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 477, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 477, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 477, + 545, + 531 + ], + "type": "text", + "content": "[14] Chuang Gan, Jeremy Schwartz, Seth Alter, Damian Mrowca, Martin Schrimpf, James Traer, Julian De Freitas, Jonas Kubilius, Abhishek Bhandwaldar, Nick Haber, et al. Threadworld: A platform for interactive multi-modal physical simulation. NeurIPS Datasets and Benchmarks Track, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 533, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 586 + ], + "type": "text", + "content": "[15] Caelan Reed Garrett, Rohan Chitnis, Rachel Holladay, Beomjoon Kim, Tom Silver, Leslie Pack Kaelbling, and Tomás Lozano-Pérez. Integrated task and motion planning. Annual review of control, robotics, and autonomous systems, 4:265-293, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 590, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 621 + ], + "type": "text", + "content": "[16] Jiayuan Gu, Devendra Singh Chaplot, Hao Su, and Jitendra Malik. Multi-skill mobile manipulation for object rearrangement. arXiv preprint arXiv:2209.02778, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "type": "text", + "content": "[17] Kourosh Hakhamaneshi, Ruihan Zhao, Albert Zhan, Pieter Abbeel, and Michael Laskin. Hierarchical few-shot imitation with skill transition models. arXiv preprint arXiv:2107.08981, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 668, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 668, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 668, + 545, + 689 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, 2016." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 692, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 712 + ], + "type": "text", + "content": "[19] Daniel Honerkamp, Tim Welschehold, and Abhinav Valada. Learning kinematic feasibility for mobile manipulation" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "text", + "content": "10860" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "type": "text", + "content": "through deep reinforcement learning. IEEE Robotics and Automation Letters, 6(4):6289-6296, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[20] Michael Janner, Qiyang Li, and Sergey Levine. Reinforcement learning as one big sequence modeling problem. In ICML 2021 Workshop on Unsupervised Reinforcement Learning, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 228 + ], + "type": "text", + "content": "[21] Dmitry Kalashnikov, Alex Irpan, Peter Pastor, Julian Ibarz, Alexander Herzog, Eric Jang, Deirdre Quillen, Ethan Holly, Mrinal Kalakrishnan, Vincent Vanhoucke, and Sergey Levine. Scalable deep reinforcement learning for vision-based robotic manipulation. In 2nd Annual Conference on Robot Learning, CoRL 2018, Zürich, Switzerland, 29-31 October 2018, Proceedings, volume 87 of Proceedings of Machine Learning Research, pages 651-673. PMLR, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 273 + ], + "type": "text", + "content": "[22] George Konidaris and Andrew Barto. Skill discovery in continuous reinforcement learning domains using skill chaining. Advances in neural information processing systems, 22, 2009." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "type": "text", + "content": "[23] Kuang-Huei Lee, Ofir Nachum, Mengjiao Yang, Lisa Lee, Daniel Freeman, Winnie Xu, Sergio Guadarrama, Ian Fischer, Eric Jang, Henryk Michalewski, et al. Multi-game decision transformers. arXiv preprint arXiv:2205.15241, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 363 + ], + "type": "text", + "content": "[24] Youngwoon Lee, Joseph J Lim, Anima Anandkumar, and Yuke Zhu. Adversarial skill chaining for long-horizon robot manipulation via terminal state regularization. arXiv preprint arXiv:2111.07999, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "type": "text", + "content": "[25] Youngwoon Lee, Shao-Hua Sun, Sriram Somasundaram, Edward S Hu, and Joseph J Lim. Composing complex skills by learning transition policies. In International Conference on Learning Representations, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "type": "text", + "content": "[26] Chengshu Li, Fei Xia, Roberto Martin-Martin, and Silvio Savarese. Hrl4in: Hierarchical reinforcement learning for interactive navigation with mobile manipulators. In Conference on Robot Learning, pages 603-616. PMLR, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 498 + ], + "type": "text", + "content": "[27] Anthony Liang, Ishika Singh, Karl Pertsch, and Jesse Thomason. Transformer adapters for robot learning. In CoRL 2022 Workshop on Pre-training Robot Learning, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "text", + "content": "[28] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. arXiv preprint arXiv:2209.07753, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "text", + "content": "[29] Maria Vittoria Minniti, Farbod Farshidian, Ruben Grandia, and Marco Hutter. Whole-body mpc for a dynamically stable mobile manipulator. IEEE Robotics and Automation Letters, 4(4):3687-3694, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "text", + "content": "[30] Mayank Mittal, David Hoeller, Farbod Farshidian, Marco Hutter, and Animesh Garg. Articulated object interaction in unknown scenes with whole-body mobile manipulation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1647-1654. IEEE, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "text", + "content": "[31] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[32] Ram Ramrakhya, Eric Undersander, Dhruv Batra, and Abhishek Das. Habitat-web: Learning embodied object-search strategies from human demonstrations at scale. In Proceed-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5173-5183, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 97, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 545, + 151 + ], + "type": "text", + "content": "[33] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv preprint arXiv:2205.06175, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 153, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 545, + 175 + ], + "type": "text", + "content": "[34] Fetch robotics. Fetch. http://fetchrobotics.com/, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 177, + 545, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 177, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 307, + 177, + 545, + 210 + ], + "type": "text", + "content": "[35] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 212, + 545, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 212, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 545, + 245 + ], + "type": "text", + "content": "[36] Archit Sharma, Shixiang Gu, Sergey Levine, Vikash Kumar, and Karol Hausman. Dynamics-aware unsupervised discovery of skills. arXiv preprint arXiv:1907.01657, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 247, + 545, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 247, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 307, + 247, + 545, + 280 + ], + "type": "text", + "content": "[37] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In Conference on Robot Learning, pages 894–906. PMLR, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 281, + 545, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 281, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 307, + 281, + 545, + 314 + ], + "type": "text", + "content": "[38] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. arXiv preprint arXiv:2209.05451, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 316, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 360 + ], + "type": "text", + "content": "[39] Richard S Sutton, Doina Precup, and Satinder Singh. Between mdps and semi-mdps: A framework for temporal abstraction in reinforcement learning. Artificial intelligence, 112(1-2):181-211, 1999." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 361, + 545, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 427 + ], + "type": "text", + "content": "[40] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, et al. Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems, 34, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 430, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 430, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 430, + 545, + 483 + ], + "type": "text", + "content": "[41] Andrew Szot, Karmesh Yadav, Alex Clegg, Vincent-Pierre Berges, Aaron Gokaslan, Angel Chang, Manolis Savva, Zsolt Kira, and Dhruv Batra. Habitat rearrangement challenge 2022. https://aihabitat.org/challenge/2022_rearrange, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 486, + 545, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 486, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 307, + 486, + 545, + 530 + ], + "type": "text", + "content": "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 532, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 545, + 575 + ], + "type": "text", + "content": "[43] Cong Wang, Qifeng Zhang, Qiyan Tian, Shuo Li, Xiaohui Wang, David Lane, Yvan Petillot, and Sen Wang. Learning mobile manipulation through deep reinforcement learning. Sensors, 20(3):939, 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 578, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 621 + ], + "type": "text", + "content": "[44] Luca Weihs, Matt Deitke, Aniruddha Kembhavi, and Roozbeh Mottaghi. Visual room rearrangement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5922-5931, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "text", + "content": "[45] Erik Wijmans, Irfan Essa, and Dhruv Batra. Ver: Scaling on-policy rl leads to the emergence of navigation in embodied rearrangement. In Advances in Neural Information Processing Systems, 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[46] Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. arXiv preprint arXiv:1911.00357, 2019." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "text", + "content": "10861" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 206 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[47] Josiah Wong, Albert Tung, Andrey Kurenkov, Ajay Mandlekar, Li Fei-Fei, Silvio Savarese, and Roberto Martin-Martin. Error-aware imitation learning from teleoperation data for mobile manipulation. In Conference on Robot Learning, pages 1367–1378. PMLR, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "text", + "content": "[48] Mengdi Xu, Yikang Shen, Shun Zhang, Yuchen Lu, Ding Zhao, Joshua Tenenbaum, and Chuang Gan. Prompting decision transformer for few-shot policy generalization. In International Conference on Machine Learning, pages 24631-24645. PMLR, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 206 + ], + "type": "text", + "content": "[49] Qinqing Zheng, Amy Zhang, and Aditya Grover. Online decision transformer. arXiv preprint arXiv:2202.05607, 2022." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "text", + "content": "10862" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Skip-Plan_ Procedure Planning in Instructional Videos via Condensed Action Space Learning/c1a3a0ec-3885-4f2f-93dd-62793c50f314_content_list.json b/2023/Skip-Plan_ Procedure Planning in Instructional Videos via Condensed Action Space Learning/c1a3a0ec-3885-4f2f-93dd-62793c50f314_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0b6e1ae111652eaca94e8d7cd84d0e46a75a6454 --- /dev/null +++ b/2023/Skip-Plan_ Procedure Planning in Instructional Videos via Condensed Action Space Learning/c1a3a0ec-3885-4f2f-93dd-62793c50f314_content_list.json @@ -0,0 +1,1395 @@ +[ + { + "type": "text", + "text": "Skip-Plan: Procedure Planning in Instructional Videos via Condensed Action Space Learning", + "text_level": 1, + "bbox": [ + 207, + 130, + 759, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhiheng Li $^{1}$ , Wenjia Geng $^{2}$ , Muheng Li $^{1}$ , Lei Chen $^{3}$ , Yansong Tang $^{2*}$ , Jiwen Lu $^{1,4}$ , Jie Zhou $^{1,4}$", + "bbox": [ + 116, + 202, + 849, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Automation, Tsinghua University", + "bbox": [ + 285, + 220, + 683, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Shenzhen International Graduate School, Tsinghua University", + "bbox": [ + 233, + 239, + 733, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Beijing University of Science and Technology", + "bbox": [ + 295, + 256, + 671, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Beijing National Research Center for Information Science and Technology", + "bbox": [ + 184, + 273, + 785, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{lizhihan21@gengwj22@,li-mh20}@mails.tsinghua.edu.cn,chenlei2022@ustb.edu.cn, {tang.yansong@sz.,lujiwen,jzhou@}tsinghua.edu.cn", + "bbox": [ + 142, + 294, + 825, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 361, + 312, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose Skip-Plan, a condensed action space learning method for procedure planning in instructional videos. Current procedure planning methods all stick to the state-action pair prediction at every timestep and generate actions adjacently. Although it coincides with human intuition, such a methodology consistently struggles with high-dimensional state supervision and error accumulation on action sequences. In this work, we abstract the procedure planning problem as a mathematical chain model. By skipping uncertain nodes and edges in action chains, we transfer long and complex sequence functions into short but reliable ones in two ways. First, we skip all the intermediate state supervision and only focus on action predictions. Second, we decompose relatively long chains into multiple short sub-chains by skipping unreliable intermediate actions. By this means, our model explores all sorts of reliable sub-relations within an action sequence in the condensed action space. Extensive experiments show Skip-Plan achieves state-of-the-art performance on the CrossTask and COIN benchmarks for procedure planning.", + "bbox": [ + 76, + 393, + 473, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 727, + 209, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given the initial and goal visual observations which are curated from instructional videos [44, 37], the task of procedure planning is to predict a sequence of actions to reach the indicated goal [7], as illustrated in Figure 1(a). This is an emerging research direction in the field of video understanding. Procedure planning has worldwide applications and assists humans with instruction-following tasks, such as cooking, assembling furniture, electronics repair, etc.", + "bbox": [ + 75, + 753, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Question: What action sequence are needed to get from $v_{start}$ to $v_{goal}$ ?", + "bbox": [ + 509, + 363, + 694, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b69948dc1487f9e8bff0863ec5ecd0094f4913a48597e658da62001680dd1a32.jpg", + "image_caption": [ + "(a) Problem Definition" + ], + "image_footnote": [], + "bbox": [ + 527, + 391, + 697, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Answer: Required action sequence", + "bbox": [ + 702, + 363, + 883, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(Time Horizon $T = 4$", + "bbox": [ + 728, + 375, + 851, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/29ec5d8dee3b24a6f4f646646fd8eaac91ab1a59a2ab7d331b5bd4d3130910d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 388, + 870, + 453 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/248d852484ca6d081083b77706bb1cec3e6b9db73d2358a2a53429c00375c233.jpg", + "image_caption": [ + "(b) Conventional Procedure Planning Paradigm" + ], + "image_footnote": [], + "bbox": [ + 519, + 477, + 862, + 547 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/57cd4e7a4e8bf0c1e11c4180e00653ee59f406f5fad470388633dd5209df88c8.jpg", + "image_caption": [ + "(c)Skip-Plan" + ], + "image_footnote": [], + "bbox": [ + 522, + 566, + 867, + 650 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0d9a2dd55e040bdd5893ea75417e20dc29eff784da94beba768ffbb0d7aad84f.jpg", + "image_caption": [ + "Figure 1. Comparisons of the conventional paradigm and our proposed Skip-Plan in procedure planning. (a) Problem definition: given the initial and goal visual observations $(v_{start}$ and $v_{goal})$ , procedure planning is to answer an action sequence to reach the indicated goal with a pre-defined time horizon $T$ . (b) Conventional paradigm predicts a state-action pair $\\{\\tilde{s}_t,\\tilde{a}_t\\}$ at every timestep $t$ , and repeat it step-by-step. (c) Our Skip-Plan skips state supervision and extracts multiple sub-chains by skipping most intermediate actions. The final action sequence is accumulated from all predicted sub-chains. Skip-Plan significantly reduces the overall compound error, indicated by the shallower darkness of the chevron arrows." + ], + "image_footnote": [], + "bbox": [ + 509, + 669, + 883, + 700 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* indicates the corresponding author.", + "bbox": [ + 94, + 886, + 290, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10297", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the past few years, various approaches have been proposed to solve the procedure planning problem. Most existing methods utilize an autoregressive model with visual state supervision to generate actions one by one [7, 4, 36]. Another type of work [41] solves it in a non-autoregressive way with language state supervision. The difficulties of procedure planning have been widely discussed in the previous works. For instance, the sequential orders in procedure planning can be altered but still sensible (e.g., switching the order of 'pour water' and 'pour alcohol' does not matter in many cases), and the visual states are unstructured and high-dimensional. Thus, we can conclude the following limitations from the existing works. (a) State supervision: the major difficulty in procedure planning is state supervision. Visual images are high-dimensional and complex, with extra misleading information such as background, illumination, viewing point, etc. (b) Step-by-step predictions: the conventional paradigm in procedure planning predicts one state-action pair at every timestep, and repeats it step-by-step, as illustrated in Figure 1(b). Even though the proposed non-autoregressive model generates whole action sequences in batches [41], it can be still regarded as a step-by-step method because the model neglects the links between non-adjacent actions. From Figure 1(b), we can see while predicting an action sequence with a time horizon of 4, the conventional model is equivalently predicting a sequence with a length of 7 (4 action nodes, 3 state nodes). More frequent predictions tend to accumulate a larger compound error along the chain, which lowers the prediction accuracy.", + "bbox": [ + 76, + 90, + 472, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these limitations, we propose Skip-Plan, a condensed action space learning method that solves the procedure planning problem through the lens of mathematics, aiming to reduce compound error along action chains. As we have discussed above, existing methods struggle with high-dimensional and complex state supervision. Thus, the compound error along the action chain is increased due to large errors in state predictions. This motivates us to skip all the state nodes in our Skip-Plan framework and only model sequence relations in the condensed action space, as demonstrated in Figure 1(c). Without the presence of state supervision, a lower compound error results in better performance in Skip-Plan. Although the dependencies between high-level actions without state supervision are very abstract, the algorithm of our Skip-Plan is highly mathematically sensible in the Markov chain model.", + "bbox": [ + 76, + 532, + 472, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another convention in procedure planning is we should learn sequential relations between adjacent actions. The Markov chain model inspires us to decouple relatively long chains into multiple short sub-chains, because the compound error can be reduced with shorter chain lengths. The problem here is how to decouple these long chains, because there are various choices of sub-chains with different lengths and we should always choose the most reliable", + "bbox": [ + 76, + 779, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ones under any condition. Given the initial and goal frames as concrete inputs, it is apparent the reliabilities of the initial and final actions are the highest, and the intermediate actions are more likely to be predicted wrong. Therefore, our Skip-Plan decomposes long chains by 'skipping' some unreliable intermediate actions, and each sub-chain in our model is composed of the initial action, one unique intermediate action, and the final action. One example for a chain length of 4 is shown in Figure 1(c). In this way, the reliable sub-relations between non-adjacent actions can be well explored by the network. To model the discriminative relations of all sub-chains, we propose sub-chain decoders, where each decoder only learns one sub-chain by the specific loss design. The final action sequence is aggregated from all sub-chain predictions by a sub-chain accumulator.", + "bbox": [ + 496, + 90, + 893, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are threefold. (a) Unlike the previous methods which put sophisticated designs on state supervision, Skip-Plan skips unreliable state predictions to boost the model performance in procedure planning. (b) In contrast to the existing methods which only focus on the links between adjacent actions, we decompose long chains into multiple reliable sub-chains by skipping some intermediate actions, and propose sub-chain decoders to learn their discriminative relations. (c) Extensive experiments show Skip-Plan achieves state-of-the-art performance on the CrossTask [44] and COIN [37] benchmarks for procedure planning.", + "bbox": [ + 496, + 318, + 893, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 512, + 640, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Procedure Planning. Procedure planning in instructional videos was first introduced by Chang et al. [7]. Recently, various methods have been conducted to solve this task. Chang et al. [7] propose a dual dynamic network to conjugate the relationships between states and actions. Sun et al. [36] utilize a GPT network to generate actions and states one by one. Bi et al. [4] use the model-based reinforcement learning with time-invariant context information to predict state-action pairs autoregressively. Different from the previous autoregressive methods with visual state supervision, Zhao et al. [41] use a non-autoregressive transformer decoder to generate action sequences with language state supervision. Unlike previous works, Skip-Plan generates actions non-adjacently without any state supervision.", + "bbox": [ + 496, + 537, + 893, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instructional Videos Analysis. Instructional videos are used for learning human behaviours because they are visual representations of human actions. Based on instructional videos, various action-related tasks have been proposed, such as action segmentation[21, 22, 13, 14, 28, 31], step localization[39, 11, 43, 33, 26], procedure segmentation[42, 19, 3, 5], video summarization/captioning[8, 35, 29, 34, 16, 18], etc. Different from these action tasks in instructional videos, we investigate procedure planning, a goal-conditioned problem that plans a sequence of actions given", + "bbox": [ + 496, + 750, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10298", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the initial and goal visual observations.", + "bbox": [ + 76, + 90, + 336, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Predictive Uncertainty on Sequence Relations. All types of questions that we are interested in can be modeled as a sequence model, such as images [25] and languages [9]. Currently, Transformer [38] is the most effective network to model sequence relations, and our model is based on the transformer decoder with a learnable query input[40, 6]. However, sequence relations are always distorted by predictive uncertainty[12, 27, 17, 30, 20], and the sources of predictive uncertainty can be various, including model uncertainty, data uncertainty, etc[2, 10, 15, 23, 32]. In this work, we pay particular attention to data uncertainty, in which the uncertainty directly stems from the data. Data uncertainty is caused by information loss when we measure a data distribution in a space with incorrect dimensions (e.g., we must lose depth information if we measure a three-dimensional object in a two-dimensional space). Therefore, we often avoid high-dimensional data for lower data uncertainty.", + "bbox": [ + 75, + 106, + 472, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 76, + 375, + 174, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 398, + 279, + 414 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We follow the definition of procedure planning proposed by Chang et al. [7]. Here, we denote the action and the state at timestep $t$ as $a_{t}$ and $s_{t}$ respectively. Given the initial visual observation $v_{start}$ and the goal visual observation $v_{goal}$ , the model generates a sequence of actions $a_{1:T}$ to reach the indicated goal. In this work, we introduce a novel formulation for the procedure planning task:", + "bbox": [ + 75, + 422, + 470, + 529 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(a _ {1: T} \\mid v _ {s}, v _ {g}\\right) = \\prod_ {t = 2} ^ {T - 1} p \\left(a _ {t} \\mid a _ {1}, a _ {T}\\right) p \\left(a _ {1}, a _ {T} \\mid v _ {s}, v _ {g}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 535, + 468, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $v_{s}$ and $v_{g}$ short for ' $v_{start}$ ' and ' $v_{goal}$ ' . We denote $p(a_{1:T}|v_{start},v_{goal})$ as the conditional probability distribution of the action sequence given the initial visual observation $v_{start}$ and the goal visual observation $v_{goal}$ , $p(a_t|a_1,a_T)$ as the conditional probability distribution of intermediate actions $(2\\leq t\\leq T - 1)$ given the initial action $a_1$ and the final action $a_{T}$ , and $p(a_{1},a_{T}|v_{start},v_{goal})$ as the conditional probability distribution of the initial action $a_1$ and the final action $a_{T}$ given the initial visual observation $v_{start}$ and the goal visual observation $v_{goal}$ . Equation 1 reveals the uniqueness of our Skip-Plan method: First, we do not rely on any intermediate state supervision $s_t$ for action generations. Second, every intermediate action prediction $a_{t}$ $(2\\leq t\\leq T - 1)$ only depends on the initial action $a_1$ and the final action $a_{T}$ . Thus, Skip-Plan does not focus on sequential relations between adjacent actions but predicts action sequences non-adjacently without state supervision.", + "bbox": [ + 75, + 582, + 472, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Action Predictions without State Supervision", + "text_level": 1, + "bbox": [ + 76, + 847, + 457, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Information does not always help in the decision-making process. Sometimes ambiguous information can even lead", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to sub-optimal results, and plausible plans can be only made based on reliable information. In procedure planning, two types of supervision can be utilized, including actions $a_{t}$ and visual states $s_{t}$ . Even though they can be directly generated from each other, the intrinsic properties of these two supervisions are largely different. We evaluate them from the perspectives of the value of information and predictive uncertainty. Visual state supervision is well known to be high-dimensional. In addition to the factor of action class, various factors such as background, viewing point, illumination, etc. can be involved in visual states. Thus, useful information in visual state supervision for action predictions is overwhelmed by many irrelevant factors. In contrast, action labels are the most condensed supervision, without being interrupted by any factors. From this point, we argue the value of information offered by state supervision is extremely low because the portion of information, which is relevant to action predictions, is too small. On the other hand, when models generate high-dimensional predictions such as visual images, large predictive errors are more likely to occur. Thus, the predictive uncertainty of intermediate visual states is significantly higher than the one of condensed action predictions. Combining the effects of the value of information and the predictive uncertainty, we argue the existence of visual state supervision hurts the predictive accuracy of sequential actions. Even though we can use the weak language state supervision to reduce some irrelevant dimensions, it is still too high-dimensional compared to condensed action labels. Therefore, we 'skip' intermediate state supervision in our Skip-Plan framework. The overall compound error along an action chain is significantly reduced without the predictive errors introduced by states.", + "bbox": [ + 496, + 90, + 893, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Getting rid of intermediate state supervision does not mean all sorts of visual information are useless. In procedure planning, the initial visual observation $v_{start}$ and the goal visual observation $v_{goal}$ are vital because action sequences are constrained by them. These visual observations are in the form of frame features in seconds. Inspired by the image analysis above, we argue the visual input information is underutilized by the previous works, especially along the time dimension. In the Skip-Plan framework, we take a shared multi-layer perceptrons (MLP) network in the Visual Input Module to learn the time information, where the MLP network is three-layered with ReLU activations. This simple method can boost the model performance by a large margin, especially for short sequences $(T = 3)$ .", + "bbox": [ + 496, + 575, + 893, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Decomposition of Long Chains", + "text_level": 1, + "bbox": [ + 500, + 800, + 772, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Long chains are not favored when a compound error is accumulating. Thus, decomposing long chains into short ones is a natural way to better model performance. Through the lens of mathematics, the only design philosophy here is that we should extract most reliable sub-chains from the", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10299", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/626640d6c67c78dd6e9559a4d27b6c9010289993c200761769d0796d28562694.jpg", + "image_caption": [ + "Figure 2. Our Skip-Plan architecture. The shared MLP network in the Visual Input Module extracts the initial and final information from the initial and goal observations ( $v_{start}$ and $v_{goal}$ ) and attaches them to the first and last dimensions of every learnable query input. A number of $T - 2$ query inputs are separately passed to each decoder in the Sub-chain Decoder Module. In each decoder, the input is processed by the self-attention block and then by the cross-attention block, where the cross-attention block is augmented by a learnable memory $M_i$ . The sub-chain decoders are separated. The $n$ th decoder outputs the sub-chain $\\{ \\tilde{a}_1, \\tilde{a}_t, \\tilde{a}_T \\}$ with $t = n + 1$ . The predicted sub-chains are supervised by two loss supervision. In Individual Sub-chain Supervision, the predicted sub-chains are supervised by the corresponding ground-truth sub-chains. In Complete Chain Supervision, the predicted sub-chains are aggregated by the Sub-chain Accumulator to generate the complete action chain, which is supervised by the complete ground-truth action chain." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 890, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "original long chain. Following this principle, we ought to find there is so much flexibility in the decoupling method: We can choose sub-chains with any lengths, and actions in the sub-chains are not necessary to be adjacent.", + "bbox": [ + 75, + 422, + 468, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The reliability of a sub-chain can be determined by the uncertainty of predictions. According to the uncertainty analysis, the reliabilities of the initial and final actions are the highest because they are the closest actions to the initial and goal visual inputs. Since the compound error is accumulated from two ends of the chain, the reliability of action predictions is decreasing while approaching the middle of the chain. Consequently, we remove all the edges between adjacent intermediate actions, and each individual intermediate action only connects to the reliable initial and final actions. This design just coincides with the term $p(a_{t}|a_{1},a_{T})$ with $2\\leq t\\leq T - 1$ in Equation 1. In this way, we generate a quantity of $T - 2$ sub-chains from an action chain with a time horizon of $T$ . The set of sub-chains is given by:", + "bbox": [ + 75, + 484, + 470, + 696 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\{(a _ {1}, a _ {2}, a _ {T}), (a _ {1}, a _ {3}, a _ {T}), \\dots , (a _ {1}, a _ {T - 1}, a _ {T}) \\}. \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 705, + 468, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each sub-chain has a length of 3, with the order of the initial action, one unique intermediate action, and the final action. This decoupling method works well for relatively long chains $(T > 3)$ .", + "bbox": [ + 75, + 734, + 468, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given multiple short sub-chains, we design a compatible network to model all their sub-relations. Since these sub-chains are parallel and independent to each other, we propose the Sub-chain Decoder Module, where each transformer decoder only models one discriminative relation of one sub-chain. Every transformer decoder is non-autoregressive with two types of inputs: a learnable query", + "bbox": [ + 75, + 795, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with the initial and goal visual features, and a learnable memory component. The novelty here is each decoder only learns one sub-relation, with a unique loss only corresponding to the three actions of that sub-chain. The query input, parameters, and memory component of every decoder are separated from other decoders'. Therefore, the number of decoders in the Sub-chain Decoder Module is equal to the number of sub-chains $(T - 2)$ extracted from the original chain. The final sequence generation is aggregated from all the sub-chain predictions through the Sub-chain Accumulator, which is an MLP network. In this way, our Skip-Plan network can model all sorts of discriminative sub-relations within the original action sequence.", + "bbox": [ + 496, + 422, + 893, + 621 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Training", + "text_level": 1, + "bbox": [ + 500, + 627, + 606, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Loss is an essential component to discriminate sub-chain decoders. Instead of conventional cross-entropy loss, we use the focal loss to supervise our sub-chain decoders. The focal loss $FL$ is formulated as[24]:", + "bbox": [ + 496, + 650, + 893, + 712 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF L = - \\sum_ {t = 1} ^ {T} a _ {t} (1 - \\tilde {a} _ {t}) ^ {\\gamma} \\log \\tilde {a} _ {t}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 717, + 890, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we denote $a_{t}$ as the ground-truth one-hot labels, $\\tilde{a}_{t}$ as the predicted probability distribution of the action sequence, and $\\gamma$ as the gamma value of focal loss. The focal loss is equivalent to the cross-entropy loss when $\\gamma = 0$ . The model training is benefited from the use of focal loss because the data distribution in procedure planning is imbalanced. In this work, we define $FL(a_{1},a_{t},a_{T})$ as the focal loss for the sub-chain with a length of 3, where the sub-chain is composed of the initial action $a_{1}$ , the action at timestep $t$ $a_{t}$ , and", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10300", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the final action $a_{T}$ . We present the focal loss of the whole action sequence as $FL(a_{1:T})$ .", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss of Individual Sub-chains. As we have discussed before, the loss of each decoder only corresponds to the actions within that sub-chain. Therefore, the individual loss of the $n$ th decoder is defined as $FL(a_{1},a_{t},a_{T})$ with $t = n + 1$ , and the loss from all sub-chain decoders can be written as:", + "bbox": [ + 75, + 121, + 468, + 195 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {N} = \\sum_ {t = 2} ^ {T - 1} F L \\left(a _ {1}, a _ {t}, a _ {T}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 205, + 468, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss of Complete Chain. After aggregating all sub-chains by the Sub-chain Accumulator, the sequence output with a length of $T$ is supervised by the focal loss of the whole action sequence:", + "bbox": [ + 75, + 257, + 468, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T} = F L \\left(a _ {1: T}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 330, + 468, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall Loss. Therefore, the total loss is defined as the sum of the loss from all sub-chains and the complete chain loss, which can be written as:", + "bbox": [ + 75, + 358, + 468, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {N} + \\mathcal {L} _ {T}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 416, + 468, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, our model can explore all sorts of discriminative but reliable sub-relations within an action sequence, while guaranteeing the final sequence generation is optimal.", + "bbox": [ + 75, + 443, + 468, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 498, + 294, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each decoder, the query input length is $T + 1$ , and the memory size is 128. The numbers of decoder layers, decoder heads, and decoder dimensions are set to be 1, 16, and 1024 respectively. All MLP networks are three-layered with ReLU activations. In the Visual Input Module, our shared MLP network on the time dimension of input frame features has the shape $[3 \\to 6 \\to 1]$ . All action classifiers in the Sub-chain Decoder Module are the same MLP networks with the shape $[1024 \\to 512 \\to n_a]$ , where $n_a$ is the number of action classes. The Sub-chain Accumulator is an MLP network with the shape $[T(T - 2) \\to 3T(T - 2) \\to T]$ . Models are optimized for 500 epochs with stochastic gradient descent. The initial learning rate is 0.02 and decays by 10 at every 50 epochs after 100 epochs. Experiments are conducted on a single GeForce RTX 3090 GPU. We also test our model using the Mindspace [1].", + "bbox": [ + 75, + 521, + 468, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 777, + 209, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Evaluation Protocols", + "text_level": 1, + "bbox": [ + 76, + 801, + 272, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our model on two real-world instructional video datasets: CrossTask[44] and COIN[37]. CrossTask contains 2763 videos with a duration of 213 hours. It consists of 18 different tasks with 7.4 steps per task. The COIN dataset spans 12 different domains related", + "bbox": [ + 75, + 824, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to our daily life. It contains 11,827 videos with 3.9 steps per video. The COIN dataset is challenging due to the diversity of its tasks and action labels (180 tasks and 779 action labels). We curate a set of plans by the sliding window approach, and adopt the standard $70\\% / 30\\%$ to create our train/test splits. The division is done on the generated samples rather than the original video. The data pre-processing procedure in our work is identical to the most widely used protocol in the previous procedure planning works (e.g., DDN [7], Ext-GAIL [4], and the main protocol in P3IV [41]) for fair comparisons.", + "bbox": [ + 496, + 90, + 890, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. Following the previous works, we used three metrics to comprehensively evaluate model performance. For each action sequence: (a) mean Intersection over Union (mIoU) regards the predicted and ground-truth action sequences as two sets, and measure the overlap between these sets. The mIoU metric only indicates whether desired actions exist in the predicted action set, agnostic to the order of actions. (b) mean Accuracy (mAcc) evaluates the correctness of the predicted action at every timestep, termed as percentages. Thus, mAcc counts the order of actions, and is more strict than mIoU. (c) Success Rate (SR) evaluates the correctness of the whole action sequence. A predicted sequence is successful only if it matches the ground-truth sequence. The success rate is the proportion of successful samples over all samples, which is more strict than mAcc.", + "bbox": [ + 496, + 258, + 890, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Scheme on Each Dataset. We specify an appropriate gamma value on each dataset to maximize model performance. Through experiments, we find the gamma values of 1.5 and 0 are the best parameters for the CrossTask and COIN datasets respectively. This is reasonable because the data distribution in the COIN dataset is more balanced due to a large quantity of action labels in diverse tasks.", + "bbox": [ + 496, + 486, + 890, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluating Procedure Planning", + "text_level": 1, + "bbox": [ + 498, + 604, + 777, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Short-horizon Prediction. We compare the model performance of the short-horizon predictions $(T = 3 / 4)$ in Table 1. As illustrated in Table 1, Skip-Plan significantly outperforms the previous state-of-the-art model at all metrics for $T = 3$ . In the CrossTask dataset, our SR and mAcc results surpass the results of the second best approach (e.g., P3IV[41]) by $6\\%$ and $11\\%$ . Our method also outperforms P3IV[41] by $8\\%$ on SR and $25\\%$ on mAcc in the COIN dataset. Such incredible improvements strongly prove Skip-Plan without state supervision suppresses predictive uncertainty effectively, and our module design on the visual input extracts initial and goal visual information thoroughly. For $T = 4$ , Skip-Plan achieves large gains of around $5\\%$ , $24\\%$ , and $7\\%$ in SR, mAcc and mIoU respectively in the COIN dataset. For the CrossTask dataset, the mAcc and mIoU of our Skip-Plan model are the best and the SR is the second best. Particularly, our mAcc is higher than P3IV's[41] by $11.48\\%$ . The methods of no state supervision and decou", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10301", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/26721f071576e71c2d14f1df142cba5917b2e8c763e928de1cca0c8615b9352d.jpg", + "table_caption": [ + "Table 1. Metric results on CrossTask and COIN for short-horizon predictions $\\left( {T \\in \\{ 3,4\\} }\\right)$ . The ’ $\\checkmark$ ’ and ’ $X$ ’ in the column State Supervision denote with and without state supervision respectively. The best and second best metric results are coloured in red and blue respectively." + ], + "table_footnote": [], + "table_body": "
HorizonsModelsState SupervisionCrossTaskCOIN
SRmAccmIoUSRmAccmIoU
T=3DDN[7]✓12.1831.2947.4813.9020.1964.78
Plate[36]✓16.0036.1765.91---
Ext-GAIL[4]✓21.2749.4661.70---
P3IV[41]✓23.3449.9673.8915.4021.6776.31
Skip-Plan✗28.8561.1874.9823.6547.1278.44
T=4DDN[7]✓5.9727.1048.4611.1317.7168.06
Plate[36]✓14.0035.2955.36---
Ext-GAIL[4]✓16.4143.0560.93---
P3IV[41]✓13.4044.1670.0111.3218.8570.53
Skip-Plan✗15.5655.6470.3016.0443.1977.07
", + "bbox": [ + 114, + 118, + 849, + 345 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1e7d012dabeaee0851c39e0164b51214d40ce84e39c0c338cd90b9f1ce6e05db.jpg", + "table_caption": [ + "Table 2. Metric results for long-horizon predictions $(T \\in \\{5,6\\})$ on CrossTask and COIN." + ], + "table_footnote": [], + "table_body": "
DatasetsHorizonsModelsSRmAccmIoU
CrossTaskT=5DDN[7]3.10--
P3IV[41]7.21--
Skip-Plan8.5552.0569.09
T=6DDN[7]1.20--
P3IV[41]4.40--
Skip-Plan5.1247.3466.11
COINT=5P3IV[41]4.2710.8168.81
Skip-Plan9.9038.9976.93
T=6P3IV[41]4.178.9767.72
Skip-Plan6.6133.6373.68
", + "bbox": [ + 78, + 393, + 465, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pling long chains both drive Skip-Plan to succeed at $T = 4$", + "bbox": [ + 76, + 628, + 467, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Long-horizon Prediction. We further evaluate Skip-Plan on long-horizon predictions ( $T = 5/6$ ). In the COIN dataset, our Skip-Plan performs extraordinarily well on the metric of mAcc, with remarkable improvements of $28\\%$ and $25\\%$ for $T = 5/6$ , compared to the second best model P3IV[41]. As for the metrics of SR and mIoU, Skip-Plan exceeds P3IV[41] by $6\\%$ and $8\\%$ at $T = 5$ , and $2\\%$ and $6\\%$ at $T = 6$ respectively. The outstanding performance we presented above further validates the success of the decoupling approach. Our Skip-Plan also achieved the best results on SR in the CrossTask dataset, surpassing P3IV[41] by $1.34\\%$ and $0.72\\%$ at $T = 5/6$ . For the long-horizon predictions, we argue mAcc and mIoU are more reliable metrics to measure the effectiveness of a model. SR is interrupted by a large variance, because data samples become insufficient with a large time horizon.", + "bbox": [ + 75, + 657, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Prerequisite Experiments", + "text_level": 1, + "bbox": [ + 500, + 369, + 732, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we carried out several prerequisite experiments for the long-chain decoupling method. They can validate some statements that we have made in Section 3.3. These experiments are foundational because they proved the prerequisite conditions for the decoupling approach to work. Thus, we name them as the prerequisite experiments.", + "bbox": [ + 496, + 396, + 890, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Error Rate Distribution along a Predicted Chain. We tend to remove the edges between unreliable nodes and only connect one unreliable node to other reliable ones. In this way, we can filter out most reliable sub-chains from the original long chain. Here, we measure the reliability of each node in terms of the error rate distribution along the non-autoregressively predicted chain. The error rate distributions along the predicted chains with $T = 3/4/5/6$ are illustrated in Figure 3. It is clear the reliabilities of the first and final action nodes are the highest and the model is more likely to make wrong predictions while approaching the middle of the chain. It is not hard to understand this phenomenon: a higher error rate occurs at intermediate action predictions because they are further away from the initial and goal visual observations. This empirical finding motivates us to remove the edges between unreliable intermediate actions and only connect each intermediate action to the reliable initial and final action nodes.", + "bbox": [ + 496, + 489, + 892, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reliability of Standalone Sub-chains. In addition to the reliability of each node, we further study the reliability of the sub-chains that we have decoupled from the original long chain. For a clearer explanation, we take a chain with a length of 5 as an example (e.g., $\\{a_1, a_2, a_3, a_4, a_5\\}$ ), and only study the reliability of the sub-chain $\\{a_1, a_3, a_5\\}$ . We consider two cases: First, we just model the relation of the whole chain and only investigate the prediction result of that sub-chain contained in the long chain. In other words, we", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10302", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b313374b0c51add4dea043815653a9f1f91625f81a3c6bea6fcb399bfacbd008.jpg", + "image_caption": [ + "Figure 3. Error rate distributions along the non-autoregressively predicted chains with $T = 3/4/5/6$ ." + ], + "image_footnote": [], + "bbox": [ + 83, + 94, + 464, + 258 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/549fb2295385e101a3bebf9758909886324c77b29a2afe3d250bdc7bfcdb276c.jpg", + "table_caption": [ + "Table 3. Reliability of standalone sub-chains vs sub-chains contained within the original long chain." + ], + "table_footnote": [], + "table_body": "
Sub-chainLoss TypeSRmAccmIoU
{a1,a2,a5}Long22.5258.2871.87
Short22.6858.7372.40
{a1,a3,a5}Long13.7356.7471.78
Short16.2959.9173.07
{a1,a4,a5}Long12.5056.5871.23
Short13.2557.6171.72
", + "bbox": [ + 78, + 333, + 465, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "formulate the loss function along the whole sequence (involving $\\{a_1, a_2, a_3, a_4, a_5\\}$ ), and only compute the metric results over the sub-chain $\\{a_1, a_3, a_5\\}$ . The second case is we only model the relation of the standalone sub-chain and check its prediction result. Equivalently, we model the loss function just involving $\\{a_1, a_3, a_5\\}$ , and still computes the metric results of the sub-chain $\\{a_1, a_3, a_5\\}$ . The metric results of these two cases are presented in Table 3, where we denote 'Long' as the first case and denote 'Short' as the second case due to the length difference in loss. We spot the metric results of standalone sub-chains ('Short') are always better than the ones contained within the whole action chain ('Long'). This empirical finding also happens in the cases of $T = 4$ and $T = 6$ . As we stated at the beginning of this section, such experimental results are foundational: The method of decoupling long chains can be only workable when standalone sub-chains can result in better metric results compared to the sub-chains contained within their original long chain. In this way, we experimentally prove the reliability of standalone sub-chains is better under any condition. It constructs the prerequisite condition to guarantee our decoupling method is workable.", + "bbox": [ + 75, + 503, + 468, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 845, + 230, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present comprehensive ablation experiments to illustrate the effectiveness of different components in our", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2b1e31ee2fb4fed44670b52f6c26facfd0a51da3a71497ed64b6914e0642e52b.jpg", + "table_caption": [ + "Table 4. Ablation experiments on CrossTask." + ], + "table_footnote": [], + "table_body": "
SRmAccmIoU
(a) With and Without State Supervision at T = 3
Language26.2858.9871.82
None28.8561.1874.98
(b) Types of layers on the Time of Input Frames at T = 3
Average Pooling24.4056.4170.87
Linear24.4056.6971.24
Conv26.5659.6074.15
MLP28.8561.1874.98
(c) Without and With Decoupling at T = 4
w/o Decoupling12.7253.1869.13
Decoupling15.5655.6470.30
(d) Without and With Decoupling at T = 5
w/o Decoupling4.9748.7165.29
Decoupling8.5552.0569.09
(e) Without and With Decoupling at T = 6
w/o Decoupling2.6244.0862.07
Decoupling5.1247.3466.11
(f) Gamma Values of Focal Loss at T = 4
γ = 013.8454.6769.74
γ = 1.015.0854.5870.09
γ = 1.515.5655.6470.30
γ = 2.014.8054.9970.20
", + "bbox": [ + 501, + 101, + 888, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Skip-Plan. All ablation experiments are carried out on the CrossTask dataset, illustrated in Table 4.", + "bbox": [ + 498, + 549, + 890, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With and Without State Supervision. We consider $T = 3$ to compare the metric results with and without intermediate state supervision. The model with intermediate state supervision is done by adding weak language state supervision, which is identical to the one proposed by P3IV[41]. The weak language state supervision in P3IV[41] is the most effective state supervision among all the previous works. In Table 4(a), we can see the model without state supervision outperforms the model with language state supervision in all metrics. It confirms the existence of intermediate state nodes does not provide much valuable information but introduces undesired noise into the chain.", + "bbox": [ + 496, + 580, + 892, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Types of Layers on the Time of Input Frames. Our model achieves significant improvements by applying a small MLP network on the time dimension of visual input frames. Here, we explore the effects of various types of layers on the time information, including average pooling, a single linear layer, 1D convolution, and our proposed MLP, as shown in Table 4(b). All the previous works overlooked the significance of initial and goal visual inputs. They used the averaging pooling operation, which performs the worst", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10303", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/033532c38ca7be1c586ef29060e8be387013f0bba3c245f68fdda090beb1e2e1.jpg", + "image_caption": [ + "Figure 4. Case study on CrossTask: our Skip-Plan predicts all actions correctly, whereas P3IV[41] fails on the last three predictions. The outlines of correct and incorrect predictions are coloured in green and red respectively." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 472, + 188 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "among all types of layers. We take $T = 3$ as an example to compare different layers, and indicate the MLP network in our Skip-Plan framework performs the best.", + "bbox": [ + 75, + 275, + 468, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Decomposition of Long Chains. We compare the performance without and with the decoupling approach at all long $T$ , ranging from 4 to 6. Complete ablation experiments demonstrate the decoupling approach in our Skip-Plan can robustly improve all the metric results at any condition.", + "bbox": [ + 75, + 320, + 468, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Focal Loss with Gamma Values We compare the experimental results with different gamma values of focal loss. We find the gamma value of 1.5 works best to overcome the imbalanced data distribution in the CrossTask dataset.", + "bbox": [ + 75, + 397, + 468, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Case Study", + "text_level": 1, + "bbox": [ + 76, + 468, + 199, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Skip-Plan and the Baseline. We compare our Skip-Plan with the previous state-of-the-art P3IV[41] in real cases, illustrated in Figure 4. Given the initial and goal visual observations, P3IV[41] only predicts the first action correctly and fails to get the other actions at $T = 4$ . In contrast, Skip-Plan extracts the initial and goal visual information thoroughly, so Skip-Plan predicts the final action correctly. Benefiting from the decoupling method, our Skip-Plan makes the correct predictions on the second and third actions.", + "bbox": [ + 75, + 491, + 468, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of Different Components in Skip-Plan. We investigate the effects of various components in Skip-Plan on model performance by real samples, which are illustrated in Figure 5. Here, the components for comparisons include the types of layers on the time dimension of visual input frames, state supervision, and the decoupling approach. For the types of layers on the time dimension of visual input frames, we compare the average pooling layer with our shared MLP network, both without state supervision and decoupling. As shown in Figure 5(b), the average pooling layer makes wrong predictions on all actions. However, the model with the shared MLP network predicts the first and last actions correctly, as illustrated in Figure 5(c). This is because the MLP network can extract the time information from the initial and goal visual observations more thoroughly than the average pooling layer. We further simulate the shared MLP network without decoupling but with language state supervision[41] in Figure 5(d). Comparing", + "bbox": [ + 75, + 628, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/87ab1d8252ae3fbd6c85244c8de2f3b09d05c88db3679ec5a57709b44f822e21.jpg", + "image_caption": [ + "Figure 5. Case study on Skip-Plan components at $T = 4$ : (a) ground truth; (b) average pooling: average pooling layer on the time dimension of visual input frames without decoupling and state supervision; (c) MLP: shared MLP network on the time dimension of visual input frames without decoupling and state supervision; (d) MLP with state supervision: shared MLP network with state supervision but without decoupling; (e) MLP with decoupling: shared MLP network with decoupling but without state supervision. The outlines of correct and incorrect action predictions are coloured in green and red respectively." + ], + "image_footnote": [], + "bbox": [ + 498, + 85, + 893, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 5(d) with Figure 5(c), we notice the addition of state supervision cannot correct the predictions of the second and the third actions, because the state supervision does not reduce the compound error. In contrast, the shared MLP network without state supervision but with decoupling can lower the compound error effectively, resulting in correct predictions of all actions, as demonstrated in Figure 5(e).", + "bbox": [ + 496, + 422, + 893, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 551, + 617, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose Skip-Plan, a condensed action space learning method for the procedure planning task. Our Skip-Plan solves the procedure planning problem mathematically in two novel ways. First, Skip-Plan skips the intermediate state supervision and only focuses on action relations. Second, Skip-Plan decomposes long action chains into multiple reliable sub-chains by skipping most uncertain intermediate actions. In these ways, our Skip-Plan model explores all sorts of sub-relations within an action chain and generates actions non-adjacently only in the condensed action space. Extensive experiments show the superior performance of our Skip-Plan method on two instructional video datasets. Skip-Plan has the potential to inspire the research field of procedure planning, since most of them heavily rely on state observations and adjacent actions at all times.", + "bbox": [ + 496, + 580, + 890, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0114903 and in part by the National Natural Science Foundation of China under Grant 62125603 and Grant 62206153, CAAI-Huawei MindSpore Open Fund.", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10304", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mindspore. .https://www.mindspore.cn/.5", + "[2] Armenii Ashukha, Alexander Lyzhov, Dmitry Molchanov, and Dmitry Vetrov. Pitfalls of in-domain uncertainty estimation and ensembling in deep learning. arXiv preprint arXiv:2002.06470, 2020. 3", + "[3] Anil Batra, Shreyank Gowda, Laura Sevilla-Lara, and Frank Keller. A closer look at temporal ordering in the segmentation of instructional videos. arXiv preprint arXiv:2209.15501, 2022. 2", + "[4] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model-based policy learning. In ICCV, pages 15611-15620, 2021. 2, 5, 6", + "[5] Nicolas Biard, Salomé Cojean, and Eric Jamet. Effects of segmentation and pacing on procedural learning by video. Comput. Hum. Behav., 89:411-417, 2018. 2", + "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 3", + "[7] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In ECCV, pages 334–350, 2020. 1, 2, 3, 5, 6", + "[8] Chekuri Choudary and Tiecheng Liu. Summarization of visual content in instructional videos. IEEE Trans. Multimed., 9(7):1443-1455, 2007. 2", + "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3", + "[10] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 3", + "[11] Mikita Dvornik, Isma Hadji, Konstantinos G Derpanis, Animesh Garg, and Allan Jepson. Drop-dtw: Aligning common signal between sequences while dropping outliers. NeuRIPS, 34:13782-13793, 2021. 2", + "[12] Jakob Gawlikowski, Cedrique Rovile Njieutcheu Tassi, Mohsin Ali, Jongseok Lee, Matthias Humt, Jianxiang Feng, Anna Kruspe, Rudolph Trielbel, Peter Jung, Ribana Roscher, et al. A survey of uncertainty in deep neural networks. arXiv preprint arXiv:2107.03342, 2021. 3", + "[13] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, Chiho Choi, and Behzad Dariush. Weakly-supervised online action segmentation in multi-view instructional videos. In CVPR, pages 13780-13790, 2022. 2", + "[14] Reza Ghoddoosian, Saif Sayed, and Vassilis Athitsos. Hierarchical modeling for task recognition and action segmentation in weakly-labeled instructional videos. In WACV, pages 1922-1932, 2022. 2", + "[15] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. arXiv preprint arXiv:1610.02136, 2016. 3" + ], + "bbox": [ + 78, + 116, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Gabriel Huang, Bo Pang, Zhenhai Zhu, Clara Rivera, and Radu Soricut. Multimodal pretraining for dense video captioning. arXiv preprint arXiv:2011.11760, 2020. 2", + "[17] Eyke Hüllermeier and Willem Waegeman. Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods. Mach. Learn, 2021. 3", + "[18] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In CVPRW, pages 958-959, 2020. 2", + "[19] Lei Ji, Chenfei Wu, Daisy Zhou, Kun Yan, Edward Cui, Xilin Chen, and Nan Duan. Learning temporal video procedure segmentation from an automatically collected large dataset. In WACV, pages 1506-1515, 2022. 2", + "[20] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? NeurIPS, 2017. 3", + "[21] Colin Lea, Michael D Flynn, Rene Vidal, Austin Reiter, and Gregory D Hager. Temporal convolutional networks for action segmentation and detection. In CVPR, pages 156-165, 2017. 2", + "[22] Muheng Li, Lei Chen, Yueqi Duan, Zhilan Hu, Jianjiang Feng, Jie Zhou, and Jiwen Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In CVPR, pages 19880-19889, 2022. 2", + "[23] Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. arXiv preprint arXiv:1706.02690, 2017. 3", + "[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 4", + "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 3", + "[26] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2", + "[27] Andrey Malinin and Mark Gales. Predictive uncertainty estimation via prior networks. NeurIPS, 2018. 3", + "[28] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In CVPR, pages 9879-9889, 2020. 2", + "[29] Medhini Narasimhan, Arsha Nagrani, Chen Sun, Michael Rubinstein, Trevor Darrell, Anna Rohrbach, and Cordelia Schmid. Tl; dw? summarizing instructional videos with task relevance and cross-modal saliency. In ECCV, pages 540-557, 2022. 2", + "[30] Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. NeurIPS, 32, 2019. 3", + "[31] AJ Piergiovanni, Anelia Angelova, Michael S Ryoo, and Irfan Essa. Unsupervised action segmentation for instructional videos. arXiv preprint arXiv:2106.03738, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10305", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Alireza Shafaei, Mark Schmidt, and James J Little. A less biased evaluation of out-of-distribution sample detectors. arXiv preprint arXiv:1809.04729, 2018. 3", + "[33] Anshul Shah, Benjamin Lundell, Harpreet Sawhney, and Rama Chellappa. Steps: Self-supervised key step extraction from unlabeled procedural videos. arXiv preprint arXiv:2301.00794, 2023. 2", + "[34] Botian Shi, Lei Ji, Yaobo Liang, Nan Duan, Peng Chen, Zhendong Niu, and Ming Zhou. Dense procedure captioning in narrated instructional videos. In ACL, pages 6382-6391, 2019. 2", + "[35] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In ICCV, pages 7464-7473, 2019. 2", + "[36] Jiankai Sun, De-An Huang, Bo Lu, Yun-Hui Liu, Bolei Zhou, and Animesh Garg. Plate: Visually-grounded planning with transformers in procedural tasks. IEEE Robot. Autom. Lett., 7(2):4924-4930, 2022. 2, 6", + "[37] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In CVPR, pages 1207-1216, 2019. 1, 2, 5", + "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3", + "[39] Mengmeng Xu, Juan-Manuel Pérez-Rúa, Victor Escorcia, Brais Martinez, Xiatian Zhu, Li Zhang, Bernard Ghanem, and Tao Xiang. Boundary-sensitive pre-training for temporal localization in videos. In ICCV, pages 7220-7230, 2021. 2", + "[40] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Temporal query networks for fine-grained video understanding. In CVPR, pages 4486-4496, 2021. 3", + "[41] He Zhao, Isma Hadji, Nikita Dvornik, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In CVPR, pages 2938-2948, 2022. 2, 5, 6, 7, 8", + "[42] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, pages 7590-7598, 2018. 2", + "[43] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In CVPR, pages 8746-8755, 2020. 2", + "[44] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In CVPR, pages 3537-3545, 2019. 1, 2, 5" + ], + "bbox": [ + 78, + 90, + 468, + 785 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10306", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Skip-Plan_ Procedure Planning in Instructional Videos via Condensed Action Space Learning/c1a3a0ec-3885-4f2f-93dd-62793c50f314_model.json b/2023/Skip-Plan_ Procedure Planning in Instructional Videos via Condensed Action Space Learning/c1a3a0ec-3885-4f2f-93dd-62793c50f314_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e7a8d8b53d81a00a23a3efabbad186c43f0a4a2c --- /dev/null +++ b/2023/Skip-Plan_ Procedure Planning in Instructional Videos via Condensed Action Space Learning/c1a3a0ec-3885-4f2f-93dd-62793c50f314_model.json @@ -0,0 +1,1903 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.209, + 0.131, + 0.761, + 0.177 + ], + "angle": 0, + "content": "Skip-Plan: Procedure Planning in Instructional Videos via Condensed Action Space Learning" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.203, + 0.851, + 0.222 + ], + "angle": 0, + "content": "Zhiheng Li\\(^{1}\\), Wenjia Geng\\(^{2}\\), Muheng Li\\(^{1}\\), Lei Chen\\(^{3}\\), Yansong Tang\\(^{2*}\\), Jiwen Lu\\(^{1,4}\\), Jie Zhou\\(^{1,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.287, + 0.222, + 0.684, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\) Department of Automation, Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.24, + 0.735, + 0.257 + ], + "angle": 0, + "content": "2 Shenzhen International Graduate School, Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.257, + 0.673, + 0.275 + ], + "angle": 0, + "content": "3 Beijing University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.275, + 0.786, + 0.293 + ], + "angle": 0, + "content": "4 Beijing National Research Center for Information Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.295, + 0.826, + 0.327 + ], + "angle": 0, + "content": "{lizhihan21@gengwj22@,li-mh20}@mails.tsinghua.edu.cn,chenlei2022@ustb.edu.cn, {tang.yansong@sz.,lujiwen,jzhou@}tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.362, + 0.313, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.395, + 0.474, + 0.699 + ], + "angle": 0, + "content": "In this paper, we propose Skip-Plan, a condensed action space learning method for procedure planning in instructional videos. Current procedure planning methods all stick to the state-action pair prediction at every timestep and generate actions adjacently. Although it coincides with human intuition, such a methodology consistently struggles with high-dimensional state supervision and error accumulation on action sequences. In this work, we abstract the procedure planning problem as a mathematical chain model. By skipping uncertain nodes and edges in action chains, we transfer long and complex sequence functions into short but reliable ones in two ways. First, we skip all the intermediate state supervision and only focus on action predictions. Second, we decompose relatively long chains into multiple short sub-chains by skipping unreliable intermediate actions. By this means, our model explores all sorts of reliable sub-relations within an action sequence in the condensed action space. Extensive experiments show Skip-Plan achieves state-of-the-art performance on the CrossTask and COIN benchmarks for procedure planning." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.728, + 0.21, + 0.743 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.754, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Given the initial and goal visual observations which are curated from instructional videos [44, 37], the task of procedure planning is to predict a sequence of actions to reach the indicated goal [7], as illustrated in Figure 1(a). This is an emerging research direction in the field of video understanding. Procedure planning has worldwide applications and assists humans with instruction-following tasks, such as cooking, assembling furniture, electronics repair, etc." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.364, + 0.695, + 0.39 + ], + "angle": 0, + "content": "Question: What action sequence are needed to get from \\( v_{start} \\) to \\( v_{goal} \\)?" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.392, + 0.698, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.704, + 0.364, + 0.884, + 0.376 + ], + "angle": 0, + "content": "Answer: Required action sequence" + }, + { + "type": "text", + "bbox": [ + 0.73, + 0.376, + 0.852, + 0.387 + ], + "angle": 0, + "content": "(Time Horizon \\(T = 4\\)" + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.389, + 0.871, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.461, + 0.761, + 0.473 + ], + "angle": 0, + "content": "(a) Problem Definition" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.478, + 0.864, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.55, + 0.825, + 0.562 + ], + "angle": 0, + "content": "(b) Conventional Procedure Planning Paradigm" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.568, + 0.868, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.666, + 0.653, + 0.741, + 0.665 + ], + "angle": 0, + "content": "(c)Skip-Plan" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.67, + 0.885, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.871 + ], + "angle": 0, + "content": "Figure 1. Comparisons of the conventional paradigm and our proposed Skip-Plan in procedure planning. (a) Problem definition: given the initial and goal visual observations \\((v_{start}\\) and \\(v_{goal})\\), procedure planning is to answer an action sequence to reach the indicated goal with a pre-defined time horizon \\(T\\). (b) Conventional paradigm predicts a state-action pair \\(\\{\\tilde{s}_t,\\tilde{a}_t\\}\\) at every timestep \\(t\\), and repeat it step-by-step. (c) Our Skip-Plan skips state supervision and extracts multiple sub-chains by skipping most intermediate actions. The final action sequence is accumulated from all predicted sub-chains. Skip-Plan significantly reduces the overall compound error, indicated by the shallower darkness of the chevron arrows." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.292, + 0.9 + ], + "angle": 0, + "content": "* indicates the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10297" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.473, + 0.531 + ], + "angle": 0, + "content": "In the past few years, various approaches have been proposed to solve the procedure planning problem. Most existing methods utilize an autoregressive model with visual state supervision to generate actions one by one [7, 4, 36]. Another type of work [41] solves it in a non-autoregressive way with language state supervision. The difficulties of procedure planning have been widely discussed in the previous works. For instance, the sequential orders in procedure planning can be altered but still sensible (e.g., switching the order of 'pour water' and 'pour alcohol' does not matter in many cases), and the visual states are unstructured and high-dimensional. Thus, we can conclude the following limitations from the existing works. (a) State supervision: the major difficulty in procedure planning is state supervision. Visual images are high-dimensional and complex, with extra misleading information such as background, illumination, viewing point, etc. (b) Step-by-step predictions: the conventional paradigm in procedure planning predicts one state-action pair at every timestep, and repeats it step-by-step, as illustrated in Figure 1(b). Even though the proposed non-autoregressive model generates whole action sequences in batches [41], it can be still regarded as a step-by-step method because the model neglects the links between non-adjacent actions. From Figure 1(b), we can see while predicting an action sequence with a time horizon of 4, the conventional model is equivalently predicting a sequence with a length of 7 (4 action nodes, 3 state nodes). More frequent predictions tend to accumulate a larger compound error along the chain, which lowers the prediction accuracy." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.534, + 0.473, + 0.777 + ], + "angle": 0, + "content": "To overcome these limitations, we propose Skip-Plan, a condensed action space learning method that solves the procedure planning problem through the lens of mathematics, aiming to reduce compound error along action chains. As we have discussed above, existing methods struggle with high-dimensional and complex state supervision. Thus, the compound error along the action chain is increased due to large errors in state predictions. This motivates us to skip all the state nodes in our Skip-Plan framework and only model sequence relations in the condensed action space, as demonstrated in Figure 1(c). Without the presence of state supervision, a lower compound error results in better performance in Skip-Plan. Although the dependencies between high-level actions without state supervision are very abstract, the algorithm of our Skip-Plan is highly mathematically sensible in the Markov chain model." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Another convention in procedure planning is we should learn sequential relations between adjacent actions. The Markov chain model inspires us to decouple relatively long chains into multiple short sub-chains, because the compound error can be reduced with shorter chain lengths. The problem here is how to decouple these long chains, because there are various choices of sub-chains with different lengths and we should always choose the most reliable" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.319 + ], + "angle": 0, + "content": "ones under any condition. Given the initial and goal frames as concrete inputs, it is apparent the reliabilities of the initial and final actions are the highest, and the intermediate actions are more likely to be predicted wrong. Therefore, our Skip-Plan decomposes long chains by 'skipping' some unreliable intermediate actions, and each sub-chain in our model is composed of the initial action, one unique intermediate action, and the final action. One example for a chain length of 4 is shown in Figure 1(c). In this way, the reliable sub-relations between non-adjacent actions can be well explored by the network. To model the discriminative relations of all sub-chains, we propose sub-chain decoders, where each decoder only learns one sub-chain by the specific loss design. The final action sequence is aggregated from all sub-chain predictions by a sub-chain accumulator." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.895, + 0.502 + ], + "angle": 0, + "content": "In summary, our contributions are threefold. (a) Unlike the previous methods which put sophisticated designs on state supervision, Skip-Plan skips unreliable state predictions to boost the model performance in procedure planning. (b) In contrast to the existing methods which only focus on the links between adjacent actions, we decompose long chains into multiple reliable sub-chains by skipping some intermediate actions, and propose sub-chain decoders to learn their discriminative relations. (c) Extensive experiments show Skip-Plan achieves state-of-the-art performance on the CrossTask [44] and COIN [37] benchmarks for procedure planning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.642, + 0.529 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.75 + ], + "angle": 0, + "content": "Procedure Planning. Procedure planning in instructional videos was first introduced by Chang et al. [7]. Recently, various methods have been conducted to solve this task. Chang et al. [7] propose a dual dynamic network to conjugate the relationships between states and actions. Sun et al. [36] utilize a GPT network to generate actions and states one by one. Bi et al. [4] use the model-based reinforcement learning with time-invariant context information to predict state-action pairs autoregressively. Different from the previous autoregressive methods with visual state supervision, Zhao et al. [41] use a non-autoregressive transformer decoder to generate action sequences with language state supervision. Unlike previous works, Skip-Plan generates actions non-adjacently without any state supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Instructional Videos Analysis. Instructional videos are used for learning human behaviours because they are visual representations of human actions. Based on instructional videos, various action-related tasks have been proposed, such as action segmentation[21, 22, 13, 14, 28, 31], step localization[39, 11, 43, 33, 26], procedure segmentation[42, 19, 3, 5], video summarization/captioning[8, 35, 29, 34, 16, 18], etc. Different from these action tasks in instructional videos, we investigate procedure planning, a goal-conditioned problem that plans a sequence of actions given" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10298" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.338, + 0.106 + ], + "angle": 0, + "content": "the initial and goal visual observations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.473, + 0.364 + ], + "angle": 0, + "content": "Predictive Uncertainty on Sequence Relations. All types of questions that we are interested in can be modeled as a sequence model, such as images [25] and languages [9]. Currently, Transformer [38] is the most effective network to model sequence relations, and our model is based on the transformer decoder with a learnable query input[40, 6]. However, sequence relations are always distorted by predictive uncertainty[12, 27, 17, 30, 20], and the sources of predictive uncertainty can be various, including model uncertainty, data uncertainty, etc[2, 10, 15, 23, 32]. In this work, we pay particular attention to data uncertainty, in which the uncertainty directly stems from the data. Data uncertainty is caused by information loss when we measure a data distribution in a space with incorrect dimensions (e.g., we must lose depth information if we measure a three-dimensional object in a two-dimensional space). Therefore, we often avoid high-dimensional data for lower data uncertainty." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.375, + 0.176, + 0.391 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.4, + 0.281, + 0.415 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.423, + 0.471, + 0.53 + ], + "angle": 0, + "content": "We follow the definition of procedure planning proposed by Chang et al. [7]. Here, we denote the action and the state at timestep \\( t \\) as \\( a_{t} \\) and \\( s_{t} \\) respectively. Given the initial visual observation \\( v_{start} \\) and the goal visual observation \\( v_{goal} \\), the model generates a sequence of actions \\( a_{1:T} \\) to reach the indicated goal. In this work, we introduce a novel formulation for the procedure planning task:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.536, + 0.47, + 0.577 + ], + "angle": 0, + "content": "\\[\np \\left(a _ {1: T} \\mid v _ {s}, v _ {g}\\right) = \\prod_ {t = 2} ^ {T - 1} p \\left(a _ {t} \\mid a _ {1}, a _ {T}\\right) p \\left(a _ {1}, a _ {T} \\mid v _ {s}, v _ {g}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.473, + 0.841 + ], + "angle": 0, + "content": "where \\( v_{s} \\) and \\( v_{g} \\) short for ' \\( v_{start} \\) ' and ' \\( v_{goal} \\) ' . We denote \\( p(a_{1:T}|v_{start},v_{goal}) \\) as the conditional probability distribution of the action sequence given the initial visual observation \\( v_{start} \\) and the goal visual observation \\( v_{goal} \\), \\( p(a_t|a_1,a_T) \\) as the conditional probability distribution of intermediate actions \\( (2\\leq t\\leq T - 1) \\) given the initial action \\( a_1 \\) and the final action \\( a_{T} \\), and \\( p(a_{1},a_{T}|v_{start},v_{goal}) \\) as the conditional probability distribution of the initial action \\( a_1 \\) and the final action \\( a_{T} \\) given the initial visual observation \\( v_{start} \\) and the goal visual observation \\( v_{goal} \\). Equation 1 reveals the uniqueness of our Skip-Plan method: First, we do not rely on any intermediate state supervision \\( s_t \\) for action generations. Second, every intermediate action prediction \\( a_{t} \\) \\( (2\\leq t\\leq T - 1) \\) only depends on the initial action \\( a_1 \\) and the final action \\( a_{T} \\). Thus, Skip-Plan does not focus on sequential relations between adjacent actions but predicts action sequences non-adjacently without state supervision." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.458, + 0.865 + ], + "angle": 0, + "content": "3.2. Action Predictions without State Supervision" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Information does not always help in the decision-making process. Sometimes ambiguous information can even lead" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.575 + ], + "angle": 0, + "content": "to sub-optimal results, and plausible plans can be only made based on reliable information. In procedure planning, two types of supervision can be utilized, including actions \\( a_{t} \\) and visual states \\( s_{t} \\). Even though they can be directly generated from each other, the intrinsic properties of these two supervisions are largely different. We evaluate them from the perspectives of the value of information and predictive uncertainty. Visual state supervision is well known to be high-dimensional. In addition to the factor of action class, various factors such as background, viewing point, illumination, etc. can be involved in visual states. Thus, useful information in visual state supervision for action predictions is overwhelmed by many irrelevant factors. In contrast, action labels are the most condensed supervision, without being interrupted by any factors. From this point, we argue the value of information offered by state supervision is extremely low because the portion of information, which is relevant to action predictions, is too small. On the other hand, when models generate high-dimensional predictions such as visual images, large predictive errors are more likely to occur. Thus, the predictive uncertainty of intermediate visual states is significantly higher than the one of condensed action predictions. Combining the effects of the value of information and the predictive uncertainty, we argue the existence of visual state supervision hurts the predictive accuracy of sequential actions. Even though we can use the weak language state supervision to reduce some irrelevant dimensions, it is still too high-dimensional compared to condensed action labels. Therefore, we 'skip' intermediate state supervision in our Skip-Plan framework. The overall compound error along an action chain is significantly reduced without the predictive errors introduced by states." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.895, + 0.789 + ], + "angle": 0, + "content": "Getting rid of intermediate state supervision does not mean all sorts of visual information are useless. In procedure planning, the initial visual observation \\( v_{start} \\) and the goal visual observation \\( v_{goal} \\) are vital because action sequences are constrained by them. These visual observations are in the form of frame features in seconds. Inspired by the image analysis above, we argue the visual input information is underutilized by the previous works, especially along the time dimension. In the Skip-Plan framework, we take a shared multi-layer perceptrons (MLP) network in the Visual Input Module to learn the time information, where the MLP network is three-layered with ReLU activations. This simple method can boost the model performance by a large margin, especially for short sequences \\( (T = 3) \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.774, + 0.817 + ], + "angle": 0, + "content": "3.3. Decomposition of Long Chains" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Long chains are not favored when a compound error is accumulating. Thus, decomposing long chains into short ones is a natural way to better model performance. Through the lens of mathematics, the only design philosophy here is that we should extract most reliable sub-chains from the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10299" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.891, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.288, + 0.895, + 0.4 + ], + "angle": 0, + "content": "Figure 2. Our Skip-Plan architecture. The shared MLP network in the Visual Input Module extracts the initial and final information from the initial and goal observations (\\( v_{start} \\) and \\( v_{goal} \\)) and attaches them to the first and last dimensions of every learnable query input. A number of \\( T - 2 \\) query inputs are separately passed to each decoder in the Sub-chain Decoder Module. In each decoder, the input is processed by the self-attention block and then by the cross-attention block, where the cross-attention block is augmented by a learnable memory \\( M_i \\). The sub-chain decoders are separated. The \\( n \\)th decoder outputs the sub-chain \\( \\{ \\tilde{a}_1, \\tilde{a}_t, \\tilde{a}_T \\} \\) with \\( t = n + 1 \\). The predicted sub-chains are supervised by two loss supervision. In Individual Sub-chain Supervision, the predicted sub-chains are supervised by the corresponding ground-truth sub-chains. In Complete Chain Supervision, the predicted sub-chains are aggregated by the Sub-chain Accumulator to generate the complete action chain, which is supervised by the complete ground-truth action chain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.47, + 0.484 + ], + "angle": 0, + "content": "original long chain. Following this principle, we ought to find there is so much flexibility in the decoupling method: We can choose sub-chains with any lengths, and actions in the sub-chains are not necessary to be adjacent." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.471, + 0.697 + ], + "angle": 0, + "content": "The reliability of a sub-chain can be determined by the uncertainty of predictions. According to the uncertainty analysis, the reliabilities of the initial and final actions are the highest because they are the closest actions to the initial and goal visual inputs. Since the compound error is accumulated from two ends of the chain, the reliability of action predictions is decreasing while approaching the middle of the chain. Consequently, we remove all the edges between adjacent intermediate actions, and each individual intermediate action only connects to the reliable initial and final actions. This design just coincides with the term \\( p(a_{t}|a_{1},a_{T}) \\) with \\( 2\\leq t\\leq T - 1 \\) in Equation 1. In this way, we generate a quantity of \\( T - 2 \\) sub-chains from an action chain with a time horizon of \\( T \\). The set of sub-chains is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.707, + 0.47, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\{(a _ {1}, a _ {2}, a _ {T}), (a _ {1}, a _ {3}, a _ {T}), \\dots , (a _ {1}, a _ {T - 1}, a _ {T}) \\}. \\quad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.795 + ], + "angle": 0, + "content": "Each sub-chain has a length of 3, with the order of the initial action, one unique intermediate action, and the final action. This decoupling method works well for relatively long chains \\((T > 3)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Given multiple short sub-chains, we design a compatible network to model all their sub-relations. Since these sub-chains are parallel and independent to each other, we propose the Sub-chain Decoder Module, where each transformer decoder only models one discriminative relation of one sub-chain. Every transformer decoder is non-autoregressive with two types of inputs: a learnable query" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.895, + 0.622 + ], + "angle": 0, + "content": "with the initial and goal visual features, and a learnable memory component. The novelty here is each decoder only learns one sub-relation, with a unique loss only corresponding to the three actions of that sub-chain. The query input, parameters, and memory component of every decoder are separated from other decoders'. Therefore, the number of decoders in the Sub-chain Decoder Module is equal to the number of sub-chains \\((T - 2)\\) extracted from the original chain. The final sequence generation is aggregated from all the sub-chain predictions through the Sub-chain Accumulator, which is an MLP network. In this way, our Skip-Plan network can model all sorts of discriminative sub-relations within the original action sequence." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.628, + 0.607, + 0.645 + ], + "angle": 0, + "content": "3.4. Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.894, + 0.713 + ], + "angle": 0, + "content": "Loss is an essential component to discriminate sub-chain decoders. Instead of conventional cross-entropy loss, we use the focal loss to supervise our sub-chain decoders. The focal loss \\( FL \\) is formulated as[24]:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.718, + 0.892, + 0.759 + ], + "angle": 0, + "content": "\\[\nF L = - \\sum_ {t = 1} ^ {T} a _ {t} (1 - \\tilde {a} _ {t}) ^ {\\gamma} \\log \\tilde {a} _ {t}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "where we denote \\(a_{t}\\) as the ground-truth one-hot labels, \\(\\tilde{a}_{t}\\) as the predicted probability distribution of the action sequence, and \\(\\gamma\\) as the gamma value of focal loss. The focal loss is equivalent to the cross-entropy loss when \\(\\gamma = 0\\). The model training is benefited from the use of focal loss because the data distribution in procedure planning is imbalanced. In this work, we define \\(FL(a_{1},a_{t},a_{T})\\) as the focal loss for the sub-chain with a length of 3, where the sub-chain is composed of the initial action \\(a_{1}\\), the action at timestep \\(t\\) \\(a_{t}\\), and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10300" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "the final action \\( a_{T} \\). We present the focal loss of the whole action sequence as \\( FL(a_{1:T}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.47, + 0.196 + ], + "angle": 0, + "content": "Loss of Individual Sub-chains. As we have discussed before, the loss of each decoder only corresponds to the actions within that sub-chain. Therefore, the individual loss of the \\( n \\)th decoder is defined as \\( FL(a_{1},a_{t},a_{T}) \\) with \\( t = n + 1 \\), and the loss from all sub-chain decoders can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.206, + 0.469, + 0.248 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {N} = \\sum_ {t = 2} ^ {T - 1} F L \\left(a _ {1}, a _ {t}, a _ {T}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.469, + 0.318 + ], + "angle": 0, + "content": "Loss of Complete Chain. After aggregating all sub-chains by the Sub-chain Accumulator, the sequence output with a length of \\( T \\) is supervised by the focal loss of the whole action sequence:" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.331, + 0.469, + 0.348 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T} = F L \\left(a _ {1: T}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.469, + 0.403 + ], + "angle": 0, + "content": "Overall Loss. Therefore, the total loss is defined as the sum of the loss from all sub-chains and the complete chain loss, which can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.417, + 0.469, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {N} + \\mathcal {L} _ {T}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.469, + 0.49 + ], + "angle": 0, + "content": "In this way, our model can explore all sorts of discriminative but reliable sub-relations within an action sequence, while guaranteeing the final sequence generation is optimal." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.499, + 0.295, + 0.515 + ], + "angle": 0, + "content": "3.5. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.765 + ], + "angle": 0, + "content": "For each decoder, the query input length is \\( T + 1 \\), and the memory size is 128. The numbers of decoder layers, decoder heads, and decoder dimensions are set to be 1, 16, and 1024 respectively. All MLP networks are three-layered with ReLU activations. In the Visual Input Module, our shared MLP network on the time dimension of input frame features has the shape \\( [3 \\to 6 \\to 1] \\). All action classifiers in the Sub-chain Decoder Module are the same MLP networks with the shape \\( [1024 \\to 512 \\to n_a] \\), where \\( n_a \\) is the number of action classes. The Sub-chain Accumulator is an MLP network with the shape \\( [T(T - 2) \\to 3T(T - 2) \\to T] \\). Models are optimized for 500 epochs with stochastic gradient descent. The initial learning rate is 0.02 and decays by 10 at every 50 epochs after 100 epochs. Experiments are conducted on a single GeForce RTX 3090 GPU. We also test our model using the Mindspace [1]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.778, + 0.21, + 0.794 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.273, + 0.817 + ], + "angle": 0, + "content": "4.1. Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Datasets. We evaluate our model on two real-world instructional video datasets: CrossTask[44] and COIN[37]. CrossTask contains 2763 videos with a duration of 213 hours. It consists of 18 different tasks with 7.4 steps per task. The COIN dataset spans 12 different domains related" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.257 + ], + "angle": 0, + "content": "to our daily life. It contains 11,827 videos with 3.9 steps per video. The COIN dataset is challenging due to the diversity of its tasks and action labels (180 tasks and 779 action labels). We curate a set of plans by the sliding window approach, and adopt the standard \\(70\\% / 30\\%\\) to create our train/test splits. The division is done on the generated samples rather than the original video. The data pre-processing procedure in our work is identical to the most widely used protocol in the previous procedure planning works (e.g., DDN [7], Ext-GAIL [4], and the main protocol in P3IV [41]) for fair comparisons." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.892, + 0.485 + ], + "angle": 0, + "content": "Metrics. Following the previous works, we used three metrics to comprehensively evaluate model performance. For each action sequence: (a) mean Intersection over Union (mIoU) regards the predicted and ground-truth action sequences as two sets, and measure the overlap between these sets. The mIoU metric only indicates whether desired actions exist in the predicted action set, agnostic to the order of actions. (b) mean Accuracy (mAcc) evaluates the correctness of the predicted action at every timestep, termed as percentages. Thus, mAcc counts the order of actions, and is more strict than mIoU. (c) Success Rate (SR) evaluates the correctness of the whole action sequence. A predicted sequence is successful only if it matches the ground-truth sequence. The success rate is the proportion of successful samples over all samples, which is more strict than mAcc." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.592 + ], + "angle": 0, + "content": "Training Scheme on Each Dataset. We specify an appropriate gamma value on each dataset to maximize model performance. Through experiments, we find the gamma values of 1.5 and 0 are the best parameters for the CrossTask and COIN datasets respectively. This is reasonable because the data distribution in the COIN dataset is more balanced due to a large quantity of action labels in diverse tasks." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.605, + 0.779, + 0.621 + ], + "angle": 0, + "content": "4.2. Evaluating Procedure Planning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Short-horizon Prediction. We compare the model performance of the short-horizon predictions \\((T = 3 / 4)\\) in Table 1. As illustrated in Table 1, Skip-Plan significantly outperforms the previous state-of-the-art model at all metrics for \\(T = 3\\). In the CrossTask dataset, our SR and mAcc results surpass the results of the second best approach (e.g., P3IV[41]) by \\(6\\%\\) and \\(11\\%\\). Our method also outperforms P3IV[41] by \\(8\\%\\) on SR and \\(25\\%\\) on mAcc in the COIN dataset. Such incredible improvements strongly prove Skip-Plan without state supervision suppresses predictive uncertainty effectively, and our module design on the visual input extracts initial and goal visual information thoroughly. For \\(T = 4\\), Skip-Plan achieves large gains of around \\(5\\%\\), \\(24\\%\\), and \\(7\\%\\) in SR, mAcc and mIoU respectively in the COIN dataset. For the CrossTask dataset, the mAcc and mIoU of our Skip-Plan model are the best and the SR is the second best. Particularly, our mAcc is higher than P3IV's[41] by \\(11.48\\%\\). The methods of no state supervision and decou" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10301" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.894, + 0.117 + ], + "angle": 0, + "content": "Table 1. Metric results on CrossTask and COIN for short-horizon predictions \\( \\left( {T \\in \\{ 3,4\\} }\\right) \\) . The ’ \\( \\checkmark \\) ’ and ’ \\( X \\) ’ in the column State Supervision denote with and without state supervision respectively. The best and second best metric results are coloured in red and blue respectively." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.119, + 0.85, + 0.346 + ], + "angle": 0, + "content": "
HorizonsModelsState SupervisionCrossTaskCOIN
SRmAccmIoUSRmAccmIoU
T=3DDN[7]✓12.1831.2947.4813.9020.1964.78
Plate[36]✓16.0036.1765.91---
Ext-GAIL[4]✓21.2749.4661.70---
P3IV[41]✓23.3449.9673.8915.4021.6776.31
Skip-Plan✗28.8561.1874.9823.6547.1278.44
T=4DDN[7]✓5.9727.1048.4611.1317.7168.06
Plate[36]✓14.0035.2955.36---
Ext-GAIL[4]✓16.4143.0560.93---
P3IV[41]✓13.4044.1670.0111.3218.8570.53
Skip-Plan✗15.5655.6470.3016.0443.1977.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.368, + 0.468, + 0.394 + ], + "angle": 0, + "content": "Table 2. Metric results for long-horizon predictions \\( (T \\in \\{5,6\\}) \\) on CrossTask and COIN." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.394, + 0.467, + 0.578 + ], + "angle": 0, + "content": "
DatasetsHorizonsModelsSRmAccmIoU
CrossTaskT=5DDN[7]3.10--
P3IV[41]7.21--
Skip-Plan8.5552.0569.09
T=6DDN[7]1.20--
P3IV[41]4.40--
Skip-Plan5.1247.3466.11
COINT=5P3IV[41]4.2710.8168.81
Skip-Plan9.9038.9976.93
T=6P3IV[41]4.178.9767.72
Skip-Plan6.6133.6373.68
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.63, + 0.468, + 0.646 + ], + "angle": 0, + "content": "pling long chains both drive Skip-Plan to succeed at \\(T = 4\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Long-horizon Prediction. We further evaluate Skip-Plan on long-horizon predictions (\\(T = 5/6\\)). In the COIN dataset, our Skip-Plan performs extraordinarily well on the metric of mAcc, with remarkable improvements of \\(28\\%\\) and \\(25\\%\\) for \\(T = 5/6\\), compared to the second best model P3IV[41]. As for the metrics of SR and mIoU, Skip-Plan exceeds P3IV[41] by \\(6\\%\\) and \\(8\\%\\) at \\(T = 5\\), and \\(2\\%\\) and \\(6\\%\\) at \\(T = 6\\) respectively. The outstanding performance we presented above further validates the success of the decoupling approach. Our Skip-Plan also achieved the best results on SR in the CrossTask dataset, surpassing P3IV[41] by \\(1.34\\%\\) and \\(0.72\\%\\) at \\(T = 5/6\\). For the long-horizon predictions, we argue mAcc and mIoU are more reliable metrics to measure the effectiveness of a model. SR is interrupted by a large variance, because data samples become insufficient with a large time horizon." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.371, + 0.733, + 0.387 + ], + "angle": 0, + "content": "4.3. Prerequisite Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.488 + ], + "angle": 0, + "content": "In this section, we carried out several prerequisite experiments for the long-chain decoupling method. They can validate some statements that we have made in Section 3.3. These experiments are foundational because they proved the prerequisite conditions for the decoupling approach to work. Thus, we name them as the prerequisite experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.893, + 0.761 + ], + "angle": 0, + "content": "Error Rate Distribution along a Predicted Chain. We tend to remove the edges between unreliable nodes and only connect one unreliable node to other reliable ones. In this way, we can filter out most reliable sub-chains from the original long chain. Here, we measure the reliability of each node in terms of the error rate distribution along the non-autoregressively predicted chain. The error rate distributions along the predicted chains with \\( T = 3/4/5/6 \\) are illustrated in Figure 3. It is clear the reliabilities of the first and final action nodes are the highest and the model is more likely to make wrong predictions while approaching the middle of the chain. It is not hard to understand this phenomenon: a higher error rate occurs at intermediate action predictions because they are further away from the initial and goal visual observations. This empirical finding motivates us to remove the edges between unreliable intermediate actions and only connect each intermediate action to the reliable initial and final action nodes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Reliability of Standalone Sub-chains. In addition to the reliability of each node, we further study the reliability of the sub-chains that we have decoupled from the original long chain. For a clearer explanation, we take a chain with a length of 5 as an example (e.g., \\(\\{a_1, a_2, a_3, a_4, a_5\\}\\)), and only study the reliability of the sub-chain \\(\\{a_1, a_3, a_5\\}\\). We consider two cases: First, we just model the relation of the whole chain and only investigate the prediction result of that sub-chain contained in the long chain. In other words, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10302" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.095, + 0.465, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.265, + 0.47, + 0.295 + ], + "angle": 0, + "content": "Figure 3. Error rate distributions along the non-autoregressively predicted chains with \\( T = 3/4/5/6 \\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.306, + 0.468, + 0.334 + ], + "angle": 0, + "content": "Table 3. Reliability of standalone sub-chains vs sub-chains contained within the original long chain." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.334, + 0.466, + 0.478 + ], + "angle": 0, + "content": "
Sub-chainLoss TypeSRmAccmIoU
{a1,a2,a5}Long22.5258.2871.87
Short22.6858.7372.40
{a1,a3,a5}Long13.7356.7471.78
Short16.2959.9173.07
{a1,a4,a5}Long12.5056.5871.23
Short13.2557.6171.72
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.837 + ], + "angle": 0, + "content": "formulate the loss function along the whole sequence (involving \\(\\{a_1, a_2, a_3, a_4, a_5\\}\\)), and only compute the metric results over the sub-chain \\(\\{a_1, a_3, a_5\\}\\). The second case is we only model the relation of the standalone sub-chain and check its prediction result. Equivalently, we model the loss function just involving \\(\\{a_1, a_3, a_5\\}\\), and still computes the metric results of the sub-chain \\(\\{a_1, a_3, a_5\\}\\). The metric results of these two cases are presented in Table 3, where we denote 'Long' as the first case and denote 'Short' as the second case due to the length difference in loss. We spot the metric results of standalone sub-chains ('Short') are always better than the ones contained within the whole action chain ('Long'). This empirical finding also happens in the cases of \\(T = 4\\) and \\(T = 6\\). As we stated at the beginning of this section, such experimental results are foundational: The method of decoupling long chains can be only workable when standalone sub-chains can result in better metric results compared to the sub-chains contained within their original long chain. In this way, we experimentally prove the reliability of standalone sub-chains is better under any condition. It constructs the prerequisite condition to guarantee our decoupling method is workable." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.231, + 0.864 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We present comprehensive ablation experiments to illustrate the effectiveness of different components in our" + }, + { + "type": "table_caption", + "bbox": [ + 0.56, + 0.088, + 0.827, + 0.102 + ], + "angle": 0, + "content": "Table 4. Ablation experiments on CrossTask." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.102, + 0.889, + 0.521 + ], + "angle": 0, + "content": "
SRmAccmIoU
(a) With and Without State Supervision at T = 3
Language26.2858.9871.82
None28.8561.1874.98
(b) Types of layers on the Time of Input Frames at T = 3
Average Pooling24.4056.4170.87
Linear24.4056.6971.24
Conv26.5659.6074.15
MLP28.8561.1874.98
(c) Without and With Decoupling at T = 4
w/o Decoupling12.7253.1869.13
Decoupling15.5655.6470.30
(d) Without and With Decoupling at T = 5
w/o Decoupling4.9748.7165.29
Decoupling8.5552.0569.09
(e) Without and With Decoupling at T = 6
w/o Decoupling2.6244.0862.07
Decoupling5.1247.3466.11
(f) Gamma Values of Focal Loss at T = 4
γ = 013.8454.6769.74
γ = 1.015.0854.5870.09
γ = 1.515.5655.6470.30
γ = 2.014.8054.9970.20
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.55, + 0.892, + 0.581 + ], + "angle": 0, + "content": "Skip-Plan. All ablation experiments are carried out on the CrossTask dataset, illustrated in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.893, + 0.763 + ], + "angle": 0, + "content": "With and Without State Supervision. We consider \\( T = 3 \\) to compare the metric results with and without intermediate state supervision. The model with intermediate state supervision is done by adding weak language state supervision, which is identical to the one proposed by P3IV[41]. The weak language state supervision in P3IV[41] is the most effective state supervision among all the previous works. In Table 4(a), we can see the model without state supervision outperforms the model with language state supervision in all metrics. It confirms the existence of intermediate state nodes does not provide much valuable information but introduces undesired noise into the chain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Types of Layers on the Time of Input Frames. Our model achieves significant improvements by applying a small MLP network on the time dimension of visual input frames. Here, we explore the effects of various types of layers on the time information, including average pooling, a single linear layer, 1D convolution, and our proposed MLP, as shown in Table 4(b). All the previous works overlooked the significance of initial and goal visual inputs. They used the averaging pooling operation, which performs the worst" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10303" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.473, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.193, + 0.471, + 0.25 + ], + "angle": 0, + "content": "Figure 4. Case study on CrossTask: our Skip-Plan predicts all actions correctly, whereas P3IV[41] fails on the last three predictions. The outlines of correct and incorrect predictions are coloured in green and red respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.47, + 0.32 + ], + "angle": 0, + "content": "among all types of layers. We take \\( T = 3 \\) as an example to compare different layers, and indicate the MLP network in our Skip-Plan framework performs the best." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.321, + 0.469, + 0.397 + ], + "angle": 0, + "content": "Decomposition of Long Chains. We compare the performance without and with the decoupling approach at all long \\( T \\), ranging from 4 to 6. Complete ablation experiments demonstrate the decoupling approach in our Skip-Plan can robustly improve all the metric results at any condition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.469, + 0.457 + ], + "angle": 0, + "content": "Focal Loss with Gamma Values We compare the experimental results with different gamma values of focal loss. We find the gamma value of 1.5 works best to overcome the imbalanced data distribution in the CrossTask dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.469, + 0.2, + 0.485 + ], + "angle": 0, + "content": "4.5. Case Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.469, + 0.628 + ], + "angle": 0, + "content": "Skip-Plan and the Baseline. We compare our Skip-Plan with the previous state-of-the-art P3IV[41] in real cases, illustrated in Figure 4. Given the initial and goal visual observations, P3IV[41] only predicts the first action correctly and fails to get the other actions at \\( T = 4 \\). In contrast, Skip-Plan extracts the initial and goal visual information thoroughly, so Skip-Plan predicts the final action correctly. Benefiting from the decoupling method, our Skip-Plan makes the correct predictions on the second and third actions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Effects of Different Components in Skip-Plan. We investigate the effects of various components in Skip-Plan on model performance by real samples, which are illustrated in Figure 5. Here, the components for comparisons include the types of layers on the time dimension of visual input frames, state supervision, and the decoupling approach. For the types of layers on the time dimension of visual input frames, we compare the average pooling layer with our shared MLP network, both without state supervision and decoupling. As shown in Figure 5(b), the average pooling layer makes wrong predictions on all actions. However, the model with the shared MLP network predicts the first and last actions correctly, as illustrated in Figure 5(c). This is because the MLP network can extract the time information from the initial and goal visual observations more thoroughly than the average pooling layer. We further simulate the shared MLP network without decoupling but with language state supervision[41] in Figure 5(d). Comparing" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.087, + 0.895, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.255, + 0.895, + 0.394 + ], + "angle": 0, + "content": "Figure 5. Case study on Skip-Plan components at \\( T = 4 \\): (a) ground truth; (b) average pooling: average pooling layer on the time dimension of visual input frames without decoupling and state supervision; (c) MLP: shared MLP network on the time dimension of visual input frames without decoupling and state supervision; (d) MLP with state supervision: shared MLP network with state supervision but without decoupling; (e) MLP with decoupling: shared MLP network with decoupling but without state supervision. The outlines of correct and incorrect action predictions are coloured in green and red respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.894, + 0.53 + ], + "angle": 0, + "content": "Figure 5(d) with Figure 5(c), we notice the addition of state supervision cannot correct the predictions of the second and the third actions, because the state supervision does not reduce the compound error. In contrast, the shared MLP network without state supervision but with decoupling can lower the compound error effectively, resulting in correct predictions of all actions, as demonstrated in Figure 5(e)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.552, + 0.619, + 0.567 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.807 + ], + "angle": 0, + "content": "In this paper, we propose Skip-Plan, a condensed action space learning method for the procedure planning task. Our Skip-Plan solves the procedure planning problem mathematically in two novel ways. First, Skip-Plan skips the intermediate state supervision and only focuses on action relations. Second, Skip-Plan decomposes long action chains into multiple reliable sub-chains by skipping most uncertain intermediate actions. In these ways, our Skip-Plan model explores all sorts of sub-relations within an action chain and generates actions non-adjacently only in the condensed action space. Extensive experiments show the superior performance of our Skip-Plan method on two instructional video datasets. Skip-Plan has the potential to inspire the research field of procedure planning, since most of them heavily rely on state observations and adjacent actions at all times." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0114903 and in part by the National Natural Science Foundation of China under Grant 62125603 and Grant 62206153, CAAI-Huawei MindSpore Open Fund." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10304" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.433, + 0.131 + ], + "angle": 0, + "content": "[1] Mindspore. .https://www.mindspore.cn/.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.134, + 0.47, + 0.188 + ], + "angle": 0, + "content": "[2] Armenii Ashukha, Alexander Lyzhov, Dmitry Molchanov, and Dmitry Vetrov. Pitfalls of in-domain uncertainty estimation and ensembling in deep learning. arXiv preprint arXiv:2002.06470, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.19, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[3] Anil Batra, Shreyank Gowda, Laura Sevilla-Lara, and Frank Keller. A closer look at temporal ordering in the segmentation of instructional videos. arXiv preprint arXiv:2209.15501, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.249, + 0.47, + 0.304 + ], + "angle": 0, + "content": "[4] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model-based policy learning. In ICCV, pages 15611-15620, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.307, + 0.471, + 0.348 + ], + "angle": 0, + "content": "[5] Nicolas Biard, Salomé Cojean, and Eric Jamet. Effects of segmentation and pacing on procedural learning by video. Comput. Hum. Behav., 89:411-417, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.351, + 0.47, + 0.406 + ], + "angle": 0, + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.409, + 0.47, + 0.464 + ], + "angle": 0, + "content": "[7] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In ECCV, pages 334–350, 2020. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.47, + 0.509 + ], + "angle": 0, + "content": "[8] Chekuri Choudary and Tiecheng Liu. Summarization of visual content in instructional videos. IEEE Trans. Multimed., 9(7):1443-1455, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.511, + 0.47, + 0.566 + ], + "angle": 0, + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.569, + 0.469, + 0.611 + ], + "angle": 0, + "content": "[10] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.469, + 0.668 + ], + "angle": 0, + "content": "[11] Mikita Dvornik, Isma Hadji, Konstantinos G Derpanis, Animesh Garg, and Allan Jepson. Drop-dtw: Aligning common signal between sequences while dropping outliers. NeuRIPS, 34:13782-13793, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.469, + 0.741 + ], + "angle": 0, + "content": "[12] Jakob Gawlikowski, Cedrique Rovile Njieutcheu Tassi, Mohsin Ali, Jongseok Lee, Matthias Humt, Jianxiang Feng, Anna Kruspe, Rudolph Trielbel, Peter Jung, Ribana Roscher, et al. A survey of uncertainty in deep neural networks. arXiv preprint arXiv:2107.03342, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[13] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, Chiho Choi, and Behzad Dariush. Weakly-supervised online action segmentation in multi-view instructional videos. In CVPR, pages 13780-13790, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[14] Reza Ghoddoosian, Saif Sayed, and Vassilis Athitsos. Hierarchical modeling for task recognition and action segmentation in weakly-labeled instructional videos. In WACV, pages 1922-1932, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. arXiv preprint arXiv:1610.02136, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[16] Gabriel Huang, Bo Pang, Zhenhai Zhu, Clara Rivera, and Radu Soricut. Multimodal pretraining for dense video captioning. arXiv preprint arXiv:2011.11760, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[17] Eyke Hüllermeier and Willem Waegeman. Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods. Mach. Learn, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[18] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In CVPRW, pages 958-959, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[19] Lei Ji, Chenfei Wu, Daisy Zhou, Kun Yan, Edward Cui, Xilin Chen, and Nan Duan. Learning temporal video procedure segmentation from an automatically collected large dataset. In WACV, pages 1506-1515, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[20] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? NeurIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[21] Colin Lea, Michael D Flynn, Rene Vidal, Austin Reiter, and Gregory D Hager. Temporal convolutional networks for action segmentation and detection. In CVPR, pages 156-165, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.419 + ], + "angle": 0, + "content": "[22] Muheng Li, Lei Chen, Yueqi Duan, Zhilan Hu, Jianjiang Feng, Jie Zhou, and Jiwen Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In CVPR, pages 19880-19889, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.421, + 0.892, + 0.462 + ], + "angle": 0, + "content": "[23] Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. arXiv preprint arXiv:1706.02690, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.463, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.506, + 0.892, + 0.561 + ], + "angle": 0, + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.563, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[26] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.661 + ], + "angle": 0, + "content": "[27] Andrey Malinin and Mark Gales. Predictive uncertainty estimation via prior networks. NeurIPS, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.662, + 0.892, + 0.717 + ], + "angle": 0, + "content": "[28] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In CVPR, pages 9879-9889, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.719, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[29] Medhini Narasimhan, Arsha Nagrani, Chen Sun, Michael Rubinstein, Trevor Darrell, Anna Rohrbach, and Cordelia Schmid. Tl; dw? summarizing instructional videos with task relevance and cross-modal saliency. In ECCV, pages 540-557, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[30] Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. NeurIPS, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[31] AJ Piergiovanni, Anelia Angelova, Michael S Ryoo, and Irfan Essa. Unsupervised action segmentation for instructional videos. arXiv preprint arXiv:2106.03738, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10305" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[32] Alireza Shafaei, Mark Schmidt, and James J Little. A less biased evaluation of out-of-distribution sample detectors. arXiv preprint arXiv:1809.04729, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[33] Anshul Shah, Benjamin Lundell, Harpreet Sawhney, and Rama Chellappa. Steps: Self-supervised key step extraction from unlabeled procedural videos. arXiv preprint arXiv:2301.00794, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[34] Botian Shi, Lei Ji, Yaobo Liang, Nan Duan, Peng Chen, Zhendong Niu, and Ming Zhou. Dense procedure captioning in narrated instructional videos. In ACL, pages 6382-6391, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[35] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In ICCV, pages 7464-7473, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[36] Jiankai Sun, De-An Huang, Bo Lu, Yun-Hui Liu, Bolei Zhou, and Animesh Garg. Plate: Visually-grounded planning with transformers in procedural tasks. IEEE Robot. Autom. Lett., 7(2):4924-4930, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.362, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[37] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In CVPR, pages 1207-1216, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.419, + 0.469, + 0.46 + ], + "angle": 0, + "content": "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.462, + 0.469, + 0.529 + ], + "angle": 0, + "content": "[39] Mengmeng Xu, Juan-Manuel Pérez-Rúa, Victor Escorcia, Brais Martinez, Xiatian Zhu, Li Zhang, Bernard Ghanem, and Tao Xiang. Boundary-sensitive pre-training for temporal localization in videos. In ICCV, pages 7220-7230, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.532, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[40] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Temporal query networks for fine-grained video understanding. In CVPR, pages 4486-4496, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.575, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[41] He Zhao, Isma Hadji, Nikita Dvornik, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In CVPR, pages 2938-2948, 2022. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.646, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[42] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, pages 7590-7598, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.688, + 0.469, + 0.728 + ], + "angle": 0, + "content": "[43] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In CVPR, pages 8746-8755, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.731, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[44] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In CVPR, pages 3537-3545, 2019. 1, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.786 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10306" + } + ] +] \ No newline at end of file diff --git a/2023/Skip-Plan_ Procedure Planning in Instructional Videos via Condensed Action Space Learning/c1a3a0ec-3885-4f2f-93dd-62793c50f314_origin.pdf b/2023/Skip-Plan_ Procedure Planning in Instructional Videos via Condensed Action Space Learning/c1a3a0ec-3885-4f2f-93dd-62793c50f314_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..15a4564ac39468590189724681b20e620b1ffa80 --- /dev/null +++ b/2023/Skip-Plan_ Procedure Planning in Instructional Videos via Condensed Action Space Learning/c1a3a0ec-3885-4f2f-93dd-62793c50f314_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab962fa755143d4c9cdaa9c41c077d1c6d7af49d4afcf94929feafe42be957c8 +size 1197553 diff --git a/2023/Skip-Plan_ Procedure Planning in Instructional Videos via Condensed Action Space Learning/full.md b/2023/Skip-Plan_ Procedure Planning in Instructional Videos via Condensed Action Space Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..75a2cdeb4bf71949c8529572f1ccc251bb77d58c --- /dev/null +++ b/2023/Skip-Plan_ Procedure Planning in Instructional Videos via Condensed Action Space Learning/full.md @@ -0,0 +1,272 @@ +# Skip-Plan: Procedure Planning in Instructional Videos via Condensed Action Space Learning + +Zhiheng Li $^{1}$ , Wenjia Geng $^{2}$ , Muheng Li $^{1}$ , Lei Chen $^{3}$ , Yansong Tang $^{2*}$ , Jiwen Lu $^{1,4}$ , Jie Zhou $^{1,4}$ + +$^{1}$ Department of Automation, Tsinghua University + +2 Shenzhen International Graduate School, Tsinghua University + +3 Beijing University of Science and Technology + +4 Beijing National Research Center for Information Science and Technology + +{lizhihan21@gengwj22@,li-mh20}@mails.tsinghua.edu.cn,chenlei2022@ustb.edu.cn, {tang.yansong@sz.,lujiwen,jzhou@}tsinghua.edu.cn + +# Abstract + +In this paper, we propose Skip-Plan, a condensed action space learning method for procedure planning in instructional videos. Current procedure planning methods all stick to the state-action pair prediction at every timestep and generate actions adjacently. Although it coincides with human intuition, such a methodology consistently struggles with high-dimensional state supervision and error accumulation on action sequences. In this work, we abstract the procedure planning problem as a mathematical chain model. By skipping uncertain nodes and edges in action chains, we transfer long and complex sequence functions into short but reliable ones in two ways. First, we skip all the intermediate state supervision and only focus on action predictions. Second, we decompose relatively long chains into multiple short sub-chains by skipping unreliable intermediate actions. By this means, our model explores all sorts of reliable sub-relations within an action sequence in the condensed action space. Extensive experiments show Skip-Plan achieves state-of-the-art performance on the CrossTask and COIN benchmarks for procedure planning. + +# 1. Introduction + +Given the initial and goal visual observations which are curated from instructional videos [44, 37], the task of procedure planning is to predict a sequence of actions to reach the indicated goal [7], as illustrated in Figure 1(a). This is an emerging research direction in the field of video understanding. Procedure planning has worldwide applications and assists humans with instruction-following tasks, such as cooking, assembling furniture, electronics repair, etc. + +Question: What action sequence are needed to get from $v_{start}$ to $v_{goal}$ ? + +![](images/b69948dc1487f9e8bff0863ec5ecd0094f4913a48597e658da62001680dd1a32.jpg) +(a) Problem Definition + +Answer: Required action sequence + +(Time Horizon $T = 4$ + +![](images/29ec5d8dee3b24a6f4f646646fd8eaac91ab1a59a2ab7d331b5bd4d3130910d8.jpg) + +![](images/248d852484ca6d081083b77706bb1cec3e6b9db73d2358a2a53429c00375c233.jpg) +(b) Conventional Procedure Planning Paradigm + +![](images/57cd4e7a4e8bf0c1e11c4180e00653ee59f406f5fad470388633dd5209df88c8.jpg) +(c)Skip-Plan + +![](images/0d9a2dd55e040bdd5893ea75417e20dc29eff784da94beba768ffbb0d7aad84f.jpg) +Figure 1. Comparisons of the conventional paradigm and our proposed Skip-Plan in procedure planning. (a) Problem definition: given the initial and goal visual observations $(v_{start}$ and $v_{goal})$ , procedure planning is to answer an action sequence to reach the indicated goal with a pre-defined time horizon $T$ . (b) Conventional paradigm predicts a state-action pair $\{\tilde{s}_t,\tilde{a}_t\}$ at every timestep $t$ , and repeat it step-by-step. (c) Our Skip-Plan skips state supervision and extracts multiple sub-chains by skipping most intermediate actions. The final action sequence is accumulated from all predicted sub-chains. Skip-Plan significantly reduces the overall compound error, indicated by the shallower darkness of the chevron arrows. + +In the past few years, various approaches have been proposed to solve the procedure planning problem. Most existing methods utilize an autoregressive model with visual state supervision to generate actions one by one [7, 4, 36]. Another type of work [41] solves it in a non-autoregressive way with language state supervision. The difficulties of procedure planning have been widely discussed in the previous works. For instance, the sequential orders in procedure planning can be altered but still sensible (e.g., switching the order of 'pour water' and 'pour alcohol' does not matter in many cases), and the visual states are unstructured and high-dimensional. Thus, we can conclude the following limitations from the existing works. (a) State supervision: the major difficulty in procedure planning is state supervision. Visual images are high-dimensional and complex, with extra misleading information such as background, illumination, viewing point, etc. (b) Step-by-step predictions: the conventional paradigm in procedure planning predicts one state-action pair at every timestep, and repeats it step-by-step, as illustrated in Figure 1(b). Even though the proposed non-autoregressive model generates whole action sequences in batches [41], it can be still regarded as a step-by-step method because the model neglects the links between non-adjacent actions. From Figure 1(b), we can see while predicting an action sequence with a time horizon of 4, the conventional model is equivalently predicting a sequence with a length of 7 (4 action nodes, 3 state nodes). More frequent predictions tend to accumulate a larger compound error along the chain, which lowers the prediction accuracy. + +To overcome these limitations, we propose Skip-Plan, a condensed action space learning method that solves the procedure planning problem through the lens of mathematics, aiming to reduce compound error along action chains. As we have discussed above, existing methods struggle with high-dimensional and complex state supervision. Thus, the compound error along the action chain is increased due to large errors in state predictions. This motivates us to skip all the state nodes in our Skip-Plan framework and only model sequence relations in the condensed action space, as demonstrated in Figure 1(c). Without the presence of state supervision, a lower compound error results in better performance in Skip-Plan. Although the dependencies between high-level actions without state supervision are very abstract, the algorithm of our Skip-Plan is highly mathematically sensible in the Markov chain model. + +Another convention in procedure planning is we should learn sequential relations between adjacent actions. The Markov chain model inspires us to decouple relatively long chains into multiple short sub-chains, because the compound error can be reduced with shorter chain lengths. The problem here is how to decouple these long chains, because there are various choices of sub-chains with different lengths and we should always choose the most reliable + +ones under any condition. Given the initial and goal frames as concrete inputs, it is apparent the reliabilities of the initial and final actions are the highest, and the intermediate actions are more likely to be predicted wrong. Therefore, our Skip-Plan decomposes long chains by 'skipping' some unreliable intermediate actions, and each sub-chain in our model is composed of the initial action, one unique intermediate action, and the final action. One example for a chain length of 4 is shown in Figure 1(c). In this way, the reliable sub-relations between non-adjacent actions can be well explored by the network. To model the discriminative relations of all sub-chains, we propose sub-chain decoders, where each decoder only learns one sub-chain by the specific loss design. The final action sequence is aggregated from all sub-chain predictions by a sub-chain accumulator. + +In summary, our contributions are threefold. (a) Unlike the previous methods which put sophisticated designs on state supervision, Skip-Plan skips unreliable state predictions to boost the model performance in procedure planning. (b) In contrast to the existing methods which only focus on the links between adjacent actions, we decompose long chains into multiple reliable sub-chains by skipping some intermediate actions, and propose sub-chain decoders to learn their discriminative relations. (c) Extensive experiments show Skip-Plan achieves state-of-the-art performance on the CrossTask [44] and COIN [37] benchmarks for procedure planning. + +# 2. Related Work + +Procedure Planning. Procedure planning in instructional videos was first introduced by Chang et al. [7]. Recently, various methods have been conducted to solve this task. Chang et al. [7] propose a dual dynamic network to conjugate the relationships between states and actions. Sun et al. [36] utilize a GPT network to generate actions and states one by one. Bi et al. [4] use the model-based reinforcement learning with time-invariant context information to predict state-action pairs autoregressively. Different from the previous autoregressive methods with visual state supervision, Zhao et al. [41] use a non-autoregressive transformer decoder to generate action sequences with language state supervision. Unlike previous works, Skip-Plan generates actions non-adjacently without any state supervision. + +Instructional Videos Analysis. Instructional videos are used for learning human behaviours because they are visual representations of human actions. Based on instructional videos, various action-related tasks have been proposed, such as action segmentation[21, 22, 13, 14, 28, 31], step localization[39, 11, 43, 33, 26], procedure segmentation[42, 19, 3, 5], video summarization/captioning[8, 35, 29, 34, 16, 18], etc. Different from these action tasks in instructional videos, we investigate procedure planning, a goal-conditioned problem that plans a sequence of actions given + +the initial and goal visual observations. + +Predictive Uncertainty on Sequence Relations. All types of questions that we are interested in can be modeled as a sequence model, such as images [25] and languages [9]. Currently, Transformer [38] is the most effective network to model sequence relations, and our model is based on the transformer decoder with a learnable query input[40, 6]. However, sequence relations are always distorted by predictive uncertainty[12, 27, 17, 30, 20], and the sources of predictive uncertainty can be various, including model uncertainty, data uncertainty, etc[2, 10, 15, 23, 32]. In this work, we pay particular attention to data uncertainty, in which the uncertainty directly stems from the data. Data uncertainty is caused by information loss when we measure a data distribution in a space with incorrect dimensions (e.g., we must lose depth information if we measure a three-dimensional object in a two-dimensional space). Therefore, we often avoid high-dimensional data for lower data uncertainty. + +# 3. Methods + +# 3.1. Problem Formulation + +We follow the definition of procedure planning proposed by Chang et al. [7]. Here, we denote the action and the state at timestep $t$ as $a_{t}$ and $s_{t}$ respectively. Given the initial visual observation $v_{start}$ and the goal visual observation $v_{goal}$ , the model generates a sequence of actions $a_{1:T}$ to reach the indicated goal. In this work, we introduce a novel formulation for the procedure planning task: + +$$ +p \left(a _ {1: T} \mid v _ {s}, v _ {g}\right) = \prod_ {t = 2} ^ {T - 1} p \left(a _ {t} \mid a _ {1}, a _ {T}\right) p \left(a _ {1}, a _ {T} \mid v _ {s}, v _ {g}\right), \tag {1} +$$ + +where $v_{s}$ and $v_{g}$ short for ' $v_{start}$ ' and ' $v_{goal}$ ' . We denote $p(a_{1:T}|v_{start},v_{goal})$ as the conditional probability distribution of the action sequence given the initial visual observation $v_{start}$ and the goal visual observation $v_{goal}$ , $p(a_t|a_1,a_T)$ as the conditional probability distribution of intermediate actions $(2\leq t\leq T - 1)$ given the initial action $a_1$ and the final action $a_{T}$ , and $p(a_{1},a_{T}|v_{start},v_{goal})$ as the conditional probability distribution of the initial action $a_1$ and the final action $a_{T}$ given the initial visual observation $v_{start}$ and the goal visual observation $v_{goal}$ . Equation 1 reveals the uniqueness of our Skip-Plan method: First, we do not rely on any intermediate state supervision $s_t$ for action generations. Second, every intermediate action prediction $a_{t}$ $(2\leq t\leq T - 1)$ only depends on the initial action $a_1$ and the final action $a_{T}$ . Thus, Skip-Plan does not focus on sequential relations between adjacent actions but predicts action sequences non-adjacently without state supervision. + +# 3.2. Action Predictions without State Supervision + +Information does not always help in the decision-making process. Sometimes ambiguous information can even lead + +to sub-optimal results, and plausible plans can be only made based on reliable information. In procedure planning, two types of supervision can be utilized, including actions $a_{t}$ and visual states $s_{t}$ . Even though they can be directly generated from each other, the intrinsic properties of these two supervisions are largely different. We evaluate them from the perspectives of the value of information and predictive uncertainty. Visual state supervision is well known to be high-dimensional. In addition to the factor of action class, various factors such as background, viewing point, illumination, etc. can be involved in visual states. Thus, useful information in visual state supervision for action predictions is overwhelmed by many irrelevant factors. In contrast, action labels are the most condensed supervision, without being interrupted by any factors. From this point, we argue the value of information offered by state supervision is extremely low because the portion of information, which is relevant to action predictions, is too small. On the other hand, when models generate high-dimensional predictions such as visual images, large predictive errors are more likely to occur. Thus, the predictive uncertainty of intermediate visual states is significantly higher than the one of condensed action predictions. Combining the effects of the value of information and the predictive uncertainty, we argue the existence of visual state supervision hurts the predictive accuracy of sequential actions. Even though we can use the weak language state supervision to reduce some irrelevant dimensions, it is still too high-dimensional compared to condensed action labels. Therefore, we 'skip' intermediate state supervision in our Skip-Plan framework. The overall compound error along an action chain is significantly reduced without the predictive errors introduced by states. + +Getting rid of intermediate state supervision does not mean all sorts of visual information are useless. In procedure planning, the initial visual observation $v_{start}$ and the goal visual observation $v_{goal}$ are vital because action sequences are constrained by them. These visual observations are in the form of frame features in seconds. Inspired by the image analysis above, we argue the visual input information is underutilized by the previous works, especially along the time dimension. In the Skip-Plan framework, we take a shared multi-layer perceptrons (MLP) network in the Visual Input Module to learn the time information, where the MLP network is three-layered with ReLU activations. This simple method can boost the model performance by a large margin, especially for short sequences $(T = 3)$ . + +# 3.3. Decomposition of Long Chains + +Long chains are not favored when a compound error is accumulating. Thus, decomposing long chains into short ones is a natural way to better model performance. Through the lens of mathematics, the only design philosophy here is that we should extract most reliable sub-chains from the + +![](images/626640d6c67c78dd6e9559a4d27b6c9010289993c200761769d0796d28562694.jpg) +Figure 2. Our Skip-Plan architecture. The shared MLP network in the Visual Input Module extracts the initial and final information from the initial and goal observations ( $v_{start}$ and $v_{goal}$ ) and attaches them to the first and last dimensions of every learnable query input. A number of $T - 2$ query inputs are separately passed to each decoder in the Sub-chain Decoder Module. In each decoder, the input is processed by the self-attention block and then by the cross-attention block, where the cross-attention block is augmented by a learnable memory $M_i$ . The sub-chain decoders are separated. The $n$ th decoder outputs the sub-chain $\{ \tilde{a}_1, \tilde{a}_t, \tilde{a}_T \}$ with $t = n + 1$ . The predicted sub-chains are supervised by two loss supervision. In Individual Sub-chain Supervision, the predicted sub-chains are supervised by the corresponding ground-truth sub-chains. In Complete Chain Supervision, the predicted sub-chains are aggregated by the Sub-chain Accumulator to generate the complete action chain, which is supervised by the complete ground-truth action chain. + +original long chain. Following this principle, we ought to find there is so much flexibility in the decoupling method: We can choose sub-chains with any lengths, and actions in the sub-chains are not necessary to be adjacent. + +The reliability of a sub-chain can be determined by the uncertainty of predictions. According to the uncertainty analysis, the reliabilities of the initial and final actions are the highest because they are the closest actions to the initial and goal visual inputs. Since the compound error is accumulated from two ends of the chain, the reliability of action predictions is decreasing while approaching the middle of the chain. Consequently, we remove all the edges between adjacent intermediate actions, and each individual intermediate action only connects to the reliable initial and final actions. This design just coincides with the term $p(a_{t}|a_{1},a_{T})$ with $2\leq t\leq T - 1$ in Equation 1. In this way, we generate a quantity of $T - 2$ sub-chains from an action chain with a time horizon of $T$ . The set of sub-chains is given by: + +$$ +\{(a _ {1}, a _ {2}, a _ {T}), (a _ {1}, a _ {3}, a _ {T}), \dots , (a _ {1}, a _ {T - 1}, a _ {T}) \}. \quad (2) +$$ + +Each sub-chain has a length of 3, with the order of the initial action, one unique intermediate action, and the final action. This decoupling method works well for relatively long chains $(T > 3)$ . + +Given multiple short sub-chains, we design a compatible network to model all their sub-relations. Since these sub-chains are parallel and independent to each other, we propose the Sub-chain Decoder Module, where each transformer decoder only models one discriminative relation of one sub-chain. Every transformer decoder is non-autoregressive with two types of inputs: a learnable query + +with the initial and goal visual features, and a learnable memory component. The novelty here is each decoder only learns one sub-relation, with a unique loss only corresponding to the three actions of that sub-chain. The query input, parameters, and memory component of every decoder are separated from other decoders'. Therefore, the number of decoders in the Sub-chain Decoder Module is equal to the number of sub-chains $(T - 2)$ extracted from the original chain. The final sequence generation is aggregated from all the sub-chain predictions through the Sub-chain Accumulator, which is an MLP network. In this way, our Skip-Plan network can model all sorts of discriminative sub-relations within the original action sequence. + +# 3.4. Training + +Loss is an essential component to discriminate sub-chain decoders. Instead of conventional cross-entropy loss, we use the focal loss to supervise our sub-chain decoders. The focal loss $FL$ is formulated as[24]: + +$$ +F L = - \sum_ {t = 1} ^ {T} a _ {t} (1 - \tilde {a} _ {t}) ^ {\gamma} \log \tilde {a} _ {t}, \tag {3} +$$ + +where we denote $a_{t}$ as the ground-truth one-hot labels, $\tilde{a}_{t}$ as the predicted probability distribution of the action sequence, and $\gamma$ as the gamma value of focal loss. The focal loss is equivalent to the cross-entropy loss when $\gamma = 0$ . The model training is benefited from the use of focal loss because the data distribution in procedure planning is imbalanced. In this work, we define $FL(a_{1},a_{t},a_{T})$ as the focal loss for the sub-chain with a length of 3, where the sub-chain is composed of the initial action $a_{1}$ , the action at timestep $t$ $a_{t}$ , and + +the final action $a_{T}$ . We present the focal loss of the whole action sequence as $FL(a_{1:T})$ . + +Loss of Individual Sub-chains. As we have discussed before, the loss of each decoder only corresponds to the actions within that sub-chain. Therefore, the individual loss of the $n$ th decoder is defined as $FL(a_{1},a_{t},a_{T})$ with $t = n + 1$ , and the loss from all sub-chain decoders can be written as: + +$$ +\mathcal {L} _ {N} = \sum_ {t = 2} ^ {T - 1} F L \left(a _ {1}, a _ {t}, a _ {T}\right). \tag {4} +$$ + +Loss of Complete Chain. After aggregating all sub-chains by the Sub-chain Accumulator, the sequence output with a length of $T$ is supervised by the focal loss of the whole action sequence: + +$$ +\mathcal {L} _ {T} = F L \left(a _ {1: T}\right). \tag {5} +$$ + +Overall Loss. Therefore, the total loss is defined as the sum of the loss from all sub-chains and the complete chain loss, which can be written as: + +$$ +\mathcal {L} = \mathcal {L} _ {N} + \mathcal {L} _ {T}. \tag {6} +$$ + +In this way, our model can explore all sorts of discriminative but reliable sub-relations within an action sequence, while guaranteeing the final sequence generation is optimal. + +# 3.5. Implementation Details + +For each decoder, the query input length is $T + 1$ , and the memory size is 128. The numbers of decoder layers, decoder heads, and decoder dimensions are set to be 1, 16, and 1024 respectively. All MLP networks are three-layered with ReLU activations. In the Visual Input Module, our shared MLP network on the time dimension of input frame features has the shape $[3 \to 6 \to 1]$ . All action classifiers in the Sub-chain Decoder Module are the same MLP networks with the shape $[1024 \to 512 \to n_a]$ , where $n_a$ is the number of action classes. The Sub-chain Accumulator is an MLP network with the shape $[T(T - 2) \to 3T(T - 2) \to T]$ . Models are optimized for 500 epochs with stochastic gradient descent. The initial learning rate is 0.02 and decays by 10 at every 50 epochs after 100 epochs. Experiments are conducted on a single GeForce RTX 3090 GPU. We also test our model using the Mindspace [1]. + +# 4. Experiments + +# 4.1. Evaluation Protocols + +Datasets. We evaluate our model on two real-world instructional video datasets: CrossTask[44] and COIN[37]. CrossTask contains 2763 videos with a duration of 213 hours. It consists of 18 different tasks with 7.4 steps per task. The COIN dataset spans 12 different domains related + +to our daily life. It contains 11,827 videos with 3.9 steps per video. The COIN dataset is challenging due to the diversity of its tasks and action labels (180 tasks and 779 action labels). We curate a set of plans by the sliding window approach, and adopt the standard $70\% / 30\%$ to create our train/test splits. The division is done on the generated samples rather than the original video. The data pre-processing procedure in our work is identical to the most widely used protocol in the previous procedure planning works (e.g., DDN [7], Ext-GAIL [4], and the main protocol in P3IV [41]) for fair comparisons. + +Metrics. Following the previous works, we used three metrics to comprehensively evaluate model performance. For each action sequence: (a) mean Intersection over Union (mIoU) regards the predicted and ground-truth action sequences as two sets, and measure the overlap between these sets. The mIoU metric only indicates whether desired actions exist in the predicted action set, agnostic to the order of actions. (b) mean Accuracy (mAcc) evaluates the correctness of the predicted action at every timestep, termed as percentages. Thus, mAcc counts the order of actions, and is more strict than mIoU. (c) Success Rate (SR) evaluates the correctness of the whole action sequence. A predicted sequence is successful only if it matches the ground-truth sequence. The success rate is the proportion of successful samples over all samples, which is more strict than mAcc. + +Training Scheme on Each Dataset. We specify an appropriate gamma value on each dataset to maximize model performance. Through experiments, we find the gamma values of 1.5 and 0 are the best parameters for the CrossTask and COIN datasets respectively. This is reasonable because the data distribution in the COIN dataset is more balanced due to a large quantity of action labels in diverse tasks. + +# 4.2. Evaluating Procedure Planning + +Short-horizon Prediction. We compare the model performance of the short-horizon predictions $(T = 3 / 4)$ in Table 1. As illustrated in Table 1, Skip-Plan significantly outperforms the previous state-of-the-art model at all metrics for $T = 3$ . In the CrossTask dataset, our SR and mAcc results surpass the results of the second best approach (e.g., P3IV[41]) by $6\%$ and $11\%$ . Our method also outperforms P3IV[41] by $8\%$ on SR and $25\%$ on mAcc in the COIN dataset. Such incredible improvements strongly prove Skip-Plan without state supervision suppresses predictive uncertainty effectively, and our module design on the visual input extracts initial and goal visual information thoroughly. For $T = 4$ , Skip-Plan achieves large gains of around $5\%$ , $24\%$ , and $7\%$ in SR, mAcc and mIoU respectively in the COIN dataset. For the CrossTask dataset, the mAcc and mIoU of our Skip-Plan model are the best and the SR is the second best. Particularly, our mAcc is higher than P3IV's[41] by $11.48\%$ . The methods of no state supervision and decou + +Table 1. Metric results on CrossTask and COIN for short-horizon predictions $\left( {T \in \{ 3,4\} }\right)$ . The ’ $\checkmark$ ’ and ’ $X$ ’ in the column State Supervision denote with and without state supervision respectively. The best and second best metric results are coloured in red and blue respectively. + +
HorizonsModelsState SupervisionCrossTaskCOIN
SRmAccmIoUSRmAccmIoU
T=3DDN[7]✓12.1831.2947.4813.9020.1964.78
Plate[36]✓16.0036.1765.91---
Ext-GAIL[4]✓21.2749.4661.70---
P3IV[41]✓23.3449.9673.8915.4021.6776.31
Skip-Plan✗28.8561.1874.9823.6547.1278.44
T=4DDN[7]✓5.9727.1048.4611.1317.7168.06
Plate[36]✓14.0035.2955.36---
Ext-GAIL[4]✓16.4143.0560.93---
P3IV[41]✓13.4044.1670.0111.3218.8570.53
Skip-Plan✗15.5655.6470.3016.0443.1977.07
+ +Table 2. Metric results for long-horizon predictions $(T \in \{5,6\})$ on CrossTask and COIN. + +
DatasetsHorizonsModelsSRmAccmIoU
CrossTaskT=5DDN[7]3.10--
P3IV[41]7.21--
Skip-Plan8.5552.0569.09
T=6DDN[7]1.20--
P3IV[41]4.40--
Skip-Plan5.1247.3466.11
COINT=5P3IV[41]4.2710.8168.81
Skip-Plan9.9038.9976.93
T=6P3IV[41]4.178.9767.72
Skip-Plan6.6133.6373.68
+ +pling long chains both drive Skip-Plan to succeed at $T = 4$ + +Long-horizon Prediction. We further evaluate Skip-Plan on long-horizon predictions ( $T = 5/6$ ). In the COIN dataset, our Skip-Plan performs extraordinarily well on the metric of mAcc, with remarkable improvements of $28\%$ and $25\%$ for $T = 5/6$ , compared to the second best model P3IV[41]. As for the metrics of SR and mIoU, Skip-Plan exceeds P3IV[41] by $6\%$ and $8\%$ at $T = 5$ , and $2\%$ and $6\%$ at $T = 6$ respectively. The outstanding performance we presented above further validates the success of the decoupling approach. Our Skip-Plan also achieved the best results on SR in the CrossTask dataset, surpassing P3IV[41] by $1.34\%$ and $0.72\%$ at $T = 5/6$ . For the long-horizon predictions, we argue mAcc and mIoU are more reliable metrics to measure the effectiveness of a model. SR is interrupted by a large variance, because data samples become insufficient with a large time horizon. + +# 4.3. Prerequisite Experiments + +In this section, we carried out several prerequisite experiments for the long-chain decoupling method. They can validate some statements that we have made in Section 3.3. These experiments are foundational because they proved the prerequisite conditions for the decoupling approach to work. Thus, we name them as the prerequisite experiments. + +Error Rate Distribution along a Predicted Chain. We tend to remove the edges between unreliable nodes and only connect one unreliable node to other reliable ones. In this way, we can filter out most reliable sub-chains from the original long chain. Here, we measure the reliability of each node in terms of the error rate distribution along the non-autoregressively predicted chain. The error rate distributions along the predicted chains with $T = 3/4/5/6$ are illustrated in Figure 3. It is clear the reliabilities of the first and final action nodes are the highest and the model is more likely to make wrong predictions while approaching the middle of the chain. It is not hard to understand this phenomenon: a higher error rate occurs at intermediate action predictions because they are further away from the initial and goal visual observations. This empirical finding motivates us to remove the edges between unreliable intermediate actions and only connect each intermediate action to the reliable initial and final action nodes. + +Reliability of Standalone Sub-chains. In addition to the reliability of each node, we further study the reliability of the sub-chains that we have decoupled from the original long chain. For a clearer explanation, we take a chain with a length of 5 as an example (e.g., $\{a_1, a_2, a_3, a_4, a_5\}$ ), and only study the reliability of the sub-chain $\{a_1, a_3, a_5\}$ . We consider two cases: First, we just model the relation of the whole chain and only investigate the prediction result of that sub-chain contained in the long chain. In other words, we + +![](images/b313374b0c51add4dea043815653a9f1f91625f81a3c6bea6fcb399bfacbd008.jpg) +Figure 3. Error rate distributions along the non-autoregressively predicted chains with $T = 3/4/5/6$ . + +Table 3. Reliability of standalone sub-chains vs sub-chains contained within the original long chain. + +
Sub-chainLoss TypeSRmAccmIoU
{a1,a2,a5}Long22.5258.2871.87
Short22.6858.7372.40
{a1,a3,a5}Long13.7356.7471.78
Short16.2959.9173.07
{a1,a4,a5}Long12.5056.5871.23
Short13.2557.6171.72
+ +formulate the loss function along the whole sequence (involving $\{a_1, a_2, a_3, a_4, a_5\}$ ), and only compute the metric results over the sub-chain $\{a_1, a_3, a_5\}$ . The second case is we only model the relation of the standalone sub-chain and check its prediction result. Equivalently, we model the loss function just involving $\{a_1, a_3, a_5\}$ , and still computes the metric results of the sub-chain $\{a_1, a_3, a_5\}$ . The metric results of these two cases are presented in Table 3, where we denote 'Long' as the first case and denote 'Short' as the second case due to the length difference in loss. We spot the metric results of standalone sub-chains ('Short') are always better than the ones contained within the whole action chain ('Long'). This empirical finding also happens in the cases of $T = 4$ and $T = 6$ . As we stated at the beginning of this section, such experimental results are foundational: The method of decoupling long chains can be only workable when standalone sub-chains can result in better metric results compared to the sub-chains contained within their original long chain. In this way, we experimentally prove the reliability of standalone sub-chains is better under any condition. It constructs the prerequisite condition to guarantee our decoupling method is workable. + +# 4.4. Ablation Study + +We present comprehensive ablation experiments to illustrate the effectiveness of different components in our + +Table 4. Ablation experiments on CrossTask. + +
SRmAccmIoU
(a) With and Without State Supervision at T = 3
Language26.2858.9871.82
None28.8561.1874.98
(b) Types of layers on the Time of Input Frames at T = 3
Average Pooling24.4056.4170.87
Linear24.4056.6971.24
Conv26.5659.6074.15
MLP28.8561.1874.98
(c) Without and With Decoupling at T = 4
w/o Decoupling12.7253.1869.13
Decoupling15.5655.6470.30
(d) Without and With Decoupling at T = 5
w/o Decoupling4.9748.7165.29
Decoupling8.5552.0569.09
(e) Without and With Decoupling at T = 6
w/o Decoupling2.6244.0862.07
Decoupling5.1247.3466.11
(f) Gamma Values of Focal Loss at T = 4
γ = 013.8454.6769.74
γ = 1.015.0854.5870.09
γ = 1.515.5655.6470.30
γ = 2.014.8054.9970.20
+ +Skip-Plan. All ablation experiments are carried out on the CrossTask dataset, illustrated in Table 4. + +With and Without State Supervision. We consider $T = 3$ to compare the metric results with and without intermediate state supervision. The model with intermediate state supervision is done by adding weak language state supervision, which is identical to the one proposed by P3IV[41]. The weak language state supervision in P3IV[41] is the most effective state supervision among all the previous works. In Table 4(a), we can see the model without state supervision outperforms the model with language state supervision in all metrics. It confirms the existence of intermediate state nodes does not provide much valuable information but introduces undesired noise into the chain. + +Types of Layers on the Time of Input Frames. Our model achieves significant improvements by applying a small MLP network on the time dimension of visual input frames. Here, we explore the effects of various types of layers on the time information, including average pooling, a single linear layer, 1D convolution, and our proposed MLP, as shown in Table 4(b). All the previous works overlooked the significance of initial and goal visual inputs. They used the averaging pooling operation, which performs the worst + +![](images/033532c38ca7be1c586ef29060e8be387013f0bba3c245f68fdda090beb1e2e1.jpg) +Figure 4. Case study on CrossTask: our Skip-Plan predicts all actions correctly, whereas P3IV[41] fails on the last three predictions. The outlines of correct and incorrect predictions are coloured in green and red respectively. + +among all types of layers. We take $T = 3$ as an example to compare different layers, and indicate the MLP network in our Skip-Plan framework performs the best. + +Decomposition of Long Chains. We compare the performance without and with the decoupling approach at all long $T$ , ranging from 4 to 6. Complete ablation experiments demonstrate the decoupling approach in our Skip-Plan can robustly improve all the metric results at any condition. + +Focal Loss with Gamma Values We compare the experimental results with different gamma values of focal loss. We find the gamma value of 1.5 works best to overcome the imbalanced data distribution in the CrossTask dataset. + +# 4.5. Case Study + +Skip-Plan and the Baseline. We compare our Skip-Plan with the previous state-of-the-art P3IV[41] in real cases, illustrated in Figure 4. Given the initial and goal visual observations, P3IV[41] only predicts the first action correctly and fails to get the other actions at $T = 4$ . In contrast, Skip-Plan extracts the initial and goal visual information thoroughly, so Skip-Plan predicts the final action correctly. Benefiting from the decoupling method, our Skip-Plan makes the correct predictions on the second and third actions. + +Effects of Different Components in Skip-Plan. We investigate the effects of various components in Skip-Plan on model performance by real samples, which are illustrated in Figure 5. Here, the components for comparisons include the types of layers on the time dimension of visual input frames, state supervision, and the decoupling approach. For the types of layers on the time dimension of visual input frames, we compare the average pooling layer with our shared MLP network, both without state supervision and decoupling. As shown in Figure 5(b), the average pooling layer makes wrong predictions on all actions. However, the model with the shared MLP network predicts the first and last actions correctly, as illustrated in Figure 5(c). This is because the MLP network can extract the time information from the initial and goal visual observations more thoroughly than the average pooling layer. We further simulate the shared MLP network without decoupling but with language state supervision[41] in Figure 5(d). Comparing + +![](images/87ab1d8252ae3fbd6c85244c8de2f3b09d05c88db3679ec5a57709b44f822e21.jpg) +Figure 5. Case study on Skip-Plan components at $T = 4$ : (a) ground truth; (b) average pooling: average pooling layer on the time dimension of visual input frames without decoupling and state supervision; (c) MLP: shared MLP network on the time dimension of visual input frames without decoupling and state supervision; (d) MLP with state supervision: shared MLP network with state supervision but without decoupling; (e) MLP with decoupling: shared MLP network with decoupling but without state supervision. The outlines of correct and incorrect action predictions are coloured in green and red respectively. + +Figure 5(d) with Figure 5(c), we notice the addition of state supervision cannot correct the predictions of the second and the third actions, because the state supervision does not reduce the compound error. In contrast, the shared MLP network without state supervision but with decoupling can lower the compound error effectively, resulting in correct predictions of all actions, as demonstrated in Figure 5(e). + +# 5. Conclusion + +In this paper, we propose Skip-Plan, a condensed action space learning method for the procedure planning task. Our Skip-Plan solves the procedure planning problem mathematically in two novel ways. First, Skip-Plan skips the intermediate state supervision and only focuses on action relations. Second, Skip-Plan decomposes long action chains into multiple reliable sub-chains by skipping most uncertain intermediate actions. In these ways, our Skip-Plan model explores all sorts of sub-relations within an action chain and generates actions non-adjacently only in the condensed action space. Extensive experiments show the superior performance of our Skip-Plan method on two instructional video datasets. Skip-Plan has the potential to inspire the research field of procedure planning, since most of them heavily rely on state observations and adjacent actions at all times. + +Acknowledgement. This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0114903 and in part by the National Natural Science Foundation of China under Grant 62125603 and Grant 62206153, CAAI-Huawei MindSpore Open Fund. + +# References + +[1] Mindspore. .https://www.mindspore.cn/.5 +[2] Armenii Ashukha, Alexander Lyzhov, Dmitry Molchanov, and Dmitry Vetrov. Pitfalls of in-domain uncertainty estimation and ensembling in deep learning. arXiv preprint arXiv:2002.06470, 2020. 3 +[3] Anil Batra, Shreyank Gowda, Laura Sevilla-Lara, and Frank Keller. A closer look at temporal ordering in the segmentation of instructional videos. arXiv preprint arXiv:2209.15501, 2022. 2 +[4] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model-based policy learning. In ICCV, pages 15611-15620, 2021. 2, 5, 6 +[5] Nicolas Biard, Salomé Cojean, and Eric Jamet. Effects of segmentation and pacing on procedural learning by video. Comput. Hum. Behav., 89:411-417, 2018. 2 +[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 3 +[7] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In ECCV, pages 334–350, 2020. 1, 2, 3, 5, 6 +[8] Chekuri Choudary and Tiecheng Liu. Summarization of visual content in instructional videos. IEEE Trans. Multimed., 9(7):1443-1455, 2007. 2 +[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3 +[10] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 3 +[11] Mikita Dvornik, Isma Hadji, Konstantinos G Derpanis, Animesh Garg, and Allan Jepson. Drop-dtw: Aligning common signal between sequences while dropping outliers. NeuRIPS, 34:13782-13793, 2021. 2 +[12] Jakob Gawlikowski, Cedrique Rovile Njieutcheu Tassi, Mohsin Ali, Jongseok Lee, Matthias Humt, Jianxiang Feng, Anna Kruspe, Rudolph Trielbel, Peter Jung, Ribana Roscher, et al. A survey of uncertainty in deep neural networks. arXiv preprint arXiv:2107.03342, 2021. 3 +[13] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, Chiho Choi, and Behzad Dariush. Weakly-supervised online action segmentation in multi-view instructional videos. In CVPR, pages 13780-13790, 2022. 2 +[14] Reza Ghoddoosian, Saif Sayed, and Vassilis Athitsos. Hierarchical modeling for task recognition and action segmentation in weakly-labeled instructional videos. In WACV, pages 1922-1932, 2022. 2 +[15] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. arXiv preprint arXiv:1610.02136, 2016. 3 + +[16] Gabriel Huang, Bo Pang, Zhenhai Zhu, Clara Rivera, and Radu Soricut. Multimodal pretraining for dense video captioning. arXiv preprint arXiv:2011.11760, 2020. 2 +[17] Eyke Hüllermeier and Willem Waegeman. Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods. Mach. Learn, 2021. 3 +[18] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In CVPRW, pages 958-959, 2020. 2 +[19] Lei Ji, Chenfei Wu, Daisy Zhou, Kun Yan, Edward Cui, Xilin Chen, and Nan Duan. Learning temporal video procedure segmentation from an automatically collected large dataset. In WACV, pages 1506-1515, 2022. 2 +[20] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? NeurIPS, 2017. 3 +[21] Colin Lea, Michael D Flynn, Rene Vidal, Austin Reiter, and Gregory D Hager. Temporal convolutional networks for action segmentation and detection. In CVPR, pages 156-165, 2017. 2 +[22] Muheng Li, Lei Chen, Yueqi Duan, Zhilan Hu, Jianjiang Feng, Jie Zhou, and Jiwen Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In CVPR, pages 19880-19889, 2022. 2 +[23] Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. arXiv preprint arXiv:1706.02690, 2017. 3 +[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 4 +[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 3 +[26] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2 +[27] Andrey Malinin and Mark Gales. Predictive uncertainty estimation via prior networks. NeurIPS, 2018. 3 +[28] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In CVPR, pages 9879-9889, 2020. 2 +[29] Medhini Narasimhan, Arsha Nagrani, Chen Sun, Michael Rubinstein, Trevor Darrell, Anna Rohrbach, and Cordelia Schmid. Tl; dw? summarizing instructional videos with task relevance and cross-modal saliency. In ECCV, pages 540-557, 2022. 2 +[30] Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. NeurIPS, 32, 2019. 3 +[31] AJ Piergiovanni, Anelia Angelova, Michael S Ryoo, and Irfan Essa. Unsupervised action segmentation for instructional videos. arXiv preprint arXiv:2106.03738, 2021. 2 + +[32] Alireza Shafaei, Mark Schmidt, and James J Little. A less biased evaluation of out-of-distribution sample detectors. arXiv preprint arXiv:1809.04729, 2018. 3 +[33] Anshul Shah, Benjamin Lundell, Harpreet Sawhney, and Rama Chellappa. Steps: Self-supervised key step extraction from unlabeled procedural videos. arXiv preprint arXiv:2301.00794, 2023. 2 +[34] Botian Shi, Lei Ji, Yaobo Liang, Nan Duan, Peng Chen, Zhendong Niu, and Ming Zhou. Dense procedure captioning in narrated instructional videos. In ACL, pages 6382-6391, 2019. 2 +[35] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In ICCV, pages 7464-7473, 2019. 2 +[36] Jiankai Sun, De-An Huang, Bo Lu, Yun-Hui Liu, Bolei Zhou, and Animesh Garg. Plate: Visually-grounded planning with transformers in procedural tasks. IEEE Robot. Autom. Lett., 7(2):4924-4930, 2022. 2, 6 +[37] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In CVPR, pages 1207-1216, 2019. 1, 2, 5 +[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3 +[39] Mengmeng Xu, Juan-Manuel Pérez-Rúa, Victor Escorcia, Brais Martinez, Xiatian Zhu, Li Zhang, Bernard Ghanem, and Tao Xiang. Boundary-sensitive pre-training for temporal localization in videos. In ICCV, pages 7220-7230, 2021. 2 +[40] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Temporal query networks for fine-grained video understanding. In CVPR, pages 4486-4496, 2021. 3 +[41] He Zhao, Isma Hadji, Nikita Dvornik, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In CVPR, pages 2938-2948, 2022. 2, 5, 6, 7, 8 +[42] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, pages 7590-7598, 2018. 2 +[43] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In CVPR, pages 8746-8755, 2020. 2 +[44] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In CVPR, pages 3537-3545, 2019. 1, 2, 5 \ No newline at end of file diff --git a/2023/Skip-Plan_ Procedure Planning in Instructional Videos via Condensed Action Space Learning/images.zip b/2023/Skip-Plan_ Procedure Planning in Instructional Videos via Condensed Action Space Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..496678be33a252e12c2b662546f9d01e34bbbfb4 --- /dev/null +++ b/2023/Skip-Plan_ Procedure Planning in Instructional Videos via Condensed Action Space Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d29c6fe0b5b1f65976998753bdea4308c709436c2d1c48e2f5c5bd5c885fd9b9 +size 488118 diff --git a/2023/Skip-Plan_ Procedure Planning in Instructional Videos via Condensed Action Space Learning/layout.json b/2023/Skip-Plan_ Procedure Planning in Instructional Videos via Condensed Action Space Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3ac4f5f27f5505188599534760ce22f693dbf5ed --- /dev/null +++ b/2023/Skip-Plan_ Procedure Planning in Instructional Videos via Condensed Action Space Learning/layout.json @@ -0,0 +1,8328 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 465, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 465, + 140 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 465, + 140 + ], + "type": "text", + "content": "Skip-Plan: Procedure Planning in Instructional Videos via Condensed Action Space Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 160, + 520, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 160, + 520, + 175 + ], + "spans": [ + { + "bbox": [ + 71, + 160, + 520, + 175 + ], + "type": "text", + "content": "Zhiheng Li" + }, + { + "bbox": [ + 71, + 160, + 520, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 160, + 520, + 175 + ], + "type": "text", + "content": ", Wenjia Geng" + }, + { + "bbox": [ + 71, + 160, + 520, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 160, + 520, + 175 + ], + "type": "text", + "content": ", Muheng Li" + }, + { + "bbox": [ + 71, + 160, + 520, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 160, + 520, + 175 + ], + "type": "text", + "content": ", Lei Chen" + }, + { + "bbox": [ + 71, + 160, + 520, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 71, + 160, + 520, + 175 + ], + "type": "text", + "content": ", Yansong Tang" + }, + { + "bbox": [ + 71, + 160, + 520, + 175 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 71, + 160, + 520, + 175 + ], + "type": "text", + "content": ", Jiwen Lu" + }, + { + "bbox": [ + 71, + 160, + 520, + 175 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 71, + 160, + 520, + 175 + ], + "type": "text", + "content": ", Jie Zhou" + }, + { + "bbox": [ + 71, + 160, + 520, + 175 + ], + "type": "inline_equation", + "content": "^{1,4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 175, + 175, + 418, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 175, + 418, + 190 + ], + "spans": [ + { + "bbox": [ + 175, + 175, + 418, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 175, + 175, + 418, + 190 + ], + "type": "text", + "content": " Department of Automation, Tsinghua University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 190, + 449, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 190, + 449, + 203 + ], + "spans": [ + { + "bbox": [ + 143, + 190, + 449, + 203 + ], + "type": "text", + "content": "2 Shenzhen International Graduate School, Tsinghua University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 181, + 203, + 411, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 203, + 411, + 217 + ], + "spans": [ + { + "bbox": [ + 181, + 203, + 411, + 217 + ], + "type": "text", + "content": "3 Beijing University of Science and Technology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 113, + 217, + 481, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 217, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 113, + 217, + 481, + 232 + ], + "type": "text", + "content": "4 Beijing National Research Center for Information Science and Technology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 87, + 233, + 505, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 233, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 87, + 233, + 505, + 258 + ], + "type": "text", + "content": "{lizhihan21@gengwj22@,li-mh20}@mails.tsinghua.edu.cn,chenlei2022@ustb.edu.cn, {tang.yansong@sz.,lujiwen,jzhou@}tsinghua.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 286, + 191, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 286, + 191, + 300 + ], + "spans": [ + { + "bbox": [ + 143, + 286, + 191, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 312, + 290, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 290, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 290, + 553 + ], + "type": "text", + "content": "In this paper, we propose Skip-Plan, a condensed action space learning method for procedure planning in instructional videos. Current procedure planning methods all stick to the state-action pair prediction at every timestep and generate actions adjacently. Although it coincides with human intuition, such a methodology consistently struggles with high-dimensional state supervision and error accumulation on action sequences. In this work, we abstract the procedure planning problem as a mathematical chain model. By skipping uncertain nodes and edges in action chains, we transfer long and complex sequence functions into short but reliable ones in two ways. First, we skip all the intermediate state supervision and only focus on action predictions. Second, we decompose relatively long chains into multiple short sub-chains by skipping unreliable intermediate actions. By this means, our model explores all sorts of reliable sub-relations within an action sequence in the condensed action space. Extensive experiments show Skip-Plan achieves state-of-the-art performance on the CrossTask and COIN benchmarks for procedure planning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 576, + 128, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 576, + 128, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 576, + 128, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 597, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 597, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 287, + 693 + ], + "type": "text", + "content": "Given the initial and goal visual observations which are curated from instructional videos [44, 37], the task of procedure planning is to predict a sequence of actions to reach the indicated goal [7], as illustrated in Figure 1(a). This is an emerging research direction in the field of video understanding. Procedure planning has worldwide applications and assists humans with instruction-following tasks, such as cooking, assembling furniture, electronics repair, etc." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 288, + 425, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 288, + 425, + 308 + ], + "spans": [ + { + "bbox": [ + 312, + 288, + 425, + 308 + ], + "type": "text", + "content": "Question: What action sequence are needed to get from " + }, + { + "bbox": [ + 312, + 288, + 425, + 308 + ], + "type": "inline_equation", + "content": "v_{start}" + }, + { + "bbox": [ + 312, + 288, + 425, + 308 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 312, + 288, + 425, + 308 + ], + "type": "inline_equation", + "content": "v_{goal}" + }, + { + "bbox": [ + 312, + 288, + 425, + 308 + ], + "type": "text", + "content": "?" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 323, + 310, + 427, + 361 + ], + "blocks": [ + { + "bbox": [ + 323, + 310, + 427, + 361 + ], + "lines": [ + { + "bbox": [ + 323, + 310, + 427, + 361 + ], + "spans": [ + { + "bbox": [ + 323, + 310, + 427, + 361 + ], + "type": "image", + "image_path": "b69948dc1487f9e8bff0863ec5ecd0094f4913a48597e658da62001680dd1a32.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 365, + 465, + 374 + ], + "lines": [ + { + "bbox": [ + 386, + 365, + 465, + 374 + ], + "spans": [ + { + "bbox": [ + 386, + 365, + 465, + 374 + ], + "type": "text", + "content": "(a) Problem Definition" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 430, + 288, + 541, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 430, + 288, + 541, + 297 + ], + "spans": [ + { + "bbox": [ + 430, + 288, + 541, + 297 + ], + "type": "text", + "content": "Answer: Required action sequence" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 446, + 297, + 521, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 446, + 297, + 521, + 306 + ], + "spans": [ + { + "bbox": [ + 446, + 297, + 521, + 306 + ], + "type": "text", + "content": "(Time Horizon " + }, + { + "bbox": [ + 446, + 297, + 521, + 306 + ], + "type": "inline_equation", + "content": "T = 4" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 449, + 308, + 533, + 359 + ], + "blocks": [ + { + "bbox": [ + 449, + 308, + 533, + 359 + ], + "lines": [ + { + "bbox": [ + 449, + 308, + 533, + 359 + ], + "spans": [ + { + "bbox": [ + 449, + 308, + 533, + 359 + ], + "type": "image", + "image_path": "29ec5d8dee3b24a6f4f646646fd8eaac91ab1a59a2ab7d331b5bd4d3130910d8.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 318, + 378, + 528, + 434 + ], + "blocks": [ + { + "bbox": [ + 318, + 378, + 528, + 434 + ], + "lines": [ + { + "bbox": [ + 318, + 378, + 528, + 434 + ], + "spans": [ + { + "bbox": [ + 318, + 378, + 528, + 434 + ], + "type": "image", + "image_path": "248d852484ca6d081083b77706bb1cec3e6b9db73d2358a2a53429c00375c233.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 435, + 504, + 445 + ], + "lines": [ + { + "bbox": [ + 339, + 435, + 504, + 445 + ], + "spans": [ + { + "bbox": [ + 339, + 435, + 504, + 445 + ], + "type": "text", + "content": "(b) Conventional Procedure Planning Paradigm" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 320, + 449, + 531, + 515 + ], + "blocks": [ + { + "bbox": [ + 320, + 449, + 531, + 515 + ], + "lines": [ + { + "bbox": [ + 320, + 449, + 531, + 515 + ], + "spans": [ + { + "bbox": [ + 320, + 449, + 531, + 515 + ], + "type": "image", + "image_path": "57cd4e7a4e8bf0c1e11c4180e00653ee59f406f5fad470388633dd5209df88c8.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 407, + 517, + 453, + 526 + ], + "lines": [ + { + "bbox": [ + 407, + 517, + 453, + 526 + ], + "spans": [ + { + "bbox": [ + 407, + 517, + 453, + 526 + ], + "type": "text", + "content": "(c)Skip-Plan" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 312, + 530, + 541, + 555 + ], + "blocks": [ + { + "bbox": [ + 312, + 530, + 541, + 555 + ], + "lines": [ + { + "bbox": [ + 312, + 530, + 541, + 555 + ], + "spans": [ + { + "bbox": [ + 312, + 530, + 541, + 555 + ], + "type": "image", + "image_path": "0d9a2dd55e040bdd5893ea75417e20dc29eff784da94beba768ffbb0d7aad84f.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "content": "Figure 1. Comparisons of the conventional paradigm and our proposed Skip-Plan in procedure planning. (a) Problem definition: given the initial and goal visual observations " + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "inline_equation", + "content": "(v_{start}" + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "inline_equation", + "content": "v_{goal})" + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "content": ", procedure planning is to answer an action sequence to reach the indicated goal with a pre-defined time horizon " + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "content": ". (b) Conventional paradigm predicts a state-action pair " + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\{\\tilde{s}_t,\\tilde{a}_t\\}" + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "content": " at every timestep " + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "content": ", and repeat it step-by-step. (c) Our Skip-Plan skips state supervision and extracts multiple sub-chains by skipping most intermediate actions. The final action sequence is accumulated from all predicted sub-chains. Skip-Plan significantly reduces the overall compound error, indicated by the shallower darkness of the chevron arrows." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 178, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 178, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 178, + 712 + ], + "type": "text", + "content": "* indicates the corresponding author." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10297" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 420 + ], + "type": "text", + "content": "In the past few years, various approaches have been proposed to solve the procedure planning problem. Most existing methods utilize an autoregressive model with visual state supervision to generate actions one by one [7, 4, 36]. Another type of work [41] solves it in a non-autoregressive way with language state supervision. The difficulties of procedure planning have been widely discussed in the previous works. For instance, the sequential orders in procedure planning can be altered but still sensible (e.g., switching the order of 'pour water' and 'pour alcohol' does not matter in many cases), and the visual states are unstructured and high-dimensional. Thus, we can conclude the following limitations from the existing works. (a) State supervision: the major difficulty in procedure planning is state supervision. Visual images are high-dimensional and complex, with extra misleading information such as background, illumination, viewing point, etc. (b) Step-by-step predictions: the conventional paradigm in procedure planning predicts one state-action pair at every timestep, and repeats it step-by-step, as illustrated in Figure 1(b). Even though the proposed non-autoregressive model generates whole action sequences in batches [41], it can be still regarded as a step-by-step method because the model neglects the links between non-adjacent actions. From Figure 1(b), we can see while predicting an action sequence with a time horizon of 4, the conventional model is equivalently predicting a sequence with a length of 7 (4 action nodes, 3 state nodes). More frequent predictions tend to accumulate a larger compound error along the chain, which lowers the prediction accuracy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 422, + 289, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 289, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 289, + 615 + ], + "type": "text", + "content": "To overcome these limitations, we propose Skip-Plan, a condensed action space learning method that solves the procedure planning problem through the lens of mathematics, aiming to reduce compound error along action chains. As we have discussed above, existing methods struggle with high-dimensional and complex state supervision. Thus, the compound error along the action chain is increased due to large errors in state predictions. This motivates us to skip all the state nodes in our Skip-Plan framework and only model sequence relations in the condensed action space, as demonstrated in Figure 1(c). Without the presence of state supervision, a lower compound error results in better performance in Skip-Plan. Although the dependencies between high-level actions without state supervision are very abstract, the algorithm of our Skip-Plan is highly mathematically sensible in the Markov chain model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 289, + 715 + ], + "type": "text", + "content": "Another convention in procedure planning is we should learn sequential relations between adjacent actions. The Markov chain model inspires us to decouple relatively long chains into multiple short sub-chains, because the compound error can be reduced with shorter chain lengths. The problem here is how to decouple these long chains, because there are various choices of sub-chains with different lengths and we should always choose the most reliable" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": "ones under any condition. Given the initial and goal frames as concrete inputs, it is apparent the reliabilities of the initial and final actions are the highest, and the intermediate actions are more likely to be predicted wrong. Therefore, our Skip-Plan decomposes long chains by 'skipping' some unreliable intermediate actions, and each sub-chain in our model is composed of the initial action, one unique intermediate action, and the final action. One example for a chain length of 4 is shown in Figure 1(c). In this way, the reliable sub-relations between non-adjacent actions can be well explored by the network. To model the discriminative relations of all sub-chains, we propose sub-chain decoders, where each decoder only learns one sub-chain by the specific loss design. The final action sequence is aggregated from all sub-chain predictions by a sub-chain accumulator." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "text", + "content": "In summary, our contributions are threefold. (a) Unlike the previous methods which put sophisticated designs on state supervision, Skip-Plan skips unreliable state predictions to boost the model performance in procedure planning. (b) In contrast to the existing methods which only focus on the links between adjacent actions, we decompose long chains into multiple reliable sub-chains by skipping some intermediate actions, and propose sub-chain decoders to learn their discriminative relations. (c) Extensive experiments show Skip-Plan achieves state-of-the-art performance on the CrossTask [44] and COIN [37] benchmarks for procedure planning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 406, + 392, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 392, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 392, + 418 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 426, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 594 + ], + "type": "text", + "content": "Procedure Planning. Procedure planning in instructional videos was first introduced by Chang et al. [7]. Recently, various methods have been conducted to solve this task. Chang et al. [7] propose a dual dynamic network to conjugate the relationships between states and actions. Sun et al. [36] utilize a GPT network to generate actions and states one by one. Bi et al. [4] use the model-based reinforcement learning with time-invariant context information to predict state-action pairs autoregressively. Different from the previous autoregressive methods with visual state supervision, Zhao et al. [41] use a non-autoregressive transformer decoder to generate action sequences with language state supervision. Unlike previous works, Skip-Plan generates actions non-adjacently without any state supervision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "content": "Instructional Videos Analysis. Instructional videos are used for learning human behaviours because they are visual representations of human actions. Based on instructional videos, various action-related tasks have been proposed, such as action segmentation[21, 22, 13, 14, 28, 31], step localization[39, 11, 43, 33, 26], procedure segmentation[42, 19, 3, 5], video summarization/captioning[8, 35, 29, 34, 16, 18], etc. Different from these action tasks in instructional videos, we investigate procedure planning, a goal-conditioned problem that plans a sequence of actions given" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10298" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 206, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 206, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 206, + 83 + ], + "type": "text", + "content": "the initial and goal visual observations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 289, + 288 + ], + "type": "text", + "content": "Predictive Uncertainty on Sequence Relations. All types of questions that we are interested in can be modeled as a sequence model, such as images [25] and languages [9]. Currently, Transformer [38] is the most effective network to model sequence relations, and our model is based on the transformer decoder with a learnable query input[40, 6]. However, sequence relations are always distorted by predictive uncertainty[12, 27, 17, 30, 20], and the sources of predictive uncertainty can be various, including model uncertainty, data uncertainty, etc[2, 10, 15, 23, 32]. In this work, we pay particular attention to data uncertainty, in which the uncertainty directly stems from the data. Data uncertainty is caused by information loss when we measure a data distribution in a space with incorrect dimensions (e.g., we must lose depth information if we measure a three-dimensional object in a two-dimensional space). Therefore, we often avoid high-dimensional data for lower data uncertainty." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 297, + 107, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 107, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 107, + 309 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 316, + 171, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 171, + 328 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 171, + 328 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 335, + 288, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 288, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 288, + 419 + ], + "type": "text", + "content": "We follow the definition of procedure planning proposed by Chang et al. [7]. Here, we denote the action and the state at timestep " + }, + { + "bbox": [ + 46, + 335, + 288, + 419 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 335, + 288, + 419 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 335, + 288, + 419 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 46, + 335, + 288, + 419 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 335, + 288, + 419 + ], + "type": "inline_equation", + "content": "s_{t}" + }, + { + "bbox": [ + 46, + 335, + 288, + 419 + ], + "type": "text", + "content": " respectively. Given the initial visual observation " + }, + { + "bbox": [ + 46, + 335, + 288, + 419 + ], + "type": "inline_equation", + "content": "v_{start}" + }, + { + "bbox": [ + 46, + 335, + 288, + 419 + ], + "type": "text", + "content": " and the goal visual observation " + }, + { + "bbox": [ + 46, + 335, + 288, + 419 + ], + "type": "inline_equation", + "content": "v_{goal}" + }, + { + "bbox": [ + 46, + 335, + 288, + 419 + ], + "type": "text", + "content": ", the model generates a sequence of actions " + }, + { + "bbox": [ + 46, + 335, + 288, + 419 + ], + "type": "inline_equation", + "content": "a_{1:T}" + }, + { + "bbox": [ + 46, + 335, + 288, + 419 + ], + "type": "text", + "content": " to reach the indicated goal. In this work, we introduce a novel formulation for the procedure planning task:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 424, + 287, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 424, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 424, + 287, + 456 + ], + "type": "interline_equation", + "content": "p \\left(a _ {1: T} \\mid v _ {s}, v _ {g}\\right) = \\prod_ {t = 2} ^ {T - 1} p \\left(a _ {t} \\mid a _ {1}, a _ {T}\\right) p \\left(a _ {1}, a _ {T} \\mid v _ {s}, v _ {g}\\right), \\tag {1}", + "image_path": "d7200cceacb8c4393f839f254709e32adb11280f1bd1bbd0e1e094215d99b670.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "v_{s}" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "v_{g}" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": " short for ' " + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "v_{start}" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": " ' and ' " + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "v_{goal}" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": " ' . We denote " + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "p(a_{1:T}|v_{start},v_{goal})" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": " as the conditional probability distribution of the action sequence given the initial visual observation " + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "v_{start}" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": " and the goal visual observation " + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "v_{goal}" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "p(a_t|a_1,a_T)" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": " as the conditional probability distribution of intermediate actions " + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "(2\\leq t\\leq T - 1)" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": " given the initial action " + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "a_1" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": " and the final action " + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "a_{T}" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "p(a_{1},a_{T}|v_{start},v_{goal})" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": " as the conditional probability distribution of the initial action " + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "a_1" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": " and the final action " + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "a_{T}" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": " given the initial visual observation " + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "v_{start}" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": " and the goal visual observation " + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "v_{goal}" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": ". Equation 1 reveals the uniqueness of our Skip-Plan method: First, we do not rely on any intermediate state supervision " + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": " for action generations. Second, every intermediate action prediction " + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "(2\\leq t\\leq T - 1)" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": " only depends on the initial action " + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "a_1" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": " and the final action " + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "inline_equation", + "content": "a_{T}" + }, + { + "bbox": [ + 46, + 461, + 289, + 666 + ], + "type": "text", + "content": ". Thus, Skip-Plan does not focus on sequential relations between adjacent actions but predicts action sequences non-adjacently without state supervision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 671, + 280, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 280, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 280, + 685 + ], + "type": "text", + "content": "3.2. Action Predictions without State Supervision" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Information does not always help in the decision-making process. Sometimes ambiguous information can even lead" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 455 + ], + "type": "text", + "content": "to sub-optimal results, and plausible plans can be only made based on reliable information. In procedure planning, two types of supervision can be utilized, including actions " + }, + { + "bbox": [ + 304, + 72, + 547, + 455 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 304, + 72, + 547, + 455 + ], + "type": "text", + "content": " and visual states " + }, + { + "bbox": [ + 304, + 72, + 547, + 455 + ], + "type": "inline_equation", + "content": "s_{t}" + }, + { + "bbox": [ + 304, + 72, + 547, + 455 + ], + "type": "text", + "content": ". Even though they can be directly generated from each other, the intrinsic properties of these two supervisions are largely different. We evaluate them from the perspectives of the value of information and predictive uncertainty. Visual state supervision is well known to be high-dimensional. In addition to the factor of action class, various factors such as background, viewing point, illumination, etc. can be involved in visual states. Thus, useful information in visual state supervision for action predictions is overwhelmed by many irrelevant factors. In contrast, action labels are the most condensed supervision, without being interrupted by any factors. From this point, we argue the value of information offered by state supervision is extremely low because the portion of information, which is relevant to action predictions, is too small. On the other hand, when models generate high-dimensional predictions such as visual images, large predictive errors are more likely to occur. Thus, the predictive uncertainty of intermediate visual states is significantly higher than the one of condensed action predictions. Combining the effects of the value of information and the predictive uncertainty, we argue the existence of visual state supervision hurts the predictive accuracy of sequential actions. Even though we can use the weak language state supervision to reduce some irrelevant dimensions, it is still too high-dimensional compared to condensed action labels. Therefore, we 'skip' intermediate state supervision in our Skip-Plan framework. The overall compound error along an action chain is significantly reduced without the predictive errors introduced by states." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 456, + 547, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 547, + 624 + ], + "type": "text", + "content": "Getting rid of intermediate state supervision does not mean all sorts of visual information are useless. In procedure planning, the initial visual observation " + }, + { + "bbox": [ + 304, + 456, + 547, + 624 + ], + "type": "inline_equation", + "content": "v_{start}" + }, + { + "bbox": [ + 304, + 456, + 547, + 624 + ], + "type": "text", + "content": " and the goal visual observation " + }, + { + "bbox": [ + 304, + 456, + 547, + 624 + ], + "type": "inline_equation", + "content": "v_{goal}" + }, + { + "bbox": [ + 304, + 456, + 547, + 624 + ], + "type": "text", + "content": " are vital because action sequences are constrained by them. These visual observations are in the form of frame features in seconds. Inspired by the image analysis above, we argue the visual input information is underutilized by the previous works, especially along the time dimension. In the Skip-Plan framework, we take a shared multi-layer perceptrons (MLP) network in the Visual Input Module to learn the time information, where the MLP network is three-layered with ReLU activations. This simple method can boost the model performance by a large margin, especially for short sequences " + }, + { + "bbox": [ + 304, + 456, + 547, + 624 + ], + "type": "inline_equation", + "content": "(T = 3)" + }, + { + "bbox": [ + 304, + 456, + 547, + 624 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 634, + 473, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 473, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 473, + 647 + ], + "type": "text", + "content": "3.3. Decomposition of Long Chains" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "Long chains are not favored when a compound error is accumulating. Thus, decomposing long chains into short ones is a natural way to better model performance. Through the lens of mathematics, the only design philosophy here is that we should extract most reliable sub-chains from the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10299" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 545, + 225 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 545, + 225 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 545, + 225 + ], + "type": "image", + "image_path": "626640d6c67c78dd6e9559a4d27b6c9010289993c200761769d0796d28562694.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 228, + 547, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 547, + 316 + ], + "type": "text", + "content": "Figure 2. Our Skip-Plan architecture. The shared MLP network in the Visual Input Module extracts the initial and final information from the initial and goal observations (" + }, + { + "bbox": [ + 46, + 228, + 547, + 316 + ], + "type": "inline_equation", + "content": "v_{start}" + }, + { + "bbox": [ + 46, + 228, + 547, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 228, + 547, + 316 + ], + "type": "inline_equation", + "content": "v_{goal}" + }, + { + "bbox": [ + 46, + 228, + 547, + 316 + ], + "type": "text", + "content": ") and attaches them to the first and last dimensions of every learnable query input. A number of " + }, + { + "bbox": [ + 46, + 228, + 547, + 316 + ], + "type": "inline_equation", + "content": "T - 2" + }, + { + "bbox": [ + 46, + 228, + 547, + 316 + ], + "type": "text", + "content": " query inputs are separately passed to each decoder in the Sub-chain Decoder Module. In each decoder, the input is processed by the self-attention block and then by the cross-attention block, where the cross-attention block is augmented by a learnable memory " + }, + { + "bbox": [ + 46, + 228, + 547, + 316 + ], + "type": "inline_equation", + "content": "M_i" + }, + { + "bbox": [ + 46, + 228, + 547, + 316 + ], + "type": "text", + "content": ". The sub-chain decoders are separated. The " + }, + { + "bbox": [ + 46, + 228, + 547, + 316 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 228, + 547, + 316 + ], + "type": "text", + "content": "th decoder outputs the sub-chain " + }, + { + "bbox": [ + 46, + 228, + 547, + 316 + ], + "type": "inline_equation", + "content": "\\{ \\tilde{a}_1, \\tilde{a}_t, \\tilde{a}_T \\}" + }, + { + "bbox": [ + 46, + 228, + 547, + 316 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 228, + 547, + 316 + ], + "type": "inline_equation", + "content": "t = n + 1" + }, + { + "bbox": [ + 46, + 228, + 547, + 316 + ], + "type": "text", + "content": ". The predicted sub-chains are supervised by two loss supervision. In Individual Sub-chain Supervision, the predicted sub-chains are supervised by the corresponding ground-truth sub-chains. In Complete Chain Supervision, the predicted sub-chains are aggregated by the Sub-chain Accumulator to generate the complete action chain, which is supervised by the complete ground-truth action chain." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 335, + 287, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 383 + ], + "type": "text", + "content": "original long chain. Following this principle, we ought to find there is so much flexibility in the decoupling method: We can choose sub-chains with any lengths, and actions in the sub-chains are not necessary to be adjacent." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 384, + 288, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 288, + 552 + ], + "type": "text", + "content": "The reliability of a sub-chain can be determined by the uncertainty of predictions. According to the uncertainty analysis, the reliabilities of the initial and final actions are the highest because they are the closest actions to the initial and goal visual inputs. Since the compound error is accumulated from two ends of the chain, the reliability of action predictions is decreasing while approaching the middle of the chain. Consequently, we remove all the edges between adjacent intermediate actions, and each individual intermediate action only connects to the reliable initial and final actions. This design just coincides with the term " + }, + { + "bbox": [ + 46, + 384, + 288, + 552 + ], + "type": "inline_equation", + "content": "p(a_{t}|a_{1},a_{T})" + }, + { + "bbox": [ + 46, + 384, + 288, + 552 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 384, + 288, + 552 + ], + "type": "inline_equation", + "content": "2\\leq t\\leq T - 1" + }, + { + "bbox": [ + 46, + 384, + 288, + 552 + ], + "type": "text", + "content": " in Equation 1. In this way, we generate a quantity of " + }, + { + "bbox": [ + 46, + 384, + 288, + 552 + ], + "type": "inline_equation", + "content": "T - 2" + }, + { + "bbox": [ + 46, + 384, + 288, + 552 + ], + "type": "text", + "content": " sub-chains from an action chain with a time horizon of " + }, + { + "bbox": [ + 46, + 384, + 288, + 552 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 384, + 288, + 552 + ], + "type": "text", + "content": ". The set of sub-chains is given by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 559, + 287, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 559, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 71, + 559, + 287, + 574 + ], + "type": "interline_equation", + "content": "\\{(a _ {1}, a _ {2}, a _ {T}), (a _ {1}, a _ {3}, a _ {T}), \\dots , (a _ {1}, a _ {T - 1}, a _ {T}) \\}. \\quad (2)", + "image_path": "a4a6fff5a96e190e85596945db70c8d600fee0e2e51d4e2b49c5fbe854d31ecc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 629 + ], + "type": "text", + "content": "Each sub-chain has a length of 3, with the order of the initial action, one unique intermediate action, and the final action. This decoupling method works well for relatively long chains " + }, + { + "bbox": [ + 46, + 582, + 287, + 629 + ], + "type": "inline_equation", + "content": "(T > 3)" + }, + { + "bbox": [ + 46, + 582, + 287, + 629 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": "Given multiple short sub-chains, we design a compatible network to model all their sub-relations. Since these sub-chains are parallel and independent to each other, we propose the Sub-chain Decoder Module, where each transformer decoder only models one discriminative relation of one sub-chain. Every transformer decoder is non-autoregressive with two types of inputs: a learnable query" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 335, + 547, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 547, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 547, + 492 + ], + "type": "text", + "content": "with the initial and goal visual features, and a learnable memory component. The novelty here is each decoder only learns one sub-relation, with a unique loss only corresponding to the three actions of that sub-chain. The query input, parameters, and memory component of every decoder are separated from other decoders'. Therefore, the number of decoders in the Sub-chain Decoder Module is equal to the number of sub-chains " + }, + { + "bbox": [ + 304, + 335, + 547, + 492 + ], + "type": "inline_equation", + "content": "(T - 2)" + }, + { + "bbox": [ + 304, + 335, + 547, + 492 + ], + "type": "text", + "content": " extracted from the original chain. The final sequence generation is aggregated from all the sub-chain predictions through the Sub-chain Accumulator, which is an MLP network. In this way, our Skip-Plan network can model all sorts of discriminative sub-relations within the original action sequence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 497, + 371, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 497, + 371, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 497, + 371, + 510 + ], + "type": "text", + "content": "3.4. Training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 515, + 547, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 547, + 564 + ], + "type": "text", + "content": "Loss is an essential component to discriminate sub-chain decoders. Instead of conventional cross-entropy loss, we use the focal loss to supervise our sub-chain decoders. The focal loss " + }, + { + "bbox": [ + 304, + 515, + 547, + 564 + ], + "type": "inline_equation", + "content": "FL" + }, + { + "bbox": [ + 304, + 515, + 547, + 564 + ], + "type": "text", + "content": " is formulated as[24]:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 358, + 568, + 545, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 568, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 358, + 568, + 545, + 601 + ], + "type": "interline_equation", + "content": "F L = - \\sum_ {t = 1} ^ {T} a _ {t} (1 - \\tilde {a} _ {t}) ^ {\\gamma} \\log \\tilde {a} _ {t}, \\tag {3}", + "image_path": "a221a4d2739e3b328c00618eb85701839d82fc4d79320acc5ec2fde203be2e1f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "where we denote " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " as the ground-truth one-hot labels, " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{a}_{t}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " as the predicted probability distribution of the action sequence, and " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " as the gamma value of focal loss. The focal loss is equivalent to the cross-entropy loss when " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\gamma = 0" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ". The model training is benefited from the use of focal loss because the data distribution in procedure planning is imbalanced. In this work, we define " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "FL(a_{1},a_{t},a_{T})" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " as the focal loss for the sub-chain with a length of 3, where the sub-chain is composed of the initial action " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "a_{1}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ", the action at timestep " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ", and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10300" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "the final action " + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "a_{T}" + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": ". We present the focal loss of the whole action sequence as " + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "FL(a_{1:T})" + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 287, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 287, + 155 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 287, + 155 + ], + "type": "text", + "content": "Loss of Individual Sub-chains. As we have discussed before, the loss of each decoder only corresponds to the actions within that sub-chain. Therefore, the individual loss of the " + }, + { + "bbox": [ + 46, + 96, + 287, + 155 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 96, + 287, + 155 + ], + "type": "text", + "content": "th decoder is defined as " + }, + { + "bbox": [ + 46, + 96, + 287, + 155 + ], + "type": "inline_equation", + "content": "FL(a_{1},a_{t},a_{T})" + }, + { + "bbox": [ + 46, + 96, + 287, + 155 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 96, + 287, + 155 + ], + "type": "inline_equation", + "content": "t = n + 1" + }, + { + "bbox": [ + 46, + 96, + 287, + 155 + ], + "type": "text", + "content": ", and the loss from all sub-chain decoders can be written as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 110, + 163, + 287, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 163, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 110, + 163, + 287, + 196 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {N} = \\sum_ {t = 2} ^ {T - 1} F L \\left(a _ {1}, a _ {t}, a _ {T}\\right). \\tag {4}", + "image_path": "46ec69771f3f6ae714844e2cea2d82245e0e42bfeddfc34d48eaab5297327f3d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 204, + 287, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 287, + 251 + ], + "type": "text", + "content": "Loss of Complete Chain. After aggregating all sub-chains by the Sub-chain Accumulator, the sequence output with a length of " + }, + { + "bbox": [ + 46, + 204, + 287, + 251 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 204, + 287, + 251 + ], + "type": "text", + "content": " is supervised by the focal loss of the whole action sequence:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 262, + 287, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 262, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 131, + 262, + 287, + 275 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T} = F L \\left(a _ {1: T}\\right). \\tag {5}", + "image_path": "9d361122e97d85d1eafdea1cbcccb66ad4019f7516af5e03f3392133c46f770e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 284, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 287, + 319 + ], + "type": "text", + "content": "Overall Loss. Therefore, the total loss is defined as the sum of the loss from all sub-chains and the complete chain loss, which can be written as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 330, + 287, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 330, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 134, + 330, + 287, + 342 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {N} + \\mathcal {L} _ {T}. \\tag {6}", + "image_path": "554d9d7487f20ed05ee565f3568ca9de613526e27fed4eb92e8257685e54f3b1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 351, + 287, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 287, + 388 + ], + "type": "text", + "content": "In this way, our model can explore all sorts of discriminative but reliable sub-relations within an action sequence, while guaranteeing the final sequence generation is optimal." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 395, + 180, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 395, + 180, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 395, + 180, + 407 + ], + "type": "text", + "content": "3.5. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "text", + "content": "For each decoder, the query input length is " + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "inline_equation", + "content": "T + 1" + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "text", + "content": ", and the memory size is 128. The numbers of decoder layers, decoder heads, and decoder dimensions are set to be 1, 16, and 1024 respectively. All MLP networks are three-layered with ReLU activations. In the Visual Input Module, our shared MLP network on the time dimension of input frame features has the shape " + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "inline_equation", + "content": "[3 \\to 6 \\to 1]" + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "text", + "content": ". All action classifiers in the Sub-chain Decoder Module are the same MLP networks with the shape " + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "inline_equation", + "content": "[1024 \\to 512 \\to n_a]" + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "inline_equation", + "content": "n_a" + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "text", + "content": " is the number of action classes. The Sub-chain Accumulator is an MLP network with the shape " + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "inline_equation", + "content": "[T(T - 2) \\to 3T(T - 2) \\to T]" + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "text", + "content": ". Models are optimized for 500 epochs with stochastic gradient descent. The initial learning rate is 0.02 and decays by 10 at every 50 epochs after 100 epochs. Experiments are conducted on a single GeForce RTX 3090 GPU. We also test our model using the Mindspace [1]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 635, + 167, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 167, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 167, + 647 + ], + "type": "text", + "content": "4.1. Evaluation Protocols" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": "Datasets. We evaluate our model on two real-world instructional video datasets: CrossTask[44] and COIN[37]. CrossTask contains 2763 videos with a duration of 213 hours. It consists of 18 different tasks with 7.4 steps per task. The COIN dataset spans 12 different domains related" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "to our daily life. It contains 11,827 videos with 3.9 steps per video. The COIN dataset is challenging due to the diversity of its tasks and action labels (180 tasks and 779 action labels). We curate a set of plans by the sliding window approach, and adopt the standard " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "70\\% / 30\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " to create our train/test splits. The division is done on the generated samples rather than the original video. The data pre-processing procedure in our work is identical to the most widely used protocol in the previous procedure planning works (e.g., DDN [7], Ext-GAIL [4], and the main protocol in P3IV [41]) for fair comparisons." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 205, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 384 + ], + "type": "text", + "content": "Metrics. Following the previous works, we used three metrics to comprehensively evaluate model performance. For each action sequence: (a) mean Intersection over Union (mIoU) regards the predicted and ground-truth action sequences as two sets, and measure the overlap between these sets. The mIoU metric only indicates whether desired actions exist in the predicted action set, agnostic to the order of actions. (b) mean Accuracy (mAcc) evaluates the correctness of the predicted action at every timestep, termed as percentages. Thus, mAcc counts the order of actions, and is more strict than mIoU. (c) Success Rate (SR) evaluates the correctness of the whole action sequence. A predicted sequence is successful only if it matches the ground-truth sequence. The success rate is the proportion of successful samples over all samples, which is more strict than mAcc." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 385, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 468 + ], + "type": "text", + "content": "Training Scheme on Each Dataset. We specify an appropriate gamma value on each dataset to maximize model performance. Through experiments, we find the gamma values of 1.5 and 0 are the best parameters for the CrossTask and COIN datasets respectively. This is reasonable because the data distribution in the COIN dataset is more balanced due to a large quantity of action labels in diverse tasks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 479, + 476, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 479, + 476, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 479, + 476, + 491 + ], + "type": "text", + "content": "4.2. Evaluating Procedure Planning" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "Short-horizon Prediction. We compare the model performance of the short-horizon predictions " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "(T = 3 / 4)" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " in Table 1. As illustrated in Table 1, Skip-Plan significantly outperforms the previous state-of-the-art model at all metrics for " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "T = 3" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": ". In the CrossTask dataset, our SR and mAcc results surpass the results of the second best approach (e.g., P3IV[41]) by " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": ". Our method also outperforms P3IV[41] by " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " on SR and " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " on mAcc in the COIN dataset. Such incredible improvements strongly prove Skip-Plan without state supervision suppresses predictive uncertainty effectively, and our module design on the visual input extracts initial and goal visual information thoroughly. For " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "T = 4" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": ", Skip-Plan achieves large gains of around " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " in SR, mAcc and mIoU respectively in the COIN dataset. For the CrossTask dataset, the mAcc and mIoU of our Skip-Plan model are the best and the SR is the second best. Particularly, our mAcc is higher than P3IV's[41] by " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "11.48\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": ". The methods of no state supervision and decou" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10301" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 94, + 520, + 274 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 547, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 547, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 547, + 92 + ], + "type": "text", + "content": "Table 1. Metric results on CrossTask and COIN for short-horizon predictions " + }, + { + "bbox": [ + 47, + 69, + 547, + 92 + ], + "type": "inline_equation", + "content": "\\left( {T \\in \\{ 3,4\\} }\\right)" + }, + { + "bbox": [ + 47, + 69, + 547, + 92 + ], + "type": "text", + "content": " . The ’ " + }, + { + "bbox": [ + 47, + 69, + 547, + 92 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 47, + 69, + 547, + 92 + ], + "type": "text", + "content": " ’ and ’ " + }, + { + "bbox": [ + 47, + 69, + 547, + 92 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 69, + 547, + 92 + ], + "type": "text", + "content": " ’ in the column State Supervision denote with and without state supervision respectively. The best and second best metric results are coloured in red and blue respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 94, + 520, + 274 + ], + "lines": [ + { + "bbox": [ + 70, + 94, + 520, + 274 + ], + "spans": [ + { + "bbox": [ + 70, + 94, + 520, + 274 + ], + "type": "table", + "html": "
HorizonsModelsState SupervisionCrossTaskCOIN
SRmAccmIoUSRmAccmIoU
T=3DDN[7]✓12.1831.2947.4813.9020.1964.78
Plate[36]✓16.0036.1765.91---
Ext-GAIL[4]✓21.2749.4661.70---
P3IV[41]✓23.3449.9673.8915.4021.6776.31
Skip-Plan✗28.8561.1874.9823.6547.1278.44
T=4DDN[7]✓5.9727.1048.4611.1317.7168.06
Plate[36]✓14.0035.2955.36---
Ext-GAIL[4]✓16.4143.0560.93---
P3IV[41]✓13.4044.1670.0111.3218.8570.53
Skip-Plan✗15.5655.6470.3016.0443.1977.07
", + "image_path": "26721f071576e71c2d14f1df142cba5917b2e8c763e928de1cca0c8615b9352d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 312, + 285, + 457 + ], + "blocks": [ + { + "bbox": [ + 47, + 291, + 286, + 312 + ], + "lines": [ + { + "bbox": [ + 47, + 291, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 291, + 286, + 312 + ], + "type": "text", + "content": "Table 2. Metric results for long-horizon predictions " + }, + { + "bbox": [ + 47, + 291, + 286, + 312 + ], + "type": "inline_equation", + "content": "(T \\in \\{5,6\\})" + }, + { + "bbox": [ + 47, + 291, + 286, + 312 + ], + "type": "text", + "content": " on CrossTask and COIN." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 312, + 285, + 457 + ], + "lines": [ + { + "bbox": [ + 48, + 312, + 285, + 457 + ], + "spans": [ + { + "bbox": [ + 48, + 312, + 285, + 457 + ], + "type": "table", + "html": "
DatasetsHorizonsModelsSRmAccmIoU
CrossTaskT=5DDN[7]3.10--
P3IV[41]7.21--
Skip-Plan8.5552.0569.09
T=6DDN[7]1.20--
P3IV[41]4.40--
Skip-Plan5.1247.3466.11
COINT=5P3IV[41]4.2710.8168.81
Skip-Plan9.9038.9976.93
T=6P3IV[41]4.178.9767.72
Skip-Plan6.6133.6373.68
", + "image_path": "1e7d012dabeaee0851c39e0164b51214d40ce84e39c0c338cd90b9f1ce6e05db.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 498, + 286, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 286, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 286, + 511 + ], + "type": "text", + "content": "pling long chains both drive Skip-Plan to succeed at " + }, + { + "bbox": [ + 47, + 498, + 286, + 511 + ], + "type": "inline_equation", + "content": "T = 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": "Long-horizon Prediction. We further evaluate Skip-Plan on long-horizon predictions (" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "inline_equation", + "content": "T = 5/6" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": "). In the COIN dataset, our Skip-Plan performs extraordinarily well on the metric of mAcc, with remarkable improvements of " + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "inline_equation", + "content": "T = 5/6" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": ", compared to the second best model P3IV[41]. As for the metrics of SR and mIoU, Skip-Plan exceeds P3IV[41] by " + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "inline_equation", + "content": "T = 5" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "inline_equation", + "content": "T = 6" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": " respectively. The outstanding performance we presented above further validates the success of the decoupling approach. Our Skip-Plan also achieved the best results on SR in the CrossTask dataset, surpassing P3IV[41] by " + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "inline_equation", + "content": "1.34\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "inline_equation", + "content": "0.72\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "inline_equation", + "content": "T = 5/6" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": ". For the long-horizon predictions, we argue mAcc and mIoU are more reliable metrics to measure the effectiveness of a model. SR is interrupted by a large variance, because data samples become insufficient with a large time horizon." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 293, + 448, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 293, + 448, + 306 + ], + "spans": [ + { + "bbox": [ + 306, + 293, + 448, + 306 + ], + "type": "text", + "content": "4.3. Prerequisite Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 314, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 545, + 386 + ], + "type": "text", + "content": "In this section, we carried out several prerequisite experiments for the long-chain decoupling method. They can validate some statements that we have made in Section 3.3. These experiments are foundational because they proved the prerequisite conditions for the decoupling approach to work. Thus, we name them as the prerequisite experiments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 388, + 546, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 546, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 546, + 602 + ], + "type": "text", + "content": "Error Rate Distribution along a Predicted Chain. We tend to remove the edges between unreliable nodes and only connect one unreliable node to other reliable ones. In this way, we can filter out most reliable sub-chains from the original long chain. Here, we measure the reliability of each node in terms of the error rate distribution along the non-autoregressively predicted chain. The error rate distributions along the predicted chains with " + }, + { + "bbox": [ + 304, + 388, + 546, + 602 + ], + "type": "inline_equation", + "content": "T = 3/4/5/6" + }, + { + "bbox": [ + 304, + 388, + 546, + 602 + ], + "type": "text", + "content": " are illustrated in Figure 3. It is clear the reliabilities of the first and final action nodes are the highest and the model is more likely to make wrong predictions while approaching the middle of the chain. It is not hard to understand this phenomenon: a higher error rate occurs at intermediate action predictions because they are further away from the initial and goal visual observations. This empirical finding motivates us to remove the edges between unreliable intermediate actions and only connect each intermediate action to the reliable initial and final action nodes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "Reliability of Standalone Sub-chains. In addition to the reliability of each node, we further study the reliability of the sub-chains that we have decoupled from the original long chain. For a clearer explanation, we take a chain with a length of 5 as an example (e.g., " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\{a_1, a_2, a_3, a_4, a_5\\}" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "), and only study the reliability of the sub-chain " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\{a_1, a_3, a_5\\}" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": ". We consider two cases: First, we just model the relation of the whole chain and only investigate the prediction result of that sub-chain contained in the long chain. In other words, we" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10302" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 75, + 284, + 205 + ], + "blocks": [ + { + "bbox": [ + 51, + 75, + 284, + 205 + ], + "lines": [ + { + "bbox": [ + 51, + 75, + 284, + 205 + ], + "spans": [ + { + "bbox": [ + 51, + 75, + 284, + 205 + ], + "type": "image", + "image_path": "b313374b0c51add4dea043815653a9f1f91625f81a3c6bea6fcb399bfacbd008.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 209, + 287, + 233 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 287, + 233 + ], + "type": "text", + "content": "Figure 3. Error rate distributions along the non-autoregressively predicted chains with " + }, + { + "bbox": [ + 46, + 209, + 287, + 233 + ], + "type": "inline_equation", + "content": "T = 3/4/5/6" + }, + { + "bbox": [ + 46, + 209, + 287, + 233 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 264, + 285, + 378 + ], + "blocks": [ + { + "bbox": [ + 47, + 242, + 286, + 264 + ], + "lines": [ + { + "bbox": [ + 47, + 242, + 286, + 264 + ], + "spans": [ + { + "bbox": [ + 47, + 242, + 286, + 264 + ], + "type": "text", + "content": "Table 3. Reliability of standalone sub-chains vs sub-chains contained within the original long chain." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 264, + 285, + 378 + ], + "lines": [ + { + "bbox": [ + 48, + 264, + 285, + 378 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 285, + 378 + ], + "type": "table", + "html": "
Sub-chainLoss TypeSRmAccmIoU
{a1,a2,a5}Long22.5258.2871.87
Short22.6858.7372.40
{a1,a3,a5}Long13.7356.7471.78
Short16.2959.9173.07
{a1,a4,a5}Long12.5056.5871.23
Short13.2557.6171.72
", + "image_path": "549fb2295385e101a3bebf9758909886324c77b29a2afe3d250bdc7bfcdb276c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 399, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 662 + ], + "type": "text", + "content": "formulate the loss function along the whole sequence (involving " + }, + { + "bbox": [ + 46, + 399, + 287, + 662 + ], + "type": "inline_equation", + "content": "\\{a_1, a_2, a_3, a_4, a_5\\}" + }, + { + "bbox": [ + 46, + 399, + 287, + 662 + ], + "type": "text", + "content": "), and only compute the metric results over the sub-chain " + }, + { + "bbox": [ + 46, + 399, + 287, + 662 + ], + "type": "inline_equation", + "content": "\\{a_1, a_3, a_5\\}" + }, + { + "bbox": [ + 46, + 399, + 287, + 662 + ], + "type": "text", + "content": ". The second case is we only model the relation of the standalone sub-chain and check its prediction result. Equivalently, we model the loss function just involving " + }, + { + "bbox": [ + 46, + 399, + 287, + 662 + ], + "type": "inline_equation", + "content": "\\{a_1, a_3, a_5\\}" + }, + { + "bbox": [ + 46, + 399, + 287, + 662 + ], + "type": "text", + "content": ", and still computes the metric results of the sub-chain " + }, + { + "bbox": [ + 46, + 399, + 287, + 662 + ], + "type": "inline_equation", + "content": "\\{a_1, a_3, a_5\\}" + }, + { + "bbox": [ + 46, + 399, + 287, + 662 + ], + "type": "text", + "content": ". The metric results of these two cases are presented in Table 3, where we denote 'Long' as the first case and denote 'Short' as the second case due to the length difference in loss. We spot the metric results of standalone sub-chains ('Short') are always better than the ones contained within the whole action chain ('Long'). This empirical finding also happens in the cases of " + }, + { + "bbox": [ + 46, + 399, + 287, + 662 + ], + "type": "inline_equation", + "content": "T = 4" + }, + { + "bbox": [ + 46, + 399, + 287, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 399, + 287, + 662 + ], + "type": "inline_equation", + "content": "T = 6" + }, + { + "bbox": [ + 46, + 399, + 287, + 662 + ], + "type": "text", + "content": ". As we stated at the beginning of this section, such experimental results are foundational: The method of decoupling long chains can be only workable when standalone sub-chains can result in better metric results compared to the sub-chains contained within their original long chain. In this way, we experimentally prove the reliability of standalone sub-chains is better under any condition. It constructs the prerequisite condition to guarantee our decoupling method is workable." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 670, + 141, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 141, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 141, + 684 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "We present comprehensive ablation experiments to illustrate the effectiveness of different components in our" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 80, + 544, + 412 + ], + "blocks": [ + { + "bbox": [ + 342, + 69, + 506, + 80 + ], + "lines": [ + { + "bbox": [ + 342, + 69, + 506, + 80 + ], + "spans": [ + { + "bbox": [ + 342, + 69, + 506, + 80 + ], + "type": "text", + "content": "Table 4. Ablation experiments on CrossTask." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 80, + 544, + 412 + ], + "lines": [ + { + "bbox": [ + 307, + 80, + 544, + 412 + ], + "spans": [ + { + "bbox": [ + 307, + 80, + 544, + 412 + ], + "type": "table", + "html": "
SRmAccmIoU
(a) With and Without State Supervision at T = 3
Language26.2858.9871.82
None28.8561.1874.98
(b) Types of layers on the Time of Input Frames at T = 3
Average Pooling24.4056.4170.87
Linear24.4056.6971.24
Conv26.5659.6074.15
MLP28.8561.1874.98
(c) Without and With Decoupling at T = 4
w/o Decoupling12.7253.1869.13
Decoupling15.5655.6470.30
(d) Without and With Decoupling at T = 5
w/o Decoupling4.9748.7165.29
Decoupling8.5552.0569.09
(e) Without and With Decoupling at T = 6
w/o Decoupling2.6244.0862.07
Decoupling5.1247.3466.11
(f) Gamma Values of Focal Loss at T = 4
γ = 013.8454.6769.74
γ = 1.015.0854.5870.09
γ = 1.515.5655.6470.30
γ = 2.014.8054.9970.20
", + "image_path": "2b1e31ee2fb4fed44670b52f6c26facfd0a51da3a71497ed64b6914e0642e52b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 435, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 435, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 305, + 435, + 545, + 460 + ], + "type": "text", + "content": "Skip-Plan. All ablation experiments are carried out on the CrossTask dataset, illustrated in Table 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "type": "text", + "content": "With and Without State Supervision. We consider " + }, + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "type": "inline_equation", + "content": "T = 3" + }, + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "type": "text", + "content": " to compare the metric results with and without intermediate state supervision. The model with intermediate state supervision is done by adding weak language state supervision, which is identical to the one proposed by P3IV[41]. The weak language state supervision in P3IV[41] is the most effective state supervision among all the previous works. In Table 4(a), we can see the model without state supervision outperforms the model with language state supervision in all metrics. It confirms the existence of intermediate state nodes does not provide much valuable information but introduces undesired noise into the chain." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Types of Layers on the Time of Input Frames. Our model achieves significant improvements by applying a small MLP network on the time dimension of visual input frames. Here, we explore the effects of various types of layers on the time information, including average pooling, a single linear layer, 1D convolution, and our proposed MLP, as shown in Table 4(b). All the previous works overlooked the significance of initial and goal visual inputs. They used the averaging pooling operation, which performs the worst" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10303" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 289, + 149 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 289, + 149 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 289, + 149 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 289, + 149 + ], + "type": "image", + "image_path": "033532c38ca7be1c586ef29060e8be387013f0bba3c245f68fdda090beb1e2e1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 152, + 288, + 198 + ], + "lines": [ + { + "bbox": [ + 46, + 152, + 288, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 288, + 198 + ], + "type": "text", + "content": "Figure 4. Case study on CrossTask: our Skip-Plan predicts all actions correctly, whereas P3IV[41] fails on the last three predictions. The outlines of correct and incorrect predictions are coloured in green and red respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 218, + 287, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 287, + 253 + ], + "type": "text", + "content": "among all types of layers. We take " + }, + { + "bbox": [ + 46, + 218, + 287, + 253 + ], + "type": "inline_equation", + "content": "T = 3" + }, + { + "bbox": [ + 46, + 218, + 287, + 253 + ], + "type": "text", + "content": " as an example to compare different layers, and indicate the MLP network in our Skip-Plan framework performs the best." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 254, + 287, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 254, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 287, + 314 + ], + "type": "text", + "content": "Decomposition of Long Chains. We compare the performance without and with the decoupling approach at all long " + }, + { + "bbox": [ + 46, + 254, + 287, + 314 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 254, + 287, + 314 + ], + "type": "text", + "content": ", ranging from 4 to 6. Complete ablation experiments demonstrate the decoupling approach in our Skip-Plan can robustly improve all the metric results at any condition." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 315, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 287, + 361 + ], + "type": "text", + "content": "Focal Loss with Gamma Values We compare the experimental results with different gamma values of focal loss. We find the gamma value of 1.5 works best to overcome the imbalanced data distribution in the CrossTask dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 371, + 122, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 371, + 122, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 122, + 384 + ], + "type": "text", + "content": "4.5. Case Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 389, + 287, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 497 + ], + "type": "text", + "content": "Skip-Plan and the Baseline. We compare our Skip-Plan with the previous state-of-the-art P3IV[41] in real cases, illustrated in Figure 4. Given the initial and goal visual observations, P3IV[41] only predicts the first action correctly and fails to get the other actions at " + }, + { + "bbox": [ + 46, + 389, + 287, + 497 + ], + "type": "inline_equation", + "content": "T = 4" + }, + { + "bbox": [ + 46, + 389, + 287, + 497 + ], + "type": "text", + "content": ". In contrast, Skip-Plan extracts the initial and goal visual information thoroughly, so Skip-Plan predicts the final action correctly. Benefiting from the decoupling method, our Skip-Plan makes the correct predictions on the second and third actions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "content": "Effects of Different Components in Skip-Plan. We investigate the effects of various components in Skip-Plan on model performance by real samples, which are illustrated in Figure 5. Here, the components for comparisons include the types of layers on the time dimension of visual input frames, state supervision, and the decoupling approach. For the types of layers on the time dimension of visual input frames, we compare the average pooling layer with our shared MLP network, both without state supervision and decoupling. As shown in Figure 5(b), the average pooling layer makes wrong predictions on all actions. However, the model with the shared MLP network predicts the first and last actions correctly, as illustrated in Figure 5(c). This is because the MLP network can extract the time information from the initial and goal visual observations more thoroughly than the average pooling layer. We further simulate the shared MLP network without decoupling but with language state supervision[41] in Figure 5(d). Comparing" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 305, + 68, + 547, + 198 + ], + "blocks": [ + { + "bbox": [ + 305, + 68, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 305, + 68, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 305, + 68, + 547, + 198 + ], + "type": "image", + "image_path": "87ab1d8252ae3fbd6c85244c8de2f3b09d05c88db3679ec5a57709b44f822e21.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 201, + 547, + 312 + ], + "lines": [ + { + "bbox": [ + 304, + 201, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 547, + 312 + ], + "type": "text", + "content": "Figure 5. Case study on Skip-Plan components at " + }, + { + "bbox": [ + 304, + 201, + 547, + 312 + ], + "type": "inline_equation", + "content": "T = 4" + }, + { + "bbox": [ + 304, + 201, + 547, + 312 + ], + "type": "text", + "content": ": (a) ground truth; (b) average pooling: average pooling layer on the time dimension of visual input frames without decoupling and state supervision; (c) MLP: shared MLP network on the time dimension of visual input frames without decoupling and state supervision; (d) MLP with state supervision: shared MLP network with state supervision but without decoupling; (e) MLP with decoupling: shared MLP network with decoupling but without state supervision. The outlines of correct and incorrect action predictions are coloured in green and red respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 335, + 547, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 547, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 547, + 419 + ], + "type": "text", + "content": "Figure 5(d) with Figure 5(c), we notice the addition of state supervision cannot correct the predictions of the second and the third actions, because the state supervision does not reduce the compound error. In contrast, the shared MLP network without state supervision but with decoupling can lower the compound error effectively, resulting in correct predictions of all actions, as demonstrated in Figure 5(e)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 437, + 378, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 437, + 378, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 378, + 449 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 460, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 639 + ], + "type": "text", + "content": "In this paper, we propose Skip-Plan, a condensed action space learning method for the procedure planning task. Our Skip-Plan solves the procedure planning problem mathematically in two novel ways. First, Skip-Plan skips the intermediate state supervision and only focuses on action relations. Second, Skip-Plan decomposes long action chains into multiple reliable sub-chains by skipping most uncertain intermediate actions. In these ways, our Skip-Plan model explores all sorts of sub-relations within an action chain and generates actions non-adjacently only in the condensed action space. Extensive experiments show the superior performance of our Skip-Plan method on two instructional video datasets. Skip-Plan has the potential to inspire the research field of procedure planning, since most of them heavily rely on state observations and adjacent actions at all times." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0114903 and in part by the National Natural Science Foundation of China under Grant 62125603 and Grant 62206153, CAAI-Huawei MindSpore Open Fund." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10304" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 92, + 264, + 103 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 264, + 103 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 264, + 103 + ], + "type": "text", + "content": "[1] Mindspore. .https://www.mindspore.cn/.5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 106, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 106, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 106, + 287, + 148 + ], + "type": "text", + "content": "[2] Armenii Ashukha, Alexander Lyzhov, Dmitry Molchanov, and Dmitry Vetrov. Pitfalls of in-domain uncertainty estimation and ensembling in deep learning. arXiv preprint arXiv:2002.06470, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 150, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 150, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 288, + 195 + ], + "type": "text", + "content": "[3] Anil Batra, Shreyank Gowda, Laura Sevilla-Lara, and Frank Keller. A closer look at temporal ordering in the segmentation of instructional videos. arXiv preprint arXiv:2209.15501, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 197, + 287, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 197, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 53, + 197, + 287, + 240 + ], + "type": "text", + "content": "[4] Jing Bi, Jiebo Luo, and Chenliang Xu. Procedure planning in instructional videos via contextual modeling and model-based policy learning. In ICCV, pages 15611-15620, 2021. 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 243, + 288, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 243, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 53, + 243, + 288, + 275 + ], + "type": "text", + "content": "[5] Nicolas Biard, Salomé Cojean, and Eric Jamet. Effects of segmentation and pacing on procedural learning by video. Comput. Hum. Behav., 89:411-417, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 277, + 287, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 277, + 287, + 321 + ], + "spans": [ + { + "bbox": [ + 53, + 277, + 287, + 321 + ], + "type": "text", + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 323, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 323, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 323, + 287, + 367 + ], + "type": "text", + "content": "[7] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In ECCV, pages 334–350, 2020. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 403 + ], + "type": "text", + "content": "[8] Chekuri Choudary and Tiecheng Liu. Summarization of visual content in instructional videos. IEEE Trans. Multimed., 9(7):1443-1455, 2007. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 404, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 287, + 448 + ], + "type": "text", + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 450, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 450, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 450, + 287, + 483 + ], + "type": "text", + "content": "[10] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 485, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 287, + 529 + ], + "type": "text", + "content": "[11] Mikita Dvornik, Isma Hadji, Konstantinos G Derpanis, Animesh Garg, and Allan Jepson. Drop-dtw: Aligning common signal between sequences while dropping outliers. NeuRIPS, 34:13782-13793, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 531, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 531, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 287, + 586 + ], + "type": "text", + "content": "[12] Jakob Gawlikowski, Cedrique Rovile Njieutcheu Tassi, Mohsin Ali, Jongseok Lee, Matthias Humt, Jianxiang Feng, Anna Kruspe, Rudolph Trielbel, Peter Jung, Ribana Roscher, et al. A survey of uncertainty in deep neural networks. arXiv preprint arXiv:2107.03342, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 588, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 287, + 632 + ], + "type": "text", + "content": "[13] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, Chiho Choi, and Behzad Dariush. Weakly-supervised online action segmentation in multi-view instructional videos. In CVPR, pages 13780-13790, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "text", + "content": "[14] Reza Ghoddoosian, Saif Sayed, and Vassilis Athitsos. Hierarchical modeling for task recognition and action segmentation in weakly-labeled instructional videos. In WACV, pages 1922-1932, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[15] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. arXiv preprint arXiv:1610.02136, 2016. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Gabriel Huang, Bo Pang, Zhenhai Zhu, Clara Rivera, and Radu Soricut. Multimodal pretraining for dense video captioning. arXiv preprint arXiv:2011.11760, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[17] Eyke Hüllermeier and Willem Waegeman. Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods. Mach. Learn, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 141, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 163 + ], + "type": "text", + "content": "[18] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In CVPRW, pages 958-959, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "type": "text", + "content": "[19] Lei Ji, Chenfei Wu, Daisy Zhou, Kun Yan, Edward Cui, Xilin Chen, and Nan Duan. Learning temporal video procedure segmentation from an automatically collected large dataset. In WACV, pages 1506-1515, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 209, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 240 + ], + "type": "text", + "content": "[20] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? NeurIPS, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "text", + "content": "[21] Colin Lea, Michael D Flynn, Rene Vidal, Austin Reiter, and Gregory D Hager. Temporal convolutional networks for action segmentation and detection. In CVPR, pages 156-165, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 287, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 331 + ], + "type": "text", + "content": "[22] Muheng Li, Lei Chen, Yueqi Duan, Zhilan Hu, Jianjiang Feng, Jie Zhou, and Jiwen Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In CVPR, pages 19880-19889, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 333, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 333, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 333, + 545, + 365 + ], + "type": "text", + "content": "[23] Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. arXiv preprint arXiv:1706.02690, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 366, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 366, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 366, + 545, + 399 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 400, + 545, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 400, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 307, + 400, + 545, + 444 + ], + "type": "text", + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "type": "text", + "content": "[26] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 501, + 545, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 523 + ], + "type": "text", + "content": "[27] Andrey Malinin and Mark Gales. Predictive uncertainty estimation via prior networks. NeurIPS, 2018. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 524, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 524, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 524, + 545, + 567 + ], + "type": "text", + "content": "[28] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In CVPR, pages 9879-9889, 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 569, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 569, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 569, + 545, + 622 + ], + "type": "text", + "content": "[29] Medhini Narasimhan, Arsha Nagrani, Chen Sun, Michael Rubinstein, Trevor Darrell, Anna Rohrbach, and Cordelia Schmid. Tl; dw? summarizing instructional videos with task relevance and cross-modal saliency. In ECCV, pages 540-557, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "text", + "content": "[30] Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. NeurIPS, 32, 2019. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[31] AJ Piergiovanni, Anelia Angelova, Michael S Ryoo, and Irfan Essa. Unsupervised action segmentation for instructional videos. arXiv preprint arXiv:2106.03738, 2021. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10305" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 622 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[32] Alireza Shafaei, Mark Schmidt, and James J Little. A less biased evaluation of out-of-distribution sample detectors. arXiv preprint arXiv:1809.04729, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "type": "text", + "content": "[33] Anshul Shah, Benjamin Lundell, Harpreet Sawhney, and Rama Chellappa. Steps: Self-supervised key step extraction from unlabeled procedural videos. arXiv preprint arXiv:2301.00794, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 287, + 195 + ], + "type": "text", + "content": "[34] Botian Shi, Lei Ji, Yaobo Liang, Nan Duan, Peng Chen, Zhendong Niu, and Ming Zhou. Dense procedure captioning in narrated instructional videos. In ACL, pages 6382-6391, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 287, + 239 + ], + "type": "text", + "content": "[35] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In ICCV, pages 7464-7473, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 241, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 287, + 285 + ], + "type": "text", + "content": "[36] Jiankai Sun, De-An Huang, Bo Lu, Yun-Hui Liu, Bolei Zhou, and Animesh Garg. Plate: Visually-grounded planning with transformers in procedural tasks. IEEE Robot. Autom. Lett., 7(2):4924-4930, 2022. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 286, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 286, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 49, + 286, + 287, + 330 + ], + "type": "text", + "content": "[37] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In CVPR, pages 1207-1216, 2019. 1, 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 331, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 331, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 49, + 331, + 287, + 364 + ], + "type": "text", + "content": "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 365, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 365, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 49, + 365, + 287, + 418 + ], + "type": "text", + "content": "[39] Mengmeng Xu, Juan-Manuel Pérez-Rúa, Victor Escorcia, Brais Martinez, Xiatian Zhu, Li Zhang, Bernard Ghanem, and Tao Xiang. Boundary-sensitive pre-training for temporal localization in videos. In ICCV, pages 7220-7230, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 421, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 421, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 421, + 287, + 453 + ], + "type": "text", + "content": "[40] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Temporal query networks for fine-grained video understanding. In CVPR, pages 4486-4496, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 455, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 455, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 49, + 455, + 287, + 509 + ], + "type": "text", + "content": "[41] He Zhao, Isma Hadji, Nikita Dvornik, Konstantinos G Derpanis, Richard P Wildes, and Allan D Jepson. P3iv: Probabilistic procedure planning from instructional videos with weak supervision. In CVPR, pages 2938-2948, 2022. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 511, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 511, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 49, + 511, + 287, + 543 + ], + "type": "text", + "content": "[42] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, pages 7590-7598, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 544, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 544, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 49, + 544, + 287, + 576 + ], + "type": "text", + "content": "[43] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In CVPR, pages 8746-8755, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 578, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 578, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 49, + 578, + 287, + 622 + ], + "type": "text", + "content": "[44] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramadan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In CVPR, pages 3537-3545, 2019. 1, 2, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10306" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SlaBins_ Fisheye Depth Estimation using Slanted Bins on Road Environments/2f72858d-bf2e-45dd-b589-9c5d6578e810_content_list.json b/2023/SlaBins_ Fisheye Depth Estimation using Slanted Bins on Road Environments/2f72858d-bf2e-45dd-b589-9c5d6578e810_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dae6978e10f08b3e844c1341e63653fb0c6f0eee --- /dev/null +++ b/2023/SlaBins_ Fisheye Depth Estimation using Slanted Bins on Road Environments/2f72858d-bf2e-45dd-b589-9c5d6578e810_content_list.json @@ -0,0 +1,1821 @@ +[ + { + "type": "text", + "text": "SlaBins: Fisheye Depth Estimation using Slanted Bins on Road Environments", + "text_level": 1, + "bbox": [ + 94, + 161, + 875, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jongsung Lee $^{1}$ , Gyeongsu Cho $^{1*}$ , Jeongin Park $^{1*}$ , Kyongjun Kim $^{1*}$ , Seongoh Lee $^{1*}$ , Jung-Hee Kim $^{2}$ , Seong-Gyun Jeong $^{2}$ , Kyungdon Joo $^{1\\dagger}$", + "bbox": [ + 157, + 206, + 815, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Artificial Intelligence Graduate School, UNIST, 242dot Inc.", + "bbox": [ + 243, + 242, + 724, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{syniez, threedv, jeonginpark, kimkj38, solee, kyungdon}@unist.ac.kr", + "bbox": [ + 181, + 262, + 783, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{junghee.kim, seonggyun.jeong}@42dot.ai", + "bbox": [ + 314, + 277, + 657, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 320, + 312, + 334 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although 3D perception for autonomous vehicles has focused on frontal-view information, more than half of fatal accidents occur due to side impacts in practice (e.g., T-bone crash). Motivated by this fact, we investigate the problem of side-view depth estimation, especially for monocular fisheye cameras, which provide wide FoV information. However, since fisheye cameras head road areas, it observes road areas mostly and results in severe distortion on object areas, such as vehicles or pedestrians. To alleviate these issues, we propose a new fisheye depth estimation network, SlaBins, that infers an accurate and dense depth map based on a geometric property of road environments; most objects are standing (i.e., orthogonal) on the road environments. Concretely, we introduce a slanted multi-cylindrical image (MCI) representation, which allows us to describe a distance as a radius to a cylindrical layer orthogonal to the ground regardless of the camera viewing direction. Based on the slanted MCI, we estimate a set of adaptive bins and a per-pixel probability map for depth estimation. Then by combining it with the estimated slanted angle of viewing direction, we directly infer a dense and accurate depth map for fisheye cameras. Experiments demonstrate that SlaBins outperforms the state-of-the-art methods in both qualitative and quantitative evaluation on the SynWoodScape and KITTI-360 depth datasets. For more information, you can visit our project page https://syniez.github.io/SlaBins/.", + "bbox": [ + 75, + 350, + 473, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 751, + 208, + 766 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the advent of autonomous driving, the importance of 3D perception keeps growing for safety [20, 25, 44]. For example, ADAS systems pervading our daily driving already, such as lane-keeping aid and forward collision", + "bbox": [ + 75, + 775, + 468, + 832 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d7972720d786bf5df51d9fa598f8e11245287d358ff6e72fae4804fcebe9ef92.jpg", + "image_caption": [ + "RGB" + ], + "image_footnote": [], + "bbox": [ + 537, + 326, + 615, + 366 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/37b580a71bf49cf0fecfaa948c403f57b7cba1bbeb5510e580a9b0bc794b9dd2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 368, + 613, + 408 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d0b8520def059b3be9dcfe75b01395c6a8eaeaf52ce61e53aa75d7cb318791cc.jpg", + "image_caption": [ + "Figure 1: SlaBins on the SynWoodScape dataset [34]. SlaBins predicts dense and accurate depth maps regardless of the viewing directions using the slanted MCI representation." + ], + "image_footnote": [], + "bbox": [ + 537, + 410, + 613, + 454 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8d8f184aef1f234e2fe02f96a914794b6d7f4134d100206e1769b7b1574c0dea.jpg", + "image_caption": [ + "Slanted MCI representation" + ], + "image_footnote": [], + "bbox": [ + 645, + 338, + 752, + 380 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7f96553eab508578605a2bb78c6bb6d48249c691d76b811d8ed157b55c3244a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 618, + 388, + 773, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1dba10f1d3bc1670f20779959c64f817370d99ce9fc28790171071aafd506e23.jpg", + "image_caption": [ + "Depth" + ], + "image_footnote": [], + "bbox": [ + 778, + 327, + 875, + 366 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1916eeae437689dd6e45be8389dd44a8a7744130cde9b6863669a882614c7778.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 778, + 368, + 875, + 409 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1197552c0da8cc880c7bd38bfff1fe4a1e72ccb4d59ea085403bfc4b563b74b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 778, + 410, + 877, + 454 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "avoidance assist, depend on various sensor modalities to observe 3D information [43, 5]. Here, one interesting point is that various sensor modalities for 3D perception, including many datasets related to autonomous driving, are focusing on the frontal view predominantly [43, 38, 2, 6]. We can easily deduce the reason by our common sense that it is important to keep an eye on the front while driving. However, according to the National Transportation Highway Safety Administration [1], $51\\%$ of fatal car accidents correspond to T-bone crashes in which one car hits the side of another, which is the second highest type of accident. This statistic supports that side-view 3D perception is important to assist drivers because they are hard to pay attention to the side view due to the physical characteristics of the people looking ahead and their position characteristics.", + "bbox": [ + 497, + 505, + 892, + 712 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motivated by this fact, 3D perception for side-view, especially fisheye-based depth estimation, has started to gain the spotlight [31, 16, 41]. Fisheye cameras have several advantages for 3D perception in road scenes. Specifically, fisheye cameras provide broader FoV information (typically greater than $180^{\\circ}$ ) in a single image, which allows us to observe road information (e.g., road, vehicles, and pedestrians) at a low cost. In addition, we can directly utilize perspective camera 3D perception approaches because mathematical mutual conversion between two camera models is possible. Thus, if we can obtain side-view depth informa", + "bbox": [ + 497, + 714, + 892, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 129, + 51, + 200, + 84 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 253, + 48, + 789, + 86 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution (alphabet order).", + "bbox": [ + 95, + 842, + 292, + 853 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 95, + 853, + 221, + 865 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8765", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion using a single fisheye camera, we can efficiently handle various vision tasks for autonomous driving.", + "bbox": [ + 75, + 125, + 468, + 153 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, despite the advantages of fisheye cameras, monocular fisheye camera-based side-view depth estimation is a challenging problem due to significant distortion. In addition to inherent issues of monocular depth estimation, such as scale ambiguity [8], severe distortion of fisheye cameras makes depth estimation more difficult by distorting scene objects (e.g., close objects look closer and distant objects look more distant abnormally). Particularly, this distortion problem could be more severe in road environments because fisheye cameras for ADAS systems generally look to the road areas, and vital objects such as vehicles and pedestrians are located at the edge of the image. That is, most regions of the image correspond to the road areas and objects contain very few pixels, so depth estimation becomes more challenging.", + "bbox": [ + 75, + 155, + 468, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To alleviate this problem, we pay attention to the geometric property of road environments; most objects are standing on the road environments. Based on this geometric characteristic, we can easily imagine the visible region of each object as a planar segment orthogonal to the road region. This approximation is connected to a multi-plane image (MPI) concept [36] that represents a given scene as the composition of multi-layered images according to depth. Note that one significant difference is that a fisheye camera looks down to the road region in our case (i.e., camera viewing direction is not parallel to the ground), unlike a camera points to a given scene in the general MPI concept. Thus, a naive MPI concept is not directly applicable.", + "bbox": [ + 75, + 363, + 468, + 543 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a new fisheye depth estimation framework, SlaBins, that exploits the geometric property of the road environments (see Fig. 1). To be specific, we adopt a multi-cylindrical image (MCI) representation; especially, we introduce a slanted MCI of which the cylindrical layer is defined as orthogonal to the road ground regardless of camera viewing direction. Based on the slanted MCI, we estimate adaptive bins and the per-pixel probability of the slanted cylindrical layers, which allows us to implicitly preserve the geometric property of the road environments and improve the depth quality near the boundary between objects and the ground region. In addition, we independently estimate the slanted angle of the camera viewing direction w.r.t. the road ground. Then, by combining this angle with the estimated depth information in the slanted MCI domain, we can directly compute a dense and accurate depth map for fisheye cameras. We validate the proposed SlaBins framework on the SynWoodScape dataset [34] and KITTI-360 depth dataset (a modified dataset of KITTI", + "bbox": [ + 75, + 543, + 468, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "360 [24] for our scenario), where SlaBins outperforms the baseline methods [3, 20, 21] and shows potential as scene representations for downstream tasks, e.g., segmentation.", + "bbox": [ + 499, + 125, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In a nutshell, our contributions are as follows:", + "bbox": [ + 517, + 168, + 823, + 180 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new fisheye depth estimation framework, SlaBins, which infers dense depth based on the geometric property of road environments.", + "- We newly introduce a slanted MCI representation as an intermediate road scene representation, which allows us to describe a given scene regardless of viewing direction and improve the depth quality on the boundary areas between objects and the road.", + "- By decoupling the slanted angle of an input image, we force the proposed model to learn the slanted MCI depth information and then seamlessly compute a fisheye depth map in a slanted-aware regression way." + ], + "bbox": [ + 499, + 186, + 890, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 499, + 383, + 640, + 397 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Monocular depth estimation. With the advent of breakthroughs in deep neural networks, such as CNN and transformer architecture, monocular depth estimation has shown impressive performance [15, 33, 21, 32, 23, 18, 40, 11, 9, 12, 29, 19, 16, 17, 20]. As one of the recent trends, several methods formulate a depth estimation task as a combination of classification and regression problems, which enables training the network more easily and shows stable performance [14, 3, 4]. For this formulation, they commonly use an MPI according to depth, which is popularly used in novel-view synthesis tasks [22, 39, 13]. The MPI notation represents depth as quantized depth values w.r.t. the uniformly quantized planes and residuals from mapped planes.", + "bbox": [ + 497, + 407, + 890, + 586 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To improve the uniformly quantized planes by uniform bins, recent research works [3, 4] present a network that predicts adaptive bins depending on the input scene and local regions in the image. AdaBins [3] computes adaptive bin-centers and per-pixel probabilities relying on each input image. Inspired by AdaBins, LocalBins [4] improves depth quality by considering both the global histogram and local-region histograms of each input image. Nevertheless, the predicted depth from the above models also depends on the camera's pose because they predict depth parallel with the camera's principal axis. This constraint makes it hard to adopt the previous models into fisheye depth estimation, which has a large FoV and uses the Euclidean distance instead of depth. In this work, we introduce and exploit a slanted MCI representation that is invariant to the viewing angle of fisheye cameras w.r.t. road environments.", + "bbox": [ + 497, + 586, + 890, + 807 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fisheye depth estimation. Despite increasing demand for fisheye cameras in the industry, like ADAS systems, there are only a few previous research works in fisheye depth estimation fields [20, 17, 19, 16, 40]. Kumar et al. [19] propose", + "bbox": [ + 497, + 809, + 890, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Basically, depth in the fisheye model is defined as the Euclidean distance, which is connected to a multi-spherical image (MSI). In this work, We adopt an MCI representation that combines the geometrical suitability of MPI and the fisheye model fitness of MSI.", + "bbox": [ + 75, + 820, + 468, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8766", + "bbox": [ + 468, + 887, + 502, + 897 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5edfc6aa98a4cee7db2ed3927bd6007ed4962c651216163c8bef773ce430c348.jpg", + "image_caption": [ + "Figure 2: Overview of the proposed SlaBins. (a) Given an input fisheye RGB image, SlaBins framework first extracts features by the encoder-decoder block. The extracted features are used as input to the slanted angle prediction module and the SlaBins module to get a depth image. (b) Illustration of the coordinates and MCI representations defined in our model. (c) Illustration of the slanted angle predictor." + ], + "image_footnote": [], + "bbox": [ + 97, + 119, + 868, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a camera-aware depth estimation network to handle the severe distortion of fisheye cameras. They encode the camera's intrinsic parameters as a tensor named camera geometry tensor and use it in the training step for camera-aware depth estimation in fisheye images. With camera geometry tensor, they propose a multi-task network for fisheye cameras, called OmniDet [20], that simultaneously estimates depth estimation, motion, semantic segmentation, and object detection for synergy. However, since OmniDet relies on the partial region in the whole image, they may miss information on image edges. Using a whole image, Yan et al. [40] distill depth ordering information from a teacher network trained on the large-scale dataset to train in fisheye images, but FoV they cover is not large (e.g., vFoV=66°). In contrast to previous approaches, we propose a fisheye depth estimation network that predicts fisheye image depth without any information loss such as crop, and narrow FoV.", + "bbox": [ + 75, + 330, + 468, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Geometric cues for depth estimation. One way to improve fisheye depth estimation is to exploit geometric cues between the camera and the world. Liu et al. [26] use a ground plane to get depth priors using camera elevation and camera parameters. Moreover, Mahjourian et al. [29] predict camera ego-motion and compute loss on both the 3D point cloud domain and the 2D image domain. In addition to driving scenes, several works use plane coefficients to give consistency on plane regions in depth estimation [30, 27]. Patil et al. [30] estimate plane coefficients instead of depth and then combine plane coefficients into depth to use plane constraints (e.g., co-planarity, plane normals) in 3D space. Inspired by the previous works [3, 30], we design a network that estimates a dense and accurate depth map for fisheye cameras based on the slanted MCI representation using some geometric constraints (i.e., orthogonality between ground and object).", + "bbox": [ + 75, + 568, + 468, + 802 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Framework", + "text_level": 1, + "bbox": [ + 77, + 815, + 284, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the geometric property of the road environments, we propose a new fisheye depth estimation frame", + "bbox": [ + 77, + 838, + 468, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "work (see Fig. 2 for the overview). Specifically, we introduce a slanted MCI representation, which allows us to describe a multi-layer cylindrical image orthogonal to the road ground regardless of the viewing direction of the camera. Based on the slanted MCI representation, our SlaBins module estimates the adaptive bin widths and per-pixel probability map in the orthogonal coordinate, which provides depth information invariant to the camera viewing direction. We then combine the estimated depth information with the slanted angle estimated by the slanted angle prediction module. Through this process, we can directly estimate a dense depth map for fisheye cameras. Our approach is built upon the previous AdaBins framework [3]. So, we call our method SlaBins.", + "bbox": [ + 497, + 328, + 890, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Slanted MCI Representation", + "text_level": 1, + "bbox": [ + 499, + 534, + 756, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As intermediate scene representation, we use multi-layer images for fisheye depth estimation. Here, we briefly compare two multi-layer image representations for fisheye images: multi-spherical image (MSI) and multi-cylindrical image (MCI). We then present a slanted MCI representation, which is a suitable representation of our problem.", + "bbox": [ + 497, + 557, + 890, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MSI vs. MCI. By nature of fisheye cameras covering over $180^{\\circ}$ , it is common to use the Euclidean distance as its depth; that is, a layer with the same depth value forms a sphere in the fisheye domain. Thus, we can represent a multi-layer image for fisheye cameras as an MSI in the spherical coordinate, which is a naive representation. However, this naive MSI representation has inherent limitations for fisheye depth estimation. We observed that such a representation can result in ambiguous depth boundaries between the ground and object regions, as shown in Fig. 3a. It may neutralize the benefit of multi-layer images by separating an object into two different layers, for example.", + "bbox": [ + 497, + 643, + 890, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the MCI representation, a cylindrical layer encodes the same Euclidean distance w.r.t. the y-axis. In other words, we can represent an object, such as a pedestrian on the road, in a single cylindrical layer (see Fig. 3b). Further", + "bbox": [ + 497, + 809, + 890, + 866 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8767", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ea4262433bf086d7a8b2ac0379dc44babe6a81a7aeb9e88c04f5fa8d519bb004.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 104, + 127, + 189, + 226 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dead83876114f702840b50d1018e067c973d44ec28768124628d0242097c6da9.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 200, + 135, + 285, + 226 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6d5259420a93cc5f1427c3e76bedce9f962d1879a1890b65f3edee92f392fd9b.jpg", + "image_caption": [ + "(c)", + "Figure 3: MSI vs. MCI representation. Illustration of an example scene and zoomed results with (a) MSI and (b) MCI, where we can observe more accurate depth in the slanted MCI representation near the boundary between pedestrians and the ground. (c) Relationship between the original cylindrical layer $C_r^o$ (gray-colored cylinder) in the world coordinate (arrows) and the slanted cylindrical layer $C_{r_\\perp}^s$ (orange-colored cylinders) in the orthogonal coordinate (dashed arrows)." + ], + "image_footnote": [], + "bbox": [ + 302, + 123, + 442, + 225 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "more, we can directly exploit the geometric properties of the road environment: most objects are standing (i.e., orthogonal) on the road environment. For this reason, we adopt the MCI as our intermediate scene representation.", + "bbox": [ + 75, + 357, + 468, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Slanted MCI representation. Let $\\mathbf{X}$ be a 3D point in the world coordinate. Then, we can represent $\\mathbf{X}$ according to the coordinate systems:", + "bbox": [ + 75, + 414, + 468, + 453 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} = \\left[ \\begin{array}{l} X \\\\ Y \\\\ Z \\end{array} \\right] = \\rho \\left[ \\begin{array}{c} \\hat {X} \\\\ \\hat {Y} \\\\ \\hat {Z} \\end{array} \\right] = \\rho \\left[ \\begin{array}{c} \\sin \\theta \\cos \\phi \\\\ \\sin \\theta \\sin \\phi \\\\ \\cos \\theta \\end{array} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 144, + 454, + 468, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{X}} = [\\hat{X},\\hat{Y},\\hat{Z}]^{\\top}$ is the normalized unit vector (i.e., ray direction), $\\rho$ indicates the Euclidean distance (i.e., depth in the fisheye domain), and $\\theta$ and $\\phi$ denote the elevation and azimuth angles, respectively.", + "bbox": [ + 75, + 505, + 468, + 562 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the cylindrical representation, we can form a cylindrical layer according to a radius $r$ as:", + "bbox": [ + 75, + 562, + 468, + 591 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {r} ^ {o} = \\left\\{\\mathbf {X} \\mid X ^ {2} + Z ^ {2} = r ^ {2}; \\mathbf {X} \\in \\mathbb {R} ^ {3} \\right\\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 144, + 595, + 468, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We call $C_r^o$ the original cylindrical layer at $r$ . This original MCI is inherently suitable for road environments and shows less distortion than MSI on ambiguous depth boundaries. However, depending on the principal axis of cameras (i.e., viewing direction) w.r.t. the road environments, the original MCI may hinder layered scene representation like the MSI.", + "bbox": [ + 75, + 619, + 468, + 701 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To alleviate this issue, we consider the 1D angle $\\alpha$ between the principal axis and the road environments. Using $\\alpha$ , we can implicitly describe the slanted MCI representation that is orthogonal to the road areas (see Fig. 3c). To this end, let $\\mathbf{X}_{\\perp}$ be a 3D point in the orthogonal coordinate, of which the XZ plane is parallel to the ground. Similarly to Eq. (2), we can form a cylindrical layer according to a radius $r_{\\perp}$ in the orthogonal coordinate as:", + "bbox": [ + 75, + 701, + 468, + 812 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {r _ {\\perp}} ^ {s} = \\left\\{\\mathbf {X} _ {\\perp} \\mid X _ {\\perp} ^ {2} + Z _ {\\perp} ^ {2} = r _ {\\perp} ^ {2}; \\mathbf {X} _ {\\perp} \\in \\mathbb {R} ^ {3} \\right\\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 126, + 817, + 468, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which corresponds to a slanted cylindrical layer in the world (camera) coordinate (see Fig. 2b for coordinate systems).", + "bbox": [ + 75, + 839, + 468, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By $\\alpha$ , we can derive the relationship between $\\mathbf{X}_{\\perp}$ in the orthogonal coordinate and $\\mathbf{X}$ in the world coordinate by:", + "bbox": [ + 497, + 125, + 890, + 152 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {\\perp} = \\left[ \\begin{array}{l} X _ {\\perp} \\\\ Y _ {\\perp} \\\\ Z _ {\\perp} \\end{array} \\right] = \\left[ \\begin{array}{c} X \\\\ Y \\cos \\alpha + Z \\sin \\alpha \\\\ - Y \\sin \\alpha + Z \\cos \\alpha \\end{array} \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 564, + 155, + 890, + 195 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we can deduce the following relationship between a 3D point $\\mathbf{X}_{\\perp}$ on $C_{r_{\\perp}}^{s}$ and its spherical representation in the camera coordinate by using Eqs. (3) and (4):", + "bbox": [ + 497, + 200, + 890, + 242 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} r _ {\\perp} ^ {2} = X _ {\\perp} ^ {2} + Z _ {\\perp} ^ {2} = X ^ {2} + (- Y \\sin \\alpha + Z \\cos \\alpha) ^ {2}, \\\\ = \\rho^ {2} \\left(\\hat {X} ^ {2} + (- \\hat {Y} \\sin \\alpha + \\hat {Z} \\cos \\alpha) ^ {2}\\right). \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 244, + 890, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, we can convert the relationship in Eq. (5) to a function $f$ as:", + "bbox": [ + 497, + 294, + 890, + 319 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf (\\tilde {\\mathbf {X}}, r _ {\\perp}, \\alpha) = \\rho , \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 636, + 317, + 890, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $r_{\\perp}$ denotes a radius in the orthogonal coordinate. The function $f$ represents that if we know slanted angle $\\alpha$ and a radius $r_{\\perp}$ at a certain pixel location on the fisheye domain, we can directly compute the corresponding depth value $\\rho$ . A more detailed explanation for Eq. (6) is available in the supplementary material. Using this relation, we implicitly model the slanted MCI representation inside the network as scene representation.", + "bbox": [ + 497, + 334, + 890, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. SlaBins framework", + "text_level": 1, + "bbox": [ + 499, + 452, + 684, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the slanted MCI representation (cf. Sec. 3.1), the proposed SlaBins framework estimates an accurate and dense depth map. SlaBins consists of three modules: Encoder-decoder block, slanted angle predictor, and SlaBins module, as shown in Fig. 2.", + "bbox": [ + 497, + 472, + 890, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concretely, given the input fisheye image, the encoder-decoder block extracts a feature map containing useful information for depth prediction. Using the extracted feature map, the SlaBins module estimates several key information, such as adaptive bin-centers and per-pixel probability map in the slanted MCI representation (i.e., the orthogonal coordinate), using a transformer architecture. It should be worth noticing that by virtue of this SlaBins module, we can implicitly preserve the geometric relationship of objects on the road environments. Then, by combining with the slanted angle estimated by the slanted angle prediction module, we can directly regress a fisheye depth map.", + "bbox": [ + 497, + 542, + 890, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Encoder-decoder block. Given input fisheye image $x_{i} \\in \\mathbb{R}^{H \\times W \\times 3}$ , where $H$ and $W$ denote the height and the width of the input, our encoder-decoder block first extract features. We use the pre-trained EfficientNet B5 [37] as the backbone for the encoder and standard feature upsampling layers for the decoder, which returns decoded features $x_{d} \\in \\mathbb{R}^{h \\times w \\times C_{d}}$ , where $h = H / 2$ , $w = W / 2$ , and $C_{d}$ indicates the feature dimension.", + "bbox": [ + 497, + 710, + 890, + 820 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Slanted angle prediction module. We design compact MLP-based classification and regression modules that predict $\\alpha$ from $x_{d}$ in a two-step. The classifier first classifies", + "bbox": [ + 497, + 824, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8768", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f9ffadc1450b2b43272998450f0ed9d8f7682ceadba8466503b64bf7bc2dd07f.jpg", + "image_caption": [ + "Figure 4: The details of the SlaBins module. Our SlaBins module consists of two components: One is the mini-ViT block that extracts the range-attention map and bin widths, which are invariant to the slanted angle $\\alpha$ from the decoded feature. The other is the SlaBins-aware regression. This module receives the outputs of mini-ViT and the slanted angle $\\alpha$ as inputs and directly converts the orthogonally estimated depth $r_{\\perp}$ for an arbitrary 3D point $\\mathbf{X}$ into the depth $\\rho$ of the spherical coordinate." + ], + "image_footnote": [], + "bbox": [ + 85, + 123, + 455, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\alpha$ in units of $5^{\\circ}$ , and the following regressor estimates the offset in the range of $(-2.5^{\\circ}, 2.5^{\\circ})$ . This two-step angle prediction enables us to estimate an accurate viewing direction angle and train more efficiently (see Fig. 2c).", + "bbox": [ + 75, + 480, + 468, + 538 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, $x_{d}$ is downsampled to $1/4$ size with two $3 \\times 3$ convolution layers and average pooling. Downsampled features are flattened (we denote the flattened feature as $x_{f}$ ) and $x_{f}$ is used as input of classifier and regressor, which are composed of 3 FC layers followed by softmax and 4 FC layers, respectively. The output of the classifier is a 19-dimensional vector and that of the regressor is a normalized angle offset in the range of $(-1^{\\circ}, 1^{\\circ})$ .", + "bbox": [ + 75, + 539, + 468, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SlaBins module. Given decoded features $x_{d}$ , the proposed SlaBins module first learns intermediate depth information, such as bin widths and probability maps in the slanted MCI representation, which are invariant to the viewing direction (i.e., slanted angle $\\alpha$ ), as shown in Fig. 4. To this end, we adopt mini-ViT (mViT) architecture in [3].", + "bbox": [ + 75, + 656, + 468, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The decoded features $x_{d}$ are fed into the mViT with learnable positional encodings. To make a fixed-size patch-embedded input, we embed the decoded features to the $C_p$ dimensional flattened tensor $x_{p}\\in \\mathbb{R}^{S\\times C_{p}}$ using patch embedding before feeding into the mViT (here, $S$ indicates $w\\times h / p^2$ and $p$ denotes the size of patch and stride). Then, the sum of flattened tensor $x_{p}$ and positional encodings are fed into the mViT and the output of the mViT has the same shape as input tensor $x_{o}\\in \\mathbb{R}^{S\\times C_{p}}$ . The first MLP head", + "bbox": [ + 75, + 741, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of the output $x_{o}$ means slanted bin widths and it is normalized into $b_{i}$ for further depth regression. The other features followed by the $1 \\times 1$ convolution layer perform pixel-wise dot product with a $3 \\times 3$ convolution output of $x_{d}$ to get the range-attention map. Softmax function with $1 \\times 1$ convolution layer is used to represent the range-attention map as a probability of each pixel $p_{k}$ in $N$ bin-centers. These normalized slanted bin-width $b_{i}$ and per-pixel probability $p_{k}$ with $\\alpha$ are used for SlaBins-aware regression.", + "bbox": [ + 497, + 125, + 890, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our final depth $\\tilde{\\rho}$ is computed by the linear combination of the two outputs from mViT (i.e., slanted bin-width $b_{i}$ and per-pixel probability $p_k$ ) and the transformation function in Eq. (6). To compute final depth, slanted depth bin-centers $c^s$ are computed with normalized bin-width $b_{i}$ (i.e., $\\sum b_{i} = 1$ ):", + "bbox": [ + 497, + 250, + 890, + 320 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc ^ {s} \\left(b _ {i}\\right) = r _ {\\perp} ^ {\\min } + \\left(r _ {\\perp} ^ {\\max } - r _ {\\perp} ^ {\\min }\\right) \\left(b _ {i} / 2 + \\Sigma_ {j = 1} ^ {i - 1} b _ {j}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 333, + 890, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $r_{\\perp}^{min}$ and $r_{\\perp}^{max}$ mean the lower and upper bound of the radius in slanted MCI representation. Then, orthogonal coordinate radius $r_{\\perp}$ is calculated using the cumulative sum of $c_k^s$ and $p_k^s$ (i.e., $\\sum_{k=1}^{N} c^s(b_k)p_k$ ) and transformed to our final depth $\\tilde{\\rho}$ using Eq. (6):", + "bbox": [ + 497, + 357, + 890, + 427 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\rho} = f \\left(\\hat {\\mathbf {X}}, \\sum_ {k = 1} ^ {N} c ^ {s} \\left(b _ {k}\\right) p _ {k}, \\alpha\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 428, + 890, + 451 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{X}}$ is a known normalized vector computed by pixel location and camera intrinsic, and $\\alpha$ denotes the estimated slanted angle by the slanted angle predictor.", + "bbox": [ + 497, + 454, + 890, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Loss function", + "text_level": 1, + "bbox": [ + 499, + 504, + 638, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we introduce a set of loss functions used for training SlaBins. Specifically, we compute loss in both orthogonal and spherical coordinates to make our SlaBins framework can estimate fisheye depth that has better discrimination between objects and road environments.", + "bbox": [ + 497, + 524, + 890, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Bin loss. Following [3], we use Chamfer loss[7] to make adaptive bins:", + "bbox": [ + 497, + 596, + 890, + 622 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {b i n s}} = \\operatorname {c h a m f e r} \\left(r _ {\\perp}, c ^ {s} (\\tilde {\\mathbf {b}})\\right) + \\operatorname {c h a m f e r} \\left(c ^ {s} (\\tilde {\\mathbf {b}})\\right), r _ {\\perp}), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 512, + 623, + 890, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{b}}$ indicates the predicted slanted bin widths from the SlaBins module, and $r_{\\perp}$ is the ground truth cylinder radius in the slanted MCI representation. In our case, to generate slanted bins (i.e., bins in the orthogonal coordinate), the ground-truth depth is converted to the ground truth radius $r_{\\perp}$ at the orthogonal coordinate by Eq. (5) and the slanted MCI.", + "bbox": [ + 497, + 644, + 890, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pixel-wise depth loss. We apply a scaled version of the Scale-Invariant (SI) log loss introduced by Eigen et al. [42]:", + "bbox": [ + 497, + 745, + 890, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p i x e l}} = \\beta \\sqrt {\\frac {1}{T} \\Sigma_ {i} g _ {i} ^ {2} - \\frac {\\lambda}{T ^ {2}} \\left(\\Sigma_ {i} g _ {i}\\right) ^ {2}} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 777, + 890, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $g_{i} = \\log \\tilde{\\rho} - \\log \\rho$ with the ground truth depth $\\rho$ and predicted depth $\\tilde{\\rho}$ . $T$ denotes the number of pixels containing valid ground truth values. Following [3], we set $\\lambda = 0.85$ and $\\beta = 10$ for all our experiments.", + "bbox": [ + 497, + 809, + 890, + 866 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8769", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/db73ed08f26011508219018b81bbf0e175173a1a18990b0fff9d37d677957c79.jpg", + "table_caption": [ + "Table 1: Quantitative evaluation on SynWoodScape [34]." + ], + "table_footnote": [], + "table_body": "
MethodRMSE ↓RMSE log ↓Abs Rel ↓Sq Rel ↓δ1 ↑δ2 ↑δ3 ↑
OmniDet* [20]5.6230.7401.0803.1330.0310.0640.753
BTS [21]2.9980.3250.1790.4590.6530.8450.932
AdaBins [3]1.5240.1040.0700.1270.9710.9930.997
SlaBins (ours)1.0400.0560.0220.0550.9880.9970.999
", + "bbox": [ + 78, + 140, + 468, + 194 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Slanted angle loss. As mentioned in Sec. 3.2, we predict slanted angle $\\alpha$ through a two-step module: classification and regression. For the classification module, we use the cross-entropy loss $\\mathcal{L}_{CE}$ . As for the regression module, we utilize the mean squared error (MSE) loss $\\mathcal{L}_{MSE}$ between the estimated residual and the ground truth residual. We define the sum of the CE loss and MES loss as the slanted angle loss:", + "bbox": [ + 75, + 199, + 468, + 307 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a n g l e}} = \\mathcal {L} _ {C E} + \\mathcal {L} _ {M S E}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 307, + 468, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Total loss. Finally, the total loss for our problem is:", + "bbox": [ + 77, + 327, + 421, + 340 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\gamma \\mathcal {L} _ {\\text {b i n s}} + \\mathcal {L} _ {\\text {p i x e l}} + \\mathcal {L} _ {\\text {a n g l e}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 350, + 468, + 365 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\gamma$ is a balancing parameter (we set $\\gamma = 0.1$ ).", + "bbox": [ + 77, + 372, + 413, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 77, + 397, + 206, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the proposed method from various perspectives. We introduce datasets and evaluation metrics in Sec. 4.1 and implementation details in Sec. 4.2. We then offer quantitative and qualitative comparisons to state-of-the-art methods in Sec. 4.3 and an ablation study in Sec. 4.4.", + "bbox": [ + 75, + 422, + 468, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and evaluation metrics", + "text_level": 1, + "bbox": [ + 77, + 500, + 356, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For evaluation, we use two fisheye depth datasets: the SynWoodScape dataset [34], which is a synthetic dataset, and the KITTI-360 depth dataset, which is a modified dataset of real-world KITTI-360 [24] for our scenario. Note that we additionally augment two datasets for our purposes and will release these datasets for research communities.", + "bbox": [ + 75, + 521, + 468, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SynWoodScape dataset [34]. The SynWoodScape dataset is a synthetic surround-view fisheye camera dataset for autonomous driving, which is a synthetic version of the real-world WoodScape dataset $[42]^2$ . This dataset provides pairs of synthetic fisheye images and depth maps for the front, left, right, and rear of the driving situation. Each view contains 500 images with $190^{\\circ}$ FoV and a size of $1280\\times 966$ .", + "bbox": [ + 75, + 604, + 468, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we augment the SynWoodScape dataset in terms of the viewing direction of fisheye cameras to validate the proposed framework. Specifically, we warp fisheye images and the corresponding depth maps by changing a total of nine viewing directions, of which rotation angles w.r.t. the x-axis are between $0^{\\circ}$ and $90^{\\circ}$ (between the viewing direction and the normal of the ground), including the original viewing direction. So, we have 18K fisheye images", + "bbox": [ + 75, + 701, + 468, + 811 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/252239aaff9fc4c7519572405c1ad22dcc76533f1e9e71ae4499af6c147d7d22.jpg", + "table_caption": [ + "Table 2: Quantitative evaluation on KITTI-360 depth." + ], + "table_footnote": [], + "table_body": "
MethodRMSE ↓RMSE log ↓Abs Rel ↓Sq Rel ↓δ1 ↑δ2 ↑δ3 ↑
OmniDet* [20]1.9170.1500.0860.2460.9080.9750.992
BTS [21]1.7870.1530.0930.2430.9060.9760.992
AdaBins [3]1.5900.1290.0740.1850.9400.9840.994
SlaBins (ours)1.5370.1300.0770.1770.9410.9830.994
", + "bbox": [ + 500, + 140, + 890, + 194 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in total. We split the augmented dataset into 14.4K training, 1.8K validation, and 1.8K test samples. Following the previous work [16], we train the model and evaluate the range of depth up to $40\\mathrm{m}$ because fisheye cameras undergo high data compression and can perform up to this range, unlike the KITTI dataset [10] covering up to $80\\mathrm{m}$ .", + "bbox": [ + 497, + 205, + 890, + 288 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "KITTI-360 depth dataset. The KITTI-360 dataset [24] is a suburban driving dataset for 2D/3D semantic and instance segmentation tasks that include various input modalities, semantic instance annotations, and precise localization. KITTI-360 provides fisheye images, raw LiDAR point clouds, and semantic and instance labels with $185^{\\circ}$ FoV of left and right in the driving situation, with a resolution of $1400\\times 1400$ . Note that since KITTI-360 does not provide the ground truth for fisheye depth estimation, we generate the corresponding ground truth using raw LiDAR data. We call this modified dataset for fisheye depth estimation as KITTI-360 depth dataset.", + "bbox": [ + 497, + 290, + 892, + 456 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Concretely, we first make dense LiDAR by aggregating LiDAR data of twenty neighbor frames. We then project the dense LiDAR point clouds on the fisheye camera image domain and conduct occlusion handling [18] because of the displacement between the fisheye camera and the LiDAR (see Fig. 6). We then consider depth values, of which distance is less than $40\\mathrm{m}$ , as the ground truth. We sample 4K samples of the training set from sequence 0 to 7, 0.4K samples of the validation set from sequence 9, and 0.4K samples of the test set from sequence 10. To make dense depth maps, we resize samples to $700\\times 700$ . In addition, as we did for SynWoodScape, we augment the viewing directions of the KITTI-360 depth dataset; the angle between the viewing direction and the normal on the ground to $30^{\\circ}$ , $50^{\\circ}$ , $70^{\\circ}$ , and $90^{\\circ}$ (the original viewing angle). So, we have 19.2K fisheye images in total. We split this augmented dataset into 16K training, 1.6K validation, and 1.6K test samples.", + "bbox": [ + 497, + 457, + 890, + 692 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We use the standard depth estimation metrics used in the previous work [21]. For error metrics, we calculate mean absolute relative error (Abs Rel), root mean square error (RMSE), root mean square log error (RMSE log), the root mean square relative error (Sq Rel). For accuracy metrics, we calculate threshold accuracy $\\delta_{i} = 1.25^{i}$ for $i\\in \\{1,2,3\\}$ . Generally, the smaller the error, the better, and the higher the accuracy, the better.", + "bbox": [ + 497, + 695, + 890, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation details", + "text_level": 1, + "bbox": [ + 499, + 815, + 712, + 831 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our proposed method is implemented in PyTorch and trained on 4 RTX A6000 GPUs. Our model is trained for", + "bbox": [ + 499, + 838, + 890, + 865 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2Unfortunately, the real-world WoodScape dataset [42] does not provide the ground truth depth data.", + "bbox": [ + 75, + 819, + 468, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "${}^{3}$ Currently,the SynWoodScape dataset is partially released to the public. We perform the experiments using the partial dataset.", + "bbox": [ + 75, + 842, + 467, + 865 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8770", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6d3c7429a7e8ad17cd1c5f6f3597db8ef6e94bd5eaf4c4c592b32150cf36ac18.jpg", + "image_caption": [ + "RGB" + ], + "image_footnote": [], + "bbox": [ + 208, + 115, + 292, + 290 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf170ac772f44ad54bf400ce288917d5be89402e6eb6226517c2730b8887ffc2.jpg", + "image_caption": [ + "OmniDet* [20]" + ], + "image_footnote": [], + "bbox": [ + 302, + 116, + 386, + 290 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b8dabc0e04cf9b8853c8a5e676afedd420da7e777af3f39bbb3a504e43ea33f6.jpg", + "image_caption": [ + "BTS [21]" + ], + "image_footnote": [], + "bbox": [ + 396, + 116, + 480, + 290 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a1f6f495b7946782a7cd798ea1f2e57ac5093d2c85012f2d41d27ab407f7821.jpg", + "image_caption": [ + "AdaBins [3]" + ], + "image_footnote": [], + "bbox": [ + 489, + 116, + 573, + 290 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cd939911703bf1ea774a6a074cd4f04fdc646cabd88727d485cc8bf8baf84281.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 583, + 116, + 665, + 290 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/436fa484293c46f7d11dfbd38d909b935b5f6a49dabde97688b9499937608064.jpg", + "image_caption": [ + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 675, + 116, + 759, + 290 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0413c375d7b9bb8d32226b85d47b58ae94b75b287ba2cf54536a5f369a2e20ca.jpg", + "image_caption": [ + "Figure 5: Qualitative comparison on the SynWoodScape dataset. Sky-blue boxes are enlarged views of specific objects to show detail. Overall, our method shows distinct boundaries compared to the other models. In particular, we can observe the detailed structure from our results, such as bicycle wheels in the first row and the boundary of two overlapped vehicles in the second row.", + "Figure 6: Qualitative comparison on the KITTI-360 depth dataset. For visualization, we crop the object region in the fisheye image. In contrast to the other methods, our method shows clear boundaries near the object regions, as shown in sky-blue and green boxes." + ], + "image_footnote": [], + "bbox": [ + 168, + 352, + 759, + 600 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "10 epochs on SynWoodScape and 20 epochs on KITTI-360 depth using AdamW [28] optimizer with weight-decay $10^{-2}$ , with a batch size of 8. We use the same 1-cycle policy [35] as AdaBins [3] for the learning rate with max $\\mathrm{lr} = 3.5 \\times 10^{-4}$ . For the first $30\\%$ of iterations, the linear learning rate warm-up technique is utilized. For the learning rate decay, the cosine annealing learning rate strategy is applied.", + "bbox": [ + 75, + 646, + 468, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison", + "text_level": 1, + "bbox": [ + 75, + 758, + 208, + 774 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our SlaBins with several baseline approaches: OmniDet [20], BTS [21], and AdaBins [3]. OmniDet is a fisheye depth estimation method using self-supervised learning. Based on the released code of the authors, we modify OmniDet to a supervised one for fairness (we denote this modified one OmniDet*). As base", + "bbox": [ + 75, + 783, + 468, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "line methods for pinhole cameras, we compare BTS and AdaBins, especially AdaBins is a bins-based fundamental method for depth estimation. It should be noted that we target depth estimation fisheye cameras, but we additionally compare it with conventional pinhole camera-based methods due to the lack of fisheye camera-based methods.", + "bbox": [ + 497, + 646, + 890, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative evaluation. Table 1 shows the quantitative comparisons on the SynWoodScape dataset. The arrow next to the metric indicates the direction of better performance. Overall, SlaBins outperforms the other baseline approaches [3, 20, 21]. In particular, RMSE and Abs-Rel errors are significantly improved; $31.7\\%$ and $68.6\\%$ decreased compared to AdaBins, which shows the effect of the slanted MCI representation in the fisheye domain.", + "bbox": [ + 497, + 737, + 890, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the KITTI-360 depth dataset, our method still outper", + "bbox": [ + 517, + 852, + 890, + 866 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8771", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2212e2af948023b583428b260616d7d55ad63e5f051987bd8e26db16f5727976.jpg", + "image_caption": [ + "Figure 7: Depth histogram on the object regions. Left: RGB data (top) and the concept of depth histogram in the orthogonal coordinate (bottom). Middle: depth maps of only object regions for each model and the ground truth. Right: depth histograms (AdaBins, original MCI, slanted MCI, ground truth in that order)." + ], + "image_footnote": [], + "bbox": [ + 109, + 123, + 435, + 275 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "forms the comparison methods (see Table 2). Specifically, RMSE and Sq-Rel errors are improved; $13.9\\%$ and $27.1\\%$ decreased compared to BTS and $3.3\\%$ and $4.3\\%$ decreased compared to AdaBins. Note that even though the KITTI-360 depth dataset captured in suburban scenes does not perfectly fit our scenario of road environments, SlaBins still shows competitive performance, which demonstrates the effect of the proposed framework.", + "bbox": [ + 75, + 344, + 468, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative evaluation. Figure 5 shows the estimated depth maps for the test scenes on SynWoodScape. Due to the nature of fisheye cameras, we can observe ambiguous boundary regions in the comparison methods. Even AdaBins, which is our baseline and a naive MSI representation by the Euclidean distance of fisheye, shows bleeding depth estimation results near the boundaries between objects and ground regions. In contrast, thanks to the slanted MCI representation, SlaBins shows accurate depth results with distinct boundaries regardless of camera viewing directions. Specifically, we can observe that sharp objects with details (e.g., human legs and bike) are preserved in our method, unlike the other methods that smooth object details. In addition, we can observe a similar trend on the KITTI-360 depth dataset, as shown in Fig. 6. Compared to the baseline methods, SlaBins shows better depth consistency even on sharp objects (e.g., third and fourth columns in Fig. 6). Additional results are available in the supplementary material.", + "bbox": [ + 75, + 459, + 468, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation study", + "text_level": 1, + "bbox": [ + 75, + 719, + 226, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-layered representation. We compare the proposed slanted MCI with the MSI used in AdaBins and the original MCI on the orthogonal coordinate. We measure the depth histogram for only object regions on the orthogonal coordinate. As shown in Fig. 7, compared to the other representations, SlaBins reveals the most similar trend (i.e., peak points) with the ground truth histogram. Furthermore, we compare slanted MCI and original MCI in terms of depth consistency. Figure 8 shows that the slanted MCI provides", + "bbox": [ + 75, + 741, + 467, + 865 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/620ba8fd308f1615bb8ba8f2c0ed397c2b10ba1a9f8cae01b712a742bb836b65.jpg", + "image_caption": [ + "Figure 8: Slanted MCI vs. original MCI in depth and clustering. All the results including the ground truth are converted to the XZ distance in the orthogonal coordinate." + ], + "image_footnote": [], + "bbox": [ + 512, + 122, + 873, + 226 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/646d053359c18a99b902eaf45cd69b457ebc9b56b3ca22b459af7d26c6432a9c.jpg", + "table_caption": [ + "Table 3: Ablation study for the slanted angle predictor on SynWoodScape and KITTI-360 depth. We report the depth error (Abs Rel) and mean angle error. C and R denote classification-only and regression-only, respectively." + ], + "table_footnote": [], + "table_body": "
MethodSynWoodScapeKITTI-360 depth
Abs Rel ↓Angle error↓Abs Rel ↓Angle error
R0.07527.0310.60448.32
C0.0260.8420.1035
C+R (ours)0.0220.0310.0770.017
", + "bbox": [ + 559, + 326, + 833, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "more consistent results in object regions both on depth and depth-based clustering results. Hence, we believe that the slanted MCI can provide robust geometric constraints for various tasks in fisheye domains.", + "bbox": [ + 497, + 396, + 890, + 451 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Slanted angle prediction. As shown in Table 3, we perform an ablation study for the angle prediction module. Unlike a conventional regression module having the advantage over classifications, in our case, it is hard to converge because of the trade-off with the depth estimation part in the SlaBins module. That is why regression-only module is poor on the angle prediction. On the contrary, our two-step prediction module, composed of classification and residual regression, shows the best performance with stable training.", + "bbox": [ + 497, + 454, + 890, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 499, + 590, + 616, + 605 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed a novel fisheye depth estimation network, called SlaBins, based on the geometric property of the road environments. We newly introduce the slanted MCI representation, which allows us to represent a given road scene as a set of cylindrical layers orthogonal to the road ground. By virtue of the slanted MCI, our network can implicitly encode scene depth information invariant to the camera viewing direction while preserving boundary information of scene objects w.r.t. the road region. Thus, we achieve state-of-the-art performance on monocular fisheye depth estimation.", + "bbox": [ + 497, + 614, + 890, + 764 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported by 42dot Inc., Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2020-0-01336, Artificial Intelligence Graduate School Program (UNIST)) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF2021R1C1C1005723).", + "bbox": [ + 497, + 769, + 890, + 865 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8772", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 124, + 171, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] T-Bone Accident Statistics And Other Legal Information. https://www.injurytriallawyer.com/faqs/ how-dangerous-are-t-bone-accidents-.cfm. 1", + "[2] Ashutosh Agarwal and Chetan Arora. Attention Attention Everywhere: Monocular Depth Prediction with Skip Attention. arXiv preprint arXiv:2210.09071, 2022. 1", + "[3] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In CVPR, 2021. 2, 3, 5, 6, 7", + "[4] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. LocalBins: Improving Depth Estimation by Learning Local Distributions. In ECCV, 2022. 2", + "[5] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, et al. PersFormer: 3D Lane Detection via Perspective Transformer and the OpenLane Benchmark. arXiv preprint arXiv:2203.11089, 2022. 1", + "[6] Liang-Chieh Chen, Huiyu Wang, and Siyuan Qiao. Scaling wide residual networks for panoptic segmentation. arXiv preprint arXiv:2011.11675, 2020. 1", + "[7] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In CVPR, 2017. 5", + "[8] Olivier Faugeras and Quang-Tuan Luong. The geometry of multiple images: the laws that govern the formation of multiple images of a scene and some of their applications. MIT press, 2001. 2", + "[9] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, 2018. 2", + "[10] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 6", + "[11] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In CVPR, 2017. 2", + "[12] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In CVPR, 2020. 2", + "[13] Yuxuan Han, Ruicheng Wang, and Jiaolong Yang. Single-View View Synthesis in the Wild with Learned Adaptive Multiplane Images. arXiv preprint arXiv:2205.11733, 2022. 2", + "[14] Adrian Johnston and Gustavo Carneiro. Self-supervised monocular trained depth estimation using self-attention and discrete disparity volume. In CVPR, 2020. 2", + "[15] Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, and Junmo Kim. Global-Local Path Networks for Monocular Depth Estimation with Vertical Cut-Depth. arXiv preprint arXiv:2201.07436, 2022. 2", + "[16] Varun Ravi Kumar, Sandesh Athni Hiremath, Markus Bach, Stefan Milz, Christian Witt, Clément Pinard, Senthil Yogamani, and Patrick Mäder. Fisheyedistancenet: Self-supervised scale-aware distance estimation using monocular" + ], + "bbox": [ + 78, + 147, + 467, + 866 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "fisheye camera for autonomous driving. In ICRA, 2020. 1, 2, 6", + "[17] Varun Ravi Kumar, Marvin Klingner, Senthil Yogamani, Markus Bach, Stefan Milz, Tim Fingscheidt, and Patrick Mäder. SVDistNet: Self-supervised near-field distance estimation on surround view fisheye cameras. IEEE Transactions on Intelligent Transportation Systems, 2021. 2", + "[18] Varun Ravi Kumar, Stefan Milz, Christian Witt, Martin Simon, Karl Amende, Johannes Petzold, Senthil Yogamani, and Timo Pech. Monocular fisheye camera depth estimation using sparse lidar supervision. In IEEE International Conference on Intelligent Transportation Systems (ITSC), 2018. 2, 6", + "[19] Varun Ravi Kumar, Senthil Yogamani, Markus Bach, Christian Witt, Stefan Milz, and Patrick Mäder. Unrectdepthnet: Self-supervised monocular depth estimation using a generic framework for handling common camera distortion models. In IROS, 2020. 2", + "[20] Varun Ravi Kumar, Senthil Yogamani, Hazem Rashed, Ganesh Sitsu, Christian Witt, Isabelle Leang, Stefan Milz, and Patrick Mader. Omnidet: Surround view cameras based multi-task visual perception network for autonomous driving. IEEE Robotics and Automation Letters, 6(2):2830-2837, 2021. 1, 2, 3, 6, 7", + "[21] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 2, 6, 7", + "[22] Jiaxin Li, Zijian Feng, Qi She, Henghui Ding, Changhu Wang, and Gim Hee Lee. Mine: Towards continuous depth spi with nerf for novel view synthesis. In ICCV, 2021. 2", + "[23] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. DepthFormer: Exploiting Long-Range Correlation and Local Information for Accurate Monocular Depth Estimation. arXiv preprint arXiv:2203.14211, 2022. 2", + "[24] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE TPAMI, 2022. 2, 6", + "[25] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETRv2: A Unified Framework for 3D Perception from Multi-Camera Images. arXiv preprint arXiv:2206.01256, 2022. 1", + "[26] Yuxuan Liu, Yuan Yixuan, and Ming Liu. Ground-aware monocular 3d object detection for autonomous driving. IEEE Robotics and Automation Letters, 6(2):919-926, 2021. 3", + "[27] Xiaoxiao Long, Cheng Lin, Lingjie Liu, Wei Li, Christian Theobalt, Ruigang Yang, and Wenping Wang. Adaptive surface normal constraint for depth estimation. In ICCV, 2021. 3", + "[28] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In *ICLR*, 2018. 7", + "[29] Reza Mahjourian, Martin Wicke, and Anelia Angelova. Unsupervised learning of depth and ego-motion from monocular video using 3d geometric constraints. In CVPR, 2018. 2, 3", + "[30] Vaishakh Patil, Christos Sakaridis, Alexander Liniger, and Luc Van Gool. P3Depth: Monocular Depth Estimation with a Piecewise Planarity Prior. In CVPR, 2022. 3" + ], + "bbox": [ + 502, + 127, + 890, + 866 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8773", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Elad Plaut, Erez Ben Yaacov, and Bat El Shlomo. 3d object detection from a single fisheye image without a single fisheye training image. In CVPR, 2021. 1", + "[32] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 2", + "[33] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI, 2020. 2", + "[34] Ahmed Rida Sekkat, Yohan Dupuis, Varun Ravi Kumar, Hazem Rashed, Senthil Yogamani, Pascal Vasseur, and Paul Honeine. SynWoodScape: Synthetic Surround-View Fisheye Camera Dataset for Autonomous Driving. IEEE Robotics and Automation Letters, 7(3):8502-8509, jul 2022. 1, 2, 6", + "[35] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, volume 11006, pages 369-386. SPIE, 2019. 7", + "[36] Richard Szeliski and Polina Golland. Stereo matching with transparency and matting. In ICCV, 1998. 2", + "[37] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 4", + "[38] Andrew Tao, Karan Sapra, and Bryan Catanzaro. Hierarchical multi-scale attention for semantic segmentation. arXiv preprint arXiv:2005.10821, 2020. 1", + "[39] Richard Tucker and Noah Snively. Single-view view synthesis with multiplane images. In CVPR, 2020. 2", + "[40] Qingan Yan, Pan Ji, Nitin Bansal, Yuxin Ma, Yuan Tian, and Yi Xu. FisheyeDistill: Self-Supervised Monocular Depth Estimation with Ordinal Distillation for Fisheye Cameras. arXiv preprint arXiv:2205.02930, 2022. 2, 3", + "[41] Yaozu Ye, Kailun Yang, Kaite Xiang, Juan Wang, and Kaiwei Wang. Universal semantic segmentation for fisheye urban driving images. In IEEE International Conference on Systems, Man, and Cybernetics (SMC), pages 648-655, 2020. 1", + "[42] Senthil Yogamani, Ciarán Hughes, Jonathan Horgan, Ganesh Sistu, Padraig Varley, Derek O'Dea, Michal Uricár, Stefan Milz, Martin Simon, Karl Amende, et al. WoodScape: A multi-task, multi-camera fisheye dataset for autonomous driving. arXiv preprint arXiv:1905.01489, 2019. 5, 6", + "[43] Yifan Zhang, Qijian Zhang, Zhiyu Zhu, Junhui Hou, and Yixuan Yuan. GLENet: Boosting 3D Object Detectors with Generative Label Uncertainty Estimation. arXiv preprint arXiv:2207.02466, 2022. 1", + "[44] Dingfu Zhou, Jin Fang, Xibin Song, Liu Liu, Junbo Yin, Yuchao Dai, Hongdong Li, and Ruigang Yang. Joint 3d instance segmentation and object detection for autonomous driving. In CVPR, 2020. 1" + ], + "bbox": [ + 78, + 125, + 468, + 800 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8774", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/SlaBins_ Fisheye Depth Estimation using Slanted Bins on Road Environments/2f72858d-bf2e-45dd-b589-9c5d6578e810_model.json b/2023/SlaBins_ Fisheye Depth Estimation using Slanted Bins on Road Environments/2f72858d-bf2e-45dd-b589-9c5d6578e810_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1fe1d4106070dbc86507a597d2f4f2c22ad99a45 --- /dev/null +++ b/2023/SlaBins_ Fisheye Depth Estimation using Slanted Bins on Road Environments/2f72858d-bf2e-45dd-b589-9c5d6578e810_model.json @@ -0,0 +1,2464 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.13, + 0.052, + 0.201, + 0.085 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.254, + 0.049, + 0.79, + 0.087 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.162, + 0.877, + 0.182 + ], + "angle": 0, + "content": "SlaBins: Fisheye Depth Estimation using Slanted Bins on Road Environments" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.207, + 0.816, + 0.241 + ], + "angle": 0, + "content": "Jongsung Lee\\(^{1}\\), Gyeongsu Cho\\(^{1*}\\), Jeongin Park\\(^{1*}\\), Kyongjun Kim\\(^{1*}\\), Seongoh Lee\\(^{1*}\\), Jung-Hee Kim\\(^{2}\\), Seong-Gyun Jeong\\(^{2}\\), Kyungdon Joo\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.243, + 0.725, + 0.26 + ], + "angle": 0, + "content": "1Artificial Intelligence Graduate School, UNIST, 242dot Inc." + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.263, + 0.784, + 0.277 + ], + "angle": 0, + "content": "{syniez, threedv, jeonginpark, kimkj38, solee, kyungdon}@unist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.278, + 0.658, + 0.29 + ], + "angle": 0, + "content": "{junghee.kim, seonggyun.jeong}@42dot.ai" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.321, + 0.314, + 0.335 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.351, + 0.474, + 0.725 + ], + "angle": 0, + "content": "Although 3D perception for autonomous vehicles has focused on frontal-view information, more than half of fatal accidents occur due to side impacts in practice (e.g., T-bone crash). Motivated by this fact, we investigate the problem of side-view depth estimation, especially for monocular fisheye cameras, which provide wide FoV information. However, since fisheye cameras head road areas, it observes road areas mostly and results in severe distortion on object areas, such as vehicles or pedestrians. To alleviate these issues, we propose a new fisheye depth estimation network, SlaBins, that infers an accurate and dense depth map based on a geometric property of road environments; most objects are standing (i.e., orthogonal) on the road environments. Concretely, we introduce a slanted multi-cylindrical image (MCI) representation, which allows us to describe a distance as a radius to a cylindrical layer orthogonal to the ground regardless of the camera viewing direction. Based on the slanted MCI, we estimate a set of adaptive bins and a per-pixel probability map for depth estimation. Then by combining it with the estimated slanted angle of viewing direction, we directly infer a dense and accurate depth map for fisheye cameras. Experiments demonstrate that SlaBins outperforms the state-of-the-art methods in both qualitative and quantitative evaluation on the SynWoodScape and KITTI-360 depth datasets. For more information, you can visit our project page https://syniez.github.io/SlaBins/." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.752, + 0.21, + 0.767 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.776, + 0.47, + 0.833 + ], + "angle": 0, + "content": "With the advent of autonomous driving, the importance of 3D perception keeps growing for safety [20, 25, 44]. For example, ADAS systems pervading our daily driving already, such as lane-keeping aid and forward collision" + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.317, + 0.582, + 0.326 + ], + "angle": 0, + "content": "RGB" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.327, + 0.616, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.369, + 0.615, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.411, + 0.615, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.327, + 0.764, + 0.338 + ], + "angle": 0, + "content": "Slanted MCI representation" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.339, + 0.753, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.389, + 0.774, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.805, + 0.317, + 0.838, + 0.326 + ], + "angle": 0, + "content": "Depth" + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.328, + 0.877, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.369, + 0.877, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.411, + 0.878, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.458, + 0.893, + 0.496 + ], + "angle": 0, + "content": "Figure 1: SlaBins on the SynWoodScape dataset [34]. SlaBins predicts dense and accurate depth maps regardless of the viewing directions using the slanted MCI representation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.893, + 0.713 + ], + "angle": 0, + "content": "avoidance assist, depend on various sensor modalities to observe 3D information [43, 5]. Here, one interesting point is that various sensor modalities for 3D perception, including many datasets related to autonomous driving, are focusing on the frontal view predominantly [43, 38, 2, 6]. We can easily deduce the reason by our common sense that it is important to keep an eye on the front while driving. However, according to the National Transportation Highway Safety Administration [1], \\(51\\%\\) of fatal car accidents correspond to T-bone crashes in which one car hits the side of another, which is the second highest type of accident. This statistic supports that side-view 3D perception is important to assist drivers because they are hard to pay attention to the side view due to the physical characteristics of the people looking ahead and their position characteristics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.894, + 0.868 + ], + "angle": 0, + "content": "Motivated by this fact, 3D perception for side-view, especially fisheye-based depth estimation, has started to gain the spotlight [31, 16, 41]. Fisheye cameras have several advantages for 3D perception in road scenes. Specifically, fisheye cameras provide broader FoV information (typically greater than \\(180^{\\circ}\\)) in a single image, which allows us to observe road information (e.g., road, vehicles, and pedestrians) at a low cost. In addition, we can directly utilize perspective camera 3D perception approaches because mathematical mutual conversion between two camera models is possible. Thus, if we can obtain side-view depth informa" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.843, + 0.293, + 0.855 + ], + "angle": 0, + "content": "*Equal contribution (alphabet order)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.855, + 0.223, + 0.866 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.843, + 0.293, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.501, + 0.899 + ], + "angle": 0, + "content": "8765" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.126, + 0.47, + 0.154 + ], + "angle": 0, + "content": "tion using a single fisheye camera, we can efficiently handle various vision tasks for autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.156, + 0.47, + 0.363 + ], + "angle": 0, + "content": "However, despite the advantages of fisheye cameras, monocular fisheye camera-based side-view depth estimation is a challenging problem due to significant distortion. In addition to inherent issues of monocular depth estimation, such as scale ambiguity [8], severe distortion of fisheye cameras makes depth estimation more difficult by distorting scene objects (e.g., close objects look closer and distant objects look more distant abnormally). Particularly, this distortion problem could be more severe in road environments because fisheye cameras for ADAS systems generally look to the road areas, and vital objects such as vehicles and pedestrians are located at the edge of the image. That is, most regions of the image correspond to the road areas and objects contain very few pixels, so depth estimation becomes more challenging." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.47, + 0.544 + ], + "angle": 0, + "content": "To alleviate this problem, we pay attention to the geometric property of road environments; most objects are standing on the road environments. Based on this geometric characteristic, we can easily imagine the visible region of each object as a planar segment orthogonal to the road region. This approximation is connected to a multi-plane image (MPI) concept [36] that represents a given scene as the composition of multi-layered images according to depth. Note that one significant difference is that a fisheye camera looks down to the road region in our case (i.e., camera viewing direction is not parallel to the ground), unlike a camera points to a given scene in the general MPI concept. Thus, a naive MPI concept is not directly applicable." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.47, + 0.807 + ], + "angle": 0, + "content": "In this work, we propose a new fisheye depth estimation framework, SlaBins, that exploits the geometric property of the road environments (see Fig. 1). To be specific, we adopt a multi-cylindrical image (MCI) representation; especially, we introduce a slanted MCI of which the cylindrical layer is defined as orthogonal to the road ground regardless of camera viewing direction. Based on the slanted MCI, we estimate adaptive bins and the per-pixel probability of the slanted cylindrical layers, which allows us to implicitly preserve the geometric property of the road environments and improve the depth quality near the boundary between objects and the ground region. In addition, we independently estimate the slanted angle of the camera viewing direction w.r.t. the road ground. Then, by combining this angle with the estimated depth information in the slanted MCI domain, we can directly compute a dense and accurate depth map for fisheye cameras. We validate the proposed SlaBins framework on the SynWoodScape dataset [34] and KITTI-360 depth dataset (a modified dataset of KITTI" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.127, + 0.892, + 0.168 + ], + "angle": 0, + "content": "360 [24] for our scenario), where SlaBins outperforms the baseline methods [3, 20, 21] and shows potential as scene representations for downstream tasks, e.g., segmentation." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.169, + 0.825, + 0.181 + ], + "angle": 0, + "content": "In a nutshell, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.187, + 0.892, + 0.229 + ], + "angle": 0, + "content": "- We propose a new fisheye depth estimation framework, SlaBins, which infers dense depth based on the geometric property of road environments." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.238, + 0.892, + 0.307 + ], + "angle": 0, + "content": "- We newly introduce a slanted MCI representation as an intermediate road scene representation, which allows us to describe a given scene regardless of viewing direction and improve the depth quality on the boundary areas between objects and the road." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.317, + 0.892, + 0.372 + ], + "angle": 0, + "content": "- By decoupling the slanted angle of an input image, we force the proposed model to learn the slanted MCI depth information and then seamlessly compute a fisheye depth map in a slanted-aware regression way." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.187, + 0.892, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.384, + 0.642, + 0.399 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.408, + 0.892, + 0.587 + ], + "angle": 0, + "content": "Monocular depth estimation. With the advent of breakthroughs in deep neural networks, such as CNN and transformer architecture, monocular depth estimation has shown impressive performance [15, 33, 21, 32, 23, 18, 40, 11, 9, 12, 29, 19, 16, 17, 20]. As one of the recent trends, several methods formulate a depth estimation task as a combination of classification and regression problems, which enables training the network more easily and shows stable performance [14, 3, 4]. For this formulation, they commonly use an MPI according to depth, which is popularly used in novel-view synthesis tasks [22, 39, 13]. The MPI notation represents depth as quantized depth values w.r.t. the uniformly quantized planes and residuals from mapped planes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.892, + 0.808 + ], + "angle": 0, + "content": "To improve the uniformly quantized planes by uniform bins, recent research works [3, 4] present a network that predicts adaptive bins depending on the input scene and local regions in the image. AdaBins [3] computes adaptive bin-centers and per-pixel probabilities relying on each input image. Inspired by AdaBins, LocalBins [4] improves depth quality by considering both the global histogram and local-region histograms of each input image. Nevertheless, the predicted depth from the above models also depends on the camera's pose because they predict depth parallel with the camera's principal axis. This constraint makes it hard to adopt the previous models into fisheye depth estimation, which has a large FoV and uses the Euclidean distance instead of depth. In this work, we introduce and exploit a slanted MCI representation that is invariant to the viewing angle of fisheye cameras w.r.t. road environments." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.811, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Fisheye depth estimation. Despite increasing demand for fisheye cameras in the industry, like ADAS systems, there are only a few previous research works in fisheye depth estimation fields [20, 17, 19, 16, 40]. Kumar et al. [19] propose" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.821, + 0.47, + 0.867 + ], + "angle": 0, + "content": "1Basically, depth in the fisheye model is defined as the Euclidean distance, which is connected to a multi-spherical image (MSI). In this work, We adopt an MCI representation that combines the geometrical suitability of MPI and the fisheye model fitness of MSI." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.503, + 0.899 + ], + "angle": 0, + "content": "8766" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.121, + 0.87, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.277, + 0.893, + 0.318 + ], + "angle": 0, + "content": "Figure 2: Overview of the proposed SlaBins. (a) Given an input fisheye RGB image, SlaBins framework first extracts features by the encoder-decoder block. The extracted features are used as input to the slanted angle prediction module and the SlaBins module to get a depth image. (b) Illustration of the coordinates and MCI representations defined in our model. (c) Illustration of the slanted angle predictor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.331, + 0.47, + 0.565 + ], + "angle": 0, + "content": "a camera-aware depth estimation network to handle the severe distortion of fisheye cameras. They encode the camera's intrinsic parameters as a tensor named camera geometry tensor and use it in the training step for camera-aware depth estimation in fisheye images. With camera geometry tensor, they propose a multi-task network for fisheye cameras, called OmniDet [20], that simultaneously estimates depth estimation, motion, semantic segmentation, and object detection for synergy. However, since OmniDet relies on the partial region in the whole image, they may miss information on image edges. Using a whole image, Yan et al. [40] distill depth ordering information from a teacher network trained on the large-scale dataset to train in fisheye images, but FoV they cover is not large (e.g., vFoV=66°). In contrast to previous approaches, we propose a fisheye depth estimation network that predicts fisheye image depth without any information loss such as crop, and narrow FoV." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.804 + ], + "angle": 0, + "content": "Geometric cues for depth estimation. One way to improve fisheye depth estimation is to exploit geometric cues between the camera and the world. Liu et al. [26] use a ground plane to get depth priors using camera elevation and camera parameters. Moreover, Mahjourian et al. [29] predict camera ego-motion and compute loss on both the 3D point cloud domain and the 2D image domain. In addition to driving scenes, several works use plane coefficients to give consistency on plane regions in depth estimation [30, 27]. Patil et al. [30] estimate plane coefficients instead of depth and then combine plane coefficients into depth to use plane constraints (e.g., co-planarity, plane normals) in 3D space. Inspired by the previous works [3, 30], we design a network that estimates a dense and accurate depth map for fisheye cameras based on the slanted MCI representation using some geometric constraints (i.e., orthogonality between ground and object)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.816, + 0.285, + 0.831 + ], + "angle": 0, + "content": "3. Proposed Framework" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.839, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Inspired by the geometric property of the road environments, we propose a new fisheye depth estimation frame" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.33, + 0.892, + 0.523 + ], + "angle": 0, + "content": "work (see Fig. 2 for the overview). Specifically, we introduce a slanted MCI representation, which allows us to describe a multi-layer cylindrical image orthogonal to the road ground regardless of the viewing direction of the camera. Based on the slanted MCI representation, our SlaBins module estimates the adaptive bin widths and per-pixel probability map in the orthogonal coordinate, which provides depth information invariant to the camera viewing direction. We then combine the estimated depth information with the slanted angle estimated by the slanted angle prediction module. Through this process, we can directly estimate a dense depth map for fisheye cameras. Our approach is built upon the previous AdaBins framework [3]. So, we call our method SlaBins." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.535, + 0.757, + 0.55 + ], + "angle": 0, + "content": "3.1. Slanted MCI Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.558, + 0.892, + 0.641 + ], + "angle": 0, + "content": "As intermediate scene representation, we use multi-layer images for fisheye depth estimation. Here, we briefly compare two multi-layer image representations for fisheye images: multi-spherical image (MSI) and multi-cylindrical image (MCI). We then present a slanted MCI representation, which is a suitable representation of our problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.81 + ], + "angle": 0, + "content": "MSI vs. MCI. By nature of fisheye cameras covering over \\(180^{\\circ}\\), it is common to use the Euclidean distance as its depth; that is, a layer with the same depth value forms a sphere in the fisheye domain. Thus, we can represent a multi-layer image for fisheye cameras as an MSI in the spherical coordinate, which is a naive representation. However, this naive MSI representation has inherent limitations for fisheye depth estimation. We observed that such a representation can result in ambiguous depth boundaries between the ground and object regions, as shown in Fig. 3a. It may neutralize the benefit of multi-layer images by separating an object into two different layers, for example." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.867 + ], + "angle": 0, + "content": "In the MCI representation, a cylindrical layer encodes the same Euclidean distance w.r.t. the y-axis. In other words, we can represent an object, such as a pedestrian on the road, in a single cylindrical layer (see Fig. 3b). Further" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.501, + 0.898 + ], + "angle": 0, + "content": "8767" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.105, + 0.128, + 0.191, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.233, + 0.157, + 0.243 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.2, + 0.136, + 0.286, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.233, + 0.252, + 0.243 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.124, + 0.443, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.233, + 0.382, + 0.243 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.245, + 0.47, + 0.346 + ], + "angle": 0, + "content": "Figure 3: MSI vs. MCI representation. Illustration of an example scene and zoomed results with (a) MSI and (b) MCI, where we can observe more accurate depth in the slanted MCI representation near the boundary between pedestrians and the ground. (c) Relationship between the original cylindrical layer \\( C_r^o \\) (gray-colored cylinder) in the world coordinate (arrows) and the slanted cylindrical layer \\( C_{r_\\perp}^s \\) (orange-colored cylinders) in the orthogonal coordinate (dashed arrows)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.469, + 0.413 + ], + "angle": 0, + "content": "more, we can directly exploit the geometric properties of the road environment: most objects are standing (i.e., orthogonal) on the road environment. For this reason, we adopt the MCI as our intermediate scene representation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.415, + 0.469, + 0.454 + ], + "angle": 0, + "content": "Slanted MCI representation. Let \\(\\mathbf{X}\\) be a 3D point in the world coordinate. Then, we can represent \\(\\mathbf{X}\\) according to the coordinate systems:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.456, + 0.469, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} = \\left[ \\begin{array}{l} X \\\\ Y \\\\ Z \\end{array} \\right] = \\rho \\left[ \\begin{array}{c} \\hat {X} \\\\ \\hat {Y} \\\\ \\hat {Z} \\end{array} \\right] = \\rho \\left[ \\begin{array}{c} \\sin \\theta \\cos \\phi \\\\ \\sin \\theta \\sin \\phi \\\\ \\cos \\theta \\end{array} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.469, + 0.564 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{X}} = [\\hat{X},\\hat{Y},\\hat{Z}]^{\\top}\\) is the normalized unit vector (i.e., ray direction), \\(\\rho\\) indicates the Euclidean distance (i.e., depth in the fisheye domain), and \\(\\theta\\) and \\(\\phi\\) denote the elevation and azimuth angles, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.469, + 0.592 + ], + "angle": 0, + "content": "Based on the cylindrical representation, we can form a cylindrical layer according to a radius \\(r\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.596, + 0.469, + 0.614 + ], + "angle": 0, + "content": "\\[\nC _ {r} ^ {o} = \\left\\{\\mathbf {X} \\mid X ^ {2} + Z ^ {2} = r ^ {2}; \\mathbf {X} \\in \\mathbb {R} ^ {3} \\right\\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.469, + 0.702 + ], + "angle": 0, + "content": "We call \\( C_r^o \\) the original cylindrical layer at \\( r \\). This original MCI is inherently suitable for road environments and shows less distortion than MSI on ambiguous depth boundaries. However, depending on the principal axis of cameras (i.e., viewing direction) w.r.t. the road environments, the original MCI may hinder layered scene representation like the MSI." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.469, + 0.813 + ], + "angle": 0, + "content": "To alleviate this issue, we consider the 1D angle \\(\\alpha\\) between the principal axis and the road environments. Using \\(\\alpha\\), we can implicitly describe the slanted MCI representation that is orthogonal to the road areas (see Fig. 3c). To this end, let \\(\\mathbf{X}_{\\perp}\\) be a 3D point in the orthogonal coordinate, of which the XZ plane is parallel to the ground. Similarly to Eq. (2), we can form a cylindrical layer according to a radius \\(r_{\\perp}\\) in the orthogonal coordinate as:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.818, + 0.469, + 0.836 + ], + "angle": 0, + "content": "\\[\nC _ {r _ {\\perp}} ^ {s} = \\left\\{\\mathbf {X} _ {\\perp} \\mid X _ {\\perp} ^ {2} + Z _ {\\perp} ^ {2} = r _ {\\perp} ^ {2}; \\mathbf {X} _ {\\perp} \\in \\mathbb {R} ^ {3} \\right\\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.469, + 0.867 + ], + "angle": 0, + "content": "which corresponds to a slanted cylindrical layer in the world (camera) coordinate (see Fig. 2b for coordinate systems)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.127, + 0.892, + 0.153 + ], + "angle": 0, + "content": "By \\(\\alpha\\), we can derive the relationship between \\(\\mathbf{X}_{\\perp}\\) in the orthogonal coordinate and \\(\\mathbf{X}\\) in the world coordinate by:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.156, + 0.892, + 0.197 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {\\perp} = \\left[ \\begin{array}{l} X _ {\\perp} \\\\ Y _ {\\perp} \\\\ Z _ {\\perp} \\end{array} \\right] = \\left[ \\begin{array}{c} X \\\\ Y \\cos \\alpha + Z \\sin \\alpha \\\\ - Y \\sin \\alpha + Z \\cos \\alpha \\end{array} \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.201, + 0.892, + 0.243 + ], + "angle": 0, + "content": "Then, we can deduce the following relationship between a 3D point \\(\\mathbf{X}_{\\perp}\\) on \\(C_{r_{\\perp}}^{s}\\) and its spherical representation in the camera coordinate by using Eqs. (3) and (4):" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.245, + 0.891, + 0.288 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} r _ {\\perp} ^ {2} = X _ {\\perp} ^ {2} + Z _ {\\perp} ^ {2} = X ^ {2} + (- Y \\sin \\alpha + Z \\cos \\alpha) ^ {2}, \\\\ = \\rho^ {2} \\left(\\hat {X} ^ {2} + (- \\hat {Y} \\sin \\alpha + \\hat {Z} \\cos \\alpha) ^ {2}\\right). \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.295, + 0.892, + 0.32 + ], + "angle": 0, + "content": "Furthermore, we can convert the relationship in Eq. (5) to a function \\( f \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.637, + 0.318, + 0.891, + 0.334 + ], + "angle": 0, + "content": "\\[\nf (\\tilde {\\mathbf {X}}, r _ {\\perp}, \\alpha) = \\rho , \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.892, + 0.445 + ], + "angle": 0, + "content": "where \\(r_{\\perp}\\) denotes a radius in the orthogonal coordinate. The function \\(f\\) represents that if we know slanted angle \\(\\alpha\\) and a radius \\(r_{\\perp}\\) at a certain pixel location on the fisheye domain, we can directly compute the corresponding depth value \\(\\rho\\). A more detailed explanation for Eq. (6) is available in the supplementary material. Using this relation, we implicitly model the slanted MCI representation inside the network as scene representation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.453, + 0.685, + 0.466 + ], + "angle": 0, + "content": "3.2. SlaBins framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.892, + 0.543 + ], + "angle": 0, + "content": "Based on the slanted MCI representation (cf. Sec. 3.1), the proposed SlaBins framework estimates an accurate and dense depth map. SlaBins consists of three modules: Encoder-decoder block, slanted angle predictor, and SlaBins module, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.892, + 0.709 + ], + "angle": 0, + "content": "Concretely, given the input fisheye image, the encoder-decoder block extracts a feature map containing useful information for depth prediction. Using the extracted feature map, the SlaBins module estimates several key information, such as adaptive bin-centers and per-pixel probability map in the slanted MCI representation (i.e., the orthogonal coordinate), using a transformer architecture. It should be worth noticing that by virtue of this SlaBins module, we can implicitly preserve the geometric relationship of objects on the road environments. Then, by combining with the slanted angle estimated by the slanted angle prediction module, we can directly regress a fisheye depth map." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.711, + 0.892, + 0.821 + ], + "angle": 0, + "content": "Encoder-decoder block. Given input fisheye image \\( x_{i} \\in \\mathbb{R}^{H \\times W \\times 3} \\), where \\( H \\) and \\( W \\) denote the height and the width of the input, our encoder-decoder block first extract features. We use the pre-trained EfficientNet B5 [37] as the backbone for the encoder and standard feature upsampling layers for the decoder, which returns decoded features \\( x_{d} \\in \\mathbb{R}^{h \\times w \\times C_{d}} \\), where \\( h = H / 2 \\), \\( w = W / 2 \\), and \\( C_{d} \\) indicates the feature dimension." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Slanted angle prediction module. We design compact MLP-based classification and regression modules that predict \\(\\alpha\\) from \\(x_{d}\\) in a two-step. The classifier first classifies" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.502, + 0.898 + ], + "angle": 0, + "content": "8768" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.124, + 0.457, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.37, + 0.473, + 0.472 + ], + "angle": 0, + "content": "Figure 4: The details of the SlaBins module. Our SlaBins module consists of two components: One is the mini-ViT block that extracts the range-attention map and bin widths, which are invariant to the slanted angle \\(\\alpha\\) from the decoded feature. The other is the SlaBins-aware regression. This module receives the outputs of mini-ViT and the slanted angle \\(\\alpha\\) as inputs and directly converts the orthogonally estimated depth \\(r_{\\perp}\\) for an arbitrary 3D point \\(\\mathbf{X}\\) into the depth \\(\\rho\\) of the spherical coordinate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.47, + 0.539 + ], + "angle": 0, + "content": "\\(\\alpha\\) in units of \\(5^{\\circ}\\), and the following regressor estimates the offset in the range of \\((-2.5^{\\circ}, 2.5^{\\circ})\\). This two-step angle prediction enables us to estimate an accurate viewing direction angle and train more efficiently (see Fig. 2c)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.47, + 0.651 + ], + "angle": 0, + "content": "Specifically, \\( x_{d} \\) is downsampled to \\( 1/4 \\) size with two \\( 3 \\times 3 \\) convolution layers and average pooling. Downsampled features are flattened (we denote the flattened feature as \\( x_{f} \\)) and \\( x_{f} \\) is used as input of classifier and regressor, which are composed of 3 FC layers followed by softmax and 4 FC layers, respectively. The output of the classifier is a 19-dimensional vector and that of the regressor is a normalized angle offset in the range of \\( (-1^{\\circ}, 1^{\\circ}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.47, + 0.74 + ], + "angle": 0, + "content": "SlaBins module. Given decoded features \\( x_{d} \\), the proposed SlaBins module first learns intermediate depth information, such as bin widths and probability maps in the slanted MCI representation, which are invariant to the viewing direction (i.e., slanted angle \\( \\alpha \\)), as shown in Fig. 4. To this end, we adopt mini-ViT (mViT) architecture in [3]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.47, + 0.868 + ], + "angle": 0, + "content": "The decoded features \\( x_{d} \\) are fed into the mViT with learnable positional encodings. To make a fixed-size patch-embedded input, we embed the decoded features to the \\( C_p \\) dimensional flattened tensor \\( x_{p}\\in \\mathbb{R}^{S\\times C_{p}} \\) using patch embedding before feeding into the mViT (here, \\( S \\) indicates \\( w\\times h / p^2 \\) and \\( p \\) denotes the size of patch and stride). Then, the sum of flattened tensor \\( x_{p} \\) and positional encodings are fed into the mViT and the output of the mViT has the same shape as input tensor \\( x_{o}\\in \\mathbb{R}^{S\\times C_{p}} \\). The first MLP head" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.126, + 0.892, + 0.251 + ], + "angle": 0, + "content": "of the output \\( x_{o} \\) means slanted bin widths and it is normalized into \\( b_{i} \\) for further depth regression. The other features followed by the \\( 1 \\times 1 \\) convolution layer perform pixel-wise dot product with a \\( 3 \\times 3 \\) convolution output of \\( x_{d} \\) to get the range-attention map. Softmax function with \\( 1 \\times 1 \\) convolution layer is used to represent the range-attention map as a probability of each pixel \\( p_{k} \\) in \\( N \\) bin-centers. These normalized slanted bin-width \\( b_{i} \\) and per-pixel probability \\( p_{k} \\) with \\( \\alpha \\) are used for SlaBins-aware regression." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.892, + 0.321 + ], + "angle": 0, + "content": "Our final depth \\(\\tilde{\\rho}\\) is computed by the linear combination of the two outputs from mViT (i.e., slanted bin-width \\(b_{i}\\) and per-pixel probability \\(p_k\\)) and the transformation function in Eq. (6). To compute final depth, slanted depth bin-centers \\(c^s\\) are computed with normalized bin-width \\(b_{i}\\) (i.e., \\(\\sum b_{i} = 1\\)):" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.334, + 0.892, + 0.353 + ], + "angle": 0, + "content": "\\[\nc ^ {s} \\left(b _ {i}\\right) = r _ {\\perp} ^ {\\min } + \\left(r _ {\\perp} ^ {\\max } - r _ {\\perp} ^ {\\min }\\right) \\left(b _ {i} / 2 + \\Sigma_ {j = 1} ^ {i - 1} b _ {j}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.892, + 0.428 + ], + "angle": 0, + "content": "where \\(r_{\\perp}^{min}\\) and \\(r_{\\perp}^{max}\\) mean the lower and upper bound of the radius in slanted MCI representation. Then, orthogonal coordinate radius \\(r_{\\perp}\\) is calculated using the cumulative sum of \\(c_k^s\\) and \\(p_k^s\\) (i.e., \\(\\sum_{k=1}^{N} c^s(b_k)p_k\\)) and transformed to our final depth \\(\\tilde{\\rho}\\) using Eq. (6):" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.429, + 0.892, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\rho} = f \\left(\\hat {\\mathbf {X}}, \\sum_ {k = 1} ^ {N} c ^ {s} \\left(b _ {k}\\right) p _ {k}, \\alpha\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.498 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{X}}\\) is a known normalized vector computed by pixel location and camera intrinsic, and \\(\\alpha\\) denotes the estimated slanted angle by the slanted angle predictor." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.505, + 0.64, + 0.518 + ], + "angle": 0, + "content": "3.3. Loss function" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.526, + 0.892, + 0.595 + ], + "angle": 0, + "content": "Here, we introduce a set of loss functions used for training SlaBins. Specifically, we compute loss in both orthogonal and spherical coordinates to make our SlaBins framework can estimate fisheye depth that has better discrimination between objects and road environments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.623 + ], + "angle": 0, + "content": "Bin loss. Following [3], we use Chamfer loss[7] to make adaptive bins:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.624, + 0.892, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {b i n s}} = \\operatorname {c h a m f e r} \\left(r _ {\\perp}, c ^ {s} (\\tilde {\\mathbf {b}})\\right) + \\operatorname {c h a m f e r} \\left(c ^ {s} (\\tilde {\\mathbf {b}})\\right), r _ {\\perp}), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.743 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{b}}\\) indicates the predicted slanted bin widths from the SlaBins module, and \\(r_{\\perp}\\) is the ground truth cylinder radius in the slanted MCI representation. In our case, to generate slanted bins (i.e., bins in the orthogonal coordinate), the ground-truth depth is converted to the ground truth radius \\(r_{\\perp}\\) at the orthogonal coordinate by Eq. (5) and the slanted MCI." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.746, + 0.892, + 0.774 + ], + "angle": 0, + "content": "Pixel-wise depth loss. We apply a scaled version of the Scale-Invariant (SI) log loss introduced by Eigen et al. [42]:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.778, + 0.892, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p i x e l}} = \\beta \\sqrt {\\frac {1}{T} \\Sigma_ {i} g _ {i} ^ {2} - \\frac {\\lambda}{T ^ {2}} \\left(\\Sigma_ {i} g _ {i}\\right) ^ {2}} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.868 + ], + "angle": 0, + "content": "where \\( g_{i} = \\log \\tilde{\\rho} - \\log \\rho \\) with the ground truth depth \\( \\rho \\) and predicted depth \\( \\tilde{\\rho} \\). \\( T \\) denotes the number of pixels containing valid ground truth values. Following [3], we set \\( \\lambda = 0.85 \\) and \\( \\beta = 10 \\) for all our experiments." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.502, + 0.898 + ], + "angle": 0, + "content": "8769" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.096, + 0.125, + 0.45, + 0.138 + ], + "angle": 0, + "content": "Table 1: Quantitative evaluation on SynWoodScape [34]." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.141, + 0.469, + 0.195 + ], + "angle": 0, + "content": "
MethodRMSE ↓RMSE log ↓Abs Rel ↓Sq Rel ↓δ1 ↑δ2 ↑δ3 ↑
OmniDet* [20]5.6230.7401.0803.1330.0310.0640.753
BTS [21]2.9980.3250.1790.4590.6530.8450.932
AdaBins [3]1.5240.1040.0700.1270.9710.9930.997
SlaBins (ours)1.0400.0560.0220.0550.9880.9970.999
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.469, + 0.308 + ], + "angle": 0, + "content": "Slanted angle loss. As mentioned in Sec. 3.2, we predict slanted angle \\(\\alpha\\) through a two-step module: classification and regression. For the classification module, we use the cross-entropy loss \\(\\mathcal{L}_{CE}\\). As for the regression module, we utilize the mean squared error (MSE) loss \\(\\mathcal{L}_{MSE}\\) between the estimated residual and the ground truth residual. We define the sum of the CE loss and MES loss as the slanted angle loss:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.308, + 0.469, + 0.322 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a n g l e}} = \\mathcal {L} _ {C E} + \\mathcal {L} _ {M S E}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.328, + 0.423, + 0.342 + ], + "angle": 0, + "content": "Total loss. Finally, the total loss for our problem is:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.351, + 0.469, + 0.366 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\gamma \\mathcal {L} _ {\\text {b i n s}} + \\mathcal {L} _ {\\text {p i x e l}} + \\mathcal {L} _ {\\text {a n g l e}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.373, + 0.414, + 0.387 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a balancing parameter (we set \\(\\gamma = 0.1\\))." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.399, + 0.208, + 0.415 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.423, + 0.469, + 0.492 + ], + "angle": 0, + "content": "We evaluate the proposed method from various perspectives. We introduce datasets and evaluation metrics in Sec. 4.1 and implementation details in Sec. 4.2. We then offer quantitative and qualitative comparisons to state-of-the-art methods in Sec. 4.3 and an ablation study in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.501, + 0.357, + 0.514 + ], + "angle": 0, + "content": "4.1. Datasets and evaluation metrics" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.522, + 0.469, + 0.604 + ], + "angle": 0, + "content": "For evaluation, we use two fisheye depth datasets: the SynWoodScape dataset [34], which is a synthetic dataset, and the KITTI-360 depth dataset, which is a modified dataset of real-world KITTI-360 [24] for our scenario. Note that we additionally augment two datasets for our purposes and will release these datasets for research communities." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.605, + 0.469, + 0.701 + ], + "angle": 0, + "content": "SynWoodScape dataset [34]. The SynWoodScape dataset is a synthetic surround-view fisheye camera dataset for autonomous driving, which is a synthetic version of the real-world WoodScape dataset \\([42]^2\\). This dataset provides pairs of synthetic fisheye images and depth maps for the front, left, right, and rear of the driving situation. Each view contains 500 images with \\(190^{\\circ}\\) FoV and a size of \\(1280\\times 966\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.702, + 0.469, + 0.812 + ], + "angle": 0, + "content": "Additionally, we augment the SynWoodScape dataset in terms of the viewing direction of fisheye cameras to validate the proposed framework. Specifically, we warp fisheye images and the corresponding depth maps by changing a total of nine viewing directions, of which rotation angles w.r.t. the x-axis are between \\(0^{\\circ}\\) and \\(90^{\\circ}\\) (between the viewing direction and the normal of the ground), including the original viewing direction. So, we have 18K fisheye images" + }, + { + "type": "table_caption", + "bbox": [ + 0.526, + 0.125, + 0.864, + 0.138 + ], + "angle": 0, + "content": "Table 2: Quantitative evaluation on KITTI-360 depth." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.141, + 0.892, + 0.195 + ], + "angle": 0, + "content": "
MethodRMSE ↓RMSE log ↓Abs Rel ↓Sq Rel ↓δ1 ↑δ2 ↑δ3 ↑
OmniDet* [20]1.9170.1500.0860.2460.9080.9750.992
BTS [21]1.7870.1530.0930.2430.9060.9760.992
AdaBins [3]1.5900.1290.0740.1850.9400.9840.994
SlaBins (ours)1.5370.1300.0770.1770.9410.9830.994
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.206, + 0.892, + 0.289 + ], + "angle": 0, + "content": "in total. We split the augmented dataset into 14.4K training, 1.8K validation, and 1.8K test samples. Following the previous work [16], we train the model and evaluate the range of depth up to \\(40\\mathrm{m}\\) because fisheye cameras undergo high data compression and can perform up to this range, unlike the KITTI dataset [10] covering up to \\(80\\mathrm{m}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.291, + 0.893, + 0.457 + ], + "angle": 0, + "content": "KITTI-360 depth dataset. The KITTI-360 dataset [24] is a suburban driving dataset for 2D/3D semantic and instance segmentation tasks that include various input modalities, semantic instance annotations, and precise localization. KITTI-360 provides fisheye images, raw LiDAR point clouds, and semantic and instance labels with \\(185^{\\circ}\\) FoV of left and right in the driving situation, with a resolution of \\(1400\\times 1400\\). Note that since KITTI-360 does not provide the ground truth for fisheye depth estimation, we generate the corresponding ground truth using raw LiDAR data. We call this modified dataset for fisheye depth estimation as KITTI-360 depth dataset." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.458, + 0.892, + 0.693 + ], + "angle": 0, + "content": "Concretely, we first make dense LiDAR by aggregating LiDAR data of twenty neighbor frames. We then project the dense LiDAR point clouds on the fisheye camera image domain and conduct occlusion handling [18] because of the displacement between the fisheye camera and the LiDAR (see Fig. 6). We then consider depth values, of which distance is less than \\(40\\mathrm{m}\\), as the ground truth. We sample 4K samples of the training set from sequence 0 to 7, 0.4K samples of the validation set from sequence 9, and 0.4K samples of the test set from sequence 10. To make dense depth maps, we resize samples to \\(700\\times 700\\). In addition, as we did for SynWoodScape, we augment the viewing directions of the KITTI-360 depth dataset; the angle between the viewing direction and the normal on the ground to \\(30^{\\circ}\\), \\(50^{\\circ}\\), \\(70^{\\circ}\\), and \\(90^{\\circ}\\) (the original viewing angle). So, we have 19.2K fisheye images in total. We split this augmented dataset into 16K training, 1.6K validation, and 1.6K test samples." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.697, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Evaluation metrics. We use the standard depth estimation metrics used in the previous work [21]. For error metrics, we calculate mean absolute relative error (Abs Rel), root mean square error (RMSE), root mean square log error (RMSE log), the root mean square relative error (Sq Rel). For accuracy metrics, we calculate threshold accuracy \\(\\delta_{i} = 1.25^{i}\\) for \\(i\\in \\{1,2,3\\}\\). Generally, the smaller the error, the better, and the higher the accuracy, the better." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.714, + 0.832 + ], + "angle": 0, + "content": "4.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.839, + 0.892, + 0.866 + ], + "angle": 0, + "content": "Our proposed method is implemented in PyTorch and trained on 4 RTX A6000 GPUs. Our model is trained for" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.82, + 0.469, + 0.843 + ], + "angle": 0, + "content": "2Unfortunately, the real-world WoodScape dataset [42] does not provide the ground truth depth data." + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.843, + 0.468, + 0.866 + ], + "angle": 0, + "content": "\\( {}^{3} \\) Currently,the SynWoodScape dataset is partially released to the public. We perform the experiments using the partial dataset." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.82, + 0.469, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.502, + 0.898 + ], + "angle": 0, + "content": "8770" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.209, + 0.116, + 0.293, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.296, + 0.267, + 0.306 + ], + "angle": 0, + "content": "RGB" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.117, + 0.387, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.305, + 0.295, + 0.386, + 0.307 + ], + "angle": 0, + "content": "OmniDet* [20]" + }, + { + "type": "image", + "bbox": [ + 0.397, + 0.117, + 0.481, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.295, + 0.464, + 0.306 + ], + "angle": 0, + "content": "BTS [21]" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.117, + 0.574, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.295, + 0.566, + 0.306 + ], + "angle": 0, + "content": "AdaBins [3]" + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.117, + 0.667, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.296, + 0.64, + 0.306 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.117, + 0.761, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.684, + 0.295, + 0.754, + 0.306 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.308, + 0.893, + 0.346 + ], + "angle": 0, + "content": "Figure 5: Qualitative comparison on the SynWoodScape dataset. Sky-blue boxes are enlarged views of specific objects to show detail. Overall, our method shows distinct boundaries compared to the other models. In particular, we can observe the detailed structure from our results, such as bicycle wheels in the first row and the boundary of two overlapped vehicles in the second row." + }, + { + "type": "image", + "bbox": [ + 0.169, + 0.353, + 0.761, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.607, + 0.893, + 0.634 + ], + "angle": 0, + "content": "Figure 6: Qualitative comparison on the KITTI-360 depth dataset. For visualization, we crop the object region in the fisheye image. In contrast to the other methods, our method shows clear boundaries near the object regions, as shown in sky-blue and green boxes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.647, + 0.47, + 0.745 + ], + "angle": 0, + "content": "10 epochs on SynWoodScape and 20 epochs on KITTI-360 depth using AdamW [28] optimizer with weight-decay \\(10^{-2}\\), with a batch size of 8. We use the same 1-cycle policy [35] as AdaBins [3] for the learning rate with max \\(\\mathrm{lr} = 3.5 \\times 10^{-4}\\). For the first \\(30\\%\\) of iterations, the linear learning rate warm-up technique is utilized. For the learning rate decay, the cosine annealing learning rate strategy is applied." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.76, + 0.21, + 0.775 + ], + "angle": 0, + "content": "4.3. Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.784, + 0.47, + 0.867 + ], + "angle": 0, + "content": "We compare our SlaBins with several baseline approaches: OmniDet [20], BTS [21], and AdaBins [3]. OmniDet is a fisheye depth estimation method using self-supervised learning. Based on the released code of the authors, we modify OmniDet to a supervised one for fairness (we denote this modified one OmniDet*). As base" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.647, + 0.892, + 0.73 + ], + "angle": 0, + "content": "line methods for pinhole cameras, we compare BTS and AdaBins, especially AdaBins is a bins-based fundamental method for depth estimation. It should be noted that we target depth estimation fisheye cameras, but we additionally compare it with conventional pinhole camera-based methods due to the lack of fisheye camera-based methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.738, + 0.892, + 0.849 + ], + "angle": 0, + "content": "Quantitative evaluation. Table 1 shows the quantitative comparisons on the SynWoodScape dataset. The arrow next to the metric indicates the direction of better performance. Overall, SlaBins outperforms the other baseline approaches [3, 20, 21]. In particular, RMSE and Abs-Rel errors are significantly improved; \\(31.7\\%\\) and \\(68.6\\%\\) decreased compared to AdaBins, which shows the effect of the slanted MCI representation in the fisheye domain." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.853, + 0.892, + 0.868 + ], + "angle": 0, + "content": "On the KITTI-360 depth dataset, our method still outper" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.501, + 0.899 + ], + "angle": 0, + "content": "8771" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.124, + 0.436, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.277, + 0.47, + 0.342 + ], + "angle": 0, + "content": "Figure 7: Depth histogram on the object regions. Left: RGB data (top) and the concept of depth histogram in the orthogonal coordinate (bottom). Middle: depth maps of only object regions for each model and the ground truth. Right: depth histograms (AdaBins, original MCI, slanted MCI, ground truth in that order)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.345, + 0.47, + 0.456 + ], + "angle": 0, + "content": "forms the comparison methods (see Table 2). Specifically, RMSE and Sq-Rel errors are improved; \\(13.9\\%\\) and \\(27.1\\%\\) decreased compared to BTS and \\(3.3\\%\\) and \\(4.3\\%\\) decreased compared to AdaBins. Note that even though the KITTI-360 depth dataset captured in suburban scenes does not perfectly fit our scenario of road environments, SlaBins still shows competitive performance, which demonstrates the effect of the proposed framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.47, + 0.709 + ], + "angle": 0, + "content": "Qualitative evaluation. Figure 5 shows the estimated depth maps for the test scenes on SynWoodScape. Due to the nature of fisheye cameras, we can observe ambiguous boundary regions in the comparison methods. Even AdaBins, which is our baseline and a naive MSI representation by the Euclidean distance of fisheye, shows bleeding depth estimation results near the boundaries between objects and ground regions. In contrast, thanks to the slanted MCI representation, SlaBins shows accurate depth results with distinct boundaries regardless of camera viewing directions. Specifically, we can observe that sharp objects with details (e.g., human legs and bike) are preserved in our method, unlike the other methods that smooth object details. In addition, we can observe a similar trend on the KITTI-360 depth dataset, as shown in Fig. 6. Compared to the baseline methods, SlaBins shows better depth consistency even on sharp objects (e.g., third and fourth columns in Fig. 6). Additional results are available in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.72, + 0.227, + 0.735 + ], + "angle": 0, + "content": "4.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.468, + 0.866 + ], + "angle": 0, + "content": "Multi-layered representation. We compare the proposed slanted MCI with the MSI used in AdaBins and the original MCI on the orthogonal coordinate. We measure the depth histogram for only object regions on the orthogonal coordinate. As shown in Fig. 7, compared to the other representations, SlaBins reveals the most similar trend (i.e., peak points) with the ground truth histogram. Furthermore, we compare slanted MCI and original MCI in terms of depth consistency. Figure 8 shows that the slanted MCI provides" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.123, + 0.875, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.229, + 0.892, + 0.267 + ], + "angle": 0, + "content": "Figure 8: Slanted MCI vs. original MCI in depth and clustering. All the results including the ground truth are converted to the XZ distance in the orthogonal coordinate." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.273, + 0.892, + 0.325 + ], + "angle": 0, + "content": "Table 3: Ablation study for the slanted angle predictor on SynWoodScape and KITTI-360 depth. We report the depth error (Abs Rel) and mean angle error. C and R denote classification-only and regression-only, respectively." + }, + { + "type": "table", + "bbox": [ + 0.561, + 0.327, + 0.835, + 0.389 + ], + "angle": 0, + "content": "
MethodSynWoodScapeKITTI-360 depth
Abs Rel ↓Angle error↓Abs Rel ↓Angle error
R0.07527.0310.60448.32
C0.0260.8420.1035
C+R (ours)0.0220.0310.0770.017
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.452 + ], + "angle": 0, + "content": "more consistent results in object regions both on depth and depth-based clustering results. Hence, we believe that the slanted MCI can provide robust geometric constraints for various tasks in fisheye domains." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.892, + 0.581 + ], + "angle": 0, + "content": "Slanted angle prediction. As shown in Table 3, we perform an ablation study for the angle prediction module. Unlike a conventional regression module having the advantage over classifications, in our case, it is hard to converge because of the trade-off with the depth estimation part in the SlaBins module. That is why regression-only module is poor on the angle prediction. On the contrary, our two-step prediction module, composed of classification and residual regression, shows the best performance with stable training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.591, + 0.618, + 0.606 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.892, + 0.766 + ], + "angle": 0, + "content": "We have proposed a novel fisheye depth estimation network, called SlaBins, based on the geometric property of the road environments. We newly introduce the slanted MCI representation, which allows us to represent a given road scene as a set of cylindrical layers orthogonal to the road ground. By virtue of the slanted MCI, our network can implicitly encode scene depth information invariant to the camera viewing direction while preserving boundary information of scene objects w.r.t. the road region. Thus, we achieve state-of-the-art performance on monocular fisheye depth estimation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.77, + 0.892, + 0.866 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported by 42dot Inc., Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2020-0-01336, Artificial Intelligence Graduate School Program (UNIST)) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF2021R1C1C1005723)." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.502, + 0.898 + ], + "angle": 0, + "content": "8772" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.125, + 0.173, + 0.139 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.148, + 0.468, + 0.198 + ], + "angle": 0, + "content": "[1] T-Bone Accident Statistics And Other Legal Information. https://www.injurytriallawyer.com/faqs/ how-dangerous-are-t-bone-accidents-.cfm. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.201, + 0.468, + 0.239 + ], + "angle": 0, + "content": "[2] Ashutosh Agarwal and Chetan Arora. Attention Attention Everywhere: Monocular Depth Prediction with Skip Attention. arXiv preprint arXiv:2210.09071, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.241, + 0.468, + 0.277 + ], + "angle": 0, + "content": "[3] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In CVPR, 2021. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.28, + 0.468, + 0.317 + ], + "angle": 0, + "content": "[4] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. LocalBins: Improving Depth Estimation by Learning Local Distributions. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.32, + 0.468, + 0.382 + ], + "angle": 0, + "content": "[5] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, et al. PersFormer: 3D Lane Detection via Perspective Transformer and the OpenLane Benchmark. arXiv preprint arXiv:2203.11089, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.384, + 0.468, + 0.422 + ], + "angle": 0, + "content": "[6] Liang-Chieh Chen, Huiyu Wang, and Siyuan Qiao. Scaling wide residual networks for panoptic segmentation. arXiv preprint arXiv:2011.11675, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.423, + 0.468, + 0.462 + ], + "angle": 0, + "content": "[7] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.463, + 0.468, + 0.514 + ], + "angle": 0, + "content": "[8] Olivier Faugeras and Quang-Tuan Luong. The geometry of multiple images: the laws that govern the formation of multiple images of a scene and some of their applications. MIT press, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.515, + 0.468, + 0.553 + ], + "angle": 0, + "content": "[9] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.555, + 0.468, + 0.592 + ], + "angle": 0, + "content": "[10] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.594, + 0.468, + 0.632 + ], + "angle": 0, + "content": "[11] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.468, + 0.671 + ], + "angle": 0, + "content": "[12] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.468, + 0.721 + ], + "angle": 0, + "content": "[13] Yuxuan Han, Ruicheng Wang, and Jiaolong Yang. Single-View View Synthesis in the Wild with Learned Adaptive Multiplane Images. arXiv preprint arXiv:2205.11733, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.724, + 0.468, + 0.762 + ], + "angle": 0, + "content": "[14] Adrian Johnston and Gustavo Carneiro. Self-supervised monocular trained depth estimation using self-attention and discrete disparity volume. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.763, + 0.468, + 0.814 + ], + "angle": 0, + "content": "[15] Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, and Junmo Kim. Global-Local Path Networks for Monocular Depth Estimation with Vertical Cut-Depth. arXiv preprint arXiv:2201.07436, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.468, + 0.867 + ], + "angle": 0, + "content": "[16] Varun Ravi Kumar, Sandesh Athni Hiremath, Markus Bach, Stefan Milz, Christian Witt, Clément Pinard, Senthil Yogamani, and Patrick Mäder. Fisheyedistancenet: Self-supervised scale-aware distance estimation using monocular" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.148, + 0.468, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.128, + 0.892, + 0.153 + ], + "angle": 0, + "content": "fisheye camera for autonomous driving. In ICRA, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.155, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[17] Varun Ravi Kumar, Marvin Klingner, Senthil Yogamani, Markus Bach, Stefan Milz, Tim Fingscheidt, and Patrick Mäder. SVDistNet: Self-supervised near-field distance estimation on surround view fisheye cameras. IEEE Transactions on Intelligent Transportation Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.218, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[18] Varun Ravi Kumar, Stefan Milz, Christian Witt, Martin Simon, Karl Amende, Johannes Petzold, Senthil Yogamani, and Timo Pech. Monocular fisheye camera depth estimation using sparse lidar supervision. In IEEE International Conference on Intelligent Transportation Systems (ITSC), 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.355 + ], + "angle": 0, + "content": "[19] Varun Ravi Kumar, Senthil Yogamani, Markus Bach, Christian Witt, Stefan Milz, and Patrick Mäder. Unrectdepthnet: Self-supervised monocular depth estimation using a generic framework for handling common camera distortion models. In IROS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.357, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[20] Varun Ravi Kumar, Senthil Yogamani, Hazem Rashed, Ganesh Sitsu, Christian Witt, Isabelle Leang, Stefan Milz, and Patrick Mader. Omnidet: Surround view cameras based multi-task visual perception network for autonomous driving. IEEE Robotics and Automation Letters, 6(2):2830-2837, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[21] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.522 + ], + "angle": 0, + "content": "[22] Jiaxin Li, Zijian Feng, Qi She, Henghui Ding, Changhu Wang, and Gim Hee Lee. Mine: Towards continuous depth spi with nerf for novel view synthesis. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.523, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[23] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. DepthFormer: Exploiting Long-Range Correlation and Local Information for Accurate Monocular Depth Estimation. arXiv preprint arXiv:2203.14211, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.574, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[24] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE TPAMI, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.612, + 0.892, + 0.662 + ], + "angle": 0, + "content": "[25] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETRv2: A Unified Framework for 3D Perception from Multi-Camera Images. arXiv preprint arXiv:2206.01256, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.663, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[26] Yuxuan Liu, Yuan Yixuan, and Ming Liu. Ground-aware monocular 3d object detection for autonomous driving. IEEE Robotics and Automation Letters, 6(2):919-926, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.701, + 0.892, + 0.75 + ], + "angle": 0, + "content": "[27] Xiaoxiao Long, Cheng Lin, Lingjie Liu, Wei Li, Christian Theobalt, Ruigang Yang, and Wenping Wang. Adaptive surface normal constraint for depth estimation. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.752, + 0.892, + 0.776 + ], + "angle": 0, + "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In *ICLR*, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.778, + 0.892, + 0.827 + ], + "angle": 0, + "content": "[29] Reza Mahjourian, Martin Wicke, and Anelia Angelova. Unsupervised learning of depth and ego-motion from monocular video using 3d geometric constraints. In CVPR, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.829, + 0.892, + 0.867 + ], + "angle": 0, + "content": "[30] Vaishakh Patil, Christos Sakaridis, Alexander Liniger, and Luc Van Gool. P3Depth: Monocular Depth Estimation with a Piecewise Planarity Prior. In CVPR, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.128, + 0.892, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.501, + 0.898 + ], + "angle": 0, + "content": "8773" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.127, + 0.469, + 0.165 + ], + "angle": 0, + "content": "[31] Elad Plaut, Erez Ben Yaacov, and Bat El Shlomo. 3d object detection from a single fisheye image without a single fisheye training image. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.167, + 0.469, + 0.192 + ], + "angle": 0, + "content": "[32] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.469, + 0.243 + ], + "angle": 0, + "content": "[33] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.245, + 0.469, + 0.308 + ], + "angle": 0, + "content": "[34] Ahmed Rida Sekkat, Yohan Dupuis, Varun Ravi Kumar, Hazem Rashed, Senthil Yogamani, Pascal Vasseur, and Paul Honeine. SynWoodScape: Synthetic Surround-View Fisheye Camera Dataset for Autonomous Driving. IEEE Robotics and Automation Letters, 7(3):8502-8509, jul 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.31, + 0.469, + 0.372 + ], + "angle": 0, + "content": "[35] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, volume 11006, pages 369-386. SPIE, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.374, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[36] Richard Szeliski and Polina Golland. Stereo matching with transparency and matting. In ICCV, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.401, + 0.469, + 0.45 + ], + "angle": 0, + "content": "[37] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.452, + 0.469, + 0.49 + ], + "angle": 0, + "content": "[38] Andrew Tao, Karan Sapra, and Bryan Catanzaro. Hierarchical multi-scale attention for semantic segmentation. arXiv preprint arXiv:2005.10821, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[39] Richard Tucker and Noah Snively. Single-view view synthesis with multiplane images. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.469, + 0.568 + ], + "angle": 0, + "content": "[40] Qingan Yan, Pan Ji, Nitin Bansal, Yuxin Ma, Yuan Tian, and Yi Xu. FisheyeDistill: Self-Supervised Monocular Depth Estimation with Ordinal Distillation for Fisheye Cameras. arXiv preprint arXiv:2205.02930, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.569, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[41] Yaozu Ye, Kailun Yang, Kaite Xiang, Juan Wang, and Kaiwei Wang. Universal semantic segmentation for fisheye urban driving images. In IEEE International Conference on Systems, Man, and Cybernetics (SMC), pages 648-655, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.697 + ], + "angle": 0, + "content": "[42] Senthil Yogamani, Ciarán Hughes, Jonathan Horgan, Ganesh Sistu, Padraig Varley, Derek O'Dea, Michal Uricár, Stefan Milz, Martin Simon, Karl Amende, et al. WoodScape: A multi-task, multi-camera fisheye dataset for autonomous driving. arXiv preprint arXiv:1905.01489, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.698, + 0.469, + 0.748 + ], + "angle": 0, + "content": "[43] Yifan Zhang, Qijian Zhang, Zhiyu Zhu, Junhui Hou, and Yixuan Yuan. GLENet: Boosting 3D Object Detectors with Generative Label Uncertainty Estimation. arXiv preprint arXiv:2207.02466, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.75, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[44] Dingfu Zhou, Jin Fang, Xibin Song, Liu Liu, Junbo Yin, Yuchao Dai, Hongdong Li, and Ruigang Yang. Joint 3d instance segmentation and object detection for autonomous driving. In CVPR, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.469, + 0.801 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.501, + 0.899 + ], + "angle": 0, + "content": "8774" + } + ] +] \ No newline at end of file diff --git a/2023/SlaBins_ Fisheye Depth Estimation using Slanted Bins on Road Environments/2f72858d-bf2e-45dd-b589-9c5d6578e810_origin.pdf b/2023/SlaBins_ Fisheye Depth Estimation using Slanted Bins on Road Environments/2f72858d-bf2e-45dd-b589-9c5d6578e810_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..92a7e79f82238131d47c8ff311de66e618c23629 --- /dev/null +++ b/2023/SlaBins_ Fisheye Depth Estimation using Slanted Bins on Road Environments/2f72858d-bf2e-45dd-b589-9c5d6578e810_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46074cd449f40be3cf5f2f3dd4361cffa07ce977def2dd495ad4397f6f2be598 +size 8866670 diff --git a/2023/SlaBins_ Fisheye Depth Estimation using Slanted Bins on Road Environments/full.md b/2023/SlaBins_ Fisheye Depth Estimation using Slanted Bins on Road Environments/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6d1101a75d98f2aa025d90fd9fc7749096ceaef1 --- /dev/null +++ b/2023/SlaBins_ Fisheye Depth Estimation using Slanted Bins on Road Environments/full.md @@ -0,0 +1,357 @@ +# SlaBins: Fisheye Depth Estimation using Slanted Bins on Road Environments + +Jongsung Lee $^{1}$ , Gyeongsu Cho $^{1*}$ , Jeongin Park $^{1*}$ , Kyongjun Kim $^{1*}$ , Seongoh Lee $^{1*}$ , Jung-Hee Kim $^{2}$ , Seong-Gyun Jeong $^{2}$ , Kyungdon Joo $^{1\dagger}$ + +1Artificial Intelligence Graduate School, UNIST, 242dot Inc. + +{syniez, threedv, jeonginpark, kimkj38, solee, kyungdon}@unist.ac.kr + +{junghee.kim, seonggyun.jeong}@42dot.ai + +# Abstract + +Although 3D perception for autonomous vehicles has focused on frontal-view information, more than half of fatal accidents occur due to side impacts in practice (e.g., T-bone crash). Motivated by this fact, we investigate the problem of side-view depth estimation, especially for monocular fisheye cameras, which provide wide FoV information. However, since fisheye cameras head road areas, it observes road areas mostly and results in severe distortion on object areas, such as vehicles or pedestrians. To alleviate these issues, we propose a new fisheye depth estimation network, SlaBins, that infers an accurate and dense depth map based on a geometric property of road environments; most objects are standing (i.e., orthogonal) on the road environments. Concretely, we introduce a slanted multi-cylindrical image (MCI) representation, which allows us to describe a distance as a radius to a cylindrical layer orthogonal to the ground regardless of the camera viewing direction. Based on the slanted MCI, we estimate a set of adaptive bins and a per-pixel probability map for depth estimation. Then by combining it with the estimated slanted angle of viewing direction, we directly infer a dense and accurate depth map for fisheye cameras. Experiments demonstrate that SlaBins outperforms the state-of-the-art methods in both qualitative and quantitative evaluation on the SynWoodScape and KITTI-360 depth datasets. For more information, you can visit our project page https://syniez.github.io/SlaBins/. + +# 1. Introduction + +With the advent of autonomous driving, the importance of 3D perception keeps growing for safety [20, 25, 44]. For example, ADAS systems pervading our daily driving already, such as lane-keeping aid and forward collision + +![](images/d7972720d786bf5df51d9fa598f8e11245287d358ff6e72fae4804fcebe9ef92.jpg) +RGB + +![](images/37b580a71bf49cf0fecfaa948c403f57b7cba1bbeb5510e580a9b0bc794b9dd2.jpg) + +![](images/d0b8520def059b3be9dcfe75b01395c6a8eaeaf52ce61e53aa75d7cb318791cc.jpg) +Figure 1: SlaBins on the SynWoodScape dataset [34]. SlaBins predicts dense and accurate depth maps regardless of the viewing directions using the slanted MCI representation. + +![](images/8d8f184aef1f234e2fe02f96a914794b6d7f4134d100206e1769b7b1574c0dea.jpg) +Slanted MCI representation + +![](images/7f96553eab508578605a2bb78c6bb6d48249c691d76b811d8ed157b55c3244a7.jpg) + +![](images/1dba10f1d3bc1670f20779959c64f817370d99ce9fc28790171071aafd506e23.jpg) +Depth + +![](images/1916eeae437689dd6e45be8389dd44a8a7744130cde9b6863669a882614c7778.jpg) + +![](images/1197552c0da8cc880c7bd38bfff1fe4a1e72ccb4d59ea085403bfc4b563b74b4.jpg) + +avoidance assist, depend on various sensor modalities to observe 3D information [43, 5]. Here, one interesting point is that various sensor modalities for 3D perception, including many datasets related to autonomous driving, are focusing on the frontal view predominantly [43, 38, 2, 6]. We can easily deduce the reason by our common sense that it is important to keep an eye on the front while driving. However, according to the National Transportation Highway Safety Administration [1], $51\%$ of fatal car accidents correspond to T-bone crashes in which one car hits the side of another, which is the second highest type of accident. This statistic supports that side-view 3D perception is important to assist drivers because they are hard to pay attention to the side view due to the physical characteristics of the people looking ahead and their position characteristics. + +Motivated by this fact, 3D perception for side-view, especially fisheye-based depth estimation, has started to gain the spotlight [31, 16, 41]. Fisheye cameras have several advantages for 3D perception in road scenes. Specifically, fisheye cameras provide broader FoV information (typically greater than $180^{\circ}$ ) in a single image, which allows us to observe road information (e.g., road, vehicles, and pedestrians) at a low cost. In addition, we can directly utilize perspective camera 3D perception approaches because mathematical mutual conversion between two camera models is possible. Thus, if we can obtain side-view depth informa + +tion using a single fisheye camera, we can efficiently handle various vision tasks for autonomous driving. + +However, despite the advantages of fisheye cameras, monocular fisheye camera-based side-view depth estimation is a challenging problem due to significant distortion. In addition to inherent issues of monocular depth estimation, such as scale ambiguity [8], severe distortion of fisheye cameras makes depth estimation more difficult by distorting scene objects (e.g., close objects look closer and distant objects look more distant abnormally). Particularly, this distortion problem could be more severe in road environments because fisheye cameras for ADAS systems generally look to the road areas, and vital objects such as vehicles and pedestrians are located at the edge of the image. That is, most regions of the image correspond to the road areas and objects contain very few pixels, so depth estimation becomes more challenging. + +To alleviate this problem, we pay attention to the geometric property of road environments; most objects are standing on the road environments. Based on this geometric characteristic, we can easily imagine the visible region of each object as a planar segment orthogonal to the road region. This approximation is connected to a multi-plane image (MPI) concept [36] that represents a given scene as the composition of multi-layered images according to depth. Note that one significant difference is that a fisheye camera looks down to the road region in our case (i.e., camera viewing direction is not parallel to the ground), unlike a camera points to a given scene in the general MPI concept. Thus, a naive MPI concept is not directly applicable. + +In this work, we propose a new fisheye depth estimation framework, SlaBins, that exploits the geometric property of the road environments (see Fig. 1). To be specific, we adopt a multi-cylindrical image (MCI) representation; especially, we introduce a slanted MCI of which the cylindrical layer is defined as orthogonal to the road ground regardless of camera viewing direction. Based on the slanted MCI, we estimate adaptive bins and the per-pixel probability of the slanted cylindrical layers, which allows us to implicitly preserve the geometric property of the road environments and improve the depth quality near the boundary between objects and the ground region. In addition, we independently estimate the slanted angle of the camera viewing direction w.r.t. the road ground. Then, by combining this angle with the estimated depth information in the slanted MCI domain, we can directly compute a dense and accurate depth map for fisheye cameras. We validate the proposed SlaBins framework on the SynWoodScape dataset [34] and KITTI-360 depth dataset (a modified dataset of KITTI + +360 [24] for our scenario), where SlaBins outperforms the baseline methods [3, 20, 21] and shows potential as scene representations for downstream tasks, e.g., segmentation. + +In a nutshell, our contributions are as follows: + +- We propose a new fisheye depth estimation framework, SlaBins, which infers dense depth based on the geometric property of road environments. +- We newly introduce a slanted MCI representation as an intermediate road scene representation, which allows us to describe a given scene regardless of viewing direction and improve the depth quality on the boundary areas between objects and the road. +- By decoupling the slanted angle of an input image, we force the proposed model to learn the slanted MCI depth information and then seamlessly compute a fisheye depth map in a slanted-aware regression way. + +# 2. Related Work + +Monocular depth estimation. With the advent of breakthroughs in deep neural networks, such as CNN and transformer architecture, monocular depth estimation has shown impressive performance [15, 33, 21, 32, 23, 18, 40, 11, 9, 12, 29, 19, 16, 17, 20]. As one of the recent trends, several methods formulate a depth estimation task as a combination of classification and regression problems, which enables training the network more easily and shows stable performance [14, 3, 4]. For this formulation, they commonly use an MPI according to depth, which is popularly used in novel-view synthesis tasks [22, 39, 13]. The MPI notation represents depth as quantized depth values w.r.t. the uniformly quantized planes and residuals from mapped planes. + +To improve the uniformly quantized planes by uniform bins, recent research works [3, 4] present a network that predicts adaptive bins depending on the input scene and local regions in the image. AdaBins [3] computes adaptive bin-centers and per-pixel probabilities relying on each input image. Inspired by AdaBins, LocalBins [4] improves depth quality by considering both the global histogram and local-region histograms of each input image. Nevertheless, the predicted depth from the above models also depends on the camera's pose because they predict depth parallel with the camera's principal axis. This constraint makes it hard to adopt the previous models into fisheye depth estimation, which has a large FoV and uses the Euclidean distance instead of depth. In this work, we introduce and exploit a slanted MCI representation that is invariant to the viewing angle of fisheye cameras w.r.t. road environments. + +Fisheye depth estimation. Despite increasing demand for fisheye cameras in the industry, like ADAS systems, there are only a few previous research works in fisheye depth estimation fields [20, 17, 19, 16, 40]. Kumar et al. [19] propose + +![](images/5edfc6aa98a4cee7db2ed3927bd6007ed4962c651216163c8bef773ce430c348.jpg) +Figure 2: Overview of the proposed SlaBins. (a) Given an input fisheye RGB image, SlaBins framework first extracts features by the encoder-decoder block. The extracted features are used as input to the slanted angle prediction module and the SlaBins module to get a depth image. (b) Illustration of the coordinates and MCI representations defined in our model. (c) Illustration of the slanted angle predictor. + +a camera-aware depth estimation network to handle the severe distortion of fisheye cameras. They encode the camera's intrinsic parameters as a tensor named camera geometry tensor and use it in the training step for camera-aware depth estimation in fisheye images. With camera geometry tensor, they propose a multi-task network for fisheye cameras, called OmniDet [20], that simultaneously estimates depth estimation, motion, semantic segmentation, and object detection for synergy. However, since OmniDet relies on the partial region in the whole image, they may miss information on image edges. Using a whole image, Yan et al. [40] distill depth ordering information from a teacher network trained on the large-scale dataset to train in fisheye images, but FoV they cover is not large (e.g., vFoV=66°). In contrast to previous approaches, we propose a fisheye depth estimation network that predicts fisheye image depth without any information loss such as crop, and narrow FoV. + +Geometric cues for depth estimation. One way to improve fisheye depth estimation is to exploit geometric cues between the camera and the world. Liu et al. [26] use a ground plane to get depth priors using camera elevation and camera parameters. Moreover, Mahjourian et al. [29] predict camera ego-motion and compute loss on both the 3D point cloud domain and the 2D image domain. In addition to driving scenes, several works use plane coefficients to give consistency on plane regions in depth estimation [30, 27]. Patil et al. [30] estimate plane coefficients instead of depth and then combine plane coefficients into depth to use plane constraints (e.g., co-planarity, plane normals) in 3D space. Inspired by the previous works [3, 30], we design a network that estimates a dense and accurate depth map for fisheye cameras based on the slanted MCI representation using some geometric constraints (i.e., orthogonality between ground and object). + +# 3. Proposed Framework + +Inspired by the geometric property of the road environments, we propose a new fisheye depth estimation frame + +work (see Fig. 2 for the overview). Specifically, we introduce a slanted MCI representation, which allows us to describe a multi-layer cylindrical image orthogonal to the road ground regardless of the viewing direction of the camera. Based on the slanted MCI representation, our SlaBins module estimates the adaptive bin widths and per-pixel probability map in the orthogonal coordinate, which provides depth information invariant to the camera viewing direction. We then combine the estimated depth information with the slanted angle estimated by the slanted angle prediction module. Through this process, we can directly estimate a dense depth map for fisheye cameras. Our approach is built upon the previous AdaBins framework [3]. So, we call our method SlaBins. + +# 3.1. Slanted MCI Representation + +As intermediate scene representation, we use multi-layer images for fisheye depth estimation. Here, we briefly compare two multi-layer image representations for fisheye images: multi-spherical image (MSI) and multi-cylindrical image (MCI). We then present a slanted MCI representation, which is a suitable representation of our problem. + +MSI vs. MCI. By nature of fisheye cameras covering over $180^{\circ}$ , it is common to use the Euclidean distance as its depth; that is, a layer with the same depth value forms a sphere in the fisheye domain. Thus, we can represent a multi-layer image for fisheye cameras as an MSI in the spherical coordinate, which is a naive representation. However, this naive MSI representation has inherent limitations for fisheye depth estimation. We observed that such a representation can result in ambiguous depth boundaries between the ground and object regions, as shown in Fig. 3a. It may neutralize the benefit of multi-layer images by separating an object into two different layers, for example. + +In the MCI representation, a cylindrical layer encodes the same Euclidean distance w.r.t. the y-axis. In other words, we can represent an object, such as a pedestrian on the road, in a single cylindrical layer (see Fig. 3b). Further + +![](images/ea4262433bf086d7a8b2ac0379dc44babe6a81a7aeb9e88c04f5fa8d519bb004.jpg) +(a) + +![](images/dead83876114f702840b50d1018e067c973d44ec28768124628d0242097c6da9.jpg) +(b) + +![](images/6d5259420a93cc5f1427c3e76bedce9f962d1879a1890b65f3edee92f392fd9b.jpg) +(c) +Figure 3: MSI vs. MCI representation. Illustration of an example scene and zoomed results with (a) MSI and (b) MCI, where we can observe more accurate depth in the slanted MCI representation near the boundary between pedestrians and the ground. (c) Relationship between the original cylindrical layer $C_r^o$ (gray-colored cylinder) in the world coordinate (arrows) and the slanted cylindrical layer $C_{r_\perp}^s$ (orange-colored cylinders) in the orthogonal coordinate (dashed arrows). + +more, we can directly exploit the geometric properties of the road environment: most objects are standing (i.e., orthogonal) on the road environment. For this reason, we adopt the MCI as our intermediate scene representation. + +Slanted MCI representation. Let $\mathbf{X}$ be a 3D point in the world coordinate. Then, we can represent $\mathbf{X}$ according to the coordinate systems: + +$$ +\mathbf {X} = \left[ \begin{array}{l} X \\ Y \\ Z \end{array} \right] = \rho \left[ \begin{array}{c} \hat {X} \\ \hat {Y} \\ \hat {Z} \end{array} \right] = \rho \left[ \begin{array}{c} \sin \theta \cos \phi \\ \sin \theta \sin \phi \\ \cos \theta \end{array} \right], \tag {1} +$$ + +where $\hat{\mathbf{X}} = [\hat{X},\hat{Y},\hat{Z}]^{\top}$ is the normalized unit vector (i.e., ray direction), $\rho$ indicates the Euclidean distance (i.e., depth in the fisheye domain), and $\theta$ and $\phi$ denote the elevation and azimuth angles, respectively. + +Based on the cylindrical representation, we can form a cylindrical layer according to a radius $r$ as: + +$$ +C _ {r} ^ {o} = \left\{\mathbf {X} \mid X ^ {2} + Z ^ {2} = r ^ {2}; \mathbf {X} \in \mathbb {R} ^ {3} \right\}. \tag {2} +$$ + +We call $C_r^o$ the original cylindrical layer at $r$ . This original MCI is inherently suitable for road environments and shows less distortion than MSI on ambiguous depth boundaries. However, depending on the principal axis of cameras (i.e., viewing direction) w.r.t. the road environments, the original MCI may hinder layered scene representation like the MSI. + +To alleviate this issue, we consider the 1D angle $\alpha$ between the principal axis and the road environments. Using $\alpha$ , we can implicitly describe the slanted MCI representation that is orthogonal to the road areas (see Fig. 3c). To this end, let $\mathbf{X}_{\perp}$ be a 3D point in the orthogonal coordinate, of which the XZ plane is parallel to the ground. Similarly to Eq. (2), we can form a cylindrical layer according to a radius $r_{\perp}$ in the orthogonal coordinate as: + +$$ +C _ {r _ {\perp}} ^ {s} = \left\{\mathbf {X} _ {\perp} \mid X _ {\perp} ^ {2} + Z _ {\perp} ^ {2} = r _ {\perp} ^ {2}; \mathbf {X} _ {\perp} \in \mathbb {R} ^ {3} \right\}, \tag {3} +$$ + +which corresponds to a slanted cylindrical layer in the world (camera) coordinate (see Fig. 2b for coordinate systems). + +By $\alpha$ , we can derive the relationship between $\mathbf{X}_{\perp}$ in the orthogonal coordinate and $\mathbf{X}$ in the world coordinate by: + +$$ +\mathbf {X} _ {\perp} = \left[ \begin{array}{l} X _ {\perp} \\ Y _ {\perp} \\ Z _ {\perp} \end{array} \right] = \left[ \begin{array}{c} X \\ Y \cos \alpha + Z \sin \alpha \\ - Y \sin \alpha + Z \cos \alpha \end{array} \right]. \tag {4} +$$ + +Then, we can deduce the following relationship between a 3D point $\mathbf{X}_{\perp}$ on $C_{r_{\perp}}^{s}$ and its spherical representation in the camera coordinate by using Eqs. (3) and (4): + +$$ +\begin{array}{l} r _ {\perp} ^ {2} = X _ {\perp} ^ {2} + Z _ {\perp} ^ {2} = X ^ {2} + (- Y \sin \alpha + Z \cos \alpha) ^ {2}, \\ = \rho^ {2} \left(\hat {X} ^ {2} + (- \hat {Y} \sin \alpha + \hat {Z} \cos \alpha) ^ {2}\right). \tag {5} \\ \end{array} +$$ + +Furthermore, we can convert the relationship in Eq. (5) to a function $f$ as: + +$$ +f (\tilde {\mathbf {X}}, r _ {\perp}, \alpha) = \rho , \tag {6} +$$ + +where $r_{\perp}$ denotes a radius in the orthogonal coordinate. The function $f$ represents that if we know slanted angle $\alpha$ and a radius $r_{\perp}$ at a certain pixel location on the fisheye domain, we can directly compute the corresponding depth value $\rho$ . A more detailed explanation for Eq. (6) is available in the supplementary material. Using this relation, we implicitly model the slanted MCI representation inside the network as scene representation. + +# 3.2. SlaBins framework + +Based on the slanted MCI representation (cf. Sec. 3.1), the proposed SlaBins framework estimates an accurate and dense depth map. SlaBins consists of three modules: Encoder-decoder block, slanted angle predictor, and SlaBins module, as shown in Fig. 2. + +Concretely, given the input fisheye image, the encoder-decoder block extracts a feature map containing useful information for depth prediction. Using the extracted feature map, the SlaBins module estimates several key information, such as adaptive bin-centers and per-pixel probability map in the slanted MCI representation (i.e., the orthogonal coordinate), using a transformer architecture. It should be worth noticing that by virtue of this SlaBins module, we can implicitly preserve the geometric relationship of objects on the road environments. Then, by combining with the slanted angle estimated by the slanted angle prediction module, we can directly regress a fisheye depth map. + +Encoder-decoder block. Given input fisheye image $x_{i} \in \mathbb{R}^{H \times W \times 3}$ , where $H$ and $W$ denote the height and the width of the input, our encoder-decoder block first extract features. We use the pre-trained EfficientNet B5 [37] as the backbone for the encoder and standard feature upsampling layers for the decoder, which returns decoded features $x_{d} \in \mathbb{R}^{h \times w \times C_{d}}$ , where $h = H / 2$ , $w = W / 2$ , and $C_{d}$ indicates the feature dimension. + +Slanted angle prediction module. We design compact MLP-based classification and regression modules that predict $\alpha$ from $x_{d}$ in a two-step. The classifier first classifies + +![](images/f9ffadc1450b2b43272998450f0ed9d8f7682ceadba8466503b64bf7bc2dd07f.jpg) +Figure 4: The details of the SlaBins module. Our SlaBins module consists of two components: One is the mini-ViT block that extracts the range-attention map and bin widths, which are invariant to the slanted angle $\alpha$ from the decoded feature. The other is the SlaBins-aware regression. This module receives the outputs of mini-ViT and the slanted angle $\alpha$ as inputs and directly converts the orthogonally estimated depth $r_{\perp}$ for an arbitrary 3D point $\mathbf{X}$ into the depth $\rho$ of the spherical coordinate. + +$\alpha$ in units of $5^{\circ}$ , and the following regressor estimates the offset in the range of $(-2.5^{\circ}, 2.5^{\circ})$ . This two-step angle prediction enables us to estimate an accurate viewing direction angle and train more efficiently (see Fig. 2c). + +Specifically, $x_{d}$ is downsampled to $1/4$ size with two $3 \times 3$ convolution layers and average pooling. Downsampled features are flattened (we denote the flattened feature as $x_{f}$ ) and $x_{f}$ is used as input of classifier and regressor, which are composed of 3 FC layers followed by softmax and 4 FC layers, respectively. The output of the classifier is a 19-dimensional vector and that of the regressor is a normalized angle offset in the range of $(-1^{\circ}, 1^{\circ})$ . + +SlaBins module. Given decoded features $x_{d}$ , the proposed SlaBins module first learns intermediate depth information, such as bin widths and probability maps in the slanted MCI representation, which are invariant to the viewing direction (i.e., slanted angle $\alpha$ ), as shown in Fig. 4. To this end, we adopt mini-ViT (mViT) architecture in [3]. + +The decoded features $x_{d}$ are fed into the mViT with learnable positional encodings. To make a fixed-size patch-embedded input, we embed the decoded features to the $C_p$ dimensional flattened tensor $x_{p}\in \mathbb{R}^{S\times C_{p}}$ using patch embedding before feeding into the mViT (here, $S$ indicates $w\times h / p^2$ and $p$ denotes the size of patch and stride). Then, the sum of flattened tensor $x_{p}$ and positional encodings are fed into the mViT and the output of the mViT has the same shape as input tensor $x_{o}\in \mathbb{R}^{S\times C_{p}}$ . The first MLP head + +of the output $x_{o}$ means slanted bin widths and it is normalized into $b_{i}$ for further depth regression. The other features followed by the $1 \times 1$ convolution layer perform pixel-wise dot product with a $3 \times 3$ convolution output of $x_{d}$ to get the range-attention map. Softmax function with $1 \times 1$ convolution layer is used to represent the range-attention map as a probability of each pixel $p_{k}$ in $N$ bin-centers. These normalized slanted bin-width $b_{i}$ and per-pixel probability $p_{k}$ with $\alpha$ are used for SlaBins-aware regression. + +Our final depth $\tilde{\rho}$ is computed by the linear combination of the two outputs from mViT (i.e., slanted bin-width $b_{i}$ and per-pixel probability $p_k$ ) and the transformation function in Eq. (6). To compute final depth, slanted depth bin-centers $c^s$ are computed with normalized bin-width $b_{i}$ (i.e., $\sum b_{i} = 1$ ): + +$$ +c ^ {s} \left(b _ {i}\right) = r _ {\perp} ^ {\min } + \left(r _ {\perp} ^ {\max } - r _ {\perp} ^ {\min }\right) \left(b _ {i} / 2 + \Sigma_ {j = 1} ^ {i - 1} b _ {j}\right), \tag {7} +$$ + +where $r_{\perp}^{min}$ and $r_{\perp}^{max}$ mean the lower and upper bound of the radius in slanted MCI representation. Then, orthogonal coordinate radius $r_{\perp}$ is calculated using the cumulative sum of $c_k^s$ and $p_k^s$ (i.e., $\sum_{k=1}^{N} c^s(b_k)p_k$ ) and transformed to our final depth $\tilde{\rho}$ using Eq. (6): + +$$ +\tilde {\rho} = f \left(\hat {\mathbf {X}}, \sum_ {k = 1} ^ {N} c ^ {s} \left(b _ {k}\right) p _ {k}, \alpha\right), \tag {8} +$$ + +where $\hat{\mathbf{X}}$ is a known normalized vector computed by pixel location and camera intrinsic, and $\alpha$ denotes the estimated slanted angle by the slanted angle predictor. + +# 3.3. Loss function + +Here, we introduce a set of loss functions used for training SlaBins. Specifically, we compute loss in both orthogonal and spherical coordinates to make our SlaBins framework can estimate fisheye depth that has better discrimination between objects and road environments. + +Bin loss. Following [3], we use Chamfer loss[7] to make adaptive bins: + +$$ +\mathcal {L} _ {\text {b i n s}} = \operatorname {c h a m f e r} \left(r _ {\perp}, c ^ {s} (\tilde {\mathbf {b}})\right) + \operatorname {c h a m f e r} \left(c ^ {s} (\tilde {\mathbf {b}})\right), r _ {\perp}), \tag {9} +$$ + +where $\tilde{\mathbf{b}}$ indicates the predicted slanted bin widths from the SlaBins module, and $r_{\perp}$ is the ground truth cylinder radius in the slanted MCI representation. In our case, to generate slanted bins (i.e., bins in the orthogonal coordinate), the ground-truth depth is converted to the ground truth radius $r_{\perp}$ at the orthogonal coordinate by Eq. (5) and the slanted MCI. + +Pixel-wise depth loss. We apply a scaled version of the Scale-Invariant (SI) log loss introduced by Eigen et al. [42]: + +$$ +\mathcal {L} _ {\text {p i x e l}} = \beta \sqrt {\frac {1}{T} \Sigma_ {i} g _ {i} ^ {2} - \frac {\lambda}{T ^ {2}} \left(\Sigma_ {i} g _ {i}\right) ^ {2}} \tag {10} +$$ + +where $g_{i} = \log \tilde{\rho} - \log \rho$ with the ground truth depth $\rho$ and predicted depth $\tilde{\rho}$ . $T$ denotes the number of pixels containing valid ground truth values. Following [3], we set $\lambda = 0.85$ and $\beta = 10$ for all our experiments. + +Table 1: Quantitative evaluation on SynWoodScape [34]. + +
MethodRMSE ↓RMSE log ↓Abs Rel ↓Sq Rel ↓δ1 ↑δ2 ↑δ3 ↑
OmniDet* [20]5.6230.7401.0803.1330.0310.0640.753
BTS [21]2.9980.3250.1790.4590.6530.8450.932
AdaBins [3]1.5240.1040.0700.1270.9710.9930.997
SlaBins (ours)1.0400.0560.0220.0550.9880.9970.999
+ +Slanted angle loss. As mentioned in Sec. 3.2, we predict slanted angle $\alpha$ through a two-step module: classification and regression. For the classification module, we use the cross-entropy loss $\mathcal{L}_{CE}$ . As for the regression module, we utilize the mean squared error (MSE) loss $\mathcal{L}_{MSE}$ between the estimated residual and the ground truth residual. We define the sum of the CE loss and MES loss as the slanted angle loss: + +$$ +\mathcal {L} _ {\text {a n g l e}} = \mathcal {L} _ {C E} + \mathcal {L} _ {M S E}. \tag {11} +$$ + +Total loss. Finally, the total loss for our problem is: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \gamma \mathcal {L} _ {\text {b i n s}} + \mathcal {L} _ {\text {p i x e l}} + \mathcal {L} _ {\text {a n g l e}}, \tag {12} +$$ + +where $\gamma$ is a balancing parameter (we set $\gamma = 0.1$ ). + +# 4. Experiments + +We evaluate the proposed method from various perspectives. We introduce datasets and evaluation metrics in Sec. 4.1 and implementation details in Sec. 4.2. We then offer quantitative and qualitative comparisons to state-of-the-art methods in Sec. 4.3 and an ablation study in Sec. 4.4. + +# 4.1. Datasets and evaluation metrics + +For evaluation, we use two fisheye depth datasets: the SynWoodScape dataset [34], which is a synthetic dataset, and the KITTI-360 depth dataset, which is a modified dataset of real-world KITTI-360 [24] for our scenario. Note that we additionally augment two datasets for our purposes and will release these datasets for research communities. + +SynWoodScape dataset [34]. The SynWoodScape dataset is a synthetic surround-view fisheye camera dataset for autonomous driving, which is a synthetic version of the real-world WoodScape dataset $[42]^2$ . This dataset provides pairs of synthetic fisheye images and depth maps for the front, left, right, and rear of the driving situation. Each view contains 500 images with $190^{\circ}$ FoV and a size of $1280\times 966$ . + +Additionally, we augment the SynWoodScape dataset in terms of the viewing direction of fisheye cameras to validate the proposed framework. Specifically, we warp fisheye images and the corresponding depth maps by changing a total of nine viewing directions, of which rotation angles w.r.t. the x-axis are between $0^{\circ}$ and $90^{\circ}$ (between the viewing direction and the normal of the ground), including the original viewing direction. So, we have 18K fisheye images + +Table 2: Quantitative evaluation on KITTI-360 depth. + +
MethodRMSE ↓RMSE log ↓Abs Rel ↓Sq Rel ↓δ1 ↑δ2 ↑δ3 ↑
OmniDet* [20]1.9170.1500.0860.2460.9080.9750.992
BTS [21]1.7870.1530.0930.2430.9060.9760.992
AdaBins [3]1.5900.1290.0740.1850.9400.9840.994
SlaBins (ours)1.5370.1300.0770.1770.9410.9830.994
+ +in total. We split the augmented dataset into 14.4K training, 1.8K validation, and 1.8K test samples. Following the previous work [16], we train the model and evaluate the range of depth up to $40\mathrm{m}$ because fisheye cameras undergo high data compression and can perform up to this range, unlike the KITTI dataset [10] covering up to $80\mathrm{m}$ . + +KITTI-360 depth dataset. The KITTI-360 dataset [24] is a suburban driving dataset for 2D/3D semantic and instance segmentation tasks that include various input modalities, semantic instance annotations, and precise localization. KITTI-360 provides fisheye images, raw LiDAR point clouds, and semantic and instance labels with $185^{\circ}$ FoV of left and right in the driving situation, with a resolution of $1400\times 1400$ . Note that since KITTI-360 does not provide the ground truth for fisheye depth estimation, we generate the corresponding ground truth using raw LiDAR data. We call this modified dataset for fisheye depth estimation as KITTI-360 depth dataset. + +Concretely, we first make dense LiDAR by aggregating LiDAR data of twenty neighbor frames. We then project the dense LiDAR point clouds on the fisheye camera image domain and conduct occlusion handling [18] because of the displacement between the fisheye camera and the LiDAR (see Fig. 6). We then consider depth values, of which distance is less than $40\mathrm{m}$ , as the ground truth. We sample 4K samples of the training set from sequence 0 to 7, 0.4K samples of the validation set from sequence 9, and 0.4K samples of the test set from sequence 10. To make dense depth maps, we resize samples to $700\times 700$ . In addition, as we did for SynWoodScape, we augment the viewing directions of the KITTI-360 depth dataset; the angle between the viewing direction and the normal on the ground to $30^{\circ}$ , $50^{\circ}$ , $70^{\circ}$ , and $90^{\circ}$ (the original viewing angle). So, we have 19.2K fisheye images in total. We split this augmented dataset into 16K training, 1.6K validation, and 1.6K test samples. + +Evaluation metrics. We use the standard depth estimation metrics used in the previous work [21]. For error metrics, we calculate mean absolute relative error (Abs Rel), root mean square error (RMSE), root mean square log error (RMSE log), the root mean square relative error (Sq Rel). For accuracy metrics, we calculate threshold accuracy $\delta_{i} = 1.25^{i}$ for $i\in \{1,2,3\}$ . Generally, the smaller the error, the better, and the higher the accuracy, the better. + +# 4.2. Implementation details + +Our proposed method is implemented in PyTorch and trained on 4 RTX A6000 GPUs. Our model is trained for + +![](images/6d3c7429a7e8ad17cd1c5f6f3597db8ef6e94bd5eaf4c4c592b32150cf36ac18.jpg) +RGB + +![](images/cf170ac772f44ad54bf400ce288917d5be89402e6eb6226517c2730b8887ffc2.jpg) +OmniDet* [20] + +![](images/b8dabc0e04cf9b8853c8a5e676afedd420da7e777af3f39bbb3a504e43ea33f6.jpg) +BTS [21] + +![](images/9a1f6f495b7946782a7cd798ea1f2e57ac5093d2c85012f2d41d27ab407f7821.jpg) +AdaBins [3] + +![](images/cd939911703bf1ea774a6a074cd4f04fdc646cabd88727d485cc8bf8baf84281.jpg) +Ours + +![](images/436fa484293c46f7d11dfbd38d909b935b5f6a49dabde97688b9499937608064.jpg) +Ground truth + +![](images/0413c375d7b9bb8d32226b85d47b58ae94b75b287ba2cf54536a5f369a2e20ca.jpg) +Figure 5: Qualitative comparison on the SynWoodScape dataset. Sky-blue boxes are enlarged views of specific objects to show detail. Overall, our method shows distinct boundaries compared to the other models. In particular, we can observe the detailed structure from our results, such as bicycle wheels in the first row and the boundary of two overlapped vehicles in the second row. +Figure 6: Qualitative comparison on the KITTI-360 depth dataset. For visualization, we crop the object region in the fisheye image. In contrast to the other methods, our method shows clear boundaries near the object regions, as shown in sky-blue and green boxes. + +10 epochs on SynWoodScape and 20 epochs on KITTI-360 depth using AdamW [28] optimizer with weight-decay $10^{-2}$ , with a batch size of 8. We use the same 1-cycle policy [35] as AdaBins [3] for the learning rate with max $\mathrm{lr} = 3.5 \times 10^{-4}$ . For the first $30\%$ of iterations, the linear learning rate warm-up technique is utilized. For the learning rate decay, the cosine annealing learning rate strategy is applied. + +# 4.3. Comparison + +We compare our SlaBins with several baseline approaches: OmniDet [20], BTS [21], and AdaBins [3]. OmniDet is a fisheye depth estimation method using self-supervised learning. Based on the released code of the authors, we modify OmniDet to a supervised one for fairness (we denote this modified one OmniDet*). As base + +line methods for pinhole cameras, we compare BTS and AdaBins, especially AdaBins is a bins-based fundamental method for depth estimation. It should be noted that we target depth estimation fisheye cameras, but we additionally compare it with conventional pinhole camera-based methods due to the lack of fisheye camera-based methods. + +Quantitative evaluation. Table 1 shows the quantitative comparisons on the SynWoodScape dataset. The arrow next to the metric indicates the direction of better performance. Overall, SlaBins outperforms the other baseline approaches [3, 20, 21]. In particular, RMSE and Abs-Rel errors are significantly improved; $31.7\%$ and $68.6\%$ decreased compared to AdaBins, which shows the effect of the slanted MCI representation in the fisheye domain. + +On the KITTI-360 depth dataset, our method still outper + +![](images/2212e2af948023b583428b260616d7d55ad63e5f051987bd8e26db16f5727976.jpg) +Figure 7: Depth histogram on the object regions. Left: RGB data (top) and the concept of depth histogram in the orthogonal coordinate (bottom). Middle: depth maps of only object regions for each model and the ground truth. Right: depth histograms (AdaBins, original MCI, slanted MCI, ground truth in that order). + +forms the comparison methods (see Table 2). Specifically, RMSE and Sq-Rel errors are improved; $13.9\%$ and $27.1\%$ decreased compared to BTS and $3.3\%$ and $4.3\%$ decreased compared to AdaBins. Note that even though the KITTI-360 depth dataset captured in suburban scenes does not perfectly fit our scenario of road environments, SlaBins still shows competitive performance, which demonstrates the effect of the proposed framework. + +Qualitative evaluation. Figure 5 shows the estimated depth maps for the test scenes on SynWoodScape. Due to the nature of fisheye cameras, we can observe ambiguous boundary regions in the comparison methods. Even AdaBins, which is our baseline and a naive MSI representation by the Euclidean distance of fisheye, shows bleeding depth estimation results near the boundaries between objects and ground regions. In contrast, thanks to the slanted MCI representation, SlaBins shows accurate depth results with distinct boundaries regardless of camera viewing directions. Specifically, we can observe that sharp objects with details (e.g., human legs and bike) are preserved in our method, unlike the other methods that smooth object details. In addition, we can observe a similar trend on the KITTI-360 depth dataset, as shown in Fig. 6. Compared to the baseline methods, SlaBins shows better depth consistency even on sharp objects (e.g., third and fourth columns in Fig. 6). Additional results are available in the supplementary material. + +# 4.4. Ablation study + +Multi-layered representation. We compare the proposed slanted MCI with the MSI used in AdaBins and the original MCI on the orthogonal coordinate. We measure the depth histogram for only object regions on the orthogonal coordinate. As shown in Fig. 7, compared to the other representations, SlaBins reveals the most similar trend (i.e., peak points) with the ground truth histogram. Furthermore, we compare slanted MCI and original MCI in terms of depth consistency. Figure 8 shows that the slanted MCI provides + +![](images/620ba8fd308f1615bb8ba8f2c0ed397c2b10ba1a9f8cae01b712a742bb836b65.jpg) +Figure 8: Slanted MCI vs. original MCI in depth and clustering. All the results including the ground truth are converted to the XZ distance in the orthogonal coordinate. + +Table 3: Ablation study for the slanted angle predictor on SynWoodScape and KITTI-360 depth. We report the depth error (Abs Rel) and mean angle error. C and R denote classification-only and regression-only, respectively. + +
MethodSynWoodScapeKITTI-360 depth
Abs Rel ↓Angle error↓Abs Rel ↓Angle error
R0.07527.0310.60448.32
C0.0260.8420.1035
C+R (ours)0.0220.0310.0770.017
+ +more consistent results in object regions both on depth and depth-based clustering results. Hence, we believe that the slanted MCI can provide robust geometric constraints for various tasks in fisheye domains. + +Slanted angle prediction. As shown in Table 3, we perform an ablation study for the angle prediction module. Unlike a conventional regression module having the advantage over classifications, in our case, it is hard to converge because of the trade-off with the depth estimation part in the SlaBins module. That is why regression-only module is poor on the angle prediction. On the contrary, our two-step prediction module, composed of classification and residual regression, shows the best performance with stable training. + +# 5. Conclusion + +We have proposed a novel fisheye depth estimation network, called SlaBins, based on the geometric property of the road environments. We newly introduce the slanted MCI representation, which allows us to represent a given road scene as a set of cylindrical layers orthogonal to the road ground. By virtue of the slanted MCI, our network can implicitly encode scene depth information invariant to the camera viewing direction while preserving boundary information of scene objects w.r.t. the road region. Thus, we achieve state-of-the-art performance on monocular fisheye depth estimation. + +Acknowledgments. This work was supported by 42dot Inc., Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2020-0-01336, Artificial Intelligence Graduate School Program (UNIST)) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF2021R1C1C1005723). + +# References + +[1] T-Bone Accident Statistics And Other Legal Information. https://www.injurytriallawyer.com/faqs/ how-dangerous-are-t-bone-accidents-.cfm. 1 +[2] Ashutosh Agarwal and Chetan Arora. Attention Attention Everywhere: Monocular Depth Prediction with Skip Attention. arXiv preprint arXiv:2210.09071, 2022. 1 +[3] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In CVPR, 2021. 2, 3, 5, 6, 7 +[4] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. LocalBins: Improving Depth Estimation by Learning Local Distributions. In ECCV, 2022. 2 +[5] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, et al. PersFormer: 3D Lane Detection via Perspective Transformer and the OpenLane Benchmark. arXiv preprint arXiv:2203.11089, 2022. 1 +[6] Liang-Chieh Chen, Huiyu Wang, and Siyuan Qiao. Scaling wide residual networks for panoptic segmentation. arXiv preprint arXiv:2011.11675, 2020. 1 +[7] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In CVPR, 2017. 5 +[8] Olivier Faugeras and Quang-Tuan Luong. The geometry of multiple images: the laws that govern the formation of multiple images of a scene and some of their applications. MIT press, 2001. 2 +[9] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, 2018. 2 +[10] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 6 +[11] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In CVPR, 2017. 2 +[12] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In CVPR, 2020. 2 +[13] Yuxuan Han, Ruicheng Wang, and Jiaolong Yang. Single-View View Synthesis in the Wild with Learned Adaptive Multiplane Images. arXiv preprint arXiv:2205.11733, 2022. 2 +[14] Adrian Johnston and Gustavo Carneiro. Self-supervised monocular trained depth estimation using self-attention and discrete disparity volume. In CVPR, 2020. 2 +[15] Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, and Junmo Kim. Global-Local Path Networks for Monocular Depth Estimation with Vertical Cut-Depth. arXiv preprint arXiv:2201.07436, 2022. 2 +[16] Varun Ravi Kumar, Sandesh Athni Hiremath, Markus Bach, Stefan Milz, Christian Witt, Clément Pinard, Senthil Yogamani, and Patrick Mäder. Fisheyedistancenet: Self-supervised scale-aware distance estimation using monocular + +fisheye camera for autonomous driving. In ICRA, 2020. 1, 2, 6 +[17] Varun Ravi Kumar, Marvin Klingner, Senthil Yogamani, Markus Bach, Stefan Milz, Tim Fingscheidt, and Patrick Mäder. SVDistNet: Self-supervised near-field distance estimation on surround view fisheye cameras. IEEE Transactions on Intelligent Transportation Systems, 2021. 2 +[18] Varun Ravi Kumar, Stefan Milz, Christian Witt, Martin Simon, Karl Amende, Johannes Petzold, Senthil Yogamani, and Timo Pech. Monocular fisheye camera depth estimation using sparse lidar supervision. In IEEE International Conference on Intelligent Transportation Systems (ITSC), 2018. 2, 6 +[19] Varun Ravi Kumar, Senthil Yogamani, Markus Bach, Christian Witt, Stefan Milz, and Patrick Mäder. Unrectdepthnet: Self-supervised monocular depth estimation using a generic framework for handling common camera distortion models. In IROS, 2020. 2 +[20] Varun Ravi Kumar, Senthil Yogamani, Hazem Rashed, Ganesh Sitsu, Christian Witt, Isabelle Leang, Stefan Milz, and Patrick Mader. Omnidet: Surround view cameras based multi-task visual perception network for autonomous driving. IEEE Robotics and Automation Letters, 6(2):2830-2837, 2021. 1, 2, 3, 6, 7 +[21] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 2, 6, 7 +[22] Jiaxin Li, Zijian Feng, Qi She, Henghui Ding, Changhu Wang, and Gim Hee Lee. Mine: Towards continuous depth spi with nerf for novel view synthesis. In ICCV, 2021. 2 +[23] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. DepthFormer: Exploiting Long-Range Correlation and Local Information for Accurate Monocular Depth Estimation. arXiv preprint arXiv:2203.14211, 2022. 2 +[24] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE TPAMI, 2022. 2, 6 +[25] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETRv2: A Unified Framework for 3D Perception from Multi-Camera Images. arXiv preprint arXiv:2206.01256, 2022. 1 +[26] Yuxuan Liu, Yuan Yixuan, and Ming Liu. Ground-aware monocular 3d object detection for autonomous driving. IEEE Robotics and Automation Letters, 6(2):919-926, 2021. 3 +[27] Xiaoxiao Long, Cheng Lin, Lingjie Liu, Wei Li, Christian Theobalt, Ruigang Yang, and Wenping Wang. Adaptive surface normal constraint for depth estimation. In ICCV, 2021. 3 +[28] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In *ICLR*, 2018. 7 +[29] Reza Mahjourian, Martin Wicke, and Anelia Angelova. Unsupervised learning of depth and ego-motion from monocular video using 3d geometric constraints. In CVPR, 2018. 2, 3 +[30] Vaishakh Patil, Christos Sakaridis, Alexander Liniger, and Luc Van Gool. P3Depth: Monocular Depth Estimation with a Piecewise Planarity Prior. In CVPR, 2022. 3 + +[31] Elad Plaut, Erez Ben Yaacov, and Bat El Shlomo. 3d object detection from a single fisheye image without a single fisheye training image. In CVPR, 2021. 1 +[32] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 2 +[33] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI, 2020. 2 +[34] Ahmed Rida Sekkat, Yohan Dupuis, Varun Ravi Kumar, Hazem Rashed, Senthil Yogamani, Pascal Vasseur, and Paul Honeine. SynWoodScape: Synthetic Surround-View Fisheye Camera Dataset for Autonomous Driving. IEEE Robotics and Automation Letters, 7(3):8502-8509, jul 2022. 1, 2, 6 +[35] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, volume 11006, pages 369-386. SPIE, 2019. 7 +[36] Richard Szeliski and Polina Golland. Stereo matching with transparency and matting. In ICCV, 1998. 2 +[37] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 4 +[38] Andrew Tao, Karan Sapra, and Bryan Catanzaro. Hierarchical multi-scale attention for semantic segmentation. arXiv preprint arXiv:2005.10821, 2020. 1 +[39] Richard Tucker and Noah Snively. Single-view view synthesis with multiplane images. In CVPR, 2020. 2 +[40] Qingan Yan, Pan Ji, Nitin Bansal, Yuxin Ma, Yuan Tian, and Yi Xu. FisheyeDistill: Self-Supervised Monocular Depth Estimation with Ordinal Distillation for Fisheye Cameras. arXiv preprint arXiv:2205.02930, 2022. 2, 3 +[41] Yaozu Ye, Kailun Yang, Kaite Xiang, Juan Wang, and Kaiwei Wang. Universal semantic segmentation for fisheye urban driving images. In IEEE International Conference on Systems, Man, and Cybernetics (SMC), pages 648-655, 2020. 1 +[42] Senthil Yogamani, Ciarán Hughes, Jonathan Horgan, Ganesh Sistu, Padraig Varley, Derek O'Dea, Michal Uricár, Stefan Milz, Martin Simon, Karl Amende, et al. WoodScape: A multi-task, multi-camera fisheye dataset for autonomous driving. arXiv preprint arXiv:1905.01489, 2019. 5, 6 +[43] Yifan Zhang, Qijian Zhang, Zhiyu Zhu, Junhui Hou, and Yixuan Yuan. GLENet: Boosting 3D Object Detectors with Generative Label Uncertainty Estimation. arXiv preprint arXiv:2207.02466, 2022. 1 +[44] Dingfu Zhou, Jin Fang, Xibin Song, Liu Liu, Junbo Yin, Yuchao Dai, Hongdong Li, and Ruigang Yang. Joint 3d instance segmentation and object detection for autonomous driving. In CVPR, 2020. 1 \ No newline at end of file diff --git a/2023/SlaBins_ Fisheye Depth Estimation using Slanted Bins on Road Environments/images.zip b/2023/SlaBins_ Fisheye Depth Estimation using Slanted Bins on Road Environments/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1214e2107d84c57f98697280270317070e3ceead --- /dev/null +++ b/2023/SlaBins_ Fisheye Depth Estimation using Slanted Bins on Road Environments/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d553319173db6293d5c4e390d4ac03472965514c816b066c7e59e5ce8809c0a +size 501670 diff --git a/2023/SlaBins_ Fisheye Depth Estimation using Slanted Bins on Road Environments/layout.json b/2023/SlaBins_ Fisheye Depth Estimation using Slanted Bins on Road Environments/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d80237f776e40d3bef970e5f8bd9321afa0901e9 --- /dev/null +++ b/2023/SlaBins_ Fisheye Depth Estimation using Slanted Bins on Road Environments/layout.json @@ -0,0 +1,10325 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 136, + 521, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 136, + 521, + 153 + ], + "spans": [ + { + "bbox": [ + 56, + 136, + 521, + 153 + ], + "type": "text", + "content": "SlaBins: Fisheye Depth Estimation using Slanted Bins on Road Environments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 174, + 485, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 174, + 485, + 202 + ], + "spans": [ + { + "bbox": [ + 94, + 174, + 485, + 202 + ], + "type": "text", + "content": "Jongsung Lee" + }, + { + "bbox": [ + 94, + 174, + 485, + 202 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 174, + 485, + 202 + ], + "type": "text", + "content": ", Gyeongsu Cho" + }, + { + "bbox": [ + 94, + 174, + 485, + 202 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 94, + 174, + 485, + 202 + ], + "type": "text", + "content": ", Jeongin Park" + }, + { + "bbox": [ + 94, + 174, + 485, + 202 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 94, + 174, + 485, + 202 + ], + "type": "text", + "content": ", Kyongjun Kim" + }, + { + "bbox": [ + 94, + 174, + 485, + 202 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 94, + 174, + 485, + 202 + ], + "type": "text", + "content": ", Seongoh Lee" + }, + { + "bbox": [ + 94, + 174, + 485, + 202 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 94, + 174, + 485, + 202 + ], + "type": "text", + "content": ", Jung-Hee Kim" + }, + { + "bbox": [ + 94, + 174, + 485, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 94, + 174, + 485, + 202 + ], + "type": "text", + "content": ", Seong-Gyun Jeong" + }, + { + "bbox": [ + 94, + 174, + 485, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 94, + 174, + 485, + 202 + ], + "type": "text", + "content": ", Kyungdon Joo" + }, + { + "bbox": [ + 94, + 174, + 485, + 202 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 145, + 204, + 431, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 204, + 431, + 218 + ], + "spans": [ + { + "bbox": [ + 145, + 204, + 431, + 218 + ], + "type": "text", + "content": "1Artificial Intelligence Graduate School, UNIST, 242dot Inc." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 108, + 221, + 466, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 221, + 466, + 233 + ], + "spans": [ + { + "bbox": [ + 108, + 221, + 466, + 233 + ], + "type": "text", + "content": "{syniez, threedv, jeonginpark, kimkj38, solee, kyungdon}@unist.ac.kr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 187, + 234, + 391, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 234, + 391, + 244 + ], + "spans": [ + { + "bbox": [ + 187, + 234, + 391, + 244 + ], + "type": "text", + "content": "{junghee.kim, seonggyun.jeong}@42dot.ai" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 139, + 270, + 186, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 270, + 186, + 282 + ], + "spans": [ + { + "bbox": [ + 139, + 270, + 186, + 282 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 295, + 282, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 295, + 282, + 610 + ], + "spans": [ + { + "bbox": [ + 45, + 295, + 282, + 610 + ], + "type": "text", + "content": "Although 3D perception for autonomous vehicles has focused on frontal-view information, more than half of fatal accidents occur due to side impacts in practice (e.g., T-bone crash). Motivated by this fact, we investigate the problem of side-view depth estimation, especially for monocular fisheye cameras, which provide wide FoV information. However, since fisheye cameras head road areas, it observes road areas mostly and results in severe distortion on object areas, such as vehicles or pedestrians. To alleviate these issues, we propose a new fisheye depth estimation network, SlaBins, that infers an accurate and dense depth map based on a geometric property of road environments; most objects are standing (i.e., orthogonal) on the road environments. Concretely, we introduce a slanted multi-cylindrical image (MCI) representation, which allows us to describe a distance as a radius to a cylindrical layer orthogonal to the ground regardless of the camera viewing direction. Based on the slanted MCI, we estimate a set of adaptive bins and a per-pixel probability map for depth estimation. Then by combining it with the estimated slanted angle of viewing direction, we directly infer a dense and accurate depth map for fisheye cameras. Experiments demonstrate that SlaBins outperforms the state-of-the-art methods in both qualitative and quantitative evaluation on the SynWoodScape and KITTI-360 depth datasets. For more information, you can visit our project page https://syniez.github.io/SlaBins/." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 633, + 124, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 124, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 124, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 653, + 279, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 653, + 279, + 701 + ], + "spans": [ + { + "bbox": [ + 45, + 653, + 279, + 701 + ], + "type": "text", + "content": "With the advent of autonomous driving, the importance of 3D perception keeps growing for safety [20, 25, 44]. For example, ADAS systems pervading our daily driving already, such as lane-keeping aid and forward collision" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 275, + 366, + 309 + ], + "blocks": [ + { + "bbox": [ + 314, + 266, + 346, + 274 + ], + "lines": [ + { + "bbox": [ + 314, + 266, + 346, + 274 + ], + "spans": [ + { + "bbox": [ + 314, + 266, + 346, + 274 + ], + "type": "text", + "content": "RGB" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 275, + 366, + 309 + ], + "lines": [ + { + "bbox": [ + 320, + 275, + 366, + 309 + ], + "spans": [ + { + "bbox": [ + 320, + 275, + 366, + 309 + ], + "type": "image", + "image_path": "d7972720d786bf5df51d9fa598f8e11245287d358ff6e72fae4804fcebe9ef92.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 320, + 310, + 365, + 344 + ], + "blocks": [ + { + "bbox": [ + 320, + 310, + 365, + 344 + ], + "lines": [ + { + "bbox": [ + 320, + 310, + 365, + 344 + ], + "spans": [ + { + "bbox": [ + 320, + 310, + 365, + 344 + ], + "type": "image", + "image_path": "37b580a71bf49cf0fecfaa948c403f57b7cba1bbeb5510e580a9b0bc794b9dd2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 320, + 346, + 365, + 383 + ], + "blocks": [ + { + "bbox": [ + 320, + 346, + 365, + 383 + ], + "lines": [ + { + "bbox": [ + 320, + 346, + 365, + 383 + ], + "spans": [ + { + "bbox": [ + 320, + 346, + 365, + 383 + ], + "type": "image", + "image_path": "d0b8520def059b3be9dcfe75b01395c6a8eaeaf52ce61e53aa75d7cb318791cc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 297, + 385, + 531, + 417 + ], + "lines": [ + { + "bbox": [ + 297, + 385, + 531, + 417 + ], + "spans": [ + { + "bbox": [ + 297, + 385, + 531, + 417 + ], + "type": "text", + "content": "Figure 1: SlaBins on the SynWoodScape dataset [34]. SlaBins predicts dense and accurate depth maps regardless of the viewing directions using the slanted MCI representation." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 384, + 285, + 448, + 320 + ], + "blocks": [ + { + "bbox": [ + 372, + 275, + 454, + 284 + ], + "lines": [ + { + "bbox": [ + 372, + 275, + 454, + 284 + ], + "spans": [ + { + "bbox": [ + 372, + 275, + 454, + 284 + ], + "type": "text", + "content": "Slanted MCI representation" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 384, + 285, + 448, + 320 + ], + "lines": [ + { + "bbox": [ + 384, + 285, + 448, + 320 + ], + "spans": [ + { + "bbox": [ + 384, + 285, + 448, + 320 + ], + "type": "image", + "image_path": "8d8f184aef1f234e2fe02f96a914794b6d7f4134d100206e1769b7b1574c0dea.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 368, + 327, + 460, + 361 + ], + "blocks": [ + { + "bbox": [ + 368, + 327, + 460, + 361 + ], + "lines": [ + { + "bbox": [ + 368, + 327, + 460, + 361 + ], + "spans": [ + { + "bbox": [ + 368, + 327, + 460, + 361 + ], + "type": "image", + "image_path": "7f96553eab508578605a2bb78c6bb6d48249c691d76b811d8ed157b55c3244a7.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 463, + 276, + 521, + 309 + ], + "blocks": [ + { + "bbox": [ + 478, + 266, + 498, + 274 + ], + "lines": [ + { + "bbox": [ + 478, + 266, + 498, + 274 + ], + "spans": [ + { + "bbox": [ + 478, + 266, + 498, + 274 + ], + "type": "text", + "content": "Depth" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 276, + 521, + 309 + ], + "lines": [ + { + "bbox": [ + 463, + 276, + 521, + 309 + ], + "spans": [ + { + "bbox": [ + 463, + 276, + 521, + 309 + ], + "type": "image", + "image_path": "1dba10f1d3bc1670f20779959c64f817370d99ce9fc28790171071aafd506e23.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 463, + 310, + 521, + 345 + ], + "blocks": [ + { + "bbox": [ + 463, + 310, + 521, + 345 + ], + "lines": [ + { + "bbox": [ + 463, + 310, + 521, + 345 + ], + "spans": [ + { + "bbox": [ + 463, + 310, + 521, + 345 + ], + "type": "image", + "image_path": "1916eeae437689dd6e45be8389dd44a8a7744130cde9b6863669a882614c7778.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 463, + 346, + 522, + 383 + ], + "blocks": [ + { + "bbox": [ + 463, + 346, + 522, + 383 + ], + "lines": [ + { + "bbox": [ + 463, + 346, + 522, + 383 + ], + "spans": [ + { + "bbox": [ + 463, + 346, + 522, + 383 + ], + "type": "image", + "image_path": "1197552c0da8cc880c7bd38bfff1fe4a1e72ccb4d59ea085403bfc4b563b74b4.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 296, + 426, + 531, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 426, + 531, + 600 + ], + "spans": [ + { + "bbox": [ + 296, + 426, + 531, + 600 + ], + "type": "text", + "content": "avoidance assist, depend on various sensor modalities to observe 3D information [43, 5]. Here, one interesting point is that various sensor modalities for 3D perception, including many datasets related to autonomous driving, are focusing on the frontal view predominantly [43, 38, 2, 6]. We can easily deduce the reason by our common sense that it is important to keep an eye on the front while driving. However, according to the National Transportation Highway Safety Administration [1], " + }, + { + "bbox": [ + 296, + 426, + 531, + 600 + ], + "type": "inline_equation", + "content": "51\\%" + }, + { + "bbox": [ + 296, + 426, + 531, + 600 + ], + "type": "text", + "content": " of fatal car accidents correspond to T-bone crashes in which one car hits the side of another, which is the second highest type of accident. This statistic supports that side-view 3D perception is important to assist drivers because they are hard to pay attention to the side view due to the physical characteristics of the people looking ahead and their position characteristics." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 296, + 602, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 602, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 602, + 531, + 730 + ], + "type": "text", + "content": "Motivated by this fact, 3D perception for side-view, especially fisheye-based depth estimation, has started to gain the spotlight [31, 16, 41]. Fisheye cameras have several advantages for 3D perception in road scenes. Specifically, fisheye cameras provide broader FoV information (typically greater than " + }, + { + "bbox": [ + 296, + 602, + 531, + 730 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 296, + 602, + 531, + 730 + ], + "type": "text", + "content": ") in a single image, which allows us to observe road information (e.g., road, vehicles, and pedestrians) at a low cost. In addition, we can directly utilize perspective camera 3D perception approaches because mathematical mutual conversion between two camera models is possible. Thus, if we can obtain side-view depth informa" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 77, + 43, + 119, + 71 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 43, + 119, + 71 + ], + "spans": [ + { + "bbox": [ + 77, + 43, + 119, + 71 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 151, + 41, + 470, + 73 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 41, + 470, + 73 + ], + "spans": [ + { + "bbox": [ + 151, + 41, + 470, + 73 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 709, + 174, + 719 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 709, + 174, + 719 + ], + "spans": [ + { + "bbox": [ + 57, + 709, + 174, + 719 + ], + "type": "text", + "content": "*Equal contribution (alphabet order)." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 57, + 719, + 132, + 729 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 719, + 132, + 729 + ], + "spans": [ + { + "bbox": [ + 57, + 719, + 132, + 729 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "8765" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 279, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 279, + 129 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 279, + 129 + ], + "type": "text", + "content": "tion using a single fisheye camera, we can efficiently handle various vision tasks for autonomous driving." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 131, + 279, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 131, + 279, + 305 + ], + "spans": [ + { + "bbox": [ + 45, + 131, + 279, + 305 + ], + "type": "text", + "content": "However, despite the advantages of fisheye cameras, monocular fisheye camera-based side-view depth estimation is a challenging problem due to significant distortion. In addition to inherent issues of monocular depth estimation, such as scale ambiguity [8], severe distortion of fisheye cameras makes depth estimation more difficult by distorting scene objects (e.g., close objects look closer and distant objects look more distant abnormally). Particularly, this distortion problem could be more severe in road environments because fisheye cameras for ADAS systems generally look to the road areas, and vital objects such as vehicles and pedestrians are located at the edge of the image. That is, most regions of the image correspond to the road areas and objects contain very few pixels, so depth estimation becomes more challenging." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 306, + 279, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 306, + 279, + 458 + ], + "spans": [ + { + "bbox": [ + 45, + 306, + 279, + 458 + ], + "type": "text", + "content": "To alleviate this problem, we pay attention to the geometric property of road environments; most objects are standing on the road environments. Based on this geometric characteristic, we can easily imagine the visible region of each object as a planar segment orthogonal to the road region. This approximation is connected to a multi-plane image (MPI) concept [36] that represents a given scene as the composition of multi-layered images according to depth. Note that one significant difference is that a fisheye camera looks down to the road region in our case (i.e., camera viewing direction is not parallel to the ground), unlike a camera points to a given scene in the general MPI concept. Thus, a naive MPI concept is not directly applicable." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 458, + 279, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 458, + 279, + 679 + ], + "spans": [ + { + "bbox": [ + 45, + 458, + 279, + 679 + ], + "type": "text", + "content": "In this work, we propose a new fisheye depth estimation framework, SlaBins, that exploits the geometric property of the road environments (see Fig. 1). To be specific, we adopt a multi-cylindrical image (MCI) representation; especially, we introduce a slanted MCI of which the cylindrical layer is defined as orthogonal to the road ground regardless of camera viewing direction. Based on the slanted MCI, we estimate adaptive bins and the per-pixel probability of the slanted cylindrical layers, which allows us to implicitly preserve the geometric property of the road environments and improve the depth quality near the boundary between objects and the ground region. In addition, we independently estimate the slanted angle of the camera viewing direction w.r.t. the road ground. Then, by combining this angle with the estimated depth information in the slanted MCI domain, we can directly compute a dense and accurate depth map for fisheye cameras. We validate the proposed SlaBins framework on the SynWoodScape dataset [34] and KITTI-360 depth dataset (a modified dataset of KITTI" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 297, + 106, + 530, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 106, + 530, + 141 + ], + "spans": [ + { + "bbox": [ + 297, + 106, + 530, + 141 + ], + "type": "text", + "content": "360 [24] for our scenario), where SlaBins outperforms the baseline methods [3, 20, 21] and shows potential as scene representations for downstream tasks, e.g., segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 142, + 490, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 142, + 490, + 152 + ], + "spans": [ + { + "bbox": [ + 308, + 142, + 490, + 152 + ], + "type": "text", + "content": "In a nutshell, our contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 297, + 157, + 530, + 313 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 297, + 157, + 530, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 157, + 530, + 192 + ], + "spans": [ + { + "bbox": [ + 297, + 157, + 530, + 192 + ], + "type": "text", + "content": "- We propose a new fisheye depth estimation framework, SlaBins, which infers dense depth based on the geometric property of road environments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 297, + 200, + 530, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 200, + 530, + 258 + ], + "spans": [ + { + "bbox": [ + 297, + 200, + 530, + 258 + ], + "type": "text", + "content": "- We newly introduce a slanted MCI representation as an intermediate road scene representation, which allows us to describe a given scene regardless of viewing direction and improve the depth quality on the boundary areas between objects and the road." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 297, + 266, + 530, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 266, + 530, + 313 + ], + "spans": [ + { + "bbox": [ + 297, + 266, + 530, + 313 + ], + "type": "text", + "content": "- By decoupling the slanted angle of an input image, we force the proposed model to learn the slanted MCI depth information and then seamlessly compute a fisheye depth map in a slanted-aware regression way." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 297, + 323, + 381, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 323, + 381, + 335 + ], + "spans": [ + { + "bbox": [ + 297, + 323, + 381, + 335 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 343, + 530, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 343, + 530, + 494 + ], + "spans": [ + { + "bbox": [ + 296, + 343, + 530, + 494 + ], + "type": "text", + "content": "Monocular depth estimation. With the advent of breakthroughs in deep neural networks, such as CNN and transformer architecture, monocular depth estimation has shown impressive performance [15, 33, 21, 32, 23, 18, 40, 11, 9, 12, 29, 19, 16, 17, 20]. As one of the recent trends, several methods formulate a depth estimation task as a combination of classification and regression problems, which enables training the network more easily and shows stable performance [14, 3, 4]. For this formulation, they commonly use an MPI according to depth, which is popularly used in novel-view synthesis tasks [22, 39, 13]. The MPI notation represents depth as quantized depth values w.r.t. the uniformly quantized planes and residuals from mapped planes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 494, + 530, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 494, + 530, + 680 + ], + "spans": [ + { + "bbox": [ + 296, + 494, + 530, + 680 + ], + "type": "text", + "content": "To improve the uniformly quantized planes by uniform bins, recent research works [3, 4] present a network that predicts adaptive bins depending on the input scene and local regions in the image. AdaBins [3] computes adaptive bin-centers and per-pixel probabilities relying on each input image. Inspired by AdaBins, LocalBins [4] improves depth quality by considering both the global histogram and local-region histograms of each input image. Nevertheless, the predicted depth from the above models also depends on the camera's pose because they predict depth parallel with the camera's principal axis. This constraint makes it hard to adopt the previous models into fisheye depth estimation, which has a large FoV and uses the Euclidean distance instead of depth. In this work, we introduce and exploit a slanted MCI representation that is invariant to the viewing angle of fisheye cameras w.r.t. road environments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "content": "Fisheye depth estimation. Despite increasing demand for fisheye cameras in the industry, like ADAS systems, there are only a few previous research works in fisheye depth estimation fields [20, 17, 19, 16, 40]. Kumar et al. [19] propose" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 45, + 691, + 279, + 730 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 691, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 691, + 279, + 730 + ], + "type": "text", + "content": "1Basically, depth in the fisheye model is defined as the Euclidean distance, which is connected to a multi-spherical image (MSI). In this work, We adopt an MCI representation that combines the geometrical suitability of MPI and the fisheye model fitness of MSI." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 279, + 747, + 299, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 299, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 299, + 756 + ], + "type": "text", + "content": "8766" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 101, + 517, + 230 + ], + "blocks": [ + { + "bbox": [ + 58, + 101, + 517, + 230 + ], + "lines": [ + { + "bbox": [ + 58, + 101, + 517, + 230 + ], + "spans": [ + { + "bbox": [ + 58, + 101, + 517, + 230 + ], + "type": "image", + "image_path": "5edfc6aa98a4cee7db2ed3927bd6007ed4962c651216163c8bef773ce430c348.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 233, + 531, + 267 + ], + "lines": [ + { + "bbox": [ + 45, + 233, + 531, + 267 + ], + "spans": [ + { + "bbox": [ + 45, + 233, + 531, + 267 + ], + "type": "text", + "content": "Figure 2: Overview of the proposed SlaBins. (a) Given an input fisheye RGB image, SlaBins framework first extracts features by the encoder-decoder block. The extracted features are used as input to the slanted angle prediction module and the SlaBins module to get a depth image. (b) Illustration of the coordinates and MCI representations defined in our model. (c) Illustration of the slanted angle predictor." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 278, + 279, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 278, + 279, + 475 + ], + "spans": [ + { + "bbox": [ + 45, + 278, + 279, + 475 + ], + "type": "text", + "content": "a camera-aware depth estimation network to handle the severe distortion of fisheye cameras. They encode the camera's intrinsic parameters as a tensor named camera geometry tensor and use it in the training step for camera-aware depth estimation in fisheye images. With camera geometry tensor, they propose a multi-task network for fisheye cameras, called OmniDet [20], that simultaneously estimates depth estimation, motion, semantic segmentation, and object detection for synergy. However, since OmniDet relies on the partial region in the whole image, they may miss information on image edges. Using a whole image, Yan et al. [40] distill depth ordering information from a teacher network trained on the large-scale dataset to train in fisheye images, but FoV they cover is not large (e.g., vFoV=66°). In contrast to previous approaches, we propose a fisheye depth estimation network that predicts fisheye image depth without any information loss such as crop, and narrow FoV." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 479, + 279, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 479, + 279, + 676 + ], + "spans": [ + { + "bbox": [ + 45, + 479, + 279, + 676 + ], + "type": "text", + "content": "Geometric cues for depth estimation. One way to improve fisheye depth estimation is to exploit geometric cues between the camera and the world. Liu et al. [26] use a ground plane to get depth priors using camera elevation and camera parameters. Moreover, Mahjourian et al. [29] predict camera ego-motion and compute loss on both the 3D point cloud domain and the 2D image domain. In addition to driving scenes, several works use plane coefficients to give consistency on plane regions in depth estimation [30, 27]. Patil et al. [30] estimate plane coefficients instead of depth and then combine plane coefficients into depth to use plane constraints (e.g., co-planarity, plane normals) in 3D space. Inspired by the previous works [3, 30], we design a network that estimates a dense and accurate depth map for fisheye cameras based on the slanted MCI representation using some geometric constraints (i.e., orthogonality between ground and object)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 687, + 169, + 699 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 687, + 169, + 699 + ], + "spans": [ + { + "bbox": [ + 46, + 687, + 169, + 699 + ], + "type": "text", + "content": "3. Proposed Framework" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 706, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 706, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 46, + 706, + 279, + 730 + ], + "type": "text", + "content": "Inspired by the geometric property of the road environments, we propose a new fisheye depth estimation frame" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 296, + 277, + 530, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 277, + 530, + 440 + ], + "spans": [ + { + "bbox": [ + 296, + 277, + 530, + 440 + ], + "type": "text", + "content": "work (see Fig. 2 for the overview). Specifically, we introduce a slanted MCI representation, which allows us to describe a multi-layer cylindrical image orthogonal to the road ground regardless of the viewing direction of the camera. Based on the slanted MCI representation, our SlaBins module estimates the adaptive bin widths and per-pixel probability map in the orthogonal coordinate, which provides depth information invariant to the camera viewing direction. We then combine the estimated depth information with the slanted angle estimated by the slanted angle prediction module. Through this process, we can directly estimate a dense depth map for fisheye cameras. Our approach is built upon the previous AdaBins framework [3]. So, we call our method SlaBins." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 297, + 450, + 450, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 450, + 450, + 463 + ], + "spans": [ + { + "bbox": [ + 297, + 450, + 450, + 463 + ], + "type": "text", + "content": "3.1. Slanted MCI Representation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 469, + 530, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 469, + 530, + 539 + ], + "spans": [ + { + "bbox": [ + 296, + 469, + 530, + 539 + ], + "type": "text", + "content": "As intermediate scene representation, we use multi-layer images for fisheye depth estimation. Here, we briefly compare two multi-layer image representations for fisheye images: multi-spherical image (MSI) and multi-cylindrical image (MCI). We then present a slanted MCI representation, which is a suitable representation of our problem." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 542, + 530, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 542, + 530, + 682 + ], + "spans": [ + { + "bbox": [ + 296, + 542, + 530, + 682 + ], + "type": "text", + "content": "MSI vs. MCI. By nature of fisheye cameras covering over " + }, + { + "bbox": [ + 296, + 542, + 530, + 682 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 296, + 542, + 530, + 682 + ], + "type": "text", + "content": ", it is common to use the Euclidean distance as its depth; that is, a layer with the same depth value forms a sphere in the fisheye domain. Thus, we can represent a multi-layer image for fisheye cameras as an MSI in the spherical coordinate, which is a naive representation. However, this naive MSI representation has inherent limitations for fisheye depth estimation. We observed that such a representation can result in ambiguous depth boundaries between the ground and object regions, as shown in Fig. 3a. It may neutralize the benefit of multi-layer images by separating an object into two different layers, for example." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "content": "In the MCI representation, a cylindrical layer encodes the same Euclidean distance w.r.t. the y-axis. In other words, we can represent an object, such as a pedestrian on the road, in a single cylindrical layer (see Fig. 3b). Further" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "8767" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 107, + 113, + 191 + ], + "blocks": [ + { + "bbox": [ + 62, + 107, + 113, + 191 + ], + "lines": [ + { + "bbox": [ + 62, + 107, + 113, + 191 + ], + "spans": [ + { + "bbox": [ + 62, + 107, + 113, + 191 + ], + "type": "image", + "image_path": "ea4262433bf086d7a8b2ac0379dc44babe6a81a7aeb9e88c04f5fa8d519bb004.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 196, + 93, + 204 + ], + "lines": [ + { + "bbox": [ + 82, + 196, + 93, + 204 + ], + "spans": [ + { + "bbox": [ + 82, + 196, + 93, + 204 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 119, + 114, + 170, + 191 + ], + "blocks": [ + { + "bbox": [ + 119, + 114, + 170, + 191 + ], + "lines": [ + { + "bbox": [ + 119, + 114, + 170, + 191 + ], + "spans": [ + { + "bbox": [ + 119, + 114, + 170, + 191 + ], + "type": "image", + "image_path": "dead83876114f702840b50d1018e067c973d44ec28768124628d0242097c6da9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 196, + 149, + 204 + ], + "lines": [ + { + "bbox": [ + 139, + 196, + 149, + 204 + ], + "spans": [ + { + "bbox": [ + 139, + 196, + 149, + 204 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 180, + 104, + 263, + 190 + ], + "blocks": [ + { + "bbox": [ + 180, + 104, + 263, + 190 + ], + "lines": [ + { + "bbox": [ + 180, + 104, + 263, + 190 + ], + "spans": [ + { + "bbox": [ + 180, + 104, + 263, + 190 + ], + "type": "image", + "image_path": "6d5259420a93cc5f1427c3e76bedce9f962d1879a1890b65f3edee92f392fd9b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 196, + 227, + 204 + ], + "lines": [ + { + "bbox": [ + 217, + 196, + 227, + 204 + ], + "spans": [ + { + "bbox": [ + 217, + 196, + 227, + 204 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 206, + 279, + 291 + ], + "lines": [ + { + "bbox": [ + 45, + 206, + 279, + 291 + ], + "spans": [ + { + "bbox": [ + 45, + 206, + 279, + 291 + ], + "type": "text", + "content": "Figure 3: MSI vs. MCI representation. Illustration of an example scene and zoomed results with (a) MSI and (b) MCI, where we can observe more accurate depth in the slanted MCI representation near the boundary between pedestrians and the ground. (c) Relationship between the original cylindrical layer " + }, + { + "bbox": [ + 45, + 206, + 279, + 291 + ], + "type": "inline_equation", + "content": "C_r^o" + }, + { + "bbox": [ + 45, + 206, + 279, + 291 + ], + "type": "text", + "content": " (gray-colored cylinder) in the world coordinate (arrows) and the slanted cylindrical layer " + }, + { + "bbox": [ + 45, + 206, + 279, + 291 + ], + "type": "inline_equation", + "content": "C_{r_\\perp}^s" + }, + { + "bbox": [ + 45, + 206, + 279, + 291 + ], + "type": "text", + "content": " (orange-colored cylinders) in the orthogonal coordinate (dashed arrows)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 301, + 279, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 301, + 279, + 347 + ], + "spans": [ + { + "bbox": [ + 45, + 301, + 279, + 347 + ], + "type": "text", + "content": "more, we can directly exploit the geometric properties of the road environment: most objects are standing (i.e., orthogonal) on the road environment. For this reason, we adopt the MCI as our intermediate scene representation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 349, + 279, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 349, + 279, + 382 + ], + "spans": [ + { + "bbox": [ + 45, + 349, + 279, + 382 + ], + "type": "text", + "content": "Slanted MCI representation. Let " + }, + { + "bbox": [ + 45, + 349, + 279, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 45, + 349, + 279, + 382 + ], + "type": "text", + "content": " be a 3D point in the world coordinate. Then, we can represent " + }, + { + "bbox": [ + 45, + 349, + 279, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 45, + 349, + 279, + 382 + ], + "type": "text", + "content": " according to the coordinate systems:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 86, + 383, + 279, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 383, + 279, + 418 + ], + "spans": [ + { + "bbox": [ + 86, + 383, + 279, + 418 + ], + "type": "interline_equation", + "content": "\\mathbf {X} = \\left[ \\begin{array}{l} X \\\\ Y \\\\ Z \\end{array} \\right] = \\rho \\left[ \\begin{array}{c} \\hat {X} \\\\ \\hat {Y} \\\\ \\hat {Z} \\end{array} \\right] = \\rho \\left[ \\begin{array}{c} \\sin \\theta \\cos \\phi \\\\ \\sin \\theta \\sin \\phi \\\\ \\cos \\theta \\end{array} \\right], \\tag {1}", + "image_path": "ab6da6354c146cd651b6f54ca8d3f4fbe95283ce3f8ae8c4792f748af7f0e692.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 426, + 279, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 426, + 279, + 474 + ], + "spans": [ + { + "bbox": [ + 45, + 426, + 279, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 426, + 279, + 474 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}} = [\\hat{X},\\hat{Y},\\hat{Z}]^{\\top}" + }, + { + "bbox": [ + 45, + 426, + 279, + 474 + ], + "type": "text", + "content": " is the normalized unit vector (i.e., ray direction), " + }, + { + "bbox": [ + 45, + 426, + 279, + 474 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 45, + 426, + 279, + 474 + ], + "type": "text", + "content": " indicates the Euclidean distance (i.e., depth in the fisheye domain), and " + }, + { + "bbox": [ + 45, + 426, + 279, + 474 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 45, + 426, + 279, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 426, + 279, + 474 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 45, + 426, + 279, + 474 + ], + "type": "text", + "content": " denote the elevation and azimuth angles, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 474, + 279, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 474, + 279, + 498 + ], + "spans": [ + { + "bbox": [ + 45, + 474, + 279, + 498 + ], + "type": "text", + "content": "Based on the cylindrical representation, we can form a cylindrical layer according to a radius " + }, + { + "bbox": [ + 45, + 474, + 279, + 498 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 45, + 474, + 279, + 498 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 86, + 501, + 279, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 501, + 279, + 516 + ], + "spans": [ + { + "bbox": [ + 86, + 501, + 279, + 516 + ], + "type": "interline_equation", + "content": "C _ {r} ^ {o} = \\left\\{\\mathbf {X} \\mid X ^ {2} + Z ^ {2} = r ^ {2}; \\mathbf {X} \\in \\mathbb {R} ^ {3} \\right\\}. \\tag {2}", + "image_path": "9d53146737e6ad941c9a9fbe281a0ed471a67e0cc274d33773b67af6560aa6f3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 45, + 522, + 279, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 522, + 279, + 591 + ], + "spans": [ + { + "bbox": [ + 45, + 522, + 279, + 591 + ], + "type": "text", + "content": "We call " + }, + { + "bbox": [ + 45, + 522, + 279, + 591 + ], + "type": "inline_equation", + "content": "C_r^o" + }, + { + "bbox": [ + 45, + 522, + 279, + 591 + ], + "type": "text", + "content": " the original cylindrical layer at " + }, + { + "bbox": [ + 45, + 522, + 279, + 591 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 45, + 522, + 279, + 591 + ], + "type": "text", + "content": ". This original MCI is inherently suitable for road environments and shows less distortion than MSI on ambiguous depth boundaries. However, depending on the principal axis of cameras (i.e., viewing direction) w.r.t. the road environments, the original MCI may hinder layered scene representation like the MSI." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 45, + 591, + 279, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 591, + 279, + 684 + ], + "spans": [ + { + "bbox": [ + 45, + 591, + 279, + 684 + ], + "type": "text", + "content": "To alleviate this issue, we consider the 1D angle " + }, + { + "bbox": [ + 45, + 591, + 279, + 684 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 45, + 591, + 279, + 684 + ], + "type": "text", + "content": " between the principal axis and the road environments. Using " + }, + { + "bbox": [ + 45, + 591, + 279, + 684 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 45, + 591, + 279, + 684 + ], + "type": "text", + "content": ", we can implicitly describe the slanted MCI representation that is orthogonal to the road areas (see Fig. 3c). To this end, let " + }, + { + "bbox": [ + 45, + 591, + 279, + 684 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\perp}" + }, + { + "bbox": [ + 45, + 591, + 279, + 684 + ], + "type": "text", + "content": " be a 3D point in the orthogonal coordinate, of which the XZ plane is parallel to the ground. Similarly to Eq. (2), we can form a cylindrical layer according to a radius " + }, + { + "bbox": [ + 45, + 591, + 279, + 684 + ], + "type": "inline_equation", + "content": "r_{\\perp}" + }, + { + "bbox": [ + 45, + 591, + 279, + 684 + ], + "type": "text", + "content": " in the orthogonal coordinate as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 75, + 688, + 279, + 703 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 688, + 279, + 703 + ], + "spans": [ + { + "bbox": [ + 75, + 688, + 279, + 703 + ], + "type": "interline_equation", + "content": "C _ {r _ {\\perp}} ^ {s} = \\left\\{\\mathbf {X} _ {\\perp} \\mid X _ {\\perp} ^ {2} + Z _ {\\perp} ^ {2} = r _ {\\perp} ^ {2}; \\mathbf {X} _ {\\perp} \\in \\mathbb {R} ^ {3} \\right\\}, \\tag {3}", + "image_path": "74a048b0c199b8dfa1f942030e3017cf19926fc180eb1063ecf5ce7cb5849abb.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 45, + 707, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 707, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 707, + 279, + 730 + ], + "type": "text", + "content": "which corresponds to a slanted cylindrical layer in the world (camera) coordinate (see Fig. 2b for coordinate systems)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 106, + 530, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 128 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 128 + ], + "type": "text", + "content": "By " + }, + { + "bbox": [ + 296, + 106, + 530, + 128 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 296, + 106, + 530, + 128 + ], + "type": "text", + "content": ", we can derive the relationship between " + }, + { + "bbox": [ + 296, + 106, + 530, + 128 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\perp}" + }, + { + "bbox": [ + 296, + 106, + 530, + 128 + ], + "type": "text", + "content": " in the orthogonal coordinate and " + }, + { + "bbox": [ + 296, + 106, + 530, + 128 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 296, + 106, + 530, + 128 + ], + "type": "text", + "content": " in the world coordinate by:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 336, + 131, + 530, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 131, + 530, + 165 + ], + "spans": [ + { + "bbox": [ + 336, + 131, + 530, + 165 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {\\perp} = \\left[ \\begin{array}{l} X _ {\\perp} \\\\ Y _ {\\perp} \\\\ Z _ {\\perp} \\end{array} \\right] = \\left[ \\begin{array}{c} X \\\\ Y \\cos \\alpha + Z \\sin \\alpha \\\\ - Y \\sin \\alpha + Z \\cos \\alpha \\end{array} \\right]. \\tag {4}", + "image_path": "73ee04f1d4d2984a2a58a82ea8902049811e0d5eacdd7b44abfc29ecf7927351.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 169, + 530, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 169, + 530, + 204 + ], + "spans": [ + { + "bbox": [ + 296, + 169, + 530, + 204 + ], + "type": "text", + "content": "Then, we can deduce the following relationship between a 3D point " + }, + { + "bbox": [ + 296, + 169, + 530, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\perp}" + }, + { + "bbox": [ + 296, + 169, + 530, + 204 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 296, + 169, + 530, + 204 + ], + "type": "inline_equation", + "content": "C_{r_{\\perp}}^{s}" + }, + { + "bbox": [ + 296, + 169, + 530, + 204 + ], + "type": "text", + "content": " and its spherical representation in the camera coordinate by using Eqs. (3) and (4):" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 206, + 530, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 530, + 242 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 530, + 242 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} r _ {\\perp} ^ {2} = X _ {\\perp} ^ {2} + Z _ {\\perp} ^ {2} = X ^ {2} + (- Y \\sin \\alpha + Z \\cos \\alpha) ^ {2}, \\\\ = \\rho^ {2} \\left(\\hat {X} ^ {2} + (- \\hat {Y} \\sin \\alpha + \\hat {Z} \\cos \\alpha) ^ {2}\\right). \\tag {5} \\\\ \\end{array}", + "image_path": "a4c9c80e1eb37b7b622b61a72a12db838dc40be575f5c0d3f10d3a1f77f814b3.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 296, + 248, + 530, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 248, + 530, + 269 + ], + "spans": [ + { + "bbox": [ + 296, + 248, + 530, + 269 + ], + "type": "text", + "content": "Furthermore, we can convert the relationship in Eq. (5) to a function " + }, + { + "bbox": [ + 296, + 248, + 530, + 269 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 296, + 248, + 530, + 269 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 379, + 267, + 530, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 267, + 530, + 281 + ], + "spans": [ + { + "bbox": [ + 379, + 267, + 530, + 281 + ], + "type": "interline_equation", + "content": "f (\\tilde {\\mathbf {X}}, r _ {\\perp}, \\alpha) = \\rho , \\tag {6}", + "image_path": "dd167718b32a0b60774febf0e417b1b60d6cd99163d43c864fc077ee5fb8f6a5.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 296, + 282, + 530, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 282, + 530, + 374 + ], + "spans": [ + { + "bbox": [ + 296, + 282, + 530, + 374 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 282, + 530, + 374 + ], + "type": "inline_equation", + "content": "r_{\\perp}" + }, + { + "bbox": [ + 296, + 282, + 530, + 374 + ], + "type": "text", + "content": " denotes a radius in the orthogonal coordinate. The function " + }, + { + "bbox": [ + 296, + 282, + 530, + 374 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 296, + 282, + 530, + 374 + ], + "type": "text", + "content": " represents that if we know slanted angle " + }, + { + "bbox": [ + 296, + 282, + 530, + 374 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 296, + 282, + 530, + 374 + ], + "type": "text", + "content": " and a radius " + }, + { + "bbox": [ + 296, + 282, + 530, + 374 + ], + "type": "inline_equation", + "content": "r_{\\perp}" + }, + { + "bbox": [ + 296, + 282, + 530, + 374 + ], + "type": "text", + "content": " at a certain pixel location on the fisheye domain, we can directly compute the corresponding depth value " + }, + { + "bbox": [ + 296, + 282, + 530, + 374 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 296, + 282, + 530, + 374 + ], + "type": "text", + "content": ". A more detailed explanation for Eq. (6) is available in the supplementary material. Using this relation, we implicitly model the slanted MCI representation inside the network as scene representation." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 297, + 381, + 407, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 381, + 407, + 392 + ], + "spans": [ + { + "bbox": [ + 297, + 381, + 407, + 392 + ], + "type": "text", + "content": "3.2. SlaBins framework" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 296, + 398, + 530, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 398, + 530, + 457 + ], + "spans": [ + { + "bbox": [ + 296, + 398, + 530, + 457 + ], + "type": "text", + "content": "Based on the slanted MCI representation (cf. Sec. 3.1), the proposed SlaBins framework estimates an accurate and dense depth map. SlaBins consists of three modules: Encoder-decoder block, slanted angle predictor, and SlaBins module, as shown in Fig. 2." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 296, + 457, + 530, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 457, + 530, + 596 + ], + "spans": [ + { + "bbox": [ + 296, + 457, + 530, + 596 + ], + "type": "text", + "content": "Concretely, given the input fisheye image, the encoder-decoder block extracts a feature map containing useful information for depth prediction. Using the extracted feature map, the SlaBins module estimates several key information, such as adaptive bin-centers and per-pixel probability map in the slanted MCI representation (i.e., the orthogonal coordinate), using a transformer architecture. It should be worth noticing that by virtue of this SlaBins module, we can implicitly preserve the geometric relationship of objects on the road environments. Then, by combining with the slanted angle estimated by the slanted angle prediction module, we can directly regress a fisheye depth map." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 296, + 598, + 530, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 598, + 530, + 691 + ], + "spans": [ + { + "bbox": [ + 296, + 598, + 530, + 691 + ], + "type": "text", + "content": "Encoder-decoder block. Given input fisheye image " + }, + { + "bbox": [ + 296, + 598, + 530, + 691 + ], + "type": "inline_equation", + "content": "x_{i} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 296, + 598, + 530, + 691 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 296, + 598, + 530, + 691 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 296, + 598, + 530, + 691 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 598, + 530, + 691 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 296, + 598, + 530, + 691 + ], + "type": "text", + "content": " denote the height and the width of the input, our encoder-decoder block first extract features. We use the pre-trained EfficientNet B5 [37] as the backbone for the encoder and standard feature upsampling layers for the decoder, which returns decoded features " + }, + { + "bbox": [ + 296, + 598, + 530, + 691 + ], + "type": "inline_equation", + "content": "x_{d} \\in \\mathbb{R}^{h \\times w \\times C_{d}}" + }, + { + "bbox": [ + 296, + 598, + 530, + 691 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 296, + 598, + 530, + 691 + ], + "type": "inline_equation", + "content": "h = H / 2" + }, + { + "bbox": [ + 296, + 598, + 530, + 691 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 598, + 530, + 691 + ], + "type": "inline_equation", + "content": "w = W / 2" + }, + { + "bbox": [ + 296, + 598, + 530, + 691 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 296, + 598, + 530, + 691 + ], + "type": "inline_equation", + "content": "C_{d}" + }, + { + "bbox": [ + 296, + 598, + 530, + 691 + ], + "type": "text", + "content": " indicates the feature dimension." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "text", + "content": "Slanted angle prediction module. We design compact MLP-based classification and regression modules that predict " + }, + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "inline_equation", + "content": "x_{d}" + }, + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "text", + "content": " in a two-step. The classifier first classifies" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "8768" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 104, + 271, + 309 + ], + "blocks": [ + { + "bbox": [ + 51, + 104, + 271, + 309 + ], + "lines": [ + { + "bbox": [ + 51, + 104, + 271, + 309 + ], + "spans": [ + { + "bbox": [ + 51, + 104, + 271, + 309 + ], + "type": "image", + "image_path": "f9ffadc1450b2b43272998450f0ed9d8f7682ceadba8466503b64bf7bc2dd07f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 311, + 281, + 397 + ], + "lines": [ + { + "bbox": [ + 45, + 311, + 281, + 397 + ], + "spans": [ + { + "bbox": [ + 45, + 311, + 281, + 397 + ], + "type": "text", + "content": "Figure 4: The details of the SlaBins module. Our SlaBins module consists of two components: One is the mini-ViT block that extracts the range-attention map and bin widths, which are invariant to the slanted angle " + }, + { + "bbox": [ + 45, + 311, + 281, + 397 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 45, + 311, + 281, + 397 + ], + "type": "text", + "content": " from the decoded feature. The other is the SlaBins-aware regression. This module receives the outputs of mini-ViT and the slanted angle " + }, + { + "bbox": [ + 45, + 311, + 281, + 397 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 45, + 311, + 281, + 397 + ], + "type": "text", + "content": " as inputs and directly converts the orthogonally estimated depth " + }, + { + "bbox": [ + 45, + 311, + 281, + 397 + ], + "type": "inline_equation", + "content": "r_{\\perp}" + }, + { + "bbox": [ + 45, + 311, + 281, + 397 + ], + "type": "text", + "content": " for an arbitrary 3D point " + }, + { + "bbox": [ + 45, + 311, + 281, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 45, + 311, + 281, + 397 + ], + "type": "text", + "content": " into the depth " + }, + { + "bbox": [ + 45, + 311, + 281, + 397 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 45, + 311, + 281, + 397 + ], + "type": "text", + "content": " of the spherical coordinate." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 405, + 279, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 405, + 279, + 453 + ], + "spans": [ + { + "bbox": [ + 45, + 405, + 279, + 453 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 45, + 405, + 279, + 453 + ], + "type": "text", + "content": " in units of " + }, + { + "bbox": [ + 45, + 405, + 279, + 453 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 45, + 405, + 279, + 453 + ], + "type": "text", + "content": ", and the following regressor estimates the offset in the range of " + }, + { + "bbox": [ + 45, + 405, + 279, + 453 + ], + "type": "inline_equation", + "content": "(-2.5^{\\circ}, 2.5^{\\circ})" + }, + { + "bbox": [ + 45, + 405, + 279, + 453 + ], + "type": "text", + "content": ". This two-step angle prediction enables us to estimate an accurate viewing direction angle and train more efficiently (see Fig. 2c)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 454, + 279, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 454, + 279, + 548 + ], + "spans": [ + { + "bbox": [ + 45, + 454, + 279, + 548 + ], + "type": "text", + "content": "Specifically, " + }, + { + "bbox": [ + 45, + 454, + 279, + 548 + ], + "type": "inline_equation", + "content": "x_{d}" + }, + { + "bbox": [ + 45, + 454, + 279, + 548 + ], + "type": "text", + "content": " is downsampled to " + }, + { + "bbox": [ + 45, + 454, + 279, + 548 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 45, + 454, + 279, + 548 + ], + "type": "text", + "content": " size with two " + }, + { + "bbox": [ + 45, + 454, + 279, + 548 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 45, + 454, + 279, + 548 + ], + "type": "text", + "content": " convolution layers and average pooling. Downsampled features are flattened (we denote the flattened feature as " + }, + { + "bbox": [ + 45, + 454, + 279, + 548 + ], + "type": "inline_equation", + "content": "x_{f}" + }, + { + "bbox": [ + 45, + 454, + 279, + 548 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 45, + 454, + 279, + 548 + ], + "type": "inline_equation", + "content": "x_{f}" + }, + { + "bbox": [ + 45, + 454, + 279, + 548 + ], + "type": "text", + "content": " is used as input of classifier and regressor, which are composed of 3 FC layers followed by softmax and 4 FC layers, respectively. The output of the classifier is a 19-dimensional vector and that of the regressor is a normalized angle offset in the range of " + }, + { + "bbox": [ + 45, + 454, + 279, + 548 + ], + "type": "inline_equation", + "content": "(-1^{\\circ}, 1^{\\circ})" + }, + { + "bbox": [ + 45, + 454, + 279, + 548 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 553, + 279, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 553, + 279, + 623 + ], + "spans": [ + { + "bbox": [ + 45, + 553, + 279, + 623 + ], + "type": "text", + "content": "SlaBins module. Given decoded features " + }, + { + "bbox": [ + 45, + 553, + 279, + 623 + ], + "type": "inline_equation", + "content": "x_{d}" + }, + { + "bbox": [ + 45, + 553, + 279, + 623 + ], + "type": "text", + "content": ", the proposed SlaBins module first learns intermediate depth information, such as bin widths and probability maps in the slanted MCI representation, which are invariant to the viewing direction (i.e., slanted angle " + }, + { + "bbox": [ + 45, + 553, + 279, + 623 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 45, + 553, + 279, + 623 + ], + "type": "text", + "content": "), as shown in Fig. 4. To this end, we adopt mini-ViT (mViT) architecture in [3]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "text", + "content": "The decoded features " + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "inline_equation", + "content": "x_{d}" + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "text", + "content": " are fed into the mViT with learnable positional encodings. To make a fixed-size patch-embedded input, we embed the decoded features to the " + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "inline_equation", + "content": "C_p" + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "text", + "content": " dimensional flattened tensor " + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "inline_equation", + "content": "x_{p}\\in \\mathbb{R}^{S\\times C_{p}}" + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "text", + "content": " using patch embedding before feeding into the mViT (here, " + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "text", + "content": " indicates " + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "inline_equation", + "content": "w\\times h / p^2" + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "text", + "content": " denotes the size of patch and stride). Then, the sum of flattened tensor " + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "inline_equation", + "content": "x_{p}" + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "text", + "content": " and positional encodings are fed into the mViT and the output of the mViT has the same shape as input tensor " + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "inline_equation", + "content": "x_{o}\\in \\mathbb{R}^{S\\times C_{p}}" + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "text", + "content": ". The first MLP head" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "text", + "content": "of the output " + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "inline_equation", + "content": "x_{o}" + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "text", + "content": " means slanted bin widths and it is normalized into " + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "text", + "content": " for further depth regression. The other features followed by the " + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "text", + "content": " convolution layer perform pixel-wise dot product with a " + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "text", + "content": " convolution output of " + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "inline_equation", + "content": "x_{d}" + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "text", + "content": " to get the range-attention map. Softmax function with " + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "text", + "content": " convolution layer is used to represent the range-attention map as a probability of each pixel " + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "inline_equation", + "content": "p_{k}" + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "text", + "content": " bin-centers. These normalized slanted bin-width " + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "text", + "content": " and per-pixel probability " + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "inline_equation", + "content": "p_{k}" + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 296, + 106, + 530, + 211 + ], + "type": "text", + "content": " are used for SlaBins-aware regression." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 211, + 530, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 211, + 530, + 270 + ], + "spans": [ + { + "bbox": [ + 296, + 211, + 530, + 270 + ], + "type": "text", + "content": "Our final depth " + }, + { + "bbox": [ + 296, + 211, + 530, + 270 + ], + "type": "inline_equation", + "content": "\\tilde{\\rho}" + }, + { + "bbox": [ + 296, + 211, + 530, + 270 + ], + "type": "text", + "content": " is computed by the linear combination of the two outputs from mViT (i.e., slanted bin-width " + }, + { + "bbox": [ + 296, + 211, + 530, + 270 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 296, + 211, + 530, + 270 + ], + "type": "text", + "content": " and per-pixel probability " + }, + { + "bbox": [ + 296, + 211, + 530, + 270 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 296, + 211, + 530, + 270 + ], + "type": "text", + "content": ") and the transformation function in Eq. (6). To compute final depth, slanted depth bin-centers " + }, + { + "bbox": [ + 296, + 211, + 530, + 270 + ], + "type": "inline_equation", + "content": "c^s" + }, + { + "bbox": [ + 296, + 211, + 530, + 270 + ], + "type": "text", + "content": " are computed with normalized bin-width " + }, + { + "bbox": [ + 296, + 211, + 530, + 270 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 296, + 211, + 530, + 270 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 296, + 211, + 530, + 270 + ], + "type": "inline_equation", + "content": "\\sum b_{i} = 1" + }, + { + "bbox": [ + 296, + 211, + 530, + 270 + ], + "type": "text", + "content": "):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 281, + 530, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 281, + 530, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 281, + 530, + 297 + ], + "type": "interline_equation", + "content": "c ^ {s} \\left(b _ {i}\\right) = r _ {\\perp} ^ {\\min } + \\left(r _ {\\perp} ^ {\\max } - r _ {\\perp} ^ {\\min }\\right) \\left(b _ {i} / 2 + \\Sigma_ {j = 1} ^ {i - 1} b _ {j}\\right), \\tag {7}", + "image_path": "e6de19e816006814613853ec0d29240643284ed86d655fbc9a4166358e036791.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 301, + 530, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 301, + 530, + 360 + ], + "spans": [ + { + "bbox": [ + 296, + 301, + 530, + 360 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 301, + 530, + 360 + ], + "type": "inline_equation", + "content": "r_{\\perp}^{min}" + }, + { + "bbox": [ + 296, + 301, + 530, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 301, + 530, + 360 + ], + "type": "inline_equation", + "content": "r_{\\perp}^{max}" + }, + { + "bbox": [ + 296, + 301, + 530, + 360 + ], + "type": "text", + "content": " mean the lower and upper bound of the radius in slanted MCI representation. Then, orthogonal coordinate radius " + }, + { + "bbox": [ + 296, + 301, + 530, + 360 + ], + "type": "inline_equation", + "content": "r_{\\perp}" + }, + { + "bbox": [ + 296, + 301, + 530, + 360 + ], + "type": "text", + "content": " is calculated using the cumulative sum of " + }, + { + "bbox": [ + 296, + 301, + 530, + 360 + ], + "type": "inline_equation", + "content": "c_k^s" + }, + { + "bbox": [ + 296, + 301, + 530, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 301, + 530, + 360 + ], + "type": "inline_equation", + "content": "p_k^s" + }, + { + "bbox": [ + 296, + 301, + 530, + 360 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 296, + 301, + 530, + 360 + ], + "type": "inline_equation", + "content": "\\sum_{k=1}^{N} c^s(b_k)p_k" + }, + { + "bbox": [ + 296, + 301, + 530, + 360 + ], + "type": "text", + "content": ") and transformed to our final depth " + }, + { + "bbox": [ + 296, + 301, + 530, + 360 + ], + "type": "inline_equation", + "content": "\\tilde{\\rho}" + }, + { + "bbox": [ + 296, + 301, + 530, + 360 + ], + "type": "text", + "content": " using Eq. (6):" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 351, + 361, + 530, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 361, + 530, + 380 + ], + "spans": [ + { + "bbox": [ + 351, + 361, + 530, + 380 + ], + "type": "interline_equation", + "content": "\\tilde {\\rho} = f \\left(\\hat {\\mathbf {X}}, \\sum_ {k = 1} ^ {N} c ^ {s} \\left(b _ {k}\\right) p _ {k}, \\alpha\\right), \\tag {8}", + "image_path": "59366fa2c9bdd9505ba93df5a643e8ce77b35e90dcf48c819f2ea404734924f6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 383, + 530, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 383, + 530, + 419 + ], + "spans": [ + { + "bbox": [ + 296, + 383, + 530, + 419 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 383, + 530, + 419 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 296, + 383, + 530, + 419 + ], + "type": "text", + "content": " is a known normalized vector computed by pixel location and camera intrinsic, and " + }, + { + "bbox": [ + 296, + 383, + 530, + 419 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 296, + 383, + 530, + 419 + ], + "type": "text", + "content": " denotes the estimated slanted angle by the slanted angle predictor." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 297, + 425, + 380, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 425, + 380, + 436 + ], + "spans": [ + { + "bbox": [ + 297, + 425, + 380, + 436 + ], + "type": "text", + "content": "3.3. Loss function" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 442, + 530, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 442, + 530, + 500 + ], + "spans": [ + { + "bbox": [ + 296, + 442, + 530, + 500 + ], + "type": "text", + "content": "Here, we introduce a set of loss functions used for training SlaBins. Specifically, we compute loss in both orthogonal and spherical coordinates to make our SlaBins framework can estimate fisheye depth that has better discrimination between objects and road environments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 502, + 530, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 502, + 530, + 524 + ], + "spans": [ + { + "bbox": [ + 296, + 502, + 530, + 524 + ], + "type": "text", + "content": "Bin loss. Following [3], we use Chamfer loss[7] to make adaptive bins:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 525, + 530, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 525, + 530, + 539 + ], + "spans": [ + { + "bbox": [ + 305, + 525, + 530, + 539 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {b i n s}} = \\operatorname {c h a m f e r} \\left(r _ {\\perp}, c ^ {s} (\\tilde {\\mathbf {b}})\\right) + \\operatorname {c h a m f e r} \\left(c ^ {s} (\\tilde {\\mathbf {b}})\\right), r _ {\\perp}), \\tag {9}", + "image_path": "0d1008d439cca3bb093d7a3b2ba47fe3500ed6fe5bdbbc3a9b3f7f9d7c5484b2.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 543, + 530, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 543, + 530, + 625 + ], + "spans": [ + { + "bbox": [ + 296, + 543, + 530, + 625 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 543, + 530, + 625 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{b}}" + }, + { + "bbox": [ + 296, + 543, + 530, + 625 + ], + "type": "text", + "content": " indicates the predicted slanted bin widths from the SlaBins module, and " + }, + { + "bbox": [ + 296, + 543, + 530, + 625 + ], + "type": "inline_equation", + "content": "r_{\\perp}" + }, + { + "bbox": [ + 296, + 543, + 530, + 625 + ], + "type": "text", + "content": " is the ground truth cylinder radius in the slanted MCI representation. In our case, to generate slanted bins (i.e., bins in the orthogonal coordinate), the ground-truth depth is converted to the ground truth radius " + }, + { + "bbox": [ + 296, + 543, + 530, + 625 + ], + "type": "inline_equation", + "content": "r_{\\perp}" + }, + { + "bbox": [ + 296, + 543, + 530, + 625 + ], + "type": "text", + "content": " at the orthogonal coordinate by Eq. (5) and the slanted MCI." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 628, + 530, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 628, + 530, + 651 + ], + "spans": [ + { + "bbox": [ + 296, + 628, + 530, + 651 + ], + "type": "text", + "content": "Pixel-wise depth loss. We apply a scaled version of the Scale-Invariant (SI) log loss introduced by Eigen et al. [42]:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 343, + 655, + 530, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 655, + 530, + 681 + ], + "spans": [ + { + "bbox": [ + 343, + 655, + 530, + 681 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p i x e l}} = \\beta \\sqrt {\\frac {1}{T} \\Sigma_ {i} g _ {i} ^ {2} - \\frac {\\lambda}{T ^ {2}} \\left(\\Sigma_ {i} g _ {i}\\right) ^ {2}} \\tag {10}", + "image_path": "b53ac1a166bbacc884c7fadf05a33289b497fc0cfc7d5530fb9b5820bf9382f0.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "inline_equation", + "content": "g_{i} = \\log \\tilde{\\rho} - \\log \\rho" + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "content": " with the ground truth depth " + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "content": " and predicted depth " + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\tilde{\\rho}" + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "content": " denotes the number of pixels containing valid ground truth values. Following [3], we set " + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\lambda = 0.85" + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\beta = 10" + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "content": " for all our experiments." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "8769" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 118, + 279, + 164 + ], + "blocks": [ + { + "bbox": [ + 57, + 105, + 267, + 116 + ], + "lines": [ + { + "bbox": [ + 57, + 105, + 267, + 116 + ], + "spans": [ + { + "bbox": [ + 57, + 105, + 267, + 116 + ], + "type": "text", + "content": "Table 1: Quantitative evaluation on SynWoodScape [34]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 118, + 279, + 164 + ], + "lines": [ + { + "bbox": [ + 47, + 118, + 279, + 164 + ], + "spans": [ + { + "bbox": [ + 47, + 118, + 279, + 164 + ], + "type": "table", + "html": "
MethodRMSE ↓RMSE log ↓Abs Rel ↓Sq Rel ↓δ1 ↑δ2 ↑δ3 ↑
OmniDet* [20]5.6230.7401.0803.1330.0310.0640.753
BTS [21]2.9980.3250.1790.4590.6530.8450.932
AdaBins [3]1.5240.1040.0700.1270.9710.9930.997
SlaBins (ours)1.0400.0560.0220.0550.9880.9970.999
", + "image_path": "db73ed08f26011508219018b81bbf0e175173a1a18990b0fff9d37d677957c79.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 168, + 279, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 168, + 279, + 259 + ], + "spans": [ + { + "bbox": [ + 45, + 168, + 279, + 259 + ], + "type": "text", + "content": "Slanted angle loss. As mentioned in Sec. 3.2, we predict slanted angle " + }, + { + "bbox": [ + 45, + 168, + 279, + 259 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 45, + 168, + 279, + 259 + ], + "type": "text", + "content": " through a two-step module: classification and regression. For the classification module, we use the cross-entropy loss " + }, + { + "bbox": [ + 45, + 168, + 279, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 45, + 168, + 279, + 259 + ], + "type": "text", + "content": ". As for the regression module, we utilize the mean squared error (MSE) loss " + }, + { + "bbox": [ + 45, + 168, + 279, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MSE}" + }, + { + "bbox": [ + 45, + 168, + 279, + 259 + ], + "type": "text", + "content": " between the estimated residual and the ground truth residual. We define the sum of the CE loss and MES loss as the slanted angle loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 259, + 279, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 259, + 279, + 271 + ], + "spans": [ + { + "bbox": [ + 111, + 259, + 279, + 271 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a n g l e}} = \\mathcal {L} _ {C E} + \\mathcal {L} _ {M S E}. \\tag {11}", + "image_path": "4ea9ec790fef23708239fa66675fd850b49784783b3eea16c462ee2af15b818a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 276, + 251, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 251, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 251, + 287 + ], + "type": "text", + "content": "Total loss. Finally, the total loss for our problem is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 295, + 279, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 295, + 279, + 308 + ], + "spans": [ + { + "bbox": [ + 90, + 295, + 279, + 308 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\gamma \\mathcal {L} _ {\\text {b i n s}} + \\mathcal {L} _ {\\text {p i x e l}} + \\mathcal {L} _ {\\text {a n g l e}}, \\tag {12}", + "image_path": "9faf5c75bbbd40f90fd6e153167a09eaec72b547b0dd468612981803188aa46e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 314, + 246, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 246, + 325 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 246, + 325 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 314, + 246, + 325 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 314, + 246, + 325 + ], + "type": "text", + "content": " is a balancing parameter (we set " + }, + { + "bbox": [ + 46, + 314, + 246, + 325 + ], + "type": "inline_equation", + "content": "\\gamma = 0.1" + }, + { + "bbox": [ + 46, + 314, + 246, + 325 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 335, + 123, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 123, + 349 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 123, + 349 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 356, + 279, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 356, + 279, + 414 + ], + "spans": [ + { + "bbox": [ + 45, + 356, + 279, + 414 + ], + "type": "text", + "content": "We evaluate the proposed method from various perspectives. We introduce datasets and evaluation metrics in Sec. 4.1 and implementation details in Sec. 4.2. We then offer quantitative and qualitative comparisons to state-of-the-art methods in Sec. 4.3 and an ablation study in Sec. 4.4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 421, + 212, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 421, + 212, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 212, + 432 + ], + "type": "text", + "content": "4.1. Datasets and evaluation metrics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 439, + 279, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 439, + 279, + 508 + ], + "spans": [ + { + "bbox": [ + 45, + 439, + 279, + 508 + ], + "type": "text", + "content": "For evaluation, we use two fisheye depth datasets: the SynWoodScape dataset [34], which is a synthetic dataset, and the KITTI-360 depth dataset, which is a modified dataset of real-world KITTI-360 [24] for our scenario. Note that we additionally augment two datasets for our purposes and will release these datasets for research communities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 509, + 279, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 509, + 279, + 590 + ], + "spans": [ + { + "bbox": [ + 45, + 509, + 279, + 590 + ], + "type": "text", + "content": "SynWoodScape dataset [34]. The SynWoodScape dataset is a synthetic surround-view fisheye camera dataset for autonomous driving, which is a synthetic version of the real-world WoodScape dataset " + }, + { + "bbox": [ + 45, + 509, + 279, + 590 + ], + "type": "inline_equation", + "content": "[42]^2" + }, + { + "bbox": [ + 45, + 509, + 279, + 590 + ], + "type": "text", + "content": ". This dataset provides pairs of synthetic fisheye images and depth maps for the front, left, right, and rear of the driving situation. Each view contains 500 images with " + }, + { + "bbox": [ + 45, + 509, + 279, + 590 + ], + "type": "inline_equation", + "content": "190^{\\circ}" + }, + { + "bbox": [ + 45, + 509, + 279, + 590 + ], + "type": "text", + "content": " FoV and a size of " + }, + { + "bbox": [ + 45, + 509, + 279, + 590 + ], + "type": "inline_equation", + "content": "1280\\times 966" + }, + { + "bbox": [ + 45, + 509, + 279, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 45, + 591, + 279, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 591, + 279, + 683 + ], + "spans": [ + { + "bbox": [ + 45, + 591, + 279, + 683 + ], + "type": "text", + "content": "Additionally, we augment the SynWoodScape dataset in terms of the viewing direction of fisheye cameras to validate the proposed framework. Specifically, we warp fisheye images and the corresponding depth maps by changing a total of nine viewing directions, of which rotation angles w.r.t. the x-axis are between " + }, + { + "bbox": [ + 45, + 591, + 279, + 683 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 45, + 591, + 279, + 683 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 591, + 279, + 683 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 45, + 591, + 279, + 683 + ], + "type": "text", + "content": " (between the viewing direction and the normal of the ground), including the original viewing direction. So, we have 18K fisheye images" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 298, + 118, + 530, + 164 + ], + "blocks": [ + { + "bbox": [ + 312, + 105, + 514, + 116 + ], + "lines": [ + { + "bbox": [ + 312, + 105, + 514, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 105, + 514, + 116 + ], + "type": "text", + "content": "Table 2: Quantitative evaluation on KITTI-360 depth." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 298, + 118, + 530, + 164 + ], + "lines": [ + { + "bbox": [ + 298, + 118, + 530, + 164 + ], + "spans": [ + { + "bbox": [ + 298, + 118, + 530, + 164 + ], + "type": "table", + "html": "
MethodRMSE ↓RMSE log ↓Abs Rel ↓Sq Rel ↓δ1 ↑δ2 ↑δ3 ↑
OmniDet* [20]1.9170.1500.0860.2460.9080.9750.992
BTS [21]1.7870.1530.0930.2430.9060.9760.992
AdaBins [3]1.5900.1290.0740.1850.9400.9840.994
SlaBins (ours)1.5370.1300.0770.1770.9410.9830.994
", + "image_path": "252239aaff9fc4c7519572405c1ad22dcc76533f1e9e71ae4499af6c147d7d22.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 173, + 530, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 173, + 530, + 243 + ], + "spans": [ + { + "bbox": [ + 296, + 173, + 530, + 243 + ], + "type": "text", + "content": "in total. We split the augmented dataset into 14.4K training, 1.8K validation, and 1.8K test samples. Following the previous work [16], we train the model and evaluate the range of depth up to " + }, + { + "bbox": [ + 296, + 173, + 530, + 243 + ], + "type": "inline_equation", + "content": "40\\mathrm{m}" + }, + { + "bbox": [ + 296, + 173, + 530, + 243 + ], + "type": "text", + "content": " because fisheye cameras undergo high data compression and can perform up to this range, unlike the KITTI dataset [10] covering up to " + }, + { + "bbox": [ + 296, + 173, + 530, + 243 + ], + "type": "inline_equation", + "content": "80\\mathrm{m}" + }, + { + "bbox": [ + 296, + 173, + 530, + 243 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 245, + 531, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 245, + 531, + 384 + ], + "spans": [ + { + "bbox": [ + 296, + 245, + 531, + 384 + ], + "type": "text", + "content": "KITTI-360 depth dataset. The KITTI-360 dataset [24] is a suburban driving dataset for 2D/3D semantic and instance segmentation tasks that include various input modalities, semantic instance annotations, and precise localization. KITTI-360 provides fisheye images, raw LiDAR point clouds, and semantic and instance labels with " + }, + { + "bbox": [ + 296, + 245, + 531, + 384 + ], + "type": "inline_equation", + "content": "185^{\\circ}" + }, + { + "bbox": [ + 296, + 245, + 531, + 384 + ], + "type": "text", + "content": " FoV of left and right in the driving situation, with a resolution of " + }, + { + "bbox": [ + 296, + 245, + 531, + 384 + ], + "type": "inline_equation", + "content": "1400\\times 1400" + }, + { + "bbox": [ + 296, + 245, + 531, + 384 + ], + "type": "text", + "content": ". Note that since KITTI-360 does not provide the ground truth for fisheye depth estimation, we generate the corresponding ground truth using raw LiDAR data. We call this modified dataset for fisheye depth estimation as KITTI-360 depth dataset." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 385, + 530, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 385, + 530, + 583 + ], + "spans": [ + { + "bbox": [ + 296, + 385, + 530, + 583 + ], + "type": "text", + "content": "Concretely, we first make dense LiDAR by aggregating LiDAR data of twenty neighbor frames. We then project the dense LiDAR point clouds on the fisheye camera image domain and conduct occlusion handling [18] because of the displacement between the fisheye camera and the LiDAR (see Fig. 6). We then consider depth values, of which distance is less than " + }, + { + "bbox": [ + 296, + 385, + 530, + 583 + ], + "type": "inline_equation", + "content": "40\\mathrm{m}" + }, + { + "bbox": [ + 296, + 385, + 530, + 583 + ], + "type": "text", + "content": ", as the ground truth. We sample 4K samples of the training set from sequence 0 to 7, 0.4K samples of the validation set from sequence 9, and 0.4K samples of the test set from sequence 10. To make dense depth maps, we resize samples to " + }, + { + "bbox": [ + 296, + 385, + 530, + 583 + ], + "type": "inline_equation", + "content": "700\\times 700" + }, + { + "bbox": [ + 296, + 385, + 530, + 583 + ], + "type": "text", + "content": ". In addition, as we did for SynWoodScape, we augment the viewing directions of the KITTI-360 depth dataset; the angle between the viewing direction and the normal on the ground to " + }, + { + "bbox": [ + 296, + 385, + 530, + 583 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 296, + 385, + 530, + 583 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 385, + 530, + 583 + ], + "type": "inline_equation", + "content": "50^{\\circ}" + }, + { + "bbox": [ + 296, + 385, + 530, + 583 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 385, + 530, + 583 + ], + "type": "inline_equation", + "content": "70^{\\circ}" + }, + { + "bbox": [ + 296, + 385, + 530, + 583 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 296, + 385, + 530, + 583 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 296, + 385, + 530, + 583 + ], + "type": "text", + "content": " (the original viewing angle). So, we have 19.2K fisheye images in total. We split this augmented dataset into 16K training, 1.6K validation, and 1.6K test samples." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 586, + 530, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 586, + 530, + 679 + ], + "spans": [ + { + "bbox": [ + 296, + 586, + 530, + 679 + ], + "type": "text", + "content": "Evaluation metrics. We use the standard depth estimation metrics used in the previous work [21]. For error metrics, we calculate mean absolute relative error (Abs Rel), root mean square error (RMSE), root mean square log error (RMSE log), the root mean square relative error (Sq Rel). For accuracy metrics, we calculate threshold accuracy " + }, + { + "bbox": [ + 296, + 586, + 530, + 679 + ], + "type": "inline_equation", + "content": "\\delta_{i} = 1.25^{i}" + }, + { + "bbox": [ + 296, + 586, + 530, + 679 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 296, + 586, + 530, + 679 + ], + "type": "inline_equation", + "content": "i\\in \\{1,2,3\\}" + }, + { + "bbox": [ + 296, + 586, + 530, + 679 + ], + "type": "text", + "content": ". Generally, the smaller the error, the better, and the higher the accuracy, the better." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 297, + 687, + 424, + 700 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 687, + 424, + 700 + ], + "spans": [ + { + "bbox": [ + 297, + 687, + 424, + 700 + ], + "type": "text", + "content": "4.2. Implementation details" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 297, + 706, + 530, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 706, + 530, + 729 + ], + "spans": [ + { + "bbox": [ + 297, + 706, + 530, + 729 + ], + "type": "text", + "content": "Our proposed method is implemented in PyTorch and trained on 4 RTX A6000 GPUs. Our model is trained for" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 45, + 690, + 279, + 709 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 690, + 279, + 709 + ], + "spans": [ + { + "bbox": [ + 45, + 690, + 279, + 709 + ], + "type": "text", + "content": "2Unfortunately, the real-world WoodScape dataset [42] does not provide the ground truth depth data." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 45, + 709, + 278, + 729 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 709, + 278, + 729 + ], + "spans": [ + { + "bbox": [ + 45, + 709, + 278, + 729 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 45, + 709, + 278, + 729 + ], + "type": "text", + "content": " Currently,the SynWoodScape dataset is partially released to the public. We perform the experiments using the partial dataset." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "8770" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 124, + 97, + 174, + 245 + ], + "blocks": [ + { + "bbox": [ + 124, + 97, + 174, + 245 + ], + "lines": [ + { + "bbox": [ + 124, + 97, + 174, + 245 + ], + "spans": [ + { + "bbox": [ + 124, + 97, + 174, + 245 + ], + "type": "image", + "image_path": "6d3c7429a7e8ad17cd1c5f6f3597db8ef6e94bd5eaf4c4c592b32150cf36ac18.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 249, + 158, + 257 + ], + "lines": [ + { + "bbox": [ + 141, + 249, + 158, + 257 + ], + "spans": [ + { + "bbox": [ + 141, + 249, + 158, + 257 + ], + "type": "text", + "content": "RGB" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 180, + 98, + 230, + 245 + ], + "blocks": [ + { + "bbox": [ + 180, + 98, + 230, + 245 + ], + "lines": [ + { + "bbox": [ + 180, + 98, + 230, + 245 + ], + "spans": [ + { + "bbox": [ + 180, + 98, + 230, + 245 + ], + "type": "image", + "image_path": "cf170ac772f44ad54bf400ce288917d5be89402e6eb6226517c2730b8887ffc2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 248, + 229, + 258 + ], + "lines": [ + { + "bbox": [ + 181, + 248, + 229, + 258 + ], + "spans": [ + { + "bbox": [ + 181, + 248, + 229, + 258 + ], + "type": "text", + "content": "OmniDet* [20]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 236, + 98, + 286, + 245 + ], + "blocks": [ + { + "bbox": [ + 236, + 98, + 286, + 245 + ], + "lines": [ + { + "bbox": [ + 236, + 98, + 286, + 245 + ], + "spans": [ + { + "bbox": [ + 236, + 98, + 286, + 245 + ], + "type": "image", + "image_path": "b8dabc0e04cf9b8853c8a5e676afedd420da7e777af3f39bbb3a504e43ea33f6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 245, + 248, + 276, + 257 + ], + "lines": [ + { + "bbox": [ + 245, + 248, + 276, + 257 + ], + "spans": [ + { + "bbox": [ + 245, + 248, + 276, + 257 + ], + "type": "text", + "content": "BTS [21]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 291, + 98, + 341, + 245 + ], + "blocks": [ + { + "bbox": [ + 291, + 98, + 341, + 245 + ], + "lines": [ + { + "bbox": [ + 291, + 98, + 341, + 245 + ], + "spans": [ + { + "bbox": [ + 291, + 98, + 341, + 245 + ], + "type": "image", + "image_path": "9a1f6f495b7946782a7cd798ea1f2e57ac5093d2c85012f2d41d27ab407f7821.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 248, + 336, + 257 + ], + "lines": [ + { + "bbox": [ + 296, + 248, + 336, + 257 + ], + "spans": [ + { + "bbox": [ + 296, + 248, + 336, + 257 + ], + "type": "text", + "content": "AdaBins [3]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 347, + 98, + 396, + 245 + ], + "blocks": [ + { + "bbox": [ + 347, + 98, + 396, + 245 + ], + "lines": [ + { + "bbox": [ + 347, + 98, + 396, + 245 + ], + "spans": [ + { + "bbox": [ + 347, + 98, + 396, + 245 + ], + "type": "image", + "image_path": "cd939911703bf1ea774a6a074cd4f04fdc646cabd88727d485cc8bf8baf84281.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 249, + 380, + 257 + ], + "lines": [ + { + "bbox": [ + 364, + 249, + 380, + 257 + ], + "spans": [ + { + "bbox": [ + 364, + 249, + 380, + 257 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 402, + 98, + 452, + 245 + ], + "blocks": [ + { + "bbox": [ + 402, + 98, + 452, + 245 + ], + "lines": [ + { + "bbox": [ + 402, + 98, + 452, + 245 + ], + "spans": [ + { + "bbox": [ + 402, + 98, + 452, + 245 + ], + "type": "image", + "image_path": "436fa484293c46f7d11dfbd38d909b935b5f6a49dabde97688b9499937608064.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 248, + 448, + 257 + ], + "lines": [ + { + "bbox": [ + 406, + 248, + 448, + 257 + ], + "spans": [ + { + "bbox": [ + 406, + 248, + 448, + 257 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 100, + 297, + 452, + 506 + ], + "blocks": [ + { + "bbox": [ + 45, + 259, + 531, + 291 + ], + "lines": [ + { + "bbox": [ + 45, + 259, + 531, + 291 + ], + "spans": [ + { + "bbox": [ + 45, + 259, + 531, + 291 + ], + "type": "text", + "content": "Figure 5: Qualitative comparison on the SynWoodScape dataset. Sky-blue boxes are enlarged views of specific objects to show detail. Overall, our method shows distinct boundaries compared to the other models. In particular, we can observe the detailed structure from our results, such as bicycle wheels in the first row and the boundary of two overlapped vehicles in the second row." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 100, + 297, + 452, + 506 + ], + "lines": [ + { + "bbox": [ + 100, + 297, + 452, + 506 + ], + "spans": [ + { + "bbox": [ + 100, + 297, + 452, + 506 + ], + "type": "image", + "image_path": "0413c375d7b9bb8d32226b85d47b58ae94b75b287ba2cf54536a5f369a2e20ca.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 511, + 531, + 533 + ], + "lines": [ + { + "bbox": [ + 45, + 511, + 531, + 533 + ], + "spans": [ + { + "bbox": [ + 45, + 511, + 531, + 533 + ], + "type": "text", + "content": "Figure 6: Qualitative comparison on the KITTI-360 depth dataset. For visualization, we crop the object region in the fisheye image. In contrast to the other methods, our method shows clear boundaries near the object regions, as shown in sky-blue and green boxes." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 45, + 544, + 279, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 544, + 279, + 627 + ], + "spans": [ + { + "bbox": [ + 45, + 544, + 279, + 627 + ], + "type": "text", + "content": "10 epochs on SynWoodScape and 20 epochs on KITTI-360 depth using AdamW [28] optimizer with weight-decay " + }, + { + "bbox": [ + 45, + 544, + 279, + 627 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 45, + 544, + 279, + 627 + ], + "type": "text", + "content": ", with a batch size of 8. We use the same 1-cycle policy [35] as AdaBins [3] for the learning rate with max " + }, + { + "bbox": [ + 45, + 544, + 279, + 627 + ], + "type": "inline_equation", + "content": "\\mathrm{lr} = 3.5 \\times 10^{-4}" + }, + { + "bbox": [ + 45, + 544, + 279, + 627 + ], + "type": "text", + "content": ". For the first " + }, + { + "bbox": [ + 45, + 544, + 279, + 627 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 45, + 544, + 279, + 627 + ], + "type": "text", + "content": " of iterations, the linear learning rate warm-up technique is utilized. For the learning rate decay, the cosine annealing learning rate strategy is applied." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 45, + 639, + 124, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 639, + 124, + 652 + ], + "spans": [ + { + "bbox": [ + 45, + 639, + 124, + 652 + ], + "type": "text", + "content": "4.3. Comparison" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": "We compare our SlaBins with several baseline approaches: OmniDet [20], BTS [21], and AdaBins [3]. OmniDet is a fisheye depth estimation method using self-supervised learning. Based on the released code of the authors, we modify OmniDet to a supervised one for fairness (we denote this modified one OmniDet*). As base" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 544, + 530, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 544, + 530, + 614 + ], + "spans": [ + { + "bbox": [ + 296, + 544, + 530, + 614 + ], + "type": "text", + "content": "line methods for pinhole cameras, we compare BTS and AdaBins, especially AdaBins is a bins-based fundamental method for depth estimation. It should be noted that we target depth estimation fisheye cameras, but we additionally compare it with conventional pinhole camera-based methods due to the lack of fisheye camera-based methods." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 621, + 530, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 621, + 530, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 621, + 530, + 714 + ], + "type": "text", + "content": "Quantitative evaluation. Table 1 shows the quantitative comparisons on the SynWoodScape dataset. The arrow next to the metric indicates the direction of better performance. Overall, SlaBins outperforms the other baseline approaches [3, 20, 21]. In particular, RMSE and Abs-Rel errors are significantly improved; " + }, + { + "bbox": [ + 296, + 621, + 530, + 714 + ], + "type": "inline_equation", + "content": "31.7\\%" + }, + { + "bbox": [ + 296, + 621, + 530, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 621, + 530, + 714 + ], + "type": "inline_equation", + "content": "68.6\\%" + }, + { + "bbox": [ + 296, + 621, + 530, + 714 + ], + "type": "text", + "content": " decreased compared to AdaBins, which shows the effect of the slanted MCI representation in the fisheye domain." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 718, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 718, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 308, + 718, + 530, + 730 + ], + "type": "text", + "content": "On the KITTI-360 depth dataset, our method still outper" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "8771" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 104, + 259, + 232 + ], + "blocks": [ + { + "bbox": [ + 65, + 104, + 259, + 232 + ], + "lines": [ + { + "bbox": [ + 65, + 104, + 259, + 232 + ], + "spans": [ + { + "bbox": [ + 65, + 104, + 259, + 232 + ], + "type": "image", + "image_path": "2212e2af948023b583428b260616d7d55ad63e5f051987bd8e26db16f5727976.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 233, + 279, + 287 + ], + "lines": [ + { + "bbox": [ + 45, + 233, + 279, + 287 + ], + "spans": [ + { + "bbox": [ + 45, + 233, + 279, + 287 + ], + "type": "text", + "content": "Figure 7: Depth histogram on the object regions. Left: RGB data (top) and the concept of depth histogram in the orthogonal coordinate (bottom). Middle: depth maps of only object regions for each model and the ground truth. Right: depth histograms (AdaBins, original MCI, slanted MCI, ground truth in that order)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 290, + 279, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 290, + 279, + 383 + ], + "spans": [ + { + "bbox": [ + 45, + 290, + 279, + 383 + ], + "type": "text", + "content": "forms the comparison methods (see Table 2). Specifically, RMSE and Sq-Rel errors are improved; " + }, + { + "bbox": [ + 45, + 290, + 279, + 383 + ], + "type": "inline_equation", + "content": "13.9\\%" + }, + { + "bbox": [ + 45, + 290, + 279, + 383 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 290, + 279, + 383 + ], + "type": "inline_equation", + "content": "27.1\\%" + }, + { + "bbox": [ + 45, + 290, + 279, + 383 + ], + "type": "text", + "content": " decreased compared to BTS and " + }, + { + "bbox": [ + 45, + 290, + 279, + 383 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 45, + 290, + 279, + 383 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 290, + 279, + 383 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 45, + 290, + 279, + 383 + ], + "type": "text", + "content": " decreased compared to AdaBins. Note that even though the KITTI-360 depth dataset captured in suburban scenes does not perfectly fit our scenario of road environments, SlaBins still shows competitive performance, which demonstrates the effect of the proposed framework." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 387, + 279, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 387, + 279, + 596 + ], + "spans": [ + { + "bbox": [ + 45, + 387, + 279, + 596 + ], + "type": "text", + "content": "Qualitative evaluation. Figure 5 shows the estimated depth maps for the test scenes on SynWoodScape. Due to the nature of fisheye cameras, we can observe ambiguous boundary regions in the comparison methods. Even AdaBins, which is our baseline and a naive MSI representation by the Euclidean distance of fisheye, shows bleeding depth estimation results near the boundaries between objects and ground regions. In contrast, thanks to the slanted MCI representation, SlaBins shows accurate depth results with distinct boundaries regardless of camera viewing directions. Specifically, we can observe that sharp objects with details (e.g., human legs and bike) are preserved in our method, unlike the other methods that smooth object details. In addition, we can observe a similar trend on the KITTI-360 depth dataset, as shown in Fig. 6. Compared to the baseline methods, SlaBins shows better depth consistency even on sharp objects (e.g., third and fourth columns in Fig. 6). Additional results are available in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 606, + 135, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 606, + 135, + 618 + ], + "spans": [ + { + "bbox": [ + 45, + 606, + 135, + 618 + ], + "type": "text", + "content": "4.4. Ablation study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 624, + 278, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 624, + 278, + 729 + ], + "spans": [ + { + "bbox": [ + 45, + 624, + 278, + 729 + ], + "type": "text", + "content": "Multi-layered representation. We compare the proposed slanted MCI with the MSI used in AdaBins and the original MCI on the orthogonal coordinate. We measure the depth histogram for only object regions on the orthogonal coordinate. As shown in Fig. 7, compared to the other representations, SlaBins reveals the most similar trend (i.e., peak points) with the ground truth histogram. Furthermore, we compare slanted MCI and original MCI in terms of depth consistency. Figure 8 shows that the slanted MCI provides" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 305, + 103, + 520, + 191 + ], + "blocks": [ + { + "bbox": [ + 305, + 103, + 520, + 191 + ], + "lines": [ + { + "bbox": [ + 305, + 103, + 520, + 191 + ], + "spans": [ + { + "bbox": [ + 305, + 103, + 520, + 191 + ], + "type": "image", + "image_path": "620ba8fd308f1615bb8ba8f2c0ed397c2b10ba1a9f8cae01b712a742bb836b65.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 192, + 530, + 224 + ], + "lines": [ + { + "bbox": [ + 296, + 192, + 530, + 224 + ], + "spans": [ + { + "bbox": [ + 296, + 192, + 530, + 224 + ], + "type": "text", + "content": "Figure 8: Slanted MCI vs. original MCI in depth and clustering. All the results including the ground truth are converted to the XZ distance in the orthogonal coordinate." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 333, + 275, + 496, + 327 + ], + "blocks": [ + { + "bbox": [ + 296, + 229, + 530, + 273 + ], + "lines": [ + { + "bbox": [ + 296, + 229, + 530, + 273 + ], + "spans": [ + { + "bbox": [ + 296, + 229, + 530, + 273 + ], + "type": "text", + "content": "Table 3: Ablation study for the slanted angle predictor on SynWoodScape and KITTI-360 depth. We report the depth error (Abs Rel) and mean angle error. C and R denote classification-only and regression-only, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 333, + 275, + 496, + 327 + ], + "lines": [ + { + "bbox": [ + 333, + 275, + 496, + 327 + ], + "spans": [ + { + "bbox": [ + 333, + 275, + 496, + 327 + ], + "type": "table", + "html": "
MethodSynWoodScapeKITTI-360 depth
Abs Rel ↓Angle error↓Abs Rel ↓Angle error
R0.07527.0310.60448.32
C0.0260.8420.1035
C+R (ours)0.0220.0310.0770.017
", + "image_path": "646d053359c18a99b902eaf45cd69b457ebc9b56b3ca22b459af7d26c6432a9c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 334, + 530, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 334, + 530, + 380 + ], + "spans": [ + { + "bbox": [ + 296, + 334, + 530, + 380 + ], + "type": "text", + "content": "more consistent results in object regions both on depth and depth-based clustering results. Hence, we believe that the slanted MCI can provide robust geometric constraints for various tasks in fisheye domains." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 383, + 530, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 383, + 530, + 489 + ], + "spans": [ + { + "bbox": [ + 296, + 383, + 530, + 489 + ], + "type": "text", + "content": "Slanted angle prediction. As shown in Table 3, we perform an ablation study for the angle prediction module. Unlike a conventional regression module having the advantage over classifications, in our case, it is hard to converge because of the trade-off with the depth estimation part in the SlaBins module. That is why regression-only module is poor on the angle prediction. On the contrary, our two-step prediction module, composed of classification and residual regression, shows the best performance with stable training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 297, + 497, + 367, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 497, + 367, + 510 + ], + "spans": [ + { + "bbox": [ + 297, + 497, + 367, + 510 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 517, + 530, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 517, + 530, + 644 + ], + "spans": [ + { + "bbox": [ + 296, + 517, + 530, + 644 + ], + "type": "text", + "content": "We have proposed a novel fisheye depth estimation network, called SlaBins, based on the geometric property of the road environments. We newly introduce the slanted MCI representation, which allows us to represent a given road scene as a set of cylindrical layers orthogonal to the road ground. By virtue of the slanted MCI, our network can implicitly encode scene depth information invariant to the camera viewing direction while preserving boundary information of scene objects w.r.t. the road region. Thus, we achieve state-of-the-art performance on monocular fisheye depth estimation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 648, + 530, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 648, + 530, + 729 + ], + "spans": [ + { + "bbox": [ + 296, + 648, + 530, + 729 + ], + "type": "text", + "content": "Acknowledgments. This work was supported by 42dot Inc., Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2020-0-01336, Artificial Intelligence Graduate School Program (UNIST)) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF2021R1C1C1005723)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "8772" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 105, + 102, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 105, + 102, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 102, + 117 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 124, + 278, + 730 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 52, + 124, + 278, + 166 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 124, + 278, + 166 + ], + "spans": [ + { + "bbox": [ + 52, + 124, + 278, + 166 + ], + "type": "text", + "content": "[1] T-Bone Accident Statistics And Other Legal Information. https://www.injurytriallawyer.com/faqs/ how-dangerous-are-t-bone-accidents-.cfm. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 52, + 169, + 278, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 169, + 278, + 201 + ], + "spans": [ + { + "bbox": [ + 52, + 169, + 278, + 201 + ], + "type": "text", + "content": "[2] Ashutosh Agarwal and Chetan Arora. Attention Attention Everywhere: Monocular Depth Prediction with Skip Attention. arXiv preprint arXiv:2210.09071, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 202, + 278, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 202, + 278, + 233 + ], + "spans": [ + { + "bbox": [ + 52, + 202, + 278, + 233 + ], + "type": "text", + "content": "[3] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In CVPR, 2021. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 235, + 278, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 235, + 278, + 266 + ], + "spans": [ + { + "bbox": [ + 52, + 235, + 278, + 266 + ], + "type": "text", + "content": "[4] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. LocalBins: Improving Depth Estimation by Learning Local Distributions. In ECCV, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 269, + 278, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 269, + 278, + 321 + ], + "spans": [ + { + "bbox": [ + 52, + 269, + 278, + 321 + ], + "type": "text", + "content": "[5] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, et al. PersFormer: 3D Lane Detection via Perspective Transformer and the OpenLane Benchmark. arXiv preprint arXiv:2203.11089, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 323, + 278, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 323, + 278, + 355 + ], + "spans": [ + { + "bbox": [ + 52, + 323, + 278, + 355 + ], + "type": "text", + "content": "[6] Liang-Chieh Chen, Huiyu Wang, and Siyuan Qiao. Scaling wide residual networks for panoptic segmentation. arXiv preprint arXiv:2011.11675, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 52, + 356, + 278, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 356, + 278, + 389 + ], + "spans": [ + { + "bbox": [ + 52, + 356, + 278, + 389 + ], + "type": "text", + "content": "[7] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In CVPR, 2017. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 389, + 278, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 389, + 278, + 432 + ], + "spans": [ + { + "bbox": [ + 52, + 389, + 278, + 432 + ], + "type": "text", + "content": "[8] Olivier Faugeras and Quang-Tuan Luong. The geometry of multiple images: the laws that govern the formation of multiple images of a scene and some of their applications. MIT press, 2001. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 52, + 433, + 278, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 433, + 278, + 465 + ], + "spans": [ + { + "bbox": [ + 52, + 433, + 278, + 465 + ], + "type": "text", + "content": "[9] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 467, + 278, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 278, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 278, + 498 + ], + "type": "text", + "content": "[10] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 500, + 278, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 500, + 278, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 500, + 278, + 532 + ], + "type": "text", + "content": "[11] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In CVPR, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 532, + 278, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 278, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 278, + 564 + ], + "type": "text", + "content": "[12] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 565, + 278, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 565, + 278, + 607 + ], + "spans": [ + { + "bbox": [ + 47, + 565, + 278, + 607 + ], + "type": "text", + "content": "[13] Yuxuan Han, Ruicheng Wang, and Jiaolong Yang. Single-View View Synthesis in the Wild with Learned Adaptive Multiplane Images. arXiv preprint arXiv:2205.11733, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 609, + 278, + 641 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 278, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 278, + 641 + ], + "type": "text", + "content": "[14] Adrian Johnston and Gustavo Carneiro. Self-supervised monocular trained depth estimation using self-attention and discrete disparity volume. In CVPR, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 642, + 278, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 278, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 278, + 685 + ], + "type": "text", + "content": "[15] Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, and Junmo Kim. Global-Local Path Networks for Monocular Depth Estimation with Vertical Cut-Depth. arXiv preprint arXiv:2201.07436, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 687, + 278, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 687, + 278, + 730 + ], + "spans": [ + { + "bbox": [ + 47, + 687, + 278, + 730 + ], + "type": "text", + "content": "[16] Varun Ravi Kumar, Sandesh Athni Hiremath, Markus Bach, Stefan Milz, Christian Witt, Clément Pinard, Senthil Yogamani, and Patrick Mäder. Fisheyedistancenet: Self-supervised scale-aware distance estimation using monocular" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 107, + 530, + 730 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 318, + 107, + 530, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 107, + 530, + 128 + ], + "spans": [ + { + "bbox": [ + 318, + 107, + 530, + 128 + ], + "type": "text", + "content": "fisheye camera for autonomous driving. In ICRA, 2020. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 130, + 530, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 130, + 530, + 181 + ], + "spans": [ + { + "bbox": [ + 299, + 130, + 530, + 181 + ], + "type": "text", + "content": "[17] Varun Ravi Kumar, Marvin Klingner, Senthil Yogamani, Markus Bach, Stefan Milz, Tim Fingscheidt, and Patrick Mäder. SVDistNet: Self-supervised near-field distance estimation on surround view fisheye cameras. IEEE Transactions on Intelligent Transportation Systems, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 183, + 530, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 183, + 530, + 245 + ], + "spans": [ + { + "bbox": [ + 299, + 183, + 530, + 245 + ], + "type": "text", + "content": "[18] Varun Ravi Kumar, Stefan Milz, Christian Witt, Martin Simon, Karl Amende, Johannes Petzold, Senthil Yogamani, and Timo Pech. Monocular fisheye camera depth estimation using sparse lidar supervision. In IEEE International Conference on Intelligent Transportation Systems (ITSC), 2018. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 246, + 530, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 246, + 530, + 298 + ], + "spans": [ + { + "bbox": [ + 299, + 246, + 530, + 298 + ], + "type": "text", + "content": "[19] Varun Ravi Kumar, Senthil Yogamani, Markus Bach, Christian Witt, Stefan Milz, and Patrick Mäder. Unrectdepthnet: Self-supervised monocular depth estimation using a generic framework for handling common camera distortion models. In IROS, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 300, + 530, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 300, + 530, + 362 + ], + "spans": [ + { + "bbox": [ + 299, + 300, + 530, + 362 + ], + "type": "text", + "content": "[20] Varun Ravi Kumar, Senthil Yogamani, Hazem Rashed, Ganesh Sitsu, Christian Witt, Isabelle Leang, Stefan Milz, and Patrick Mader. Omnidet: Surround view cameras based multi-task visual perception network for autonomous driving. IEEE Robotics and Automation Letters, 6(2):2830-2837, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 364, + 530, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 364, + 530, + 405 + ], + "spans": [ + { + "bbox": [ + 299, + 364, + 530, + 405 + ], + "type": "text", + "content": "[21] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 2, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 407, + 530, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 407, + 530, + 439 + ], + "spans": [ + { + "bbox": [ + 299, + 407, + 530, + 439 + ], + "type": "text", + "content": "[22] Jiaxin Li, Zijian Feng, Qi She, Henghui Ding, Changhu Wang, and Gim Hee Lee. Mine: Towards continuous depth spi with nerf for novel view synthesis. In ICCV, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 440, + 530, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 440, + 530, + 481 + ], + "spans": [ + { + "bbox": [ + 299, + 440, + 530, + 481 + ], + "type": "text", + "content": "[23] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. DepthFormer: Exploiting Long-Range Correlation and Local Information for Accurate Monocular Depth Estimation. arXiv preprint arXiv:2203.14211, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 483, + 530, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 483, + 530, + 514 + ], + "spans": [ + { + "bbox": [ + 299, + 483, + 530, + 514 + ], + "type": "text", + "content": "[24] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE TPAMI, 2022. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 515, + 530, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 515, + 530, + 557 + ], + "spans": [ + { + "bbox": [ + 299, + 515, + 530, + 557 + ], + "type": "text", + "content": "[25] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETRv2: A Unified Framework for 3D Perception from Multi-Camera Images. arXiv preprint arXiv:2206.01256, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 299, + 558, + 530, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 558, + 530, + 588 + ], + "spans": [ + { + "bbox": [ + 299, + 558, + 530, + 588 + ], + "type": "text", + "content": "[26] Yuxuan Liu, Yuan Yixuan, and Ming Liu. Ground-aware monocular 3d object detection for autonomous driving. IEEE Robotics and Automation Letters, 6(2):919-926, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 299, + 590, + 530, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 590, + 530, + 631 + ], + "spans": [ + { + "bbox": [ + 299, + 590, + 530, + 631 + ], + "type": "text", + "content": "[27] Xiaoxiao Long, Cheng Lin, Lingjie Liu, Wei Li, Christian Theobalt, Ruigang Yang, and Wenping Wang. Adaptive surface normal constraint for depth estimation. In ICCV, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 299, + 633, + 530, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 633, + 530, + 653 + ], + "spans": [ + { + "bbox": [ + 299, + 633, + 530, + 653 + ], + "type": "text", + "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In *ICLR*, 2018. 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 299, + 655, + 530, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 655, + 530, + 696 + ], + "spans": [ + { + "bbox": [ + 299, + 655, + 530, + 696 + ], + "type": "text", + "content": "[29] Reza Mahjourian, Martin Wicke, and Anelia Angelova. Unsupervised learning of depth and ego-motion from monocular video using 3d geometric constraints. In CVPR, 2018. 2, 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 299, + 698, + 530, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 698, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 299, + 698, + 530, + 730 + ], + "type": "text", + "content": "[30] Vaishakh Patil, Christos Sakaridis, Alexander Liniger, and Luc Van Gool. P3Depth: Monocular Depth Estimation with a Piecewise Planarity Prior. In CVPR, 2022. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "8773" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 674 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "type": "text", + "content": "[31] Elad Plaut, Erez Ben Yaacov, and Bat El Shlomo. 3d object detection from a single fisheye image without a single fisheye training image. In CVPR, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 140, + 279, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 279, + 161 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 279, + 161 + ], + "type": "text", + "content": "[32] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 163, + 279, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 279, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 279, + 204 + ], + "type": "text", + "content": "[33] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 206, + 279, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 206, + 279, + 259 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 279, + 259 + ], + "type": "text", + "content": "[34] Ahmed Rida Sekkat, Yohan Dupuis, Varun Ravi Kumar, Hazem Rashed, Senthil Yogamani, Pascal Vasseur, and Paul Honeine. SynWoodScape: Synthetic Surround-View Fisheye Camera Dataset for Autonomous Driving. IEEE Robotics and Automation Letters, 7(3):8502-8509, jul 2022. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 261, + 279, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 261, + 279, + 313 + ], + "spans": [ + { + "bbox": [ + 47, + 261, + 279, + 313 + ], + "type": "text", + "content": "[35] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, volume 11006, pages 369-386. SPIE, 2019. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 314, + 279, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 314, + 279, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 279, + 335 + ], + "type": "text", + "content": "[36] Richard Szeliski and Polina Golland. Stereo matching with transparency and matting. In ICCV, 1998. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 337, + 279, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 337, + 279, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 337, + 279, + 378 + ], + "type": "text", + "content": "[37] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 380, + 279, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 279, + 412 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 279, + 412 + ], + "type": "text", + "content": "[38] Andrew Tao, Karan Sapra, and Bryan Catanzaro. Hierarchical multi-scale attention for semantic segmentation. arXiv preprint arXiv:2005.10821, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 413, + 279, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 413, + 279, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 413, + 279, + 434 + ], + "type": "text", + "content": "[39] Richard Tucker and Noah Snively. Single-view view synthesis with multiplane images. In CVPR, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 436, + 279, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 279, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 279, + 478 + ], + "type": "text", + "content": "[40] Qingan Yan, Pan Ji, Nitin Bansal, Yuxin Ma, Yuan Tian, and Yi Xu. FisheyeDistill: Self-Supervised Monocular Depth Estimation with Ordinal Distillation for Fisheye Cameras. arXiv preprint arXiv:2205.02930, 2022. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 479, + 279, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 279, + 531 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 279, + 531 + ], + "type": "text", + "content": "[41] Yaozu Ye, Kailun Yang, Kaite Xiang, Juan Wang, and Kaiwei Wang. Universal semantic segmentation for fisheye urban driving images. In IEEE International Conference on Systems, Man, and Cybernetics (SMC), pages 648-655, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 532, + 279, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 279, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 279, + 586 + ], + "type": "text", + "content": "[42] Senthil Yogamani, Ciarán Hughes, Jonathan Horgan, Ganesh Sistu, Padraig Varley, Derek O'Dea, Michal Uricár, Stefan Milz, Martin Simon, Karl Amende, et al. WoodScape: A multi-task, multi-camera fisheye dataset for autonomous driving. arXiv preprint arXiv:1905.01489, 2019. 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 587, + 279, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 279, + 629 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 279, + 629 + ], + "type": "text", + "content": "[43] Yifan Zhang, Qijian Zhang, Zhiyu Zhu, Junhui Hou, and Yixuan Yuan. GLENet: Boosting 3D Object Detectors with Generative Label Uncertainty Estimation. arXiv preprint arXiv:2207.02466, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 631, + 279, + 674 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 631, + 279, + 674 + ], + "spans": [ + { + "bbox": [ + 47, + 631, + 279, + 674 + ], + "type": "text", + "content": "[44] Dingfu Zhou, Jin Fang, Xibin Song, Liu Liu, Junbo Yin, Yuchao Dai, Hongdong Li, and Ruigang Yang. Joint 3d instance segmentation and object detection for autonomous driving. In CVPR, 2020. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "8774" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning/231f8e4e-8ec6-40cc-a455-514ae44e6fa1_content_list.json b/2023/Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning/231f8e4e-8ec6-40cc-a455-514ae44e6fa1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d21e2828fa4ab77c692314d7c6f976920ad8d316 --- /dev/null +++ b/2023/Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning/231f8e4e-8ec6-40cc-a455-514ae44e6fa1_content_list.json @@ -0,0 +1,1826 @@ +[ + { + "type": "text", + "text": "Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning", + "text_level": 1, + "bbox": [ + 166, + 130, + 802, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiang Yuan Gong Cheng * Kebing Yan Qinghua Zeng Junwei Han School of Automation, Northwestern Polytechnical University, Xi'an, China {shaunyuan, kebingyan, zengqinghua}@mail.nwpu.edu.cn, {gcheng, jhan}@nwpu.edu.cn", + "bbox": [ + 133, + 203, + 833, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The past few years have witnessed the immense success of object detection, while current excellent detectors struggle on tackling size-limited instances. Concretely, the well-known challenge of low overlaps between the priors and object regions leads to a constrained sample pool for optimization, and the paucity of discriminative information further aggravates the recognition. To alleviate the aforementioned issues, we propose CFINet, a two-stage framework tailored for small object detection based on the Coarse-to-fine pipeline and Feature Imitation learning. Firstly, we introduce Coarse-to-fine RPN (CRPN) to ensure sufficient and high-quality proposals for small objects through the dynamic anchor selection strategy and cascade regression. Then, we equip the conventional detection head with a Feature Imitation (FI) branch to facilitate the region representations of size-limited instances that perplex the model in an imitation manner. Moreover, an auxiliary imitation loss following supervised contrastive learning paradigm is devised to optimize this branch. When integrated with Faster RCNN, CFINet achieves state-of-the-art performance on the large-scale small object detection benchmarks, SODA-D and SODA-A, underscoring its superiority over baseline detector and other mainstream detection approaches. Code is available at https://github.com/shaunyuan22/CFINet.", + "bbox": [ + 76, + 323, + 473, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 715, + 209, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Small object detection (SOD) aims to classify and localize the instances with limited regions, which plays an important role in a wide range of scenarios, such as pedestrian detection, autonomous driving, and intelligent surveillance understanding, to name a few [42, 28, 22, 19, 36, 2, 20]. Compared to the generic object detection which has been", + "bbox": [ + 75, + 739, + 468, + 832 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c37fa509a0e67644577a7ce2fba591e7ecec0af059b50881524a43feb10d684c.jpg", + "image_caption": [ + "Figure 1. Distribution of maximum IoU of anchors matched to each ground-truth instance in SODA-D [9] train-set, where extremely Small (eS), relatively Small (rS), and generally Small (gS) correspond to three area subsets in SODA [9] with the ranges (0, 144], (144, 400] and (400, 1024]. The smaller the objects are, the lower IoU the matched anchors have, hence the commonly used positive IoU threshold (0.7) is too rigorous for small objects." + ], + "image_footnote": [], + "bbox": [ + 503, + 290, + 640, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c95f8e6f6e719c3c11a981aa942d2e24e7e5f5292ce20f61c636f77ca6ff9040.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 290, + 767, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9307cc36ff082cb685b821386ade2a48547d27db4f897e62781ae5052b023836.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 291, + 890, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "extensively studied, SOD task receives relatively little attention and good solutions are still scarce so far. Moreover, generic detectors [29, 6, 25, 31, 7, 3, 26, 5] usually struggle on handling small objects due to two inherent challenges: the insufficient and low-quality samples for training and the uncertain prediction of RoIs (Region of Interests).", + "bbox": [ + 496, + 551, + 893, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "First, current prevailing detectors exploit either overlap-based [29, 26] or distance-based [31] strategies to select the positive priors of objects for training. However, small instances usually occupy an extremely limited area, therefore the region overlaps between densely arranged anchors and ground truth boxes are significantly small and far from the commonly used positive IoU (Intersection-over-Union) threshold, as in Figure 1. In other words, the existing positive sample criterion is overly stringent when applied to small/tiny objects, resulting in a restricted number of samples available for optimization. An intuitive approach involves reducing the threshold for defining a positive sample [47]. However, while this can lead to an increase in the number of positive samples, it often at the expense of overall sample quality, in which low-quality samples disrupt optimization and incur a trivial regression solution. Worse still, this is actually contradictory to the purpose of proposal", + "bbox": [ + 496, + 643, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: gcheng@nwpu.edu.cn", + "bbox": [ + 94, + 839, + 344, + 851 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This paper focuses on the detection of \"pure\" small objects, where the scales of all the objects are distributed within a relatively tight range [45, 34, 9], which is distinct from the known small objects in multi-scale object detection [24].", + "bbox": [ + 76, + 851, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6317", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/f63e2c88f04a097accad854273716ee9dda513183982d9b818b2a52a3e4b7d47.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. The performances of Cascade RPN [32] compared to the baseline (a vanilla Faster RCNN [29]). The ctr/ign ratio denotes the sampling region in first regression stage of Cascade RPN. The results are tested on the SODA-D [9] test-set and with a ResNet-50 [17] as the backbone." + ], + "table_body": "
ctrl/ign ratioAPAP50AP75APesSAPrsSAPgSAPN
Baseline28.959.424.113.825.734.543.0
0.2/0.529.156.525.912.525.535.444.7
0.5/0.829.557.826.013.526.235.845.0
0.8/1.027.554.124.311.223.833.742.5
", + "bbox": [ + 81, + 88, + 467, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "network, i.e., guaranteeing the recall and ease the burden of subsequent work.", + "bbox": [ + 76, + 244, + 467, + 275 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To sum up, current prevailing priors-to-proposals paradigms that heavily depend on the overlap or distance metric have inherent limitations in detecting small objects, and nowadays devised assignment or sampling schemes contribute minimally to this problem [9, 40]. Since the proposals play such a crucial role in two-stage detectors, so how about the improved Region Proposal Network (RPN) variants [13, 32, 33] meet small object detection? Following this line, we take Cascade RPN [32], one of the most superior proposal network for generic object detection, to perform preliminary experiments and the results are shown in Table 1. While the auxiliary regression phase provides refined priors with better initialization for subsequent regression, the final results remain somewhat unsatisfactory. Specifically, the improvement mainly comes from the larger objects and the $AP_{eS}$ as well as $AP_{rS}$ actually decrease significantly instead, indicating that the region-based sampling strategy is inclined to large instances which further dominate the proposal network. Meanwhile, enlarging the sample region contributes little (even negative) to this condition (see the bottom row in Table 1). Therefore, the coarse-to-fine pipeline has the potential to surmount the barrier of conventional prior-to-proposal paradigm, but the crux lies in dedicating sufficient attention to small instances.", + "bbox": [ + 76, + 277, + 467, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, small objects usually lack discriminative information and distorted structures, leading the inclination of model to give ambiguous even incorrect predictions [1, 11]. Meanwhile, there are a certain amount of large instances embodying clear visual cues and better discrimination. Building upon this observation, several works proposed to bridge the representation gap between small objects and large ones, and most of them [2, 22, 28, 1] rely on Generative Adversarial Network (GAN) [14] or similarity learning [19, 36] to super-resolve/restore the features of size-limited instances under the guidance of large ones that are deemed to be visually authentic. However, these approaches overlook the fact: high quality $\\neq$ large size meanwhile small size $\\neq$ low quality. In other words, the criterion for humans and the model to decide whether a sample is competent to be a good example is distinct. For the latter, it is dynamic and should be adjusted according to", + "bbox": [ + 76, + 643, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the current optimization of the detector. Moreover, efforts in this line have to resort to sophisticated training strategies or additional models, which is time-consuming and break the conventional end-to-end paradigm.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Putting the above parts together, we propose a two-stage small object detector CFINet based on the coarse-to-fine pipeline and feature imitation learning. Concretely, enlightened by the multi-stage proposal generation scheme in Cascade RPN, we devise Coarse-to-fine RPN (CRPN). It firstly employs an dynamic anchor selection strategy to mine potential priors to conduct coarse regression, and henceforth, these refined anchors will be classified and regressed by the region proposal network. In addition, we extend the conventional classification-and-regression setting with an auxiliary Feature Imitation (FI) branch, which can leverage the regional features of high-quality instances to guide the learning of those objects with uncertain/mistaken predictions, and a loss function based on the Supervised Contrastive Learning (SCL) [18] is designed to optimize the whole process. The main contributions of this paper are summarized as follows:", + "bbox": [ + 496, + 152, + 890, + 406 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A coarse-to-fine proposal generation pipeline named CRPN was built to perform anchor-to-proposal procedure, where an area-based anchor mining strategy and cascade regression empower the high-quality proposals for small instances.", + "- An auxiliary Feature Imitation (FI) branch was introduced to enrich the representations of low-quality instances perplexing the model under the supervision of high-quality instances, and this novel branch is optimized by a tailored loss function based on SCL.", + "- The experiment results on the SODA-D and SODA-A datasets exhibit the superiority of our CFINet to detect these instances with extremely limited sizes." + ], + "bbox": [ + 517, + 415, + 890, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 648, + 648, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Anchor Refinement and Region Proposals. Two-stage anchor-based approaches heavily depend on the high-quality proposals [33, 32]. Towards this goal, RPN was first introduced in Faster RCNN [29] to produce proposals in a fully convolutional network, and this simple yet effective design facilitates end-to-end model optimization. Following RPN, [13] proposed to iteratively regress the predefined anchors. GA-RPN [33] discards the uniform anchoring strategy and formulates the anchor generation in two steps: first determining the locations which may contain objects and on which the anchor scales are predicted then. By installing a multi-stage anchor-to-proposal strategy and alleviating the misalignment between the refined anchors and image features, Cascade RPN [32] enables high-quality proposal generation. Unfortunately, current proposal-oriented", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6318", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/70bd0e14cf5547c9f29db918e444f9f99a5355b3bfc96bfb9fd021ab91b8c27e.jpg", + "image_caption": [ + "Figure 2. The overall architecture of CFINet. In Coarse-to-fine RPN (CRPN), the area-based Anchor Mining strategy ensures sufficient candidates for instances of various sizes (small: orange boxes, large: green boxes) based on the dynamic pos threshold, which will be then used to obtain the coarse proposals. After that, the alignment between coarse proposals and corresponding features is enabled by the Adaptive Convolution before feeding into the RPN to produce high-quality proposals (blue boxes). The Feature Imitation (FI) branch is devised to facilitate the representations of small instances, in which the RoI features of uncertain/mistaken predictions will be pulled to their counterparts of exemplar feature set in embedding space (throughout the Feat2Embed module), while pushed apart from the exemplar features of other categories and background. And the exemplar features are collected based on the model predictions using the proposed quality indicator, i.e., Instance Quality $(IQ)$ . We tailor a Feature Imitation loss function $\\mathcal{L}_{\\mathrm{FI}}$ to optimize this auxiliary branch. Note that we only exhibit single-level Feature Pyramid Network (FPN) [25] feature for clear illustration." + ], + "image_footnote": [], + "bbox": [ + 80, + 85, + 493, + 351 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e11300be834b1610a6e82ed6ecf56fbdfbff87f116a3d2dd9121205f9c4315a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 85, + 911, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "frameworks fail to produce high-quality proposals for instances with limited regions, and the root cause lies in the notoriously low overlaps between the objects and priors [40]. Different from the above methods, our coarse-to-fine proposal pipeline could exploit the potential of multi-stage refinement paradigm, thereby guaranteeing both the quantity and quality of proposals for instances with extremely limited sizes.", + "bbox": [ + 75, + 502, + 467, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature Imitation for Small Object Detection. One of the major challenges to detect small objects is the low-quality representation [45, 40, 9], while large instances often with clear structures and discriminative features. Hence, a series of efforts have been made to boost the semantic representations of small/tiny instances by mining the intrinsic correlations between small and large objects. Based on the generative adversarial paradigm, Perceptual GAN [22] designs a generator that is optimized to produce high-quality representations of small instances to fool the subsequent discriminator. Bai et al. [1] devised a novel pipeline to restore a clear face from the inputting blurry one. Noh et al. [28] further introduced precise supervision for the super-resolution process of small objects. Moreover, Wu et al. [36] and Kim et al. [19] both exploited similarity learning to force the features of small-scale pedestrians close to that of the large-scale ones which are obtained by an additional model. The existence of super-resolution branch or offline", + "bbox": [ + 75, + 628, + 467, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "feature bank hampers the end-to-end optimization while our method updates the exemplar features in an online fashion, which guarantees the diversity of high-quality feature sets thereby getting rid of the collapse issue.", + "bbox": [ + 496, + 502, + 890, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive Learning for Object Detection. The recent explosion of self-supervised learning mainly comes from its Contrastive Learning fashion, and several works have extended this paradigm into detection fields. Detco [38] is an effective self-supervised framework for object detection which utilizes the image and its local patches to conduct contrastive learning. Wu et al. [37] applied contrastive learning to object detection under smoky conditions. Though contrastive learning has recently received considerable interests [4, 16, 35], the potential of utilizing contrastive learning for better representation of small objects has not yet been investigated to date.", + "bbox": [ + 496, + 565, + 890, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Method", + "text_level": 1, + "bbox": [ + 500, + 766, + 629, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section presents the details about CFINet. We start with a discussion about the inherent limitations of Cascade RPN when confronting small objects, then our coarse-to-fine high-quality proposal generation pipeline tailored for size-limited instances is introduced. Afterwards, we elucidate the architecture of newly designed Feature Imitation branch, also with the optimization and training procedure.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6319", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall architecture of CFINet is shown in Figure 2.", + "bbox": [ + 76, + 90, + 446, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Towards Better Proposals", + "text_level": 1, + "bbox": [ + 76, + 116, + 310, + 132 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Limitations of Cascade RPN. High-quality proposals play a pivotal role in two-stage detectors, but need heuristic anchor settings. Casdade RPN [32] discards this conventional setup by placing one single anchor on each feature point and conducting multi-stage refinement. Though exhibiting superior performance on objects of general scales, Cascade RPN fails to tackle extremely small objects well due to its inherent limitations. Concretely, the distance metric used in first-stage regression cannot guarantee sufficient potential anchors for small objects who have significantly small center region. Moreover, Cascade RPN only marks eligible anchors on a single pyramid level as positive, while this heuristic scheme simply discards those possible anchors at other levels which can still convey the existence and rough location information of small objects [42].", + "bbox": [ + 75, + 140, + 468, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Coarse-to-fine RPN. To remedy the aforementioned issues of Cascade RPN when handling small instances, we propose Coarse-to-fine RPN and the detailed structure is in Figure 2. First, we design an area-based anchor selection strategy to enable the instances of various sizes could have (relatively) adequate potential anchors. Concretely, for an object box with the width $w$ and the height $h$ , any anchors who have an IoU larger than $T_{\\mathrm{a}}$ will be regarded as positive for the coarse regression, and $T_{\\mathrm{a}}$ is formulated as follow:", + "bbox": [ + 75, + 367, + 468, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {\\mathrm {a}} = \\max (0. 2 5, 0. 2 0 + \\gamma \\cdot \\log \\frac {\\sqrt {w \\cdot h}}{1 2}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 513, + 468, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma$ denotes a scale factor and is set to default 0.15 in our experiments, and the term 12 actually corresponds to the minimal area definition of SODA dataset [9], which enables adequate samples for extreme-size objects and can be tuned for different datasets. Moreover, $\\gamma$ and max operation keep the optimization from being overwhelmed by the low-quality priors. Taking IoU as the criterion to mine potential anchors, the optimization inconsistency in multi-stage regression of Cascade RPN can be averted. Meanwhile, the model determines the positive sample in a more smooth way on top of the proposed continuous threshold.", + "bbox": [ + 75, + 551, + 468, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Distinct from Cascade RPN, we preserve anchors of all Feature Pyramid Network (FPN) [25] levels $\\{P_2,P_3,P_4,P_5\\}$ to perform first-stage regression. In this way, we could mine sufficient potential anchors for extremely small instances and meanwhile, larger instances still can obtain proper attention since the anchors matched to them have naturally higher IoUs, as discussed in Figure 1. After the first-stage regression, we then capture the offsets inside the regressed boxes and input them with the feature maps to RPN, in which the Adaptive Convolution [32] will be exploited to align the features and conduct second-stage regression and foreground-background classification.", + "bbox": [ + 75, + 719, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Loss Function. The training objective of our CRPN is:", + "bbox": [ + 500, + 90, + 864, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C R P N}} = \\alpha_ {1} \\left(\\mathcal {L} _ {\\mathrm {r e g}} ^ {\\mathrm {c}} + \\mathcal {L} _ {\\mathrm {r e g}} ^ {\\mathrm {f}}\\right) + \\alpha_ {2} \\mathcal {L} _ {\\mathrm {c l s}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 119, + 890, + 138 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we use cross-entropy loss and IoU Loss [44] as $\\mathcal{L}_{\\mathrm{cls}}$ and $\\mathcal{L}_{\\mathrm{reg}}$ , respectively. The c and f in Eq. (2) indicate the coarse-stage and fine-stage in our CRPN, and noting that we only do classification in the latter stage. The loss weights $\\alpha_{1}$ and $\\alpha_{2}$ are set to 9.0 and 0.9, respectively.", + "bbox": [ + 496, + 143, + 890, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Feature Imitation for Small Object Detection", + "text_level": 1, + "bbox": [ + 498, + 228, + 879, + 243 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Efforts on exploiting the intrinsic correlations between objects of different scales to boost the representations of small objects have been made, but most of them fail to the effectiveness and diversity. Specifically, the majority of previous methods [22, 2, 1, 28] resort to GAN to super-resolve the representations of small instances. This calls for the sophisticated training schemes and is prone to fabricate fake textures and artifacts [11]. Another line of efforts turn to the similarity learning which either has to construct offline feature bank in a cumbersome way [19], or directly leverages $\\mathcal{L}_2$ norm to similarity measurement between different RoI features [36], potentially leading the feature collapse issue: the region features after amending could have high-similarity but lost their own characteristics. This homogenization in feature space actually impairs the generality and robustness of the model.", + "bbox": [ + 496, + 250, + 890, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To mitigate the collapse risks and avoid the memory burden as well as enable the end-to-end optimization, we devise a Feature Imitation (FI) head (see Figure 2). Most importantly, instead of solely taking large-scale objects as the guidance of this procedure, we consider the model response in current state for each instance, thereby constructing a dynamic and currently optimized feature bank of proper exemplars in an online fashion. The FI branch mainly composes an Exemplar Feature Set and a Feature-to-Embedding (Feat2Embed) module, where the former reserves the RoI features of high-quality exemplars and the latter projects the input to the embedding space. Next we elucidate the details about our Feature Imitation branch.", + "bbox": [ + 496, + 493, + 890, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "What is a proper exemplar? As we discussed above, an exemplar is vital in the imitation learning. To determine the most representative/proper/high-quality examples which can deliver authentic guidance/supervision for small objects confusing the model at this moment, we first introduce a simple quality indicator for an instance. Given a ground-truth (GT) object $g = (c^{*}, b^{*})$ , where $c^{*}$ and $b^{*}$ denote its label and bounding box coordinates. Assuming the detection head outputs a prediction set $S = \\{C_{i}, IoU_{i}\\}_{i=1,2,\\dots,M}$ for $g$ , in which $C_{i} \\in \\mathbb{R}^{N+1}$ indicates the predicted classification vector and $IoU_{i}$ stands for the IoU of predicted box to GT, and $N$ is the number of foreground classes. Then we can obtain the potential high-quality set $S' = \\{(C_{j}, IoU_{i}) | \\arg \\max C_{j} = 0$", + "bbox": [ + 496, + 688, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6320", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Training of Feature Imitation branch.", + "text_level": 1, + "bbox": [ + 80, + 90, + 411, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input:", + "text_level": 1, + "bbox": [ + 83, + 108, + 130, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The set of GT boxes $\\mathcal{G} = \\{c_i^*, b_i^*\\}_{i=1,2,\\dots,T}$ and corresponding RoI features $\\{\\pmb{x}_i^g\\}_{i=1,2,\\dots,T}$ in current batch;", + "bbox": [ + 104, + 123, + 467, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The set of exemplar features $\\mathcal{E} = \\{\\mathcal{E}_i\\}_{c = 1,2,\\dots,N}$", + "bbox": [ + 106, + 152, + 429, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The set of background RoI features in current batch $\\mathcal{X}_{\\mathrm{bg}}$", + "bbox": [ + 106, + 167, + 468, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The threshold of high-quality $T_{\\mathrm{hq}}$", + "bbox": [ + 106, + 198, + 334, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The number of pos/neg samples $N_{\\mathrm{pos}}$ and $N_{\\mathrm{neg}}$ ;", + "bbox": [ + 106, + 213, + 424, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The transformation function $\\Gamma$ ;", + "bbox": [ + 106, + 229, + 313, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output:", + "text_level": 1, + "bbox": [ + 83, + 243, + 142, + 257 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "The set of positive embeddings $\\mathcal{P}_{\\mathrm{pos}}$ and negative embeddings $\\mathcal{P}_{\\mathrm{neg}}$", + "1: Initialize the set of positive features $\\mathcal{X}_{\\mathrm{pos}}$ and negative features $\\mathcal{X}_{\\mathrm{neg}}$ with $\\varnothing$ ;", + "2: for $g$ in $\\mathcal{G}$ do", + "3: Compute the $IQ$ of current $g$ according to Eq. (3)", + "4: $\\mathcal{X}_{\\mathrm{neg}}^q \\gets$ sample $N_{\\mathrm{neg}}$ features from $\\mathcal{X}_{\\mathrm{bg}} \\cup \\mathcal{E} \\backslash \\mathcal{E}_{c^*}$", + "5: if $I\\tilde{Q} \\geq T_{\\mathrm{hq}}$ then", + "6: $\\mathcal{E}_{c^*}\\gets \\pmb {x}_i^g$", + "7: $\\mathcal{X}_{\\mathrm{pos}}^g\\gets \\Gamma (\\pmb{x}_i^g)$", + "8: else", + "9: $\\mathcal{X}_{\\mathrm{pos}}^q\\gets$ sample $N_{\\mathrm{pos}}$ features from $\\mathcal{E}_{c^{*}}$", + "10: end if", + "11: $\\mathcal{X}_{\\mathrm{pos}} = \\mathcal{X}_{\\mathrm{pos}}\\cup \\mathcal{X}_{\\mathrm{pos}}^{g},\\mathcal{X}_{\\mathrm{neg}} = \\mathcal{X}_{\\mathrm{neg}}\\cup \\mathcal{X}_{\\mathrm{neg}}^{g}$", + "12: Apply Eq. (4) to $\\mathcal{X}_{\\mathrm{pos}}$ and $\\mathcal{X}_{\\mathrm{neg}}^q$ to obtain $\\mathcal{P}_{\\mathrm{pos}}$ and $\\mathcal{P}_{\\mathrm{neg}}$", + "13: end for", + "14: return $\\mathcal{P}_{\\mathrm{pos}}$ and $\\mathcal{P}_{\\mathrm{neg}}$" + ], + "bbox": [ + 83, + 258, + 468, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\left.c^{*}\\right\\rbrace_{j = 1,2,\\dots,M^{\\prime}}$ where $M^{\\prime}\\leq M$ . Now the Instance Quality of an object $g$ is defined as:", + "bbox": [ + 75, + 561, + 468, + 593 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI Q = \\frac {1}{\\left| \\mathcal {S} ^ {\\prime} \\right|} \\sum_ {j = 1} ^ {M ^ {\\prime}} C _ {j, c ^ {*}} \\cdot I o U _ {j} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 604, + 468, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The $IQ$ of a GT serves as an indicator of the current model's detection capability, enabling us to capture the high-quality exemplars who have precise localization and high-confidence classification scores, and the instances confusing the model often fail to either of them. By setting appropriate threshold, we can select proper instances to build the teacher feature-set and perform the imitation process.", + "bbox": [ + 75, + 657, + 468, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feat2Embed Module. Instead of directly measuring the similarity between different RoI features [36], we first embed these features with the simple Feat2Embed module. The input of FI branch is the region feature $\\boldsymbol{x}_i \\in \\mathbb{R}^{H \\times W \\times C}$ obtained by the RoI-wise operation, e.g., RoI Align, which will be first processed by three consecutive $3 \\times 3$ convolutional layers (with no padding operation) to abstract compact representations. It is worth noting that we update parameters during the extraction of current regional features", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d04745ae166de47b0b2ba865072db57ac92a8597fd2d88c781162db9b3cef9e9.jpg", + "image_caption": [ + "Figure 3. Size distribution of instances in (a) SODA-D and (b) SODA-A, where the absolute size corresponds to the square root of the object area." + ], + "image_footnote": [], + "bbox": [ + 506, + 89, + 697, + 255 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/219a7179ae45b50039c02f5d28aab6d38c0eab04734d461cba9198129f1ea40d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 89, + 887, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and freeze parameters during the extraction of exemplar ones (see Feat2Embed module in Figure 2), resulting in improved stability in performance. Subsequently, the intermediate features will be mapped to the embedding space on top of a two-layer perceptron and the embedding layer with the dimension of 128, in which the dimension of hidden layers is set to 512. We have also investigated various design choices and structures for our Feat2Embed module, and detailed information can be found in the Supplementary Materials. Finally, the output of the Feature Imitation branch is defined as:", + "bbox": [ + 496, + 328, + 890, + 492 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {i} = \\Theta_ {\\mathrm {F I}} (\\boldsymbol {x} _ {i}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 511, + 890, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Theta_{\\mathrm{FI}}$ denotes the parameters of Feature Imitation branch to be optimized.", + "bbox": [ + 496, + 536, + 890, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss Function. The objective of our FI head is simple: calculating the similarity between the RoI feature of proposal and that of the stored high-quality instances in embedding space, thereby pulling the features of those instances that confuse the model close to the exemplar ones of belonging category, while pushing apart from that of other categories and backgrounds. To this end, we propose a loss function based on Supervised Contrastive Learning [18] which extends the contrastive learning setup and allows multiple positive samples for an anchor object by exploiting the accessible label information. The loss function tailored for our FI branch is as follows:", + "bbox": [ + 496, + 566, + 890, + 746 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {F I}} = \\frac {- 1}{| \\mathcal {P} _ {\\text {p o s}} |} \\sum_ {j} \\sum_ {v _ {\\mathrm {p}} \\in \\mathcal {P} _ {\\text {p o s}}} \\log \\frac {\\exp \\left(v _ {j} \\cdot v _ {\\mathrm {p}} / \\tau\\right)}{\\sum_ {v _ {i} \\in \\mathcal {P}} \\exp \\left(v _ {j} \\cdot v _ {i} / \\tau\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 772, + 890, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{P} = \\mathcal{P}_{\\mathrm{pos}} \\cup \\mathcal{P}_{\\mathrm{neg}}$ denotes the sample set, while $\\mathcal{P}_{\\mathrm{pos}}$ and $\\mathcal{P}_{\\mathrm{neg}}$ represent the positive and negative set respectively and they have the same cardinality ideally, and $v_{\\mathrm{p}}$ and $v_{\\mathrm{n}}$ are the positive and negative sample from $\\mathcal{P}_{\\mathrm{pos}}$ and $\\mathcal{P}_{\\mathrm{neg}}$ . Moreover, $j$ indexes the current proposal and $\\tau$ indicates the", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6321", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/40f893fba3037cd8707f76cc8b085b67afbf11018a88c277cc0a41d713113564.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPublicationScheduleAPAP50AP75APesSAPrsSAPgsSAPN
One-stage
RetianNet [26]ICCV'171×28.257.623.711.925.234.144.2
FCOS [31]ICCV'191×23.949.519.96.919.430.940.9
ATSS [46]CVPR'201×26.855.622.111.723.932.241.3
YOLOX [12]ArXiv'2170e26.753.423.013.625.130.930.4
DyHead [10]CVPR'211×27.556.123.212.424.433.041.9
Keypoint-based
CornerNet [21]ECCV'182×24.649.521.76.520.532.243.8
CenterNet [48]ArXiv'1970e21.548.815.65.116.229.642.4
RepPoints [43]ICCV'191×28.055.624.710.123.835.145.3
Query-based
Deformable-DETR [49]ICLR'2050e19.244.813.76.315.424.934.2
Sparse RCNN [30]CVPR'211×24.250.320.38.820.430.239.4
Two-stage
Baseline [29]NeurIPS'151×28.959.424.113.825.734.543.0
Cascade RPN [32]NeurIPS'191×29.156.525.912.525.535.444.7
RFLA [40]ECCV'221×29.760.225.213.226.935.444.6
CFINet (ours)-1×30.760.826.714.727.836.444.6
", + "bbox": [ + 140, + 87, + 826, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison with state-of-the-art detection approaches on the SODA-D test-set, where 'Baseline' refers to Faster RCNN [29], serving as the baseline for the two-stage methods in the table. All the methods are trained on a ResNet-50 [17], except YOLOX (CSP-Darknet) [12] and CornerNet (HourglassNet-104) [21]. 'Schedule' denotes the number of epochs for training, in which $1 \\times$ corresponds to 12 epochs and $50\\mathrm{e}$ indicates 50 epochs.", + "bbox": [ + 75, + 364, + 890, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "temperature which plays a crucial part in contrastive learning and needs to be well designed, and we conduct ablation studies (see Table 9) to determine the optimal setting in our framework. The total loss function is presented:", + "bbox": [ + 75, + 444, + 468, + 506 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C R P N}} + \\mathcal {L} _ {\\mathrm {c l s}} + \\mathcal {L} _ {\\mathrm {r e g}} + \\alpha_ {3} \\mathcal {L} _ {\\mathrm {F I}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 531, + 468, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in which $\\mathcal{L}_{\\mathrm{cls}}$ and $\\mathcal{L}_{\\mathrm{reg}}$ are the original losses of detection head, and the $\\alpha_{3}$ is utilized to scale the weight of Feature Imitation part. With contrastive learning setups, not only can we fulfill the imitation learning but also prevent the collapse issue, thereby boosting the representations of small instances effectively. Moreover, the imitation process is only installed in training phase and will not slow the pace of inference.", + "bbox": [ + 75, + 565, + 468, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. Next we elucidate the training details of FI branch. The exemplar set $\\mathcal{E} = \\{\\mathcal{E}_i\\}_{c=1,2,\\dots,N}$ containing high-quality features of $N$ foreground categories and $\\mathcal{E}_i = \\{\\pmb{x}_{i,j}\\}_{j=1,2,\\dots,N_i}$ corresponding to the exemplar features of the $i$ -th class, and $N_i$ represents its size. We use $T_{\\mathrm{hq}}$ to pick out those high-quality instances which are compatible to be a good exemplar, and in practice, we set a bound value to the number of high-quality predictions of an instance to filter the effect of fluctuation of the network. The function $\\Gamma$ is used to augment features for high-quality instances, i.e., the positive features for a high-quality instance are the transformations of itself. The overall training procedure of FI branch is shown in Alg. 1 and more details please refer to the Supplementary Materials.", + "bbox": [ + 75, + 688, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 443, + 632, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 500, + 468, + 594, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our method, we perform extensive experiments on the recently released large-scale benchmark tailored for small object detection: SODA [9], including SODA-D and SODA-A.", + "bbox": [ + 496, + 492, + 890, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SODA-D. Focusing on the driving scenario, SODA-D comprises 24828 high-quality images and 278433 instances distributed across nine categories: people, rider, bicycle, motor, vehicle, traffic-sign, traffic-light, traffic-camera, and warning-cone. One of the most distinctive strengths of SODA-D is its diversity in terms of period, geographical locations, weather conditions, camera viewpoints, etc.", + "bbox": [ + 496, + 551, + 890, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SODA-A. SODA-A contains 872069 objects with oriented box annotations in 2513 aerial images and encompassing nine classes: airplane, helicopter, small-vehicle, large-vehicle, ship, container, storage-tank, swimming-pool, and windmill. The instances in SODA-A can appear in arbitrary orientations and are with significant density variations. To be specific, the average number of instances per image in SODA-A is about 350.", + "bbox": [ + 496, + 657, + 890, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As a specialized benchmark for small object detection, the instances in SODA are tiny, with most of them having an average size ranging from 10 to 30 pixels (see Figure 3). In contrast to conventional datasets for object detection, SODA includes extensive ignore annotations, aimed at filtering out instances that are either too large or are challenging to be identified deterministically due to heavy occlusion or lens flare. This procedure helps the model focus on valu", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6322", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9e46b7b661e6bb368fdfa73572c4426fc7b1c96fd8160263e85eec698d3bcad0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPublicationScheduleAPAP50AP75APesSAPrsSAPgsSAPN
One-stage
Rotated RetinaNet [26]ICCV'171×26.863.416.29.122.035.428.2
S2A-Net [15]TGRS'221×28.369.613.110.222.835.829.5
Oriented RepPoints [23]CVPR'221×26.358.819.09.422.632.428.5
DHRec [27]TPAMI'221×30.168.819.810.624.640.334.6
Two-stage
Baseline [29]NeurIPS'151×32.570.124.311.927.342.234.4
Gliding Vertex [41]TPAMI'211×31.770.822.611.727.041.133.8
Oriented RCNN [39]ICCV'211×34.470.728.612.528.644.536.7
DODet [8]TGRS'221×31.668.123.411.326.341.033.5
CFINet (ours)-1×34.473.126.113.529.344.035.9
", + "bbox": [ + 140, + 87, + 826, + 262 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06c1161da153eb4f95a2e9c739c16ec16b6d5504a7b01e07ec71e047ce779adc.jpg", + "image_caption": [ + "(a) Faster RCNN", + "Figure 4. The average number of high-quality proposals generated by (a) RPN, (b) Cascade RPN and (c) CRPN for instances in extremely Small $(eS)$ , relatively Small $(rS)$ , generally Small $(gS)$ , and Normal $(N)$ subsets, respectively. Noting that a proposal who has an IoU larger than 0.5 to any ground-truth boxes will be registered as a high-quality proposal." + ], + "image_footnote": [], + "bbox": [ + 140, + 321, + 364, + 472 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/21b07048a8cfdb5503efcd5a8738b9797431028ee9a8637befbe54e87cc7df8f.jpg", + "image_caption": [ + "(b) Cascade RPN" + ], + "image_footnote": [], + "bbox": [ + 388, + 321, + 596, + 470 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0ce2af594a95f3116ebfb7e7537487d53d9f1699b1b9c8a3158f6b09ec7ba564.jpg", + "image_caption": [ + "(c) CRPN" + ], + "image_footnote": [], + "bbox": [ + 620, + 321, + 828, + 472 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2d36f626c13a93bc637e10645430291a47152aa3a3e621978425760cee9f6b60.jpg", + "table_caption": [ + "Table 3. Comparison with state-of-the-art detection approaches on the SODA-A test-set, where 'Baseline' refers to Rotated Faster RCNN [29], serving as the baseline for the two-stage methods in the table. Other settings are consistent with Table 2." + ], + "table_footnote": [], + "table_body": "
Proposal MethodAR\\( AR_{eS} \\)\\( AR_{rS} \\)\\( AR_{gS} \\)\\( AR_N \\)
RPN [29]41.224.038.347.357.1
RPN-0.541.324.238.547.354.1
GA-RPN [33]42.124.139.248.956.2
Cascade RPN [32]41.822.838.248.757.1
CRPN42.624.638.949.156.9
", + "bbox": [ + 78, + 542, + 483, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "able small instances. The objects in SODA are divided into Small and Normal according to their areas, in which Small is further split into three subsets: extremely Small (eS), relatively Small (rS) and generally Small (gS). The evaluation metric of SODA follows that of COCO [24], namely averaging the precision over 10 IoU thresholds ranging from 0.5 to 0.95 (with an interval of 0.05), specifically focusing on Small objects.", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/83329ae4a9bd5d28c0b3b5b83ff406be8ca1ee2d87c553b22f16f15ac03d57d5.jpg", + "table_caption": [ + "Table 4. Average Recall $(AR)$ performances of our CRPN and its counterparts on the SODA-D test-set. All the methods are with Faster RCNN (ResNet-50) as the baseline and trained for a $1\\times$ schedule, in which RPN denotes the vanilla version of Faster RCNN whose positive threshold in RPN stage is set 0.7, and RPN-0.5 represents the version with 0.5 as its positive threshold of RPN. The results are tested with 300 proposals per image." + ], + "table_footnote": [], + "table_body": "
BaselineCRPNFIAPAPeSAPrSAPgS
✓28.913.825.734.5
✓✓30.314.327.336.1
✓✓29.514.426.335.1
✓✓✓30.714.727.836.4
", + "bbox": [ + 506, + 542, + 883, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation analysis of our method, in which 'Baseline' denotes the vanilla Faster RCNN, CRPN and FI indicate Coarse-to-fine RPN and Feature Imitation branch, respectively.", + "bbox": [ + 498, + 623, + 890, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 680, + 718, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the following experiments, unless specified, we use train-set to conduct the training and leave test-set to performance comparisons and ablation studies. Considering that the images in SODA enjoy a very high resolution $(\\sim 4000\\times 3000)$ , we first split the original images into a series of $800\\times 800$ patches with a stride of 650, and similar to [9] these patches will be resized to $1200\\times 1200$ during training and testing. All the experiments in this paper are conducted on a single RTX 3090 with the batch size of 4. Only random flip involved in data augmentation. We train all the models with a $1\\times$ schedule (a bunch of 12 epochs), and the learning rate is set to 0.01 which decays after epoch 8 and epoch 11 by 0.1. The default optimizer is SGD with", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6323", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b6b25e1b8bbafe0b141b51706c15cbe6e0adab44b45f059547c08ca9b2d1c2da.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
StrategyAPAPeSAPrSAPgS
0.2029.913.726.835.9
0.4030.114.126.936.2
Ours30.314.327.336.1
", + "bbox": [ + 138, + 88, + 408, + 147 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f25e08745268b1c9ee14c8fb808060436693097e4e21d6c609aebcd91ec63eee.jpg", + "table_caption": [ + "Table 6. Different definitions about positive anchor for CRPN, in which 'Ours' denotes the proposed dynamic strategy." + ], + "table_footnote": [], + "table_body": "
α3APAPeSAPrSAPgS
0.2530.614.727.636.2
0.5030.714.727.836.4
0.7530.414.027.536.2
", + "bbox": [ + 138, + 191, + 408, + 256 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the momentum of 0.9 and the weight decay of 0.0001. We use ResNet-50 [17] with FPN [25] for all models.", + "bbox": [ + 76, + 287, + 468, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 76, + 329, + 215, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To exhibit the effectiveness of our method, we conduct a thorough comparison with current representative approaches on the SODA-D and SODA-A.", + "bbox": [ + 76, + 352, + 468, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2 represents the results of our method and several mainstream approaches on the SODA-D test-set. Integrating with Faster RCNN [29], our CFINet achieves state-of-the-art performance with an overall $AP$ of $30.7\\%$ , and outperforms the baseline model with $1.8\\%$ points. When delving into specific metrics, our method exhibits clear predominance, particularly on the most challenging metrics $AP_{es}$ and $AP_{rs}$ . Moreover, CFINet exceeds the tailored small object detection method RFLA [40] by a significant margin (1.0% on $AP$ , 1.5% on $AP_{es}$ , and 0.9% on $AP_{rs}$ ) when both taking Faster RCNN as the baseline. Actually, RFLA sacrifices the performance on extremely Small instances (with a decrease of $0.6\\%$ points when compared to Faster RCNN).", + "bbox": [ + 75, + 398, + 468, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On the SODA-A test-set, CFINet also achieves the best result and shows great advantage in comparison to other solutions especially on $AP_{eS}$ (see Table 3), indicating its superiority and generality. Furthermore, albeit exhibiting advantage on $AP_{75}$ metric and larger instances, Oriented RCNN [39] lags largely behind our approach on $AP_{50}$ (73.1% vs. 70.7%) and $AP_{eS}$ (13.5% vs. 12.5%).", + "bbox": [ + 75, + 609, + 468, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Effectiveness of CRPN", + "text_level": 1, + "bbox": [ + 76, + 724, + 285, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "One of the main designs in this paper is the Coarse-to-fine RPN, which is based on the observation that current fixed overlaps-based sampling paradigm is inappropriate for small instances due to the inherent contradictions, while the refined designs of RPN could partially reduce this barrier but still fail to satisfactory results. Here, we conduct thorough analyses to demonstrate the capability of our CRPN to generate high-quality proposals for size-limited instances.", + "bbox": [ + 75, + 750, + 468, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We first exhibit the recall performances of our CRPN and its counterparts in Table 4, from which we can see that low", + "bbox": [ + 76, + 869, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/21b84337df99ab436ef25bcada39158b0d0a98851de0f9a2fd972e62ee259e62.jpg", + "table_caption": [ + "Table 7. The effect of loss weight for Feature Imitation branch." + ], + "table_footnote": [], + "table_body": "
ThqAPAPeSAPrSAPgS
0.5030.314.027.336.0
0.5530.614.327.336.5
0.6530.714.727.836.4
0.7030.514.627.336.3
", + "bbox": [ + 571, + 88, + 821, + 161 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f0d737675412a9fd982fd21eb482587139c7438dc1bf0cbed7539b5108b8ebdd.jpg", + "table_caption": [ + "Table 8. The investigation of the criterion to be an exemplar instance." + ], + "table_footnote": [], + "table_body": "
τAPAPeSAPrSAPgS
0.1030.314.127.336.1
0.5030.414.427.236.2
0.6030.714.727.836.4
0.8030.214.027.336.0
", + "bbox": [ + 570, + 204, + 821, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. The choices of temperature $\\tau$ to the final performance.", + "bbox": [ + 500, + 280, + 872, + 294 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ering the positive threshold slightly improves the average recall while sacrificing the performance of larger instances $(AR_{N}$ experiences a sharp decline from $57.1\\%$ to $54.1\\%$ ). GA-RPN [33] as well as Cascade RPN [32] both fail to better results since their patterns incline to large instances as we discussed before. In comparison with RPN and its variants, our CRPN demonstrates superior performance on objects in the Small, while achieving comparable results on Normal instances. This validates our assumption that refined proposal networks tend to favor larger objects.", + "bbox": [ + 496, + 321, + 890, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conjecture that one of the most challenging issues towards accurate small object detection is the scarcity of high-quality samples, which is also the major motivation behind the design of CRPN. Hence, we intuitively compare the baseline RPN, Cascade RPN and our CRPN about the number of high-quality samples. In Figure 4, our CRPN generates more high-quality proposals compared to the other competitors. Interestingly, CRPN can dynamically shift the focus along the training: at the beginning, the model concentrates more on large objects which are conducive for early optimization while as the training goes, the model gradually shifts its attention to objects with small sizes that are usually not handled well before. This is interpretable since the instances having extremely limited sizes are with more uncertainties and fitting them in early-phase is not an optimal choice for the detector.", + "bbox": [ + 496, + 472, + 892, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation and Discussion", + "text_level": 1, + "bbox": [ + 500, + 724, + 720, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this part, we conduct ablation studies as well as comprehensive discussions to attest the importance of the CRPN and FI branch, and moreover, determine the appropriate settings of our approach. All the experiments of this section are conducted on the SODA-D test-set.", + "bbox": [ + 496, + 750, + 890, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Investigation of Designed Components. We first perform ablation experiments to verify the effectiveness of two modules. As in Table 5, our CRPN and FI can both improve the performances steadily, while the introduction of feature imitation branch is conducive for the recognition of size-", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6324", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e5d4b8aa12b8a76b267c49ca40f941f184c53f1520ec2603aae038f11f9123fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 89, + 493, + 309 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f768988861a961f0b9e6c85feb36fda818ebc3bee8c8b0b8db68e1cdb90500cd.jpg", + "image_caption": [ + "people" + ], + "image_footnote": [], + "bbox": [ + 122, + 314, + 147, + 325 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/341b2c87ca4f9ad8262dd4f74f21cdf09ea19eb8bc6b0f232d789b2cd91bfdba.jpg", + "image_caption": [ + "rider", + "Figure 5. Qualitative results of our method on the SODA-D test-set. Only predictions with confidence scores larger than 0.3 are demonstrated and the masked bounding boxes represent ignore regions. Best viewed in color and zoom-in windows." + ], + "image_footnote": [], + "bbox": [ + 194, + 314, + 214, + 325 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/cf692c67adc8cde0f975c194db3521c7837419b8592302fb8aae9ddde520cbff.jpg", + "image_caption": [ + "bicycle" + ], + "image_footnote": [], + "bbox": [ + 259, + 314, + 281, + 325 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/842dc621bb1adaf4d03d7025959f6314ade178f76812e41710e44c553c5d894f.jpg", + "image_caption": [ + "motor" + ], + "image_footnote": [], + "bbox": [ + 334, + 314, + 357, + 325 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/32ab6006046aa2afdf1851f67ff4746a3a31ce59b16cdb41bd73b2ce4a949798.jpg", + "image_caption": [ + "vehicle" + ], + "image_footnote": [], + "bbox": [ + 406, + 315, + 429, + 325 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/afeb7780a1a48b43ac48917ec4dce1e80a3f1b3e5b522e08444792753f8130e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 90, + 910, + 311 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/216e7c7af19ce4309f803c82af3de7c8297cef2022bc2fb70f19bc73adf83f76.jpg", + "image_caption": [ + "c-sign" + ], + "image_footnote": [], + "bbox": [ + 498, + 315, + 522, + 325 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4c16ae33eac43f790dfeb505957673672182147820f96898ac3c99bcb458eb3e.jpg", + "image_caption": [ + "traffic-light" + ], + "image_footnote": [], + "bbox": [ + 571, + 315, + 594, + 325 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/90c053259362cf6f50d580ac9723840ce8543c177c5a7171e5cc6c6547c5520c.jpg", + "image_caption": [ + "traffic-camera" + ], + "image_footnote": [], + "bbox": [ + 665, + 315, + 687, + 325 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e8212f3ddfaa25b94322f9cf2cae84358d98ee9c57b1f692ae14be81d7e658ec.jpg", + "image_caption": [ + "ing-cone" + ], + "image_footnote": [], + "bbox": [ + 790, + 315, + 813, + 325 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "limited instances. Consistently, the integration of FI and CRPN achieves the best result, as CRPN is capable of generating sufficient high-quality proposals (as shown in Figure 4) thereby offering a more accurate indication of instance quality and the potential to act as an exemplar.", + "bbox": [ + 75, + 382, + 468, + 457 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fixed or Dynamic. A natural idea is directly setting fixed positive threshold to obtain more anchors for one-stage regression of CRPN. Here we show that our simple-yet-effective area-based anchor mining strategy can achieve the best performance. In Table 6, when the IoU threshold of a positive potential anchor for first-stage regression drops to 0.20, the $AP_{eS}$ is only $13.7\\%$ and this could be attributed to low-quality samples and the predominance of large instances as discussed before. The proposed area-based anchor mining strategy could mitigate this problem and obtain the best overall accuracy.", + "bbox": [ + 75, + 459, + 468, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Weights of Feature Imitation Loss. In this section, we analyze the impact of the weight parameter of FI branch (namely the hyper-parameter $\\alpha_{3}$ ) on the model. As shown in Table 7, paying too little or excessive attention both deteriorates the final performance, hence we set $\\alpha_{3}$ to 0.5 in our experiments to ensure overall accuracy.", + "bbox": [ + 75, + 627, + 468, + 717 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The Criterion of Being An Exemplar. The quality of the exemplar plays a pivotal role in the imitation process [19, 36]. Next we discuss the choices for capturing a high-quality exemplar to build the feature set. In Table 8, the lower $T_{\\mathrm{hq}}$ involves more exemplars and updates the teacher set more frequently while the higher $T_{\\mathrm{hq}}$ does exactly the opposite. It can be seen that increasing the $T_{\\mathrm{hq}}$ from 0.5 to 0.65 could facilitate the imitation process and when the $T_{\\mathrm{hq}}$ reaches 0.70, the overall $AP$ drops instead. This may originates that the earliest samples stored in the feature set are inappropriate for current state, because the optimization is dynamic and the model is evolving hence the criterion of", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "being an exemplar has already changed.", + "bbox": [ + 498, + 382, + 764, + 397 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The Temperature. The temperature is paramount for contrastive learning [18] and we conduct a series of experiments to verify the best choice for $\\tau$ . From Table 9, when $\\tau$ ranges from 0.10 to 0.80, the overall performance increases first and then decreases to $30.2\\%$ , therefore we choose 0.6 for our method.", + "bbox": [ + 498, + 398, + 890, + 487 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Visualization. We demonstrate the visualization results of example images from SODA-D test-set in Figure 5 to intuitively show the capability of our detector when detecting the small instances.", + "bbox": [ + 498, + 488, + 890, + 547 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 563, + 617, + 578 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we proposed CFINet, a two-stage detector based on the Coarse-to-fine Region Proposal Network and Feature Imitation setups, in which the former can produce sufficient high-quality proposals for small instances particularly for those with extremely limited sizes. Then the novel detection head on top of the feature imitation branch facilitates the representations of small objects posing challenges to the model under the contrastive learning paradigm. The experiments results show that our method achieves state-of-the-art performance on the large-scale small object detection datasets SODA-D and SODA-A. In the future, a more flexible and general indicator of instance quality is worth investigating.", + "bbox": [ + 496, + 589, + 890, + 785 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 799, + 658, + 816 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Science Foundation of China under Grant 62136007 and Grant U20B2068, and in part by the Natural Science Basic Research Program of Shaanxi under Grants 2021JC-16 and 2023-JC-ZD-36.", + "bbox": [ + 498, + 825, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6325", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. Finding tiny faces in the wild with generative adversarial network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 21-30, 2018.", + "[2] Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. Sod-mtgan: Small object detection via multi-task generative adversarial network. In Proceedings of the European Conference on Computer Vision, pages 206-221, 2018.", + "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European Conference on Computer Vision, pages 213-229, 2020.", + "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Cconference on Machine Learning, pages 1597-1607, 2020.", + "[5] Gong Cheng, Junwei Han, Peicheng Zhou, and Dong Xu. Learning rotation-invariant and fisher discriminative convolutional neural networks for object detection. IEEE Transactions on Image Processing, 28(1):265-278, 2019.", + "[6] Gong Cheng, Qingyang Li, Guangxing Wang, Xingxing Xie, Lingtong Min, and Junwei Han. Sfrnet: Fine-grained oriented object recognition via separate feature refinement. IEEE Transactions on Geoscience and Remote Sensing, 61:1-10, 2023.", + "[7] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022.", + "[8] Gong Cheng, Yanqing Yao, Shengyang Li, Ke Li, Xingxing Xie, Jiabao Wang, Xiwen Yao, and Junwei Han. Dual-aligned oriented detector. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022.", + "[9] Gong Cheng, Xiang Yuan, Xiwen Yao, Kebing Yan, Qinghua Zeng, Xingxing Xie, and Junwei Han. Towards large-scale small object detection: Survey and benchmarks. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-20, 2023.", + "[10] Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unifying object detection heads with attentions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7373-7382, 2021.", + "[11] Chunfang Deng, Mengmeng Wang, Liang Liu, Yong Liu, and Yunliang Jiang. Extended feature pyramid network for small object detection. IEEE Transactions on Multimedia, 24:1968-1979, 2021.", + "[12] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021.", + "[13] Spyros Gidaris and Nikos Komodakis. Attend refine repeat: Active box proposal generation via in-out localization. arXiv preprint arXiv:1606.04446, 2016." + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020.", + "[15] Jiaming Han, Jian Ding, Jie Li, and Gui-Song Xia. Align deep features for oriented object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2021.", + "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020.", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016.", + "[18] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020.", + "[19] Jung Uk Kim, Sungjune Park, and Yong Man Ro. Robust small-scale pedestrian detection with cued recall via memory learning. In Proceedings of the IEEE International Conference on Computer Vision, pages 3050-3059, 2021.", + "[20] Chunbo Lang, Gong Cheng, Binfei Tu, Chao Li, and Junwei Han. Base and meta: A new perspective on few-shot segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2023.", + "[21] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In ECCV, pages 734-750, 2018.", + "[22] Jianan Li, Xiaodan Liang, Yunchao Wei, Tingfa Xu, Jiashi Feng, and Shuicheng Yan. Perceptual generative adversarial networks for small object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1222-1230, 2017.", + "[23] Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented reploints for aerial object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1829-1838, 2022.", + "[24] Tsung Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, pages 740-755, 2014.", + "[25] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2117-2125, 2017.", + "[26] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. pages 2980-2988, 2017.", + "[27] Guangtao Nie and Hua Huang. Multi-oriented object detection in aerial images with double horizontal rectangles. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4932-4944, 2022." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6326", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Junhyug Noh, Wonho Bae, Wonhee Lee, Jinhwan Seo, and Gunhee Kim. Better to follow, follow to be better: Towards precise supervision of feature super-resolution for small object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 9725-9734, 2019.", + "[29] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28, 2015.", + "[30] Peize Sun et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 14449-14458, 2021.", + "[31] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. pages 9627-9636, 2019.", + "[32] Thang Vu, Hyunjun Jang, Trung X Pham, and Chang Yoo. Cascade rpn: Delving into high-quality region proposal network with adaptive convolution. Advances in Neural Information Processing Systems, 32, 2019.", + "[33] Jiaqi Wang, Kai Chen, Shuo Yang, Chen Change Loy, and Dahua Lin. Region proposal by guided anchoring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2965-2974, 2019.", + "[34] Jinwang Wang, Wen Yang, Haowen Guo, Ruixiang Zhang, and Gui-Song Xia. Tiny object detection in aerial images. In IEEE International Conference on Pattern Recognition, pages 3791-3798, 2021.", + "[35] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 7303-7313, 2021.", + "[36] Jialian Wu, Chunluan Zhou, Qian Zhang, Ming Yang, and Junsong Yuan. Self-mimic learning for small-scale pedestrian detection. In Proceedings of the ACM International Conference on Multimedia, pages 2012-2020, 2020.", + "[37] Wei Wu, Hao Chang, Yonghua Zheng, Zhu Li, Zhiwen Chen, and Ziheng Zhang. Contrastive learning-based robust object detection under smoky conditions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4295-4302, 2022.", + "[38] Enze Xie, Jian Ding, Wenhai Wang, Xiaohang Zhan, Hang Xu, Peize Sun, Zhenguo Li, and Ping Luo. Detco: Unsupervised contrastive learning for object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 8392-8401, 2021.", + "[39] Xingxing Xie, Gong Cheng, Jiabao Wang, Xiwen Yao, and Junwei Han. Oriented r-cnn for object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 3520-3529, 2021.", + "[40] Chang Xu, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Rfla: Gaussian receptive based label assignment for tiny object detection. In Proceedings of the European Conference on Computer Vision, 2022.", + "[41] Yongchao Xu, Mingtao Fu, Qimeng Wang, Yukang Wang, Kai Chen, Gui-Song Xia, and Xiang Bai. Gliding vertex on" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the horizontal bounding box for multi-oriented object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(4):1452-1459, 2020.", + "[42] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. Querydet: Cascaded sparse query for accelerating high-resolution small object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 13668-13677, 2022.", + "[43] Ze Yang, Shaohui Liu, Han Hu, Liwei Wang, and Stephen Lin. Repoints: Point set representation for object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 9657-9666, 2019.", + "[44] Jiahui Yu, Yuning Jiang, Zhangyang Wang, Zhimin Cao, and Thomas Huang. Unitbox: An advanced object detection network. In Proceedings of the ACM International Conference on Multimedia, pages 516-520, 2016.", + "[45] Xuehui Yu, Yuqi Gong, Nan Jiang, Qixiang Ye, and Zhenjun Han. Scale match for tiny person detection. In IEEE Conference on Winter Conference on Applications of Computer Vision, pages 1257-1265, 2020.", + "[46] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9759-9768, 2020.", + "[47] Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, and Stan Z Li. S3fd: Single shot scale-invariant face detector. In Proceedings of the IEEE International Conference on Computer Vision, pages 192-201, 2017.", + "[48] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019.", + "[49] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2020." + ], + "bbox": [ + 501, + 92, + 890, + 585 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6327", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning/231f8e4e-8ec6-40cc-a455-514ae44e6fa1_model.json b/2023/Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning/231f8e4e-8ec6-40cc-a455-514ae44e6fa1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0c978c6eb6c8459f7ce2c73b758974631355389d --- /dev/null +++ b/2023/Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning/231f8e4e-8ec6-40cc-a455-514ae44e6fa1_model.json @@ -0,0 +1,2642 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.168, + 0.131, + 0.803, + 0.177 + ], + "angle": 0, + "content": "Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.204, + 0.834, + 0.258 + ], + "angle": 0, + "content": "Xiang Yuan Gong Cheng * Kebing Yan Qinghua Zeng Junwei Han School of Automation, Northwestern Polytechnical University, Xi'an, China {shaunyuan, kebingyan, zengqinghua}@mail.nwpu.edu.cn, {gcheng, jhan}@nwpu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.324, + 0.474, + 0.702 + ], + "angle": 0, + "content": "The past few years have witnessed the immense success of object detection, while current excellent detectors struggle on tackling size-limited instances. Concretely, the well-known challenge of low overlaps between the priors and object regions leads to a constrained sample pool for optimization, and the paucity of discriminative information further aggravates the recognition. To alleviate the aforementioned issues, we propose CFINet, a two-stage framework tailored for small object detection based on the Coarse-to-fine pipeline and Feature Imitation learning. Firstly, we introduce Coarse-to-fine RPN (CRPN) to ensure sufficient and high-quality proposals for small objects through the dynamic anchor selection strategy and cascade regression. Then, we equip the conventional detection head with a Feature Imitation (FI) branch to facilitate the region representations of size-limited instances that perplex the model in an imitation manner. Moreover, an auxiliary imitation loss following supervised contrastive learning paradigm is devised to optimize this branch. When integrated with Faster RCNN, CFINet achieves state-of-the-art performance on the large-scale small object detection benchmarks, SODA-D and SODA-A, underscoring its superiority over baseline detector and other mainstream detection approaches. Code is available at https://github.com/shaunyuan22/CFINet." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.716, + 0.21, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.741, + 0.47, + 0.833 + ], + "angle": 0, + "content": "Small object detection (SOD) aims to classify and localize the instances with limited regions, which plays an important role in a wide range of scenarios, such as pedestrian detection, autonomous driving, and intelligent surveillance understanding, to name a few [42, 28, 22, 19, 36, 2, 20]. Compared to the generic object detection which has been" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.291, + 0.641, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.291, + 0.768, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.292, + 0.892, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.445, + 0.895, + 0.544 + ], + "angle": 0, + "content": "Figure 1. Distribution of maximum IoU of anchors matched to each ground-truth instance in SODA-D [9] train-set, where extremely Small (eS), relatively Small (rS), and generally Small (gS) correspond to three area subsets in SODA [9] with the ranges (0, 144], (144, 400] and (400, 1024]. The smaller the objects are, the lower IoU the matched anchors have, hence the commonly used positive IoU threshold (0.7) is too rigorous for small objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.894, + 0.644 + ], + "angle": 0, + "content": "extensively studied, SOD task receives relatively little attention and good solutions are still scarce so far. Moreover, generic detectors [29, 6, 25, 31, 7, 3, 26, 5] usually struggle on handling small objects due to two inherent challenges: the insufficient and low-quality samples for training and the uncertain prediction of RoIs (Region of Interests)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.895, + 0.903 + ], + "angle": 0, + "content": "First, current prevailing detectors exploit either overlap-based [29, 26] or distance-based [31] strategies to select the positive priors of objects for training. However, small instances usually occupy an extremely limited area, therefore the region overlaps between densely arranged anchors and ground truth boxes are significantly small and far from the commonly used positive IoU (Intersection-over-Union) threshold, as in Figure 1. In other words, the existing positive sample criterion is overly stringent when applied to small/tiny objects, resulting in a restricted number of samples available for optimization. An intuitive approach involves reducing the threshold for defining a positive sample [47]. However, while this can lead to an increase in the number of positive samples, it often at the expense of overall sample quality, in which low-quality samples disrupt optimization and incur a trivial regression solution. Worse still, this is actually contradictory to the purpose of proposal" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.84, + 0.346, + 0.852 + ], + "angle": 0, + "content": "*Corresponding author: gcheng@nwpu.edu.cn" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.852, + 0.47, + 0.9 + ], + "angle": 0, + "content": "This paper focuses on the detection of \"pure\" small objects, where the scales of all the objects are distributed within a relatively tight range [45, 34, 9], which is distinct from the known small objects in multi-scale object detection [24]." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.84, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6317" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.152 + ], + "angle": 0, + "content": "
ctrl/ign ratioAPAP50AP75APesSAPrsSAPgSAPN
Baseline28.959.424.113.825.734.543.0
0.2/0.529.156.525.912.525.535.444.7
0.5/0.829.557.826.013.526.235.845.0
0.8/1.027.554.124.311.223.833.742.5
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.162, + 0.468, + 0.232 + ], + "angle": 0, + "content": "Table 1. The performances of Cascade RPN [32] compared to the baseline (a vanilla Faster RCNN [29]). The ctr/ign ratio denotes the sampling region in first regression stage of Cascade RPN. The results are tested on the SODA-D [9] test-set and with a ResNet-50 [17] as the backbone." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.245, + 0.468, + 0.276 + ], + "angle": 0, + "content": "network, i.e., guaranteeing the recall and ease the burden of subsequent work." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.279, + 0.468, + 0.64 + ], + "angle": 0, + "content": "To sum up, current prevailing priors-to-proposals paradigms that heavily depend on the overlap or distance metric have inherent limitations in detecting small objects, and nowadays devised assignment or sampling schemes contribute minimally to this problem [9, 40]. Since the proposals play such a crucial role in two-stage detectors, so how about the improved Region Proposal Network (RPN) variants [13, 32, 33] meet small object detection? Following this line, we take Cascade RPN [32], one of the most superior proposal network for generic object detection, to perform preliminary experiments and the results are shown in Table 1. While the auxiliary regression phase provides refined priors with better initialization for subsequent regression, the final results remain somewhat unsatisfactory. Specifically, the improvement mainly comes from the larger objects and the \\(AP_{eS}\\) as well as \\(AP_{rS}\\) actually decrease significantly instead, indicating that the region-based sampling strategy is inclined to large instances which further dominate the proposal network. Meanwhile, enlarging the sample region contributes little (even negative) to this condition (see the bottom row in Table 1). Therefore, the coarse-to-fine pipeline has the potential to surmount the barrier of conventional prior-to-proposal paradigm, but the crux lies in dedicating sufficient attention to small instances." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.645, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Second, small objects usually lack discriminative information and distorted structures, leading the inclination of model to give ambiguous even incorrect predictions [1, 11]. Meanwhile, there are a certain amount of large instances embodying clear visual cues and better discrimination. Building upon this observation, several works proposed to bridge the representation gap between small objects and large ones, and most of them [2, 22, 28, 1] rely on Generative Adversarial Network (GAN) [14] or similarity learning [19, 36] to super-resolve/restore the features of size-limited instances under the guidance of large ones that are deemed to be visually authentic. However, these approaches overlook the fact: high quality \\(\\neq\\) large size meanwhile small size \\(\\neq\\) low quality. In other words, the criterion for humans and the model to decide whether a sample is competent to be a good example is distinct. For the latter, it is dynamic and should be adjusted according to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "the current optimization of the detector. Moreover, efforts in this line have to resort to sophisticated training strategies or additional models, which is time-consuming and break the conventional end-to-end paradigm." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.892, + 0.407 + ], + "angle": 0, + "content": "Putting the above parts together, we propose a two-stage small object detector CFINet based on the coarse-to-fine pipeline and feature imitation learning. Concretely, enlightened by the multi-stage proposal generation scheme in Cascade RPN, we devise Coarse-to-fine RPN (CRPN). It firstly employs an dynamic anchor selection strategy to mine potential priors to conduct coarse regression, and henceforth, these refined anchors will be classified and regressed by the region proposal network. In addition, we extend the conventional classification-and-regression setting with an auxiliary Feature Imitation (FI) branch, which can leverage the regional features of high-quality instances to guide the learning of those objects with uncertain/mistaken predictions, and a loss function based on the Supervised Contrastive Learning (SCL) [18] is designed to optimize the whole process. The main contributions of this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.416, + 0.892, + 0.491 + ], + "angle": 0, + "content": "- A coarse-to-fine proposal generation pipeline named CRPN was built to perform anchor-to-proposal procedure, where an area-based anchor mining strategy and cascade regression empower the high-quality proposals for small instances." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.5, + 0.892, + 0.576 + ], + "angle": 0, + "content": "- An auxiliary Feature Imitation (FI) branch was introduced to enrich the representations of low-quality instances perplexing the model under the supervision of high-quality instances, and this novel branch is optimized by a tailored loss function based on SCL." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.584, + 0.892, + 0.63 + ], + "angle": 0, + "content": "- The experiment results on the SODA-D and SODA-A datasets exhibit the superiority of our CFINet to detect these instances with extremely limited sizes." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.416, + 0.892, + 0.63 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.65, + 0.665 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Anchor Refinement and Region Proposals. Two-stage anchor-based approaches heavily depend on the high-quality proposals [33, 32]. Towards this goal, RPN was first introduced in Faster RCNN [29] to produce proposals in a fully convolutional network, and this simple yet effective design facilitates end-to-end model optimization. Following RPN, [13] proposed to iteratively regress the predefined anchors. GA-RPN [33] discards the uniform anchoring strategy and formulates the anchor generation in two steps: first determining the locations which may contain objects and on which the anchor scales are predicted then. By installing a multi-stage anchor-to-proposal strategy and alleviating the misalignment between the refined anchors and image features, Cascade RPN [32] enables high-quality proposal generation. Unfortunately, current proposal-oriented" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6318" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.087, + 0.495, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.087, + 0.912, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.353, + 0.892, + 0.478 + ], + "angle": 0, + "content": "Figure 2. The overall architecture of CFINet. In Coarse-to-fine RPN (CRPN), the area-based Anchor Mining strategy ensures sufficient candidates for instances of various sizes (small: orange boxes, large: green boxes) based on the dynamic pos threshold, which will be then used to obtain the coarse proposals. After that, the alignment between coarse proposals and corresponding features is enabled by the Adaptive Convolution before feeding into the RPN to produce high-quality proposals (blue boxes). The Feature Imitation (FI) branch is devised to facilitate the representations of small instances, in which the RoI features of uncertain/mistaken predictions will be pulled to their counterparts of exemplar feature set in embedding space (throughout the Feat2Embed module), while pushed apart from the exemplar features of other categories and background. And the exemplar features are collected based on the model predictions using the proposed quality indicator, i.e., Instance Quality \\((IQ)\\). We tailor a Feature Imitation loss function \\(\\mathcal{L}_{\\mathrm{FI}}\\) to optimize this auxiliary branch. Note that we only exhibit single-level Feature Pyramid Network (FPN) [25] feature for clear illustration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.468, + 0.623 + ], + "angle": 0, + "content": "frameworks fail to produce high-quality proposals for instances with limited regions, and the root cause lies in the notoriously low overlaps between the objects and priors [40]. Different from the above methods, our coarse-to-fine proposal pipeline could exploit the potential of multi-stage refinement paradigm, thereby guaranteeing both the quantity and quality of proposals for instances with extremely limited sizes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Feature Imitation for Small Object Detection. One of the major challenges to detect small objects is the low-quality representation [45, 40, 9], while large instances often with clear structures and discriminative features. Hence, a series of efforts have been made to boost the semantic representations of small/tiny instances by mining the intrinsic correlations between small and large objects. Based on the generative adversarial paradigm, Perceptual GAN [22] designs a generator that is optimized to produce high-quality representations of small instances to fool the subsequent discriminator. Bai et al. [1] devised a novel pipeline to restore a clear face from the inputting blurry one. Noh et al. [28] further introduced precise supervision for the super-resolution process of small objects. Moreover, Wu et al. [36] and Kim et al. [19] both exploited similarity learning to force the features of small-scale pedestrians close to that of the large-scale ones which are obtained by an additional model. The existence of super-resolution branch or offline" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.892, + 0.563 + ], + "angle": 0, + "content": "feature bank hampers the end-to-end optimization while our method updates the exemplar features in an online fashion, which guarantees the diversity of high-quality feature sets thereby getting rid of the collapse issue." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.892, + 0.747 + ], + "angle": 0, + "content": "Contrastive Learning for Object Detection. The recent explosion of self-supervised learning mainly comes from its Contrastive Learning fashion, and several works have extended this paradigm into detection fields. Detco [38] is an effective self-supervised framework for object detection which utilizes the image and its local patches to conduct contrastive learning. Wu et al. [37] applied contrastive learning to object detection under smoky conditions. Though contrastive learning has recently received considerable interests [4, 16, 35], the potential of utilizing contrastive learning for better representation of small objects has not yet been investigated to date." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.767, + 0.63, + 0.783 + ], + "angle": 0, + "content": "3. Our Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This section presents the details about CFINet. We start with a discussion about the inherent limitations of Cascade RPN when confronting small objects, then our coarse-to-fine high-quality proposal generation pipeline tailored for size-limited instances is introduced. Afterwards, we elucidate the architecture of newly designed Feature Imitation branch, also with the optimization and training procedure." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6319" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.447, + 0.107 + ], + "angle": 0, + "content": "The overall architecture of CFINet is shown in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.117, + 0.312, + 0.133 + ], + "angle": 0, + "content": "3.1. Towards Better Proposals" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.141, + 0.47, + 0.367 + ], + "angle": 0, + "content": "Limitations of Cascade RPN. High-quality proposals play a pivotal role in two-stage detectors, but need heuristic anchor settings. Casdade RPN [32] discards this conventional setup by placing one single anchor on each feature point and conducting multi-stage refinement. Though exhibiting superior performance on objects of general scales, Cascade RPN fails to tackle extremely small objects well due to its inherent limitations. Concretely, the distance metric used in first-stage regression cannot guarantee sufficient potential anchors for small objects who have significantly small center region. Moreover, Cascade RPN only marks eligible anchors on a single pyramid level as positive, while this heuristic scheme simply discards those possible anchors at other levels which can still convey the existence and rough location information of small objects [42]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.47, + 0.504 + ], + "angle": 0, + "content": "Coarse-to-fine RPN. To remedy the aforementioned issues of Cascade RPN when handling small instances, we propose Coarse-to-fine RPN and the detailed structure is in Figure 2. First, we design an area-based anchor selection strategy to enable the instances of various sizes could have (relatively) adequate potential anchors. Concretely, for an object box with the width \\( w \\) and the height \\( h \\), any anchors who have an IoU larger than \\( T_{\\mathrm{a}} \\) will be regarded as positive for the coarse regression, and \\( T_{\\mathrm{a}} \\) is formulated as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.515, + 0.469, + 0.548 + ], + "angle": 0, + "content": "\\[\nT _ {\\mathrm {a}} = \\max (0. 2 5, 0. 2 0 + \\gamma \\cdot \\log \\frac {\\sqrt {w \\cdot h}}{1 2}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.469, + 0.719 + ], + "angle": 0, + "content": "where \\(\\gamma\\) denotes a scale factor and is set to default 0.15 in our experiments, and the term 12 actually corresponds to the minimal area definition of SODA dataset [9], which enables adequate samples for extreme-size objects and can be tuned for different datasets. Moreover, \\(\\gamma\\) and max operation keep the optimization from being overwhelmed by the low-quality priors. Taking IoU as the criterion to mine potential anchors, the optimization inconsistency in multi-stage regression of Cascade RPN can be averted. Meanwhile, the model determines the positive sample in a more smooth way on top of the proposed continuous threshold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Distinct from Cascade RPN, we preserve anchors of all Feature Pyramid Network (FPN) [25] levels \\(\\{P_2,P_3,P_4,P_5\\}\\) to perform first-stage regression. In this way, we could mine sufficient potential anchors for extremely small instances and meanwhile, larger instances still can obtain proper attention since the anchors matched to them have naturally higher IoUs, as discussed in Figure 1. After the first-stage regression, we then capture the offsets inside the regressed boxes and input them with the feature maps to RPN, in which the Adaptive Convolution [32] will be exploited to align the features and conduct second-stage regression and foreground-background classification." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.866, + 0.107 + ], + "angle": 0, + "content": "Loss Function. The training objective of our CRPN is:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.12, + 0.892, + 0.14 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C R P N}} = \\alpha_ {1} \\left(\\mathcal {L} _ {\\mathrm {r e g}} ^ {\\mathrm {c}} + \\mathcal {L} _ {\\mathrm {r e g}} ^ {\\mathrm {f}}\\right) + \\alpha_ {2} \\mathcal {L} _ {\\mathrm {c l s}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.144, + 0.892, + 0.22 + ], + "angle": 0, + "content": "where we use cross-entropy loss and IoU Loss [44] as \\(\\mathcal{L}_{\\mathrm{cls}}\\) and \\(\\mathcal{L}_{\\mathrm{reg}}\\), respectively. The c and f in Eq. (2) indicate the coarse-stage and fine-stage in our CRPN, and noting that we only do classification in the latter stage. The loss weights \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\) are set to 9.0 and 0.9, respectively." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.229, + 0.88, + 0.244 + ], + "angle": 0, + "content": "3.2. Feature Imitation for Small Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.892, + 0.492 + ], + "angle": 0, + "content": "Efforts on exploiting the intrinsic correlations between objects of different scales to boost the representations of small objects have been made, but most of them fail to the effectiveness and diversity. Specifically, the majority of previous methods [22, 2, 1, 28] resort to GAN to super-resolve the representations of small instances. This calls for the sophisticated training schemes and is prone to fabricate fake textures and artifacts [11]. Another line of efforts turn to the similarity learning which either has to construct offline feature bank in a cumbersome way [19], or directly leverages \\(\\mathcal{L}_2\\) norm to similarity measurement between different RoI features [36], potentially leading the feature collapse issue: the region features after amending could have high-similarity but lost their own characteristics. This homogenization in feature space actually impairs the generality and robustness of the model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.688 + ], + "angle": 0, + "content": "To mitigate the collapse risks and avoid the memory burden as well as enable the end-to-end optimization, we devise a Feature Imitation (FI) head (see Figure 2). Most importantly, instead of solely taking large-scale objects as the guidance of this procedure, we consider the model response in current state for each instance, thereby constructing a dynamic and currently optimized feature bank of proper exemplars in an online fashion. The FI branch mainly composes an Exemplar Feature Set and a Feature-to-Embedding (Feat2Embed) module, where the former reserves the RoI features of high-quality exemplars and the latter projects the input to the embedding space. Next we elucidate the details about our Feature Imitation branch." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.902 + ], + "angle": 0, + "content": "What is a proper exemplar? As we discussed above, an exemplar is vital in the imitation learning. To determine the most representative/proper/high-quality examples which can deliver authentic guidance/supervision for small objects confusing the model at this moment, we first introduce a simple quality indicator for an instance. Given a ground-truth (GT) object \\( g = (c^{*}, b^{*}) \\), where \\( c^{*} \\) and \\( b^{*} \\) denote its label and bounding box coordinates. Assuming the detection head outputs a prediction set \\( S = \\{C_{i}, IoU_{i}\\}_{i=1,2,\\dots,M} \\) for \\( g \\), in which \\( C_{i} \\in \\mathbb{R}^{N+1} \\) indicates the predicted classification vector and \\( IoU_{i} \\) stands for the IoU of predicted box to GT, and \\( N \\) is the number of foreground classes. Then we can obtain the potential high-quality set \\( S' = \\{(C_{j}, IoU_{i}) | \\arg \\max C_{j} = 0 \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6320" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.081, + 0.092, + 0.413, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 Training of Feature Imitation branch." + }, + { + "type": "title", + "bbox": [ + 0.084, + 0.109, + 0.131, + 0.123 + ], + "angle": 0, + "content": "Input:" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.124, + 0.468, + 0.154 + ], + "angle": 0, + "content": "The set of GT boxes \\(\\mathcal{G} = \\{c_i^*, b_i^*\\}_{i=1,2,\\dots,T}\\) and corresponding RoI features \\(\\{\\pmb{x}_i^g\\}_{i=1,2,\\dots,T}\\) in current batch;" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.154, + 0.431, + 0.169 + ], + "angle": 0, + "content": "The set of exemplar features \\(\\mathcal{E} = \\{\\mathcal{E}_i\\}_{c = 1,2,\\dots,N}\\)" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.169, + 0.469, + 0.199 + ], + "angle": 0, + "content": "The set of background RoI features in current batch \\(\\mathcal{X}_{\\mathrm{bg}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.199, + 0.336, + 0.214 + ], + "angle": 0, + "content": "The threshold of high-quality \\( T_{\\mathrm{hq}} \\)" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.214, + 0.425, + 0.23 + ], + "angle": 0, + "content": "The number of pos/neg samples \\(N_{\\mathrm{pos}}\\) and \\(N_{\\mathrm{neg}}\\);" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.23, + 0.314, + 0.243 + ], + "angle": 0, + "content": "The transformation function \\(\\Gamma\\);" + }, + { + "type": "title", + "bbox": [ + 0.084, + 0.244, + 0.143, + 0.258 + ], + "angle": 0, + "content": "Output:" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.26, + 0.469, + 0.29 + ], + "angle": 0, + "content": "The set of positive embeddings \\(\\mathcal{P}_{\\mathrm{pos}}\\) and negative embeddings \\(\\mathcal{P}_{\\mathrm{neg}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.29, + 0.469, + 0.32 + ], + "angle": 0, + "content": "1: Initialize the set of positive features \\(\\mathcal{X}_{\\mathrm{pos}}\\) and negative features \\(\\mathcal{X}_{\\mathrm{neg}}\\) with \\(\\varnothing\\);" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.321, + 0.197, + 0.334 + ], + "angle": 0, + "content": "2: for \\( g \\) in \\( \\mathcal{G} \\) do" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.335, + 0.451, + 0.35 + ], + "angle": 0, + "content": "3: Compute the \\( IQ \\) of current \\( g \\) according to Eq. (3)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.35, + 0.445, + 0.366 + ], + "angle": 0, + "content": "4: \\(\\mathcal{X}_{\\mathrm{neg}}^q \\gets\\) sample \\(N_{\\mathrm{neg}}\\) features from \\(\\mathcal{X}_{\\mathrm{bg}} \\cup \\mathcal{E} \\backslash \\mathcal{E}_{c^*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.365, + 0.243, + 0.381 + ], + "angle": 0, + "content": "5: if \\( I\\tilde{Q} \\geq T_{\\mathrm{hq}} \\) then" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.381, + 0.208, + 0.395 + ], + "angle": 0, + "content": "6: \\(\\mathcal{E}_{c^*}\\gets \\pmb {x}_i^g\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.395, + 0.24, + 0.411 + ], + "angle": 0, + "content": "7: \\(\\mathcal{X}_{\\mathrm{pos}}^g\\gets \\Gamma (\\pmb{x}_i^g)\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.412, + 0.153, + 0.424 + ], + "angle": 0, + "content": "8: else" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.425, + 0.398, + 0.441 + ], + "angle": 0, + "content": "9: \\(\\mathcal{X}_{\\mathrm{pos}}^q\\gets\\) sample \\(N_{\\mathrm{pos}}\\) features from \\(\\mathcal{E}_{c^{*}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.441, + 0.168, + 0.453 + ], + "angle": 0, + "content": "10: end if" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.455, + 0.404, + 0.472 + ], + "angle": 0, + "content": "11: \\(\\mathcal{X}_{\\mathrm{pos}} = \\mathcal{X}_{\\mathrm{pos}}\\cup \\mathcal{X}_{\\mathrm{pos}}^{g},\\mathcal{X}_{\\mathrm{neg}} = \\mathcal{X}_{\\mathrm{neg}}\\cup \\mathcal{X}_{\\mathrm{neg}}^{g}\\)" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.472, + 0.469, + 0.502 + ], + "angle": 0, + "content": "12: Apply Eq. (4) to \\(\\mathcal{X}_{\\mathrm{pos}}\\) and \\(\\mathcal{X}_{\\mathrm{neg}}^q\\) to obtain \\(\\mathcal{P}_{\\mathrm{pos}}\\) and \\(\\mathcal{P}_{\\mathrm{neg}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.502, + 0.16, + 0.514 + ], + "angle": 0, + "content": "13: end for" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.516, + 0.258, + 0.532 + ], + "angle": 0, + "content": "14: return \\(\\mathcal{P}_{\\mathrm{pos}}\\) and \\(\\mathcal{P}_{\\mathrm{neg}}\\)" + }, + { + "type": "list", + "bbox": [ + 0.084, + 0.26, + 0.469, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.469, + 0.594 + ], + "angle": 0, + "content": "\\(\\left.c^{*}\\right\\rbrace_{j = 1,2,\\dots,M^{\\prime}}\\) where \\(M^{\\prime}\\leq M\\) . Now the Instance Quality of an object \\(g\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.606, + 0.469, + 0.652 + ], + "angle": 0, + "content": "\\[\nI Q = \\frac {1}{\\left| \\mathcal {S} ^ {\\prime} \\right|} \\sum_ {j = 1} ^ {M ^ {\\prime}} C _ {j, c ^ {*}} \\cdot I o U _ {j} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.469, + 0.764 + ], + "angle": 0, + "content": "The \\( IQ \\) of a GT serves as an indicator of the current model's detection capability, enabling us to capture the high-quality exemplars who have precise localization and high-confidence classification scores, and the instances confusing the model often fail to either of them. By setting appropriate threshold, we can select proper instances to build the teacher feature-set and perform the imitation process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Feat2Embed Module. Instead of directly measuring the similarity between different RoI features [36], we first embed these features with the simple Feat2Embed module. The input of FI branch is the region feature \\( \\boldsymbol{x}_i \\in \\mathbb{R}^{H \\times W \\times C} \\) obtained by the RoI-wise operation, e.g., RoI Align, which will be first processed by three consecutive \\( 3 \\times 3 \\) convolutional layers (with no padding operation) to abstract compact representations. It is worth noting that we update parameters during the extraction of current regional features" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.09, + 0.699, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.09, + 0.888, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.26, + 0.892, + 0.303 + ], + "angle": 0, + "content": "Figure 3. Size distribution of instances in (a) SODA-D and (b) SODA-A, where the absolute size corresponds to the square root of the object area." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.329, + 0.892, + 0.493 + ], + "angle": 0, + "content": "and freeze parameters during the extraction of exemplar ones (see Feat2Embed module in Figure 2), resulting in improved stability in performance. Subsequently, the intermediate features will be mapped to the embedding space on top of a two-layer perceptron and the embedding layer with the dimension of 128, in which the dimension of hidden layers is set to 512. We have also investigated various design choices and structures for our Feat2Embed module, and detailed information can be found in the Supplementary Materials. Finally, the output of the Feature Imitation branch is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.512, + 0.891, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {i} = \\Theta_ {\\mathrm {F I}} (\\boldsymbol {x} _ {i}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.891, + 0.566 + ], + "angle": 0, + "content": "where \\(\\Theta_{\\mathrm{FI}}\\) denotes the parameters of Feature Imitation branch to be optimized." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.747 + ], + "angle": 0, + "content": "Loss Function. The objective of our FI head is simple: calculating the similarity between the RoI feature of proposal and that of the stored high-quality instances in embedding space, thereby pulling the features of those instances that confuse the model close to the exemplar ones of belonging category, while pushing apart from that of other categories and backgrounds. To this end, we propose a loss function based on Supervised Contrastive Learning [18] which extends the contrastive learning setup and allows multiple positive samples for an anchor object by exploiting the accessible label information. The loss function tailored for our FI branch is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.773, + 0.891, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {F I}} = \\frac {- 1}{| \\mathcal {P} _ {\\text {p o s}} |} \\sum_ {j} \\sum_ {v _ {\\mathrm {p}} \\in \\mathcal {P} _ {\\text {p o s}}} \\log \\frac {\\exp \\left(v _ {j} \\cdot v _ {\\mathrm {p}} / \\tau\\right)}{\\sum_ {v _ {i} \\in \\mathcal {P}} \\exp \\left(v _ {j} \\cdot v _ {i} / \\tau\\right)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{P} = \\mathcal{P}_{\\mathrm{pos}} \\cup \\mathcal{P}_{\\mathrm{neg}}\\) denotes the sample set, while \\(\\mathcal{P}_{\\mathrm{pos}}\\) and \\(\\mathcal{P}_{\\mathrm{neg}}\\) represent the positive and negative set respectively and they have the same cardinality ideally, and \\(v_{\\mathrm{p}}\\) and \\(v_{\\mathrm{n}}\\) are the positive and negative sample from \\(\\mathcal{P}_{\\mathrm{pos}}\\) and \\(\\mathcal{P}_{\\mathrm{neg}}\\). Moreover, \\(j\\) indexes the current proposal and \\(\\tau\\) indicates the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6321" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.141, + 0.088, + 0.828, + 0.362 + ], + "angle": 0, + "content": "
MethodPublicationScheduleAPAP50AP75APesSAPrsSAPgsSAPN
One-stage
RetianNet [26]ICCV'171×28.257.623.711.925.234.144.2
FCOS [31]ICCV'191×23.949.519.96.919.430.940.9
ATSS [46]CVPR'201×26.855.622.111.723.932.241.3
YOLOX [12]ArXiv'2170e26.753.423.013.625.130.930.4
DyHead [10]CVPR'211×27.556.123.212.424.433.041.9
Keypoint-based
CornerNet [21]ECCV'182×24.649.521.76.520.532.243.8
CenterNet [48]ArXiv'1970e21.548.815.65.116.229.642.4
RepPoints [43]ICCV'191×28.055.624.710.123.835.145.3
Query-based
Deformable-DETR [49]ICLR'2050e19.244.813.76.315.424.934.2
Sparse RCNN [30]CVPR'211×24.250.320.38.820.430.239.4
Two-stage
Baseline [29]NeurIPS'151×28.959.424.113.825.734.543.0
Cascade RPN [32]NeurIPS'191×29.156.525.912.525.535.444.7
RFLA [40]ECCV'221×29.760.225.213.226.935.444.6
CFINet (ours)-1×30.760.826.714.727.836.444.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.365, + 0.892, + 0.421 + ], + "angle": 0, + "content": "Table 2. Comparison with state-of-the-art detection approaches on the SODA-D test-set, where 'Baseline' refers to Faster RCNN [29], serving as the baseline for the two-stage methods in the table. All the methods are trained on a ResNet-50 [17], except YOLOX (CSP-Darknet) [12] and CornerNet (HourglassNet-104) [21]. 'Schedule' denotes the number of epochs for training, in which \\(1 \\times\\) corresponds to 12 epochs and \\(50\\mathrm{e}\\) indicates 50 epochs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.47, + 0.507 + ], + "angle": 0, + "content": "temperature which plays a crucial part in contrastive learning and needs to be well designed, and we conduct ablation studies (see Table 9) to determine the optimal setting in our framework. The total loss function is presented:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.532, + 0.47, + 0.549 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C R P N}} + \\mathcal {L} _ {\\mathrm {c l s}} + \\mathcal {L} _ {\\mathrm {r e g}} + \\alpha_ {3} \\mathcal {L} _ {\\mathrm {F I}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.47, + 0.686 + ], + "angle": 0, + "content": "in which \\(\\mathcal{L}_{\\mathrm{cls}}\\) and \\(\\mathcal{L}_{\\mathrm{reg}}\\) are the original losses of detection head, and the \\(\\alpha_{3}\\) is utilized to scale the weight of Feature Imitation part. With contrastive learning setups, not only can we fulfill the imitation learning but also prevent the collapse issue, thereby boosting the representations of small instances effectively. Moreover, the imitation process is only installed in training phase and will not slow the pace of inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Training. Next we elucidate the training details of FI branch. The exemplar set \\(\\mathcal{E} = \\{\\mathcal{E}_i\\}_{c=1,2,\\dots,N}\\) containing high-quality features of \\(N\\) foreground categories and \\(\\mathcal{E}_i = \\{\\pmb{x}_{i,j}\\}_{j=1,2,\\dots,N_i}\\) corresponding to the exemplar features of the \\(i\\)-th class, and \\(N_i\\) represents its size. We use \\(T_{\\mathrm{hq}}\\) to pick out those high-quality instances which are compatible to be a good exemplar, and in practice, we set a bound value to the number of high-quality predictions of an instance to filter the effect of fluctuation of the network. The function \\(\\Gamma\\) is used to augment features for high-quality instances, i.e., the positive features for a high-quality instance are the transformations of itself. The overall training procedure of FI branch is shown in Alg. 1 and more details please refer to the Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.444, + 0.633, + 0.461 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.469, + 0.596, + 0.484 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.553 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our method, we perform extensive experiments on the recently released large-scale benchmark tailored for small object detection: SODA [9], including SODA-D and SODA-A." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.658 + ], + "angle": 0, + "content": "SODA-D. Focusing on the driving scenario, SODA-D comprises 24828 high-quality images and 278433 instances distributed across nine categories: people, rider, bicycle, motor, vehicle, traffic-sign, traffic-light, traffic-camera, and warning-cone. One of the most distinctive strengths of SODA-D is its diversity in terms of period, geographical locations, weather conditions, camera viewpoints, etc." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.779 + ], + "angle": 0, + "content": "SODA-A. SODA-A contains 872069 objects with oriented box annotations in 2513 aerial images and encompassing nine classes: airplane, helicopter, small-vehicle, large-vehicle, ship, container, storage-tank, swimming-pool, and windmill. The instances in SODA-A can appear in arbitrary orientations and are with significant density variations. To be specific, the average number of instances per image in SODA-A is about 350." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As a specialized benchmark for small object detection, the instances in SODA are tiny, with most of them having an average size ranging from 10 to 30 pixels (see Figure 3). In contrast to conventional datasets for object detection, SODA includes extensive ignore annotations, aimed at filtering out instances that are either too large or are challenging to be identified deterministically due to heavy occlusion or lens flare. This procedure helps the model focus on valu" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6322" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.141, + 0.088, + 0.828, + 0.263 + ], + "angle": 0, + "content": "
MethodPublicationScheduleAPAP50AP75APesSAPrsSAPgsSAPN
One-stage
Rotated RetinaNet [26]ICCV'171×26.863.416.29.122.035.428.2
S2A-Net [15]TGRS'221×28.369.613.110.222.835.829.5
Oriented RepPoints [23]CVPR'221×26.358.819.09.422.632.428.5
DHRec [27]TPAMI'221×30.168.819.810.624.640.334.6
Two-stage
Baseline [29]NeurIPS'151×32.570.124.311.927.342.234.4
Gliding Vertex [41]TPAMI'211×31.770.822.611.727.041.133.8
Oriented RCNN [39]ICCV'211×34.470.728.612.528.644.536.7
DODet [8]TGRS'221×31.668.123.411.326.341.033.5
CFINet (ours)-1×34.473.126.113.529.344.035.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.266, + 0.893, + 0.296 + ], + "angle": 0, + "content": "Table 3. Comparison with state-of-the-art detection approaches on the SODA-A test-set, where 'Baseline' refers to Rotated Faster RCNN [29], serving as the baseline for the two-stage methods in the table. Other settings are consistent with Table 2." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.311, + 0.317, + 0.322 + ], + "angle": 0, + "content": "(a) Faster RCNN" + }, + { + "type": "image", + "bbox": [ + 0.142, + 0.323, + 0.365, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.446, + 0.311, + 0.55, + 0.322 + ], + "angle": 0, + "content": "(b) Cascade RPN" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.323, + 0.597, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.311, + 0.758, + 0.322 + ], + "angle": 0, + "content": "(c) CRPN" + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.323, + 0.829, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.477, + 0.893, + 0.52 + ], + "angle": 0, + "content": "Figure 4. The average number of high-quality proposals generated by (a) RPN, (b) Cascade RPN and (c) CRPN for instances in extremely Small \\((eS)\\), relatively Small \\((rS)\\), generally Small \\((gS)\\), and Normal \\((N)\\) subsets, respectively. Noting that a proposal who has an IoU larger than 0.5 to any ground-truth boxes will be registered as a high-quality proposal." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.544, + 0.485, + 0.636 + ], + "angle": 0, + "content": "
Proposal MethodAR\\( AR_{eS} \\)\\( AR_{rS} \\)\\( AR_{gS} \\)\\( AR_N \\)
RPN [29]41.224.038.347.357.1
RPN-0.541.324.238.547.354.1
GA-RPN [33]42.124.139.248.956.2
Cascade RPN [32]41.822.838.248.757.1
CRPN42.624.638.949.156.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.641, + 0.47, + 0.74 + ], + "angle": 0, + "content": "Table 4. Average Recall \\((AR)\\) performances of our CRPN and its counterparts on the SODA-D test-set. All the methods are with Faster RCNN (ResNet-50) as the baseline and trained for a \\(1\\times\\) schedule, in which RPN denotes the vanilla version of Faster RCNN whose positive threshold in RPN stage is set 0.7, and RPN-0.5 represents the version with 0.5 as its positive threshold of RPN. The results are tested with 300 proposals per image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "able small instances. The objects in SODA are divided into Small and Normal according to their areas, in which Small is further split into three subsets: extremely Small (eS), relatively Small (rS) and generally Small (gS). The evaluation metric of SODA follows that of COCO [24], namely averaging the precision over 10 IoU thresholds ranging from 0.5 to 0.95 (with an interval of 0.05), specifically focusing on Small objects." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.544, + 0.885, + 0.622 + ], + "angle": 0, + "content": "
BaselineCRPNFIAPAPeSAPrSAPgS
✓28.913.825.734.5
✓✓30.314.327.336.1
✓✓29.514.426.335.1
✓✓✓30.714.727.836.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.624, + 0.892, + 0.667 + ], + "angle": 0, + "content": "Table 5. Ablation analysis of our method, in which 'Baseline' denotes the vanilla Faster RCNN, CRPN and FI indicate Coarse-to-fine RPN and Feature Imitation branch, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.719, + 0.697 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In the following experiments, unless specified, we use train-set to conduct the training and leave test-set to performance comparisons and ablation studies. Considering that the images in SODA enjoy a very high resolution \\((\\sim 4000\\times 3000)\\), we first split the original images into a series of \\(800\\times 800\\) patches with a stride of 650, and similar to [9] these patches will be resized to \\(1200\\times 1200\\) during training and testing. All the experiments in this paper are conducted on a single RTX 3090 with the batch size of 4. Only random flip involved in data augmentation. We train all the models with a \\(1\\times\\) schedule (a bunch of 12 epochs), and the learning rate is set to 0.01 which decays after epoch 8 and epoch 11 by 0.1. The default optimizer is SGD with" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6323" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.139, + 0.089, + 0.409, + 0.148 + ], + "angle": 0, + "content": "
StrategyAPAPeSAPrSAPgS
0.2029.913.726.835.9
0.4030.114.126.936.2
Ours30.314.327.336.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.151, + 0.47, + 0.18 + ], + "angle": 0, + "content": "Table 6. Different definitions about positive anchor for CRPN, in which 'Ours' denotes the proposed dynamic strategy." + }, + { + "type": "table", + "bbox": [ + 0.139, + 0.192, + 0.41, + 0.257 + ], + "angle": 0, + "content": "
α3APAPeSAPrSAPgS
0.2530.614.727.636.2
0.5030.714.727.836.4
0.7530.414.027.536.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.26, + 0.45, + 0.273 + ], + "angle": 0, + "content": "Table 7. The effect of loss weight for Feature Imitation branch." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.289, + 0.469, + 0.318 + ], + "angle": 0, + "content": "the momentum of 0.9 and the weight decay of 0.0001. We use ResNet-50 [17] with FPN [25] for all models." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.33, + 0.216, + 0.344 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.353, + 0.469, + 0.398 + ], + "angle": 0, + "content": "To exhibit the effectiveness of our method, we conduct a thorough comparison with current representative approaches on the SODA-D and SODA-A." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.469, + 0.609 + ], + "angle": 0, + "content": "Table 2 represents the results of our method and several mainstream approaches on the SODA-D test-set. Integrating with Faster RCNN [29], our CFINet achieves state-of-the-art performance with an overall \\(AP\\) of \\(30.7\\%\\), and outperforms the baseline model with \\(1.8\\%\\) points. When delving into specific metrics, our method exhibits clear predominance, particularly on the most challenging metrics \\(AP_{es}\\) and \\(AP_{rs}\\). Moreover, CFINet exceeds the tailored small object detection method RFLA [40] by a significant margin (1.0% on \\(AP\\), 1.5% on \\(AP_{es}\\), and 0.9% on \\(AP_{rs}\\)) when both taking Faster RCNN as the baseline. Actually, RFLA sacrifices the performance on extremely Small instances (with a decrease of \\(0.6\\%\\) points when compared to Faster RCNN)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.469, + 0.716 + ], + "angle": 0, + "content": "On the SODA-A test-set, CFINet also achieves the best result and shows great advantage in comparison to other solutions especially on \\( AP_{eS} \\) (see Table 3), indicating its superiority and generality. Furthermore, albeit exhibiting advantage on \\( AP_{75} \\) metric and larger instances, Oriented RCNN [39] lags largely behind our approach on \\( AP_{50} \\) (73.1% vs. 70.7%) and \\( AP_{eS} \\) (13.5% vs. 12.5%)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.287, + 0.74 + ], + "angle": 0, + "content": "4.4. Effectiveness of CRPN" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.87 + ], + "angle": 0, + "content": "One of the main designs in this paper is the Coarse-to-fine RPN, which is based on the observation that current fixed overlaps-based sampling paradigm is inappropriate for small instances due to the inherent contradictions, while the refined designs of RPN could partially reduce this barrier but still fail to satisfactory results. Here, we conduct thorough analyses to demonstrate the capability of our CRPN to generate high-quality proposals for size-limited instances." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.9 + ], + "angle": 0, + "content": "We first exhibit the recall performances of our CRPN and its counterparts in Table 4, from which we can see that low" + }, + { + "type": "table", + "bbox": [ + 0.572, + 0.089, + 0.822, + 0.162 + ], + "angle": 0, + "content": "
ThqAPAPeSAPrSAPgS
0.5030.314.027.336.0
0.5530.614.327.336.5
0.6530.714.727.836.4
0.7030.514.627.336.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.166, + 0.892, + 0.191 + ], + "angle": 0, + "content": "Table 8. The investigation of the criterion to be an exemplar instance." + }, + { + "type": "table", + "bbox": [ + 0.571, + 0.205, + 0.822, + 0.278 + ], + "angle": 0, + "content": "
τAPAPeSAPrSAPgS
0.1030.314.127.336.1
0.5030.414.427.236.2
0.6030.714.727.836.4
0.8030.214.027.336.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.281, + 0.874, + 0.295 + ], + "angle": 0, + "content": "Table 9. The choices of temperature \\( \\tau \\) to the final performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.892, + 0.473 + ], + "angle": 0, + "content": "ering the positive threshold slightly improves the average recall while sacrificing the performance of larger instances \\((AR_{N}\\) experiences a sharp decline from \\(57.1\\%\\) to \\(54.1\\%\\)). GA-RPN [33] as well as Cascade RPN [32] both fail to better results since their patterns incline to large instances as we discussed before. In comparison with RPN and its variants, our CRPN demonstrates superior performance on objects in the Small, while achieving comparable results on Normal instances. This validates our assumption that refined proposal networks tend to favor larger objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.893, + 0.715 + ], + "angle": 0, + "content": "We conjecture that one of the most challenging issues towards accurate small object detection is the scarcity of high-quality samples, which is also the major motivation behind the design of CRPN. Hence, we intuitively compare the baseline RPN, Cascade RPN and our CRPN about the number of high-quality samples. In Figure 4, our CRPN generates more high-quality proposals compared to the other competitors. Interestingly, CRPN can dynamically shift the focus along the training: at the beginning, the model concentrates more on large objects which are conducive for early optimization while as the training goes, the model gradually shifts its attention to objects with small sizes that are usually not handled well before. This is interpretable since the instances having extremely limited sizes are with more uncertainties and fitting them in early-phase is not an optimal choice for the detector." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.722, + 0.74 + ], + "angle": 0, + "content": "4.5. Ablation and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.824 + ], + "angle": 0, + "content": "In this part, we conduct ablation studies as well as comprehensive discussions to attest the importance of the CRPN and FI branch, and moreover, determine the appropriate settings of our approach. All the experiments of this section are conducted on the SODA-D test-set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Investigation of Designed Components. We first perform ablation experiments to verify the effectiveness of two modules. As in Table 5, our CRPN and FI can both improve the performances steadily, while the introduction of feature imitation branch is conducive for the recognition of size-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6324" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.495, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.315, + 0.148, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.316, + 0.179, + 0.326 + ], + "angle": 0, + "content": "people" + }, + { + "type": "image", + "bbox": [ + 0.195, + 0.315, + 0.215, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.316, + 0.243, + 0.326 + ], + "angle": 0, + "content": "rider" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.315, + 0.282, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.284, + 0.316, + 0.316, + 0.326 + ], + "angle": 0, + "content": "bicycle" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.315, + 0.359, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.316, + 0.387, + 0.326 + ], + "angle": 0, + "content": "motor" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.316, + 0.431, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.433, + 0.316, + 0.462, + 0.326 + ], + "angle": 0, + "content": "vehicle" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.091, + 0.911, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.316, + 0.524, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.316, + 0.55, + 0.326 + ], + "angle": 0, + "content": "c-sign" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.316, + 0.595, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.316, + 0.645, + 0.326 + ], + "angle": 0, + "content": "traffic-light" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.316, + 0.688, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.69, + 0.316, + 0.748, + 0.326 + ], + "angle": 0, + "content": "traffic-camera" + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.316, + 0.814, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.316, + 0.852, + 0.326 + ], + "angle": 0, + "content": "ing-cone" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.33, + 0.892, + 0.358 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of our method on the SODA-D test-set. Only predictions with confidence scores larger than 0.3 are demonstrated and the masked bounding boxes represent ignore regions. Best viewed in color and zoom-in windows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.469, + 0.458 + ], + "angle": 0, + "content": "limited instances. Consistently, the integration of FI and CRPN achieves the best result, as CRPN is capable of generating sufficient high-quality proposals (as shown in Figure 4) thereby offering a more accurate indication of instance quality and the potential to act as an exemplar." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.469, + 0.626 + ], + "angle": 0, + "content": "Fixed or Dynamic. A natural idea is directly setting fixed positive threshold to obtain more anchors for one-stage regression of CRPN. Here we show that our simple-yet-effective area-based anchor mining strategy can achieve the best performance. In Table 6, when the IoU threshold of a positive potential anchor for first-stage regression drops to 0.20, the \\(AP_{eS}\\) is only \\(13.7\\%\\) and this could be attributed to low-quality samples and the predominance of large instances as discussed before. The proposed area-based anchor mining strategy could mitigate this problem and obtain the best overall accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.469, + 0.718 + ], + "angle": 0, + "content": "Weights of Feature Imitation Loss. In this section, we analyze the impact of the weight parameter of FI branch (namely the hyper-parameter \\(\\alpha_{3}\\)) on the model. As shown in Table 7, paying too little or excessive attention both deteriorates the final performance, hence we set \\(\\alpha_{3}\\) to 0.5 in our experiments to ensure overall accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The Criterion of Being An Exemplar. The quality of the exemplar plays a pivotal role in the imitation process [19, 36]. Next we discuss the choices for capturing a high-quality exemplar to build the feature set. In Table 8, the lower \\( T_{\\mathrm{hq}} \\) involves more exemplars and updates the teacher set more frequently while the higher \\( T_{\\mathrm{hq}} \\) does exactly the opposite. It can be seen that increasing the \\( T_{\\mathrm{hq}} \\) from 0.5 to 0.65 could facilitate the imitation process and when the \\( T_{\\mathrm{hq}} \\) reaches 0.70, the overall \\( AP \\) drops instead. This may originates that the earliest samples stored in the feature set are inappropriate for current state, because the optimization is dynamic and the model is evolving hence the criterion of" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.383, + 0.765, + 0.398 + ], + "angle": 0, + "content": "being an exemplar has already changed." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.399, + 0.892, + 0.488 + ], + "angle": 0, + "content": "The Temperature. The temperature is paramount for contrastive learning [18] and we conduct a series of experiments to verify the best choice for \\(\\tau\\). From Table 9, when \\(\\tau\\) ranges from 0.10 to 0.80, the overall performance increases first and then decreases to \\(30.2\\%\\), therefore we choose 0.6 for our method." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.489, + 0.892, + 0.549 + ], + "angle": 0, + "content": "Visualization. We demonstrate the visualization results of example images from SODA-D test-set in Figure 5 to intuitively show the capability of our detector when detecting the small instances." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.564, + 0.618, + 0.579 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.892, + 0.786 + ], + "angle": 0, + "content": "In this paper, we proposed CFINet, a two-stage detector based on the Coarse-to-fine Region Proposal Network and Feature Imitation setups, in which the former can produce sufficient high-quality proposals for small instances particularly for those with extremely limited sizes. Then the novel detection head on top of the feature imitation branch facilitates the representations of small objects posing challenges to the model under the contrastive learning paradigm. The experiments results show that our method achieves state-of-the-art performance on the large-scale small object detection datasets SODA-D and SODA-A. In the future, a more flexible and general indicator of instance quality is worth investigating." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.659, + 0.817 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work was supported in part by the National Science Foundation of China under Grant 62136007 and Grant U20B2068, and in part by the Natural Science Basic Research Program of Shaanxi under Grants 2021JC-16 and 2023-JC-ZD-36." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "6325" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. Finding tiny faces in the wild with generative adversarial network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 21-30, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. Sod-mtgan: Small object detection via multi-task generative adversarial network. In Proceedings of the European Conference on Computer Vision, pages 206-221, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.313 + ], + "angle": 0, + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European Conference on Computer Vision, pages 213-229, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.47, + 0.372 + ], + "angle": 0, + "content": "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Cconference on Machine Learning, pages 1597-1607, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.469, + 0.429 + ], + "angle": 0, + "content": "[5] Gong Cheng, Junwei Han, Peicheng Zhou, and Dong Xu. Learning rotation-invariant and fisher discriminative convolutional neural networks for object detection. IEEE Transactions on Image Processing, 28(1):265-278, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.469, + 0.499 + ], + "angle": 0, + "content": "[6] Gong Cheng, Qingyang Li, Guangxing Wang, Xingxing Xie, Lingtong Min, and Junwei Han. Sfrnet: Fine-grained oriented object recognition via separate feature refinement. IEEE Transactions on Geoscience and Remote Sensing, 61:1-10, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.469, + 0.557 + ], + "angle": 0, + "content": "[7] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[8] Gong Cheng, Yanqing Yao, Shengyang Li, Ke Li, Xingxing Xie, Jiabao Wang, Xiwen Yao, and Junwei Han. Dual-aligned oriented detector. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[9] Gong Cheng, Xiang Yuan, Xiwen Yao, Kebing Yan, Qinghua Zeng, Xingxing Xie, and Junwei Han. Towards large-scale small object detection: Survey and benchmarks. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-20, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[10] Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unifying object detection heads with attentions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7373-7382, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[11] Chunfang Deng, Mengmeng Wang, Liang Liu, Yong Liu, and Yunliang Jiang. Extended feature pyramid network for small object detection. IEEE Transactions on Multimedia, 24:1968-1979, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[12] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Spyros Gidaris and Nikos Komodakis. Attend refine repeat: Active box proposal generation via in-out localization. arXiv preprint arXiv:1606.04446, 2016." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[15] Jiaming Han, Jian Ding, Jie Li, and Gui-Song Xia. Align deep features for oriented object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[18] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[19] Jung Uk Kim, Sungjune Park, and Yong Man Ro. Robust small-scale pedestrian detection with cued recall via memory learning. In Proceedings of the IEEE International Conference on Computer Vision, pages 3050-3059, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[20] Chunbo Lang, Gong Cheng, Binfei Tu, Chao Li, and Junwei Han. Base and meta: A new perspective on few-shot segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.533 + ], + "angle": 0, + "content": "[21] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In ECCV, pages 734-750, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.534, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[22] Jianan Li, Xiaodan Liang, Yunchao Wei, Tingfa Xu, Jiashi Feng, and Shuicheng Yan. Perceptual generative adversarial networks for small object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1222-1230, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[23] Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented reploints for aerial object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1829-1838, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[24] Tsung Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, pages 740-755, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2117-2125, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. pages 2980-2988, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Guangtao Nie and Hua Huang. Multi-oriented object detection in aerial images with double horizontal rectangles. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4932-4944, 2022." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6326" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[28] Junhyug Noh, Wonho Bae, Wonhee Lee, Jinhwan Seo, and Gunhee Kim. Better to follow, follow to be better: Towards precise supervision of feature super-resolution for small object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 9725-9734, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[29] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.275 + ], + "angle": 0, + "content": "[30] Peize Sun et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 14449-14458, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.47, + 0.318 + ], + "angle": 0, + "content": "[31] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. pages 9627-9636, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.47, + 0.375 + ], + "angle": 0, + "content": "[32] Thang Vu, Hyunjun Jang, Trung X Pham, and Chang Yoo. Cascade rpn: Delving into high-quality region proposal network with adaptive convolution. Advances in Neural Information Processing Systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.47, + 0.432 + ], + "angle": 0, + "content": "[33] Jiaqi Wang, Kai Chen, Shuo Yang, Chen Change Loy, and Dahua Lin. Region proposal by guided anchoring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2965-2974, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[34] Jinwang Wang, Wen Yang, Haowen Guo, Ruixiang Zhang, and Gui-Song Xia. Tiny object detection in aerial images. In IEEE International Conference on Pattern Recognition, pages 3791-3798, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[35] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 7303-7313, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.617 + ], + "angle": 0, + "content": "[36] Jialian Wu, Chunluan Zhou, Qian Zhang, Ming Yang, and Junsong Yuan. Self-mimic learning for small-scale pedestrian detection. In Proceedings of the ACM International Conference on Multimedia, pages 2012-2020, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[37] Wei Wu, Hao Chang, Yonghua Zheng, Zhu Li, Zhiwen Chen, and Ziheng Zhang. Contrastive learning-based robust object detection under smoky conditions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4295-4302, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[38] Enze Xie, Jian Ding, Wenhai Wang, Xiaohang Zhan, Hang Xu, Peize Sun, Zhenguo Li, and Ping Luo. Detco: Unsupervised contrastive learning for object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 8392-8401, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[39] Xingxing Xie, Gong Cheng, Jiabao Wang, Xiwen Yao, and Junwei Han. Oriented r-cnn for object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 3520-3529, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[40] Chang Xu, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Rfla: Gaussian receptive based label assignment for tiny object detection. In Proceedings of the European Conference on Computer Vision, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[41] Yongchao Xu, Mingtao Fu, Qimeng Wang, Yukang Wang, Kai Chen, Gui-Song Xia, and Xiang Bai. Gliding vertex on" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "the horizontal bounding box for multi-oriented object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(4):1452-1459, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[42] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. Querydet: Cascaded sparse query for accelerating high-resolution small object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 13668-13677, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[43] Ze Yang, Shaohui Liu, Han Hu, Liwei Wang, and Stephen Lin. Repoints: Point set representation for object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 9657-9666, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[44] Jiahui Yu, Yuning Jiang, Zhangyang Wang, Zhimin Cao, and Thomas Huang. Unitbox: An advanced object detection network. In Proceedings of the ACM International Conference on Multimedia, pages 516-520, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[45] Xuehui Yu, Yuqi Gong, Nan Jiang, Qixiang Ye, and Zhenjun Han. Scale match for tiny person detection. In IEEE Conference on Winter Conference on Applications of Computer Vision, pages 1257-1265, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[46] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9759-9768, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[47] Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, and Stan Z Li. S3fd: Single shot scale-invariant face detector. In Proceedings of the IEEE International Conference on Computer Vision, pages 192-201, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[48] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[49] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6327" + } + ] +] \ No newline at end of file diff --git a/2023/Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning/231f8e4e-8ec6-40cc-a455-514ae44e6fa1_origin.pdf b/2023/Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning/231f8e4e-8ec6-40cc-a455-514ae44e6fa1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..36a257c76fda53f0c5dc5ff6ea07a88f0cdd5071 --- /dev/null +++ b/2023/Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning/231f8e4e-8ec6-40cc-a455-514ae44e6fa1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc37a66464acc28dba3cfa69a55671327c85f4bc69e7d23aa8c7b545029d731c +size 9924314 diff --git a/2023/Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning/full.md b/2023/Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0a31ffa1eb673ff5581f64247c64b5ecf2a321b5 --- /dev/null +++ b/2023/Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning/full.md @@ -0,0 +1,364 @@ +# Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning + +Xiang Yuan Gong Cheng * Kebing Yan Qinghua Zeng Junwei Han School of Automation, Northwestern Polytechnical University, Xi'an, China {shaunyuan, kebingyan, zengqinghua}@mail.nwpu.edu.cn, {gcheng, jhan}@nwpu.edu.cn + +# Abstract + +The past few years have witnessed the immense success of object detection, while current excellent detectors struggle on tackling size-limited instances. Concretely, the well-known challenge of low overlaps between the priors and object regions leads to a constrained sample pool for optimization, and the paucity of discriminative information further aggravates the recognition. To alleviate the aforementioned issues, we propose CFINet, a two-stage framework tailored for small object detection based on the Coarse-to-fine pipeline and Feature Imitation learning. Firstly, we introduce Coarse-to-fine RPN (CRPN) to ensure sufficient and high-quality proposals for small objects through the dynamic anchor selection strategy and cascade regression. Then, we equip the conventional detection head with a Feature Imitation (FI) branch to facilitate the region representations of size-limited instances that perplex the model in an imitation manner. Moreover, an auxiliary imitation loss following supervised contrastive learning paradigm is devised to optimize this branch. When integrated with Faster RCNN, CFINet achieves state-of-the-art performance on the large-scale small object detection benchmarks, SODA-D and SODA-A, underscoring its superiority over baseline detector and other mainstream detection approaches. Code is available at https://github.com/shaunyuan22/CFINet. + +# 1. Introduction + +Small object detection (SOD) aims to classify and localize the instances with limited regions, which plays an important role in a wide range of scenarios, such as pedestrian detection, autonomous driving, and intelligent surveillance understanding, to name a few [42, 28, 22, 19, 36, 2, 20]. Compared to the generic object detection which has been + +![](images/c37fa509a0e67644577a7ce2fba591e7ecec0af059b50881524a43feb10d684c.jpg) +Figure 1. Distribution of maximum IoU of anchors matched to each ground-truth instance in SODA-D [9] train-set, where extremely Small (eS), relatively Small (rS), and generally Small (gS) correspond to three area subsets in SODA [9] with the ranges (0, 144], (144, 400] and (400, 1024]. The smaller the objects are, the lower IoU the matched anchors have, hence the commonly used positive IoU threshold (0.7) is too rigorous for small objects. + +![](images/c95f8e6f6e719c3c11a981aa942d2e24e7e5f5292ce20f61c636f77ca6ff9040.jpg) + +![](images/9307cc36ff082cb685b821386ade2a48547d27db4f897e62781ae5052b023836.jpg) + +extensively studied, SOD task receives relatively little attention and good solutions are still scarce so far. Moreover, generic detectors [29, 6, 25, 31, 7, 3, 26, 5] usually struggle on handling small objects due to two inherent challenges: the insufficient and low-quality samples for training and the uncertain prediction of RoIs (Region of Interests). + +First, current prevailing detectors exploit either overlap-based [29, 26] or distance-based [31] strategies to select the positive priors of objects for training. However, small instances usually occupy an extremely limited area, therefore the region overlaps between densely arranged anchors and ground truth boxes are significantly small and far from the commonly used positive IoU (Intersection-over-Union) threshold, as in Figure 1. In other words, the existing positive sample criterion is overly stringent when applied to small/tiny objects, resulting in a restricted number of samples available for optimization. An intuitive approach involves reducing the threshold for defining a positive sample [47]. However, while this can lead to an increase in the number of positive samples, it often at the expense of overall sample quality, in which low-quality samples disrupt optimization and incur a trivial regression solution. Worse still, this is actually contradictory to the purpose of proposal + +
ctrl/ign ratioAPAP50AP75APesSAPrsSAPgSAPN
Baseline28.959.424.113.825.734.543.0
0.2/0.529.156.525.912.525.535.444.7
0.5/0.829.557.826.013.526.235.845.0
0.8/1.027.554.124.311.223.833.742.5
+ +Table 1. The performances of Cascade RPN [32] compared to the baseline (a vanilla Faster RCNN [29]). The ctr/ign ratio denotes the sampling region in first regression stage of Cascade RPN. The results are tested on the SODA-D [9] test-set and with a ResNet-50 [17] as the backbone. + +network, i.e., guaranteeing the recall and ease the burden of subsequent work. + +To sum up, current prevailing priors-to-proposals paradigms that heavily depend on the overlap or distance metric have inherent limitations in detecting small objects, and nowadays devised assignment or sampling schemes contribute minimally to this problem [9, 40]. Since the proposals play such a crucial role in two-stage detectors, so how about the improved Region Proposal Network (RPN) variants [13, 32, 33] meet small object detection? Following this line, we take Cascade RPN [32], one of the most superior proposal network for generic object detection, to perform preliminary experiments and the results are shown in Table 1. While the auxiliary regression phase provides refined priors with better initialization for subsequent regression, the final results remain somewhat unsatisfactory. Specifically, the improvement mainly comes from the larger objects and the $AP_{eS}$ as well as $AP_{rS}$ actually decrease significantly instead, indicating that the region-based sampling strategy is inclined to large instances which further dominate the proposal network. Meanwhile, enlarging the sample region contributes little (even negative) to this condition (see the bottom row in Table 1). Therefore, the coarse-to-fine pipeline has the potential to surmount the barrier of conventional prior-to-proposal paradigm, but the crux lies in dedicating sufficient attention to small instances. + +Second, small objects usually lack discriminative information and distorted structures, leading the inclination of model to give ambiguous even incorrect predictions [1, 11]. Meanwhile, there are a certain amount of large instances embodying clear visual cues and better discrimination. Building upon this observation, several works proposed to bridge the representation gap between small objects and large ones, and most of them [2, 22, 28, 1] rely on Generative Adversarial Network (GAN) [14] or similarity learning [19, 36] to super-resolve/restore the features of size-limited instances under the guidance of large ones that are deemed to be visually authentic. However, these approaches overlook the fact: high quality $\neq$ large size meanwhile small size $\neq$ low quality. In other words, the criterion for humans and the model to decide whether a sample is competent to be a good example is distinct. For the latter, it is dynamic and should be adjusted according to + +the current optimization of the detector. Moreover, efforts in this line have to resort to sophisticated training strategies or additional models, which is time-consuming and break the conventional end-to-end paradigm. + +Putting the above parts together, we propose a two-stage small object detector CFINet based on the coarse-to-fine pipeline and feature imitation learning. Concretely, enlightened by the multi-stage proposal generation scheme in Cascade RPN, we devise Coarse-to-fine RPN (CRPN). It firstly employs an dynamic anchor selection strategy to mine potential priors to conduct coarse regression, and henceforth, these refined anchors will be classified and regressed by the region proposal network. In addition, we extend the conventional classification-and-regression setting with an auxiliary Feature Imitation (FI) branch, which can leverage the regional features of high-quality instances to guide the learning of those objects with uncertain/mistaken predictions, and a loss function based on the Supervised Contrastive Learning (SCL) [18] is designed to optimize the whole process. The main contributions of this paper are summarized as follows: + +- A coarse-to-fine proposal generation pipeline named CRPN was built to perform anchor-to-proposal procedure, where an area-based anchor mining strategy and cascade regression empower the high-quality proposals for small instances. +- An auxiliary Feature Imitation (FI) branch was introduced to enrich the representations of low-quality instances perplexing the model under the supervision of high-quality instances, and this novel branch is optimized by a tailored loss function based on SCL. +- The experiment results on the SODA-D and SODA-A datasets exhibit the superiority of our CFINet to detect these instances with extremely limited sizes. + +# 2. Related Works + +Anchor Refinement and Region Proposals. Two-stage anchor-based approaches heavily depend on the high-quality proposals [33, 32]. Towards this goal, RPN was first introduced in Faster RCNN [29] to produce proposals in a fully convolutional network, and this simple yet effective design facilitates end-to-end model optimization. Following RPN, [13] proposed to iteratively regress the predefined anchors. GA-RPN [33] discards the uniform anchoring strategy and formulates the anchor generation in two steps: first determining the locations which may contain objects and on which the anchor scales are predicted then. By installing a multi-stage anchor-to-proposal strategy and alleviating the misalignment between the refined anchors and image features, Cascade RPN [32] enables high-quality proposal generation. Unfortunately, current proposal-oriented + +![](images/70bd0e14cf5547c9f29db918e444f9f99a5355b3bfc96bfb9fd021ab91b8c27e.jpg) +Figure 2. The overall architecture of CFINet. In Coarse-to-fine RPN (CRPN), the area-based Anchor Mining strategy ensures sufficient candidates for instances of various sizes (small: orange boxes, large: green boxes) based on the dynamic pos threshold, which will be then used to obtain the coarse proposals. After that, the alignment between coarse proposals and corresponding features is enabled by the Adaptive Convolution before feeding into the RPN to produce high-quality proposals (blue boxes). The Feature Imitation (FI) branch is devised to facilitate the representations of small instances, in which the RoI features of uncertain/mistaken predictions will be pulled to their counterparts of exemplar feature set in embedding space (throughout the Feat2Embed module), while pushed apart from the exemplar features of other categories and background. And the exemplar features are collected based on the model predictions using the proposed quality indicator, i.e., Instance Quality $(IQ)$ . We tailor a Feature Imitation loss function $\mathcal{L}_{\mathrm{FI}}$ to optimize this auxiliary branch. Note that we only exhibit single-level Feature Pyramid Network (FPN) [25] feature for clear illustration. + +![](images/e11300be834b1610a6e82ed6ecf56fbdfbff87f116a3d2dd9121205f9c4315a8.jpg) + +frameworks fail to produce high-quality proposals for instances with limited regions, and the root cause lies in the notoriously low overlaps between the objects and priors [40]. Different from the above methods, our coarse-to-fine proposal pipeline could exploit the potential of multi-stage refinement paradigm, thereby guaranteeing both the quantity and quality of proposals for instances with extremely limited sizes. + +Feature Imitation for Small Object Detection. One of the major challenges to detect small objects is the low-quality representation [45, 40, 9], while large instances often with clear structures and discriminative features. Hence, a series of efforts have been made to boost the semantic representations of small/tiny instances by mining the intrinsic correlations between small and large objects. Based on the generative adversarial paradigm, Perceptual GAN [22] designs a generator that is optimized to produce high-quality representations of small instances to fool the subsequent discriminator. Bai et al. [1] devised a novel pipeline to restore a clear face from the inputting blurry one. Noh et al. [28] further introduced precise supervision for the super-resolution process of small objects. Moreover, Wu et al. [36] and Kim et al. [19] both exploited similarity learning to force the features of small-scale pedestrians close to that of the large-scale ones which are obtained by an additional model. The existence of super-resolution branch or offline + +feature bank hampers the end-to-end optimization while our method updates the exemplar features in an online fashion, which guarantees the diversity of high-quality feature sets thereby getting rid of the collapse issue. + +Contrastive Learning for Object Detection. The recent explosion of self-supervised learning mainly comes from its Contrastive Learning fashion, and several works have extended this paradigm into detection fields. Detco [38] is an effective self-supervised framework for object detection which utilizes the image and its local patches to conduct contrastive learning. Wu et al. [37] applied contrastive learning to object detection under smoky conditions. Though contrastive learning has recently received considerable interests [4, 16, 35], the potential of utilizing contrastive learning for better representation of small objects has not yet been investigated to date. + +# 3. Our Method + +This section presents the details about CFINet. We start with a discussion about the inherent limitations of Cascade RPN when confronting small objects, then our coarse-to-fine high-quality proposal generation pipeline tailored for size-limited instances is introduced. Afterwards, we elucidate the architecture of newly designed Feature Imitation branch, also with the optimization and training procedure. + +The overall architecture of CFINet is shown in Figure 2. + +# 3.1. Towards Better Proposals + +Limitations of Cascade RPN. High-quality proposals play a pivotal role in two-stage detectors, but need heuristic anchor settings. Casdade RPN [32] discards this conventional setup by placing one single anchor on each feature point and conducting multi-stage refinement. Though exhibiting superior performance on objects of general scales, Cascade RPN fails to tackle extremely small objects well due to its inherent limitations. Concretely, the distance metric used in first-stage regression cannot guarantee sufficient potential anchors for small objects who have significantly small center region. Moreover, Cascade RPN only marks eligible anchors on a single pyramid level as positive, while this heuristic scheme simply discards those possible anchors at other levels which can still convey the existence and rough location information of small objects [42]. + +Coarse-to-fine RPN. To remedy the aforementioned issues of Cascade RPN when handling small instances, we propose Coarse-to-fine RPN and the detailed structure is in Figure 2. First, we design an area-based anchor selection strategy to enable the instances of various sizes could have (relatively) adequate potential anchors. Concretely, for an object box with the width $w$ and the height $h$ , any anchors who have an IoU larger than $T_{\mathrm{a}}$ will be regarded as positive for the coarse regression, and $T_{\mathrm{a}}$ is formulated as follow: + +$$ +T _ {\mathrm {a}} = \max (0. 2 5, 0. 2 0 + \gamma \cdot \log \frac {\sqrt {w \cdot h}}{1 2}), \tag {1} +$$ + +where $\gamma$ denotes a scale factor and is set to default 0.15 in our experiments, and the term 12 actually corresponds to the minimal area definition of SODA dataset [9], which enables adequate samples for extreme-size objects and can be tuned for different datasets. Moreover, $\gamma$ and max operation keep the optimization from being overwhelmed by the low-quality priors. Taking IoU as the criterion to mine potential anchors, the optimization inconsistency in multi-stage regression of Cascade RPN can be averted. Meanwhile, the model determines the positive sample in a more smooth way on top of the proposed continuous threshold. + +Distinct from Cascade RPN, we preserve anchors of all Feature Pyramid Network (FPN) [25] levels $\{P_2,P_3,P_4,P_5\}$ to perform first-stage regression. In this way, we could mine sufficient potential anchors for extremely small instances and meanwhile, larger instances still can obtain proper attention since the anchors matched to them have naturally higher IoUs, as discussed in Figure 1. After the first-stage regression, we then capture the offsets inside the regressed boxes and input them with the feature maps to RPN, in which the Adaptive Convolution [32] will be exploited to align the features and conduct second-stage regression and foreground-background classification. + +Loss Function. The training objective of our CRPN is: + +$$ +\mathcal {L} _ {\mathrm {C R P N}} = \alpha_ {1} \left(\mathcal {L} _ {\mathrm {r e g}} ^ {\mathrm {c}} + \mathcal {L} _ {\mathrm {r e g}} ^ {\mathrm {f}}\right) + \alpha_ {2} \mathcal {L} _ {\mathrm {c l s}}, \tag {2} +$$ + +where we use cross-entropy loss and IoU Loss [44] as $\mathcal{L}_{\mathrm{cls}}$ and $\mathcal{L}_{\mathrm{reg}}$ , respectively. The c and f in Eq. (2) indicate the coarse-stage and fine-stage in our CRPN, and noting that we only do classification in the latter stage. The loss weights $\alpha_{1}$ and $\alpha_{2}$ are set to 9.0 and 0.9, respectively. + +# 3.2. Feature Imitation for Small Object Detection + +Efforts on exploiting the intrinsic correlations between objects of different scales to boost the representations of small objects have been made, but most of them fail to the effectiveness and diversity. Specifically, the majority of previous methods [22, 2, 1, 28] resort to GAN to super-resolve the representations of small instances. This calls for the sophisticated training schemes and is prone to fabricate fake textures and artifacts [11]. Another line of efforts turn to the similarity learning which either has to construct offline feature bank in a cumbersome way [19], or directly leverages $\mathcal{L}_2$ norm to similarity measurement between different RoI features [36], potentially leading the feature collapse issue: the region features after amending could have high-similarity but lost their own characteristics. This homogenization in feature space actually impairs the generality and robustness of the model. + +To mitigate the collapse risks and avoid the memory burden as well as enable the end-to-end optimization, we devise a Feature Imitation (FI) head (see Figure 2). Most importantly, instead of solely taking large-scale objects as the guidance of this procedure, we consider the model response in current state for each instance, thereby constructing a dynamic and currently optimized feature bank of proper exemplars in an online fashion. The FI branch mainly composes an Exemplar Feature Set and a Feature-to-Embedding (Feat2Embed) module, where the former reserves the RoI features of high-quality exemplars and the latter projects the input to the embedding space. Next we elucidate the details about our Feature Imitation branch. + +What is a proper exemplar? As we discussed above, an exemplar is vital in the imitation learning. To determine the most representative/proper/high-quality examples which can deliver authentic guidance/supervision for small objects confusing the model at this moment, we first introduce a simple quality indicator for an instance. Given a ground-truth (GT) object $g = (c^{*}, b^{*})$ , where $c^{*}$ and $b^{*}$ denote its label and bounding box coordinates. Assuming the detection head outputs a prediction set $S = \{C_{i}, IoU_{i}\}_{i=1,2,\dots,M}$ for $g$ , in which $C_{i} \in \mathbb{R}^{N+1}$ indicates the predicted classification vector and $IoU_{i}$ stands for the IoU of predicted box to GT, and $N$ is the number of foreground classes. Then we can obtain the potential high-quality set $S' = \{(C_{j}, IoU_{i}) | \arg \max C_{j} = 0$ + +# Algorithm 1 Training of Feature Imitation branch. + +# Input: + +The set of GT boxes $\mathcal{G} = \{c_i^*, b_i^*\}_{i=1,2,\dots,T}$ and corresponding RoI features $\{\pmb{x}_i^g\}_{i=1,2,\dots,T}$ in current batch; + +The set of exemplar features $\mathcal{E} = \{\mathcal{E}_i\}_{c = 1,2,\dots,N}$ + +The set of background RoI features in current batch $\mathcal{X}_{\mathrm{bg}}$ + +The threshold of high-quality $T_{\mathrm{hq}}$ + +The number of pos/neg samples $N_{\mathrm{pos}}$ and $N_{\mathrm{neg}}$ ; + +The transformation function $\Gamma$ ; + +# Output: + +The set of positive embeddings $\mathcal{P}_{\mathrm{pos}}$ and negative embeddings $\mathcal{P}_{\mathrm{neg}}$ +1: Initialize the set of positive features $\mathcal{X}_{\mathrm{pos}}$ and negative features $\mathcal{X}_{\mathrm{neg}}$ with $\varnothing$ ; +2: for $g$ in $\mathcal{G}$ do +3: Compute the $IQ$ of current $g$ according to Eq. (3) +4: $\mathcal{X}_{\mathrm{neg}}^q \gets$ sample $N_{\mathrm{neg}}$ features from $\mathcal{X}_{\mathrm{bg}} \cup \mathcal{E} \backslash \mathcal{E}_{c^*}$ +5: if $I\tilde{Q} \geq T_{\mathrm{hq}}$ then +6: $\mathcal{E}_{c^*}\gets \pmb {x}_i^g$ +7: $\mathcal{X}_{\mathrm{pos}}^g\gets \Gamma (\pmb{x}_i^g)$ +8: else +9: $\mathcal{X}_{\mathrm{pos}}^q\gets$ sample $N_{\mathrm{pos}}$ features from $\mathcal{E}_{c^{*}}$ +10: end if +11: $\mathcal{X}_{\mathrm{pos}} = \mathcal{X}_{\mathrm{pos}}\cup \mathcal{X}_{\mathrm{pos}}^{g},\mathcal{X}_{\mathrm{neg}} = \mathcal{X}_{\mathrm{neg}}\cup \mathcal{X}_{\mathrm{neg}}^{g}$ +12: Apply Eq. (4) to $\mathcal{X}_{\mathrm{pos}}$ and $\mathcal{X}_{\mathrm{neg}}^q$ to obtain $\mathcal{P}_{\mathrm{pos}}$ and $\mathcal{P}_{\mathrm{neg}}$ +13: end for +14: return $\mathcal{P}_{\mathrm{pos}}$ and $\mathcal{P}_{\mathrm{neg}}$ + +$\left.c^{*}\right\rbrace_{j = 1,2,\dots,M^{\prime}}$ where $M^{\prime}\leq M$ . Now the Instance Quality of an object $g$ is defined as: + +$$ +I Q = \frac {1}{\left| \mathcal {S} ^ {\prime} \right|} \sum_ {j = 1} ^ {M ^ {\prime}} C _ {j, c ^ {*}} \cdot I o U _ {j} \tag {3} +$$ + +The $IQ$ of a GT serves as an indicator of the current model's detection capability, enabling us to capture the high-quality exemplars who have precise localization and high-confidence classification scores, and the instances confusing the model often fail to either of them. By setting appropriate threshold, we can select proper instances to build the teacher feature-set and perform the imitation process. + +Feat2Embed Module. Instead of directly measuring the similarity between different RoI features [36], we first embed these features with the simple Feat2Embed module. The input of FI branch is the region feature $\boldsymbol{x}_i \in \mathbb{R}^{H \times W \times C}$ obtained by the RoI-wise operation, e.g., RoI Align, which will be first processed by three consecutive $3 \times 3$ convolutional layers (with no padding operation) to abstract compact representations. It is worth noting that we update parameters during the extraction of current regional features + +![](images/d04745ae166de47b0b2ba865072db57ac92a8597fd2d88c781162db9b3cef9e9.jpg) +Figure 3. Size distribution of instances in (a) SODA-D and (b) SODA-A, where the absolute size corresponds to the square root of the object area. + +![](images/219a7179ae45b50039c02f5d28aab6d38c0eab04734d461cba9198129f1ea40d.jpg) + +and freeze parameters during the extraction of exemplar ones (see Feat2Embed module in Figure 2), resulting in improved stability in performance. Subsequently, the intermediate features will be mapped to the embedding space on top of a two-layer perceptron and the embedding layer with the dimension of 128, in which the dimension of hidden layers is set to 512. We have also investigated various design choices and structures for our Feat2Embed module, and detailed information can be found in the Supplementary Materials. Finally, the output of the Feature Imitation branch is defined as: + +$$ +\boldsymbol {v} _ {i} = \Theta_ {\mathrm {F I}} (\boldsymbol {x} _ {i}), \tag {4} +$$ + +where $\Theta_{\mathrm{FI}}$ denotes the parameters of Feature Imitation branch to be optimized. + +Loss Function. The objective of our FI head is simple: calculating the similarity between the RoI feature of proposal and that of the stored high-quality instances in embedding space, thereby pulling the features of those instances that confuse the model close to the exemplar ones of belonging category, while pushing apart from that of other categories and backgrounds. To this end, we propose a loss function based on Supervised Contrastive Learning [18] which extends the contrastive learning setup and allows multiple positive samples for an anchor object by exploiting the accessible label information. The loss function tailored for our FI branch is as follows: + +$$ +\mathcal {L} _ {\mathrm {F I}} = \frac {- 1}{| \mathcal {P} _ {\text {p o s}} |} \sum_ {j} \sum_ {v _ {\mathrm {p}} \in \mathcal {P} _ {\text {p o s}}} \log \frac {\exp \left(v _ {j} \cdot v _ {\mathrm {p}} / \tau\right)}{\sum_ {v _ {i} \in \mathcal {P}} \exp \left(v _ {j} \cdot v _ {i} / \tau\right)}, \tag {5} +$$ + +where $\mathcal{P} = \mathcal{P}_{\mathrm{pos}} \cup \mathcal{P}_{\mathrm{neg}}$ denotes the sample set, while $\mathcal{P}_{\mathrm{pos}}$ and $\mathcal{P}_{\mathrm{neg}}$ represent the positive and negative set respectively and they have the same cardinality ideally, and $v_{\mathrm{p}}$ and $v_{\mathrm{n}}$ are the positive and negative sample from $\mathcal{P}_{\mathrm{pos}}$ and $\mathcal{P}_{\mathrm{neg}}$ . Moreover, $j$ indexes the current proposal and $\tau$ indicates the + +
MethodPublicationScheduleAPAP50AP75APesSAPrsSAPgsSAPN
One-stage
RetianNet [26]ICCV'171×28.257.623.711.925.234.144.2
FCOS [31]ICCV'191×23.949.519.96.919.430.940.9
ATSS [46]CVPR'201×26.855.622.111.723.932.241.3
YOLOX [12]ArXiv'2170e26.753.423.013.625.130.930.4
DyHead [10]CVPR'211×27.556.123.212.424.433.041.9
Keypoint-based
CornerNet [21]ECCV'182×24.649.521.76.520.532.243.8
CenterNet [48]ArXiv'1970e21.548.815.65.116.229.642.4
RepPoints [43]ICCV'191×28.055.624.710.123.835.145.3
Query-based
Deformable-DETR [49]ICLR'2050e19.244.813.76.315.424.934.2
Sparse RCNN [30]CVPR'211×24.250.320.38.820.430.239.4
Two-stage
Baseline [29]NeurIPS'151×28.959.424.113.825.734.543.0
Cascade RPN [32]NeurIPS'191×29.156.525.912.525.535.444.7
RFLA [40]ECCV'221×29.760.225.213.226.935.444.6
CFINet (ours)-1×30.760.826.714.727.836.444.6
+ +Table 2. Comparison with state-of-the-art detection approaches on the SODA-D test-set, where 'Baseline' refers to Faster RCNN [29], serving as the baseline for the two-stage methods in the table. All the methods are trained on a ResNet-50 [17], except YOLOX (CSP-Darknet) [12] and CornerNet (HourglassNet-104) [21]. 'Schedule' denotes the number of epochs for training, in which $1 \times$ corresponds to 12 epochs and $50\mathrm{e}$ indicates 50 epochs. + +temperature which plays a crucial part in contrastive learning and needs to be well designed, and we conduct ablation studies (see Table 9) to determine the optimal setting in our framework. The total loss function is presented: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {C R P N}} + \mathcal {L} _ {\mathrm {c l s}} + \mathcal {L} _ {\mathrm {r e g}} + \alpha_ {3} \mathcal {L} _ {\mathrm {F I}}, \tag {6} +$$ + +in which $\mathcal{L}_{\mathrm{cls}}$ and $\mathcal{L}_{\mathrm{reg}}$ are the original losses of detection head, and the $\alpha_{3}$ is utilized to scale the weight of Feature Imitation part. With contrastive learning setups, not only can we fulfill the imitation learning but also prevent the collapse issue, thereby boosting the representations of small instances effectively. Moreover, the imitation process is only installed in training phase and will not slow the pace of inference. + +Training. Next we elucidate the training details of FI branch. The exemplar set $\mathcal{E} = \{\mathcal{E}_i\}_{c=1,2,\dots,N}$ containing high-quality features of $N$ foreground categories and $\mathcal{E}_i = \{\pmb{x}_{i,j}\}_{j=1,2,\dots,N_i}$ corresponding to the exemplar features of the $i$ -th class, and $N_i$ represents its size. We use $T_{\mathrm{hq}}$ to pick out those high-quality instances which are compatible to be a good exemplar, and in practice, we set a bound value to the number of high-quality predictions of an instance to filter the effect of fluctuation of the network. The function $\Gamma$ is used to augment features for high-quality instances, i.e., the positive features for a high-quality instance are the transformations of itself. The overall training procedure of FI branch is shown in Alg. 1 and more details please refer to the Supplementary Materials. + +# 4. Experiments + +# 4.1. Dataset + +To evaluate the effectiveness of our method, we perform extensive experiments on the recently released large-scale benchmark tailored for small object detection: SODA [9], including SODA-D and SODA-A. + +SODA-D. Focusing on the driving scenario, SODA-D comprises 24828 high-quality images and 278433 instances distributed across nine categories: people, rider, bicycle, motor, vehicle, traffic-sign, traffic-light, traffic-camera, and warning-cone. One of the most distinctive strengths of SODA-D is its diversity in terms of period, geographical locations, weather conditions, camera viewpoints, etc. + +SODA-A. SODA-A contains 872069 objects with oriented box annotations in 2513 aerial images and encompassing nine classes: airplane, helicopter, small-vehicle, large-vehicle, ship, container, storage-tank, swimming-pool, and windmill. The instances in SODA-A can appear in arbitrary orientations and are with significant density variations. To be specific, the average number of instances per image in SODA-A is about 350. + +As a specialized benchmark for small object detection, the instances in SODA are tiny, with most of them having an average size ranging from 10 to 30 pixels (see Figure 3). In contrast to conventional datasets for object detection, SODA includes extensive ignore annotations, aimed at filtering out instances that are either too large or are challenging to be identified deterministically due to heavy occlusion or lens flare. This procedure helps the model focus on valu + +
MethodPublicationScheduleAPAP50AP75APesSAPrsSAPgsSAPN
One-stage
Rotated RetinaNet [26]ICCV'171×26.863.416.29.122.035.428.2
S2A-Net [15]TGRS'221×28.369.613.110.222.835.829.5
Oriented RepPoints [23]CVPR'221×26.358.819.09.422.632.428.5
DHRec [27]TPAMI'221×30.168.819.810.624.640.334.6
Two-stage
Baseline [29]NeurIPS'151×32.570.124.311.927.342.234.4
Gliding Vertex [41]TPAMI'211×31.770.822.611.727.041.133.8
Oriented RCNN [39]ICCV'211×34.470.728.612.528.644.536.7
DODet [8]TGRS'221×31.668.123.411.326.341.033.5
CFINet (ours)-1×34.473.126.113.529.344.035.9
+ +![](images/06c1161da153eb4f95a2e9c739c16ec16b6d5504a7b01e07ec71e047ce779adc.jpg) +(a) Faster RCNN +Figure 4. The average number of high-quality proposals generated by (a) RPN, (b) Cascade RPN and (c) CRPN for instances in extremely Small $(eS)$ , relatively Small $(rS)$ , generally Small $(gS)$ , and Normal $(N)$ subsets, respectively. Noting that a proposal who has an IoU larger than 0.5 to any ground-truth boxes will be registered as a high-quality proposal. + +![](images/21b07048a8cfdb5503efcd5a8738b9797431028ee9a8637befbe54e87cc7df8f.jpg) +(b) Cascade RPN + +![](images/0ce2af594a95f3116ebfb7e7537487d53d9f1699b1b9c8a3158f6b09ec7ba564.jpg) +(c) CRPN + +Table 3. Comparison with state-of-the-art detection approaches on the SODA-A test-set, where 'Baseline' refers to Rotated Faster RCNN [29], serving as the baseline for the two-stage methods in the table. Other settings are consistent with Table 2. + +
Proposal MethodAR\( AR_{eS} \)\( AR_{rS} \)\( AR_{gS} \)\( AR_N \)
RPN [29]41.224.038.347.357.1
RPN-0.541.324.238.547.354.1
GA-RPN [33]42.124.139.248.956.2
Cascade RPN [32]41.822.838.248.757.1
CRPN42.624.638.949.156.9
+ +able small instances. The objects in SODA are divided into Small and Normal according to their areas, in which Small is further split into three subsets: extremely Small (eS), relatively Small (rS) and generally Small (gS). The evaluation metric of SODA follows that of COCO [24], namely averaging the precision over 10 IoU thresholds ranging from 0.5 to 0.95 (with an interval of 0.05), specifically focusing on Small objects. + +Table 4. Average Recall $(AR)$ performances of our CRPN and its counterparts on the SODA-D test-set. All the methods are with Faster RCNN (ResNet-50) as the baseline and trained for a $1\times$ schedule, in which RPN denotes the vanilla version of Faster RCNN whose positive threshold in RPN stage is set 0.7, and RPN-0.5 represents the version with 0.5 as its positive threshold of RPN. The results are tested with 300 proposals per image. + +
BaselineCRPNFIAPAPeSAPrSAPgS
✓28.913.825.734.5
✓✓30.314.327.336.1
✓✓29.514.426.335.1
✓✓✓30.714.727.836.4
+ +Table 5. Ablation analysis of our method, in which 'Baseline' denotes the vanilla Faster RCNN, CRPN and FI indicate Coarse-to-fine RPN and Feature Imitation branch, respectively. + +# 4.2. Implementation Details + +In the following experiments, unless specified, we use train-set to conduct the training and leave test-set to performance comparisons and ablation studies. Considering that the images in SODA enjoy a very high resolution $(\sim 4000\times 3000)$ , we first split the original images into a series of $800\times 800$ patches with a stride of 650, and similar to [9] these patches will be resized to $1200\times 1200$ during training and testing. All the experiments in this paper are conducted on a single RTX 3090 with the batch size of 4. Only random flip involved in data augmentation. We train all the models with a $1\times$ schedule (a bunch of 12 epochs), and the learning rate is set to 0.01 which decays after epoch 8 and epoch 11 by 0.1. The default optimizer is SGD with + +
StrategyAPAPeSAPrSAPgS
0.2029.913.726.835.9
0.4030.114.126.936.2
Ours30.314.327.336.1
+ +Table 6. Different definitions about positive anchor for CRPN, in which 'Ours' denotes the proposed dynamic strategy. + +
α3APAPeSAPrSAPgS
0.2530.614.727.636.2
0.5030.714.727.836.4
0.7530.414.027.536.2
+ +the momentum of 0.9 and the weight decay of 0.0001. We use ResNet-50 [17] with FPN [25] for all models. + +# 4.3. Main Results + +To exhibit the effectiveness of our method, we conduct a thorough comparison with current representative approaches on the SODA-D and SODA-A. + +Table 2 represents the results of our method and several mainstream approaches on the SODA-D test-set. Integrating with Faster RCNN [29], our CFINet achieves state-of-the-art performance with an overall $AP$ of $30.7\%$ , and outperforms the baseline model with $1.8\%$ points. When delving into specific metrics, our method exhibits clear predominance, particularly on the most challenging metrics $AP_{es}$ and $AP_{rs}$ . Moreover, CFINet exceeds the tailored small object detection method RFLA [40] by a significant margin (1.0% on $AP$ , 1.5% on $AP_{es}$ , and 0.9% on $AP_{rs}$ ) when both taking Faster RCNN as the baseline. Actually, RFLA sacrifices the performance on extremely Small instances (with a decrease of $0.6\%$ points when compared to Faster RCNN). + +On the SODA-A test-set, CFINet also achieves the best result and shows great advantage in comparison to other solutions especially on $AP_{eS}$ (see Table 3), indicating its superiority and generality. Furthermore, albeit exhibiting advantage on $AP_{75}$ metric and larger instances, Oriented RCNN [39] lags largely behind our approach on $AP_{50}$ (73.1% vs. 70.7%) and $AP_{eS}$ (13.5% vs. 12.5%). + +# 4.4. Effectiveness of CRPN + +One of the main designs in this paper is the Coarse-to-fine RPN, which is based on the observation that current fixed overlaps-based sampling paradigm is inappropriate for small instances due to the inherent contradictions, while the refined designs of RPN could partially reduce this barrier but still fail to satisfactory results. Here, we conduct thorough analyses to demonstrate the capability of our CRPN to generate high-quality proposals for size-limited instances. + +We first exhibit the recall performances of our CRPN and its counterparts in Table 4, from which we can see that low + +Table 7. The effect of loss weight for Feature Imitation branch. + +
ThqAPAPeSAPrSAPgS
0.5030.314.027.336.0
0.5530.614.327.336.5
0.6530.714.727.836.4
0.7030.514.627.336.3
+ +Table 8. The investigation of the criterion to be an exemplar instance. + +
τAPAPeSAPrSAPgS
0.1030.314.127.336.1
0.5030.414.427.236.2
0.6030.714.727.836.4
0.8030.214.027.336.0
+ +Table 9. The choices of temperature $\tau$ to the final performance. + +ering the positive threshold slightly improves the average recall while sacrificing the performance of larger instances $(AR_{N}$ experiences a sharp decline from $57.1\%$ to $54.1\%$ ). GA-RPN [33] as well as Cascade RPN [32] both fail to better results since their patterns incline to large instances as we discussed before. In comparison with RPN and its variants, our CRPN demonstrates superior performance on objects in the Small, while achieving comparable results on Normal instances. This validates our assumption that refined proposal networks tend to favor larger objects. + +We conjecture that one of the most challenging issues towards accurate small object detection is the scarcity of high-quality samples, which is also the major motivation behind the design of CRPN. Hence, we intuitively compare the baseline RPN, Cascade RPN and our CRPN about the number of high-quality samples. In Figure 4, our CRPN generates more high-quality proposals compared to the other competitors. Interestingly, CRPN can dynamically shift the focus along the training: at the beginning, the model concentrates more on large objects which are conducive for early optimization while as the training goes, the model gradually shifts its attention to objects with small sizes that are usually not handled well before. This is interpretable since the instances having extremely limited sizes are with more uncertainties and fitting them in early-phase is not an optimal choice for the detector. + +# 4.5. Ablation and Discussion + +In this part, we conduct ablation studies as well as comprehensive discussions to attest the importance of the CRPN and FI branch, and moreover, determine the appropriate settings of our approach. All the experiments of this section are conducted on the SODA-D test-set. + +Investigation of Designed Components. We first perform ablation experiments to verify the effectiveness of two modules. As in Table 5, our CRPN and FI can both improve the performances steadily, while the introduction of feature imitation branch is conducive for the recognition of size- + +![](images/e5d4b8aa12b8a76b267c49ca40f941f184c53f1520ec2603aae038f11f9123fb.jpg) + +![](images/f768988861a961f0b9e6c85feb36fda818ebc3bee8c8b0b8db68e1cdb90500cd.jpg) +people + +![](images/341b2c87ca4f9ad8262dd4f74f21cdf09ea19eb8bc6b0f232d789b2cd91bfdba.jpg) +rider +Figure 5. Qualitative results of our method on the SODA-D test-set. Only predictions with confidence scores larger than 0.3 are demonstrated and the masked bounding boxes represent ignore regions. Best viewed in color and zoom-in windows. + +![](images/cf692c67adc8cde0f975c194db3521c7837419b8592302fb8aae9ddde520cbff.jpg) +bicycle + +![](images/842dc621bb1adaf4d03d7025959f6314ade178f76812e41710e44c553c5d894f.jpg) +motor + +![](images/32ab6006046aa2afdf1851f67ff4746a3a31ce59b16cdb41bd73b2ce4a949798.jpg) +vehicle + +![](images/afeb7780a1a48b43ac48917ec4dce1e80a3f1b3e5b522e08444792753f8130e9.jpg) + +![](images/216e7c7af19ce4309f803c82af3de7c8297cef2022bc2fb70f19bc73adf83f76.jpg) +c-sign + +![](images/4c16ae33eac43f790dfeb505957673672182147820f96898ac3c99bcb458eb3e.jpg) +traffic-light + +![](images/90c053259362cf6f50d580ac9723840ce8543c177c5a7171e5cc6c6547c5520c.jpg) +traffic-camera + +![](images/e8212f3ddfaa25b94322f9cf2cae84358d98ee9c57b1f692ae14be81d7e658ec.jpg) +ing-cone + +limited instances. Consistently, the integration of FI and CRPN achieves the best result, as CRPN is capable of generating sufficient high-quality proposals (as shown in Figure 4) thereby offering a more accurate indication of instance quality and the potential to act as an exemplar. + +Fixed or Dynamic. A natural idea is directly setting fixed positive threshold to obtain more anchors for one-stage regression of CRPN. Here we show that our simple-yet-effective area-based anchor mining strategy can achieve the best performance. In Table 6, when the IoU threshold of a positive potential anchor for first-stage regression drops to 0.20, the $AP_{eS}$ is only $13.7\%$ and this could be attributed to low-quality samples and the predominance of large instances as discussed before. The proposed area-based anchor mining strategy could mitigate this problem and obtain the best overall accuracy. + +Weights of Feature Imitation Loss. In this section, we analyze the impact of the weight parameter of FI branch (namely the hyper-parameter $\alpha_{3}$ ) on the model. As shown in Table 7, paying too little or excessive attention both deteriorates the final performance, hence we set $\alpha_{3}$ to 0.5 in our experiments to ensure overall accuracy. + +The Criterion of Being An Exemplar. The quality of the exemplar plays a pivotal role in the imitation process [19, 36]. Next we discuss the choices for capturing a high-quality exemplar to build the feature set. In Table 8, the lower $T_{\mathrm{hq}}$ involves more exemplars and updates the teacher set more frequently while the higher $T_{\mathrm{hq}}$ does exactly the opposite. It can be seen that increasing the $T_{\mathrm{hq}}$ from 0.5 to 0.65 could facilitate the imitation process and when the $T_{\mathrm{hq}}$ reaches 0.70, the overall $AP$ drops instead. This may originates that the earliest samples stored in the feature set are inappropriate for current state, because the optimization is dynamic and the model is evolving hence the criterion of + +being an exemplar has already changed. + +The Temperature. The temperature is paramount for contrastive learning [18] and we conduct a series of experiments to verify the best choice for $\tau$ . From Table 9, when $\tau$ ranges from 0.10 to 0.80, the overall performance increases first and then decreases to $30.2\%$ , therefore we choose 0.6 for our method. + +Visualization. We demonstrate the visualization results of example images from SODA-D test-set in Figure 5 to intuitively show the capability of our detector when detecting the small instances. + +# 5. Conclusion + +In this paper, we proposed CFINet, a two-stage detector based on the Coarse-to-fine Region Proposal Network and Feature Imitation setups, in which the former can produce sufficient high-quality proposals for small instances particularly for those with extremely limited sizes. Then the novel detection head on top of the feature imitation branch facilitates the representations of small objects posing challenges to the model under the contrastive learning paradigm. The experiments results show that our method achieves state-of-the-art performance on the large-scale small object detection datasets SODA-D and SODA-A. In the future, a more flexible and general indicator of instance quality is worth investigating. + +# Acknowledgments + +This work was supported in part by the National Science Foundation of China under Grant 62136007 and Grant U20B2068, and in part by the Natural Science Basic Research Program of Shaanxi under Grants 2021JC-16 and 2023-JC-ZD-36. + +# References + +[1] Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. Finding tiny faces in the wild with generative adversarial network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 21-30, 2018. +[2] Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. Sod-mtgan: Small object detection via multi-task generative adversarial network. In Proceedings of the European Conference on Computer Vision, pages 206-221, 2018. +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European Conference on Computer Vision, pages 213-229, 2020. +[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Cconference on Machine Learning, pages 1597-1607, 2020. +[5] Gong Cheng, Junwei Han, Peicheng Zhou, and Dong Xu. Learning rotation-invariant and fisher discriminative convolutional neural networks for object detection. IEEE Transactions on Image Processing, 28(1):265-278, 2019. +[6] Gong Cheng, Qingyang Li, Guangxing Wang, Xingxing Xie, Lingtong Min, and Junwei Han. Sfrnet: Fine-grained oriented object recognition via separate feature refinement. IEEE Transactions on Geoscience and Remote Sensing, 61:1-10, 2023. +[7] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. +[8] Gong Cheng, Yanqing Yao, Shengyang Li, Ke Li, Xingxing Xie, Jiabao Wang, Xiwen Yao, and Junwei Han. Dual-aligned oriented detector. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. +[9] Gong Cheng, Xiang Yuan, Xiwen Yao, Kebing Yan, Qinghua Zeng, Xingxing Xie, and Junwei Han. Towards large-scale small object detection: Survey and benchmarks. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-20, 2023. +[10] Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unifying object detection heads with attentions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7373-7382, 2021. +[11] Chunfang Deng, Mengmeng Wang, Liang Liu, Yong Liu, and Yunliang Jiang. Extended feature pyramid network for small object detection. IEEE Transactions on Multimedia, 24:1968-1979, 2021. +[12] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. +[13] Spyros Gidaris and Nikos Komodakis. Attend refine repeat: Active box proposal generation via in-out localization. arXiv preprint arXiv:1606.04446, 2016. + +[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. +[15] Jiaming Han, Jian Ding, Jie Li, and Gui-Song Xia. Align deep features for oriented object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2021. +[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. +[18] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. +[19] Jung Uk Kim, Sungjune Park, and Yong Man Ro. Robust small-scale pedestrian detection with cued recall via memory learning. In Proceedings of the IEEE International Conference on Computer Vision, pages 3050-3059, 2021. +[20] Chunbo Lang, Gong Cheng, Binfei Tu, Chao Li, and Junwei Han. Base and meta: A new perspective on few-shot segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2023. +[21] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In ECCV, pages 734-750, 2018. +[22] Jianan Li, Xiaodan Liang, Yunchao Wei, Tingfa Xu, Jiashi Feng, and Shuicheng Yan. Perceptual generative adversarial networks for small object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1222-1230, 2017. +[23] Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented reploints for aerial object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1829-1838, 2022. +[24] Tsung Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, pages 740-755, 2014. +[25] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2117-2125, 2017. +[26] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. pages 2980-2988, 2017. +[27] Guangtao Nie and Hua Huang. Multi-oriented object detection in aerial images with double horizontal rectangles. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4932-4944, 2022. + +[28] Junhyug Noh, Wonho Bae, Wonhee Lee, Jinhwan Seo, and Gunhee Kim. Better to follow, follow to be better: Towards precise supervision of feature super-resolution for small object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 9725-9734, 2019. +[29] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28, 2015. +[30] Peize Sun et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 14449-14458, 2021. +[31] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. pages 9627-9636, 2019. +[32] Thang Vu, Hyunjun Jang, Trung X Pham, and Chang Yoo. Cascade rpn: Delving into high-quality region proposal network with adaptive convolution. Advances in Neural Information Processing Systems, 32, 2019. +[33] Jiaqi Wang, Kai Chen, Shuo Yang, Chen Change Loy, and Dahua Lin. Region proposal by guided anchoring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2965-2974, 2019. +[34] Jinwang Wang, Wen Yang, Haowen Guo, Ruixiang Zhang, and Gui-Song Xia. Tiny object detection in aerial images. In IEEE International Conference on Pattern Recognition, pages 3791-3798, 2021. +[35] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 7303-7313, 2021. +[36] Jialian Wu, Chunluan Zhou, Qian Zhang, Ming Yang, and Junsong Yuan. Self-mimic learning for small-scale pedestrian detection. In Proceedings of the ACM International Conference on Multimedia, pages 2012-2020, 2020. +[37] Wei Wu, Hao Chang, Yonghua Zheng, Zhu Li, Zhiwen Chen, and Ziheng Zhang. Contrastive learning-based robust object detection under smoky conditions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4295-4302, 2022. +[38] Enze Xie, Jian Ding, Wenhai Wang, Xiaohang Zhan, Hang Xu, Peize Sun, Zhenguo Li, and Ping Luo. Detco: Unsupervised contrastive learning for object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 8392-8401, 2021. +[39] Xingxing Xie, Gong Cheng, Jiabao Wang, Xiwen Yao, and Junwei Han. Oriented r-cnn for object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 3520-3529, 2021. +[40] Chang Xu, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Rfla: Gaussian receptive based label assignment for tiny object detection. In Proceedings of the European Conference on Computer Vision, 2022. +[41] Yongchao Xu, Mingtao Fu, Qimeng Wang, Yukang Wang, Kai Chen, Gui-Song Xia, and Xiang Bai. Gliding vertex on + +the horizontal bounding box for multi-oriented object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(4):1452-1459, 2020. +[42] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. Querydet: Cascaded sparse query for accelerating high-resolution small object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 13668-13677, 2022. +[43] Ze Yang, Shaohui Liu, Han Hu, Liwei Wang, and Stephen Lin. Repoints: Point set representation for object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 9657-9666, 2019. +[44] Jiahui Yu, Yuning Jiang, Zhangyang Wang, Zhimin Cao, and Thomas Huang. Unitbox: An advanced object detection network. In Proceedings of the ACM International Conference on Multimedia, pages 516-520, 2016. +[45] Xuehui Yu, Yuqi Gong, Nan Jiang, Qixiang Ye, and Zhenjun Han. Scale match for tiny person detection. In IEEE Conference on Winter Conference on Applications of Computer Vision, pages 1257-1265, 2020. +[46] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9759-9768, 2020. +[47] Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, and Stan Z Li. S3fd: Single shot scale-invariant face detector. In Proceedings of the IEEE International Conference on Computer Vision, pages 192-201, 2017. +[48] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. +[49] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2020. \ No newline at end of file diff --git a/2023/Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning/images.zip b/2023/Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3a8b1263986cf618be3efefc2efc21855a44a7a9 --- /dev/null +++ b/2023/Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52229e4fabd75330c05758da2cffdfd0d55477bb689a39143b7e9d53c3c18d80 +size 694444 diff --git a/2023/Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning/layout.json b/2023/Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..36988a668a5ea9455c6f71419b50f299862a594a --- /dev/null +++ b/2023/Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning/layout.json @@ -0,0 +1,10619 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 491, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 491, + 140 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 491, + 140 + ], + "type": "text", + "content": "Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 161, + 510, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 161, + 510, + 204 + ], + "spans": [ + { + "bbox": [ + 82, + 161, + 510, + 204 + ], + "type": "text", + "content": "Xiang Yuan Gong Cheng * Kebing Yan Qinghua Zeng Junwei Han School of Automation, Northwestern Polytechnical University, Xi'an, China {shaunyuan, kebingyan, zengqinghua}@mail.nwpu.edu.cn, {gcheng, jhan}@nwpu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 256, + 290, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 256, + 290, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 256, + 290, + 555 + ], + "type": "text", + "content": "The past few years have witnessed the immense success of object detection, while current excellent detectors struggle on tackling size-limited instances. Concretely, the well-known challenge of low overlaps between the priors and object regions leads to a constrained sample pool for optimization, and the paucity of discriminative information further aggravates the recognition. To alleviate the aforementioned issues, we propose CFINet, a two-stage framework tailored for small object detection based on the Coarse-to-fine pipeline and Feature Imitation learning. Firstly, we introduce Coarse-to-fine RPN (CRPN) to ensure sufficient and high-quality proposals for small objects through the dynamic anchor selection strategy and cascade regression. Then, we equip the conventional detection head with a Feature Imitation (FI) branch to facilitate the region representations of size-limited instances that perplex the model in an imitation manner. Moreover, an auxiliary imitation loss following supervised contrastive learning paradigm is devised to optimize this branch. When integrated with Faster RCNN, CFINet achieves state-of-the-art performance on the large-scale small object detection benchmarks, SODA-D and SODA-A, underscoring its superiority over baseline detector and other mainstream detection approaches. Code is available at https://github.com/shaunyuan22/CFINet." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 586, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 287, + 659 + ], + "type": "text", + "content": "Small object detection (SOD) aims to classify and localize the instances with limited regions, which plays an important role in a wide range of scenarios, such as pedestrian detection, autonomous driving, and intelligent surveillance understanding, to name a few [42, 28, 22, 19, 36, 2, 20]. Compared to the generic object detection which has been" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 230, + 392, + 350 + ], + "blocks": [ + { + "bbox": [ + 308, + 230, + 392, + 350 + ], + "lines": [ + { + "bbox": [ + 308, + 230, + 392, + 350 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 392, + 350 + ], + "type": "image", + "image_path": "c37fa509a0e67644577a7ce2fba591e7ecec0af059b50881524a43feb10d684c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 352, + 547, + 430 + ], + "lines": [ + { + "bbox": [ + 304, + 352, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 547, + 430 + ], + "type": "text", + "content": "Figure 1. Distribution of maximum IoU of anchors matched to each ground-truth instance in SODA-D [9] train-set, where extremely Small (eS), relatively Small (rS), and generally Small (gS) correspond to three area subsets in SODA [9] with the ranges (0, 144], (144, 400] and (400, 1024]. The smaller the objects are, the lower IoU the matched anchors have, hence the commonly used positive IoU threshold (0.7) is too rigorous for small objects." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 392, + 230, + 470, + 350 + ], + "blocks": [ + { + "bbox": [ + 392, + 230, + 470, + 350 + ], + "lines": [ + { + "bbox": [ + 392, + 230, + 470, + 350 + ], + "spans": [ + { + "bbox": [ + 392, + 230, + 470, + 350 + ], + "type": "image", + "image_path": "c95f8e6f6e719c3c11a981aa942d2e24e7e5f5292ce20f61c636f77ca6ff9040.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 470, + 231, + 545, + 350 + ], + "blocks": [ + { + "bbox": [ + 470, + 231, + 545, + 350 + ], + "lines": [ + { + "bbox": [ + 470, + 231, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 470, + 231, + 545, + 350 + ], + "type": "image", + "image_path": "9307cc36ff082cb685b821386ade2a48547d27db4f897e62781ae5052b023836.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "type": "text", + "content": "extensively studied, SOD task receives relatively little attention and good solutions are still scarce so far. Moreover, generic detectors [29, 6, 25, 31, 7, 3, 26, 5] usually struggle on handling small objects due to two inherent challenges: the insufficient and low-quality samples for training and the uncertain prediction of RoIs (Region of Interests)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": "First, current prevailing detectors exploit either overlap-based [29, 26] or distance-based [31] strategies to select the positive priors of objects for training. However, small instances usually occupy an extremely limited area, therefore the region overlaps between densely arranged anchors and ground truth boxes are significantly small and far from the commonly used positive IoU (Intersection-over-Union) threshold, as in Figure 1. In other words, the existing positive sample criterion is overly stringent when applied to small/tiny objects, resulting in a restricted number of samples available for optimization. An intuitive approach involves reducing the threshold for defining a positive sample [47]. However, while this can lead to an increase in the number of positive samples, it often at the expense of overall sample quality, in which low-quality samples disrupt optimization and incur a trivial regression solution. Worse still, this is actually contradictory to the purpose of proposal" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 665, + 211, + 674 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 665, + 211, + 674 + ], + "spans": [ + { + "bbox": [ + 58, + 665, + 211, + 674 + ], + "type": "text", + "content": "*Corresponding author: gcheng@nwpu.edu.cn" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 674, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 674, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 674, + 287, + 712 + ], + "type": "text", + "content": "This paper focuses on the detection of \"pure\" small objects, where the scales of all the objects are distributed within a relatively tight range [45, 34, 9], which is distinct from the known small objects in multi-scale object detection [24]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6317" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 120 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 120 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 120 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 120 + ], + "type": "table", + "html": "
ctrl/ign ratioAPAP50AP75APesSAPrsSAPgSAPN
Baseline28.959.424.113.825.734.543.0
0.2/0.529.156.525.912.525.535.444.7
0.5/0.829.557.826.013.526.235.845.0
0.8/1.027.554.124.311.223.833.742.5
", + "image_path": "f63e2c88f04a097accad854273716ee9dda513183982d9b818b2a52a3e4b7d47.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 128, + 286, + 183 + ], + "lines": [ + { + "bbox": [ + 47, + 128, + 286, + 183 + ], + "spans": [ + { + "bbox": [ + 47, + 128, + 286, + 183 + ], + "type": "text", + "content": "Table 1. The performances of Cascade RPN [32] compared to the baseline (a vanilla Faster RCNN [29]). The ctr/ign ratio denotes the sampling region in first regression stage of Cascade RPN. The results are tested on the SODA-D [9] test-set and with a ResNet-50 [17] as the backbone." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 194, + 286, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 194, + 286, + 218 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 286, + 218 + ], + "type": "text", + "content": "network, i.e., guaranteeing the recall and ease the burden of subsequent work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 220, + 286, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 220, + 286, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 220, + 286, + 506 + ], + "type": "text", + "content": "To sum up, current prevailing priors-to-proposals paradigms that heavily depend on the overlap or distance metric have inherent limitations in detecting small objects, and nowadays devised assignment or sampling schemes contribute minimally to this problem [9, 40]. Since the proposals play such a crucial role in two-stage detectors, so how about the improved Region Proposal Network (RPN) variants [13, 32, 33] meet small object detection? Following this line, we take Cascade RPN [32], one of the most superior proposal network for generic object detection, to perform preliminary experiments and the results are shown in Table 1. While the auxiliary regression phase provides refined priors with better initialization for subsequent regression, the final results remain somewhat unsatisfactory. Specifically, the improvement mainly comes from the larger objects and the " + }, + { + "bbox": [ + 47, + 220, + 286, + 506 + ], + "type": "inline_equation", + "content": "AP_{eS}" + }, + { + "bbox": [ + 47, + 220, + 286, + 506 + ], + "type": "text", + "content": " as well as " + }, + { + "bbox": [ + 47, + 220, + 286, + 506 + ], + "type": "inline_equation", + "content": "AP_{rS}" + }, + { + "bbox": [ + 47, + 220, + 286, + 506 + ], + "type": "text", + "content": " actually decrease significantly instead, indicating that the region-based sampling strategy is inclined to large instances which further dominate the proposal network. Meanwhile, enlarging the sample region contributes little (even negative) to this condition (see the bottom row in Table 1). Therefore, the coarse-to-fine pipeline has the potential to surmount the barrier of conventional prior-to-proposal paradigm, but the crux lies in dedicating sufficient attention to small instances." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 510, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 286, + 713 + ], + "type": "text", + "content": "Second, small objects usually lack discriminative information and distorted structures, leading the inclination of model to give ambiguous even incorrect predictions [1, 11]. Meanwhile, there are a certain amount of large instances embodying clear visual cues and better discrimination. Building upon this observation, several works proposed to bridge the representation gap between small objects and large ones, and most of them [2, 22, 28, 1] rely on Generative Adversarial Network (GAN) [14] or similarity learning [19, 36] to super-resolve/restore the features of size-limited instances under the guidance of large ones that are deemed to be visually authentic. However, these approaches overlook the fact: high quality " + }, + { + "bbox": [ + 47, + 510, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\neq" + }, + { + "bbox": [ + 47, + 510, + 286, + 713 + ], + "type": "text", + "content": " large size meanwhile small size " + }, + { + "bbox": [ + 47, + 510, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\neq" + }, + { + "bbox": [ + 47, + 510, + 286, + 713 + ], + "type": "text", + "content": " low quality. In other words, the criterion for humans and the model to decide whether a sample is competent to be a good example is distinct. For the latter, it is dynamic and should be adjusted according to" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "the current optimization of the detector. Moreover, efforts in this line have to resort to sophisticated training strategies or additional models, which is time-consuming and break the conventional end-to-end paradigm." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 121, + 545, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 545, + 322 + ], + "type": "text", + "content": "Putting the above parts together, we propose a two-stage small object detector CFINet based on the coarse-to-fine pipeline and feature imitation learning. Concretely, enlightened by the multi-stage proposal generation scheme in Cascade RPN, we devise Coarse-to-fine RPN (CRPN). It firstly employs an dynamic anchor selection strategy to mine potential priors to conduct coarse regression, and henceforth, these refined anchors will be classified and regressed by the region proposal network. In addition, we extend the conventional classification-and-regression setting with an auxiliary Feature Imitation (FI) branch, which can leverage the regional features of high-quality instances to guide the learning of those objects with uncertain/mistaken predictions, and a loss function based on the Supervised Contrastive Learning (SCL) [18] is designed to optimize the whole process. The main contributions of this paper are summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 329, + 545, + 498 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 317, + 329, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 329, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 317, + 329, + 545, + 388 + ], + "type": "text", + "content": "- A coarse-to-fine proposal generation pipeline named CRPN was built to perform anchor-to-proposal procedure, where an area-based anchor mining strategy and cascade regression empower the high-quality proposals for small instances." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 396, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 396, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 317, + 396, + 545, + 456 + ], + "type": "text", + "content": "- An auxiliary Feature Imitation (FI) branch was introduced to enrich the representations of low-quality instances perplexing the model under the supervision of high-quality instances, and this novel branch is optimized by a tailored loss function based on SCL." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 462, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 462, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 462, + 545, + 498 + ], + "type": "text", + "content": "- The experiment results on the SODA-D and SODA-A datasets exhibit the superiority of our CFINet to detect these instances with extremely limited sizes." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 514, + 397, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 397, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 397, + 526 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "Anchor Refinement and Region Proposals. Two-stage anchor-based approaches heavily depend on the high-quality proposals [33, 32]. Towards this goal, RPN was first introduced in Faster RCNN [29] to produce proposals in a fully convolutional network, and this simple yet effective design facilitates end-to-end model optimization. Following RPN, [13] proposed to iteratively regress the predefined anchors. GA-RPN [33] discards the uniform anchoring strategy and formulates the anchor generation in two steps: first determining the locations which may contain objects and on which the anchor scales are predicted then. By installing a multi-stage anchor-to-proposal strategy and alleviating the misalignment between the refined anchors and image features, Cascade RPN [32] enables high-quality proposal generation. Unfortunately, current proposal-oriented" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6318" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 68, + 302, + 278 + ], + "blocks": [ + { + "bbox": [ + 49, + 68, + 302, + 278 + ], + "lines": [ + { + "bbox": [ + 49, + 68, + 302, + 278 + ], + "spans": [ + { + "bbox": [ + 49, + 68, + 302, + 278 + ], + "type": "image", + "image_path": "70bd0e14cf5547c9f29db918e444f9f99a5355b3bfc96bfb9fd021ab91b8c27e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 279, + 545, + 378 + ], + "lines": [ + { + "bbox": [ + 46, + 279, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 545, + 378 + ], + "type": "text", + "content": "Figure 2. The overall architecture of CFINet. In Coarse-to-fine RPN (CRPN), the area-based Anchor Mining strategy ensures sufficient candidates for instances of various sizes (small: orange boxes, large: green boxes) based on the dynamic pos threshold, which will be then used to obtain the coarse proposals. After that, the alignment between coarse proposals and corresponding features is enabled by the Adaptive Convolution before feeding into the RPN to produce high-quality proposals (blue boxes). The Feature Imitation (FI) branch is devised to facilitate the representations of small instances, in which the RoI features of uncertain/mistaken predictions will be pulled to their counterparts of exemplar feature set in embedding space (throughout the Feat2Embed module), while pushed apart from the exemplar features of other categories and background. And the exemplar features are collected based on the model predictions using the proposed quality indicator, i.e., Instance Quality " + }, + { + "bbox": [ + 46, + 279, + 545, + 378 + ], + "type": "inline_equation", + "content": "(IQ)" + }, + { + "bbox": [ + 46, + 279, + 545, + 378 + ], + "type": "text", + "content": ". We tailor a Feature Imitation loss function " + }, + { + "bbox": [ + 46, + 279, + 545, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FI}}" + }, + { + "bbox": [ + 46, + 279, + 545, + 378 + ], + "type": "text", + "content": " to optimize this auxiliary branch. Note that we only exhibit single-level Feature Pyramid Network (FPN) [25] feature for clear illustration." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 304, + 68, + 558, + 277 + ], + "blocks": [ + { + "bbox": [ + 304, + 68, + 558, + 277 + ], + "lines": [ + { + "bbox": [ + 304, + 68, + 558, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 68, + 558, + 277 + ], + "type": "image", + "image_path": "e11300be834b1610a6e82ed6ecf56fbdfbff87f116a3d2dd9121205f9c4315a8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 398, + 286, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 286, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 286, + 493 + ], + "type": "text", + "content": "frameworks fail to produce high-quality proposals for instances with limited regions, and the root cause lies in the notoriously low overlaps between the objects and priors [40]. Different from the above methods, our coarse-to-fine proposal pipeline could exploit the potential of multi-stage refinement paradigm, thereby guaranteeing both the quantity and quality of proposals for instances with extremely limited sizes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 286, + 712 + ], + "type": "text", + "content": "Feature Imitation for Small Object Detection. One of the major challenges to detect small objects is the low-quality representation [45, 40, 9], while large instances often with clear structures and discriminative features. Hence, a series of efforts have been made to boost the semantic representations of small/tiny instances by mining the intrinsic correlations between small and large objects. Based on the generative adversarial paradigm, Perceptual GAN [22] designs a generator that is optimized to produce high-quality representations of small instances to fool the subsequent discriminator. Bai et al. [1] devised a novel pipeline to restore a clear face from the inputting blurry one. Noh et al. [28] further introduced precise supervision for the super-resolution process of small objects. Moreover, Wu et al. [36] and Kim et al. [19] both exploited similarity learning to force the features of small-scale pedestrians close to that of the large-scale ones which are obtained by an additional model. The existence of super-resolution branch or offline" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 398, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 545, + 445 + ], + "type": "text", + "content": "feature bank hampers the end-to-end optimization while our method updates the exemplar features in an online fashion, which guarantees the diversity of high-quality feature sets thereby getting rid of the collapse issue." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 448, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 545, + 591 + ], + "type": "text", + "content": "Contrastive Learning for Object Detection. The recent explosion of self-supervised learning mainly comes from its Contrastive Learning fashion, and several works have extended this paradigm into detection fields. Detco [38] is an effective self-supervised framework for object detection which utilizes the image and its local patches to conduct contrastive learning. Wu et al. [37] applied contrastive learning to object detection under smoky conditions. Though contrastive learning has recently received considerable interests [4, 16, 35], the potential of utilizing contrastive learning for better representation of small objects has not yet been investigated to date." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 607, + 385, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 607, + 385, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 607, + 385, + 620 + ], + "type": "text", + "content": "3. Our Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "This section presents the details about CFINet. We start with a discussion about the inherent limitations of Cascade RPN when confronting small objects, then our coarse-to-fine high-quality proposal generation pipeline tailored for size-limited instances is introduced. Afterwards, we elucidate the architecture of newly designed Feature Imitation branch, also with the optimization and training procedure." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6319" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 273, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 273, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 273, + 84 + ], + "type": "text", + "content": "The overall architecture of CFINet is shown in Figure 2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 92, + 190, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 92, + 190, + 105 + ], + "spans": [ + { + "bbox": [ + 47, + 92, + 190, + 105 + ], + "type": "text", + "content": "3.1. Towards Better Proposals" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 111, + 287, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 111, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 46, + 111, + 287, + 290 + ], + "type": "text", + "content": "Limitations of Cascade RPN. High-quality proposals play a pivotal role in two-stage detectors, but need heuristic anchor settings. Casdade RPN [32] discards this conventional setup by placing one single anchor on each feature point and conducting multi-stage refinement. Though exhibiting superior performance on objects of general scales, Cascade RPN fails to tackle extremely small objects well due to its inherent limitations. Concretely, the distance metric used in first-stage regression cannot guarantee sufficient potential anchors for small objects who have significantly small center region. Moreover, Cascade RPN only marks eligible anchors on a single pyramid level as positive, while this heuristic scheme simply discards those possible anchors at other levels which can still convey the existence and rough location information of small objects [42]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 291, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 287, + 399 + ], + "type": "text", + "content": "Coarse-to-fine RPN. To remedy the aforementioned issues of Cascade RPN when handling small instances, we propose Coarse-to-fine RPN and the detailed structure is in Figure 2. First, we design an area-based anchor selection strategy to enable the instances of various sizes could have (relatively) adequate potential anchors. Concretely, for an object box with the width " + }, + { + "bbox": [ + 46, + 291, + 287, + 399 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 291, + 287, + 399 + ], + "type": "text", + "content": " and the height " + }, + { + "bbox": [ + 46, + 291, + 287, + 399 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 291, + 287, + 399 + ], + "type": "text", + "content": ", any anchors who have an IoU larger than " + }, + { + "bbox": [ + 46, + 291, + 287, + 399 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{a}}" + }, + { + "bbox": [ + 46, + 291, + 287, + 399 + ], + "type": "text", + "content": " will be regarded as positive for the coarse regression, and " + }, + { + "bbox": [ + 46, + 291, + 287, + 399 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{a}}" + }, + { + "bbox": [ + 46, + 291, + 287, + 399 + ], + "type": "text", + "content": " is formulated as follow:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 407, + 287, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 407, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 83, + 407, + 287, + 434 + ], + "type": "interline_equation", + "content": "T _ {\\mathrm {a}} = \\max (0. 2 5, 0. 2 0 + \\gamma \\cdot \\log \\frac {\\sqrt {w \\cdot h}}{1 2}), \\tag {1}", + "image_path": "1b6a5d7ba04669ae9c0897bab2ff9aefcfd15ac6a7191a0e2565d4a06b44cbc0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 569 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 437, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 437, + 287, + 569 + ], + "type": "text", + "content": " denotes a scale factor and is set to default 0.15 in our experiments, and the term 12 actually corresponds to the minimal area definition of SODA dataset [9], which enables adequate samples for extreme-size objects and can be tuned for different datasets. Moreover, " + }, + { + "bbox": [ + 46, + 437, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 437, + 287, + 569 + ], + "type": "text", + "content": " and max operation keep the optimization from being overwhelmed by the low-quality priors. Taking IoU as the criterion to mine potential anchors, the optimization inconsistency in multi-stage regression of Cascade RPN can be averted. Meanwhile, the model determines the positive sample in a more smooth way on top of the proposed continuous threshold." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": "Distinct from Cascade RPN, we preserve anchors of all Feature Pyramid Network (FPN) [25] levels " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{P_2,P_3,P_4,P_5\\}" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": " to perform first-stage regression. In this way, we could mine sufficient potential anchors for extremely small instances and meanwhile, larger instances still can obtain proper attention since the anchors matched to them have naturally higher IoUs, as discussed in Figure 1. After the first-stage regression, we then capture the offsets inside the regressed boxes and input them with the feature maps to RPN, in which the Adaptive Convolution [32] will be exploited to align the features and conduct second-stage regression and foreground-background classification." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 529, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 529, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 529, + 84 + ], + "type": "text", + "content": "Loss Function. The training objective of our CRPN is:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 347, + 95, + 545, + 110 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 95, + 545, + 110 + ], + "spans": [ + { + "bbox": [ + 347, + 95, + 545, + 110 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C R P N}} = \\alpha_ {1} \\left(\\mathcal {L} _ {\\mathrm {r e g}} ^ {\\mathrm {c}} + \\mathcal {L} _ {\\mathrm {r e g}} ^ {\\mathrm {f}}\\right) + \\alpha_ {2} \\mathcal {L} _ {\\mathrm {c l s}}, \\tag {2}", + "image_path": "d8852bda313d073d73fbeb3f7fb7b2e72559eda345d8042f319c2c95d24b4fcd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 114, + 545, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 114, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 545, + 174 + ], + "type": "text", + "content": "where we use cross-entropy loss and IoU Loss [44] as " + }, + { + "bbox": [ + 304, + 114, + 545, + 174 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 304, + 114, + 545, + 174 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 114, + 545, + 174 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 304, + 114, + 545, + 174 + ], + "type": "text", + "content": ", respectively. The c and f in Eq. (2) indicate the coarse-stage and fine-stage in our CRPN, and noting that we only do classification in the latter stage. The loss weights " + }, + { + "bbox": [ + 304, + 114, + 545, + 174 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 304, + 114, + 545, + 174 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 114, + 545, + 174 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 304, + 114, + 545, + 174 + ], + "type": "text", + "content": " are set to 9.0 and 0.9, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 181, + 538, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 181, + 538, + 193 + ], + "spans": [ + { + "bbox": [ + 305, + 181, + 538, + 193 + ], + "type": "text", + "content": "3.2. Feature Imitation for Small Object Detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 198, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 389 + ], + "type": "text", + "content": "Efforts on exploiting the intrinsic correlations between objects of different scales to boost the representations of small objects have been made, but most of them fail to the effectiveness and diversity. Specifically, the majority of previous methods [22, 2, 1, 28] resort to GAN to super-resolve the representations of small instances. This calls for the sophisticated training schemes and is prone to fabricate fake textures and artifacts [11]. Another line of efforts turn to the similarity learning which either has to construct offline feature bank in a cumbersome way [19], or directly leverages " + }, + { + "bbox": [ + 304, + 198, + 545, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 304, + 198, + 545, + 389 + ], + "type": "text", + "content": " norm to similarity measurement between different RoI features [36], potentially leading the feature collapse issue: the region features after amending could have high-similarity but lost their own characteristics. This homogenization in feature space actually impairs the generality and robustness of the model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 391, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 544 + ], + "type": "text", + "content": "To mitigate the collapse risks and avoid the memory burden as well as enable the end-to-end optimization, we devise a Feature Imitation (FI) head (see Figure 2). Most importantly, instead of solely taking large-scale objects as the guidance of this procedure, we consider the model response in current state for each instance, thereby constructing a dynamic and currently optimized feature bank of proper exemplars in an online fashion. The FI branch mainly composes an Exemplar Feature Set and a Feature-to-Embedding (Feat2Embed) module, where the former reserves the RoI features of high-quality exemplars and the latter projects the input to the embedding space. Next we elucidate the details about our Feature Imitation branch." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "text", + "content": "What is a proper exemplar? As we discussed above, an exemplar is vital in the imitation learning. To determine the most representative/proper/high-quality examples which can deliver authentic guidance/supervision for small objects confusing the model at this moment, we first introduce a simple quality indicator for an instance. Given a ground-truth (GT) object " + }, + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "inline_equation", + "content": "g = (c^{*}, b^{*})" + }, + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "inline_equation", + "content": "c^{*}" + }, + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "inline_equation", + "content": "b^{*}" + }, + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "text", + "content": " denote its label and bounding box coordinates. Assuming the detection head outputs a prediction set " + }, + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "inline_equation", + "content": "S = \\{C_{i}, IoU_{i}\\}_{i=1,2,\\dots,M}" + }, + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "text", + "content": ", in which " + }, + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "inline_equation", + "content": "C_{i} \\in \\mathbb{R}^{N+1}" + }, + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "text", + "content": " indicates the predicted classification vector and " + }, + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "inline_equation", + "content": "IoU_{i}" + }, + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "text", + "content": " stands for the IoU of predicted box to GT, and " + }, + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "text", + "content": " is the number of foreground classes. Then we can obtain the potential high-quality set " + }, + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "inline_equation", + "content": "S' = \\{(C_{j}, IoU_{i}) | \\arg \\max C_{j} = 0" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6320" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 252, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 252, + 83 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 252, + 83 + ], + "type": "text", + "content": "Algorithm 1 Training of Feature Imitation branch." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 51, + 86, + 80, + 97 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 86, + 80, + 97 + ], + "spans": [ + { + "bbox": [ + 51, + 86, + 80, + 97 + ], + "type": "text", + "content": "Input:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 98, + 286, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 98, + 286, + 121 + ], + "spans": [ + { + "bbox": [ + 64, + 98, + 286, + 121 + ], + "type": "text", + "content": "The set of GT boxes " + }, + { + "bbox": [ + 64, + 98, + 286, + 121 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{c_i^*, b_i^*\\}_{i=1,2,\\dots,T}" + }, + { + "bbox": [ + 64, + 98, + 286, + 121 + ], + "type": "text", + "content": " and corresponding RoI features " + }, + { + "bbox": [ + 64, + 98, + 286, + 121 + ], + "type": "inline_equation", + "content": "\\{\\pmb{x}_i^g\\}_{i=1,2,\\dots,T}" + }, + { + "bbox": [ + 64, + 98, + 286, + 121 + ], + "type": "text", + "content": " in current batch;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 121, + 263, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 121, + 263, + 133 + ], + "spans": [ + { + "bbox": [ + 65, + 121, + 263, + 133 + ], + "type": "text", + "content": "The set of exemplar features " + }, + { + "bbox": [ + 65, + 121, + 263, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{\\mathcal{E}_i\\}_{c = 1,2,\\dots,N}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 133, + 287, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 133, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 65, + 133, + 287, + 157 + ], + "type": "text", + "content": "The set of background RoI features in current batch " + }, + { + "bbox": [ + 65, + 133, + 287, + 157 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{bg}}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 157, + 205, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 157, + 205, + 169 + ], + "spans": [ + { + "bbox": [ + 65, + 157, + 205, + 169 + ], + "type": "text", + "content": "The threshold of high-quality " + }, + { + "bbox": [ + 65, + 157, + 205, + 169 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{hq}}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 65, + 169, + 260, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 169, + 260, + 182 + ], + "spans": [ + { + "bbox": [ + 65, + 169, + 260, + 182 + ], + "type": "text", + "content": "The number of pos/neg samples " + }, + { + "bbox": [ + 65, + 169, + 260, + 182 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{pos}}" + }, + { + "bbox": [ + 65, + 169, + 260, + 182 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 65, + 169, + 260, + 182 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{neg}}" + }, + { + "bbox": [ + 65, + 169, + 260, + 182 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 65, + 182, + 192, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 182, + 192, + 192 + ], + "spans": [ + { + "bbox": [ + 65, + 182, + 192, + 192 + ], + "type": "text", + "content": "The transformation function " + }, + { + "bbox": [ + 65, + 182, + 192, + 192 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 65, + 182, + 192, + 192 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 51, + 193, + 87, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 193, + 87, + 204 + ], + "spans": [ + { + "bbox": [ + 51, + 193, + 87, + 204 + ], + "type": "text", + "content": "Output:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 51, + 205, + 287, + 421 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 64, + 205, + 287, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 205, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 64, + 205, + 287, + 229 + ], + "type": "text", + "content": "The set of positive embeddings " + }, + { + "bbox": [ + 64, + 205, + 287, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 64, + 205, + 287, + 229 + ], + "type": "text", + "content": " and negative embeddings " + }, + { + "bbox": [ + 64, + 205, + 287, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{neg}}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 229, + 287, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 229, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 53, + 229, + 287, + 253 + ], + "type": "text", + "content": "1: Initialize the set of positive features " + }, + { + "bbox": [ + 53, + 229, + 287, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 53, + 229, + 287, + 253 + ], + "type": "text", + "content": " and negative features " + }, + { + "bbox": [ + 53, + 229, + 287, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{neg}}" + }, + { + "bbox": [ + 53, + 229, + 287, + 253 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 53, + 229, + 287, + 253 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 53, + 229, + 287, + 253 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 254, + 120, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 254, + 120, + 264 + ], + "spans": [ + { + "bbox": [ + 53, + 254, + 120, + 264 + ], + "type": "text", + "content": "2: for " + }, + { + "bbox": [ + 53, + 254, + 120, + 264 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 53, + 254, + 120, + 264 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 53, + 254, + 120, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 53, + 254, + 120, + 264 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 265, + 276, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 265, + 276, + 277 + ], + "spans": [ + { + "bbox": [ + 53, + 265, + 276, + 277 + ], + "type": "text", + "content": "3: Compute the " + }, + { + "bbox": [ + 53, + 265, + 276, + 277 + ], + "type": "inline_equation", + "content": "IQ" + }, + { + "bbox": [ + 53, + 265, + 276, + 277 + ], + "type": "text", + "content": " of current " + }, + { + "bbox": [ + 53, + 265, + 276, + 277 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 53, + 265, + 276, + 277 + ], + "type": "text", + "content": " according to Eq. (3)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 277, + 272, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 277, + 272, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 277, + 272, + 289 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 53, + 277, + 272, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{neg}}^q \\gets" + }, + { + "bbox": [ + 53, + 277, + 272, + 289 + ], + "type": "text", + "content": " sample " + }, + { + "bbox": [ + 53, + 277, + 272, + 289 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{neg}}" + }, + { + "bbox": [ + 53, + 277, + 272, + 289 + ], + "type": "text", + "content": " features from " + }, + { + "bbox": [ + 53, + 277, + 272, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{bg}} \\cup \\mathcal{E} \\backslash \\mathcal{E}_{c^*}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 289, + 148, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 289, + 148, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 289, + 148, + 301 + ], + "type": "text", + "content": "5: if " + }, + { + "bbox": [ + 53, + 289, + 148, + 301 + ], + "type": "inline_equation", + "content": "I\\tilde{Q} \\geq T_{\\mathrm{hq}}" + }, + { + "bbox": [ + 53, + 289, + 148, + 301 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 301, + 127, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 301, + 127, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 301, + 127, + 312 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 53, + 301, + 127, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{c^*}\\gets \\pmb {x}_i^g" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 312, + 146, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 312, + 146, + 325 + ], + "spans": [ + { + "bbox": [ + 53, + 312, + 146, + 325 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 53, + 312, + 146, + 325 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{pos}}^g\\gets \\Gamma (\\pmb{x}_i^g)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 53, + 326, + 93, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 326, + 93, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 326, + 93, + 335 + ], + "type": "text", + "content": "8: else" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 53, + 336, + 243, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 243, + 349 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 243, + 349 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 53, + 336, + 243, + 349 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{pos}}^q\\gets" + }, + { + "bbox": [ + 53, + 336, + 243, + 349 + ], + "type": "text", + "content": " sample " + }, + { + "bbox": [ + 53, + 336, + 243, + 349 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{pos}}" + }, + { + "bbox": [ + 53, + 336, + 243, + 349 + ], + "type": "text", + "content": " features from " + }, + { + "bbox": [ + 53, + 336, + 243, + 349 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{c^{*}}" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 51, + 349, + 102, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 349, + 102, + 358 + ], + "spans": [ + { + "bbox": [ + 51, + 349, + 102, + 358 + ], + "type": "text", + "content": "10: end if" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 51, + 360, + 247, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 360, + 247, + 373 + ], + "spans": [ + { + "bbox": [ + 51, + 360, + 247, + 373 + ], + "type": "text", + "content": "11: " + }, + { + "bbox": [ + 51, + 360, + 247, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{pos}} = \\mathcal{X}_{\\mathrm{pos}}\\cup \\mathcal{X}_{\\mathrm{pos}}^{g},\\mathcal{X}_{\\mathrm{neg}} = \\mathcal{X}_{\\mathrm{neg}}\\cup \\mathcal{X}_{\\mathrm{neg}}^{g}" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 51, + 373, + 287, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 373, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 51, + 373, + 287, + 397 + ], + "type": "text", + "content": "12: Apply Eq. (4) to " + }, + { + "bbox": [ + 51, + 373, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 51, + 373, + 287, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 51, + 373, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{neg}}^q" + }, + { + "bbox": [ + 51, + 373, + 287, + 397 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 51, + 373, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 51, + 373, + 287, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 51, + 373, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{neg}}" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 51, + 397, + 97, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 397, + 97, + 407 + ], + "spans": [ + { + "bbox": [ + 51, + 397, + 97, + 407 + ], + "type": "text", + "content": "13: end for" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 51, + 408, + 157, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 408, + 157, + 421 + ], + "spans": [ + { + "bbox": [ + 51, + 408, + 157, + 421 + ], + "type": "text", + "content": "14: return " + }, + { + "bbox": [ + 51, + 408, + 157, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 51, + 408, + 157, + 421 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 51, + 408, + 157, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{neg}}" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 445, + 287, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 470 + ], + "type": "inline_equation", + "content": "\\left.c^{*}\\right\\rbrace_{j = 1,2,\\dots,M^{\\prime}}" + }, + { + "bbox": [ + 46, + 445, + 287, + 470 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 445, + 287, + 470 + ], + "type": "inline_equation", + "content": "M^{\\prime}\\leq M" + }, + { + "bbox": [ + 46, + 445, + 287, + 470 + ], + "type": "text", + "content": " . Now the Instance Quality of an object " + }, + { + "bbox": [ + 46, + 445, + 287, + 470 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 445, + 287, + 470 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 111, + 479, + 287, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 479, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 111, + 479, + 287, + 516 + ], + "type": "interline_equation", + "content": "I Q = \\frac {1}{\\left| \\mathcal {S} ^ {\\prime} \\right|} \\sum_ {j = 1} ^ {M ^ {\\prime}} C _ {j, c ^ {*}} \\cdot I o U _ {j} \\tag {3}", + "image_path": "69a5b1e24f7308b83e0ea525fb464da2f5eb5e195ceb0616db770f7af48ab38f.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "inline_equation", + "content": "IQ" + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "content": " of a GT serves as an indicator of the current model's detection capability, enabling us to capture the high-quality exemplars who have precise localization and high-confidence classification scores, and the instances confusing the model often fail to either of them. By setting appropriate threshold, we can select proper instances to build the teacher feature-set and perform the imitation process." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Feat2Embed Module. Instead of directly measuring the similarity between different RoI features [36], we first embed these features with the simple Feat2Embed module. The input of FI branch is the region feature " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " obtained by the RoI-wise operation, e.g., RoI Align, which will be first processed by three consecutive " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " convolutional layers (with no padding operation) to abstract compact representations. It is worth noting that we update parameters during the extraction of current regional features" + } + ] + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 427, + 202 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 427, + 202 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 427, + 202 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 427, + 202 + ], + "type": "image", + "image_path": "d04745ae166de47b0b2ba865072db57ac92a8597fd2d88c781162db9b3cef9e9.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 205, + 545, + 239 + ], + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 239 + ], + "type": "text", + "content": "Figure 3. Size distribution of instances in (a) SODA-D and (b) SODA-A, where the absolute size corresponds to the square root of the object area." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 433, + 71, + 543, + 202 + ], + "blocks": [ + { + "bbox": [ + 433, + 71, + 543, + 202 + ], + "lines": [ + { + "bbox": [ + 433, + 71, + 543, + 202 + ], + "spans": [ + { + "bbox": [ + 433, + 71, + 543, + 202 + ], + "type": "image", + "image_path": "219a7179ae45b50039c02f5d28aab6d38c0eab04734d461cba9198129f1ea40d.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 260, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 260, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 260, + 545, + 390 + ], + "type": "text", + "content": "and freeze parameters during the extraction of exemplar ones (see Feat2Embed module in Figure 2), resulting in improved stability in performance. Subsequently, the intermediate features will be mapped to the embedding space on top of a two-layer perceptron and the embedding layer with the dimension of 128, in which the dimension of hidden layers is set to 512. We have also investigated various design choices and structures for our Feat2Embed module, and detailed information can be found in the Supplementary Materials. Finally, the output of the Feature Imitation branch is defined as:" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 394, + 405, + 545, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 405, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 394, + 405, + 545, + 418 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {i} = \\Theta_ {\\mathrm {F I}} (\\boldsymbol {x} _ {i}), \\tag {4}", + "image_path": "31e2d802c32cf04476159dbeb3bc0557b3e3dc3350e454564a650be35ff14082.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 425, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 425, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathrm{FI}}" + }, + { + "bbox": [ + 304, + 425, + 545, + 448 + ], + "type": "text", + "content": " denotes the parameters of Feature Imitation branch to be optimized." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 449, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 591 + ], + "type": "text", + "content": "Loss Function. The objective of our FI head is simple: calculating the similarity between the RoI feature of proposal and that of the stored high-quality instances in embedding space, thereby pulling the features of those instances that confuse the model close to the exemplar ones of belonging category, while pushing apart from that of other categories and backgrounds. To this end, we propose a loss function based on Supervised Contrastive Learning [18] which extends the contrastive learning setup and allows multiple positive samples for an anchor object by exploiting the accessible label information. The loss function tailored for our FI branch is as follows:" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 315, + 612, + 545, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 612, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 315, + 612, + 545, + 653 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {F I}} = \\frac {- 1}{| \\mathcal {P} _ {\\text {p o s}} |} \\sum_ {j} \\sum_ {v _ {\\mathrm {p}} \\in \\mathcal {P} _ {\\text {p o s}}} \\log \\frac {\\exp \\left(v _ {j} \\cdot v _ {\\mathrm {p}} / \\tau\\right)}{\\sum_ {v _ {i} \\in \\mathcal {P}} \\exp \\left(v _ {j} \\cdot v _ {i} / \\tau\\right)}, \\tag {5}", + "image_path": "2db04d9cf2c88f6d2b975baeb821eb4277f2e2dbd8ba234d6d5b72a71b26446e.jpg" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\mathcal{P}_{\\mathrm{pos}} \\cup \\mathcal{P}_{\\mathrm{neg}}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " denotes the sample set, while " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{neg}}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " represent the positive and negative set respectively and they have the same cardinality ideally, and " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{p}}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{n}}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " are the positive and negative sample from " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{neg}}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ". Moreover, " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " indexes the current proposal and " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " indicates the" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6321" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 86, + 69, + 506, + 286 + ], + "blocks": [ + { + "bbox": [ + 86, + 69, + 506, + 286 + ], + "lines": [ + { + "bbox": [ + 86, + 69, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 86, + 69, + 506, + 286 + ], + "type": "table", + "html": "
MethodPublicationScheduleAPAP50AP75APesSAPrsSAPgsSAPN
One-stage
RetianNet [26]ICCV'171×28.257.623.711.925.234.144.2
FCOS [31]ICCV'191×23.949.519.96.919.430.940.9
ATSS [46]CVPR'201×26.855.622.111.723.932.241.3
YOLOX [12]ArXiv'2170e26.753.423.013.625.130.930.4
DyHead [10]CVPR'211×27.556.123.212.424.433.041.9
Keypoint-based
CornerNet [21]ECCV'182×24.649.521.76.520.532.243.8
CenterNet [48]ArXiv'1970e21.548.815.65.116.229.642.4
RepPoints [43]ICCV'191×28.055.624.710.123.835.145.3
Query-based
Deformable-DETR [49]ICLR'2050e19.244.813.76.315.424.934.2
Sparse RCNN [30]CVPR'211×24.250.320.38.820.430.239.4
Two-stage
Baseline [29]NeurIPS'151×28.959.424.113.825.734.543.0
Cascade RPN [32]NeurIPS'191×29.156.525.912.525.535.444.7
RFLA [40]ECCV'221×29.760.225.213.226.935.444.6
CFINet (ours)-1×30.760.826.714.727.836.444.6
", + "image_path": "40f893fba3037cd8707f76cc8b085b67afbf11018a88c277cc0a41d713113564.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 289, + 545, + 333 + ], + "lines": [ + { + "bbox": [ + 46, + 289, + 545, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 545, + 333 + ], + "type": "text", + "content": "Table 2. Comparison with state-of-the-art detection approaches on the SODA-D test-set, where 'Baseline' refers to Faster RCNN [29], serving as the baseline for the two-stage methods in the table. All the methods are trained on a ResNet-50 [17], except YOLOX (CSP-Darknet) [12] and CornerNet (HourglassNet-104) [21]. 'Schedule' denotes the number of epochs for training, in which " + }, + { + "bbox": [ + 46, + 289, + 545, + 333 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 46, + 289, + 545, + 333 + ], + "type": "text", + "content": " corresponds to 12 epochs and " + }, + { + "bbox": [ + 46, + 289, + 545, + 333 + ], + "type": "inline_equation", + "content": "50\\mathrm{e}" + }, + { + "bbox": [ + 46, + 289, + 545, + 333 + ], + "type": "text", + "content": " indicates 50 epochs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 352, + 287, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 401 + ], + "type": "text", + "content": "temperature which plays a crucial part in contrastive learning and needs to be well designed, and we conduct ablation studies (see Table 9) to determine the optimal setting in our framework. The total loss function is presented:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 421, + 287, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 421, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 90, + 421, + 287, + 434 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C R P N}} + \\mathcal {L} _ {\\mathrm {c l s}} + \\mathcal {L} _ {\\mathrm {r e g}} + \\alpha_ {3} \\mathcal {L} _ {\\mathrm {F I}}, \\tag {6}", + "image_path": "d8ad46fedb92b7b8b573942e86e6b9f4c8bd8b5b79b928a79bb2a239dbe3e3ea.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 448, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 287, + 543 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 46, + 448, + 287, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 46, + 448, + 287, + 543 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 448, + 287, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 46, + 448, + 287, + 543 + ], + "type": "text", + "content": " are the original losses of detection head, and the " + }, + { + "bbox": [ + 46, + 448, + 287, + 543 + ], + "type": "inline_equation", + "content": "\\alpha_{3}" + }, + { + "bbox": [ + 46, + 448, + 287, + 543 + ], + "type": "text", + "content": " is utilized to scale the weight of Feature Imitation part. With contrastive learning setups, not only can we fulfill the imitation learning but also prevent the collapse issue, thereby boosting the representations of small instances effectively. Moreover, the imitation process is only installed in training phase and will not slow the pace of inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": "Training. Next we elucidate the training details of FI branch. The exemplar set " + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{\\mathcal{E}_i\\}_{c=1,2,\\dots,N}" + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": " containing high-quality features of " + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": " foreground categories and " + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_i = \\{\\pmb{x}_{i,j}\\}_{j=1,2,\\dots,N_i}" + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": " corresponding to the exemplar features of the " + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": "-th class, and " + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "N_i" + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": " represents its size. We use " + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{hq}}" + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": " to pick out those high-quality instances which are compatible to be a good exemplar, and in practice, we set a bound value to the number of high-quality predictions of an instance to filter the effect of fluctuation of the network. The function " + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": " is used to augment features for high-quality instances, i.e., the positive features for a high-quality instance are the transformations of itself. The overall training procedure of FI branch is shown in Alg. 1 and more details please refer to the Supplementary Materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 351, + 387, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 351, + 387, + 365 + ], + "spans": [ + { + "bbox": [ + 306, + 351, + 387, + 365 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 371, + 364, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 364, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 364, + 383 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 390, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 437 + ], + "type": "text", + "content": "To evaluate the effectiveness of our method, we perform extensive experiments on the recently released large-scale benchmark tailored for small object detection: SODA [9], including SODA-D and SODA-A." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "type": "text", + "content": "SODA-D. Focusing on the driving scenario, SODA-D comprises 24828 high-quality images and 278433 instances distributed across nine categories: people, rider, bicycle, motor, vehicle, traffic-sign, traffic-light, traffic-camera, and warning-cone. One of the most distinctive strengths of SODA-D is its diversity in terms of period, geographical locations, weather conditions, camera viewpoints, etc." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 616 + ], + "type": "text", + "content": "SODA-A. SODA-A contains 872069 objects with oriented box annotations in 2513 aerial images and encompassing nine classes: airplane, helicopter, small-vehicle, large-vehicle, ship, container, storage-tank, swimming-pool, and windmill. The instances in SODA-A can appear in arbitrary orientations and are with significant density variations. To be specific, the average number of instances per image in SODA-A is about 350." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "As a specialized benchmark for small object detection, the instances in SODA are tiny, with most of them having an average size ranging from 10 to 30 pixels (see Figure 3). In contrast to conventional datasets for object detection, SODA includes extensive ignore annotations, aimed at filtering out instances that are either too large or are challenging to be identified deterministically due to heavy occlusion or lens flare. This procedure helps the model focus on valu" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6322" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 86, + 69, + 506, + 208 + ], + "blocks": [ + { + "bbox": [ + 86, + 69, + 506, + 208 + ], + "lines": [ + { + "bbox": [ + 86, + 69, + 506, + 208 + ], + "spans": [ + { + "bbox": [ + 86, + 69, + 506, + 208 + ], + "type": "table", + "html": "
MethodPublicationScheduleAPAP50AP75APesSAPrsSAPgsSAPN
One-stage
Rotated RetinaNet [26]ICCV'171×26.863.416.29.122.035.428.2
S2A-Net [15]TGRS'221×28.369.613.110.222.835.829.5
Oriented RepPoints [23]CVPR'221×26.358.819.09.422.632.428.5
DHRec [27]TPAMI'221×30.168.819.810.624.640.334.6
Two-stage
Baseline [29]NeurIPS'151×32.570.124.311.927.342.234.4
Gliding Vertex [41]TPAMI'211×31.770.822.611.727.041.133.8
Oriented RCNN [39]ICCV'211×34.470.728.612.528.644.536.7
DODet [8]TGRS'221×31.668.123.411.326.341.033.5
CFINet (ours)-1×34.473.126.113.529.344.035.9
", + "image_path": "9e46b7b661e6bb368fdfa73572c4426fc7b1c96fd8160263e85eec698d3bcad0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 86, + 255, + 223, + 374 + ], + "blocks": [ + { + "bbox": [ + 130, + 246, + 194, + 255 + ], + "lines": [ + { + "bbox": [ + 130, + 246, + 194, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 194, + 255 + ], + "type": "text", + "content": "(a) Faster RCNN" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 86, + 255, + 223, + 374 + ], + "lines": [ + { + "bbox": [ + 86, + 255, + 223, + 374 + ], + "spans": [ + { + "bbox": [ + 86, + 255, + 223, + 374 + ], + "type": "image", + "image_path": "06c1161da153eb4f95a2e9c739c16ec16b6d5504a7b01e07ec71e047ce779adc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 377, + 546, + 411 + ], + "lines": [ + { + "bbox": [ + 46, + 377, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 546, + 411 + ], + "type": "text", + "content": "Figure 4. The average number of high-quality proposals generated by (a) RPN, (b) Cascade RPN and (c) CRPN for instances in extremely Small " + }, + { + "bbox": [ + 46, + 377, + 546, + 411 + ], + "type": "inline_equation", + "content": "(eS)" + }, + { + "bbox": [ + 46, + 377, + 546, + 411 + ], + "type": "text", + "content": ", relatively Small " + }, + { + "bbox": [ + 46, + 377, + 546, + 411 + ], + "type": "inline_equation", + "content": "(rS)" + }, + { + "bbox": [ + 46, + 377, + 546, + 411 + ], + "type": "text", + "content": ", generally Small " + }, + { + "bbox": [ + 46, + 377, + 546, + 411 + ], + "type": "inline_equation", + "content": "(gS)" + }, + { + "bbox": [ + 46, + 377, + 546, + 411 + ], + "type": "text", + "content": ", and Normal " + }, + { + "bbox": [ + 46, + 377, + 546, + 411 + ], + "type": "inline_equation", + "content": "(N)" + }, + { + "bbox": [ + 46, + 377, + 546, + 411 + ], + "type": "text", + "content": " subsets, respectively. Noting that a proposal who has an IoU larger than 0.5 to any ground-truth boxes will be registered as a high-quality proposal." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 238, + 255, + 365, + 373 + ], + "blocks": [ + { + "bbox": [ + 272, + 246, + 336, + 255 + ], + "lines": [ + { + "bbox": [ + 272, + 246, + 336, + 255 + ], + "spans": [ + { + "bbox": [ + 272, + 246, + 336, + 255 + ], + "type": "text", + "content": "(b) Cascade RPN" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 238, + 255, + 365, + 373 + ], + "lines": [ + { + "bbox": [ + 238, + 255, + 365, + 373 + ], + "spans": [ + { + "bbox": [ + 238, + 255, + 365, + 373 + ], + "type": "image", + "image_path": "21b07048a8cfdb5503efcd5a8738b9797431028ee9a8637befbe54e87cc7df8f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 380, + 255, + 507, + 374 + ], + "blocks": [ + { + "bbox": [ + 429, + 246, + 463, + 255 + ], + "lines": [ + { + "bbox": [ + 429, + 246, + 463, + 255 + ], + "spans": [ + { + "bbox": [ + 429, + 246, + 463, + 255 + ], + "type": "text", + "content": "(c) CRPN" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 380, + 255, + 507, + 374 + ], + "lines": [ + { + "bbox": [ + 380, + 255, + 507, + 374 + ], + "spans": [ + { + "bbox": [ + 380, + 255, + 507, + 374 + ], + "type": "image", + "image_path": "0ce2af594a95f3116ebfb7e7537487d53d9f1699b1b9c8a3158f6b09ec7ba564.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 48, + 430, + 296, + 503 + ], + "blocks": [ + { + "bbox": [ + 46, + 210, + 546, + 234 + ], + "lines": [ + { + "bbox": [ + 46, + 210, + 546, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 546, + 234 + ], + "type": "text", + "content": "Table 3. Comparison with state-of-the-art detection approaches on the SODA-A test-set, where 'Baseline' refers to Rotated Faster RCNN [29], serving as the baseline for the two-stage methods in the table. Other settings are consistent with Table 2." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 430, + 296, + 503 + ], + "lines": [ + { + "bbox": [ + 48, + 430, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 430, + 296, + 503 + ], + "type": "table", + "html": "
Proposal MethodAR\\( AR_{eS} \\)\\( AR_{rS} \\)\\( AR_{gS} \\)\\( AR_N \\)
RPN [29]41.224.038.347.357.1
RPN-0.541.324.238.547.354.1
GA-RPN [33]42.124.139.248.956.2
Cascade RPN [32]41.822.838.248.757.1
CRPN42.624.638.949.156.9
", + "image_path": "2d36f626c13a93bc637e10645430291a47152aa3a3e621978425760cee9f6b60.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "able small instances. The objects in SODA are divided into Small and Normal according to their areas, in which Small is further split into three subsets: extremely Small (eS), relatively Small (rS) and generally Small (gS). The evaluation metric of SODA follows that of COCO [24], namely averaging the precision over 10 IoU thresholds ranging from 0.5 to 0.95 (with an interval of 0.05), specifically focusing on Small objects." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 310, + 430, + 541, + 492 + ], + "blocks": [ + { + "bbox": [ + 46, + 507, + 287, + 586 + ], + "lines": [ + { + "bbox": [ + 46, + 507, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 287, + 586 + ], + "type": "text", + "content": "Table 4. Average Recall " + }, + { + "bbox": [ + 46, + 507, + 287, + 586 + ], + "type": "inline_equation", + "content": "(AR)" + }, + { + "bbox": [ + 46, + 507, + 287, + 586 + ], + "type": "text", + "content": " performances of our CRPN and its counterparts on the SODA-D test-set. All the methods are with Faster RCNN (ResNet-50) as the baseline and trained for a " + }, + { + "bbox": [ + 46, + 507, + 287, + 586 + ], + "type": "inline_equation", + "content": "1\\times" + }, + { + "bbox": [ + 46, + 507, + 287, + 586 + ], + "type": "text", + "content": " schedule, in which RPN denotes the vanilla version of Faster RCNN whose positive threshold in RPN stage is set 0.7, and RPN-0.5 represents the version with 0.5 as its positive threshold of RPN. The results are tested with 300 proposals per image." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 430, + 541, + 492 + ], + "lines": [ + { + "bbox": [ + 310, + 430, + 541, + 492 + ], + "spans": [ + { + "bbox": [ + 310, + 430, + 541, + 492 + ], + "type": "table", + "html": "
BaselineCRPNFIAPAPeSAPrSAPgS
✓28.913.825.734.5
✓✓30.314.327.336.1
✓✓29.514.426.335.1
✓✓✓30.714.727.836.4
", + "image_path": "83329ae4a9bd5d28c0b3b5b83ff406be8ca1ee2d87c553b22f16f15ac03d57d5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 494, + 545, + 528 + ], + "lines": [ + { + "bbox": [ + 305, + 494, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 305, + 494, + 545, + 528 + ], + "type": "text", + "content": "Table 5. Ablation analysis of our method, in which 'Baseline' denotes the vanilla Faster RCNN, CRPN and FI indicate Coarse-to-fine RPN and Feature Imitation branch, respectively." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 539, + 440, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 440, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 440, + 552 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "In the following experiments, unless specified, we use train-set to conduct the training and leave test-set to performance comparisons and ablation studies. Considering that the images in SODA enjoy a very high resolution " + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "inline_equation", + "content": "(\\sim 4000\\times 3000)" + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": ", we first split the original images into a series of " + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "inline_equation", + "content": "800\\times 800" + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": " patches with a stride of 650, and similar to [9] these patches will be resized to " + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "inline_equation", + "content": "1200\\times 1200" + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": " during training and testing. All the experiments in this paper are conducted on a single RTX 3090 with the batch size of 4. Only random flip involved in data augmentation. We train all the models with a " + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "inline_equation", + "content": "1\\times" + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": " schedule (a bunch of 12 epochs), and the learning rate is set to 0.01 which decays after epoch 8 and epoch 11 by 0.1. The default optimizer is SGD with" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6323" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 70, + 250, + 117 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 250, + 117 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 250, + 117 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 250, + 117 + ], + "type": "table", + "html": "
StrategyAPAPeSAPrSAPgS
0.2029.913.726.835.9
0.4030.114.126.936.2
Ours30.314.327.336.1
", + "image_path": "b6b25e1b8bbafe0b141b51706c15cbe6e0adab44b45f059547c08ca9b2d1c2da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 85, + 152, + 250, + 203 + ], + "blocks": [ + { + "bbox": [ + 47, + 119, + 287, + 142 + ], + "lines": [ + { + "bbox": [ + 47, + 119, + 287, + 142 + ], + "spans": [ + { + "bbox": [ + 47, + 119, + 287, + 142 + ], + "type": "text", + "content": "Table 6. Different definitions about positive anchor for CRPN, in which 'Ours' denotes the proposed dynamic strategy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 152, + 250, + 203 + ], + "lines": [ + { + "bbox": [ + 85, + 152, + 250, + 203 + ], + "spans": [ + { + "bbox": [ + 85, + 152, + 250, + 203 + ], + "type": "table", + "html": "
α3APAPeSAPrSAPgS
0.2530.614.727.636.2
0.5030.714.727.836.4
0.7530.414.027.536.2
", + "image_path": "f25e08745268b1c9ee14c8fb808060436693097e4e21d6c609aebcd91ec63eee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 228, + 287, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 287, + 251 + ], + "type": "text", + "content": "the momentum of 0.9 and the weight decay of 0.0001. We use ResNet-50 [17] with FPN [25] for all models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 261, + 132, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 261, + 132, + 272 + ], + "spans": [ + { + "bbox": [ + 47, + 261, + 132, + 272 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 279, + 287, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 279, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 47, + 279, + 287, + 315 + ], + "type": "text", + "content": "To exhibit the effectiveness of our method, we conduct a thorough comparison with current representative approaches on the SODA-D and SODA-A." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "type": "text", + "content": "Table 2 represents the results of our method and several mainstream approaches on the SODA-D test-set. Integrating with Faster RCNN [29], our CFINet achieves state-of-the-art performance with an overall " + }, + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "type": "inline_equation", + "content": "30.7\\%" + }, + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "type": "text", + "content": ", and outperforms the baseline model with " + }, + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "type": "text", + "content": " points. When delving into specific metrics, our method exhibits clear predominance, particularly on the most challenging metrics " + }, + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "type": "inline_equation", + "content": "AP_{es}" + }, + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "type": "inline_equation", + "content": "AP_{rs}" + }, + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "type": "text", + "content": ". Moreover, CFINet exceeds the tailored small object detection method RFLA [40] by a significant margin (1.0% on " + }, + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "type": "text", + "content": ", 1.5% on " + }, + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "type": "inline_equation", + "content": "AP_{es}" + }, + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "type": "text", + "content": ", and 0.9% on " + }, + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "type": "inline_equation", + "content": "AP_{rs}" + }, + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "type": "text", + "content": ") when both taking Faster RCNN as the baseline. Actually, RFLA sacrifices the performance on extremely Small instances (with a decrease of " + }, + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 46, + 316, + 287, + 482 + ], + "type": "text", + "content": " points when compared to Faster RCNN)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": "On the SODA-A test-set, CFINet also achieves the best result and shows great advantage in comparison to other solutions especially on " + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "inline_equation", + "content": "AP_{eS}" + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": " (see Table 3), indicating its superiority and generality. Furthermore, albeit exhibiting advantage on " + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "inline_equation", + "content": "AP_{75}" + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": " metric and larger instances, Oriented RCNN [39] lags largely behind our approach on " + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "inline_equation", + "content": "AP_{50}" + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": " (73.1% vs. 70.7%) and " + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "inline_equation", + "content": "AP_{eS}" + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": " (13.5% vs. 12.5%)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 574, + 175, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 175, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 175, + 586 + ], + "type": "text", + "content": "4.4. Effectiveness of CRPN" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 689 + ], + "type": "text", + "content": "One of the main designs in this paper is the Coarse-to-fine RPN, which is based on the observation that current fixed overlaps-based sampling paradigm is inappropriate for small instances due to the inherent contradictions, while the refined designs of RPN could partially reduce this barrier but still fail to satisfactory results. Here, we conduct thorough analyses to demonstrate the capability of our CRPN to generate high-quality proposals for size-limited instances." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "content": "We first exhibit the recall performances of our CRPN and its counterparts in Table 4, from which we can see that low" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 350, + 70, + 503, + 128 + ], + "blocks": [ + { + "bbox": [ + 47, + 205, + 275, + 216 + ], + "lines": [ + { + "bbox": [ + 47, + 205, + 275, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 275, + 216 + ], + "type": "text", + "content": "Table 7. The effect of loss weight for Feature Imitation branch." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 350, + 70, + 503, + 128 + ], + "lines": [ + { + "bbox": [ + 350, + 70, + 503, + 128 + ], + "spans": [ + { + "bbox": [ + 350, + 70, + 503, + 128 + ], + "type": "table", + "html": "
ThqAPAPeSAPrSAPgS
0.5030.314.027.336.0
0.5530.614.327.336.5
0.6530.714.727.836.4
0.7030.514.627.336.3
", + "image_path": "21b84337df99ab436ef25bcada39158b0d0a98851de0f9a2fd972e62ee259e62.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 349, + 162, + 503, + 220 + ], + "blocks": [ + { + "bbox": [ + 306, + 131, + 545, + 151 + ], + "lines": [ + { + "bbox": [ + 306, + 131, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 306, + 131, + 545, + 151 + ], + "type": "text", + "content": "Table 8. The investigation of the criterion to be an exemplar instance." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 349, + 162, + 503, + 220 + ], + "lines": [ + { + "bbox": [ + 349, + 162, + 503, + 220 + ], + "spans": [ + { + "bbox": [ + 349, + 162, + 503, + 220 + ], + "type": "table", + "html": "
τAPAPeSAPrSAPgS
0.1030.314.127.336.1
0.5030.414.427.236.2
0.6030.714.727.836.4
0.8030.214.027.336.0
", + "image_path": "f0d737675412a9fd982fd21eb482587139c7438dc1bf0cbed7539b5108b8ebdd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 222, + 534, + 233 + ], + "lines": [ + { + "bbox": [ + 306, + 222, + 534, + 233 + ], + "spans": [ + { + "bbox": [ + 306, + 222, + 534, + 233 + ], + "type": "text", + "content": "Table 9. The choices of temperature " + }, + { + "bbox": [ + 306, + 222, + 534, + 233 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 306, + 222, + 534, + 233 + ], + "type": "text", + "content": " to the final performance." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 255, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 374 + ], + "type": "text", + "content": "ering the positive threshold slightly improves the average recall while sacrificing the performance of larger instances " + }, + { + "bbox": [ + 304, + 255, + 545, + 374 + ], + "type": "inline_equation", + "content": "(AR_{N}" + }, + { + "bbox": [ + 304, + 255, + 545, + 374 + ], + "type": "text", + "content": " experiences a sharp decline from " + }, + { + "bbox": [ + 304, + 255, + 545, + 374 + ], + "type": "inline_equation", + "content": "57.1\\%" + }, + { + "bbox": [ + 304, + 255, + 545, + 374 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 255, + 545, + 374 + ], + "type": "inline_equation", + "content": "54.1\\%" + }, + { + "bbox": [ + 304, + 255, + 545, + 374 + ], + "type": "text", + "content": "). GA-RPN [33] as well as Cascade RPN [32] both fail to better results since their patterns incline to large instances as we discussed before. In comparison with RPN and its variants, our CRPN demonstrates superior performance on objects in the Small, while achieving comparable results on Normal instances. This validates our assumption that refined proposal networks tend to favor larger objects." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 374, + 546, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 546, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 546, + 566 + ], + "type": "text", + "content": "We conjecture that one of the most challenging issues towards accurate small object detection is the scarcity of high-quality samples, which is also the major motivation behind the design of CRPN. Hence, we intuitively compare the baseline RPN, Cascade RPN and our CRPN about the number of high-quality samples. In Figure 4, our CRPN generates more high-quality proposals compared to the other competitors. Interestingly, CRPN can dynamically shift the focus along the training: at the beginning, the model concentrates more on large objects which are conducive for early optimization while as the training goes, the model gradually shifts its attention to objects with small sizes that are usually not handled well before. This is interpretable since the instances having extremely limited sizes are with more uncertainties and fitting them in early-phase is not an optimal choice for the detector." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 574, + 441, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 441, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 441, + 586 + ], + "type": "text", + "content": "4.5. Ablation and Discussion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 594, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 652 + ], + "type": "text", + "content": "In this part, we conduct ablation studies as well as comprehensive discussions to attest the importance of the CRPN and FI branch, and moreover, determine the appropriate settings of our approach. All the experiments of this section are conducted on the SODA-D test-set." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "Investigation of Designed Components. We first perform ablation experiments to verify the effectiveness of two modules. As in Table 5, our CRPN and FI can both improve the performances steadily, while the introduction of feature imitation branch is conducive for the recognition of size-" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6324" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 302, + 245 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 302, + 245 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 302, + 245 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 302, + 245 + ], + "type": "image", + "image_path": "e5d4b8aa12b8a76b267c49ca40f941f184c53f1520ec2603aae038f11f9123fb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 75, + 249, + 90, + 258 + ], + "blocks": [ + { + "bbox": [ + 75, + 249, + 90, + 258 + ], + "lines": [ + { + "bbox": [ + 75, + 249, + 90, + 258 + ], + "spans": [ + { + "bbox": [ + 75, + 249, + 90, + 258 + ], + "type": "image", + "image_path": "f768988861a961f0b9e6c85feb36fda818ebc3bee8c8b0b8db68e1cdb90500cd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 250, + 109, + 258 + ], + "lines": [ + { + "bbox": [ + 91, + 250, + 109, + 258 + ], + "spans": [ + { + "bbox": [ + 91, + 250, + 109, + 258 + ], + "type": "text", + "content": "people" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 119, + 249, + 131, + 258 + ], + "blocks": [ + { + "bbox": [ + 119, + 249, + 131, + 258 + ], + "lines": [ + { + "bbox": [ + 119, + 249, + 131, + 258 + ], + "spans": [ + { + "bbox": [ + 119, + 249, + 131, + 258 + ], + "type": "image", + "image_path": "341b2c87ca4f9ad8262dd4f74f21cdf09ea19eb8bc6b0f232d789b2cd91bfdba.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 250, + 148, + 258 + ], + "lines": [ + { + "bbox": [ + 132, + 250, + 148, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 250, + 148, + 258 + ], + "type": "text", + "content": "rider" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 261, + 545, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 261, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 545, + 283 + ], + "type": "text", + "content": "Figure 5. Qualitative results of our method on the SODA-D test-set. Only predictions with confidence scores larger than 0.3 are demonstrated and the masked bounding boxes represent ignore regions. Best viewed in color and zoom-in windows." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 159, + 249, + 172, + 258 + ], + "blocks": [ + { + "bbox": [ + 159, + 249, + 172, + 258 + ], + "lines": [ + { + "bbox": [ + 159, + 249, + 172, + 258 + ], + "spans": [ + { + "bbox": [ + 159, + 249, + 172, + 258 + ], + "type": "image", + "image_path": "cf692c67adc8cde0f975c194db3521c7837419b8592302fb8aae9ddde520cbff.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 250, + 193, + 258 + ], + "lines": [ + { + "bbox": [ + 173, + 250, + 193, + 258 + ], + "spans": [ + { + "bbox": [ + 173, + 250, + 193, + 258 + ], + "type": "text", + "content": "bicycle" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 205, + 249, + 219, + 258 + ], + "blocks": [ + { + "bbox": [ + 205, + 249, + 219, + 258 + ], + "lines": [ + { + "bbox": [ + 205, + 249, + 219, + 258 + ], + "spans": [ + { + "bbox": [ + 205, + 249, + 219, + 258 + ], + "type": "image", + "image_path": "842dc621bb1adaf4d03d7025959f6314ade178f76812e41710e44c553c5d894f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 250, + 236, + 258 + ], + "lines": [ + { + "bbox": [ + 220, + 250, + 236, + 258 + ], + "spans": [ + { + "bbox": [ + 220, + 250, + 236, + 258 + ], + "type": "text", + "content": "motor" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 249, + 250, + 263, + 258 + ], + "blocks": [ + { + "bbox": [ + 249, + 250, + 263, + 258 + ], + "lines": [ + { + "bbox": [ + 249, + 250, + 263, + 258 + ], + "spans": [ + { + "bbox": [ + 249, + 250, + 263, + 258 + ], + "type": "image", + "image_path": "32ab6006046aa2afdf1851f67ff4746a3a31ce59b16cdb41bd73b2ce4a949798.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 250, + 282, + 258 + ], + "lines": [ + { + "bbox": [ + 264, + 250, + 282, + 258 + ], + "spans": [ + { + "bbox": [ + 264, + 250, + 282, + 258 + ], + "type": "text", + "content": "vehicle" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 304, + 72, + 557, + 247 + ], + "blocks": [ + { + "bbox": [ + 304, + 72, + 557, + 247 + ], + "lines": [ + { + "bbox": [ + 304, + 72, + 557, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 557, + 247 + ], + "type": "image", + "image_path": "afeb7780a1a48b43ac48917ec4dce1e80a3f1b3e5b522e08444792753f8130e9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 305, + 250, + 320, + 258 + ], + "blocks": [ + { + "bbox": [ + 305, + 250, + 320, + 258 + ], + "lines": [ + { + "bbox": [ + 305, + 250, + 320, + 258 + ], + "spans": [ + { + "bbox": [ + 305, + 250, + 320, + 258 + ], + "type": "image", + "image_path": "216e7c7af19ce4309f803c82af3de7c8297cef2022bc2fb70f19bc73adf83f76.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 250, + 336, + 258 + ], + "lines": [ + { + "bbox": [ + 321, + 250, + 336, + 258 + ], + "spans": [ + { + "bbox": [ + 321, + 250, + 336, + 258 + ], + "type": "text", + "content": "c-sign" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 350, + 250, + 364, + 258 + ], + "blocks": [ + { + "bbox": [ + 350, + 250, + 364, + 258 + ], + "lines": [ + { + "bbox": [ + 350, + 250, + 364, + 258 + ], + "spans": [ + { + "bbox": [ + 350, + 250, + 364, + 258 + ], + "type": "image", + "image_path": "4c16ae33eac43f790dfeb505957673672182147820f96898ac3c99bcb458eb3e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 250, + 394, + 258 + ], + "lines": [ + { + "bbox": [ + 365, + 250, + 394, + 258 + ], + "spans": [ + { + "bbox": [ + 365, + 250, + 394, + 258 + ], + "type": "text", + "content": "traffic-light" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 407, + 250, + 421, + 258 + ], + "blocks": [ + { + "bbox": [ + 407, + 250, + 421, + 258 + ], + "lines": [ + { + "bbox": [ + 407, + 250, + 421, + 258 + ], + "spans": [ + { + "bbox": [ + 407, + 250, + 421, + 258 + ], + "type": "image", + "image_path": "90c053259362cf6f50d580ac9723840ce8543c177c5a7171e5cc6c6547c5520c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 250, + 457, + 258 + ], + "lines": [ + { + "bbox": [ + 422, + 250, + 457, + 258 + ], + "spans": [ + { + "bbox": [ + 422, + 250, + 457, + 258 + ], + "type": "text", + "content": "traffic-camera" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 484, + 250, + 498, + 258 + ], + "blocks": [ + { + "bbox": [ + 484, + 250, + 498, + 258 + ], + "lines": [ + { + "bbox": [ + 484, + 250, + 498, + 258 + ], + "spans": [ + { + "bbox": [ + 484, + 250, + 498, + 258 + ], + "type": "image", + "image_path": "e8212f3ddfaa25b94322f9cf2cae84358d98ee9c57b1f692ae14be81d7e658ec.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 499, + 250, + 521, + 258 + ], + "lines": [ + { + "bbox": [ + 499, + 250, + 521, + 258 + ], + "spans": [ + { + "bbox": [ + 499, + 250, + 521, + 258 + ], + "type": "text", + "content": "ing-cone" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 303, + 287, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 287, + 362 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 287, + 362 + ], + "type": "text", + "content": "limited instances. Consistently, the integration of FI and CRPN achieves the best result, as CRPN is capable of generating sufficient high-quality proposals (as shown in Figure 4) thereby offering a more accurate indication of instance quality and the potential to act as an exemplar." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 364, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 287, + 495 + ], + "type": "text", + "content": "Fixed or Dynamic. A natural idea is directly setting fixed positive threshold to obtain more anchors for one-stage regression of CRPN. Here we show that our simple-yet-effective area-based anchor mining strategy can achieve the best performance. In Table 6, when the IoU threshold of a positive potential anchor for first-stage regression drops to 0.20, the " + }, + { + "bbox": [ + 46, + 364, + 287, + 495 + ], + "type": "inline_equation", + "content": "AP_{eS}" + }, + { + "bbox": [ + 46, + 364, + 287, + 495 + ], + "type": "text", + "content": " is only " + }, + { + "bbox": [ + 46, + 364, + 287, + 495 + ], + "type": "inline_equation", + "content": "13.7\\%" + }, + { + "bbox": [ + 46, + 364, + 287, + 495 + ], + "type": "text", + "content": " and this could be attributed to low-quality samples and the predominance of large instances as discussed before. The proposed area-based anchor mining strategy could mitigate this problem and obtain the best overall accuracy." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 497, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 568 + ], + "type": "text", + "content": "Weights of Feature Imitation Loss. In this section, we analyze the impact of the weight parameter of FI branch (namely the hyper-parameter " + }, + { + "bbox": [ + 46, + 497, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\alpha_{3}" + }, + { + "bbox": [ + 46, + 497, + 287, + 568 + ], + "type": "text", + "content": ") on the model. As shown in Table 7, paying too little or excessive attention both deteriorates the final performance, hence we set " + }, + { + "bbox": [ + 46, + 497, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\alpha_{3}" + }, + { + "bbox": [ + 46, + 497, + 287, + 568 + ], + "type": "text", + "content": " to 0.5 in our experiments to ensure overall accuracy." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "The Criterion of Being An Exemplar. The quality of the exemplar plays a pivotal role in the imitation process [19, 36]. Next we discuss the choices for capturing a high-quality exemplar to build the feature set. In Table 8, the lower " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{hq}}" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": " involves more exemplars and updates the teacher set more frequently while the higher " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{hq}}" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": " does exactly the opposite. It can be seen that increasing the " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{hq}}" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": " from 0.5 to 0.65 could facilitate the imitation process and when the " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{hq}}" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": " reaches 0.70, the overall " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": " drops instead. This may originates that the earliest samples stored in the feature set are inappropriate for current state, because the optimization is dynamic and the model is evolving hence the criterion of" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 303, + 468, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 303, + 468, + 315 + ], + "spans": [ + { + "bbox": [ + 305, + 303, + 468, + 315 + ], + "type": "text", + "content": "being an exemplar has already changed." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 316, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 316, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 545, + 386 + ], + "type": "text", + "content": "The Temperature. The temperature is paramount for contrastive learning [18] and we conduct a series of experiments to verify the best choice for " + }, + { + "bbox": [ + 305, + 316, + 545, + 386 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 316, + 545, + 386 + ], + "type": "text", + "content": ". From Table 9, when " + }, + { + "bbox": [ + 305, + 316, + 545, + 386 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 316, + 545, + 386 + ], + "type": "text", + "content": " ranges from 0.10 to 0.80, the overall performance increases first and then decreases to " + }, + { + "bbox": [ + 305, + 316, + 545, + 386 + ], + "type": "inline_equation", + "content": "30.2\\%" + }, + { + "bbox": [ + 305, + 316, + 545, + 386 + ], + "type": "text", + "content": ", therefore we choose 0.6 for our method." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 387, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 387, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 387, + 545, + 434 + ], + "type": "text", + "content": "Visualization. We demonstrate the visualization results of example images from SODA-D test-set in Figure 5 to intuitively show the capability of our detector when detecting the small instances." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 446, + 378, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 446, + 378, + 458 + ], + "spans": [ + { + "bbox": [ + 306, + 446, + 378, + 458 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 467, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 545, + 622 + ], + "type": "text", + "content": "In this paper, we proposed CFINet, a two-stage detector based on the Coarse-to-fine Region Proposal Network and Feature Imitation setups, in which the former can produce sufficient high-quality proposals for small instances particularly for those with extremely limited sizes. Then the novel detection head on top of the feature imitation branch facilitates the representations of small objects posing challenges to the model under the contrastive learning paradigm. The experiments results show that our method achieves state-of-the-art performance on the large-scale small object detection datasets SODA-D and SODA-A. In the future, a more flexible and general indicator of instance quality is worth investigating." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 633, + 403, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 403, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 403, + 647 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 712 + ], + "type": "text", + "content": "This work was supported in part by the National Science Foundation of China under Grant 62136007 and Grant U20B2068, and in part by the Natural Science Basic Research Program of Shaanxi under Grants 2021JC-16 and 2023-JC-ZD-36." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6325" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. Finding tiny faces in the wild with generative adversarial network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 21-30, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "text", + "content": "[2] Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. Sod-mtgan: Small object detection via multi-task generative adversarial network. In Proceedings of the European Conference on Computer Vision, pages 206-221, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 287, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 247 + ], + "type": "text", + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European Conference on Computer Vision, pages 213-229, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 287, + 294 + ], + "type": "text", + "content": "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Cconference on Machine Learning, pages 1597-1607, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 287, + 339 + ], + "type": "text", + "content": "[5] Gong Cheng, Junwei Han, Peicheng Zhou, and Dong Xu. Learning rotation-invariant and fisher discriminative convolutional neural networks for object detection. IEEE Transactions on Image Processing, 28(1):265-278, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 341, + 287, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 287, + 395 + ], + "type": "text", + "content": "[6] Gong Cheng, Qingyang Li, Guangxing Wang, Xingxing Xie, Lingtong Min, and Junwei Han. Sfrnet: Fine-grained oriented object recognition via separate feature refinement. IEEE Transactions on Geoscience and Remote Sensing, 61:1-10, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 396, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 287, + 441 + ], + "type": "text", + "content": "[7] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "type": "text", + "content": "[8] Gong Cheng, Yanqing Yao, Shengyang Li, Ke Li, Xingxing Xie, Jiabao Wang, Xiwen Yao, and Junwei Han. Dual-aligned oriented detector. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 542 + ], + "type": "text", + "content": "[9] Gong Cheng, Xiang Yuan, Xiwen Yao, Kebing Yan, Qinghua Zeng, Xingxing Xie, and Junwei Han. Towards large-scale small object detection: Survey and benchmarks. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-20, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "text", + "content": "[10] Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unifying object detection heads with attentions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7373-7382, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "type": "text", + "content": "[11] Chunfang Deng, Mengmeng Wang, Liang Liu, Yong Liu, and Yunliang Jiang. Extended feature pyramid network for small object detection. IEEE Transactions on Multimedia, 24:1968-1979, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "text", + "content": "[12] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[13] Spyros Gidaris and Nikos Komodakis. Attend refine repeat: Active box proposal generation via in-out localization. arXiv preprint arXiv:1606.04446, 2016." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[15] Jiaming Han, Jian Ding, Jie Li, and Gui-Song Xia. Align deep features for oriented object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "type": "text", + "content": "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 254, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 306 + ], + "type": "text", + "content": "[18] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "type": "text", + "content": "[19] Jung Uk Kim, Sungjune Park, and Yong Man Ro. Robust small-scale pedestrian detection with cued recall via memory learning. In Proceedings of the IEEE International Conference on Computer Vision, pages 3050-3059, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "type": "text", + "content": "[20] Chunbo Lang, Gong Cheng, Binfei Tu, Chao Li, and Junwei Han. Base and meta: A new perspective on few-shot segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 399, + 545, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 422 + ], + "type": "text", + "content": "[21] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In ECCV, pages 734-750, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 422, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 476 + ], + "type": "text", + "content": "[22] Jianan Li, Xiaodan Liang, Yunchao Wei, Tingfa Xu, Jiashi Feng, and Shuicheng Yan. Perceptual generative adversarial networks for small object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1222-1230, 2017." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "text", + "content": "[23] Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented reploints for aerial object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1829-1838, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "type": "text", + "content": "[24] Tsung Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, pages 740-755, 2014." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 580, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 545, + 633 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2117-2125, 2017." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. pages 2980-2988, 2017." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[27] Guangtao Nie and Hua Huang. Multi-oriented object detection in aerial images with double horizontal rectangles. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4932-4944, 2022." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6326" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[28] Junhyug Noh, Wonho Bae, Wonhee Lee, Jinhwan Seo, and Gunhee Kim. Better to follow, follow to be better: Towards precise supervision of feature super-resolution for small object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 9725-9734, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[29] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28, 2015." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 217 + ], + "type": "text", + "content": "[30] Peize Sun et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 14449-14458, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 251 + ], + "type": "text", + "content": "[31] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. pages 9627-9636, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "type": "text", + "content": "[32] Thang Vu, Hyunjun Jang, Trung X Pham, and Chang Yoo. Cascade rpn: Delving into high-quality region proposal network with adaptive convolution. Advances in Neural Information Processing Systems, 32, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "type": "text", + "content": "[33] Jiaqi Wang, Kai Chen, Shuo Yang, Chen Change Loy, and Dahua Lin. Region proposal by guided anchoring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2965-2974, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "type": "text", + "content": "[34] Jinwang Wang, Wen Yang, Haowen Guo, Ruixiang Zhang, and Gui-Song Xia. Tiny object detection in aerial images. In IEEE International Conference on Pattern Recognition, pages 3791-3798, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "type": "text", + "content": "[35] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 7303-7313, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "type": "text", + "content": "[36] Jialian Wu, Chunluan Zhou, Qian Zhang, Ming Yang, and Junsong Yuan. Self-mimic learning for small-scale pedestrian detection. In Proceedings of the ACM International Conference on Multimedia, pages 2012-2020, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "text", + "content": "[37] Wei Wu, Hao Chang, Yonghua Zheng, Zhu Li, Zhiwen Chen, and Ziheng Zhang. Contrastive learning-based robust object detection under smoky conditions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4295-4302, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 600 + ], + "type": "text", + "content": "[38] Enze Xie, Jian Ding, Wenhai Wang, Xiaohang Zhan, Hang Xu, Peize Sun, Zhenguo Li, and Ping Luo. Detco: Unsupervised contrastive learning for object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 8392-8401, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "text", + "content": "[39] Xingxing Xie, Gong Cheng, Jiabao Wang, Xiwen Yao, and Junwei Han. Oriented r-cnn for object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 3520-3529, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[40] Chang Xu, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Rfla: Gaussian receptive based label assignment for tiny object detection. In Proceedings of the European Conference on Computer Vision, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[41] Yongchao Xu, Mingtao Fu, Qimeng Wang, Yukang Wang, Kai Chen, Gui-Song Xia, and Xiang Bai. Gliding vertex on" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 464 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "the horizontal bounding box for multi-oriented object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(4):1452-1459, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "text", + "content": "[42] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. Querydet: Cascaded sparse query for accelerating high-resolution small object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 13668-13677, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[43] Ze Yang, Shaohui Liu, Han Hu, Liwei Wang, and Stephen Lin. Repoints: Point set representation for object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 9657-9666, 2019." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "text", + "content": "[44] Jiahui Yu, Yuning Jiang, Zhangyang Wang, Zhimin Cao, and Thomas Huang. Unitbox: An advanced object detection network. In Proceedings of the ACM International Conference on Multimedia, pages 516-520, 2016." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "type": "text", + "content": "[45] Xuehui Yu, Yuqi Gong, Nan Jiang, Qixiang Ye, and Zhenjun Han. Scale match for tiny person detection. In IEEE Conference on Winter Conference on Applications of Computer Vision, pages 1257-1265, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "type": "text", + "content": "[46] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9759-9768, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "text", + "content": "[47] Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, and Stan Z Li. S3fd: Single shot scale-invariant face detector. In Proceedings of the IEEE International Conference on Computer Vision, pages 192-201, 2017." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 399, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 420 + ], + "type": "text", + "content": "[48] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "type": "text", + "content": "[49] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2020." + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6327" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Smoothness Similarity Regularization for Few-Shot GAN Adaptation/e521c479-803a-4c4d-bf48-e946e8872ca9_content_list.json b/2023/Smoothness Similarity Regularization for Few-Shot GAN Adaptation/e521c479-803a-4c4d-bf48-e946e8872ca9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b3194f30cf95fdcbb72a84ab9dddab2daf827698 --- /dev/null +++ b/2023/Smoothness Similarity Regularization for Few-Shot GAN Adaptation/e521c479-803a-4c4d-bf48-e946e8872ca9_content_list.json @@ -0,0 +1,1355 @@ +[ + { + "type": "text", + "text": "Smoothness Similarity Regularization for Few-Shot GAN Adaptation", + "text_level": 1, + "bbox": [ + 135, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vadim Sushko $^{1,2}$ Ruyu Wang $^{1}$ Juergen Gall $^{2,3}$", + "bbox": [ + 261, + 179, + 704, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Bosch Center for Artificial Intelligence $^{2}$ University of Bonn", + "bbox": [ + 267, + 200, + 702, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Lamarr Institute for Machine Learning and Artificial Intelligence", + "bbox": [ + 267, + 217, + 702, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "vad221@gmail.com ruyu.wang@de.bosch.com gall@iai.uni-bonn.de", + "bbox": [ + 181, + 237, + 785, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 279, + 312, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The task of few-shot GAN adaptation aims to adapt a pre-trained GAN model to a small dataset with very few training images. While existing methods perform well when the dataset for pre-training is structurally similar to the target dataset, the approaches suffer from training instabilities or memorization issues when the objects in the two domains have a very different structure. To mitigate this limitation, we propose a new smoothness similarity regularization that transfers the inherently learned smoothness of the pre-trained GAN to the few-shot target domain even if the two domains are very different. We evaluate our approach by adapting an unconditional and a class-conditional GAN to diverse few-shot target domains. Our proposed method significantly outperforms prior few-shot GAN adaptation methods in the challenging case of structurally dissimilar source-target domains, while performing on par with the state of the art for similar source-target domains.", + "bbox": [ + 75, + 305, + 473, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 573, + 207, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative adversarial networks (GANs) have been shown to be powerful at various image synthesis tasks [4, 28, 3, 13, 27, 26]. The success of these models is in large part enabled by the availability of large datasets for training, typically consisting of thousands of images. However, there are many applications and computer vision tasks such as one-shot or few-shot learning [1, 33], out-of-distribution detection [24], or long-tailed recognition tasks [8] where the number of available training images is very low.", + "bbox": [ + 75, + 598, + 468, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since training a GAN from scratch on very few samples does not perform well as shown in Fig. 1, a common strategy is to fine-tune a pre-trained GAN model on the few-shot dataset, typically employing additional regularization losses to penalize the degradation of the diversity [23, 37]. This approach, referred to as few-shot GAN adaptation, performs well when the target domain is structurally very similar to the dataset that has been used for pre-training, e.g., photographs vs. sketches of human faces. However, the performance drastically degrades in case of large dissimilarities between the source and target domain as shown in Fig. 1.", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/21e06df4d0f06758a31fb4db90ba718d82a6587a51c4aa78b1cfae86feea72a4.jpg", + "image_caption": [ + "Figure 1. Training a GAN model $G$ on a few-shot dataset (row 1) from scratch fails due to training instabilities (row 2). We thus aim to adapt a GAN $G_{s}$ that has been pre-trained on a large dataset like LSUN-Church (row 3) to the target few-shot dataset $(G_{t})$ . While fine-tuning [23] does not perform well either if source and target are dissimilar (row 4), our approach generates diverse and realistic images (row 5) by transferring the smoothness properties of $G_{s}$ ." + ], + "image_footnote": [], + "bbox": [ + 501, + 276, + 890, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Such dissimilarities are a major bottleneck of using GANs in other disciplines like medicine, production, or crop science, where there is a lack of large datasets due to privacy, confidentiality, or simply lack of data. Motivated by this fact, we extend the protocol for few-shot GAN adaptation by investigating also pairs of datasets that are very different like churches and shells as shown in Fig. 1.", + "bbox": [ + 496, + 656, + 890, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To improve few-shot GAN adaptation in the case of structurally dissimilar pairs, we propose a new GAN adaptation strategy. Firstly, we propose a new smoothness similarity regularization for the generator. Our key observation is that pre-trained GAN generators, regardless of the exact structure of objects in the pre-training dataset, learn well-structured and smooth latent spaces. For example, prior works demonstrated that various local shifts in the latent space can lead to interpretable and smooth transitions of", + "bbox": [ + 496, + 763, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7073", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "output images, such as translation of objects in the scene or changing their size [34, 9, 30]. As we show in our experiments, the proposed smoothness similarity regularization enables the transfer of this desirable property to other few-shot image domains without compromising the synthesis quality. Secondly, to overcome overfitting issues, we revisit the adversarial loss function of the discriminator and propose a simple yet efficient modification by computing the loss at different layers of the discriminator. This leads to the mitigation of overfitting and a more stabilized adaptation of the model to diverse target domains.", + "bbox": [ + 80, + 90, + 467, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our approach by adapting an unconditional [15] and a class-conditional GAN [2] to diverse few-shot target domains. Our model significantly outperforms previous state-of-the-art methods in image quality and diversity in the challenging case of dissimilar source and target domains, while performing on par with the state of the art on structurally similar dataset pairs. In summary, our contributions are as follows: (i) We extend the evaluation protocol for few-shot GAN adaptation by including new dataset pairs that are structurally much less similar than was considered in prior work. (ii) We propose a new smoothness similarity regularization, which enables diverse synthesis in the target domain by transferring the learned smoothness of a pretrained GAN. (iii) We revisit the adversarial loss function of the discriminator to stabilize few-shot GAN adaptation across diverse target domains. (iv) Our proposed model enables high-quality synthesis in the challenging case of dissimilar source and target domains, significantly outperforming prior methods. In addition, we show that our method can be applied to different classes of GAN architectures, including unconditional and class-conditional GAN models.", + "bbox": [ + 80, + 258, + 467, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 80, + 588, + 215, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the image generation problem in the low data regime, existing works mainly follow three research lines - one-shot, low-shot, and few-shot learning. One-shot generation methods [29, 31] focus on leveraging the internal patch distribution within a single image, however, their extension to capture the distribution of a small collection of images is non-trivial. In low-shot learning [41], several works [41, 12] proposed to mitigate the limited-data-induced overfitting issue by adapting data augmentations to the generative networks. Others [18, 5] stabilized the training process and reduced overfitting by revising the network design. Despite the promising performance in many low data regimes (typically having $100+$ images), these low-shot methods fail in the extremely few-shot setting (e.g., 10 images). Our work lies in the scope of few-shot learning.", + "bbox": [ + 80, + 613, + 467, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-shot image synthesis. Conventional few-shot learning aims at learning a discriminative classifier under limited data scenarios. In the context of image synthesis with GANs, the goal instead is to produce diverse new im", + "bbox": [ + 80, + 840, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ages from the learned distribution while preventing overfitting to the few training samples. A straightforward way is to treat it as a domain adaptation problem and incorporate the commonly used transfer learning technique, i.e., fine-tuning, to ease the need for data. However, naive fine-tuning (TGAN) [36] often suffers from overfitting and results in poor performance. Researchers proposed remedies such as mining suitable parts of the latent space before fine-tuning [35] or restricting weight updates, for example, updating only the BatchNorm parameters of the generator [22], penalizing drastic changes in important weights [17], or freezing the earliest layers of the discriminator (FreezeD) [20]. More recent works focused on introducing different regularizations to preserve specific knowledge from the pretrained model and prevent diversity degradation [42]. For example, CDC [23] proposed to preserve the pair-wise perceptual similarity between samples from the source domain and to transfer it to the target domain, while RSSA [37] designed a novel consistency term to align the structural information between source and target domains. Although the two aforementioned methods constitute the current state of the art in few-shot generative learning, their assumptions impose strong constraints on the structure of the few-shot target domain. As we show in experiments, they fail in the more challenging regime when the source and target domains are not restrictively similar. Most recently, [39] proposed to replace prior knowledge preservation criteria with adaptation-aware kernel modulation (AdAM), which relaxed the source-target proximity requirement of previous methods to some extent. In this work, we take a step further and introduce a new regularization term to preserve the generator's smoothness properties that are not limited to a specific domain, enabling successful adaptation between image domains of unprecedented structural dissimilarity.", + "bbox": [ + 503, + 90, + 888, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Smoothness of image generators. Smooth transitions in the latent space are an important property for generative models, where it is believed to be a sign of a well-conditioned generator. Models trained on large datasets naturally possess this property with or without explicit regularization [2, 15]. For example, StyleGANv2 [15] introduced a regularization based on the perceptual path length measure (PPL) [14], which encourages that a fixed-size step in the latent space results in a fixed-magnitude change in the image space. However, achieving a smooth mapping of the generator is difficult for few-shot image synthesis since there are not enough training samples. Thus, MixDL [16] sought to alleviate the \"staircase\" latent space interpolations, i.e., jumps between training samples, by introducing a continuous coefficient vector and enforcing smooth interpolations between training images. Although the two above regularizers aim to encourage smoother interpolations between training samples and thus mitigate mode collapse, they are not designed to take advantage of the available pre-training", + "bbox": [ + 503, + 613, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7074", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6f3dc01222538e80f247fe1115e4e0e41498a62ab8f02299a5901c836d68de64.jpg", + "image_caption": [ + "Figure 2. Given a pre-trained generator $G_{s}$ , the proposed smoothness similarity regularization preserves the learned smoothness of $G_{s}$ while adapting it to a target domain with very few images. To mitigate overfitting to the target domain, the discriminator loss utilizes features at various layers and automatically adjusts the impact of different semantic scales to the similarity of the source and target domain." + ], + "image_footnote": [], + "bbox": [ + 78, + 80, + 880, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "knowledge. In contrast, in this work we develop a new smoothness similarity regularization that leverages the well-structured latent space of a pre-trained GAN generator. In effect, our approach enables high-quality few-shot image synthesis by transferring smooth and realistic image transitions of pre-trained GANs to diverse few-shot domains.", + "bbox": [ + 75, + 282, + 470, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 387, + 168, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the task of few-shot GAN adaptation, we are given a small target dataset $T$ and a pre-trained GAN model, consisting of a discriminator $D$ and a generator $G_{s}$ , which produces an image $x = G_{s}(z)$ from a continuous input variable $z$ , such as a random noise vector or a continuous class embedding. The goal is to adapt the generator to the target dataset such that it generates diverse and realistic images in the domain of the target dataset as shown in Fig. 1. We denote the adapted target generator by $G_{t}$ .", + "bbox": [ + 75, + 412, + 468, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve few-shot synthesis with a high image quality and diversity, our model should adhere to the following two properties. Firstly, the generator $G_{t}$ should not only memorize and generate the target images, which will be addressed by the smoothness similarity regularization (Sec. 3.1). Secondly, the discriminator $D$ must avoid overfitting to the few target images in order to provide useful supervision for $G_{t}$ (Sec. 3.2). The overview of our method is shown in Fig. 2.", + "bbox": [ + 75, + 549, + 468, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Smoothness similarity regularization for $G_{t}$", + "text_level": 1, + "bbox": [ + 76, + 680, + 444, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In a low data regime like ours, $G_{t}$ can easily overfit to the target dataset $T$ and collapse to reproducing only the few modes represented in the training data. When walking in the latent space of such a generator, one would observe \"staircase\" patterns, where minor shifts in the latent space cause discontinuous transitions in the output image space (as shown in row 4 of Fig. 5). Naturally, to achieve a synthesis of high diversity, it is desirable for $G_{t}$ to avoid such discontinuities, as having smoother image transitions allows to generate intermediate samples that can exhibit novel features. Therefore, in our model we aim to encourage $G_{t}$ to produce smooth latent space interpolations, in which all the intermediate images are realistic.", + "bbox": [ + 75, + 703, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach is based on the observation that GANs trained on large datasets tend to have a well-structured latent space [34, 9, 30], in which different latent space directions can lead to smooth and interpretable image transitions. For example, in a generator pre-trained on a large dataset of churches, latent directions can emerge causing smooth zooming or translation of churches (see Fig. 2). Our observation is that the nature of such image transitions (e.g., zooming or translation) is remarkably general. Thus, we propose a regularizer that utilizes this smoothness property of the source generator $G_{s}$ as a cue while adapting it to another image domain, which can be very different from the domain that was used for pre-training. For example, as shown in Fig. 2, the same latent directions of churches can cause similar zooming or translation effects on shells.", + "bbox": [ + 496, + 282, + 890, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mathematically, the smoothness of the generator can be represented via a Jacobian matrix $J_{G^l}(z) = ||\\partial G^l (z) / \\partial z||$ , quantifying how the generator's intermediate features after the $l$ -th block change under local shifts in the latent space. As we want the same latent shift to cause perceptually similar image transitions in the source and target domains, we design a regularization term that brings the Jacobian matrices of $G_s^l$ and $G_t^l$ closer together. As the computation of full Jacobian matrices is expensive, we use an unbiased estimator of their products with a Gaussian vector [6, 15], which can be computed with standard back-propagation:", + "bbox": [ + 496, + 508, + 892, + 675 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nJ _ {G ^ {l}} ^ {T} (z) \\cdot y = \\mathbb {E} _ {(y) \\sim N (0, 1)} \\nabla_ {z} \\langle G ^ {l} (z), y \\rangle , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 683, + 890, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $y$ is a Gaussian tensor of the same shape as $G^l$ . Our smoothness similarity regularization is then expressed as:", + "bbox": [ + 496, + 712, + 890, + 742 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S S} = \\lambda_ {S S} \\cdot \\mathbb {E} _ {(z, y) \\sim N (0, 1)} \\| \\nabla_ {z} \\langle G _ {s} ^ {l} (z), y \\rangle - \\nabla_ {z} \\langle G _ {t} ^ {l} (z), y \\rangle \\| _ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 750, + 890, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\lambda_{SS}$ steers the impact of the regularizer. As shown in Fig. 2, the smoothness similarity regularization depends on both generators, but only $G_{t}$ is updated. It is interesting to note that the Jacobian matrix is also used for the path length regularization [15], which forces $J_{G}(z)$ to be orthogonal up to a global scale at any $z$ . While this alternative regularizer also induces some form of smoothness, it does not transfer the inherently learned smoothness", + "bbox": [ + 496, + 780, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7075", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of a pre-trained GAN. We show in Sec. 4.1 that it struggles to enforce the realism of intermediate images. Furthermore, our approach shares the motivation with some prior regularization approaches that use noise perturbations to enforce diversity [23, 37]. In contrast to Eq. 2, these approaches incorporate non-gradient components, e.g., assuming similarity of images $G_{s}(z)\\leftrightarrow G_{t}(z)$ or distributions $d(G_{t}(z_{1}),G_{t}(z_{2}))\\leftrightarrow d(G_{s}(z_{1}),G_{s}(z_{2}))$ . As such assumptions are violated when source and target domains are dissimilar, they perform poorly compared to our smoothness similarity regularization $\\mathcal{L}_{SS}$ as shown in the experiments.", + "bbox": [ + 75, + 90, + 472, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Revisiting the $D$ adversarial loss", + "text_level": 1, + "bbox": [ + 76, + 265, + 362, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To identify what kind of image transitions look realistic for the target domain, $G_{t}$ requires strong supervision from the discriminator on image realism at different semantic scales. This includes the colors and textures of objects, as well as object shapes, especially if their distribution is different from the shapes of objects in the source domain. Learning the concept of image realism in low data regimes is, however, challenging due to the problem of overfitting.", + "bbox": [ + 75, + 287, + 468, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Typically, a GAN discriminator consists of several consecutive blocks $\\{D^i\\}_{i = 1}^N$ and computes for each given image $x$ a real/fake logit after the last block $l = s^{N}\\circ D^{N}(x)$ where $s^N$ is a final processing layer such as a convolution. When adapting such a discriminator to a very small dataset, it is prone to memorizing the training set [32], leading to mode collapse and poor diversity of synthesized images [23]. A possible solution [23, 37] to overcome memorization is to use variants of the PatchGAN discriminator [11], discarding the latest discriminator layers: $l = s^{k}\\circ D^{k}(x),k < N$ . This solution allows to adapt colors and textures of generated images to the target domain while avoiding the memorization problem. However, it naturally has a limited capacity to learn more high-level semantic scene properties such as the shapes of objects, which we show in the experiments.", + "bbox": [ + 75, + 409, + 470, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to avoid memorization, and yet to balance the adaptation of colors, textures, and shapes of generated objects to a new domain, we hypothesize that a more flexible attention to different levels of image realism is required by the discriminator. To this end, we perform a simple yet efficient modification to the loss function of the discriminator. Given a discriminator $\\{D^i\\}_{i = 1}^N$ and its adversarial loss function $\\mathcal{L}_D(l)$ used for pre-training (e.g., cross-entropy or hinge loss), we design the discriminator to produce real/fake logits after each discriminator's block, and correspondingly compute the loss as the average across all blocks:", + "bbox": [ + 75, + 650, + 468, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a l l} (x) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\mathcal {D}} \\left[ l ^ {i} (x) \\right], \\quad l ^ {i} (x) = s ^ {i} \\circ D ^ {i} (x). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 821, + 468, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the new objective, $D$ is given more freedom to utilize the features extracted at different scales to compute the", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "loss. Our finding is that $D$ dynamically adapts the magnitude of the loss at each scale to the target domain, without explicit supervision (see Fig. 6). Consequently, we observe a strong overall stabilization effect on the adaptation performance across diverse source-target dataset pairs.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 184, + 633, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To demonstrate that our approach for few-shot GAN adaptation can be applied to unconditional and class-conditional GANs, we selected for each category a popular GAN architecture: unconditional StyleGANv2 [15] and class-conditional BigGAN [2]. For both models, we test our approach on a variety of source-target domain pairs. We focus on 10-shot target adaptation in the main paper, but we provide results for 1-shot and 5-shot adaptation in the supplementary material. For fair comparisons with prior works, most of our ablations and comparisons are conducted with StyleGANv2.", + "bbox": [ + 496, + 210, + 890, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Adaptation of unconditional GAN", + "text_level": 1, + "bbox": [ + 498, + 388, + 797, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. In contrast to previous works that mostly considered pairs of similar datasets like Face $\\rightarrow$ Sketch and Face $\\rightarrow$ Sunglasses, we extend the protocol by including structurally dissimilar pairs of source and target domains, which is a more challenging task and is our primary interest. As source generators, we use StyleGANv2 checkpoints pretrained on FFHQ [14], LSUN-Church, and LSUN-Horse [38]. For the target datasets, we selected 10-shot subsets of various commonly used few-shot datasets, such as Anime-Face, Shells, or Pokemons [41, 18]. Results on more datasets are shown in the supplementary material.", + "bbox": [ + 496, + 414, + 890, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training details. We fine-tune StyleGANv2 using the $\\mathcal{L}_{SS}$ and $\\mathcal{L}_{all}$ loss terms as presented in Sec. 3. For the smoothness similarity regularization, we use the intermediate features $G^{l}$ at resolution $(32\\times 32)$ and set $\\lambda_{SS} = 5.0$ . We follow [23] in choosing all the other hyperparameters, such as image resolution $(256\\times 256)$ , learning rates, and batch size. Our experiments across all datasets use the same model configuration and set of hyperparameters.", + "bbox": [ + 496, + 580, + 890, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Baselines. We compare our method to most recent few-shot GAN adaptation approaches: TGAN [36], FreezeD [20], CDC [23], RSSA [37], and AdAM [39]. In addition, we compare our proposed smoothness similarity regularizer $\\mathcal{L}_{SS}$ to other regularization techniques: path length regularization (PPL) [15] and MixDL [16].", + "bbox": [ + 496, + 703, + 890, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation. In low data regimes, it is necessary to judge results both in quality and diversity aspects, as there is a trade-off between them [25, 32]. We measure the quality with FID [10] between a held-out validation set and a generated set of the same size. Following [23], we evaluate diversity with the intra-LPIPS, clustering generated images according to their nearest training samples and computing", + "bbox": [ + 496, + 795, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7076", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0ecb19495f802ecc0a2969200d6588e807c13f045f459d9dadef380b8bc339ce.jpg", + "image_caption": [ + "10-shot Anime-Faces" + ], + "image_footnote": [], + "bbox": [ + 96, + 118, + 223, + 344 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/755207d03d3422ecc244f636825bc5b35919ee3a036e0f615e21e6a62b5b9732.jpg", + "image_caption": [ + "TAN Freezed CDC RSSA AdAM Ours Source TGAN Freezed CDC RSSA AdAM Ours Source" + ], + "image_footnote": [], + "bbox": [ + 236, + 85, + 857, + 400 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e381ce5a57f8934ec2c9d3cd42500c3f3b6555e7a51c20067500f40114d8ae19.jpg", + "image_caption": [ + "10-shot Shells" + ], + "image_footnote": [], + "bbox": [ + 96, + 438, + 223, + 662 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/04c4721c37231e94ed46e5170d28281e7f0fb359a385802ce1dc6501d079b875.jpg", + "image_caption": [ + "Figure 3. Visual comparison to prior methods on Face $\\rightarrow$ Anime and Church $\\rightarrow$ Shells, the source-target dataset pairs with a dissimilar structure (e.g., shapes of objects). In this challenging regime, we observe that prior methods suffer from training instabilities, memorization issues, or inability to adapt the shapes of objects to the new domain. In contrast, our method generates images that look realistic, flexibly combine features of different target images, and transfer the variation of images from the source domain to the target domain." + ], + "image_footnote": [], + "bbox": [ + 236, + 402, + 857, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the average LPIPS [40] of all the clusters. We train all models for 30k epochs in case of dissimilar domain pairs and for 5k on closer domains, evaluating metrics every 1k epochs. Final checkpoints in all experiments correspond to best FID.", + "bbox": [ + 76, + 787, + 470, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results with dissimilar source-target domains. We first present our results on the source-target domain pairs with dissimilar structure: Face $\\rightarrow$ Anime, Church $\\rightarrow$ Shells,", + "bbox": [ + 76, + 854, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and Horse $\\rightarrow$ Pokemon (see Fig. 3 and supplementary material). Our general observation from Fig. 3 is that in this challenging regime prior methods suffer either from training instabilities, memorization issues, or inability to adapt the shape of objects to the new domain. For example, for Face $\\rightarrow$ Anime, despite an apparent correspondence between the two domains, none of the prior methods success-", + "bbox": [ + 496, + 787, + 890, + 893 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7077", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5b307de5acf42f68a394289aca640efaafb14b63e815e25ccf920ca7a8a3e69a.jpg", + "image_caption": [ + "Figure 4. Visual comparison to most recent prior methods on Face $\\rightarrow$ Sketch and Church $\\rightarrow$ Sunglasses, the dataset pairs depicting similar image domains. In this regime, our method performs on par with previous state of the art. See Table 2 for a quantitative comparison." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 483, + 343 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b62d35529d84d6944182a608b06629941dee6a06949cd931ac22659e88d5d19c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 90, + 890, + 343 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/eb65083a6f086631adfccd00a3db9254374e9c26c0f10eda34a10982c704a398.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFace→AnimeChurch→ShellsHorse→Pokemons
FID↓LPIPS↑FID↓LPIPS↑FID↓LPIPS↑
TGAN [36]153.20.29205.30.22115.00.52
FreezeD [20]112.40.22180.80.27123.30.49
CDC [23]140.20.50187.90.48109.50.55
RSSA [37]133.20.37182.40.44117.30.54
AdAM [39]116.40.42152.40.28106.50.55
Ours97.30.57140.50.5384.10.61
", + "bbox": [ + 75, + 386, + 477, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "fully transfers the distribution of head poses to the anime style, e.g., overfitting too strongly to the 10 provided samples (FreezeD), failing to adapt the shape of faces to the style of anime (CDC), or not generating high-quality animenefaces due to instabilities (TGAN, RSSA, AdAM). Similarly, for $\\text{Church} \\rightarrow \\text{Shells}$ , we observe that prior methods produce only copies of the example shells (FreezeD, AdAM), generate shells of unrealistic church-like shapes (CDC, RSSA), or suffer from instabilities (TGAN). In contrast, our method achieves high-quality synthesis, in which the generated images (i) look like realistic anime-faces and shells; (ii) flexibly combine features observed in different target images (e.g., anime hair color can be combined with various eye colors or background styles); and (iii) meaningfully transfer the variation of images from the source domain (e.g., generated shells adjust to the positions and shapes of churches).", + "bbox": [ + 75, + 551, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The quantitative comparison in Table 1 confirms our analysis, where our method achieves the best quality and diversity scores across all datasets. We note a high average relative improvement of more than $18\\%$ and $11\\%$ in FID and LPIPS compared to the highest scores achieved by prior methods. Overall, we conclude that our method significantly improves over prior works on few-shot GAN", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c3b7d2a77f832660f7911d73017159830eb5015a2b0b9692870d2296e039d5a6.jpg", + "table_caption": [ + "Table 1. Comparison of the adaptation performance in case of dissimilar source-target domains. Bold denotes best performance." + ], + "table_footnote": [], + "table_body": "
MethodFace→SketchFace→Sunglasses
FID↓LPIPS↑FID↓LPIPS↑
TGAN [36]54.20.3836.80.56
FreezeD [20]48.80.3232.00.59
CDC [23]54.20.4030.50.59
RSSA [37]61.40.4536.30.58
AdAM [39]56.30.3731.10.60
Ours45.20.4427.50.60
", + "bbox": [ + 545, + 386, + 841, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison in case of structurally close source-target domains. Bold denotes best performance.", + "bbox": [ + 500, + 510, + 888, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "adaptation with dissimilar source and target domains.", + "bbox": [ + 500, + 553, + 852, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results with close source-target domains. Next, we follow the evaluation of prior works and compare the models on similar source and target domains, such as adaptation of human faces to a different style. The visual results for Face $\\rightarrow$ Sketch and Face $\\rightarrow$ Sunglasses are shown in Fig. 4. Our method successfully performs the few-shot adaptation in this setting, adapting the colors and textures of faces to the gray-scale sketch domain, or adding a novel attribute (sunglasses). We note that our method is not explicitly designed to transfer all the details of a face from the source domain, thus changes in the generated images like facial hair are expected. Yet, we observe that our method generally does not lose distinctive features of faces in source images, performing on par with previous state-of-the-art methods. The quantitative comparison is provided in Table $2^{1}$ : on both datasets our method achieves the best FID scores and performs on par with the best performer in LPIPS.", + "bbox": [ + 496, + 569, + 890, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablations. We demonstrate the importance of our proposed loss terms in Fig. 5, which shows latent space interpolations of trained models and their similarity to the pre", + "bbox": [ + 496, + 827, + 890, + 873 + ], + "page_idx": 5 + }, + { + "type": "aside_text", + "text": "Ours AdAM RSSA CDC Source Real", + "bbox": [ + 68, + 93, + 86, + 334 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{1}$ FID evaluation differs from prior works (discussed in suppl. material).", + "bbox": [ + 517, + 886, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7078", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4755dec36406e5b280bc981bfe1ca884e4f1926cf0de83e7373ff4f0e24b556f.jpg", + "image_caption": [ + "Figure 5. Latent space interpolations of the source generator and the ablation models from Tables 3-4. Leftmost and rightmost columns show the used $D$ loss and $G$ smoothness regularization." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 465, + 411 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/59945f8af071a35a73994ada6762141a9fd9fecc7277d0fbe4aabbff3496a4a9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
D lossSmooth reg. for GFace→AnimeChurch→Shells
FID↓LPIPS↑FID↓LPIPS↑
StyleGANv2-178.00.21243.80.17
StyleGANv2SS (ours)180.70.61252.80.62
PatchGAN [23]-145.20.37183.10.31
PatchGAN [23]SS (ours)132.20.55184.20.56
Lall (ours)-116.40.36175.40.43
Lall (ours)SS (ours)97.30.57140.50.53
", + "bbox": [ + 94, + 460, + 447, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "trained source model $G_{s}$ (row 1). Firstly, we note that the plain StyleGANv2 model (row 2) suffers from instabilities in our low data regime, achieving poor image quality and diversity and having \"staircase\"-like latent space interpolations. Applying $\\mathcal{L}_{SS}$ without $\\mathcal{L}_{all}$ (row 3) helps to achieve diverse synthesis with smooth interpolations, but it is not enough to achieve good image quality. On the other hand, using $\\mathcal{L}_{all}$ (row 4) helps to overcome instabilities and improve image quality, but it cannot maintain smooth interpolations and high diversity. Finally, our full model (row 5) allows a higher-quality, diverse synthesis with smooth and realistic latent space interpolations. Note how the image transitions mimic the behaviour of the source model (churches and shells change shapes and positions similarly), allowing to achieve diverse and realistic synthesis.", + "bbox": [ + 75, + 612, + 468, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The effect of $\\mathcal{L}_{all}$ is further demonstrated in Fig. 6, where we show the contribution of different $D$ blocks to the adversarial loss at different epochs. We note the ability of the discriminator to identify correct loss contributions", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/598b03f082b3faefeeaa68bdfddecd7ce535e606436fbd31756c97770494dad4.jpg", + "image_caption": [ + "Figure 6. The contribution of features at different $D$ blocks to the adversarial loss function $\\mathcal{L}_{all}$ . For two closer image domains (the left plot), the network concentrates mostly on earlier layers to compute the loss, while for less similar domains the network learns to use the later layers representing more high-level $D$ features." + ], + "image_footnote": [], + "bbox": [ + 498, + 87, + 890, + 190 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a11703e5a96ab9c17b011bd9f47f246a68a6d1cc24de1e1a4b70dff39b3ca586.jpg", + "table_caption": [ + "Table 3. Impact of ${\\mathcal{L}}_{\\text{all }}$ and ${\\mathcal{L}}_{SS}$ . Bold denotes best performance." + ], + "table_footnote": [], + "table_body": "
D lossSmooth reg. for GFace→AnimeChurch→Shells
FID↓LPIPS↑FID↓LPIPS↑
Lall(ours)-116.40.36175.40.43
Lall(ours)PPL [14]107.80.46179.40.44
Lall(ours)MixDL [16]105.90.50150.40.51
Lall(ours)SS (ours)97.30.57140.50.53
", + "bbox": [ + 519, + 272, + 864, + 362 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Comparison of smoothness similarity regularization $\\mathcal{L}_{SS}$ with other regularizers. Bold denotes best performance.", + "bbox": [ + 500, + 363, + 890, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "adaptively for different source-target domain pairs. For example for Face $\\rightarrow$ Anime, the network concentrates mostly on the earliest $D$ blocks to adapt the colors and textures of faces to a new style. In contrast, for the more distant domains Church $\\rightarrow$ Shells, the network learns to attribute a higher weight to the later blocks to also adapt higher-level features, such as shapes of objects. In effect, we observe a stabilized adaptation of colors, textures, and shapes of objects across diverse source-target pairs. Using PatchGAN [23] as discriminator loss does not achieve such a balance as it focuses mostly on lower-scale features (row 6 in Fig. 5).", + "bbox": [ + 496, + 412, + 892, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our observations are confirmed by the quantitative study in Table 3: without $\\mathcal{L}_{SS}$ the model does not achieve high diversity (high LPIPS), while $\\mathcal{L}_{all}$ is necessary for high image quality (low FID). We conclude that both our proposed loss terms are important to achieve high-quality synthesis. More ablations on $\\mathcal{L}_{SS}$ and $\\mathcal{L}_{all}$ can be found in the supplementary material.", + "bbox": [ + 496, + 580, + 890, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Lastly, Table 4 provides a comparison of our proposed $\\mathcal{L}_{SS}$ loss term to other regularizers: path length regularization (PPL) [14] and MixDL [16]. While all regularizers help to achieve smoother latent space interpolations and thus improve the quality and diversity metrics, our smoothness similarity regularization enables the highest performance in both FID and LPIPS. While our approach transfers the learned smoothness of the source generator to the target domain, PPL and MixDL resort to gradually interpolating between the provided training samples, which leads to latent space interpolations that either look unrealistic or lack diversity (rows 7-8 in Fig. 5). This demonstrates that transferring smoothness from a pre-trained generator is beneficial to enforce image transitions that are realistic and diverse.", + "bbox": [ + 496, + 688, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7079", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/89f2496245519eeb260bee46876711f937542fa6af06bf6e628c2eac6006e71b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 76, + 104, + 480, + 137 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b1024f67326930cdb76f583a23ccb5cb80713e8a6af5592c5537c890752315b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 485, + 104, + 888, + 136 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7ec0596780f910ed1bc441fbc7016076181f7b5bdb666831af14b45941a3cc73.jpg", + "image_caption": [ + "10-Shot adaptation results of our method on class-conditional BigGAN [2], pre-trained on ImageNet", + "Figure 7. 10-shot adaptation results for the class-conditional BigGAN [2] pre-trained on ImageNet. While simple fine-tuning (FT) suffers from training instabilities and mode collapse, our method helps to achieve much higher image quality and diversity, transferring smooth and realistic image transitions from the source domain, e.g., objects smoothly changing their locations, size, and shape." + ], + "image_footnote": [], + "bbox": [ + 73, + 155, + 898, + 383 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7804531f3a34bc00220a25f7776482a41f8eb282bcd7943d27404069c4818f8f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
D lossSmooth reg. for GImageNet→FlowersImageNet→Pokemons
FID↓LPIPS↑FID↓LPIPS↑
BigGAN-213.30.29226.80.15
BigGANSS (ours)225.60.47208.30.47
Lall (ours)-123.90.28129.40.27
Lall (ours)SS (ours)106.40.5589.60.56
", + "bbox": [ + 76, + 435, + 475, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation on the performance when adapting the class-conditional BigGAN model [2] pre-trained on ImageNet.", + "bbox": [ + 75, + 530, + 468, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Adaptation of class-conditional GAN", + "text_level": 1, + "bbox": [ + 76, + 573, + 397, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our approach is not limited to unconditional GANs, but it can also be applied to a class-conditional GAN model. We selected BigGAN [2] for our experiments as it is a popular backbone architecture for class-conditional image synthesis on ImageNet [7]. We make two modifications to enable the adaptation of the model to unconditional few-shot datasets. Firstly, we remove the conditioning of the discriminator via the projection layer [19]. Secondly, we treat the generator's learned continuous class embedding as part of the latent space, thus sampling a Gaussian vector in the joint noise-class space at each fine-tuning epoch. This way, the generator produces an image based on a single input vector in an unconditional fashion. We then fine-tune the pre-trained model using our loss terms $\\mathcal{L}_{SS}$ and $\\mathcal{L}_{all}$ as presented in Sec. 3. We use image resolution $256\\times 256$ and batch size of 32. The hyperparameters for $\\mathcal{L}_{SS}$ are the same as for StyleGANv2: intermediate features $G^{l}$ at resolution $(32\\times 32)$ and $\\lambda_{SS} = 5.0$ . We train for 30k epochs and select checkpoints by best FID.", + "bbox": [ + 75, + 597, + 468, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. As the source generator, we use the Big-", + "bbox": [ + 96, + 885, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GAN checkpoint pre-trained on class-conditional ImageNet at resolution $256 \\times 256$ . We demonstrate 10-shot adaptation results with two commonly used few-shot generation datasets: Oxford-Flowers [21] and Pokémon [18]. We use the same model configuration for both datasets.", + "bbox": [ + 498, + 436, + 892, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. Fig. 7 demonstrates latent space interpolations of the source and target generators. We note that a simple fine-tuning of BigGAN suffers from training instabilities and mode collapse. In contrast, our method successfully adapts BigGAN to generate diverse images in the target domains. We highlight that our method transfers smooth and realistic image transitions from the well-learned BigGAN's noise-class space, despite significant dissimilarities between ImageNet and the few-shot datasets, in particular Pokemons. For example, it can be noticed how the latent space interpolations in the target domains mimic the source domain, e.g., the generated flowers andPokemons change their position and size similarly to dogs and wolves $(5^{\\mathrm{th}}-$ $10^{\\mathrm{th}}$ columns in Fig. 7) or stretch their shape to mimic the proportions of busses $(11^{\\mathrm{th}} - 14^{\\mathrm{th}}$ columns).", + "bbox": [ + 496, + 518, + 892, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5 shows the importance of our proposed loss terms. Our observations are consistent with the ablations with StylGANv2: $\\mathcal{L}_{all}$ is necessary to avoid instabilities and achieve a good image quality (low FID), while $\\mathcal{L}_{SS}$ is required to achieve smooth latent space interpolations and good diversity (high LPIPS). We conclude that our method successfully extends to the adaptation of class-conditional models, where target domains benefit from the rich noise-class space learned on a multi-class dataset such as ImageNet. More details and results are provided in the supplementary material.", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Real samples - 10-Shot Flowers", + "bbox": [ + 176, + 89, + 392, + 103 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Real samples - 10-Shot Pokemons", + "bbox": [ + 576, + 89, + 805, + 103 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7080", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 89, + 194, + 104 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we presented a new method for few-shot adaptation of GAN models. It transfers the smooth latent space of a pre-trained GAN, which was trained on a large dataset, to a new domain with very few images. We addressed the case of few-shot GAN adaptation when the source and target domains are structurally dissimilar, which is a common issue in applications. Our extensive results demonstrate that in this setting our approach outperforms previous works in terms of image quality and diversity.", + "bbox": [ + 75, + 114, + 472, + 252 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 76, + 263, + 236, + 279 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The work has been supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC 2070 -390732324 and the ERC Consolidator Grant FORHUE (101044724). We thank Jinhui Yi for providing and analyzing the sugar beet data used in supplementary experiments.", + "bbox": [ + 75, + 287, + 468, + 378 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 388, + 174, + 405 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-shot segmentation without meta-learning: A good transductive inference is all you need? In Conference on Computer Vision and Pattern Recognition (CVPR), 2021.", + "[2] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In International Conference on Learning Representations (ICLR), 2019.", + "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[4] Yunjoy Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020.", + "[5] Kaiwen Cui, Jiaxing Huang, Zhipeng Luo, Gongjie Zhang, Fangneng Zhan, and Shijian Lu. Genco: Generative constraining for generative adversarial networks with limited data. In Conference on Artificial Intelligence (AAAI), 2021.", + "[6] Yann Dauphin, Harm De Vries, and Yoshua Bengio. Equilibrated adaptive learning rates for non-convex optimization. In Advances in Neural Information Processing Systems (NeurIPS), 2015.", + "[7] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In Conference on Computer Vision and Pattern Recognition (CVPR), 2009.", + "[8] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + ], + "bbox": [ + 86, + 415, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In Advances in Neural Information Processing Systems (NeurIPS), 2020.", + "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems (NeurIPS), 2017.", + "[11] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017.", + "[12] Tero Karras, Miika Aittala, Janne Hellsten, S. Laine, J. Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Advances in Neural Information Processing Systems (NeurIPS), 2020.", + "[13] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Advances in Neural Information Processing Systems (NeurIPS), 2021.", + "[14] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019.", + "[15] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020.", + "[16] Chaerin Kong, Jeesoo Kim, Donghoon Han, and Nojun Kwak. Few-shot image generation with mixup-based distance learning. In European Conference on Computer Vision (ECCV), 2022.", + "[17] Yijun Li, Richard Zhang, Jingwan Cynthia Lu, and Eli Shechtman. Few-shot image generation with elastic weight consolidation. In Advances in Neural Information Processing Systems (NeurIPS), 2020.", + "[18] Bingchen Liu, Yizhe Zhu, Kunpeng Song, and Ahmed Elgammal. Towards faster and stabilized gan training for high-fidelity few-shot image synthesis. In International Conference on Learning Representations (ICLR), 2021.", + "[19] Takeru Miyato and Masanori Koyama. cGANs with projection discriminator. In International Conference on Learning Representations (ICLR), 2018.", + "[20] Sangwoo Mo, Minsu Cho, and Jinwoo Shin. Freeze discriminator: A simple baseline for fine-tuning gans. In CVPR AI for Content Creation Workshop, 2020.", + "[21] Maria-Elena Nilsback and Andrew Zisserman. A visual vocabulary for flower classification. In Conference on Computer Vision and Pattern Recognition (CVPR), 2006.", + "[22] Atsuhiro Noguchi and T. Harada. Image generation from small datasets via batch statistics adaptation. In International Conference on Computer Vision (ICCV), 2019.", + "[23] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7081", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Jie Ren, Peter J Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark Depristo, Joshua Dillon, and Balaji Lakshminarayanan. Likelihood ratios for out-of-distribution detection. Advances in Neural Information Processing Systems (NeurIPS), 2019.", + "[25] Esther Robb, Wen-Sheng Chu, Abhishek Kumar, and Jia-Bin Huang. Few-shot adaptation of generative adversarial networks. arXiv:2010.11943, 2021.", + "[26] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis. arXiv preprint arXiv:2301.09515, 2023.", + "[27] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleganxI: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH, 2022.", + "[28] Edgar Schonfeld, Vadim Sushko, Dan Zhang, Juergen Gall, Bernt Schiele, and Anna Khoreva. You only need adversarial supervision for semantic image synthesis. In International Conference on Learning Representations (ICLR), 2021.", + "[29] Tamar Rott Shaham, Tali Dekel, and T. Michaeli. SinGAN: Learning a generative model from a single natural image. In International Conference on Computer Vision (ICCV), 2019.", + "[30] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021.", + "[31] Vadim Sushko, Juergen Gall, and Anna Khoreva. One-Shot GAN: Learning to generate samples from single images and videos. In Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2021.", + "[32] Vadim Sushko, Dan Zhang, Juergen Gall, and Anna Khoreva. Learning to generate novel scene compositions from single images and videos. arXiv:2103.13389, 2021.", + "[33] Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. Transactions on Pattern Analysis and Machine Intelligence, 2020.", + "[34] Andrey Voynov and Artem Babenko. Unsupervised discovery of interpretable directions in the gan latent space. In International conference on machine learning (ICML), 2020.", + "[35] Yaxing Wang, Abel Gonzalez-Garcia, David Berga, L. Herranz, F. Khan, and Joost van de Weijer. Minegan: Effective knowledge transfer from gans to target domains with few images. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020.", + "[36] Yaxing Wang, Chenshen Wu, L. Herranz, Joost van de Weijer, Abel Gonzalez-Garcia, and B. Raducanu. Transferring gans: generating images from limited data. In European Conference on Computer Vision (ECCV), 2018.", + "[37] Jiayu Xiao, Liang Li, Chaofei Wang, Zheng-Jun Zha, and Qingming Huang. Few shot generative model adaption via relaxed spatial structural alignment. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[38] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv:1506.03365, 2015." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Zhao Yunqing, Keshigeyan Chandrasegaran, Milad Abdol-lahzadeh, and Ngai-man Cheung. Few-shot image generation via adaptation-aware kernel modulation. In Advances in Neural Information Processing Systems (NeurIPS), 2022.", + "[40] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Conference on Computer Vision and Pattern Recognition (CVPR), 2018.", + "[41] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. In Conference on Neural Information Processing Systems (NeurIPS), 2020.", + "[42] Yunqing Zhao, Henghui Ding, Houjing Huang, and Ngai-Man Cheung. A closer look at few-shot image generation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + ], + "bbox": [ + 501, + 90, + 892, + 316 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7082", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Smoothness Similarity Regularization for Few-Shot GAN Adaptation/e521c479-803a-4c4d-bf48-e946e8872ca9_model.json b/2023/Smoothness Similarity Regularization for Few-Shot GAN Adaptation/e521c479-803a-4c4d-bf48-e946e8872ca9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0ddd5c84718762f92e4851decf4d0f125ac22698 --- /dev/null +++ b/2023/Smoothness Similarity Regularization for Few-Shot GAN Adaptation/e521c479-803a-4c4d-bf48-e946e8872ca9_model.json @@ -0,0 +1,1848 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.131, + 0.834, + 0.154 + ], + "angle": 0, + "content": "Smoothness Similarity Regularization for Few-Shot GAN Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.18, + 0.705, + 0.2 + ], + "angle": 0, + "content": "Vadim Sushko\\(^{1,2}\\) Ruyu Wang\\(^{1}\\) Juergen Gall\\(^{2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.201, + 0.703, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\) Bosch Center for Artificial Intelligence \\(^{2}\\) University of Bonn" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.218, + 0.703, + 0.234 + ], + "angle": 0, + "content": "\\(^{3}\\)Lamarr Institute for Machine Learning and Artificial Intelligence" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.238, + 0.786, + 0.251 + ], + "angle": 0, + "content": "vad221@gmail.com ruyu.wang@de.bosch.com gall@iai.uni-bonn.de" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.28, + 0.313, + 0.294 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.474, + 0.563 + ], + "angle": 0, + "content": "The task of few-shot GAN adaptation aims to adapt a pre-trained GAN model to a small dataset with very few training images. While existing methods perform well when the dataset for pre-training is structurally similar to the target dataset, the approaches suffer from training instabilities or memorization issues when the objects in the two domains have a very different structure. To mitigate this limitation, we propose a new smoothness similarity regularization that transfers the inherently learned smoothness of the pre-trained GAN to the few-shot target domain even if the two domains are very different. We evaluate our approach by adapting an unconditional and a class-conditional GAN to diverse few-shot target domains. Our proposed method significantly outperforms prior few-shot GAN adaptation methods in the challenging case of structurally dissimilar source-target domains, while performing on par with the state of the art for similar source-target domains." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.574, + 0.208, + 0.589 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Generative adversarial networks (GANs) have been shown to be powerful at various image synthesis tasks [4, 28, 3, 13, 27, 26]. The success of these models is in large part enabled by the availability of large datasets for training, typically consisting of thousands of images. However, there are many applications and computer vision tasks such as one-shot or few-shot learning [1, 33], out-of-distribution detection [24], or long-tailed recognition tasks [8] where the number of available training images is very low." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Since training a GAN from scratch on very few samples does not perform well as shown in Fig. 1, a common strategy is to fine-tune a pre-trained GAN model on the few-shot dataset, typically employing additional regularization losses to penalize the degradation of the diversity [23, 37]. This approach, referred to as few-shot GAN adaptation, performs well when the target domain is structurally very similar to the dataset that has been used for pre-training, e.g., photographs vs. sketches of human faces. However, the performance drastically degrades in case of large dissimilarities between the source and target domain as shown in Fig. 1." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.277, + 0.891, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.55, + 0.894, + 0.648 + ], + "angle": 0, + "content": "Figure 1. Training a GAN model \\(G\\) on a few-shot dataset (row 1) from scratch fails due to training instabilities (row 2). We thus aim to adapt a GAN \\(G_{s}\\) that has been pre-trained on a large dataset like LSUN-Church (row 3) to the target few-shot dataset \\((G_{t})\\). While fine-tuning [23] does not perform well either if source and target are dissimilar (row 4), our approach generates diverse and realistic images (row 5) by transferring the smoothness properties of \\(G_{s}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.764 + ], + "angle": 0, + "content": "Such dissimilarities are a major bottleneck of using GANs in other disciplines like medicine, production, or crop science, where there is a lack of large datasets due to privacy, confidentiality, or simply lack of data. Motivated by this fact, we extend the protocol for few-shot GAN adaptation by investigating also pairs of datasets that are very different like churches and shells as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.902 + ], + "angle": 0, + "content": "To improve few-shot GAN adaptation in the case of structurally dissimilar pairs, we propose a new GAN adaptation strategy. Firstly, we propose a new smoothness similarity regularization for the generator. Our key observation is that pre-trained GAN generators, regardless of the exact structure of objects in the pre-training dataset, learn well-structured and smooth latent spaces. For example, prior works demonstrated that various local shifts in the latent space can lead to interpretable and smooth transitions of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7073" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.257 + ], + "angle": 0, + "content": "output images, such as translation of objects in the scene or changing their size [34, 9, 30]. As we show in our experiments, the proposed smoothness similarity regularization enables the transfer of this desirable property to other few-shot image domains without compromising the synthesis quality. Secondly, to overcome overfitting issues, we revisit the adversarial loss function of the discriminator and propose a simple yet efficient modification by computing the loss at different layers of the discriminator. This leads to the mitigation of overfitting and a more stabilized adaptation of the model to diverse target domains." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.259, + 0.468, + 0.573 + ], + "angle": 0, + "content": "We evaluate our approach by adapting an unconditional [15] and a class-conditional GAN [2] to diverse few-shot target domains. Our model significantly outperforms previous state-of-the-art methods in image quality and diversity in the challenging case of dissimilar source and target domains, while performing on par with the state of the art on structurally similar dataset pairs. In summary, our contributions are as follows: (i) We extend the evaluation protocol for few-shot GAN adaptation by including new dataset pairs that are structurally much less similar than was considered in prior work. (ii) We propose a new smoothness similarity regularization, which enables diverse synthesis in the target domain by transferring the learned smoothness of a pretrained GAN. (iii) We revisit the adversarial loss function of the discriminator to stabilize few-shot GAN adaptation across diverse target domains. (iv) Our proposed model enables high-quality synthesis in the challenging case of dissimilar source and target domains, significantly outperforming prior methods. In addition, we show that our method can be applied to different classes of GAN architectures, including unconditional and class-conditional GAN models." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.589, + 0.216, + 0.604 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.614, + 0.468, + 0.839 + ], + "angle": 0, + "content": "To address the image generation problem in the low data regime, existing works mainly follow three research lines - one-shot, low-shot, and few-shot learning. One-shot generation methods [29, 31] focus on leveraging the internal patch distribution within a single image, however, their extension to capture the distribution of a small collection of images is non-trivial. In low-shot learning [41], several works [41, 12] proposed to mitigate the limited-data-induced overfitting issue by adapting data augmentations to the generative networks. Others [18, 5] stabilized the training process and reduced overfitting by revising the network design. Despite the promising performance in many low data regimes (typically having \\(100+\\) images), these low-shot methods fail in the extremely few-shot setting (e.g., 10 images). Our work lies in the scope of few-shot learning." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.841, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Few-shot image synthesis. Conventional few-shot learning aims at learning a discriminative classifier under limited data scenarios. In the context of image synthesis with GANs, the goal instead is to produce diverse new im" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.89, + 0.605 + ], + "angle": 0, + "content": "ages from the learned distribution while preventing overfitting to the few training samples. A straightforward way is to treat it as a domain adaptation problem and incorporate the commonly used transfer learning technique, i.e., fine-tuning, to ease the need for data. However, naive fine-tuning (TGAN) [36] often suffers from overfitting and results in poor performance. Researchers proposed remedies such as mining suitable parts of the latent space before fine-tuning [35] or restricting weight updates, for example, updating only the BatchNorm parameters of the generator [22], penalizing drastic changes in important weights [17], or freezing the earliest layers of the discriminator (FreezeD) [20]. More recent works focused on introducing different regularizations to preserve specific knowledge from the pretrained model and prevent diversity degradation [42]. For example, CDC [23] proposed to preserve the pair-wise perceptual similarity between samples from the source domain and to transfer it to the target domain, while RSSA [37] designed a novel consistency term to align the structural information between source and target domains. Although the two aforementioned methods constitute the current state of the art in few-shot generative learning, their assumptions impose strong constraints on the structure of the few-shot target domain. As we show in experiments, they fail in the more challenging regime when the source and target domains are not restrictively similar. Most recently, [39] proposed to replace prior knowledge preservation criteria with adaptation-aware kernel modulation (AdAM), which relaxed the source-target proximity requirement of previous methods to some extent. In this work, we take a step further and introduce a new regularization term to preserve the generator's smoothness properties that are not limited to a specific domain, enabling successful adaptation between image domains of unprecedented structural dissimilarity." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.614, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Smoothness of image generators. Smooth transitions in the latent space are an important property for generative models, where it is believed to be a sign of a well-conditioned generator. Models trained on large datasets naturally possess this property with or without explicit regularization [2, 15]. For example, StyleGANv2 [15] introduced a regularization based on the perceptual path length measure (PPL) [14], which encourages that a fixed-size step in the latent space results in a fixed-magnitude change in the image space. However, achieving a smooth mapping of the generator is difficult for few-shot image synthesis since there are not enough training samples. Thus, MixDL [16] sought to alleviate the \"staircase\" latent space interpolations, i.e., jumps between training samples, by introducing a continuous coefficient vector and enforcing smooth interpolations between training images. Although the two above regularizers aim to encourage smoother interpolations between training samples and thus mitigate mode collapse, they are not designed to take advantage of the available pre-training" + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.956 + ], + "angle": 0, + "content": "7074" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.082, + 0.881, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.227, + 0.895, + 0.27 + ], + "angle": 0, + "content": "Figure 2. Given a pre-trained generator \\( G_{s} \\), the proposed smoothness similarity regularization preserves the learned smoothness of \\( G_{s} \\) while adapting it to a target domain with very few images. To mitigate overfitting to the target domain, the discriminator loss utilizes features at various layers and automatically adjusts the impact of different semantic scales to the similarity of the source and target domain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.471, + 0.373 + ], + "angle": 0, + "content": "knowledge. In contrast, in this work we develop a new smoothness similarity regularization that leverages the well-structured latent space of a pre-trained GAN generator. In effect, our approach enables high-quality few-shot image synthesis by transferring smooth and realistic image transitions of pre-trained GANs to diverse few-shot domains." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.388, + 0.169, + 0.403 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.47, + 0.549 + ], + "angle": 0, + "content": "In the task of few-shot GAN adaptation, we are given a small target dataset \\( T \\) and a pre-trained GAN model, consisting of a discriminator \\( D \\) and a generator \\( G_{s} \\), which produces an image \\( x = G_{s}(z) \\) from a continuous input variable \\( z \\), such as a random noise vector or a continuous class embedding. The goal is to adapt the generator to the target dataset such that it generates diverse and realistic images in the domain of the target dataset as shown in Fig. 1. We denote the adapted target generator by \\( G_{t} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.47, + 0.671 + ], + "angle": 0, + "content": "To achieve few-shot synthesis with a high image quality and diversity, our model should adhere to the following two properties. Firstly, the generator \\( G_{t} \\) should not only memorize and generate the target images, which will be addressed by the smoothness similarity regularization (Sec. 3.1). Secondly, the discriminator \\( D \\) must avoid overfitting to the few target images in order to provide useful supervision for \\( G_{t} \\) (Sec. 3.2). The overview of our method is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.445, + 0.697 + ], + "angle": 0, + "content": "3.1. Smoothness similarity regularization for \\( G_{t} \\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In a low data regime like ours, \\( G_{t} \\) can easily overfit to the target dataset \\( T \\) and collapse to reproducing only the few modes represented in the training data. When walking in the latent space of such a generator, one would observe \"staircase\" patterns, where minor shifts in the latent space cause discontinuous transitions in the output image space (as shown in row 4 of Fig. 5). Naturally, to achieve a synthesis of high diversity, it is desirable for \\( G_{t} \\) to avoid such discontinuities, as having smoother image transitions allows to generate intermediate samples that can exhibit novel features. Therefore, in our model we aim to encourage \\( G_{t} \\) to produce smooth latent space interpolations, in which all the intermediate images are realistic." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.283, + 0.892, + 0.509 + ], + "angle": 0, + "content": "Our approach is based on the observation that GANs trained on large datasets tend to have a well-structured latent space [34, 9, 30], in which different latent space directions can lead to smooth and interpretable image transitions. For example, in a generator pre-trained on a large dataset of churches, latent directions can emerge causing smooth zooming or translation of churches (see Fig. 2). Our observation is that the nature of such image transitions (e.g., zooming or translation) is remarkably general. Thus, we propose a regularizer that utilizes this smoothness property of the source generator \\(G_{s}\\) as a cue while adapting it to another image domain, which can be very different from the domain that was used for pre-training. For example, as shown in Fig. 2, the same latent directions of churches can cause similar zooming or translation effects on shells." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.893, + 0.676 + ], + "angle": 0, + "content": "Mathematically, the smoothness of the generator can be represented via a Jacobian matrix \\( J_{G^l}(z) = ||\\partial G^l (z) / \\partial z|| \\), quantifying how the generator's intermediate features after the \\( l \\)-th block change under local shifts in the latent space. As we want the same latent shift to cause perceptually similar image transitions in the source and target domains, we design a regularization term that brings the Jacobian matrices of \\( G_s^l \\) and \\( G_t^l \\) closer together. As the computation of full Jacobian matrices is expensive, we use an unbiased estimator of their products with a Gaussian vector [6, 15], which can be computed with standard back-propagation:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.684, + 0.892, + 0.704 + ], + "angle": 0, + "content": "\\[\nJ _ {G ^ {l}} ^ {T} (z) \\cdot y = \\mathbb {E} _ {(y) \\sim N (0, 1)} \\nabla_ {z} \\langle G ^ {l} (z), y \\rangle , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.892, + 0.743 + ], + "angle": 0, + "content": "where \\(y\\) is a Gaussian tensor of the same shape as \\(G^l\\). Our smoothness similarity regularization is then expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.503, + 0.75, + 0.892, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S S} = \\lambda_ {S S} \\cdot \\mathbb {E} _ {(z, y) \\sim N (0, 1)} \\| \\nabla_ {z} \\langle G _ {s} ^ {l} (z), y \\rangle - \\nabla_ {z} \\langle G _ {t} ^ {l} (z), y \\rangle \\| _ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.893, + 0.901 + ], + "angle": 0, + "content": "where \\(\\lambda_{SS}\\) steers the impact of the regularizer. As shown in Fig. 2, the smoothness similarity regularization depends on both generators, but only \\(G_{t}\\) is updated. It is interesting to note that the Jacobian matrix is also used for the path length regularization [15], which forces \\(J_{G}(z)\\) to be orthogonal up to a global scale at any \\(z\\). While this alternative regularizer also induces some form of smoothness, it does not transfer the inherently learned smoothness" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7075" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.259 + ], + "angle": 0, + "content": "of a pre-trained GAN. We show in Sec. 4.1 that it struggles to enforce the realism of intermediate images. Furthermore, our approach shares the motivation with some prior regularization approaches that use noise perturbations to enforce diversity [23, 37]. In contrast to Eq. 2, these approaches incorporate non-gradient components, e.g., assuming similarity of images \\(G_{s}(z)\\leftrightarrow G_{t}(z)\\) or distributions \\(d(G_{t}(z_{1}),G_{t}(z_{2}))\\leftrightarrow d(G_{s}(z_{1}),G_{s}(z_{2}))\\) . As such assumptions are violated when source and target domains are dissimilar, they perform poorly compared to our smoothness similarity regularization \\(\\mathcal{L}_{SS}\\) as shown in the experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.266, + 0.363, + 0.281 + ], + "angle": 0, + "content": "3.2. Revisiting the \\(D\\) adversarial loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.47, + 0.41 + ], + "angle": 0, + "content": "To identify what kind of image transitions look realistic for the target domain, \\( G_{t} \\) requires strong supervision from the discriminator on image realism at different semantic scales. This includes the colors and textures of objects, as well as object shapes, especially if their distribution is different from the shapes of objects in the source domain. Learning the concept of image realism in low data regimes is, however, challenging due to the problem of overfitting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.471, + 0.65 + ], + "angle": 0, + "content": "Typically, a GAN discriminator consists of several consecutive blocks \\(\\{D^i\\}_{i = 1}^N\\) and computes for each given image \\(x\\) a real/fake logit after the last block \\(l = s^{N}\\circ D^{N}(x)\\) where \\(s^N\\) is a final processing layer such as a convolution. When adapting such a discriminator to a very small dataset, it is prone to memorizing the training set [32], leading to mode collapse and poor diversity of synthesized images [23]. A possible solution [23, 37] to overcome memorization is to use variants of the PatchGAN discriminator [11], discarding the latest discriminator layers: \\(l = s^{k}\\circ D^{k}(x),k < N\\). This solution allows to adapt colors and textures of generated images to the target domain while avoiding the memorization problem. However, it naturally has a limited capacity to learn more high-level semantic scene properties such as the shapes of objects, which we show in the experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.47, + 0.817 + ], + "angle": 0, + "content": "In order to avoid memorization, and yet to balance the adaptation of colors, textures, and shapes of generated objects to a new domain, we hypothesize that a more flexible attention to different levels of image realism is required by the discriminator. To this end, we perform a simple yet efficient modification to the loss function of the discriminator. Given a discriminator \\(\\{D^i\\}_{i = 1}^N\\) and its adversarial loss function \\(\\mathcal{L}_D(l)\\) used for pre-training (e.g., cross-entropy or hinge loss), we design the discriminator to produce real/fake logits after each discriminator's block, and correspondingly compute the loss as the average across all blocks:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.823, + 0.47, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a l l} (x) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\mathcal {D}} \\left[ l ^ {i} (x) \\right], \\quad l ^ {i} (x) = s ^ {i} \\circ D ^ {i} (x). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "With the new objective, \\( D \\) is given more freedom to utilize the features extracted at different scales to compute the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "loss. Our finding is that \\( D \\) dynamically adapts the magnitude of the loss at each scale to the target domain, without explicit supervision (see Fig. 6). Consequently, we observe a strong overall stabilization effect on the adaptation performance across diverse source-target dataset pairs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.185, + 0.634, + 0.202 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.211, + 0.892, + 0.378 + ], + "angle": 0, + "content": "To demonstrate that our approach for few-shot GAN adaptation can be applied to unconditional and class-conditional GANs, we selected for each category a popular GAN architecture: unconditional StyleGANv2 [15] and class-conditional BigGAN [2]. For both models, we test our approach on a variety of source-target domain pairs. We focus on 10-shot target adaptation in the main paper, but we provide results for 1-shot and 5-shot adaptation in the supplementary material. For fair comparisons with prior works, most of our ablations and comparisons are conducted with StyleGANv2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.39, + 0.798, + 0.405 + ], + "angle": 0, + "content": "4.1. Adaptation of unconditional GAN" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.892, + 0.581 + ], + "angle": 0, + "content": "Datasets. In contrast to previous works that mostly considered pairs of similar datasets like Face \\(\\rightarrow\\) Sketch and Face \\(\\rightarrow\\) Sunglasses, we extend the protocol by including structurally dissimilar pairs of source and target domains, which is a more challenging task and is our primary interest. As source generators, we use StyleGANv2 checkpoints pretrained on FFHQ [14], LSUN-Church, and LSUN-Horse [38]. For the target datasets, we selected 10-shot subsets of various commonly used few-shot datasets, such as Anime-Face, Shells, or Pokemons [41, 18]. Results on more datasets are shown in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Training details. We fine-tune StyleGANv2 using the \\(\\mathcal{L}_{SS}\\) and \\(\\mathcal{L}_{all}\\) loss terms as presented in Sec. 3. For the smoothness similarity regularization, we use the intermediate features \\(G^{l}\\) at resolution \\((32\\times 32)\\) and set \\(\\lambda_{SS} = 5.0\\). We follow [23] in choosing all the other hyperparameters, such as image resolution \\((256\\times 256)\\), learning rates, and batch size. Our experiments across all datasets use the same model configuration and set of hyperparameters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Baselines. We compare our method to most recent few-shot GAN adaptation approaches: TGAN [36], FreezeD [20], CDC [23], RSSA [37], and AdAM [39]. In addition, we compare our proposed smoothness similarity regularizer \\(\\mathcal{L}_{SS}\\) to other regularization techniques: path length regularization (PPL) [15] and MixDL [16]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Evaluation. In low data regimes, it is necessary to judge results both in quality and diversity aspects, as there is a trade-off between them [25, 32]. We measure the quality with FID [10] between a held-out validation set and a generated set of the same size. Following [23], we evaluate diversity with the intra-LPIPS, clustering generated images according to their nearest training samples and computing" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7076" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.119, + 0.225, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.349, + 0.195, + 0.373 + ], + "angle": 0, + "content": "10-shot Anime-Faces" + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.086, + 0.859, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.439, + 0.225, + 0.664 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.669, + 0.179, + 0.693 + ], + "angle": 0, + "content": "10-shot Shells" + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.404, + 0.859, + 0.72 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.862, + 0.087, + 0.882, + 0.715 + ], + "angle": 0, + "content": "TAN Freezed CDC RSSA AdAM Ours Source TGAN Freezed CDC RSSA AdAM Ours Source" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.723, + 0.894, + 0.78 + ], + "angle": 0, + "content": "Figure 3. Visual comparison to prior methods on Face \\(\\rightarrow\\) Anime and Church \\(\\rightarrow\\) Shells, the source-target dataset pairs with a dissimilar structure (e.g., shapes of objects). In this challenging regime, we observe that prior methods suffer from training instabilities, memorization issues, or inability to adapt the shapes of objects to the new domain. In contrast, our method generates images that look realistic, flexibly combine features of different target images, and transfer the variation of images from the source domain to the target domain." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.789, + 0.471, + 0.85 + ], + "angle": 0, + "content": "the average LPIPS [40] of all the clusters. We train all models for 30k epochs in case of dissimilar domain pairs and for 5k on closer domains, evaluating metrics every 1k epochs. Final checkpoints in all experiments correspond to best FID." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Results with dissimilar source-target domains. We first present our results on the source-target domain pairs with dissimilar structure: Face \\(\\rightarrow\\) Anime, Church \\(\\rightarrow\\) Shells," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.892, + 0.895 + ], + "angle": 0, + "content": "and Horse \\(\\rightarrow\\) Pokemon (see Fig. 3 and supplementary material). Our general observation from Fig. 3 is that in this challenging regime prior methods suffer either from training instabilities, memorization issues, or inability to adapt the shape of objects to the new domain. For example, for Face \\(\\rightarrow\\) Anime, despite an apparent correspondence between the two domains, none of the prior methods success-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7077" + } + ], + [ + { + "type": "aside_text", + "bbox": [ + 0.07, + 0.094, + 0.088, + 0.335 + ], + "angle": 270, + "content": "Ours AdAM RSSA CDC Source Real" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.091, + 0.484, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.348, + 0.892, + 0.378 + ], + "angle": 0, + "content": "Figure 4. Visual comparison to most recent prior methods on Face \\(\\rightarrow\\) Sketch and Church \\(\\rightarrow\\) Sunglasses, the dataset pairs depicting similar image domains. In this regime, our method performs on par with previous state of the art. See Table 2 for a quantitative comparison." + }, + { + "type": "table", + "bbox": [ + 0.076, + 0.387, + 0.478, + 0.505 + ], + "angle": 0, + "content": "
MethodFace→AnimeChurch→ShellsHorse→Pokemons
FID↓LPIPS↑FID↓LPIPS↑FID↓LPIPS↑
TGAN [36]153.20.29205.30.22115.00.52
FreezeD [20]112.40.22180.80.27123.30.49
CDC [23]140.20.50187.90.48109.50.55
RSSA [37]133.20.37182.40.44117.30.54
AdAM [39]116.40.42152.40.28106.50.55
Ours97.30.57140.50.5384.10.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.511, + 0.47, + 0.54 + ], + "angle": 0, + "content": "Table 1. Comparison of the adaptation performance in case of dissimilar source-target domains. Bold denotes best performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.794 + ], + "angle": 0, + "content": "fully transfers the distribution of head poses to the anime style, e.g., overfitting too strongly to the 10 provided samples (FreezeD), failing to adapt the shape of faces to the style of anime (CDC), or not generating high-quality animenefaces due to instabilities (TGAN, RSSA, AdAM). Similarly, for \\( \\text{Church} \\rightarrow \\text{Shells} \\), we observe that prior methods produce only copies of the example shells (FreezeD, AdAM), generate shells of unrealistic church-like shapes (CDC, RSSA), or suffer from instabilities (TGAN). In contrast, our method achieves high-quality synthesis, in which the generated images (i) look like realistic anime-faces and shells; (ii) flexibly combine features observed in different target images (e.g., anime hair color can be combined with various eye colors or background styles); and (iii) meaningfully transfer the variation of images from the source domain (e.g., generated shells adjust to the positions and shapes of churches)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The quantitative comparison in Table 1 confirms our analysis, where our method achieves the best quality and diversity scores across all datasets. We note a high average relative improvement of more than \\(18\\%\\) and \\(11\\%\\) in FID and LPIPS compared to the highest scores achieved by prior methods. Overall, we conclude that our method significantly improves over prior works on few-shot GAN" + }, + { + "type": "table", + "bbox": [ + 0.546, + 0.387, + 0.842, + 0.505 + ], + "angle": 0, + "content": "
MethodFace→SketchFace→Sunglasses
FID↓LPIPS↑FID↓LPIPS↑
TGAN [36]54.20.3836.80.56
FreezeD [20]48.80.3232.00.59
CDC [23]54.20.4030.50.59
RSSA [37]61.40.4536.30.58
AdAM [39]56.30.3731.10.60
Ours45.20.4427.50.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.511, + 0.89, + 0.54 + ], + "angle": 0, + "content": "Table 2. Comparison in case of structurally close source-target domains. Bold denotes best performance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.554, + 0.854, + 0.569 + ], + "angle": 0, + "content": "adaptation with dissimilar source and target domains." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.892, + 0.827 + ], + "angle": 0, + "content": "Results with close source-target domains. Next, we follow the evaluation of prior works and compare the models on similar source and target domains, such as adaptation of human faces to a different style. The visual results for Face \\(\\rightarrow\\) Sketch and Face \\(\\rightarrow\\) Sunglasses are shown in Fig. 4. Our method successfully performs the few-shot adaptation in this setting, adapting the colors and textures of faces to the gray-scale sketch domain, or adding a novel attribute (sunglasses). We note that our method is not explicitly designed to transfer all the details of a face from the source domain, thus changes in the generated images like facial hair are expected. Yet, we observe that our method generally does not lose distinctive features of faces in source images, performing on par with previous state-of-the-art methods. The quantitative comparison is provided in Table \\(2^{1}\\): on both datasets our method achieves the best FID scores and performs on par with the best performer in LPIPS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.828, + 0.892, + 0.874 + ], + "angle": 0, + "content": "Ablations. We demonstrate the importance of our proposed loss terms in Fig. 5, which shows latent space interpolations of trained models and their similarity to the pre" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.891, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)FID evaluation differs from prior works (discussed in suppl. material)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7078" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.466, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.415, + 0.47, + 0.457 + ], + "angle": 0, + "content": "Figure 5. Latent space interpolations of the source generator and the ablation models from Tables 3-4. Leftmost and rightmost columns show the used \\( D \\) loss and \\( G \\) smoothness regularization." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.462, + 0.449, + 0.581 + ], + "angle": 0, + "content": "
D lossSmooth reg. for GFace→AnimeChurch→Shells
FID↓LPIPS↑FID↓LPIPS↑
StyleGANv2-178.00.21243.80.17
StyleGANv2SS (ours)180.70.61252.80.62
PatchGAN [23]-145.20.37183.10.31
PatchGAN [23]SS (ours)132.20.55184.20.56
Lall (ours)-116.40.36175.40.43
Lall (ours)SS (ours)97.30.57140.50.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.581, + 0.468, + 0.595 + ], + "angle": 0, + "content": "Table 3. Impact of \\( {\\mathcal{L}}_{\\text{all }} \\) and \\( {\\mathcal{L}}_{SS} \\) . Bold denotes best performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.839 + ], + "angle": 0, + "content": "trained source model \\( G_{s} \\) (row 1). Firstly, we note that the plain StyleGANv2 model (row 2) suffers from instabilities in our low data regime, achieving poor image quality and diversity and having \"staircase\"-like latent space interpolations. Applying \\( \\mathcal{L}_{SS} \\) without \\( \\mathcal{L}_{all} \\) (row 3) helps to achieve diverse synthesis with smooth interpolations, but it is not enough to achieve good image quality. On the other hand, using \\( \\mathcal{L}_{all} \\) (row 4) helps to overcome instabilities and improve image quality, but it cannot maintain smooth interpolations and high diversity. Finally, our full model (row 5) allows a higher-quality, diverse synthesis with smooth and realistic latent space interpolations. Note how the image transitions mimic the behaviour of the source model (churches and shells change shapes and positions similarly), allowing to achieve diverse and realistic synthesis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The effect of \\(\\mathcal{L}_{all}\\) is further demonstrated in Fig. 6, where we show the contribution of different \\(D\\) blocks to the adversarial loss at different epochs. We note the ability of the discriminator to identify correct loss contributions" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.088, + 0.891, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.195, + 0.893, + 0.265 + ], + "angle": 0, + "content": "Figure 6. The contribution of features at different \\( D \\) blocks to the adversarial loss function \\( \\mathcal{L}_{all} \\). For two closer image domains (the left plot), the network concentrates mostly on earlier layers to compute the loss, while for less similar domains the network learns to use the later layers representing more high-level \\( D \\) features." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.273, + 0.866, + 0.363 + ], + "angle": 0, + "content": "
D lossSmooth reg. for GFace→AnimeChurch→Shells
FID↓LPIPS↑FID↓LPIPS↑
Lall(ours)-116.40.36175.40.43
Lall(ours)PPL [14]107.80.46179.40.44
Lall(ours)MixDL [16]105.90.50150.40.51
Lall(ours)SS (ours)97.30.57140.50.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.364, + 0.891, + 0.391 + ], + "angle": 0, + "content": "Table 4. Comparison of smoothness similarity regularization \\(\\mathcal{L}_{SS}\\) with other regularizers. Bold denotes best performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.893, + 0.579 + ], + "angle": 0, + "content": "adaptively for different source-target domain pairs. For example for Face \\(\\rightarrow\\) Anime, the network concentrates mostly on the earliest \\(D\\) blocks to adapt the colors and textures of faces to a new style. In contrast, for the more distant domains Church \\(\\rightarrow\\) Shells, the network learns to attribute a higher weight to the later blocks to also adapt higher-level features, such as shapes of objects. In effect, we observe a stabilized adaptation of colors, textures, and shapes of objects across diverse source-target pairs. Using PatchGAN [23] as discriminator loss does not achieve such a balance as it focuses mostly on lower-scale features (row 6 in Fig. 5)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.687 + ], + "angle": 0, + "content": "Our observations are confirmed by the quantitative study in Table 3: without \\(\\mathcal{L}_{SS}\\) the model does not achieve high diversity (high LPIPS), while \\(\\mathcal{L}_{all}\\) is necessary for high image quality (low FID). We conclude that both our proposed loss terms are important to achieve high-quality synthesis. More ablations on \\(\\mathcal{L}_{SS}\\) and \\(\\mathcal{L}_{all}\\) can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Lastly, Table 4 provides a comparison of our proposed \\(\\mathcal{L}_{SS}\\) loss term to other regularizers: path length regularization (PPL) [14] and MixDL [16]. While all regularizers help to achieve smoother latent space interpolations and thus improve the quality and diversity metrics, our smoothness similarity regularization enables the highest performance in both FID and LPIPS. While our approach transfers the learned smoothness of the source generator to the target domain, PPL and MixDL resort to gradually interpolating between the provided training samples, which leads to latent space interpolations that either look unrealistic or lack diversity (rows 7-8 in Fig. 5). This demonstrates that transferring smoothness from a pre-trained generator is beneficial to enforce image transitions that are realistic and diverse." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7079" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.178, + 0.09, + 0.393, + 0.104 + ], + "angle": 0, + "content": "Real samples - 10-Shot Flowers" + }, + { + "type": "header", + "bbox": [ + 0.578, + 0.09, + 0.807, + 0.104 + ], + "angle": 0, + "content": "Real samples - 10-Shot Pokemons" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.105, + 0.482, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.105, + 0.89, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.141, + 0.825, + 0.156 + ], + "angle": 0, + "content": "10-Shot adaptation results of our method on class-conditional BigGAN [2], pre-trained on ImageNet" + }, + { + "type": "image", + "bbox": [ + 0.075, + 0.156, + 0.899, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.388, + 0.894, + 0.43 + ], + "angle": 0, + "content": "Figure 7. 10-shot adaptation results for the class-conditional BigGAN [2] pre-trained on ImageNet. While simple fine-tuning (FT) suffers from training instabilities and mode collapse, our method helps to achieve much higher image quality and diversity, transferring smooth and realistic image transitions from the source domain, e.g., objects smoothly changing their locations, size, and shape." + }, + { + "type": "table", + "bbox": [ + 0.077, + 0.436, + 0.476, + 0.525 + ], + "angle": 0, + "content": "
D lossSmooth reg. for GImageNet→FlowersImageNet→Pokemons
FID↓LPIPS↑FID↓LPIPS↑
BigGAN-213.30.29226.80.15
BigGANSS (ours)225.60.47208.30.47
Lall (ours)-123.90.28129.40.27
Lall (ours)SS (ours)106.40.5589.60.56
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.531, + 0.47, + 0.56 + ], + "angle": 0, + "content": "Table 5. Ablation on the performance when adapting the class-conditional BigGAN model [2] pre-trained on ImageNet." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.574, + 0.398, + 0.59 + ], + "angle": 0, + "content": "4.2. Adaptation of class-conditional GAN" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.884 + ], + "angle": 0, + "content": "Our approach is not limited to unconditional GANs, but it can also be applied to a class-conditional GAN model. We selected BigGAN [2] for our experiments as it is a popular backbone architecture for class-conditional image synthesis on ImageNet [7]. We make two modifications to enable the adaptation of the model to unconditional few-shot datasets. Firstly, we remove the conditioning of the discriminator via the projection layer [19]. Secondly, we treat the generator's learned continuous class embedding as part of the latent space, thus sampling a Gaussian vector in the joint noise-class space at each fine-tuning epoch. This way, the generator produces an image based on a single input vector in an unconditional fashion. We then fine-tune the pre-trained model using our loss terms \\(\\mathcal{L}_{SS}\\) and \\(\\mathcal{L}_{all}\\) as presented in Sec. 3. We use image resolution \\(256\\times 256\\) and batch size of 32. The hyperparameters for \\(\\mathcal{L}_{SS}\\) are the same as for StyleGANv2: intermediate features \\(G^{l}\\) at resolution \\((32\\times 32)\\) and \\(\\lambda_{SS} = 5.0\\). We train for 30k epochs and select checkpoints by best FID." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Datasets. As the source generator, we use the Big-" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.438, + 0.893, + 0.514 + ], + "angle": 0, + "content": "GAN checkpoint pre-trained on class-conditional ImageNet at resolution \\(256 \\times 256\\). We demonstrate 10-shot adaptation results with two commonly used few-shot generation datasets: Oxford-Flowers [21] and Pokémon [18]. We use the same model configuration for both datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.893, + 0.746 + ], + "angle": 0, + "content": "Results. Fig. 7 demonstrates latent space interpolations of the source and target generators. We note that a simple fine-tuning of BigGAN suffers from training instabilities and mode collapse. In contrast, our method successfully adapts BigGAN to generate diverse images in the target domains. We highlight that our method transfers smooth and realistic image transitions from the well-learned BigGAN's noise-class space, despite significant dissimilarities between ImageNet and the few-shot datasets, in particular Pokemons. For example, it can be noticed how the latent space interpolations in the target domains mimic the source domain, e.g., the generated flowers andPokemons change their position and size similarly to dogs and wolves \\((5^{\\mathrm{th}}-\\) \\(10^{\\mathrm{th}}\\) columns in Fig. 7) or stretch their shape to mimic the proportions of busses \\((11^{\\mathrm{th}} - 14^{\\mathrm{th}}\\) columns)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Table 5 shows the importance of our proposed loss terms. Our observations are consistent with the ablations with StylGANv2: \\(\\mathcal{L}_{all}\\) is necessary to avoid instabilities and achieve a good image quality (low FID), while \\(\\mathcal{L}_{SS}\\) is required to achieve smooth latent space interpolations and good diversity (high LPIPS). We conclude that our method successfully extends to the adaptation of class-conditional models, where target domains benefit from the rich noise-class space learned on a multi-class dataset such as ImageNet. More details and results are provided in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7080" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.196, + 0.106 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.473, + 0.253 + ], + "angle": 0, + "content": "In this work, we presented a new method for few-shot adaptation of GAN models. It transfers the smooth latent space of a pre-trained GAN, which was trained on a large dataset, to a new domain with very few images. We addressed the case of few-shot GAN adaptation when the source and target domains are structurally dissimilar, which is a common issue in applications. Our extensive results demonstrate that in this setting our approach outperforms previous works in terms of image quality and diversity." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.264, + 0.238, + 0.28 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.47, + 0.38 + ], + "angle": 0, + "content": "The work has been supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC 2070 -390732324 and the ERC Consolidator Grant FORHUE (101044724). We thank Jinhui Yi for providing and analyzing the sugar beet data used in supplementary experiments." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.39, + 0.175, + 0.406 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.47, + 0.484 + ], + "angle": 0, + "content": "[1] Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-shot segmentation without meta-learning: A good transductive inference is all you need? In Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.485, + 0.472, + 0.539 + ], + "angle": 0, + "content": "[2] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In International Conference on Learning Representations (ICLR), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.541, + 0.47, + 0.623 + ], + "angle": 0, + "content": "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.624, + 0.47, + 0.678 + ], + "angle": 0, + "content": "[4] Yunjoy Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.679, + 0.47, + 0.734 + ], + "angle": 0, + "content": "[5] Kaiwen Cui, Jiaxing Huang, Zhipeng Luo, Gongjie Zhang, Fangneng Zhan, and Shijian Lu. Genco: Generative constraining for generative adversarial networks with limited data. In Conference on Artificial Intelligence (AAAI), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.735, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[6] Yann Dauphin, Harm De Vries, and Yoshua Bengio. Equilibrated adaptive learning rates for non-convex optimization. In Advances in Neural Information Processing Systems (NeurIPS), 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.79, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[7] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In Conference on Computer Vision and Pattern Recognition (CVPR), 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[8] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.416, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[9] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In Advances in Neural Information Processing Systems (NeurIPS), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems (NeurIPS), 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[11] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[12] Tero Karras, Miika Aittala, Janne Hellsten, S. Laine, J. Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Advances in Neural Information Processing Systems (NeurIPS), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[13] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Advances in Neural Information Processing Systems (NeurIPS), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[14] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[15] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[16] Chaerin Kong, Jeesoo Kim, Donghoon Han, and Nojun Kwak. Few-shot image generation with mixup-based distance learning. In European Conference on Computer Vision (ECCV), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[17] Yijun Li, Richard Zhang, Jingwan Cynthia Lu, and Eli Shechtman. Few-shot image generation with elastic weight consolidation. In Advances in Neural Information Processing Systems (NeurIPS), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[18] Bingchen Liu, Yizhe Zhu, Kunpeng Song, and Ahmed Elgammal. Towards faster and stabilized gan training for high-fidelity few-shot image synthesis. In International Conference on Learning Representations (ICLR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[19] Takeru Miyato and Masanori Koyama. cGANs with projection discriminator. In International Conference on Learning Representations (ICLR), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.706, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[20] Sangwoo Mo, Minsu Cho, and Jinwoo Shin. Freeze discriminator: A simple baseline for fine-tuning gans. In CVPR AI for Content Creation Workshop, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[21] Maria-Elena Nilsback and Andrew Zisserman. A visual vocabulary for flower classification. In Conference on Computer Vision and Pattern Recognition (CVPR), 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[22] Atsuhiro Noguchi and T. Harada. Image generation from small datasets via batch statistics adaptation. In International Conference on Computer Vision (ICCV), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[23] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7081" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[24] Jie Ren, Peter J Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark Depristo, Joshua Dillon, and Balaji Lakshminarayanan. Likelihood ratios for out-of-distribution detection. Advances in Neural Information Processing Systems (NeurIPS), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[25] Esther Robb, Wen-Sheng Chu, Abhishek Kumar, and Jia-Bin Huang. Few-shot adaptation of generative adversarial networks. arXiv:2010.11943, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.209, + 0.471, + 0.263 + ], + "angle": 0, + "content": "[26] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis. arXiv preprint arXiv:2301.09515, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.266, + 0.47, + 0.307 + ], + "angle": 0, + "content": "[27] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleganxI: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.31, + 0.47, + 0.365 + ], + "angle": 0, + "content": "[28] Edgar Schonfeld, Vadim Sushko, Dan Zhang, Juergen Gall, Bernt Schiele, and Anna Khoreva. You only need adversarial supervision for semantic image synthesis. In International Conference on Learning Representations (ICLR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.368, + 0.47, + 0.41 + ], + "angle": 0, + "content": "[29] Tamar Rott Shaham, Tali Dekel, and T. Michaeli. SinGAN: Learning a generative model from a single natural image. In International Conference on Computer Vision (ICCV), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.412, + 0.47, + 0.453 + ], + "angle": 0, + "content": "[30] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.47, + 0.512 + ], + "angle": 0, + "content": "[31] Vadim Sushko, Juergen Gall, and Anna Khoreva. One-Shot GAN: Learning to generate samples from single images and videos. In Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.514, + 0.47, + 0.555 + ], + "angle": 0, + "content": "[32] Vadim Sushko, Dan Zhang, Juergen Gall, and Anna Khoreva. Learning to generate novel scene compositions from single images and videos. arXiv:2103.13389, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.557, + 0.47, + 0.613 + ], + "angle": 0, + "content": "[33] Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. Transactions on Pattern Analysis and Machine Intelligence, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.615, + 0.47, + 0.657 + ], + "angle": 0, + "content": "[34] Andrey Voynov and Artem Babenko. Unsupervised discovery of interpretable directions in the gan latent space. In International conference on machine learning (ICML), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.47, + 0.728 + ], + "angle": 0, + "content": "[35] Yaxing Wang, Abel Gonzalez-Garcia, David Berga, L. Herranz, F. Khan, and Joost van de Weijer. Minegan: Effective knowledge transfer from gans to target domains with few images. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[36] Yaxing Wang, Chenshen Wu, L. Herranz, Joost van de Weijer, Abel Gonzalez-Garcia, and B. Raducanu. Transferring gans: generating images from limited data. In European Conference on Computer Vision (ECCV), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[37] Jiayu Xiao, Liang Li, Chaofei Wang, Zheng-Jun Zha, and Qingming Huang. Few shot generative model adaption via relaxed spatial structural alignment. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[38] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv:1506.03365, 2015." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.148 + ], + "angle": 0, + "content": "[39] Zhao Yunqing, Keshigeyan Chandrasegaran, Milad Abdol-lahzadeh, and Ngai-man Cheung. Few-shot image generation via adaptation-aware kernel modulation. In Advances in Neural Information Processing Systems (NeurIPS), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.205 + ], + "angle": 0, + "content": "[40] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Conference on Computer Vision and Pattern Recognition (CVPR), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.893, + 0.261 + ], + "angle": 0, + "content": "[41] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. In Conference on Neural Information Processing Systems (NeurIPS), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.893, + 0.318 + ], + "angle": 0, + "content": "[42] Yunqing Zhao, Henghui Ding, Houjing Huang, and Ngai-Man Cheung. A closer look at few-shot image generation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7082" + } + ] +] \ No newline at end of file diff --git a/2023/Smoothness Similarity Regularization for Few-Shot GAN Adaptation/e521c479-803a-4c4d-bf48-e946e8872ca9_origin.pdf b/2023/Smoothness Similarity Regularization for Few-Shot GAN Adaptation/e521c479-803a-4c4d-bf48-e946e8872ca9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1d07c786d47ed44cd30447eb2205bcff1061de9a --- /dev/null +++ b/2023/Smoothness Similarity Regularization for Few-Shot GAN Adaptation/e521c479-803a-4c4d-bf48-e946e8872ca9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8455761c17f88b720e670b6c58a66f5353d2f258f7726bc99399acccd50a5501 +size 7428312 diff --git a/2023/Smoothness Similarity Regularization for Few-Shot GAN Adaptation/full.md b/2023/Smoothness Similarity Regularization for Few-Shot GAN Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d8cbc818df6178f2039748d2a9848c47d7e70a98 --- /dev/null +++ b/2023/Smoothness Similarity Regularization for Few-Shot GAN Adaptation/full.md @@ -0,0 +1,248 @@ +# Smoothness Similarity Regularization for Few-Shot GAN Adaptation + +Vadim Sushko $^{1,2}$ Ruyu Wang $^{1}$ Juergen Gall $^{2,3}$ + +$^{1}$ Bosch Center for Artificial Intelligence $^{2}$ University of Bonn + +$^{3}$ Lamarr Institute for Machine Learning and Artificial Intelligence + +vad221@gmail.com ruyu.wang@de.bosch.com gall@iai.uni-bonn.de + +# Abstract + +The task of few-shot GAN adaptation aims to adapt a pre-trained GAN model to a small dataset with very few training images. While existing methods perform well when the dataset for pre-training is structurally similar to the target dataset, the approaches suffer from training instabilities or memorization issues when the objects in the two domains have a very different structure. To mitigate this limitation, we propose a new smoothness similarity regularization that transfers the inherently learned smoothness of the pre-trained GAN to the few-shot target domain even if the two domains are very different. We evaluate our approach by adapting an unconditional and a class-conditional GAN to diverse few-shot target domains. Our proposed method significantly outperforms prior few-shot GAN adaptation methods in the challenging case of structurally dissimilar source-target domains, while performing on par with the state of the art for similar source-target domains. + +# 1. Introduction + +Generative adversarial networks (GANs) have been shown to be powerful at various image synthesis tasks [4, 28, 3, 13, 27, 26]. The success of these models is in large part enabled by the availability of large datasets for training, typically consisting of thousands of images. However, there are many applications and computer vision tasks such as one-shot or few-shot learning [1, 33], out-of-distribution detection [24], or long-tailed recognition tasks [8] where the number of available training images is very low. + +Since training a GAN from scratch on very few samples does not perform well as shown in Fig. 1, a common strategy is to fine-tune a pre-trained GAN model on the few-shot dataset, typically employing additional regularization losses to penalize the degradation of the diversity [23, 37]. This approach, referred to as few-shot GAN adaptation, performs well when the target domain is structurally very similar to the dataset that has been used for pre-training, e.g., photographs vs. sketches of human faces. However, the performance drastically degrades in case of large dissimilarities between the source and target domain as shown in Fig. 1. + +![](images/21e06df4d0f06758a31fb4db90ba718d82a6587a51c4aa78b1cfae86feea72a4.jpg) +Figure 1. Training a GAN model $G$ on a few-shot dataset (row 1) from scratch fails due to training instabilities (row 2). We thus aim to adapt a GAN $G_{s}$ that has been pre-trained on a large dataset like LSUN-Church (row 3) to the target few-shot dataset $(G_{t})$ . While fine-tuning [23] does not perform well either if source and target are dissimilar (row 4), our approach generates diverse and realistic images (row 5) by transferring the smoothness properties of $G_{s}$ . + +Such dissimilarities are a major bottleneck of using GANs in other disciplines like medicine, production, or crop science, where there is a lack of large datasets due to privacy, confidentiality, or simply lack of data. Motivated by this fact, we extend the protocol for few-shot GAN adaptation by investigating also pairs of datasets that are very different like churches and shells as shown in Fig. 1. + +To improve few-shot GAN adaptation in the case of structurally dissimilar pairs, we propose a new GAN adaptation strategy. Firstly, we propose a new smoothness similarity regularization for the generator. Our key observation is that pre-trained GAN generators, regardless of the exact structure of objects in the pre-training dataset, learn well-structured and smooth latent spaces. For example, prior works demonstrated that various local shifts in the latent space can lead to interpretable and smooth transitions of + +output images, such as translation of objects in the scene or changing their size [34, 9, 30]. As we show in our experiments, the proposed smoothness similarity regularization enables the transfer of this desirable property to other few-shot image domains without compromising the synthesis quality. Secondly, to overcome overfitting issues, we revisit the adversarial loss function of the discriminator and propose a simple yet efficient modification by computing the loss at different layers of the discriminator. This leads to the mitigation of overfitting and a more stabilized adaptation of the model to diverse target domains. + +We evaluate our approach by adapting an unconditional [15] and a class-conditional GAN [2] to diverse few-shot target domains. Our model significantly outperforms previous state-of-the-art methods in image quality and diversity in the challenging case of dissimilar source and target domains, while performing on par with the state of the art on structurally similar dataset pairs. In summary, our contributions are as follows: (i) We extend the evaluation protocol for few-shot GAN adaptation by including new dataset pairs that are structurally much less similar than was considered in prior work. (ii) We propose a new smoothness similarity regularization, which enables diverse synthesis in the target domain by transferring the learned smoothness of a pretrained GAN. (iii) We revisit the adversarial loss function of the discriminator to stabilize few-shot GAN adaptation across diverse target domains. (iv) Our proposed model enables high-quality synthesis in the challenging case of dissimilar source and target domains, significantly outperforming prior methods. In addition, we show that our method can be applied to different classes of GAN architectures, including unconditional and class-conditional GAN models. + +# 2. Related Work + +To address the image generation problem in the low data regime, existing works mainly follow three research lines - one-shot, low-shot, and few-shot learning. One-shot generation methods [29, 31] focus on leveraging the internal patch distribution within a single image, however, their extension to capture the distribution of a small collection of images is non-trivial. In low-shot learning [41], several works [41, 12] proposed to mitigate the limited-data-induced overfitting issue by adapting data augmentations to the generative networks. Others [18, 5] stabilized the training process and reduced overfitting by revising the network design. Despite the promising performance in many low data regimes (typically having $100+$ images), these low-shot methods fail in the extremely few-shot setting (e.g., 10 images). Our work lies in the scope of few-shot learning. + +Few-shot image synthesis. Conventional few-shot learning aims at learning a discriminative classifier under limited data scenarios. In the context of image synthesis with GANs, the goal instead is to produce diverse new im + +ages from the learned distribution while preventing overfitting to the few training samples. A straightforward way is to treat it as a domain adaptation problem and incorporate the commonly used transfer learning technique, i.e., fine-tuning, to ease the need for data. However, naive fine-tuning (TGAN) [36] often suffers from overfitting and results in poor performance. Researchers proposed remedies such as mining suitable parts of the latent space before fine-tuning [35] or restricting weight updates, for example, updating only the BatchNorm parameters of the generator [22], penalizing drastic changes in important weights [17], or freezing the earliest layers of the discriminator (FreezeD) [20]. More recent works focused on introducing different regularizations to preserve specific knowledge from the pretrained model and prevent diversity degradation [42]. For example, CDC [23] proposed to preserve the pair-wise perceptual similarity between samples from the source domain and to transfer it to the target domain, while RSSA [37] designed a novel consistency term to align the structural information between source and target domains. Although the two aforementioned methods constitute the current state of the art in few-shot generative learning, their assumptions impose strong constraints on the structure of the few-shot target domain. As we show in experiments, they fail in the more challenging regime when the source and target domains are not restrictively similar. Most recently, [39] proposed to replace prior knowledge preservation criteria with adaptation-aware kernel modulation (AdAM), which relaxed the source-target proximity requirement of previous methods to some extent. In this work, we take a step further and introduce a new regularization term to preserve the generator's smoothness properties that are not limited to a specific domain, enabling successful adaptation between image domains of unprecedented structural dissimilarity. + +Smoothness of image generators. Smooth transitions in the latent space are an important property for generative models, where it is believed to be a sign of a well-conditioned generator. Models trained on large datasets naturally possess this property with or without explicit regularization [2, 15]. For example, StyleGANv2 [15] introduced a regularization based on the perceptual path length measure (PPL) [14], which encourages that a fixed-size step in the latent space results in a fixed-magnitude change in the image space. However, achieving a smooth mapping of the generator is difficult for few-shot image synthesis since there are not enough training samples. Thus, MixDL [16] sought to alleviate the "staircase" latent space interpolations, i.e., jumps between training samples, by introducing a continuous coefficient vector and enforcing smooth interpolations between training images. Although the two above regularizers aim to encourage smoother interpolations between training samples and thus mitigate mode collapse, they are not designed to take advantage of the available pre-training + +![](images/6f3dc01222538e80f247fe1115e4e0e41498a62ab8f02299a5901c836d68de64.jpg) +Figure 2. Given a pre-trained generator $G_{s}$ , the proposed smoothness similarity regularization preserves the learned smoothness of $G_{s}$ while adapting it to a target domain with very few images. To mitigate overfitting to the target domain, the discriminator loss utilizes features at various layers and automatically adjusts the impact of different semantic scales to the similarity of the source and target domain. + +knowledge. In contrast, in this work we develop a new smoothness similarity regularization that leverages the well-structured latent space of a pre-trained GAN generator. In effect, our approach enables high-quality few-shot image synthesis by transferring smooth and realistic image transitions of pre-trained GANs to diverse few-shot domains. + +# 3. Method + +In the task of few-shot GAN adaptation, we are given a small target dataset $T$ and a pre-trained GAN model, consisting of a discriminator $D$ and a generator $G_{s}$ , which produces an image $x = G_{s}(z)$ from a continuous input variable $z$ , such as a random noise vector or a continuous class embedding. The goal is to adapt the generator to the target dataset such that it generates diverse and realistic images in the domain of the target dataset as shown in Fig. 1. We denote the adapted target generator by $G_{t}$ . + +To achieve few-shot synthesis with a high image quality and diversity, our model should adhere to the following two properties. Firstly, the generator $G_{t}$ should not only memorize and generate the target images, which will be addressed by the smoothness similarity regularization (Sec. 3.1). Secondly, the discriminator $D$ must avoid overfitting to the few target images in order to provide useful supervision for $G_{t}$ (Sec. 3.2). The overview of our method is shown in Fig. 2. + +# 3.1. Smoothness similarity regularization for $G_{t}$ + +In a low data regime like ours, $G_{t}$ can easily overfit to the target dataset $T$ and collapse to reproducing only the few modes represented in the training data. When walking in the latent space of such a generator, one would observe "staircase" patterns, where minor shifts in the latent space cause discontinuous transitions in the output image space (as shown in row 4 of Fig. 5). Naturally, to achieve a synthesis of high diversity, it is desirable for $G_{t}$ to avoid such discontinuities, as having smoother image transitions allows to generate intermediate samples that can exhibit novel features. Therefore, in our model we aim to encourage $G_{t}$ to produce smooth latent space interpolations, in which all the intermediate images are realistic. + +Our approach is based on the observation that GANs trained on large datasets tend to have a well-structured latent space [34, 9, 30], in which different latent space directions can lead to smooth and interpretable image transitions. For example, in a generator pre-trained on a large dataset of churches, latent directions can emerge causing smooth zooming or translation of churches (see Fig. 2). Our observation is that the nature of such image transitions (e.g., zooming or translation) is remarkably general. Thus, we propose a regularizer that utilizes this smoothness property of the source generator $G_{s}$ as a cue while adapting it to another image domain, which can be very different from the domain that was used for pre-training. For example, as shown in Fig. 2, the same latent directions of churches can cause similar zooming or translation effects on shells. + +Mathematically, the smoothness of the generator can be represented via a Jacobian matrix $J_{G^l}(z) = ||\partial G^l (z) / \partial z||$ , quantifying how the generator's intermediate features after the $l$ -th block change under local shifts in the latent space. As we want the same latent shift to cause perceptually similar image transitions in the source and target domains, we design a regularization term that brings the Jacobian matrices of $G_s^l$ and $G_t^l$ closer together. As the computation of full Jacobian matrices is expensive, we use an unbiased estimator of their products with a Gaussian vector [6, 15], which can be computed with standard back-propagation: + +$$ +J _ {G ^ {l}} ^ {T} (z) \cdot y = \mathbb {E} _ {(y) \sim N (0, 1)} \nabla_ {z} \langle G ^ {l} (z), y \rangle , \tag {1} +$$ + +where $y$ is a Gaussian tensor of the same shape as $G^l$ . Our smoothness similarity regularization is then expressed as: + +$$ +\mathcal {L} _ {S S} = \lambda_ {S S} \cdot \mathbb {E} _ {(z, y) \sim N (0, 1)} \| \nabla_ {z} \langle G _ {s} ^ {l} (z), y \rangle - \nabla_ {z} \langle G _ {t} ^ {l} (z), y \rangle \| _ {2}, \tag {2} +$$ + +where $\lambda_{SS}$ steers the impact of the regularizer. As shown in Fig. 2, the smoothness similarity regularization depends on both generators, but only $G_{t}$ is updated. It is interesting to note that the Jacobian matrix is also used for the path length regularization [15], which forces $J_{G}(z)$ to be orthogonal up to a global scale at any $z$ . While this alternative regularizer also induces some form of smoothness, it does not transfer the inherently learned smoothness + +of a pre-trained GAN. We show in Sec. 4.1 that it struggles to enforce the realism of intermediate images. Furthermore, our approach shares the motivation with some prior regularization approaches that use noise perturbations to enforce diversity [23, 37]. In contrast to Eq. 2, these approaches incorporate non-gradient components, e.g., assuming similarity of images $G_{s}(z)\leftrightarrow G_{t}(z)$ or distributions $d(G_{t}(z_{1}),G_{t}(z_{2}))\leftrightarrow d(G_{s}(z_{1}),G_{s}(z_{2}))$ . As such assumptions are violated when source and target domains are dissimilar, they perform poorly compared to our smoothness similarity regularization $\mathcal{L}_{SS}$ as shown in the experiments. + +# 3.2. Revisiting the $D$ adversarial loss + +To identify what kind of image transitions look realistic for the target domain, $G_{t}$ requires strong supervision from the discriminator on image realism at different semantic scales. This includes the colors and textures of objects, as well as object shapes, especially if their distribution is different from the shapes of objects in the source domain. Learning the concept of image realism in low data regimes is, however, challenging due to the problem of overfitting. + +Typically, a GAN discriminator consists of several consecutive blocks $\{D^i\}_{i = 1}^N$ and computes for each given image $x$ a real/fake logit after the last block $l = s^{N}\circ D^{N}(x)$ where $s^N$ is a final processing layer such as a convolution. When adapting such a discriminator to a very small dataset, it is prone to memorizing the training set [32], leading to mode collapse and poor diversity of synthesized images [23]. A possible solution [23, 37] to overcome memorization is to use variants of the PatchGAN discriminator [11], discarding the latest discriminator layers: $l = s^{k}\circ D^{k}(x),k < N$ . This solution allows to adapt colors and textures of generated images to the target domain while avoiding the memorization problem. However, it naturally has a limited capacity to learn more high-level semantic scene properties such as the shapes of objects, which we show in the experiments. + +In order to avoid memorization, and yet to balance the adaptation of colors, textures, and shapes of generated objects to a new domain, we hypothesize that a more flexible attention to different levels of image realism is required by the discriminator. To this end, we perform a simple yet efficient modification to the loss function of the discriminator. Given a discriminator $\{D^i\}_{i = 1}^N$ and its adversarial loss function $\mathcal{L}_D(l)$ used for pre-training (e.g., cross-entropy or hinge loss), we design the discriminator to produce real/fake logits after each discriminator's block, and correspondingly compute the loss as the average across all blocks: + +$$ +\mathcal {L} _ {a l l} (x) = \frac {1}{N} \sum_ {i = 1} ^ {N} \mathcal {L} _ {\mathcal {D}} \left[ l ^ {i} (x) \right], \quad l ^ {i} (x) = s ^ {i} \circ D ^ {i} (x). \tag {3} +$$ + +With the new objective, $D$ is given more freedom to utilize the features extracted at different scales to compute the + +loss. Our finding is that $D$ dynamically adapts the magnitude of the loss at each scale to the target domain, without explicit supervision (see Fig. 6). Consequently, we observe a strong overall stabilization effect on the adaptation performance across diverse source-target dataset pairs. + +# 4. Experiments + +To demonstrate that our approach for few-shot GAN adaptation can be applied to unconditional and class-conditional GANs, we selected for each category a popular GAN architecture: unconditional StyleGANv2 [15] and class-conditional BigGAN [2]. For both models, we test our approach on a variety of source-target domain pairs. We focus on 10-shot target adaptation in the main paper, but we provide results for 1-shot and 5-shot adaptation in the supplementary material. For fair comparisons with prior works, most of our ablations and comparisons are conducted with StyleGANv2. + +# 4.1. Adaptation of unconditional GAN + +Datasets. In contrast to previous works that mostly considered pairs of similar datasets like Face $\rightarrow$ Sketch and Face $\rightarrow$ Sunglasses, we extend the protocol by including structurally dissimilar pairs of source and target domains, which is a more challenging task and is our primary interest. As source generators, we use StyleGANv2 checkpoints pretrained on FFHQ [14], LSUN-Church, and LSUN-Horse [38]. For the target datasets, we selected 10-shot subsets of various commonly used few-shot datasets, such as Anime-Face, Shells, or Pokemons [41, 18]. Results on more datasets are shown in the supplementary material. + +Training details. We fine-tune StyleGANv2 using the $\mathcal{L}_{SS}$ and $\mathcal{L}_{all}$ loss terms as presented in Sec. 3. For the smoothness similarity regularization, we use the intermediate features $G^{l}$ at resolution $(32\times 32)$ and set $\lambda_{SS} = 5.0$ . We follow [23] in choosing all the other hyperparameters, such as image resolution $(256\times 256)$ , learning rates, and batch size. Our experiments across all datasets use the same model configuration and set of hyperparameters. + +Baselines. We compare our method to most recent few-shot GAN adaptation approaches: TGAN [36], FreezeD [20], CDC [23], RSSA [37], and AdAM [39]. In addition, we compare our proposed smoothness similarity regularizer $\mathcal{L}_{SS}$ to other regularization techniques: path length regularization (PPL) [15] and MixDL [16]. + +Evaluation. In low data regimes, it is necessary to judge results both in quality and diversity aspects, as there is a trade-off between them [25, 32]. We measure the quality with FID [10] between a held-out validation set and a generated set of the same size. Following [23], we evaluate diversity with the intra-LPIPS, clustering generated images according to their nearest training samples and computing + +![](images/0ecb19495f802ecc0a2969200d6588e807c13f045f459d9dadef380b8bc339ce.jpg) +10-shot Anime-Faces + +![](images/755207d03d3422ecc244f636825bc5b35919ee3a036e0f615e21e6a62b5b9732.jpg) +TAN Freezed CDC RSSA AdAM Ours Source TGAN Freezed CDC RSSA AdAM Ours Source + +![](images/e381ce5a57f8934ec2c9d3cd42500c3f3b6555e7a51c20067500f40114d8ae19.jpg) +10-shot Shells + +![](images/04c4721c37231e94ed46e5170d28281e7f0fb359a385802ce1dc6501d079b875.jpg) +Figure 3. Visual comparison to prior methods on Face $\rightarrow$ Anime and Church $\rightarrow$ Shells, the source-target dataset pairs with a dissimilar structure (e.g., shapes of objects). In this challenging regime, we observe that prior methods suffer from training instabilities, memorization issues, or inability to adapt the shapes of objects to the new domain. In contrast, our method generates images that look realistic, flexibly combine features of different target images, and transfer the variation of images from the source domain to the target domain. + +the average LPIPS [40] of all the clusters. We train all models for 30k epochs in case of dissimilar domain pairs and for 5k on closer domains, evaluating metrics every 1k epochs. Final checkpoints in all experiments correspond to best FID. + +Results with dissimilar source-target domains. We first present our results on the source-target domain pairs with dissimilar structure: Face $\rightarrow$ Anime, Church $\rightarrow$ Shells, + +and Horse $\rightarrow$ Pokemon (see Fig. 3 and supplementary material). Our general observation from Fig. 3 is that in this challenging regime prior methods suffer either from training instabilities, memorization issues, or inability to adapt the shape of objects to the new domain. For example, for Face $\rightarrow$ Anime, despite an apparent correspondence between the two domains, none of the prior methods success- + +![](images/5b307de5acf42f68a394289aca640efaafb14b63e815e25ccf920ca7a8a3e69a.jpg) +Figure 4. Visual comparison to most recent prior methods on Face $\rightarrow$ Sketch and Church $\rightarrow$ Sunglasses, the dataset pairs depicting similar image domains. In this regime, our method performs on par with previous state of the art. See Table 2 for a quantitative comparison. + +![](images/b62d35529d84d6944182a608b06629941dee6a06949cd931ac22659e88d5d19c.jpg) + +
MethodFace→AnimeChurch→ShellsHorse→Pokemons
FID↓LPIPS↑FID↓LPIPS↑FID↓LPIPS↑
TGAN [36]153.20.29205.30.22115.00.52
FreezeD [20]112.40.22180.80.27123.30.49
CDC [23]140.20.50187.90.48109.50.55
RSSA [37]133.20.37182.40.44117.30.54
AdAM [39]116.40.42152.40.28106.50.55
Ours97.30.57140.50.5384.10.61
+ +fully transfers the distribution of head poses to the anime style, e.g., overfitting too strongly to the 10 provided samples (FreezeD), failing to adapt the shape of faces to the style of anime (CDC), or not generating high-quality animenefaces due to instabilities (TGAN, RSSA, AdAM). Similarly, for $\text{Church} \rightarrow \text{Shells}$ , we observe that prior methods produce only copies of the example shells (FreezeD, AdAM), generate shells of unrealistic church-like shapes (CDC, RSSA), or suffer from instabilities (TGAN). In contrast, our method achieves high-quality synthesis, in which the generated images (i) look like realistic anime-faces and shells; (ii) flexibly combine features observed in different target images (e.g., anime hair color can be combined with various eye colors or background styles); and (iii) meaningfully transfer the variation of images from the source domain (e.g., generated shells adjust to the positions and shapes of churches). + +The quantitative comparison in Table 1 confirms our analysis, where our method achieves the best quality and diversity scores across all datasets. We note a high average relative improvement of more than $18\%$ and $11\%$ in FID and LPIPS compared to the highest scores achieved by prior methods. Overall, we conclude that our method significantly improves over prior works on few-shot GAN + +Table 1. Comparison of the adaptation performance in case of dissimilar source-target domains. Bold denotes best performance. + +
MethodFace→SketchFace→Sunglasses
FID↓LPIPS↑FID↓LPIPS↑
TGAN [36]54.20.3836.80.56
FreezeD [20]48.80.3232.00.59
CDC [23]54.20.4030.50.59
RSSA [37]61.40.4536.30.58
AdAM [39]56.30.3731.10.60
Ours45.20.4427.50.60
+ +Table 2. Comparison in case of structurally close source-target domains. Bold denotes best performance. + +adaptation with dissimilar source and target domains. + +Results with close source-target domains. Next, we follow the evaluation of prior works and compare the models on similar source and target domains, such as adaptation of human faces to a different style. The visual results for Face $\rightarrow$ Sketch and Face $\rightarrow$ Sunglasses are shown in Fig. 4. Our method successfully performs the few-shot adaptation in this setting, adapting the colors and textures of faces to the gray-scale sketch domain, or adding a novel attribute (sunglasses). We note that our method is not explicitly designed to transfer all the details of a face from the source domain, thus changes in the generated images like facial hair are expected. Yet, we observe that our method generally does not lose distinctive features of faces in source images, performing on par with previous state-of-the-art methods. The quantitative comparison is provided in Table $2^{1}$ : on both datasets our method achieves the best FID scores and performs on par with the best performer in LPIPS. + +Ablations. We demonstrate the importance of our proposed loss terms in Fig. 5, which shows latent space interpolations of trained models and their similarity to the pre + +![](images/4755dec36406e5b280bc981bfe1ca884e4f1926cf0de83e7373ff4f0e24b556f.jpg) +Figure 5. Latent space interpolations of the source generator and the ablation models from Tables 3-4. Leftmost and rightmost columns show the used $D$ loss and $G$ smoothness regularization. + +
D lossSmooth reg. for GFace→AnimeChurch→Shells
FID↓LPIPS↑FID↓LPIPS↑
StyleGANv2-178.00.21243.80.17
StyleGANv2SS (ours)180.70.61252.80.62
PatchGAN [23]-145.20.37183.10.31
PatchGAN [23]SS (ours)132.20.55184.20.56
Lall (ours)-116.40.36175.40.43
Lall (ours)SS (ours)97.30.57140.50.53
+ +trained source model $G_{s}$ (row 1). Firstly, we note that the plain StyleGANv2 model (row 2) suffers from instabilities in our low data regime, achieving poor image quality and diversity and having "staircase"-like latent space interpolations. Applying $\mathcal{L}_{SS}$ without $\mathcal{L}_{all}$ (row 3) helps to achieve diverse synthesis with smooth interpolations, but it is not enough to achieve good image quality. On the other hand, using $\mathcal{L}_{all}$ (row 4) helps to overcome instabilities and improve image quality, but it cannot maintain smooth interpolations and high diversity. Finally, our full model (row 5) allows a higher-quality, diverse synthesis with smooth and realistic latent space interpolations. Note how the image transitions mimic the behaviour of the source model (churches and shells change shapes and positions similarly), allowing to achieve diverse and realistic synthesis. + +The effect of $\mathcal{L}_{all}$ is further demonstrated in Fig. 6, where we show the contribution of different $D$ blocks to the adversarial loss at different epochs. We note the ability of the discriminator to identify correct loss contributions + +![](images/598b03f082b3faefeeaa68bdfddecd7ce535e606436fbd31756c97770494dad4.jpg) +Figure 6. The contribution of features at different $D$ blocks to the adversarial loss function $\mathcal{L}_{all}$ . For two closer image domains (the left plot), the network concentrates mostly on earlier layers to compute the loss, while for less similar domains the network learns to use the later layers representing more high-level $D$ features. + +Table 3. Impact of ${\mathcal{L}}_{\text{all }}$ and ${\mathcal{L}}_{SS}$ . Bold denotes best performance. + +
D lossSmooth reg. for GFace→AnimeChurch→Shells
FID↓LPIPS↑FID↓LPIPS↑
Lall(ours)-116.40.36175.40.43
Lall(ours)PPL [14]107.80.46179.40.44
Lall(ours)MixDL [16]105.90.50150.40.51
Lall(ours)SS (ours)97.30.57140.50.53
+ +Table 4. Comparison of smoothness similarity regularization $\mathcal{L}_{SS}$ with other regularizers. Bold denotes best performance. + +adaptively for different source-target domain pairs. For example for Face $\rightarrow$ Anime, the network concentrates mostly on the earliest $D$ blocks to adapt the colors and textures of faces to a new style. In contrast, for the more distant domains Church $\rightarrow$ Shells, the network learns to attribute a higher weight to the later blocks to also adapt higher-level features, such as shapes of objects. In effect, we observe a stabilized adaptation of colors, textures, and shapes of objects across diverse source-target pairs. Using PatchGAN [23] as discriminator loss does not achieve such a balance as it focuses mostly on lower-scale features (row 6 in Fig. 5). + +Our observations are confirmed by the quantitative study in Table 3: without $\mathcal{L}_{SS}$ the model does not achieve high diversity (high LPIPS), while $\mathcal{L}_{all}$ is necessary for high image quality (low FID). We conclude that both our proposed loss terms are important to achieve high-quality synthesis. More ablations on $\mathcal{L}_{SS}$ and $\mathcal{L}_{all}$ can be found in the supplementary material. + +Lastly, Table 4 provides a comparison of our proposed $\mathcal{L}_{SS}$ loss term to other regularizers: path length regularization (PPL) [14] and MixDL [16]. While all regularizers help to achieve smoother latent space interpolations and thus improve the quality and diversity metrics, our smoothness similarity regularization enables the highest performance in both FID and LPIPS. While our approach transfers the learned smoothness of the source generator to the target domain, PPL and MixDL resort to gradually interpolating between the provided training samples, which leads to latent space interpolations that either look unrealistic or lack diversity (rows 7-8 in Fig. 5). This demonstrates that transferring smoothness from a pre-trained generator is beneficial to enforce image transitions that are realistic and diverse. + +![](images/89f2496245519eeb260bee46876711f937542fa6af06bf6e628c2eac6006e71b.jpg) + +![](images/b1024f67326930cdb76f583a23ccb5cb80713e8a6af5592c5537c890752315b1.jpg) + +![](images/7ec0596780f910ed1bc441fbc7016076181f7b5bdb666831af14b45941a3cc73.jpg) +10-Shot adaptation results of our method on class-conditional BigGAN [2], pre-trained on ImageNet +Figure 7. 10-shot adaptation results for the class-conditional BigGAN [2] pre-trained on ImageNet. While simple fine-tuning (FT) suffers from training instabilities and mode collapse, our method helps to achieve much higher image quality and diversity, transferring smooth and realistic image transitions from the source domain, e.g., objects smoothly changing their locations, size, and shape. + +
D lossSmooth reg. for GImageNet→FlowersImageNet→Pokemons
FID↓LPIPS↑FID↓LPIPS↑
BigGAN-213.30.29226.80.15
BigGANSS (ours)225.60.47208.30.47
Lall (ours)-123.90.28129.40.27
Lall (ours)SS (ours)106.40.5589.60.56
+ +Table 5. Ablation on the performance when adapting the class-conditional BigGAN model [2] pre-trained on ImageNet. + +# 4.2. Adaptation of class-conditional GAN + +Our approach is not limited to unconditional GANs, but it can also be applied to a class-conditional GAN model. We selected BigGAN [2] for our experiments as it is a popular backbone architecture for class-conditional image synthesis on ImageNet [7]. We make two modifications to enable the adaptation of the model to unconditional few-shot datasets. Firstly, we remove the conditioning of the discriminator via the projection layer [19]. Secondly, we treat the generator's learned continuous class embedding as part of the latent space, thus sampling a Gaussian vector in the joint noise-class space at each fine-tuning epoch. This way, the generator produces an image based on a single input vector in an unconditional fashion. We then fine-tune the pre-trained model using our loss terms $\mathcal{L}_{SS}$ and $\mathcal{L}_{all}$ as presented in Sec. 3. We use image resolution $256\times 256$ and batch size of 32. The hyperparameters for $\mathcal{L}_{SS}$ are the same as for StyleGANv2: intermediate features $G^{l}$ at resolution $(32\times 32)$ and $\lambda_{SS} = 5.0$ . We train for 30k epochs and select checkpoints by best FID. + +Datasets. As the source generator, we use the Big- + +GAN checkpoint pre-trained on class-conditional ImageNet at resolution $256 \times 256$ . We demonstrate 10-shot adaptation results with two commonly used few-shot generation datasets: Oxford-Flowers [21] and Pokémon [18]. We use the same model configuration for both datasets. + +Results. Fig. 7 demonstrates latent space interpolations of the source and target generators. We note that a simple fine-tuning of BigGAN suffers from training instabilities and mode collapse. In contrast, our method successfully adapts BigGAN to generate diverse images in the target domains. We highlight that our method transfers smooth and realistic image transitions from the well-learned BigGAN's noise-class space, despite significant dissimilarities between ImageNet and the few-shot datasets, in particular Pokemons. For example, it can be noticed how the latent space interpolations in the target domains mimic the source domain, e.g., the generated flowers andPokemons change their position and size similarly to dogs and wolves $(5^{\mathrm{th}}-$ $10^{\mathrm{th}}$ columns in Fig. 7) or stretch their shape to mimic the proportions of busses $(11^{\mathrm{th}} - 14^{\mathrm{th}}$ columns). + +Table 5 shows the importance of our proposed loss terms. Our observations are consistent with the ablations with StylGANv2: $\mathcal{L}_{all}$ is necessary to avoid instabilities and achieve a good image quality (low FID), while $\mathcal{L}_{SS}$ is required to achieve smooth latent space interpolations and good diversity (high LPIPS). We conclude that our method successfully extends to the adaptation of class-conditional models, where target domains benefit from the rich noise-class space learned on a multi-class dataset such as ImageNet. More details and results are provided in the supplementary material. + +# 5. Conclusion + +In this work, we presented a new method for few-shot adaptation of GAN models. It transfers the smooth latent space of a pre-trained GAN, which was trained on a large dataset, to a new domain with very few images. We addressed the case of few-shot GAN adaptation when the source and target domains are structurally dissimilar, which is a common issue in applications. Our extensive results demonstrate that in this setting our approach outperforms previous works in terms of image quality and diversity. + +# Acknowledgement + +The work has been supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC 2070 -390732324 and the ERC Consolidator Grant FORHUE (101044724). We thank Jinhui Yi for providing and analyzing the sugar beet data used in supplementary experiments. + +# References + +[1] Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-shot segmentation without meta-learning: A good transductive inference is all you need? In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. +[2] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In International Conference on Learning Representations (ICLR), 2019. +[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[4] Yunjoy Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. +[5] Kaiwen Cui, Jiaxing Huang, Zhipeng Luo, Gongjie Zhang, Fangneng Zhan, and Shijian Lu. Genco: Generative constraining for generative adversarial networks with limited data. In Conference on Artificial Intelligence (AAAI), 2021. +[6] Yann Dauphin, Harm De Vries, and Yoshua Bengio. Equilibrated adaptive learning rates for non-convex optimization. In Advances in Neural Information Processing Systems (NeurIPS), 2015. +[7] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In Conference on Computer Vision and Pattern Recognition (CVPR), 2009. +[8] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. + +[9] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In Advances in Neural Information Processing Systems (NeurIPS), 2020. +[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems (NeurIPS), 2017. +[11] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. +[12] Tero Karras, Miika Aittala, Janne Hellsten, S. Laine, J. Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Advances in Neural Information Processing Systems (NeurIPS), 2020. +[13] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Advances in Neural Information Processing Systems (NeurIPS), 2021. +[14] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. +[15] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. +[16] Chaerin Kong, Jeesoo Kim, Donghoon Han, and Nojun Kwak. Few-shot image generation with mixup-based distance learning. In European Conference on Computer Vision (ECCV), 2022. +[17] Yijun Li, Richard Zhang, Jingwan Cynthia Lu, and Eli Shechtman. Few-shot image generation with elastic weight consolidation. In Advances in Neural Information Processing Systems (NeurIPS), 2020. +[18] Bingchen Liu, Yizhe Zhu, Kunpeng Song, and Ahmed Elgammal. Towards faster and stabilized gan training for high-fidelity few-shot image synthesis. In International Conference on Learning Representations (ICLR), 2021. +[19] Takeru Miyato and Masanori Koyama. cGANs with projection discriminator. In International Conference on Learning Representations (ICLR), 2018. +[20] Sangwoo Mo, Minsu Cho, and Jinwoo Shin. Freeze discriminator: A simple baseline for fine-tuning gans. In CVPR AI for Content Creation Workshop, 2020. +[21] Maria-Elena Nilsback and Andrew Zisserman. A visual vocabulary for flower classification. In Conference on Computer Vision and Pattern Recognition (CVPR), 2006. +[22] Atsuhiro Noguchi and T. Harada. Image generation from small datasets via batch statistics adaptation. In International Conference on Computer Vision (ICCV), 2019. +[23] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. + +[24] Jie Ren, Peter J Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark Depristo, Joshua Dillon, and Balaji Lakshminarayanan. Likelihood ratios for out-of-distribution detection. Advances in Neural Information Processing Systems (NeurIPS), 2019. +[25] Esther Robb, Wen-Sheng Chu, Abhishek Kumar, and Jia-Bin Huang. Few-shot adaptation of generative adversarial networks. arXiv:2010.11943, 2021. +[26] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis. arXiv preprint arXiv:2301.09515, 2023. +[27] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleganxI: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH, 2022. +[28] Edgar Schonfeld, Vadim Sushko, Dan Zhang, Juergen Gall, Bernt Schiele, and Anna Khoreva. You only need adversarial supervision for semantic image synthesis. In International Conference on Learning Representations (ICLR), 2021. +[29] Tamar Rott Shaham, Tali Dekel, and T. Michaeli. SinGAN: Learning a generative model from a single natural image. In International Conference on Computer Vision (ICCV), 2019. +[30] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. +[31] Vadim Sushko, Juergen Gall, and Anna Khoreva. One-Shot GAN: Learning to generate samples from single images and videos. In Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2021. +[32] Vadim Sushko, Dan Zhang, Juergen Gall, and Anna Khoreva. Learning to generate novel scene compositions from single images and videos. arXiv:2103.13389, 2021. +[33] Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. Transactions on Pattern Analysis and Machine Intelligence, 2020. +[34] Andrey Voynov and Artem Babenko. Unsupervised discovery of interpretable directions in the gan latent space. In International conference on machine learning (ICML), 2020. +[35] Yaxing Wang, Abel Gonzalez-Garcia, David Berga, L. Herranz, F. Khan, and Joost van de Weijer. Minegan: Effective knowledge transfer from gans to target domains with few images. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. +[36] Yaxing Wang, Chenshen Wu, L. Herranz, Joost van de Weijer, Abel Gonzalez-Garcia, and B. Raducanu. Transferring gans: generating images from limited data. In European Conference on Computer Vision (ECCV), 2018. +[37] Jiayu Xiao, Liang Li, Chaofei Wang, Zheng-Jun Zha, and Qingming Huang. Few shot generative model adaption via relaxed spatial structural alignment. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[38] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv:1506.03365, 2015. + +[39] Zhao Yunqing, Keshigeyan Chandrasegaran, Milad Abdol-lahzadeh, and Ngai-man Cheung. Few-shot image generation via adaptation-aware kernel modulation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. +[40] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Conference on Computer Vision and Pattern Recognition (CVPR), 2018. +[41] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. In Conference on Neural Information Processing Systems (NeurIPS), 2020. +[42] Yunqing Zhao, Henghui Ding, Houjing Huang, and Ngai-Man Cheung. A closer look at few-shot image generation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. \ No newline at end of file diff --git a/2023/Smoothness Similarity Regularization for Few-Shot GAN Adaptation/images.zip b/2023/Smoothness Similarity Regularization for Few-Shot GAN Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..38896d1e053632a9309ed9399d0da6c0e4573706 --- /dev/null +++ b/2023/Smoothness Similarity Regularization for Few-Shot GAN Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b239aae2468c42c714a57ba5d239f54a790a23425b4e3cb06fd0592cbec08a3a +size 1216301 diff --git a/2023/Smoothness Similarity Regularization for Few-Shot GAN Adaptation/layout.json b/2023/Smoothness Similarity Regularization for Few-Shot GAN Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..717762d1dae6f937f31a7bde3e9c5b460c41db3e --- /dev/null +++ b/2023/Smoothness Similarity Regularization for Few-Shot GAN Adaptation/layout.json @@ -0,0 +1,7902 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "type": "text", + "content": "Smoothness Similarity Regularization for Few-Shot GAN Adaptation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 142, + 431, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 142, + 431, + 158 + ], + "spans": [ + { + "bbox": [ + 160, + 142, + 431, + 158 + ], + "type": "text", + "content": "Vadim Sushko" + }, + { + "bbox": [ + 160, + 142, + 431, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 160, + 142, + 431, + 158 + ], + "type": "text", + "content": " Ruyu Wang" + }, + { + "bbox": [ + 160, + 142, + 431, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 160, + 142, + 431, + 158 + ], + "type": "text", + "content": " Juergen Gall" + }, + { + "bbox": [ + 160, + 142, + 431, + 158 + ], + "type": "inline_equation", + "content": "^{2,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 164, + 159, + 430, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 159, + 430, + 171 + ], + "spans": [ + { + "bbox": [ + 164, + 159, + 430, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 164, + 159, + 430, + 171 + ], + "type": "text", + "content": " Bosch Center for Artificial Intelligence " + }, + { + "bbox": [ + 164, + 159, + 430, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 164, + 159, + 430, + 171 + ], + "type": "text", + "content": " University of Bonn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 164, + 172, + 430, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 172, + 430, + 185 + ], + "spans": [ + { + "bbox": [ + 164, + 172, + 430, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 164, + 172, + 430, + 185 + ], + "type": "text", + "content": "Lamarr Institute for Machine Learning and Artificial Intelligence" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 188, + 481, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 188, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 111, + 188, + 481, + 198 + ], + "type": "text", + "content": "vad221@gmail.com ruyu.wang@de.bosch.com gall@iai.uni-bonn.de" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 221, + 191, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 221, + 191, + 232 + ], + "spans": [ + { + "bbox": [ + 143, + 221, + 191, + 232 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 242, + 290, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 290, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 290, + 445 + ], + "type": "text", + "content": "The task of few-shot GAN adaptation aims to adapt a pre-trained GAN model to a small dataset with very few training images. While existing methods perform well when the dataset for pre-training is structurally similar to the target dataset, the approaches suffer from training instabilities or memorization issues when the objects in the two domains have a very different structure. To mitigate this limitation, we propose a new smoothness similarity regularization that transfers the inherently learned smoothness of the pre-trained GAN to the few-shot target domain even if the two domains are very different. We evaluate our approach by adapting an unconditional and a class-conditional GAN to diverse few-shot target domains. Our proposed method significantly outperforms prior few-shot GAN adaptation methods in the challenging case of structurally dissimilar source-target domains, while performing on par with the state of the art for similar source-target domains." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 454, + 127, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 127, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 127, + 466 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "content": "Generative adversarial networks (GANs) have been shown to be powerful at various image synthesis tasks [4, 28, 3, 13, 27, 26]. The success of these models is in large part enabled by the availability of large datasets for training, typically consisting of thousands of images. However, there are many applications and computer vision tasks such as one-shot or few-shot learning [1, 33], out-of-distribution detection [24], or long-tailed recognition tasks [8] where the number of available training images is very low." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "Since training a GAN from scratch on very few samples does not perform well as shown in Fig. 1, a common strategy is to fine-tune a pre-trained GAN model on the few-shot dataset, typically employing additional regularization losses to penalize the degradation of the diversity [23, 37]. This approach, referred to as few-shot GAN adaptation, performs well when the target domain is structurally very similar to the dataset that has been used for pre-training, e.g., photographs vs. sketches of human faces. However, the performance drastically degrades in case of large dissimilarities between the source and target domain as shown in Fig. 1." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 219, + 545, + 434 + ], + "blocks": [ + { + "bbox": [ + 307, + 219, + 545, + 434 + ], + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 434 + ], + "type": "image", + "image_path": "21e06df4d0f06758a31fb4db90ba718d82a6587a51c4aa78b1cfae86feea72a4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 435, + 547, + 513 + ], + "lines": [ + { + "bbox": [ + 305, + 435, + 547, + 513 + ], + "spans": [ + { + "bbox": [ + 305, + 435, + 547, + 513 + ], + "type": "text", + "content": "Figure 1. Training a GAN model " + }, + { + "bbox": [ + 305, + 435, + 547, + 513 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 305, + 435, + 547, + 513 + ], + "type": "text", + "content": " on a few-shot dataset (row 1) from scratch fails due to training instabilities (row 2). We thus aim to adapt a GAN " + }, + { + "bbox": [ + 305, + 435, + 547, + 513 + ], + "type": "inline_equation", + "content": "G_{s}" + }, + { + "bbox": [ + 305, + 435, + 547, + 513 + ], + "type": "text", + "content": " that has been pre-trained on a large dataset like LSUN-Church (row 3) to the target few-shot dataset " + }, + { + "bbox": [ + 305, + 435, + 547, + 513 + ], + "type": "inline_equation", + "content": "(G_{t})" + }, + { + "bbox": [ + 305, + 435, + 547, + 513 + ], + "type": "text", + "content": ". While fine-tuning [23] does not perform well either if source and target are dissimilar (row 4), our approach generates diverse and realistic images (row 5) by transferring the smoothness properties of " + }, + { + "bbox": [ + 305, + 435, + 547, + 513 + ], + "type": "inline_equation", + "content": "G_{s}" + }, + { + "bbox": [ + 305, + 435, + 547, + 513 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 520, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 605 + ], + "type": "text", + "content": "Such dissimilarities are a major bottleneck of using GANs in other disciplines like medicine, production, or crop science, where there is a lack of large datasets due to privacy, confidentiality, or simply lack of data. Motivated by this fact, we extend the protocol for few-shot GAN adaptation by investigating also pairs of datasets that are very different like churches and shells as shown in Fig. 1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "To improve few-shot GAN adaptation in the case of structurally dissimilar pairs, we propose a new GAN adaptation strategy. Firstly, we propose a new smoothness similarity regularization for the generator. Our key observation is that pre-trained GAN generators, regardless of the exact structure of objects in the pre-training dataset, learn well-structured and smooth latent spaces. For example, prior works demonstrated that various local shifts in the latent space can lead to interpretable and smooth transitions of" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7073" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 203 + ], + "type": "text", + "content": "output images, such as translation of objects in the scene or changing their size [34, 9, 30]. As we show in our experiments, the proposed smoothness similarity regularization enables the transfer of this desirable property to other few-shot image domains without compromising the synthesis quality. Secondly, to overcome overfitting issues, we revisit the adversarial loss function of the discriminator and propose a simple yet efficient modification by computing the loss at different layers of the discriminator. This leads to the mitigation of overfitting and a more stabilized adaptation of the model to diverse target domains." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 205, + 286, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 205, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 205, + 286, + 453 + ], + "type": "text", + "content": "We evaluate our approach by adapting an unconditional [15] and a class-conditional GAN [2] to diverse few-shot target domains. Our model significantly outperforms previous state-of-the-art methods in image quality and diversity in the challenging case of dissimilar source and target domains, while performing on par with the state of the art on structurally similar dataset pairs. In summary, our contributions are as follows: (i) We extend the evaluation protocol for few-shot GAN adaptation by including new dataset pairs that are structurally much less similar than was considered in prior work. (ii) We propose a new smoothness similarity regularization, which enables diverse synthesis in the target domain by transferring the learned smoothness of a pretrained GAN. (iii) We revisit the adversarial loss function of the discriminator to stabilize few-shot GAN adaptation across diverse target domains. (iv) Our proposed model enables high-quality synthesis in the challenging case of dissimilar source and target domains, significantly outperforming prior methods. In addition, we show that our method can be applied to different classes of GAN architectures, including unconditional and class-conditional GAN models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 466, + 132, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 466, + 132, + 478 + ], + "spans": [ + { + "bbox": [ + 49, + 466, + 132, + 478 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 486, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 486, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 49, + 486, + 286, + 664 + ], + "type": "text", + "content": "To address the image generation problem in the low data regime, existing works mainly follow three research lines - one-shot, low-shot, and few-shot learning. One-shot generation methods [29, 31] focus on leveraging the internal patch distribution within a single image, however, their extension to capture the distribution of a small collection of images is non-trivial. In low-shot learning [41], several works [41, 12] proposed to mitigate the limited-data-induced overfitting issue by adapting data augmentations to the generative networks. Others [18, 5] stabilized the training process and reduced overfitting by revising the network design. Despite the promising performance in many low data regimes (typically having " + }, + { + "bbox": [ + 49, + 486, + 286, + 664 + ], + "type": "inline_equation", + "content": "100+" + }, + { + "bbox": [ + 49, + 486, + 286, + 664 + ], + "type": "text", + "content": " images), these low-shot methods fail in the extremely few-shot setting (e.g., 10 images). Our work lies in the scope of few-shot learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 666, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 666, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 666, + 286, + 712 + ], + "type": "text", + "content": "Few-shot image synthesis. Conventional few-shot learning aims at learning a discriminative classifier under limited data scenarios. In the context of image synthesis with GANs, the goal instead is to produce diverse new im" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 72, + 544, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 479 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 479 + ], + "type": "text", + "content": "ages from the learned distribution while preventing overfitting to the few training samples. A straightforward way is to treat it as a domain adaptation problem and incorporate the commonly used transfer learning technique, i.e., fine-tuning, to ease the need for data. However, naive fine-tuning (TGAN) [36] often suffers from overfitting and results in poor performance. Researchers proposed remedies such as mining suitable parts of the latent space before fine-tuning [35] or restricting weight updates, for example, updating only the BatchNorm parameters of the generator [22], penalizing drastic changes in important weights [17], or freezing the earliest layers of the discriminator (FreezeD) [20]. More recent works focused on introducing different regularizations to preserve specific knowledge from the pretrained model and prevent diversity degradation [42]. For example, CDC [23] proposed to preserve the pair-wise perceptual similarity between samples from the source domain and to transfer it to the target domain, while RSSA [37] designed a novel consistency term to align the structural information between source and target domains. Although the two aforementioned methods constitute the current state of the art in few-shot generative learning, their assumptions impose strong constraints on the structure of the few-shot target domain. As we show in experiments, they fail in the more challenging regime when the source and target domains are not restrictively similar. Most recently, [39] proposed to replace prior knowledge preservation criteria with adaptation-aware kernel modulation (AdAM), which relaxed the source-target proximity requirement of previous methods to some extent. In this work, we take a step further and introduce a new regularization term to preserve the generator's smoothness properties that are not limited to a specific domain, enabling successful adaptation between image domains of unprecedented structural dissimilarity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 486, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 486, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 486, + 544, + 712 + ], + "type": "text", + "content": "Smoothness of image generators. Smooth transitions in the latent space are an important property for generative models, where it is believed to be a sign of a well-conditioned generator. Models trained on large datasets naturally possess this property with or without explicit regularization [2, 15]. For example, StyleGANv2 [15] introduced a regularization based on the perceptual path length measure (PPL) [14], which encourages that a fixed-size step in the latent space results in a fixed-magnitude change in the image space. However, achieving a smooth mapping of the generator is difficult for few-shot image synthesis since there are not enough training samples. Thus, MixDL [16] sought to alleviate the \"staircase\" latent space interpolations, i.e., jumps between training samples, by introducing a continuous coefficient vector and enforcing smooth interpolations between training images. Although the two above regularizers aim to encourage smoother interpolations between training samples and thus mitigate mode collapse, they are not designed to take advantage of the available pre-training" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "7074" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 64, + 539, + 178 + ], + "blocks": [ + { + "bbox": [ + 48, + 64, + 539, + 178 + ], + "lines": [ + { + "bbox": [ + 48, + 64, + 539, + 178 + ], + "spans": [ + { + "bbox": [ + 48, + 64, + 539, + 178 + ], + "type": "image", + "image_path": "6f3dc01222538e80f247fe1115e4e0e41498a62ab8f02299a5901c836d68de64.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 179, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 46, + 179, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 547, + 213 + ], + "type": "text", + "content": "Figure 2. Given a pre-trained generator " + }, + { + "bbox": [ + 46, + 179, + 547, + 213 + ], + "type": "inline_equation", + "content": "G_{s}" + }, + { + "bbox": [ + 46, + 179, + 547, + 213 + ], + "type": "text", + "content": ", the proposed smoothness similarity regularization preserves the learned smoothness of " + }, + { + "bbox": [ + 46, + 179, + 547, + 213 + ], + "type": "inline_equation", + "content": "G_{s}" + }, + { + "bbox": [ + 46, + 179, + 547, + 213 + ], + "type": "text", + "content": " while adapting it to a target domain with very few images. To mitigate overfitting to the target domain, the discriminator loss utilizes features at various layers and automatically adjusts the impact of different semantic scales to the similarity of the source and target domain." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 224, + 288, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 288, + 295 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 288, + 295 + ], + "type": "text", + "content": "knowledge. In contrast, in this work we develop a new smoothness similarity regularization that leverages the well-structured latent space of a pre-trained GAN generator. In effect, our approach enables high-quality few-shot image synthesis by transferring smooth and realistic image transitions of pre-trained GANs to diverse few-shot domains." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 307, + 103, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 307, + 103, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 307, + 103, + 319 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "text", + "content": "In the task of few-shot GAN adaptation, we are given a small target dataset " + }, + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "text", + "content": " and a pre-trained GAN model, consisting of a discriminator " + }, + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "text", + "content": " and a generator " + }, + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "inline_equation", + "content": "G_{s}" + }, + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "text", + "content": ", which produces an image " + }, + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "inline_equation", + "content": "x = G_{s}(z)" + }, + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "text", + "content": " from a continuous input variable " + }, + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "text", + "content": ", such as a random noise vector or a continuous class embedding. The goal is to adapt the generator to the target dataset such that it generates diverse and realistic images in the domain of the target dataset as shown in Fig. 1. We denote the adapted target generator by " + }, + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 435, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 531 + ], + "type": "text", + "content": "To achieve few-shot synthesis with a high image quality and diversity, our model should adhere to the following two properties. Firstly, the generator " + }, + { + "bbox": [ + 46, + 435, + 287, + 531 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 46, + 435, + 287, + 531 + ], + "type": "text", + "content": " should not only memorize and generate the target images, which will be addressed by the smoothness similarity regularization (Sec. 3.1). Secondly, the discriminator " + }, + { + "bbox": [ + 46, + 435, + 287, + 531 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 435, + 287, + 531 + ], + "type": "text", + "content": " must avoid overfitting to the few target images in order to provide useful supervision for " + }, + { + "bbox": [ + 46, + 435, + 287, + 531 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 46, + 435, + 287, + 531 + ], + "type": "text", + "content": " (Sec. 3.2). The overview of our method is shown in Fig. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 539, + 272, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 272, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 272, + 552 + ], + "type": "text", + "content": "3.1. Smoothness similarity regularization for " + }, + { + "bbox": [ + 47, + 539, + 272, + 552 + ], + "type": "inline_equation", + "content": "G_{t}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": "In a low data regime like ours, " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": " can easily overfit to the target dataset " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": " and collapse to reproducing only the few modes represented in the training data. When walking in the latent space of such a generator, one would observe \"staircase\" patterns, where minor shifts in the latent space cause discontinuous transitions in the output image space (as shown in row 4 of Fig. 5). Naturally, to achieve a synthesis of high diversity, it is desirable for " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": " to avoid such discontinuities, as having smoother image transitions allows to generate intermediate samples that can exhibit novel features. Therefore, in our model we aim to encourage " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": " to produce smooth latent space interpolations, in which all the intermediate images are realistic." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 224, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 545, + 403 + ], + "type": "text", + "content": "Our approach is based on the observation that GANs trained on large datasets tend to have a well-structured latent space [34, 9, 30], in which different latent space directions can lead to smooth and interpretable image transitions. For example, in a generator pre-trained on a large dataset of churches, latent directions can emerge causing smooth zooming or translation of churches (see Fig. 2). Our observation is that the nature of such image transitions (e.g., zooming or translation) is remarkably general. Thus, we propose a regularizer that utilizes this smoothness property of the source generator " + }, + { + "bbox": [ + 304, + 224, + 545, + 403 + ], + "type": "inline_equation", + "content": "G_{s}" + }, + { + "bbox": [ + 304, + 224, + 545, + 403 + ], + "type": "text", + "content": " as a cue while adapting it to another image domain, which can be very different from the domain that was used for pre-training. For example, as shown in Fig. 2, the same latent directions of churches can cause similar zooming or translation effects on shells." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 403, + 546, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 546, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 546, + 535 + ], + "type": "text", + "content": "Mathematically, the smoothness of the generator can be represented via a Jacobian matrix " + }, + { + "bbox": [ + 304, + 403, + 546, + 535 + ], + "type": "inline_equation", + "content": "J_{G^l}(z) = ||\\partial G^l (z) / \\partial z||" + }, + { + "bbox": [ + 304, + 403, + 546, + 535 + ], + "type": "text", + "content": ", quantifying how the generator's intermediate features after the " + }, + { + "bbox": [ + 304, + 403, + 546, + 535 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 403, + 546, + 535 + ], + "type": "text", + "content": "-th block change under local shifts in the latent space. As we want the same latent shift to cause perceptually similar image transitions in the source and target domains, we design a regularization term that brings the Jacobian matrices of " + }, + { + "bbox": [ + 304, + 403, + 546, + 535 + ], + "type": "inline_equation", + "content": "G_s^l" + }, + { + "bbox": [ + 304, + 403, + 546, + 535 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 403, + 546, + 535 + ], + "type": "inline_equation", + "content": "G_t^l" + }, + { + "bbox": [ + 304, + 403, + 546, + 535 + ], + "type": "text", + "content": " closer together. As the computation of full Jacobian matrices is expensive, we use an unbiased estimator of their products with a Gaussian vector [6, 15], which can be computed with standard back-propagation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 344, + 541, + 545, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 541, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 344, + 541, + 545, + 557 + ], + "type": "interline_equation", + "content": "J _ {G ^ {l}} ^ {T} (z) \\cdot y = \\mathbb {E} _ {(y) \\sim N (0, 1)} \\nabla_ {z} \\langle G ^ {l} (z), y \\rangle , \\tag {1}", + "image_path": "10ca29036a5bde11a8b46131c67feb51cca37907774c08849b56412b262e0cb8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 564, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 545, + 588 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 564, + 545, + 588 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 564, + 545, + 588 + ], + "type": "text", + "content": " is a Gaussian tensor of the same shape as " + }, + { + "bbox": [ + 304, + 564, + 545, + 588 + ], + "type": "inline_equation", + "content": "G^l" + }, + { + "bbox": [ + 304, + 564, + 545, + 588 + ], + "type": "text", + "content": ". Our smoothness similarity regularization is then expressed as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 594, + 545, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 594, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 307, + 594, + 545, + 617 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S S} = \\lambda_ {S S} \\cdot \\mathbb {E} _ {(z, y) \\sim N (0, 1)} \\| \\nabla_ {z} \\langle G _ {s} ^ {l} (z), y \\rangle - \\nabla_ {z} \\langle G _ {t} ^ {l} (z), y \\rangle \\| _ {2}, \\tag {2}", + "image_path": "5f0ad310273d7b567d05c26e217cf908f0ff9e858a92cb78e4873b0382bd0d1b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{SS}" + }, + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "type": "text", + "content": " steers the impact of the regularizer. As shown in Fig. 2, the smoothness similarity regularization depends on both generators, but only " + }, + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "type": "text", + "content": " is updated. It is interesting to note that the Jacobian matrix is also used for the path length regularization [15], which forces " + }, + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "type": "inline_equation", + "content": "J_{G}(z)" + }, + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "type": "text", + "content": " to be orthogonal up to a global scale at any " + }, + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "type": "text", + "content": ". While this alternative regularizer also induces some form of smoothness, it does not transfer the inherently learned smoothness" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7075" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": "of a pre-trained GAN. We show in Sec. 4.1 that it struggles to enforce the realism of intermediate images. Furthermore, our approach shares the motivation with some prior regularization approaches that use noise perturbations to enforce diversity [23, 37]. In contrast to Eq. 2, these approaches incorporate non-gradient components, e.g., assuming similarity of images " + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "inline_equation", + "content": "G_{s}(z)\\leftrightarrow G_{t}(z)" + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": " or distributions " + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "inline_equation", + "content": "d(G_{t}(z_{1}),G_{t}(z_{2}))\\leftrightarrow d(G_{s}(z_{1}),G_{s}(z_{2}))" + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": " . As such assumptions are violated when source and target domains are dissimilar, they perform poorly compared to our smoothness similarity regularization " + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SS}" + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": " as shown in the experiments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 210, + 222, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 210, + 222, + 222 + ], + "spans": [ + { + "bbox": [ + 47, + 210, + 222, + 222 + ], + "type": "text", + "content": "3.2. Revisiting the " + }, + { + "bbox": [ + 47, + 210, + 222, + 222 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 210, + 222, + 222 + ], + "type": "text", + "content": " adversarial loss" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 228, + 287, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 324 + ], + "type": "text", + "content": "To identify what kind of image transitions look realistic for the target domain, " + }, + { + "bbox": [ + 46, + 228, + 287, + 324 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 46, + 228, + 287, + 324 + ], + "type": "text", + "content": " requires strong supervision from the discriminator on image realism at different semantic scales. This includes the colors and textures of objects, as well as object shapes, especially if their distribution is different from the shapes of objects in the source domain. Learning the concept of image realism in low data regimes is, however, challenging due to the problem of overfitting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 324, + 288, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 288, + 514 + ], + "type": "text", + "content": "Typically, a GAN discriminator consists of several consecutive blocks " + }, + { + "bbox": [ + 46, + 324, + 288, + 514 + ], + "type": "inline_equation", + "content": "\\{D^i\\}_{i = 1}^N" + }, + { + "bbox": [ + 46, + 324, + 288, + 514 + ], + "type": "text", + "content": " and computes for each given image " + }, + { + "bbox": [ + 46, + 324, + 288, + 514 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 324, + 288, + 514 + ], + "type": "text", + "content": " a real/fake logit after the last block " + }, + { + "bbox": [ + 46, + 324, + 288, + 514 + ], + "type": "inline_equation", + "content": "l = s^{N}\\circ D^{N}(x)" + }, + { + "bbox": [ + 46, + 324, + 288, + 514 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 324, + 288, + 514 + ], + "type": "inline_equation", + "content": "s^N" + }, + { + "bbox": [ + 46, + 324, + 288, + 514 + ], + "type": "text", + "content": " is a final processing layer such as a convolution. When adapting such a discriminator to a very small dataset, it is prone to memorizing the training set [32], leading to mode collapse and poor diversity of synthesized images [23]. A possible solution [23, 37] to overcome memorization is to use variants of the PatchGAN discriminator [11], discarding the latest discriminator layers: " + }, + { + "bbox": [ + 46, + 324, + 288, + 514 + ], + "type": "inline_equation", + "content": "l = s^{k}\\circ D^{k}(x),k < N" + }, + { + "bbox": [ + 46, + 324, + 288, + 514 + ], + "type": "text", + "content": ". This solution allows to adapt colors and textures of generated images to the target domain while avoiding the memorization problem. However, it naturally has a limited capacity to learn more high-level semantic scene properties such as the shapes of objects, which we show in the experiments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 515, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 287, + 647 + ], + "type": "text", + "content": "In order to avoid memorization, and yet to balance the adaptation of colors, textures, and shapes of generated objects to a new domain, we hypothesize that a more flexible attention to different levels of image realism is required by the discriminator. To this end, we perform a simple yet efficient modification to the loss function of the discriminator. Given a discriminator " + }, + { + "bbox": [ + 46, + 515, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\{D^i\\}_{i = 1}^N" + }, + { + "bbox": [ + 46, + 515, + 287, + 647 + ], + "type": "text", + "content": " and its adversarial loss function " + }, + { + "bbox": [ + 46, + 515, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D(l)" + }, + { + "bbox": [ + 46, + 515, + 287, + 647 + ], + "type": "text", + "content": " used for pre-training (e.g., cross-entropy or hinge loss), we design the discriminator to produce real/fake logits after each discriminator's block, and correspondingly compute the loss as the average across all blocks:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 651, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 651, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 58, + 651, + 287, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a l l} (x) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\mathcal {D}} \\left[ l ^ {i} (x) \\right], \\quad l ^ {i} (x) = s ^ {i} \\circ D ^ {i} (x). \\tag {3}", + "image_path": "9d8a52ef63b75d3067644f9e7118140499124dd079446655217f5dfcbf2ba1e9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "With the new objective, " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": " is given more freedom to utilize the features extracted at different scales to compute the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "loss. Our finding is that " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " dynamically adapts the magnitude of the loss at each scale to the target domain, without explicit supervision (see Fig. 6). Consequently, we observe a strong overall stabilization effect on the adaptation performance across diverse source-target dataset pairs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 146, + 388, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 146, + 388, + 159 + ], + "spans": [ + { + "bbox": [ + 306, + 146, + 388, + 159 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 167, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 167, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 545, + 299 + ], + "type": "text", + "content": "To demonstrate that our approach for few-shot GAN adaptation can be applied to unconditional and class-conditional GANs, we selected for each category a popular GAN architecture: unconditional StyleGANv2 [15] and class-conditional BigGAN [2]. For both models, we test our approach on a variety of source-target domain pairs. We focus on 10-shot target adaptation in the main paper, but we provide results for 1-shot and 5-shot adaptation in the supplementary material. For fair comparisons with prior works, most of our ablations and comparisons are conducted with StyleGANv2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 308, + 488, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 308, + 488, + 320 + ], + "spans": [ + { + "bbox": [ + 305, + 308, + 488, + 320 + ], + "type": "text", + "content": "4.1. Adaptation of unconditional GAN" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 328, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 460 + ], + "type": "text", + "content": "Datasets. In contrast to previous works that mostly considered pairs of similar datasets like Face " + }, + { + "bbox": [ + 304, + 328, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 328, + 545, + 460 + ], + "type": "text", + "content": " Sketch and Face " + }, + { + "bbox": [ + 304, + 328, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 328, + 545, + 460 + ], + "type": "text", + "content": " Sunglasses, we extend the protocol by including structurally dissimilar pairs of source and target domains, which is a more challenging task and is our primary interest. As source generators, we use StyleGANv2 checkpoints pretrained on FFHQ [14], LSUN-Church, and LSUN-Horse [38]. For the target datasets, we selected 10-shot subsets of various commonly used few-shot datasets, such as Anime-Face, Shells, or Pokemons [41, 18]. Results on more datasets are shown in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "text", + "content": "Training details. We fine-tune StyleGANv2 using the " + }, + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SS}" + }, + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{all}" + }, + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "text", + "content": " loss terms as presented in Sec. 3. For the smoothness similarity regularization, we use the intermediate features " + }, + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "inline_equation", + "content": "G^{l}" + }, + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "text", + "content": " at resolution " + }, + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "inline_equation", + "content": "(32\\times 32)" + }, + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "text", + "content": " and set " + }, + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "inline_equation", + "content": "\\lambda_{SS} = 5.0" + }, + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "text", + "content": ". We follow [23] in choosing all the other hyperparameters, such as image resolution " + }, + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "inline_equation", + "content": "(256\\times 256)" + }, + { + "bbox": [ + 304, + 460, + 545, + 556 + ], + "type": "text", + "content": ", learning rates, and batch size. Our experiments across all datasets use the same model configuration and set of hyperparameters." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 557, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 629 + ], + "type": "text", + "content": "Baselines. We compare our method to most recent few-shot GAN adaptation approaches: TGAN [36], FreezeD [20], CDC [23], RSSA [37], and AdAM [39]. In addition, we compare our proposed smoothness similarity regularizer " + }, + { + "bbox": [ + 304, + 557, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SS}" + }, + { + "bbox": [ + 304, + 557, + 545, + 629 + ], + "type": "text", + "content": " to other regularization techniques: path length regularization (PPL) [15] and MixDL [16]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": "Evaluation. In low data regimes, it is necessary to judge results both in quality and diversity aspects, as there is a trade-off between them [25, 32]. We measure the quality with FID [10] between a held-out validation set and a generated set of the same size. Following [23], we evaluate diversity with the intra-LPIPS, clustering generated images according to their nearest training samples and computing" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7076" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 94, + 137, + 273 + ], + "blocks": [ + { + "bbox": [ + 59, + 94, + 137, + 273 + ], + "lines": [ + { + "bbox": [ + 59, + 94, + 137, + 273 + ], + "spans": [ + { + "bbox": [ + 59, + 94, + 137, + 273 + ], + "type": "image", + "image_path": "0ecb19495f802ecc0a2969200d6588e807c13f045f459d9dadef380b8bc339ce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 276, + 119, + 295 + ], + "lines": [ + { + "bbox": [ + 69, + 276, + 119, + 295 + ], + "spans": [ + { + "bbox": [ + 69, + 276, + 119, + 295 + ], + "type": "text", + "content": "10-shot Anime-Faces" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 145, + 68, + 525, + 317 + ], + "blocks": [ + { + "bbox": [ + 145, + 68, + 525, + 317 + ], + "lines": [ + { + "bbox": [ + 145, + 68, + 525, + 317 + ], + "spans": [ + { + "bbox": [ + 145, + 68, + 525, + 317 + ], + "type": "image", + "image_path": "755207d03d3422ecc244f636825bc5b35919ee3a036e0f615e21e6a62b5b9732.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 527, + 68, + 539, + 566 + ], + "lines": [ + { + "bbox": [ + 527, + 68, + 539, + 566 + ], + "spans": [ + { + "bbox": [ + 527, + 68, + 539, + 566 + ], + "type": "text", + "content": "TAN Freezed CDC RSSA AdAM Ours Source TGAN Freezed CDC RSSA AdAM Ours Source" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 347, + 137, + 525 + ], + "blocks": [ + { + "bbox": [ + 59, + 347, + 137, + 525 + ], + "lines": [ + { + "bbox": [ + 59, + 347, + 137, + 525 + ], + "spans": [ + { + "bbox": [ + 59, + 347, + 137, + 525 + ], + "type": "image", + "image_path": "e381ce5a57f8934ec2c9d3cd42500c3f3b6555e7a51c20067500f40114d8ae19.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 529, + 109, + 548 + ], + "lines": [ + { + "bbox": [ + 81, + 529, + 109, + 548 + ], + "spans": [ + { + "bbox": [ + 81, + 529, + 109, + 548 + ], + "type": "text", + "content": "10-shot Shells" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 145, + 319, + 525, + 570 + ], + "blocks": [ + { + "bbox": [ + 145, + 319, + 525, + 570 + ], + "lines": [ + { + "bbox": [ + 145, + 319, + 525, + 570 + ], + "spans": [ + { + "bbox": [ + 145, + 319, + 525, + 570 + ], + "type": "image", + "image_path": "04c4721c37231e94ed46e5170d28281e7f0fb359a385802ce1dc6501d079b875.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 572, + 547, + 617 + ], + "lines": [ + { + "bbox": [ + 46, + 572, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 572, + 547, + 617 + ], + "type": "text", + "content": "Figure 3. Visual comparison to prior methods on Face " + }, + { + "bbox": [ + 46, + 572, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 572, + 547, + 617 + ], + "type": "text", + "content": " Anime and Church " + }, + { + "bbox": [ + 46, + 572, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 572, + 547, + 617 + ], + "type": "text", + "content": " Shells, the source-target dataset pairs with a dissimilar structure (e.g., shapes of objects). In this challenging regime, we observe that prior methods suffer from training instabilities, memorization issues, or inability to adapt the shapes of objects to the new domain. In contrast, our method generates images that look realistic, flexibly combine features of different target images, and transfer the variation of images from the source domain to the target domain." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 624, + 288, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 288, + 673 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 288, + 673 + ], + "type": "text", + "content": "the average LPIPS [40] of all the clusters. We train all models for 30k epochs in case of dissimilar domain pairs and for 5k on closer domains, evaluating metrics every 1k epochs. Final checkpoints in all experiments correspond to best FID." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": "Results with dissimilar source-target domains. We first present our results on the source-target domain pairs with dissimilar structure: Face " + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": " Anime, Church " + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": " Shells," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 624, + 545, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 708 + ], + "type": "text", + "content": "and Horse " + }, + { + "bbox": [ + 304, + 624, + 545, + 708 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 624, + 545, + 708 + ], + "type": "text", + "content": " Pokemon (see Fig. 3 and supplementary material). Our general observation from Fig. 3 is that in this challenging regime prior methods suffer either from training instabilities, memorization issues, or inability to adapt the shape of objects to the new domain. For example, for Face " + }, + { + "bbox": [ + 304, + 624, + 545, + 708 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 624, + 545, + 708 + ], + "type": "text", + "content": " Anime, despite an apparent correspondence between the two domains, none of the prior methods success-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7077" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 296, + 272 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 296, + 272 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 296, + 272 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 296, + 272 + ], + "type": "image", + "image_path": "5b307de5acf42f68a394289aca640efaafb14b63e815e25ccf920ca7a8a3e69a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 275, + 545, + 299 + ], + "lines": [ + { + "bbox": [ + 46, + 275, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 545, + 299 + ], + "type": "text", + "content": "Figure 4. Visual comparison to most recent prior methods on Face " + }, + { + "bbox": [ + 46, + 275, + 545, + 299 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 275, + 545, + 299 + ], + "type": "text", + "content": " Sketch and Church " + }, + { + "bbox": [ + 46, + 275, + 545, + 299 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 275, + 545, + 299 + ], + "type": "text", + "content": " Sunglasses, the dataset pairs depicting similar image domains. In this regime, our method performs on par with previous state of the art. See Table 2 for a quantitative comparison." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 305, + 72, + 545, + 272 + ], + "blocks": [ + { + "bbox": [ + 305, + 72, + 545, + 272 + ], + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 272 + ], + "type": "image", + "image_path": "b62d35529d84d6944182a608b06629941dee6a06949cd931ac22659e88d5d19c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 46, + 306, + 292, + 399 + ], + "blocks": [ + { + "bbox": [ + 46, + 306, + 292, + 399 + ], + "lines": [ + { + "bbox": [ + 46, + 306, + 292, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 292, + 399 + ], + "type": "table", + "html": "
MethodFace→AnimeChurch→ShellsHorse→Pokemons
FID↓LPIPS↑FID↓LPIPS↑FID↓LPIPS↑
TGAN [36]153.20.29205.30.22115.00.52
FreezeD [20]112.40.22180.80.27123.30.49
CDC [23]140.20.50187.90.48109.50.55
RSSA [37]133.20.37182.40.44117.30.54
AdAM [39]116.40.42152.40.28106.50.55
Ours97.30.57140.50.5384.10.61
", + "image_path": "eb65083a6f086631adfccd00a3db9254374e9c26c0f10eda34a10982c704a398.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "text", + "content": "fully transfers the distribution of head poses to the anime style, e.g., overfitting too strongly to the 10 provided samples (FreezeD), failing to adapt the shape of faces to the style of anime (CDC), or not generating high-quality animenefaces due to instabilities (TGAN, RSSA, AdAM). Similarly, for " + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\text{Church} \\rightarrow \\text{Shells}" + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "text", + "content": ", we observe that prior methods produce only copies of the example shells (FreezeD, AdAM), generate shells of unrealistic church-like shapes (CDC, RSSA), or suffer from instabilities (TGAN). In contrast, our method achieves high-quality synthesis, in which the generated images (i) look like realistic anime-faces and shells; (ii) flexibly combine features observed in different target images (e.g., anime hair color can be combined with various eye colors or background styles); and (iii) meaningfully transfer the variation of images from the source domain (e.g., generated shells adjust to the positions and shapes of churches)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "The quantitative comparison in Table 1 confirms our analysis, where our method achieves the best quality and diversity scores across all datasets. We note a high average relative improvement of more than " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " in FID and LPIPS compared to the highest scores achieved by prior methods. Overall, we conclude that our method significantly improves over prior works on few-shot GAN" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 334, + 306, + 515, + 399 + ], + "blocks": [ + { + "bbox": [ + 47, + 404, + 287, + 427 + ], + "lines": [ + { + "bbox": [ + 47, + 404, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 47, + 404, + 287, + 427 + ], + "type": "text", + "content": "Table 1. Comparison of the adaptation performance in case of dissimilar source-target domains. Bold denotes best performance." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 306, + 515, + 399 + ], + "lines": [ + { + "bbox": [ + 334, + 306, + 515, + 399 + ], + "spans": [ + { + "bbox": [ + 334, + 306, + 515, + 399 + ], + "type": "table", + "html": "
MethodFace→SketchFace→Sunglasses
FID↓LPIPS↑FID↓LPIPS↑
TGAN [36]54.20.3836.80.56
FreezeD [20]48.80.3232.00.59
CDC [23]54.20.4030.50.59
RSSA [37]61.40.4536.30.58
AdAM [39]56.30.3731.10.60
Ours45.20.4427.50.60
", + "image_path": "c3b7d2a77f832660f7911d73017159830eb5015a2b0b9692870d2296e039d5a6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 404, + 544, + 427 + ], + "lines": [ + { + "bbox": [ + 306, + 404, + 544, + 427 + ], + "spans": [ + { + "bbox": [ + 306, + 404, + 544, + 427 + ], + "type": "text", + "content": "Table 2. Comparison in case of structurally close source-target domains. Bold denotes best performance." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 438, + 522, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 438, + 522, + 450 + ], + "spans": [ + { + "bbox": [ + 306, + 438, + 522, + 450 + ], + "type": "text", + "content": "adaptation with dissimilar source and target domains." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 451, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 545, + 654 + ], + "type": "text", + "content": "Results with close source-target domains. Next, we follow the evaluation of prior works and compare the models on similar source and target domains, such as adaptation of human faces to a different style. The visual results for Face " + }, + { + "bbox": [ + 304, + 451, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 451, + 545, + 654 + ], + "type": "text", + "content": " Sketch and Face " + }, + { + "bbox": [ + 304, + 451, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 451, + 545, + 654 + ], + "type": "text", + "content": " Sunglasses are shown in Fig. 4. Our method successfully performs the few-shot adaptation in this setting, adapting the colors and textures of faces to the gray-scale sketch domain, or adding a novel attribute (sunglasses). We note that our method is not explicitly designed to transfer all the details of a face from the source domain, thus changes in the generated images like facial hair are expected. Yet, we observe that our method generally does not lose distinctive features of faces in source images, performing on par with previous state-of-the-art methods. The quantitative comparison is provided in Table " + }, + { + "bbox": [ + 304, + 451, + 545, + 654 + ], + "type": "inline_equation", + "content": "2^{1}" + }, + { + "bbox": [ + 304, + 451, + 545, + 654 + ], + "type": "text", + "content": ": on both datasets our method achieves the best FID scores and performs on par with the best performer in LPIPS." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 655, + 545, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 655, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 304, + 655, + 545, + 692 + ], + "type": "text", + "content": "Ablations. We demonstrate the importance of our proposed loss terms in Fig. 5, which shows latent space interpolations of trained models and their similarity to the pre" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 42, + 74, + 53, + 265 + ], + "type": "aside_text", + "angle": 270, + "lines": [ + { + "bbox": [ + 42, + 74, + 53, + 265 + ], + "spans": [ + { + "bbox": [ + 42, + 74, + 53, + 265 + ], + "type": "text", + "content": "Ours AdAM RSSA CDC Source Real" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 317, + 702, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 545, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 317, + 702, + 545, + 713 + ], + "type": "text", + "content": "FID evaluation differs from prior works (discussed in suppl. material)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7078" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 285, + 326 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 285, + 326 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 285, + 326 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 285, + 326 + ], + "type": "image", + "image_path": "4755dec36406e5b280bc981bfe1ca884e4f1926cf0de83e7373ff4f0e24b556f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 328, + 287, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 361 + ], + "type": "text", + "content": "Figure 5. Latent space interpolations of the source generator and the ablation models from Tables 3-4. Leftmost and rightmost columns show the used " + }, + { + "bbox": [ + 46, + 328, + 287, + 361 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 328, + 287, + 361 + ], + "type": "text", + "content": " loss and " + }, + { + "bbox": [ + 46, + 328, + 287, + 361 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 328, + 287, + 361 + ], + "type": "text", + "content": " smoothness regularization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 365, + 274, + 460 + ], + "blocks": [ + { + "bbox": [ + 58, + 365, + 274, + 460 + ], + "lines": [ + { + "bbox": [ + 58, + 365, + 274, + 460 + ], + "spans": [ + { + "bbox": [ + 58, + 365, + 274, + 460 + ], + "type": "table", + "html": "
D lossSmooth reg. for GFace→AnimeChurch→Shells
FID↓LPIPS↑FID↓LPIPS↑
StyleGANv2-178.00.21243.80.17
StyleGANv2SS (ours)180.70.61252.80.62
PatchGAN [23]-145.20.37183.10.31
PatchGAN [23]SS (ours)132.20.55184.20.56
Lall (ours)-116.40.36175.40.43
Lall (ours)SS (ours)97.30.57140.50.53
", + "image_path": "59945f8af071a35a73994ada6762141a9fd9fecc7277d0fbe4aabbff3496a4a9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "text", + "content": "trained source model " + }, + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "inline_equation", + "content": "G_{s}" + }, + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "text", + "content": " (row 1). Firstly, we note that the plain StyleGANv2 model (row 2) suffers from instabilities in our low data regime, achieving poor image quality and diversity and having \"staircase\"-like latent space interpolations. Applying " + }, + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SS}" + }, + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "text", + "content": " without " + }, + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{all}" + }, + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "text", + "content": " (row 3) helps to achieve diverse synthesis with smooth interpolations, but it is not enough to achieve good image quality. On the other hand, using " + }, + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{all}" + }, + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "text", + "content": " (row 4) helps to overcome instabilities and improve image quality, but it cannot maintain smooth interpolations and high diversity. Finally, our full model (row 5) allows a higher-quality, diverse synthesis with smooth and realistic latent space interpolations. Note how the image transitions mimic the behaviour of the source model (churches and shells change shapes and positions similarly), allowing to achieve diverse and realistic synthesis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "The effect of " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{all}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " is further demonstrated in Fig. 6, where we show the contribution of different " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " blocks to the adversarial loss at different epochs. We note the ability of the discriminator to identify correct loss contributions" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 305, + 69, + 545, + 151 + ], + "blocks": [ + { + "bbox": [ + 305, + 69, + 545, + 151 + ], + "lines": [ + { + "bbox": [ + 305, + 69, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 305, + 69, + 545, + 151 + ], + "type": "image", + "image_path": "598b03f082b3faefeeaa68bdfddecd7ce535e606436fbd31756c97770494dad4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 154, + 546, + 209 + ], + "lines": [ + { + "bbox": [ + 305, + 154, + 546, + 209 + ], + "spans": [ + { + "bbox": [ + 305, + 154, + 546, + 209 + ], + "type": "text", + "content": "Figure 6. The contribution of features at different " + }, + { + "bbox": [ + 305, + 154, + 546, + 209 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 154, + 546, + 209 + ], + "type": "text", + "content": " blocks to the adversarial loss function " + }, + { + "bbox": [ + 305, + 154, + 546, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{all}" + }, + { + "bbox": [ + 305, + 154, + 546, + 209 + ], + "type": "text", + "content": ". For two closer image domains (the left plot), the network concentrates mostly on earlier layers to compute the loss, while for less similar domains the network learns to use the later layers representing more high-level " + }, + { + "bbox": [ + 305, + 154, + 546, + 209 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 154, + 546, + 209 + ], + "type": "text", + "content": " features." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 318, + 216, + 529, + 287 + ], + "blocks": [ + { + "bbox": [ + 47, + 460, + 286, + 471 + ], + "lines": [ + { + "bbox": [ + 47, + 460, + 286, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 286, + 471 + ], + "type": "text", + "content": "Table 3. Impact of " + }, + { + "bbox": [ + 47, + 460, + 286, + 471 + ], + "type": "inline_equation", + "content": "{\\mathcal{L}}_{\\text{all }}" + }, + { + "bbox": [ + 47, + 460, + 286, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 460, + 286, + 471 + ], + "type": "inline_equation", + "content": "{\\mathcal{L}}_{SS}" + }, + { + "bbox": [ + 47, + 460, + 286, + 471 + ], + "type": "text", + "content": " . Bold denotes best performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 216, + 529, + 287 + ], + "lines": [ + { + "bbox": [ + 318, + 216, + 529, + 287 + ], + "spans": [ + { + "bbox": [ + 318, + 216, + 529, + 287 + ], + "type": "table", + "html": "
D lossSmooth reg. for GFace→AnimeChurch→Shells
FID↓LPIPS↑FID↓LPIPS↑
Lall(ours)-116.40.36175.40.43
Lall(ours)PPL [14]107.80.46179.40.44
Lall(ours)MixDL [16]105.90.50150.40.51
Lall(ours)SS (ours)97.30.57140.50.53
", + "image_path": "a11703e5a96ab9c17b011bd9f47f246a68a6d1cc24de1e1a4b70dff39b3ca586.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 288, + 545, + 309 + ], + "lines": [ + { + "bbox": [ + 306, + 288, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 306, + 288, + 545, + 309 + ], + "type": "text", + "content": "Table 4. Comparison of smoothness similarity regularization " + }, + { + "bbox": [ + 306, + 288, + 545, + 309 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SS}" + }, + { + "bbox": [ + 306, + 288, + 545, + 309 + ], + "type": "text", + "content": " with other regularizers. Bold denotes best performance." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 327, + 546, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 546, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 546, + 458 + ], + "type": "text", + "content": "adaptively for different source-target domain pairs. For example for Face " + }, + { + "bbox": [ + 304, + 327, + 546, + 458 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 327, + 546, + 458 + ], + "type": "text", + "content": " Anime, the network concentrates mostly on the earliest " + }, + { + "bbox": [ + 304, + 327, + 546, + 458 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 327, + 546, + 458 + ], + "type": "text", + "content": " blocks to adapt the colors and textures of faces to a new style. In contrast, for the more distant domains Church " + }, + { + "bbox": [ + 304, + 327, + 546, + 458 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 327, + 546, + 458 + ], + "type": "text", + "content": " Shells, the network learns to attribute a higher weight to the later blocks to also adapt higher-level features, such as shapes of objects. In effect, we observe a stabilized adaptation of colors, textures, and shapes of objects across diverse source-target pairs. Using PatchGAN [23] as discriminator loss does not achieve such a balance as it focuses mostly on lower-scale features (row 6 in Fig. 5)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "text", + "content": "Our observations are confirmed by the quantitative study in Table 3: without " + }, + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SS}" + }, + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "text", + "content": " the model does not achieve high diversity (high LPIPS), while " + }, + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{all}" + }, + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "text", + "content": " is necessary for high image quality (low FID). We conclude that both our proposed loss terms are important to achieve high-quality synthesis. More ablations on " + }, + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SS}" + }, + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{all}" + }, + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "text", + "content": " can be found in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": "Lastly, Table 4 provides a comparison of our proposed " + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SS}" + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": " loss term to other regularizers: path length regularization (PPL) [14] and MixDL [16]. While all regularizers help to achieve smoother latent space interpolations and thus improve the quality and diversity metrics, our smoothness similarity regularization enables the highest performance in both FID and LPIPS. While our approach transfers the learned smoothness of the source generator to the target domain, PPL and MixDL resort to gradually interpolating between the provided training samples, which leads to latent space interpolations that either look unrealistic or lack diversity (rows 7-8 in Fig. 5). This demonstrates that transferring smoothness from a pre-trained generator is beneficial to enforce image transitions that are realistic and diverse." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7079" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 83, + 294, + 109 + ], + "blocks": [ + { + "bbox": [ + 47, + 83, + 294, + 109 + ], + "lines": [ + { + "bbox": [ + 47, + 83, + 294, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 83, + 294, + 109 + ], + "type": "image", + "image_path": "89f2496245519eeb260bee46876711f937542fa6af06bf6e628c2eac6006e71b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 297, + 83, + 544, + 108 + ], + "blocks": [ + { + "bbox": [ + 297, + 83, + 544, + 108 + ], + "lines": [ + { + "bbox": [ + 297, + 83, + 544, + 108 + ], + "spans": [ + { + "bbox": [ + 297, + 83, + 544, + 108 + ], + "type": "image", + "image_path": "b1024f67326930cdb76f583a23ccb5cb80713e8a6af5592c5537c890752315b1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 45, + 123, + 550, + 304 + ], + "blocks": [ + { + "bbox": [ + 102, + 111, + 504, + 123 + ], + "lines": [ + { + "bbox": [ + 102, + 111, + 504, + 123 + ], + "spans": [ + { + "bbox": [ + 102, + 111, + 504, + 123 + ], + "type": "text", + "content": "10-Shot adaptation results of our method on class-conditional BigGAN [2], pre-trained on ImageNet" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 123, + 550, + 304 + ], + "lines": [ + { + "bbox": [ + 45, + 123, + 550, + 304 + ], + "spans": [ + { + "bbox": [ + 45, + 123, + 550, + 304 + ], + "type": "image", + "image_path": "7ec0596780f910ed1bc441fbc7016076181f7b5bdb666831af14b45941a3cc73.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 307, + 547, + 340 + ], + "lines": [ + { + "bbox": [ + 46, + 307, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 547, + 340 + ], + "type": "text", + "content": "Figure 7. 10-shot adaptation results for the class-conditional BigGAN [2] pre-trained on ImageNet. While simple fine-tuning (FT) suffers from training instabilities and mode collapse, our method helps to achieve much higher image quality and diversity, transferring smooth and realistic image transitions from the source domain, e.g., objects smoothly changing their locations, size, and shape." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 47, + 345, + 291, + 415 + ], + "blocks": [ + { + "bbox": [ + 47, + 345, + 291, + 415 + ], + "lines": [ + { + "bbox": [ + 47, + 345, + 291, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 345, + 291, + 415 + ], + "type": "table", + "html": "
D lossSmooth reg. for GImageNet→FlowersImageNet→Pokemons
FID↓LPIPS↑FID↓LPIPS↑
BigGAN-213.30.29226.80.15
BigGANSS (ours)225.60.47208.30.47
Lall (ours)-123.90.28129.40.27
Lall (ours)SS (ours)106.40.5589.60.56
", + "image_path": "7804531f3a34bc00220a25f7776482a41f8eb282bcd7943d27404069c4818f8f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 420, + 287, + 443 + ], + "lines": [ + { + "bbox": [ + 46, + 420, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 287, + 443 + ], + "type": "text", + "content": "Table 5. Ablation on the performance when adapting the class-conditional BigGAN model [2] pre-trained on ImageNet." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 454, + 243, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 243, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 243, + 467 + ], + "type": "text", + "content": "4.2. Adaptation of class-conditional GAN" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 473, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 700 + ], + "type": "text", + "content": "Our approach is not limited to unconditional GANs, but it can also be applied to a class-conditional GAN model. We selected BigGAN [2] for our experiments as it is a popular backbone architecture for class-conditional image synthesis on ImageNet [7]. We make two modifications to enable the adaptation of the model to unconditional few-shot datasets. Firstly, we remove the conditioning of the discriminator via the projection layer [19]. Secondly, we treat the generator's learned continuous class embedding as part of the latent space, thus sampling a Gaussian vector in the joint noise-class space at each fine-tuning epoch. This way, the generator produces an image based on a single input vector in an unconditional fashion. We then fine-tune the pre-trained model using our loss terms " + }, + { + "bbox": [ + 46, + 473, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SS}" + }, + { + "bbox": [ + 46, + 473, + 287, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 473, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{all}" + }, + { + "bbox": [ + 46, + 473, + 287, + 700 + ], + "type": "text", + "content": " as presented in Sec. 3. We use image resolution " + }, + { + "bbox": [ + 46, + 473, + 287, + 700 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 46, + 473, + 287, + 700 + ], + "type": "text", + "content": " and batch size of 32. The hyperparameters for " + }, + { + "bbox": [ + 46, + 473, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SS}" + }, + { + "bbox": [ + 46, + 473, + 287, + 700 + ], + "type": "text", + "content": " are the same as for StyleGANv2: intermediate features " + }, + { + "bbox": [ + 46, + 473, + 287, + 700 + ], + "type": "inline_equation", + "content": "G^{l}" + }, + { + "bbox": [ + 46, + 473, + 287, + 700 + ], + "type": "text", + "content": " at resolution " + }, + { + "bbox": [ + 46, + 473, + 287, + 700 + ], + "type": "inline_equation", + "content": "(32\\times 32)" + }, + { + "bbox": [ + 46, + 473, + 287, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 473, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\lambda_{SS} = 5.0" + }, + { + "bbox": [ + 46, + 473, + 287, + 700 + ], + "type": "text", + "content": ". We train for 30k epochs and select checkpoints by best FID." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": "Datasets. As the source generator, we use the Big-" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 346, + 546, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 546, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 546, + 407 + ], + "type": "text", + "content": "GAN checkpoint pre-trained on class-conditional ImageNet at resolution " + }, + { + "bbox": [ + 305, + 346, + 546, + 407 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 305, + 346, + 546, + 407 + ], + "type": "text", + "content": ". We demonstrate 10-shot adaptation results with two commonly used few-shot generation datasets: Oxford-Flowers [21] and Pokémon [18]. We use the same model configuration for both datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 411, + 546, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 546, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 546, + 590 + ], + "type": "text", + "content": "Results. Fig. 7 demonstrates latent space interpolations of the source and target generators. We note that a simple fine-tuning of BigGAN suffers from training instabilities and mode collapse. In contrast, our method successfully adapts BigGAN to generate diverse images in the target domains. We highlight that our method transfers smooth and realistic image transitions from the well-learned BigGAN's noise-class space, despite significant dissimilarities between ImageNet and the few-shot datasets, in particular Pokemons. For example, it can be noticed how the latent space interpolations in the target domains mimic the source domain, e.g., the generated flowers andPokemons change their position and size similarly to dogs and wolves " + }, + { + "bbox": [ + 304, + 411, + 546, + 590 + ], + "type": "inline_equation", + "content": "(5^{\\mathrm{th}}-" + }, + { + "bbox": [ + 304, + 411, + 546, + 590 + ], + "type": "inline_equation", + "content": "10^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 411, + 546, + 590 + ], + "type": "text", + "content": " columns in Fig. 7) or stretch their shape to mimic the proportions of busses " + }, + { + "bbox": [ + 304, + 411, + 546, + 590 + ], + "type": "inline_equation", + "content": "(11^{\\mathrm{th}} - 14^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 411, + 546, + 590 + ], + "type": "text", + "content": " columns)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "Table 5 shows the importance of our proposed loss terms. Our observations are consistent with the ablations with StylGANv2: " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{all}" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " is necessary to avoid instabilities and achieve a good image quality (low FID), while " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SS}" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " is required to achieve smooth latent space interpolations and good diversity (high LPIPS). We conclude that our method successfully extends to the adaptation of class-conditional models, where target domains benefit from the rich noise-class space learned on a multi-class dataset such as ImageNet. More details and results are provided in the supplementary material." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 108, + 71, + 240, + 82 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 71, + 240, + 82 + ], + "spans": [ + { + "bbox": [ + 108, + 71, + 240, + 82 + ], + "type": "text", + "content": "Real samples - 10-Shot Flowers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 353, + 71, + 493, + 82 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 71, + 493, + 82 + ], + "spans": [ + { + "bbox": [ + 353, + 71, + 493, + 82 + ], + "type": "text", + "content": "Real samples - 10-Shot Pokemons" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7080" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 119, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 119, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 119, + 83 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 200 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 200 + ], + "type": "text", + "content": "In this work, we presented a new method for few-shot adaptation of GAN models. It transfers the smooth latent space of a pre-trained GAN, which was trained on a large dataset, to a new domain with very few images. We addressed the case of few-shot GAN adaptation when the source and target domains are structurally dissimilar, which is a common issue in applications. Our extensive results demonstrate that in this setting our approach outperforms previous works in terms of image quality and diversity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 209, + 145, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 209, + 145, + 221 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 145, + 221 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 228, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 300 + ], + "type": "text", + "content": "The work has been supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC 2070 -390732324 and the ERC Consolidator Grant FORHUE (101044724). We thank Jinhui Yi for providing and analyzing the sugar beet data used in supplementary experiments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 308, + 107, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 107, + 321 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 107, + 321 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 329, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 329, + 287, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 287, + 383 + ], + "type": "text", + "content": "[1] Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-shot segmentation without meta-learning: A good transductive inference is all you need? In Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 384, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 384, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 53, + 384, + 288, + 426 + ], + "type": "text", + "content": "[2] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In International Conference on Learning Representations (ICLR), 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 428, + 287, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 428, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 53, + 428, + 287, + 493 + ], + "type": "text", + "content": "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 494, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 494, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 53, + 494, + 287, + 536 + ], + "type": "text", + "content": "[4] Yunjoy Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 537, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 537, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 53, + 537, + 287, + 581 + ], + "type": "text", + "content": "[5] Kaiwen Cui, Jiaxing Huang, Zhipeng Luo, Gongjie Zhang, Fangneng Zhan, and Shijian Lu. Genco: Generative constraining for generative adversarial networks with limited data. In Conference on Artificial Intelligence (AAAI), 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 582, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 582, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 582, + 287, + 624 + ], + "type": "text", + "content": "[6] Yann Dauphin, Harm De Vries, and Yoshua Bengio. Equilibrated adaptive learning rates for non-convex optimization. In Advances in Neural Information Processing Systems (NeurIPS), 2015." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 625, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 625, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 53, + 625, + 287, + 669 + ], + "type": "text", + "content": "[7] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In Conference on Computer Vision and Pattern Recognition (CVPR), 2009." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "type": "text", + "content": "[8] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "text", + "content": "[9] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In Advances in Neural Information Processing Systems (NeurIPS), 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems (NeurIPS), 2017." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 172, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 215 + ], + "type": "text", + "content": "[11] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 217, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 259 + ], + "type": "text", + "content": "[12] Tero Karras, Miika Aittala, Janne Hellsten, S. Laine, J. Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Advances in Neural Information Processing Systems (NeurIPS), 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 304 + ], + "type": "text", + "content": "[13] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Advances in Neural Information Processing Systems (NeurIPS), 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 305, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 545, + 348 + ], + "type": "text", + "content": "[14] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "text", + "content": "[15] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 393, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 436 + ], + "type": "text", + "content": "[16] Chaerin Kong, Jeesoo Kim, Donghoon Han, and Nojun Kwak. Few-shot image generation with mixup-based distance learning. In European Conference on Computer Vision (ECCV), 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "type": "text", + "content": "[17] Yijun Li, Richard Zhang, Jingwan Cynthia Lu, and Eli Shechtman. Few-shot image generation with elastic weight consolidation. In Advances in Neural Information Processing Systems (NeurIPS), 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "type": "text", + "content": "[18] Bingchen Liu, Yizhe Zhu, Kunpeng Song, and Ahmed Elgammal. Towards faster and stabilized gan training for high-fidelity few-shot image synthesis. In International Conference on Learning Representations (ICLR), 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 526, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 545, + 557 + ], + "type": "text", + "content": "[19] Takeru Miyato and Masanori Koyama. cGANs with projection discriminator. In International Conference on Learning Representations (ICLR), 2018." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 559, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 592 + ], + "type": "text", + "content": "[20] Sangwoo Mo, Minsu Cho, and Jinwoo Shin. Freeze discriminator: A simple baseline for fine-tuning gans. In CVPR AI for Content Creation Workshop, 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 593, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 545, + 624 + ], + "type": "text", + "content": "[21] Maria-Elena Nilsback and Andrew Zisserman. A visual vocabulary for flower classification. In Conference on Computer Vision and Pattern Recognition (CVPR), 2006." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 625, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 658 + ], + "type": "text", + "content": "[22] Atsuhiro Noguchi and T. Harada. Image generation from small datasets via batch statistics adaptation. In International Conference on Computer Vision (ICCV), 2019." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[23] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7081" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[24] Jie Ren, Peter J Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark Depristo, Joshua Dillon, and Balaji Lakshminarayanan. Likelihood ratios for out-of-distribution detection. Advances in Neural Information Processing Systems (NeurIPS), 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "type": "text", + "content": "[25] Esther Robb, Wen-Sheng Chu, Abhishek Kumar, and Jia-Bin Huang. Few-shot adaptation of generative adversarial networks. arXiv:2010.11943, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 165, + 288, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 165, + 288, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 288, + 208 + ], + "type": "text", + "content": "[26] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis. arXiv preprint arXiv:2301.09515, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 210, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 210, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 287, + 243 + ], + "type": "text", + "content": "[27] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleganxI: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 245, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 245, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 48, + 245, + 287, + 289 + ], + "type": "text", + "content": "[28] Edgar Schonfeld, Vadim Sushko, Dan Zhang, Juergen Gall, Bernt Schiele, and Anna Khoreva. You only need adversarial supervision for semantic image synthesis. In International Conference on Learning Representations (ICLR), 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 291, + 287, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 291, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 48, + 291, + 287, + 324 + ], + "type": "text", + "content": "[29] Tamar Rott Shaham, Tali Dekel, and T. Michaeli. SinGAN: Learning a generative model from a single natural image. In International Conference on Computer Vision (ICCV), 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 326, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 326, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 326, + 287, + 358 + ], + "type": "text", + "content": "[30] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 287, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 405 + ], + "type": "text", + "content": "[31] Vadim Sushko, Juergen Gall, and Anna Khoreva. One-Shot GAN: Learning to generate samples from single images and videos. In Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 407, + 287, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 407, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 48, + 407, + 287, + 439 + ], + "type": "text", + "content": "[32] Vadim Sushko, Dan Zhang, Juergen Gall, and Anna Khoreva. Learning to generate novel scene compositions from single images and videos. arXiv:2103.13389, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 441, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 441, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 441, + 287, + 485 + ], + "type": "text", + "content": "[33] Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. Transactions on Pattern Analysis and Machine Intelligence, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 487, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 520 + ], + "type": "text", + "content": "[34] Andrey Voynov and Artem Babenko. Unsupervised discovery of interpretable directions in the gan latent space. In International conference on machine learning (ICML), 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 521, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 576 + ], + "type": "text", + "content": "[35] Yaxing Wang, Abel Gonzalez-Garcia, David Berga, L. Herranz, F. Khan, and Joost van de Weijer. Minegan: Effective knowledge transfer from gans to target domains with few images. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "type": "text", + "content": "[36] Yaxing Wang, Chenshen Wu, L. Herranz, Joost van de Weijer, Abel Gonzalez-Garcia, and B. Raducanu. Transferring gans: generating images from limited data. In European Conference on Computer Vision (ECCV), 2018." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "text", + "content": "[37] Jiayu Xiao, Liang Li, Chaofei Wang, Zheng-Jun Zha, and Qingming Huang. Few shot generative model adaption via relaxed spatial structural alignment. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[38] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv:1506.03365, 2015." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 251 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 117 + ], + "type": "text", + "content": "[39] Zhao Yunqing, Keshigeyan Chandrasegaran, Milad Abdol-lahzadeh, and Ngai-man Cheung. Few-shot image generation via adaptation-aware kernel modulation. In Advances in Neural Information Processing Systems (NeurIPS), 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 546, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 162 + ], + "type": "text", + "content": "[40] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Conference on Computer Vision and Pattern Recognition (CVPR), 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 546, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 546, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 546, + 206 + ], + "type": "text", + "content": "[41] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. In Conference on Neural Information Processing Systems (NeurIPS), 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 208, + 546, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 546, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 546, + 251 + ], + "type": "text", + "content": "[42] Yunqing Zhao, Henghui Ding, Houjing Huang, and Ngai-Man Cheung. A closer look at few-shot image generation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7082" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Snow Removal in Video_ A New Dataset and A Novel Method/5ae50b63-d59f-4dc4-bf7f-4d6ac34210b4_content_list.json b/2023/Snow Removal in Video_ A New Dataset and A Novel Method/5ae50b63-d59f-4dc4-bf7f-4d6ac34210b4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8ed556fe0dab5f0e50bb9d623f30fa1d097039bc --- /dev/null +++ b/2023/Snow Removal in Video_ A New Dataset and A Novel Method/5ae50b63-d59f-4dc4-bf7f-4d6ac34210b4_content_list.json @@ -0,0 +1,2084 @@ +[ + { + "type": "text", + "text": "Snow Removal in Video: A New Dataset and A Novel Method", + "text_level": 1, + "bbox": [ + 174, + 130, + 794, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haoyu Chen $^{1}$ , Jingjing Ren $^{1}$ , Jinjin Gu $^{2}$ , Hongtao Wu $^{1}$ , Xuequan Lu $^{3}$ , Haoming Cai $^{4}$ , Lei Zhu $^{1,5}$", + "bbox": [ + 241, + 179, + 727, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Hong Kong University of Science and Technology (Guangzhou) 2The University of Sydney", + "bbox": [ + 191, + 217, + 777, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ La Trobe University $^{4}$ The University of Maryland", + "bbox": [ + 323, + 234, + 645, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5The Hong Kong University of Science and Technology", + "bbox": [ + 318, + 253, + 651, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project page: https://haoyuchen.com/VideoDesnowing", + "bbox": [ + 256, + 272, + 710, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Snowfall is a common weather phenomenon that can severely affect computer vision tasks by obscuring objects and scenes. However, existing deep learning-based snow removal methods are designed for single images only. In this paper, we target a more complex task - video snow removal, which aims to restore the clear video from the snowy video. To facilitate this task, we propose the first high-quality video dataset, which simulates realistic physical characteristics of snow and haze using a rendering engine and augmentation techniques. We also develop a deep learning framework for video snow removal. Specifically, we propose a snow-query temporal aggregation module and a snow-aware contrastive learning loss function. The module aggregates features between video frames and removes snow effectively, while the loss function helps identify and eliminate snow features. We conduct extensive experiments and demonstrate that our proposed dataset is more realistic than previous datasets, and the models trained on it achieve better performance in real-world snowing images. Our proposed method outperforms state-of-the-art video and image-based methods on both synthetic and real snowy videos.", + "bbox": [ + 76, + 354, + 472, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 720, + 209, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Snowfall is a common weather phenomenon that can significantly affect visibility in photographs, interfering with and reducing the accuracy of computer vision tasks such as target detection [34], tracking [14], and autonomous driving [33]. Image and video restoration [39, 23, 48, 37, 46, 7, 5, 26, 40, 43, 42, 41, 45, 31] is a widely studied direction, but research on image and video snow removal is still limited. Removing snow is crucial for improving accuracy and robustness of computer vision tasks. Additionally, snow can obstruct objects, decreasing image and video quality. Snow", + "bbox": [ + 75, + 750, + 472, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "removal is a challenging task that requires distinguishing snowflake regions from the background and restoring obscured scenes, complicated by the complex geometric shape and texture of snowflakes. Moreover, snowfall often comes with haze, which further obscures the images.", + "bbox": [ + 496, + 323, + 893, + 398 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contrast with single image, the video provides richer information about the dynamic features of the scene, i.e., a consecutive frame sequence on the temporal dimension. Although recent image snow removal algorithms [25, 11, 12, 50, 8, 10] have achieved good results, applying these algorithms directly to video snow removal is inappropriate because image algorithms ignore temporal information in video frames and this information could be very useful to improve snow removal outcomes. As such, it is significant to conduct in-depth research and exploration for snow video data. Previous studies have explored some aspects of video desnowing, such as methods based on non-deep learning [32, 18] or online learning [22]. However, deep learning-based video desnowing remains an under-researched area. A major challenge for applying deep learning methods is the lack of high-quality video desnowing datasets. In the era of powerful deep learning, this becomes an urgent problem that needs to be addressed. In fact, the lack of appropriate datasets has significantly slowed down the progress of deep learning-based video snow removal research, which has been a bottleneck in this field.", + "bbox": [ + 496, + 400, + 895, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To this end, we present the first high-quality annotated video dataset for video snow removal: Realistic Video DeSnowing Dataset (RVSD). RVSD contains a total of 110 pairs of videos. Each pair contains snowy and hazy videos and corresponding snow-free and haze-free ground truth videos. Unlike previous image datasets that simulate snow using Photoshop, we use a rendering engine and various augmentation techniques to generate snow with diverse and realistic physical properties. This results in more realistic and varied synthesized videos, which improve the model's performance on real-world data. To further enhance the model's performance, we also consider the effect of haze", + "bbox": [ + 496, + 719, + 895, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13211", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/32dd73bf420f912e63db375ae0a0858c395f4433b54006127e7ca1c151e8b6a3.jpg", + "image_caption": [ + "a.City Day" + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 179, + 186 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3de314e6c425295fe6ccad00951d1985e084e36bcc8768b4d1bd2bd0624200e7.jpg", + "image_caption": [ + "b. City Night" + ], + "image_footnote": [], + "bbox": [ + 181, + 90, + 277, + 186 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1fa1488917077fdbdbe1616848314e000e355dbd8777a9abe7a88c92257b702d.jpg", + "image_caption": [ + "c. Nature", + "Figure 1: Samples of the proposed dataset (left) and previous datasets (right). Refer to the supplementary to see the videos." + ], + "image_footnote": [], + "bbox": [ + 279, + 90, + 354, + 186 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f73943a08fd6883b84647a4324019f9359578b359b1cc79d6caf50dc09b3fffd.jpg", + "image_caption": [ + "Snow100K" + ], + "image_footnote": [], + "bbox": [ + 395, + 90, + 493, + 186 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/24ac02178277a49b6e6b62340d6cb379dd2a48329ca06bd1fe8ccc25ff642b1b.jpg", + "image_caption": [ + "SnowCityScape" + ], + "image_footnote": [], + "bbox": [ + 495, + 90, + 591, + 186 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/046b13f627357807135358a66d498d33867906575fd9be8dd46cf19167597a18.jpg", + "image_caption": [ + "SnowCityScape" + ], + "image_footnote": [], + "bbox": [ + 594, + 90, + 691, + 186 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fb8ddd6cb5f9548b7b829f9bd117c025f9db7303d7fa749ffd8ae4f7319d8bdc.jpg", + "image_caption": [ + "SRRS" + ], + "image_footnote": [], + "bbox": [ + 694, + 90, + 790, + 186 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ddeb05d42fb0086e2a5e834f424fbbb94fdd1d106d74b0c01a1dd8141357c18b.jpg", + "image_caption": [ + "CSD" + ], + "image_footnote": [], + "bbox": [ + 792, + 90, + 888, + 186 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in snowy conditions. Previous methods for fog synthesis relied on mathematical models with limitations, such as requiring accurate depth information and assuming fixed and uniform haze characteristics. However, real-world snowy conditions involve dynamic and heterogeneous haze. To overcome these limitations, we propose a new haze model that uses Unreal Engine 5 to render storm haze and obtain a haze layer. We then fuse the haze layer with the input video after data augmentation to produce more diverse and realistic haze, better simulating real snowy conditions. We conduct experiments to show that our proposed dataset has more-realistic snow scenes than existing datasets, and models trained on our dataset achieve better snow removal performance on real snowy images. Therefore, our dataset can facilitate the training and evaluation of video snow removal algorithms, and advance video snow removal research.", + "bbox": [ + 75, + 253, + 468, + 493 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Trivially applying the video processing methods on our dataset may result in suboptimal since they do not consider snow and haze properties. Moreover, there are no existing deep learning-based methods for video snow removal. Therefore, we propose the first deep learning-based framework in this work. Our framework consists of three main steps. First, we perform preliminary snow removal on single frames to reduce the effect of snow occlusion on optical flow estimation and video feature alignment. Second, we fuse the video frames and perceive and remove the snow. We introduce a novel \"snow-query temporal aggregation module\", which aggregates features across video frames while perceiving and removing snow. Third, we design a new \"snow-aware contrastive loss\", which leverages the prior knowledge that snowflakes have different shapes, motion directions, and other characteristics in different videos. This loss function helps identify and remove snow features more accurately. We conduct extensive experiments to show that our proposed method outperforms the state-of-the-art methods. Our contributions are summarized as follows:", + "bbox": [ + 75, + 494, + 468, + 795 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the first high-quality video dataset (RVSD) for video snow removal. We use the Unreal Engine and various augmentation techniques to produce snow and haze with diverse and realistic physical characteristics.", + "- Comprehensive experiments show that the visual effect" + ], + "bbox": [ + 94, + 801, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of our proposed snow removal dataset is more realistic than previous datasets, and the models trained on our proposed dataset can achieve better performance in real-world snowy images.", + "bbox": [ + 529, + 253, + 890, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- A deep learning-based video snow removal framework is proposed. The snow-aware temporal aggregation module and snow-aware contrastive learning loss are introduced, to effectively remove snow. Experiments prove that our method enables the best performance.", + "bbox": [ + 517, + 324, + 890, + 401 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 425, + 640, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Desnowing Datasets", + "text_level": 1, + "bbox": [ + 500, + 450, + 689, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are several single-image snow removal datasets available. Snow100K [25] first used PhotoShop to synthesize snow. In order to model different types of snow, SnowKITTI2012 [50] and SnowCityScapes [50] contains three levels of snow, including light, medium, and heavy snow. However, these datasets do not take into account the veiling effect of snowy weather, the models trained on these datasets cannot remove haze and have limited practical performance in real snowy scenarios. SRRS [11] firstly simulated the veiling effect inspired by the Koschmieder model, and then used PhotoShop to synthesize the snow streaks. CSD [12] added Gaussian blur to the above two steps to better simulate the real snow image. Nearly all methods use PhotoShop to synthesize snow. As a result, the shape and style of snow are relatively homogeneous, which substantially limits the model's performance on realistic and complex snowing scenes. A quick comparison of these datasets can be found in Table 1. To the best of our knowledge, there is no dataset for video desnowing, which significantly hinders the research of video desnowing.", + "bbox": [ + 496, + 473, + 890, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Image Desnowing", + "text_level": 1, + "bbox": [ + 500, + 786, + 671, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early image snow removal methods tend to perform snow removal based on physical priors. [1] performed snow detection and removal by calculating the histogram of orientations of snow streaks. [29] used features on saturation and visibility to remove rain and snow from a single image. In recent years, deep learning has refreshed the snow removal", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13212", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/e643d0e5be3c375b275f841562b0959e3b0427755c8f4db25889f6da86e7306a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetVenueType# imagesResolutionSnow synthesisStatic HazeDynamic HazeIlluminationUser Study (1 - 10)
Snow100K [25]TIP 2018image100K≤ 640 × 640PhotoshopXXX3.98
SnowCityScapes[50]TIP 2021image2K+2K512 × 256PhotoshopXXX4.18
SnowKITTI2012 [50]TIP 2021image1.5K+1K884 × 256PhotoshopXXX3.72
SRRS [11]ECCV 2020image15K640 × 480Photoshop✓XX4.20
CSD [12]ICCV 2021image10K640 × 480Photoshop✓XX4.67
Ours-video-480p - 4KUE Rendering✓✓✓6.94
", + "bbox": [ + 78, + 89, + 643, + 178 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/fcb9466f644126ffda0b6690b972a04a50625694f4582b9d37232f89ea5a305a.jpg", + "table_caption": [ + "Table 1: Comparison with previous desnowing datasets." + ], + "table_footnote": [], + "table_body": "
Scenes# videos# frames# frames (512×512)
city day405,12131,020
city night202,37823,878
nature433,52632,526
other74073,012
In total11011,43290,436
", + "bbox": [ + 656, + 89, + 890, + 178 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2: Statistics of our dataset.", + "bbox": [ + 661, + 180, + 880, + 194 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "task. The first deep learning based snow removal method DesnowNet [25] proposed a multi-stage, multi-scale design to remove translucent and opaque snow streaks. JSTASR [11] proposed a joint size and transparency-aware snow removal model that can classify snow particles based on their size and remove snow. HDCW-Net [12] used a hierarchical decomposition paradigm in which a dual-tree wavelet transform and a wavelet loss are used. It also proposed a discriminative feature for snow removal called contradict channel. DDMSNet [50] introduced semantic and depth information to learn semantic-aware and geometry-aware representations for snow removal. SnowFormer [9] used a vision transformer architecture that fully combines local and global information, and obtains superior results on multiple datasets.", + "bbox": [ + 75, + 222, + 472, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Video Desnowing", + "text_level": 1, + "bbox": [ + 76, + 459, + 246, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Before the success of deep learning-based methods, traditional computer vision techniques were used for video-based snow removal. Ren et al. [32] utilized the low-rank assumption of the background to separate sparse and dense snow, and addressed heavy snow in dynamic scenes. Kim et al. [18] considered global and local motion, as well as snowflakes of various sizes, in their snowflake removal algorithm with low-rank matrix completion. Due to the rapid improvement of deep learning models and their ability to study complex patterns, some researchers have proposed the use of deep learning for video-based snow removal. [44] utilized self-adaptive snow detection and a patch-based Gaussian mixture model, adept at removing both sparse and dense snowflakes from videos. Li et al. [21, 22] proposed a method for dynamic backgrounds where snow was encoded and removed using an online multi-scale convolutional sparse coding model.", + "bbox": [ + 75, + 482, + 470, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Realistic Video Desnowing Dataset", + "text_level": 1, + "bbox": [ + 76, + 753, + 393, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous snow and haze synthesis methods suffer from a lack of variation and poor physical accuracy. In contrast, our approach employs a 3D rendering engine and a range of augmentations, together with a more realistic fusion method to create snow and haze with more diverse, complex, and accurate physical properties. As a result, the synthesized video is more realistic and diverse, enabling our model to enjoy better generalization ability to real-world data. We", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a2523a3842090bd29425f3304a23cf2c779414b3d56dd64644f4a7d147092538.jpg", + "image_caption": [ + "Figure 2: Box plots of the user study on the visual realism perception scores of different datasets. Higher scores represent better perceptual realism." + ], + "image_footnote": [], + "bbox": [ + 500, + 222, + 723, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "demonstrate this in Section 5.1.", + "bbox": [ + 500, + 366, + 710, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Collection of Clean Videos", + "text_level": 1, + "bbox": [ + 500, + 390, + 738, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike previous work that only considered daytime video, we collect both daytime and nighttime videos in our dataset, which is unlike previous works. We apply different add-snow methods for daytime or nighttime videos to enhance the realism of the generated results. We will discuss these methods in the next section. Our dataset covers a variety of scenes such as buildings, streets, nature scenes, close shots, distant views, overhead and elevation angles to simulate the rich scenes of real snow videos. As shown in Table 2, we have collected 110 videos in total, among which 80 videos are used for training and 30 videos for testing. To accommodate the diverse resolutions of online videos, our video resolutions range from $480\\mathrm{p}$ to $4\\mathrm{k}$ . This enables the model to produce better results on videos with more sources and resolutions.", + "bbox": [ + 496, + 414, + 893, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Haze Rendering", + "text_level": 1, + "bbox": [ + 500, + 650, + 661, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous studies have employed a fog synthesis method based on the atmospheric scattering model [20]: $I_{haze}(x) = J(x)t(x) + A(1 - t(x))$ , where $I_{haze}(x)$ is hazy image, $J(x)$ is the haze-free image, $A$ denotes the global atmospheric light, and $t(x)$ is the transmission matrix. This model has both strengths and weaknesses. One of the strengths is that the formula is intuitive and concise, making it easy to embed into network design. For limitations, it relies on depth information, and the generated haze exhibits a relatively uniform pattern. More importantly, there is a fundamental difference between snow-generated haze on a snowy day and conventional haze. Snow haze is often dynamic and follows the movement of snow, as illustrated in Figure 4. Therefore, using traditional haze synthesis methods that assume a homogeneous haze pattern may", + "bbox": [ + 496, + 674, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13213", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1f22993164006276ad80b34809152f86f08a836fbb8fadbce77595690e2a085b.jpg", + "image_caption": [ + "Figure 3: Snow-free videos with rich scenes are selected first. Then, we use two haze models to generate depth-related static haze and Unreal Engine (UE) rendered dynamic haze that is more realistic. After that, we use the UE to render snow layers with realistic physical characteristics. Further the snow layers are dynamically blended according to the background time after complex augmentation. These complex operations make the results more realistic with diverse features." + ], + "image_footnote": [], + "bbox": [ + 122, + 88, + 844, + 255 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/09004c2c7292247fd322e48992d8db2ba9791ada351a7a483efeb05d021d86bd.jpg", + "image_caption": [ + "Figure 4: Dynamic haze in a snowstorm." + ], + "image_footnote": [], + "bbox": [ + 78, + 354, + 468, + 406 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5d26b40a8818d9ba3c7ce85df13c1d03803c0a1fab35b2f527dc23ff81535c56.jpg", + "image_caption": [ + "Figure 5: Fusion weight for daytime and nighttime." + ], + "image_footnote": [], + "bbox": [ + 78, + 444, + 450, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "not be accurate in simulating the dynamic and varying snow haze that occurs during a snowstorm. We propose a novel haze model that leverages the capabilities of the Unreal Engine to render the haze of snowstorms and obtain different haze layers $H$ . The dynamic haze of the storms is then integrated with the input video following data augmentation. As a result, our approach produces a more diverse and realistic haze that more closely resembles real snow days and nights. This enables the models trained on our data to handle more diverse and dynamic snow haze. Our composite haze model is defined as follows:", + "bbox": [ + 75, + 564, + 468, + 728 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {h a z e} (x) = J (x) t (x) + A (1 - t (x)) + \\operatorname {A u g} (H), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 729, + 468, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where Aug denotes data augmentation.", + "bbox": [ + 76, + 748, + 336, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Snow Rendering", + "text_level": 1, + "bbox": [ + 76, + 771, + 240, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In previous snow production work [11, 25, 50, 12], snowflakes and snow streaks were synthesized using Photoshop, resulting in relatively homogeneous patterns. In contrast, our work employs Unreal Engine to synthesize snow. As a 3D computer graphics engine, Unreal Engine enables the creation of snow with more realistic and complex physical features, including snowflake flip, spatial changes from", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0bef36974df88c23b1126bf9cfeecb19c4d988fe00ff8557a5580b026f82fd95.jpg", + "image_caption": [ + "Figure 6: The visibility of daytime snow and nighttime snow varies depending on the background." + ], + "image_footnote": [], + "bbox": [ + 501, + 353, + 893, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "far to near, and perspective changes, among others. We use various 3D snow models to manipulate the physical properties of the snow, resulting in diverse and complex shapes that more closely resemble real-life snow. More details of the features are shown in Table 3. Once the snow layer $S$ is generated, it needs to be blended with the clean video. We augment the snow layer by adjusting the color temperature, brightness, contrast, and applying random affine transformations. Then we add different levels of blur, such as Gaussian blur and motion blur. Finally, we consider the temporal properties of the video background based on time of day. During the day, especially when the sky is in the background, snowflakes tend to blend with the bright sky, resulting in snowflakes that are not easily distinguishable. At night, snowflakes are more visible in bright places, such as those illuminated by headlights or streetlights, while they tend to be less visible in places without light. We show some real samples in Figure 6. Based on this property, we utilize Eq. (2) to perform the integration of snow layer and video background.", + "bbox": [ + 496, + 561, + 890, + 863 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha (x) = \\left\\{ \\begin{array}{c} \\sigma (- (V (x) - \\gamma) \\times \\beta), \\text {i f d a y t i m e} \\\\ \\sigma ((V (x) - \\gamma) \\times \\beta), \\text {i f n i g h t t i m e} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 864, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13214", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/182db053e1f19820c50fd29cda071aa1d1c2a70d7236e73791892e7a9bf9b54f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FeatureDetails
Snow Speedby setting gravity intensity, wind speed, air friction coefficient and other parameters, snow can move at different speeds
Snow Intensityby setting the number of particles in the snow model, our dataset contains snow of various degrees.
Snow Typeby employing various snow models and adjusting various parameters, combined with the Unreal Engine to impart realistic physical features to the snowflakes (such as turbulence, rotation, etc.), the resulting snow has a highly diverse range of shapes, sizes, and characteristics.
Motion directionhead-on, back-on, left-facing, right-facing, and many other angles
Illuminationlights have different intensities and colors, such as different warm and cold tones
Veiling effectdepth-related static haze, as well as rendered dynamic complex haze.
Scenescity, street, day & night, nature
", + "bbox": [ + 78, + 89, + 468, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3: Features of our proposed dataset.", + "bbox": [ + 133, + 252, + 410, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we first convert the RGB image to the HSV color space, where $V = \\max(R, G, B)$ , represents the largest color component. We obtain a corresponding blending weight based on the pixel's brightness. $v$ denotes the value of the V channel in the HSV space, which is normalized to the range of 0 to 1. $\\gamma$ and $\\beta$ are two parameters artificially adjusted based on the specific video, and $\\sigma$ represents the softmax function. Figure 5 shows some examples. Our dynamic video snow model is defined as follows:", + "bbox": [ + 75, + 292, + 468, + 428 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nZ (x) = I _ {h a z e} (x) + \\alpha (x) \\operatorname {A u g} (S (x)). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 430, + 468, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where Aug denotes data augmentation. $Z$ is the final output with both snow and haze.", + "bbox": [ + 75, + 446, + 468, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Snow-Aware Video Desnowing Network", + "text_level": 1, + "bbox": [ + 75, + 489, + 431, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Network Architecture", + "text_level": 1, + "bbox": [ + 76, + 513, + 282, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our network architecture is composed of three main modules: the encoder, the snow-query temporal aggregation module, and the decoder, as shown in Figure 7. To mitigate the impact of snow occlusion on video frame alignment and fusion, the encoder takes single-frame images as input and generates coarse features $F^{C}$ , which are then transformed into coarse desnowed images by the recovery module. However, residual snow may still remain in the video frames due to the lack of utilization of video information. Therefore, it is imperative to incorporate multi-frame video information to further remove residual snow. To achieve this, the Snow-query Temporal Aggregation Module takes in continuous video frames and the features obtained by the encoder. This module aligns and fuses video features, and can perceive the location of snow, which facilitates more effective snow removal. Then this module outputs continuous refined video features $F^{F}$ , which are fed into the decoder. Finally, the recovery module is employed to obtain the final desnowed video.", + "bbox": [ + 75, + 537, + 468, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recovery Module. To simultaneously remove snow and snow-induced haze, we adopt the atmospheric scattering model to decompose images containing both snow and haze into three parts: snow feature $S$ , the global atmospheric light $A$ , and transmission matrix $T$ . We perform", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "this operation in the feature space of the image. Here, $S \\in \\mathbb{R}^{N \\times C \\times H \\times W}$ , while $A \\in \\mathbb{R}^{N \\times 1 \\times H \\times W}$ and $T \\in \\mathbb{R}^{N \\times C \\times H \\times W}$ . $N$ is the number of frames, $C$ is the number of feature channel, $H$ and $W$ are the height and width. The recovery process can be formulated as:", + "bbox": [ + 498, + 90, + 890, + 165 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF (x) = \\frac {E (I (x)) - S (x) - (1 - T (x)) A (x)}{T (x)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 166, + 890, + 199 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nJ (x) = M (F (x)),\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 200, + 671, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I(x)$ is the input snowy image, $F(x)$ is the output desnowed feature, $E(\\cdot)$ is the feature extractor, $M(\\cdot)$ is a convolution layer that maps the feature into image space, and $J(x)$ is the snow-free image.", + "bbox": [ + 496, + 218, + 890, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Coarse Desnowing of Single Video Frames Prior to Alignment. Video restoration networks frequently employ optical flow or deformable convolution to integrate information from multiple frames for feature extraction. However, this is not entirely suitable for video snow removal owing to the significant masking of the video background by snowflakes and snow lines, which disturb temporal feature aggregation. The position discrepancies between snowflakes and snow lines in adjacent frames, especially in heavily snowing videos, pose a challenge to alignment methods in estimating accurate motion parameters. Thus, coarse desnowing of single frames before temporal feature attending is crucial for effectively integrating information between frames in video snow removal later on. This step does not cause any information loss for the further step. The reason is that the desnowing inputs of the further step are the features of original snowy video frames, instead of the desnowed images of the first step.", + "bbox": [ + 496, + 279, + 890, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 9 shows the results of optical flow computation on video frames with and without snow. Optical flow computation on snow-covered video frames shows significant errors, which may mislead the alignment of video frame features. Our idea is to obtain coarse desnowed frames before computing optical flows. Specifically, the encoder takes a single snowy frame as input and outputs coarse features $S^C$ , $A^C$ , and $T^C$ . The recovery module then restores the coarse desnowed frame feature $F^C$ and its corresponding desnowed frame $x^C$ .", + "bbox": [ + 496, + 551, + 890, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Snow-Query Temporal Aggregation", + "text_level": 1, + "bbox": [ + 500, + 710, + 808, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal Aggregation Module. To fully utilize the features between video frames, we use bidirectional propagation to independently propagate the frames' features forward and backward in time dimension. This technique enables the information to flow in both directions, which can help capture more spatial-temporal context and improve the model's accuracy. Given coarse desnowed frame $x_{i}^{C}$ and its neighboring frame $x_{i-1}^{C}$ and $x_{i+1}^{C}$ , we use SpyNet [30] to estimate the optical flows $O_{t+1 \\to t}$ and $O_{t-1 \\to t}$ . To obtain the corresponding features propagated from the neighboring frames, which are denoted as $h_{t-1}^{f}$ and $h_{t+1}^{b}$ , we have", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13215", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6a3b07189f814bc2aeddae9f4f5fcb1a5a2b7267d3a3e121ca06d6d4f6d9e0ad.jpg", + "image_caption": [ + "Figure 7: Snow-Aware Video Desnowing Network (SVDNet). The architecture is composed of three modules: the encoder, the snow-query temporal aggregation module, and the decoder. To avoid the effect of snow occlusion on video alignment, the encoder first uses a single frame to generate coarse feature $F^C$ and desnowed image $x^C$ . Then the Snow-query Temporal Aggregation Module aligns and fuses video features, detects snow, and outputs continuous fine video features $F^F$ . These features are then fed into the decoder, and the recovery module is used to obtain the final desnowed video." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 890, + 275 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/05e77ff5934c7740a3aac6cf61c8aac096cdf12166ca21dfe460efb5c3c72a24.jpg", + "image_caption": [ + "Figure 8: Snow-Query Temporal Aggregation Module." + ], + "image_footnote": [], + "bbox": [ + 78, + 388, + 472, + 616 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nh _ {t} ^ {b} = H _ {b} \\left(F _ {t} ^ {C}, O _ {t + 1 \\rightarrow t}, h _ {t + 1} ^ {b}, T _ {t} ^ {C}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 675, + 392, + 694 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nh _ {t} ^ {f} = H _ {f} \\left(F _ {t} ^ {C}, O _ {t - 1 \\rightarrow t}, h _ {t - 1} ^ {f}, T _ {t} ^ {C}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 693, + 468, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $H_{b}$ and $H_{f}$ denote the backward and forward propagation branches.", + "bbox": [ + 76, + 726, + 468, + 755 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nw _ {t} ^ {\\{b, f \\}} = W \\left(h _ {t \\pm 1}, O _ {t \\pm 1 \\rightarrow t}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 756, + 333, + 776 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nh _ {t} ^ {\\{b, f \\}} = \\operatorname {M S A} \\left(\\operatorname {c a t} \\left(w _ {t} ^ {\\{b, f \\}}, F _ {t} ^ {C}, T _ {t} ^ {C}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 777, + 468, + 805 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nh _ {t} = \\operatorname {c a t} \\left(h _ {t} ^ {f}, h _ {t} ^ {b}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 806, + 281, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $W$ denotes feature warping, and MSA denotes multi-head self-attention.", + "bbox": [ + 76, + 825, + 468, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Snow-Query Cross-Attention Module. The snow features for the same video are always similar. Thus, we can further use the existing snow features to match and guide the", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8593678d02f16642073ff64ee6d7f42dde3dd45f8b3393435830af8e92d01d11.jpg", + "image_caption": [ + "Figure 9: Incorrect estimation of optical flow due to snow." + ], + "image_footnote": [], + "bbox": [ + 503, + 388, + 893, + 483 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fb7bef89128f082c4e886bf104a14bd026a7c75354e6631a2f4194bdf3635194.jpg", + "image_caption": [ + "Figure 10: t-SNE visualization of features learned with/without the contrastive loss. Our contrastive loss function effectively distinguishes different types of snow while without it, their feature distributions highly overlap." + ], + "image_footnote": [], + "bbox": [ + 501, + 527, + 890, + 661 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "network to find the remaining snow. We show this module in Figure 8. First, we perform patch embedding for $h_t$ and corresponding snow features $S_t^C$ . We consider each patch of size $2 \\times 2 \\times C$ as a token. Then, a linear embedding layer is applied to project the features of each token to an arbitrary dimension. The key design is using snow features as queries. The snow features can help find the residual snow regions in the video feature. Specifically, the queries obtained from the snow feature and the key obtained from the", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13216", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bf277ede75ba1f761a9a6662bb5540a85939b4183a51af12f8679712d9609633.jpg", + "image_caption": [ + "Figure 11: Visual results on the real-world images. The same model is trained separately on different datasets for image snow removal. The model trained on our proposed dataset achieves the best results for snow and haze removal on real images." + ], + "image_footnote": [], + "bbox": [ + 80, + 87, + 893, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "video feature are multiplied to obtain a correlation matrix. The coordinates with high correlation represent the residual snow in the video, which enables the network to locate the residual snow and further remove it.", + "bbox": [ + 75, + 314, + 468, + 375 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C r o s s A t t n} \\left(\\mathcal {Q} _ {s}, \\mathcal {K} _ {v}, \\mathcal {V} _ {v}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {Q} _ {s} \\mathcal {K} _ {v}}{\\sqrt {d}} + p\\right) \\mathcal {V} _ {v}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 373, + 468, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $Q_{s}, K_{v}, V_{v}$ are the snow query, video key, and value matrices. $d$ is the dimension. $p$ is the relative position embedding.", + "bbox": [ + 75, + 417, + 468, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Snow-Aware Contrastive Loss", + "text_level": 1, + "bbox": [ + 76, + 470, + 344, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We introduce a prior for snow-covered images: the snow features vary across videos based on the snow's angle and properties (e.g., density, shape, direction, and speed), but are consistent within the same video, as shown in Figure 7. We design a snow-aware contrastive learning paradigm. It uses snow features from different frames of the same video as positive samples $S^{+}$ and snow features from different videos as negative samples $S^{-}$ . By pulling positive samples closer together and pushing negative samples farther apart in feature space, the network can more accurately identify snow features and improve snow removal. The loss function for contrastive learning is formulated as:", + "bbox": [ + 75, + 494, + 468, + 674 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o n t}} \\left(S, S ^ {+}, S ^ {-}\\right) = \\\\ - \\log \\left[ \\frac {\\exp \\left(D \\left(S ^ {+} , S ^ {-}\\right) / \\tau\\right)}{\\exp \\left(D \\left(S ^ {+} , S ^ {-}\\right) / \\tau\\right) + \\sum_ {r = 1} ^ {N} \\exp \\left(D \\left(S ^ {+} , S ^ {-}\\right) / \\tau\\right)} \\right] \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 676, + 478, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $D(x,y) = \\Psi (x)\\cdot \\Psi (y)$ . $\\Psi (\\cdot)$ is the two-layer MLP function which implements feature projection. $\\tau$ is the scale temperature which is set to 0.1. $N$ denotes the total number of negative samples.", + "bbox": [ + 75, + 750, + 468, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To illustrate the effect of the proposed snow-aware contrastive learning loss on snow feature learning, we visualize the learned features using t-SNE for three distinct types of snow with different motion directions and snow stripe sizes, as shown in Figure 10. Without the proposed contrastive learning, different snow types have highly similar and over", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "lapping feature distributions. In contrast, with the contrastive learning loss, the features of different snow types are clearly distinguished, demonstrating the ability of the method to accurately identify diverse snow features. The overall loss function is formulated as: $\\mathcal{L} = \\mathcal{L}_1 + \\lambda \\mathcal{L}_{cont}$ where $\\lambda$ is set to 0.1.", + "bbox": [ + 496, + 314, + 893, + 405 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 498, + 425, + 633, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Proposed Dataset vs. Previous Datasets", + "text_level": 1, + "bbox": [ + 498, + 452, + 836, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Objective Evaluation. Previous work [4, 51] has demonstrated the significant impact of training set data on the generalization performance of the model. To demonstrate that the snow scenes in our created dataset are more realistic and closer to real-world scenarios than those in other synthetic datasets, we train the same model on different datasets and test it on real snowy images. The results show that the model trained on our proposed dataset performs better in handling real snow, thus validating the authenticity and usefulness of our dataset. Figure 11 displays two samples. In the first row, the model trained on our dataset successfully removes all snow residues, whereas models trained on other datasets fail to completely eliminate snow, leaving numerous snow residues in the output images. Since the Snow100K, SnowCityScapes, and SnowKITTI2012 datasets do not include the effect of snow haze, the models trained on these datasets cannot remove snow haze. In the second row, the models trained on the Snow100K, SnowCityScapes, and SnowKITTI2012 datasets remove slightly larger snowflakes but preserve many smaller ones. Furthermore, these models mistakenly remove white parallelogram backgrounds, mistaking them for snowflakes. Conversely, SRRS and CSD remove smaller snowflakes but fail to remove larger ones. In contrast, our proposed dataset facilitates the training of a model that effectively removes snowflakes at all sizes. It does not misidentify the white background as snow. Since there is no ground truth for real-world images, we conduct a user study to compare the", + "bbox": [ + 496, + 477, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13217", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6c436a64193c6d1b2f86bd1fb918bebcb81dc21583cf833d3ce556a0841c6353.jpg", + "image_caption": [ + "Figure 12: Distributions of mean opinion scores obtained by models trained on different datasets (testing on real-world snow images)." + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 472, + 188 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "desnowing results of models which are trained on different datasets, as shown in Figure 12. Model trained on our dataset has the best desnowing performance on real-world images.", + "bbox": [ + 75, + 271, + 468, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Subjective Evaluation. To demonstrate that our dataset produces snow images that are closer to real snow in human visual perception than previous datasets, we conduct a user study to compare the perceived realism of snow generated from different datasets. The results are shown in Figure 2, dataset with higher scores indicating greater fidelity. Thanks to our proposed complex, specially designed dataset construction method, our dataset is scored much higher than the other datasets.", + "bbox": [ + 75, + 333, + 470, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Performance Evaluation", + "text_level": 1, + "bbox": [ + 76, + 481, + 302, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation Details. Each input image is randomly cropped to a spatial resolution of $256 \\times 256$ , and the number of the total training iterations is $500\\mathrm{K}$ . During the training phase, we adopt the Adam optimizer [19] with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.99$ . The initial learning rate is set to $2 \\times 10^{-4}$ and reduced by half at the milestone of $100\\mathrm{K}$ iterations and $150\\mathrm{K}$ iterations. The batch size is set to 4. PyTorch [28] is used to implement our model with 4 RTX 3090 GPUs.", + "bbox": [ + 75, + 505, + 468, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compared Methods. We compare our network against 18 state-of-the-art methods, including three single-image desnowing models, two image adverse weather restoration models, six image restoration methods, two video deraining methods, and five video restoration methods. Please refer to Table 5 for more details. Moreover, we follow [16, 15] to employ PSNR, SSIM [38], and LPIPS [52] to quantitatively compared different methods.", + "bbox": [ + 75, + 627, + 468, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative Comparison. Table 5 reports the quantitative results of our network and 17 compared methods. From these quantitative results, we can find that the performance of video-based models is generally inferior to that of single-image models. The major reason behind is that the severe occlusion of the video background by snowflakes hinders the alignment and fusion of the video frames. Among all compared methods, Restormer has the largest PSNR score of 24.34 and the smallest 0.1164, while MPRNet has the largest SSIM score of 0.8960. More importantly, our net", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/106c2bed74bdf4eddc5c6d964c65b50faecf3f0586a60f65a0845d70f4755d55.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRestormerTransWeatherS2VDRVRTBasicVSR++Ours
E_warp ↓(×10-3)1.5942.3922.0093.4252.6131.203
VFID ↓0.09380.15110.16860.22070.15390.0492
", + "bbox": [ + 503, + 89, + 890, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1fccd8f1e1dfb2049e5ed4d1bcf72f80f0e27caa69def37f9cb9f80ef3dc0449.jpg", + "table_caption": [ + "Table 4: Results of temporal consistency preservation." + ], + "table_footnote": [], + "table_body": "
TypeMethodVenuePSNR ↑SSIM ↑LPIPS ↓
Image DesnowingJSTASR [11]ECCV 202022.080.82800.2336
HDCW-Net [12]ICCV 202122.630.85920.2010
SnowFormer [8]arXiv 202224.010.89390.1219
Image Adverse WeatherTransWeather [35]CVPR 202223.110.85430.2086
TKL [13]CVPR 202223.050.85890.2027
Image RestorationSwinIR [23]CVPR 202122.510.85620.2105
MPRNet [49]CVPR 202124.270.89600.1266
Uformer [39]CVPR 202223.610.87300.1706
DGUNet [27]CVPR 202224.180.89850.1238
NAFNet [6]ECCV 202224.010.88380.1472
Restormer [48]CVPR 202224.340.89290.1164
Video DerainingS2VD [47]CVPR 202122.950.85900.1856
RDD [36]ECCV 202222.970.87420.1631
Video RestorationEDVR [37]CVPRW 201917.930.57900.3771
BasicVSR [2]CVPR 202122.460.84730.2087
IconVSR [2]CVPR 202122.350.84820.2034
BasicVSR++ [3]CVPR 202222.640.86180.1868
RVRT [24]NeurlPS 202220.900.79740.2977
SVDNet (Ours)-25.060.92100.0842
", + "bbox": [ + 503, + 183, + 888, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5: Quantitative results on the proposed video dataset.", + "bbox": [ + 500, + 431, + 890, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "work further outperform these methods in terms of PSNR, SSIM, and LPIPS scores. It improves the PSNR score from 24.34 to 25.06, the SSIM score from 0.8960 to 0.9210, and the LPIPS score from 0.1164 to 0.0842. It demonstrates the effectiveness of our proposed framework, which first obtains a coarse snow removal result for a single frame image to reduce the challenging alignment issue on the snowy video frames, and then fully utilizes the temporal information between video frames for video desnowing.", + "bbox": [ + 496, + 474, + 890, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further verify the temporal consistency preservation, we utilize a widely-used metric $E_{\\text{warp}}$ to compute the temporal consistency, and VFID [17] to compute video restoration performance. Table below shows our method has the smallest $E_{\\text{warp}}$ and VFID. Our method can achieve consistent desnowing between frames.", + "bbox": [ + 496, + 612, + 890, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual comparisons on synthetic videos. Figure 13 visually compares video desnowing results produced by our network and state-of-the-art methods. We can find that existing methods tend to maintain snowflakes (e.g., TKL, TransWeather), or haze (e.g., Restomer), or both snowflakes and haze (i.e., RVRT) in their desnowing results. On the contrary, our method can effectively remove both snowflakes and haze, and better preserve background details.", + "bbox": [ + 496, + 703, + 890, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual comparisons on real-world videos. Moreover, Figure 14 shows the visual comparisons between our network and state-of-the-art methods in terms of input frames from real-world snowstorm videos with both heavy snow and haze. We can find that our method can effectively remove", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13218", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f58e5f0032cc5cf935036c23af03def374ca884cc2c68e723915cb60a82f441.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 88, + 196, + 160 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8b65bab7ea4fad30b21806a96664127f68502af3edcaff070fdcecdcb9031eb3.jpg", + "image_caption": [ + "TKL [13]" + ], + "image_footnote": [], + "bbox": [ + 197, + 88, + 312, + 161 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3d5614415600dedbda16c5cb2cf2541ef58b393190ae2c7225d20466f01fbbbe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 88, + 426, + 161 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/775bcbe1f5b8f820706f060f7c6d68616ebcf4dbce1719f2298720ccd41e24ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 428, + 88, + 542, + 161 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b1c82358c4a06d9b227c8920f56888bd4f5eb533437d980e42b8e484cc4d8c53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 88, + 656, + 161 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a2f8fef4027292335e8696c318056a28c7ae95f626ffbb44e530487488c18976.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 88, + 772, + 161 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9eaa5d5b0500203c25687d5440a1d7094e22ab431e11c7eb72e8c47b0c71d37b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 88, + 887, + 161 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2f3c69faaca36ba166e933a4a243ce97635567aaf5aa9f76a9996fe0a904d020.jpg", + "image_caption": [ + "Input Frame", + "S2VD [47]", + "Figure 13: Visual video desnowing results produced by our network and state-of-the-art methods." + ], + "image_footnote": [], + "bbox": [ + 83, + 172, + 196, + 244 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/53caa05dec53d0f360ef223673273f20133036ed9f8b458ebe831bd834e2d628.jpg", + "image_caption": [ + "RDD [36]" + ], + "image_footnote": [], + "bbox": [ + 197, + 172, + 312, + 244 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/cc46f46f0318eed6ab50472f49abf23ef0bd600dd3dde40db0c02d5096228685.jpg", + "image_caption": [ + "TransWeather [35]", + "RVRT [24]" + ], + "image_footnote": [], + "bbox": [ + 313, + 172, + 426, + 244 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/576b1a935a986d2f5064f000e1dafafa14bbe0ec3ca617a6d854914165967238.jpg", + "image_caption": [ + "Snowformer [8]", + "IconVSR [2]" + ], + "image_footnote": [], + "bbox": [ + 428, + 172, + 542, + 244 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/db52a41ac806d5411c6727e9356709068505380725e0a0bfd8092b26d7392a11.jpg", + "image_caption": [ + "Restormer [48]", + "BasicVSR++ [3]" + ], + "image_footnote": [], + "bbox": [ + 542, + 172, + 656, + 244 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5a696d3cdf283181e8b6ae8fc7532ae53e7c342b4526780ecb15acb543f773bc.jpg", + "image_caption": [ + "MPRNet [49]", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 658, + 172, + 772, + 244 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0cc95c97bb83078a52a57b7418638907170f19d8c430b36e5abdd2ae06a8fd59.jpg", + "image_caption": [ + "DGUNet [27]", + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 774, + 172, + 887, + 244 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5067a693dac14c7b2450b704644c91ff247c9f65d8ce6554187dd9f28439a511.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 308, + 176, + 381 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4e5339b483474826997b7b09cad4c3eb22f02facd2149b83864343d00fbbdf0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 308, + 272, + 381 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b862fdd9a66d2b4bd5eb0444ee47bafe6349100b8d946f910439832c8032ef20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 308, + 369, + 381 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/804247cd4a7c4ee81a34be0a037b797e31b031cb95353b316400721b790a2261.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 308, + 465, + 381 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/510966f35b2a527de801f85a3357ccd0ee22480832308085001da67c47cbd34f.jpg", + "image_caption": [ + "Input Frame", + "S2VD [47]", + "Figure 14: Visual comparison on real snowstorm video." + ], + "image_footnote": [], + "bbox": [ + 83, + 393, + 176, + 465 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c74014ae602bab47ed84249d62f13b8ef6030d2722e1c604963e0cdc1cea9aac.jpg", + "image_caption": [ + "MPRNet [49]", + "IconVSR [2]" + ], + "image_footnote": [], + "bbox": [ + 179, + 393, + 272, + 465 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4fcc296e79418007c6e2a86979ccebf4e00e97306100ca7f7b699c168e0bc953.jpg", + "image_caption": [ + "DGUNet [27]", + "BasicVSR++[3]" + ], + "image_footnote": [], + "bbox": [ + 274, + 393, + 369, + 465 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6b358618061b3712cbd970e1b0d0a4b090cc3748307db32020005f5c45fe6b0d.jpg", + "image_caption": [ + "RDD [36]", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 370, + 393, + 465, + 465 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/c8aaa7efef0cc15ad7487d1a9c262d4730cea6d876e582fd101859d78ab1db34.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IndexTemporal AggregationSnow-QueryContrastive LossPSNR ↑SSIM ↑LPIPS ↓
M1(basic)24.160.89900.1125
M2✓24.580.91300.0934
M3✓✓24.800.91080.9004
M4✓✓24.840.91810.0882
Ours✓✓✓25.060.92100.0842
", + "bbox": [ + 81, + 520, + 467, + 606 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 6: Quantitative results of the ablation study.", + "bbox": [ + 107, + 628, + 437, + 645 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "both the haze and snowflakes and betterrecover the obscured background details, while compared methods tend to main parts of haze and snowflakes.", + "bbox": [ + 75, + 672, + 468, + 718 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ablation Study. We perform ablation study experiments to verify the temporal aggregation module, the snow-query cross-attention module, and the snow-aware contrastive loss of our network. To do so, we first built a basic model (see M1 of Table 6) by removing these three major components from our network, and then add the temporal aggregation module into M1 to construct M2. Then we add the snow-query cross-attention module into M2 to build M3, and add the snow-aware contrastive loss into M2 to build M4. Table 6 reports the PSNR, SSIM, and LPIPS results of our network and four baseline networks. Specifically, compared to M1, M2 improves the PSNR score from 24.16", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "to 24.58, the SSIM score from 0.8990 to 0.9130, and the LPIPS score from 0.1125 to 0.0934. It shows that the temporal aggregation module incurs a better video desnowing performance. Then, we find that M3 and M4 have a superior PSNR, SSIM, and LPIPS performance over M2. It demonstrates the effectiveness of the snow-aware cross-attention module or snow-aware contrastive loss for enhancing the video desnowing performance. Moreover, our method has larger PSNR and SSIM scores and a smaller LPIPS score than M3 and M4. It shows that combing three components together has the best performance of video desnowing.", + "bbox": [ + 496, + 310, + 893, + 477 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 488, + 617, + 503 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The first contribution of our work is to synthesize the first high-quality video desnowing dataset, which is more realistic than previous image datasets. Then, we devise a deep learning-based framework that incorporates a snow-query temporal aggregation module and a snow-aware contrastive learning loss. Experimental results show that our network outperforms state-of-the-art methods in terms of video denosing on synthetic and real-world snowy videos.", + "bbox": [ + 496, + 513, + 893, + 635 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgment. This work was supported by the National Natural Science Foundation of China (Grant No. 61902275), and Guangzhou Municipal Science and Technology Project (Grant No. 2023A03J0671).", + "bbox": [ + 496, + 637, + 893, + 698 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 709, + 596, + 724 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jérémie Bossu, Nicolas Hautiere, and Jean-Philippe Tarel. Rain or snow detection in image sequences through use of a histogram of orientation of streaks. International journal of computer vision, 93:348–367, 2011.", + "[2] Kelvin C.K. Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2021.", + "[3] Kelvin C.K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. BasicVSR++: Improving video superresolution with enhanced propagation and alignment. In" + ], + "bbox": [ + 506, + 733, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13219", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Conference on Computer Vision and Pattern Recognition, 2022.", + "[4] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1692-1703, June 2023.", + "[5] Haoyu Chen, Jinjin Gu, and Zhi Zhang. Attention in attention network for image super-resolution, 2021.", + "[6] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 17-33. Springer, 2022.", + "[7] Sixiang Chen, Tian Ye, Yun Liu, and Erkang Chen. Dualformer: Hybrid self-attention transformer for efficient image restoration. arXiv preprint arXiv:2210.01069, 2022.", + "[8] Sixiang Chen, Tian Ye, Yun Liu, Erkang Chen, Jun Shi, and Jingchun Zhou. Snowformer: Scale-aware transformer via context interaction for single image desnowing. arXiv preprint arXiv:2208.09703, 2022.", + "[9] Sixiang Chen, Tian Ye, Yun Liu, Erkang Chen, Jun Shi, and Jingchun Zhou. Snowformer: Scale-aware transformer via context interaction for single image desnowing. arXiv preprint arXiv:2208.09703, 2022.", + "[10] Sixiang Chen, Tian Ye, Yun Liu, Taodong Liao, Jingxia Jiang, Erkang Chen, and Peng Chen. Msp-former: Multiscale projection transformer for single image desnowing. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023.", + "[11] Wei-Ting Chen, Hao-Yu Fang, Jian-Jiun Ding, Cheng-Che Tsai, and Sy-Yen Kuo. Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 754-770. Springer, 2020.", + "[12] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4196-4205, 2021.", + "[13] Wei-Ting Chen, Zhi-Kai Huang, Cheng-Che Tsai, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17653–17662, 2022.", + "[14] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5374-5383, 2019." + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Image quality assessment for perceptual image restoration: A new dataset, benchmark and metric. arXiv preprint arXiv:2011.15002, 2020.", + "[16] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In European Conference on Computer Vision, pages 633-651. Springer, 2020.", + "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017.", + "[18] Jin-Hwan Kim, Jae-Young Sim, and Chang-Su Kim. Video deraining and desnowing using temporal correlation and low-rank matrix completion. IEEE Transactions on Image Processing, 24(9):2658-2670, 2015.", + "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.", + "[20] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2018.", + "[21] Minghan Li, Xiangyong Cao, Qian Zhao, Lei Zhang, Chenqiang Gao, and Deyu Meng. Video rain/snow removal by transformed online multiscale convolutional sparse coding. arXiv preprint arXiv:1909.06148, 2019.", + "[22] Minghan Li, Xiangyong Cao, Qian Zhao, Lei Zhang, and Deyu Meng. Online rain/snow removal from surveillance videos. IEEE Transactions on Image Processing, 30:2029-2044, 2021.", + "[23] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021.", + "[24] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent video restoration transformer with guided deformable attention. arXiv preprint arXiv:2206.02146, 2022.", + "[25] Yun-Fu Liu, Da-Wei Jaw, Shih-Chia Huang, and Jenq-Neng Hwang. Desnownet: Context-aware deep network for snow removal. IEEE Transactions on Image Processing, 27(6):3064-3073, 2018.", + "[26] Xin Luo, Yunan Zhu, Shunxin Xu, and Dong Liu. On the effectiveness of spectral discriminators for perceptual quality improvement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023.", + "[27] Chong Mou, Qian Wang, and Jian Zhang. Deep generalized unfolding networks for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17399-17410, 2022.", + "[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13220", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019.", + "[29] Soo-Chang Pei, Yu-Tai Tsai, and Chen-Yu Lee. Removing rain and snow in a single image using saturation and visibility features. In 2014 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 1-6. IEEE, 2014.", + "[30] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4161-4170, 2017.", + "[31] Jingjing Ren, Qingqing Zheng, Yuanyuan Zhao, Xuemiao Xu, and Chen Li. Dlformer: Discrete latent transformer for video inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3511-3520, June 2022.", + "[32] Weihong Ren, Jiandong Tian, Zhi Han, Antoni Chan, and Yandong Tang. Video desnowing and deraining based on matrix decomposition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4210-4219, 2017.", + "[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020.", + "[34] Mingxing Tan, Ruoming Pang, and Quoc V Le. Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10781-10790, 2020.", + "[35] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022.", + "[36] Shuai Wang, Lei Zhu, Huazhu Fu, Jing Qin, Carola-Bibiane Schonlieb, Wei Feng, and Song Wang. Rethinking video rain streak removal: A new synthesis model and a deraining network with video rain prior. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIX, pages 565-582. Springer, 2022.", + "[37] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019.", + "[38] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004.", + "[39] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Zeyu Xiao, Xueyang Fu, Jie Huang, Zhen Cheng, and Zhiwei Xiong. Space-time distillation for video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2113-2122, June 2021.", + "[41] Zeyu Xiao, Yutong Liu, Ruisheng Gao, and Zhiwei Xiong. Cutmib: Boosting light field super-resolution via multi-view image blending. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1672-1682, June 2023.", + "[42] Zeyu Xiao, Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. $\\text{Eva}^2$ : Event-assisted video frame interpolation via cross-modal alignment and aggregation. IEEE Transactions on Computational Imaging, 8:1145-1158, 2022.", + "[43] Zeyu Xiao, Zhiwei Xiong, Xueyang Fu, Dong Liu, and Zheng-Jun Zha. Space-time video super-resolution using temporal profiles. In Proceedings of the 28th ACM International Conference on Multimedia, pages 664-672, 2020.", + "[44] Bin Yang, Zhenhong Jia, Jie Yang, and Nikola K Kasabov. Video snow removal based on self-adaptation snow detection and patch-based gaussian mixture model. IEEE Access, 8:160188-160201, 2020.", + "[45] Yijun Yang, Angelica Aviles-Rivero, Huazhu Fu, Ye Liu, Weiming Wang, and Lei Zhu. Video adverse-weather-component suppression network via weather messenger and adversarial backpropagation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023.", + "[46] Tian Ye, Sixiang Chen, Yun Liu, Yi Ye, Erkang Chen, and Yuche Li. Underwater light field retention: Neural rendering for underwater imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 488-497, 2022.", + "[47] Zongsheng Yue, Jianwen Xie, Qian Zhao, and Deyu Meng. Semi-supervised video deraining with dynamical rain generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 642-652, 2021.", + "[48] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022.", + "[49] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021.", + "[50] Kaihao Zhang, Rongqing Li, Yanjiang Yu, Wenhan Luo, and Changsheng Li. Deep dense multi-scale network for snow removal using semantic and depth priors. IEEE Transactions on Image Processing, 30:7419-7431, 2021.", + "[51] Ruofan Zhang, Jinjin Gu, Haoyu Chen, Chao Dong, Yulun Zhang, and Wenming Yang. Crafting training degradation distribution for the accuracy-generalization trade-off. arXiv preprint arXiv:2305.18107, 2023." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13221", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "ref_text", + "text": "[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018.", + "bbox": [ + 78, + 90, + 470, + 136 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "13222", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Snow Removal in Video_ A New Dataset and A Novel Method/5ae50b63-d59f-4dc4-bf7f-4d6ac34210b4_model.json b/2023/Snow Removal in Video_ A New Dataset and A Novel Method/5ae50b63-d59f-4dc4-bf7f-4d6ac34210b4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7ca68b3f1b8f3d7289902d35c0370494a14d0b3e --- /dev/null +++ b/2023/Snow Removal in Video_ A New Dataset and A Novel Method/5ae50b63-d59f-4dc4-bf7f-4d6ac34210b4_model.json @@ -0,0 +1,2941 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.175, + 0.131, + 0.795, + 0.152 + ], + "angle": 0, + "content": "Snow Removal in Video: A New Dataset and A Novel Method" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.18, + 0.728, + 0.217 + ], + "angle": 0, + "content": "Haoyu Chen\\(^{1}\\), Jingjing Ren\\(^{1}\\), Jinjin Gu\\(^{2}\\), Hongtao Wu\\(^{1}\\), Xuequan Lu\\(^{3}\\), Haoming Cai\\(^{4}\\), Lei Zhu\\(^{1,5}\\)" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.218, + 0.778, + 0.234 + ], + "angle": 0, + "content": "1The Hong Kong University of Science and Technology (Guangzhou) 2The University of Sydney" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.236, + 0.646, + 0.251 + ], + "angle": 0, + "content": "\\(^{3}\\)La Trobe University \\(^{4}\\)The University of Maryland" + }, + { + "type": "text", + "bbox": [ + 0.319, + 0.254, + 0.652, + 0.269 + ], + "angle": 0, + "content": "5The Hong Kong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.273, + 0.712, + 0.287 + ], + "angle": 0, + "content": "Project page: https://haoyuchen.com/VideoDesnowing" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.356, + 0.473, + 0.687 + ], + "angle": 0, + "content": "Snowfall is a common weather phenomenon that can severely affect computer vision tasks by obscuring objects and scenes. However, existing deep learning-based snow removal methods are designed for single images only. In this paper, we target a more complex task - video snow removal, which aims to restore the clear video from the snowy video. To facilitate this task, we propose the first high-quality video dataset, which simulates realistic physical characteristics of snow and haze using a rendering engine and augmentation techniques. We also develop a deep learning framework for video snow removal. Specifically, we propose a snow-query temporal aggregation module and a snow-aware contrastive learning loss function. The module aggregates features between video frames and removes snow effectively, while the loss function helps identify and eliminate snow features. We conduct extensive experiments and demonstrate that our proposed dataset is more realistic than previous datasets, and the models trained on it achieve better performance in real-world snowing images. Our proposed method outperforms state-of-the-art video and image-based methods on both synthetic and real snowy videos." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.722, + 0.21, + 0.739 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Snowfall is a common weather phenomenon that can significantly affect visibility in photographs, interfering with and reducing the accuracy of computer vision tasks such as target detection [34], tracking [14], and autonomous driving [33]. Image and video restoration [39, 23, 48, 37, 46, 7, 5, 26, 40, 43, 42, 41, 45, 31] is a widely studied direction, but research on image and video snow removal is still limited. Removing snow is crucial for improving accuracy and robustness of computer vision tasks. Additionally, snow can obstruct objects, decreasing image and video quality. Snow" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.895, + 0.399 + ], + "angle": 0, + "content": "removal is a challenging task that requires distinguishing snowflake regions from the background and restoring obscured scenes, complicated by the complex geometric shape and texture of snowflakes. Moreover, snowfall often comes with haze, which further obscures the images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.896, + 0.717 + ], + "angle": 0, + "content": "Contrast with single image, the video provides richer information about the dynamic features of the scene, i.e., a consecutive frame sequence on the temporal dimension. Although recent image snow removal algorithms [25, 11, 12, 50, 8, 10] have achieved good results, applying these algorithms directly to video snow removal is inappropriate because image algorithms ignore temporal information in video frames and this information could be very useful to improve snow removal outcomes. As such, it is significant to conduct in-depth research and exploration for snow video data. Previous studies have explored some aspects of video desnowing, such as methods based on non-deep learning [32, 18] or online learning [22]. However, deep learning-based video desnowing remains an under-researched area. A major challenge for applying deep learning methods is the lack of high-quality video desnowing datasets. In the era of powerful deep learning, this becomes an urgent problem that needs to be addressed. In fact, the lack of appropriate datasets has significantly slowed down the progress of deep learning-based video snow removal research, which has been a bottleneck in this field." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.897, + 0.901 + ], + "angle": 0, + "content": "To this end, we present the first high-quality annotated video dataset for video snow removal: Realistic Video DeSnowing Dataset (RVSD). RVSD contains a total of 110 pairs of videos. Each pair contains snowy and hazy videos and corresponding snow-free and haze-free ground truth videos. Unlike previous image datasets that simulate snow using Photoshop, we use a rendering engine and various augmentation techniques to generate snow with diverse and realistic physical properties. This results in more realistic and varied synthesized videos, which improve the model's performance on real-world data. To further enhance the model's performance, we also consider the effect of haze" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13211" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.181, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.189, + 0.158, + 0.199 + ], + "angle": 0, + "content": "a.City Day" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.091, + 0.279, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.198, + 0.189, + 0.262, + 0.199 + ], + "angle": 0, + "content": "b. City Night" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.091, + 0.355, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.305, + 0.189, + 0.352, + 0.199 + ], + "angle": 0, + "content": "c. Nature" + }, + { + "type": "image", + "bbox": [ + 0.397, + 0.091, + 0.495, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.189, + 0.474, + 0.198 + ], + "angle": 0, + "content": "Snow100K" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.091, + 0.593, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.189, + 0.585, + 0.199 + ], + "angle": 0, + "content": "SnowCityScape" + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.091, + 0.692, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.189, + 0.683, + 0.199 + ], + "angle": 0, + "content": "SnowCityScape" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.091, + 0.791, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.727, + 0.189, + 0.758, + 0.198 + ], + "angle": 0, + "content": "SRRS" + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.091, + 0.89, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.828, + 0.189, + 0.854, + 0.198 + ], + "angle": 0, + "content": "CSD" + }, + { + "type": "image_caption", + "bbox": [ + 0.08, + 0.212, + 0.887, + 0.227 + ], + "angle": 0, + "content": "Figure 1: Samples of the proposed dataset (left) and previous datasets (right). Refer to the supplementary to see the videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.254, + 0.469, + 0.494 + ], + "angle": 0, + "content": "in snowy conditions. Previous methods for fog synthesis relied on mathematical models with limitations, such as requiring accurate depth information and assuming fixed and uniform haze characteristics. However, real-world snowy conditions involve dynamic and heterogeneous haze. To overcome these limitations, we propose a new haze model that uses Unreal Engine 5 to render storm haze and obtain a haze layer. We then fuse the haze layer with the input video after data augmentation to produce more diverse and realistic haze, better simulating real snowy conditions. We conduct experiments to show that our proposed dataset has more-realistic snow scenes than existing datasets, and models trained on our dataset achieve better snow removal performance on real snowy images. Therefore, our dataset can facilitate the training and evaluation of video snow removal algorithms, and advance video snow removal research." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.496, + 0.469, + 0.796 + ], + "angle": 0, + "content": "Trivially applying the video processing methods on our dataset may result in suboptimal since they do not consider snow and haze properties. Moreover, there are no existing deep learning-based methods for video snow removal. Therefore, we propose the first deep learning-based framework in this work. Our framework consists of three main steps. First, we perform preliminary snow removal on single frames to reduce the effect of snow occlusion on optical flow estimation and video feature alignment. Second, we fuse the video frames and perceive and remove the snow. We introduce a novel \"snow-query temporal aggregation module\", which aggregates features across video frames while perceiving and removing snow. Third, we design a new \"snow-aware contrastive loss\", which leverages the prior knowledge that snowflakes have different shapes, motion directions, and other characteristics in different videos. This loss function helps identify and remove snow features more accurately. We conduct extensive experiments to show that our proposed method outperforms the state-of-the-art methods. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.803, + 0.469, + 0.877 + ], + "angle": 0, + "content": "- We propose the first high-quality video dataset (RVSD) for video snow removal. We use the Unreal Engine and various augmentation techniques to produce snow and haze with diverse and realistic physical characteristics." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "- Comprehensive experiments show that the visual effect" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.803, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.254, + 0.892, + 0.314 + ], + "angle": 0, + "content": "of our proposed snow removal dataset is more realistic than previous datasets, and the models trained on our proposed dataset can achieve better performance in real-world snowy images." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.325, + 0.892, + 0.402 + ], + "angle": 0, + "content": "- A deep learning-based video snow removal framework is proposed. The snow-aware temporal aggregation module and snow-aware contrastive learning loss are introduced, to effectively remove snow. Experiments prove that our method enables the best performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.426, + 0.641, + 0.442 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.451, + 0.691, + 0.467 + ], + "angle": 0, + "content": "2.1. Desnowing Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.892, + 0.777 + ], + "angle": 0, + "content": "There are several single-image snow removal datasets available. Snow100K [25] first used PhotoShop to synthesize snow. In order to model different types of snow, SnowKITTI2012 [50] and SnowCityScapes [50] contains three levels of snow, including light, medium, and heavy snow. However, these datasets do not take into account the veiling effect of snowy weather, the models trained on these datasets cannot remove haze and have limited practical performance in real snowy scenarios. SRRS [11] firstly simulated the veiling effect inspired by the Koschmieder model, and then used PhotoShop to synthesize the snow streaks. CSD [12] added Gaussian blur to the above two steps to better simulate the real snow image. Nearly all methods use PhotoShop to synthesize snow. As a result, the shape and style of snow are relatively homogeneous, which substantially limits the model's performance on realistic and complex snowing scenes. A quick comparison of these datasets can be found in Table 1. To the best of our knowledge, there is no dataset for video desnowing, which significantly hinders the research of video desnowing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.673, + 0.804 + ], + "angle": 0, + "content": "2.2. Image Desnowing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Early image snow removal methods tend to perform snow removal based on physical priors. [1] performed snow detection and removal by calculating the histogram of orientations of snow streaks. [29] used features on saturation and visibility to remove rain and snow from a single image. In recent years, deep learning has refreshed the snow removal" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13212" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.09, + 0.644, + 0.179 + ], + "angle": 0, + "content": "
DatasetVenueType# imagesResolutionSnow synthesisStatic HazeDynamic HazeIlluminationUser Study (1 - 10)
Snow100K [25]TIP 2018image100K≤ 640 × 640PhotoshopXXX3.98
SnowCityScapes[50]TIP 2021image2K+2K512 × 256PhotoshopXXX4.18
SnowKITTI2012 [50]TIP 2021image1.5K+1K884 × 256PhotoshopXXX3.72
SRRS [11]ECCV 2020image15K640 × 480Photoshop✓XX4.20
CSD [12]ICCV 2021image10K640 × 480Photoshop✓XX4.67
Ours-video-480p - 4KUE Rendering✓✓✓6.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.175, + 0.181, + 0.544, + 0.197 + ], + "angle": 0, + "content": "Table 1: Comparison with previous desnowing datasets." + }, + { + "type": "table", + "bbox": [ + 0.657, + 0.09, + 0.891, + 0.179 + ], + "angle": 0, + "content": "
Scenes# videos# frames# frames (512×512)
city day405,12131,020
city night202,37823,878
nature433,52632,526
other74073,012
In total11011,43290,436
" + }, + { + "type": "table_caption", + "bbox": [ + 0.663, + 0.181, + 0.881, + 0.195 + ], + "angle": 0, + "content": "Table 2: Statistics of our dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.223, + 0.473, + 0.45 + ], + "angle": 0, + "content": "task. The first deep learning based snow removal method DesnowNet [25] proposed a multi-stage, multi-scale design to remove translucent and opaque snow streaks. JSTASR [11] proposed a joint size and transparency-aware snow removal model that can classify snow particles based on their size and remove snow. HDCW-Net [12] used a hierarchical decomposition paradigm in which a dual-tree wavelet transform and a wavelet loss are used. It also proposed a discriminative feature for snow removal called contradict channel. DDMSNet [50] introduced semantic and depth information to learn semantic-aware and geometry-aware representations for snow removal. SnowFormer [9] used a vision transformer architecture that fully combines local and global information, and obtains superior results on multiple datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.46, + 0.248, + 0.476 + ], + "angle": 0, + "content": "2.3. Video Desnowing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.472, + 0.741 + ], + "angle": 0, + "content": "Before the success of deep learning-based methods, traditional computer vision techniques were used for video-based snow removal. Ren et al. [32] utilized the low-rank assumption of the background to separate sparse and dense snow, and addressed heavy snow in dynamic scenes. Kim et al. [18] considered global and local motion, as well as snowflakes of various sizes, in their snowflake removal algorithm with low-rank matrix completion. Due to the rapid improvement of deep learning models and their ability to study complex patterns, some researchers have proposed the use of deep learning for video-based snow removal. [44] utilized self-adaptive snow detection and a patch-based Gaussian mixture model, adept at removing both sparse and dense snowflakes from videos. Li et al. [21, 22] proposed a method for dynamic backgrounds where snow was encoded and removed using an online multi-scale convolutional sparse coding model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.754, + 0.394, + 0.772 + ], + "angle": 0, + "content": "3. Realistic Video Desnowing Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Previous snow and haze synthesis methods suffer from a lack of variation and poor physical accuracy. In contrast, our approach employs a 3D rendering engine and a range of augmentations, together with a more realistic fusion method to create snow and haze with more diverse, complex, and accurate physical properties. As a result, the synthesized video is more realistic and diverse, enabling our model to enjoy better generalization ability to real-world data. We" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.223, + 0.725, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.221, + 0.895, + 0.326 + ], + "angle": 0, + "content": "Figure 2: Box plots of the user study on the visual realism perception scores of different datasets. Higher scores represent better perceptual realism." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.367, + 0.712, + 0.381 + ], + "angle": 0, + "content": "demonstrate this in Section 5.1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.391, + 0.74, + 0.407 + ], + "angle": 0, + "content": "3.1. Collection of Clean Videos" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.895, + 0.642 + ], + "angle": 0, + "content": "Unlike previous work that only considered daytime video, we collect both daytime and nighttime videos in our dataset, which is unlike previous works. We apply different add-snow methods for daytime or nighttime videos to enhance the realism of the generated results. We will discuss these methods in the next section. Our dataset covers a variety of scenes such as buildings, streets, nature scenes, close shots, distant views, overhead and elevation angles to simulate the rich scenes of real snow videos. As shown in Table 2, we have collected 110 videos in total, among which 80 videos are used for training and 30 videos for testing. To accommodate the diverse resolutions of online videos, our video resolutions range from \\(480\\mathrm{p}\\) to \\(4\\mathrm{k}\\). This enables the model to produce better results on videos with more sources and resolutions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.663, + 0.668 + ], + "angle": 0, + "content": "3.2. Haze Rendering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Previous studies have employed a fog synthesis method based on the atmospheric scattering model [20]: \\( I_{haze}(x) = J(x)t(x) + A(1 - t(x)) \\), where \\( I_{haze}(x) \\) is hazy image, \\( J(x) \\) is the haze-free image, \\( A \\) denotes the global atmospheric light, and \\( t(x) \\) is the transmission matrix. This model has both strengths and weaknesses. One of the strengths is that the formula is intuitive and concise, making it easy to embed into network design. For limitations, it relies on depth information, and the generated haze exhibits a relatively uniform pattern. More importantly, there is a fundamental difference between snow-generated haze on a snowy day and conventional haze. Snow haze is often dynamic and follows the movement of snow, as illustrated in Figure 4. Therefore, using traditional haze synthesis methods that assume a homogeneous haze pattern may" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13213" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.124, + 0.089, + 0.845, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.269, + 0.895, + 0.332 + ], + "angle": 0, + "content": "Figure 3: Snow-free videos with rich scenes are selected first. Then, we use two haze models to generate depth-related static haze and Unreal Engine (UE) rendered dynamic haze that is more realistic. After that, we use the UE to render snow layers with realistic physical characteristics. Further the snow layers are dynamically blended according to the background time after complex augmentation. These complex operations make the results more realistic with diverse features." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.355, + 0.47, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.418, + 0.408, + 0.433 + ], + "angle": 0, + "content": "Figure 4: Dynamic haze in a snowstorm." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.445, + 0.451, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.524, + 0.443, + 0.54 + ], + "angle": 0, + "content": "Figure 5: Fusion weight for daytime and nighttime." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.47, + 0.729 + ], + "angle": 0, + "content": "not be accurate in simulating the dynamic and varying snow haze that occurs during a snowstorm. We propose a novel haze model that leverages the capabilities of the Unreal Engine to render the haze of snowstorms and obtain different haze layers \\( H \\). The dynamic haze of the storms is then integrated with the input video following data augmentation. As a result, our approach produces a more diverse and realistic haze that more closely resembles real snow days and nights. This enables the models trained on our data to handle more diverse and dynamic snow haze. Our composite haze model is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.731, + 0.469, + 0.748 + ], + "angle": 0, + "content": "\\[\nI _ {h a z e} (x) = J (x) t (x) + A (1 - t (x)) + \\operatorname {A u g} (H), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.749, + 0.338, + 0.763 + ], + "angle": 0, + "content": "where Aug denotes data augmentation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.772, + 0.241, + 0.789 + ], + "angle": 0, + "content": "3.3. Snow Rendering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In previous snow production work [11, 25, 50, 12], snowflakes and snow streaks were synthesized using Photoshop, resulting in relatively homogeneous patterns. In contrast, our work employs Unreal Engine to synthesize snow. As a 3D computer graphics engine, Unreal Engine enables the creation of snow with more realistic and complex physical features, including snowflake flip, spatial changes from" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.354, + 0.895, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.483, + 0.892, + 0.514 + ], + "angle": 0, + "content": "Figure 6: The visibility of daytime snow and nighttime snow varies depending on the background." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.864 + ], + "angle": 0, + "content": "far to near, and perspective changes, among others. We use various 3D snow models to manipulate the physical properties of the snow, resulting in diverse and complex shapes that more closely resemble real-life snow. More details of the features are shown in Table 3. Once the snow layer \\( S \\) is generated, it needs to be blended with the clean video. We augment the snow layer by adjusting the color temperature, brightness, contrast, and applying random affine transformations. Then we add different levels of blur, such as Gaussian blur and motion blur. Finally, we consider the temporal properties of the video background based on time of day. During the day, especially when the sky is in the background, snowflakes tend to blend with the bright sky, resulting in snowflakes that are not easily distinguishable. At night, snowflakes are more visible in bright places, such as those illuminated by headlights or streetlights, while they tend to be less visible in places without light. We show some real samples in Figure 6. Based on this property, we utilize Eq. (2) to perform the integration of snow layer and video background." + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.865, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\alpha (x) = \\left\\{ \\begin{array}{c} \\sigma (- (V (x) - \\gamma) \\times \\beta), \\text {i f d a y t i m e} \\\\ \\sigma ((V (x) - \\gamma) \\times \\beta), \\text {i f n i g h t t i m e} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13214" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.09, + 0.47, + 0.23 + ], + "angle": 0, + "content": "
FeatureDetails
Snow Speedby setting gravity intensity, wind speed, air friction coefficient and other parameters, snow can move at different speeds
Snow Intensityby setting the number of particles in the snow model, our dataset contains snow of various degrees.
Snow Typeby employing various snow models and adjusting various parameters, combined with the Unreal Engine to impart realistic physical features to the snowflakes (such as turbulence, rotation, etc.), the resulting snow has a highly diverse range of shapes, sizes, and characteristics.
Motion directionhead-on, back-on, left-facing, right-facing, and many other angles
Illuminationlights have different intensities and colors, such as different warm and cold tones
Veiling effectdepth-related static haze, as well as rendered dynamic complex haze.
Scenescity, street, day & night, nature
" + }, + { + "type": "table_caption", + "bbox": [ + 0.134, + 0.253, + 0.411, + 0.268 + ], + "angle": 0, + "content": "Table 3: Features of our proposed dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.294, + 0.469, + 0.429 + ], + "angle": 0, + "content": "Specifically, we first convert the RGB image to the HSV color space, where \\( V = \\max(R, G, B) \\), represents the largest color component. We obtain a corresponding blending weight based on the pixel's brightness. \\( v \\) denotes the value of the V channel in the HSV space, which is normalized to the range of 0 to 1. \\( \\gamma \\) and \\( \\beta \\) are two parameters artificially adjusted based on the specific video, and \\( \\sigma \\) represents the softmax function. Figure 5 shows some examples. Our dynamic video snow model is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.431, + 0.469, + 0.447 + ], + "angle": 0, + "content": "\\[\nZ (x) = I _ {h a z e} (x) + \\alpha (x) \\operatorname {A u g} (S (x)). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.469, + 0.476 + ], + "angle": 0, + "content": "where Aug denotes data augmentation. \\( Z \\) is the final output with both snow and haze." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.491, + 0.433, + 0.508 + ], + "angle": 0, + "content": "4. Snow-Aware Video Desnowing Network" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.283, + 0.53 + ], + "angle": 0, + "content": "4.1. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Our network architecture is composed of three main modules: the encoder, the snow-query temporal aggregation module, and the decoder, as shown in Figure 7. To mitigate the impact of snow occlusion on video frame alignment and fusion, the encoder takes single-frame images as input and generates coarse features \\( F^{C} \\), which are then transformed into coarse desnowed images by the recovery module. However, residual snow may still remain in the video frames due to the lack of utilization of video information. Therefore, it is imperative to incorporate multi-frame video information to further remove residual snow. To achieve this, the Snow-query Temporal Aggregation Module takes in continuous video frames and the features obtained by the encoder. This module aligns and fuses video features, and can perceive the location of snow, which facilitates more effective snow removal. Then this module outputs continuous refined video features \\( F^{F} \\), which are fed into the decoder. Finally, the recovery module is employed to obtain the final desnowed video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Recovery Module. To simultaneously remove snow and snow-induced haze, we adopt the atmospheric scattering model to decompose images containing both snow and haze into three parts: snow feature \\( S \\), the global atmospheric light \\( A \\), and transmission matrix \\( T \\). We perform" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "this operation in the feature space of the image. Here, \\(S \\in \\mathbb{R}^{N \\times C \\times H \\times W}\\), while \\(A \\in \\mathbb{R}^{N \\times 1 \\times H \\times W}\\) and \\(T \\in \\mathbb{R}^{N \\times C \\times H \\times W}\\). \\(N\\) is the number of frames, \\(C\\) is the number of feature channel, \\(H\\) and \\(W\\) are the height and width. The recovery process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.167, + 0.892, + 0.2 + ], + "angle": 0, + "content": "\\[\nF (x) = \\frac {E (I (x)) - S (x) - (1 - T (x)) A (x)}{T (x)}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.202, + 0.673, + 0.218 + ], + "angle": 0, + "content": "\\[\nJ (x) = M (F (x)),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.219, + 0.892, + 0.279 + ], + "angle": 0, + "content": "where \\( I(x) \\) is the input snowy image, \\( F(x) \\) is the output desnowed feature, \\( E(\\cdot) \\) is the feature extractor, \\( M(\\cdot) \\) is a convolution layer that maps the feature into image space, and \\( J(x) \\) is the snow-free image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.28, + 0.892, + 0.551 + ], + "angle": 0, + "content": "Coarse Desnowing of Single Video Frames Prior to Alignment. Video restoration networks frequently employ optical flow or deformable convolution to integrate information from multiple frames for feature extraction. However, this is not entirely suitable for video snow removal owing to the significant masking of the video background by snowflakes and snow lines, which disturb temporal feature aggregation. The position discrepancies between snowflakes and snow lines in adjacent frames, especially in heavily snowing videos, pose a challenge to alignment methods in estimating accurate motion parameters. Thus, coarse desnowing of single frames before temporal feature attending is crucial for effectively integrating information between frames in video snow removal later on. This step does not cause any information loss for the further step. The reason is that the desnowing inputs of the further step are the features of original snowy video frames, instead of the desnowed images of the first step." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.702 + ], + "angle": 0, + "content": "Figure 9 shows the results of optical flow computation on video frames with and without snow. Optical flow computation on snow-covered video frames shows significant errors, which may mislead the alignment of video frame features. Our idea is to obtain coarse desnowed frames before computing optical flows. Specifically, the encoder takes a single snowy frame as input and outputs coarse features \\( S^C \\), \\( A^C \\), and \\( T^C \\). The recovery module then restores the coarse desnowed frame feature \\( F^C \\) and its corresponding desnowed frame \\( x^C \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.81, + 0.728 + ], + "angle": 0, + "content": "4.2. Snow-Query Temporal Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Temporal Aggregation Module. To fully utilize the features between video frames, we use bidirectional propagation to independently propagate the frames' features forward and backward in time dimension. This technique enables the information to flow in both directions, which can help capture more spatial-temporal context and improve the model's accuracy. Given coarse desnowed frame \\( x_{i}^{C} \\) and its neighboring frame \\( x_{i-1}^{C} \\) and \\( x_{i+1}^{C} \\), we use SpyNet [30] to estimate the optical flows \\( O_{t+1 \\to t} \\) and \\( O_{t-1 \\to t} \\). To obtain the corresponding features propagated from the neighboring frames, which are denoted as \\( h_{t-1}^{f} \\) and \\( h_{t+1}^{b} \\), we have" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13215" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.891, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.29, + 0.897, + 0.368 + ], + "angle": 0, + "content": "Figure 7: Snow-Aware Video Desnowing Network (SVDNet). The architecture is composed of three modules: the encoder, the snow-query temporal aggregation module, and the decoder. To avoid the effect of snow occlusion on video alignment, the encoder first uses a single frame to generate coarse feature \\( F^C \\) and desnowed image \\( x^C \\). Then the Snow-query Temporal Aggregation Module aligns and fuses video features, detects snow, and outputs continuous fine video features \\( F^F \\). These features are then fed into the decoder, and the recovery module is used to obtain the final desnowed video." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.39, + 0.473, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.625, + 0.455, + 0.641 + ], + "angle": 0, + "content": "Figure 8: Snow-Query Temporal Aggregation Module." + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.676, + 0.393, + 0.695 + ], + "angle": 0, + "content": "\\[\nh _ {t} ^ {b} = H _ {b} \\left(F _ {t} ^ {C}, O _ {t + 1 \\rightarrow t}, h _ {t + 1} ^ {b}, T _ {t} ^ {C}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.694, + 0.469, + 0.723 + ], + "angle": 0, + "content": "\\[\nh _ {t} ^ {f} = H _ {f} \\left(F _ {t} ^ {C}, O _ {t - 1 \\rightarrow t}, h _ {t - 1} ^ {f}, T _ {t} ^ {C}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.727, + 0.47, + 0.756 + ], + "angle": 0, + "content": "where \\( H_{b} \\) and \\( H_{f} \\) denote the backward and forward propagation branches." + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.757, + 0.334, + 0.777 + ], + "angle": 0, + "content": "\\[\nw _ {t} ^ {\\{b, f \\}} = W \\left(h _ {t \\pm 1}, O _ {t \\pm 1 \\rightarrow t}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.779, + 0.469, + 0.806 + ], + "angle": 0, + "content": "\\[\nh _ {t} ^ {\\{b, f \\}} = \\operatorname {M S A} \\left(\\operatorname {c a t} \\left(w _ {t} ^ {\\{b, f \\}}, F _ {t} ^ {C}, T _ {t} ^ {C}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.808, + 0.282, + 0.826 + ], + "angle": 0, + "content": "\\[\nh _ {t} = \\operatorname {c a t} \\left(h _ {t} ^ {f}, h _ {t} ^ {b}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.469, + 0.855 + ], + "angle": 0, + "content": "where \\( W \\) denotes feature warping, and MSA denotes multi-head self-attention." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Snow-Query Cross-Attention Module. The snow features for the same video are always similar. Thus, we can further use the existing snow features to match and guide the" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.39, + 0.895, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.492, + 0.889, + 0.508 + ], + "angle": 0, + "content": "Figure 9: Incorrect estimation of optical flow due to snow." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.529, + 0.892, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.67, + 0.894, + 0.732 + ], + "angle": 0, + "content": "Figure 10: t-SNE visualization of features learned with/without the contrastive loss. Our contrastive loss function effectively distinguishes different types of snow while without it, their feature distributions highly overlap." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.901 + ], + "angle": 0, + "content": "network to find the remaining snow. We show this module in Figure 8. First, we perform patch embedding for \\( h_t \\) and corresponding snow features \\( S_t^C \\). We consider each patch of size \\( 2 \\times 2 \\times C \\) as a token. Then, a linear embedding layer is applied to project the features of each token to an arbitrary dimension. The key design is using snow features as queries. The snow features can help find the residual snow regions in the video feature. Specifically, the queries obtained from the snow feature and the key obtained from the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13216" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.088, + 0.895, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.258, + 0.895, + 0.29 + ], + "angle": 0, + "content": "Figure 11: Visual results on the real-world images. The same model is trained separately on different datasets for image snow removal. The model trained on our proposed dataset achieves the best results for snow and haze removal on real images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.315, + 0.47, + 0.375 + ], + "angle": 0, + "content": "video feature are multiplied to obtain a correlation matrix. The coordinates with high correlation represent the residual snow in the video, which enables the network to locate the residual snow and further remove it." + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.374, + 0.469, + 0.417 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C r o s s A t t n} \\left(\\mathcal {Q} _ {s}, \\mathcal {K} _ {v}, \\mathcal {V} _ {v}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {Q} _ {s} \\mathcal {K} _ {v}}{\\sqrt {d}} + p\\right) \\mathcal {V} _ {v}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.47, + 0.464 + ], + "angle": 0, + "content": "where \\( Q_{s}, K_{v}, V_{v} \\) are the snow query, video key, and value matrices. \\( d \\) is the dimension. \\( p \\) is the relative position embedding." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.472, + 0.345, + 0.487 + ], + "angle": 0, + "content": "4.3. Snow-Aware Contrastive Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.496, + 0.47, + 0.675 + ], + "angle": 0, + "content": "We introduce a prior for snow-covered images: the snow features vary across videos based on the snow's angle and properties (e.g., density, shape, direction, and speed), but are consistent within the same video, as shown in Figure 7. We design a snow-aware contrastive learning paradigm. It uses snow features from different frames of the same video as positive samples \\( S^{+} \\) and snow features from different videos as negative samples \\( S^{-} \\). By pulling positive samples closer together and pushing negative samples farther apart in feature space, the network can more accurately identify snow features and improve snow removal. The loss function for contrastive learning is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.677, + 0.48, + 0.748 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o n t}} \\left(S, S ^ {+}, S ^ {-}\\right) = \\\\ - \\log \\left[ \\frac {\\exp \\left(D \\left(S ^ {+} , S ^ {-}\\right) / \\tau\\right)}{\\exp \\left(D \\left(S ^ {+} , S ^ {-}\\right) / \\tau\\right) + \\sum_ {r = 1} ^ {N} \\exp \\left(D \\left(S ^ {+} , S ^ {-}\\right) / \\tau\\right)} \\right] \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.81 + ], + "angle": 0, + "content": "where \\(D(x,y) = \\Psi (x)\\cdot \\Psi (y)\\). \\(\\Psi (\\cdot)\\) is the two-layer MLP function which implements feature projection. \\(\\tau\\) is the scale temperature which is set to 0.1. \\(N\\) denotes the total number of negative samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To illustrate the effect of the proposed snow-aware contrastive learning loss on snow feature learning, we visualize the learned features using t-SNE for three distinct types of snow with different motion directions and snow stripe sizes, as shown in Figure 10. Without the proposed contrastive learning, different snow types have highly similar and over" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.315, + 0.895, + 0.406 + ], + "angle": 0, + "content": "lapping feature distributions. In contrast, with the contrastive learning loss, the features of different snow types are clearly distinguished, demonstrating the ability of the method to accurately identify diverse snow features. The overall loss function is formulated as: \\(\\mathcal{L} = \\mathcal{L}_1 + \\lambda \\mathcal{L}_{cont}\\) where \\(\\lambda\\) is set to 0.1." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.426, + 0.634, + 0.443 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.453, + 0.838, + 0.469 + ], + "angle": 0, + "content": "5.1. Proposed Dataset vs. Previous Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Objective Evaluation. Previous work [4, 51] has demonstrated the significant impact of training set data on the generalization performance of the model. To demonstrate that the snow scenes in our created dataset are more realistic and closer to real-world scenarios than those in other synthetic datasets, we train the same model on different datasets and test it on real snowy images. The results show that the model trained on our proposed dataset performs better in handling real snow, thus validating the authenticity and usefulness of our dataset. Figure 11 displays two samples. In the first row, the model trained on our dataset successfully removes all snow residues, whereas models trained on other datasets fail to completely eliminate snow, leaving numerous snow residues in the output images. Since the Snow100K, SnowCityScapes, and SnowKITTI2012 datasets do not include the effect of snow haze, the models trained on these datasets cannot remove snow haze. In the second row, the models trained on the Snow100K, SnowCityScapes, and SnowKITTI2012 datasets remove slightly larger snowflakes but preserve many smaller ones. Furthermore, these models mistakenly remove white parallelogram backgrounds, mistaking them for snowflakes. Conversely, SRRS and CSD remove smaller snowflakes but fail to remove larger ones. In contrast, our proposed dataset facilitates the training of a model that effectively removes snowflakes at all sizes. It does not misidentify the white background as snow. Since there is no ground truth for real-world images, we conduct a user study to compare the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13217" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.088, + 0.473, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.198, + 0.471, + 0.245 + ], + "angle": 0, + "content": "Figure 12: Distributions of mean opinion scores obtained by models trained on different datasets (testing on real-world snow images)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.272, + 0.47, + 0.332 + ], + "angle": 0, + "content": "desnowing results of models which are trained on different datasets, as shown in Figure 12. Model trained on our dataset has the best desnowing performance on real-world images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.471, + 0.471 + ], + "angle": 0, + "content": "Subjective Evaluation. To demonstrate that our dataset produces snow images that are closer to real snow in human visual perception than previous datasets, we conduct a user study to compare the perceived realism of snow generated from different datasets. The results are shown in Figure 2, dataset with higher scores indicating greater fidelity. Thanks to our proposed complex, specially designed dataset construction method, our dataset is scored much higher than the other datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.482, + 0.303, + 0.497 + ], + "angle": 0, + "content": "5.2. Performance Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.627 + ], + "angle": 0, + "content": "Implementation Details. Each input image is randomly cropped to a spatial resolution of \\(256 \\times 256\\), and the number of the total training iterations is \\(500\\mathrm{K}\\). During the training phase, we adopt the Adam optimizer [19] with \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.99\\). The initial learning rate is set to \\(2 \\times 10^{-4}\\) and reduced by half at the milestone of \\(100\\mathrm{K}\\) iterations and \\(150\\mathrm{K}\\) iterations. The batch size is set to 4. PyTorch [28] is used to implement our model with 4 RTX 3090 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.749 + ], + "angle": 0, + "content": "Compared Methods. We compare our network against 18 state-of-the-art methods, including three single-image desnowing models, two image adverse weather restoration models, six image restoration methods, two video deraining methods, and five video restoration methods. Please refer to Table 5 for more details. Moreover, we follow [16, 15] to employ PSNR, SSIM [38], and LPIPS [52] to quantitatively compared different methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Quantitative Comparison. Table 5 reports the quantitative results of our network and 17 compared methods. From these quantitative results, we can find that the performance of video-based models is generally inferior to that of single-image models. The major reason behind is that the severe occlusion of the video background by snowflakes hinders the alignment and fusion of the video frames. Among all compared methods, Restormer has the largest PSNR score of 24.34 and the smallest 0.1164, while MPRNet has the largest SSIM score of 0.8960. More importantly, our net" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.09, + 0.891, + 0.13 + ], + "angle": 0, + "content": "
MethodRestormerTransWeatherS2VDRVRTBasicVSR++Ours
E_warp ↓(×10-3)1.5942.3922.0093.4252.6131.203
VFID ↓0.09380.15110.16860.22070.15390.0492
" + }, + { + "type": "table_caption", + "bbox": [ + 0.516, + 0.152, + 0.874, + 0.168 + ], + "angle": 0, + "content": "Table 4: Results of temporal consistency preservation." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.184, + 0.89, + 0.41 + ], + "angle": 0, + "content": "
TypeMethodVenuePSNR ↑SSIM ↑LPIPS ↓
Image DesnowingJSTASR [11]ECCV 202022.080.82800.2336
HDCW-Net [12]ICCV 202122.630.85920.2010
SnowFormer [8]arXiv 202224.010.89390.1219
Image Adverse WeatherTransWeather [35]CVPR 202223.110.85430.2086
TKL [13]CVPR 202223.050.85890.2027
Image RestorationSwinIR [23]CVPR 202122.510.85620.2105
MPRNet [49]CVPR 202124.270.89600.1266
Uformer [39]CVPR 202223.610.87300.1706
DGUNet [27]CVPR 202224.180.89850.1238
NAFNet [6]ECCV 202224.010.88380.1472
Restormer [48]CVPR 202224.340.89290.1164
Video DerainingS2VD [47]CVPR 202122.950.85900.1856
RDD [36]ECCV 202222.970.87420.1631
Video RestorationEDVR [37]CVPRW 201917.930.57900.3771
BasicVSR [2]CVPR 202122.460.84730.2087
IconVSR [2]CVPR 202122.350.84820.2034
BasicVSR++ [3]CVPR 202222.640.86180.1868
RVRT [24]NeurlPS 202220.900.79740.2977
SVDNet (Ours)-25.060.92100.0842
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.433, + 0.891, + 0.449 + ], + "angle": 0, + "content": "Table 5: Quantitative results on the proposed video dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.892, + 0.612 + ], + "angle": 0, + "content": "work further outperform these methods in terms of PSNR, SSIM, and LPIPS scores. It improves the PSNR score from 24.34 to 25.06, the SSIM score from 0.8960 to 0.9210, and the LPIPS score from 0.1164 to 0.0842. It demonstrates the effectiveness of our proposed framework, which first obtains a coarse snow removal result for a single frame image to reduce the challenging alignment issue on the snowy video frames, and then fully utilizes the temporal information between video frames for video desnowing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.703 + ], + "angle": 0, + "content": "To further verify the temporal consistency preservation, we utilize a widely-used metric \\( E_{\\text{warp}} \\) to compute the temporal consistency, and VFID [17] to compute video restoration performance. Table below shows our method has the smallest \\( E_{\\text{warp}} \\) and VFID. Our method can achieve consistent desnowing between frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Visual comparisons on synthetic videos. Figure 13 visually compares video desnowing results produced by our network and state-of-the-art methods. We can find that existing methods tend to maintain snowflakes (e.g., TKL, TransWeather), or haze (e.g., Restomer), or both snowflakes and haze (i.e., RVRT) in their desnowing results. On the contrary, our method can effectively remove both snowflakes and haze, and better preserve background details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Visual comparisons on real-world videos. Moreover, Figure 14 shows the visual comparisons between our network and state-of-the-art methods in terms of input frames from real-world snowstorm videos with both heavy snow and haze. We can find that our method can effectively remove" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13218" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.197, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.163, + 0.171, + 0.173 + ], + "angle": 0, + "content": "Input Frame" + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.089, + 0.313, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.163, + 0.278, + 0.172 + ], + "angle": 0, + "content": "TKL [13]" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.089, + 0.428, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.327, + 0.163, + 0.414, + 0.173 + ], + "angle": 0, + "content": "TransWeather [35]" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.089, + 0.544, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.448, + 0.163, + 0.523, + 0.173 + ], + "angle": 0, + "content": "Snowformer [8]" + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.089, + 0.658, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.163, + 0.636, + 0.173 + ], + "angle": 0, + "content": "Restormer [48]" + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.089, + 0.774, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.163, + 0.748, + 0.173 + ], + "angle": 0, + "content": "MPRNet [49]" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.089, + 0.888, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.798, + 0.163, + 0.863, + 0.173 + ], + "angle": 0, + "content": "DGUNet [27]" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.174, + 0.197, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.246, + 0.167, + 0.257 + ], + "angle": 0, + "content": "S2VD [47]" + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.174, + 0.313, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.246, + 0.279, + 0.257 + ], + "angle": 0, + "content": "RDD [36]" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.174, + 0.428, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.344, + 0.246, + 0.398, + 0.257 + ], + "angle": 0, + "content": "RVRT [24]" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.174, + 0.543, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.246, + 0.516, + 0.257 + ], + "angle": 0, + "content": "IconVSR [2]" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.174, + 0.658, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.246, + 0.641, + 0.257 + ], + "angle": 0, + "content": "BasicVSR++ [3]" + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.174, + 0.774, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.703, + 0.246, + 0.728, + 0.257 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.174, + 0.888, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.798, + 0.246, + 0.863, + 0.257 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.164, + 0.269, + 0.804, + 0.285 + ], + "angle": 0, + "content": "Figure 13: Visual video desnowing results produced by our network and state-of-the-art methods." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.309, + 0.178, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.384, + 0.158, + 0.393 + ], + "angle": 0, + "content": "Input Frame" + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.309, + 0.274, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.196, + 0.384, + 0.257, + 0.393 + ], + "angle": 0, + "content": "MPRNet [49]" + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.309, + 0.37, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.384, + 0.352, + 0.393 + ], + "angle": 0, + "content": "DGUNet [27]" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.309, + 0.466, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.384, + 0.44, + 0.393 + ], + "angle": 0, + "content": "RDD [36]" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.394, + 0.178, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.467, + 0.155, + 0.477 + ], + "angle": 0, + "content": "S2VD [47]" + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.394, + 0.274, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.467, + 0.255, + 0.477 + ], + "angle": 0, + "content": "IconVSR [2]" + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.394, + 0.37, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.467, + 0.359, + 0.477 + ], + "angle": 0, + "content": "BasicVSR++[3]" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.394, + 0.466, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.468, + 0.43, + 0.477 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.489, + 0.461, + 0.505 + ], + "angle": 0, + "content": "Figure 14: Visual comparison on real snowstorm video." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.521, + 0.468, + 0.607 + ], + "angle": 0, + "content": "
IndexTemporal AggregationSnow-QueryContrastive LossPSNR ↑SSIM ↑LPIPS ↓
M1(basic)24.160.89900.1125
M2✓24.580.91300.0934
M3✓✓24.800.91080.9004
M4✓✓24.840.91810.0882
Ours✓✓✓25.060.92100.0842
" + }, + { + "type": "table_caption", + "bbox": [ + 0.108, + 0.63, + 0.438, + 0.646 + ], + "angle": 0, + "content": "Table 6: Quantitative results of the ablation study." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.719 + ], + "angle": 0, + "content": "both the haze and snowflakes and betterrecover the obscured background details, while compared methods tend to main parts of haze and snowflakes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Ablation Study. We perform ablation study experiments to verify the temporal aggregation module, the snow-query cross-attention module, and the snow-aware contrastive loss of our network. To do so, we first built a basic model (see M1 of Table 6) by removing these three major components from our network, and then add the temporal aggregation module into M1 to construct M2. Then we add the snow-query cross-attention module into M2 to build M3, and add the snow-aware contrastive loss into M2 to build M4. Table 6 reports the PSNR, SSIM, and LPIPS results of our network and four baseline networks. Specifically, compared to M1, M2 improves the PSNR score from 24.16" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.895, + 0.478 + ], + "angle": 0, + "content": "to 24.58, the SSIM score from 0.8990 to 0.9130, and the LPIPS score from 0.1125 to 0.0934. It shows that the temporal aggregation module incurs a better video desnowing performance. Then, we find that M3 and M4 have a superior PSNR, SSIM, and LPIPS performance over M2. It demonstrates the effectiveness of the snow-aware cross-attention module or snow-aware contrastive loss for enhancing the video desnowing performance. Moreover, our method has larger PSNR and SSIM scores and a smaller LPIPS score than M3 and M4. It shows that combing three components together has the best performance of video desnowing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.489, + 0.619, + 0.505 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.514, + 0.895, + 0.636 + ], + "angle": 0, + "content": "The first contribution of our work is to synthesize the first high-quality video desnowing dataset, which is more realistic than previous image datasets. Then, we devise a deep learning-based framework that incorporates a snow-query temporal aggregation module and a snow-aware contrastive learning loss. Experimental results show that our network outperforms state-of-the-art methods in terms of video denosing on synthetic and real-world snowy videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.894, + 0.699 + ], + "angle": 0, + "content": "Acknowledgment. This work was supported by the National Natural Science Foundation of China (Grant No. 61902275), and Guangzhou Municipal Science and Technology Project (Grant No. 2023A03J0671)." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.71, + 0.597, + 0.725 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.734, + 0.895, + 0.789 + ], + "angle": 0, + "content": "[1] Jérémie Bossu, Nicolas Hautiere, and Jean-Philippe Tarel. Rain or snow detection in image sequences through use of a histogram of orientation of streaks. International journal of computer vision, 93:348–367, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.79, + 0.894, + 0.859 + ], + "angle": 0, + "content": "[2] Kelvin C.K. Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.86, + 0.895, + 0.902 + ], + "angle": 0, + "content": "[3] Kelvin C.K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. BasicVSR++: Improving video superresolution with enhanced propagation and alignment. In" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.734, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13219" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "IEEE Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[4] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1692-1703, June 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.205, + 0.468, + 0.232 + ], + "angle": 0, + "content": "[5] Haoyu Chen, Jinjin Gu, and Zhi Zhang. Attention in attention network for image super-resolution, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.233, + 0.468, + 0.302 + ], + "angle": 0, + "content": "[6] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 17-33. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.303, + 0.468, + 0.344 + ], + "angle": 0, + "content": "[7] Sixiang Chen, Tian Ye, Yun Liu, and Erkang Chen. Dualformer: Hybrid self-attention transformer for efficient image restoration. arXiv preprint arXiv:2210.01069, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.345, + 0.468, + 0.399 + ], + "angle": 0, + "content": "[8] Sixiang Chen, Tian Ye, Yun Liu, Erkang Chen, Jun Shi, and Jingchun Zhou. Snowformer: Scale-aware transformer via context interaction for single image desnowing. arXiv preprint arXiv:2208.09703, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.401, + 0.468, + 0.455 + ], + "angle": 0, + "content": "[9] Sixiang Chen, Tian Ye, Yun Liu, Erkang Chen, Jun Shi, and Jingchun Zhou. Snowformer: Scale-aware transformer via context interaction for single image desnowing. arXiv preprint arXiv:2208.09703, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.457, + 0.468, + 0.538 + ], + "angle": 0, + "content": "[10] Sixiang Chen, Tian Ye, Yun Liu, Taodong Liao, Jingxia Jiang, Erkang Chen, and Peng Chen. Msp-former: Multiscale projection transformer for single image desnowing. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.54, + 0.468, + 0.636 + ], + "angle": 0, + "content": "[11] Wei-Ting Chen, Hao-Yu Fang, Jian-Jiun Ding, Cheng-Che Tsai, and Sy-Yen Kuo. Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 754-770. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.638, + 0.468, + 0.719 + ], + "angle": 0, + "content": "[12] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4196-4205, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.468, + 0.815 + ], + "angle": 0, + "content": "[13] Wei-Ting Chen, Zhi-Kai Huang, Cheng-Che Tsai, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17653–17662, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[14] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5374-5383, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[15] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Image quality assessment for perceptual image restoration: A new dataset, benchmark and metric. arXiv preprint arXiv:2011.15002, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.15, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[16] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In European Conference on Computer Vision, pages 633-651. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[18] Jin-Hwan Kim, Jae-Young Sim, and Chang-Su Kim. Video deraining and desnowing using temporal correlation and low-rank matrix completion. IEEE Transactions on Image Processing, 24(9):2658-2670, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.348, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[20] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[21] Minghan Li, Xiangyong Cao, Qian Zhao, Lei Zhang, Chenqiang Gao, and Deyu Meng. Video rain/snow removal by transformed online multiscale convolutional sparse coding. arXiv preprint arXiv:1909.06148, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[22] Minghan Li, Xiangyong Cao, Qian Zhao, Lei Zhang, and Deyu Meng. Online rain/snow removal from surveillance videos. IEEE Transactions on Image Processing, 30:2029-2044, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[23] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.7 + ], + "angle": 0, + "content": "[24] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent video restoration transformer with guided deformable attention. arXiv preprint arXiv:2206.02146, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.703, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[25] Yun-Fu Liu, Da-Wei Jaw, Shih-Chia Huang, and Jenq-Neng Hwang. Desnownet: Context-aware deep network for snow removal. IEEE Transactions on Image Processing, 27(6):3064-3073, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[26] Xin Luo, Yunan Zhu, Shunxin Xu, and Dong Liu. On the effectiveness of spectral discriminators for perceptual quality improvement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[27] Chong Mou, Qian Wang, and Jian Zhang. Deep generalized unfolding networks for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17399-17410, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13220" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.202 + ], + "angle": 0, + "content": "[29] Soo-Chang Pei, Yu-Tai Tsai, and Chen-Yu Lee. Removing rain and snow in a single image using saturation and visibility features. In 2014 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 1-6. IEEE, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[30] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4161-4170, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[31] Jingjing Ren, Qingqing Zheng, Yuanyuan Zhao, Xuemiao Xu, and Chen Li. Dlformer: Discrete latent transformer for video inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3511-3520, June 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.398 + ], + "angle": 0, + "content": "[32] Weihong Ren, Jiandong Tian, Zhi Han, Antoni Chan, and Yandong Tang. Video desnowing and deraining based on matrix decomposition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4210-4219, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.47, + 0.538 + ], + "angle": 0, + "content": "[34] Mingxing Tan, Ruoming Pang, and Quoc V Le. Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10781-10790, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.54, + 0.469, + 0.608 + ], + "angle": 0, + "content": "[35] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[36] Shuai Wang, Lei Zhu, Huazhu Fu, Jing Qin, Carola-Bibiane Schonlieb, Wei Feng, and Song Wang. Rethinking video rain streak removal: A new synthesis model and a deraining network with video rain prior. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIX, pages 565-582. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[37] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[38] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[40] Zeyu Xiao, Xueyang Fu, Jie Huang, Zhen Cheng, and Zhiwei Xiong. Space-time distillation for video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2113-2122, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[41] Zeyu Xiao, Yutong Liu, Ruisheng Gao, and Zhiwei Xiong. Cutmib: Boosting light field super-resolution via multi-view image blending. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1672-1682, June 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.292 + ], + "angle": 0, + "content": "[42] Zeyu Xiao, Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. \\( \\text{Eva}^2 \\): Event-assisted video frame interpolation via cross-modal alignment and aggregation. IEEE Transactions on Computational Imaging, 8:1145-1158, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.296, + 0.892, + 0.351 + ], + "angle": 0, + "content": "[43] Zeyu Xiao, Zhiwei Xiong, Xueyang Fu, Dong Liu, and Zheng-Jun Zha. Space-time video super-resolution using temporal profiles. In Proceedings of the 28th ACM International Conference on Multimedia, pages 664-672, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.354, + 0.892, + 0.408 + ], + "angle": 0, + "content": "[44] Bin Yang, Zhenhong Jia, Jie Yang, and Nikola K Kasabov. Video snow removal based on self-adaptation snow detection and patch-based gaussian mixture model. IEEE Access, 8:160188-160201, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.412, + 0.892, + 0.493 + ], + "angle": 0, + "content": "[45] Yijun Yang, Angelica Aviles-Rivero, Huazhu Fu, Ye Liu, Weiming Wang, and Lei Zhu. Video adverse-weather-component suppression network via weather messenger and adversarial backpropagation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[46] Tian Ye, Sixiang Chen, Yun Liu, Yi Ye, Erkang Chen, and Yuche Li. Underwater light field retention: Neural rendering for underwater imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 488-497, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.571, + 0.892, + 0.626 + ], + "angle": 0, + "content": "[47] Zongsheng Yue, Jianwen Xie, Qian Zhao, and Deyu Meng. Semi-supervised video deraining with dynamical rain generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 642-652, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.629, + 0.892, + 0.711 + ], + "angle": 0, + "content": "[48] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.715, + 0.892, + 0.784 + ], + "angle": 0, + "content": "[49] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.787, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[50] Kaihao Zhang, Rongqing Li, Yanjiang Yu, Wenhan Luo, and Changsheng Li. Deep dense multi-scale network for snow removal using semantic and depth priors. IEEE Transactions on Image Processing, 30:7419-7431, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[51] Ruofan Zhang, Jinjin Gu, Haoyu Chen, Chao Dong, Yulun Zhang, and Wenming Yang. Crafting training degradation distribution for the accuracy-generalization trade-off. arXiv preprint arXiv:2305.18107, 2023." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13221" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.472, + 0.137 + ], + "angle": 0, + "content": "[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13222" + } + ] +] \ No newline at end of file diff --git a/2023/Snow Removal in Video_ A New Dataset and A Novel Method/5ae50b63-d59f-4dc4-bf7f-4d6ac34210b4_origin.pdf b/2023/Snow Removal in Video_ A New Dataset and A Novel Method/5ae50b63-d59f-4dc4-bf7f-4d6ac34210b4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..14220cf64c5c3efbf0c232c5800d421a9577c0ba --- /dev/null +++ b/2023/Snow Removal in Video_ A New Dataset and A Novel Method/5ae50b63-d59f-4dc4-bf7f-4d6ac34210b4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c3f0661ba9de7a3c7dd0bf1527bfba5c3633d93eaaf43f70856eec5878700da +size 11882976 diff --git a/2023/Snow Removal in Video_ A New Dataset and A Novel Method/full.md b/2023/Snow Removal in Video_ A New Dataset and A Novel Method/full.md new file mode 100644 index 0000000000000000000000000000000000000000..555ee0033543c7a7e5de5f4dbf791cf89c6bb9ed --- /dev/null +++ b/2023/Snow Removal in Video_ A New Dataset and A Novel Method/full.md @@ -0,0 +1,420 @@ +# Snow Removal in Video: A New Dataset and A Novel Method + +Haoyu Chen $^{1}$ , Jingjing Ren $^{1}$ , Jinjin Gu $^{2}$ , Hongtao Wu $^{1}$ , Xuequan Lu $^{3}$ , Haoming Cai $^{4}$ , Lei Zhu $^{1,5}$ + +1The Hong Kong University of Science and Technology (Guangzhou) 2The University of Sydney + +$^{3}$ La Trobe University $^{4}$ The University of Maryland + +5The Hong Kong University of Science and Technology + +Project page: https://haoyuchen.com/VideoDesnowing + +# Abstract + +Snowfall is a common weather phenomenon that can severely affect computer vision tasks by obscuring objects and scenes. However, existing deep learning-based snow removal methods are designed for single images only. In this paper, we target a more complex task - video snow removal, which aims to restore the clear video from the snowy video. To facilitate this task, we propose the first high-quality video dataset, which simulates realistic physical characteristics of snow and haze using a rendering engine and augmentation techniques. We also develop a deep learning framework for video snow removal. Specifically, we propose a snow-query temporal aggregation module and a snow-aware contrastive learning loss function. The module aggregates features between video frames and removes snow effectively, while the loss function helps identify and eliminate snow features. We conduct extensive experiments and demonstrate that our proposed dataset is more realistic than previous datasets, and the models trained on it achieve better performance in real-world snowing images. Our proposed method outperforms state-of-the-art video and image-based methods on both synthetic and real snowy videos. + +# 1. Introduction + +Snowfall is a common weather phenomenon that can significantly affect visibility in photographs, interfering with and reducing the accuracy of computer vision tasks such as target detection [34], tracking [14], and autonomous driving [33]. Image and video restoration [39, 23, 48, 37, 46, 7, 5, 26, 40, 43, 42, 41, 45, 31] is a widely studied direction, but research on image and video snow removal is still limited. Removing snow is crucial for improving accuracy and robustness of computer vision tasks. Additionally, snow can obstruct objects, decreasing image and video quality. Snow + +removal is a challenging task that requires distinguishing snowflake regions from the background and restoring obscured scenes, complicated by the complex geometric shape and texture of snowflakes. Moreover, snowfall often comes with haze, which further obscures the images. + +Contrast with single image, the video provides richer information about the dynamic features of the scene, i.e., a consecutive frame sequence on the temporal dimension. Although recent image snow removal algorithms [25, 11, 12, 50, 8, 10] have achieved good results, applying these algorithms directly to video snow removal is inappropriate because image algorithms ignore temporal information in video frames and this information could be very useful to improve snow removal outcomes. As such, it is significant to conduct in-depth research and exploration for snow video data. Previous studies have explored some aspects of video desnowing, such as methods based on non-deep learning [32, 18] or online learning [22]. However, deep learning-based video desnowing remains an under-researched area. A major challenge for applying deep learning methods is the lack of high-quality video desnowing datasets. In the era of powerful deep learning, this becomes an urgent problem that needs to be addressed. In fact, the lack of appropriate datasets has significantly slowed down the progress of deep learning-based video snow removal research, which has been a bottleneck in this field. + +To this end, we present the first high-quality annotated video dataset for video snow removal: Realistic Video DeSnowing Dataset (RVSD). RVSD contains a total of 110 pairs of videos. Each pair contains snowy and hazy videos and corresponding snow-free and haze-free ground truth videos. Unlike previous image datasets that simulate snow using Photoshop, we use a rendering engine and various augmentation techniques to generate snow with diverse and realistic physical properties. This results in more realistic and varied synthesized videos, which improve the model's performance on real-world data. To further enhance the model's performance, we also consider the effect of haze + +![](images/32dd73bf420f912e63db375ae0a0858c395f4433b54006127e7ca1c151e8b6a3.jpg) +a.City Day + +![](images/3de314e6c425295fe6ccad00951d1985e084e36bcc8768b4d1bd2bd0624200e7.jpg) +b. City Night + +![](images/1fa1488917077fdbdbe1616848314e000e355dbd8777a9abe7a88c92257b702d.jpg) +c. Nature +Figure 1: Samples of the proposed dataset (left) and previous datasets (right). Refer to the supplementary to see the videos. + +![](images/f73943a08fd6883b84647a4324019f9359578b359b1cc79d6caf50dc09b3fffd.jpg) +Snow100K + +![](images/24ac02178277a49b6e6b62340d6cb379dd2a48329ca06bd1fe8ccc25ff642b1b.jpg) +SnowCityScape + +![](images/046b13f627357807135358a66d498d33867906575fd9be8dd46cf19167597a18.jpg) +SnowCityScape + +![](images/fb8ddd6cb5f9548b7b829f9bd117c025f9db7303d7fa749ffd8ae4f7319d8bdc.jpg) +SRRS + +![](images/ddeb05d42fb0086e2a5e834f424fbbb94fdd1d106d74b0c01a1dd8141357c18b.jpg) +CSD + +in snowy conditions. Previous methods for fog synthesis relied on mathematical models with limitations, such as requiring accurate depth information and assuming fixed and uniform haze characteristics. However, real-world snowy conditions involve dynamic and heterogeneous haze. To overcome these limitations, we propose a new haze model that uses Unreal Engine 5 to render storm haze and obtain a haze layer. We then fuse the haze layer with the input video after data augmentation to produce more diverse and realistic haze, better simulating real snowy conditions. We conduct experiments to show that our proposed dataset has more-realistic snow scenes than existing datasets, and models trained on our dataset achieve better snow removal performance on real snowy images. Therefore, our dataset can facilitate the training and evaluation of video snow removal algorithms, and advance video snow removal research. + +Trivially applying the video processing methods on our dataset may result in suboptimal since they do not consider snow and haze properties. Moreover, there are no existing deep learning-based methods for video snow removal. Therefore, we propose the first deep learning-based framework in this work. Our framework consists of three main steps. First, we perform preliminary snow removal on single frames to reduce the effect of snow occlusion on optical flow estimation and video feature alignment. Second, we fuse the video frames and perceive and remove the snow. We introduce a novel "snow-query temporal aggregation module", which aggregates features across video frames while perceiving and removing snow. Third, we design a new "snow-aware contrastive loss", which leverages the prior knowledge that snowflakes have different shapes, motion directions, and other characteristics in different videos. This loss function helps identify and remove snow features more accurately. We conduct extensive experiments to show that our proposed method outperforms the state-of-the-art methods. Our contributions are summarized as follows: + +- We propose the first high-quality video dataset (RVSD) for video snow removal. We use the Unreal Engine and various augmentation techniques to produce snow and haze with diverse and realistic physical characteristics. +- Comprehensive experiments show that the visual effect + +of our proposed snow removal dataset is more realistic than previous datasets, and the models trained on our proposed dataset can achieve better performance in real-world snowy images. + +- A deep learning-based video snow removal framework is proposed. The snow-aware temporal aggregation module and snow-aware contrastive learning loss are introduced, to effectively remove snow. Experiments prove that our method enables the best performance. + +# 2. Related Work + +# 2.1. Desnowing Datasets + +There are several single-image snow removal datasets available. Snow100K [25] first used PhotoShop to synthesize snow. In order to model different types of snow, SnowKITTI2012 [50] and SnowCityScapes [50] contains three levels of snow, including light, medium, and heavy snow. However, these datasets do not take into account the veiling effect of snowy weather, the models trained on these datasets cannot remove haze and have limited practical performance in real snowy scenarios. SRRS [11] firstly simulated the veiling effect inspired by the Koschmieder model, and then used PhotoShop to synthesize the snow streaks. CSD [12] added Gaussian blur to the above two steps to better simulate the real snow image. Nearly all methods use PhotoShop to synthesize snow. As a result, the shape and style of snow are relatively homogeneous, which substantially limits the model's performance on realistic and complex snowing scenes. A quick comparison of these datasets can be found in Table 1. To the best of our knowledge, there is no dataset for video desnowing, which significantly hinders the research of video desnowing. + +# 2.2. Image Desnowing + +Early image snow removal methods tend to perform snow removal based on physical priors. [1] performed snow detection and removal by calculating the histogram of orientations of snow streaks. [29] used features on saturation and visibility to remove rain and snow from a single image. In recent years, deep learning has refreshed the snow removal + +
DatasetVenueType# imagesResolutionSnow synthesisStatic HazeDynamic HazeIlluminationUser Study (1 - 10)
Snow100K [25]TIP 2018image100K≤ 640 × 640PhotoshopXXX3.98
SnowCityScapes[50]TIP 2021image2K+2K512 × 256PhotoshopXXX4.18
SnowKITTI2012 [50]TIP 2021image1.5K+1K884 × 256PhotoshopXXX3.72
SRRS [11]ECCV 2020image15K640 × 480Photoshop✓XX4.20
CSD [12]ICCV 2021image10K640 × 480Photoshop✓XX4.67
Ours-video-480p - 4KUE Rendering✓✓✓6.94
+ +Table 1: Comparison with previous desnowing datasets. + +
Scenes# videos# frames# frames (512×512)
city day405,12131,020
city night202,37823,878
nature433,52632,526
other74073,012
In total11011,43290,436
+ +Table 2: Statistics of our dataset. + +task. The first deep learning based snow removal method DesnowNet [25] proposed a multi-stage, multi-scale design to remove translucent and opaque snow streaks. JSTASR [11] proposed a joint size and transparency-aware snow removal model that can classify snow particles based on their size and remove snow. HDCW-Net [12] used a hierarchical decomposition paradigm in which a dual-tree wavelet transform and a wavelet loss are used. It also proposed a discriminative feature for snow removal called contradict channel. DDMSNet [50] introduced semantic and depth information to learn semantic-aware and geometry-aware representations for snow removal. SnowFormer [9] used a vision transformer architecture that fully combines local and global information, and obtains superior results on multiple datasets. + +# 2.3. Video Desnowing + +Before the success of deep learning-based methods, traditional computer vision techniques were used for video-based snow removal. Ren et al. [32] utilized the low-rank assumption of the background to separate sparse and dense snow, and addressed heavy snow in dynamic scenes. Kim et al. [18] considered global and local motion, as well as snowflakes of various sizes, in their snowflake removal algorithm with low-rank matrix completion. Due to the rapid improvement of deep learning models and their ability to study complex patterns, some researchers have proposed the use of deep learning for video-based snow removal. [44] utilized self-adaptive snow detection and a patch-based Gaussian mixture model, adept at removing both sparse and dense snowflakes from videos. Li et al. [21, 22] proposed a method for dynamic backgrounds where snow was encoded and removed using an online multi-scale convolutional sparse coding model. + +# 3. Realistic Video Desnowing Dataset + +Previous snow and haze synthesis methods suffer from a lack of variation and poor physical accuracy. In contrast, our approach employs a 3D rendering engine and a range of augmentations, together with a more realistic fusion method to create snow and haze with more diverse, complex, and accurate physical properties. As a result, the synthesized video is more realistic and diverse, enabling our model to enjoy better generalization ability to real-world data. We + +![](images/a2523a3842090bd29425f3304a23cf2c779414b3d56dd64644f4a7d147092538.jpg) +Figure 2: Box plots of the user study on the visual realism perception scores of different datasets. Higher scores represent better perceptual realism. + +demonstrate this in Section 5.1. + +# 3.1. Collection of Clean Videos + +Unlike previous work that only considered daytime video, we collect both daytime and nighttime videos in our dataset, which is unlike previous works. We apply different add-snow methods for daytime or nighttime videos to enhance the realism of the generated results. We will discuss these methods in the next section. Our dataset covers a variety of scenes such as buildings, streets, nature scenes, close shots, distant views, overhead and elevation angles to simulate the rich scenes of real snow videos. As shown in Table 2, we have collected 110 videos in total, among which 80 videos are used for training and 30 videos for testing. To accommodate the diverse resolutions of online videos, our video resolutions range from $480\mathrm{p}$ to $4\mathrm{k}$ . This enables the model to produce better results on videos with more sources and resolutions. + +# 3.2. Haze Rendering + +Previous studies have employed a fog synthesis method based on the atmospheric scattering model [20]: $I_{haze}(x) = J(x)t(x) + A(1 - t(x))$ , where $I_{haze}(x)$ is hazy image, $J(x)$ is the haze-free image, $A$ denotes the global atmospheric light, and $t(x)$ is the transmission matrix. This model has both strengths and weaknesses. One of the strengths is that the formula is intuitive and concise, making it easy to embed into network design. For limitations, it relies on depth information, and the generated haze exhibits a relatively uniform pattern. More importantly, there is a fundamental difference between snow-generated haze on a snowy day and conventional haze. Snow haze is often dynamic and follows the movement of snow, as illustrated in Figure 4. Therefore, using traditional haze synthesis methods that assume a homogeneous haze pattern may + +![](images/1f22993164006276ad80b34809152f86f08a836fbb8fadbce77595690e2a085b.jpg) +Figure 3: Snow-free videos with rich scenes are selected first. Then, we use two haze models to generate depth-related static haze and Unreal Engine (UE) rendered dynamic haze that is more realistic. After that, we use the UE to render snow layers with realistic physical characteristics. Further the snow layers are dynamically blended according to the background time after complex augmentation. These complex operations make the results more realistic with diverse features. + +![](images/09004c2c7292247fd322e48992d8db2ba9791ada351a7a483efeb05d021d86bd.jpg) +Figure 4: Dynamic haze in a snowstorm. + +![](images/5d26b40a8818d9ba3c7ce85df13c1d03803c0a1fab35b2f527dc23ff81535c56.jpg) +Figure 5: Fusion weight for daytime and nighttime. + +not be accurate in simulating the dynamic and varying snow haze that occurs during a snowstorm. We propose a novel haze model that leverages the capabilities of the Unreal Engine to render the haze of snowstorms and obtain different haze layers $H$ . The dynamic haze of the storms is then integrated with the input video following data augmentation. As a result, our approach produces a more diverse and realistic haze that more closely resembles real snow days and nights. This enables the models trained on our data to handle more diverse and dynamic snow haze. Our composite haze model is defined as follows: + +$$ +I _ {h a z e} (x) = J (x) t (x) + A (1 - t (x)) + \operatorname {A u g} (H), \tag {1} +$$ + +where Aug denotes data augmentation. + +# 3.3. Snow Rendering + +In previous snow production work [11, 25, 50, 12], snowflakes and snow streaks were synthesized using Photoshop, resulting in relatively homogeneous patterns. In contrast, our work employs Unreal Engine to synthesize snow. As a 3D computer graphics engine, Unreal Engine enables the creation of snow with more realistic and complex physical features, including snowflake flip, spatial changes from + +![](images/0bef36974df88c23b1126bf9cfeecb19c4d988fe00ff8557a5580b026f82fd95.jpg) +Figure 6: The visibility of daytime snow and nighttime snow varies depending on the background. + +far to near, and perspective changes, among others. We use various 3D snow models to manipulate the physical properties of the snow, resulting in diverse and complex shapes that more closely resemble real-life snow. More details of the features are shown in Table 3. Once the snow layer $S$ is generated, it needs to be blended with the clean video. We augment the snow layer by adjusting the color temperature, brightness, contrast, and applying random affine transformations. Then we add different levels of blur, such as Gaussian blur and motion blur. Finally, we consider the temporal properties of the video background based on time of day. During the day, especially when the sky is in the background, snowflakes tend to blend with the bright sky, resulting in snowflakes that are not easily distinguishable. At night, snowflakes are more visible in bright places, such as those illuminated by headlights or streetlights, while they tend to be less visible in places without light. We show some real samples in Figure 6. Based on this property, we utilize Eq. (2) to perform the integration of snow layer and video background. + +$$ +\alpha (x) = \left\{ \begin{array}{c} \sigma (- (V (x) - \gamma) \times \beta), \text {i f d a y t i m e} \\ \sigma ((V (x) - \gamma) \times \beta), \text {i f n i g h t t i m e} \end{array} \right. \tag {2} +$$ + +
FeatureDetails
Snow Speedby setting gravity intensity, wind speed, air friction coefficient and other parameters, snow can move at different speeds
Snow Intensityby setting the number of particles in the snow model, our dataset contains snow of various degrees.
Snow Typeby employing various snow models and adjusting various parameters, combined with the Unreal Engine to impart realistic physical features to the snowflakes (such as turbulence, rotation, etc.), the resulting snow has a highly diverse range of shapes, sizes, and characteristics.
Motion directionhead-on, back-on, left-facing, right-facing, and many other angles
Illuminationlights have different intensities and colors, such as different warm and cold tones
Veiling effectdepth-related static haze, as well as rendered dynamic complex haze.
Scenescity, street, day & night, nature
+ +Table 3: Features of our proposed dataset. + +Specifically, we first convert the RGB image to the HSV color space, where $V = \max(R, G, B)$ , represents the largest color component. We obtain a corresponding blending weight based on the pixel's brightness. $v$ denotes the value of the V channel in the HSV space, which is normalized to the range of 0 to 1. $\gamma$ and $\beta$ are two parameters artificially adjusted based on the specific video, and $\sigma$ represents the softmax function. Figure 5 shows some examples. Our dynamic video snow model is defined as follows: + +$$ +Z (x) = I _ {h a z e} (x) + \alpha (x) \operatorname {A u g} (S (x)). \tag {3} +$$ + +where Aug denotes data augmentation. $Z$ is the final output with both snow and haze. + +# 4. Snow-Aware Video Desnowing Network + +# 4.1. Network Architecture + +Our network architecture is composed of three main modules: the encoder, the snow-query temporal aggregation module, and the decoder, as shown in Figure 7. To mitigate the impact of snow occlusion on video frame alignment and fusion, the encoder takes single-frame images as input and generates coarse features $F^{C}$ , which are then transformed into coarse desnowed images by the recovery module. However, residual snow may still remain in the video frames due to the lack of utilization of video information. Therefore, it is imperative to incorporate multi-frame video information to further remove residual snow. To achieve this, the Snow-query Temporal Aggregation Module takes in continuous video frames and the features obtained by the encoder. This module aligns and fuses video features, and can perceive the location of snow, which facilitates more effective snow removal. Then this module outputs continuous refined video features $F^{F}$ , which are fed into the decoder. Finally, the recovery module is employed to obtain the final desnowed video. + +Recovery Module. To simultaneously remove snow and snow-induced haze, we adopt the atmospheric scattering model to decompose images containing both snow and haze into three parts: snow feature $S$ , the global atmospheric light $A$ , and transmission matrix $T$ . We perform + +this operation in the feature space of the image. Here, $S \in \mathbb{R}^{N \times C \times H \times W}$ , while $A \in \mathbb{R}^{N \times 1 \times H \times W}$ and $T \in \mathbb{R}^{N \times C \times H \times W}$ . $N$ is the number of frames, $C$ is the number of feature channel, $H$ and $W$ are the height and width. The recovery process can be formulated as: + +$$ +F (x) = \frac {E (I (x)) - S (x) - (1 - T (x)) A (x)}{T (x)}, \tag {4} +$$ + +$$ +J (x) = M (F (x)), +$$ + +where $I(x)$ is the input snowy image, $F(x)$ is the output desnowed feature, $E(\cdot)$ is the feature extractor, $M(\cdot)$ is a convolution layer that maps the feature into image space, and $J(x)$ is the snow-free image. + +Coarse Desnowing of Single Video Frames Prior to Alignment. Video restoration networks frequently employ optical flow or deformable convolution to integrate information from multiple frames for feature extraction. However, this is not entirely suitable for video snow removal owing to the significant masking of the video background by snowflakes and snow lines, which disturb temporal feature aggregation. The position discrepancies between snowflakes and snow lines in adjacent frames, especially in heavily snowing videos, pose a challenge to alignment methods in estimating accurate motion parameters. Thus, coarse desnowing of single frames before temporal feature attending is crucial for effectively integrating information between frames in video snow removal later on. This step does not cause any information loss for the further step. The reason is that the desnowing inputs of the further step are the features of original snowy video frames, instead of the desnowed images of the first step. + +Figure 9 shows the results of optical flow computation on video frames with and without snow. Optical flow computation on snow-covered video frames shows significant errors, which may mislead the alignment of video frame features. Our idea is to obtain coarse desnowed frames before computing optical flows. Specifically, the encoder takes a single snowy frame as input and outputs coarse features $S^C$ , $A^C$ , and $T^C$ . The recovery module then restores the coarse desnowed frame feature $F^C$ and its corresponding desnowed frame $x^C$ . + +# 4.2. Snow-Query Temporal Aggregation + +Temporal Aggregation Module. To fully utilize the features between video frames, we use bidirectional propagation to independently propagate the frames' features forward and backward in time dimension. This technique enables the information to flow in both directions, which can help capture more spatial-temporal context and improve the model's accuracy. Given coarse desnowed frame $x_{i}^{C}$ and its neighboring frame $x_{i-1}^{C}$ and $x_{i+1}^{C}$ , we use SpyNet [30] to estimate the optical flows $O_{t+1 \to t}$ and $O_{t-1 \to t}$ . To obtain the corresponding features propagated from the neighboring frames, which are denoted as $h_{t-1}^{f}$ and $h_{t+1}^{b}$ , we have + +![](images/6a3b07189f814bc2aeddae9f4f5fcb1a5a2b7267d3a3e121ca06d6d4f6d9e0ad.jpg) +Figure 7: Snow-Aware Video Desnowing Network (SVDNet). The architecture is composed of three modules: the encoder, the snow-query temporal aggregation module, and the decoder. To avoid the effect of snow occlusion on video alignment, the encoder first uses a single frame to generate coarse feature $F^C$ and desnowed image $x^C$ . Then the Snow-query Temporal Aggregation Module aligns and fuses video features, detects snow, and outputs continuous fine video features $F^F$ . These features are then fed into the decoder, and the recovery module is used to obtain the final desnowed video. + +![](images/05e77ff5934c7740a3aac6cf61c8aac096cdf12166ca21dfe460efb5c3c72a24.jpg) +Figure 8: Snow-Query Temporal Aggregation Module. + +$$ +h _ {t} ^ {b} = H _ {b} \left(F _ {t} ^ {C}, O _ {t + 1 \rightarrow t}, h _ {t + 1} ^ {b}, T _ {t} ^ {C}\right), +$$ + +$$ +h _ {t} ^ {f} = H _ {f} \left(F _ {t} ^ {C}, O _ {t - 1 \rightarrow t}, h _ {t - 1} ^ {f}, T _ {t} ^ {C}\right), \tag {5} +$$ + +where $H_{b}$ and $H_{f}$ denote the backward and forward propagation branches. + +$$ +w _ {t} ^ {\{b, f \}} = W \left(h _ {t \pm 1}, O _ {t \pm 1 \rightarrow t}\right), +$$ + +$$ +h _ {t} ^ {\{b, f \}} = \operatorname {M S A} \left(\operatorname {c a t} \left(w _ {t} ^ {\{b, f \}}, F _ {t} ^ {C}, T _ {t} ^ {C}\right)\right), \tag {6} +$$ + +$$ +h _ {t} = \operatorname {c a t} \left(h _ {t} ^ {f}, h _ {t} ^ {b}\right), +$$ + +where $W$ denotes feature warping, and MSA denotes multi-head self-attention. + +Snow-Query Cross-Attention Module. The snow features for the same video are always similar. Thus, we can further use the existing snow features to match and guide the + +![](images/8593678d02f16642073ff64ee6d7f42dde3dd45f8b3393435830af8e92d01d11.jpg) +Figure 9: Incorrect estimation of optical flow due to snow. + +![](images/fb7bef89128f082c4e886bf104a14bd026a7c75354e6631a2f4194bdf3635194.jpg) +Figure 10: t-SNE visualization of features learned with/without the contrastive loss. Our contrastive loss function effectively distinguishes different types of snow while without it, their feature distributions highly overlap. + +network to find the remaining snow. We show this module in Figure 8. First, we perform patch embedding for $h_t$ and corresponding snow features $S_t^C$ . We consider each patch of size $2 \times 2 \times C$ as a token. Then, a linear embedding layer is applied to project the features of each token to an arbitrary dimension. The key design is using snow features as queries. The snow features can help find the residual snow regions in the video feature. Specifically, the queries obtained from the snow feature and the key obtained from the + +![](images/bf277ede75ba1f761a9a6662bb5540a85939b4183a51af12f8679712d9609633.jpg) +Figure 11: Visual results on the real-world images. The same model is trained separately on different datasets for image snow removal. The model trained on our proposed dataset achieves the best results for snow and haze removal on real images. + +video feature are multiplied to obtain a correlation matrix. The coordinates with high correlation represent the residual snow in the video, which enables the network to locate the residual snow and further remove it. + +$$ +\operatorname {C r o s s A t t n} \left(\mathcal {Q} _ {s}, \mathcal {K} _ {v}, \mathcal {V} _ {v}\right) = \operatorname {S o f t m a x} \left(\frac {\mathcal {Q} _ {s} \mathcal {K} _ {v}}{\sqrt {d}} + p\right) \mathcal {V} _ {v}, \tag {7} +$$ + +where $Q_{s}, K_{v}, V_{v}$ are the snow query, video key, and value matrices. $d$ is the dimension. $p$ is the relative position embedding. + +# 4.3. Snow-Aware Contrastive Loss + +We introduce a prior for snow-covered images: the snow features vary across videos based on the snow's angle and properties (e.g., density, shape, direction, and speed), but are consistent within the same video, as shown in Figure 7. We design a snow-aware contrastive learning paradigm. It uses snow features from different frames of the same video as positive samples $S^{+}$ and snow features from different videos as negative samples $S^{-}$ . By pulling positive samples closer together and pushing negative samples farther apart in feature space, the network can more accurately identify snow features and improve snow removal. The loss function for contrastive learning is formulated as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {c o n t}} \left(S, S ^ {+}, S ^ {-}\right) = \\ - \log \left[ \frac {\exp \left(D \left(S ^ {+} , S ^ {-}\right) / \tau\right)}{\exp \left(D \left(S ^ {+} , S ^ {-}\right) / \tau\right) + \sum_ {r = 1} ^ {N} \exp \left(D \left(S ^ {+} , S ^ {-}\right) / \tau\right)} \right] \tag {8} \\ \end{array} +$$ + +where $D(x,y) = \Psi (x)\cdot \Psi (y)$ . $\Psi (\cdot)$ is the two-layer MLP function which implements feature projection. $\tau$ is the scale temperature which is set to 0.1. $N$ denotes the total number of negative samples. + +To illustrate the effect of the proposed snow-aware contrastive learning loss on snow feature learning, we visualize the learned features using t-SNE for three distinct types of snow with different motion directions and snow stripe sizes, as shown in Figure 10. Without the proposed contrastive learning, different snow types have highly similar and over + +lapping feature distributions. In contrast, with the contrastive learning loss, the features of different snow types are clearly distinguished, demonstrating the ability of the method to accurately identify diverse snow features. The overall loss function is formulated as: $\mathcal{L} = \mathcal{L}_1 + \lambda \mathcal{L}_{cont}$ where $\lambda$ is set to 0.1. + +# 5. Experiments + +# 5.1. Proposed Dataset vs. Previous Datasets + +Objective Evaluation. Previous work [4, 51] has demonstrated the significant impact of training set data on the generalization performance of the model. To demonstrate that the snow scenes in our created dataset are more realistic and closer to real-world scenarios than those in other synthetic datasets, we train the same model on different datasets and test it on real snowy images. The results show that the model trained on our proposed dataset performs better in handling real snow, thus validating the authenticity and usefulness of our dataset. Figure 11 displays two samples. In the first row, the model trained on our dataset successfully removes all snow residues, whereas models trained on other datasets fail to completely eliminate snow, leaving numerous snow residues in the output images. Since the Snow100K, SnowCityScapes, and SnowKITTI2012 datasets do not include the effect of snow haze, the models trained on these datasets cannot remove snow haze. In the second row, the models trained on the Snow100K, SnowCityScapes, and SnowKITTI2012 datasets remove slightly larger snowflakes but preserve many smaller ones. Furthermore, these models mistakenly remove white parallelogram backgrounds, mistaking them for snowflakes. Conversely, SRRS and CSD remove smaller snowflakes but fail to remove larger ones. In contrast, our proposed dataset facilitates the training of a model that effectively removes snowflakes at all sizes. It does not misidentify the white background as snow. Since there is no ground truth for real-world images, we conduct a user study to compare the + +![](images/6c436a64193c6d1b2f86bd1fb918bebcb81dc21583cf833d3ce556a0841c6353.jpg) +Figure 12: Distributions of mean opinion scores obtained by models trained on different datasets (testing on real-world snow images). + +desnowing results of models which are trained on different datasets, as shown in Figure 12. Model trained on our dataset has the best desnowing performance on real-world images. + +Subjective Evaluation. To demonstrate that our dataset produces snow images that are closer to real snow in human visual perception than previous datasets, we conduct a user study to compare the perceived realism of snow generated from different datasets. The results are shown in Figure 2, dataset with higher scores indicating greater fidelity. Thanks to our proposed complex, specially designed dataset construction method, our dataset is scored much higher than the other datasets. + +# 5.2. Performance Evaluation + +Implementation Details. Each input image is randomly cropped to a spatial resolution of $256 \times 256$ , and the number of the total training iterations is $500\mathrm{K}$ . During the training phase, we adopt the Adam optimizer [19] with $\beta_{1} = 0.9$ and $\beta_{2} = 0.99$ . The initial learning rate is set to $2 \times 10^{-4}$ and reduced by half at the milestone of $100\mathrm{K}$ iterations and $150\mathrm{K}$ iterations. The batch size is set to 4. PyTorch [28] is used to implement our model with 4 RTX 3090 GPUs. + +Compared Methods. We compare our network against 18 state-of-the-art methods, including three single-image desnowing models, two image adverse weather restoration models, six image restoration methods, two video deraining methods, and five video restoration methods. Please refer to Table 5 for more details. Moreover, we follow [16, 15] to employ PSNR, SSIM [38], and LPIPS [52] to quantitatively compared different methods. + +Quantitative Comparison. Table 5 reports the quantitative results of our network and 17 compared methods. From these quantitative results, we can find that the performance of video-based models is generally inferior to that of single-image models. The major reason behind is that the severe occlusion of the video background by snowflakes hinders the alignment and fusion of the video frames. Among all compared methods, Restormer has the largest PSNR score of 24.34 and the smallest 0.1164, while MPRNet has the largest SSIM score of 0.8960. More importantly, our net + +
MethodRestormerTransWeatherS2VDRVRTBasicVSR++Ours
E_warp ↓(×10-3)1.5942.3922.0093.4252.6131.203
VFID ↓0.09380.15110.16860.22070.15390.0492
+ +Table 4: Results of temporal consistency preservation. + +
TypeMethodVenuePSNR ↑SSIM ↑LPIPS ↓
Image DesnowingJSTASR [11]ECCV 202022.080.82800.2336
HDCW-Net [12]ICCV 202122.630.85920.2010
SnowFormer [8]arXiv 202224.010.89390.1219
Image Adverse WeatherTransWeather [35]CVPR 202223.110.85430.2086
TKL [13]CVPR 202223.050.85890.2027
Image RestorationSwinIR [23]CVPR 202122.510.85620.2105
MPRNet [49]CVPR 202124.270.89600.1266
Uformer [39]CVPR 202223.610.87300.1706
DGUNet [27]CVPR 202224.180.89850.1238
NAFNet [6]ECCV 202224.010.88380.1472
Restormer [48]CVPR 202224.340.89290.1164
Video DerainingS2VD [47]CVPR 202122.950.85900.1856
RDD [36]ECCV 202222.970.87420.1631
Video RestorationEDVR [37]CVPRW 201917.930.57900.3771
BasicVSR [2]CVPR 202122.460.84730.2087
IconVSR [2]CVPR 202122.350.84820.2034
BasicVSR++ [3]CVPR 202222.640.86180.1868
RVRT [24]NeurlPS 202220.900.79740.2977
SVDNet (Ours)-25.060.92100.0842
+ +Table 5: Quantitative results on the proposed video dataset. + +work further outperform these methods in terms of PSNR, SSIM, and LPIPS scores. It improves the PSNR score from 24.34 to 25.06, the SSIM score from 0.8960 to 0.9210, and the LPIPS score from 0.1164 to 0.0842. It demonstrates the effectiveness of our proposed framework, which first obtains a coarse snow removal result for a single frame image to reduce the challenging alignment issue on the snowy video frames, and then fully utilizes the temporal information between video frames for video desnowing. + +To further verify the temporal consistency preservation, we utilize a widely-used metric $E_{\text{warp}}$ to compute the temporal consistency, and VFID [17] to compute video restoration performance. Table below shows our method has the smallest $E_{\text{warp}}$ and VFID. Our method can achieve consistent desnowing between frames. + +Visual comparisons on synthetic videos. Figure 13 visually compares video desnowing results produced by our network and state-of-the-art methods. We can find that existing methods tend to maintain snowflakes (e.g., TKL, TransWeather), or haze (e.g., Restomer), or both snowflakes and haze (i.e., RVRT) in their desnowing results. On the contrary, our method can effectively remove both snowflakes and haze, and better preserve background details. + +Visual comparisons on real-world videos. Moreover, Figure 14 shows the visual comparisons between our network and state-of-the-art methods in terms of input frames from real-world snowstorm videos with both heavy snow and haze. We can find that our method can effectively remove + +![](images/6f58e5f0032cc5cf935036c23af03def374ca884cc2c68e723915cb60a82f441.jpg) + +![](images/8b65bab7ea4fad30b21806a96664127f68502af3edcaff070fdcecdcb9031eb3.jpg) +TKL [13] + +![](images/3d5614415600dedbda16c5cb2cf2541ef58b393190ae2c7225d20466f01fbbbe.jpg) + +![](images/775bcbe1f5b8f820706f060f7c6d68616ebcf4dbce1719f2298720ccd41e24ea.jpg) + +![](images/b1c82358c4a06d9b227c8920f56888bd4f5eb533437d980e42b8e484cc4d8c53.jpg) + +![](images/a2f8fef4027292335e8696c318056a28c7ae95f626ffbb44e530487488c18976.jpg) + +![](images/9eaa5d5b0500203c25687d5440a1d7094e22ab431e11c7eb72e8c47b0c71d37b.jpg) + +![](images/2f3c69faaca36ba166e933a4a243ce97635567aaf5aa9f76a9996fe0a904d020.jpg) +Input Frame +S2VD [47] +Figure 13: Visual video desnowing results produced by our network and state-of-the-art methods. + +![](images/53caa05dec53d0f360ef223673273f20133036ed9f8b458ebe831bd834e2d628.jpg) +RDD [36] + +![](images/cc46f46f0318eed6ab50472f49abf23ef0bd600dd3dde40db0c02d5096228685.jpg) +TransWeather [35] +RVRT [24] + +![](images/576b1a935a986d2f5064f000e1dafafa14bbe0ec3ca617a6d854914165967238.jpg) +Snowformer [8] +IconVSR [2] + +![](images/db52a41ac806d5411c6727e9356709068505380725e0a0bfd8092b26d7392a11.jpg) +Restormer [48] +BasicVSR++ [3] + +![](images/5a696d3cdf283181e8b6ae8fc7532ae53e7c342b4526780ecb15acb543f773bc.jpg) +MPRNet [49] +Ours + +![](images/0cc95c97bb83078a52a57b7418638907170f19d8c430b36e5abdd2ae06a8fd59.jpg) +DGUNet [27] +Ground Truth + +![](images/5067a693dac14c7b2450b704644c91ff247c9f65d8ce6554187dd9f28439a511.jpg) + +![](images/4e5339b483474826997b7b09cad4c3eb22f02facd2149b83864343d00fbbdf0f.jpg) + +![](images/b862fdd9a66d2b4bd5eb0444ee47bafe6349100b8d946f910439832c8032ef20.jpg) + +![](images/804247cd4a7c4ee81a34be0a037b797e31b031cb95353b316400721b790a2261.jpg) + +![](images/510966f35b2a527de801f85a3357ccd0ee22480832308085001da67c47cbd34f.jpg) +Input Frame +S2VD [47] +Figure 14: Visual comparison on real snowstorm video. + +![](images/c74014ae602bab47ed84249d62f13b8ef6030d2722e1c604963e0cdc1cea9aac.jpg) +MPRNet [49] +IconVSR [2] + +![](images/4fcc296e79418007c6e2a86979ccebf4e00e97306100ca7f7b699c168e0bc953.jpg) +DGUNet [27] +BasicVSR++[3] + +![](images/6b358618061b3712cbd970e1b0d0a4b090cc3748307db32020005f5c45fe6b0d.jpg) +RDD [36] +Ours + +
IndexTemporal AggregationSnow-QueryContrastive LossPSNR ↑SSIM ↑LPIPS ↓
M1(basic)24.160.89900.1125
M2✓24.580.91300.0934
M3✓✓24.800.91080.9004
M4✓✓24.840.91810.0882
Ours✓✓✓25.060.92100.0842
+ +Table 6: Quantitative results of the ablation study. + +both the haze and snowflakes and betterrecover the obscured background details, while compared methods tend to main parts of haze and snowflakes. + +Ablation Study. We perform ablation study experiments to verify the temporal aggregation module, the snow-query cross-attention module, and the snow-aware contrastive loss of our network. To do so, we first built a basic model (see M1 of Table 6) by removing these three major components from our network, and then add the temporal aggregation module into M1 to construct M2. Then we add the snow-query cross-attention module into M2 to build M3, and add the snow-aware contrastive loss into M2 to build M4. Table 6 reports the PSNR, SSIM, and LPIPS results of our network and four baseline networks. Specifically, compared to M1, M2 improves the PSNR score from 24.16 + +to 24.58, the SSIM score from 0.8990 to 0.9130, and the LPIPS score from 0.1125 to 0.0934. It shows that the temporal aggregation module incurs a better video desnowing performance. Then, we find that M3 and M4 have a superior PSNR, SSIM, and LPIPS performance over M2. It demonstrates the effectiveness of the snow-aware cross-attention module or snow-aware contrastive loss for enhancing the video desnowing performance. Moreover, our method has larger PSNR and SSIM scores and a smaller LPIPS score than M3 and M4. It shows that combing three components together has the best performance of video desnowing. + +# 6. Conclusion + +The first contribution of our work is to synthesize the first high-quality video desnowing dataset, which is more realistic than previous image datasets. Then, we devise a deep learning-based framework that incorporates a snow-query temporal aggregation module and a snow-aware contrastive learning loss. Experimental results show that our network outperforms state-of-the-art methods in terms of video denosing on synthetic and real-world snowy videos. + +Acknowledgment. This work was supported by the National Natural Science Foundation of China (Grant No. 61902275), and Guangzhou Municipal Science and Technology Project (Grant No. 2023A03J0671). + +# References + +[1] Jérémie Bossu, Nicolas Hautiere, and Jean-Philippe Tarel. Rain or snow detection in image sequences through use of a histogram of orientation of streaks. International journal of computer vision, 93:348–367, 2011. +[2] Kelvin C.K. Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2021. +[3] Kelvin C.K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. BasicVSR++: Improving video superresolution with enhanced propagation and alignment. In + +IEEE Conference on Computer Vision and Pattern Recognition, 2022. +[4] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1692-1703, June 2023. +[5] Haoyu Chen, Jinjin Gu, and Zhi Zhang. Attention in attention network for image super-resolution, 2021. +[6] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 17-33. Springer, 2022. +[7] Sixiang Chen, Tian Ye, Yun Liu, and Erkang Chen. Dualformer: Hybrid self-attention transformer for efficient image restoration. arXiv preprint arXiv:2210.01069, 2022. +[8] Sixiang Chen, Tian Ye, Yun Liu, Erkang Chen, Jun Shi, and Jingchun Zhou. Snowformer: Scale-aware transformer via context interaction for single image desnowing. arXiv preprint arXiv:2208.09703, 2022. +[9] Sixiang Chen, Tian Ye, Yun Liu, Erkang Chen, Jun Shi, and Jingchun Zhou. Snowformer: Scale-aware transformer via context interaction for single image desnowing. arXiv preprint arXiv:2208.09703, 2022. +[10] Sixiang Chen, Tian Ye, Yun Liu, Taodong Liao, Jingxia Jiang, Erkang Chen, and Peng Chen. Msp-former: Multiscale projection transformer for single image desnowing. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. +[11] Wei-Ting Chen, Hao-Yu Fang, Jian-Jiun Ding, Cheng-Che Tsai, and Sy-Yen Kuo. Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 754-770. Springer, 2020. +[12] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4196-4205, 2021. +[13] Wei-Ting Chen, Zhi-Kai Huang, Cheng-Che Tsai, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17653–17662, 2022. +[14] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5374-5383, 2019. + +[15] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Image quality assessment for perceptual image restoration: A new dataset, benchmark and metric. arXiv preprint arXiv:2011.15002, 2020. +[16] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In European Conference on Computer Vision, pages 633-651. Springer, 2020. +[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. +[18] Jin-Hwan Kim, Jae-Young Sim, and Chang-Su Kim. Video deraining and desnowing using temporal correlation and low-rank matrix completion. IEEE Transactions on Image Processing, 24(9):2658-2670, 2015. +[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. +[20] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2018. +[21] Minghan Li, Xiangyong Cao, Qian Zhao, Lei Zhang, Chenqiang Gao, and Deyu Meng. Video rain/snow removal by transformed online multiscale convolutional sparse coding. arXiv preprint arXiv:1909.06148, 2019. +[22] Minghan Li, Xiangyong Cao, Qian Zhao, Lei Zhang, and Deyu Meng. Online rain/snow removal from surveillance videos. IEEE Transactions on Image Processing, 30:2029-2044, 2021. +[23] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. +[24] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent video restoration transformer with guided deformable attention. arXiv preprint arXiv:2206.02146, 2022. +[25] Yun-Fu Liu, Da-Wei Jaw, Shih-Chia Huang, and Jenq-Neng Hwang. Desnownet: Context-aware deep network for snow removal. IEEE Transactions on Image Processing, 27(6):3064-3073, 2018. +[26] Xin Luo, Yunan Zhu, Shunxin Xu, and Dong Liu. On the effectiveness of spectral discriminators for perceptual quality improvement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. +[27] Chong Mou, Qian Wang, and Jian Zhang. Deep generalized unfolding networks for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17399-17410, 2022. +[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming + +Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. +[29] Soo-Chang Pei, Yu-Tai Tsai, and Chen-Yu Lee. Removing rain and snow in a single image using saturation and visibility features. In 2014 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 1-6. IEEE, 2014. +[30] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4161-4170, 2017. +[31] Jingjing Ren, Qingqing Zheng, Yuanyuan Zhao, Xuemiao Xu, and Chen Li. Dlformer: Discrete latent transformer for video inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3511-3520, June 2022. +[32] Weihong Ren, Jiandong Tian, Zhi Han, Antoni Chan, and Yandong Tang. Video desnowing and deraining based on matrix decomposition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4210-4219, 2017. +[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. +[34] Mingxing Tan, Ruoming Pang, and Quoc V Le. Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10781-10790, 2020. +[35] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. +[36] Shuai Wang, Lei Zhu, Huazhu Fu, Jing Qin, Carola-Bibiane Schonlieb, Wei Feng, and Song Wang. Rethinking video rain streak removal: A new synthesis model and a deraining network with video rain prior. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIX, pages 565-582. Springer, 2022. +[37] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. +[38] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. +[39] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. + +[40] Zeyu Xiao, Xueyang Fu, Jie Huang, Zhen Cheng, and Zhiwei Xiong. Space-time distillation for video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2113-2122, June 2021. +[41] Zeyu Xiao, Yutong Liu, Ruisheng Gao, and Zhiwei Xiong. Cutmib: Boosting light field super-resolution via multi-view image blending. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1672-1682, June 2023. +[42] Zeyu Xiao, Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. $\text{Eva}^2$ : Event-assisted video frame interpolation via cross-modal alignment and aggregation. IEEE Transactions on Computational Imaging, 8:1145-1158, 2022. +[43] Zeyu Xiao, Zhiwei Xiong, Xueyang Fu, Dong Liu, and Zheng-Jun Zha. Space-time video super-resolution using temporal profiles. In Proceedings of the 28th ACM International Conference on Multimedia, pages 664-672, 2020. +[44] Bin Yang, Zhenhong Jia, Jie Yang, and Nikola K Kasabov. Video snow removal based on self-adaptation snow detection and patch-based gaussian mixture model. IEEE Access, 8:160188-160201, 2020. +[45] Yijun Yang, Angelica Aviles-Rivero, Huazhu Fu, Ye Liu, Weiming Wang, and Lei Zhu. Video adverse-weather-component suppression network via weather messenger and adversarial backpropagation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. +[46] Tian Ye, Sixiang Chen, Yun Liu, Yi Ye, Erkang Chen, and Yuche Li. Underwater light field retention: Neural rendering for underwater imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 488-497, 2022. +[47] Zongsheng Yue, Jianwen Xie, Qian Zhao, and Deyu Meng. Semi-supervised video deraining with dynamical rain generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 642-652, 2021. +[48] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. +[49] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. +[50] Kaihao Zhang, Rongqing Li, Yanjiang Yu, Wenhan Luo, and Changsheng Li. Deep dense multi-scale network for snow removal using semantic and depth priors. IEEE Transactions on Image Processing, 30:7419-7431, 2021. +[51] Ruofan Zhang, Jinjin Gu, Haoyu Chen, Chao Dong, Yulun Zhang, and Wenming Yang. Crafting training degradation distribution for the accuracy-generalization trade-off. arXiv preprint arXiv:2305.18107, 2023. + +[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. \ No newline at end of file diff --git a/2023/Snow Removal in Video_ A New Dataset and A Novel Method/images.zip b/2023/Snow Removal in Video_ A New Dataset and A Novel Method/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..acc45b281aae609a332691b0a0f31ebdf8d27a63 --- /dev/null +++ b/2023/Snow Removal in Video_ A New Dataset and A Novel Method/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:972ab6df3bb02ab51ca6362f24ab87a163ec7716723fb448a5fd2b36a6302ae6 +size 906934 diff --git a/2023/Snow Removal in Video_ A New Dataset and A Novel Method/layout.json b/2023/Snow Removal in Video_ A New Dataset and A Novel Method/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..95b697a1a56ff6dbf523966fe914ffd2c202dce6 --- /dev/null +++ b/2023/Snow Removal in Video_ A New Dataset and A Novel Method/layout.json @@ -0,0 +1,10991 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 107, + 103, + 486, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 103, + 486, + 120 + ], + "spans": [ + { + "bbox": [ + 107, + 103, + 486, + 120 + ], + "type": "text", + "content": "Snow Removal in Video: A New Dataset and A Novel Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 148, + 142, + 445, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 142, + 445, + 171 + ], + "spans": [ + { + "bbox": [ + 148, + 142, + 445, + 171 + ], + "type": "text", + "content": "Haoyu Chen" + }, + { + "bbox": [ + 148, + 142, + 445, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 142, + 445, + 171 + ], + "type": "text", + "content": ", Jingjing Ren" + }, + { + "bbox": [ + 148, + 142, + 445, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 142, + 445, + 171 + ], + "type": "text", + "content": ", Jinjin Gu" + }, + { + "bbox": [ + 148, + 142, + 445, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 148, + 142, + 445, + 171 + ], + "type": "text", + "content": ", Hongtao Wu" + }, + { + "bbox": [ + 148, + 142, + 445, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 142, + 445, + 171 + ], + "type": "text", + "content": ", Xuequan Lu" + }, + { + "bbox": [ + 148, + 142, + 445, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 148, + 142, + 445, + 171 + ], + "type": "text", + "content": ", Haoming Cai" + }, + { + "bbox": [ + 148, + 142, + 445, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 148, + 142, + 445, + 171 + ], + "type": "text", + "content": ", Lei Zhu" + }, + { + "bbox": [ + 148, + 142, + 445, + 171 + ], + "type": "inline_equation", + "content": "^{1,5}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 172, + 476, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 172, + 476, + 185 + ], + "spans": [ + { + "bbox": [ + 117, + 172, + 476, + 185 + ], + "type": "text", + "content": "1The Hong Kong University of Science and Technology (Guangzhou) 2The University of Sydney" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 198, + 186, + 395, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 186, + 395, + 198 + ], + "spans": [ + { + "bbox": [ + 198, + 186, + 395, + 198 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 198, + 186, + 395, + 198 + ], + "type": "text", + "content": "La Trobe University " + }, + { + "bbox": [ + 198, + 186, + 395, + 198 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 198, + 186, + 395, + 198 + ], + "type": "text", + "content": "The University of Maryland" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 195, + 201, + 399, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 201, + 399, + 213 + ], + "spans": [ + { + "bbox": [ + 195, + 201, + 399, + 213 + ], + "type": "text", + "content": "5The Hong Kong University of Science and Technology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 157, + 216, + 435, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 216, + 435, + 227 + ], + "spans": [ + { + "bbox": [ + 157, + 216, + 435, + 227 + ], + "type": "text", + "content": "Project page: https://haoyuchen.com/VideoDesnowing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 281, + 289, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 289, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 289, + 544 + ], + "type": "text", + "content": "Snowfall is a common weather phenomenon that can severely affect computer vision tasks by obscuring objects and scenes. However, existing deep learning-based snow removal methods are designed for single images only. In this paper, we target a more complex task - video snow removal, which aims to restore the clear video from the snowy video. To facilitate this task, we propose the first high-quality video dataset, which simulates realistic physical characteristics of snow and haze using a rendering engine and augmentation techniques. We also develop a deep learning framework for video snow removal. Specifically, we propose a snow-query temporal aggregation module and a snow-aware contrastive learning loss function. The module aggregates features between video frames and removes snow effectively, while the loss function helps identify and eliminate snow features. We conduct extensive experiments and demonstrate that our proposed dataset is more realistic than previous datasets, and the models trained on it achieve better performance in real-world snowing images. Our proposed method outperforms state-of-the-art video and image-based methods on both synthetic and real snowy videos." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 571, + 128, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 571, + 128, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 571, + 128, + 585 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": "Snowfall is a common weather phenomenon that can significantly affect visibility in photographs, interfering with and reducing the accuracy of computer vision tasks such as target detection [34], tracking [14], and autonomous driving [33]. Image and video restoration [39, 23, 48, 37, 46, 7, 5, 26, 40, 43, 42, 41, 45, 31] is a widely studied direction, but research on image and video snow removal is still limited. Removing snow is crucial for improving accuracy and robustness of computer vision tasks. Additionally, snow can obstruct objects, decreasing image and video quality. Snow" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 256, + 547, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 547, + 316 + ], + "type": "text", + "content": "removal is a challenging task that requires distinguishing snowflake regions from the background and restoring obscured scenes, complicated by the complex geometric shape and texture of snowflakes. Moreover, snowfall often comes with haze, which further obscures the images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 317, + 548, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 548, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 548, + 567 + ], + "type": "text", + "content": "Contrast with single image, the video provides richer information about the dynamic features of the scene, i.e., a consecutive frame sequence on the temporal dimension. Although recent image snow removal algorithms [25, 11, 12, 50, 8, 10] have achieved good results, applying these algorithms directly to video snow removal is inappropriate because image algorithms ignore temporal information in video frames and this information could be very useful to improve snow removal outcomes. As such, it is significant to conduct in-depth research and exploration for snow video data. Previous studies have explored some aspects of video desnowing, such as methods based on non-deep learning [32, 18] or online learning [22]. However, deep learning-based video desnowing remains an under-researched area. A major challenge for applying deep learning methods is the lack of high-quality video desnowing datasets. In the era of powerful deep learning, this becomes an urgent problem that needs to be addressed. In fact, the lack of appropriate datasets has significantly slowed down the progress of deep learning-based video snow removal research, which has been a bottleneck in this field." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": "To this end, we present the first high-quality annotated video dataset for video snow removal: Realistic Video DeSnowing Dataset (RVSD). RVSD contains a total of 110 pairs of videos. Each pair contains snowy and hazy videos and corresponding snow-free and haze-free ground truth videos. Unlike previous image datasets that simulate snow using Photoshop, we use a rendering engine and various augmentation techniques to generate snow with diverse and realistic physical properties. This results in more realistic and varied synthesized videos, which improve the model's performance on real-world data. To further enhance the model's performance, we also consider the effect of haze" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13211" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 110, + 148 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 110, + 148 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 110, + 148 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 110, + 148 + ], + "type": "image", + "image_path": "32dd73bf420f912e63db375ae0a0858c395f4433b54006127e7ca1c151e8b6a3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 149, + 96, + 157 + ], + "lines": [ + { + "bbox": [ + 62, + 149, + 96, + 157 + ], + "spans": [ + { + "bbox": [ + 62, + 149, + 96, + 157 + ], + "type": "text", + "content": "a.City Day" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 111, + 72, + 170, + 148 + ], + "blocks": [ + { + "bbox": [ + 111, + 72, + 170, + 148 + ], + "lines": [ + { + "bbox": [ + 111, + 72, + 170, + 148 + ], + "spans": [ + { + "bbox": [ + 111, + 72, + 170, + 148 + ], + "type": "image", + "image_path": "3de314e6c425295fe6ccad00951d1985e084e36bcc8768b4d1bd2bd0624200e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 149, + 160, + 157 + ], + "lines": [ + { + "bbox": [ + 121, + 149, + 160, + 157 + ], + "spans": [ + { + "bbox": [ + 121, + 149, + 160, + 157 + ], + "type": "text", + "content": "b. City Night" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 171, + 72, + 217, + 148 + ], + "blocks": [ + { + "bbox": [ + 171, + 72, + 217, + 148 + ], + "lines": [ + { + "bbox": [ + 171, + 72, + 217, + 148 + ], + "spans": [ + { + "bbox": [ + 171, + 72, + 217, + 148 + ], + "type": "image", + "image_path": "1fa1488917077fdbdbe1616848314e000e355dbd8777a9abe7a88c92257b702d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 149, + 215, + 157 + ], + "lines": [ + { + "bbox": [ + 186, + 149, + 215, + 157 + ], + "spans": [ + { + "bbox": [ + 186, + 149, + 215, + 157 + ], + "type": "text", + "content": "c. Nature" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 167, + 542, + 179 + ], + "lines": [ + { + "bbox": [ + 48, + 167, + 542, + 179 + ], + "spans": [ + { + "bbox": [ + 48, + 167, + 542, + 179 + ], + "type": "text", + "content": "Figure 1: Samples of the proposed dataset (left) and previous datasets (right). Refer to the supplementary to see the videos." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 242, + 72, + 302, + 148 + ], + "blocks": [ + { + "bbox": [ + 242, + 72, + 302, + 148 + ], + "lines": [ + { + "bbox": [ + 242, + 72, + 302, + 148 + ], + "spans": [ + { + "bbox": [ + 242, + 72, + 302, + 148 + ], + "type": "image", + "image_path": "f73943a08fd6883b84647a4324019f9359578b359b1cc79d6caf50dc09b3fffd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 149, + 290, + 156 + ], + "lines": [ + { + "bbox": [ + 255, + 149, + 290, + 156 + ], + "spans": [ + { + "bbox": [ + 255, + 149, + 290, + 156 + ], + "type": "text", + "content": "Snow100K" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 303, + 72, + 362, + 148 + ], + "blocks": [ + { + "bbox": [ + 303, + 72, + 362, + 148 + ], + "lines": [ + { + "bbox": [ + 303, + 72, + 362, + 148 + ], + "spans": [ + { + "bbox": [ + 303, + 72, + 362, + 148 + ], + "type": "image", + "image_path": "24ac02178277a49b6e6b62340d6cb379dd2a48329ca06bd1fe8ccc25ff642b1b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 149, + 358, + 157 + ], + "lines": [ + { + "bbox": [ + 309, + 149, + 358, + 157 + ], + "spans": [ + { + "bbox": [ + 309, + 149, + 358, + 157 + ], + "type": "text", + "content": "SnowCityScape" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 364, + 72, + 423, + 148 + ], + "blocks": [ + { + "bbox": [ + 364, + 72, + 423, + 148 + ], + "lines": [ + { + "bbox": [ + 364, + 72, + 423, + 148 + ], + "spans": [ + { + "bbox": [ + 364, + 72, + 423, + 148 + ], + "type": "image", + "image_path": "046b13f627357807135358a66d498d33867906575fd9be8dd46cf19167597a18.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 149, + 417, + 157 + ], + "lines": [ + { + "bbox": [ + 369, + 149, + 417, + 157 + ], + "spans": [ + { + "bbox": [ + 369, + 149, + 417, + 157 + ], + "type": "text", + "content": "SnowCityScape" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 425, + 72, + 484, + 148 + ], + "blocks": [ + { + "bbox": [ + 425, + 72, + 484, + 148 + ], + "lines": [ + { + "bbox": [ + 425, + 72, + 484, + 148 + ], + "spans": [ + { + "bbox": [ + 425, + 72, + 484, + 148 + ], + "type": "image", + "image_path": "fb8ddd6cb5f9548b7b829f9bd117c025f9db7303d7fa749ffd8ae4f7319d8bdc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 149, + 463, + 156 + ], + "lines": [ + { + "bbox": [ + 444, + 149, + 463, + 156 + ], + "spans": [ + { + "bbox": [ + 444, + 149, + 463, + 156 + ], + "type": "text", + "content": "SRRS" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 485, + 72, + 544, + 148 + ], + "blocks": [ + { + "bbox": [ + 485, + 72, + 544, + 148 + ], + "lines": [ + { + "bbox": [ + 485, + 72, + 544, + 148 + ], + "spans": [ + { + "bbox": [ + 485, + 72, + 544, + 148 + ], + "type": "image", + "image_path": "ddeb05d42fb0086e2a5e834f424fbbb94fdd1d106d74b0c01a1dd8141357c18b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 506, + 149, + 522, + 156 + ], + "lines": [ + { + "bbox": [ + 506, + 149, + 522, + 156 + ], + "spans": [ + { + "bbox": [ + 506, + 149, + 522, + 156 + ], + "type": "text", + "content": "CSD" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 201, + 287, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 287, + 391 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 287, + 391 + ], + "type": "text", + "content": "in snowy conditions. Previous methods for fog synthesis relied on mathematical models with limitations, such as requiring accurate depth information and assuming fixed and uniform haze characteristics. However, real-world snowy conditions involve dynamic and heterogeneous haze. To overcome these limitations, we propose a new haze model that uses Unreal Engine 5 to render storm haze and obtain a haze layer. We then fuse the haze layer with the input video after data augmentation to produce more diverse and realistic haze, better simulating real snowy conditions. We conduct experiments to show that our proposed dataset has more-realistic snow scenes than existing datasets, and models trained on our dataset achieve better snow removal performance on real snowy images. Therefore, our dataset can facilitate the training and evaluation of video snow removal algorithms, and advance video snow removal research." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 392, + 287, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 287, + 630 + ], + "type": "text", + "content": "Trivially applying the video processing methods on our dataset may result in suboptimal since they do not consider snow and haze properties. Moreover, there are no existing deep learning-based methods for video snow removal. Therefore, we propose the first deep learning-based framework in this work. Our framework consists of three main steps. First, we perform preliminary snow removal on single frames to reduce the effect of snow occlusion on optical flow estimation and video feature alignment. Second, we fuse the video frames and perceive and remove the snow. We introduce a novel \"snow-query temporal aggregation module\", which aggregates features across video frames while perceiving and removing snow. Third, we design a new \"snow-aware contrastive loss\", which leverages the prior knowledge that snowflakes have different shapes, motion directions, and other characteristics in different videos. This loss function helps identify and remove snow features more accurately. We conduct extensive experiments to show that our proposed method outperforms the state-of-the-art methods. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 58, + 635, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 58, + 635, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 635, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 58, + 635, + 287, + 694 + ], + "type": "text", + "content": "- We propose the first high-quality video dataset (RVSD) for video snow removal. We use the Unreal Engine and various augmentation techniques to produce snow and haze with diverse and realistic physical characteristics." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 58, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 701, + 287, + 713 + ], + "type": "text", + "content": "- Comprehensive experiments show that the visual effect" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 324, + 201, + 545, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 201, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 324, + 201, + 545, + 248 + ], + "type": "text", + "content": "of our proposed snow removal dataset is more realistic than previous datasets, and the models trained on our proposed dataset can achieve better performance in real-world snowy images." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 257, + 545, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 257, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 317, + 257, + 545, + 318 + ], + "type": "text", + "content": "- A deep learning-based video snow removal framework is proposed. The snow-aware temporal aggregation module and snow-aware contrastive learning loss are introduced, to effectively remove snow. Experiments prove that our method enables the best performance." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 337, + 392, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 337, + 392, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 337, + 392, + 350 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 357, + 422, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 357, + 422, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 357, + 422, + 369 + ], + "type": "text", + "content": "2.1. Desnowing Datasets" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 375, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 545, + 615 + ], + "type": "text", + "content": "There are several single-image snow removal datasets available. Snow100K [25] first used PhotoShop to synthesize snow. In order to model different types of snow, SnowKITTI2012 [50] and SnowCityScapes [50] contains three levels of snow, including light, medium, and heavy snow. However, these datasets do not take into account the veiling effect of snowy weather, the models trained on these datasets cannot remove haze and have limited practical performance in real snowy scenarios. SRRS [11] firstly simulated the veiling effect inspired by the Koschmieder model, and then used PhotoShop to synthesize the snow streaks. CSD [12] added Gaussian blur to the above two steps to better simulate the real snow image. Nearly all methods use PhotoShop to synthesize snow. As a result, the shape and style of snow are relatively homogeneous, which substantially limits the model's performance on realistic and complex snowing scenes. A quick comparison of these datasets can be found in Table 1. To the best of our knowledge, there is no dataset for video desnowing, which significantly hinders the research of video desnowing." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 623, + 411, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 411, + 636 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 411, + 636 + ], + "type": "text", + "content": "2.2. Image Desnowing" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Early image snow removal methods tend to perform snow removal based on physical priors. [1] performed snow detection and removal by calculating the histogram of orientations of snow streaks. [29] used features on saturation and visibility to remove rain and snow from a single image. In recent years, deep learning has refreshed the snow removal" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13212" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 394, + 141 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 394, + 141 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 394, + 141 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 394, + 141 + ], + "type": "table", + "html": "
DatasetVenueType# imagesResolutionSnow synthesisStatic HazeDynamic HazeIlluminationUser Study (1 - 10)
Snow100K [25]TIP 2018image100K≤ 640 × 640PhotoshopXXX3.98
SnowCityScapes[50]TIP 2021image2K+2K512 × 256PhotoshopXXX4.18
SnowKITTI2012 [50]TIP 2021image1.5K+1K884 × 256PhotoshopXXX3.72
SRRS [11]ECCV 2020image15K640 × 480Photoshop✓XX4.20
CSD [12]ICCV 2021image10K640 × 480Photoshop✓XX4.67
Ours-video-480p - 4KUE Rendering✓✓✓6.94
", + "image_path": "e643d0e5be3c375b275f841562b0959e3b0427755c8f4db25889f6da86e7306a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 402, + 71, + 545, + 141 + ], + "blocks": [ + { + "bbox": [ + 107, + 143, + 332, + 156 + ], + "lines": [ + { + "bbox": [ + 107, + 143, + 332, + 156 + ], + "spans": [ + { + "bbox": [ + 107, + 143, + 332, + 156 + ], + "type": "text", + "content": "Table 1: Comparison with previous desnowing datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 402, + 71, + 545, + 141 + ], + "lines": [ + { + "bbox": [ + 402, + 71, + 545, + 141 + ], + "spans": [ + { + "bbox": [ + 402, + 71, + 545, + 141 + ], + "type": "table", + "html": "
Scenes# videos# frames# frames (512×512)
city day405,12131,020
city night202,37823,878
nature433,52632,526
other74073,012
In total11011,43290,436
", + "image_path": "fcb9466f644126ffda0b6690b972a04a50625694f4582b9d37232f89ea5a305a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 405, + 143, + 539, + 154 + ], + "lines": [ + { + "bbox": [ + 405, + 143, + 539, + 154 + ], + "spans": [ + { + "bbox": [ + 405, + 143, + 539, + 154 + ], + "type": "text", + "content": "Table 2: Statistics of our dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 176, + 289, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 176, + 289, + 356 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 289, + 356 + ], + "type": "text", + "content": "task. The first deep learning based snow removal method DesnowNet [25] proposed a multi-stage, multi-scale design to remove translucent and opaque snow streaks. JSTASR [11] proposed a joint size and transparency-aware snow removal model that can classify snow particles based on their size and remove snow. HDCW-Net [12] used a hierarchical decomposition paradigm in which a dual-tree wavelet transform and a wavelet loss are used. It also proposed a discriminative feature for snow removal called contradict channel. DDMSNet [50] introduced semantic and depth information to learn semantic-aware and geometry-aware representations for snow removal. SnowFormer [9] used a vision transformer architecture that fully combines local and global information, and obtains superior results on multiple datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 364, + 151, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 364, + 151, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 364, + 151, + 376 + ], + "type": "text", + "content": "2.3. Video Desnowing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 382, + 288, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 288, + 586 + ], + "type": "text", + "content": "Before the success of deep learning-based methods, traditional computer vision techniques were used for video-based snow removal. Ren et al. [32] utilized the low-rank assumption of the background to separate sparse and dense snow, and addressed heavy snow in dynamic scenes. Kim et al. [18] considered global and local motion, as well as snowflakes of various sizes, in their snowflake removal algorithm with low-rank matrix completion. Due to the rapid improvement of deep learning models and their ability to study complex patterns, some researchers have proposed the use of deep learning for video-based snow removal. [44] utilized self-adaptive snow detection and a patch-based Gaussian mixture model, adept at removing both sparse and dense snowflakes from videos. Li et al. [21, 22] proposed a method for dynamic backgrounds where snow was encoded and removed using an online multi-scale convolutional sparse coding model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 597, + 241, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 241, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 241, + 611 + ], + "type": "text", + "content": "3. Realistic Video Desnowing Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "Previous snow and haze synthesis methods suffer from a lack of variation and poor physical accuracy. In contrast, our approach employs a 3D rendering engine and a range of augmentations, together with a more realistic fusion method to create snow and haze with more diverse, complex, and accurate physical properties. As a result, the synthesized video is more realistic and diverse, enabling our model to enjoy better generalization ability to real-world data. We" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 176, + 443, + 269 + ], + "blocks": [ + { + "bbox": [ + 306, + 176, + 443, + 269 + ], + "lines": [ + { + "bbox": [ + 306, + 176, + 443, + 269 + ], + "spans": [ + { + "bbox": [ + 306, + 176, + 443, + 269 + ], + "type": "image", + "image_path": "a2523a3842090bd29425f3304a23cf2c779414b3d56dd64644f4a7d147092538.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 175, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 447, + 175, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 447, + 175, + 547, + 258 + ], + "type": "text", + "content": "Figure 2: Box plots of the user study on the visual realism perception scores of different datasets. Higher scores represent better perceptual realism." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 290, + 435, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 290, + 435, + 301 + ], + "spans": [ + { + "bbox": [ + 306, + 290, + 435, + 301 + ], + "type": "text", + "content": "demonstrate this in Section 5.1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 309, + 452, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 309, + 452, + 322 + ], + "spans": [ + { + "bbox": [ + 306, + 309, + 452, + 322 + ], + "type": "text", + "content": "3.1. Collection of Clean Videos" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 328, + 547, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 547, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 547, + 508 + ], + "type": "text", + "content": "Unlike previous work that only considered daytime video, we collect both daytime and nighttime videos in our dataset, which is unlike previous works. We apply different add-snow methods for daytime or nighttime videos to enhance the realism of the generated results. We will discuss these methods in the next section. Our dataset covers a variety of scenes such as buildings, streets, nature scenes, close shots, distant views, overhead and elevation angles to simulate the rich scenes of real snow videos. As shown in Table 2, we have collected 110 videos in total, among which 80 videos are used for training and 30 videos for testing. To accommodate the diverse resolutions of online videos, our video resolutions range from " + }, + { + "bbox": [ + 304, + 328, + 547, + 508 + ], + "type": "inline_equation", + "content": "480\\mathrm{p}" + }, + { + "bbox": [ + 304, + 328, + 547, + 508 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 328, + 547, + 508 + ], + "type": "inline_equation", + "content": "4\\mathrm{k}" + }, + { + "bbox": [ + 304, + 328, + 547, + 508 + ], + "type": "text", + "content": ". This enables the model to produce better results on videos with more sources and resolutions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 515, + 405, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 405, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 405, + 529 + ], + "type": "text", + "content": "3.2. Haze Rendering" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": "Previous studies have employed a fog synthesis method based on the atmospheric scattering model [20]: " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "I_{haze}(x) = J(x)t(x) + A(1 - t(x))" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "I_{haze}(x)" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " is hazy image, " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "J(x)" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " is the haze-free image, " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " denotes the global atmospheric light, and " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "t(x)" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " is the transmission matrix. This model has both strengths and weaknesses. One of the strengths is that the formula is intuitive and concise, making it easy to embed into network design. For limitations, it relies on depth information, and the generated haze exhibits a relatively uniform pattern. More importantly, there is a fundamental difference between snow-generated haze on a snowy day and conventional haze. Snow haze is often dynamic and follows the movement of snow, as illustrated in Figure 4. Therefore, using traditional haze synthesis methods that assume a homogeneous haze pattern may" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13213" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 70, + 517, + 202 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 517, + 202 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 517, + 202 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 517, + 202 + ], + "type": "image", + "image_path": "1f22993164006276ad80b34809152f86f08a836fbb8fadbce77595690e2a085b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 547, + 262 + ], + "type": "text", + "content": "Figure 3: Snow-free videos with rich scenes are selected first. Then, we use two haze models to generate depth-related static haze and Unreal Engine (UE) rendered dynamic haze that is more realistic. After that, we use the UE to render snow layers with realistic physical characteristics. Further the snow layers are dynamically blended according to the background time after complex augmentation. These complex operations make the results more realistic with diverse features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 281, + 287, + 322 + ], + "blocks": [ + { + "bbox": [ + 48, + 281, + 287, + 322 + ], + "lines": [ + { + "bbox": [ + 48, + 281, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 281, + 287, + 322 + ], + "type": "image", + "image_path": "09004c2c7292247fd322e48992d8db2ba9791ada351a7a483efeb05d021d86bd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 331, + 249, + 342 + ], + "lines": [ + { + "bbox": [ + 84, + 331, + 249, + 342 + ], + "spans": [ + { + "bbox": [ + 84, + 331, + 249, + 342 + ], + "type": "text", + "content": "Figure 4: Dynamic haze in a snowstorm." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 48, + 352, + 276, + 407 + ], + "blocks": [ + { + "bbox": [ + 48, + 352, + 276, + 407 + ], + "lines": [ + { + "bbox": [ + 48, + 352, + 276, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 352, + 276, + 407 + ], + "type": "image", + "image_path": "5d26b40a8818d9ba3c7ce85df13c1d03803c0a1fab35b2f527dc23ff81535c56.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 415, + 271, + 427 + ], + "lines": [ + { + "bbox": [ + 63, + 415, + 271, + 427 + ], + "spans": [ + { + "bbox": [ + 63, + 415, + 271, + 427 + ], + "type": "text", + "content": "Figure 5: Fusion weight for daytime and nighttime." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 447, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 287, + 577 + ], + "type": "text", + "content": "not be accurate in simulating the dynamic and varying snow haze that occurs during a snowstorm. We propose a novel haze model that leverages the capabilities of the Unreal Engine to render the haze of snowstorms and obtain different haze layers " + }, + { + "bbox": [ + 46, + 447, + 287, + 577 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 447, + 287, + 577 + ], + "type": "text", + "content": ". The dynamic haze of the storms is then integrated with the input video following data augmentation. As a result, our approach produces a more diverse and realistic haze that more closely resembles real snow days and nights. This enables the models trained on our data to handle more diverse and dynamic snow haze. Our composite haze model is defined as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 578, + 287, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 578, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 61, + 578, + 287, + 592 + ], + "type": "interline_equation", + "content": "I _ {h a z e} (x) = J (x) t (x) + A (1 - t (x)) + \\operatorname {A u g} (H), \\tag {1}", + "image_path": "e82c14402ea874e00a7e047fffb660070c9eb2122610b9750a53ee0b89b08b79.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 593, + 206, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 593, + 206, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 593, + 206, + 604 + ], + "type": "text", + "content": "where Aug denotes data augmentation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 611, + 147, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 147, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 147, + 624 + ], + "type": "text", + "content": "3.3. Snow Rendering" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "In previous snow production work [11, 25, 50, 12], snowflakes and snow streaks were synthesized using Photoshop, resulting in relatively homogeneous patterns. In contrast, our work employs Unreal Engine to synthesize snow. As a 3D computer graphics engine, Unreal Engine enables the creation of snow with more realistic and complex physical features, including snowflake flip, spatial changes from" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 280, + 547, + 375 + ], + "blocks": [ + { + "bbox": [ + 307, + 280, + 547, + 375 + ], + "lines": [ + { + "bbox": [ + 307, + 280, + 547, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 280, + 547, + 375 + ], + "type": "image", + "image_path": "0bef36974df88c23b1126bf9cfeecb19c4d988fe00ff8557a5580b026f82fd95.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 382, + 545, + 407 + ], + "lines": [ + { + "bbox": [ + 305, + 382, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 382, + 545, + 407 + ], + "type": "text", + "content": "Figure 6: The visibility of daytime snow and nighttime snow varies depending on the background." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 445, + 545, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 684 + ], + "type": "text", + "content": "far to near, and perspective changes, among others. We use various 3D snow models to manipulate the physical properties of the snow, resulting in diverse and complex shapes that more closely resemble real-life snow. More details of the features are shown in Table 3. Once the snow layer " + }, + { + "bbox": [ + 304, + 445, + 545, + 684 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 445, + 545, + 684 + ], + "type": "text", + "content": " is generated, it needs to be blended with the clean video. We augment the snow layer by adjusting the color temperature, brightness, contrast, and applying random affine transformations. Then we add different levels of blur, such as Gaussian blur and motion blur. Finally, we consider the temporal properties of the video background based on time of day. During the day, especially when the sky is in the background, snowflakes tend to blend with the bright sky, resulting in snowflakes that are not easily distinguishable. At night, snowflakes are more visible in bright places, such as those illuminated by headlights or streetlights, while they tend to be less visible in places without light. We show some real samples in Figure 6. Based on this property, we utilize Eq. (2) to perform the integration of snow layer and video background." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 330, + 685, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 685, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 330, + 685, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\alpha (x) = \\left\\{ \\begin{array}{c} \\sigma (- (V (x) - \\gamma) \\times \\beta), \\text {i f d a y t i m e} \\\\ \\sigma ((V (x) - \\gamma) \\times \\beta), \\text {i f n i g h t t i m e} \\end{array} \\right. \\tag {2}", + "image_path": "0f675b2f2c657efb5d944b1ab1bfb77ee6925efe37406bb630ab57ed9ebd8412.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13214" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 287, + 182 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 287, + 182 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 287, + 182 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 287, + 182 + ], + "type": "table", + "html": "
FeatureDetails
Snow Speedby setting gravity intensity, wind speed, air friction coefficient and other parameters, snow can move at different speeds
Snow Intensityby setting the number of particles in the snow model, our dataset contains snow of various degrees.
Snow Typeby employing various snow models and adjusting various parameters, combined with the Unreal Engine to impart realistic physical features to the snowflakes (such as turbulence, rotation, etc.), the resulting snow has a highly diverse range of shapes, sizes, and characteristics.
Motion directionhead-on, back-on, left-facing, right-facing, and many other angles
Illuminationlights have different intensities and colors, such as different warm and cold tones
Veiling effectdepth-related static haze, as well as rendered dynamic complex haze.
Scenescity, street, day & night, nature
", + "image_path": "182db053e1f19820c50fd29cda071aa1d1c2a70d7236e73791892e7a9bf9b54f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 82, + 200, + 251, + 212 + ], + "lines": [ + { + "bbox": [ + 82, + 200, + 251, + 212 + ], + "spans": [ + { + "bbox": [ + 82, + 200, + 251, + 212 + ], + "type": "text", + "content": "Table 3: Features of our proposed dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 232, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 287, + 339 + ], + "type": "text", + "content": "Specifically, we first convert the RGB image to the HSV color space, where " + }, + { + "bbox": [ + 46, + 232, + 287, + 339 + ], + "type": "inline_equation", + "content": "V = \\max(R, G, B)" + }, + { + "bbox": [ + 46, + 232, + 287, + 339 + ], + "type": "text", + "content": ", represents the largest color component. We obtain a corresponding blending weight based on the pixel's brightness. " + }, + { + "bbox": [ + 46, + 232, + 287, + 339 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 232, + 287, + 339 + ], + "type": "text", + "content": " denotes the value of the V channel in the HSV space, which is normalized to the range of 0 to 1. " + }, + { + "bbox": [ + 46, + 232, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 232, + 287, + 339 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 232, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 232, + 287, + 339 + ], + "type": "text", + "content": " are two parameters artificially adjusted based on the specific video, and " + }, + { + "bbox": [ + 46, + 232, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 232, + 287, + 339 + ], + "type": "text", + "content": " represents the softmax function. Figure 5 shows some examples. Our dynamic video snow model is defined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 341, + 287, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 341, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 89, + 341, + 287, + 354 + ], + "type": "interline_equation", + "content": "Z (x) = I _ {h a z e} (x) + \\alpha (x) \\operatorname {A u g} (S (x)). \\tag {3}", + "image_path": "2677ee3409a847b6e78a0de51ca301b54ada01ccf5747bb0f8ff7c5064c42831.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 354, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 376 + ], + "type": "text", + "content": "where Aug denotes data augmentation. " + }, + { + "bbox": [ + 46, + 354, + 287, + 376 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 354, + 287, + 376 + ], + "type": "text", + "content": " is the final output with both snow and haze." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 388, + 264, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 264, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 264, + 402 + ], + "type": "text", + "content": "4. Snow-Aware Video Desnowing Network" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 407, + 173, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 173, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 173, + 419 + ], + "type": "text", + "content": "4.1. Network Architecture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 426, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 652 + ], + "type": "text", + "content": "Our network architecture is composed of three main modules: the encoder, the snow-query temporal aggregation module, and the decoder, as shown in Figure 7. To mitigate the impact of snow occlusion on video frame alignment and fusion, the encoder takes single-frame images as input and generates coarse features " + }, + { + "bbox": [ + 46, + 426, + 287, + 652 + ], + "type": "inline_equation", + "content": "F^{C}" + }, + { + "bbox": [ + 46, + 426, + 287, + 652 + ], + "type": "text", + "content": ", which are then transformed into coarse desnowed images by the recovery module. However, residual snow may still remain in the video frames due to the lack of utilization of video information. Therefore, it is imperative to incorporate multi-frame video information to further remove residual snow. To achieve this, the Snow-query Temporal Aggregation Module takes in continuous video frames and the features obtained by the encoder. This module aligns and fuses video features, and can perceive the location of snow, which facilitates more effective snow removal. Then this module outputs continuous refined video features " + }, + { + "bbox": [ + 46, + 426, + 287, + 652 + ], + "type": "inline_equation", + "content": "F^{F}" + }, + { + "bbox": [ + 46, + 426, + 287, + 652 + ], + "type": "text", + "content": ", which are fed into the decoder. Finally, the recovery module is employed to obtain the final desnowed video." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Recovery Module. To simultaneously remove snow and snow-induced haze, we adopt the atmospheric scattering model to decompose images containing both snow and haze into three parts: snow feature " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": ", the global atmospheric light " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": ", and transmission matrix " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": ". We perform" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "text", + "content": "this operation in the feature space of the image. Here, " + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{N \\times C \\times H \\times W}" + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{N \\times 1 \\times H \\times W}" + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "T \\in \\mathbb{R}^{N \\times C \\times H \\times W}" + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "text", + "content": " is the number of frames, " + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "text", + "content": " is the number of feature channel, " + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "text", + "content": " are the height and width. The recovery process can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 331, + 132, + 545, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 132, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 331, + 132, + 545, + 158 + ], + "type": "interline_equation", + "content": "F (x) = \\frac {E (I (x)) - S (x) - (1 - T (x)) A (x)}{T (x)}, \\tag {4}", + "image_path": "dbfd48c41e5c3d9b6929bff1eafce013e339f397a6608674484ca36978e86d5d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 333, + 159, + 411, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 159, + 411, + 172 + ], + "spans": [ + { + "bbox": [ + 333, + 159, + 411, + 172 + ], + "type": "interline_equation", + "content": "J (x) = M (F (x)),", + "image_path": "9bb9f21ee42d7c68eb63295ef7aefa56283e6a3730227863fcc23fef24e4e468.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 173, + 545, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 173, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 545, + 220 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 173, + 545, + 220 + ], + "type": "inline_equation", + "content": "I(x)" + }, + { + "bbox": [ + 304, + 173, + 545, + 220 + ], + "type": "text", + "content": " is the input snowy image, " + }, + { + "bbox": [ + 304, + 173, + 545, + 220 + ], + "type": "inline_equation", + "content": "F(x)" + }, + { + "bbox": [ + 304, + 173, + 545, + 220 + ], + "type": "text", + "content": " is the output desnowed feature, " + }, + { + "bbox": [ + 304, + 173, + 545, + 220 + ], + "type": "inline_equation", + "content": "E(\\cdot)" + }, + { + "bbox": [ + 304, + 173, + 545, + 220 + ], + "type": "text", + "content": " is the feature extractor, " + }, + { + "bbox": [ + 304, + 173, + 545, + 220 + ], + "type": "inline_equation", + "content": "M(\\cdot)" + }, + { + "bbox": [ + 304, + 173, + 545, + 220 + ], + "type": "text", + "content": " is a convolution layer that maps the feature into image space, and " + }, + { + "bbox": [ + 304, + 173, + 545, + 220 + ], + "type": "inline_equation", + "content": "J(x)" + }, + { + "bbox": [ + 304, + 173, + 545, + 220 + ], + "type": "text", + "content": " is the snow-free image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 221, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 221, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 545, + 436 + ], + "type": "text", + "content": "Coarse Desnowing of Single Video Frames Prior to Alignment. Video restoration networks frequently employ optical flow or deformable convolution to integrate information from multiple frames for feature extraction. However, this is not entirely suitable for video snow removal owing to the significant masking of the video background by snowflakes and snow lines, which disturb temporal feature aggregation. The position discrepancies between snowflakes and snow lines in adjacent frames, especially in heavily snowing videos, pose a challenge to alignment methods in estimating accurate motion parameters. Thus, coarse desnowing of single frames before temporal feature attending is crucial for effectively integrating information between frames in video snow removal later on. This step does not cause any information loss for the further step. The reason is that the desnowing inputs of the further step are the features of original snowy video frames, instead of the desnowed images of the first step." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 437, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 555 + ], + "type": "text", + "content": "Figure 9 shows the results of optical flow computation on video frames with and without snow. Optical flow computation on snow-covered video frames shows significant errors, which may mislead the alignment of video frame features. Our idea is to obtain coarse desnowed frames before computing optical flows. Specifically, the encoder takes a single snowy frame as input and outputs coarse features " + }, + { + "bbox": [ + 304, + 437, + 545, + 555 + ], + "type": "inline_equation", + "content": "S^C" + }, + { + "bbox": [ + 304, + 437, + 545, + 555 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 437, + 545, + 555 + ], + "type": "inline_equation", + "content": "A^C" + }, + { + "bbox": [ + 304, + 437, + 545, + 555 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 437, + 545, + 555 + ], + "type": "inline_equation", + "content": "T^C" + }, + { + "bbox": [ + 304, + 437, + 545, + 555 + ], + "type": "text", + "content": ". The recovery module then restores the coarse desnowed frame feature " + }, + { + "bbox": [ + 304, + 437, + 545, + 555 + ], + "type": "inline_equation", + "content": "F^C" + }, + { + "bbox": [ + 304, + 437, + 545, + 555 + ], + "type": "text", + "content": " and its corresponding desnowed frame " + }, + { + "bbox": [ + 304, + 437, + 545, + 555 + ], + "type": "inline_equation", + "content": "x^C" + }, + { + "bbox": [ + 304, + 437, + 545, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 563, + 495, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 495, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 495, + 576 + ], + "type": "text", + "content": "4.2. Snow-Query Temporal Aggregation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Temporal Aggregation Module. To fully utilize the features between video frames, we use bidirectional propagation to independently propagate the frames' features forward and backward in time dimension. This technique enables the information to flow in both directions, which can help capture more spatial-temporal context and improve the model's accuracy. Given coarse desnowed frame " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{i}^{C}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " and its neighboring frame " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{i-1}^{C}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{i+1}^{C}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", we use SpyNet [30] to estimate the optical flows " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "O_{t+1 \\to t}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "O_{t-1 \\to t}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ". To obtain the corresponding features propagated from the neighboring frames, which are denoted as " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "h_{t-1}^{f}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "h_{t+1}^{b}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13215" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 545, + 218 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 545, + 218 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 545, + 218 + ], + "type": "image", + "image_path": "6a3b07189f814bc2aeddae9f4f5fcb1a5a2b7267d3a3e121ca06d6d4f6d9e0ad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 229, + 548, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 548, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 548, + 291 + ], + "type": "text", + "content": "Figure 7: Snow-Aware Video Desnowing Network (SVDNet). The architecture is composed of three modules: the encoder, the snow-query temporal aggregation module, and the decoder. To avoid the effect of snow occlusion on video alignment, the encoder first uses a single frame to generate coarse feature " + }, + { + "bbox": [ + 46, + 229, + 548, + 291 + ], + "type": "inline_equation", + "content": "F^C" + }, + { + "bbox": [ + 46, + 229, + 548, + 291 + ], + "type": "text", + "content": " and desnowed image " + }, + { + "bbox": [ + 46, + 229, + 548, + 291 + ], + "type": "inline_equation", + "content": "x^C" + }, + { + "bbox": [ + 46, + 229, + 548, + 291 + ], + "type": "text", + "content": ". Then the Snow-query Temporal Aggregation Module aligns and fuses video features, detects snow, and outputs continuous fine video features " + }, + { + "bbox": [ + 46, + 229, + 548, + 291 + ], + "type": "inline_equation", + "content": "F^F" + }, + { + "bbox": [ + 46, + 229, + 548, + 291 + ], + "type": "text", + "content": ". These features are then fed into the decoder, and the recovery module is used to obtain the final desnowed video." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 308, + 289, + 488 + ], + "blocks": [ + { + "bbox": [ + 48, + 308, + 289, + 488 + ], + "lines": [ + { + "bbox": [ + 48, + 308, + 289, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 289, + 488 + ], + "type": "image", + "image_path": "05e77ff5934c7740a3aac6cf61c8aac096cdf12166ca21dfe460efb5c3c72a24.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 495, + 278, + 507 + ], + "lines": [ + { + "bbox": [ + 55, + 495, + 278, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 278, + 507 + ], + "type": "text", + "content": "Figure 8: Snow-Query Temporal Aggregation Module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 535, + 240, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 535, + 240, + 550 + ], + "spans": [ + { + "bbox": [ + 91, + 535, + 240, + 550 + ], + "type": "interline_equation", + "content": "h _ {t} ^ {b} = H _ {b} \\left(F _ {t} ^ {C}, O _ {t + 1 \\rightarrow t}, h _ {t + 1} ^ {b}, T _ {t} ^ {C}\\right),", + "image_path": "1ec54184120cdb9c68622909f4a8123c111d4a4b52e7e6ea82afa87252b30e4c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 549, + 287, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 549, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 91, + 549, + 287, + 572 + ], + "type": "interline_equation", + "content": "h _ {t} ^ {f} = H _ {f} \\left(F _ {t} ^ {C}, O _ {t - 1 \\rightarrow t}, h _ {t - 1} ^ {f}, T _ {t} ^ {C}\\right), \\tag {5}", + "image_path": "88f3fa257e105b17c66d035c45e794abdfd9a1ca542ae46155c91f3dfd619ee9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 575, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 287, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 575, + 287, + 598 + ], + "type": "inline_equation", + "content": "H_{b}" + }, + { + "bbox": [ + 47, + 575, + 287, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 575, + 287, + 598 + ], + "type": "inline_equation", + "content": "H_{f}" + }, + { + "bbox": [ + 47, + 575, + 287, + 598 + ], + "type": "text", + "content": " denote the backward and forward propagation branches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 80, + 599, + 204, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 599, + 204, + 615 + ], + "spans": [ + { + "bbox": [ + 80, + 599, + 204, + 615 + ], + "type": "interline_equation", + "content": "w _ {t} ^ {\\{b, f \\}} = W \\left(h _ {t \\pm 1}, O _ {t \\pm 1 \\rightarrow t}\\right),", + "image_path": "fcb13755b4764f03190e803ab0709108024b853f6e40fa65362057b84b3df88f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 82, + 616, + 287, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 616, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 82, + 616, + 287, + 638 + ], + "type": "interline_equation", + "content": "h _ {t} ^ {\\{b, f \\}} = \\operatorname {M S A} \\left(\\operatorname {c a t} \\left(w _ {t} ^ {\\{b, f \\}}, F _ {t} ^ {C}, T _ {t} ^ {C}\\right)\\right), \\tag {6}", + "image_path": "33c1f8ee053f53d0908d2edddcf28c4fc1b82ff6831f31283ac8fbe7bcad88f5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 97, + 639, + 172, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 639, + 172, + 654 + ], + "spans": [ + { + "bbox": [ + 97, + 639, + 172, + 654 + ], + "type": "interline_equation", + "content": "h _ {t} = \\operatorname {c a t} \\left(h _ {t} ^ {f}, h _ {t} ^ {b}\\right),", + "image_path": "713f6195d76264dffe21ff82df78fd1d1d25890fd7033a487a94a25e6d12cf1f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 654, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 654, + 287, + 677 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 47, + 654, + 287, + 677 + ], + "type": "text", + "content": " denotes feature warping, and MSA denotes multi-head self-attention." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "Snow-Query Cross-Attention Module. The snow features for the same video are always similar. Thus, we can further use the existing snow features to match and guide the" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 308, + 308, + 547, + 383 + ], + "blocks": [ + { + "bbox": [ + 308, + 308, + 547, + 383 + ], + "lines": [ + { + "bbox": [ + 308, + 308, + 547, + 383 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 547, + 383 + ], + "type": "image", + "image_path": "8593678d02f16642073ff64ee6d7f42dde3dd45f8b3393435830af8e92d01d11.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 389, + 544, + 402 + ], + "lines": [ + { + "bbox": [ + 308, + 389, + 544, + 402 + ], + "spans": [ + { + "bbox": [ + 308, + 389, + 544, + 402 + ], + "type": "text", + "content": "Figure 9: Incorrect estimation of optical flow due to snow." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 418, + 545, + 524 + ], + "blocks": [ + { + "bbox": [ + 307, + 418, + 545, + 524 + ], + "lines": [ + { + "bbox": [ + 307, + 418, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 307, + 418, + 545, + 524 + ], + "type": "image", + "image_path": "fb7bef89128f082c4e886bf104a14bd026a7c75354e6631a2f4194bdf3635194.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 530, + 547, + 579 + ], + "lines": [ + { + "bbox": [ + 305, + 530, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 305, + 530, + 547, + 579 + ], + "type": "text", + "content": "Figure 10: t-SNE visualization of features learned with/without the contrastive loss. Our contrastive loss function effectively distinguishes different types of snow while without it, their feature distributions highly overlap." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "network to find the remaining snow. We show this module in Figure 8. First, we perform patch embedding for " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " and corresponding snow features " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "S_t^C" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": ". We consider each patch of size " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "2 \\times 2 \\times C" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " as a token. Then, a linear embedding layer is applied to project the features of each token to an arbitrary dimension. The key design is using snow features as queries. The snow features can help find the residual snow regions in the video feature. Specifically, the queries obtained from the snow feature and the key obtained from the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13216" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 69, + 547, + 198 + ], + "blocks": [ + { + "bbox": [ + 49, + 69, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 49, + 69, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 49, + 69, + 547, + 198 + ], + "type": "image", + "image_path": "bf277ede75ba1f761a9a6662bb5540a85939b4183a51af12f8679712d9609633.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 204, + 547, + 229 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 547, + 229 + ], + "type": "text", + "content": "Figure 11: Visual results on the real-world images. The same model is trained separately on different datasets for image snow removal. The model trained on our proposed dataset achieves the best results for snow and haze removal on real images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 249, + 287, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 249, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 287, + 297 + ], + "type": "text", + "content": "video feature are multiplied to obtain a correlation matrix. The coordinates with high correlation represent the residual snow in the video, which enables the network to locate the residual snow and further remove it." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 296, + 287, + 330 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 287, + 330 + ], + "type": "interline_equation", + "content": "\\operatorname {C r o s s A t t n} \\left(\\mathcal {Q} _ {s}, \\mathcal {K} _ {v}, \\mathcal {V} _ {v}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {Q} _ {s} \\mathcal {K} _ {v}}{\\sqrt {d}} + p\\right) \\mathcal {V} _ {v}, \\tag {7}", + "image_path": "51a147a3cbcec08250a815f3d87eee4c6f71ce3429ab40a2e58769123f71b188.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "type": "inline_equation", + "content": "Q_{s}, K_{v}, V_{v}" + }, + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "type": "text", + "content": " are the snow query, video key, and value matrices. " + }, + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "type": "text", + "content": " is the dimension. " + }, + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "type": "text", + "content": " is the relative position embedding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 373, + 211, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 211, + 385 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 211, + 385 + ], + "type": "text", + "content": "4.3. Snow-Aware Contrastive Loss" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 392, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 287, + 534 + ], + "type": "text", + "content": "We introduce a prior for snow-covered images: the snow features vary across videos based on the snow's angle and properties (e.g., density, shape, direction, and speed), but are consistent within the same video, as shown in Figure 7. We design a snow-aware contrastive learning paradigm. It uses snow features from different frames of the same video as positive samples " + }, + { + "bbox": [ + 46, + 392, + 287, + 534 + ], + "type": "inline_equation", + "content": "S^{+}" + }, + { + "bbox": [ + 46, + 392, + 287, + 534 + ], + "type": "text", + "content": " and snow features from different videos as negative samples " + }, + { + "bbox": [ + 46, + 392, + 287, + 534 + ], + "type": "inline_equation", + "content": "S^{-}" + }, + { + "bbox": [ + 46, + 392, + 287, + 534 + ], + "type": "text", + "content": ". By pulling positive samples closer together and pushing negative samples farther apart in feature space, the network can more accurately identify snow features and improve snow removal. The loss function for contrastive learning is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 536, + 293, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 293, + 592 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 293, + 592 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {c o n t}} \\left(S, S ^ {+}, S ^ {-}\\right) = \\\\ - \\log \\left[ \\frac {\\exp \\left(D \\left(S ^ {+} , S ^ {-}\\right) / \\tau\\right)}{\\exp \\left(D \\left(S ^ {+} , S ^ {-}\\right) / \\tau\\right) + \\sum_ {r = 1} ^ {N} \\exp \\left(D \\left(S ^ {+} , S ^ {-}\\right) / \\tau\\right)} \\right] \\tag {8} \\\\ \\end{array}", + "image_path": "c30cce7b0c0c939f54d95ddfe98f7e6303a355d3fd3749b85e737a492c4b0c31.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "type": "inline_equation", + "content": "D(x,y) = \\Psi (x)\\cdot \\Psi (y)" + }, + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\Psi (\\cdot)" + }, + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "type": "text", + "content": " is the two-layer MLP function which implements feature projection. " + }, + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "type": "text", + "content": " is the scale temperature which is set to 0.1. " + }, + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "type": "text", + "content": " denotes the total number of negative samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "To illustrate the effect of the proposed snow-aware contrastive learning loss on snow feature learning, we visualize the learned features using t-SNE for three distinct types of snow with different motion directions and snow stripe sizes, as shown in Figure 10. Without the proposed contrastive learning, different snow types have highly similar and over" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 249, + 547, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 249, + 547, + 321 + ], + "spans": [ + { + "bbox": [ + 304, + 249, + 547, + 321 + ], + "type": "text", + "content": "lapping feature distributions. In contrast, with the contrastive learning loss, the features of different snow types are clearly distinguished, demonstrating the ability of the method to accurately identify diverse snow features. The overall loss function is formulated as: " + }, + { + "bbox": [ + 304, + 249, + 547, + 321 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_1 + \\lambda \\mathcal{L}_{cont}" + }, + { + "bbox": [ + 304, + 249, + 547, + 321 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 249, + 547, + 321 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 249, + 547, + 321 + ], + "type": "text", + "content": " is set to 0.1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 337, + 388, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 337, + 388, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 337, + 388, + 350 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 358, + 512, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 358, + 512, + 371 + ], + "spans": [ + { + "bbox": [ + 305, + 358, + 512, + 371 + ], + "type": "text", + "content": "5.1. Proposed Dataset vs. Previous Datasets" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 378, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 713 + ], + "type": "text", + "content": "Objective Evaluation. Previous work [4, 51] has demonstrated the significant impact of training set data on the generalization performance of the model. To demonstrate that the snow scenes in our created dataset are more realistic and closer to real-world scenarios than those in other synthetic datasets, we train the same model on different datasets and test it on real snowy images. The results show that the model trained on our proposed dataset performs better in handling real snow, thus validating the authenticity and usefulness of our dataset. Figure 11 displays two samples. In the first row, the model trained on our dataset successfully removes all snow residues, whereas models trained on other datasets fail to completely eliminate snow, leaving numerous snow residues in the output images. Since the Snow100K, SnowCityScapes, and SnowKITTI2012 datasets do not include the effect of snow haze, the models trained on these datasets cannot remove snow haze. In the second row, the models trained on the Snow100K, SnowCityScapes, and SnowKITTI2012 datasets remove slightly larger snowflakes but preserve many smaller ones. Furthermore, these models mistakenly remove white parallelogram backgrounds, mistaking them for snowflakes. Conversely, SRRS and CSD remove smaller snowflakes but fail to remove larger ones. In contrast, our proposed dataset facilitates the training of a model that effectively removes snowflakes at all sizes. It does not misidentify the white background as snow. Since there is no ground truth for real-world images, we conduct a user study to compare the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "13217" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 289, + 149 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 289, + 149 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 289, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 289, + 149 + ], + "type": "image", + "image_path": "6c436a64193c6d1b2f86bd1fb918bebcb81dc21583cf833d3ce556a0841c6353.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 156, + 288, + 194 + ], + "lines": [ + { + "bbox": [ + 46, + 156, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 288, + 194 + ], + "type": "text", + "content": "Figure 12: Distributions of mean opinion scores obtained by models trained on different datasets (testing on real-world snow images)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 215, + 287, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 215, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 287, + 262 + ], + "type": "text", + "content": "desnowing results of models which are trained on different datasets, as shown in Figure 12. Model trained on our dataset has the best desnowing performance on real-world images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 264, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 288, + 373 + ], + "type": "text", + "content": "Subjective Evaluation. To demonstrate that our dataset produces snow images that are closer to real snow in human visual perception than previous datasets, we conduct a user study to compare the perceived realism of snow generated from different datasets. The results are shown in Figure 2, dataset with higher scores indicating greater fidelity. Thanks to our proposed complex, specially designed dataset construction method, our dataset is scored much higher than the other datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 381, + 185, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 185, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 185, + 393 + ], + "type": "text", + "content": "5.2. Performance Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 400, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 496 + ], + "type": "text", + "content": "Implementation Details. Each input image is randomly cropped to a spatial resolution of " + }, + { + "bbox": [ + 46, + 400, + 287, + 496 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 400, + 287, + 496 + ], + "type": "text", + "content": ", and the number of the total training iterations is " + }, + { + "bbox": [ + 46, + 400, + 287, + 496 + ], + "type": "inline_equation", + "content": "500\\mathrm{K}" + }, + { + "bbox": [ + 46, + 400, + 287, + 496 + ], + "type": "text", + "content": ". During the training phase, we adopt the Adam optimizer [19] with " + }, + { + "bbox": [ + 46, + 400, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 46, + 400, + 287, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 400, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.99" + }, + { + "bbox": [ + 46, + 400, + 287, + 496 + ], + "type": "text", + "content": ". The initial learning rate is set to " + }, + { + "bbox": [ + 46, + 400, + 287, + 496 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 400, + 287, + 496 + ], + "type": "text", + "content": " and reduced by half at the milestone of " + }, + { + "bbox": [ + 46, + 400, + 287, + 496 + ], + "type": "inline_equation", + "content": "100\\mathrm{K}" + }, + { + "bbox": [ + 46, + 400, + 287, + 496 + ], + "type": "text", + "content": " iterations and " + }, + { + "bbox": [ + 46, + 400, + 287, + 496 + ], + "type": "inline_equation", + "content": "150\\mathrm{K}" + }, + { + "bbox": [ + 46, + 400, + 287, + 496 + ], + "type": "text", + "content": " iterations. The batch size is set to 4. PyTorch [28] is used to implement our model with 4 RTX 3090 GPUs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "type": "text", + "content": "Compared Methods. We compare our network against 18 state-of-the-art methods, including three single-image desnowing models, two image adverse weather restoration models, six image restoration methods, two video deraining methods, and five video restoration methods. Please refer to Table 5 for more details. Moreover, we follow [16, 15] to employ PSNR, SSIM [38], and LPIPS [52] to quantitatively compared different methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Quantitative Comparison. Table 5 reports the quantitative results of our network and 17 compared methods. From these quantitative results, we can find that the performance of video-based models is generally inferior to that of single-image models. The major reason behind is that the severe occlusion of the video background by snowflakes hinders the alignment and fusion of the video frames. Among all compared methods, Restormer has the largest PSNR score of 24.34 and the smallest 0.1164, while MPRNet has the largest SSIM score of 0.8960. More importantly, our net" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 71, + 545, + 102 + ], + "blocks": [ + { + "bbox": [ + 308, + 71, + 545, + 102 + ], + "lines": [ + { + "bbox": [ + 308, + 71, + 545, + 102 + ], + "spans": [ + { + "bbox": [ + 308, + 71, + 545, + 102 + ], + "type": "table", + "html": "
MethodRestormerTransWeatherS2VDRVRTBasicVSR++Ours
E_warp ↓(×10-3)1.5942.3922.0093.4252.6131.203
VFID ↓0.09380.15110.16860.22070.15390.0492
", + "image_path": "106c2bed74bdf4eddc5c6d964c65b50faecf3f0586a60f65a0845d70f4755d55.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 145, + 544, + 324 + ], + "blocks": [ + { + "bbox": [ + 315, + 120, + 534, + 133 + ], + "lines": [ + { + "bbox": [ + 315, + 120, + 534, + 133 + ], + "spans": [ + { + "bbox": [ + 315, + 120, + 534, + 133 + ], + "type": "text", + "content": "Table 4: Results of temporal consistency preservation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 145, + 544, + 324 + ], + "lines": [ + { + "bbox": [ + 308, + 145, + 544, + 324 + ], + "spans": [ + { + "bbox": [ + 308, + 145, + 544, + 324 + ], + "type": "table", + "html": "
TypeMethodVenuePSNR ↑SSIM ↑LPIPS ↓
Image DesnowingJSTASR [11]ECCV 202022.080.82800.2336
HDCW-Net [12]ICCV 202122.630.85920.2010
SnowFormer [8]arXiv 202224.010.89390.1219
Image Adverse WeatherTransWeather [35]CVPR 202223.110.85430.2086
TKL [13]CVPR 202223.050.85890.2027
Image RestorationSwinIR [23]CVPR 202122.510.85620.2105
MPRNet [49]CVPR 202124.270.89600.1266
Uformer [39]CVPR 202223.610.87300.1706
DGUNet [27]CVPR 202224.180.89850.1238
NAFNet [6]ECCV 202224.010.88380.1472
Restormer [48]CVPR 202224.340.89290.1164
Video DerainingS2VD [47]CVPR 202122.950.85900.1856
RDD [36]ECCV 202222.970.87420.1631
Video RestorationEDVR [37]CVPRW 201917.930.57900.3771
BasicVSR [2]CVPR 202122.460.84730.2087
IconVSR [2]CVPR 202122.350.84820.2034
BasicVSR++ [3]CVPR 202222.640.86180.1868
RVRT [24]NeurlPS 202220.900.79740.2977
SVDNet (Ours)-25.060.92100.0842
", + "image_path": "1fccd8f1e1dfb2049e5ed4d1bcf72f80f0e27caa69def37f9cb9f80ef3dc0449.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 342, + 545, + 355 + ], + "lines": [ + { + "bbox": [ + 306, + 342, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 545, + 355 + ], + "type": "text", + "content": "Table 5: Quantitative results on the proposed video dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "type": "text", + "content": "work further outperform these methods in terms of PSNR, SSIM, and LPIPS scores. It improves the PSNR score from 24.34 to 25.06, the SSIM score from 0.8960 to 0.9210, and the LPIPS score from 0.1164 to 0.0842. It demonstrates the effectiveness of our proposed framework, which first obtains a coarse snow removal result for a single frame image to reduce the challenging alignment issue on the snowy video frames, and then fully utilizes the temporal information between video frames for video desnowing." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 485, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 556 + ], + "type": "text", + "content": "To further verify the temporal consistency preservation, we utilize a widely-used metric " + }, + { + "bbox": [ + 304, + 485, + 545, + 556 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}" + }, + { + "bbox": [ + 304, + 485, + 545, + 556 + ], + "type": "text", + "content": " to compute the temporal consistency, and VFID [17] to compute video restoration performance. Table below shows our method has the smallest " + }, + { + "bbox": [ + 304, + 485, + 545, + 556 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}" + }, + { + "bbox": [ + 304, + 485, + 545, + 556 + ], + "type": "text", + "content": " and VFID. Our method can achieve consistent desnowing between frames." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "content": "Visual comparisons on synthetic videos. Figure 13 visually compares video desnowing results produced by our network and state-of-the-art methods. We can find that existing methods tend to maintain snowflakes (e.g., TKL, TransWeather), or haze (e.g., Restomer), or both snowflakes and haze (i.e., RVRT) in their desnowing results. On the contrary, our method can effectively remove both snowflakes and haze, and better preserve background details." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Visual comparisons on real-world videos. Moreover, Figure 14 shows the visual comparisons between our network and state-of-the-art methods in terms of input frames from real-world snowstorm videos with both heavy snow and haze. We can find that our method can effectively remove" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13218" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 120, + 127 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 120, + 127 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 120, + 127 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 120, + 127 + ], + "type": "image", + "image_path": "6f58e5f0032cc5cf935036c23af03def374ca884cc2c68e723915cb60a82f441.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 121, + 70, + 191, + 128 + ], + "blocks": [ + { + "bbox": [ + 121, + 70, + 191, + 128 + ], + "lines": [ + { + "bbox": [ + 121, + 70, + 191, + 128 + ], + "spans": [ + { + "bbox": [ + 121, + 70, + 191, + 128 + ], + "type": "image", + "image_path": "8b65bab7ea4fad30b21806a96664127f68502af3edcaff070fdcecdcb9031eb3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 129, + 170, + 136 + ], + "lines": [ + { + "bbox": [ + 141, + 129, + 170, + 136 + ], + "spans": [ + { + "bbox": [ + 141, + 129, + 170, + 136 + ], + "type": "text", + "content": "TKL [13]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 192, + 70, + 261, + 128 + ], + "blocks": [ + { + "bbox": [ + 192, + 70, + 261, + 128 + ], + "lines": [ + { + "bbox": [ + 192, + 70, + 261, + 128 + ], + "spans": [ + { + "bbox": [ + 192, + 70, + 261, + 128 + ], + "type": "image", + "image_path": "3d5614415600dedbda16c5cb2cf2541ef58b393190ae2c7225d20466f01fbbbe.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 262, + 70, + 332, + 128 + ], + "blocks": [ + { + "bbox": [ + 262, + 70, + 332, + 128 + ], + "lines": [ + { + "bbox": [ + 262, + 70, + 332, + 128 + ], + "spans": [ + { + "bbox": [ + 262, + 70, + 332, + 128 + ], + "type": "image", + "image_path": "775bcbe1f5b8f820706f060f7c6d68616ebcf4dbce1719f2298720ccd41e24ea.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 332, + 70, + 402, + 128 + ], + "blocks": [ + { + "bbox": [ + 332, + 70, + 402, + 128 + ], + "lines": [ + { + "bbox": [ + 332, + 70, + 402, + 128 + ], + "spans": [ + { + "bbox": [ + 332, + 70, + 402, + 128 + ], + "type": "image", + "image_path": "b1c82358c4a06d9b227c8920f56888bd4f5eb533437d980e42b8e484cc4d8c53.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 403, + 70, + 473, + 128 + ], + "blocks": [ + { + "bbox": [ + 403, + 70, + 473, + 128 + ], + "lines": [ + { + "bbox": [ + 403, + 70, + 473, + 128 + ], + "spans": [ + { + "bbox": [ + 403, + 70, + 473, + 128 + ], + "type": "image", + "image_path": "a2f8fef4027292335e8696c318056a28c7ae95f626ffbb44e530487488c18976.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 474, + 70, + 543, + 128 + ], + "blocks": [ + { + "bbox": [ + 474, + 70, + 543, + 128 + ], + "lines": [ + { + "bbox": [ + 474, + 70, + 543, + 128 + ], + "spans": [ + { + "bbox": [ + 474, + 70, + 543, + 128 + ], + "type": "image", + "image_path": "9eaa5d5b0500203c25687d5440a1d7094e22ab431e11c7eb72e8c47b0c71d37b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 51, + 137, + 120, + 194 + ], + "blocks": [ + { + "bbox": [ + 67, + 129, + 104, + 137 + ], + "lines": [ + { + "bbox": [ + 67, + 129, + 104, + 137 + ], + "spans": [ + { + "bbox": [ + 67, + 129, + 104, + 137 + ], + "type": "text", + "content": "Input Frame" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 137, + 120, + 194 + ], + "lines": [ + { + "bbox": [ + 51, + 137, + 120, + 194 + ], + "spans": [ + { + "bbox": [ + 51, + 137, + 120, + 194 + ], + "type": "image", + "image_path": "2f3c69faaca36ba166e933a4a243ce97635567aaf5aa9f76a9996fe0a904d020.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 194, + 102, + 203 + ], + "lines": [ + { + "bbox": [ + 69, + 194, + 102, + 203 + ], + "spans": [ + { + "bbox": [ + 69, + 194, + 102, + 203 + ], + "type": "text", + "content": "S2VD [47]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 100, + 213, + 492, + 225 + ], + "lines": [ + { + "bbox": [ + 100, + 213, + 492, + 225 + ], + "spans": [ + { + "bbox": [ + 100, + 213, + 492, + 225 + ], + "type": "text", + "content": "Figure 13: Visual video desnowing results produced by our network and state-of-the-art methods." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 121, + 137, + 191, + 194 + ], + "blocks": [ + { + "bbox": [ + 121, + 137, + 191, + 194 + ], + "lines": [ + { + "bbox": [ + 121, + 137, + 191, + 194 + ], + "spans": [ + { + "bbox": [ + 121, + 137, + 191, + 194 + ], + "type": "image", + "image_path": "53caa05dec53d0f360ef223673273f20133036ed9f8b458ebe831bd834e2d628.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 194, + 170, + 203 + ], + "lines": [ + { + "bbox": [ + 141, + 194, + 170, + 203 + ], + "spans": [ + { + "bbox": [ + 141, + 194, + 170, + 203 + ], + "type": "text", + "content": "RDD [36]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 192, + 137, + 261, + 194 + ], + "blocks": [ + { + "bbox": [ + 200, + 129, + 253, + 137 + ], + "lines": [ + { + "bbox": [ + 200, + 129, + 253, + 137 + ], + "spans": [ + { + "bbox": [ + 200, + 129, + 253, + 137 + ], + "type": "text", + "content": "TransWeather [35]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 192, + 137, + 261, + 194 + ], + "lines": [ + { + "bbox": [ + 192, + 137, + 261, + 194 + ], + "spans": [ + { + "bbox": [ + 192, + 137, + 261, + 194 + ], + "type": "image", + "image_path": "cc46f46f0318eed6ab50472f49abf23ef0bd600dd3dde40db0c02d5096228685.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 210, + 194, + 243, + 203 + ], + "lines": [ + { + "bbox": [ + 210, + 194, + 243, + 203 + ], + "spans": [ + { + "bbox": [ + 210, + 194, + 243, + 203 + ], + "type": "text", + "content": "RVRT [24]" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 262, + 137, + 332, + 194 + ], + "blocks": [ + { + "bbox": [ + 274, + 129, + 320, + 137 + ], + "lines": [ + { + "bbox": [ + 274, + 129, + 320, + 137 + ], + "spans": [ + { + "bbox": [ + 274, + 129, + 320, + 137 + ], + "type": "text", + "content": "Snowformer [8]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 262, + 137, + 332, + 194 + ], + "lines": [ + { + "bbox": [ + 262, + 137, + 332, + 194 + ], + "spans": [ + { + "bbox": [ + 262, + 137, + 332, + 194 + ], + "type": "image", + "image_path": "576b1a935a986d2f5064f000e1dafafa14bbe0ec3ca617a6d854914165967238.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 278, + 194, + 315, + 203 + ], + "lines": [ + { + "bbox": [ + 278, + 194, + 315, + 203 + ], + "spans": [ + { + "bbox": [ + 278, + 194, + 315, + 203 + ], + "type": "text", + "content": "IconVSR [2]" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 332, + 137, + 402, + 194 + ], + "blocks": [ + { + "bbox": [ + 345, + 129, + 389, + 137 + ], + "lines": [ + { + "bbox": [ + 345, + 129, + 389, + 137 + ], + "spans": [ + { + "bbox": [ + 345, + 129, + 389, + 137 + ], + "type": "text", + "content": "Restormer [48]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 332, + 137, + 402, + 194 + ], + "lines": [ + { + "bbox": [ + 332, + 137, + 402, + 194 + ], + "spans": [ + { + "bbox": [ + 332, + 137, + 402, + 194 + ], + "type": "image", + "image_path": "db52a41ac806d5411c6727e9356709068505380725e0a0bfd8092b26d7392a11.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 194, + 392, + 203 + ], + "lines": [ + { + "bbox": [ + 343, + 194, + 392, + 203 + ], + "spans": [ + { + "bbox": [ + 343, + 194, + 392, + 203 + ], + "type": "text", + "content": "BasicVSR++ [3]" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 403, + 137, + 473, + 194 + ], + "blocks": [ + { + "bbox": [ + 417, + 129, + 457, + 137 + ], + "lines": [ + { + "bbox": [ + 417, + 129, + 457, + 137 + ], + "spans": [ + { + "bbox": [ + 417, + 129, + 457, + 137 + ], + "type": "text", + "content": "MPRNet [49]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 403, + 137, + 473, + 194 + ], + "lines": [ + { + "bbox": [ + 403, + 137, + 473, + 194 + ], + "spans": [ + { + "bbox": [ + 403, + 137, + 473, + 194 + ], + "type": "image", + "image_path": "5a696d3cdf283181e8b6ae8fc7532ae53e7c342b4526780ecb15acb543f773bc.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 194, + 445, + 203 + ], + "lines": [ + { + "bbox": [ + 430, + 194, + 445, + 203 + ], + "spans": [ + { + "bbox": [ + 430, + 194, + 445, + 203 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 474, + 137, + 543, + 194 + ], + "blocks": [ + { + "bbox": [ + 488, + 129, + 528, + 137 + ], + "lines": [ + { + "bbox": [ + 488, + 129, + 528, + 137 + ], + "spans": [ + { + "bbox": [ + 488, + 129, + 528, + 137 + ], + "type": "text", + "content": "DGUNet [27]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 474, + 137, + 543, + 194 + ], + "lines": [ + { + "bbox": [ + 474, + 137, + 543, + 194 + ], + "spans": [ + { + "bbox": [ + 474, + 137, + 543, + 194 + ], + "type": "image", + "image_path": "0cc95c97bb83078a52a57b7418638907170f19d8c430b36e5abdd2ae06a8fd59.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 194, + 528, + 203 + ], + "lines": [ + { + "bbox": [ + 488, + 194, + 528, + 203 + ], + "spans": [ + { + "bbox": [ + 488, + 194, + 528, + 203 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 51, + 244, + 108, + 302 + ], + "blocks": [ + { + "bbox": [ + 51, + 244, + 108, + 302 + ], + "lines": [ + { + "bbox": [ + 51, + 244, + 108, + 302 + ], + "spans": [ + { + "bbox": [ + 51, + 244, + 108, + 302 + ], + "type": "image", + "image_path": "5067a693dac14c7b2450b704644c91ff247c9f65d8ce6554187dd9f28439a511.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 110, + 244, + 167, + 302 + ], + "blocks": [ + { + "bbox": [ + 110, + 244, + 167, + 302 + ], + "lines": [ + { + "bbox": [ + 110, + 244, + 167, + 302 + ], + "spans": [ + { + "bbox": [ + 110, + 244, + 167, + 302 + ], + "type": "image", + "image_path": "4e5339b483474826997b7b09cad4c3eb22f02facd2149b83864343d00fbbdf0f.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 168, + 244, + 226, + 302 + ], + "blocks": [ + { + "bbox": [ + 168, + 244, + 226, + 302 + ], + "lines": [ + { + "bbox": [ + 168, + 244, + 226, + 302 + ], + "spans": [ + { + "bbox": [ + 168, + 244, + 226, + 302 + ], + "type": "image", + "image_path": "b862fdd9a66d2b4bd5eb0444ee47bafe6349100b8d946f910439832c8032ef20.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 227, + 244, + 285, + 302 + ], + "blocks": [ + { + "bbox": [ + 227, + 244, + 285, + 302 + ], + "lines": [ + { + "bbox": [ + 227, + 244, + 285, + 302 + ], + "spans": [ + { + "bbox": [ + 227, + 244, + 285, + 302 + ], + "type": "image", + "image_path": "804247cd4a7c4ee81a34be0a037b797e31b031cb95353b316400721b790a2261.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 51, + 312, + 108, + 369 + ], + "blocks": [ + { + "bbox": [ + 63, + 304, + 96, + 311 + ], + "lines": [ + { + "bbox": [ + 63, + 304, + 96, + 311 + ], + "spans": [ + { + "bbox": [ + 63, + 304, + 96, + 311 + ], + "type": "text", + "content": "Input Frame" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 312, + 108, + 369 + ], + "lines": [ + { + "bbox": [ + 51, + 312, + 108, + 369 + ], + "spans": [ + { + "bbox": [ + 51, + 312, + 108, + 369 + ], + "type": "image", + "image_path": "510966f35b2a527de801f85a3357ccd0ee22480832308085001da67c47cbd34f.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 369, + 94, + 377 + ], + "lines": [ + { + "bbox": [ + 64, + 369, + 94, + 377 + ], + "spans": [ + { + "bbox": [ + 64, + 369, + 94, + 377 + ], + "type": "text", + "content": "S2VD [47]" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 387, + 282, + 399 + ], + "lines": [ + { + "bbox": [ + 52, + 387, + 282, + 399 + ], + "spans": [ + { + "bbox": [ + 52, + 387, + 282, + 399 + ], + "type": "text", + "content": "Figure 14: Visual comparison on real snowstorm video." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 110, + 312, + 167, + 369 + ], + "blocks": [ + { + "bbox": [ + 119, + 304, + 157, + 311 + ], + "lines": [ + { + "bbox": [ + 119, + 304, + 157, + 311 + ], + "spans": [ + { + "bbox": [ + 119, + 304, + 157, + 311 + ], + "type": "text", + "content": "MPRNet [49]" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 110, + 312, + 167, + 369 + ], + "lines": [ + { + "bbox": [ + 110, + 312, + 167, + 369 + ], + "spans": [ + { + "bbox": [ + 110, + 312, + 167, + 369 + ], + "type": "image", + "image_path": "c74014ae602bab47ed84249d62f13b8ef6030d2722e1c604963e0cdc1cea9aac.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 120, + 369, + 156, + 377 + ], + "lines": [ + { + "bbox": [ + 120, + 369, + 156, + 377 + ], + "spans": [ + { + "bbox": [ + 120, + 369, + 156, + 377 + ], + "type": "text", + "content": "IconVSR [2]" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 168, + 312, + 226, + 369 + ], + "blocks": [ + { + "bbox": [ + 178, + 304, + 215, + 311 + ], + "lines": [ + { + "bbox": [ + 178, + 304, + 215, + 311 + ], + "spans": [ + { + "bbox": [ + 178, + 304, + 215, + 311 + ], + "type": "text", + "content": "DGUNet [27]" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 168, + 312, + 226, + 369 + ], + "lines": [ + { + "bbox": [ + 168, + 312, + 226, + 369 + ], + "spans": [ + { + "bbox": [ + 168, + 312, + 226, + 369 + ], + "type": "image", + "image_path": "4fcc296e79418007c6e2a86979ccebf4e00e97306100ca7f7b699c168e0bc953.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 369, + 219, + 377 + ], + "lines": [ + { + "bbox": [ + 174, + 369, + 219, + 377 + ], + "spans": [ + { + "bbox": [ + 174, + 369, + 219, + 377 + ], + "type": "text", + "content": "BasicVSR++[3]" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 227, + 312, + 285, + 369 + ], + "blocks": [ + { + "bbox": [ + 241, + 304, + 269, + 311 + ], + "lines": [ + { + "bbox": [ + 241, + 304, + 269, + 311 + ], + "spans": [ + { + "bbox": [ + 241, + 304, + 269, + 311 + ], + "type": "text", + "content": "RDD [36]" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 227, + 312, + 285, + 369 + ], + "lines": [ + { + "bbox": [ + 227, + 312, + 285, + 369 + ], + "spans": [ + { + "bbox": [ + 227, + 312, + 285, + 369 + ], + "type": "image", + "image_path": "6b358618061b3712cbd970e1b0d0a4b090cc3748307db32020005f5c45fe6b0d.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 370, + 263, + 377 + ], + "lines": [ + { + "bbox": [ + 247, + 370, + 263, + 377 + ], + "spans": [ + { + "bbox": [ + 247, + 370, + 263, + 377 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "table", + "bbox": [ + 50, + 412, + 286, + 480 + ], + "blocks": [ + { + "bbox": [ + 50, + 412, + 286, + 480 + ], + "lines": [ + { + "bbox": [ + 50, + 412, + 286, + 480 + ], + "spans": [ + { + "bbox": [ + 50, + 412, + 286, + 480 + ], + "type": "table", + "html": "
IndexTemporal AggregationSnow-QueryContrastive LossPSNR ↑SSIM ↑LPIPS ↓
M1(basic)24.160.89900.1125
M2✓24.580.91300.0934
M3✓✓24.800.91080.9004
M4✓✓24.840.91810.0882
Ours✓✓✓25.060.92100.0842
", + "image_path": "c8aaa7efef0cc15ad7487d1a9c262d4730cea6d876e582fd101859d78ab1db34.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "table_body" + } + ], + "index": 46 + }, + { + "bbox": [ + 66, + 498, + 268, + 511 + ], + "lines": [ + { + "bbox": [ + 66, + 498, + 268, + 511 + ], + "spans": [ + { + "bbox": [ + 66, + 498, + 268, + 511 + ], + "type": "text", + "content": "Table 6: Quantitative results of the ablation study." + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 533, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 569 + ], + "type": "text", + "content": "both the haze and snowflakes and betterrecover the obscured background details, while compared methods tend to main parts of haze and snowflakes." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "Ablation Study. We perform ablation study experiments to verify the temporal aggregation module, the snow-query cross-attention module, and the snow-aware contrastive loss of our network. To do so, we first built a basic model (see M1 of Table 6) by removing these three major components from our network, and then add the temporal aggregation module into M1 to construct M2. Then we add the snow-query cross-attention module into M2 to build M3, and add the snow-aware contrastive loss into M2 to build M4. Table 6 reports the PSNR, SSIM, and LPIPS results of our network and four baseline networks. Specifically, compared to M1, M2 improves the PSNR score from 24.16" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 304, + 246, + 547, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 547, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 547, + 378 + ], + "type": "text", + "content": "to 24.58, the SSIM score from 0.8990 to 0.9130, and the LPIPS score from 0.1125 to 0.0934. It shows that the temporal aggregation module incurs a better video desnowing performance. Then, we find that M3 and M4 have a superior PSNR, SSIM, and LPIPS performance over M2. It demonstrates the effectiveness of the snow-aware cross-attention module or snow-aware contrastive loss for enhancing the video desnowing performance. Moreover, our method has larger PSNR and SSIM scores and a smaller LPIPS score than M3 and M4. It shows that combing three components together has the best performance of video desnowing." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 306, + 387, + 378, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 387, + 378, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 387, + 378, + 399 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 407, + 547, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 547, + 503 + ], + "type": "text", + "content": "The first contribution of our work is to synthesize the first high-quality video desnowing dataset, which is more realistic than previous image datasets. Then, we devise a deep learning-based framework that incorporates a snow-query temporal aggregation module and a snow-aware contrastive learning loss. Experimental results show that our network outperforms state-of-the-art methods in terms of video denosing on synthetic and real-world snowy videos." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 505, + 547, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 547, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 547, + 553 + ], + "type": "text", + "content": "Acknowledgment. This work was supported by the National Natural Science Foundation of China (Grant No. 61902275), and Guangzhou Municipal Science and Technology Project (Grant No. 2023A03J0671)." + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 306, + 562, + 365, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 365, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 365, + 574 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 310, + 581, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 58, + "blocks": [ + { + "bbox": [ + 310, + 581, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 581, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 310, + 581, + 547, + 624 + ], + "type": "text", + "content": "[1] Jérémie Bossu, Nicolas Hautiere, and Jean-Philippe Tarel. Rain or snow detection in image sequences through use of a histogram of orientation of streaks. International journal of computer vision, 93:348–367, 2011." + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 311, + 625, + 547, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 625, + 547, + 680 + ], + "spans": [ + { + "bbox": [ + 311, + 625, + 547, + 680 + ], + "type": "text", + "content": "[2] Kelvin C.K. Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2021." + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 311, + 681, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 681, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 311, + 681, + 547, + 714 + ], + "type": "text", + "content": "[3] Kelvin C.K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. BasicVSR++: Improving video superresolution with enhanced propagation and alignment. In" + } + ] + } + ], + "index": 57 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13219" + } + ] + } + ], + "index": 59 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "text", + "content": "IEEE Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 287, + 160 + ], + "type": "text", + "content": "[4] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1692-1703, June 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 162, + 286, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 162, + 286, + 183 + ], + "spans": [ + { + "bbox": [ + 54, + 162, + 286, + 183 + ], + "type": "text", + "content": "[5] Haoyu Chen, Jinjin Gu, and Zhi Zhang. Attention in attention network for image super-resolution, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 184, + 286, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 184, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 54, + 184, + 286, + 239 + ], + "type": "text", + "content": "[6] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 17-33. Springer, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 239, + 286, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 286, + 272 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 286, + 272 + ], + "type": "text", + "content": "[7] Sixiang Chen, Tian Ye, Yun Liu, and Erkang Chen. Dualformer: Hybrid self-attention transformer for efficient image restoration. arXiv preprint arXiv:2210.01069, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 273, + 286, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 273, + 286, + 316 + ], + "spans": [ + { + "bbox": [ + 54, + 273, + 286, + 316 + ], + "type": "text", + "content": "[8] Sixiang Chen, Tian Ye, Yun Liu, Erkang Chen, Jun Shi, and Jingchun Zhou. Snowformer: Scale-aware transformer via context interaction for single image desnowing. arXiv preprint arXiv:2208.09703, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 317, + 286, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 317, + 286, + 360 + ], + "spans": [ + { + "bbox": [ + 54, + 317, + 286, + 360 + ], + "type": "text", + "content": "[9] Sixiang Chen, Tian Ye, Yun Liu, Erkang Chen, Jun Shi, and Jingchun Zhou. Snowformer: Scale-aware transformer via context interaction for single image desnowing. arXiv preprint arXiv:2208.09703, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 286, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 286, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 286, + 426 + ], + "type": "text", + "content": "[10] Sixiang Chen, Tian Ye, Yun Liu, Taodong Liao, Jingxia Jiang, Erkang Chen, and Peng Chen. Msp-former: Multiscale projection transformer for single image desnowing. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 427, + 286, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 427, + 286, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 427, + 286, + 503 + ], + "type": "text", + "content": "[11] Wei-Ting Chen, Hao-Yu Fang, Jian-Jiun Ding, Cheng-Che Tsai, and Sy-Yen Kuo. Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 754-770. Springer, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 505, + 286, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 505, + 286, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 505, + 286, + 569 + ], + "type": "text", + "content": "[12] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4196-4205, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 571, + 286, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 286, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 286, + 645 + ], + "type": "text", + "content": "[13] Wei-Ting Chen, Zhi-Kai Huang, Cheng-Che Tsai, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17653–17662, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 712 + ], + "type": "text", + "content": "[14] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5374-5383, 2019." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[15] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Image quality assessment for perceptual image restoration: A new dataset, benchmark and metric. arXiv preprint arXiv:2011.15002, 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 118, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 173 + ], + "type": "text", + "content": "[16] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In European Conference on Computer Vision, pages 633-651. Springer, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 175, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 228 + ], + "type": "text", + "content": "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 230, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 274 + ], + "type": "text", + "content": "[18] Jin-Hwan Kim, Jae-Young Sim, and Chang-Su Kim. Video deraining and desnowing using temporal correlation and low-rank matrix completion. IEEE Transactions on Image Processing, 24(9):2658-2670, 2015." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 275, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 275, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 308, + 275, + 545, + 306 + ], + "type": "text", + "content": "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "type": "text", + "content": "[20] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "type": "text", + "content": "[21] Minghan Li, Xiangyong Cao, Qian Zhao, Lei Zhang, Chenqiang Gao, and Deyu Meng. Video rain/snow removal by transformed online multiscale convolutional sparse coding. arXiv preprint arXiv:1909.06148, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "type": "text", + "content": "[22] Minghan Li, Xiangyong Cao, Qian Zhao, Lei Zhang, and Deyu Meng. Online rain/snow removal from surveillance videos. IEEE Transactions on Image Processing, 30:2029-2044, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "type": "text", + "content": "[23] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 501, + 545, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 554 + ], + "type": "text", + "content": "[24] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent video restoration transformer with guided deformable attention. arXiv preprint arXiv:2206.02146, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 556, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 556, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 556, + 545, + 599 + ], + "type": "text", + "content": "[25] Yun-Fu Liu, Da-Wei Jaw, Shih-Chia Huang, and Jenq-Neng Hwang. Desnownet: Context-aware deep network for snow removal. IEEE Transactions on Image Processing, 27(6):3064-3073, 2018." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 601, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 645 + ], + "type": "text", + "content": "[26] Xin Luo, Yunan Zhu, Shunxin Xu, and Dong Liu. On the effectiveness of spectral discriminators for perceptual quality improvement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[27] Chong Mou, Qian Wang, and Jian Zhang. Deep generalized unfolding networks for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17399-17410, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13220" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 159 + ], + "type": "text", + "content": "[29] Soo-Chang Pei, Yu-Tai Tsai, and Chen-Yu Lee. Removing rain and snow in a single image using saturation and visibility features. In 2014 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 1-6. IEEE, 2014." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 287, + 205 + ], + "type": "text", + "content": "[30] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4161-4170, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "type": "text", + "content": "[31] Jingjing Ren, Qingqing Zheng, Yuanyuan Zhao, Xuemiao Xu, and Chen Li. Dlformer: Discrete latent transformer for video inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3511-3520, June 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 315 + ], + "type": "text", + "content": "[32] Weihong Ren, Jiandong Tian, Zhi Han, Antoni Chan, and Yandong Tang. Video desnowing and deraining based on matrix decomposition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4210-4219, 2017." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 381 + ], + "type": "text", + "content": "[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "type": "text", + "content": "[34] Mingxing Tan, Ruoming Pang, and Quoc V Le. Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10781-10790, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 427, + 287, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 427, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 48, + 427, + 287, + 481 + ], + "type": "text", + "content": "[35] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 482, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 557 + ], + "type": "text", + "content": "[36] Shuai Wang, Lei Zhu, Huazhu Fu, Jing Qin, Carola-Bibiane Schonlieb, Wei Feng, and Song Wang. Rethinking video rain streak removal: A new synthesis model and a deraining network with video rain prior. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIX, pages 565-582. Springer, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "type": "text", + "content": "[37] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "text", + "content": "[38] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[39] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[40] Zeyu Xiao, Xueyang Fu, Jie Huang, Zhen Cheng, and Zhiwei Xiong. Space-time distillation for video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2113-2122, June 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 130, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 185 + ], + "type": "text", + "content": "[41] Zeyu Xiao, Yutong Liu, Ruisheng Gao, and Zhiwei Xiong. Cutmib: Boosting light field super-resolution via multi-view image blending. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1672-1682, June 2023." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 186, + 545, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 231 + ], + "type": "text", + "content": "[42] Zeyu Xiao, Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. " + }, + { + "bbox": [ + 307, + 186, + 545, + 231 + ], + "type": "inline_equation", + "content": "\\text{Eva}^2" + }, + { + "bbox": [ + 307, + 186, + 545, + 231 + ], + "type": "text", + "content": ": Event-assisted video frame interpolation via cross-modal alignment and aggregation. IEEE Transactions on Computational Imaging, 8:1145-1158, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 234, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 234, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 234, + 545, + 277 + ], + "type": "text", + "content": "[43] Zeyu Xiao, Zhiwei Xiong, Xueyang Fu, Dong Liu, and Zheng-Jun Zha. Space-time video super-resolution using temporal profiles. In Proceedings of the 28th ACM International Conference on Multimedia, pages 664-672, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 280, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 280, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 280, + 545, + 323 + ], + "type": "text", + "content": "[44] Bin Yang, Zhenhong Jia, Jie Yang, and Nikola K Kasabov. Video snow removal based on self-adaptation snow detection and patch-based gaussian mixture model. IEEE Access, 8:160188-160201, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 326, + 545, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 326, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 307, + 326, + 545, + 390 + ], + "type": "text", + "content": "[45] Yijun Yang, Angelica Aviles-Rivero, Huazhu Fu, Ye Liu, Weiming Wang, and Lei Zhu. Video adverse-weather-component suppression network via weather messenger and adversarial backpropagation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 394, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 449 + ], + "type": "text", + "content": "[46] Tian Ye, Sixiang Chen, Yun Liu, Yi Ye, Erkang Chen, and Yuche Li. Underwater light field retention: Neural rendering for underwater imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 488-497, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 452, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 452, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 452, + 545, + 495 + ], + "type": "text", + "content": "[47] Zongsheng Yue, Jianwen Xie, Qian Zhao, and Deyu Meng. Semi-supervised video deraining with dynamical rain generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 642-652, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 498, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 563 + ], + "type": "text", + "content": "[48] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 566, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 566, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 566, + 545, + 620 + ], + "type": "text", + "content": "[49] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 623, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 623, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 623, + 545, + 666 + ], + "type": "text", + "content": "[50] Kaihao Zhang, Rongqing Li, Yanjiang Yu, Wenhan Luo, and Changsheng Li. Deep dense multi-scale network for snow removal using semantic and depth priors. IEEE Transactions on Image Processing, 30:7419-7431, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[51] Ruofan Zhang, Jinjin Gu, Haoyu Chen, Chao Dong, Yulun Zhang, and Wenming Yang. Crafting training degradation distribution for the accuracy-generalization trade-off. arXiv preprint arXiv:2305.18107, 2023." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13221" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 108 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 108 + ], + "type": "text", + "content": "[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018." + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13222" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SoDaCam_ Software-defined Cameras via Single-Photon Imaging/ce1d43a9-f5c5-4d36-93a7-015029e926c0_content_list.json b/2023/SoDaCam_ Software-defined Cameras via Single-Photon Imaging/ce1d43a9-f5c5-4d36-93a7-015029e926c0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6970e54261de66541d5ef8668dc14ace8fb02e4c --- /dev/null +++ b/2023/SoDaCam_ Software-defined Cameras via Single-Photon Imaging/ce1d43a9-f5c5-4d36-93a7-015029e926c0_content_list.json @@ -0,0 +1,2393 @@ +[ + { + "type": "text", + "text": "SoDaCam: Software-defined Cameras via Single-Photon Imaging", + "text_level": 1, + "bbox": [ + 153, + 130, + 818, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Varun Sundar†", + "bbox": [ + 207, + 180, + 326, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "vsundar4@cs.wisc.edu", + "bbox": [ + 176, + 202, + 357, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Andrei Ardelean†", + "bbox": [ + 418, + 180, + 558, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a.ardelean@epfl.ch", + "bbox": [ + 408, + 202, + 570, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tristan Swedish", + "bbox": [ + 637, + 180, + 774, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tristan@ubicept.com", + "bbox": [ + 620, + 202, + 792, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Claudio Bruschini†", + "bbox": [ + 215, + 223, + 369, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{claudio.bruschini, edoardo.charbon}@epfl.ch", + "bbox": [ + 184, + 244, + 570, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mohit Gupta†,§", + "bbox": [ + 640, + 222, + 763, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mohitg@cs.wisc.edu", + "bbox": [ + 620, + 244, + 784, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "†University of Wisconsin-Madison ‡École polytechnique fédérale de Lausanne §Ubicept", + "bbox": [ + 122, + 266, + 844, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 320, + 313, + 335 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reinterpretable cameras are defined by their post-processing capabilities that exceed traditional imaging. We present \"SoDaCam\" that provides reinterpretable cameras at the granularity of photons, from photon-cubes acquired by single-photon devices. Photon-cubes represent the spatio-temporal detections of photons as a sequence of binary frames, at frame-rates as high as $100\\mathrm{kHz}$ . We show that simple transformations of the photon-cube, or photon-cube projections, provide the functionality of numerous imaging systems including: exposure bracketing, flutter shutter cameras, video compressive systems, event cameras, and even cameras that move during exposure. Our photon-cube projections offer the flexibility of being software-defined constructs that are only limited by what is computable, and shot-noise. We exploit this flexibility to provide new capabilities for the emulated cameras. As an added benefit, our projections provide camera-dependent compression of photon-cubes, which we demonstrate using an implementation of our projections on a novel compute architecture that is designed for single-photon imaging.", + "bbox": [ + 75, + 352, + 473, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 681, + 209, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Throughout the history of imaging, sensing technologies and the corresponding processing have developed hand-in-hand. In fact, sensing technologies have, to some extent, defined the scope of processing captured data. In the film era, instances of such processing included dodging and burning. The advent of digital cameras provided processing at the granularity of pixels and paved the way for modern computer vision. Light field cameras [34, 78], by sampling the plenoptic function [2], allowed post-capture processing at", + "bbox": [ + 75, + 705, + 468, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the granularity of light rays, enabling novel functionalities such as refocusing photos after-capture. The logical limit of post-capture processing, given the fundamental quantization of light, would be at the level of individual photons. What would imaging look like if we could perform computational processing on individual photons?", + "bbox": [ + 500, + 321, + 890, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we show that photon data captured by a new class of single-photon detectors, called single-photon avalanche diodes (SPADs), makes it possible to emulate a wide range of imaging modalities such as exposure bracketing [12], video compressive systems [38, 55] and event cameras [52, 60]. A user then has the flexibility to choose one (or even multiple) of these functionalities post-capture (Fig. 1 (top)). SPAD arrays can operate as extremely high frame-rate photon detectors ( $\\sim 100\\mathrm{kHz}$ ), producing a temporal sequence of binary frames called a photon-cube [16]. We show that computing photon-cube projections, which are simple linear and shift operations, can reinterpret the photon-cube to achieve novel post-capture imaging functionalities in a software-defined manner (Fig. 1 (middle)).", + "bbox": [ + 496, + 412, + 892, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As case studies, we emulate three distinct imagers: high-speed video compressive imaging; event cameras which respond to dynamic scene content; and motion projections which emulate sensor motion, without any real camera movement. Fig. 1 (bottom) shows the outputs of these cameras that are derived from the same photon-cube.", + "bbox": [ + 496, + 625, + 890, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Computing photon-cube projections. One way to obtain photon-cube projections is to read the entire photon-cube off the SPAD array and then perform relevant computations off-chip; we adopt this strategy for our experiments in Secs. 6.1 and 6.2. While reasonable for certain applications, reading out photon-cubes requires an exorbitant data-bandwidth, which can be up to 100 Gbps for a 1 MPixel array—well beyond the capacity of existing data peripherals. Such readout considerations will become center stage as large-format SPAD arrays are fabricated [48, 49].", + "bbox": [ + 496, + 734, + 890, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "An alternative is to avoid transferring the entire photon-", + "bbox": [ + 519, + 886, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: Varun Sundar. This research was supported in parts by NSF CAREER award 1943149, NSF award CNS-2107060, and the Swiss National Science Foundation grant 200021_166289. We also thank Paul Mos for providing access to SwissSPAD2 acquisition software.", + "bbox": [ + 75, + 851, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8165", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ac3cc844e8b4994e3c96d64b90cda8bef579a768f693e45dc790917fe510e938.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 88, + 890, + 253 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/517d947cd97d8ef1890b948664efad8ff9c7df797efb8f796c851fc0a82dba69.jpg", + "image_caption": [ + "Capabilities of photon-cube projections" + ], + "image_footnote": [], + "bbox": [ + 84, + 255, + 890, + 415 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d7196f7c9edadd2ca73635dcac2fe8f2ff409c3d5295acaaa5ae0b6845806c80.jpg", + "image_caption": [ + "Coded exposure" + ], + "image_footnote": [], + "bbox": [ + 84, + 446, + 267, + 518 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dcb2d461ccca256ef02b4a145c5a5ee01e3c3521a0e828babfb30d37ed461e52.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 284, + 452, + 354, + 515 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1f7993e7b5524f5c2010889f3c15273ef20bac7a4baafbbd6f137b6d25424784.jpg", + "image_caption": [ + "$12 \\times$ video reconstruction (raster order) Snapshot compressive imager (Sec. 5.1)" + ], + "image_footnote": [], + "bbox": [ + 84, + 539, + 357, + 612 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e52254981b0ce021b1dc902ea604da4da8a713e730883b34b1ab21a327e82312.jpg", + "image_caption": [ + "3D scatter plot, $0\\leq t\\leq 0.1s$" + ], + "image_footnote": [], + "bbox": [ + 390, + 445, + 576, + 518 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c398e22cbebc862536d57707aec275a589f0709aa54198923293d85b6757594d.jpg", + "image_caption": [ + "Event image, $0.5 \\leq t \\leq 0.53$ s Event camera (Sec. 5.2)", + "Figure 1: (top) SoDaCam can emulate a variety of cameras from the photon-cubes acquired by single-photon devices. (middle) Photon-cubes represent the spatio-temporal detection of photons as a sequence of binary frames. Projections of the photon-cube, when computed either on or off-chip, result in reinterpretable and software-defined cameras. We demonstrate the versatility of photon-cube projections on a real dynamic scene: a die falls on a table, bounces, spins in the air and later ricochets off a nearby toy top. (bottom) The cameras emulated by our photon-cube projections can produce: a $12 \\times$ high-speed video from a single compressive snapshot, event-stream representations of two time intervals (blue and red depict positive and negative spikes respectively), an image where the die appears stationary, as well as a motion-deblurred image." + ], + "image_footnote": [], + "bbox": [ + 388, + 539, + 576, + 612 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/094aa9f7ec7d9cf191839e995fc0d270dba4e4c82d208c93b1925c638e477261.jpg", + "image_caption": [ + "Static camera" + ], + "image_footnote": [], + "bbox": [ + 607, + 446, + 736, + 518 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8c24a2275f618d791908dd8fee6b565343c6620eb9310bf98410846f85bf3cd7.jpg", + "image_caption": [ + "Parabolic trajectory Motion Pro" + ], + "image_footnote": [], + "bbox": [ + 607, + 540, + 736, + 612 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8bd7d4db4e90149d22171812ec99872fdd68a2f1a432339e077df1406e4e1e8c.jpg", + "image_caption": [ + "Linear trajectory" + ], + "image_footnote": [], + "bbox": [ + 756, + 446, + 885, + 518 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/effe023a8ca759e86bd7881337a3dcf4ea424fcc72fd6118da0fb4f2da2cc3b2.jpg", + "image_caption": [ + "Parabolic deblurred" + ], + "image_footnote": [], + "bbox": [ + 756, + 540, + 885, + 612 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cube by computing projections near sensor. As a proof-of-concept, we implement photon-cube projections on Ultra-Phase [5], a recently-developed programmable SPAD imager with independent processing cores that have dedicated RAM and instruction memory. We show, in Sec. 6.3, that computing projections on-chip greatly reduces sensor readout and, as a consequence, power consumption.", + "bbox": [ + 75, + 784, + 468, + 890 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implications: Toward a photon-level software-defined camera. The photon-cube projections introduced in this paper are computational constructs that provide a realization of software-defined cameras or SoDaCam. Being software-defined, SoDaCam can emulate multiple cameras simultaneously without additional hardware complexity. SoDaCam, by going beyond baked in hardware choices,", + "bbox": [ + 496, + 784, + 890, + 890 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8166", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "unlocks hitherto unseen capabilities—such as 2000 FPS video from $25\\mathrm{Hz}$ readout (Fig. 7); event imaging in very low-light conditions (Fig. 9); and motion stacks, which are a stack of images where in each image, objects only in certain velocity ranges appear sharp (Fig. 6).", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Limitations. The SPAD array [67] used in this work has a relatively low spatial resolution $(512 \\times 256)$ , and a low fill-factor $(\\sim 10\\%)$ owing to the lack of microlenses in the prototype used. Similarly, the near-sensor processor that we use has limited capabilities compared to off-chip processors. However, with rapid progress in the development of single-photon cameras [48, 49] and increasing interest in near-sensor processors, we anticipate that many of these shortcomings will be addressed in the upcoming years.", + "bbox": [ + 75, + 185, + 470, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 335, + 218, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reinterpretable imaging has previously been explored at the granularity of light rays [3], by modulating the plenoptic function, and at the level of spatio-temporal voxels [20], by using fast per-pixel shutters. SoDaCam represents a logical culmination of reinterpretability at the level of photon detections, that facilitates multiple post-capture imaging functionalities.", + "bbox": [ + 75, + 359, + 468, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Programmable imaging using a digital micromirror device was first introduced in Nayar et al. [51] to perform precapture radiometric manipulations. Modern programmable cameras are typically near-sensor processors [8, 69, 71] that can perform limited operations in analog [46, 55, 69], while more complex operations [9, 45] occur after analog-to-digital conversion (ADC). In contrast, by performing post-capture computations directly on photon detections, we can perform complex operations without incurring the read-noise penalty that is associated with ADC.", + "bbox": [ + 75, + 484, + 468, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Passive single-photon imaging. Only recently have SPADs been utilized as passive imaging devices, with applications in high-dynamic range imaging [29, 30, 37, 50], motion-compensation [31, 58], burst photography [41, 42] and object tracking [22]. Compared to compute-intensive burst-photography methods [41], our proposed techniques involve lightweight computations that can be performed near sensor. These computations can also be performed using other single-photon imagers such as Jots [15, 17], which feature higher sensor resolution and photon-efficiency [40], albeit at lower frame-rates and higher read noise.", + "bbox": [ + 75, + 654, + 468, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reducing the readout of SPADs. Several data reduction strategies have been proposed in the context of SPADs that are used to timestamp incident photons, including: coarse histograms [13, 23, 56], compressive histograms [21], and", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "measuring differential time-of-arrivals [70, 77]. When SPADs are operated as photon-detectors, multi-bit counting [48], or summing binary frames, can reduce readout. While compression is not our main objective, we show that photon-cube projections act as camera-specific compression schemes that dramatically reduce sensor readout.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background: Single-Photon Imaging Model", + "text_level": 1, + "bbox": [ + 498, + 196, + 890, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A SPAD array captures incident light as a photon-cube: a temporal sequence of binary frames that represents the pixel-wise detection of photons across their respective exposure windows. We can model the stochastic arrival of photons as a Poisson process [72], allowing us to treat spatio-temporal values of the photon-cube as independent Bernoulli random variables with", + "bbox": [ + 496, + 223, + 890, + 327 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Pr \\left\\{B _ {t} (\\mathbf {x}) = 1 \\right\\} = 1 - e ^ {- (\\eta \\Phi (\\mathbf {x}, t) + r _ {q}) w _ {\\exp}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 339, + 890, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $B_{t}(\\mathbf{x})$ represents the value of the photon-cube at pixel $\\mathbf{x}$ and exposure index $1 \\leq t \\leq T$ , which receives a mean incident flux of intensity $\\Phi (\\mathbf{x},t)$ across its exposure of duration $w_{\\mathrm{exp}}$ . Additionally, $\\eta$ is the photon detection efficiency of the SPAD, and $r_q$ denotes the sensor's dark count rate—which is the rate of spurious counts unrelated to incident photons. While individual binary frames are extremely noisy, the temporal sum of the photon-cube", + "bbox": [ + 496, + 369, + 890, + 492 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {\\text {s u m}} (\\mathbf {x}) := \\sum_ {t = 1} ^ {T} B _ {t} (\\mathbf {x}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 503, + 890, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "can produce an 'image' of the scene that is sharp in static regions, but blurry in dynamic regions (Fig. 2 (top)). Indeed, in static regions, the sum-image can be used to derive a maximum likelihood estimator of the scene intensity [4], given by $\\hat{\\Phi} (\\mathbf{x}) = -\\ln (1 - T^{-1}\\mathcal{I}_{\\mathrm{sum}}(\\mathbf{x})) / \\eta w_{\\mathrm{exp}} - r_q / w_{\\mathrm{exp}}$", + "bbox": [ + 496, + 556, + 890, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Projections of the Photon-Cube", + "text_level": 1, + "bbox": [ + 498, + 647, + 787, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The temporal sum described in Eq. (2) is a simple instance of projections of a photon-cube. Our key observation is that it is possible to compute a wide range of photon-cube projections, each of which emulates a unique sensing modality post-capture—including modalities that are difficult to achieve with conventional cameras. For example, varying the number of bit-planes that are summed over emulates exposure bracketing [12, 43], which is typically used for HDR imaging. Compared to conventional exposure bracketing, the emulated exposure stack, being software-defined, does not require spatial and temporal registration, which can often be error-prone. Fig. 2 (top) shows an example of an exposure stack computed from a photon-cube.", + "bbox": [ + 496, + 672, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Going further, we can gradually increase the complexity of the projections. For example, consider a coded exposure", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8167", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/88db00ad5a71bae5abb411466fd9c30ff92884915b3f654f9556302ee6ee7913.jpg", + "image_caption": [ + "Short exposure" + ], + "image_footnote": [], + "bbox": [ + 84, + 88, + 197, + 162 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b21ff920af31b12a2ecd1327561bfbdc5f5b85a5c4b93f69209d9fa6d380248d.jpg", + "image_caption": [ + "Medium exposure Exposure Stack" + ], + "image_footnote": [], + "bbox": [ + 209, + 88, + 323, + 162 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ce91a8a8d2d809a185b9730f0d6e542eea36b63f8784704cdc8972c7b21f1090.jpg", + "image_caption": [ + "Long exposure" + ], + "image_footnote": [], + "bbox": [ + 334, + 88, + 450, + 162 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bcdae6bd9a273b24708fe8f67fbac54a45d79cd3f0f19bcc69765cf09a74220f.jpg", + "image_caption": [ + "Spatially-varying exposure" + ], + "image_footnote": [], + "bbox": [ + 83, + 199, + 197, + 273 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1ddfa74e7cdeb3cff1045a6a45807535dd0f29d46ce77d71a177dd9d542a59fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 198, + 300, + 248 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/59c3c43b9ed1c36873b8e5511fae2fa0c159cfdf2fa3fdb72c7a8ac9f34f4962.jpg", + "image_caption": [ + "Figure 2: Coded exposures from photon-cubes. (top) An exposure stack with sum-images computed using 250, 500, and 1000 bit-planes. Short exposures are noisy while long exposures exhibit motion-blur. (bottom) Spatially-varying exposure that uses a quad pattern [32] (see inset), and a video compressive frame that uses 16 random binary masks to modulate the photon-cube. Zoom-in to see details." + ], + "image_footnote": [], + "bbox": [ + 233, + 250, + 299, + 299 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/348b598fcc997840b7770eb9f2714bfe7e8a29982af3ca7db9b49972480c320d.jpg", + "image_caption": [ + "Video compressive frame" + ], + "image_footnote": [], + "bbox": [ + 334, + 199, + 450, + 275 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "projection that multiplexes bit-planes with a temporal code", + "bbox": [ + 76, + 449, + 465, + 465 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {\\text {f l u t t e r}} (\\mathbf {x}) := \\sum_ {t = 1} ^ {T} C _ {t} B _ {t} (\\mathbf {x}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 472, + 468, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $C_t$ is the temporal code. An example of globally-coded exposures is the flutter shutter camera [53], which uses pseudo-random binary codes for motion-deblurring.", + "bbox": [ + 75, + 521, + 468, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More general coded exposures can be obtained via spatially-varying temporal coding patterns $C_t(\\mathbf{x})$ :", + "bbox": [ + 75, + 566, + 468, + 595 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {\\text {c o d e d}} (\\mathbf {x}) := \\sum_ {t = 1} ^ {T} C _ {t} (\\mathbf {x}) B _ {t} (\\mathbf {x}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 604, + 468, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 2 (bottom) shows spatially-varying exposures that use a quad (Bayer-like) spatial pattern and random binary masks. With photon-cubes, we can perform spatially-varying coding without bulky spatial light modulators, similar to focalplane sensor-processors [46, 69]. Moreover, we can capture multiple coded exposures simultaneously, which is challenging to realize in existing sensors. In Sec. 5.1, we describe coding patterns for video compressive sensing.", + "bbox": [ + 75, + 652, + 468, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatial and temporal gradients form the building blocks of several computer vision algorithms [7, 11, 24, 26, 39]. Given this, another projection of interest is temporal contrast, i.e., a derivative filter preceded by a smoothing filter:", + "bbox": [ + 75, + 773, + 468, + 834 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {\\text {c o n t r a s t}} (\\mathbf {x}, t) := D _ {t} \\circ G * B _ {t} (\\mathbf {x}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 849, + 468, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D_{t}$ is the difference operator, $G$ could be exponential or Gaussian smoothing, $\\circ$ denotes function composition,", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and * denotes convolution. Due to their sparse nature, gradients form the basis of bandwidth- and power-efficient event cameras [10, 14, 36, 60], which we emulate in Sec. 5.2.", + "bbox": [ + 498, + 90, + 890, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "So far, we have considered projections taken only along the time axis. Next, we consider a more general class of spatio-temporal projections that lead to novel functionalities. For instance, computing a simple projection, such as the temporal sum, along arbitrary spatio-temporal directions emulates sensor motion during exposure time [33], but without moving the sensor. We achieve this by shifting bitplanes and computing their sum:", + "bbox": [ + 498, + 136, + 890, + 257 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {\\text {s h i f t}} (\\mathbf {x}) := \\sum_ {t = 1} ^ {T} B _ {t} (\\mathbf {x} + \\mathbf {r} (t)), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 266, + 890, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{r}$ is a discretized 2D trajectory that determines sensor motion. Outside a software-defined framework, such projections are hard to realize without physical actuators. We describe the capabilities of motion projections in Sec. 5.3.", + "bbox": [ + 496, + 316, + 890, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, the proposed photon-cube projections are simple linear and shift operators that lead to a diverse set of post-capture imaging functionalities. These projections pave the way for future 'swiss-army-knife' imaging systems that achieve multiple functionalities (e.g., event cameras, high-speed cameras, conventional cameras, HDR cameras) simultaneously with a single sensor. Finally, these projections can be computed efficiently in an online manner, which makes on-chip implementation viable (Sec. 6.3).", + "bbox": [ + 496, + 377, + 890, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At this point, we note that a key enabling factor of photon-cube projections is the extremely high temporal-sampling rate of SPADs. Indeed, the temporal sampling rate determines key aspects of sensor emulation, such as the discretization of temporal derivatives and motion trajectories. This raises a natural question: can we use conventional high-speed cameras for computing projections?", + "bbox": [ + 496, + 512, + 890, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Trade-off between frame-rate and SNR. In principle, photon-cube projections can be computed using regular (CMOS or CCD based) high-speed cameras. Unfortunately, each frame captured by a high-speed camera incurs a read-noise penalty, which increases with the camera's frame-rate [6]. In fact, the read noise levels of high-speed cameras [1] can be $10 - 30 \\times$ higher than consumer cameras [28]. Coupled with the low per-frame incident flux at high frame-rates, high levels of read noise result in extremely low SNRs. In contrast, SPADs do not incur a per-frame read noise and are limited only by the fundamental photon noise. Hence, for the post-capture software-defined functionalities proposed here, it is imperative to use SPADs.", + "bbox": [ + 496, + 635, + 890, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Emulating Cameras from Photon-Cubes", + "text_level": 1, + "bbox": [ + 498, + 844, + 861, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sec. 4 presented the concept of photon-cube projections, and its potential for achieving multiple post-capture imag-", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8168", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a2900e485e032033d45c703a41e5a3da27819274f01258de20b1e17687313898.jpg", + "image_caption": [ + "Figure 3: Modulating masks for video compressive sensing. (left) A single VCS measurement temporally compresses a sequence of frames using binary random masks. (center) Two-bucket cameras capture an additional measurement by using the complementary mask sequence. (right) We propose using multi-bucket captures by randomly choosing an active bucket for each frame. Both two-bucket and multi-bucket captures have $100\\%$ light efficiency. All masks are visualized here for $16 \\times 16$ pixels." + ], + "image_footnote": [], + "bbox": [ + 80, + 92, + 196, + 204 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/329176052390257c785be133c10c0d93034f35f2074c8fdefd7de845acc46489.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 207, + 90, + 326, + 204 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1ec8b5fdaec299ac3fa57af842c242c1b4b073e83881f95d970329dd37809727.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 88, + 465, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing functionalities. As case studies, we now demonstrate three imaging modalities: video compressive sensing, event cameras, and motion-projection cameras. These modalities have been well-studied over several years; in particular, there exist active research communities around video compressive sensing and event cameras today. We also show new variants of these imaging systems that arise from the software-defined nature of photon-cube projections.", + "bbox": [ + 75, + 380, + 467, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Video Compressive Sensing", + "text_level": 1, + "bbox": [ + 76, + 513, + 323, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video compressive systems optically multiplex light with random binary masks, such as the patterns in Fig. 3 (left). As discussed in the previous section, such multiplexing can be achieved computationally using photon-cubes.", + "bbox": [ + 75, + 537, + 468, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Two-bucket cameras. One drawback of capturing coded measurements is the light loss due to blocking of incident light. To prevent loss of light, coded two-bucket cameras [69] capture an additional measurement that is modulated by the complementary mask sequence (Fig. 3 (center)). Such measurements recover higher quality frames, even after accounting for the extra readout [35, 61]. Two-bucket captures can be readily derived from photon-cubes, by implementing Eq. (4) with the additional mask sequence.", + "bbox": [ + 75, + 619, + 467, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-bucket cameras. We can extend the idea of two-bucket captures to multi-bucket captures by accumulating bit-planes in one of $k$ buckets that is randomly chosen at each time instant and pixel location. Since multiplexing is performed computationally, we do not face any loss in photoreceptive area that [59, 68] hampers existing multi-bucket sensors. Multi-bucket captures can reconstruct a large number of frames by better conditioning video recovery and", + "bbox": [ + 75, + 779, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fdeb5aad951fe92184d1b86c9c064d8f9671f1bff55af3ce9963cd79b23afca2.jpg", + "image_caption": [ + "$\\log (x)$ vs $1 - e^{-x}$" + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 669, + 179 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0f2f2a0ea4d297cf3dff34efd3a939b2da3654b3a0c873c36ada8fe8ac5c8fb4.jpg", + "image_caption": [ + "Spatio-temporal plot" + ], + "image_footnote": [], + "bbox": [ + 669, + 90, + 870, + 179 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6d6626c534dae7f70bd1c71edb0d94ca56af365011cd170ff6542b41a3c960e0.jpg", + "image_caption": [ + "$\\tau = 0.45, \\beta = 0.95$", + "5 MEvent/s", + "Effect of contrast-threshold, exponential decay", + "Figure 4: Event stream from photon-cubes. (top left) By exploiting the non-linear response curve of SPADs to encode brightness, we can avoid the underflow issues of a log-response. We visualize events generated from photon-cubes using a 3D scatter plot of polarities (top right, 14000 bit-planes), and frame accumulation of events (bottom, 1200 bit-planes). Blue and red denote positive and negative spikes respectively. The event images also show the effect of varying the contrast threshold $\\tau$ and exponential decay $\\beta$ -larger values yield a less noisy but sparser event stream." + ], + "image_footnote": [], + "bbox": [ + 506, + 205, + 627, + 273 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/de92179dc5d733509ee24dfe131bba7dd85d3d05a3af8b7033c6657f06ca72c7.jpg", + "image_caption": [ + "$\\tau = 0.25, \\beta = 0.95$", + "25 MEvent/s" + ], + "image_footnote": [], + "bbox": [ + 635, + 205, + 756, + 273 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/06734ec02cad1a28bf0b71f3b94d69fa076e60210fcdc765fed2634d9606b14b.jpg", + "image_caption": [ + "$\\tau = 0.45, \\beta = 0.98$", + "2.5 MEvent/s" + ], + "image_footnote": [], + "bbox": [ + 763, + 207, + 883, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "provide extreme high-speed video imaging. Fig. 3 (right) shows the modulating masks for a four-bucket capture.", + "bbox": [ + 498, + 511, + 890, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2.Event Cameras", + "text_level": 1, + "bbox": [ + 500, + 553, + 653, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we describe emulation of event-cameras, which capture changes in light intensity and are conceptually similar to the temporal contrast projection introduced in Eq. (5). Physical implementations of event sensors [10, 14, 36, 60] generate a photoreceptor voltage $V(\\mathbf{x}, t)$ with a logarithmic response to incident flux $\\Phi(\\mathbf{x}, t)$ , and output an event $(\\mathbf{x}, t, p)$ when this voltage deviates sufficiently from a reference voltage $V_{\\mathrm{ref}}(\\mathbf{x})$ :", + "bbox": [ + 496, + 578, + 890, + 700 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left| V (\\mathbf {x}, t) - V _ {\\text {r e f}} (\\mathbf {x}) \\right| > \\tau , \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 712, + 890, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ is called the contrast-threshold and $p = \\mathrm{sign}(V(\\mathbf{x},t) - V_{\\mathrm{ref}}(\\mathbf{x}))$ encodes the polarity of the event. Once an event is generated, $V_{\\mathrm{ref}}(\\mathbf{x})$ is updated to $V(\\mathbf{x},t)$ . Eq. (7), for a smoothly-varying flux intensity, thresholds a function of the temporal gradient, i.e., $\\partial_t\\log (\\Phi (\\mathbf{x},t))$", + "bbox": [ + 496, + 742, + 890, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From bit-planes to event streams. To produce events from SPAD frames, we compute an exponential moving average (EMA) of the bit-planes, as $\\mu_t(\\mathbf{x}) = (1 - \\beta)B_t(\\mathbf{x}) + \\beta \\mu_{t - 1}(\\mathbf{x})$ where $\\mu_t(\\mathbf{x})$ is the EMA, $\\beta$ is the smoothing", + "bbox": [ + 496, + 839, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8169", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e6656c4cc37d7fcdd415afb7de28cc059369c90e3c76c3dc0a9ae037f0872d8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 88, + 263, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e0a08c526f44f3d877f4c3f47dae2342229790c25a9af49896fbf19b92bc213d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 160, + 256, + 204 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4f07bcadec453ebbbb6d8c55fd2d414878ed1f3e92d89c7ac201bb81c330f8d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 88, + 464, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fc72fb42f026bf3e2e660bf720973e961cff5ae13bf1eb0a86b5a434c3e1574a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 284, + 159, + 457, + 205 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c1069f1838814529b3a092bebf6aa903e55fbb56a6bf29b0c91199a5ba9c0ec3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 210, + 263, + 281 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7e017e1f64b6f1df9f7f6122d4c06c01395bd01338fba431a34cfe05f8789365.jpg", + "image_caption": [ + "Parabolic trajectory" + ], + "image_footnote": [], + "bbox": [ + 83, + 282, + 256, + 313 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f5fdd82605fb6883608f0bde223d092297f6e17b647864f20f3c0f02bc4d5689.jpg", + "image_caption": [ + "Figure 6: Motion stack. Computing multiple linear projections with different trajectories can produce a stack of images where objects with matching velocity are sharp. Here, we show a traffic scene involving four cars that have four different velocities. By suitably altering the slope of the linear trajectory, we can produce images where only one of the cars appear sharp at a time. We indicate the slope of the trajectories chosen and the objects that are \"in-focus\"." + ], + "image_footnote": [], + "bbox": [ + 282, + 210, + 464, + 281 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6d9b5bad09bb7c41561d2b68c602b314790e14ef4ff4e33bf35d06198d26e7e7.jpg", + "image_caption": [ + "Parabolic deblurred", + "Figure 5: Motion projections. (top) Integrating along a linear trajectory in the photon-cube changes the apparent image-space velocity of scene objects. Details are seen for (top left) the case when static, and (top right) the metallic tape when the sensor translates along the $x$ -axis. (bottom) A parabolic integration trajectory results in a motion-invariant image, resulting in similar blur kernels for all objects. (bottom right) Deblurring with the resultant shift-invariant point spread function (shown in inset) produces a sharp image." + ], + "image_footnote": [], + "bbox": [ + 312, + 284, + 421, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "factor, and $B_{t}$ is a bit-plane. We generate an event when $\\mu_t(\\mathbf{x})$ deviates from $\\mu_{\\mathrm{ref}}(\\mathbf{x})$ by at least $\\tau$ :", + "bbox": [ + 75, + 508, + 468, + 541 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left| h \\left(\\mu_ {t} (\\mathbf {x})\\right) - h \\left(\\mu_ {\\text {r e f}} (\\mathbf {x})\\right) \\right| > \\tau , \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 551, + 468, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $h$ is a scalar function applied to the EMA. We can see that Eq. (8) thresholds temporal contrast, by observing the role played by the EMA and the difference operator.", + "bbox": [ + 75, + 579, + 468, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Setting $h$ to be the logarithm of the flux MLE mimics Eq. (7). However, since the log-scale is used to prevent sensor saturation, a simpler alternative is to use the non-saturating response curve of SPAD pixels $(h(x) = x)$ . The response curve takes the form of $1 - \\exp (-\\alpha \\Phi (\\mathbf{x},t))$ , where $\\alpha$ is a flux-independent constant. As a major advantage, this response curve avoids the underflow issues of the log function that can occur in low-light scenarios [62].", + "bbox": [ + 75, + 626, + 468, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The SPAD's frame-rate determines the time-stamp resolution of emulated events. In Fig. 4, we show the events generated from a photon-cube acquired at a frame-rate of $96.8\\mathrm{kHz}$ resulting in a time-stamp resolution of $\\sim 10~\\mu \\mathrm{s}$ that is comparable to those of existing event cameras.", + "bbox": [ + 75, + 746, + 468, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "How do SPAD-events differ from the output of a regular event camera? The main difference is the expression of temporal contrast, given by $\\partial_t h$ , is now $-\\partial_t \\exp(-\\alpha \\Phi(\\mathbf{x}, t))$ , instead of $\\partial_t \\log(\\Phi(\\mathbf{x}, t))$ . This dif", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1709c7778e3cf88a2a004695e25143a4ad7fa0a2fd2d5f915e20ffbaa82ef9bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 88, + 697, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1c41b3c328fb8d6d2a989938fa554a706049d4f82ec3037b4759f94902af11c0.jpg", + "image_caption": [ + "$160\\hat{x} +447\\hat{y}$ pixels/sec" + ], + "image_footnote": [], + "bbox": [ + 511, + 186, + 697, + 261 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d4a1b9cda9e01fe1ab01df568296f6ea507966676fdf96c631718e543222b385.jpg", + "image_caption": [ + "$100\\hat{x} + 282\\hat{y}$ pixels/sec" + ], + "image_footnote": [], + "bbox": [ + 702, + 88, + 887, + 164 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f60191071b30b5e4fa35b5ee6cd7761d55a5f580686286e62f6ac2af1c7b5ce5.jpg", + "image_caption": [ + "$420\\hat{x} + 1177\\hat{y}$ pixels/sec" + ], + "image_footnote": [], + "bbox": [ + 702, + 186, + 887, + 261 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ference poses no compatibility issues for a large class of event-vision algorithms that utilize a grid of events [54, 57, 66] or brightness changes [18]. We show examples of downstream applications using SPAD-events in Supplementary Note 2. Finally, SPAD-events can be easily augmented with spatially- and temporally-aligned intensity information—a synergistic combination that has been exploited by several recent event-vision works [18, 25, 74].", + "bbox": [ + 496, + 436, + 890, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Motion Projections", + "text_level": 1, + "bbox": [ + 500, + 566, + 683, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Having described the emulation of cameras that capture coded exposures and temporal contrasts, we now shift our attention to cameras that emulate sensor motion during exposure, viz. motion cameras. We describe two useful trajectories when emulating motion cameras using Eq. (6).", + "bbox": [ + 496, + 590, + 890, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Linear trajectory. The simplest sensor trajectory involves linear motion, where $\\mathbf{r}(t) = (bt + c)\\hat{\\mathbf{p}}$ for some constants $b, c \\in \\mathbb{R}$ and unit vector $\\hat{\\mathbf{p}}$ . As Fig. 5 (top row) shows, this can change the scene's frame of reference: making moving objects appear stationary and vice-versa.", + "bbox": [ + 496, + 685, + 890, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Motion-invariant parabolic projection. If motion is along $\\hat{\\mathbf{p}}$ , parabolic integration produces a motion-invariant image [33]—all objects, irrespective of their velocity, are blurred by the same point spread function (PSF), up to a linear shift. Thus, a deblurred parabolic capture produces a sharp image of all velocity groups (Fig. 5 (bottom row)). The parabolic trajectory is given by $\\mathbf{r}(t) = (at^2 + bt + c)\\hat{\\mathbf{p}}$ . We choose $a$ based on the maximum object velocity and $b, c$", + "bbox": [ + 496, + 779, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8170", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f3e5ed354caccbee3543c8c2950ba4e8c8629eadd0e5ca9b85c43134055c8fa5.jpg", + "image_caption": [ + "Figure 7: High-speed videography at 2000 FPS of a tennis ball dropped into a bowl of water, from a $25\\mathrm{Hz}$ readout. The conventional capture provides a visualization of the scene dynamics. It is challenging to reconstruct a large number of frames from a single compressive snapshot. Multi-bucket captures recover frames with significantly greater detail, such as the crown of water surrounding the ball. We include more sequences (e.g., a bursting balloon) in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 78, + 89, + 888, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b47c6bfabc84357280f9d81926085a1217888186ea2f7f273cda943887699918.jpg", + "image_caption": [ + "Figure 8: Deblurring of traffic scenes using motion projections. Linear projections can recover details of moving objects if their velocity is known. When only the motion direction is known (e.g., road's orientation), a sharp image can be obtained by either deblurring a parabolic projection or by blending multiple randomly-sampled linear projections. We quantitatively compare against the compute- and bandwidth-expensive Quanta Burst Photography [41], based on PSNR and LPIPS [76]." + ], + "image_footnote": [], + "bbox": [ + 78, + 309, + 890, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "so the parabola's vertex lies at $T / 2$ . We readily obtain the PSF by applying the parabolic integration to a delta input. Upon deconvolution using the PSF, a parabolic projection provides the optimal SNR for a blur-free image from single capture when only the direction of velocity is known.", + "bbox": [ + 75, + 628, + 470, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ensembling linear projections. Finally, we leverage the flexibility of photon-cubes to compute multiple linear projections, as seen in Fig. 6. This produces a stack of images where one velocity group is motion-blur free at a time—or a 'motion stack', analogous to a focal stack. This novel construct can be used to compensate motion by blending stack images using cues such as blur orientation or optical flow.", + "bbox": [ + 75, + 724, + 468, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Hardware and Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 845, + 408, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We design a range of experiments to demonstrate the versatility of photon-cube projections: both when computa", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tions occur after readout (Secs. 6.1 and 6.2), and when they are performed near-sensor on-chip (Sec. 6.3). All photon-cubes were acquired using the SwissSPAD2 array [67], operated using one of two sub-arrays, each having $512 \\times 256$ pixels, and at a frame-rate of $96.8\\mathrm{kHz}$ . For the on-chip experiments, we use the UltraPhase compute architecture to interface with photon-cubes acquired by the SwissSPAD2.", + "bbox": [ + 496, + 628, + 890, + 737 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. SoDaCam Capabilities", + "text_level": 1, + "bbox": [ + 500, + 753, + 710, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "High-speed compressive imaging. We reconstruct 80 frames from compressive snapshots that are emulated at $25\\mathrm{Hz}$ , resulting in a 2000 FPS video. We decode compressive snapshots using a plug-and-play (PnP) approach, PnP-FastDVDNet [73]. As Fig. 7 shows, it is challenging to recover a large number of frames from a single compressive measurement. Using the proposed multi-bucket scheme significantly improves the quality of video reconstruction.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8171", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/59f8080ba282867c4d22838a60bcb7bf1ec710645f571ad142bc47a111e1f536.jpg", + "image_caption": [ + "Sufficient Light 158 lux, 1.03 PPP" + ], + "image_footnote": [], + "bbox": [ + 109, + 125, + 179, + 178 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/88404c8016c9e7b8df7a345335bc5cddce2a9ea9cda51452cafd8b65fcf20dc6.jpg", + "image_caption": [ + "Low light 0.9 lux, 0.1 PPP" + ], + "image_footnote": [], + "bbox": [ + 238, + 125, + 308, + 178 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2e3dd248446e1dcf04390d0feac2a6f77fedf801d07f7225b49927d03b5adc91.jpg", + "image_caption": [ + "Very low light 0.3 lux, 0.03 PPP" + ], + "image_footnote": [], + "bbox": [ + 364, + 125, + 434, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3d6d717eafad9874614643283bd060c59b7b2269c31598f334f5ac790798c780.jpg", + "image_caption": [ + "SPAD-emulated events" + ], + "image_footnote": [], + "bbox": [ + 86, + 205, + 205, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f5aed45fe02cb7294012691749a1d12badfcc30b39ba6065ce34022d30d3b17c.jpg", + "image_caption": [ + "DSLR captures (30 Hz)" + ], + "image_footnote": [], + "bbox": [ + 212, + 205, + 331, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7aaf9f1ac84da73d2749050a24ba8720f6f533a418b11100f73bd01bd4526604.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 205, + 460, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/51a862585f198d34166b38a0ca7ba90c408291b25d248fa3ada9aa1b48592835.jpg", + "image_caption": [ + "Prophesee EVK4 events" + ], + "image_footnote": [], + "bbox": [ + 86, + 325, + 205, + 417 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/278f1ad503ec66b55ab202dfac536d76fedc1977f137bc3f23a353c99bdc5570.jpg", + "image_caption": [ + "Figure 9: Comparison to a state-of-the-art event camera. SPAD-events can capture temporal gradients even when the light-level is reduced by $500\\times$ , by benefiting from their single-photon sensitivity and bounded brightness response curve. In contrast, low-light induces blur and deteriorates the Prophesee's event stream. As a measure of the light-level, we report the PPP (photons per pixel) averaged across bit-planes and a light meter's reading at the sensor location." + ], + "image_footnote": [], + "bbox": [ + 212, + 325, + 331, + 417 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7f31fb95ba25496365b8b6c3062f4c2c2648bb3887b2ca3b0f8f9987def73c73.jpg", + "image_caption": [ + "Figure 10: Comparison against conventional high-speed acquisition at $4000\\mathrm{Hz}$ . (top) SPAD projections recover a $16\\times$ compressive video and an event image of a spinning roulette wheel. (middle) Read-noise corrupts the incident flux in the Infinicam high-speed camera, removing details in frames which are compressed on-the-fly. (bottom) Although using a larger aperture to admit more light recovers some detail, noise and compression artifacts still persist." + ], + "image_footnote": [], + "bbox": [ + 344, + 325, + 460, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While multi-bucket captures require more bandwidth, this can be partially amortized by coding only dynamic regions, which we show in Supplementary Note 1.", + "bbox": [ + 75, + 598, + 468, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Motion projections on a traffic scene. Fig. 8 shows two traffic scenes captured using a $50~\\mathrm{mm}$ focal length lens and at $30\\mathrm{Hz}$ emulation. When object velocity is known, a linear projection can make moving objects appear stationary. If only the velocity direction is known (e.g., road's orientation in Fig. 8), a parabolic projection provides a sharp reconstruction of all objects. We deblur parabolic captures using PnP-DnCNN [75]. We offer an improvement by randomly sampling 8 linear projections along the velocity direction and blending them using the optical flow predicted by RAFT [65] between two short-exposures.", + "bbox": [ + 75, + 666, + 468, + 833 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Low-light event imaging. Fig. 9 compares event-image visualizations of SPAD and that of a state-of-the-art commercial event sensor (Prophesee EVK4), across various", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/85c7f483ef5ff3d49b0a9ca1ef3518a6f86d20b34b7f4a18599f97c31eb0fbfa.jpg", + "image_caption": [ + "Compressive frame" + ], + "image_footnote": [], + "bbox": [ + 508, + 108, + 632, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/68caba06bceae84361f249d39cf5f17d9ffc27556e84ddeb3858dbe92f168567.jpg", + "image_caption": [ + "Recovered frame" + ], + "image_footnote": [], + "bbox": [ + 637, + 108, + 761, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/859ef4cd2e6e9b055864c9f0db9b5ac5bfbab380bdf1043f5349407b1f08122c.jpg", + "image_caption": [ + "Event image", + "SwissSPAD2 projections" + ], + "image_footnote": [], + "bbox": [ + 766, + 109, + 890, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a38bfe0d98ac5a3b148c42422f35949a0dec357730c02c225786bf3bfc442759.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 199, + 630, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/056222d39ccb53124b482ec6010df2cf88ee1633d237d1d959147f7cf6642679.jpg", + "image_caption": [ + "Infinicam projections" + ], + "image_footnote": [], + "bbox": [ + 637, + 199, + 761, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e925881cb7d9e2b08ded24c1094f58849cb7f14efae3f5031fea13ee6d6d9954.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 200, + 890, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/52001fca6b9e05c38e133049fb6968a1865da47d1dd78376857ab2acdcc46ade.jpg", + "image_caption": [ + "Infinicam projections (16× more light)" + ], + "image_footnote": [], + "bbox": [ + 508, + 291, + 630, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/907f2d1c8ff784ad88f54f3683feb65f3d6169d52a504b789e6879a94aecae98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 291, + 761, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8b5c82a105fcdde8453fbb76374f86659d472e4651b4ef7f960369e67b1d5c1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 291, + 890, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "light levels, with an accumulation period of $33\\mathrm{ms}$ . For a fair comparison, we bin the Prophesee's events in blocks of $2\\times 2$ pixels and use a smaller aperture to account for the lower fill-factor of the SPAD. We tuned event-generation parameters (contrast threshold, integrator decay rate) of both cameras at each light level. Low light induces blur and deteriorates the Prophesee's event stream. In contrast, SPAD-events continue to capture temporal gradients, due to the SPAD's low-light capabilities and its brightness-encoding response curve. We include an ablative study of brightness-encoding functions in Supplementary Note 2.", + "bbox": [ + 496, + 534, + 890, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our observations are in concurrence with recent works that examine the low-light performance of event cameras [19, 27], and show that SPAD-events can provide neuromorphic vision in these challenging-SNR scenarios.", + "bbox": [ + 496, + 700, + 890, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2. Comparison to High-Speed Cameras", + "text_level": 1, + "bbox": [ + 500, + 771, + 816, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Recall, as previously discussed in Sec. 4, that read-noise limits the per-frame SNR of high-speed cameras. To demonstrate this limitation, we compute projections using the $4\\mathrm{kHz}$ acquisition of the Photon Infinicam, a conventional high-speed camera, at a resolution of $1246\\times 240$ pixels. We operate the SwissSPAD2 and the Infinicam at ambient light conditions using the same lens specifications.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8172", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb1db41027ec422ffa4fcf6c16153765de028fa94fa48c064e8b750ec88e2fce.jpg", + "image_caption": [ + "UltraPhase board" + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 200, + 196 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/86f53554ad579b4c3582097706391466bfcd1ae17199cd42f715cb10148a4149.jpg", + "image_caption": [ + "Power and bandwidth measurement", + "Figure 11: Power and bandwidth requirements when computing photon-cube projections on UltraPhase [5] (left), a recent compute architecture designed for single-photon imaging, at $40\\mathrm{Hz}$ readout. (right) Our projections act as a compression scheme for photon-cubes, resulting in dramatically reduced sensor-readout and power consumption." + ], + "image_footnote": [], + "bbox": [ + 207, + 88, + 460, + 198 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As Fig. 10 shows, read noise corrupts the incident signal in Infinicam and makes it impossible to derive any useful projections. The read noise could be averaged out to some extent if the Infinicam did not perform compression-on-the-fly, but compression is central to the camera's working and enables readout over USB. Using a larger aperture to admit more light improves the quality of computed projections, but the video reconstruction and event image remain considerably worse than corresponding outputs of the SPAD.", + "bbox": [ + 75, + 345, + 468, + 482 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6.3. Bandwidth and Power Implications", + "text_level": 1, + "bbox": [ + 76, + 494, + 383, + 512 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While Sec. 6.1 has demonstrated the capabilities of photon-cube projections, we now show that our projections can also be obtained in a bandwidth-efficient manner via near-sensor computations. We implement photon-cube projections on UltraPhase (Fig. 11 (left)), a novel compute architecture designed for single-photon imaging. UltraPhase consists of $3 \\times 6$ processing cores, each of which interfaces with $4 \\times 4$ pixels, and can be 3D stacked beneath a SPAD array. We include visualizations and programming details of a few example projections in Supplementary Note 5.", + "bbox": [ + 75, + 520, + 468, + 671 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We measure the readout and power consumption of UltraPhase when computing projections on 2500 bit-planes of the falling die sequence (Fig. 1). The projections include: VCS with 16 random binary masks, an event camera, a linear projection and a combination of the three. We output projections at 12-bit depth and calculate metrics based on the clock cycles required for both compute and readout. As seen in Fig. 11 (right), computing projections on-chip dramatically reduces sensor-readout and power consumption as compared to reading out the photon-cube. Finally, similar to existing event cameras, SPAD-events have a resource footprint that reflects the underlying scene dynamics.", + "bbox": [ + 75, + 672, + 468, + 853 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In summary, our on-chip experiments show that performing computations near-sensor can increase the viability of single-photon imaging in resource constrained settings.", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7. Discussion and Future Outlook", + "text_level": 1, + "bbox": [ + 500, + 89, + 784, + 106 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "SoDaCam provides a realization of reinterpretable software-defined cameras [2, 3, 20, 34, 51] at the fine temporal resolution of SPAD-acquired photon-cubes. The proposed computations, or photon-cube projections, can match and in some cases, surpass the capabilities of existing imaging systems. The software-defined nature of photon-cube projections provides functionalities that may be difficult to achieve in conventional sensors. These projections can reduce the readout and power-consumption of SPAD arrays and potentially spur widespread adoption of single-photon imaging in the consumer domain. Finally, future chip-to-chip communication standards may also make it feasible to compute projections on a camera image signal processor.", + "bbox": [ + 496, + 114, + 890, + 313 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Adding color to SoDaCam. One way to add color is by overlaying color filter arrays (CFAs) and perform demosaicing on the computed photon-cube projection: depending on the projection, demosaicing could be relatively simple or more complex. As a reference, Bayer CFAs have been considered in the context of both video compressive sensing [73] and event cameras [64]. Incorporating CFAs with motion projections requires careful considerations, e.g., avoiding integrating across pixel locations of differing color.", + "bbox": [ + 496, + 334, + 890, + 470 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Future outlook on SPAD characteristics. A key SPAD characteristic that determines several properties of emulated cameras is the frame rate. While no fundamental limitations prevent SPADs from being operated at the frame-rates utilized in this work ( $\\sim 100\\mathrm{kHz}$ ), sensor readout and power constraints can preclude high speeds, especially in high-resolution SPAD arrays. Photon-cube projections can enable future large-format SPADs to preserve high-speed information with modest resource requirements.", + "bbox": [ + 496, + 492, + 890, + 628 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A platform for comparing cameras. Comparing imaging modalities can be quite challenging since hardware realizations of sensors can differ in numerous aspects, such as their quantum efficiency, fill factor, pixel pitch and array resolution. By emulating their imaging models, SoDaCam can serve as a platform for hardware-agnostic comparisons; for instance, determining operating conditions where one imaging modality is advantageous over another.", + "bbox": [ + 496, + 651, + 890, + 772 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A Cambrian explosion of new cameras. Besides comparing cameras, by virtue of being software-defined, SoDa-Cam can also make it significantly easier to prototype and deploy new unconventional imaging models, and even facilitate sensor-in-the-loop optimization [44, 47, 63] by tailoring photon-cube projections for downstream computer-vision tasks. This is an exciting future line of research.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8173", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Phantom-v2640. https://www.phantomhighspeed.com/products/cameraul ultrahigh4mpx/v2640. Accessed: 2023-01-28. 4", + "[2] E. H. Adelson, J. R. Bergen, et al. The plenoptic function and the elements of early vision. Computational models of visual processing, 1(2):3-20, 1991. 1, 9", + "[3] A. Agrawal, A. Veeraraghavan, and R. Raskar. Reinterpretable imager: Towards variable post-capture space, angle and time resolution in photography. In Computer Graphics Forum, volume 29, pages 763-772. Wiley Online Library, 2010. 3, 9", + "[4] I. M. Antolovic, S. Burri, C. Bruschini, R. Hoebe, and E. Charbon. Nonuniformity analysis of a 65-kpixel CMOS SPAD imager. IEEE Transactions on Electron Devices, 63 (1):57-64, 2016. doi: 10.1109/TED.2015.2458295. 3", + "[5] A. Ardelean. Computational Imaging SPAD Cameras. PhD thesis, École polytechnique fédérale de Lausanne, 2023. 2, 9", + "[6] A. Boukhayma, A. Peizerat, and C. Enz. A sub-0.5 electron read noise VGA image sensor in a standard CMOS process. IEEE Journal of Solid-State Circuits, 2016. 4", + "[7] J. Canny. A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-8(6):679-698, 1986. doi: 10.1109/TPAMI.1986.4767851. 4", + "[8] S. J. Carey, A. Lopich, D. R. Barr, B. Wang, and P. Dudek. A 100,000 fps vision sensor with embedded 535GOPS/W $256 \\times 256$ SIMD processor array. In 2013 Symposium on VLSI Circuits, pages C182-C183, 2013. 3", + "[9] J. Chen, S. J. Carey, and P. Dudek. Feature extraction using a portable vision system. In IEEE/RSJ Int. Conf. Intell. Robots Syst., Workshop Vis.-based Agile Auton. Navigation UAVs, volume 2, 2017. 3", + "[10] S. Chen and M. Guo. Live demonstration: Celex-v: A 1m pixel multi-mode event-based sensor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2019. 4, 5", + "[11] N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), volume 1, pages 886-893. IEEE, 2005. 4", + "[12] P. E. Debevec and J. Malik. Recovering high dynamic range radiance maps from photographs. In Proceedings of the 24th Annual Conference on Computer Graphics and Interactive Techniques, page 369-378. ACM Press/Addison-Wesley Publishing Co., 1997. doi: 10.1145/258734.258884. URL https://doi.org/10.1145/258734.258884. 1, 3", + "[13] F. M. Della Rocca, H. Mai, S. W. Hutchings, T. Al Abbas, K. Buckbee, A. Tsiamis, P. Lomax, I. Gyongy, N. A. Dutton, and R. K. Henderson. A $128 \\times 128$ SPAD motion-triggered time-of-flight image sensor with in-pixel histogram and column-parallel vision processor. IEEE Journal of Solid-State Circuits, 55(7):1762-1775, 2020. 3", + "[14] T. Finateu, A. Niwa, D. Matolin, K. Tsuchimoto, A. Mascheroni, E. Reynaud, P. Mostafalu, F. Brady, L. Chotard, F. LeGoff, H. Takahashi, H. Wakabayashi," + ], + "bbox": [ + 78, + 114, + 470, + 888 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Y. Oike, and C. Posch. 5.10 a $1280 \\times 720$ back-illuminated stacked temporal contrast event-based vision sensor with $4.86\\mu m$ pixels, 1.066GEPS readout, programmable event-rate controller and compressive data-formating pipeline. In 2020 IEEE International Solid-State Circuits Conference - (ISSCC), pages 112-114, 2020. doi: 10.1109/ISSCC19947.2020.9063149. 4, 5", + "[15] E. R. Fossum. What to do with sub-diffraction-limit (SDL) pixels?—a proposal for a gigapixel digital film sensor (DFS). In IEEE Workshop on Charge-Coupled Devices and Advanced Image Sensors, pages 214–217, 2005. 3", + "[16] E. R. Fossum. The quanta image sensor (QIS): Concepts and challenges. In Imaging and Applied Optics, page JTuE1. Optica Publishing Group, 2011. doi: 10.1364/COSI.2011. JTuE1. URL http://opg.optica.org/abstract.cfm?URI=COSI-2011-JTuE1.1", + "[17] E. R. Fossum, J. Ma, S. Masoodian, L. Anzagira, and R. Zizza. The quanta image sensor: Every photon counts. Sensors, 16(8):1260, 2016. 3", + "[18] D. Gehrig, H. Rebecq, G. Gallego, and D. Scaramuzza. Eklt: Asynchronous photometric feature tracking using events and frames. International Journal of Computer Vision, 128(3): 601-618, 2020. 6", + "[19] R. Graca and T. Delbruck. Unraveling the paradox of intensity-dependent DVS pixel noise. arXiv preprint arXiv:2109.08640, 2021. 8", + "[20] M. Gupta, A. Agrawal, A. Veeraraghavan, and S. G. Narasimhan. Flexible voxels for motion-aware videography. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part I 11, pages 100-114. Springer, 2010. 3, 9", + "[21] F. Gutierrez-Barragan, A. Ingle, T. Seets, M. Gupta, and A. Velten. Compressive single-photon 3D cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17854-17864, June 2022. 3", + "[22] I. Gyongy, N. A. Dutton, and R. K. Henderson. Single-photon tracking for high-speed vision. Sensors, 18(2):323, 2018. 3", + "[23] I. Gyongy, S. W. Hutchings, A. Halimi, M. Tyler, S. Chan, F. Zhu, S. McLaughlin, R. K. Henderson, and J. Leach. High-speed 3D sensing via hybrid-mode imaging and guided upsampling. Optica, 7(10):1253-1260, 2020. 3", + "[24] C. Harris, M. Stephens, et al. A combined corner and edge detector. In *Alvey vision conference*, volume 15, pages 10-5244. CiteSeer, 1988. 4", + "[25] J. Hidalgo-Carrió, G. Gallego, and D. Scaramuzza. Event-aided direct sparse odometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5781-5790, June 2022. 6", + "[26] B. K. Horn and B. G. Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185-203, 1981. 4", + "[27] Y. Hu, S.-C. Liu, and T. Delbruck. v2e: From video frames to realistic DVS events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1312-1321, June 2021. 8", + "[28] J. Igual. Photographic noise performance measures based on" + ], + "bbox": [ + 503, + 92, + 890, + 893 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8174", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "raw files analysis of consumer cameras. Electronics, 8(11): 1284, 2019. 4", + "[29] A. Ingle, A. Velten, and M. Gupta. High Flux Passive Imaging With Single-Photon Sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 3", + "[30] A. Ingle, T. Seets, M. Buttafava, S. Gupta, A. Tosi, M. Gupta, and A. Velten. Passive inter-photon imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021. 3", + "[31] K. Iwabuchi, Y. Kameda, and T. Hamamoto. Image quality improvements based on motion-based deblurring for single-photon imaging. IEEE Access, 9:30080-30094, 2021. doi: 10.1109/ACCESS.2021.3059293. 3", + "[32] Y. Jiang, I. Choi, J. Jiang, and J. Gu. HDR video reconstruction with tri-exposure quad-bayer sensors. arXiv preprint arXiv:2103.10982, 2021. 4", + "[33] A. Levin, P. Sand, T. S. Cho, F. Durand, and W. T. Freeman. Motion-invariant photography. ACM Transactions on Graphics (TOG), 27(3):1-9, 2008. 4, 6", + "[34] M. Levoy and P. Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 31-42, 1996. 1, 9", + "[35] Y. Li, M. Qi, R. Gulve, M. Wei, R. Genov, K. N. Kutulakos, and W. Heidrich. End-to-end video compressive sensing using anderson-accelerated unrolled networks. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12, 2020. doi: 10.1109/ICCP48838.2020.9105237. 5", + "[36] P. Lichtsteiner. 64x64 event-driven logarithmic temporal derivative silicon retina. In Program 2003 IEEE Workshop on CCD and AIS, 2003. 4, 5", + "[37] Y. Liu, F. Gutierrez-Barragan, A. Ingle, M. Gupta, and A. Velten. Single-photon camera guided extreme dynamic range imaging. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1575-1585, January 2022. 3", + "[38] P. Llull, X. Liao, X. Yuan, J. Yang, D. Kittle, L. Carin, G. Sapiro, and D. J. Brady. Coded aperture compressive temporal imaging. Opt. Express, 21(9):10526-10545, May 2013. doi: 10.1364/OE.21.010526. URL https://opg.optica.org/oe/abstract.cfm?URI=oe-21-9-10526.1", + "[39] B. D. Lucas and T. Kanade. An iterative image registration technique with an application to stereo vision. In *IJCAI'81: 7th international joint conference on Artificial intelligence*, volume 2, pages 674–679, 1981. 4", + "[40] J. Ma, S. Masoodian, D. A. Starkey, and E. R. Fossum. Photon-number-resolving megapixel image sensor at room temperature without avalanche gain. Optica, 4(12): 1474-1481, Dec 2017. doi: 10.1364/OPTICA.4.001474. URL http://www.osapublishing.org/optica/abstract.cfm?URI=optica-4-12-1474.3", + "[41] S. Ma, S. Gupta, A. C. Ulku, C. Bruschini, E. Charbon, and M. Gupta. Quanta burst photography. ACM Transactions on Graphics, 39(4):1-16, July 2020. ISSN 0730-0301, 1557-7368. 3, 7", + "[42] S. Ma, P. Mos, E. Charbon, and M. Gupta. Burst vi" + ], + "bbox": [ + 78, + 90, + 470, + 895 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sion using single-photon cameras. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5375-5385, January 2023. 3", + "[43] S. Mann and R. Picard. Beingundigital'with digital cameras. MIT Media Lab Perceptual, 1:2, 1994. 3", + "[44] J. N. Martel, L. K. Mueller, S. J. Carey, P. Dudek, and G. Wetzstein. Neural sensors: Learning pixel exposures for HDR imaging and video compressive sensing with programmable sensors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(7):1642-1653, 2020. 9", + "[45] J. N. P. Martel, L. K. Müller, S. J. Carey, and P. Dudek. Parallel HDR tone mapping and auto-focus on a cellular processor array vision chip. In 2016 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1430-1433, 2016. doi: 10.1109/ISCAS.2016.7527519. 3", + "[46] J. N. P. Martel, L. K. Müller, S. J. Carey, and P. Dudek. High-speed depth from focus on a programmable vision chip using a focus tunable lens. In 2017 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-4, 2017. doi: 10.1109/ISCAS.2017.8050548.3, 4", + "[47] C. A. Metzler, H. Ikoma, Y. Peng, and G. Wetzstein. Deep optics for single-shot high-dynamic-range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 9", + "[48] K. Morimoto, A. Ardelean, M.-L. Wu, A. C. Ulku, I. M. Antolovic, C. Bruschini, and E. Charbon. Megapixel time-gated SPAD image sensor for 2D and 3D imaging applications. Optica, 7(4):346-354, Apr. 2020. 1, 3", + "[49] K. Morimoto, J. Iwata, M. Shinohara, H. Sekine, A. Abdelghafar, H. Tsuchiya, Y. Kuroda, K. Tojima, W. Endo, Y. Maehashi, Y. Ota, T. Sasago, S. Maekawa, S. Hikosaka, T. Kanou, A. Kato, T. Tezuka, S. Yoshizaki, T. Ogawa, K. Uehira, A. Ehara, F. Inui, Y. Matsuno, K. Sakurai, and T. Ichikawa. 3.2 megapixel 3D-stacked charge focusing SPAD for low-light imaging and depth sensing. In 2021 IEEE International Electron Devices Meeting (IEDM), pages 20.2.1-20.2.4, 2021. doi: 10.1109/IEDM19574.2021.9720605.1, 3", + "[50] S. Namiki, S. Sato, Y. Kameda, and T. Hamamoto. Imaging method using multi-threshold pattern for photon detection of quanta image sensor. In International Workshop on Advanced Imaging Technology (IWAIT) 2022, volume 12177, page 1217702. SPIE, 2022. 3", + "[51] S. K. Nayar, V. Branzoi, and T. E. Boult. Programmable imaging: Towards a flexible camera. International Journal of Computer Vision, 70:7-22, 2006. 3, 9", + "[52] C. Posch, D. Matolin, and R. Wohlgenannt. A QVGA 143 dB Dynamic Range Frame-Free PWM Image Sensor With Lossless Pixel-Level Video Compression and Time-Domain CDS. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2011. doi: 10.1109/JSSC.2010.2085952. 1", + "[53] R. Raskar, A. Agrawal, and J. Tumblin. Coded exposure photography: motion deblurring using fluttered shutter. In Acm Siggraph 2006 Papers, pages 795-804. 2006. 4", + "[54] H. Rebecq, R. Ranftl, V. Koltun, and D. Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 6" + ], + "bbox": [ + 501, + 92, + 890, + 895 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8175", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] D. Reddy, A. Veeraraghavan, and R. Chellappa. P2C2: Programmable pixel compressive camera for high speed imaging. In CVPR 2011, pages 329-336, 2011. doi: 10.1109/CVPR.2011.5995542. 1, 3", + "[56] X. Ren, P. W. Connolly, A. Halimi, Y. Altmann, S. McLaughlin, I. Gyongy, R. K. Henderson, and G. S. Buller. High-resolution depth profiling using a range-gated CMOS SPAD quanta image sensor. Optics express, 26(5):5541-5557, 2018. 3", + "[57] C. Scheerlinck, H. Rebecq, D. Gehrig, N. Barnes, R. Mahony, and D. Scaramuzza. Fast image reconstruction with an event camera. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), March 2020. 6", + "[58] T. Seets, A. Ingle, M. Laurenzis, and A. Velten. Motion adaptive deblurring with single-photon cameras. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1945-1954, January 2021. 3", + "[59] M.-W. Seo, Y. Shirakawa, Y. Masuda, Y. Kawata, K. Kagawa, K. Yasutomi, and S. Kawahito. 4.3 a programmable sub-nanosecond time-gated 4-tap lock-in pixel cmos image sensor for real-time fluorescence lifetime imaging microscopy. In 2017 IEEE International Solid-State Circuits Conference (ISSCC), pages 70-71, 2017. doi: 10.1109/ISSCC.2017.7870265. 5", + "[60] T. Serrano-Gotarredona and B. Linares-Barranco. A $128 \\times 128$ $1.5\\%$ contrast sensitivity $0.9\\%$ FPN $3\\mu s$ latency $4\\mathrm{mW}$ asynchronous frame-free dynamic vision sensor using transimpedance preamplifiers. IEEE Journal of Solid-State Circuits, 48(3):827-838, 2013. 1, 4, 5", + "[61] P. Shedligeri, A. S., and K. Mitra. A unified framework for compressive video recovery from coded exposure techniques. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1600-1609, January 2021. 5", + "[62] C. Shi, N. Song, W. Li, Y. Li, B. Wei, H. Liu, and J. Jin. A review of event-based indoor positioning and navigation. 2022. 6", + "[63] V. Sitzmann, S. Diamond, Y. Peng, X. Dun, S. Boyd, W. Heidrich, F. Heide, and G. Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and super-resolution imaging. ACM Transactions on Graphics (TOG), 37(4):114, 2018. 9", + "[64] G. Taverni, D. Paul Moeys, C. Li, C. Cavaco, V. Motsnyi, D. San Segundo Bello, and T. Delbruck. Front and back illuminated dynamic and active pixel vision sensors comparison. IEEE Transactions on Circuits and Systems II: Express Briefs, 65(5):677-681, 2018. doi: 10.1109/TCSII.2018.2824899. 9", + "[65] Z. Teed and J. Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European Conference on Computer Vision, 2020. 8", + "[66] S. Tulyakov, D. Gehrig, S. Georgoulis, J. Erbach, M. Gehrig, Y. Li, and D. Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16155-16164, June 2021. 6" + ], + "bbox": [ + 78, + 90, + 470, + 895 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] A. C. Ulku, C. Bruschini, I. M. Antolovic, Y. Kuo, R. Ankri, S. Weiss, X. Michalet, and E. Charbon. A $512 \\times 512$ SPAD Image Sensor With Integrated Gating for Widefield FLIM. IEEE Journal of Selected Topics in Quantum Electronics, 25 (1):1-12, Jan. 2019. ISSN 1077-260X, 1558-4542. doi: 10. 1109/JSTQE.2018.2867439. 3, 7", + "[68] G. Wan, X. Li, G. Agranov, M. Levoy, and M. Horowitz. CMOS image sensors with multi-bucket pixels for computational photography. IEEE Journal of Solid-State Circuits, 47 (4):1031-1042, 2012. doi: 10.1109/JSSC.2012.2185189. 5", + "[69] M. Wei, N. Sarhangnejad, Z. Xia, N. Gusev, N. Katic, R. Genov, and K. N. Kutulakos. Coded two-bucket cameras for computer vision. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 3, 4, 5", + "[70] M. White, S. Ghajari, T. Zhang, A. Dave, A. Veeraraghavan, and A. Molnar. A differential SPAD array architecture in $0.18\\mu \\mathrm{m}$ CMOS for HDR imaging. In 2022 IEEE International Symposium on Circuits and Systems (ISCAS), pages 292-296, 2022. doi: 10.1109/ISCAS48785.2022.9937558.3", + "[71] T. Yamazaki, H. Katayama, S. Uehara, A. Nose, M. Kobayashi, S. Shida, M. Odahara, K. Takamiya, Y. Hisamatsu, S. Matsumoto, L. Miyashita, Y. Watanabe, T. Izawa, Y. Muramatsu, and M. Ishikawa. 4.9 a 1ms high-speed vision chip with 3D-stacked 140GOPS column-parallel PEs for spatio-temporal image processing. In 2017 IEEE International Solid-State Circuits Conference (ISSCC), pages 82–83, 2017. doi: 10.1109/ISSCC.2017.7870271. 3", + "[72] F. Yang, Y. M. Lu, L. Sbaiz, and M. Vetterli. Bits from photons: Oversampled image acquisition using binary poisson statistics. IEEE Transactions on Image Processing, 21(4): 1421-1436, 2012. doi: 10.1109/TIP.2011.2179306. 3", + "[73] X. Yuan, Y. Liu, J. Suo, F. Durand, and Q. Dai. Plug-and-play algorithms for video snapshot compressive imaging. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):7093-7111, 2022. doi: 10.1109/TPAMI.2021.3099035.7, 9", + "[74] J. Zhang, X. Yang, Y. Fu, X. Wei, B. Yin, and B. Dong. Object tracking by jointly exploiting frame and event domain. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13043-13052, 2021. 6", + "[75] K. Zhang, Y. Li, W. Zuo, L. Zhang, L. Van Gool, and R. Timofte. Plug-and-play image restoration with deep denoiser prior. arXiv preprint, 2020. 8", + "[76] R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 7", + "[77] T. Zhang, M. J. White, A. Dave, S. Ghajari, A. Raghuram, A. C. Molnar, and A. Veeraraghavan. First arrival differential LiDAR. In 2022 IEEE International Conference on Computational Photography (ICCP), pages 1-12, 2022. doi: 10.1109/ICCP54855.2022.9887683. 3", + "[78] W. Zhang and L. Lin. Light field flow estimation based on occlusion detection. Journal of Computer and Communications, 5(3):1-9, 2017. 1" + ], + "bbox": [ + 501, + 92, + 890, + 878 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "8176", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/SoDaCam_ Software-defined Cameras via Single-Photon Imaging/ce1d43a9-f5c5-4d36-93a7-015029e926c0_model.json b/2023/SoDaCam_ Software-defined Cameras via Single-Photon Imaging/ce1d43a9-f5c5-4d36-93a7-015029e926c0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4c76904b6fd21853d9482508804084521b08e4cf --- /dev/null +++ b/2023/SoDaCam_ Software-defined Cameras via Single-Photon Imaging/ce1d43a9-f5c5-4d36-93a7-015029e926c0_model.json @@ -0,0 +1,3557 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.155, + 0.131, + 0.819, + 0.154 + ], + "angle": 0, + "content": "SoDaCam: Software-defined Cameras via Single-Photon Imaging" + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.181, + 0.327, + 0.198 + ], + "angle": 0, + "content": "Varun Sundar†" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.203, + 0.358, + 0.215 + ], + "angle": 0, + "content": "vsundar4@cs.wisc.edu" + }, + { + "type": "text", + "bbox": [ + 0.419, + 0.181, + 0.56, + 0.198 + ], + "angle": 0, + "content": "Andrei Ardelean†" + }, + { + "type": "text", + "bbox": [ + 0.409, + 0.203, + 0.571, + 0.216 + ], + "angle": 0, + "content": "a.ardelean@epfl.ch" + }, + { + "type": "text", + "bbox": [ + 0.638, + 0.181, + 0.775, + 0.198 + ], + "angle": 0, + "content": "Tristan Swedish" + }, + { + "type": "text", + "bbox": [ + 0.621, + 0.203, + 0.793, + 0.216 + ], + "angle": 0, + "content": "tristan@ubicept.com" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.224, + 0.37, + 0.241 + ], + "angle": 0, + "content": "Claudio Bruschini†" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.245, + 0.571, + 0.26 + ], + "angle": 0, + "content": "{claudio.bruschini, edoardo.charbon}@epfl.ch" + }, + { + "type": "text", + "bbox": [ + 0.642, + 0.223, + 0.764, + 0.243 + ], + "angle": 0, + "content": "Mohit Gupta†,§" + }, + { + "type": "text", + "bbox": [ + 0.622, + 0.245, + 0.785, + 0.259 + ], + "angle": 0, + "content": "mohitg@cs.wisc.edu" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.267, + 0.846, + 0.287 + ], + "angle": 0, + "content": "†University of Wisconsin-Madison ‡École polytechnique fédérale de Lausanne §Ubicept" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.321, + 0.314, + 0.337 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.474, + 0.656 + ], + "angle": 0, + "content": "Reinterpretable cameras are defined by their post-processing capabilities that exceed traditional imaging. We present \"SoDaCam\" that provides reinterpretable cameras at the granularity of photons, from photon-cubes acquired by single-photon devices. Photon-cubes represent the spatio-temporal detections of photons as a sequence of binary frames, at frame-rates as high as \\(100\\mathrm{kHz}\\). We show that simple transformations of the photon-cube, or photon-cube projections, provide the functionality of numerous imaging systems including: exposure bracketing, flutter shutter cameras, video compressive systems, event cameras, and even cameras that move during exposure. Our photon-cube projections offer the flexibility of being software-defined constructs that are only limited by what is computable, and shot-noise. We exploit this flexibility to provide new capabilities for the emulated cameras. As an added benefit, our projections provide camera-dependent compression of photon-cubes, which we demonstrate using an implementation of our projections on a novel compute architecture that is designed for single-photon imaging." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.682, + 0.21, + 0.699 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.707, + 0.47, + 0.846 + ], + "angle": 0, + "content": "Throughout the history of imaging, sensing technologies and the corresponding processing have developed hand-in-hand. In fact, sensing technologies have, to some extent, defined the scope of processing captured data. In the film era, instances of such processing included dodging and burning. The advent of digital cameras provided processing at the granularity of pixels and paved the way for modern computer vision. Light field cameras [34, 78], by sampling the plenoptic function [2], allowed post-capture processing at" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.322, + 0.892, + 0.413 + ], + "angle": 0, + "content": "the granularity of light rays, enabling novel functionalities such as refocusing photos after-capture. The logical limit of post-capture processing, given the fundamental quantization of light, would be at the level of individual photons. What would imaging look like if we could perform computational processing on individual photons?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.893, + 0.625 + ], + "angle": 0, + "content": "In this work, we show that photon data captured by a new class of single-photon detectors, called single-photon avalanche diodes (SPADs), makes it possible to emulate a wide range of imaging modalities such as exposure bracketing [12], video compressive systems [38, 55] and event cameras [52, 60]. A user then has the flexibility to choose one (or even multiple) of these functionalities post-capture (Fig. 1 (top)). SPAD arrays can operate as extremely high frame-rate photon detectors (\\(\\sim 100\\mathrm{kHz}\\)), producing a temporal sequence of binary frames called a photon-cube [16]. We show that computing photon-cube projections, which are simple linear and shift operations, can reinterpret the photon-cube to achieve novel post-capture imaging functionalities in a software-defined manner (Fig. 1 (middle))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.892, + 0.716 + ], + "angle": 0, + "content": "As case studies, we emulate three distinct imagers: high-speed video compressive imaging; event cameras which respond to dynamic scene content; and motion projections which emulate sensor motion, without any real camera movement. Fig. 1 (bottom) shows the outputs of these cameras that are derived from the same photon-cube." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.886 + ], + "angle": 0, + "content": "Computing photon-cube projections. One way to obtain photon-cube projections is to read the entire photon-cube off the SPAD array and then perform relevant computations off-chip; we adopt this strategy for our experiments in Secs. 6.1 and 6.2. While reasonable for certain applications, reading out photon-cubes requires an exorbitant data-bandwidth, which can be up to 100 Gbps for a 1 MPixel array—well beyond the capacity of existing data peripherals. Such readout considerations will become center stage as large-format SPAD arrays are fabricated [48, 49]." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.887, + 0.892, + 0.902 + ], + "angle": 0, + "content": "An alternative is to avoid transferring the entire photon-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.852, + 0.47, + 0.902 + ], + "angle": 0, + "content": "*Corresponding author: Varun Sundar. This research was supported in parts by NSF CAREER award 1943149, NSF award CNS-2107060, and the Swiss National Science Foundation grant 200021_166289. We also thank Paul Mos for providing access to SwissSPAD2 acquisition software." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8165" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.891, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.256, + 0.891, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.357, + 0.424, + 0.615, + 0.438 + ], + "angle": 0, + "content": "Capabilities of photon-cube projections" + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.447, + 0.268, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.523, + 0.268, + 0.535 + ], + "angle": 0, + "content": "Coded exposure" + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.453, + 0.356, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.54, + 0.358, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.616, + 0.352, + 0.645 + ], + "angle": 0, + "content": "\\(12 \\times\\) video reconstruction (raster order) Snapshot compressive imager (Sec. 5.1)" + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.446, + 0.577, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.397, + 0.522, + 0.574, + 0.535 + ], + "angle": 0, + "content": "3D scatter plot, \\(0\\leq t\\leq 0.1s\\)" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.54, + 0.577, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.397, + 0.615, + 0.58, + 0.645 + ], + "angle": 0, + "content": "Event image, \\(0.5 \\leq t \\leq 0.53\\) s Event camera (Sec. 5.2)" + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.447, + 0.737, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.523, + 0.712, + 0.534 + ], + "angle": 0, + "content": "Static camera" + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.541, + 0.738, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.615, + 0.732, + 0.645 + ], + "angle": 0, + "content": "Parabolic trajectory Motion Pro" + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.447, + 0.887, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.523, + 0.871, + 0.535 + ], + "angle": 0, + "content": "Linear trajectory" + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.541, + 0.887, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.615, + 0.879, + 0.627 + ], + "angle": 0, + "content": "Parabolic deblurred" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.657, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Figure 1: (top) SoDaCam can emulate a variety of cameras from the photon-cubes acquired by single-photon devices. (middle) Photon-cubes represent the spatio-temporal detection of photons as a sequence of binary frames. Projections of the photon-cube, when computed either on or off-chip, result in reinterpretable and software-defined cameras. We demonstrate the versatility of photon-cube projections on a real dynamic scene: a die falls on a table, bounces, spins in the air and later ricochets off a nearby toy top. (bottom) The cameras emulated by our photon-cube projections can produce: a \\(12 \\times\\) high-speed video from a single compressive snapshot, event-stream representations of two time intervals (blue and red depict positive and negative spikes respectively), an image where the die appears stationary, as well as a motion-deblurred image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.785, + 0.47, + 0.891 + ], + "angle": 0, + "content": "cube by computing projections near sensor. As a proof-of-concept, we implement photon-cube projections on Ultra-Phase [5], a recently-developed programmable SPAD imager with independent processing cores that have dedicated RAM and instruction memory. We show, in Sec. 6.3, that computing projections on-chip greatly reduces sensor readout and, as a consequence, power consumption." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.785, + 0.892, + 0.891 + ], + "angle": 0, + "content": "Implications: Toward a photon-level software-defined camera. The photon-cube projections introduced in this paper are computational constructs that provide a realization of software-defined cameras or SoDaCam. Being software-defined, SoDaCam can emulate multiple cameras simultaneously without additional hardware complexity. SoDaCam, by going beyond baked in hardware choices," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8166" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.471, + 0.168 + ], + "angle": 0, + "content": "unlocks hitherto unseen capabilities—such as 2000 FPS video from \\(25\\mathrm{Hz}\\) readout (Fig. 7); event imaging in very low-light conditions (Fig. 9); and motion stacks, which are a stack of images where in each image, objects only in certain velocity ranges appear sharp (Fig. 6)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.186, + 0.471, + 0.323 + ], + "angle": 0, + "content": "Limitations. The SPAD array [67] used in this work has a relatively low spatial resolution \\((512 \\times 256)\\), and a low fill-factor \\((\\sim 10\\%)\\) owing to the lack of microlenses in the prototype used. Similarly, the near-sensor processor that we use has limited capabilities compared to off-chip processors. However, with rapid progress in the development of single-photon cameras [48, 49] and increasing interest in near-sensor processors, we anticipate that many of these shortcomings will be addressed in the upcoming years." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.336, + 0.22, + 0.351 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.361, + 0.47, + 0.468 + ], + "angle": 0, + "content": "Reinterpretable imaging has previously been explored at the granularity of light rays [3], by modulating the plenoptic function, and at the level of spatio-temporal voxels [20], by using fast per-pixel shutters. SoDaCam represents a logical culmination of reinterpretability at the level of photon detections, that facilitates multiple post-capture imaging functionalities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.47, + 0.637 + ], + "angle": 0, + "content": "Programmable imaging using a digital micromirror device was first introduced in Nayar et al. [51] to perform precapture radiometric manipulations. Modern programmable cameras are typically near-sensor processors [8, 69, 71] that can perform limited operations in analog [46, 55, 69], while more complex operations [9, 45] occur after analog-to-digital conversion (ADC). In contrast, by performing post-capture computations directly on photon detections, we can perform complex operations without incurring the read-noise penalty that is associated with ADC." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.47, + 0.822 + ], + "angle": 0, + "content": "Passive single-photon imaging. Only recently have SPADs been utilized as passive imaging devices, with applications in high-dynamic range imaging [29, 30, 37, 50], motion-compensation [31, 58], burst photography [41, 42] and object tracking [22]. Compared to compute-intensive burst-photography methods [41], our proposed techniques involve lightweight computations that can be performed near sensor. These computations can also be performed using other single-photon imagers such as Jots [15, 17], which feature higher sensor resolution and photon-efficiency [40], albeit at lower frame-rates and higher read noise." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Reducing the readout of SPADs. Several data reduction strategies have been proposed in the context of SPADs that are used to timestamp incident photons, including: coarse histograms [13, 23, 56], compressive histograms [21], and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "measuring differential time-of-arrivals [70, 77]. When SPADs are operated as photon-detectors, multi-bit counting [48], or summing binary frames, can reduce readout. While compression is not our main objective, we show that photon-cube projections act as camera-specific compression schemes that dramatically reduce sensor readout." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.198, + 0.892, + 0.216 + ], + "angle": 0, + "content": "3. Background: Single-Photon Imaging Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.224, + 0.892, + 0.328 + ], + "angle": 0, + "content": "A SPAD array captures incident light as a photon-cube: a temporal sequence of binary frames that represents the pixel-wise detection of photons across their respective exposure windows. We can model the stochastic arrival of photons as a Poisson process [72], allowing us to treat spatio-temporal values of the photon-cube as independent Bernoulli random variables with" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.34, + 0.892, + 0.359 + ], + "angle": 0, + "content": "\\[\n\\Pr \\left\\{B _ {t} (\\mathbf {x}) = 1 \\right\\} = 1 - e ^ {- (\\eta \\Phi (\\mathbf {x}, t) + r _ {q}) w _ {\\exp}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.892, + 0.493 + ], + "angle": 0, + "content": "where \\( B_{t}(\\mathbf{x}) \\) represents the value of the photon-cube at pixel \\( \\mathbf{x} \\) and exposure index \\( 1 \\leq t \\leq T \\), which receives a mean incident flux of intensity \\( \\Phi (\\mathbf{x},t) \\) across its exposure of duration \\( w_{\\mathrm{exp}} \\). Additionally, \\( \\eta \\) is the photon detection efficiency of the SPAD, and \\( r_q \\) denotes the sensor's dark count rate—which is the rate of spurious counts unrelated to incident photons. While individual binary frames are extremely noisy, the temporal sum of the photon-cube" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.505, + 0.892, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {\\text {s u m}} (\\mathbf {x}) := \\sum_ {t = 1} ^ {T} B _ {t} (\\mathbf {x}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.892, + 0.635 + ], + "angle": 0, + "content": "can produce an 'image' of the scene that is sharp in static regions, but blurry in dynamic regions (Fig. 2 (top)). Indeed, in static regions, the sum-image can be used to derive a maximum likelihood estimator of the scene intensity [4], given by \\(\\hat{\\Phi} (\\mathbf{x}) = -\\ln (1 - T^{-1}\\mathcal{I}_{\\mathrm{sum}}(\\mathbf{x})) / \\eta w_{\\mathrm{exp}} - r_q / w_{\\mathrm{exp}}\\)" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.648, + 0.788, + 0.666 + ], + "angle": 0, + "content": "4. Projections of the Photon-Cube" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.87 + ], + "angle": 0, + "content": "The temporal sum described in Eq. (2) is a simple instance of projections of a photon-cube. Our key observation is that it is possible to compute a wide range of photon-cube projections, each of which emulates a unique sensing modality post-capture—including modalities that are difficult to achieve with conventional cameras. For example, varying the number of bit-planes that are summed over emulates exposure bracketing [12, 43], which is typically used for HDR imaging. Compared to conventional exposure bracketing, the emulated exposure stack, being software-defined, does not require spatial and temporal registration, which can often be error-prone. Fig. 2 (top) shows an example of an exposure stack computed from a photon-cube." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Going further, we can gradually increase the complexity of the projections. For example, consider a coded exposure" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8167" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.085, + 0.089, + 0.199, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.165, + 0.189, + 0.176 + ], + "angle": 0, + "content": "Short exposure" + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.089, + 0.325, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.165, + 0.322, + 0.19 + ], + "angle": 0, + "content": "Medium exposure Exposure Stack" + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.089, + 0.451, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.165, + 0.437, + 0.177 + ], + "angle": 0, + "content": "Long exposure" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.2, + 0.199, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.276, + 0.194, + 0.303 + ], + "angle": 0, + "content": "Spatially-varying exposure" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.199, + 0.301, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.25, + 0.3, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.2, + 0.452, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.277, + 0.455, + 0.302 + ], + "angle": 0, + "content": "Video compressive frame" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.319, + 0.471, + 0.425 + ], + "angle": 0, + "content": "Figure 2: Coded exposures from photon-cubes. (top) An exposure stack with sum-images computed using 250, 500, and 1000 bit-planes. Short exposures are noisy while long exposures exhibit motion-blur. (bottom) Spatially-varying exposure that uses a quad pattern [32] (see inset), and a video compressive frame that uses 16 random binary masks to modulate the photon-cube. Zoom-in to see details." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.45, + 0.466, + 0.466 + ], + "angle": 0, + "content": "projection that multiplexes bit-planes with a temporal code" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.473, + 0.469, + 0.514 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {\\text {f l u t t e r}} (\\mathbf {x}) := \\sum_ {t = 1} ^ {T} C _ {t} B _ {t} (\\mathbf {x}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.469, + 0.566 + ], + "angle": 0, + "content": "where \\( C_t \\) is the temporal code. An example of globally-coded exposures is the flutter shutter camera [53], which uses pseudo-random binary codes for motion-deblurring." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.469, + 0.597 + ], + "angle": 0, + "content": "More general coded exposures can be obtained via spatially-varying temporal coding patterns \\( C_t(\\mathbf{x}) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.606, + 0.469, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {\\text {c o d e d}} (\\mathbf {x}) := \\sum_ {t = 1} ^ {T} C _ {t} (\\mathbf {x}) B _ {t} (\\mathbf {x}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.654, + 0.469, + 0.774 + ], + "angle": 0, + "content": "Fig. 2 (bottom) shows spatially-varying exposures that use a quad (Bayer-like) spatial pattern and random binary masks. With photon-cubes, we can perform spatially-varying coding without bulky spatial light modulators, similar to focalplane sensor-processors [46, 69]. Moreover, we can capture multiple coded exposures simultaneously, which is challenging to realize in existing sensors. In Sec. 5.1, we describe coding patterns for video compressive sensing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.775, + 0.469, + 0.835 + ], + "angle": 0, + "content": "Spatial and temporal gradients form the building blocks of several computer vision algorithms [7, 11, 24, 26, 39]. Given this, another projection of interest is temporal contrast, i.e., a derivative filter preceded by a smoothing filter:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.85, + 0.469, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {\\text {c o n t r a s t}} (\\mathbf {x}, t) := D _ {t} \\circ G * B _ {t} (\\mathbf {x}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(D_{t}\\) is the difference operator, \\(G\\) could be exponential or Gaussian smoothing, \\(\\circ\\) denotes function composition," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "and * denotes convolution. Due to their sparse nature, gradients form the basis of bandwidth- and power-efficient event cameras [10, 14, 36, 60], which we emulate in Sec. 5.2." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.137, + 0.892, + 0.258 + ], + "angle": 0, + "content": "So far, we have considered projections taken only along the time axis. Next, we consider a more general class of spatio-temporal projections that lead to novel functionalities. For instance, computing a simple projection, such as the temporal sum, along arbitrary spatio-temporal directions emulates sensor motion during exposure time [33], but without moving the sensor. We achieve this by shifting bitplanes and computing their sum:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.267, + 0.892, + 0.308 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {\\text {s h i f t}} (\\mathbf {x}) := \\sum_ {t = 1} ^ {T} B _ {t} (\\mathbf {x} + \\mathbf {r} (t)), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.892, + 0.377 + ], + "angle": 0, + "content": "where \\(\\mathbf{r}\\) is a discretized 2D trajectory that determines sensor motion. Outside a software-defined framework, such projections are hard to realize without physical actuators. We describe the capabilities of motion projections in Sec. 5.3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.892, + 0.513 + ], + "angle": 0, + "content": "In summary, the proposed photon-cube projections are simple linear and shift operators that lead to a diverse set of post-capture imaging functionalities. These projections pave the way for future 'swiss-army-knife' imaging systems that achieve multiple functionalities (e.g., event cameras, high-speed cameras, conventional cameras, HDR cameras) simultaneously with a single sensor. Finally, these projections can be computed efficiently in an online manner, which makes on-chip implementation viable (Sec. 6.3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.513, + 0.892, + 0.619 + ], + "angle": 0, + "content": "At this point, we note that a key enabling factor of photon-cube projections is the extremely high temporal-sampling rate of SPADs. Indeed, the temporal sampling rate determines key aspects of sensor emulation, such as the discretization of temporal derivatives and motion trajectories. This raises a natural question: can we use conventional high-speed cameras for computing projections?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.892, + 0.833 + ], + "angle": 0, + "content": "Trade-off between frame-rate and SNR. In principle, photon-cube projections can be computed using regular (CMOS or CCD based) high-speed cameras. Unfortunately, each frame captured by a high-speed camera incurs a read-noise penalty, which increases with the camera's frame-rate [6]. In fact, the read noise levels of high-speed cameras [1] can be \\(10 - 30 \\times\\) higher than consumer cameras [28]. Coupled with the low per-frame incident flux at high frame-rates, high levels of read noise result in extremely low SNRs. In contrast, SPADs do not incur a per-frame read noise and are limited only by the fundamental photon noise. Hence, for the post-capture software-defined functionalities proposed here, it is imperative to use SPADs." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.845, + 0.862, + 0.863 + ], + "angle": 0, + "content": "5. Emulating Cameras from Photon-Cubes" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Sec. 4 presented the concept of photon-cube projections, and its potential for achieving multiple post-capture imag-" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8168" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.093, + 0.197, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.209, + 0.092, + 0.327, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.089, + 0.467, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.221, + 0.47, + 0.357 + ], + "angle": 0, + "content": "Figure 3: Modulating masks for video compressive sensing. (left) A single VCS measurement temporally compresses a sequence of frames using binary random masks. (center) Two-bucket cameras capture an additional measurement by using the complementary mask sequence. (right) We propose using multi-bucket captures by randomly choosing an active bucket for each frame. Both two-bucket and multi-bucket captures have \\(100\\%\\) light efficiency. All masks are visualized here for \\(16 \\times 16\\) pixels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.468, + 0.502 + ], + "angle": 0, + "content": "ing functionalities. As case studies, we now demonstrate three imaging modalities: video compressive sensing, event cameras, and motion-projection cameras. These modalities have been well-studied over several years; in particular, there exist active research communities around video compressive sensing and event cameras today. We also show new variants of these imaging systems that arise from the software-defined nature of photon-cube projections." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.514, + 0.325, + 0.53 + ], + "angle": 0, + "content": "5.1. Video Compressive Sensing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.6 + ], + "angle": 0, + "content": "Video compressive systems optically multiplex light with random binary masks, such as the patterns in Fig. 3 (left). As discussed in the previous section, such multiplexing can be achieved computationally using photon-cubes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.468, + 0.759 + ], + "angle": 0, + "content": "Two-bucket cameras. One drawback of capturing coded measurements is the light loss due to blocking of incident light. To prevent loss of light, coded two-bucket cameras [69] capture an additional measurement that is modulated by the complementary mask sequence (Fig. 3 (center)). Such measurements recover higher quality frames, even after accounting for the extra readout [35, 61]. Two-bucket captures can be readily derived from photon-cubes, by implementing Eq. (4) with the additional mask sequence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Multi-bucket cameras. We can extend the idea of two-bucket captures to multi-bucket captures by accumulating bit-planes in one of \\(k\\) buckets that is randomly chosen at each time instant and pixel location. Since multiplexing is performed computationally, we do not face any loss in photoreceptive area that [59, 68] hampers existing multi-bucket sensors. Multi-bucket captures can reconstruct a large number of frames by better conditioning video recovery and" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.091, + 0.671, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.18, + 0.65, + 0.194 + ], + "angle": 0, + "content": "\\(\\log (x)\\) vs \\(1 - e^{-x}\\)" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.091, + 0.871, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.18, + 0.838, + 0.194 + ], + "angle": 0, + "content": "Spatio-temporal plot" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.207, + 0.629, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.277, + 0.626, + 0.287 + ], + "angle": 0, + "content": "\\(\\tau = 0.45, \\beta = 0.95\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.289, + 0.605, + 0.3 + ], + "angle": 0, + "content": "5 MEvent/s" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.207, + 0.757, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.277, + 0.75, + 0.287 + ], + "angle": 0, + "content": "\\(\\tau = 0.25, \\beta = 0.95\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.289, + 0.736, + 0.3 + ], + "angle": 0, + "content": "25 MEvent/s" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.208, + 0.884, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.277, + 0.878, + 0.287 + ], + "angle": 0, + "content": "\\(\\tau = 0.45, \\beta = 0.98\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.289, + 0.864, + 0.3 + ], + "angle": 0, + "content": "2.5 MEvent/s" + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.303, + 0.845, + 0.318 + ], + "angle": 0, + "content": "Effect of contrast-threshold, exponential decay" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.333, + 0.892, + 0.484 + ], + "angle": 0, + "content": "Figure 4: Event stream from photon-cubes. (top left) By exploiting the non-linear response curve of SPADs to encode brightness, we can avoid the underflow issues of a log-response. We visualize events generated from photon-cubes using a 3D scatter plot of polarities (top right, 14000 bit-planes), and frame accumulation of events (bottom, 1200 bit-planes). Blue and red denote positive and negative spikes respectively. The event images also show the effect of varying the contrast threshold \\(\\tau\\) and exponential decay \\(\\beta\\)-larger values yield a less noisy but sparser event stream." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.512, + 0.892, + 0.543 + ], + "angle": 0, + "content": "provide extreme high-speed video imaging. Fig. 3 (right) shows the modulating masks for a four-bucket capture." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.554, + 0.655, + 0.569 + ], + "angle": 0, + "content": "5.2.Event Cameras" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.701 + ], + "angle": 0, + "content": "Next, we describe emulation of event-cameras, which capture changes in light intensity and are conceptually similar to the temporal contrast projection introduced in Eq. (5). Physical implementations of event sensors [10, 14, 36, 60] generate a photoreceptor voltage \\( V(\\mathbf{x}, t) \\) with a logarithmic response to incident flux \\( \\Phi(\\mathbf{x}, t) \\), and output an event \\( (\\mathbf{x}, t, p) \\) when this voltage deviates sufficiently from a reference voltage \\( V_{\\mathrm{ref}}(\\mathbf{x}) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.713, + 0.892, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\left| V (\\mathbf {x}, t) - V _ {\\text {r e f}} (\\mathbf {x}) \\right| > \\tau , \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.743, + 0.892, + 0.819 + ], + "angle": 0, + "content": "where \\(\\tau\\) is called the contrast-threshold and \\(p = \\mathrm{sign}(V(\\mathbf{x},t) - V_{\\mathrm{ref}}(\\mathbf{x}))\\) encodes the polarity of the event. Once an event is generated, \\(V_{\\mathrm{ref}}(\\mathbf{x})\\) is updated to \\(V(\\mathbf{x},t)\\). Eq. (7), for a smoothly-varying flux intensity, thresholds a function of the temporal gradient, i.e., \\(\\partial_t\\log (\\Phi (\\mathbf{x},t))\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.903 + ], + "angle": 0, + "content": "From bit-planes to event streams. To produce events from SPAD frames, we compute an exponential moving average (EMA) of the bit-planes, as \\(\\mu_t(\\mathbf{x}) = (1 - \\beta)B_t(\\mathbf{x}) + \\beta \\mu_{t - 1}(\\mathbf{x})\\) where \\(\\mu_t(\\mathbf{x})\\) is the EMA, \\(\\beta\\) is the smoothing" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8169" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.264, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.161, + 0.258, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.089, + 0.465, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.16, + 0.458, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.211, + 0.264, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.283, + 0.258, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.318, + 0.236, + 0.332 + ], + "angle": 0, + "content": "Parabolic trajectory" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.211, + 0.465, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.285, + 0.422, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.311, + 0.318, + 0.438, + 0.33 + ], + "angle": 0, + "content": "Parabolic deblurred" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.347, + 0.473, + 0.484 + ], + "angle": 0, + "content": "Figure 5: Motion projections. (top) Integrating along a linear trajectory in the photon-cube changes the apparent image-space velocity of scene objects. Details are seen for (top left) the case when static, and (top right) the metallic tape when the sensor translates along the \\(x\\)-axis. (bottom) A parabolic integration trajectory results in a motion-invariant image, resulting in similar blur kernels for all objects. (bottom right) Deblurring with the resultant shift-invariant point spread function (shown in inset) produces a sharp image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.47, + 0.542 + ], + "angle": 0, + "content": "factor, and \\(B_{t}\\) is a bit-plane. We generate an event when \\(\\mu_t(\\mathbf{x})\\) deviates from \\(\\mu_{\\mathrm{ref}}(\\mathbf{x})\\) by at least \\(\\tau\\):" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.552, + 0.469, + 0.57 + ], + "angle": 0, + "content": "\\[\n\\left| h \\left(\\mu_ {t} (\\mathbf {x})\\right) - h \\left(\\mu_ {\\text {r e f}} (\\mathbf {x})\\right) \\right| > \\tau , \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.469, + 0.626 + ], + "angle": 0, + "content": "where \\( h \\) is a scalar function applied to the EMA. We can see that Eq. (8) thresholds temporal contrast, by observing the role played by the EMA and the difference operator." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.469, + 0.746 + ], + "angle": 0, + "content": "Setting \\( h \\) to be the logarithm of the flux MLE mimics Eq. (7). However, since the log-scale is used to prevent sensor saturation, a simpler alternative is to use the non-saturating response curve of SPAD pixels \\( (h(x) = x) \\). The response curve takes the form of \\( 1 - \\exp (-\\alpha \\Phi (\\mathbf{x},t)) \\), where \\( \\alpha \\) is a flux-independent constant. As a major advantage, this response curve avoids the underflow issues of the log function that can occur in low-light scenarios [62]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.747, + 0.47, + 0.822 + ], + "angle": 0, + "content": "The SPAD's frame-rate determines the time-stamp resolution of emulated events. In Fig. 4, we show the events generated from a photon-cube acquired at a frame-rate of \\(96.8\\mathrm{kHz}\\) resulting in a time-stamp resolution of \\(\\sim 10~\\mu \\mathrm{s}\\) that is comparable to those of existing event cameras." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "How do SPAD-events differ from the output of a regular event camera? The main difference is the expression of temporal contrast, given by \\(\\partial_t h\\), is now \\(-\\partial_t \\exp(-\\alpha \\Phi(\\mathbf{x}, t))\\), instead of \\(\\partial_t \\log(\\Phi(\\mathbf{x}, t))\\). This dif" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.698, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.187, + 0.698, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.264, + 0.681, + 0.278 + ], + "angle": 0, + "content": "\\(160\\hat{x} +447\\hat{y}\\) pixels/sec" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.089, + 0.888, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.166, + 0.873, + 0.181 + ], + "angle": 0, + "content": "\\(100\\hat{x} + 282\\hat{y}\\) pixels/sec" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.187, + 0.888, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.718, + 0.264, + 0.879, + 0.278 + ], + "angle": 0, + "content": "\\(420\\hat{x} + 1177\\hat{y}\\) pixels/sec" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.289, + 0.892, + 0.41 + ], + "angle": 0, + "content": "Figure 6: Motion stack. Computing multiple linear projections with different trajectories can produce a stack of images where objects with matching velocity are sharp. Here, we show a traffic scene involving four cars that have four different velocities. By suitably altering the slope of the linear trajectory, we can produce images where only one of the cars appear sharp at a time. We indicate the slope of the trajectories chosen and the objects that are \"in-focus\"." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.558 + ], + "angle": 0, + "content": "ference poses no compatibility issues for a large class of event-vision algorithms that utilize a grid of events [54, 57, 66] or brightness changes [18]. We show examples of downstream applications using SPAD-events in Supplementary Note 2. Finally, SPAD-events can be easily augmented with spatially- and temporally-aligned intensity information—a synergistic combination that has been exploited by several recent event-vision works [18, 25, 74]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.568, + 0.684, + 0.584 + ], + "angle": 0, + "content": "5.3. Motion Projections" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.892, + 0.668 + ], + "angle": 0, + "content": "Having described the emulation of cameras that capture coded exposures and temporal contrasts, we now shift our attention to cameras that emulate sensor motion during exposure, viz. motion cameras. We describe two useful trajectories when emulating motion cameras using Eq. (6)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.762 + ], + "angle": 0, + "content": "Linear trajectory. The simplest sensor trajectory involves linear motion, where \\(\\mathbf{r}(t) = (bt + c)\\hat{\\mathbf{p}}\\) for some constants \\(b, c \\in \\mathbb{R}\\) and unit vector \\(\\hat{\\mathbf{p}}\\). As Fig. 5 (top row) shows, this can change the scene's frame of reference: making moving objects appear stationary and vice-versa." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Motion-invariant parabolic projection. If motion is along \\(\\hat{\\mathbf{p}}\\), parabolic integration produces a motion-invariant image [33]—all objects, irrespective of their velocity, are blurred by the same point spread function (PSF), up to a linear shift. Thus, a deblurred parabolic capture produces a sharp image of all velocity groups (Fig. 5 (bottom row)). The parabolic trajectory is given by \\(\\mathbf{r}(t) = (at^2 + bt + c)\\hat{\\mathbf{p}}\\). We choose \\(a\\) based on the maximum object velocity and \\(b, c\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8170" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.09, + 0.89, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.239, + 0.895, + 0.302 + ], + "angle": 0, + "content": "Figure 7: High-speed videography at 2000 FPS of a tennis ball dropped into a bowl of water, from a \\(25\\mathrm{Hz}\\) readout. The conventional capture provides a visualization of the scene dynamics. It is challenging to reconstruct a large number of frames from a single compressive snapshot. Multi-bucket captures recover frames with significantly greater detail, such as the crown of water surrounding the ball. We include more sequences (e.g., a bursting balloon) in the supplementary material." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.31, + 0.892, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.543, + 0.895, + 0.605 + ], + "angle": 0, + "content": "Figure 8: Deblurring of traffic scenes using motion projections. Linear projections can recover details of moving objects if their velocity is known. When only the motion direction is known (e.g., road's orientation), a sharp image can be obtained by either deblurring a parabolic projection or by blending multiple randomly-sampled linear projections. We quantitatively compare against the compute- and bandwidth-expensive Quanta Burst Photography [41], based on PSNR and LPIPS [76]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.471, + 0.707 + ], + "angle": 0, + "content": "so the parabola's vertex lies at \\( T / 2 \\). We readily obtain the PSF by applying the parabolic integration to a delta input. Upon deconvolution using the PSF, a parabolic projection provides the optimal SNR for a blur-free image from single capture when only the direction of velocity is known." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.725, + 0.47, + 0.831 + ], + "angle": 0, + "content": "Ensembling linear projections. Finally, we leverage the flexibility of photon-cubes to compute multiple linear projections, as seen in Fig. 6. This produces a stack of images where one velocity group is motion-blur free at a time—or a 'motion stack', analogous to a focal stack. This novel construct can be used to compensate motion by blending stack images using cues such as blur orientation or optical flow." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.41, + 0.863 + ], + "angle": 0, + "content": "6. Hardware and Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We design a range of experiments to demonstrate the versatility of photon-cube projections: both when computa" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.738 + ], + "angle": 0, + "content": "tions occur after readout (Secs. 6.1 and 6.2), and when they are performed near-sensor on-chip (Sec. 6.3). All photon-cubes were acquired using the SwissSPAD2 array [67], operated using one of two sub-arrays, each having \\(512 \\times 256\\) pixels, and at a frame-rate of \\(96.8\\mathrm{kHz}\\). For the on-chip experiments, we use the UltraPhase compute architecture to interface with photon-cubes acquired by the SwissSPAD2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.754, + 0.712, + 0.77 + ], + "angle": 0, + "content": "6.1. SoDaCam Capabilities" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "High-speed compressive imaging. We reconstruct 80 frames from compressive snapshots that are emulated at \\(25\\mathrm{Hz}\\), resulting in a 2000 FPS video. We decode compressive snapshots using a plug-and-play (PnP) approach, PnP-FastDVDNet [73]. As Fig. 7 shows, it is challenging to recover a large number of frames from a single compressive measurement. Using the proposed multi-bucket scheme significantly improves the quality of video reconstruction." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8171" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.096, + 0.204, + 0.124 + ], + "angle": 0, + "content": "Sufficient Light 158 lux, 1.03 PPP" + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.126, + 0.18, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.219, + 0.096, + 0.323, + 0.123 + ], + "angle": 0, + "content": "Low light 0.9 lux, 0.1 PPP" + }, + { + "type": "image", + "bbox": [ + 0.239, + 0.125, + 0.309, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.096, + 0.458, + 0.124 + ], + "angle": 0, + "content": "Very low light 0.3 lux, 0.03 PPP" + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.125, + 0.436, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.183, + 0.347, + 0.197 + ], + "angle": 0, + "content": "DSLR captures (30 Hz)" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.207, + 0.207, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.207, + 0.332, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.345, + 0.207, + 0.462, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.199, + 0.304, + 0.351, + 0.317 + ], + "angle": 0, + "content": "SPAD-emulated events" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.326, + 0.206, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.326, + 0.333, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.345, + 0.326, + 0.462, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.196, + 0.424, + 0.354, + 0.437 + ], + "angle": 0, + "content": "Prophesee EVK4 events" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.449, + 0.47, + 0.57 + ], + "angle": 0, + "content": "Figure 9: Comparison to a state-of-the-art event camera. SPAD-events can capture temporal gradients even when the light-level is reduced by \\(500\\times\\), by benefiting from their single-photon sensitivity and bounded brightness response curve. In contrast, low-light induces blur and deteriorates the Prophesee's event stream. As a measure of the light-level, we report the PPP (photons per pixel) averaged across bit-planes and a light meter's reading at the sensor location." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.645 + ], + "angle": 0, + "content": "While multi-bucket captures require more bandwidth, this can be partially amortized by coding only dynamic regions, which we show in Supplementary Note 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.47, + 0.834 + ], + "angle": 0, + "content": "Motion projections on a traffic scene. Fig. 8 shows two traffic scenes captured using a \\(50~\\mathrm{mm}\\) focal length lens and at \\(30\\mathrm{Hz}\\) emulation. When object velocity is known, a linear projection can make moving objects appear stationary. If only the velocity direction is known (e.g., road's orientation in Fig. 8), a parabolic projection provides a sharp reconstruction of all objects. We deblur parabolic captures using PnP-DnCNN [75]. We offer an improvement by randomly sampling 8 linear projections along the velocity direction and blending them using the optical flow predicted by RAFT [65] between two short-exposures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Low-light event imaging. Fig. 9 compares event-image visualizations of SPAD and that of a state-of-the-art commercial event sensor (Prophesee EVK4), across various" + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.096, + 0.63, + 0.107 + ], + "angle": 0, + "content": "Compressive frame" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.109, + 0.633, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.096, + 0.751, + 0.107 + ], + "angle": 0, + "content": "Recovered frame" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.109, + 0.762, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.096, + 0.867, + 0.108 + ], + "angle": 0, + "content": "Event image" + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.11, + 0.892, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.18, + 0.773, + 0.193 + ], + "angle": 0, + "content": "SwissSPAD2 projections" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.2, + 0.632, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.2, + 0.762, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.201, + 0.892, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.271, + 0.765, + 0.285 + ], + "angle": 0, + "content": "Infinicam projections" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.292, + 0.632, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.292, + 0.762, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.292, + 0.892, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.363, + 0.817, + 0.376 + ], + "angle": 0, + "content": "Infinicam projections (16× more light)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.387, + 0.893, + 0.508 + ], + "angle": 0, + "content": "Figure 10: Comparison against conventional high-speed acquisition at \\(4000\\mathrm{Hz}\\). (top) SPAD projections recover a \\(16\\times\\) compressive video and an event image of a spinning roulette wheel. (middle) Read-noise corrupts the incident flux in the Infinicam high-speed camera, removing details in frames which are compressed on-the-fly. (bottom) Although using a larger aperture to admit more light recovers some detail, noise and compression artifacts still persist." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.701 + ], + "angle": 0, + "content": "light levels, with an accumulation period of \\(33\\mathrm{ms}\\). For a fair comparison, we bin the Prophesee's events in blocks of \\(2\\times 2\\) pixels and use a smaller aperture to account for the lower fill-factor of the SPAD. We tuned event-generation parameters (contrast threshold, integrator decay rate) of both cameras at each light level. Low light induces blur and deteriorates the Prophesee's event stream. In contrast, SPAD-events continue to capture temporal gradients, due to the SPAD's low-light capabilities and its brightness-encoding response curve. We include an ablative study of brightness-encoding functions in Supplementary Note 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Our observations are in concurrence with recent works that examine the low-light performance of event cameras [19, 27], and show that SPAD-events can provide neuromorphic vision in these challenging-SNR scenarios." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.818, + 0.788 + ], + "angle": 0, + "content": "6.2. Comparison to High-Speed Cameras" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recall, as previously discussed in Sec. 4, that read-noise limits the per-frame SNR of high-speed cameras. To demonstrate this limitation, we compute projections using the \\(4\\mathrm{kHz}\\) acquisition of the Photon Infinicam, a conventional high-speed camera, at a resolution of \\(1246\\times 240\\) pixels. We operate the SwissSPAD2 and the Infinicam at ambient light conditions using the same lens specifications." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8172" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.201, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.201, + 0.201, + 0.214 + ], + "angle": 0, + "content": "UltraPhase board" + }, + { + "type": "image", + "bbox": [ + 0.208, + 0.089, + 0.462, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.201, + 0.452, + 0.214 + ], + "angle": 0, + "content": "Power and bandwidth measurement" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.23, + 0.47, + 0.322 + ], + "angle": 0, + "content": "Figure 11: Power and bandwidth requirements when computing photon-cube projections on UltraPhase [5] (left), a recent compute architecture designed for single-photon imaging, at \\(40\\mathrm{Hz}\\) readout. (right) Our projections act as a compression scheme for photon-cubes, resulting in dramatically reduced sensor-readout and power consumption." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.346, + 0.47, + 0.483 + ], + "angle": 0, + "content": "As Fig. 10 shows, read noise corrupts the incident signal in Infinicam and makes it impossible to derive any useful projections. The read noise could be averaged out to some extent if the Infinicam did not perform compression-on-the-fly, but compression is central to the camera's working and enables readout over USB. Using a larger aperture to admit more light improves the quality of computed projections, but the video reconstruction and event image remain considerably worse than corresponding outputs of the SPAD." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.496, + 0.385, + 0.513 + ], + "angle": 0, + "content": "6.3. Bandwidth and Power Implications" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.47, + 0.672 + ], + "angle": 0, + "content": "While Sec. 6.1 has demonstrated the capabilities of photon-cube projections, we now show that our projections can also be obtained in a bandwidth-efficient manner via near-sensor computations. We implement photon-cube projections on UltraPhase (Fig. 11 (left)), a novel compute architecture designed for single-photon imaging. UltraPhase consists of \\(3 \\times 6\\) processing cores, each of which interfaces with \\(4 \\times 4\\) pixels, and can be 3D stacked beneath a SPAD array. We include visualizations and programming details of a few example projections in Supplementary Note 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.854 + ], + "angle": 0, + "content": "We measure the readout and power consumption of UltraPhase when computing projections on 2500 bit-planes of the falling die sequence (Fig. 1). The projections include: VCS with 16 random binary masks, an event camera, a linear projection and a combination of the three. We output projections at 12-bit depth and calculate metrics based on the clock cycles required for both compute and readout. As seen in Fig. 11 (right), computing projections on-chip dramatically reduces sensor-readout and power consumption as compared to reading out the photon-cube. Finally, similar to existing event cameras, SPAD-events have a resource footprint that reflects the underlying scene dynamics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In summary, our on-chip experiments show that performing computations near-sensor can increase the viability of single-photon imaging in resource constrained settings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.785, + 0.107 + ], + "angle": 0, + "content": "7. Discussion and Future Outlook" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.892, + 0.314 + ], + "angle": 0, + "content": "SoDaCam provides a realization of reinterpretable software-defined cameras [2, 3, 20, 34, 51] at the fine temporal resolution of SPAD-acquired photon-cubes. The proposed computations, or photon-cube projections, can match and in some cases, surpass the capabilities of existing imaging systems. The software-defined nature of photon-cube projections provides functionalities that may be difficult to achieve in conventional sensors. These projections can reduce the readout and power-consumption of SPAD arrays and potentially spur widespread adoption of single-photon imaging in the consumer domain. Finally, future chip-to-chip communication standards may also make it feasible to compute projections on a camera image signal processor." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.892, + 0.472 + ], + "angle": 0, + "content": "Adding color to SoDaCam. One way to add color is by overlaying color filter arrays (CFAs) and perform demosaicing on the computed photon-cube projection: depending on the projection, demosaicing could be relatively simple or more complex. As a reference, Bayer CFAs have been considered in the context of both video compressive sensing [73] and event cameras [64]. Incorporating CFAs with motion projections requires careful considerations, e.g., avoiding integrating across pixel locations of differing color." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.63 + ], + "angle": 0, + "content": "Future outlook on SPAD characteristics. A key SPAD characteristic that determines several properties of emulated cameras is the frame rate. While no fundamental limitations prevent SPADs from being operated at the frame-rates utilized in this work (\\(\\sim 100\\mathrm{kHz}\\)), sensor readout and power constraints can preclude high speeds, especially in high-resolution SPAD arrays. Photon-cube projections can enable future large-format SPADs to preserve high-speed information with modest resource requirements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.892, + 0.773 + ], + "angle": 0, + "content": "A platform for comparing cameras. Comparing imaging modalities can be quite challenging since hardware realizations of sensors can differ in numerous aspects, such as their quantum efficiency, fill factor, pixel pitch and array resolution. By emulating their imaging models, SoDaCam can serve as a platform for hardware-agnostic comparisons; for instance, determining operating conditions where one imaging modality is advantageous over another." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "A Cambrian explosion of new cameras. Besides comparing cameras, by virtue of being software-defined, SoDa-Cam can also make it significantly easier to prototype and deploy new unconventional imaging models, and even facilitate sensor-in-the-loop optimization [44, 47, 63] by tailoring photon-cube projections for downstream computer-vision tasks. This is an exciting future line of research." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8173" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.116, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Phantom-v2640. https://www.phantomhighspeed.com/products/cameraul ultrahigh4mpx/v2640. Accessed: 2023-01-28. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.157, + 0.471, + 0.198 + ], + "angle": 0, + "content": "[2] E. H. Adelson, J. R. Bergen, et al. The plenoptic function and the elements of early vision. Computational models of visual processing, 1(2):3-20, 1991. 1, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.199, + 0.469, + 0.266 + ], + "angle": 0, + "content": "[3] A. Agrawal, A. Veeraraghavan, and R. Raskar. Reinterpretable imager: Towards variable post-capture space, angle and time resolution in photography. In Computer Graphics Forum, volume 29, pages 763-772. Wiley Online Library, 2010. 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.268, + 0.469, + 0.322 + ], + "angle": 0, + "content": "[4] I. M. Antolovic, S. Burri, C. Bruschini, R. Hoebe, and E. Charbon. Nonuniformity analysis of a 65-kpixel CMOS SPAD imager. IEEE Transactions on Electron Devices, 63 (1):57-64, 2016. doi: 10.1109/TED.2015.2458295. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.324, + 0.469, + 0.351 + ], + "angle": 0, + "content": "[5] A. Ardelean. Computational Imaging SPAD Cameras. PhD thesis, École polytechnique fédérale de Lausanne, 2023. 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.352, + 0.468, + 0.391 + ], + "angle": 0, + "content": "[6] A. Boukhayma, A. Peizerat, and C. Enz. A sub-0.5 electron read noise VGA image sensor in a standard CMOS process. IEEE Journal of Solid-State Circuits, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.393, + 0.468, + 0.446 + ], + "angle": 0, + "content": "[7] J. Canny. A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-8(6):679-698, 1986. doi: 10.1109/TPAMI.1986.4767851. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.448, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[8] S. J. Carey, A. Lopich, D. R. Barr, B. Wang, and P. Dudek. A 100,000 fps vision sensor with embedded 535GOPS/W \\(256 \\times 256\\) SIMD processor array. In 2013 Symposium on VLSI Circuits, pages C182-C183, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.504, + 0.469, + 0.557 + ], + "angle": 0, + "content": "[9] J. Chen, S. J. Carey, and P. Dudek. Feature extraction using a portable vision system. In IEEE/RSJ Int. Conf. Intell. Robots Syst., Workshop Vis.-based Agile Auton. Navigation UAVs, volume 2, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.559, + 0.47, + 0.613 + ], + "angle": 0, + "content": "[10] S. Chen and M. Guo. Live demonstration: Celex-v: A 1m pixel multi-mode event-based sensor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2019. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.614, + 0.469, + 0.668 + ], + "angle": 0, + "content": "[11] N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), volume 1, pages 886-893. IEEE, 2005. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.669, + 0.469, + 0.763 + ], + "angle": 0, + "content": "[12] P. E. Debevec and J. Malik. Recovering high dynamic range radiance maps from photographs. In Proceedings of the 24th Annual Conference on Computer Graphics and Interactive Techniques, page 369-378. ACM Press/Addison-Wesley Publishing Co., 1997. doi: 10.1145/258734.258884. URL https://doi.org/10.1145/258734.258884. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.766, + 0.469, + 0.847 + ], + "angle": 0, + "content": "[13] F. M. Della Rocca, H. Mai, S. W. Hutchings, T. Al Abbas, K. Buckbee, A. Tsiamis, P. Lomax, I. Gyongy, N. A. Dutton, and R. K. Henderson. A \\(128 \\times 128\\) SPAD motion-triggered time-of-flight image sensor with in-pixel histogram and column-parallel vision processor. IEEE Journal of Solid-State Circuits, 55(7):1762-1775, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.849, + 0.469, + 0.89 + ], + "angle": 0, + "content": "[14] T. Finateu, A. Niwa, D. Matolin, K. Tsuchimoto, A. Mascheroni, E. Reynaud, P. Mostafalu, F. Brady, L. Chotard, F. LeGoff, H. Takahashi, H. Wakabayashi," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.471, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.188 + ], + "angle": 0, + "content": "Y. Oike, and C. Posch. 5.10 a \\(1280 \\times 720\\) back-illuminated stacked temporal contrast event-based vision sensor with \\(4.86\\mu m\\) pixels, 1.066GEPS readout, programmable event-rate controller and compressive data-formating pipeline. In 2020 IEEE International Solid-State Circuits Conference - (ISSCC), pages 112-114, 2020. doi: 10.1109/ISSCC19947.2020.9063149. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[15] E. R. Fossum. What to do with sub-diffraction-limit (SDL) pixels?—a proposal for a gigapixel digital film sensor (DFS). In IEEE Workshop on Charge-Coupled Devices and Advanced Image Sensors, pages 214–217, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.246, + 0.892, + 0.313 + ], + "angle": 0, + "content": "[16] E. R. Fossum. The quanta image sensor (QIS): Concepts and challenges. In Imaging and Applied Optics, page JTuE1. Optica Publishing Group, 2011. doi: 10.1364/COSI.2011. JTuE1. URL http://opg.optica.org/abstract.cfm?URI=COSI-2011-JTuE1.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.315, + 0.892, + 0.355 + ], + "angle": 0, + "content": "[17] E. R. Fossum, J. Ma, S. Masoodian, L. Anzagira, and R. Zizza. The quanta image sensor: Every photon counts. Sensors, 16(8):1260, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.356, + 0.892, + 0.409 + ], + "angle": 0, + "content": "[18] D. Gehrig, H. Rebecq, G. Gallego, and D. Scaramuzza. Eklt: Asynchronous photometric feature tracking using events and frames. International Journal of Computer Vision, 128(3): 601-618, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.412, + 0.892, + 0.451 + ], + "angle": 0, + "content": "[19] R. Graca and T. Delbruck. Unraveling the paradox of intensity-dependent DVS pixel noise. arXiv preprint arXiv:2109.08640, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.453, + 0.892, + 0.535 + ], + "angle": 0, + "content": "[20] M. Gupta, A. Agrawal, A. Veeraraghavan, and S. G. Narasimhan. Flexible voxels for motion-aware videography. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part I 11, pages 100-114. Springer, 2010. 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.536, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[21] F. Gutierrez-Barragan, A. Ingle, T. Seets, M. Gupta, and A. Velten. Compressive single-photon 3D cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17854-17864, June 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.605, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[22] I. Gyongy, N. A. Dutton, and R. K. Henderson. Single-photon tracking for high-speed vision. Sensors, 18(2):323, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[23] I. Gyongy, S. W. Hutchings, A. Halimi, M. Tyler, S. Chan, F. Zhu, S. McLaughlin, R. K. Henderson, and J. Leach. High-speed 3D sensing via hybrid-mode imaging and guided upsampling. Optica, 7(10):1253-1260, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.702, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[24] C. Harris, M. Stephens, et al. A combined corner and edge detector. In *Alvey vision conference*, volume 15, pages 10-5244. CiteSeer, 1988. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.743, + 0.892, + 0.798 + ], + "angle": 0, + "content": "[25] J. Hidalgo-Carrió, G. Gallego, and D. Scaramuzza. Event-aided direct sparse odometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5781-5790, June 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.799, + 0.892, + 0.826 + ], + "angle": 0, + "content": "[26] B. K. Horn and B. G. Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185-203, 1981. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.827, + 0.892, + 0.88 + ], + "angle": 0, + "content": "[27] Y. Hu, S.-C. Liu, and T. Delbruck. v2e: From video frames to realistic DVS events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1312-1321, June 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.882, + 0.892, + 0.895 + ], + "angle": 0, + "content": "[28] J. Igual. Photographic noise performance measures based on" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8174" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "raw files analysis of consumer cameras. Electronics, 8(11): 1284, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.472, + 0.175 + ], + "angle": 0, + "content": "[29] A. Ingle, A. Velten, and M. Gupta. High Flux Passive Imaging With Single-Photon Sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[30] A. Ingle, T. Seets, M. Buttafava, S. Gupta, A. Tosi, M. Gupta, and A. Velten. Passive inter-photon imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.285 + ], + "angle": 0, + "content": "[31] K. Iwabuchi, Y. Kameda, and T. Hamamoto. Image quality improvements based on motion-based deblurring for single-photon imaging. IEEE Access, 9:30080-30094, 2021. doi: 10.1109/ACCESS.2021.3059293. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.287, + 0.469, + 0.327 + ], + "angle": 0, + "content": "[32] Y. Jiang, I. Choi, J. Jiang, and J. Gu. HDR video reconstruction with tri-exposure quad-bayer sensors. arXiv preprint arXiv:2103.10982, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.328, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[33] A. Levin, P. Sand, T. S. Cho, F. Durand, and W. T. Freeman. Motion-invariant photography. ACM Transactions on Graphics (TOG), 27(3):1-9, 2008. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.469, + 0.411 + ], + "angle": 0, + "content": "[34] M. Levoy and P. Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 31-42, 1996. 1, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.412, + 0.469, + 0.493 + ], + "angle": 0, + "content": "[35] Y. Li, M. Qi, R. Gulve, M. Wei, R. Genov, K. N. Kutulakos, and W. Heidrich. End-to-end video compressive sensing using anderson-accelerated unrolled networks. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12, 2020. doi: 10.1109/ICCP48838.2020.9105237. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.494, + 0.469, + 0.535 + ], + "angle": 0, + "content": "[36] P. Lichtsteiner. 64x64 event-driven logarithmic temporal derivative silicon retina. In Program 2003 IEEE Workshop on CCD and AIS, 2003. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.536, + 0.469, + 0.605 + ], + "angle": 0, + "content": "[37] Y. Liu, F. Gutierrez-Barragan, A. Ingle, M. Gupta, and A. Velten. Single-photon camera guided extreme dynamic range imaging. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1575-1585, January 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.606, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[38] P. Llull, X. Liao, X. Yuan, J. Yang, D. Kittle, L. Carin, G. Sapiro, and D. J. Brady. Coded aperture compressive temporal imaging. Opt. Express, 21(9):10526-10545, May 2013. doi: 10.1364/OE.21.010526. URL https://opg.optica.org/oe/abstract.cfm?URI=oe-21-9-10526.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[39] B. D. Lucas and T. Kanade. An iterative image registration technique with an application to stereo vision. In *IJCAI'81: 7th international joint conference on Artificial intelligence*, volume 2, pages 674–679, 1981. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.469, + 0.826 + ], + "angle": 0, + "content": "[40] J. Ma, S. Masoodian, D. A. Starkey, and E. R. Fossum. Photon-number-resolving megapixel image sensor at room temperature without avalanche gain. Optica, 4(12): 1474-1481, Dec 2017. doi: 10.1364/OPTICA.4.001474. URL http://www.osapublishing.org/optica/abstract.cfm?URI=optica-4-12-1474.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.827, + 0.469, + 0.88 + ], + "angle": 0, + "content": "[41] S. Ma, S. Gupta, A. C. Ulku, C. Bruschini, E. Charbon, and M. Gupta. Quanta burst photography. ACM Transactions on Graphics, 39(4):1-16, July 2020. ISSN 0730-0301, 1557-7368. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.882, + 0.469, + 0.896 + ], + "angle": 0, + "content": "[42] S. Ma, P. Mos, E. Charbon, and M. Gupta. Burst vi" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "sion using single-photon cameras. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5375-5385, January 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[43] S. Mann and R. Picard. Beingundigital'with digital cameras. MIT Media Lab Perceptual, 1:2, 1994. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[44] J. N. Martel, L. K. Mueller, S. J. Carey, P. Dudek, and G. Wetzstein. Neural sensors: Learning pixel exposures for HDR imaging and video compressive sensing with programmable sensors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(7):1642-1653, 2020. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[45] J. N. P. Martel, L. K. Müller, S. J. Carey, and P. Dudek. Parallel HDR tone mapping and auto-focus on a cellular processor array vision chip. In 2016 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1430-1433, 2016. doi: 10.1109/ISCAS.2016.7527519. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.301, + 0.892, + 0.369 + ], + "angle": 0, + "content": "[46] J. N. P. Martel, L. K. Müller, S. J. Carey, and P. Dudek. High-speed depth from focus on a programmable vision chip using a focus tunable lens. In 2017 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-4, 2017. doi: 10.1109/ISCAS.2017.8050548.3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.37, + 0.892, + 0.425 + ], + "angle": 0, + "content": "[47] C. A. Metzler, H. Ikoma, Y. Peng, and G. Wetzstein. Deep optics for single-shot high-dynamic-range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.426, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[48] K. Morimoto, A. Ardelean, M.-L. Wu, A. C. Ulku, I. M. Antolovic, C. Bruschini, and E. Charbon. Megapixel time-gated SPAD image sensor for 2D and 3D imaging applications. Optica, 7(4):346-354, Apr. 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.481, + 0.892, + 0.618 + ], + "angle": 0, + "content": "[49] K. Morimoto, J. Iwata, M. Shinohara, H. Sekine, A. Abdelghafar, H. Tsuchiya, Y. Kuroda, K. Tojima, W. Endo, Y. Maehashi, Y. Ota, T. Sasago, S. Maekawa, S. Hikosaka, T. Kanou, A. Kato, T. Tezuka, S. Yoshizaki, T. Ogawa, K. Uehira, A. Ehara, F. Inui, Y. Matsuno, K. Sakurai, and T. Ichikawa. 3.2 megapixel 3D-stacked charge focusing SPAD for low-light imaging and depth sensing. In 2021 IEEE International Electron Devices Meeting (IEDM), pages 20.2.1-20.2.4, 2021. doi: 10.1109/IEDM19574.2021.9720605.1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[50] S. Namiki, S. Sato, Y. Kameda, and T. Hamamoto. Imaging method using multi-threshold pattern for photon detection of quanta image sensor. In International Workshop on Advanced Imaging Technology (IWAIT) 2022, volume 12177, page 1217702. SPIE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[51] S. K. Nayar, V. Branzoi, and T. E. Boult. Programmable imaging: Towards a flexible camera. International Journal of Computer Vision, 70:7-22, 2006. 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.731, + 0.892, + 0.797 + ], + "angle": 0, + "content": "[52] C. Posch, D. Matolin, and R. Wohlgenannt. A QVGA 143 dB Dynamic Range Frame-Free PWM Image Sensor With Lossless Pixel-Level Video Compression and Time-Domain CDS. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2011. doi: 10.1109/JSSC.2010.2085952. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.799, + 0.892, + 0.841 + ], + "angle": 0, + "content": "[53] R. Raskar, A. Agrawal, and J. Tumblin. Coded exposure photography: motion deblurring using fluttered shutter. In Acm Siggraph 2006 Papers, pages 795-804. 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.842, + 0.892, + 0.896 + ], + "angle": 0, + "content": "[54] H. Rebecq, R. Ranftl, V. Koltun, and D. Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "8175" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[55] D. Reddy, A. Veeraraghavan, and R. Chellappa. P2C2: Programmable pixel compressive camera for high speed imaging. In CVPR 2011, pages 329-336, 2011. doi: 10.1109/CVPR.2011.5995542. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.472, + 0.215 + ], + "angle": 0, + "content": "[56] X. Ren, P. W. Connolly, A. Halimi, Y. Altmann, S. McLaughlin, I. Gyongy, R. K. Henderson, and G. S. Buller. High-resolution depth profiling using a range-gated CMOS SPAD quanta image sensor. Optics express, 26(5):5541-5557, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.217, + 0.471, + 0.285 + ], + "angle": 0, + "content": "[57] C. Scheerlinck, H. Rebecq, D. Gehrig, N. Barnes, R. Mahony, and D. Scaramuzza. Fast image reconstruction with an event camera. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), March 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.287, + 0.472, + 0.354 + ], + "angle": 0, + "content": "[58] T. Seets, A. Ingle, M. Laurenzis, and A. Velten. Motion adaptive deblurring with single-photon cameras. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1945-1954, January 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.355, + 0.471, + 0.451 + ], + "angle": 0, + "content": "[59] M.-W. Seo, Y. Shirakawa, Y. Masuda, Y. Kawata, K. Kagawa, K. Yasutomi, and S. Kawahito. 4.3 a programmable sub-nanosecond time-gated 4-tap lock-in pixel cmos image sensor for real-time fluorescence lifetime imaging microscopy. In 2017 IEEE International Solid-State Circuits Conference (ISSCC), pages 70-71, 2017. doi: 10.1109/ISSCC.2017.7870265. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.452, + 0.471, + 0.521 + ], + "angle": 0, + "content": "[60] T. Serrano-Gotarredona and B. Linares-Barranco. A \\(128 \\times 128\\) \\(1.5\\%\\) contrast sensitivity \\(0.9\\%\\) FPN \\(3\\mu s\\) latency \\(4\\mathrm{mW}\\) asynchronous frame-free dynamic vision sensor using transimpedance preamplifiers. IEEE Journal of Solid-State Circuits, 48(3):827-838, 2013. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.522, + 0.471, + 0.59 + ], + "angle": 0, + "content": "[61] P. Shedligeri, A. S., and K. Mitra. A unified framework for compressive video recovery from coded exposure techniques. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1600-1609, January 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.471, + 0.631 + ], + "angle": 0, + "content": "[62] C. Shi, N. Song, W. Li, Y. Li, B. Wei, H. Liu, and J. Jin. A review of event-based indoor positioning and navigation. 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.471, + 0.701 + ], + "angle": 0, + "content": "[63] V. Sitzmann, S. Diamond, Y. Peng, X. Dun, S. Boyd, W. Heidrich, F. Heide, and G. Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and super-resolution imaging. ACM Transactions on Graphics (TOG), 37(4):114, 2018. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.471, + 0.783 + ], + "angle": 0, + "content": "[64] G. Taverni, D. Paul Moeys, C. Li, C. Cavaco, V. Motsnyi, D. San Segundo Bello, and T. Delbruck. Front and back illuminated dynamic and active pixel vision sensors comparison. IEEE Transactions on Circuits and Systems II: Express Briefs, 65(5):677-681, 2018. doi: 10.1109/TCSII.2018.2824899. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.785, + 0.471, + 0.825 + ], + "angle": 0, + "content": "[65] Z. Teed and J. Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European Conference on Computer Vision, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.826, + 0.471, + 0.896 + ], + "angle": 0, + "content": "[66] S. Tulyakov, D. Gehrig, S. Georgoulis, J. Erbach, M. Gehrig, Y. Li, and D. Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16155-16164, June 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[67] A. C. Ulku, C. Bruschini, I. M. Antolovic, Y. Kuo, R. Ankri, S. Weiss, X. Michalet, and E. Charbon. A \\(512 \\times 512\\) SPAD Image Sensor With Integrated Gating for Widefield FLIM. IEEE Journal of Selected Topics in Quantum Electronics, 25 (1):1-12, Jan. 2019. ISSN 1077-260X, 1558-4542. doi: 10. 1109/JSTQE.2018.2867439. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[68] G. Wan, X. Li, G. Agranov, M. Levoy, and M. Horowitz. CMOS image sensors with multi-bucket pixels for computational photography. IEEE Journal of Solid-State Circuits, 47 (4):1031-1042, 2012. doi: 10.1109/JSSC.2012.2185189. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.232, + 0.892, + 0.298 + ], + "angle": 0, + "content": "[69] M. Wei, N. Sarhangnejad, Z. Xia, N. Gusev, N. Katic, R. Genov, and K. N. Kutulakos. Coded two-bucket cameras for computer vision. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.3, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[70] M. White, S. Ghajari, T. Zhang, A. Dave, A. Veeraraghavan, and A. Molnar. A differential SPAD array architecture in \\(0.18\\mu \\mathrm{m}\\) CMOS for HDR imaging. In 2022 IEEE International Symposium on Circuits and Systems (ISCAS), pages 292-296, 2022. doi: 10.1109/ISCAS48785.2022.9937558.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.383, + 0.892, + 0.493 + ], + "angle": 0, + "content": "[71] T. Yamazaki, H. Katayama, S. Uehara, A. Nose, M. Kobayashi, S. Shida, M. Odahara, K. Takamiya, Y. Hisamatsu, S. Matsumoto, L. Miyashita, Y. Watanabe, T. Izawa, Y. Muramatsu, and M. Ishikawa. 4.9 a 1ms high-speed vision chip with 3D-stacked 140GOPS column-parallel PEs for spatio-temporal image processing. In 2017 IEEE International Solid-State Circuits Conference (ISSCC), pages 82–83, 2017. doi: 10.1109/ISSCC.2017.7870271. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.494, + 0.892, + 0.549 + ], + "angle": 0, + "content": "[72] F. Yang, Y. M. Lu, L. Sbaiz, and M. Vetterli. Bits from photons: Oversampled image acquisition using binary poisson statistics. IEEE Transactions on Image Processing, 21(4): 1421-1436, 2012. doi: 10.1109/TIP.2011.2179306. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.55, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[73] X. Yuan, Y. Liu, J. Suo, F. Durand, and Q. Dai. Plug-and-play algorithms for video snapshot compressive imaging. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):7093-7111, 2022. doi: 10.1109/TPAMI.2021.3099035.7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[74] J. Zhang, X. Yang, Y. Fu, X. Wei, B. Yin, and B. Dong. Object tracking by jointly exploiting frame and event domain. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13043-13052, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[75] K. Zhang, Y. Li, W. Zuo, L. Zhang, L. Van Gool, and R. Timofte. Plug-and-play image restoration with deep denoiser prior. arXiv preprint, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.716, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[76] R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.771, + 0.892, + 0.839 + ], + "angle": 0, + "content": "[77] T. Zhang, M. J. White, A. Dave, S. Ghajari, A. Raghuram, A. C. Molnar, and A. Veeraraghavan. First arrival differential LiDAR. In 2022 IEEE International Conference on Computational Photography (ICCP), pages 1-12, 2022. doi: 10.1109/ICCP54855.2022.9887683. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.84, + 0.892, + 0.88 + ], + "angle": 0, + "content": "[78] W. Zhang and L. Lin. Light field flow estimation based on occlusion detection. Journal of Computer and Communications, 5(3):1-9, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.88 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8176" + } + ] +] \ No newline at end of file diff --git a/2023/SoDaCam_ Software-defined Cameras via Single-Photon Imaging/ce1d43a9-f5c5-4d36-93a7-015029e926c0_origin.pdf b/2023/SoDaCam_ Software-defined Cameras via Single-Photon Imaging/ce1d43a9-f5c5-4d36-93a7-015029e926c0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b7070a4481237fb3f6347ba555a7805f24466941 --- /dev/null +++ b/2023/SoDaCam_ Software-defined Cameras via Single-Photon Imaging/ce1d43a9-f5c5-4d36-93a7-015029e926c0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5dc64c3568c4b0eaabd76c1d8080006afd4b8dcad6b23abffc0e33e151096285 +size 9821467 diff --git a/2023/SoDaCam_ Software-defined Cameras via Single-Photon Imaging/full.md b/2023/SoDaCam_ Software-defined Cameras via Single-Photon Imaging/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1e5109c507eb6a1cdf8fba0a2976fb0b7d2a3353 --- /dev/null +++ b/2023/SoDaCam_ Software-defined Cameras via Single-Photon Imaging/full.md @@ -0,0 +1,487 @@ +# SoDaCam: Software-defined Cameras via Single-Photon Imaging + +Varun Sundar† + +vsundar4@cs.wisc.edu + +Andrei Ardelean† + +a.ardelean@epfl.ch + +Tristan Swedish + +tristan@ubicept.com + +Claudio Bruschini† + +{claudio.bruschini, edoardo.charbon}@epfl.ch + +Mohit Gupta†,§ + +mohitg@cs.wisc.edu + +†University of Wisconsin-Madison ‡École polytechnique fédérale de Lausanne §Ubicept + +# Abstract + +Reinterpretable cameras are defined by their post-processing capabilities that exceed traditional imaging. We present "SoDaCam" that provides reinterpretable cameras at the granularity of photons, from photon-cubes acquired by single-photon devices. Photon-cubes represent the spatio-temporal detections of photons as a sequence of binary frames, at frame-rates as high as $100\mathrm{kHz}$ . We show that simple transformations of the photon-cube, or photon-cube projections, provide the functionality of numerous imaging systems including: exposure bracketing, flutter shutter cameras, video compressive systems, event cameras, and even cameras that move during exposure. Our photon-cube projections offer the flexibility of being software-defined constructs that are only limited by what is computable, and shot-noise. We exploit this flexibility to provide new capabilities for the emulated cameras. As an added benefit, our projections provide camera-dependent compression of photon-cubes, which we demonstrate using an implementation of our projections on a novel compute architecture that is designed for single-photon imaging. + +# 1. Introduction + +Throughout the history of imaging, sensing technologies and the corresponding processing have developed hand-in-hand. In fact, sensing technologies have, to some extent, defined the scope of processing captured data. In the film era, instances of such processing included dodging and burning. The advent of digital cameras provided processing at the granularity of pixels and paved the way for modern computer vision. Light field cameras [34, 78], by sampling the plenoptic function [2], allowed post-capture processing at + +the granularity of light rays, enabling novel functionalities such as refocusing photos after-capture. The logical limit of post-capture processing, given the fundamental quantization of light, would be at the level of individual photons. What would imaging look like if we could perform computational processing on individual photons? + +In this work, we show that photon data captured by a new class of single-photon detectors, called single-photon avalanche diodes (SPADs), makes it possible to emulate a wide range of imaging modalities such as exposure bracketing [12], video compressive systems [38, 55] and event cameras [52, 60]. A user then has the flexibility to choose one (or even multiple) of these functionalities post-capture (Fig. 1 (top)). SPAD arrays can operate as extremely high frame-rate photon detectors ( $\sim 100\mathrm{kHz}$ ), producing a temporal sequence of binary frames called a photon-cube [16]. We show that computing photon-cube projections, which are simple linear and shift operations, can reinterpret the photon-cube to achieve novel post-capture imaging functionalities in a software-defined manner (Fig. 1 (middle)). + +As case studies, we emulate three distinct imagers: high-speed video compressive imaging; event cameras which respond to dynamic scene content; and motion projections which emulate sensor motion, without any real camera movement. Fig. 1 (bottom) shows the outputs of these cameras that are derived from the same photon-cube. + +Computing photon-cube projections. One way to obtain photon-cube projections is to read the entire photon-cube off the SPAD array and then perform relevant computations off-chip; we adopt this strategy for our experiments in Secs. 6.1 and 6.2. While reasonable for certain applications, reading out photon-cubes requires an exorbitant data-bandwidth, which can be up to 100 Gbps for a 1 MPixel array—well beyond the capacity of existing data peripherals. Such readout considerations will become center stage as large-format SPAD arrays are fabricated [48, 49]. + +An alternative is to avoid transferring the entire photon- + +![](images/ac3cc844e8b4994e3c96d64b90cda8bef579a768f693e45dc790917fe510e938.jpg) + +![](images/517d947cd97d8ef1890b948664efad8ff9c7df797efb8f796c851fc0a82dba69.jpg) +Capabilities of photon-cube projections + +![](images/d7196f7c9edadd2ca73635dcac2fe8f2ff409c3d5295acaaa5ae0b6845806c80.jpg) +Coded exposure + +![](images/dcb2d461ccca256ef02b4a145c5a5ee01e3c3521a0e828babfb30d37ed461e52.jpg) + +![](images/1f7993e7b5524f5c2010889f3c15273ef20bac7a4baafbbd6f137b6d25424784.jpg) +$12 \times$ video reconstruction (raster order) Snapshot compressive imager (Sec. 5.1) + +![](images/e52254981b0ce021b1dc902ea604da4da8a713e730883b34b1ab21a327e82312.jpg) +3D scatter plot, $0\leq t\leq 0.1s$ + +![](images/c398e22cbebc862536d57707aec275a589f0709aa54198923293d85b6757594d.jpg) +Event image, $0.5 \leq t \leq 0.53$ s Event camera (Sec. 5.2) +Figure 1: (top) SoDaCam can emulate a variety of cameras from the photon-cubes acquired by single-photon devices. (middle) Photon-cubes represent the spatio-temporal detection of photons as a sequence of binary frames. Projections of the photon-cube, when computed either on or off-chip, result in reinterpretable and software-defined cameras. We demonstrate the versatility of photon-cube projections on a real dynamic scene: a die falls on a table, bounces, spins in the air and later ricochets off a nearby toy top. (bottom) The cameras emulated by our photon-cube projections can produce: a $12 \times$ high-speed video from a single compressive snapshot, event-stream representations of two time intervals (blue and red depict positive and negative spikes respectively), an image where the die appears stationary, as well as a motion-deblurred image. + +![](images/094aa9f7ec7d9cf191839e995fc0d270dba4e4c82d208c93b1925c638e477261.jpg) +Static camera + +![](images/8c24a2275f618d791908dd8fee6b565343c6620eb9310bf98410846f85bf3cd7.jpg) +Parabolic trajectory Motion Pro + +![](images/8bd7d4db4e90149d22171812ec99872fdd68a2f1a432339e077df1406e4e1e8c.jpg) +Linear trajectory + +![](images/effe023a8ca759e86bd7881337a3dcf4ea424fcc72fd6118da0fb4f2da2cc3b2.jpg) +Parabolic deblurred + +cube by computing projections near sensor. As a proof-of-concept, we implement photon-cube projections on Ultra-Phase [5], a recently-developed programmable SPAD imager with independent processing cores that have dedicated RAM and instruction memory. We show, in Sec. 6.3, that computing projections on-chip greatly reduces sensor readout and, as a consequence, power consumption. + +Implications: Toward a photon-level software-defined camera. The photon-cube projections introduced in this paper are computational constructs that provide a realization of software-defined cameras or SoDaCam. Being software-defined, SoDaCam can emulate multiple cameras simultaneously without additional hardware complexity. SoDaCam, by going beyond baked in hardware choices, + +unlocks hitherto unseen capabilities—such as 2000 FPS video from $25\mathrm{Hz}$ readout (Fig. 7); event imaging in very low-light conditions (Fig. 9); and motion stacks, which are a stack of images where in each image, objects only in certain velocity ranges appear sharp (Fig. 6). + +Limitations. The SPAD array [67] used in this work has a relatively low spatial resolution $(512 \times 256)$ , and a low fill-factor $(\sim 10\%)$ owing to the lack of microlenses in the prototype used. Similarly, the near-sensor processor that we use has limited capabilities compared to off-chip processors. However, with rapid progress in the development of single-photon cameras [48, 49] and increasing interest in near-sensor processors, we anticipate that many of these shortcomings will be addressed in the upcoming years. + +# 2. Related Work + +Reinterpretable imaging has previously been explored at the granularity of light rays [3], by modulating the plenoptic function, and at the level of spatio-temporal voxels [20], by using fast per-pixel shutters. SoDaCam represents a logical culmination of reinterpretability at the level of photon detections, that facilitates multiple post-capture imaging functionalities. + +Programmable imaging using a digital micromirror device was first introduced in Nayar et al. [51] to perform precapture radiometric manipulations. Modern programmable cameras are typically near-sensor processors [8, 69, 71] that can perform limited operations in analog [46, 55, 69], while more complex operations [9, 45] occur after analog-to-digital conversion (ADC). In contrast, by performing post-capture computations directly on photon detections, we can perform complex operations without incurring the read-noise penalty that is associated with ADC. + +Passive single-photon imaging. Only recently have SPADs been utilized as passive imaging devices, with applications in high-dynamic range imaging [29, 30, 37, 50], motion-compensation [31, 58], burst photography [41, 42] and object tracking [22]. Compared to compute-intensive burst-photography methods [41], our proposed techniques involve lightweight computations that can be performed near sensor. These computations can also be performed using other single-photon imagers such as Jots [15, 17], which feature higher sensor resolution and photon-efficiency [40], albeit at lower frame-rates and higher read noise. + +Reducing the readout of SPADs. Several data reduction strategies have been proposed in the context of SPADs that are used to timestamp incident photons, including: coarse histograms [13, 23, 56], compressive histograms [21], and + +measuring differential time-of-arrivals [70, 77]. When SPADs are operated as photon-detectors, multi-bit counting [48], or summing binary frames, can reduce readout. While compression is not our main objective, we show that photon-cube projections act as camera-specific compression schemes that dramatically reduce sensor readout. + +# 3. Background: Single-Photon Imaging Model + +A SPAD array captures incident light as a photon-cube: a temporal sequence of binary frames that represents the pixel-wise detection of photons across their respective exposure windows. We can model the stochastic arrival of photons as a Poisson process [72], allowing us to treat spatio-temporal values of the photon-cube as independent Bernoulli random variables with + +$$ +\Pr \left\{B _ {t} (\mathbf {x}) = 1 \right\} = 1 - e ^ {- (\eta \Phi (\mathbf {x}, t) + r _ {q}) w _ {\exp}}, \tag {1} +$$ + +where $B_{t}(\mathbf{x})$ represents the value of the photon-cube at pixel $\mathbf{x}$ and exposure index $1 \leq t \leq T$ , which receives a mean incident flux of intensity $\Phi (\mathbf{x},t)$ across its exposure of duration $w_{\mathrm{exp}}$ . Additionally, $\eta$ is the photon detection efficiency of the SPAD, and $r_q$ denotes the sensor's dark count rate—which is the rate of spurious counts unrelated to incident photons. While individual binary frames are extremely noisy, the temporal sum of the photon-cube + +$$ +\mathcal {I} _ {\text {s u m}} (\mathbf {x}) := \sum_ {t = 1} ^ {T} B _ {t} (\mathbf {x}), \tag {2} +$$ + +can produce an 'image' of the scene that is sharp in static regions, but blurry in dynamic regions (Fig. 2 (top)). Indeed, in static regions, the sum-image can be used to derive a maximum likelihood estimator of the scene intensity [4], given by $\hat{\Phi} (\mathbf{x}) = -\ln (1 - T^{-1}\mathcal{I}_{\mathrm{sum}}(\mathbf{x})) / \eta w_{\mathrm{exp}} - r_q / w_{\mathrm{exp}}$ + +# 4. Projections of the Photon-Cube + +The temporal sum described in Eq. (2) is a simple instance of projections of a photon-cube. Our key observation is that it is possible to compute a wide range of photon-cube projections, each of which emulates a unique sensing modality post-capture—including modalities that are difficult to achieve with conventional cameras. For example, varying the number of bit-planes that are summed over emulates exposure bracketing [12, 43], which is typically used for HDR imaging. Compared to conventional exposure bracketing, the emulated exposure stack, being software-defined, does not require spatial and temporal registration, which can often be error-prone. Fig. 2 (top) shows an example of an exposure stack computed from a photon-cube. + +Going further, we can gradually increase the complexity of the projections. For example, consider a coded exposure + +![](images/88db00ad5a71bae5abb411466fd9c30ff92884915b3f654f9556302ee6ee7913.jpg) +Short exposure + +![](images/b21ff920af31b12a2ecd1327561bfbdc5f5b85a5c4b93f69209d9fa6d380248d.jpg) +Medium exposure Exposure Stack + +![](images/ce91a8a8d2d809a185b9730f0d6e542eea36b63f8784704cdc8972c7b21f1090.jpg) +Long exposure + +![](images/bcdae6bd9a273b24708fe8f67fbac54a45d79cd3f0f19bcc69765cf09a74220f.jpg) +Spatially-varying exposure + +![](images/1ddfa74e7cdeb3cff1045a6a45807535dd0f29d46ce77d71a177dd9d542a59fc.jpg) + +![](images/59c3c43b9ed1c36873b8e5511fae2fa0c159cfdf2fa3fdb72c7a8ac9f34f4962.jpg) +Figure 2: Coded exposures from photon-cubes. (top) An exposure stack with sum-images computed using 250, 500, and 1000 bit-planes. Short exposures are noisy while long exposures exhibit motion-blur. (bottom) Spatially-varying exposure that uses a quad pattern [32] (see inset), and a video compressive frame that uses 16 random binary masks to modulate the photon-cube. Zoom-in to see details. + +![](images/348b598fcc997840b7770eb9f2714bfe7e8a29982af3ca7db9b49972480c320d.jpg) +Video compressive frame + +projection that multiplexes bit-planes with a temporal code + +$$ +\mathcal {I} _ {\text {f l u t t e r}} (\mathbf {x}) := \sum_ {t = 1} ^ {T} C _ {t} B _ {t} (\mathbf {x}), \tag {3} +$$ + +where $C_t$ is the temporal code. An example of globally-coded exposures is the flutter shutter camera [53], which uses pseudo-random binary codes for motion-deblurring. + +More general coded exposures can be obtained via spatially-varying temporal coding patterns $C_t(\mathbf{x})$ : + +$$ +\mathcal {I} _ {\text {c o d e d}} (\mathbf {x}) := \sum_ {t = 1} ^ {T} C _ {t} (\mathbf {x}) B _ {t} (\mathbf {x}). \tag {4} +$$ + +Fig. 2 (bottom) shows spatially-varying exposures that use a quad (Bayer-like) spatial pattern and random binary masks. With photon-cubes, we can perform spatially-varying coding without bulky spatial light modulators, similar to focalplane sensor-processors [46, 69]. Moreover, we can capture multiple coded exposures simultaneously, which is challenging to realize in existing sensors. In Sec. 5.1, we describe coding patterns for video compressive sensing. + +Spatial and temporal gradients form the building blocks of several computer vision algorithms [7, 11, 24, 26, 39]. Given this, another projection of interest is temporal contrast, i.e., a derivative filter preceded by a smoothing filter: + +$$ +\mathcal {I} _ {\text {c o n t r a s t}} (\mathbf {x}, t) := D _ {t} \circ G * B _ {t} (\mathbf {x}), \tag {5} +$$ + +where $D_{t}$ is the difference operator, $G$ could be exponential or Gaussian smoothing, $\circ$ denotes function composition, + +and * denotes convolution. Due to their sparse nature, gradients form the basis of bandwidth- and power-efficient event cameras [10, 14, 36, 60], which we emulate in Sec. 5.2. + +So far, we have considered projections taken only along the time axis. Next, we consider a more general class of spatio-temporal projections that lead to novel functionalities. For instance, computing a simple projection, such as the temporal sum, along arbitrary spatio-temporal directions emulates sensor motion during exposure time [33], but without moving the sensor. We achieve this by shifting bitplanes and computing their sum: + +$$ +\mathcal {I} _ {\text {s h i f t}} (\mathbf {x}) := \sum_ {t = 1} ^ {T} B _ {t} (\mathbf {x} + \mathbf {r} (t)), \tag {6} +$$ + +where $\mathbf{r}$ is a discretized 2D trajectory that determines sensor motion. Outside a software-defined framework, such projections are hard to realize without physical actuators. We describe the capabilities of motion projections in Sec. 5.3. + +In summary, the proposed photon-cube projections are simple linear and shift operators that lead to a diverse set of post-capture imaging functionalities. These projections pave the way for future 'swiss-army-knife' imaging systems that achieve multiple functionalities (e.g., event cameras, high-speed cameras, conventional cameras, HDR cameras) simultaneously with a single sensor. Finally, these projections can be computed efficiently in an online manner, which makes on-chip implementation viable (Sec. 6.3). + +At this point, we note that a key enabling factor of photon-cube projections is the extremely high temporal-sampling rate of SPADs. Indeed, the temporal sampling rate determines key aspects of sensor emulation, such as the discretization of temporal derivatives and motion trajectories. This raises a natural question: can we use conventional high-speed cameras for computing projections? + +Trade-off between frame-rate and SNR. In principle, photon-cube projections can be computed using regular (CMOS or CCD based) high-speed cameras. Unfortunately, each frame captured by a high-speed camera incurs a read-noise penalty, which increases with the camera's frame-rate [6]. In fact, the read noise levels of high-speed cameras [1] can be $10 - 30 \times$ higher than consumer cameras [28]. Coupled with the low per-frame incident flux at high frame-rates, high levels of read noise result in extremely low SNRs. In contrast, SPADs do not incur a per-frame read noise and are limited only by the fundamental photon noise. Hence, for the post-capture software-defined functionalities proposed here, it is imperative to use SPADs. + +# 5. Emulating Cameras from Photon-Cubes + +Sec. 4 presented the concept of photon-cube projections, and its potential for achieving multiple post-capture imag- + +![](images/a2900e485e032033d45c703a41e5a3da27819274f01258de20b1e17687313898.jpg) +Figure 3: Modulating masks for video compressive sensing. (left) A single VCS measurement temporally compresses a sequence of frames using binary random masks. (center) Two-bucket cameras capture an additional measurement by using the complementary mask sequence. (right) We propose using multi-bucket captures by randomly choosing an active bucket for each frame. Both two-bucket and multi-bucket captures have $100\%$ light efficiency. All masks are visualized here for $16 \times 16$ pixels. + +![](images/329176052390257c785be133c10c0d93034f35f2074c8fdefd7de845acc46489.jpg) + +![](images/1ec8b5fdaec299ac3fa57af842c242c1b4b073e83881f95d970329dd37809727.jpg) + +ing functionalities. As case studies, we now demonstrate three imaging modalities: video compressive sensing, event cameras, and motion-projection cameras. These modalities have been well-studied over several years; in particular, there exist active research communities around video compressive sensing and event cameras today. We also show new variants of these imaging systems that arise from the software-defined nature of photon-cube projections. + +# 5.1. Video Compressive Sensing + +Video compressive systems optically multiplex light with random binary masks, such as the patterns in Fig. 3 (left). As discussed in the previous section, such multiplexing can be achieved computationally using photon-cubes. + +Two-bucket cameras. One drawback of capturing coded measurements is the light loss due to blocking of incident light. To prevent loss of light, coded two-bucket cameras [69] capture an additional measurement that is modulated by the complementary mask sequence (Fig. 3 (center)). Such measurements recover higher quality frames, even after accounting for the extra readout [35, 61]. Two-bucket captures can be readily derived from photon-cubes, by implementing Eq. (4) with the additional mask sequence. + +Multi-bucket cameras. We can extend the idea of two-bucket captures to multi-bucket captures by accumulating bit-planes in one of $k$ buckets that is randomly chosen at each time instant and pixel location. Since multiplexing is performed computationally, we do not face any loss in photoreceptive area that [59, 68] hampers existing multi-bucket sensors. Multi-bucket captures can reconstruct a large number of frames by better conditioning video recovery and + +![](images/fdeb5aad951fe92184d1b86c9c064d8f9671f1bff55af3ce9963cd79b23afca2.jpg) +$\log (x)$ vs $1 - e^{-x}$ + +![](images/0f2f2a0ea4d297cf3dff34efd3a939b2da3654b3a0c873c36ada8fe8ac5c8fb4.jpg) +Spatio-temporal plot + +![](images/6d6626c534dae7f70bd1c71edb0d94ca56af365011cd170ff6542b41a3c960e0.jpg) +$\tau = 0.45, \beta = 0.95$ +5 MEvent/s +Effect of contrast-threshold, exponential decay +Figure 4: Event stream from photon-cubes. (top left) By exploiting the non-linear response curve of SPADs to encode brightness, we can avoid the underflow issues of a log-response. We visualize events generated from photon-cubes using a 3D scatter plot of polarities (top right, 14000 bit-planes), and frame accumulation of events (bottom, 1200 bit-planes). Blue and red denote positive and negative spikes respectively. The event images also show the effect of varying the contrast threshold $\tau$ and exponential decay $\beta$ -larger values yield a less noisy but sparser event stream. + +![](images/de92179dc5d733509ee24dfe131bba7dd85d3d05a3af8b7033c6657f06ca72c7.jpg) +$\tau = 0.25, \beta = 0.95$ +25 MEvent/s + +![](images/06734ec02cad1a28bf0b71f3b94d69fa076e60210fcdc765fed2634d9606b14b.jpg) +$\tau = 0.45, \beta = 0.98$ +2.5 MEvent/s + +provide extreme high-speed video imaging. Fig. 3 (right) shows the modulating masks for a four-bucket capture. + +# 5.2.Event Cameras + +Next, we describe emulation of event-cameras, which capture changes in light intensity and are conceptually similar to the temporal contrast projection introduced in Eq. (5). Physical implementations of event sensors [10, 14, 36, 60] generate a photoreceptor voltage $V(\mathbf{x}, t)$ with a logarithmic response to incident flux $\Phi(\mathbf{x}, t)$ , and output an event $(\mathbf{x}, t, p)$ when this voltage deviates sufficiently from a reference voltage $V_{\mathrm{ref}}(\mathbf{x})$ : + +$$ +\left| V (\mathbf {x}, t) - V _ {\text {r e f}} (\mathbf {x}) \right| > \tau , \tag {7} +$$ + +where $\tau$ is called the contrast-threshold and $p = \mathrm{sign}(V(\mathbf{x},t) - V_{\mathrm{ref}}(\mathbf{x}))$ encodes the polarity of the event. Once an event is generated, $V_{\mathrm{ref}}(\mathbf{x})$ is updated to $V(\mathbf{x},t)$ . Eq. (7), for a smoothly-varying flux intensity, thresholds a function of the temporal gradient, i.e., $\partial_t\log (\Phi (\mathbf{x},t))$ + +From bit-planes to event streams. To produce events from SPAD frames, we compute an exponential moving average (EMA) of the bit-planes, as $\mu_t(\mathbf{x}) = (1 - \beta)B_t(\mathbf{x}) + \beta \mu_{t - 1}(\mathbf{x})$ where $\mu_t(\mathbf{x})$ is the EMA, $\beta$ is the smoothing + +![](images/e6656c4cc37d7fcdd415afb7de28cc059369c90e3c76c3dc0a9ae037f0872d8c.jpg) + +![](images/e0a08c526f44f3d877f4c3f47dae2342229790c25a9af49896fbf19b92bc213d.jpg) + +![](images/4f07bcadec453ebbbb6d8c55fd2d414878ed1f3e92d89c7ac201bb81c330f8d0.jpg) + +![](images/fc72fb42f026bf3e2e660bf720973e961cff5ae13bf1eb0a86b5a434c3e1574a.jpg) + +![](images/c1069f1838814529b3a092bebf6aa903e55fbb56a6bf29b0c91199a5ba9c0ec3.jpg) + +![](images/7e017e1f64b6f1df9f7f6122d4c06c01395bd01338fba431a34cfe05f8789365.jpg) +Parabolic trajectory + +![](images/f5fdd82605fb6883608f0bde223d092297f6e17b647864f20f3c0f02bc4d5689.jpg) +Figure 6: Motion stack. Computing multiple linear projections with different trajectories can produce a stack of images where objects with matching velocity are sharp. Here, we show a traffic scene involving four cars that have four different velocities. By suitably altering the slope of the linear trajectory, we can produce images where only one of the cars appear sharp at a time. We indicate the slope of the trajectories chosen and the objects that are "in-focus". + +![](images/6d9b5bad09bb7c41561d2b68c602b314790e14ef4ff4e33bf35d06198d26e7e7.jpg) +Parabolic deblurred +Figure 5: Motion projections. (top) Integrating along a linear trajectory in the photon-cube changes the apparent image-space velocity of scene objects. Details are seen for (top left) the case when static, and (top right) the metallic tape when the sensor translates along the $x$ -axis. (bottom) A parabolic integration trajectory results in a motion-invariant image, resulting in similar blur kernels for all objects. (bottom right) Deblurring with the resultant shift-invariant point spread function (shown in inset) produces a sharp image. + +factor, and $B_{t}$ is a bit-plane. We generate an event when $\mu_t(\mathbf{x})$ deviates from $\mu_{\mathrm{ref}}(\mathbf{x})$ by at least $\tau$ : + +$$ +\left| h \left(\mu_ {t} (\mathbf {x})\right) - h \left(\mu_ {\text {r e f}} (\mathbf {x})\right) \right| > \tau , \tag {8} +$$ + +where $h$ is a scalar function applied to the EMA. We can see that Eq. (8) thresholds temporal contrast, by observing the role played by the EMA and the difference operator. + +Setting $h$ to be the logarithm of the flux MLE mimics Eq. (7). However, since the log-scale is used to prevent sensor saturation, a simpler alternative is to use the non-saturating response curve of SPAD pixels $(h(x) = x)$ . The response curve takes the form of $1 - \exp (-\alpha \Phi (\mathbf{x},t))$ , where $\alpha$ is a flux-independent constant. As a major advantage, this response curve avoids the underflow issues of the log function that can occur in low-light scenarios [62]. + +The SPAD's frame-rate determines the time-stamp resolution of emulated events. In Fig. 4, we show the events generated from a photon-cube acquired at a frame-rate of $96.8\mathrm{kHz}$ resulting in a time-stamp resolution of $\sim 10~\mu \mathrm{s}$ that is comparable to those of existing event cameras. + +How do SPAD-events differ from the output of a regular event camera? The main difference is the expression of temporal contrast, given by $\partial_t h$ , is now $-\partial_t \exp(-\alpha \Phi(\mathbf{x}, t))$ , instead of $\partial_t \log(\Phi(\mathbf{x}, t))$ . This dif + +![](images/1709c7778e3cf88a2a004695e25143a4ad7fa0a2fd2d5f915e20ffbaa82ef9bc.jpg) + +![](images/1c41b3c328fb8d6d2a989938fa554a706049d4f82ec3037b4759f94902af11c0.jpg) +$160\hat{x} +447\hat{y}$ pixels/sec + +![](images/d4a1b9cda9e01fe1ab01df568296f6ea507966676fdf96c631718e543222b385.jpg) +$100\hat{x} + 282\hat{y}$ pixels/sec + +![](images/f60191071b30b5e4fa35b5ee6cd7761d55a5f580686286e62f6ac2af1c7b5ce5.jpg) +$420\hat{x} + 1177\hat{y}$ pixels/sec + +ference poses no compatibility issues for a large class of event-vision algorithms that utilize a grid of events [54, 57, 66] or brightness changes [18]. We show examples of downstream applications using SPAD-events in Supplementary Note 2. Finally, SPAD-events can be easily augmented with spatially- and temporally-aligned intensity information—a synergistic combination that has been exploited by several recent event-vision works [18, 25, 74]. + +# 5.3. Motion Projections + +Having described the emulation of cameras that capture coded exposures and temporal contrasts, we now shift our attention to cameras that emulate sensor motion during exposure, viz. motion cameras. We describe two useful trajectories when emulating motion cameras using Eq. (6). + +Linear trajectory. The simplest sensor trajectory involves linear motion, where $\mathbf{r}(t) = (bt + c)\hat{\mathbf{p}}$ for some constants $b, c \in \mathbb{R}$ and unit vector $\hat{\mathbf{p}}$ . As Fig. 5 (top row) shows, this can change the scene's frame of reference: making moving objects appear stationary and vice-versa. + +Motion-invariant parabolic projection. If motion is along $\hat{\mathbf{p}}$ , parabolic integration produces a motion-invariant image [33]—all objects, irrespective of their velocity, are blurred by the same point spread function (PSF), up to a linear shift. Thus, a deblurred parabolic capture produces a sharp image of all velocity groups (Fig. 5 (bottom row)). The parabolic trajectory is given by $\mathbf{r}(t) = (at^2 + bt + c)\hat{\mathbf{p}}$ . We choose $a$ based on the maximum object velocity and $b, c$ + +![](images/f3e5ed354caccbee3543c8c2950ba4e8c8629eadd0e5ca9b85c43134055c8fa5.jpg) +Figure 7: High-speed videography at 2000 FPS of a tennis ball dropped into a bowl of water, from a $25\mathrm{Hz}$ readout. The conventional capture provides a visualization of the scene dynamics. It is challenging to reconstruct a large number of frames from a single compressive snapshot. Multi-bucket captures recover frames with significantly greater detail, such as the crown of water surrounding the ball. We include more sequences (e.g., a bursting balloon) in the supplementary material. + +![](images/b47c6bfabc84357280f9d81926085a1217888186ea2f7f273cda943887699918.jpg) +Figure 8: Deblurring of traffic scenes using motion projections. Linear projections can recover details of moving objects if their velocity is known. When only the motion direction is known (e.g., road's orientation), a sharp image can be obtained by either deblurring a parabolic projection or by blending multiple randomly-sampled linear projections. We quantitatively compare against the compute- and bandwidth-expensive Quanta Burst Photography [41], based on PSNR and LPIPS [76]. + +so the parabola's vertex lies at $T / 2$ . We readily obtain the PSF by applying the parabolic integration to a delta input. Upon deconvolution using the PSF, a parabolic projection provides the optimal SNR for a blur-free image from single capture when only the direction of velocity is known. + +Ensembling linear projections. Finally, we leverage the flexibility of photon-cubes to compute multiple linear projections, as seen in Fig. 6. This produces a stack of images where one velocity group is motion-blur free at a time—or a 'motion stack', analogous to a focal stack. This novel construct can be used to compensate motion by blending stack images using cues such as blur orientation or optical flow. + +# 6. Hardware and Experimental Results + +We design a range of experiments to demonstrate the versatility of photon-cube projections: both when computa + +tions occur after readout (Secs. 6.1 and 6.2), and when they are performed near-sensor on-chip (Sec. 6.3). All photon-cubes were acquired using the SwissSPAD2 array [67], operated using one of two sub-arrays, each having $512 \times 256$ pixels, and at a frame-rate of $96.8\mathrm{kHz}$ . For the on-chip experiments, we use the UltraPhase compute architecture to interface with photon-cubes acquired by the SwissSPAD2. + +# 6.1. SoDaCam Capabilities + +High-speed compressive imaging. We reconstruct 80 frames from compressive snapshots that are emulated at $25\mathrm{Hz}$ , resulting in a 2000 FPS video. We decode compressive snapshots using a plug-and-play (PnP) approach, PnP-FastDVDNet [73]. As Fig. 7 shows, it is challenging to recover a large number of frames from a single compressive measurement. Using the proposed multi-bucket scheme significantly improves the quality of video reconstruction. + +![](images/59f8080ba282867c4d22838a60bcb7bf1ec710645f571ad142bc47a111e1f536.jpg) +Sufficient Light 158 lux, 1.03 PPP + +![](images/88404c8016c9e7b8df7a345335bc5cddce2a9ea9cda51452cafd8b65fcf20dc6.jpg) +Low light 0.9 lux, 0.1 PPP + +![](images/2e3dd248446e1dcf04390d0feac2a6f77fedf801d07f7225b49927d03b5adc91.jpg) +Very low light 0.3 lux, 0.03 PPP + +![](images/3d6d717eafad9874614643283bd060c59b7b2269c31598f334f5ac790798c780.jpg) +SPAD-emulated events + +![](images/f5aed45fe02cb7294012691749a1d12badfcc30b39ba6065ce34022d30d3b17c.jpg) +DSLR captures (30 Hz) + +![](images/7aaf9f1ac84da73d2749050a24ba8720f6f533a418b11100f73bd01bd4526604.jpg) + +![](images/51a862585f198d34166b38a0ca7ba90c408291b25d248fa3ada9aa1b48592835.jpg) +Prophesee EVK4 events + +![](images/278f1ad503ec66b55ab202dfac536d76fedc1977f137bc3f23a353c99bdc5570.jpg) +Figure 9: Comparison to a state-of-the-art event camera. SPAD-events can capture temporal gradients even when the light-level is reduced by $500\times$ , by benefiting from their single-photon sensitivity and bounded brightness response curve. In contrast, low-light induces blur and deteriorates the Prophesee's event stream. As a measure of the light-level, we report the PPP (photons per pixel) averaged across bit-planes and a light meter's reading at the sensor location. + +![](images/7f31fb95ba25496365b8b6c3062f4c2c2648bb3887b2ca3b0f8f9987def73c73.jpg) +Figure 10: Comparison against conventional high-speed acquisition at $4000\mathrm{Hz}$ . (top) SPAD projections recover a $16\times$ compressive video and an event image of a spinning roulette wheel. (middle) Read-noise corrupts the incident flux in the Infinicam high-speed camera, removing details in frames which are compressed on-the-fly. (bottom) Although using a larger aperture to admit more light recovers some detail, noise and compression artifacts still persist. + +While multi-bucket captures require more bandwidth, this can be partially amortized by coding only dynamic regions, which we show in Supplementary Note 1. + +Motion projections on a traffic scene. Fig. 8 shows two traffic scenes captured using a $50~\mathrm{mm}$ focal length lens and at $30\mathrm{Hz}$ emulation. When object velocity is known, a linear projection can make moving objects appear stationary. If only the velocity direction is known (e.g., road's orientation in Fig. 8), a parabolic projection provides a sharp reconstruction of all objects. We deblur parabolic captures using PnP-DnCNN [75]. We offer an improvement by randomly sampling 8 linear projections along the velocity direction and blending them using the optical flow predicted by RAFT [65] between two short-exposures. + +Low-light event imaging. Fig. 9 compares event-image visualizations of SPAD and that of a state-of-the-art commercial event sensor (Prophesee EVK4), across various + +![](images/85c7f483ef5ff3d49b0a9ca1ef3518a6f86d20b34b7f4a18599f97c31eb0fbfa.jpg) +Compressive frame + +![](images/68caba06bceae84361f249d39cf5f17d9ffc27556e84ddeb3858dbe92f168567.jpg) +Recovered frame + +![](images/859ef4cd2e6e9b055864c9f0db9b5ac5bfbab380bdf1043f5349407b1f08122c.jpg) +Event image +SwissSPAD2 projections + +![](images/a38bfe0d98ac5a3b148c42422f35949a0dec357730c02c225786bf3bfc442759.jpg) + +![](images/056222d39ccb53124b482ec6010df2cf88ee1633d237d1d959147f7cf6642679.jpg) +Infinicam projections + +![](images/e925881cb7d9e2b08ded24c1094f58849cb7f14efae3f5031fea13ee6d6d9954.jpg) + +![](images/52001fca6b9e05c38e133049fb6968a1865da47d1dd78376857ab2acdcc46ade.jpg) +Infinicam projections (16× more light) + +![](images/907f2d1c8ff784ad88f54f3683feb65f3d6169d52a504b789e6879a94aecae98.jpg) + +![](images/8b5c82a105fcdde8453fbb76374f86659d472e4651b4ef7f960369e67b1d5c1e.jpg) + +light levels, with an accumulation period of $33\mathrm{ms}$ . For a fair comparison, we bin the Prophesee's events in blocks of $2\times 2$ pixels and use a smaller aperture to account for the lower fill-factor of the SPAD. We tuned event-generation parameters (contrast threshold, integrator decay rate) of both cameras at each light level. Low light induces blur and deteriorates the Prophesee's event stream. In contrast, SPAD-events continue to capture temporal gradients, due to the SPAD's low-light capabilities and its brightness-encoding response curve. We include an ablative study of brightness-encoding functions in Supplementary Note 2. + +Our observations are in concurrence with recent works that examine the low-light performance of event cameras [19, 27], and show that SPAD-events can provide neuromorphic vision in these challenging-SNR scenarios. + +# 6.2. Comparison to High-Speed Cameras + +Recall, as previously discussed in Sec. 4, that read-noise limits the per-frame SNR of high-speed cameras. To demonstrate this limitation, we compute projections using the $4\mathrm{kHz}$ acquisition of the Photon Infinicam, a conventional high-speed camera, at a resolution of $1246\times 240$ pixels. We operate the SwissSPAD2 and the Infinicam at ambient light conditions using the same lens specifications. + +![](images/eb1db41027ec422ffa4fcf6c16153765de028fa94fa48c064e8b750ec88e2fce.jpg) +UltraPhase board + +![](images/86f53554ad579b4c3582097706391466bfcd1ae17199cd42f715cb10148a4149.jpg) +Power and bandwidth measurement +Figure 11: Power and bandwidth requirements when computing photon-cube projections on UltraPhase [5] (left), a recent compute architecture designed for single-photon imaging, at $40\mathrm{Hz}$ readout. (right) Our projections act as a compression scheme for photon-cubes, resulting in dramatically reduced sensor-readout and power consumption. + +As Fig. 10 shows, read noise corrupts the incident signal in Infinicam and makes it impossible to derive any useful projections. The read noise could be averaged out to some extent if the Infinicam did not perform compression-on-the-fly, but compression is central to the camera's working and enables readout over USB. Using a larger aperture to admit more light improves the quality of computed projections, but the video reconstruction and event image remain considerably worse than corresponding outputs of the SPAD. + +# 6.3. Bandwidth and Power Implications + +While Sec. 6.1 has demonstrated the capabilities of photon-cube projections, we now show that our projections can also be obtained in a bandwidth-efficient manner via near-sensor computations. We implement photon-cube projections on UltraPhase (Fig. 11 (left)), a novel compute architecture designed for single-photon imaging. UltraPhase consists of $3 \times 6$ processing cores, each of which interfaces with $4 \times 4$ pixels, and can be 3D stacked beneath a SPAD array. We include visualizations and programming details of a few example projections in Supplementary Note 5. + +We measure the readout and power consumption of UltraPhase when computing projections on 2500 bit-planes of the falling die sequence (Fig. 1). The projections include: VCS with 16 random binary masks, an event camera, a linear projection and a combination of the three. We output projections at 12-bit depth and calculate metrics based on the clock cycles required for both compute and readout. As seen in Fig. 11 (right), computing projections on-chip dramatically reduces sensor-readout and power consumption as compared to reading out the photon-cube. Finally, similar to existing event cameras, SPAD-events have a resource footprint that reflects the underlying scene dynamics. + +In summary, our on-chip experiments show that performing computations near-sensor can increase the viability of single-photon imaging in resource constrained settings. + +# 7. Discussion and Future Outlook + +SoDaCam provides a realization of reinterpretable software-defined cameras [2, 3, 20, 34, 51] at the fine temporal resolution of SPAD-acquired photon-cubes. The proposed computations, or photon-cube projections, can match and in some cases, surpass the capabilities of existing imaging systems. The software-defined nature of photon-cube projections provides functionalities that may be difficult to achieve in conventional sensors. These projections can reduce the readout and power-consumption of SPAD arrays and potentially spur widespread adoption of single-photon imaging in the consumer domain. Finally, future chip-to-chip communication standards may also make it feasible to compute projections on a camera image signal processor. + +Adding color to SoDaCam. One way to add color is by overlaying color filter arrays (CFAs) and perform demosaicing on the computed photon-cube projection: depending on the projection, demosaicing could be relatively simple or more complex. As a reference, Bayer CFAs have been considered in the context of both video compressive sensing [73] and event cameras [64]. Incorporating CFAs with motion projections requires careful considerations, e.g., avoiding integrating across pixel locations of differing color. + +Future outlook on SPAD characteristics. A key SPAD characteristic that determines several properties of emulated cameras is the frame rate. While no fundamental limitations prevent SPADs from being operated at the frame-rates utilized in this work ( $\sim 100\mathrm{kHz}$ ), sensor readout and power constraints can preclude high speeds, especially in high-resolution SPAD arrays. Photon-cube projections can enable future large-format SPADs to preserve high-speed information with modest resource requirements. + +A platform for comparing cameras. Comparing imaging modalities can be quite challenging since hardware realizations of sensors can differ in numerous aspects, such as their quantum efficiency, fill factor, pixel pitch and array resolution. By emulating their imaging models, SoDaCam can serve as a platform for hardware-agnostic comparisons; for instance, determining operating conditions where one imaging modality is advantageous over another. + +A Cambrian explosion of new cameras. Besides comparing cameras, by virtue of being software-defined, SoDa-Cam can also make it significantly easier to prototype and deploy new unconventional imaging models, and even facilitate sensor-in-the-loop optimization [44, 47, 63] by tailoring photon-cube projections for downstream computer-vision tasks. This is an exciting future line of research. + +# References + +[1] Phantom-v2640. https://www.phantomhighspeed.com/products/cameraul ultrahigh4mpx/v2640. Accessed: 2023-01-28. 4 +[2] E. H. Adelson, J. R. Bergen, et al. The plenoptic function and the elements of early vision. Computational models of visual processing, 1(2):3-20, 1991. 1, 9 +[3] A. Agrawal, A. Veeraraghavan, and R. Raskar. Reinterpretable imager: Towards variable post-capture space, angle and time resolution in photography. In Computer Graphics Forum, volume 29, pages 763-772. Wiley Online Library, 2010. 3, 9 +[4] I. M. Antolovic, S. Burri, C. Bruschini, R. Hoebe, and E. Charbon. Nonuniformity analysis of a 65-kpixel CMOS SPAD imager. IEEE Transactions on Electron Devices, 63 (1):57-64, 2016. doi: 10.1109/TED.2015.2458295. 3 +[5] A. Ardelean. Computational Imaging SPAD Cameras. PhD thesis, École polytechnique fédérale de Lausanne, 2023. 2, 9 +[6] A. Boukhayma, A. Peizerat, and C. Enz. A sub-0.5 electron read noise VGA image sensor in a standard CMOS process. IEEE Journal of Solid-State Circuits, 2016. 4 +[7] J. Canny. A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-8(6):679-698, 1986. doi: 10.1109/TPAMI.1986.4767851. 4 +[8] S. J. Carey, A. Lopich, D. R. Barr, B. Wang, and P. Dudek. A 100,000 fps vision sensor with embedded 535GOPS/W $256 \times 256$ SIMD processor array. In 2013 Symposium on VLSI Circuits, pages C182-C183, 2013. 3 +[9] J. Chen, S. J. Carey, and P. Dudek. Feature extraction using a portable vision system. In IEEE/RSJ Int. Conf. Intell. Robots Syst., Workshop Vis.-based Agile Auton. Navigation UAVs, volume 2, 2017. 3 +[10] S. Chen and M. Guo. Live demonstration: Celex-v: A 1m pixel multi-mode event-based sensor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2019. 4, 5 +[11] N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), volume 1, pages 886-893. IEEE, 2005. 4 +[12] P. E. Debevec and J. Malik. Recovering high dynamic range radiance maps from photographs. In Proceedings of the 24th Annual Conference on Computer Graphics and Interactive Techniques, page 369-378. ACM Press/Addison-Wesley Publishing Co., 1997. doi: 10.1145/258734.258884. URL https://doi.org/10.1145/258734.258884. 1, 3 +[13] F. M. Della Rocca, H. Mai, S. W. Hutchings, T. Al Abbas, K. Buckbee, A. Tsiamis, P. Lomax, I. Gyongy, N. A. Dutton, and R. K. Henderson. A $128 \times 128$ SPAD motion-triggered time-of-flight image sensor with in-pixel histogram and column-parallel vision processor. IEEE Journal of Solid-State Circuits, 55(7):1762-1775, 2020. 3 +[14] T. Finateu, A. Niwa, D. Matolin, K. Tsuchimoto, A. Mascheroni, E. Reynaud, P. Mostafalu, F. Brady, L. Chotard, F. LeGoff, H. Takahashi, H. Wakabayashi, + +Y. Oike, and C. Posch. 5.10 a $1280 \times 720$ back-illuminated stacked temporal contrast event-based vision sensor with $4.86\mu m$ pixels, 1.066GEPS readout, programmable event-rate controller and compressive data-formating pipeline. In 2020 IEEE International Solid-State Circuits Conference - (ISSCC), pages 112-114, 2020. doi: 10.1109/ISSCC19947.2020.9063149. 4, 5 +[15] E. R. Fossum. What to do with sub-diffraction-limit (SDL) pixels?—a proposal for a gigapixel digital film sensor (DFS). In IEEE Workshop on Charge-Coupled Devices and Advanced Image Sensors, pages 214–217, 2005. 3 +[16] E. R. Fossum. The quanta image sensor (QIS): Concepts and challenges. In Imaging and Applied Optics, page JTuE1. Optica Publishing Group, 2011. doi: 10.1364/COSI.2011. JTuE1. URL http://opg.optica.org/abstract.cfm?URI=COSI-2011-JTuE1.1 +[17] E. R. Fossum, J. Ma, S. Masoodian, L. Anzagira, and R. Zizza. The quanta image sensor: Every photon counts. Sensors, 16(8):1260, 2016. 3 +[18] D. Gehrig, H. Rebecq, G. Gallego, and D. Scaramuzza. Eklt: Asynchronous photometric feature tracking using events and frames. International Journal of Computer Vision, 128(3): 601-618, 2020. 6 +[19] R. Graca and T. Delbruck. Unraveling the paradox of intensity-dependent DVS pixel noise. arXiv preprint arXiv:2109.08640, 2021. 8 +[20] M. Gupta, A. Agrawal, A. Veeraraghavan, and S. G. Narasimhan. Flexible voxels for motion-aware videography. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part I 11, pages 100-114. Springer, 2010. 3, 9 +[21] F. Gutierrez-Barragan, A. Ingle, T. Seets, M. Gupta, and A. Velten. Compressive single-photon 3D cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17854-17864, June 2022. 3 +[22] I. Gyongy, N. A. Dutton, and R. K. Henderson. Single-photon tracking for high-speed vision. Sensors, 18(2):323, 2018. 3 +[23] I. Gyongy, S. W. Hutchings, A. Halimi, M. Tyler, S. Chan, F. Zhu, S. McLaughlin, R. K. Henderson, and J. Leach. High-speed 3D sensing via hybrid-mode imaging and guided upsampling. Optica, 7(10):1253-1260, 2020. 3 +[24] C. Harris, M. Stephens, et al. A combined corner and edge detector. In *Alvey vision conference*, volume 15, pages 10-5244. CiteSeer, 1988. 4 +[25] J. Hidalgo-Carrió, G. Gallego, and D. Scaramuzza. Event-aided direct sparse odometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5781-5790, June 2022. 6 +[26] B. K. Horn and B. G. Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185-203, 1981. 4 +[27] Y. Hu, S.-C. Liu, and T. Delbruck. v2e: From video frames to realistic DVS events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1312-1321, June 2021. 8 +[28] J. Igual. Photographic noise performance measures based on + +raw files analysis of consumer cameras. Electronics, 8(11): 1284, 2019. 4 +[29] A. Ingle, A. Velten, and M. Gupta. High Flux Passive Imaging With Single-Photon Sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 3 +[30] A. Ingle, T. Seets, M. Buttafava, S. Gupta, A. Tosi, M. Gupta, and A. Velten. Passive inter-photon imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021. 3 +[31] K. Iwabuchi, Y. Kameda, and T. Hamamoto. Image quality improvements based on motion-based deblurring for single-photon imaging. IEEE Access, 9:30080-30094, 2021. doi: 10.1109/ACCESS.2021.3059293. 3 +[32] Y. Jiang, I. Choi, J. Jiang, and J. Gu. HDR video reconstruction with tri-exposure quad-bayer sensors. arXiv preprint arXiv:2103.10982, 2021. 4 +[33] A. Levin, P. Sand, T. S. Cho, F. Durand, and W. T. Freeman. Motion-invariant photography. ACM Transactions on Graphics (TOG), 27(3):1-9, 2008. 4, 6 +[34] M. Levoy and P. Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 31-42, 1996. 1, 9 +[35] Y. Li, M. Qi, R. Gulve, M. Wei, R. Genov, K. N. Kutulakos, and W. Heidrich. End-to-end video compressive sensing using anderson-accelerated unrolled networks. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12, 2020. doi: 10.1109/ICCP48838.2020.9105237. 5 +[36] P. Lichtsteiner. 64x64 event-driven logarithmic temporal derivative silicon retina. In Program 2003 IEEE Workshop on CCD and AIS, 2003. 4, 5 +[37] Y. Liu, F. Gutierrez-Barragan, A. Ingle, M. Gupta, and A. Velten. Single-photon camera guided extreme dynamic range imaging. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1575-1585, January 2022. 3 +[38] P. Llull, X. Liao, X. Yuan, J. Yang, D. Kittle, L. Carin, G. Sapiro, and D. J. Brady. Coded aperture compressive temporal imaging. Opt. Express, 21(9):10526-10545, May 2013. doi: 10.1364/OE.21.010526. URL https://opg.optica.org/oe/abstract.cfm?URI=oe-21-9-10526.1 +[39] B. D. Lucas and T. Kanade. An iterative image registration technique with an application to stereo vision. In *IJCAI'81: 7th international joint conference on Artificial intelligence*, volume 2, pages 674–679, 1981. 4 +[40] J. Ma, S. Masoodian, D. A. Starkey, and E. R. Fossum. Photon-number-resolving megapixel image sensor at room temperature without avalanche gain. Optica, 4(12): 1474-1481, Dec 2017. doi: 10.1364/OPTICA.4.001474. URL http://www.osapublishing.org/optica/abstract.cfm?URI=optica-4-12-1474.3 +[41] S. Ma, S. Gupta, A. C. Ulku, C. Bruschini, E. Charbon, and M. Gupta. Quanta burst photography. ACM Transactions on Graphics, 39(4):1-16, July 2020. ISSN 0730-0301, 1557-7368. 3, 7 +[42] S. Ma, P. Mos, E. Charbon, and M. Gupta. Burst vi + +sion using single-photon cameras. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5375-5385, January 2023. 3 +[43] S. Mann and R. Picard. Beingundigital'with digital cameras. MIT Media Lab Perceptual, 1:2, 1994. 3 +[44] J. N. Martel, L. K. Mueller, S. J. Carey, P. Dudek, and G. Wetzstein. Neural sensors: Learning pixel exposures for HDR imaging and video compressive sensing with programmable sensors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(7):1642-1653, 2020. 9 +[45] J. N. P. Martel, L. K. Müller, S. J. Carey, and P. Dudek. Parallel HDR tone mapping and auto-focus on a cellular processor array vision chip. In 2016 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1430-1433, 2016. doi: 10.1109/ISCAS.2016.7527519. 3 +[46] J. N. P. Martel, L. K. Müller, S. J. Carey, and P. Dudek. High-speed depth from focus on a programmable vision chip using a focus tunable lens. In 2017 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-4, 2017. doi: 10.1109/ISCAS.2017.8050548.3, 4 +[47] C. A. Metzler, H. Ikoma, Y. Peng, and G. Wetzstein. Deep optics for single-shot high-dynamic-range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 9 +[48] K. Morimoto, A. Ardelean, M.-L. Wu, A. C. Ulku, I. M. Antolovic, C. Bruschini, and E. Charbon. Megapixel time-gated SPAD image sensor for 2D and 3D imaging applications. Optica, 7(4):346-354, Apr. 2020. 1, 3 +[49] K. Morimoto, J. Iwata, M. Shinohara, H. Sekine, A. Abdelghafar, H. Tsuchiya, Y. Kuroda, K. Tojima, W. Endo, Y. Maehashi, Y. Ota, T. Sasago, S. Maekawa, S. Hikosaka, T. Kanou, A. Kato, T. Tezuka, S. Yoshizaki, T. Ogawa, K. Uehira, A. Ehara, F. Inui, Y. Matsuno, K. Sakurai, and T. Ichikawa. 3.2 megapixel 3D-stacked charge focusing SPAD for low-light imaging and depth sensing. In 2021 IEEE International Electron Devices Meeting (IEDM), pages 20.2.1-20.2.4, 2021. doi: 10.1109/IEDM19574.2021.9720605.1, 3 +[50] S. Namiki, S. Sato, Y. Kameda, and T. Hamamoto. Imaging method using multi-threshold pattern for photon detection of quanta image sensor. In International Workshop on Advanced Imaging Technology (IWAIT) 2022, volume 12177, page 1217702. SPIE, 2022. 3 +[51] S. K. Nayar, V. Branzoi, and T. E. Boult. Programmable imaging: Towards a flexible camera. International Journal of Computer Vision, 70:7-22, 2006. 3, 9 +[52] C. Posch, D. Matolin, and R. Wohlgenannt. A QVGA 143 dB Dynamic Range Frame-Free PWM Image Sensor With Lossless Pixel-Level Video Compression and Time-Domain CDS. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2011. doi: 10.1109/JSSC.2010.2085952. 1 +[53] R. Raskar, A. Agrawal, and J. Tumblin. Coded exposure photography: motion deblurring using fluttered shutter. In Acm Siggraph 2006 Papers, pages 795-804. 2006. 4 +[54] H. Rebecq, R. Ranftl, V. Koltun, and D. Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 6 + +[55] D. Reddy, A. Veeraraghavan, and R. Chellappa. P2C2: Programmable pixel compressive camera for high speed imaging. In CVPR 2011, pages 329-336, 2011. doi: 10.1109/CVPR.2011.5995542. 1, 3 +[56] X. Ren, P. W. Connolly, A. Halimi, Y. Altmann, S. McLaughlin, I. Gyongy, R. K. Henderson, and G. S. Buller. High-resolution depth profiling using a range-gated CMOS SPAD quanta image sensor. Optics express, 26(5):5541-5557, 2018. 3 +[57] C. Scheerlinck, H. Rebecq, D. Gehrig, N. Barnes, R. Mahony, and D. Scaramuzza. Fast image reconstruction with an event camera. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), March 2020. 6 +[58] T. Seets, A. Ingle, M. Laurenzis, and A. Velten. Motion adaptive deblurring with single-photon cameras. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1945-1954, January 2021. 3 +[59] M.-W. Seo, Y. Shirakawa, Y. Masuda, Y. Kawata, K. Kagawa, K. Yasutomi, and S. Kawahito. 4.3 a programmable sub-nanosecond time-gated 4-tap lock-in pixel cmos image sensor for real-time fluorescence lifetime imaging microscopy. In 2017 IEEE International Solid-State Circuits Conference (ISSCC), pages 70-71, 2017. doi: 10.1109/ISSCC.2017.7870265. 5 +[60] T. Serrano-Gotarredona and B. Linares-Barranco. A $128 \times 128$ $1.5\%$ contrast sensitivity $0.9\%$ FPN $3\mu s$ latency $4\mathrm{mW}$ asynchronous frame-free dynamic vision sensor using transimpedance preamplifiers. IEEE Journal of Solid-State Circuits, 48(3):827-838, 2013. 1, 4, 5 +[61] P. Shedligeri, A. S., and K. Mitra. A unified framework for compressive video recovery from coded exposure techniques. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1600-1609, January 2021. 5 +[62] C. Shi, N. Song, W. Li, Y. Li, B. Wei, H. Liu, and J. Jin. A review of event-based indoor positioning and navigation. 2022. 6 +[63] V. Sitzmann, S. Diamond, Y. Peng, X. Dun, S. Boyd, W. Heidrich, F. Heide, and G. Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and super-resolution imaging. ACM Transactions on Graphics (TOG), 37(4):114, 2018. 9 +[64] G. Taverni, D. Paul Moeys, C. Li, C. Cavaco, V. Motsnyi, D. San Segundo Bello, and T. Delbruck. Front and back illuminated dynamic and active pixel vision sensors comparison. IEEE Transactions on Circuits and Systems II: Express Briefs, 65(5):677-681, 2018. doi: 10.1109/TCSII.2018.2824899. 9 +[65] Z. Teed and J. Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European Conference on Computer Vision, 2020. 8 +[66] S. Tulyakov, D. Gehrig, S. Georgoulis, J. Erbach, M. Gehrig, Y. Li, and D. Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16155-16164, June 2021. 6 + +[67] A. C. Ulku, C. Bruschini, I. M. Antolovic, Y. Kuo, R. Ankri, S. Weiss, X. Michalet, and E. Charbon. A $512 \times 512$ SPAD Image Sensor With Integrated Gating for Widefield FLIM. IEEE Journal of Selected Topics in Quantum Electronics, 25 (1):1-12, Jan. 2019. ISSN 1077-260X, 1558-4542. doi: 10. 1109/JSTQE.2018.2867439. 3, 7 +[68] G. Wan, X. Li, G. Agranov, M. Levoy, and M. Horowitz. CMOS image sensors with multi-bucket pixels for computational photography. IEEE Journal of Solid-State Circuits, 47 (4):1031-1042, 2012. doi: 10.1109/JSSC.2012.2185189. 5 +[69] M. Wei, N. Sarhangnejad, Z. Xia, N. Gusev, N. Katic, R. Genov, and K. N. Kutulakos. Coded two-bucket cameras for computer vision. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 3, 4, 5 +[70] M. White, S. Ghajari, T. Zhang, A. Dave, A. Veeraraghavan, and A. Molnar. A differential SPAD array architecture in $0.18\mu \mathrm{m}$ CMOS for HDR imaging. In 2022 IEEE International Symposium on Circuits and Systems (ISCAS), pages 292-296, 2022. doi: 10.1109/ISCAS48785.2022.9937558.3 +[71] T. Yamazaki, H. Katayama, S. Uehara, A. Nose, M. Kobayashi, S. Shida, M. Odahara, K. Takamiya, Y. Hisamatsu, S. Matsumoto, L. Miyashita, Y. Watanabe, T. Izawa, Y. Muramatsu, and M. Ishikawa. 4.9 a 1ms high-speed vision chip with 3D-stacked 140GOPS column-parallel PEs for spatio-temporal image processing. In 2017 IEEE International Solid-State Circuits Conference (ISSCC), pages 82–83, 2017. doi: 10.1109/ISSCC.2017.7870271. 3 +[72] F. Yang, Y. M. Lu, L. Sbaiz, and M. Vetterli. Bits from photons: Oversampled image acquisition using binary poisson statistics. IEEE Transactions on Image Processing, 21(4): 1421-1436, 2012. doi: 10.1109/TIP.2011.2179306. 3 +[73] X. Yuan, Y. Liu, J. Suo, F. Durand, and Q. Dai. Plug-and-play algorithms for video snapshot compressive imaging. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):7093-7111, 2022. doi: 10.1109/TPAMI.2021.3099035.7, 9 +[74] J. Zhang, X. Yang, Y. Fu, X. Wei, B. Yin, and B. Dong. Object tracking by jointly exploiting frame and event domain. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13043-13052, 2021. 6 +[75] K. Zhang, Y. Li, W. Zuo, L. Zhang, L. Van Gool, and R. Timofte. Plug-and-play image restoration with deep denoiser prior. arXiv preprint, 2020. 8 +[76] R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 7 +[77] T. Zhang, M. J. White, A. Dave, S. Ghajari, A. Raghuram, A. C. Molnar, and A. Veeraraghavan. First arrival differential LiDAR. In 2022 IEEE International Conference on Computational Photography (ICCP), pages 1-12, 2022. doi: 10.1109/ICCP54855.2022.9887683. 3 +[78] W. Zhang and L. Lin. Light field flow estimation based on occlusion detection. Journal of Computer and Communications, 5(3):1-9, 2017. 1 \ No newline at end of file diff --git a/2023/SoDaCam_ Software-defined Cameras via Single-Photon Imaging/images.zip b/2023/SoDaCam_ Software-defined Cameras via Single-Photon Imaging/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f81b493129f7df7b6fb2fc02002a54bd1f58199c --- /dev/null +++ b/2023/SoDaCam_ Software-defined Cameras via Single-Photon Imaging/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db3a21639ad9cf6d28f2259233c08b25b21e5ebc59a1fbf1aaf40f333dc8c4e1 +size 819273 diff --git a/2023/SoDaCam_ Software-defined Cameras via Single-Photon Imaging/layout.json b/2023/SoDaCam_ Software-defined Cameras via Single-Photon Imaging/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9725e45d56d4a48886628295ccb21d7d4b9f65c0 --- /dev/null +++ b/2023/SoDaCam_ Software-defined Cameras via Single-Photon Imaging/layout.json @@ -0,0 +1,13175 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 103, + 501, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 103, + 501, + 121 + ], + "spans": [ + { + "bbox": [ + 94, + 103, + 501, + 121 + ], + "type": "text", + "content": "SoDaCam: Software-defined Cameras via Single-Photon Imaging" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 143, + 200, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 143, + 200, + 156 + ], + "spans": [ + { + "bbox": [ + 127, + 143, + 200, + 156 + ], + "type": "text", + "content": "Varun Sundar†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 108, + 160, + 219, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 160, + 219, + 170 + ], + "spans": [ + { + "bbox": [ + 108, + 160, + 219, + 170 + ], + "type": "text", + "content": "vsundar4@cs.wisc.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 256, + 143, + 342, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 143, + 342, + 156 + ], + "spans": [ + { + "bbox": [ + 256, + 143, + 342, + 156 + ], + "type": "text", + "content": "Andrei Ardelean†" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 250, + 160, + 349, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 160, + 349, + 171 + ], + "spans": [ + { + "bbox": [ + 250, + 160, + 349, + 171 + ], + "type": "text", + "content": "a.ardelean@epfl.ch" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 390, + 143, + 474, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 143, + 474, + 156 + ], + "spans": [ + { + "bbox": [ + 390, + 143, + 474, + 156 + ], + "type": "text", + "content": "Tristan Swedish" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 380, + 160, + 485, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 160, + 485, + 171 + ], + "spans": [ + { + "bbox": [ + 380, + 160, + 485, + 171 + ], + "type": "text", + "content": "tristan@ubicept.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 177, + 226, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 177, + 226, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 177, + 226, + 190 + ], + "type": "text", + "content": "Claudio Bruschini†" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 113, + 194, + 349, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 194, + 349, + 205 + ], + "spans": [ + { + "bbox": [ + 113, + 194, + 349, + 205 + ], + "type": "text", + "content": "{claudio.bruschini, edoardo.charbon}@epfl.ch" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 392, + 176, + 467, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 176, + 467, + 192 + ], + "spans": [ + { + "bbox": [ + 392, + 176, + 467, + 192 + ], + "type": "text", + "content": "Mohit Gupta†,§" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 380, + 194, + 480, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 194, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 380, + 194, + 480, + 205 + ], + "type": "text", + "content": "mohitg@cs.wisc.edu" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 75, + 211, + 517, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 211, + 517, + 227 + ], + "spans": [ + { + "bbox": [ + 75, + 211, + 517, + 227 + ], + "type": "text", + "content": "†University of Wisconsin-Madison ‡École polytechnique fédérale de Lausanne §Ubicept" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 143, + 254, + 192, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 254, + 192, + 266 + ], + "spans": [ + { + "bbox": [ + 143, + 254, + 192, + 266 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 279, + 290, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 290, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 290, + 519 + ], + "type": "text", + "content": "Reinterpretable cameras are defined by their post-processing capabilities that exceed traditional imaging. We present \"SoDaCam\" that provides reinterpretable cameras at the granularity of photons, from photon-cubes acquired by single-photon devices. Photon-cubes represent the spatio-temporal detections of photons as a sequence of binary frames, at frame-rates as high as " + }, + { + "bbox": [ + 46, + 279, + 290, + 519 + ], + "type": "inline_equation", + "content": "100\\mathrm{kHz}" + }, + { + "bbox": [ + 46, + 279, + 290, + 519 + ], + "type": "text", + "content": ". We show that simple transformations of the photon-cube, or photon-cube projections, provide the functionality of numerous imaging systems including: exposure bracketing, flutter shutter cameras, video compressive systems, event cameras, and even cameras that move during exposure. Our photon-cube projections offer the flexibility of being software-defined constructs that are only limited by what is computable, and shot-noise. We exploit this flexibility to provide new capabilities for the emulated cameras. As an added benefit, our projections provide camera-dependent compression of photon-cubes, which we demonstrate using an implementation of our projections on a novel compute architecture that is designed for single-photon imaging." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 540, + 128, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 128, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 128, + 553 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 559, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 287, + 670 + ], + "type": "text", + "content": "Throughout the history of imaging, sensing technologies and the corresponding processing have developed hand-in-hand. In fact, sensing technologies have, to some extent, defined the scope of processing captured data. In the film era, instances of such processing included dodging and burning. The advent of digital cameras provided processing at the granularity of pixels and paved the way for modern computer vision. Light field cameras [34, 78], by sampling the plenoptic function [2], allowed post-capture processing at" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 255, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 255, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 306, + 255, + 545, + 327 + ], + "type": "text", + "content": "the granularity of light rays, enabling novel functionalities such as refocusing photos after-capture. The logical limit of post-capture processing, given the fundamental quantization of light, would be at the level of individual photons. What would imaging look like if we could perform computational processing on individual photons?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 327, + 546, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 546, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 546, + 495 + ], + "type": "text", + "content": "In this work, we show that photon data captured by a new class of single-photon detectors, called single-photon avalanche diodes (SPADs), makes it possible to emulate a wide range of imaging modalities such as exposure bracketing [12], video compressive systems [38, 55] and event cameras [52, 60]. A user then has the flexibility to choose one (or even multiple) of these functionalities post-capture (Fig. 1 (top)). SPAD arrays can operate as extremely high frame-rate photon detectors (" + }, + { + "bbox": [ + 304, + 327, + 546, + 495 + ], + "type": "inline_equation", + "content": "\\sim 100\\mathrm{kHz}" + }, + { + "bbox": [ + 304, + 327, + 546, + 495 + ], + "type": "text", + "content": "), producing a temporal sequence of binary frames called a photon-cube [16]. We show that computing photon-cube projections, which are simple linear and shift operations, can reinterpret the photon-cube to achieve novel post-capture imaging functionalities in a software-defined manner (Fig. 1 (middle))." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "type": "text", + "content": "As case studies, we emulate three distinct imagers: high-speed video compressive imaging; event cameras which respond to dynamic scene content; and motion projections which emulate sensor motion, without any real camera movement. Fig. 1 (bottom) shows the outputs of these cameras that are derived from the same photon-cube." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 582, + 545, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 701 + ], + "type": "text", + "content": "Computing photon-cube projections. One way to obtain photon-cube projections is to read the entire photon-cube off the SPAD array and then perform relevant computations off-chip; we adopt this strategy for our experiments in Secs. 6.1 and 6.2. While reasonable for certain applications, reading out photon-cubes requires an exorbitant data-bandwidth, which can be up to 100 Gbps for a 1 MPixel array—well beyond the capacity of existing data peripherals. Such readout considerations will become center stage as large-format SPAD arrays are fabricated [48, 49]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 318, + 702, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 702, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 702, + 545, + 714 + ], + "type": "text", + "content": "An alternative is to avoid transferring the entire photon-" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 674, + 287, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 674, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 674, + 287, + 714 + ], + "type": "text", + "content": "*Corresponding author: Varun Sundar. This research was supported in parts by NSF CAREER award 1943149, NSF award CNS-2107060, and the Swiss National Science Foundation grant 200021_166289. We also thank Paul Mos for providing access to SwissSPAD2 acquisition software." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8165" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 545, + 201 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 545, + 201 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 545, + 201 + ], + "type": "image", + "image_path": "ac3cc844e8b4994e3c96d64b90cda8bef579a768f693e45dc790917fe510e938.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 202, + 545, + 329 + ], + "blocks": [ + { + "bbox": [ + 52, + 202, + 545, + 329 + ], + "lines": [ + { + "bbox": [ + 52, + 202, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 52, + 202, + 545, + 329 + ], + "type": "image", + "image_path": "517d947cd97d8ef1890b948664efad8ff9c7df797efb8f796c851fc0a82dba69.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 218, + 335, + 376, + 346 + ], + "lines": [ + { + "bbox": [ + 218, + 335, + 376, + 346 + ], + "spans": [ + { + "bbox": [ + 218, + 335, + 376, + 346 + ], + "type": "text", + "content": "Capabilities of photon-cube projections" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 52, + 354, + 164, + 411 + ], + "blocks": [ + { + "bbox": [ + 52, + 354, + 164, + 411 + ], + "lines": [ + { + "bbox": [ + 52, + 354, + 164, + 411 + ], + "spans": [ + { + "bbox": [ + 52, + 354, + 164, + 411 + ], + "type": "image", + "image_path": "d7196f7c9edadd2ca73635dcac2fe8f2ff409c3d5295acaaa5ae0b6845806c80.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 414, + 164, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 164, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 164, + 423 + ], + "type": "text", + "content": "Coded exposure" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 174, + 358, + 217, + 408 + ], + "blocks": [ + { + "bbox": [ + 174, + 358, + 217, + 408 + ], + "lines": [ + { + "bbox": [ + 174, + 358, + 217, + 408 + ], + "spans": [ + { + "bbox": [ + 174, + 358, + 217, + 408 + ], + "type": "image", + "image_path": "dcb2d461ccca256ef02b4a145c5a5ee01e3c3521a0e828babfb30d37ed461e52.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 52, + 427, + 219, + 485 + ], + "blocks": [ + { + "bbox": [ + 52, + 427, + 219, + 485 + ], + "lines": [ + { + "bbox": [ + 52, + 427, + 219, + 485 + ], + "spans": [ + { + "bbox": [ + 52, + 427, + 219, + 485 + ], + "type": "image", + "image_path": "1f7993e7b5524f5c2010889f3c15273ef20bac7a4baafbbd6f137b6d25424784.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 487, + 215, + 510 + ], + "lines": [ + { + "bbox": [ + 59, + 487, + 215, + 510 + ], + "spans": [ + { + "bbox": [ + 59, + 487, + 215, + 510 + ], + "type": "inline_equation", + "content": "12 \\times" + }, + { + "bbox": [ + 59, + 487, + 215, + 510 + ], + "type": "text", + "content": " video reconstruction (raster order) Snapshot compressive imager (Sec. 5.1)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 239, + 353, + 353, + 411 + ], + "blocks": [ + { + "bbox": [ + 239, + 353, + 353, + 411 + ], + "lines": [ + { + "bbox": [ + 239, + 353, + 353, + 411 + ], + "spans": [ + { + "bbox": [ + 239, + 353, + 353, + 411 + ], + "type": "image", + "image_path": "e52254981b0ce021b1dc902ea604da4da8a713e730883b34b1ab21a327e82312.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 413, + 351, + 423 + ], + "lines": [ + { + "bbox": [ + 242, + 413, + 351, + 423 + ], + "spans": [ + { + "bbox": [ + 242, + 413, + 351, + 423 + ], + "type": "text", + "content": "3D scatter plot, " + }, + { + "bbox": [ + 242, + 413, + 351, + 423 + ], + "type": "inline_equation", + "content": "0\\leq t\\leq 0.1s" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 238, + 427, + 353, + 485 + ], + "blocks": [ + { + "bbox": [ + 238, + 427, + 353, + 485 + ], + "lines": [ + { + "bbox": [ + 238, + 427, + 353, + 485 + ], + "spans": [ + { + "bbox": [ + 238, + 427, + 353, + 485 + ], + "type": "image", + "image_path": "c398e22cbebc862536d57707aec275a589f0709aa54198923293d85b6757594d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 487, + 354, + 510 + ], + "lines": [ + { + "bbox": [ + 242, + 487, + 354, + 510 + ], + "spans": [ + { + "bbox": [ + 242, + 487, + 354, + 510 + ], + "type": "text", + "content": "Event image, " + }, + { + "bbox": [ + 242, + 487, + 354, + 510 + ], + "type": "inline_equation", + "content": "0.5 \\leq t \\leq 0.53" + }, + { + "bbox": [ + 242, + 487, + 354, + 510 + ], + "type": "text", + "content": " s Event camera (Sec. 5.2)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 520, + 545, + 604 + ], + "lines": [ + { + "bbox": [ + 46, + 520, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 545, + 604 + ], + "type": "text", + "content": "Figure 1: (top) SoDaCam can emulate a variety of cameras from the photon-cubes acquired by single-photon devices. (middle) Photon-cubes represent the spatio-temporal detection of photons as a sequence of binary frames. Projections of the photon-cube, when computed either on or off-chip, result in reinterpretable and software-defined cameras. We demonstrate the versatility of photon-cube projections on a real dynamic scene: a die falls on a table, bounces, spins in the air and later ricochets off a nearby toy top. (bottom) The cameras emulated by our photon-cube projections can produce: a " + }, + { + "bbox": [ + 46, + 520, + 545, + 604 + ], + "type": "inline_equation", + "content": "12 \\times" + }, + { + "bbox": [ + 46, + 520, + 545, + 604 + ], + "type": "text", + "content": " high-speed video from a single compressive snapshot, event-stream representations of two time intervals (blue and red depict positive and negative spikes respectively), an image where the die appears stationary, as well as a motion-deblurred image." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 372, + 354, + 451, + 411 + ], + "blocks": [ + { + "bbox": [ + 372, + 354, + 451, + 411 + ], + "lines": [ + { + "bbox": [ + 372, + 354, + 451, + 411 + ], + "spans": [ + { + "bbox": [ + 372, + 354, + 451, + 411 + ], + "type": "image", + "image_path": "094aa9f7ec7d9cf191839e995fc0d270dba4e4c82d208c93b1925c638e477261.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 414, + 435, + 422 + ], + "lines": [ + { + "bbox": [ + 387, + 414, + 435, + 422 + ], + "spans": [ + { + "bbox": [ + 387, + 414, + 435, + 422 + ], + "type": "text", + "content": "Static camera" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 372, + 428, + 451, + 485 + ], + "blocks": [ + { + "bbox": [ + 372, + 428, + 451, + 485 + ], + "lines": [ + { + "bbox": [ + 372, + 428, + 451, + 485 + ], + "spans": [ + { + "bbox": [ + 372, + 428, + 451, + 485 + ], + "type": "image", + "image_path": "8c24a2275f618d791908dd8fee6b565343c6620eb9310bf98410846f85bf3cd7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 487, + 447, + 510 + ], + "lines": [ + { + "bbox": [ + 378, + 487, + 447, + 510 + ], + "spans": [ + { + "bbox": [ + 378, + 487, + 447, + 510 + ], + "type": "text", + "content": "Parabolic trajectory Motion Pro" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 463, + 354, + 542, + 411 + ], + "blocks": [ + { + "bbox": [ + 463, + 354, + 542, + 411 + ], + "lines": [ + { + "bbox": [ + 463, + 354, + 542, + 411 + ], + "spans": [ + { + "bbox": [ + 463, + 354, + 542, + 411 + ], + "type": "image", + "image_path": "8bd7d4db4e90149d22171812ec99872fdd68a2f1a432339e077df1406e4e1e8c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 414, + 533, + 423 + ], + "lines": [ + { + "bbox": [ + 473, + 414, + 533, + 423 + ], + "spans": [ + { + "bbox": [ + 473, + 414, + 533, + 423 + ], + "type": "text", + "content": "Linear trajectory" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 463, + 428, + 542, + 485 + ], + "blocks": [ + { + "bbox": [ + 463, + 428, + 542, + 485 + ], + "lines": [ + { + "bbox": [ + 463, + 428, + 542, + 485 + ], + "spans": [ + { + "bbox": [ + 463, + 428, + 542, + 485 + ], + "type": "image", + "image_path": "effe023a8ca759e86bd7881337a3dcf4ea424fcc72fd6118da0fb4f2da2cc3b2.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 487, + 537, + 496 + ], + "lines": [ + { + "bbox": [ + 468, + 487, + 537, + 496 + ], + "spans": [ + { + "bbox": [ + 468, + 487, + 537, + 496 + ], + "type": "text", + "content": "Parabolic deblurred" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 621, + 287, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 621, + 287, + 705 + ], + "spans": [ + { + "bbox": [ + 46, + 621, + 287, + 705 + ], + "type": "text", + "content": "cube by computing projections near sensor. As a proof-of-concept, we implement photon-cube projections on Ultra-Phase [5], a recently-developed programmable SPAD imager with independent processing cores that have dedicated RAM and instruction memory. We show, in Sec. 6.3, that computing projections on-chip greatly reduces sensor readout and, as a consequence, power consumption." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 621, + 545, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 621, + 545, + 705 + ], + "spans": [ + { + "bbox": [ + 304, + 621, + 545, + 705 + ], + "type": "text", + "content": "Implications: Toward a photon-level software-defined camera. The photon-cube projections introduced in this paper are computational constructs that provide a realization of software-defined cameras or SoDaCam. Being software-defined, SoDaCam can emulate multiple cameras simultaneously without additional hardware complexity. SoDaCam, by going beyond baked in hardware choices," + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8166" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "unlocks hitherto unseen capabilities—such as 2000 FPS video from " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "25\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": " readout (Fig. 7); event imaging in very low-light conditions (Fig. 9); and motion stacks, which are a stack of images where in each image, objects only in certain velocity ranges appear sharp (Fig. 6)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 147, + 288, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 147, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 288, + 255 + ], + "type": "text", + "content": "Limitations. The SPAD array [67] used in this work has a relatively low spatial resolution " + }, + { + "bbox": [ + 46, + 147, + 288, + 255 + ], + "type": "inline_equation", + "content": "(512 \\times 256)" + }, + { + "bbox": [ + 46, + 147, + 288, + 255 + ], + "type": "text", + "content": ", and a low fill-factor " + }, + { + "bbox": [ + 46, + 147, + 288, + 255 + ], + "type": "inline_equation", + "content": "(\\sim 10\\%)" + }, + { + "bbox": [ + 46, + 147, + 288, + 255 + ], + "type": "text", + "content": " owing to the lack of microlenses in the prototype used. Similarly, the near-sensor processor that we use has limited capabilities compared to off-chip processors. However, with rapid progress in the development of single-photon cameras [48, 49] and increasing interest in near-sensor processors, we anticipate that many of these shortcomings will be addressed in the upcoming years." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 266, + 134, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 134, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 134, + 277 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 285, + 287, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 370 + ], + "type": "text", + "content": "Reinterpretable imaging has previously been explored at the granularity of light rays [3], by modulating the plenoptic function, and at the level of spatio-temporal voxels [20], by using fast per-pixel shutters. SoDaCam represents a logical culmination of reinterpretability at the level of photon detections, that facilitates multiple post-capture imaging functionalities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 384, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 504 + ], + "type": "text", + "content": "Programmable imaging using a digital micromirror device was first introduced in Nayar et al. [51] to perform precapture radiometric manipulations. Modern programmable cameras are typically near-sensor processors [8, 69, 71] that can perform limited operations in analog [46, 55, 69], while more complex operations [9, 45] occur after analog-to-digital conversion (ADC). In contrast, by performing post-capture computations directly on photon detections, we can perform complex operations without incurring the read-noise penalty that is associated with ADC." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 518, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 651 + ], + "type": "text", + "content": "Passive single-photon imaging. Only recently have SPADs been utilized as passive imaging devices, with applications in high-dynamic range imaging [29, 30, 37, 50], motion-compensation [31, 58], burst photography [41, 42] and object tracking [22]. Compared to compute-intensive burst-photography methods [41], our proposed techniques involve lightweight computations that can be performed near sensor. These computations can also be performed using other single-photon imagers such as Jots [15, 17], which feature higher sensor resolution and photon-efficiency [40], albeit at lower frame-rates and higher read noise." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Reducing the readout of SPADs. Several data reduction strategies have been proposed in the context of SPADs that are used to timestamp incident photons, including: coarse histograms [13, 23, 56], compressive histograms [21], and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "measuring differential time-of-arrivals [70, 77]. When SPADs are operated as photon-detectors, multi-bit counting [48], or summing binary frames, can reduce readout. While compression is not our main objective, we show that photon-cube projections act as camera-specific compression schemes that dramatically reduce sensor readout." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 156, + 545, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 156, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 305, + 156, + 545, + 171 + ], + "type": "text", + "content": "3. Background: Single-Photon Imaging Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 177, + 545, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 177, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 304, + 177, + 545, + 259 + ], + "type": "text", + "content": "A SPAD array captures incident light as a photon-cube: a temporal sequence of binary frames that represents the pixel-wise detection of photons across their respective exposure windows. We can model the stochastic arrival of photons as a Poisson process [72], allowing us to treat spatio-temporal values of the photon-cube as independent Bernoulli random variables with" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 339, + 269, + 545, + 284 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 269, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 339, + 269, + 545, + 284 + ], + "type": "interline_equation", + "content": "\\Pr \\left\\{B _ {t} (\\mathbf {x}) = 1 \\right\\} = 1 - e ^ {- (\\eta \\Phi (\\mathbf {x}, t) + r _ {q}) w _ {\\exp}}, \\tag {1}", + "image_path": "f90e3f92642d1e7ee876f29480fd18f782e2f8e4ce7ad0725da7e3ae5f2ca32d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 293, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 390 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 293, + 545, + 390 + ], + "type": "inline_equation", + "content": "B_{t}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 293, + 545, + 390 + ], + "type": "text", + "content": " represents the value of the photon-cube at pixel " + }, + { + "bbox": [ + 304, + 293, + 545, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 293, + 545, + 390 + ], + "type": "text", + "content": " and exposure index " + }, + { + "bbox": [ + 304, + 293, + 545, + 390 + ], + "type": "inline_equation", + "content": "1 \\leq t \\leq T" + }, + { + "bbox": [ + 304, + 293, + 545, + 390 + ], + "type": "text", + "content": ", which receives a mean incident flux of intensity " + }, + { + "bbox": [ + 304, + 293, + 545, + 390 + ], + "type": "inline_equation", + "content": "\\Phi (\\mathbf{x},t)" + }, + { + "bbox": [ + 304, + 293, + 545, + 390 + ], + "type": "text", + "content": " across its exposure of duration " + }, + { + "bbox": [ + 304, + 293, + 545, + 390 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{exp}}" + }, + { + "bbox": [ + 304, + 293, + 545, + 390 + ], + "type": "text", + "content": ". Additionally, " + }, + { + "bbox": [ + 304, + 293, + 545, + 390 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 293, + 545, + 390 + ], + "type": "text", + "content": " is the photon detection efficiency of the SPAD, and " + }, + { + "bbox": [ + 304, + 293, + 545, + 390 + ], + "type": "inline_equation", + "content": "r_q" + }, + { + "bbox": [ + 304, + 293, + 545, + 390 + ], + "type": "text", + "content": " denotes the sensor's dark count rate—which is the rate of spurious counts unrelated to incident photons. While individual binary frames are extremely noisy, the temporal sum of the photon-cube" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 377, + 399, + 545, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 399, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 377, + 399, + 545, + 433 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {\\text {s u m}} (\\mathbf {x}) := \\sum_ {t = 1} ^ {T} B _ {t} (\\mathbf {x}), \\tag {2}", + "image_path": "5d1b0fffeb33ece9c00d6b83851eead7f4324a9586bb6239557f14711861867e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 441, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 502 + ], + "type": "text", + "content": "can produce an 'image' of the scene that is sharp in static regions, but blurry in dynamic regions (Fig. 2 (top)). Indeed, in static regions, the sum-image can be used to derive a maximum likelihood estimator of the scene intensity [4], given by " + }, + { + "bbox": [ + 304, + 441, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\hat{\\Phi} (\\mathbf{x}) = -\\ln (1 - T^{-1}\\mathcal{I}_{\\mathrm{sum}}(\\mathbf{x})) / \\eta w_{\\mathrm{exp}} - r_q / w_{\\mathrm{exp}}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 513, + 482, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 513, + 482, + 527 + ], + "spans": [ + { + "bbox": [ + 305, + 513, + 482, + 527 + ], + "type": "text", + "content": "4. Projections of the Photon-Cube" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 533, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 689 + ], + "type": "text", + "content": "The temporal sum described in Eq. (2) is a simple instance of projections of a photon-cube. Our key observation is that it is possible to compute a wide range of photon-cube projections, each of which emulates a unique sensing modality post-capture—including modalities that are difficult to achieve with conventional cameras. For example, varying the number of bit-planes that are summed over emulates exposure bracketing [12, 43], which is typically used for HDR imaging. Compared to conventional exposure bracketing, the emulated exposure stack, being software-defined, does not require spatial and temporal registration, which can often be error-prone. Fig. 2 (top) shows an example of an exposure stack computed from a photon-cube." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Going further, we can gradually increase the complexity of the projections. For example, consider a coded exposure" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8167" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 70, + 121, + 129 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 121, + 129 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 121, + 129 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 121, + 129 + ], + "type": "image", + "image_path": "88db00ad5a71bae5abb411466fd9c30ff92884915b3f654f9556302ee6ee7913.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 130, + 115, + 139 + ], + "lines": [ + { + "bbox": [ + 61, + 130, + 115, + 139 + ], + "spans": [ + { + "bbox": [ + 61, + 130, + 115, + 139 + ], + "type": "text", + "content": "Short exposure" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 128, + 70, + 198, + 129 + ], + "blocks": [ + { + "bbox": [ + 128, + 70, + 198, + 129 + ], + "lines": [ + { + "bbox": [ + 128, + 70, + 198, + 129 + ], + "spans": [ + { + "bbox": [ + 128, + 70, + 198, + 129 + ], + "type": "image", + "image_path": "b21ff920af31b12a2ecd1327561bfbdc5f5b85a5c4b93f69209d9fa6d380248d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 130, + 197, + 150 + ], + "lines": [ + { + "bbox": [ + 132, + 130, + 197, + 150 + ], + "spans": [ + { + "bbox": [ + 132, + 130, + 197, + 150 + ], + "type": "text", + "content": "Medium exposure Exposure Stack" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 205, + 70, + 276, + 129 + ], + "blocks": [ + { + "bbox": [ + 205, + 70, + 276, + 129 + ], + "lines": [ + { + "bbox": [ + 205, + 70, + 276, + 129 + ], + "spans": [ + { + "bbox": [ + 205, + 70, + 276, + 129 + ], + "type": "image", + "image_path": "ce91a8a8d2d809a185b9730f0d6e542eea36b63f8784704cdc8972c7b21f1090.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 130, + 267, + 140 + ], + "lines": [ + { + "bbox": [ + 215, + 130, + 267, + 140 + ], + "spans": [ + { + "bbox": [ + 215, + 130, + 267, + 140 + ], + "type": "text", + "content": "Long exposure" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 51, + 158, + 121, + 217 + ], + "blocks": [ + { + "bbox": [ + 51, + 158, + 121, + 217 + ], + "lines": [ + { + "bbox": [ + 51, + 158, + 121, + 217 + ], + "spans": [ + { + "bbox": [ + 51, + 158, + 121, + 217 + ], + "type": "image", + "image_path": "bcdae6bd9a273b24708fe8f67fbac54a45d79cd3f0f19bcc69765cf09a74220f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 218, + 118, + 239 + ], + "lines": [ + { + "bbox": [ + 53, + 218, + 118, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 218, + 118, + 239 + ], + "type": "text", + "content": "Spatially-varying exposure" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 143, + 157, + 184, + 197 + ], + "blocks": [ + { + "bbox": [ + 143, + 157, + 184, + 197 + ], + "lines": [ + { + "bbox": [ + 143, + 157, + 184, + 197 + ], + "spans": [ + { + "bbox": [ + 143, + 157, + 184, + 197 + ], + "type": "image", + "image_path": "1ddfa74e7cdeb3cff1045a6a45807535dd0f29d46ce77d71a177dd9d542a59fc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 143, + 198, + 183, + 237 + ], + "blocks": [ + { + "bbox": [ + 143, + 198, + 183, + 237 + ], + "lines": [ + { + "bbox": [ + 143, + 198, + 183, + 237 + ], + "spans": [ + { + "bbox": [ + 143, + 198, + 183, + 237 + ], + "type": "image", + "image_path": "59c3c43b9ed1c36873b8e5511fae2fa0c159cfdf2fa3fdb72c7a8ac9f34f4962.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 252, + 288, + 336 + ], + "lines": [ + { + "bbox": [ + 46, + 252, + 288, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 288, + 336 + ], + "type": "text", + "content": "Figure 2: Coded exposures from photon-cubes. (top) An exposure stack with sum-images computed using 250, 500, and 1000 bit-planes. Short exposures are noisy while long exposures exhibit motion-blur. (bottom) Spatially-varying exposure that uses a quad pattern [32] (see inset), and a video compressive frame that uses 16 random binary masks to modulate the photon-cube. Zoom-in to see details." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 205, + 158, + 276, + 218 + ], + "blocks": [ + { + "bbox": [ + 205, + 158, + 276, + 218 + ], + "lines": [ + { + "bbox": [ + 205, + 158, + 276, + 218 + ], + "spans": [ + { + "bbox": [ + 205, + 158, + 276, + 218 + ], + "type": "image", + "image_path": "348b598fcc997840b7770eb9f2714bfe7e8a29982af3ca7db9b49972480c320d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 219, + 278, + 239 + ], + "lines": [ + { + "bbox": [ + 205, + 219, + 278, + 239 + ], + "spans": [ + { + "bbox": [ + 205, + 219, + 278, + 239 + ], + "type": "text", + "content": "Video compressive frame" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 356, + 285, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 285, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 285, + 369 + ], + "type": "text", + "content": "projection that multiplexes bit-planes with a temporal code" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 111, + 374, + 287, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 374, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 111, + 374, + 287, + 407 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {\\text {f l u t t e r}} (\\mathbf {x}) := \\sum_ {t = 1} ^ {T} C _ {t} B _ {t} (\\mathbf {x}), \\tag {3}", + "image_path": "813f4d14fcf1b88dd3fa63cb489f4d6f5034a72a2b735d06ef7cbdc91f4f119e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 413, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 413, + 287, + 448 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 46, + 413, + 287, + 448 + ], + "type": "text", + "content": " is the temporal code. An example of globally-coded exposures is the flutter shutter camera [53], which uses pseudo-random binary codes for motion-deblurring." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 449, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 472 + ], + "type": "text", + "content": "More general coded exposures can be obtained via spatially-varying temporal coding patterns " + }, + { + "bbox": [ + 46, + 449, + 287, + 472 + ], + "type": "inline_equation", + "content": "C_t(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 449, + 287, + 472 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 479, + 287, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 479, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 104, + 479, + 287, + 512 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {\\text {c o d e d}} (\\mathbf {x}) := \\sum_ {t = 1} ^ {T} C _ {t} (\\mathbf {x}) B _ {t} (\\mathbf {x}). \\tag {4}", + "image_path": "c299cf41d1213b2f2468238f8ddc3f16ae22e46ee0e697c086cddf90be86953d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 517, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 613 + ], + "type": "text", + "content": "Fig. 2 (bottom) shows spatially-varying exposures that use a quad (Bayer-like) spatial pattern and random binary masks. With photon-cubes, we can perform spatially-varying coding without bulky spatial light modulators, similar to focalplane sensor-processors [46, 69]. Moreover, we can capture multiple coded exposures simultaneously, which is challenging to realize in existing sensors. In Sec. 5.1, we describe coding patterns for video compressive sensing." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 613, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 287, + 661 + ], + "type": "text", + "content": "Spatial and temporal gradients form the building blocks of several computer vision algorithms [7, 11, 24, 26, 39]. Given this, another projection of interest is temporal contrast, i.e., a derivative filter preceded by a smoothing filter:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 99, + 673, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 673, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 99, + 673, + 287, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {\\text {c o n t r a s t}} (\\mathbf {x}, t) := D _ {t} \\circ G * B _ {t} (\\mathbf {x}), \\tag {5}", + "image_path": "dc5259e456fa0016ba1273aa06d70c515b26333d9868d0ef129dfbcdb72a46f2.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": " is the difference operator, " + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": " could be exponential or Gaussian smoothing, " + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": " denotes function composition," + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": "and * denotes convolution. Due to their sparse nature, gradients form the basis of bandwidth- and power-efficient event cameras [10, 14, 36, 60], which we emulate in Sec. 5.2." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 108, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 108, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 305, + 108, + 545, + 204 + ], + "type": "text", + "content": "So far, we have considered projections taken only along the time axis. Next, we consider a more general class of spatio-temporal projections that lead to novel functionalities. For instance, computing a simple projection, such as the temporal sum, along arbitrary spatio-temporal directions emulates sensor motion during exposure time [33], but without moving the sensor. We achieve this by shifting bitplanes and computing their sum:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 361, + 211, + 545, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 211, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 361, + 211, + 545, + 243 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {\\text {s h i f t}} (\\mathbf {x}) := \\sum_ {t = 1} ^ {T} B _ {t} (\\mathbf {x} + \\mathbf {r} (t)), \\tag {6}", + "image_path": "237009e8a9a718f4a464abf25f70168a3c37cd3e69af289b3feb5c0840c96566.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 251, + 545, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 298 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 251, + 545, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 304, + 251, + 545, + 298 + ], + "type": "text", + "content": " is a discretized 2D trajectory that determines sensor motion. Outside a software-defined framework, such projections are hard to realize without physical actuators. We describe the capabilities of motion projections in Sec. 5.3." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 299, + 545, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 545, + 406 + ], + "type": "text", + "content": "In summary, the proposed photon-cube projections are simple linear and shift operators that lead to a diverse set of post-capture imaging functionalities. These projections pave the way for future 'swiss-army-knife' imaging systems that achieve multiple functionalities (e.g., event cameras, high-speed cameras, conventional cameras, HDR cameras) simultaneously with a single sensor. Finally, these projections can be computed efficiently in an online manner, which makes on-chip implementation viable (Sec. 6.3)." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 406, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 406, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 545, + 490 + ], + "type": "text", + "content": "At this point, we note that a key enabling factor of photon-cube projections is the extremely high temporal-sampling rate of SPADs. Indeed, the temporal sampling rate determines key aspects of sensor emulation, such as the discretization of temporal derivatives and motion trajectories. This raises a natural question: can we use conventional high-speed cameras for computing projections?" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 503, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 545, + 659 + ], + "type": "text", + "content": "Trade-off between frame-rate and SNR. In principle, photon-cube projections can be computed using regular (CMOS or CCD based) high-speed cameras. Unfortunately, each frame captured by a high-speed camera incurs a read-noise penalty, which increases with the camera's frame-rate [6]. In fact, the read noise levels of high-speed cameras [1] can be " + }, + { + "bbox": [ + 304, + 503, + 545, + 659 + ], + "type": "inline_equation", + "content": "10 - 30 \\times" + }, + { + "bbox": [ + 304, + 503, + 545, + 659 + ], + "type": "text", + "content": " higher than consumer cameras [28]. Coupled with the low per-frame incident flux at high frame-rates, high levels of read noise result in extremely low SNRs. In contrast, SPADs do not incur a per-frame read noise and are limited only by the fundamental photon noise. Hence, for the post-capture software-defined functionalities proposed here, it is imperative to use SPADs." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 669, + 527, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 669, + 527, + 683 + ], + "spans": [ + { + "bbox": [ + 305, + 669, + 527, + 683 + ], + "type": "text", + "content": "5. Emulating Cameras from Photon-Cubes" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Sec. 4 presented the concept of photon-cube projections, and its potential for achieving multiple post-capture imag-" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8168" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 73, + 120, + 162 + ], + "blocks": [ + { + "bbox": [ + 49, + 73, + 120, + 162 + ], + "lines": [ + { + "bbox": [ + 49, + 73, + 120, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 73, + 120, + 162 + ], + "type": "image", + "image_path": "a2900e485e032033d45c703a41e5a3da27819274f01258de20b1e17687313898.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 175, + 287, + 282 + ], + "lines": [ + { + "bbox": [ + 46, + 175, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 287, + 282 + ], + "type": "text", + "content": "Figure 3: Modulating masks for video compressive sensing. (left) A single VCS measurement temporally compresses a sequence of frames using binary random masks. (center) Two-bucket cameras capture an additional measurement by using the complementary mask sequence. (right) We propose using multi-bucket captures by randomly choosing an active bucket for each frame. Both two-bucket and multi-bucket captures have " + }, + { + "bbox": [ + 46, + 175, + 287, + 282 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 175, + 287, + 282 + ], + "type": "text", + "content": " light efficiency. All masks are visualized here for " + }, + { + "bbox": [ + 46, + 175, + 287, + 282 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 46, + 175, + 287, + 282 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 127, + 72, + 200, + 162 + ], + "blocks": [ + { + "bbox": [ + 127, + 72, + 200, + 162 + ], + "lines": [ + { + "bbox": [ + 127, + 72, + 200, + 162 + ], + "spans": [ + { + "bbox": [ + 127, + 72, + 200, + 162 + ], + "type": "image", + "image_path": "329176052390257c785be133c10c0d93034f35f2074c8fdefd7de845acc46489.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 209, + 70, + 285, + 162 + ], + "blocks": [ + { + "bbox": [ + 209, + 70, + 285, + 162 + ], + "lines": [ + { + "bbox": [ + 209, + 70, + 285, + 162 + ], + "spans": [ + { + "bbox": [ + 209, + 70, + 285, + 162 + ], + "type": "image", + "image_path": "1ec8b5fdaec299ac3fa57af842c242c1b4b073e83881f95d970329dd37809727.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 301, + 286, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 286, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 286, + 397 + ], + "type": "text", + "content": "ing functionalities. As case studies, we now demonstrate three imaging modalities: video compressive sensing, event cameras, and motion-projection cameras. These modalities have been well-studied over several years; in particular, there exist active research communities around video compressive sensing and event cameras today. We also show new variants of these imaging systems that arise from the software-defined nature of photon-cube projections." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 407, + 198, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 198, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 198, + 419 + ], + "type": "text", + "content": "5.1. Video Compressive Sensing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "type": "text", + "content": "Video compressive systems optically multiplex light with random binary masks, such as the patterns in Fig. 3 (left). As discussed in the previous section, such multiplexing can be achieved computationally using photon-cubes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 491, + 286, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 286, + 601 + ], + "type": "text", + "content": "Two-bucket cameras. One drawback of capturing coded measurements is the light loss due to blocking of incident light. To prevent loss of light, coded two-bucket cameras [69] capture an additional measurement that is modulated by the complementary mask sequence (Fig. 3 (center)). Such measurements recover higher quality frames, even after accounting for the extra readout [35, 61]. Two-bucket captures can be readily derived from photon-cubes, by implementing Eq. (4) with the additional mask sequence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": "Multi-bucket cameras. We can extend the idea of two-bucket captures to multi-bucket captures by accumulating bit-planes in one of " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": " buckets that is randomly chosen at each time instant and pixel location. Since multiplexing is performed computationally, we do not face any loss in photoreceptive area that [59, 68] hampers existing multi-bucket sensors. Multi-bucket captures can reconstruct a large number of frames by better conditioning video recovery and" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 410, + 142 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 410, + 142 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 410, + 142 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 410, + 142 + ], + "type": "image", + "image_path": "fdeb5aad951fe92184d1b86c9c064d8f9671f1bff55af3ce9963cd79b23afca2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 142, + 397, + 153 + ], + "lines": [ + { + "bbox": [ + 324, + 142, + 397, + 153 + ], + "spans": [ + { + "bbox": [ + 324, + 142, + 397, + 153 + ], + "type": "inline_equation", + "content": "\\log (x)" + }, + { + "bbox": [ + 324, + 142, + 397, + 153 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 324, + 142, + 397, + 153 + ], + "type": "inline_equation", + "content": "1 - e^{-x}" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 410, + 72, + 533, + 142 + ], + "blocks": [ + { + "bbox": [ + 410, + 72, + 533, + 142 + ], + "lines": [ + { + "bbox": [ + 410, + 72, + 533, + 142 + ], + "spans": [ + { + "bbox": [ + 410, + 72, + 533, + 142 + ], + "type": "image", + "image_path": "0f2f2a0ea4d297cf3dff34efd3a939b2da3654b3a0c873c36ada8fe8ac5c8fb4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 142, + 512, + 153 + ], + "lines": [ + { + "bbox": [ + 429, + 142, + 512, + 153 + ], + "spans": [ + { + "bbox": [ + 429, + 142, + 512, + 153 + ], + "type": "text", + "content": "Spatio-temporal plot" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 163, + 384, + 217 + ], + "blocks": [ + { + "bbox": [ + 310, + 163, + 384, + 217 + ], + "lines": [ + { + "bbox": [ + 310, + 163, + 384, + 217 + ], + "spans": [ + { + "bbox": [ + 310, + 163, + 384, + 217 + ], + "type": "image", + "image_path": "6d6626c534dae7f70bd1c71edb0d94ca56af365011cd170ff6542b41a3c960e0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 219, + 383, + 227 + ], + "lines": [ + { + "bbox": [ + 315, + 219, + 383, + 227 + ], + "spans": [ + { + "bbox": [ + 315, + 219, + 383, + 227 + ], + "type": "inline_equation", + "content": "\\tau = 0.45, \\beta = 0.95" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 326, + 228, + 370, + 237 + ], + "lines": [ + { + "bbox": [ + 326, + 228, + 370, + 237 + ], + "spans": [ + { + "bbox": [ + 326, + 228, + 370, + 237 + ], + "type": "text", + "content": "5 MEvent/s" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 333, + 239, + 517, + 251 + ], + "lines": [ + { + "bbox": [ + 333, + 239, + 517, + 251 + ], + "spans": [ + { + "bbox": [ + 333, + 239, + 517, + 251 + ], + "type": "text", + "content": "Effect of contrast-threshold, exponential decay" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 263, + 545, + 383 + ], + "lines": [ + { + "bbox": [ + 304, + 263, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 545, + 383 + ], + "type": "text", + "content": "Figure 4: Event stream from photon-cubes. (top left) By exploiting the non-linear response curve of SPADs to encode brightness, we can avoid the underflow issues of a log-response. We visualize events generated from photon-cubes using a 3D scatter plot of polarities (top right, 14000 bit-planes), and frame accumulation of events (bottom, 1200 bit-planes). Blue and red denote positive and negative spikes respectively. The event images also show the effect of varying the contrast threshold " + }, + { + "bbox": [ + 304, + 263, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 263, + 545, + 383 + ], + "type": "text", + "content": " and exponential decay " + }, + { + "bbox": [ + 304, + 263, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 263, + 545, + 383 + ], + "type": "text", + "content": "-larger values yield a less noisy but sparser event stream." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 389, + 163, + 463, + 217 + ], + "blocks": [ + { + "bbox": [ + 389, + 163, + 463, + 217 + ], + "lines": [ + { + "bbox": [ + 389, + 163, + 463, + 217 + ], + "spans": [ + { + "bbox": [ + 389, + 163, + 463, + 217 + ], + "type": "image", + "image_path": "de92179dc5d733509ee24dfe131bba7dd85d3d05a3af8b7033c6657f06ca72c7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 219, + 459, + 227 + ], + "lines": [ + { + "bbox": [ + 392, + 219, + 459, + 227 + ], + "spans": [ + { + "bbox": [ + 392, + 219, + 459, + 227 + ], + "type": "inline_equation", + "content": "\\tau = 0.25, \\beta = 0.95" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 401, + 228, + 450, + 237 + ], + "lines": [ + { + "bbox": [ + 401, + 228, + 450, + 237 + ], + "spans": [ + { + "bbox": [ + 401, + 228, + 450, + 237 + ], + "type": "text", + "content": "25 MEvent/s" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 467, + 164, + 541, + 217 + ], + "blocks": [ + { + "bbox": [ + 467, + 164, + 541, + 217 + ], + "lines": [ + { + "bbox": [ + 467, + 164, + 541, + 217 + ], + "spans": [ + { + "bbox": [ + 467, + 164, + 541, + 217 + ], + "type": "image", + "image_path": "06734ec02cad1a28bf0b71f3b94d69fa076e60210fcdc765fed2634d9606b14b.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 219, + 537, + 227 + ], + "lines": [ + { + "bbox": [ + 470, + 219, + 537, + 227 + ], + "spans": [ + { + "bbox": [ + 470, + 219, + 537, + 227 + ], + "type": "inline_equation", + "content": "\\tau = 0.45, \\beta = 0.98" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 478, + 228, + 528, + 237 + ], + "lines": [ + { + "bbox": [ + 478, + 228, + 528, + 237 + ], + "spans": [ + { + "bbox": [ + 478, + 228, + 528, + 237 + ], + "type": "text", + "content": "2.5 MEvent/s" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 405, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 405, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 405, + 545, + 430 + ], + "type": "text", + "content": "provide extreme high-speed video imaging. Fig. 3 (right) shows the modulating masks for a four-bucket capture." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 438, + 400, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 438, + 400, + 450 + ], + "spans": [ + { + "bbox": [ + 306, + 438, + 400, + 450 + ], + "type": "text", + "content": "5.2.Event Cameras" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 458, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 555 + ], + "type": "text", + "content": "Next, we describe emulation of event-cameras, which capture changes in light intensity and are conceptually similar to the temporal contrast projection introduced in Eq. (5). Physical implementations of event sensors [10, 14, 36, 60] generate a photoreceptor voltage " + }, + { + "bbox": [ + 304, + 458, + 545, + 555 + ], + "type": "inline_equation", + "content": "V(\\mathbf{x}, t)" + }, + { + "bbox": [ + 304, + 458, + 545, + 555 + ], + "type": "text", + "content": " with a logarithmic response to incident flux " + }, + { + "bbox": [ + 304, + 458, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\Phi(\\mathbf{x}, t)" + }, + { + "bbox": [ + 304, + 458, + 545, + 555 + ], + "type": "text", + "content": ", and output an event " + }, + { + "bbox": [ + 304, + 458, + 545, + 555 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}, t, p)" + }, + { + "bbox": [ + 304, + 458, + 545, + 555 + ], + "type": "text", + "content": " when this voltage deviates sufficiently from a reference voltage " + }, + { + "bbox": [ + 304, + 458, + 545, + 555 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{ref}}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 458, + 545, + 555 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 376, + 564, + 545, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 564, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 376, + 564, + 545, + 578 + ], + "type": "interline_equation", + "content": "\\left| V (\\mathbf {x}, t) - V _ {\\text {r e f}} (\\mathbf {x}) \\right| > \\tau , \\tag {7}", + "image_path": "bc14a7e529e1c4f650e7d78afa870f8e5e62425fc39e2c9502b7f48c05a8e76f.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 588, + 545, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 545, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 545, + 648 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 588, + 545, + 648 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 588, + 545, + 648 + ], + "type": "text", + "content": " is called the contrast-threshold and " + }, + { + "bbox": [ + 304, + 588, + 545, + 648 + ], + "type": "inline_equation", + "content": "p = \\mathrm{sign}(V(\\mathbf{x},t) - V_{\\mathrm{ref}}(\\mathbf{x}))" + }, + { + "bbox": [ + 304, + 588, + 545, + 648 + ], + "type": "text", + "content": " encodes the polarity of the event. Once an event is generated, " + }, + { + "bbox": [ + 304, + 588, + 545, + 648 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{ref}}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 588, + 545, + 648 + ], + "type": "text", + "content": " is updated to " + }, + { + "bbox": [ + 304, + 588, + 545, + 648 + ], + "type": "inline_equation", + "content": "V(\\mathbf{x},t)" + }, + { + "bbox": [ + 304, + 588, + 545, + 648 + ], + "type": "text", + "content": ". Eq. (7), for a smoothly-varying flux intensity, thresholds a function of the temporal gradient, i.e., " + }, + { + "bbox": [ + 304, + 588, + 545, + 648 + ], + "type": "inline_equation", + "content": "\\partial_t\\log (\\Phi (\\mathbf{x},t))" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": "From bit-planes to event streams. To produce events from SPAD frames, we compute an exponential moving average (EMA) of the bit-planes, as " + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mu_t(\\mathbf{x}) = (1 - \\beta)B_t(\\mathbf{x}) + \\beta \\mu_{t - 1}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mu_t(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": " is the EMA, " + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": " is the smoothing" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8169" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 161, + 125 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 161, + 125 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 161, + 125 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 161, + 125 + ], + "type": "image", + "image_path": "e6656c4cc37d7fcdd415afb7de28cc059369c90e3c76c3dc0a9ae037f0872d8c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 127, + 157, + 162 + ], + "blocks": [ + { + "bbox": [ + 51, + 127, + 157, + 162 + ], + "lines": [ + { + "bbox": [ + 51, + 127, + 157, + 162 + ], + "spans": [ + { + "bbox": [ + 51, + 127, + 157, + 162 + ], + "type": "image", + "image_path": "e0a08c526f44f3d877f4c3f47dae2342229790c25a9af49896fbf19b92bc213d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 284, + 125 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 284, + 125 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 284, + 125 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 284, + 125 + ], + "type": "image", + "image_path": "4f07bcadec453ebbbb6d8c55fd2d414878ed1f3e92d89c7ac201bb81c330f8d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 174, + 126, + 280, + 163 + ], + "blocks": [ + { + "bbox": [ + 174, + 126, + 280, + 163 + ], + "lines": [ + { + "bbox": [ + 174, + 126, + 280, + 163 + ], + "spans": [ + { + "bbox": [ + 174, + 126, + 280, + 163 + ], + "type": "image", + "image_path": "fc72fb42f026bf3e2e660bf720973e961cff5ae13bf1eb0a86b5a434c3e1574a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 167, + 161, + 223 + ], + "blocks": [ + { + "bbox": [ + 50, + 167, + 161, + 223 + ], + "lines": [ + { + "bbox": [ + 50, + 167, + 161, + 223 + ], + "spans": [ + { + "bbox": [ + 50, + 167, + 161, + 223 + ], + "type": "image", + "image_path": "c1069f1838814529b3a092bebf6aa903e55fbb56a6bf29b0c91199a5ba9c0ec3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 51, + 224, + 157, + 248 + ], + "blocks": [ + { + "bbox": [ + 51, + 224, + 157, + 248 + ], + "lines": [ + { + "bbox": [ + 51, + 224, + 157, + 248 + ], + "spans": [ + { + "bbox": [ + 51, + 224, + 157, + 248 + ], + "type": "image", + "image_path": "7e017e1f64b6f1df9f7f6122d4c06c01395bd01338fba431a34cfe05f8789365.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 251, + 144, + 262 + ], + "lines": [ + { + "bbox": [ + 66, + 251, + 144, + 262 + ], + "spans": [ + { + "bbox": [ + 66, + 251, + 144, + 262 + ], + "type": "text", + "content": "Parabolic trajectory" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 173, + 167, + 284, + 223 + ], + "blocks": [ + { + "bbox": [ + 173, + 167, + 284, + 223 + ], + "lines": [ + { + "bbox": [ + 173, + 167, + 284, + 223 + ], + "spans": [ + { + "bbox": [ + 173, + 167, + 284, + 223 + ], + "type": "image", + "image_path": "f5fdd82605fb6883608f0bde223d092297f6e17b647864f20f3c0f02bc4d5689.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 228, + 545, + 324 + ], + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 324 + ], + "type": "text", + "content": "Figure 6: Motion stack. Computing multiple linear projections with different trajectories can produce a stack of images where objects with matching velocity are sharp. Here, we show a traffic scene involving four cars that have four different velocities. By suitably altering the slope of the linear trajectory, we can produce images where only one of the cars appear sharp at a time. We indicate the slope of the trajectories chosen and the objects that are \"in-focus\"." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 191, + 225, + 258, + 248 + ], + "blocks": [ + { + "bbox": [ + 191, + 225, + 258, + 248 + ], + "lines": [ + { + "bbox": [ + 191, + 225, + 258, + 248 + ], + "spans": [ + { + "bbox": [ + 191, + 225, + 258, + 248 + ], + "type": "image", + "image_path": "6d9b5bad09bb7c41561d2b68c602b314790e14ef4ff4e33bf35d06198d26e7e7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 251, + 268, + 261 + ], + "lines": [ + { + "bbox": [ + 190, + 251, + 268, + 261 + ], + "spans": [ + { + "bbox": [ + 190, + 251, + 268, + 261 + ], + "type": "text", + "content": "Parabolic deblurred" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 274, + 289, + 383 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 289, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 289, + 383 + ], + "type": "text", + "content": "Figure 5: Motion projections. (top) Integrating along a linear trajectory in the photon-cube changes the apparent image-space velocity of scene objects. Details are seen for (top left) the case when static, and (top right) the metallic tape when the sensor translates along the " + }, + { + "bbox": [ + 46, + 274, + 289, + 383 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 274, + 289, + 383 + ], + "type": "text", + "content": "-axis. (bottom) A parabolic integration trajectory results in a motion-invariant image, resulting in similar blur kernels for all objects. (bottom right) Deblurring with the resultant shift-invariant point spread function (shown in inset) produces a sharp image." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 403, + 287, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 429 + ], + "type": "text", + "content": "factor, and " + }, + { + "bbox": [ + 46, + 403, + 287, + 429 + ], + "type": "inline_equation", + "content": "B_{t}" + }, + { + "bbox": [ + 46, + 403, + 287, + 429 + ], + "type": "text", + "content": " is a bit-plane. We generate an event when " + }, + { + "bbox": [ + 46, + 403, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\mu_t(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 403, + 287, + 429 + ], + "type": "text", + "content": " deviates from " + }, + { + "bbox": [ + 46, + 403, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathrm{ref}}(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 403, + 287, + 429 + ], + "type": "text", + "content": " by at least " + }, + { + "bbox": [ + 46, + 403, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 403, + 287, + 429 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 437, + 287, + 451 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 437, + 287, + 451 + ], + "spans": [ + { + "bbox": [ + 107, + 437, + 287, + 451 + ], + "type": "interline_equation", + "content": "\\left| h \\left(\\mu_ {t} (\\mathbf {x})\\right) - h \\left(\\mu_ {\\text {r e f}} (\\mathbf {x})\\right) \\right| > \\tau , \\tag {8}", + "image_path": "df1d5c0ea78ec15dc85a0354f509eb5c63877d8c0d0693603bdc2ab494280dec.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 459, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 459, + 287, + 495 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 459, + 287, + 495 + ], + "type": "text", + "content": " is a scalar function applied to the EMA. We can see that Eq. (8) thresholds temporal contrast, by observing the role played by the EMA and the difference operator." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 496, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 590 + ], + "type": "text", + "content": "Setting " + }, + { + "bbox": [ + 46, + 496, + 287, + 590 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 496, + 287, + 590 + ], + "type": "text", + "content": " to be the logarithm of the flux MLE mimics Eq. (7). However, since the log-scale is used to prevent sensor saturation, a simpler alternative is to use the non-saturating response curve of SPAD pixels " + }, + { + "bbox": [ + 46, + 496, + 287, + 590 + ], + "type": "inline_equation", + "content": "(h(x) = x)" + }, + { + "bbox": [ + 46, + 496, + 287, + 590 + ], + "type": "text", + "content": ". The response curve takes the form of " + }, + { + "bbox": [ + 46, + 496, + 287, + 590 + ], + "type": "inline_equation", + "content": "1 - \\exp (-\\alpha \\Phi (\\mathbf{x},t))" + }, + { + "bbox": [ + 46, + 496, + 287, + 590 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 496, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 496, + 287, + 590 + ], + "type": "text", + "content": " is a flux-independent constant. As a major advantage, this response curve avoids the underflow issues of the log function that can occur in low-light scenarios [62]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 591, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 591, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 287, + 651 + ], + "type": "text", + "content": "The SPAD's frame-rate determines the time-stamp resolution of emulated events. In Fig. 4, we show the events generated from a photon-cube acquired at a frame-rate of " + }, + { + "bbox": [ + 46, + 591, + 287, + 651 + ], + "type": "inline_equation", + "content": "96.8\\mathrm{kHz}" + }, + { + "bbox": [ + 46, + 591, + 287, + 651 + ], + "type": "text", + "content": " resulting in a time-stamp resolution of " + }, + { + "bbox": [ + 46, + 591, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\sim 10~\\mu \\mathrm{s}" + }, + { + "bbox": [ + 46, + 591, + 287, + 651 + ], + "type": "text", + "content": " that is comparable to those of existing event cameras." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "How do SPAD-events differ from the output of a regular event camera? The main difference is the expression of temporal contrast, given by " + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\partial_t h" + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": ", is now " + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "inline_equation", + "content": "-\\partial_t \\exp(-\\alpha \\Phi(\\mathbf{x}, t))" + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": ", instead of " + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\partial_t \\log(\\Phi(\\mathbf{x}, t))" + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": ". This dif" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 427, + 144 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 427, + 144 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 427, + 144 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 427, + 144 + ], + "type": "image", + "image_path": "1709c7778e3cf88a2a004695e25143a4ad7fa0a2fd2d5f915e20ffbaa82ef9bc.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 313, + 148, + 427, + 207 + ], + "blocks": [ + { + "bbox": [ + 313, + 148, + 427, + 207 + ], + "lines": [ + { + "bbox": [ + 313, + 148, + 427, + 207 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 427, + 207 + ], + "type": "image", + "image_path": "1c41b3c328fb8d6d2a989938fa554a706049d4f82ec3037b4759f94902af11c0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 209, + 416, + 220 + ], + "lines": [ + { + "bbox": [ + 326, + 209, + 416, + 220 + ], + "spans": [ + { + "bbox": [ + 326, + 209, + 416, + 220 + ], + "type": "inline_equation", + "content": "160\\hat{x} +447\\hat{y}" + }, + { + "bbox": [ + 326, + 209, + 416, + 220 + ], + "type": "text", + "content": " pixels/sec" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 430, + 70, + 543, + 130 + ], + "blocks": [ + { + "bbox": [ + 430, + 70, + 543, + 130 + ], + "lines": [ + { + "bbox": [ + 430, + 70, + 543, + 130 + ], + "spans": [ + { + "bbox": [ + 430, + 70, + 543, + 130 + ], + "type": "image", + "image_path": "d4a1b9cda9e01fe1ab01df568296f6ea507966676fdf96c631718e543222b385.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 131, + 534, + 143 + ], + "lines": [ + { + "bbox": [ + 441, + 131, + 534, + 143 + ], + "spans": [ + { + "bbox": [ + 441, + 131, + 534, + 143 + ], + "type": "inline_equation", + "content": "100\\hat{x} + 282\\hat{y}" + }, + { + "bbox": [ + 441, + 131, + 534, + 143 + ], + "type": "text", + "content": " pixels/sec" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 430, + 148, + 543, + 207 + ], + "blocks": [ + { + "bbox": [ + 430, + 148, + 543, + 207 + ], + "lines": [ + { + "bbox": [ + 430, + 148, + 543, + 207 + ], + "spans": [ + { + "bbox": [ + 430, + 148, + 543, + 207 + ], + "type": "image", + "image_path": "f60191071b30b5e4fa35b5ee6cd7761d55a5f580686286e62f6ac2af1c7b5ce5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 439, + 209, + 537, + 220 + ], + "lines": [ + { + "bbox": [ + 439, + 209, + 537, + 220 + ], + "spans": [ + { + "bbox": [ + 439, + 209, + 537, + 220 + ], + "type": "inline_equation", + "content": "420\\hat{x} + 1177\\hat{y}" + }, + { + "bbox": [ + 439, + 209, + 537, + 220 + ], + "type": "text", + "content": " pixels/sec" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 346, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 441 + ], + "type": "text", + "content": "ference poses no compatibility issues for a large class of event-vision algorithms that utilize a grid of events [54, 57, 66] or brightness changes [18]. We show examples of downstream applications using SPAD-events in Supplementary Note 2. Finally, SPAD-events can be easily augmented with spatially- and temporally-aligned intensity information—a synergistic combination that has been exploited by several recent event-vision works [18, 25, 74]." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 449, + 418, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 418, + 462 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 418, + 462 + ], + "type": "text", + "content": "5.3. Motion Projections" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 468, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 529 + ], + "type": "text", + "content": "Having described the emulation of cameras that capture coded exposures and temporal contrasts, we now shift our attention to cameras that emulate sensor motion during exposure, viz. motion cameras. We describe two useful trajectories when emulating motion cameras using Eq. (6)." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 543, + 545, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 545, + 603 + ], + "type": "text", + "content": "Linear trajectory. The simplest sensor trajectory involves linear motion, where " + }, + { + "bbox": [ + 304, + 543, + 545, + 603 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(t) = (bt + c)\\hat{\\mathbf{p}}" + }, + { + "bbox": [ + 304, + 543, + 545, + 603 + ], + "type": "text", + "content": " for some constants " + }, + { + "bbox": [ + 304, + 543, + 545, + 603 + ], + "type": "inline_equation", + "content": "b, c \\in \\mathbb{R}" + }, + { + "bbox": [ + 304, + 543, + 545, + 603 + ], + "type": "text", + "content": " and unit vector " + }, + { + "bbox": [ + 304, + 543, + 545, + 603 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}" + }, + { + "bbox": [ + 304, + 543, + 545, + 603 + ], + "type": "text", + "content": ". As Fig. 5 (top row) shows, this can change the scene's frame of reference: making moving objects appear stationary and vice-versa." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": "Motion-invariant parabolic projection. If motion is along " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": ", parabolic integration produces a motion-invariant image [33]—all objects, irrespective of their velocity, are blurred by the same point spread function (PSF), up to a linear shift. Thus, a deblurred parabolic capture produces a sharp image of all velocity groups (Fig. 5 (bottom row)). The parabolic trajectory is given by " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(t) = (at^2 + bt + c)\\hat{\\mathbf{p}}" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": ". We choose " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": " based on the maximum object velocity and " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "b, c" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8170" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 71, + 544, + 180 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 544, + 180 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 544, + 180 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 544, + 180 + ], + "type": "image", + "image_path": "f3e5ed354caccbee3543c8c2950ba4e8c8629eadd0e5ca9b85c43134055c8fa5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 189, + 547, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 189, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 547, + 239 + ], + "type": "text", + "content": "Figure 7: High-speed videography at 2000 FPS of a tennis ball dropped into a bowl of water, from a " + }, + { + "bbox": [ + 46, + 189, + 547, + 239 + ], + "type": "inline_equation", + "content": "25\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 189, + 547, + 239 + ], + "type": "text", + "content": " readout. The conventional capture provides a visualization of the scene dynamics. It is challenging to reconstruct a large number of frames from a single compressive snapshot. Multi-bucket captures recover frames with significantly greater detail, such as the crown of water surrounding the ball. We include more sequences (e.g., a bursting balloon) in the supplementary material." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 245, + 545, + 421 + ], + "blocks": [ + { + "bbox": [ + 48, + 245, + 545, + 421 + ], + "lines": [ + { + "bbox": [ + 48, + 245, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 245, + 545, + 421 + ], + "type": "image", + "image_path": "b47c6bfabc84357280f9d81926085a1217888186ea2f7f273cda943887699918.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 430, + 547, + 479 + ], + "lines": [ + { + "bbox": [ + 46, + 430, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 547, + 479 + ], + "type": "text", + "content": "Figure 8: Deblurring of traffic scenes using motion projections. Linear projections can recover details of moving objects if their velocity is known. When only the motion direction is known (e.g., road's orientation), a sharp image can be obtained by either deblurring a parabolic projection or by blending multiple randomly-sampled linear projections. We quantitatively compare against the compute- and bandwidth-expensive Quanta Burst Photography [41], based on PSNR and LPIPS [76]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 498, + 288, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 559 + ], + "type": "text", + "content": "so the parabola's vertex lies at " + }, + { + "bbox": [ + 46, + 498, + 288, + 559 + ], + "type": "inline_equation", + "content": "T / 2" + }, + { + "bbox": [ + 46, + 498, + 288, + 559 + ], + "type": "text", + "content": ". We readily obtain the PSF by applying the parabolic integration to a delta input. Upon deconvolution using the PSF, a parabolic projection provides the optimal SNR for a blur-free image from single capture when only the direction of velocity is known." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 574, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 287, + 658 + ], + "type": "text", + "content": "Ensembling linear projections. Finally, we leverage the flexibility of photon-cubes to compute multiple linear projections, as seen in Fig. 6. This produces a stack of images where one velocity group is motion-blur free at a time—or a 'motion stack', analogous to a focal stack. This novel construct can be used to compensate motion by blending stack images using cues such as blur orientation or optical flow." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 670, + 250, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 250, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 250, + 683 + ], + "type": "text", + "content": "6. Hardware and Experimental Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "We design a range of experiments to demonstrate the versatility of photon-cube projections: both when computa" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 498, + 545, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 584 + ], + "type": "text", + "content": "tions occur after readout (Secs. 6.1 and 6.2), and when they are performed near-sensor on-chip (Sec. 6.3). All photon-cubes were acquired using the SwissSPAD2 array [67], operated using one of two sub-arrays, each having " + }, + { + "bbox": [ + 304, + 498, + 545, + 584 + ], + "type": "inline_equation", + "content": "512 \\times 256" + }, + { + "bbox": [ + 304, + 498, + 545, + 584 + ], + "type": "text", + "content": " pixels, and at a frame-rate of " + }, + { + "bbox": [ + 304, + 498, + 545, + 584 + ], + "type": "inline_equation", + "content": "96.8\\mathrm{kHz}" + }, + { + "bbox": [ + 304, + 498, + 545, + 584 + ], + "type": "text", + "content": ". For the on-chip experiments, we use the UltraPhase compute architecture to interface with photon-cubes acquired by the SwissSPAD2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 597, + 435, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 435, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 435, + 609 + ], + "type": "text", + "content": "6.1. SoDaCam Capabilities" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "High-speed compressive imaging. We reconstruct 80 frames from compressive snapshots that are emulated at " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "25\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ", resulting in a 2000 FPS video. We decode compressive snapshots using a plug-and-play (PnP) approach, PnP-FastDVDNet [73]. As Fig. 7 shows, it is challenging to recover a large number of frames from a single compressive measurement. Using the proposed multi-bucket scheme significantly improves the quality of video reconstruction." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "8171" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 99, + 110, + 141 + ], + "blocks": [ + { + "bbox": [ + 56, + 76, + 124, + 98 + ], + "lines": [ + { + "bbox": [ + 56, + 76, + 124, + 98 + ], + "spans": [ + { + "bbox": [ + 56, + 76, + 124, + 98 + ], + "type": "text", + "content": "Sufficient Light 158 lux, 1.03 PPP" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 99, + 110, + 141 + ], + "lines": [ + { + "bbox": [ + 67, + 99, + 110, + 141 + ], + "spans": [ + { + "bbox": [ + 67, + 99, + 110, + 141 + ], + "type": "image", + "image_path": "59f8080ba282867c4d22838a60bcb7bf1ec710645f571ad142bc47a111e1f536.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 146, + 99, + 189, + 141 + ], + "blocks": [ + { + "bbox": [ + 134, + 76, + 197, + 97 + ], + "lines": [ + { + "bbox": [ + 134, + 76, + 197, + 97 + ], + "spans": [ + { + "bbox": [ + 134, + 76, + 197, + 97 + ], + "type": "text", + "content": "Low light 0.9 lux, 0.1 PPP" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 146, + 99, + 189, + 141 + ], + "lines": [ + { + "bbox": [ + 146, + 99, + 189, + 141 + ], + "spans": [ + { + "bbox": [ + 146, + 99, + 189, + 141 + ], + "type": "image", + "image_path": "88404c8016c9e7b8df7a345335bc5cddce2a9ea9cda51452cafd8b65fcf20dc6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 223, + 99, + 266, + 140 + ], + "blocks": [ + { + "bbox": [ + 212, + 76, + 280, + 98 + ], + "lines": [ + { + "bbox": [ + 212, + 76, + 280, + 98 + ], + "spans": [ + { + "bbox": [ + 212, + 76, + 280, + 98 + ], + "type": "text", + "content": "Very low light 0.3 lux, 0.03 PPP" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 223, + 99, + 266, + 140 + ], + "lines": [ + { + "bbox": [ + 223, + 99, + 266, + 140 + ], + "spans": [ + { + "bbox": [ + 223, + 99, + 266, + 140 + ], + "type": "image", + "image_path": "2e3dd248446e1dcf04390d0feac2a6f77fedf801d07f7225b49927d03b5adc91.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 53, + 163, + 126, + 236 + ], + "blocks": [ + { + "bbox": [ + 53, + 163, + 126, + 236 + ], + "lines": [ + { + "bbox": [ + 53, + 163, + 126, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 126, + 236 + ], + "type": "image", + "image_path": "3d6d717eafad9874614643283bd060c59b7b2269c31598f334f5ac790798c780.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 240, + 214, + 251 + ], + "lines": [ + { + "bbox": [ + 121, + 240, + 214, + 251 + ], + "spans": [ + { + "bbox": [ + 121, + 240, + 214, + 251 + ], + "type": "text", + "content": "SPAD-emulated events" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 130, + 163, + 203, + 236 + ], + "blocks": [ + { + "bbox": [ + 123, + 144, + 212, + 156 + ], + "lines": [ + { + "bbox": [ + 123, + 144, + 212, + 156 + ], + "spans": [ + { + "bbox": [ + 123, + 144, + 212, + 156 + ], + "type": "text", + "content": "DSLR captures (30 Hz)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 163, + 203, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 163, + 203, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 163, + 203, + 236 + ], + "type": "image", + "image_path": "f5aed45fe02cb7294012691749a1d12badfcc30b39ba6065ce34022d30d3b17c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 211, + 163, + 282, + 236 + ], + "blocks": [ + { + "bbox": [ + 211, + 163, + 282, + 236 + ], + "lines": [ + { + "bbox": [ + 211, + 163, + 282, + 236 + ], + "spans": [ + { + "bbox": [ + 211, + 163, + 282, + 236 + ], + "type": "image", + "image_path": "7aaf9f1ac84da73d2749050a24ba8720f6f533a418b11100f73bd01bd4526604.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 53, + 258, + 126, + 331 + ], + "blocks": [ + { + "bbox": [ + 53, + 258, + 126, + 331 + ], + "lines": [ + { + "bbox": [ + 53, + 258, + 126, + 331 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 126, + 331 + ], + "type": "image", + "image_path": "51a862585f198d34166b38a0ca7ba90c408291b25d248fa3ada9aa1b48592835.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 335, + 216, + 346 + ], + "lines": [ + { + "bbox": [ + 119, + 335, + 216, + 346 + ], + "spans": [ + { + "bbox": [ + 119, + 335, + 216, + 346 + ], + "type": "text", + "content": "Prophesee EVK4 events" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 130, + 258, + 203, + 331 + ], + "blocks": [ + { + "bbox": [ + 130, + 258, + 203, + 331 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 203, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 203, + 331 + ], + "type": "image", + "image_path": "278f1ad503ec66b55ab202dfac536d76fedc1977f137bc3f23a353c99bdc5570.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 355, + 287, + 451 + ], + "lines": [ + { + "bbox": [ + 46, + 355, + 287, + 451 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 287, + 451 + ], + "type": "text", + "content": "Figure 9: Comparison to a state-of-the-art event camera. SPAD-events can capture temporal gradients even when the light-level is reduced by " + }, + { + "bbox": [ + 46, + 355, + 287, + 451 + ], + "type": "inline_equation", + "content": "500\\times" + }, + { + "bbox": [ + 46, + 355, + 287, + 451 + ], + "type": "text", + "content": ", by benefiting from their single-photon sensitivity and bounded brightness response curve. In contrast, low-light induces blur and deteriorates the Prophesee's event stream. As a measure of the light-level, we report the PPP (photons per pixel) averaged across bit-planes and a light meter's reading at the sensor location." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 211, + 258, + 282, + 331 + ], + "blocks": [ + { + "bbox": [ + 211, + 258, + 282, + 331 + ], + "lines": [ + { + "bbox": [ + 211, + 258, + 282, + 331 + ], + "spans": [ + { + "bbox": [ + 211, + 258, + 282, + 331 + ], + "type": "image", + "image_path": "7f31fb95ba25496365b8b6c3062f4c2c2648bb3887b2ca3b0f8f9987def73c73.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 306, + 546, + 402 + ], + "lines": [ + { + "bbox": [ + 305, + 306, + 546, + 402 + ], + "spans": [ + { + "bbox": [ + 305, + 306, + 546, + 402 + ], + "type": "text", + "content": "Figure 10: Comparison against conventional high-speed acquisition at " + }, + { + "bbox": [ + 305, + 306, + 546, + 402 + ], + "type": "inline_equation", + "content": "4000\\mathrm{Hz}" + }, + { + "bbox": [ + 305, + 306, + 546, + 402 + ], + "type": "text", + "content": ". (top) SPAD projections recover a " + }, + { + "bbox": [ + 305, + 306, + 546, + 402 + ], + "type": "inline_equation", + "content": "16\\times" + }, + { + "bbox": [ + 305, + 306, + 546, + 402 + ], + "type": "text", + "content": " compressive video and an event image of a spinning roulette wheel. (middle) Read-noise corrupts the incident flux in the Infinicam high-speed camera, removing details in frames which are compressed on-the-fly. (bottom) Although using a larger aperture to admit more light recovers some detail, noise and compression artifacts still persist." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 474, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 510 + ], + "type": "text", + "content": "While multi-bucket captures require more bandwidth, this can be partially amortized by coding only dynamic regions, which we show in Supplementary Note 1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 528, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 287, + 660 + ], + "type": "text", + "content": "Motion projections on a traffic scene. Fig. 8 shows two traffic scenes captured using a " + }, + { + "bbox": [ + 46, + 528, + 287, + 660 + ], + "type": "inline_equation", + "content": "50~\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 528, + 287, + 660 + ], + "type": "text", + "content": " focal length lens and at " + }, + { + "bbox": [ + 46, + 528, + 287, + 660 + ], + "type": "inline_equation", + "content": "30\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 528, + 287, + 660 + ], + "type": "text", + "content": " emulation. When object velocity is known, a linear projection can make moving objects appear stationary. If only the velocity direction is known (e.g., road's orientation in Fig. 8), a parabolic projection provides a sharp reconstruction of all objects. We deblur parabolic captures using PnP-DnCNN [75]. We offer an improvement by randomly sampling 8 linear projections along the velocity direction and blending them using the optical flow predicted by RAFT [65] between two short-exposures." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Low-light event imaging. Fig. 9 compares event-image visualizations of SPAD and that of a state-of-the-art commercial event sensor (Prophesee EVK4), across various" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 311, + 86, + 387, + 140 + ], + "blocks": [ + { + "bbox": [ + 312, + 76, + 385, + 84 + ], + "lines": [ + { + "bbox": [ + 312, + 76, + 385, + 84 + ], + "spans": [ + { + "bbox": [ + 312, + 76, + 385, + 84 + ], + "type": "text", + "content": "Compressive frame" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 311, + 86, + 387, + 140 + ], + "lines": [ + { + "bbox": [ + 311, + 86, + 387, + 140 + ], + "spans": [ + { + "bbox": [ + 311, + 86, + 387, + 140 + ], + "type": "image", + "image_path": "85c7f483ef5ff3d49b0a9ca1ef3518a6f86d20b34b7f4a18599f97c31eb0fbfa.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 390, + 86, + 466, + 140 + ], + "blocks": [ + { + "bbox": [ + 396, + 76, + 459, + 84 + ], + "lines": [ + { + "bbox": [ + 396, + 76, + 459, + 84 + ], + "spans": [ + { + "bbox": [ + 396, + 76, + 459, + 84 + ], + "type": "text", + "content": "Recovered frame" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 390, + 86, + 466, + 140 + ], + "lines": [ + { + "bbox": [ + 390, + 86, + 466, + 140 + ], + "spans": [ + { + "bbox": [ + 390, + 86, + 466, + 140 + ], + "type": "image", + "image_path": "68caba06bceae84361f249d39cf5f17d9ffc27556e84ddeb3858dbe92f168567.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 469, + 87, + 545, + 140 + ], + "blocks": [ + { + "bbox": [ + 484, + 76, + 530, + 85 + ], + "lines": [ + { + "bbox": [ + 484, + 76, + 530, + 85 + ], + "spans": [ + { + "bbox": [ + 484, + 76, + 530, + 85 + ], + "type": "text", + "content": "Event image" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 87, + 545, + 140 + ], + "lines": [ + { + "bbox": [ + 469, + 87, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 469, + 87, + 545, + 140 + ], + "type": "image", + "image_path": "859ef4cd2e6e9b055864c9f0db9b5ac5bfbab380bdf1043f5349407b1f08122c.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 142, + 473, + 152 + ], + "lines": [ + { + "bbox": [ + 384, + 142, + 473, + 152 + ], + "spans": [ + { + "bbox": [ + 384, + 142, + 473, + 152 + ], + "type": "text", + "content": "SwissSPAD2 projections" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 311, + 158, + 386, + 213 + ], + "blocks": [ + { + "bbox": [ + 311, + 158, + 386, + 213 + ], + "lines": [ + { + "bbox": [ + 311, + 158, + 386, + 213 + ], + "spans": [ + { + "bbox": [ + 311, + 158, + 386, + 213 + ], + "type": "image", + "image_path": "a38bfe0d98ac5a3b148c42422f35949a0dec357730c02c225786bf3bfc442759.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 390, + 158, + 466, + 213 + ], + "blocks": [ + { + "bbox": [ + 390, + 158, + 466, + 213 + ], + "lines": [ + { + "bbox": [ + 390, + 158, + 466, + 213 + ], + "spans": [ + { + "bbox": [ + 390, + 158, + 466, + 213 + ], + "type": "image", + "image_path": "056222d39ccb53124b482ec6010df2cf88ee1633d237d1d959147f7cf6642679.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 214, + 468, + 225 + ], + "lines": [ + { + "bbox": [ + 389, + 214, + 468, + 225 + ], + "spans": [ + { + "bbox": [ + 389, + 214, + 468, + 225 + ], + "type": "text", + "content": "Infinicam projections" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 470, + 159, + 545, + 213 + ], + "blocks": [ + { + "bbox": [ + 470, + 159, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 470, + 159, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 470, + 159, + 545, + 213 + ], + "type": "image", + "image_path": "e925881cb7d9e2b08ded24c1094f58849cb7f14efae3f5031fea13ee6d6d9954.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 311, + 231, + 386, + 285 + ], + "blocks": [ + { + "bbox": [ + 311, + 231, + 386, + 285 + ], + "lines": [ + { + "bbox": [ + 311, + 231, + 386, + 285 + ], + "spans": [ + { + "bbox": [ + 311, + 231, + 386, + 285 + ], + "type": "image", + "image_path": "52001fca6b9e05c38e133049fb6968a1865da47d1dd78376857ab2acdcc46ade.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 287, + 500, + 297 + ], + "lines": [ + { + "bbox": [ + 356, + 287, + 500, + 297 + ], + "spans": [ + { + "bbox": [ + 356, + 287, + 500, + 297 + ], + "type": "text", + "content": "Infinicam projections (16× more light)" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 390, + 231, + 466, + 285 + ], + "blocks": [ + { + "bbox": [ + 390, + 231, + 466, + 285 + ], + "lines": [ + { + "bbox": [ + 390, + 231, + 466, + 285 + ], + "spans": [ + { + "bbox": [ + 390, + 231, + 466, + 285 + ], + "type": "image", + "image_path": "907f2d1c8ff784ad88f54f3683feb65f3d6169d52a504b789e6879a94aecae98.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 470, + 231, + 545, + 285 + ], + "blocks": [ + { + "bbox": [ + 470, + 231, + 545, + 285 + ], + "lines": [ + { + "bbox": [ + 470, + 231, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 470, + 231, + 545, + 285 + ], + "type": "image", + "image_path": "8b5c82a105fcdde8453fbb76374f86659d472e4651b4ef7f960369e67b1d5c1e.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 423, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 555 + ], + "type": "text", + "content": "light levels, with an accumulation period of " + }, + { + "bbox": [ + 304, + 423, + 545, + 555 + ], + "type": "inline_equation", + "content": "33\\mathrm{ms}" + }, + { + "bbox": [ + 304, + 423, + 545, + 555 + ], + "type": "text", + "content": ". For a fair comparison, we bin the Prophesee's events in blocks of " + }, + { + "bbox": [ + 304, + 423, + 545, + 555 + ], + "type": "inline_equation", + "content": "2\\times 2" + }, + { + "bbox": [ + 304, + 423, + 545, + 555 + ], + "type": "text", + "content": " pixels and use a smaller aperture to account for the lower fill-factor of the SPAD. We tuned event-generation parameters (contrast threshold, integrator decay rate) of both cameras at each light level. Low light induces blur and deteriorates the Prophesee's event stream. In contrast, SPAD-events continue to capture temporal gradients, due to the SPAD's low-light capabilities and its brightness-encoding response curve. We include an ablative study of brightness-encoding functions in Supplementary Note 2." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 555, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 604 + ], + "type": "text", + "content": "Our observations are in concurrence with recent works that examine the low-light performance of event cameras [19, 27], and show that SPAD-events can provide neuromorphic vision in these challenging-SNR scenarios." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 306, + 611, + 500, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 500, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 500, + 624 + ], + "type": "text", + "content": "6.2. Comparison to High-Speed Cameras" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Recall, as previously discussed in Sec. 4, that read-noise limits the per-frame SNR of high-speed cameras. To demonstrate this limitation, we compute projections using the " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "4\\mathrm{kHz}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " acquisition of the Photon Infinicam, a conventional high-speed camera, at a resolution of " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "1246\\times 240" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " pixels. We operate the SwissSPAD2 and the Infinicam at ambient light conditions using the same lens specifications." + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8172" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 123, + 156 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 123, + 156 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 123, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 123, + 156 + ], + "type": "image", + "image_path": "eb1db41027ec422ffa4fcf6c16153765de028fa94fa48c064e8b750ec88e2fce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 159, + 123, + 169 + ], + "lines": [ + { + "bbox": [ + 54, + 159, + 123, + 169 + ], + "spans": [ + { + "bbox": [ + 54, + 159, + 123, + 169 + ], + "type": "text", + "content": "UltraPhase board" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 127, + 70, + 282, + 157 + ], + "blocks": [ + { + "bbox": [ + 127, + 70, + 282, + 157 + ], + "lines": [ + { + "bbox": [ + 127, + 70, + 282, + 157 + ], + "spans": [ + { + "bbox": [ + 127, + 70, + 282, + 157 + ], + "type": "image", + "image_path": "86f53554ad579b4c3582097706391466bfcd1ae17199cd42f715cb10148a4149.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 159, + 276, + 169 + ], + "lines": [ + { + "bbox": [ + 135, + 159, + 276, + 169 + ], + "spans": [ + { + "bbox": [ + 135, + 159, + 276, + 169 + ], + "type": "text", + "content": "Power and bandwidth measurement" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 182, + 287, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 287, + 255 + ], + "type": "text", + "content": "Figure 11: Power and bandwidth requirements when computing photon-cube projections on UltraPhase [5] (left), a recent compute architecture designed for single-photon imaging, at " + }, + { + "bbox": [ + 46, + 182, + 287, + 255 + ], + "type": "inline_equation", + "content": "40\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 182, + 287, + 255 + ], + "type": "text", + "content": " readout. (right) Our projections act as a compression scheme for photon-cubes, resulting in dramatically reduced sensor-readout and power consumption." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 274, + 287, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 287, + 382 + ], + "type": "text", + "content": "As Fig. 10 shows, read noise corrupts the incident signal in Infinicam and makes it impossible to derive any useful projections. The read noise could be averaged out to some extent if the Infinicam did not perform compression-on-the-fly, but compression is central to the camera's working and enables readout over USB. Using a larger aperture to admit more light improves the quality of computed projections, but the video reconstruction and event image remain considerably worse than corresponding outputs of the SPAD." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 392, + 235, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 235, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 235, + 406 + ], + "type": "text", + "content": "6.3. Bandwidth and Power Implications" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 412, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 532 + ], + "type": "text", + "content": "While Sec. 6.1 has demonstrated the capabilities of photon-cube projections, we now show that our projections can also be obtained in a bandwidth-efficient manner via near-sensor computations. We implement photon-cube projections on UltraPhase (Fig. 11 (left)), a novel compute architecture designed for single-photon imaging. UltraPhase consists of " + }, + { + "bbox": [ + 46, + 412, + 287, + 532 + ], + "type": "inline_equation", + "content": "3 \\times 6" + }, + { + "bbox": [ + 46, + 412, + 287, + 532 + ], + "type": "text", + "content": " processing cores, each of which interfaces with " + }, + { + "bbox": [ + 46, + 412, + 287, + 532 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 46, + 412, + 287, + 532 + ], + "type": "text", + "content": " pixels, and can be 3D stacked beneath a SPAD array. We include visualizations and programming details of a few example projections in Supplementary Note 5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 533, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 676 + ], + "type": "text", + "content": "We measure the readout and power consumption of UltraPhase when computing projections on 2500 bit-planes of the falling die sequence (Fig. 1). The projections include: VCS with 16 random binary masks, an event camera, a linear projection and a combination of the three. We output projections at 12-bit depth and calculate metrics based on the clock cycles required for both compute and readout. As seen in Fig. 11 (right), computing projections on-chip dramatically reduces sensor-readout and power consumption as compared to reading out the photon-cube. Finally, similar to existing event cameras, SPAD-events have a resource footprint that reflects the underlying scene dynamics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "In summary, our on-chip experiments show that performing computations near-sensor can increase the viability of single-photon imaging in resource constrained settings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 71, + 480, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 480, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 480, + 84 + ], + "type": "text", + "content": "7. Discussion and Future Outlook" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 91, + 545, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 248 + ], + "type": "text", + "content": "SoDaCam provides a realization of reinterpretable software-defined cameras [2, 3, 20, 34, 51] at the fine temporal resolution of SPAD-acquired photon-cubes. The proposed computations, or photon-cube projections, can match and in some cases, surpass the capabilities of existing imaging systems. The software-defined nature of photon-cube projections provides functionalities that may be difficult to achieve in conventional sensors. These projections can reduce the readout and power-consumption of SPAD arrays and potentially spur widespread adoption of single-photon imaging in the consumer domain. Finally, future chip-to-chip communication standards may also make it feasible to compute projections on a camera image signal processor." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 265, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 545, + 373 + ], + "type": "text", + "content": "Adding color to SoDaCam. One way to add color is by overlaying color filter arrays (CFAs) and perform demosaicing on the computed photon-cube projection: depending on the projection, demosaicing could be relatively simple or more complex. As a reference, Bayer CFAs have been considered in the context of both video compressive sensing [73] and event cameras [64]. Incorporating CFAs with motion projections requires careful considerations, e.g., avoiding integrating across pixel locations of differing color." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 390, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 498 + ], + "type": "text", + "content": "Future outlook on SPAD characteristics. A key SPAD characteristic that determines several properties of emulated cameras is the frame rate. While no fundamental limitations prevent SPADs from being operated at the frame-rates utilized in this work (" + }, + { + "bbox": [ + 304, + 390, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\sim 100\\mathrm{kHz}" + }, + { + "bbox": [ + 304, + 390, + 545, + 498 + ], + "type": "text", + "content": "), sensor readout and power constraints can preclude high speeds, especially in high-resolution SPAD arrays. Photon-cube projections can enable future large-format SPADs to preserve high-speed information with modest resource requirements." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 516, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 545, + 612 + ], + "type": "text", + "content": "A platform for comparing cameras. Comparing imaging modalities can be quite challenging since hardware realizations of sensors can differ in numerous aspects, such as their quantum efficiency, fill factor, pixel pitch and array resolution. By emulating their imaging models, SoDaCam can serve as a platform for hardware-agnostic comparisons; for instance, determining operating conditions where one imaging modality is advantageous over another." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "A Cambrian explosion of new cameras. Besides comparing cameras, by virtue of being software-defined, SoDa-Cam can also make it significantly easier to prototype and deploy new unconventional imaging models, and even facilitate sensor-in-the-loop optimization [44, 47, 63] by tailoring photon-cube projections for downstream computer-vision tasks. This is an exciting future line of research." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8173" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 704 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Phantom-v2640. https://www.phantomhighspeed.com/products/cameraul ultrahigh4mpx/v2640. Accessed: 2023-01-28. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 156 + ], + "type": "text", + "content": "[2] E. H. Adelson, J. R. Bergen, et al. The plenoptic function and the elements of early vision. Computational models of visual processing, 1(2):3-20, 1991. 1, 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 157, + 287, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 157, + 287, + 210 + ], + "spans": [ + { + "bbox": [ + 54, + 157, + 287, + 210 + ], + "type": "text", + "content": "[3] A. Agrawal, A. Veeraraghavan, and R. Raskar. Reinterpretable imager: Towards variable post-capture space, angle and time resolution in photography. In Computer Graphics Forum, volume 29, pages 763-772. Wiley Online Library, 2010. 3, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 212, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 212, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 54, + 212, + 287, + 255 + ], + "type": "text", + "content": "[4] I. M. Antolovic, S. Burri, C. Bruschini, R. Hoebe, and E. Charbon. Nonuniformity analysis of a 65-kpixel CMOS SPAD imager. IEEE Transactions on Electron Devices, 63 (1):57-64, 2016. doi: 10.1109/TED.2015.2458295. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 256, + 287, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 256, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 53, + 256, + 287, + 277 + ], + "type": "text", + "content": "[5] A. Ardelean. Computational Imaging SPAD Cameras. PhD thesis, École polytechnique fédérale de Lausanne, 2023. 2, 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 278, + 286, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 278, + 286, + 309 + ], + "spans": [ + { + "bbox": [ + 53, + 278, + 286, + 309 + ], + "type": "text", + "content": "[6] A. Boukhayma, A. Peizerat, and C. Enz. A sub-0.5 electron read noise VGA image sensor in a standard CMOS process. IEEE Journal of Solid-State Circuits, 2016. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 311, + 286, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 311, + 286, + 353 + ], + "spans": [ + { + "bbox": [ + 54, + 311, + 286, + 353 + ], + "type": "text", + "content": "[7] J. Canny. A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-8(6):679-698, 1986. doi: 10.1109/TPAMI.1986.4767851. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 354, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 287, + 397 + ], + "type": "text", + "content": "[8] S. J. Carey, A. Lopich, D. R. Barr, B. Wang, and P. Dudek. A 100,000 fps vision sensor with embedded 535GOPS/W " + }, + { + "bbox": [ + 53, + 354, + 287, + 397 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 53, + 354, + 287, + 397 + ], + "type": "text", + "content": " SIMD processor array. In 2013 Symposium on VLSI Circuits, pages C182-C183, 2013. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 399, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 287, + 441 + ], + "type": "text", + "content": "[9] J. Chen, S. J. Carey, and P. Dudek. Feature extraction using a portable vision system. In IEEE/RSJ Int. Conf. Intell. Robots Syst., Workshop Vis.-based Agile Auton. Navigation UAVs, volume 2, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 442, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 442, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 442, + 287, + 485 + ], + "type": "text", + "content": "[10] S. Chen and M. Guo. Live demonstration: Celex-v: A 1m pixel multi-mode event-based sensor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2019. 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 486, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 486, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 486, + 287, + 529 + ], + "type": "text", + "content": "[11] N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), volume 1, pages 886-893. IEEE, 2005. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 529, + 287, + 604 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 529, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 48, + 529, + 287, + 604 + ], + "type": "text", + "content": "[12] P. E. Debevec and J. Malik. Recovering high dynamic range radiance maps from photographs. In Proceedings of the 24th Annual Conference on Computer Graphics and Interactive Techniques, page 369-378. ACM Press/Addison-Wesley Publishing Co., 1997. doi: 10.1145/258734.258884. URL https://doi.org/10.1145/258734.258884. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 606, + 287, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 606, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 48, + 606, + 287, + 670 + ], + "type": "text", + "content": "[13] F. M. Della Rocca, H. Mai, S. W. Hutchings, T. Al Abbas, K. Buckbee, A. Tsiamis, P. Lomax, I. Gyongy, N. A. Dutton, and R. K. Henderson. A " + }, + { + "bbox": [ + 48, + 606, + 287, + 670 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 48, + 606, + 287, + 670 + ], + "type": "text", + "content": " SPAD motion-triggered time-of-flight image sensor with in-pixel histogram and column-parallel vision processor. IEEE Journal of Solid-State Circuits, 55(7):1762-1775, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 672, + 287, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 672, + 287, + 704 + ], + "spans": [ + { + "bbox": [ + 48, + 672, + 287, + 704 + ], + "type": "text", + "content": "[14] T. Finateu, A. Niwa, D. Matolin, K. Tsuchimoto, A. Mascheroni, E. Reynaud, P. Mostafalu, F. Brady, L. Chotard, F. LeGoff, H. Takahashi, H. Wakabayashi," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 708 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 148 + ], + "type": "text", + "content": "Y. Oike, and C. Posch. 5.10 a " + }, + { + "bbox": [ + 326, + 73, + 545, + 148 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 326, + 73, + 545, + 148 + ], + "type": "text", + "content": " back-illuminated stacked temporal contrast event-based vision sensor with " + }, + { + "bbox": [ + 326, + 73, + 545, + 148 + ], + "type": "inline_equation", + "content": "4.86\\mu m" + }, + { + "bbox": [ + 326, + 73, + 545, + 148 + ], + "type": "text", + "content": " pixels, 1.066GEPS readout, programmable event-rate controller and compressive data-formating pipeline. In 2020 IEEE International Solid-State Circuits Conference - (ISSCC), pages 112-114, 2020. doi: 10.1109/ISSCC19947.2020.9063149. 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 150, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 193 + ], + "type": "text", + "content": "[15] E. R. Fossum. What to do with sub-diffraction-limit (SDL) pixels?—a proposal for a gigapixel digital film sensor (DFS). In IEEE Workshop on Charge-Coupled Devices and Advanced Image Sensors, pages 214–217, 2005. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 194, + 545, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 194, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 308, + 194, + 545, + 247 + ], + "type": "text", + "content": "[16] E. R. Fossum. The quanta image sensor (QIS): Concepts and challenges. In Imaging and Applied Optics, page JTuE1. Optica Publishing Group, 2011. doi: 10.1364/COSI.2011. JTuE1. URL http://opg.optica.org/abstract.cfm?URI=COSI-2011-JTuE1.1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 249, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 249, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 249, + 545, + 281 + ], + "type": "text", + "content": "[17] E. R. Fossum, J. Ma, S. Masoodian, L. Anzagira, and R. Zizza. The quanta image sensor: Every photon counts. Sensors, 16(8):1260, 2016. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 281, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 281, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 308, + 281, + 545, + 323 + ], + "type": "text", + "content": "[18] D. Gehrig, H. Rebecq, G. Gallego, and D. Scaramuzza. Eklt: Asynchronous photometric feature tracking using events and frames. International Journal of Computer Vision, 128(3): 601-618, 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 326, + 545, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 326, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 308, + 326, + 545, + 357 + ], + "type": "text", + "content": "[19] R. Graca and T. Delbruck. Unraveling the paradox of intensity-dependent DVS pixel noise. arXiv preprint arXiv:2109.08640, 2021. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 358, + 545, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 358, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 308, + 358, + 545, + 423 + ], + "type": "text", + "content": "[20] M. Gupta, A. Agrawal, A. Veeraraghavan, and S. G. Narasimhan. Flexible voxels for motion-aware videography. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part I 11, pages 100-114. Springer, 2010. 3, 9" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 424, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 424, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 424, + 545, + 477 + ], + "type": "text", + "content": "[21] F. Gutierrez-Barragan, A. Ingle, T. Seets, M. Gupta, and A. Velten. Compressive single-photon 3D cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17854-17864, June 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 479, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 545, + 510 + ], + "type": "text", + "content": "[22] I. Gyongy, N. A. Dutton, and R. K. Henderson. Single-photon tracking for high-speed vision. Sensors, 18(2):323, 2018. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 511, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 545, + 555 + ], + "type": "text", + "content": "[23] I. Gyongy, S. W. Hutchings, A. Halimi, M. Tyler, S. Chan, F. Zhu, S. McLaughlin, R. K. Henderson, and J. Leach. High-speed 3D sensing via hybrid-mode imaging and guided upsampling. Optica, 7(10):1253-1260, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 555, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 555, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 555, + 545, + 586 + ], + "type": "text", + "content": "[24] C. Harris, M. Stephens, et al. A combined corner and edge detector. In *Alvey vision conference*, volume 15, pages 10-5244. CiteSeer, 1988. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 588, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 588, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 588, + 545, + 632 + ], + "type": "text", + "content": "[25] J. Hidalgo-Carrió, G. Gallego, and D. Scaramuzza. Event-aided direct sparse odometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5781-5790, June 2022. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 632, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 632, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 308, + 632, + 545, + 654 + ], + "type": "text", + "content": "[26] B. K. Horn and B. G. Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185-203, 1981. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 654, + 545, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 654, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 308, + 654, + 545, + 696 + ], + "type": "text", + "content": "[27] Y. Hu, S.-C. Liu, and T. Delbruck. v2e: From video frames to realistic DVS events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1312-1321, June 2021. 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 698, + 545, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 698, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 308, + 698, + 545, + 708 + ], + "type": "text", + "content": "[28] J. Igual. Photographic noise performance measures based on" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8174" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 709 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "raw files analysis of consumer cameras. Electronics, 8(11): 1284, 2019. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 288, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 288, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 288, + 138 + ], + "type": "text", + "content": "[29] A. Ingle, A. Velten, and M. Gupta. High Flux Passive Imaging With Single-Photon Sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "text", + "content": "[30] A. Ingle, T. Seets, M. Buttafava, S. Gupta, A. Tosi, M. Gupta, and A. Velten. Passive inter-photon imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 225 + ], + "type": "text", + "content": "[31] K. Iwabuchi, Y. Kameda, and T. Hamamoto. Image quality improvements based on motion-based deblurring for single-photon imaging. IEEE Access, 9:30080-30094, 2021. doi: 10.1109/ACCESS.2021.3059293. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 227, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 227, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 48, + 227, + 287, + 258 + ], + "type": "text", + "content": "[32] Y. Jiang, I. Choi, J. Jiang, and J. Gu. HDR video reconstruction with tri-exposure quad-bayer sensors. arXiv preprint arXiv:2103.10982, 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 259, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 259, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 259, + 287, + 293 + ], + "type": "text", + "content": "[33] A. Levin, P. Sand, T. S. Cho, F. Durand, and W. T. Freeman. Motion-invariant photography. ACM Transactions on Graphics (TOG), 27(3):1-9, 2008. 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 293, + 287, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 325 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 325 + ], + "type": "text", + "content": "[34] M. Levoy and P. Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 31-42, 1996. 1, 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 326, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 326, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 48, + 326, + 287, + 390 + ], + "type": "text", + "content": "[35] Y. Li, M. Qi, R. Gulve, M. Wei, R. Genov, K. N. Kutulakos, and W. Heidrich. End-to-end video compressive sensing using anderson-accelerated unrolled networks. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12, 2020. doi: 10.1109/ICCP48838.2020.9105237. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 391, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 391, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 48, + 391, + 287, + 423 + ], + "type": "text", + "content": "[36] P. Lichtsteiner. 64x64 event-driven logarithmic temporal derivative silicon retina. In Program 2003 IEEE Workshop on CCD and AIS, 2003. 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 424, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 424, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 424, + 287, + 479 + ], + "type": "text", + "content": "[37] Y. Liu, F. Gutierrez-Barragan, A. Ingle, M. Gupta, and A. Velten. Single-photon camera guided extreme dynamic range imaging. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1575-1585, January 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "type": "text", + "content": "[38] P. Llull, X. Liao, X. Yuan, J. Yang, D. Kittle, L. Carin, G. Sapiro, and D. J. Brady. Coded aperture compressive temporal imaging. Opt. Express, 21(9):10526-10545, May 2013. doi: 10.1364/OE.21.010526. URL https://opg.optica.org/oe/abstract.cfm?URI=oe-21-9-10526.1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "text", + "content": "[39] B. D. Lucas and T. Kanade. An iterative image registration technique with an application to stereo vision. In *IJCAI'81: 7th international joint conference on Artificial intelligence*, volume 2, pages 674–679, 1981. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 589, + 287, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 654 + ], + "type": "text", + "content": "[40] J. Ma, S. Masoodian, D. A. Starkey, and E. R. Fossum. Photon-number-resolving megapixel image sensor at room temperature without avalanche gain. Optica, 4(12): 1474-1481, Dec 2017. doi: 10.1364/OPTICA.4.001474. URL http://www.osapublishing.org/optica/abstract.cfm?URI=optica-4-12-1474.3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 654, + 287, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 654, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 48, + 654, + 287, + 696 + ], + "type": "text", + "content": "[41] S. Ma, S. Gupta, A. C. Ulku, C. Bruschini, E. Charbon, and M. Gupta. Quanta burst photography. ACM Transactions on Graphics, 39(4):1-16, July 2020. ISSN 0730-0301, 1557-7368. 3, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 698, + 287, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 698, + 287, + 709 + ], + "spans": [ + { + "bbox": [ + 48, + 698, + 287, + 709 + ], + "type": "text", + "content": "[42] S. Ma, P. Mos, E. Charbon, and M. Gupta. Burst vi" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 709 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "sion using single-photon cameras. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5375-5385, January 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "type": "text", + "content": "[43] S. Mann and R. Picard. Beingundigital'with digital cameras. MIT Media Lab Perceptual, 1:2, 1994. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[44] J. N. Martel, L. K. Mueller, S. J. Carey, P. Dudek, and G. Wetzstein. Neural sensors: Learning pixel exposures for HDR imaging and video compressive sensing with programmable sensors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(7):1642-1653, 2020. 9" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "type": "text", + "content": "[45] J. N. P. Martel, L. K. Müller, S. J. Carey, and P. Dudek. Parallel HDR tone mapping and auto-focus on a cellular processor array vision chip. In 2016 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1430-1433, 2016. doi: 10.1109/ISCAS.2016.7527519. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 238, + 545, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 238, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 308, + 238, + 545, + 292 + ], + "type": "text", + "content": "[46] J. N. P. Martel, L. K. Müller, S. J. Carey, and P. Dudek. High-speed depth from focus on a programmable vision chip using a focus tunable lens. In 2017 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-4, 2017. doi: 10.1109/ISCAS.2017.8050548.3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 293, + 545, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 293, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 308, + 293, + 545, + 336 + ], + "type": "text", + "content": "[47] C. A. Metzler, H. Ikoma, Y. Peng, and G. Wetzstein. Deep optics for single-shot high-dynamic-range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 9" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 337, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 337, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 337, + 545, + 380 + ], + "type": "text", + "content": "[48] K. Morimoto, A. Ardelean, M.-L. Wu, A. C. Ulku, I. M. Antolovic, C. Bruschini, and E. Charbon. Megapixel time-gated SPAD image sensor for 2D and 3D imaging applications. Optica, 7(4):346-354, Apr. 2020. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 380, + 545, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 380, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 308, + 380, + 545, + 489 + ], + "type": "text", + "content": "[49] K. Morimoto, J. Iwata, M. Shinohara, H. Sekine, A. Abdelghafar, H. Tsuchiya, Y. Kuroda, K. Tojima, W. Endo, Y. Maehashi, Y. Ota, T. Sasago, S. Maekawa, S. Hikosaka, T. Kanou, A. Kato, T. Tezuka, S. Yoshizaki, T. Ogawa, K. Uehira, A. Ehara, F. Inui, Y. Matsuno, K. Sakurai, and T. Ichikawa. 3.2 megapixel 3D-stacked charge focusing SPAD for low-light imaging and depth sensing. In 2021 IEEE International Electron Devices Meeting (IEDM), pages 20.2.1-20.2.4, 2021. doi: 10.1109/IEDM19574.2021.9720605.1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "type": "text", + "content": "[50] S. Namiki, S. Sato, Y. Kameda, and T. Hamamoto. Imaging method using multi-threshold pattern for photon detection of quanta image sensor. In International Workshop on Advanced Imaging Technology (IWAIT) 2022, volume 12177, page 1217702. SPIE, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "type": "text", + "content": "[51] S. K. Nayar, V. Branzoi, and T. E. Boult. Programmable imaging: Towards a flexible camera. International Journal of Computer Vision, 70:7-22, 2006. 3, 9" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 578, + 545, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 545, + 631 + ], + "type": "text", + "content": "[52] C. Posch, D. Matolin, and R. Wohlgenannt. A QVGA 143 dB Dynamic Range Frame-Free PWM Image Sensor With Lossless Pixel-Level Video Compression and Time-Domain CDS. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2011. doi: 10.1109/JSSC.2010.2085952. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 632, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 632, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 308, + 632, + 545, + 666 + ], + "type": "text", + "content": "[53] R. Raskar, A. Agrawal, and J. Tumblin. Coded exposure photography: motion deblurring using fluttered shutter. In Acm Siggraph 2006 Papers, pages 795-804. 2006. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 666, + 545, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 666, + 545, + 709 + ], + "spans": [ + { + "bbox": [ + 308, + 666, + 545, + 709 + ], + "type": "text", + "content": "[54] H. Rebecq, R. Ranftl, V. Koltun, and D. Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8175" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 709 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[55] D. Reddy, A. Veeraraghavan, and R. Chellappa. P2C2: Programmable pixel compressive camera for high speed imaging. In CVPR 2011, pages 329-336, 2011. doi: 10.1109/CVPR.2011.5995542. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 288, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 288, + 170 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 288, + 170 + ], + "type": "text", + "content": "[56] X. Ren, P. W. Connolly, A. Halimi, Y. Altmann, S. McLaughlin, I. Gyongy, R. K. Henderson, and G. S. Buller. High-resolution depth profiling using a range-gated CMOS SPAD quanta image sensor. Optics express, 26(5):5541-5557, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 171, + 288, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 171, + 288, + 225 + ], + "spans": [ + { + "bbox": [ + 48, + 171, + 288, + 225 + ], + "type": "text", + "content": "[57] C. Scheerlinck, H. Rebecq, D. Gehrig, N. Barnes, R. Mahony, and D. Scaramuzza. Fast image reconstruction with an event camera. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), March 2020. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 227, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 227, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 48, + 227, + 288, + 280 + ], + "type": "text", + "content": "[58] T. Seets, A. Ingle, M. Laurenzis, and A. Velten. Motion adaptive deblurring with single-photon cameras. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1945-1954, January 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 281, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 281, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 281, + 288, + 357 + ], + "type": "text", + "content": "[59] M.-W. Seo, Y. Shirakawa, Y. Masuda, Y. Kawata, K. Kagawa, K. Yasutomi, and S. Kawahito. 4.3 a programmable sub-nanosecond time-gated 4-tap lock-in pixel cmos image sensor for real-time fluorescence lifetime imaging microscopy. In 2017 IEEE International Solid-State Circuits Conference (ISSCC), pages 70-71, 2017. doi: 10.1109/ISSCC.2017.7870265. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 357, + 288, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 357, + 288, + 412 + ], + "spans": [ + { + "bbox": [ + 48, + 357, + 288, + 412 + ], + "type": "text", + "content": "[60] T. Serrano-Gotarredona and B. Linares-Barranco. A " + }, + { + "bbox": [ + 48, + 357, + 288, + 412 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 48, + 357, + 288, + 412 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 48, + 357, + 288, + 412 + ], + "type": "text", + "content": " contrast sensitivity " + }, + { + "bbox": [ + 48, + 357, + 288, + 412 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 48, + 357, + 288, + 412 + ], + "type": "text", + "content": " FPN " + }, + { + "bbox": [ + 48, + 357, + 288, + 412 + ], + "type": "inline_equation", + "content": "3\\mu s" + }, + { + "bbox": [ + 48, + 357, + 288, + 412 + ], + "type": "text", + "content": " latency " + }, + { + "bbox": [ + 48, + 357, + 288, + 412 + ], + "type": "inline_equation", + "content": "4\\mathrm{mW}" + }, + { + "bbox": [ + 48, + 357, + 288, + 412 + ], + "type": "text", + "content": " asynchronous frame-free dynamic vision sensor using transimpedance preamplifiers. IEEE Journal of Solid-State Circuits, 48(3):827-838, 2013. 1, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 413, + 288, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 413, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 48, + 413, + 288, + 467 + ], + "type": "text", + "content": "[61] P. Shedligeri, A. S., and K. Mitra. A unified framework for compressive video recovery from coded exposure techniques. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1600-1609, January 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 468, + 288, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 288, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 288, + 499 + ], + "type": "text", + "content": "[62] C. Shi, N. Song, W. Li, Y. Li, B. Wei, H. Liu, and J. Jin. A review of event-based indoor positioning and navigation. 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 500, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 288, + 555 + ], + "type": "text", + "content": "[63] V. Sitzmann, S. Diamond, Y. Peng, X. Dun, S. Boyd, W. Heidrich, F. Heide, and G. Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and super-resolution imaging. ACM Transactions on Graphics (TOG), 37(4):114, 2018. 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 555, + 288, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 288, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 288, + 620 + ], + "type": "text", + "content": "[64] G. Taverni, D. Paul Moeys, C. Li, C. Cavaco, V. Motsnyi, D. San Segundo Bello, and T. Delbruck. Front and back illuminated dynamic and active pixel vision sensors comparison. IEEE Transactions on Circuits and Systems II: Express Briefs, 65(5):677-681, 2018. doi: 10.1109/TCSII.2018.2824899. 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 621, + 288, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 621, + 288, + 653 + ], + "spans": [ + { + "bbox": [ + 48, + 621, + 288, + 653 + ], + "type": "text", + "content": "[65] Z. Teed and J. Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European Conference on Computer Vision, 2020. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 654, + 288, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 654, + 288, + 709 + ], + "spans": [ + { + "bbox": [ + 48, + 654, + 288, + 709 + ], + "type": "text", + "content": "[66] S. Tulyakov, D. Gehrig, S. Georgoulis, J. Erbach, M. Gehrig, Y. Li, and D. Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16155-16164, June 2021. 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 696 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[67] A. C. Ulku, C. Bruschini, I. M. Antolovic, Y. Kuo, R. Ankri, S. Weiss, X. Michalet, and E. Charbon. A " + }, + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": " SPAD Image Sensor With Integrated Gating for Widefield FLIM. IEEE Journal of Selected Topics in Quantum Electronics, 25 (1):1-12, Jan. 2019. ISSN 1077-260X, 1558-4542. doi: 10. 1109/JSTQE.2018.2867439. 3, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 140, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 182 + ], + "type": "text", + "content": "[68] G. Wan, X. Li, G. Agranov, M. Levoy, and M. Horowitz. CMOS image sensors with multi-bucket pixels for computational photography. IEEE Journal of Solid-State Circuits, 47 (4):1031-1042, 2012. doi: 10.1109/JSSC.2012.2185189. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 183, + 545, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 183, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 308, + 183, + 545, + 236 + ], + "type": "text", + "content": "[69] M. Wei, N. Sarhangnejad, Z. Xia, N. Gusev, N. Katic, R. Genov, and K. N. Kutulakos. Coded two-bucket cameras for computer vision. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 3, 4, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 237, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 237, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 308, + 237, + 545, + 303 + ], + "type": "text", + "content": "[70] M. White, S. Ghajari, T. Zhang, A. Dave, A. Veeraraghavan, and A. Molnar. A differential SPAD array architecture in " + }, + { + "bbox": [ + 308, + 237, + 545, + 303 + ], + "type": "inline_equation", + "content": "0.18\\mu \\mathrm{m}" + }, + { + "bbox": [ + 308, + 237, + 545, + 303 + ], + "type": "text", + "content": " CMOS for HDR imaging. In 2022 IEEE International Symposium on Circuits and Systems (ISCAS), pages 292-296, 2022. doi: 10.1109/ISCAS48785.2022.9937558.3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 303, + 545, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 303, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 308, + 303, + 545, + 390 + ], + "type": "text", + "content": "[71] T. Yamazaki, H. Katayama, S. Uehara, A. Nose, M. Kobayashi, S. Shida, M. Odahara, K. Takamiya, Y. Hisamatsu, S. Matsumoto, L. Miyashita, Y. Watanabe, T. Izawa, Y. Muramatsu, and M. Ishikawa. 4.9 a 1ms high-speed vision chip with 3D-stacked 140GOPS column-parallel PEs for spatio-temporal image processing. In 2017 IEEE International Solid-State Circuits Conference (ISSCC), pages 82–83, 2017. doi: 10.1109/ISSCC.2017.7870271. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 391, + 545, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 391, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 308, + 391, + 545, + 434 + ], + "type": "text", + "content": "[72] F. Yang, Y. M. Lu, L. Sbaiz, and M. Vetterli. Bits from photons: Oversampled image acquisition using binary poisson statistics. IEEE Transactions on Image Processing, 21(4): 1421-1436, 2012. doi: 10.1109/TIP.2011.2179306. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 435, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 435, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 435, + 545, + 488 + ], + "type": "text", + "content": "[73] X. Yuan, Y. Liu, J. Suo, F. Durand, and Q. Dai. Plug-and-play algorithms for video snapshot compressive imaging. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):7093-7111, 2022. doi: 10.1109/TPAMI.2021.3099035.7, 9" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 489, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 545, + 533 + ], + "type": "text", + "content": "[74] J. Zhang, X. Yang, Y. Fu, X. Wei, B. Yin, and B. Dong. Object tracking by jointly exploiting frame and event domain. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13043-13052, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 534, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 566 + ], + "type": "text", + "content": "[75] K. Zhang, Y. Li, W. Zuo, L. Zhang, L. Van Gool, and R. Timofte. Plug-and-play image restoration with deep denoiser prior. arXiv preprint, 2020. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 567, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 609 + ], + "type": "text", + "content": "[76] R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 610, + 545, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 610, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 308, + 610, + 545, + 664 + ], + "type": "text", + "content": "[77] T. Zhang, M. J. White, A. Dave, S. Ghajari, A. Raghuram, A. C. Molnar, and A. Veeraraghavan. First arrival differential LiDAR. In 2022 IEEE International Conference on Computational Photography (ICCP), pages 1-12, 2022. doi: 10.1109/ICCP54855.2022.9887683. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 665, + 545, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 665, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 308, + 665, + 545, + 696 + ], + "type": "text", + "content": "[78] W. Zhang and L. Lin. Light field flow estimation based on occlusion detection. Journal of Computer and Communications, 5(3):1-9, 2017. 1" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "8176" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Social Diffusion_ Long-term Multiple Human Motion Anticipation/e840bc61-69a6-4f9d-9e0d-f25143903533_content_list.json b/2023/Social Diffusion_ Long-term Multiple Human Motion Anticipation/e840bc61-69a6-4f9d-9e0d-f25143903533_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..544e09d5e55de1a21ff35a9fb08f695fdb8a8fc1 --- /dev/null +++ b/2023/Social Diffusion_ Long-term Multiple Human Motion Anticipation/e840bc61-69a6-4f9d-9e0d-f25143903533_content_list.json @@ -0,0 +1,1887 @@ +[ + { + "type": "text", + "text": "Social Diffusion: Long-term Multiple Human Motion Anticipation", + "text_level": 1, + "bbox": [ + 148, + 130, + 818, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Julian Tanke\\*, Linguang Zhang?, Amy Zhao?, Chengcheng Tang?, Yujun Cai?, Lezi Wang?, Po-Chen Wu?, Juergen Gall1,3, and Cem Keskin2", + "bbox": [ + 83, + 191, + 883, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Bonn 2Reality Labs Research \n3Lamarr Institute for Machine Learning and Artificial Intelligence", + "bbox": [ + 222, + 239, + 743, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{tanke,gall}@iai.uni-bonn.de {linguang,xamyzhao,chengcheng.tang,yujuncai,wanglezi,pochenwu,cemkeskin}@meta.com", + "bbox": [ + 129, + 277, + 843, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 349, + 313, + 364 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose Social Diffusion, a novel method for short-term and long-term forecasting of the motion of multiple persons as well as their social interactions. Jointly forecasting motions for multiple persons involved in social activities is inherently a challenging problem due to the interdependencies between individuals. In this work, we leverage a diffusion model conditioned on motion histories and causal temporal convolutional networks to forecast individually and contextually plausible motions for all participants. The contextual plausibility is achieved via an order-invariant aggregation function. As a second contribution, we design a new evaluation protocol that measures the plausibility of social interactions which we evaluate on the Haggling dataset, which features a challenging social activity where people are actively taking turns to talk and switching their attention. We evaluate our approach on four datasets for multi-person forecasting where our approach outperforms the state-of-the-art in terms of motion realism and contextual plausibility.", + "bbox": [ + 75, + 381, + 472, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 681, + 207, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding and anticipating social interactions in groups of people is a challenging and highly relevant topic [35, 9, 32, 45, 47, 5, 34]. For instance, it is essential for socially-compliant robots [44], but it is also relevant for neuroscience and social sciences since it allows to develop computational models on how the behavior of other persons is perceived and how it changes the own behavior.", + "bbox": [ + 75, + 708, + 468, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Forecasting realistic social interactions, however, is very challenging for two reasons. First, social interactions tend to last for tens of seconds [33] or even minutes - much longer than the prediction from most of the existing human", + "bbox": [ + 75, + 815, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "motion anticipation models [23, 69, 3, 22, 27, 42, 49, 11, 25, 37, 4, 14, 41, 40]. Second, social interactions consist of interdependent motions [54, 36], which requires modeling the relationships among all individuals. For example, in conversational turn-taking, a person's turn to talk highly depends on the start/end of the others' speaking. While multi-person motion anticipation has emerged as a new topic, current approaches [23, 69, 3] do not pay much attention on complex social interactions. For instance, they do not preserve the social role of individuals in a group such that the interactions become socially implausible over time.", + "bbox": [ + 500, + 349, + 890, + 517 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the limitations of existing models, we propose Social Diffusion to predict motions of multiple people and ensure contextually plausible interactions, as shown in Fig. 1. To this end, we learn the distribution of human motion by leveraging a diffusion model [39, 28, 52, 58, 24, 59, 62, 73]. To enforce information exchange among people, which is critical to predicting contextually plausible interactions, we introduce an order-invariant aggregation function to aggregate motion features from all people. For inference, we feed back the input sequence to the signal during the reverse-diffusion steps to condition the motion generation on the past motion. Our method is fully convolutional which allows us to generate sequences of arbitrary size. This allows us to not just forecast the next few seconds of an input motion but also to forecast social interactions that last longer. Furthermore, our approach is very flexible in the sense that the number of persons during training and inference can differ. To the best of our knowledge, our approach represents the first diffusion model that produces multi-person motions at the same time.", + "bbox": [ + 496, + 518, + 892, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a second contribution, we propose a new evaluation protocol for social interactions based on Symbolic Social Cues, which measures whether the forecast motion is socially plausible. Our key observation is that the probabilities of transitions between social interaction states are highly", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done partially while Julian was at Reality Labs Research.", + "bbox": [ + 94, + 886, + 434, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9601", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ce52e1e5601acd91e4fbea13d22232ae67f1981ecc07eda16150bcf2b7facdb3.jpg", + "image_caption": [ + "Figure 1: We propose an approach for multi-person motion anticipation: given a sequence of human social interactions (blue/red skeletons), the proposed model forecasts multi-person motions where the social roles are preserved and the interactions are socially plausible." + ], + "image_footnote": [], + "bbox": [ + 83, + 85, + 890, + 209 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "correlated with the plausibility of predicted social interactions. In a conversation, for example, a person usually starts talking only when a peer stops talking. To evaluate predicted motions, we first build the state transition graph by extracting states from the motions. We then treat the state transition graph as a probability distribution and compare it to the real data distribution.", + "bbox": [ + 75, + 277, + 468, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For evaluation, we use the Haggling dataset [30] which comprises 175 videos of well-defined triadic social interactions. In contrast to other existing multi-person human motion datasets [43, 66, 69], the persons have different social roles that impact their behavior. We furthermore evaluate our approach on the MuPoTS-3D [43], 3DPW [67], and CMU-Mocap [1, 69] dataset. On all four datasets, our approach outperforms the state of the art for multi-person human motion forecasting.", + "bbox": [ + 75, + 383, + 468, + 520 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contribution is two-fold:", + "bbox": [ + 96, + 521, + 372, + 535 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose Social Diffusion, the first stochastic multiperson motion anticipation model that outperforms the state of the art on common multi-person motion anticipation datasets.", + "2. We propose a novel social interaction evaluation protocol that considers not only the validity of poses but also the plausibility of social interactions." + ], + "bbox": [ + 88, + 546, + 468, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 678, + 218, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single Person Human Motion Prediction Human motion prediction, which typically refers to generating motion sequences given a prefix segment, has been extensively studied in the past few years. Most recent papers focus on single person 3D predictions. Due to the inherent temporal nature of human motion, various temporal-based methods, such as recurrent neural networks (RNNs) [18, 22, 27, 42, 49, 48, 61, 72], temporal convolutional networks [11, 25, 37], transformers [4, 14, 12] and graph neural networks (GNNs) [14, 41, 40] have been used for this task. For long-term predictions, auto-regressive models that operate in the discrete space [46, 38, 61] have shown success, where a long prediction sequence can be obtained without", + "bbox": [ + 75, + 703, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "converging to mean poses. Recently, anchor-based methods [16, 68] have been proposed; these focus on forecasting characteristic anchor poses rather than the entire sequence auto-regressively. To achieve better interactions with environments, [7, 70, 68, 74] proposed various ways to include contextual information into human motion predictions. Since generated motions can be controlled with a high-level guidance such as action class or text, some approaches [50, 13] used Variational Auto-Encoders [58] to solve this problem. Going beyond single human prediction, we predict the motions and interactions of multiple humans.", + "bbox": [ + 496, + 277, + 892, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-person Anticipation Modeling multi-person interactions has been a long standing problem [46, 23, 2, 65, 20, 30, 46, 63, 75]. For instance, $\\mathrm{DR}^2\\mathrm{N}$ [60] predicts the activities of multiple people given a past video sequence. For a given frame, personal relationships between candidates are estimated using a graph attention network (GAT) [65] while temporal relationships are predicted using recurrent neural networks. Recently, Wang et al. [69] addressed multi-person 3D motion trajectory prediction via a Multi-Range Transformers framework. Guo et al. [23] introduced Cross-Interaction Attention to jointly model highly expressive dance sequences. Joo et al. [30] introduced a triadic haggling game for social signal prediction, based on Panoptic Studio [29, 31]. In their work they predict the motion of a single person given the motion of the others. Similarly, [46] predicts motions based on the other actors' behavior. Crucially, both methods only predict the motion of a single individual, given other persons social signals.", + "bbox": [ + 496, + 446, + 893, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models Diffusion models [24, 57, 59] belong to the family of probabilistic generative models, which convert the training data successively to Gaussian noise, and then learn to recover the data by reversing this noising process. Diffusion models have emerged as powerful deep generative models with breakthroughs in many applications, including image synthesis [24, 15], segmentation [10], and natural language processing [8, 26]. For conditioned generation, [15] introduced classifier-guided diffusion, and [55] takes the inpainted images as denoised by the model. More recently, [62, 73] have suggested diffusion models for motion generation; however, they are limited to single human", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9602", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cbf1d81ef2fc39b6828648edcfe6213bf3769da7a72fb791a16c4d47cd156a1a.jpg", + "image_caption": [ + "Figure 2: Model overview: the reverse diffusion process $g$ consists of a fully convolutional causal encoder $e$ and a fully convolutional causal decoder $d$ that produces a denoised motion sequence $\\hat{\\mathbf{x}}_0^i \\equiv \\hat{\\mathbf{x}}_0^{i,1:N}$ for person $i$ , given the bottleneck state $h^i$ and the aggregation function $\\Gamma(\\{\\mathbf{h}_j \\mid j \\in 1, \\dots, p\\})$ over all people in the scene." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 478, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D motion prediction. To the best of our knowledge, we are the first to build a stochastic multi-person motion anticipation model that can predict very long-term future motions.", + "bbox": [ + 75, + 417, + 468, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Social Diffusion Model", + "text_level": 1, + "bbox": [ + 76, + 474, + 292, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 1, we aim to forecast the motion of multiple persons that interact with each other. The forecast motion should be realistic and socially plausible. For instance, not all persons should talk at the same time. Formally, we represent a human motion sequence with $p$ people of length $N$ as $\\mathbf{X}^{1:N} \\in \\mathbb{R}^{N \\times p \\times \\delta}$ where $\\delta$ represents the dimension of the individual pose vector at a given frame. Our goal is then to predict the future motion $\\hat{\\mathbf{X}}^{n+1:N}$ for all people, given their past motions $\\mathbf{X}^{1:n}$ :", + "bbox": [ + 75, + 500, + 468, + 637 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {X}} ^ {n + 1: N} = \\operatorname {S D M} \\left(\\mathbf {X} ^ {1: n}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 643, + 468, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Before we describe the proposed Social Diffusion Model (SDM) in Section 3.2, we will briefly describe a generic diffusion model [24] in Section 3.1. In Section 4, we will then introduce the social interaction evaluation protocol.", + "bbox": [ + 75, + 669, + 468, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Diffusion Model", + "text_level": 1, + "bbox": [ + 76, + 738, + 238, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A latent representation $\\mathbf{X}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{1})$ is obtained via a $T$ step Markov Gaussian noising process $q(\\mathbf{X}_T|\\mathbf{X}_0)$ where $\\mathbf{X}_0\\equiv \\mathbf{X}^{1:N}$ is a real motion sequence from the training set. The Markov Gaussian noising process can be written in closed form:", + "bbox": [ + 75, + 761, + 468, + 835 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {X} _ {t} \\mid \\mathbf {X} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {X} _ {t}; \\sqrt {\\alpha_ {t}} \\mathbf {X} _ {0}, (1 - \\alpha_ {t}) \\mathbf {I}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 845, + 468, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{t} \\in (0,1)$ is a step-dependent fixed hyperparameter. To sample from the generative model, we learn", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to invert the noising step using the generator function $g$ :", + "bbox": [ + 498, + 90, + 890, + 107 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {X}} _ {0, t} = g \\left(\\mathbf {X} _ {t}, t\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 128, + 890, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The key contribution of our model is the novel generator function $g$ , which models social interactions over time and will be described in Section 3.2. Following [24, 62, 51], the loss during training is defined by:", + "bbox": [ + 496, + 157, + 893, + 219 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {\\mathbf {X} _ {0} \\sim p (\\mathbf {X}), t \\sim [ 1, T ]} \\left[ \\left| \\left| \\mathbf {X} _ {0} - g \\left(\\mathbf {X} _ {t}, t\\right) \\right| \\right| \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 226, + 890, + 261 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For inference, we reverse-iterate over Equation (3), starting at sampling step $T$ and latent representation $\\mathbf{X}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{1})$ . At each iteration $t$ , we slowly denoise the motion sequence using (2) and (3):", + "bbox": [ + 496, + 268, + 893, + 330 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {X}} _ {0, t - 1} = g \\left(q \\left(\\hat {\\mathbf {X}} _ {t - 1} \\mid \\hat {\\mathbf {X}} _ {0, t}\\right), t - 1\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 337, + 890, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The final denoised motion is obtained when $t = 1$ .", + "bbox": [ + 498, + 366, + 834, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multi-person Motion Generator", + "text_level": 1, + "bbox": [ + 500, + 388, + 782, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The diffusion model described in Section 3.1 produces unconditioned motion. In order to use the model for motion forecasting, we need to condition the model on the observed motion sequence of all persons $\\mathbf{X}_0^{1:n} = \\mathbf{X}^{1:n}$ . To this end, we modify the inference sampling (5) to also include past motion as follows:", + "bbox": [ + 496, + 412, + 890, + 502 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {X}} _ {0, t - 1} = g \\left(q \\left(\\hat {\\mathbf {X}} _ {t - 1} \\mid \\mathbf {X} ^ {1: n} \\cup \\hat {\\mathbf {X}} _ {0, t} ^ {n + 1: N}\\right), t - 1\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 511, + 890, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The reverse diffusion process $g(\\mathbf{X}_t, t)$ consists of three components, a causal temporal convolutional encoder $e$ , a causal temporal convolutional decoder $d$ , and an order-invariant function $\\Gamma$ that aggregates the interaction of the different persons, see Figure 2.", + "bbox": [ + 496, + 540, + 890, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The encoder $e$ and the decoder $d$ process each individual sequence independently while $\\Gamma$ ensures that information flows between all persons in the scene. Formally, given $\\mathbf{x}_t^i = \\hat{\\mathbf{X}}_t^{i,1:N} \\in \\mathbb{R}^{N \\times \\delta}$ , which is the motion sequence for a single person $i$ at scheduled noisng step $t$ , we obtain for each person $i$ in the scene a bottleneck encoding $\\mathbf{h}^i$ :", + "bbox": [ + 496, + 616, + 890, + 705 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} ^ {i} = e \\left(\\mathbf {x} _ {t} ^ {i}, t\\right) \\forall i \\in 1, \\dots , p. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 715, + 890, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The encoder $e$ consists of a 4-layer temporal convolutional network where each layer progressively reduces the temporal resolution by half via striding. In each layer, the noising step $t$ is fed via sinusoidal positional encoding [64]. To produce the denoised motion $\\hat{\\mathbf{x}}_0^i$ , the decoder $d$ can be utilized as follows:", + "bbox": [ + 496, + 743, + 890, + 832 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {0} ^ {i} = d \\left(\\mathbf {h} ^ {i}, t, \\Gamma \\left(\\mathbf {h} ^ {j} \\forall j \\in 1, \\dots , p\\right)\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 842, + 890, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $d$ is a 4-layer temporal convolutional network and each layer progressively doubles the temporal resolution via", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9603", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "linear upsampling. As for the encoder, the noising step $t$ is fed via sinusoidal positional encoding to each layer. In addition, the output of the order-invariant aggregation function $\\Gamma$ is concatenated to $\\mathbf{h}^{\\mathrm{i}}$ before passing it to the first convolutional layer. The estimated motion sequences $\\hat{\\mathbf{x}}_0^i$ of each person $i$ at noising step $t$ are then concatenated to obtain $\\hat{\\mathbf{X}}_{0,t}$ and the approach proceeds to the next step $t - 1$ .", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The order-invariant aggregation function $\\Gamma$ passes information from other people in the scene. In our experiments, we evaluate two aggregation functions, averaging $(\\Gamma_{\\mathbb{E}})$ over all people and multi-headed attention [64] $(\\Gamma_{\\mathrm{attn}})$ :", + "bbox": [ + 75, + 196, + 468, + 258 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Gamma_ {\\mathrm {a t t n}} (\\mathbf {H}) = \\operatorname {M u l t i H e a d} (\\mathbf {H}) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 265, + 468, + 282 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Gamma_ {\\mathbb {E}} (\\mathbf {H}) = \\frac {1}{p} \\sum_ {i = 1} ^ {p} \\mathbf {h} ^ {i} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 286, + 468, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{H} = \\{h^i\\}_{i=1}^p$ are the embeddings of all layers of the encoder and for all people in the scene. MultiHead $(\\mathbf{H})$ calculates the self-attention across all people for a given frame.", + "bbox": [ + 76, + 333, + 468, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 387, + 294, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We follow state-of-the-art diffusion models [24, 39, 62] and use the cosine variance schedule. We set the number of diffusion steps to $T = 1000$ . The encoder $e$ consists of four layers of convolutional blocks with kernel size 3 and stride 2. The decoder $d$ consists of four layers of convolutional blocks with additional upsampling layers that upsample the input sequence by factor two using linear interpolation. We standardize the training data to have zero mean and standard deviation one. We normalize all poses by splitting pose and global translation: each pose is transformed into a hipcentric coordinate frame and the pose is concatenated with the global rotation and translation to form a $\\delta$ dimensional pose vector.", + "bbox": [ + 75, + 411, + 468, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Symbolic Social Cues Protocol", + "text_level": 1, + "bbox": [ + 76, + 619, + 356, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We are interested in anticipating the social interactions among multiple people. Multi-person social interactions consist of several intricate and complex behaviours such as paying attention to a specific person [56] and turn-taking [54, 36], which usually take tens of seconds or even minutes. Current state-of-the-art multi-person motion anticipation methods [3, 69] calculate the Mean Per Joint Positional Error (MPJPE) using the ground-truth sequence, which is only meaningful for short time horizons of around one second [6, 21, 53, 61, 71]. More important, however, is that it does not measure the realism of social interaction.", + "bbox": [ + 75, + 643, + 468, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We thus propose the Symbolic Social Cues Protocol (SSCP), which divides the social interactions into a set of discrete interaction classes. In SSP, we define a social signal function", + "bbox": [ + 75, + 811, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC ^ {1: N} = s \\left(\\mathbf {X} ^ {1: N}\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 878, + 468, + 896 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which takes as input a multi-person motion sequence $\\mathbf{X}^{1:N}\\in \\mathbb{R}^{N\\times p\\times \\delta}$ and produces a discrete symbolic representation $C^{1:N} = \\{c_n\\}_{n = 1}^N$ , where $c_{n}\\in \\{1,\\dots,m\\}$ and $m$ represents the total number of symbolic states. A symbolic state is a unique summary of the current state of interaction, e.g., a person is talking and another person is listening. Given a test set $\\mathcal{X} = \\{\\mathbf{X}_i^{1:N}\\}_{i = 1}^K$ with $K$ sequences, we can now calculate the probability distribution $p_{\\mathrm{sscp}}$ over the social state transitions.", + "bbox": [ + 498, + 90, + 890, + 226 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\mathrm {s s c p}} (\\mathcal {X}) = \\frac {1}{\\zeta} \\sum_ {i = 1} ^ {K} \\operatorname {s t m} \\left(s \\left(\\mathbf {X} _ {i} ^ {1: N}\\right)\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 234, + 890, + 276 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{stm}(C_i^{1:N})$ produces the $m\\times m$ state transition matrix for the discrete symbolic sequence $C^{1:N}$ and $\\zeta = \\sum_{i=1}^{K}\\sum_{m',m''}\\mathrm{stm}(C^{1:N})_{m',m''}$ is a normalization constant to ensure that $p_{\\mathrm{sscp}}$ is a valid probability distribution.", + "bbox": [ + 498, + 287, + 890, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To evaluate a motion anticipation model $f$ , we predict the future motion for all $K$ test sequences from a fixed start frame $n$ until the end of the sequence $N$ :", + "bbox": [ + 498, + 349, + 890, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {X}} ^ {n + 1: N} = \\left\\{f \\left(\\mathbf {X} _ {i} ^ {1: n}\\right) \\right\\} _ {i = 1} ^ {K} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 402, + 890, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We can now calculate the distance between the generated and ground-truth social motion distribution:", + "bbox": [ + 498, + 431, + 890, + 462 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {\\mathrm {J S D}} \\left(p _ {\\text {s s c p}} \\left(\\mathcal {X} ^ {n: N}\\right), p _ {\\text {s s c p}} \\left(\\hat {\\mathcal {X}} ^ {n: N}\\right)\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 472, + 890, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$D_{\\mathrm{JSD}}$ is the squared Jensen-Shannon distance [17, 19]:", + "bbox": [ + 500, + 501, + 890, + 517 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {\\mathrm {J S D}} (p \\| q) = \\sqrt {\\frac {\\left(D _ {\\mathrm {K L}} (p \\| \\frac {p + q}{2}) + D _ {\\mathrm {K L}} (q \\| \\frac {p + q}{2})\\right)}{2}} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 540, + 890, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p$ and $q$ are probability distributions and $D_{\\mathrm{KL}}$ is the Kullback-Leibler divergence. Note that we compare the generated motion $\\hat{\\mathcal{X}}^{n + 1:N}$ to the test set $\\mathcal{X}^{n + 1:N}$ with the same start frame as the data distribution might shift across time.", + "bbox": [ + 498, + 590, + 890, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 679, + 633, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate our approach on four datasets. Following [69], we report the Mean Per Joint Positional Error (MPJPE) in global and local aligned coordinates for the multi-person human motion datasets MuPoTS-3D [43], 3DPW [67] and CMU-Mocap [1, 69]. MuPoTS-3D [43] contains recordings of 2 to 3 persons in workout settings. Interactions between the persons are rare. 3DPW [67] contains recordings of 1 to 2 persons and the sequences cover a wide range of different activities. The level of interactions range from no interaction and little interaction, like two persons walking, to close interactions such as dancing. CMU [69] combines the motion of different sequences from the CMU-Mocap dataset [1]. The composition of motion", + "bbox": [ + 496, + 703, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9604", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6b9b39020ac79ca53eaacf698fb15ac23c04088ea147a007a04469452421f5d0.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 86, + 93, + 240, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1c897da6acc91b72d6a44cd9612768f56ff0ed37df0ed6489ee1db582f7ba97f.jpg", + "image_caption": [ + "Figure 3: A sample frame from the Haggling dataset [30] for evaluating social interactions. (a): 3D poses in a haggling sequence. Blue limbs represent the left body side while red limbs represent the right body side. The buyer's attention is indicated as green arrow. (b): a sample video frame from a haggling sequence." + ], + "image_footnote": [], + "bbox": [ + 277, + 108, + 467, + 250 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a3a592b1dae0819fe7503d7051d04ff9b3ef3d0ce5c73b097ddaa5ef6fdb9aee.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCMU-MocapMuPoTS-3D3DPW
1s2s3s1s2s3s1s2s3s
LTD [41]1.372.193.261.191.812.344.677.108.71
HRI [40]1.492.603.070.941.682.294.076.328.01
SP [2]1.152.713.900.921.672.514.177.179.27
MRT [69]0.961.572.180.891.592.223.876.127.83
Ours0.741.061.341.151.291.441.642.723.55
", + "bbox": [ + 78, + 369, + 468, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: MPJPE $\\downarrow$ in dm on different datasets.", + "bbox": [ + 120, + 460, + 423, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sequences, however, does not reflect realistic interactions. For these datasets, 1 second of human motion is observed and 1-3 seconds need to be forecast. As our model is generative, we sample 8 samples for each test sequence and report the average over all 8 samples. The source code including trained evaluation models and the dataset transformation script is publicly available1.", + "bbox": [ + 75, + 492, + 467, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Haggling Forecasting Dataset", + "text_level": 1, + "bbox": [ + 76, + 609, + 339, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the three datasets contain multiple persons, but very few social interactions, we prepared a new dataset for multi-person forecasting in the context of social interactions. For this, we utilize the Haggling dataset [30] where 122 participants play a social game with two sellers trying to sell their products to a buyer. Each game lasts one minute and contains interesting triadic interactions such as turn-taking and attention changes. A sample scene is shown in Figure 3. The dataset consists of 135 training sequences and 40 test sequences, sampled at $30\\mathrm{Hz}$ . Some of the 3D poses in the dataset are noisy as they have been estimated [29, 31] and we manually correct them. In total, the dataset consists of 234, 907 training and 69, 951 test frames, each with three people. For more details, please see the supplementary material.", + "bbox": [ + 75, + 633, + 467, + 857 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For evaluation, we take the first $10\\%$ of a sequence", + "bbox": [ + 94, + 859, + 467, + 875 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as observation and forecast the remaining $90\\%$ of the sequence, but we also evaluate the motion at intermediate frames ranging from frame 1 to frame 1300. It needs to be noted that long-term forecasting is highly relevant for neuroscience and social sciences since it allows to develop computational models on how the behavior of other persons is perceived and how it changes the own behavior. For measuring the plausibility of the forecast motion of each individual, we use the Normalized Directional Motion Similarity (NDMS) [61] since the measure not only considers static poses but also the motion of the forecast sequence. Furthermore, it can be applied to sequences of any length. NDMS, however, does not measure if the social interactions are plausible.", + "bbox": [ + 500, + 90, + 890, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For evaluating the social motion quality, we utilize the proposed Symbolic Social Cues Protocol described in Section 4. To this end, we need to specify the classes of social interactions. A haggling activity is composed of the sellers trying to convince the buyer to purchase their products and the buyer switching attention between the sellers. Throughout the game, certain social patterns emerged [56, 36]:", + "bbox": [ + 496, + 303, + 890, + 409 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Most of the time, only a single person speaks", + "2. For almost all frames, at least one person is talking", + "3. The sellers speak roughly the same amount of time while the buyer seldom talks", + "4. The buyer pays attention (looks at) to whoever talks", + "5. The sellers take turns to speak but sometimes interrupt each other" + ], + "bbox": [ + 511, + 415, + 890, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the well-defined structure of the task and the emerging social behaviors, we reduce the haggling game to two key signals:", + "bbox": [ + 496, + 537, + 890, + 584 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- talking: defines who is talking", + "- attention: defines who of the two sellers has the buyer's attention" + ], + "bbox": [ + 517, + 590, + 888, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given all possible combinations (e.g., both sellers can talk at the same time, or nobody talks), we end up with 16 possible states for each frame and formulate the social interactions as a symbolic representation over time. Note that we have to distinguish between left/right seller to catch events such as attention switching. Please see the supplementary material for more details.", + "bbox": [ + 496, + 645, + 890, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We define the social signal function $s(\\mathbf{X}^{1:N})$ as Equation (11), which takes a sequence of multi-person motion $\\mathbf{X}^{1:N}$ as input and generates one of the 16 distinct states per frame. For the social signal function $s$ to work on any haggling motion sequence, we need to determine three pieces of information:", + "bbox": [ + 496, + 751, + 890, + 840 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. who the buyer is,", + "2. whom the buyer is paying attention to,", + "3. whether someone is speaking." + ], + "bbox": [ + 511, + 849, + 787, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "", + "bbox": [ + 94, + 886, + 441, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9605", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1226f927eb08124c4d7851bc98d60e5b06f042f0a3a32508c1b1b80aef792b89.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCMU-MocapMuPoTS-3D3DPW
RootPoseRootPoseRootPose
1s2s3s1s2s3s1s2s3s1s2s3s1s2s3s
LTD [41]0.971.732.620.981.211.370.891.391.910.881.141.314.286.798.41
HRI [40]0.962.063.111.051.371.580.661.302.160.731.071.303.676.428.64
SP [2]0.962.012.961.031.411.710.961.382.210.721.081.303.766.869.07
MRT [69]0.601.121.710.791.051.220.671.251.860.690.991.193.425.697.30
Ours0.721.101.440.380.460.491.141.281.420.590.640.671.662.763.59
", + "bbox": [ + 127, + 88, + 841, + 185 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9baabd959d37a56a7ba1730459ec6c95dafcceb9e99060f1492cf2d1030ed631.jpg", + "table_caption": [ + "Table 2: MPJPE $\\downarrow$ in dm for root joint and pose. The lowest error is in bold and the second lowest is underscored." + ], + "table_footnote": [], + "table_body": "
Frame1510152025306012025050075010001300
MRT [69]0.6240.2780.1940.2120.2240.2150.2150.2180.2050.1800.1290.0790.0620.047
Ours (Γ∅)0.6440.2800.2060.2150.2250.2260.2290.2330.2290.2260.2290.2270.2250.226
Ours (Γattn)0.6390.2800.1990.2130.2240.2290.2320.2270.2290.2230.2330.2280.2330.223
Ours* (ΓE)0.6400.2790.2040.2160.2270.2280.2300.2330.2270.2220.2300.2290.2300.226
", + "bbox": [ + 99, + 208, + 869, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3: Per-frame average NDMS ↑ score on the Haggling dataset. The highest score is in bold and the second highest is underscored.", + "bbox": [ + 75, + 287, + 892, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To solve (1) we train a simple buyer detection network, consisting of three layers of bi-directional Gated Recurrent Units, which gets as input a haggling motion and outputs the likelihood of each participant being the buyer. In Table 5, we report our accuracy of this approach. We see that the buyer detector correctly identifies the buyer all the time on the test set.", + "bbox": [ + 75, + 333, + 467, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For (2), we define the buyer's attention as whomever they look at, which can be easily calculated from the 3D body pose:", + "bbox": [ + 75, + 438, + 467, + 483 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {i \\in \\{\\text {l e f t}, \\text {r i g h t} \\}} {\\operatorname {a r g m i n}} \\left[ n ^ {T} \\left(\\frac {d _ {i}}{| d _ {i} |}\\right) \\right] \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 494, + 467, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $d_{\\mathrm{left}}$ and $d_{\\mathrm{right}}$ are the directional vectors from the buyer nose to the left and right seller nose, respectively, projected onto the ground plane and $n$ is the 2D unit vector that is perpendicular to the left eye→right eye vector of the buyer, projected onto the ground plane.", + "bbox": [ + 75, + 541, + 467, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Last but not least, to determine if someone is speaking we utilize an off-the-shelf action classification network consisting of three layers of bi-directional Gated Recurrent Units. For training, we use the annotation of the Haggling dataset [30] which indicates if a person is speaking or not. The classifier achieves $87\\%$ accuracy in speech detection on the test set.", + "bbox": [ + 75, + 617, + 467, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Multi-Person Forecasting", + "text_level": 1, + "bbox": [ + 76, + 732, + 308, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first report the results for the multi-person human motion datasets MuPoTS-3D $[43]^2$ , 3DPW $[67]^1$ and CMU-Mocap [1, 69]. We follow [69] and report the Mean Per Joint Positional Error (MPJPE) using global coordinates in Table 1. Our approach outperforms the methods LTD [41], HRI [40], SP [2], and MRT [69] by a large margin. While some methods perform better for the first second on MuPoTS-3D [43], our approach achieves a much lower", + "bbox": [ + 75, + 756, + 467, + 876 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "error for all other settings and datasets. On the most difficult dataset 3DPW [67], the error is reduced by $57.6\\%$ , $55.6\\%$ , and $54.7\\%$ for 1, 2, and 3 seconds, respectively. As in [69], we also report the error of the position of the root joint and pose error in local coordinates, i.e., setting the root position for all frames to zero, in Table 2. The results show that our approach forecasts by far the most accurate poses and outperforms the state of the art by a large margin on all datasets. Only the position of the root joint is slightly better estimated by other methods at the beginning of the datasets MuPoTS-3D and CMU-Mocap. At 3 seconds, our approach also achieves the lowest root joint error for all three datasets.", + "bbox": [ + 500, + 333, + 890, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Multi-Person Forecasting in the Context of Social Interactions", + "text_level": 1, + "bbox": [ + 500, + 527, + 890, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the remaining experiments, we evaluate our approach on the newly prepared Haggling dataset since it contains more social interactions as the other datasets. We compare our approach to Multi-Range Transformers (MRT) [69], which performed better than other approaches in Section 5.2. We used the publicly available source code and adjusted the approach to work with $30\\mathrm{Hz}$ . We kept all other settings as is.", + "bbox": [ + 496, + 566, + 890, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first report the per-frame NDMS (higher is better) at different frames in Table 3. Our experiments show that MRT [69] is capable of generating realistic motion for a few seconds. However, after 120 frames $(4s)$ the NDMS score drops significantly. This is caused by the auto-regressive motion forecasting strategy adopted by MRT, which results in error accumulation over time. In contrast, our method continues to predict motions with good NDMS scores well into the future. As discussed in Section 3.2, we compare different variants of the aggregation function $\\Gamma$ . In terms of $\\Gamma$ , there are no major differences in terms of NDMS, but they all perform much better than MRT [69]. In Table 4, we report the average NDMS score over the entire forecast sequence. We observe that $\\Gamma_{\\emptyset}$ , i.e., using no aggregation", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2Data access and processing was conducted at University of Bonn", + "bbox": [ + 94, + 886, + 442, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9606", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/709b96924f85b85bbe64e7fb4265dcd2ab2eef26a158055020590513ef8ce370.jpg", + "image_caption": [ + "Figure 4: Our model is capable of generating realistic motion for 7 people from the Ultimatum sequence of Panoptic Studio [29], even though it was trained only on the triadic Haggling dataset. The Ultimatum sequence shares similarities with the Haggling dataset such as persons taking turns and talking to each other." + ], + "image_footnote": [], + "bbox": [ + 150, + 90, + 826, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b45c9a22f0fe144b9ebd6dce917f20f856ccfc66c6c2e1c8ed3efb2e41e59d9c.jpg", + "image_caption": [ + "(a) Train" + ], + "image_footnote": [], + "bbox": [ + 78, + 292, + 205, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/056c3f8fa3a1fd91bbf99d322c31945972cb75601fa53c01b603d9f04bfe4db9.jpg", + "image_caption": [ + "(b) Test" + ], + "image_footnote": [], + "bbox": [ + 210, + 294, + 336, + 392 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/facd8100c3df6afb7dae8aa33c167cf9907e223f8682552b37055c2069be6ce2.jpg", + "image_caption": [ + "(c) MRT [69]" + ], + "image_footnote": [], + "bbox": [ + 341, + 294, + 468, + 392 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cd5486f2618286e99fd02a717172e1f952528762bc94c6659bb78edc1d25392f.jpg", + "image_caption": [ + "(d) Ours $(\\Gamma_{\\emptyset})$" + ], + "image_footnote": [], + "bbox": [ + 78, + 412, + 205, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b7e675e708f7ced81c92bb2c06402d795927a22c5c7abefb1953335f4d43589.jpg", + "image_caption": [ + "(e) Ours $(\\Gamma_{\\mathrm{att}})$" + ], + "image_footnote": [], + "bbox": [ + 209, + 412, + 336, + 512 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eafb11b4ab620b18df8e59ae751ba1e65bf5c300866930aece59300ed1b2fb6b.jpg", + "image_caption": [ + "(f) Ours $(\\Gamma_{\\mathbb{E}})$", + "Figure 5: State transition matrices for the states defined in Section 5.1 for the ground-truth of the train (a) and test set (b). The other transition matrices are obtained by the forecast of the current state-of-the-art model MRT [69] on the test set (c) and our model without context (d), with context via attention (e), and with context via averaging (f). The more similar the transition matrix is to the test set (b), the closer it matches the test motion." + ], + "image_footnote": [], + "bbox": [ + 341, + 412, + 468, + 512 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dc9962a46508ee2be10cbb321a91471fd306d29bba5da7ce4f61472f60a7d655.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Train[69]Ours (Γ∅)Ours (Γattn)Ours*(ΓR)
NDMS ↑-0.10150.23010.22700.2297
SSCP ↓0.09990.48390.35760.32780.3252
", + "bbox": [ + 78, + 696, + 482, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "information yields a slightly higher NDMS score than the other versions, but the differences are very small. However, $\\Gamma_{\\mathrm{attn}}$ and $\\Gamma_{\\mathbb{E}}$ forecast much more plausible social interactions (lower SSCP) as we will discuss next.", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ef9e533d790c8371790df9729d635773d76d045bd133993ce89cdd3118968002.jpg", + "table_caption": [ + "Table 4: Per-frame average NDMS $\\uparrow$ and SSCP $\\downarrow$ on the Haggling dataset." + ], + "table_footnote": [], + "table_body": "
Test setMRT [69]Ours (Γ∅)Ours (Γattn)Ours* (ΓE)
1.00.32500.88120.88750.8938
", + "bbox": [ + 500, + 294, + 893, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5: Accuracy $\\uparrow$ of the buyer detection network. For MRT [69], the model randomly selects a person to be the buyer, as there is a $1/3$ chance of selecting the buyer with random chance.", + "bbox": [ + 498, + 343, + 893, + 405 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3.1 Social Motion Evaluation", + "text_level": 1, + "bbox": [ + 500, + 420, + 732, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The SSCP scores are presented in Table 4 where a lower value corresponds to more plausible forecast social interactions. All our proposed variants outperform the state-of-the-art approach MRT [69]. Using no aggregation information $(\\Gamma_{\\emptyset})$ performs worse than the aggregation functions $\\Gamma_{\\mathrm{attn}}$ (9) and $\\Gamma_{\\mathbb{E}}$ (10), which is expected since the aggregation function passes information from other people in the scene. The average aggregation $(\\Gamma_{\\mathbb{E}})$ performs slightly better than the more complex multi-headed attention approach $(\\Gamma_{\\mathrm{attn}})$ . We conjecture that averaging bottleneck encodings over all people introduces an inductive bias to pay the same attention to everyone, which works well for modelling the haggling game. For completeness, we also report the SSCP score of the training set.", + "bbox": [ + 496, + 446, + 890, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We can draw some insights about what motion each model generates by looking at the state probability transition matrices in Figure 5. For example, MRT [69] (Figure 5d) produces mostly self-loops (diagonal of transition matrix) indicating that the motion gets stuck over time. When no context information is provided $(\\Gamma_{\\emptyset})$ , our method produces motions where all three people are talking at the same time, as can be seen in Figure 5d, where the last two entries (red arrow) in the transition matrix represent states with all three people talking. This is sensible as the model sees two sellers and only one buyer during training and thus it is more likely to produce motion that resembles a seller, who talks most of the time. When the context is provided, our approach overcomes this limitation as expected and rarely produces motion where all three people are talking at the same time as shown in Figures 5e and 5f.", + "bbox": [ + 496, + 659, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9607", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These observations are also confirmed when measuring the buyer detection accuracy on the forecast motion, which is reported in Table 5. The detector fails to identify the correct buyer in the sequences that are forecast by MRT [69] and it nearly chooses the buyer at random with $1/3$ accuracy. This confirms that MRT does not forecast socially consistent sequences where the social role of the persons, namely buyer or seller, is preserved. In contrast, our method predicts motion where the buyer can be easily determined most of the time. As for the SSCP scores reported in Table 4, $\\Gamma_{\\mathbb{E}}$ performs best.", + "bbox": [ + 75, + 90, + 472, + 256 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, the aggregation function $\\Gamma_{\\mathbb{E}}$ outperforms the other aggregation functions on the Haggling dataset [30] as it produces the most socially plausible motion according to our Symbolic Social Cues Protocol while also generating highly plausible 3D body motion.", + "bbox": [ + 75, + 258, + 470, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 345, + 230, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Average velocity over time Freezing or unrealistically expanding motion are common failures in human motion anticipation. While NDMS [61] penalizes in contrast to MPJPE errors in the velocity, visualizing the average motion velocity can give interesting insights. In Figure 6, we plot the average velocity over all frames for all our summary function variants, the test set, and the state-of-the-art method MRT [69]. Note that the beginning of the sequence has a higher velocity due to people walking into the scene. We observe that MRT suffers from error accumulation caused by the auto-regressive inference scheme. The velocity produced by our motion tightly follows the test set velocity for roughly 250 frames after which the test set velocity is slightly larger. We attribute this to the higher degree of stochasticity of real motion, which results in sudden jerks and swings that increase the average velocity.", + "bbox": [ + 75, + 369, + 472, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "NDMS score over time In Figure 7, we visualize how the NDMS scores of all proposed variants and MRT evolve over time. For reference, we also calculate the NDMS score of the training data which is guaranteed to be realistic. Note that NDMS is 1 for the observed part of the test sequences. As shown in the figure, our method achieves almost the same level of realism as the training data while the quality of MRT slowly degenerates over time.", + "bbox": [ + 75, + 613, + 472, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Anticipating more than three people We have trained and evaluated our model on the Haggling dataset where each sequence consists of triadic interaction. However, the fully convolutional nature of our approach as well as the order-invariance of the summarization function $\\Gamma$ allow us to forecast any number of people. To demonstrate this capability, we predict 7 people from the Ultimatum sequence of Panoptic Studio [29] using only the model trained on the Haggling dataset. This works well because the Haggling and Ultimatum sequences share many social behaviors, such as turn-taking, talking, and paying attention while", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac00d8c38b4b636ae1aa8bd642f4cd9abdcd02230489aed81afffeaebd4d4790.jpg", + "image_caption": [ + "Figure 6: Average velocity over time for the entire test motion $\\mathcal{X}$ and generated motions $\\hat{\\mathcal{X}}$ . The x-axis represents the frames while the y-axis represents the (log) average velocity of the data." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 890, + 267 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5a351a73378648bf0e2250f7ae2b295144fbb2e523be95ef2281e01aa6347924.jpg", + "image_caption": [ + "Figure 7: Average NDMS [61] score $\\uparrow$ of all proposed variants, MRT, and the train set over time." + ], + "image_footnote": [], + "bbox": [ + 501, + 345, + 890, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "standing in a circle. Our results can be seen in Figure 4 where our model is able to predict realistic motion for 7 people, even though it was trained only on the triadic Haggling dataset. More results are provided in the supplementary material.", + "bbox": [ + 496, + 585, + 893, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 676, + 619, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present Social Diffusion, a stochastic multi-person motion anticipation model. The approach not only forecasts realistic motions on the individual level, but also plausible social interactions where the social roles of individuals are preserved over time. The approach is very flexible. It can be used for short and long-term forecasting and can be applied to larger groups than observed during training. As a second contribution, we proposed a new evaluation protocol to measure the realism of forecast social interactions. We furthermore derived a dataset for multi-person social interaction forecasting from the Haggling dataset [30] where the persons have different social roles that impact their behavior. We evaluated our approach", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9608", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "on four multi-person datasets and demonstrated that our approach outperforms the state-of-the-art for short-term and long-term anticipation both in realism of forecast motion and social interaction. The approach has still some limitations. For instance, the global positions of the root joints can be better estimated. Future directions also include extending the model to predict motions of dynamic groups of people, e.g., at a cocktail party where any individual can freely disengage from the current conversation group and join another one.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 76, + 260, + 236, + 277 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Juergen Gall has been supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) GA 1927/4-2 (FOR 2535 Anticipating Human Behavior), by the Federal Ministry of Education and Research (BMBF) under grant no. 01IS22094E WEST-AI, the project iBehave (receiving funding from the programme \"Netzwerke 2021\", an initiative of the Ministry of Culture and Science of the State of Northrhine Westphalia), and the ERC Consolidator Grant FORHUE (101044724). The sole responsibility for the content of this publication lies with the authors.", + "bbox": [ + 75, + 286, + 472, + 454 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 76, + 469, + 174, + 486 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] CMU. Carnegie-Mellon Mocap Database.", + "[2] Vida Adeli, Ehsan Adeli, Ian Reid, Juan Carlos Niebles, and Hamid Rezatofighi. Socially and contextually aware human motion and pose forecasting. Robotics and Automation Letters, 2020.", + "[3] Vida Adeli, Mahsa Ehsanpour, Ian Reid, Juan Carlos Niebles, Silvio Savarese, Ehsan Adeli, and Hamid Rezatofighi. Tripod: Human trajectory and pose dynamics forecasting in the wild. In International Conference on Computer Vision, 2021.", + "[4] Emre Aksan, Manuel Kaufmann, Peng Cao, and Otmar Hilliges. A spatio-temporal transformer for 3d human motion prediction. In International Conference on 3D Vision, 2021.", + "[5] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social lstm: Human trajectory prediction in crowded spaces. In Conference on Computer Vision and Pattern Recognition, 2016.", + "[6] Sadegh Aliakbarian, Fatemeh Sadat Saleh, Mathieu Salzmann, Lars Petersson, and Stephen Gould. A stochastic conditioning scheme for diverse human motion prediction. In Conference on Computer Vision and Pattern Recognition, 2020.", + "[7] Eduardo Alvarado, Damien Rohmer, and Marie-Paule Cani. Generating upper-body motion for real-time characters making their way through dynamic environments. Eurographics Symposium on Computer Animation, 2022." + ], + "bbox": [ + 84, + 496, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[8] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 2021.", + "[9] Timur Bagautdinov, Alexandre Alahi, François Fleuret, Pascal Fua, and Silvio Savarese. Social scene understanding: End-to-end multi-person action localization and collective activity recognition. In Conference on Computer Vision and Pattern Recognition, 2017.", + "[10] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021.", + "[11] Judith Butepage, Michael J Black, Danica Kragic, and Hedvig Kjellstrom. Deep representation learning for human motion prediction and classification. In Conference on Computer Vision and Pattern Recognition, 2017.", + "[12] Yujun Cai, Lin Huang, Yiwei Wang, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Xu Yang, Yiheng Zhu, Xiaohui Shen, et al. Learning progressive joint propagation for human motion prediction. In European Conference on Computer Vision, 2020.", + "[13] Yujun Cai, Yiwei Wang, Yiheng Zhu, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Chuanxia Zheng, Sijie Yan, Henghui Ding, et al. A unified 3d human motion synthesis model via conditional variational auto-encoder. In International Conference on Computer Vision, 2021.", + "[14] Lujing Chen, Rui Liu, Xin Yang, Dongsheng Zhou, Qiang Zhang, and Xiaopeng Wei. Sttg-net: a spatio-temporal network for human motion prediction based on transformer and graph convolution network. Visual Computing for Industry, Biomedicine, and Art, 2022.", + "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 2021.", + "[16] Christian Diller, Thomas Funkhouser, and Angela Dai. Forecasting characteristic 3d poses of human actions. Conference on Computer Vision and Pattern Recognition, 2022.", + "[17] Dominik Maria Endres and Johannes E Schindelin. A new metric for probability distributions. Transactions on Information theory, 2003.", + "[18] Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models for human dynamics. In International Conference on Computer Vision, 2015.", + "[19] Bent Fuglede and Flemming Topsoe. Jensen-Shannon divergence and Hilbert space embedding. In International Symposium on Information Theory, 2004.", + "[20] Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In Conference on Computer Vision and Pattern Recognition, 2020.", + "[21] Anand Gopalakrishnan, Ankur Mali, Dan Kifer, Lee Giles, and Alexander G Ororbia. A neural temporal model for human motion prediction. In Conference on Computer Vision and Pattern Recognition, 2019." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9609", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Liang-Yan Gui, Yu-Xiong Wang, Xiaodan Liang, and José MF Moura. Adversarial geometry-aware human motion prediction. In European Conference on Computer Vision, 2018.", + "[23] Wen Guo, Xiaoyu Bie, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Multi-person extreme motion prediction. In Conference on Computer Vision and Pattern Recognition, 2022.", + "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 2020.", + "[25] Daniel Holden, Jun Saito, Taku Komura, and Thomas Joyce. Learning motion manifolds with convolutional autoencoders. In SIGGRAPH Asia technical briefs. 2015.", + "[26] Emiel Hoogeboom, Didrik Nielsen, Priyank Jaini, Patrick Forre, and Max Welling. Argmax flows and multinomial diffusion: Learning categorical distributions. Advances in Neural Information Processing Systems, 2021.", + "[27] Ashesh Jain, Amir R Zamir, Silvio Savarese, and Ashutosh Saxena. Structural-rnn: Deep learning on spatio-temporal graphs. In Conference on Computer Vision and Pattern Recognition, 2016.", + "[28] Michael Janner, Yilun Du, Joshua B Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. International Conference on Learning Representations, 2022.", + "[29] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic Studio: A Massively Multiview System for Social Motion Capture. In International Conference on Computer Vision, 2015.", + "[30] Hanbyul Joo, Tomas Simon, Mina Cikara, and Yaser Sheikh. Towards social artificial intelligence: Nonverbal social signal prediction in a triadic interaction. In Conference on Computer Vision and Pattern Recognition, 2019.", + "[31] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Scott Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. Transactions on Pattern Analysis and Machine Intelligence, 2017.", + "[32] Julia Kantorovitch, Janne Väre, Vesa Pehkonen, Arto Laikari, and Heikki Seppälä. An assistive household robot-doing more than just cleaning. Journal of Assistive Technologies, 2014.", + "[33] Anne Keitel, Moritz M Daum, et al. The use of intonation for turn anticipation in observed conversations without visual signals as source of information. Frontiers in psychology, 2015.", + "[34] Lukasz Kidziński, Bryan Yang, Jennifer L Hicks, Apoorva Rajagopal, Scott L Delp, and Michael H Schwartz. Deep neural networks enable quantitative movement analysis using single-camera videos. Nature communications, 2020.", + "[35] Vineet Kosaraju, Amir Sadeghian, Roberto Martin-Martin, Ian Reid, Hamid Rezatofighi, and Silvio Savarese. Socialbigat: Multimodal trajectory forecasting using bicycle-gan and graph attention networks. Advances in Neural Information Processing Systems, 2019." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Stephen C Levinson and Francisco Torreira. Timing in turn-taking and its implications for processing models of language. Frontiers in psychology, 2015.", + "[37] Chen Li, Zhen Zhang, Wee Sun Lee, and Gim Hee Lee. Convolutional sequence to sequence model for human dynamics. In Conference on Computer Vision and Pattern Recognition, 2018.", + "[38] Thomas Lucas*, Fabien Baradel*, Philippe Weinzaepfel, and Grégory Rogez. Posegpt: Quantization-based 3d human motion generation and forecasting. In European Conference on Computer Vision, 2022.", + "[39] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Conference on Computer Vision and Pattern Recognition, 2022.", + "[40] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. History repeats itself: Human motion prediction via motion attention. In European Conference on Computer Vision, 2020.", + "[41] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies for human motion prediction. International Conference on Computer Vision, 2019.", + "[42] Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Conference on Computer Vision and Pattern Recognition, 2017.", + "[43] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In International Conference on 3D Vision, 2018.", + "[44] Ronja Möller, Antonino Furnari, Sebastiano Battiato, Aki Härma, and Giovanni Maria Farinella. A survey on human-aware robot navigation. Robotics and Autonomous Systems, 2021.", + "[45] Brendan Tran Morris and Mohan Manubhai Trivedi. A survey of vision-based trajectory learning and analysis for surveillance. Transactions on Circuits and Systems for Video Technology, 2008.", + "[46] Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. Learning to listen: Modeling non-deterministic dyadic facial motion. In Conference on Computer Vision and Pattern Recognition, 2022.", + "[47] Sangmin Oh, Anthony Hoogs, Amitha Perera, Naresh Cunteor, Chia-Chih Chen, Jong Taek Lee, Saurajit Mukherjee, JK Aggarwal, Hyungtae Lee, Larry Davis, et al. A large-scale benchmark dataset for event recognition in surveillance video. In Conference on Computer Vision and Pattern Recognition, 2011.", + "[48] Dario Pavllo, Christoph Feichtenhofer, Michael Auli, and David Grangier. Modeling human motion with quaternion-based neural networks. International Journal of Computer Vision, 2020.", + "[49] Dario Pavllo, David Grangier, and Michael Auli. *Quaternion: A quaternion-based recurrent model for human motion*. In *British Machine Vision Conference*, 2018." + ], + "bbox": [ + 501, + 90, + 892, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9610", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. European Conference on Computer Vision, 2022.", + "[51] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.", + "[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition, 2022.", + "[53] Alejandro Hernandez Ruiz, Juergen Gall, and Francesc Moreno-Noguer. Human motion prediction via spatiotemporal inpainting. International Conference on Computer Vision, 2019.", + "[54] Harvey Sacks, Emanuel A Schegloff, and Gail Jefferson. A simplest systematics for the organization of turn taking for conversation. In Studies in the organization of conversational interaction. 1978.", + "[55] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In Transactions on Graphics, 2022.", + "[56] Keren Shavit-Cohen and Elana Zion Golumbic. The dynamics of attention shifts among concurrent speech in a naturalistic multi-speaker virtual environment. Frontiers in Human Neuroscience, 2019.", + "[57] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, 2015.", + "[58] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 2015.", + "[59] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020.", + "[60] Chen Sun, Abhinav Shrivastava, Carl Vondrick, Rahul Sukthankar, Kevin Murphy, and Cordelia Schmid. Relational action forecasting. In Conference on Computer Vision and Pattern Recognition, 2019.", + "[61] Julian Tanke, Chintan Zaveri, and Juergen Gall. Intention-based long-term human motion anticipation. In International Conference on 3D Vision, 2021.", + "[62] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022.", + "[63] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 2017.", + "[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 2017." + ], + "bbox": [ + 78, + 90, + 467, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. International Conference on Learning Representations, 2017.", + "[66] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European Conference on Computer Vision, 2018.", + "[67] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European conference on computer vision, 2018.", + "[68] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In Conference on Computer Vision and Pattern Recognition, 2022.", + "[69] Jiashun Wang, Huazhe Xu, Medhini Narasimhan, and Xiaolong Wang. Multi-person 3d motion prediction with multirange transformers. Advances in Neural Information Processing Systems, 2021.", + "[70] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In Conference on Computer Vision and Pattern Recognition, 2021.", + "[71] Ye Yuan and Kris Kitani. DLOW: Diversifying latent flows for diverse human motion prediction. In European Conference on Computer Vision, 2020.", + "[72] Jason Y Zhang, Panna Felsen, Angjoo Kanazawa, and Jitendra Malik. Predicting 3d human dynamics from video. In International Conference on Computer Vision, 2019.", + "[73] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022.", + "[74] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J Black, and Siyu Tang. Generating 3d people in scenes without people. In Conference on computer vision and pattern recognition, 2020.", + "[75] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3d scenes. In Conference on Computer Vision and Pattern Recognition, 2022." + ], + "bbox": [ + 501, + 92, + 890, + 671 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9611", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Social Diffusion_ Long-term Multiple Human Motion Anticipation/e840bc61-69a6-4f9d-9e0d-f25143903533_model.json b/2023/Social Diffusion_ Long-term Multiple Human Motion Anticipation/e840bc61-69a6-4f9d-9e0d-f25143903533_model.json new file mode 100644 index 0000000000000000000000000000000000000000..63a6a92557298d2cc453ec81eddfa15a08f3d83a --- /dev/null +++ b/2023/Social Diffusion_ Long-term Multiple Human Motion Anticipation/e840bc61-69a6-4f9d-9e0d-f25143903533_model.json @@ -0,0 +1,2829 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.15, + 0.131, + 0.82, + 0.154 + ], + "angle": 0, + "content": "Social Diffusion: Long-term Multiple Human Motion Anticipation" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.193, + 0.885, + 0.23 + ], + "angle": 0, + "content": "Julian Tanke\\*, Linguang Zhang?, Amy Zhao?, Chengcheng Tang?, Yujun Cai?, Lezi Wang?, Po-Chen Wu?, Juergen Gall1,3, and Cem Keskin2" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.241, + 0.745, + 0.277 + ], + "angle": 0, + "content": "1University of Bonn 2Reality Labs Research \n3Lamarr Institute for Machine Learning and Artificial Intelligence" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.278, + 0.844, + 0.305 + ], + "angle": 0, + "content": "{tanke,gall}@iai.uni-bonn.de {linguang,xamyzhao,chengcheng.tang,yujuncai,wanglezi,pochenwu,cemkeskin}@meta.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.35, + 0.314, + 0.366 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.473, + 0.669 + ], + "angle": 0, + "content": "We propose Social Diffusion, a novel method for short-term and long-term forecasting of the motion of multiple persons as well as their social interactions. Jointly forecasting motions for multiple persons involved in social activities is inherently a challenging problem due to the interdependencies between individuals. In this work, we leverage a diffusion model conditioned on motion histories and causal temporal convolutional networks to forecast individually and contextually plausible motions for all participants. The contextual plausibility is achieved via an order-invariant aggregation function. As a second contribution, we design a new evaluation protocol that measures the plausibility of social interactions which we evaluate on the Haggling dataset, which features a challenging social activity where people are actively taking turns to talk and switching their attention. We evaluate our approach on four datasets for multi-person forecasting where our approach outperforms the state-of-the-art in terms of motion realism and contextual plausibility." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.683, + 0.208, + 0.699 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.709, + 0.47, + 0.815 + ], + "angle": 0, + "content": "Understanding and anticipating social interactions in groups of people is a challenging and highly relevant topic [35, 9, 32, 45, 47, 5, 34]. For instance, it is essential for socially-compliant robots [44], but it is also relevant for neuroscience and social sciences since it allows to develop computational models on how the behavior of other persons is perceived and how it changes the own behavior." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.816, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Forecasting realistic social interactions, however, is very challenging for two reasons. First, social interactions tend to last for tens of seconds [33] or even minutes - much longer than the prediction from most of the existing human" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.351, + 0.892, + 0.518 + ], + "angle": 0, + "content": "motion anticipation models [23, 69, 3, 22, 27, 42, 49, 11, 25, 37, 4, 14, 41, 40]. Second, social interactions consist of interdependent motions [54, 36], which requires modeling the relationships among all individuals. For example, in conversational turn-taking, a person's turn to talk highly depends on the start/end of the others' speaking. While multi-person motion anticipation has emerged as a new topic, current approaches [23, 69, 3] do not pay much attention on complex social interactions. For instance, they do not preserve the social role of individuals in a group such that the interactions become socially implausible over time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.893, + 0.822 + ], + "angle": 0, + "content": "To address the limitations of existing models, we propose Social Diffusion to predict motions of multiple people and ensure contextually plausible interactions, as shown in Fig. 1. To this end, we learn the distribution of human motion by leveraging a diffusion model [39, 28, 52, 58, 24, 59, 62, 73]. To enforce information exchange among people, which is critical to predicting contextually plausible interactions, we introduce an order-invariant aggregation function to aggregate motion features from all people. For inference, we feed back the input sequence to the signal during the reverse-diffusion steps to condition the motion generation on the past motion. Our method is fully convolutional which allows us to generate sequences of arbitrary size. This allows us to not just forecast the next few seconds of an input motion but also to forecast social interactions that last longer. Furthermore, our approach is very flexible in the sense that the number of persons during training and inference can differ. To the best of our knowledge, our approach represents the first diffusion model that produces multi-person motions at the same time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "As a second contribution, we propose a new evaluation protocol for social interactions based on Symbolic Social Cues, which measures whether the forecast motion is socially plausible. Our key observation is that the probabilities of transitions between social interaction states are highly" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.436, + 0.901 + ], + "angle": 0, + "content": "*Work done partially while Julian was at Reality Labs Research." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9601" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.087, + 0.891, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.222, + 0.893, + 0.267 + ], + "angle": 0, + "content": "Figure 1: We propose an approach for multi-person motion anticipation: given a sequence of human social interactions (blue/red skeletons), the proposed model forecasts multi-person motions where the social roles are preserved and the interactions are socially plausible." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.279, + 0.47, + 0.384 + ], + "angle": 0, + "content": "correlated with the plausibility of predicted social interactions. In a conversation, for example, a person usually starts talking only when a peer stops talking. To evaluate predicted motions, we first build the state transition graph by extracting states from the motions. We then treat the state transition graph as a probability distribution and compare it to the real data distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.47, + 0.521 + ], + "angle": 0, + "content": "For evaluation, we use the Haggling dataset [30] which comprises 175 videos of well-defined triadic social interactions. In contrast to other existing multi-person human motion datasets [43, 66, 69], the persons have different social roles that impact their behavior. We furthermore evaluate our approach on the MuPoTS-3D [43], 3DPW [67], and CMU-Mocap [1, 69] dataset. On all four datasets, our approach outperforms the state of the art for multi-person human motion forecasting." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.522, + 0.373, + 0.536 + ], + "angle": 0, + "content": "In summary, our contribution is two-fold:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.547, + 0.47, + 0.609 + ], + "angle": 0, + "content": "1. We propose Social Diffusion, the first stochastic multiperson motion anticipation model that outperforms the state of the art on common multi-person motion anticipation datasets." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.619, + 0.47, + 0.665 + ], + "angle": 0, + "content": "2. We propose a novel social interaction evaluation protocol that considers not only the validity of poses but also the plausibility of social interactions." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.547, + 0.47, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.679, + 0.22, + 0.695 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Single Person Human Motion Prediction Human motion prediction, which typically refers to generating motion sequences given a prefix segment, has been extensively studied in the past few years. Most recent papers focus on single person 3D predictions. Due to the inherent temporal nature of human motion, various temporal-based methods, such as recurrent neural networks (RNNs) [18, 22, 27, 42, 49, 48, 61, 72], temporal convolutional networks [11, 25, 37], transformers [4, 14, 12] and graph neural networks (GNNs) [14, 41, 40] have been used for this task. For long-term predictions, auto-regressive models that operate in the discrete space [46, 38, 61] have shown success, where a long prediction sequence can be obtained without" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.279, + 0.893, + 0.446 + ], + "angle": 0, + "content": "converging to mean poses. Recently, anchor-based methods [16, 68] have been proposed; these focus on forecasting characteristic anchor poses rather than the entire sequence auto-regressively. To achieve better interactions with environments, [7, 70, 68, 74] proposed various ways to include contextual information into human motion predictions. Since generated motions can be controlled with a high-level guidance such as action class or text, some approaches [50, 13] used Variational Auto-Encoders [58] to solve this problem. Going beyond single human prediction, we predict the motions and interactions of multiple humans." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.895, + 0.719 + ], + "angle": 0, + "content": "Multi-person Anticipation Modeling multi-person interactions has been a long standing problem [46, 23, 2, 65, 20, 30, 46, 63, 75]. For instance, \\(\\mathrm{DR}^2\\mathrm{N}\\) [60] predicts the activities of multiple people given a past video sequence. For a given frame, personal relationships between candidates are estimated using a graph attention network (GAT) [65] while temporal relationships are predicted using recurrent neural networks. Recently, Wang et al. [69] addressed multi-person 3D motion trajectory prediction via a Multi-Range Transformers framework. Guo et al. [23] introduced Cross-Interaction Attention to jointly model highly expressive dance sequences. Joo et al. [30] introduced a triadic haggling game for social signal prediction, based on Panoptic Studio [29, 31]. In their work they predict the motion of a single person given the motion of the others. Similarly, [46] predicts motions based on the other actors' behavior. Crucially, both methods only predict the motion of a single individual, given other persons social signals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Diffusion Models Diffusion models [24, 57, 59] belong to the family of probabilistic generative models, which convert the training data successively to Gaussian noise, and then learn to recover the data by reversing this noising process. Diffusion models have emerged as powerful deep generative models with breakthroughs in many applications, including image synthesis [24, 15], segmentation [10], and natural language processing [8, 26]. For conditioned generation, [15] introduced classifier-guided diffusion, and [55] takes the inpainted images as denoised by the model. More recently, [62, 73] have suggested diffusion models for motion generation; however, they are limited to single human" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9602" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.48, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.303, + 0.473, + 0.396 + ], + "angle": 0, + "content": "Figure 2: Model overview: the reverse diffusion process \\( g \\) consists of a fully convolutional causal encoder \\( e \\) and a fully convolutional causal decoder \\( d \\) that produces a denoised motion sequence \\( \\hat{\\mathbf{x}}_0^i \\equiv \\hat{\\mathbf{x}}_0^{i,1:N} \\) for person \\( i \\), given the bottleneck state \\( h^i \\) and the aggregation function \\( \\Gamma(\\{\\mathbf{h}_j \\mid j \\in 1, \\dots, p\\}) \\) over all people in the scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.47, + 0.465 + ], + "angle": 0, + "content": "3D motion prediction. To the best of our knowledge, we are the first to build a stochastic multi-person motion anticipation model that can predict very long-term future motions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.476, + 0.294, + 0.492 + ], + "angle": 0, + "content": "3. Social Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.47, + 0.638 + ], + "angle": 0, + "content": "As illustrated in Fig. 1, we aim to forecast the motion of multiple persons that interact with each other. The forecast motion should be realistic and socially plausible. For instance, not all persons should talk at the same time. Formally, we represent a human motion sequence with \\( p \\) people of length \\( N \\) as \\( \\mathbf{X}^{1:N} \\in \\mathbb{R}^{N \\times p \\times \\delta} \\) where \\( \\delta \\) represents the dimension of the individual pose vector at a given frame. Our goal is then to predict the future motion \\( \\hat{\\mathbf{X}}^{n+1:N} \\) for all people, given their past motions \\( \\mathbf{X}^{1:n} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.644, + 0.469, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {X}} ^ {n + 1: N} = \\operatorname {S D M} \\left(\\mathbf {X} ^ {1: n}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.47, + 0.732 + ], + "angle": 0, + "content": "Before we describe the proposed Social Diffusion Model (SDM) in Section 3.2, we will briefly describe a generic diffusion model [24] in Section 3.1. In Section 4, we will then introduce the social interaction evaluation protocol." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.739, + 0.24, + 0.754 + ], + "angle": 0, + "content": "3.1. Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.762, + 0.47, + 0.837 + ], + "angle": 0, + "content": "A latent representation \\(\\mathbf{X}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{1})\\) is obtained via a \\(T\\) step Markov Gaussian noising process \\(q(\\mathbf{X}_T|\\mathbf{X}_0)\\) where \\(\\mathbf{X}_0\\equiv \\mathbf{X}^{1:N}\\) is a real motion sequence from the training set. The Markov Gaussian noising process can be written in closed form:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.846, + 0.469, + 0.864 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {X} _ {t} \\mid \\mathbf {X} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {X} _ {t}; \\sqrt {\\alpha_ {t}} \\mathbf {X} _ {0}, (1 - \\alpha_ {t}) \\mathbf {I}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\(\\alpha_{t} \\in (0,1)\\) is a step-dependent fixed hyperparameter. To sample from the generative model, we learn" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.108 + ], + "angle": 0, + "content": "to invert the noising step using the generator function \\(g\\):" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.13, + 0.892, + 0.15 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {X}} _ {0, t} = g \\left(\\mathbf {X} _ {t}, t\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.159, + 0.894, + 0.22 + ], + "angle": 0, + "content": "The key contribution of our model is the novel generator function \\( g \\), which models social interactions over time and will be described in Section 3.2. Following [24, 62, 51], the loss during training is defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.227, + 0.892, + 0.262 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {\\mathbf {X} _ {0} \\sim p (\\mathbf {X}), t \\sim [ 1, T ]} \\left[ \\left| \\left| \\mathbf {X} _ {0} - g \\left(\\mathbf {X} _ {t}, t\\right) \\right| \\right| \\right] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.269, + 0.894, + 0.331 + ], + "angle": 0, + "content": "For inference, we reverse-iterate over Equation (3), starting at sampling step \\(T\\) and latent representation \\(\\mathbf{X}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{1})\\). At each iteration \\(t\\), we slowly denoise the motion sequence using (2) and (3):" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.338, + 0.892, + 0.357 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {X}} _ {0, t - 1} = g \\left(q \\left(\\hat {\\mathbf {X}} _ {t - 1} \\mid \\hat {\\mathbf {X}} _ {0, t}\\right), t - 1\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.367, + 0.835, + 0.381 + ], + "angle": 0, + "content": "The final denoised motion is obtained when \\( t = 1 \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.39, + 0.783, + 0.407 + ], + "angle": 0, + "content": "3.2. Multi-person Motion Generator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.892, + 0.503 + ], + "angle": 0, + "content": "The diffusion model described in Section 3.1 produces unconditioned motion. In order to use the model for motion forecasting, we need to condition the model on the observed motion sequence of all persons \\(\\mathbf{X}_0^{1:n} = \\mathbf{X}^{1:n}\\). To this end, we modify the inference sampling (5) to also include past motion as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.512, + 0.892, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {X}} _ {0, t - 1} = g \\left(q \\left(\\hat {\\mathbf {X}} _ {t - 1} \\mid \\mathbf {X} ^ {1: n} \\cup \\hat {\\mathbf {X}} _ {0, t} ^ {n + 1: N}\\right), t - 1\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.541, + 0.892, + 0.616 + ], + "angle": 0, + "content": "The reverse diffusion process \\( g(\\mathbf{X}_t, t) \\) consists of three components, a causal temporal convolutional encoder \\( e \\), a causal temporal convolutional decoder \\( d \\), and an order-invariant function \\( \\Gamma \\) that aggregates the interaction of the different persons, see Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.617, + 0.892, + 0.707 + ], + "angle": 0, + "content": "The encoder \\(e\\) and the decoder \\(d\\) process each individual sequence independently while \\(\\Gamma\\) ensures that information flows between all persons in the scene. Formally, given \\(\\mathbf{x}_t^i = \\hat{\\mathbf{X}}_t^{i,1:N} \\in \\mathbb{R}^{N \\times \\delta}\\), which is the motion sequence for a single person \\(i\\) at scheduled noisng step \\(t\\), we obtain for each person \\(i\\) in the scene a bottleneck encoding \\(\\mathbf{h}^i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.716, + 0.892, + 0.735 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} ^ {i} = e \\left(\\mathbf {x} _ {t} ^ {i}, t\\right) \\forall i \\in 1, \\dots , p. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.744, + 0.892, + 0.833 + ], + "angle": 0, + "content": "The encoder \\( e \\) consists of a 4-layer temporal convolutional network where each layer progressively reduces the temporal resolution by half via striding. In each layer, the noising step \\( t \\) is fed via sinusoidal positional encoding [64]. To produce the denoised motion \\( \\hat{\\mathbf{x}}_0^i \\), the decoder \\( d \\) can be utilized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.843, + 0.892, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {0} ^ {i} = d \\left(\\mathbf {h} ^ {i}, t, \\Gamma \\left(\\mathbf {h} ^ {j} \\forall j \\in 1, \\dots , p\\right)\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(d\\) is a 4-layer temporal convolutional network and each layer progressively doubles the temporal resolution via" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9603" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "linear upsampling. As for the encoder, the noising step \\( t \\) is fed via sinusoidal positional encoding to each layer. In addition, the output of the order-invariant aggregation function \\( \\Gamma \\) is concatenated to \\( \\mathbf{h}^{\\mathrm{i}} \\) before passing it to the first convolutional layer. The estimated motion sequences \\( \\hat{\\mathbf{x}}_0^i \\) of each person \\( i \\) at noising step \\( t \\) are then concatenated to obtain \\( \\hat{\\mathbf{X}}_{0,t} \\) and the approach proceeds to the next step \\( t - 1 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.47, + 0.259 + ], + "angle": 0, + "content": "The order-invariant aggregation function \\(\\Gamma\\) passes information from other people in the scene. In our experiments, we evaluate two aggregation functions, averaging \\((\\Gamma_{\\mathbb{E}})\\) over all people and multi-headed attention [64] \\((\\Gamma_{\\mathrm{attn}})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.266, + 0.469, + 0.283 + ], + "angle": 0, + "content": "\\[\n\\Gamma_ {\\mathrm {a t t n}} (\\mathbf {H}) = \\operatorname {M u l t i H e a d} (\\mathbf {H}) \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.287, + 0.469, + 0.326 + ], + "angle": 0, + "content": "\\[\n\\Gamma_ {\\mathbb {E}} (\\mathbf {H}) = \\frac {1}{p} \\sum_ {i = 1} ^ {p} \\mathbf {h} ^ {i} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.334, + 0.469, + 0.381 + ], + "angle": 0, + "content": "where \\(\\mathbf{H} = \\{h^i\\}_{i=1}^p\\) are the embeddings of all layers of the encoder and for all people in the scene. MultiHead \\((\\mathbf{H})\\) calculates the self-attention across all people for a given frame." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.388, + 0.295, + 0.404 + ], + "angle": 0, + "content": "3.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.47, + 0.609 + ], + "angle": 0, + "content": "We follow state-of-the-art diffusion models [24, 39, 62] and use the cosine variance schedule. We set the number of diffusion steps to \\( T = 1000 \\). The encoder \\( e \\) consists of four layers of convolutional blocks with kernel size 3 and stride 2. The decoder \\( d \\) consists of four layers of convolutional blocks with additional upsampling layers that upsample the input sequence by factor two using linear interpolation. We standardize the training data to have zero mean and standard deviation one. We normalize all poses by splitting pose and global translation: each pose is transformed into a hipcentric coordinate frame and the pose is concatenated with the global rotation and translation to form a \\( \\delta \\) dimensional pose vector." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.62, + 0.357, + 0.637 + ], + "angle": 0, + "content": "4. Symbolic Social Cues Protocol" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.469, + 0.81 + ], + "angle": 0, + "content": "We are interested in anticipating the social interactions among multiple people. Multi-person social interactions consist of several intricate and complex behaviours such as paying attention to a specific person [56] and turn-taking [54, 36], which usually take tens of seconds or even minutes. Current state-of-the-art multi-person motion anticipation methods [3, 69] calculate the Mean Per Joint Positional Error (MPJPE) using the ground-truth sequence, which is only meaningful for short time horizons of around one second [6, 21, 53, 61, 71]. More important, however, is that it does not measure the realism of social interaction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.812, + 0.47, + 0.871 + ], + "angle": 0, + "content": "We thus propose the Symbolic Social Cues Protocol (SSCP), which divides the social interactions into a set of discrete interaction classes. In SSP, we define a social signal function" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.879, + 0.469, + 0.897 + ], + "angle": 0, + "content": "\\[\nC ^ {1: N} = s \\left(\\mathbf {X} ^ {1: N}\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "which takes as input a multi-person motion sequence \\(\\mathbf{X}^{1:N}\\in \\mathbb{R}^{N\\times p\\times \\delta}\\) and produces a discrete symbolic representation \\(C^{1:N} = \\{c_n\\}_{n = 1}^N\\) , where \\(c_{n}\\in \\{1,\\dots,m\\}\\) and \\(m\\) represents the total number of symbolic states. A symbolic state is a unique summary of the current state of interaction, e.g., a person is talking and another person is listening. Given a test set \\(\\mathcal{X} = \\{\\mathbf{X}_i^{1:N}\\}_{i = 1}^K\\) with \\(K\\) sequences, we can now calculate the probability distribution \\(p_{\\mathrm{sscp}}\\) over the social state transitions." + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.236, + 0.892, + 0.277 + ], + "angle": 0, + "content": "\\[\np _ {\\mathrm {s s c p}} (\\mathcal {X}) = \\frac {1}{\\zeta} \\sum_ {i = 1} ^ {K} \\operatorname {s t m} \\left(s \\left(\\mathbf {X} _ {i} ^ {1: N}\\right)\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.288, + 0.892, + 0.349 + ], + "angle": 0, + "content": "where \\(\\mathrm{stm}(C_i^{1:N})\\) produces the \\(m\\times m\\) state transition matrix for the discrete symbolic sequence \\(C^{1:N}\\) and \\(\\zeta = \\sum_{i=1}^{K}\\sum_{m',m''}\\mathrm{stm}(C^{1:N})_{m',m''}\\) is a normalization constant to ensure that \\(p_{\\mathrm{sscp}}\\) is a valid probability distribution." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.35, + 0.892, + 0.394 + ], + "angle": 0, + "content": "To evaluate a motion anticipation model \\( f \\), we predict the future motion for all \\( K \\) test sequences from a fixed start frame \\( n \\) until the end of the sequence \\( N \\):" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.403, + 0.891, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {X}} ^ {n + 1: N} = \\left\\{f \\left(\\mathbf {X} _ {i} ^ {1: n}\\right) \\right\\} _ {i = 1} ^ {K} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.433, + 0.892, + 0.463 + ], + "angle": 0, + "content": "We can now calculate the distance between the generated and ground-truth social motion distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.473, + 0.891, + 0.492 + ], + "angle": 0, + "content": "\\[\nD _ {\\mathrm {J S D}} \\left(p _ {\\text {s s c p}} \\left(\\mathcal {X} ^ {n: N}\\right), p _ {\\text {s s c p}} \\left(\\hat {\\mathcal {X}} ^ {n: N}\\right)\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.502, + 0.891, + 0.518 + ], + "angle": 0, + "content": "\\(D_{\\mathrm{JSD}}\\) is the squared Jensen-Shannon distance [17, 19]:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.541, + 0.891, + 0.581 + ], + "angle": 0, + "content": "\\[\nD _ {\\mathrm {J S D}} (p \\| q) = \\sqrt {\\frac {\\left(D _ {\\mathrm {K L}} (p \\| \\frac {p + q}{2}) + D _ {\\mathrm {K L}} (q \\| \\frac {p + q}{2})\\right)}{2}} \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.591, + 0.892, + 0.665 + ], + "angle": 0, + "content": "where \\( p \\) and \\( q \\) are probability distributions and \\( D_{\\mathrm{KL}} \\) is the Kullback-Leibler divergence. Note that we compare the generated motion \\( \\hat{\\mathcal{X}}^{n + 1:N} \\) to the test set \\( \\mathcal{X}^{n + 1:N} \\) with the same start frame as the data distribution might shift across time." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.634, + 0.696 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We evaluate our approach on four datasets. Following [69], we report the Mean Per Joint Positional Error (MPJPE) in global and local aligned coordinates for the multi-person human motion datasets MuPoTS-3D [43], 3DPW [67] and CMU-Mocap [1, 69]. MuPoTS-3D [43] contains recordings of 2 to 3 persons in workout settings. Interactions between the persons are rare. 3DPW [67] contains recordings of 1 to 2 persons and the sequences cover a wide range of different activities. The level of interactions range from no interaction and little interaction, like two persons walking, to close interactions such as dancing. CMU [69] combines the motion of different sequences from the CMU-Mocap dataset [1]. The composition of motion" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9604" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.094, + 0.241, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.237, + 0.171, + 0.249 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.109, + 0.468, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.262, + 0.47, + 0.352 + ], + "angle": 0, + "content": "Figure 3: A sample frame from the Haggling dataset [30] for evaluating social interactions. (a): 3D poses in a haggling sequence. Blue limbs represent the left body side while red limbs represent the right body side. The buyer's attention is indicated as green arrow. (b): a sample video frame from a haggling sequence." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.37, + 0.47, + 0.456 + ], + "angle": 0, + "content": "
MethodCMU-MocapMuPoTS-3D3DPW
1s2s3s1s2s3s1s2s3s
LTD [41]1.372.193.261.191.812.344.677.108.71
HRI [40]1.492.603.070.941.682.294.076.328.01
SP [2]1.152.713.900.921.672.514.177.179.27
MRT [69]0.961.572.180.891.592.223.876.127.83
Ours0.741.061.341.151.291.441.642.723.55
" + }, + { + "type": "table_caption", + "bbox": [ + 0.121, + 0.462, + 0.424, + 0.476 + ], + "angle": 0, + "content": "Table 1: MPJPE \\( \\downarrow \\) in dm on different datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.468, + 0.598 + ], + "angle": 0, + "content": "sequences, however, does not reflect realistic interactions. For these datasets, 1 second of human motion is observed and 1-3 seconds need to be forecast. As our model is generative, we sample 8 samples for each test sequence and report the average over all 8 samples. The source code including trained evaluation models and the dataset transformation script is publicly available1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.61, + 0.34, + 0.627 + ], + "angle": 0, + "content": "5.1. Haggling Forecasting Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.634, + 0.468, + 0.858 + ], + "angle": 0, + "content": "Since the three datasets contain multiple persons, but very few social interactions, we prepared a new dataset for multi-person forecasting in the context of social interactions. For this, we utilize the Haggling dataset [30] where 122 participants play a social game with two sellers trying to sell their products to a buyer. Each game lasts one minute and contains interesting triadic interactions such as turn-taking and attention changes. A sample scene is shown in Figure 3. The dataset consists of 135 training sequences and 40 test sequences, sampled at \\(30\\mathrm{Hz}\\). Some of the 3D poses in the dataset are noisy as they have been estimated [29, 31] and we manually correct them. In total, the dataset consists of 234, 907 training and 69, 951 test frames, each with three people. For more details, please see the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.861, + 0.468, + 0.875 + ], + "angle": 0, + "content": "For evaluation, we take the first \\(10\\%\\) of a sequence" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.302 + ], + "angle": 0, + "content": "as observation and forecast the remaining \\(90\\%\\) of the sequence, but we also evaluate the motion at intermediate frames ranging from frame 1 to frame 1300. It needs to be noted that long-term forecasting is highly relevant for neuroscience and social sciences since it allows to develop computational models on how the behavior of other persons is perceived and how it changes the own behavior. For measuring the plausibility of the forecast motion of each individual, we use the Normalized Directional Motion Similarity (NDMS) [61] since the measure not only considers static poses but also the motion of the forecast sequence. Furthermore, it can be applied to sequences of any length. NDMS, however, does not measure if the social interactions are plausible." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.41 + ], + "angle": 0, + "content": "For evaluating the social motion quality, we utilize the proposed Symbolic Social Cues Protocol described in Section 4. To this end, we need to specify the classes of social interactions. A haggling activity is composed of the sellers trying to convince the buyer to purchase their products and the buyer switching attention between the sellers. Throughout the game, certain social patterns emerged [56, 36]:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.416, + 0.833, + 0.431 + ], + "angle": 0, + "content": "1. Most of the time, only a single person speaks" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.434, + 0.87, + 0.45 + ], + "angle": 0, + "content": "2. For almost all frames, at least one person is talking" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.89, + 0.481 + ], + "angle": 0, + "content": "3. The sellers speak roughly the same amount of time while the buyer seldom talks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.876, + 0.5 + ], + "angle": 0, + "content": "4. The buyer pays attention (looks at) to whoever talks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.891, + 0.531 + ], + "angle": 0, + "content": "5. The sellers take turns to speak but sometimes interrupt each other" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.416, + 0.891, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.585 + ], + "angle": 0, + "content": "Given the well-defined structure of the task and the emerging social behaviors, we reduce the haggling game to two key signals:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.592, + 0.738, + 0.607 + ], + "angle": 0, + "content": "- talking: defines who is talking" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.61, + 0.89, + 0.639 + ], + "angle": 0, + "content": "- attention: defines who of the two sellers has the buyer's attention" + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.592, + 0.89, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.646, + 0.892, + 0.751 + ], + "angle": 0, + "content": "Given all possible combinations (e.g., both sellers can talk at the same time, or nobody talks), we end up with 16 possible states for each frame and formulate the social interactions as a symbolic representation over time. Note that we have to distinguish between left/right seller to catch events such as attention switching. Please see the supplementary material for more details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.752, + 0.892, + 0.842 + ], + "angle": 0, + "content": "We define the social signal function \\( s(\\mathbf{X}^{1:N}) \\) as Equation (11), which takes a sequence of multi-person motion \\( \\mathbf{X}^{1:N} \\) as input and generates one of the 16 distinct states per frame. For the social signal function \\( s \\) to work on any haggling motion sequence, we need to determine three pieces of information:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.851, + 0.65, + 0.865 + ], + "angle": 0, + "content": "1. who the buyer is," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.868, + 0.788, + 0.884 + ], + "angle": 0, + "content": "2. whom the buyer is paying attention to," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.887, + 0.733, + 0.901 + ], + "angle": 0, + "content": "3. whether someone is speaking." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.851, + 0.788, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.442, + 0.9 + ], + "angle": 0, + "content": "" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9605" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.129, + 0.089, + 0.843, + 0.186 + ], + "angle": 0, + "content": "
MethodCMU-MocapMuPoTS-3D3DPW
RootPoseRootPoseRootPose
1s2s3s1s2s3s1s2s3s1s2s3s1s2s3s
LTD [41]0.971.732.620.981.211.370.891.391.910.881.141.314.286.798.41
HRI [40]0.962.063.111.051.371.580.661.302.160.731.071.303.676.428.64
SP [2]0.962.012.961.031.411.710.961.382.210.721.081.303.766.869.07
MRT [69]0.601.121.710.791.051.220.671.251.860.690.991.193.425.697.30
Ours0.721.101.440.380.460.491.141.281.420.590.640.671.662.763.59
" + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.191, + 0.854, + 0.206 + ], + "angle": 0, + "content": "Table 2: MPJPE \\( \\downarrow \\) in dm for root joint and pose. The lowest error is in bold and the second lowest is underscored." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.209, + 0.87, + 0.284 + ], + "angle": 0, + "content": "
Frame1510152025306012025050075010001300
MRT [69]0.6240.2780.1940.2120.2240.2150.2150.2180.2050.1800.1290.0790.0620.047
Ours (Γ∅)0.6440.2800.2060.2150.2250.2260.2290.2330.2290.2260.2290.2270.2250.226
Ours (Γattn)0.6390.2800.1990.2130.2240.2290.2320.2270.2290.2230.2330.2280.2330.223
Ours* (ΓE)0.6400.2790.2040.2160.2270.2280.2300.2330.2270.2220.2300.2290.2300.226
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.289, + 0.893, + 0.318 + ], + "angle": 0, + "content": "Table 3: Per-frame average NDMS ↑ score on the Haggling dataset. The highest score is in bold and the second highest is underscored." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.468, + 0.437 + ], + "angle": 0, + "content": "To solve (1) we train a simple buyer detection network, consisting of three layers of bi-directional Gated Recurrent Units, which gets as input a haggling motion and outputs the likelihood of each participant being the buyer. In Table 5, we report our accuracy of this approach. We see that the buyer detector correctly identifies the buyer all the time on the test set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.468, + 0.484 + ], + "angle": 0, + "content": "For (2), we define the buyer's attention as whomever they look at, which can be easily calculated from the 3D body pose:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.495, + 0.468, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\underset {i \\in \\{\\text {l e f t}, \\text {r i g h t} \\}} {\\operatorname {a r g m i n}} \\left[ n ^ {T} \\left(\\frac {d _ {i}}{| d _ {i} |}\\right) \\right] \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.468, + 0.617 + ], + "angle": 0, + "content": "where \\( d_{\\mathrm{left}} \\) and \\( d_{\\mathrm{right}} \\) are the directional vectors from the buyer nose to the left and right seller nose, respectively, projected onto the ground plane and \\( n \\) is the 2D unit vector that is perpendicular to the left eye→right eye vector of the buyer, projected onto the ground plane." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.618, + 0.468, + 0.722 + ], + "angle": 0, + "content": "Last but not least, to determine if someone is speaking we utilize an off-the-shelf action classification network consisting of three layers of bi-directional Gated Recurrent Units. For training, we use the annotation of the Haggling dataset [30] which indicates if a person is speaking or not. The classifier achieves \\(87\\%\\) accuracy in speech detection on the test set." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.733, + 0.31, + 0.75 + ], + "angle": 0, + "content": "5.2. Multi-Person Forecasting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.468, + 0.877 + ], + "angle": 0, + "content": "We first report the results for the multi-person human motion datasets MuPoTS-3D \\([43]^2\\), 3DPW \\([67]^1\\) and CMU-Mocap [1, 69]. We follow [69] and report the Mean Per Joint Positional Error (MPJPE) using global coordinates in Table 1. Our approach outperforms the methods LTD [41], HRI [40], SP [2], and MRT [69] by a large margin. While some methods perform better for the first second on MuPoTS-3D [43], our approach achieves a much lower" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.334, + 0.892, + 0.515 + ], + "angle": 0, + "content": "error for all other settings and datasets. On the most difficult dataset 3DPW [67], the error is reduced by \\(57.6\\%\\), \\(55.6\\%\\), and \\(54.7\\%\\) for 1, 2, and 3 seconds, respectively. As in [69], we also report the error of the position of the root joint and pose error in local coordinates, i.e., setting the root position for all frames to zero, in Table 2. The results show that our approach forecasts by far the most accurate poses and outperforms the state of the art by a large margin on all datasets. Only the position of the root joint is slightly better estimated by other methods at the beginning of the datasets MuPoTS-3D and CMU-Mocap. At 3 seconds, our approach also achieves the lowest root joint error for all three datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.528, + 0.892, + 0.558 + ], + "angle": 0, + "content": "5.3. Multi-Person Forecasting in the Context of Social Interactions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.688 + ], + "angle": 0, + "content": "For the remaining experiments, we evaluate our approach on the newly prepared Haggling dataset since it contains more social interactions as the other datasets. We compare our approach to Multi-Range Transformers (MRT) [69], which performed better than other approaches in Section 5.2. We used the publicly available source code and adjusted the approach to work with \\(30\\mathrm{Hz}\\). We kept all other settings as is." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We first report the per-frame NDMS (higher is better) at different frames in Table 3. Our experiments show that MRT [69] is capable of generating realistic motion for a few seconds. However, after 120 frames \\((4s)\\) the NDMS score drops significantly. This is caused by the auto-regressive motion forecasting strategy adopted by MRT, which results in error accumulation over time. In contrast, our method continues to predict motions with good NDMS scores well into the future. As discussed in Section 3.2, we compare different variants of the aggregation function \\(\\Gamma\\). In terms of \\(\\Gamma\\), there are no major differences in terms of NDMS, but they all perform much better than MRT [69]. In Table 4, we report the average NDMS score over the entire forecast sequence. We observe that \\(\\Gamma_{\\emptyset}\\), i.e., using no aggregation" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.444, + 0.901 + ], + "angle": 0, + "content": "2Data access and processing was conducted at University of Bonn" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9606" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.151, + 0.092, + 0.827, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.225, + 0.895, + 0.272 + ], + "angle": 0, + "content": "Figure 4: Our model is capable of generating realistic motion for 7 people from the Ultimatum sequence of Panoptic Studio [29], even though it was trained only on the triadic Haggling dataset. The Ultimatum sequence shares similarities with the Haggling dataset such as persons taking turns and talking to each other." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.294, + 0.207, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.398, + 0.171, + 0.411 + ], + "angle": 0, + "content": "(a) Train" + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.295, + 0.338, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.398, + 0.298, + 0.411 + ], + "angle": 0, + "content": "(b) Test" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.295, + 0.469, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.398, + 0.446, + 0.411 + ], + "angle": 0, + "content": "(c) MRT [69]" + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.414, + 0.206, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.519, + 0.185, + 0.532 + ], + "angle": 0, + "content": "(d) Ours \\((\\Gamma_{\\emptyset})\\)" + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.414, + 0.338, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.52, + 0.319, + 0.533 + ], + "angle": 0, + "content": "(e) Ours \\((\\Gamma_{\\mathrm{att}})\\)" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.414, + 0.469, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.52, + 0.446, + 0.533 + ], + "angle": 0, + "content": "(f) Ours \\((\\Gamma_{\\mathbb{E}})\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.545, + 0.473, + 0.665 + ], + "angle": 0, + "content": "Figure 5: State transition matrices for the states defined in Section 5.1 for the ground-truth of the train (a) and test set (b). The other transition matrices are obtained by the forecast of the current state-of-the-art model MRT [69] on the test set (c) and our model without context (d), with context via attention (e), and with context via averaging (f). The more similar the transition matrix is to the test set (b), the closer it matches the test motion." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.698, + 0.483, + 0.758 + ], + "angle": 0, + "content": "
Train[69]Ours (Γ∅)Ours (Γattn)Ours*(ΓR)
NDMS ↑-0.10150.23010.22700.2297
SSCP ↓0.09990.48390.35760.32780.3252
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.762, + 0.47, + 0.794 + ], + "angle": 0, + "content": "Table 4: Per-frame average NDMS \\(\\uparrow\\) and SSCP \\(\\downarrow\\) on the Haggling dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "information yields a slightly higher NDMS score than the other versions, but the differences are very small. However, \\(\\Gamma_{\\mathrm{attn}}\\) and \\(\\Gamma_{\\mathbb{E}}\\) forecast much more plausible social interactions (lower SSCP) as we will discuss next." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.295, + 0.895, + 0.34 + ], + "angle": 0, + "content": "
Test setMRT [69]Ours (Γ∅)Ours (Γattn)Ours* (ΓE)
1.00.32500.88120.88750.8938
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.344, + 0.894, + 0.406 + ], + "angle": 0, + "content": "Table 5: Accuracy \\(\\uparrow\\) of the buyer detection network. For MRT [69], the model randomly selects a person to be the buyer, as there is a \\(1/3\\) chance of selecting the buyer with random chance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.421, + 0.733, + 0.437 + ], + "angle": 0, + "content": "5.3.1 Social Motion Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.658 + ], + "angle": 0, + "content": "The SSCP scores are presented in Table 4 where a lower value corresponds to more plausible forecast social interactions. All our proposed variants outperform the state-of-the-art approach MRT [69]. Using no aggregation information \\((\\Gamma_{\\emptyset})\\) performs worse than the aggregation functions \\(\\Gamma_{\\mathrm{attn}}\\) (9) and \\(\\Gamma_{\\mathbb{E}}\\) (10), which is expected since the aggregation function passes information from other people in the scene. The average aggregation \\((\\Gamma_{\\mathbb{E}})\\) performs slightly better than the more complex multi-headed attention approach \\((\\Gamma_{\\mathrm{attn}})\\). We conjecture that averaging bottleneck encodings over all people introduces an inductive bias to pay the same attention to everyone, which works well for modelling the haggling game. For completeness, we also report the SSCP score of the training set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We can draw some insights about what motion each model generates by looking at the state probability transition matrices in Figure 5. For example, MRT [69] (Figure 5d) produces mostly self-loops (diagonal of transition matrix) indicating that the motion gets stuck over time. When no context information is provided \\((\\Gamma_{\\emptyset})\\), our method produces motions where all three people are talking at the same time, as can be seen in Figure 5d, where the last two entries (red arrow) in the transition matrix represent states with all three people talking. This is sensible as the model sees two sellers and only one buyer during training and thus it is more likely to produce motion that resembles a seller, who talks most of the time. When the context is provided, our approach overcomes this limitation as expected and rarely produces motion where all three people are talking at the same time as shown in Figures 5e and 5f." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9607" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.257 + ], + "angle": 0, + "content": "These observations are also confirmed when measuring the buyer detection accuracy on the forecast motion, which is reported in Table 5. The detector fails to identify the correct buyer in the sequences that are forecast by MRT [69] and it nearly chooses the buyer at random with \\(1/3\\) accuracy. This confirms that MRT does not forecast socially consistent sequences where the social role of the persons, namely buyer or seller, is preserved. In contrast, our method predicts motion where the buyer can be easily determined most of the time. As for the SSCP scores reported in Table 4, \\(\\Gamma_{\\mathbb{E}}\\) performs best." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.472, + 0.336 + ], + "angle": 0, + "content": "In summary, the aggregation function \\(\\Gamma_{\\mathbb{E}}\\) outperforms the other aggregation functions on the Haggling dataset [30] as it produces the most socially plausible motion according to our Symbolic Social Cues Protocol while also generating highly plausible 3D body motion." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.347, + 0.231, + 0.363 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.473, + 0.614 + ], + "angle": 0, + "content": "Average velocity over time Freezing or unrealistically expanding motion are common failures in human motion anticipation. While NDMS [61] penalizes in contrast to MPJPE errors in the velocity, visualizing the average motion velocity can give interesting insights. In Figure 6, we plot the average velocity over all frames for all our summary function variants, the test set, and the state-of-the-art method MRT [69]. Note that the beginning of the sequence has a higher velocity due to people walking into the scene. We observe that MRT suffers from error accumulation caused by the auto-regressive inference scheme. The velocity produced by our motion tightly follows the test set velocity for roughly 250 frames after which the test set velocity is slightly larger. We attribute this to the higher degree of stochasticity of real motion, which results in sudden jerks and swings that increase the average velocity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.734 + ], + "angle": 0, + "content": "NDMS score over time In Figure 7, we visualize how the NDMS scores of all proposed variants and MRT evolve over time. For reference, we also calculate the NDMS score of the training data which is guaranteed to be realistic. Note that NDMS is 1 for the observed part of the test sequences. As shown in the figure, our method achieves almost the same level of realism as the training data while the quality of MRT slowly degenerates over time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Anticipating more than three people We have trained and evaluated our model on the Haggling dataset where each sequence consists of triadic interaction. However, the fully convolutional nature of our approach as well as the order-invariance of the summarization function \\(\\Gamma\\) allow us to forecast any number of people. To demonstrate this capability, we predict 7 people from the Ultimatum sequence of Panoptic Studio [29] using only the model trained on the Haggling dataset. This works well because the Haggling and Ultimatum sequences share many social behaviors, such as turn-taking, talking, and paying attention while" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.891, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.28, + 0.895, + 0.342 + ], + "angle": 0, + "content": "Figure 6: Average velocity over time for the entire test motion \\(\\mathcal{X}\\) and generated motions \\(\\hat{\\mathcal{X}}\\). The x-axis represents the frames while the y-axis represents the (log) average velocity of the data." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.346, + 0.891, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.57 + ], + "angle": 0, + "content": "Figure 7: Average NDMS [61] score \\(\\uparrow\\) of all proposed variants, MRT, and the train set over time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.586, + 0.895, + 0.662 + ], + "angle": 0, + "content": "standing in a circle. Our results can be seen in Figure 4 where our model is able to predict realistic motion for 7 people, even though it was trained only on the triadic Haggling dataset. More results are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.678, + 0.62, + 0.695 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this work, we present Social Diffusion, a stochastic multi-person motion anticipation model. The approach not only forecasts realistic motions on the individual level, but also plausible social interactions where the social roles of individuals are preserved over time. The approach is very flexible. It can be used for short and long-term forecasting and can be applied to larger groups than observed during training. As a second contribution, we proposed a new evaluation protocol to measure the realism of forecast social interactions. We furthermore derived a dataset for multi-person social interaction forecasting from the Haggling dataset [30] where the persons have different social roles that impact their behavior. We evaluated our approach" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9608" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "on four multi-person datasets and demonstrated that our approach outperforms the state-of-the-art for short-term and long-term anticipation both in realism of forecast motion and social interaction. The approach has still some limitations. For instance, the global positions of the root joints can be better estimated. Future directions also include extending the model to predict motions of dynamic groups of people, e.g., at a cocktail party where any individual can freely disengage from the current conversation group and join another one." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.261, + 0.238, + 0.278 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.287, + 0.473, + 0.455 + ], + "angle": 0, + "content": "Juergen Gall has been supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) GA 1927/4-2 (FOR 2535 Anticipating Human Behavior), by the Federal Ministry of Education and Research (BMBF) under grant no. 01IS22094E WEST-AI, the project iBehave (receiving funding from the programme \"Netzwerke 2021\", an initiative of the Ministry of Culture and Science of the State of Northrhine Westphalia), and the ERC Consolidator Grant FORHUE (101044724). The sole responsibility for the content of this publication lies with the authors." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.47, + 0.175, + 0.487 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.497, + 0.362, + 0.511 + ], + "angle": 0, + "content": "[1] CMU. Carnegie-Mellon Mocap Database." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.513, + 0.469, + 0.569 + ], + "angle": 0, + "content": "[2] Vida Adeli, Ehsan Adeli, Ian Reid, Juan Carlos Niebles, and Hamid Rezatofighi. Socially and contextually aware human motion and pose forecasting. Robotics and Automation Letters, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.572, + 0.469, + 0.642 + ], + "angle": 0, + "content": "[3] Vida Adeli, Mahsa Ehsanpour, Ian Reid, Juan Carlos Niebles, Silvio Savarese, Ehsan Adeli, and Hamid Rezatofighi. Tripod: Human trajectory and pose dynamics forecasting in the wild. In International Conference on Computer Vision, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.644, + 0.469, + 0.697 + ], + "angle": 0, + "content": "[4] Emre Aksan, Manuel Kaufmann, Peng Cao, and Otmar Hilliges. A spatio-temporal transformer for 3d human motion prediction. In International Conference on 3D Vision, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.702, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[5] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social lstm: Human trajectory prediction in crowded spaces. In Conference on Computer Vision and Pattern Recognition, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.773, + 0.469, + 0.841 + ], + "angle": 0, + "content": "[6] Sadegh Aliakbarian, Fatemeh Sadat Saleh, Mathieu Salzmann, Lars Petersson, and Stephen Gould. A stochastic conditioning scheme for diverse human motion prediction. In Conference on Computer Vision and Pattern Recognition, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.845, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[7] Eduardo Alvarado, Damien Rohmer, and Marie-Paule Cani. Generating upper-body motion for real-time characters making their way through dynamic environments. Eurographics Symposium on Computer Animation, 2022." + }, + { + "type": "list", + "bbox": [ + 0.086, + 0.497, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[8] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.151, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[9] Timur Bagautdinov, Alexandre Alahi, François Fleuret, Pascal Fua, and Silvio Savarese. Social scene understanding: End-to-end multi-person action localization and collective activity recognition. In Conference on Computer Vision and Pattern Recognition, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[10] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.28, + 0.892, + 0.336 + ], + "angle": 0, + "content": "[11] Judith Butepage, Michael J Black, Danica Kragic, and Hedvig Kjellstrom. Deep representation learning for human motion prediction and classification. In Conference on Computer Vision and Pattern Recognition, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.338, + 0.892, + 0.407 + ], + "angle": 0, + "content": "[12] Yujun Cai, Lin Huang, Yiwei Wang, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Xu Yang, Yiheng Zhu, Xiaohui Shen, et al. Learning progressive joint propagation for human motion prediction. In European Conference on Computer Vision, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.409, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[13] Yujun Cai, Yiwei Wang, Yiheng Zhu, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Chuanxia Zheng, Sijie Yan, Henghui Ding, et al. A unified 3d human motion synthesis model via conditional variational auto-encoder. In International Conference on Computer Vision, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.482, + 0.892, + 0.55 + ], + "angle": 0, + "content": "[14] Lujing Chen, Rui Liu, Xin Yang, Dongsheng Zhou, Qiang Zhang, and Xiaopeng Wei. Sttg-net: a spatio-temporal network for human motion prediction based on transformer and graph convolution network. Visual Computing for Industry, Biomedicine, and Art, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.595 + ], + "angle": 0, + "content": "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.598, + 0.892, + 0.639 + ], + "angle": 0, + "content": "[16] Christian Diller, Thomas Funkhouser, and Angela Dai. Forecasting characteristic 3d poses of human actions. Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.641, + 0.892, + 0.683 + ], + "angle": 0, + "content": "[17] Dominik Maria Endres and Johannes E Schindelin. A new metric for probability distributions. Transactions on Information theory, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.686, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[18] Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models for human dynamics. In International Conference on Computer Vision, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.73, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[19] Bent Fuglede and Flemming Topsoe. Jensen-Shannon divergence and Hilbert space embedding. In International Symposium on Information Theory, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[20] Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In Conference on Computer Vision and Pattern Recognition, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[21] Anand Gopalakrishnan, Ankur Mali, Dan Kifer, Lee Giles, and Alexander G Ororbia. A neural temporal model for human motion prediction. In Conference on Computer Vision and Pattern Recognition, 2019." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9609" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[22] Liang-Yan Gui, Yu-Xiong Wang, Xiaodan Liang, and José MF Moura. Adversarial geometry-aware human motion prediction. In European Conference on Computer Vision, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[23] Wen Guo, Xiaoyu Bie, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Multi-person extreme motion prediction. In Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[25] Daniel Holden, Jun Saito, Taku Komura, and Thomas Joyce. Learning motion manifolds with convolutional autoencoders. In SIGGRAPH Asia technical briefs. 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.289, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[26] Emiel Hoogeboom, Didrik Nielsen, Priyank Jaini, Patrick Forre, and Max Welling. Argmax flows and multinomial diffusion: Learning categorical distributions. Advances in Neural Information Processing Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.4 + ], + "angle": 0, + "content": "[27] Ashesh Jain, Amir R Zamir, Silvio Savarese, and Ashutosh Saxena. Structural-rnn: Deep learning on spatio-temporal graphs. In Conference on Computer Vision and Pattern Recognition, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[28] Michael Janner, Yilun Du, Joshua B Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.471, + 0.525 + ], + "angle": 0, + "content": "[29] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic Studio: A Massively Multiview System for Social Motion Capture. In International Conference on Computer Vision, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.471, + 0.58 + ], + "angle": 0, + "content": "[30] Hanbyul Joo, Tomas Simon, Mina Cikara, and Yaser Sheikh. Towards social artificial intelligence: Nonverbal social signal prediction in a triadic interaction. In Conference on Computer Vision and Pattern Recognition, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.471, + 0.664 + ], + "angle": 0, + "content": "[31] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Scott Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. Transactions on Pattern Analysis and Machine Intelligence, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.471, + 0.72 + ], + "angle": 0, + "content": "[32] Julia Kantorovitch, Janne Väre, Vesa Pehkonen, Arto Laikari, and Heikki Seppälä. An assistive household robot-doing more than just cleaning. Journal of Assistive Technologies, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.471, + 0.775 + ], + "angle": 0, + "content": "[33] Anne Keitel, Moritz M Daum, et al. The use of intonation for turn anticipation in observed conversations without visual signals as source of information. Frontiers in psychology, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.471, + 0.832 + ], + "angle": 0, + "content": "[34] Lukasz Kidziński, Bryan Yang, Jennifer L Hicks, Apoorva Rajagopal, Scott L Delp, and Michael H Schwartz. Deep neural networks enable quantitative movement analysis using single-camera videos. Nature communications, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[35] Vineet Kosaraju, Amir Sadeghian, Roberto Martin-Martin, Ian Reid, Hamid Rezatofighi, and Silvio Savarese. Socialbigat: Multimodal trajectory forecasting using bicycle-gan and graph attention networks. Advances in Neural Information Processing Systems, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.135 + ], + "angle": 0, + "content": "[36] Stephen C Levinson and Francisco Torreira. Timing in turn-taking and its implications for processing models of language. Frontiers in psychology, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.893, + 0.192 + ], + "angle": 0, + "content": "[37] Chen Li, Zhen Zhang, Wee Sun Lee, and Gim Hee Lee. Convolutional sequence to sequence model for human dynamics. In Conference on Computer Vision and Pattern Recognition, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.893, + 0.25 + ], + "angle": 0, + "content": "[38] Thomas Lucas*, Fabien Baradel*, Philippe Weinzaepfel, and Grégory Rogez. Posegpt: Quantization-based 3d human motion generation and forecasting. In European Conference on Computer Vision, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.251, + 0.893, + 0.308 + ], + "angle": 0, + "content": "[39] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.31, + 0.893, + 0.353 + ], + "angle": 0, + "content": "[40] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. History repeats itself: Human motion prediction via motion attention. In European Conference on Computer Vision, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.354, + 0.893, + 0.41 + ], + "angle": 0, + "content": "[41] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies for human motion prediction. International Conference on Computer Vision, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.411, + 0.893, + 0.467 + ], + "angle": 0, + "content": "[42] Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Conference on Computer Vision and Pattern Recognition, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.893, + 0.539 + ], + "angle": 0, + "content": "[43] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In International Conference on 3D Vision, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.541, + 0.893, + 0.597 + ], + "angle": 0, + "content": "[44] Ronja Möller, Antonino Furnari, Sebastiano Battiato, Aki Härma, and Giovanni Maria Farinella. A survey on human-aware robot navigation. Robotics and Autonomous Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.6, + 0.893, + 0.656 + ], + "angle": 0, + "content": "[45] Brendan Tran Morris and Mohan Manubhai Trivedi. A survey of vision-based trajectory learning and analysis for surveillance. Transactions on Circuits and Systems for Video Technology, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.893, + 0.714 + ], + "angle": 0, + "content": "[46] Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. Learning to listen: Modeling non-deterministic dyadic facial motion. In Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.715, + 0.893, + 0.799 + ], + "angle": 0, + "content": "[47] Sangmin Oh, Anthony Hoogs, Amitha Perera, Naresh Cunteor, Chia-Chih Chen, Jong Taek Lee, Saurajit Mukherjee, JK Aggarwal, Hyungtae Lee, Larry Davis, et al. A large-scale benchmark dataset for event recognition in surveillance video. In Conference on Computer Vision and Pattern Recognition, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.801, + 0.893, + 0.857 + ], + "angle": 0, + "content": "[48] Dario Pavllo, Christoph Feichtenhofer, Michael Auli, and David Grangier. Modeling human motion with quaternion-based neural networks. International Journal of Computer Vision, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.858, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[49] Dario Pavllo, David Grangier, and Michael Auli. *Quaternion: A quaternion-based recurrent model for human motion*. In *British Machine Vision Conference*, 2018." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "9610" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[50] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. European Conference on Computer Vision, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[51] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.194, + 0.468, + 0.249 + ], + "angle": 0, + "content": "[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.252, + 0.468, + 0.307 + ], + "angle": 0, + "content": "[53] Alejandro Hernandez Ruiz, Juergen Gall, and Francesc Moreno-Noguer. Human motion prediction via spatiotemporal inpainting. International Conference on Computer Vision, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.31, + 0.468, + 0.364 + ], + "angle": 0, + "content": "[54] Harvey Sacks, Emanuel A Schegloff, and Gail Jefferson. A simplest systematics for the organization of turn taking for conversation. In Studies in the organization of conversational interaction. 1978." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.367, + 0.468, + 0.422 + ], + "angle": 0, + "content": "[55] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In Transactions on Graphics, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.425, + 0.468, + 0.479 + ], + "angle": 0, + "content": "[56] Keren Shavit-Cohen and Elana Zion Golumbic. The dynamics of attention shifts among concurrent speech in a naturalistic multi-speaker virtual environment. Frontiers in Human Neuroscience, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.483, + 0.468, + 0.538 + ], + "angle": 0, + "content": "[57] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.541, + 0.468, + 0.595 + ], + "angle": 0, + "content": "[58] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.599, + 0.468, + 0.653 + ], + "angle": 0, + "content": "[59] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.656, + 0.468, + 0.711 + ], + "angle": 0, + "content": "[60] Chen Sun, Abhinav Shrivastava, Carl Vondrick, Rahul Sukthankar, Kevin Murphy, and Cordelia Schmid. Relational action forecasting. In Conference on Computer Vision and Pattern Recognition, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.714, + 0.468, + 0.755 + ], + "angle": 0, + "content": "[61] Julian Tanke, Chintan Zaveri, and Juergen Gall. Intention-based long-term human motion anticipation. In International Conference on 3D Vision, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.758, + 0.468, + 0.799 + ], + "angle": 0, + "content": "[62] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.802, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[63] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 2017." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[65] Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. International Conference on Learning Representations, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.15, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[66] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European Conference on Computer Vision, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[67] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European conference on computer vision, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[68] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[69] Jiashun Wang, Huazhe Xu, Medhini Narasimhan, and Xiaolong Wang. Multi-person 3d motion prediction with multirange transformers. Advances in Neural Information Processing Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.376, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[70] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In Conference on Computer Vision and Pattern Recognition, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[71] Ye Yuan and Kris Kitani. DLOW: Diversifying latent flows for diverse human motion prediction. In European Conference on Computer Vision, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[72] Jason Y Zhang, Panna Felsen, Angjoo Kanazawa, and Jitendra Malik. Predicting 3d human dynamics from video. In International Conference on Computer Vision, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[73] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.575, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[74] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J Black, and Siyu Tang. Generating 3d people in scenes without people. In Conference on computer vision and pattern recognition, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[75] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3d scenes. In Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9611" + } + ] +] \ No newline at end of file diff --git a/2023/Social Diffusion_ Long-term Multiple Human Motion Anticipation/e840bc61-69a6-4f9d-9e0d-f25143903533_origin.pdf b/2023/Social Diffusion_ Long-term Multiple Human Motion Anticipation/e840bc61-69a6-4f9d-9e0d-f25143903533_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6462a78c2c0cd823dd92af197972239863c9c93c --- /dev/null +++ b/2023/Social Diffusion_ Long-term Multiple Human Motion Anticipation/e840bc61-69a6-4f9d-9e0d-f25143903533_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9a2fe920705a4ed3b73e36416cf4d976c75ecc64310f85c6ab71f4b15413eb6 +size 2733424 diff --git a/2023/Social Diffusion_ Long-term Multiple Human Motion Anticipation/full.md b/2023/Social Diffusion_ Long-term Multiple Human Motion Anticipation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..793f62a63d6f852029641da6f72fd7ce23633b41 --- /dev/null +++ b/2023/Social Diffusion_ Long-term Multiple Human Motion Anticipation/full.md @@ -0,0 +1,399 @@ +# Social Diffusion: Long-term Multiple Human Motion Anticipation + +Julian Tanke\*, Linguang Zhang?, Amy Zhao?, Chengcheng Tang?, Yujun Cai?, Lezi Wang?, Po-Chen Wu?, Juergen Gall1,3, and Cem Keskin2 + +1University of Bonn 2Reality Labs Research +3Lamarr Institute for Machine Learning and Artificial Intelligence + +{tanke,gall}@iai.uni-bonn.de {linguang,xamyzhao,chengcheng.tang,yujuncai,wanglezi,pochenwu,cemkeskin}@meta.com + +# Abstract + +We propose Social Diffusion, a novel method for short-term and long-term forecasting of the motion of multiple persons as well as their social interactions. Jointly forecasting motions for multiple persons involved in social activities is inherently a challenging problem due to the interdependencies between individuals. In this work, we leverage a diffusion model conditioned on motion histories and causal temporal convolutional networks to forecast individually and contextually plausible motions for all participants. The contextual plausibility is achieved via an order-invariant aggregation function. As a second contribution, we design a new evaluation protocol that measures the plausibility of social interactions which we evaluate on the Haggling dataset, which features a challenging social activity where people are actively taking turns to talk and switching their attention. We evaluate our approach on four datasets for multi-person forecasting where our approach outperforms the state-of-the-art in terms of motion realism and contextual plausibility. + +# 1. Introduction + +Understanding and anticipating social interactions in groups of people is a challenging and highly relevant topic [35, 9, 32, 45, 47, 5, 34]. For instance, it is essential for socially-compliant robots [44], but it is also relevant for neuroscience and social sciences since it allows to develop computational models on how the behavior of other persons is perceived and how it changes the own behavior. + +Forecasting realistic social interactions, however, is very challenging for two reasons. First, social interactions tend to last for tens of seconds [33] or even minutes - much longer than the prediction from most of the existing human + +motion anticipation models [23, 69, 3, 22, 27, 42, 49, 11, 25, 37, 4, 14, 41, 40]. Second, social interactions consist of interdependent motions [54, 36], which requires modeling the relationships among all individuals. For example, in conversational turn-taking, a person's turn to talk highly depends on the start/end of the others' speaking. While multi-person motion anticipation has emerged as a new topic, current approaches [23, 69, 3] do not pay much attention on complex social interactions. For instance, they do not preserve the social role of individuals in a group such that the interactions become socially implausible over time. + +To address the limitations of existing models, we propose Social Diffusion to predict motions of multiple people and ensure contextually plausible interactions, as shown in Fig. 1. To this end, we learn the distribution of human motion by leveraging a diffusion model [39, 28, 52, 58, 24, 59, 62, 73]. To enforce information exchange among people, which is critical to predicting contextually plausible interactions, we introduce an order-invariant aggregation function to aggregate motion features from all people. For inference, we feed back the input sequence to the signal during the reverse-diffusion steps to condition the motion generation on the past motion. Our method is fully convolutional which allows us to generate sequences of arbitrary size. This allows us to not just forecast the next few seconds of an input motion but also to forecast social interactions that last longer. Furthermore, our approach is very flexible in the sense that the number of persons during training and inference can differ. To the best of our knowledge, our approach represents the first diffusion model that produces multi-person motions at the same time. + +As a second contribution, we propose a new evaluation protocol for social interactions based on Symbolic Social Cues, which measures whether the forecast motion is socially plausible. Our key observation is that the probabilities of transitions between social interaction states are highly + +![](images/ce52e1e5601acd91e4fbea13d22232ae67f1981ecc07eda16150bcf2b7facdb3.jpg) +Figure 1: We propose an approach for multi-person motion anticipation: given a sequence of human social interactions (blue/red skeletons), the proposed model forecasts multi-person motions where the social roles are preserved and the interactions are socially plausible. + +correlated with the plausibility of predicted social interactions. In a conversation, for example, a person usually starts talking only when a peer stops talking. To evaluate predicted motions, we first build the state transition graph by extracting states from the motions. We then treat the state transition graph as a probability distribution and compare it to the real data distribution. + +For evaluation, we use the Haggling dataset [30] which comprises 175 videos of well-defined triadic social interactions. In contrast to other existing multi-person human motion datasets [43, 66, 69], the persons have different social roles that impact their behavior. We furthermore evaluate our approach on the MuPoTS-3D [43], 3DPW [67], and CMU-Mocap [1, 69] dataset. On all four datasets, our approach outperforms the state of the art for multi-person human motion forecasting. + +In summary, our contribution is two-fold: + +1. We propose Social Diffusion, the first stochastic multiperson motion anticipation model that outperforms the state of the art on common multi-person motion anticipation datasets. +2. We propose a novel social interaction evaluation protocol that considers not only the validity of poses but also the plausibility of social interactions. + +# 2. Related Work + +Single Person Human Motion Prediction Human motion prediction, which typically refers to generating motion sequences given a prefix segment, has been extensively studied in the past few years. Most recent papers focus on single person 3D predictions. Due to the inherent temporal nature of human motion, various temporal-based methods, such as recurrent neural networks (RNNs) [18, 22, 27, 42, 49, 48, 61, 72], temporal convolutional networks [11, 25, 37], transformers [4, 14, 12] and graph neural networks (GNNs) [14, 41, 40] have been used for this task. For long-term predictions, auto-regressive models that operate in the discrete space [46, 38, 61] have shown success, where a long prediction sequence can be obtained without + +converging to mean poses. Recently, anchor-based methods [16, 68] have been proposed; these focus on forecasting characteristic anchor poses rather than the entire sequence auto-regressively. To achieve better interactions with environments, [7, 70, 68, 74] proposed various ways to include contextual information into human motion predictions. Since generated motions can be controlled with a high-level guidance such as action class or text, some approaches [50, 13] used Variational Auto-Encoders [58] to solve this problem. Going beyond single human prediction, we predict the motions and interactions of multiple humans. + +Multi-person Anticipation Modeling multi-person interactions has been a long standing problem [46, 23, 2, 65, 20, 30, 46, 63, 75]. For instance, $\mathrm{DR}^2\mathrm{N}$ [60] predicts the activities of multiple people given a past video sequence. For a given frame, personal relationships between candidates are estimated using a graph attention network (GAT) [65] while temporal relationships are predicted using recurrent neural networks. Recently, Wang et al. [69] addressed multi-person 3D motion trajectory prediction via a Multi-Range Transformers framework. Guo et al. [23] introduced Cross-Interaction Attention to jointly model highly expressive dance sequences. Joo et al. [30] introduced a triadic haggling game for social signal prediction, based on Panoptic Studio [29, 31]. In their work they predict the motion of a single person given the motion of the others. Similarly, [46] predicts motions based on the other actors' behavior. Crucially, both methods only predict the motion of a single individual, given other persons social signals. + +Diffusion Models Diffusion models [24, 57, 59] belong to the family of probabilistic generative models, which convert the training data successively to Gaussian noise, and then learn to recover the data by reversing this noising process. Diffusion models have emerged as powerful deep generative models with breakthroughs in many applications, including image synthesis [24, 15], segmentation [10], and natural language processing [8, 26]. For conditioned generation, [15] introduced classifier-guided diffusion, and [55] takes the inpainted images as denoised by the model. More recently, [62, 73] have suggested diffusion models for motion generation; however, they are limited to single human + +![](images/cbf1d81ef2fc39b6828648edcfe6213bf3769da7a72fb791a16c4d47cd156a1a.jpg) +Figure 2: Model overview: the reverse diffusion process $g$ consists of a fully convolutional causal encoder $e$ and a fully convolutional causal decoder $d$ that produces a denoised motion sequence $\hat{\mathbf{x}}_0^i \equiv \hat{\mathbf{x}}_0^{i,1:N}$ for person $i$ , given the bottleneck state $h^i$ and the aggregation function $\Gamma(\{\mathbf{h}_j \mid j \in 1, \dots, p\})$ over all people in the scene. + +3D motion prediction. To the best of our knowledge, we are the first to build a stochastic multi-person motion anticipation model that can predict very long-term future motions. + +# 3. Social Diffusion Model + +As illustrated in Fig. 1, we aim to forecast the motion of multiple persons that interact with each other. The forecast motion should be realistic and socially plausible. For instance, not all persons should talk at the same time. Formally, we represent a human motion sequence with $p$ people of length $N$ as $\mathbf{X}^{1:N} \in \mathbb{R}^{N \times p \times \delta}$ where $\delta$ represents the dimension of the individual pose vector at a given frame. Our goal is then to predict the future motion $\hat{\mathbf{X}}^{n+1:N}$ for all people, given their past motions $\mathbf{X}^{1:n}$ : + +$$ +\hat {\mathbf {X}} ^ {n + 1: N} = \operatorname {S D M} \left(\mathbf {X} ^ {1: n}\right). \tag {1} +$$ + +Before we describe the proposed Social Diffusion Model (SDM) in Section 3.2, we will briefly describe a generic diffusion model [24] in Section 3.1. In Section 4, we will then introduce the social interaction evaluation protocol. + +# 3.1. Diffusion Model + +A latent representation $\mathbf{X}_T\sim \mathcal{N}(\mathbf{0},\mathbf{1})$ is obtained via a $T$ step Markov Gaussian noising process $q(\mathbf{X}_T|\mathbf{X}_0)$ where $\mathbf{X}_0\equiv \mathbf{X}^{1:N}$ is a real motion sequence from the training set. The Markov Gaussian noising process can be written in closed form: + +$$ +q \left(\mathbf {X} _ {t} \mid \mathbf {X} _ {0}\right) = \mathcal {N} \left(\mathbf {X} _ {t}; \sqrt {\alpha_ {t}} \mathbf {X} _ {0}, (1 - \alpha_ {t}) \mathbf {I}\right) \tag {2} +$$ + +where $\alpha_{t} \in (0,1)$ is a step-dependent fixed hyperparameter. To sample from the generative model, we learn + +to invert the noising step using the generator function $g$ : + +$$ +\hat {\mathbf {X}} _ {0, t} = g \left(\mathbf {X} _ {t}, t\right) \tag {3} +$$ + +The key contribution of our model is the novel generator function $g$ , which models social interactions over time and will be described in Section 3.2. Following [24, 62, 51], the loss during training is defined by: + +$$ +\mathcal {L} = \mathbb {E} _ {\mathbf {X} _ {0} \sim p (\mathbf {X}), t \sim [ 1, T ]} \left[ \left| \left| \mathbf {X} _ {0} - g \left(\mathbf {X} _ {t}, t\right) \right| \right| \right] \tag {4} +$$ + +For inference, we reverse-iterate over Equation (3), starting at sampling step $T$ and latent representation $\mathbf{X}_T\sim \mathcal{N}(\mathbf{0},\mathbf{1})$ . At each iteration $t$ , we slowly denoise the motion sequence using (2) and (3): + +$$ +\hat {\mathbf {X}} _ {0, t - 1} = g \left(q \left(\hat {\mathbf {X}} _ {t - 1} \mid \hat {\mathbf {X}} _ {0, t}\right), t - 1\right) \tag {5} +$$ + +The final denoised motion is obtained when $t = 1$ . + +# 3.2. Multi-person Motion Generator + +The diffusion model described in Section 3.1 produces unconditioned motion. In order to use the model for motion forecasting, we need to condition the model on the observed motion sequence of all persons $\mathbf{X}_0^{1:n} = \mathbf{X}^{1:n}$ . To this end, we modify the inference sampling (5) to also include past motion as follows: + +$$ +\hat {\mathbf {X}} _ {0, t - 1} = g \left(q \left(\hat {\mathbf {X}} _ {t - 1} \mid \mathbf {X} ^ {1: n} \cup \hat {\mathbf {X}} _ {0, t} ^ {n + 1: N}\right), t - 1\right) \tag {6} +$$ + +The reverse diffusion process $g(\mathbf{X}_t, t)$ consists of three components, a causal temporal convolutional encoder $e$ , a causal temporal convolutional decoder $d$ , and an order-invariant function $\Gamma$ that aggregates the interaction of the different persons, see Figure 2. + +The encoder $e$ and the decoder $d$ process each individual sequence independently while $\Gamma$ ensures that information flows between all persons in the scene. Formally, given $\mathbf{x}_t^i = \hat{\mathbf{X}}_t^{i,1:N} \in \mathbb{R}^{N \times \delta}$ , which is the motion sequence for a single person $i$ at scheduled noisng step $t$ , we obtain for each person $i$ in the scene a bottleneck encoding $\mathbf{h}^i$ : + +$$ +\mathbf {h} ^ {i} = e \left(\mathbf {x} _ {t} ^ {i}, t\right) \forall i \in 1, \dots , p. \tag {7} +$$ + +The encoder $e$ consists of a 4-layer temporal convolutional network where each layer progressively reduces the temporal resolution by half via striding. In each layer, the noising step $t$ is fed via sinusoidal positional encoding [64]. To produce the denoised motion $\hat{\mathbf{x}}_0^i$ , the decoder $d$ can be utilized as follows: + +$$ +\hat {\mathbf {x}} _ {0} ^ {i} = d \left(\mathbf {h} ^ {i}, t, \Gamma \left(\mathbf {h} ^ {j} \forall j \in 1, \dots , p\right)\right) \tag {8} +$$ + +where $d$ is a 4-layer temporal convolutional network and each layer progressively doubles the temporal resolution via + +linear upsampling. As for the encoder, the noising step $t$ is fed via sinusoidal positional encoding to each layer. In addition, the output of the order-invariant aggregation function $\Gamma$ is concatenated to $\mathbf{h}^{\mathrm{i}}$ before passing it to the first convolutional layer. The estimated motion sequences $\hat{\mathbf{x}}_0^i$ of each person $i$ at noising step $t$ are then concatenated to obtain $\hat{\mathbf{X}}_{0,t}$ and the approach proceeds to the next step $t - 1$ . + +The order-invariant aggregation function $\Gamma$ passes information from other people in the scene. In our experiments, we evaluate two aggregation functions, averaging $(\Gamma_{\mathbb{E}})$ over all people and multi-headed attention [64] $(\Gamma_{\mathrm{attn}})$ : + +$$ +\Gamma_ {\mathrm {a t t n}} (\mathbf {H}) = \operatorname {M u l t i H e a d} (\mathbf {H}) \tag {9} +$$ + +$$ +\Gamma_ {\mathbb {E}} (\mathbf {H}) = \frac {1}{p} \sum_ {i = 1} ^ {p} \mathbf {h} ^ {i} \tag {10} +$$ + +where $\mathbf{H} = \{h^i\}_{i=1}^p$ are the embeddings of all layers of the encoder and for all people in the scene. MultiHead $(\mathbf{H})$ calculates the self-attention across all people for a given frame. + +# 3.3. Implementation Details + +We follow state-of-the-art diffusion models [24, 39, 62] and use the cosine variance schedule. We set the number of diffusion steps to $T = 1000$ . The encoder $e$ consists of four layers of convolutional blocks with kernel size 3 and stride 2. The decoder $d$ consists of four layers of convolutional blocks with additional upsampling layers that upsample the input sequence by factor two using linear interpolation. We standardize the training data to have zero mean and standard deviation one. We normalize all poses by splitting pose and global translation: each pose is transformed into a hipcentric coordinate frame and the pose is concatenated with the global rotation and translation to form a $\delta$ dimensional pose vector. + +# 4. Symbolic Social Cues Protocol + +We are interested in anticipating the social interactions among multiple people. Multi-person social interactions consist of several intricate and complex behaviours such as paying attention to a specific person [56] and turn-taking [54, 36], which usually take tens of seconds or even minutes. Current state-of-the-art multi-person motion anticipation methods [3, 69] calculate the Mean Per Joint Positional Error (MPJPE) using the ground-truth sequence, which is only meaningful for short time horizons of around one second [6, 21, 53, 61, 71]. More important, however, is that it does not measure the realism of social interaction. + +We thus propose the Symbolic Social Cues Protocol (SSCP), which divides the social interactions into a set of discrete interaction classes. In SSP, we define a social signal function + +$$ +C ^ {1: N} = s \left(\mathbf {X} ^ {1: N}\right) \tag {11} +$$ + +which takes as input a multi-person motion sequence $\mathbf{X}^{1:N}\in \mathbb{R}^{N\times p\times \delta}$ and produces a discrete symbolic representation $C^{1:N} = \{c_n\}_{n = 1}^N$ , where $c_{n}\in \{1,\dots,m\}$ and $m$ represents the total number of symbolic states. A symbolic state is a unique summary of the current state of interaction, e.g., a person is talking and another person is listening. Given a test set $\mathcal{X} = \{\mathbf{X}_i^{1:N}\}_{i = 1}^K$ with $K$ sequences, we can now calculate the probability distribution $p_{\mathrm{sscp}}$ over the social state transitions. + +$$ +p _ {\mathrm {s s c p}} (\mathcal {X}) = \frac {1}{\zeta} \sum_ {i = 1} ^ {K} \operatorname {s t m} \left(s \left(\mathbf {X} _ {i} ^ {1: N}\right)\right) \tag {12} +$$ + +where $\mathrm{stm}(C_i^{1:N})$ produces the $m\times m$ state transition matrix for the discrete symbolic sequence $C^{1:N}$ and $\zeta = \sum_{i=1}^{K}\sum_{m',m''}\mathrm{stm}(C^{1:N})_{m',m''}$ is a normalization constant to ensure that $p_{\mathrm{sscp}}$ is a valid probability distribution. + +To evaluate a motion anticipation model $f$ , we predict the future motion for all $K$ test sequences from a fixed start frame $n$ until the end of the sequence $N$ : + +$$ +\hat {\mathcal {X}} ^ {n + 1: N} = \left\{f \left(\mathbf {X} _ {i} ^ {1: n}\right) \right\} _ {i = 1} ^ {K} \tag {13} +$$ + +We can now calculate the distance between the generated and ground-truth social motion distribution: + +$$ +D _ {\mathrm {J S D}} \left(p _ {\text {s s c p}} \left(\mathcal {X} ^ {n: N}\right), p _ {\text {s s c p}} \left(\hat {\mathcal {X}} ^ {n: N}\right)\right). \tag {14} +$$ + +$D_{\mathrm{JSD}}$ is the squared Jensen-Shannon distance [17, 19]: + +$$ +D _ {\mathrm {J S D}} (p \| q) = \sqrt {\frac {\left(D _ {\mathrm {K L}} (p \| \frac {p + q}{2}) + D _ {\mathrm {K L}} (q \| \frac {p + q}{2})\right)}{2}} \tag {15} +$$ + +where $p$ and $q$ are probability distributions and $D_{\mathrm{KL}}$ is the Kullback-Leibler divergence. Note that we compare the generated motion $\hat{\mathcal{X}}^{n + 1:N}$ to the test set $\mathcal{X}^{n + 1:N}$ with the same start frame as the data distribution might shift across time. + +# 5. Experiments + +We evaluate our approach on four datasets. Following [69], we report the Mean Per Joint Positional Error (MPJPE) in global and local aligned coordinates for the multi-person human motion datasets MuPoTS-3D [43], 3DPW [67] and CMU-Mocap [1, 69]. MuPoTS-3D [43] contains recordings of 2 to 3 persons in workout settings. Interactions between the persons are rare. 3DPW [67] contains recordings of 1 to 2 persons and the sequences cover a wide range of different activities. The level of interactions range from no interaction and little interaction, like two persons walking, to close interactions such as dancing. CMU [69] combines the motion of different sequences from the CMU-Mocap dataset [1]. The composition of motion + +![](images/6b9b39020ac79ca53eaacf698fb15ac23c04088ea147a007a04469452421f5d0.jpg) +(a) + +![](images/1c897da6acc91b72d6a44cd9612768f56ff0ed37df0ed6489ee1db582f7ba97f.jpg) +Figure 3: A sample frame from the Haggling dataset [30] for evaluating social interactions. (a): 3D poses in a haggling sequence. Blue limbs represent the left body side while red limbs represent the right body side. The buyer's attention is indicated as green arrow. (b): a sample video frame from a haggling sequence. + +
MethodCMU-MocapMuPoTS-3D3DPW
1s2s3s1s2s3s1s2s3s
LTD [41]1.372.193.261.191.812.344.677.108.71
HRI [40]1.492.603.070.941.682.294.076.328.01
SP [2]1.152.713.900.921.672.514.177.179.27
MRT [69]0.961.572.180.891.592.223.876.127.83
Ours0.741.061.341.151.291.441.642.723.55
+ +Table 1: MPJPE $\downarrow$ in dm on different datasets. + +sequences, however, does not reflect realistic interactions. For these datasets, 1 second of human motion is observed and 1-3 seconds need to be forecast. As our model is generative, we sample 8 samples for each test sequence and report the average over all 8 samples. The source code including trained evaluation models and the dataset transformation script is publicly available1. + +# 5.1. Haggling Forecasting Dataset + +Since the three datasets contain multiple persons, but very few social interactions, we prepared a new dataset for multi-person forecasting in the context of social interactions. For this, we utilize the Haggling dataset [30] where 122 participants play a social game with two sellers trying to sell their products to a buyer. Each game lasts one minute and contains interesting triadic interactions such as turn-taking and attention changes. A sample scene is shown in Figure 3. The dataset consists of 135 training sequences and 40 test sequences, sampled at $30\mathrm{Hz}$ . Some of the 3D poses in the dataset are noisy as they have been estimated [29, 31] and we manually correct them. In total, the dataset consists of 234, 907 training and 69, 951 test frames, each with three people. For more details, please see the supplementary material. + +For evaluation, we take the first $10\%$ of a sequence + +as observation and forecast the remaining $90\%$ of the sequence, but we also evaluate the motion at intermediate frames ranging from frame 1 to frame 1300. It needs to be noted that long-term forecasting is highly relevant for neuroscience and social sciences since it allows to develop computational models on how the behavior of other persons is perceived and how it changes the own behavior. For measuring the plausibility of the forecast motion of each individual, we use the Normalized Directional Motion Similarity (NDMS) [61] since the measure not only considers static poses but also the motion of the forecast sequence. Furthermore, it can be applied to sequences of any length. NDMS, however, does not measure if the social interactions are plausible. + +For evaluating the social motion quality, we utilize the proposed Symbolic Social Cues Protocol described in Section 4. To this end, we need to specify the classes of social interactions. A haggling activity is composed of the sellers trying to convince the buyer to purchase their products and the buyer switching attention between the sellers. Throughout the game, certain social patterns emerged [56, 36]: + +1. Most of the time, only a single person speaks +2. For almost all frames, at least one person is talking +3. The sellers speak roughly the same amount of time while the buyer seldom talks +4. The buyer pays attention (looks at) to whoever talks +5. The sellers take turns to speak but sometimes interrupt each other + +Given the well-defined structure of the task and the emerging social behaviors, we reduce the haggling game to two key signals: + +- talking: defines who is talking +- attention: defines who of the two sellers has the buyer's attention + +Given all possible combinations (e.g., both sellers can talk at the same time, or nobody talks), we end up with 16 possible states for each frame and formulate the social interactions as a symbolic representation over time. Note that we have to distinguish between left/right seller to catch events such as attention switching. Please see the supplementary material for more details. + +We define the social signal function $s(\mathbf{X}^{1:N})$ as Equation (11), which takes a sequence of multi-person motion $\mathbf{X}^{1:N}$ as input and generates one of the 16 distinct states per frame. For the social signal function $s$ to work on any haggling motion sequence, we need to determine three pieces of information: + +1. who the buyer is, +2. whom the buyer is paying attention to, +3. whether someone is speaking. + +
MethodCMU-MocapMuPoTS-3D3DPW
RootPoseRootPoseRootPose
1s2s3s1s2s3s1s2s3s1s2s3s1s2s3s
LTD [41]0.971.732.620.981.211.370.891.391.910.881.141.314.286.798.41
HRI [40]0.962.063.111.051.371.580.661.302.160.731.071.303.676.428.64
SP [2]0.962.012.961.031.411.710.961.382.210.721.081.303.766.869.07
MRT [69]0.601.121.710.791.051.220.671.251.860.690.991.193.425.697.30
Ours0.721.101.440.380.460.491.141.281.420.590.640.671.662.763.59
+ +Table 2: MPJPE $\downarrow$ in dm for root joint and pose. The lowest error is in bold and the second lowest is underscored. + +
Frame1510152025306012025050075010001300
MRT [69]0.6240.2780.1940.2120.2240.2150.2150.2180.2050.1800.1290.0790.0620.047
Ours (Γ∅)0.6440.2800.2060.2150.2250.2260.2290.2330.2290.2260.2290.2270.2250.226
Ours (Γattn)0.6390.2800.1990.2130.2240.2290.2320.2270.2290.2230.2330.2280.2330.223
Ours* (ΓE)0.6400.2790.2040.2160.2270.2280.2300.2330.2270.2220.2300.2290.2300.226
+ +Table 3: Per-frame average NDMS ↑ score on the Haggling dataset. The highest score is in bold and the second highest is underscored. + +To solve (1) we train a simple buyer detection network, consisting of three layers of bi-directional Gated Recurrent Units, which gets as input a haggling motion and outputs the likelihood of each participant being the buyer. In Table 5, we report our accuracy of this approach. We see that the buyer detector correctly identifies the buyer all the time on the test set. + +For (2), we define the buyer's attention as whomever they look at, which can be easily calculated from the 3D body pose: + +$$ +\underset {i \in \{\text {l e f t}, \text {r i g h t} \}} {\operatorname {a r g m i n}} \left[ n ^ {T} \left(\frac {d _ {i}}{| d _ {i} |}\right) \right] \tag {16} +$$ + +where $d_{\mathrm{left}}$ and $d_{\mathrm{right}}$ are the directional vectors from the buyer nose to the left and right seller nose, respectively, projected onto the ground plane and $n$ is the 2D unit vector that is perpendicular to the left eye→right eye vector of the buyer, projected onto the ground plane. + +Last but not least, to determine if someone is speaking we utilize an off-the-shelf action classification network consisting of three layers of bi-directional Gated Recurrent Units. For training, we use the annotation of the Haggling dataset [30] which indicates if a person is speaking or not. The classifier achieves $87\%$ accuracy in speech detection on the test set. + +# 5.2. Multi-Person Forecasting + +We first report the results for the multi-person human motion datasets MuPoTS-3D $[43]^2$ , 3DPW $[67]^1$ and CMU-Mocap [1, 69]. We follow [69] and report the Mean Per Joint Positional Error (MPJPE) using global coordinates in Table 1. Our approach outperforms the methods LTD [41], HRI [40], SP [2], and MRT [69] by a large margin. While some methods perform better for the first second on MuPoTS-3D [43], our approach achieves a much lower + +error for all other settings and datasets. On the most difficult dataset 3DPW [67], the error is reduced by $57.6\%$ , $55.6\%$ , and $54.7\%$ for 1, 2, and 3 seconds, respectively. As in [69], we also report the error of the position of the root joint and pose error in local coordinates, i.e., setting the root position for all frames to zero, in Table 2. The results show that our approach forecasts by far the most accurate poses and outperforms the state of the art by a large margin on all datasets. Only the position of the root joint is slightly better estimated by other methods at the beginning of the datasets MuPoTS-3D and CMU-Mocap. At 3 seconds, our approach also achieves the lowest root joint error for all three datasets. + +# 5.3. Multi-Person Forecasting in the Context of Social Interactions + +For the remaining experiments, we evaluate our approach on the newly prepared Haggling dataset since it contains more social interactions as the other datasets. We compare our approach to Multi-Range Transformers (MRT) [69], which performed better than other approaches in Section 5.2. We used the publicly available source code and adjusted the approach to work with $30\mathrm{Hz}$ . We kept all other settings as is. + +We first report the per-frame NDMS (higher is better) at different frames in Table 3. Our experiments show that MRT [69] is capable of generating realistic motion for a few seconds. However, after 120 frames $(4s)$ the NDMS score drops significantly. This is caused by the auto-regressive motion forecasting strategy adopted by MRT, which results in error accumulation over time. In contrast, our method continues to predict motions with good NDMS scores well into the future. As discussed in Section 3.2, we compare different variants of the aggregation function $\Gamma$ . In terms of $\Gamma$ , there are no major differences in terms of NDMS, but they all perform much better than MRT [69]. In Table 4, we report the average NDMS score over the entire forecast sequence. We observe that $\Gamma_{\emptyset}$ , i.e., using no aggregation + +![](images/709b96924f85b85bbe64e7fb4265dcd2ab2eef26a158055020590513ef8ce370.jpg) +Figure 4: Our model is capable of generating realistic motion for 7 people from the Ultimatum sequence of Panoptic Studio [29], even though it was trained only on the triadic Haggling dataset. The Ultimatum sequence shares similarities with the Haggling dataset such as persons taking turns and talking to each other. + +![](images/b45c9a22f0fe144b9ebd6dce917f20f856ccfc66c6c2e1c8ed3efb2e41e59d9c.jpg) +(a) Train + +![](images/056c3f8fa3a1fd91bbf99d322c31945972cb75601fa53c01b603d9f04bfe4db9.jpg) +(b) Test + +![](images/facd8100c3df6afb7dae8aa33c167cf9907e223f8682552b37055c2069be6ce2.jpg) +(c) MRT [69] + +![](images/cd5486f2618286e99fd02a717172e1f952528762bc94c6659bb78edc1d25392f.jpg) +(d) Ours $(\Gamma_{\emptyset})$ + +![](images/5b7e675e708f7ced81c92bb2c06402d795927a22c5c7abefb1953335f4d43589.jpg) +(e) Ours $(\Gamma_{\mathrm{att}})$ + +![](images/eafb11b4ab620b18df8e59ae751ba1e65bf5c300866930aece59300ed1b2fb6b.jpg) +(f) Ours $(\Gamma_{\mathbb{E}})$ +Figure 5: State transition matrices for the states defined in Section 5.1 for the ground-truth of the train (a) and test set (b). The other transition matrices are obtained by the forecast of the current state-of-the-art model MRT [69] on the test set (c) and our model without context (d), with context via attention (e), and with context via averaging (f). The more similar the transition matrix is to the test set (b), the closer it matches the test motion. + +
Train[69]Ours (Γ∅)Ours (Γattn)Ours*(ΓR)
NDMS ↑-0.10150.23010.22700.2297
SSCP ↓0.09990.48390.35760.32780.3252
+ +information yields a slightly higher NDMS score than the other versions, but the differences are very small. However, $\Gamma_{\mathrm{attn}}$ and $\Gamma_{\mathbb{E}}$ forecast much more plausible social interactions (lower SSCP) as we will discuss next. + +Table 4: Per-frame average NDMS $\uparrow$ and SSCP $\downarrow$ on the Haggling dataset. + +
Test setMRT [69]Ours (Γ∅)Ours (Γattn)Ours* (ΓE)
1.00.32500.88120.88750.8938
+ +Table 5: Accuracy $\uparrow$ of the buyer detection network. For MRT [69], the model randomly selects a person to be the buyer, as there is a $1/3$ chance of selecting the buyer with random chance. + +# 5.3.1 Social Motion Evaluation + +The SSCP scores are presented in Table 4 where a lower value corresponds to more plausible forecast social interactions. All our proposed variants outperform the state-of-the-art approach MRT [69]. Using no aggregation information $(\Gamma_{\emptyset})$ performs worse than the aggregation functions $\Gamma_{\mathrm{attn}}$ (9) and $\Gamma_{\mathbb{E}}$ (10), which is expected since the aggregation function passes information from other people in the scene. The average aggregation $(\Gamma_{\mathbb{E}})$ performs slightly better than the more complex multi-headed attention approach $(\Gamma_{\mathrm{attn}})$ . We conjecture that averaging bottleneck encodings over all people introduces an inductive bias to pay the same attention to everyone, which works well for modelling the haggling game. For completeness, we also report the SSCP score of the training set. + +We can draw some insights about what motion each model generates by looking at the state probability transition matrices in Figure 5. For example, MRT [69] (Figure 5d) produces mostly self-loops (diagonal of transition matrix) indicating that the motion gets stuck over time. When no context information is provided $(\Gamma_{\emptyset})$ , our method produces motions where all three people are talking at the same time, as can be seen in Figure 5d, where the last two entries (red arrow) in the transition matrix represent states with all three people talking. This is sensible as the model sees two sellers and only one buyer during training and thus it is more likely to produce motion that resembles a seller, who talks most of the time. When the context is provided, our approach overcomes this limitation as expected and rarely produces motion where all three people are talking at the same time as shown in Figures 5e and 5f. + +These observations are also confirmed when measuring the buyer detection accuracy on the forecast motion, which is reported in Table 5. The detector fails to identify the correct buyer in the sequences that are forecast by MRT [69] and it nearly chooses the buyer at random with $1/3$ accuracy. This confirms that MRT does not forecast socially consistent sequences where the social role of the persons, namely buyer or seller, is preserved. In contrast, our method predicts motion where the buyer can be easily determined most of the time. As for the SSCP scores reported in Table 4, $\Gamma_{\mathbb{E}}$ performs best. + +In summary, the aggregation function $\Gamma_{\mathbb{E}}$ outperforms the other aggregation functions on the Haggling dataset [30] as it produces the most socially plausible motion according to our Symbolic Social Cues Protocol while also generating highly plausible 3D body motion. + +# 5.4. Ablation Study + +Average velocity over time Freezing or unrealistically expanding motion are common failures in human motion anticipation. While NDMS [61] penalizes in contrast to MPJPE errors in the velocity, visualizing the average motion velocity can give interesting insights. In Figure 6, we plot the average velocity over all frames for all our summary function variants, the test set, and the state-of-the-art method MRT [69]. Note that the beginning of the sequence has a higher velocity due to people walking into the scene. We observe that MRT suffers from error accumulation caused by the auto-regressive inference scheme. The velocity produced by our motion tightly follows the test set velocity for roughly 250 frames after which the test set velocity is slightly larger. We attribute this to the higher degree of stochasticity of real motion, which results in sudden jerks and swings that increase the average velocity. + +NDMS score over time In Figure 7, we visualize how the NDMS scores of all proposed variants and MRT evolve over time. For reference, we also calculate the NDMS score of the training data which is guaranteed to be realistic. Note that NDMS is 1 for the observed part of the test sequences. As shown in the figure, our method achieves almost the same level of realism as the training data while the quality of MRT slowly degenerates over time. + +Anticipating more than three people We have trained and evaluated our model on the Haggling dataset where each sequence consists of triadic interaction. However, the fully convolutional nature of our approach as well as the order-invariance of the summarization function $\Gamma$ allow us to forecast any number of people. To demonstrate this capability, we predict 7 people from the Ultimatum sequence of Panoptic Studio [29] using only the model trained on the Haggling dataset. This works well because the Haggling and Ultimatum sequences share many social behaviors, such as turn-taking, talking, and paying attention while + +![](images/ac00d8c38b4b636ae1aa8bd642f4cd9abdcd02230489aed81afffeaebd4d4790.jpg) +Figure 6: Average velocity over time for the entire test motion $\mathcal{X}$ and generated motions $\hat{\mathcal{X}}$ . The x-axis represents the frames while the y-axis represents the (log) average velocity of the data. + +![](images/5a351a73378648bf0e2250f7ae2b295144fbb2e523be95ef2281e01aa6347924.jpg) +Figure 7: Average NDMS [61] score $\uparrow$ of all proposed variants, MRT, and the train set over time. + +standing in a circle. Our results can be seen in Figure 4 where our model is able to predict realistic motion for 7 people, even though it was trained only on the triadic Haggling dataset. More results are provided in the supplementary material. + +# 6. Conclusion + +In this work, we present Social Diffusion, a stochastic multi-person motion anticipation model. The approach not only forecasts realistic motions on the individual level, but also plausible social interactions where the social roles of individuals are preserved over time. The approach is very flexible. It can be used for short and long-term forecasting and can be applied to larger groups than observed during training. As a second contribution, we proposed a new evaluation protocol to measure the realism of forecast social interactions. We furthermore derived a dataset for multi-person social interaction forecasting from the Haggling dataset [30] where the persons have different social roles that impact their behavior. We evaluated our approach + +on four multi-person datasets and demonstrated that our approach outperforms the state-of-the-art for short-term and long-term anticipation both in realism of forecast motion and social interaction. The approach has still some limitations. For instance, the global positions of the root joints can be better estimated. Future directions also include extending the model to predict motions of dynamic groups of people, e.g., at a cocktail party where any individual can freely disengage from the current conversation group and join another one. + +# Acknowledgement + +Juergen Gall has been supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) GA 1927/4-2 (FOR 2535 Anticipating Human Behavior), by the Federal Ministry of Education and Research (BMBF) under grant no. 01IS22094E WEST-AI, the project iBehave (receiving funding from the programme "Netzwerke 2021", an initiative of the Ministry of Culture and Science of the State of Northrhine Westphalia), and the ERC Consolidator Grant FORHUE (101044724). The sole responsibility for the content of this publication lies with the authors. + +# References + +[1] CMU. Carnegie-Mellon Mocap Database. +[2] Vida Adeli, Ehsan Adeli, Ian Reid, Juan Carlos Niebles, and Hamid Rezatofighi. Socially and contextually aware human motion and pose forecasting. Robotics and Automation Letters, 2020. +[3] Vida Adeli, Mahsa Ehsanpour, Ian Reid, Juan Carlos Niebles, Silvio Savarese, Ehsan Adeli, and Hamid Rezatofighi. Tripod: Human trajectory and pose dynamics forecasting in the wild. In International Conference on Computer Vision, 2021. +[4] Emre Aksan, Manuel Kaufmann, Peng Cao, and Otmar Hilliges. A spatio-temporal transformer for 3d human motion prediction. In International Conference on 3D Vision, 2021. +[5] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social lstm: Human trajectory prediction in crowded spaces. In Conference on Computer Vision and Pattern Recognition, 2016. +[6] Sadegh Aliakbarian, Fatemeh Sadat Saleh, Mathieu Salzmann, Lars Petersson, and Stephen Gould. A stochastic conditioning scheme for diverse human motion prediction. In Conference on Computer Vision and Pattern Recognition, 2020. +[7] Eduardo Alvarado, Damien Rohmer, and Marie-Paule Cani. Generating upper-body motion for real-time characters making their way through dynamic environments. Eurographics Symposium on Computer Animation, 2022. + +[8] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 2021. +[9] Timur Bagautdinov, Alexandre Alahi, François Fleuret, Pascal Fua, and Silvio Savarese. Social scene understanding: End-to-end multi-person action localization and collective activity recognition. In Conference on Computer Vision and Pattern Recognition, 2017. +[10] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. +[11] Judith Butepage, Michael J Black, Danica Kragic, and Hedvig Kjellstrom. Deep representation learning for human motion prediction and classification. In Conference on Computer Vision and Pattern Recognition, 2017. +[12] Yujun Cai, Lin Huang, Yiwei Wang, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Xu Yang, Yiheng Zhu, Xiaohui Shen, et al. Learning progressive joint propagation for human motion prediction. In European Conference on Computer Vision, 2020. +[13] Yujun Cai, Yiwei Wang, Yiheng Zhu, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Chuanxia Zheng, Sijie Yan, Henghui Ding, et al. A unified 3d human motion synthesis model via conditional variational auto-encoder. In International Conference on Computer Vision, 2021. +[14] Lujing Chen, Rui Liu, Xin Yang, Dongsheng Zhou, Qiang Zhang, and Xiaopeng Wei. Sttg-net: a spatio-temporal network for human motion prediction based on transformer and graph convolution network. Visual Computing for Industry, Biomedicine, and Art, 2022. +[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 2021. +[16] Christian Diller, Thomas Funkhouser, and Angela Dai. Forecasting characteristic 3d poses of human actions. Conference on Computer Vision and Pattern Recognition, 2022. +[17] Dominik Maria Endres and Johannes E Schindelin. A new metric for probability distributions. Transactions on Information theory, 2003. +[18] Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models for human dynamics. In International Conference on Computer Vision, 2015. +[19] Bent Fuglede and Flemming Topsoe. Jensen-Shannon divergence and Hilbert space embedding. In International Symposium on Information Theory, 2004. +[20] Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In Conference on Computer Vision and Pattern Recognition, 2020. +[21] Anand Gopalakrishnan, Ankur Mali, Dan Kifer, Lee Giles, and Alexander G Ororbia. A neural temporal model for human motion prediction. In Conference on Computer Vision and Pattern Recognition, 2019. + +[22] Liang-Yan Gui, Yu-Xiong Wang, Xiaodan Liang, and José MF Moura. Adversarial geometry-aware human motion prediction. In European Conference on Computer Vision, 2018. +[23] Wen Guo, Xiaoyu Bie, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Multi-person extreme motion prediction. In Conference on Computer Vision and Pattern Recognition, 2022. +[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 2020. +[25] Daniel Holden, Jun Saito, Taku Komura, and Thomas Joyce. Learning motion manifolds with convolutional autoencoders. In SIGGRAPH Asia technical briefs. 2015. +[26] Emiel Hoogeboom, Didrik Nielsen, Priyank Jaini, Patrick Forre, and Max Welling. Argmax flows and multinomial diffusion: Learning categorical distributions. Advances in Neural Information Processing Systems, 2021. +[27] Ashesh Jain, Amir R Zamir, Silvio Savarese, and Ashutosh Saxena. Structural-rnn: Deep learning on spatio-temporal graphs. In Conference on Computer Vision and Pattern Recognition, 2016. +[28] Michael Janner, Yilun Du, Joshua B Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. International Conference on Learning Representations, 2022. +[29] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic Studio: A Massively Multiview System for Social Motion Capture. In International Conference on Computer Vision, 2015. +[30] Hanbyul Joo, Tomas Simon, Mina Cikara, and Yaser Sheikh. Towards social artificial intelligence: Nonverbal social signal prediction in a triadic interaction. In Conference on Computer Vision and Pattern Recognition, 2019. +[31] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Scott Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. Transactions on Pattern Analysis and Machine Intelligence, 2017. +[32] Julia Kantorovitch, Janne Väre, Vesa Pehkonen, Arto Laikari, and Heikki Seppälä. An assistive household robot-doing more than just cleaning. Journal of Assistive Technologies, 2014. +[33] Anne Keitel, Moritz M Daum, et al. The use of intonation for turn anticipation in observed conversations without visual signals as source of information. Frontiers in psychology, 2015. +[34] Lukasz Kidziński, Bryan Yang, Jennifer L Hicks, Apoorva Rajagopal, Scott L Delp, and Michael H Schwartz. Deep neural networks enable quantitative movement analysis using single-camera videos. Nature communications, 2020. +[35] Vineet Kosaraju, Amir Sadeghian, Roberto Martin-Martin, Ian Reid, Hamid Rezatofighi, and Silvio Savarese. Socialbigat: Multimodal trajectory forecasting using bicycle-gan and graph attention networks. Advances in Neural Information Processing Systems, 2019. + +[36] Stephen C Levinson and Francisco Torreira. Timing in turn-taking and its implications for processing models of language. Frontiers in psychology, 2015. +[37] Chen Li, Zhen Zhang, Wee Sun Lee, and Gim Hee Lee. Convolutional sequence to sequence model for human dynamics. In Conference on Computer Vision and Pattern Recognition, 2018. +[38] Thomas Lucas*, Fabien Baradel*, Philippe Weinzaepfel, and Grégory Rogez. Posegpt: Quantization-based 3d human motion generation and forecasting. In European Conference on Computer Vision, 2022. +[39] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Conference on Computer Vision and Pattern Recognition, 2022. +[40] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. History repeats itself: Human motion prediction via motion attention. In European Conference on Computer Vision, 2020. +[41] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies for human motion prediction. International Conference on Computer Vision, 2019. +[42] Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Conference on Computer Vision and Pattern Recognition, 2017. +[43] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In International Conference on 3D Vision, 2018. +[44] Ronja Möller, Antonino Furnari, Sebastiano Battiato, Aki Härma, and Giovanni Maria Farinella. A survey on human-aware robot navigation. Robotics and Autonomous Systems, 2021. +[45] Brendan Tran Morris and Mohan Manubhai Trivedi. A survey of vision-based trajectory learning and analysis for surveillance. Transactions on Circuits and Systems for Video Technology, 2008. +[46] Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. Learning to listen: Modeling non-deterministic dyadic facial motion. In Conference on Computer Vision and Pattern Recognition, 2022. +[47] Sangmin Oh, Anthony Hoogs, Amitha Perera, Naresh Cunteor, Chia-Chih Chen, Jong Taek Lee, Saurajit Mukherjee, JK Aggarwal, Hyungtae Lee, Larry Davis, et al. A large-scale benchmark dataset for event recognition in surveillance video. In Conference on Computer Vision and Pattern Recognition, 2011. +[48] Dario Pavllo, Christoph Feichtenhofer, Michael Auli, and David Grangier. Modeling human motion with quaternion-based neural networks. International Journal of Computer Vision, 2020. +[49] Dario Pavllo, David Grangier, and Michael Auli. *Quaternion: A quaternion-based recurrent model for human motion*. In *British Machine Vision Conference*, 2018. + +[50] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. European Conference on Computer Vision, 2022. +[51] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. +[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition, 2022. +[53] Alejandro Hernandez Ruiz, Juergen Gall, and Francesc Moreno-Noguer. Human motion prediction via spatiotemporal inpainting. International Conference on Computer Vision, 2019. +[54] Harvey Sacks, Emanuel A Schegloff, and Gail Jefferson. A simplest systematics for the organization of turn taking for conversation. In Studies in the organization of conversational interaction. 1978. +[55] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In Transactions on Graphics, 2022. +[56] Keren Shavit-Cohen and Elana Zion Golumbic. The dynamics of attention shifts among concurrent speech in a naturalistic multi-speaker virtual environment. Frontiers in Human Neuroscience, 2019. +[57] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, 2015. +[58] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 2015. +[59] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. +[60] Chen Sun, Abhinav Shrivastava, Carl Vondrick, Rahul Sukthankar, Kevin Murphy, and Cordelia Schmid. Relational action forecasting. In Conference on Computer Vision and Pattern Recognition, 2019. +[61] Julian Tanke, Chintan Zaveri, and Juergen Gall. Intention-based long-term human motion anticipation. In International Conference on 3D Vision, 2021. +[62] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. +[63] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 2017. +[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 2017. + +[65] Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. International Conference on Learning Representations, 2017. +[66] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European Conference on Computer Vision, 2018. +[67] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European conference on computer vision, 2018. +[68] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In Conference on Computer Vision and Pattern Recognition, 2022. +[69] Jiashun Wang, Huazhe Xu, Medhini Narasimhan, and Xiaolong Wang. Multi-person 3d motion prediction with multirange transformers. Advances in Neural Information Processing Systems, 2021. +[70] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In Conference on Computer Vision and Pattern Recognition, 2021. +[71] Ye Yuan and Kris Kitani. DLOW: Diversifying latent flows for diverse human motion prediction. In European Conference on Computer Vision, 2020. +[72] Jason Y Zhang, Panna Felsen, Angjoo Kanazawa, and Jitendra Malik. Predicting 3d human dynamics from video. In International Conference on Computer Vision, 2019. +[73] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. +[74] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J Black, and Siyu Tang. Generating 3d people in scenes without people. In Conference on computer vision and pattern recognition, 2020. +[75] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3d scenes. In Conference on Computer Vision and Pattern Recognition, 2022. \ No newline at end of file diff --git a/2023/Social Diffusion_ Long-term Multiple Human Motion Anticipation/images.zip b/2023/Social Diffusion_ Long-term Multiple Human Motion Anticipation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..df05fce94c9c97935a96169c86af27f724476ce8 --- /dev/null +++ b/2023/Social Diffusion_ Long-term Multiple Human Motion Anticipation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7704196f514ee2fd2aaacc49cdb621f0e87f24f17cf0704cbc2b11e697bfd868 +size 409509 diff --git a/2023/Social Diffusion_ Long-term Multiple Human Motion Anticipation/layout.json b/2023/Social Diffusion_ Long-term Multiple Human Motion Anticipation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ec4b9e552525ebbd6dbd756203f04c7bcc8c2918 --- /dev/null +++ b/2023/Social Diffusion_ Long-term Multiple Human Motion Anticipation/layout.json @@ -0,0 +1,11069 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 103, + 501, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 103, + 501, + 121 + ], + "spans": [ + { + "bbox": [ + 91, + 103, + 501, + 121 + ], + "type": "text", + "content": "Social Diffusion: Long-term Multiple Human Motion Anticipation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 152, + 541, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 152, + 541, + 182 + ], + "spans": [ + { + "bbox": [ + 51, + 152, + 541, + 182 + ], + "type": "text", + "content": "Julian Tanke\\*, Linguang Zhang?, Amy Zhao?, Chengcheng Tang?, Yujun Cai?, Lezi Wang?, Po-Chen Wu?, Juergen Gall1,3, and Cem Keskin2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 190, + 455, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 190, + 455, + 219 + ], + "spans": [ + { + "bbox": [ + 136, + 190, + 455, + 219 + ], + "type": "text", + "content": "1University of Bonn 2Reality Labs Research \n3Lamarr Institute for Machine Learning and Artificial Intelligence" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 220, + 516, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 220, + 516, + 241 + ], + "spans": [ + { + "bbox": [ + 79, + 220, + 516, + 241 + ], + "type": "text", + "content": "{tanke,gall}@iai.uni-bonn.de {linguang,xamyzhao,chengcheng.tang,yujuncai,wanglezi,pochenwu,cemkeskin}@meta.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 277, + 192, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 277, + 192, + 289 + ], + "spans": [ + { + "bbox": [ + 143, + 277, + 192, + 289 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 302, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 289, + 529 + ], + "type": "text", + "content": "We propose Social Diffusion, a novel method for short-term and long-term forecasting of the motion of multiple persons as well as their social interactions. Jointly forecasting motions for multiple persons involved in social activities is inherently a challenging problem due to the interdependencies between individuals. In this work, we leverage a diffusion model conditioned on motion histories and causal temporal convolutional networks to forecast individually and contextually plausible motions for all participants. The contextual plausibility is achieved via an order-invariant aggregation function. As a second contribution, we design a new evaluation protocol that measures the plausibility of social interactions which we evaluate on the Haggling dataset, which features a challenging social activity where people are actively taking turns to talk and switching their attention. We evaluate our approach on four datasets for multi-person forecasting where our approach outperforms the state-of-the-art in terms of motion realism and contextual plausibility." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 540, + 127, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 127, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 127, + 553 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 561, + 287, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 561, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 46, + 561, + 287, + 645 + ], + "type": "text", + "content": "Understanding and anticipating social interactions in groups of people is a challenging and highly relevant topic [35, 9, 32, 45, 47, 5, 34]. For instance, it is essential for socially-compliant robots [44], but it is also relevant for neuroscience and social sciences since it allows to develop computational models on how the behavior of other persons is perceived and how it changes the own behavior." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 646, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 646, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 646, + 287, + 694 + ], + "type": "text", + "content": "Forecasting realistic social interactions, however, is very challenging for two reasons. First, social interactions tend to last for tens of seconds [33] or even minutes - much longer than the prediction from most of the existing human" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 277, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 277, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 306, + 277, + 545, + 410 + ], + "type": "text", + "content": "motion anticipation models [23, 69, 3, 22, 27, 42, 49, 11, 25, 37, 4, 14, 41, 40]. Second, social interactions consist of interdependent motions [54, 36], which requires modeling the relationships among all individuals. For example, in conversational turn-taking, a person's turn to talk highly depends on the start/end of the others' speaking. While multi-person motion anticipation has emerged as a new topic, current approaches [23, 69, 3] do not pay much attention on complex social interactions. For instance, they do not preserve the social role of individuals in a group such that the interactions become socially implausible over time." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 411, + 546, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 546, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 546, + 651 + ], + "type": "text", + "content": "To address the limitations of existing models, we propose Social Diffusion to predict motions of multiple people and ensure contextually plausible interactions, as shown in Fig. 1. To this end, we learn the distribution of human motion by leveraging a diffusion model [39, 28, 52, 58, 24, 59, 62, 73]. To enforce information exchange among people, which is critical to predicting contextually plausible interactions, we introduce an order-invariant aggregation function to aggregate motion features from all people. For inference, we feed back the input sequence to the signal during the reverse-diffusion steps to condition the motion generation on the past motion. Our method is fully convolutional which allows us to generate sequences of arbitrary size. This allows us to not just forecast the next few seconds of an input motion but also to forecast social interactions that last longer. Furthermore, our approach is very flexible in the sense that the number of persons during training and inference can differ. To the best of our knowledge, our approach represents the first diffusion model that produces multi-person motions at the same time." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "As a second contribution, we propose a new evaluation protocol for social interactions based on Symbolic Social Cues, which measures whether the forecast motion is socially plausible. Our key observation is that the probabilities of transitions between social interaction states are highly" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 266, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 266, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 266, + 713 + ], + "type": "text", + "content": "*Work done partially while Julian was at Reality Labs Research." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9601" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 68, + 545, + 166 + ], + "blocks": [ + { + "bbox": [ + 51, + 68, + 545, + 166 + ], + "lines": [ + { + "bbox": [ + 51, + 68, + 545, + 166 + ], + "spans": [ + { + "bbox": [ + 51, + 68, + 545, + 166 + ], + "type": "image", + "image_path": "ce52e1e5601acd91e4fbea13d22232ae67f1981ecc07eda16150bcf2b7facdb3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 175, + 546, + 211 + ], + "lines": [ + { + "bbox": [ + 46, + 175, + 546, + 211 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 546, + 211 + ], + "type": "text", + "content": "Figure 1: We propose an approach for multi-person motion anticipation: given a sequence of human social interactions (blue/red skeletons), the proposed model forecasts multi-person motions where the social roles are preserved and the interactions are socially plausible." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 220, + 287, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 287, + 304 + ], + "type": "text", + "content": "correlated with the plausibility of predicted social interactions. In a conversation, for example, a person usually starts talking only when a peer stops talking. To evaluate predicted motions, we first build the state transition graph by extracting states from the motions. We then treat the state transition graph as a probability distribution and compare it to the real data distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 304, + 287, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 412 + ], + "type": "text", + "content": "For evaluation, we use the Haggling dataset [30] which comprises 175 videos of well-defined triadic social interactions. In contrast to other existing multi-person human motion datasets [43, 66, 69], the persons have different social roles that impact their behavior. We furthermore evaluate our approach on the MuPoTS-3D [43], 3DPW [67], and CMU-Mocap [1, 69] dataset. On all four datasets, our approach outperforms the state of the art for multi-person human motion forecasting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 413, + 228, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 413, + 228, + 424 + ], + "spans": [ + { + "bbox": [ + 59, + 413, + 228, + 424 + ], + "type": "text", + "content": "In summary, our contribution is two-fold:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 433, + 287, + 526 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 433, + 287, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 287, + 482 + ], + "type": "text", + "content": "1. We propose Social Diffusion, the first stochastic multiperson motion anticipation model that outperforms the state of the art on common multi-person motion anticipation datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 490, + 287, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 490, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 54, + 490, + 287, + 526 + ], + "type": "text", + "content": "2. We propose a novel social interaction evaluation protocol that considers not only the validity of poses but also the plausibility of social interactions." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 537, + 134, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 134, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 134, + 550 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": "Single Person Human Motion Prediction Human motion prediction, which typically refers to generating motion sequences given a prefix segment, has been extensively studied in the past few years. Most recent papers focus on single person 3D predictions. Due to the inherent temporal nature of human motion, various temporal-based methods, such as recurrent neural networks (RNNs) [18, 22, 27, 42, 49, 48, 61, 72], temporal convolutional networks [11, 25, 37], transformers [4, 14, 12] and graph neural networks (GNNs) [14, 41, 40] have been used for this task. For long-term predictions, auto-regressive models that operate in the discrete space [46, 38, 61] have shown success, where a long prediction sequence can be obtained without" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 220, + 546, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 546, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 546, + 353 + ], + "type": "text", + "content": "converging to mean poses. Recently, anchor-based methods [16, 68] have been proposed; these focus on forecasting characteristic anchor poses rather than the entire sequence auto-regressively. To achieve better interactions with environments, [7, 70, 68, 74] proposed various ways to include contextual information into human motion predictions. Since generated motions can be controlled with a high-level guidance such as action class or text, some approaches [50, 13] used Variational Auto-Encoders [58] to solve this problem. Going beyond single human prediction, we predict the motions and interactions of multiple humans." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 354, + 547, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 569 + ], + "type": "text", + "content": "Multi-person Anticipation Modeling multi-person interactions has been a long standing problem [46, 23, 2, 65, 20, 30, 46, 63, 75]. For instance, " + }, + { + "bbox": [ + 304, + 354, + 547, + 569 + ], + "type": "inline_equation", + "content": "\\mathrm{DR}^2\\mathrm{N}" + }, + { + "bbox": [ + 304, + 354, + 547, + 569 + ], + "type": "text", + "content": " [60] predicts the activities of multiple people given a past video sequence. For a given frame, personal relationships between candidates are estimated using a graph attention network (GAT) [65] while temporal relationships are predicted using recurrent neural networks. Recently, Wang et al. [69] addressed multi-person 3D motion trajectory prediction via a Multi-Range Transformers framework. Guo et al. [23] introduced Cross-Interaction Attention to jointly model highly expressive dance sequences. Joo et al. [30] introduced a triadic haggling game for social signal prediction, based on Panoptic Studio [29, 31]. In their work they predict the motion of a single person given the motion of the others. Similarly, [46] predicts motions based on the other actors' behavior. Crucially, both methods only predict the motion of a single individual, given other persons social signals." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "Diffusion Models Diffusion models [24, 57, 59] belong to the family of probabilistic generative models, which convert the training data successively to Gaussian noise, and then learn to recover the data by reversing this noising process. Diffusion models have emerged as powerful deep generative models with breakthroughs in many applications, including image synthesis [24, 15], segmentation [10], and natural language processing [8, 26]. For conditioned generation, [15] introduced classifier-guided diffusion, and [55] takes the inpainted images as denoised by the model. More recently, [62, 73] have suggested diffusion models for motion generation; however, they are limited to single human" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9602" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 293, + 233 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 293, + 233 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 293, + 233 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 293, + 233 + ], + "type": "image", + "image_path": "cbf1d81ef2fc39b6828648edcfe6213bf3769da7a72fb791a16c4d47cd156a1a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 239, + 289, + 313 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 289, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 289, + 313 + ], + "type": "text", + "content": "Figure 2: Model overview: the reverse diffusion process " + }, + { + "bbox": [ + 46, + 239, + 289, + 313 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 239, + 289, + 313 + ], + "type": "text", + "content": " consists of a fully convolutional causal encoder " + }, + { + "bbox": [ + 46, + 239, + 289, + 313 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 239, + 289, + 313 + ], + "type": "text", + "content": " and a fully convolutional causal decoder " + }, + { + "bbox": [ + 46, + 239, + 289, + 313 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 239, + 289, + 313 + ], + "type": "text", + "content": " that produces a denoised motion sequence " + }, + { + "bbox": [ + 46, + 239, + 289, + 313 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0^i \\equiv \\hat{\\mathbf{x}}_0^{i,1:N}" + }, + { + "bbox": [ + 46, + 239, + 289, + 313 + ], + "type": "text", + "content": " for person " + }, + { + "bbox": [ + 46, + 239, + 289, + 313 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 239, + 289, + 313 + ], + "type": "text", + "content": ", given the bottleneck state " + }, + { + "bbox": [ + 46, + 239, + 289, + 313 + ], + "type": "inline_equation", + "content": "h^i" + }, + { + "bbox": [ + 46, + 239, + 289, + 313 + ], + "type": "text", + "content": " and the aggregation function " + }, + { + "bbox": [ + 46, + 239, + 289, + 313 + ], + "type": "inline_equation", + "content": "\\Gamma(\\{\\mathbf{h}_j \\mid j \\in 1, \\dots, p\\})" + }, + { + "bbox": [ + 46, + 239, + 289, + 313 + ], + "type": "text", + "content": " over all people in the scene." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 331, + 287, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 368 + ], + "type": "text", + "content": "3D motion prediction. To the best of our knowledge, we are the first to build a stochastic multi-person motion anticipation model that can predict very long-term future motions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 376, + 179, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 179, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 179, + 389 + ], + "type": "text", + "content": "3. Social Diffusion Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "text", + "content": "As illustrated in Fig. 1, we aim to forecast the motion of multiple persons that interact with each other. The forecast motion should be realistic and socially plausible. For instance, not all persons should talk at the same time. Formally, we represent a human motion sequence with " + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "text", + "content": " people of length " + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{1:N} \\in \\mathbb{R}^{N \\times p \\times \\delta}" + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "text", + "content": " represents the dimension of the individual pose vector at a given frame. Our goal is then to predict the future motion " + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^{n+1:N}" + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "text", + "content": " for all people, given their past motions " + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{1:n}" + }, + { + "bbox": [ + 46, + 396, + 287, + 505 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 510, + 287, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 510, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 115, + 510, + 287, + 525 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {X}} ^ {n + 1: N} = \\operatorname {S D M} \\left(\\mathbf {X} ^ {1: n}\\right). \\tag {1}", + "image_path": "b9aee1965853f83673b369df4ee52032c534cc5b399e9bcc70bd5e2133407d1c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 530, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 287, + 579 + ], + "type": "text", + "content": "Before we describe the proposed Social Diffusion Model (SDM) in Section 3.2, we will briefly describe a generic diffusion model [24] in Section 3.1. In Section 4, we will then introduce the social interaction evaluation protocol." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 585, + 146, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 146, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 146, + 597 + ], + "type": "text", + "content": "3.1. Diffusion Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 603, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 603, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 603, + 287, + 662 + ], + "type": "text", + "content": "A latent representation " + }, + { + "bbox": [ + 46, + 603, + 287, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{1})" + }, + { + "bbox": [ + 46, + 603, + 287, + 662 + ], + "type": "text", + "content": " is obtained via a " + }, + { + "bbox": [ + 46, + 603, + 287, + 662 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 603, + 287, + 662 + ], + "type": "text", + "content": " step Markov Gaussian noising process " + }, + { + "bbox": [ + 46, + 603, + 287, + 662 + ], + "type": "inline_equation", + "content": "q(\\mathbf{X}_T|\\mathbf{X}_0)" + }, + { + "bbox": [ + 46, + 603, + 287, + 662 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 603, + 287, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0\\equiv \\mathbf{X}^{1:N}" + }, + { + "bbox": [ + 46, + 603, + 287, + 662 + ], + "type": "text", + "content": " is a real motion sequence from the training set. The Markov Gaussian noising process can be written in closed form:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 85, + 670, + 287, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 670, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 85, + 670, + 287, + 684 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {X} _ {t} \\mid \\mathbf {X} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {X} _ {t}; \\sqrt {\\alpha_ {t}} \\mathbf {X} _ {0}, (1 - \\alpha_ {t}) \\mathbf {I}\\right) \\tag {2}", + "image_path": "b3f9d3000bb64ab0f748002e68e5e8ee8ea163f3c1fcba2406fc48bdb8351830.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{t} \\in (0,1)" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " is a step-dependent fixed hyperparameter. To sample from the generative model, we learn" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 85 + ], + "type": "text", + "content": "to invert the noising step using the generator function " + }, + { + "bbox": [ + 305, + 72, + 545, + 85 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 305, + 72, + 545, + 85 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 391, + 102, + 545, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 102, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 391, + 102, + 545, + 118 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {X}} _ {0, t} = g \\left(\\mathbf {X} _ {t}, t\\right) \\tag {3}", + "image_path": "9484bc13ea02a7b5d4cc6beccd31861f39c1e13852a72959a045618c6b1297bd.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 125, + 547, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 547, + 174 + ], + "type": "text", + "content": "The key contribution of our model is the novel generator function " + }, + { + "bbox": [ + 304, + 125, + 547, + 174 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 125, + 547, + 174 + ], + "type": "text", + "content": ", which models social interactions over time and will be described in Section 3.2. Following [24, 62, 51], the loss during training is defined by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 337, + 179, + 545, + 207 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 179, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 337, + 179, + 545, + 207 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {\\mathbf {X} _ {0} \\sim p (\\mathbf {X}), t \\sim [ 1, T ]} \\left[ \\left| \\left| \\mathbf {X} _ {0} - g \\left(\\mathbf {X} _ {t}, t\\right) \\right| \\right| \\right] \\tag {4}", + "image_path": "6831083c57e260b6e0e4514d63c66d872ccbf11c1890fb01c704875232a5389f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 213, + 547, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 262 + ], + "type": "text", + "content": "For inference, we reverse-iterate over Equation (3), starting at sampling step " + }, + { + "bbox": [ + 304, + 213, + 547, + 262 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 213, + 547, + 262 + ], + "type": "text", + "content": " and latent representation " + }, + { + "bbox": [ + 304, + 213, + 547, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{1})" + }, + { + "bbox": [ + 304, + 213, + 547, + 262 + ], + "type": "text", + "content": ". At each iteration " + }, + { + "bbox": [ + 304, + 213, + 547, + 262 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 213, + 547, + 262 + ], + "type": "text", + "content": ", we slowly denoise the motion sequence using (2) and (3):" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 355, + 267, + 545, + 282 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 267, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 355, + 267, + 545, + 282 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {X}} _ {0, t - 1} = g \\left(q \\left(\\hat {\\mathbf {X}} _ {t - 1} \\mid \\hat {\\mathbf {X}} _ {0, t}\\right), t - 1\\right) \\tag {5}", + "image_path": "63103415abb12eb4cd3034f952e7c25843f7aaeb74944600dc83e6dabdbeca76.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 290, + 511, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 290, + 511, + 301 + ], + "spans": [ + { + "bbox": [ + 305, + 290, + 511, + 301 + ], + "type": "text", + "content": "The final denoised motion is obtained when " + }, + { + "bbox": [ + 305, + 290, + 511, + 301 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 305, + 290, + 511, + 301 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 308, + 479, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 479, + 322 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 479, + 322 + ], + "type": "text", + "content": "3.2. Multi-person Motion Generator" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 327, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 398 + ], + "type": "text", + "content": "The diffusion model described in Section 3.1 produces unconditioned motion. In order to use the model for motion forecasting, we need to condition the model on the observed motion sequence of all persons " + }, + { + "bbox": [ + 304, + 327, + 545, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0^{1:n} = \\mathbf{X}^{1:n}" + }, + { + "bbox": [ + 304, + 327, + 545, + 398 + ], + "type": "text", + "content": ". To this end, we modify the inference sampling (5) to also include past motion as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 332, + 405, + 545, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 405, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 332, + 405, + 545, + 421 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {X}} _ {0, t - 1} = g \\left(q \\left(\\hat {\\mathbf {X}} _ {t - 1} \\mid \\mathbf {X} ^ {1: n} \\cup \\hat {\\mathbf {X}} _ {0, t} ^ {n + 1: N}\\right), t - 1\\right) \\tag {6}", + "image_path": "06acd7e562f385a338d67def6cd99f27fc14ee93b474a31f30b079fed13b1e1e.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 428, + 545, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 428, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 428, + 545, + 487 + ], + "type": "text", + "content": "The reverse diffusion process " + }, + { + "bbox": [ + 304, + 428, + 545, + 487 + ], + "type": "inline_equation", + "content": "g(\\mathbf{X}_t, t)" + }, + { + "bbox": [ + 304, + 428, + 545, + 487 + ], + "type": "text", + "content": " consists of three components, a causal temporal convolutional encoder " + }, + { + "bbox": [ + 304, + 428, + 545, + 487 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 428, + 545, + 487 + ], + "type": "text", + "content": ", a causal temporal convolutional decoder " + }, + { + "bbox": [ + 304, + 428, + 545, + 487 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 428, + 545, + 487 + ], + "type": "text", + "content": ", and an order-invariant function " + }, + { + "bbox": [ + 304, + 428, + 545, + 487 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 304, + 428, + 545, + 487 + ], + "type": "text", + "content": " that aggregates the interaction of the different persons, see Figure 2." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 488, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 488, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 545, + 559 + ], + "type": "text", + "content": "The encoder " + }, + { + "bbox": [ + 304, + 488, + 545, + 559 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 488, + 545, + 559 + ], + "type": "text", + "content": " and the decoder " + }, + { + "bbox": [ + 304, + 488, + 545, + 559 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 488, + 545, + 559 + ], + "type": "text", + "content": " process each individual sequence independently while " + }, + { + "bbox": [ + 304, + 488, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 304, + 488, + 545, + 559 + ], + "type": "text", + "content": " ensures that information flows between all persons in the scene. Formally, given " + }, + { + "bbox": [ + 304, + 488, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^i = \\hat{\\mathbf{X}}_t^{i,1:N} \\in \\mathbb{R}^{N \\times \\delta}" + }, + { + "bbox": [ + 304, + 488, + 545, + 559 + ], + "type": "text", + "content": ", which is the motion sequence for a single person " + }, + { + "bbox": [ + 304, + 488, + 545, + 559 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 488, + 545, + 559 + ], + "type": "text", + "content": " at scheduled noisng step " + }, + { + "bbox": [ + 304, + 488, + 545, + 559 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 488, + 545, + 559 + ], + "type": "text", + "content": ", we obtain for each person " + }, + { + "bbox": [ + 304, + 488, + 545, + 559 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 488, + 545, + 559 + ], + "type": "text", + "content": " in the scene a bottleneck encoding " + }, + { + "bbox": [ + 304, + 488, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^i" + }, + { + "bbox": [ + 304, + 488, + 545, + 559 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 369, + 567, + 545, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 567, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 369, + 567, + 545, + 582 + ], + "type": "interline_equation", + "content": "\\mathbf {h} ^ {i} = e \\left(\\mathbf {x} _ {t} ^ {i}, t\\right) \\forall i \\in 1, \\dots , p. \\tag {7}", + "image_path": "7a24613ddcecb4f2b298c4cad7baeab16bc2a6d6f98192e0a97f94703eee9270.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 589, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 589, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 589, + 545, + 659 + ], + "type": "text", + "content": "The encoder " + }, + { + "bbox": [ + 304, + 589, + 545, + 659 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 589, + 545, + 659 + ], + "type": "text", + "content": " consists of a 4-layer temporal convolutional network where each layer progressively reduces the temporal resolution by half via striding. In each layer, the noising step " + }, + { + "bbox": [ + 304, + 589, + 545, + 659 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 589, + 545, + 659 + ], + "type": "text", + "content": " is fed via sinusoidal positional encoding [64]. To produce the denoised motion " + }, + { + "bbox": [ + 304, + 589, + 545, + 659 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0^i" + }, + { + "bbox": [ + 304, + 589, + 545, + 659 + ], + "type": "text", + "content": ", the decoder " + }, + { + "bbox": [ + 304, + 589, + 545, + 659 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 589, + 545, + 659 + ], + "type": "text", + "content": " can be utilized as follows:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 355, + 667, + 545, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 667, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 355, + 667, + 545, + 683 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {0} ^ {i} = d \\left(\\mathbf {h} ^ {i}, t, \\Gamma \\left(\\mathbf {h} ^ {j} \\forall j \\in 1, \\dots , p\\right)\\right) \\tag {8}", + "image_path": "137efbf2d9d97be1131751164fe4293c6fc22f52362ee3d3d3ee3b5bce316c90.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " is a 4-layer temporal convolutional network and each layer progressively doubles the temporal resolution via" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9603" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "linear upsampling. As for the encoder, the noising step " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " is fed via sinusoidal positional encoding to each layer. In addition, the output of the order-invariant aggregation function " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " is concatenated to " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^{\\mathrm{i}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " before passing it to the first convolutional layer. The estimated motion sequences " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0^i" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " of each person " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " at noising step " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " are then concatenated to obtain " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{0,t}" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " and the approach proceeds to the next step " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 287, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 287, + 205 + ], + "type": "text", + "content": "The order-invariant aggregation function " + }, + { + "bbox": [ + 46, + 156, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 156, + 287, + 205 + ], + "type": "text", + "content": " passes information from other people in the scene. In our experiments, we evaluate two aggregation functions, averaging " + }, + { + "bbox": [ + 46, + 156, + 287, + 205 + ], + "type": "inline_equation", + "content": "(\\Gamma_{\\mathbb{E}})" + }, + { + "bbox": [ + 46, + 156, + 287, + 205 + ], + "type": "text", + "content": " over all people and multi-headed attention [64] " + }, + { + "bbox": [ + 46, + 156, + 287, + 205 + ], + "type": "inline_equation", + "content": "(\\Gamma_{\\mathrm{attn}})" + }, + { + "bbox": [ + 46, + 156, + 287, + 205 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 110, + 210, + 287, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 210, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 110, + 210, + 287, + 224 + ], + "type": "interline_equation", + "content": "\\Gamma_ {\\mathrm {a t t n}} (\\mathbf {H}) = \\operatorname {M u l t i H e a d} (\\mathbf {H}) \\tag {9}", + "image_path": "902daa7180eeeb3f717dd6f8e27caa79553c61543cbb49fd9bc07155da134bc7.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 227, + 287, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 227, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 110, + 227, + 287, + 258 + ], + "type": "interline_equation", + "content": "\\Gamma_ {\\mathbb {E}} (\\mathbf {H}) = \\frac {1}{p} \\sum_ {i = 1} ^ {p} \\mathbf {h} ^ {i} \\tag {10}", + "image_path": "1893d025795258e69db28b8ad417e4c84831d89c73f7488afb8a4e81338ad47b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 264, + 287, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 264, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 264, + 287, + 301 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 264, + 287, + 301 + ], + "type": "inline_equation", + "content": "\\mathbf{H} = \\{h^i\\}_{i=1}^p" + }, + { + "bbox": [ + 47, + 264, + 287, + 301 + ], + "type": "text", + "content": " are the embeddings of all layers of the encoder and for all people in the scene. MultiHead " + }, + { + "bbox": [ + 47, + 264, + 287, + 301 + ], + "type": "inline_equation", + "content": "(\\mathbf{H})" + }, + { + "bbox": [ + 47, + 264, + 287, + 301 + ], + "type": "text", + "content": " calculates the self-attention across all people for a given frame." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 307, + 180, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 307, + 180, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 307, + 180, + 319 + ], + "type": "text", + "content": "3.3. Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 326, + 287, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 287, + 482 + ], + "type": "text", + "content": "We follow state-of-the-art diffusion models [24, 39, 62] and use the cosine variance schedule. We set the number of diffusion steps to " + }, + { + "bbox": [ + 46, + 326, + 287, + 482 + ], + "type": "inline_equation", + "content": "T = 1000" + }, + { + "bbox": [ + 46, + 326, + 287, + 482 + ], + "type": "text", + "content": ". The encoder " + }, + { + "bbox": [ + 46, + 326, + 287, + 482 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 326, + 287, + 482 + ], + "type": "text", + "content": " consists of four layers of convolutional blocks with kernel size 3 and stride 2. The decoder " + }, + { + "bbox": [ + 46, + 326, + 287, + 482 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 326, + 287, + 482 + ], + "type": "text", + "content": " consists of four layers of convolutional blocks with additional upsampling layers that upsample the input sequence by factor two using linear interpolation. We standardize the training data to have zero mean and standard deviation one. We normalize all poses by splitting pose and global translation: each pose is transformed into a hipcentric coordinate frame and the pose is concatenated with the global rotation and translation to form a " + }, + { + "bbox": [ + 46, + 326, + 287, + 482 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 326, + 287, + 482 + ], + "type": "text", + "content": " dimensional pose vector." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 491, + 218, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 218, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 218, + 504 + ], + "type": "text", + "content": "4. Symbolic Social Cues Protocol" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "content": "We are interested in anticipating the social interactions among multiple people. Multi-person social interactions consist of several intricate and complex behaviours such as paying attention to a specific person [56] and turn-taking [54, 36], which usually take tens of seconds or even minutes. Current state-of-the-art multi-person motion anticipation methods [3, 69] calculate the Mean Per Joint Positional Error (MPJPE) using the ground-truth sequence, which is only meaningful for short time horizons of around one second [6, 21, 53, 61, 71]. More important, however, is that it does not measure the realism of social interaction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 643, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 643, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 643, + 287, + 689 + ], + "type": "text", + "content": "We thus propose the Symbolic Social Cues Protocol (SSCP), which divides the social interactions into a set of discrete interaction classes. In SSP, we define a social signal function" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 696, + 287, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 696, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 130, + 696, + 287, + 710 + ], + "type": "interline_equation", + "content": "C ^ {1: N} = s \\left(\\mathbf {X} ^ {1: N}\\right) \\tag {11}", + "image_path": "634c517f71552f260209257cb9401e54b9234a0219c3837a882e6c7d8d156feb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "text", + "content": "which takes as input a multi-person motion sequence " + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{1:N}\\in \\mathbb{R}^{N\\times p\\times \\delta}" + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "text", + "content": " and produces a discrete symbolic representation " + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "C^{1:N} = \\{c_n\\}_{n = 1}^N" + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "c_{n}\\in \\{1,\\dots,m\\}" + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "text", + "content": " represents the total number of symbolic states. A symbolic state is a unique summary of the current state of interaction, e.g., a person is talking and another person is listening. Given a test set " + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{\\mathbf{X}_i^{1:N}\\}_{i = 1}^K" + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "text", + "content": " sequences, we can now calculate the probability distribution " + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{sscp}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 179 + ], + "type": "text", + "content": " over the social state transitions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 356, + 186, + 545, + 219 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 186, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 356, + 186, + 545, + 219 + ], + "type": "interline_equation", + "content": "p _ {\\mathrm {s s c p}} (\\mathcal {X}) = \\frac {1}{\\zeta} \\sum_ {i = 1} ^ {K} \\operatorname {s t m} \\left(s \\left(\\mathbf {X} _ {i} ^ {1: N}\\right)\\right) \\tag {12}", + "image_path": "b993a88bf1895279ddc00765a6084d9f5ea59bcc62cde11d47617295389e2f2c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 228, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 228, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 545, + 276 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 228, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\mathrm{stm}(C_i^{1:N})" + }, + { + "bbox": [ + 305, + 228, + 545, + 276 + ], + "type": "text", + "content": " produces the " + }, + { + "bbox": [ + 305, + 228, + 545, + 276 + ], + "type": "inline_equation", + "content": "m\\times m" + }, + { + "bbox": [ + 305, + 228, + 545, + 276 + ], + "type": "text", + "content": " state transition matrix for the discrete symbolic sequence " + }, + { + "bbox": [ + 305, + 228, + 545, + 276 + ], + "type": "inline_equation", + "content": "C^{1:N}" + }, + { + "bbox": [ + 305, + 228, + 545, + 276 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 228, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\zeta = \\sum_{i=1}^{K}\\sum_{m',m''}\\mathrm{stm}(C^{1:N})_{m',m''}" + }, + { + "bbox": [ + 305, + 228, + 545, + 276 + ], + "type": "text", + "content": " is a normalization constant to ensure that " + }, + { + "bbox": [ + 305, + 228, + 545, + 276 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{sscp}}" + }, + { + "bbox": [ + 305, + 228, + 545, + 276 + ], + "type": "text", + "content": " is a valid probability distribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 277, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 277, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 545, + 312 + ], + "type": "text", + "content": "To evaluate a motion anticipation model " + }, + { + "bbox": [ + 305, + 277, + 545, + 312 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 277, + 545, + 312 + ], + "type": "text", + "content": ", we predict the future motion for all " + }, + { + "bbox": [ + 305, + 277, + 545, + 312 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 277, + 545, + 312 + ], + "type": "text", + "content": " test sequences from a fixed start frame " + }, + { + "bbox": [ + 305, + 277, + 545, + 312 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 277, + 545, + 312 + ], + "type": "text", + "content": " until the end of the sequence " + }, + { + "bbox": [ + 305, + 277, + 545, + 312 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 277, + 545, + 312 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 372, + 319, + 545, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 319, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 372, + 319, + 545, + 334 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {X}} ^ {n + 1: N} = \\left\\{f \\left(\\mathbf {X} _ {i} ^ {1: n}\\right) \\right\\} _ {i = 1} ^ {K} \\tag {13}", + "image_path": "66f09afcc08160f59af0175eaa3b0bfd69cd36b28fc96279b9502c087476f26b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 342, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 342, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 545, + 366 + ], + "type": "text", + "content": "We can now calculate the distance between the generated and ground-truth social motion distribution:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 353, + 374, + 545, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 374, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 353, + 374, + 545, + 389 + ], + "type": "interline_equation", + "content": "D _ {\\mathrm {J S D}} \\left(p _ {\\text {s s c p}} \\left(\\mathcal {X} ^ {n: N}\\right), p _ {\\text {s s c p}} \\left(\\hat {\\mathcal {X}} ^ {n: N}\\right)\\right). \\tag {14}", + "image_path": "1375921734bc7b721a980e57c2da0bf7cc7857b889ad74f67c9742523d4b23b2.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 397, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 397, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 306, + 397, + 545, + 410 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{JSD}}" + }, + { + "bbox": [ + 306, + 397, + 545, + 410 + ], + "type": "text", + "content": " is the squared Jensen-Shannon distance [17, 19]:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 428, + 545, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 428, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 315, + 428, + 545, + 460 + ], + "type": "interline_equation", + "content": "D _ {\\mathrm {J S D}} (p \\| q) = \\sqrt {\\frac {\\left(D _ {\\mathrm {K L}} (p \\| \\frac {p + q}{2}) + D _ {\\mathrm {K L}} (q \\| \\frac {p + q}{2})\\right)}{2}} \\tag {15}", + "image_path": "e4231d78716c280ef24c2406f19b889df9831438f030ecd33bc912f9f8a988fd.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 468, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 468, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 305, + 468, + 545, + 526 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 468, + 545, + 526 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 468, + 545, + 526 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 468, + 545, + 526 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 305, + 468, + 545, + 526 + ], + "type": "text", + "content": " are probability distributions and " + }, + { + "bbox": [ + 305, + 468, + 545, + 526 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{KL}}" + }, + { + "bbox": [ + 305, + 468, + 545, + 526 + ], + "type": "text", + "content": " is the Kullback-Leibler divergence. Note that we compare the generated motion " + }, + { + "bbox": [ + 305, + 468, + 545, + 526 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{X}}^{n + 1:N}" + }, + { + "bbox": [ + 305, + 468, + 545, + 526 + ], + "type": "text", + "content": " to the test set " + }, + { + "bbox": [ + 305, + 468, + 545, + 526 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^{n + 1:N}" + }, + { + "bbox": [ + 305, + 468, + 545, + 526 + ], + "type": "text", + "content": " with the same start frame as the data distribution might shift across time." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 538, + 388, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 388, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 388, + 551 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": "We evaluate our approach on four datasets. Following [69], we report the Mean Per Joint Positional Error (MPJPE) in global and local aligned coordinates for the multi-person human motion datasets MuPoTS-3D [43], 3DPW [67] and CMU-Mocap [1, 69]. MuPoTS-3D [43] contains recordings of 2 to 3 persons in workout settings. Interactions between the persons are rare. 3DPW [67] contains recordings of 1 to 2 persons and the sequences cover a wide range of different activities. The level of interactions range from no interaction and little interaction, like two persons walking, to close interactions such as dancing. CMU [69] combines the motion of different sequences from the CMU-Mocap dataset [1]. The composition of motion" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9604" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 74, + 147, + 178 + ], + "blocks": [ + { + "bbox": [ + 53, + 74, + 147, + 178 + ], + "lines": [ + { + "bbox": [ + 53, + 74, + 147, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 74, + 147, + 178 + ], + "type": "image", + "image_path": "6b9b39020ac79ca53eaacf698fb15ac23c04088ea147a007a04469452421f5d0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 187, + 104, + 197 + ], + "lines": [ + { + "bbox": [ + 93, + 187, + 104, + 197 + ], + "spans": [ + { + "bbox": [ + 93, + 187, + 104, + 197 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 170, + 86, + 286, + 198 + ], + "blocks": [ + { + "bbox": [ + 170, + 86, + 286, + 198 + ], + "lines": [ + { + "bbox": [ + 170, + 86, + 286, + 198 + ], + "spans": [ + { + "bbox": [ + 170, + 86, + 286, + 198 + ], + "type": "image", + "image_path": "1c897da6acc91b72d6a44cd9612768f56ff0ed37df0ed6489ee1db582f7ba97f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 207, + 287, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 207, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 287, + 278 + ], + "type": "text", + "content": "Figure 3: A sample frame from the Haggling dataset [30] for evaluating social interactions. (a): 3D poses in a haggling sequence. Blue limbs represent the left body side while red limbs represent the right body side. The buyer's attention is indicated as green arrow. (b): a sample video frame from a haggling sequence." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 293, + 287, + 361 + ], + "blocks": [ + { + "bbox": [ + 48, + 293, + 287, + 361 + ], + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 361 + ], + "type": "table", + "html": "
MethodCMU-MocapMuPoTS-3D3DPW
1s2s3s1s2s3s1s2s3s
LTD [41]1.372.193.261.191.812.344.677.108.71
HRI [40]1.492.603.070.941.682.294.076.328.01
SP [2]1.152.713.900.921.672.514.177.179.27
MRT [69]0.961.572.180.891.592.223.876.127.83
Ours0.741.061.341.151.291.441.642.723.55
", + "image_path": "a3a592b1dae0819fe7503d7051d04ff9b3ef3d0ce5c73b097ddaa5ef6fdb9aee.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 365, + 259, + 376 + ], + "lines": [ + { + "bbox": [ + 74, + 365, + 259, + 376 + ], + "spans": [ + { + "bbox": [ + 74, + 365, + 259, + 376 + ], + "type": "text", + "content": "Table 1: MPJPE " + }, + { + "bbox": [ + 74, + 365, + 259, + 376 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 74, + 365, + 259, + 376 + ], + "type": "text", + "content": " in dm on different datasets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "text", + "content": "sequences, however, does not reflect realistic interactions. For these datasets, 1 second of human motion is observed and 1-3 seconds need to be forecast. As our model is generative, we sample 8 samples for each test sequence and report the average over all 8 samples. The source code including trained evaluation models and the dataset transformation script is publicly available1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 483, + 208, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 208, + 496 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 208, + 496 + ], + "type": "text", + "content": "5.1. Haggling Forecasting Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 502, + 286, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 286, + 679 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 286, + 679 + ], + "type": "text", + "content": "Since the three datasets contain multiple persons, but very few social interactions, we prepared a new dataset for multi-person forecasting in the context of social interactions. For this, we utilize the Haggling dataset [30] where 122 participants play a social game with two sellers trying to sell their products to a buyer. Each game lasts one minute and contains interesting triadic interactions such as turn-taking and attention changes. A sample scene is shown in Figure 3. The dataset consists of 135 training sequences and 40 test sequences, sampled at " + }, + { + "bbox": [ + 46, + 502, + 286, + 679 + ], + "type": "inline_equation", + "content": "30\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 502, + 286, + 679 + ], + "type": "text", + "content": ". Some of the 3D poses in the dataset are noisy as they have been estimated [29, 31] and we manually correct them. In total, the dataset consists of 234, 907 training and 69, 951 test frames, each with three people. For more details, please see the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 681, + 286, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 681, + 286, + 693 + ], + "spans": [ + { + "bbox": [ + 58, + 681, + 286, + 693 + ], + "type": "text", + "content": "For evaluation, we take the first " + }, + { + "bbox": [ + 58, + 681, + 286, + 693 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 58, + 681, + 286, + 693 + ], + "type": "text", + "content": " of a sequence" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 239 + ], + "type": "text", + "content": "as observation and forecast the remaining " + }, + { + "bbox": [ + 306, + 72, + 545, + 239 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 306, + 72, + 545, + 239 + ], + "type": "text", + "content": " of the sequence, but we also evaluate the motion at intermediate frames ranging from frame 1 to frame 1300. It needs to be noted that long-term forecasting is highly relevant for neuroscience and social sciences since it allows to develop computational models on how the behavior of other persons is perceived and how it changes the own behavior. For measuring the plausibility of the forecast motion of each individual, we use the Normalized Directional Motion Similarity (NDMS) [61] since the measure not only considers static poses but also the motion of the forecast sequence. Furthermore, it can be applied to sequences of any length. NDMS, however, does not measure if the social interactions are plausible." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 240, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 324 + ], + "type": "text", + "content": "For evaluating the social motion quality, we utilize the proposed Symbolic Social Cues Protocol described in Section 4. To this end, we need to specify the classes of social interactions. A haggling activity is composed of the sellers trying to convince the buyer to purchase their products and the buyer switching attention between the sellers. Throughout the game, certain social patterns emerged [56, 36]:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 329, + 545, + 420 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 313, + 329, + 509, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 509, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 509, + 341 + ], + "type": "text", + "content": "1. Most of the time, only a single person speaks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 343, + 532, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 532, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 532, + 356 + ], + "type": "text", + "content": "2. For almost all frames, at least one person is talking" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 357, + 544, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 544, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 544, + 380 + ], + "type": "text", + "content": "3. The sellers speak roughly the same amount of time while the buyer seldom talks" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 384, + 536, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 536, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 536, + 396 + ], + "type": "text", + "content": "4. The buyer pays attention (looks at) to whoever talks" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 398, + 545, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 545, + 420 + ], + "type": "text", + "content": "5. The sellers take turns to speak but sometimes interrupt each other" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 426, + 545, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 463 + ], + "type": "text", + "content": "Given the well-defined structure of the task and the emerging social behaviors, we reduce the haggling game to two key signals:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 468, + 544, + 506 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 317, + 468, + 451, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 451, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 451, + 480 + ], + "type": "text", + "content": "- talking: defines who is talking" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 483, + 544, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 483, + 544, + 506 + ], + "spans": [ + { + "bbox": [ + 317, + 483, + 544, + 506 + ], + "type": "text", + "content": "- attention: defines who of the two sellers has the buyer's attention" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 511, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 511, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 511, + 545, + 594 + ], + "type": "text", + "content": "Given all possible combinations (e.g., both sellers can talk at the same time, or nobody talks), we end up with 16 possible states for each frame and formulate the social interactions as a symbolic representation over time. Note that we have to distinguish between left/right seller to catch events such as attention switching. Please see the supplementary material for more details." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 595, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 595, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 595, + 545, + 666 + ], + "type": "text", + "content": "We define the social signal function " + }, + { + "bbox": [ + 304, + 595, + 545, + 666 + ], + "type": "inline_equation", + "content": "s(\\mathbf{X}^{1:N})" + }, + { + "bbox": [ + 304, + 595, + 545, + 666 + ], + "type": "text", + "content": " as Equation (11), which takes a sequence of multi-person motion " + }, + { + "bbox": [ + 304, + 595, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{1:N}" + }, + { + "bbox": [ + 304, + 595, + 545, + 666 + ], + "type": "text", + "content": " as input and generates one of the 16 distinct states per frame. For the social signal function " + }, + { + "bbox": [ + 304, + 595, + 545, + 666 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 595, + 545, + 666 + ], + "type": "text", + "content": " to work on any haggling motion sequence, we need to determine three pieces of information:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 673, + 482, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 313, + 673, + 397, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 397, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 397, + 685 + ], + "type": "text", + "content": "1. who the buyer is," + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 687, + 482, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 687, + 482, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 687, + 482, + 700 + ], + "type": "text", + "content": "2. whom the buyer is paying attention to," + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 702, + 448, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 702, + 448, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 702, + 448, + 713 + ], + "type": "text", + "content": "3. whether someone is speaking." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 270, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 270, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 270, + 712 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9605" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 78, + 70, + 515, + 147 + ], + "blocks": [ + { + "bbox": [ + 78, + 70, + 515, + 147 + ], + "lines": [ + { + "bbox": [ + 78, + 70, + 515, + 147 + ], + "spans": [ + { + "bbox": [ + 78, + 70, + 515, + 147 + ], + "type": "table", + "html": "
MethodCMU-MocapMuPoTS-3D3DPW
RootPoseRootPoseRootPose
1s2s3s1s2s3s1s2s3s1s2s3s1s2s3s
LTD [41]0.971.732.620.981.211.370.891.391.910.881.141.314.286.798.41
HRI [40]0.962.063.111.051.371.580.661.302.160.731.071.303.676.428.64
SP [2]0.962.012.961.031.411.710.961.382.210.721.081.303.766.869.07
MRT [69]0.601.121.710.791.051.220.671.251.860.690.991.193.425.697.30
Ours0.721.101.440.380.460.491.141.281.420.590.640.671.662.763.59
", + "image_path": "1226f927eb08124c4d7851bc98d60e5b06f042f0a3a32508c1b1b80aef792b89.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 165, + 532, + 224 + ], + "blocks": [ + { + "bbox": [ + 67, + 151, + 522, + 163 + ], + "lines": [ + { + "bbox": [ + 67, + 151, + 522, + 163 + ], + "spans": [ + { + "bbox": [ + 67, + 151, + 522, + 163 + ], + "type": "text", + "content": "Table 2: MPJPE " + }, + { + "bbox": [ + 67, + 151, + 522, + 163 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 67, + 151, + 522, + 163 + ], + "type": "text", + "content": " in dm for root joint and pose. The lowest error is in bold and the second lowest is underscored." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 165, + 532, + 224 + ], + "lines": [ + { + "bbox": [ + 61, + 165, + 532, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 165, + 532, + 224 + ], + "type": "table", + "html": "
Frame1510152025306012025050075010001300
MRT [69]0.6240.2780.1940.2120.2240.2150.2150.2180.2050.1800.1290.0790.0620.047
Ours (Γ∅)0.6440.2800.2060.2150.2250.2260.2290.2330.2290.2260.2290.2270.2250.226
Ours (Γattn)0.6390.2800.1990.2130.2240.2290.2320.2270.2290.2230.2330.2280.2330.223
Ours* (ΓE)0.6400.2790.2040.2160.2270.2280.2300.2330.2270.2220.2300.2290.2300.226
", + "image_path": "9baabd959d37a56a7ba1730459ec6c95dafcceb9e99060f1492cf2d1030ed631.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 228, + 546, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 546, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 546, + 251 + ], + "type": "text", + "content": "Table 3: Per-frame average NDMS ↑ score on the Haggling dataset. The highest score is in bold and the second highest is underscored." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 264, + 286, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 286, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 286, + 346 + ], + "type": "text", + "content": "To solve (1) we train a simple buyer detection network, consisting of three layers of bi-directional Gated Recurrent Units, which gets as input a haggling motion and outputs the likelihood of each participant being the buyer. In Table 5, we report our accuracy of this approach. We see that the buyer detector correctly identifies the buyer all the time on the test set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 347, + 286, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 286, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 286, + 383 + ], + "type": "text", + "content": "For (2), we define the buyer's attention as whomever they look at, which can be easily calculated from the 3D body pose:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 392, + 286, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 392, + 286, + 420 + ], + "spans": [ + { + "bbox": [ + 113, + 392, + 286, + 420 + ], + "type": "interline_equation", + "content": "\\underset {i \\in \\{\\text {l e f t}, \\text {r i g h t} \\}} {\\operatorname {a r g m i n}} \\left[ n ^ {T} \\left(\\frac {d _ {i}}{| d _ {i} |}\\right) \\right] \\tag {16}", + "image_path": "c52fd34fdb0d1a5f60bf23717545b11f99021025445b82a50db61e2d21ed949e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 429, + 286, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 286, + 488 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 286, + 488 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 429, + 286, + 488 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{left}}" + }, + { + "bbox": [ + 46, + 429, + 286, + 488 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 429, + 286, + 488 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{right}}" + }, + { + "bbox": [ + 46, + 429, + 286, + 488 + ], + "type": "text", + "content": " are the directional vectors from the buyer nose to the left and right seller nose, respectively, projected onto the ground plane and " + }, + { + "bbox": [ + 46, + 429, + 286, + 488 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 429, + 286, + 488 + ], + "type": "text", + "content": " is the 2D unit vector that is perpendicular to the left eye→right eye vector of the buyer, projected onto the ground plane." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 489, + 286, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 489, + 286, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 286, + 571 + ], + "type": "text", + "content": "Last but not least, to determine if someone is speaking we utilize an off-the-shelf action classification network consisting of three layers of bi-directional Gated Recurrent Units. For training, we use the annotation of the Haggling dataset [30] which indicates if a person is speaking or not. The classifier achieves " + }, + { + "bbox": [ + 46, + 489, + 286, + 571 + ], + "type": "inline_equation", + "content": "87\\%" + }, + { + "bbox": [ + 46, + 489, + 286, + 571 + ], + "type": "text", + "content": " accuracy in speech detection on the test set." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 580, + 189, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 580, + 189, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 580, + 189, + 594 + ], + "type": "text", + "content": "5.2. Multi-Person Forecasting" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 599, + 286, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 286, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 286, + 694 + ], + "type": "text", + "content": "We first report the results for the multi-person human motion datasets MuPoTS-3D " + }, + { + "bbox": [ + 46, + 599, + 286, + 694 + ], + "type": "inline_equation", + "content": "[43]^2" + }, + { + "bbox": [ + 46, + 599, + 286, + 694 + ], + "type": "text", + "content": ", 3DPW " + }, + { + "bbox": [ + 46, + 599, + 286, + 694 + ], + "type": "inline_equation", + "content": "[67]^1" + }, + { + "bbox": [ + 46, + 599, + 286, + 694 + ], + "type": "text", + "content": " and CMU-Mocap [1, 69]. We follow [69] and report the Mean Per Joint Positional Error (MPJPE) using global coordinates in Table 1. Our approach outperforms the methods LTD [41], HRI [40], SP [2], and MRT [69] by a large margin. While some methods perform better for the first second on MuPoTS-3D [43], our approach achieves a much lower" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 264, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 264, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 264, + 545, + 407 + ], + "type": "text", + "content": "error for all other settings and datasets. On the most difficult dataset 3DPW [67], the error is reduced by " + }, + { + "bbox": [ + 306, + 264, + 545, + 407 + ], + "type": "inline_equation", + "content": "57.6\\%" + }, + { + "bbox": [ + 306, + 264, + 545, + 407 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 264, + 545, + 407 + ], + "type": "inline_equation", + "content": "55.6\\%" + }, + { + "bbox": [ + 306, + 264, + 545, + 407 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 306, + 264, + 545, + 407 + ], + "type": "inline_equation", + "content": "54.7\\%" + }, + { + "bbox": [ + 306, + 264, + 545, + 407 + ], + "type": "text", + "content": " for 1, 2, and 3 seconds, respectively. As in [69], we also report the error of the position of the root joint and pose error in local coordinates, i.e., setting the root position for all frames to zero, in Table 2. The results show that our approach forecasts by far the most accurate poses and outperforms the state of the art by a large margin on all datasets. Only the position of the root joint is slightly better estimated by other methods at the beginning of the datasets MuPoTS-3D and CMU-Mocap. At 3 seconds, our approach also achieves the lowest root joint error for all three datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 418, + 545, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 545, + 441 + ], + "type": "text", + "content": "5.3. Multi-Person Forecasting in the Context of Social Interactions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "content": "For the remaining experiments, we evaluate our approach on the newly prepared Haggling dataset since it contains more social interactions as the other datasets. We compare our approach to Multi-Range Transformers (MRT) [69], which performed better than other approaches in Section 5.2. We used the publicly available source code and adjusted the approach to work with " + }, + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "inline_equation", + "content": "30\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "content": ". We kept all other settings as is." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "We first report the per-frame NDMS (higher is better) at different frames in Table 3. Our experiments show that MRT [69] is capable of generating realistic motion for a few seconds. However, after 120 frames " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "(4s)" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " the NDMS score drops significantly. This is caused by the auto-regressive motion forecasting strategy adopted by MRT, which results in error accumulation over time. In contrast, our method continues to predict motions with good NDMS scores well into the future. As discussed in Section 3.2, we compare different variants of the aggregation function " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ". In terms of " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ", there are no major differences in terms of NDMS, but they all perform much better than MRT [69]. In Table 4, we report the average NDMS score over the entire forecast sequence. We observe that " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\emptyset}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ", i.e., using no aggregation" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 271, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 271, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 271, + 713 + ], + "type": "text", + "content": "2Data access and processing was conducted at University of Bonn" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9606" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 92, + 72, + 506, + 167 + ], + "blocks": [ + { + "bbox": [ + 92, + 72, + 506, + 167 + ], + "lines": [ + { + "bbox": [ + 92, + 72, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 92, + 72, + 506, + 167 + ], + "type": "image", + "image_path": "709b96924f85b85bbe64e7fb4265dcd2ab2eef26a158055020590513ef8ce370.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 178, + 547, + 215 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 547, + 215 + ], + "type": "text", + "content": "Figure 4: Our model is capable of generating realistic motion for 7 people from the Ultimatum sequence of Panoptic Studio [29], even though it was trained only on the triadic Haggling dataset. The Ultimatum sequence shares similarities with the Haggling dataset such as persons taking turns and talking to each other." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 232, + 126, + 312 + ], + "blocks": [ + { + "bbox": [ + 48, + 232, + 126, + 312 + ], + "lines": [ + { + "bbox": [ + 48, + 232, + 126, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 126, + 312 + ], + "type": "image", + "image_path": "b45c9a22f0fe144b9ebd6dce917f20f856ccfc66c6c2e1c8ed3efb2e41e59d9c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 315, + 104, + 325 + ], + "lines": [ + { + "bbox": [ + 70, + 315, + 104, + 325 + ], + "spans": [ + { + "bbox": [ + 70, + 315, + 104, + 325 + ], + "type": "text", + "content": "(a) Train" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 129, + 233, + 206, + 311 + ], + "blocks": [ + { + "bbox": [ + 129, + 233, + 206, + 311 + ], + "lines": [ + { + "bbox": [ + 129, + 233, + 206, + 311 + ], + "spans": [ + { + "bbox": [ + 129, + 233, + 206, + 311 + ], + "type": "image", + "image_path": "056c3f8fa3a1fd91bbf99d322c31945972cb75601fa53c01b603d9f04bfe4db9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 315, + 182, + 325 + ], + "lines": [ + { + "bbox": [ + 153, + 315, + 182, + 325 + ], + "spans": [ + { + "bbox": [ + 153, + 315, + 182, + 325 + ], + "type": "text", + "content": "(b) Test" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 209, + 233, + 287, + 311 + ], + "blocks": [ + { + "bbox": [ + 209, + 233, + 287, + 311 + ], + "lines": [ + { + "bbox": [ + 209, + 233, + 287, + 311 + ], + "spans": [ + { + "bbox": [ + 209, + 233, + 287, + 311 + ], + "type": "image", + "image_path": "facd8100c3df6afb7dae8aa33c167cf9907e223f8682552b37055c2069be6ce2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 315, + 272, + 325 + ], + "lines": [ + { + "bbox": [ + 223, + 315, + 272, + 325 + ], + "spans": [ + { + "bbox": [ + 223, + 315, + 272, + 325 + ], + "type": "text", + "content": "(c) MRT [69]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 48, + 327, + 126, + 407 + ], + "blocks": [ + { + "bbox": [ + 48, + 327, + 126, + 407 + ], + "lines": [ + { + "bbox": [ + 48, + 327, + 126, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 126, + 407 + ], + "type": "image", + "image_path": "cd5486f2618286e99fd02a717172e1f952528762bc94c6659bb78edc1d25392f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 411, + 113, + 421 + ], + "lines": [ + { + "bbox": [ + 61, + 411, + 113, + 421 + ], + "spans": [ + { + "bbox": [ + 61, + 411, + 113, + 421 + ], + "type": "text", + "content": "(d) Ours " + }, + { + "bbox": [ + 61, + 411, + 113, + 421 + ], + "type": "inline_equation", + "content": "(\\Gamma_{\\emptyset})" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 128, + 327, + 206, + 406 + ], + "blocks": [ + { + "bbox": [ + 128, + 327, + 206, + 406 + ], + "lines": [ + { + "bbox": [ + 128, + 327, + 206, + 406 + ], + "spans": [ + { + "bbox": [ + 128, + 327, + 206, + 406 + ], + "type": "image", + "image_path": "5b7e675e708f7ced81c92bb2c06402d795927a22c5c7abefb1953335f4d43589.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 411, + 195, + 422 + ], + "lines": [ + { + "bbox": [ + 140, + 411, + 195, + 422 + ], + "spans": [ + { + "bbox": [ + 140, + 411, + 195, + 422 + ], + "type": "text", + "content": "(e) Ours " + }, + { + "bbox": [ + 140, + 411, + 195, + 422 + ], + "type": "inline_equation", + "content": "(\\Gamma_{\\mathrm{att}})" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 209, + 327, + 287, + 406 + ], + "blocks": [ + { + "bbox": [ + 209, + 327, + 287, + 406 + ], + "lines": [ + { + "bbox": [ + 209, + 327, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 209, + 327, + 287, + 406 + ], + "type": "image", + "image_path": "eafb11b4ab620b18df8e59ae751ba1e65bf5c300866930aece59300ed1b2fb6b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 411, + 272, + 422 + ], + "lines": [ + { + "bbox": [ + 223, + 411, + 272, + 422 + ], + "spans": [ + { + "bbox": [ + 223, + 411, + 272, + 422 + ], + "type": "text", + "content": "(f) Ours " + }, + { + "bbox": [ + 223, + 411, + 272, + 422 + ], + "type": "inline_equation", + "content": "(\\Gamma_{\\mathbb{E}})" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 431, + 289, + 526 + ], + "lines": [ + { + "bbox": [ + 46, + 431, + 289, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 289, + 526 + ], + "type": "text", + "content": "Figure 5: State transition matrices for the states defined in Section 5.1 for the ground-truth of the train (a) and test set (b). The other transition matrices are obtained by the forecast of the current state-of-the-art model MRT [69] on the test set (c) and our model without context (d), with context via attention (e), and with context via averaging (f). The more similar the transition matrix is to the test set (b), the closer it matches the test motion." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 48, + 552, + 295, + 600 + ], + "blocks": [ + { + "bbox": [ + 48, + 552, + 295, + 600 + ], + "lines": [ + { + "bbox": [ + 48, + 552, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 552, + 295, + 600 + ], + "type": "table", + "html": "
Train[69]Ours (Γ∅)Ours (Γattn)Ours*(ΓR)
NDMS ↑-0.10150.23010.22700.2297
SSCP ↓0.09990.48390.35760.32780.3252
", + "image_path": "dc9962a46508ee2be10cbb321a91471fd306d29bba5da7ce4f61472f60a7d655.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "information yields a slightly higher NDMS score than the other versions, but the differences are very small. However, " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathrm{attn}}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathbb{E}}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " forecast much more plausible social interactions (lower SSCP) as we will discuss next." + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 306, + 233, + 547, + 269 + ], + "blocks": [ + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "lines": [ + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "type": "text", + "content": "Table 4: Per-frame average NDMS " + }, + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "type": "text", + "content": " and SSCP " + }, + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 603, + 287, + 628 + ], + "type": "text", + "content": " on the Haggling dataset." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 233, + 547, + 269 + ], + "lines": [ + { + "bbox": [ + 306, + 233, + 547, + 269 + ], + "spans": [ + { + "bbox": [ + 306, + 233, + 547, + 269 + ], + "type": "table", + "html": "
Test setMRT [69]Ours (Γ∅)Ours (Γattn)Ours* (ΓE)
1.00.32500.88120.88750.8938
", + "image_path": "ef9e533d790c8371790df9729d635773d76d045bd133993ce89cdd3118968002.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 272, + 547, + 321 + ], + "lines": [ + { + "bbox": [ + 305, + 272, + 547, + 321 + ], + "spans": [ + { + "bbox": [ + 305, + 272, + 547, + 321 + ], + "type": "text", + "content": "Table 5: Accuracy " + }, + { + "bbox": [ + 305, + 272, + 547, + 321 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 305, + 272, + 547, + 321 + ], + "type": "text", + "content": " of the buyer detection network. For MRT [69], the model randomly selects a person to be the buyer, as there is a " + }, + { + "bbox": [ + 305, + 272, + 547, + 321 + ], + "type": "inline_equation", + "content": "1/3" + }, + { + "bbox": [ + 305, + 272, + 547, + 321 + ], + "type": "text", + "content": " chance of selecting the buyer with random chance." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 333, + 448, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 333, + 448, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 333, + 448, + 346 + ], + "type": "text", + "content": "5.3.1 Social Motion Evaluation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "text", + "content": "The SSCP scores are presented in Table 4 where a lower value corresponds to more plausible forecast social interactions. All our proposed variants outperform the state-of-the-art approach MRT [69]. Using no aggregation information " + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "inline_equation", + "content": "(\\Gamma_{\\emptyset})" + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "text", + "content": " performs worse than the aggregation functions " + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathrm{attn}}" + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "text", + "content": " (9) and " + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathbb{E}}" + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "text", + "content": " (10), which is expected since the aggregation function passes information from other people in the scene. The average aggregation " + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "inline_equation", + "content": "(\\Gamma_{\\mathbb{E}})" + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "text", + "content": " performs slightly better than the more complex multi-headed attention approach " + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "inline_equation", + "content": "(\\Gamma_{\\mathrm{attn}})" + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "text", + "content": ". We conjecture that averaging bottleneck encodings over all people introduces an inductive bias to pay the same attention to everyone, which works well for modelling the haggling game. For completeness, we also report the SSCP score of the training set." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": "We can draw some insights about what motion each model generates by looking at the state probability transition matrices in Figure 5. For example, MRT [69] (Figure 5d) produces mostly self-loops (diagonal of transition matrix) indicating that the motion gets stuck over time. When no context information is provided " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "(\\Gamma_{\\emptyset})" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": ", our method produces motions where all three people are talking at the same time, as can be seen in Figure 5d, where the last two entries (red arrow) in the transition matrix represent states with all three people talking. This is sensible as the model sees two sellers and only one buyer during training and thus it is more likely to produce motion that resembles a seller, who talks most of the time. When the context is provided, our approach overcomes this limitation as expected and rarely produces motion where all three people are talking at the same time as shown in Figures 5e and 5f." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9607" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": "These observations are also confirmed when measuring the buyer detection accuracy on the forecast motion, which is reported in Table 5. The detector fails to identify the correct buyer in the sequences that are forecast by MRT [69] and it nearly chooses the buyer at random with " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "1/3" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": " accuracy. This confirms that MRT does not forecast socially consistent sequences where the social role of the persons, namely buyer or seller, is preserved. In contrast, our method predicts motion where the buyer can be easily determined most of the time. As for the SSCP scores reported in Table 4, " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathbb{E}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": " performs best." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 288, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 288, + 266 + ], + "type": "text", + "content": "In summary, the aggregation function " + }, + { + "bbox": [ + 46, + 205, + 288, + 266 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathbb{E}}" + }, + { + "bbox": [ + 46, + 205, + 288, + 266 + ], + "type": "text", + "content": " outperforms the other aggregation functions on the Haggling dataset [30] as it produces the most socially plausible motion according to our Symbolic Social Cues Protocol while also generating highly plausible 3D body motion." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 274, + 141, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 274, + 141, + 287 + ], + "spans": [ + { + "bbox": [ + 47, + 274, + 141, + 287 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 293, + 289, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 289, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 289, + 486 + ], + "type": "text", + "content": "Average velocity over time Freezing or unrealistically expanding motion are common failures in human motion anticipation. While NDMS [61] penalizes in contrast to MPJPE errors in the velocity, visualizing the average motion velocity can give interesting insights. In Figure 6, we plot the average velocity over all frames for all our summary function variants, the test set, and the state-of-the-art method MRT [69]. Note that the beginning of the sequence has a higher velocity due to people walking into the scene. We observe that MRT suffers from error accumulation caused by the auto-regressive inference scheme. The velocity produced by our motion tightly follows the test set velocity for roughly 250 frames after which the test set velocity is slightly larger. We attribute this to the higher degree of stochasticity of real motion, which results in sudden jerks and swings that increase the average velocity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 486, + 289, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 581 + ], + "type": "text", + "content": "NDMS score over time In Figure 7, we visualize how the NDMS scores of all proposed variants and MRT evolve over time. For reference, we also calculate the NDMS score of the training data which is guaranteed to be realistic. Note that NDMS is 1 for the observed part of the test sequences. As shown in the figure, our method achieves almost the same level of realism as the training data while the quality of MRT slowly degenerates over time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Anticipating more than three people We have trained and evaluated our model on the Haggling dataset where each sequence consists of triadic interaction. However, the fully convolutional nature of our approach as well as the order-invariance of the summarization function " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " allow us to forecast any number of people. To demonstrate this capability, we predict 7 people from the Ultimatum sequence of Panoptic Studio [29] using only the model trained on the Haggling dataset. This works well because the Haggling and Ultimatum sequences share many social behaviors, such as turn-taking, talking, and paying attention while" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 545, + 212 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 545, + 212 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 545, + 212 + ], + "type": "image", + "image_path": "ac00d8c38b4b636ae1aa8bd642f4cd9abdcd02230489aed81afffeaebd4d4790.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 221, + 547, + 270 + ], + "lines": [ + { + "bbox": [ + 304, + 221, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 547, + 270 + ], + "type": "text", + "content": "Figure 6: Average velocity over time for the entire test motion " + }, + { + "bbox": [ + 304, + 221, + 547, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 221, + 547, + 270 + ], + "type": "text", + "content": " and generated motions " + }, + { + "bbox": [ + 304, + 221, + 547, + 270 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{X}}" + }, + { + "bbox": [ + 304, + 221, + 547, + 270 + ], + "type": "text", + "content": ". The x-axis represents the frames while the y-axis represents the (log) average velocity of the data." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 274, + 545, + 417 + ], + "blocks": [ + { + "bbox": [ + 307, + 274, + 545, + 417 + ], + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 417 + ], + "type": "image", + "image_path": "5a351a73378648bf0e2250f7ae2b295144fbb2e523be95ef2281e01aa6347924.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 426, + 545, + 451 + ], + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 451 + ], + "type": "text", + "content": "Figure 7: Average NDMS [61] score " + }, + { + "bbox": [ + 304, + 426, + 545, + 451 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 304, + 426, + 545, + 451 + ], + "type": "text", + "content": " of all proposed variants, MRT, and the train set over time." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 464, + 547, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 547, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 547, + 524 + ], + "type": "text", + "content": "standing in a circle. Our results can be seen in Figure 4 where our model is able to predict realistic motion for 7 people, even though it was trained only on the triadic Haggling dataset. More results are provided in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 536, + 379, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 379, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 379, + 550 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "In this work, we present Social Diffusion, a stochastic multi-person motion anticipation model. The approach not only forecasts realistic motions on the individual level, but also plausible social interactions where the social roles of individuals are preserved over time. The approach is very flexible. It can be used for short and long-term forecasting and can be applied to larger groups than observed during training. As a second contribution, we proposed a new evaluation protocol to measure the realism of forecast social interactions. We furthermore derived a dataset for multi-person social interaction forecasting from the Haggling dataset [30] where the persons have different social roles that impact their behavior. We evaluated our approach" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9608" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "on four multi-person datasets and demonstrated that our approach outperforms the state-of-the-art for short-term and long-term anticipation both in realism of forecast motion and social interaction. The approach has still some limitations. For instance, the global positions of the root joints can be better estimated. Future directions also include extending the model to predict motions of dynamic groups of people, e.g., at a cocktail party where any individual can freely disengage from the current conversation group and join another one." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 206, + 145, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 206, + 145, + 220 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 145, + 220 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 227, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 227, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 289, + 360 + ], + "type": "text", + "content": "Juergen Gall has been supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) GA 1927/4-2 (FOR 2535 Anticipating Human Behavior), by the Federal Ministry of Education and Research (BMBF) under grant no. 01IS22094E WEST-AI, the project iBehave (receiving funding from the programme \"Netzwerke 2021\", an initiative of the Ministry of Culture and Science of the State of Northrhine Westphalia), and the ERC Consolidator Grant FORHUE (101044724). The sole responsibility for the content of this publication lies with the authors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 372, + 107, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 107, + 385 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 107, + 385 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 393, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 53, + 393, + 221, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 393, + 221, + 404 + ], + "spans": [ + { + "bbox": [ + 53, + 393, + 221, + 404 + ], + "type": "text", + "content": "[1] CMU. Carnegie-Mellon Mocap Database." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 406, + 287, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 406, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 53, + 406, + 287, + 450 + ], + "type": "text", + "content": "[2] Vida Adeli, Ehsan Adeli, Ian Reid, Juan Carlos Niebles, and Hamid Rezatofighi. Socially and contextually aware human motion and pose forecasting. Robotics and Automation Letters, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 453, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 287, + 508 + ], + "type": "text", + "content": "[3] Vida Adeli, Mahsa Ehsanpour, Ian Reid, Juan Carlos Niebles, Silvio Savarese, Ehsan Adeli, and Hamid Rezatofighi. Tripod: Human trajectory and pose dynamics forecasting in the wild. In International Conference on Computer Vision, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 510, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 287, + 552 + ], + "type": "text", + "content": "[4] Emre Aksan, Manuel Kaufmann, Peng Cao, and Otmar Hilliges. A spatio-temporal transformer for 3d human motion prediction. In International Conference on 3D Vision, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 555, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 555, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 53, + 555, + 287, + 609 + ], + "type": "text", + "content": "[5] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social lstm: Human trajectory prediction in crowded spaces. In Conference on Computer Vision and Pattern Recognition, 2016." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 612, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 612, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 53, + 612, + 287, + 666 + ], + "type": "text", + "content": "[6] Sadegh Aliakbarian, Fatemeh Sadat Saleh, Mathieu Salzmann, Lars Petersson, and Stephen Gould. A stochastic conditioning scheme for diverse human motion prediction. In Conference on Computer Vision and Pattern Recognition, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 52, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 52, + 669, + 287, + 713 + ], + "type": "text", + "content": "[7] Eduardo Alvarado, Damien Rohmer, and Marie-Paule Cani. Generating upper-body motion for real-time characters making their way through dynamic environments. Eurographics Symposium on Computer Animation, 2022." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 311, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 311, + 73, + 545, + 117 + ], + "type": "text", + "content": "[8] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 119, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 119, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 311, + 119, + 545, + 174 + ], + "type": "text", + "content": "[9] Timur Bagautdinov, Alexandre Alahi, François Fleuret, Pascal Fua, and Silvio Savarese. Social scene understanding: End-to-end multi-person action localization and collective activity recognition. In Conference on Computer Vision and Pattern Recognition, 2017." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 175, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 219 + ], + "type": "text", + "content": "[10] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 221, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 221, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 221, + 545, + 266 + ], + "type": "text", + "content": "[11] Judith Butepage, Michael J Black, Danica Kragic, and Hedvig Kjellstrom. Deep representation learning for human motion prediction and classification. In Conference on Computer Vision and Pattern Recognition, 2017." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 267, + 545, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 267, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 307, + 267, + 545, + 322 + ], + "type": "text", + "content": "[12] Yujun Cai, Lin Huang, Yiwei Wang, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Xu Yang, Yiheng Zhu, Xiaohui Shen, et al. Learning progressive joint propagation for human motion prediction. In European Conference on Computer Vision, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 323, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 323, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 323, + 545, + 380 + ], + "type": "text", + "content": "[13] Yujun Cai, Yiwei Wang, Yiheng Zhu, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Chuanxia Zheng, Sijie Yan, Henghui Ding, et al. A unified 3d human motion synthesis model via conditional variational auto-encoder. In International Conference on Computer Vision, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 381, + 545, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 381, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 307, + 381, + 545, + 435 + ], + "type": "text", + "content": "[14] Lujing Chen, Rui Liu, Xin Yang, Dongsheng Zhou, Qiang Zhang, and Xiaopeng Wei. Sttg-net: a spatio-temporal network for human motion prediction based on transformer and graph convolution network. Visual Computing for Industry, Biomedicine, and Art, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 437, + 545, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 471 + ], + "type": "text", + "content": "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 473, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 473, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 307, + 473, + 545, + 506 + ], + "type": "text", + "content": "[16] Christian Diller, Thomas Funkhouser, and Angela Dai. Forecasting characteristic 3d poses of human actions. Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 507, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 507, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 307, + 507, + 545, + 540 + ], + "type": "text", + "content": "[17] Dominik Maria Endres and Johannes E Schindelin. A new metric for probability distributions. Transactions on Information theory, 2003." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 543, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 543, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 307, + 543, + 545, + 576 + ], + "type": "text", + "content": "[18] Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models for human dynamics. In International Conference on Computer Vision, 2015." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "type": "text", + "content": "[19] Bent Fuglede and Flemming Topsoe. Jensen-Shannon divergence and Hilbert space embedding. In International Symposium on Information Theory, 2004." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "text", + "content": "[20] Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In Conference on Computer Vision and Pattern Recognition, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[21] Anand Gopalakrishnan, Ankur Mali, Dan Kifer, Lee Giles, and Alexander G Ororbia. A neural temporal model for human motion prediction. In Conference on Computer Vision and Pattern Recognition, 2019." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9609" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[22] Liang-Yan Gui, Yu-Xiong Wang, Xiaodan Liang, and José MF Moura. Adversarial geometry-aware human motion prediction. In European Conference on Computer Vision, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "type": "text", + "content": "[23] Wen Guo, Xiaoyu Bie, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Multi-person extreme motion prediction. In Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 195 + ], + "type": "text", + "content": "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 228 + ], + "type": "text", + "content": "[25] Daniel Holden, Jun Saito, Taku Komura, and Thomas Joyce. Learning motion manifolds with convolutional autoencoders. In SIGGRAPH Asia technical briefs. 2015." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "text", + "content": "[26] Emiel Hoogeboom, Didrik Nielsen, Priyank Jaini, Patrick Forre, and Max Welling. Argmax flows and multinomial diffusion: Learning categorical distributions. Advances in Neural Information Processing Systems, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "type": "text", + "content": "[27] Ashesh Jain, Amir R Zamir, Silvio Savarese, and Ashutosh Saxena. Structural-rnn: Deep learning on spatio-temporal graphs. In Conference on Computer Vision and Pattern Recognition, 2016." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "type": "text", + "content": "[28] Michael Janner, Yilun Du, Joshua B Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 288, + 415 + ], + "type": "text", + "content": "[29] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic Studio: A Massively Multiview System for Social Motion Capture. In International Conference on Computer Vision, 2015." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 459 + ], + "type": "text", + "content": "[30] Hanbyul Joo, Tomas Simon, Mina Cikara, and Yaser Sheikh. Towards social artificial intelligence: Nonverbal social signal prediction in a triadic interaction. In Conference on Computer Vision and Pattern Recognition, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 459, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 288, + 525 + ], + "type": "text", + "content": "[31] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Scott Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. Transactions on Pattern Analysis and Machine Intelligence, 2017." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 288, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 570 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 570 + ], + "type": "text", + "content": "[32] Julia Kantorovitch, Janne Väre, Vesa Pehkonen, Arto Laikari, and Heikki Seppälä. An assistive household robot-doing more than just cleaning. Journal of Assistive Technologies, 2014." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 288, + 613 + ], + "type": "text", + "content": "[33] Anne Keitel, Moritz M Daum, et al. The use of intonation for turn anticipation in observed conversations without visual signals as source of information. Frontiers in psychology, 2015." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "type": "text", + "content": "[34] Lukasz Kidziński, Bryan Yang, Jennifer L Hicks, Apoorva Rajagopal, Scott L Delp, and Michael H Schwartz. Deep neural networks enable quantitative movement analysis using single-camera videos. Nature communications, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[35] Vineet Kosaraju, Amir Sadeghian, Roberto Martin-Martin, Ian Reid, Hamid Rezatofighi, and Silvio Savarese. Socialbigat: Multimodal trajectory forecasting using bicycle-gan and graph attention networks. Advances in Neural Information Processing Systems, 2019." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "type": "text", + "content": "[36] Stephen C Levinson and Francisco Torreira. Timing in turn-taking and its implications for processing models of language. Frontiers in psychology, 2015." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 546, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 546, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 546, + 152 + ], + "type": "text", + "content": "[37] Chen Li, Zhen Zhang, Wee Sun Lee, and Gim Hee Lee. Convolutional sequence to sequence model for human dynamics. In Conference on Computer Vision and Pattern Recognition, 2018." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 546, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 546, + 198 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 546, + 198 + ], + "type": "text", + "content": "[38] Thomas Lucas*, Fabien Baradel*, Philippe Weinzaepfel, and Grégory Rogez. Posegpt: Quantization-based 3d human motion generation and forecasting. In European Conference on Computer Vision, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 198, + 546, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 546, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 546, + 243 + ], + "type": "text", + "content": "[39] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 245, + 546, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 245, + 546, + 279 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 546, + 279 + ], + "type": "text", + "content": "[40] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. History repeats itself: Human motion prediction via motion attention. In European Conference on Computer Vision, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 280, + 546, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 280, + 546, + 324 + ], + "spans": [ + { + "bbox": [ + 307, + 280, + 546, + 324 + ], + "type": "text", + "content": "[41] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies for human motion prediction. International Conference on Computer Vision, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 325, + 546, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 325, + 546, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 325, + 546, + 369 + ], + "type": "text", + "content": "[42] Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Conference on Computer Vision and Pattern Recognition, 2017." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 371, + 546, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 546, + 426 + ], + "type": "text", + "content": "[43] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In International Conference on 3D Vision, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 428, + 546, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 428, + 546, + 472 + ], + "spans": [ + { + "bbox": [ + 307, + 428, + 546, + 472 + ], + "type": "text", + "content": "[44] Ronja Möller, Antonino Furnari, Sebastiano Battiato, Aki Härma, and Giovanni Maria Farinella. A survey on human-aware robot navigation. Robotics and Autonomous Systems, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 475, + 546, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 546, + 519 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 546, + 519 + ], + "type": "text", + "content": "[45] Brendan Tran Morris and Mohan Manubhai Trivedi. A survey of vision-based trajectory learning and analysis for surveillance. Transactions on Circuits and Systems for Video Technology, 2008." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 520, + 546, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 546, + 565 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 546, + 565 + ], + "type": "text", + "content": "[46] Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. Learning to listen: Modeling non-deterministic dyadic facial motion. In Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 566, + 546, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 566, + 546, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 566, + 546, + 632 + ], + "type": "text", + "content": "[47] Sangmin Oh, Anthony Hoogs, Amitha Perera, Naresh Cunteor, Chia-Chih Chen, Jong Taek Lee, Saurajit Mukherjee, JK Aggarwal, Hyungtae Lee, Larry Davis, et al. A large-scale benchmark dataset for event recognition in surveillance video. In Conference on Computer Vision and Pattern Recognition, 2011." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 634, + 546, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 634, + 546, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 634, + 546, + 678 + ], + "type": "text", + "content": "[48] Dario Pavllo, Christoph Feichtenhofer, Michael Auli, and David Grangier. Modeling human motion with quaternion-based neural networks. International Journal of Computer Vision, 2020." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 679, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 679, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 679, + 546, + 713 + ], + "type": "text", + "content": "[49] Dario Pavllo, David Grangier, and Michael Auli. *Quaternion: A quaternion-based recurrent model for human motion*. In *British Machine Vision Conference*, 2018." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9610" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[50] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. European Conference on Computer Vision, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "type": "text", + "content": "[51] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 153, + 286, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 153, + 286, + 197 + ], + "spans": [ + { + "bbox": [ + 49, + 153, + 286, + 197 + ], + "type": "text", + "content": "[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 199, + 286, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 199, + 286, + 243 + ], + "spans": [ + { + "bbox": [ + 49, + 199, + 286, + 243 + ], + "type": "text", + "content": "[53] Alejandro Hernandez Ruiz, Juergen Gall, and Francesc Moreno-Noguer. Human motion prediction via spatiotemporal inpainting. International Conference on Computer Vision, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 245, + 286, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 245, + 286, + 288 + ], + "spans": [ + { + "bbox": [ + 49, + 245, + 286, + 288 + ], + "type": "text", + "content": "[54] Harvey Sacks, Emanuel A Schegloff, and Gail Jefferson. A simplest systematics for the organization of turn taking for conversation. In Studies in the organization of conversational interaction. 1978." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 290, + 286, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 290, + 286, + 334 + ], + "spans": [ + { + "bbox": [ + 49, + 290, + 286, + 334 + ], + "type": "text", + "content": "[55] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In Transactions on Graphics, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 336, + 286, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 336, + 286, + 379 + ], + "spans": [ + { + "bbox": [ + 49, + 336, + 286, + 379 + ], + "type": "text", + "content": "[56] Keren Shavit-Cohen and Elana Zion Golumbic. The dynamics of attention shifts among concurrent speech in a naturalistic multi-speaker virtual environment. Frontiers in Human Neuroscience, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 382, + 286, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 382, + 286, + 426 + ], + "spans": [ + { + "bbox": [ + 49, + 382, + 286, + 426 + ], + "type": "text", + "content": "[57] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, 2015." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 428, + 286, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 428, + 286, + 471 + ], + "spans": [ + { + "bbox": [ + 49, + 428, + 286, + 471 + ], + "type": "text", + "content": "[58] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 2015." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 474, + 286, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 474, + 286, + 517 + ], + "spans": [ + { + "bbox": [ + 49, + 474, + 286, + 517 + ], + "type": "text", + "content": "[59] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 519, + 286, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 519, + 286, + 563 + ], + "spans": [ + { + "bbox": [ + 49, + 519, + 286, + 563 + ], + "type": "text", + "content": "[60] Chen Sun, Abhinav Shrivastava, Carl Vondrick, Rahul Sukthankar, Kevin Murphy, and Cordelia Schmid. Relational action forecasting. In Conference on Computer Vision and Pattern Recognition, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 565, + 286, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 565, + 286, + 597 + ], + "spans": [ + { + "bbox": [ + 49, + 565, + 286, + 597 + ], + "type": "text", + "content": "[61] Julian Tanke, Chintan Zaveri, and Juergen Gall. Intention-based long-term human motion anticipation. In International Conference on 3D Vision, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 600, + 286, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 600, + 286, + 632 + ], + "spans": [ + { + "bbox": [ + 49, + 600, + 286, + 632 + ], + "type": "text", + "content": "[62] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 635, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 635, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 49, + 635, + 286, + 667 + ], + "type": "text", + "content": "[63] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 2017." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 670, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 286, + 713 + ], + "type": "text", + "content": "[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 2017." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 532 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[65] Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. International Conference on Learning Representations, 2017." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 162 + ], + "type": "text", + "content": "[66] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European Conference on Computer Vision, 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "type": "text", + "content": "[67] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European conference on computer vision, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 251 + ], + "type": "text", + "content": "[68] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 253, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 296 + ], + "type": "text", + "content": "[69] Jiashun Wang, Huazhe Xu, Medhini Narasimhan, and Xiaolong Wang. Multi-person 3d motion prediction with multirange transformers. Advances in Neural Information Processing Systems, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 297, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 341 + ], + "type": "text", + "content": "[70] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In Conference on Computer Vision and Pattern Recognition, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 342, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 374 + ], + "type": "text", + "content": "[71] Ye Yuan and Kris Kitani. DLOW: Diversifying latent flows for diverse human motion prediction. In European Conference on Computer Vision, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 376, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 409 + ], + "type": "text", + "content": "[72] Jason Y Zhang, Panna Felsen, Angjoo Kanazawa, and Jitendra Malik. Predicting 3d human dynamics from video. In International Conference on Computer Vision, 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "type": "text", + "content": "[73] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "type": "text", + "content": "[74] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J Black, and Siyu Tang. Generating 3d people in scenes without people. In Conference on computer vision and pattern recognition, 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 500, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 532 + ], + "type": "text", + "content": "[75] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3d scenes. In Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "9611" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Sound Localization from Motion_ Jointly Learning Sound Direction and Camera Rotation/ebfa790a-8a91-4c3c-9e52-be53fe87e14f_content_list.json b/2023/Sound Localization from Motion_ Jointly Learning Sound Direction and Camera Rotation/ebfa790a-8a91-4c3c-9e52-be53fe87e14f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4ff8041f221e0614ff89e008ceb9064e518f2912 --- /dev/null +++ b/2023/Sound Localization from Motion_ Jointly Learning Sound Direction and Camera Rotation/ebfa790a-8a91-4c3c-9e52-be53fe87e14f_content_list.json @@ -0,0 +1,1786 @@ +[ + { + "type": "text", + "text": "Sound Localization from Motion: Jointly Learning Sound Direction and Camera Rotation", + "text_level": 1, + "bbox": [ + 202, + 130, + 767, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziyang Chen", + "bbox": [ + 269, + 203, + 377, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shengyi Qian", + "bbox": [ + 416, + 203, + 527, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Andrew Owens", + "bbox": [ + 568, + 204, + 694, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Michigan", + "bbox": [ + 383, + 233, + 571, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 285, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The images and sounds that we perceive undergo subtle but geometrically consistent changes as we rotate our heads. In this paper, we use these cues to solve a problem we call Sound Localization from Motion (SLfM): jointly estimating camera rotation and localizing sound sources. We learn to solve these tasks solely through self-supervision. A visual model predicts camera rotation from a pair of images, while an audio model predicts the direction of sound sources from binaural sounds. We train these models to generate predictions that agree with one another. At test time, the models can be deployed independently. To obtain a feature representation that is well-suited to solving this challenging problem, we also propose a method for learning an audio-visual representation through cross-view binauralization: estimating binaural sound from one view, given images and sound from another. Our model can successfully estimate accurate rotations on both real and synthetic scenes, and localize sound sources with accuracy competitive with state-of-the-art self-supervised approaches. Project site: https://ificl.github.io/SLfM.", + "bbox": [ + 75, + 318, + 473, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 630, + 209, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As you rotate your head, the images and sounds that you perceive change in geometrically consistent ways. For example, after turning to the right, a sound source that was directly in front of you will become louder in your left ear and quieter in your right, while simultaneously the visual scene will move right-to-left across your visual field (Fig. 1).", + "bbox": [ + 75, + 657, + 470, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We hypothesize that these co-occurring audio and visual signals provide \"free\" supervision that captures geometry, including the motion made by a camera and the direction of sound sources. These are each core problems in machine perception, but are largely studied separately, often using supervised methods that rely on difficult-to-acquire labeled training data, such as annotated sound directions. We take inspiration from self-supervised approaches to structure from motion [104], which learn to estimate 3D structure and camera pose by solving both tasks simultaneously.", + "bbox": [ + 75, + 750, + 472, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db541dd8a9a951af38cf10e58cc73dfe15c211eb8e2078d40fec774a8fb1bdfc.jpg", + "image_caption": [ + "Figure 1: Images and sounds change in geometrically consistent ways. For example, when we rotate to the right, a sound source that is initially in front of us becomes louder in our left ear. We use these cues to jointly train models for two tasks: localizing sounds from binaural audio and estimating camera rotation from images. The two models are trained entirely through self-supervision, by learning to produce outputs that agree with one other." + ], + "image_footnote": [], + "bbox": [ + 501, + 282, + 890, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Analogously, we propose a problem we call sound localization from motion (SLfM): jointly estimating camera rotation from images and the sound direction from binaural audio. By solving both tasks simultaneously, we avoid the need for labeled training data. Our models provide each other with self-supervision: a visual model predicts the rotation angle between pairs of images, while an audio model predicts the azimuth of sound sources. We force their predictions to agree with one another, such that changes in rotation are consistent with changes in sound direction and binaural cues. After training, the models can be deployed independently, without multimodal data at test time.", + "bbox": [ + 496, + 625, + 893, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This is a challenging task that requires perceiving motion in images and binaural cues in audio. Our second contribution is a method for learning representations that are well-suited to this task through cross-view binauralization. We train a network to convert mono to binaural sound for one viewpoint, given an audio-visual pair sampled from another", + "bbox": [ + 496, + 809, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7897", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "viewpoint. Since the sound source is not necessarily visible in the images, the only way to successfully solve this pretext task is by analyzing the changes in the camera pose and predicting how they affect the sound direction.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "All components of our model are entirely self-supervised and are trained solely on unlabeled audio-visual data. Our results suggest that paired audio-visual data provides a useful and complementary signal for learning about geometry. In contrast to other audio or visual self-supervised pose estimation methods, we obtain supervision from abundantly available audio data, thus avoiding the need of 3D ground truth or correspondences between pixels [104, 106] or audio samples [17]. Through experiments, we show:", + "bbox": [ + 75, + 152, + 470, + 287 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Paired audio-visual data provides a supervisory signal for pose estimation tasks.", + "- We obtain competitive performance with state-of-the-art self-supervised sound localization methods [17].", + "- We obtain strong rotation estimation performance, and our model generalizes to Stanford2D3D [6] dataset, where it is competitive with classic sparse feature matching methods.", + "- The features we learn through our pretext task outperform other representations for our downstream tasks." + ], + "bbox": [ + 76, + 290, + 468, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 441, + 218, + 458 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Audio for spatial perception. Recent works have explored the use of sound for spatial understanding. Purushwalkam et al. [71] reconstructed floor plans in simulated environments [12]. Chen et al. [18] used ambient sounds from environments to learn about scene structures. Konno et al. [48] integrated sound localization to visual SfM while do not jointly learn them. Other work learns representations for spatial audio-visual tasks. Yang et al. [97] predicted whether stereo channels are swapped in a video, and Morgado et al. [63] solved a spatial alignment task. The learned representations are then used to improve localization, up-mixing, and segmentation models. In contrast, we learn camera pose and sound localization solely from self-supervision, obtaining angular predictions without labeled data. Other work uses echolocation sounds to learn representations [29, 96] and predict depth maps [19, 70] and estimate camera poses [96] using labeled data. In contrast, our proposed approach jointly learns binaural sound localization and camera pose through passive audio sensing, without supervision.", + "bbox": [ + 75, + 467, + 470, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Acoustic synthesis and spatialization. Researchers have explored visually-guided sound synthesis [28, 33, 42, 21] and text-guided audio synthesis [49, 95, 40]. Additionally, researchers have investigated generating realistic environmental acoustics using visual information [11, 84, 15, 57]. Chen et al. [14] introduced the novel-view acoustic synthesis task, which synthesizes binaural sound at the target view using audio and visual information from a source view. Liang et al. [50] proposed an audio-visual neural field in", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "real-world audio-visual scenes. Many recent works have proposed to generate spatial audio from mono audio using visual cues [64, 30, 74, 93, 51, 103, 32], or the relative pose between sound sources and the receiver [78, 41]. Inspired by these works, our feature learning approach learns spatial representations through an audio prediction task.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Binaural sound localization. Humans have the ability to localize sound sources from binaural sound [80]. Traditional approaches estimate interaural time delays via cross-correlation using hand-crafted features [47], factorization methods [82], or loudness differences between ears [77, 90]. Chen et al. [17] adapted methods from self-supervised visual tracking to the problem of binaural sound localization. Similarly, we estimate direction through self-supervision. However, we obtain our supervision through cross-modal supervision from vision instead of from correspondence cues. Moreover, we also obtain visual camera rotation estimation through our learning process. Francl [27] learned representations of sound location with a contrastive loss where positive and negative examples are selected based on the extent of head movements. Other works have used supervised learning techniques with labeled data to localize sound sources in reverberant environments [1, 89, 94]. Unlike these methods, our model learns 3D sound localization without labels.", + "bbox": [ + 496, + 189, + 893, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Camera pose estimation. Traditional methods for camera pose estimation are based on finding correspondences between images and then solving an optimization problem using constraints from multi-view geometry [36]. These include structure from motion methods that estimate full pose [83] and camera rotation [8]. Recent methods have directly predicted camera pose using neural networks, including methods that use photos [72, 45, 59, 44, 55] or RGB-D scans [99, 98, 22, 23]. Our setup is similar to work that learns relative camera poses from sparse views [43, 9], and we use their network architectures. However, we learn the camera pose through cross-modal supervision from audio, rather than from labels. Our work is also closely related to methods that learn structure from motion through self-supervision [104, 106], such as by jointly learning models that perform depth and camera pose estimation with photoconsistency constraints. In contrast, our visual model's learning signal comes solely from audio-based supervision, and we jointly learn audio localization.", + "bbox": [ + 496, + 469, + 893, + 756 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Audio-visual learning. A number of works have focused on learning multimodal representations for audio and vision, taking into account semantic correspondence and temporal synchronization [67, 92, 7, 65, 68, 2, 60]. Other approaches study audio-visual sound localization [39, 61, 16, 62], source separation [56, 31, 58, 87], active speaker detection [3, 86, 5], navigation [12, 13, 10] and forensics [105, 34, 24]. Our focus, in contrast, is on utilizing multi-view audio-visual signals to learn geometry.", + "bbox": [ + 496, + 763, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7898", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0af377614fad8cce4404f7daff0ffe16d7c754f65cb725801f3cd0a6a852a88d.jpg", + "image_caption": [ + "(a) Cross-view audio prediction" + ], + "image_footnote": [], + "bbox": [ + 76, + 75, + 620, + 224 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7aec18f353123bcdd088dfa7097260c35e09a003197f9dadb80260e5aa046005.jpg", + "image_caption": [ + "(b) Sound localization from motion", + "Figure 2: Method overview. (a) We learn a feature representation by predicting how changes in images lead to changes in sound in a cross-view binauralization pretext task. We convert mono sound to binaural sound at a target viewpoint, after conditioning the model on observations from a source viewpoint. (b) We use the representation to jointly solve two pose estimation tasks: visual rotation estimation and binaural sound localization. We train the visual rotation angle, $\\phi_{s,t}$ , to be consistent with the difference in predicted sound angles $\\theta_s$ and $\\theta_t$ ." + ], + "image_footnote": [], + "bbox": [ + 633, + 75, + 890, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Sound Localization from Motion", + "text_level": 1, + "bbox": [ + 75, + 324, + 375, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We address the sound localization from motion (SLfM) task: predicting the azimuth of a sound source from binaural audio and camera rotation from two images. First, we present a self-supervised representation that can be used to solve this downstream task. Then, we show how the representation can be used to solve the task.", + "bbox": [ + 75, + 349, + 468, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Learning representation via spatialization", + "text_level": 1, + "bbox": [ + 75, + 450, + 433, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We learn an audio-visual representation that conveys spatial information by solving a cross-view binauralization task: converting mono sound to stereo for one viewpoint, given an audio-visual pair sampled from another viewpoint (Fig. 2a). In order to successfully solve the task, a model must implicitly estimate the sound direction in the source view and predict how the change in viewpoint will affect the sound in the target view. A key difference between this task and traditional, single-view binauralization [30] is that the sound source need not be visible in any of the images. Hence, the task cannot be solved from the target audio-visual pair alone.", + "bbox": [ + 75, + 472, + 468, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We binauralize the sound $\\mathbf{a}_t$ at the target view through an audio predictor $\\mathcal{F}_{\\theta}$ . To make our representation suitable for downstream tasks, we factorize it into visual features $f_{v}(\\mathbf{v}_{s},\\mathbf{v}_{t})$ , which are intended to create features relevant to relative pose, and audio features $f_{a}(\\mathbf{a}_{s})$ , which capture sound localization cues. We predict the binauralized audio $\\hat{\\mathbf{a}}_t$ at the target view from mono audio $\\bar{\\mathbf{a}}_t$ , the visual change of $(\\mathbf{v}_s,\\mathbf{v}_t)$ and binaural sound $\\mathbf{a}_s$ heard at the source viewpoint:", + "bbox": [ + 75, + 638, + 470, + 760 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {a}} _ {t} = \\mathcal {F} _ {\\theta} \\left(\\bar {\\mathbf {a}} _ {t}, f _ {v} \\left(\\mathbf {v} _ {s}, \\mathbf {v} _ {t}\\right), f _ {a} \\left(\\mathbf {a} _ {s}\\right)\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 771, + 468, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We represent audio $\\mathbf{a}$ as a spectrogram $\\mathbf{A}$ using short-time Fourier transform (STFT). Following Gao et al. [30], given the mix of stereo audio $\\bar{\\mathbf{A}}_t = \\mathrm{STFT}(\\mathbf{a}_t^L + \\mathbf{a}_t^R)$ , we predict the difference of two channels $\\mathbf{A}_t = \\mathrm{STFT}(\\mathbf{a}_t^L - \\mathbf{a}_t^R)$ . We optimize the $L1$ loss between predicted spectrogram $\\hat{\\mathbf{A}}_t$ and ground-truth spectrogram $\\mathbf{A}_t$ :", + "bbox": [ + 75, + 794, + 468, + 886 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p r e t e x t}} = \\left| \\left| \\hat {\\mathbf {A}} _ {t} - \\mathbf {A} _ {t} \\right| \\right| _ {1}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 888, + 468, + 906 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-view binauralization. Following Zhou et al. [104], we improve our representations by binauralizing sounds at $N$ different target viewpoints, using observations from a single source viewpoint $s$ . We hypothesize that jointly solving spatialization problems from a single viewpoint for multiple target viewpoints would require the model to make more accurate predictions of sound source locations, thereby improving the estimation of view changes:", + "bbox": [ + 498, + 325, + 893, + 446 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p r e t e x t}} = \\frac {1}{N} \\sum_ {i} ^ {N} \\left| \\left| \\mathcal {F} _ {\\theta} \\left(\\bar {\\mathbf {A}} _ {i}, f _ {v} \\left(\\mathbf {v} _ {s}, \\mathbf {v} _ {i}\\right), f _ {a} \\left(\\mathbf {a} _ {s}\\right)\\right) - \\mathbf {A} _ {i} \\right| \\right| _ {1}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 460, + 890, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Estimating pose and localizing sounds", + "text_level": 1, + "bbox": [ + 500, + 518, + 826, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now address the problem of learning models for sound localization and pose estimation, using our self-supervised audio-visual features. Given two views, we predict sound directions and relative rotation. We train the model to make these two predictions consistent with one another, while using simple binaural constraints to resolve ambiguities.", + "bbox": [ + 496, + 542, + 890, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We are given images $\\mathbf{v}_s$ and $\\mathbf{v}_t$ (rotated views recorded at the same position) and learned visual embedding $f_v$ . We predict the scalar rotation angle $\\phi_{s,t}$ via the encoder $g_v$ :", + "bbox": [ + 496, + 633, + 890, + 680 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {s, t} = g _ {v} \\left(f _ {v} \\left(\\mathbf {v} _ {s}, \\mathbf {v} _ {t}\\right)\\right), R _ {s, t} = \\left[ \\begin{array}{l l} \\cos \\phi_ {s, t} & - \\sin \\phi_ {s, t} \\\\ \\sin \\phi_ {s, t} & \\cos \\phi_ {s, t} \\end{array} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 694, + 890, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $R_{s,t}$ is 2D rotation matrix of $\\phi_{s,t}$ . Following common practice in indoor scene reconstruction, we give the camera a fixed downward tilt [102, 73, 101] and only estimate azimuth [43]. This is also a common assumption in audio localization [1, 89], since azimuth has strong binaural cues.", + "bbox": [ + 496, + 739, + 893, + 815 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similarly, we predict the azimuths of the sound sources using audio features and the encoder $g_{a}$ :", + "bbox": [ + 498, + 816, + 890, + 847 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {i} = g _ {a} \\left(f _ {a} (\\mathbf {a} _ {i})\\right), \\quad \\mathbf {r} _ {i} = \\left[ \\begin{array}{l l} \\cos \\theta_ {i} & \\sin \\theta_ {i} \\end{array} \\right] ^ {\\top}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 861, + 890, + 882 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where we represent the azimuth as a vector $\\mathbf{r}_i$", + "bbox": [ + 500, + 896, + 802, + 910 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7899", + "bbox": [ + 482, + 949, + 514, + 960 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-modal geometric consistency. When the camera is rotated, the sound source ought to rotate in the opposite direction (Fig. 2b). For example, a $30^{\\circ}$ clockwise camera rotation should result in a $30^{\\circ}$ counterclockwise rotation in sound direction. Such a constraint could be converted into a loss:", + "bbox": [ + 75, + 90, + 472, + 179 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r o t}} = \\left\\| \\mathbf {r} _ {s} - R _ {s, t} \\mathbf {r} _ {t} \\right\\| ^ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 180, + 470, + 198 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, a well-known ambiguity called front-back confusion [77, 25] exists in binaural sound perception: one cannot generally tell whether a sound is in front of the view, or behind them. To address that, we use permutation invariant training [100] (PIT), and allow the model to use either the predicted sound direction or its reflection about the $x$ axis without penalty. This results in the loss:", + "bbox": [ + 75, + 204, + 473, + 309 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {g e o}} = \\min _ {\\substack {\\hat {\\mathbf {r}} _ {s} \\in \\{\\mathbf {r} _ {s}, Q \\mathbf {r} _ {s} \\} \\\\ \\hat {\\mathbf {r}} _ {t} \\in \\{\\mathbf {r} _ {t}, Q \\mathbf {r} _ {t} \\}}} \\| \\hat {\\mathbf {r}} _ {s} - R _ {s, t} \\hat {\\mathbf {r}} _ {t} \\| ^ {2}, \\tag{7}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 316, + 470, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q = \\begin{bmatrix} 1 & 0 \\\\ 0 & -1 \\end{bmatrix}$ reflects the sound direction.", + "bbox": [ + 78, + 359, + 411, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As a consequence of this ambiguity, there are also two possible solutions for the visual rotation model, since the visual rotation matrices can be mirrored about the $x$ axis. For example, one can create a solution with equal loss by multiplying the rotations and sound directions by $Q$ . We discuss this ambiguity in more depth in Sec. 4.4.", + "bbox": [ + 75, + 398, + 470, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Incorporating binaural observations. Without additional constraints, the solution is ambiguous, and may collapse into a trivial solution (e.g., predicting zero for all three angles). To avoid this, we force the model to agree with a simple binaural cue based on interaural intensity difference (IID). We predict whether the sound is to the left or right of the viewer, based on whether it is louder in the left or right microphone: $d = \\mathrm{sign}(\\log \\left|\\frac{\\mathbf{A}^L}{\\mathbf{A}^R}\\right|)$ , where $|\\mathbf{A}|$ is the magnitude of the spectrogram A. We perform this left/right test at each timestep in the spectrogram and then pool via majority voting (see supplementary for details). We penalize predictions that are inconsistent with these \"left or right\" observations:", + "bbox": [ + 75, + 494, + 472, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {b i n a u r a l}} = \\mathcal {L} _ {\\text {B C E}} (\\sin \\theta_ {i}, d _ {i}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 686, + 468, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{BCE}}$ is binary cross entropy loss.", + "bbox": [ + 75, + 710, + 349, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Encouraging symmetry. To help regularize the model, we also add symmetry constraints. For sound localization, swapping the left and right channels of the audio ought to result in a prediction in the opposite direction since the binaural cues are reversed. For rotation estimation, the relative pose between images $s$ and $t$ should invert the pose from $t$ to $s$ . We encourage both constraints via a loss:", + "bbox": [ + 75, + 731, + 472, + 838 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s y m}} = \\left| \\theta + \\theta_ {\\text {f l i p}} \\right| + \\left| \\phi_ {s, t} + \\phi_ {t, s} \\right|, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 849, + 468, + 867 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/99ddbdc91cdffb8b78ffbf33341c969bb81d918b52309982c38bbb2560a362de.jpg", + "image_caption": [ + "Figure 3: Cross-view binauralization architecture. We take a mono spectrogram as input and fuse the audio and visual features from the audio and visual encoder respectively to synthesize the binaural spectrogram at the target viewpoint." + ], + "image_footnote": [], + "bbox": [ + 501, + 74, + 893, + 231 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta_{\\mathrm{flip}}$ is the prediction of sound angle $\\theta$ using audio with swapped audio channels, and $\\phi_{s,t}$ and $\\phi_{t,s}$ are the predicted rotations between cameras $s$ and $t$ .", + "bbox": [ + 498, + 311, + 893, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall loss. We combine these constraints to obtain an overall loss:", + "bbox": [ + 498, + 369, + 890, + 398 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda \\mathcal {L} _ {\\mathrm {g e o}} + \\mathcal {L} _ {\\text {b i n a u r a l}} + \\mathcal {L} _ {\\text {s y m}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 409, + 890, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is the weight for the geometric loss.", + "bbox": [ + 503, + 440, + 799, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 470, + 633, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We have introduced a self-supervised method to learn camera pose and sound localization from audio-visual data. In experiments, we first evaluate how well our learned representation captures spatial information. We then evaluate how well our method learns camera pose and sound localization by comparing it with baselines. Finally, we show generalization to indoor panorama images Stanford2D3D [6] and in-the-wild binaural audio [17].", + "bbox": [ + 498, + 497, + 893, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Implementations", + "text_level": 1, + "bbox": [ + 500, + 632, + 666, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual pose encoder. We follow recent pose estimation work [9, 43] and build a Siamese-style visual pose network $f_{v}$ with ResNet-18 [38] as the backbone. We compute dense 4D correlation volumes between the features from the third residual layer and then encode them by convolution layers. We resize images to $320 \\times 240$ and encode a pair of images into 512-d features. We use an MLP $g_{v}$ to map visual features to 1-d logits for our SLfM models.", + "bbox": [ + 498, + 654, + 893, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Binaural audio encoder. We obtain binaural audio embeddings $f_{a}(\\cdot)$ using ResNet-18 [38] that operates on spectrograms. We covert the two-channel waveform of length $L$ to a spectrogram representation of size $256 \\times 256 \\times 4$ using short-time Fourier transform (STFT), where we keep both the magnitude and phase of spectrograms. We extract 512-d features of binaural sound with $f_{a}$ and map them to 1-d logits using an MLP $g_{a}$ .", + "bbox": [ + 498, + 789, + 893, + 911 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1Work on self-supervised SfM has similar ambiguities [104], and deals with them by adding analogous constraints, such as photometric consistency.", + "bbox": [ + 75, + 875, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7900", + "bbox": [ + 482, + 949, + 514, + 960 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Audio prediction model. We adopt the light-weighted audio-visual U-Net [30] to perform binauralization. We feed in spectrograms of size $256 \\times 256 \\times 2$ and predict the target spectrograms. We concatenate the visual pose features $f_{v}(\\cdot)$ and audio features $f_{a}(\\cdot)$ at the bottleneck of U-Net. We show the architecture of our models in Fig. 3. Please see supp. for more implementation details.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Dataset", + "text_level": 1, + "bbox": [ + 76, + 205, + 171, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since there is no public multi-view audio-visual dataset with camera poses and sound direction ground truth, we use the SoundSpaces 2.0 platform [15] to create a dataset. Our 3D scenes come from Habitat-Matterport 3D dataset (HM3D) [76], which is a large dataset of real 3D scenes. This setup allows us to have photorealistic images and high-quality spatial audio with real-world acoustics phenomenon (e.g., reverberation), as well as providing the ground-truth camera pose and sound directions that can be used for evaluation. We call this dataset HM3D-SS.", + "bbox": [ + 75, + 229, + 472, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We generate binaural Room Impulse Responses (RIRs) and images with a $60^{\\circ}$ field of view, using 100 scenes of HM3D [76]. For each audio-visual example, we randomly place sound sources in the scene with a height range of (0.7, 1.7) meters, and sample 4 different rotated viewpoints at one location within 4 meters. The rotations are limited to $(10^{\\circ}, 90^{\\circ})$ relative to the source viewpoints. We follow the standard practice to set the height to agents to be $1.5\\mathrm{m}$ and lock a downward tilt angle [43, 101, 73, 102]. We render the binaural RIRs and images given the position of agents and sound sources. We obtain binaural audio by convolving binaural RIRs with mono audio samples from LibriSpeech [69] and Free Music Archive [20]. To ensure that the evaluation tests the model's pose estimation abilities, rather than its ability to visually localize sound sources, the sound sources are not visible on screen.", + "bbox": [ + 75, + 380, + 472, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We create 50K audio-visual pairs from 200K viewpoints. The audio was rendered with average reverberation of $\\mathrm{RT}_{60} = 0.4\\mathrm{s}$ (see supp. for details). We divided our data into 81/9/10 scenes for the train/val/test, respectively.", + "bbox": [ + 75, + 621, + 470, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Evaluating the learned representation", + "text_level": 1, + "bbox": [ + 76, + 690, + 403, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we directly evaluate the quality of our learned features for rotation estimation and sound localization via linear probing with labeled data (rather than learning them jointly through self-supervision).", + "bbox": [ + 75, + 713, + 470, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines and ablations. We compare our model with several baselines that use alternative pretext tasks: 1) AVSA [63]: it learns spatial cues by training a model to spatially align video and audio clips extracted from different viewing angles. We adapt this model to our dataset and train with 4 different views; 2) RotNCE [27]: it applies contrastive learning on the audio from different angles and uses annotations of the agent's rotation to select positive", + "bbox": [ + 75, + 779, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and negative samples, which results in learning audio spatial representation. For baselines, we use the same architecture for feature extractors to ensure fair comparisons.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To determine if we utilize visual and audio features from different views to solve the binauralization task, we also study some variants of our models: 1) Ours-NoA: we only provide visual features for the binauralization task; 2) Ours-NoV: which only uses audio from the other view to spatialize sounds; 3) Ours-GTRot: we provide ground-truth rotation embedding instead of features from visual frames.", + "bbox": [ + 496, + 136, + 893, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Besides the mono-to-binaural task, we also experiment with another objective: predicting the right channel from the left channel. We train our L2R model with the same setup as our M2B model. (Please see supp. for pretext results.)", + "bbox": [ + 496, + 242, + 893, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Downstream tasks. We assess the quality of spatial representations we learned from our pretext tasks in two downstream tasks: relative camera rotation and 3D sound localization. We formulate them as classification problems, where angles are categorized into 64 bins, and we use accuracy as the evaluation metric. To evaluate the learned features, we freeze them and train a linear classifier on the downstream tasks. We compare the performance of our features with those learned from RotNCE [27], AVSA [63], ImageNet [37], and random features, and report the results in Tab. 1. Our approach outperforms the baselines in both tasks, indicating that we learn better spatial representations. Furthermore, our linear probe models show comparable performance against the supervised method which can be regarded as approximate upper bounds for our models, suggesting that our pretext tasks help learn a useful representation. Please see supp. for experiments on FreeMusic [20].", + "bbox": [ + 496, + 308, + 895, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Emerging camera pose from audio prompting. To help understand the strong performance of our self-supervised features, we asked whether we could use the cross-view binauralization model alone to estimate camera rotation. Inspired by prompting in vision and language models [75], we", + "bbox": [ + 496, + 571, + 893, + 648 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d6e46f3615fe29d8bbb9f3520c201da0e9d102552ff848042d2ff86913d3df5b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelAudio Loc. Acc (%) ↑Camera Rot. Acc (%) ↑
LibriSpeechRandom feature4.84.7
ImageNet [38]+Random-56.3
RotNCE [27]50.9-
AVSA [63]71.26.5
Ours-NoA-9.6
Ours-NoV53.9-
Ours-GTRot70.6-
Ours-L2R (3 views)78.576.1
Ours (2 views)74.580.0
Ours (3 views)75.481.3
Supervised81.595.8
", + "bbox": [ + 514, + 656, + 879, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: Downstream task performance on HM3D-SS dataset. We report linear probe performance on the audio localization and camera rotation downstream tasks.", + "bbox": [ + 496, + 863, + 893, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7901", + "bbox": [ + 482, + 949, + 513, + 960 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c934e608082ec2c72295e625821dd146fc1966214a19ef3b6ba7d33653e35bdf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelAudio angle MAE (°) ↓Camera angle MAE (°) ↓
LibriSpeechChance40.2829.41
SIFT [54]-12.2
Ours w/o Reflect28.0826.80
Ours-Prompt-9.13
Ours-L2R3.220.80
Ours3.170.77
Ours-Front4.483.08
Ours-GTRot1.83-
Superglue [81]-2.47
Supervised1.710.46
FreeMusicChance40.8129.41
SIFT [54]-12.2
Ours w/o Reflect28.2426.81
Ours-L2R3.180.81
Ours3.370.84
Ours-Front3.992.49
Ours-GTRot1.96-
Superglue [81]-2.47
Supervised2.500.46
", + "bbox": [ + 101, + 70, + 447, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "obtain rough estimates of camera rotation by providing our model with carefully-provided inputs. Given a pair of images $(\\mathbf{v}_s,\\mathbf{v}_t)$ , we create a synthetic binaural audio prompt, $\\mathbf{a}_s$ . We then ask our model to generate the binaural sound $\\hat{\\mathbf{a}}_t$ for the target viewpoint. By analyzing the IID cues in $\\hat{\\mathbf{a}}_t$ , we can estimate the model's implicitly predicted camera pose. To do this, we find the nearest neighbor of our generated audio $\\hat{\\mathbf{a}}_t$ , using a database of synthetically generated audio with known sound directions. Please refer to supp. for more details. Our method achieves the mean absolute error of $9.13^{\\circ}$ on HM3D-SS dataset where chance is $29.41^{\\circ}$ (Tab. 2). Our approach can also generalize to Stanford2D3D [6] dataset, where we can achieve mean absolute error of $9.93^{\\circ}$ (Tab. 3).", + "bbox": [ + 75, + 470, + 472, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Evaluating SLfM", + "text_level": 1, + "bbox": [ + 76, + 676, + 246, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluated our SLfM models on the HM3D-SS dataset. We use the mean absolute error of angle in degrees as evaluation metrics. To avoid sound field ambiguity, we filtered out samples with sound angles outside of $(-90^{\\circ}, 90^{\\circ})$ for the", + "bbox": [ + 75, + 699, + 470, + 762 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/faa58b777188becf848cf49185abfe74a4bf1acc96383789f8f4efdf03ee17b0.jpg", + "image_caption": [ + "Figure 4: Qualitative results on real-world examples. We show our predictions on Stanford2D3D [6] and In-the-wild audio [17]. Green denotes accurate predictions." + ], + "image_footnote": [], + "bbox": [ + 78, + 768, + 305, + 858 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c88cdb70d7fb5ce1c5690c0479f41d2bff5c24a5534a56c9466f58e8fbe82010.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 770, + 467, + 858 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7da04a9e23a66d10293de4127afc3d375781214f80fda1a2501969cf0e8fcd89.jpg", + "table_caption": [ + "Table 2: Sound localization from motion results on HM3D-SS. We evaluate our SLfM models on each modality independently." + ], + "table_footnote": [], + "table_body": "
In the wild audio [17]
ModelAcc (%) ↑
GCC-PHAT [47]77.2
IID [17]75.4
IID (Ours)82.1
MonoCLR [17]87.4
StereoCRW [17]87.2
Ours-L2R84.9
Ours84.0
", + "bbox": [ + 504, + 71, + 689, + 205 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2863461e8b54c4039edde554fa0bd998d3f3a0236008f720fc07aeeb0d89e4f1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Stanford2D3D [6]
ModelRot Err. (°) ↓
MeanMed.
SIFT [54]16.40.06
LoFTR [85]6.101.13
Superglue [81]5.070.07
Ours-Prompt9.939.02
Ours-L2R1.120.71
Ours1.140.67
", + "bbox": [ + 707, + 74, + 893, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3: Evaluation of our SLfM models on the real-world data. We evaluate our audio localization model on the In-the-wild audio [17] (left) and our camera pose model on the Stanford2D3D [6] (right). Rot. denote camera rotation.", + "bbox": [ + 496, + 210, + 893, + 267 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "evaluation set. When training our SLfM models, we use our best-performing features, i.e., pretext tasks trained with 3 views. We freeze learned audio and visual features and only train multi-layer perceptrons on top of them and evaluate them independently on each modality.", + "bbox": [ + 496, + 284, + 893, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines and ablations. For relative camera pose estimation, we compare our models with sparse feature matching using SIFT [54] and SuperGlue [81], followed by rotation fitting. We use those methods to detect key points, run Lowe's ratio test, and use RANSAC with five-point algorithm to recover the camera pose [54, 36, 66, 26]. For the sound localization task, we compared with time delay estimation methods: the popular GCC-PHAT [47] and the recent self-supervised method StereoCRW [17]. As far as we know, there are no other baselines that can estimate poses and sound source directions without labels.", + "bbox": [ + 496, + 364, + 893, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also compared several variations of our method, including Ours-Front, where we filter out the samples with sound sources behind the viewers to remove binaural ambiguity during training, and an oracle model Ours-GTRot, which uses ground-truth rotation angles instead.", + "bbox": [ + 496, + 531, + 893, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. We show our results in Tab. 2. Our models can predict the azimuths of sound sources, obtaining strong performance without any labels. Without providing reflection invariance (Eq. (7)), the model failed to learn reasonable geometric due to the audio ambiguity. Our self-supervised", + "bbox": [ + 496, + 612, + 893, + 688 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fb78b25bd41a287098b8e407f3c1ae28ef900d53731b9cde3e9364f7012ddef6.jpg", + "image_caption": [ + "Figure 5: Real video. We play music using a speaker and record a video using iPhone with a binaural microphone (different from training examples). We show the predicted camera rotation and sound direction over time. They change smoothly and match our camera motion. Please see our project webpage for video results." + ], + "image_footnote": [], + "bbox": [ + 501, + 695, + 890, + 830 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7902", + "bbox": [ + 482, + 949, + 514, + 960 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ba0393c760b52f00cb2e09a8a47564c6a1937c782a66b1523372d929855d79a7.jpg", + "image_caption": [ + "Figure 6: Qualitative results on HM3D-SS. Given two images and their corresponding audio at the source viewpoints, our approach can individually predict the relative camera pose and localize sound locations accurately. We visualize our camera rotation prediction with blue bar and our sound angle prediction with orange bar. To highlight the subtle differences in the waveforms, we color code the amplitude. Please see the project webpage for more video results." + ], + "image_footnote": [], + "bbox": [ + 80, + 71, + 890, + 281 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "model achieves comparable performance against our oracle model (Ours-GTRot) and supervised models, indicating we estimate camera poses and localize sound accurately. We show some qualitative results on HM3D-SS in Fig. 6 with LibriSpeech samples [69].", + "bbox": [ + 75, + 362, + 472, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalization to other datasets. We further demonstrate the generalization ability of our models by experimenting with out-of-distribution, real-world data. We evaluate our camera rotation model on Stanford2D3D [6] with real indoor RGB images (Tab. 3). We obtain image pairs by cropping from panoramas. Although our model is trained on renderings of HM3D [76], it obtains strong generalization ability. Compared with rotation estimation based on SIFT [54] and SuperGlue [81], our model is significantly better on mean rotation error. We also report median rotation error to be consistent with prior works [9, 43]. However, SIFT [54] and Superglue [81] have a very low median error. This is likely due to the all-or-nothing nature of feature matching-based approaches, which either produce highly accurate predictions if the matches are correct (especially for \"easy\" cases with small amounts of rotation) or else produce gross errors.", + "bbox": [ + 75, + 446, + 472, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also evaluate our sound localization model on the in-the-wild binaural audio [17]. We use binary accuracy as the metric for left-or-right direction classification accounting for the fact that microphone baselines are unknown for internet videos. For an apples-to-apples comparison to prior work, we retrained our model using 0.51s length of audio for the binaural audio encoder $f_{a}(\\cdot)$ . As shown in Tab. 3, our model obtains similar performance to StereoCRW, a state-of-the-art self-supervised time delay method, suggesting we have a strong capability for sound localization. We show qualitative results in Fig. 4. We also perform both tasks on a self-recorded video (Fig. 5) of a rotating camera and a binaural microphone. We show the mean and standard deviation of predictions in 1.0s windows.", + "bbox": [ + 75, + 688, + 473, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Handling ambiguity. In binaural sound perception, there is a fundamental ambiguity that whether the sound is in front of or behind us. It leads to multiple solutions with equal loss in our model, mirroring sound sources and negating rotation angles with flipped $z$ axis (Fig. 7). These two solutions differ in that a visual rotation angle either indicates a clockwise or counterclockwise rotation. For evaluation, we convert \"backwards\" counterclockwise predictions to clockwise predictions by simply providing the model with pairs of input frames, then negating the model's outputs if the angle is the opposite of the expected direction2.", + "bbox": [ + 496, + 362, + 893, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Generalization to more complex scenarios", + "text_level": 1, + "bbox": [ + 498, + 537, + 857, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We investigate whether our approach generalizes to more complex scenarios, such as with multiple sound sources or when translation is included in camera motion.", + "bbox": [ + 496, + 561, + 893, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multiple sound sources. We evaluate our representations and SLfM models on more complex scenes containing multiple sound sources. We train our models with two source sources placed in the scenes. One of them is dominant and", + "bbox": [ + 496, + 611, + 893, + 672 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a369f32c41ea1be4885ae2b44e5496a654f5b3c613ab97b8dcdccd68b3d49923.jpg", + "image_caption": [ + "Figure 7: Mirror ambiguity. The $+z$ and $+x$ axes represent forward and rightward directions. Both solutions have the same loss." + ], + "image_footnote": [], + "bbox": [ + 501, + 714, + 679, + 830 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/528096eb7f258d146a61014842594bc58c8971bfb3ba6e990a3b0ea38b9a50bb.jpg", + "image_caption": [ + "Figure 8: Robustness to reverberation. We study the effect of reverberation on our SLfM model." + ], + "image_footnote": [], + "bbox": [ + 712, + 715, + 890, + 830 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2Similar to ambiguities in SfM where reconstructions can be reoriented such that the sky is in the positive $y$ direction [36, 35, 4].", + "bbox": [ + 500, + 680, + 890, + 705 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7903", + "bbox": [ + 482, + 949, + 514, + 960 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ff2232dbec0f6fab7c95c7ddd56807ceb05f052012826069cfafe7378ccdb183.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMulti-source Acc (%) ↑Small trans. Acc (%) ↑
Aud.Rot.Aud.Rot.
Random feature4.54.75.73.8
ImageNet [38]+Random-56.3-23.0
RotNCE [27]35.8---
AVSA [63]55.96.765.46.1
Ours (3 views)59.277.972.249.6
Supervised74.595.881.663.5
", + "bbox": [ + 101, + 73, + 444, + 213 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e037ac7b412f889c94cbf4173cd5757c5290495344d34f8b3afcc9934e01dfe5.jpg", + "table_caption": [ + "Table 4: Generalization to more complex scenarios. We evaluate the audio and visual features learned from more complex scenarios via linear probing. Aud. and Rot. denote audio localization and camera rotation respectively." + ], + "table_footnote": [], + "table_body": "
ModelMulti-source MAE (°) ↓Small trans. MAE (°) ↓
Aud.Rot.Aud.Rot.
Chance40.4629.4140.2829.41
SIFT [54]-12.2-12.2
Ours7.670.716.281.04
Ours-GTRot5.81-4.24-
Superglue [81]-2.47-2.47
Supervised3.600.461.710.46
", + "bbox": [ + 116, + 280, + 431, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "our sound localization target. We report the results in Tab. 4 and Tab. 5. Our model learns better representations than baselines and achieves accurate predictions of the azimuth of the source and camera pose even in challenging scenarios with multiple sound sources.", + "bbox": [ + 75, + 497, + 468, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Translation in camera motions. To study how small translations in the camera motion could affect our models, we generate 50K pairs of audio-visual data with both rotation and translation change, limiting the uniformly sampled translation to 0.5 meters. We train our models with LibriSpeech samples [69]. For our linear probing evaluation, we measure the ability of features to handle complex examples by testing on the dataset that has translation. For our SLfM model, we study our ability to learn from noisy data, thus we evaluate it on the rotation-only examples. As the results are shown in Tab. 4 and Tab. 5, we successfully learn useful features and obtain accurate rotation and sound direction predictions despite the presence of translation. Since we jointly learn the audio and visual representations, it can negatively impact the learning of one modality when another one becomes harder.", + "bbox": [ + 75, + 580, + 470, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 818, + 225, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness to reverberation. We study how our SLfM models perform under the different reverberation configurations. We used SoundSpaces [15] to create audio with average reverberation $\\mathrm{RT}_{60} \\in \\{0.1, 0.2, 0.3, 0.4, 0.5\\}$ while", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dd5728e202d45ea862879aa48f746fff14199316565d1aefb2644cdc40395a98.jpg", + "table_caption": [ + "Table 5: SLfM results on more complex scenarios. We evaluate our model on the version of HM3D-SS with two sound sources. We also evaluate our model trained with small translations on rotation-only examples." + ], + "table_footnote": [], + "table_body": "
ModelLossesAudio angle MAE (°) ↓Camera angle MAE (°) ↓
LgeoLbinauralLsym
Ours✓37.6029.20
✓✓37.5229.17
✓✓3.586.99
✓✓✓3.170.77
", + "bbox": [ + 504, + 73, + 887, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6: Ablation experiments on our SLfM losses. We evaluate our SLfM models with different combinations of losses.", + "bbox": [ + 498, + 172, + 890, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "keeping the visual signals the same. We train our model on each setting. As shown in Fig. 8, our performance decreases as the level of reverberation increases, where audio becomes more challenging during both training and testing. Please see supp. for the study on representations.", + "bbox": [ + 496, + 210, + 890, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Losses for SLfM. We study the necessity of our proposed loss functions in Tab. 6. Our models fail to learn accurate pose estimation without binaural loss or symmetric loss. It highlights the crucial role of these losses.", + "bbox": [ + 496, + 292, + 890, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 366, + 617, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed the sound localization from motion (SLfM) problem, and provided a self-supervised method for solving it. We also presented a method for learning audio-visual features that convey sound directions and camera rotation, which we show are well-suited to solving the SLfM task. Despite learning our models solely from unlabeled audio-visual data, we obtain strong performance on a variety of benchmarks, including rotation estimation on the Stanford2D3D [6] dataset and \"in the wild\" sound direction estimation [17]. Our results suggest that the subtle correlations between sights and binaural sounds that result from rotational motion provide a useful (and previously unused) learning signal. We see our work as opening new directions in self-supervised geometry estimation and feature learning that use sound as a complementary source of supervision. We will release code, data, and models upon acceptance.", + "bbox": [ + 496, + 391, + 893, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Broader Impacts. Our work has several limitations. First, while we evaluate our models on real images and sounds, we train on data from simulators, due to a lack of available relevant data. We note that this is common practice in visual 3D reconstruction [43, 91, 52]. Second, we assume that the 3D scene and sound sources are stationary. Third, we do not evaluate our model on extreme viewpoint changes [55], which requires reasoning about images that have little or no overlap.", + "bbox": [ + 496, + 638, + 893, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. We would like to thank Ang Cao, Tiange Luo, and Linyi Jin for their helpful discussion. We thank Changan Chen for his help with SoundSpaces 2.0. This work was funded in part by DARPA Semafor and Sony. The views, opinions and/or findings expressed are those of the authors and should not be interpreted as representing the official views or policies of the Department of Defense or the U.S. Government.", + "bbox": [ + 496, + 779, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7904", + "bbox": [ + 482, + 949, + 514, + 960 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sharath Adavanne, Archontis Politis, and Tuomas Virtanen. Direction of arrival estimation for multiple sound sources using convolutional recurrent neural network. In 2018 26th European Signal Processing Conference (EUSIPCO), pages 1462-1466. IEEE, 2018. 2, 3", + "[2] Triantafyllos Afouras, Yuki M Asano, Francois Fagan, Andrea Vedaldi, and Florian Metze. Self-supervised object detection from audio-visual correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10575–10586, 2022. 2", + "[3] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. arXiv preprint arXiv:2008.04237, 2020. 2", + "[4] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54(10):105-112, 2011. 7", + "[5] Juan León Alcazar, Fabian Caba, Ali K Thabet, and Bernard Ghanem. Maas: Multi-modal assignation for active speaker detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 265-274, 2021. 2", + "[6] Iro Armeni, Sasha Sax, Amir R Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105, 2017. 2, 4, 6, 7, 8", + "[7] Yuki Asano, Mandela Patrick, Christian Rupprecht, and Andrea Vedaldi. Labelling unlabelled videos from scratch with multi-modal self-supervision. Advances in Neural Information Processing Systems, 33:4660-4671, 2020. 2", + "[8] Matthew Brown and David G Lowe. Automatic panoramic image stitching using invariant features. International journal of computer vision, 74:59-73, 2007. 2", + "[9] Ruojin Cai, Bharath Hariharan, Noah Snavely, and Hadar Averbuch-Elor. Extreme rotation estimation using dense correlation volumes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14566-14575, 2021. 2, 4, 7", + "[10] Changan Chen, Ziad Al-Halah, and Kristen Grauman. Semantic audio-visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15516–15525, 2021. 2", + "[11] Changan Chen, Ruohan Gao, Paul Calamia, and Kristen Grauman. Visual acoustic matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18858-18868, 2022. 2", + "[12] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part VI 16, pages 17–36. Springer, 2020. 2", + "[13] Changan Chen, Sagnik Majumder, Ziad Al-Halah, Ruohan Gao, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Learning to set waypoints for audio-visual navigation. arXiv preprint arXiv:2008.09622, 2020. 2" + ], + "bbox": [ + 86, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Changan Chen, Alexander Richard, Roman Shapovalov, Vamsi Krishna Ithapu, Natalia Neverova, Kristen Grauman, and Andrea Vedaldi. Novel-view acoustic synthesis. arXiv preprint arXiv:2301.08730, 2023. 2", + "[15] Changan Chen, Carl Schissler, Sanchit Garg, Philip Kobernik, Alexander Clegg, Paul Calamia, Dhruv Batra, Philip W Robinson, and Kristen Grauman. Soundspaces 2.0: A simulation platform for visual-acoustic learning. arXiv preprint arXiv:2206.08312, 2022. 2, 5, 8, 1", + "[16] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16867-16876, 2021. 2", + "[17] Ziyang Chen, David F Fouhey, and Andrew Owens. Sound localization by self-supervised time delay estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVI, pages 489-508. Springer, 2022. 2, 4, 6, 7, 8, 1", + "[18] Ziyang Chen, Xixi Hu, and Andrew Owens. Structure from silence: Learning scene structure from ambient sound. arXiv preprint arXiv:2111.05846, 2021. 2", + "[19] Jesper Haahr Christensen, Sascha Hornauer, and Stella Yu. Batvision with gcc-phat features for better sound to vision predictions. arXiv preprint arXiv:2006.07995, 2020. 2", + "[20] Michael Defferrard, Kirell Benzi, Pierre Vandergheynst, and Xavier Bresson. Fma: A dataset for music analysis. arXiv preprint arXiv:1612.01840, 2016. 5, 1", + "[21] Yuexi Du, Ziyang Chen, Justin Salamon, Bryan Russell, and Andrew Owens. Conditional generation of audio from video via foley analogies. Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[22] Mohamed El Banani, Luya Gao, and Justin Johnson. Unsupervisedr&r: Unsupervised point cloud registration via differentiable rendering. In CVPR, 2021. 2", + "[23] Mohamed El Banani and Justin Johnson. Bootstrap your own correspondences. In ICCV, 2021. 2", + "[24] Chao Feng, Ziyang Chen, and Andrew Owens. Self-supervised video forensics by audio-visual anomaly detection. arXiv preprint arXiv:2301.01767, 2023. 2", + "[25] Tim Fischer, Marco Caversaccio, and Wilhelm Wimmer. A front-back confusion metric in horizontal sound localization: The fbc score. In ACM Symposium on Applied Perception 2020, pages 1-5, 2020. 4", + "[26] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981. 6, 1", + "[27] Andrew Francl. Modeling and Evaluating Human Sound Localization in the Natural Environment. PhD thesis, Massachusetts Institute of Technology, 2022. 2, 5, 8, 1", + "[28] Chuang Gan, Deng Huang, Peihao Chen, Joshua B Tenenbaum, and Antonio Torralba. Foley music: Learning to generate music from videos. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 758-775. Springer, 2020. 2" + ], + "bbox": [ + 509, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7905", + "bbox": [ + 482, + 949, + 514, + 960 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Ruohan Gao, Changan Chen, Ziad Al-Halah, Carl Schissler, and Kristen Grauman. Visualechoes: Spatial image representation learning through echolocation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16, pages 658-676. Springer, 2020. 2", + "[30] Ruohan Gao and Kristen Grauman. 2.5 d visual sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 324-333, 2019. 2, 3, 5", + "[31] Ruohan Gao and Kristen Grauman. Visualvoice: Audiovisual speech separation with cross-modal consistency. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15490-15500. IEEE, 2021. 2", + "[32] Rishabh Garg, Ruohan Gao, and Kristen Grauman. Geometry-aware multi-task learning for binaural audio generation from video. arXiv preprint arXiv:2111.10882, 2021. 2", + "[33] Sanchita Ghose and John Jeffrey Prevost. Auto Foley: Artificial synthesis of synchronized sound tracks for silent videos with deep learning. IEEE Transactions on Multimedia, 23:1895-1907, 2020. 2", + "[34] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14950-14962, 2022. 2", + "[35] R. I. Hartley. Chirality. International Journal of Computer Vision (IJCV), 1998. 7", + "[36] R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. 2, 6, 7", + "[37] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In International Conference on Computer Vision (ICCV), 2015. 5", + "[38] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition (CVPR), 2016. 4, 5, 8, 1", + "[39] Xixi Hu, Ziyang Chen, and Andrew Owens. Mix and localize: Localizing sound sources in mixtures. Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[40] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. arXiv preprint arXiv:2301.12661, 2023. 2", + "[41] Wen Chin Huang, Dejan Markovic, Alexander Richard, Israel Dejene Gebru, and Anjali Menon. End-to-end binaural speech synthesis. arXiv preprint arXiv:2207.03697, 2022. 2", + "[42] Vladimir Iashin and Esa Rahtu. Taming visually guided sound generation. arXiv preprint arXiv:2110.08791, 2021. 2", + "[43] Linyi Jin, Shengyi Qian, Andrew Owens, and David F Fouhey. Planar surface reconstruction from sparse views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12991-13000, 2021. 2, 3, 4, 5, 7, 8" + ], + "bbox": [ + 86, + 90, + 472, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Linyi Jin, Jianming Zhang, Yannick Hold-Geoffroy, Oliver Wang, Kevin Matzen, Matthew Sticha, and David F Fouhey. Perspective fields for single image camera calibration. arXiv, 2022. 2", + "[45] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In ICCV, 2015. 2", + "[46] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representation, 2015. 2", + "[47] Charles Knapp and Grifford Carter. The generalized correlation method for estimation of time delay. IEEE transactions on acoustics, speech, and signal processing, 24(4):320-327, 1976. 2, 6", + "[48] Takashi Konno, Kenji Nishida, Katsutoshi Itoyama, and Kazuhiro Nakadai. Audio-visual sfm towards 4d reconstruction under dynamic scenes. 2022. 2", + "[49] Felix Kreuk, Gabriel Synnaeve, Adam Polyak, Uriel Singer, Alexandre Défossez, Jade Copet, Devi Parikh, Yaniv Taigman, and Yossi Adi. Audiogen: Textually guided audio generation. arXiv preprint arXiv:2209.15352, 2022. 2", + "[50] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Av-nerf: Learning neural fields for real-world audio-visual scene synthesis. arXiv preprint arXiv:2302.02088, 2023. 2", + "[51] Yan-Bo Lin and Yu-Chiang Frank Wang. Exploiting audiovisual consistency with partial supervision for spatial audio generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2056–2063, 2021. 2", + "[52] Zhi-Hao Lin, Wei-Chiu Ma, Hao-Yu Hsu, Yu-Chiang Frank Wang, and Shenlong Wang. Neurmips: Neural mixture of planar experts for view synthesis. In CVPR, 2022. 8", + "[53] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 2", + "[54] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 6, 7, 8", + "[55] Wei-Chiu Ma, Anqi Joyce Yang, Shenlong Wang, Raquel Urtasun, and Antonio Torralba. Virtual correspondence: Humans as a cue for extreme-view geometry. In CVPR, 2022. 2, 8", + "[56] Sagnik Majumder, Ziad Al-Halah, and Kristen Grauman. Move2hear: Active audio-visual source separation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 275–285, 2021. 2", + "[57] Sagnik Majumder, Changan Chen, Ziad Al-Halah, and Kristen Grauman. Few-shot audio-visual learning of environment acoustics. arXiv preprint arXiv:2206.04006, 2022. 2", + "[58] Sagnik Majumder and Kristen Grauman. Active audio-visual separation of dynamic sound sources. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXIX, pages 551–569. Springer, 2022. 2", + "[59] Iaroslav Melekhov, Juha Ylioinas, Juho Kannala, and Esa Rahtu. Relative camera pose estimation using convolutional" + ], + "bbox": [ + 509, + 90, + 893, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7906", + "bbox": [ + 482, + 949, + 516, + 960 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "neural networks. In Advanced Concepts for Intelligent Vision Systems: 18th International Conference, ACIVS 2017, Antwerp, Belgium, September 18-21, 2017, Proceedings 18, pages 675-687. Springer, 2017. 2", + "[60] Himangi Mittal, Pedro Morgado, Unnat Jain, and Abhinav Gupta. Learning state-aware visual representations from audible interactions. arXiv preprint arXiv:2209.13583, 2022. 2", + "[61] Shentong Mo and Pedro Morgado. A closer look at weakly-supervised audio-visual source localization. arXiv preprint arXiv:2209.09634, 2022. 2", + "[62] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 218-234. Springer, 2022. 2", + "[63] Pedro Morgado, Yi Li, and Nuno Nvasconcelos. Learning representations from audio-visual spatial alignment. Advances in Neural Information Processing Systems, 33:4733-4744, 2020. 2, 5, 8, 1", + "[64] Pedro Morgado, Nuno Vasconcelos, Timothy Langlois, and Oliver Wang. Self-supervised generation of spatial audio for 360 video. arXiv preprint arXiv:1809.02587, 2018. 2", + "[65] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12475-12486, 2021. 2", + "[66] David Nister. An efficient solution to the five-point relative pose problem. IEEE transactions on pattern analysis and machine intelligence, 26(6):756-770, 2004. 6", + "[67] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. European Conference on Computer Vision (ECCV), 2018. 2", + "[68] Andrew Owens, Phillip Isola, Josh McDermott, Antonio Torralba, Edward H Adelson, and William T Freeman. Visually indicated sounds. In Computer Vision and Pattern Recognition (CVPR), 2016. 2", + "[69] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE, 2015. 5, 7, 8, 1, 2", + "[70] Kranti Kumar Parida, Siddharth Srivastava, and Gaurav Sharma. Beyond image to depth: Improving depth prediction using echoes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8268-8277, 2021. 2", + "[71] Senthil Purushwalkam, Sebastia Vicenc Amengual Gari, Vamsi Krishna Ithapu, Carl Schissler, Philip Robinson, Abhinav Gupta, and Kristen Grauman. Audio-visual floorplan reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1183–1192, 2021. 2", + "[72] Shengyi Qian, Linyi Jin, and David F Fouhey. Associative3d: Volumetric reconstruction from sparse views. In ECCV, 2020. 2" + ], + "bbox": [ + 86, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[73] Shengyi Qian, Alexander Kirillov, Nikhila Ravi, Devendra Singh Chaplot, Justin Johnson, David F Fouhey, and Georgia Gkioxari. Recognizing scenes from novel viewpoints. arXiv preprint arXiv:2112.01520, 2021. 3, 5", + "[74] Kranthi Kumar Rachavarapu, Vignesh Sundaresha, AN Rajagopalan, et al. Localize to binauralize: Audio spatialization from visual sound source localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1930-1939, 2021. 2", + "[75] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[76] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 5, 7, 2", + "[77] Lord Rayleigh. Xii. on our perception of sound direction. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 13(74):214-232, 1907. 2, 4", + "[78] Alexander Richard, Dejan Markovic, Israel D Gebru, Steven Krenn, Gladstone Alexander Butler, Fernando Torre, and Yaser Sheikh. Neural synthesis of binaural speech from mono audio. In International Conference on Learning Representations, 2021. 2", + "[79] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2", + "[80] Lord Rayleigh O.M. Pres. R.S. Xii. on our perception of sound direction. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 1907. 2", + "[81] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4938–4947, 2020. 6, 7, 8", + "[82] Ralph Schmidt. Multiple emitter location and signal parameter estimation. IEEE transactions on antennas and propagation, 34(3):276-280, 1986. 2", + "[83] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2", + "[84] Nikhil Singh, Jeff Mentch, Jerry Ng, Matthew Beveridge, and Iddo Drori. Image2reverb: Cross-modal reverb impulse response synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 286-295, 2021. 2" + ], + "bbox": [ + 509, + 90, + 893, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7907", + "bbox": [ + 482, + 949, + 514, + 960 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[85] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 6", + "[86] Ruijie Tao, Zexu Pan, Rohan Kumar Das, Xinyuan Qian, Mike Zheng Shou, and Haizhou Li. Is someone speaking? exploring long-term temporal features for audio-visual active speaker detection. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3927–3935, 2021. 2", + "[87] Efthymios Tzinis, Scott Wisdom, Tal Remez, and John R Hershey. Audioscopev2: Audio-visual attention architectures for calibrated open-domain on-screen sound separation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXVII, pages 368–385. Springer, 2022. 2", + "[88] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 2", + "[89] Paolo Vecchiotti, Ning Ma, Stefano Squartini, and Guy J Brown. End-to-end binaural sound localisation from the raw waveform. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 451-455. IEEE, 2019. 2, 3", + "[90] DeLiang Wang and Guy J Brown. Computational auditory scene analysis: Principles, algorithms, and applications. Wiley-IEEE press, 2006. 2", + "[91] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. In ECCV, 2022. 8", + "[92] Fanyi Xiao, Yong Jae Lee, Kristen Grauman, Jitendra Malik, and Christoph Feichtenhofer. Audiovisual slowfast networks for video recognition. arXiv preprint arXiv:2001.08740, 2020. 2", + "[93] Xudong Xu, Hang Zhou, Ziwei Liu, Bo Dai, Xiaogang Wang, and Dahua Lin. Visually informed binaural audio generation without binaural audios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15485-15494, 2021. 2", + "[94] Nelson Yalta, Kazuhiro Nakadai, and Tetsuya Ogata. Sound source localization using deep learning models. Journal of Robotics and Mechatronics, 29(1):37-48, 2017. 2", + "[95] Dongchao Yang, Jianwei Yu, Helin Wang, Wen Wang, Chao Weng, Yuexian Zou, and Dong Yu. Diffsound: Discrete diffusion model for text-to-sound generation. arXiv preprint arXiv:2207.09983, 2022. 2", + "[96] Karren Yang, Michael Firman, Eric Brachmann, and Clément Godard. Camera pose estimation and localization with active audio sensing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 271-291. Springer, 2022. 2", + "[97] Karren Yang, Bryan Russell, and Justin Salamon. Telling left from right: Learning spatial correspondence of sight" + ], + "bbox": [ + 86, + 90, + 472, + 902 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9932-9941, 2020. 2", + "[98] Zhenpei Yang, Jeffrey Z Pan, Linjie Luo, Xiaowei Zhou, Kristen Grauman, and Qixing Huang. Extreme relative pose estimation for rgb-d scans via scene completion. In CVPR, 2019. 2", + "[99] Zhenpei Yang, Siming Yan, and Qixing Huang. Extreme relative pose network under hybrid representations. In CVPR, 2020. 2", + "[100] Dong Yu, Morten Kolbæk, Zheng-Hua Tan, and Jesper Jensen. Permutation invariant training of deep models for speaker-independent multi-talker speech separation. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 241-245. IEEE, 2017. 4", + "[101] Yinda Zhang, Shuran Song, Ersin Yumer, Manolis Savva, Joon-Young Lee, Hailin Jin, and Thomas Funkhouser. Physically-based rendering for indoor scene understanding using convolutional neural networks. In CVPR, 2017. 3, 5", + "[102] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In ICCV, 2021. 3, 5", + "[103] Hang Zhou, Xudong Xu, Dahua Lin, Xiaogang Wang, and Ziwei Liu. Sep-stereo: Visually guided stereophonic audio generation by associating source separation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2", + "[104] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, 2017. 1, 2, 3, 4", + "[105] Yipin Zhou and Ser-Nam Lim. Joint audio-visual deepfake detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14800–14809, 2021. 2", + "[106] Yuliang Zou, Pan Ji, Quoc-Huy Tran, Jia-Bin Huang, and Manmohan Chandraker. Learning monocular visual odometry via self-supervised long-term modeling. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV, pages 710-727. Springer, 2020. 2" + ], + "bbox": [ + 501, + 92, + 893, + 671 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "7908", + "bbox": [ + 482, + 949, + 516, + 960 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Sound Localization from Motion_ Jointly Learning Sound Direction and Camera Rotation/ebfa790a-8a91-4c3c-9e52-be53fe87e14f_model.json b/2023/Sound Localization from Motion_ Jointly Learning Sound Direction and Camera Rotation/ebfa790a-8a91-4c3c-9e52-be53fe87e14f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a32b6bf68c8844a5b6f0f5f10ffadc6101e5af48 --- /dev/null +++ b/2023/Sound Localization from Motion_ Jointly Learning Sound Direction and Camera Rotation/ebfa790a-8a91-4c3c-9e52-be53fe87e14f_model.json @@ -0,0 +1,2963 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.203, + 0.131, + 0.768, + 0.177 + ], + "angle": 0, + "content": "Sound Localization from Motion: Jointly Learning Sound Direction and Camera Rotation" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.204, + 0.379, + 0.223 + ], + "angle": 0, + "content": "Ziyang Chen" + }, + { + "type": "text", + "bbox": [ + 0.417, + 0.204, + 0.529, + 0.222 + ], + "angle": 0, + "content": "Shengyi Qian" + }, + { + "type": "text", + "bbox": [ + 0.569, + 0.205, + 0.695, + 0.221 + ], + "angle": 0, + "content": "Andrew Owens" + }, + { + "type": "text", + "bbox": [ + 0.384, + 0.234, + 0.573, + 0.253 + ], + "angle": 0, + "content": "University of Michigan" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.286, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.598 + ], + "angle": 0, + "content": "The images and sounds that we perceive undergo subtle but geometrically consistent changes as we rotate our heads. In this paper, we use these cues to solve a problem we call Sound Localization from Motion (SLfM): jointly estimating camera rotation and localizing sound sources. We learn to solve these tasks solely through self-supervision. A visual model predicts camera rotation from a pair of images, while an audio model predicts the direction of sound sources from binaural sounds. We train these models to generate predictions that agree with one another. At test time, the models can be deployed independently. To obtain a feature representation that is well-suited to solving this challenging problem, we also propose a method for learning an audio-visual representation through cross-view binauralization: estimating binaural sound from one view, given images and sound from another. Our model can successfully estimate accurate rotations on both real and synthetic scenes, and localize sound sources with accuracy competitive with state-of-the-art self-supervised approaches. Project site: https://ificl.github.io/SLfM." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.631, + 0.21, + 0.646 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.471, + 0.749 + ], + "angle": 0, + "content": "As you rotate your head, the images and sounds that you perceive change in geometrically consistent ways. For example, after turning to the right, a sound source that was directly in front of you will become louder in your left ear and quieter in your right, while simultaneously the visual scene will move right-to-left across your visual field (Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.902 + ], + "angle": 0, + "content": "We hypothesize that these co-occurring audio and visual signals provide \"free\" supervision that captures geometry, including the motion made by a camera and the direction of sound sources. These are each core problems in machine perception, but are largely studied separately, often using supervised methods that rely on difficult-to-acquire labeled training data, such as annotated sound directions. We take inspiration from self-supervised approaches to structure from motion [104], which learn to estimate 3D structure and camera pose by solving both tasks simultaneously." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.284, + 0.892, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.517, + 0.895, + 0.616 + ], + "angle": 0, + "content": "Figure 1: Images and sounds change in geometrically consistent ways. For example, when we rotate to the right, a sound source that is initially in front of us becomes louder in our left ear. We use these cues to jointly train models for two tasks: localizing sounds from binaural audio and estimating camera rotation from images. The two models are trained entirely through self-supervision, by learning to produce outputs that agree with one other." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.895, + 0.807 + ], + "angle": 0, + "content": "Analogously, we propose a problem we call sound localization from motion (SLfM): jointly estimating camera rotation from images and the sound direction from binaural audio. By solving both tasks simultaneously, we avoid the need for labeled training data. Our models provide each other with self-supervision: a visual model predicts the rotation angle between pairs of images, while an audio model predicts the azimuth of sound sources. We force their predictions to agree with one another, such that changes in rotation are consistent with changes in sound direction and binaural cues. After training, the models can be deployed independently, without multimodal data at test time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.897, + 0.902 + ], + "angle": 0, + "content": "This is a challenging task that requires perceiving motion in images and binaural cues in audio. Our second contribution is a method for learning representations that are well-suited to this task through cross-view binauralization. We train a network to convert mono to binaural sound for one viewpoint, given an audio-visual pair sampled from another" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "7897" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "viewpoint. Since the sound source is not necessarily visible in the images, the only way to successfully solve this pretext task is by analyzing the changes in the camera pose and predicting how they affect the sound direction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.289 + ], + "angle": 0, + "content": "All components of our model are entirely self-supervised and are trained solely on unlabeled audio-visual data. Our results suggest that paired audio-visual data provides a useful and complementary signal for learning about geometry. In contrast to other audio or visual self-supervised pose estimation methods, we obtain supervision from abundantly available audio data, thus avoiding the need of 3D ground truth or correspondences between pixels [104, 106] or audio samples [17]. Through experiments, we show:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.291, + 0.47, + 0.32 + ], + "angle": 0, + "content": "- Paired audio-visual data provides a supervisory signal for pose estimation tasks." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.322, + 0.469, + 0.351 + ], + "angle": 0, + "content": "- We obtain competitive performance with state-of-the-art self-supervised sound localization methods [17]." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.352, + 0.47, + 0.397 + ], + "angle": 0, + "content": "- We obtain strong rotation estimation performance, and our model generalizes to Stanford2D3D [6] dataset, where it is competitive with classic sparse feature matching methods." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.398, + 0.469, + 0.426 + ], + "angle": 0, + "content": "- The features we learn through our pretext task outperform other representations for our downstream tasks." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.291, + 0.47, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.443, + 0.219, + 0.459 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.468, + 0.471, + 0.755 + ], + "angle": 0, + "content": "Audio for spatial perception. Recent works have explored the use of sound for spatial understanding. Purushwalkam et al. [71] reconstructed floor plans in simulated environments [12]. Chen et al. [18] used ambient sounds from environments to learn about scene structures. Konno et al. [48] integrated sound localization to visual SfM while do not jointly learn them. Other work learns representations for spatial audio-visual tasks. Yang et al. [97] predicted whether stereo channels are swapped in a video, and Morgado et al. [63] solved a spatial alignment task. The learned representations are then used to improve localization, up-mixing, and segmentation models. In contrast, we learn camera pose and sound localization solely from self-supervision, obtaining angular predictions without labeled data. Other work uses echolocation sounds to learn representations [29, 96] and predict depth maps [19, 70] and estimate camera poses [96] using labeled data. In contrast, our proposed approach jointly learns binaural sound localization and camera pose through passive audio sensing, without supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Acoustic synthesis and spatialization. Researchers have explored visually-guided sound synthesis [28, 33, 42, 21] and text-guided audio synthesis [49, 95, 40]. Additionally, researchers have investigated generating realistic environmental acoustics using visual information [11, 84, 15, 57]. Chen et al. [14] introduced the novel-view acoustic synthesis task, which synthesizes binaural sound at the target view using audio and visual information from a source view. Liang et al. [50] proposed an audio-visual neural field in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "real-world audio-visual scenes. Many recent works have proposed to generate spatial audio from mono audio using visual cues [64, 30, 74, 93, 51, 103, 32], or the relative pose between sound sources and the receiver [78, 41]. Inspired by these works, our feature learning approach learns spatial representations through an audio prediction task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.19, + 0.895, + 0.462 + ], + "angle": 0, + "content": "Binaural sound localization. Humans have the ability to localize sound sources from binaural sound [80]. Traditional approaches estimate interaural time delays via cross-correlation using hand-crafted features [47], factorization methods [82], or loudness differences between ears [77, 90]. Chen et al. [17] adapted methods from self-supervised visual tracking to the problem of binaural sound localization. Similarly, we estimate direction through self-supervision. However, we obtain our supervision through cross-modal supervision from vision instead of from correspondence cues. Moreover, we also obtain visual camera rotation estimation through our learning process. Francl [27] learned representations of sound location with a contrastive loss where positive and negative examples are selected based on the extent of head movements. Other works have used supervised learning techniques with labeled data to localize sound sources in reverberant environments [1, 89, 94]. Unlike these methods, our model learns 3D sound localization without labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.895, + 0.757 + ], + "angle": 0, + "content": "Camera pose estimation. Traditional methods for camera pose estimation are based on finding correspondences between images and then solving an optimization problem using constraints from multi-view geometry [36]. These include structure from motion methods that estimate full pose [83] and camera rotation [8]. Recent methods have directly predicted camera pose using neural networks, including methods that use photos [72, 45, 59, 44, 55] or RGB-D scans [99, 98, 22, 23]. Our setup is similar to work that learns relative camera poses from sparse views [43, 9], and we use their network architectures. However, we learn the camera pose through cross-modal supervision from audio, rather than from labels. Our work is also closely related to methods that learn structure from motion through self-supervision [104, 106], such as by jointly learning models that perform depth and camera pose estimation with photoconsistency constraints. In contrast, our visual model's learning signal comes solely from audio-based supervision, and we jointly learn audio localization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Audio-visual learning. A number of works have focused on learning multimodal representations for audio and vision, taking into account semantic correspondence and temporal synchronization [67, 92, 7, 65, 68, 2, 60]. Other approaches study audio-visual sound localization [39, 61, 16, 62], source separation [56, 31, 58, 87], active speaker detection [3, 86, 5], navigation [12, 13, 10] and forensics [105, 34, 24]. Our focus, in contrast, is on utilizing multi-view audio-visual signals to learn geometry." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "7898" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.077, + 0.621, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.232, + 0.432, + 0.247 + ], + "angle": 0, + "content": "(a) Cross-view audio prediction" + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.077, + 0.891, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.232, + 0.877, + 0.246 + ], + "angle": 0, + "content": "(b) Sound localization from motion" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.253, + 0.894, + 0.31 + ], + "angle": 0, + "content": "Figure 2: Method overview. (a) We learn a feature representation by predicting how changes in images lead to changes in sound in a cross-view binauralization pretext task. We convert mono sound to binaural sound at a target viewpoint, after conditioning the model on observations from a source viewpoint. (b) We use the representation to jointly solve two pose estimation tasks: visual rotation estimation and binaural sound localization. We train the visual rotation angle, \\(\\phi_{s,t}\\), to be consistent with the difference in predicted sound angles \\(\\theta_s\\) and \\(\\theta_t\\)." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.325, + 0.376, + 0.342 + ], + "angle": 0, + "content": "3. Sound Localization from Motion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.47, + 0.44 + ], + "angle": 0, + "content": "We address the sound localization from motion (SLfM) task: predicting the azimuth of a sound source from binaural audio and camera rotation from two images. First, we present a self-supervised representation that can be used to solve this downstream task. Then, we show how the representation can be used to solve the task." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.451, + 0.434, + 0.467 + ], + "angle": 0, + "content": "3.1. Learning representation via spatialization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.47, + 0.639 + ], + "angle": 0, + "content": "We learn an audio-visual representation that conveys spatial information by solving a cross-view binauralization task: converting mono sound to stereo for one viewpoint, given an audio-visual pair sampled from another viewpoint (Fig. 2a). In order to successfully solve the task, a model must implicitly estimate the sound direction in the source view and predict how the change in viewpoint will affect the sound in the target view. A key difference between this task and traditional, single-view binauralization [30] is that the sound source need not be visible in any of the images. Hence, the task cannot be solved from the target audio-visual pair alone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.471, + 0.761 + ], + "angle": 0, + "content": "We binauralize the sound \\(\\mathbf{a}_t\\) at the target view through an audio predictor \\(\\mathcal{F}_{\\theta}\\). To make our representation suitable for downstream tasks, we factorize it into visual features \\(f_{v}(\\mathbf{v}_{s},\\mathbf{v}_{t})\\), which are intended to create features relevant to relative pose, and audio features \\(f_{a}(\\mathbf{a}_{s})\\), which capture sound localization cues. We predict the binauralized audio \\(\\hat{\\mathbf{a}}_t\\) at the target view from mono audio \\(\\bar{\\mathbf{a}}_t\\), the visual change of \\((\\mathbf{v}_s,\\mathbf{v}_t)\\) and binaural sound \\(\\mathbf{a}_s\\) heard at the source viewpoint:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.772, + 0.47, + 0.788 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {a}} _ {t} = \\mathcal {F} _ {\\theta} \\left(\\bar {\\mathbf {a}} _ {t}, f _ {v} \\left(\\mathbf {v} _ {s}, \\mathbf {v} _ {t}\\right), f _ {a} \\left(\\mathbf {a} _ {s}\\right)\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.887 + ], + "angle": 0, + "content": "We represent audio \\( \\mathbf{a} \\) as a spectrogram \\( \\mathbf{A} \\) using short-time Fourier transform (STFT). Following Gao et al. [30], given the mix of stereo audio \\( \\bar{\\mathbf{A}}_t = \\mathrm{STFT}(\\mathbf{a}_t^L + \\mathbf{a}_t^R) \\), we predict the difference of two channels \\( \\mathbf{A}_t = \\mathrm{STFT}(\\mathbf{a}_t^L - \\mathbf{a}_t^R) \\). We optimize the \\( L1 \\) loss between predicted spectrogram \\( \\hat{\\mathbf{A}}_t \\) and ground-truth spectrogram \\( \\mathbf{A}_t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.89, + 0.47, + 0.907 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p r e t e x t}} = \\left| \\left| \\hat {\\mathbf {A}} _ {t} - \\mathbf {A} _ {t} \\right| \\right| _ {1}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.326, + 0.895, + 0.448 + ], + "angle": 0, + "content": "Multi-view binauralization. Following Zhou et al. [104], we improve our representations by binauralizing sounds at \\(N\\) different target viewpoints, using observations from a single source viewpoint \\(s\\). We hypothesize that jointly solving spatialization problems from a single viewpoint for multiple target viewpoints would require the model to make more accurate predictions of sound source locations, thereby improving the estimation of view changes:" + }, + { + "type": "equation", + "bbox": [ + 0.505, + 0.461, + 0.892, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p r e t e x t}} = \\frac {1}{N} \\sum_ {i} ^ {N} \\left| \\left| \\mathcal {F} _ {\\theta} \\left(\\bar {\\mathbf {A}} _ {i}, f _ {v} \\left(\\mathbf {v} _ {s}, \\mathbf {v} _ {i}\\right), f _ {a} \\left(\\mathbf {a} _ {s}\\right)\\right) - \\mathbf {A} _ {i} \\right| \\right| _ {1}. \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.519, + 0.827, + 0.535 + ], + "angle": 0, + "content": "3.2. Estimating pose and localizing sounds" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.892, + 0.634 + ], + "angle": 0, + "content": "We now address the problem of learning models for sound localization and pose estimation, using our self-supervised audio-visual features. Given two views, we predict sound directions and relative rotation. We train the model to make these two predictions consistent with one another, while using simple binaural constraints to resolve ambiguities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.635, + 0.892, + 0.681 + ], + "angle": 0, + "content": "We are given images \\(\\mathbf{v}_s\\) and \\(\\mathbf{v}_t\\) (rotated views recorded at the same position) and learned visual embedding \\(f_v\\). We predict the scalar rotation angle \\(\\phi_{s,t}\\) via the encoder \\(g_v\\):" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.695, + 0.892, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\phi_ {s, t} = g _ {v} \\left(f _ {v} \\left(\\mathbf {v} _ {s}, \\mathbf {v} _ {t}\\right)\\right), R _ {s, t} = \\left[ \\begin{array}{l l} \\cos \\phi_ {s, t} & - \\sin \\phi_ {s, t} \\\\ \\sin \\phi_ {s, t} & \\cos \\phi_ {s, t} \\end{array} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.741, + 0.894, + 0.816 + ], + "angle": 0, + "content": "where \\( R_{s,t} \\) is 2D rotation matrix of \\( \\phi_{s,t} \\). Following common practice in indoor scene reconstruction, we give the camera a fixed downward tilt [102, 73, 101] and only estimate azimuth [43]. This is also a common assumption in audio localization [1, 89], since azimuth has strong binaural cues." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.818, + 0.892, + 0.848 + ], + "angle": 0, + "content": "Similarly, we predict the azimuths of the sound sources using audio features and the encoder \\( g_{a} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.862, + 0.892, + 0.883 + ], + "angle": 0, + "content": "\\[\n\\theta_ {i} = g _ {a} \\left(f _ {a} (\\mathbf {a} _ {i})\\right), \\quad \\mathbf {r} _ {i} = \\left[ \\begin{array}{l l} \\cos \\theta_ {i} & \\sin \\theta_ {i} \\end{array} \\right] ^ {\\top}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.897, + 0.803, + 0.911 + ], + "angle": 0, + "content": "where we represent the azimuth as a vector \\(\\mathbf{r}_i\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.95, + 0.516, + 0.962 + ], + "angle": 0, + "content": "7899" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.18 + ], + "angle": 0, + "content": "Cross-modal geometric consistency. When the camera is rotated, the sound source ought to rotate in the opposite direction (Fig. 2b). For example, a \\(30^{\\circ}\\) clockwise camera rotation should result in a \\(30^{\\circ}\\) counterclockwise rotation in sound direction. Such a constraint could be converted into a loss:" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.181, + 0.471, + 0.199 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r o t}} = \\left\\| \\mathbf {r} _ {s} - R _ {s, t} \\mathbf {r} _ {t} \\right\\| ^ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.205, + 0.474, + 0.31 + ], + "angle": 0, + "content": "However, a well-known ambiguity called front-back confusion [77, 25] exists in binaural sound perception: one cannot generally tell whether a sound is in front of the view, or behind them. To address that, we use permutation invariant training [100] (PIT), and allow the model to use either the predicted sound direction or its reflection about the \\( x \\) axis without penalty. This results in the loss:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.318, + 0.471, + 0.354 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {g e o}} = \\min _ {\\substack {\\hat {\\mathbf {r}} _ {s} \\in \\{\\mathbf {r} _ {s}, Q \\mathbf {r} _ {s} \\} \\\\ \\hat {\\mathbf {r}} _ {t} \\in \\{\\mathbf {r} _ {t}, Q \\mathbf {r} _ {t} \\}}} \\| \\hat {\\mathbf {r}} _ {s} - R _ {s, t} \\hat {\\mathbf {r}} _ {t} \\| ^ {2}, \\tag{7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.361, + 0.413, + 0.395 + ], + "angle": 0, + "content": "where \\(Q = \\begin{bmatrix} 1 & 0 \\\\ 0 & -1 \\end{bmatrix}\\) reflects the sound direction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.471, + 0.49 + ], + "angle": 0, + "content": "As a consequence of this ambiguity, there are also two possible solutions for the visual rotation model, since the visual rotation matrices can be mirrored about the \\( x \\) axis. For example, one can create a solution with equal loss by multiplying the rotations and sound directions by \\( Q \\). We discuss this ambiguity in more depth in Sec. 4.4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.495, + 0.473, + 0.682 + ], + "angle": 0, + "content": "Incorporating binaural observations. Without additional constraints, the solution is ambiguous, and may collapse into a trivial solution (e.g., predicting zero for all three angles). To avoid this, we force the model to agree with a simple binaural cue based on interaural intensity difference (IID). We predict whether the sound is to the left or right of the viewer, based on whether it is louder in the left or right microphone: \\( d = \\mathrm{sign}(\\log \\left|\\frac{\\mathbf{A}^L}{\\mathbf{A}^R}\\right|) \\), where \\( |\\mathbf{A}| \\) is the magnitude of the spectrogram A. We perform this left/right test at each timestep in the spectrogram and then pool via majority voting (see supplementary for details). We penalize predictions that are inconsistent with these \"left or right\" observations:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.688, + 0.47, + 0.705 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {b i n a u r a l}} = \\mathcal {L} _ {\\text {B C E}} (\\sin \\theta_ {i}, d _ {i}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.351, + 0.728 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{BCE}}\\) is binary cross entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.473, + 0.839 + ], + "angle": 0, + "content": "Encouraging symmetry. To help regularize the model, we also add symmetry constraints. For sound localization, swapping the left and right channels of the audio ought to result in a prediction in the opposite direction since the binaural cues are reversed. For rotation estimation, the relative pose between images \\( s \\) and \\( t \\) should invert the pose from \\( t \\) to \\( s \\). We encourage both constraints via a loss:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.85, + 0.47, + 0.868 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s y m}} = \\left| \\theta + \\theta_ {\\text {f l i p}} \\right| + \\left| \\phi_ {s, t} + \\phi_ {t, s} \\right|, \\tag {9}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.471, + 0.902 + ], + "angle": 0, + "content": "1Work on self-supervised SfM has similar ambiguities [104], and deals with them by adding analogous constraints, such as photometric consistency." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.075, + 0.895, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.235, + 0.895, + 0.292 + ], + "angle": 0, + "content": "Figure 3: Cross-view binauralization architecture. We take a mono spectrogram as input and fuse the audio and visual features from the audio and visual encoder respectively to synthesize the binaural spectrogram at the target viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.312, + 0.895, + 0.357 + ], + "angle": 0, + "content": "where \\(\\theta_{\\mathrm{flip}}\\) is the prediction of sound angle \\(\\theta\\) using audio with swapped audio channels, and \\(\\phi_{s,t}\\) and \\(\\phi_{t,s}\\) are the predicted rotations between cameras \\(s\\) and \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.371, + 0.892, + 0.4 + ], + "angle": 0, + "content": "Overall loss. We combine these constraints to obtain an overall loss:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.41, + 0.892, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda \\mathcal {L} _ {\\mathrm {g e o}} + \\mathcal {L} _ {\\text {b i n a u r a l}} + \\mathcal {L} _ {\\text {s y m}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.441, + 0.8, + 0.457 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the weight for the geometric loss." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.471, + 0.634, + 0.488 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.498, + 0.895, + 0.619 + ], + "angle": 0, + "content": "We have introduced a self-supervised method to learn camera pose and sound localization from audio-visual data. In experiments, we first evaluate how well our learned representation captures spatial information. We then evaluate how well our method learns camera pose and sound localization by comparing it with baselines. Finally, we show generalization to indoor panorama images Stanford2D3D [6] and in-the-wild binaural audio [17]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.633, + 0.667, + 0.649 + ], + "angle": 0, + "content": "4.1. Implementations" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.655, + 0.894, + 0.777 + ], + "angle": 0, + "content": "Visual pose encoder. We follow recent pose estimation work [9, 43] and build a Siamese-style visual pose network \\( f_{v} \\) with ResNet-18 [38] as the backbone. We compute dense 4D correlation volumes between the features from the third residual layer and then encode them by convolution layers. We resize images to \\( 320 \\times 240 \\) and encode a pair of images into 512-d features. We use an MLP \\( g_{v} \\) to map visual features to 1-d logits for our SLfM models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.79, + 0.895, + 0.912 + ], + "angle": 0, + "content": "Binaural audio encoder. We obtain binaural audio embeddings \\( f_{a}(\\cdot) \\) using ResNet-18 [38] that operates on spectrograms. We covert the two-channel waveform of length \\( L \\) to a spectrogram representation of size \\( 256 \\times 256 \\times 4 \\) using short-time Fourier transform (STFT), where we keep both the magnitude and phase of spectrograms. We extract 512-d features of binaural sound with \\( f_{a} \\) and map them to 1-d logits using an MLP \\( g_{a} \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.95, + 0.516, + 0.962 + ], + "angle": 0, + "content": "7900" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "Audio prediction model. We adopt the light-weighted audio-visual U-Net [30] to perform binauralization. We feed in spectrograms of size \\(256 \\times 256 \\times 2\\) and predict the target spectrograms. We concatenate the visual pose features \\(f_{v}(\\cdot)\\) and audio features \\(f_{a}(\\cdot)\\) at the bottleneck of U-Net. We show the architecture of our models in Fig. 3. Please see supp. for more implementation details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.207, + 0.173, + 0.22 + ], + "angle": 0, + "content": "4.2. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.473, + 0.379 + ], + "angle": 0, + "content": "Since there is no public multi-view audio-visual dataset with camera poses and sound direction ground truth, we use the SoundSpaces 2.0 platform [15] to create a dataset. Our 3D scenes come from Habitat-Matterport 3D dataset (HM3D) [76], which is a large dataset of real 3D scenes. This setup allows us to have photorealistic images and high-quality spatial audio with real-world acoustics phenomenon (e.g., reverberation), as well as providing the ground-truth camera pose and sound directions that can be used for evaluation. We call this dataset HM3D-SS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.473, + 0.621 + ], + "angle": 0, + "content": "We generate binaural Room Impulse Responses (RIRs) and images with a \\(60^{\\circ}\\) field of view, using 100 scenes of HM3D [76]. For each audio-visual example, we randomly place sound sources in the scene with a height range of (0.7, 1.7) meters, and sample 4 different rotated viewpoints at one location within 4 meters. The rotations are limited to \\((10^{\\circ}, 90^{\\circ})\\) relative to the source viewpoints. We follow the standard practice to set the height to agents to be \\(1.5\\mathrm{m}\\) and lock a downward tilt angle [43, 101, 73, 102]. We render the binaural RIRs and images given the position of agents and sound sources. We obtain binaural audio by convolving binaural RIRs with mono audio samples from LibriSpeech [69] and Free Music Archive [20]. To ensure that the evaluation tests the model's pose estimation abilities, rather than its ability to visually localize sound sources, the sound sources are not visible on screen." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.622, + 0.472, + 0.682 + ], + "angle": 0, + "content": "We create 50K audio-visual pairs from 200K viewpoints. The audio was rendered with average reverberation of \\(\\mathrm{RT}_{60} = 0.4\\mathrm{s}\\) (see supp. for details). We divided our data into 81/9/10 scenes for the train/val/test, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.691, + 0.405, + 0.707 + ], + "angle": 0, + "content": "4.3. Evaluating the learned representation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.714, + 0.471, + 0.775 + ], + "angle": 0, + "content": "First, we directly evaluate the quality of our learned features for rotation estimation and sound localization via linear probing with labeled data (rather than learning them jointly through self-supervision)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Baselines and ablations. We compare our model with several baselines that use alternative pretext tasks: 1) AVSA [63]: it learns spatial cues by training a model to spatially align video and audio clips extracted from different viewing angles. We adapt this model to our dataset and train with 4 different views; 2) RotNCE [27]: it applies contrastive learning on the audio from different angles and uses annotations of the agent's rotation to select positive" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "and negative samples, which results in learning audio spatial representation. For baselines, we use the same architecture for feature extractors to ensure fair comparisons." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.895, + 0.242 + ], + "angle": 0, + "content": "To determine if we utilize visual and audio features from different views to solve the binauralization task, we also study some variants of our models: 1) Ours-NoA: we only provide visual features for the binauralization task; 2) Ours-NoV: which only uses audio from the other view to spatialize sounds; 3) Ours-GTRot: we provide ground-truth rotation embedding instead of features from visual frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.895, + 0.304 + ], + "angle": 0, + "content": "Besides the mono-to-binaural task, we also experiment with another objective: predicting the right channel from the left channel. We train our L2R model with the same setup as our M2B model. (Please see supp. for pretext results.)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.309, + 0.897, + 0.568 + ], + "angle": 0, + "content": "Downstream tasks. We assess the quality of spatial representations we learned from our pretext tasks in two downstream tasks: relative camera rotation and 3D sound localization. We formulate them as classification problems, where angles are categorized into 64 bins, and we use accuracy as the evaluation metric. To evaluate the learned features, we freeze them and train a linear classifier on the downstream tasks. We compare the performance of our features with those learned from RotNCE [27], AVSA [63], ImageNet [37], and random features, and report the results in Tab. 1. Our approach outperforms the baselines in both tasks, indicating that we learn better spatial representations. Furthermore, our linear probe models show comparable performance against the supervised method which can be regarded as approximate upper bounds for our models, suggesting that our pretext tasks help learn a useful representation. Please see supp. for experiments on FreeMusic [20]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.573, + 0.895, + 0.65 + ], + "angle": 0, + "content": "Emerging camera pose from audio prompting. To help understand the strong performance of our self-supervised features, we asked whether we could use the cross-view binauralization model alone to estimate camera rotation. Inspired by prompting in vision and language models [75], we" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.657, + 0.88, + 0.86 + ], + "angle": 0, + "content": "
ModelAudio Loc. Acc (%) ↑Camera Rot. Acc (%) ↑
LibriSpeechRandom feature4.84.7
ImageNet [38]+Random-56.3
RotNCE [27]50.9-
AVSA [63]71.26.5
Ours-NoA-9.6
Ours-NoV53.9-
Ours-GTRot70.6-
Ours-L2R (3 views)78.576.1
Ours (2 views)74.580.0
Ours (3 views)75.481.3
Supervised81.595.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.864, + 0.895, + 0.905 + ], + "angle": 0, + "content": "Table 1: Downstream task performance on HM3D-SS dataset. We report linear probe performance on the audio localization and camera rotation downstream tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.95, + 0.514, + 0.962 + ], + "angle": 0, + "content": "7901" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.071, + 0.448, + 0.42 + ], + "angle": 0, + "content": "
ModelAudio angle MAE (°) ↓Camera angle MAE (°) ↓
LibriSpeechChance40.2829.41
SIFT [54]-12.2
Ours w/o Reflect28.0826.80
Ours-Prompt-9.13
Ours-L2R3.220.80
Ours3.170.77
Ours-Front4.483.08
Ours-GTRot1.83-
Superglue [81]-2.47
Supervised1.710.46
FreeMusicChance40.8129.41
SIFT [54]-12.2
Ours w/o Reflect28.2426.81
Ours-L2R3.180.81
Ours3.370.84
Ours-Front3.992.49
Ours-GTRot1.96-
Superglue [81]-2.47
Supervised2.500.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.425, + 0.472, + 0.455 + ], + "angle": 0, + "content": "Table 2: Sound localization from motion results on HM3D-SS. We evaluate our SLfM models on each modality independently." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.471, + 0.473, + 0.669 + ], + "angle": 0, + "content": "obtain rough estimates of camera rotation by providing our model with carefully-provided inputs. Given a pair of images \\((\\mathbf{v}_s,\\mathbf{v}_t)\\), we create a synthetic binaural audio prompt, \\(\\mathbf{a}_s\\). We then ask our model to generate the binaural sound \\(\\hat{\\mathbf{a}}_t\\) for the target viewpoint. By analyzing the IID cues in \\(\\hat{\\mathbf{a}}_t\\), we can estimate the model's implicitly predicted camera pose. To do this, we find the nearest neighbor of our generated audio \\(\\hat{\\mathbf{a}}_t\\), using a database of synthetically generated audio with known sound directions. Please refer to supp. for more details. Our method achieves the mean absolute error of \\(9.13^{\\circ}\\) on HM3D-SS dataset where chance is \\(29.41^{\\circ}\\) (Tab. 2). Our approach can also generalize to Stanford2D3D [6] dataset, where we can achieve mean absolute error of \\(9.93^{\\circ}\\) (Tab. 3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.677, + 0.248, + 0.693 + ], + "angle": 0, + "content": "4.4. Evaluating SLfM" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.472, + 0.763 + ], + "angle": 0, + "content": "We evaluated our SLfM models on the HM3D-SS dataset. We use the mean absolute error of angle in degrees as evaluation metrics. To avoid sound field ambiguity, we filtered out samples with sound angles outside of \\((-90^{\\circ}, 90^{\\circ})\\) for the" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.77, + 0.307, + 0.859 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.771, + 0.468, + 0.859 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.864, + 0.473, + 0.907 + ], + "angle": 0, + "content": "Figure 4: Qualitative results on real-world examples. We show our predictions on Stanford2D3D [6] and In-the-wild audio [17]. Green denotes accurate predictions." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.072, + 0.691, + 0.207 + ], + "angle": 0, + "content": "
In the wild audio [17]
ModelAcc (%) ↑
GCC-PHAT [47]77.2
IID [17]75.4
IID (Ours)82.1
MonoCLR [17]87.4
StereoCRW [17]87.2
Ours-L2R84.9
Ours84.0
" + }, + { + "type": "table", + "bbox": [ + 0.708, + 0.075, + 0.894, + 0.207 + ], + "angle": 0, + "content": "
Stanford2D3D [6]
ModelRot Err. (°) ↓
MeanMed.
SIFT [54]16.40.06
LoFTR [85]6.101.13
Superglue [81]5.070.07
Ours-Prompt9.939.02
Ours-L2R1.120.71
Ours1.140.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.211, + 0.895, + 0.268 + ], + "angle": 0, + "content": "Table 3: Evaluation of our SLfM models on the real-world data. We evaluate our audio localization model on the In-the-wild audio [17] (left) and our camera pose model on the Stanford2D3D [6] (right). Rot. denote camera rotation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.285, + 0.894, + 0.361 + ], + "angle": 0, + "content": "evaluation set. When training our SLfM models, we use our best-performing features, i.e., pretext tasks trained with 3 views. We freeze learned audio and visual features and only train multi-layer perceptrons on top of them and evaluate them independently on each modality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.895, + 0.531 + ], + "angle": 0, + "content": "Baselines and ablations. For relative camera pose estimation, we compare our models with sparse feature matching using SIFT [54] and SuperGlue [81], followed by rotation fitting. We use those methods to detect key points, run Lowe's ratio test, and use RANSAC with five-point algorithm to recover the camera pose [54, 36, 66, 26]. For the sound localization task, we compared with time delay estimation methods: the popular GCC-PHAT [47] and the recent self-supervised method StereoCRW [17]. As far as we know, there are no other baselines that can estimate poses and sound source directions without labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.894, + 0.607 + ], + "angle": 0, + "content": "We also compared several variations of our method, including Ours-Front, where we filter out the samples with sound sources behind the viewers to remove binaural ambiguity during training, and an oracle model Ours-GTRot, which uses ground-truth rotation angles instead." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.894, + 0.689 + ], + "angle": 0, + "content": "Results. We show our results in Tab. 2. Our models can predict the azimuths of sound sources, obtaining strong performance without any labels. Without providing reflection invariance (Eq. (7)), the model failed to learn reasonable geometric due to the audio ambiguity. Our self-supervised" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.696, + 0.892, + 0.832 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.836, + 0.895, + 0.907 + ], + "angle": 0, + "content": "Figure 5: Real video. We play music using a speaker and record a video using iPhone with a binaural microphone (different from training examples). We show the predicted camera rotation and sound direction over time. They change smoothly and match our camera motion. Please see our project webpage for video results." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.95, + 0.516, + 0.962 + ], + "angle": 0, + "content": "7902" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.073, + 0.891, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.289, + 0.897, + 0.348 + ], + "angle": 0, + "content": "Figure 6: Qualitative results on HM3D-SS. Given two images and their corresponding audio at the source viewpoints, our approach can individually predict the relative camera pose and localize sound locations accurately. We visualize our camera rotation prediction with blue bar and our sound angle prediction with orange bar. To highlight the subtle differences in the waveforms, we color code the amplitude. Please see the project webpage for more video results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.473, + 0.44 + ], + "angle": 0, + "content": "model achieves comparable performance against our oracle model (Ours-GTRot) and supervised models, indicating we estimate camera poses and localize sound accurately. We show some qualitative results on HM3D-SS in Fig. 6 with LibriSpeech samples [69]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.473, + 0.689 + ], + "angle": 0, + "content": "Generalization to other datasets. We further demonstrate the generalization ability of our models by experimenting with out-of-distribution, real-world data. We evaluate our camera rotation model on Stanford2D3D [6] with real indoor RGB images (Tab. 3). We obtain image pairs by cropping from panoramas. Although our model is trained on renderings of HM3D [76], it obtains strong generalization ability. Compared with rotation estimation based on SIFT [54] and SuperGlue [81], our model is significantly better on mean rotation error. We also report median rotation error to be consistent with prior works [9, 43]. However, SIFT [54] and Superglue [81] have a very low median error. This is likely due to the all-or-nothing nature of feature matching-based approaches, which either produce highly accurate predictions if the matches are correct (especially for \"easy\" cases with small amounts of rotation) or else produce gross errors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.474, + 0.901 + ], + "angle": 0, + "content": "We also evaluate our sound localization model on the in-the-wild binaural audio [17]. We use binary accuracy as the metric for left-or-right direction classification accounting for the fact that microphone baselines are unknown for internet videos. For an apples-to-apples comparison to prior work, we retrained our model using 0.51s length of audio for the binaural audio encoder \\( f_{a}(\\cdot) \\). As shown in Tab. 3, our model obtains similar performance to StereoCRW, a state-of-the-art self-supervised time delay method, suggesting we have a strong capability for sound localization. We show qualitative results in Fig. 4. We also perform both tasks on a self-recorded video (Fig. 5) of a rotating camera and a binaural microphone. We show the mean and standard deviation of predictions in 1.0s windows." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.895, + 0.53 + ], + "angle": 0, + "content": "Handling ambiguity. In binaural sound perception, there is a fundamental ambiguity that whether the sound is in front of or behind us. It leads to multiple solutions with equal loss in our model, mirroring sound sources and negating rotation angles with flipped \\( z \\) axis (Fig. 7). These two solutions differ in that a visual rotation angle either indicates a clockwise or counterclockwise rotation. For evaluation, we convert \"backwards\" counterclockwise predictions to clockwise predictions by simply providing the model with pairs of input frames, then negating the model's outputs if the angle is the opposite of the expected direction2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.538, + 0.859, + 0.556 + ], + "angle": 0, + "content": "4.5. Generalization to more complex scenarios" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.895, + 0.607 + ], + "angle": 0, + "content": "We investigate whether our approach generalizes to more complex scenarios, such as with multiple sound sources or when translation is included in camera motion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.895, + 0.673 + ], + "angle": 0, + "content": "Multiple sound sources. We evaluate our representations and SLfM models on more complex scenes containing multiple sound sources. We train our models with two source sources placed in the scenes. One of them is dominant and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.681, + 0.892, + 0.707 + ], + "angle": 0, + "content": "2Similar to ambiguities in SfM where reconstructions can be reoriented such that the sky is in the positive \\(y\\) direction [36, 35, 4]." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.715, + 0.681, + 0.832 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.836, + 0.684, + 0.906 + ], + "angle": 0, + "content": "Figure 7: Mirror ambiguity. The \\( +z \\) and \\( +x \\) axes represent forward and rightward directions. Both solutions have the same loss." + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.716, + 0.892, + 0.832 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.837, + 0.895, + 0.892 + ], + "angle": 0, + "content": "Figure 8: Robustness to reverberation. We study the effect of reverberation on our SLfM model." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.95, + 0.516, + 0.962 + ], + "angle": 0, + "content": "7903" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.074, + 0.445, + 0.214 + ], + "angle": 0, + "content": "
ModelMulti-source Acc (%) ↑Small trans. Acc (%) ↑
Aud.Rot.Aud.Rot.
Random feature4.54.75.73.8
ImageNet [38]+Random-56.3-23.0
RotNCE [27]35.8---
AVSA [63]55.96.765.46.1
Ours (3 views)59.277.972.249.6
Supervised74.595.881.663.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.22, + 0.47, + 0.276 + ], + "angle": 0, + "content": "Table 4: Generalization to more complex scenarios. We evaluate the audio and visual features learned from more complex scenarios via linear probing. Aud. and Rot. denote audio localization and camera rotation respectively." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.281, + 0.432, + 0.43 + ], + "angle": 0, + "content": "
ModelMulti-source MAE (°) ↓Small trans. MAE (°) ↓
Aud.Rot.Aud.Rot.
Chance40.4629.4140.2829.41
SIFT [54]-12.2-12.2
Ours7.670.716.281.04
Ours-GTRot5.81-4.24-
Superglue [81]-2.47-2.47
Supervised3.600.461.710.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.435, + 0.47, + 0.492 + ], + "angle": 0, + "content": "Table 5: SLfM results on more complex scenarios. We evaluate our model on the version of HM3D-SS with two sound sources. We also evaluate our model trained with small translations on rotation-only examples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.498, + 0.47, + 0.575 + ], + "angle": 0, + "content": "our sound localization target. We report the results in Tab. 4 and Tab. 5. Our model learns better representations than baselines and achieves accurate predictions of the azimuth of the source and camera pose even in challenging scenarios with multiple sound sources." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.471, + 0.809 + ], + "angle": 0, + "content": "Translation in camera motions. To study how small translations in the camera motion could affect our models, we generate 50K pairs of audio-visual data with both rotation and translation change, limiting the uniformly sampled translation to 0.5 meters. We train our models with LibriSpeech samples [69]. For our linear probing evaluation, we measure the ability of features to handle complex examples by testing on the dataset that has translation. For our SLfM model, we study our ability to learn from noisy data, thus we evaluate it on the rotation-only examples. As the results are shown in Tab. 4 and Tab. 5, we successfully learn useful features and obtain accurate rotation and sound direction predictions despite the presence of translation. Since we jointly learn the audio and visual representations, it can negatively impact the learning of one modality when another one becomes harder." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.819, + 0.227, + 0.835 + ], + "angle": 0, + "content": "4.6. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Robustness to reverberation. We study how our SLfM models perform under the different reverberation configurations. We used SoundSpaces [15] to create audio with average reverberation \\(\\mathrm{RT}_{60} \\in \\{0.1, 0.2, 0.3, 0.4, 0.5\\}\\) while" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.074, + 0.888, + 0.169 + ], + "angle": 0, + "content": "
ModelLossesAudio angle MAE (°) ↓Camera angle MAE (°) ↓
LgeoLbinauralLsym
Ours✓37.6029.20
✓✓37.5229.17
✓✓3.586.99
✓✓✓3.170.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.174, + 0.892, + 0.201 + ], + "angle": 0, + "content": "Table 6: Ablation experiments on our SLfM losses. We evaluate our SLfM models with different combinations of losses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.212, + 0.892, + 0.289 + ], + "angle": 0, + "content": "keeping the visual signals the same. We train our model on each setting. As shown in Fig. 8, our performance decreases as the level of reverberation increases, where audio becomes more challenging during both training and testing. Please see supp. for the study on representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.293, + 0.892, + 0.354 + ], + "angle": 0, + "content": "Losses for SLfM. We study the necessity of our proposed loss functions in Tab. 6. Our models fail to learn accurate pose estimation without binaural loss or symmetric loss. It highlights the crucial role of these losses." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.367, + 0.619, + 0.383 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.392, + 0.895, + 0.634 + ], + "angle": 0, + "content": "In this paper, we proposed the sound localization from motion (SLfM) problem, and provided a self-supervised method for solving it. We also presented a method for learning audio-visual features that convey sound directions and camera rotation, which we show are well-suited to solving the SLfM task. Despite learning our models solely from unlabeled audio-visual data, we obtain strong performance on a variety of benchmarks, including rotation estimation on the Stanford2D3D [6] dataset and \"in the wild\" sound direction estimation [17]. Our results suggest that the subtle correlations between sights and binaural sounds that result from rotational motion provide a useful (and previously unused) learning signal. We see our work as opening new directions in self-supervised geometry estimation and feature learning that use sound as a complementary source of supervision. We will release code, data, and models upon acceptance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.894, + 0.775 + ], + "angle": 0, + "content": "Limitations and Broader Impacts. Our work has several limitations. First, while we evaluate our models on real images and sounds, we train on data from simulators, due to a lack of available relevant data. We note that this is common practice in visual 3D reconstruction [43, 91, 52]. Second, we assume that the 3D scene and sound sources are stationary. Third, we do not evaluate our model on extreme viewpoint changes [55], which requires reasoning about images that have little or no overlap." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. We would like to thank Ang Cao, Tiange Luo, and Linyi Jin for their helpful discussion. We thank Changan Chen for his help with SoundSpaces 2.0. This work was funded in part by DARPA Semafor and Sony. The views, opinions and/or findings expressed are those of the authors and should not be interpreted as representing the official views or policies of the Department of Defense or the U.S. Government." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.95, + 0.516, + 0.962 + ], + "angle": 0, + "content": "7904" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Sharath Adavanne, Archontis Politis, and Tuomas Virtanen. Direction of arrival estimation for multiple sound sources using convolutional recurrent neural network. In 2018 26th European Signal Processing Conference (EUSIPCO), pages 1462-1466. IEEE, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.186, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Triantafyllos Afouras, Yuki M Asano, Francois Fagan, Andrea Vedaldi, and Florian Metze. Self-supervised object detection from audio-visual correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10575–10586, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.256, + 0.471, + 0.31 + ], + "angle": 0, + "content": "[3] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. arXiv preprint arXiv:2008.04237, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.312, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[4] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54(10):105-112, 2011. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.368, + 0.47, + 0.424 + ], + "angle": 0, + "content": "[5] Juan León Alcazar, Fabian Caba, Ali K Thabet, and Bernard Ghanem. Maas: Multi-modal assignation for active speaker detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 265-274, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.425, + 0.471, + 0.466 + ], + "angle": 0, + "content": "[6] Iro Armeni, Sasha Sax, Amir R Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105, 2017. 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.468, + 0.471, + 0.522 + ], + "angle": 0, + "content": "[7] Yuki Asano, Mandela Patrick, Christian Rupprecht, and Andrea Vedaldi. Labelling unlabelled videos from scratch with multi-modal self-supervision. Advances in Neural Information Processing Systems, 33:4660-4671, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.524, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[8] Matthew Brown and David G Lowe. Automatic panoramic image stitching using invariant features. International journal of computer vision, 74:59-73, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.566, + 0.47, + 0.634 + ], + "angle": 0, + "content": "[9] Ruojin Cai, Bharath Hariharan, Noah Snavely, and Hadar Averbuch-Elor. Extreme rotation estimation using dense correlation volumes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14566-14575, 2021. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.636, + 0.471, + 0.69 + ], + "angle": 0, + "content": "[10] Changan Chen, Ziad Al-Halah, and Kristen Grauman. Semantic audio-visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15516–15525, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.692, + 0.471, + 0.747 + ], + "angle": 0, + "content": "[11] Changan Chen, Ruohan Gao, Paul Calamia, and Kristen Grauman. Visual acoustic matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18858-18868, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.748, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[12] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part VI 16, pages 17–36. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[13] Changan Chen, Sagnik Majumder, Ziad Al-Halah, Ruohan Gao, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Learning to set waypoints for audio-visual navigation. arXiv preprint arXiv:2008.09622, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.894, + 0.148 + ], + "angle": 0, + "content": "[14] Changan Chen, Alexander Richard, Roman Shapovalov, Vamsi Krishna Ithapu, Natalia Neverova, Kristen Grauman, and Andrea Vedaldi. Novel-view acoustic synthesis. arXiv preprint arXiv:2301.08730, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.149, + 0.894, + 0.219 + ], + "angle": 0, + "content": "[15] Changan Chen, Carl Schissler, Sanchit Garg, Philip Kobernik, Alexander Clegg, Paul Calamia, Dhruv Batra, Philip W Robinson, and Kristen Grauman. Soundspaces 2.0: A simulation platform for visual-acoustic learning. arXiv preprint arXiv:2206.08312, 2022. 2, 5, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.221, + 0.894, + 0.29 + ], + "angle": 0, + "content": "[16] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16867-16876, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.291, + 0.893, + 0.36 + ], + "angle": 0, + "content": "[17] Ziyang Chen, David F Fouhey, and Andrew Owens. Sound localization by self-supervised time delay estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVI, pages 489-508. Springer, 2022. 2, 4, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.361, + 0.893, + 0.403 + ], + "angle": 0, + "content": "[18] Ziyang Chen, Xixi Hu, and Andrew Owens. Structure from silence: Learning scene structure from ambient sound. arXiv preprint arXiv:2111.05846, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.405, + 0.893, + 0.446 + ], + "angle": 0, + "content": "[19] Jesper Haahr Christensen, Sascha Hornauer, and Stella Yu. Batvision with gcc-phat features for better sound to vision predictions. arXiv preprint arXiv:2006.07995, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.447, + 0.893, + 0.489 + ], + "angle": 0, + "content": "[20] Michael Defferrard, Kirell Benzi, Pierre Vandergheynst, and Xavier Bresson. Fma: A dataset for music analysis. arXiv preprint arXiv:1612.01840, 2016. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.49, + 0.893, + 0.545 + ], + "angle": 0, + "content": "[21] Yuexi Du, Ziyang Chen, Justin Salamon, Bryan Russell, and Andrew Owens. Conditional generation of audio from video via foley analogies. Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.547, + 0.893, + 0.588 + ], + "angle": 0, + "content": "[22] Mohamed El Banani, Luya Gao, and Justin Johnson. Unsupervisedr&r: Unsupervised point cloud registration via differentiable rendering. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.59, + 0.893, + 0.617 + ], + "angle": 0, + "content": "[23] Mohamed El Banani and Justin Johnson. Bootstrap your own correspondences. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.619, + 0.894, + 0.66 + ], + "angle": 0, + "content": "[24] Chao Feng, Ziyang Chen, and Andrew Owens. Self-supervised video forensics by audio-visual anomaly detection. arXiv preprint arXiv:2301.01767, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.661, + 0.893, + 0.717 + ], + "angle": 0, + "content": "[25] Tim Fischer, Marco Caversaccio, and Wilhelm Wimmer. A front-back confusion metric in horizontal sound localization: The fbc score. In ACM Symposium on Applied Perception 2020, pages 1-5, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.718, + 0.893, + 0.773 + ], + "angle": 0, + "content": "[26] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.775, + 0.893, + 0.817 + ], + "angle": 0, + "content": "[27] Andrew Francl. Modeling and Evaluating Human Sound Localization in the Natural Environment. PhD thesis, Massachusetts Institute of Technology, 2022. 2, 5, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.818, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[28] Chuang Gan, Deng Huang, Peihao Chen, Joshua B Tenenbaum, and Antonio Torralba. Foley music: Learning to generate music from videos. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 758-775. Springer, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.95, + 0.516, + 0.962 + ], + "angle": 0, + "content": "7905" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.473, + 0.176 + ], + "angle": 0, + "content": "[29] Ruohan Gao, Changan Chen, Ziad Al-Halah, Carl Schissler, and Kristen Grauman. Visualechoes: Spatial image representation learning through echolocation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16, pages 658-676. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.177, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[30] Ruohan Gao and Kristen Grauman. 2.5 d visual sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 324-333, 2019. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.233, + 0.472, + 0.3 + ], + "angle": 0, + "content": "[31] Ruohan Gao and Kristen Grauman. Visualvoice: Audiovisual speech separation with cross-modal consistency. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15490-15500. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.301, + 0.472, + 0.357 + ], + "angle": 0, + "content": "[32] Rishabh Garg, Ruohan Gao, and Kristen Grauman. Geometry-aware multi-task learning for binaural audio generation from video. arXiv preprint arXiv:2111.10882, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.358, + 0.472, + 0.412 + ], + "angle": 0, + "content": "[33] Sanchita Ghose and John Jeffrey Prevost. Auto Foley: Artificial synthesis of synchronized sound tracks for silent videos with deep learning. IEEE Transactions on Multimedia, 23:1895-1907, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.413, + 0.472, + 0.483 + ], + "angle": 0, + "content": "[34] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14950-14962, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.484, + 0.472, + 0.51 + ], + "angle": 0, + "content": "[35] R. I. Hartley. Chirality. International Journal of Computer Vision (IJCV), 1998. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.511, + 0.472, + 0.552 + ], + "angle": 0, + "content": "[36] R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.553, + 0.472, + 0.608 + ], + "angle": 0, + "content": "[37] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In International Conference on Computer Vision (ICCV), 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.609, + 0.472, + 0.65 + ], + "angle": 0, + "content": "[38] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition (CVPR), 2016. 4, 5, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.651, + 0.472, + 0.692 + ], + "angle": 0, + "content": "[39] Xixi Hu, Ziyang Chen, and Andrew Owens. Mix and localize: Localizing sound sources in mixtures. Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.692, + 0.472, + 0.761 + ], + "angle": 0, + "content": "[40] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. arXiv preprint arXiv:2301.12661, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.762, + 0.472, + 0.803 + ], + "angle": 0, + "content": "[41] Wen Chin Huang, Dejan Markovic, Alexander Richard, Israel Dejene Gebru, and Anjali Menon. End-to-end binaural speech synthesis. arXiv preprint arXiv:2207.03697, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.804, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[42] Vladimir Iashin and Esa Rahtu. Taming visually guided sound generation. arXiv preprint arXiv:2110.08791, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.845, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[43] Linyi Jin, Shengyi Qian, Andrew Owens, and David F Fouhey. Planar surface reconstruction from sparse views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12991-13000, 2021. 2, 3, 4, 5, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.092, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.092, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[44] Linyi Jin, Jianming Zhang, Yannick Hold-Geoffroy, Oliver Wang, Kevin Matzen, Matthew Sticha, and David F Fouhey. Perspective fields for single image camera calibration. arXiv, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.149, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[45] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In ICCV, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.192, + 0.894, + 0.234 + ], + "angle": 0, + "content": "[46] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representation, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.235, + 0.894, + 0.29 + ], + "angle": 0, + "content": "[47] Charles Knapp and Grifford Carter. The generalized correlation method for estimation of time delay. IEEE transactions on acoustics, speech, and signal processing, 24(4):320-327, 1976. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.291, + 0.894, + 0.333 + ], + "angle": 0, + "content": "[48] Takashi Konno, Kenji Nishida, Katsutoshi Itoyama, and Kazuhiro Nakadai. Audio-visual sfm towards 4d reconstruction under dynamic scenes. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.334, + 0.894, + 0.39 + ], + "angle": 0, + "content": "[49] Felix Kreuk, Gabriel Synnaeve, Adam Polyak, Uriel Singer, Alexandre Défossez, Jade Copet, Devi Parikh, Yaniv Taigman, and Yossi Adi. Audiogen: Textually guided audio generation. arXiv preprint arXiv:2209.15352, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.391, + 0.894, + 0.446 + ], + "angle": 0, + "content": "[50] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Av-nerf: Learning neural fields for real-world audio-visual scene synthesis. arXiv preprint arXiv:2302.02088, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.448, + 0.894, + 0.515 + ], + "angle": 0, + "content": "[51] Yan-Bo Lin and Yu-Chiang Frank Wang. Exploiting audiovisual consistency with partial supervision for spatial audio generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2056–2063, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.517, + 0.894, + 0.56 + ], + "angle": 0, + "content": "[52] Zhi-Hao Lin, Wei-Chiu Ma, Hao-Yu Hsu, Yu-Chiang Frank Wang, and Shenlong Wang. Neurmips: Neural mixture of planar experts for view synthesis. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.561, + 0.894, + 0.589 + ], + "angle": 0, + "content": "[53] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.59, + 0.894, + 0.631 + ], + "angle": 0, + "content": "[54] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.632, + 0.894, + 0.687 + ], + "angle": 0, + "content": "[55] Wei-Chiu Ma, Anqi Joyce Yang, Shenlong Wang, Raquel Urtasun, and Antonio Torralba. Virtual correspondence: Humans as a cue for extreme-view geometry. In CVPR, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.689, + 0.894, + 0.746 + ], + "angle": 0, + "content": "[56] Sagnik Majumder, Ziad Al-Halah, and Kristen Grauman. Move2hear: Active audio-visual source separation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 275–285, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.746, + 0.894, + 0.8 + ], + "angle": 0, + "content": "[57] Sagnik Majumder, Changan Chen, Ziad Al-Halah, and Kristen Grauman. Few-shot audio-visual learning of environment acoustics. arXiv preprint arXiv:2206.04006, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.802, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[58] Sagnik Majumder and Kristen Grauman. Active audio-visual separation of dynamic sound sources. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXIX, pages 551–569. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.872, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[59] Iaroslav Melekhov, Juha Ylioinas, Juho Kannala, and Esa Rahtu. Relative camera pose estimation using convolutional" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.092, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.95, + 0.517, + 0.962 + ], + "angle": 0, + "content": "7906" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.472, + 0.148 + ], + "angle": 0, + "content": "neural networks. In Advanced Concepts for Intelligent Vision Systems: 18th International Conference, ACIVS 2017, Antwerp, Belgium, September 18-21, 2017, Proceedings 18, pages 675-687. Springer, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.15, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[60] Himangi Mittal, Pedro Morgado, Unnat Jain, and Abhinav Gupta. Learning state-aware visual representations from audible interactions. arXiv preprint arXiv:2209.13583, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.207, + 0.472, + 0.248 + ], + "angle": 0, + "content": "[61] Shentong Mo and Pedro Morgado. A closer look at weakly-supervised audio-visual source localization. arXiv preprint arXiv:2209.09634, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.25, + 0.472, + 0.317 + ], + "angle": 0, + "content": "[62] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 218-234. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.32, + 0.472, + 0.375 + ], + "angle": 0, + "content": "[63] Pedro Morgado, Yi Li, and Nuno Nvasconcelos. Learning representations from audio-visual spatial alignment. Advances in Neural Information Processing Systems, 33:4733-4744, 2020. 2, 5, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.377, + 0.472, + 0.418 + ], + "angle": 0, + "content": "[64] Pedro Morgado, Nuno Vasconcelos, Timothy Langlois, and Oliver Wang. Self-supervised generation of spatial audio for 360 video. arXiv preprint arXiv:1809.02587, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.42, + 0.472, + 0.488 + ], + "angle": 0, + "content": "[65] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12475-12486, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.491, + 0.472, + 0.533 + ], + "angle": 0, + "content": "[66] David Nister. An efficient solution to the five-point relative pose problem. IEEE transactions on pattern analysis and machine intelligence, 26(6):756-770, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.535, + 0.472, + 0.576 + ], + "angle": 0, + "content": "[67] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. European Conference on Computer Vision (ECCV), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.577, + 0.472, + 0.632 + ], + "angle": 0, + "content": "[68] Andrew Owens, Phillip Isola, Josh McDermott, Antonio Torralba, Edward H Adelson, and William T Freeman. Visually indicated sounds. In Computer Vision and Pattern Recognition (CVPR), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.634, + 0.472, + 0.702 + ], + "angle": 0, + "content": "[69] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE, 2015. 5, 7, 8, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.704, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[70] Kranti Kumar Parida, Siddharth Srivastava, and Gaurav Sharma. Beyond image to depth: Improving depth prediction using echoes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8268-8277, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.775, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[71] Senthil Purushwalkam, Sebastia Vicenc Amengual Gari, Vamsi Krishna Ithapu, Carl Schissler, Philip Robinson, Abhinav Gupta, and Kristen Grauman. Audio-visual floorplan reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1183–1192, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.859, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[72] Shengyi Qian, Linyi Jin, and David F Fouhey. Associative3d: Volumetric reconstruction from sparse views. In ECCV, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.092, + 0.894, + 0.148 + ], + "angle": 0, + "content": "[73] Shengyi Qian, Alexander Kirillov, Nikhila Ravi, Devendra Singh Chaplot, Justin Johnson, David F Fouhey, and Georgia Gkioxari. Recognizing scenes from novel viewpoints. arXiv preprint arXiv:2112.01520, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.15, + 0.894, + 0.219 + ], + "angle": 0, + "content": "[74] Kranthi Kumar Rachavarapu, Vignesh Sundaresha, AN Rajagopalan, et al. Localize to binauralize: Audio spatialization from visual sound source localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1930-1939, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.221, + 0.894, + 0.304 + ], + "angle": 0, + "content": "[75] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.305, + 0.894, + 0.416 + ], + "angle": 0, + "content": "[76] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 5, 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.418, + 0.894, + 0.472 + ], + "angle": 0, + "content": "[77] Lord Rayleigh. Xii. on our perception of sound direction. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 13(74):214-232, 1907. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.475, + 0.894, + 0.544 + ], + "angle": 0, + "content": "[78] Alexander Richard, Dejan Markovic, Israel D Gebru, Steven Krenn, Gladstone Alexander Butler, Fernando Torre, and Yaser Sheikh. Neural synthesis of binaural speech from mono audio. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.546, + 0.894, + 0.615 + ], + "angle": 0, + "content": "[79] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.617, + 0.894, + 0.659 + ], + "angle": 0, + "content": "[80] Lord Rayleigh O.M. Pres. R.S. Xii. on our perception of sound direction. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 1907. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.66, + 0.894, + 0.73 + ], + "angle": 0, + "content": "[81] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4938–4947, 2020. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.732, + 0.894, + 0.773 + ], + "angle": 0, + "content": "[82] Ralph Schmidt. Multiple emitter location and signal parameter estimation. IEEE transactions on antennas and propagation, 34(3):276-280, 1986. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.775, + 0.894, + 0.829 + ], + "angle": 0, + "content": "[83] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.831, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[84] Nikhil Singh, Jeff Mentch, Jerry Ng, Matthew Beveridge, and Iddo Drori. Image2reverb: Cross-modal reverb impulse response synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 286-295, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.092, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.95, + 0.516, + 0.962 + ], + "angle": 0, + "content": "7907" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[85] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.163, + 0.473, + 0.246 + ], + "angle": 0, + "content": "[86] Ruijie Tao, Zexu Pan, Rohan Kumar Das, Xinyuan Qian, Mike Zheng Shou, and Haizhou Li. Is someone speaking? exploring long-term temporal features for audio-visual active speaker detection. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3927–3935, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.248, + 0.473, + 0.332 + ], + "angle": 0, + "content": "[87] Efthymios Tzinis, Scott Wisdom, Tal Remez, and John R Hershey. Audioscopev2: Audio-visual attention architectures for calibrated open-domain on-screen sound separation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXVII, pages 368–385. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.334, + 0.473, + 0.389 + ], + "angle": 0, + "content": "[88] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.39, + 0.472, + 0.459 + ], + "angle": 0, + "content": "[89] Paolo Vecchiotti, Ning Ma, Stefano Squartini, and Guy J Brown. End-to-end binaural sound localisation from the raw waveform. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 451-455. IEEE, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.461, + 0.472, + 0.502 + ], + "angle": 0, + "content": "[90] DeLiang Wang and Guy J Brown. Computational auditory scene analysis: Principles, algorithms, and applications. Wiley-IEEE press, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.504, + 0.472, + 0.559 + ], + "angle": 0, + "content": "[91] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. In ECCV, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.561, + 0.472, + 0.615 + ], + "angle": 0, + "content": "[92] Fanyi Xiao, Yong Jae Lee, Kristen Grauman, Jitendra Malik, and Christoph Feichtenhofer. Audiovisual slowfast networks for video recognition. arXiv preprint arXiv:2001.08740, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.617, + 0.472, + 0.687 + ], + "angle": 0, + "content": "[93] Xudong Xu, Hang Zhou, Ziwei Liu, Bo Dai, Xiaogang Wang, and Dahua Lin. Visually informed binaural audio generation without binaural audios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15485-15494, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.689, + 0.472, + 0.73 + ], + "angle": 0, + "content": "[94] Nelson Yalta, Kazuhiro Nakadai, and Tetsuya Ogata. Sound source localization using deep learning models. Journal of Robotics and Mechatronics, 29(1):37-48, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.732, + 0.472, + 0.786 + ], + "angle": 0, + "content": "[95] Dongchao Yang, Jianwei Yu, Helin Wang, Wen Wang, Chao Weng, Yuexian Zou, and Dong Yu. Diffsound: Discrete diffusion model for text-to-sound generation. arXiv preprint arXiv:2207.09983, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.789, + 0.472, + 0.87 + ], + "angle": 0, + "content": "[96] Karren Yang, Michael Firman, Eric Brachmann, and Clément Godard. Camera pose estimation and localization with active audio sensing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 271-291. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.472, + 0.903 + ], + "angle": 0, + "content": "[97] Karren Yang, Bryan Russell, and Justin Salamon. Telling left from right: Learning spatial correspondence of sight" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.092, + 0.473, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9932-9941, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.136, + 0.895, + 0.19 + ], + "angle": 0, + "content": "[98] Zhenpei Yang, Jeffrey Z Pan, Linjie Luo, Xiaowei Zhou, Kristen Grauman, and Qixing Huang. Extreme relative pose estimation for rgb-d scans via scene completion. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.192, + 0.895, + 0.233 + ], + "angle": 0, + "content": "[99] Zhenpei Yang, Siming Yan, and Qixing Huang. Extreme relative pose network under hybrid representations. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.895, + 0.304 + ], + "angle": 0, + "content": "[100] Dong Yu, Morten Kolbæk, Zheng-Hua Tan, and Jesper Jensen. Permutation invariant training of deep models for speaker-independent multi-talker speech separation. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 241-245. IEEE, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.895, + 0.361 + ], + "angle": 0, + "content": "[101] Yinda Zhang, Shuran Song, Ersin Yumer, Manolis Savva, Joon-Young Lee, Hailin Jin, and Thomas Funkhouser. Physically-based rendering for indoor scene understanding using convolutional neural networks. In CVPR, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.895, + 0.416 + ], + "angle": 0, + "content": "[102] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In ICCV, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.895, + 0.487 + ], + "angle": 0, + "content": "[103] Hang Zhou, Xudong Xu, Dahua Lin, Xiaogang Wang, and Ziwei Liu. Sep-stereo: Visually guided stereophonic audio generation by associating source separation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.489, + 0.894, + 0.53 + ], + "angle": 0, + "content": "[104] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, 2017. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.532, + 0.895, + 0.585 + ], + "angle": 0, + "content": "[105] Yipin Zhou and Ser-Nam Lim. Joint audio-visual deepfake detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14800–14809, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.588, + 0.895, + 0.672 + ], + "angle": 0, + "content": "[106] Yuliang Zou, Pan Ji, Quoc-Huy Tran, Jia-Bin Huang, and Manmohan Chandraker. Learning monocular visual odometry via self-supervised long-term modeling. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV, pages 710-727. Springer, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.95, + 0.517, + 0.962 + ], + "angle": 0, + "content": "7908" + } + ] +] \ No newline at end of file diff --git a/2023/Sound Localization from Motion_ Jointly Learning Sound Direction and Camera Rotation/ebfa790a-8a91-4c3c-9e52-be53fe87e14f_origin.pdf b/2023/Sound Localization from Motion_ Jointly Learning Sound Direction and Camera Rotation/ebfa790a-8a91-4c3c-9e52-be53fe87e14f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3c45da69295942c285fd470556064e564254a4c3 --- /dev/null +++ b/2023/Sound Localization from Motion_ Jointly Learning Sound Direction and Camera Rotation/ebfa790a-8a91-4c3c-9e52-be53fe87e14f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:166f456bd64bf03c33103371fb5d757969c9cb09f6d5462446f23249cb85ebbb +size 5920944 diff --git a/2023/Sound Localization from Motion_ Jointly Learning Sound Direction and Camera Rotation/full.md b/2023/Sound Localization from Motion_ Jointly Learning Sound Direction and Camera Rotation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c92585942aae547bad96b3eb247146225ce4c669 --- /dev/null +++ b/2023/Sound Localization from Motion_ Jointly Learning Sound Direction and Camera Rotation/full.md @@ -0,0 +1,390 @@ +# Sound Localization from Motion: Jointly Learning Sound Direction and Camera Rotation + +Ziyang Chen + +Shengyi Qian + +Andrew Owens + +University of Michigan + +# Abstract + +The images and sounds that we perceive undergo subtle but geometrically consistent changes as we rotate our heads. In this paper, we use these cues to solve a problem we call Sound Localization from Motion (SLfM): jointly estimating camera rotation and localizing sound sources. We learn to solve these tasks solely through self-supervision. A visual model predicts camera rotation from a pair of images, while an audio model predicts the direction of sound sources from binaural sounds. We train these models to generate predictions that agree with one another. At test time, the models can be deployed independently. To obtain a feature representation that is well-suited to solving this challenging problem, we also propose a method for learning an audio-visual representation through cross-view binauralization: estimating binaural sound from one view, given images and sound from another. Our model can successfully estimate accurate rotations on both real and synthetic scenes, and localize sound sources with accuracy competitive with state-of-the-art self-supervised approaches. Project site: https://ificl.github.io/SLfM. + +# 1. Introduction + +As you rotate your head, the images and sounds that you perceive change in geometrically consistent ways. For example, after turning to the right, a sound source that was directly in front of you will become louder in your left ear and quieter in your right, while simultaneously the visual scene will move right-to-left across your visual field (Fig. 1). + +We hypothesize that these co-occurring audio and visual signals provide "free" supervision that captures geometry, including the motion made by a camera and the direction of sound sources. These are each core problems in machine perception, but are largely studied separately, often using supervised methods that rely on difficult-to-acquire labeled training data, such as annotated sound directions. We take inspiration from self-supervised approaches to structure from motion [104], which learn to estimate 3D structure and camera pose by solving both tasks simultaneously. + +![](images/db541dd8a9a951af38cf10e58cc73dfe15c211eb8e2078d40fec774a8fb1bdfc.jpg) +Figure 1: Images and sounds change in geometrically consistent ways. For example, when we rotate to the right, a sound source that is initially in front of us becomes louder in our left ear. We use these cues to jointly train models for two tasks: localizing sounds from binaural audio and estimating camera rotation from images. The two models are trained entirely through self-supervision, by learning to produce outputs that agree with one other. + +Analogously, we propose a problem we call sound localization from motion (SLfM): jointly estimating camera rotation from images and the sound direction from binaural audio. By solving both tasks simultaneously, we avoid the need for labeled training data. Our models provide each other with self-supervision: a visual model predicts the rotation angle between pairs of images, while an audio model predicts the azimuth of sound sources. We force their predictions to agree with one another, such that changes in rotation are consistent with changes in sound direction and binaural cues. After training, the models can be deployed independently, without multimodal data at test time. + +This is a challenging task that requires perceiving motion in images and binaural cues in audio. Our second contribution is a method for learning representations that are well-suited to this task through cross-view binauralization. We train a network to convert mono to binaural sound for one viewpoint, given an audio-visual pair sampled from another + +viewpoint. Since the sound source is not necessarily visible in the images, the only way to successfully solve this pretext task is by analyzing the changes in the camera pose and predicting how they affect the sound direction. + +All components of our model are entirely self-supervised and are trained solely on unlabeled audio-visual data. Our results suggest that paired audio-visual data provides a useful and complementary signal for learning about geometry. In contrast to other audio or visual self-supervised pose estimation methods, we obtain supervision from abundantly available audio data, thus avoiding the need of 3D ground truth or correspondences between pixels [104, 106] or audio samples [17]. Through experiments, we show: + +- Paired audio-visual data provides a supervisory signal for pose estimation tasks. +- We obtain competitive performance with state-of-the-art self-supervised sound localization methods [17]. +- We obtain strong rotation estimation performance, and our model generalizes to Stanford2D3D [6] dataset, where it is competitive with classic sparse feature matching methods. +- The features we learn through our pretext task outperform other representations for our downstream tasks. + +# 2. Related Work + +Audio for spatial perception. Recent works have explored the use of sound for spatial understanding. Purushwalkam et al. [71] reconstructed floor plans in simulated environments [12]. Chen et al. [18] used ambient sounds from environments to learn about scene structures. Konno et al. [48] integrated sound localization to visual SfM while do not jointly learn them. Other work learns representations for spatial audio-visual tasks. Yang et al. [97] predicted whether stereo channels are swapped in a video, and Morgado et al. [63] solved a spatial alignment task. The learned representations are then used to improve localization, up-mixing, and segmentation models. In contrast, we learn camera pose and sound localization solely from self-supervision, obtaining angular predictions without labeled data. Other work uses echolocation sounds to learn representations [29, 96] and predict depth maps [19, 70] and estimate camera poses [96] using labeled data. In contrast, our proposed approach jointly learns binaural sound localization and camera pose through passive audio sensing, without supervision. + +Acoustic synthesis and spatialization. Researchers have explored visually-guided sound synthesis [28, 33, 42, 21] and text-guided audio synthesis [49, 95, 40]. Additionally, researchers have investigated generating realistic environmental acoustics using visual information [11, 84, 15, 57]. Chen et al. [14] introduced the novel-view acoustic synthesis task, which synthesizes binaural sound at the target view using audio and visual information from a source view. Liang et al. [50] proposed an audio-visual neural field in + +real-world audio-visual scenes. Many recent works have proposed to generate spatial audio from mono audio using visual cues [64, 30, 74, 93, 51, 103, 32], or the relative pose between sound sources and the receiver [78, 41]. Inspired by these works, our feature learning approach learns spatial representations through an audio prediction task. + +Binaural sound localization. Humans have the ability to localize sound sources from binaural sound [80]. Traditional approaches estimate interaural time delays via cross-correlation using hand-crafted features [47], factorization methods [82], or loudness differences between ears [77, 90]. Chen et al. [17] adapted methods from self-supervised visual tracking to the problem of binaural sound localization. Similarly, we estimate direction through self-supervision. However, we obtain our supervision through cross-modal supervision from vision instead of from correspondence cues. Moreover, we also obtain visual camera rotation estimation through our learning process. Francl [27] learned representations of sound location with a contrastive loss where positive and negative examples are selected based on the extent of head movements. Other works have used supervised learning techniques with labeled data to localize sound sources in reverberant environments [1, 89, 94]. Unlike these methods, our model learns 3D sound localization without labels. + +Camera pose estimation. Traditional methods for camera pose estimation are based on finding correspondences between images and then solving an optimization problem using constraints from multi-view geometry [36]. These include structure from motion methods that estimate full pose [83] and camera rotation [8]. Recent methods have directly predicted camera pose using neural networks, including methods that use photos [72, 45, 59, 44, 55] or RGB-D scans [99, 98, 22, 23]. Our setup is similar to work that learns relative camera poses from sparse views [43, 9], and we use their network architectures. However, we learn the camera pose through cross-modal supervision from audio, rather than from labels. Our work is also closely related to methods that learn structure from motion through self-supervision [104, 106], such as by jointly learning models that perform depth and camera pose estimation with photoconsistency constraints. In contrast, our visual model's learning signal comes solely from audio-based supervision, and we jointly learn audio localization. + +Audio-visual learning. A number of works have focused on learning multimodal representations for audio and vision, taking into account semantic correspondence and temporal synchronization [67, 92, 7, 65, 68, 2, 60]. Other approaches study audio-visual sound localization [39, 61, 16, 62], source separation [56, 31, 58, 87], active speaker detection [3, 86, 5], navigation [12, 13, 10] and forensics [105, 34, 24]. Our focus, in contrast, is on utilizing multi-view audio-visual signals to learn geometry. + +![](images/0af377614fad8cce4404f7daff0ffe16d7c754f65cb725801f3cd0a6a852a88d.jpg) +(a) Cross-view audio prediction + +![](images/7aec18f353123bcdd088dfa7097260c35e09a003197f9dadb80260e5aa046005.jpg) +(b) Sound localization from motion +Figure 2: Method overview. (a) We learn a feature representation by predicting how changes in images lead to changes in sound in a cross-view binauralization pretext task. We convert mono sound to binaural sound at a target viewpoint, after conditioning the model on observations from a source viewpoint. (b) We use the representation to jointly solve two pose estimation tasks: visual rotation estimation and binaural sound localization. We train the visual rotation angle, $\phi_{s,t}$ , to be consistent with the difference in predicted sound angles $\theta_s$ and $\theta_t$ . + +# 3. Sound Localization from Motion + +We address the sound localization from motion (SLfM) task: predicting the azimuth of a sound source from binaural audio and camera rotation from two images. First, we present a self-supervised representation that can be used to solve this downstream task. Then, we show how the representation can be used to solve the task. + +# 3.1. Learning representation via spatialization + +We learn an audio-visual representation that conveys spatial information by solving a cross-view binauralization task: converting mono sound to stereo for one viewpoint, given an audio-visual pair sampled from another viewpoint (Fig. 2a). In order to successfully solve the task, a model must implicitly estimate the sound direction in the source view and predict how the change in viewpoint will affect the sound in the target view. A key difference between this task and traditional, single-view binauralization [30] is that the sound source need not be visible in any of the images. Hence, the task cannot be solved from the target audio-visual pair alone. + +We binauralize the sound $\mathbf{a}_t$ at the target view through an audio predictor $\mathcal{F}_{\theta}$ . To make our representation suitable for downstream tasks, we factorize it into visual features $f_{v}(\mathbf{v}_{s},\mathbf{v}_{t})$ , which are intended to create features relevant to relative pose, and audio features $f_{a}(\mathbf{a}_{s})$ , which capture sound localization cues. We predict the binauralized audio $\hat{\mathbf{a}}_t$ at the target view from mono audio $\bar{\mathbf{a}}_t$ , the visual change of $(\mathbf{v}_s,\mathbf{v}_t)$ and binaural sound $\mathbf{a}_s$ heard at the source viewpoint: + +$$ +\hat {\mathbf {a}} _ {t} = \mathcal {F} _ {\theta} \left(\bar {\mathbf {a}} _ {t}, f _ {v} \left(\mathbf {v} _ {s}, \mathbf {v} _ {t}\right), f _ {a} \left(\mathbf {a} _ {s}\right)\right). \tag {1} +$$ + +We represent audio $\mathbf{a}$ as a spectrogram $\mathbf{A}$ using short-time Fourier transform (STFT). Following Gao et al. [30], given the mix of stereo audio $\bar{\mathbf{A}}_t = \mathrm{STFT}(\mathbf{a}_t^L + \mathbf{a}_t^R)$ , we predict the difference of two channels $\mathbf{A}_t = \mathrm{STFT}(\mathbf{a}_t^L - \mathbf{a}_t^R)$ . We optimize the $L1$ loss between predicted spectrogram $\hat{\mathbf{A}}_t$ and ground-truth spectrogram $\mathbf{A}_t$ : + +$$ +\mathcal {L} _ {\text {p r e t e x t}} = \left| \left| \hat {\mathbf {A}} _ {t} - \mathbf {A} _ {t} \right| \right| _ {1}. \tag {2} +$$ + +Multi-view binauralization. Following Zhou et al. [104], we improve our representations by binauralizing sounds at $N$ different target viewpoints, using observations from a single source viewpoint $s$ . We hypothesize that jointly solving spatialization problems from a single viewpoint for multiple target viewpoints would require the model to make more accurate predictions of sound source locations, thereby improving the estimation of view changes: + +$$ +\mathcal {L} _ {\text {p r e t e x t}} = \frac {1}{N} \sum_ {i} ^ {N} \left| \left| \mathcal {F} _ {\theta} \left(\bar {\mathbf {A}} _ {i}, f _ {v} \left(\mathbf {v} _ {s}, \mathbf {v} _ {i}\right), f _ {a} \left(\mathbf {a} _ {s}\right)\right) - \mathbf {A} _ {i} \right| \right| _ {1}. \tag {3} +$$ + +# 3.2. Estimating pose and localizing sounds + +We now address the problem of learning models for sound localization and pose estimation, using our self-supervised audio-visual features. Given two views, we predict sound directions and relative rotation. We train the model to make these two predictions consistent with one another, while using simple binaural constraints to resolve ambiguities. + +We are given images $\mathbf{v}_s$ and $\mathbf{v}_t$ (rotated views recorded at the same position) and learned visual embedding $f_v$ . We predict the scalar rotation angle $\phi_{s,t}$ via the encoder $g_v$ : + +$$ +\phi_ {s, t} = g _ {v} \left(f _ {v} \left(\mathbf {v} _ {s}, \mathbf {v} _ {t}\right)\right), R _ {s, t} = \left[ \begin{array}{l l} \cos \phi_ {s, t} & - \sin \phi_ {s, t} \\ \sin \phi_ {s, t} & \cos \phi_ {s, t} \end{array} \right], \tag {4} +$$ + +where $R_{s,t}$ is 2D rotation matrix of $\phi_{s,t}$ . Following common practice in indoor scene reconstruction, we give the camera a fixed downward tilt [102, 73, 101] and only estimate azimuth [43]. This is also a common assumption in audio localization [1, 89], since azimuth has strong binaural cues. + +Similarly, we predict the azimuths of the sound sources using audio features and the encoder $g_{a}$ : + +$$ +\theta_ {i} = g _ {a} \left(f _ {a} (\mathbf {a} _ {i})\right), \quad \mathbf {r} _ {i} = \left[ \begin{array}{l l} \cos \theta_ {i} & \sin \theta_ {i} \end{array} \right] ^ {\top}, \tag {5} +$$ + +where we represent the azimuth as a vector $\mathbf{r}_i$ + +Cross-modal geometric consistency. When the camera is rotated, the sound source ought to rotate in the opposite direction (Fig. 2b). For example, a $30^{\circ}$ clockwise camera rotation should result in a $30^{\circ}$ counterclockwise rotation in sound direction. Such a constraint could be converted into a loss: + +$$ +\mathcal {L} _ {\text {r o t}} = \left\| \mathbf {r} _ {s} - R _ {s, t} \mathbf {r} _ {t} \right\| ^ {2}. \tag {6} +$$ + +However, a well-known ambiguity called front-back confusion [77, 25] exists in binaural sound perception: one cannot generally tell whether a sound is in front of the view, or behind them. To address that, we use permutation invariant training [100] (PIT), and allow the model to use either the predicted sound direction or its reflection about the $x$ axis without penalty. This results in the loss: + +$$ +\mathcal {L} _ {\mathrm {g e o}} = \min _ {\substack {\hat {\mathbf {r}} _ {s} \in \{\mathbf {r} _ {s}, Q \mathbf {r} _ {s} \} \\ \hat {\mathbf {r}} _ {t} \in \{\mathbf {r} _ {t}, Q \mathbf {r} _ {t} \}}} \| \hat {\mathbf {r}} _ {s} - R _ {s, t} \hat {\mathbf {r}} _ {t} \| ^ {2}, \tag{7} +$$ + +where $Q = \begin{bmatrix} 1 & 0 \\ 0 & -1 \end{bmatrix}$ reflects the sound direction. + +As a consequence of this ambiguity, there are also two possible solutions for the visual rotation model, since the visual rotation matrices can be mirrored about the $x$ axis. For example, one can create a solution with equal loss by multiplying the rotations and sound directions by $Q$ . We discuss this ambiguity in more depth in Sec. 4.4. + +Incorporating binaural observations. Without additional constraints, the solution is ambiguous, and may collapse into a trivial solution (e.g., predicting zero for all three angles). To avoid this, we force the model to agree with a simple binaural cue based on interaural intensity difference (IID). We predict whether the sound is to the left or right of the viewer, based on whether it is louder in the left or right microphone: $d = \mathrm{sign}(\log \left|\frac{\mathbf{A}^L}{\mathbf{A}^R}\right|)$ , where $|\mathbf{A}|$ is the magnitude of the spectrogram A. We perform this left/right test at each timestep in the spectrogram and then pool via majority voting (see supplementary for details). We penalize predictions that are inconsistent with these "left or right" observations: + +$$ +\mathcal {L} _ {\text {b i n a u r a l}} = \mathcal {L} _ {\text {B C E}} (\sin \theta_ {i}, d _ {i}), \tag {8} +$$ + +where $\mathcal{L}_{\mathrm{BCE}}$ is binary cross entropy loss. + +Encouraging symmetry. To help regularize the model, we also add symmetry constraints. For sound localization, swapping the left and right channels of the audio ought to result in a prediction in the opposite direction since the binaural cues are reversed. For rotation estimation, the relative pose between images $s$ and $t$ should invert the pose from $t$ to $s$ . We encourage both constraints via a loss: + +$$ +\mathcal {L} _ {\text {s y m}} = \left| \theta + \theta_ {\text {f l i p}} \right| + \left| \phi_ {s, t} + \phi_ {t, s} \right|, \tag {9} +$$ + +![](images/99ddbdc91cdffb8b78ffbf33341c969bb81d918b52309982c38bbb2560a362de.jpg) +Figure 3: Cross-view binauralization architecture. We take a mono spectrogram as input and fuse the audio and visual features from the audio and visual encoder respectively to synthesize the binaural spectrogram at the target viewpoint. + +where $\theta_{\mathrm{flip}}$ is the prediction of sound angle $\theta$ using audio with swapped audio channels, and $\phi_{s,t}$ and $\phi_{t,s}$ are the predicted rotations between cameras $s$ and $t$ . + +Overall loss. We combine these constraints to obtain an overall loss: + +$$ +\mathcal {L} = \lambda \mathcal {L} _ {\mathrm {g e o}} + \mathcal {L} _ {\text {b i n a u r a l}} + \mathcal {L} _ {\text {s y m}}, \tag {10} +$$ + +where $\lambda$ is the weight for the geometric loss. + +# 4. Experiments + +We have introduced a self-supervised method to learn camera pose and sound localization from audio-visual data. In experiments, we first evaluate how well our learned representation captures spatial information. We then evaluate how well our method learns camera pose and sound localization by comparing it with baselines. Finally, we show generalization to indoor panorama images Stanford2D3D [6] and in-the-wild binaural audio [17]. + +# 4.1. Implementations + +Visual pose encoder. We follow recent pose estimation work [9, 43] and build a Siamese-style visual pose network $f_{v}$ with ResNet-18 [38] as the backbone. We compute dense 4D correlation volumes between the features from the third residual layer and then encode them by convolution layers. We resize images to $320 \times 240$ and encode a pair of images into 512-d features. We use an MLP $g_{v}$ to map visual features to 1-d logits for our SLfM models. + +Binaural audio encoder. We obtain binaural audio embeddings $f_{a}(\cdot)$ using ResNet-18 [38] that operates on spectrograms. We covert the two-channel waveform of length $L$ to a spectrogram representation of size $256 \times 256 \times 4$ using short-time Fourier transform (STFT), where we keep both the magnitude and phase of spectrograms. We extract 512-d features of binaural sound with $f_{a}$ and map them to 1-d logits using an MLP $g_{a}$ . + +Audio prediction model. We adopt the light-weighted audio-visual U-Net [30] to perform binauralization. We feed in spectrograms of size $256 \times 256 \times 2$ and predict the target spectrograms. We concatenate the visual pose features $f_{v}(\cdot)$ and audio features $f_{a}(\cdot)$ at the bottleneck of U-Net. We show the architecture of our models in Fig. 3. Please see supp. for more implementation details. + +# 4.2. Dataset + +Since there is no public multi-view audio-visual dataset with camera poses and sound direction ground truth, we use the SoundSpaces 2.0 platform [15] to create a dataset. Our 3D scenes come from Habitat-Matterport 3D dataset (HM3D) [76], which is a large dataset of real 3D scenes. This setup allows us to have photorealistic images and high-quality spatial audio with real-world acoustics phenomenon (e.g., reverberation), as well as providing the ground-truth camera pose and sound directions that can be used for evaluation. We call this dataset HM3D-SS. + +We generate binaural Room Impulse Responses (RIRs) and images with a $60^{\circ}$ field of view, using 100 scenes of HM3D [76]. For each audio-visual example, we randomly place sound sources in the scene with a height range of (0.7, 1.7) meters, and sample 4 different rotated viewpoints at one location within 4 meters. The rotations are limited to $(10^{\circ}, 90^{\circ})$ relative to the source viewpoints. We follow the standard practice to set the height to agents to be $1.5\mathrm{m}$ and lock a downward tilt angle [43, 101, 73, 102]. We render the binaural RIRs and images given the position of agents and sound sources. We obtain binaural audio by convolving binaural RIRs with mono audio samples from LibriSpeech [69] and Free Music Archive [20]. To ensure that the evaluation tests the model's pose estimation abilities, rather than its ability to visually localize sound sources, the sound sources are not visible on screen. + +We create 50K audio-visual pairs from 200K viewpoints. The audio was rendered with average reverberation of $\mathrm{RT}_{60} = 0.4\mathrm{s}$ (see supp. for details). We divided our data into 81/9/10 scenes for the train/val/test, respectively. + +# 4.3. Evaluating the learned representation + +First, we directly evaluate the quality of our learned features for rotation estimation and sound localization via linear probing with labeled data (rather than learning them jointly through self-supervision). + +Baselines and ablations. We compare our model with several baselines that use alternative pretext tasks: 1) AVSA [63]: it learns spatial cues by training a model to spatially align video and audio clips extracted from different viewing angles. We adapt this model to our dataset and train with 4 different views; 2) RotNCE [27]: it applies contrastive learning on the audio from different angles and uses annotations of the agent's rotation to select positive + +and negative samples, which results in learning audio spatial representation. For baselines, we use the same architecture for feature extractors to ensure fair comparisons. + +To determine if we utilize visual and audio features from different views to solve the binauralization task, we also study some variants of our models: 1) Ours-NoA: we only provide visual features for the binauralization task; 2) Ours-NoV: which only uses audio from the other view to spatialize sounds; 3) Ours-GTRot: we provide ground-truth rotation embedding instead of features from visual frames. + +Besides the mono-to-binaural task, we also experiment with another objective: predicting the right channel from the left channel. We train our L2R model with the same setup as our M2B model. (Please see supp. for pretext results.) + +Downstream tasks. We assess the quality of spatial representations we learned from our pretext tasks in two downstream tasks: relative camera rotation and 3D sound localization. We formulate them as classification problems, where angles are categorized into 64 bins, and we use accuracy as the evaluation metric. To evaluate the learned features, we freeze them and train a linear classifier on the downstream tasks. We compare the performance of our features with those learned from RotNCE [27], AVSA [63], ImageNet [37], and random features, and report the results in Tab. 1. Our approach outperforms the baselines in both tasks, indicating that we learn better spatial representations. Furthermore, our linear probe models show comparable performance against the supervised method which can be regarded as approximate upper bounds for our models, suggesting that our pretext tasks help learn a useful representation. Please see supp. for experiments on FreeMusic [20]. + +Emerging camera pose from audio prompting. To help understand the strong performance of our self-supervised features, we asked whether we could use the cross-view binauralization model alone to estimate camera rotation. Inspired by prompting in vision and language models [75], we + +
ModelAudio Loc. Acc (%) ↑Camera Rot. Acc (%) ↑
LibriSpeechRandom feature4.84.7
ImageNet [38]+Random-56.3
RotNCE [27]50.9-
AVSA [63]71.26.5
Ours-NoA-9.6
Ours-NoV53.9-
Ours-GTRot70.6-
Ours-L2R (3 views)78.576.1
Ours (2 views)74.580.0
Ours (3 views)75.481.3
Supervised81.595.8
+ +Table 1: Downstream task performance on HM3D-SS dataset. We report linear probe performance on the audio localization and camera rotation downstream tasks. + +
ModelAudio angle MAE (°) ↓Camera angle MAE (°) ↓
LibriSpeechChance40.2829.41
SIFT [54]-12.2
Ours w/o Reflect28.0826.80
Ours-Prompt-9.13
Ours-L2R3.220.80
Ours3.170.77
Ours-Front4.483.08
Ours-GTRot1.83-
Superglue [81]-2.47
Supervised1.710.46
FreeMusicChance40.8129.41
SIFT [54]-12.2
Ours w/o Reflect28.2426.81
Ours-L2R3.180.81
Ours3.370.84
Ours-Front3.992.49
Ours-GTRot1.96-
Superglue [81]-2.47
Supervised2.500.46
+ +obtain rough estimates of camera rotation by providing our model with carefully-provided inputs. Given a pair of images $(\mathbf{v}_s,\mathbf{v}_t)$ , we create a synthetic binaural audio prompt, $\mathbf{a}_s$ . We then ask our model to generate the binaural sound $\hat{\mathbf{a}}_t$ for the target viewpoint. By analyzing the IID cues in $\hat{\mathbf{a}}_t$ , we can estimate the model's implicitly predicted camera pose. To do this, we find the nearest neighbor of our generated audio $\hat{\mathbf{a}}_t$ , using a database of synthetically generated audio with known sound directions. Please refer to supp. for more details. Our method achieves the mean absolute error of $9.13^{\circ}$ on HM3D-SS dataset where chance is $29.41^{\circ}$ (Tab. 2). Our approach can also generalize to Stanford2D3D [6] dataset, where we can achieve mean absolute error of $9.93^{\circ}$ (Tab. 3). + +# 4.4. Evaluating SLfM + +We evaluated our SLfM models on the HM3D-SS dataset. We use the mean absolute error of angle in degrees as evaluation metrics. To avoid sound field ambiguity, we filtered out samples with sound angles outside of $(-90^{\circ}, 90^{\circ})$ for the + +![](images/faa58b777188becf848cf49185abfe74a4bf1acc96383789f8f4efdf03ee17b0.jpg) +Figure 4: Qualitative results on real-world examples. We show our predictions on Stanford2D3D [6] and In-the-wild audio [17]. Green denotes accurate predictions. + +![](images/c88cdb70d7fb5ce1c5690c0479f41d2bff5c24a5534a56c9466f58e8fbe82010.jpg) + +Table 2: Sound localization from motion results on HM3D-SS. We evaluate our SLfM models on each modality independently. + +
In the wild audio [17]
ModelAcc (%) ↑
GCC-PHAT [47]77.2
IID [17]75.4
IID (Ours)82.1
MonoCLR [17]87.4
StereoCRW [17]87.2
Ours-L2R84.9
Ours84.0
+ +
Stanford2D3D [6]
ModelRot Err. (°) ↓
MeanMed.
SIFT [54]16.40.06
LoFTR [85]6.101.13
Superglue [81]5.070.07
Ours-Prompt9.939.02
Ours-L2R1.120.71
Ours1.140.67
+ +Table 3: Evaluation of our SLfM models on the real-world data. We evaluate our audio localization model on the In-the-wild audio [17] (left) and our camera pose model on the Stanford2D3D [6] (right). Rot. denote camera rotation. + +evaluation set. When training our SLfM models, we use our best-performing features, i.e., pretext tasks trained with 3 views. We freeze learned audio and visual features and only train multi-layer perceptrons on top of them and evaluate them independently on each modality. + +Baselines and ablations. For relative camera pose estimation, we compare our models with sparse feature matching using SIFT [54] and SuperGlue [81], followed by rotation fitting. We use those methods to detect key points, run Lowe's ratio test, and use RANSAC with five-point algorithm to recover the camera pose [54, 36, 66, 26]. For the sound localization task, we compared with time delay estimation methods: the popular GCC-PHAT [47] and the recent self-supervised method StereoCRW [17]. As far as we know, there are no other baselines that can estimate poses and sound source directions without labels. + +We also compared several variations of our method, including Ours-Front, where we filter out the samples with sound sources behind the viewers to remove binaural ambiguity during training, and an oracle model Ours-GTRot, which uses ground-truth rotation angles instead. + +Results. We show our results in Tab. 2. Our models can predict the azimuths of sound sources, obtaining strong performance without any labels. Without providing reflection invariance (Eq. (7)), the model failed to learn reasonable geometric due to the audio ambiguity. Our self-supervised + +![](images/fb78b25bd41a287098b8e407f3c1ae28ef900d53731b9cde3e9364f7012ddef6.jpg) +Figure 5: Real video. We play music using a speaker and record a video using iPhone with a binaural microphone (different from training examples). We show the predicted camera rotation and sound direction over time. They change smoothly and match our camera motion. Please see our project webpage for video results. + +![](images/ba0393c760b52f00cb2e09a8a47564c6a1937c782a66b1523372d929855d79a7.jpg) +Figure 6: Qualitative results on HM3D-SS. Given two images and their corresponding audio at the source viewpoints, our approach can individually predict the relative camera pose and localize sound locations accurately. We visualize our camera rotation prediction with blue bar and our sound angle prediction with orange bar. To highlight the subtle differences in the waveforms, we color code the amplitude. Please see the project webpage for more video results. + +model achieves comparable performance against our oracle model (Ours-GTRot) and supervised models, indicating we estimate camera poses and localize sound accurately. We show some qualitative results on HM3D-SS in Fig. 6 with LibriSpeech samples [69]. + +Generalization to other datasets. We further demonstrate the generalization ability of our models by experimenting with out-of-distribution, real-world data. We evaluate our camera rotation model on Stanford2D3D [6] with real indoor RGB images (Tab. 3). We obtain image pairs by cropping from panoramas. Although our model is trained on renderings of HM3D [76], it obtains strong generalization ability. Compared with rotation estimation based on SIFT [54] and SuperGlue [81], our model is significantly better on mean rotation error. We also report median rotation error to be consistent with prior works [9, 43]. However, SIFT [54] and Superglue [81] have a very low median error. This is likely due to the all-or-nothing nature of feature matching-based approaches, which either produce highly accurate predictions if the matches are correct (especially for "easy" cases with small amounts of rotation) or else produce gross errors. + +We also evaluate our sound localization model on the in-the-wild binaural audio [17]. We use binary accuracy as the metric for left-or-right direction classification accounting for the fact that microphone baselines are unknown for internet videos. For an apples-to-apples comparison to prior work, we retrained our model using 0.51s length of audio for the binaural audio encoder $f_{a}(\cdot)$ . As shown in Tab. 3, our model obtains similar performance to StereoCRW, a state-of-the-art self-supervised time delay method, suggesting we have a strong capability for sound localization. We show qualitative results in Fig. 4. We also perform both tasks on a self-recorded video (Fig. 5) of a rotating camera and a binaural microphone. We show the mean and standard deviation of predictions in 1.0s windows. + +Handling ambiguity. In binaural sound perception, there is a fundamental ambiguity that whether the sound is in front of or behind us. It leads to multiple solutions with equal loss in our model, mirroring sound sources and negating rotation angles with flipped $z$ axis (Fig. 7). These two solutions differ in that a visual rotation angle either indicates a clockwise or counterclockwise rotation. For evaluation, we convert "backwards" counterclockwise predictions to clockwise predictions by simply providing the model with pairs of input frames, then negating the model's outputs if the angle is the opposite of the expected direction2. + +# 4.5. Generalization to more complex scenarios + +We investigate whether our approach generalizes to more complex scenarios, such as with multiple sound sources or when translation is included in camera motion. + +Multiple sound sources. We evaluate our representations and SLfM models on more complex scenes containing multiple sound sources. We train our models with two source sources placed in the scenes. One of them is dominant and + +![](images/a369f32c41ea1be4885ae2b44e5496a654f5b3c613ab97b8dcdccd68b3d49923.jpg) +Figure 7: Mirror ambiguity. The $+z$ and $+x$ axes represent forward and rightward directions. Both solutions have the same loss. + +![](images/528096eb7f258d146a61014842594bc58c8971bfb3ba6e990a3b0ea38b9a50bb.jpg) +Figure 8: Robustness to reverberation. We study the effect of reverberation on our SLfM model. + +
ModelMulti-source Acc (%) ↑Small trans. Acc (%) ↑
Aud.Rot.Aud.Rot.
Random feature4.54.75.73.8
ImageNet [38]+Random-56.3-23.0
RotNCE [27]35.8---
AVSA [63]55.96.765.46.1
Ours (3 views)59.277.972.249.6
Supervised74.595.881.663.5
+ +Table 4: Generalization to more complex scenarios. We evaluate the audio and visual features learned from more complex scenarios via linear probing. Aud. and Rot. denote audio localization and camera rotation respectively. + +
ModelMulti-source MAE (°) ↓Small trans. MAE (°) ↓
Aud.Rot.Aud.Rot.
Chance40.4629.4140.2829.41
SIFT [54]-12.2-12.2
Ours7.670.716.281.04
Ours-GTRot5.81-4.24-
Superglue [81]-2.47-2.47
Supervised3.600.461.710.46
+ +our sound localization target. We report the results in Tab. 4 and Tab. 5. Our model learns better representations than baselines and achieves accurate predictions of the azimuth of the source and camera pose even in challenging scenarios with multiple sound sources. + +Translation in camera motions. To study how small translations in the camera motion could affect our models, we generate 50K pairs of audio-visual data with both rotation and translation change, limiting the uniformly sampled translation to 0.5 meters. We train our models with LibriSpeech samples [69]. For our linear probing evaluation, we measure the ability of features to handle complex examples by testing on the dataset that has translation. For our SLfM model, we study our ability to learn from noisy data, thus we evaluate it on the rotation-only examples. As the results are shown in Tab. 4 and Tab. 5, we successfully learn useful features and obtain accurate rotation and sound direction predictions despite the presence of translation. Since we jointly learn the audio and visual representations, it can negatively impact the learning of one modality when another one becomes harder. + +# 4.6. Ablation study + +Robustness to reverberation. We study how our SLfM models perform under the different reverberation configurations. We used SoundSpaces [15] to create audio with average reverberation $\mathrm{RT}_{60} \in \{0.1, 0.2, 0.3, 0.4, 0.5\}$ while + +Table 5: SLfM results on more complex scenarios. We evaluate our model on the version of HM3D-SS with two sound sources. We also evaluate our model trained with small translations on rotation-only examples. + +
ModelLossesAudio angle MAE (°) ↓Camera angle MAE (°) ↓
LgeoLbinauralLsym
Ours✓37.6029.20
✓✓37.5229.17
✓✓3.586.99
✓✓✓3.170.77
+ +Table 6: Ablation experiments on our SLfM losses. We evaluate our SLfM models with different combinations of losses. + +keeping the visual signals the same. We train our model on each setting. As shown in Fig. 8, our performance decreases as the level of reverberation increases, where audio becomes more challenging during both training and testing. Please see supp. for the study on representations. + +Losses for SLfM. We study the necessity of our proposed loss functions in Tab. 6. Our models fail to learn accurate pose estimation without binaural loss or symmetric loss. It highlights the crucial role of these losses. + +# 5. Conclusion + +In this paper, we proposed the sound localization from motion (SLfM) problem, and provided a self-supervised method for solving it. We also presented a method for learning audio-visual features that convey sound directions and camera rotation, which we show are well-suited to solving the SLfM task. Despite learning our models solely from unlabeled audio-visual data, we obtain strong performance on a variety of benchmarks, including rotation estimation on the Stanford2D3D [6] dataset and "in the wild" sound direction estimation [17]. Our results suggest that the subtle correlations between sights and binaural sounds that result from rotational motion provide a useful (and previously unused) learning signal. We see our work as opening new directions in self-supervised geometry estimation and feature learning that use sound as a complementary source of supervision. We will release code, data, and models upon acceptance. + +Limitations and Broader Impacts. Our work has several limitations. First, while we evaluate our models on real images and sounds, we train on data from simulators, due to a lack of available relevant data. We note that this is common practice in visual 3D reconstruction [43, 91, 52]. Second, we assume that the 3D scene and sound sources are stationary. Third, we do not evaluate our model on extreme viewpoint changes [55], which requires reasoning about images that have little or no overlap. + +Acknowledgements. We would like to thank Ang Cao, Tiange Luo, and Linyi Jin for their helpful discussion. We thank Changan Chen for his help with SoundSpaces 2.0. This work was funded in part by DARPA Semafor and Sony. The views, opinions and/or findings expressed are those of the authors and should not be interpreted as representing the official views or policies of the Department of Defense or the U.S. Government. + +# References + +[1] Sharath Adavanne, Archontis Politis, and Tuomas Virtanen. Direction of arrival estimation for multiple sound sources using convolutional recurrent neural network. In 2018 26th European Signal Processing Conference (EUSIPCO), pages 1462-1466. IEEE, 2018. 2, 3 +[2] Triantafyllos Afouras, Yuki M Asano, Francois Fagan, Andrea Vedaldi, and Florian Metze. Self-supervised object detection from audio-visual correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10575–10586, 2022. 2 +[3] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. arXiv preprint arXiv:2008.04237, 2020. 2 +[4] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54(10):105-112, 2011. 7 +[5] Juan León Alcazar, Fabian Caba, Ali K Thabet, and Bernard Ghanem. Maas: Multi-modal assignation for active speaker detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 265-274, 2021. 2 +[6] Iro Armeni, Sasha Sax, Amir R Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105, 2017. 2, 4, 6, 7, 8 +[7] Yuki Asano, Mandela Patrick, Christian Rupprecht, and Andrea Vedaldi. Labelling unlabelled videos from scratch with multi-modal self-supervision. Advances in Neural Information Processing Systems, 33:4660-4671, 2020. 2 +[8] Matthew Brown and David G Lowe. Automatic panoramic image stitching using invariant features. International journal of computer vision, 74:59-73, 2007. 2 +[9] Ruojin Cai, Bharath Hariharan, Noah Snavely, and Hadar Averbuch-Elor. Extreme rotation estimation using dense correlation volumes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14566-14575, 2021. 2, 4, 7 +[10] Changan Chen, Ziad Al-Halah, and Kristen Grauman. Semantic audio-visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15516–15525, 2021. 2 +[11] Changan Chen, Ruohan Gao, Paul Calamia, and Kristen Grauman. Visual acoustic matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18858-18868, 2022. 2 +[12] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part VI 16, pages 17–36. Springer, 2020. 2 +[13] Changan Chen, Sagnik Majumder, Ziad Al-Halah, Ruohan Gao, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Learning to set waypoints for audio-visual navigation. arXiv preprint arXiv:2008.09622, 2020. 2 + +[14] Changan Chen, Alexander Richard, Roman Shapovalov, Vamsi Krishna Ithapu, Natalia Neverova, Kristen Grauman, and Andrea Vedaldi. Novel-view acoustic synthesis. arXiv preprint arXiv:2301.08730, 2023. 2 +[15] Changan Chen, Carl Schissler, Sanchit Garg, Philip Kobernik, Alexander Clegg, Paul Calamia, Dhruv Batra, Philip W Robinson, and Kristen Grauman. Soundspaces 2.0: A simulation platform for visual-acoustic learning. arXiv preprint arXiv:2206.08312, 2022. 2, 5, 8, 1 +[16] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16867-16876, 2021. 2 +[17] Ziyang Chen, David F Fouhey, and Andrew Owens. Sound localization by self-supervised time delay estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVI, pages 489-508. Springer, 2022. 2, 4, 6, 7, 8, 1 +[18] Ziyang Chen, Xixi Hu, and Andrew Owens. Structure from silence: Learning scene structure from ambient sound. arXiv preprint arXiv:2111.05846, 2021. 2 +[19] Jesper Haahr Christensen, Sascha Hornauer, and Stella Yu. Batvision with gcc-phat features for better sound to vision predictions. arXiv preprint arXiv:2006.07995, 2020. 2 +[20] Michael Defferrard, Kirell Benzi, Pierre Vandergheynst, and Xavier Bresson. Fma: A dataset for music analysis. arXiv preprint arXiv:1612.01840, 2016. 5, 1 +[21] Yuexi Du, Ziyang Chen, Justin Salamon, Bryan Russell, and Andrew Owens. Conditional generation of audio from video via foley analogies. Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[22] Mohamed El Banani, Luya Gao, and Justin Johnson. Unsupervisedr&r: Unsupervised point cloud registration via differentiable rendering. In CVPR, 2021. 2 +[23] Mohamed El Banani and Justin Johnson. Bootstrap your own correspondences. In ICCV, 2021. 2 +[24] Chao Feng, Ziyang Chen, and Andrew Owens. Self-supervised video forensics by audio-visual anomaly detection. arXiv preprint arXiv:2301.01767, 2023. 2 +[25] Tim Fischer, Marco Caversaccio, and Wilhelm Wimmer. A front-back confusion metric in horizontal sound localization: The fbc score. In ACM Symposium on Applied Perception 2020, pages 1-5, 2020. 4 +[26] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981. 6, 1 +[27] Andrew Francl. Modeling and Evaluating Human Sound Localization in the Natural Environment. PhD thesis, Massachusetts Institute of Technology, 2022. 2, 5, 8, 1 +[28] Chuang Gan, Deng Huang, Peihao Chen, Joshua B Tenenbaum, and Antonio Torralba. Foley music: Learning to generate music from videos. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 758-775. Springer, 2020. 2 + +[29] Ruohan Gao, Changan Chen, Ziad Al-Halah, Carl Schissler, and Kristen Grauman. Visualechoes: Spatial image representation learning through echolocation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16, pages 658-676. Springer, 2020. 2 +[30] Ruohan Gao and Kristen Grauman. 2.5 d visual sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 324-333, 2019. 2, 3, 5 +[31] Ruohan Gao and Kristen Grauman. Visualvoice: Audiovisual speech separation with cross-modal consistency. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15490-15500. IEEE, 2021. 2 +[32] Rishabh Garg, Ruohan Gao, and Kristen Grauman. Geometry-aware multi-task learning for binaural audio generation from video. arXiv preprint arXiv:2111.10882, 2021. 2 +[33] Sanchita Ghose and John Jeffrey Prevost. Auto Foley: Artificial synthesis of synchronized sound tracks for silent videos with deep learning. IEEE Transactions on Multimedia, 23:1895-1907, 2020. 2 +[34] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14950-14962, 2022. 2 +[35] R. I. Hartley. Chirality. International Journal of Computer Vision (IJCV), 1998. 7 +[36] R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. 2, 6, 7 +[37] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In International Conference on Computer Vision (ICCV), 2015. 5 +[38] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition (CVPR), 2016. 4, 5, 8, 1 +[39] Xixi Hu, Ziyang Chen, and Andrew Owens. Mix and localize: Localizing sound sources in mixtures. Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[40] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. arXiv preprint arXiv:2301.12661, 2023. 2 +[41] Wen Chin Huang, Dejan Markovic, Alexander Richard, Israel Dejene Gebru, and Anjali Menon. End-to-end binaural speech synthesis. arXiv preprint arXiv:2207.03697, 2022. 2 +[42] Vladimir Iashin and Esa Rahtu. Taming visually guided sound generation. arXiv preprint arXiv:2110.08791, 2021. 2 +[43] Linyi Jin, Shengyi Qian, Andrew Owens, and David F Fouhey. Planar surface reconstruction from sparse views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12991-13000, 2021. 2, 3, 4, 5, 7, 8 + +[44] Linyi Jin, Jianming Zhang, Yannick Hold-Geoffroy, Oliver Wang, Kevin Matzen, Matthew Sticha, and David F Fouhey. Perspective fields for single image camera calibration. arXiv, 2022. 2 +[45] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In ICCV, 2015. 2 +[46] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representation, 2015. 2 +[47] Charles Knapp and Grifford Carter. The generalized correlation method for estimation of time delay. IEEE transactions on acoustics, speech, and signal processing, 24(4):320-327, 1976. 2, 6 +[48] Takashi Konno, Kenji Nishida, Katsutoshi Itoyama, and Kazuhiro Nakadai. Audio-visual sfm towards 4d reconstruction under dynamic scenes. 2022. 2 +[49] Felix Kreuk, Gabriel Synnaeve, Adam Polyak, Uriel Singer, Alexandre Défossez, Jade Copet, Devi Parikh, Yaniv Taigman, and Yossi Adi. Audiogen: Textually guided audio generation. arXiv preprint arXiv:2209.15352, 2022. 2 +[50] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Av-nerf: Learning neural fields for real-world audio-visual scene synthesis. arXiv preprint arXiv:2302.02088, 2023. 2 +[51] Yan-Bo Lin and Yu-Chiang Frank Wang. Exploiting audiovisual consistency with partial supervision for spatial audio generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2056–2063, 2021. 2 +[52] Zhi-Hao Lin, Wei-Chiu Ma, Hao-Yu Hsu, Yu-Chiang Frank Wang, and Shenlong Wang. Neurmips: Neural mixture of planar experts for view synthesis. In CVPR, 2022. 8 +[53] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 2 +[54] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 6, 7, 8 +[55] Wei-Chiu Ma, Anqi Joyce Yang, Shenlong Wang, Raquel Urtasun, and Antonio Torralba. Virtual correspondence: Humans as a cue for extreme-view geometry. In CVPR, 2022. 2, 8 +[56] Sagnik Majumder, Ziad Al-Halah, and Kristen Grauman. Move2hear: Active audio-visual source separation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 275–285, 2021. 2 +[57] Sagnik Majumder, Changan Chen, Ziad Al-Halah, and Kristen Grauman. Few-shot audio-visual learning of environment acoustics. arXiv preprint arXiv:2206.04006, 2022. 2 +[58] Sagnik Majumder and Kristen Grauman. Active audio-visual separation of dynamic sound sources. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXIX, pages 551–569. Springer, 2022. 2 +[59] Iaroslav Melekhov, Juha Ylioinas, Juho Kannala, and Esa Rahtu. Relative camera pose estimation using convolutional + +neural networks. In Advanced Concepts for Intelligent Vision Systems: 18th International Conference, ACIVS 2017, Antwerp, Belgium, September 18-21, 2017, Proceedings 18, pages 675-687. Springer, 2017. 2 +[60] Himangi Mittal, Pedro Morgado, Unnat Jain, and Abhinav Gupta. Learning state-aware visual representations from audible interactions. arXiv preprint arXiv:2209.13583, 2022. 2 +[61] Shentong Mo and Pedro Morgado. A closer look at weakly-supervised audio-visual source localization. arXiv preprint arXiv:2209.09634, 2022. 2 +[62] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 218-234. Springer, 2022. 2 +[63] Pedro Morgado, Yi Li, and Nuno Nvasconcelos. Learning representations from audio-visual spatial alignment. Advances in Neural Information Processing Systems, 33:4733-4744, 2020. 2, 5, 8, 1 +[64] Pedro Morgado, Nuno Vasconcelos, Timothy Langlois, and Oliver Wang. Self-supervised generation of spatial audio for 360 video. arXiv preprint arXiv:1809.02587, 2018. 2 +[65] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12475-12486, 2021. 2 +[66] David Nister. An efficient solution to the five-point relative pose problem. IEEE transactions on pattern analysis and machine intelligence, 26(6):756-770, 2004. 6 +[67] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. European Conference on Computer Vision (ECCV), 2018. 2 +[68] Andrew Owens, Phillip Isola, Josh McDermott, Antonio Torralba, Edward H Adelson, and William T Freeman. Visually indicated sounds. In Computer Vision and Pattern Recognition (CVPR), 2016. 2 +[69] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE, 2015. 5, 7, 8, 1, 2 +[70] Kranti Kumar Parida, Siddharth Srivastava, and Gaurav Sharma. Beyond image to depth: Improving depth prediction using echoes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8268-8277, 2021. 2 +[71] Senthil Purushwalkam, Sebastia Vicenc Amengual Gari, Vamsi Krishna Ithapu, Carl Schissler, Philip Robinson, Abhinav Gupta, and Kristen Grauman. Audio-visual floorplan reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1183–1192, 2021. 2 +[72] Shengyi Qian, Linyi Jin, and David F Fouhey. Associative3d: Volumetric reconstruction from sparse views. In ECCV, 2020. 2 + +[73] Shengyi Qian, Alexander Kirillov, Nikhila Ravi, Devendra Singh Chaplot, Justin Johnson, David F Fouhey, and Georgia Gkioxari. Recognizing scenes from novel viewpoints. arXiv preprint arXiv:2112.01520, 2021. 3, 5 +[74] Kranthi Kumar Rachavarapu, Vignesh Sundaresha, AN Rajagopalan, et al. Localize to binauralize: Audio spatialization from visual sound source localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1930-1939, 2021. 2 +[75] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[76] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 5, 7, 2 +[77] Lord Rayleigh. Xii. on our perception of sound direction. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 13(74):214-232, 1907. 2, 4 +[78] Alexander Richard, Dejan Markovic, Israel D Gebru, Steven Krenn, Gladstone Alexander Butler, Fernando Torre, and Yaser Sheikh. Neural synthesis of binaural speech from mono audio. In International Conference on Learning Representations, 2021. 2 +[79] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2 +[80] Lord Rayleigh O.M. Pres. R.S. Xii. on our perception of sound direction. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 1907. 2 +[81] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4938–4947, 2020. 6, 7, 8 +[82] Ralph Schmidt. Multiple emitter location and signal parameter estimation. IEEE transactions on antennas and propagation, 34(3):276-280, 1986. 2 +[83] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2 +[84] Nikhil Singh, Jeff Mentch, Jerry Ng, Matthew Beveridge, and Iddo Drori. Image2reverb: Cross-modal reverb impulse response synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 286-295, 2021. 2 + +[85] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 6 +[86] Ruijie Tao, Zexu Pan, Rohan Kumar Das, Xinyuan Qian, Mike Zheng Shou, and Haizhou Li. Is someone speaking? exploring long-term temporal features for audio-visual active speaker detection. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3927–3935, 2021. 2 +[87] Efthymios Tzinis, Scott Wisdom, Tal Remez, and John R Hershey. Audioscopev2: Audio-visual attention architectures for calibrated open-domain on-screen sound separation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXVII, pages 368–385. Springer, 2022. 2 +[88] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 2 +[89] Paolo Vecchiotti, Ning Ma, Stefano Squartini, and Guy J Brown. End-to-end binaural sound localisation from the raw waveform. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 451-455. IEEE, 2019. 2, 3 +[90] DeLiang Wang and Guy J Brown. Computational auditory scene analysis: Principles, algorithms, and applications. Wiley-IEEE press, 2006. 2 +[91] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. In ECCV, 2022. 8 +[92] Fanyi Xiao, Yong Jae Lee, Kristen Grauman, Jitendra Malik, and Christoph Feichtenhofer. Audiovisual slowfast networks for video recognition. arXiv preprint arXiv:2001.08740, 2020. 2 +[93] Xudong Xu, Hang Zhou, Ziwei Liu, Bo Dai, Xiaogang Wang, and Dahua Lin. Visually informed binaural audio generation without binaural audios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15485-15494, 2021. 2 +[94] Nelson Yalta, Kazuhiro Nakadai, and Tetsuya Ogata. Sound source localization using deep learning models. Journal of Robotics and Mechatronics, 29(1):37-48, 2017. 2 +[95] Dongchao Yang, Jianwei Yu, Helin Wang, Wen Wang, Chao Weng, Yuexian Zou, and Dong Yu. Diffsound: Discrete diffusion model for text-to-sound generation. arXiv preprint arXiv:2207.09983, 2022. 2 +[96] Karren Yang, Michael Firman, Eric Brachmann, and Clément Godard. Camera pose estimation and localization with active audio sensing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 271-291. Springer, 2022. 2 +[97] Karren Yang, Bryan Russell, and Justin Salamon. Telling left from right: Learning spatial correspondence of sight + +and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9932-9941, 2020. 2 +[98] Zhenpei Yang, Jeffrey Z Pan, Linjie Luo, Xiaowei Zhou, Kristen Grauman, and Qixing Huang. Extreme relative pose estimation for rgb-d scans via scene completion. In CVPR, 2019. 2 +[99] Zhenpei Yang, Siming Yan, and Qixing Huang. Extreme relative pose network under hybrid representations. In CVPR, 2020. 2 +[100] Dong Yu, Morten Kolbæk, Zheng-Hua Tan, and Jesper Jensen. Permutation invariant training of deep models for speaker-independent multi-talker speech separation. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 241-245. IEEE, 2017. 4 +[101] Yinda Zhang, Shuran Song, Ersin Yumer, Manolis Savva, Joon-Young Lee, Hailin Jin, and Thomas Funkhouser. Physically-based rendering for indoor scene understanding using convolutional neural networks. In CVPR, 2017. 3, 5 +[102] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In ICCV, 2021. 3, 5 +[103] Hang Zhou, Xudong Xu, Dahua Lin, Xiaogang Wang, and Ziwei Liu. Sep-stereo: Visually guided stereophonic audio generation by associating source separation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2 +[104] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, 2017. 1, 2, 3, 4 +[105] Yipin Zhou and Ser-Nam Lim. Joint audio-visual deepfake detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14800–14809, 2021. 2 +[106] Yuliang Zou, Pan Ji, Quoc-Huy Tran, Jia-Bin Huang, and Manmohan Chandraker. Learning monocular visual odometry via self-supervised long-term modeling. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV, pages 710-727. Springer, 2020. 2 \ No newline at end of file diff --git a/2023/Sound Localization from Motion_ Jointly Learning Sound Direction and Camera Rotation/images.zip b/2023/Sound Localization from Motion_ Jointly Learning Sound Direction and Camera Rotation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cb59df9d006cdba9c4df04fc45d637713ed61a2a --- /dev/null +++ b/2023/Sound Localization from Motion_ Jointly Learning Sound Direction and Camera Rotation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a8ae91a0401208c938566c2d499033282fe0d938d3559989a12ceef5b8b349c +size 542130 diff --git a/2023/Sound Localization from Motion_ Jointly Learning Sound Direction and Camera Rotation/layout.json b/2023/Sound Localization from Motion_ Jointly Learning Sound Direction and Camera Rotation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dbbd5a5b021ef27b8c6ba2938ac6f3538000d182 --- /dev/null +++ b/2023/Sound Localization from Motion_ Jointly Learning Sound Direction and Camera Rotation/layout.json @@ -0,0 +1,10569 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 124, + 103, + 470, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 103, + 470, + 140 + ], + "spans": [ + { + "bbox": [ + 124, + 103, + 470, + 140 + ], + "type": "text", + "content": "Sound Localization from Motion: Jointly Learning Sound Direction and Camera Rotation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 165, + 161, + 231, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 161, + 231, + 176 + ], + "spans": [ + { + "bbox": [ + 165, + 161, + 231, + 176 + ], + "type": "text", + "content": "Ziyang Chen" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 255, + 161, + 323, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 161, + 323, + 175 + ], + "spans": [ + { + "bbox": [ + 255, + 161, + 323, + 175 + ], + "type": "text", + "content": "Shengyi Qian" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 348, + 162, + 425, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 162, + 425, + 175 + ], + "spans": [ + { + "bbox": [ + 348, + 162, + 425, + 175 + ], + "type": "text", + "content": "Andrew Owens" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 235, + 185, + 350, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 185, + 350, + 200 + ], + "spans": [ + { + "bbox": [ + 235, + 185, + 350, + 200 + ], + "type": "text", + "content": "University of Michigan" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 252, + 290, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 473 + ], + "type": "text", + "content": "The images and sounds that we perceive undergo subtle but geometrically consistent changes as we rotate our heads. In this paper, we use these cues to solve a problem we call Sound Localization from Motion (SLfM): jointly estimating camera rotation and localizing sound sources. We learn to solve these tasks solely through self-supervision. A visual model predicts camera rotation from a pair of images, while an audio model predicts the direction of sound sources from binaural sounds. We train these models to generate predictions that agree with one another. At test time, the models can be deployed independently. To obtain a feature representation that is well-suited to solving this challenging problem, we also propose a method for learning an audio-visual representation through cross-view binauralization: estimating binaural sound from one view, given images and sound from another. Our model can successfully estimate accurate rotations on both real and synthetic scenes, and localize sound sources with accuracy competitive with state-of-the-art self-supervised approaches. Project site: https://ificl.github.io/SLfM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 499, + 128, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 128, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 128, + 511 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 521, + 288, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 593 + ], + "type": "text", + "content": "As you rotate your head, the images and sounds that you perceive change in geometrically consistent ways. For example, after turning to the right, a sound source that was directly in front of you will become louder in your left ear and quieter in your right, while simultaneously the visual scene will move right-to-left across your visual field (Fig. 1)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": "We hypothesize that these co-occurring audio and visual signals provide \"free\" supervision that captures geometry, including the motion made by a camera and the direction of sound sources. These are each core problems in machine perception, but are largely studied separately, often using supervised methods that rely on difficult-to-acquire labeled training data, such as annotated sound directions. We take inspiration from self-supervised approaches to structure from motion [104], which learn to estimate 3D structure and camera pose by solving both tasks simultaneously." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 224, + 545, + 406 + ], + "blocks": [ + { + "bbox": [ + 307, + 224, + 545, + 406 + ], + "lines": [ + { + "bbox": [ + 307, + 224, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 307, + 224, + 545, + 406 + ], + "type": "image", + "image_path": "db541dd8a9a951af38cf10e58cc73dfe15c211eb8e2078d40fec774a8fb1bdfc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 409, + 547, + 487 + ], + "lines": [ + { + "bbox": [ + 305, + 409, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 305, + 409, + 547, + 487 + ], + "type": "text", + "content": "Figure 1: Images and sounds change in geometrically consistent ways. For example, when we rotate to the right, a sound source that is initially in front of us becomes louder in our left ear. We use these cues to jointly train models for two tasks: localizing sounds from binaural audio and estimating camera rotation from images. The two models are trained entirely through self-supervision, by learning to produce outputs that agree with one other." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 495, + 547, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 639 + ], + "type": "text", + "content": "Analogously, we propose a problem we call sound localization from motion (SLfM): jointly estimating camera rotation from images and the sound direction from binaural audio. By solving both tasks simultaneously, we avoid the need for labeled training data. Our models provide each other with self-supervision: a visual model predicts the rotation angle between pairs of images, while an audio model predicts the azimuth of sound sources. We force their predictions to agree with one another, such that changes in rotation are consistent with changes in sound direction and binaural cues. After training, the models can be deployed independently, without multimodal data at test time." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 641, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 548, + 714 + ], + "type": "text", + "content": "This is a challenging task that requires perceiving motion in images and binaural cues in audio. Our second contribution is a method for learning representations that are well-suited to this task through cross-view binauralization. We train a network to convert mono to binaural sound for one viewpoint, given an audio-visual pair sampled from another" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "7897" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "viewpoint. Since the sound source is not necessarily visible in the images, the only way to successfully solve this pretext task is by analyzing the changes in the camera pose and predicting how they affect the sound direction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "text", + "content": "All components of our model are entirely self-supervised and are trained solely on unlabeled audio-visual data. Our results suggest that paired audio-visual data provides a useful and complementary signal for learning about geometry. In contrast to other audio or visual self-supervised pose estimation methods, we obtain supervision from abundantly available audio data, thus avoiding the need of 3D ground truth or correspondences between pixels [104, 106] or audio samples [17]. Through experiments, we show:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 230, + 287, + 337 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 230, + 287, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 230, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 47, + 230, + 287, + 253 + ], + "type": "text", + "content": "- Paired audio-visual data provides a supervisory signal for pose estimation tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 255, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 287, + 277 + ], + "type": "text", + "content": "- We obtain competitive performance with state-of-the-art self-supervised sound localization methods [17]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 278, + 287, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 278, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 287, + 314 + ], + "type": "text", + "content": "- We obtain strong rotation estimation performance, and our model generalizes to Stanford2D3D [6] dataset, where it is competitive with classic sparse feature matching methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 315, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 315, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 47, + 315, + 287, + 337 + ], + "type": "text", + "content": "- The features we learn through our pretext task outperform other representations for our downstream tasks." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 350, + 134, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 134, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 134, + 363 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 370, + 288, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 370, + 288, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 370, + 288, + 597 + ], + "type": "text", + "content": "Audio for spatial perception. Recent works have explored the use of sound for spatial understanding. Purushwalkam et al. [71] reconstructed floor plans in simulated environments [12]. Chen et al. [18] used ambient sounds from environments to learn about scene structures. Konno et al. [48] integrated sound localization to visual SfM while do not jointly learn them. Other work learns representations for spatial audio-visual tasks. Yang et al. [97] predicted whether stereo channels are swapped in a video, and Morgado et al. [63] solved a spatial alignment task. The learned representations are then used to improve localization, up-mixing, and segmentation models. In contrast, we learn camera pose and sound localization solely from self-supervision, obtaining angular predictions without labeled data. Other work uses echolocation sounds to learn representations [29, 96] and predict depth maps [19, 70] and estimate camera poses [96] using labeled data. In contrast, our proposed approach jointly learns binaural sound localization and camera pose through passive audio sensing, without supervision." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "Acoustic synthesis and spatialization. Researchers have explored visually-guided sound synthesis [28, 33, 42, 21] and text-guided audio synthesis [49, 95, 40]. Additionally, researchers have investigated generating realistic environmental acoustics using visual information [11, 84, 15, 57]. Chen et al. [14] introduced the novel-view acoustic synthesis task, which synthesizes binaural sound at the target view using audio and visual information from a source view. Liang et al. [50] proposed an audio-visual neural field in" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "real-world audio-visual scenes. Many recent works have proposed to generate spatial audio from mono audio using visual cues [64, 30, 74, 93, 51, 103, 32], or the relative pose between sound sources and the receiver [78, 41]. Inspired by these works, our feature learning approach learns spatial representations through an audio prediction task." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 150, + 547, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 150, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 150, + 547, + 365 + ], + "type": "text", + "content": "Binaural sound localization. Humans have the ability to localize sound sources from binaural sound [80]. Traditional approaches estimate interaural time delays via cross-correlation using hand-crafted features [47], factorization methods [82], or loudness differences between ears [77, 90]. Chen et al. [17] adapted methods from self-supervised visual tracking to the problem of binaural sound localization. Similarly, we estimate direction through self-supervision. However, we obtain our supervision through cross-modal supervision from vision instead of from correspondence cues. Moreover, we also obtain visual camera rotation estimation through our learning process. Francl [27] learned representations of sound location with a contrastive loss where positive and negative examples are selected based on the extent of head movements. Other works have used supervised learning techniques with labeled data to localize sound sources in reverberant environments [1, 89, 94]. Unlike these methods, our model learns 3D sound localization without labels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 372, + 547, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 547, + 599 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 547, + 599 + ], + "type": "text", + "content": "Camera pose estimation. Traditional methods for camera pose estimation are based on finding correspondences between images and then solving an optimization problem using constraints from multi-view geometry [36]. These include structure from motion methods that estimate full pose [83] and camera rotation [8]. Recent methods have directly predicted camera pose using neural networks, including methods that use photos [72, 45, 59, 44, 55] or RGB-D scans [99, 98, 22, 23]. Our setup is similar to work that learns relative camera poses from sparse views [43, 9], and we use their network architectures. However, we learn the camera pose through cross-modal supervision from audio, rather than from labels. Our work is also closely related to methods that learn structure from motion through self-supervision [104, 106], such as by jointly learning models that perform depth and camera pose estimation with photoconsistency constraints. In contrast, our visual model's learning signal comes solely from audio-based supervision, and we jointly learn audio localization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "Audio-visual learning. A number of works have focused on learning multimodal representations for audio and vision, taking into account semantic correspondence and temporal synchronization [67, 92, 7, 65, 68, 2, 60]. Other approaches study audio-visual sound localization [39, 61, 16, 62], source separation [56, 31, 58, 87], active speaker detection [3, 86, 5], navigation [12, 13, 10] and forensics [105, 34, 24]. Our focus, in contrast, is on utilizing multi-view audio-visual signals to learn geometry." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "7898" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 60, + 380, + 178 + ], + "blocks": [ + { + "bbox": [ + 47, + 60, + 380, + 178 + ], + "lines": [ + { + "bbox": [ + 47, + 60, + 380, + 178 + ], + "spans": [ + { + "bbox": [ + 47, + 60, + 380, + 178 + ], + "type": "image", + "image_path": "0af377614fad8cce4404f7daff0ffe16d7c754f65cb725801f3cd0a6a852a88d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 183, + 264, + 195 + ], + "lines": [ + { + "bbox": [ + 136, + 183, + 264, + 195 + ], + "spans": [ + { + "bbox": [ + 136, + 183, + 264, + 195 + ], + "type": "text", + "content": "(a) Cross-view audio prediction" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 388, + 60, + 545, + 176 + ], + "blocks": [ + { + "bbox": [ + 388, + 60, + 545, + 176 + ], + "lines": [ + { + "bbox": [ + 388, + 60, + 545, + 176 + ], + "spans": [ + { + "bbox": [ + 388, + 60, + 545, + 176 + ], + "type": "image", + "image_path": "7aec18f353123bcdd088dfa7097260c35e09a003197f9dadb80260e5aa046005.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 183, + 536, + 194 + ], + "lines": [ + { + "bbox": [ + 392, + 183, + 536, + 194 + ], + "spans": [ + { + "bbox": [ + 392, + 183, + 536, + 194 + ], + "type": "text", + "content": "(b) Sound localization from motion" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 200, + 547, + 245 + ], + "lines": [ + { + "bbox": [ + 45, + 200, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 45, + 200, + 547, + 245 + ], + "type": "text", + "content": "Figure 2: Method overview. (a) We learn a feature representation by predicting how changes in images lead to changes in sound in a cross-view binauralization pretext task. We convert mono sound to binaural sound at a target viewpoint, after conditioning the model on observations from a source viewpoint. (b) We use the representation to jointly solve two pose estimation tasks: visual rotation estimation and binaural sound localization. We train the visual rotation angle, " + }, + { + "bbox": [ + 45, + 200, + 547, + 245 + ], + "type": "inline_equation", + "content": "\\phi_{s,t}" + }, + { + "bbox": [ + 45, + 200, + 547, + 245 + ], + "type": "text", + "content": ", to be consistent with the difference in predicted sound angles " + }, + { + "bbox": [ + 45, + 200, + 547, + 245 + ], + "type": "inline_equation", + "content": "\\theta_s" + }, + { + "bbox": [ + 45, + 200, + 547, + 245 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 200, + 547, + 245 + ], + "type": "inline_equation", + "content": "\\theta_t" + }, + { + "bbox": [ + 45, + 200, + 547, + 245 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 257, + 230, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 230, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 230, + 270 + ], + "type": "text", + "content": "3. Sound Localization from Motion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 277, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 348 + ], + "type": "text", + "content": "We address the sound localization from motion (SLfM) task: predicting the azimuth of a sound source from binaural audio and camera rotation from two images. First, we present a self-supervised representation that can be used to solve this downstream task. Then, we show how the representation can be used to solve the task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 357, + 265, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 265, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 265, + 369 + ], + "type": "text", + "content": "3.1. Learning representation via spatialization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 374, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 506 + ], + "type": "text", + "content": "We learn an audio-visual representation that conveys spatial information by solving a cross-view binauralization task: converting mono sound to stereo for one viewpoint, given an audio-visual pair sampled from another viewpoint (Fig. 2a). In order to successfully solve the task, a model must implicitly estimate the sound direction in the source view and predict how the change in viewpoint will affect the sound in the target view. A key difference between this task and traditional, single-view binauralization [30] is that the sound source need not be visible in any of the images. Hence, the task cannot be solved from the target audio-visual pair alone." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 506, + 288, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 288, + 602 + ], + "type": "text", + "content": "We binauralize the sound " + }, + { + "bbox": [ + 46, + 506, + 288, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t" + }, + { + "bbox": [ + 46, + 506, + 288, + 602 + ], + "type": "text", + "content": " at the target view through an audio predictor " + }, + { + "bbox": [ + 46, + 506, + 288, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 46, + 506, + 288, + 602 + ], + "type": "text", + "content": ". To make our representation suitable for downstream tasks, we factorize it into visual features " + }, + { + "bbox": [ + 46, + 506, + 288, + 602 + ], + "type": "inline_equation", + "content": "f_{v}(\\mathbf{v}_{s},\\mathbf{v}_{t})" + }, + { + "bbox": [ + 46, + 506, + 288, + 602 + ], + "type": "text", + "content": ", which are intended to create features relevant to relative pose, and audio features " + }, + { + "bbox": [ + 46, + 506, + 288, + 602 + ], + "type": "inline_equation", + "content": "f_{a}(\\mathbf{a}_{s})" + }, + { + "bbox": [ + 46, + 506, + 288, + 602 + ], + "type": "text", + "content": ", which capture sound localization cues. We predict the binauralized audio " + }, + { + "bbox": [ + 46, + 506, + 288, + 602 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}_t" + }, + { + "bbox": [ + 46, + 506, + 288, + 602 + ], + "type": "text", + "content": " at the target view from mono audio " + }, + { + "bbox": [ + 46, + 506, + 288, + 602 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{a}}_t" + }, + { + "bbox": [ + 46, + 506, + 288, + 602 + ], + "type": "text", + "content": ", the visual change of " + }, + { + "bbox": [ + 46, + 506, + 288, + 602 + ], + "type": "inline_equation", + "content": "(\\mathbf{v}_s,\\mathbf{v}_t)" + }, + { + "bbox": [ + 46, + 506, + 288, + 602 + ], + "type": "text", + "content": " and binaural sound " + }, + { + "bbox": [ + 46, + 506, + 288, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_s" + }, + { + "bbox": [ + 46, + 506, + 288, + 602 + ], + "type": "text", + "content": " heard at the source viewpoint:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 97, + 611, + 287, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 611, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 97, + 611, + 287, + 624 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {a}} _ {t} = \\mathcal {F} _ {\\theta} \\left(\\bar {\\mathbf {a}} _ {t}, f _ {v} \\left(\\mathbf {v} _ {s}, \\mathbf {v} _ {t}\\right), f _ {a} \\left(\\mathbf {a} _ {s}\\right)\\right). \\tag {1}", + "image_path": "b56c5ec4d7904804fcf16d13f90cd56fd74738f444451640eaee5a7a59b5aa4e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 629, + 287, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 702 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 702 + ], + "type": "text", + "content": "We represent audio " + }, + { + "bbox": [ + 46, + 629, + 287, + 702 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 46, + 629, + 287, + 702 + ], + "type": "text", + "content": " as a spectrogram " + }, + { + "bbox": [ + 46, + 629, + 287, + 702 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 629, + 287, + 702 + ], + "type": "text", + "content": " using short-time Fourier transform (STFT). Following Gao et al. [30], given the mix of stereo audio " + }, + { + "bbox": [ + 46, + 629, + 287, + 702 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{A}}_t = \\mathrm{STFT}(\\mathbf{a}_t^L + \\mathbf{a}_t^R)" + }, + { + "bbox": [ + 46, + 629, + 287, + 702 + ], + "type": "text", + "content": ", we predict the difference of two channels " + }, + { + "bbox": [ + 46, + 629, + 287, + 702 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_t = \\mathrm{STFT}(\\mathbf{a}_t^L - \\mathbf{a}_t^R)" + }, + { + "bbox": [ + 46, + 629, + 287, + 702 + ], + "type": "text", + "content": ". We optimize the " + }, + { + "bbox": [ + 46, + 629, + 287, + 702 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 46, + 629, + 287, + 702 + ], + "type": "text", + "content": " loss between predicted spectrogram " + }, + { + "bbox": [ + 46, + 629, + 287, + 702 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}_t" + }, + { + "bbox": [ + 46, + 629, + 287, + 702 + ], + "type": "text", + "content": " and ground-truth spectrogram " + }, + { + "bbox": [ + 46, + 629, + 287, + 702 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_t" + }, + { + "bbox": [ + 46, + 629, + 287, + 702 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 704, + 287, + 718 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 704, + 287, + 718 + ], + "spans": [ + { + "bbox": [ + 115, + 704, + 287, + 718 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p r e t e x t}} = \\left| \\left| \\hat {\\mathbf {A}} _ {t} - \\mathbf {A} _ {t} \\right| \\right| _ {1}. \\tag {2}", + "image_path": "8484978a8bcca4f0e0e979a83fefb014d3429c71f4194b77314be2c6bfbb0aee.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 258, + 547, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 258, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 547, + 354 + ], + "type": "text", + "content": "Multi-view binauralization. Following Zhou et al. [104], we improve our representations by binauralizing sounds at " + }, + { + "bbox": [ + 305, + 258, + 547, + 354 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 258, + 547, + 354 + ], + "type": "text", + "content": " different target viewpoints, using observations from a single source viewpoint " + }, + { + "bbox": [ + 305, + 258, + 547, + 354 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 305, + 258, + 547, + 354 + ], + "type": "text", + "content": ". We hypothesize that jointly solving spatialization problems from a single viewpoint for multiple target viewpoints would require the model to make more accurate predictions of sound source locations, thereby improving the estimation of view changes:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 309, + 365, + 545, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 365, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 309, + 365, + 545, + 407 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p r e t e x t}} = \\frac {1}{N} \\sum_ {i} ^ {N} \\left| \\left| \\mathcal {F} _ {\\theta} \\left(\\bar {\\mathbf {A}} _ {i}, f _ {v} \\left(\\mathbf {v} _ {s}, \\mathbf {v} _ {i}\\right), f _ {a} \\left(\\mathbf {a} _ {s}\\right)\\right) - \\mathbf {A} _ {i} \\right| \\right| _ {1}. \\tag {3}", + "image_path": "24fbb7489dd9415028e973c7ee2404296b05191157244bcfe7ae6eb6e422107a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 411, + 506, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 411, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 306, + 411, + 506, + 423 + ], + "type": "text", + "content": "3.2. Estimating pose and localizing sounds" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 430, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 502 + ], + "type": "text", + "content": "We now address the problem of learning models for sound localization and pose estimation, using our self-supervised audio-visual features. Given two views, we predict sound directions and relative rotation. We train the model to make these two predictions consistent with one another, while using simple binaural constraints to resolve ambiguities." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 502, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 545, + 539 + ], + "type": "text", + "content": "We are given images " + }, + { + "bbox": [ + 304, + 502, + 545, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_s" + }, + { + "bbox": [ + 304, + 502, + 545, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 502, + 545, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t" + }, + { + "bbox": [ + 304, + 502, + 545, + 539 + ], + "type": "text", + "content": " (rotated views recorded at the same position) and learned visual embedding " + }, + { + "bbox": [ + 304, + 502, + 545, + 539 + ], + "type": "inline_equation", + "content": "f_v" + }, + { + "bbox": [ + 304, + 502, + 545, + 539 + ], + "type": "text", + "content": ". We predict the scalar rotation angle " + }, + { + "bbox": [ + 304, + 502, + 545, + 539 + ], + "type": "inline_equation", + "content": "\\phi_{s,t}" + }, + { + "bbox": [ + 304, + 502, + 545, + 539 + ], + "type": "text", + "content": " via the encoder " + }, + { + "bbox": [ + 304, + 502, + 545, + 539 + ], + "type": "inline_equation", + "content": "g_v" + }, + { + "bbox": [ + 304, + 502, + 545, + 539 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 550, + 545, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 550, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 312, + 550, + 545, + 586 + ], + "type": "interline_equation", + "content": "\\phi_ {s, t} = g _ {v} \\left(f _ {v} \\left(\\mathbf {v} _ {s}, \\mathbf {v} _ {t}\\right)\\right), R _ {s, t} = \\left[ \\begin{array}{l l} \\cos \\phi_ {s, t} & - \\sin \\phi_ {s, t} \\\\ \\sin \\phi_ {s, t} & \\cos \\phi_ {s, t} \\end{array} \\right], \\tag {4}", + "image_path": "74c5ddc072081c5067b18d11892df35b63278d7482c25beac2ab0b7465b60d16.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 586, + 547, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 547, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 547, + 646 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 586, + 547, + 646 + ], + "type": "inline_equation", + "content": "R_{s,t}" + }, + { + "bbox": [ + 304, + 586, + 547, + 646 + ], + "type": "text", + "content": " is 2D rotation matrix of " + }, + { + "bbox": [ + 304, + 586, + 547, + 646 + ], + "type": "inline_equation", + "content": "\\phi_{s,t}" + }, + { + "bbox": [ + 304, + 586, + 547, + 646 + ], + "type": "text", + "content": ". Following common practice in indoor scene reconstruction, we give the camera a fixed downward tilt [102, 73, 101] and only estimate azimuth [43]. This is also a common assumption in audio localization [1, 89], since azimuth has strong binaural cues." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 647, + 545, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 647, + 545, + 671 + ], + "spans": [ + { + "bbox": [ + 305, + 647, + 545, + 671 + ], + "type": "text", + "content": "Similarly, we predict the azimuths of the sound sources using audio features and the encoder " + }, + { + "bbox": [ + 305, + 647, + 545, + 671 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 305, + 647, + 545, + 671 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 334, + 682, + 545, + 699 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 682, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 334, + 682, + 545, + 699 + ], + "type": "interline_equation", + "content": "\\theta_ {i} = g _ {a} \\left(f _ {a} (\\mathbf {a} _ {i})\\right), \\quad \\mathbf {r} _ {i} = \\left[ \\begin{array}{l l} \\cos \\theta_ {i} & \\sin \\theta_ {i} \\end{array} \\right] ^ {\\top}, \\tag {5}", + "image_path": "1bc974a1155ea97f7249353dd4cc292e00c3143280a505238845591a683f4527.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 710, + 491, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 710, + 491, + 721 + ], + "spans": [ + { + "bbox": [ + 306, + 710, + 491, + 721 + ], + "type": "text", + "content": "where we represent the azimuth as a vector " + }, + { + "bbox": [ + 306, + 710, + 491, + 721 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "spans": [ + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "type": "text", + "content": "7899" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 142 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 142 + ], + "type": "text", + "content": "Cross-modal geometric consistency. When the camera is rotated, the sound source ought to rotate in the opposite direction (Fig. 2b). For example, a " + }, + { + "bbox": [ + 46, + 72, + 289, + 142 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 46, + 72, + 289, + 142 + ], + "type": "text", + "content": " clockwise camera rotation should result in a " + }, + { + "bbox": [ + 46, + 72, + 289, + 142 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 46, + 72, + 289, + 142 + ], + "type": "text", + "content": " counterclockwise rotation in sound direction. Such a constraint could be converted into a loss:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 118, + 143, + 288, + 157 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 143, + 288, + 157 + ], + "spans": [ + { + "bbox": [ + 118, + 143, + 288, + 157 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r o t}} = \\left\\| \\mathbf {r} _ {s} - R _ {s, t} \\mathbf {r} _ {t} \\right\\| ^ {2}. \\tag {6}", + "image_path": "8b71a9a2eea50dff6e2a70ec40c60076cf3a9eba82b4b71922ed5e5390d651f3.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 162, + 290, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 162, + 290, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 162, + 290, + 245 + ], + "type": "text", + "content": "However, a well-known ambiguity called front-back confusion [77, 25] exists in binaural sound perception: one cannot generally tell whether a sound is in front of the view, or behind them. To address that, we use permutation invariant training [100] (PIT), and allow the model to use either the predicted sound direction or its reflection about the " + }, + { + "bbox": [ + 46, + 162, + 290, + 245 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 162, + 290, + 245 + ], + "type": "text", + "content": " axis without penalty. This results in the loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 251, + 288, + 280 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 251, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 93, + 251, + 288, + 280 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {g e o}} = \\min _ {\\substack {\\hat {\\mathbf {r}} _ {s} \\in \\{\\mathbf {r} _ {s}, Q \\mathbf {r} _ {s} \\} \\\\ \\hat {\\mathbf {r}} _ {t} \\in \\{\\mathbf {r} _ {t}, Q \\mathbf {r} _ {t} \\}}} \\| \\hat {\\mathbf {r}} _ {s} - R _ {s, t} \\hat {\\mathbf {r}} _ {t} \\| ^ {2}, \\tag{7}", + "image_path": "5407f1f72d8a8fa5eba8952a43445dca09f1b3aeb7c309fc3fd603f9891ed7c2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 285, + 252, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 252, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 252, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 48, + 285, + 252, + 312 + ], + "type": "inline_equation", + "content": "Q = \\begin{bmatrix} 1 & 0 \\\\ 0 & -1 \\end{bmatrix}" + }, + { + "bbox": [ + 48, + 285, + 252, + 312 + ], + "type": "text", + "content": " reflects the sound direction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 316, + 288, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 388 + ], + "type": "text", + "content": "As a consequence of this ambiguity, there are also two possible solutions for the visual rotation model, since the visual rotation matrices can be mirrored about the " + }, + { + "bbox": [ + 46, + 316, + 288, + 388 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 316, + 288, + 388 + ], + "type": "text", + "content": " axis. For example, one can create a solution with equal loss by multiplying the rotations and sound directions by " + }, + { + "bbox": [ + 46, + 316, + 288, + 388 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 316, + 288, + 388 + ], + "type": "text", + "content": ". We discuss this ambiguity in more depth in Sec. 4.4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 392, + 289, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 289, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 289, + 540 + ], + "type": "text", + "content": "Incorporating binaural observations. Without additional constraints, the solution is ambiguous, and may collapse into a trivial solution (e.g., predicting zero for all three angles). To avoid this, we force the model to agree with a simple binaural cue based on interaural intensity difference (IID). We predict whether the sound is to the left or right of the viewer, based on whether it is louder in the left or right microphone: " + }, + { + "bbox": [ + 46, + 392, + 289, + 540 + ], + "type": "inline_equation", + "content": "d = \\mathrm{sign}(\\log \\left|\\frac{\\mathbf{A}^L}{\\mathbf{A}^R}\\right|)" + }, + { + "bbox": [ + 46, + 392, + 289, + 540 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 392, + 289, + 540 + ], + "type": "inline_equation", + "content": "|\\mathbf{A}|" + }, + { + "bbox": [ + 46, + 392, + 289, + 540 + ], + "type": "text", + "content": " is the magnitude of the spectrogram A. We perform this left/right test at each timestep in the spectrogram and then pool via majority voting (see supplementary for details). We penalize predictions that are inconsistent with these \"left or right\" observations:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 544, + 287, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 544, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 104, + 544, + 287, + 558 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {b i n a u r a l}} = \\mathcal {L} _ {\\text {B C E}} (\\sin \\theta_ {i}, d _ {i}), \\tag {8}", + "image_path": "a4258a1ef6b179acc73256a87619ef1703c6ca87b6822f61cf63261dc645a0f4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 563, + 214, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 214, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 214, + 576 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 563, + 214, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BCE}}" + }, + { + "bbox": [ + 46, + 563, + 214, + 576 + ], + "type": "text", + "content": " is binary cross entropy loss." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 579, + 289, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 289, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 289, + 664 + ], + "type": "text", + "content": "Encouraging symmetry. To help regularize the model, we also add symmetry constraints. For sound localization, swapping the left and right channels of the audio ought to result in a prediction in the opposite direction since the binaural cues are reversed. For rotation estimation, the relative pose between images " + }, + { + "bbox": [ + 46, + 579, + 289, + 664 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 579, + 289, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 579, + 289, + 664 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 579, + 289, + 664 + ], + "type": "text", + "content": " should invert the pose from " + }, + { + "bbox": [ + 46, + 579, + 289, + 664 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 579, + 289, + 664 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 579, + 289, + 664 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 579, + 289, + 664 + ], + "type": "text", + "content": ". We encourage both constraints via a loss:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 96, + 673, + 287, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 673, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 96, + 673, + 287, + 687 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s y m}} = \\left| \\theta + \\theta_ {\\text {f l i p}} \\right| + \\left| \\phi_ {s, t} + \\phi_ {t, s} \\right|, \\tag {9}", + "image_path": "947712f24f92087dc3d6f770d52b0c85d33ab0fab24ba53ec7b6dd7959f6d74e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 59, + 547, + 183 + ], + "blocks": [ + { + "bbox": [ + 307, + 59, + 547, + 183 + ], + "lines": [ + { + "bbox": [ + 307, + 59, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 59, + 547, + 183 + ], + "type": "image", + "image_path": "99ddbdc91cdffb8b78ffbf33341c969bb81d918b52309982c38bbb2560a362de.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 186, + 547, + 231 + ], + "lines": [ + { + "bbox": [ + 305, + 186, + 547, + 231 + ], + "spans": [ + { + "bbox": [ + 305, + 186, + 547, + 231 + ], + "type": "text", + "content": "Figure 3: Cross-view binauralization architecture. We take a mono spectrogram as input and fuse the audio and visual features from the audio and visual encoder respectively to synthesize the binaural spectrogram at the target viewpoint." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 247, + 547, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 247, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 547, + 282 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 247, + 547, + 282 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{flip}}" + }, + { + "bbox": [ + 305, + 247, + 547, + 282 + ], + "type": "text", + "content": " is the prediction of sound angle " + }, + { + "bbox": [ + 305, + 247, + 547, + 282 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 247, + 547, + 282 + ], + "type": "text", + "content": " using audio with swapped audio channels, and " + }, + { + "bbox": [ + 305, + 247, + 547, + 282 + ], + "type": "inline_equation", + "content": "\\phi_{s,t}" + }, + { + "bbox": [ + 305, + 247, + 547, + 282 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 247, + 547, + 282 + ], + "type": "inline_equation", + "content": "\\phi_{t,s}" + }, + { + "bbox": [ + 305, + 247, + 547, + 282 + ], + "type": "text", + "content": " are the predicted rotations between cameras " + }, + { + "bbox": [ + 305, + 247, + 547, + 282 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 305, + 247, + 547, + 282 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 247, + 547, + 282 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 247, + 547, + 282 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 293, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 545, + 316 + ], + "type": "text", + "content": "Overall loss. We combine these constraints to obtain an overall loss:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 359, + 324, + 545, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 324, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 359, + 324, + 545, + 338 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda \\mathcal {L} _ {\\mathrm {g e o}} + \\mathcal {L} _ {\\text {b i n a u r a l}} + \\mathcal {L} _ {\\text {s y m}}, \\tag {10}", + "image_path": "718ca81afad6a36a1cbe76b8a12e0c9b227d7ed502921adfef0b7d1e2f037ed0.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 349, + 489, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 349, + 489, + 361 + ], + "spans": [ + { + "bbox": [ + 308, + 349, + 489, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 308, + 349, + 489, + 361 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 308, + 349, + 489, + 361 + ], + "type": "text", + "content": " is the weight for the geometric loss." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 373, + 388, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 388, + 386 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 388, + 386 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 394, + 547, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 394, + 547, + 490 + ], + "spans": [ + { + "bbox": [ + 305, + 394, + 547, + 490 + ], + "type": "text", + "content": "We have introduced a self-supervised method to learn camera pose and sound localization from audio-visual data. In experiments, we first evaluate how well our learned representation captures spatial information. We then evaluate how well our method learns camera pose and sound localization by comparing it with baselines. Finally, we show generalization to indoor panorama images Stanford2D3D [6] and in-the-wild binaural audio [17]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 501, + 408, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 501, + 408, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 501, + 408, + 514 + ], + "type": "text", + "content": "4.1. Implementations" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 518, + 547, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 518, + 547, + 615 + ], + "spans": [ + { + "bbox": [ + 305, + 518, + 547, + 615 + ], + "type": "text", + "content": "Visual pose encoder. We follow recent pose estimation work [9, 43] and build a Siamese-style visual pose network " + }, + { + "bbox": [ + 305, + 518, + 547, + 615 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 305, + 518, + 547, + 615 + ], + "type": "text", + "content": " with ResNet-18 [38] as the backbone. We compute dense 4D correlation volumes between the features from the third residual layer and then encode them by convolution layers. We resize images to " + }, + { + "bbox": [ + 305, + 518, + 547, + 615 + ], + "type": "inline_equation", + "content": "320 \\times 240" + }, + { + "bbox": [ + 305, + 518, + 547, + 615 + ], + "type": "text", + "content": " and encode a pair of images into 512-d features. We use an MLP " + }, + { + "bbox": [ + 305, + 518, + 547, + 615 + ], + "type": "inline_equation", + "content": "g_{v}" + }, + { + "bbox": [ + 305, + 518, + 547, + 615 + ], + "type": "text", + "content": " to map visual features to 1-d logits for our SLfM models." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 625, + 547, + 722 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 625, + 547, + 722 + ], + "spans": [ + { + "bbox": [ + 305, + 625, + 547, + 722 + ], + "type": "text", + "content": "Binaural audio encoder. We obtain binaural audio embeddings " + }, + { + "bbox": [ + 305, + 625, + 547, + 722 + ], + "type": "inline_equation", + "content": "f_{a}(\\cdot)" + }, + { + "bbox": [ + 305, + 625, + 547, + 722 + ], + "type": "text", + "content": " using ResNet-18 [38] that operates on spectrograms. We covert the two-channel waveform of length " + }, + { + "bbox": [ + 305, + 625, + 547, + 722 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 625, + 547, + 722 + ], + "type": "text", + "content": " to a spectrogram representation of size " + }, + { + "bbox": [ + 305, + 625, + 547, + 722 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 4" + }, + { + "bbox": [ + 305, + 625, + 547, + 722 + ], + "type": "text", + "content": " using short-time Fourier transform (STFT), where we keep both the magnitude and phase of spectrograms. We extract 512-d features of binaural sound with " + }, + { + "bbox": [ + 305, + 625, + 547, + 722 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 305, + 625, + 547, + 722 + ], + "type": "text", + "content": " and map them to 1-d logits using an MLP " + }, + { + "bbox": [ + 305, + 625, + 547, + 722 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 305, + 625, + 547, + 722 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 693, + 288, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 288, + 714 + ], + "type": "text", + "content": "1Work on self-supervised SfM has similar ambiguities [104], and deals with them by adding analogous constraints, such as photometric consistency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "spans": [ + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "type": "text", + "content": "7900" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "Audio prediction model. We adopt the light-weighted audio-visual U-Net [30] to perform binauralization. We feed in spectrograms of size " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 2" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " and predict the target spectrograms. We concatenate the visual pose features " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "f_{v}(\\cdot)" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " and audio features " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "f_{a}(\\cdot)" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " at the bottleneck of U-Net. We show the architecture of our models in Fig. 3. Please see supp. for more implementation details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 163, + 105, + 174 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 105, + 174 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 105, + 174 + ], + "type": "text", + "content": "4.2. Dataset" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 182, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 289, + 300 + ], + "type": "text", + "content": "Since there is no public multi-view audio-visual dataset with camera poses and sound direction ground truth, we use the SoundSpaces 2.0 platform [15] to create a dataset. Our 3D scenes come from Habitat-Matterport 3D dataset (HM3D) [76], which is a large dataset of real 3D scenes. This setup allows us to have photorealistic images and high-quality spatial audio with real-world acoustics phenomenon (e.g., reverberation), as well as providing the ground-truth camera pose and sound directions that can be used for evaluation. We call this dataset HM3D-SS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 301, + 289, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 289, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 289, + 491 + ], + "type": "text", + "content": "We generate binaural Room Impulse Responses (RIRs) and images with a " + }, + { + "bbox": [ + 46, + 301, + 289, + 491 + ], + "type": "inline_equation", + "content": "60^{\\circ}" + }, + { + "bbox": [ + 46, + 301, + 289, + 491 + ], + "type": "text", + "content": " field of view, using 100 scenes of HM3D [76]. For each audio-visual example, we randomly place sound sources in the scene with a height range of (0.7, 1.7) meters, and sample 4 different rotated viewpoints at one location within 4 meters. The rotations are limited to " + }, + { + "bbox": [ + 46, + 301, + 289, + 491 + ], + "type": "inline_equation", + "content": "(10^{\\circ}, 90^{\\circ})" + }, + { + "bbox": [ + 46, + 301, + 289, + 491 + ], + "type": "text", + "content": " relative to the source viewpoints. We follow the standard practice to set the height to agents to be " + }, + { + "bbox": [ + 46, + 301, + 289, + 491 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{m}" + }, + { + "bbox": [ + 46, + 301, + 289, + 491 + ], + "type": "text", + "content": " and lock a downward tilt angle [43, 101, 73, 102]. We render the binaural RIRs and images given the position of agents and sound sources. We obtain binaural audio by convolving binaural RIRs with mono audio samples from LibriSpeech [69] and Free Music Archive [20]. To ensure that the evaluation tests the model's pose estimation abilities, rather than its ability to visually localize sound sources, the sound sources are not visible on screen." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 492, + 288, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 492, + 288, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 492, + 288, + 540 + ], + "type": "text", + "content": "We create 50K audio-visual pairs from 200K viewpoints. The audio was rendered with average reverberation of " + }, + { + "bbox": [ + 46, + 492, + 288, + 540 + ], + "type": "inline_equation", + "content": "\\mathrm{RT}_{60} = 0.4\\mathrm{s}" + }, + { + "bbox": [ + 46, + 492, + 288, + 540 + ], + "type": "text", + "content": " (see supp. for details). We divided our data into 81/9/10 scenes for the train/val/test, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 547, + 247, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 547, + 247, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 547, + 247, + 559 + ], + "type": "text", + "content": "4.3. Evaluating the learned representation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 565, + 288, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 565, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 565, + 288, + 613 + ], + "type": "text", + "content": "First, we directly evaluate the quality of our learned features for rotation estimation and sound localization via linear probing with labeled data (rather than learning them jointly through self-supervision)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": "Baselines and ablations. We compare our model with several baselines that use alternative pretext tasks: 1) AVSA [63]: it learns spatial cues by training a model to spatially align video and audio clips extracted from different viewing angles. We adapt this model to our dataset and train with 4 different views; 2) RotNCE [27]: it applies contrastive learning on the audio from different angles and uses annotations of the agent's rotation to select positive" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "and negative samples, which results in learning audio spatial representation. For baselines, we use the same architecture for feature extractors to ensure fair comparisons." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 108, + 547, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 547, + 191 + ], + "type": "text", + "content": "To determine if we utilize visual and audio features from different views to solve the binauralization task, we also study some variants of our models: 1) Ours-NoA: we only provide visual features for the binauralization task; 2) Ours-NoV: which only uses audio from the other view to spatialize sounds; 3) Ours-GTRot: we provide ground-truth rotation embedding instead of features from visual frames." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 192, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 547, + 240 + ], + "type": "text", + "content": "Besides the mono-to-binaural task, we also experiment with another objective: predicting the right channel from the left channel. We train our L2R model with the same setup as our M2B model. (Please see supp. for pretext results.)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 244, + 548, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 244, + 548, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 548, + 449 + ], + "type": "text", + "content": "Downstream tasks. We assess the quality of spatial representations we learned from our pretext tasks in two downstream tasks: relative camera rotation and 3D sound localization. We formulate them as classification problems, where angles are categorized into 64 bins, and we use accuracy as the evaluation metric. To evaluate the learned features, we freeze them and train a linear classifier on the downstream tasks. We compare the performance of our features with those learned from RotNCE [27], AVSA [63], ImageNet [37], and random features, and report the results in Tab. 1. Our approach outperforms the baselines in both tasks, indicating that we learn better spatial representations. Furthermore, our linear probe models show comparable performance against the supervised method which can be regarded as approximate upper bounds for our models, suggesting that our pretext tasks help learn a useful representation. Please see supp. for experiments on FreeMusic [20]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 453, + 547, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 547, + 514 + ], + "type": "text", + "content": "Emerging camera pose from audio prompting. To help understand the strong performance of our self-supervised features, we asked whether we could use the cross-view binauralization model alone to estimate camera rotation. Inspired by prompting in vision and language models [75], we" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 315, + 520, + 538, + 681 + ], + "blocks": [ + { + "bbox": [ + 315, + 520, + 538, + 681 + ], + "lines": [ + { + "bbox": [ + 315, + 520, + 538, + 681 + ], + "spans": [ + { + "bbox": [ + 315, + 520, + 538, + 681 + ], + "type": "table", + "html": "
ModelAudio Loc. Acc (%) ↑Camera Rot. Acc (%) ↑
LibriSpeechRandom feature4.84.7
ImageNet [38]+Random-56.3
RotNCE [27]50.9-
AVSA [63]71.26.5
Ours-NoA-9.6
Ours-NoV53.9-
Ours-GTRot70.6-
Ours-L2R (3 views)78.576.1
Ours (2 views)74.580.0
Ours (3 views)75.481.3
Supervised81.595.8
", + "image_path": "d6e46f3615fe29d8bbb9f3520c201da0e9d102552ff848042d2ff86913d3df5b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 684, + 547, + 716 + ], + "lines": [ + { + "bbox": [ + 304, + 684, + 547, + 716 + ], + "spans": [ + { + "bbox": [ + 304, + 684, + 547, + 716 + ], + "type": "text", + "content": "Table 1: Downstream task performance on HM3D-SS dataset. We report linear probe performance on the audio localization and camera rotation downstream tasks." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 752, + 314, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 752, + 314, + 761 + ], + "spans": [ + { + "bbox": [ + 295, + 752, + 314, + 761 + ], + "type": "text", + "content": "7901" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 56, + 274, + 332 + ], + "blocks": [ + { + "bbox": [ + 62, + 56, + 274, + 332 + ], + "lines": [ + { + "bbox": [ + 62, + 56, + 274, + 332 + ], + "spans": [ + { + "bbox": [ + 62, + 56, + 274, + 332 + ], + "type": "table", + "html": "
ModelAudio angle MAE (°) ↓Camera angle MAE (°) ↓
LibriSpeechChance40.2829.41
SIFT [54]-12.2
Ours w/o Reflect28.0826.80
Ours-Prompt-9.13
Ours-L2R3.220.80
Ours3.170.77
Ours-Front4.483.08
Ours-GTRot1.83-
Superglue [81]-2.47
Supervised1.710.46
FreeMusicChance40.8129.41
SIFT [54]-12.2
Ours w/o Reflect28.2426.81
Ours-L2R3.180.81
Ours3.370.84
Ours-Front3.992.49
Ours-GTRot1.96-
Superglue [81]-2.47
Supervised2.500.46
", + "image_path": "c934e608082ec2c72295e625821dd146fc1966214a19ef3b6ba7d33653e35bdf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 373, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 289, + 529 + ], + "type": "text", + "content": "obtain rough estimates of camera rotation by providing our model with carefully-provided inputs. Given a pair of images " + }, + { + "bbox": [ + 46, + 373, + 289, + 529 + ], + "type": "inline_equation", + "content": "(\\mathbf{v}_s,\\mathbf{v}_t)" + }, + { + "bbox": [ + 46, + 373, + 289, + 529 + ], + "type": "text", + "content": ", we create a synthetic binaural audio prompt, " + }, + { + "bbox": [ + 46, + 373, + 289, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_s" + }, + { + "bbox": [ + 46, + 373, + 289, + 529 + ], + "type": "text", + "content": ". We then ask our model to generate the binaural sound " + }, + { + "bbox": [ + 46, + 373, + 289, + 529 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}_t" + }, + { + "bbox": [ + 46, + 373, + 289, + 529 + ], + "type": "text", + "content": " for the target viewpoint. By analyzing the IID cues in " + }, + { + "bbox": [ + 46, + 373, + 289, + 529 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}_t" + }, + { + "bbox": [ + 46, + 373, + 289, + 529 + ], + "type": "text", + "content": ", we can estimate the model's implicitly predicted camera pose. To do this, we find the nearest neighbor of our generated audio " + }, + { + "bbox": [ + 46, + 373, + 289, + 529 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}_t" + }, + { + "bbox": [ + 46, + 373, + 289, + 529 + ], + "type": "text", + "content": ", using a database of synthetically generated audio with known sound directions. Please refer to supp. for more details. Our method achieves the mean absolute error of " + }, + { + "bbox": [ + 46, + 373, + 289, + 529 + ], + "type": "inline_equation", + "content": "9.13^{\\circ}" + }, + { + "bbox": [ + 46, + 373, + 289, + 529 + ], + "type": "text", + "content": " on HM3D-SS dataset where chance is " + }, + { + "bbox": [ + 46, + 373, + 289, + 529 + ], + "type": "inline_equation", + "content": "29.41^{\\circ}" + }, + { + "bbox": [ + 46, + 373, + 289, + 529 + ], + "type": "text", + "content": " (Tab. 2). Our approach can also generalize to Stanford2D3D [6] dataset, where we can achieve mean absolute error of " + }, + { + "bbox": [ + 46, + 373, + 289, + 529 + ], + "type": "inline_equation", + "content": "9.93^{\\circ}" + }, + { + "bbox": [ + 46, + 373, + 289, + 529 + ], + "type": "text", + "content": " (Tab. 3)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 536, + 151, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 151, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 151, + 548 + ], + "type": "text", + "content": "4.4. Evaluating SLfM" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 554, + 288, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 288, + 604 + ], + "type": "text", + "content": "We evaluated our SLfM models on the HM3D-SS dataset. We use the mean absolute error of angle in degrees as evaluation metrics. To avoid sound field ambiguity, we filtered out samples with sound angles outside of " + }, + { + "bbox": [ + 46, + 554, + 288, + 604 + ], + "type": "inline_equation", + "content": "(-90^{\\circ}, 90^{\\circ})" + }, + { + "bbox": [ + 46, + 554, + 288, + 604 + ], + "type": "text", + "content": " for the" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 609, + 187, + 680 + ], + "blocks": [ + { + "bbox": [ + 48, + 609, + 187, + 680 + ], + "lines": [ + { + "bbox": [ + 48, + 609, + 187, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 609, + 187, + 680 + ], + "type": "image", + "image_path": "faa58b777188becf848cf49185abfe74a4bf1acc96383789f8f4efdf03ee17b0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 684, + 289, + 718 + ], + "lines": [ + { + "bbox": [ + 46, + 684, + 289, + 718 + ], + "spans": [ + { + "bbox": [ + 46, + 684, + 289, + 718 + ], + "type": "text", + "content": "Figure 4: Qualitative results on real-world examples. We show our predictions on Stanford2D3D [6] and In-the-wild audio [17]. Green denotes accurate predictions." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 195, + 610, + 286, + 680 + ], + "blocks": [ + { + "bbox": [ + 195, + 610, + 286, + 680 + ], + "lines": [ + { + "bbox": [ + 195, + 610, + 286, + 680 + ], + "spans": [ + { + "bbox": [ + 195, + 610, + 286, + 680 + ], + "type": "image", + "image_path": "c88cdb70d7fb5ce1c5690c0479f41d2bff5c24a5534a56c9466f58e8fbe82010.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 57, + 422, + 163 + ], + "blocks": [ + { + "bbox": [ + 47, + 336, + 288, + 360 + ], + "lines": [ + { + "bbox": [ + 47, + 336, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 288, + 360 + ], + "type": "text", + "content": "Table 2: Sound localization from motion results on HM3D-SS. We evaluate our SLfM models on each modality independently." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 57, + 422, + 163 + ], + "lines": [ + { + "bbox": [ + 309, + 57, + 422, + 163 + ], + "spans": [ + { + "bbox": [ + 309, + 57, + 422, + 163 + ], + "type": "table", + "html": "
In the wild audio [17]
ModelAcc (%) ↑
GCC-PHAT [47]77.2
IID [17]75.4
IID (Ours)82.1
MonoCLR [17]87.4
StereoCRW [17]87.2
Ours-L2R84.9
Ours84.0
", + "image_path": "7da04a9e23a66d10293de4127afc3d375781214f80fda1a2501969cf0e8fcd89.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 433, + 59, + 547, + 163 + ], + "blocks": [ + { + "bbox": [ + 433, + 59, + 547, + 163 + ], + "lines": [ + { + "bbox": [ + 433, + 59, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 433, + 59, + 547, + 163 + ], + "type": "table", + "html": "
Stanford2D3D [6]
ModelRot Err. (°) ↓
MeanMed.
SIFT [54]16.40.06
LoFTR [85]6.101.13
Superglue [81]5.070.07
Ours-Prompt9.939.02
Ours-L2R1.120.71
Ours1.140.67
", + "image_path": "2863461e8b54c4039edde554fa0bd998d3f3a0236008f720fc07aeeb0d89e4f1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 167, + 547, + 212 + ], + "lines": [ + { + "bbox": [ + 304, + 167, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 547, + 212 + ], + "type": "text", + "content": "Table 3: Evaluation of our SLfM models on the real-world data. We evaluate our audio localization model on the In-the-wild audio [17] (left) and our camera pose model on the Stanford2D3D [6] (right). Rot. denote camera rotation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 225, + 547, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 225, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 225, + 547, + 285 + ], + "type": "text", + "content": "evaluation set. When training our SLfM models, we use our best-performing features, i.e., pretext tasks trained with 3 views. We freeze learned audio and visual features and only train multi-layer perceptrons on top of them and evaluate them independently on each modality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 289, + 547, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 547, + 420 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 547, + 420 + ], + "type": "text", + "content": "Baselines and ablations. For relative camera pose estimation, we compare our models with sparse feature matching using SIFT [54] and SuperGlue [81], followed by rotation fitting. We use those methods to detect key points, run Lowe's ratio test, and use RANSAC with five-point algorithm to recover the camera pose [54, 36, 66, 26]. For the sound localization task, we compared with time delay estimation methods: the popular GCC-PHAT [47] and the recent self-supervised method StereoCRW [17]. As far as we know, there are no other baselines that can estimate poses and sound source directions without labels." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 421, + 547, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 547, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 547, + 480 + ], + "type": "text", + "content": "We also compared several variations of our method, including Ours-Front, where we filter out the samples with sound sources behind the viewers to remove binaural ambiguity during training, and an oracle model Ours-GTRot, which uses ground-truth rotation angles instead." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 485, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 545 + ], + "type": "text", + "content": "Results. We show our results in Tab. 2. Our models can predict the azimuths of sound sources, obtaining strong performance without any labels. Without providing reflection invariance (Eq. (7)), the model failed to learn reasonable geometric due to the audio ambiguity. Our self-supervised" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 551, + 545, + 658 + ], + "blocks": [ + { + "bbox": [ + 307, + 551, + 545, + 658 + ], + "lines": [ + { + "bbox": [ + 307, + 551, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 551, + 545, + 658 + ], + "type": "image", + "image_path": "fb78b25bd41a287098b8e407f3c1ae28ef900d53731b9cde3e9364f7012ddef6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 662, + 547, + 718 + ], + "lines": [ + { + "bbox": [ + 304, + 662, + 547, + 718 + ], + "spans": [ + { + "bbox": [ + 304, + 662, + 547, + 718 + ], + "type": "text", + "content": "Figure 5: Real video. We play music using a speaker and record a video using iPhone with a binaural microphone (different from training examples). We show the predicted camera rotation and sound direction over time. They change smoothly and match our camera motion. Please see our project webpage for video results." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "spans": [ + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "type": "text", + "content": "7902" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 57, + 545, + 223 + ], + "blocks": [ + { + "bbox": [ + 49, + 57, + 545, + 223 + ], + "lines": [ + { + "bbox": [ + 49, + 57, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 49, + 57, + 545, + 223 + ], + "type": "image", + "image_path": "ba0393c760b52f00cb2e09a8a47564c6a1937c782a66b1523372d929855d79a7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 228, + 548, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 548, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 548, + 275 + ], + "type": "text", + "content": "Figure 6: Qualitative results on HM3D-SS. Given two images and their corresponding audio at the source viewpoints, our approach can individually predict the relative camera pose and localize sound locations accurately. We visualize our camera rotation prediction with blue bar and our sound angle prediction with orange bar. To highlight the subtle differences in the waveforms, we color code the amplitude. Please see the project webpage for more video results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 287, + 289, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 289, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 289, + 348 + ], + "type": "text", + "content": "model achieves comparable performance against our oracle model (Ours-GTRot) and supervised models, indicating we estimate camera poses and localize sound accurately. We show some qualitative results on HM3D-SS in Fig. 6 with LibriSpeech samples [69]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 354, + 289, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 289, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 289, + 545 + ], + "type": "text", + "content": "Generalization to other datasets. We further demonstrate the generalization ability of our models by experimenting with out-of-distribution, real-world data. We evaluate our camera rotation model on Stanford2D3D [6] with real indoor RGB images (Tab. 3). We obtain image pairs by cropping from panoramas. Although our model is trained on renderings of HM3D [76], it obtains strong generalization ability. Compared with rotation estimation based on SIFT [54] and SuperGlue [81], our model is significantly better on mean rotation error. We also report median rotation error to be consistent with prior works [9, 43]. However, SIFT [54] and Superglue [81] have a very low median error. This is likely due to the all-or-nothing nature of feature matching-based approaches, which either produce highly accurate predictions if the matches are correct (especially for \"easy\" cases with small amounts of rotation) or else produce gross errors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 290, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 290, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 290, + 713 + ], + "type": "text", + "content": "We also evaluate our sound localization model on the in-the-wild binaural audio [17]. We use binary accuracy as the metric for left-or-right direction classification accounting for the fact that microphone baselines are unknown for internet videos. For an apples-to-apples comparison to prior work, we retrained our model using 0.51s length of audio for the binaural audio encoder " + }, + { + "bbox": [ + 46, + 545, + 290, + 713 + ], + "type": "inline_equation", + "content": "f_{a}(\\cdot)" + }, + { + "bbox": [ + 46, + 545, + 290, + 713 + ], + "type": "text", + "content": ". As shown in Tab. 3, our model obtains similar performance to StereoCRW, a state-of-the-art self-supervised time delay method, suggesting we have a strong capability for sound localization. We show qualitative results in Fig. 4. We also perform both tasks on a self-recorded video (Fig. 5) of a rotating camera and a binaural microphone. We show the mean and standard deviation of predictions in 1.0s windows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 287, + 547, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 547, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 547, + 419 + ], + "type": "text", + "content": "Handling ambiguity. In binaural sound perception, there is a fundamental ambiguity that whether the sound is in front of or behind us. It leads to multiple solutions with equal loss in our model, mirroring sound sources and negating rotation angles with flipped " + }, + { + "bbox": [ + 304, + 287, + 547, + 419 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 287, + 547, + 419 + ], + "type": "text", + "content": " axis (Fig. 7). These two solutions differ in that a visual rotation angle either indicates a clockwise or counterclockwise rotation. For evaluation, we convert \"backwards\" counterclockwise predictions to clockwise predictions by simply providing the model with pairs of input frames, then negating the model's outputs if the angle is the opposite of the expected direction2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 426, + 525, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 426, + 525, + 440 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 525, + 440 + ], + "type": "text", + "content": "4.5. Generalization to more complex scenarios" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 445, + 547, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 547, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 547, + 480 + ], + "type": "text", + "content": "We investigate whether our approach generalizes to more complex scenarios, such as with multiple sound sources or when translation is included in camera motion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 484, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 547, + 533 + ], + "type": "text", + "content": "Multiple sound sources. We evaluate our representations and SLfM models on more complex scenes containing multiple sound sources. We train our models with two source sources placed in the scenes. One of them is dominant and" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 566, + 416, + 658 + ], + "blocks": [ + { + "bbox": [ + 307, + 566, + 416, + 658 + ], + "lines": [ + { + "bbox": [ + 307, + 566, + 416, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 566, + 416, + 658 + ], + "type": "image", + "image_path": "a369f32c41ea1be4885ae2b44e5496a654f5b3c613ab97b8dcdccd68b3d49923.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 662, + 418, + 717 + ], + "lines": [ + { + "bbox": [ + 305, + 662, + 418, + 717 + ], + "spans": [ + { + "bbox": [ + 305, + 662, + 418, + 717 + ], + "type": "text", + "content": "Figure 7: Mirror ambiguity. The " + }, + { + "bbox": [ + 305, + 662, + 418, + 717 + ], + "type": "inline_equation", + "content": "+z" + }, + { + "bbox": [ + 305, + 662, + 418, + 717 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 662, + 418, + 717 + ], + "type": "inline_equation", + "content": "+x" + }, + { + "bbox": [ + 305, + 662, + 418, + 717 + ], + "type": "text", + "content": " axes represent forward and rightward directions. Both solutions have the same loss." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 436, + 567, + 545, + 658 + ], + "blocks": [ + { + "bbox": [ + 436, + 567, + 545, + 658 + ], + "lines": [ + { + "bbox": [ + 436, + 567, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 436, + 567, + 545, + 658 + ], + "type": "image", + "image_path": "528096eb7f258d146a61014842594bc58c8971bfb3ba6e990a3b0ea38b9a50bb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 662, + 547, + 706 + ], + "lines": [ + { + "bbox": [ + 435, + 662, + 547, + 706 + ], + "spans": [ + { + "bbox": [ + 435, + 662, + 547, + 706 + ], + "type": "text", + "content": "Figure 8: Robustness to reverberation. We study the effect of reverberation on our SLfM model." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 539, + 545, + 559 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 545, + 559 + ], + "type": "text", + "content": "2Similar to ambiguities in SfM where reconstructions can be reoriented such that the sky is in the positive " + }, + { + "bbox": [ + 306, + 539, + 545, + 559 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 306, + 539, + 545, + 559 + ], + "type": "text", + "content": " direction [36, 35, 4]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "spans": [ + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "type": "text", + "content": "7903" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 58, + 272, + 169 + ], + "blocks": [ + { + "bbox": [ + 62, + 58, + 272, + 169 + ], + "lines": [ + { + "bbox": [ + 62, + 58, + 272, + 169 + ], + "spans": [ + { + "bbox": [ + 62, + 58, + 272, + 169 + ], + "type": "table", + "html": "
ModelMulti-source Acc (%) ↑Small trans. Acc (%) ↑
Aud.Rot.Aud.Rot.
Random feature4.54.75.73.8
ImageNet [38]+Random-56.3-23.0
RotNCE [27]35.8---
AVSA [63]55.96.765.46.1
Ours (3 views)59.277.972.249.6
Supervised74.595.881.663.5
", + "image_path": "ff2232dbec0f6fab7c95c7ddd56807ceb05f052012826069cfafe7378ccdb183.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 71, + 222, + 264, + 340 + ], + "blocks": [ + { + "bbox": [ + 46, + 174, + 287, + 218 + ], + "lines": [ + { + "bbox": [ + 46, + 174, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 287, + 218 + ], + "type": "text", + "content": "Table 4: Generalization to more complex scenarios. We evaluate the audio and visual features learned from more complex scenarios via linear probing. Aud. and Rot. denote audio localization and camera rotation respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 222, + 264, + 340 + ], + "lines": [ + { + "bbox": [ + 71, + 222, + 264, + 340 + ], + "spans": [ + { + "bbox": [ + 71, + 222, + 264, + 340 + ], + "type": "table", + "html": "
ModelMulti-source MAE (°) ↓Small trans. MAE (°) ↓
Aud.Rot.Aud.Rot.
Chance40.4629.4140.2829.41
SIFT [54]-12.2-12.2
Ours7.670.716.281.04
Ours-GTRot5.81-4.24-
Superglue [81]-2.47-2.47
Supervised3.600.461.710.46
", + "image_path": "e037ac7b412f889c94cbf4173cd5757c5290495344d34f8b3afcc9934e01dfe5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 394, + 287, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 394, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 287, + 455 + ], + "type": "text", + "content": "our sound localization target. We report the results in Tab. 4 and Tab. 5. Our model learns better representations than baselines and achieves accurate predictions of the azimuth of the source and camera pose even in challenging scenarios with multiple sound sources." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 460, + 288, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 288, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 288, + 640 + ], + "type": "text", + "content": "Translation in camera motions. To study how small translations in the camera motion could affect our models, we generate 50K pairs of audio-visual data with both rotation and translation change, limiting the uniformly sampled translation to 0.5 meters. We train our models with LibriSpeech samples [69]. For our linear probing evaluation, we measure the ability of features to handle complex examples by testing on the dataset that has translation. For our SLfM model, we study our ability to learn from noisy data, thus we evaluate it on the rotation-only examples. As the results are shown in Tab. 4 and Tab. 5, we successfully learn useful features and obtain accurate rotation and sound direction predictions despite the presence of translation. Since we jointly learn the audio and visual representations, it can negatively impact the learning of one modality when another one becomes harder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 648, + 138, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 648, + 138, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 648, + 138, + 661 + ], + "type": "text", + "content": "4.6. Ablation study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "Robustness to reverberation. We study how our SLfM models perform under the different reverberation configurations. We used SoundSpaces [15] to create audio with average reverberation " + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{RT}_{60} \\in \\{0.1, 0.2, 0.3, 0.4, 0.5\\}" + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": " while" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 58, + 543, + 133 + ], + "blocks": [ + { + "bbox": [ + 46, + 344, + 287, + 389 + ], + "lines": [ + { + "bbox": [ + 46, + 344, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 287, + 389 + ], + "type": "text", + "content": "Table 5: SLfM results on more complex scenarios. We evaluate our model on the version of HM3D-SS with two sound sources. We also evaluate our model trained with small translations on rotation-only examples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 58, + 543, + 133 + ], + "lines": [ + { + "bbox": [ + 309, + 58, + 543, + 133 + ], + "spans": [ + { + "bbox": [ + 309, + 58, + 543, + 133 + ], + "type": "table", + "html": "
ModelLossesAudio angle MAE (°) ↓Camera angle MAE (°) ↓
LgeoLbinauralLsym
Ours✓37.6029.20
✓✓37.5229.17
✓✓3.586.99
✓✓✓3.170.77
", + "image_path": "dd5728e202d45ea862879aa48f746fff14199316565d1aefb2644cdc40395a98.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 137, + 545, + 159 + ], + "lines": [ + { + "bbox": [ + 305, + 137, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 305, + 137, + 545, + 159 + ], + "type": "text", + "content": "Table 6: Ablation experiments on our SLfM losses. We evaluate our SLfM models with different combinations of losses." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 167, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 167, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 545, + 228 + ], + "type": "text", + "content": "keeping the visual signals the same. We train our model on each setting. As shown in Fig. 8, our performance decreases as the level of reverberation increases, where audio becomes more challenging during both training and testing. Please see supp. for the study on representations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 232, + 545, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 280 + ], + "type": "text", + "content": "Losses for SLfM. We study the necessity of our proposed loss functions in Tab. 6. Our models fail to learn accurate pose estimation without binaural loss or symmetric loss. It highlights the crucial role of these losses." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 290, + 378, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 290, + 378, + 303 + ], + "spans": [ + { + "bbox": [ + 306, + 290, + 378, + 303 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 310, + 547, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 547, + 502 + ], + "type": "text", + "content": "In this paper, we proposed the sound localization from motion (SLfM) problem, and provided a self-supervised method for solving it. We also presented a method for learning audio-visual features that convey sound directions and camera rotation, which we show are well-suited to solving the SLfM task. Despite learning our models solely from unlabeled audio-visual data, we obtain strong performance on a variety of benchmarks, including rotation estimation on the Stanford2D3D [6] dataset and \"in the wild\" sound direction estimation [17]. Our results suggest that the subtle correlations between sights and binaural sounds that result from rotational motion provide a useful (and previously unused) learning signal. We see our work as opening new directions in self-supervised geometry estimation and feature learning that use sound as a complementary source of supervision. We will release code, data, and models upon acceptance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 506, + 547, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 613 + ], + "type": "text", + "content": "Limitations and Broader Impacts. Our work has several limitations. First, while we evaluate our models on real images and sounds, we train on data from simulators, due to a lack of available relevant data. We note that this is common practice in visual 3D reconstruction [43, 91, 52]. Second, we assume that the 3D scene and sound sources are stationary. Third, we do not evaluate our model on extreme viewpoint changes [55], which requires reasoning about images that have little or no overlap." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgements. We would like to thank Ang Cao, Tiange Luo, and Linyi Jin for their helpful discussion. We thank Changan Chen for his help with SoundSpaces 2.0. This work was funded in part by DARPA Semafor and Sony. The views, opinions and/or findings expressed are those of the authors and should not be interpreted as representing the official views or policies of the Department of Defense or the U.S. Government." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "spans": [ + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "type": "text", + "content": "7904" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 58, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Sharath Adavanne, Archontis Politis, and Tuomas Virtanen. Direction of arrival estimation for multiple sound sources using convolutional recurrent neural network. In 2018 26th European Signal Processing Conference (EUSIPCO), pages 1462-1466. IEEE, 2018. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 147, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 147, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 58, + 147, + 288, + 201 + ], + "type": "text", + "content": "[2] Triantafyllos Afouras, Yuki M Asano, Francois Fagan, Andrea Vedaldi, and Florian Metze. Self-supervised object detection from audio-visual correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10575–10586, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 202, + 288, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 202, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 58, + 202, + 288, + 245 + ], + "type": "text", + "content": "[3] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. arXiv preprint arXiv:2008.04237, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 247, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 247, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 58, + 247, + 288, + 290 + ], + "type": "text", + "content": "[4] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54(10):105-112, 2011. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 291, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 291, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 58, + 291, + 287, + 335 + ], + "type": "text", + "content": "[5] Juan León Alcazar, Fabian Caba, Ali K Thabet, and Bernard Ghanem. Maas: Multi-modal assignation for active speaker detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 265-274, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 336, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 336, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 58, + 336, + 288, + 369 + ], + "type": "text", + "content": "[6] Iro Armeni, Sasha Sax, Amir R Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105, 2017. 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 370, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 370, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 58, + 370, + 288, + 413 + ], + "type": "text", + "content": "[7] Yuki Asano, Mandela Patrick, Christian Rupprecht, and Andrea Vedaldi. Labelling unlabelled videos from scratch with multi-modal self-supervision. Advances in Neural Information Processing Systems, 33:4660-4671, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 415, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 415, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 58, + 415, + 288, + 446 + ], + "type": "text", + "content": "[8] Matthew Brown and David G Lowe. Automatic panoramic image stitching using invariant features. International journal of computer vision, 74:59-73, 2007. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 448, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 448, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 58, + 448, + 287, + 502 + ], + "type": "text", + "content": "[9] Ruojin Cai, Bharath Hariharan, Noah Snavely, and Hadar Averbuch-Elor. Extreme rotation estimation using dense correlation volumes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14566-14575, 2021. 2, 4, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 503, + 288, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 503, + 288, + 546 + ], + "spans": [ + { + "bbox": [ + 53, + 503, + 288, + 546 + ], + "type": "text", + "content": "[10] Changan Chen, Ziad Al-Halah, and Kristen Grauman. Semantic audio-visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15516–15525, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 548, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 548, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 53, + 548, + 288, + 591 + ], + "type": "text", + "content": "[11] Changan Chen, Ruohan Gao, Paul Calamia, and Kristen Grauman. Visual acoustic matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18858-18868, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 592, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 592, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 53, + 592, + 288, + 669 + ], + "type": "text", + "content": "[12] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part VI 16, pages 17–36. Springer, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "type": "text", + "content": "[13] Changan Chen, Sagnik Majumder, Ziad Al-Halah, Ruohan Gao, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Learning to set waypoints for audio-visual navigation. arXiv preprint arXiv:2008.09622, 2020. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 312, + 73, + 547, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 547, + 117 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 547, + 117 + ], + "type": "text", + "content": "[14] Changan Chen, Alexander Richard, Roman Shapovalov, Vamsi Krishna Ithapu, Natalia Neverova, Kristen Grauman, and Andrea Vedaldi. Novel-view acoustic synthesis. arXiv preprint arXiv:2301.08730, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 118, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 547, + 173 + ], + "type": "text", + "content": "[15] Changan Chen, Carl Schissler, Sanchit Garg, Philip Kobernik, Alexander Clegg, Paul Calamia, Dhruv Batra, Philip W Robinson, and Kristen Grauman. Soundspaces 2.0: A simulation platform for visual-acoustic learning. arXiv preprint arXiv:2206.08312, 2022. 2, 5, 8, 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 175, + 547, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 175, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 312, + 175, + 547, + 229 + ], + "type": "text", + "content": "[16] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16867-16876, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 230, + 546, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 230, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 312, + 230, + 546, + 285 + ], + "type": "text", + "content": "[17] Ziyang Chen, David F Fouhey, and Andrew Owens. Sound localization by self-supervised time delay estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVI, pages 489-508. Springer, 2022. 2, 4, 6, 7, 8, 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 285, + 546, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 285, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 312, + 285, + 546, + 319 + ], + "type": "text", + "content": "[18] Ziyang Chen, Xixi Hu, and Andrew Owens. Structure from silence: Learning scene structure from ambient sound. arXiv preprint arXiv:2111.05846, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 320, + 546, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 320, + 546, + 353 + ], + "spans": [ + { + "bbox": [ + 312, + 320, + 546, + 353 + ], + "type": "text", + "content": "[19] Jesper Haahr Christensen, Sascha Hornauer, and Stella Yu. Batvision with gcc-phat features for better sound to vision predictions. arXiv preprint arXiv:2006.07995, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 354, + 546, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 354, + 546, + 387 + ], + "spans": [ + { + "bbox": [ + 312, + 354, + 546, + 387 + ], + "type": "text", + "content": "[20] Michael Defferrard, Kirell Benzi, Pierre Vandergheynst, and Xavier Bresson. Fma: A dataset for music analysis. arXiv preprint arXiv:1612.01840, 2016. 5, 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 388, + 546, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 388, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 312, + 388, + 546, + 431 + ], + "type": "text", + "content": "[21] Yuexi Du, Ziyang Chen, Justin Salamon, Bryan Russell, and Andrew Owens. Conditional generation of audio from video via foley analogies. Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 433, + 546, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 433, + 546, + 465 + ], + "spans": [ + { + "bbox": [ + 312, + 433, + 546, + 465 + ], + "type": "text", + "content": "[22] Mohamed El Banani, Luya Gao, and Justin Johnson. Unsupervisedr&r: Unsupervised point cloud registration via differentiable rendering. In CVPR, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 467, + 546, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 467, + 546, + 488 + ], + "spans": [ + { + "bbox": [ + 312, + 467, + 546, + 488 + ], + "type": "text", + "content": "[23] Mohamed El Banani and Justin Johnson. Bootstrap your own correspondences. In ICCV, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 490, + 547, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 490, + 547, + 522 + ], + "spans": [ + { + "bbox": [ + 312, + 490, + 547, + 522 + ], + "type": "text", + "content": "[24] Chao Feng, Ziyang Chen, and Andrew Owens. Self-supervised video forensics by audio-visual anomaly detection. arXiv preprint arXiv:2301.01767, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 523, + 546, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 523, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 312, + 523, + 546, + 567 + ], + "type": "text", + "content": "[25] Tim Fischer, Marco Caversaccio, and Wilhelm Wimmer. A front-back confusion metric in horizontal sound localization: The fbc score. In ACM Symposium on Applied Perception 2020, pages 1-5, 2020. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 568, + 546, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 568, + 546, + 612 + ], + "spans": [ + { + "bbox": [ + 312, + 568, + 546, + 612 + ], + "type": "text", + "content": "[26] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981. 6, 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 613, + 546, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 613, + 546, + 647 + ], + "spans": [ + { + "bbox": [ + 312, + 613, + 546, + 647 + ], + "type": "text", + "content": "[27] Andrew Francl. Modeling and Evaluating Human Sound Localization in the Natural Environment. PhD thesis, Massachusetts Institute of Technology, 2022. 2, 5, 8, 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 647, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 647, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 647, + 547, + 712 + ], + "type": "text", + "content": "[28] Chuang Gan, Deng Huang, Peihao Chen, Joshua B Tenenbaum, and Antonio Torralba. Foley music: Learning to generate music from videos. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 758-775. Springer, 2020. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "spans": [ + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "type": "text", + "content": "7905" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 72, + 289, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 289, + 139 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 289, + 139 + ], + "type": "text", + "content": "[29] Ruohan Gao, Changan Chen, Ziad Al-Halah, Carl Schissler, and Kristen Grauman. Visualechoes: Spatial image representation learning through echolocation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16, pages 658-676. Springer, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 140, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 140, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 53, + 140, + 288, + 183 + ], + "type": "text", + "content": "[30] Ruohan Gao and Kristen Grauman. 2.5 d visual sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 324-333, 2019. 2, 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 184, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 184, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 184, + 288, + 237 + ], + "type": "text", + "content": "[31] Ruohan Gao and Kristen Grauman. Visualvoice: Audiovisual speech separation with cross-modal consistency. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15490-15500. IEEE, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 238, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 288, + 282 + ], + "type": "text", + "content": "[32] Rishabh Garg, Ruohan Gao, and Kristen Grauman. Geometry-aware multi-task learning for binaural audio generation from video. arXiv preprint arXiv:2111.10882, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 283, + 288, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 283, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 53, + 283, + 288, + 326 + ], + "type": "text", + "content": "[33] Sanchita Ghose and John Jeffrey Prevost. Auto Foley: Artificial synthesis of synchronized sound tracks for silent videos with deep learning. IEEE Transactions on Multimedia, 23:1895-1907, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 327, + 288, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 327, + 288, + 382 + ], + "spans": [ + { + "bbox": [ + 53, + 327, + 288, + 382 + ], + "type": "text", + "content": "[34] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14950-14962, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 383, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 383, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 383, + 288, + 403 + ], + "type": "text", + "content": "[35] R. I. Hartley. Chirality. International Journal of Computer Vision (IJCV), 1998. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 404, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 288, + 437 + ], + "type": "text", + "content": "[36] R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 437, + 288, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 437, + 288, + 481 + ], + "spans": [ + { + "bbox": [ + 53, + 437, + 288, + 481 + ], + "type": "text", + "content": "[37] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In International Conference on Computer Vision (ICCV), 2015. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 482, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 482, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 53, + 482, + 288, + 514 + ], + "type": "text", + "content": "[38] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition (CVPR), 2016. 4, 5, 8, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 515, + 288, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 515, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 53, + 515, + 288, + 548 + ], + "type": "text", + "content": "[39] Xixi Hu, Ziyang Chen, and Andrew Owens. Mix and localize: Localizing sound sources in mixtures. Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 548, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 548, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 53, + 548, + 288, + 602 + ], + "type": "text", + "content": "[40] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. arXiv preprint arXiv:2301.12661, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 603, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 603, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 603, + 288, + 635 + ], + "type": "text", + "content": "[41] Wen Chin Huang, Dejan Markovic, Alexander Richard, Israel Dejene Gebru, and Anjali Menon. End-to-end binaural speech synthesis. arXiv preprint arXiv:2207.03697, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 636, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 636, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 636, + 288, + 667 + ], + "type": "text", + "content": "[42] Vladimir Iashin and Esa Rahtu. Taming visually guided sound generation. arXiv preprint arXiv:2110.08791, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 669, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 669, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 669, + 288, + 714 + ], + "type": "text", + "content": "[43] Linyi Jin, Shengyi Qian, Andrew Owens, and David F Fouhey. Planar surface reconstruction from sparse views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12991-13000, 2021. 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 72, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 312, + 72, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 547, + 116 + ], + "type": "text", + "content": "[44] Linyi Jin, Jianming Zhang, Yannick Hold-Geoffroy, Oliver Wang, Kevin Matzen, Matthew Sticha, and David F Fouhey. Perspective fields for single image camera calibration. arXiv, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 118, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 547, + 150 + ], + "type": "text", + "content": "[45] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In ICCV, 2015. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 152, + 547, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 547, + 185 + ], + "type": "text", + "content": "[46] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representation, 2015. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 186, + 547, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 186, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 312, + 186, + 547, + 229 + ], + "type": "text", + "content": "[47] Charles Knapp and Grifford Carter. The generalized correlation method for estimation of time delay. IEEE transactions on acoustics, speech, and signal processing, 24(4):320-327, 1976. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 230, + 547, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 230, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 312, + 230, + 547, + 263 + ], + "type": "text", + "content": "[48] Takashi Konno, Kenji Nishida, Katsutoshi Itoyama, and Kazuhiro Nakadai. Audio-visual sfm towards 4d reconstruction under dynamic scenes. 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 264, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 264, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 312, + 264, + 547, + 308 + ], + "type": "text", + "content": "[49] Felix Kreuk, Gabriel Synnaeve, Adam Polyak, Uriel Singer, Alexandre Défossez, Jade Copet, Devi Parikh, Yaniv Taigman, and Yossi Adi. Audiogen: Textually guided audio generation. arXiv preprint arXiv:2209.15352, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 309, + 547, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 309, + 547, + 353 + ], + "spans": [ + { + "bbox": [ + 312, + 309, + 547, + 353 + ], + "type": "text", + "content": "[50] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Av-nerf: Learning neural fields for real-world audio-visual scene synthesis. arXiv preprint arXiv:2302.02088, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 354, + 547, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 354, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 312, + 354, + 547, + 407 + ], + "type": "text", + "content": "[51] Yan-Bo Lin and Yu-Chiang Frank Wang. Exploiting audiovisual consistency with partial supervision for spatial audio generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2056–2063, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 409, + 547, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 409, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 312, + 409, + 547, + 443 + ], + "type": "text", + "content": "[52] Zhi-Hao Lin, Wei-Chiu Ma, Hao-Yu Hsu, Yu-Chiang Frank Wang, and Shenlong Wang. Neurmips: Neural mixture of planar experts for view synthesis. In CVPR, 2022. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 444, + 547, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 444, + 547, + 466 + ], + "spans": [ + { + "bbox": [ + 312, + 444, + 547, + 466 + ], + "type": "text", + "content": "[53] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 467, + 547, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 467, + 547, + 499 + ], + "spans": [ + { + "bbox": [ + 312, + 467, + 547, + 499 + ], + "type": "text", + "content": "[54] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 500, + 547, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 500, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 312, + 500, + 547, + 544 + ], + "type": "text", + "content": "[55] Wei-Chiu Ma, Anqi Joyce Yang, Shenlong Wang, Raquel Urtasun, and Antonio Torralba. Virtual correspondence: Humans as a cue for extreme-view geometry. In CVPR, 2022. 2, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 545, + 547, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 547, + 590 + ], + "type": "text", + "content": "[56] Sagnik Majumder, Ziad Al-Halah, and Kristen Grauman. Move2hear: Active audio-visual source separation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 275–285, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 590, + 547, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 590, + 547, + 633 + ], + "spans": [ + { + "bbox": [ + 312, + 590, + 547, + 633 + ], + "type": "text", + "content": "[57] Sagnik Majumder, Changan Chen, Ziad Al-Halah, and Kristen Grauman. Few-shot audio-visual learning of environment acoustics. arXiv preprint arXiv:2206.04006, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 635, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 635, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 547, + 689 + ], + "type": "text", + "content": "[58] Sagnik Majumder and Kristen Grauman. Active audio-visual separation of dynamic sound sources. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXIX, pages 551–569. Springer, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 690, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 690, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 690, + 547, + 714 + ], + "type": "text", + "content": "[59] Iaroslav Melekhov, Juha Ylioinas, Juho Kannala, and Esa Rahtu. Relative camera pose estimation using convolutional" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 752, + 316, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 752, + 316, + 761 + ], + "spans": [ + { + "bbox": [ + 295, + 752, + 316, + 761 + ], + "type": "text", + "content": "7906" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 72, + 288, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 288, + 117 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 288, + 117 + ], + "type": "text", + "content": "neural networks. In Advanced Concepts for Intelligent Vision Systems: 18th International Conference, ACIVS 2017, Antwerp, Belgium, September 18-21, 2017, Proceedings 18, pages 675-687. Springer, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 288, + 162 + ], + "type": "text", + "content": "[60] Himangi Mittal, Pedro Morgado, Unnat Jain, and Abhinav Gupta. Learning state-aware visual representations from audible interactions. arXiv preprint arXiv:2209.13583, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 163, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 163, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 288, + 196 + ], + "type": "text", + "content": "[61] Shentong Mo and Pedro Morgado. A closer look at weakly-supervised audio-visual source localization. arXiv preprint arXiv:2209.09634, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 198, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 198, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 288, + 251 + ], + "type": "text", + "content": "[62] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 218-234. Springer, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 253, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 253, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 253, + 288, + 297 + ], + "type": "text", + "content": "[63] Pedro Morgado, Yi Li, and Nuno Nvasconcelos. Learning representations from audio-visual spatial alignment. Advances in Neural Information Processing Systems, 33:4733-4744, 2020. 2, 5, 8, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 298, + 288, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 298, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 53, + 298, + 288, + 331 + ], + "type": "text", + "content": "[64] Pedro Morgado, Nuno Vasconcelos, Timothy Langlois, and Oliver Wang. Self-supervised generation of spatial audio for 360 video. arXiv preprint arXiv:1809.02587, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 332, + 288, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 332, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 53, + 332, + 288, + 386 + ], + "type": "text", + "content": "[65] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12475-12486, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 388, + 288, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 288, + 422 + ], + "type": "text", + "content": "[66] David Nister. An efficient solution to the five-point relative pose problem. IEEE transactions on pattern analysis and machine intelligence, 26(6):756-770, 2004. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 423, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 423, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 423, + 288, + 456 + ], + "type": "text", + "content": "[67] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. European Conference on Computer Vision (ECCV), 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 456, + 288, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 456, + 288, + 500 + ], + "spans": [ + { + "bbox": [ + 53, + 456, + 288, + 500 + ], + "type": "text", + "content": "[68] Andrew Owens, Phillip Isola, Josh McDermott, Antonio Torralba, Edward H Adelson, and William T Freeman. Visually indicated sounds. In Computer Vision and Pattern Recognition (CVPR), 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 502, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 288, + 555 + ], + "type": "text", + "content": "[69] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE, 2015. 5, 7, 8, 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 557, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 557, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 53, + 557, + 288, + 612 + ], + "type": "text", + "content": "[70] Kranti Kumar Parida, Siddharth Srivastava, and Gaurav Sharma. Beyond image to depth: Improving depth prediction using echoes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8268-8277, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 613, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 613, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 53, + 613, + 288, + 678 + ], + "type": "text", + "content": "[71] Senthil Purushwalkam, Sebastia Vicenc Amengual Gari, Vamsi Krishna Ithapu, Carl Schissler, Philip Robinson, Abhinav Gupta, and Kristen Grauman. Audio-visual floorplan reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1183–1192, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 680, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 680, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 680, + 288, + 712 + ], + "type": "text", + "content": "[72] Shengyi Qian, Linyi Jin, and David F Fouhey. Associative3d: Volumetric reconstruction from sparse views. In ECCV, 2020. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 72, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 312, + 72, + 547, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 547, + 117 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 547, + 117 + ], + "type": "text", + "content": "[73] Shengyi Qian, Alexander Kirillov, Nikhila Ravi, Devendra Singh Chaplot, Justin Johnson, David F Fouhey, and Georgia Gkioxari. Recognizing scenes from novel viewpoints. arXiv preprint arXiv:2112.01520, 2021. 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 118, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 547, + 173 + ], + "type": "text", + "content": "[74] Kranthi Kumar Rachavarapu, Vignesh Sundaresha, AN Rajagopalan, et al. Localize to binauralize: Audio spatialization from visual sound source localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1930-1939, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 175, + 547, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 175, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 312, + 175, + 547, + 240 + ], + "type": "text", + "content": "[75] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 241, + 547, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 241, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 312, + 241, + 547, + 329 + ], + "type": "text", + "content": "[76] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 5, 7, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 331, + 547, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 331, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 312, + 331, + 547, + 373 + ], + "type": "text", + "content": "[77] Lord Rayleigh. Xii. on our perception of sound direction. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 13(74):214-232, 1907. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 376, + 547, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 376, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 312, + 376, + 547, + 430 + ], + "type": "text", + "content": "[78] Alexander Richard, Dejan Markovic, Israel D Gebru, Steven Krenn, Gladstone Alexander Butler, Fernando Torre, and Yaser Sheikh. Neural synthesis of binaural speech from mono audio. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 432, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 432, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 312, + 432, + 547, + 487 + ], + "type": "text", + "content": "[79] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 488, + 547, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 488, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 312, + 488, + 547, + 521 + ], + "type": "text", + "content": "[80] Lord Rayleigh O.M. Pres. R.S. Xii. on our perception of sound direction. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 1907. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 522, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 522, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 312, + 522, + 547, + 578 + ], + "type": "text", + "content": "[81] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4938–4947, 2020. 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 579, + 547, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 579, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 312, + 579, + 547, + 612 + ], + "type": "text", + "content": "[82] Ralph Schmidt. Multiple emitter location and signal parameter estimation. IEEE transactions on antennas and propagation, 34(3):276-280, 1986. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 613, + 547, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 613, + 547, + 656 + ], + "spans": [ + { + "bbox": [ + 312, + 613, + 547, + 656 + ], + "type": "text", + "content": "[83] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 658, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 658, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 658, + 547, + 712 + ], + "type": "text", + "content": "[84] Nikhil Singh, Jeff Mentch, Jerry Ng, Matthew Beveridge, and Iddo Drori. Image2reverb: Cross-modal reverb impulse response synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 286-295, 2021. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "spans": [ + { + "bbox": [ + 295, + 752, + 315, + 761 + ], + "type": "text", + "content": "7907" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 289, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 288, + 127 + ], + "type": "text", + "content": "[85] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 129, + 289, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 129, + 289, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 129, + 289, + 194 + ], + "type": "text", + "content": "[86] Ruijie Tao, Zexu Pan, Rohan Kumar Das, Xinyuan Qian, Mike Zheng Shou, and Haizhou Li. Is someone speaking? exploring long-term temporal features for audio-visual active speaker detection. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3927–3935, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 196, + 289, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 196, + 289, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 196, + 289, + 262 + ], + "type": "text", + "content": "[87] Efthymios Tzinis, Scott Wisdom, Tal Remez, and John R Hershey. Audioscopev2: Audio-visual attention architectures for calibrated open-domain on-screen sound separation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXVII, pages 368–385. Springer, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 264, + 289, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 289, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 289, + 308 + ], + "type": "text", + "content": "[88] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 308, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 308, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 53, + 308, + 288, + 363 + ], + "type": "text", + "content": "[89] Paolo Vecchiotti, Ning Ma, Stefano Squartini, and Guy J Brown. End-to-end binaural sound localisation from the raw waveform. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 451-455. IEEE, 2019. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 365, + 288, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 365, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 53, + 365, + 288, + 397 + ], + "type": "text", + "content": "[90] DeLiang Wang and Guy J Brown. Computational auditory scene analysis: Principles, algorithms, and applications. Wiley-IEEE press, 2006. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 399, + 288, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 288, + 442 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 288, + 442 + ], + "type": "text", + "content": "[91] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. In ECCV, 2022. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 444, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 444, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 444, + 288, + 487 + ], + "type": "text", + "content": "[92] Fanyi Xiao, Yong Jae Lee, Kristen Grauman, Jitendra Malik, and Christoph Feichtenhofer. Audiovisual slowfast networks for video recognition. arXiv preprint arXiv:2001.08740, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 488, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 488, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 488, + 288, + 544 + ], + "type": "text", + "content": "[93] Xudong Xu, Hang Zhou, Ziwei Liu, Bo Dai, Xiaogang Wang, and Dahua Lin. Visually informed binaural audio generation without binaural audios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15485-15494, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 545, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 545, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 53, + 545, + 288, + 578 + ], + "type": "text", + "content": "[94] Nelson Yalta, Kazuhiro Nakadai, and Tetsuya Ogata. Sound source localization using deep learning models. Journal of Robotics and Mechatronics, 29(1):37-48, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 579, + 288, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 579, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 53, + 579, + 288, + 622 + ], + "type": "text", + "content": "[95] Dongchao Yang, Jianwei Yu, Helin Wang, Wen Wang, Chao Weng, Yuexian Zou, and Dong Yu. Diffsound: Discrete diffusion model for text-to-sound generation. arXiv preprint arXiv:2207.09983, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 624, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 624, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 624, + 288, + 689 + ], + "type": "text", + "content": "[96] Karren Yang, Michael Firman, Eric Brachmann, and Clément Godard. Camera pose estimation and localization with active audio sensing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII, pages 271-291. Springer, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 288, + 715 + ], + "type": "text", + "content": "[97] Karren Yang, Bryan Russell, and Justin Salamon. Telling left from right: Learning spatial correspondence of sight" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 532 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 315, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 547, + 106 + ], + "type": "text", + "content": "and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9932-9941, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 107, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 547, + 150 + ], + "type": "text", + "content": "[98] Zhenpei Yang, Jeffrey Z Pan, Linjie Luo, Xiaowei Zhou, Kristen Grauman, and Qixing Huang. Extreme relative pose estimation for rgb-d scans via scene completion. In CVPR, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 152, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 547, + 184 + ], + "type": "text", + "content": "[99] Zhenpei Yang, Siming Yan, and Qixing Huang. Extreme relative pose network under hybrid representations. In CVPR, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 186, + 547, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 240 + ], + "type": "text", + "content": "[100] Dong Yu, Morten Kolbæk, Zheng-Hua Tan, and Jesper Jensen. Permutation invariant training of deep models for speaker-independent multi-talker speech separation. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 241-245. IEEE, 2017. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 241, + 547, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 547, + 285 + ], + "type": "text", + "content": "[101] Yinda Zhang, Shuran Song, Ersin Yumer, Manolis Savva, Joon-Young Lee, Hailin Jin, and Thomas Funkhouser. Physically-based rendering for indoor scene understanding using convolutional neural networks. In CVPR, 2017. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 286, + 547, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 547, + 329 + ], + "type": "text", + "content": "[102] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In ICCV, 2021. 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 331, + 547, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 547, + 385 + ], + "type": "text", + "content": "[103] Hang Zhou, Xudong Xu, Dahua Lin, Xiaogang Wang, and Ziwei Liu. Sep-stereo: Visually guided stereophonic audio generation by associating source separation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 387, + 547, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 387, + 547, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 387, + 547, + 419 + ], + "type": "text", + "content": "[104] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, 2017. 1, 2, 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 421, + 547, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 421, + 547, + 463 + ], + "spans": [ + { + "bbox": [ + 308, + 421, + 547, + 463 + ], + "type": "text", + "content": "[105] Yipin Zhou and Ser-Nam Lim. Joint audio-visual deepfake detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14800–14809, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 465, + 547, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 465, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 465, + 547, + 532 + ], + "type": "text", + "content": "[106] Yuliang Zou, Pan Ji, Quoc-Huy Tran, Jia-Bin Huang, and Manmohan Chandraker. Learning monocular visual odometry via self-supervised long-term modeling. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV, pages 710-727. Springer, 2020. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 752, + 316, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 752, + 316, + 761 + ], + "spans": [ + { + "bbox": [ + 295, + 752, + 316, + 761 + ], + "type": "text", + "content": "7908" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Sound Source Localization is All about Cross-Modal Alignment/7b57a83c-7e70-4280-a6ee-f0f550384305_content_list.json b/2023/Sound Source Localization is All about Cross-Modal Alignment/7b57a83c-7e70-4280-a6ee-f0f550384305_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..efe93699c77c7b31ed6b0ab85ecbbbf569db4982 --- /dev/null +++ b/2023/Sound Source Localization is All about Cross-Modal Alignment/7b57a83c-7e70-4280-a6ee-f0f550384305_content_list.json @@ -0,0 +1,1576 @@ +[ + { + "type": "text", + "text": "Sound Source Localization is All about Cross-Modal Alignment", + "text_level": 1, + "bbox": [ + 163, + 130, + 805, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Arda Senocak $^{1*}$ Hyeonggon Ryu $^{1*}$ Junsik Kim $^{2*}$ Tae-Hyun Oh $^{3,4}$ Hanspeter Pfister $^{2}$ Joon Son Chung $^{1}$", + "bbox": [ + 189, + 179, + 763, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Korea Advanced Institute of Science and Technology $^{2}$ Harvard University", + "bbox": [ + 176, + 220, + 792, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Dept. of Electrical Engineering and Grad. School of Artificial Intelligence, POSTECH", + "bbox": [ + 138, + 239, + 828, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Institute for Convergence Research and Education in Advanced Technology, Yonsei University", + "bbox": [ + 107, + 257, + 861, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans can easily perceive the direction of sound sources in a visual scene, termed sound source localization. Recent studies on learning-based sound source localization have mainly explored the problem from a localization perspective. However, prior arts and existing benchmarks do not account for a more important aspect of the problem, cross-modal semantic understanding, which is essential for genuine sound source localization. Cross-modal semantic understanding is important in understanding semantically mismatched audio-visual events, e.g., silent objects, or off-screen sounds. To account for this, we propose a cross-modal alignment task as a joint task with sound source localization to better learn the interaction between audio and visual modalities. Thereby, we achieve high localization performance with strong cross-modal semantic understanding. Our method outperforms the state-of-the-art approaches in both sound source localization and cross-modal retrieval. Our work suggests that jointly tackling both tasks is necessary to conquer genuine sound source localization.", + "bbox": [ + 75, + 357, + 473, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 654, + 209, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans can easily perceive where the sound comes from in a scene. We naturally attend to the sounding direction and associate incoming audio-visual signals to understand the event. To achieve human-level audio-visual perception, sound source localization in visual scenes has been extensively studied [50, 51, 4, 47, 8, 35, 31, 33, 53, 54, 52, 36, 39, 38, 20]. Motivated by that humans learn from natural audio-visual correspondences without explicit supervision, most of the studies have been developed on a fundamental assumption that audio and visual signals are temporally correlated. With the assumption, losses of the sound source localization task are modeled by audio-visual correspondence as a self-supervision signal and are implemented by con", + "bbox": [ + 75, + 679, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/112abaefa8581f7e1d760d65eec2c80fc28af22ae6b6a062f2f72a0614e7dd21.jpg", + "image_caption": [ + "Figure 1. A conceptual difference between prior approaches and our alignment-based sound source localization." + ], + "image_footnote": [], + "bbox": [ + 501, + 309, + 890, + 421 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "trasting audio-visual pairs, i.e., contrastive learning.", + "bbox": [ + 498, + 462, + 841, + 477 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While these approaches appear to be unsupervised methods, they strongly rely on partial supervision information; e.g., using supervisedly pretrained vision networks [50, 51, 47, 53, 54, 20] and visual objectness estimators for postprocessing [39, 38]. Without leveraging such strong initial representations, the performance is degraded. Thus, the previous methods are not purely self-supervised approaches. Even further, there are recent studies [45, 39, 38] that point out visual objectness bias in existing sound source localization benchmarks and exploit the objectness prior to improve the localization accuracy. They show that, even without interaction between visual and audio signals, a model may achieve strong accuracy in localization by only referring visual signals alone, which is not the true intention of the sound source localization task. In short, the current evaluation and setting of the sound source localization do not capture the true sound source localization performance.", + "bbox": [ + 496, + 478, + 892, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we first sort out evaluating sound source localization methods by introducing a cross-modal retrieval task as an auxiliary evaluation task. By this task, we can measure whether the learned representation has the capability to accurately interact between audio and visual modalities; i.e., more fine-grained audio-visual correspondence which is essential for genuine sound source localization. This aspect has been missed in existing sound source localization benchmarks. Indeed, our experiments show that higher sound localization performance does not guarantee higher cross-modal retrieval performance.", + "bbox": [ + 496, + 734, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 76, + 886, + 334, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7777", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Second, given this additional criterion, we revisit the importance of semantic understanding shared across audio and visual modalities in both sound source localization and cross-modal retrieval. In the previous methods [50, 51, 54, 47], the cross-modal semantic alignment is induced by instance-level cross-modal contrastive learning, i.e., cross-modal instance discrimination between visual and audio features. However, they are aided by labels or supervisedly pretrained encoder for easing challenging cross-modal feature alignment. Instead, our method learns from scratch supporting the lack of guidance by incorporating multiple positive samples into cross-modal contrastive learning. Specifically, we construct a positive set for each modality using both multi-view [10] and conceptually similar samples [17]. Thereby, we enhance feature alignment and achieve high localization performance and strong cross-modal semantic understanding.", + "bbox": [ + 75, + 90, + 472, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our method on the VGG-SS and SoundNet-Flickr benchmarks for sound source localization and cross-modal retrieval. As aforementioned, the sound source localization task is closely related to the cross-modal retrieval task, but our experiments show that existing works have a weak performance correlation between them. This implies that we need to evaluate both tasks for evaluating the genuine sound source localization. The proposed method performs favorably against the recent state-of-the-art approaches in both tasks.", + "bbox": [ + 75, + 347, + 472, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize the contributions of our work as follows:", + "bbox": [ + 96, + 500, + 467, + 513 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We analyze that sound source localization benchmarks are not capable of evaluating cross-modal semantic understanding, thereby sound source localization methods may perform poorly in cross-modal retrieval tasks.", + "- We propose semantic alignment to improve cross-modal semantic understanding of sound source localization models.", + "- We expand semantic alignment with multi-views and conceptually similar samples which leads to state-of-the-art performance on both sound source localization and cross-modal retrieval." + ], + "bbox": [ + 76, + 518, + 468, + 690 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 695, + 215, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sound source localization. Sound source localization in visual scenes has been investigated by exploiting correspondences between audio and visual modalities. The most widely used approach for sound source localization is cross-modal attention [50, 51, 57] with contrastive loss [13, 29, 42]. Later, the attention-based method is improved by intra-frame hard sample mining [8], iterative contrastive learning with pseudo labels [35], feature regularization [36], positive mining [52], negative free learning [54] with stop-gradient operation [12], or momentum encoders [38].", + "bbox": [ + 75, + 713, + 468, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some sound localization approaches exploit additional semantic labels [47, 33, 53] or object prior [39, 63]. Semantic labels are used to pretrain audio and vision encoders with classification loss [33, 53] or refine audio-visual feature alignment [47]. A more explicit way to refine localization output is to use object prior. EZVSL [39] proposes post-processing to combine attention based localization output with a pretrained visual feature activation map. Similarly, Xuan et al. [63] propose to combine off-the-shelf object proposals with attention based sound localization results. However, postprocessing by object prior may generate a false positive output as it is solely based on vision without audio-visual interaction.", + "bbox": [ + 496, + 90, + 890, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to the localization, there has been an attempt to localize sounding objects and recover the separated sounds simultaneously, also known as the cocktail party problem [27, 37]. The separation of sound mixture is achieved by predicting masks of spectrogram guided by visual features [19, 1, 64, 23, 62, 21, 2, 65, 24, 58, 56]. Furthermore, a number of recent papers are presented on audio-visual navigation for a given sound source [7, 22].", + "bbox": [ + 496, + 287, + 892, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised representation learning. In a broader categorization, sound source localization belongs to self-supervised multimodal learning. Our work is also relevant to self-supervised audio-visual representation learning, and other multimodal learning studies.", + "bbox": [ + 496, + 416, + 890, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive learning aims to learn robust representations from large-scale raw data without annotations. Recent representation learning approaches [60, 10, 28, 11] use instance discrimination by contrastive learning [13, 29, 42] as a pretext task with notable advancements in visual recognition tasks. Recently, positive mining by nearest-neighbor search are used to learn representations of images [17, 18, 61], videos [26], neural recordings [6], and text-image [34]. In this work, we expand the previous works by incorporating both multi-views and conceptually similar samples into audio-visual modalities for cross-modal feature alignment.", + "bbox": [ + 496, + 492, + 892, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A series of audio-visual representation learning studies have shown that audio and visual contents in a video are correlated, therefore a visual representation can be learned by sound prediction [44] or audio representation can be distilled from visual representation [5, 55]. Later, a variety of joint audio-visual representation learning methods are proposed with an assumption that there is a semantic [3, 30, 41, 40] or temporal [14, 43, 32, 15] correspondence between them. However, simply learning sound source localization by audio-visual correspondence with instance discrimination ignores the semantic similarity of audio-visual contents among samples, introducing false negatives or positives. In order to mitigate this issue, clustering [30], sampling [41], weighting [40], and hard mining [32] are proposed. Similarly, in this work, we go beyond instance discrimination by using multiple positive samples", + "bbox": [ + 496, + 659, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "Typically, an image encoder is pretrained on ImageNet [16] and an audio encoder is pretrained on AudioSet [25] in supervised ways.", + "bbox": [ + 71, + 875, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7778", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/33bd5916b5d2db143c4350848ae6aaa9bf92f1f972ccfd9924bf6d94e623e3cf.jpg", + "image_caption": [ + "Figure 2. Our sound source localization framework. Our model constructs multiple positive pairs with augmentation and Nearest Neighbor Search (Conceptually Similar Samples). By using these newly constructed 9 pairs, our model employs spatial localization, $s_L$ , and semantic feature alignment, $s_A$ , for each pair to learn a better sound source localization ability." + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 893, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to enforce semantic understanding across modalities.", + "bbox": [ + 76, + 352, + 428, + 368 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 382, + 168, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 407, + 217, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive learning learns representation by containing positive and negative pairs. Given an encoded query sample $q$ and its encoded positive pair $k^{+}$ and negative pairs $k$ , the loss can be defined as:", + "bbox": [ + 75, + 439, + 470, + 500 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = - \\log \\frac {\\exp (q \\cdot k ^ {+} / \\tau)}{\\sum_ {i} \\exp (q \\cdot k _ {i} / \\tau)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 510, + 468, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau$ is the temperature parameter.", + "bbox": [ + 76, + 556, + 328, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-modal contrastive learning extends contrastive learning across multiple modalities. In sound source localization, audio-visual correspondence is used to define positive and negative cross-modal pairs. With an audio-visual dataset $\\mathcal{D} = \\{(v_i,a_i):i = 1,\\dots,N\\}$ and its encoded features $\\mathbf{v}_i = f_v(v_i)$ and $\\mathbf{a}_i = f_a(a_i)$ , cross-modal contrastive learning loss is defined as:", + "bbox": [ + 75, + 579, + 468, + 685 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i} = - \\log \\frac {\\exp (s (\\mathbf {v} _ {i} , \\mathbf {a} _ {i}) / \\tau)}{\\sum_ {j} \\exp (s (\\mathbf {v} _ {i} , \\mathbf {a} _ {j}) / \\tau)} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 696, + 468, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $s$ is a cross-modal similarity function. The cross-modal contrastive loss Eq. (2) can be extended to symmetric form [48] as used in a few previous works [39, 38].", + "bbox": [ + 75, + 744, + 468, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Cross-Modal Feature Alignment", + "text_level": 1, + "bbox": [ + 76, + 801, + 362, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider both spatial localization and semantic feature alignment for sound source localization. To this end, we use two different similarity functions $s_L$ and $s_A$ for contrastive learning (Eq. (2)), $s_L$ for localization and $s_A$ for cross-modal feature alignment.", + "bbox": [ + 75, + 824, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent studies rely on audio-visual spatial correspondence maps to learn sound source localization by contrasting them. Given a spatial visual feature $\\mathbf{v} \\in \\mathbb{R}^{c \\times h \\times w}$ and audio feature $\\mathbf{a} \\in \\mathbb{R}^{\\mathbf{c}}$ , audio-visual similarity with a correspondence map can be calculated as follows:", + "bbox": [ + 498, + 352, + 890, + 428 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns _ {L} (\\mathbf {v}, \\mathbf {a}) = \\sum_ {x y \\in M} \\frac {1}{| M |} \\frac {\\mathbf {v} ^ {x y} \\cdot \\mathbf {a}}{\\| \\mathbf {v} ^ {x y} \\| \\| \\mathbf {a} \\|} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 436, + 890, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{v}^{xy}$ is a feature vector at location $(x,y)$ , and $M$ is an optional binary mask when an annotation or pseudomask [8, 36] is available. Since we assume no supervision for sound source localization, we do not use any mask, therefore, $M = \\mathbf{1}$ .", + "bbox": [ + 498, + 484, + 890, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The contrastive loss with localization similarity $s_L$ enforces location dependent alignment giving sparse but strong audio-visual correspondence which enables to perform localization. However, our empirical studies on cross-modal retrieval indicate that strong localization performance does not guarantee semantic understanding. To overcome the low semantic understanding in recent studies, we propose to add instance-level contrastive loss. Instance-level contrasting encapsulates the whole context in a scene, enforcing better audio-visual semantic alignment. However, instance-level contrasting may smooth out spatial discriminativeness learned by Eq. (3). Inspired by SimCLR [10], we adopt a projection layer to align audio-visual semantics in a projection space. The projection layer separates the latent space of localization and semantic alignment, thereby preventing the alignment loss smoothing out the spatial discriminativeness. The similarity function for cross-modal feature alignment is defined as follows:", + "bbox": [ + 496, + 561, + 892, + 833 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns _ {A} (\\mathbf {v}, \\mathbf {a}) = \\frac {p _ {v} (\\operatorname {a v g} (\\mathbf {v})) \\cdot p _ {a} (\\mathbf {a})}{\\| p _ {v} (\\operatorname {a v g} (\\mathbf {v})) \\| \\| p _ {a} \\mathbf {a} \\|} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 842, + 890, + 877 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{avg}(\\cdot)$ is spatial average pooling, $p_v$ is a projection", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7779", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "layer for visual features, and $p_a$ is a projection layer for audio features.", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Expanding with Multiple Positive Samples", + "text_level": 1, + "bbox": [ + 76, + 133, + 437, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Typically, contrastive learning contrasts between one positive pair and multiple negative pairs as shown in Eq. (1). In audio-visual learning, by an audio-visual correspondence assumption, an audio-image pair from the same clip is used as a positive pair while negative pairs are sampled from different clips. However, single-instance discrimination may not be sufficient to achieve strong cross-modal alignment. In this section, we expand contrastive learning beyond single instance discrimination by positive set construction and pairing them. To construct a positive set, we incorporate both hand-crafted positive and conceptual positive samples for each modality. Later, we adjust the contrastive learning to incorporate multiple positive pairs to enforce cross-modal alignment.", + "bbox": [ + 75, + 157, + 468, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Obtaining hand-crafted positive samples. Using randomly augmented samples as positive multi-view pairs are widely adopted in self-supervised representation learning, i.e., instance discrimination. Similarly, we extend a single anchor audio-image pair to multiple positive pairs by applying simple augmentations on image and audio samples separately. While we utilize common image transformations on images, we apply temporal shifting to audiOs. It is worth noting that sound source localization task learns from the underlying semantic consistency rather than subtle time differences as in videos. Thus, a slight shift in the audio may not alter contextual information significantly. As a result of hand-crafted multi-view positive pair generation, we obtain additional $\\mathbf{v}^{aug}$ and $\\mathbf{a}^{aug}$ samples.", + "bbox": [ + 75, + 378, + 468, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Obtaining conceptual positive samples. Apart from manually created augmented views, we additionally expand our positive set with conceptually similar samples. The sampling strategy with nearest neighbor search can be performed in a various way, such as on-the-fly sampling [17, 49, 61, 34], sampling by pretrained encoders [52], or guided sampling [26, 18] using another modality. For selecting our conceptually similar samples, we utilize pretrained encoders. Note that pretrained encoders trained either with supervised or self-supervised learning are effective in positive sample mining as shown in the experiment section. By employing readily available image and audio encoders, we use the $k$ -nearest neighborhood search to sample semantically similar samples in both modalities. In particular, given a pair of image and audio, we compute cosine similarity with all other samples and choose the top- $k$ most similar samples among the training set for each modality. From a set of $k$ samples, we randomly select one sample to obtain conceptually similar samples for each modality, $\\mathbf{v}^{conc}$ and $\\mathbf{a}^{conc}$ . By utilizing the conceptually similar samples as", + "bbox": [ + 75, + 598, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "positive samples, our model expands semantic understanding.", + "bbox": [ + 498, + 90, + 888, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pair Construction. Once we obtain the conceptual and hand-crafted positive samples for each modality, we proceed to create 9 distinct audio-visual pairs by pairing $\\mathbf{V} = \\{\\mathbf{v},\\mathbf{v}^{aug},\\mathbf{v}^{conc}\\}$ and $\\mathbf{A} = \\{\\mathbf{a},\\mathbf{a}^{aug},\\mathbf{a}^{conc}\\}$ . This is done to ensure semantic alignment and consistency between them through contrastive learning. The negative pairs are randomly paired from the remaining samples in a training set. It is worth noting that some of these pairs are a combination of hand-crafted and conceptually similar samples, which further enhances the feature alignment of our model during training.", + "bbox": [ + 498, + 128, + 890, + 295 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Training", + "text_level": 1, + "bbox": [ + 500, + 306, + 602, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our loss formulation incorporates both localization and instance-level similarity functions with multiple positive pairs constructed by augmentation and conceptually similar sample search. The final loss term is defined as follows:", + "bbox": [ + 498, + 330, + 890, + 390 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {i} = - \\sum_ {\\mathbf {v} _ {i} \\in \\mathbf {V}} \\sum_ {\\mathbf {a} _ {i} \\in \\mathbf {A}} \\left[ \\log \\frac {\\exp \\left(s _ {L} \\left(\\mathbf {v} _ {i} , \\mathbf {a} _ {i}\\right) / \\tau\\right)}{\\sum_ {j} \\exp \\left(s _ {L} \\left(\\mathbf {v} _ {i} , \\mathbf {a} _ {j}\\right) / \\tau\\right)} \\right. \\tag {5} \\\\ \\left. + \\log \\frac {\\exp \\left(s _ {A} \\left(\\mathbf {v} _ {i} , \\mathbf {a} _ {i}\\right) / \\tau\\right)}{\\sum_ {j} \\exp \\left(s _ {A} \\left(\\mathbf {v} _ {i} , \\mathbf {a} _ {j}\\right) / \\tau\\right)} \\right] \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 402, + 890, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{V}$ and $\\mathbf{A}$ indicate positive sample sets.", + "bbox": [ + 500, + 500, + 800, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 530, + 633, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our proposed method for sound source localization is validated through experiments conducted on VG-GSound [9] and SoundNet-Flickr [5]. First, we conduct a quantitative analysis to evaluate the accuracy of the localization, cross-modal retrieval, and the impact of various components of our model. Then, we visualize our sound source localization results across different categories of sounds.", + "bbox": [ + 498, + 555, + 890, + 675 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 500, + 686, + 674, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. Our method is trained using the VGGSound [9] and SoundNet-Flickr-144K [50, 51]. VGGSound is an audio-visual dataset containing around $\\sim 200\\mathrm{K}$ videos. SoundNet-Flickr-144K set is the subset of SoundNetFlickr [5]. After training, we test the sound localization performance with VGG-SS [8] and SoundNet-Flickr-Test [50] datasets for the main experiments. These evaluation sets have bounding box annotations of sound sources for $\\sim 5\\mathrm{K}$ and 250 samples, respectively. Moreover, we employ the AVSBench [66] and Extended VGGSound/SoundNetFlickr [38] datasets for additional evaluations. AVSBench dataset provides binary segmentation maps that show the", + "bbox": [ + 498, + 719, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7780", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/1bfe09549e50b43a3f68f7a79b85158d6c86751effefa6ed88d110656aa2a23c.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Quantitative results on the VGG-SS and SoundNet-Flickr test sets. All models are trained with 144K samples from VGG-Sound and tested on VGG-SS and SoundNet-Flickr. $\\dagger$ is the result of the model released on the official project page. SLAVC [38] does not provide AUC scores." + ], + "table_body": "
MethodPre. VisionVGG-SSFlickr-SoundNet
cIoU ↑AUC ↑cIoU ↑AUC ↑
Attention [50]CVPR18✓18.5030.2066.0055.80
CoarseToFine [47]ECCV20✓29.1034.80--
LCBM [53]WACV22✓32.2036.60--
LVS [8]†CVPR21✗30.3036.4072.4057.80
LVS [8]CVPR21✗34.4038.2071.9058.20
HardPos [52]ICASSP22✗34.6038.0076.8059.20
SSPL (w/o PCM) [54]CVPR22✓27.0034.8073.9060.20
SSPL (w/ PCM) [54]CVPR22✓33.9038.0076.7060.50
EZ-VSL (w/o OGL) [39]ECCV22✓35.9638.2078.3161.74
SSL-TIE [36]ACMM22✗38.6339.6579.5061.20
SLAVC (w/o OGL) [38]NeurIPS22✓37.7939.4083.60-
Ours
↓ NN Search w/ Supervised Pre. Encoders✗39.9440.0279.6063.44
↓ NN Search w/ Self-Supervised Pre. Encoders✗39.2039.7079.2063.00
with OGL:
EZ-VSL (w/ OGL) [39]ECCV22✓38.8539.5483.9463.60
SLAVC (w/ OGL) [38]NeurIPS22✓39.80-86.00-
Ours (w/ OGL)
↓ NN Search w/ Supervised Pre. Encoders✗42.6441.4882.4064.60
↓ NN Search w/ Self-Supervised Pre. Encoders✗42.4741.4282.8064.48
with Optical Flow:
HearTheFlow [20]WACV23✓39.4040.0084.8064.00
", + "bbox": [ + 80, + 88, + 470, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "audio-visually correspondent pixels for roughly 5k five-second videos belonging to 23 categories. Lastly, the Extended VGGSound /SoundNet-Flickr dataset, proposed by [38], is used to understand non-visible sound sources.", + "bbox": [ + 75, + 410, + 468, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We use two ResNet18 models for both audio and vision encoding. Unlike prior approaches, we do not fine-tune (or use a pretrained) a visual encoder from ImageNet pretrained weights. Instead, we train both the audio and vision encoders from scratch. We preprocess images and audios following the previous works [8, 52]. To create multiple pairs, we utilize both NN search and generic augmentation approaches. For NN search, we experiment on two different setups to retrieve k conceptually similar samples: (1) For supervisedly pretrained encoder experiments, We employ ResNet and VGGSound models pretrained on ImageNet and VGGSound respectively, (2) For self-supervisedly pretrained encoder experiments, we utilize the CLIP [48] Vision Encoder and Wav2CLIP [59] Audio Encoder. We use $k = 1000$ for the experiments. To perform image augmentations, we follow the augmentations used in SimCLR [10]. For audios, we randomly select timewindow shifts in a time axis. The model is trained for 50 epochs with Adam Optimizer and a learning rate of 0.0001. $\\tau$ is set to 0.07 in contrastive learning.", + "bbox": [ + 75, + 478, + 470, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 76, + 792, + 272, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison with strong baselines. In this section, we conduct a comparative analysis of our sound source localization method against existing approaches. We carry out our evaluations in two settings, following previous approaches. Firstly, we train our model on VGGSound-144K", + "bbox": [ + 75, + 825, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1a68b3b7007f70b34d80dbee402a0ff096f01f2f753b7993cb4c6df78e1410a9.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Quantitative results on the SoundNet-Flickr test set. All models are trained and tested on the SoundNet-Flickr 144K dataset. $\\dagger$ is the result of the model from the official project page. SLAVC [38] does not provide results with SoundNet-Flickr 144K." + ], + "table_body": "
MethodPre. VisioncIoU ↑AUC ↑
Attention[50]CVPR18✓66.0055.80
DMC[30]CVPR19✓67.1056.80
LVS [8]†CVPR21X67.2056.20
LVS [8]CVPR21X69.9057.30
HardPos [52]ICASSP22X75.2059.70
SSPL (w/o PCM) [54]CVPR22✓69.9058.00
SSPL (w/ PCM) [54]CVPR22✓75.9061.00
EZ-VSL (w/o OGL) [39]ECCV22✓71.8958.81
SSL-TIE [36]ACMM22X81.5061.10
SLAVC (w/o OGL) [38]NeurIPS22✓--
Ours
↓ NN Search w/ Supervised Pre. EncodersX85.2062.20
↓ NN Search w/ Self-Supervised Pre. EncodersX84.8062.66
with OGL:
EZ-VSL (w/ OGL) [39]ECCV22✓83.1363.06
SLAVC (w/ OGL) [38]NeurIPS22✓--
Ours (w/ OGL)
↓ NN Search w/ Supervised Pre. EncodersX84.0064.16
↓ NN Search w/ Self-Supervised Pre. EncodersX84.4064.38
with Optical Flow:
HearTheFlow [20]WACV23✓86.5063.90
", + "bbox": [ + 504, + 88, + 893, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and evaluate it on VGG-SS and SoundNet-Flickr test sets. Secondly, we train our model on SoundNet-Flickr-144K and evaluate it on the SoundNet-Flickr test set. It is important to note that all the compared models are trained using the same amount of data. AVEL [57], AVObject [2], and LCBM [53] models rely on video input, and as such, they cannot be evaluated on the SoundNet-Flickr dataset, which contains static image and audio pairs. We present our results in Table 1 and Table 2.", + "bbox": [ + 496, + 444, + 890, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our proposed model achieves higher performance compared to prior approaches on both test sets. Specifically, it yields a $+2.15\\%$ cIoU and $+0.6\\%$ AUC improvement on VGGSS, as well as a $+3.7\\%$ cIoU improvement on SoundNet-Flickr compared to the state-of-the-art methods that uses pretrained vision encoder. It is worth highlighting that unlike the majority of previous works, our proposed model does not utilize a vision encoder pretrained on ImageNet in a sound source localization backbone. This is because, as discussed in Mo et al. [38], using supervisedly pretrained vision encoders makes the sound source localization problem a weakly supervised problem. However, it is worth noting that even without using a pretrained vision encoder, our method achieves state-of-the-art performance on both experiments that are presented in Table 1 and Table 2. We demonstrate the performance of our model with the pretrained models learned through supervised learning (NN Search w/ Supervised Pre. Encoders) and with models that are pretrained through self-supervised learning (NN Search w/ Self-Supervised Pre. Encoders) in NN Search module. As the results indicate, using self-supervised pre", + "bbox": [ + 496, + 583, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7781", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/489603953d1cbec4b8dae37db97d607e698c7d0a1f21db171399ca4ea762ded5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Test ClassMethodPre. VisioncIoU ↑AUC ↑
Heard 110LVS [8]CVPR21X28.9036.20
EZ-VSL(w/o OGL) [39]ECCV22✓31.8636.19
SLAVC(w/o OGL) [38]NeurlPS22✓35.84-
OursX38.3139.05
with OGL:
EZ-VSL(w/o OGL) [39]ECCV22✓37.2538.97
SLAVC(w/o OGL) [38]NeurlPS22✓38.22-
Ours(w/ OGL)X41.8540.93
with Optical Flow:
HearTheFlow [20]WACV23✓37.3038.60
Unheard 110LVS [8]CVPR21X26.3034.70
EZ-VSL(w/o OGL) [39]ECCV22✓32.6636.72
SLAVC(w/o OGL) [38]NeurlPS22✓36.50-
OursX39.1139.80
with OGL:
EZ-VSL(w/o OGL) [39]ECCV22✓39.5739.60
SLAVC(w/o OGL) [38]NeurlPS22✓38.87-
Ours(w/ OGL)✓42.9441.54
with Optical Flow:
HearTheFlow [20]WACV23✓39.3040.00
", + "bbox": [ + 96, + 88, + 455, + 285 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/17c4691d1129feefc6a3eb79024699fd9250fc974e1c9e9600b99f34ccdce179.jpg", + "table_caption": [ + "Table 3. Comparison results on open-set audio-visual localization experiments trained and tested on the splits of [8, 39, 20]." + ], + "table_footnote": [], + "table_body": "
Test ClassMethodPre. VisioncIoU ↑AUC ↑
Heard 110SSSL-TIE [36]ACMM22X39.0040.30
OursX41.2041.00
Unheard 110SSSL-TIE [36]ACMM22X36.5038.60
OursX36.9038.59
", + "bbox": [ + 109, + 327, + 439, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "trained encoders in NN Search performs on par with the supervised pretrained encoders in NN Search. This shows that our model does not depend on supervised pretrained encoders for the NN search module and can utilize any type of pretrained encoder feature for nearest neighbor search. Note that these pretrained encoders are not used in the backbone networks of the sound source localization module but only in the NN Search Module, as illustrated in Figure 2.", + "bbox": [ + 75, + 429, + 467, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also discuss the methods employed by previous studies, such as SSPL [54] which utilizes a sub-module called PCM to reduce the impact of background noise, HTF [20] which utilizes Optical Flow, and EZ-VSL [39] which refines its initial audio-visual localization outcomes through object guidance obtained from an ImageNet pretrained visual encoder. Our model, on the other hand, and any of its variations do not require any task-specific modules or operations to achieve the state-of-the-art (SOTA) results. This suggests that using additional semantic and multi-view correspondence, as well as feature alignment, provides more varied and robust supervision for better aligned audio and visual features, as opposed to using task-specific approaches.", + "bbox": [ + 75, + 551, + 467, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The quantitative results presented in Table 1 and Table 2 also showcase the performance of previous methods that utilize object guidance to evaluate their final sound source localizations. Our model outperforms all previous methods that employ object guidance on the VGG-SS test set and achieves comparable results on the SoundNet-Flickr test set, even though our model does not use object guided refinement (OGL). Additionally, we acknowledge that the addition of OGL to our audio-visual localization results in improvement on the VGGSS test set, while degrading perfor", + "bbox": [ + 75, + 750, + 467, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7d60cb960c0e2b6b681424b7958fc2453c7563e26e3ccce1ed9dfd444b841718.jpg", + "table_caption": [ + "Table 4. Comparison results on open set audio-visual localization experiments trained and tested on the splits of [36]." + ], + "table_footnote": [], + "table_body": "
Test SetMethodPre. VisionmIoU ↑F-Score ↑
S4LVS (w/o OGL) [8]CVPR21X26.933.6
EZ-VSL (w/o OGL) [39]ECCV22✓27.634.2
SLAVC (w/o OGL) [38]NeuRTPS22✓28.134.6
Ours (w/o OGL)
↓ NN Search w/ Supervised Pre. EncodersX29.635.9
↓ NN Search w/ Self-Supervised Pre. EncodersX29.335.6
", + "bbox": [ + 504, + 88, + 890, + 167 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/337b012838eae12d4d2878a284dc4310aef39f5a97d9bc1b9a44f971200893f9.jpg", + "table_caption": [ + "Table 5. Quantitative results on AVS Bench S4 dataset. All models are trained on the VGGSound 144K dataset." + ], + "table_footnote": [], + "table_body": "
ModelPre. VisionA → II → A
R@1R@5R@10R@1R@5R@10
LVS [8]CVPR21X3.8712.3520.734.9014.2921.37
EZ-VSL [39]ECCV22✓5.0115.7324.8114.233.5145.18
SSL-TIE [36]MM22X10.2930.6843.7612.7629.5839.72
SLAVC [38]NeurIPS22✓4.7713.0819.106.1221.1632.12
Ours
↓ NN Search w/ Supervised Pre. EncodersX16.4736.9949.0020.0942.3853.66
↓ NN Search w/ Self-Supervised Pre. EncodersX14.3137.8149.1718.0038.3949.02
", + "bbox": [ + 501, + 219, + 895, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 6. Summary of retrieval recall scores for all models. All of the models are trained on VGGSound 144K data and retrieval is performed on entire VGG-SS dataset, containing $\\sim 5\\mathrm{K}$ samples.", + "bbox": [ + 500, + 306, + 890, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mance on the SoundNet-Flickr test set. In contrast, prior methods see modest improvements when utilizing OGL. This can be explained by the fact that our model is already accurately localizing the sounding objects, and object guidance can interfere with localization results by introducing visual regions that are not sounding (refer to Section 4.4 for visual results). Unlike prior methods, we do not use OGL in our architecture for the remainder of this paper, unless it is being directly compared with OGL-based methods.", + "bbox": [ + 496, + 361, + 890, + 497 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, in comparison to HearTheFlow, which utilizes an additional Optical Flow modality, our method outperforms it on the VGGSS test set, and achieves slightly lower performance on the SoundNet-Flickr test set without utilizing any additional modalities, but instead relying on better audio-visual correspondence and alignment.", + "bbox": [ + 496, + 500, + 890, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Open Set Audio-Visual Localization. The study by Chen et al. [8] evaluates the generalization ability of sound source localization methods in an open set scenario. This involves testing the models on categories that are both present in the training data (heard) and categories that are not present (unheard). To accomplish this, 110 randomly selected categories from the VGGSound dataset are used for training, while another disjoint set of 110 categories are reserved for evaluation to ensure the model have never seen or heard them before. It should be noted that not all previous works use the same train/test splits. While some works, including [8, 39, 38, 46], share the same splits,[36] uses a different split. Therefore, to ensure a fair comparison, we conduct experiments on both splits, evaluating on test samples from both heard and unheard categories. The results are shown in Table 3 and Table 4. Our model outperforms existing approaches on both categories, regardless of the train/test splits used. Specifically, in Table 3, our model (w/o OGL) even surpasses the other models that use OGL. Previous approaches draw different conclusions from", + "bbox": [ + 496, + 598, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7782", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e4bad1492f02903387a789ff0d6eb5b172b37e82f0fae4a24dd9d658d0d8a7a7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPre. VisionExtended Flickr-SoundNetExtended VGG-SS
AP ↑max-F1 ↑LocAcc ↑AP ↑max-F1 ↑LocAcc ↑
CoarseToFine [47]ECCV20✓0.0038.2047.200.0019.8021.93
LVS [8]CVPR21✗9.8017.9019.605.159.9010.43
Attention10k [50]CVPR18✓15.9824.0034.166.7013.1014.04
DMC [30]CVPR19✓25.5641.8052.8011.5320.3022.63
DSOL [31]NeuIPS20✓38.3249.4072.9116.8425.6026.87
OGL [39]ECCV22-40.2055.7077.2018.7330.9036.58
EZ-VSL (w/o OGL) [39]ECCV22✓46.3054.6066.4024.5530.9031.58
SLAVC (w/o OGL) [38]NeuIPS22✓51.6359.1083.6032.9540.0037.79
Ours
↓ NN Search w/ Supervised Pre. Encoders✗64.4366.9079.6034.7340.7039.94
↓ NN Search w/ Self-Supervised Pre. Encoders✗62.6766.1079.2033.0940.0039.20
", + "bbox": [ + 80, + 88, + 467, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "these open set experiments. While some conclude that their models have strong generalization ability because their performance in unheard categories is higher than heard categories [39, 38, 46], the other works that cannot achieve the same trend discuss that this is expected since their models are dealing with unseen categories [36]. However, our results show that these conclusions are highly dependent on the chosen train/test splits. Our model performs better than existing works in both splits, but there is no uniform trend in between two splits. While our method performs better on unheard categories in the splits of [8, 39, 38, 46], it performs worse on unheard categories in the split of [36]. Therefore, we conclude that the observed trends are highly dependent on the randomly selected train/test splits.", + "bbox": [ + 75, + 263, + 470, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "AVSBench [66]. To demonstrate the precise sound localization ability of our model, we conduct experiments on the AVSBench S4 dataset. The dataset's objective is to detect audio-visual correspondence and correlation at the pixel level. To make a fair comparison, we use some of the self-supervised sound source localization methods mentioned earlier. All models are trained on VGGSound-144K and directly assessed on the AVSBench S4 dataset without any further fine-tuning (zero-shot setting). Our results, which are presented in Table 5, indicate that our method achieves the highest performance, as in the previous experiments.", + "bbox": [ + 75, + 483, + 468, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Retrieval. We evaluate sound localization models on the VGG-SS dataset for cross-modal retrieval. As shown in Table 6, our method clearly outperforms other state-of-the-art methods. One interesting observation is that EZ-VSL [39] notably performs better than SLAVC [38] on cross-modal retrieval, while SLAVC performs better on sound source localization in Table 1. This shows that with the current benchmark evaluations, better sound localization performance does not guarantee better audio-visual semantic understanding, thereby we need to additionally evaluate sound source localization methods on cross-modal understanding tasks. Another observation is that the performance gap between our method and the strongest competitor SSL-TIE [36] is notably larger on cross-modal retrieval than sound source localization. This is due to the strong cross-modal feature alignment of our method that is over", + "bbox": [ + 75, + 657, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b7ec74897f50ce3c467909d7316d744e3480b483192f98571ddc41c4aeff23c6.jpg", + "table_caption": [ + "Table 7. Quantitative results on the Extended VGG-SS and Extended SoundNet-Flickr sets. All models are trained with 144K samples from VGG-Sound. The results of the prior approaches are obtained from [38]." + ], + "table_footnote": [], + "table_body": "
SemanticMulti-ViewFeature AlignmentcIoU ↑AUC ↑
(A)✓✓✓39.9440.02
(B)✓✓✗39.1039.44
(C)✓✗✓38.7539.34
(D)✓✗✗38.2438.90
(E)✗✓✓38.3039.38
(F)✗✓✗37.7239.19
(G)✗✗✓34.9337.94
(H)✗✗✗34.2237.67
", + "bbox": [ + 504, + 88, + 890, + 203 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8. Ablation studies on our proposed method to see the impact of each main component.", + "bbox": [ + 500, + 205, + 890, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "looked in the sound source localization benchmarks.", + "bbox": [ + 500, + 255, + 846, + 270 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Extended Flickr and VGG-SS datasets. The prior study [38] points out that the current sound source localization benchmarks overlook false positive detection. It is because the evaluation samples always contain at least a sounding object in a scene; thus cannot capture false positive outputs, e.g., silent objects or off-screen sounds. To analyze false positive detection, Mo and Morgado [38] extended the benchmarks with non-audible, non-visible, and mismatched audio-visual samples. The expectation is that a sound source localization model should not localize any objects when audio-visual semantics do not match.", + "bbox": [ + 496, + 277, + 890, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The experiment with the extended datasets in Table 7 shows that our method performs favorably against state-of-the-art competitors. Our method performs better than the competing methods in false positive detection measured by AP and max-F1, while SLAVC [38] achieves better localization performance on Extended Flickr-SoundNet. As both false positive detection and cross-modal retrieval require cross-modal interaction, our method shows strong performance on both tasks.", + "bbox": [ + 496, + 446, + 892, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Results", + "text_level": 1, + "bbox": [ + 500, + 595, + 663, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct a series of experiments in order to verify our design choices and make further analysis. To save computational time and resources, we primarily perform ablation studies by training our model on VGGSound-144K with NN Search w/ Supervised Pre. Encoders setup and evaluating it on VGG-SS. Results are in Table 8.", + "bbox": [ + 496, + 619, + 890, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Semantic and Multi-View Invariance. In order to understand the impact of each type of invariance (consistency), we analyze the performance of our model with different type of invariance methodologies in Table 8. As the results of (C vs. E) and (D vs. F) reveal, using semantically similar samples (semantic invariance) produces better performance $(+0.45\\%)$ and $+0.5\\%$ on cIoU respectively compared to augmented multi-view invariance. Moreover, as the results of (A vs. C) and (A vs. E) depict, the combination of these two different types of invariance complement each other and and further enhances the model's performance. Using pair combination of these two different", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7783", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2b7baad645a34fd19154d0134d9422c38aedcd09c7a9caf896a17d43b6ad4c5f.jpg", + "image_caption": [ + "Figure 3. Sound Localization Results on VGG-SS (top) and SoundNet-Flickr (bottom)." + ], + "image_footnote": [], + "bbox": [ + 81, + 98, + 490, + 373 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/19523afffaeccb4433899c6fbcace7e5da95cdbe3f1e2cb22fe36193fc607b48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 99, + 888, + 373 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5100caa3a583462fe503ccb57d03ac9cd173ca38161f28a5df47f684843ea094.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
k in k-NN10301005001000
cIoU ↑38.8038.8239.4639.9039.94
AUC ↑39.5139.6739.9340.0040.02
", + "bbox": [ + 96, + 407, + 439, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Varying k in conceptually similar sample selection.", + "bbox": [ + 76, + 467, + 447, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "types of consistency elements provides additional supervi sions, invariance and alignments, leading to a more robust representation space and improve sound localization performance.", + "bbox": [ + 76, + 492, + 468, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Feature Alignment. We perform controlled experiments to verify the effect of the feature alignment strategy, and the results are presented in Table 8. Comparing the performance of the proposed model with and without feature alignment, (A vs. B), highlights the importance of this strategy to boost the performance. Further, examining the results of experiments (C vs. D) and (E vs. F) reveals that feature alignment provides additional gains irrespective of the consistency types. These findings indicate that global feature-based alignment helps the optimization of audio-visual correspondence.", + "bbox": [ + 76, + 560, + 468, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of $k$ in conceptually similar sample selection. Selecting an appropriate $k$ value for sampling nearest neighbors is crucial. If this value is set too high, it may result in noisy samples that could disrupt the learning phase. Conversely, if the value is set too low, only very similar samples to the anchor will be provided and it limits semantic invariance. Nevertheless, when compared to Table 8 (E), we observe performance gain throughout the range of $k$ used for the ablation study. Table 9 shows an ablative evaluation of the effect of $k$ value used to select neighborhood samples. The results indicate that an optimal choice is $k = 1000$ . This", + "bbox": [ + 76, + 734, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cfc727a4fc22eeae7874a4d683920c4574791ebd9ce59d0695b65c37667021ce.jpg", + "image_caption": [ + "Figure 4. OGL degrades our sound localization results on SoundNet-Flickr." + ], + "image_footnote": [], + "bbox": [ + 511, + 415, + 880, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "choice of $k$ can be explained by the fact that it provides a balance between semantic similarity and sufficient diversity.", + "bbox": [ + 500, + 517, + 890, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 500, + 558, + 681, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we visualize and compare our sound localization results with the recent prior works on standard benchmarks, namely on VGG-SS and SoundNet-Flickr. The visualized samples in Figure 3 show that localized regions of the proposed method are more compact and accurately aligns with the sounding objects than the other methods. For instance, small size musical instrument is localized accurately compared to the recent methods in the top right column.", + "bbox": [ + 496, + 582, + 890, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also compare our localization results with and without object-guided localization (OGL). As shown in Figure 4, OGL deteriorates our sound localization outputs. OGL captures objectness in a scene, thereby tending to attend to any distinctive objects regardless of whether it is the sound source or not. Therefore, OGL can be helpful when localization totally fails because of the objectness bias in the benchmarks, but it is harmful when the localization is accurate which is the case for the examples shown. This result is consistent with the quantitative result in Table 2, showing that our method with OGL performs worse.", + "bbox": [ + 496, + 718, + 890, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Throughout the paper, we discuss the importance of", + "bbox": [ + 519, + 885, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7784", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1083664842bda8e3a257421b52ffda1b5208ee05bd0b99922367f75c9f5d836e.jpg", + "image_caption": [ + "Figure 5. Interactive Sound Localization of Ours and SLAVC [38]. Our model correctly follows the cross-modal interaction for given different sounds." + ], + "image_footnote": [], + "bbox": [ + 83, + 94, + 467, + 327 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "cross-modal semantic understanding. We demonstrate interactiveness of our method across modalities in Figure 5. Genuine sound source localization should be able to localize objects that are correlated with the sound. To visualize cross-modal interaction, we synthetically pair up the same image with different sounds of objects that are visible in a scene. The examples demonstrate that the proposed method can localize different objects depending on the contexts of sounds, while the competing method can not.", + "bbox": [ + 75, + 397, + 472, + 534 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 566, + 194, + 583 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we investigate cross-modal semantic understanding that has been overlooked in sound source localization studies. We observe that higher sound source localization performance on the current benchmark does not necessarily show higher performance in cross-modal retrieval, despite its causal relevance in reality. To enforce strong understanding of audio-visual semantic matching while maintaining localization capability, we propose semantic alignment with multi-views of audio-visual pairs in a simple yet effective way. The ablation study shows that strong semantic alignment is achieved when both semantic alignment loss and enriched positive pairs are used. We extensively evaluate our method on sound source localization benchmarks including cross-dataset and open-set settings. Moreover, our analyses on cross-modal retrieval and false positive detection verify that the proposed method has strong capability in cross-modal interaction. Our study suggests that sound localization methods should be evaluated not only on localization benchmarks but also on cross-modal understanding tasks.", + "bbox": [ + 75, + 598, + 472, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 90, + 671, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2023-00212845, Multimodal Speech Processing for Human-Computer Interaction). H. Pfister and J. Kim were partially supported by NIH grant R01HD104969. T.-H. Oh was partially supported by IITP grant funded by the Korea government (MSIT) (No. 2021-0-02068, Artificial Intelligence Innovation Hub; No. 2022-0-00290, Visual Intelligence for Space-Time Understanding and Generation based on Multi-layered Visual Common Sense).", + "bbox": [ + 496, + 114, + 893, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 501, + 295, + 596, + 310 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. The conversation: Deep audio-visual speech enhancement. In INTERSPEECH, pages 3244–3248, 2018. 2", + "[2] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. In European Conference on Computer Vision, 2020. 2, 5", + "[3] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In IEEE International Conference on Computer Vision, 2017. 2", + "[4] Relja Arandjelović and Andrew Zisserman. Objects that sound. In European Conference on Computer Vision, 2018. 1", + "[5] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Soundnet: Learning sound representations from unlabeled video. Advances in Neural Information Processing Systems, NeurIPS, 2016. 2, 4", + "[6] Mehdi Azabou, Mohammad Gheshlaghi Azar, Ran Liu, ChiHeng Lin, Erik C Johnson, Kiran Bhaskaran-Nair, Max Dabagia, Bernardo Avila-Pires, Lindsey Kitchell, Keith B Hengen, et al. Mine your own view: Self-supervised learning through across-sample prediction. arXiv preprint arXiv:2102.10106, 2021. 2", + "[7] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In European Conference on Computer Vision, 2020. 2", + "[8] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 3, 4, 5, 6, 7", + "[9] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2020. 4", + "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, 2020. 2, 3, 5" + ], + "bbox": [ + 501, + 319, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7785", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2", + "[12] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[13] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In IEEE Conference on Computer Vision and Pattern Recognition, 2005. 2", + "[14] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asia Conference on Computer Vision, pages 251-263. Springer, 2017. 2", + "[15] Soo-Whan Chung, Hong Goo Kang, and Joon Son Chung. Seeing voices and hearing voices: learning discriminative embeddings using cross-modal self-supervision. In INTER-SPEECH, 2020. 2", + "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition, 2009. 2", + "[17] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In IEEE International Conference on Computer Vision, 2021. 2, 4", + "[18] Mohamed El Banani, Karan Desai, and Justin Johnson. Learning Visual Representations via Language-Guided Sampling. In CVPR, 2022. 2, 4", + "[19] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. ACM Transactions on Graphics (Proceedings of SIGGRAPH), 2018. 2", + "[20] Dennis Fedorishin, Deen Dayal Mohan, Bhavin Jawade, Srirangaraj Setlur, and Venu Govindaraju. Hear the flow: Optical flow-based self-supervised visual sound source localization. In IEEE Winter Conf. on Applications of Computer Vision (WACV), 2023. 1, 5, 6", + "[21] Chuang Gan, Deng Huang, Hang Zhao, Joshua B. Tenenbaum, and Antonio Torralba. Music gesture for visual sound separation. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[22] Chuang Gan, Yiwei Zhang, Jiajun Wu, Boqing Gong, and Joshua B. Tenenbaum. Look, listen, and act: Towards audiovisual embodied navigation. In IEEE International Conference on Robotics and Automation, 2020. 2", + "[23] Ruohan Gao and Kristen Grauman. Co-separating sounds of visual objects. In IEEE International Conference on Computer Vision, 2019. 2", + "[24] Ruohan Gao and Kristen Grauman. Visualvoice: Audiovisual speech separation with cross-modal consistency. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[25] Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal," + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2017. 2", + "[26] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. In Advances in Neural Information Processing Systems, NeurIPS, 2020. 2, 4", + "[27] Simon Haykin and Zhe Chen. The cocktail party problem. Neural computation, 17(9):1875-1902, 2005. 2", + "[28] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[29] Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In International workshop on similarity-based pattern recognition, 2015. 2", + "[30] Di Hu, Feiping Nie, and Xuelong Li. Deep multimodal clustering for unsupervised audiovisual learning. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2, 5, 7", + "[31] Di Hu, Rui Qian, Minyue Jiang, Xiao Tan, Shilei Wen, Errui Ding, Weiyao Lin, and Dejing Dou. Discriminative sounding objects localization via self-supervised audiovisual matching. Advances in Neural Information Processing Systems, NeurIPS, 2020. 1, 7", + "[32] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. Advances in Neural Information Processing Systems, 31, 2018. 2", + "[33] Sizhe Li, Yapeng Tian, and Chenliang Xu. Space-time memory network for sounding object localization in videos. In British Machine Vision Conference, 2021. 1, 2", + "[34] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In International Conference on Learning Representations, 2022. 2, 4", + "[35] Yan-Bo Lin, Hung-Yu Tseng, Hsin-Ying Lee, Yen-Yu Lin, and Ming-Hsuan Yang. Unsupervised sound localization via iterative contrastive learning. arXiv preprint arXiv:2104.00315, 2021. 1, 2", + "[36] Jinxiang Liu, Chen Ju, Weidi Xie, and Ya Zhang. Exploiting transformation invariance and equivariance for self-supervised sound localisation. In ACM International Conference on Multimedia, 2022. 1, 2, 3, 5, 6, 7", + "[37] Josh H McDermott. The cocktail party problem. Current Biology, 19(22):R1024-R1027, 2009. 2", + "[38] Shentong Mo and Pedro Morgado. A closer look at weakly-supervised audio-visual source localization. Advances in Neural Information Processing Systems, NeurIPS, 2022. 1, 2, 3, 4, 5, 6, 7, 9", + "[39] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In European Conference on Computer Vision, 2022. 1, 2, 3, 5, 6, 7", + "[40] Pedro Morgado, Ishan Misra, and Nuno Vasconcelos. Robust audio-visual instance discrimination. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7786", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[42] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2", + "[43] Andrew Owens and Alexei A. Efros. Audio-visual scene analysis with self-supervised multisensory features. In European Conference on Computer Vision, 2018. 2", + "[44] Andrew Owens, Jiajun Wu, Josh H McDermott, William T Freeman, and Antonio Torralba. Ambient sound provides supervision for visual learning. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 801-816. Springer, 2016. 2", + "[45] Takashi Oya, Shohei Iwase, Ryota Natsume, Takahiro Itazuri, Shugo Yamaguchi, and Shigeo Morishima. Do we need sound for sound source localization? In Asia Conference on Computer Vision, 2020. 1", + "[46] Sooyoung Park, Arda Senocak, and Joon Son Chung. MarginNCE: Robust sound localization with a negative margin. In arXiv preprint arXiv:2211.01966, 2022. 6, 7", + "[47] Rui Qian, Di Hu, Heinrich Dinkel, Mengyue Wu, Ning Xu, and Weiyao Lin. Multiple sound sources localization from coarse to fine. In European Conference on Computer Vision, 2020. 1, 2, 5, 7", + "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 3, 5", + "[49] Hyeonggon Ryu, Arda Senocak, In So Kweon, and Joon Son Chung. Hindi as a second language: Improving visually grounded speech with semantically similar samples. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2023. 4", + "[50] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound source in visual scenes. In IEEE Conference on Computer Vision and Pattern Recognition, 2018. 1, 2, 4, 5, 7", + "[51] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound sources in visual scenes: Analysis and applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(5):1605-1619, 2021. 1, 2, 4", + "[52] Arda Senocak, Hyeonggon Ryu, Junsik Kim, and In So Kweon. Learning sound localization better from semantically similar samples. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2022. 1, 2, 4, 5", + "[53] Arda Senocak, Hyeonggon Ryu, Junsik Kim, and In So Kweon. Less can be more: Sound source localization with a classification model. In IEEE Winter Conf. on Applications of Computer Vision (WACV), 2022. 1, 2, 5", + "[54] Zengjie Song, Yuxi Wang, Junsong Fan, Tieniu Tan, and Zhaoxiang Zhang. Self-supervised predictive learning: A" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "negative-free method for sound source localization in visual scenes. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 5, 6", + "[55] Kim Sung-Bin, Arda Senocak, Hyunwoo Ha, Andrew Owens, and Tae-Hyun Oh. Sound to visual scene generation by audio-to-visual latent alignment. In IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[56] Yapeng Tian, Di Hu, and Chenliang Xu. Cyclic co-learning of sounding object visual grounding and sound separation. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[57] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In European Conference on Computer Vision, 2018. 2, 5", + "[58] Efthymios Tzinis, Scott Wisdom, Aren Jansen, Shawn Hershey, Tal Remez, Daniel P. W. Ellis, and John R. Hershey. Into the wild with audioscope: Unsupervised audio-visual separation of on-screen sounds. In International Conference on Learning Representations, 2021. 2", + "[59] Ho-Hsiang Wu, Prem Seetharaman, Kundan Kumar, and Juan Pablo Bello. Wav2CLIP: Learning robust audio representations from clip. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2022. 5", + "[60] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2", + "[61] Haohang Xu, Xiaopeng Zhang, Hao Li, Lingxi Xie, Wenrui Dai, Hongkai Xiong, and Qi Tian. Seed the views: Hierarchical semantic alignment for contrastive representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 4", + "[62] Xudong Xu, Bo Dai, and Lin Dahua. Recursive visual sound separation using minus-plus net. In IEEE International Conference on Computer Vision, 2019. 2", + "[63] Hanyu Xuan, Zhiliang Wu, Jian Yang, Yan Yan, and Xavier Alameda-Pineda. A proposal-based paradigm for self-supervised sound source localization in videos. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1029-1038, 2022. 2", + "[64] Hang Zhao, Chuang Gan, Wei-Chiu Ma, and Antonio Torralba. The sound of motions. In IEEE International Conference on Computer Vision, 2019. 2", + "[65] Hang Zhou, Xudong Xu, Lin Dahua, Xiaogang Wang, and Ziwei Liu. Sep-stereo: Visually guided stereophonic audio generation by associating source separation. In European Conference on Computer Vision, 2020. 2", + "[66] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In European Conference on Computer Vision, 2022. 4, 7" + ], + "bbox": [ + 503, + 92, + 890, + 825 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7787", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Sound Source Localization is All about Cross-Modal Alignment/7b57a83c-7e70-4280-a6ee-f0f550384305_model.json b/2023/Sound Source Localization is All about Cross-Modal Alignment/7b57a83c-7e70-4280-a6ee-f0f550384305_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dd2425a6e558fc71bf8c6a05823ad908a45a540b --- /dev/null +++ b/2023/Sound Source Localization is All about Cross-Modal Alignment/7b57a83c-7e70-4280-a6ee-f0f550384305_model.json @@ -0,0 +1,2334 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.165, + 0.131, + 0.806, + 0.154 + ], + "angle": 0, + "content": "Sound Source Localization is All about Cross-Modal Alignment" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.18, + 0.764, + 0.218 + ], + "angle": 0, + "content": "Arda Senocak\\(^{1*}\\) Hyeonggon Ryu\\(^{1*}\\) Junsik Kim\\(^{2*}\\) Tae-Hyun Oh\\(^{3,4}\\) Hanspeter Pfister\\(^{2}\\) Joon Son Chung\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.222, + 0.793, + 0.241 + ], + "angle": 0, + "content": "\\(^{1}\\) Korea Advanced Institute of Science and Technology \\(^{2}\\) Harvard University" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.241, + 0.83, + 0.258 + ], + "angle": 0, + "content": "\\(^{3}\\)Dept. of Electrical Engineering and Grad. School of Artificial Intelligence, POSTECH" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.258, + 0.862, + 0.277 + ], + "angle": 0, + "content": "\\(^{4}\\)Institute for Convergence Research and Education in Advanced Technology, Yonsei University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.326, + 0.314, + 0.342 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.474, + 0.645 + ], + "angle": 0, + "content": "Humans can easily perceive the direction of sound sources in a visual scene, termed sound source localization. Recent studies on learning-based sound source localization have mainly explored the problem from a localization perspective. However, prior arts and existing benchmarks do not account for a more important aspect of the problem, cross-modal semantic understanding, which is essential for genuine sound source localization. Cross-modal semantic understanding is important in understanding semantically mismatched audio-visual events, e.g., silent objects, or off-screen sounds. To account for this, we propose a cross-modal alignment task as a joint task with sound source localization to better learn the interaction between audio and visual modalities. Thereby, we achieve high localization performance with strong cross-modal semantic understanding. Our method outperforms the state-of-the-art approaches in both sound source localization and cross-modal retrieval. Our work suggests that jointly tackling both tasks is necessary to conquer genuine sound source localization." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.655, + 0.21, + 0.67 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Humans can easily perceive where the sound comes from in a scene. We naturally attend to the sounding direction and associate incoming audio-visual signals to understand the event. To achieve human-level audio-visual perception, sound source localization in visual scenes has been extensively studied [50, 51, 4, 47, 8, 35, 31, 33, 53, 54, 52, 36, 39, 38, 20]. Motivated by that humans learn from natural audio-visual correspondences without explicit supervision, most of the studies have been developed on a fundamental assumption that audio and visual signals are temporally correlated. With the assumption, losses of the sound source localization task are modeled by audio-visual correspondence as a self-supervision signal and are implemented by con" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.31, + 0.892, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.425, + 0.892, + 0.453 + ], + "angle": 0, + "content": "Figure 1. A conceptual difference between prior approaches and our alignment-based sound source localization." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.463, + 0.843, + 0.478 + ], + "angle": 0, + "content": "trasting audio-visual pairs, i.e., contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.893, + 0.734 + ], + "angle": 0, + "content": "While these approaches appear to be unsupervised methods, they strongly rely on partial supervision information; e.g., using supervisedly pretrained vision networks [50, 51, 47, 53, 54, 20] and visual objectness estimators for postprocessing [39, 38]. Without leveraging such strong initial representations, the performance is degraded. Thus, the previous methods are not purely self-supervised approaches. Even further, there are recent studies [45, 39, 38] that point out visual objectness bias in existing sound source localization benchmarks and exploit the objectness prior to improve the localization accuracy. They show that, even without interaction between visual and audio signals, a model may achieve strong accuracy in localization by only referring visual signals alone, which is not the true intention of the sound source localization task. In short, the current evaluation and setting of the sound source localization do not capture the true sound source localization performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.902 + ], + "angle": 0, + "content": "In this work, we first sort out evaluating sound source localization methods by introducing a cross-modal retrieval task as an auxiliary evaluation task. By this task, we can measure whether the learned representation has the capability to accurately interact between audio and visual modalities; i.e., more fine-grained audio-visual correspondence which is essential for genuine sound source localization. This aspect has been missed in existing sound source localization benchmarks. Indeed, our experiments show that higher sound localization performance does not guarantee higher cross-modal retrieval performance." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.887, + 0.336, + 0.901 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7777" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.348 + ], + "angle": 0, + "content": "Second, given this additional criterion, we revisit the importance of semantic understanding shared across audio and visual modalities in both sound source localization and cross-modal retrieval. In the previous methods [50, 51, 54, 47], the cross-modal semantic alignment is induced by instance-level cross-modal contrastive learning, i.e., cross-modal instance discrimination between visual and audio features. However, they are aided by labels or supervisedly pretrained encoder for easing challenging cross-modal feature alignment. Instead, our method learns from scratch supporting the lack of guidance by incorporating multiple positive samples into cross-modal contrastive learning. Specifically, we construct a positive set for each modality using both multi-view [10] and conceptually similar samples [17]. Thereby, we enhance feature alignment and achieve high localization performance and strong cross-modal semantic understanding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.473, + 0.499 + ], + "angle": 0, + "content": "We evaluate our method on the VGG-SS and SoundNet-Flickr benchmarks for sound source localization and cross-modal retrieval. As aforementioned, the sound source localization task is closely related to the cross-modal retrieval task, but our experiments show that existing works have a weak performance correlation between them. This implies that we need to evaluate both tasks for evaluating the genuine sound source localization. The proposed method performs favorably against the recent state-of-the-art approaches in both tasks." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.5, + 0.468, + 0.514 + ], + "angle": 0, + "content": "We summarize the contributions of our work as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.519, + 0.469, + 0.579 + ], + "angle": 0, + "content": "- We analyze that sound source localization benchmarks are not capable of evaluating cross-modal semantic understanding, thereby sound source localization methods may perform poorly in cross-modal retrieval tasks." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.583, + 0.469, + 0.627 + ], + "angle": 0, + "content": "- We propose semantic alignment to improve cross-modal semantic understanding of sound source localization models." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.631, + 0.469, + 0.691 + ], + "angle": 0, + "content": "- We expand semantic alignment with multi-views and conceptually similar samples which leads to state-of-the-art performance on both sound source localization and cross-modal retrieval." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.519, + 0.469, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.696, + 0.216, + 0.71 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.714, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Sound source localization. Sound source localization in visual scenes has been investigated by exploiting correspondences between audio and visual modalities. The most widely used approach for sound source localization is cross-modal attention [50, 51, 57] with contrastive loss [13, 29, 42]. Later, the attention-based method is improved by intra-frame hard sample mining [8], iterative contrastive learning with pseudo labels [35], feature regularization [36], positive mining [52], negative free learning [54] with stop-gradient operation [12], or momentum encoders [38]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.072, + 0.875, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Typically, an image encoder is pretrained on ImageNet [16] and an audio encoder is pretrained on AudioSet [25] in supervised ways." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.287 + ], + "angle": 0, + "content": "Some sound localization approaches exploit additional semantic labels [47, 33, 53] or object prior [39, 63]. Semantic labels are used to pretrain audio and vision encoders with classification loss [33, 53] or refine audio-visual feature alignment [47]. A more explicit way to refine localization output is to use object prior. EZVSL [39] proposes post-processing to combine attention based localization output with a pretrained visual feature activation map. Similarly, Xuan et al. [63] propose to combine off-the-shelf object proposals with attention based sound localization results. However, postprocessing by object prior may generate a false positive output as it is solely based on vision without audio-visual interaction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.893, + 0.41 + ], + "angle": 0, + "content": "In addition to the localization, there has been an attempt to localize sounding objects and recover the separated sounds simultaneously, also known as the cocktail party problem [27, 37]. The separation of sound mixture is achieved by predicting masks of spectrogram guided by visual features [19, 1, 64, 23, 62, 21, 2, 65, 24, 58, 56]. Furthermore, a number of recent papers are presented on audio-visual navigation for a given sound source [7, 22]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.892, + 0.493 + ], + "angle": 0, + "content": "Self-supervised representation learning. In a broader categorization, sound source localization belongs to self-supervised multimodal learning. Our work is also relevant to self-supervised audio-visual representation learning, and other multimodal learning studies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.893, + 0.659 + ], + "angle": 0, + "content": "Contrastive learning aims to learn robust representations from large-scale raw data without annotations. Recent representation learning approaches [60, 10, 28, 11] use instance discrimination by contrastive learning [13, 29, 42] as a pretext task with notable advancements in visual recognition tasks. Recently, positive mining by nearest-neighbor search are used to learn representations of images [17, 18, 61], videos [26], neural recordings [6], and text-image [34]. In this work, we expand the previous works by incorporating both multi-views and conceptually similar samples into audio-visual modalities for cross-modal feature alignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.894, + 0.902 + ], + "angle": 0, + "content": "A series of audio-visual representation learning studies have shown that audio and visual contents in a video are correlated, therefore a visual representation can be learned by sound prediction [44] or audio representation can be distilled from visual representation [5, 55]. Later, a variety of joint audio-visual representation learning methods are proposed with an assumption that there is a semantic [3, 30, 41, 40] or temporal [14, 43, 32, 15] correspondence between them. However, simply learning sound source localization by audio-visual correspondence with instance discrimination ignores the semantic similarity of audio-visual contents among samples, introducing false negatives or positives. In order to mitigate this issue, clustering [30], sampling [41], weighting [40], and hard mining [32] are proposed. Similarly, in this work, we go beyond instance discrimination by using multiple positive samples" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7778" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.088, + 0.895, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.285, + 0.895, + 0.329 + ], + "angle": 0, + "content": "Figure 2. Our sound source localization framework. Our model constructs multiple positive pairs with augmentation and Nearest Neighbor Search (Conceptually Similar Samples). By using these newly constructed 9 pairs, our model employs spatial localization, \\( s_L \\), and semantic feature alignment, \\( s_A \\), for each pair to learn a better sound source localization ability." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.353, + 0.429, + 0.369 + ], + "angle": 0, + "content": "to enforce semantic understanding across modalities." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.383, + 0.169, + 0.399 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.409, + 0.218, + 0.423 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.471, + 0.501 + ], + "angle": 0, + "content": "Contrastive learning learns representation by containing positive and negative pairs. Given an encoded query sample \\( q \\) and its encoded positive pair \\( k^{+} \\) and negative pairs \\( k \\), the loss can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.511, + 0.47, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = - \\log \\frac {\\exp (q \\cdot k ^ {+} / \\tau)}{\\sum_ {i} \\exp (q \\cdot k _ {i} / \\tau)} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.557, + 0.33, + 0.572 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the temperature parameter." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.47, + 0.686 + ], + "angle": 0, + "content": "Cross-modal contrastive learning extends contrastive learning across multiple modalities. In sound source localization, audio-visual correspondence is used to define positive and negative cross-modal pairs. With an audio-visual dataset \\(\\mathcal{D} = \\{(v_i,a_i):i = 1,\\dots,N\\}\\) and its encoded features \\(\\mathbf{v}_i = f_v(v_i)\\) and \\(\\mathbf{a}_i = f_a(a_i)\\), cross-modal contrastive learning loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.698, + 0.47, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i} = - \\log \\frac {\\exp (s (\\mathbf {v} _ {i} , \\mathbf {a} _ {i}) / \\tau)}{\\sum_ {j} \\exp (s (\\mathbf {v} _ {i} , \\mathbf {a} _ {j}) / \\tau)} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.745, + 0.47, + 0.79 + ], + "angle": 0, + "content": "where \\( s \\) is a cross-modal similarity function. The cross-modal contrastive loss Eq. (2) can be extended to symmetric form [48] as used in a few previous works [39, 38]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.363, + 0.818 + ], + "angle": 0, + "content": "3.2. Cross-Modal Feature Alignment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We consider both spatial localization and semantic feature alignment for sound source localization. To this end, we use two different similarity functions \\( s_L \\) and \\( s_A \\) for contrastive learning (Eq. (2)), \\( s_L \\) for localization and \\( s_A \\) for cross-modal feature alignment." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.353, + 0.892, + 0.429 + ], + "angle": 0, + "content": "Recent studies rely on audio-visual spatial correspondence maps to learn sound source localization by contrasting them. Given a spatial visual feature \\(\\mathbf{v} \\in \\mathbb{R}^{c \\times h \\times w}\\) and audio feature \\(\\mathbf{a} \\in \\mathbb{R}^{\\mathbf{c}}\\), audio-visual similarity with a correspondence map can be calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.437, + 0.892, + 0.476 + ], + "angle": 0, + "content": "\\[\ns _ {L} (\\mathbf {v}, \\mathbf {a}) = \\sum_ {x y \\in M} \\frac {1}{| M |} \\frac {\\mathbf {v} ^ {x y} \\cdot \\mathbf {a}}{\\| \\mathbf {v} ^ {x y} \\| \\| \\mathbf {a} \\|} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.485, + 0.892, + 0.56 + ], + "angle": 0, + "content": "where \\(\\mathbf{v}^{xy}\\) is a feature vector at location \\((x,y)\\), and \\(M\\) is an optional binary mask when an annotation or pseudomask [8, 36] is available. Since we assume no supervision for sound source localization, we do not use any mask, therefore, \\(M = \\mathbf{1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.893, + 0.834 + ], + "angle": 0, + "content": "The contrastive loss with localization similarity \\( s_L \\) enforces location dependent alignment giving sparse but strong audio-visual correspondence which enables to perform localization. However, our empirical studies on cross-modal retrieval indicate that strong localization performance does not guarantee semantic understanding. To overcome the low semantic understanding in recent studies, we propose to add instance-level contrastive loss. Instance-level contrasting encapsulates the whole context in a scene, enforcing better audio-visual semantic alignment. However, instance-level contrasting may smooth out spatial discriminativeness learned by Eq. (3). Inspired by SimCLR [10], we adopt a projection layer to align audio-visual semantics in a projection space. The projection layer separates the latent space of localization and semantic alignment, thereby preventing the alignment loss smoothing out the spatial discriminativeness. The similarity function for cross-modal feature alignment is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.843, + 0.892, + 0.878 + ], + "angle": 0, + "content": "\\[\ns _ {A} (\\mathbf {v}, \\mathbf {a}) = \\frac {p _ {v} (\\operatorname {a v g} (\\mathbf {v})) \\cdot p _ {a} (\\mathbf {a})}{\\| p _ {v} (\\operatorname {a v g} (\\mathbf {v})) \\| \\| p _ {a} \\mathbf {a} \\|} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathrm{avg}(\\cdot)\\) is spatial average pooling, \\(p_v\\) is a projection" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7779" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "layer for visual features, and \\( p_a \\) is a projection layer for audio features." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.134, + 0.438, + 0.151 + ], + "angle": 0, + "content": "3.3. Expanding with Multiple Positive Samples" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.159, + 0.47, + 0.37 + ], + "angle": 0, + "content": "Typically, contrastive learning contrasts between one positive pair and multiple negative pairs as shown in Eq. (1). In audio-visual learning, by an audio-visual correspondence assumption, an audio-image pair from the same clip is used as a positive pair while negative pairs are sampled from different clips. However, single-instance discrimination may not be sufficient to achieve strong cross-modal alignment. In this section, we expand contrastive learning beyond single instance discrimination by positive set construction and pairing them. To construct a positive set, we incorporate both hand-crafted positive and conceptual positive samples for each modality. Later, we adjust the contrastive learning to incorporate multiple positive pairs to enforce cross-modal alignment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.47, + 0.59 + ], + "angle": 0, + "content": "Obtaining hand-crafted positive samples. Using randomly augmented samples as positive multi-view pairs are widely adopted in self-supervised representation learning, i.e., instance discrimination. Similarly, we extend a single anchor audio-image pair to multiple positive pairs by applying simple augmentations on image and audio samples separately. While we utilize common image transformations on images, we apply temporal shifting to audiOs. It is worth noting that sound source localization task learns from the underlying semantic consistency rather than subtle time differences as in videos. Thus, a slight shift in the audio may not alter contextual information significantly. As a result of hand-crafted multi-view positive pair generation, we obtain additional \\(\\mathbf{v}^{aug}\\) and \\(\\mathbf{a}^{aug}\\) samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Obtaining conceptual positive samples. Apart from manually created augmented views, we additionally expand our positive set with conceptually similar samples. The sampling strategy with nearest neighbor search can be performed in a various way, such as on-the-fly sampling [17, 49, 61, 34], sampling by pretrained encoders [52], or guided sampling [26, 18] using another modality. For selecting our conceptually similar samples, we utilize pretrained encoders. Note that pretrained encoders trained either with supervised or self-supervised learning are effective in positive sample mining as shown in the experiment section. By employing readily available image and audio encoders, we use the \\(k\\)-nearest neighborhood search to sample semantically similar samples in both modalities. In particular, given a pair of image and audio, we compute cosine similarity with all other samples and choose the top-\\(k\\) most similar samples among the training set for each modality. From a set of \\(k\\) samples, we randomly select one sample to obtain conceptually similar samples for each modality, \\(\\mathbf{v}^{conc}\\) and \\(\\mathbf{a}^{conc}\\). By utilizing the conceptually similar samples as" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.122 + ], + "angle": 0, + "content": "positive samples, our model expands semantic understanding." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.13, + 0.892, + 0.296 + ], + "angle": 0, + "content": "Pair Construction. Once we obtain the conceptual and hand-crafted positive samples for each modality, we proceed to create 9 distinct audio-visual pairs by pairing \\(\\mathbf{V} = \\{\\mathbf{v},\\mathbf{v}^{aug},\\mathbf{v}^{conc}\\}\\) and \\(\\mathbf{A} = \\{\\mathbf{a},\\mathbf{a}^{aug},\\mathbf{a}^{conc}\\}\\). This is done to ensure semantic alignment and consistency between them through contrastive learning. The negative pairs are randomly paired from the remaining samples in a training set. It is worth noting that some of these pairs are a combination of hand-crafted and conceptually similar samples, which further enhances the feature alignment of our model during training." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.307, + 0.604, + 0.323 + ], + "angle": 0, + "content": "3.4. Training" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.331, + 0.892, + 0.391 + ], + "angle": 0, + "content": "Our loss formulation incorporates both localization and instance-level similarity functions with multiple positive pairs constructed by augmentation and conceptually similar sample search. The final loss term is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.403, + 0.892, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {i} = - \\sum_ {\\mathbf {v} _ {i} \\in \\mathbf {V}} \\sum_ {\\mathbf {a} _ {i} \\in \\mathbf {A}} \\left[ \\log \\frac {\\exp \\left(s _ {L} \\left(\\mathbf {v} _ {i} , \\mathbf {a} _ {i}\\right) / \\tau\\right)}{\\sum_ {j} \\exp \\left(s _ {L} \\left(\\mathbf {v} _ {i} , \\mathbf {a} _ {j}\\right) / \\tau\\right)} \\right. \\tag {5} \\\\ \\left. + \\log \\frac {\\exp \\left(s _ {A} \\left(\\mathbf {v} _ {i} , \\mathbf {a} _ {i}\\right) / \\tau\\right)}{\\sum_ {j} \\exp \\left(s _ {A} \\left(\\mathbf {v} _ {i} , \\mathbf {a} _ {j}\\right) / \\tau\\right)} \\right] \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.5, + 0.802, + 0.515 + ], + "angle": 0, + "content": "where \\(\\mathbf{V}\\) and \\(\\mathbf{A}\\) indicate positive sample sets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.531, + 0.634, + 0.547 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.556, + 0.892, + 0.676 + ], + "angle": 0, + "content": "Our proposed method for sound source localization is validated through experiments conducted on VG-GSound [9] and SoundNet-Flickr [5]. First, we conduct a quantitative analysis to evaluate the accuracy of the localization, cross-modal retrieval, and the impact of various components of our model. Then, we visualize our sound source localization results across different categories of sounds." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.688, + 0.676, + 0.705 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Datasets. Our method is trained using the VGGSound [9] and SoundNet-Flickr-144K [50, 51]. VGGSound is an audio-visual dataset containing around \\(\\sim 200\\mathrm{K}\\) videos. SoundNet-Flickr-144K set is the subset of SoundNetFlickr [5]. After training, we test the sound localization performance with VGG-SS [8] and SoundNet-Flickr-Test [50] datasets for the main experiments. These evaluation sets have bounding box annotations of sound sources for \\(\\sim 5\\mathrm{K}\\) and 250 samples, respectively. Moreover, we employ the AVSBench [66] and Extended VGGSound/SoundNetFlickr [38] datasets for additional evaluations. AVSBench dataset provides binary segmentation maps that show the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7780" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.472, + 0.322 + ], + "angle": 0, + "content": "
MethodPre. VisionVGG-SSFlickr-SoundNet
cIoU ↑AUC ↑cIoU ↑AUC ↑
Attention [50]CVPR18✓18.5030.2066.0055.80
CoarseToFine [47]ECCV20✓29.1034.80--
LCBM [53]WACV22✓32.2036.60--
LVS [8]†CVPR21✗30.3036.4072.4057.80
LVS [8]CVPR21✗34.4038.2071.9058.20
HardPos [52]ICASSP22✗34.6038.0076.8059.20
SSPL (w/o PCM) [54]CVPR22✓27.0034.8073.9060.20
SSPL (w/ PCM) [54]CVPR22✓33.9038.0076.7060.50
EZ-VSL (w/o OGL) [39]ECCV22✓35.9638.2078.3161.74
SSL-TIE [36]ACMM22✗38.6339.6579.5061.20
SLAVC (w/o OGL) [38]NeurIPS22✓37.7939.4083.60-
Ours
↓ NN Search w/ Supervised Pre. Encoders✗39.9440.0279.6063.44
↓ NN Search w/ Self-Supervised Pre. Encoders✗39.2039.7079.2063.00
with OGL:
EZ-VSL (w/ OGL) [39]ECCV22✓38.8539.5483.9463.60
SLAVC (w/ OGL) [38]NeurIPS22✓39.80-86.00-
Ours (w/ OGL)
↓ NN Search w/ Supervised Pre. Encoders✗42.6441.4882.4064.60
↓ NN Search w/ Self-Supervised Pre. Encoders✗42.4741.4282.8064.48
with Optical Flow:
HearTheFlow [20]WACV23✓39.4040.0084.8064.00
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.322, + 0.47, + 0.391 + ], + "angle": 0, + "content": "Table 1. Quantitative results on the VGG-SS and SoundNet-Flickr test sets. All models are trained with 144K samples from VGG-Sound and tested on VGG-SS and SoundNet-Flickr. \\(\\dagger\\) is the result of the model released on the official project page. SLAVC [38] does not provide AUC scores." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.47, + 0.471 + ], + "angle": 0, + "content": "audio-visually correspondent pixels for roughly 5k five-second videos belonging to 23 categories. Lastly, the Extended VGGSound /SoundNet-Flickr dataset, proposed by [38], is used to understand non-visible sound sources." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.472, + 0.782 + ], + "angle": 0, + "content": "Implementation details. We use two ResNet18 models for both audio and vision encoding. Unlike prior approaches, we do not fine-tune (or use a pretrained) a visual encoder from ImageNet pretrained weights. Instead, we train both the audio and vision encoders from scratch. We preprocess images and audios following the previous works [8, 52]. To create multiple pairs, we utilize both NN search and generic augmentation approaches. For NN search, we experiment on two different setups to retrieve k conceptually similar samples: (1) For supervisedly pretrained encoder experiments, We employ ResNet and VGGSound models pretrained on ImageNet and VGGSound respectively, (2) For self-supervisedly pretrained encoder experiments, we utilize the CLIP [48] Vision Encoder and Wav2CLIP [59] Audio Encoder. We use \\(k = 1000\\) for the experiments. To perform image augmentations, we follow the augmentations used in SimCLR [10]. For audios, we randomly select timewindow shifts in a time axis. The model is trained for 50 epochs with Adam Optimizer and a learning rate of 0.0001. \\(\\tau\\) is set to 0.07 in contrastive learning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.794, + 0.273, + 0.81 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Comparison with strong baselines. In this section, we conduct a comparative analysis of our sound source localization method against existing approaches. We carry out our evaluations in two settings, following previous approaches. Firstly, we train our model on VGGSound-144K" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.895, + 0.363 + ], + "angle": 0, + "content": "
MethodPre. VisioncIoU ↑AUC ↑
Attention[50]CVPR18✓66.0055.80
DMC[30]CVPR19✓67.1056.80
LVS [8]†CVPR21X67.2056.20
LVS [8]CVPR21X69.9057.30
HardPos [52]ICASSP22X75.2059.70
SSPL (w/o PCM) [54]CVPR22✓69.9058.00
SSPL (w/ PCM) [54]CVPR22✓75.9061.00
EZ-VSL (w/o OGL) [39]ECCV22✓71.8958.81
SSL-TIE [36]ACMM22X81.5061.10
SLAVC (w/o OGL) [38]NeurIPS22✓--
Ours
↓ NN Search w/ Supervised Pre. EncodersX85.2062.20
↓ NN Search w/ Self-Supervised Pre. EncodersX84.8062.66
with OGL:
EZ-VSL (w/ OGL) [39]ECCV22✓83.1363.06
SLAVC (w/ OGL) [38]NeurIPS22✓--
Ours (w/ OGL)
↓ NN Search w/ Supervised Pre. EncodersX84.0064.16
↓ NN Search w/ Self-Supervised Pre. EncodersX84.4064.38
with Optical Flow:
HearTheFlow [20]WACV23✓86.5063.90
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.366, + 0.895, + 0.422 + ], + "angle": 0, + "content": "Table 2. Quantitative results on the SoundNet-Flickr test set. All models are trained and tested on the SoundNet-Flickr 144K dataset. \\(\\dagger\\) is the result of the model from the official project page. SLAVC [38] does not provide results with SoundNet-Flickr 144K." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.581 + ], + "angle": 0, + "content": "and evaluate it on VGG-SS and SoundNet-Flickr test sets. Secondly, we train our model on SoundNet-Flickr-144K and evaluate it on the SoundNet-Flickr test set. It is important to note that all the compared models are trained using the same amount of data. AVEL [57], AVObject [2], and LCBM [53] models rely on video input, and as such, they cannot be evaluated on the SoundNet-Flickr dataset, which contains static image and audio pairs. We present our results in Table 1 and Table 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Our proposed model achieves higher performance compared to prior approaches on both test sets. Specifically, it yields a \\(+2.15\\%\\) cIoU and \\(+0.6\\%\\) AUC improvement on VGGSS, as well as a \\(+3.7\\%\\) cIoU improvement on SoundNet-Flickr compared to the state-of-the-art methods that uses pretrained vision encoder. It is worth highlighting that unlike the majority of previous works, our proposed model does not utilize a vision encoder pretrained on ImageNet in a sound source localization backbone. This is because, as discussed in Mo et al. [38], using supervisedly pretrained vision encoders makes the sound source localization problem a weakly supervised problem. However, it is worth noting that even without using a pretrained vision encoder, our method achieves state-of-the-art performance on both experiments that are presented in Table 1 and Table 2. We demonstrate the performance of our model with the pretrained models learned through supervised learning (NN Search w/ Supervised Pre. Encoders) and with models that are pretrained through self-supervised learning (NN Search w/ Self-Supervised Pre. Encoders) in NN Search module. As the results indicate, using self-supervised pre" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7781" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.457, + 0.286 + ], + "angle": 0, + "content": "
Test ClassMethodPre. VisioncIoU ↑AUC ↑
Heard 110LVS [8]CVPR21X28.9036.20
EZ-VSL(w/o OGL) [39]ECCV22✓31.8636.19
SLAVC(w/o OGL) [38]NeurlPS22✓35.84-
OursX38.3139.05
with OGL:
EZ-VSL(w/o OGL) [39]ECCV22✓37.2538.97
SLAVC(w/o OGL) [38]NeurlPS22✓38.22-
Ours(w/ OGL)X41.8540.93
with Optical Flow:
HearTheFlow [20]WACV23✓37.3038.60
Unheard 110LVS [8]CVPR21X26.3034.70
EZ-VSL(w/o OGL) [39]ECCV22✓32.6636.72
SLAVC(w/o OGL) [38]NeurlPS22✓36.50-
OursX39.1139.80
with OGL:
EZ-VSL(w/o OGL) [39]ECCV22✓39.5739.60
SLAVC(w/o OGL) [38]NeurlPS22✓38.87-
Ours(w/ OGL)✓42.9441.54
with Optical Flow:
HearTheFlow [20]WACV23✓39.3040.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.287, + 0.468, + 0.315 + ], + "angle": 0, + "content": "Table 3. Comparison results on open-set audio-visual localization experiments trained and tested on the splits of [8, 39, 20]." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.328, + 0.441, + 0.383 + ], + "angle": 0, + "content": "
Test ClassMethodPre. VisioncIoU ↑AUC ↑
Heard 110SSSL-TIE [36]ACMM22X39.0040.30
OursX41.2041.00
Unheard 110SSSL-TIE [36]ACMM22X36.5038.60
OursX36.9038.59
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.383, + 0.468, + 0.411 + ], + "angle": 0, + "content": "Table 4. Comparison results on open set audio-visual localization experiments trained and tested on the splits of [36]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.468, + 0.551 + ], + "angle": 0, + "content": "trained encoders in NN Search performs on par with the supervised pretrained encoders in NN Search. This shows that our model does not depend on supervised pretrained encoders for the NN search module and can utilize any type of pretrained encoder feature for nearest neighbor search. Note that these pretrained encoders are not used in the backbone networks of the sound source localization module but only in the NN Search Module, as illustrated in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.468, + 0.749 + ], + "angle": 0, + "content": "We also discuss the methods employed by previous studies, such as SSPL [54] which utilizes a sub-module called PCM to reduce the impact of background noise, HTF [20] which utilizes Optical Flow, and EZ-VSL [39] which refines its initial audio-visual localization outcomes through object guidance obtained from an ImageNet pretrained visual encoder. Our model, on the other hand, and any of its variations do not require any task-specific modules or operations to achieve the state-of-the-art (SOTA) results. This suggests that using additional semantic and multi-view correspondence, as well as feature alignment, provides more varied and robust supervision for better aligned audio and visual features, as opposed to using task-specific approaches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.902 + ], + "angle": 0, + "content": "The quantitative results presented in Table 1 and Table 2 also showcase the performance of previous methods that utilize object guidance to evaluate their final sound source localizations. Our model outperforms all previous methods that employ object guidance on the VGG-SS test set and achieves comparable results on the SoundNet-Flickr test set, even though our model does not use object guided refinement (OGL). Additionally, we acknowledge that the addition of OGL to our audio-visual localization results in improvement on the VGGSS test set, while degrading perfor" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.891, + 0.169 + ], + "angle": 0, + "content": "
Test SetMethodPre. VisionmIoU ↑F-Score ↑
S4LVS (w/o OGL) [8]CVPR21X26.933.6
EZ-VSL (w/o OGL) [39]ECCV22✓27.634.2
SLAVC (w/o OGL) [38]NeuRTPS22✓28.134.6
Ours (w/o OGL)
↓ NN Search w/ Supervised Pre. EncodersX29.635.9
↓ NN Search w/ Self-Supervised Pre. EncodersX29.335.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.18, + 0.892, + 0.207 + ], + "angle": 0, + "content": "Table 5. Quantitative results on AVS Bench S4 dataset. All models are trained on the VGGSound 144K dataset." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.22, + 0.897, + 0.308 + ], + "angle": 0, + "content": "
ModelPre. VisionA → II → A
R@1R@5R@10R@1R@5R@10
LVS [8]CVPR21X3.8712.3520.734.9014.2921.37
EZ-VSL [39]ECCV22✓5.0115.7324.8114.233.5145.18
SSL-TIE [36]MM22X10.2930.6843.7612.7629.5839.72
SLAVC [38]NeurIPS22✓4.7713.0819.106.1221.1632.12
Ours
↓ NN Search w/ Supervised Pre. EncodersX16.4736.9949.0020.0942.3853.66
↓ NN Search w/ Self-Supervised Pre. EncodersX14.3137.8149.1718.0038.3949.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.308, + 0.892, + 0.349 + ], + "angle": 0, + "content": "Table 6. Summary of retrieval recall scores for all models. All of the models are trained on VGGSound 144K data and retrieval is performed on entire VGG-SS dataset, containing \\(\\sim 5\\mathrm{K}\\) samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.892, + 0.498 + ], + "angle": 0, + "content": "mance on the SoundNet-Flickr test set. In contrast, prior methods see modest improvements when utilizing OGL. This can be explained by the fact that our model is already accurately localizing the sounding objects, and object guidance can interfere with localization results by introducing visual regions that are not sounding (refer to Section 4.4 for visual results). Unlike prior methods, we do not use OGL in our architecture for the remainder of this paper, unless it is being directly compared with OGL-based methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.892, + 0.59 + ], + "angle": 0, + "content": "Finally, in comparison to HearTheFlow, which utilizes an additional Optical Flow modality, our method outperforms it on the VGGSS test set, and achieves slightly lower performance on the SoundNet-Flickr test set without utilizing any additional modalities, but instead relying on better audio-visual correspondence and alignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Open Set Audio-Visual Localization. The study by Chen et al. [8] evaluates the generalization ability of sound source localization methods in an open set scenario. This involves testing the models on categories that are both present in the training data (heard) and categories that are not present (unheard). To accomplish this, 110 randomly selected categories from the VGGSound dataset are used for training, while another disjoint set of 110 categories are reserved for evaluation to ensure the model have never seen or heard them before. It should be noted that not all previous works use the same train/test splits. While some works, including [8, 39, 38, 46], share the same splits,[36] uses a different split. Therefore, to ensure a fair comparison, we conduct experiments on both splits, evaluating on test samples from both heard and unheard categories. The results are shown in Table 3 and Table 4. Our model outperforms existing approaches on both categories, regardless of the train/test splits used. Specifically, in Table 3, our model (w/o OGL) even surpasses the other models that use OGL. Previous approaches draw different conclusions from" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7782" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.193 + ], + "angle": 0, + "content": "
MethodPre. VisionExtended Flickr-SoundNetExtended VGG-SS
AP ↑max-F1 ↑LocAcc ↑AP ↑max-F1 ↑LocAcc ↑
CoarseToFine [47]ECCV20✓0.0038.2047.200.0019.8021.93
LVS [8]CVPR21✗9.8017.9019.605.159.9010.43
Attention10k [50]CVPR18✓15.9824.0034.166.7013.1014.04
DMC [30]CVPR19✓25.5641.8052.8011.5320.3022.63
DSOL [31]NeuIPS20✓38.3249.4072.9116.8425.6026.87
OGL [39]ECCV22-40.2055.7077.2018.7330.9036.58
EZ-VSL (w/o OGL) [39]ECCV22✓46.3054.6066.4024.5530.9031.58
SLAVC (w/o OGL) [38]NeuIPS22✓51.6359.1083.6032.9540.0037.79
Ours
↓ NN Search w/ Supervised Pre. Encoders✗64.4366.9079.6034.7340.7039.94
↓ NN Search w/ Self-Supervised Pre. Encoders✗62.6766.1079.2033.0940.0039.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.195, + 0.47, + 0.251 + ], + "angle": 0, + "content": "Table 7. Quantitative results on the Extended VGG-SS and Extended SoundNet-Flickr sets. All models are trained with 144K samples from VGG-Sound. The results of the prior approaches are obtained from [38]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.264, + 0.471, + 0.476 + ], + "angle": 0, + "content": "these open set experiments. While some conclude that their models have strong generalization ability because their performance in unheard categories is higher than heard categories [39, 38, 46], the other works that cannot achieve the same trend discuss that this is expected since their models are dealing with unseen categories [36]. However, our results show that these conclusions are highly dependent on the chosen train/test splits. Our model performs better than existing works in both splits, but there is no uniform trend in between two splits. While our method performs better on unheard categories in the splits of [8, 39, 38, 46], it performs worse on unheard categories in the split of [36]. Therefore, we conclude that the observed trends are highly dependent on the randomly selected train/test splits." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.484, + 0.47, + 0.651 + ], + "angle": 0, + "content": "AVSBench [66]. To demonstrate the precise sound localization ability of our model, we conduct experiments on the AVSBench S4 dataset. The dataset's objective is to detect audio-visual correspondence and correlation at the pixel level. To make a fair comparison, we use some of the self-supervised sound source localization methods mentioned earlier. All models are trained on VGGSound-144K and directly assessed on the AVSBench S4 dataset without any further fine-tuning (zero-shot setting). Our results, which are presented in Table 5, indicate that our method achieves the highest performance, as in the previous experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Retrieval. We evaluate sound localization models on the VGG-SS dataset for cross-modal retrieval. As shown in Table 6, our method clearly outperforms other state-of-the-art methods. One interesting observation is that EZ-VSL [39] notably performs better than SLAVC [38] on cross-modal retrieval, while SLAVC performs better on sound source localization in Table 1. This shows that with the current benchmark evaluations, better sound localization performance does not guarantee better audio-visual semantic understanding, thereby we need to additionally evaluate sound source localization methods on cross-modal understanding tasks. Another observation is that the performance gap between our method and the strongest competitor SSL-TIE [36] is notably larger on cross-modal retrieval than sound source localization. This is due to the strong cross-modal feature alignment of our method that is over" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.891, + 0.204 + ], + "angle": 0, + "content": "
SemanticMulti-ViewFeature AlignmentcIoU ↑AUC ↑
(A)✓✓✓39.9440.02
(B)✓✓✗39.1039.44
(C)✓✗✓38.7539.34
(D)✓✗✗38.2438.90
(E)✗✓✓38.3039.38
(F)✗✓✗37.7239.19
(G)✗✗✓34.9337.94
(H)✗✗✗34.2237.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.207, + 0.892, + 0.235 + ], + "angle": 0, + "content": "Table 8. Ablation studies on our proposed method to see the impact of each main component." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.256, + 0.847, + 0.271 + ], + "angle": 0, + "content": "looked in the sound source localization benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.279, + 0.892, + 0.445 + ], + "angle": 0, + "content": "Extended Flickr and VGG-SS datasets. The prior study [38] points out that the current sound source localization benchmarks overlook false positive detection. It is because the evaluation samples always contain at least a sounding object in a scene; thus cannot capture false positive outputs, e.g., silent objects or off-screen sounds. To analyze false positive detection, Mo and Morgado [38] extended the benchmarks with non-audible, non-visible, and mismatched audio-visual samples. The expectation is that a sound source localization model should not localize any objects when audio-visual semantics do not match." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.893, + 0.582 + ], + "angle": 0, + "content": "The experiment with the extended datasets in Table 7 shows that our method performs favorably against state-of-the-art competitors. Our method performs better than the competing methods in false positive detection measured by AP and max-F1, while SLAVC [38] achieves better localization performance on Extended Flickr-SoundNet. As both false positive detection and cross-modal retrieval require cross-modal interaction, our method shows strong performance on both tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.596, + 0.665, + 0.611 + ], + "angle": 0, + "content": "4.3. Ablation Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.892, + 0.711 + ], + "angle": 0, + "content": "We conduct a series of experiments in order to verify our design choices and make further analysis. To save computational time and resources, we primarily perform ablation studies by training our model on VGGSound-144K with NN Search w/ Supervised Pre. Encoders setup and evaluating it on VGG-SS. Results are in Table 8." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Impact of Semantic and Multi-View Invariance. In order to understand the impact of each type of invariance (consistency), we analyze the performance of our model with different type of invariance methodologies in Table 8. As the results of (C vs. E) and (D vs. F) reveal, using semantically similar samples (semantic invariance) produces better performance \\((+0.45\\%)\\) and \\(+0.5\\%\\) on cIoU respectively compared to augmented multi-view invariance. Moreover, as the results of (A vs. C) and (A vs. E) depict, the combination of these two different types of invariance complement each other and and further enhances the model's performance. Using pair combination of these two different" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7783" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.099, + 0.491, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.1, + 0.89, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.209, + 0.387, + 0.75, + 0.401 + ], + "angle": 0, + "content": "Figure 3. Sound Localization Results on VGG-SS (top) and SoundNet-Flickr (bottom)." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.409, + 0.44, + 0.466 + ], + "angle": 0, + "content": "
k in k-NN10301005001000
cIoU ↑38.8038.8239.4639.9039.94
AUC ↑39.5139.6739.9340.0040.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.468, + 0.448, + 0.482 + ], + "angle": 0, + "content": "Table 9. Varying k in conceptually similar sample selection." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.493, + 0.469, + 0.553 + ], + "angle": 0, + "content": "types of consistency elements provides additional supervi sions, invariance and alignments, leading to a more robust representation space and improve sound localization performance." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.561, + 0.469, + 0.728 + ], + "angle": 0, + "content": "Impact of Feature Alignment. We perform controlled experiments to verify the effect of the feature alignment strategy, and the results are presented in Table 8. Comparing the performance of the proposed model with and without feature alignment, (A vs. B), highlights the importance of this strategy to boost the performance. Further, examining the results of experiments (C vs. D) and (E vs. F) reveals that feature alignment provides additional gains irrespective of the consistency types. These findings indicate that global feature-based alignment helps the optimization of audio-visual correspondence." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.735, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Impact of \\( k \\) in conceptually similar sample selection. Selecting an appropriate \\( k \\) value for sampling nearest neighbors is crucial. If this value is set too high, it may result in noisy samples that could disrupt the learning phase. Conversely, if the value is set too low, only very similar samples to the anchor will be provided and it limits semantic invariance. Nevertheless, when compared to Table 8 (E), we observe performance gain throughout the range of \\( k \\) used for the ablation study. Table 9 shows an ablative evaluation of the effect of \\( k \\) value used to select neighborhood samples. The results indicate that an optimal choice is \\( k = 1000 \\). This" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.416, + 0.882, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.48, + 0.892, + 0.507 + ], + "angle": 0, + "content": "Figure 4. OGL degrades our sound localization results on SoundNet-Flickr." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.518, + 0.892, + 0.548 + ], + "angle": 0, + "content": "choice of \\( k \\) can be explained by the fact that it provides a balance between semantic similarity and sufficient diversity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.559, + 0.683, + 0.575 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.717 + ], + "angle": 0, + "content": "In this section, we visualize and compare our sound localization results with the recent prior works on standard benchmarks, namely on VGG-SS and SoundNet-Flickr. The visualized samples in Figure 3 show that localized regions of the proposed method are more compact and accurately aligns with the sounding objects than the other methods. For instance, small size musical instrument is localized accurately compared to the recent methods in the top right column." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.885 + ], + "angle": 0, + "content": "We also compare our localization results with and without object-guided localization (OGL). As shown in Figure 4, OGL deteriorates our sound localization outputs. OGL captures objectness in a scene, thereby tending to attend to any distinctive objects regardless of whether it is the sound source or not. Therefore, OGL can be helpful when localization totally fails because of the objectness bias in the benchmarks, but it is harmful when the localization is accurate which is the case for the examples shown. This result is consistent with the quantitative result in Table 2, showing that our method with OGL performs worse." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Throughout the paper, we discuss the importance of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7784" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.095, + 0.468, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.332, + 0.47, + 0.375 + ], + "angle": 0, + "content": "Figure 5. Interactive Sound Localization of Ours and SLAVC [38]. Our model correctly follows the cross-modal interaction for given different sounds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.473, + 0.535 + ], + "angle": 0, + "content": "cross-modal semantic understanding. We demonstrate interactiveness of our method across modalities in Figure 5. Genuine sound source localization should be able to localize objects that are correlated with the sound. To visualize cross-modal interaction, we synthetically pair up the same image with different sounds of objects that are visible in a scene. The examples demonstrate that the proposed method can localize different objects depending on the contexts of sounds, while the competing method can not." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.567, + 0.196, + 0.584 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In this work, we investigate cross-modal semantic understanding that has been overlooked in sound source localization studies. We observe that higher sound source localization performance on the current benchmark does not necessarily show higher performance in cross-modal retrieval, despite its causal relevance in reality. To enforce strong understanding of audio-visual semantic matching while maintaining localization capability, we propose semantic alignment with multi-views of audio-visual pairs in a simple yet effective way. The ablation study shows that strong semantic alignment is achieved when both semantic alignment loss and enriched positive pairs are used. We extensively evaluate our method on sound source localization benchmarks including cross-dataset and open-set settings. Moreover, our analyses on cross-modal retrieval and false positive detection verify that the proposed method has strong capability in cross-modal interaction. Our study suggests that sound localization methods should be evaluated not only on localization benchmarks but also on cross-modal understanding tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.673, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.895, + 0.283 + ], + "angle": 0, + "content": "This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2023-00212845, Multimodal Speech Processing for Human-Computer Interaction). H. Pfister and J. Kim were partially supported by NIH grant R01HD104969. T.-H. Oh was partially supported by IITP grant funded by the Korea government (MSIT) (No. 2021-0-02068, Artificial Intelligence Innovation Hub; No. 2022-0-00290, Visual Intelligence for Space-Time Understanding and Generation based on Multi-layered Visual Common Sense)." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.296, + 0.597, + 0.311 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.32, + 0.892, + 0.363 + ], + "angle": 0, + "content": "[1] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. The conversation: Deep audio-visual speech enhancement. In INTERSPEECH, pages 3244–3248, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.364, + 0.894, + 0.419 + ], + "angle": 0, + "content": "[2] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. In European Conference on Computer Vision, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.421, + 0.892, + 0.462 + ], + "angle": 0, + "content": "[3] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In IEEE International Conference on Computer Vision, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.464, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[4] Relja Arandjelović and Andrew Zisserman. Objects that sound. In European Conference on Computer Vision, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.507, + 0.892, + 0.561 + ], + "angle": 0, + "content": "[5] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Soundnet: Learning sound representations from unlabeled video. Advances in Neural Information Processing Systems, NeurIPS, 2016. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.564, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[6] Mehdi Azabou, Mohammad Gheshlaghi Azar, Ran Liu, ChiHeng Lin, Erik C Johnson, Kiran Bhaskaran-Nair, Max Dabagia, Bernardo Avila-Pires, Lindsey Kitchell, Keith B Hengen, et al. Mine your own view: Self-supervised learning through across-sample prediction. arXiv preprint arXiv:2102.10106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.648, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[7] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In European Conference on Computer Vision, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.719, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[8] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[9] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, 2020. 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.32, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7785" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[11] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.177 + ], + "angle": 0, + "content": "[12] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[13] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In IEEE Conference on Computer Vision and Pattern Recognition, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.236, + 0.469, + 0.277 + ], + "angle": 0, + "content": "[14] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asia Conference on Computer Vision, pages 251-263. Springer, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.279, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[15] Soo-Whan Chung, Hong Goo Kang, and Joon Son Chung. Seeing voices and hearing voices: learning discriminative embeddings using cross-modal self-supervision. In INTER-SPEECH, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.335, + 0.468, + 0.39 + ], + "angle": 0, + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.392, + 0.468, + 0.461 + ], + "angle": 0, + "content": "[17] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In IEEE International Conference on Computer Vision, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.463, + 0.468, + 0.504 + ], + "angle": 0, + "content": "[18] Mohamed El Banani, Karan Desai, and Justin Johnson. Learning Visual Representations via Language-Guided Sampling. In CVPR, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.506, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[19] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. ACM Transactions on Graphics (Proceedings of SIGGRAPH), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.59, + 0.469, + 0.658 + ], + "angle": 0, + "content": "[20] Dennis Fedorishin, Deen Dayal Mohan, Bhavin Jawade, Srirangaraj Setlur, and Venu Govindaraju. Hear the flow: Optical flow-based self-supervised visual sound source localization. In IEEE Winter Conf. on Applications of Computer Vision (WACV), 2023. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.661, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[21] Chuang Gan, Deng Huang, Hang Zhao, Joshua B. Tenenbaum, and Antonio Torralba. Music gesture for visual sound separation. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.718, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[22] Chuang Gan, Yiwei Zhang, Jiajun Wu, Boqing Gong, and Joshua B. Tenenbaum. Look, listen, and act: Towards audiovisual embodied navigation. In IEEE International Conference on Robotics and Automation, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.774, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[23] Ruohan Gao and Kristen Grauman. Co-separating sounds of visual objects. In IEEE International Conference on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.817, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[24] Ruohan Gao and Kristen Grauman. Visualvoice: Audiovisual speech separation with cross-modal consistency. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[25] Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[26] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. In Advances in Neural Information Processing Systems, NeurIPS, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.891, + 0.232 + ], + "angle": 0, + "content": "[27] Simon Haykin and Zhe Chen. The cocktail party problem. Neural computation, 17(9):1875-1902, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[28] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[29] Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In International workshop on similarity-based pattern recognition, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[30] Di Hu, Feiping Nie, and Xuelong Li. Deep multimodal clustering for unsupervised audiovisual learning. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.454 + ], + "angle": 0, + "content": "[31] Di Hu, Rui Qian, Minyue Jiang, Xiao Tan, Shilei Wen, Errui Ding, Weiyao Lin, and Dejing Dou. Discriminative sounding objects localization via self-supervised audiovisual matching. Advances in Neural Information Processing Systems, NeurIPS, 2020. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[32] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. Advances in Neural Information Processing Systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[33] Sizhe Li, Yapeng Tian, and Chenliang Xu. Space-time memory network for sounding object localization in videos. In British Machine Vision Conference, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[34] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In International Conference on Learning Representations, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[35] Yan-Bo Lin, Hung-Yu Tseng, Hsin-Ying Lee, Yen-Yu Lin, and Ming-Hsuan Yang. Unsupervised sound localization via iterative contrastive learning. arXiv preprint arXiv:2104.00315, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[36] Jinxiang Liu, Chen Ju, Weidi Xie, and Ya Zhang. Exploiting transformation invariance and equivariance for self-supervised sound localisation. In ACM International Conference on Multimedia, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[37] Josh H McDermott. The cocktail party problem. Current Biology, 19(22):R1024-R1027, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[38] Shentong Mo and Pedro Morgado. A closer look at weakly-supervised audio-visual source localization. Advances in Neural Information Processing Systems, NeurIPS, 2022. 1, 2, 3, 4, 5, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[39] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In European Conference on Computer Vision, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[40] Pedro Morgado, Ishan Misra, and Nuno Vasconcelos. Robust audio-visual instance discrimination. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7786" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[41] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.15, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[42] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.193, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[43] Andrew Owens and Alexei A. Efros. Audio-visual scene analysis with self-supervised multisensory features. In European Conference on Computer Vision, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.236, + 0.469, + 0.319 + ], + "angle": 0, + "content": "[44] Andrew Owens, Jiajun Wu, Josh H McDermott, William T Freeman, and Antonio Torralba. Ambient sound provides supervision for visual learning. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 801-816. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.321, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[45] Takashi Oya, Shohei Iwase, Ryota Natsume, Takahiro Itazuri, Shugo Yamaguchi, and Shigeo Morishima. Do we need sound for sound source localization? In Asia Conference on Computer Vision, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.378, + 0.469, + 0.418 + ], + "angle": 0, + "content": "[46] Sooyoung Park, Arda Senocak, and Joon Son Chung. MarginNCE: Robust sound localization with a negative margin. In arXiv preprint arXiv:2211.01966, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.42, + 0.469, + 0.475 + ], + "angle": 0, + "content": "[47] Rui Qian, Di Hu, Heinrich Dinkel, Mengyue Wu, Ning Xu, and Weiyao Lin. Multiple sound sources localization from coarse to fine. In European Conference on Computer Vision, 2020. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.477, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.562, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[49] Hyeonggon Ryu, Arda Senocak, In So Kweon, and Joon Son Chung. Hindi as a second language: Improving visually grounded speech with semantically similar samples. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.633, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[50] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound source in visual scenes. In IEEE Conference on Computer Vision and Pattern Recognition, 2018. 1, 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.689, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[51] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound sources in visual scenes: Analysis and applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(5):1605-1619, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.76, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[52] Arda Senocak, Hyeonggon Ryu, Junsik Kim, and In So Kweon. Learning sound localization better from semantically similar samples. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2022. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[53] Arda Senocak, Hyeonggon Ryu, Junsik Kim, and In So Kweon. Less can be more: Sound source localization with a classification model. In IEEE Winter Conf. on Applications of Computer Vision (WACV), 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[54] Zengjie Song, Yuxi Wang, Junsong Fan, Tieniu Tan, and Zhaoxiang Zhang. Self-supervised predictive learning: A" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "negative-free method for sound source localization in visual scenes. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[55] Kim Sung-Bin, Arda Senocak, Hyunwoo Ha, Andrew Owens, and Tae-Hyun Oh. Sound to visual scene generation by audio-to-visual latent alignment. In IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[56] Yapeng Tian, Di Hu, and Chenliang Xu. Cyclic co-learning of sounding object visual grounding and sound separation. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[57] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In European Conference on Computer Vision, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.306, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[58] Efthymios Tzinis, Scott Wisdom, Aren Jansen, Shawn Hershey, Tal Remez, Daniel P. W. Ellis, and John R. Hershey. Into the wild with audioscope: Unsupervised audio-visual separation of on-screen sounds. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.376, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[59] Ho-Hsiang Wu, Prem Seetharaman, Kundan Kumar, and Juan Pablo Bello. Wav2CLIP: Learning robust audio representations from clip. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[60] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.489, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[61] Haohang Xu, Xiaopeng Zhang, Hao Li, Lingxi Xie, Wenrui Dai, Hongkai Xiong, and Qi Tian. Seed the views: Hierarchical semantic alignment for contrastive representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.56, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[62] Xudong Xu, Bo Dai, and Lin Dahua. Recursive visual sound separation using minus-plus net. In IEEE International Conference on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.603, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[63] Hanyu Xuan, Zhiliang Wu, Jian Yang, Yan Yan, and Xavier Alameda-Pineda. A proposal-based paradigm for self-supervised sound source localization in videos. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1029-1038, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.673, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[64] Hang Zhao, Chuang Gan, Wei-Chiu Ma, and Antonio Torralba. The sound of motions. In IEEE International Conference on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.716, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[65] Hang Zhou, Xudong Xu, Lin Dahua, Xiaogang Wang, and Ziwei Liu. Sep-stereo: Visually guided stereophonic audio generation by associating source separation. In European Conference on Computer Vision, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.772, + 0.892, + 0.827 + ], + "angle": 0, + "content": "[66] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In European Conference on Computer Vision, 2022. 4, 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.827 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "7787" + } + ] +] \ No newline at end of file diff --git a/2023/Sound Source Localization is All about Cross-Modal Alignment/7b57a83c-7e70-4280-a6ee-f0f550384305_origin.pdf b/2023/Sound Source Localization is All about Cross-Modal Alignment/7b57a83c-7e70-4280-a6ee-f0f550384305_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b16e51c38fbf3b9aa16c260292678c4a2ca66c91 --- /dev/null +++ b/2023/Sound Source Localization is All about Cross-Modal Alignment/7b57a83c-7e70-4280-a6ee-f0f550384305_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f66c8955ffd2ca7a28b2843d06238cab3eb6569244e556bee98c2b08cc2cf279 +size 2060551 diff --git a/2023/Sound Source Localization is All about Cross-Modal Alignment/full.md b/2023/Sound Source Localization is All about Cross-Modal Alignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..99dea60eef4c33183c4481db8b5b203c927ef013 --- /dev/null +++ b/2023/Sound Source Localization is All about Cross-Modal Alignment/full.md @@ -0,0 +1,317 @@ +# Sound Source Localization is All about Cross-Modal Alignment + +Arda Senocak $^{1*}$ Hyeonggon Ryu $^{1*}$ Junsik Kim $^{2*}$ Tae-Hyun Oh $^{3,4}$ Hanspeter Pfister $^{2}$ Joon Son Chung $^{1}$ + +$^{1}$ Korea Advanced Institute of Science and Technology $^{2}$ Harvard University + +$^{3}$ Dept. of Electrical Engineering and Grad. School of Artificial Intelligence, POSTECH + +$^{4}$ Institute for Convergence Research and Education in Advanced Technology, Yonsei University + +# Abstract + +Humans can easily perceive the direction of sound sources in a visual scene, termed sound source localization. Recent studies on learning-based sound source localization have mainly explored the problem from a localization perspective. However, prior arts and existing benchmarks do not account for a more important aspect of the problem, cross-modal semantic understanding, which is essential for genuine sound source localization. Cross-modal semantic understanding is important in understanding semantically mismatched audio-visual events, e.g., silent objects, or off-screen sounds. To account for this, we propose a cross-modal alignment task as a joint task with sound source localization to better learn the interaction between audio and visual modalities. Thereby, we achieve high localization performance with strong cross-modal semantic understanding. Our method outperforms the state-of-the-art approaches in both sound source localization and cross-modal retrieval. Our work suggests that jointly tackling both tasks is necessary to conquer genuine sound source localization. + +# 1. Introduction + +Humans can easily perceive where the sound comes from in a scene. We naturally attend to the sounding direction and associate incoming audio-visual signals to understand the event. To achieve human-level audio-visual perception, sound source localization in visual scenes has been extensively studied [50, 51, 4, 47, 8, 35, 31, 33, 53, 54, 52, 36, 39, 38, 20]. Motivated by that humans learn from natural audio-visual correspondences without explicit supervision, most of the studies have been developed on a fundamental assumption that audio and visual signals are temporally correlated. With the assumption, losses of the sound source localization task are modeled by audio-visual correspondence as a self-supervision signal and are implemented by con + +![](images/112abaefa8581f7e1d760d65eec2c80fc28af22ae6b6a062f2f72a0614e7dd21.jpg) +Figure 1. A conceptual difference between prior approaches and our alignment-based sound source localization. + +trasting audio-visual pairs, i.e., contrastive learning. + +While these approaches appear to be unsupervised methods, they strongly rely on partial supervision information; e.g., using supervisedly pretrained vision networks [50, 51, 47, 53, 54, 20] and visual objectness estimators for postprocessing [39, 38]. Without leveraging such strong initial representations, the performance is degraded. Thus, the previous methods are not purely self-supervised approaches. Even further, there are recent studies [45, 39, 38] that point out visual objectness bias in existing sound source localization benchmarks and exploit the objectness prior to improve the localization accuracy. They show that, even without interaction between visual and audio signals, a model may achieve strong accuracy in localization by only referring visual signals alone, which is not the true intention of the sound source localization task. In short, the current evaluation and setting of the sound source localization do not capture the true sound source localization performance. + +In this work, we first sort out evaluating sound source localization methods by introducing a cross-modal retrieval task as an auxiliary evaluation task. By this task, we can measure whether the learned representation has the capability to accurately interact between audio and visual modalities; i.e., more fine-grained audio-visual correspondence which is essential for genuine sound source localization. This aspect has been missed in existing sound source localization benchmarks. Indeed, our experiments show that higher sound localization performance does not guarantee higher cross-modal retrieval performance. + +Second, given this additional criterion, we revisit the importance of semantic understanding shared across audio and visual modalities in both sound source localization and cross-modal retrieval. In the previous methods [50, 51, 54, 47], the cross-modal semantic alignment is induced by instance-level cross-modal contrastive learning, i.e., cross-modal instance discrimination between visual and audio features. However, they are aided by labels or supervisedly pretrained encoder for easing challenging cross-modal feature alignment. Instead, our method learns from scratch supporting the lack of guidance by incorporating multiple positive samples into cross-modal contrastive learning. Specifically, we construct a positive set for each modality using both multi-view [10] and conceptually similar samples [17]. Thereby, we enhance feature alignment and achieve high localization performance and strong cross-modal semantic understanding. + +We evaluate our method on the VGG-SS and SoundNet-Flickr benchmarks for sound source localization and cross-modal retrieval. As aforementioned, the sound source localization task is closely related to the cross-modal retrieval task, but our experiments show that existing works have a weak performance correlation between them. This implies that we need to evaluate both tasks for evaluating the genuine sound source localization. The proposed method performs favorably against the recent state-of-the-art approaches in both tasks. + +We summarize the contributions of our work as follows: + +- We analyze that sound source localization benchmarks are not capable of evaluating cross-modal semantic understanding, thereby sound source localization methods may perform poorly in cross-modal retrieval tasks. +- We propose semantic alignment to improve cross-modal semantic understanding of sound source localization models. +- We expand semantic alignment with multi-views and conceptually similar samples which leads to state-of-the-art performance on both sound source localization and cross-modal retrieval. + +# 2. Related work + +Sound source localization. Sound source localization in visual scenes has been investigated by exploiting correspondences between audio and visual modalities. The most widely used approach for sound source localization is cross-modal attention [50, 51, 57] with contrastive loss [13, 29, 42]. Later, the attention-based method is improved by intra-frame hard sample mining [8], iterative contrastive learning with pseudo labels [35], feature regularization [36], positive mining [52], negative free learning [54] with stop-gradient operation [12], or momentum encoders [38]. + +Some sound localization approaches exploit additional semantic labels [47, 33, 53] or object prior [39, 63]. Semantic labels are used to pretrain audio and vision encoders with classification loss [33, 53] or refine audio-visual feature alignment [47]. A more explicit way to refine localization output is to use object prior. EZVSL [39] proposes post-processing to combine attention based localization output with a pretrained visual feature activation map. Similarly, Xuan et al. [63] propose to combine off-the-shelf object proposals with attention based sound localization results. However, postprocessing by object prior may generate a false positive output as it is solely based on vision without audio-visual interaction. + +In addition to the localization, there has been an attempt to localize sounding objects and recover the separated sounds simultaneously, also known as the cocktail party problem [27, 37]. The separation of sound mixture is achieved by predicting masks of spectrogram guided by visual features [19, 1, 64, 23, 62, 21, 2, 65, 24, 58, 56]. Furthermore, a number of recent papers are presented on audio-visual navigation for a given sound source [7, 22]. + +Self-supervised representation learning. In a broader categorization, sound source localization belongs to self-supervised multimodal learning. Our work is also relevant to self-supervised audio-visual representation learning, and other multimodal learning studies. + +Contrastive learning aims to learn robust representations from large-scale raw data without annotations. Recent representation learning approaches [60, 10, 28, 11] use instance discrimination by contrastive learning [13, 29, 42] as a pretext task with notable advancements in visual recognition tasks. Recently, positive mining by nearest-neighbor search are used to learn representations of images [17, 18, 61], videos [26], neural recordings [6], and text-image [34]. In this work, we expand the previous works by incorporating both multi-views and conceptually similar samples into audio-visual modalities for cross-modal feature alignment. + +A series of audio-visual representation learning studies have shown that audio and visual contents in a video are correlated, therefore a visual representation can be learned by sound prediction [44] or audio representation can be distilled from visual representation [5, 55]. Later, a variety of joint audio-visual representation learning methods are proposed with an assumption that there is a semantic [3, 30, 41, 40] or temporal [14, 43, 32, 15] correspondence between them. However, simply learning sound source localization by audio-visual correspondence with instance discrimination ignores the semantic similarity of audio-visual contents among samples, introducing false negatives or positives. In order to mitigate this issue, clustering [30], sampling [41], weighting [40], and hard mining [32] are proposed. Similarly, in this work, we go beyond instance discrimination by using multiple positive samples + +![](images/33bd5916b5d2db143c4350848ae6aaa9bf92f1f972ccfd9924bf6d94e623e3cf.jpg) +Figure 2. Our sound source localization framework. Our model constructs multiple positive pairs with augmentation and Nearest Neighbor Search (Conceptually Similar Samples). By using these newly constructed 9 pairs, our model employs spatial localization, $s_L$ , and semantic feature alignment, $s_A$ , for each pair to learn a better sound source localization ability. + +to enforce semantic understanding across modalities. + +# 3. Method + +# 3.1. Preliminaries + +Contrastive learning learns representation by containing positive and negative pairs. Given an encoded query sample $q$ and its encoded positive pair $k^{+}$ and negative pairs $k$ , the loss can be defined as: + +$$ +\mathcal {L} = - \log \frac {\exp (q \cdot k ^ {+} / \tau)}{\sum_ {i} \exp (q \cdot k _ {i} / \tau)} \tag {1} +$$ + +where $\tau$ is the temperature parameter. + +Cross-modal contrastive learning extends contrastive learning across multiple modalities. In sound source localization, audio-visual correspondence is used to define positive and negative cross-modal pairs. With an audio-visual dataset $\mathcal{D} = \{(v_i,a_i):i = 1,\dots,N\}$ and its encoded features $\mathbf{v}_i = f_v(v_i)$ and $\mathbf{a}_i = f_a(a_i)$ , cross-modal contrastive learning loss is defined as: + +$$ +\mathcal {L} _ {i} = - \log \frac {\exp (s (\mathbf {v} _ {i} , \mathbf {a} _ {i}) / \tau)}{\sum_ {j} \exp (s (\mathbf {v} _ {i} , \mathbf {a} _ {j}) / \tau)} \tag {2} +$$ + +where $s$ is a cross-modal similarity function. The cross-modal contrastive loss Eq. (2) can be extended to symmetric form [48] as used in a few previous works [39, 38]. + +# 3.2. Cross-Modal Feature Alignment + +We consider both spatial localization and semantic feature alignment for sound source localization. To this end, we use two different similarity functions $s_L$ and $s_A$ for contrastive learning (Eq. (2)), $s_L$ for localization and $s_A$ for cross-modal feature alignment. + +Recent studies rely on audio-visual spatial correspondence maps to learn sound source localization by contrasting them. Given a spatial visual feature $\mathbf{v} \in \mathbb{R}^{c \times h \times w}$ and audio feature $\mathbf{a} \in \mathbb{R}^{\mathbf{c}}$ , audio-visual similarity with a correspondence map can be calculated as follows: + +$$ +s _ {L} (\mathbf {v}, \mathbf {a}) = \sum_ {x y \in M} \frac {1}{| M |} \frac {\mathbf {v} ^ {x y} \cdot \mathbf {a}}{\| \mathbf {v} ^ {x y} \| \| \mathbf {a} \|} \tag {3} +$$ + +where $\mathbf{v}^{xy}$ is a feature vector at location $(x,y)$ , and $M$ is an optional binary mask when an annotation or pseudomask [8, 36] is available. Since we assume no supervision for sound source localization, we do not use any mask, therefore, $M = \mathbf{1}$ . + +The contrastive loss with localization similarity $s_L$ enforces location dependent alignment giving sparse but strong audio-visual correspondence which enables to perform localization. However, our empirical studies on cross-modal retrieval indicate that strong localization performance does not guarantee semantic understanding. To overcome the low semantic understanding in recent studies, we propose to add instance-level contrastive loss. Instance-level contrasting encapsulates the whole context in a scene, enforcing better audio-visual semantic alignment. However, instance-level contrasting may smooth out spatial discriminativeness learned by Eq. (3). Inspired by SimCLR [10], we adopt a projection layer to align audio-visual semantics in a projection space. The projection layer separates the latent space of localization and semantic alignment, thereby preventing the alignment loss smoothing out the spatial discriminativeness. The similarity function for cross-modal feature alignment is defined as follows: + +$$ +s _ {A} (\mathbf {v}, \mathbf {a}) = \frac {p _ {v} (\operatorname {a v g} (\mathbf {v})) \cdot p _ {a} (\mathbf {a})}{\| p _ {v} (\operatorname {a v g} (\mathbf {v})) \| \| p _ {a} \mathbf {a} \|} \tag {4} +$$ + +where $\mathrm{avg}(\cdot)$ is spatial average pooling, $p_v$ is a projection + +layer for visual features, and $p_a$ is a projection layer for audio features. + +# 3.3. Expanding with Multiple Positive Samples + +Typically, contrastive learning contrasts between one positive pair and multiple negative pairs as shown in Eq. (1). In audio-visual learning, by an audio-visual correspondence assumption, an audio-image pair from the same clip is used as a positive pair while negative pairs are sampled from different clips. However, single-instance discrimination may not be sufficient to achieve strong cross-modal alignment. In this section, we expand contrastive learning beyond single instance discrimination by positive set construction and pairing them. To construct a positive set, we incorporate both hand-crafted positive and conceptual positive samples for each modality. Later, we adjust the contrastive learning to incorporate multiple positive pairs to enforce cross-modal alignment. + +Obtaining hand-crafted positive samples. Using randomly augmented samples as positive multi-view pairs are widely adopted in self-supervised representation learning, i.e., instance discrimination. Similarly, we extend a single anchor audio-image pair to multiple positive pairs by applying simple augmentations on image and audio samples separately. While we utilize common image transformations on images, we apply temporal shifting to audiOs. It is worth noting that sound source localization task learns from the underlying semantic consistency rather than subtle time differences as in videos. Thus, a slight shift in the audio may not alter contextual information significantly. As a result of hand-crafted multi-view positive pair generation, we obtain additional $\mathbf{v}^{aug}$ and $\mathbf{a}^{aug}$ samples. + +Obtaining conceptual positive samples. Apart from manually created augmented views, we additionally expand our positive set with conceptually similar samples. The sampling strategy with nearest neighbor search can be performed in a various way, such as on-the-fly sampling [17, 49, 61, 34], sampling by pretrained encoders [52], or guided sampling [26, 18] using another modality. For selecting our conceptually similar samples, we utilize pretrained encoders. Note that pretrained encoders trained either with supervised or self-supervised learning are effective in positive sample mining as shown in the experiment section. By employing readily available image and audio encoders, we use the $k$ -nearest neighborhood search to sample semantically similar samples in both modalities. In particular, given a pair of image and audio, we compute cosine similarity with all other samples and choose the top- $k$ most similar samples among the training set for each modality. From a set of $k$ samples, we randomly select one sample to obtain conceptually similar samples for each modality, $\mathbf{v}^{conc}$ and $\mathbf{a}^{conc}$ . By utilizing the conceptually similar samples as + +positive samples, our model expands semantic understanding. + +Pair Construction. Once we obtain the conceptual and hand-crafted positive samples for each modality, we proceed to create 9 distinct audio-visual pairs by pairing $\mathbf{V} = \{\mathbf{v},\mathbf{v}^{aug},\mathbf{v}^{conc}\}$ and $\mathbf{A} = \{\mathbf{a},\mathbf{a}^{aug},\mathbf{a}^{conc}\}$ . This is done to ensure semantic alignment and consistency between them through contrastive learning. The negative pairs are randomly paired from the remaining samples in a training set. It is worth noting that some of these pairs are a combination of hand-crafted and conceptually similar samples, which further enhances the feature alignment of our model during training. + +# 3.4. Training + +Our loss formulation incorporates both localization and instance-level similarity functions with multiple positive pairs constructed by augmentation and conceptually similar sample search. The final loss term is defined as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {i} = - \sum_ {\mathbf {v} _ {i} \in \mathbf {V}} \sum_ {\mathbf {a} _ {i} \in \mathbf {A}} \left[ \log \frac {\exp \left(s _ {L} \left(\mathbf {v} _ {i} , \mathbf {a} _ {i}\right) / \tau\right)}{\sum_ {j} \exp \left(s _ {L} \left(\mathbf {v} _ {i} , \mathbf {a} _ {j}\right) / \tau\right)} \right. \tag {5} \\ \left. + \log \frac {\exp \left(s _ {A} \left(\mathbf {v} _ {i} , \mathbf {a} _ {i}\right) / \tau\right)}{\sum_ {j} \exp \left(s _ {A} \left(\mathbf {v} _ {i} , \mathbf {a} _ {j}\right) / \tau\right)} \right] \\ \end{array} +$$ + +where $\mathbf{V}$ and $\mathbf{A}$ indicate positive sample sets. + +# 4. Experiments + +Our proposed method for sound source localization is validated through experiments conducted on VG-GSound [9] and SoundNet-Flickr [5]. First, we conduct a quantitative analysis to evaluate the accuracy of the localization, cross-modal retrieval, and the impact of various components of our model. Then, we visualize our sound source localization results across different categories of sounds. + +# 4.1. Experiment Setup + +Datasets. Our method is trained using the VGGSound [9] and SoundNet-Flickr-144K [50, 51]. VGGSound is an audio-visual dataset containing around $\sim 200\mathrm{K}$ videos. SoundNet-Flickr-144K set is the subset of SoundNetFlickr [5]. After training, we test the sound localization performance with VGG-SS [8] and SoundNet-Flickr-Test [50] datasets for the main experiments. These evaluation sets have bounding box annotations of sound sources for $\sim 5\mathrm{K}$ and 250 samples, respectively. Moreover, we employ the AVSBench [66] and Extended VGGSound/SoundNetFlickr [38] datasets for additional evaluations. AVSBench dataset provides binary segmentation maps that show the + +
MethodPre. VisionVGG-SSFlickr-SoundNet
cIoU ↑AUC ↑cIoU ↑AUC ↑
Attention [50]CVPR18✓18.5030.2066.0055.80
CoarseToFine [47]ECCV20✓29.1034.80--
LCBM [53]WACV22✓32.2036.60--
LVS [8]†CVPR21✗30.3036.4072.4057.80
LVS [8]CVPR21✗34.4038.2071.9058.20
HardPos [52]ICASSP22✗34.6038.0076.8059.20
SSPL (w/o PCM) [54]CVPR22✓27.0034.8073.9060.20
SSPL (w/ PCM) [54]CVPR22✓33.9038.0076.7060.50
EZ-VSL (w/o OGL) [39]ECCV22✓35.9638.2078.3161.74
SSL-TIE [36]ACMM22✗38.6339.6579.5061.20
SLAVC (w/o OGL) [38]NeurIPS22✓37.7939.4083.60-
Ours
↓ NN Search w/ Supervised Pre. Encoders✗39.9440.0279.6063.44
↓ NN Search w/ Self-Supervised Pre. Encoders✗39.2039.7079.2063.00
with OGL:
EZ-VSL (w/ OGL) [39]ECCV22✓38.8539.5483.9463.60
SLAVC (w/ OGL) [38]NeurIPS22✓39.80-86.00-
Ours (w/ OGL)
↓ NN Search w/ Supervised Pre. Encoders✗42.6441.4882.4064.60
↓ NN Search w/ Self-Supervised Pre. Encoders✗42.4741.4282.8064.48
with Optical Flow:
HearTheFlow [20]WACV23✓39.4040.0084.8064.00
+ +Table 1. Quantitative results on the VGG-SS and SoundNet-Flickr test sets. All models are trained with 144K samples from VGG-Sound and tested on VGG-SS and SoundNet-Flickr. $\dagger$ is the result of the model released on the official project page. SLAVC [38] does not provide AUC scores. + +audio-visually correspondent pixels for roughly 5k five-second videos belonging to 23 categories. Lastly, the Extended VGGSound /SoundNet-Flickr dataset, proposed by [38], is used to understand non-visible sound sources. + +Implementation details. We use two ResNet18 models for both audio and vision encoding. Unlike prior approaches, we do not fine-tune (or use a pretrained) a visual encoder from ImageNet pretrained weights. Instead, we train both the audio and vision encoders from scratch. We preprocess images and audios following the previous works [8, 52]. To create multiple pairs, we utilize both NN search and generic augmentation approaches. For NN search, we experiment on two different setups to retrieve k conceptually similar samples: (1) For supervisedly pretrained encoder experiments, We employ ResNet and VGGSound models pretrained on ImageNet and VGGSound respectively, (2) For self-supervisedly pretrained encoder experiments, we utilize the CLIP [48] Vision Encoder and Wav2CLIP [59] Audio Encoder. We use $k = 1000$ for the experiments. To perform image augmentations, we follow the augmentations used in SimCLR [10]. For audios, we randomly select timewindow shifts in a time axis. The model is trained for 50 epochs with Adam Optimizer and a learning rate of 0.0001. $\tau$ is set to 0.07 in contrastive learning. + +# 4.2. Quantitative Results + +Comparison with strong baselines. In this section, we conduct a comparative analysis of our sound source localization method against existing approaches. We carry out our evaluations in two settings, following previous approaches. Firstly, we train our model on VGGSound-144K + +
MethodPre. VisioncIoU ↑AUC ↑
Attention[50]CVPR18✓66.0055.80
DMC[30]CVPR19✓67.1056.80
LVS [8]†CVPR21X67.2056.20
LVS [8]CVPR21X69.9057.30
HardPos [52]ICASSP22X75.2059.70
SSPL (w/o PCM) [54]CVPR22✓69.9058.00
SSPL (w/ PCM) [54]CVPR22✓75.9061.00
EZ-VSL (w/o OGL) [39]ECCV22✓71.8958.81
SSL-TIE [36]ACMM22X81.5061.10
SLAVC (w/o OGL) [38]NeurIPS22✓--
Ours
↓ NN Search w/ Supervised Pre. EncodersX85.2062.20
↓ NN Search w/ Self-Supervised Pre. EncodersX84.8062.66
with OGL:
EZ-VSL (w/ OGL) [39]ECCV22✓83.1363.06
SLAVC (w/ OGL) [38]NeurIPS22✓--
Ours (w/ OGL)
↓ NN Search w/ Supervised Pre. EncodersX84.0064.16
↓ NN Search w/ Self-Supervised Pre. EncodersX84.4064.38
with Optical Flow:
HearTheFlow [20]WACV23✓86.5063.90
+ +Table 2. Quantitative results on the SoundNet-Flickr test set. All models are trained and tested on the SoundNet-Flickr 144K dataset. $\dagger$ is the result of the model from the official project page. SLAVC [38] does not provide results with SoundNet-Flickr 144K. + +and evaluate it on VGG-SS and SoundNet-Flickr test sets. Secondly, we train our model on SoundNet-Flickr-144K and evaluate it on the SoundNet-Flickr test set. It is important to note that all the compared models are trained using the same amount of data. AVEL [57], AVObject [2], and LCBM [53] models rely on video input, and as such, they cannot be evaluated on the SoundNet-Flickr dataset, which contains static image and audio pairs. We present our results in Table 1 and Table 2. + +Our proposed model achieves higher performance compared to prior approaches on both test sets. Specifically, it yields a $+2.15\%$ cIoU and $+0.6\%$ AUC improvement on VGGSS, as well as a $+3.7\%$ cIoU improvement on SoundNet-Flickr compared to the state-of-the-art methods that uses pretrained vision encoder. It is worth highlighting that unlike the majority of previous works, our proposed model does not utilize a vision encoder pretrained on ImageNet in a sound source localization backbone. This is because, as discussed in Mo et al. [38], using supervisedly pretrained vision encoders makes the sound source localization problem a weakly supervised problem. However, it is worth noting that even without using a pretrained vision encoder, our method achieves state-of-the-art performance on both experiments that are presented in Table 1 and Table 2. We demonstrate the performance of our model with the pretrained models learned through supervised learning (NN Search w/ Supervised Pre. Encoders) and with models that are pretrained through self-supervised learning (NN Search w/ Self-Supervised Pre. Encoders) in NN Search module. As the results indicate, using self-supervised pre + +
Test ClassMethodPre. VisioncIoU ↑AUC ↑
Heard 110LVS [8]CVPR21X28.9036.20
EZ-VSL(w/o OGL) [39]ECCV22✓31.8636.19
SLAVC(w/o OGL) [38]NeurlPS22✓35.84-
OursX38.3139.05
with OGL:
EZ-VSL(w/o OGL) [39]ECCV22✓37.2538.97
SLAVC(w/o OGL) [38]NeurlPS22✓38.22-
Ours(w/ OGL)X41.8540.93
with Optical Flow:
HearTheFlow [20]WACV23✓37.3038.60
Unheard 110LVS [8]CVPR21X26.3034.70
EZ-VSL(w/o OGL) [39]ECCV22✓32.6636.72
SLAVC(w/o OGL) [38]NeurlPS22✓36.50-
OursX39.1139.80
with OGL:
EZ-VSL(w/o OGL) [39]ECCV22✓39.5739.60
SLAVC(w/o OGL) [38]NeurlPS22✓38.87-
Ours(w/ OGL)✓42.9441.54
with Optical Flow:
HearTheFlow [20]WACV23✓39.3040.00
+ +Table 3. Comparison results on open-set audio-visual localization experiments trained and tested on the splits of [8, 39, 20]. + +
Test ClassMethodPre. VisioncIoU ↑AUC ↑
Heard 110SSSL-TIE [36]ACMM22X39.0040.30
OursX41.2041.00
Unheard 110SSSL-TIE [36]ACMM22X36.5038.60
OursX36.9038.59
+ +trained encoders in NN Search performs on par with the supervised pretrained encoders in NN Search. This shows that our model does not depend on supervised pretrained encoders for the NN search module and can utilize any type of pretrained encoder feature for nearest neighbor search. Note that these pretrained encoders are not used in the backbone networks of the sound source localization module but only in the NN Search Module, as illustrated in Figure 2. + +We also discuss the methods employed by previous studies, such as SSPL [54] which utilizes a sub-module called PCM to reduce the impact of background noise, HTF [20] which utilizes Optical Flow, and EZ-VSL [39] which refines its initial audio-visual localization outcomes through object guidance obtained from an ImageNet pretrained visual encoder. Our model, on the other hand, and any of its variations do not require any task-specific modules or operations to achieve the state-of-the-art (SOTA) results. This suggests that using additional semantic and multi-view correspondence, as well as feature alignment, provides more varied and robust supervision for better aligned audio and visual features, as opposed to using task-specific approaches. + +The quantitative results presented in Table 1 and Table 2 also showcase the performance of previous methods that utilize object guidance to evaluate their final sound source localizations. Our model outperforms all previous methods that employ object guidance on the VGG-SS test set and achieves comparable results on the SoundNet-Flickr test set, even though our model does not use object guided refinement (OGL). Additionally, we acknowledge that the addition of OGL to our audio-visual localization results in improvement on the VGGSS test set, while degrading perfor + +Table 4. Comparison results on open set audio-visual localization experiments trained and tested on the splits of [36]. + +
Test SetMethodPre. VisionmIoU ↑F-Score ↑
S4LVS (w/o OGL) [8]CVPR21X26.933.6
EZ-VSL (w/o OGL) [39]ECCV22✓27.634.2
SLAVC (w/o OGL) [38]NeuRTPS22✓28.134.6
Ours (w/o OGL)
↓ NN Search w/ Supervised Pre. EncodersX29.635.9
↓ NN Search w/ Self-Supervised Pre. EncodersX29.335.6
+ +Table 5. Quantitative results on AVS Bench S4 dataset. All models are trained on the VGGSound 144K dataset. + +
ModelPre. VisionA → II → A
R@1R@5R@10R@1R@5R@10
LVS [8]CVPR21X3.8712.3520.734.9014.2921.37
EZ-VSL [39]ECCV22✓5.0115.7324.8114.233.5145.18
SSL-TIE [36]MM22X10.2930.6843.7612.7629.5839.72
SLAVC [38]NeurIPS22✓4.7713.0819.106.1221.1632.12
Ours
↓ NN Search w/ Supervised Pre. EncodersX16.4736.9949.0020.0942.3853.66
↓ NN Search w/ Self-Supervised Pre. EncodersX14.3137.8149.1718.0038.3949.02
+ +Table 6. Summary of retrieval recall scores for all models. All of the models are trained on VGGSound 144K data and retrieval is performed on entire VGG-SS dataset, containing $\sim 5\mathrm{K}$ samples. + +mance on the SoundNet-Flickr test set. In contrast, prior methods see modest improvements when utilizing OGL. This can be explained by the fact that our model is already accurately localizing the sounding objects, and object guidance can interfere with localization results by introducing visual regions that are not sounding (refer to Section 4.4 for visual results). Unlike prior methods, we do not use OGL in our architecture for the remainder of this paper, unless it is being directly compared with OGL-based methods. + +Finally, in comparison to HearTheFlow, which utilizes an additional Optical Flow modality, our method outperforms it on the VGGSS test set, and achieves slightly lower performance on the SoundNet-Flickr test set without utilizing any additional modalities, but instead relying on better audio-visual correspondence and alignment. + +Open Set Audio-Visual Localization. The study by Chen et al. [8] evaluates the generalization ability of sound source localization methods in an open set scenario. This involves testing the models on categories that are both present in the training data (heard) and categories that are not present (unheard). To accomplish this, 110 randomly selected categories from the VGGSound dataset are used for training, while another disjoint set of 110 categories are reserved for evaluation to ensure the model have never seen or heard them before. It should be noted that not all previous works use the same train/test splits. While some works, including [8, 39, 38, 46], share the same splits,[36] uses a different split. Therefore, to ensure a fair comparison, we conduct experiments on both splits, evaluating on test samples from both heard and unheard categories. The results are shown in Table 3 and Table 4. Our model outperforms existing approaches on both categories, regardless of the train/test splits used. Specifically, in Table 3, our model (w/o OGL) even surpasses the other models that use OGL. Previous approaches draw different conclusions from + +
MethodPre. VisionExtended Flickr-SoundNetExtended VGG-SS
AP ↑max-F1 ↑LocAcc ↑AP ↑max-F1 ↑LocAcc ↑
CoarseToFine [47]ECCV20✓0.0038.2047.200.0019.8021.93
LVS [8]CVPR21✗9.8017.9019.605.159.9010.43
Attention10k [50]CVPR18✓15.9824.0034.166.7013.1014.04
DMC [30]CVPR19✓25.5641.8052.8011.5320.3022.63
DSOL [31]NeuIPS20✓38.3249.4072.9116.8425.6026.87
OGL [39]ECCV22-40.2055.7077.2018.7330.9036.58
EZ-VSL (w/o OGL) [39]ECCV22✓46.3054.6066.4024.5530.9031.58
SLAVC (w/o OGL) [38]NeuIPS22✓51.6359.1083.6032.9540.0037.79
Ours
↓ NN Search w/ Supervised Pre. Encoders✗64.4366.9079.6034.7340.7039.94
↓ NN Search w/ Self-Supervised Pre. Encoders✗62.6766.1079.2033.0940.0039.20
+ +these open set experiments. While some conclude that their models have strong generalization ability because their performance in unheard categories is higher than heard categories [39, 38, 46], the other works that cannot achieve the same trend discuss that this is expected since their models are dealing with unseen categories [36]. However, our results show that these conclusions are highly dependent on the chosen train/test splits. Our model performs better than existing works in both splits, but there is no uniform trend in between two splits. While our method performs better on unheard categories in the splits of [8, 39, 38, 46], it performs worse on unheard categories in the split of [36]. Therefore, we conclude that the observed trends are highly dependent on the randomly selected train/test splits. + +AVSBench [66]. To demonstrate the precise sound localization ability of our model, we conduct experiments on the AVSBench S4 dataset. The dataset's objective is to detect audio-visual correspondence and correlation at the pixel level. To make a fair comparison, we use some of the self-supervised sound source localization methods mentioned earlier. All models are trained on VGGSound-144K and directly assessed on the AVSBench S4 dataset without any further fine-tuning (zero-shot setting). Our results, which are presented in Table 5, indicate that our method achieves the highest performance, as in the previous experiments. + +Retrieval. We evaluate sound localization models on the VGG-SS dataset for cross-modal retrieval. As shown in Table 6, our method clearly outperforms other state-of-the-art methods. One interesting observation is that EZ-VSL [39] notably performs better than SLAVC [38] on cross-modal retrieval, while SLAVC performs better on sound source localization in Table 1. This shows that with the current benchmark evaluations, better sound localization performance does not guarantee better audio-visual semantic understanding, thereby we need to additionally evaluate sound source localization methods on cross-modal understanding tasks. Another observation is that the performance gap between our method and the strongest competitor SSL-TIE [36] is notably larger on cross-modal retrieval than sound source localization. This is due to the strong cross-modal feature alignment of our method that is over + +Table 7. Quantitative results on the Extended VGG-SS and Extended SoundNet-Flickr sets. All models are trained with 144K samples from VGG-Sound. The results of the prior approaches are obtained from [38]. + +
SemanticMulti-ViewFeature AlignmentcIoU ↑AUC ↑
(A)✓✓✓39.9440.02
(B)✓✓✗39.1039.44
(C)✓✗✓38.7539.34
(D)✓✗✗38.2438.90
(E)✗✓✓38.3039.38
(F)✗✓✗37.7239.19
(G)✗✗✓34.9337.94
(H)✗✗✗34.2237.67
+ +Table 8. Ablation studies on our proposed method to see the impact of each main component. + +looked in the sound source localization benchmarks. + +Extended Flickr and VGG-SS datasets. The prior study [38] points out that the current sound source localization benchmarks overlook false positive detection. It is because the evaluation samples always contain at least a sounding object in a scene; thus cannot capture false positive outputs, e.g., silent objects or off-screen sounds. To analyze false positive detection, Mo and Morgado [38] extended the benchmarks with non-audible, non-visible, and mismatched audio-visual samples. The expectation is that a sound source localization model should not localize any objects when audio-visual semantics do not match. + +The experiment with the extended datasets in Table 7 shows that our method performs favorably against state-of-the-art competitors. Our method performs better than the competing methods in false positive detection measured by AP and max-F1, while SLAVC [38] achieves better localization performance on Extended Flickr-SoundNet. As both false positive detection and cross-modal retrieval require cross-modal interaction, our method shows strong performance on both tasks. + +# 4.3. Ablation Results + +We conduct a series of experiments in order to verify our design choices and make further analysis. To save computational time and resources, we primarily perform ablation studies by training our model on VGGSound-144K with NN Search w/ Supervised Pre. Encoders setup and evaluating it on VGG-SS. Results are in Table 8. + +Impact of Semantic and Multi-View Invariance. In order to understand the impact of each type of invariance (consistency), we analyze the performance of our model with different type of invariance methodologies in Table 8. As the results of (C vs. E) and (D vs. F) reveal, using semantically similar samples (semantic invariance) produces better performance $(+0.45\%)$ and $+0.5\%$ on cIoU respectively compared to augmented multi-view invariance. Moreover, as the results of (A vs. C) and (A vs. E) depict, the combination of these two different types of invariance complement each other and and further enhances the model's performance. Using pair combination of these two different + +![](images/2b7baad645a34fd19154d0134d9422c38aedcd09c7a9caf896a17d43b6ad4c5f.jpg) +Figure 3. Sound Localization Results on VGG-SS (top) and SoundNet-Flickr (bottom). + +![](images/19523afffaeccb4433899c6fbcace7e5da95cdbe3f1e2cb22fe36193fc607b48.jpg) + +
k in k-NN10301005001000
cIoU ↑38.8038.8239.4639.9039.94
AUC ↑39.5139.6739.9340.0040.02
+ +Table 9. Varying k in conceptually similar sample selection. + +types of consistency elements provides additional supervi sions, invariance and alignments, leading to a more robust representation space and improve sound localization performance. + +Impact of Feature Alignment. We perform controlled experiments to verify the effect of the feature alignment strategy, and the results are presented in Table 8. Comparing the performance of the proposed model with and without feature alignment, (A vs. B), highlights the importance of this strategy to boost the performance. Further, examining the results of experiments (C vs. D) and (E vs. F) reveals that feature alignment provides additional gains irrespective of the consistency types. These findings indicate that global feature-based alignment helps the optimization of audio-visual correspondence. + +Impact of $k$ in conceptually similar sample selection. Selecting an appropriate $k$ value for sampling nearest neighbors is crucial. If this value is set too high, it may result in noisy samples that could disrupt the learning phase. Conversely, if the value is set too low, only very similar samples to the anchor will be provided and it limits semantic invariance. Nevertheless, when compared to Table 8 (E), we observe performance gain throughout the range of $k$ used for the ablation study. Table 9 shows an ablative evaluation of the effect of $k$ value used to select neighborhood samples. The results indicate that an optimal choice is $k = 1000$ . This + +![](images/cfc727a4fc22eeae7874a4d683920c4574791ebd9ce59d0695b65c37667021ce.jpg) +Figure 4. OGL degrades our sound localization results on SoundNet-Flickr. + +choice of $k$ can be explained by the fact that it provides a balance between semantic similarity and sufficient diversity. + +# 4.4. Qualitative Results + +In this section, we visualize and compare our sound localization results with the recent prior works on standard benchmarks, namely on VGG-SS and SoundNet-Flickr. The visualized samples in Figure 3 show that localized regions of the proposed method are more compact and accurately aligns with the sounding objects than the other methods. For instance, small size musical instrument is localized accurately compared to the recent methods in the top right column. + +We also compare our localization results with and without object-guided localization (OGL). As shown in Figure 4, OGL deteriorates our sound localization outputs. OGL captures objectness in a scene, thereby tending to attend to any distinctive objects regardless of whether it is the sound source or not. Therefore, OGL can be helpful when localization totally fails because of the objectness bias in the benchmarks, but it is harmful when the localization is accurate which is the case for the examples shown. This result is consistent with the quantitative result in Table 2, showing that our method with OGL performs worse. + +Throughout the paper, we discuss the importance of + +![](images/1083664842bda8e3a257421b52ffda1b5208ee05bd0b99922367f75c9f5d836e.jpg) +Figure 5. Interactive Sound Localization of Ours and SLAVC [38]. Our model correctly follows the cross-modal interaction for given different sounds. + +cross-modal semantic understanding. We demonstrate interactiveness of our method across modalities in Figure 5. Genuine sound source localization should be able to localize objects that are correlated with the sound. To visualize cross-modal interaction, we synthetically pair up the same image with different sounds of objects that are visible in a scene. The examples demonstrate that the proposed method can localize different objects depending on the contexts of sounds, while the competing method can not. + +# 5. Conclusion + +In this work, we investigate cross-modal semantic understanding that has been overlooked in sound source localization studies. We observe that higher sound source localization performance on the current benchmark does not necessarily show higher performance in cross-modal retrieval, despite its causal relevance in reality. To enforce strong understanding of audio-visual semantic matching while maintaining localization capability, we propose semantic alignment with multi-views of audio-visual pairs in a simple yet effective way. The ablation study shows that strong semantic alignment is achieved when both semantic alignment loss and enriched positive pairs are used. We extensively evaluate our method on sound source localization benchmarks including cross-dataset and open-set settings. Moreover, our analyses on cross-modal retrieval and false positive detection verify that the proposed method has strong capability in cross-modal interaction. Our study suggests that sound localization methods should be evaluated not only on localization benchmarks but also on cross-modal understanding tasks. + +# 6. Acknowledgment + +This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2023-00212845, Multimodal Speech Processing for Human-Computer Interaction). H. Pfister and J. Kim were partially supported by NIH grant R01HD104969. T.-H. Oh was partially supported by IITP grant funded by the Korea government (MSIT) (No. 2021-0-02068, Artificial Intelligence Innovation Hub; No. 2022-0-00290, Visual Intelligence for Space-Time Understanding and Generation based on Multi-layered Visual Common Sense). + +# References + +[1] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. The conversation: Deep audio-visual speech enhancement. In INTERSPEECH, pages 3244–3248, 2018. 2 +[2] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. In European Conference on Computer Vision, 2020. 2, 5 +[3] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In IEEE International Conference on Computer Vision, 2017. 2 +[4] Relja Arandjelović and Andrew Zisserman. Objects that sound. In European Conference on Computer Vision, 2018. 1 +[5] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Soundnet: Learning sound representations from unlabeled video. Advances in Neural Information Processing Systems, NeurIPS, 2016. 2, 4 +[6] Mehdi Azabou, Mohammad Gheshlaghi Azar, Ran Liu, ChiHeng Lin, Erik C Johnson, Kiran Bhaskaran-Nair, Max Dabagia, Bernardo Avila-Pires, Lindsey Kitchell, Keith B Hengen, et al. Mine your own view: Self-supervised learning through across-sample prediction. arXiv preprint arXiv:2102.10106, 2021. 2 +[7] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In European Conference on Computer Vision, 2020. 2 +[8] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 3, 4, 5, 6, 7 +[9] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2020. 4 +[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, 2020. 2, 3, 5 + +[11] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2 +[12] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2 +[13] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In IEEE Conference on Computer Vision and Pattern Recognition, 2005. 2 +[14] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asia Conference on Computer Vision, pages 251-263. Springer, 2017. 2 +[15] Soo-Whan Chung, Hong Goo Kang, and Joon Son Chung. Seeing voices and hearing voices: learning discriminative embeddings using cross-modal self-supervision. In INTER-SPEECH, 2020. 2 +[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition, 2009. 2 +[17] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In IEEE International Conference on Computer Vision, 2021. 2, 4 +[18] Mohamed El Banani, Karan Desai, and Justin Johnson. Learning Visual Representations via Language-Guided Sampling. In CVPR, 2022. 2, 4 +[19] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. ACM Transactions on Graphics (Proceedings of SIGGRAPH), 2018. 2 +[20] Dennis Fedorishin, Deen Dayal Mohan, Bhavin Jawade, Srirangaraj Setlur, and Venu Govindaraju. Hear the flow: Optical flow-based self-supervised visual sound source localization. In IEEE Winter Conf. on Applications of Computer Vision (WACV), 2023. 1, 5, 6 +[21] Chuang Gan, Deng Huang, Hang Zhao, Joshua B. Tenenbaum, and Antonio Torralba. Music gesture for visual sound separation. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2 +[22] Chuang Gan, Yiwei Zhang, Jiajun Wu, Boqing Gong, and Joshua B. Tenenbaum. Look, listen, and act: Towards audiovisual embodied navigation. In IEEE International Conference on Robotics and Automation, 2020. 2 +[23] Ruohan Gao and Kristen Grauman. Co-separating sounds of visual objects. In IEEE International Conference on Computer Vision, 2019. 2 +[24] Ruohan Gao and Kristen Grauman. Visualvoice: Audiovisual speech separation with cross-modal consistency. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2 +[25] Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal, + +and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2017. 2 +[26] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. In Advances in Neural Information Processing Systems, NeurIPS, 2020. 2, 4 +[27] Simon Haykin and Zhe Chen. The cocktail party problem. Neural computation, 17(9):1875-1902, 2005. 2 +[28] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2 +[29] Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In International workshop on similarity-based pattern recognition, 2015. 2 +[30] Di Hu, Feiping Nie, and Xuelong Li. Deep multimodal clustering for unsupervised audiovisual learning. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2, 5, 7 +[31] Di Hu, Rui Qian, Minyue Jiang, Xiao Tan, Shilei Wen, Errui Ding, Weiyao Lin, and Dejing Dou. Discriminative sounding objects localization via self-supervised audiovisual matching. Advances in Neural Information Processing Systems, NeurIPS, 2020. 1, 7 +[32] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. Advances in Neural Information Processing Systems, 31, 2018. 2 +[33] Sizhe Li, Yapeng Tian, and Chenliang Xu. Space-time memory network for sounding object localization in videos. In British Machine Vision Conference, 2021. 1, 2 +[34] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In International Conference on Learning Representations, 2022. 2, 4 +[35] Yan-Bo Lin, Hung-Yu Tseng, Hsin-Ying Lee, Yen-Yu Lin, and Ming-Hsuan Yang. Unsupervised sound localization via iterative contrastive learning. arXiv preprint arXiv:2104.00315, 2021. 1, 2 +[36] Jinxiang Liu, Chen Ju, Weidi Xie, and Ya Zhang. Exploiting transformation invariance and equivariance for self-supervised sound localisation. In ACM International Conference on Multimedia, 2022. 1, 2, 3, 5, 6, 7 +[37] Josh H McDermott. The cocktail party problem. Current Biology, 19(22):R1024-R1027, 2009. 2 +[38] Shentong Mo and Pedro Morgado. A closer look at weakly-supervised audio-visual source localization. Advances in Neural Information Processing Systems, NeurIPS, 2022. 1, 2, 3, 4, 5, 6, 7, 9 +[39] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In European Conference on Computer Vision, 2022. 1, 2, 3, 5, 6, 7 +[40] Pedro Morgado, Ishan Misra, and Nuno Vasconcelos. Robust audio-visual instance discrimination. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2 + +[41] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2 +[42] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2 +[43] Andrew Owens and Alexei A. Efros. Audio-visual scene analysis with self-supervised multisensory features. In European Conference on Computer Vision, 2018. 2 +[44] Andrew Owens, Jiajun Wu, Josh H McDermott, William T Freeman, and Antonio Torralba. Ambient sound provides supervision for visual learning. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 801-816. Springer, 2016. 2 +[45] Takashi Oya, Shohei Iwase, Ryota Natsume, Takahiro Itazuri, Shugo Yamaguchi, and Shigeo Morishima. Do we need sound for sound source localization? In Asia Conference on Computer Vision, 2020. 1 +[46] Sooyoung Park, Arda Senocak, and Joon Son Chung. MarginNCE: Robust sound localization with a negative margin. In arXiv preprint arXiv:2211.01966, 2022. 6, 7 +[47] Rui Qian, Di Hu, Heinrich Dinkel, Mengyue Wu, Ning Xu, and Weiyao Lin. Multiple sound sources localization from coarse to fine. In European Conference on Computer Vision, 2020. 1, 2, 5, 7 +[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 3, 5 +[49] Hyeonggon Ryu, Arda Senocak, In So Kweon, and Joon Son Chung. Hindi as a second language: Improving visually grounded speech with semantically similar samples. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2023. 4 +[50] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound source in visual scenes. In IEEE Conference on Computer Vision and Pattern Recognition, 2018. 1, 2, 4, 5, 7 +[51] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound sources in visual scenes: Analysis and applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(5):1605-1619, 2021. 1, 2, 4 +[52] Arda Senocak, Hyeonggon Ryu, Junsik Kim, and In So Kweon. Learning sound localization better from semantically similar samples. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2022. 1, 2, 4, 5 +[53] Arda Senocak, Hyeonggon Ryu, Junsik Kim, and In So Kweon. Less can be more: Sound source localization with a classification model. In IEEE Winter Conf. on Applications of Computer Vision (WACV), 2022. 1, 2, 5 +[54] Zengjie Song, Yuxi Wang, Junsong Fan, Tieniu Tan, and Zhaoxiang Zhang. Self-supervised predictive learning: A + +negative-free method for sound source localization in visual scenes. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 5, 6 +[55] Kim Sung-Bin, Arda Senocak, Hyunwoo Ha, Andrew Owens, and Tae-Hyun Oh. Sound to visual scene generation by audio-to-visual latent alignment. In IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2 +[56] Yapeng Tian, Di Hu, and Chenliang Xu. Cyclic co-learning of sounding object visual grounding and sound separation. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2 +[57] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In European Conference on Computer Vision, 2018. 2, 5 +[58] Efthymios Tzinis, Scott Wisdom, Aren Jansen, Shawn Hershey, Tal Remez, Daniel P. W. Ellis, and John R. Hershey. Into the wild with audioscope: Unsupervised audio-visual separation of on-screen sounds. In International Conference on Learning Representations, 2021. 2 +[59] Ho-Hsiang Wu, Prem Seetharaman, Kundan Kumar, and Juan Pablo Bello. Wav2CLIP: Learning robust audio representations from clip. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2022. 5 +[60] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2 +[61] Haohang Xu, Xiaopeng Zhang, Hao Li, Lingxi Xie, Wenrui Dai, Hongkai Xiong, and Qi Tian. Seed the views: Hierarchical semantic alignment for contrastive representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 4 +[62] Xudong Xu, Bo Dai, and Lin Dahua. Recursive visual sound separation using minus-plus net. In IEEE International Conference on Computer Vision, 2019. 2 +[63] Hanyu Xuan, Zhiliang Wu, Jian Yang, Yan Yan, and Xavier Alameda-Pineda. A proposal-based paradigm for self-supervised sound source localization in videos. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1029-1038, 2022. 2 +[64] Hang Zhao, Chuang Gan, Wei-Chiu Ma, and Antonio Torralba. The sound of motions. In IEEE International Conference on Computer Vision, 2019. 2 +[65] Hang Zhou, Xudong Xu, Lin Dahua, Xiaogang Wang, and Ziwei Liu. Sep-stereo: Visually guided stereophonic audio generation by associating source separation. In European Conference on Computer Vision, 2020. 2 +[66] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In European Conference on Computer Vision, 2022. 4, 7 \ No newline at end of file diff --git a/2023/Sound Source Localization is All about Cross-Modal Alignment/images.zip b/2023/Sound Source Localization is All about Cross-Modal Alignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f8af6ea6a0db610cd0f6fc57b9c1a8f97dc16b4f --- /dev/null +++ b/2023/Sound Source Localization is All about Cross-Modal Alignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd89a17cc6b4073151dee76ad5ab9055507a86e98b8d474b161f48463fb6d794 +size 664955 diff --git a/2023/Sound Source Localization is All about Cross-Modal Alignment/layout.json b/2023/Sound Source Localization is All about Cross-Modal Alignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1a8536591661ce1ec102a7d0fa9e7801f76d9f49 --- /dev/null +++ b/2023/Sound Source Localization is All about Cross-Modal Alignment/layout.json @@ -0,0 +1,8357 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 103, + 493, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 103, + 493, + 121 + ], + "spans": [ + { + "bbox": [ + 100, + 103, + 493, + 121 + ], + "type": "text", + "content": "Sound Source Localization is All about Cross-Modal Alignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 142, + 467, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 142, + 467, + 172 + ], + "spans": [ + { + "bbox": [ + 116, + 142, + 467, + 172 + ], + "type": "text", + "content": "Arda Senocak" + }, + { + "bbox": [ + 116, + 142, + 467, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 116, + 142, + 467, + 172 + ], + "type": "text", + "content": " Hyeonggon Ryu" + }, + { + "bbox": [ + 116, + 142, + 467, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 116, + 142, + 467, + 172 + ], + "type": "text", + "content": " Junsik Kim" + }, + { + "bbox": [ + 116, + 142, + 467, + 172 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 116, + 142, + 467, + 172 + ], + "type": "text", + "content": " Tae-Hyun Oh" + }, + { + "bbox": [ + 116, + 142, + 467, + 172 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 116, + 142, + 467, + 172 + ], + "type": "text", + "content": " Hanspeter Pfister" + }, + { + "bbox": [ + 116, + 142, + 467, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 116, + 142, + 467, + 172 + ], + "type": "text", + "content": " Joon Son Chung" + }, + { + "bbox": [ + 116, + 142, + 467, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 108, + 175, + 485, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 175, + 485, + 190 + ], + "spans": [ + { + "bbox": [ + 108, + 175, + 485, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 175, + 485, + 190 + ], + "type": "text", + "content": " Korea Advanced Institute of Science and Technology " + }, + { + "bbox": [ + 108, + 175, + 485, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 175, + 485, + 190 + ], + "type": "text", + "content": " Harvard University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 190, + 507, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 190, + 507, + 204 + ], + "spans": [ + { + "bbox": [ + 85, + 190, + 507, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 85, + 190, + 507, + 204 + ], + "type": "text", + "content": "Dept. of Electrical Engineering and Grad. School of Artificial Intelligence, POSTECH" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 204, + 527, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 204, + 527, + 219 + ], + "spans": [ + { + "bbox": [ + 66, + 204, + 527, + 219 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 66, + 204, + 527, + 219 + ], + "type": "text", + "content": "Institute for Convergence Research and Education in Advanced Technology, Yonsei University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 258, + 192, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 270 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 270 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 283, + 290, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 290, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 290, + 510 + ], + "type": "text", + "content": "Humans can easily perceive the direction of sound sources in a visual scene, termed sound source localization. Recent studies on learning-based sound source localization have mainly explored the problem from a localization perspective. However, prior arts and existing benchmarks do not account for a more important aspect of the problem, cross-modal semantic understanding, which is essential for genuine sound source localization. Cross-modal semantic understanding is important in understanding semantically mismatched audio-visual events, e.g., silent objects, or off-screen sounds. To account for this, we propose a cross-modal alignment task as a joint task with sound source localization to better learn the interaction between audio and visual modalities. Thereby, we achieve high localization performance with strong cross-modal semantic understanding. Our method outperforms the state-of-the-art approaches in both sound source localization and cross-modal retrieval. Our work suggests that jointly tackling both tasks is necessary to conquer genuine sound source localization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 518, + 128, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 128, + 530 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 128, + 530 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 538, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 287, + 694 + ], + "type": "text", + "content": "Humans can easily perceive where the sound comes from in a scene. We naturally attend to the sounding direction and associate incoming audio-visual signals to understand the event. To achieve human-level audio-visual perception, sound source localization in visual scenes has been extensively studied [50, 51, 4, 47, 8, 35, 31, 33, 53, 54, 52, 36, 39, 38, 20]. Motivated by that humans learn from natural audio-visual correspondences without explicit supervision, most of the studies have been developed on a fundamental assumption that audio and visual signals are temporally correlated. With the assumption, losses of the sound source localization task are modeled by audio-visual correspondence as a self-supervision signal and are implemented by con" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 245, + 545, + 334 + ], + "blocks": [ + { + "bbox": [ + 307, + 245, + 545, + 334 + ], + "lines": [ + { + "bbox": [ + 307, + 245, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 545, + 334 + ], + "type": "image", + "image_path": "112abaefa8581f7e1d760d65eec2c80fc28af22ae6b6a062f2f72a0614e7dd21.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 336, + 545, + 358 + ], + "lines": [ + { + "bbox": [ + 305, + 336, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 336, + 545, + 358 + ], + "type": "text", + "content": "Figure 1. A conceptual difference between prior approaches and our alignment-based sound source localization." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 366, + 515, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 366, + 515, + 378 + ], + "spans": [ + { + "bbox": [ + 305, + 366, + 515, + 378 + ], + "type": "text", + "content": "trasting audio-visual pairs, i.e., contrastive learning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 379, + 546, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 546, + 581 + ], + "type": "text", + "content": "While these approaches appear to be unsupervised methods, they strongly rely on partial supervision information; e.g., using supervisedly pretrained vision networks [50, 51, 47, 53, 54, 20] and visual objectness estimators for postprocessing [39, 38]. Without leveraging such strong initial representations, the performance is degraded. Thus, the previous methods are not purely self-supervised approaches. Even further, there are recent studies [45, 39, 38] that point out visual objectness bias in existing sound source localization benchmarks and exploit the objectness prior to improve the localization accuracy. They show that, even without interaction between visual and audio signals, a model may achieve strong accuracy in localization by only referring visual signals alone, which is not the true intention of the sound source localization task. In short, the current evaluation and setting of the sound source localization do not capture the true sound source localization performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": "In this work, we first sort out evaluating sound source localization methods by introducing a cross-modal retrieval task as an auxiliary evaluation task. By this task, we can measure whether the learned representation has the capability to accurately interact between audio and visual modalities; i.e., more fine-grained audio-visual correspondence which is essential for genuine sound source localization. This aspect has been missed in existing sound source localization benchmarks. Indeed, our experiments show that higher sound localization performance does not guarantee higher cross-modal retrieval performance." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 702, + 205, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 702, + 205, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 702, + 205, + 713 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7777" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": "Second, given this additional criterion, we revisit the importance of semantic understanding shared across audio and visual modalities in both sound source localization and cross-modal retrieval. In the previous methods [50, 51, 54, 47], the cross-modal semantic alignment is induced by instance-level cross-modal contrastive learning, i.e., cross-modal instance discrimination between visual and audio features. However, they are aided by labels or supervisedly pretrained encoder for easing challenging cross-modal feature alignment. Instead, our method learns from scratch supporting the lack of guidance by incorporating multiple positive samples into cross-modal contrastive learning. Specifically, we construct a positive set for each modality using both multi-view [10] and conceptually similar samples [17]. Thereby, we enhance feature alignment and achieve high localization performance and strong cross-modal semantic understanding." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 275, + 289, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 289, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 289, + 395 + ], + "type": "text", + "content": "We evaluate our method on the VGG-SS and SoundNet-Flickr benchmarks for sound source localization and cross-modal retrieval. As aforementioned, the sound source localization task is closely related to the cross-modal retrieval task, but our experiments show that existing works have a weak performance correlation between them. This implies that we need to evaluate both tasks for evaluating the genuine sound source localization. The proposed method performs favorably against the recent state-of-the-art approaches in both tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 396, + 286, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 396, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 59, + 396, + 286, + 407 + ], + "type": "text", + "content": "We summarize the contributions of our work as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 411, + 287, + 547 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 411, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 287, + 458 + ], + "type": "text", + "content": "- We analyze that sound source localization benchmarks are not capable of evaluating cross-modal semantic understanding, thereby sound source localization methods may perform poorly in cross-modal retrieval tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 461, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 461, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 47, + 461, + 287, + 496 + ], + "type": "text", + "content": "- We propose semantic alignment to improve cross-modal semantic understanding of sound source localization models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 499, + 287, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 287, + 547 + ], + "type": "text", + "content": "- We expand semantic alignment with multi-views and conceptually similar samples which leads to state-of-the-art performance on both sound source localization and cross-modal retrieval." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 551, + 132, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 132, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 132, + 562 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 565, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 565, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 565, + 287, + 685 + ], + "type": "text", + "content": "Sound source localization. Sound source localization in visual scenes has been investigated by exploiting correspondences between audio and visual modalities. The most widely used approach for sound source localization is cross-modal attention [50, 51, 57] with contrastive loss [13, 29, 42]. Later, the attention-based method is improved by intra-frame hard sample mining [8], iterative contrastive learning with pseudo labels [35], feature regularization [36], positive mining [52], negative free learning [54] with stop-gradient operation [12], or momentum encoders [38]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": "Some sound localization approaches exploit additional semantic labels [47, 33, 53] or object prior [39, 63]. Semantic labels are used to pretrain audio and vision encoders with classification loss [33, 53] or refine audio-visual feature alignment [47]. A more explicit way to refine localization output is to use object prior. EZVSL [39] proposes post-processing to combine attention based localization output with a pretrained visual feature activation map. Similarly, Xuan et al. [63] propose to combine off-the-shelf object proposals with attention based sound localization results. However, postprocessing by object prior may generate a false positive output as it is solely based on vision without audio-visual interaction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 228, + 546, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 546, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 546, + 324 + ], + "type": "text", + "content": "In addition to the localization, there has been an attempt to localize sounding objects and recover the separated sounds simultaneously, also known as the cocktail party problem [27, 37]. The separation of sound mixture is achieved by predicting masks of spectrogram guided by visual features [19, 1, 64, 23, 62, 21, 2, 65, 24, 58, 56]. Furthermore, a number of recent papers are presented on audio-visual navigation for a given sound source [7, 22]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 330, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 545, + 390 + ], + "type": "text", + "content": "Self-supervised representation learning. In a broader categorization, sound source localization belongs to self-supervised multimodal learning. Our work is also relevant to self-supervised audio-visual representation learning, and other multimodal learning studies." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 390, + 546, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 546, + 521 + ], + "type": "text", + "content": "Contrastive learning aims to learn robust representations from large-scale raw data without annotations. Recent representation learning approaches [60, 10, 28, 11] use instance discrimination by contrastive learning [13, 29, 42] as a pretext task with notable advancements in visual recognition tasks. Recently, positive mining by nearest-neighbor search are used to learn representations of images [17, 18, 61], videos [26], neural recordings [6], and text-image [34]. In this work, we expand the previous works by incorporating both multi-views and conceptually similar samples into audio-visual modalities for cross-modal feature alignment." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "content": "A series of audio-visual representation learning studies have shown that audio and visual contents in a video are correlated, therefore a visual representation can be learned by sound prediction [44] or audio representation can be distilled from visual representation [5, 55]. Later, a variety of joint audio-visual representation learning methods are proposed with an assumption that there is a semantic [3, 30, 41, 40] or temporal [14, 43, 32, 15] correspondence between them. However, simply learning sound source localization by audio-visual correspondence with instance discrimination ignores the semantic similarity of audio-visual contents among samples, introducing false negatives or positives. In order to mitigate this issue, clustering [30], sampling [41], weighting [40], and hard mining [32] are proposed. Similarly, in this work, we go beyond instance discrimination by using multiple positive samples" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 44, + 693, + 288, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 44, + 693, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 44, + 693, + 288, + 715 + ], + "type": "text", + "content": "Typically, an image encoder is pretrained on ImageNet [16] and an audio encoder is pretrained on AudioSet [25] in supervised ways." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7778" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 547, + 224 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 547, + 224 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 547, + 224 + ], + "type": "image", + "image_path": "33bd5916b5d2db143c4350848ae6aaa9bf92f1f972ccfd9924bf6d94e623e3cf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 225, + 547, + 260 + ], + "lines": [ + { + "bbox": [ + 46, + 225, + 547, + 260 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 547, + 260 + ], + "type": "text", + "content": "Figure 2. Our sound source localization framework. Our model constructs multiple positive pairs with augmentation and Nearest Neighbor Search (Conceptually Similar Samples). By using these newly constructed 9 pairs, our model employs spatial localization, " + }, + { + "bbox": [ + 46, + 225, + 547, + 260 + ], + "type": "inline_equation", + "content": "s_L" + }, + { + "bbox": [ + 46, + 225, + 547, + 260 + ], + "type": "text", + "content": ", and semantic feature alignment, " + }, + { + "bbox": [ + 46, + 225, + 547, + 260 + ], + "type": "inline_equation", + "content": "s_A" + }, + { + "bbox": [ + 46, + 225, + 547, + 260 + ], + "type": "text", + "content": ", for each pair to learn a better sound source localization ability." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 279, + 262, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 279, + 262, + 292 + ], + "spans": [ + { + "bbox": [ + 47, + 279, + 262, + 292 + ], + "type": "text", + "content": "to enforce semantic understanding across modalities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 303, + 103, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 303, + 103, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 303, + 103, + 316 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 323, + 133, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 133, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 133, + 335 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 348, + 288, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 288, + 396 + ], + "type": "text", + "content": "Contrastive learning learns representation by containing positive and negative pairs. Given an encoded query sample " + }, + { + "bbox": [ + 46, + 348, + 288, + 396 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 348, + 288, + 396 + ], + "type": "text", + "content": " and its encoded positive pair " + }, + { + "bbox": [ + 46, + 348, + 288, + 396 + ], + "type": "inline_equation", + "content": "k^{+}" + }, + { + "bbox": [ + 46, + 348, + 288, + 396 + ], + "type": "text", + "content": " and negative pairs " + }, + { + "bbox": [ + 46, + 348, + 288, + 396 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 348, + 288, + 396 + ], + "type": "text", + "content": ", the loss can be defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 108, + 404, + 287, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 404, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 108, + 404, + 287, + 433 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = - \\log \\frac {\\exp (q \\cdot k ^ {+} / \\tau)}{\\sum_ {i} \\exp (q \\cdot k _ {i} / \\tau)} \\tag {1}", + "image_path": "7c8cb12887d27269bd204a37b4dc31a8fa506193a45b56007d57006d36024d63.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 441, + 201, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 201, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 201, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 441, + 201, + 453 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 441, + 201, + 453 + ], + "type": "text", + "content": " is the temperature parameter." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 459, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 543 + ], + "type": "text", + "content": "Cross-modal contrastive learning extends contrastive learning across multiple modalities. In sound source localization, audio-visual correspondence is used to define positive and negative cross-modal pairs. With an audio-visual dataset " + }, + { + "bbox": [ + 46, + 459, + 287, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(v_i,a_i):i = 1,\\dots,N\\}" + }, + { + "bbox": [ + 46, + 459, + 287, + 543 + ], + "type": "text", + "content": " and its encoded features " + }, + { + "bbox": [ + 46, + 459, + 287, + 543 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i = f_v(v_i)" + }, + { + "bbox": [ + 46, + 459, + 287, + 543 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 459, + 287, + 543 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i = f_a(a_i)" + }, + { + "bbox": [ + 46, + 459, + 287, + 543 + ], + "type": "text", + "content": ", cross-modal contrastive learning loss is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 99, + 552, + 287, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 552, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 99, + 552, + 287, + 581 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i} = - \\log \\frac {\\exp (s (\\mathbf {v} _ {i} , \\mathbf {a} _ {i}) / \\tau)}{\\sum_ {j} \\exp (s (\\mathbf {v} _ {i} , \\mathbf {a} _ {j}) / \\tau)} \\tag {2}", + "image_path": "0b8bfaea9325e9988ecfc78cdb76e32b171499f59ce48679a62a60b6243567ec.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 590, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 625 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 590, + 287, + 625 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 590, + 287, + 625 + ], + "type": "text", + "content": " is a cross-modal similarity function. The cross-modal contrastive loss Eq. (2) can be extended to symmetric form [48] as used in a few previous works [39, 38]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 635, + 222, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 222, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 222, + 647 + ], + "type": "text", + "content": "3.2. Cross-Modal Feature Alignment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": "We consider both spatial localization and semantic feature alignment for sound source localization. To this end, we use two different similarity functions " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "s_L" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "s_A" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " for contrastive learning (Eq. (2)), " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "s_L" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " for localization and " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "s_A" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " for cross-modal feature alignment." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 279, + 545, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 279, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 305, + 279, + 545, + 339 + ], + "type": "text", + "content": "Recent studies rely on audio-visual spatial correspondence maps to learn sound source localization by contrasting them. Given a spatial visual feature " + }, + { + "bbox": [ + 305, + 279, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\in \\mathbb{R}^{c \\times h \\times w}" + }, + { + "bbox": [ + 305, + 279, + 545, + 339 + ], + "type": "text", + "content": " and audio feature " + }, + { + "bbox": [ + 305, + 279, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{a} \\in \\mathbb{R}^{\\mathbf{c}}" + }, + { + "bbox": [ + 305, + 279, + 545, + 339 + ], + "type": "text", + "content": ", audio-visual similarity with a correspondence map can be calculated as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 356, + 346, + 545, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 346, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 356, + 346, + 545, + 376 + ], + "type": "interline_equation", + "content": "s _ {L} (\\mathbf {v}, \\mathbf {a}) = \\sum_ {x y \\in M} \\frac {1}{| M |} \\frac {\\mathbf {v} ^ {x y} \\cdot \\mathbf {a}}{\\| \\mathbf {v} ^ {x y} \\| \\| \\mathbf {a} \\|} \\tag {3}", + "image_path": "2180a1eca9f04264335473b44e89f42b0e86f53b7ba233591f4446ea4c2fe6dc.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 384, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 545, + 443 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 384, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^{xy}" + }, + { + "bbox": [ + 305, + 384, + 545, + 443 + ], + "type": "text", + "content": " is a feature vector at location " + }, + { + "bbox": [ + 305, + 384, + 545, + 443 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 305, + 384, + 545, + 443 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 384, + 545, + 443 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 384, + 545, + 443 + ], + "type": "text", + "content": " is an optional binary mask when an annotation or pseudomask [8, 36] is available. Since we assume no supervision for sound source localization, we do not use any mask, therefore, " + }, + { + "bbox": [ + 305, + 384, + 545, + 443 + ], + "type": "inline_equation", + "content": "M = \\mathbf{1}" + }, + { + "bbox": [ + 305, + 384, + 545, + 443 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 445, + 546, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 546, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 546, + 660 + ], + "type": "text", + "content": "The contrastive loss with localization similarity " + }, + { + "bbox": [ + 304, + 445, + 546, + 660 + ], + "type": "inline_equation", + "content": "s_L" + }, + { + "bbox": [ + 304, + 445, + 546, + 660 + ], + "type": "text", + "content": " enforces location dependent alignment giving sparse but strong audio-visual correspondence which enables to perform localization. However, our empirical studies on cross-modal retrieval indicate that strong localization performance does not guarantee semantic understanding. To overcome the low semantic understanding in recent studies, we propose to add instance-level contrastive loss. Instance-level contrasting encapsulates the whole context in a scene, enforcing better audio-visual semantic alignment. However, instance-level contrasting may smooth out spatial discriminativeness learned by Eq. (3). Inspired by SimCLR [10], we adopt a projection layer to align audio-visual semantics in a projection space. The projection layer separates the latent space of localization and semantic alignment, thereby preventing the alignment loss smoothing out the spatial discriminativeness. The similarity function for cross-modal feature alignment is defined as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 358, + 667, + 545, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 667, + 545, + 695 + ], + "spans": [ + { + "bbox": [ + 358, + 667, + 545, + 695 + ], + "type": "interline_equation", + "content": "s _ {A} (\\mathbf {v}, \\mathbf {a}) = \\frac {p _ {v} (\\operatorname {a v g} (\\mathbf {v})) \\cdot p _ {a} (\\mathbf {a})}{\\| p _ {v} (\\operatorname {a v g} (\\mathbf {v})) \\| \\| p _ {a} \\mathbf {a} \\|} \\tag {4}", + "image_path": "36a87b15a00b9f6c69ecc4aa514a033cfa707fbb976a2b4fae04a1d30f2b7b13.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{avg}(\\cdot)" + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": " is spatial average pooling, " + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "inline_equation", + "content": "p_v" + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": " is a projection" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7779" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "layer for visual features, and " + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "p_a" + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": " is a projection layer for audio features." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 106, + 268, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 268, + 119 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 268, + 119 + ], + "type": "text", + "content": "3.3. Expanding with Multiple Positive Samples" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 125, + 287, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 125, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 125, + 287, + 293 + ], + "type": "text", + "content": "Typically, contrastive learning contrasts between one positive pair and multiple negative pairs as shown in Eq. (1). In audio-visual learning, by an audio-visual correspondence assumption, an audio-image pair from the same clip is used as a positive pair while negative pairs are sampled from different clips. However, single-instance discrimination may not be sufficient to achieve strong cross-modal alignment. In this section, we expand contrastive learning beyond single instance discrimination by positive set construction and pairing them. To construct a positive set, we incorporate both hand-crafted positive and conceptual positive samples for each modality. Later, we adjust the contrastive learning to incorporate multiple positive pairs to enforce cross-modal alignment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 300, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 467 + ], + "type": "text", + "content": "Obtaining hand-crafted positive samples. Using randomly augmented samples as positive multi-view pairs are widely adopted in self-supervised representation learning, i.e., instance discrimination. Similarly, we extend a single anchor audio-image pair to multiple positive pairs by applying simple augmentations on image and audio samples separately. While we utilize common image transformations on images, we apply temporal shifting to audiOs. It is worth noting that sound source localization task learns from the underlying semantic consistency rather than subtle time differences as in videos. Thus, a slight shift in the audio may not alter contextual information significantly. As a result of hand-crafted multi-view positive pair generation, we obtain additional " + }, + { + "bbox": [ + 46, + 300, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^{aug}" + }, + { + "bbox": [ + 46, + 300, + 287, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 300, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{a}^{aug}" + }, + { + "bbox": [ + 46, + 300, + 287, + 467 + ], + "type": "text", + "content": " samples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 474, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 288, + 713 + ], + "type": "text", + "content": "Obtaining conceptual positive samples. Apart from manually created augmented views, we additionally expand our positive set with conceptually similar samples. The sampling strategy with nearest neighbor search can be performed in a various way, such as on-the-fly sampling [17, 49, 61, 34], sampling by pretrained encoders [52], or guided sampling [26, 18] using another modality. For selecting our conceptually similar samples, we utilize pretrained encoders. Note that pretrained encoders trained either with supervised or self-supervised learning are effective in positive sample mining as shown in the experiment section. By employing readily available image and audio encoders, we use the " + }, + { + "bbox": [ + 46, + 474, + 288, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 474, + 288, + 713 + ], + "type": "text", + "content": "-nearest neighborhood search to sample semantically similar samples in both modalities. In particular, given a pair of image and audio, we compute cosine similarity with all other samples and choose the top-" + }, + { + "bbox": [ + 46, + 474, + 288, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 474, + 288, + 713 + ], + "type": "text", + "content": " most similar samples among the training set for each modality. From a set of " + }, + { + "bbox": [ + 46, + 474, + 288, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 474, + 288, + 713 + ], + "type": "text", + "content": " samples, we randomly select one sample to obtain conceptually similar samples for each modality, " + }, + { + "bbox": [ + 46, + 474, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^{conc}" + }, + { + "bbox": [ + 46, + 474, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 474, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{a}^{conc}" + }, + { + "bbox": [ + 46, + 474, + 288, + 713 + ], + "type": "text", + "content": ". By utilizing the conceptually similar samples as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "content": "positive samples, our model expands semantic understanding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 102, + 545, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 102, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 305, + 102, + 545, + 234 + ], + "type": "text", + "content": "Pair Construction. Once we obtain the conceptual and hand-crafted positive samples for each modality, we proceed to create 9 distinct audio-visual pairs by pairing " + }, + { + "bbox": [ + 305, + 102, + 545, + 234 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\{\\mathbf{v},\\mathbf{v}^{aug},\\mathbf{v}^{conc}\\}" + }, + { + "bbox": [ + 305, + 102, + 545, + 234 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 102, + 545, + 234 + ], + "type": "inline_equation", + "content": "\\mathbf{A} = \\{\\mathbf{a},\\mathbf{a}^{aug},\\mathbf{a}^{conc}\\}" + }, + { + "bbox": [ + 305, + 102, + 545, + 234 + ], + "type": "text", + "content": ". This is done to ensure semantic alignment and consistency between them through contrastive learning. The negative pairs are randomly paired from the remaining samples in a training set. It is worth noting that some of these pairs are a combination of hand-crafted and conceptually similar samples, which further enhances the feature alignment of our model during training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 243, + 369, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 243, + 369, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 369, + 255 + ], + "type": "text", + "content": "3.4. Training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 262, + 545, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 262, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 545, + 309 + ], + "type": "text", + "content": "Our loss formulation incorporates both localization and instance-level similarity functions with multiple positive pairs constructed by augmentation and conceptually similar sample search. The final loss term is defined as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 322, + 319, + 545, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 319, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 322, + 319, + 545, + 387 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {i} = - \\sum_ {\\mathbf {v} _ {i} \\in \\mathbf {V}} \\sum_ {\\mathbf {a} _ {i} \\in \\mathbf {A}} \\left[ \\log \\frac {\\exp \\left(s _ {L} \\left(\\mathbf {v} _ {i} , \\mathbf {a} _ {i}\\right) / \\tau\\right)}{\\sum_ {j} \\exp \\left(s _ {L} \\left(\\mathbf {v} _ {i} , \\mathbf {a} _ {j}\\right) / \\tau\\right)} \\right. \\tag {5} \\\\ \\left. + \\log \\frac {\\exp \\left(s _ {A} \\left(\\mathbf {v} _ {i} , \\mathbf {a} _ {i}\\right) / \\tau\\right)}{\\sum_ {j} \\exp \\left(s _ {A} \\left(\\mathbf {v} _ {i} , \\mathbf {a} _ {j}\\right) / \\tau\\right)} \\right] \\\\ \\end{array}", + "image_path": "094edd218574d410c3a92ee0c6683face1edc8fcf1ee82b090b5db841fa30671.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 396, + 490, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 396, + 490, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 396, + 490, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 396, + 490, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 306, + 396, + 490, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 396, + 490, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 306, + 396, + 490, + 407 + ], + "type": "text", + "content": " indicate positive sample sets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 420, + 388, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 420, + 388, + 433 + ], + "spans": [ + { + "bbox": [ + 306, + 420, + 388, + 433 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 440, + 545, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 440, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 305, + 440, + 545, + 535 + ], + "type": "text", + "content": "Our proposed method for sound source localization is validated through experiments conducted on VG-GSound [9] and SoundNet-Flickr [5]. First, we conduct a quantitative analysis to evaluate the accuracy of the localization, cross-modal retrieval, and the impact of various components of our model. Then, we visualize our sound source localization results across different categories of sounds." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 544, + 413, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 413, + 558 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 413, + 558 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": "Datasets. Our method is trained using the VGGSound [9] and SoundNet-Flickr-144K [50, 51]. VGGSound is an audio-visual dataset containing around " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sim 200\\mathrm{K}" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": " videos. SoundNet-Flickr-144K set is the subset of SoundNetFlickr [5]. After training, we test the sound localization performance with VGG-SS [8] and SoundNet-Flickr-Test [50] datasets for the main experiments. These evaluation sets have bounding box annotations of sound sources for " + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sim 5\\mathrm{K}" + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": " and 250 samples, respectively. Moreover, we employ the AVSBench [66] and Extended VGGSound/SoundNetFlickr [38] datasets for additional evaluations. AVSBench dataset provides binary segmentation maps that show the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7780" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 288, + 255 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 288, + 255 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 288, + 255 + ], + "type": "table", + "html": "
MethodPre. VisionVGG-SSFlickr-SoundNet
cIoU ↑AUC ↑cIoU ↑AUC ↑
Attention [50]CVPR18✓18.5030.2066.0055.80
CoarseToFine [47]ECCV20✓29.1034.80--
LCBM [53]WACV22✓32.2036.60--
LVS [8]†CVPR21✗30.3036.4072.4057.80
LVS [8]CVPR21✗34.4038.2071.9058.20
HardPos [52]ICASSP22✗34.6038.0076.8059.20
SSPL (w/o PCM) [54]CVPR22✓27.0034.8073.9060.20
SSPL (w/ PCM) [54]CVPR22✓33.9038.0076.7060.50
EZ-VSL (w/o OGL) [39]ECCV22✓35.9638.2078.3161.74
SSL-TIE [36]ACMM22✗38.6339.6579.5061.20
SLAVC (w/o OGL) [38]NeurIPS22✓37.7939.4083.60-
Ours
↓ NN Search w/ Supervised Pre. Encoders✗39.9440.0279.6063.44
↓ NN Search w/ Self-Supervised Pre. Encoders✗39.2039.7079.2063.00
with OGL:
EZ-VSL (w/ OGL) [39]ECCV22✓38.8539.5483.9463.60
SLAVC (w/ OGL) [38]NeurIPS22✓39.80-86.00-
Ours (w/ OGL)
↓ NN Search w/ Supervised Pre. Encoders✗42.6441.4882.4064.60
↓ NN Search w/ Self-Supervised Pre. Encoders✗42.4741.4282.8064.48
with Optical Flow:
HearTheFlow [20]WACV23✓39.4040.0084.8064.00
", + "image_path": "1bfe09549e50b43a3f68f7a79b85158d6c86751effefa6ed88d110656aa2a23c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 255, + 287, + 309 + ], + "lines": [ + { + "bbox": [ + 47, + 255, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 287, + 309 + ], + "type": "text", + "content": "Table 1. Quantitative results on the VGG-SS and SoundNet-Flickr test sets. All models are trained with 144K samples from VGG-Sound and tested on VGG-SS and SoundNet-Flickr. " + }, + { + "bbox": [ + 47, + 255, + 287, + 309 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 47, + 255, + 287, + 309 + ], + "type": "text", + "content": " is the result of the model released on the official project page. SLAVC [38] does not provide AUC scores." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 325, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 287, + 373 + ], + "type": "text", + "content": "audio-visually correspondent pixels for roughly 5k five-second videos belonging to 23 categories. Lastly, the Extended VGGSound /SoundNet-Flickr dataset, proposed by [38], is used to understand non-visible sound sources." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 379, + 288, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 288, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 288, + 619 + ], + "type": "text", + "content": "Implementation details. We use two ResNet18 models for both audio and vision encoding. Unlike prior approaches, we do not fine-tune (or use a pretrained) a visual encoder from ImageNet pretrained weights. Instead, we train both the audio and vision encoders from scratch. We preprocess images and audios following the previous works [8, 52]. To create multiple pairs, we utilize both NN search and generic augmentation approaches. For NN search, we experiment on two different setups to retrieve k conceptually similar samples: (1) For supervisedly pretrained encoder experiments, We employ ResNet and VGGSound models pretrained on ImageNet and VGGSound respectively, (2) For self-supervisedly pretrained encoder experiments, we utilize the CLIP [48] Vision Encoder and Wav2CLIP [59] Audio Encoder. We use " + }, + { + "bbox": [ + 46, + 379, + 288, + 619 + ], + "type": "inline_equation", + "content": "k = 1000" + }, + { + "bbox": [ + 46, + 379, + 288, + 619 + ], + "type": "text", + "content": " for the experiments. To perform image augmentations, we follow the augmentations used in SimCLR [10]. For audios, we randomly select timewindow shifts in a time axis. The model is trained for 50 epochs with Adam Optimizer and a learning rate of 0.0001. " + }, + { + "bbox": [ + 46, + 379, + 288, + 619 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 379, + 288, + 619 + ], + "type": "text", + "content": " is set to 0.07 in contrastive learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 628, + 167, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 167, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 167, + 641 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "content": "Comparison with strong baselines. In this section, we conduct a comparative analysis of our sound source localization method against existing approaches. We carry out our evaluations in two settings, following previous approaches. Firstly, we train our model on VGGSound-144K" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 547, + 287 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 547, + 287 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 547, + 287 + ], + "type": "table", + "html": "
MethodPre. VisioncIoU ↑AUC ↑
Attention[50]CVPR18✓66.0055.80
DMC[30]CVPR19✓67.1056.80
LVS [8]†CVPR21X67.2056.20
LVS [8]CVPR21X69.9057.30
HardPos [52]ICASSP22X75.2059.70
SSPL (w/o PCM) [54]CVPR22✓69.9058.00
SSPL (w/ PCM) [54]CVPR22✓75.9061.00
EZ-VSL (w/o OGL) [39]ECCV22✓71.8958.81
SSL-TIE [36]ACMM22X81.5061.10
SLAVC (w/o OGL) [38]NeurIPS22✓--
Ours
↓ NN Search w/ Supervised Pre. EncodersX85.2062.20
↓ NN Search w/ Self-Supervised Pre. EncodersX84.8062.66
with OGL:
EZ-VSL (w/ OGL) [39]ECCV22✓83.1363.06
SLAVC (w/ OGL) [38]NeurIPS22✓--
Ours (w/ OGL)
↓ NN Search w/ Supervised Pre. EncodersX84.0064.16
↓ NN Search w/ Self-Supervised Pre. EncodersX84.4064.38
with Optical Flow:
HearTheFlow [20]WACV23✓86.5063.90
", + "image_path": "1a68b3b7007f70b34d80dbee402a0ff096f01f2f753b7993cb4c6df78e1410a9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 289, + 547, + 334 + ], + "lines": [ + { + "bbox": [ + 305, + 289, + 547, + 334 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 547, + 334 + ], + "type": "text", + "content": "Table 2. Quantitative results on the SoundNet-Flickr test set. All models are trained and tested on the SoundNet-Flickr 144K dataset. " + }, + { + "bbox": [ + 305, + 289, + 547, + 334 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 305, + 289, + 547, + 334 + ], + "type": "text", + "content": " is the result of the model from the official project page. SLAVC [38] does not provide results with SoundNet-Flickr 144K." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 352, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 460 + ], + "type": "text", + "content": "and evaluate it on VGG-SS and SoundNet-Flickr test sets. Secondly, we train our model on SoundNet-Flickr-144K and evaluate it on the SoundNet-Flickr test set. It is important to note that all the compared models are trained using the same amount of data. AVEL [57], AVObject [2], and LCBM [53] models rely on video input, and as such, they cannot be evaluated on the SoundNet-Flickr dataset, which contains static image and audio pairs. We present our results in Table 1 and Table 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": "Our proposed model achieves higher performance compared to prior approaches on both test sets. Specifically, it yields a " + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "inline_equation", + "content": "+2.15\\%" + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": " cIoU and " + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "inline_equation", + "content": "+0.6\\%" + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": " AUC improvement on VGGSS, as well as a " + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "inline_equation", + "content": "+3.7\\%" + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": " cIoU improvement on SoundNet-Flickr compared to the state-of-the-art methods that uses pretrained vision encoder. It is worth highlighting that unlike the majority of previous works, our proposed model does not utilize a vision encoder pretrained on ImageNet in a sound source localization backbone. This is because, as discussed in Mo et al. [38], using supervisedly pretrained vision encoders makes the sound source localization problem a weakly supervised problem. However, it is worth noting that even without using a pretrained vision encoder, our method achieves state-of-the-art performance on both experiments that are presented in Table 1 and Table 2. We demonstrate the performance of our model with the pretrained models learned through supervised learning (NN Search w/ Supervised Pre. Encoders) and with models that are pretrained through self-supervised learning (NN Search w/ Self-Supervised Pre. Encoders) in NN Search module. As the results indicate, using self-supervised pre" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "7781" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 279, + 226 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 279, + 226 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 279, + 226 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 279, + 226 + ], + "type": "table", + "html": "
Test ClassMethodPre. VisioncIoU ↑AUC ↑
Heard 110LVS [8]CVPR21X28.9036.20
EZ-VSL(w/o OGL) [39]ECCV22✓31.8636.19
SLAVC(w/o OGL) [38]NeurlPS22✓35.84-
OursX38.3139.05
with OGL:
EZ-VSL(w/o OGL) [39]ECCV22✓37.2538.97
SLAVC(w/o OGL) [38]NeurlPS22✓38.22-
Ours(w/ OGL)X41.8540.93
with Optical Flow:
HearTheFlow [20]WACV23✓37.3038.60
Unheard 110LVS [8]CVPR21X26.3034.70
EZ-VSL(w/o OGL) [39]ECCV22✓32.6636.72
SLAVC(w/o OGL) [38]NeurlPS22✓36.50-
OursX39.1139.80
with OGL:
EZ-VSL(w/o OGL) [39]ECCV22✓39.5739.60
SLAVC(w/o OGL) [38]NeurlPS22✓38.87-
Ours(w/ OGL)✓42.9441.54
with Optical Flow:
HearTheFlow [20]WACV23✓39.3040.00
", + "image_path": "489603953d1cbec4b8dae37db97d607e698c7d0a1f21db171399ca4ea762ded5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 67, + 259, + 269, + 303 + ], + "blocks": [ + { + "bbox": [ + 47, + 227, + 286, + 249 + ], + "lines": [ + { + "bbox": [ + 47, + 227, + 286, + 249 + ], + "spans": [ + { + "bbox": [ + 47, + 227, + 286, + 249 + ], + "type": "text", + "content": "Table 3. Comparison results on open-set audio-visual localization experiments trained and tested on the splits of [8, 39, 20]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 259, + 269, + 303 + ], + "lines": [ + { + "bbox": [ + 67, + 259, + 269, + 303 + ], + "spans": [ + { + "bbox": [ + 67, + 259, + 269, + 303 + ], + "type": "table", + "html": "
Test ClassMethodPre. VisioncIoU ↑AUC ↑
Heard 110SSSL-TIE [36]ACMM22X39.0040.30
OursX41.2041.00
Unheard 110SSSL-TIE [36]ACMM22X36.5038.60
OursX36.9038.59
", + "image_path": "17c4691d1129feefc6a3eb79024699fd9250fc974e1c9e9600b99f34ccdce179.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 340, + 286, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 286, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 286, + 436 + ], + "type": "text", + "content": "trained encoders in NN Search performs on par with the supervised pretrained encoders in NN Search. This shows that our model does not depend on supervised pretrained encoders for the NN search module and can utilize any type of pretrained encoder feature for nearest neighbor search. Note that these pretrained encoders are not used in the backbone networks of the sound source localization module but only in the NN Search Module, as illustrated in Figure 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 286, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 286, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 286, + 593 + ], + "type": "text", + "content": "We also discuss the methods employed by previous studies, such as SSPL [54] which utilizes a sub-module called PCM to reduce the impact of background noise, HTF [20] which utilizes Optical Flow, and EZ-VSL [39] which refines its initial audio-visual localization outcomes through object guidance obtained from an ImageNet pretrained visual encoder. Our model, on the other hand, and any of its variations do not require any task-specific modules or operations to achieve the state-of-the-art (SOTA) results. This suggests that using additional semantic and multi-view correspondence, as well as feature alignment, provides more varied and robust supervision for better aligned audio and visual features, as opposed to using task-specific approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 714 + ], + "type": "text", + "content": "The quantitative results presented in Table 1 and Table 2 also showcase the performance of previous methods that utilize object guidance to evaluate their final sound source localizations. Our model outperforms all previous methods that employ object guidance on the VGG-SS test set and achieves comparable results on the SoundNet-Flickr test set, even though our model does not use object guided refinement (OGL). Additionally, we acknowledge that the addition of OGL to our audio-visual localization results in improvement on the VGGSS test set, while degrading perfor" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 545, + 133 + ], + "blocks": [ + { + "bbox": [ + 47, + 303, + 286, + 325 + ], + "lines": [ + { + "bbox": [ + 47, + 303, + 286, + 325 + ], + "spans": [ + { + "bbox": [ + 47, + 303, + 286, + 325 + ], + "type": "text", + "content": "Table 4. Comparison results on open set audio-visual localization experiments trained and tested on the splits of [36]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 545, + 133 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 133 + ], + "type": "table", + "html": "
Test SetMethodPre. VisionmIoU ↑F-Score ↑
S4LVS (w/o OGL) [8]CVPR21X26.933.6
EZ-VSL (w/o OGL) [39]ECCV22✓27.634.2
SLAVC (w/o OGL) [38]NeuRTPS22✓28.134.6
Ours (w/o OGL)
↓ NN Search w/ Supervised Pre. EncodersX29.635.9
↓ NN Search w/ Self-Supervised Pre. EncodersX29.335.6
", + "image_path": "7d60cb960c0e2b6b681424b7958fc2453c7563e26e3ccce1ed9dfd444b841718.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 174, + 548, + 243 + ], + "blocks": [ + { + "bbox": [ + 306, + 142, + 545, + 163 + ], + "lines": [ + { + "bbox": [ + 306, + 142, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 306, + 142, + 545, + 163 + ], + "type": "text", + "content": "Table 5. Quantitative results on AVS Bench S4 dataset. All models are trained on the VGGSound 144K dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 174, + 548, + 243 + ], + "lines": [ + { + "bbox": [ + 307, + 174, + 548, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 548, + 243 + ], + "type": "table", + "html": "
ModelPre. VisionA → II → A
R@1R@5R@10R@1R@5R@10
LVS [8]CVPR21X3.8712.3520.734.9014.2921.37
EZ-VSL [39]ECCV22✓5.0115.7324.8114.233.5145.18
SSL-TIE [36]MM22X10.2930.6843.7612.7629.5839.72
SLAVC [38]NeurIPS22✓4.7713.0819.106.1221.1632.12
Ours
↓ NN Search w/ Supervised Pre. EncodersX16.4736.9949.0020.0942.3853.66
↓ NN Search w/ Self-Supervised Pre. EncodersX14.3137.8149.1718.0038.3949.02
", + "image_path": "337b012838eae12d4d2878a284dc4310aef39f5a97d9bc1b9a44f971200893f9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 243, + 545, + 276 + ], + "lines": [ + { + "bbox": [ + 306, + 243, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 545, + 276 + ], + "type": "text", + "content": "Table 6. Summary of retrieval recall scores for all models. All of the models are trained on VGGSound 144K data and retrieval is performed on entire VGG-SS dataset, containing " + }, + { + "bbox": [ + 306, + 243, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\sim 5\\mathrm{K}" + }, + { + "bbox": [ + 306, + 243, + 545, + 276 + ], + "type": "text", + "content": " samples." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 286, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 545, + 394 + ], + "type": "text", + "content": "mance on the SoundNet-Flickr test set. In contrast, prior methods see modest improvements when utilizing OGL. This can be explained by the fact that our model is already accurately localizing the sounding objects, and object guidance can interfere with localization results by introducing visual regions that are not sounding (refer to Section 4.4 for visual results). Unlike prior methods, we do not use OGL in our architecture for the remainder of this paper, unless it is being directly compared with OGL-based methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 396, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 467 + ], + "type": "text", + "content": "Finally, in comparison to HearTheFlow, which utilizes an additional Optical Flow modality, our method outperforms it on the VGGSS test set, and achieves slightly lower performance on the SoundNet-Flickr test set without utilizing any additional modalities, but instead relying on better audio-visual correspondence and alignment." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": "Open Set Audio-Visual Localization. The study by Chen et al. [8] evaluates the generalization ability of sound source localization methods in an open set scenario. This involves testing the models on categories that are both present in the training data (heard) and categories that are not present (unheard). To accomplish this, 110 randomly selected categories from the VGGSound dataset are used for training, while another disjoint set of 110 categories are reserved for evaluation to ensure the model have never seen or heard them before. It should be noted that not all previous works use the same train/test splits. While some works, including [8, 39, 38, 46], share the same splits,[36] uses a different split. Therefore, to ensure a fair comparison, we conduct experiments on both splits, evaluating on test samples from both heard and unheard categories. The results are shown in Table 3 and Table 4. Our model outperforms existing approaches on both categories, regardless of the train/test splits used. Specifically, in Table 3, our model (w/o OGL) even surpasses the other models that use OGL. Previous approaches draw different conclusions from" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7782" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 286, + 152 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 152 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 152 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 152 + ], + "type": "table", + "html": "
MethodPre. VisionExtended Flickr-SoundNetExtended VGG-SS
AP ↑max-F1 ↑LocAcc ↑AP ↑max-F1 ↑LocAcc ↑
CoarseToFine [47]ECCV20✓0.0038.2047.200.0019.8021.93
LVS [8]CVPR21✗9.8017.9019.605.159.9010.43
Attention10k [50]CVPR18✓15.9824.0034.166.7013.1014.04
DMC [30]CVPR19✓25.5641.8052.8011.5320.3022.63
DSOL [31]NeuIPS20✓38.3249.4072.9116.8425.6026.87
OGL [39]ECCV22-40.2055.7077.2018.7330.9036.58
EZ-VSL (w/o OGL) [39]ECCV22✓46.3054.6066.4024.5530.9031.58
SLAVC (w/o OGL) [38]NeuIPS22✓51.6359.1083.6032.9540.0037.79
Ours
↓ NN Search w/ Supervised Pre. Encoders✗64.4366.9079.6034.7340.7039.94
↓ NN Search w/ Self-Supervised Pre. Encoders✗62.6766.1079.2033.0940.0039.20
", + "image_path": "e4bad1492f02903387a789ff0d6eb5b172b37e82f0fae4a24dd9d658d0d8a7a7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 209, + 288, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 288, + 376 + ], + "type": "text", + "content": "these open set experiments. While some conclude that their models have strong generalization ability because their performance in unheard categories is higher than heard categories [39, 38, 46], the other works that cannot achieve the same trend discuss that this is expected since their models are dealing with unseen categories [36]. However, our results show that these conclusions are highly dependent on the chosen train/test splits. Our model performs better than existing works in both splits, but there is no uniform trend in between two splits. While our method performs better on unheard categories in the splits of [8, 39, 38, 46], it performs worse on unheard categories in the split of [36]. Therefore, we conclude that the observed trends are highly dependent on the randomly selected train/test splits." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 383, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 383, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 383, + 287, + 515 + ], + "type": "text", + "content": "AVSBench [66]. To demonstrate the precise sound localization ability of our model, we conduct experiments on the AVSBench S4 dataset. The dataset's objective is to detect audio-visual correspondence and correlation at the pixel level. To make a fair comparison, we use some of the self-supervised sound source localization methods mentioned earlier. All models are trained on VGGSound-144K and directly assessed on the AVSBench S4 dataset without any further fine-tuning (zero-shot setting). Our results, which are presented in Table 5, indicate that our method achieves the highest performance, as in the previous experiments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 521, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 713 + ], + "type": "text", + "content": "Retrieval. We evaluate sound localization models on the VGG-SS dataset for cross-modal retrieval. As shown in Table 6, our method clearly outperforms other state-of-the-art methods. One interesting observation is that EZ-VSL [39] notably performs better than SLAVC [38] on cross-modal retrieval, while SLAVC performs better on sound source localization in Table 1. This shows that with the current benchmark evaluations, better sound localization performance does not guarantee better audio-visual semantic understanding, thereby we need to additionally evaluate sound source localization methods on cross-modal understanding tasks. Another observation is that the performance gap between our method and the strongest competitor SSL-TIE [36] is notably larger on cross-modal retrieval than sound source localization. This is due to the strong cross-modal feature alignment of our method that is over" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 545, + 161 + ], + "blocks": [ + { + "bbox": [ + 47, + 154, + 287, + 198 + ], + "lines": [ + { + "bbox": [ + 47, + 154, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 154, + 287, + 198 + ], + "type": "text", + "content": "Table 7. Quantitative results on the Extended VGG-SS and Extended SoundNet-Flickr sets. All models are trained with 144K samples from VGG-Sound. The results of the prior approaches are obtained from [38]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 545, + 161 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 161 + ], + "type": "table", + "html": "
SemanticMulti-ViewFeature AlignmentcIoU ↑AUC ↑
(A)✓✓✓39.9440.02
(B)✓✓✗39.1039.44
(C)✓✗✓38.7539.34
(D)✓✗✗38.2438.90
(E)✗✓✓38.3039.38
(F)✗✓✗37.7239.19
(G)✗✗✓34.9337.94
(H)✗✗✗34.2237.67
", + "image_path": "b7ec74897f50ce3c467909d7316d744e3480b483192f98571ddc41c4aeff23c6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 163, + 545, + 186 + ], + "lines": [ + { + "bbox": [ + 306, + 163, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 306, + 163, + 545, + 186 + ], + "type": "text", + "content": "Table 8. Ablation studies on our proposed method to see the impact of each main component." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 202, + 518, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 202, + 518, + 214 + ], + "spans": [ + { + "bbox": [ + 306, + 202, + 518, + 214 + ], + "type": "text", + "content": "looked in the sound source localization benchmarks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 220, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 352 + ], + "type": "text", + "content": "Extended Flickr and VGG-SS datasets. The prior study [38] points out that the current sound source localization benchmarks overlook false positive detection. It is because the evaluation samples always contain at least a sounding object in a scene; thus cannot capture false positive outputs, e.g., silent objects or off-screen sounds. To analyze false positive detection, Mo and Morgado [38] extended the benchmarks with non-audible, non-visible, and mismatched audio-visual samples. The expectation is that a sound source localization model should not localize any objects when audio-visual semantics do not match." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 354, + 546, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 546, + 460 + ], + "type": "text", + "content": "The experiment with the extended datasets in Table 7 shows that our method performs favorably against state-of-the-art competitors. Our method performs better than the competing methods in false positive detection measured by AP and max-F1, while SLAVC [38] achieves better localization performance on Extended Flickr-SoundNet. As both false positive detection and cross-modal retrieval require cross-modal interaction, our method shows strong performance on both tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 472, + 406, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 472, + 406, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 406, + 483 + ], + "type": "text", + "content": "4.3. Ablation Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 491, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 563 + ], + "type": "text", + "content": "We conduct a series of experiments in order to verify our design choices and make further analysis. To save computational time and resources, we primarily perform ablation studies by training our model on VGGSound-144K with NN Search w/ Supervised Pre. Encoders setup and evaluating it on VGG-SS. Results are in Table 8." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Impact of Semantic and Multi-View Invariance. In order to understand the impact of each type of invariance (consistency), we analyze the performance of our model with different type of invariance methodologies in Table 8. As the results of (C vs. E) and (D vs. F) reveal, using semantically similar samples (semantic invariance) produces better performance " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "(+0.45\\%)" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "+0.5\\%" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " on cIoU respectively compared to augmented multi-view invariance. Moreover, as the results of (A vs. C) and (A vs. E) depict, the combination of these two different types of invariance complement each other and and further enhances the model's performance. Using pair combination of these two different" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7783" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 78, + 300, + 296 + ], + "blocks": [ + { + "bbox": [ + 50, + 78, + 300, + 296 + ], + "lines": [ + { + "bbox": [ + 50, + 78, + 300, + 296 + ], + "spans": [ + { + "bbox": [ + 50, + 78, + 300, + 296 + ], + "type": "image", + "image_path": "2b7baad645a34fd19154d0134d9422c38aedcd09c7a9caf896a17d43b6ad4c5f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 306, + 459, + 317 + ], + "lines": [ + { + "bbox": [ + 127, + 306, + 459, + 317 + ], + "spans": [ + { + "bbox": [ + 127, + 306, + 459, + 317 + ], + "type": "text", + "content": "Figure 3. Sound Localization Results on VGG-SS (top) and SoundNet-Flickr (bottom)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 307, + 79, + 544, + 296 + ], + "blocks": [ + { + "bbox": [ + 307, + 79, + 544, + 296 + ], + "lines": [ + { + "bbox": [ + 307, + 79, + 544, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 79, + 544, + 296 + ], + "type": "image", + "image_path": "19523afffaeccb4433899c6fbcace7e5da95cdbe3f1e2cb22fe36193fc607b48.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 59, + 323, + 269, + 369 + ], + "blocks": [ + { + "bbox": [ + 59, + 323, + 269, + 369 + ], + "lines": [ + { + "bbox": [ + 59, + 323, + 269, + 369 + ], + "spans": [ + { + "bbox": [ + 59, + 323, + 269, + 369 + ], + "type": "table", + "html": "
k in k-NN10301005001000
cIoU ↑38.8038.8239.4639.9039.94
AUC ↑39.5139.6739.9340.0040.02
", + "image_path": "5100caa3a583462fe503ccb57d03ac9cd173ca38161f28a5df47f684843ea094.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 370, + 274, + 381 + ], + "lines": [ + { + "bbox": [ + 47, + 370, + 274, + 381 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 274, + 381 + ], + "type": "text", + "content": "Table 9. Varying k in conceptually similar sample selection." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 390, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 390, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 390, + 287, + 437 + ], + "type": "text", + "content": "types of consistency elements provides additional supervi sions, invariance and alignments, leading to a more robust representation space and improve sound localization performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 444, + 287, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 287, + 576 + ], + "type": "text", + "content": "Impact of Feature Alignment. We perform controlled experiments to verify the effect of the feature alignment strategy, and the results are presented in Table 8. Comparing the performance of the proposed model with and without feature alignment, (A vs. B), highlights the importance of this strategy to boost the performance. Further, examining the results of experiments (C vs. D) and (E vs. F) reveals that feature alignment provides additional gains irrespective of the consistency types. These findings indicate that global feature-based alignment helps the optimization of audio-visual correspondence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 287, + 713 + ], + "type": "text", + "content": "Impact of " + }, + { + "bbox": [ + 47, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 582, + 287, + 713 + ], + "type": "text", + "content": " in conceptually similar sample selection. Selecting an appropriate " + }, + { + "bbox": [ + 47, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 582, + 287, + 713 + ], + "type": "text", + "content": " value for sampling nearest neighbors is crucial. If this value is set too high, it may result in noisy samples that could disrupt the learning phase. Conversely, if the value is set too low, only very similar samples to the anchor will be provided and it limits semantic invariance. Nevertheless, when compared to Table 8 (E), we observe performance gain throughout the range of " + }, + { + "bbox": [ + 47, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 582, + 287, + 713 + ], + "type": "text", + "content": " used for the ablation study. Table 9 shows an ablative evaluation of the effect of " + }, + { + "bbox": [ + 47, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 582, + 287, + 713 + ], + "type": "text", + "content": " value used to select neighborhood samples. The results indicate that an optimal choice is " + }, + { + "bbox": [ + 47, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "k = 1000" + }, + { + "bbox": [ + 47, + 582, + 287, + 713 + ], + "type": "text", + "content": ". This" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 313, + 329, + 539, + 376 + ], + "blocks": [ + { + "bbox": [ + 313, + 329, + 539, + 376 + ], + "lines": [ + { + "bbox": [ + 313, + 329, + 539, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 539, + 376 + ], + "type": "image", + "image_path": "cfc727a4fc22eeae7874a4d683920c4574791ebd9ce59d0695b65c37667021ce.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 380, + 545, + 401 + ], + "lines": [ + { + "bbox": [ + 306, + 380, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 545, + 401 + ], + "type": "text", + "content": "Figure 4. OGL degrades our sound localization results on SoundNet-Flickr." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 410, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 410, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 306, + 410, + 545, + 434 + ], + "type": "text", + "content": "choice of " + }, + { + "bbox": [ + 306, + 410, + 545, + 434 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 410, + 545, + 434 + ], + "type": "text", + "content": " can be explained by the fact that it provides a balance between semantic similarity and sufficient diversity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 442, + 417, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 442, + 417, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 442, + 417, + 455 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 461, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 567 + ], + "type": "text", + "content": "In this section, we visualize and compare our sound localization results with the recent prior works on standard benchmarks, namely on VGG-SS and SoundNet-Flickr. The visualized samples in Figure 3 show that localized regions of the proposed method are more compact and accurately aligns with the sounding objects than the other methods. For instance, small size musical instrument is localized accurately compared to the recent methods in the top right column." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "type": "text", + "content": "We also compare our localization results with and without object-guided localization (OGL). As shown in Figure 4, OGL deteriorates our sound localization outputs. OGL captures objectness in a scene, thereby tending to attend to any distinctive objects regardless of whether it is the sound source or not. Therefore, OGL can be helpful when localization totally fails because of the objectness bias in the benchmarks, but it is harmful when the localization is accurate which is the case for the examples shown. This result is consistent with the quantitative result in Table 2, showing that our method with OGL performs worse." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "content": "Throughout the paper, we discuss the importance of" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7784" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 75, + 286, + 259 + ], + "blocks": [ + { + "bbox": [ + 51, + 75, + 286, + 259 + ], + "lines": [ + { + "bbox": [ + 51, + 75, + 286, + 259 + ], + "spans": [ + { + "bbox": [ + 51, + 75, + 286, + 259 + ], + "type": "image", + "image_path": "1083664842bda8e3a257421b52ffda1b5208ee05bd0b99922367f75c9f5d836e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 262, + 287, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 297 + ], + "type": "text", + "content": "Figure 5. Interactive Sound Localization of Ours and SLAVC [38]. Our model correctly follows the cross-modal interaction for given different sounds." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 315, + 289, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 289, + 423 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 289, + 423 + ], + "type": "text", + "content": "cross-modal semantic understanding. We demonstrate interactiveness of our method across modalities in Figure 5. Genuine sound source localization should be able to localize objects that are correlated with the sound. To visualize cross-modal interaction, we synthetically pair up the same image with different sounds of objects that are visible in a scene. The examples demonstrate that the proposed method can localize different objects depending on the contexts of sounds, while the competing method can not." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 449, + 119, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 119, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 119, + 462 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "content": "In this work, we investigate cross-modal semantic understanding that has been overlooked in sound source localization studies. We observe that higher sound source localization performance on the current benchmark does not necessarily show higher performance in cross-modal retrieval, despite its causal relevance in reality. To enforce strong understanding of audio-visual semantic matching while maintaining localization capability, we propose semantic alignment with multi-views of audio-visual pairs in a simple yet effective way. The ablation study shows that strong semantic alignment is achieved when both semantic alignment loss and enriched positive pairs are used. We extensively evaluate our method on sound source localization benchmarks including cross-dataset and open-set settings. Moreover, our analyses on cross-modal retrieval and false positive detection verify that the proposed method has strong capability in cross-modal interaction. Our study suggests that sound localization methods should be evaluated not only on localization benchmarks but also on cross-modal understanding tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 411, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 411, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 411, + 85 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 91, + 547, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 224 + ], + "type": "text", + "content": "This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2023-00212845, Multimodal Speech Processing for Human-Computer Interaction). H. Pfister and J. Kim were partially supported by NIH grant R01HD104969. T.-H. Oh was partially supported by IITP grant funded by the Korea government (MSIT) (No. 2021-0-02068, Artificial Intelligence Innovation Hub; No. 2022-0-00290, Visual Intelligence for Space-Time Understanding and Generation based on Multi-layered Visual Common Sense)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 234, + 365, + 246 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 234, + 365, + 246 + ], + "spans": [ + { + "bbox": [ + 307, + 234, + 365, + 246 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 253, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 310, + 253, + 545, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 253, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 310, + 253, + 545, + 287 + ], + "type": "text", + "content": "[1] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. The conversation: Deep audio-visual speech enhancement. In INTERSPEECH, pages 3244–3248, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 311, + 288, + 547, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 288, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 311, + 288, + 547, + 331 + ], + "type": "text", + "content": "[2] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. In European Conference on Computer Vision, 2020. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 311, + 333, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 333, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 311, + 333, + 545, + 365 + ], + "type": "text", + "content": "[3] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In IEEE International Conference on Computer Vision, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 367, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 367, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 311, + 367, + 545, + 399 + ], + "type": "text", + "content": "[4] Relja Arandjelović and Andrew Zisserman. Objects that sound. In European Conference on Computer Vision, 2018. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 401, + 545, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 401, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 311, + 401, + 545, + 444 + ], + "type": "text", + "content": "[5] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Soundnet: Learning sound representations from unlabeled video. Advances in Neural Information Processing Systems, NeurIPS, 2016. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 446, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 446, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 311, + 446, + 545, + 511 + ], + "type": "text", + "content": "[6] Mehdi Azabou, Mohammad Gheshlaghi Azar, Ran Liu, ChiHeng Lin, Erik C Johnson, Kiran Bhaskaran-Nair, Max Dabagia, Bernardo Avila-Pires, Lindsey Kitchell, Keith B Hengen, et al. Mine your own view: Self-supervised learning through across-sample prediction. arXiv preprint arXiv:2102.10106, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 513, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 513, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 311, + 513, + 545, + 567 + ], + "type": "text", + "content": "[7] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In European Conference on Computer Vision, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 569, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 569, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 311, + 569, + 545, + 621 + ], + "type": "text", + "content": "[8] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 311, + 624, + 545, + 668 + ], + "type": "text", + "content": "[9] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2020. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, 2020. 2, 3, 5" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7785" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[11] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "type": "text", + "content": "[12] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 287, + 185 + ], + "type": "text", + "content": "[13] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In IEEE Conference on Computer Vision and Pattern Recognition, 2005. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 287, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 287, + 219 + ], + "type": "text", + "content": "[14] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asia Conference on Computer Vision, pages 251-263. Springer, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 287, + 262 + ], + "type": "text", + "content": "[15] Soo-Whan Chung, Hong Goo Kang, and Joon Son Chung. Seeing voices and hearing voices: learning discriminative embeddings using cross-modal self-supervision. In INTER-SPEECH, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 265, + 286, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 265, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 265, + 286, + 308 + ], + "type": "text", + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition, 2009. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 310, + 286, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 310, + 286, + 365 + ], + "spans": [ + { + "bbox": [ + 49, + 310, + 286, + 365 + ], + "type": "text", + "content": "[17] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In IEEE International Conference on Computer Vision, 2021. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 366, + 286, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 366, + 286, + 399 + ], + "spans": [ + { + "bbox": [ + 49, + 366, + 286, + 399 + ], + "type": "text", + "content": "[18] Mohamed El Banani, Karan Desai, and Justin Johnson. Learning Visual Representations via Language-Guided Sampling. In CVPR, 2022. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 400, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 400, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 49, + 400, + 287, + 464 + ], + "type": "text", + "content": "[19] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. ACM Transactions on Graphics (Proceedings of SIGGRAPH), 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 467, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 467, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 49, + 467, + 287, + 521 + ], + "type": "text", + "content": "[20] Dennis Fedorishin, Deen Dayal Mohan, Bhavin Jawade, Srirangaraj Setlur, and Venu Govindaraju. Hear the flow: Optical flow-based self-supervised visual sound source localization. In IEEE Winter Conf. on Applications of Computer Vision (WACV), 2023. 1, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 523, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 523, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 49, + 523, + 287, + 566 + ], + "type": "text", + "content": "[21] Chuang Gan, Deng Huang, Hang Zhao, Joshua B. Tenenbaum, and Antonio Torralba. Music gesture for visual sound separation. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 568, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 568, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 49, + 568, + 287, + 610 + ], + "type": "text", + "content": "[22] Chuang Gan, Yiwei Zhang, Jiajun Wu, Boqing Gong, and Joshua B. Tenenbaum. Look, listen, and act: Towards audiovisual embodied navigation. In IEEE International Conference on Robotics and Automation, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 613, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 613, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 49, + 613, + 287, + 645 + ], + "type": "text", + "content": "[23] Ruohan Gao and Kristen Grauman. Co-separating sounds of visual objects. In IEEE International Conference on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "text", + "content": "[24] Ruohan Gao and Kristen Grauman. Visualvoice: Audiovisual speech separation with cross-modal consistency. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 691, + 287, + 713 + ], + "type": "text", + "content": "[25] Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 115 + ], + "type": "text", + "content": "and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 160 + ], + "type": "text", + "content": "[26] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. In Advances in Neural Information Processing Systems, NeurIPS, 2020. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 162, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 183 + ], + "type": "text", + "content": "[27] Simon Haykin and Zhe Chen. The cocktail party problem. Neural computation, 17(9):1875-1902, 2005. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 185, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 545, + 227 + ], + "type": "text", + "content": "[28] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "type": "text", + "content": "[29] Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In International workshop on similarity-based pattern recognition, 2015. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 261, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 303 + ], + "type": "text", + "content": "[30] Di Hu, Feiping Nie, and Xuelong Li. Deep multimodal clustering for unsupervised audiovisual learning. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2, 5, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 305, + 545, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 359 + ], + "type": "text", + "content": "[31] Di Hu, Rui Qian, Minyue Jiang, Xiao Tan, Shilei Wen, Errui Ding, Weiyao Lin, and Dejing Dou. Discriminative sounding objects localization via self-supervised audiovisual matching. Advances in Neural Information Processing Systems, NeurIPS, 2020. 1, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 361, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 403 + ], + "type": "text", + "content": "[32] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. Advances in Neural Information Processing Systems, 31, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 405, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 545, + 437 + ], + "type": "text", + "content": "[33] Sizhe Li, Yapeng Tian, and Chenliang Xu. Space-time memory network for sounding object localization in videos. In British Machine Vision Conference, 2021. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 438, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 491 + ], + "type": "text", + "content": "[34] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In International Conference on Learning Representations, 2022. 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 493, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 535 + ], + "type": "text", + "content": "[35] Yan-Bo Lin, Hung-Yu Tseng, Hsin-Ying Lee, Yen-Yu Lin, and Ming-Hsuan Yang. Unsupervised sound localization via iterative contrastive learning. arXiv preprint arXiv:2104.00315, 2021. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "type": "text", + "content": "[36] Jinxiang Liu, Chen Ju, Weidi Xie, and Ya Zhang. Exploiting transformation invariance and equivariance for self-supervised sound localisation. In ACM International Conference on Multimedia, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 582, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 602 + ], + "type": "text", + "content": "[37] Josh H McDermott. The cocktail party problem. Current Biology, 19(22):R1024-R1027, 2009. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 604, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 645 + ], + "type": "text", + "content": "[38] Shentong Mo and Pedro Morgado. A closer look at weakly-supervised audio-visual source localization. Advances in Neural Information Processing Systems, NeurIPS, 2022. 1, 2, 3, 4, 5, 6, 7, 9" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "text", + "content": "[39] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In European Conference on Computer Vision, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[40] Pedro Morgado, Ishan Misra, and Nuno Vasconcelos. Robust audio-visual instance discrimination. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7786" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[41] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 151 + ], + "type": "text", + "content": "[42] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 287, + 185 + ], + "type": "text", + "content": "[43] Andrew Owens and Alexei A. Efros. Audio-visual scene analysis with self-supervised multisensory features. In European Conference on Computer Vision, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 287, + 252 + ], + "type": "text", + "content": "[44] Andrew Owens, Jiajun Wu, Josh H McDermott, William T Freeman, and Antonio Torralba. Ambient sound provides supervision for visual learning. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 801-816. Springer, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 254, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 254, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 254, + 287, + 297 + ], + "type": "text", + "content": "[45] Takashi Oya, Shohei Iwase, Ryota Natsume, Takahiro Itazuri, Shugo Yamaguchi, and Shigeo Morishima. Do we need sound for sound source localization? In Asia Conference on Computer Vision, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 299, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 299, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 49, + 299, + 287, + 331 + ], + "type": "text", + "content": "[46] Sooyoung Park, Arda Senocak, and Joon Son Chung. MarginNCE: Robust sound localization with a negative margin. In arXiv preprint arXiv:2211.01966, 2022. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 332, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 332, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 49, + 332, + 287, + 376 + ], + "type": "text", + "content": "[47] Rui Qian, Di Hu, Heinrich Dinkel, Mengyue Wu, Ning Xu, and Weiyao Lin. Multiple sound sources localization from coarse to fine. In European Conference on Computer Vision, 2020. 1, 2, 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 377, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 377, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 49, + 377, + 287, + 442 + ], + "type": "text", + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 445, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 445, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 49, + 445, + 287, + 498 + ], + "type": "text", + "content": "[49] Hyeonggon Ryu, Arda Senocak, In So Kweon, and Joon Son Chung. Hindi as a second language: Improving visually grounded speech with semantically similar samples. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2023. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 501, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 501, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 49, + 501, + 287, + 544 + ], + "type": "text", + "content": "[50] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound source in visual scenes. In IEEE Conference on Computer Vision and Pattern Recognition, 2018. 1, 2, 4, 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 545, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 545, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 49, + 545, + 287, + 599 + ], + "type": "text", + "content": "[51] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound sources in visual scenes: Analysis and applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(5):1605-1619, 2021. 1, 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 601, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 601, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 49, + 601, + 287, + 645 + ], + "type": "text", + "content": "[52] Arda Senocak, Hyeonggon Ryu, Junsik Kim, and In So Kweon. Learning sound localization better from semantically similar samples. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2022. 1, 2, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "text", + "content": "[53] Arda Senocak, Hyeonggon Ryu, Junsik Kim, and In So Kweon. Less can be more: Sound source localization with a classification model. In IEEE Winter Conf. on Applications of Computer Vision (WACV), 2022. 1, 2, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "text", + "content": "[54] Zengjie Song, Yuxi Wang, Junsong Fan, Tieniu Tan, and Zhaoxiang Zhang. Self-supervised predictive learning: A" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 654 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "negative-free method for sound source localization in visual scenes. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "text", + "content": "[55] Kim Sung-Bin, Arda Senocak, Hyunwoo Ha, Andrew Owens, and Tae-Hyun Oh. Sound to visual scene generation by audio-to-visual latent alignment. In IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 152, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 194 + ], + "type": "text", + "content": "[56] Yapeng Tian, Di Hu, and Chenliang Xu. Cyclic co-learning of sounding object visual grounding and sound separation. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 239 + ], + "type": "text", + "content": "[57] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In European Conference on Computer Vision, 2018. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 242, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 242, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 242, + 545, + 296 + ], + "type": "text", + "content": "[58] Efthymios Tzinis, Scott Wisdom, Aren Jansen, Shawn Hershey, Tal Remez, Daniel P. W. Ellis, and John R. Hershey. Into the wild with audioscope: Unsupervised audio-visual separation of on-screen sounds. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 297, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 341 + ], + "type": "text", + "content": "[59] Ho-Hsiang Wu, Prem Seetharaman, Kundan Kumar, and Juan Pablo Bello. Wav2CLIP: Learning robust audio representations from clip. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 342, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 385 + ], + "type": "text", + "content": "[60] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 387, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 387, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 387, + 545, + 442 + ], + "type": "text", + "content": "[61] Haohang Xu, Xiaopeng Zhang, Hao Li, Lingxi Xie, Wenrui Dai, Hongkai Xiong, and Qi Tian. Seed the views: Hierarchical semantic alignment for contrastive representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 443, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 443, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 443, + 545, + 475 + ], + "type": "text", + "content": "[62] Xudong Xu, Bo Dai, and Lin Dahua. Recursive visual sound separation using minus-plus net. In IEEE International Conference on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 477, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 477, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 308, + 477, + 545, + 531 + ], + "type": "text", + "content": "[63] Hanyu Xuan, Zhiliang Wu, Jian Yang, Yan Yan, and Xavier Alameda-Pineda. A proposal-based paradigm for self-supervised sound source localization in videos. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1029-1038, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 533, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 545, + 564 + ], + "type": "text", + "content": "[64] Hang Zhao, Chuang Gan, Wei-Chiu Ma, and Antonio Torralba. The sound of motions. In IEEE International Conference on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 567, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 609 + ], + "type": "text", + "content": "[65] Hang Zhou, Xudong Xu, Lin Dahua, Xiaogang Wang, and Ziwei Liu. Sep-stereo: Visually guided stereophonic audio generation by associating source separation. In European Conference on Computer Vision, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 611, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 611, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 308, + 611, + 545, + 654 + ], + "type": "text", + "content": "[66] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In European Conference on Computer Vision, 2022. 4, 7" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7787" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Source-free Depth for Object Pop-out/1ac228ed-431b-40e3-b444-e5a574a99abb_content_list.json b/2023/Source-free Depth for Object Pop-out/1ac228ed-431b-40e3-b444-e5a574a99abb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1c0e85c450e7f70b09494d6d5e00ae3b5e589ded --- /dev/null +++ b/2023/Source-free Depth for Object Pop-out/1ac228ed-431b-40e3-b444-e5a574a99abb_content_list.json @@ -0,0 +1,1663 @@ +[ + { + "type": "text", + "text": "Source-free Depth for Object Pop-out", + "text_level": 1, + "bbox": [ + 292, + 130, + 676, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zongwei Wu $^{1,2,3}$ Danda Pani Paudel $^{1,4}$ Deng-Ping Fan $^{1*}$ Jingjing Wang $^{5}$ Shuo Wang $^{1}$ Cédric Demonceaux $^{2,6}$ Radu Timofte $^{3}$ Luc Van Gool $^{1,4}$", + "bbox": [ + 109, + 178, + 854, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ CVL, ETH Zurich $^{2}$ University of Burgundy, CNRS, ICB $^{3}$ Computer Vision Lab, CAIDAS & IFI, University of Wurzburg $^{4}$ INSAIT, Sofia University $^{5}$ AUST $^{6}$ University of Lorraine, CNRS, Inria, Loria", + "bbox": [ + 101, + 217, + 869, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Depth cues are known to be useful for visual perception. However, direct measurement of depth is often impracticable. Fortunately, though, modern learning-based methods offer promising depth maps by inference in the wild. In this work, we adapt such depth inference models for object segmentation using the objects' \"pop-out\" prior in 3D. The \"pop-out\" is a simple composition prior that assumes objects reside on the background surface. Such compositional prior allows us to reason about objects in the 3D space. More specifically, we adapt the inferred depth maps such that objects can be localized using only 3D information. Such separation, however, requires knowledge about contact surface which we learn using the weak supervision of the segmentation mask. Our intermediate representation of contact surface, and thereby reasoning about objects purely in 3D, allows us to better transfer the depth knowledge into semantics. The proposed adaptation method uses only the depth model without needing the source data used for training, making the learning process efficient and practical. Our experiments on eight datasets of two challenging tasks, namely salient object detection and camouflaged object detection, consistently demonstrate the benefit of our method in terms of both performance and generalizability. The source code is publicly available at https://github.com/Zongwei97/PopNet.", + "bbox": [ + 75, + 319, + 473, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 727, + 209, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The 3D knowledge of the scene is long known to be complementary to the task of visual perception [12, 16, 56, 67, 91]. Often in practice, though, visual perception needs to be carried out using only 2D images. Given multiple images, 3D geometry may be recovered using the structure-from-motion techniques [26, 41, 55, 86]. Such inversion, however, is not compatible when only a single image is available. Under such circumstances, image inversion", + "bbox": [ + 75, + 753, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bc00a305be8a8a03db526ed5dc79a002452bdf98338452b280083f379e51b4fb.jpg", + "image_caption": [ + "Figure 1. Depth to semantics conversion using the object pop-out prior. For input RGB and source-free depth pair, we learn contact surface. The obtained contact surface is then used to separate objects and backgrounds to derive pseudo semantics for supervision." + ], + "image_footnote": [], + "bbox": [ + 501, + 290, + 890, + 428 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to depth map is usually done using learning-based methods [13,45,50,51,65], which have shown unparalleled success in the recent years. Unfortunately, the learning-based methods may not offer high-quality depth maps due to the generalization deficiency across domains [68, 73].", + "bbox": [ + 496, + 506, + 892, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite poor generalization, the knowledge gained in one domain is shown to be useful in other close-by domains. This utility is harnessed by performing the so-called domain adaptation (DA) [2,3,6,53,61,89]. In fact, it has been shown recently that DA methods can efficiently transfer knowledge using only the prediction models, i.e., without requiring access to the data where the model is trained – also known as the source-free domain adaptation (SDA) [25,28,39,71,77]. The SDA methods are of gripping interest due to their efficiency and privacy promises.", + "bbox": [ + 496, + 582, + 893, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The most existing SDA methods make one or both of these implicit assumptions: (a) a similar (as that of the source) supervising task at the target [28, 75, 77]; (b) task of discrete (and known) label space [27, 29, 36, 71]. The former assumption not only makes the source and target domains easier to compare but also potentially keeps the two domains closer. The latter assumption allows performing SDA by self-training where the discrete labels facilitate reasoning about the model's confidence. The self-training is then performed by boosting the confidence at the target for some picked reliable examples.", + "bbox": [ + 496, + 734, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author: Deng-Ping Fan (dengpfan@gmail.com)", + "bbox": [ + 94, + 886, + 437, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1032", + "bbox": [ + 483, + 924, + 514, + 936 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f208f75faa12ac29e7bda8c0fdd583158bc75bf9757271f90a8896e8d21f815a.jpg", + "image_caption": [ + "Cross-Task Cross-Dataset Analysis with Depth Pop-out Prior", + "Figure 2. The performance gained in F-measure using our method over the established baselines. Our method $(\\triangle, \\square)$ significantly improves baselines on 8 datasets and 2 tasks (each task on 4 datasets- SOD: left two; COD: right two). We also compare 24 methods $(\\bullet)$ , where our method offers state-of-the-art results despite their task specialization. Note that all methods are connected using lines to illustrate their performance fluctuations across datasets. Please, refer Tables 1 & 2 and Section 4.3 for more details and discussions." + ], + "image_footnote": [], + "bbox": [ + 81, + 107, + 893, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we aim for source-free transfer of depth knowledge for object detection. Such transfer is desired to assist in locating the object by depth cues and to exploit the depth knowledge despite the domain gap. The addressed problem setting differs from the standard SDA in terms of (a) the source and target tasks' difference; (b) and the continuous label space of the depth. These differences (with standard SDA) make our task at hand very challenging, which is addressed for the first time in this paper, up to our knowledge. To address such a challenging problem, we rely on the \"pop-out\" prior, which allows us to reason about the object's location directly in the 3D space. The \"pop-out\" is a simple composition prior that assumes objects reside on the background surface. A graphical illustration of the used \"pop-out\" prior, using the results obtained by our method, is given in Figure 1.", + "bbox": [ + 75, + 303, + 468, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The pop-out prior for image composition was successfully used by Kang et al. in [24]. An early work of Treisman has provided an in-depth study of such prior in [62]. In this work, we rely on the same compositional foundation of these works and exploit the pop-out before transferring the depth of knowledge across domains. Although our motivation comes from these early works, our experimental setup largely differs from theirs. We differ in terms of not only depth knowledge transfer across domains (without source data) but also in target supervision using only semantics.", + "bbox": [ + 75, + 549, + 468, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed method exploits the source-free depth to map it into a space where objects in depth stand out better against the background. This mapping is used for object and background separation using a learned contact surface between them. Such separation allows us to derive the semantic masks which can be directly compared against the ground truth for supervision. Using this supervision at the target, we can minimize the domain and task gap between the source and target. The overall framework of our method that performs cross-task cross-domain knowledge transfer by using the intermediate representation of the pop-out space is shown in Figure 3. As can be seen, we first leverage an object popping network to encourage the object", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to jump out from the source-free depth. Then, we introduce another network i.e., the segmentation with contact surface, to localize the object and predict the contact surface. These learning modules are jointly trained in an end-to-end manner, transferring the source-free depths into intermediate representations which are adapted to the target task, i.e., object detection. To evaluate the proposed method, we conducted exhaustive experiments on eight datasets of two challenging tasks, namely salient object detection and camouflaged object detection. In both tasks, our method significantly improves the established baselines and offers state-of-the-art results at the same time, whose overview can be seen in Figure 2. The major contributions are as follows:", + "bbox": [ + 496, + 303, + 890, + 500 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Our problem of transferring source-free depth knowledge across domains and tasks is practical and novel.", + "- Our method relies on our object pop-out prior for visual understanding, which is simple and effective.", + "- Results of our method in two different tasks are significantly better than the baselines and existing models." + ], + "bbox": [ + 517, + 502, + 890, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 603, + 648, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Source-free Adaptation: Knowledge transfer by domain adaptation without access to the source data has recently gathered vast interest [1,25,28,39,71,77], due to the privacy, practicality, and efficiency reasons. We observe accessing the source data for transferring the depth knowledge learned by the off-the-shelf knowledge models [13, 51] is particularly impractical due to their multi-stage training on various datasets. The existing source-free adaptation methods either use generative [28, 29, 34], pseudo-label [25, 37, 64], or other customized [54, 76] approaches. In this work, we use the pseudo-label-based approach. However, using existing methods is not straightforward in our setting due to the tasks' difference between source and target. With our pop-out technique that provides the pseudo semantics, the source-free depth is better transferred across tasks.", + "bbox": [ + 496, + 628, + 890, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Salient Object Detection (SOD): Saliency detection aims to detect and segment the most prominent region within an image that visually attracts human attention [11]. A", + "bbox": [ + 496, + 854, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1033", + "bbox": [ + 483, + 924, + 514, + 936 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/451fab1cda9c67f92d3f0a55b31ef68d5b893ceeda5b7501ee15edbedad854ef.jpg", + "image_caption": [ + "Figure 3. Our proposed framework, termed PopNet, is composed of a source-free network, an object popping network, a segmentation with contact surface, and an object separation module. The source-free depth network generates pseudo-depth in an off-the-shelf manner (Section 3.1). The object popping network converts the source-free depth into the popped-out depth of objects, bridging the cross-domain and cross-task gaps (Section 3.2). The segmentation network uses this depth to estimate the object's mask and contact surface (Section 3.3). The object separation module then converts popped-out depth to the second mask of objects using the contact surface (Section 3.4). We compare both semantic masks against ground truth to supervise the whole pipeline in an end-to-end manner." + ], + "image_footnote": [], + "bbox": [ + 117, + 87, + 851, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "number of works have shown that saliency can be an auxiliary step for different vision tasks such as object tracking [93], object detection [58], etc. Conventional saliency works are unimodal, i.e., they only require RGB images as input. In generic and common settings, RGB-based models [38, 70, 84] have already achieved very promising results. More recently, several works [10,20,49,66,68,88,92] exploit depth maps as additional clues to the 3D geometry since the depth can provide more truthful information on the object boundary as well as the scale awareness. These 3D features further improve the detection accuracy and performance in challenging scenarios [22, 33, 69, 80].", + "bbox": [ + 75, + 381, + 468, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Camouflaged Object Detection (COD): Camouflage detection aims to find the preying object within an image. For computer vision society, primary works [9,32,47] often compare COD with SOD. A number of works [8,9,74,79] have shown that simply extending saliency models [70] on COD will lead to undesired results, which is mainly caused by the nature of target object, i.e., concealed or prominent. Hence, to constrain the attention on the concealed objects, several works come up with different perceptual systems that mimic human behavior vis-a-vis camouflaged objects, such as three-stage localize-segment-rank strategy [42]; iterative refinement [21], which is similar to repeatedly looking on the images; zooming into possible regions [47, 59]. Others [18, 40, 60, 82, 83, 90, 94, 95] deeply explore the texture difference with the help of the gradient [18], frequency [90], edges [60, 94], and probability [32, 74].", + "bbox": [ + 75, + 565, + 468, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The latest psychological studies [5, 7] have shown that human perception can naturally benefit from the depth cues to understand the scene: (A) the smooth variation within the object can contribute to alleviating the fake edges and preserving the object structure; (B) the depth discontinuity on the object boundary can make the segmentation easier. In", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "spired by these observations, we aim to explore the source-free depth for both SOD and COD tasks. To tackle the domain gap for source-free depth maps, we propose to jointly finetune the source-free depth together with the semantic network in an end-to-end manner, with both self-supervised loss and weak semantic supervision.", + "bbox": [ + 496, + 381, + 890, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed PopNet", + "text_level": 1, + "bbox": [ + 498, + 489, + 669, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an input RGB image $I$ with size $I \\in \\mathbb{R}^{3 \\times H \\times W}$ , where $H$ and $W$ are the spatial resolutions, i.e., height and width, of the image, our objective is to predict the semantic mask $\\tilde{S} \\in \\mathbb{R}^{H \\times W}$ for object detection. As shown in Figure 3, the input image $I$ is firstly fed into a frozen-weight depth network to generate the source-free depth $D_{sf} \\in \\mathbb{R}^{H \\times W}$ (Section 3.1). Then the mimicked multi-modal images are fed together into the depth popping network to compute the intermediate popped-out depth $D_{po} \\in \\mathbb{R}^{H \\times W}$ (Section 3.2). This intermediate representation, as well as the RGB image $I$ , is later processed by the segmentation network and transformed into a contact surface $D_{c} \\in \\mathbb{R}^{H \\times W}$ and a semantics prediction $\\tilde{S} \\in \\mathbb{R}^{H \\times W}$ (Section 3.3). On the one hand, the semantics prediction is directly supervised by the ground truth mask $GT$ , denoted as $G$ , which is similar to conventional segmentation supervision. On the other hand, we further explore the contact surface to pop the object out of the background by means of our object separation module (Section 3.4). This transfers the geometric cues into pseudo semantics and leads to another level of supervision.", + "bbox": [ + 496, + 513, + 890, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Source-free Depth Network", + "text_level": 1, + "bbox": [ + 500, + 830, + 746, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In a practical setting, the GT depth is not always available. Therefore, we generate the source-free depth $D_{sf}$ in an off-the-shelf manner to mimic the multi-modal input.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1034", + "bbox": [ + 483, + 924, + 514, + 936 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We choose the state-of-the-art DPT model [50] with frozen weights as our source-free depth network, which offers us promising depth at the target. This choice is made upon its generalization capability as suggested in [51]. To obtain the depth map with the highest quality possible, by enhancing local details, we apply the boosting method [45] together with DPT. Despite the plausible results achieved by learning-based methods, the obtained source-free depth does not always offer high-quality geometric cues due to the domain gap. Therefore, we leverage the geometric and semantic priors to jointly finetune the source-free depth.", + "bbox": [ + 75, + 90, + 472, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Object Popping Network", + "text_level": 1, + "bbox": [ + 76, + 268, + 305, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Network Architecture: We build a depth popping network to refine/smooth the source-free depth. The popping network follows the encoder-decoder design with skip connection as shown in Figure 4. In our case, we simply concatenate RGB and source-free depth at the input side to form a 4-channel input and feed them into the popping network. The encoder extracts semantic cues and generates five-scale outputs. Our decoder is composed of Conv2D, BN, ReLU, and upsampling layers. Following U-Net [52], we build a skip connection through simple addition.", + "bbox": [ + 75, + 291, + 468, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Structure Preserving: To supervise our popping network, we first guide the depth refinement with the help of generated pseudo-depth. We only leverage structural similarity since we aim to detect, preserve, and extract object structure from the intermediate representation. We use the following SSIM loss [13] for structural similarity.", + "bbox": [ + 75, + 444, + 468, + 535 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d e p} = S S I M \\left(D _ {p o}, D _ {s f}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 547, + 468, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Local Depth Smoothing: In addition to the pseudo-depth supervision, we propose two losses to constrain the depth also by semantics. We assume that the objects' structure should be distinguishable from the background, i.e., it should be smooth within the object region and sharp on the bounding pixels. Hence, we propose to leverage the weak semantic cues together with the geometric priors. Specifically, we first introduce a local smoothness loss. The locality is defined by the ground truth semantics $G$ . Technically, we mask out background pixels with element-wise multiplication to suppress the inactive area through $D_{obj} = D_{po} \\otimes G$ . Let $\\nabla_x$ and $\\nabla_y$ be the Sobel operations. Then, our local loss $\\mathcal{L}_{loc}$ is expressed as:", + "bbox": [ + 75, + 577, + 468, + 773 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c} \\overrightarrow {n (p)} = (- \\nabla_ {x} (D _ {o b j} (p)); - \\nabla_ {y} (D _ {o b j} (p)); 1); \\\\ \\mathcal {L} _ {l o c} = \\sum_ {p} \\sum_ {q \\in \\mathcal {N} (p)} 1 - \\operatorname {c o s i n e} (\\overrightarrow {n (p)}; \\overrightarrow {n (q)}), \\end{array} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 784, + 468, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\overrightarrow{n}$ stands for the normal, $p$ denotes the pixels within the object region, $\\mathcal{N}(p)$ is the neighboring pixels, and cosine is the cosine similarity between two vectors. This way, our", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9e4cfb368af124dd46d4911338806fb8feac354156afccce5a7ffabc2555d4bb.jpg", + "image_caption": [ + "Figure 4. Our object popping network maps RGB-D inputs into popped-out depths. This network is supervised using a combination of structure preserving, local depth smoothing, and depth edge sharpening losses $\\mathcal{L}_{dep}$ , $\\mathcal{L}_{loc}$ , and $\\mathcal{L}_{wtv}$ , respectively." + ], + "image_footnote": [], + "bbox": [ + 501, + 87, + 890, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "local loss only works on the object area, making the object structure consistent within the target region. Applying local smoothness loss reduces the depth noise at the object level.", + "bbox": [ + 498, + 311, + 890, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Depth Edge Sharpening: In addition to the local smoothness, we also use edge sharpening. The edge sharpening loss is formulated as a weighted total variation. For this, we first compute the edge-aware weight $w(p)$ at any pixel $p$ as,", + "bbox": [ + 498, + 358, + 890, + 420 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw (p) = \\left\\{ \\begin{array}{l l} w _ {0}, & \\text {i f} \\nabla_ {x} (G (p)) ^ {2} + \\nabla_ {y} (G (p)) ^ {2} \\neq 0, \\\\ w _ {0} + \\gamma , & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 440, + 890, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_0$ is a pre-defined non-zero weight and $\\gamma$ is an additional weight for boundary pixels. In our setting, we choose $w_0$ as the normalized (by the image size) count of the boundary pixels, and we set $\\gamma = 0.5$ . We adopt the square form such that the large gradients play more important roles. Our weighted total-variation loss is given by,", + "bbox": [ + 498, + 489, + 890, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {w T v} = \\sum_ {p} \\sum_ {q \\in \\mathcal {N} (p)} w (p) \\cdot | | D _ {p o} (p) - D _ {p o} (q) | | _ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 588, + 890, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our weighted total-variance loss differs from the conventional edge loss, due to our weighting function. More specifically, our weighting function relies on semantic boundaries, unlike the commonly used image gradients [14, 17]. Our motivation for using semantic boundaries instead of image gradients comes from our interest in performing object detection under challenging conditions, such as camouflaged objects. In such cases, image gradients may result in misleading weights. At first glance, our loss function seems to be similar to semantic-guided depth estimation methods [4, 35, 72]. However, [72] uses GT depths, and [4, 35] use multi-frames for supervision. Unfortunately, such supervision is not possible in our setup. Note that we exploit single-view source-free depth while only using semantic ground truth for supervision. Furthermore, we transfer the source-free depths' knowledge despite their domain gap. Existing regularizations between depth and semantics are already exploited in our method, which is shown to be", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1035", + "bbox": [ + 483, + 924, + 514, + 936 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "complimentary to our pop-out prior. We argue that our network architecture that exploits the pop-out prior is not trivial while simultaneously being generic and easy to use.", + "bbox": [ + 76, + 90, + 468, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The total objective function $\\mathcal{L}_{pop}$ to supervise our object popping network is given by,", + "bbox": [ + 76, + 136, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p o p} = \\mathcal {L} _ {d e p} + \\lambda_ {1} \\cdot \\mathcal {L} _ {l o c} + \\lambda_ {2} \\cdot \\mathcal {L} _ {w t v}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 178, + 468, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ are the hyperparameters.", + "bbox": [ + 76, + 205, + 354, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Segmentation with Contact Surface", + "text_level": 1, + "bbox": [ + 76, + 229, + 385, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The smoothness and edge losses encourage homogenizing the object structure, making it noticeable from the background. We now aim to further enlarge the object-background distance to make the object structure jump out. Specifically, we use an RGB-D segmentation network as shown in Figure 5. The main component of our segmentation network is a three-stream RGB-D network with some fusion design. In our setting, we choose [92] as our baseline since it is one of the SOTA RGB-D methods for saliency detection. We add a surface head to learn the depth of the contact surface $D_{c}$ , which has the same resolution as the input depth $D_{po}$ . Our surface head is composed of ConvLayer (Conv2D, BN, ReLU) and a Conv2D, which first decodes the feature maps and then transfers them into a 1-D map.", + "bbox": [ + 75, + 253, + 470, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Object Separation", + "text_level": 1, + "bbox": [ + 76, + 474, + 253, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using the previously discussed contact surface, in this section, we aim to separate the object from its background. At this point, we make an assumption that pixels in front of the contact surface belong to objects. The remaining pixels belong to the background. This assumption allows us to explicitly transfer the 3D knowledge into 2D semantics. Let the predicted depth of the contact surface be $D_{c} \\in \\mathbb{R}^{H \\times W}$ . We obtain the pseudo semantics $S_{s}$ , using the popped-out depth $D_{po}$ and surface's depth $D_{c}$ , as:", + "bbox": [ + 75, + 497, + 468, + 633 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {s} = \\operatorname {s i g m o i d} \\left(\\sigma \\cdot \\left(D _ {p o} - D _ {c}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 645, + 468, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma$ is a scalar value that controls the slope of the sigmoid function. In our experiments, we use $\\sigma = 10$ to perform soft-thresholding, mimicking the desired hard one for binary outputs. Such soft thresholding facilitates the gradient back-propagation required for training. Finally, we minimize the gap between the pseudo semantics $S_{s}$ and the GT semantics $G$ with binary cross-entropy (BCE):", + "bbox": [ + 75, + 672, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s e p} = B C E \\left(S _ {s}, G\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 790, + 468, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Overall Loss Function", + "text_level": 1, + "bbox": [ + 76, + 816, + 284, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Both of our trainable modules, i.e., object popping and segmentation networks, of our framework (Figure 3) are trained in an end-to-end manner. Therefore, the overall loss function consists of three parts: our depth popping loss", + "bbox": [ + 76, + 839, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ee2d166a5549719ca4c61146184fb0125def3e986be754f9511e618701074cb3.jpg", + "image_caption": [ + "Figure 5. Our segmentation network uses a basic RGB-D three-stream network [92]. In addition to the conventional semantic head, we learn to predict pixel-wise contact surface. The contact surface is later used to transfer depth knowledge to semantics." + ], + "image_footnote": [], + "bbox": [ + 501, + 87, + 890, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{pop}$ , our object separation loss $\\mathcal{L}_{sep}$ , and the conventional semantic loss $\\mathcal{L}_{sem}$ from the RGB-D baseline network. The total loss $\\mathcal{L}_{total}$ used for training is given by,", + "bbox": [ + 498, + 304, + 890, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p o p}} + \\alpha_ {1} \\cdot \\mathcal {L} _ {\\text {s e p}} + \\alpha_ {2} \\cdot \\mathcal {L} _ {\\text {s e m}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 359, + 890, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{1}$ and $\\alpha_{2}$ are hyperparameters.", + "bbox": [ + 498, + 387, + 754, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remarks: Our losses work in a complementary manner. The $\\mathcal{L}_{pop}$ plays the same role as a smooth filter as in image smoothing. It removes the noisy depth response due to the domain gap while preserving the object structure with the help of a weak semantic label. Hence, the smoothed background becomes less informative while the object region becomes uniform, making it easily detectable. These functionalities contribute to transferring the source-free depth into popped-out depth, as desired. Such a process brings objects above the background surface, despite their distance from the camera and other distracting surfaces. The $\\mathcal{L}_{sep}$ , on the other hand, fully benefits from the \"pop-out\" prior to segment the object from the background. With the help of the learned contact surface, this loss enlarges the foreground-background distance by pulling them in opposite directions. Such pulling results in a binary-like mask, which effectively bridges the gap between geometric knowledge and semantics. Finally, both depth-transferred and learned semantics are compared to the ground truths for supervised training.", + "bbox": [ + 496, + 402, + 890, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 500, + 715, + 586, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 741, + 689, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset Preparation: To better illustrate the generalizability of our approach, we evaluate the effectiveness of our approach on both SOD and COD benchmarks. We choose four widely used RGB-D SOD datasets, i.e., NLPR [48], NJUK [23], STERE [46], and SIP [10], as well as four COD datasets, i.e., CAMO [30], CHAMELEON [57], COD10K [9], and NC4K [42]. For SOD datasets, we conduct experiments with both GT depth and source-free depth. We follow the conventional learning protocol [19, 69, 92]", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1036", + "bbox": [ + 483, + 924, + 514, + 935 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/944bd52b4d893475fbbfec3ef17c0d68f0f818e186a4cdd61d65a30363d949d6.jpg", + "table_caption": [ + "Table 1. Quantitative comparison on RGB-D SOD datasets. $\\uparrow (\\downarrow)$ denotes that the higher (lower) is better. We use the Mean Absolute Error $(M)$ , max F-measure $(F_{m})$ , S-measure $(S_{m})$ , and max E-measure $(E_{m})$ as evaluation metrics. G.D. stands for GT Depth. Bold denotes the best performance." + ], + "table_footnote": [], + "table_body": "
G.D.Public.Dataset MetricNLPR [48]NJUK [23]STERE [46]SIP [10]
M↓Fm↑Sm↑Em↑M↓Fm↑Sm↑Em↑M↓Fm↑Sm↑Em↑M↓Fm↑Sm↑Em↑
Performance of RGB-D Models Trained with Source-free Depth
XMM21 [85]DFM-Net.027.909.914.944.046.903.895.927.042.906.903.934.067.873.850.891
XTIP22 [63]DCMF.027.915.921.943.044.908.903.929.041.909.907.931.067.873.853.893
XCVPR22 [21]SegMAR.024.923.920.952.036.921.909.941.037.916.907.936.052.893.872.914
XCVPR22 [47]ZoomNet.023.916.919.944.037.926.914.940.037.918.909.938.054.891.868.909
XOursPopNet.022.925.926.956.031.931.920.949.032.922.916.947.046.911.885.926
Performance of RGB-D Models Trained with GT Depth
✓TIP21 [87]BIANet.032.888.900.930.056.878.867.898.048.898.895.918.091.816.802.847
✓TIP21 [33]HAINet.024.920.924.956.037.924.911.940.040.917.907.938.052.907.879.917
✓TNNLS21 [10]D3Net.029.904.911.942.046.909.899.927.044.902.906.925.063.880.860.897
✓ECCV22 [31]SPSN.023.917.923.956.032.927.918.949.035.909.906.941.043.910.891.932
✓OursPopNet.019.927.932.963.030.936.924.952.033.924.917.947.040.923.897.937
", + "bbox": [ + 76, + 128, + 890, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and use 700 images from NLPR and 1,485 images from NJUK for training. The rest are used for testing. For the unimodal COD dataset, we compare with both RGB COD models and RGB-D SOD models retrained on the COD datasets with the same source-free depth $D_{sf}$ . We follow the conventional training/testing protocol [8, 9, 21, 42, 47] and use 3,040 images from COD10K and 1,000 images from CAMO for training. The rest are used for testing.", + "bbox": [ + 75, + 310, + 468, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics: We evaluate the performance with four generally-recognized metrics: Mean Absolute Error $(M)$ , max F-measure $(F_{m})$ , S-measure $(S_{m})$ , and max E-measure $(E_{m})$ . All the object segmentation masks are trained or downloaded from the official resources. To make a fair comparison, we evaluate the prediction semantics with the standardized evaluation protocol as [92].", + "bbox": [ + 75, + 431, + 468, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details: Our model is implemented based on Pytorch with a V100 GPU. We use the Adam algorithm as an optimizer. The learning rate is initialized to $1e - 4$ and is further divided by 10 every 60 epochs. We set the input resolution to $352 \\times 352$ resolution for RGB and depth. A detailed comparison with higher resolution can be found in Table 3. During training, conventional data augmentation such as random flipping, rotating, and border clipping are adopted. The training takes around 6 hours for RGB-SOD tasks and 12 hours for COD tasks for 100 epochs.", + "bbox": [ + 75, + 537, + 470, + 689 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons", + "text_level": 1, + "bbox": [ + 76, + 696, + 217, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with RGB-D SOD Models: We present in Table 1 the performance on SOD benchmarks with our source-free depth or with GT depth. It can be seen that our model with source-free depth achieves very competitive performance compared to many RGB-D models with GT depths. Our method with GT depth also outperforms the SOTA counterparts. The qualitative comparison can be found in Fig. 6. Note that we also retrain the SOTA unimodal COD models SegMAR [21] and ZoomNet [47] on the SOD dataset only with RGB images. We show that our model with source-free depth outperforms these counterparts, showing that our method can better generalize across", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a7c7c99ae722615d051e65dcf5584a4ec196b3e6b178e04ae235ac4098b11e60.jpg", + "image_caption": [ + "Figure 6. Qualitative Comparison with GT Depth. Our method outperforms all counterparts while dealing with multi-objects, large depth variation, and visually-mixed foreground-background." + ], + "image_footnote": [], + "bbox": [ + 501, + 309, + 890, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "different tasks with favorably better performance.", + "bbox": [ + 498, + 505, + 826, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with COD Models: We present in Table 2 the performance of the most competitive SOTA methods, including task-specific COD models as well as retrained RGB-D SOD models with source-free depth. For a fair comparison, we retrain all RGB-D methods, SegMAR [21], and ZoomNet [47] in an end-to-end manner on the same resolution images as ours. Some RGB-only methods perform even better than many RGB-D methods, mainly due to their COD task-specific designs and the lack of ground-truth depths required for RGB-D methods. It is important to note that competing RGB-only COD methods do not perform favorably on SOD tasks. Please, refer to Table 1 & 3 to observe their poor cross-tasks generalization.", + "bbox": [ + 496, + 521, + 890, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Towards Higher Resolution: Previous studies have shown that the image resolution may influence the model performance [43, 47, 73, 83]. For example, the current SOTA COD method ZoomNet [47] is with main scale $384^2$ and implies the highest resolution of $(384 \\times 1.5)^2 = 576^2$ , as it operates on $0.5 \\times 1 \\times$ and $1.5 \\times$ scales. To make a fair comparison, we retrain the model with the same resolution $(352^2$ or $512^2)$ as ours. We show in Table3-ZoomNet* that the results deteriorate as expected. Compared to these counterparts, our method offers a good trade-off between accuracy and efficiency. More comparisons on SOD and COD benchmarks can be found in the supplementary material.", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1037", + "bbox": [ + 483, + 924, + 514, + 936 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ab569332d5085f97f013e7fb9c93c3e93a24e06a2879d19a2e76034e71533c4e.jpg", + "table_caption": [ + "Table 2. Quantitative comparison on COD datasets. Pseudo stands for source-free depth used for RGB-D methods." + ], + "table_footnote": [], + "table_body": "
Pseudo Public.Dataset MetricCAMO [30]CHAMELEON [57]COD10K [9]NC4K [42]
M ↓Fm ↑Sm ↑Em ↑M ↓Fm ↑Sm ↑Em ↑M ↓Fm ↑Sm ↑Em ↑M ↓Fm ↑Sm ↑Em ↑
Performance of RGB COD Models
X CVPR20 [9]SINet.099.762.751.790.044.845.868.908.051.708.771.832.058.804.808.873
X CVPR21 [42]SLSR.080.791.787.843.030.866.889.938.037.756.804.854.048.836.839.898
X CVPR21 [78]MGL-R.088.791.775.820.031.868.893.932.035.767.813.874.053.828.832.876
X CVPR21 [44]PFNet.085.793.782.845.033.859.882.927.040.747.800.880.053.820.829.891
X CVPR21 [32]UJSC.072.812.800.861.030.874.891.948.035.761.808.886.047.838.841.900
X IJCAI21 [59]C2FNet.079.802.796.856.032.871.888.936.036.764.813.894.049.831.838.898
X ICCV21 [74]UGTR.086.800.783.829.031.862.887.926.036.769.816.873.052.831.839.884
X CVPR22 [21]SegMAR.080.799.794.857.032.871.887.935.039.750.799.876.050.828.836.893
X CVPR22 [47]ZoomNet.074.818.801.858.033.829.859.915.034.771.808.872.045.841.843.893
Performance of RGB-D Models Retrained with Source-free Depth
✓ MM21 [80]CDINet.100.638.732.766.036.787.879.903.044.610.778.821.067.697.793.830
✓ CVPR21 [19]DCF.089.724.749.834.037.821.850.923.040.685.766.864.061.765.791.878
✓ ICCV21 [81]CMINet.087.798.782.827.032.881.891.930.039.768.811.868.053.832.839.888
✓ ICCV21 [92]SPNet.083.807.783.831.033.872.888.930.037.776.808.869.054.828.825.874
✓ TIP22 [63]DCMF.115.737.728.757.059.807.830.853.063.679.748.776.077.782.794.820
✓ ECCV22 [31]SPSN.084.782.773.829.032.866.887.932.042.727.789.854.059.803.813.867
✓ OursPopNet.073.821.806.869.022.893.910.962.031.789.827.897.043.852.852.908
", + "bbox": [ + 76, + 99, + 893, + 347 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/313a4908b5d16c66ace5c12b698a042f87c9b8097dd66ad77b2c96d0d0ad1135.jpg", + "image_caption": [ + "Image" + ], + "image_footnote": [], + "bbox": [ + 78, + 364, + 153, + 476 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc2c24b00d10f08b7bdd7dd8f2964a1fd652a889b097e89c340d9031beb183a4.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 153, + 364, + 225, + 476 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/02e0ff4e078e32ba845abdbfe3c8f40e834636ad230a059bf03c238dd086bd8a.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 225, + 364, + 300, + 476 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c2f5971a18abf0bc7be5d0538c3d65b89663431c1cc977dfa8c96e7b47d0feb.jpg", + "image_caption": [ + "ZoomNet", + "Figure 7. Qualitative comparison. Our method can better preserve the object structure compared to other counterparts, especially while dealing with objects with occlusion (two first rows). Our method performs favorably with multiple objects (last row). Better to zoom in." + ], + "image_footnote": [], + "bbox": [ + 300, + 364, + 372, + 476 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/151245e316fa0105cc34499f11c453927f83187910a3a9ab87897251828f94a6.jpg", + "image_caption": [ + "UGTR" + ], + "image_footnote": [], + "bbox": [ + 374, + 364, + 447, + 476 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f9bc22628844ea3d45b3b01616fdfda185a8ef4ad372170fcd5231926718472d.jpg", + "image_caption": [ + "UJSC" + ], + "image_footnote": [], + "bbox": [ + 449, + 364, + 522, + 476 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b82db1fbe06a43661851b35db1db411cd5eef8b20f4dbe670b0f0677c5528f21.jpg", + "image_caption": [ + "C2FNet" + ], + "image_footnote": [], + "bbox": [ + 522, + 364, + 596, + 476 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ce3b530f65c1510010c26bc45654e97d0cb3cf9a76f72ed608331043141d7995.jpg", + "image_caption": [ + "PFNet", + "MGL-R" + ], + "image_footnote": [], + "bbox": [ + 596, + 364, + 741, + 476 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/24fff8867aec7c6c7d44ac4e9c274c5984dd0ea2789e2b6032edbedd0d2b9efd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 364, + 743, + 476 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a7c88e08c35af38c8a95c03fc87ee3a3fcf879dd6fffb226d4dd8f82e51a2e84.jpg", + "image_caption": [ + "SLSR", + "SINet" + ], + "image_footnote": [], + "bbox": [ + 743, + 364, + 888, + 476 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 834, + 364, + 888, + 476 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dbd30ca728e0bdc57c014372bf7486b4640f242b948a2ee4ae7c2929c510bd46.jpg", + "table_caption": [ + "Table 3. End-to-end comparison with different resolutions on SOD benchmarks. Our method with source-free depth generalizes significantly better compared to SOTA COD models." + ], + "table_footnote": [], + "table_body": "
ModelSizeFlops (G)NJUK [23]SIP [10]
M↓Fm↑Sm↑Em↑M↓Fm↑Sm↑Em↑
SegMAR [21]352267.3.036.921.909.941.052.893.872.914
ZoomNet [47]3522167.8.037.926.914.940.054.891.868.909
Ours3522228.8.031.931.920.949.046.911.885.926
SegMAR [21]5122142.4.035.927.914.943.050.899.878.917
ZoomNet [47]5122353.4.036.926.915.942.052.895.873.910
Ours5122484.0.031.933.922.951.044.911.890.927
", + "bbox": [ + 80, + 570, + 465, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparisons: Figure 7 presents the output of our network on challenging cases. It can be seen while dealing with objects occluded by thin objects ( $1^{st} - 4^{th}$ rows), our method can accurately reason about the segmentation masks closer to the GT. We also achieve better performance while dealing with multiple objects (last row). More discussions on multiple objects can be found in Table 6.", + "bbox": [ + 75, + 680, + 468, + 785 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 800, + 230, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Loss: In this section, we conduct experiments on analyzing the effectiveness of the proposed losses. The quantitative results of different loss combinations are provided in Table 4. It can be seen that each proposed loss behaves properly, i.e., improving the performance compared to the baseline. More", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e7b360863ad44dcf4cddebac3906002fece7ccf2bc4fb9209df448c668c90a9d.jpg", + "table_caption": [ + "Table 4. Ablation study on the proposed losses." + ], + "table_footnote": [], + "table_body": "
\\( \\mathcal{L}_{dep} \\)\\( \\mathcal{L}_{loc} \\)\\( \\mathcal{L}_{wtv} \\)\\( \\mathcal{L}_{sep} \\)SIP [10]NC4K [42]
\\( M\\downarrow \\)\\( F_m\\uparrow \\)\\( S_m\\uparrow \\)\\( E_m\\uparrow \\)\\( M\\downarrow \\)\\( F_m\\uparrow \\)\\( S_m\\uparrow \\)\\( E_m\\uparrow \\)
----.048.903.884.922.052.832.832.893
✓---.046.907.889.925.051.833.839.895
-✓--.045.908.893.929.048.837.844.898
--✓-.046.906.891.927.050.833.841.894
---✓.043.914.893.933.048.840.848.900
✓✓--.044.911.893.928.049.837.844.897
✓-✓-.046.909.893.927.046.840.845.898
✓-✓✓.040.918.897.935.045.848.849.904
✓✓-✓.042.916.894.931.044.850.850.906
✓✓✓✓.040.923.897.937.043.852.852.908
", + "bbox": [ + 506, + 542, + 885, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "discussions and ablation studies on the hyperparameters can be found in the supplementary material.", + "bbox": [ + 496, + 688, + 890, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Object Popping Network as Plug-in: Our object popping network can be easily adapted with different encoders and with different existing RGB-D models. For example, with a ResNet-18 [15] encoder and convolution-based decoder, our poping only costs around 12.7M additional learning parameters or 48.7 MB model size. We show in Table 5 that our method can favorably improve performance over the baseline with less than an extra $10\\%$ GFlops.", + "bbox": [ + 496, + 719, + 890, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pop-out Under Reduced Training Data: Here, we are interested in analyzing source-free depth's benefits. Therefore, we conduct different experiments by reducing the training data. As shown in Figure 8(left), when both our", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1038", + "bbox": [ + 483, + 924, + 514, + 936 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Generalization and cost over different RGB-D baselines. Dataset Metric Flops Param S1P [10] NC4K [42] M $\\downarrow$ ${F}_{m} \\uparrow$ ${S}_{m} \\uparrow$ ${E}_{m} \\uparrow$ ${M\\downarrow }$ ${F}_{m} \\uparrow$ ${S}_{m} \\uparrow$ ${E}_{m} \\uparrow$ +Ours 363.2 59.8 .053 .899 .874 .919 .057 .809 .804 .872 373.7 72.5 .051 .910 .886 .923 .055 .814 .811 .878 SPNet [92] 149.0 150.4 .044 .911 .887 .914 .054 .828 .825 .874 +Ours 159.5 163.1 .042 .917 .894 .932 .044 .851 .851 .905", + "bbox": [ + 78, + 87, + 467, + 165 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/228225256ac391b291cd02e59acc6601ce70bd569a76eddd625e1174b20fb95d.jpg", + "image_caption": [ + "Figure 8. Benefit of our method on reducing training data. When our network is trained only with $25\\%$ data, the performance remains competitive compared to the baseline, i.e., the absolute performance gains are $+0.4\\%$ max F-measure and $+0.1\\%$ S-measure on NC4K compared to the baseline trained with all data." + ], + "image_footnote": [], + "bbox": [ + 81, + 178, + 467, + 325 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c0240fe18a2b8fcba530dfb9589bae2379566a86fac6320b5455982b53e654cc.jpg", + "image_caption": [ + "Figure 9. Histogram on the gain. Please zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 78, + 406, + 431, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "PopNet and our baseline are trained with all data, our PopNet can lead to absolute improvements with $5.6\\%$ in max F-measure and with $3.9\\%$ in S-measure on COD10K dataset. While our PopNet is trained with only $25\\%$ data, it can still achieve competitive performance compared to our baseline trained with all data. Similar phenomena can be observed in the NC4K dataset as shown in Figure 8(right). To conclude, our method can efficiently explore the geometric prior and significantly reduce the required training data volume.", + "bbox": [ + 75, + 579, + 468, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Gain over baseline: With depth cues, as shown in Figure 9, we boost the performance in 3125 over 4121 images ( $\\sim 75\\%$ cases). We also show in Table 6 that our network performs favorably over baselines with single or multiple objects. Our method may fail when the source-free depth is clueless. This mainly happens when the object is well concealed and fools the depth network. However, such cases are also challenging, even for humans.", + "bbox": [ + 75, + 717, + 468, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Are RGB-D Methods Better Than RGB-only Methods?", + "text_level": 1, + "bbox": [ + 76, + 839, + 467, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "RGB-D methods are indeed better than RGB-only methods provided GT depth. However, only a few RGB-D methods can benefit from the source-free depth. For example,", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3dfe58a2ff7eb57aaa9ec93e81d023aa58244b72a2a1fa92e4a2cf8811a37c6b.jpg", + "table_caption": [ + "Table 6. Multi-Object performance on NC4K [42] with size ${512}^{2}$ ." + ], + "table_footnote": [], + "table_body": "
Obj. Nbr. (%)MetricSingle (92%)Two (6%)More (2%)Overall
M↓Fm↑M↓Fm↑M↓Fm↑M↓Fm↑
RGB Baseline.054.828.067.811.091.738.056.825
+ Dxf.050.842.063.821.093.746.051.839
Ours.040.864.051.847.079.767.042.861
", + "bbox": [ + 503, + 103, + 887, + 159 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cb29a7a53d2039aa839450ea039ff04e04c737fc9c2d2acc131dd4124efa0215.jpg", + "table_caption": [ + "Table 7. Performance with RGB- $D_{sf}$ model vs. with RGB-only baseline. $D_{sf}$ stands for source-free depth." + ], + "table_footnote": [], + "table_body": "
DatasetCOD10K [9]NC4K [42]
MetricM↓Fm↑Sm↑Em↑M↓Fm↑Sm↑Em↑
DASNet [88].041.643.793.864.055.747.830.879
+ Dsf.041.642.796.858.055.743.830.874
SPNet [92].040.743.801.867.052.846.833.883
+ Dsf.037.776.808.869.054.828.825.874
", + "bbox": [ + 501, + 196, + 890, + 263 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "as shown in Table 7, DASNet [88] achieves poorer performance compared to RGB baseline when trained with source-free depth. Similarly, even for one of the SOTA RGB-D models, SPNet [92], the performance on NC4K dataset with RGB-only input is better than with additional source-free depth. Moreover, when provided with source-free depth, none of the existing RGB-D methods outperform the best performing RGB-only method (e.g., ZoomNet [47]) on COD. The same observation was also made on SOD as well. This could be because of the domain gap coupled with the fusion design, among others. We also found it non-trivial to extend the most well-performing RGB-only methods to the RGB-D case. Note that our PopNet performs better than all existing RGB-only and RGB-D methods, with source-free or GT depth maps.", + "bbox": [ + 496, + 287, + 892, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 537, + 617, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We demonstrate a successful case of cross-domain cross-task depth to semantics knowledge transfer using only the source model. In this paper, the source-free depth at the target offered by a given source model is used. The proposed method learns to transfer knowledge from depth to semantics using the objects' pop-out prior. We facilitate our network to use such prior by designing a novel network architecture. The designed network reasons about the objects by popping them out from the provided depth maps. This process is followed by separating objects from the background using the learned contact surface. We show that the joint learning of object pop-out and contact surface can be successfully supervised using the target semantics. Exhaustive experiments on SOD and COD benchmarks show the successful transfer of depth knowledge to the target, in terms of improved performance and generalization.", + "bbox": [ + 496, + 565, + 890, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement The authors thank the anonymous reviewers and ACs for their tremendous efforts and helpful comments. This research is financed in part by the Conseil Régional de Bourgogne-Franche-Comté, Toyota Motor Europe (research project TRACE-Zurich), and the Alexander von Humboldt Foundation.", + "bbox": [ + 496, + 810, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1039", + "bbox": [ + 483, + 924, + 514, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Peshal Agarwal, Danda Pani Paudel, Jan-Nico Zaech, and Luc Van Gool. Unsupervised robust domain adaptation without source data. In WACV, 2022. 2", + "[2] Amir Atapour-Abarghouei and Toby P Breckon. Real-time monocular depth estimation using synthetic data with domain adaptation via image style transfer. In IEEE CVPR, 2018. 1", + "[3] Adriano Cardace, Luca De Luigi, Pierluigi Zama Ramirez, Samuele Salti, and Luigi Di Stefano. Plugging self-supervised monocular depth into unsupervised domain adaptation for semantic segmentation. In WACV, 2022. 1", + "[4] Po-Yi Chen, Alexander H Liu, Yen-Cheng Liu, and Yu-Chiang Frank Wang. Towards scene understanding: Unsupervised monocular depth estimation with semantic-aware representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 4", + "[5] Innes C Cuthill, Martin Stevens, Jenna Sheppard, Tracey Maddocks, C Alejandro Parraga, and Tom S Troscianko. Disruptive coloration and background pattern matching. Nature, 434(7029):72-74, 2005. 3", + "[6] Jiahua Dong, Zhen Fang, Anjin Liu, Gan Sun, and Tongliang Liu. Confident anchor-induced multi-source free domain adaptation. NeurIPS, 2021. 1", + "[7] Aliya El Nagar, Daniel Osorio, Sarah Zylinski, and Steven M Sait. Visual perception and camouflage response to 3d backgrounds and cast shadows in the european cuttlefish, sepia officinalis. JEB, 224(11), 2021. 3", + "[8] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE TPAMI, 44(10):6024-6042, 2022. 3, 6", + "[9] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In IEEE CVPR, 2020. 3, 5, 6, 7, 8", + "[10] Deng-Ping Fan, Zheng Lin, Zhao Zhang, Menglong Zhu, and Ming-Ming Cheng. Rethinking RGB-D salient object detection: Models, datasets, and large-scale benchmarks. IEEE TNNLS, 32(5):2075-2089, 2021. 3, 5, 6, 7, 8", + "[11] Deng-Ping Fan, Jing Zhang, Gang Xu, Ming-Ming Cheng, and Ling Shao. Salient objects in clutter. IEEE TPAMI, 2023. 2", + "[12] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In IEEE CVPR, 2019. 1", + "[13] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In IEEE ICCV, 2019. 1, 2, 4", + "[14] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In IEEE CVPR, 2022. 4", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE CVPR, 2016. 7", + "[16] Ian P Howard. Perceiving in depth, volume 1: basic mechanisms. Oxford University Press, 2012. 1" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Junjie Hu, Mete Ozay, Yan Zhang, and Takayuki Okatani. Revisiting single image depth estimation: Toward higher resolution maps with accurate object boundaries. In WACV, 2019. 4", + "[18] Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection. MIR, 2023. 3", + "[19] Wei Ji, Jingjing Li, Shuang Yu, Miao Zhang, Yongri Piao, Shunyu Yao, Qi Bi, Kai Ma, Yefeng Zheng, Huchuan Lu, et al. Calibrated RGB-D salient object detection. In IEEE CVPR, 2021. 5, 7", + "[20] Wei Ji, Jingjing Li, Miao Zhang, Yongri Piao, and Huchuan Lu. Accurate RGB-D salient object detection via collaborative learning. In ECCV, 2020. 3", + "[21] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In IEEE CVPR, 2022, 3, 6, 7", + "[22] Wen-Da Jin, Jun Xu, Qi Han, Yi Zhang, and Ming-Ming Cheng. Cdnet: Complementary depth network for rgb-d salient object detection. IEEE TIP, 30:3376-3390, 2021. 3", + "[23] Ran Ju, Ling Ge, Wenjing Geng, Tongwei Ren, and Gangshan Wu. Depth saliency based on anisotropic centersurround difference. In IEEE ICIP, 2014. 5, 6, 7", + "[24] Hongwen Kang, Alexei A Efros, Martial Hebert, and Takeo Kanade. Image composition for object pop-out. In IEEE ICCVW, 2009. 2", + "[25] Youngeun Kim, Donghyeon Cho, Kyeongtak Han, Priyadarshini Panda, and Sungeun Hong. Domain adaptation without source data. IEEE TAI, 2(6):508-518, 2021. 1, 2", + "[26] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In IEEE CVPR, 2021. 1", + "[27] Jogendra Nath Kundu, Akshay R Kulkarni, Suvaansh Bhambri, Deepesh Mehta, Shreyas Anand Kulkarni, Varun Jampani, and Venkatesh Babu Radhakrishnan. Balancing discriminability and transferability for source-free domain adaptation. In ICML, 2022. 1", + "[28] Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In IEEE CVPR, 2020. 1, 2", + "[29] Vinod K Kurmi, Venkatesh K Subramanian, and Vinay P Namboodiri. Domain impression: A source data free domain adaptation method. In WACV, 2021. 1, 2", + "[30] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, MinhTriet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. CVU, 184:45-56, 2019. 5, 7", + "[31] Minhyeok Lee, Chaewon Park, Suhwan Cho, and Sangyoun Lee. Spsn: Superpixel prototype sampling network for rgb-d salient object detection. In ECCV, 2022. 6, 7", + "[32] Aixuan Li, Jing Zhang, Yunqiu Lv, Bowen Liu, Tong Zhang, and Yuchao Dai. Uncertainty-aware joint salient object and camouflaged object detection. In IEEE CVPR, 2021. 3, 7", + "[33] Gongyang Li, Zhi Liu, Minyu Chen, Zhen Bai, Weisi Lin, and Haibin Ling. Hierarchical alternate interaction network" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1040", + "bbox": [ + 483, + 925, + 514, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "forrgb-d salient object detection.IEEE TIP,30:3528-3542, 2021.3,6,8", + "[34] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In IEEE CVPR, 2020. 2", + "[35] Rui Li, Danna Xue, Shaolin Su, Xiantuo He, Qing Mao, Yu Zhu, Jinqiu Sun, and Yanning Zhang. Learning depth via leveraging semantics: Self-supervised monocular depth estimation with both implicit and explicit semantic guidance. Pattern Recognition (PR), page 109297, 2023. 4", + "[36] Xinhao Li, Jingjing Li, Lei Zhu, Guoqing Wang, and Zi Huang. Imbalanced source-free domain adaptation. In ACM MM, 2021. 1", + "[37] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020. 2", + "[38] Jiang-Jiang Liu, Qibin Hou, Ming-Ming Cheng, Jiashi Feng, and Jianmin Jiang. A simple pooling-based design for real-time salient object detection. In IEEE CVPR, 2019. 3", + "[39] Yuang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. In IEEE CVPR, 2021. 1, 2", + "[40] Yukang Lu, Dingyao Min, Keren Fu, and Qijun Zhao. Depth-cooperated trimodal network for video salient object detection. In ICIP. IEEE, 2022. 3", + "[41] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM TOG, 39(4):1-13, 2020. 1", + "[42] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In IEEE CVPR, 2021. 3, 5, 6, 7, 8", + "[43] Daniel McKee, Zitong Zhan, Bing Shuai, Davide Modolo, Joseph Tighe, and Svetlana Lazebnik. Transfer of representations to video label propagation: Implementation factors matter. arXiv preprint arXiv:2203.05553, 2022. 6", + "[44] Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan. Camouflaged object segmentation with distraction mining. In IEEE CVPR, 2021. 7", + "[45] S Mahdi H Miangoleh, Sebastian Dille, Long Mai, Sylvain Paris, and Yagiz Aksoy. Boosting monocular depth estimation models to high-resolution via content-adaptive multi-resolution merging. In IEEE CVPR, 2021. 1, 4", + "[46] Yuzhen Niu, Yujie Geng, Xueqing Li, and Feng Liu. Leveraging stereopsis for saliency analysis. In IEEE CVPR, 2012. 5, 6", + "[47] Youwei Pang, Xiaqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In IEEE CVPR, 2022. 3, 6, 7, 8", + "[48] Houwen Peng, Bing Li, Weihua Xiong, Weiming Hu, and Rongrong Ji. RGBD salient object detection: a benchmark and algorithms. In ECCV, 2014. 5, 6", + "[49] Yongri Piao, Zhengkun Rong, Miao Zhang, Weisong Ren, and Huchuan Lu. A2dele: Adaptive and attentive depth distiller for efficient RGB-D salient object detection. In IEEE CVPR, 2020. 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In IEEE ICCV, 2021. 1, 4", + "[51] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI, 44(3):1623-1637, 2022. 1, 2, 4", + "[52] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4", + "[53] Suman Saha, Anton Obukhov, Danda Pani Paudel, Menelaos Kanakis, Yuhua Chen, Stamatios Georgoulis, and Luc Van Gool. Learning to relate depth and semantics for unsupervised domain adaptation. In IEEE CVPR, 2021. 1", + "[54] Roshni Sahoo, Divya Shanmugam, and John Guttag. Unsupervised domain adaptation in the absence of source data. arXiv preprint arXiv:2007.10233, 2020. 2", + "[55] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE CVPR, 2016. 1", + "[56] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from RGBD images. In ECCV, 2012. 1", + "[57] Przemysław Skurowski, Hassan Abdulameer, J Blaszczyk, Tomasz Depta, Adam Kornacki, and P Koziel. Animal camouflage analysis: Chameleon database. Unpublished manuscript, 2(6):7, 2018. 5, 7", + "[58] Hwanjun Song, Eunyoung Kim, Varun Jampan, Deqing Sun, Jae-Gil Lee, and Ming-Hsuan Yang. Exploiting scene depth for object detection with multimodal transformers. In BMVC, 2021. 3", + "[59] Yujia Sun, Geng Chen, Tao Zhou, Yi Zhang, and Nian Liu. Context-aware cross-level fusion network for camouflaged object detection. In *IJCAI*, 2021. 3, 7", + "[60] Yujia Sun, Shuo Wang, Chenglizhao Chen, and Tian-Zhu Xiang. Boundary-guided camouflaged object detection. In IJ-CAI, 2022. 3", + "[61] Alessio Tonioni, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Unsupervised domain adaptation for depth prediction from images. IEEE TPAMI, 42(10):2396-2409, 2019. 1", + "[62] Anne Treisman. Preattentive processing in vision. ICVGIP, 31(2):156-177, 1985. 2", + "[63] Fengyun Wang, Jinshan Pan, Shoukun Xu, and Jinhui Tang. Learning discriminative cross-modality features for RGB-D saliency detection. IEEE TIP, 31:1285-1297, 2022. 6, 7", + "[64] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In IEEE CVPR, 2022. 2", + "[65] Cho-Ying Wu, Jialiang Wang, Michael Hall, Ulrich Neumann, and Shuochen Su. Toward practical monocular indoor depth estimation. In IEEE CVPR, 2022. 1", + "[66] Yu-Huan Wu, Yun Liu, Jun Xu, Jia-Wang Bian, Yu-Chao Gu, and Ming-Ming Cheng. Mobilesal: Extremely efficient rgb-d salient object detection. IEEE TPAMI, 44(12):10261-10269, 2022. 3", + "[67] Zongwei Wu, Guillaume Allibert, Christophe Stolz, and Cédric Demonceaux. Depth-adapted CNN for RGB-D cameras. In ACCV, 2020. 1" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1041", + "bbox": [ + 483, + 925, + 513, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[68] Zongwei Wu, Guillaume Allibert, Christophe Stolz, Chao Ma, and Cédric Demonceaux. Modality-guided subnetwork for salient object detection. In 3DV, 2021. 1, 3", + "[69] Zongwei Wu, Shriarulmozhivarman Gobichettipalayam, Brahim Tamadazte, Guillaume Allibert, Danda Pani Paudel, and Cedric Demonceaux. Robust RGB-D fusion for saliency detection. 3DV, 2022. 3, 5", + "[70] Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In IEEE CVPR, 2019. 3", + "[71] Haifeng Xia, Handong Zhao, and Zhengming Ding. Adaptive adversarial network for source-free domain adaptation. In IEEE ICCV, 2021. 1, 2", + "[72] Ke Xian, Jianming Zhang, Oliver Wang, Long Mai, Zhe Lin, and Zhiguo Cao. Structure-guided ranking loss for single image depth prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 4", + "[73] Mochu Xiang, Jing Zhang, Yunqiu Lv, Aixuan Li, Yiran Zhong, and Yuchao Dai. Exploring depth contribution for camouflaged object detection. arXiv e-prints, pages arXiv-2106, 2021. 1, 6", + "[74] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping Fan. Uncertainty-guided transformer reasoning for camouflaged object detection. In IEEE ICCV, 2021. 3, 7", + "[75] Shiqi Yang, Joost van de Weijer, Luis Herranz, Shangling Jui, et al. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. NeurIPS, 2021. 1", + "[76] Shiqi Yang, Yaxing Wang, Joost van de Weijer, and Luis Herranz. Unsupervised domain adaptation without source data by casting a bait. arXiv preprint arXiv:2010.12427, 2020. 2", + "[77] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In IEEE ICCV, 2021. 1, 2", + "[78] Qiang Zhai, Xin Li, Fan Yang, Chenglizhao Chen, Hong Cheng, and Deng-Ping Fan. Mutual graph learning for camouflaged object detection. In IEEE CVPR, 2021. 7", + "[79] Wei Zhai, Yang Cao, HaiYong Xie, and Zheng-Jun Zha. Deep texton-coherence network for camouflaged object detection. IEEE TMM, 2022. 3", + "[80] Chen Zhang, Runmin Cong, Qinwei Lin, Lin Ma, Feng Li, Yao Zhao, and Sam Kwong. Cross-modality discrepant interaction network for RGB-D salient object detection. In ACM MM, 2021. 3, 7", + "[81] Jing Zhang, Deng-Ping Fan, Yuchao Dai, Xin Yu, Yiran Zhong, Nick Barnes, and Ling Shao. RGB-D saliency detection via cascaded mutual information minimization. In IEEE ICCV, 2021. 7", + "[82] Jing Zhang, Yunqiu Lv, Mochu Xiang, Aixuan Li, Yuchao Dai, and Yiran Zhong. Depth-guided camouflaged object detection. CoRR, abs/2106.13217, 2021. 3", + "[83] Miao Zhang, Shuang Xu, Yongri Piao, Dongxiang Shi, Shusen Lin, and Huchuan Lu. Preynet: Preying on camouflaged objects. In ACM MM, 2022. 3, 6", + "[84] Pingping Zhang, Dong Wang, Huchuan Lu, Hongyu Wang, and Baocai Yin. Learning uncertain convolutional features for accurate saliency detection. In IEEE ICCV, 2017. 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[85] Wenbo Zhang, Ge-Peng Ji, Zhuo Wang, Keren Fu, and Qijun Zhao. Depth quality-inspired feature manipulation for efficient RGB-D salient object detection. In ACM MM, 2021. 6", + "[86] Zhoutong Zhang, Forrester Cole, Richard Tucker, William T Freeman, and Tali Dekel. Consistent depth of moving objects in video. ACM TOG, 40(4):1-12, 2021. 1", + "[87] Zhao Zhang, Zheng Lin, Jun Xu, Wen-Da Jin, Shao-Ping Lu, and Deng-Ping Fan. Bilateral attention network forrgb-d salient object detection. IEEE TIP, 30:1949-1961, 2021. 6", + "[88] Jiawei Zhao, Yifan Zhao, Jia Li, and Xiaowu Chen. Is depth really necessary for salient object detection? In ACM MM, 2020. 3, 8", + "[89] Shanshan Zhao, Huan Fu, Mingming Gong, and Dacheng Tao. Geometry-aware symmetric domain adaptation for monocular depth estimation. In IEEE CVPR, 2019. 1", + "[90] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In IEEE CVPR, 2022. 3", + "[91] Tao Zhou, Deng-Ping Fan, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Rgb-d salient object detection: A survey. CVMJ, pages 1-33, 2021. 1", + "[92] Tao Zhou, Huazhu Fu, Geng Chen, Yi Zhou, Deng-Ping Fan, and Ling Shao. Specificity-preserving RGB-D saliency detection. In IEEE ICCV, 2021. 3, 5, 6, 7, 8", + "[93] Zikun Zhou, Wenjie Pei, Xin Li, Hongpeng Wang, Feng Zheng, and Zhenyu He. Saliency-associated object tracking. In IEEE ICCV, 2021. 3", + "[94] Hongwei Zhu, Peng Li, Haoran Xie, Xuefeng Yan, Dong Liang, Dapeng Chen, Mingqiang Wei, and Jing Qin. I can find you! boundary-guided separated attention network for camouflaged object detection. In AAAI, 2022. 3", + "[95] Jinchao Zhu, Xiaoyu Zhang, Shuo Zhang, and Junnan Liu. Inferring camouflaged objects by texture-aware interactive guidance network. In AAAI, 2021. 3" + ], + "bbox": [ + 501, + 90, + 892, + 602 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1042", + "bbox": [ + 483, + 924, + 514, + 936 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Source-free Depth for Object Pop-out/1ac228ed-431b-40e3-b444-e5a574a99abb_model.json b/2023/Source-free Depth for Object Pop-out/1ac228ed-431b-40e3-b444-e5a574a99abb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a96391ca48a4ab98143974c4a37cd9b8fb99479f --- /dev/null +++ b/2023/Source-free Depth for Object Pop-out/1ac228ed-431b-40e3-b444-e5a574a99abb_model.json @@ -0,0 +1,2829 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.294, + 0.131, + 0.677, + 0.154 + ], + "angle": 0, + "content": "Source-free Depth for Object Pop-out" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.179, + 0.856, + 0.217 + ], + "angle": 0, + "content": "Zongwei Wu\\(^{1,2,3}\\) Danda Pani Paudel\\(^{1,4}\\) Deng-Ping Fan\\(^{1*}\\) Jingjing Wang\\(^{5}\\) Shuo Wang\\(^{1}\\) Cédric Demonceaux\\(^{2,6}\\) Radu Timofte\\(^{3}\\) Luc Van Gool\\(^{1,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.218, + 0.87, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\) CVL, ETH Zurich \\(^{2}\\) University of Burgundy, CNRS, ICB \\(^{3}\\) Computer Vision Lab, CAIDAS & IFI, University of Wurzburg \\(^{4}\\) INSAIT, Sofia University \\(^{5}\\) AUST \\(^{6}\\) University of Lorraine, CNRS, Inria, Loria" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.304 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.474, + 0.699 + ], + "angle": 0, + "content": "Depth cues are known to be useful for visual perception. However, direct measurement of depth is often impracticable. Fortunately, though, modern learning-based methods offer promising depth maps by inference in the wild. In this work, we adapt such depth inference models for object segmentation using the objects' \"pop-out\" prior in 3D. The \"pop-out\" is a simple composition prior that assumes objects reside on the background surface. Such compositional prior allows us to reason about objects in the 3D space. More specifically, we adapt the inferred depth maps such that objects can be localized using only 3D information. Such separation, however, requires knowledge about contact surface which we learn using the weak supervision of the segmentation mask. Our intermediate representation of contact surface, and thereby reasoning about objects purely in 3D, allows us to better transfer the depth knowledge into semantics. The proposed adaptation method uses only the depth model without needing the source data used for training, making the learning process efficient and practical. Our experiments on eight datasets of two challenging tasks, namely salient object detection and camouflaged object detection, consistently demonstrate the benefit of our method in terms of both performance and generalizability. The source code is publicly available at https://github.com/Zongwei97/PopNet." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.728, + 0.21, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.754, + 0.47, + 0.876 + ], + "angle": 0, + "content": "The 3D knowledge of the scene is long known to be complementary to the task of visual perception [12, 16, 56, 67, 91]. Often in practice, though, visual perception needs to be carried out using only 2D images. Given multiple images, 3D geometry may be recovered using the structure-from-motion techniques [26, 41, 55, 86]. Such inversion, however, is not compatible when only a single image is available. Under such circumstances, image inversion" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.291, + 0.891, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.432, + 0.895, + 0.488 + ], + "angle": 0, + "content": "Figure 1. Depth to semantics conversion using the object pop-out prior. For input RGB and source-free depth pair, we learn contact surface. The obtained contact surface is then used to separate objects and backgrounds to derive pseudo semantics for supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.893, + 0.583 + ], + "angle": 0, + "content": "to depth map is usually done using learning-based methods [13,45,50,51,65], which have shown unparalleled success in the recent years. Unfortunately, the learning-based methods may not offer high-quality depth maps due to the generalization deficiency across domains [68, 73]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.895, + 0.734 + ], + "angle": 0, + "content": "Despite poor generalization, the knowledge gained in one domain is shown to be useful in other close-by domains. This utility is harnessed by performing the so-called domain adaptation (DA) [2,3,6,53,61,89]. In fact, it has been shown recently that DA methods can efficiently transfer knowledge using only the prediction models, i.e., without requiring access to the data where the model is trained – also known as the source-free domain adaptation (SDA) [25,28,39,71,77]. The SDA methods are of gripping interest due to their efficiency and privacy promises." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.896, + 0.902 + ], + "angle": 0, + "content": "The most existing SDA methods make one or both of these implicit assumptions: (a) a similar (as that of the source) supervising task at the target [28, 75, 77]; (b) task of discrete (and known) label space [27, 29, 36, 71]. The former assumption not only makes the source and target domains easier to compare but also potentially keeps the two domains closer. The latter assumption allows performing SDA by self-training where the discrete labels facilitate reasoning about the model's confidence. The self-training is then performed by boosting the confidence at the target for some picked reliable examples." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.438, + 0.902 + ], + "angle": 0, + "content": "*Corresponding Author: Deng-Ping Fan (dengpfan@gmail.com)" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "1032" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.094, + 0.688, + 0.107 + ], + "angle": 0, + "content": "Cross-Task Cross-Dataset Analysis with Depth Pop-out Prior" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.108, + 0.894, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.231, + 0.893, + 0.286 + ], + "angle": 0, + "content": "Figure 2. The performance gained in F-measure using our method over the established baselines. Our method \\((\\triangle, \\square)\\) significantly improves baselines on 8 datasets and 2 tasks (each task on 4 datasets- SOD: left two; COD: right two). We also compare 24 methods \\((\\bullet)\\), where our method offers state-of-the-art results despite their task specialization. Note that all methods are connected using lines to illustrate their performance fluctuations across datasets. Please, refer Tables 1 & 2 and Section 4.3 for more details and discussions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.47, + 0.546 + ], + "angle": 0, + "content": "In this work, we aim for source-free transfer of depth knowledge for object detection. Such transfer is desired to assist in locating the object by depth cues and to exploit the depth knowledge despite the domain gap. The addressed problem setting differs from the standard SDA in terms of (a) the source and target tasks' difference; (b) and the continuous label space of the depth. These differences (with standard SDA) make our task at hand very challenging, which is addressed for the first time in this paper, up to our knowledge. To address such a challenging problem, we rely on the \"pop-out\" prior, which allows us to reason about the object's location directly in the 3D space. The \"pop-out\" is a simple composition prior that assumes objects reside on the background surface. A graphical illustration of the used \"pop-out\" prior, using the results obtained by our method, is given in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.47, + 0.702 + ], + "angle": 0, + "content": "The pop-out prior for image composition was successfully used by Kang et al. in [24]. An early work of Treisman has provided an in-depth study of such prior in [62]. In this work, we rely on the same compositional foundation of these works and exploit the pop-out before transferring the depth of knowledge across domains. Although our motivation comes from these early works, our experimental setup largely differs from theirs. We differ in terms of not only depth knowledge transfer across domains (without source data) but also in target supervision using only semantics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The proposed method exploits the source-free depth to map it into a space where objects in depth stand out better against the background. This mapping is used for object and background separation using a learned contact surface between them. Such separation allows us to derive the semantic masks which can be directly compared against the ground truth for supervision. Using this supervision at the target, we can minimize the domain and task gap between the source and target. The overall framework of our method that performs cross-task cross-domain knowledge transfer by using the intermediate representation of the pop-out space is shown in Figure 3. As can be seen, we first leverage an object popping network to encourage the object" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.501 + ], + "angle": 0, + "content": "to jump out from the source-free depth. Then, we introduce another network i.e., the segmentation with contact surface, to localize the object and predict the contact surface. These learning modules are jointly trained in an end-to-end manner, transferring the source-free depths into intermediate representations which are adapted to the target task, i.e., object detection. To evaluate the proposed method, we conducted exhaustive experiments on eight datasets of two challenging tasks, namely salient object detection and camouflaged object detection. In both tasks, our method significantly improves the established baselines and offers state-of-the-art results at the same time, whose overview can be seen in Figure 2. The major contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.503, + 0.891, + 0.533 + ], + "angle": 0, + "content": "- Our problem of transferring source-free depth knowledge across domains and tasks is practical and novel." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.534, + 0.89, + 0.563 + ], + "angle": 0, + "content": "- Our method relies on our object pop-out prior for visual understanding, which is simple and effective." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.564, + 0.89, + 0.593 + ], + "angle": 0, + "content": "- Results of our method in two different tasks are significantly better than the baselines and existing models." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.503, + 0.891, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.65, + 0.62 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Source-free Adaptation: Knowledge transfer by domain adaptation without access to the source data has recently gathered vast interest [1,25,28,39,71,77], due to the privacy, practicality, and efficiency reasons. We observe accessing the source data for transferring the depth knowledge learned by the off-the-shelf knowledge models [13, 51] is particularly impractical due to their multi-stage training on various datasets. The existing source-free adaptation methods either use generative [28, 29, 34], pseudo-label [25, 37, 64], or other customized [54, 76] approaches. In this work, we use the pseudo-label-based approach. However, using existing methods is not straightforward in our setting due to the tasks' difference between source and target. With our pop-out technique that provides the pseudo semantics, the source-free depth is better transferred across tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Salient Object Detection (SOD): Saliency detection aims to detect and segment the most prominent region within an image that visually attracts human attention [11]. A" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "1033" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.088, + 0.852, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.281, + 0.892, + 0.366 + ], + "angle": 0, + "content": "Figure 3. Our proposed framework, termed PopNet, is composed of a source-free network, an object popping network, a segmentation with contact surface, and an object separation module. The source-free depth network generates pseudo-depth in an off-the-shelf manner (Section 3.1). The object popping network converts the source-free depth into the popped-out depth of objects, bridging the cross-domain and cross-task gaps (Section 3.2). The segmentation network uses this depth to estimate the object's mask and contact surface (Section 3.3). The object separation module then converts popped-out depth to the second mask of objects using the contact surface (Section 3.4). We compare both semantic masks against ground truth to supervise the whole pipeline in an end-to-end manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.47, + 0.563 + ], + "angle": 0, + "content": "number of works have shown that saliency can be an auxiliary step for different vision tasks such as object tracking [93], object detection [58], etc. Conventional saliency works are unimodal, i.e., they only require RGB images as input. In generic and common settings, RGB-based models [38, 70, 84] have already achieved very promising results. More recently, several works [10,20,49,66,68,88,92] exploit depth maps as additional clues to the 3D geometry since the depth can provide more truthful information on the object boundary as well as the scale awareness. These 3D features further improve the detection accuracy and performance in challenging scenarios [22, 33, 69, 80]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Camouflaged Object Detection (COD): Camouflage detection aims to find the preying object within an image. For computer vision society, primary works [9,32,47] often compare COD with SOD. A number of works [8,9,74,79] have shown that simply extending saliency models [70] on COD will lead to undesired results, which is mainly caused by the nature of target object, i.e., concealed or prominent. Hence, to constrain the attention on the concealed objects, several works come up with different perceptual systems that mimic human behavior vis-a-vis camouflaged objects, such as three-stage localize-segment-rank strategy [42]; iterative refinement [21], which is similar to repeatedly looking on the images; zooming into possible regions [47, 59]. Others [18, 40, 60, 82, 83, 90, 94, 95] deeply explore the texture difference with the help of the gradient [18], frequency [90], edges [60, 94], and probability [32, 74]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The latest psychological studies [5, 7] have shown that human perception can naturally benefit from the depth cues to understand the scene: (A) the smooth variation within the object can contribute to alleviating the fake edges and preserving the object structure; (B) the depth discontinuity on the object boundary can make the segmentation easier. In" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.892, + 0.473 + ], + "angle": 0, + "content": "spired by these observations, we aim to explore the source-free depth for both SOD and COD tasks. To tackle the domain gap for source-free depth maps, we propose to jointly finetune the source-free depth together with the semantic network in an end-to-end manner, with both self-supervised loss and weak semantic supervision." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.49, + 0.67, + 0.508 + ], + "angle": 0, + "content": "3. Proposed PopNet" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.892, + 0.82 + ], + "angle": 0, + "content": "Given an input RGB image \\(I\\) with size \\(I \\in \\mathbb{R}^{3 \\times H \\times W}\\), where \\(H\\) and \\(W\\) are the spatial resolutions, i.e., height and width, of the image, our objective is to predict the semantic mask \\(\\tilde{S} \\in \\mathbb{R}^{H \\times W}\\) for object detection. As shown in Figure 3, the input image \\(I\\) is firstly fed into a frozen-weight depth network to generate the source-free depth \\(D_{sf} \\in \\mathbb{R}^{H \\times W}\\) (Section 3.1). Then the mimicked multi-modal images are fed together into the depth popping network to compute the intermediate popped-out depth \\(D_{po} \\in \\mathbb{R}^{H \\times W}\\) (Section 3.2). This intermediate representation, as well as the RGB image \\(I\\), is later processed by the segmentation network and transformed into a contact surface \\(D_{c} \\in \\mathbb{R}^{H \\times W}\\) and a semantics prediction \\(\\tilde{S} \\in \\mathbb{R}^{H \\times W}\\) (Section 3.3). On the one hand, the semantics prediction is directly supervised by the ground truth mask \\(GT\\), denoted as \\(G\\), which is similar to conventional segmentation supervision. On the other hand, we further explore the contact surface to pop the object out of the background by means of our object separation module (Section 3.4). This transfers the geometric cues into pseudo semantics and leads to another level of supervision." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.747, + 0.848 + ], + "angle": 0, + "content": "3.1. Source-free Depth Network" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In a practical setting, the GT depth is not always available. Therefore, we generate the source-free depth \\( D_{sf} \\) in an off-the-shelf manner to mimic the multi-modal input." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "1034" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.259 + ], + "angle": 0, + "content": "We choose the state-of-the-art DPT model [50] with frozen weights as our source-free depth network, which offers us promising depth at the target. This choice is made upon its generalization capability as suggested in [51]. To obtain the depth map with the highest quality possible, by enhancing local details, we apply the boosting method [45] together with DPT. Despite the plausible results achieved by learning-based methods, the obtained source-free depth does not always offer high-quality geometric cues due to the domain gap. Therefore, we leverage the geometric and semantic priors to jointly finetune the source-free depth." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.269, + 0.307, + 0.285 + ], + "angle": 0, + "content": "3.2. Object Popping Network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.47, + 0.442 + ], + "angle": 0, + "content": "Network Architecture: We build a depth popping network to refine/smooth the source-free depth. The popping network follows the encoder-decoder design with skip connection as shown in Figure 4. In our case, we simply concatenate RGB and source-free depth at the input side to form a 4-channel input and feed them into the popping network. The encoder extracts semantic cues and generates five-scale outputs. Our decoder is composed of Conv2D, BN, ReLU, and upsampling layers. Following U-Net [52], we build a skip connection through simple addition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.47, + 0.536 + ], + "angle": 0, + "content": "Structure Preserving: To supervise our popping network, we first guide the depth refinement with the help of generated pseudo-depth. We only leverage structural similarity since we aim to detect, preserve, and extract object structure from the intermediate representation. We use the following SSIM loss [13] for structural similarity." + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.549, + 0.469, + 0.566 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d e p} = S S I M \\left(D _ {p o}, D _ {s f}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.47, + 0.775 + ], + "angle": 0, + "content": "Local Depth Smoothing: In addition to the pseudo-depth supervision, we propose two losses to constrain the depth also by semantics. We assume that the objects' structure should be distinguishable from the background, i.e., it should be smooth within the object region and sharp on the bounding pixels. Hence, we propose to leverage the weak semantic cues together with the geometric priors. Specifically, we first introduce a local smoothness loss. The locality is defined by the ground truth semantics \\( G \\). Technically, we mask out background pixels with element-wise multiplication to suppress the inactive area through \\( D_{obj} = D_{po} \\otimes G \\). Let \\( \\nabla_x \\) and \\( \\nabla_y \\) be the Sobel operations. Then, our local loss \\( \\mathcal{L}_{loc} \\) is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.785, + 0.47, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c} \\overrightarrow {n (p)} = (- \\nabla_ {x} (D _ {o b j} (p)); - \\nabla_ {y} (D _ {o b j} (p)); 1); \\\\ \\mathcal {L} _ {l o c} = \\sum_ {p} \\sum_ {q \\in \\mathcal {N} (p)} 1 - \\operatorname {c o s i n e} (\\overrightarrow {n (p)}; \\overrightarrow {n (q)}), \\end{array} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.855, + 0.47, + 0.903 + ], + "angle": 0, + "content": "where \\(\\overrightarrow{n}\\) stands for the normal, \\(p\\) denotes the pixels within the object region, \\(\\mathcal{N}(p)\\) is the neighboring pixels, and cosine is the cosine similarity between two vectors. This way, our" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.088, + 0.892, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.238, + 0.892, + 0.296 + ], + "angle": 0, + "content": "Figure 4. Our object popping network maps RGB-D inputs into popped-out depths. This network is supervised using a combination of structure preserving, local depth smoothing, and depth edge sharpening losses \\(\\mathcal{L}_{dep}\\), \\(\\mathcal{L}_{loc}\\), and \\(\\mathcal{L}_{wtv}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.312, + 0.892, + 0.357 + ], + "angle": 0, + "content": "local loss only works on the object area, making the object structure consistent within the target region. Applying local smoothness loss reduces the depth noise at the object level." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.359, + 0.892, + 0.421 + ], + "angle": 0, + "content": "Depth Edge Sharpening: In addition to the local smoothness, we also use edge sharpening. The edge sharpening loss is formulated as a weighted total variation. For this, we first compute the edge-aware weight \\( w(p) \\) at any pixel \\( p \\) as," + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.441, + 0.892, + 0.482 + ], + "angle": 0, + "content": "\\[\nw (p) = \\left\\{ \\begin{array}{l l} w _ {0}, & \\text {i f} \\nabla_ {x} (G (p)) ^ {2} + \\nabla_ {y} (G (p)) ^ {2} \\neq 0, \\\\ w _ {0} + \\gamma , & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.49, + 0.892, + 0.581 + ], + "angle": 0, + "content": "where \\( w_0 \\) is a pre-defined non-zero weight and \\( \\gamma \\) is an additional weight for boundary pixels. In our setting, we choose \\( w_0 \\) as the normalized (by the image size) count of the boundary pixels, and we set \\( \\gamma = 0.5 \\). We adopt the square form such that the large gradients play more important roles. Our weighted total-variation loss is given by," + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.589, + 0.892, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {w T v} = \\sum_ {p} \\sum_ {q \\in \\mathcal {N} (p)} w (p) \\cdot | | D _ {p o} (p) - D _ {p o} (q) | | _ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our weighted total-variance loss differs from the conventional edge loss, due to our weighting function. More specifically, our weighting function relies on semantic boundaries, unlike the commonly used image gradients [14, 17]. Our motivation for using semantic boundaries instead of image gradients comes from our interest in performing object detection under challenging conditions, such as camouflaged objects. In such cases, image gradients may result in misleading weights. At first glance, our loss function seems to be similar to semantic-guided depth estimation methods [4, 35, 72]. However, [72] uses GT depths, and [4, 35] use multi-frames for supervision. Unfortunately, such supervision is not possible in our setup. Note that we exploit single-view source-free depth while only using semantic ground truth for supervision. Furthermore, we transfer the source-free depths' knowledge despite their domain gap. Existing regularizations between depth and semantics are already exploited in our method, which is shown to be" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.925, + 0.515, + 0.937 + ], + "angle": 0, + "content": "1035" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "complimentary to our pop-out prior. We argue that our network architecture that exploits the pop-out prior is not trivial while simultaneously being generic and easy to use." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.137, + 0.47, + 0.168 + ], + "angle": 0, + "content": "The total objective function \\(\\mathcal{L}_{pop}\\) to supervise our object popping network is given by," + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.179, + 0.469, + 0.196 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p o p} = \\mathcal {L} _ {d e p} + \\lambda_ {1} \\cdot \\mathcal {L} _ {l o c} + \\lambda_ {2} \\cdot \\mathcal {L} _ {w t v}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.206, + 0.355, + 0.222 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are the hyperparameters." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.231, + 0.386, + 0.246 + ], + "angle": 0, + "content": "3.3. Segmentation with Contact Surface" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.254, + 0.471, + 0.466 + ], + "angle": 0, + "content": "The smoothness and edge losses encourage homogenizing the object structure, making it noticeable from the background. We now aim to further enlarge the object-background distance to make the object structure jump out. Specifically, we use an RGB-D segmentation network as shown in Figure 5. The main component of our segmentation network is a three-stream RGB-D network with some fusion design. In our setting, we choose [92] as our baseline since it is one of the SOTA RGB-D methods for saliency detection. We add a surface head to learn the depth of the contact surface \\( D_{c} \\), which has the same resolution as the input depth \\( D_{po} \\). Our surface head is composed of ConvLayer (Conv2D, BN, ReLU) and a Conv2D, which first decodes the feature maps and then transfers them into a 1-D map." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.475, + 0.254, + 0.491 + ], + "angle": 0, + "content": "3.4. Object Separation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.498, + 0.47, + 0.634 + ], + "angle": 0, + "content": "Using the previously discussed contact surface, in this section, we aim to separate the object from its background. At this point, we make an assumption that pixels in front of the contact surface belong to objects. The remaining pixels belong to the background. This assumption allows us to explicitly transfer the 3D knowledge into 2D semantics. Let the predicted depth of the contact surface be \\( D_{c} \\in \\mathbb{R}^{H \\times W} \\). We obtain the pseudo semantics \\( S_{s} \\), using the popped-out depth \\( D_{po} \\) and surface's depth \\( D_{c} \\), as:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.646, + 0.469, + 0.663 + ], + "angle": 0, + "content": "\\[\nS _ {s} = \\operatorname {s i g m o i d} \\left(\\sigma \\cdot \\left(D _ {p o} - D _ {c}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.779 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is a scalar value that controls the slope of the sigmoid function. In our experiments, we use \\(\\sigma = 10\\) to perform soft-thresholding, mimicking the desired hard one for binary outputs. Such soft thresholding facilitates the gradient back-propagation required for training. Finally, we minimize the gap between the pseudo semantics \\(S_{s}\\) and the GT semantics \\(G\\) with binary cross-entropy (BCE):" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.791, + 0.469, + 0.807 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s e p} = B C E \\left(S _ {s}, G\\right). \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.817, + 0.285, + 0.832 + ], + "angle": 0, + "content": "3.5. Overall Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Both of our trainable modules, i.e., object popping and segmentation networks, of our framework (Figure 3) are trained in an end-to-end manner. Therefore, the overall loss function consists of three parts: our depth popping loss" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.088, + 0.892, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.232, + 0.892, + 0.288 + ], + "angle": 0, + "content": "Figure 5. Our segmentation network uses a basic RGB-D three-stream network [92]. In addition to the conventional semantic head, we learn to predict pixel-wise contact surface. The contact surface is later used to transfer depth knowledge to semantics." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.305, + 0.892, + 0.351 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{pop}\\), our object separation loss \\(\\mathcal{L}_{sep}\\), and the conventional semantic loss \\(\\mathcal{L}_{sem}\\) from the RGB-D baseline network. The total loss \\(\\mathcal{L}_{total}\\) used for training is given by," + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.361, + 0.892, + 0.378 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p o p}} + \\alpha_ {1} \\cdot \\mathcal {L} _ {\\text {s e p}} + \\alpha_ {2} \\cdot \\mathcal {L} _ {\\text {s e m}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.388, + 0.756, + 0.402 + ], + "angle": 0, + "content": "where \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\) are hyperparameters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.892, + 0.705 + ], + "angle": 0, + "content": "Remarks: Our losses work in a complementary manner. The \\(\\mathcal{L}_{pop}\\) plays the same role as a smooth filter as in image smoothing. It removes the noisy depth response due to the domain gap while preserving the object structure with the help of a weak semantic label. Hence, the smoothed background becomes less informative while the object region becomes uniform, making it easily detectable. These functionalities contribute to transferring the source-free depth into popped-out depth, as desired. Such a process brings objects above the background surface, despite their distance from the camera and other distracting surfaces. The \\(\\mathcal{L}_{sep}\\), on the other hand, fully benefits from the \"pop-out\" prior to segment the object from the background. With the help of the learned contact surface, this loss enlarges the foreground-background distance by pulling them in opposite directions. Such pulling results in a binary-like mask, which effectively bridges the gap between geometric knowledge and semantics. Finally, both depth-transferred and learned semantics are compared to the ground truths for supervised training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.717, + 0.587, + 0.732 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.691, + 0.759 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Dataset Preparation: To better illustrate the generalizability of our approach, we evaluate the effectiveness of our approach on both SOD and COD benchmarks. We choose four widely used RGB-D SOD datasets, i.e., NLPR [48], NJUK [23], STERE [46], and SIP [10], as well as four COD datasets, i.e., CAMO [30], CHAMELEON [57], COD10K [9], and NC4K [42]. For SOD datasets, we conduct experiments with both GT depth and source-free depth. We follow the conventional learning protocol [19, 69, 92]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.925, + 0.516, + 0.936 + ], + "angle": 0, + "content": "1036" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.895, + 0.129 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on RGB-D SOD datasets. \\(\\uparrow (\\downarrow)\\) denotes that the higher (lower) is better. We use the Mean Absolute Error \\((M)\\), max F-measure \\((F_{m})\\), S-measure \\((S_{m})\\), and max E-measure \\((E_{m})\\) as evaluation metrics. G.D. stands for GT Depth. Bold denotes the best performance." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.129, + 0.892, + 0.304 + ], + "angle": 0, + "content": "
G.D.Public.Dataset MetricNLPR [48]NJUK [23]STERE [46]SIP [10]
M↓Fm↑Sm↑Em↑M↓Fm↑Sm↑Em↑M↓Fm↑Sm↑Em↑M↓Fm↑Sm↑Em↑
Performance of RGB-D Models Trained with Source-free Depth
XMM21 [85]DFM-Net.027.909.914.944.046.903.895.927.042.906.903.934.067.873.850.891
XTIP22 [63]DCMF.027.915.921.943.044.908.903.929.041.909.907.931.067.873.853.893
XCVPR22 [21]SegMAR.024.923.920.952.036.921.909.941.037.916.907.936.052.893.872.914
XCVPR22 [47]ZoomNet.023.916.919.944.037.926.914.940.037.918.909.938.054.891.868.909
XOursPopNet.022.925.926.956.031.931.920.949.032.922.916.947.046.911.885.926
Performance of RGB-D Models Trained with GT Depth
✓TIP21 [87]BIANet.032.888.900.930.056.878.867.898.048.898.895.918.091.816.802.847
✓TIP21 [33]HAINet.024.920.924.956.037.924.911.940.040.917.907.938.052.907.879.917
✓TNNLS21 [10]D3Net.029.904.911.942.046.909.899.927.044.902.906.925.063.880.860.897
✓ECCV22 [31]SPSN.023.917.923.956.032.927.918.949.035.909.906.941.043.910.891.932
✓OursPopNet.019.927.932.963.030.936.924.952.033.924.917.947.040.923.897.937
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.47, + 0.432 + ], + "angle": 0, + "content": "and use 700 images from NLPR and 1,485 images from NJUK for training. The rest are used for testing. For the unimodal COD dataset, we compare with both RGB COD models and RGB-D SOD models retrained on the COD datasets with the same source-free depth \\( D_{sf} \\). We follow the conventional training/testing protocol [8, 9, 21, 42, 47] and use 3,040 images from COD10K and 1,000 images from CAMO for training. The rest are used for testing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.47, + 0.537 + ], + "angle": 0, + "content": "Evaluation Metrics: We evaluate the performance with four generally-recognized metrics: Mean Absolute Error \\((M)\\), max F-measure \\((F_{m})\\), S-measure \\((S_{m})\\), and max E-measure \\((E_{m})\\). All the object segmentation masks are trained or downloaded from the official resources. To make a fair comparison, we evaluate the prediction semantics with the standardized evaluation protocol as [92]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.471, + 0.69 + ], + "angle": 0, + "content": "Implementation Details: Our model is implemented based on Pytorch with a V100 GPU. We use the Adam algorithm as an optimizer. The learning rate is initialized to \\( 1e - 4 \\) and is further divided by 10 every 60 epochs. We set the input resolution to \\( 352 \\times 352 \\) resolution for RGB and depth. A detailed comparison with higher resolution can be found in Table 3. During training, conventional data augmentation such as random flipping, rotating, and border clipping are adopted. The training takes around 6 hours for RGB-SOD tasks and 12 hours for COD tasks for 100 epochs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.218, + 0.713 + ], + "angle": 0, + "content": "4.2. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Comparison with RGB-D SOD Models: We present in Table 1 the performance on SOD benchmarks with our source-free depth or with GT depth. It can be seen that our model with source-free depth achieves very competitive performance compared to many RGB-D models with GT depths. Our method with GT depth also outperforms the SOTA counterparts. The qualitative comparison can be found in Fig. 6. Note that we also retrain the SOTA unimodal COD models SegMAR [21] and ZoomNet [47] on the SOD dataset only with RGB images. We show that our model with source-free depth outperforms these counterparts, showing that our method can better generalize across" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.31, + 0.892, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.454, + 0.892, + 0.496 + ], + "angle": 0, + "content": "Figure 6. Qualitative Comparison with GT Depth. Our method outperforms all counterparts while dealing with multi-objects, large depth variation, and visually-mixed foreground-background." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.506, + 0.828, + 0.521 + ], + "angle": 0, + "content": "different tasks with favorably better performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Comparison with COD Models: We present in Table 2 the performance of the most competitive SOTA methods, including task-specific COD models as well as retrained RGB-D SOD models with source-free depth. For a fair comparison, we retrain all RGB-D methods, SegMAR [21], and ZoomNet [47] in an end-to-end manner on the same resolution images as ours. Some RGB-only methods perform even better than many RGB-D methods, mainly due to their COD task-specific designs and the lack of ground-truth depths required for RGB-D methods. It is important to note that competing RGB-only COD methods do not perform favorably on SOD tasks. Please, refer to Table 1 & 3 to observe their poor cross-tasks generalization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Towards Higher Resolution: Previous studies have shown that the image resolution may influence the model performance [43, 47, 73, 83]. For example, the current SOTA COD method ZoomNet [47] is with main scale \\(384^2\\) and implies the highest resolution of \\((384 \\times 1.5)^2 = 576^2\\), as it operates on \\(0.5 \\times 1 \\times\\) and \\(1.5 \\times\\) scales. To make a fair comparison, we retrain the model with the same resolution \\((352^2\\) or \\(512^2)\\) as ours. We show in Table3-ZoomNet* that the results deteriorate as expected. Compared to these counterparts, our method offers a good trade-off between accuracy and efficiency. More comparisons on SOD and COD benchmarks can be found in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "1037" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.088, + 0.823, + 0.1 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on COD datasets. Pseudo stands for source-free depth used for RGB-D methods." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.101, + 0.895, + 0.348 + ], + "angle": 0, + "content": "
Pseudo Public.Dataset MetricCAMO [30]CHAMELEON [57]COD10K [9]NC4K [42]
M ↓Fm ↑Sm ↑Em ↑M ↓Fm ↑Sm ↑Em ↑M ↓Fm ↑Sm ↑Em ↑M ↓Fm ↑Sm ↑Em ↑
Performance of RGB COD Models
X CVPR20 [9]SINet.099.762.751.790.044.845.868.908.051.708.771.832.058.804.808.873
X CVPR21 [42]SLSR.080.791.787.843.030.866.889.938.037.756.804.854.048.836.839.898
X CVPR21 [78]MGL-R.088.791.775.820.031.868.893.932.035.767.813.874.053.828.832.876
X CVPR21 [44]PFNet.085.793.782.845.033.859.882.927.040.747.800.880.053.820.829.891
X CVPR21 [32]UJSC.072.812.800.861.030.874.891.948.035.761.808.886.047.838.841.900
X IJCAI21 [59]C2FNet.079.802.796.856.032.871.888.936.036.764.813.894.049.831.838.898
X ICCV21 [74]UGTR.086.800.783.829.031.862.887.926.036.769.816.873.052.831.839.884
X CVPR22 [21]SegMAR.080.799.794.857.032.871.887.935.039.750.799.876.050.828.836.893
X CVPR22 [47]ZoomNet.074.818.801.858.033.829.859.915.034.771.808.872.045.841.843.893
Performance of RGB-D Models Retrained with Source-free Depth
✓ MM21 [80]CDINet.100.638.732.766.036.787.879.903.044.610.778.821.067.697.793.830
✓ CVPR21 [19]DCF.089.724.749.834.037.821.850.923.040.685.766.864.061.765.791.878
✓ ICCV21 [81]CMINet.087.798.782.827.032.881.891.930.039.768.811.868.053.832.839.888
✓ ICCV21 [92]SPNet.083.807.783.831.033.872.888.930.037.776.808.869.054.828.825.874
✓ TIP22 [63]DCMF.115.737.728.757.059.807.830.853.063.679.748.776.077.782.794.820
✓ ECCV22 [31]SPSN.084.782.773.829.032.866.887.932.042.727.789.854.059.803.813.867
✓ OursPopNet.073.821.806.869.022.893.910.962.031.789.827.897.043.852.852.908
" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.365, + 0.154, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.478, + 0.138, + 0.49 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.155, + 0.365, + 0.226, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.478, + 0.202, + 0.488 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.365, + 0.301, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.478, + 0.282, + 0.489 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.365, + 0.374, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.309, + 0.478, + 0.37, + 0.489 + ], + "angle": 0, + "content": "ZoomNet" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.365, + 0.448, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.478, + 0.434, + 0.489 + ], + "angle": 0, + "content": "UGTR" + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.365, + 0.523, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.466, + 0.478, + 0.5, + 0.489 + ], + "angle": 0, + "content": "UJSC" + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.365, + 0.597, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.478, + 0.582, + 0.489 + ], + "angle": 0, + "content": "C2FNet" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.365, + 0.743, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.478, + 0.652, + 0.489 + ], + "angle": 0, + "content": "PFNet" + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.365, + 0.744, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.478, + 0.731, + 0.489 + ], + "angle": 0, + "content": "MGL-R" + }, + { + "type": "image", + "bbox": [ + 0.745, + 0.365, + 0.89, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.478, + 0.796, + 0.489 + ], + "angle": 0, + "content": "SLSR" + }, + { + "type": "image", + "bbox": [ + 0.835, + 0.365, + 0.89, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.835, + 0.478, + 0.872, + 0.489 + ], + "angle": 0, + "content": "SINet" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.491, + 0.891, + 0.52 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison. Our method can better preserve the object structure compared to other counterparts, especially while dealing with objects with occlusion (two first rows). Our method performs favorably with multiple objects (last row). Better to zoom in." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.53, + 0.469, + 0.571 + ], + "angle": 0, + "content": "Table 3. End-to-end comparison with different resolutions on SOD benchmarks. Our method with source-free depth generalizes significantly better compared to SOTA COD models." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.571, + 0.466, + 0.661 + ], + "angle": 0, + "content": "
ModelSizeFlops (G)NJUK [23]SIP [10]
M↓Fm↑Sm↑Em↑M↓Fm↑Sm↑Em↑
SegMAR [21]352267.3.036.921.909.941.052.893.872.914
ZoomNet [47]3522167.8.037.926.914.940.054.891.868.909
Ours3522228.8.031.931.920.949.046.911.885.926
SegMAR [21]5122142.4.035.927.914.943.050.899.878.917
ZoomNet [47]5122353.4.036.926.915.942.052.895.873.910
Ours5122484.0.031.933.922.951.044.911.890.927
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.47, + 0.786 + ], + "angle": 0, + "content": "Qualitative Comparisons: Figure 7 presents the output of our network on challenging cases. It can be seen while dealing with objects occluded by thin objects (\\(1^{st} - 4^{th}\\) rows), our method can accurately reason about the segmentation masks closer to the GT. We also achieve better performance while dealing with multiple objects (last row). More discussions on multiple objects can be found in Table 6." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.231, + 0.817 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Loss: In this section, we conduct experiments on analyzing the effectiveness of the proposed losses. The quantitative results of different loss combinations are provided in Table 4. It can be seen that each proposed loss behaves properly, i.e., improving the performance compared to the baseline. More" + }, + { + "type": "table_caption", + "bbox": [ + 0.554, + 0.53, + 0.836, + 0.543 + ], + "angle": 0, + "content": "Table 4. Ablation study on the proposed losses." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.543, + 0.887, + 0.673 + ], + "angle": 0, + "content": "
\\( \\mathcal{L}_{dep} \\)\\( \\mathcal{L}_{loc} \\)\\( \\mathcal{L}_{wtv} \\)\\( \\mathcal{L}_{sep} \\)SIP [10]NC4K [42]
\\( M\\downarrow \\)\\( F_m\\uparrow \\)\\( S_m\\uparrow \\)\\( E_m\\uparrow \\)\\( M\\downarrow \\)\\( F_m\\uparrow \\)\\( S_m\\uparrow \\)\\( E_m\\uparrow \\)
----.048.903.884.922.052.832.832.893
✓---.046.907.889.925.051.833.839.895
-✓--.045.908.893.929.048.837.844.898
--✓-.046.906.891.927.050.833.841.894
---✓.043.914.893.933.048.840.848.900
✓✓--.044.911.893.928.049.837.844.897
✓-✓-.046.909.893.927.046.840.845.898
✓-✓✓.040.918.897.935.045.848.849.904
✓✓-✓.042.916.894.931.044.850.850.906
✓✓✓✓.040.923.897.937.043.852.852.908
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.891, + 0.719 + ], + "angle": 0, + "content": "discussions and ablation studies on the hyperparameters can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Object Popping Network as Plug-in: Our object popping network can be easily adapted with different encoders and with different existing RGB-D models. For example, with a ResNet-18 [15] encoder and convolution-based decoder, our poping only costs around 12.7M additional learning parameters or 48.7 MB model size. We show in Table 5 that our method can favorably improve performance over the baseline with less than an extra \\(10\\%\\) GFlops." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Pop-out Under Reduced Training Data: Here, we are interested in analyzing source-free depth's benefits. Therefore, we conduct different experiments by reducing the training data. As shown in Figure 8(left), when both our" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "1038" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.079, + 0.088, + 0.468, + 0.166 + ], + "angle": 0, + "content": "Table 5. Generalization and cost over different RGB-D baselines. Dataset Metric Flops Param S1P [10] NC4K [42] M \\( \\downarrow \\) \\( {F}_{m} \\uparrow \\) \\( {S}_{m} \\uparrow \\) \\( {E}_{m} \\uparrow \\) \\( {M\\downarrow } \\) \\( {F}_{m} \\uparrow \\) \\( {S}_{m} \\uparrow \\) \\( {E}_{m} \\uparrow \\) +Ours 363.2 59.8 .053 .899 .874 .919 .057 .809 .804 .872 373.7 72.5 .051 .910 .886 .923 .055 .814 .811 .878 SPNet [92] 149.0 150.4 .044 .911 .887 .914 .054 .828 .825 .874 +Ours 159.5 163.1 .042 .917 .894 .932 .044 .851 .851 .905" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.179, + 0.468, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.329, + 0.47, + 0.399 + ], + "angle": 0, + "content": "Figure 8. Benefit of our method on reducing training data. When our network is trained only with \\(25\\%\\) data, the performance remains competitive compared to the baseline, i.e., the absolute performance gains are \\(+0.4\\%\\) max F-measure and \\(+0.1\\%\\) S-measure on NC4K compared to the baseline trained with all data." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.407, + 0.432, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.542, + 0.432, + 0.555 + ], + "angle": 0, + "content": "Figure 9. Histogram on the gain. Please zoom in for details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.47, + 0.716 + ], + "angle": 0, + "content": "PopNet and our baseline are trained with all data, our PopNet can lead to absolute improvements with \\(5.6\\%\\) in max F-measure and with \\(3.9\\%\\) in S-measure on COD10K dataset. While our PopNet is trained with only \\(25\\%\\) data, it can still achieve competitive performance compared to our baseline trained with all data. Similar phenomena can be observed in the NC4K dataset as shown in Figure 8(right). To conclude, our method can efficiently explore the geometric prior and significantly reduce the required training data volume." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Gain over baseline: With depth cues, as shown in Figure 9, we boost the performance in 3125 over 4121 images (\\(\\sim 75\\%\\) cases). We also show in Table 6 that our network performs favorably over baselines with single or multiple objects. Our method may fail when the source-free depth is clueless. This mainly happens when the object is well concealed and fools the depth network. However, such cases are also challenging, even for humans." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.84, + 0.468, + 0.855 + ], + "angle": 0, + "content": "Are RGB-D Methods Better Than RGB-only Methods?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "RGB-D methods are indeed better than RGB-only methods provided GT depth. However, only a few RGB-D methods can benefit from the source-free depth. For example," + }, + { + "type": "table_caption", + "bbox": [ + 0.501, + 0.089, + 0.888, + 0.103 + ], + "angle": 0, + "content": "Table 6. Multi-Object performance on NC4K [42] with size \\( {512}^{2} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.104, + 0.888, + 0.16 + ], + "angle": 0, + "content": "
Obj. Nbr. (%)MetricSingle (92%)Two (6%)More (2%)Overall
M↓Fm↑M↓Fm↑M↓Fm↑M↓Fm↑
RGB Baseline.054.828.067.811.091.738.056.825
+ Dxf.050.842.063.821.093.746.051.839
Ours.040.864.051.847.079.767.042.861
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.168, + 0.892, + 0.196 + ], + "angle": 0, + "content": "Table 7. Performance with RGB- \\(D_{sf}\\) model vs. with RGB-only baseline. \\(D_{sf}\\) stands for source-free depth." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.197, + 0.892, + 0.264 + ], + "angle": 0, + "content": "
DatasetCOD10K [9]NC4K [42]
MetricM↓Fm↑Sm↑Em↑M↓Fm↑Sm↑Em↑
DASNet [88].041.643.793.864.055.747.830.879
+ Dsf.041.642.796.858.055.743.830.874
SPNet [92].040.743.801.867.052.846.833.883
+ Dsf.037.776.808.869.054.828.825.874
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.893, + 0.516 + ], + "angle": 0, + "content": "as shown in Table 7, DASNet [88] achieves poorer performance compared to RGB baseline when trained with source-free depth. Similarly, even for one of the SOTA RGB-D models, SPNet [92], the performance on NC4K dataset with RGB-only input is better than with additional source-free depth. Moreover, when provided with source-free depth, none of the existing RGB-D methods outperform the best performing RGB-only method (e.g., ZoomNet [47]) on COD. The same observation was also made on SOD as well. This could be because of the domain gap coupled with the fusion design, among others. We also found it non-trivial to extend the most well-performing RGB-only methods to the RGB-D case. Note that our PopNet performs better than all existing RGB-only and RGB-D methods, with source-free or GT depth maps." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.538, + 0.619, + 0.554 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.892, + 0.807 + ], + "angle": 0, + "content": "We demonstrate a successful case of cross-domain cross-task depth to semantics knowledge transfer using only the source model. In this paper, the source-free depth at the target offered by a given source model is used. The proposed method learns to transfer knowledge from depth to semantics using the objects' pop-out prior. We facilitate our network to use such prior by designing a novel network architecture. The designed network reasons about the objects by popping them out from the provided depth maps. This process is followed by separating objects from the background using the learned contact surface. We show that the joint learning of object pop-out and contact surface can be successfully supervised using the target semantics. Exhaustive experiments on SOD and COD benchmarks show the successful transfer of depth knowledge to the target, in terms of improved performance and generalization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement The authors thank the anonymous reviewers and ACs for their tremendous efforts and helpful comments. This research is financed in part by the Conseil Régional de Bourgogne-Franche-Comté, Toyota Motor Europe (research project TRACE-Zurich), and the Alexander von Humboldt Foundation." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "1039" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Peshal Agarwal, Danda Pani Paudel, Jan-Nico Zaech, and Luc Van Gool. Unsupervised robust domain adaptation without source data. In WACV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.471, + 0.213 + ], + "angle": 0, + "content": "[2] Amir Atapour-Abarghouei and Toby P Breckon. Real-time monocular depth estimation using synthetic data with domain adaptation via image style transfer. In IEEE CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.471, + 0.271 + ], + "angle": 0, + "content": "[3] Adriano Cardace, Luca De Luigi, Pierluigi Zama Ramirez, Samuele Salti, and Luigi Di Stefano. Plugging self-supervised monocular depth into unsupervised domain adaptation for semantic segmentation. In WACV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.471, + 0.355 + ], + "angle": 0, + "content": "[4] Po-Yi Chen, Alexander H Liu, Yen-Cheng Liu, and Yu-Chiang Frank Wang. Towards scene understanding: Unsupervised monocular depth estimation with semantic-aware representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.357, + 0.47, + 0.412 + ], + "angle": 0, + "content": "[5] Innes C Cuthill, Martin Stevens, Jenna Sheppard, Tracey Maddocks, C Alejandro Parraga, and Tom S Troscianko. Disruptive coloration and background pattern matching. Nature, 434(7029):72-74, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.414, + 0.47, + 0.455 + ], + "angle": 0, + "content": "[6] Jiahua Dong, Zhen Fang, Anjin Liu, Gan Sun, and Tongliang Liu. Confident anchor-induced multi-source free domain adaptation. NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.457, + 0.47, + 0.512 + ], + "angle": 0, + "content": "[7] Aliya El Nagar, Daniel Osorio, Sarah Zylinski, and Steven M Sait. Visual perception and camouflage response to 3d backgrounds and cast shadows in the european cuttlefish, sepia officinalis. JEB, 224(11), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.514, + 0.47, + 0.555 + ], + "angle": 0, + "content": "[8] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE TPAMI, 44(10):6024-6042, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.557, + 0.47, + 0.599 + ], + "angle": 0, + "content": "[9] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In IEEE CVPR, 2020. 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.47, + 0.655 + ], + "angle": 0, + "content": "[10] Deng-Ping Fan, Zheng Lin, Zhao Zhang, Menglong Zhu, and Ming-Ming Cheng. Rethinking RGB-D salient object detection: Models, datasets, and large-scale benchmarks. IEEE TNNLS, 32(5):2075-2089, 2021. 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.469, + 0.697 + ], + "angle": 0, + "content": "[11] Deng-Ping Fan, Jing Zhang, Gang Xu, Ming-Ming Cheng, and Ling Shao. Salient objects in clutter. IEEE TPAMI, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.47, + 0.742 + ], + "angle": 0, + "content": "[12] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In IEEE CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[13] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In IEEE ICCV, 2019. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.47, + 0.828 + ], + "angle": 0, + "content": "[14] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In IEEE CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE CVPR, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[16] Ian P Howard. Perceiving in depth, volume 1: basic mechanisms. Oxford University Press, 2012. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[17] Junjie Hu, Mete Ozay, Yan Zhang, and Takayuki Okatani. Revisiting single image depth estimation: Toward higher resolution maps with accurate object boundaries. In WACV, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[18] Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection. MIR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[19] Wei Ji, Jingjing Li, Shuang Yu, Miao Zhang, Yongri Piao, Shunyu Yao, Qi Bi, Kai Ma, Yefeng Zheng, Huchuan Lu, et al. Calibrated RGB-D salient object detection. In IEEE CVPR, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[20] Wei Ji, Jingjing Li, Miao Zhang, Yongri Piao, and Huchuan Lu. Accurate RGB-D salient object detection via collaborative learning. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[21] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In IEEE CVPR, 2022, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.361, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[22] Wen-Da Jin, Jun Xu, Qi Han, Yi Zhang, and Ming-Ming Cheng. Cdnet: Complementary depth network for rgb-d salient object detection. IEEE TIP, 30:3376-3390, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[23] Ran Ju, Ling Ge, Wenjing Geng, Tongwei Ren, and Gangshan Wu. Depth saliency based on anisotropic centersurround difference. In IEEE ICIP, 2014. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[24] Hongwen Kang, Alexei A Efros, Martial Hebert, and Takeo Kanade. Image composition for object pop-out. In IEEE ICCVW, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[25] Youngeun Kim, Donghyeon Cho, Kyeongtak Han, Priyadarshini Panda, and Sungeun Hong. Domain adaptation without source data. IEEE TAI, 2(6):508-518, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[26] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In IEEE CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[27] Jogendra Nath Kundu, Akshay R Kulkarni, Suvaansh Bhambri, Deepesh Mehta, Shreyas Anand Kulkarni, Varun Jampani, and Venkatesh Babu Radhakrishnan. Balancing discriminability and transferability for source-free domain adaptation. In ICML, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[28] Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In IEEE CVPR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[29] Vinod K Kurmi, Venkatesh K Subramanian, and Vinay P Namboodiri. Domain impression: A source data free domain adaptation method. In WACV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[30] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, MinhTriet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. CVU, 184:45-56, 2019. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.787, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[31] Minhyeok Lee, Chaewon Park, Suhwan Cho, and Sangyoun Lee. Spsn: Superpixel prototype sampling network for rgb-d salient object detection. In ECCV, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.83, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[32] Aixuan Li, Jing Zhang, Yunqiu Lv, Bowen Liu, Tong Zhang, and Yuchao Dai. Uncertainty-aware joint salient object and camouflaged object detection. In IEEE CVPR, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[33] Gongyang Li, Zhi Liu, Minyu Chen, Zhen Bai, Weisi Lin, and Haibin Ling. Hierarchical alternate interaction network" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.926, + 0.516, + 0.937 + ], + "angle": 0, + "content": "1040" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "forrgb-d salient object detection.IEEE TIP,30:3528-3542, 2021.3,6,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.162 + ], + "angle": 0, + "content": "[34] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In IEEE CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[35] Rui Li, Danna Xue, Shaolin Su, Xiantuo He, Qing Mao, Yu Zhu, Jinqiu Sun, and Yanning Zhang. Learning depth via leveraging semantics: Self-supervised monocular depth estimation with both implicit and explicit semantic guidance. Pattern Recognition (PR), page 109297, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[36] Xinhao Li, Jingjing Li, Lei Zhu, Guoqing Wang, and Zi Huang. Imbalanced source-free domain adaptation. In ACM MM, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.319 + ], + "angle": 0, + "content": "[37] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.468, + 0.361 + ], + "angle": 0, + "content": "[38] Jiang-Jiang Liu, Qibin Hou, Ming-Ming Cheng, Jiashi Feng, and Jianmin Jiang. A simple pooling-based design for real-time salient object detection. In IEEE CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.468, + 0.403 + ], + "angle": 0, + "content": "[39] Yuang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. In IEEE CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.468, + 0.446 + ], + "angle": 0, + "content": "[40] Yukang Lu, Dingyao Min, Keren Fu, and Qijun Zhao. Depth-cooperated trimodal network for video salient object detection. In ICIP. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[41] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM TOG, 39(4):1-13, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.545 + ], + "angle": 0, + "content": "[42] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In IEEE CVPR, 2021. 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.603 + ], + "angle": 0, + "content": "[43] Daniel McKee, Zitong Zhan, Bing Shuai, Davide Modolo, Joseph Tighe, and Svetlana Lazebnik. Transfer of representations to video label propagation: Implementation factors matter. arXiv preprint arXiv:2203.05553, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[44] Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan. Camouflaged object segmentation with distraction mining. In IEEE CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[45] S Mahdi H Miangoleh, Sebastian Dille, Long Mai, Sylvain Paris, and Yagiz Aksoy. Boosting monocular depth estimation models to high-resolution via content-adaptive multi-resolution merging. In IEEE CVPR, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[46] Yuzhen Niu, Yujie Geng, Xueqing Li, and Feng Liu. Leveraging stereopsis for saliency analysis. In IEEE CVPR, 2012. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[47] Youwei Pang, Xiaqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In IEEE CVPR, 2022. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[48] Houwen Peng, Bing Li, Weihua Xiong, Weiming Hu, and Rongrong Ji. RGBD salient object detection: a benchmark and algorithms. In ECCV, 2014. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[49] Yongri Piao, Zhengkun Rong, Miao Zhang, Weisong Ren, and Huchuan Lu. A2dele: Adaptive and attentive depth distiller for efficient RGB-D salient object detection. In IEEE CVPR, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[50] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In IEEE ICCV, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[51] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI, 44(3):1623-1637, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[52] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[53] Suman Saha, Anton Obukhov, Danda Pani Paudel, Menelaos Kanakis, Yuhua Chen, Stamatios Georgoulis, and Luc Van Gool. Learning to relate depth and semantics for unsupervised domain adaptation. In IEEE CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[54] Roshni Sahoo, Divya Shanmugam, and John Guttag. Unsupervised domain adaptation in the absence of source data. arXiv preprint arXiv:2007.10233, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[55] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE CVPR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[56] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from RGBD images. In ECCV, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[57] Przemysław Skurowski, Hassan Abdulameer, J Blaszczyk, Tomasz Depta, Adam Kornacki, and P Koziel. Animal camouflage analysis: Chameleon database. Unpublished manuscript, 2(6):7, 2018. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[58] Hwanjun Song, Eunyoung Kim, Varun Jampan, Deqing Sun, Jae-Gil Lee, and Ming-Hsuan Yang. Exploiting scene depth for object detection with multimodal transformers. In BMVC, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[59] Yujia Sun, Geng Chen, Tao Zhou, Yi Zhang, and Nian Liu. Context-aware cross-level fusion network for camouflaged object detection. In *IJCAI*, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[60] Yujia Sun, Shuo Wang, Chenglizhao Chen, and Tian-Zhu Xiang. Boundary-guided camouflaged object detection. In IJ-CAI, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[61] Alessio Tonioni, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Unsupervised domain adaptation for depth prediction from images. IEEE TPAMI, 42(10):2396-2409, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[62] Anne Treisman. Preattentive processing in vision. ICVGIP, 31(2):156-177, 1985. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[63] Fengyun Wang, Jinshan Pan, Shoukun Xu, and Jinhui Tang. Learning discriminative cross-modality features for RGB-D saliency detection. IEEE TIP, 31:1285-1297, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[64] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In IEEE CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[65] Cho-Ying Wu, Jialiang Wang, Michael Hall, Ulrich Neumann, and Shuochen Su. Toward practical monocular indoor depth estimation. In IEEE CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[66] Yu-Huan Wu, Yun Liu, Jun Xu, Jia-Wang Bian, Yu-Chao Gu, and Ming-Ming Cheng. Mobilesal: Extremely efficient rgb-d salient object detection. IEEE TPAMI, 44(12):10261-10269, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[67] Zongwei Wu, Guillaume Allibert, Christophe Stolz, and Cédric Demonceaux. Depth-adapted CNN for RGB-D cameras. In ACCV, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.926, + 0.514, + 0.937 + ], + "angle": 0, + "content": "1041" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[68] Zongwei Wu, Guillaume Allibert, Christophe Stolz, Chao Ma, and Cédric Demonceaux. Modality-guided subnetwork for salient object detection. In 3DV, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[69] Zongwei Wu, Shriarulmozhivarman Gobichettipalayam, Brahim Tamadazte, Guillaume Allibert, Danda Pani Paudel, and Cedric Demonceaux. Robust RGB-D fusion for saliency detection. 3DV, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[70] Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In IEEE CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.232, + 0.47, + 0.273 + ], + "angle": 0, + "content": "[71] Haifeng Xia, Handong Zhao, and Zhengming Ding. Adaptive adversarial network for source-free domain adaptation. In IEEE ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[72] Ke Xian, Jianming Zhang, Oliver Wang, Long Mai, Zhe Lin, and Zhiguo Cao. Structure-guided ranking loss for single image depth prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.47, + 0.398 + ], + "angle": 0, + "content": "[73] Mochu Xiang, Jing Zhang, Yunqiu Lv, Aixuan Li, Yiran Zhong, and Yuchao Dai. Exploring depth contribution for camouflaged object detection. arXiv e-prints, pages arXiv-2106, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.47, + 0.453 + ], + "angle": 0, + "content": "[74] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping Fan. Uncertainty-guided transformer reasoning for camouflaged object detection. In IEEE ICCV, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.47, + 0.496 + ], + "angle": 0, + "content": "[75] Shiqi Yang, Joost van de Weijer, Luis Herranz, Shangling Jui, et al. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.47, + 0.538 + ], + "angle": 0, + "content": "[76] Shiqi Yang, Yaxing Wang, Joost van de Weijer, and Luis Herranz. Unsupervised domain adaptation without source data by casting a bait. arXiv preprint arXiv:2010.12427, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.47, + 0.58 + ], + "angle": 0, + "content": "[77] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In IEEE ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.47, + 0.622 + ], + "angle": 0, + "content": "[78] Qiang Zhai, Xin Li, Fan Yang, Chenglizhao Chen, Hong Cheng, and Deng-Ping Fan. Mutual graph learning for camouflaged object detection. In IEEE CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.47, + 0.663 + ], + "angle": 0, + "content": "[79] Wei Zhai, Yang Cao, HaiYong Xie, and Zheng-Jun Zha. Deep texton-coherence network for camouflaged object detection. IEEE TMM, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[80] Chen Zhang, Runmin Cong, Qinwei Lin, Lin Ma, Feng Li, Yao Zhao, and Sam Kwong. Cross-modality discrepant interaction network for RGB-D salient object detection. In ACM MM, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[81] Jing Zhang, Deng-Ping Fan, Yuchao Dai, Xin Yu, Yiran Zhong, Nick Barnes, and Ling Shao. RGB-D saliency detection via cascaded mutual information minimization. In IEEE ICCV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.816 + ], + "angle": 0, + "content": "[82] Jing Zhang, Yunqiu Lv, Mochu Xiang, Aixuan Li, Yuchao Dai, and Yiran Zhong. Depth-guided camouflaged object detection. CoRR, abs/2106.13217, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.859 + ], + "angle": 0, + "content": "[83] Miao Zhang, Shuang Xu, Yongri Piao, Dongxiang Shi, Shusen Lin, and Huchuan Lu. Preynet: Preying on camouflaged objects. In ACM MM, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[84] Pingping Zhang, Dong Wang, Huchuan Lu, Hongyu Wang, and Baocai Yin. Learning uncertain convolutional features for accurate saliency detection. In IEEE ICCV, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[85] Wenbo Zhang, Ge-Peng Ji, Zhuo Wang, Keren Fu, and Qijun Zhao. Depth quality-inspired feature manipulation for efficient RGB-D salient object detection. In ACM MM, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.19 + ], + "angle": 0, + "content": "[86] Zhoutong Zhang, Forrester Cole, Richard Tucker, William T Freeman, and Tali Dekel. Consistent depth of moving objects in video. ACM TOG, 40(4):1-12, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.893, + 0.246 + ], + "angle": 0, + "content": "[87] Zhao Zhang, Zheng Lin, Jun Xu, Wen-Da Jin, Shao-Ping Lu, and Deng-Ping Fan. Bilateral attention network forrgb-d salient object detection. IEEE TIP, 30:1949-1961, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.893, + 0.289 + ], + "angle": 0, + "content": "[88] Jiawei Zhao, Yifan Zhao, Jia Li, and Xiaowu Chen. Is depth really necessary for salient object detection? In ACM MM, 2020. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.893, + 0.333 + ], + "angle": 0, + "content": "[89] Shanshan Zhao, Huan Fu, Mingming Gong, and Dacheng Tao. Geometry-aware symmetric domain adaptation for monocular depth estimation. In IEEE CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.893, + 0.376 + ], + "angle": 0, + "content": "[90] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In IEEE CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.893, + 0.418 + ], + "angle": 0, + "content": "[91] Tao Zhou, Deng-Ping Fan, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Rgb-d salient object detection: A survey. CVMJ, pages 1-33, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.893, + 0.46 + ], + "angle": 0, + "content": "[92] Tao Zhou, Huazhu Fu, Geng Chen, Yi Zhou, Deng-Ping Fan, and Ling Shao. Specificity-preserving RGB-D saliency detection. In IEEE ICCV, 2021. 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.893, + 0.503 + ], + "angle": 0, + "content": "[93] Zikun Zhou, Wenjie Pei, Xin Li, Hongpeng Wang, Feng Zheng, and Zhenyu He. Saliency-associated object tracking. In IEEE ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.893, + 0.56 + ], + "angle": 0, + "content": "[94] Hongwei Zhu, Peng Li, Haoran Xie, Xuefeng Yan, Dong Liang, Dapeng Chen, Mingqiang Wei, and Jing Qin. I can find you! boundary-guided separated attention network for camouflaged object detection. In AAAI, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.893, + 0.603 + ], + "angle": 0, + "content": "[95] Jinchao Zhu, Xiaoyu Zhang, Shuo Zhang, and Junnan Liu. Inferring camouflaged objects by texture-aware interactive guidance network. In AAAI, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.603 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "1042" + } + ] +] \ No newline at end of file diff --git a/2023/Source-free Depth for Object Pop-out/1ac228ed-431b-40e3-b444-e5a574a99abb_origin.pdf b/2023/Source-free Depth for Object Pop-out/1ac228ed-431b-40e3-b444-e5a574a99abb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1285c9c0808ec101481d40fbecd5b42d71d86670 --- /dev/null +++ b/2023/Source-free Depth for Object Pop-out/1ac228ed-431b-40e3-b444-e5a574a99abb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b861a6dcf4b2934574c452d1b575386e792013598c115df7f16b186536be9265 +size 3862487 diff --git a/2023/Source-free Depth for Object Pop-out/full.md b/2023/Source-free Depth for Object Pop-out/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d7f1ddb1f083170f36d2ed0f0717504996e4ace0 --- /dev/null +++ b/2023/Source-free Depth for Object Pop-out/full.md @@ -0,0 +1,368 @@ +# Source-free Depth for Object Pop-out + +Zongwei Wu $^{1,2,3}$ Danda Pani Paudel $^{1,4}$ Deng-Ping Fan $^{1*}$ Jingjing Wang $^{5}$ Shuo Wang $^{1}$ Cédric Demonceaux $^{2,6}$ Radu Timofte $^{3}$ Luc Van Gool $^{1,4}$ + +$^{1}$ CVL, ETH Zurich $^{2}$ University of Burgundy, CNRS, ICB $^{3}$ Computer Vision Lab, CAIDAS & IFI, University of Wurzburg $^{4}$ INSAIT, Sofia University $^{5}$ AUST $^{6}$ University of Lorraine, CNRS, Inria, Loria + +# Abstract + +Depth cues are known to be useful for visual perception. However, direct measurement of depth is often impracticable. Fortunately, though, modern learning-based methods offer promising depth maps by inference in the wild. In this work, we adapt such depth inference models for object segmentation using the objects' "pop-out" prior in 3D. The "pop-out" is a simple composition prior that assumes objects reside on the background surface. Such compositional prior allows us to reason about objects in the 3D space. More specifically, we adapt the inferred depth maps such that objects can be localized using only 3D information. Such separation, however, requires knowledge about contact surface which we learn using the weak supervision of the segmentation mask. Our intermediate representation of contact surface, and thereby reasoning about objects purely in 3D, allows us to better transfer the depth knowledge into semantics. The proposed adaptation method uses only the depth model without needing the source data used for training, making the learning process efficient and practical. Our experiments on eight datasets of two challenging tasks, namely salient object detection and camouflaged object detection, consistently demonstrate the benefit of our method in terms of both performance and generalizability. The source code is publicly available at https://github.com/Zongwei97/PopNet. + +# 1. Introduction + +The 3D knowledge of the scene is long known to be complementary to the task of visual perception [12, 16, 56, 67, 91]. Often in practice, though, visual perception needs to be carried out using only 2D images. Given multiple images, 3D geometry may be recovered using the structure-from-motion techniques [26, 41, 55, 86]. Such inversion, however, is not compatible when only a single image is available. Under such circumstances, image inversion + +![](images/bc00a305be8a8a03db526ed5dc79a002452bdf98338452b280083f379e51b4fb.jpg) +Figure 1. Depth to semantics conversion using the object pop-out prior. For input RGB and source-free depth pair, we learn contact surface. The obtained contact surface is then used to separate objects and backgrounds to derive pseudo semantics for supervision. + +to depth map is usually done using learning-based methods [13,45,50,51,65], which have shown unparalleled success in the recent years. Unfortunately, the learning-based methods may not offer high-quality depth maps due to the generalization deficiency across domains [68, 73]. + +Despite poor generalization, the knowledge gained in one domain is shown to be useful in other close-by domains. This utility is harnessed by performing the so-called domain adaptation (DA) [2,3,6,53,61,89]. In fact, it has been shown recently that DA methods can efficiently transfer knowledge using only the prediction models, i.e., without requiring access to the data where the model is trained – also known as the source-free domain adaptation (SDA) [25,28,39,71,77]. The SDA methods are of gripping interest due to their efficiency and privacy promises. + +The most existing SDA methods make one or both of these implicit assumptions: (a) a similar (as that of the source) supervising task at the target [28, 75, 77]; (b) task of discrete (and known) label space [27, 29, 36, 71]. The former assumption not only makes the source and target domains easier to compare but also potentially keeps the two domains closer. The latter assumption allows performing SDA by self-training where the discrete labels facilitate reasoning about the model's confidence. The self-training is then performed by boosting the confidence at the target for some picked reliable examples. + +![](images/f208f75faa12ac29e7bda8c0fdd583158bc75bf9757271f90a8896e8d21f815a.jpg) +Cross-Task Cross-Dataset Analysis with Depth Pop-out Prior +Figure 2. The performance gained in F-measure using our method over the established baselines. Our method $(\triangle, \square)$ significantly improves baselines on 8 datasets and 2 tasks (each task on 4 datasets- SOD: left two; COD: right two). We also compare 24 methods $(\bullet)$ , where our method offers state-of-the-art results despite their task specialization. Note that all methods are connected using lines to illustrate their performance fluctuations across datasets. Please, refer Tables 1 & 2 and Section 4.3 for more details and discussions. + +In this work, we aim for source-free transfer of depth knowledge for object detection. Such transfer is desired to assist in locating the object by depth cues and to exploit the depth knowledge despite the domain gap. The addressed problem setting differs from the standard SDA in terms of (a) the source and target tasks' difference; (b) and the continuous label space of the depth. These differences (with standard SDA) make our task at hand very challenging, which is addressed for the first time in this paper, up to our knowledge. To address such a challenging problem, we rely on the "pop-out" prior, which allows us to reason about the object's location directly in the 3D space. The "pop-out" is a simple composition prior that assumes objects reside on the background surface. A graphical illustration of the used "pop-out" prior, using the results obtained by our method, is given in Figure 1. + +The pop-out prior for image composition was successfully used by Kang et al. in [24]. An early work of Treisman has provided an in-depth study of such prior in [62]. In this work, we rely on the same compositional foundation of these works and exploit the pop-out before transferring the depth of knowledge across domains. Although our motivation comes from these early works, our experimental setup largely differs from theirs. We differ in terms of not only depth knowledge transfer across domains (without source data) but also in target supervision using only semantics. + +The proposed method exploits the source-free depth to map it into a space where objects in depth stand out better against the background. This mapping is used for object and background separation using a learned contact surface between them. Such separation allows us to derive the semantic masks which can be directly compared against the ground truth for supervision. Using this supervision at the target, we can minimize the domain and task gap between the source and target. The overall framework of our method that performs cross-task cross-domain knowledge transfer by using the intermediate representation of the pop-out space is shown in Figure 3. As can be seen, we first leverage an object popping network to encourage the object + +to jump out from the source-free depth. Then, we introduce another network i.e., the segmentation with contact surface, to localize the object and predict the contact surface. These learning modules are jointly trained in an end-to-end manner, transferring the source-free depths into intermediate representations which are adapted to the target task, i.e., object detection. To evaluate the proposed method, we conducted exhaustive experiments on eight datasets of two challenging tasks, namely salient object detection and camouflaged object detection. In both tasks, our method significantly improves the established baselines and offers state-of-the-art results at the same time, whose overview can be seen in Figure 2. The major contributions are as follows: + +- Our problem of transferring source-free depth knowledge across domains and tasks is practical and novel. +- Our method relies on our object pop-out prior for visual understanding, which is simple and effective. +- Results of our method in two different tasks are significantly better than the baselines and existing models. + +# 2. Related Works + +Source-free Adaptation: Knowledge transfer by domain adaptation without access to the source data has recently gathered vast interest [1,25,28,39,71,77], due to the privacy, practicality, and efficiency reasons. We observe accessing the source data for transferring the depth knowledge learned by the off-the-shelf knowledge models [13, 51] is particularly impractical due to their multi-stage training on various datasets. The existing source-free adaptation methods either use generative [28, 29, 34], pseudo-label [25, 37, 64], or other customized [54, 76] approaches. In this work, we use the pseudo-label-based approach. However, using existing methods is not straightforward in our setting due to the tasks' difference between source and target. With our pop-out technique that provides the pseudo semantics, the source-free depth is better transferred across tasks. + +Salient Object Detection (SOD): Saliency detection aims to detect and segment the most prominent region within an image that visually attracts human attention [11]. A + +![](images/451fab1cda9c67f92d3f0a55b31ef68d5b893ceeda5b7501ee15edbedad854ef.jpg) +Figure 3. Our proposed framework, termed PopNet, is composed of a source-free network, an object popping network, a segmentation with contact surface, and an object separation module. The source-free depth network generates pseudo-depth in an off-the-shelf manner (Section 3.1). The object popping network converts the source-free depth into the popped-out depth of objects, bridging the cross-domain and cross-task gaps (Section 3.2). The segmentation network uses this depth to estimate the object's mask and contact surface (Section 3.3). The object separation module then converts popped-out depth to the second mask of objects using the contact surface (Section 3.4). We compare both semantic masks against ground truth to supervise the whole pipeline in an end-to-end manner. + +number of works have shown that saliency can be an auxiliary step for different vision tasks such as object tracking [93], object detection [58], etc. Conventional saliency works are unimodal, i.e., they only require RGB images as input. In generic and common settings, RGB-based models [38, 70, 84] have already achieved very promising results. More recently, several works [10,20,49,66,68,88,92] exploit depth maps as additional clues to the 3D geometry since the depth can provide more truthful information on the object boundary as well as the scale awareness. These 3D features further improve the detection accuracy and performance in challenging scenarios [22, 33, 69, 80]. + +Camouflaged Object Detection (COD): Camouflage detection aims to find the preying object within an image. For computer vision society, primary works [9,32,47] often compare COD with SOD. A number of works [8,9,74,79] have shown that simply extending saliency models [70] on COD will lead to undesired results, which is mainly caused by the nature of target object, i.e., concealed or prominent. Hence, to constrain the attention on the concealed objects, several works come up with different perceptual systems that mimic human behavior vis-a-vis camouflaged objects, such as three-stage localize-segment-rank strategy [42]; iterative refinement [21], which is similar to repeatedly looking on the images; zooming into possible regions [47, 59]. Others [18, 40, 60, 82, 83, 90, 94, 95] deeply explore the texture difference with the help of the gradient [18], frequency [90], edges [60, 94], and probability [32, 74]. + +The latest psychological studies [5, 7] have shown that human perception can naturally benefit from the depth cues to understand the scene: (A) the smooth variation within the object can contribute to alleviating the fake edges and preserving the object structure; (B) the depth discontinuity on the object boundary can make the segmentation easier. In + +spired by these observations, we aim to explore the source-free depth for both SOD and COD tasks. To tackle the domain gap for source-free depth maps, we propose to jointly finetune the source-free depth together with the semantic network in an end-to-end manner, with both self-supervised loss and weak semantic supervision. + +# 3. Proposed PopNet + +Given an input RGB image $I$ with size $I \in \mathbb{R}^{3 \times H \times W}$ , where $H$ and $W$ are the spatial resolutions, i.e., height and width, of the image, our objective is to predict the semantic mask $\tilde{S} \in \mathbb{R}^{H \times W}$ for object detection. As shown in Figure 3, the input image $I$ is firstly fed into a frozen-weight depth network to generate the source-free depth $D_{sf} \in \mathbb{R}^{H \times W}$ (Section 3.1). Then the mimicked multi-modal images are fed together into the depth popping network to compute the intermediate popped-out depth $D_{po} \in \mathbb{R}^{H \times W}$ (Section 3.2). This intermediate representation, as well as the RGB image $I$ , is later processed by the segmentation network and transformed into a contact surface $D_{c} \in \mathbb{R}^{H \times W}$ and a semantics prediction $\tilde{S} \in \mathbb{R}^{H \times W}$ (Section 3.3). On the one hand, the semantics prediction is directly supervised by the ground truth mask $GT$ , denoted as $G$ , which is similar to conventional segmentation supervision. On the other hand, we further explore the contact surface to pop the object out of the background by means of our object separation module (Section 3.4). This transfers the geometric cues into pseudo semantics and leads to another level of supervision. + +# 3.1. Source-free Depth Network + +In a practical setting, the GT depth is not always available. Therefore, we generate the source-free depth $D_{sf}$ in an off-the-shelf manner to mimic the multi-modal input. + +We choose the state-of-the-art DPT model [50] with frozen weights as our source-free depth network, which offers us promising depth at the target. This choice is made upon its generalization capability as suggested in [51]. To obtain the depth map with the highest quality possible, by enhancing local details, we apply the boosting method [45] together with DPT. Despite the plausible results achieved by learning-based methods, the obtained source-free depth does not always offer high-quality geometric cues due to the domain gap. Therefore, we leverage the geometric and semantic priors to jointly finetune the source-free depth. + +# 3.2. Object Popping Network + +Network Architecture: We build a depth popping network to refine/smooth the source-free depth. The popping network follows the encoder-decoder design with skip connection as shown in Figure 4. In our case, we simply concatenate RGB and source-free depth at the input side to form a 4-channel input and feed them into the popping network. The encoder extracts semantic cues and generates five-scale outputs. Our decoder is composed of Conv2D, BN, ReLU, and upsampling layers. Following U-Net [52], we build a skip connection through simple addition. + +Structure Preserving: To supervise our popping network, we first guide the depth refinement with the help of generated pseudo-depth. We only leverage structural similarity since we aim to detect, preserve, and extract object structure from the intermediate representation. We use the following SSIM loss [13] for structural similarity. + +$$ +\mathcal {L} _ {d e p} = S S I M \left(D _ {p o}, D _ {s f}\right). \tag {1} +$$ + +Local Depth Smoothing: In addition to the pseudo-depth supervision, we propose two losses to constrain the depth also by semantics. We assume that the objects' structure should be distinguishable from the background, i.e., it should be smooth within the object region and sharp on the bounding pixels. Hence, we propose to leverage the weak semantic cues together with the geometric priors. Specifically, we first introduce a local smoothness loss. The locality is defined by the ground truth semantics $G$ . Technically, we mask out background pixels with element-wise multiplication to suppress the inactive area through $D_{obj} = D_{po} \otimes G$ . Let $\nabla_x$ and $\nabla_y$ be the Sobel operations. Then, our local loss $\mathcal{L}_{loc}$ is expressed as: + +$$ +\begin{array}{c} \overrightarrow {n (p)} = (- \nabla_ {x} (D _ {o b j} (p)); - \nabla_ {y} (D _ {o b j} (p)); 1); \\ \mathcal {L} _ {l o c} = \sum_ {p} \sum_ {q \in \mathcal {N} (p)} 1 - \operatorname {c o s i n e} (\overrightarrow {n (p)}; \overrightarrow {n (q)}), \end{array} \tag {2} +$$ + +where $\overrightarrow{n}$ stands for the normal, $p$ denotes the pixels within the object region, $\mathcal{N}(p)$ is the neighboring pixels, and cosine is the cosine similarity between two vectors. This way, our + +![](images/9e4cfb368af124dd46d4911338806fb8feac354156afccce5a7ffabc2555d4bb.jpg) +Figure 4. Our object popping network maps RGB-D inputs into popped-out depths. This network is supervised using a combination of structure preserving, local depth smoothing, and depth edge sharpening losses $\mathcal{L}_{dep}$ , $\mathcal{L}_{loc}$ , and $\mathcal{L}_{wtv}$ , respectively. + +local loss only works on the object area, making the object structure consistent within the target region. Applying local smoothness loss reduces the depth noise at the object level. + +Depth Edge Sharpening: In addition to the local smoothness, we also use edge sharpening. The edge sharpening loss is formulated as a weighted total variation. For this, we first compute the edge-aware weight $w(p)$ at any pixel $p$ as, + +$$ +w (p) = \left\{ \begin{array}{l l} w _ {0}, & \text {i f} \nabla_ {x} (G (p)) ^ {2} + \nabla_ {y} (G (p)) ^ {2} \neq 0, \\ w _ {0} + \gamma , & \text {o t h e r w i s e}, \end{array} \right. \tag {3} +$$ + +where $w_0$ is a pre-defined non-zero weight and $\gamma$ is an additional weight for boundary pixels. In our setting, we choose $w_0$ as the normalized (by the image size) count of the boundary pixels, and we set $\gamma = 0.5$ . We adopt the square form such that the large gradients play more important roles. Our weighted total-variation loss is given by, + +$$ +\mathcal {L} _ {w T v} = \sum_ {p} \sum_ {q \in \mathcal {N} (p)} w (p) \cdot | | D _ {p o} (p) - D _ {p o} (q) | | _ {2}. \tag {4} +$$ + +Our weighted total-variance loss differs from the conventional edge loss, due to our weighting function. More specifically, our weighting function relies on semantic boundaries, unlike the commonly used image gradients [14, 17]. Our motivation for using semantic boundaries instead of image gradients comes from our interest in performing object detection under challenging conditions, such as camouflaged objects. In such cases, image gradients may result in misleading weights. At first glance, our loss function seems to be similar to semantic-guided depth estimation methods [4, 35, 72]. However, [72] uses GT depths, and [4, 35] use multi-frames for supervision. Unfortunately, such supervision is not possible in our setup. Note that we exploit single-view source-free depth while only using semantic ground truth for supervision. Furthermore, we transfer the source-free depths' knowledge despite their domain gap. Existing regularizations between depth and semantics are already exploited in our method, which is shown to be + +complimentary to our pop-out prior. We argue that our network architecture that exploits the pop-out prior is not trivial while simultaneously being generic and easy to use. + +The total objective function $\mathcal{L}_{pop}$ to supervise our object popping network is given by, + +$$ +\mathcal {L} _ {p o p} = \mathcal {L} _ {d e p} + \lambda_ {1} \cdot \mathcal {L} _ {l o c} + \lambda_ {2} \cdot \mathcal {L} _ {w t v}, \tag {5} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are the hyperparameters. + +# 3.3. Segmentation with Contact Surface + +The smoothness and edge losses encourage homogenizing the object structure, making it noticeable from the background. We now aim to further enlarge the object-background distance to make the object structure jump out. Specifically, we use an RGB-D segmentation network as shown in Figure 5. The main component of our segmentation network is a three-stream RGB-D network with some fusion design. In our setting, we choose [92] as our baseline since it is one of the SOTA RGB-D methods for saliency detection. We add a surface head to learn the depth of the contact surface $D_{c}$ , which has the same resolution as the input depth $D_{po}$ . Our surface head is composed of ConvLayer (Conv2D, BN, ReLU) and a Conv2D, which first decodes the feature maps and then transfers them into a 1-D map. + +# 3.4. Object Separation + +Using the previously discussed contact surface, in this section, we aim to separate the object from its background. At this point, we make an assumption that pixels in front of the contact surface belong to objects. The remaining pixels belong to the background. This assumption allows us to explicitly transfer the 3D knowledge into 2D semantics. Let the predicted depth of the contact surface be $D_{c} \in \mathbb{R}^{H \times W}$ . We obtain the pseudo semantics $S_{s}$ , using the popped-out depth $D_{po}$ and surface's depth $D_{c}$ , as: + +$$ +S _ {s} = \operatorname {s i g m o i d} \left(\sigma \cdot \left(D _ {p o} - D _ {c}\right)\right), \tag {6} +$$ + +where $\sigma$ is a scalar value that controls the slope of the sigmoid function. In our experiments, we use $\sigma = 10$ to perform soft-thresholding, mimicking the desired hard one for binary outputs. Such soft thresholding facilitates the gradient back-propagation required for training. Finally, we minimize the gap between the pseudo semantics $S_{s}$ and the GT semantics $G$ with binary cross-entropy (BCE): + +$$ +\mathcal {L} _ {s e p} = B C E \left(S _ {s}, G\right). \tag {7} +$$ + +# 3.5. Overall Loss Function + +Both of our trainable modules, i.e., object popping and segmentation networks, of our framework (Figure 3) are trained in an end-to-end manner. Therefore, the overall loss function consists of three parts: our depth popping loss + +![](images/ee2d166a5549719ca4c61146184fb0125def3e986be754f9511e618701074cb3.jpg) +Figure 5. Our segmentation network uses a basic RGB-D three-stream network [92]. In addition to the conventional semantic head, we learn to predict pixel-wise contact surface. The contact surface is later used to transfer depth knowledge to semantics. + +$\mathcal{L}_{pop}$ , our object separation loss $\mathcal{L}_{sep}$ , and the conventional semantic loss $\mathcal{L}_{sem}$ from the RGB-D baseline network. The total loss $\mathcal{L}_{total}$ used for training is given by, + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {p o p}} + \alpha_ {1} \cdot \mathcal {L} _ {\text {s e p}} + \alpha_ {2} \cdot \mathcal {L} _ {\text {s e m}}, \tag {8} +$$ + +where $\alpha_{1}$ and $\alpha_{2}$ are hyperparameters. + +Remarks: Our losses work in a complementary manner. The $\mathcal{L}_{pop}$ plays the same role as a smooth filter as in image smoothing. It removes the noisy depth response due to the domain gap while preserving the object structure with the help of a weak semantic label. Hence, the smoothed background becomes less informative while the object region becomes uniform, making it easily detectable. These functionalities contribute to transferring the source-free depth into popped-out depth, as desired. Such a process brings objects above the background surface, despite their distance from the camera and other distracting surfaces. The $\mathcal{L}_{sep}$ , on the other hand, fully benefits from the "pop-out" prior to segment the object from the background. With the help of the learned contact surface, this loss enlarges the foreground-background distance by pulling them in opposite directions. Such pulling results in a binary-like mask, which effectively bridges the gap between geometric knowledge and semantics. Finally, both depth-transferred and learned semantics are compared to the ground truths for supervised training. + +# 4. Results + +# 4.1. Experimental Setup + +Dataset Preparation: To better illustrate the generalizability of our approach, we evaluate the effectiveness of our approach on both SOD and COD benchmarks. We choose four widely used RGB-D SOD datasets, i.e., NLPR [48], NJUK [23], STERE [46], and SIP [10], as well as four COD datasets, i.e., CAMO [30], CHAMELEON [57], COD10K [9], and NC4K [42]. For SOD datasets, we conduct experiments with both GT depth and source-free depth. We follow the conventional learning protocol [19, 69, 92] + +Table 1. Quantitative comparison on RGB-D SOD datasets. $\uparrow (\downarrow)$ denotes that the higher (lower) is better. We use the Mean Absolute Error $(M)$ , max F-measure $(F_{m})$ , S-measure $(S_{m})$ , and max E-measure $(E_{m})$ as evaluation metrics. G.D. stands for GT Depth. Bold denotes the best performance. + +
G.D.Public.Dataset MetricNLPR [48]NJUK [23]STERE [46]SIP [10]
M↓Fm↑Sm↑Em↑M↓Fm↑Sm↑Em↑M↓Fm↑Sm↑Em↑M↓Fm↑Sm↑Em↑
Performance of RGB-D Models Trained with Source-free Depth
XMM21 [85]DFM-Net.027.909.914.944.046.903.895.927.042.906.903.934.067.873.850.891
XTIP22 [63]DCMF.027.915.921.943.044.908.903.929.041.909.907.931.067.873.853.893
XCVPR22 [21]SegMAR.024.923.920.952.036.921.909.941.037.916.907.936.052.893.872.914
XCVPR22 [47]ZoomNet.023.916.919.944.037.926.914.940.037.918.909.938.054.891.868.909
XOursPopNet.022.925.926.956.031.931.920.949.032.922.916.947.046.911.885.926
Performance of RGB-D Models Trained with GT Depth
✓TIP21 [87]BIANet.032.888.900.930.056.878.867.898.048.898.895.918.091.816.802.847
✓TIP21 [33]HAINet.024.920.924.956.037.924.911.940.040.917.907.938.052.907.879.917
✓TNNLS21 [10]D3Net.029.904.911.942.046.909.899.927.044.902.906.925.063.880.860.897
✓ECCV22 [31]SPSN.023.917.923.956.032.927.918.949.035.909.906.941.043.910.891.932
✓OursPopNet.019.927.932.963.030.936.924.952.033.924.917.947.040.923.897.937
+ +and use 700 images from NLPR and 1,485 images from NJUK for training. The rest are used for testing. For the unimodal COD dataset, we compare with both RGB COD models and RGB-D SOD models retrained on the COD datasets with the same source-free depth $D_{sf}$ . We follow the conventional training/testing protocol [8, 9, 21, 42, 47] and use 3,040 images from COD10K and 1,000 images from CAMO for training. The rest are used for testing. + +Evaluation Metrics: We evaluate the performance with four generally-recognized metrics: Mean Absolute Error $(M)$ , max F-measure $(F_{m})$ , S-measure $(S_{m})$ , and max E-measure $(E_{m})$ . All the object segmentation masks are trained or downloaded from the official resources. To make a fair comparison, we evaluate the prediction semantics with the standardized evaluation protocol as [92]. + +Implementation Details: Our model is implemented based on Pytorch with a V100 GPU. We use the Adam algorithm as an optimizer. The learning rate is initialized to $1e - 4$ and is further divided by 10 every 60 epochs. We set the input resolution to $352 \times 352$ resolution for RGB and depth. A detailed comparison with higher resolution can be found in Table 3. During training, conventional data augmentation such as random flipping, rotating, and border clipping are adopted. The training takes around 6 hours for RGB-SOD tasks and 12 hours for COD tasks for 100 epochs. + +# 4.2. Comparisons + +Comparison with RGB-D SOD Models: We present in Table 1 the performance on SOD benchmarks with our source-free depth or with GT depth. It can be seen that our model with source-free depth achieves very competitive performance compared to many RGB-D models with GT depths. Our method with GT depth also outperforms the SOTA counterparts. The qualitative comparison can be found in Fig. 6. Note that we also retrain the SOTA unimodal COD models SegMAR [21] and ZoomNet [47] on the SOD dataset only with RGB images. We show that our model with source-free depth outperforms these counterparts, showing that our method can better generalize across + +![](images/a7c7c99ae722615d051e65dcf5584a4ec196b3e6b178e04ae235ac4098b11e60.jpg) +Figure 6. Qualitative Comparison with GT Depth. Our method outperforms all counterparts while dealing with multi-objects, large depth variation, and visually-mixed foreground-background. + +different tasks with favorably better performance. + +Comparison with COD Models: We present in Table 2 the performance of the most competitive SOTA methods, including task-specific COD models as well as retrained RGB-D SOD models with source-free depth. For a fair comparison, we retrain all RGB-D methods, SegMAR [21], and ZoomNet [47] in an end-to-end manner on the same resolution images as ours. Some RGB-only methods perform even better than many RGB-D methods, mainly due to their COD task-specific designs and the lack of ground-truth depths required for RGB-D methods. It is important to note that competing RGB-only COD methods do not perform favorably on SOD tasks. Please, refer to Table 1 & 3 to observe their poor cross-tasks generalization. + +Towards Higher Resolution: Previous studies have shown that the image resolution may influence the model performance [43, 47, 73, 83]. For example, the current SOTA COD method ZoomNet [47] is with main scale $384^2$ and implies the highest resolution of $(384 \times 1.5)^2 = 576^2$ , as it operates on $0.5 \times 1 \times$ and $1.5 \times$ scales. To make a fair comparison, we retrain the model with the same resolution $(352^2$ or $512^2)$ as ours. We show in Table3-ZoomNet* that the results deteriorate as expected. Compared to these counterparts, our method offers a good trade-off between accuracy and efficiency. More comparisons on SOD and COD benchmarks can be found in the supplementary material. + +Table 2. Quantitative comparison on COD datasets. Pseudo stands for source-free depth used for RGB-D methods. + +
Pseudo Public.Dataset MetricCAMO [30]CHAMELEON [57]COD10K [9]NC4K [42]
M ↓Fm ↑Sm ↑Em ↑M ↓Fm ↑Sm ↑Em ↑M ↓Fm ↑Sm ↑Em ↑M ↓Fm ↑Sm ↑Em ↑
Performance of RGB COD Models
X CVPR20 [9]SINet.099.762.751.790.044.845.868.908.051.708.771.832.058.804.808.873
X CVPR21 [42]SLSR.080.791.787.843.030.866.889.938.037.756.804.854.048.836.839.898
X CVPR21 [78]MGL-R.088.791.775.820.031.868.893.932.035.767.813.874.053.828.832.876
X CVPR21 [44]PFNet.085.793.782.845.033.859.882.927.040.747.800.880.053.820.829.891
X CVPR21 [32]UJSC.072.812.800.861.030.874.891.948.035.761.808.886.047.838.841.900
X IJCAI21 [59]C2FNet.079.802.796.856.032.871.888.936.036.764.813.894.049.831.838.898
X ICCV21 [74]UGTR.086.800.783.829.031.862.887.926.036.769.816.873.052.831.839.884
X CVPR22 [21]SegMAR.080.799.794.857.032.871.887.935.039.750.799.876.050.828.836.893
X CVPR22 [47]ZoomNet.074.818.801.858.033.829.859.915.034.771.808.872.045.841.843.893
Performance of RGB-D Models Retrained with Source-free Depth
✓ MM21 [80]CDINet.100.638.732.766.036.787.879.903.044.610.778.821.067.697.793.830
✓ CVPR21 [19]DCF.089.724.749.834.037.821.850.923.040.685.766.864.061.765.791.878
✓ ICCV21 [81]CMINet.087.798.782.827.032.881.891.930.039.768.811.868.053.832.839.888
✓ ICCV21 [92]SPNet.083.807.783.831.033.872.888.930.037.776.808.869.054.828.825.874
✓ TIP22 [63]DCMF.115.737.728.757.059.807.830.853.063.679.748.776.077.782.794.820
✓ ECCV22 [31]SPSN.084.782.773.829.032.866.887.932.042.727.789.854.059.803.813.867
✓ OursPopNet.073.821.806.869.022.893.910.962.031.789.827.897.043.852.852.908
+ +![](images/313a4908b5d16c66ace5c12b698a042f87c9b8097dd66ad77b2c96d0d0ad1135.jpg) +Image + +![](images/fc2c24b00d10f08b7bdd7dd8f2964a1fd652a889b097e89c340d9031beb183a4.jpg) +GT + +![](images/02e0ff4e078e32ba845abdbfe3c8f40e834636ad230a059bf03c238dd086bd8a.jpg) +Ours + +![](images/3c2f5971a18abf0bc7be5d0538c3d65b89663431c1cc977dfa8c96e7b47d0feb.jpg) +ZoomNet +Figure 7. Qualitative comparison. Our method can better preserve the object structure compared to other counterparts, especially while dealing with objects with occlusion (two first rows). Our method performs favorably with multiple objects (last row). Better to zoom in. + +![](images/151245e316fa0105cc34499f11c453927f83187910a3a9ab87897251828f94a6.jpg) +UGTR + +![](images/f9bc22628844ea3d45b3b01616fdfda185a8ef4ad372170fcd5231926718472d.jpg) +UJSC + +![](images/b82db1fbe06a43661851b35db1db411cd5eef8b20f4dbe670b0f0677c5528f21.jpg) +C2FNet + +![](images/ce3b530f65c1510010c26bc45654e97d0cb3cf9a76f72ed608331043141d7995.jpg) +PFNet +MGL-R + +![](images/24fff8867aec7c6c7d44ac4e9c274c5984dd0ea2789e2b6032edbedd0d2b9efd.jpg) + +![](images/a7c88e08c35af38c8a95c03fc87ee3a3fcf879dd6fffb226d4dd8f82e51a2e84.jpg) +SLSR +SINet + +Table 3. End-to-end comparison with different resolutions on SOD benchmarks. Our method with source-free depth generalizes significantly better compared to SOTA COD models. + +
ModelSizeFlops (G)NJUK [23]SIP [10]
M↓Fm↑Sm↑Em↑M↓Fm↑Sm↑Em↑
SegMAR [21]352267.3.036.921.909.941.052.893.872.914
ZoomNet [47]3522167.8.037.926.914.940.054.891.868.909
Ours3522228.8.031.931.920.949.046.911.885.926
SegMAR [21]5122142.4.035.927.914.943.050.899.878.917
ZoomNet [47]5122353.4.036.926.915.942.052.895.873.910
Ours5122484.0.031.933.922.951.044.911.890.927
+ +Qualitative Comparisons: Figure 7 presents the output of our network on challenging cases. It can be seen while dealing with objects occluded by thin objects ( $1^{st} - 4^{th}$ rows), our method can accurately reason about the segmentation masks closer to the GT. We also achieve better performance while dealing with multiple objects (last row). More discussions on multiple objects can be found in Table 6. + +# 4.3. Ablation Study + +Loss: In this section, we conduct experiments on analyzing the effectiveness of the proposed losses. The quantitative results of different loss combinations are provided in Table 4. It can be seen that each proposed loss behaves properly, i.e., improving the performance compared to the baseline. More + +Table 4. Ablation study on the proposed losses. + +
\( \mathcal{L}_{dep} \)\( \mathcal{L}_{loc} \)\( \mathcal{L}_{wtv} \)\( \mathcal{L}_{sep} \)SIP [10]NC4K [42]
\( M\downarrow \)\( F_m\uparrow \)\( S_m\uparrow \)\( E_m\uparrow \)\( M\downarrow \)\( F_m\uparrow \)\( S_m\uparrow \)\( E_m\uparrow \)
----.048.903.884.922.052.832.832.893
✓---.046.907.889.925.051.833.839.895
-✓--.045.908.893.929.048.837.844.898
--✓-.046.906.891.927.050.833.841.894
---✓.043.914.893.933.048.840.848.900
✓✓--.044.911.893.928.049.837.844.897
✓-✓-.046.909.893.927.046.840.845.898
✓-✓✓.040.918.897.935.045.848.849.904
✓✓-✓.042.916.894.931.044.850.850.906
✓✓✓✓.040.923.897.937.043.852.852.908
+ +discussions and ablation studies on the hyperparameters can be found in the supplementary material. + +Object Popping Network as Plug-in: Our object popping network can be easily adapted with different encoders and with different existing RGB-D models. For example, with a ResNet-18 [15] encoder and convolution-based decoder, our poping only costs around 12.7M additional learning parameters or 48.7 MB model size. We show in Table 5 that our method can favorably improve performance over the baseline with less than an extra $10\%$ GFlops. + +Pop-out Under Reduced Training Data: Here, we are interested in analyzing source-free depth's benefits. Therefore, we conduct different experiments by reducing the training data. As shown in Figure 8(left), when both our + +Table 5. Generalization and cost over different RGB-D baselines. Dataset Metric Flops Param S1P [10] NC4K [42] M $\downarrow$ ${F}_{m} \uparrow$ ${S}_{m} \uparrow$ ${E}_{m} \uparrow$ ${M\downarrow }$ ${F}_{m} \uparrow$ ${S}_{m} \uparrow$ ${E}_{m} \uparrow$ +Ours 363.2 59.8 .053 .899 .874 .919 .057 .809 .804 .872 373.7 72.5 .051 .910 .886 .923 .055 .814 .811 .878 SPNet [92] 149.0 150.4 .044 .911 .887 .914 .054 .828 .825 .874 +Ours 159.5 163.1 .042 .917 .894 .932 .044 .851 .851 .905 + +![](images/228225256ac391b291cd02e59acc6601ce70bd569a76eddd625e1174b20fb95d.jpg) +Figure 8. Benefit of our method on reducing training data. When our network is trained only with $25\%$ data, the performance remains competitive compared to the baseline, i.e., the absolute performance gains are $+0.4\%$ max F-measure and $+0.1\%$ S-measure on NC4K compared to the baseline trained with all data. + +![](images/c0240fe18a2b8fcba530dfb9589bae2379566a86fac6320b5455982b53e654cc.jpg) +Figure 9. Histogram on the gain. Please zoom in for details. + +PopNet and our baseline are trained with all data, our PopNet can lead to absolute improvements with $5.6\%$ in max F-measure and with $3.9\%$ in S-measure on COD10K dataset. While our PopNet is trained with only $25\%$ data, it can still achieve competitive performance compared to our baseline trained with all data. Similar phenomena can be observed in the NC4K dataset as shown in Figure 8(right). To conclude, our method can efficiently explore the geometric prior and significantly reduce the required training data volume. + +Gain over baseline: With depth cues, as shown in Figure 9, we boost the performance in 3125 over 4121 images ( $\sim 75\%$ cases). We also show in Table 6 that our network performs favorably over baselines with single or multiple objects. Our method may fail when the source-free depth is clueless. This mainly happens when the object is well concealed and fools the depth network. However, such cases are also challenging, even for humans. + +# Are RGB-D Methods Better Than RGB-only Methods? + +RGB-D methods are indeed better than RGB-only methods provided GT depth. However, only a few RGB-D methods can benefit from the source-free depth. For example, + +Table 6. Multi-Object performance on NC4K [42] with size ${512}^{2}$ . + +
Obj. Nbr. (%)MetricSingle (92%)Two (6%)More (2%)Overall
M↓Fm↑M↓Fm↑M↓Fm↑M↓Fm↑
RGB Baseline.054.828.067.811.091.738.056.825
+ Dxf.050.842.063.821.093.746.051.839
Ours.040.864.051.847.079.767.042.861
+ +Table 7. Performance with RGB- $D_{sf}$ model vs. with RGB-only baseline. $D_{sf}$ stands for source-free depth. + +
DatasetCOD10K [9]NC4K [42]
MetricM↓Fm↑Sm↑Em↑M↓Fm↑Sm↑Em↑
DASNet [88].041.643.793.864.055.747.830.879
+ Dsf.041.642.796.858.055.743.830.874
SPNet [92].040.743.801.867.052.846.833.883
+ Dsf.037.776.808.869.054.828.825.874
+ +as shown in Table 7, DASNet [88] achieves poorer performance compared to RGB baseline when trained with source-free depth. Similarly, even for one of the SOTA RGB-D models, SPNet [92], the performance on NC4K dataset with RGB-only input is better than with additional source-free depth. Moreover, when provided with source-free depth, none of the existing RGB-D methods outperform the best performing RGB-only method (e.g., ZoomNet [47]) on COD. The same observation was also made on SOD as well. This could be because of the domain gap coupled with the fusion design, among others. We also found it non-trivial to extend the most well-performing RGB-only methods to the RGB-D case. Note that our PopNet performs better than all existing RGB-only and RGB-D methods, with source-free or GT depth maps. + +# 5. Conclusion + +We demonstrate a successful case of cross-domain cross-task depth to semantics knowledge transfer using only the source model. In this paper, the source-free depth at the target offered by a given source model is used. The proposed method learns to transfer knowledge from depth to semantics using the objects' pop-out prior. We facilitate our network to use such prior by designing a novel network architecture. The designed network reasons about the objects by popping them out from the provided depth maps. This process is followed by separating objects from the background using the learned contact surface. We show that the joint learning of object pop-out and contact surface can be successfully supervised using the target semantics. Exhaustive experiments on SOD and COD benchmarks show the successful transfer of depth knowledge to the target, in terms of improved performance and generalization. + +Acknowledgement The authors thank the anonymous reviewers and ACs for their tremendous efforts and helpful comments. This research is financed in part by the Conseil Régional de Bourgogne-Franche-Comté, Toyota Motor Europe (research project TRACE-Zurich), and the Alexander von Humboldt Foundation. + +# References + +[1] Peshal Agarwal, Danda Pani Paudel, Jan-Nico Zaech, and Luc Van Gool. Unsupervised robust domain adaptation without source data. In WACV, 2022. 2 +[2] Amir Atapour-Abarghouei and Toby P Breckon. Real-time monocular depth estimation using synthetic data with domain adaptation via image style transfer. In IEEE CVPR, 2018. 1 +[3] Adriano Cardace, Luca De Luigi, Pierluigi Zama Ramirez, Samuele Salti, and Luigi Di Stefano. Plugging self-supervised monocular depth into unsupervised domain adaptation for semantic segmentation. In WACV, 2022. 1 +[4] Po-Yi Chen, Alexander H Liu, Yen-Cheng Liu, and Yu-Chiang Frank Wang. Towards scene understanding: Unsupervised monocular depth estimation with semantic-aware representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 4 +[5] Innes C Cuthill, Martin Stevens, Jenna Sheppard, Tracey Maddocks, C Alejandro Parraga, and Tom S Troscianko. Disruptive coloration and background pattern matching. Nature, 434(7029):72-74, 2005. 3 +[6] Jiahua Dong, Zhen Fang, Anjin Liu, Gan Sun, and Tongliang Liu. Confident anchor-induced multi-source free domain adaptation. NeurIPS, 2021. 1 +[7] Aliya El Nagar, Daniel Osorio, Sarah Zylinski, and Steven M Sait. Visual perception and camouflage response to 3d backgrounds and cast shadows in the european cuttlefish, sepia officinalis. JEB, 224(11), 2021. 3 +[8] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE TPAMI, 44(10):6024-6042, 2022. 3, 6 +[9] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In IEEE CVPR, 2020. 3, 5, 6, 7, 8 +[10] Deng-Ping Fan, Zheng Lin, Zhao Zhang, Menglong Zhu, and Ming-Ming Cheng. Rethinking RGB-D salient object detection: Models, datasets, and large-scale benchmarks. IEEE TNNLS, 32(5):2075-2089, 2021. 3, 5, 6, 7, 8 +[11] Deng-Ping Fan, Jing Zhang, Gang Xu, Ming-Ming Cheng, and Ling Shao. Salient objects in clutter. IEEE TPAMI, 2023. 2 +[12] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In IEEE CVPR, 2019. 1 +[13] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In IEEE ICCV, 2019. 1, 2, 4 +[14] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In IEEE CVPR, 2022. 4 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE CVPR, 2016. 7 +[16] Ian P Howard. Perceiving in depth, volume 1: basic mechanisms. Oxford University Press, 2012. 1 + +[17] Junjie Hu, Mete Ozay, Yan Zhang, and Takayuki Okatani. Revisiting single image depth estimation: Toward higher resolution maps with accurate object boundaries. In WACV, 2019. 4 +[18] Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection. MIR, 2023. 3 +[19] Wei Ji, Jingjing Li, Shuang Yu, Miao Zhang, Yongri Piao, Shunyu Yao, Qi Bi, Kai Ma, Yefeng Zheng, Huchuan Lu, et al. Calibrated RGB-D salient object detection. In IEEE CVPR, 2021. 5, 7 +[20] Wei Ji, Jingjing Li, Miao Zhang, Yongri Piao, and Huchuan Lu. Accurate RGB-D salient object detection via collaborative learning. In ECCV, 2020. 3 +[21] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In IEEE CVPR, 2022, 3, 6, 7 +[22] Wen-Da Jin, Jun Xu, Qi Han, Yi Zhang, and Ming-Ming Cheng. Cdnet: Complementary depth network for rgb-d salient object detection. IEEE TIP, 30:3376-3390, 2021. 3 +[23] Ran Ju, Ling Ge, Wenjing Geng, Tongwei Ren, and Gangshan Wu. Depth saliency based on anisotropic centersurround difference. In IEEE ICIP, 2014. 5, 6, 7 +[24] Hongwen Kang, Alexei A Efros, Martial Hebert, and Takeo Kanade. Image composition for object pop-out. In IEEE ICCVW, 2009. 2 +[25] Youngeun Kim, Donghyeon Cho, Kyeongtak Han, Priyadarshini Panda, and Sungeun Hong. Domain adaptation without source data. IEEE TAI, 2(6):508-518, 2021. 1, 2 +[26] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In IEEE CVPR, 2021. 1 +[27] Jogendra Nath Kundu, Akshay R Kulkarni, Suvaansh Bhambri, Deepesh Mehta, Shreyas Anand Kulkarni, Varun Jampani, and Venkatesh Babu Radhakrishnan. Balancing discriminability and transferability for source-free domain adaptation. In ICML, 2022. 1 +[28] Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In IEEE CVPR, 2020. 1, 2 +[29] Vinod K Kurmi, Venkatesh K Subramanian, and Vinay P Namboodiri. Domain impression: A source data free domain adaptation method. In WACV, 2021. 1, 2 +[30] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, MinhTriet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. CVU, 184:45-56, 2019. 5, 7 +[31] Minhyeok Lee, Chaewon Park, Suhwan Cho, and Sangyoun Lee. Spsn: Superpixel prototype sampling network for rgb-d salient object detection. In ECCV, 2022. 6, 7 +[32] Aixuan Li, Jing Zhang, Yunqiu Lv, Bowen Liu, Tong Zhang, and Yuchao Dai. Uncertainty-aware joint salient object and camouflaged object detection. In IEEE CVPR, 2021. 3, 7 +[33] Gongyang Li, Zhi Liu, Minyu Chen, Zhen Bai, Weisi Lin, and Haibin Ling. Hierarchical alternate interaction network + +forrgb-d salient object detection.IEEE TIP,30:3528-3542, 2021.3,6,8 +[34] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In IEEE CVPR, 2020. 2 +[35] Rui Li, Danna Xue, Shaolin Su, Xiantuo He, Qing Mao, Yu Zhu, Jinqiu Sun, and Yanning Zhang. Learning depth via leveraging semantics: Self-supervised monocular depth estimation with both implicit and explicit semantic guidance. Pattern Recognition (PR), page 109297, 2023. 4 +[36] Xinhao Li, Jingjing Li, Lei Zhu, Guoqing Wang, and Zi Huang. Imbalanced source-free domain adaptation. In ACM MM, 2021. 1 +[37] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020. 2 +[38] Jiang-Jiang Liu, Qibin Hou, Ming-Ming Cheng, Jiashi Feng, and Jianmin Jiang. A simple pooling-based design for real-time salient object detection. In IEEE CVPR, 2019. 3 +[39] Yuang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. In IEEE CVPR, 2021. 1, 2 +[40] Yukang Lu, Dingyao Min, Keren Fu, and Qijun Zhao. Depth-cooperated trimodal network for video salient object detection. In ICIP. IEEE, 2022. 3 +[41] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM TOG, 39(4):1-13, 2020. 1 +[42] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In IEEE CVPR, 2021. 3, 5, 6, 7, 8 +[43] Daniel McKee, Zitong Zhan, Bing Shuai, Davide Modolo, Joseph Tighe, and Svetlana Lazebnik. Transfer of representations to video label propagation: Implementation factors matter. arXiv preprint arXiv:2203.05553, 2022. 6 +[44] Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan. Camouflaged object segmentation with distraction mining. In IEEE CVPR, 2021. 7 +[45] S Mahdi H Miangoleh, Sebastian Dille, Long Mai, Sylvain Paris, and Yagiz Aksoy. Boosting monocular depth estimation models to high-resolution via content-adaptive multi-resolution merging. In IEEE CVPR, 2021. 1, 4 +[46] Yuzhen Niu, Yujie Geng, Xueqing Li, and Feng Liu. Leveraging stereopsis for saliency analysis. In IEEE CVPR, 2012. 5, 6 +[47] Youwei Pang, Xiaqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In IEEE CVPR, 2022. 3, 6, 7, 8 +[48] Houwen Peng, Bing Li, Weihua Xiong, Weiming Hu, and Rongrong Ji. RGBD salient object detection: a benchmark and algorithms. In ECCV, 2014. 5, 6 +[49] Yongri Piao, Zhengkun Rong, Miao Zhang, Weisong Ren, and Huchuan Lu. A2dele: Adaptive and attentive depth distiller for efficient RGB-D salient object detection. In IEEE CVPR, 2020. 3 + +[50] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In IEEE ICCV, 2021. 1, 4 +[51] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI, 44(3):1623-1637, 2022. 1, 2, 4 +[52] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4 +[53] Suman Saha, Anton Obukhov, Danda Pani Paudel, Menelaos Kanakis, Yuhua Chen, Stamatios Georgoulis, and Luc Van Gool. Learning to relate depth and semantics for unsupervised domain adaptation. In IEEE CVPR, 2021. 1 +[54] Roshni Sahoo, Divya Shanmugam, and John Guttag. Unsupervised domain adaptation in the absence of source data. arXiv preprint arXiv:2007.10233, 2020. 2 +[55] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE CVPR, 2016. 1 +[56] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from RGBD images. In ECCV, 2012. 1 +[57] Przemysław Skurowski, Hassan Abdulameer, J Blaszczyk, Tomasz Depta, Adam Kornacki, and P Koziel. Animal camouflage analysis: Chameleon database. Unpublished manuscript, 2(6):7, 2018. 5, 7 +[58] Hwanjun Song, Eunyoung Kim, Varun Jampan, Deqing Sun, Jae-Gil Lee, and Ming-Hsuan Yang. Exploiting scene depth for object detection with multimodal transformers. In BMVC, 2021. 3 +[59] Yujia Sun, Geng Chen, Tao Zhou, Yi Zhang, and Nian Liu. Context-aware cross-level fusion network for camouflaged object detection. In *IJCAI*, 2021. 3, 7 +[60] Yujia Sun, Shuo Wang, Chenglizhao Chen, and Tian-Zhu Xiang. Boundary-guided camouflaged object detection. In IJ-CAI, 2022. 3 +[61] Alessio Tonioni, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Unsupervised domain adaptation for depth prediction from images. IEEE TPAMI, 42(10):2396-2409, 2019. 1 +[62] Anne Treisman. Preattentive processing in vision. ICVGIP, 31(2):156-177, 1985. 2 +[63] Fengyun Wang, Jinshan Pan, Shoukun Xu, and Jinhui Tang. Learning discriminative cross-modality features for RGB-D saliency detection. IEEE TIP, 31:1285-1297, 2022. 6, 7 +[64] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In IEEE CVPR, 2022. 2 +[65] Cho-Ying Wu, Jialiang Wang, Michael Hall, Ulrich Neumann, and Shuochen Su. Toward practical monocular indoor depth estimation. In IEEE CVPR, 2022. 1 +[66] Yu-Huan Wu, Yun Liu, Jun Xu, Jia-Wang Bian, Yu-Chao Gu, and Ming-Ming Cheng. Mobilesal: Extremely efficient rgb-d salient object detection. IEEE TPAMI, 44(12):10261-10269, 2022. 3 +[67] Zongwei Wu, Guillaume Allibert, Christophe Stolz, and Cédric Demonceaux. Depth-adapted CNN for RGB-D cameras. In ACCV, 2020. 1 + +[68] Zongwei Wu, Guillaume Allibert, Christophe Stolz, Chao Ma, and Cédric Demonceaux. Modality-guided subnetwork for salient object detection. In 3DV, 2021. 1, 3 +[69] Zongwei Wu, Shriarulmozhivarman Gobichettipalayam, Brahim Tamadazte, Guillaume Allibert, Danda Pani Paudel, and Cedric Demonceaux. Robust RGB-D fusion for saliency detection. 3DV, 2022. 3, 5 +[70] Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In IEEE CVPR, 2019. 3 +[71] Haifeng Xia, Handong Zhao, and Zhengming Ding. Adaptive adversarial network for source-free domain adaptation. In IEEE ICCV, 2021. 1, 2 +[72] Ke Xian, Jianming Zhang, Oliver Wang, Long Mai, Zhe Lin, and Zhiguo Cao. Structure-guided ranking loss for single image depth prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 4 +[73] Mochu Xiang, Jing Zhang, Yunqiu Lv, Aixuan Li, Yiran Zhong, and Yuchao Dai. Exploring depth contribution for camouflaged object detection. arXiv e-prints, pages arXiv-2106, 2021. 1, 6 +[74] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping Fan. Uncertainty-guided transformer reasoning for camouflaged object detection. In IEEE ICCV, 2021. 3, 7 +[75] Shiqi Yang, Joost van de Weijer, Luis Herranz, Shangling Jui, et al. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. NeurIPS, 2021. 1 +[76] Shiqi Yang, Yaxing Wang, Joost van de Weijer, and Luis Herranz. Unsupervised domain adaptation without source data by casting a bait. arXiv preprint arXiv:2010.12427, 2020. 2 +[77] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In IEEE ICCV, 2021. 1, 2 +[78] Qiang Zhai, Xin Li, Fan Yang, Chenglizhao Chen, Hong Cheng, and Deng-Ping Fan. Mutual graph learning for camouflaged object detection. In IEEE CVPR, 2021. 7 +[79] Wei Zhai, Yang Cao, HaiYong Xie, and Zheng-Jun Zha. Deep texton-coherence network for camouflaged object detection. IEEE TMM, 2022. 3 +[80] Chen Zhang, Runmin Cong, Qinwei Lin, Lin Ma, Feng Li, Yao Zhao, and Sam Kwong. Cross-modality discrepant interaction network for RGB-D salient object detection. In ACM MM, 2021. 3, 7 +[81] Jing Zhang, Deng-Ping Fan, Yuchao Dai, Xin Yu, Yiran Zhong, Nick Barnes, and Ling Shao. RGB-D saliency detection via cascaded mutual information minimization. In IEEE ICCV, 2021. 7 +[82] Jing Zhang, Yunqiu Lv, Mochu Xiang, Aixuan Li, Yuchao Dai, and Yiran Zhong. Depth-guided camouflaged object detection. CoRR, abs/2106.13217, 2021. 3 +[83] Miao Zhang, Shuang Xu, Yongri Piao, Dongxiang Shi, Shusen Lin, and Huchuan Lu. Preynet: Preying on camouflaged objects. In ACM MM, 2022. 3, 6 +[84] Pingping Zhang, Dong Wang, Huchuan Lu, Hongyu Wang, and Baocai Yin. Learning uncertain convolutional features for accurate saliency detection. In IEEE ICCV, 2017. 3 + +[85] Wenbo Zhang, Ge-Peng Ji, Zhuo Wang, Keren Fu, and Qijun Zhao. Depth quality-inspired feature manipulation for efficient RGB-D salient object detection. In ACM MM, 2021. 6 +[86] Zhoutong Zhang, Forrester Cole, Richard Tucker, William T Freeman, and Tali Dekel. Consistent depth of moving objects in video. ACM TOG, 40(4):1-12, 2021. 1 +[87] Zhao Zhang, Zheng Lin, Jun Xu, Wen-Da Jin, Shao-Ping Lu, and Deng-Ping Fan. Bilateral attention network forrgb-d salient object detection. IEEE TIP, 30:1949-1961, 2021. 6 +[88] Jiawei Zhao, Yifan Zhao, Jia Li, and Xiaowu Chen. Is depth really necessary for salient object detection? In ACM MM, 2020. 3, 8 +[89] Shanshan Zhao, Huan Fu, Mingming Gong, and Dacheng Tao. Geometry-aware symmetric domain adaptation for monocular depth estimation. In IEEE CVPR, 2019. 1 +[90] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In IEEE CVPR, 2022. 3 +[91] Tao Zhou, Deng-Ping Fan, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Rgb-d salient object detection: A survey. CVMJ, pages 1-33, 2021. 1 +[92] Tao Zhou, Huazhu Fu, Geng Chen, Yi Zhou, Deng-Ping Fan, and Ling Shao. Specificity-preserving RGB-D saliency detection. In IEEE ICCV, 2021. 3, 5, 6, 7, 8 +[93] Zikun Zhou, Wenjie Pei, Xin Li, Hongpeng Wang, Feng Zheng, and Zhenyu He. Saliency-associated object tracking. In IEEE ICCV, 2021. 3 +[94] Hongwei Zhu, Peng Li, Haoran Xie, Xuefeng Yan, Dong Liang, Dapeng Chen, Mingqiang Wei, and Jing Qin. I can find you! boundary-guided separated attention network for camouflaged object detection. In AAAI, 2022. 3 +[95] Jinchao Zhu, Xiaoyu Zhang, Shuo Zhang, and Junnan Liu. Inferring camouflaged objects by texture-aware interactive guidance network. In AAAI, 2021. 3 \ No newline at end of file diff --git a/2023/Source-free Depth for Object Pop-out/images.zip b/2023/Source-free Depth for Object Pop-out/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..32b4e199b26c37d47a2d5686d3430f380fe96687 --- /dev/null +++ b/2023/Source-free Depth for Object Pop-out/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0514633382c1da6b81cd236acdf04d575a1bec74d5872f192f5219644f92987a +size 853317 diff --git a/2023/Source-free Depth for Object Pop-out/layout.json b/2023/Source-free Depth for Object Pop-out/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..784a4c9ee74c139b141b7cbe00436b6c72a9412c --- /dev/null +++ b/2023/Source-free Depth for Object Pop-out/layout.json @@ -0,0 +1,10731 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 179, + 103, + 414, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 103, + 414, + 121 + ], + "spans": [ + { + "bbox": [ + 179, + 103, + 414, + 121 + ], + "type": "text", + "content": "Source-free Depth for Object Pop-out" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 141, + 523, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 141, + 523, + 171 + ], + "spans": [ + { + "bbox": [ + 67, + 141, + 523, + 171 + ], + "type": "text", + "content": "Zongwei Wu" + }, + { + "bbox": [ + 67, + 141, + 523, + 171 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 67, + 141, + 523, + 171 + ], + "type": "text", + "content": " Danda Pani Paudel" + }, + { + "bbox": [ + 67, + 141, + 523, + 171 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 67, + 141, + 523, + 171 + ], + "type": "text", + "content": " Deng-Ping Fan" + }, + { + "bbox": [ + 67, + 141, + 523, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 67, + 141, + 523, + 171 + ], + "type": "text", + "content": " Jingjing Wang" + }, + { + "bbox": [ + 67, + 141, + 523, + 171 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 67, + 141, + 523, + 171 + ], + "type": "text", + "content": " Shuo Wang" + }, + { + "bbox": [ + 67, + 141, + 523, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 141, + 523, + 171 + ], + "type": "text", + "content": " Cédric Demonceaux" + }, + { + "bbox": [ + 67, + 141, + 523, + 171 + ], + "type": "inline_equation", + "content": "^{2,6}" + }, + { + "bbox": [ + 67, + 141, + 523, + 171 + ], + "type": "text", + "content": " Radu Timofte" + }, + { + "bbox": [ + 67, + 141, + 523, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 141, + 523, + 171 + ], + "type": "text", + "content": " Luc Van Gool" + }, + { + "bbox": [ + 67, + 141, + 523, + 171 + ], + "type": "inline_equation", + "content": "^{1,4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 172, + 532, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 172, + 532, + 200 + ], + "spans": [ + { + "bbox": [ + 62, + 172, + 532, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 172, + 532, + 200 + ], + "type": "text", + "content": " CVL, ETH Zurich " + }, + { + "bbox": [ + 62, + 172, + 532, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 172, + 532, + 200 + ], + "type": "text", + "content": " University of Burgundy, CNRS, ICB " + }, + { + "bbox": [ + 62, + 172, + 532, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 172, + 532, + 200 + ], + "type": "text", + "content": " Computer Vision Lab, CAIDAS & IFI, University of Wurzburg " + }, + { + "bbox": [ + 62, + 172, + 532, + 200 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 62, + 172, + 532, + 200 + ], + "type": "text", + "content": " INSAIT, Sofia University " + }, + { + "bbox": [ + 62, + 172, + 532, + 200 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 62, + 172, + 532, + 200 + ], + "type": "text", + "content": " AUST " + }, + { + "bbox": [ + 62, + 172, + 532, + 200 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 62, + 172, + 532, + 200 + ], + "type": "text", + "content": " University of Lorraine, CNRS, Inria, Loria" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 192, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 240 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 240 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 253, + 290, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 290, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 290, + 553 + ], + "type": "text", + "content": "Depth cues are known to be useful for visual perception. However, direct measurement of depth is often impracticable. Fortunately, though, modern learning-based methods offer promising depth maps by inference in the wild. In this work, we adapt such depth inference models for object segmentation using the objects' \"pop-out\" prior in 3D. The \"pop-out\" is a simple composition prior that assumes objects reside on the background surface. Such compositional prior allows us to reason about objects in the 3D space. More specifically, we adapt the inferred depth maps such that objects can be localized using only 3D information. Such separation, however, requires knowledge about contact surface which we learn using the weak supervision of the segmentation mask. Our intermediate representation of contact surface, and thereby reasoning about objects purely in 3D, allows us to better transfer the depth knowledge into semantics. The proposed adaptation method uses only the depth model without needing the source data used for training, making the learning process efficient and practical. Our experiments on eight datasets of two challenging tasks, namely salient object detection and camouflaged object detection, consistently demonstrate the benefit of our method in terms of both performance and generalizability. The source code is publicly available at https://github.com/Zongwei97/PopNet." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 576, + 128, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 576, + 128, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 576, + 128, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 597, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 597, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 287, + 693 + ], + "type": "text", + "content": "The 3D knowledge of the scene is long known to be complementary to the task of visual perception [12, 16, 56, 67, 91]. Often in practice, though, visual perception needs to be carried out using only 2D images. Given multiple images, 3D geometry may be recovered using the structure-from-motion techniques [26, 41, 55, 86]. Such inversion, however, is not compatible when only a single image is available. Under such circumstances, image inversion" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 230, + 545, + 339 + ], + "blocks": [ + { + "bbox": [ + 307, + 230, + 545, + 339 + ], + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 339 + ], + "type": "image", + "image_path": "bc00a305be8a8a03db526ed5dc79a002452bdf98338452b280083f379e51b4fb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 342, + 547, + 386 + ], + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 386 + ], + "type": "text", + "content": "Figure 1. Depth to semantics conversion using the object pop-out prior. For input RGB and source-free depth pair, we learn contact surface. The obtained contact surface is then used to separate objects and backgrounds to derive pseudo semantics for supervision." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 401, + 546, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 546, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 546, + 461 + ], + "type": "text", + "content": "to depth map is usually done using learning-based methods [13,45,50,51,65], which have shown unparalleled success in the recent years. Unfortunately, the learning-based methods may not offer high-quality depth maps due to the generalization deficiency across domains [68, 73]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 461, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 547, + 581 + ], + "type": "text", + "content": "Despite poor generalization, the knowledge gained in one domain is shown to be useful in other close-by domains. This utility is harnessed by performing the so-called domain adaptation (DA) [2,3,6,53,61,89]. In fact, it has been shown recently that DA methods can efficiently transfer knowledge using only the prediction models, i.e., without requiring access to the data where the model is trained – also known as the source-free domain adaptation (SDA) [25,28,39,71,77]. The SDA methods are of gripping interest due to their efficiency and privacy promises." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 548, + 714 + ], + "type": "text", + "content": "The most existing SDA methods make one or both of these implicit assumptions: (a) a similar (as that of the source) supervising task at the target [28, 75, 77]; (b) task of discrete (and known) label space [27, 29, 36, 71]. The former assumption not only makes the source and target domains easier to compare but also potentially keeps the two domains closer. The latter assumption allows performing SDA by self-training where the discrete labels facilitate reasoning about the model's confidence. The self-training is then performed by boosting the confidence at the target for some picked reliable examples." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 268, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 268, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 268, + 714 + ], + "type": "text", + "content": "*Corresponding Author: Deng-Ping Fan (dengpfan@gmail.com)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "text", + "content": "1032" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 85, + 547, + 178 + ], + "blocks": [ + { + "bbox": [ + 194, + 74, + 421, + 84 + ], + "lines": [ + { + "bbox": [ + 194, + 74, + 421, + 84 + ], + "spans": [ + { + "bbox": [ + 194, + 74, + 421, + 84 + ], + "type": "text", + "content": "Cross-Task Cross-Dataset Analysis with Depth Pop-out Prior" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 85, + 547, + 178 + ], + "lines": [ + { + "bbox": [ + 50, + 85, + 547, + 178 + ], + "spans": [ + { + "bbox": [ + 50, + 85, + 547, + 178 + ], + "type": "image", + "image_path": "f208f75faa12ac29e7bda8c0fdd583158bc75bf9757271f90a8896e8d21f815a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 182, + 546, + 226 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 546, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 546, + 226 + ], + "type": "text", + "content": "Figure 2. The performance gained in F-measure using our method over the established baselines. Our method " + }, + { + "bbox": [ + 46, + 182, + 546, + 226 + ], + "type": "inline_equation", + "content": "(\\triangle, \\square)" + }, + { + "bbox": [ + 46, + 182, + 546, + 226 + ], + "type": "text", + "content": " significantly improves baselines on 8 datasets and 2 tasks (each task on 4 datasets- SOD: left two; COD: right two). We also compare 24 methods " + }, + { + "bbox": [ + 46, + 182, + 546, + 226 + ], + "type": "inline_equation", + "content": "(\\bullet)" + }, + { + "bbox": [ + 46, + 182, + 546, + 226 + ], + "type": "text", + "content": ", where our method offers state-of-the-art results despite their task specialization. Note that all methods are connected using lines to illustrate their performance fluctuations across datasets. Please, refer Tables 1 & 2 and Section 4.3 for more details and discussions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 240, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 287, + 432 + ], + "type": "text", + "content": "In this work, we aim for source-free transfer of depth knowledge for object detection. Such transfer is desired to assist in locating the object by depth cues and to exploit the depth knowledge despite the domain gap. The addressed problem setting differs from the standard SDA in terms of (a) the source and target tasks' difference; (b) and the continuous label space of the depth. These differences (with standard SDA) make our task at hand very challenging, which is addressed for the first time in this paper, up to our knowledge. To address such a challenging problem, we rely on the \"pop-out\" prior, which allows us to reason about the object's location directly in the 3D space. The \"pop-out\" is a simple composition prior that assumes objects reside on the background surface. A graphical illustration of the used \"pop-out\" prior, using the results obtained by our method, is given in Figure 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 435, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 555 + ], + "type": "text", + "content": "The pop-out prior for image composition was successfully used by Kang et al. in [24]. An early work of Treisman has provided an in-depth study of such prior in [62]. In this work, we rely on the same compositional foundation of these works and exploit the pop-out before transferring the depth of knowledge across domains. Although our motivation comes from these early works, our experimental setup largely differs from theirs. We differ in terms of not only depth knowledge transfer across domains (without source data) but also in target supervision using only semantics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "The proposed method exploits the source-free depth to map it into a space where objects in depth stand out better against the background. This mapping is used for object and background separation using a learned contact surface between them. Such separation allows us to derive the semantic masks which can be directly compared against the ground truth for supervision. Using this supervision at the target, we can minimize the domain and task gap between the source and target. The overall framework of our method that performs cross-task cross-domain knowledge transfer by using the intermediate representation of the pop-out space is shown in Figure 3. As can be seen, we first leverage an object popping network to encourage the object" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 240, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 396 + ], + "type": "text", + "content": "to jump out from the source-free depth. Then, we introduce another network i.e., the segmentation with contact surface, to localize the object and predict the contact surface. These learning modules are jointly trained in an end-to-end manner, transferring the source-free depths into intermediate representations which are adapted to the target task, i.e., object detection. To evaluate the proposed method, we conducted exhaustive experiments on eight datasets of two challenging tasks, namely salient object detection and camouflaged object detection. In both tasks, our method significantly improves the established baselines and offers state-of-the-art results at the same time, whose overview can be seen in Figure 2. The major contributions are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 398, + 545, + 469 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 317, + 398, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 398, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 317, + 398, + 545, + 422 + ], + "type": "text", + "content": "- Our problem of transferring source-free depth knowledge across domains and tasks is practical and novel." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 422, + 544, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 544, + 445 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 544, + 445 + ], + "type": "text", + "content": "- Our method relies on our object pop-out prior for visual understanding, which is simple and effective." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 446, + 544, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 446, + 544, + 469 + ], + "spans": [ + { + "bbox": [ + 317, + 446, + 544, + 469 + ], + "type": "text", + "content": "- Results of our method in two different tasks are significantly better than the baselines and existing models." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 478, + 397, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 397, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 397, + 491 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 676 + ], + "type": "text", + "content": "Source-free Adaptation: Knowledge transfer by domain adaptation without access to the source data has recently gathered vast interest [1,25,28,39,71,77], due to the privacy, practicality, and efficiency reasons. We observe accessing the source data for transferring the depth knowledge learned by the off-the-shelf knowledge models [13, 51] is particularly impractical due to their multi-stage training on various datasets. The existing source-free adaptation methods either use generative [28, 29, 34], pseudo-label [25, 37, 64], or other customized [54, 76] approaches. In this work, we use the pseudo-label-based approach. However, using existing methods is not straightforward in our setting due to the tasks' difference between source and target. With our pop-out technique that provides the pseudo semantics, the source-free depth is better transferred across tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": "Salient Object Detection (SOD): Saliency detection aims to detect and segment the most prominent region within an image that visually attracts human attention [11]. A" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "text", + "content": "1033" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 69, + 521, + 220 + ], + "blocks": [ + { + "bbox": [ + 72, + 69, + 521, + 220 + ], + "lines": [ + { + "bbox": [ + 72, + 69, + 521, + 220 + ], + "spans": [ + { + "bbox": [ + 72, + 69, + 521, + 220 + ], + "type": "image", + "image_path": "451fab1cda9c67f92d3f0a55b31ef68d5b893ceeda5b7501ee15edbedad854ef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 222, + 545, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 222, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 545, + 289 + ], + "type": "text", + "content": "Figure 3. Our proposed framework, termed PopNet, is composed of a source-free network, an object popping network, a segmentation with contact surface, and an object separation module. The source-free depth network generates pseudo-depth in an off-the-shelf manner (Section 3.1). The object popping network converts the source-free depth into the popped-out depth of objects, bridging the cross-domain and cross-task gaps (Section 3.2). The segmentation network uses this depth to estimate the object's mask and contact surface (Section 3.3). The object separation module then converts popped-out depth to the second mask of objects using the contact surface (Section 3.4). We compare both semantic masks against ground truth to supervise the whole pipeline in an end-to-end manner." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 302, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 287, + 445 + ], + "type": "text", + "content": "number of works have shown that saliency can be an auxiliary step for different vision tasks such as object tracking [93], object detection [58], etc. Conventional saliency works are unimodal, i.e., they only require RGB images as input. In generic and common settings, RGB-based models [38, 70, 84] have already achieved very promising results. More recently, several works [10,20,49,66,68,88,92] exploit depth maps as additional clues to the 3D geometry since the depth can provide more truthful information on the object boundary as well as the scale awareness. These 3D features further improve the detection accuracy and performance in challenging scenarios [22, 33, 69, 80]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 448, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 287, + 640 + ], + "type": "text", + "content": "Camouflaged Object Detection (COD): Camouflage detection aims to find the preying object within an image. For computer vision society, primary works [9,32,47] often compare COD with SOD. A number of works [8,9,74,79] have shown that simply extending saliency models [70] on COD will lead to undesired results, which is mainly caused by the nature of target object, i.e., concealed or prominent. Hence, to constrain the attention on the concealed objects, several works come up with different perceptual systems that mimic human behavior vis-a-vis camouflaged objects, such as three-stage localize-segment-rank strategy [42]; iterative refinement [21], which is similar to repeatedly looking on the images; zooming into possible regions [47, 59]. Others [18, 40, 60, 82, 83, 90, 94, 95] deeply explore the texture difference with the help of the gradient [18], frequency [90], edges [60, 94], and probability [32, 74]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "The latest psychological studies [5, 7] have shown that human perception can naturally benefit from the depth cues to understand the scene: (A) the smooth variation within the object can contribute to alleviating the fake edges and preserving the object structure; (B) the depth discontinuity on the object boundary can make the segmentation easier. In" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 302, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 545, + 374 + ], + "type": "text", + "content": "spired by these observations, we aim to explore the source-free depth for both SOD and COD tasks. To tackle the domain gap for source-free depth maps, we propose to jointly finetune the source-free depth together with the semantic network in an end-to-end manner, with both self-supervised loss and weak semantic supervision." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 388, + 410, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 388, + 410, + 402 + ], + "spans": [ + { + "bbox": [ + 305, + 388, + 410, + 402 + ], + "type": "text", + "content": "3. Proposed PopNet" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "text", + "content": "Given an input RGB image " + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "text", + "content": " with size " + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "text", + "content": " are the spatial resolutions, i.e., height and width, of the image, our objective is to predict the semantic mask " + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\tilde{S} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "text", + "content": " for object detection. As shown in Figure 3, the input image " + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "text", + "content": " is firstly fed into a frozen-weight depth network to generate the source-free depth " + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "inline_equation", + "content": "D_{sf} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "text", + "content": " (Section 3.1). Then the mimicked multi-modal images are fed together into the depth popping network to compute the intermediate popped-out depth " + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "inline_equation", + "content": "D_{po} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "text", + "content": " (Section 3.2). This intermediate representation, as well as the RGB image " + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "text", + "content": ", is later processed by the segmentation network and transformed into a contact surface " + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "inline_equation", + "content": "D_{c} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "text", + "content": " and a semantics prediction " + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\tilde{S} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "text", + "content": " (Section 3.3). On the one hand, the semantics prediction is directly supervised by the ground truth mask " + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "inline_equation", + "content": "GT" + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 407, + 545, + 649 + ], + "type": "text", + "content": ", which is similar to conventional segmentation supervision. On the other hand, we further explore the contact surface to pop the object out of the background by means of our object separation module (Section 3.4). This transfers the geometric cues into pseudo semantics and leads to another level of supervision." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 658, + 457, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 457, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 457, + 671 + ], + "type": "text", + "content": "3.1. Source-free Depth Network" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "In a practical setting, the GT depth is not always available. Therefore, we generate the source-free depth " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "D_{sf}" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " in an off-the-shelf manner to mimic the multi-modal input." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "text", + "content": "1034" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": "We choose the state-of-the-art DPT model [50] with frozen weights as our source-free depth network, which offers us promising depth at the target. This choice is made upon its generalization capability as suggested in [51]. To obtain the depth map with the highest quality possible, by enhancing local details, we apply the boosting method [45] together with DPT. Despite the plausible results achieved by learning-based methods, the obtained source-free depth does not always offer high-quality geometric cues due to the domain gap. Therefore, we leverage the geometric and semantic priors to jointly finetune the source-free depth." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 213, + 187, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 187, + 225 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 187, + 225 + ], + "type": "text", + "content": "3.2. Object Popping Network" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 231, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 287, + 350 + ], + "type": "text", + "content": "Network Architecture: We build a depth popping network to refine/smooth the source-free depth. The popping network follows the encoder-decoder design with skip connection as shown in Figure 4. In our case, we simply concatenate RGB and source-free depth at the input side to form a 4-channel input and feed them into the popping network. The encoder extracts semantic cues and generates five-scale outputs. Our decoder is composed of Conv2D, BN, ReLU, and upsampling layers. Following U-Net [52], we build a skip connection through simple addition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 352, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 424 + ], + "type": "text", + "content": "Structure Preserving: To supervise our popping network, we first guide the depth refinement with the help of generated pseudo-depth. We only leverage structural similarity since we aim to detect, preserve, and extract object structure from the intermediate representation. We use the following SSIM loss [13] for structural similarity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 434, + 287, + 448 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 434, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 116, + 434, + 287, + 448 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d e p} = S S I M \\left(D _ {p o}, D _ {s f}\\right). \\tag {1}", + "image_path": "0cc932aad9d4463aa0fc1e5f3aaff45f93a0bb29d6c11475abb32f7c1ec7c270.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 457, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 613 + ], + "type": "text", + "content": "Local Depth Smoothing: In addition to the pseudo-depth supervision, we propose two losses to constrain the depth also by semantics. We assume that the objects' structure should be distinguishable from the background, i.e., it should be smooth within the object region and sharp on the bounding pixels. Hence, we propose to leverage the weak semantic cues together with the geometric priors. Specifically, we first introduce a local smoothness loss. The locality is defined by the ground truth semantics " + }, + { + "bbox": [ + 46, + 457, + 287, + 613 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 457, + 287, + 613 + ], + "type": "text", + "content": ". Technically, we mask out background pixels with element-wise multiplication to suppress the inactive area through " + }, + { + "bbox": [ + 46, + 457, + 287, + 613 + ], + "type": "inline_equation", + "content": "D_{obj} = D_{po} \\otimes G" + }, + { + "bbox": [ + 46, + 457, + 287, + 613 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 46, + 457, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\nabla_x" + }, + { + "bbox": [ + 46, + 457, + 287, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 457, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\nabla_y" + }, + { + "bbox": [ + 46, + 457, + 287, + 613 + ], + "type": "text", + "content": " be the Sobel operations. Then, our local loss " + }, + { + "bbox": [ + 46, + 457, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{loc}" + }, + { + "bbox": [ + 46, + 457, + 287, + 613 + ], + "type": "text", + "content": " is expressed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 621, + 287, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 621, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 77, + 621, + 287, + 668 + ], + "type": "interline_equation", + "content": "\\begin{array}{c} \\overrightarrow {n (p)} = (- \\nabla_ {x} (D _ {o b j} (p)); - \\nabla_ {y} (D _ {o b j} (p)); 1); \\\\ \\mathcal {L} _ {l o c} = \\sum_ {p} \\sum_ {q \\in \\mathcal {N} (p)} 1 - \\operatorname {c o s i n e} (\\overrightarrow {n (p)}; \\overrightarrow {n (q)}), \\end{array} \\tag {2}", + "image_path": "eda446e7640339f7a22bf1afa83edcd95b43c2f1f210df59f0b67ebcd3b26857.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\overrightarrow{n}" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": " stands for the normal, " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": " denotes the pixels within the object region, " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(p)" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": " is the neighboring pixels, and cosine is the cosine similarity between two vectors. This way, our" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 69, + 545, + 186 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 545, + 186 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 545, + 186 + ], + "type": "image", + "image_path": "9e4cfb368af124dd46d4911338806fb8feac354156afccce5a7ffabc2555d4bb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 188, + 545, + 234 + ], + "lines": [ + { + "bbox": [ + 305, + 188, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 305, + 188, + 545, + 234 + ], + "type": "text", + "content": "Figure 4. Our object popping network maps RGB-D inputs into popped-out depths. This network is supervised using a combination of structure preserving, local depth smoothing, and depth edge sharpening losses " + }, + { + "bbox": [ + 305, + 188, + 545, + 234 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dep}" + }, + { + "bbox": [ + 305, + 188, + 545, + 234 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 188, + 545, + 234 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{loc}" + }, + { + "bbox": [ + 305, + 188, + 545, + 234 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 188, + 545, + 234 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{wtv}" + }, + { + "bbox": [ + 305, + 188, + 545, + 234 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 247, + 545, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 247, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 545, + 282 + ], + "type": "text", + "content": "local loss only works on the object area, making the object structure consistent within the target region. Applying local smoothness loss reduces the depth noise at the object level." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 284, + 545, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 284, + 545, + 333 + ], + "spans": [ + { + "bbox": [ + 305, + 284, + 545, + 333 + ], + "type": "text", + "content": "Depth Edge Sharpening: In addition to the local smoothness, we also use edge sharpening. The edge sharpening loss is formulated as a weighted total variation. For this, we first compute the edge-aware weight " + }, + { + "bbox": [ + 305, + 284, + 545, + 333 + ], + "type": "inline_equation", + "content": "w(p)" + }, + { + "bbox": [ + 305, + 284, + 545, + 333 + ], + "type": "text", + "content": " at any pixel " + }, + { + "bbox": [ + 305, + 284, + 545, + 333 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 284, + 545, + 333 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 349, + 545, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 349, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 314, + 349, + 545, + 381 + ], + "type": "interline_equation", + "content": "w (p) = \\left\\{ \\begin{array}{l l} w _ {0}, & \\text {i f} \\nabla_ {x} (G (p)) ^ {2} + \\nabla_ {y} (G (p)) ^ {2} \\neq 0, \\\\ w _ {0} + \\gamma , & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {3}", + "image_path": "3be82e84c5775955fd48345396502760aee2f5a2325304f6b3fb4973eb78fdc2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 388, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 388, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 305, + 388, + 545, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 388, + 545, + 460 + ], + "type": "inline_equation", + "content": "w_0" + }, + { + "bbox": [ + 305, + 388, + 545, + 460 + ], + "type": "text", + "content": " is a pre-defined non-zero weight and " + }, + { + "bbox": [ + 305, + 388, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 388, + 545, + 460 + ], + "type": "text", + "content": " is an additional weight for boundary pixels. In our setting, we choose " + }, + { + "bbox": [ + 305, + 388, + 545, + 460 + ], + "type": "inline_equation", + "content": "w_0" + }, + { + "bbox": [ + 305, + 388, + 545, + 460 + ], + "type": "text", + "content": " as the normalized (by the image size) count of the boundary pixels, and we set " + }, + { + "bbox": [ + 305, + 388, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\gamma = 0.5" + }, + { + "bbox": [ + 305, + 388, + 545, + 460 + ], + "type": "text", + "content": ". We adopt the square form such that the large gradients play more important roles. Our weighted total-variation loss is given by," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 333, + 466, + 545, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 466, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 333, + 466, + 545, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {w T v} = \\sum_ {p} \\sum_ {q \\in \\mathcal {N} (p)} w (p) \\cdot | | D _ {p o} (p) - D _ {p o} (q) | | _ {2}. \\tag {4}", + "image_path": "f098008ec9b0db74cbcb03a60a31fa1d0edade0a8c527a3e85ac3a02e33346b9.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "Our weighted total-variance loss differs from the conventional edge loss, due to our weighting function. More specifically, our weighting function relies on semantic boundaries, unlike the commonly used image gradients [14, 17]. Our motivation for using semantic boundaries instead of image gradients comes from our interest in performing object detection under challenging conditions, such as camouflaged objects. In such cases, image gradients may result in misleading weights. At first glance, our loss function seems to be similar to semantic-guided depth estimation methods [4, 35, 72]. However, [72] uses GT depths, and [4, 35] use multi-frames for supervision. Unfortunately, such supervision is not possible in our setup. Note that we exploit single-view source-free depth while only using semantic ground truth for supervision. Furthermore, we transfer the source-free depths' knowledge despite their domain gap. Existing regularizations between depth and semantics are already exploited in our method, which is shown to be" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "text", + "content": "1035" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": "complimentary to our pop-out prior. We argue that our network architecture that exploits the pop-out prior is not trivial while simultaneously being generic and easy to use." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 108, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 108, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 108, + 287, + 133 + ], + "type": "text", + "content": "The total objective function " + }, + { + "bbox": [ + 47, + 108, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pop}" + }, + { + "bbox": [ + 47, + 108, + 287, + 133 + ], + "type": "text", + "content": " to supervise our object popping network is given by," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 94, + 141, + 287, + 155 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 141, + 287, + 155 + ], + "spans": [ + { + "bbox": [ + 94, + 141, + 287, + 155 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p o p} = \\mathcal {L} _ {d e p} + \\lambda_ {1} \\cdot \\mathcal {L} _ {l o c} + \\lambda_ {2} \\cdot \\mathcal {L} _ {w t v}, \\tag {5}", + "image_path": "66c6d87751063c97bc893aae8c7a40c644daecf1868da96049009fcc7c185148.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 163, + 217, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 217, + 175 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 217, + 175 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 163, + 217, + 175 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 47, + 163, + 217, + 175 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 163, + 217, + 175 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 47, + 163, + 217, + 175 + ], + "type": "text", + "content": " are the hyperparameters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 182, + 236, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 182, + 236, + 194 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 236, + 194 + ], + "type": "text", + "content": "3.3. Segmentation with Contact Surface" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 201, + 288, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 288, + 369 + ], + "type": "text", + "content": "The smoothness and edge losses encourage homogenizing the object structure, making it noticeable from the background. We now aim to further enlarge the object-background distance to make the object structure jump out. Specifically, we use an RGB-D segmentation network as shown in Figure 5. The main component of our segmentation network is a three-stream RGB-D network with some fusion design. In our setting, we choose [92] as our baseline since it is one of the SOTA RGB-D methods for saliency detection. We add a surface head to learn the depth of the contact surface " + }, + { + "bbox": [ + 46, + 201, + 288, + 369 + ], + "type": "inline_equation", + "content": "D_{c}" + }, + { + "bbox": [ + 46, + 201, + 288, + 369 + ], + "type": "text", + "content": ", which has the same resolution as the input depth " + }, + { + "bbox": [ + 46, + 201, + 288, + 369 + ], + "type": "inline_equation", + "content": "D_{po}" + }, + { + "bbox": [ + 46, + 201, + 288, + 369 + ], + "type": "text", + "content": ". Our surface head is composed of ConvLayer (Conv2D, BN, ReLU) and a Conv2D, which first decodes the feature maps and then transfers them into a 1-D map." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 376, + 155, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 155, + 388 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 155, + 388 + ], + "type": "text", + "content": "3.4. Object Separation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "type": "text", + "content": "Using the previously discussed contact surface, in this section, we aim to separate the object from its background. At this point, we make an assumption that pixels in front of the contact surface belong to objects. The remaining pixels belong to the background. This assumption allows us to explicitly transfer the 3D knowledge into 2D semantics. Let the predicted depth of the contact surface be " + }, + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "type": "inline_equation", + "content": "D_{c} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "type": "text", + "content": ". We obtain the pseudo semantics " + }, + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "type": "inline_equation", + "content": "S_{s}" + }, + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "type": "text", + "content": ", using the popped-out depth " + }, + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "type": "inline_equation", + "content": "D_{po}" + }, + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "type": "text", + "content": " and surface's depth " + }, + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "type": "inline_equation", + "content": "D_{c}" + }, + { + "bbox": [ + 46, + 394, + 287, + 502 + ], + "type": "text", + "content": ", as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 511, + 287, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 511, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 104, + 511, + 287, + 525 + ], + "type": "interline_equation", + "content": "S _ {s} = \\operatorname {s i g m o i d} \\left(\\sigma \\cdot \\left(D _ {p o} - D _ {c}\\right)\\right), \\tag {6}", + "image_path": "a9c6f49b1ffccb98efa4fa9381b855993d0dcc326e083587720ada7dd601deab.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "content": " is a scalar value that controls the slope of the sigmoid function. In our experiments, we use " + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\sigma = 10" + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "content": " to perform soft-thresholding, mimicking the desired hard one for binary outputs. Such soft thresholding facilitates the gradient back-propagation required for training. Finally, we minimize the gap between the pseudo semantics " + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "inline_equation", + "content": "S_{s}" + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "content": " and the GT semantics " + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "content": " with binary cross-entropy (BCE):" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 125, + 626, + 287, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 626, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 125, + 626, + 287, + 639 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s e p} = B C E \\left(S _ {s}, G\\right). \\tag {7}", + "image_path": "f5898f4ba009a9a0411b1ab18f3acad2107b320b6d498adfa9365dbeefab799d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 647, + 174, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 174, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 174, + 658 + ], + "type": "text", + "content": "3.5. Overall Loss Function" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": "Both of our trainable modules, i.e., object popping and segmentation networks, of our framework (Figure 3) are trained in an end-to-end manner. Therefore, the overall loss function consists of three parts: our depth popping loss" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 69, + 545, + 182 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 545, + 182 + ], + "type": "image", + "image_path": "ee2d166a5549719ca4c61146184fb0125def3e986be754f9511e618701074cb3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 183, + 545, + 228 + ], + "lines": [ + { + "bbox": [ + 305, + 183, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 305, + 183, + 545, + 228 + ], + "type": "text", + "content": "Figure 5. Our segmentation network uses a basic RGB-D three-stream network [92]. In addition to the conventional semantic head, we learn to predict pixel-wise contact surface. The contact surface is later used to transfer depth knowledge to semantics." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 241, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 241, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 241, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pop}" + }, + { + "bbox": [ + 305, + 241, + 545, + 277 + ], + "type": "text", + "content": ", our object separation loss " + }, + { + "bbox": [ + 305, + 241, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sep}" + }, + { + "bbox": [ + 305, + 241, + 545, + 277 + ], + "type": "text", + "content": ", and the conventional semantic loss " + }, + { + "bbox": [ + 305, + 241, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sem}" + }, + { + "bbox": [ + 305, + 241, + 545, + 277 + ], + "type": "text", + "content": " from the RGB-D baseline network. The total loss " + }, + { + "bbox": [ + 305, + 241, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{total}" + }, + { + "bbox": [ + 305, + 241, + 545, + 277 + ], + "type": "text", + "content": " used for training is given by," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 350, + 285, + 545, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 285, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 350, + 285, + 545, + 299 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p o p}} + \\alpha_ {1} \\cdot \\mathcal {L} _ {\\text {s e p}} + \\alpha_ {2} \\cdot \\mathcal {L} _ {\\text {s e m}}, \\tag {8}", + "image_path": "ec3467a207bad3ed41408987ae3d6fa4ad4c489292daef835cc5a3791cc82082.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 307, + 462, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 307, + 462, + 318 + ], + "spans": [ + { + "bbox": [ + 305, + 307, + 462, + 318 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 307, + 462, + 318 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 305, + 307, + 462, + 318 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 307, + 462, + 318 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 305, + 307, + 462, + 318 + ], + "type": "text", + "content": " are hyperparameters." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 319, + 545, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 558 + ], + "type": "text", + "content": "Remarks: Our losses work in a complementary manner. The " + }, + { + "bbox": [ + 304, + 319, + 545, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pop}" + }, + { + "bbox": [ + 304, + 319, + 545, + 558 + ], + "type": "text", + "content": " plays the same role as a smooth filter as in image smoothing. It removes the noisy depth response due to the domain gap while preserving the object structure with the help of a weak semantic label. Hence, the smoothed background becomes less informative while the object region becomes uniform, making it easily detectable. These functionalities contribute to transferring the source-free depth into popped-out depth, as desired. Such a process brings objects above the background surface, despite their distance from the camera and other distracting surfaces. The " + }, + { + "bbox": [ + 304, + 319, + 545, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sep}" + }, + { + "bbox": [ + 304, + 319, + 545, + 558 + ], + "type": "text", + "content": ", on the other hand, fully benefits from the \"pop-out\" prior to segment the object from the background. With the help of the learned contact surface, this loss enlarges the foreground-background distance by pulling them in opposite directions. Such pulling results in a binary-like mask, which effectively bridges the gap between geometric knowledge and semantics. Finally, both depth-transferred and learned semantics are compared to the ground truths for supervised training." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 567, + 359, + 579 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 567, + 359, + 579 + ], + "spans": [ + { + "bbox": [ + 306, + 567, + 359, + 579 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 587, + 422, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 422, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 422, + 601 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Dataset Preparation: To better illustrate the generalizability of our approach, we evaluate the effectiveness of our approach on both SOD and COD benchmarks. We choose four widely used RGB-D SOD datasets, i.e., NLPR [48], NJUK [23], STERE [46], and SIP [10], as well as four COD datasets, i.e., CAMO [30], CHAMELEON [57], COD10K [9], and NC4K [42]. For SOD datasets, we conduct experiments with both GT depth and source-free depth. We follow the conventional learning protocol [19, 69, 92]" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 732, + 315, + 741 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 732, + 315, + 741 + ], + "spans": [ + { + "bbox": [ + 296, + 732, + 315, + 741 + ], + "type": "text", + "content": "1036" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 102, + 545, + 240 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 547, + 102 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 547, + 102 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 547, + 102 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on RGB-D SOD datasets. " + }, + { + "bbox": [ + 47, + 69, + 547, + 102 + ], + "type": "inline_equation", + "content": "\\uparrow (\\downarrow)" + }, + { + "bbox": [ + 47, + 69, + 547, + 102 + ], + "type": "text", + "content": " denotes that the higher (lower) is better. We use the Mean Absolute Error " + }, + { + "bbox": [ + 47, + 69, + 547, + 102 + ], + "type": "inline_equation", + "content": "(M)" + }, + { + "bbox": [ + 47, + 69, + 547, + 102 + ], + "type": "text", + "content": ", max F-measure " + }, + { + "bbox": [ + 47, + 69, + 547, + 102 + ], + "type": "inline_equation", + "content": "(F_{m})" + }, + { + "bbox": [ + 47, + 69, + 547, + 102 + ], + "type": "text", + "content": ", S-measure " + }, + { + "bbox": [ + 47, + 69, + 547, + 102 + ], + "type": "inline_equation", + "content": "(S_{m})" + }, + { + "bbox": [ + 47, + 69, + 547, + 102 + ], + "type": "text", + "content": ", and max E-measure " + }, + { + "bbox": [ + 47, + 69, + 547, + 102 + ], + "type": "inline_equation", + "content": "(E_{m})" + }, + { + "bbox": [ + 47, + 69, + 547, + 102 + ], + "type": "text", + "content": " as evaluation metrics. G.D. stands for GT Depth. Bold denotes the best performance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 102, + 545, + 240 + ], + "lines": [ + { + "bbox": [ + 47, + 102, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 545, + 240 + ], + "type": "table", + "html": "
G.D.Public.Dataset MetricNLPR [48]NJUK [23]STERE [46]SIP [10]
M↓Fm↑Sm↑Em↑M↓Fm↑Sm↑Em↑M↓Fm↑Sm↑Em↑M↓Fm↑Sm↑Em↑
Performance of RGB-D Models Trained with Source-free Depth
XMM21 [85]DFM-Net.027.909.914.944.046.903.895.927.042.906.903.934.067.873.850.891
XTIP22 [63]DCMF.027.915.921.943.044.908.903.929.041.909.907.931.067.873.853.893
XCVPR22 [21]SegMAR.024.923.920.952.036.921.909.941.037.916.907.936.052.893.872.914
XCVPR22 [47]ZoomNet.023.916.919.944.037.926.914.940.037.918.909.938.054.891.868.909
XOursPopNet.022.925.926.956.031.931.920.949.032.922.916.947.046.911.885.926
Performance of RGB-D Models Trained with GT Depth
✓TIP21 [87]BIANet.032.888.900.930.056.878.867.898.048.898.895.918.091.816.802.847
✓TIP21 [33]HAINet.024.920.924.956.037.924.911.940.040.917.907.938.052.907.879.917
✓TNNLS21 [10]D3Net.029.904.911.942.046.909.899.927.044.902.906.925.063.880.860.897
✓ECCV22 [31]SPSN.023.917.923.956.032.927.918.949.035.909.906.941.043.910.891.932
✓OursPopNet.019.927.932.963.030.936.924.952.033.924.917.947.040.923.897.937
", + "image_path": "944bd52b4d893475fbbfec3ef17c0d68f0f818e186a4cdd61d65a30363d949d6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 246, + 287, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 287, + 342 + ], + "type": "text", + "content": "and use 700 images from NLPR and 1,485 images from NJUK for training. The rest are used for testing. For the unimodal COD dataset, we compare with both RGB COD models and RGB-D SOD models retrained on the COD datasets with the same source-free depth " + }, + { + "bbox": [ + 46, + 246, + 287, + 342 + ], + "type": "inline_equation", + "content": "D_{sf}" + }, + { + "bbox": [ + 46, + 246, + 287, + 342 + ], + "type": "text", + "content": ". We follow the conventional training/testing protocol [8, 9, 21, 42, 47] and use 3,040 images from COD10K and 1,000 images from CAMO for training. The rest are used for testing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 342, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 425 + ], + "type": "text", + "content": "Evaluation Metrics: We evaluate the performance with four generally-recognized metrics: Mean Absolute Error " + }, + { + "bbox": [ + 46, + 342, + 287, + 425 + ], + "type": "inline_equation", + "content": "(M)" + }, + { + "bbox": [ + 46, + 342, + 287, + 425 + ], + "type": "text", + "content": ", max F-measure " + }, + { + "bbox": [ + 46, + 342, + 287, + 425 + ], + "type": "inline_equation", + "content": "(F_{m})" + }, + { + "bbox": [ + 46, + 342, + 287, + 425 + ], + "type": "text", + "content": ", S-measure " + }, + { + "bbox": [ + 46, + 342, + 287, + 425 + ], + "type": "inline_equation", + "content": "(S_{m})" + }, + { + "bbox": [ + 46, + 342, + 287, + 425 + ], + "type": "text", + "content": ", and max E-measure " + }, + { + "bbox": [ + 46, + 342, + 287, + 425 + ], + "type": "inline_equation", + "content": "(E_{m})" + }, + { + "bbox": [ + 46, + 342, + 287, + 425 + ], + "type": "text", + "content": ". All the object segmentation masks are trained or downloaded from the official resources. To make a fair comparison, we evaluate the prediction semantics with the standardized evaluation protocol as [92]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 288, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 546 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 546 + ], + "type": "text", + "content": "Implementation Details: Our model is implemented based on Pytorch with a V100 GPU. We use the Adam algorithm as an optimizer. The learning rate is initialized to " + }, + { + "bbox": [ + 46, + 426, + 288, + 546 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 46, + 426, + 288, + 546 + ], + "type": "text", + "content": " and is further divided by 10 every 60 epochs. We set the input resolution to " + }, + { + "bbox": [ + 46, + 426, + 288, + 546 + ], + "type": "inline_equation", + "content": "352 \\times 352" + }, + { + "bbox": [ + 46, + 426, + 288, + 546 + ], + "type": "text", + "content": " resolution for RGB and depth. A detailed comparison with higher resolution can be found in Table 3. During training, conventional data augmentation such as random flipping, rotating, and border clipping are adopted. The training takes around 6 hours for RGB-SOD tasks and 12 hours for COD tasks for 100 epochs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 552, + 133, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 133, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 133, + 564 + ], + "type": "text", + "content": "4.2. Comparisons" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Comparison with RGB-D SOD Models: We present in Table 1 the performance on SOD benchmarks with our source-free depth or with GT depth. It can be seen that our model with source-free depth achieves very competitive performance compared to many RGB-D models with GT depths. Our method with GT depth also outperforms the SOTA counterparts. The qualitative comparison can be found in Fig. 6. Note that we also retrain the SOTA unimodal COD models SegMAR [21] and ZoomNet [47] on the SOD dataset only with RGB images. We show that our model with source-free depth outperforms these counterparts, showing that our method can better generalize across" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 245, + 545, + 358 + ], + "blocks": [ + { + "bbox": [ + 307, + 245, + 545, + 358 + ], + "lines": [ + { + "bbox": [ + 307, + 245, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 545, + 358 + ], + "type": "image", + "image_path": "a7c7c99ae722615d051e65dcf5584a4ec196b3e6b178e04ae235ac4098b11e60.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 359, + 545, + 392 + ], + "lines": [ + { + "bbox": [ + 305, + 359, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 359, + 545, + 392 + ], + "type": "text", + "content": "Figure 6. Qualitative Comparison with GT Depth. Our method outperforms all counterparts while dealing with multi-objects, large depth variation, and visually-mixed foreground-background." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 400, + 506, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 400, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 305, + 400, + 506, + 412 + ], + "type": "text", + "content": "different tasks with favorably better performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 413, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 568 + ], + "type": "text", + "content": "Comparison with COD Models: We present in Table 2 the performance of the most competitive SOTA methods, including task-specific COD models as well as retrained RGB-D SOD models with source-free depth. For a fair comparison, we retrain all RGB-D methods, SegMAR [21], and ZoomNet [47] in an end-to-end manner on the same resolution images as ours. Some RGB-only methods perform even better than many RGB-D methods, mainly due to their COD task-specific designs and the lack of ground-truth depths required for RGB-D methods. It is important to note that competing RGB-only COD methods do not perform favorably on SOD tasks. Please, refer to Table 1 & 3 to observe their poor cross-tasks generalization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Towards Higher Resolution: Previous studies have shown that the image resolution may influence the model performance [43, 47, 73, 83]. For example, the current SOTA COD method ZoomNet [47] is with main scale " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "384^2" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " and implies the highest resolution of " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "(384 \\times 1.5)^2 = 576^2" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ", as it operates on " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "0.5 \\times 1 \\times" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "1.5 \\times" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " scales. To make a fair comparison, we retrain the model with the same resolution " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "(352^2" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "512^2)" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " as ours. We show in Table3-ZoomNet* that the results deteriorate as expected. Compared to these counterparts, our method offers a good trade-off between accuracy and efficiency. More comparisons on SOD and COD benchmarks can be found in the supplementary material." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "text", + "content": "1037" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 79, + 547, + 275 + ], + "blocks": [ + { + "bbox": [ + 88, + 69, + 503, + 79 + ], + "lines": [ + { + "bbox": [ + 88, + 69, + 503, + 79 + ], + "spans": [ + { + "bbox": [ + 88, + 69, + 503, + 79 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on COD datasets. Pseudo stands for source-free depth used for RGB-D methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 79, + 547, + 275 + ], + "lines": [ + { + "bbox": [ + 47, + 79, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 79, + 547, + 275 + ], + "type": "table", + "html": "
Pseudo Public.Dataset MetricCAMO [30]CHAMELEON [57]COD10K [9]NC4K [42]
M ↓Fm ↑Sm ↑Em ↑M ↓Fm ↑Sm ↑Em ↑M ↓Fm ↑Sm ↑Em ↑M ↓Fm ↑Sm ↑Em ↑
Performance of RGB COD Models
X CVPR20 [9]SINet.099.762.751.790.044.845.868.908.051.708.771.832.058.804.808.873
X CVPR21 [42]SLSR.080.791.787.843.030.866.889.938.037.756.804.854.048.836.839.898
X CVPR21 [78]MGL-R.088.791.775.820.031.868.893.932.035.767.813.874.053.828.832.876
X CVPR21 [44]PFNet.085.793.782.845.033.859.882.927.040.747.800.880.053.820.829.891
X CVPR21 [32]UJSC.072.812.800.861.030.874.891.948.035.761.808.886.047.838.841.900
X IJCAI21 [59]C2FNet.079.802.796.856.032.871.888.936.036.764.813.894.049.831.838.898
X ICCV21 [74]UGTR.086.800.783.829.031.862.887.926.036.769.816.873.052.831.839.884
X CVPR22 [21]SegMAR.080.799.794.857.032.871.887.935.039.750.799.876.050.828.836.893
X CVPR22 [47]ZoomNet.074.818.801.858.033.829.859.915.034.771.808.872.045.841.843.893
Performance of RGB-D Models Retrained with Source-free Depth
✓ MM21 [80]CDINet.100.638.732.766.036.787.879.903.044.610.778.821.067.697.793.830
✓ CVPR21 [19]DCF.089.724.749.834.037.821.850.923.040.685.766.864.061.765.791.878
✓ ICCV21 [81]CMINet.087.798.782.827.032.881.891.930.039.768.811.868.053.832.839.888
✓ ICCV21 [92]SPNet.083.807.783.831.033.872.888.930.037.776.808.869.054.828.825.874
✓ TIP22 [63]DCMF.115.737.728.757.059.807.830.853.063.679.748.776.077.782.794.820
✓ ECCV22 [31]SPSN.084.782.773.829.032.866.887.932.042.727.789.854.059.803.813.867
✓ OursPopNet.073.821.806.869.022.893.910.962.031.789.827.897.043.852.852.908
", + "image_path": "ab569332d5085f97f013e7fb9c93c3e93a24e06a2879d19a2e76034e71533c4e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 289, + 94, + 377 + ], + "blocks": [ + { + "bbox": [ + 48, + 289, + 94, + 377 + ], + "lines": [ + { + "bbox": [ + 48, + 289, + 94, + 377 + ], + "spans": [ + { + "bbox": [ + 48, + 289, + 94, + 377 + ], + "type": "image", + "image_path": "313a4908b5d16c66ace5c12b698a042f87c9b8097dd66ad77b2c96d0d0ad1135.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 378, + 84, + 388 + ], + "lines": [ + { + "bbox": [ + 60, + 378, + 84, + 388 + ], + "spans": [ + { + "bbox": [ + 60, + 378, + 84, + 388 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 94, + 289, + 138, + 377 + ], + "blocks": [ + { + "bbox": [ + 94, + 289, + 138, + 377 + ], + "lines": [ + { + "bbox": [ + 94, + 289, + 138, + 377 + ], + "spans": [ + { + "bbox": [ + 94, + 289, + 138, + 377 + ], + "type": "image", + "image_path": "fc2c24b00d10f08b7bdd7dd8f2964a1fd652a889b097e89c340d9031beb183a4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 378, + 123, + 386 + ], + "lines": [ + { + "bbox": [ + 110, + 378, + 123, + 386 + ], + "spans": [ + { + "bbox": [ + 110, + 378, + 123, + 386 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 138, + 289, + 184, + 377 + ], + "blocks": [ + { + "bbox": [ + 138, + 289, + 184, + 377 + ], + "lines": [ + { + "bbox": [ + 138, + 289, + 184, + 377 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 184, + 377 + ], + "type": "image", + "image_path": "02e0ff4e078e32ba845abdbfe3c8f40e834636ad230a059bf03c238dd086bd8a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 378, + 172, + 387 + ], + "lines": [ + { + "bbox": [ + 153, + 378, + 172, + 387 + ], + "spans": [ + { + "bbox": [ + 153, + 378, + 172, + 387 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 184, + 289, + 228, + 377 + ], + "blocks": [ + { + "bbox": [ + 184, + 289, + 228, + 377 + ], + "lines": [ + { + "bbox": [ + 184, + 289, + 228, + 377 + ], + "spans": [ + { + "bbox": [ + 184, + 289, + 228, + 377 + ], + "type": "image", + "image_path": "3c2f5971a18abf0bc7be5d0538c3d65b89663431c1cc977dfa8c96e7b47d0feb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 378, + 226, + 387 + ], + "lines": [ + { + "bbox": [ + 189, + 378, + 226, + 387 + ], + "spans": [ + { + "bbox": [ + 189, + 378, + 226, + 387 + ], + "type": "text", + "content": "ZoomNet" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 388, + 545, + 411 + ], + "lines": [ + { + "bbox": [ + 46, + 388, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 545, + 411 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison. Our method can better preserve the object structure compared to other counterparts, especially while dealing with objects with occlusion (two first rows). Our method performs favorably with multiple objects (last row). Better to zoom in." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 229, + 289, + 274, + 377 + ], + "blocks": [ + { + "bbox": [ + 229, + 289, + 274, + 377 + ], + "lines": [ + { + "bbox": [ + 229, + 289, + 274, + 377 + ], + "spans": [ + { + "bbox": [ + 229, + 289, + 274, + 377 + ], + "type": "image", + "image_path": "151245e316fa0105cc34499f11c453927f83187910a3a9ab87897251828f94a6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 378, + 265, + 387 + ], + "lines": [ + { + "bbox": [ + 239, + 378, + 265, + 387 + ], + "spans": [ + { + "bbox": [ + 239, + 378, + 265, + 387 + ], + "type": "text", + "content": "UGTR" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 275, + 289, + 320, + 377 + ], + "blocks": [ + { + "bbox": [ + 275, + 289, + 320, + 377 + ], + "lines": [ + { + "bbox": [ + 275, + 289, + 320, + 377 + ], + "spans": [ + { + "bbox": [ + 275, + 289, + 320, + 377 + ], + "type": "image", + "image_path": "f9bc22628844ea3d45b3b01616fdfda185a8ef4ad372170fcd5231926718472d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 378, + 306, + 387 + ], + "lines": [ + { + "bbox": [ + 285, + 378, + 306, + 387 + ], + "spans": [ + { + "bbox": [ + 285, + 378, + 306, + 387 + ], + "type": "text", + "content": "UJSC" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 320, + 289, + 365, + 377 + ], + "blocks": [ + { + "bbox": [ + 320, + 289, + 365, + 377 + ], + "lines": [ + { + "bbox": [ + 320, + 289, + 365, + 377 + ], + "spans": [ + { + "bbox": [ + 320, + 289, + 365, + 377 + ], + "type": "image", + "image_path": "b82db1fbe06a43661851b35db1db411cd5eef8b20f4dbe670b0f0677c5528f21.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 378, + 356, + 387 + ], + "lines": [ + { + "bbox": [ + 326, + 378, + 356, + 387 + ], + "spans": [ + { + "bbox": [ + 326, + 378, + 356, + 387 + ], + "type": "text", + "content": "C2FNet" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 365, + 289, + 454, + 377 + ], + "blocks": [ + { + "bbox": [ + 365, + 289, + 454, + 377 + ], + "lines": [ + { + "bbox": [ + 365, + 289, + 454, + 377 + ], + "spans": [ + { + "bbox": [ + 365, + 289, + 454, + 377 + ], + "type": "image", + "image_path": "ce3b530f65c1510010c26bc45654e97d0cb3cf9a76f72ed608331043141d7995.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 378, + 399, + 387 + ], + "lines": [ + { + "bbox": [ + 374, + 378, + 399, + 387 + ], + "spans": [ + { + "bbox": [ + 374, + 378, + 399, + 387 + ], + "type": "text", + "content": "PFNet" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 415, + 378, + 447, + 387 + ], + "lines": [ + { + "bbox": [ + 415, + 378, + 447, + 387 + ], + "spans": [ + { + "bbox": [ + 415, + 378, + 447, + 387 + ], + "type": "text", + "content": "MGL-R" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 400, + 289, + 455, + 377 + ], + "blocks": [ + { + "bbox": [ + 400, + 289, + 455, + 377 + ], + "lines": [ + { + "bbox": [ + 400, + 289, + 455, + 377 + ], + "spans": [ + { + "bbox": [ + 400, + 289, + 455, + 377 + ], + "type": "image", + "image_path": "24fff8867aec7c6c7d44ac4e9c274c5984dd0ea2789e2b6032edbedd0d2b9efd.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 455, + 289, + 544, + 377 + ], + "blocks": [ + { + "bbox": [ + 455, + 289, + 544, + 377 + ], + "lines": [ + { + "bbox": [ + 455, + 289, + 544, + 377 + ], + "spans": [ + { + "bbox": [ + 455, + 289, + 544, + 377 + ], + "type": "image", + "image_path": "a7c88e08c35af38c8a95c03fc87ee3a3fcf879dd6fffb226d4dd8f82e51a2e84.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 378, + 487, + 387 + ], + "lines": [ + { + "bbox": [ + 463, + 378, + 487, + 387 + ], + "spans": [ + { + "bbox": [ + 463, + 378, + 487, + 387 + ], + "type": "text", + "content": "SLSR" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 511, + 378, + 533, + 387 + ], + "lines": [ + { + "bbox": [ + 511, + 378, + 533, + 387 + ], + "spans": [ + { + "bbox": [ + 511, + 378, + 533, + 387 + ], + "type": "text", + "content": "SINet" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 511, + 289, + 544, + 377 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 511, + 289, + 544, + 377 + ], + "spans": [ + { + "bbox": [ + 511, + 289, + 544, + 377 + ], + "type": "image", + "image_path": "2b52de0fb3bdeb9c02382dfcb9a4df68a87b116295f15943b09f09cd796f74b7.jpg" + } + ] + } + ], + "index": 22, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 49, + 452, + 285, + 523 + ], + "blocks": [ + { + "bbox": [ + 46, + 419, + 287, + 452 + ], + "lines": [ + { + "bbox": [ + 46, + 419, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 287, + 452 + ], + "type": "text", + "content": "Table 3. End-to-end comparison with different resolutions on SOD benchmarks. Our method with source-free depth generalizes significantly better compared to SOTA COD models." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 452, + 285, + 523 + ], + "lines": [ + { + "bbox": [ + 49, + 452, + 285, + 523 + ], + "spans": [ + { + "bbox": [ + 49, + 452, + 285, + 523 + ], + "type": "table", + "html": "
ModelSizeFlops (G)NJUK [23]SIP [10]
M↓Fm↑Sm↑Em↑M↓Fm↑Sm↑Em↑
SegMAR [21]352267.3.036.921.909.941.052.893.872.914
ZoomNet [47]3522167.8.037.926.914.940.054.891.868.909
Ours3522228.8.031.931.920.949.046.911.885.926
SegMAR [21]5122142.4.035.927.914.943.050.899.878.917
ZoomNet [47]5122353.4.036.926.915.942.052.895.873.910
Ours5122484.0.031.933.922.951.044.911.890.927
", + "image_path": "dbd30ca728e0bdc57c014372bf7486b4640f242b948a2ee4ae7c2929c510bd46.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 539, + 287, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 622 + ], + "type": "text", + "content": "Qualitative Comparisons: Figure 7 presents the output of our network on challenging cases. It can be seen while dealing with objects occluded by thin objects (" + }, + { + "bbox": [ + 46, + 539, + 287, + 622 + ], + "type": "inline_equation", + "content": "1^{st} - 4^{th}" + }, + { + "bbox": [ + 46, + 539, + 287, + 622 + ], + "type": "text", + "content": " rows), our method can accurately reason about the segmentation masks closer to the GT. We also achieve better performance while dealing with multiple objects (last row). More discussions on multiple objects can be found in Table 6." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 47, + 634, + 141, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 141, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 141, + 647 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Loss: In this section, we conduct experiments on analyzing the effectiveness of the proposed losses. The quantitative results of different loss combinations are provided in Table 4. It can be seen that each proposed loss behaves properly, i.e., improving the performance compared to the baseline. More" + } + ] + } + ], + "index": 29 + }, + { + "type": "table", + "bbox": [ + 310, + 430, + 542, + 533 + ], + "blocks": [ + { + "bbox": [ + 339, + 419, + 511, + 430 + ], + "lines": [ + { + "bbox": [ + 339, + 419, + 511, + 430 + ], + "spans": [ + { + "bbox": [ + 339, + 419, + 511, + 430 + ], + "type": "text", + "content": "Table 4. Ablation study on the proposed losses." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 430, + 542, + 533 + ], + "lines": [ + { + "bbox": [ + 310, + 430, + 542, + 533 + ], + "spans": [ + { + "bbox": [ + 310, + 430, + 542, + 533 + ], + "type": "table", + "html": "
\\( \\mathcal{L}_{dep} \\)\\( \\mathcal{L}_{loc} \\)\\( \\mathcal{L}_{wtv} \\)\\( \\mathcal{L}_{sep} \\)SIP [10]NC4K [42]
\\( M\\downarrow \\)\\( F_m\\uparrow \\)\\( S_m\\uparrow \\)\\( E_m\\uparrow \\)\\( M\\downarrow \\)\\( F_m\\uparrow \\)\\( S_m\\uparrow \\)\\( E_m\\uparrow \\)
----.048.903.884.922.052.832.832.893
✓---.046.907.889.925.051.833.839.895
-✓--.045.908.893.929.048.837.844.898
--✓-.046.906.891.927.050.833.841.894
---✓.043.914.893.933.048.840.848.900
✓✓--.044.911.893.928.049.837.844.897
✓-✓-.046.909.893.927.046.840.845.898
✓-✓✓.040.918.897.935.045.848.849.904
✓✓-✓.042.916.894.931.044.850.850.906
✓✓✓✓.040.923.897.937.043.852.852.908
", + "image_path": "e7b360863ad44dcf4cddebac3906002fece7ccf2bc4fb9209df448c668c90a9d.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 545, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 569 + ], + "type": "text", + "content": "discussions and ablation studies on the hyperparameters can be found in the supplementary material." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "text", + "content": "Object Popping Network as Plug-in: Our object popping network can be easily adapted with different encoders and with different existing RGB-D models. For example, with a ResNet-18 [15] encoder and convolution-based decoder, our poping only costs around 12.7M additional learning parameters or 48.7 MB model size. We show in Table 5 that our method can favorably improve performance over the baseline with less than an extra " + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "text", + "content": " GFlops." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Pop-out Under Reduced Training Data: Here, we are interested in analyzing source-free depth's benefits. Therefore, we conduct different experiments by reducing the training data. As shown in Figure 8(left), when both our" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "text", + "content": "1038" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 69, + 286, + 131 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 286, + 131 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 286, + 131 + ], + "type": "text", + "content": "Table 5. Generalization and cost over different RGB-D baselines. Dataset Metric Flops Param S1P [10] NC4K [42] M " + }, + { + "bbox": [ + 48, + 69, + 286, + 131 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 48, + 69, + 286, + 131 + ], + "type": "inline_equation", + "content": "{F}_{m} \\uparrow" + }, + { + "bbox": [ + 48, + 69, + 286, + 131 + ], + "type": "inline_equation", + "content": "{S}_{m} \\uparrow" + }, + { + "bbox": [ + 48, + 69, + 286, + 131 + ], + "type": "inline_equation", + "content": "{E}_{m} \\uparrow" + }, + { + "bbox": [ + 48, + 69, + 286, + 131 + ], + "type": "inline_equation", + "content": "{M\\downarrow }" + }, + { + "bbox": [ + 48, + 69, + 286, + 131 + ], + "type": "inline_equation", + "content": "{F}_{m} \\uparrow" + }, + { + "bbox": [ + 48, + 69, + 286, + 131 + ], + "type": "inline_equation", + "content": "{S}_{m} \\uparrow" + }, + { + "bbox": [ + 48, + 69, + 286, + 131 + ], + "type": "inline_equation", + "content": "{E}_{m} \\uparrow" + }, + { + "bbox": [ + 48, + 69, + 286, + 131 + ], + "type": "text", + "content": " +Ours 363.2 59.8 .053 .899 .874 .919 .057 .809 .804 .872 373.7 72.5 .051 .910 .886 .923 .055 .814 .811 .878 SPNet [92] 149.0 150.4 .044 .911 .887 .914 .054 .828 .825 .874 +Ours 159.5 163.1 .042 .917 .894 .932 .044 .851 .851 .905" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 141, + 286, + 258 + ], + "blocks": [ + { + "bbox": [ + 50, + 141, + 286, + 258 + ], + "lines": [ + { + "bbox": [ + 50, + 141, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 50, + 141, + 286, + 258 + ], + "type": "image", + "image_path": "228225256ac391b291cd02e59acc6601ce70bd569a76eddd625e1174b20fb95d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 260, + 287, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 260, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 287, + 316 + ], + "type": "text", + "content": "Figure 8. Benefit of our method on reducing training data. When our network is trained only with " + }, + { + "bbox": [ + 46, + 260, + 287, + 316 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 260, + 287, + 316 + ], + "type": "text", + "content": " data, the performance remains competitive compared to the baseline, i.e., the absolute performance gains are " + }, + { + "bbox": [ + 46, + 260, + 287, + 316 + ], + "type": "inline_equation", + "content": "+0.4\\%" + }, + { + "bbox": [ + 46, + 260, + 287, + 316 + ], + "type": "text", + "content": " max F-measure and " + }, + { + "bbox": [ + 46, + 260, + 287, + 316 + ], + "type": "inline_equation", + "content": "+0.1\\%" + }, + { + "bbox": [ + 46, + 260, + 287, + 316 + ], + "type": "text", + "content": " S-measure on NC4K compared to the baseline trained with all data." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 322, + 264, + 429 + ], + "blocks": [ + { + "bbox": [ + 48, + 322, + 264, + 429 + ], + "lines": [ + { + "bbox": [ + 48, + 322, + 264, + 429 + ], + "spans": [ + { + "bbox": [ + 48, + 322, + 264, + 429 + ], + "type": "image", + "image_path": "c0240fe18a2b8fcba530dfb9589bae2379566a86fac6320b5455982b53e654cc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 429, + 264, + 439 + ], + "lines": [ + { + "bbox": [ + 47, + 429, + 264, + 439 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 264, + 439 + ], + "type": "text", + "content": "Figure 9. Histogram on the gain. Please zoom in for details." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 459, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 567 + ], + "type": "text", + "content": "PopNet and our baseline are trained with all data, our PopNet can lead to absolute improvements with " + }, + { + "bbox": [ + 46, + 459, + 287, + 567 + ], + "type": "inline_equation", + "content": "5.6\\%" + }, + { + "bbox": [ + 46, + 459, + 287, + 567 + ], + "type": "text", + "content": " in max F-measure and with " + }, + { + "bbox": [ + 46, + 459, + 287, + 567 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 46, + 459, + 287, + 567 + ], + "type": "text", + "content": " in S-measure on COD10K dataset. While our PopNet is trained with only " + }, + { + "bbox": [ + 46, + 459, + 287, + 567 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 459, + 287, + 567 + ], + "type": "text", + "content": " data, it can still achieve competitive performance compared to our baseline trained with all data. Similar phenomena can be observed in the NC4K dataset as shown in Figure 8(right). To conclude, our method can efficiently explore the geometric prior and significantly reduce the required training data volume." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 568, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 664 + ], + "type": "text", + "content": "Gain over baseline: With depth cues, as shown in Figure 9, we boost the performance in 3125 over 4121 images (" + }, + { + "bbox": [ + 46, + 568, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\sim 75\\%" + }, + { + "bbox": [ + 46, + 568, + 287, + 664 + ], + "type": "text", + "content": " cases). We also show in Table 6 that our network performs favorably over baselines with single or multiple objects. Our method may fail when the source-free depth is clueless. This mainly happens when the object is well concealed and fools the depth network. However, such cases are also challenging, even for humans." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 665, + 286, + 677 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 286, + 677 + ], + "type": "text", + "content": "Are RGB-D Methods Better Than RGB-only Methods?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "RGB-D methods are indeed better than RGB-only methods provided GT depth. However, only a few RGB-D methods can benefit from the source-free depth. For example," + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 82, + 543, + 126 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 543, + 81 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 543, + 81 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 543, + 81 + ], + "type": "text", + "content": "Table 6. Multi-Object performance on NC4K [42] with size " + }, + { + "bbox": [ + 306, + 70, + 543, + 81 + ], + "type": "inline_equation", + "content": "{512}^{2}" + }, + { + "bbox": [ + 306, + 70, + 543, + 81 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 82, + 543, + 126 + ], + "lines": [ + { + "bbox": [ + 308, + 82, + 543, + 126 + ], + "spans": [ + { + "bbox": [ + 308, + 82, + 543, + 126 + ], + "type": "table", + "html": "
Obj. Nbr. (%)MetricSingle (92%)Two (6%)More (2%)Overall
M↓Fm↑M↓Fm↑M↓Fm↑M↓Fm↑
RGB Baseline.054.828.067.811.091.738.056.825
+ Dxf.050.842.063.821.093.746.051.839
Ours.040.864.051.847.079.767.042.861
", + "image_path": "3dfe58a2ff7eb57aaa9ec93e81d023aa58244b72a2a1fa92e4a2cf8811a37c6b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 307, + 156, + 545, + 209 + ], + "blocks": [ + { + "bbox": [ + 306, + 133, + 545, + 155 + ], + "lines": [ + { + "bbox": [ + 306, + 133, + 545, + 155 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 545, + 155 + ], + "type": "text", + "content": "Table 7. Performance with RGB- " + }, + { + "bbox": [ + 306, + 133, + 545, + 155 + ], + "type": "inline_equation", + "content": "D_{sf}" + }, + { + "bbox": [ + 306, + 133, + 545, + 155 + ], + "type": "text", + "content": " model vs. with RGB-only baseline. " + }, + { + "bbox": [ + 306, + 133, + 545, + 155 + ], + "type": "inline_equation", + "content": "D_{sf}" + }, + { + "bbox": [ + 306, + 133, + 545, + 155 + ], + "type": "text", + "content": " stands for source-free depth." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 156, + 545, + 209 + ], + "lines": [ + { + "bbox": [ + 307, + 156, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 156, + 545, + 209 + ], + "type": "table", + "html": "
DatasetCOD10K [9]NC4K [42]
MetricM↓Fm↑Sm↑Em↑M↓Fm↑Sm↑Em↑
DASNet [88].041.643.793.864.055.747.830.879
+ Dsf.041.642.796.858.055.743.830.874
SPNet [92].040.743.801.867.052.846.833.883
+ Dsf.037.776.808.869.054.828.825.874
", + "image_path": "cb29a7a53d2039aa839450ea039ff04e04c737fc9c2d2acc131dd4124efa0215.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 228, + 546, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 546, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 546, + 408 + ], + "type": "text", + "content": "as shown in Table 7, DASNet [88] achieves poorer performance compared to RGB baseline when trained with source-free depth. Similarly, even for one of the SOTA RGB-D models, SPNet [92], the performance on NC4K dataset with RGB-only input is better than with additional source-free depth. Moreover, when provided with source-free depth, none of the existing RGB-D methods outperform the best performing RGB-only method (e.g., ZoomNet [47]) on COD. The same observation was also made on SOD as well. This could be because of the domain gap coupled with the fusion design, among others. We also found it non-trivial to extend the most well-performing RGB-only methods to the RGB-D case. Note that our PopNet performs better than all existing RGB-only and RGB-D methods, with source-free or GT depth maps." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 426, + 378, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 378, + 438 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 378, + 438 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 448, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 545, + 639 + ], + "type": "text", + "content": "We demonstrate a successful case of cross-domain cross-task depth to semantics knowledge transfer using only the source model. In this paper, the source-free depth at the target offered by a given source model is used. The proposed method learns to transfer knowledge from depth to semantics using the objects' pop-out prior. We facilitate our network to use such prior by designing a novel network architecture. The designed network reasons about the objects by popping them out from the provided depth maps. This process is followed by separating objects from the background using the learned contact surface. We show that the joint learning of object pop-out and contact surface can be successfully supervised using the target semantics. Exhaustive experiments on SOD and COD benchmarks show the successful transfer of depth knowledge to the target, in terms of improved performance and generalization." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgement The authors thank the anonymous reviewers and ACs for their tremendous efforts and helpful comments. This research is financed in part by the Conseil Régional de Bourgogne-Franche-Comté, Toyota Motor Europe (research project TRACE-Zurich), and the Alexander von Humboldt Foundation." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "text", + "content": "1039" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Peshal Agarwal, Danda Pani Paudel, Jan-Nico Zaech, and Luc Van Gool. Unsupervised robust domain adaptation without source data. In WACV, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "type": "text", + "content": "[2] Amir Atapour-Abarghouei and Toby P Breckon. Real-time monocular depth estimation using synthetic data with domain adaptation via image style transfer. In IEEE CVPR, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 288, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 288, + 214 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 288, + 214 + ], + "type": "text", + "content": "[3] Adriano Cardace, Luca De Luigi, Pierluigi Zama Ramirez, Samuele Salti, and Luigi Di Stefano. Plugging self-supervised monocular depth into unsupervised domain adaptation for semantic segmentation. In WACV, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 288, + 281 + ], + "type": "text", + "content": "[4] Po-Yi Chen, Alexander H Liu, Yen-Cheng Liu, and Yu-Chiang Frank Wang. Towards scene understanding: Unsupervised monocular depth estimation with semantic-aware representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 282, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 282, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 53, + 282, + 287, + 326 + ], + "type": "text", + "content": "[5] Innes C Cuthill, Martin Stevens, Jenna Sheppard, Tracey Maddocks, C Alejandro Parraga, and Tom S Troscianko. Disruptive coloration and background pattern matching. Nature, 434(7029):72-74, 2005. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 327, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 327, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 53, + 327, + 287, + 360 + ], + "type": "text", + "content": "[6] Jiahua Dong, Zhen Fang, Anjin Liu, Gan Sun, and Tongliang Liu. Confident anchor-induced multi-source free domain adaptation. NeurIPS, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 361, + 287, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 361, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 53, + 361, + 287, + 405 + ], + "type": "text", + "content": "[7] Aliya El Nagar, Daniel Osorio, Sarah Zylinski, and Steven M Sait. Visual perception and camouflage response to 3d backgrounds and cast shadows in the european cuttlefish, sepia officinalis. JEB, 224(11), 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 407, + 287, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 287, + 439 + ], + "type": "text", + "content": "[8] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE TPAMI, 44(10):6024-6042, 2022. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 441, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 441, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 441, + 287, + 474 + ], + "type": "text", + "content": "[9] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In IEEE CVPR, 2020. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 475, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 287, + 518 + ], + "type": "text", + "content": "[10] Deng-Ping Fan, Zheng Lin, Zhao Zhang, Menglong Zhu, and Ming-Ming Cheng. Rethinking RGB-D salient object detection: Models, datasets, and large-scale benchmarks. IEEE TNNLS, 32(5):2075-2089, 2021. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 520, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 552 + ], + "type": "text", + "content": "[11] Deng-Ping Fan, Jing Zhang, Gang Xu, Ming-Ming Cheng, and Ling Shao. Salient objects in clutter. IEEE TPAMI, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 554, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 554, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 554, + 287, + 587 + ], + "type": "text", + "content": "[12] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In IEEE CVPR, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "type": "text", + "content": "[13] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In IEEE ICCV, 2019. 1, 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 623, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 287, + 655 + ], + "type": "text", + "content": "[14] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In IEEE CVPR, 2022. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE CVPR, 2016. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[16] Ian P Howard. Perceiving in depth, volume 1: basic mechanisms. Oxford University Press, 2012. 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[17] Junjie Hu, Mete Ozay, Yan Zhang, and Takayuki Okatani. Revisiting single image depth estimation: Toward higher resolution maps with accurate object boundaries. In WACV, 2019. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "text", + "content": "[18] Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection. MIR, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[19] Wei Ji, Jingjing Li, Shuang Yu, Miao Zhang, Yongri Piao, Shunyu Yao, Qi Bi, Kai Ma, Yefeng Zheng, Huchuan Lu, et al. Calibrated RGB-D salient object detection. In IEEE CVPR, 2021. 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 207, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 240 + ], + "type": "text", + "content": "[20] Wei Ji, Jingjing Li, Miao Zhang, Yongri Piao, and Huchuan Lu. Accurate RGB-D salient object detection via collaborative learning. In ECCV, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 241, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 284 + ], + "type": "text", + "content": "[21] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In IEEE CVPR, 2022, 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 285, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 545, + 319 + ], + "type": "text", + "content": "[22] Wen-Da Jin, Jun Xu, Qi Han, Yi Zhang, and Ming-Ming Cheng. Cdnet: Complementary depth network for rgb-d salient object detection. IEEE TIP, 30:3376-3390, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 319, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 352 + ], + "type": "text", + "content": "[23] Ran Ju, Ling Ge, Wenjing Geng, Tongwei Ren, and Gangshan Wu. Depth saliency based on anisotropic centersurround difference. In IEEE ICIP, 2014. 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 354, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 386 + ], + "type": "text", + "content": "[24] Hongwen Kang, Alexei A Efros, Martial Hebert, and Takeo Kanade. Image composition for object pop-out. In IEEE ICCVW, 2009. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "text", + "content": "[25] Youngeun Kim, Donghyeon Cho, Kyeongtak Han, Priyadarshini Panda, and Sungeun Hong. Domain adaptation without source data. IEEE TAI, 2(6):508-518, 2021. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 432, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 545, + 454 + ], + "type": "text", + "content": "[26] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In IEEE CVPR, 2021. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 455, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 510 + ], + "type": "text", + "content": "[27] Jogendra Nath Kundu, Akshay R Kulkarni, Suvaansh Bhambri, Deepesh Mehta, Shreyas Anand Kulkarni, Varun Jampani, and Venkatesh Babu Radhakrishnan. Balancing discriminability and transferability for source-free domain adaptation. In ICML, 2022. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 511, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 544 + ], + "type": "text", + "content": "[28] Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In IEEE CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 544, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 578 + ], + "type": "text", + "content": "[29] Vinod K Kurmi, Venkatesh K Subramanian, and Vinay P Namboodiri. Domain impression: A source data free domain adaptation method. In WACV, 2021. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 578, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 621 + ], + "type": "text", + "content": "[30] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, MinhTriet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. CVU, 184:45-56, 2019. 5, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 623, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 623, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 623, + 545, + 655 + ], + "type": "text", + "content": "[31] Minhyeok Lee, Chaewon Park, Suhwan Cho, and Sangyoun Lee. Spsn: Superpixel prototype sampling network for rgb-d salient object detection. In ECCV, 2022. 6, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 657, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 657, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 657, + 545, + 689 + ], + "type": "text", + "content": "[32] Aixuan Li, Jing Zhang, Yunqiu Lv, Bowen Liu, Tong Zhang, and Yuchao Dai. Uncertainty-aware joint salient object and camouflaged object detection. In IEEE CVPR, 2021. 3, 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[33] Gongyang Li, Zhi Liu, Minyu Chen, Zhen Bai, Weisi Lin, and Haibin Ling. Hierarchical alternate interaction network" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 733, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 733, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 296, + 733, + 315, + 742 + ], + "type": "text", + "content": "1040" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "forrgb-d salient object detection.IEEE TIP,30:3528-3542, 2021.3,6,8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "type": "text", + "content": "[34] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In IEEE CVPR, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 185 + ], + "type": "text", + "content": "[35] Rui Li, Danna Xue, Shaolin Su, Xiantuo He, Qing Mao, Yu Zhu, Jinqiu Sun, and Yanning Zhang. Learning depth via leveraging semantics: Self-supervised monocular depth estimation with both implicit and explicit semantic guidance. Pattern Recognition (PR), page 109297, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 217 + ], + "type": "text", + "content": "[36] Xinhao Li, Jingjing Li, Lei Zhu, Guoqing Wang, and Zi Huang. Imbalanced source-free domain adaptation. In ACM MM, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "type": "text", + "content": "[37] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 254, + 286, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 286, + 285 + ], + "type": "text", + "content": "[38] Jiang-Jiang Liu, Qibin Hou, Ming-Ming Cheng, Jiashi Feng, and Jianmin Jiang. A simple pooling-based design for real-time salient object detection. In IEEE CVPR, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 287, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 286, + 319 + ], + "type": "text", + "content": "[39] Yuang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. In IEEE CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 321, + 286, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 286, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 286, + 353 + ], + "type": "text", + "content": "[40] Yukang Lu, Dingyao Min, Keren Fu, and Qijun Zhao. Depth-cooperated trimodal network for video salient object detection. In ICIP. IEEE, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 355, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 387 + ], + "type": "text", + "content": "[41] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM TOG, 39(4):1-13, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 388, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 431 + ], + "type": "text", + "content": "[42] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In IEEE CVPR, 2021. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 434, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 477 + ], + "type": "text", + "content": "[43] Daniel McKee, Zitong Zhan, Bing Shuai, Davide Modolo, Joseph Tighe, and Svetlana Lazebnik. Transfer of representations to video label propagation: Implementation factors matter. arXiv preprint arXiv:2203.05553, 2022. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 479, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 510 + ], + "type": "text", + "content": "[44] Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan. Camouflaged object segmentation with distraction mining. In IEEE CVPR, 2021. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "text", + "content": "[45] S Mahdi H Miangoleh, Sebastian Dille, Long Mai, Sylvain Paris, and Yagiz Aksoy. Boosting monocular depth estimation models to high-resolution via content-adaptive multi-resolution merging. In IEEE CVPR, 2021. 1, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 557, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 588 + ], + "type": "text", + "content": "[46] Yuzhen Niu, Yujie Geng, Xueqing Li, and Feng Liu. Leveraging stereopsis for saliency analysis. In IEEE CVPR, 2012. 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "text", + "content": "[47] Youwei Pang, Xiaqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In IEEE CVPR, 2022. 3, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 636, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 668 + ], + "type": "text", + "content": "[48] Houwen Peng, Bing Li, Weihua Xiong, Weiming Hu, and Rongrong Ji. RGBD salient object detection: a benchmark and algorithms. In ECCV, 2014. 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[49] Yongri Piao, Zhengkun Rong, Miao Zhang, Weisong Ren, and Huchuan Lu. A2dele: Adaptive and attentive depth distiller for efficient RGB-D salient object detection. In IEEE CVPR, 2020. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[50] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In IEEE ICCV, 2021. 1, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[51] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI, 44(3):1623-1637, 2022. 1, 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 152, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 183 + ], + "type": "text", + "content": "[52] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "text", + "content": "[53] Suman Saha, Anton Obukhov, Danda Pani Paudel, Menelaos Kanakis, Yuhua Chen, Stamatios Georgoulis, and Luc Van Gool. Learning to relate depth and semantics for unsupervised domain adaptation. In IEEE CVPR, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "text", + "content": "[54] Roshni Sahoo, Divya Shanmugam, and John Guttag. Unsupervised domain adaptation in the absence of source data. arXiv preprint arXiv:2007.10233, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 264, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 285 + ], + "type": "text", + "content": "[55] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE CVPR, 2016. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 287, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 319 + ], + "type": "text", + "content": "[56] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from RGBD images. In ECCV, 2012. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 321, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 364 + ], + "type": "text", + "content": "[57] Przemysław Skurowski, Hassan Abdulameer, J Blaszczyk, Tomasz Depta, Adam Kornacki, and P Koziel. Animal camouflage analysis: Chameleon database. Unpublished manuscript, 2(6):7, 2018. 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "type": "text", + "content": "[58] Hwanjun Song, Eunyoung Kim, Varun Jampan, Deqing Sun, Jae-Gil Lee, and Ming-Hsuan Yang. Exploiting scene depth for object detection with multimodal transformers. In BMVC, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 411, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 442 + ], + "type": "text", + "content": "[59] Yujia Sun, Geng Chen, Tao Zhou, Yi Zhang, and Nian Liu. Context-aware cross-level fusion network for camouflaged object detection. In *IJCAI*, 2021. 3, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 444, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 475 + ], + "type": "text", + "content": "[60] Yujia Sun, Shuo Wang, Chenglizhao Chen, and Tian-Zhu Xiang. Boundary-guided camouflaged object detection. In IJ-CAI, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 478, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 520 + ], + "type": "text", + "content": "[61] Alessio Tonioni, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Unsupervised domain adaptation for depth prediction from images. IEEE TPAMI, 42(10):2396-2409, 2019. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 522, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 544 + ], + "type": "text", + "content": "[62] Anne Treisman. Preattentive processing in vision. ICVGIP, 31(2):156-177, 1985. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "text", + "content": "[63] Fengyun Wang, Jinshan Pan, Shoukun Xu, and Jinhui Tang. Learning discriminative cross-modality features for RGB-D saliency detection. IEEE TIP, 31:1285-1297, 2022. 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 579, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 600 + ], + "type": "text", + "content": "[64] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In IEEE CVPR, 2022. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 602, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 634 + ], + "type": "text", + "content": "[65] Cho-Ying Wu, Jialiang Wang, Michael Hall, Ulrich Neumann, and Shuochen Su. Toward practical monocular indoor depth estimation. In IEEE CVPR, 2022. 1" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 636, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 677 + ], + "type": "text", + "content": "[66] Yu-Huan Wu, Yun Liu, Jun Xu, Jia-Wang Bian, Yu-Chao Gu, and Ming-Ming Cheng. Mobilesal: Extremely efficient rgb-d salient object detection. IEEE TPAMI, 44(12):10261-10269, 2022. 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[67] Zongwei Wu, Guillaume Allibert, Christophe Stolz, and Cédric Demonceaux. Depth-adapted CNN for RGB-D cameras. In ACCV, 2020. 1" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 733, + 314, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 733, + 314, + 742 + ], + "spans": [ + { + "bbox": [ + 296, + 733, + 314, + 742 + ], + "type": "text", + "content": "1041" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[68] Zongwei Wu, Guillaume Allibert, Christophe Stolz, Chao Ma, and Cédric Demonceaux. Modality-guided subnetwork for salient object detection. In 3DV, 2021. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "text", + "content": "[69] Zongwei Wu, Shriarulmozhivarman Gobichettipalayam, Brahim Tamadazte, Guillaume Allibert, Danda Pani Paudel, and Cedric Demonceaux. Robust RGB-D fusion for saliency detection. 3DV, 2022. 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "text", + "content": "[70] Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In IEEE CVPR, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 183, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 183, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 183, + 287, + 216 + ], + "type": "text", + "content": "[71] Haifeng Xia, Handong Zhao, and Zhengming Ding. Adaptive adversarial network for source-free domain adaptation. In IEEE ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "type": "text", + "content": "[72] Ke Xian, Jianming Zhang, Oliver Wang, Long Mai, Zhe Lin, and Zhiguo Cao. Structure-guided ranking loss for single image depth prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 271, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 287, + 315 + ], + "type": "text", + "content": "[73] Mochu Xiang, Jing Zhang, Yunqiu Lv, Aixuan Li, Yiran Zhong, and Yuchao Dai. Exploring depth contribution for camouflaged object detection. arXiv e-prints, pages arXiv-2106, 2021. 1, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 316, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 358 + ], + "type": "text", + "content": "[74] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping Fan. Uncertainty-guided transformer reasoning for camouflaged object detection. In IEEE ICCV, 2021. 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "type": "text", + "content": "[75] Shiqi Yang, Joost van de Weijer, Luis Herranz, Shangling Jui, et al. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. NeurIPS, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 393, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 426 + ], + "type": "text", + "content": "[76] Shiqi Yang, Yaxing Wang, Joost van de Weijer, and Luis Herranz. Unsupervised domain adaptation without source data by casting a bait. arXiv preprint arXiv:2010.12427, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "type": "text", + "content": "[77] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In IEEE ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "type": "text", + "content": "[78] Qiang Zhai, Xin Li, Fan Yang, Chenglizhao Chen, Hong Cheng, and Deng-Ping Fan. Mutual graph learning for camouflaged object detection. In IEEE CVPR, 2021. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 492, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 525 + ], + "type": "text", + "content": "[79] Wei Zhai, Yang Cao, HaiYong Xie, and Zheng-Jun Zha. Deep texton-coherence network for camouflaged object detection. IEEE TMM, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "text", + "content": "[80] Chen Zhang, Runmin Cong, Qinwei Lin, Lin Ma, Feng Li, Yao Zhao, and Sam Kwong. Cross-modality discrepant interaction network for RGB-D salient object detection. In ACM MM, 2021. 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "type": "text", + "content": "[81] Jing Zhang, Deng-Ping Fan, Yuchao Dai, Xin Yu, Yiran Zhong, Nick Barnes, and Ling Shao. RGB-D saliency detection via cascaded mutual information minimization. In IEEE ICCV, 2021. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "type": "text", + "content": "[82] Jing Zhang, Yunqiu Lv, Mochu Xiang, Aixuan Li, Yuchao Dai, and Yiran Zhong. Depth-guided camouflaged object detection. CoRR, abs/2106.13217, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "text", + "content": "[83] Miao Zhang, Shuang Xu, Yongri Piao, Dongxiang Shi, Shusen Lin, and Huchuan Lu. Preynet: Preying on camouflaged objects. In ACM MM, 2022. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[84] Pingping Zhang, Dong Wang, Huchuan Lu, Hongyu Wang, and Baocai Yin. Learning uncertain convolutional features for accurate saliency detection. In IEEE ICCV, 2017. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 477 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "text", + "content": "[85] Wenbo Zhang, Ge-Peng Ji, Zhuo Wang, Keren Fu, and Qijun Zhao. Depth quality-inspired feature manipulation for efficient RGB-D salient object detection. In ACM MM, 2021. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 546, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 150 + ], + "type": "text", + "content": "[86] Zhoutong Zhang, Forrester Cole, Richard Tucker, William T Freeman, and Tali Dekel. Consistent depth of moving objects in video. ACM TOG, 40(4):1-12, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 151, + 546, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 546, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 546, + 194 + ], + "type": "text", + "content": "[87] Zhao Zhang, Zheng Lin, Jun Xu, Wen-Da Jin, Shao-Ping Lu, and Deng-Ping Fan. Bilateral attention network forrgb-d salient object detection. IEEE TIP, 30:1949-1961, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 196, + 546, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 546, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 546, + 228 + ], + "type": "text", + "content": "[88] Jiawei Zhao, Yifan Zhao, Jia Li, and Xiaowu Chen. Is depth really necessary for salient object detection? In ACM MM, 2020. 3, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 230, + 546, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 546, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 546, + 263 + ], + "type": "text", + "content": "[89] Shanshan Zhao, Huan Fu, Mingming Gong, and Dacheng Tao. Geometry-aware symmetric domain adaptation for monocular depth estimation. In IEEE CVPR, 2019. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 264, + 546, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 546, + 297 + ], + "type": "text", + "content": "[90] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In IEEE CVPR, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 298, + 546, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 546, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 546, + 331 + ], + "type": "text", + "content": "[91] Tao Zhou, Deng-Ping Fan, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Rgb-d salient object detection: A survey. CVMJ, pages 1-33, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 332, + 546, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 546, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 546, + 364 + ], + "type": "text", + "content": "[92] Tao Zhou, Huazhu Fu, Geng Chen, Yi Zhou, Deng-Ping Fan, and Ling Shao. Specificity-preserving RGB-D saliency detection. In IEEE ICCV, 2021. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 365, + 546, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 546, + 398 + ], + "type": "text", + "content": "[93] Zikun Zhou, Wenjie Pei, Xin Li, Hongpeng Wang, Feng Zheng, and Zhenyu He. Saliency-associated object tracking. In IEEE ICCV, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 399, + 546, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 546, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 546, + 443 + ], + "type": "text", + "content": "[94] Hongwei Zhu, Peng Li, Haoran Xie, Xuefeng Yan, Dong Liang, Dapeng Chen, Mingqiang Wei, and Jing Qin. I can find you! boundary-guided separated attention network for camouflaged object detection. In AAAI, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 444, + 546, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 546, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 546, + 477 + ], + "type": "text", + "content": "[95] Jinchao Zhu, Xiaoyu Zhang, Shuo Zhang, and Junnan Liu. Inferring camouflaged objects by texture-aware interactive guidance network. In AAAI, 2021. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 296, + 732, + 315, + 742 + ], + "type": "text", + "content": "1042" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Source-free Domain Adaptive Human Pose Estimation/6db7e049-3b2d-4942-adce-fb8455dd18b2_content_list.json b/2023/Source-free Domain Adaptive Human Pose Estimation/6db7e049-3b2d-4942-adce-fb8455dd18b2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bc1677efc5454ab57ede17e7a40f7004e8a18d4e --- /dev/null +++ b/2023/Source-free Domain Adaptive Human Pose Estimation/6db7e049-3b2d-4942-adce-fb8455dd18b2_content_list.json @@ -0,0 +1,1690 @@ +[ + { + "type": "text", + "text": "Source-free Domain Adaptive Human Pose Estimation", + "text_level": 1, + "bbox": [ + 209, + 130, + 759, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qucheng Peng, Ce Zheng, Chen Chen \nCenter for Research in Computer Vision, University of Central Florida", + "bbox": [ + 205, + 181, + 763, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{qucheng.peng,ce.zheng}@ucf.edu，chen.chen@crcv.ucf.edu", + "bbox": [ + 241, + 218, + 727, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human Pose Estimation (HPE) is widely used in various fields, including motion analysis, healthcare, and virtual reality. However, the great expenses of labeled real-world datasets present a significant challenge for HPE. To overcome this, one approach is to train HPE models on synthetic datasets and then perform domain adaptation (DA) on real-world data. Unfortunately, existing DA methods for HPE neglect data privacy and security by using both source and target data in the adaptation process.", + "bbox": [ + 73, + 300, + 470, + 436 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To this end, we propose a new task, named source-free domain adaptive HPE, which aims to address the challenges of cross-domain learning of HPE without access to source data during the adaptation process. We further propose a novel framework that consists of three models: source model, intermediate model, and target model, which explores the task from both source-protect and target-relevant perspectives. The source-protect module preserves source information more effectively while resisting noise, and the target-relevant module reduces the sparsity of spatial representations by building a novel spatial probability space, and pose-specific contrastive learning and information maximization are proposed on the basis of this space. Comprehensive experiments on several domain adaptive HPE benchmarks show that the proposed method outperforms existing approaches by a considerable margin. The codes are available at https://github.com/davidpengucf/SFDAHPE.", + "bbox": [ + 73, + 436, + 470, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 738, + 209, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The accuracy of 2D human pose estimation (HPE) from monocular images has been significantly improved through the use of deep neural networks. However, obtaining sufficient labeled data for training such models is a laborious and time-consuming task, particularly for real-world datasets that require diverse appearances and poses with ground truth annotations. In contrast, the development of computer graphics and game engines has facilitated the generation of annotated synthetic datasets that can generate a", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "virtually unlimited number of labeled samples. Nevertheless, models trained on synthetic data may not perform well on real-world data due to the domain gap or differences in data distribution. To address this challenge, domain adaptation (DA) is essential to enable synthetic data-trained models to generalize to real-world data.", + "bbox": [ + 496, + 268, + 893, + 361 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, source-free domain adaptation has gained significant attention due to its practicality in real-world scenarios, where using labeled source data during the adaptation is expensive or not feasible due to data privacy and security concerns. However, existing source-free domain adaptation methods in classification tasks may not be directly applicable to more complex tasks, such as 2D human pose estimation (HPE). This is because HPE requires estimating the spatial locations of different body joints, instead of just the object's category as classification does, making it more challenging to adapt from source to target domains. Therefore, there is a pressing need to develop effective source-free domain adaptation techniques specifically tailored to HPE tasks. This motivates us to propose a new task named source-free domain adaptive HPE in this paper.", + "bbox": [ + 496, + 366, + 893, + 594 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The task of source-free domain adaptive HPE presents challenges from both the source-free DA and pose estimation perspectives. Specifically, in traditional DA tasks, source data can participate in the adaptation process, allowing for a focus on reducing domain shifts between source and target. However, in source-free DA, only the pretrained source model can be used during adaptation, leading to catastrophic forgetting of the source domain. Directly applying the source pretrained model can also introduce noise due to domain shifts. Therefore, balancing the absorption of the source model's knowledge with the resistance of the source model's noise is critical. Current methods tend to adopt a progressive replacement strategy to alleviate the runoff of source information. For example, [21, 22] use the previous model to supervise ongoing adaptation, thereby evoking the source-related memory, while [32, 6] use the exponential moving average (EMA) strategy to slow down the forgetting of the source. Although these methods alleviate the negative impact of noise successfully, they modify the overall parameters of the pretrained source model", + "bbox": [ + 496, + 598, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4826", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "significantly, leading to the disappearance of source-related knowledge and representations during the adaptation process. Therefore, it is necessary to develop source-protect modules that can protect source information more effectively while resisting noise at the same time.", + "bbox": [ + 78, + 90, + 467, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, we outline another challenge of source-free domain adaptive HPE: the sparsity of keypoints (i.e. key body joints) in an image. When it comes to domain adaptation (DA) in classification, achieving distribution alignment to reduce domain gap is a straightforward process as the outputs can be represented in a probability space that includes all the categories. However, the same cannot be said for Human Pose Estimation (HPE) based on a complete image, as each keypoint has a specific distance from others based on bone lengths, and the number of keypoints $K$ (e.g., 14 for human pose and 21 for hand pose) is significantly smaller than the number of pixels in an image (e.g., $512 \\times 512 \\approx 0.26\\mathrm{M}$ ), making it a challenging task. Although heatmaps [34, 33] are commonly used to predict keypoints (heatmap representation-based 2D HPE is considered as the mainstream approach for HPE) while reducing dimensionality in HPE, the number of pixels inside the heatmap is still considerable (e.g. $128 \\times 128 \\approx 16\\mathrm{K}$ ) in comparison to $K$ . Therefore, building a spatial probability space for the keypoints based on either the entire image or the heatmap is not ideal. [12] proposes constructing a low-dimensional distribution based solely on the $K$ keypoints, effectively transforming the HPE task into a $K$ -classification task. However, this spatial distribution relies on clear decision boundaries between distinctive keypoints, which is not reasonable in HPE. In other words, all samples can be classified to a certain category in classification, but most pixels inside an image (e.g. more than $97\\%$ if the fault tolerance is decided by the metric PCK@0.05) can not be assigned to a certain keypoint in HPE, but mispredictions may occur at these unpresented locations. Hence, it is necessary to develop a novel spatial probability space that reduces sparsity while providing complete representations of the prediction space.", + "bbox": [ + 80, + 171, + 467, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned task-specific challenges, we propose a novel framework that incorporates source-free domain adaptive HPE from both source-protect and target-relevant perspectives. Our framework consists of three models: the source model, the intermediate model, and the target model. Source model preserves source information, and target model absorbs knowledge from target data, while intermediate model interacts with both of them to reduce the domain gap between source and target implicitly, thereby used for the final inference. In the source-protect modules, we conduct knowledge transfer between the source model and the intermediate model to prevent catastrophic forgetting of the source-related information. We transfer knowledge from the source model to the intermediate model while keeping the source model's feature extractor fixed. To re", + "bbox": [ + 80, + 674, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sist noise, we propose a new residual loss and finetune the source model's regressor. In the target-relevant modules, we conduct executions between the intermediate model and the target model to mitigate domain shifts between the source and target domains. We build a new spatial probability space for HPE to alleviate the sparsity problem. On the basis of this space, we propose contrastive learning and information maximization designed specifically for human pose estimation. Our contributions are summarized as follows:", + "bbox": [ + 503, + 90, + 888, + 224 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We construct a new task named source-free domain adaptive human pose estimation, which focuses on the cross-domain learning of HPE without access to source data during the adaptation process.", + "- We propose a new framework consisting of three models: source model, intermediate model, and target model, which explores the task from both source-protect and target-relevant perspectives.", + "- We conduct comprehensive experiments on several domain adaptive HPE benchmarks and the results show that our method outperforms the existing approaches by a considerable margin." + ], + "bbox": [ + 501, + 227, + 888, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 503, + 422, + 638, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human Pose Estimation. The heatmap-based method [29, 37] is currently a widely used technique in 2D HPE, which involves generating heatmaps by placing 2D Gaussian kernels on potential keypoints, followed by converting these heatmaps to coordinates. For instance, [3] uses global- and refine-nets, while [39] simplifies the network structures through the Simple Baselines technique. Some approaches have focused on utilizing features at different scales [30], while others aim to improve the quality of heatmaps by extending Gaussians to second order [40] or using coupled embeddings for heatmap regression [37]. In this study, we adopt the same heatmap-based HPE methods as other domain adaptive HPE works [12, 16].", + "bbox": [ + 503, + 446, + 888, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "General Domain Adaptation. There are two main categories of methods for achieving general domain adaptation, which requires access to both the source and target data during adaptation. The first category is metric-based methods that explicitly measure the discrepancy between the source and target domains. For example, [24] uses maximum mean discrepancy, and [35] applies deep domain confusion. Recent work, such as [4], jointly performs clustering and discrimination for alignment using contrastive learning [2, 13]. The second category is GAN-based [7] methods that build a min-max game for two players related to the source and target domains. For example, [5, 28] adopts domain confusion to confuse the two players, while [31, 27] shows that discriminative clustering on the target benefits adaptation. However, access to source data is required in the general domain adaptation setting, which raises concerns about data privacy and security.", + "bbox": [ + 503, + 645, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4827", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Source-Free Domain Adaptation. Source-Free has become the mainstream paradigm for alleviating concerns about data privacy and security in domain adaptation, as shown in the work of [20, 21, 26]. There are two technique routes under the source-free setting: self-supervision and virtual source transfer. For the self-supervised methods, [21] is the most representative, which introduces information maximization to assist adaptation. Other self-supervised methods include [38], which treats the problem from a discriminative perspective, and [1, 26], which proposes online pseudo label refinement. In virtual source methods, most build GANs to generate virtual source data. Examples include conditional GANs used in [18], and collaborative GANs in [20], which achieve better generations. However, most source-free DA methods for classification are not applicable to HPE tasks due to the sparsity of keypoints' distributions.", + "bbox": [ + 75, + 90, + 472, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Domain Adaptive Human Pose Estimation. There are two main categories of methods used for domain-adaptive HPE. The first category involves a shared structure, where the weights of the networks from both the source pretrain and target adaptation are shared. CC-SSL [25] employs a single end-to-end trained network, RegDA [12] utilizes one shared feature extractor and two separate regressors, and TransPar [8] emphasizes transferable parameters using a similar structure. The second category is the unshared structure, where the pretrain and adaptation create a teacher-student paradigm. In this category, MDAM [19] addresses this structure alongside a novel pseudo-label strategy, UniFrame [16] modifies the classic Mean-Teacher [32] model by combining it with style transfer [10], and MarsDA [14] applies a teacher-student approach to edit RegDA. However, all these existing domain adaptive HPE approaches are not source-free, which raises concerns about data privacy and security. Therefore, it is necessary to consider source-free domain adaptive HPE. To our knowledge, we are the first work to study this challenging problem.", + "bbox": [ + 75, + 349, + 472, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 683, + 169, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary of 2D HPE", + "text_level": 1, + "bbox": [ + 76, + 709, + 294, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In 2D HPE, we are given a labeled pose dataset $\\mathcal{D} = (x_i,y_i)$ , where $x_{i}\\in \\mathbb{R}^{C\\times H\\times W}$ is an image and $y_{i}\\in \\mathbb{R}^{K\\times 2}$ are the corresponding keypoint coordinates. Here, $H$ and $W$ represent the height and width of the image, respectively, and $C$ is the number of channels. Additionally, $K$ represents the number of keypoints. To ensure smoother training, most existing 2D methods represent the keypoint coordinates as heatmaps $HM_{i}\\in \\mathbb{R}^{K\\times H^{\\prime}\\times W^{\\prime}}$ , where $H^{\\prime}$ and $W^{\\prime}$ are the height and width of the heatmaps, respectively. These heatmaps are used to represent the coordinates in a spatial way. The process of how the model works is defined", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/99393e82b2670ff4be632a31185e86033b4e3ced3a013c36bcad381f13d55625.jpg", + "image_caption": [ + "Figure 1. Process of heatmap-based 2D HPE. The model includes two components: the feature extractor $G$ and the regressor $F$ . After feeding the input image to the model, we obtain $K$ output heatmaps corresponding to $K$ keypoints." + ], + "image_footnote": [], + "bbox": [ + 501, + 85, + 893, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in Fig. 1. Ground truth heatmaps are generated by a 2D Gaussian centered at the ground truth joint location, with one heatmap for each keypoint. We denote the transform from coordinates to heatmaps as $T: \\mathbb{R}^{K \\times 2} \\to \\mathbb{R}^{K \\times H' \\times W'}$ . The training processes are conducted between the output heatmaps and ground truth heatmaps using MSE for supervised learning. In the inference stage, well-trained supervised models output heatmaps that predict the probability of joints occurring at each pixel. For each heatmap, the pixel with the highest probability is considered to be the predicted location of the joint.", + "bbox": [ + 496, + 282, + 893, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Problem Statement", + "text_level": 1, + "bbox": [ + 500, + 458, + 684, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the source-free domain adaptive HPE setting, we have a source domain dataset $\\mathcal{S} = \\{(x_i^s, y_i^s)\\}_{i=1}^{n_s}$ with $n_s$ labeled pose samples. $x_i^s \\in \\mathbb{R}^{C \\times H \\times W}$ is the source image and $y_i^s \\in \\mathbb{R}^{K \\times 2}$ is the corresponding pose annotation. Besides, there exists a target domain dataset $\\mathcal{T} = \\{x_i^t\\}_{i=1}^{n_t}$ that includes $n_t$ unlabeled pose samples. The source and target domains share the same label space but lie in different distributions (e.g. synthetic vs. real). The training procedure is split into two stages: the pretrain stage and the adaptation stage. In the pretrain stage, the labeled source domain dataset $\\mathcal{S}$ is used to train the source model $f^{sr}$ . However, what distinguishes source-free domain adaptive HPE from general domain adaptive HPE is the adaptation stage, where only the source model $f^{sr}$ and target domain dataset $\\mathcal{T}$ are available for training, and the source domain dataset cannot be used anymore. Our objective is to develop a model that performs well on the target dataset.", + "bbox": [ + 496, + 482, + 893, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Pipeline of Our Method", + "text_level": 1, + "bbox": [ + 500, + 747, + 718, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3.1 Source Model Pretrain", + "text_level": 1, + "bbox": [ + 500, + 771, + 714, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For DA in HPE, labeled source data are applied for pretraining. Same as most existing domain adaptive pose estimation methods [25, 19, 12, 16], the network is composed of a feature extractor $G$ (such as ResNet backbone) and a pose regressor $F$ , so the source model can be represented as $f^{sr} = F^{sr}(G^{sr}(\\cdot))$ . Based on the supervised heatmap-based loss (MSE loss), we propose the overall objective for", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4828", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c960532eded9a22c3b979ce012a59d9294c563557223b317552fadd1d1649de8.jpg", + "image_caption": [ + "Figure 2. Overall framework of our method. It includes three models, each with one feature extractor and one regressor. The adaptation contains two steps: Step A and Step B. Step A is applied for source-protect adaptation, while Step B aims at target-relevant adaptation. The solid arrows show the forward process of associated losses, and the dotted arrows show the backward updating process of associated losses. EMA means the exponential moving average process from target model to intermediate model to update intermediate model's weights." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 467, + 359 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/417679ac09e08dce32c594ba5d8369484112f2f1450c11c82be088228c035e47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 87, + 893, + 361 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "source pretraining:", + "bbox": [ + 76, + 443, + 205, + 458 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p r e t r a i n}} = \\mathbb {E} _ {\\left(x _ {i} ^ {s}, y _ {i} ^ {s}\\right) \\in \\mathcal {S}} \\mathcal {L} _ {\\text {m s e}} \\left(F ^ {s r} \\left(G ^ {s r} \\left(x _ {i} ^ {s}\\right)\\right), T \\left(y _ {i} ^ {s}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 465, + 468, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{mse}$ is the Mean Squared Error (MSE) loss between the heatmap representations of prediction and ground truth, and $T$ is the transform from coordinates to heatmaps defined in Sec. 3.1.", + "bbox": [ + 75, + 489, + 470, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 Adaptation Framework", + "text_level": 1, + "bbox": [ + 76, + 569, + 295, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As Fig. 2 shows, our method contains three models: source model $f^{sr} = F^{sr}(G^{sr}(\\cdot))$ , intermediate model $f^{in} = F^{in}(G^{in}(\\cdot))$ , and target model $f^{tg} = F^{tg}(G^{tg}(\\cdot))$ . Each model includes one feature extractor and one regressor. Assume the heatmaps generated by source model as $H_{i}^{sr} = F^{sr}(G^{sr}(x_{i}^{t})) = [h_{i,1}^{sr}, h_{i,2}^{sr}, \\dots, h_{i,K}^{sr}]^{\\top}$ , and those from intermediate model and target model respectively as $H_{i}^{in} = F^{in}(G^{in}(x_{i}^{t})) = [h_{i,1}^{in}, h_{i,2}^{in}, \\dots, h_{i,K}^{in}]^{\\top}$ and $H_{i}^{tg} = F^{tg}(G^{tg}(x_{i}^{t})) = [h_{i,1}^{tg}, h_{i,2}^{tg}, \\dots, h_{i,K}^{tg}]^{\\top}$ , where $K$ is the number of keypoints. We also assume the keypoint indexing space as $\\mathcal{P} = [1,2,\\dots,K]$ . The source model is used to maintain source information, while the target model is applied to learn target knowledge. The intermediate model is the transition model between these two models. It serves as a bridge between source and target as it learns from source and target simultaneously, thus eliminating the domain gap implicitly. Therefore, it outperforms both the source and target models and is used for the final inference. When starting the adaptation, $f^{in}$ and $f^{tg}$ are initialized from the source model $f^{sr}$ .", + "bbox": [ + 75, + 592, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a single adaptation iteration, two steps are executed as shown in Fig. 2. Step A is for the source-protect adaptation, and Step B aims to adapt in a target-relevant way.", + "bbox": [ + 498, + 441, + 890, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Step A, our goal is to overcome catastrophic forgetting on source domain, while eliminating the noise from the source model due to domain shift. Therefore, the source model's feature extractor $G^{sr}$ is fixed to retain source knowledge, while its regressor $F^{sr}$ keeps updating to improve regression accuracy, hence reducing noises in the outputs. On the opposite, the intermediate model's regressor $F^{in}$ is fixed, but its feature extractor $G^{in}$ continues to update. That's because we want the feature extractor to learn more source knowledge on the representation level, while the noise from the source model due to domain shifts can be inhibited.", + "bbox": [ + 496, + 508, + 890, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Step B, we aim to conduct the adaptation between the intermediate model and the target model in a target-relevant manner. During the process, several pose-specific techniques are utilized to minimize the discrepancy between these two models. Though both $f^{in}$ and $f^{tg}$ are updated at this step, $f^{tg}$ is updated via back-propagation while $f^{in}$ is via exponential moving average from $f^{tg}$ , since averaging model weights over training steps tends to produce a more accurate model than using the final weights directly [32].", + "bbox": [ + 496, + 712, + 890, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following two subsections, we introduce the detailed techniques used for Step A and Step B.", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4829", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7494d1f6f42d043899528e8a9b02de3dfb7e3ece876fe587967bc2611903c160.jpg", + "image_caption": [ + "Figure 3. Proposed residual loss. Here we choose two heatmaps at the same location generated from source model and intermediate model for illustration. For simplicity, here the heatmap's scale is set as $3 \\times 3$ , while the actual size is $64 \\times 64$ ." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 472, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3 Source-protect Modules", + "text_level": 1, + "bbox": [ + 76, + 335, + 297, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this part, we propose modules related to source-free-specific adaptation. The biggest challenge is to distill knowledge from the source model and resist the potential noise due to domain shift. First, we introduce the finetune loss that calibrates on the source model:", + "bbox": [ + 75, + 359, + 468, + 434 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f t} = \\mathbb {E} _ {x _ {i} ^ {t} \\in \\mathcal {T}} \\mathcal {L} _ {m s e} \\left(F ^ {s r} \\left(G ^ {s r} \\left(x _ {i} ^ {t}\\right)\\right), F ^ {i n} \\left(G ^ {i n} \\left(x _ {i} ^ {t}\\right)\\right)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 446, + 468, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To be notified, we hope the regression results of the source model can be more accurate, while the source knowledge can be retained. Thus, the feature extractor is fixed so that the source representations continue to be generated, and the regressor is being updated:", + "bbox": [ + 75, + 477, + 468, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {F ^ {s r}} ^ {t + 1} = \\theta_ {F ^ {s r}} ^ {t} - \\lambda^ {s r} \\frac {\\partial L _ {f t}}{\\partial \\theta_ {F ^ {s r}} ^ {t}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 563, + 468, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta_F^s$ is the weights of the source model's regressor $F^s$ , while $t$ denotes the training step and $\\lambda^s$ is the source model's learning rate. However, improving source model is not enough. What's more, we propose the residual loss to improve intermediate model.", + "bbox": [ + 75, + 607, + 468, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The application of Equation 2 facilitates mutual knowledge transfer between the source and intermediate models. Apart from the pixel with the highest confidence in a heatmap, the other pixels also contain valuable information that benefits the knowledge transfer, and that's why we propose residual loss in Fig. 3. From the example in the figure, the most confident pixel (confidence $= 1.0$ ) in the source heatmap is removed, along with the pixel at the same location which has a confidence of 0.8 in the intermediate heatmap marked with red. The pair marked with blue follows the same rule but on the basis of intermediate heatmap. Then we build adaptive residual heatmaps denoted as $\\hat{h}_{i,j}^{sr}$ and $\\hat{h}_{i,j}^{in}$ . Moreover, we employ KL divergence to ensure that $\\hat{h}_{i,j}^{in}$ approaches $\\hat{h}_{i,j}^{sr}$ , thereby preserving the source in", + "bbox": [ + 75, + 681, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/18126c75a76efac6603c7be223e6f0e5486647a6d6f1b8a459a89950007e3a71.jpg", + "image_caption": [ + "Figure 4. Illustration of the projection of a heatmap. In order to reduce the sparsity of a heatmap, we project it into a horizontal vector and a vertical vector. For simplicity, here the heatmap's scale is set as $4 \\times 4$ and the vector's scale as $4 \\times 1$ , while the actual heatmap's size is $64 \\times 64$ and the vector's size is $64 \\times 1$ ." + ], + "image_footnote": [], + "bbox": [ + 501, + 87, + 875, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "formation. Based on these, we deduce the residual loss as:", + "bbox": [ + 500, + 291, + 883, + 305 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e s} = \\mathbb {E} _ {x _ {i} ^ {t} \\in \\mathcal {T}} \\mathbb {E} _ {j \\in \\mathcal {P}} \\mathrm {D} _ {\\mathrm {K L}} \\left(\\sigma \\left(\\hat {h} _ {i, j} ^ {s r} / \\tau\\right) \\mid \\mid \\sigma \\left(\\hat {h} _ {i, j} ^ {i n} / \\tau\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 316, + 890, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{D}_{\\mathrm{KL}}$ is the KL divergence, and $\\sigma (\\cdot)$ is the softmax function. Besides, $\\tau$ is the temperature used to scale the residual heatmaps and is empirically set to 0.3. By combining the finetune loss and the residual loss, we get the objective function with the optimization process for updating the intermediate model:", + "bbox": [ + 498, + 347, + 890, + 436 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i n} = \\mathcal {L} _ {f t} + \\alpha \\mathcal {L} _ {r e s}, \\quad \\theta_ {G ^ {i n}} ^ {t + 1} = \\theta_ {G ^ {i n}} ^ {t} - \\lambda^ {i n} \\frac {\\partial L _ {i n}}{\\partial \\theta_ {G ^ {i n}} ^ {t}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 446, + 890, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $t$ is the training step, and $\\lambda^{in}$ is the learning rate for $G^{in}$ . Moreover, $\\alpha$ is the trade-off hyperparameter between these two losses.", + "bbox": [ + 500, + 493, + 890, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.4 Target-relevant Modules", + "text_level": 1, + "bbox": [ + 500, + 559, + 723, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section introduces a module that is associated with target information. As previously mentioned, the sparsity of keypoints in the image presents a significant challenge in our task. To address this issue, we propose using projected horizontal and vertical vectors of heatmaps as spatial representations, as shown in Fig. 4. This approach reduces the sparsity of the spatial probability space effectively since the number of overall digits inside the heatmap is $64 \\times 64 \\approx 4K$ , while the number of units in each vector is 64. Therefore, we can effectively reduce the sparsity of the spatial probability space. With 64 units in each vector, the spatial probability space size is acceptable compared to the number of keypoints.", + "bbox": [ + 496, + 583, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Despite the reduction in sparsity, the completeness of spatial representations is not affected. That is because all the coordinates can always be divided into two representations, horizontally and vertically. In such a case, our spatial representation is better than the $K$ -classification probability space proposed by [12], which only includes the keypoint-relevant information and does not take the units in the heatmap that are not keypoints into consideration.", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4830", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/92ecc105b5653a9cc0a04a1838ecbb0834757d4eb0bdd568f766d052fb9cae0c.jpg", + "image_caption": [ + "Figure 5. Illustration of building positive and negative keypoint pairs for contrastive learning between the output heatmaps of the intermediate model and target model." + ], + "image_footnote": [], + "bbox": [ + 78, + 85, + 472, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 4, we introduce the projection of a heatmap from both horizontal and vertical directions. For the heatmap $h_{i,j}^{in}$ generated by the intermediate model from $i$ th sample's $j$ th keypoint, its projection can be represented as $proj( h_{i,j}^{in}) = (v_{i,jx}^{in}, v_{i,jy}^{in})$ , where $v_{i,jx}^{in}$ is the horizontal vector and $v_{i,jy}^{in}$ is the vertical vector. This vector pair is the foundation of the following modules and losses that we propose.", + "bbox": [ + 76, + 290, + 468, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, we propose the pose-specific contrastive loss between the intermediate model and target model. According to the observation from [12], wrong predictions are usually located at other keypoints' positions. In such a case, improving regression can be achieved by minimizing the discrepancy of heatmaps at the same location and maximizing the discrepancy of heatmaps at different locations. And that's how we define the positive and negative pair in Fig. 5. Denote the similarity of two heatmaps generated by the intermediate model and target model respectively as:", + "bbox": [ + 75, + 397, + 470, + 537 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns i m \\left(h _ {i, j} ^ {i n}, h _ {i, k} ^ {t g}\\right) = \\frac {1}{2} \\left(\\frac {v _ {i , j x} ^ {i n} \\cdot v _ {i , k x} ^ {t g}}{| | v _ {i , j x} ^ {i n} | | | | | v _ {i , k x} ^ {t g} | |} + \\frac {v _ {i , j y} ^ {i n} \\cdot v _ {i , k y} ^ {t g}}{| | v _ {i , j y} ^ {i n} | | | | | v _ {i , k y} ^ {t g} | |}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 542, + 468, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On the basis of the similarity, we define the overall pose-specific contrastive loss as:", + "bbox": [ + 76, + 585, + 468, + 617 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c s t} = - \\mathbb {E} _ {x _ {i} ^ {t} \\in \\mathcal {T}} \\mathbb {E} _ {j \\in \\mathcal {P}} \\log \\frac {\\exp (s i m (h _ {i , j} ^ {i n} , h _ {i , j} ^ {t g}))}{\\Sigma_ {k \\in \\mathcal {P}} \\exp (s i m (h _ {i , j} ^ {i n} , h _ {i , k} ^ {t g}))}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 623, + 468, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we propose a pose-specific self-supervised technique called pose-specific information maximization. Information maximization is very helpful to source-free domain adaptation for classification [21] because it makes the target outputs individually certain and globally diverse, hence mitigating the domain gap implicitly. However, the improvements are limited due to the sparsity of heatmaps. With the help of projected vectors, we refine the information maximization loss as:", + "bbox": [ + 75, + 666, + 470, + 801 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i m} = \\mathbb {E} _ {x _ {i} ^ {t} \\in \\mathcal {T}} \\left(\\mathcal {L} _ {e n t x} \\left(x _ {i} ^ {t}\\right) + \\mathcal {L} _ {e n t y} \\left(x _ {i} ^ {t}\\right) - \\mathcal {L} _ {d i v} \\left(x _ {i} ^ {t}\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 88, + 811, + 468, + 830 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\mathcal {L} _ {e n t x} \\left(x _ {i} ^ {t}\\right) = - \\mathbb {E} _ {j \\in \\mathcal {P}} \\left[ \\sigma \\left(v _ {i, j x} ^ {t g}\\right) \\log \\sigma \\left(v _ {i, j x} ^ {t g}\\right) \\right], \\\\ \\mathcal {L} _ {e n t y} \\left(x _ {i} ^ {t}\\right) = - \\mathbb {E} _ {j \\in \\mathcal {P}} \\left[ \\sigma \\left(v _ {i, j y} ^ {t g}\\right) \\log \\sigma \\left(v _ {i, j y} ^ {t g}\\right) \\right], \\\\ \\mathcal {L} _ {d i v} \\left(x _ {i} ^ {t}\\right) = - \\left[ \\mathbb {E} _ {j \\in \\mathcal {P}} \\sigma \\left(h _ {i, j} ^ {t g}\\right) \\right] \\log \\left[ \\mathbb {E} _ {j \\in \\mathcal {P}} \\sigma \\left(h _ {i, j} ^ {t g}\\right) \\right]. \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 839, + 468, + 905 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here $\\mathcal{L}_{entx}$ is the entropy of horizontal vectors, and $\\mathcal{L}_{enty}$ is the entropy of vertical vectors. The last term in Equation 8 is the diversity-promoting objective. Moreover, we introduce the consistency loss between the two models' outputs:", + "bbox": [ + 498, + 90, + 890, + 148 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c o n} = \\mathbb {E} _ {\\left(x _ {i} ^ {s}, y _ {i} ^ {s}\\right) \\in \\mathcal {S}} \\mathcal {L} _ {m s e} \\left(F ^ {i n} \\left(G ^ {i n} \\left(x _ {i} ^ {s}\\right)\\right), F ^ {t g} \\left(G ^ {t g} \\left(x _ {i} ^ {s}\\right)\\right)\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 155, + 890, + 174 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By summarizing above, we conclude the overall objective loss function for the target model as:", + "bbox": [ + 498, + 181, + 890, + 212 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t g t} = \\mathcal {L} _ {c o n} + \\beta \\mathcal {L} _ {c s t} + \\gamma \\mathcal {L} _ {i m}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 223, + 890, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\beta$ and $\\gamma$ are trade-off parameters.", + "bbox": [ + 500, + 250, + 764, + 265 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, the intermediate model is updated via the exponential moving average strategy (EMA) in this step:", + "bbox": [ + 498, + 265, + 890, + 296 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {f ^ {i n}} ^ {t + 1} = \\eta \\theta_ {f ^ {i n}} ^ {t} + (1 - \\eta) \\theta_ {f ^ {t g}} ^ {t}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 304, + 890, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $t$ denotes the step of training and $\\eta$ denotes the smoothing coefficient set to 0.999 by default.", + "bbox": [ + 498, + 333, + 890, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 377, + 632, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We use three human pose datasets and three hand pose datasets to validate our approach. SURREAL [36] is utilized as the source dataset and encompasses six million synthetic human pose images. Human3.6M [11], one of the target human pose datasets, is a frequently used real-world dataset in the community, consisting of 3.6 million images distributed into seven folds. According to previous research [12, 16], S1, S5, S6, S7, and S8 are designated as the training set, whereas S9 and S11 serve as the testing set. The other target dataset is Leeds Sports Pose [15] (LSP), which is a real-world dataset with 2,000 images, and we use all of them for adaptation. We focus on two domain adaptation tasks: SURREAL $\\rightarrow$ Human3.6M and SURREAL $\\rightarrow$ LSP for human pose tasks.", + "bbox": [ + 496, + 402, + 890, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Rendered Hand Pose Dataset [42] (RHD) is the source hand dataset with 43,986 synthetic hand images, of which 41,258 are allocated for training, while the remaining 2,728 images are utilized for validation. Hand-3D-Studio [41] (H3D) is one of the two target hand datasets, including 22,000 real-world frames. For training, we employed 18,800 frames and reserved the rest for testing, following the protocol outlined in previous research papers [12, 16]. The other target hand dataset, FreiHand [43], provides 130k real-world images, and we employed all of them for our adaptation task. Our focus is on two domain adaptation tasks: RHD $\\rightarrow$ H3D and RHD $\\rightarrow$ FreiHand.", + "bbox": [ + 496, + 613, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We adopt Simple Baseline [39] with ResNet-101 [9] as the HPE backbone, following previous works [12, 16]. In the pretrain process, we conduct 40 epochs, each with 500 iterations, using the Adam [17] optimizer with an initial learning rate of 1e-4. The learning rate decreases to 1e-5 at 25 epochs. For the adaptation process, we execute 40 epochs with 30,000 iterations. In step", + "bbox": [ + 496, + 795, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4831", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A, we choose an initial learning rate of 1e-4 for $F^{sr}$ and 1e-5 for $G^{in}$ . Step B sets the initial learning rate to 1e-4 for $G^{tg}$ and 1e-3 for $F^{tg}$ . We use the same annealing strategy for the learning rate scheduler as in [12]. As for hyperparameters, we select $\\alpha = 0.7$ , $\\beta = 0.5$ , and $\\gamma = 0.85$ . Moreover, the intermediate model is used for the final inference.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Main Results", + "text_level": 1, + "bbox": [ + 76, + 194, + 215, + 209 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. Four general domain adaptive pose estimation methods CC-SSL [25], MDAM [19], RegDA [12] and UniFrame [16] with SOTA performances are chosen as baselines for comparison. We list these baselines that need to access source data to demonstrate that our proposed source-free method's performance remains competitive when compared to general domain adaptive pose estimation methods. We build source-free baselines by either adapting domain-adaptive HPE methods or source-free DA methods for classification methods. We first introduce two domain adaptive HPE methods. RegDA-SF is based on [12], which originally contains three steps and the first step uses source data. We replace the first step with an exponential moving averaging strategy to obtain the new method. Uniframe-SF is on the foundation of [16]. What we do is remove the style transfer modules between source and target so that it becomes a source-free method. The rest are adapted from source-free DA methods for classification methods. SHOT [21] is a classic source-free domain adaptation method for image classification, and we adapt its mutual information maximization module for HPE by treating the normalized heatmaps as output probabilities in classification. MMT is based on [6], which is a source-free adaptation method that applies two models mutually learning from each other. Here we replace its soft classification loss with the soft consistency loss. SHOT++ [22] is the advanced version of [21], which uses a special unsupervised representation learning technique predicting image rotations. We also add this module to our adopted baseline.", + "bbox": [ + 75, + 218, + 470, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. To evaluate the accuracy of the proposed approach, we adopt the Percentage of Correct Keypoint (PCK) metric, as used in previous work. We set the ratio of correct predictions to be $5\\%$ , and report PCK@0.05 in Tables 1-4. In addition to overall keypoint accuracy, we divide joints into several part segments and measure their performance using specific metrics. For the 21-keypoint hand skeleton, we use the Metacarpophalangeal (MCP), Proximal Interphalangeal (PIP), Distal Interphalangeal (DIP), and Finger-tip (Fin) metrics. For the 18-keypoint human skeleton, we select the Shoulder (Sld), Elbow (Elb), Wrist, Hip, Knee, and Ankle metrics. These segment-specific metrics enable a more detailed evaluation of the models' performances.", + "bbox": [ + 75, + 657, + 468, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. We present the results of our proposed method on four different tasks: hand tasks RHD $\\rightarrow$ H3D and RHD $\\rightarrow$ FreiHand, and human tasks", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/52dd9a0b9ac544c96776f48053da6308df90da21df08677829392b1a06d7019d.jpg", + "table_caption": [ + "Table 1. PCK@0.05 on RHD $\\rightarrow$ H3D Task" + ], + "table_footnote": [], + "table_body": "
MethodSFMCPPIPDIPFinAll
Source-only-67.464.263.354.861.8
CC-SSL [25] (CVPR'20)×81.579.974.464.075.1
MDAM [19] (CVPR'21)×82.379.672.361.574.1
RegDA [12] (CVPR'21)×79.674.471.262.972.5
UniFrame [16] (ECCV'22)×86.784.678.968.179.6
RegDA-SF [12] (CVPR'21)✓71.266.866.258.566.9
UniFrame-SF [16] (ECCV'22)✓84.884.277.068.077.9
SHOT [21] (ICML'20)✓77.569.870.664.273.7
MMT [6] (ICLR'20)✓78.774.172.565.375.3
SHOT++ [22] (TPAMI'22)✓85.385.178.267.678.9
Ours✓88.489.280.971.482.2
", + "bbox": [ + 506, + 99, + 883, + 224 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b3f8396f972f37fea017451c78bec96bfadfcb237c52d70ddd48daf0fe7cc52e.jpg", + "table_caption": [ + "Table 2. PCK@0.05 on RHD $\\rightarrow$ FreiHand Task" + ], + "table_footnote": [], + "table_body": "
MethodSFMCPPIPDIPFinAll
Source-only-35.250.154.850.746.8
CC-SSL [25] (CVPR'20)×37.448.250.146.543.8
MDAM [19] (CVPR'21)×32.348.151.747.345.1
RegDA [12] (CVPR'21)×40.955.058.253.151.1
UniFrame [16] (ECCV'22)×43.564.067.462.458.5
RegDA-SF [12] (CVPR'21)✓38.652.957.654.349.5
UniFrame-SF [16] (ECCV'22)✓40.662.561.060.255.7
SHOT [21] (ICML'20)✓40.461.360.558.153.3
MMT [6] (ICLR'20)✓39.660.460.057.852.6
SHOT++ [22] (TPAMI'22)✓41.062.862.759.355.8
Ours✓43.765.966.663.158.8
", + "bbox": [ + 506, + 250, + 883, + 375 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/90284d7370f296641aba911d6091b6a2d056fe8adaf10c16ac4bc652b9cdc4b4.jpg", + "table_caption": [ + "Table 3. PCK@0.05 on SURREAL $\\rightarrow$ Human3.6M Task" + ], + "table_footnote": [], + "table_body": "
MethodSFSldElbWristHipKneeAnkleAll
Source-only-69.475.466.437.977.377.767.3
CC-SSL [25] (CVPR'20)×44.368.555.222.262.357.851.7
MDAM [19] (CVPR'21)×51.783.168.917.779.476.662.9
RegDA [12] (CVPR'21)×73.386.472.854.882.084.475.6
UniFrame [16] (ECCV'22)×78.189.681.152.685.387.179.0
RegDA-SF [12] (CVPR'21)✓70.682.069.843.379.179.471.5
UniFrame-SF [16] (ECCV'22)✓74.384.872.545.981.585.275.1
SHOT [21] (ICML'20)✓73.083.271.946.779.882.473.3
MMT [6] (ICLR'20)✓73.283.572.445.180.883.973.9
SHOT++ [22] (TPAMI'22)✓75.085.276.745.382.885.075.3
Ours✓77.988.880.452.384.286.978.7
", + "bbox": [ + 506, + 401, + 883, + 502 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2d963df699c052f16b73f16f014c6dd1ffe71eb29fc2cd94f62c2479efdff703.jpg", + "table_caption": [ + "Table 4. PCK@0.05 on SURREAL $\\rightarrow$ LSP Task" + ], + "table_footnote": [], + "table_body": "
MethodSFSldElbWristHipKneeAnkleAll
Source-only-51.565.062.968.068.767.463.9
CC-SSL [25] (CVPR'20)×36.866.363.959.667.370.460.7
MDAM [19] (CVPR'21)×61.477.775.565.876.778.369.2
RegDA [12] (CVPR'21)×62.776.771.181.080.375.374.6
UniFrame [16] (ECCV'22)×69.284.983.385.584.784.382.0
RegDA-SF [12] (CVPR'21)✓54.870.567.665.473.270.066.5
UniFrame-SF [16] (ECCV'22)✓68.480.579.182.780.881.078.8
SHOT [21] (ICML'20)✓63.572.766.578.479.773.272.4
MMT [6] (ICLR'20)✓60.970.970.381.179.372.871.5
SHOT++ [22] (TPAMI'22)✓69.581.780.984.082.379.779.9
Ours✓70.785.483.886.685.285.083.2
", + "bbox": [ + 506, + 529, + 883, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SURREAL $\\rightarrow$ Human3.6M and SURREAL $\\rightarrow$ LSP. Tables 1 and 2 show the results for the hand tasks, where our method achieves state-of-the-art performance, surpassing the second best baseline SHOT++ with a significant lead of $3.3\\%$ and $3.0\\%$ , respectively. Moreover, our method outperforms the SHOT++ model by $4.1\\%$ and $3.9\\%$ on the PIP metric, respectively. For the human tasks, Tables 3 and 4 display the results. Our model achieves a considerable margin of $3.4\\%$ and $3.3\\%$ over the second best baseline SHOT++ for SURREAL $\\rightarrow$ Human3.6M and SURREAL $\\rightarrow$ LSP, respectively. Specifically, our method outperforms SHOT++ by $7.0\\%$ on the Wrist for the adaptation on Human3.6M and $5.3\\%$ on the Ankle for the adaptation on LSP. Furthermore, our approach remains competitive when compared with general domain adaptive HPE methods that need to access source data, as it surpasses UniFrame by $2.6\\%$", + "bbox": [ + 496, + 657, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4832", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/abaf841689c52e53750234650482f5f71a190ce65115d68e21a368fe06924024.jpg", + "image_caption": [ + "Source Only", + "UniFrame-SF", + "SHOT++", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 80, + 90, + 480, + 279 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 236, + 90, + 480, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 318, + 90, + 480, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f7a0788a9588aeda9fde1278823e96c31285fcb4d9ccb89ada1d47ef9174530c.jpg", + "image_caption": [ + "Figure 6. Qualitative results on H3D dataset (left) and LSP dataset (right)." + ], + "image_footnote": [], + "bbox": [ + 483, + 90, + 890, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "on RHD $\\rightarrow$ H3D, $0.3\\%$ on RHD $\\rightarrow$ FreiHand, and $1.2\\%$ on SURREAL $\\rightarrow$ LSP.", + "bbox": [ + 75, + 330, + 467, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results. Figs. 6 shows qualitative results of H3D on the left side and LSP on the right side. We use Source only, UniFrame-SF [16], SHOT++ [22], Ours, and Ground Truth for qualitative comparison. It is evident that our method outperforms other baselines significantly.", + "bbox": [ + 75, + 362, + 468, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Ablation Study on Framework", + "text_level": 1, + "bbox": [ + 76, + 449, + 346, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method contains two modules in Step A & Step B as the source-protect module (SP) and the target-relevant module (TR) separately, and here we focus on their functions. Moreover, we use MMT [6] as the baseline, which is referred in Table 1-4. Table 5 and 6 show the ablation study of frameworks on RHD $\\rightarrow$ H3D and SURREAL $\\rightarrow$ Human3.6M with different combinations of these modules.", + "bbox": [ + 75, + 473, + 468, + 579 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/74e37d18fdfec44a04d305e254589b2c0c02955fccd26db99fc6ae0797992137.jpg", + "table_caption": [ + "Table 5. Ablation of Frameworks on RHD $\\rightarrow$ H3D" + ], + "table_footnote": [], + "table_body": "
MethodMCPPIPDIPFinAll
MMT [6]78.774.172.565.375.3
MMT [6] +TR81.377.576.067.779.1
SP+TR (Ours)88.489.280.971.482.2
", + "bbox": [ + 78, + 601, + 468, + 667 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/274dacdfa8cda0f2546b000acb895bb9cf0253ec67af014b71ee321ddb2e4f32.jpg", + "table_caption": [ + "Table 6. Ablation of Frameworks on SURREAL $\\rightarrow$ Human3.6M" + ], + "table_footnote": [], + "table_body": "
MethodSldElbWristHipKneeAnkleAll
MMT [6]73.283.572.445.180.883.973.9
MMT [6] +TR74.486.575.748.981.385.176.2
SP+TR (Ours)77.988.880.452.384.286.978.7
", + "bbox": [ + 78, + 709, + 468, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results clearly demonstrate that both SP and TR contribute to improving the model's performance. Specifically, TR enhances the model's accuracy by $3.8\\%$ on RHD $\\rightarrow$ H3D and $2.3\\%$ on SURREAL $\\rightarrow$ Human3.6M, while SP leads to an improvement of $4.1\\%$ on RHD $\\rightarrow$ H3D and $2.5\\%$ on SURREAL $\\rightarrow$ Human3.6M compared to MMT. Notably, the two proposed modules provide similar levels of improvement.", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study on Proposed Losses", + "text_level": 1, + "bbox": [ + 500, + 330, + 805, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We performed a detailed ablation study on the three proposed losses, namely $\\mathcal{L}_{\\text{res}}$ , $\\mathcal{L}_{\\text{cst}}$ , and $\\mathcal{L}_{\\text{im}}$ , using the RHD→H3D and SURREAL→Human3.6M tasks. Tables 7 and 8 present the results. As we have previously discussed the functions and advantages of the source-protect and target-relevant modules in Sec. 4.2, we retained $\\mathcal{L}_{\\text{ft}}$ from the source-protect module and $\\mathcal{L}_{\\text{con}}$ from the target-relevant module in the Baseline.", + "bbox": [ + 496, + 358, + 890, + 479 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f03393a127d5078c4203a1db4a3f15ad0f7b4782d6eb64b144aa812807b4d5b4.jpg", + "table_caption": [ + "Table 7. Ablation of Losses on RHD $\\rightarrow$ H3D" + ], + "table_footnote": [], + "table_body": "
MethodMCPPIPDIPFinAll
Baseline85.386.878.568.179.4
Lres85.587.979.370.780.1
Lcst85.589.078.470.180.9
Lim85.887.278.769.880.3
Lcst&Lim86.988.880.170.381.4
Lres&Lcst&Lim88.489.280.971.482.2
", + "bbox": [ + 501, + 508, + 890, + 606 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bddd39bfb5c679fe09694acb77cd40f98456a27fa4517e524831ce37a112ea65.jpg", + "table_caption": [ + "Table 8. Ablation of Losses on SURREAL $\\rightarrow$ Human3.6M" + ], + "table_footnote": [], + "table_body": "
MethodSldElbWristHipKneeAnkleAll
Baseline75.486.176.846.583.085.075.8
Lres76.086.977.547.883.385.276.6
Lcst76.687.478.850.683.586.077.2
Lim76.287.178.349.783.285.776.9
Lcst&Lim77.387.779.251.283.686.478.0
Lres&Lcst&Lim77.988.880.452.384.286.978.7
", + "bbox": [ + 501, + 661, + 890, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We observe that each loss is able to boost the model's performance. Simply applying $\\mathcal{L}_{res}$ leads to a increase of $0.7\\%$ in RHD $\\rightarrow$ H3D and $0.8\\%$ in SURREAL $\\rightarrow$ Human3.6M. $\\mathcal{L}_{cst}$ causes an improvement of $1.5\\%$ in RHD $\\rightarrow$ H3D and $1.4\\%$ in SURREAL $\\rightarrow$ Human3.6M. As for $\\mathcal{L}_{im}$ , adding it causes a rise of $0.9\\%$ in RHD $\\rightarrow$ H3D and $1.1\\%$ in SURREAL $\\rightarrow$ Human3.6M. Besides, it is noticed that $\\mathcal{L}_{cst}$ plays a more important role than $\\mathcal{L}_{res}$ or $\\mathcal{L}_{im}$ .", + "bbox": [ + 496, + 763, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4833", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation of Sparsity Reduction", + "text_level": 1, + "bbox": [ + 76, + 90, + 352, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we investigate the effectiveness of our proposed sparsity reduction strategy. Specifically, we compare heatmap-based methods and vector-based methods for two loss functions: contrastive learning $(\\mathcal{L}_{cst})$ and information maximization $(\\mathcal{L}_{im})$ . We denote heatmap-based contrastive learning and information maximization as HBCL and HBIM, respectively, and vector-based contrastive learning and information maximization as VBCL and VBIM, respectively. Two tasks are selected and the results of these comparisons are presented in Tab. 9 and Tab. 10.", + "bbox": [ + 75, + 114, + 470, + 280 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/d22f528f41860fcb00913450ec48e424882d77db71c3a1214095febd6b38a4c3.jpg", + "table_caption": [ + "Table 9. Ablation of Sparsity Reduction on RHD $\\rightarrow$ H3D" + ], + "table_footnote": [], + "table_body": "
MethodMCPPIPDIPFinAll
MMT [6]78.774.172.565.375.3
MMT [6] + HBCL77.473.070.765.574.2
MMT [6] + VBCL81.776.674.467.978.7
MMT [6] + HBIM79.374.772.965.875.9
MMT [6] + VBIM80.376.173.666.577.5
", + "bbox": [ + 78, + 304, + 468, + 393 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/888a6dbdf919b579c242133023a49ed7e86f6ff9a74f841b7083db7bd61d0e3d.jpg", + "table_caption": [ + "Table 10. Ablation of Sparsity Reduction on SURREAL $\\rightarrow$ Human3.6M" + ], + "table_footnote": [], + "table_body": "
MethodSldElbWristHipKneeAnkleAll
MMT [6]73.283.572.445.180.883.973.9
MMT [6] + HBCL71.782.071.844.680.282.473.0
MMT [6] + VBCL75.687.075.149.283.386.376.7
MMT [6] + HBIM73.984.672.746.081.183.074.3
MMT [6] + VBIM75.386.474.348.682.485.776.0
", + "bbox": [ + 78, + 449, + 467, + 518 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Based on the results, it is evident that HBCL does not improve the model's performance significantly, whereas VBCL provides significant assistance. For instance, MMT+HBCL shows a decrease of $1.1\\%$ on the RHD $\\rightarrow$ H3D task, while MMT+VBCL exhibits an increase of $3.4\\%$ . Additionally, while HBIM can enhance the model's performance, it falls short of surpassing VBIM. For example, MMT+HBIM shows an increase of $0.6\\%$ on the RHD $\\rightarrow$ H3D task, whereas MMT+VBIM achieves an increase of $2.2\\%$ . In summary, it becomes evident that utilizing the vector-based spatial probability space is crucial for achieving optimal results.", + "bbox": [ + 75, + 537, + 468, + 720 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.5. Parameter Analysis", + "text_level": 1, + "bbox": [ + 76, + 731, + 264, + 748 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6c32538c30e8e130cc9a8da2626a57ea8de41f1cd745fad740f5e9ac387b5ede.jpg", + "image_caption": [ + "Figure 7. Parameter Analysis on RHD $\\rightarrow$ H3D (best viewed in color). a: Analysis on $\\alpha$ . b: Analysis on $\\beta$ . c: Analysis on $\\gamma$ ." + ], + "image_footnote": [], + "bbox": [ + 76, + 766, + 467, + 868 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use RHD $\\rightarrow$ H3D to illustrate the sensitivity of $\\alpha$ in Equation 5 and $\\beta$ and $\\gamma$ in Equation 11. The results are shown in Fig. 7. From it, we observe that $\\alpha = 0.7$ , $\\beta = 0.5$ , and $\\gamma = 0.85$ is the best choice. Moreover, it is noticed that the decreases from different parameters are limited to $0.2\\%$ for $\\alpha$ , $\\beta$ and $\\gamma$ . Therefore, our method shows stable performance over hyperparameters.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Qualitative Results on the Unseen Dataset COCO", + "text_level": 1, + "bbox": [ + 498, + 217, + 890, + 250 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/649bf5e302635bb29057003cbc74098a117511111957a00fef641b083dcc4fc3.jpg", + "image_caption": [ + "Figure 8. Qualitative results on the COCO dataset. Here we compare our method trained from SURREAL $\\rightarrow$ LSP with the source-only model." + ], + "image_footnote": [], + "bbox": [ + 500, + 276, + 890, + 422 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we showcase qualitative results on the COCO dataset [23], which was not utilized in our previous human dataset studies. It's important to note that our 14-keypoint poses are annotated differently from the 17-keypoint poses in COCO, which hinders providing quantitative results on COCO using its own metrics or APIs. Despite this limitation, we believe that our qualitative findings hold significant value. For our analysis, we leverage the model trained on the SURREAL $\\rightarrow$ LSP task in conjunction with the pre-trained SURREAL model. The results, presented in Fig. 8, demonstrate the model's capacity to generalize effectively to unseen datasets, implying that the adaptation process boosts the model's generalization ability beyond its original training dataset.", + "bbox": [ + 496, + 489, + 890, + 700 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 720, + 617, + 736 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose a new task named source-free domain adaptive human pose estimation that places an emphasis on the privacy of source data. Additionally, we propose a new framework that includes source-protect and target-relevant modules, which aim to alleviate the issues of catastrophic forgetting of source and the sparsity of spatial distributions, respectively. Our approach is evaluated on hand and human pose datasets through extensive experiments, demonstrating that it outperforms state-of-the-art methods by a considerable margin.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4834", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 295-305, June 2022.", + "[2] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020.", + "[3] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7103-7112, 2018.", + "[4] Wanxia Deng, Qing Liao, Lingjun Zhao, Deke Guo, Gangyao Kuang, Dewen Hu, and Li Liu. Joint clustering and discriminative feature alignment for unsupervised domain adaptation. IEEE Transactions on Image Processing, pages 7842-7855, 2021.", + "[5] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International Conference on Machine Learning, pages 1180-1189, 2015.", + "[6] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual meanteaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In International Conference on Learning Representations, 2020.", + "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020.", + "[8] Zhongyi Han, Haoliang Sun, and Yilong Yin. Learning transferable parameters for unsupervised domain adaptation. IEEE Transactions on Image Processing, 31:6424-6439, 2022.", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016.", + "[10] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 1501-1510, 2017.", + "[11] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, 2013.", + "[12] Junguang Jiang, Yifei Ji, Ximei Wang, Yufeng Liu, Jianmin Wang, and Mingsheng Long. Regressive domain adaptation for unsupervised keypoint detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6780-6789, 2021.", + "[13] Ruijie Jiang, Thuan Nguyen, Prakash Ishwar, and Shuchin Aeron. Supervised contrastive learning with hard negative samples. arXiv preprint arXiv:2209.00078, 2022.", + "[14] Rui Jin, Jing Zhang, Jianyu Yang, and Dacheng Tao. Multibranch adversarial regression for domain adaptive hand" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pose estimation. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):6125-6136, 2022.", + "[15] Sam Johnson and Mark Everingham. Clustered pose and nonlinear appearance models for human pose estimation. In British Machine Vision Conference, volume 2, page 5. Cite-seer, 2010.", + "[16] Donghyun Kim, Kaihong Wang, Kate Saenko, Margrit Betke, and Stan Sclaroff. A unified framework for domain adaptive pose estimation. arXiv preprint arXiv:2204.00172, 2022.", + "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.", + "[18] Vinod K Kurmi, Venkatesh K Subramanian, Vinay P Namboodiri, , and . Domain impression: A source data free domain adaptation method. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 615-625, 2021.", + "[19] Chen Li and Gim Hee Lee. From synthetic to real: Unsupervised domain adaptation for animal pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1482-1491, 2021.", + "[20] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9638-9647, 2020.", + "[21] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International Conference on Machine Learning, pages 6028-6039. PMLR, 2020.", + "[22] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8602-8617, 2022.", + "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, pages 740-755. Springer, 2014.", + "[24] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International Conference on Machine Learning, pages 97-105, 2015.", + "[25] Jiteng Mu, Weichao Qiu, Gregory D Hager, and Alan L Yuille. Learning from synthetic animals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12386-12395, 2020.", + "[26] Qucheng Peng, Zhengming Ding, Lingjuan Lyu, Lichao Sun, and Chen Chen. Toward better target representation for source-free and black-box domain adaptation. arXiv preprint arXiv:2208.10531, 2022.", + "[27] Ekkasit Pinyoanuntapong, Ayman Ali, Kalvik Jakkala, Pu Wang, Minwoo Lee, Qucheng Peng, Chen Chen, and Zhi Sun. Gaitsada: Self-aligned domain adaptation for mmwave gait recognition. arXiv preprint arXiv:2301.13384, 2023." + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4835", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3723-3732, 2018.", + "[29] Dahu Shi, Xing Wei, Liangqi Li, Ye Ren, and Wenming Tan. End-to-end multi-person pose estimation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11069-11078, 2022.", + "[30] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5693-5703, 2019.", + "[31] Hui Tang, Ke Chen, and Kui Jia. Unsupervised domain adaptation via structurally regularized deep clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8725-8735, 2020.", + "[32] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in Neural Information Processing Systems, 30, 2017.", + "[33] Jonathan Tompson, Ross Goroshin, Arjun Jain, Yann LeCun, and Christoph Bregler. Efficient object localization using convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 648-656, 2015.", + "[34] Jonathan J. Thompson, Arjun Jain, Yann LeCun, and Christoph Bregler. Joint training of a convolutional network and a graphical model for human pose estimation. Advances in neural information processing systems, 27, 2014.", + "[35] Eric Tzeng, Judy Hoffman, Ning Zhang, Kate Saenko, and Trevor Darrell. Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:1412.3474, 2014.", + "[36] Gul Varol, Javier Romero, Xavier Martin, Naureen Mahmood, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning from synthetic humans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 109-117, 2017.", + "[37] Haixin Wang, Lu Zhou, Yingying Chen, Ming Tang, and Jinqiao Wang. Regularizing vector embedding in bottom-up human pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI, pages 107-122. Springer, 2022.", + "[38] Haifeng Xia, Handong Zhao, and Zhengming Ding. Adaptive adversarial network for source-free domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9010-9019, 2021.", + "[39] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In Proceedings of the European Conference on Computer Vision (ECCV), pages 466-481, 2018.", + "[40] Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, and Ce Zhu. Distribution-aware coordinate representation for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7093-7102, 2020." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Zhengyi Zhao, Tianyao Wang, Siyu Xia, and Yangang Wang. Hand-3d-studio: A new multi-view system for 3d hand reconstruction. In ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2478-2482, 2020.", + "[42] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In Proceedings of the IEEE International Conference on Computer Vision, pages 4903-4911, 2017.", + "[43] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 813-822, 2019." + ], + "bbox": [ + 501, + 90, + 892, + 289 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4836", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Source-free Domain Adaptive Human Pose Estimation/6db7e049-3b2d-4942-adce-fb8455dd18b2_model.json b/2023/Source-free Domain Adaptive Human Pose Estimation/6db7e049-3b2d-4942-adce-fb8455dd18b2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..06091233126ca5cd92f1bb638a109b90729e645e --- /dev/null +++ b/2023/Source-free Domain Adaptive Human Pose Estimation/6db7e049-3b2d-4942-adce-fb8455dd18b2_model.json @@ -0,0 +1,2290 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.21, + 0.131, + 0.761, + 0.152 + ], + "angle": 0, + "content": "Source-free Domain Adaptive Human Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.182, + 0.764, + 0.217 + ], + "angle": 0, + "content": "Qucheng Peng, Ce Zheng, Chen Chen \nCenter for Research in Computer Vision, University of Central Florida" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.219, + 0.728, + 0.234 + ], + "angle": 0, + "content": "{qucheng.peng,ce.zheng}@ucf.edu，chen.chen@crcv.ucf.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.301, + 0.471, + 0.437 + ], + "angle": 0, + "content": "Human Pose Estimation (HPE) is widely used in various fields, including motion analysis, healthcare, and virtual reality. However, the great expenses of labeled real-world datasets present a significant challenge for HPE. To overcome this, one approach is to train HPE models on synthetic datasets and then perform domain adaptation (DA) on real-world data. Unfortunately, existing DA methods for HPE neglect data privacy and security by using both source and target data in the adaptation process." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.438, + 0.471, + 0.71 + ], + "angle": 0, + "content": "To this end, we propose a new task, named source-free domain adaptive HPE, which aims to address the challenges of cross-domain learning of HPE without access to source data during the adaptation process. We further propose a novel framework that consists of three models: source model, intermediate model, and target model, which explores the task from both source-protect and target-relevant perspectives. The source-protect module preserves source information more effectively while resisting noise, and the target-relevant module reduces the sparsity of spatial representations by building a novel spatial probability space, and pose-specific contrastive learning and information maximization are proposed on the basis of this space. Comprehensive experiments on several domain adaptive HPE benchmarks show that the proposed method outperforms existing approaches by a considerable margin. The codes are available at https://github.com/davidpengucf/SFDAHPE." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.739, + 0.21, + 0.755 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The accuracy of 2D human pose estimation (HPE) from monocular images has been significantly improved through the use of deep neural networks. However, obtaining sufficient labeled data for training such models is a laborious and time-consuming task, particularly for real-world datasets that require diverse appearances and poses with ground truth annotations. In contrast, the development of computer graphics and game engines has facilitated the generation of annotated synthetic datasets that can generate a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.895, + 0.362 + ], + "angle": 0, + "content": "virtually unlimited number of labeled samples. Nevertheless, models trained on synthetic data may not perform well on real-world data due to the domain gap or differences in data distribution. To address this challenge, domain adaptation (DA) is essential to enable synthetic data-trained models to generalize to real-world data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.895, + 0.595 + ], + "angle": 0, + "content": "Recently, source-free domain adaptation has gained significant attention due to its practicality in real-world scenarios, where using labeled source data during the adaptation is expensive or not feasible due to data privacy and security concerns. However, existing source-free domain adaptation methods in classification tasks may not be directly applicable to more complex tasks, such as 2D human pose estimation (HPE). This is because HPE requires estimating the spatial locations of different body joints, instead of just the object's category as classification does, making it more challenging to adapt from source to target domains. Therefore, there is a pressing need to develop effective source-free domain adaptation techniques specifically tailored to HPE tasks. This motivates us to propose a new task named source-free domain adaptive HPE in this paper." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The task of source-free domain adaptive HPE presents challenges from both the source-free DA and pose estimation perspectives. Specifically, in traditional DA tasks, source data can participate in the adaptation process, allowing for a focus on reducing domain shifts between source and target. However, in source-free DA, only the pretrained source model can be used during adaptation, leading to catastrophic forgetting of the source domain. Directly applying the source pretrained model can also introduce noise due to domain shifts. Therefore, balancing the absorption of the source model's knowledge with the resistance of the source model's noise is critical. Current methods tend to adopt a progressive replacement strategy to alleviate the runoff of source information. For example, [21, 22] use the previous model to supervise ongoing adaptation, thereby evoking the source-related memory, while [32, 6] use the exponential moving average (EMA) strategy to slow down the forgetting of the source. Although these methods alleviate the negative impact of noise successfully, they modify the overall parameters of the pretrained source model" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4826" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.166 + ], + "angle": 0, + "content": "significantly, leading to the disappearance of source-related knowledge and representations during the adaptation process. Therefore, it is necessary to develop source-protect modules that can protect source information more effectively while resisting noise at the same time." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.172, + 0.468, + 0.67 + ], + "angle": 0, + "content": "Moreover, we outline another challenge of source-free domain adaptive HPE: the sparsity of keypoints (i.e. key body joints) in an image. When it comes to domain adaptation (DA) in classification, achieving distribution alignment to reduce domain gap is a straightforward process as the outputs can be represented in a probability space that includes all the categories. However, the same cannot be said for Human Pose Estimation (HPE) based on a complete image, as each keypoint has a specific distance from others based on bone lengths, and the number of keypoints \\( K \\) (e.g., 14 for human pose and 21 for hand pose) is significantly smaller than the number of pixels in an image (e.g., \\( 512 \\times 512 \\approx 0.26\\mathrm{M} \\)), making it a challenging task. Although heatmaps [34, 33] are commonly used to predict keypoints (heatmap representation-based 2D HPE is considered as the mainstream approach for HPE) while reducing dimensionality in HPE, the number of pixels inside the heatmap is still considerable (e.g. \\( 128 \\times 128 \\approx 16\\mathrm{K} \\)) in comparison to \\( K \\). Therefore, building a spatial probability space for the keypoints based on either the entire image or the heatmap is not ideal. [12] proposes constructing a low-dimensional distribution based solely on the \\( K \\) keypoints, effectively transforming the HPE task into a \\( K \\)-classification task. However, this spatial distribution relies on clear decision boundaries between distinctive keypoints, which is not reasonable in HPE. In other words, all samples can be classified to a certain category in classification, but most pixels inside an image (e.g. more than \\( 97\\% \\) if the fault tolerance is decided by the metric PCK@0.05) can not be assigned to a certain keypoint in HPE, but mispredictions may occur at these unpresented locations. Hence, it is necessary to develop a novel spatial probability space that reduces sparsity while providing complete representations of the prediction space." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.675, + 0.468, + 0.9 + ], + "angle": 0, + "content": "To address the aforementioned task-specific challenges, we propose a novel framework that incorporates source-free domain adaptive HPE from both source-protect and target-relevant perspectives. Our framework consists of three models: the source model, the intermediate model, and the target model. Source model preserves source information, and target model absorbs knowledge from target data, while intermediate model interacts with both of them to reduce the domain gap between source and target implicitly, thereby used for the final inference. In the source-protect modules, we conduct knowledge transfer between the source model and the intermediate model to prevent catastrophic forgetting of the source-related information. We transfer knowledge from the source model to the intermediate model while keeping the source model's feature extractor fixed. To re" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.89, + 0.226 + ], + "angle": 0, + "content": "sist noise, we propose a new residual loss and finetune the source model's regressor. In the target-relevant modules, we conduct executions between the intermediate model and the target model to mitigate domain shifts between the source and target domains. We build a new spatial probability space for HPE to alleviate the sparsity problem. On the basis of this space, we propose contrastive learning and information maximization designed specifically for human pose estimation. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.228, + 0.89, + 0.287 + ], + "angle": 0, + "content": "- We construct a new task named source-free domain adaptive human pose estimation, which focuses on the cross-domain learning of HPE without access to source data during the adaptation process." + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.289, + 0.89, + 0.348 + ], + "angle": 0, + "content": "- We propose a new framework consisting of three models: source model, intermediate model, and target model, which explores the task from both source-protect and target-relevant perspectives." + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.349, + 0.89, + 0.408 + ], + "angle": 0, + "content": "- We conduct comprehensive experiments on several domain adaptive HPE benchmarks and the results show that our method outperforms the existing approaches by a considerable margin." + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.228, + 0.89, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.423, + 0.64, + 0.438 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.448, + 0.89, + 0.644 + ], + "angle": 0, + "content": "Human Pose Estimation. The heatmap-based method [29, 37] is currently a widely used technique in 2D HPE, which involves generating heatmaps by placing 2D Gaussian kernels on potential keypoints, followed by converting these heatmaps to coordinates. For instance, [3] uses global- and refine-nets, while [39] simplifies the network structures through the Simple Baselines technique. Some approaches have focused on utilizing features at different scales [30], while others aim to improve the quality of heatmaps by extending Gaussians to second order [40] or using coupled embeddings for heatmap regression [37]. In this study, we adopt the same heatmap-based HPE methods as other domain adaptive HPE works [12, 16]." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.646, + 0.89, + 0.9 + ], + "angle": 0, + "content": "General Domain Adaptation. There are two main categories of methods for achieving general domain adaptation, which requires access to both the source and target data during adaptation. The first category is metric-based methods that explicitly measure the discrepancy between the source and target domains. For example, [24] uses maximum mean discrepancy, and [35] applies deep domain confusion. Recent work, such as [4], jointly performs clustering and discrimination for alignment using contrastive learning [2, 13]. The second category is GAN-based [7] methods that build a min-max game for two players related to the source and target domains. For example, [5, 28] adopts domain confusion to confuse the two players, while [31, 27] shows that discriminative clustering on the target benefits adaptation. However, access to source data is required in the general domain adaptation setting, which raises concerns about data privacy and security." + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.956 + ], + "angle": 0, + "content": "4827" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.348 + ], + "angle": 0, + "content": "Source-Free Domain Adaptation. Source-Free has become the mainstream paradigm for alleviating concerns about data privacy and security in domain adaptation, as shown in the work of [20, 21, 26]. There are two technique routes under the source-free setting: self-supervision and virtual source transfer. For the self-supervised methods, [21] is the most representative, which introduces information maximization to assist adaptation. Other self-supervised methods include [38], which treats the problem from a discriminative perspective, and [1, 26], which proposes online pseudo label refinement. In virtual source methods, most build GANs to generate virtual source data. Examples include conditional GANs used in [18], and collaborative GANs in [20], which achieve better generations. However, most source-free DA methods for classification are not applicable to HPE tasks due to the sparsity of keypoints' distributions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.473, + 0.668 + ], + "angle": 0, + "content": "Domain Adaptive Human Pose Estimation. There are two main categories of methods used for domain-adaptive HPE. The first category involves a shared structure, where the weights of the networks from both the source pretrain and target adaptation are shared. CC-SSL [25] employs a single end-to-end trained network, RegDA [12] utilizes one shared feature extractor and two separate regressors, and TransPar [8] emphasizes transferable parameters using a similar structure. The second category is the unshared structure, where the pretrain and adaptation create a teacher-student paradigm. In this category, MDAM [19] addresses this structure alongside a novel pseudo-label strategy, UniFrame [16] modifies the classic Mean-Teacher [32] model by combining it with style transfer [10], and MarsDA [14] applies a teacher-student approach to edit RegDA. However, all these existing domain adaptive HPE approaches are not source-free, which raises concerns about data privacy and security. Therefore, it is necessary to consider source-free domain adaptive HPE. To our knowledge, we are the first work to study this challenging problem." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.684, + 0.17, + 0.7 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.71, + 0.295, + 0.726 + ], + "angle": 0, + "content": "3.1. Preliminary of 2D HPE" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In 2D HPE, we are given a labeled pose dataset \\(\\mathcal{D} = (x_i,y_i)\\), where \\(x_{i}\\in \\mathbb{R}^{C\\times H\\times W}\\) is an image and \\(y_{i}\\in \\mathbb{R}^{K\\times 2}\\) are the corresponding keypoint coordinates. Here, \\(H\\) and \\(W\\) represent the height and width of the image, respectively, and \\(C\\) is the number of channels. Additionally, \\(K\\) represents the number of keypoints. To ensure smoother training, most existing 2D methods represent the keypoint coordinates as heatmaps \\(HM_{i}\\in \\mathbb{R}^{K\\times H^{\\prime}\\times W^{\\prime}}\\), where \\(H^{\\prime}\\) and \\(W^{\\prime}\\) are the height and width of the heatmaps, respectively. These heatmaps are used to represent the coordinates in a spatial way. The process of how the model works is defined" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.087, + 0.895, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.203, + 0.895, + 0.26 + ], + "angle": 0, + "content": "Figure 1. Process of heatmap-based 2D HPE. The model includes two components: the feature extractor \\( G \\) and the regressor \\( F \\). After feeding the input image to the model, we obtain \\( K \\) output heatmaps corresponding to \\( K \\) keypoints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.283, + 0.895, + 0.452 + ], + "angle": 0, + "content": "in Fig. 1. Ground truth heatmaps are generated by a 2D Gaussian centered at the ground truth joint location, with one heatmap for each keypoint. We denote the transform from coordinates to heatmaps as \\( T: \\mathbb{R}^{K \\times 2} \\to \\mathbb{R}^{K \\times H' \\times W'} \\). The training processes are conducted between the output heatmaps and ground truth heatmaps using MSE for supervised learning. In the inference stage, well-trained supervised models output heatmaps that predict the probability of joints occurring at each pixel. For each heatmap, the pixel with the highest probability is considered to be the predicted location of the joint." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.459, + 0.686, + 0.473 + ], + "angle": 0, + "content": "3.2. Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.483, + 0.895, + 0.74 + ], + "angle": 0, + "content": "In the source-free domain adaptive HPE setting, we have a source domain dataset \\( \\mathcal{S} = \\{(x_i^s, y_i^s)\\}_{i=1}^{n_s} \\) with \\( n_s \\) labeled pose samples. \\( x_i^s \\in \\mathbb{R}^{C \\times H \\times W} \\) is the source image and \\( y_i^s \\in \\mathbb{R}^{K \\times 2} \\) is the corresponding pose annotation. Besides, there exists a target domain dataset \\( \\mathcal{T} = \\{x_i^t\\}_{i=1}^{n_t} \\) that includes \\( n_t \\) unlabeled pose samples. The source and target domains share the same label space but lie in different distributions (e.g. synthetic vs. real). The training procedure is split into two stages: the pretrain stage and the adaptation stage. In the pretrain stage, the labeled source domain dataset \\( \\mathcal{S} \\) is used to train the source model \\( f^{sr} \\). However, what distinguishes source-free domain adaptive HPE from general domain adaptive HPE is the adaptation stage, where only the source model \\( f^{sr} \\) and target domain dataset \\( \\mathcal{T} \\) are available for training, and the source domain dataset cannot be used anymore. Our objective is to develop a model that performs well on the target dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.748, + 0.72, + 0.765 + ], + "angle": 0, + "content": "3.3. Pipeline of Our Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.715, + 0.786 + ], + "angle": 0, + "content": "3.3.1 Source Model Pretrain" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.903 + ], + "angle": 0, + "content": "For DA in HPE, labeled source data are applied for pretraining. Same as most existing domain adaptive pose estimation methods [25, 19, 12, 16], the network is composed of a feature extractor \\( G \\) (such as ResNet backbone) and a pose regressor \\( F \\), so the source model can be represented as \\( f^{sr} = F^{sr}(G^{sr}(\\cdot)) \\). Based on the supervised heatmap-based loss (MSE loss), we propose the overall objective for" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4828" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.468, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.088, + 0.895, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.363, + 0.895, + 0.419 + ], + "angle": 0, + "content": "Figure 2. Overall framework of our method. It includes three models, each with one feature extractor and one regressor. The adaptation contains two steps: Step A and Step B. Step A is applied for source-protect adaptation, while Step B aims at target-relevant adaptation. The solid arrows show the forward process of associated losses, and the dotted arrows show the backward updating process of associated losses. EMA means the exponential moving average process from target model to intermediate model to update intermediate model's weights." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.444, + 0.206, + 0.459 + ], + "angle": 0, + "content": "source pretraining:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.466, + 0.469, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p r e t r a i n}} = \\mathbb {E} _ {\\left(x _ {i} ^ {s}, y _ {i} ^ {s}\\right) \\in \\mathcal {S}} \\mathcal {L} _ {\\text {m s e}} \\left(F ^ {s r} \\left(G ^ {s r} \\left(x _ {i} ^ {s}\\right)\\right), T \\left(y _ {i} ^ {s}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.471, + 0.552 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{mse}\\) is the Mean Squared Error (MSE) loss between the heatmap representations of prediction and ground truth, and \\(T\\) is the transform from coordinates to heatmaps defined in Sec. 3.1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.57, + 0.296, + 0.585 + ], + "angle": 0, + "content": "3.3.2 Adaptation Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.473, + 0.903 + ], + "angle": 0, + "content": "As Fig. 2 shows, our method contains three models: source model \\( f^{sr} = F^{sr}(G^{sr}(\\cdot)) \\), intermediate model \\( f^{in} = F^{in}(G^{in}(\\cdot)) \\), and target model \\( f^{tg} = F^{tg}(G^{tg}(\\cdot)) \\). Each model includes one feature extractor and one regressor. Assume the heatmaps generated by source model as \\( H_{i}^{sr} = F^{sr}(G^{sr}(x_{i}^{t})) = [h_{i,1}^{sr}, h_{i,2}^{sr}, \\dots, h_{i,K}^{sr}]^{\\top} \\), and those from intermediate model and target model respectively as \\( H_{i}^{in} = F^{in}(G^{in}(x_{i}^{t})) = [h_{i,1}^{in}, h_{i,2}^{in}, \\dots, h_{i,K}^{in}]^{\\top} \\) and \\( H_{i}^{tg} = F^{tg}(G^{tg}(x_{i}^{t})) = [h_{i,1}^{tg}, h_{i,2}^{tg}, \\dots, h_{i,K}^{tg}]^{\\top} \\), where \\( K \\) is the number of keypoints. We also assume the keypoint indexing space as \\( \\mathcal{P} = [1,2,\\dots,K] \\). The source model is used to maintain source information, while the target model is applied to learn target knowledge. The intermediate model is the transition model between these two models. It serves as a bridge between source and target as it learns from source and target simultaneously, thus eliminating the domain gap implicitly. Therefore, it outperforms both the source and target models and is used for the final inference. When starting the adaptation, \\( f^{in} \\) and \\( f^{tg} \\) are initialized from the source model \\( f^{sr} \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.443, + 0.892, + 0.489 + ], + "angle": 0, + "content": "For a single adaptation iteration, two steps are executed as shown in Fig. 2. Step A is for the source-protect adaptation, and Step B aims to adapt in a target-relevant way." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.892, + 0.691 + ], + "angle": 0, + "content": "In Step A, our goal is to overcome catastrophic forgetting on source domain, while eliminating the noise from the source model due to domain shift. Therefore, the source model's feature extractor \\( G^{sr} \\) is fixed to retain source knowledge, while its regressor \\( F^{sr} \\) keeps updating to improve regression accuracy, hence reducing noises in the outputs. On the opposite, the intermediate model's regressor \\( F^{in} \\) is fixed, but its feature extractor \\( G^{in} \\) continues to update. That's because we want the feature extractor to learn more source knowledge on the representation level, while the noise from the source model due to domain shifts can be inhibited." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.892, + 0.851 + ], + "angle": 0, + "content": "In Step B, we aim to conduct the adaptation between the intermediate model and the target model in a target-relevant manner. During the process, several pose-specific techniques are utilized to minimize the discrepancy between these two models. Though both \\( f^{in} \\) and \\( f^{tg} \\) are updated at this step, \\( f^{tg} \\) is updated via back-propagation while \\( f^{in} \\) is via exponential moving average from \\( f^{tg} \\), since averaging model weights over training steps tends to produce a more accurate model than using the final weights directly [32]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In the following two subsections, we introduce the detailed techniques used for Step A and Step B." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4829" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.473, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.257, + 0.47, + 0.311 + ], + "angle": 0, + "content": "Figure 3. Proposed residual loss. Here we choose two heatmaps at the same location generated from source model and intermediate model for illustration. For simplicity, here the heatmap's scale is set as \\(3 \\times 3\\), while the actual size is \\(64 \\times 64\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.337, + 0.299, + 0.352 + ], + "angle": 0, + "content": "3.3.3 Source-protect Modules" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.361, + 0.469, + 0.435 + ], + "angle": 0, + "content": "In this part, we propose modules related to source-free-specific adaptation. The biggest challenge is to distill knowledge from the source model and resist the potential noise due to domain shift. First, we introduce the finetune loss that calibrates on the source model:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.447, + 0.469, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f t} = \\mathbb {E} _ {x _ {i} ^ {t} \\in \\mathcal {T}} \\mathcal {L} _ {m s e} \\left(F ^ {s r} \\left(G ^ {s r} \\left(x _ {i} ^ {t}\\right)\\right), F ^ {i n} \\left(G ^ {i n} \\left(x _ {i} ^ {t}\\right)\\right)\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.469, + 0.553 + ], + "angle": 0, + "content": "To be notified, we hope the regression results of the source model can be more accurate, while the source knowledge can be retained. Thus, the feature extractor is fixed so that the source representations continue to be generated, and the regressor is being updated:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.564, + 0.469, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\theta_ {F ^ {s r}} ^ {t + 1} = \\theta_ {F ^ {s r}} ^ {t} - \\lambda^ {s r} \\frac {\\partial L _ {f t}}{\\partial \\theta_ {F ^ {s r}} ^ {t}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.469, + 0.681 + ], + "angle": 0, + "content": "where \\(\\theta_F^s\\) is the weights of the source model's regressor \\(F^s\\), while \\(t\\) denotes the training step and \\(\\lambda^s\\) is the source model's learning rate. However, improving source model is not enough. What's more, we propose the residual loss to improve intermediate model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.469, + 0.903 + ], + "angle": 0, + "content": "The application of Equation 2 facilitates mutual knowledge transfer between the source and intermediate models. Apart from the pixel with the highest confidence in a heatmap, the other pixels also contain valuable information that benefits the knowledge transfer, and that's why we propose residual loss in Fig. 3. From the example in the figure, the most confident pixel (confidence \\(= 1.0\\)) in the source heatmap is removed, along with the pixel at the same location which has a confidence of 0.8 in the intermediate heatmap marked with red. The pair marked with blue follows the same rule but on the basis of intermediate heatmap. Then we build adaptive residual heatmaps denoted as \\(\\hat{h}_{i,j}^{sr}\\) and \\(\\hat{h}_{i,j}^{in}\\). Moreover, we employ KL divergence to ensure that \\(\\hat{h}_{i,j}^{in}\\) approaches \\(\\hat{h}_{i,j}^{sr}\\), thereby preserving the source in" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.088, + 0.877, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.196, + 0.892, + 0.266 + ], + "angle": 0, + "content": "Figure 4. Illustration of the projection of a heatmap. In order to reduce the sparsity of a heatmap, we project it into a horizontal vector and a vertical vector. For simplicity, here the heatmap's scale is set as \\(4 \\times 4\\) and the vector's scale as \\(4 \\times 1\\), while the actual heatmap's size is \\(64 \\times 64\\) and the vector's size is \\(64 \\times 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.292, + 0.885, + 0.306 + ], + "angle": 0, + "content": "formation. Based on these, we deduce the residual loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.318, + 0.892, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e s} = \\mathbb {E} _ {x _ {i} ^ {t} \\in \\mathcal {T}} \\mathbb {E} _ {j \\in \\mathcal {P}} \\mathrm {D} _ {\\mathrm {K L}} \\left(\\sigma \\left(\\hat {h} _ {i, j} ^ {s r} / \\tau\\right) \\mid \\mid \\sigma \\left(\\hat {h} _ {i, j} ^ {i n} / \\tau\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.348, + 0.892, + 0.437 + ], + "angle": 0, + "content": "where \\(\\mathrm{D}_{\\mathrm{KL}}\\) is the KL divergence, and \\(\\sigma (\\cdot)\\) is the softmax function. Besides, \\(\\tau\\) is the temperature used to scale the residual heatmaps and is empirically set to 0.3. By combining the finetune loss and the residual loss, we get the objective function with the optimization process for updating the intermediate model:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.448, + 0.891, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i n} = \\mathcal {L} _ {f t} + \\alpha \\mathcal {L} _ {r e s}, \\quad \\theta_ {G ^ {i n}} ^ {t + 1} = \\theta_ {G ^ {i n}} ^ {t} - \\lambda^ {i n} \\frac {\\partial L _ {i n}}{\\partial \\theta_ {G ^ {i n}} ^ {t}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.494, + 0.892, + 0.537 + ], + "angle": 0, + "content": "where \\( t \\) is the training step, and \\( \\lambda^{in} \\) is the learning rate for \\( G^{in} \\). Moreover, \\( \\alpha \\) is the trade-off hyperparameter between these two losses." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.725, + 0.574 + ], + "angle": 0, + "content": "3.3.4 Target-relevant Modules" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.78 + ], + "angle": 0, + "content": "This section introduces a module that is associated with target information. As previously mentioned, the sparsity of keypoints in the image presents a significant challenge in our task. To address this issue, we propose using projected horizontal and vertical vectors of heatmaps as spatial representations, as shown in Fig. 4. This approach reduces the sparsity of the spatial probability space effectively since the number of overall digits inside the heatmap is \\(64 \\times 64 \\approx 4K\\), while the number of units in each vector is 64. Therefore, we can effectively reduce the sparsity of the spatial probability space. With 64 units in each vector, the spatial probability space size is acceptable compared to the number of keypoints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Despite the reduction in sparsity, the completeness of spatial representations is not affected. That is because all the coordinates can always be divided into two representations, horizontally and vertically. In such a case, our spatial representation is better than the \\(K\\)-classification probability space proposed by [12], which only includes the keypoint-relevant information and does not take the units in the heatmap that are not keypoints into consideration." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4830" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.086, + 0.473, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.225, + 0.47, + 0.267 + ], + "angle": 0, + "content": "Figure 5. Illustration of building positive and negative keypoint pairs for contrastive learning between the output heatmaps of the intermediate model and target model." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.291, + 0.47, + 0.398 + ], + "angle": 0, + "content": "In Fig. 4, we introduce the projection of a heatmap from both horizontal and vertical directions. For the heatmap \\( h_{i,j}^{in} \\) generated by the intermediate model from \\( i \\)th sample's \\( j \\)th keypoint, its projection can be represented as \\( proj( h_{i,j}^{in}) = (v_{i,jx}^{in}, v_{i,jy}^{in}) \\), where \\( v_{i,jx}^{in} \\) is the horizontal vector and \\( v_{i,jy}^{in} \\) is the vertical vector. This vector pair is the foundation of the following modules and losses that we propose." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.471, + 0.538 + ], + "angle": 0, + "content": "First, we propose the pose-specific contrastive loss between the intermediate model and target model. According to the observation from [12], wrong predictions are usually located at other keypoints' positions. In such a case, improving regression can be achieved by minimizing the discrepancy of heatmaps at the same location and maximizing the discrepancy of heatmaps at different locations. And that's how we define the positive and negative pair in Fig. 5. Denote the similarity of two heatmaps generated by the intermediate model and target model respectively as:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.544, + 0.47, + 0.582 + ], + "angle": 0, + "content": "\\[\ns i m \\left(h _ {i, j} ^ {i n}, h _ {i, k} ^ {t g}\\right) = \\frac {1}{2} \\left(\\frac {v _ {i , j x} ^ {i n} \\cdot v _ {i , k x} ^ {t g}}{| | v _ {i , j x} ^ {i n} | | | | | v _ {i , k x} ^ {t g} | |} + \\frac {v _ {i , j y} ^ {i n} \\cdot v _ {i , k y} ^ {t g}}{| | v _ {i , j y} ^ {i n} | | | | | v _ {i , k y} ^ {t g} | |}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.587, + 0.47, + 0.618 + ], + "angle": 0, + "content": "On the basis of the similarity, we define the overall pose-specific contrastive loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.624, + 0.47, + 0.661 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c s t} = - \\mathbb {E} _ {x _ {i} ^ {t} \\in \\mathcal {T}} \\mathbb {E} _ {j \\in \\mathcal {P}} \\log \\frac {\\exp (s i m (h _ {i , j} ^ {i n} , h _ {i , j} ^ {t g}))}{\\Sigma_ {k \\in \\mathcal {P}} \\exp (s i m (h _ {i , j} ^ {i n} , h _ {i , k} ^ {t g}))}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.471, + 0.803 + ], + "angle": 0, + "content": "Next, we propose a pose-specific self-supervised technique called pose-specific information maximization. Information maximization is very helpful to source-free domain adaptation for classification [21] because it makes the target outputs individually certain and globally diverse, hence mitigating the domain gap implicitly. However, the improvements are limited due to the sparsity of heatmaps. With the help of projected vectors, we refine the information maximization loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.089, + 0.812, + 0.47, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i m} = \\mathbb {E} _ {x _ {i} ^ {t} \\in \\mathcal {T}} \\left(\\mathcal {L} _ {e n t x} \\left(x _ {i} ^ {t}\\right) + \\mathcal {L} _ {e n t y} \\left(x _ {i} ^ {t}\\right) - \\mathcal {L} _ {d i v} \\left(x _ {i} ^ {t}\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.84, + 0.469, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\mathcal {L} _ {e n t x} \\left(x _ {i} ^ {t}\\right) = - \\mathbb {E} _ {j \\in \\mathcal {P}} \\left[ \\sigma \\left(v _ {i, j x} ^ {t g}\\right) \\log \\sigma \\left(v _ {i, j x} ^ {t g}\\right) \\right], \\\\ \\mathcal {L} _ {e n t y} \\left(x _ {i} ^ {t}\\right) = - \\mathbb {E} _ {j \\in \\mathcal {P}} \\left[ \\sigma \\left(v _ {i, j y} ^ {t g}\\right) \\log \\sigma \\left(v _ {i, j y} ^ {t g}\\right) \\right], \\\\ \\mathcal {L} _ {d i v} \\left(x _ {i} ^ {t}\\right) = - \\left[ \\mathbb {E} _ {j \\in \\mathcal {P}} \\sigma \\left(h _ {i, j} ^ {t g}\\right) \\right] \\log \\left[ \\mathbb {E} _ {j \\in \\mathcal {P}} \\sigma \\left(h _ {i, j} ^ {t g}\\right) \\right]. \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.15 + ], + "angle": 0, + "content": "Here \\(\\mathcal{L}_{entx}\\) is the entropy of horizontal vectors, and \\(\\mathcal{L}_{enty}\\) is the entropy of vertical vectors. The last term in Equation 8 is the diversity-promoting objective. Moreover, we introduce the consistency loss between the two models' outputs:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.156, + 0.892, + 0.175 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c o n} = \\mathbb {E} _ {\\left(x _ {i} ^ {s}, y _ {i} ^ {s}\\right) \\in \\mathcal {S}} \\mathcal {L} _ {m s e} \\left(F ^ {i n} \\left(G ^ {i n} \\left(x _ {i} ^ {s}\\right)\\right), F ^ {t g} \\left(G ^ {t g} \\left(x _ {i} ^ {s}\\right)\\right)\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.182, + 0.892, + 0.213 + ], + "angle": 0, + "content": "By summarizing above, we conclude the overall objective loss function for the target model as:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.224, + 0.892, + 0.241 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t g t} = \\mathcal {L} _ {c o n} + \\beta \\mathcal {L} _ {c s t} + \\gamma \\mathcal {L} _ {i m}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.251, + 0.766, + 0.266 + ], + "angle": 0, + "content": "where \\(\\beta\\) and \\(\\gamma\\) are trade-off parameters." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.266, + 0.892, + 0.297 + ], + "angle": 0, + "content": "Moreover, the intermediate model is updated via the exponential moving average strategy (EMA) in this step:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.305, + 0.892, + 0.327 + ], + "angle": 0, + "content": "\\[\n\\theta_ {f ^ {i n}} ^ {t + 1} = \\eta \\theta_ {f ^ {i n}} ^ {t} + (1 - \\eta) \\theta_ {f ^ {t g}} ^ {t}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.334, + 0.892, + 0.365 + ], + "angle": 0, + "content": "where \\( t \\) denotes the step of training and \\( \\eta \\) denotes the smoothing coefficient set to 0.999 by default." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.378, + 0.633, + 0.395 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.892, + 0.614 + ], + "angle": 0, + "content": "Datasets. We use three human pose datasets and three hand pose datasets to validate our approach. SURREAL [36] is utilized as the source dataset and encompasses six million synthetic human pose images. Human3.6M [11], one of the target human pose datasets, is a frequently used real-world dataset in the community, consisting of 3.6 million images distributed into seven folds. According to previous research [12, 16], S1, S5, S6, S7, and S8 are designated as the training set, whereas S9 and S11 serve as the testing set. The other target dataset is Leeds Sports Pose [15] (LSP), which is a real-world dataset with 2,000 images, and we use all of them for adaptation. We focus on two domain adaptation tasks: SURREAL \\(\\rightarrow\\) Human3.6M and SURREAL \\(\\rightarrow\\) LSP for human pose tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Rendered Hand Pose Dataset [42] (RHD) is the source hand dataset with 43,986 synthetic hand images, of which 41,258 are allocated for training, while the remaining 2,728 images are utilized for validation. Hand-3D-Studio [41] (H3D) is one of the two target hand datasets, including 22,000 real-world frames. For training, we employed 18,800 frames and reserved the rest for testing, following the protocol outlined in previous research papers [12, 16]. The other target hand dataset, FreiHand [43], provides 130k real-world images, and we employed all of them for our adaptation task. Our focus is on two domain adaptation tasks: RHD \\(\\rightarrow\\) H3D and RHD \\(\\rightarrow\\) FreiHand." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. We adopt Simple Baseline [39] with ResNet-101 [9] as the HPE backbone, following previous works [12, 16]. In the pretrain process, we conduct 40 epochs, each with 500 iterations, using the Adam [17] optimizer with an initial learning rate of 1e-4. The learning rate decreases to 1e-5 at 25 epochs. For the adaptation process, we execute 40 epochs with 30,000 iterations. In step" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4831" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "A, we choose an initial learning rate of 1e-4 for \\( F^{sr} \\) and 1e-5 for \\( G^{in} \\). Step B sets the initial learning rate to 1e-4 for \\( G^{tg} \\) and 1e-3 for \\( F^{tg} \\). We use the same annealing strategy for the learning rate scheduler as in [12]. As for hyperparameters, we select \\( \\alpha = 0.7 \\), \\( \\beta = 0.5 \\), and \\( \\gamma = 0.85 \\). Moreover, the intermediate model is used for the final inference." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.195, + 0.216, + 0.21 + ], + "angle": 0, + "content": "4.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.219, + 0.471, + 0.657 + ], + "angle": 0, + "content": "Baselines. Four general domain adaptive pose estimation methods CC-SSL [25], MDAM [19], RegDA [12] and UniFrame [16] with SOTA performances are chosen as baselines for comparison. We list these baselines that need to access source data to demonstrate that our proposed source-free method's performance remains competitive when compared to general domain adaptive pose estimation methods. We build source-free baselines by either adapting domain-adaptive HPE methods or source-free DA methods for classification methods. We first introduce two domain adaptive HPE methods. RegDA-SF is based on [12], which originally contains three steps and the first step uses source data. We replace the first step with an exponential moving averaging strategy to obtain the new method. Uniframe-SF is on the foundation of [16]. What we do is remove the style transfer modules between source and target so that it becomes a source-free method. The rest are adapted from source-free DA methods for classification methods. SHOT [21] is a classic source-free domain adaptation method for image classification, and we adapt its mutual information maximization module for HPE by treating the normalized heatmaps as output probabilities in classification. MMT is based on [6], which is a source-free adaptation method that applies two models mutually learning from each other. Here we replace its soft classification loss with the soft consistency loss. SHOT++ [22] is the advanced version of [21], which uses a special unsupervised representation learning technique predicting image rotations. We also add this module to our adopted baseline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Metrics. To evaluate the accuracy of the proposed approach, we adopt the Percentage of Correct Keypoint (PCK) metric, as used in previous work. We set the ratio of correct predictions to be \\(5\\%\\), and report PCK@0.05 in Tables 1-4. In addition to overall keypoint accuracy, we divide joints into several part segments and measure their performance using specific metrics. For the 21-keypoint hand skeleton, we use the Metacarpophalangeal (MCP), Proximal Interphalangeal (PIP), Distal Interphalangeal (DIP), and Finger-tip (Fin) metrics. For the 18-keypoint human skeleton, we select the Shoulder (Sld), Elbow (Elb), Wrist, Hip, Knee, and Ankle metrics. These segment-specific metrics enable a more detailed evaluation of the models' performances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. We present the results of our proposed method on four different tasks: hand tasks RHD \\(\\rightarrow\\) H3D and RHD \\(\\rightarrow\\) FreiHand, and human tasks" + }, + { + "type": "table_caption", + "bbox": [ + 0.567, + 0.088, + 0.825, + 0.1 + ], + "angle": 0, + "content": "Table 1. PCK@0.05 on RHD \\( \\rightarrow \\) H3D Task" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.1, + 0.885, + 0.226 + ], + "angle": 0, + "content": "
MethodSFMCPPIPDIPFinAll
Source-only-67.464.263.354.861.8
CC-SSL [25] (CVPR'20)×81.579.974.464.075.1
MDAM [19] (CVPR'21)×82.379.672.361.574.1
RegDA [12] (CVPR'21)×79.674.471.262.972.5
UniFrame [16] (ECCV'22)×86.784.678.968.179.6
RegDA-SF [12] (CVPR'21)✓71.266.866.258.566.9
UniFrame-SF [16] (ECCV'22)✓84.884.277.068.077.9
SHOT [21] (ICML'20)✓77.569.870.664.273.7
MMT [6] (ICLR'20)✓78.774.172.565.375.3
SHOT++ [22] (TPAMI'22)✓85.385.178.267.678.9
Ours✓88.489.280.971.482.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.554, + 0.239, + 0.838, + 0.251 + ], + "angle": 0, + "content": "Table 2. PCK@0.05 on RHD \\(\\rightarrow\\) FreiHand Task" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.251, + 0.885, + 0.376 + ], + "angle": 0, + "content": "
MethodSFMCPPIPDIPFinAll
Source-only-35.250.154.850.746.8
CC-SSL [25] (CVPR'20)×37.448.250.146.543.8
MDAM [19] (CVPR'21)×32.348.151.747.345.1
RegDA [12] (CVPR'21)×40.955.058.253.151.1
UniFrame [16] (ECCV'22)×43.564.067.462.458.5
RegDA-SF [12] (CVPR'21)✓38.652.957.654.349.5
UniFrame-SF [16] (ECCV'22)✓40.662.561.060.255.7
SHOT [21] (ICML'20)✓40.461.360.558.153.3
MMT [6] (ICLR'20)✓39.660.460.057.852.6
SHOT++ [22] (TPAMI'22)✓41.062.862.759.355.8
Ours✓43.765.966.663.158.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.526, + 0.39, + 0.866, + 0.402 + ], + "angle": 0, + "content": "Table 3. PCK@0.05 on SURREAL \\(\\rightarrow\\) Human3.6M Task" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.402, + 0.885, + 0.503 + ], + "angle": 0, + "content": "
MethodSFSldElbWristHipKneeAnkleAll
Source-only-69.475.466.437.977.377.767.3
CC-SSL [25] (CVPR'20)×44.368.555.222.262.357.851.7
MDAM [19] (CVPR'21)×51.783.168.917.779.476.662.9
RegDA [12] (CVPR'21)×73.386.472.854.882.084.475.6
UniFrame [16] (ECCV'22)×78.189.681.152.685.387.179.0
RegDA-SF [12] (CVPR'21)✓70.682.069.843.379.179.471.5
UniFrame-SF [16] (ECCV'22)✓74.384.872.545.981.585.275.1
SHOT [21] (ICML'20)✓73.083.271.946.779.882.473.3
MMT [6] (ICLR'20)✓73.283.572.445.180.883.973.9
SHOT++ [22] (TPAMI'22)✓75.085.276.745.382.885.075.3
Ours✓77.988.880.452.384.286.978.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.551, + 0.518, + 0.842, + 0.53 + ], + "angle": 0, + "content": "Table 4. PCK@0.05 on SURREAL \\(\\rightarrow\\) LSP Task" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.53, + 0.885, + 0.631 + ], + "angle": 0, + "content": "
MethodSFSldElbWristHipKneeAnkleAll
Source-only-51.565.062.968.068.767.463.9
CC-SSL [25] (CVPR'20)×36.866.363.959.667.370.460.7
MDAM [19] (CVPR'21)×61.477.775.565.876.778.369.2
RegDA [12] (CVPR'21)×62.776.771.181.080.375.374.6
UniFrame [16] (ECCV'22)×69.284.983.385.584.784.382.0
RegDA-SF [12] (CVPR'21)✓54.870.567.665.473.270.066.5
UniFrame-SF [16] (ECCV'22)✓68.480.579.182.780.881.078.8
SHOT [21] (ICML'20)✓63.572.766.578.479.773.272.4
MMT [6] (ICLR'20)✓60.970.970.381.179.372.871.5
SHOT++ [22] (TPAMI'22)✓69.581.780.984.082.379.779.9
Ours✓70.785.483.886.685.285.083.2
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.901 + ], + "angle": 0, + "content": "SURREAL \\(\\rightarrow\\) Human3.6M and SURREAL \\(\\rightarrow\\) LSP. Tables 1 and 2 show the results for the hand tasks, where our method achieves state-of-the-art performance, surpassing the second best baseline SHOT++ with a significant lead of \\(3.3\\%\\) and \\(3.0\\%\\), respectively. Moreover, our method outperforms the SHOT++ model by \\(4.1\\%\\) and \\(3.9\\%\\) on the PIP metric, respectively. For the human tasks, Tables 3 and 4 display the results. Our model achieves a considerable margin of \\(3.4\\%\\) and \\(3.3\\%\\) over the second best baseline SHOT++ for SURREAL \\(\\rightarrow\\) Human3.6M and SURREAL \\(\\rightarrow\\) LSP, respectively. Specifically, our method outperforms SHOT++ by \\(7.0\\%\\) on the Wrist for the adaptation on Human3.6M and \\(5.3\\%\\) on the Ankle for the adaptation on LSP. Furthermore, our approach remains competitive when compared with general domain adaptive HPE methods that need to access source data, as it surpasses UniFrame by \\(2.6\\%\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4832" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.091, + 0.482, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.278, + 0.155, + 0.289 + ], + "angle": 0, + "content": "Source Only" + }, + { + "type": "image_caption", + "bbox": [ + 0.164, + 0.279, + 0.236, + 0.288 + ], + "angle": 0, + "content": "UniFrame-SF" + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.091, + 0.482, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.28, + 0.306, + 0.287 + ], + "angle": 0, + "content": "SHOT++" + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.091, + 0.482, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.28, + 0.372, + 0.287 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.091, + 0.892, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.293, + 0.7, + 0.307 + ], + "angle": 0, + "content": "Figure 6. Qualitative results on H3D dataset (left) and LSP dataset (right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.468, + 0.36 + ], + "angle": 0, + "content": "on RHD \\(\\rightarrow\\) H3D, \\(0.3\\%\\) on RHD \\(\\rightarrow\\) FreiHand, and \\(1.2\\%\\) on SURREAL \\(\\rightarrow\\) LSP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.469, + 0.439 + ], + "angle": 0, + "content": "Qualitative Results. Figs. 6 shows qualitative results of H3D on the left side and LSP on the right side. We use Source only, UniFrame-SF [16], SHOT++ [22], Ours, and Ground Truth for qualitative comparison. It is evident that our method outperforms other baselines significantly." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.45, + 0.348, + 0.466 + ], + "angle": 0, + "content": "4.2. Ablation Study on Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.469, + 0.58 + ], + "angle": 0, + "content": "Our method contains two modules in Step A & Step B as the source-protect module (SP) and the target-relevant module (TR) separately, and here we focus on their functions. Moreover, we use MMT [6] as the baseline, which is referred in Table 1-4. Table 5 and 6 show the ablation study of frameworks on RHD \\(\\rightarrow\\) H3D and SURREAL \\(\\rightarrow\\) Human3.6M with different combinations of these modules." + }, + { + "type": "table_caption", + "bbox": [ + 0.121, + 0.59, + 0.424, + 0.602 + ], + "angle": 0, + "content": "Table 5. Ablation of Frameworks on RHD \\( \\rightarrow \\) H3D" + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.602, + 0.469, + 0.668 + ], + "angle": 0, + "content": "
MethodMCPPIPDIPFinAll
MMT [6]78.774.172.565.375.3
MMT [6] +TR81.377.576.067.779.1
SP+TR (Ours)88.489.280.971.482.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.697, + 0.464, + 0.709 + ], + "angle": 0, + "content": "Table 6. Ablation of Frameworks on SURREAL \\( \\rightarrow \\) Human3.6M" + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.71, + 0.469, + 0.76 + ], + "angle": 0, + "content": "
MethodSldElbWristHipKneeAnkleAll
MMT [6]73.283.572.445.180.883.973.9
MMT [6] +TR74.486.575.748.981.385.176.2
SP+TR (Ours)77.988.880.452.384.286.978.7
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The results clearly demonstrate that both SP and TR contribute to improving the model's performance. Specifically, TR enhances the model's accuracy by \\(3.8\\%\\) on RHD \\(\\rightarrow\\) H3D and \\(2.3\\%\\) on SURREAL \\(\\rightarrow\\) Human3.6M, while SP leads to an improvement of \\(4.1\\%\\) on RHD \\(\\rightarrow\\) H3D and \\(2.5\\%\\) on SURREAL \\(\\rightarrow\\) Human3.6M compared to MMT. Notably, the two proposed modules provide similar levels of improvement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.331, + 0.807, + 0.348 + ], + "angle": 0, + "content": "4.3. Ablation Study on Proposed Losses" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.892, + 0.48 + ], + "angle": 0, + "content": "We performed a detailed ablation study on the three proposed losses, namely \\(\\mathcal{L}_{\\text{res}}\\), \\(\\mathcal{L}_{\\text{cst}}\\), and \\(\\mathcal{L}_{\\text{im}}\\), using the RHD→H3D and SURREAL→Human3.6M tasks. Tables 7 and 8 present the results. As we have previously discussed the functions and advantages of the source-protect and target-relevant modules in Sec. 4.2, we retained \\(\\mathcal{L}_{\\text{ft}}\\) from the source-protect module and \\(\\mathcal{L}_{\\text{con}}\\) from the target-relevant module in the Baseline." + }, + { + "type": "table_caption", + "bbox": [ + 0.561, + 0.499, + 0.831, + 0.511 + ], + "angle": 0, + "content": "Table 7. Ablation of Losses on RHD \\( \\rightarrow \\) H3D" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.51, + 0.892, + 0.607 + ], + "angle": 0, + "content": "
MethodMCPPIPDIPFinAll
Baseline85.386.878.568.179.4
Lres85.587.979.370.780.1
Lcst85.589.078.470.180.9
Lim85.887.278.769.880.3
Lcst&Lim86.988.880.170.381.4
Lres&Lcst&Lim88.489.280.971.482.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.65, + 0.872, + 0.662 + ], + "angle": 0, + "content": "Table 8. Ablation of Losses on SURREAL \\( \\rightarrow \\) Human3.6M" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.662, + 0.892, + 0.737 + ], + "angle": 0, + "content": "
MethodSldElbWristHipKneeAnkleAll
Baseline75.486.176.846.583.085.075.8
Lres76.086.977.547.883.385.276.6
Lcst76.687.478.850.683.586.077.2
Lim76.287.178.349.783.285.776.9
Lcst&Lim77.387.779.251.283.686.478.0
Lres&Lcst&Lim77.988.880.452.384.286.978.7
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We observe that each loss is able to boost the model's performance. Simply applying \\(\\mathcal{L}_{res}\\) leads to a increase of \\(0.7\\%\\) in RHD \\(\\rightarrow\\) H3D and \\(0.8\\%\\) in SURREAL \\(\\rightarrow\\) Human3.6M. \\(\\mathcal{L}_{cst}\\) causes an improvement of \\(1.5\\%\\) in RHD \\(\\rightarrow\\) H3D and \\(1.4\\%\\) in SURREAL \\(\\rightarrow\\) Human3.6M. As for \\(\\mathcal{L}_{im}\\), adding it causes a rise of \\(0.9\\%\\) in RHD \\(\\rightarrow\\) H3D and \\(1.1\\%\\) in SURREAL \\(\\rightarrow\\) Human3.6M. Besides, it is noticed that \\(\\mathcal{L}_{cst}\\) plays a more important role than \\(\\mathcal{L}_{res}\\) or \\(\\mathcal{L}_{im}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "4833" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.353, + 0.108 + ], + "angle": 0, + "content": "4.4. Ablation of Sparsity Reduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.471, + 0.281 + ], + "angle": 0, + "content": "In this section, we investigate the effectiveness of our proposed sparsity reduction strategy. Specifically, we compare heatmap-based methods and vector-based methods for two loss functions: contrastive learning \\((\\mathcal{L}_{cst})\\) and information maximization \\((\\mathcal{L}_{im})\\). We denote heatmap-based contrastive learning and information maximization as HBCL and HBIM, respectively, and vector-based contrastive learning and information maximization as VBCL and VBIM, respectively. Two tasks are selected and the results of these comparisons are presented in Tab. 9 and Tab. 10." + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.292, + 0.445, + 0.304 + ], + "angle": 0, + "content": "Table 9. Ablation of Sparsity Reduction on RHD \\( \\rightarrow \\) H3D" + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.305, + 0.47, + 0.395 + ], + "angle": 0, + "content": "
MethodMCPPIPDIPFinAll
MMT [6]78.774.172.565.375.3
MMT [6] + HBCL77.473.070.765.574.2
MMT [6] + VBCL81.776.674.467.978.7
MMT [6] + HBIM79.374.772.965.875.9
MMT [6] + VBIM80.376.173.666.577.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.424, + 0.468, + 0.45 + ], + "angle": 0, + "content": "Table 10. Ablation of Sparsity Reduction on SURREAL \\(\\rightarrow\\) Human3.6M" + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.45, + 0.468, + 0.519 + ], + "angle": 0, + "content": "
MethodSldElbWristHipKneeAnkleAll
MMT [6]73.283.572.445.180.883.973.9
MMT [6] + HBCL71.782.071.844.680.282.473.0
MMT [6] + VBCL75.687.075.149.283.386.376.7
MMT [6] + HBIM73.984.672.746.081.183.074.3
MMT [6] + VBIM75.386.474.348.682.485.776.0
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.47, + 0.721 + ], + "angle": 0, + "content": "Based on the results, it is evident that HBCL does not improve the model's performance significantly, whereas VBCL provides significant assistance. For instance, MMT+HBCL shows a decrease of \\(1.1\\%\\) on the RHD \\(\\rightarrow\\) H3D task, while MMT+VBCL exhibits an increase of \\(3.4\\%\\). Additionally, while HBIM can enhance the model's performance, it falls short of surpassing VBIM. For example, MMT+HBIM shows an increase of \\(0.6\\%\\) on the RHD \\(\\rightarrow\\) H3D task, whereas MMT+VBIM achieves an increase of \\(2.2\\%\\). In summary, it becomes evident that utilizing the vector-based spatial probability space is crucial for achieving optimal results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.732, + 0.265, + 0.749 + ], + "angle": 0, + "content": "4.5. Parameter Analysis" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.767, + 0.468, + 0.869 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.899 + ], + "angle": 0, + "content": "Figure 7. Parameter Analysis on RHD \\(\\rightarrow\\) H3D (best viewed in color). a: Analysis on \\(\\alpha\\). b: Analysis on \\(\\beta\\). c: Analysis on \\(\\gamma\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "We use RHD \\(\\rightarrow\\) H3D to illustrate the sensitivity of \\(\\alpha\\) in Equation 5 and \\(\\beta\\) and \\(\\gamma\\) in Equation 11. The results are shown in Fig. 7. From it, we observe that \\(\\alpha = 0.7\\), \\(\\beta = 0.5\\), and \\(\\gamma = 0.85\\) is the best choice. Moreover, it is noticed that the decreases from different parameters are limited to \\(0.2\\%\\) for \\(\\alpha\\), \\(\\beta\\) and \\(\\gamma\\). Therefore, our method shows stable performance over hyperparameters." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.218, + 0.892, + 0.251 + ], + "angle": 0, + "content": "5. Qualitative Results on the Unseen Dataset COCO" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.277, + 0.892, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.426, + 0.892, + 0.468 + ], + "angle": 0, + "content": "Figure 8. Qualitative results on the COCO dataset. Here we compare our method trained from SURREAL \\(\\rightarrow\\) LSP with the source-only model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.892, + 0.702 + ], + "angle": 0, + "content": "In this section, we showcase qualitative results on the COCO dataset [23], which was not utilized in our previous human dataset studies. It's important to note that our 14-keypoint poses are annotated differently from the 17-keypoint poses in COCO, which hinders providing quantitative results on COCO using its own metrics or APIs. Despite this limitation, we believe that our qualitative findings hold significant value. For our analysis, we leverage the model trained on the SURREAL \\(\\rightarrow\\) LSP task in conjunction with the pre-trained SURREAL model. The results, presented in Fig. 8, demonstrate the model's capacity to generalize effectively to unseen datasets, implying that the adaptation process boosts the model's generalization ability beyond its original training dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.722, + 0.619, + 0.737 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a new task named source-free domain adaptive human pose estimation that places an emphasis on the privacy of source data. Additionally, we propose a new framework that includes source-protect and target-relevant modules, which aim to alleviate the issues of catastrophic forgetting of source and the sparsity of spatial distributions, respectively. Our approach is evaluated on hand and human pose datasets through extensive experiments, demonstrating that it outperforms state-of-the-art methods by a considerable margin." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4834" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 295-305, June 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.47, + 0.227 + ], + "angle": 0, + "content": "[2] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.47, + 0.297 + ], + "angle": 0, + "content": "[3] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7103-7112, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.47, + 0.367 + ], + "angle": 0, + "content": "[4] Wanxia Deng, Qing Liao, Lingjun Zhao, Deke Guo, Gangyao Kuang, Dewen Hu, and Li Liu. Joint clustering and discriminative feature alignment for unsupervised domain adaptation. IEEE Transactions on Image Processing, pages 7842-7855, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.47, + 0.41 + ], + "angle": 0, + "content": "[5] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International Conference on Machine Learning, pages 1180-1189, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.47, + 0.466 + ], + "angle": 0, + "content": "[6] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual meanteaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In International Conference on Learning Representations, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.47, + 0.522 + ], + "angle": 0, + "content": "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.47, + 0.577 + ], + "angle": 0, + "content": "[8] Zhongyi Han, Haoliang Sun, and Yilong Yin. Learning transferable parameters for unsupervised domain adaptation. IEEE Transactions on Image Processing, 31:6424-6439, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.47, + 0.634 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.635, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[10] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 1501-1510, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[11] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[12] Junguang Jiang, Yifei Ji, Ximei Wang, Yufeng Liu, Jianmin Wang, and Mingsheng Long. Regressive domain adaptation for unsupervised keypoint detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6780-6789, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.873 + ], + "angle": 0, + "content": "[13] Ruijie Jiang, Thuan Nguyen, Prakash Ishwar, and Shuchin Aeron. Supervised contrastive learning with hard negative samples. arXiv preprint arXiv:2209.00078, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Rui Jin, Jing Zhang, Jianyu Yang, and Dacheng Tao. Multibranch adversarial regression for domain adaptive hand" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "pose estimation. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):6125-6136, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[15] Sam Johnson and Mark Everingham. Clustered pose and nonlinear appearance models for human pose estimation. In British Machine Vision Conference, volume 2, page 5. Cite-seer, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[16] Donghyun Kim, Kaihong Wang, Kate Saenko, Margrit Betke, and Stan Sclaroff. A unified framework for domain adaptive pose estimation. arXiv preprint arXiv:2204.00172, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[18] Vinod K Kurmi, Venkatesh K Subramanian, Vinay P Namboodiri, , and . Domain impression: A source data free domain adaptation method. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 615-625, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.405 + ], + "angle": 0, + "content": "[19] Chen Li and Gim Hee Lee. From synthetic to real: Unsupervised domain adaptation for animal pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1482-1491, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.407, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[20] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9638-9647, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[21] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International Conference on Machine Learning, pages 6028-6039. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[22] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8602-8617, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, pages 740-755. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[24] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International Conference on Machine Learning, pages 97-105, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[25] Jiteng Mu, Weichao Qiu, Gregory D Hager, and Alan L Yuille. Learning from synthetic animals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12386-12395, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[26] Qucheng Peng, Zhengming Ding, Lingjuan Lyu, Lichao Sun, and Chen Chen. Toward better target representation for source-free and black-box domain adaptation. arXiv preprint arXiv:2208.10531, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Ekkasit Pinyoanuntapong, Ayman Ali, Kalvik Jakkala, Pu Wang, Minwoo Lee, Qucheng Peng, Chen Chen, and Zhi Sun. Gaitsada: Self-aligned domain adaptation for mmwave gait recognition. arXiv preprint arXiv:2301.13384, 2023." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4835" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[28] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3723-3732, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.22 + ], + "angle": 0, + "content": "[29] Dahu Shi, Xing Wei, Liangqi Li, Ye Ren, and Wenming Tan. End-to-end multi-person pose estimation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11069-11078, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.29 + ], + "angle": 0, + "content": "[30] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5693-5703, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.348 + ], + "angle": 0, + "content": "[31] Hui Tang, Ke Chen, and Kui Jia. Unsupervised domain adaptation via structurally regularized deep clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8725-8735, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.47, + 0.404 + ], + "angle": 0, + "content": "[32] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in Neural Information Processing Systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.47, + 0.473 + ], + "angle": 0, + "content": "[33] Jonathan Tompson, Ross Goroshin, Arjun Jain, Yann LeCun, and Christoph Bregler. Efficient object localization using convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 648-656, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.47, + 0.532 + ], + "angle": 0, + "content": "[34] Jonathan J. Thompson, Arjun Jain, Yann LeCun, and Christoph Bregler. Joint training of a convolutional network and a graphical model for human pose estimation. Advances in neural information processing systems, 27, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.47, + 0.575 + ], + "angle": 0, + "content": "[35] Eric Tzeng, Judy Hoffman, Ning Zhang, Kate Saenko, and Trevor Darrell. Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:1412.3474, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.47, + 0.646 + ], + "angle": 0, + "content": "[36] Gul Varol, Javier Romero, Xavier Martin, Naureen Mahmood, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning from synthetic humans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 109-117, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.47, + 0.717 + ], + "angle": 0, + "content": "[37] Haixin Wang, Lu Zhou, Yingying Chen, Ming Tang, and Jinqiao Wang. Regularizing vector embedding in bottom-up human pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI, pages 107-122. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[38] Haifeng Xia, Handong Zhao, and Zhengming Ding. Adaptive adversarial network for source-free domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9010-9019, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[39] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In Proceedings of the European Conference on Computer Vision (ECCV), pages 466-481, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[40] Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, and Ce Zhu. Distribution-aware coordinate representation for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7093-7102, 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[41] Zhengyi Zhao, Tianyao Wang, Siyu Xia, and Yangang Wang. Hand-3d-studio: A new multi-view system for 3d hand reconstruction. In ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2478-2482, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.219 + ], + "angle": 0, + "content": "[42] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In Proceedings of the IEEE International Conference on Computer Vision, pages 4903-4911, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.893, + 0.29 + ], + "angle": 0, + "content": "[43] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 813-822, 2019." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.956 + ], + "angle": 0, + "content": "4836" + } + ] +] \ No newline at end of file diff --git a/2023/Source-free Domain Adaptive Human Pose Estimation/6db7e049-3b2d-4942-adce-fb8455dd18b2_origin.pdf b/2023/Source-free Domain Adaptive Human Pose Estimation/6db7e049-3b2d-4942-adce-fb8455dd18b2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b2989dcbcf0c68a5963692e6dc3c482f626975ff --- /dev/null +++ b/2023/Source-free Domain Adaptive Human Pose Estimation/6db7e049-3b2d-4942-adce-fb8455dd18b2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4eb611967e0695bdf5c38feb2b23679513b05f50004a1bce38180f7a03ee03af +size 3320919 diff --git a/2023/Source-free Domain Adaptive Human Pose Estimation/full.md b/2023/Source-free Domain Adaptive Human Pose Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0a3e9b5a78b40feb1fb75f8cca0d3051e1eae8f2 --- /dev/null +++ b/2023/Source-free Domain Adaptive Human Pose Estimation/full.md @@ -0,0 +1,342 @@ +# Source-free Domain Adaptive Human Pose Estimation + +Qucheng Peng, Ce Zheng, Chen Chen +Center for Research in Computer Vision, University of Central Florida + +{qucheng.peng,ce.zheng}@ucf.edu，chen.chen@crcv.ucf.edu + +# Abstract + +Human Pose Estimation (HPE) is widely used in various fields, including motion analysis, healthcare, and virtual reality. However, the great expenses of labeled real-world datasets present a significant challenge for HPE. To overcome this, one approach is to train HPE models on synthetic datasets and then perform domain adaptation (DA) on real-world data. Unfortunately, existing DA methods for HPE neglect data privacy and security by using both source and target data in the adaptation process. + +To this end, we propose a new task, named source-free domain adaptive HPE, which aims to address the challenges of cross-domain learning of HPE without access to source data during the adaptation process. We further propose a novel framework that consists of three models: source model, intermediate model, and target model, which explores the task from both source-protect and target-relevant perspectives. The source-protect module preserves source information more effectively while resisting noise, and the target-relevant module reduces the sparsity of spatial representations by building a novel spatial probability space, and pose-specific contrastive learning and information maximization are proposed on the basis of this space. Comprehensive experiments on several domain adaptive HPE benchmarks show that the proposed method outperforms existing approaches by a considerable margin. The codes are available at https://github.com/davidpengucf/SFDAHPE. + +# 1. Introduction + +The accuracy of 2D human pose estimation (HPE) from monocular images has been significantly improved through the use of deep neural networks. However, obtaining sufficient labeled data for training such models is a laborious and time-consuming task, particularly for real-world datasets that require diverse appearances and poses with ground truth annotations. In contrast, the development of computer graphics and game engines has facilitated the generation of annotated synthetic datasets that can generate a + +virtually unlimited number of labeled samples. Nevertheless, models trained on synthetic data may not perform well on real-world data due to the domain gap or differences in data distribution. To address this challenge, domain adaptation (DA) is essential to enable synthetic data-trained models to generalize to real-world data. + +Recently, source-free domain adaptation has gained significant attention due to its practicality in real-world scenarios, where using labeled source data during the adaptation is expensive or not feasible due to data privacy and security concerns. However, existing source-free domain adaptation methods in classification tasks may not be directly applicable to more complex tasks, such as 2D human pose estimation (HPE). This is because HPE requires estimating the spatial locations of different body joints, instead of just the object's category as classification does, making it more challenging to adapt from source to target domains. Therefore, there is a pressing need to develop effective source-free domain adaptation techniques specifically tailored to HPE tasks. This motivates us to propose a new task named source-free domain adaptive HPE in this paper. + +The task of source-free domain adaptive HPE presents challenges from both the source-free DA and pose estimation perspectives. Specifically, in traditional DA tasks, source data can participate in the adaptation process, allowing for a focus on reducing domain shifts between source and target. However, in source-free DA, only the pretrained source model can be used during adaptation, leading to catastrophic forgetting of the source domain. Directly applying the source pretrained model can also introduce noise due to domain shifts. Therefore, balancing the absorption of the source model's knowledge with the resistance of the source model's noise is critical. Current methods tend to adopt a progressive replacement strategy to alleviate the runoff of source information. For example, [21, 22] use the previous model to supervise ongoing adaptation, thereby evoking the source-related memory, while [32, 6] use the exponential moving average (EMA) strategy to slow down the forgetting of the source. Although these methods alleviate the negative impact of noise successfully, they modify the overall parameters of the pretrained source model + +significantly, leading to the disappearance of source-related knowledge and representations during the adaptation process. Therefore, it is necessary to develop source-protect modules that can protect source information more effectively while resisting noise at the same time. + +Moreover, we outline another challenge of source-free domain adaptive HPE: the sparsity of keypoints (i.e. key body joints) in an image. When it comes to domain adaptation (DA) in classification, achieving distribution alignment to reduce domain gap is a straightforward process as the outputs can be represented in a probability space that includes all the categories. However, the same cannot be said for Human Pose Estimation (HPE) based on a complete image, as each keypoint has a specific distance from others based on bone lengths, and the number of keypoints $K$ (e.g., 14 for human pose and 21 for hand pose) is significantly smaller than the number of pixels in an image (e.g., $512 \times 512 \approx 0.26\mathrm{M}$ ), making it a challenging task. Although heatmaps [34, 33] are commonly used to predict keypoints (heatmap representation-based 2D HPE is considered as the mainstream approach for HPE) while reducing dimensionality in HPE, the number of pixels inside the heatmap is still considerable (e.g. $128 \times 128 \approx 16\mathrm{K}$ ) in comparison to $K$ . Therefore, building a spatial probability space for the keypoints based on either the entire image or the heatmap is not ideal. [12] proposes constructing a low-dimensional distribution based solely on the $K$ keypoints, effectively transforming the HPE task into a $K$ -classification task. However, this spatial distribution relies on clear decision boundaries between distinctive keypoints, which is not reasonable in HPE. In other words, all samples can be classified to a certain category in classification, but most pixels inside an image (e.g. more than $97\%$ if the fault tolerance is decided by the metric PCK@0.05) can not be assigned to a certain keypoint in HPE, but mispredictions may occur at these unpresented locations. Hence, it is necessary to develop a novel spatial probability space that reduces sparsity while providing complete representations of the prediction space. + +To address the aforementioned task-specific challenges, we propose a novel framework that incorporates source-free domain adaptive HPE from both source-protect and target-relevant perspectives. Our framework consists of three models: the source model, the intermediate model, and the target model. Source model preserves source information, and target model absorbs knowledge from target data, while intermediate model interacts with both of them to reduce the domain gap between source and target implicitly, thereby used for the final inference. In the source-protect modules, we conduct knowledge transfer between the source model and the intermediate model to prevent catastrophic forgetting of the source-related information. We transfer knowledge from the source model to the intermediate model while keeping the source model's feature extractor fixed. To re + +sist noise, we propose a new residual loss and finetune the source model's regressor. In the target-relevant modules, we conduct executions between the intermediate model and the target model to mitigate domain shifts between the source and target domains. We build a new spatial probability space for HPE to alleviate the sparsity problem. On the basis of this space, we propose contrastive learning and information maximization designed specifically for human pose estimation. Our contributions are summarized as follows: + +- We construct a new task named source-free domain adaptive human pose estimation, which focuses on the cross-domain learning of HPE without access to source data during the adaptation process. +- We propose a new framework consisting of three models: source model, intermediate model, and target model, which explores the task from both source-protect and target-relevant perspectives. +- We conduct comprehensive experiments on several domain adaptive HPE benchmarks and the results show that our method outperforms the existing approaches by a considerable margin. + +# 2. Related Work + +Human Pose Estimation. The heatmap-based method [29, 37] is currently a widely used technique in 2D HPE, which involves generating heatmaps by placing 2D Gaussian kernels on potential keypoints, followed by converting these heatmaps to coordinates. For instance, [3] uses global- and refine-nets, while [39] simplifies the network structures through the Simple Baselines technique. Some approaches have focused on utilizing features at different scales [30], while others aim to improve the quality of heatmaps by extending Gaussians to second order [40] or using coupled embeddings for heatmap regression [37]. In this study, we adopt the same heatmap-based HPE methods as other domain adaptive HPE works [12, 16]. + +General Domain Adaptation. There are two main categories of methods for achieving general domain adaptation, which requires access to both the source and target data during adaptation. The first category is metric-based methods that explicitly measure the discrepancy between the source and target domains. For example, [24] uses maximum mean discrepancy, and [35] applies deep domain confusion. Recent work, such as [4], jointly performs clustering and discrimination for alignment using contrastive learning [2, 13]. The second category is GAN-based [7] methods that build a min-max game for two players related to the source and target domains. For example, [5, 28] adopts domain confusion to confuse the two players, while [31, 27] shows that discriminative clustering on the target benefits adaptation. However, access to source data is required in the general domain adaptation setting, which raises concerns about data privacy and security. + +Source-Free Domain Adaptation. Source-Free has become the mainstream paradigm for alleviating concerns about data privacy and security in domain adaptation, as shown in the work of [20, 21, 26]. There are two technique routes under the source-free setting: self-supervision and virtual source transfer. For the self-supervised methods, [21] is the most representative, which introduces information maximization to assist adaptation. Other self-supervised methods include [38], which treats the problem from a discriminative perspective, and [1, 26], which proposes online pseudo label refinement. In virtual source methods, most build GANs to generate virtual source data. Examples include conditional GANs used in [18], and collaborative GANs in [20], which achieve better generations. However, most source-free DA methods for classification are not applicable to HPE tasks due to the sparsity of keypoints' distributions. + +Domain Adaptive Human Pose Estimation. There are two main categories of methods used for domain-adaptive HPE. The first category involves a shared structure, where the weights of the networks from both the source pretrain and target adaptation are shared. CC-SSL [25] employs a single end-to-end trained network, RegDA [12] utilizes one shared feature extractor and two separate regressors, and TransPar [8] emphasizes transferable parameters using a similar structure. The second category is the unshared structure, where the pretrain and adaptation create a teacher-student paradigm. In this category, MDAM [19] addresses this structure alongside a novel pseudo-label strategy, UniFrame [16] modifies the classic Mean-Teacher [32] model by combining it with style transfer [10], and MarsDA [14] applies a teacher-student approach to edit RegDA. However, all these existing domain adaptive HPE approaches are not source-free, which raises concerns about data privacy and security. Therefore, it is necessary to consider source-free domain adaptive HPE. To our knowledge, we are the first work to study this challenging problem. + +# 3. Method + +# 3.1. Preliminary of 2D HPE + +In 2D HPE, we are given a labeled pose dataset $\mathcal{D} = (x_i,y_i)$ , where $x_{i}\in \mathbb{R}^{C\times H\times W}$ is an image and $y_{i}\in \mathbb{R}^{K\times 2}$ are the corresponding keypoint coordinates. Here, $H$ and $W$ represent the height and width of the image, respectively, and $C$ is the number of channels. Additionally, $K$ represents the number of keypoints. To ensure smoother training, most existing 2D methods represent the keypoint coordinates as heatmaps $HM_{i}\in \mathbb{R}^{K\times H^{\prime}\times W^{\prime}}$ , where $H^{\prime}$ and $W^{\prime}$ are the height and width of the heatmaps, respectively. These heatmaps are used to represent the coordinates in a spatial way. The process of how the model works is defined + +![](images/99393e82b2670ff4be632a31185e86033b4e3ced3a013c36bcad381f13d55625.jpg) +Figure 1. Process of heatmap-based 2D HPE. The model includes two components: the feature extractor $G$ and the regressor $F$ . After feeding the input image to the model, we obtain $K$ output heatmaps corresponding to $K$ keypoints. + +in Fig. 1. Ground truth heatmaps are generated by a 2D Gaussian centered at the ground truth joint location, with one heatmap for each keypoint. We denote the transform from coordinates to heatmaps as $T: \mathbb{R}^{K \times 2} \to \mathbb{R}^{K \times H' \times W'}$ . The training processes are conducted between the output heatmaps and ground truth heatmaps using MSE for supervised learning. In the inference stage, well-trained supervised models output heatmaps that predict the probability of joints occurring at each pixel. For each heatmap, the pixel with the highest probability is considered to be the predicted location of the joint. + +# 3.2. Problem Statement + +In the source-free domain adaptive HPE setting, we have a source domain dataset $\mathcal{S} = \{(x_i^s, y_i^s)\}_{i=1}^{n_s}$ with $n_s$ labeled pose samples. $x_i^s \in \mathbb{R}^{C \times H \times W}$ is the source image and $y_i^s \in \mathbb{R}^{K \times 2}$ is the corresponding pose annotation. Besides, there exists a target domain dataset $\mathcal{T} = \{x_i^t\}_{i=1}^{n_t}$ that includes $n_t$ unlabeled pose samples. The source and target domains share the same label space but lie in different distributions (e.g. synthetic vs. real). The training procedure is split into two stages: the pretrain stage and the adaptation stage. In the pretrain stage, the labeled source domain dataset $\mathcal{S}$ is used to train the source model $f^{sr}$ . However, what distinguishes source-free domain adaptive HPE from general domain adaptive HPE is the adaptation stage, where only the source model $f^{sr}$ and target domain dataset $\mathcal{T}$ are available for training, and the source domain dataset cannot be used anymore. Our objective is to develop a model that performs well on the target dataset. + +# 3.3. Pipeline of Our Method + +# 3.3.1 Source Model Pretrain + +For DA in HPE, labeled source data are applied for pretraining. Same as most existing domain adaptive pose estimation methods [25, 19, 12, 16], the network is composed of a feature extractor $G$ (such as ResNet backbone) and a pose regressor $F$ , so the source model can be represented as $f^{sr} = F^{sr}(G^{sr}(\cdot))$ . Based on the supervised heatmap-based loss (MSE loss), we propose the overall objective for + +![](images/c960532eded9a22c3b979ce012a59d9294c563557223b317552fadd1d1649de8.jpg) +Figure 2. Overall framework of our method. It includes three models, each with one feature extractor and one regressor. The adaptation contains two steps: Step A and Step B. Step A is applied for source-protect adaptation, while Step B aims at target-relevant adaptation. The solid arrows show the forward process of associated losses, and the dotted arrows show the backward updating process of associated losses. EMA means the exponential moving average process from target model to intermediate model to update intermediate model's weights. + +![](images/417679ac09e08dce32c594ba5d8369484112f2f1450c11c82be088228c035e47.jpg) + +source pretraining: + +$$ +\mathcal {L} _ {\text {p r e t r a i n}} = \mathbb {E} _ {\left(x _ {i} ^ {s}, y _ {i} ^ {s}\right) \in \mathcal {S}} \mathcal {L} _ {\text {m s e}} \left(F ^ {s r} \left(G ^ {s r} \left(x _ {i} ^ {s}\right)\right), T \left(y _ {i} ^ {s}\right)\right), \tag {1} +$$ + +where $\mathcal{L}_{mse}$ is the Mean Squared Error (MSE) loss between the heatmap representations of prediction and ground truth, and $T$ is the transform from coordinates to heatmaps defined in Sec. 3.1. + +# 3.3.2 Adaptation Framework + +As Fig. 2 shows, our method contains three models: source model $f^{sr} = F^{sr}(G^{sr}(\cdot))$ , intermediate model $f^{in} = F^{in}(G^{in}(\cdot))$ , and target model $f^{tg} = F^{tg}(G^{tg}(\cdot))$ . Each model includes one feature extractor and one regressor. Assume the heatmaps generated by source model as $H_{i}^{sr} = F^{sr}(G^{sr}(x_{i}^{t})) = [h_{i,1}^{sr}, h_{i,2}^{sr}, \dots, h_{i,K}^{sr}]^{\top}$ , and those from intermediate model and target model respectively as $H_{i}^{in} = F^{in}(G^{in}(x_{i}^{t})) = [h_{i,1}^{in}, h_{i,2}^{in}, \dots, h_{i,K}^{in}]^{\top}$ and $H_{i}^{tg} = F^{tg}(G^{tg}(x_{i}^{t})) = [h_{i,1}^{tg}, h_{i,2}^{tg}, \dots, h_{i,K}^{tg}]^{\top}$ , where $K$ is the number of keypoints. We also assume the keypoint indexing space as $\mathcal{P} = [1,2,\dots,K]$ . The source model is used to maintain source information, while the target model is applied to learn target knowledge. The intermediate model is the transition model between these two models. It serves as a bridge between source and target as it learns from source and target simultaneously, thus eliminating the domain gap implicitly. Therefore, it outperforms both the source and target models and is used for the final inference. When starting the adaptation, $f^{in}$ and $f^{tg}$ are initialized from the source model $f^{sr}$ . + +For a single adaptation iteration, two steps are executed as shown in Fig. 2. Step A is for the source-protect adaptation, and Step B aims to adapt in a target-relevant way. + +In Step A, our goal is to overcome catastrophic forgetting on source domain, while eliminating the noise from the source model due to domain shift. Therefore, the source model's feature extractor $G^{sr}$ is fixed to retain source knowledge, while its regressor $F^{sr}$ keeps updating to improve regression accuracy, hence reducing noises in the outputs. On the opposite, the intermediate model's regressor $F^{in}$ is fixed, but its feature extractor $G^{in}$ continues to update. That's because we want the feature extractor to learn more source knowledge on the representation level, while the noise from the source model due to domain shifts can be inhibited. + +In Step B, we aim to conduct the adaptation between the intermediate model and the target model in a target-relevant manner. During the process, several pose-specific techniques are utilized to minimize the discrepancy between these two models. Though both $f^{in}$ and $f^{tg}$ are updated at this step, $f^{tg}$ is updated via back-propagation while $f^{in}$ is via exponential moving average from $f^{tg}$ , since averaging model weights over training steps tends to produce a more accurate model than using the final weights directly [32]. + +In the following two subsections, we introduce the detailed techniques used for Step A and Step B. + +![](images/7494d1f6f42d043899528e8a9b02de3dfb7e3ece876fe587967bc2611903c160.jpg) +Figure 3. Proposed residual loss. Here we choose two heatmaps at the same location generated from source model and intermediate model for illustration. For simplicity, here the heatmap's scale is set as $3 \times 3$ , while the actual size is $64 \times 64$ . + +# 3.3.3 Source-protect Modules + +In this part, we propose modules related to source-free-specific adaptation. The biggest challenge is to distill knowledge from the source model and resist the potential noise due to domain shift. First, we introduce the finetune loss that calibrates on the source model: + +$$ +\mathcal {L} _ {f t} = \mathbb {E} _ {x _ {i} ^ {t} \in \mathcal {T}} \mathcal {L} _ {m s e} \left(F ^ {s r} \left(G ^ {s r} \left(x _ {i} ^ {t}\right)\right), F ^ {i n} \left(G ^ {i n} \left(x _ {i} ^ {t}\right)\right)\right). \tag {2} +$$ + +To be notified, we hope the regression results of the source model can be more accurate, while the source knowledge can be retained. Thus, the feature extractor is fixed so that the source representations continue to be generated, and the regressor is being updated: + +$$ +\theta_ {F ^ {s r}} ^ {t + 1} = \theta_ {F ^ {s r}} ^ {t} - \lambda^ {s r} \frac {\partial L _ {f t}}{\partial \theta_ {F ^ {s r}} ^ {t}}, \tag {3} +$$ + +where $\theta_F^s$ is the weights of the source model's regressor $F^s$ , while $t$ denotes the training step and $\lambda^s$ is the source model's learning rate. However, improving source model is not enough. What's more, we propose the residual loss to improve intermediate model. + +The application of Equation 2 facilitates mutual knowledge transfer between the source and intermediate models. Apart from the pixel with the highest confidence in a heatmap, the other pixels also contain valuable information that benefits the knowledge transfer, and that's why we propose residual loss in Fig. 3. From the example in the figure, the most confident pixel (confidence $= 1.0$ ) in the source heatmap is removed, along with the pixel at the same location which has a confidence of 0.8 in the intermediate heatmap marked with red. The pair marked with blue follows the same rule but on the basis of intermediate heatmap. Then we build adaptive residual heatmaps denoted as $\hat{h}_{i,j}^{sr}$ and $\hat{h}_{i,j}^{in}$ . Moreover, we employ KL divergence to ensure that $\hat{h}_{i,j}^{in}$ approaches $\hat{h}_{i,j}^{sr}$ , thereby preserving the source in + +![](images/18126c75a76efac6603c7be223e6f0e5486647a6d6f1b8a459a89950007e3a71.jpg) +Figure 4. Illustration of the projection of a heatmap. In order to reduce the sparsity of a heatmap, we project it into a horizontal vector and a vertical vector. For simplicity, here the heatmap's scale is set as $4 \times 4$ and the vector's scale as $4 \times 1$ , while the actual heatmap's size is $64 \times 64$ and the vector's size is $64 \times 1$ . + +formation. Based on these, we deduce the residual loss as: + +$$ +\mathcal {L} _ {r e s} = \mathbb {E} _ {x _ {i} ^ {t} \in \mathcal {T}} \mathbb {E} _ {j \in \mathcal {P}} \mathrm {D} _ {\mathrm {K L}} \left(\sigma \left(\hat {h} _ {i, j} ^ {s r} / \tau\right) \mid \mid \sigma \left(\hat {h} _ {i, j} ^ {i n} / \tau\right)\right), \tag {4} +$$ + +where $\mathrm{D}_{\mathrm{KL}}$ is the KL divergence, and $\sigma (\cdot)$ is the softmax function. Besides, $\tau$ is the temperature used to scale the residual heatmaps and is empirically set to 0.3. By combining the finetune loss and the residual loss, we get the objective function with the optimization process for updating the intermediate model: + +$$ +\mathcal {L} _ {i n} = \mathcal {L} _ {f t} + \alpha \mathcal {L} _ {r e s}, \quad \theta_ {G ^ {i n}} ^ {t + 1} = \theta_ {G ^ {i n}} ^ {t} - \lambda^ {i n} \frac {\partial L _ {i n}}{\partial \theta_ {G ^ {i n}} ^ {t}}, \tag {5} +$$ + +where $t$ is the training step, and $\lambda^{in}$ is the learning rate for $G^{in}$ . Moreover, $\alpha$ is the trade-off hyperparameter between these two losses. + +# 3.3.4 Target-relevant Modules + +This section introduces a module that is associated with target information. As previously mentioned, the sparsity of keypoints in the image presents a significant challenge in our task. To address this issue, we propose using projected horizontal and vertical vectors of heatmaps as spatial representations, as shown in Fig. 4. This approach reduces the sparsity of the spatial probability space effectively since the number of overall digits inside the heatmap is $64 \times 64 \approx 4K$ , while the number of units in each vector is 64. Therefore, we can effectively reduce the sparsity of the spatial probability space. With 64 units in each vector, the spatial probability space size is acceptable compared to the number of keypoints. + +Despite the reduction in sparsity, the completeness of spatial representations is not affected. That is because all the coordinates can always be divided into two representations, horizontally and vertically. In such a case, our spatial representation is better than the $K$ -classification probability space proposed by [12], which only includes the keypoint-relevant information and does not take the units in the heatmap that are not keypoints into consideration. + +![](images/92ecc105b5653a9cc0a04a1838ecbb0834757d4eb0bdd568f766d052fb9cae0c.jpg) +Figure 5. Illustration of building positive and negative keypoint pairs for contrastive learning between the output heatmaps of the intermediate model and target model. + +In Fig. 4, we introduce the projection of a heatmap from both horizontal and vertical directions. For the heatmap $h_{i,j}^{in}$ generated by the intermediate model from $i$ th sample's $j$ th keypoint, its projection can be represented as $proj( h_{i,j}^{in}) = (v_{i,jx}^{in}, v_{i,jy}^{in})$ , where $v_{i,jx}^{in}$ is the horizontal vector and $v_{i,jy}^{in}$ is the vertical vector. This vector pair is the foundation of the following modules and losses that we propose. + +First, we propose the pose-specific contrastive loss between the intermediate model and target model. According to the observation from [12], wrong predictions are usually located at other keypoints' positions. In such a case, improving regression can be achieved by minimizing the discrepancy of heatmaps at the same location and maximizing the discrepancy of heatmaps at different locations. And that's how we define the positive and negative pair in Fig. 5. Denote the similarity of two heatmaps generated by the intermediate model and target model respectively as: + +$$ +s i m \left(h _ {i, j} ^ {i n}, h _ {i, k} ^ {t g}\right) = \frac {1}{2} \left(\frac {v _ {i , j x} ^ {i n} \cdot v _ {i , k x} ^ {t g}}{| | v _ {i , j x} ^ {i n} | | | | | v _ {i , k x} ^ {t g} | |} + \frac {v _ {i , j y} ^ {i n} \cdot v _ {i , k y} ^ {t g}}{| | v _ {i , j y} ^ {i n} | | | | | v _ {i , k y} ^ {t g} | |}\right). \tag {6} +$$ + +On the basis of the similarity, we define the overall pose-specific contrastive loss as: + +$$ +\mathcal {L} _ {c s t} = - \mathbb {E} _ {x _ {i} ^ {t} \in \mathcal {T}} \mathbb {E} _ {j \in \mathcal {P}} \log \frac {\exp (s i m (h _ {i , j} ^ {i n} , h _ {i , j} ^ {t g}))}{\Sigma_ {k \in \mathcal {P}} \exp (s i m (h _ {i , j} ^ {i n} , h _ {i , k} ^ {t g}))}. \tag {7} +$$ + +Next, we propose a pose-specific self-supervised technique called pose-specific information maximization. Information maximization is very helpful to source-free domain adaptation for classification [21] because it makes the target outputs individually certain and globally diverse, hence mitigating the domain gap implicitly. However, the improvements are limited due to the sparsity of heatmaps. With the help of projected vectors, we refine the information maximization loss as: + +$$ +\mathcal {L} _ {i m} = \mathbb {E} _ {x _ {i} ^ {t} \in \mathcal {T}} \left(\mathcal {L} _ {e n t x} \left(x _ {i} ^ {t}\right) + \mathcal {L} _ {e n t y} \left(x _ {i} ^ {t}\right) - \mathcal {L} _ {d i v} \left(x _ {i} ^ {t}\right)\right), \tag {8} +$$ + +$$ +\left\{ \begin{array}{l} \mathcal {L} _ {e n t x} \left(x _ {i} ^ {t}\right) = - \mathbb {E} _ {j \in \mathcal {P}} \left[ \sigma \left(v _ {i, j x} ^ {t g}\right) \log \sigma \left(v _ {i, j x} ^ {t g}\right) \right], \\ \mathcal {L} _ {e n t y} \left(x _ {i} ^ {t}\right) = - \mathbb {E} _ {j \in \mathcal {P}} \left[ \sigma \left(v _ {i, j y} ^ {t g}\right) \log \sigma \left(v _ {i, j y} ^ {t g}\right) \right], \\ \mathcal {L} _ {d i v} \left(x _ {i} ^ {t}\right) = - \left[ \mathbb {E} _ {j \in \mathcal {P}} \sigma \left(h _ {i, j} ^ {t g}\right) \right] \log \left[ \mathbb {E} _ {j \in \mathcal {P}} \sigma \left(h _ {i, j} ^ {t g}\right) \right]. \end{array} \right. \tag {9} +$$ + +Here $\mathcal{L}_{entx}$ is the entropy of horizontal vectors, and $\mathcal{L}_{enty}$ is the entropy of vertical vectors. The last term in Equation 8 is the diversity-promoting objective. Moreover, we introduce the consistency loss between the two models' outputs: + +$$ +\mathcal {L} _ {c o n} = \mathbb {E} _ {\left(x _ {i} ^ {s}, y _ {i} ^ {s}\right) \in \mathcal {S}} \mathcal {L} _ {m s e} \left(F ^ {i n} \left(G ^ {i n} \left(x _ {i} ^ {s}\right)\right), F ^ {t g} \left(G ^ {t g} \left(x _ {i} ^ {s}\right)\right)\right). \tag {10} +$$ + +By summarizing above, we conclude the overall objective loss function for the target model as: + +$$ +\mathcal {L} _ {t g t} = \mathcal {L} _ {c o n} + \beta \mathcal {L} _ {c s t} + \gamma \mathcal {L} _ {i m}, \tag {11} +$$ + +where $\beta$ and $\gamma$ are trade-off parameters. + +Moreover, the intermediate model is updated via the exponential moving average strategy (EMA) in this step: + +$$ +\theta_ {f ^ {i n}} ^ {t + 1} = \eta \theta_ {f ^ {i n}} ^ {t} + (1 - \eta) \theta_ {f ^ {t g}} ^ {t}, \tag {12} +$$ + +where $t$ denotes the step of training and $\eta$ denotes the smoothing coefficient set to 0.999 by default. + +# 4. Experiments + +Datasets. We use three human pose datasets and three hand pose datasets to validate our approach. SURREAL [36] is utilized as the source dataset and encompasses six million synthetic human pose images. Human3.6M [11], one of the target human pose datasets, is a frequently used real-world dataset in the community, consisting of 3.6 million images distributed into seven folds. According to previous research [12, 16], S1, S5, S6, S7, and S8 are designated as the training set, whereas S9 and S11 serve as the testing set. The other target dataset is Leeds Sports Pose [15] (LSP), which is a real-world dataset with 2,000 images, and we use all of them for adaptation. We focus on two domain adaptation tasks: SURREAL $\rightarrow$ Human3.6M and SURREAL $\rightarrow$ LSP for human pose tasks. + +Rendered Hand Pose Dataset [42] (RHD) is the source hand dataset with 43,986 synthetic hand images, of which 41,258 are allocated for training, while the remaining 2,728 images are utilized for validation. Hand-3D-Studio [41] (H3D) is one of the two target hand datasets, including 22,000 real-world frames. For training, we employed 18,800 frames and reserved the rest for testing, following the protocol outlined in previous research papers [12, 16]. The other target hand dataset, FreiHand [43], provides 130k real-world images, and we employed all of them for our adaptation task. Our focus is on two domain adaptation tasks: RHD $\rightarrow$ H3D and RHD $\rightarrow$ FreiHand. + +Implementation Details. We adopt Simple Baseline [39] with ResNet-101 [9] as the HPE backbone, following previous works [12, 16]. In the pretrain process, we conduct 40 epochs, each with 500 iterations, using the Adam [17] optimizer with an initial learning rate of 1e-4. The learning rate decreases to 1e-5 at 25 epochs. For the adaptation process, we execute 40 epochs with 30,000 iterations. In step + +A, we choose an initial learning rate of 1e-4 for $F^{sr}$ and 1e-5 for $G^{in}$ . Step B sets the initial learning rate to 1e-4 for $G^{tg}$ and 1e-3 for $F^{tg}$ . We use the same annealing strategy for the learning rate scheduler as in [12]. As for hyperparameters, we select $\alpha = 0.7$ , $\beta = 0.5$ , and $\gamma = 0.85$ . Moreover, the intermediate model is used for the final inference. + +# 4.1. Main Results + +Baselines. Four general domain adaptive pose estimation methods CC-SSL [25], MDAM [19], RegDA [12] and UniFrame [16] with SOTA performances are chosen as baselines for comparison. We list these baselines that need to access source data to demonstrate that our proposed source-free method's performance remains competitive when compared to general domain adaptive pose estimation methods. We build source-free baselines by either adapting domain-adaptive HPE methods or source-free DA methods for classification methods. We first introduce two domain adaptive HPE methods. RegDA-SF is based on [12], which originally contains three steps and the first step uses source data. We replace the first step with an exponential moving averaging strategy to obtain the new method. Uniframe-SF is on the foundation of [16]. What we do is remove the style transfer modules between source and target so that it becomes a source-free method. The rest are adapted from source-free DA methods for classification methods. SHOT [21] is a classic source-free domain adaptation method for image classification, and we adapt its mutual information maximization module for HPE by treating the normalized heatmaps as output probabilities in classification. MMT is based on [6], which is a source-free adaptation method that applies two models mutually learning from each other. Here we replace its soft classification loss with the soft consistency loss. SHOT++ [22] is the advanced version of [21], which uses a special unsupervised representation learning technique predicting image rotations. We also add this module to our adopted baseline. + +Metrics. To evaluate the accuracy of the proposed approach, we adopt the Percentage of Correct Keypoint (PCK) metric, as used in previous work. We set the ratio of correct predictions to be $5\%$ , and report PCK@0.05 in Tables 1-4. In addition to overall keypoint accuracy, we divide joints into several part segments and measure their performance using specific metrics. For the 21-keypoint hand skeleton, we use the Metacarpophalangeal (MCP), Proximal Interphalangeal (PIP), Distal Interphalangeal (DIP), and Finger-tip (Fin) metrics. For the 18-keypoint human skeleton, we select the Shoulder (Sld), Elbow (Elb), Wrist, Hip, Knee, and Ankle metrics. These segment-specific metrics enable a more detailed evaluation of the models' performances. + +Quantitative Results. We present the results of our proposed method on four different tasks: hand tasks RHD $\rightarrow$ H3D and RHD $\rightarrow$ FreiHand, and human tasks + +Table 1. PCK@0.05 on RHD $\rightarrow$ H3D Task + +
MethodSFMCPPIPDIPFinAll
Source-only-67.464.263.354.861.8
CC-SSL [25] (CVPR'20)×81.579.974.464.075.1
MDAM [19] (CVPR'21)×82.379.672.361.574.1
RegDA [12] (CVPR'21)×79.674.471.262.972.5
UniFrame [16] (ECCV'22)×86.784.678.968.179.6
RegDA-SF [12] (CVPR'21)✓71.266.866.258.566.9
UniFrame-SF [16] (ECCV'22)✓84.884.277.068.077.9
SHOT [21] (ICML'20)✓77.569.870.664.273.7
MMT [6] (ICLR'20)✓78.774.172.565.375.3
SHOT++ [22] (TPAMI'22)✓85.385.178.267.678.9
Ours✓88.489.280.971.482.2
+ +Table 2. PCK@0.05 on RHD $\rightarrow$ FreiHand Task + +
MethodSFMCPPIPDIPFinAll
Source-only-35.250.154.850.746.8
CC-SSL [25] (CVPR'20)×37.448.250.146.543.8
MDAM [19] (CVPR'21)×32.348.151.747.345.1
RegDA [12] (CVPR'21)×40.955.058.253.151.1
UniFrame [16] (ECCV'22)×43.564.067.462.458.5
RegDA-SF [12] (CVPR'21)✓38.652.957.654.349.5
UniFrame-SF [16] (ECCV'22)✓40.662.561.060.255.7
SHOT [21] (ICML'20)✓40.461.360.558.153.3
MMT [6] (ICLR'20)✓39.660.460.057.852.6
SHOT++ [22] (TPAMI'22)✓41.062.862.759.355.8
Ours✓43.765.966.663.158.8
+ +Table 3. PCK@0.05 on SURREAL $\rightarrow$ Human3.6M Task + +
MethodSFSldElbWristHipKneeAnkleAll
Source-only-69.475.466.437.977.377.767.3
CC-SSL [25] (CVPR'20)×44.368.555.222.262.357.851.7
MDAM [19] (CVPR'21)×51.783.168.917.779.476.662.9
RegDA [12] (CVPR'21)×73.386.472.854.882.084.475.6
UniFrame [16] (ECCV'22)×78.189.681.152.685.387.179.0
RegDA-SF [12] (CVPR'21)✓70.682.069.843.379.179.471.5
UniFrame-SF [16] (ECCV'22)✓74.384.872.545.981.585.275.1
SHOT [21] (ICML'20)✓73.083.271.946.779.882.473.3
MMT [6] (ICLR'20)✓73.283.572.445.180.883.973.9
SHOT++ [22] (TPAMI'22)✓75.085.276.745.382.885.075.3
Ours✓77.988.880.452.384.286.978.7
+ +Table 4. PCK@0.05 on SURREAL $\rightarrow$ LSP Task + +
MethodSFSldElbWristHipKneeAnkleAll
Source-only-51.565.062.968.068.767.463.9
CC-SSL [25] (CVPR'20)×36.866.363.959.667.370.460.7
MDAM [19] (CVPR'21)×61.477.775.565.876.778.369.2
RegDA [12] (CVPR'21)×62.776.771.181.080.375.374.6
UniFrame [16] (ECCV'22)×69.284.983.385.584.784.382.0
RegDA-SF [12] (CVPR'21)✓54.870.567.665.473.270.066.5
UniFrame-SF [16] (ECCV'22)✓68.480.579.182.780.881.078.8
SHOT [21] (ICML'20)✓63.572.766.578.479.773.272.4
MMT [6] (ICLR'20)✓60.970.970.381.179.372.871.5
SHOT++ [22] (TPAMI'22)✓69.581.780.984.082.379.779.9
Ours✓70.785.483.886.685.285.083.2
+ +SURREAL $\rightarrow$ Human3.6M and SURREAL $\rightarrow$ LSP. Tables 1 and 2 show the results for the hand tasks, where our method achieves state-of-the-art performance, surpassing the second best baseline SHOT++ with a significant lead of $3.3\%$ and $3.0\%$ , respectively. Moreover, our method outperforms the SHOT++ model by $4.1\%$ and $3.9\%$ on the PIP metric, respectively. For the human tasks, Tables 3 and 4 display the results. Our model achieves a considerable margin of $3.4\%$ and $3.3\%$ over the second best baseline SHOT++ for SURREAL $\rightarrow$ Human3.6M and SURREAL $\rightarrow$ LSP, respectively. Specifically, our method outperforms SHOT++ by $7.0\%$ on the Wrist for the adaptation on Human3.6M and $5.3\%$ on the Ankle for the adaptation on LSP. Furthermore, our approach remains competitive when compared with general domain adaptive HPE methods that need to access source data, as it surpasses UniFrame by $2.6\%$ + +![](images/abaf841689c52e53750234650482f5f71a190ce65115d68e21a368fe06924024.jpg) +Source Only +UniFrame-SF +SHOT++ +Ours + +![](images/f7a0788a9588aeda9fde1278823e96c31285fcb4d9ccb89ada1d47ef9174530c.jpg) +Figure 6. Qualitative results on H3D dataset (left) and LSP dataset (right). + +on RHD $\rightarrow$ H3D, $0.3\%$ on RHD $\rightarrow$ FreiHand, and $1.2\%$ on SURREAL $\rightarrow$ LSP. + +Qualitative Results. Figs. 6 shows qualitative results of H3D on the left side and LSP on the right side. We use Source only, UniFrame-SF [16], SHOT++ [22], Ours, and Ground Truth for qualitative comparison. It is evident that our method outperforms other baselines significantly. + +# 4.2. Ablation Study on Framework + +Our method contains two modules in Step A & Step B as the source-protect module (SP) and the target-relevant module (TR) separately, and here we focus on their functions. Moreover, we use MMT [6] as the baseline, which is referred in Table 1-4. Table 5 and 6 show the ablation study of frameworks on RHD $\rightarrow$ H3D and SURREAL $\rightarrow$ Human3.6M with different combinations of these modules. + +Table 5. Ablation of Frameworks on RHD $\rightarrow$ H3D + +
MethodMCPPIPDIPFinAll
MMT [6]78.774.172.565.375.3
MMT [6] +TR81.377.576.067.779.1
SP+TR (Ours)88.489.280.971.482.2
+ +Table 6. Ablation of Frameworks on SURREAL $\rightarrow$ Human3.6M + +
MethodSldElbWristHipKneeAnkleAll
MMT [6]73.283.572.445.180.883.973.9
MMT [6] +TR74.486.575.748.981.385.176.2
SP+TR (Ours)77.988.880.452.384.286.978.7
+ +The results clearly demonstrate that both SP and TR contribute to improving the model's performance. Specifically, TR enhances the model's accuracy by $3.8\%$ on RHD $\rightarrow$ H3D and $2.3\%$ on SURREAL $\rightarrow$ Human3.6M, while SP leads to an improvement of $4.1\%$ on RHD $\rightarrow$ H3D and $2.5\%$ on SURREAL $\rightarrow$ Human3.6M compared to MMT. Notably, the two proposed modules provide similar levels of improvement. + +# 4.3. Ablation Study on Proposed Losses + +We performed a detailed ablation study on the three proposed losses, namely $\mathcal{L}_{\text{res}}$ , $\mathcal{L}_{\text{cst}}$ , and $\mathcal{L}_{\text{im}}$ , using the RHD→H3D and SURREAL→Human3.6M tasks. Tables 7 and 8 present the results. As we have previously discussed the functions and advantages of the source-protect and target-relevant modules in Sec. 4.2, we retained $\mathcal{L}_{\text{ft}}$ from the source-protect module and $\mathcal{L}_{\text{con}}$ from the target-relevant module in the Baseline. + +Table 7. Ablation of Losses on RHD $\rightarrow$ H3D + +
MethodMCPPIPDIPFinAll
Baseline85.386.878.568.179.4
Lres85.587.979.370.780.1
Lcst85.589.078.470.180.9
Lim85.887.278.769.880.3
Lcst&Lim86.988.880.170.381.4
Lres&Lcst&Lim88.489.280.971.482.2
+ +Table 8. Ablation of Losses on SURREAL $\rightarrow$ Human3.6M + +
MethodSldElbWristHipKneeAnkleAll
Baseline75.486.176.846.583.085.075.8
Lres76.086.977.547.883.385.276.6
Lcst76.687.478.850.683.586.077.2
Lim76.287.178.349.783.285.776.9
Lcst&Lim77.387.779.251.283.686.478.0
Lres&Lcst&Lim77.988.880.452.384.286.978.7
+ +We observe that each loss is able to boost the model's performance. Simply applying $\mathcal{L}_{res}$ leads to a increase of $0.7\%$ in RHD $\rightarrow$ H3D and $0.8\%$ in SURREAL $\rightarrow$ Human3.6M. $\mathcal{L}_{cst}$ causes an improvement of $1.5\%$ in RHD $\rightarrow$ H3D and $1.4\%$ in SURREAL $\rightarrow$ Human3.6M. As for $\mathcal{L}_{im}$ , adding it causes a rise of $0.9\%$ in RHD $\rightarrow$ H3D and $1.1\%$ in SURREAL $\rightarrow$ Human3.6M. Besides, it is noticed that $\mathcal{L}_{cst}$ plays a more important role than $\mathcal{L}_{res}$ or $\mathcal{L}_{im}$ . + +# 4.4. Ablation of Sparsity Reduction + +In this section, we investigate the effectiveness of our proposed sparsity reduction strategy. Specifically, we compare heatmap-based methods and vector-based methods for two loss functions: contrastive learning $(\mathcal{L}_{cst})$ and information maximization $(\mathcal{L}_{im})$ . We denote heatmap-based contrastive learning and information maximization as HBCL and HBIM, respectively, and vector-based contrastive learning and information maximization as VBCL and VBIM, respectively. Two tasks are selected and the results of these comparisons are presented in Tab. 9 and Tab. 10. + +Table 9. Ablation of Sparsity Reduction on RHD $\rightarrow$ H3D + +
MethodMCPPIPDIPFinAll
MMT [6]78.774.172.565.375.3
MMT [6] + HBCL77.473.070.765.574.2
MMT [6] + VBCL81.776.674.467.978.7
MMT [6] + HBIM79.374.772.965.875.9
MMT [6] + VBIM80.376.173.666.577.5
+ +Table 10. Ablation of Sparsity Reduction on SURREAL $\rightarrow$ Human3.6M + +
MethodSldElbWristHipKneeAnkleAll
MMT [6]73.283.572.445.180.883.973.9
MMT [6] + HBCL71.782.071.844.680.282.473.0
MMT [6] + VBCL75.687.075.149.283.386.376.7
MMT [6] + HBIM73.984.672.746.081.183.074.3
MMT [6] + VBIM75.386.474.348.682.485.776.0
+ +Based on the results, it is evident that HBCL does not improve the model's performance significantly, whereas VBCL provides significant assistance. For instance, MMT+HBCL shows a decrease of $1.1\%$ on the RHD $\rightarrow$ H3D task, while MMT+VBCL exhibits an increase of $3.4\%$ . Additionally, while HBIM can enhance the model's performance, it falls short of surpassing VBIM. For example, MMT+HBIM shows an increase of $0.6\%$ on the RHD $\rightarrow$ H3D task, whereas MMT+VBIM achieves an increase of $2.2\%$ . In summary, it becomes evident that utilizing the vector-based spatial probability space is crucial for achieving optimal results. + +# 4.5. Parameter Analysis + +![](images/6c32538c30e8e130cc9a8da2626a57ea8de41f1cd745fad740f5e9ac387b5ede.jpg) +Figure 7. Parameter Analysis on RHD $\rightarrow$ H3D (best viewed in color). a: Analysis on $\alpha$ . b: Analysis on $\beta$ . c: Analysis on $\gamma$ . + +We use RHD $\rightarrow$ H3D to illustrate the sensitivity of $\alpha$ in Equation 5 and $\beta$ and $\gamma$ in Equation 11. The results are shown in Fig. 7. From it, we observe that $\alpha = 0.7$ , $\beta = 0.5$ , and $\gamma = 0.85$ is the best choice. Moreover, it is noticed that the decreases from different parameters are limited to $0.2\%$ for $\alpha$ , $\beta$ and $\gamma$ . Therefore, our method shows stable performance over hyperparameters. + +# 5. Qualitative Results on the Unseen Dataset COCO + +![](images/649bf5e302635bb29057003cbc74098a117511111957a00fef641b083dcc4fc3.jpg) +Figure 8. Qualitative results on the COCO dataset. Here we compare our method trained from SURREAL $\rightarrow$ LSP with the source-only model. + +In this section, we showcase qualitative results on the COCO dataset [23], which was not utilized in our previous human dataset studies. It's important to note that our 14-keypoint poses are annotated differently from the 17-keypoint poses in COCO, which hinders providing quantitative results on COCO using its own metrics or APIs. Despite this limitation, we believe that our qualitative findings hold significant value. For our analysis, we leverage the model trained on the SURREAL $\rightarrow$ LSP task in conjunction with the pre-trained SURREAL model. The results, presented in Fig. 8, demonstrate the model's capacity to generalize effectively to unseen datasets, implying that the adaptation process boosts the model's generalization ability beyond its original training dataset. + +# 6. Conclusion + +In this paper, we propose a new task named source-free domain adaptive human pose estimation that places an emphasis on the privacy of source data. Additionally, we propose a new framework that includes source-protect and target-relevant modules, which aim to alleviate the issues of catastrophic forgetting of source and the sparsity of spatial distributions, respectively. Our approach is evaluated on hand and human pose datasets through extensive experiments, demonstrating that it outperforms state-of-the-art methods by a considerable margin. + +# References + +[1] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 295-305, June 2022. +[2] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. +[3] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7103-7112, 2018. +[4] Wanxia Deng, Qing Liao, Lingjun Zhao, Deke Guo, Gangyao Kuang, Dewen Hu, and Li Liu. Joint clustering and discriminative feature alignment for unsupervised domain adaptation. IEEE Transactions on Image Processing, pages 7842-7855, 2021. +[5] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International Conference on Machine Learning, pages 1180-1189, 2015. +[6] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual meanteaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In International Conference on Learning Representations, 2020. +[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. +[8] Zhongyi Han, Haoliang Sun, and Yilong Yin. Learning transferable parameters for unsupervised domain adaptation. IEEE Transactions on Image Processing, 31:6424-6439, 2022. +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. +[10] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 1501-1510, 2017. +[11] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, 2013. +[12] Junguang Jiang, Yifei Ji, Ximei Wang, Yufeng Liu, Jianmin Wang, and Mingsheng Long. Regressive domain adaptation for unsupervised keypoint detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6780-6789, 2021. +[13] Ruijie Jiang, Thuan Nguyen, Prakash Ishwar, and Shuchin Aeron. Supervised contrastive learning with hard negative samples. arXiv preprint arXiv:2209.00078, 2022. +[14] Rui Jin, Jing Zhang, Jianyu Yang, and Dacheng Tao. Multibranch adversarial regression for domain adaptive hand + +pose estimation. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):6125-6136, 2022. +[15] Sam Johnson and Mark Everingham. Clustered pose and nonlinear appearance models for human pose estimation. In British Machine Vision Conference, volume 2, page 5. Cite-seer, 2010. +[16] Donghyun Kim, Kaihong Wang, Kate Saenko, Margrit Betke, and Stan Sclaroff. A unified framework for domain adaptive pose estimation. arXiv preprint arXiv:2204.00172, 2022. +[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. +[18] Vinod K Kurmi, Venkatesh K Subramanian, Vinay P Namboodiri, , and . Domain impression: A source data free domain adaptation method. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 615-625, 2021. +[19] Chen Li and Gim Hee Lee. From synthetic to real: Unsupervised domain adaptation for animal pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1482-1491, 2021. +[20] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9638-9647, 2020. +[21] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International Conference on Machine Learning, pages 6028-6039. PMLR, 2020. +[22] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8602-8617, 2022. +[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, pages 740-755. Springer, 2014. +[24] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International Conference on Machine Learning, pages 97-105, 2015. +[25] Jiteng Mu, Weichao Qiu, Gregory D Hager, and Alan L Yuille. Learning from synthetic animals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12386-12395, 2020. +[26] Qucheng Peng, Zhengming Ding, Lingjuan Lyu, Lichao Sun, and Chen Chen. Toward better target representation for source-free and black-box domain adaptation. arXiv preprint arXiv:2208.10531, 2022. +[27] Ekkasit Pinyoanuntapong, Ayman Ali, Kalvik Jakkala, Pu Wang, Minwoo Lee, Qucheng Peng, Chen Chen, and Zhi Sun. Gaitsada: Self-aligned domain adaptation for mmwave gait recognition. arXiv preprint arXiv:2301.13384, 2023. + +[28] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3723-3732, 2018. +[29] Dahu Shi, Xing Wei, Liangqi Li, Ye Ren, and Wenming Tan. End-to-end multi-person pose estimation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11069-11078, 2022. +[30] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5693-5703, 2019. +[31] Hui Tang, Ke Chen, and Kui Jia. Unsupervised domain adaptation via structurally regularized deep clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8725-8735, 2020. +[32] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in Neural Information Processing Systems, 30, 2017. +[33] Jonathan Tompson, Ross Goroshin, Arjun Jain, Yann LeCun, and Christoph Bregler. Efficient object localization using convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 648-656, 2015. +[34] Jonathan J. Thompson, Arjun Jain, Yann LeCun, and Christoph Bregler. Joint training of a convolutional network and a graphical model for human pose estimation. Advances in neural information processing systems, 27, 2014. +[35] Eric Tzeng, Judy Hoffman, Ning Zhang, Kate Saenko, and Trevor Darrell. Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:1412.3474, 2014. +[36] Gul Varol, Javier Romero, Xavier Martin, Naureen Mahmood, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning from synthetic humans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 109-117, 2017. +[37] Haixin Wang, Lu Zhou, Yingying Chen, Ming Tang, and Jinqiao Wang. Regularizing vector embedding in bottom-up human pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI, pages 107-122. Springer, 2022. +[38] Haifeng Xia, Handong Zhao, and Zhengming Ding. Adaptive adversarial network for source-free domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9010-9019, 2021. +[39] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In Proceedings of the European Conference on Computer Vision (ECCV), pages 466-481, 2018. +[40] Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, and Ce Zhu. Distribution-aware coordinate representation for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7093-7102, 2020. + +[41] Zhengyi Zhao, Tianyao Wang, Siyu Xia, and Yangang Wang. Hand-3d-studio: A new multi-view system for 3d hand reconstruction. In ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2478-2482, 2020. +[42] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In Proceedings of the IEEE International Conference on Computer Vision, pages 4903-4911, 2017. +[43] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 813-822, 2019. \ No newline at end of file diff --git a/2023/Source-free Domain Adaptive Human Pose Estimation/images.zip b/2023/Source-free Domain Adaptive Human Pose Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2ad1fdb05fea19736de6473d545f543d48c5956f --- /dev/null +++ b/2023/Source-free Domain Adaptive Human Pose Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:899fad6f5a31815fc8be5a805a620bf23c89e847cd7056ca8d1d082f65ce4adf +size 836093 diff --git a/2023/Source-free Domain Adaptive Human Pose Estimation/layout.json b/2023/Source-free Domain Adaptive Human Pose Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..20cd27bcba379333793155f9fbeea935cbddd307 --- /dev/null +++ b/2023/Source-free Domain Adaptive Human Pose Estimation/layout.json @@ -0,0 +1,10942 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 128, + 103, + 465, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 103, + 465, + 120 + ], + "spans": [ + { + "bbox": [ + 128, + 103, + 465, + 120 + ], + "type": "text", + "content": "Source-free Domain Adaptive Human Pose Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 144, + 467, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 144, + 467, + 171 + ], + "spans": [ + { + "bbox": [ + 126, + 144, + 467, + 171 + ], + "type": "text", + "content": "Qucheng Peng, Ce Zheng, Chen Chen \nCenter for Research in Computer Vision, University of Central Florida" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 148, + 173, + 445, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 173, + 445, + 185 + ], + "spans": [ + { + "bbox": [ + 148, + 173, + 445, + 185 + ], + "type": "text", + "content": "{qucheng.peng,ce.zheng}@ucf.edu，chen.chen@crcv.ucf.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 238, + 288, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 238, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 45, + 238, + 288, + 346 + ], + "type": "text", + "content": "Human Pose Estimation (HPE) is widely used in various fields, including motion analysis, healthcare, and virtual reality. However, the great expenses of labeled real-world datasets present a significant challenge for HPE. To overcome this, one approach is to train HPE models on synthetic datasets and then perform domain adaptation (DA) on real-world data. Unfortunately, existing DA methods for HPE neglect data privacy and security by using both source and target data in the adaptation process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 346, + 288, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 346, + 288, + 562 + ], + "spans": [ + { + "bbox": [ + 45, + 346, + 288, + 562 + ], + "type": "text", + "content": "To this end, we propose a new task, named source-free domain adaptive HPE, which aims to address the challenges of cross-domain learning of HPE without access to source data during the adaptation process. We further propose a novel framework that consists of three models: source model, intermediate model, and target model, which explores the task from both source-protect and target-relevant perspectives. The source-protect module preserves source information more effectively while resisting noise, and the target-relevant module reduces the sparsity of spatial representations by building a novel spatial probability space, and pose-specific contrastive learning and information maximization are proposed on the basis of this space. Comprehensive experiments on several domain adaptive HPE benchmarks show that the proposed method outperforms existing approaches by a considerable margin. The codes are available at https://github.com/davidpengucf/SFDAHPE." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 585, + 128, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 128, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 128, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "The accuracy of 2D human pose estimation (HPE) from monocular images has been significantly improved through the use of deep neural networks. However, obtaining sufficient labeled data for training such models is a laborious and time-consuming task, particularly for real-world datasets that require diverse appearances and poses with ground truth annotations. In contrast, the development of computer graphics and game engines has facilitated the generation of annotated synthetic datasets that can generate a" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 213, + 547, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 286 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 286 + ], + "type": "text", + "content": "virtually unlimited number of labeled samples. Nevertheless, models trained on synthetic data may not perform well on real-world data due to the domain gap or differences in data distribution. To address this challenge, domain adaptation (DA) is essential to enable synthetic data-trained models to generalize to real-world data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 290, + 547, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 547, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 547, + 471 + ], + "type": "text", + "content": "Recently, source-free domain adaptation has gained significant attention due to its practicality in real-world scenarios, where using labeled source data during the adaptation is expensive or not feasible due to data privacy and security concerns. However, existing source-free domain adaptation methods in classification tasks may not be directly applicable to more complex tasks, such as 2D human pose estimation (HPE). This is because HPE requires estimating the spatial locations of different body joints, instead of just the object's category as classification does, making it more challenging to adapt from source to target domains. Therefore, there is a pressing need to develop effective source-free domain adaptation techniques specifically tailored to HPE tasks. This motivates us to propose a new task named source-free domain adaptive HPE in this paper." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "content": "The task of source-free domain adaptive HPE presents challenges from both the source-free DA and pose estimation perspectives. Specifically, in traditional DA tasks, source data can participate in the adaptation process, allowing for a focus on reducing domain shifts between source and target. However, in source-free DA, only the pretrained source model can be used during adaptation, leading to catastrophic forgetting of the source domain. Directly applying the source pretrained model can also introduce noise due to domain shifts. Therefore, balancing the absorption of the source model's knowledge with the resistance of the source model's noise is critical. Current methods tend to adopt a progressive replacement strategy to alleviate the runoff of source information. For example, [21, 22] use the previous model to supervise ongoing adaptation, thereby evoking the source-related memory, while [32, 6] use the exponential moving average (EMA) strategy to slow down the forgetting of the source. Although these methods alleviate the negative impact of noise successfully, they modify the overall parameters of the pretrained source model" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4826" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 131 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 131 + ], + "type": "text", + "content": "significantly, leading to the disappearance of source-related knowledge and representations during the adaptation process. Therefore, it is necessary to develop source-protect modules that can protect source information more effectively while resisting noise at the same time." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 136, + 286, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 136, + 286, + 530 + ], + "spans": [ + { + "bbox": [ + 49, + 136, + 286, + 530 + ], + "type": "text", + "content": "Moreover, we outline another challenge of source-free domain adaptive HPE: the sparsity of keypoints (i.e. key body joints) in an image. When it comes to domain adaptation (DA) in classification, achieving distribution alignment to reduce domain gap is a straightforward process as the outputs can be represented in a probability space that includes all the categories. However, the same cannot be said for Human Pose Estimation (HPE) based on a complete image, as each keypoint has a specific distance from others based on bone lengths, and the number of keypoints " + }, + { + "bbox": [ + 49, + 136, + 286, + 530 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 49, + 136, + 286, + 530 + ], + "type": "text", + "content": " (e.g., 14 for human pose and 21 for hand pose) is significantly smaller than the number of pixels in an image (e.g., " + }, + { + "bbox": [ + 49, + 136, + 286, + 530 + ], + "type": "inline_equation", + "content": "512 \\times 512 \\approx 0.26\\mathrm{M}" + }, + { + "bbox": [ + 49, + 136, + 286, + 530 + ], + "type": "text", + "content": "), making it a challenging task. Although heatmaps [34, 33] are commonly used to predict keypoints (heatmap representation-based 2D HPE is considered as the mainstream approach for HPE) while reducing dimensionality in HPE, the number of pixels inside the heatmap is still considerable (e.g. " + }, + { + "bbox": [ + 49, + 136, + 286, + 530 + ], + "type": "inline_equation", + "content": "128 \\times 128 \\approx 16\\mathrm{K}" + }, + { + "bbox": [ + 49, + 136, + 286, + 530 + ], + "type": "text", + "content": ") in comparison to " + }, + { + "bbox": [ + 49, + 136, + 286, + 530 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 49, + 136, + 286, + 530 + ], + "type": "text", + "content": ". Therefore, building a spatial probability space for the keypoints based on either the entire image or the heatmap is not ideal. [12] proposes constructing a low-dimensional distribution based solely on the " + }, + { + "bbox": [ + 49, + 136, + 286, + 530 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 49, + 136, + 286, + 530 + ], + "type": "text", + "content": " keypoints, effectively transforming the HPE task into a " + }, + { + "bbox": [ + 49, + 136, + 286, + 530 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 49, + 136, + 286, + 530 + ], + "type": "text", + "content": "-classification task. However, this spatial distribution relies on clear decision boundaries between distinctive keypoints, which is not reasonable in HPE. In other words, all samples can be classified to a certain category in classification, but most pixels inside an image (e.g. more than " + }, + { + "bbox": [ + 49, + 136, + 286, + 530 + ], + "type": "inline_equation", + "content": "97\\%" + }, + { + "bbox": [ + 49, + 136, + 286, + 530 + ], + "type": "text", + "content": " if the fault tolerance is decided by the metric PCK@0.05) can not be assigned to a certain keypoint in HPE, but mispredictions may occur at these unpresented locations. Hence, it is necessary to develop a novel spatial probability space that reduces sparsity while providing complete representations of the prediction space." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 534, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 534, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 534, + 286, + 712 + ], + "type": "text", + "content": "To address the aforementioned task-specific challenges, we propose a novel framework that incorporates source-free domain adaptive HPE from both source-protect and target-relevant perspectives. Our framework consists of three models: the source model, the intermediate model, and the target model. Source model preserves source information, and target model absorbs knowledge from target data, while intermediate model interacts with both of them to reduce the domain gap between source and target implicitly, thereby used for the final inference. In the source-protect modules, we conduct knowledge transfer between the source model and the intermediate model to prevent catastrophic forgetting of the source-related information. We transfer knowledge from the source model to the intermediate model while keeping the source model's feature extractor fixed. To re" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 308, + 72, + 544, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 178 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 178 + ], + "type": "text", + "content": "sist noise, we propose a new residual loss and finetune the source model's regressor. In the target-relevant modules, we conduct executions between the intermediate model and the target model to mitigate domain shifts between the source and target domains. We build a new spatial probability space for HPE to alleviate the sparsity problem. On the basis of this space, we propose contrastive learning and information maximization designed specifically for human pose estimation. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 307, + 180, + 544, + 323 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 307, + 180, + 544, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 180, + 544, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 180, + 544, + 227 + ], + "type": "text", + "content": "- We construct a new task named source-free domain adaptive human pose estimation, which focuses on the cross-domain learning of HPE without access to source data during the adaptation process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 307, + 228, + 544, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 544, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 544, + 275 + ], + "type": "text", + "content": "- We propose a new framework consisting of three models: source model, intermediate model, and target model, which explores the task from both source-protect and target-relevant perspectives." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 307, + 276, + 544, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 544, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 544, + 323 + ], + "type": "text", + "content": "- We conduct comprehensive experiments on several domain adaptive HPE benchmarks and the results show that our method outperforms the existing approaches by a considerable margin." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 308, + 335, + 391, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 335, + 391, + 346 + ], + "spans": [ + { + "bbox": [ + 308, + 335, + 391, + 346 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 354, + 544, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 544, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 544, + 510 + ], + "type": "text", + "content": "Human Pose Estimation. The heatmap-based method [29, 37] is currently a widely used technique in 2D HPE, which involves generating heatmaps by placing 2D Gaussian kernels on potential keypoints, followed by converting these heatmaps to coordinates. For instance, [3] uses global- and refine-nets, while [39] simplifies the network structures through the Simple Baselines technique. Some approaches have focused on utilizing features at different scales [30], while others aim to improve the quality of heatmaps by extending Gaussians to second order [40] or using coupled embeddings for heatmap regression [37]. In this study, we adopt the same heatmap-based HPE methods as other domain adaptive HPE works [12, 16]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 511, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 544, + 712 + ], + "type": "text", + "content": "General Domain Adaptation. There are two main categories of methods for achieving general domain adaptation, which requires access to both the source and target data during adaptation. The first category is metric-based methods that explicitly measure the discrepancy between the source and target domains. For example, [24] uses maximum mean discrepancy, and [35] applies deep domain confusion. Recent work, such as [4], jointly performs clustering and discrimination for alignment using contrastive learning [2, 13]. The second category is GAN-based [7] methods that build a min-max game for two players related to the source and target domains. For example, [5, 28] adopts domain confusion to confuse the two players, while [31, 27] shows that discriminative clustering on the target benefits adaptation. However, access to source data is required in the general domain adaptation setting, which raises concerns about data privacy and security." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "4827" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": "Source-Free Domain Adaptation. Source-Free has become the mainstream paradigm for alleviating concerns about data privacy and security in domain adaptation, as shown in the work of [20, 21, 26]. There are two technique routes under the source-free setting: self-supervision and virtual source transfer. For the self-supervised methods, [21] is the most representative, which introduces information maximization to assist adaptation. Other self-supervised methods include [38], which treats the problem from a discriminative perspective, and [1, 26], which proposes online pseudo label refinement. In virtual source methods, most build GANs to generate virtual source data. Examples include conditional GANs used in [18], and collaborative GANs in [20], which achieve better generations. However, most source-free DA methods for classification are not applicable to HPE tasks due to the sparsity of keypoints' distributions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 277, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 289, + 529 + ], + "type": "text", + "content": "Domain Adaptive Human Pose Estimation. There are two main categories of methods used for domain-adaptive HPE. The first category involves a shared structure, where the weights of the networks from both the source pretrain and target adaptation are shared. CC-SSL [25] employs a single end-to-end trained network, RegDA [12] utilizes one shared feature extractor and two separate regressors, and TransPar [8] emphasizes transferable parameters using a similar structure. The second category is the unshared structure, where the pretrain and adaptation create a teacher-student paradigm. In this category, MDAM [19] addresses this structure alongside a novel pseudo-label strategy, UniFrame [16] modifies the classic Mean-Teacher [32] model by combining it with style transfer [10], and MarsDA [14] applies a teacher-student approach to edit RegDA. However, all these existing domain adaptive HPE approaches are not source-free, which raises concerns about data privacy and security. Therefore, it is necessary to consider source-free domain adaptive HPE. To our knowledge, we are the first work to study this challenging problem." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 541, + 104, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 541, + 104, + 554 + ], + "spans": [ + { + "bbox": [ + 47, + 541, + 104, + 554 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 562, + 180, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 180, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 180, + 574 + ], + "type": "text", + "content": "3.1. Preliminary of 2D HPE" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "In 2D HPE, we are given a labeled pose dataset " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = (x_i,y_i)" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "x_{i}\\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " is an image and " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "y_{i}\\in \\mathbb{R}^{K\\times 2}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " are the corresponding keypoint coordinates. Here, " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " represent the height and width of the image, respectively, and " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " is the number of channels. Additionally, " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " represents the number of keypoints. To ensure smoother training, most existing 2D methods represent the keypoint coordinates as heatmaps " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "HM_{i}\\in \\mathbb{R}^{K\\times H^{\\prime}\\times W^{\\prime}}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "H^{\\prime}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "W^{\\prime}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " are the height and width of the heatmaps, respectively. These heatmaps are used to represent the coordinates in a spatial way. The process of how the model works is defined" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 68, + 547, + 159 + ], + "blocks": [ + { + "bbox": [ + 307, + 68, + 547, + 159 + ], + "lines": [ + { + "bbox": [ + 307, + 68, + 547, + 159 + ], + "spans": [ + { + "bbox": [ + 307, + 68, + 547, + 159 + ], + "type": "image", + "image_path": "99393e82b2670ff4be632a31185e86033b4e3ced3a013c36bcad381f13d55625.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 160, + 547, + 205 + ], + "lines": [ + { + "bbox": [ + 304, + 160, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 160, + 547, + 205 + ], + "type": "text", + "content": "Figure 1. Process of heatmap-based 2D HPE. The model includes two components: the feature extractor " + }, + { + "bbox": [ + 304, + 160, + 547, + 205 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 160, + 547, + 205 + ], + "type": "text", + "content": " and the regressor " + }, + { + "bbox": [ + 304, + 160, + 547, + 205 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 160, + 547, + 205 + ], + "type": "text", + "content": ". After feeding the input image to the model, we obtain " + }, + { + "bbox": [ + 304, + 160, + 547, + 205 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 160, + 547, + 205 + ], + "type": "text", + "content": " output heatmaps corresponding to " + }, + { + "bbox": [ + 304, + 160, + 547, + 205 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 160, + 547, + 205 + ], + "type": "text", + "content": " keypoints." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 224, + 547, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 547, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 547, + 357 + ], + "type": "text", + "content": "in Fig. 1. Ground truth heatmaps are generated by a 2D Gaussian centered at the ground truth joint location, with one heatmap for each keypoint. We denote the transform from coordinates to heatmaps as " + }, + { + "bbox": [ + 304, + 224, + 547, + 357 + ], + "type": "inline_equation", + "content": "T: \\mathbb{R}^{K \\times 2} \\to \\mathbb{R}^{K \\times H' \\times W'}" + }, + { + "bbox": [ + 304, + 224, + 547, + 357 + ], + "type": "text", + "content": ". The training processes are conducted between the output heatmaps and ground truth heatmaps using MSE for supervised learning. In the inference stage, well-trained supervised models output heatmaps that predict the probability of joints occurring at each pixel. For each heatmap, the pixel with the highest probability is considered to be the predicted location of the joint." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 363, + 419, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 363, + 419, + 374 + ], + "spans": [ + { + "bbox": [ + 306, + 363, + 419, + 374 + ], + "type": "text", + "content": "3.2. Problem Statement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "text", + "content": "In the source-free domain adaptive HPE setting, we have a source domain dataset " + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{(x_i^s, y_i^s)\\}_{i=1}^{n_s}" + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "inline_equation", + "content": "n_s" + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "text", + "content": " labeled pose samples. " + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "inline_equation", + "content": "x_i^s \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "text", + "content": " is the source image and " + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "inline_equation", + "content": "y_i^s \\in \\mathbb{R}^{K \\times 2}" + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "text", + "content": " is the corresponding pose annotation. Besides, there exists a target domain dataset " + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{x_i^t\\}_{i=1}^{n_t}" + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "text", + "content": " that includes " + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "inline_equation", + "content": "n_t" + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "text", + "content": " unlabeled pose samples. The source and target domains share the same label space but lie in different distributions (e.g. synthetic vs. real). The training procedure is split into two stages: the pretrain stage and the adaptation stage. In the pretrain stage, the labeled source domain dataset " + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "text", + "content": " is used to train the source model " + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "inline_equation", + "content": "f^{sr}" + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "text", + "content": ". However, what distinguishes source-free domain adaptive HPE from general domain adaptive HPE is the adaptation stage, where only the source model " + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "inline_equation", + "content": "f^{sr}" + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "text", + "content": " and target domain dataset " + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 382, + 547, + 586 + ], + "type": "text", + "content": " are available for training, and the source domain dataset cannot be used anymore. Our objective is to develop a model that performs well on the target dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 592, + 440, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 592, + 440, + 605 + ], + "spans": [ + { + "bbox": [ + 306, + 592, + 440, + 605 + ], + "type": "text", + "content": "3.3. Pipeline of Our Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 611, + 437, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 437, + 622 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 437, + 622 + ], + "type": "text", + "content": "3.3.1 Source Model Pretrain" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "For DA in HPE, labeled source data are applied for pretraining. Same as most existing domain adaptive pose estimation methods [25, 19, 12, 16], the network is composed of a feature extractor " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " (such as ResNet backbone) and a pose regressor " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": ", so the source model can be represented as " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "f^{sr} = F^{sr}(G^{sr}(\\cdot))" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": ". Based on the supervised heatmap-based loss (MSE loss), we propose the overall objective for" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4828" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 286, + 285 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 286, + 285 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 286, + 285 + ], + "type": "image", + "image_path": "c960532eded9a22c3b979ce012a59d9294c563557223b317552fadd1d1649de8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 287, + 547, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 287, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 547, + 331 + ], + "type": "text", + "content": "Figure 2. Overall framework of our method. It includes three models, each with one feature extractor and one regressor. The adaptation contains two steps: Step A and Step B. Step A is applied for source-protect adaptation, while Step B aims at target-relevant adaptation. The solid arrows show the forward process of associated losses, and the dotted arrows show the backward updating process of associated losses. EMA means the exponential moving average process from target model to intermediate model to update intermediate model's weights." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 294, + 69, + 547, + 286 + ], + "blocks": [ + { + "bbox": [ + 294, + 69, + 547, + 286 + ], + "lines": [ + { + "bbox": [ + 294, + 69, + 547, + 286 + ], + "spans": [ + { + "bbox": [ + 294, + 69, + 547, + 286 + ], + "type": "image", + "image_path": "417679ac09e08dce32c594ba5d8369484112f2f1450c11c82be088228c035e47.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 351, + 126, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 351, + 126, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 351, + 126, + 363 + ], + "type": "text", + "content": "source pretraining:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 369, + 287, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 369, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 52, + 369, + 287, + 384 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p r e t r a i n}} = \\mathbb {E} _ {\\left(x _ {i} ^ {s}, y _ {i} ^ {s}\\right) \\in \\mathcal {S}} \\mathcal {L} _ {\\text {m s e}} \\left(F ^ {s r} \\left(G ^ {s r} \\left(x _ {i} ^ {s}\\right)\\right), T \\left(y _ {i} ^ {s}\\right)\\right), \\tag {1}", + "image_path": "e7adf996cfd9228caadebc305dec5931672b1bbfcfdf751777cba3f16fd108a5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 388, + 288, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 288, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 388, + 288, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mse}" + }, + { + "bbox": [ + 46, + 388, + 288, + 437 + ], + "type": "text", + "content": " is the Mean Squared Error (MSE) loss between the heatmap representations of prediction and ground truth, and " + }, + { + "bbox": [ + 46, + 388, + 288, + 437 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 388, + 288, + 437 + ], + "type": "text", + "content": " is the transform from coordinates to heatmaps defined in Sec. 3.1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 451, + 181, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 451, + 181, + 463 + ], + "spans": [ + { + "bbox": [ + 47, + 451, + 181, + 463 + ], + "type": "text", + "content": "3.3.2 Adaptation Framework" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "text", + "content": "As Fig. 2 shows, our method contains three models: source model " + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "inline_equation", + "content": "f^{sr} = F^{sr}(G^{sr}(\\cdot))" + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "text", + "content": ", intermediate model " + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "inline_equation", + "content": "f^{in} = F^{in}(G^{in}(\\cdot))" + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "text", + "content": ", and target model " + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "inline_equation", + "content": "f^{tg} = F^{tg}(G^{tg}(\\cdot))" + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "text", + "content": ". Each model includes one feature extractor and one regressor. Assume the heatmaps generated by source model as " + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "inline_equation", + "content": "H_{i}^{sr} = F^{sr}(G^{sr}(x_{i}^{t})) = [h_{i,1}^{sr}, h_{i,2}^{sr}, \\dots, h_{i,K}^{sr}]^{\\top}" + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "text", + "content": ", and those from intermediate model and target model respectively as " + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "inline_equation", + "content": "H_{i}^{in} = F^{in}(G^{in}(x_{i}^{t})) = [h_{i,1}^{in}, h_{i,2}^{in}, \\dots, h_{i,K}^{in}]^{\\top}" + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "inline_equation", + "content": "H_{i}^{tg} = F^{tg}(G^{tg}(x_{i}^{t})) = [h_{i,1}^{tg}, h_{i,2}^{tg}, \\dots, h_{i,K}^{tg}]^{\\top}" + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "text", + "content": " is the number of keypoints. We also assume the keypoint indexing space as " + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = [1,2,\\dots,K]" + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "text", + "content": ". The source model is used to maintain source information, while the target model is applied to learn target knowledge. The intermediate model is the transition model between these two models. It serves as a bridge between source and target as it learns from source and target simultaneously, thus eliminating the domain gap implicitly. Therefore, it outperforms both the source and target models and is used for the final inference. When starting the adaptation, " + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "inline_equation", + "content": "f^{in}" + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "inline_equation", + "content": "f^{tg}" + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "text", + "content": " are initialized from the source model " + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "inline_equation", + "content": "f^{sr}" + }, + { + "bbox": [ + 46, + 469, + 289, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 350, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 350, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 305, + 350, + 545, + 387 + ], + "type": "text", + "content": "For a single adaptation iteration, two steps are executed as shown in Fig. 2. Step A is for the source-protect adaptation, and Step B aims to adapt in a target-relevant way." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 403, + 545, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 547 + ], + "type": "text", + "content": "In Step A, our goal is to overcome catastrophic forgetting on source domain, while eliminating the noise from the source model due to domain shift. Therefore, the source model's feature extractor " + }, + { + "bbox": [ + 304, + 403, + 545, + 547 + ], + "type": "inline_equation", + "content": "G^{sr}" + }, + { + "bbox": [ + 304, + 403, + 545, + 547 + ], + "type": "text", + "content": " is fixed to retain source knowledge, while its regressor " + }, + { + "bbox": [ + 304, + 403, + 545, + 547 + ], + "type": "inline_equation", + "content": "F^{sr}" + }, + { + "bbox": [ + 304, + 403, + 545, + 547 + ], + "type": "text", + "content": " keeps updating to improve regression accuracy, hence reducing noises in the outputs. On the opposite, the intermediate model's regressor " + }, + { + "bbox": [ + 304, + 403, + 545, + 547 + ], + "type": "inline_equation", + "content": "F^{in}" + }, + { + "bbox": [ + 304, + 403, + 545, + 547 + ], + "type": "text", + "content": " is fixed, but its feature extractor " + }, + { + "bbox": [ + 304, + 403, + 545, + 547 + ], + "type": "inline_equation", + "content": "G^{in}" + }, + { + "bbox": [ + 304, + 403, + 545, + 547 + ], + "type": "text", + "content": " continues to update. That's because we want the feature extractor to learn more source knowledge on the representation level, while the noise from the source model due to domain shifts can be inhibited." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "type": "text", + "content": "In Step B, we aim to conduct the adaptation between the intermediate model and the target model in a target-relevant manner. During the process, several pose-specific techniques are utilized to minimize the discrepancy between these two models. Though both " + }, + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "type": "inline_equation", + "content": "f^{in}" + }, + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "type": "inline_equation", + "content": "f^{tg}" + }, + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "type": "text", + "content": " are updated at this step, " + }, + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "type": "inline_equation", + "content": "f^{tg}" + }, + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "type": "text", + "content": " is updated via back-propagation while " + }, + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "type": "inline_equation", + "content": "f^{in}" + }, + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "type": "text", + "content": " is via exponential moving average from " + }, + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "type": "inline_equation", + "content": "f^{tg}" + }, + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "type": "text", + "content": ", since averaging model weights over training steps tends to produce a more accurate model than using the final weights directly [32]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "In the following two subsections, we introduce the detailed techniques used for Step A and Step B." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4829" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 289, + 201 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 289, + 201 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 289, + 201 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 289, + 201 + ], + "type": "image", + "image_path": "7494d1f6f42d043899528e8a9b02de3dfb7e3ece876fe587967bc2611903c160.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 203, + 287, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 203, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 287, + 246 + ], + "type": "text", + "content": "Figure 3. Proposed residual loss. Here we choose two heatmaps at the same location generated from source model and intermediate model for illustration. For simplicity, here the heatmap's scale is set as " + }, + { + "bbox": [ + 46, + 203, + 287, + 246 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 203, + 287, + 246 + ], + "type": "text", + "content": ", while the actual size is " + }, + { + "bbox": [ + 46, + 203, + 287, + 246 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 203, + 287, + 246 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 266, + 182, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 182, + 278 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 182, + 278 + ], + "type": "text", + "content": "3.3.3 Source-protect Modules" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 285, + 287, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 344 + ], + "type": "text", + "content": "In this part, we propose modules related to source-free-specific adaptation. The biggest challenge is to distill knowledge from the source model and resist the potential noise due to domain shift. First, we introduce the finetune loss that calibrates on the source model:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 354, + 287, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 287, + 369 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f t} = \\mathbb {E} _ {x _ {i} ^ {t} \\in \\mathcal {T}} \\mathcal {L} _ {m s e} \\left(F ^ {s r} \\left(G ^ {s r} \\left(x _ {i} ^ {t}\\right)\\right), F ^ {i n} \\left(G ^ {i n} \\left(x _ {i} ^ {t}\\right)\\right)\\right). \\tag {2}", + "image_path": "96918e38b02334a01dc923908da39f252a12197ee6bdcd9ec4dee77cbdcf2ccd.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 378, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 287, + 437 + ], + "type": "text", + "content": "To be notified, we hope the regression results of the source model can be more accurate, while the source knowledge can be retained. Thus, the feature extractor is fixed so that the source representations continue to be generated, and the regressor is being updated:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 446, + 287, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 446, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 111, + 446, + 287, + 472 + ], + "type": "interline_equation", + "content": "\\theta_ {F ^ {s r}} ^ {t + 1} = \\theta_ {F ^ {s r}} ^ {t} - \\lambda^ {s r} \\frac {\\partial L _ {f t}}{\\partial \\theta_ {F ^ {s r}} ^ {t}}, \\tag {3}", + "image_path": "63de8b89f19b1a9d16c79decf7c35b15a9cd2c432f45fed23fe361b9bb13bbf1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 481, + 287, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 539 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 481, + 287, + 539 + ], + "type": "inline_equation", + "content": "\\theta_F^s" + }, + { + "bbox": [ + 46, + 481, + 287, + 539 + ], + "type": "text", + "content": " is the weights of the source model's regressor " + }, + { + "bbox": [ + 46, + 481, + 287, + 539 + ], + "type": "inline_equation", + "content": "F^s" + }, + { + "bbox": [ + 46, + 481, + 287, + 539 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 46, + 481, + 287, + 539 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 481, + 287, + 539 + ], + "type": "text", + "content": " denotes the training step and " + }, + { + "bbox": [ + 46, + 481, + 287, + 539 + ], + "type": "inline_equation", + "content": "\\lambda^s" + }, + { + "bbox": [ + 46, + 481, + 287, + 539 + ], + "type": "text", + "content": " is the source model's learning rate. However, improving source model is not enough. What's more, we propose the residual loss to improve intermediate model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 540, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 715 + ], + "type": "text", + "content": "The application of Equation 2 facilitates mutual knowledge transfer between the source and intermediate models. Apart from the pixel with the highest confidence in a heatmap, the other pixels also contain valuable information that benefits the knowledge transfer, and that's why we propose residual loss in Fig. 3. From the example in the figure, the most confident pixel (confidence " + }, + { + "bbox": [ + 46, + 540, + 287, + 715 + ], + "type": "inline_equation", + "content": "= 1.0" + }, + { + "bbox": [ + 46, + 540, + 287, + 715 + ], + "type": "text", + "content": ") in the source heatmap is removed, along with the pixel at the same location which has a confidence of 0.8 in the intermediate heatmap marked with red. The pair marked with blue follows the same rule but on the basis of intermediate heatmap. Then we build adaptive residual heatmaps denoted as " + }, + { + "bbox": [ + 46, + 540, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\hat{h}_{i,j}^{sr}" + }, + { + "bbox": [ + 46, + 540, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 540, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\hat{h}_{i,j}^{in}" + }, + { + "bbox": [ + 46, + 540, + 287, + 715 + ], + "type": "text", + "content": ". Moreover, we employ KL divergence to ensure that " + }, + { + "bbox": [ + 46, + 540, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\hat{h}_{i,j}^{in}" + }, + { + "bbox": [ + 46, + 540, + 287, + 715 + ], + "type": "text", + "content": " approaches " + }, + { + "bbox": [ + 46, + 540, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\hat{h}_{i,j}^{sr}" + }, + { + "bbox": [ + 46, + 540, + 287, + 715 + ], + "type": "text", + "content": ", thereby preserving the source in" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 69, + 536, + 152 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 536, + 152 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 536, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 536, + 152 + ], + "type": "image", + "image_path": "18126c75a76efac6603c7be223e6f0e5486647a6d6f1b8a459a89950007e3a71.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 155, + 545, + 210 + ], + "lines": [ + { + "bbox": [ + 305, + 155, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 305, + 155, + 545, + 210 + ], + "type": "text", + "content": "Figure 4. Illustration of the projection of a heatmap. In order to reduce the sparsity of a heatmap, we project it into a horizontal vector and a vertical vector. For simplicity, here the heatmap's scale is set as " + }, + { + "bbox": [ + 305, + 155, + 545, + 210 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 305, + 155, + 545, + 210 + ], + "type": "text", + "content": " and the vector's scale as " + }, + { + "bbox": [ + 305, + 155, + 545, + 210 + ], + "type": "inline_equation", + "content": "4 \\times 1" + }, + { + "bbox": [ + 305, + 155, + 545, + 210 + ], + "type": "text", + "content": ", while the actual heatmap's size is " + }, + { + "bbox": [ + 305, + 155, + 545, + 210 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 305, + 155, + 545, + 210 + ], + "type": "text", + "content": " and the vector's size is " + }, + { + "bbox": [ + 305, + 155, + 545, + 210 + ], + "type": "inline_equation", + "content": "64 \\times 1" + }, + { + "bbox": [ + 305, + 155, + 545, + 210 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 231, + 541, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 231, + 541, + 242 + ], + "spans": [ + { + "bbox": [ + 306, + 231, + 541, + 242 + ], + "type": "text", + "content": "formation. Based on these, we deduce the residual loss as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 251, + 545, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 251, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 320, + 251, + 545, + 268 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e s} = \\mathbb {E} _ {x _ {i} ^ {t} \\in \\mathcal {T}} \\mathbb {E} _ {j \\in \\mathcal {P}} \\mathrm {D} _ {\\mathrm {K L}} \\left(\\sigma \\left(\\hat {h} _ {i, j} ^ {s r} / \\tau\\right) \\mid \\mid \\sigma \\left(\\hat {h} _ {i, j} ^ {i n} / \\tau\\right)\\right), \\tag {4}", + "image_path": "16177f32f0373e2e4387ee6677c34a80bac1ac444ae42077cf84c640573e5209.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 275, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 275, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 275, + 545, + 346 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 275, + 545, + 346 + ], + "type": "inline_equation", + "content": "\\mathrm{D}_{\\mathrm{KL}}" + }, + { + "bbox": [ + 305, + 275, + 545, + 346 + ], + "type": "text", + "content": " is the KL divergence, and " + }, + { + "bbox": [ + 305, + 275, + 545, + 346 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 305, + 275, + 545, + 346 + ], + "type": "text", + "content": " is the softmax function. Besides, " + }, + { + "bbox": [ + 305, + 275, + 545, + 346 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 275, + 545, + 346 + ], + "type": "text", + "content": " is the temperature used to scale the residual heatmaps and is empirically set to 0.3. By combining the finetune loss and the residual loss, we get the objective function with the optimization process for updating the intermediate model:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 354, + 545, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 354, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 318, + 354, + 545, + 380 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i n} = \\mathcal {L} _ {f t} + \\alpha \\mathcal {L} _ {r e s}, \\quad \\theta_ {G ^ {i n}} ^ {t + 1} = \\theta_ {G ^ {i n}} ^ {t} - \\lambda^ {i n} \\frac {\\partial L _ {i n}}{\\partial \\theta_ {G ^ {i n}} ^ {t}}, \\tag {5}", + "image_path": "d54dfe96cb4cd022e757055146b6397659bf1507a6cc1efcfee7ee8389bf77db.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 391, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 391, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 306, + 391, + 545, + 425 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 391, + 545, + 425 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 306, + 391, + 545, + 425 + ], + "type": "text", + "content": " is the training step, and " + }, + { + "bbox": [ + 306, + 391, + 545, + 425 + ], + "type": "inline_equation", + "content": "\\lambda^{in}" + }, + { + "bbox": [ + 306, + 391, + 545, + 425 + ], + "type": "text", + "content": " is the learning rate for " + }, + { + "bbox": [ + 306, + 391, + 545, + 425 + ], + "type": "inline_equation", + "content": "G^{in}" + }, + { + "bbox": [ + 306, + 391, + 545, + 425 + ], + "type": "text", + "content": ". Moreover, " + }, + { + "bbox": [ + 306, + 391, + 545, + 425 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 306, + 391, + 545, + 425 + ], + "type": "text", + "content": " is the trade-off hyperparameter between these two losses." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 443, + 443, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 443, + 454 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 443, + 454 + ], + "type": "text", + "content": "3.3.4 Target-relevant Modules" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "type": "text", + "content": "This section introduces a module that is associated with target information. As previously mentioned, the sparsity of keypoints in the image presents a significant challenge in our task. To address this issue, we propose using projected horizontal and vertical vectors of heatmaps as spatial representations, as shown in Fig. 4. This approach reduces the sparsity of the spatial probability space effectively since the number of overall digits inside the heatmap is " + }, + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "type": "inline_equation", + "content": "64 \\times 64 \\approx 4K" + }, + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "type": "text", + "content": ", while the number of units in each vector is 64. Therefore, we can effectively reduce the sparsity of the spatial probability space. With 64 units in each vector, the spatial probability space size is acceptable compared to the number of keypoints." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Despite the reduction in sparsity, the completeness of spatial representations is not affected. That is because all the coordinates can always be divided into two representations, horizontally and vertically. In such a case, our spatial representation is better than the " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "-classification probability space proposed by [12], which only includes the keypoint-relevant information and does not take the units in the heatmap that are not keypoints into consideration." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4830" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 68, + 289, + 177 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 289, + 177 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 289, + 177 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 289, + 177 + ], + "type": "image", + "image_path": "92ecc105b5653a9cc0a04a1838ecbb0834757d4eb0bdd568f766d052fb9cae0c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 178, + 287, + 211 + ], + "lines": [ + { + "bbox": [ + 47, + 178, + 287, + 211 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 287, + 211 + ], + "type": "text", + "content": "Figure 5. Illustration of building positive and negative keypoint pairs for contrastive learning between the output heatmaps of the intermediate model and target model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 230, + 287, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 230, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 47, + 230, + 287, + 315 + ], + "type": "text", + "content": "In Fig. 4, we introduce the projection of a heatmap from both horizontal and vertical directions. For the heatmap " + }, + { + "bbox": [ + 47, + 230, + 287, + 315 + ], + "type": "inline_equation", + "content": "h_{i,j}^{in}" + }, + { + "bbox": [ + 47, + 230, + 287, + 315 + ], + "type": "text", + "content": " generated by the intermediate model from " + }, + { + "bbox": [ + 47, + 230, + 287, + 315 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 230, + 287, + 315 + ], + "type": "text", + "content": "th sample's " + }, + { + "bbox": [ + 47, + 230, + 287, + 315 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 230, + 287, + 315 + ], + "type": "text", + "content": "th keypoint, its projection can be represented as " + }, + { + "bbox": [ + 47, + 230, + 287, + 315 + ], + "type": "inline_equation", + "content": "proj( h_{i,j}^{in}) = (v_{i,jx}^{in}, v_{i,jy}^{in})" + }, + { + "bbox": [ + 47, + 230, + 287, + 315 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 230, + 287, + 315 + ], + "type": "inline_equation", + "content": "v_{i,jx}^{in}" + }, + { + "bbox": [ + 47, + 230, + 287, + 315 + ], + "type": "text", + "content": " is the horizontal vector and " + }, + { + "bbox": [ + 47, + 230, + 287, + 315 + ], + "type": "inline_equation", + "content": "v_{i,jy}^{in}" + }, + { + "bbox": [ + 47, + 230, + 287, + 315 + ], + "type": "text", + "content": " is the vertical vector. This vector pair is the foundation of the following modules and losses that we propose." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 315, + 288, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 288, + 426 + ], + "type": "text", + "content": "First, we propose the pose-specific contrastive loss between the intermediate model and target model. According to the observation from [12], wrong predictions are usually located at other keypoints' positions. In such a case, improving regression can be achieved by minimizing the discrepancy of heatmaps at the same location and maximizing the discrepancy of heatmaps at different locations. And that's how we define the positive and negative pair in Fig. 5. Denote the similarity of two heatmaps generated by the intermediate model and target model respectively as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 430, + 287, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 287, + 460 + ], + "type": "interline_equation", + "content": "s i m \\left(h _ {i, j} ^ {i n}, h _ {i, k} ^ {t g}\\right) = \\frac {1}{2} \\left(\\frac {v _ {i , j x} ^ {i n} \\cdot v _ {i , k x} ^ {t g}}{| | v _ {i , j x} ^ {i n} | | | | | v _ {i , k x} ^ {t g} | |} + \\frac {v _ {i , j y} ^ {i n} \\cdot v _ {i , k y} ^ {t g}}{| | v _ {i , j y} ^ {i n} | | | | | v _ {i , k y} ^ {t g} | |}\\right). \\tag {6}", + "image_path": "f416ca5773327d79588f51886f37e35834ea43a9101c90975f39507875d2c7e6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 464, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 287, + 489 + ], + "type": "text", + "content": "On the basis of the similarity, we define the overall pose-specific contrastive loss as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 494, + 287, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 494, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 59, + 494, + 287, + 523 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c s t} = - \\mathbb {E} _ {x _ {i} ^ {t} \\in \\mathcal {T}} \\mathbb {E} _ {j \\in \\mathcal {P}} \\log \\frac {\\exp (s i m (h _ {i , j} ^ {i n} , h _ {i , j} ^ {t g}))}{\\Sigma_ {k \\in \\mathcal {P}} \\exp (s i m (h _ {i , j} ^ {i n} , h _ {i , k} ^ {t g}))}. \\tag {7}", + "image_path": "7f2e3b70f0dd9eaf5faf1feae23b1233c3e237d21d7fee5d5bf511e332db9660.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 528, + 288, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 288, + 635 + ], + "type": "text", + "content": "Next, we propose a pose-specific self-supervised technique called pose-specific information maximization. Information maximization is very helpful to source-free domain adaptation for classification [21] because it makes the target outputs individually certain and globally diverse, hence mitigating the domain gap implicitly. However, the improvements are limited due to the sparsity of heatmaps. With the help of projected vectors, we refine the information maximization loss as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 643, + 287, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 643, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 54, + 643, + 287, + 658 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i m} = \\mathbb {E} _ {x _ {i} ^ {t} \\in \\mathcal {T}} \\left(\\mathcal {L} _ {e n t x} \\left(x _ {i} ^ {t}\\right) + \\mathcal {L} _ {e n t y} \\left(x _ {i} ^ {t}\\right) - \\mathcal {L} _ {d i v} \\left(x _ {i} ^ {t}\\right)\\right), \\tag {8}", + "image_path": "3b4e30f1b0078d54fdccbc8d3f4e388bf91105ef45e3c825616b43f17a8159df.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 665, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 665, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 59, + 665, + 287, + 717 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\mathcal {L} _ {e n t x} \\left(x _ {i} ^ {t}\\right) = - \\mathbb {E} _ {j \\in \\mathcal {P}} \\left[ \\sigma \\left(v _ {i, j x} ^ {t g}\\right) \\log \\sigma \\left(v _ {i, j x} ^ {t g}\\right) \\right], \\\\ \\mathcal {L} _ {e n t y} \\left(x _ {i} ^ {t}\\right) = - \\mathbb {E} _ {j \\in \\mathcal {P}} \\left[ \\sigma \\left(v _ {i, j y} ^ {t g}\\right) \\log \\sigma \\left(v _ {i, j y} ^ {t g}\\right) \\right], \\\\ \\mathcal {L} _ {d i v} \\left(x _ {i} ^ {t}\\right) = - \\left[ \\mathbb {E} _ {j \\in \\mathcal {P}} \\sigma \\left(h _ {i, j} ^ {t g}\\right) \\right] \\log \\left[ \\mathbb {E} _ {j \\in \\mathcal {P}} \\sigma \\left(h _ {i, j} ^ {t g}\\right) \\right]. \\end{array} \\right. \\tag {9}", + "image_path": "3c8a362f7b7acb49d7c78b71b2b9044f0b71458efdb988ccfe2e31a3e3d8d64f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 118 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 305, + 72, + 545, + 118 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{entx}" + }, + { + "bbox": [ + 305, + 72, + 545, + 118 + ], + "type": "text", + "content": " is the entropy of horizontal vectors, and " + }, + { + "bbox": [ + 305, + 72, + 545, + 118 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{enty}" + }, + { + "bbox": [ + 305, + 72, + 545, + 118 + ], + "type": "text", + "content": " is the entropy of vertical vectors. The last term in Equation 8 is the diversity-promoting objective. Moreover, we introduce the consistency loss between the two models' outputs:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 123, + 545, + 138 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 123, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 311, + 123, + 545, + 138 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c o n} = \\mathbb {E} _ {\\left(x _ {i} ^ {s}, y _ {i} ^ {s}\\right) \\in \\mathcal {S}} \\mathcal {L} _ {m s e} \\left(F ^ {i n} \\left(G ^ {i n} \\left(x _ {i} ^ {s}\\right)\\right), F ^ {t g} \\left(G ^ {t g} \\left(x _ {i} ^ {s}\\right)\\right)\\right). \\tag {10}", + "image_path": "e52e3d13e7d3fa7d65b6d4ea82da6ec26547a2d2fe73e2321b300aabaa3301c5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 144, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 144, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 545, + 168 + ], + "type": "text", + "content": "By summarizing above, we conclude the overall objective loss function for the target model as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 361, + 177, + 545, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 177, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 361, + 177, + 545, + 190 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t g t} = \\mathcal {L} _ {c o n} + \\beta \\mathcal {L} _ {c s t} + \\gamma \\mathcal {L} _ {i m}, \\tag {11}", + "image_path": "36a73fac4f18d4ca100d4c99b34d96020d513316cbf47ec80df35a70bb0ff6fd.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 198, + 468, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 198, + 468, + 210 + ], + "spans": [ + { + "bbox": [ + 306, + 198, + 468, + 210 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 198, + 468, + 210 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 306, + 198, + 468, + 210 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 198, + 468, + 210 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 306, + 198, + 468, + 210 + ], + "type": "text", + "content": " are trade-off parameters." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 210, + 545, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 210, + 545, + 235 + ], + "spans": [ + { + "bbox": [ + 305, + 210, + 545, + 235 + ], + "type": "text", + "content": "Moreover, the intermediate model is updated via the exponential moving average strategy (EMA) in this step:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 365, + 241, + 545, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 241, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 365, + 241, + 545, + 258 + ], + "type": "interline_equation", + "content": "\\theta_ {f ^ {i n}} ^ {t + 1} = \\eta \\theta_ {f ^ {i n}} ^ {t} + (1 - \\eta) \\theta_ {f ^ {t g}} ^ {t}, \\tag {12}", + "image_path": "43cbf85c1f5b542ae3a409f9a826028343e9bf00c16b5aa03f9fee71d0934754.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 264, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 264, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 264, + 545, + 289 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 264, + 545, + 289 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 264, + 545, + 289 + ], + "type": "text", + "content": " denotes the step of training and " + }, + { + "bbox": [ + 305, + 264, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 305, + 264, + 545, + 289 + ], + "type": "text", + "content": " denotes the smoothing coefficient set to 0.999 by default." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 299, + 387, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 299, + 387, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 299, + 387, + 312 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 319, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 486 + ], + "type": "text", + "content": "Datasets. We use three human pose datasets and three hand pose datasets to validate our approach. SURREAL [36] is utilized as the source dataset and encompasses six million synthetic human pose images. Human3.6M [11], one of the target human pose datasets, is a frequently used real-world dataset in the community, consisting of 3.6 million images distributed into seven folds. According to previous research [12, 16], S1, S5, S6, S7, and S8 are designated as the training set, whereas S9 and S11 serve as the testing set. The other target dataset is Leeds Sports Pose [15] (LSP), which is a real-world dataset with 2,000 images, and we use all of them for adaptation. We focus on two domain adaptation tasks: SURREAL " + }, + { + "bbox": [ + 304, + 319, + 545, + 486 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 319, + 545, + 486 + ], + "type": "text", + "content": " Human3.6M and SURREAL " + }, + { + "bbox": [ + 304, + 319, + 545, + 486 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 319, + 545, + 486 + ], + "type": "text", + "content": " LSP for human pose tasks." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 486, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 628 + ], + "type": "text", + "content": "Rendered Hand Pose Dataset [42] (RHD) is the source hand dataset with 43,986 synthetic hand images, of which 41,258 are allocated for training, while the remaining 2,728 images are utilized for validation. Hand-3D-Studio [41] (H3D) is one of the two target hand datasets, including 22,000 real-world frames. For training, we employed 18,800 frames and reserved the rest for testing, following the protocol outlined in previous research papers [12, 16]. The other target hand dataset, FreiHand [43], provides 130k real-world images, and we employed all of them for our adaptation task. Our focus is on two domain adaptation tasks: RHD " + }, + { + "bbox": [ + 304, + 486, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 486, + 545, + 628 + ], + "type": "text", + "content": " H3D and RHD " + }, + { + "bbox": [ + 304, + 486, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 486, + 545, + 628 + ], + "type": "text", + "content": " FreiHand." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": "Implementation Details. We adopt Simple Baseline [39] with ResNet-101 [9] as the HPE backbone, following previous works [12, 16]. In the pretrain process, we conduct 40 epochs, each with 500 iterations, using the Adam [17] optimizer with an initial learning rate of 1e-4. The learning rate decreases to 1e-5 at 25 epochs. For the adaptation process, we execute 40 epochs with 30,000 iterations. In step" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4831" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "A, we choose an initial learning rate of 1e-4 for " + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "inline_equation", + "content": "F^{sr}" + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": " and 1e-5 for " + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "inline_equation", + "content": "G^{in}" + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": ". Step B sets the initial learning rate to 1e-4 for " + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "inline_equation", + "content": "G^{tg}" + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": " and 1e-3 for " + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "inline_equation", + "content": "F^{tg}" + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": ". We use the same annealing strategy for the learning rate scheduler as in [12]. As for hyperparameters, we select " + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "inline_equation", + "content": "\\alpha = 0.7" + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "inline_equation", + "content": "\\beta = 0.5" + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "inline_equation", + "content": "\\gamma = 0.85" + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": ". Moreover, the intermediate model is used for the final inference." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 154, + 132, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 154, + 132, + 166 + ], + "spans": [ + { + "bbox": [ + 47, + 154, + 132, + 166 + ], + "type": "text", + "content": "4.1. Main Results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 173, + 288, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 173, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 173, + 288, + 520 + ], + "type": "text", + "content": "Baselines. Four general domain adaptive pose estimation methods CC-SSL [25], MDAM [19], RegDA [12] and UniFrame [16] with SOTA performances are chosen as baselines for comparison. We list these baselines that need to access source data to demonstrate that our proposed source-free method's performance remains competitive when compared to general domain adaptive pose estimation methods. We build source-free baselines by either adapting domain-adaptive HPE methods or source-free DA methods for classification methods. We first introduce two domain adaptive HPE methods. RegDA-SF is based on [12], which originally contains three steps and the first step uses source data. We replace the first step with an exponential moving averaging strategy to obtain the new method. Uniframe-SF is on the foundation of [16]. What we do is remove the style transfer modules between source and target so that it becomes a source-free method. The rest are adapted from source-free DA methods for classification methods. SHOT [21] is a classic source-free domain adaptation method for image classification, and we adapt its mutual information maximization module for HPE by treating the normalized heatmaps as output probabilities in classification. MMT is based on [6], which is a source-free adaptation method that applies two models mutually learning from each other. Here we replace its soft classification loss with the soft consistency loss. SHOT++ [22] is the advanced version of [21], which uses a special unsupervised representation learning technique predicting image rotations. We also add this module to our adopted baseline." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 521, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 676 + ], + "type": "text", + "content": "Metrics. To evaluate the accuracy of the proposed approach, we adopt the Percentage of Correct Keypoint (PCK) metric, as used in previous work. We set the ratio of correct predictions to be " + }, + { + "bbox": [ + 46, + 521, + 287, + 676 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 676 + ], + "type": "text", + "content": ", and report PCK@0.05 in Tables 1-4. In addition to overall keypoint accuracy, we divide joints into several part segments and measure their performance using specific metrics. For the 21-keypoint hand skeleton, we use the Metacarpophalangeal (MCP), Proximal Interphalangeal (PIP), Distal Interphalangeal (DIP), and Finger-tip (Fin) metrics. For the 18-keypoint human skeleton, we select the Shoulder (Sld), Elbow (Elb), Wrist, Hip, Knee, and Ankle metrics. These segment-specific metrics enable a more detailed evaluation of the models' performances." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Quantitative Results. We present the results of our proposed method on four different tasks: hand tasks RHD " + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": " H3D and RHD " + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": " FreiHand, and human tasks" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 310, + 79, + 541, + 178 + ], + "blocks": [ + { + "bbox": [ + 347, + 69, + 504, + 79 + ], + "lines": [ + { + "bbox": [ + 347, + 69, + 504, + 79 + ], + "spans": [ + { + "bbox": [ + 347, + 69, + 504, + 79 + ], + "type": "text", + "content": "Table 1. PCK@0.05 on RHD " + }, + { + "bbox": [ + 347, + 69, + 504, + 79 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 347, + 69, + 504, + 79 + ], + "type": "text", + "content": " H3D Task" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 79, + 541, + 178 + ], + "lines": [ + { + "bbox": [ + 310, + 79, + 541, + 178 + ], + "spans": [ + { + "bbox": [ + 310, + 79, + 541, + 178 + ], + "type": "table", + "html": "
MethodSFMCPPIPDIPFinAll
Source-only-67.464.263.354.861.8
CC-SSL [25] (CVPR'20)×81.579.974.464.075.1
MDAM [19] (CVPR'21)×82.379.672.361.574.1
RegDA [12] (CVPR'21)×79.674.471.262.972.5
UniFrame [16] (ECCV'22)×86.784.678.968.179.6
RegDA-SF [12] (CVPR'21)✓71.266.866.258.566.9
UniFrame-SF [16] (ECCV'22)✓84.884.277.068.077.9
SHOT [21] (ICML'20)✓77.569.870.664.273.7
MMT [6] (ICLR'20)✓78.774.172.565.375.3
SHOT++ [22] (TPAMI'22)✓85.385.178.267.678.9
Ours✓88.489.280.971.482.2
", + "image_path": "52dd9a0b9ac544c96776f48053da6308df90da21df08677829392b1a06d7019d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 310, + 198, + 541, + 297 + ], + "blocks": [ + { + "bbox": [ + 339, + 189, + 512, + 198 + ], + "lines": [ + { + "bbox": [ + 339, + 189, + 512, + 198 + ], + "spans": [ + { + "bbox": [ + 339, + 189, + 512, + 198 + ], + "type": "text", + "content": "Table 2. PCK@0.05 on RHD " + }, + { + "bbox": [ + 339, + 189, + 512, + 198 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 339, + 189, + 512, + 198 + ], + "type": "text", + "content": " FreiHand Task" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 198, + 541, + 297 + ], + "lines": [ + { + "bbox": [ + 310, + 198, + 541, + 297 + ], + "spans": [ + { + "bbox": [ + 310, + 198, + 541, + 297 + ], + "type": "table", + "html": "
MethodSFMCPPIPDIPFinAll
Source-only-35.250.154.850.746.8
CC-SSL [25] (CVPR'20)×37.448.250.146.543.8
MDAM [19] (CVPR'21)×32.348.151.747.345.1
RegDA [12] (CVPR'21)×40.955.058.253.151.1
UniFrame [16] (ECCV'22)×43.564.067.462.458.5
RegDA-SF [12] (CVPR'21)✓38.652.957.654.349.5
UniFrame-SF [16] (ECCV'22)✓40.662.561.060.255.7
SHOT [21] (ICML'20)✓40.461.360.558.153.3
MMT [6] (ICLR'20)✓39.660.460.057.852.6
SHOT++ [22] (TPAMI'22)✓41.062.862.759.355.8
Ours✓43.765.966.663.158.8
", + "image_path": "b3f8396f972f37fea017451c78bec96bfadfcb237c52d70ddd48daf0fe7cc52e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 310, + 318, + 541, + 398 + ], + "blocks": [ + { + "bbox": [ + 321, + 308, + 529, + 318 + ], + "lines": [ + { + "bbox": [ + 321, + 308, + 529, + 318 + ], + "spans": [ + { + "bbox": [ + 321, + 308, + 529, + 318 + ], + "type": "text", + "content": "Table 3. PCK@0.05 on SURREAL " + }, + { + "bbox": [ + 321, + 308, + 529, + 318 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 321, + 308, + 529, + 318 + ], + "type": "text", + "content": " Human3.6M Task" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 318, + 541, + 398 + ], + "lines": [ + { + "bbox": [ + 310, + 318, + 541, + 398 + ], + "spans": [ + { + "bbox": [ + 310, + 318, + 541, + 398 + ], + "type": "table", + "html": "
MethodSFSldElbWristHipKneeAnkleAll
Source-only-69.475.466.437.977.377.767.3
CC-SSL [25] (CVPR'20)×44.368.555.222.262.357.851.7
MDAM [19] (CVPR'21)×51.783.168.917.779.476.662.9
RegDA [12] (CVPR'21)×73.386.472.854.882.084.475.6
UniFrame [16] (ECCV'22)×78.189.681.152.685.387.179.0
RegDA-SF [12] (CVPR'21)✓70.682.069.843.379.179.471.5
UniFrame-SF [16] (ECCV'22)✓74.384.872.545.981.585.275.1
SHOT [21] (ICML'20)✓73.083.271.946.779.882.473.3
MMT [6] (ICLR'20)✓73.283.572.445.180.883.973.9
SHOT++ [22] (TPAMI'22)✓75.085.276.745.382.885.075.3
Ours✓77.988.880.452.384.286.978.7
", + "image_path": "90284d7370f296641aba911d6091b6a2d056fe8adaf10c16ac4bc652b9cdc4b4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 310, + 419, + 541, + 499 + ], + "blocks": [ + { + "bbox": [ + 337, + 410, + 515, + 419 + ], + "lines": [ + { + "bbox": [ + 337, + 410, + 515, + 419 + ], + "spans": [ + { + "bbox": [ + 337, + 410, + 515, + 419 + ], + "type": "text", + "content": "Table 4. PCK@0.05 on SURREAL " + }, + { + "bbox": [ + 337, + 410, + 515, + 419 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 337, + 410, + 515, + 419 + ], + "type": "text", + "content": " LSP Task" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 419, + 541, + 499 + ], + "lines": [ + { + "bbox": [ + 310, + 419, + 541, + 499 + ], + "spans": [ + { + "bbox": [ + 310, + 419, + 541, + 499 + ], + "type": "table", + "html": "
MethodSFSldElbWristHipKneeAnkleAll
Source-only-51.565.062.968.068.767.463.9
CC-SSL [25] (CVPR'20)×36.866.363.959.667.370.460.7
MDAM [19] (CVPR'21)×61.477.775.565.876.778.369.2
RegDA [12] (CVPR'21)×62.776.771.181.080.375.374.6
UniFrame [16] (ECCV'22)×69.284.983.385.584.784.382.0
RegDA-SF [12] (CVPR'21)✓54.870.567.665.473.270.066.5
UniFrame-SF [16] (ECCV'22)✓68.480.579.182.780.881.078.8
SHOT [21] (ICML'20)✓63.572.766.578.479.773.272.4
MMT [6] (ICLR'20)✓60.970.970.381.179.372.871.5
SHOT++ [22] (TPAMI'22)✓69.581.780.984.082.379.779.9
Ours✓70.785.483.886.685.285.083.2
", + "image_path": "2d963df699c052f16b73f16f014c6dd1ffe71eb29fc2cd94f62c2479efdff703.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": "SURREAL " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " Human3.6M and SURREAL " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " LSP. Tables 1 and 2 show the results for the hand tasks, where our method achieves state-of-the-art performance, surpassing the second best baseline SHOT++ with a significant lead of " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": ", respectively. Moreover, our method outperforms the SHOT++ model by " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " on the PIP metric, respectively. For the human tasks, Tables 3 and 4 display the results. Our model achieves a considerable margin of " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " over the second best baseline SHOT++ for SURREAL " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " Human3.6M and SURREAL " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " LSP, respectively. Specifically, our method outperforms SHOT++ by " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " on the Wrist for the adaptation on Human3.6M and " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "5.3\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " on the Ankle for the adaptation on LSP. Furthermore, our approach remains competitive when compared with general domain adaptive HPE methods that need to access source data, as it surpasses UniFrame by " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "2.6\\%" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4832" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 72, + 294, + 221 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 294, + 221 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 294, + 221 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 294, + 221 + ], + "type": "image", + "image_path": "abaf841689c52e53750234650482f5f71a190ce65115d68e21a368fe06924024.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 220, + 94, + 228 + ], + "lines": [ + { + "bbox": [ + 53, + 220, + 94, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 94, + 228 + ], + "type": "text", + "content": "Source Only" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 100, + 220, + 144, + 228 + ], + "lines": [ + { + "bbox": [ + 100, + 220, + 144, + 228 + ], + "spans": [ + { + "bbox": [ + 100, + 220, + 144, + 228 + ], + "type": "text", + "content": "UniFrame-SF" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 157, + 221, + 187, + 227 + ], + "lines": [ + { + "bbox": [ + 157, + 221, + 187, + 227 + ], + "spans": [ + { + "bbox": [ + 157, + 221, + 187, + 227 + ], + "type": "text", + "content": "SHOT++" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 211, + 221, + 227, + 227 + ], + "lines": [ + { + "bbox": [ + 211, + 221, + 227, + 227 + ], + "spans": [ + { + "bbox": [ + 211, + 221, + 227, + 227 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 72, + 294, + 220 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 72, + 294, + 220 + ], + "spans": [ + { + "bbox": [ + 145, + 72, + 294, + 220 + ], + "type": "image", + "image_path": "3662e08c4f5453cd4b6f55c95d9e5775af76710af0d46db3d430ce27e2113bb2.jpg" + } + ] + } + ], + "index": 3, + "type": "text" + }, + { + "bbox": [ + 195, + 72, + 294, + 220 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 72, + 294, + 220 + ], + "spans": [ + { + "bbox": [ + 195, + 72, + 294, + 220 + ], + "type": "image", + "image_path": "a29ce9a99d43ee6237511a4b46c4e96098db158bcee1961a0c3ffed75f17f82a.jpg" + } + ] + } + ], + "index": 5, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 296, + 72, + 545, + 228 + ], + "blocks": [ + { + "bbox": [ + 296, + 72, + 545, + 228 + ], + "lines": [ + { + "bbox": [ + 296, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 296, + 72, + 545, + 228 + ], + "type": "image", + "image_path": "f7a0788a9588aeda9fde1278823e96c31285fcb4d9ccb89ada1d47ef9174530c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 232, + 428, + 243 + ], + "lines": [ + { + "bbox": [ + 160, + 232, + 428, + 243 + ], + "spans": [ + { + "bbox": [ + 160, + 232, + 428, + 243 + ], + "type": "text", + "content": "Figure 6. Qualitative results on H3D dataset (left) and LSP dataset (right)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 262, + 286, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 286, + 285 + ], + "type": "text", + "content": "on RHD " + }, + { + "bbox": [ + 46, + 262, + 286, + 285 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 262, + 286, + 285 + ], + "type": "text", + "content": " H3D, " + }, + { + "bbox": [ + 46, + 262, + 286, + 285 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 46, + 262, + 286, + 285 + ], + "type": "text", + "content": " on RHD " + }, + { + "bbox": [ + 46, + 262, + 286, + 285 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 262, + 286, + 285 + ], + "type": "text", + "content": " FreiHand, and " + }, + { + "bbox": [ + 46, + 262, + 286, + 285 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 46, + 262, + 286, + 285 + ], + "type": "text", + "content": " on SURREAL " + }, + { + "bbox": [ + 46, + 262, + 286, + 285 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 262, + 286, + 285 + ], + "type": "text", + "content": " LSP." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 287, + 287, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 287, + 347 + ], + "type": "text", + "content": "Qualitative Results. Figs. 6 shows qualitative results of H3D on the left side and LSP on the right side. We use Source only, UniFrame-SF [16], SHOT++ [22], Ours, and Ground Truth for qualitative comparison. It is evident that our method outperforms other baselines significantly." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 356, + 212, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 212, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 212, + 369 + ], + "type": "text", + "content": "4.2. Ablation Study on Framework" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 375, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 287, + 459 + ], + "type": "text", + "content": "Our method contains two modules in Step A & Step B as the source-protect module (SP) and the target-relevant module (TR) separately, and here we focus on their functions. Moreover, we use MMT [6] as the baseline, which is referred in Table 1-4. Table 5 and 6 show the ablation study of frameworks on RHD " + }, + { + "bbox": [ + 46, + 375, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 375, + 287, + 459 + ], + "type": "text", + "content": " H3D and SURREAL " + }, + { + "bbox": [ + 46, + 375, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 375, + 287, + 459 + ], + "type": "text", + "content": " Human3.6M with different combinations of these modules." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 48, + 476, + 287, + 529 + ], + "blocks": [ + { + "bbox": [ + 74, + 467, + 259, + 476 + ], + "lines": [ + { + "bbox": [ + 74, + 467, + 259, + 476 + ], + "spans": [ + { + "bbox": [ + 74, + 467, + 259, + 476 + ], + "type": "text", + "content": "Table 5. Ablation of Frameworks on RHD " + }, + { + "bbox": [ + 74, + 467, + 259, + 476 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 74, + 467, + 259, + 476 + ], + "type": "text", + "content": " H3D" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 476, + 287, + 529 + ], + "lines": [ + { + "bbox": [ + 48, + 476, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 287, + 529 + ], + "type": "table", + "html": "
MethodMCPPIPDIPFinAll
MMT [6]78.774.172.565.375.3
MMT [6] +TR81.377.576.067.779.1
SP+TR (Ours)88.489.280.971.482.2
", + "image_path": "74e37d18fdfec44a04d305e254589b2c0c02955fccd26db99fc6ae0797992137.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 48, + 562, + 287, + 601 + ], + "blocks": [ + { + "bbox": [ + 47, + 552, + 283, + 561 + ], + "lines": [ + { + "bbox": [ + 47, + 552, + 283, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 283, + 561 + ], + "type": "text", + "content": "Table 6. Ablation of Frameworks on SURREAL " + }, + { + "bbox": [ + 47, + 552, + 283, + 561 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 47, + 552, + 283, + 561 + ], + "type": "text", + "content": " Human3.6M" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 562, + 287, + 601 + ], + "lines": [ + { + "bbox": [ + 48, + 562, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 562, + 287, + 601 + ], + "type": "table", + "html": "
MethodSldElbWristHipKneeAnkleAll
MMT [6]73.283.572.445.180.883.973.9
MMT [6] +TR74.486.575.748.981.385.176.2
SP+TR (Ours)77.988.880.452.384.286.978.7
", + "image_path": "274dacdfa8cda0f2546b000acb895bb9cf0253ec67af014b71ee321ddb2e4f32.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "The results clearly demonstrate that both SP and TR contribute to improving the model's performance. Specifically, TR enhances the model's accuracy by " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " on RHD " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " H3D and " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " on SURREAL " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " Human3.6M, while SP leads to an improvement of " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " on RHD " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " H3D and " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " on SURREAL " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " Human3.6M compared to MMT. Notably, the two proposed modules provide similar levels of improvement." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 262, + 493, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 493, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 493, + 275 + ], + "type": "text", + "content": "4.3. Ablation Study on Proposed Losses" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 284, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 545, + 380 + ], + "type": "text", + "content": "We performed a detailed ablation study on the three proposed losses, namely " + }, + { + "bbox": [ + 304, + 284, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{res}}" + }, + { + "bbox": [ + 304, + 284, + 545, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 284, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{cst}}" + }, + { + "bbox": [ + 304, + 284, + 545, + 380 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 284, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{im}}" + }, + { + "bbox": [ + 304, + 284, + 545, + 380 + ], + "type": "text", + "content": ", using the RHD→H3D and SURREAL→Human3.6M tasks. Tables 7 and 8 present the results. As we have previously discussed the functions and advantages of the source-protect and target-relevant modules in Sec. 4.2, we retained " + }, + { + "bbox": [ + 304, + 284, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{ft}}" + }, + { + "bbox": [ + 304, + 284, + 545, + 380 + ], + "type": "text", + "content": " from the source-protect module and " + }, + { + "bbox": [ + 304, + 284, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{con}}" + }, + { + "bbox": [ + 304, + 284, + 545, + 380 + ], + "type": "text", + "content": " from the target-relevant module in the Baseline." + } + ] + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 307, + 403, + 545, + 480 + ], + "blocks": [ + { + "bbox": [ + 343, + 395, + 508, + 404 + ], + "lines": [ + { + "bbox": [ + 343, + 395, + 508, + 404 + ], + "spans": [ + { + "bbox": [ + 343, + 395, + 508, + 404 + ], + "type": "text", + "content": "Table 7. Ablation of Losses on RHD " + }, + { + "bbox": [ + 343, + 395, + 508, + 404 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 343, + 395, + 508, + 404 + ], + "type": "text", + "content": " H3D" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 403, + 545, + 480 + ], + "lines": [ + { + "bbox": [ + 307, + 403, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 403, + 545, + 480 + ], + "type": "table", + "html": "
MethodMCPPIPDIPFinAll
Baseline85.386.878.568.179.4
Lres85.587.979.370.780.1
Lcst85.589.078.470.180.9
Lim85.887.278.769.880.3
Lcst&Lim86.988.880.170.381.4
Lres&Lcst&Lim88.489.280.971.482.2
", + "image_path": "f03393a127d5078c4203a1db4a3f15ad0f7b4782d6eb64b144aa812807b4d5b4.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 307, + 524, + 545, + 583 + ], + "blocks": [ + { + "bbox": [ + 318, + 514, + 533, + 524 + ], + "lines": [ + { + "bbox": [ + 318, + 514, + 533, + 524 + ], + "spans": [ + { + "bbox": [ + 318, + 514, + 533, + 524 + ], + "type": "text", + "content": "Table 8. Ablation of Losses on SURREAL " + }, + { + "bbox": [ + 318, + 514, + 533, + 524 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 318, + 514, + 533, + 524 + ], + "type": "text", + "content": " Human3.6M" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 524, + 545, + 583 + ], + "lines": [ + { + "bbox": [ + 307, + 524, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 307, + 524, + 545, + 583 + ], + "type": "table", + "html": "
MethodSldElbWristHipKneeAnkleAll
Baseline75.486.176.846.583.085.075.8
Lres76.086.977.547.883.385.276.6
Lcst76.687.478.850.683.586.077.2
Lim76.287.178.349.783.285.776.9
Lcst&Lim77.387.779.251.283.686.478.0
Lres&Lcst&Lim77.988.880.452.384.286.978.7
", + "image_path": "bddd39bfb5c679fe09694acb77cd40f98456a27fa4517e524831ce37a112ea65.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": "We observe that each loss is able to boost the model's performance. Simply applying " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{res}" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " leads to a increase of " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " in RHD " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " H3D and " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " in SURREAL " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " Human3.6M. " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cst}" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " causes an improvement of " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " in RHD " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " H3D and " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " in SURREAL " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " Human3.6M. As for " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{im}" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": ", adding it causes a rise of " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " in RHD " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " H3D and " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " in SURREAL " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " Human3.6M. Besides, it is noticed that " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cst}" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " plays a more important role than " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{res}" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{im}" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4833" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 216, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 216, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 216, + 85 + ], + "type": "text", + "content": "4.4. Ablation of Sparsity Reduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 288, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 288, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 288, + 222 + ], + "type": "text", + "content": "In this section, we investigate the effectiveness of our proposed sparsity reduction strategy. Specifically, we compare heatmap-based methods and vector-based methods for two loss functions: contrastive learning " + }, + { + "bbox": [ + 46, + 91, + 288, + 222 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{cst})" + }, + { + "bbox": [ + 46, + 91, + 288, + 222 + ], + "type": "text", + "content": " and information maximization " + }, + { + "bbox": [ + 46, + 91, + 288, + 222 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{im})" + }, + { + "bbox": [ + 46, + 91, + 288, + 222 + ], + "type": "text", + "content": ". We denote heatmap-based contrastive learning and information maximization as HBCL and HBIM, respectively, and vector-based contrastive learning and information maximization as VBCL and VBIM, respectively. Two tasks are selected and the results of these comparisons are presented in Tab. 9 and Tab. 10." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 241, + 287, + 312 + ], + "blocks": [ + { + "bbox": [ + 62, + 231, + 272, + 240 + ], + "lines": [ + { + "bbox": [ + 62, + 231, + 272, + 240 + ], + "spans": [ + { + "bbox": [ + 62, + 231, + 272, + 240 + ], + "type": "text", + "content": "Table 9. Ablation of Sparsity Reduction on RHD " + }, + { + "bbox": [ + 62, + 231, + 272, + 240 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 62, + 231, + 272, + 240 + ], + "type": "text", + "content": " H3D" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 241, + 287, + 312 + ], + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 312 + ], + "type": "table", + "html": "
MethodMCPPIPDIPFinAll
MMT [6]78.774.172.565.375.3
MMT [6] + HBCL77.473.070.765.574.2
MMT [6] + VBCL81.776.674.467.978.7
MMT [6] + HBIM79.374.772.965.875.9
MMT [6] + VBIM80.376.173.666.577.5
", + "image_path": "d22f528f41860fcb00913450ec48e424882d77db71c3a1214095febd6b38a4c3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 356, + 286, + 411 + ], + "blocks": [ + { + "bbox": [ + 47, + 335, + 286, + 356 + ], + "lines": [ + { + "bbox": [ + 47, + 335, + 286, + 356 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 286, + 356 + ], + "type": "text", + "content": "Table 10. Ablation of Sparsity Reduction on SURREAL " + }, + { + "bbox": [ + 47, + 335, + 286, + 356 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 47, + 335, + 286, + 356 + ], + "type": "text", + "content": " Human3.6M" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 356, + 286, + 411 + ], + "lines": [ + { + "bbox": [ + 48, + 356, + 286, + 411 + ], + "spans": [ + { + "bbox": [ + 48, + 356, + 286, + 411 + ], + "type": "table", + "html": "
MethodSldElbWristHipKneeAnkleAll
MMT [6]73.283.572.445.180.883.973.9
MMT [6] + HBCL71.782.071.844.680.282.473.0
MMT [6] + VBCL75.687.075.149.283.386.376.7
MMT [6] + HBIM73.984.672.746.081.183.074.3
MMT [6] + VBIM75.386.474.348.682.485.776.0
", + "image_path": "888a6dbdf919b579c242133023a49ed7e86f6ff9a74f841b7083db7bd61d0e3d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 426, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 571 + ], + "type": "text", + "content": "Based on the results, it is evident that HBCL does not improve the model's performance significantly, whereas VBCL provides significant assistance. For instance, MMT+HBCL shows a decrease of " + }, + { + "bbox": [ + 46, + 426, + 287, + 571 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 46, + 426, + 287, + 571 + ], + "type": "text", + "content": " on the RHD " + }, + { + "bbox": [ + 46, + 426, + 287, + 571 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 426, + 287, + 571 + ], + "type": "text", + "content": " H3D task, while MMT+VBCL exhibits an increase of " + }, + { + "bbox": [ + 46, + 426, + 287, + 571 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 46, + 426, + 287, + 571 + ], + "type": "text", + "content": ". Additionally, while HBIM can enhance the model's performance, it falls short of surpassing VBIM. For example, MMT+HBIM shows an increase of " + }, + { + "bbox": [ + 46, + 426, + 287, + 571 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 46, + 426, + 287, + 571 + ], + "type": "text", + "content": " on the RHD " + }, + { + "bbox": [ + 46, + 426, + 287, + 571 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 426, + 287, + 571 + ], + "type": "text", + "content": " H3D task, whereas MMT+VBIM achieves an increase of " + }, + { + "bbox": [ + 46, + 426, + 287, + 571 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 46, + 426, + 287, + 571 + ], + "type": "text", + "content": ". In summary, it becomes evident that utilizing the vector-based spatial probability space is crucial for achieving optimal results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 579, + 162, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 162, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 162, + 593 + ], + "type": "text", + "content": "4.5. Parameter Analysis" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 47, + 607, + 286, + 688 + ], + "blocks": [ + { + "bbox": [ + 47, + 607, + 286, + 688 + ], + "lines": [ + { + "bbox": [ + 47, + 607, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 47, + 607, + 286, + 688 + ], + "type": "image", + "image_path": "6c32538c30e8e130cc9a8da2626a57ea8de41f1cd745fad740f5e9ac387b5ede.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 689, + 286, + 712 + ], + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 712 + ], + "type": "text", + "content": "Figure 7. Parameter Analysis on RHD " + }, + { + "bbox": [ + 46, + 689, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 689, + 286, + 712 + ], + "type": "text", + "content": " H3D (best viewed in color). a: Analysis on " + }, + { + "bbox": [ + 46, + 689, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 689, + 286, + 712 + ], + "type": "text", + "content": ". b: Analysis on " + }, + { + "bbox": [ + 46, + 689, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 689, + 286, + 712 + ], + "type": "text", + "content": ". c: Analysis on " + }, + { + "bbox": [ + 46, + 689, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 689, + 286, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "We use RHD " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " H3D to illustrate the sensitivity of " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " in Equation 5 and " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " in Equation 11. The results are shown in Fig. 7. From it, we observe that " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\alpha = 0.7" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\beta = 0.5" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\gamma = 0.85" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " is the best choice. Moreover, it is noticed that the decreases from different parameters are limited to " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ". Therefore, our method shows stable performance over hyperparameters." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 172, + 545, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 172, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 305, + 172, + 545, + 198 + ], + "type": "text", + "content": "5. Qualitative Results on the Unseen Dataset COCO" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 306, + 219, + 545, + 335 + ], + "blocks": [ + { + "bbox": [ + 306, + 219, + 545, + 335 + ], + "lines": [ + { + "bbox": [ + 306, + 219, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 219, + 545, + 335 + ], + "type": "image", + "image_path": "649bf5e302635bb29057003cbc74098a117511111957a00fef641b083dcc4fc3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 337, + 545, + 370 + ], + "lines": [ + { + "bbox": [ + 305, + 337, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 305, + 337, + 545, + 370 + ], + "type": "text", + "content": "Figure 8. Qualitative results on the COCO dataset. Here we compare our method trained from SURREAL " + }, + { + "bbox": [ + 305, + 337, + 545, + 370 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 305, + 337, + 545, + 370 + ], + "type": "text", + "content": " LSP with the source-only model." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 388, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 555 + ], + "type": "text", + "content": "In this section, we showcase qualitative results on the COCO dataset [23], which was not utilized in our previous human dataset studies. It's important to note that our 14-keypoint poses are annotated differently from the 17-keypoint poses in COCO, which hinders providing quantitative results on COCO using its own metrics or APIs. Despite this limitation, we believe that our qualitative findings hold significant value. For our analysis, we leverage the model trained on the SURREAL " + }, + { + "bbox": [ + 304, + 388, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 388, + 545, + 555 + ], + "type": "text", + "content": " LSP task in conjunction with the pre-trained SURREAL model. The results, presented in Fig. 8, demonstrate the model's capacity to generalize effectively to unseen datasets, implying that the adaptation process boosts the model's generalization ability beyond its original training dataset." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 571, + 378, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 571, + 378, + 583 + ], + "spans": [ + { + "bbox": [ + 306, + 571, + 378, + 583 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we propose a new task named source-free domain adaptive human pose estimation that places an emphasis on the privacy of source data. Additionally, we propose a new framework that includes source-protect and target-relevant modules, which aim to alleviate the issues of catastrophic forgetting of source and the sparsity of spatial distributions, respectively. Our approach is evaluated on hand and human pose datasets through extensive experiments, demonstrating that it outperforms state-of-the-art methods by a considerable margin." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4834" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 295-305, June 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 287, + 179 + ], + "type": "text", + "content": "[2] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 235 + ], + "type": "text", + "content": "[3] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7103-7112, 2018." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 290 + ], + "type": "text", + "content": "[4] Wanxia Deng, Qing Liao, Lingjun Zhao, Deke Guo, Gangyao Kuang, Dewen Hu, and Li Liu. Joint clustering and discriminative feature alignment for unsupervised domain adaptation. IEEE Transactions on Image Processing, pages 7842-7855, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 287, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 324 + ], + "type": "text", + "content": "[5] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International Conference on Machine Learning, pages 1180-1189, 2015." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 325, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 287, + 369 + ], + "type": "text", + "content": "[6] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual meanteaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In International Conference on Learning Representations, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "text", + "content": "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 287, + 456 + ], + "type": "text", + "content": "[8] Zhongyi Han, Haoliang Sun, and Yilong Yin. Learning transferable parameters for unsupervised domain adaptation. IEEE Transactions on Image Processing, 31:6424-6439, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 458, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 287, + 502 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 502, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 546 + ], + "type": "text", + "content": "[10] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 1501-1510, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "type": "text", + "content": "[11] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, 2013." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "text", + "content": "[12] Junguang Jiang, Yifei Ji, Ximei Wang, Yufeng Liu, Jianmin Wang, and Mingsheng Long. Regressive domain adaptation for unsupervised keypoint detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6780-6789, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "type": "text", + "content": "[13] Ruijie Jiang, Thuan Nguyen, Prakash Ishwar, and Shuchin Aeron. Supervised contrastive learning with hard negative samples. arXiv preprint arXiv:2209.00078, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[14] Rui Jin, Jing Zhang, Jianyu Yang, and Dacheng Tao. Multibranch adversarial regression for domain adaptive hand" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "pose estimation. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):6125-6136, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[15] Sam Johnson and Mark Everingham. Clustered pose and nonlinear appearance models for human pose estimation. In British Machine Vision Conference, volume 2, page 5. Cite-seer, 2010." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "text", + "content": "[16] Donghyun Kim, Kaihong Wang, Kate Saenko, Margrit Betke, and Stan Sclaroff. A unified framework for domain adaptive pose estimation. arXiv preprint arXiv:2204.00172, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 186, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 218 + ], + "type": "text", + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 220, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 275 + ], + "type": "text", + "content": "[18] Vinod K Kurmi, Venkatesh K Subramanian, Vinay P Namboodiri, , and . Domain impression: A source data free domain adaptation method. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 615-625, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 276, + 545, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 320 + ], + "type": "text", + "content": "[19] Chen Li and Gim Hee Lee. From synthetic to real: Unsupervised domain adaptation for animal pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1482-1491, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 322, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 322, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 322, + 545, + 374 + ], + "type": "text", + "content": "[20] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9638-9647, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 376, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 421 + ], + "type": "text", + "content": "[21] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International Conference on Machine Learning, pages 6028-6039. PMLR, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 422, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 477 + ], + "type": "text", + "content": "[22] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8602-8617, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 478, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 533 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, pages 740-755. Springer, 2014." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "type": "text", + "content": "[24] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International Conference on Machine Learning, pages 97-105, 2015." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 579, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 624 + ], + "type": "text", + "content": "[25] Jiteng Mu, Weichao Qiu, Gregory D Hager, and Alan L Yuille. Learning from synthetic animals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12386-12395, 2020." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "type": "text", + "content": "[26] Qucheng Peng, Zhengming Ding, Lingjuan Lyu, Lichao Sun, and Chen Chen. Toward better target representation for source-free and black-box domain adaptation. arXiv preprint arXiv:2208.10531, 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "text", + "content": "[27] Ekkasit Pinyoanuntapong, Ayman Ali, Kalvik Jakkala, Pu Wang, Minwoo Lee, Qucheng Peng, Chen Chen, and Zhi Sun. Gaitsada: Self-aligned domain adaptation for mmwave gait recognition. arXiv preprint arXiv:2301.13384, 2023." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4835" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[28] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3723-3732, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 174 + ], + "type": "text", + "content": "[29] Dahu Shi, Xing Wei, Liangqi Li, Ye Ren, and Wenming Tan. End-to-end multi-person pose estimation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11069-11078, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "type": "text", + "content": "[30] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5693-5703, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 275 + ], + "type": "text", + "content": "[31] Hui Tang, Ke Chen, and Kui Jia. Unsupervised domain adaptation via structurally regularized deep clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8725-8735, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "type": "text", + "content": "[32] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in Neural Information Processing Systems, 30, 2017." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "type": "text", + "content": "[33] Jonathan Tompson, Ross Goroshin, Arjun Jain, Yann LeCun, and Christoph Bregler. Efficient object localization using convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 648-656, 2015." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "type": "text", + "content": "[34] Jonathan J. Thompson, Arjun Jain, Yann LeCun, and Christoph Bregler. Joint training of a convolutional network and a graphical model for human pose estimation. Advances in neural information processing systems, 27, 2014." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 455 + ], + "type": "text", + "content": "[35] Eric Tzeng, Judy Hoffman, Ning Zhang, Kate Saenko, and Trevor Darrell. Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:1412.3474, 2014." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 511 + ], + "type": "text", + "content": "[36] Gul Varol, Javier Romero, Xavier Martin, Naureen Mahmood, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning from synthetic humans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 109-117, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "text", + "content": "[37] Haixin Wang, Lu Zhou, Yingying Chen, Ming Tang, and Jinqiao Wang. Regularizing vector embedding in bottom-up human pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI, pages 107-122. Springer, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "type": "text", + "content": "[38] Haifeng Xia, Handong Zhao, and Zhengming Ding. Adaptive adversarial network for source-free domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9010-9019, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "type": "text", + "content": "[39] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In Proceedings of the European Conference on Computer Vision (ECCV), pages 466-481, 2018." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[40] Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, and Ce Zhu. Distribution-aware coordinate representation for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7093-7102, 2020." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 229 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "type": "text", + "content": "[41] Zhengyi Zhao, Tianyao Wang, Siyu Xia, and Yangang Wang. Hand-3d-studio: A new multi-view system for 3d hand reconstruction. In ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2478-2482, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "type": "text", + "content": "[42] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In Proceedings of the IEEE International Conference on Computer Vision, pages 4903-4911, 2017." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 174, + 546, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 546, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 546, + 229 + ], + "type": "text", + "content": "[43] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 813-822, 2019." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "4836" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Space Engage_ Collaborative Space Supervision for Contrastive-Based Semi-Supervised Semantic Segmentation/5953e602-104c-4dc8-b226-b90796937563_content_list.json b/2023/Space Engage_ Collaborative Space Supervision for Contrastive-Based Semi-Supervised Semantic Segmentation/5953e602-104c-4dc8-b226-b90796937563_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..89c545186cee19431d87095261a823c2f7fcd069 --- /dev/null +++ b/2023/Space Engage_ Collaborative Space Supervision for Contrastive-Based Semi-Supervised Semantic Segmentation/5953e602-104c-4dc8-b226-b90796937563_content_list.json @@ -0,0 +1,2034 @@ +[ + { + "type": "text", + "text": "Space Engage: Collaborative Space Supervision for Contrastive-based Semi-Supervised Semantic Segmentation", + "text_level": 1, + "bbox": [ + 130, + 130, + 839, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Changqi Wang $^{1*}$ , Haoyu Xie $^{1*}$ , Yuhui Yuan $^{2}$ , Chong Fu $^{1,4\\dagger}$ , Xiangyu Yue $^{3}$", + "bbox": [ + 197, + 202, + 769, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Engineering, Northeastern University, Shenyang, China \n $^{2}$ Microsoft Research Asia $^{3}$ The Chinese University of Hong Kong", + "bbox": [ + 125, + 227, + 841, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Key Laboratory of Intelligent Computing in Medical Image, Ministry of Education, NEU, China", + "bbox": [ + 98, + 263, + 870, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2101668@stu.neu.edu.cn, xiehaoyu@stumail.neu.edu.cn, yuyua@microsoft.com", + "bbox": [ + 163, + 292, + 805, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fuchong@mail.neu.edu.cn, xyyue@ie.cuhk.edu.hk", + "bbox": [ + 284, + 310, + 684, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 359, + 313, + 376 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semi-Supervised Semantic Segmentation (S4) aims to train a segmentation model with limited labeled images and a substantial volume of unlabeled images. To improve the robustness of representations, powerful methods introduce a pixel-wise contrastive learning approach in latent space (i.e., representation space) that aggregates the representations to their prototypes in a fully supervised manner. However, previous contrastive-based S4 methods merely rely on the supervision from the model's output (logits) in logit space during unlabeled training. In contrast, we utilize the outputs in both logit space and representation space to obtain supervision in a collaborative way. The supervision from two spaces plays two roles: 1) reduces the risk of over-fitting to incorrect semantic information in logits with the help of representations; 2) enhances the knowledge exchange between the two spaces. Furthermore, unlike previous approaches, we use the similarity between representations and prototypes as a new indicator to tilt training those under-performing representations and achieve a more efficient contrastive learning process. Results on two public benchmarks demonstrate the competitive performance of our method compared with state-of-the-art methods.", + "bbox": [ + 75, + 390, + 473, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 750, + 209, + 766 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic segmentation is a fundamental task in computer vision, aiming to classify each pixel in an image. Significant progress [22, 4] has been made in training on high-quality labeled images using segmentation models composed of an encoder and a segmentation head. However, annotating images is expensive and time-consuming. Semi", + "bbox": [ + 75, + 776, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/59bfb879373b542a433ac5a4aaa2599ca463101edce0c030da9bf2c5369d8d07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 359, + 888, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eeb11bfac1f0b1f95d9c58b267eceba51bb7932527e4ba689ef3e2719777a707.jpg", + "image_caption": [ + "Figure 1. We enhance the knowledge exchange between the logit and representation spaces. Orange and blue represent different classes. Top: Existing contrastive-based S4 methods overlook the semantic information in representation space. Bottom: Our method uses dual-space collaborative supervision." + ], + "image_footnote": [], + "bbox": [ + 504, + 474, + 887, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "supervised Semantic Segmentation (S4) alleviates the thirst for annotation by leveraging unlabeled images to train segmentation models.", + "bbox": [ + 498, + 715, + 892, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most existing works learn from unlabeled images via self-training [44, 37, 15] or consistency regularization [40, 38, 13] strategies, both of which retrain the model with its predictions on unlabeled images. Recently, great success has been achieved by introducing pixel-wise contrastive learning to semantic segmentation, which endows the model with a stronger features-extracting ability by accessing a more discriminative representation space. Specifically, these methods [45, 23] project each pixel to repre", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*equal contribution", + "bbox": [ + 94, + 875, + 202, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ corresponding author", + "bbox": [ + 96, + 887, + 217, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "931", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sensation space as a representation and regularize it in a fully supervised manner, i.e., aggregating the representations with the same class and separating them with different classes. In semi-supervised settings, due to limited labels, most methods [1, 32, 46] obtain supervision from the model's output logits in logit space during the unlabeled training process. However, recent contrastive-based semantic segmentation methods [1, 32, 46, 45] mainly focus on the learning process in logit space while only taking that in representation space as an auxiliary task. The unidirectional supervision makes training dominated by the predicted logits, leading to the neglect of information in the representation space. We argue that this kind of single-space supervision may incorrectly provide semantic guidance to representation learning and fails to facilitate knowledge exchange between the two spaces (see Sec. 5.1).", + "bbox": [ + 80, + 90, + 467, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we extend the single-space supervision to a dual-space supervision for contrastive-based S4 and propose Collaborative Space Supervision (CSS). Our key insight is to: i) utilize the semantic information in representations to obtain more reliable guidance during unlabeled training, and to enhance knowledge exchange between two spaces; ii) provide a more accurate reference for the model's performance on predicting each representation to tilt training those under-performing representations. To achieve objective i), we obtain dense semantic predictions by retrieving the nearest class prototype for each representation in the representation space and then engage with predictions from the logits for collaborative supervision of the model. For objective ii), we measure the similarity between the representations and prototypes and use the similarity after a normalization operation as the indicator for guiding the learning process in the representation space. Unlike previous works that utilize confidence as the indicator to involve representation learning, the similarity directly reflects the confusion level between representations and prototypes, resulting in more efficient representation learning.", + "bbox": [ + 80, + 333, + 467, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our main contributions are three-fold: 1) The dual-space collaboration for contrastive-based S4, enhances the knowledge exchange between the logit and representation spaces. 2) Utilizing similarity to provide a more accurate reference for the model's performance in representation learning. 3) Extensive experiments on two S4 benchmarks demonstrate the effectiveness of our method.", + "bbox": [ + 80, + 651, + 467, + 756 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 80, + 773, + 223, + 789 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Semi-supervised Semantic Segmentation", + "text_level": 1, + "bbox": [ + 80, + 800, + 421, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The aim of S4 is to train a segmentation model with the semi-supervised setting (i.e., a few labeled images and a large number of unlabeled images) to classify each pixel in an entire image. The critical issue of S4 is how to leverage unlabeled images to train the model. Some methods [25,", + "bbox": [ + 80, + 824, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "27, 30, 35] based on GANs [16], adversarial training [36], and consistency regularization paradigm [38, 13, 40, 7, 56]. Meanwhile, self-training [28, 44, 49, 61, 53] is also a striking paradigm, which always generates pseudo-labels from model and retransmits the model with the combined supervision of human annotations and pseudo-labels. One essential issue of self-training is the accuracy of pseudo-labels. Some methods [29, 33, 14, 42, 52] try to polish pseudo-labels and provide reliable guidance. Some methods [21, 47, 24, 17] focus on the class-imbalance problems in the dataset and try to alleviate the negative effect from class-biased pseudolabels generated by the model pre-trained on imbalanced labeled images. We build our framework based on the self-training and additionally explore semantic information among different images.", + "bbox": [ + 501, + 90, + 888, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Pixel-wise Contrastive Learning", + "text_level": 1, + "bbox": [ + 503, + 328, + 779, + 344 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pixel-wise contrastive learning explores semantic relations not only in the individual image but also among different images. Different from instance-wise contrastive learning [19, 6, 3], pixel-wise contrastive learning [50, 55, 5, 58] project each pixel to the representation in representation space with the cooperation of encoder and representation head. Representations are then aggregated in their prototypes and are separated from each other in different classes. In semi-supervised settings, most methods [1, 32, 46, 48] use pseudo-labels based on logits to provide semantic information contrastive learning process during training on unlabeled images. Meanwhile, the confidence of logit is used as an indicator to involve the contrastive learning process, e.g., [32] uses the hard representations whose corresponding logit confidence is lower than a threshold to contrast for effective training. As opposed to the above methods, we use collaborative space supervision for contrastive learning on unlabeled images and use a new indicator to involve the contrastive learning progress.", + "bbox": [ + 501, + 352, + 888, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Prototype-based Learning", + "text_level": 1, + "bbox": [ + 503, + 650, + 735, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prototype-based learning has been widely studied in few-shot learning [41, 11, 31, 34] and unsupervised domain adaption [54, 26, 43, 39, 57]. Recently, it is restudied in semantic segmentation as known as a non-parametric prototype-based classifier [60]. Concretely, the classes in the dataset are presented by a set of non-learnable prototypes, and the dense semantic predictions are thus achieved by assigning the output features to its most similar prototype. Under semi-supervised settings, some methods [51] maintain the consistency between predictions from a linear predictor and a prototype-based predictor. The two predictors are followed by the encoder and project the features to logit space and representation space, respectively. In this work, we combine the semantic information in the logit and representation spaces to provide supervision in a collabora", + "bbox": [ + 501, + 674, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "932", + "bbox": [ + 486, + 946, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tive way during semi-supervised learning.", + "bbox": [ + 76, + 90, + 354, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 122, + 210, + 138 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the S4 task, we have a small labeled set $\\mathcal{D}_l = \\{(\\pmb{x}_i^l,\\pmb{y}_i^l)\\}_{i=1}^{N_l}$ and a large unlabeled set $\\mathcal{D}_u = \\{\\pmb{x}_i^u\\}_{i=1}^{N_u}$ , where $\\pmb{x}_i^l,\\pmb{x}_i^u\\in \\mathbb{R}^{H\\times W\\times 3}$ , $H,W$ denote the height, and the width, respectively. And ground truth $\\pmb{y}_i^l\\in \\{0,1\\}^{H\\times W\\times |C|}$ with the set of class $C$ . The goal is to boost model performance with $\\mathcal{D}_u$ . The base model consists of an encoder $f(\\cdot)$ and a segmentation head $g(\\cdot)$ , which projects features to the logit space $\\mathbb{R}^{H\\times W\\times |C|}$ . We adopt Self-Training (ST) and pixel-wise contrastive learning to our framework, as described in Sec. 3.1. The supervision for $\\mathcal{D}_u$ is produced by the collaboration between the logit and representation spaces, as described in Sec. 3.3.", + "bbox": [ + 76, + 148, + 472, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. ST and Pixel-wise Contrastive Learning", + "text_level": 1, + "bbox": [ + 76, + 340, + 419, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main idea of self-training is to pre-train a model on labeled images and use it to produce pseudo-labels as supervision for unlabeled images. A typical framework is the teacher-student framework [44], which consists of a student model and a teacher model. Both the student model and the teacher model are constructed by an encoder and a segmentation head. Parameters of the student model are optimized via Stochastic Gradient Decent (SGD) while parameters of the teacher model are updated by the Exponential Moving Average (EMA) of student model parameters. We denote the encoder and the segmentation head in the student model by $f(\\cdot)$ and $g(\\cdot)$ while denoting those in the teacher model by $f^{\\prime}(\\cdot)$ and $g^{\\prime}(\\cdot)$ . The pseudo-labels $\\hat{\\pmb{y}}_i^{u,lgt}$ are produced based on the teacher model's output logits $\\hat{\\pmb{p}}_i^u = g'(f'(x_i^u))$ in logit space, formulated as:", + "bbox": [ + 75, + 364, + 472, + 594 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {y}} _ {i} ^ {u, l g t} = \\mathbf {1} _ {c} (\\underset {c} {\\arg \\max } \\{\\hat {\\boldsymbol {p}} _ {i, c} ^ {u} \\} _ {c \\in C}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 606, + 468, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{1}_c(\\cdot)$ denotes the one-hot encoding of class $c$ .", + "bbox": [ + 76, + 643, + 423, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to enhance the ability of the model itself to extract features, recent works [32, 46, 1] additionally employ pixel-wise contrastive learning and introduce a representation head to both teacher and student models. We denote the representation head in the student model as $h(\\cdot)$ and that in the teacher model as $h^{\\prime}(\\cdot)$ . The pixel $x_{i}$ of the class $c$ are projected as representations $z_{ci}$ in representation space by the cooperating of $f(\\cdot)$ and $h(\\cdot)$ , i.e., $z_{ci} = h(f(x_i))$ . And the representation $z_{ci}$ is then aggregated to its class centroid (prototype) while separated from representations in different classes $z_{\\tilde{c} i}$ (negatives). The semantic guidance for contrastive learning is from the combination of ground truth $y_{i}^{l}$ and pseudo-labels $\\hat{y}_i^{u,lgt}$ in logit space. Moreover, in order to emphasize the reliable and crucial aspects during unlabeled and contrastive learning, a sampling strategy is adopted to select valid pixels $x_{i}$ according to their corre", + "bbox": [ + 75, + 659, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sponding confidence, i.e. the student model's output logits $\\pmb{p}_i$ after a Softmax operation.", + "bbox": [ + 496, + 90, + 890, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Discussion. In recent works [46, 1, 32], the supervision of unlabeled images is derived solely from the logit space. This overlooks the potential benefits of the supervision from the representation space, leading to two potential limitations: 1) the pseudo-labels $\\hat{y}_i^{u,lgt}$ obtained from the logit space may contain noise and miss the opportunity to be corrected by semantic information from the representation space; 2) since the confidence from logit space is used as the indicator $\\hat{j}_i$ for the sampling strategy, learning in the representation space may not be critical enough due to the different confusing parts between the two spaces.", + "bbox": [ + 496, + 121, + 893, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To mitigate these limitations, we produce pseudo-labels from the representation space and combine them with pseudo-labels from the logit space to provide higher-quality supervision during unlabeled training. Meanwhile, we obtain a new indicator from the representation space for the more effective sampling strategy.", + "bbox": [ + 496, + 287, + 893, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Supervision from Representation Space", + "text_level": 1, + "bbox": [ + 498, + 386, + 839, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we detail the approach to obtain the pseudo-labels from the representation space. Meanwhile, we simultaneously access a new indicator for the sampling strategy in representation spaces, which provides a critical reference in the contrastive learning process.", + "bbox": [ + 496, + 409, + 890, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we first build a set of prototypes for each class and obtain the pseudo-labels by retrieving the nearest prototype for each representation. We calculate the centroid of all representations in the current class $c$ as the prototype $\\rho_{c}$ , which is formulated as:", + "bbox": [ + 496, + 484, + 890, + 560 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {c} = \\frac {1}{N _ {c}} \\sum_ {i} ^ {N _ {c}} z _ {i} ^ {\\prime}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 569, + 890, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N_{c}$ is the total number of representations of current class $c$ and $\\pmb{z}_i^{\\prime}$ is the representation projected by the cooperation of the $f^{\\prime}(\\cdot)$ and $h^\\prime (\\cdot)$ . Meanwhile, to include more representation information, we update the prototype along the sequential iterations with EMA as follows:", + "bbox": [ + 496, + 614, + 890, + 689 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\rho} _ {c} (t) = \\alpha \\hat {\\rho} _ {c} (t - 1) + (1 - \\alpha) \\rho_ {c} (t), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 700, + 890, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{\\rho}_c(t),\\hat{\\rho}_c(t - 1)$ mean the current $t^{th}$ prototype and last $(t - 1)^{th}$ prototype in iterations, $\\rho_{c}(t)$ means the prototype calculated by Eq.2 in current iteration, and $\\alpha$ is a hyper-parameter that controls the updating speed. Thus, the pseudo-label from the representation space is achieved by:", + "bbox": [ + 496, + 724, + 890, + 801 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {y}} _ {i} ^ {u, r e p} = \\mathbf {1} _ {\\hat {c}} (\\hat {c}), w i t h \\hat {c} = \\underset {c} {\\arg \\max } \\left\\{\\operatorname {s i m} \\left(\\boldsymbol {z} _ {i} ^ {\\prime}, \\hat {\\boldsymbol {\\rho}} _ {c} (t)\\right) \\right\\} _ {c \\in C}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 808, + 890, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $sim(\\cdot)$ is defined as the cosine similarity.", + "bbox": [ + 496, + 839, + 818, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As for the indicator for the sampling strategy in the representation space, we use the Softmax function on the similarity among the representation and all prototypes, which is", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "933", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9460f85ac9f86af3b5a8ad3d7cde59fdaa955bb3d1ce27cd70bd3f30f15d9484.jpg", + "image_caption": [ + "Figure 2. Overview of our framework. Our training pipeline consists of learning in two spaces: logit space and representation space. The pseudo-labels $\\hat{\\pmb{y}}_i^u$ during unlabeled training are produced by the collaboration of two spaces with mix pseudo-labeling strategy (1) or cross pseudo-labeling strategy (2). The indicator for representation learning is produced by similarity $(s_1, s_2, \\text{and} s_3)$ ." + ], + "image_footnote": [], + "bbox": [ + 122, + 89, + 846, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "followed as:", + "bbox": [ + 76, + 406, + 161, + 420 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {j} _ {i} ^ {u, r e p} = \\frac {e ^ {s i m \\left(\\boldsymbol {z} _ {\\boldsymbol {c} i} , \\hat {\\rho} _ {\\boldsymbol {c}} (t)\\right) / \\tau}}{e ^ {s i m \\left(\\boldsymbol {z} _ {\\boldsymbol {c} i} , \\hat {\\rho} _ {\\boldsymbol {c}} (t)\\right) / \\tau} + \\sum_ {\\tilde {c} \\in \\tilde {C}} e ^ {s i m \\left(\\boldsymbol {z} _ {\\boldsymbol {c} i} , \\hat {\\rho} _ {\\tilde {c}} (t)\\right) / \\tau}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 428, + 468, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\rho}_{\\tilde{c}}(t)$ means the prototype with different classes from $z_{ci}$ and $\\tau$ is a hyper-parameter. Different from using confidence from logit space as the indicator to involve representation learning [32, 46], the Softmax similarity directly helps the model to discover the confusion between representations and their prototypes and focus on them during the subsequent training.", + "bbox": [ + 75, + 470, + 468, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Collaboration Between Two Spaces", + "text_level": 1, + "bbox": [ + 76, + 585, + 382, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the pseudo-labels in two spaces, we propose two pseudo-labeling strategies to strengthen the collaboration between two spaces and obtain more reliable pseudo-labels.", + "bbox": [ + 75, + 609, + 468, + 656 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Mix pseudo-labeling. To mitigate the negative effects of inherent noise from both two spaces during pseudolabeling, we adopt the mix pseudo-labeling strategy that only considers the mutually agreeable pseudolabels between the two spaces. Specifically, we define the set of final pseudo-labels as $\\hat{Y}^u = \\hat{Y}^{u,lgt} \\cap \\hat{Y}^{u,rep}$ , where $\\hat{\\mathbf{y}}_i^{u,lgt} \\in \\hat{Y}^{u,lgt}$ and $\\hat{\\mathbf{y}}_i^{u,rep} \\in \\hat{Y}^{u,rep}$ .", + "- Cross pseudo-labeling. Inspired by recent researches [7, 38] that maintain consistency among the predictions of the same image across different models or decoders in different views, we propose a cross pseudolabeling strategy that leverages pseudo-labels from one space to supervise the other. Specifically, we use pseudo-labels $\\hat{\\pmb{y}}_i^{u,rep}$ to supervise the logit space, and vice versa." + ], + "bbox": [ + 94, + 664, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The strengths of using pseudo-labels from two spaces in the collaborative way are twofold: 1) obtaining more reliable supervision during unlabeled training, and 2) enabling the strengths of learning in different spaces to complement each other. Since the learning in different feature spaces concentrates on different parts of features, i.e., the logit space mainly focuses on the most discriminative part of features while the representation space treats all parts of features equally, the performance of pseudo-labels from two spaces varies across different classes or regions of images. Therefore, our collaborative pseudo-labeling strategies exchange knowledge between two spaces and provide higher-quality supervision during unlabeled training. The experimental proof is in Sec. 5.1.", + "bbox": [ + 496, + 406, + 890, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As for indicators, we use confidence as the indicator $\\hat{j}_i^{u,lgt}$ for learning logit space and Softmax similarity as the indicator $\\hat{j}_i^{u,rep}$ for learning representation space. We argue that the confusing parts of learning in both two spaces are varied due to the different parts of features being concentrated in each space. Therefore, this strategy allows the learning in different spaces to focus on their own confusing parts, which can be more effective than using a single indicator when mining confusing parts for both spaces in the training process. The experimental proof is in Sec. 5.2.", + "bbox": [ + 496, + 619, + 892, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Training Objective", + "text_level": 1, + "bbox": [ + 500, + 784, + 681, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the indicators $\\hat{j}_i^{u,lgt}$ and $\\hat{j}_i^{u,rep}$ , we adopt some threshold sampling strategies. In logit space, we set a threshold $\\delta_u$ during unlabeled learning and logits $\\hat{p}_i^u$ whose indicator $\\hat{j}_i^{u,lgt}$ is higher than $\\delta_u$ will be regarded as the valid logits in logit space. In representation space, our sample strategy can be divided into three parts: 1) Valid Sam", + "bbox": [ + 496, + 806, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "934", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pling Strategy. Similar to the sampling strategy in logit space, a threshold $\\delta_w$ is used to sample representations whose indicator $\\hat{j}_i^{u,rep}$ is higher than $\\delta_w$ . 2) Hard Sampling Strategy. We adopt the hard sampling strategy for tilting to train those confusing representations. Specifically, we set a threshold $\\delta_s$ to sample representations whose indicator $\\hat{j}_i^{u,rep}$ is lower than $\\delta_s$ . 3) Negative Sampling Strategy. We sample negatives according to the similarity between the prototype of current class $c$ and other prototypes. Concretely, the negatives are more likely to be sampled if its prototype is more similar to the prototype of the current class.", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cooperated with the ground truth $\\pmb{y}_i^l$ , pseudo-labels $\\pmb{y}_i^u$ produced from two spaces in a collaborative way, and different sampling strategies in two spaces, the total learning object is composed with a supervised loss $\\mathcal{L}_s$ , an unsupervised loss $\\mathcal{L}_u$ , and a contrastive loss $\\mathcal{L}_c$ as follows:", + "bbox": [ + 76, + 272, + 468, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {s} + \\mathcal {L} _ {u} + \\lambda_ {c} \\mathcal {L} _ {c}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 363, + 468, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{c}$ is used to tune the contribution between logit space and representation space. Specifically, $\\mathcal{L}_s$ and $\\mathcal{L}_u$ are constructed by the Cross-Entropy (CE) $\\ell_{ce}$ and can be formulated as:", + "bbox": [ + 76, + 392, + 472, + 453 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\frac {1}{| \\mathcal {B} _ {l} |} \\sum_ {\\boldsymbol {x} _ {i} ^ {l} \\in \\mathcal {B} _ {l}} \\ell_ {c e} \\left(\\boldsymbol {p} _ {i} ^ {l}, \\boldsymbol {y} _ {i} ^ {l}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 462, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {u} = \\frac {1}{| \\hat {\\mathcal {B}} _ {u} |} \\sum_ {\\boldsymbol {x} _ {i} ^ {u} \\in \\hat {\\mathcal {B}} _ {u}} \\ell_ {c e} \\left(\\boldsymbol {p} _ {i} ^ {u}, \\hat {\\boldsymbol {y}} _ {i} ^ {u}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 513, + 468, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{B}_l$ denotes the labeled images in a mini-batch and $\\hat{\\mathcal{B}}_u$ is the subsets that sampled from unlabeled images in a mini-batch according to the sampling strategy. Meanwhile, the contrastive loss $\\mathcal{L}_c$ is formulated as:", + "bbox": [ + 76, + 560, + 468, + 621 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = - \\frac {1}{| C | \\times | \\hat {\\mathcal {Z}} _ {c} |} \\sum_ {c \\in C} \\sum_ {\\boldsymbol {z} _ {c i} \\in \\hat {\\mathcal {Z}} _ {c}} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 631, + 468, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl o g [ \\frac {e ^ {s (\\pmb {z} _ {c i} , \\hat {\\pmb {\\rho}} _ {c}) / \\tau}}{e ^ {s (\\pmb {z} _ {c i} , \\hat {\\pmb {\\rho}} _ {c} (t))) / \\tau} + \\sum_ {\\tilde {e} \\in \\tilde {C}} \\sum_ {\\pmb {z} _ {\\tilde {e} i} \\in \\hat {\\pmb {Z}} _ {\\tilde {c}}} e ^ {s (\\pmb {z} _ {c i} , \\pmb {z} _ {\\tilde {c} i}) / \\tau} ],\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 670, + 434, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathcal{Z}}_c$ is the subset sampled from the set of the representations which belong to class $c$ according to the sampling strategy, $\\hat{\\mathcal{Z}}_{\\tilde{c}}$ is the subset sampled from the set of the representations which bot belong to class $c$ , $\\tilde{C}$ denotes the subset of $C$ with class $c$ removed, and the supervision information comes from the final pseudo-labels $\\hat{\\pmb{y}}_i^u$ after the pseudo-labeling strategies.", + "bbox": [ + 75, + 717, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The whole framework is shown in Fig. 2. All the pseudocode of producing pseudo-labels and their indicators from the logit and representation spaces is shown in Algorithm 1, and the pseudo-code of the mix pseudo-labeling strategy is shown in Algorithm 2. The pseudo-code is PyTorch-like.", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Pseudo-code of producing pseudo-labels and indicators", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network: Teacher's encoder $f'$ , segmentation head $g'$ , representation head $h'$ .", + "bbox": [ + 498, + 119, + 890, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Unlabeled mini-batch $B_{u}$ consists of $X^{u}$ , the set of prototypes $\\{\\hat{\\rho}_c(t)\\}_{c\\in C}$ .", + "bbox": [ + 498, + 151, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for $X_{u} \\in B_{u}$ do", + "2: # Predict logits from unlabeled images.", + "3: $\\hat{P}^{u,lgt}\\gets g'(f'(X^u))$", + "4: # Produce labels and indicators based on logits with Eq. 1.", + "5: $\\hat{Y}^{u,lgt},\\hat{J}^{u,lgt}\\gets lgt\\text{-} pseudo(\\tilde{P}^{u,lgt})$", + "6: # Predict representations from unlabeled images.", + "7: $\\hat{R}^{u,rep}\\gets h'(f'(X^u))$", + "8: # Produce labels and indicators based on representations with Eq. 4 and Eq. 5.", + "9: $\\hat{Y}^{u,rep},\\hat{J}^{u,rep}\\gets rep\\text{-} pseudo(\\hat{R}^{u,rep},\\{\\hat{\\rho}_c(t)\\}_{c\\in C})$", + "10: end for" + ], + "bbox": [ + 509, + 181, + 890, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 2 Pseudo-code of mix pseudo-labeling strategy", + "bbox": [ + 500, + 381, + 888, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Pseudo-labels from logit space $\\hat{Y}^{u,lgt}$ and from representation space $\\hat{Y}^{u,rep}$ . Indicators $\\hat{J}^{u,lgt}$ and $\\hat{J}^{u,rep}$ .", + "bbox": [ + 500, + 400, + 890, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Notation: Threshold $\\delta_{u}$ for sampling strategy in $\\mathcal{L}_u$ , weak threshold $\\delta_w$ .", + "bbox": [ + 500, + 431, + 890, + 460 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: # Sampling valid pseudo-labels according to the sampling strategy in logit space.", + "2: $\\hat{Y}_{val}^{u,lgt}\\gets$ sample_lgt( $\\hat{Y}^{u,lgt},\\hat{J}^{u,lgt},\\delta_u)$", + "3: # Sampling valid pseudo-labels according to the valid sampling strategy in logit space.", + "4: $\\hat{Y}_{val}^{u,rep}\\gets$ sample_rep( $\\hat{Y}^{u,rep},\\hat{J}^{u,rep},\\delta_w)$", + "5: # Produce the mask for mix pseudo-labels.", + "6: $mask_{y} \\gets \\hat{Y}^{u,rep}.$ eq( $\\hat{Y}^{u,lgt}$ )", + "7: # Produce mix pseudo-labels according to the mask.", + "8: $\\hat{Y}^u\\gets$ mask_pseudo $(\\hat{Y}^{u,lgt},\\text{mask}_y)$" + ], + "bbox": [ + 509, + 462, + 890, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Discussions", + "text_level": 1, + "bbox": [ + 500, + 648, + 624, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We discuss the relations between our framework and other most-related S4 frameworks.", + "bbox": [ + 498, + 672, + 890, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compare with PCR [51]. The key insight of PCR and our method lies in the promotion of consistency between the output in the logit and representation spaces. However, PCR exclusively relies on pseudo-labels derived from logit space as the guidance in the logit and representation spaces during the unsupervised training process, thereby neglecting the exploitation of the extensive semantic information inherent in representations and prototypes. In contrast, our method adopts a collaborative way by combining pseudo-labels obtained from both the logit and representation spaces, which improves the quality of pseudo-labels and enhances the knowledge exchange between the two spaces. In addition, since these two approaches are orthogonal, one can employ", + "bbox": [ + 496, + 703, + 892, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "935", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "our pseudo-labeling and indicator strategies within the PCR framework, or alternatively incorporate multiple prototypes and consistency loss into our approach. Such combinations may yield further improvements in performance, however, also increase the computational complexity.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compare with CPS [7]. CPS contains two segmentation models initialized differently. It leverages pseudo-labels generated by one model to supervise the other, thereby maintaining the consistency between the output of the two models in the logit space. In a similar vein to CPS, our cross pseudo-labeling strategy in Sec. 3.3 is also motivated by the objective of maintaining consistency during unsupervised training. However, our method is distinct from CPS in that we focus on preserving consistency between the logit space and the representation space. This distinction confers a distinct advantage in terms of memory efficiency, as we solely introduce a MLP as the representation head, rather than introducing an additional segmentation model as in CPS.", + "bbox": [ + 75, + 167, + 470, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compare with other contrastive based S4 works [32, 1, 46, 23]. We adhere to prior works to build our prototype, describe in Eq. 2. Different from them, we update our prototypes iteratively, similar to [60]. Our sampling strategies are based on the threshold, similar to [32, 23, 45]. However, it is worth noting that our indicators which serve as the basis for comparison with the threshold, are derived from both the logit and representation spaces. This stands in contrast to prior approaches where indicators solely originate from the logit space. The efficacy of this modification is substantiated and extensively discussed in Sec. 5.2.", + "bbox": [ + 75, + 364, + 470, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 547, + 209, + 564 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 76, + 573, + 158, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on PASCAL VOC 2012 dataset [12], Cityscapes dataset [9], ADE20K dataset [59], and COCO-Stuff 10K dataset [2] to validate the effectiveness of our proposed method. The original PASCAL VOC 2012 dataset contains 1,464 labeled images in train set and 1,449 images for validation in val set. Following [7, 37], we additionally introduce 9,118 images from SBD [18] as training images. Since the labels in SBD are coarsely annotated, following [46], we use both classic VOC train set (1,464 candidate labeled images) and blender VOC train set (10,582 candidate labeled images). Cityscapes dataset is a dataset for urban scene understanding, which contains 2,975 images in train set and 500 images in val set. ADE20K has 20,210 and 2,000 images in train and val set, with 150 classes in total. COCO-Stuff 10K has 9,000 images in train set and 1,000 images in val set, with 181 classes. The approach of preprocessing labels is followed by MMSegmentation [8].", + "bbox": [ + 75, + 597, + 470, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Network structure. We use Deeplabv3+ [4] with ResNet-101 [20] pre-trained on ImageNet [10] as our network struc", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ture. The segmentation and representation head are composed of Conv-BN-ReLU-Conv.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. For training on PASCAL VOC 2012 dataset and COCO-Stuff 10K dataset, we set the learning rate as 0.0064, weight decay as 0.0005, crop size as 512 $\\times$ 512, batch size as 16, and a total of 40,000 iterations. For training on the Cityscapes dataset, we set the learning rate as 0.0038, weight decay as 0.0005, crop size as $768 \\times$ 768, batch size as 8, and a total of 80,000 iterations. For training on ADE20K dataset, we set the learning rate as 0.0064, weight decay as 0.0005, crop size as $512 \\times 512$ , batch size as 16, and a total of 80,000 iterations. We use the poly scheduling to decay the learning rate during the training process: $lr = lr_{base} \\times (1 - \\frac{epoch}{total\\_epoch})^{0.9}$ . We use the mean of Intersection over Union (mIoU) as the metric in evaluation. We use the sliding window strategy to evaluate the performance of our method on the Cityscapes dataset, following [7]. In addition, when adopting our cross-labeling strategy, due to the requirement of a set of high-quality prototypes when classifying each representation, we first solely use the supervision from logit space for 20 epochs to initialize the prototypes.", + "bbox": [ + 496, + 121, + 892, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with Existing Methods", + "text_level": 1, + "bbox": [ + 498, + 436, + 808, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this subsection, we first reproduce three baselines: MT [44], CutMix [15], and a same contrastive-based framework with us but with only logit space pseudo-labels and indicators (Baseline) on classic VOC train set. Meanwhile, we make the comparison of our method with mix (CSS (mix)) and cross (CSS (crs.)) pseudo-labeling strategy on blender VOC train set and Cityscapes train set with following recent SOTA S4 methods: CCT [38], CPS [7], U²PL [46], ST++ [52], PRCL [48], PCR [51], and PSMT [33]. Since the data split will dramatically affect the performance in S4, i.e., choosing labeled images plays an important role in the final results, we conduct experiments with three different data splits and report the mean value and standard deviation (blue color numbers). Since the mix pseudo-labeling strategy has better performance, we only use CSS (mix) when compared with SOTAs. Meanwhile, we use ResNet-101 with deep stem blocks as our network structure when compared with SOTAs. Since there is no uniform data split, we use the data splits in U²PL [46]. We use the OHEM loss when training Cityscapes.", + "bbox": [ + 496, + 460, + 890, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on PASCAL VOC 2012. Tab. 1 shows the comparison with our baselines on Classic PASCAL VOC 2012 set. Our method consistently outperforms baselines with an acceptable standard deviation on all label rates. Tab. 2 shows the comparison with the SOTAs on PASCAL VOC 2012.", + "bbox": [ + 496, + 763, + 890, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on Cityscapes. Tab. 3 shows the performance of our method on Cityscapes.", + "bbox": [ + 496, + 839, + 890, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on ADE20K and COCO-Stuff 10K. Tab. 4 shows the results of our method on ADE20K and COCO-Stuff.", + "bbox": [ + 498, + 869, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "936", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bd9a2c5ea5cf427244589964771a74592fcaf2ff1121de0bb81f410ba58da557.jpg", + "table_caption": [ + "Table 1. Results on classic VOC train set with four different label rate. Labeled data splits are from the original VOC train set. All approaches are reproduced with three runs and three different data splits." + ], + "table_footnote": [], + "table_body": "
Pascal VOC 2012 (Classic)
Method92183366732
Sup.51.57±3.5854.69±2.4464.86±1.0470.77±0.76
MT58.92±2.9961.63±1.7666.79±0.5371.58±0.51
CutMix65.82±3.6067.91±1.4772.53±0.5074.08±0.49
Baseline66.91±4.2170.32±1.8673.97±1.0476.49±0.58
CSS(crs.)67.03±4.5871.41±1.6774.47±1.0877.08±0.37
CSS(mix)68.09±4.8971.93±1.8874.91±1.1277.57±0.73
", + "bbox": [ + 80, + 142, + 467, + 258 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f12f4e0946f0045f47e675aba0f182275f9550f7b8d6dc81af541f4daaf59611.jpg", + "table_caption": [ + "Table 2. Results on blender VOC train set. All the results from the recent papers [46, 52, 13, 29, 33, 51]. Labeled data is from the augmented VOC train set and the data splits are from [46, 33]." + ], + "table_footnote": [], + "table_body": "
Pascal VOC 2012 (Blender)
Method662132326465291
CCT [38]71.8673.6876.5177.40
CPS [7]74.4876.4477.6878.64
U2PL [46]77.2179.0179.3080.50
ST++ [52]74.7077.9077.90-
PRCL [48]76.9678.1679.0279.59
PCR [51]78.6080.7180.7880.91
PSMT [33]75.5078.2078.7279.76
CSS (mix)78.7379.5480.8281.06
", + "bbox": [ + 80, + 311, + 465, + 452 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e48f517d8521fc2662cf1f45eaf7a4bb3d4985a65647d5cddde92dbb9cd217aa.jpg", + "table_caption": [ + "Table 3. Results on Cityscapes. The model is trained on the Cityscapes train set, which consists of 2,975 samples in total, and tested on Cityscapes val set. And all the results from the recent papers [46, 51, 33]." + ], + "table_footnote": [], + "table_body": "
Cityscapes
Method1863727441488
CCT [38]69.3274.1275.9978.10
CPS [7]69.7874.3174.5876.82
U2PL [46]70.3074.3776.4779.05
PCR [51]73.4176.3178.4079.11
PSMT [33]-76.8977.6079.09
CSS (mix)74.0276.9377.9479.62
", + "bbox": [ + 80, + 521, + 465, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Ablative Study", + "text_level": 1, + "bbox": [ + 76, + 660, + 225, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The main contribution of our work lies in 1) collaborative pseudo-labeling strategies and 2) a new indicator for representation learning. To further prove the effectiveness of our proposed method, we conduct ablative studies on these two points. We choose Deeplabv3+ with ResNet-101 pre-trained on ImageNet as our backbone and leverage 92 labeled images and 183 labeled images in PASCAL VOC 2012. The other settings are the same as those in Sec. 4.", + "bbox": [ + 75, + 686, + 468, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Effectiveness of Collaborative pseudo-labeling", + "text_level": 1, + "bbox": [ + 76, + 816, + 465, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quality of pseudo-labels. To illustrate the superiority of using pseudo-labels from two spaces in a collaborative way as supervision, we conduct experiments to show the quality of pseudo-labels obtained 1) from logit space (lgt.), 2) from", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9ec7b34a9c8edda15bbb118fc8aad3b89f26be65e4f81c35fb50a8784b65bfc6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 89, + 566, + 137 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/885b6ec9ed1b4342c72e6b0a35d62753823bf6a702037312c0b75bcb4e800fde.jpg", + "image_caption": [ + "(b) GT" + ], + "image_footnote": [], + "bbox": [ + 566, + 90, + 630, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f53cf14542852989825603154b7a11bfa60e5a33b958812402bd582f13415e77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 90, + 692, + 137 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3bd1a658daae03d39c76792d3e531523ef9aba32cfb9563c71d41f96e2723fcd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 90, + 758, + 137 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2101bf8846b2951f47194ea6ee2e90ba11e34bf91205d42f088d7fb709551c9e.jpg", + "image_caption": [ + "(d) lgt. label", + "Figure 3. Differences between pseudo-labels in different spaces." + ], + "image_footnote": [], + "bbox": [ + 694, + 138, + 758, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3311a85ac4b5d87c7ad39fbef4c91b440d5bb9a920abb2bace77c69987926f12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 90, + 821, + 137 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de9a6e7cf32de80262c4629e83a94e3f231d7f7965cfc96a2b7ec931a5a5022f.jpg", + "image_caption": [ + "(e) rep. mask" + ], + "image_footnote": [], + "bbox": [ + 761, + 138, + 823, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7fbcb16361f71a321b1f1851c62aef74216a3d53756d95c792d51d6eff4f5f3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 826, + 90, + 888, + 137 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c9a2c8b867d5390141cce1649bf2e365f208bc51a70272403e5b3aa16c997ef.jpg", + "image_caption": [ + "(f) rep. label" + ], + "image_footnote": [], + "bbox": [ + 826, + 138, + 888, + 186 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "representation space (rep.), and 3) from the mix pseudolabeling strategy (mix). The pseudo-labels are sampled with corresponding sampling strategies in Sec. 3.4. Tab. 5 illustrates the IoU of pseudo-labels for each class on PASCAL VOC 2012 with 92 labeled images. The results clearly indicate that employing pseudo-labels from the representation space enhances the accuracy of the final pseudo-labels in most classes. This improvement is particularly evident in classes that are originally under-performing, such as the IoU improvement of $11.42\\%$ for the chair class and $11.58\\%$ for the sofa class.", + "bbox": [ + 496, + 250, + 890, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Meanwhile, we also visualize the pseudo-labels obtained from logit space (lgt.) and representation space (rep.) in Fig. 3. Fig. 3 (c) and (e) show the masks for pseudolabels produced by the sampling strategies. In particular, the white color represents the valid pixels used during unlabeled learning, while the black color indicates the discarded pixels. Fig. 3 (d) and (f) are the pseudo-labels we obtained from two spaces. The figure clearly illustrates the differences between pseudo-labels produced by different spaces. For example, the parts of the instance edge in pseudo-labels are usually discarded since they are challenging for learning in logit space. However, pseudo-labels from representation space will easily tackle this problem (first row). The pseudo-labels from representation space are more inaccurate in some complex scenes, which be resolved by combining pseudo-labels from logit space (second row).", + "bbox": [ + 496, + 416, + 890, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We mainly attribute the differences in pseudo-label to the differing concentrations of learning in the two spaces. Specifically, learning in the logit space primarily emphasizes the most discriminative part of features, whereas that in the representation space treats each part of features equally. As a result, learning in the logit space may overlook minor feature differences, leading to sub-optimal performance in predicting instance edges and distinguishing between similar classes (e.g., chair and sofa). Conversely, learning in the representation space produces balanced performance across all image parts and classes. However, this can lead to erroneous predictions for classes with high intra-class variance (e.g., background). By leveraging pseudo-labels from the two spaces, we capitalize on the strengths of the learning in each space and enhance the knowledge exchange between the two spaces.", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "937", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/230773c88f7d9bc98d41e4e8bafbf593a794fa97ce9c78ee832021623c76db00.jpg", + "table_caption": [ + "Table 4. Results on ADE20K and COCO-Stuff 10K with four different label rates and single data splits." + ], + "table_footnote": [], + "table_body": "
ADE20KCOCO-Stuff 10K
Method1/161/81/41/21/161/81/41/2
Cutmix28.9131.8934.6239.1523.0926.1429.8830.91
Baseline31.1133.2436.4639.9126.5727.9128.9731.08
CSS (crs.)32.0133.9237.8439.9727.0628.8530.5931.73
CSS (mix)32.5534.2137.0140.8527.4629.5129.9831.91
", + "bbox": [ + 80, + 101, + 890, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/65b884998334dfe4b377e7bb07c9be9e444b9401c620a45999ce72efde4772b2.jpg", + "table_caption": [ + "Table 5. The quality of pseudo-labels from different pseudo-labeling strategies. The pseudo-labels are sampled by sampling strategies." + ], + "table_footnote": [], + "table_body": "
sourcebackaero.bicy.birdboatbott.buscarcatchaircow
lgt.96.7895.1477.4793.3881.3787.5496.7695.4894.474.0992.15
rep.90.7196.5061.4275.7553.3054.6584.4680.5591.1128.0388.16
mix96.66↓0.1298.12↑2.9882.76↑5.2994.47↑1.0985.12↑3.7589.94↑2.4097.03↑0.2795.98↑0.5094.65↑0.1819.01↑14.9294.81↑2.66
sourcetabeldoghorsemotorpers.plantsheepsofatraintvmIoU
lgt.58.3494.1293.0191.3793.6850.3391.1018.1686.6564.8978.87
rep.52.2386.9773.2288.7091.7556.1366.7835.2588.5162.6171.57
mix64.35↑6.0194.33↑0.2193.65↑0.6491.50↑0.1393.01↓0.6755.64↑5.3191.56↑0.4629.74↑11.5893.54↑6.8969.48↑4.5982.17↑3.33
", + "bbox": [ + 80, + 213, + 890, + 329 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cc1af88b0a1528668fa93525f0174368487d95d601cc419cc2e4a68e11c8e349.jpg", + "table_caption": [ + "Table 6. Results on pseudo-labels from different sources on two different label rates." + ], + "table_footnote": [], + "table_body": "
source92 labels183 labels
logit space67.1170.32
representation space64.2067.52
mix pseudo-labeling68.41↑1.3072.74↑2.42
cross pseudo-labeling67.85↑0.7471.98↑1.66
", + "bbox": [ + 78, + 377, + 467, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results of different strategies To investigate the involvement of different pseudo-labeling strategies, we conduct the experiments as follows: 1) Using pseudo-labels from logit space. 2) Using pseudo-labels from representation space. 3) Using mix pseudo-labeling strategy. 4) Using cross pseudolabeling strategy. Tab. 6 shows the effectiveness of our proposed strategy in two different label rates. The results show that the performances of experiments with collaborative pseudo-labeling strategies are better than the ones whose pseudo-labels come from a single space with two different label rates, which proves the effectiveness of our proposed collaboration between the two spaces. It is worth noting that even though the quality of pseudo-labels from representation space is lower than that from logit space, the performance of the model is also boosted by using the cross pseudo-labeling strategy to maintain consistency between the predictions in two spaces. In addition, our method with the mix pseudo-labeling strategy outperforms that with the cross pseudo-labeling strategy.", + "bbox": [ + 76, + 474, + 468, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Effectiveness of the Indicator", + "text_level": 1, + "bbox": [ + 76, + 771, + 336, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation of merely using confidence. To explain the limitation of merely using confidence for learning in the logit and representation spaces, we conduct experiments to show the relations between confidence and similarity. The similarity is the cosine similarity between representations and prototypes, which directly shows the confusion level between the representation and the prototype of its class.", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b808f15724ab1f45e36ba846846ae5ac6fdbccf2c1e21314c20fd4b0b040d92f.jpg", + "image_caption": [ + "(a)", + "Figure 4. Relations between similarity and confidence." + ], + "image_footnote": [], + "bbox": [ + 501, + 357, + 678, + 446 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d4e6d66393eb772fd8063bec986ad111d15dd70005f4e8f98ccee0b69d3f013b.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 679, + 358, + 888, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 4 (a) shows the comparison between the confidence of each prediction and the corresponding similarity. We use the class person for demonstrating in our experiments. It shows clearly that even though fixing the confidence into a small range (from 0.8 to 0.81 in our settings), the similarity varies. Meanwhile, in Fig. 4 (b), different color bars stand for different intervals of confidence, and the lines denote the mean similarity between the prototype and each representation whose corresponding confidence is in the current interval. Fig. 4 (b) illustrates that the mean similarity of the class fluctuates when its interval of confidence rises. Both two figures imply that confidence is not able to represent the confusion level between representations and prototypes since there are no direct and close relations between confidence and similarity.", + "bbox": [ + 496, + 505, + 890, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 5 visualizes the similarity and confidence of an image in both logit (lgt.) and representation (rep.) spaces, indicating the varying levels of confusion in the same region when learning in different spaces.", + "bbox": [ + 496, + 732, + 890, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also attribute it to the different concentrations of learning two spaces, i.e., the confusing region in one space can be more readily addressed in the other space.", + "bbox": [ + 496, + 794, + 890, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Thus, it is inappropriate to choose confidence as the indicator to involve representation learning, e.g., sampling more hard samples with threshold and indicator. In contrast, our indicator directly employs similarity between the represen", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "938", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/afbb859f0b9726ddd8dd67ba3373e08a7863d8d794adf298f134bddb32e33570.jpg", + "image_caption": [ + "high", + "(a) Image", + "Figure 5. Visualization of the confusing part in different spaces." + ], + "image_footnote": [], + "bbox": [ + 80, + 112, + 174, + 186 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/7c004d2fc7084f03c4164fb131f750be95def112c593c4cadd15172f5de04802.jpg", + "image_caption": [ + "(b) GT" + ], + "image_footnote": [], + "bbox": [ + 178, + 114, + 271, + 186 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6e11740b2e78d6466f069457b1c094909b6eee02dbd9ecf3470bfe3b2de4de31.jpg", + "image_caption": [ + "(c) lgt. space" + ], + "image_footnote": [], + "bbox": [ + 272, + 114, + 366, + 186 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/692662af87d13f12ef88a56656a8888fa45fb2d2e7820fe9a4ac44e5f98eb755.jpg", + "image_caption": [ + "(c) rep. space" + ], + "image_footnote": [], + "bbox": [ + 367, + 114, + 465, + 186 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "tation and prototype of its class, which directly reflects the confusion level in representation learning. It is more accurate to use similarity as the indicator to sample hard and critical samples in representation learning.", + "bbox": [ + 75, + 253, + 467, + 314 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Results of different indicators. Tab. 7 shows the impact of using different indicators. We conduct experiments on two different label rates (92 and 183) and three different indicators for pseudo-labels: only confidence (conf.), only similarity (smlr.), and confidence for learning logit space while similarity for learning representation space (mix). We use two different approaches to obtain pseudo-labels: from logit space only (seg label) and from mix pseudo-labeling strategy (mix label). It is clear that using both confidence and similarity to involve the learning in their own spaces obtains the best performance.", + "bbox": [ + 75, + 314, + 467, + 481 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/04b71f1e4c0077e41db1af48f3d8f649b8a5e862424f976d0715f10b250eb1a8.jpg", + "table_caption": [ + "Table 7. Results on indicators from different spaces on two different label rates." + ], + "table_footnote": [], + "table_body": "
92 labels183 labels
sourceseg labelmix labelseg labelmix label
conf.67.1168.3370.3271.92
smlr.66.1666.8968.9069.25
mix67.80↑0.6968.41↑1.3071.50↑1.1872.74↑2.42
", + "bbox": [ + 78, + 517, + 467, + 590 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.3. Ablation study of Components", + "text_level": 1, + "bbox": [ + 76, + 617, + 346, + 633 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we conduct experiments to introduce our components in CSS step by step, with results shown in Tab. 8. Our baseline is the conventional contrastive-based S4, achieving mIoU of $67.11\\%$ on 92 labels and $70.32\\%$ on 183 labels. Mix and cross means the pseudo-labels are from the mix pseudo-labeling strategy and cross pseudo-labeling strategy while the indicator is still the confidence in two spaces. Ind means we use the different indicators in different spaces while the pseudo-labels are from logit space. The last two rows represent our proposed two pseudo-labeling strategies with indicators from two spaces.", + "bbox": [ + 75, + 641, + 467, + 806 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 76, + 816, + 259, + 832 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fig. 6 shows the qualitative results of different methods on PASCAL VOC 2012 with 92 labeled images. Baseline means the conventional contrastive-based method. Compared with the original self-training methods (CutMix),", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/47d7024d5d72288596e8438e9812b8c51186e8af4abaa69f8eeab86b859fd826.jpg", + "table_caption": [ + "Table 8. Ablation study on different components of our CSS." + ], + "table_footnote": [], + "table_body": "
component92 labels183 labels
baseline67.1170.32
mix68.33↑1.2271.92↑1.60
cross67.21↑0.1070.87↑0.55
ind67.80↑0.6971.50↑1.18
mix + ind68.41↑1.3072.74↑2.42
cross + ind67.85↑0.7471.98↑1.66
", + "bbox": [ + 503, + 99, + 888, + 200 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "thanks to introducing pixel-wise contrastive learning, the baseline, and our method perform better in some ambiguous regions. Furthermore, benefiting from the supervision of two spaces and different indicators in different spaces, our method outperforms the baseline.", + "bbox": [ + 496, + 228, + 890, + 303 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0087841697414687131256fde037c71be6289bf332eac50c0f368691b3531be0.jpg", + "image_caption": [ + "images. Yellow boxes highlight the main differences." + ], + "image_footnote": [], + "bbox": [ + 503, + 316, + 888, + 526 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 585, + 617, + 602 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose two collaborative pseudolabeling strategies to take full use of the semantic information in the representation space and enhance the knowledge exchange between the logit and representation spaces. Moreover, we employ a new indicator for the learning process in the representation space. Extensive experiments demonstrate that our pseudo-labeling strategies obtain more reliable supervision during unlabeled training and our indicator helps the model to concentrate on more critical parts during representation learning.", + "bbox": [ + 496, + 612, + 890, + 762 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Future work: In this paper, we employ pseudo-labeling strategies to utilize the semantic information in both logit and representation spaces. In the future, we will investigate more powerful strategies to enhance the knowledge exchange between two spaces.", + "bbox": [ + 496, + 763, + 890, + 839 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "939", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Inigo Alonso, Alberto Sabater, David Ferstl, Luis Montesano, and Ana C. Murillo. Semi-supervised semantic segmentation with pixel-level contrastive learning from a class-wise memory bank. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8219-8228, October 2021.", + "[2] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018.", + "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020.", + "[4] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018.", + "[5] Mu Chen, Zhedong Zheng, Yi Yang, and Tat-Seng Chua. Pipa: Pixel-and patch-wise self-supervised learning for domain adaptive semantic segmentation. arXiv preprint arXiv:2211.07609, 2022.", + "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020.", + "[7] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021.", + "[8] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.", + "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016.", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009.", + "[11] Nanqing Dong and Eric P Xing. Few-shot semantic segmentation with prototype learning. In BMVC, volume 3, 2018.", + "[12] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010.", + "[13] Jiashuo Fan, Bin Gao, Huan Jin, and Lihui Jiang. Ucc: Uncertainty guided cross-head co-training for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition (CVPR), pages 9947-9956, June 2022.", + "[14] Zhengyang Feng, Qianyu Zhou, Qiqi Gu, Xin Tan, Guangliang Cheng, Xuequan Lu, Jianping Shi, and Lizhuang Ma. Dmt: Dynamic mutual training for semi-supervised learning. Pattern Recognition, page 108777, 2022.", + "[15] Geoff French, Timo Aila, Samuli Laine, Michal Mackiewicz, and Graham Finlayson. Semi-supervised semantic segmentation needs strong, high-dimensional perturbations. In In 31th British Machine Vision Conference, 2020.", + "[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020.", + "[17] Dayan Guan, Jiaxing Huang, Aoran Xiao, and Shijian Lu. Unbiased subclass regularization for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9968-9978, 2022.", + "[18] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In 2011 international conference on computer vision, pages 991-998. IEEE, 2011.", + "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020.", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.", + "[21] Ruifei He, Jihan Yang, and Xiaojuan Qi. Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6930-6940, October 2021.", + "[22] Judy Hoffman, Dequan Wang, Fisher Yu, and Trevor Darrell. Fcts in the wild: Pixel-level adversarial and constraint-based adaptation. arXiv preprint arXiv:1612.02649, 2016.", + "[23] Hanzhe Hu, Jinshi Cui, and Liwei Wang. Region-aware contrastive learning for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16291-16301, 2021.", + "[24] Hanzhe Hu, Fangyun Wei, Han Hu, Qiwei Ye, Jinshi Cui, and Liwei Wang. Semi-supervised semantic segmentation via adaptive equalization learning. Advances in Neural Information Processing Systems, 34:22106-22118, 2021.", + "[25] Wei-Chih Hung, Yi-Hsuan Tsai, Yan-Ting Liou, Yen-Yu Lin, and Ming-Hsuan Yang. Adversarial learning for semi-supervised semantic segmentation. arXiv preprint arXiv:1802.07934, 2018.", + "[26] Zhengkai Jiang, Yuxi Li, Ceyuan Yang, Peng Gao, Yabiao Wang, Ying Tai, and Chengjie Wang. Prototypical contrast adaptation for domain adaptive semantic segmentation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 36-54. Springer, 2022." + ], + "bbox": [ + 503, + 93, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "940", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Tarun Kalluri, Girish Varma, Manmohan Chandraker, and CV Jawahar. Universal semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5259-5270, 2019.", + "[28] Rihuan Ke, Angelica I Aviles-Rivero, Saurabh Pandey, Saikumar Reddy, and Carola-Bibiane Schonlieb. A three-stage self-training framework for semi-supervised semantic segmentation. IEEE Transactions on Image Processing, 31:1805-1815, 2022.", + "[29] Donghyeon Kwon and Suha Kwak. Semi-supervised semantic segmentation with error localization network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9957-9967, 2022.", + "[30] Daiqing Li, Junlin Yang, Karsten Kreis, Antonio Torralba, and Sanja Fidler. Semantic segmentation with generative models: Semi-supervised learning and strong out-of-domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8300-8311, 2021.", + "[31] Jinlu Liu, Liang Song, and Yongqiang Qin. Prototype rectification for few-shot learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16, pages 741–756. Springer, 2020.", + "[32] Shikun Liu, Shuaifeng Zhi, Edward Johns, and Andrew Davison. Bootstrapping semantic segmentation with regional contrast. In International Conference on Learning Representations, 2022.", + "[33] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4258-4267, 2022.", + "[34] Binjie Mao, Xinbang Zhang, Lingfeng Wang, Qian Zhang, Shiming Xiang, and Chunhong Pan. Learning from the target: Dual prototype network for few shot semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1953-1961, 2022.", + "[35] Sudhanshu Mittal, Maxim Tatarchenko, and Thomas Brox. Semi-supervised semantic segmentation with high-and low-level consistency. IEEE transactions on pattern analysis and machine intelligence, 43(4):1369-1379, 2019.", + "[36] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979-1993, 2018.", + "[37] Viktor Olsson, Wilhelm Tranheden, Juliano Pinto, and Lennart Svensson. Classmix: Segmentation-based data augmentation for semi-supervised learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1369-1378, 2021.", + "[38] Yassine Ouali, Celine Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross-consistency training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Yingwei Pan, Ting Yao, Yehao Li, Yu Wang, Chong-Wah Ngo, and Tao Mei. Transferrable prototypical networks for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2239-2247, 2019.", + "[40] Jizong Peng, Guillermo Estrada, Marco Pedersoli, and Christian Desrosiers. Deep co-training for semi-supervised image segmentation. Pattern Recognition, 107:107269, 2020.", + "[41] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017.", + "[42] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020.", + "[43] Korawat Tanwisuth, Xinjie Fan, Huangjie Zheng, Shujian Zhang, Hao Zhang, Bo Chen, and Mingyuan Zhou. A prototype-oriented framework for unsupervised domain adaptation. Advances in Neural Information Processing Systems, 34:17194-17208, 2021.", + "[44] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017.", + "[45] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7303-7313, 2021.", + "[46] Yuchao Wang, Haochen Wang, Yujun Shen, Jingjing Fei, Wei Li, Guoqiang Jin, Liwei Wu, Rui Zhao, and Xinyi Le. Semi-supervised semantic segmentation using unreliable pseudo labels. In Proceedings of the IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[47] Chen Wei, Kihyuk Sohn, Clayton Mellina, Alan Yuille, and Fan Yang. Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10857-10866, June 2021.", + "[48] Haoyu Xie, Changqi Wang, Mingkai Zheng, Minjing Dong, Shan You, and Chang Xu. Boosting semi-supervised semantic segmentation with probabilistic representations. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2938-2946, 2023.", + "[49] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020.", + "[50] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021." + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "941", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Hai-Ming Xu, Lingqiao Liu, Qiuchen Bian, and Zhen Yang. Semi-supervised semantic segmentation with prototype-based consistency regularization. Advances in Neural Information Processing Systems, 2022.", + "[52] Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. St++: Make self-training work better for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4268-4277, 2022.", + "[53] Jianlong Yuan, Yifan Liu, Chunhua Shen, Zhibin Wang, and Hao Li. A simple baseline for semi-supervised semantic segmentation with strong data augmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8229-8238, 2021.", + "[54] Xiangyu Yue, Zangwei Zheng, Shanghang Zhang, Yang Gao, Trevor Darrell, Kurt Keutzer, and Alberto Sangiovanni Vincentelli. Prototypical cross-domain self-supervised learning for few-shot unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13834-13844, 2021.", + "[55] Xiangyun Zhao, Raviteja Vemulapalli, Philip Andrew Mansfield, Boqing Gong, Bradley Green, Lior Shapira, and Ying Wu. Contrastive learning for label efficient semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10623-10633, October 2021.", + "[56] Xu Zheng, Yunhao Luo, Hao Wang, Chong Fu, and Lin Wang. Transformer-cnn cohort: Semi-supervised semantic segmentation by the best of both students, 2022.", + "[57] Xu Zheng, Jinjing Zhu, Yexin Liu, Zidong Cao, Chong Fu, and Lin Wang. Both style and distortion matter: Dualpath unsupervised domain adaptation for panoramic semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1285-1295, June 2023.", + "[58] Yuanyi Zhong, Bodi Yuan, Hong Wu, Zhiqiang Yuan, Jian Peng, and Yu-Xiong Wang. Pixel contrastive-consistent semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7273-7282, 2021.", + "[59] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017.", + "[60] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A prototype view. In CVPR, 2022.", + "[61] Yanning Zhou, Hang Xu, Wei Zhang, Bin Gao, and Pheng-Ann Heng. C3-semiseg: Contrastive semi-supervised segmentation via cross-set learning and dynamic classbalancing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7036-7045, 2021." + ], + "bbox": [ + 78, + 90, + 468, + 838 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "942", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Space Engage_ Collaborative Space Supervision for Contrastive-Based Semi-Supervised Semantic Segmentation/5953e602-104c-4dc8-b226-b90796937563_model.json b/2023/Space Engage_ Collaborative Space Supervision for Contrastive-Based Semi-Supervised Semantic Segmentation/5953e602-104c-4dc8-b226-b90796937563_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9a95c63b0fcb11c5e0c2a5da86580f8e7ac44e33 --- /dev/null +++ b/2023/Space Engage_ Collaborative Space Supervision for Contrastive-Based Semi-Supervised Semantic Segmentation/5953e602-104c-4dc8-b226-b90796937563_model.json @@ -0,0 +1,3007 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.132, + 0.131, + 0.84, + 0.177 + ], + "angle": 0, + "content": "Space Engage: Collaborative Space Supervision for Contrastive-based Semi-Supervised Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.198, + 0.203, + 0.771, + 0.223 + ], + "angle": 0, + "content": "Changqi Wang\\(^{1*}\\), Haoyu Xie\\(^{1*}\\), Yuhui Yuan\\(^{2}\\), Chong Fu\\(^{1,4\\dagger}\\), Xiangyu Yue\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.228, + 0.843, + 0.265 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Computer Science and Engineering, Northeastern University, Shenyang, China \n\\(^{2}\\) Microsoft Research Asia \\(^{3}\\) The Chinese University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.265, + 0.871, + 0.283 + ], + "angle": 0, + "content": "4 Key Laboratory of Intelligent Computing in Medical Image, Ministry of Education, NEU, China" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.293, + 0.807, + 0.307 + ], + "angle": 0, + "content": "2101668@stu.neu.edu.cn, xiehaoyu@stumail.neu.edu.cn, yuyua@microsoft.com" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.311, + 0.686, + 0.325 + ], + "angle": 0, + "content": "fuchong@mail.neu.edu.cn, xyyue@ie.cuhk.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.36, + 0.314, + 0.377 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.391, + 0.474, + 0.725 + ], + "angle": 0, + "content": "Semi-Supervised Semantic Segmentation (S4) aims to train a segmentation model with limited labeled images and a substantial volume of unlabeled images. To improve the robustness of representations, powerful methods introduce a pixel-wise contrastive learning approach in latent space (i.e., representation space) that aggregates the representations to their prototypes in a fully supervised manner. However, previous contrastive-based S4 methods merely rely on the supervision from the model's output (logits) in logit space during unlabeled training. In contrast, we utilize the outputs in both logit space and representation space to obtain supervision in a collaborative way. The supervision from two spaces plays two roles: 1) reduces the risk of over-fitting to incorrect semantic information in logits with the help of representations; 2) enhances the knowledge exchange between the two spaces. Furthermore, unlike previous approaches, we use the similarity between representations and prototypes as a new indicator to tilt training those under-performing representations and achieve a more efficient contrastive learning process. Results on two public benchmarks demonstrate the competitive performance of our method compared with state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.751, + 0.21, + 0.767 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.777, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Semantic segmentation is a fundamental task in computer vision, aiming to classify each pixel in an image. Significant progress [22, 4] has been made in training on high-quality labeled images using segmentation models composed of an encoder and a segmentation head. However, annotating images is expensive and time-consuming. Semi" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.36, + 0.889, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.476, + 0.888, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.613, + 0.895, + 0.684 + ], + "angle": 0, + "content": "Figure 1. We enhance the knowledge exchange between the logit and representation spaces. Orange and blue represent different classes. Top: Existing contrastive-based S4 methods overlook the semantic information in representation space. Bottom: Our method uses dual-space collaborative supervision." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.716, + 0.893, + 0.761 + ], + "angle": 0, + "content": "supervised Semantic Segmentation (S4) alleviates the thirst for annotation by leveraging unlabeled images to train segmentation models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Most existing works learn from unlabeled images via self-training [44, 37, 15] or consistency regularization [40, 38, 13] strategies, both of which retrain the model with its predictions on unlabeled images. Recently, great success has been achieved by introducing pixel-wise contrastive learning to semantic segmentation, which endows the model with a stronger features-extracting ability by accessing a more discriminative representation space. Specifically, these methods [45, 23] project each pixel to repre" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.876, + 0.203, + 0.888 + ], + "angle": 0, + "content": "*equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.218, + 0.901 + ], + "angle": 0, + "content": "\\(\\dagger\\) corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.876, + 0.218, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "931" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.332 + ], + "angle": 0, + "content": "sensation space as a representation and regularize it in a fully supervised manner, i.e., aggregating the representations with the same class and separating them with different classes. In semi-supervised settings, due to limited labels, most methods [1, 32, 46] obtain supervision from the model's output logits in logit space during the unlabeled training process. However, recent contrastive-based semantic segmentation methods [1, 32, 46, 45] mainly focus on the learning process in logit space while only taking that in representation space as an auxiliary task. The unidirectional supervision makes training dominated by the predicted logits, leading to the neglect of information in the representation space. We argue that this kind of single-space supervision may incorrectly provide semantic guidance to representation learning and fails to facilitate knowledge exchange between the two spaces (see Sec. 5.1)." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.334, + 0.468, + 0.651 + ], + "angle": 0, + "content": "In this work, we extend the single-space supervision to a dual-space supervision for contrastive-based S4 and propose Collaborative Space Supervision (CSS). Our key insight is to: i) utilize the semantic information in representations to obtain more reliable guidance during unlabeled training, and to enhance knowledge exchange between two spaces; ii) provide a more accurate reference for the model's performance on predicting each representation to tilt training those under-performing representations. To achieve objective i), we obtain dense semantic predictions by retrieving the nearest class prototype for each representation in the representation space and then engage with predictions from the logits for collaborative supervision of the model. For objective ii), we measure the similarity between the representations and prototypes and use the similarity after a normalization operation as the indicator for guiding the learning process in the representation space. Unlike previous works that utilize confidence as the indicator to involve representation learning, the similarity directly reflects the confusion level between representations and prototypes, resulting in more efficient representation learning." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.652, + 0.468, + 0.757 + ], + "angle": 0, + "content": "To summarize, our main contributions are three-fold: 1) The dual-space collaboration for contrastive-based S4, enhances the knowledge exchange between the logit and representation spaces. 2) Utilizing similarity to provide a more accurate reference for the model's performance in representation learning. 3) Extensive experiments on two S4 benchmarks demonstrate the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.775, + 0.225, + 0.79 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.801, + 0.422, + 0.817 + ], + "angle": 0, + "content": "2.1. Semi-supervised Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.825, + 0.468, + 0.9 + ], + "angle": 0, + "content": "The aim of S4 is to train a segmentation model with the semi-supervised setting (i.e., a few labeled images and a large number of unlabeled images) to classify each pixel in an entire image. The critical issue of S4 is how to leverage unlabeled images to train the model. Some methods [25," + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.092, + 0.89, + 0.318 + ], + "angle": 0, + "content": "27, 30, 35] based on GANs [16], adversarial training [36], and consistency regularization paradigm [38, 13, 40, 7, 56]. Meanwhile, self-training [28, 44, 49, 61, 53] is also a striking paradigm, which always generates pseudo-labels from model and retransmits the model with the combined supervision of human annotations and pseudo-labels. One essential issue of self-training is the accuracy of pseudo-labels. Some methods [29, 33, 14, 42, 52] try to polish pseudo-labels and provide reliable guidance. Some methods [21, 47, 24, 17] focus on the class-imbalance problems in the dataset and try to alleviate the negative effect from class-biased pseudolabels generated by the model pre-trained on imbalanced labeled images. We build our framework based on the self-training and additionally explore semantic information among different images." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.329, + 0.781, + 0.345 + ], + "angle": 0, + "content": "2.2. Pixel-wise Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.353, + 0.89, + 0.64 + ], + "angle": 0, + "content": "Pixel-wise contrastive learning explores semantic relations not only in the individual image but also among different images. Different from instance-wise contrastive learning [19, 6, 3], pixel-wise contrastive learning [50, 55, 5, 58] project each pixel to the representation in representation space with the cooperation of encoder and representation head. Representations are then aggregated in their prototypes and are separated from each other in different classes. In semi-supervised settings, most methods [1, 32, 46, 48] use pseudo-labels based on logits to provide semantic information contrastive learning process during training on unlabeled images. Meanwhile, the confidence of logit is used as an indicator to involve the contrastive learning process, e.g., [32] uses the hard representations whose corresponding logit confidence is lower than a threshold to contrast for effective training. As opposed to the above methods, we use collaborative space supervision for contrastive learning on unlabeled images and use a new indicator to involve the contrastive learning progress." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.651, + 0.736, + 0.667 + ], + "angle": 0, + "content": "2.3. Prototype-based Learning" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.675, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Prototype-based learning has been widely studied in few-shot learning [41, 11, 31, 34] and unsupervised domain adaption [54, 26, 43, 39, 57]. Recently, it is restudied in semantic segmentation as known as a non-parametric prototype-based classifier [60]. Concretely, the classes in the dataset are presented by a set of non-learnable prototypes, and the dense semantic predictions are thus achieved by assigning the output features to its most similar prototype. Under semi-supervised settings, some methods [51] maintain the consistency between predictions from a linear predictor and a prototype-based predictor. The two predictors are followed by the encoder and project the features to logit space and representation space, respectively. In this work, we combine the semantic information in the logit and representation spaces to provide supervision in a collabora" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.947, + 0.51, + 0.956 + ], + "angle": 0, + "content": "932" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.355, + 0.107 + ], + "angle": 0, + "content": "tive way during semi-supervised learning." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.123, + 0.212, + 0.14 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.149, + 0.473, + 0.332 + ], + "angle": 0, + "content": "In the S4 task, we have a small labeled set \\(\\mathcal{D}_l = \\{(\\pmb{x}_i^l,\\pmb{y}_i^l)\\}_{i=1}^{N_l}\\) and a large unlabeled set \\(\\mathcal{D}_u = \\{\\pmb{x}_i^u\\}_{i=1}^{N_u}\\), where \\(\\pmb{x}_i^l,\\pmb{x}_i^u\\in \\mathbb{R}^{H\\times W\\times 3}\\), \\(H,W\\) denote the height, and the width, respectively. And ground truth \\(\\pmb{y}_i^l\\in \\{0,1\\}^{H\\times W\\times |C|}\\) with the set of class \\(C\\). The goal is to boost model performance with \\(\\mathcal{D}_u\\). The base model consists of an encoder \\(f(\\cdot)\\) and a segmentation head \\(g(\\cdot)\\), which projects features to the logit space \\(\\mathbb{R}^{H\\times W\\times |C|}\\). We adopt Self-Training (ST) and pixel-wise contrastive learning to our framework, as described in Sec. 3.1. The supervision for \\(\\mathcal{D}_u\\) is produced by the collaboration between the logit and representation spaces, as described in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.342, + 0.421, + 0.359 + ], + "angle": 0, + "content": "3.1. ST and Pixel-wise Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.473, + 0.595 + ], + "angle": 0, + "content": "The main idea of self-training is to pre-train a model on labeled images and use it to produce pseudo-labels as supervision for unlabeled images. A typical framework is the teacher-student framework [44], which consists of a student model and a teacher model. Both the student model and the teacher model are constructed by an encoder and a segmentation head. Parameters of the student model are optimized via Stochastic Gradient Decent (SGD) while parameters of the teacher model are updated by the Exponential Moving Average (EMA) of student model parameters. We denote the encoder and the segmentation head in the student model by \\( f(\\cdot) \\) and \\( g(\\cdot) \\) while denoting those in the teacher model by \\( f^{\\prime}(\\cdot) \\) and \\( g^{\\prime}(\\cdot) \\). The pseudo-labels \\( \\hat{\\pmb{y}}_i^{u,lgt} \\) are produced based on the teacher model's output logits \\( \\hat{\\pmb{p}}_i^u = g'(f'(x_i^u)) \\) in logit space, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.607, + 0.47, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {y}} _ {i} ^ {u, l g t} = \\mathbf {1} _ {c} (\\underset {c} {\\arg \\max } \\{\\hat {\\boldsymbol {p}} _ {i, c} ^ {u} \\} _ {c \\in C}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.644, + 0.424, + 0.659 + ], + "angle": 0, + "content": "where \\(\\mathbf{1}_c(\\cdot)\\) denotes the one-hot encoding of class \\(c\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.902 + ], + "angle": 0, + "content": "In order to enhance the ability of the model itself to extract features, recent works [32, 46, 1] additionally employ pixel-wise contrastive learning and introduce a representation head to both teacher and student models. We denote the representation head in the student model as \\( h(\\cdot) \\) and that in the teacher model as \\( h^{\\prime}(\\cdot) \\). The pixel \\( x_{i} \\) of the class \\( c \\) are projected as representations \\( z_{ci} \\) in representation space by the cooperating of \\( f(\\cdot) \\) and \\( h(\\cdot) \\), i.e., \\( z_{ci} = h(f(x_i)) \\). And the representation \\( z_{ci} \\) is then aggregated to its class centroid (prototype) while separated from representations in different classes \\( z_{\\tilde{c} i} \\) (negatives). The semantic guidance for contrastive learning is from the combination of ground truth \\( y_{i}^{l} \\) and pseudo-labels \\( \\hat{y}_i^{u,lgt} \\) in logit space. Moreover, in order to emphasize the reliable and crucial aspects during unlabeled and contrastive learning, a sampling strategy is adopted to select valid pixels \\( x_{i} \\) according to their corre" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "sponding confidence, i.e. the student model's output logits \\( \\pmb{p}_i \\) after a Softmax operation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.894, + 0.288 + ], + "angle": 0, + "content": "Discussion. In recent works [46, 1, 32], the supervision of unlabeled images is derived solely from the logit space. This overlooks the potential benefits of the supervision from the representation space, leading to two potential limitations: 1) the pseudo-labels \\(\\hat{y}_i^{u,lgt}\\) obtained from the logit space may contain noise and miss the opportunity to be corrected by semantic information from the representation space; 2) since the confidence from logit space is used as the indicator \\(\\hat{j}_i\\) for the sampling strategy, learning in the representation space may not be critical enough due to the different confusing parts between the two spaces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.894, + 0.38 + ], + "angle": 0, + "content": "To mitigate these limitations, we produce pseudo-labels from the representation space and combine them with pseudo-labels from the logit space to provide higher-quality supervision during unlabeled training. Meanwhile, we obtain a new indicator from the representation space for the more effective sampling strategy." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.387, + 0.84, + 0.404 + ], + "angle": 0, + "content": "3.2. Supervision from Representation Space" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.892, + 0.485 + ], + "angle": 0, + "content": "In this section, we detail the approach to obtain the pseudo-labels from the representation space. Meanwhile, we simultaneously access a new indicator for the sampling strategy in representation spaces, which provides a critical reference in the contrastive learning process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.892, + 0.561 + ], + "angle": 0, + "content": "Specifically, we first build a set of prototypes for each class and obtain the pseudo-labels by retrieving the nearest prototype for each representation. We calculate the centroid of all representations in the current class \\( c \\) as the prototype \\( \\rho_{c} \\), which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.57, + 0.892, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\rho_ {c} = \\frac {1}{N _ {c}} \\sum_ {i} ^ {N _ {c}} z _ {i} ^ {\\prime}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.892, + 0.69 + ], + "angle": 0, + "content": "where \\(N_{c}\\) is the total number of representations of current class \\(c\\) and \\(\\pmb{z}_i^{\\prime}\\) is the representation projected by the cooperation of the \\(f^{\\prime}(\\cdot)\\) and \\(h^\\prime (\\cdot)\\). Meanwhile, to include more representation information, we update the prototype along the sequential iterations with EMA as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.701, + 0.892, + 0.718 + ], + "angle": 0, + "content": "\\[\n\\hat {\\rho} _ {c} (t) = \\alpha \\hat {\\rho} _ {c} (t - 1) + (1 - \\alpha) \\rho_ {c} (t), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.726, + 0.892, + 0.803 + ], + "angle": 0, + "content": "where \\(\\hat{\\rho}_c(t),\\hat{\\rho}_c(t - 1)\\) mean the current \\(t^{th}\\) prototype and last \\((t - 1)^{th}\\) prototype in iterations, \\(\\rho_{c}(t)\\) means the prototype calculated by Eq.2 in current iteration, and \\(\\alpha\\) is a hyper-parameter that controls the updating speed. Thus, the pseudo-label from the representation space is achieved by:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.809, + 0.892, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {y}} _ {i} ^ {u, r e p} = \\mathbf {1} _ {\\hat {c}} (\\hat {c}), w i t h \\hat {c} = \\underset {c} {\\arg \\max } \\left\\{\\operatorname {s i m} \\left(\\boldsymbol {z} _ {i} ^ {\\prime}, \\hat {\\boldsymbol {\\rho}} _ {c} (t)\\right) \\right\\} _ {c \\in C}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.819, + 0.856 + ], + "angle": 0, + "content": "where \\(sim(\\cdot)\\) is defined as the cosine similarity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As for the indicator for the sampling strategy in the representation space, we use the Softmax function on the similarity among the representation and all prototypes, which is" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "933" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.09, + 0.848, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.339, + 0.895, + 0.385 + ], + "angle": 0, + "content": "Figure 2. Overview of our framework. Our training pipeline consists of learning in two spaces: logit space and representation space. The pseudo-labels \\(\\hat{\\pmb{y}}_i^u\\) during unlabeled training are produced by the collaboration of two spaces with mix pseudo-labeling strategy (1) or cross pseudo-labeling strategy (2). The indicator for representation learning is produced by similarity \\((s_1, s_2, \\text{and} s_3)\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.407, + 0.163, + 0.421 + ], + "angle": 0, + "content": "followed as:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.429, + 0.47, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\hat {j} _ {i} ^ {u, r e p} = \\frac {e ^ {s i m \\left(\\boldsymbol {z} _ {\\boldsymbol {c} i} , \\hat {\\rho} _ {\\boldsymbol {c}} (t)\\right) / \\tau}}{e ^ {s i m \\left(\\boldsymbol {z} _ {\\boldsymbol {c} i} , \\hat {\\rho} _ {\\boldsymbol {c}} (t)\\right) / \\tau} + \\sum_ {\\tilde {c} \\in \\tilde {C}} e ^ {s i m \\left(\\boldsymbol {z} _ {\\boldsymbol {c} i} , \\hat {\\rho} _ {\\tilde {c}} (t)\\right) / \\tau}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.47, + 0.58 + ], + "angle": 0, + "content": "where \\(\\hat{\\rho}_{\\tilde{c}}(t)\\) means the prototype with different classes from \\(z_{ci}\\) and \\(\\tau\\) is a hyper-parameter. Different from using confidence from logit space as the indicator to involve representation learning [32, 46], the Softmax similarity directly helps the model to discover the confusion between representations and their prototypes and focus on them during the subsequent training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.587, + 0.383, + 0.604 + ], + "angle": 0, + "content": "3.3. Collaboration Between Two Spaces" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.47, + 0.657 + ], + "angle": 0, + "content": "With the pseudo-labels in two spaces, we propose two pseudo-labeling strategies to strengthen the collaboration between two spaces and obtain more reliable pseudo-labels." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.665, + 0.47, + 0.773 + ], + "angle": 0, + "content": "- Mix pseudo-labeling. To mitigate the negative effects of inherent noise from both two spaces during pseudolabeling, we adopt the mix pseudo-labeling strategy that only considers the mutually agreeable pseudolabels between the two spaces. Specifically, we define the set of final pseudo-labels as \\(\\hat{Y}^u = \\hat{Y}^{u,lgt} \\cap \\hat{Y}^{u,rep}\\), where \\(\\hat{\\mathbf{y}}_i^{u,lgt} \\in \\hat{Y}^{u,lgt}\\) and \\(\\hat{\\mathbf{y}}_i^{u,rep} \\in \\hat{Y}^{u,rep}\\)." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.781, + 0.471, + 0.901 + ], + "angle": 0, + "content": "- Cross pseudo-labeling. Inspired by recent researches [7, 38] that maintain consistency among the predictions of the same image across different models or decoders in different views, we propose a cross pseudolabeling strategy that leverages pseudo-labels from one space to supervise the other. Specifically, we use pseudo-labels \\(\\hat{\\pmb{y}}_i^{u,rep}\\) to supervise the logit space, and vice versa." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.665, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.892, + 0.618 + ], + "angle": 0, + "content": "The strengths of using pseudo-labels from two spaces in the collaborative way are twofold: 1) obtaining more reliable supervision during unlabeled training, and 2) enabling the strengths of learning in different spaces to complement each other. Since the learning in different feature spaces concentrates on different parts of features, i.e., the logit space mainly focuses on the most discriminative part of features while the representation space treats all parts of features equally, the performance of pseudo-labels from two spaces varies across different classes or regions of images. Therefore, our collaborative pseudo-labeling strategies exchange knowledge between two spaces and provide higher-quality supervision during unlabeled training. The experimental proof is in Sec. 5.1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.893, + 0.773 + ], + "angle": 0, + "content": "As for indicators, we use confidence as the indicator \\(\\hat{j}_i^{u,lgt}\\) for learning logit space and Softmax similarity as the indicator \\(\\hat{j}_i^{u,rep}\\) for learning representation space. We argue that the confusing parts of learning in both two spaces are varied due to the different parts of features being concentrated in each space. Therefore, this strategy allows the learning in different spaces to focus on their own confusing parts, which can be more effective than using a single indicator when mining confusing parts for both spaces in the training process. The experimental proof is in Sec. 5.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.682, + 0.801 + ], + "angle": 0, + "content": "3.4. Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.807, + 0.892, + 0.903 + ], + "angle": 0, + "content": "With the indicators \\(\\hat{j}_i^{u,lgt}\\) and \\(\\hat{j}_i^{u,rep}\\), we adopt some threshold sampling strategies. In logit space, we set a threshold \\(\\delta_u\\) during unlabeled learning and logits \\(\\hat{p}_i^u\\) whose indicator \\(\\hat{j}_i^{u,lgt}\\) is higher than \\(\\delta_u\\) will be regarded as the valid logits in logit space. In representation space, our sample strategy can be divided into three parts: 1) Valid Sam" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "934" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.272 + ], + "angle": 0, + "content": "pling Strategy. Similar to the sampling strategy in logit space, a threshold \\(\\delta_w\\) is used to sample representations whose indicator \\(\\hat{j}_i^{u,rep}\\) is higher than \\(\\delta_w\\). 2) Hard Sampling Strategy. We adopt the hard sampling strategy for tilting to train those confusing representations. Specifically, we set a threshold \\(\\delta_s\\) to sample representations whose indicator \\(\\hat{j}_i^{u,rep}\\) is lower than \\(\\delta_s\\). 3) Negative Sampling Strategy. We sample negatives according to the similarity between the prototype of current class \\(c\\) and other prototypes. Concretely, the negatives are more likely to be sampled if its prototype is more similar to the prototype of the current class." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.273, + 0.47, + 0.35 + ], + "angle": 0, + "content": "Cooperated with the ground truth \\( \\pmb{y}_i^l \\), pseudo-labels \\( \\pmb{y}_i^u \\) produced from two spaces in a collaborative way, and different sampling strategies in two spaces, the total learning object is composed with a supervised loss \\( \\mathcal{L}_s \\), an unsupervised loss \\( \\mathcal{L}_u \\), and a contrastive loss \\( \\mathcal{L}_c \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.364, + 0.47, + 0.38 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {s} + \\mathcal {L} _ {u} + \\lambda_ {c} \\mathcal {L} _ {c}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.393, + 0.473, + 0.454 + ], + "angle": 0, + "content": "where \\(\\lambda_{c}\\) is used to tune the contribution between logit space and representation space. Specifically, \\(\\mathcal{L}_s\\) and \\(\\mathcal{L}_u\\) are constructed by the Cross-Entropy (CE) \\(\\ell_{ce}\\) and can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.463, + 0.47, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\frac {1}{| \\mathcal {B} _ {l} |} \\sum_ {\\boldsymbol {x} _ {i} ^ {l} \\in \\mathcal {B} _ {l}} \\ell_ {c e} \\left(\\boldsymbol {p} _ {i} ^ {l}, \\boldsymbol {y} _ {i} ^ {l}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.515, + 0.47, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {u} = \\frac {1}{| \\hat {\\mathcal {B}} _ {u} |} \\sum_ {\\boldsymbol {x} _ {i} ^ {u} \\in \\hat {\\mathcal {B}} _ {u}} \\ell_ {c e} \\left(\\boldsymbol {p} _ {i} ^ {u}, \\hat {\\boldsymbol {y}} _ {i} ^ {u}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.561, + 0.47, + 0.622 + ], + "angle": 0, + "content": "where \\(\\mathcal{B}_l\\) denotes the labeled images in a mini-batch and \\(\\hat{\\mathcal{B}}_u\\) is the subsets that sampled from unlabeled images in a mini-batch according to the sampling strategy. Meanwhile, the contrastive loss \\(\\mathcal{L}_c\\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.632, + 0.47, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = - \\frac {1}{| C | \\times | \\hat {\\mathcal {Z}} _ {c} |} \\sum_ {c \\in C} \\sum_ {\\boldsymbol {z} _ {c i} \\in \\hat {\\mathcal {Z}} _ {c}} \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.671, + 0.436, + 0.705 + ], + "angle": 0, + "content": "\\[\nl o g [ \\frac {e ^ {s (\\pmb {z} _ {c i} , \\hat {\\pmb {\\rho}} _ {c}) / \\tau}}{e ^ {s (\\pmb {z} _ {c i} , \\hat {\\pmb {\\rho}} _ {c} (t))) / \\tau} + \\sum_ {\\tilde {e} \\in \\tilde {C}} \\sum_ {\\pmb {z} _ {\\tilde {e} i} \\in \\hat {\\pmb {Z}} _ {\\tilde {c}}} e ^ {s (\\pmb {z} _ {c i} , \\pmb {z} _ {\\tilde {c} i}) / \\tau} ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.825 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathcal{Z}}_c\\) is the subset sampled from the set of the representations which belong to class \\(c\\) according to the sampling strategy, \\(\\hat{\\mathcal{Z}}_{\\tilde{c}}\\) is the subset sampled from the set of the representations which bot belong to class \\(c\\), \\(\\tilde{C}\\) denotes the subset of \\(C\\) with class \\(c\\) removed, and the supervision information comes from the final pseudo-labels \\(\\hat{\\pmb{y}}_i^u\\) after the pseudo-labeling strategies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The whole framework is shown in Fig. 2. All the pseudocode of producing pseudo-labels and their indicators from the logit and representation spaces is shown in Algorithm 1, and the pseudo-code of the mix pseudo-labeling strategy is shown in Algorithm 2. The pseudo-code is PyTorch-like." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Algorithm 1 Pseudo-code of producing pseudo-labels and indicators" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.121, + 0.892, + 0.15 + ], + "angle": 0, + "content": "Network: Teacher's encoder \\( f' \\), segmentation head \\( g' \\), representation head \\( h' \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.152, + 0.892, + 0.182 + ], + "angle": 0, + "content": "Input: Unlabeled mini-batch \\( B_{u} \\) consists of \\( X^{u} \\), the set of prototypes \\( \\{\\hat{\\rho}_c(t)\\}_{c\\in C} \\)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.183, + 0.643, + 0.196 + ], + "angle": 0, + "content": "1: for \\( X_{u} \\in B_{u} \\) do" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.197, + 0.805, + 0.211 + ], + "angle": 0, + "content": "2: # Predict logits from unlabeled images." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.212, + 0.691, + 0.227 + ], + "angle": 0, + "content": "3: \\(\\hat{P}^{u,lgt}\\gets g'(f'(X^u))\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.228, + 0.892, + 0.256 + ], + "angle": 0, + "content": "4: # Produce labels and indicators based on logits with Eq. 1." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.257, + 0.79, + 0.273 + ], + "angle": 0, + "content": "5: \\(\\hat{Y}^{u,lgt},\\hat{J}^{u,lgt}\\gets lgt\\text{-} pseudo(\\tilde{P}^{u,lgt})\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.273, + 0.867, + 0.287 + ], + "angle": 0, + "content": "6: # Predict representations from unlabeled images." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.287, + 0.695, + 0.303 + ], + "angle": 0, + "content": "7: \\(\\hat{R}^{u,rep}\\gets h'(f'(X^u))\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.303, + 0.891, + 0.332 + ], + "angle": 0, + "content": "8: # Produce labels and indicators based on representations with Eq. 4 and Eq. 5." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.332, + 0.888, + 0.349 + ], + "angle": 0, + "content": "9: \\(\\hat{Y}^{u,rep},\\hat{J}^{u,rep}\\gets rep\\text{-} pseudo(\\hat{R}^{u,rep},\\{\\hat{\\rho}_c(t)\\}_{c\\in C})\\)" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.349, + 0.585, + 0.361 + ], + "angle": 0, + "content": "10: end for" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.183, + 0.892, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.382, + 0.889, + 0.398 + ], + "angle": 0, + "content": "Algorithm 2 Pseudo-code of mix pseudo-labeling strategy" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.401, + 0.891, + 0.432 + ], + "angle": 0, + "content": "Input: Pseudo-labels from logit space \\(\\hat{Y}^{u,lgt}\\) and from representation space \\(\\hat{Y}^{u,rep}\\). Indicators \\(\\hat{J}^{u,lgt}\\) and \\(\\hat{J}^{u,rep}\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.433, + 0.892, + 0.462 + ], + "angle": 0, + "content": "Notation: Threshold \\(\\delta_{u}\\) for sampling strategy in \\(\\mathcal{L}_u\\), weak threshold \\(\\delta_w\\)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.463, + 0.892, + 0.493 + ], + "angle": 0, + "content": "1: # Sampling valid pseudo-labels according to the sampling strategy in logit space." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.494, + 0.799, + 0.512 + ], + "angle": 0, + "content": "2: \\(\\hat{Y}_{val}^{u,lgt}\\gets\\) sample_lgt( \\(\\hat{Y}^{u,lgt},\\hat{J}^{u,lgt},\\delta_u)\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.512, + 0.892, + 0.54 + ], + "angle": 0, + "content": "3: # Sampling valid pseudo-labels according to the valid sampling strategy in logit space." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.54, + 0.815, + 0.558 + ], + "angle": 0, + "content": "4: \\(\\hat{Y}_{val}^{u,rep}\\gets\\) sample_rep( \\(\\hat{Y}^{u,rep},\\hat{J}^{u,rep},\\delta_w)\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.558, + 0.81, + 0.571 + ], + "angle": 0, + "content": "5: # Produce the mask for mix pseudo-labels." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.571, + 0.722, + 0.588 + ], + "angle": 0, + "content": "6: \\(mask_{y} \\gets \\hat{Y}^{u,rep}.\\) eq(\\(\\hat{Y}^{u,lgt}\\))" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.588, + 0.872, + 0.601 + ], + "angle": 0, + "content": "7: # Produce mix pseudo-labels according to the mask." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.601, + 0.781, + 0.618 + ], + "angle": 0, + "content": "8: \\(\\hat{Y}^u\\gets\\) mask_pseudo \\((\\hat{Y}^{u,lgt},\\text{mask}_y)\\)" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.463, + 0.892, + 0.618 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.65, + 0.625, + 0.665 + ], + "angle": 0, + "content": "3.5. Discussions" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.674, + 0.892, + 0.703 + ], + "angle": 0, + "content": "We discuss the relations between our framework and other most-related S4 frameworks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Compare with PCR [51]. The key insight of PCR and our method lies in the promotion of consistency between the output in the logit and representation spaces. However, PCR exclusively relies on pseudo-labels derived from logit space as the guidance in the logit and representation spaces during the unsupervised training process, thereby neglecting the exploitation of the extensive semantic information inherent in representations and prototypes. In contrast, our method adopts a collaborative way by combining pseudo-labels obtained from both the logit and representation spaces, which improves the quality of pseudo-labels and enhances the knowledge exchange between the two spaces. In addition, since these two approaches are orthogonal, one can employ" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "935" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "our pseudo-labeling and indicator strategies within the PCR framework, or alternatively incorporate multiple prototypes and consistency loss into our approach. Such combinations may yield further improvements in performance, however, also increase the computational complexity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.471, + 0.365 + ], + "angle": 0, + "content": "Compare with CPS [7]. CPS contains two segmentation models initialized differently. It leverages pseudo-labels generated by one model to supervise the other, thereby maintaining the consistency between the output of the two models in the logit space. In a similar vein to CPS, our cross pseudo-labeling strategy in Sec. 3.3 is also motivated by the objective of maintaining consistency during unsupervised training. However, our method is distinct from CPS in that we focus on preserving consistency between the logit space and the representation space. This distinction confers a distinct advantage in terms of memory efficiency, as we solely introduce a MLP as the representation head, rather than introducing an additional segmentation model as in CPS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.471, + 0.532 + ], + "angle": 0, + "content": "Compare with other contrastive based S4 works [32, 1, 46, 23]. We adhere to prior works to build our prototype, describe in Eq. 2. Different from them, we update our prototypes iteratively, similar to [60]. Our sampling strategies are based on the threshold, similar to [32, 23, 45]. However, it is worth noting that our indicators which serve as the basis for comparison with the threshold, are derived from both the logit and representation spaces. This stands in contrast to prior approaches where indicators solely originate from the logit space. The efficacy of this modification is substantiated and extensively discussed in Sec. 5.2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.548, + 0.21, + 0.565 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.574, + 0.16, + 0.59 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.471, + 0.87 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on PASCAL VOC 2012 dataset [12], Cityscapes dataset [9], ADE20K dataset [59], and COCO-Stuff 10K dataset [2] to validate the effectiveness of our proposed method. The original PASCAL VOC 2012 dataset contains 1,464 labeled images in train set and 1,449 images for validation in val set. Following [7, 37], we additionally introduce 9,118 images from SBD [18] as training images. Since the labels in SBD are coarsely annotated, following [46], we use both classic VOC train set (1,464 candidate labeled images) and blender VOC train set (10,582 candidate labeled images). Cityscapes dataset is a dataset for urban scene understanding, which contains 2,975 images in train set and 500 images in val set. ADE20K has 20,210 and 2,000 images in train and val set, with 150 classes in total. COCO-Stuff 10K has 9,000 images in train set and 1,000 images in val set, with 181 classes. The approach of preprocessing labels is followed by MMSegmentation [8]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Network structure. We use Deeplabv3+ [4] with ResNet-101 [20] pre-trained on ImageNet [10] as our network struc" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "ture. The segmentation and representation head are composed of Conv-BN-ReLU-Conv." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.893, + 0.426 + ], + "angle": 0, + "content": "Implementation details. For training on PASCAL VOC 2012 dataset and COCO-Stuff 10K dataset, we set the learning rate as 0.0064, weight decay as 0.0005, crop size as 512 \\(\\times\\) 512, batch size as 16, and a total of 40,000 iterations. For training on the Cityscapes dataset, we set the learning rate as 0.0038, weight decay as 0.0005, crop size as \\(768 \\times\\) 768, batch size as 8, and a total of 80,000 iterations. For training on ADE20K dataset, we set the learning rate as 0.0064, weight decay as 0.0005, crop size as \\(512 \\times 512\\), batch size as 16, and a total of 80,000 iterations. We use the poly scheduling to decay the learning rate during the training process: \\(lr = lr_{base} \\times (1 - \\frac{epoch}{total\\_epoch})^{0.9}\\). We use the mean of Intersection over Union (mIoU) as the metric in evaluation. We use the sliding window strategy to evaluate the performance of our method on the Cityscapes dataset, following [7]. In addition, when adopting our cross-labeling strategy, due to the requirement of a set of high-quality prototypes when classifying each representation, we first solely use the supervision from logit space for 20 epochs to initialize the prototypes." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.437, + 0.809, + 0.453 + ], + "angle": 0, + "content": "4.2. Comparison with Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.763 + ], + "angle": 0, + "content": "In this subsection, we first reproduce three baselines: MT [44], CutMix [15], and a same contrastive-based framework with us but with only logit space pseudo-labels and indicators (Baseline) on classic VOC train set. Meanwhile, we make the comparison of our method with mix (CSS (mix)) and cross (CSS (crs.)) pseudo-labeling strategy on blender VOC train set and Cityscapes train set with following recent SOTA S4 methods: CCT [38], CPS [7], U²PL [46], ST++ [52], PRCL [48], PCR [51], and PSMT [33]. Since the data split will dramatically affect the performance in S4, i.e., choosing labeled images plays an important role in the final results, we conduct experiments with three different data splits and report the mean value and standard deviation (blue color numbers). Since the mix pseudo-labeling strategy has better performance, we only use CSS (mix) when compared with SOTAs. Meanwhile, we use ResNet-101 with deep stem blocks as our network structure when compared with SOTAs. Since there is no uniform data split, we use the data splits in U²PL [46]. We use the OHEM loss when training Cityscapes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Results on PASCAL VOC 2012. Tab. 1 shows the comparison with our baselines on Classic PASCAL VOC 2012 set. Our method consistently outperforms baselines with an acceptable standard deviation on all label rates. Tab. 2 shows the comparison with the SOTAs on PASCAL VOC 2012." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Results on Cityscapes. Tab. 3 shows the performance of our method on Cityscapes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Results on ADE20K and COCO-Stuff 10K. Tab. 4 shows the results of our method on ADE20K and COCO-Stuff." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "936" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.47, + 0.142 + ], + "angle": 0, + "content": "Table 1. Results on classic VOC train set with four different label rate. Labeled data splits are from the original VOC train set. All approaches are reproduced with three runs and three different data splits." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.143, + 0.468, + 0.259 + ], + "angle": 0, + "content": "
Pascal VOC 2012 (Classic)
Method92183366732
Sup.51.57±3.5854.69±2.4464.86±1.0470.77±0.76
MT58.92±2.9961.63±1.7666.79±0.5371.58±0.51
CutMix65.82±3.6067.91±1.4772.53±0.5074.08±0.49
Baseline66.91±4.2170.32±1.8673.97±1.0476.49±0.58
CSS(crs.)67.03±4.5871.41±1.6774.47±1.0877.08±0.37
CSS(mix)68.09±4.8971.93±1.8874.91±1.1277.57±0.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.271, + 0.468, + 0.312 + ], + "angle": 0, + "content": "Table 2. Results on blender VOC train set. All the results from the recent papers [46, 52, 13, 29, 33, 51]. Labeled data is from the augmented VOC train set and the data splits are from [46, 33]." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.313, + 0.467, + 0.453 + ], + "angle": 0, + "content": "
Pascal VOC 2012 (Blender)
Method662132326465291
CCT [38]71.8673.6876.5177.40
CPS [7]74.4876.4477.6878.64
U2PL [46]77.2179.0179.3080.50
ST++ [52]74.7077.9077.90-
PRCL [48]76.9678.1679.0279.59
PCR [51]78.6080.7180.7880.91
PSMT [33]75.5078.2078.7279.76
CSS (mix)78.7379.5480.8281.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.467, + 0.469, + 0.522 + ], + "angle": 0, + "content": "Table 3. Results on Cityscapes. The model is trained on the Cityscapes train set, which consists of 2,975 samples in total, and tested on Cityscapes val set. And all the results from the recent papers [46, 51, 33]." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.522, + 0.467, + 0.636 + ], + "angle": 0, + "content": "
Cityscapes
Method1863727441488
CCT [38]69.3274.1275.9978.10
CPS [7]69.7874.3174.5876.82
U2PL [46]70.3074.3776.4779.05
PCR [51]73.4176.3178.4079.11
PSMT [33]-76.8977.6079.09
CSS (mix)74.0276.9377.9479.62
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.661, + 0.226, + 0.679 + ], + "angle": 0, + "content": "5. Ablative Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.469, + 0.807 + ], + "angle": 0, + "content": "The main contribution of our work lies in 1) collaborative pseudo-labeling strategies and 2) a new indicator for representation learning. To further prove the effectiveness of our proposed method, we conduct ablative studies on these two points. We choose Deeplabv3+ with ResNet-101 pre-trained on ImageNet as our backbone and leverage 92 labeled images and 183 labeled images in PASCAL VOC 2012. The other settings are the same as those in Sec. 4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.466, + 0.833 + ], + "angle": 0, + "content": "5.1. Effectiveness of Collaborative pseudo-labeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Quality of pseudo-labels. To illustrate the superiority of using pseudo-labels from two spaces in a collaborative way as supervision, we conduct experiments to show the quality of pseudo-labels obtained 1) from logit space (lgt.), 2) from" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.568, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.091, + 0.631, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.19, + 0.615, + 0.201 + ], + "angle": 0, + "content": "(b) GT" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.091, + 0.694, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.091, + 0.759, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.14, + 0.759, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.19, + 0.755, + 0.2 + ], + "angle": 0, + "content": "(d) lgt. label" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.091, + 0.823, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.14, + 0.825, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.19, + 0.824, + 0.2 + ], + "angle": 0, + "content": "(e) rep. mask" + }, + { + "type": "image", + "bbox": [ + 0.828, + 0.091, + 0.89, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.828, + 0.14, + 0.89, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.829, + 0.19, + 0.887, + 0.2 + ], + "angle": 0, + "content": "(f) rep. label" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.208, + 0.88, + 0.222 + ], + "angle": 0, + "content": "Figure 3. Differences between pseudo-labels in different spaces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.25, + 0.892, + 0.414 + ], + "angle": 0, + "content": "representation space (rep.), and 3) from the mix pseudolabeling strategy (mix). The pseudo-labels are sampled with corresponding sampling strategies in Sec. 3.4. Tab. 5 illustrates the IoU of pseudo-labels for each class on PASCAL VOC 2012 with 92 labeled images. The results clearly indicate that employing pseudo-labels from the representation space enhances the accuracy of the final pseudo-labels in most classes. This improvement is particularly evident in classes that are originally under-performing, such as the IoU improvement of \\(11.42\\%\\) for the chair class and \\(11.58\\%\\) for the sofa class." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Meanwhile, we also visualize the pseudo-labels obtained from logit space (lgt.) and representation space (rep.) in Fig. 3. Fig. 3 (c) and (e) show the masks for pseudolabels produced by the sampling strategies. In particular, the white color represents the valid pixels used during unlabeled learning, while the black color indicates the discarded pixels. Fig. 3 (d) and (f) are the pseudo-labels we obtained from two spaces. The figure clearly illustrates the differences between pseudo-labels produced by different spaces. For example, the parts of the instance edge in pseudo-labels are usually discarded since they are challenging for learning in logit space. However, pseudo-labels from representation space will easily tackle this problem (first row). The pseudo-labels from representation space are more inaccurate in some complex scenes, which be resolved by combining pseudo-labels from logit space (second row)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We mainly attribute the differences in pseudo-label to the differing concentrations of learning in the two spaces. Specifically, learning in the logit space primarily emphasizes the most discriminative part of features, whereas that in the representation space treats each part of features equally. As a result, learning in the logit space may overlook minor feature differences, leading to sub-optimal performance in predicting instance edges and distinguishing between similar classes (e.g., chair and sofa). Conversely, learning in the representation space produces balanced performance across all image parts and classes. However, this can lead to erroneous predictions for classes with high intra-class variance (e.g., background). By leveraging pseudo-labels from the two spaces, we capitalize on the strengths of the learning in each space and enhance the knowledge exchange between the two spaces." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "937" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.178, + 0.088, + 0.791, + 0.101 + ], + "angle": 0, + "content": "Table 4. Results on ADE20K and COCO-Stuff 10K with four different label rates and single data splits." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.102, + 0.891, + 0.188 + ], + "angle": 0, + "content": "
ADE20KCOCO-Stuff 10K
Method1/161/81/41/21/161/81/41/2
Cutmix28.9131.8934.6239.1523.0926.1429.8830.91
Baseline31.1133.2436.4639.9126.5727.9128.9731.08
CSS (crs.)32.0133.9237.8439.9727.0628.8530.5931.73
CSS (mix)32.5534.2137.0140.8527.4629.5129.9831.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.201, + 0.869, + 0.214 + ], + "angle": 0, + "content": "Table 5. The quality of pseudo-labels from different pseudo-labeling strategies. The pseudo-labels are sampled by sampling strategies." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.214, + 0.891, + 0.33 + ], + "angle": 0, + "content": "
sourcebackaero.bicy.birdboatbott.buscarcatchaircow
lgt.96.7895.1477.4793.3881.3787.5496.7695.4894.474.0992.15
rep.90.7196.5061.4275.7553.3054.6584.4680.5591.1128.0388.16
mix96.66↓0.1298.12↑2.9882.76↑5.2994.47↑1.0985.12↑3.7589.94↑2.4097.03↑0.2795.98↑0.5094.65↑0.1819.01↑14.9294.81↑2.66
sourcetabeldoghorsemotorpers.plantsheepsofatraintvmIoU
lgt.58.3494.1293.0191.3793.6850.3391.1018.1686.6564.8978.87
rep.52.2386.9773.2288.7091.7556.1366.7835.2588.5162.6171.57
mix64.35↑6.0194.33↑0.2193.65↑0.6491.50↑0.1393.01↓0.6755.64↑5.3191.56↑0.4629.74↑11.5893.54↑6.8969.48↑4.5982.17↑3.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.353, + 0.468, + 0.378 + ], + "angle": 0, + "content": "Table 6. Results on pseudo-labels from different sources on two different label rates." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.378, + 0.468, + 0.451 + ], + "angle": 0, + "content": "
source92 labels183 labels
logit space67.1170.32
representation space64.2067.52
mix pseudo-labeling68.41↑1.3072.74↑2.42
cross pseudo-labeling67.85↑0.7471.98↑1.66
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.476, + 0.47, + 0.763 + ], + "angle": 0, + "content": "Results of different strategies To investigate the involvement of different pseudo-labeling strategies, we conduct the experiments as follows: 1) Using pseudo-labels from logit space. 2) Using pseudo-labels from representation space. 3) Using mix pseudo-labeling strategy. 4) Using cross pseudolabeling strategy. Tab. 6 shows the effectiveness of our proposed strategy in two different label rates. The results show that the performances of experiments with collaborative pseudo-labeling strategies are better than the ones whose pseudo-labels come from a single space with two different label rates, which proves the effectiveness of our proposed collaboration between the two spaces. It is worth noting that even though the quality of pseudo-labels from representation space is lower than that from logit space, the performance of the model is also boosted by using the cross pseudo-labeling strategy to maintain consistency between the predictions in two spaces. In addition, our method with the mix pseudo-labeling strategy outperforms that with the cross pseudo-labeling strategy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.337, + 0.787 + ], + "angle": 0, + "content": "5.2. Effectiveness of the Indicator" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Limitation of merely using confidence. To explain the limitation of merely using confidence for learning in the logit and representation spaces, we conduct experiments to show the relations between confidence and similarity. The similarity is the cosine similarity between representations and prototypes, which directly shows the confusion level between the representation and the prototype of its class." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.358, + 0.679, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.448, + 0.602, + 0.457 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.359, + 0.89, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.448, + 0.783, + 0.459 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.465, + 0.855, + 0.479 + ], + "angle": 0, + "content": "Figure 4. Relations between similarity and confidence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Fig. 4 (a) shows the comparison between the confidence of each prediction and the corresponding similarity. We use the class person for demonstrating in our experiments. It shows clearly that even though fixing the confidence into a small range (from 0.8 to 0.81 in our settings), the similarity varies. Meanwhile, in Fig. 4 (b), different color bars stand for different intervals of confidence, and the lines denote the mean similarity between the prototype and each representation whose corresponding confidence is in the current interval. Fig. 4 (b) illustrates that the mean similarity of the class fluctuates when its interval of confidence rises. Both two figures imply that confidence is not able to represent the confusion level between representations and prototypes since there are no direct and close relations between confidence and similarity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Fig. 5 visualizes the similarity and confidence of an image in both logit (lgt.) and representation (rep.) spaces, indicating the varying levels of confusion in the same region when learning in different spaces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.839 + ], + "angle": 0, + "content": "We also attribute it to the different concentrations of learning two spaces, i.e., the confusing region in one space can be more readily addressed in the other space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Thus, it is inappropriate to choose confidence as the indicator to involve representation learning, e.g., sampling more hard samples with threshold and indicator. In contrast, our indicator directly employs similarity between the represen" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "938" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.095, + 0.11, + 0.104 + ], + "angle": 0, + "content": "high" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.113, + 0.176, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.192, + 0.154, + 0.203 + ], + "angle": 0, + "content": "(a) Image" + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.115, + 0.272, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.206, + 0.192, + 0.243, + 0.203 + ], + "angle": 0, + "content": "(b) GT" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.115, + 0.367, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.192, + 0.356, + 0.204 + ], + "angle": 0, + "content": "(c) lgt. space" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.115, + 0.466, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.192, + 0.453, + 0.204 + ], + "angle": 0, + "content": "(c) rep. space" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.215, + 0.456, + 0.23 + ], + "angle": 0, + "content": "Figure 5. Visualization of the confusing part in different spaces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.254, + 0.468, + 0.315 + ], + "angle": 0, + "content": "tation and prototype of its class, which directly reflects the confusion level in representation learning. It is more accurate to use similarity as the indicator to sample hard and critical samples in representation learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.315, + 0.468, + 0.482 + ], + "angle": 0, + "content": "Results of different indicators. Tab. 7 shows the impact of using different indicators. We conduct experiments on two different label rates (92 and 183) and three different indicators for pseudo-labels: only confidence (conf.), only similarity (smlr.), and confidence for learning logit space while similarity for learning representation space (mix). We use two different approaches to obtain pseudo-labels: from logit space only (seg label) and from mix pseudo-labeling strategy (mix label). It is clear that using both confidence and similarity to involve the learning in their own spaces obtains the best performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.493, + 0.468, + 0.518 + ], + "angle": 0, + "content": "Table 7. Results on indicators from different spaces on two different label rates." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.518, + 0.468, + 0.591 + ], + "angle": 0, + "content": "
92 labels183 labels
sourceseg labelmix labelseg labelmix label
conf.67.1168.3370.3271.92
smlr.66.1666.8968.9069.25
mix67.80↑0.6968.41↑1.3071.50↑1.1872.74↑2.42
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.618, + 0.347, + 0.634 + ], + "angle": 0, + "content": "5.3. Ablation study of Components" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.468, + 0.807 + ], + "angle": 0, + "content": "In this section, we conduct experiments to introduce our components in CSS step by step, with results shown in Tab. 8. Our baseline is the conventional contrastive-based S4, achieving mIoU of \\(67.11\\%\\) on 92 labels and \\(70.32\\%\\) on 183 labels. Mix and cross means the pseudo-labels are from the mix pseudo-labeling strategy and cross pseudo-labeling strategy while the indicator is still the confidence in two spaces. Ind means we use the different indicators in different spaces while the pseudo-labels are from logit space. The last two rows represent our proposed two pseudo-labeling strategies with indicators from two spaces." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.26, + 0.833 + ], + "angle": 0, + "content": "5.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Fig. 6 shows the qualitative results of different methods on PASCAL VOC 2012 with 92 labeled images. Baseline means the conventional contrastive-based method. Compared with the original self-training methods (CutMix)," + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.088, + 0.862, + 0.1 + ], + "angle": 0, + "content": "Table 8. Ablation study on different components of our CSS." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.101, + 0.889, + 0.202 + ], + "angle": 0, + "content": "
component92 labels183 labels
baseline67.1170.32
mix68.33↑1.2271.92↑1.60
cross67.21↑0.1070.87↑0.55
ind67.80↑0.6971.50↑1.18
mix + ind68.41↑1.3072.74↑2.42
cross + ind67.85↑0.7471.98↑1.66
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.229, + 0.892, + 0.304 + ], + "angle": 0, + "content": "thanks to introducing pixel-wise contrastive learning, the baseline, and our method perform better in some ambiguous regions. Furthermore, benefiting from the supervision of two spaces and different indicators in different spaces, our method outperforms the baseline." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.317, + 0.89, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.544, + 0.892, + 0.559 + ], + "angle": 0, + "content": "images. Yellow boxes highlight the main differences." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.587, + 0.619, + 0.603 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.763 + ], + "angle": 0, + "content": "In this paper, we propose two collaborative pseudolabeling strategies to take full use of the semantic information in the representation space and enhance the knowledge exchange between the logit and representation spaces. Moreover, we employ a new indicator for the learning process in the representation space. Extensive experiments demonstrate that our pseudo-labeling strategies obtain more reliable supervision during unlabeled training and our indicator helps the model to concentrate on more critical parts during representation learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Future work: In this paper, we employ pseudo-labeling strategies to utilize the semantic information in both logit and representation spaces. In the future, we will investigate more powerful strategies to enhance the knowledge exchange between two spaces." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "939" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[1] Inigo Alonso, Alberto Sabater, David Ferstl, Luis Montesano, and Ana C. Murillo. Semi-supervised semantic segmentation with pixel-level contrastive learning from a class-wise memory bank. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8219-8228, October 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.47, + 0.324 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.395 + ], + "angle": 0, + "content": "[4] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.47, + 0.45 + ], + "angle": 0, + "content": "[5] Mu Chen, Zhedong Zheng, Yi Yang, and Tat-Seng Chua. Pipa: Pixel-and patch-wise self-supervised learning for domain adaptive semantic segmentation. arXiv preprint arXiv:2211.07609, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.47, + 0.507 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.47, + 0.577 + ], + "angle": 0, + "content": "[7] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.468, + 0.633 + ], + "angle": 0, + "content": "[8] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.468, + 0.718 + ], + "angle": 0, + "content": "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.468, + 0.802 + ], + "angle": 0, + "content": "[11] Nanqing Dong and Eric P Xing. Few-shot semantic segmentation with prototype learning. In BMVC, volume 3, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[12] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Jiashuo Fan, Bin Gao, Huan Jin, and Lihui Jiang. Ucc: Uncertainty guided cross-head co-training for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), pages 9947-9956, June 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[14] Zhengyang Feng, Qianyu Zhou, Qiqi Gu, Xin Tan, Guangliang Cheng, Xuequan Lu, Jianping Shi, and Lizhuang Ma. Dmt: Dynamic mutual training for semi-supervised learning. Pattern Recognition, page 108777, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[15] Geoff French, Timo Aila, Samuli Laine, Michal Mackiewicz, and Graham Finlayson. Semi-supervised semantic segmentation needs strong, high-dimensional perturbations. In In 31th British Machine Vision Conference, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[17] Dayan Guan, Jiaxing Huang, Aoran Xiao, and Shijian Lu. Unbiased subclass regularization for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9968-9978, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[18] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In 2011 international conference on computer vision, pages 991-998. IEEE, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[21] Ruifei He, Jihan Yang, and Xiaojuan Qi. Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6930-6940, October 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[22] Judy Hoffman, Dequan Wang, Fisher Yu, and Trevor Darrell. Fcts in the wild: Pixel-level adversarial and constraint-based adaptation. arXiv preprint arXiv:1612.02649, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[23] Hanzhe Hu, Jinshi Cui, and Liwei Wang. Region-aware contrastive learning for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16291-16301, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[24] Hanzhe Hu, Fangyun Wei, Han Hu, Qiwei Ye, Jinshi Cui, and Liwei Wang. Semi-supervised semantic segmentation via adaptive equalization learning. Advances in Neural Information Processing Systems, 34:22106-22118, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[25] Wei-Chih Hung, Yi-Hsuan Tsai, Yan-Ting Liou, Yen-Yu Lin, and Ming-Hsuan Yang. Adversarial learning for semi-supervised semantic segmentation. arXiv preprint arXiv:1802.07934, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Zhengkai Jiang, Yuxi Li, Ceyuan Yang, Peng Gao, Yabiao Wang, Ying Tai, and Chengjie Wang. Prototypical contrast adaptation for domain adaptive semantic segmentation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 36-54. Springer, 2022." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "940" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[27] Tarun Kalluri, Girish Varma, Manmohan Chandraker, and CV Jawahar. Universal semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5259-5270, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[28] Rihuan Ke, Angelica I Aviles-Rivero, Saurabh Pandey, Saikumar Reddy, and Carola-Bibiane Schonlieb. A three-stage self-training framework for semi-supervised semantic segmentation. IEEE Transactions on Image Processing, 31:1805-1815, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[29] Donghyeon Kwon and Suha Kwak. Semi-supervised semantic segmentation with error localization network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9957-9967, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.471, + 0.356 + ], + "angle": 0, + "content": "[30] Daiqing Li, Junlin Yang, Karsten Kreis, Antonio Torralba, and Sanja Fidler. Semantic segmentation with generative models: Semi-supervised learning and strong out-of-domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8300-8311, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.471, + 0.425 + ], + "angle": 0, + "content": "[31] Jinlu Liu, Liang Song, and Yongqiang Qin. Prototype rectification for few-shot learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16, pages 741–756. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[32] Shikun Liu, Shuaifeng Zhi, Edward Johns, and Andrew Davison. Bootstrapping semantic segmentation with regional contrast. In International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[33] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4258-4267, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.471, + 0.636 + ], + "angle": 0, + "content": "[34] Binjie Mao, Xinbang Zhang, Lingfeng Wang, Qian Zhang, Shiming Xiang, and Chunhong Pan. Learning from the target: Dual prototype network for few shot semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1953-1961, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.471, + 0.691 + ], + "angle": 0, + "content": "[35] Sudhanshu Mittal, Maxim Tatarchenko, and Thomas Brox. Semi-supervised semantic segmentation with high-and low-level consistency. IEEE transactions on pattern analysis and machine intelligence, 43(4):1369-1379, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.471, + 0.76 + ], + "angle": 0, + "content": "[36] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979-1993, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[37] Viktor Olsson, Wilhelm Tranheden, Juliano Pinto, and Lennart Svensson. Classmix: Segmentation-based data augmentation for semi-supervised learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1369-1378, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[38] Yassine Ouali, Celine Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross-consistency training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[39] Yingwei Pan, Ting Yao, Yehao Li, Yu Wang, Chong-Wah Ngo, and Tao Mei. Transferrable prototypical networks for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2239-2247, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.204 + ], + "angle": 0, + "content": "[40] Jizong Peng, Guillermo Estrada, Marco Pedersoli, and Christian Desrosiers. Deep co-training for semi-supervised image segmentation. Pattern Recognition, 107:107269, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.893, + 0.245 + ], + "angle": 0, + "content": "[41] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.893, + 0.327 + ], + "angle": 0, + "content": "[42] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.893, + 0.398 + ], + "angle": 0, + "content": "[43] Korawat Tanwisuth, Xinjie Fan, Huangjie Zheng, Shujian Zhang, Hao Zhang, Bo Chen, and Mingyuan Zhou. A prototype-oriented framework for unsupervised domain adaptation. Advances in Neural Information Processing Systems, 34:17194-17208, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.893, + 0.455 + ], + "angle": 0, + "content": "[44] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.457, + 0.893, + 0.524 + ], + "angle": 0, + "content": "[45] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7303-7313, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.893, + 0.607 + ], + "angle": 0, + "content": "[46] Yuchao Wang, Haochen Wang, Yujun Shen, Jingjing Fei, Wei Li, Guoqiang Jin, Liwei Wu, Rui Zhao, and Xinyi Le. Semi-supervised semantic segmentation using unreliable pseudo labels. In Proceedings of the IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.893, + 0.69 + ], + "angle": 0, + "content": "[47] Chen Wei, Kihyuk Sohn, Clayton Mellina, Alan Yuille, and Fan Yang. Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10857-10866, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.893, + 0.761 + ], + "angle": 0, + "content": "[48] Haoyu Xie, Changqi Wang, Mingkai Zheng, Minjing Dong, Shan You, and Chang Xu. Boosting semi-supervised semantic segmentation with probabilistic representations. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2938-2946, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.893, + 0.83 + ], + "angle": 0, + "content": "[49] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.893, + 0.9 + ], + "angle": 0, + "content": "[50] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "941" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[51] Hai-Ming Xu, Lingqiao Liu, Qiuchen Bian, and Zhen Yang. Semi-supervised semantic segmentation with prototype-based consistency regularization. Advances in Neural Information Processing Systems, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[52] Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. St++: Make self-training work better for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4268-4277, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.289 + ], + "angle": 0, + "content": "[53] Jianlong Yuan, Yifan Liu, Chunhua Shen, Zhibin Wang, and Hao Li. A simple baseline for semi-supervised semantic segmentation with strong data augmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8229-8238, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.373 + ], + "angle": 0, + "content": "[54] Xiangyu Yue, Zangwei Zheng, Shanghang Zhang, Yang Gao, Trevor Darrell, Kurt Keutzer, and Alberto Sangiovanni Vincentelli. Prototypical cross-domain self-supervised learning for few-shot unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13834-13844, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.375, + 0.47, + 0.457 + ], + "angle": 0, + "content": "[55] Xiangyun Zhao, Raviteja Vemulapalli, Philip Andrew Mansfield, Boqing Gong, Bradley Green, Lior Shapira, and Ying Wu. Contrastive learning for label efficient semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10623-10633, October 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.459, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[56] Xu Zheng, Yunhao Luo, Hao Wang, Chong Fu, and Lin Wang. Transformer-cnn cohort: Semi-supervised semantic segmentation by the best of both students, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.502, + 0.469, + 0.584 + ], + "angle": 0, + "content": "[57] Xu Zheng, Jinjing Zhu, Yexin Liu, Zidong Cao, Chong Fu, and Lin Wang. Both style and distortion matter: Dualpath unsupervised domain adaptation for panoramic semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1285-1295, June 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.586, + 0.47, + 0.655 + ], + "angle": 0, + "content": "[58] Yuanyi Zhong, Bodi Yuan, Hong Wu, Zhiqiang Yuan, Jian Peng, and Yu-Xiong Wang. Pixel contrastive-consistent semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7273-7282, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.469, + 0.724 + ], + "angle": 0, + "content": "[59] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.727, + 0.469, + 0.768 + ], + "angle": 0, + "content": "[60] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A prototype view. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.77, + 0.469, + 0.839 + ], + "angle": 0, + "content": "[61] Yanning Zhou, Hang Xu, Wei Zhang, Bin Gao, and Pheng-Ann Heng. C3-semiseg: Contrastive semi-supervised segmentation via cross-set learning and dynamic classbalancing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7036-7045, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "942" + } + ] +] \ No newline at end of file diff --git a/2023/Space Engage_ Collaborative Space Supervision for Contrastive-Based Semi-Supervised Semantic Segmentation/5953e602-104c-4dc8-b226-b90796937563_origin.pdf b/2023/Space Engage_ Collaborative Space Supervision for Contrastive-Based Semi-Supervised Semantic Segmentation/5953e602-104c-4dc8-b226-b90796937563_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c76f615415d6bd24c5c41d91ae31779eaedfc486 --- /dev/null +++ b/2023/Space Engage_ Collaborative Space Supervision for Contrastive-Based Semi-Supervised Semantic Segmentation/5953e602-104c-4dc8-b226-b90796937563_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8344a28aea272033d6a8c4df4632d8c0638cb43b26a250b9094314b7ef984bd4 +size 17713806 diff --git a/2023/Space Engage_ Collaborative Space Supervision for Contrastive-Based Semi-Supervised Semantic Segmentation/full.md b/2023/Space Engage_ Collaborative Space Supervision for Contrastive-Based Semi-Supervised Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ccbb30064f2e7f29212dcefd37c7e6b4f7e51a93 --- /dev/null +++ b/2023/Space Engage_ Collaborative Space Supervision for Contrastive-Based Semi-Supervised Semantic Segmentation/full.md @@ -0,0 +1,421 @@ +# Space Engage: Collaborative Space Supervision for Contrastive-based Semi-Supervised Semantic Segmentation + +Changqi Wang $^{1*}$ , Haoyu Xie $^{1*}$ , Yuhui Yuan $^{2}$ , Chong Fu $^{1,4\dagger}$ , Xiangyu Yue $^{3}$ + +$^{1}$ School of Computer Science and Engineering, Northeastern University, Shenyang, China + $^{2}$ Microsoft Research Asia $^{3}$ The Chinese University of Hong Kong + +4 Key Laboratory of Intelligent Computing in Medical Image, Ministry of Education, NEU, China + +2101668@stu.neu.edu.cn, xiehaoyu@stumail.neu.edu.cn, yuyua@microsoft.com + +fuchong@mail.neu.edu.cn, xyyue@ie.cuhk.edu.hk + +# Abstract + +Semi-Supervised Semantic Segmentation (S4) aims to train a segmentation model with limited labeled images and a substantial volume of unlabeled images. To improve the robustness of representations, powerful methods introduce a pixel-wise contrastive learning approach in latent space (i.e., representation space) that aggregates the representations to their prototypes in a fully supervised manner. However, previous contrastive-based S4 methods merely rely on the supervision from the model's output (logits) in logit space during unlabeled training. In contrast, we utilize the outputs in both logit space and representation space to obtain supervision in a collaborative way. The supervision from two spaces plays two roles: 1) reduces the risk of over-fitting to incorrect semantic information in logits with the help of representations; 2) enhances the knowledge exchange between the two spaces. Furthermore, unlike previous approaches, we use the similarity between representations and prototypes as a new indicator to tilt training those under-performing representations and achieve a more efficient contrastive learning process. Results on two public benchmarks demonstrate the competitive performance of our method compared with state-of-the-art methods. + +# 1. Introduction + +Semantic segmentation is a fundamental task in computer vision, aiming to classify each pixel in an image. Significant progress [22, 4] has been made in training on high-quality labeled images using segmentation models composed of an encoder and a segmentation head. However, annotating images is expensive and time-consuming. Semi + +![](images/59bfb879373b542a433ac5a4aaa2599ca463101edce0c030da9bf2c5369d8d07.jpg) + +![](images/eeb11bfac1f0b1f95d9c58b267eceba51bb7932527e4ba689ef3e2719777a707.jpg) +Figure 1. We enhance the knowledge exchange between the logit and representation spaces. Orange and blue represent different classes. Top: Existing contrastive-based S4 methods overlook the semantic information in representation space. Bottom: Our method uses dual-space collaborative supervision. + +supervised Semantic Segmentation (S4) alleviates the thirst for annotation by leveraging unlabeled images to train segmentation models. + +Most existing works learn from unlabeled images via self-training [44, 37, 15] or consistency regularization [40, 38, 13] strategies, both of which retrain the model with its predictions on unlabeled images. Recently, great success has been achieved by introducing pixel-wise contrastive learning to semantic segmentation, which endows the model with a stronger features-extracting ability by accessing a more discriminative representation space. Specifically, these methods [45, 23] project each pixel to repre + +sensation space as a representation and regularize it in a fully supervised manner, i.e., aggregating the representations with the same class and separating them with different classes. In semi-supervised settings, due to limited labels, most methods [1, 32, 46] obtain supervision from the model's output logits in logit space during the unlabeled training process. However, recent contrastive-based semantic segmentation methods [1, 32, 46, 45] mainly focus on the learning process in logit space while only taking that in representation space as an auxiliary task. The unidirectional supervision makes training dominated by the predicted logits, leading to the neglect of information in the representation space. We argue that this kind of single-space supervision may incorrectly provide semantic guidance to representation learning and fails to facilitate knowledge exchange between the two spaces (see Sec. 5.1). + +In this work, we extend the single-space supervision to a dual-space supervision for contrastive-based S4 and propose Collaborative Space Supervision (CSS). Our key insight is to: i) utilize the semantic information in representations to obtain more reliable guidance during unlabeled training, and to enhance knowledge exchange between two spaces; ii) provide a more accurate reference for the model's performance on predicting each representation to tilt training those under-performing representations. To achieve objective i), we obtain dense semantic predictions by retrieving the nearest class prototype for each representation in the representation space and then engage with predictions from the logits for collaborative supervision of the model. For objective ii), we measure the similarity between the representations and prototypes and use the similarity after a normalization operation as the indicator for guiding the learning process in the representation space. Unlike previous works that utilize confidence as the indicator to involve representation learning, the similarity directly reflects the confusion level between representations and prototypes, resulting in more efficient representation learning. + +To summarize, our main contributions are three-fold: 1) The dual-space collaboration for contrastive-based S4, enhances the knowledge exchange between the logit and representation spaces. 2) Utilizing similarity to provide a more accurate reference for the model's performance in representation learning. 3) Extensive experiments on two S4 benchmarks demonstrate the effectiveness of our method. + +# 2. Related Works + +# 2.1. Semi-supervised Semantic Segmentation + +The aim of S4 is to train a segmentation model with the semi-supervised setting (i.e., a few labeled images and a large number of unlabeled images) to classify each pixel in an entire image. The critical issue of S4 is how to leverage unlabeled images to train the model. Some methods [25, + +27, 30, 35] based on GANs [16], adversarial training [36], and consistency regularization paradigm [38, 13, 40, 7, 56]. Meanwhile, self-training [28, 44, 49, 61, 53] is also a striking paradigm, which always generates pseudo-labels from model and retransmits the model with the combined supervision of human annotations and pseudo-labels. One essential issue of self-training is the accuracy of pseudo-labels. Some methods [29, 33, 14, 42, 52] try to polish pseudo-labels and provide reliable guidance. Some methods [21, 47, 24, 17] focus on the class-imbalance problems in the dataset and try to alleviate the negative effect from class-biased pseudolabels generated by the model pre-trained on imbalanced labeled images. We build our framework based on the self-training and additionally explore semantic information among different images. + +# 2.2. Pixel-wise Contrastive Learning + +Pixel-wise contrastive learning explores semantic relations not only in the individual image but also among different images. Different from instance-wise contrastive learning [19, 6, 3], pixel-wise contrastive learning [50, 55, 5, 58] project each pixel to the representation in representation space with the cooperation of encoder and representation head. Representations are then aggregated in their prototypes and are separated from each other in different classes. In semi-supervised settings, most methods [1, 32, 46, 48] use pseudo-labels based on logits to provide semantic information contrastive learning process during training on unlabeled images. Meanwhile, the confidence of logit is used as an indicator to involve the contrastive learning process, e.g., [32] uses the hard representations whose corresponding logit confidence is lower than a threshold to contrast for effective training. As opposed to the above methods, we use collaborative space supervision for contrastive learning on unlabeled images and use a new indicator to involve the contrastive learning progress. + +# 2.3. Prototype-based Learning + +Prototype-based learning has been widely studied in few-shot learning [41, 11, 31, 34] and unsupervised domain adaption [54, 26, 43, 39, 57]. Recently, it is restudied in semantic segmentation as known as a non-parametric prototype-based classifier [60]. Concretely, the classes in the dataset are presented by a set of non-learnable prototypes, and the dense semantic predictions are thus achieved by assigning the output features to its most similar prototype. Under semi-supervised settings, some methods [51] maintain the consistency between predictions from a linear predictor and a prototype-based predictor. The two predictors are followed by the encoder and project the features to logit space and representation space, respectively. In this work, we combine the semantic information in the logit and representation spaces to provide supervision in a collabora + +tive way during semi-supervised learning. + +# 3. Methodology + +In the S4 task, we have a small labeled set $\mathcal{D}_l = \{(\pmb{x}_i^l,\pmb{y}_i^l)\}_{i=1}^{N_l}$ and a large unlabeled set $\mathcal{D}_u = \{\pmb{x}_i^u\}_{i=1}^{N_u}$ , where $\pmb{x}_i^l,\pmb{x}_i^u\in \mathbb{R}^{H\times W\times 3}$ , $H,W$ denote the height, and the width, respectively. And ground truth $\pmb{y}_i^l\in \{0,1\}^{H\times W\times |C|}$ with the set of class $C$ . The goal is to boost model performance with $\mathcal{D}_u$ . The base model consists of an encoder $f(\cdot)$ and a segmentation head $g(\cdot)$ , which projects features to the logit space $\mathbb{R}^{H\times W\times |C|}$ . We adopt Self-Training (ST) and pixel-wise contrastive learning to our framework, as described in Sec. 3.1. The supervision for $\mathcal{D}_u$ is produced by the collaboration between the logit and representation spaces, as described in Sec. 3.3. + +# 3.1. ST and Pixel-wise Contrastive Learning + +The main idea of self-training is to pre-train a model on labeled images and use it to produce pseudo-labels as supervision for unlabeled images. A typical framework is the teacher-student framework [44], which consists of a student model and a teacher model. Both the student model and the teacher model are constructed by an encoder and a segmentation head. Parameters of the student model are optimized via Stochastic Gradient Decent (SGD) while parameters of the teacher model are updated by the Exponential Moving Average (EMA) of student model parameters. We denote the encoder and the segmentation head in the student model by $f(\cdot)$ and $g(\cdot)$ while denoting those in the teacher model by $f^{\prime}(\cdot)$ and $g^{\prime}(\cdot)$ . The pseudo-labels $\hat{\pmb{y}}_i^{u,lgt}$ are produced based on the teacher model's output logits $\hat{\pmb{p}}_i^u = g'(f'(x_i^u))$ in logit space, formulated as: + +$$ +\hat {\boldsymbol {y}} _ {i} ^ {u, l g t} = \mathbf {1} _ {c} (\underset {c} {\arg \max } \{\hat {\boldsymbol {p}} _ {i, c} ^ {u} \} _ {c \in C}), \tag {1} +$$ + +where $\mathbf{1}_c(\cdot)$ denotes the one-hot encoding of class $c$ . + +In order to enhance the ability of the model itself to extract features, recent works [32, 46, 1] additionally employ pixel-wise contrastive learning and introduce a representation head to both teacher and student models. We denote the representation head in the student model as $h(\cdot)$ and that in the teacher model as $h^{\prime}(\cdot)$ . The pixel $x_{i}$ of the class $c$ are projected as representations $z_{ci}$ in representation space by the cooperating of $f(\cdot)$ and $h(\cdot)$ , i.e., $z_{ci} = h(f(x_i))$ . And the representation $z_{ci}$ is then aggregated to its class centroid (prototype) while separated from representations in different classes $z_{\tilde{c} i}$ (negatives). The semantic guidance for contrastive learning is from the combination of ground truth $y_{i}^{l}$ and pseudo-labels $\hat{y}_i^{u,lgt}$ in logit space. Moreover, in order to emphasize the reliable and crucial aspects during unlabeled and contrastive learning, a sampling strategy is adopted to select valid pixels $x_{i}$ according to their corre + +sponding confidence, i.e. the student model's output logits $\pmb{p}_i$ after a Softmax operation. + +Discussion. In recent works [46, 1, 32], the supervision of unlabeled images is derived solely from the logit space. This overlooks the potential benefits of the supervision from the representation space, leading to two potential limitations: 1) the pseudo-labels $\hat{y}_i^{u,lgt}$ obtained from the logit space may contain noise and miss the opportunity to be corrected by semantic information from the representation space; 2) since the confidence from logit space is used as the indicator $\hat{j}_i$ for the sampling strategy, learning in the representation space may not be critical enough due to the different confusing parts between the two spaces. + +To mitigate these limitations, we produce pseudo-labels from the representation space and combine them with pseudo-labels from the logit space to provide higher-quality supervision during unlabeled training. Meanwhile, we obtain a new indicator from the representation space for the more effective sampling strategy. + +# 3.2. Supervision from Representation Space + +In this section, we detail the approach to obtain the pseudo-labels from the representation space. Meanwhile, we simultaneously access a new indicator for the sampling strategy in representation spaces, which provides a critical reference in the contrastive learning process. + +Specifically, we first build a set of prototypes for each class and obtain the pseudo-labels by retrieving the nearest prototype for each representation. We calculate the centroid of all representations in the current class $c$ as the prototype $\rho_{c}$ , which is formulated as: + +$$ +\rho_ {c} = \frac {1}{N _ {c}} \sum_ {i} ^ {N _ {c}} z _ {i} ^ {\prime}, \tag {2} +$$ + +where $N_{c}$ is the total number of representations of current class $c$ and $\pmb{z}_i^{\prime}$ is the representation projected by the cooperation of the $f^{\prime}(\cdot)$ and $h^\prime (\cdot)$ . Meanwhile, to include more representation information, we update the prototype along the sequential iterations with EMA as follows: + +$$ +\hat {\rho} _ {c} (t) = \alpha \hat {\rho} _ {c} (t - 1) + (1 - \alpha) \rho_ {c} (t), \tag {3} +$$ + +where $\hat{\rho}_c(t),\hat{\rho}_c(t - 1)$ mean the current $t^{th}$ prototype and last $(t - 1)^{th}$ prototype in iterations, $\rho_{c}(t)$ means the prototype calculated by Eq.2 in current iteration, and $\alpha$ is a hyper-parameter that controls the updating speed. Thus, the pseudo-label from the representation space is achieved by: + +$$ +\hat {\boldsymbol {y}} _ {i} ^ {u, r e p} = \mathbf {1} _ {\hat {c}} (\hat {c}), w i t h \hat {c} = \underset {c} {\arg \max } \left\{\operatorname {s i m} \left(\boldsymbol {z} _ {i} ^ {\prime}, \hat {\boldsymbol {\rho}} _ {c} (t)\right) \right\} _ {c \in C}, \tag {4} +$$ + +where $sim(\cdot)$ is defined as the cosine similarity. + +As for the indicator for the sampling strategy in the representation space, we use the Softmax function on the similarity among the representation and all prototypes, which is + +![](images/9460f85ac9f86af3b5a8ad3d7cde59fdaa955bb3d1ce27cd70bd3f30f15d9484.jpg) +Figure 2. Overview of our framework. Our training pipeline consists of learning in two spaces: logit space and representation space. The pseudo-labels $\hat{\pmb{y}}_i^u$ during unlabeled training are produced by the collaboration of two spaces with mix pseudo-labeling strategy (1) or cross pseudo-labeling strategy (2). The indicator for representation learning is produced by similarity $(s_1, s_2, \text{and} s_3)$ . + +followed as: + +$$ +\hat {j} _ {i} ^ {u, r e p} = \frac {e ^ {s i m \left(\boldsymbol {z} _ {\boldsymbol {c} i} , \hat {\rho} _ {\boldsymbol {c}} (t)\right) / \tau}}{e ^ {s i m \left(\boldsymbol {z} _ {\boldsymbol {c} i} , \hat {\rho} _ {\boldsymbol {c}} (t)\right) / \tau} + \sum_ {\tilde {c} \in \tilde {C}} e ^ {s i m \left(\boldsymbol {z} _ {\boldsymbol {c} i} , \hat {\rho} _ {\tilde {c}} (t)\right) / \tau}}, \tag {5} +$$ + +where $\hat{\rho}_{\tilde{c}}(t)$ means the prototype with different classes from $z_{ci}$ and $\tau$ is a hyper-parameter. Different from using confidence from logit space as the indicator to involve representation learning [32, 46], the Softmax similarity directly helps the model to discover the confusion between representations and their prototypes and focus on them during the subsequent training. + +# 3.3. Collaboration Between Two Spaces + +With the pseudo-labels in two spaces, we propose two pseudo-labeling strategies to strengthen the collaboration between two spaces and obtain more reliable pseudo-labels. + +- Mix pseudo-labeling. To mitigate the negative effects of inherent noise from both two spaces during pseudolabeling, we adopt the mix pseudo-labeling strategy that only considers the mutually agreeable pseudolabels between the two spaces. Specifically, we define the set of final pseudo-labels as $\hat{Y}^u = \hat{Y}^{u,lgt} \cap \hat{Y}^{u,rep}$ , where $\hat{\mathbf{y}}_i^{u,lgt} \in \hat{Y}^{u,lgt}$ and $\hat{\mathbf{y}}_i^{u,rep} \in \hat{Y}^{u,rep}$ . +- Cross pseudo-labeling. Inspired by recent researches [7, 38] that maintain consistency among the predictions of the same image across different models or decoders in different views, we propose a cross pseudolabeling strategy that leverages pseudo-labels from one space to supervise the other. Specifically, we use pseudo-labels $\hat{\pmb{y}}_i^{u,rep}$ to supervise the logit space, and vice versa. + +The strengths of using pseudo-labels from two spaces in the collaborative way are twofold: 1) obtaining more reliable supervision during unlabeled training, and 2) enabling the strengths of learning in different spaces to complement each other. Since the learning in different feature spaces concentrates on different parts of features, i.e., the logit space mainly focuses on the most discriminative part of features while the representation space treats all parts of features equally, the performance of pseudo-labels from two spaces varies across different classes or regions of images. Therefore, our collaborative pseudo-labeling strategies exchange knowledge between two spaces and provide higher-quality supervision during unlabeled training. The experimental proof is in Sec. 5.1. + +As for indicators, we use confidence as the indicator $\hat{j}_i^{u,lgt}$ for learning logit space and Softmax similarity as the indicator $\hat{j}_i^{u,rep}$ for learning representation space. We argue that the confusing parts of learning in both two spaces are varied due to the different parts of features being concentrated in each space. Therefore, this strategy allows the learning in different spaces to focus on their own confusing parts, which can be more effective than using a single indicator when mining confusing parts for both spaces in the training process. The experimental proof is in Sec. 5.2. + +# 3.4. Training Objective + +With the indicators $\hat{j}_i^{u,lgt}$ and $\hat{j}_i^{u,rep}$ , we adopt some threshold sampling strategies. In logit space, we set a threshold $\delta_u$ during unlabeled learning and logits $\hat{p}_i^u$ whose indicator $\hat{j}_i^{u,lgt}$ is higher than $\delta_u$ will be regarded as the valid logits in logit space. In representation space, our sample strategy can be divided into three parts: 1) Valid Sam + +pling Strategy. Similar to the sampling strategy in logit space, a threshold $\delta_w$ is used to sample representations whose indicator $\hat{j}_i^{u,rep}$ is higher than $\delta_w$ . 2) Hard Sampling Strategy. We adopt the hard sampling strategy for tilting to train those confusing representations. Specifically, we set a threshold $\delta_s$ to sample representations whose indicator $\hat{j}_i^{u,rep}$ is lower than $\delta_s$ . 3) Negative Sampling Strategy. We sample negatives according to the similarity between the prototype of current class $c$ and other prototypes. Concretely, the negatives are more likely to be sampled if its prototype is more similar to the prototype of the current class. + +Cooperated with the ground truth $\pmb{y}_i^l$ , pseudo-labels $\pmb{y}_i^u$ produced from two spaces in a collaborative way, and different sampling strategies in two spaces, the total learning object is composed with a supervised loss $\mathcal{L}_s$ , an unsupervised loss $\mathcal{L}_u$ , and a contrastive loss $\mathcal{L}_c$ as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {s} + \mathcal {L} _ {u} + \lambda_ {c} \mathcal {L} _ {c}, \tag {6} +$$ + +where $\lambda_{c}$ is used to tune the contribution between logit space and representation space. Specifically, $\mathcal{L}_s$ and $\mathcal{L}_u$ are constructed by the Cross-Entropy (CE) $\ell_{ce}$ and can be formulated as: + +$$ +\mathcal {L} _ {s} = \frac {1}{| \mathcal {B} _ {l} |} \sum_ {\boldsymbol {x} _ {i} ^ {l} \in \mathcal {B} _ {l}} \ell_ {c e} \left(\boldsymbol {p} _ {i} ^ {l}, \boldsymbol {y} _ {i} ^ {l}\right), \tag {7} +$$ + +$$ +\mathcal {L} _ {u} = \frac {1}{| \hat {\mathcal {B}} _ {u} |} \sum_ {\boldsymbol {x} _ {i} ^ {u} \in \hat {\mathcal {B}} _ {u}} \ell_ {c e} \left(\boldsymbol {p} _ {i} ^ {u}, \hat {\boldsymbol {y}} _ {i} ^ {u}\right), \tag {8} +$$ + +where $\mathcal{B}_l$ denotes the labeled images in a mini-batch and $\hat{\mathcal{B}}_u$ is the subsets that sampled from unlabeled images in a mini-batch according to the sampling strategy. Meanwhile, the contrastive loss $\mathcal{L}_c$ is formulated as: + +$$ +\mathcal {L} _ {c} = - \frac {1}{| C | \times | \hat {\mathcal {Z}} _ {c} |} \sum_ {c \in C} \sum_ {\boldsymbol {z} _ {c i} \in \hat {\mathcal {Z}} _ {c}} \tag {9} +$$ + +$$ +l o g [ \frac {e ^ {s (\pmb {z} _ {c i} , \hat {\pmb {\rho}} _ {c}) / \tau}}{e ^ {s (\pmb {z} _ {c i} , \hat {\pmb {\rho}} _ {c} (t))) / \tau} + \sum_ {\tilde {e} \in \tilde {C}} \sum_ {\pmb {z} _ {\tilde {e} i} \in \hat {\pmb {Z}} _ {\tilde {c}}} e ^ {s (\pmb {z} _ {c i} , \pmb {z} _ {\tilde {c} i}) / \tau} ], +$$ + +where $\hat{\mathcal{Z}}_c$ is the subset sampled from the set of the representations which belong to class $c$ according to the sampling strategy, $\hat{\mathcal{Z}}_{\tilde{c}}$ is the subset sampled from the set of the representations which bot belong to class $c$ , $\tilde{C}$ denotes the subset of $C$ with class $c$ removed, and the supervision information comes from the final pseudo-labels $\hat{\pmb{y}}_i^u$ after the pseudo-labeling strategies. + +The whole framework is shown in Fig. 2. All the pseudocode of producing pseudo-labels and their indicators from the logit and representation spaces is shown in Algorithm 1, and the pseudo-code of the mix pseudo-labeling strategy is shown in Algorithm 2. The pseudo-code is PyTorch-like. + +Algorithm 1 Pseudo-code of producing pseudo-labels and indicators + +Network: Teacher's encoder $f'$ , segmentation head $g'$ , representation head $h'$ . + +Input: Unlabeled mini-batch $B_{u}$ consists of $X^{u}$ , the set of prototypes $\{\hat{\rho}_c(t)\}_{c\in C}$ . + +1: for $X_{u} \in B_{u}$ do +2: # Predict logits from unlabeled images. +3: $\hat{P}^{u,lgt}\gets g'(f'(X^u))$ +4: # Produce labels and indicators based on logits with Eq. 1. +5: $\hat{Y}^{u,lgt},\hat{J}^{u,lgt}\gets lgt\text{-} pseudo(\tilde{P}^{u,lgt})$ +6: # Predict representations from unlabeled images. +7: $\hat{R}^{u,rep}\gets h'(f'(X^u))$ +8: # Produce labels and indicators based on representations with Eq. 4 and Eq. 5. +9: $\hat{Y}^{u,rep},\hat{J}^{u,rep}\gets rep\text{-} pseudo(\hat{R}^{u,rep},\{\hat{\rho}_c(t)\}_{c\in C})$ +10: end for + +Algorithm 2 Pseudo-code of mix pseudo-labeling strategy + +Input: Pseudo-labels from logit space $\hat{Y}^{u,lgt}$ and from representation space $\hat{Y}^{u,rep}$ . Indicators $\hat{J}^{u,lgt}$ and $\hat{J}^{u,rep}$ . + +Notation: Threshold $\delta_{u}$ for sampling strategy in $\mathcal{L}_u$ , weak threshold $\delta_w$ . + +1: # Sampling valid pseudo-labels according to the sampling strategy in logit space. +2: $\hat{Y}_{val}^{u,lgt}\gets$ sample_lgt( $\hat{Y}^{u,lgt},\hat{J}^{u,lgt},\delta_u)$ +3: # Sampling valid pseudo-labels according to the valid sampling strategy in logit space. +4: $\hat{Y}_{val}^{u,rep}\gets$ sample_rep( $\hat{Y}^{u,rep},\hat{J}^{u,rep},\delta_w)$ +5: # Produce the mask for mix pseudo-labels. +6: $mask_{y} \gets \hat{Y}^{u,rep}.$ eq( $\hat{Y}^{u,lgt}$ ) +7: # Produce mix pseudo-labels according to the mask. +8: $\hat{Y}^u\gets$ mask_pseudo $(\hat{Y}^{u,lgt},\text{mask}_y)$ + +# 3.5. Discussions + +We discuss the relations between our framework and other most-related S4 frameworks. + +Compare with PCR [51]. The key insight of PCR and our method lies in the promotion of consistency between the output in the logit and representation spaces. However, PCR exclusively relies on pseudo-labels derived from logit space as the guidance in the logit and representation spaces during the unsupervised training process, thereby neglecting the exploitation of the extensive semantic information inherent in representations and prototypes. In contrast, our method adopts a collaborative way by combining pseudo-labels obtained from both the logit and representation spaces, which improves the quality of pseudo-labels and enhances the knowledge exchange between the two spaces. In addition, since these two approaches are orthogonal, one can employ + +our pseudo-labeling and indicator strategies within the PCR framework, or alternatively incorporate multiple prototypes and consistency loss into our approach. Such combinations may yield further improvements in performance, however, also increase the computational complexity. + +Compare with CPS [7]. CPS contains two segmentation models initialized differently. It leverages pseudo-labels generated by one model to supervise the other, thereby maintaining the consistency between the output of the two models in the logit space. In a similar vein to CPS, our cross pseudo-labeling strategy in Sec. 3.3 is also motivated by the objective of maintaining consistency during unsupervised training. However, our method is distinct from CPS in that we focus on preserving consistency between the logit space and the representation space. This distinction confers a distinct advantage in terms of memory efficiency, as we solely introduce a MLP as the representation head, rather than introducing an additional segmentation model as in CPS. + +Compare with other contrastive based S4 works [32, 1, 46, 23]. We adhere to prior works to build our prototype, describe in Eq. 2. Different from them, we update our prototypes iteratively, similar to [60]. Our sampling strategies are based on the threshold, similar to [32, 23, 45]. However, it is worth noting that our indicators which serve as the basis for comparison with the threshold, are derived from both the logit and representation spaces. This stands in contrast to prior approaches where indicators solely originate from the logit space. The efficacy of this modification is substantiated and extensively discussed in Sec. 5.2. + +# 4. Experiments + +# 4.1. Setup + +Datasets. We conduct experiments on PASCAL VOC 2012 dataset [12], Cityscapes dataset [9], ADE20K dataset [59], and COCO-Stuff 10K dataset [2] to validate the effectiveness of our proposed method. The original PASCAL VOC 2012 dataset contains 1,464 labeled images in train set and 1,449 images for validation in val set. Following [7, 37], we additionally introduce 9,118 images from SBD [18] as training images. Since the labels in SBD are coarsely annotated, following [46], we use both classic VOC train set (1,464 candidate labeled images) and blender VOC train set (10,582 candidate labeled images). Cityscapes dataset is a dataset for urban scene understanding, which contains 2,975 images in train set and 500 images in val set. ADE20K has 20,210 and 2,000 images in train and val set, with 150 classes in total. COCO-Stuff 10K has 9,000 images in train set and 1,000 images in val set, with 181 classes. The approach of preprocessing labels is followed by MMSegmentation [8]. + +Network structure. We use Deeplabv3+ [4] with ResNet-101 [20] pre-trained on ImageNet [10] as our network struc + +ture. The segmentation and representation head are composed of Conv-BN-ReLU-Conv. + +Implementation details. For training on PASCAL VOC 2012 dataset and COCO-Stuff 10K dataset, we set the learning rate as 0.0064, weight decay as 0.0005, crop size as 512 $\times$ 512, batch size as 16, and a total of 40,000 iterations. For training on the Cityscapes dataset, we set the learning rate as 0.0038, weight decay as 0.0005, crop size as $768 \times$ 768, batch size as 8, and a total of 80,000 iterations. For training on ADE20K dataset, we set the learning rate as 0.0064, weight decay as 0.0005, crop size as $512 \times 512$ , batch size as 16, and a total of 80,000 iterations. We use the poly scheduling to decay the learning rate during the training process: $lr = lr_{base} \times (1 - \frac{epoch}{total\_epoch})^{0.9}$ . We use the mean of Intersection over Union (mIoU) as the metric in evaluation. We use the sliding window strategy to evaluate the performance of our method on the Cityscapes dataset, following [7]. In addition, when adopting our cross-labeling strategy, due to the requirement of a set of high-quality prototypes when classifying each representation, we first solely use the supervision from logit space for 20 epochs to initialize the prototypes. + +# 4.2. Comparison with Existing Methods + +In this subsection, we first reproduce three baselines: MT [44], CutMix [15], and a same contrastive-based framework with us but with only logit space pseudo-labels and indicators (Baseline) on classic VOC train set. Meanwhile, we make the comparison of our method with mix (CSS (mix)) and cross (CSS (crs.)) pseudo-labeling strategy on blender VOC train set and Cityscapes train set with following recent SOTA S4 methods: CCT [38], CPS [7], U²PL [46], ST++ [52], PRCL [48], PCR [51], and PSMT [33]. Since the data split will dramatically affect the performance in S4, i.e., choosing labeled images plays an important role in the final results, we conduct experiments with three different data splits and report the mean value and standard deviation (blue color numbers). Since the mix pseudo-labeling strategy has better performance, we only use CSS (mix) when compared with SOTAs. Meanwhile, we use ResNet-101 with deep stem blocks as our network structure when compared with SOTAs. Since there is no uniform data split, we use the data splits in U²PL [46]. We use the OHEM loss when training Cityscapes. + +Results on PASCAL VOC 2012. Tab. 1 shows the comparison with our baselines on Classic PASCAL VOC 2012 set. Our method consistently outperforms baselines with an acceptable standard deviation on all label rates. Tab. 2 shows the comparison with the SOTAs on PASCAL VOC 2012. + +Results on Cityscapes. Tab. 3 shows the performance of our method on Cityscapes. + +Results on ADE20K and COCO-Stuff 10K. Tab. 4 shows the results of our method on ADE20K and COCO-Stuff. + +Table 1. Results on classic VOC train set with four different label rate. Labeled data splits are from the original VOC train set. All approaches are reproduced with three runs and three different data splits. + +
Pascal VOC 2012 (Classic)
Method92183366732
Sup.51.57±3.5854.69±2.4464.86±1.0470.77±0.76
MT58.92±2.9961.63±1.7666.79±0.5371.58±0.51
CutMix65.82±3.6067.91±1.4772.53±0.5074.08±0.49
Baseline66.91±4.2170.32±1.8673.97±1.0476.49±0.58
CSS(crs.)67.03±4.5871.41±1.6774.47±1.0877.08±0.37
CSS(mix)68.09±4.8971.93±1.8874.91±1.1277.57±0.73
+ +Table 2. Results on blender VOC train set. All the results from the recent papers [46, 52, 13, 29, 33, 51]. Labeled data is from the augmented VOC train set and the data splits are from [46, 33]. + +
Pascal VOC 2012 (Blender)
Method662132326465291
CCT [38]71.8673.6876.5177.40
CPS [7]74.4876.4477.6878.64
U2PL [46]77.2179.0179.3080.50
ST++ [52]74.7077.9077.90-
PRCL [48]76.9678.1679.0279.59
PCR [51]78.6080.7180.7880.91
PSMT [33]75.5078.2078.7279.76
CSS (mix)78.7379.5480.8281.06
+ +Table 3. Results on Cityscapes. The model is trained on the Cityscapes train set, which consists of 2,975 samples in total, and tested on Cityscapes val set. And all the results from the recent papers [46, 51, 33]. + +
Cityscapes
Method1863727441488
CCT [38]69.3274.1275.9978.10
CPS [7]69.7874.3174.5876.82
U2PL [46]70.3074.3776.4779.05
PCR [51]73.4176.3178.4079.11
PSMT [33]-76.8977.6079.09
CSS (mix)74.0276.9377.9479.62
+ +# 5. Ablative Study + +The main contribution of our work lies in 1) collaborative pseudo-labeling strategies and 2) a new indicator for representation learning. To further prove the effectiveness of our proposed method, we conduct ablative studies on these two points. We choose Deeplabv3+ with ResNet-101 pre-trained on ImageNet as our backbone and leverage 92 labeled images and 183 labeled images in PASCAL VOC 2012. The other settings are the same as those in Sec. 4. + +# 5.1. Effectiveness of Collaborative pseudo-labeling + +Quality of pseudo-labels. To illustrate the superiority of using pseudo-labels from two spaces in a collaborative way as supervision, we conduct experiments to show the quality of pseudo-labels obtained 1) from logit space (lgt.), 2) from + +![](images/9ec7b34a9c8edda15bbb118fc8aad3b89f26be65e4f81c35fb50a8784b65bfc6.jpg) + +![](images/885b6ec9ed1b4342c72e6b0a35d62753823bf6a702037312c0b75bcb4e800fde.jpg) +(b) GT + +![](images/f53cf14542852989825603154b7a11bfa60e5a33b958812402bd582f13415e77.jpg) + +![](images/3bd1a658daae03d39c76792d3e531523ef9aba32cfb9563c71d41f96e2723fcd.jpg) + +![](images/2101bf8846b2951f47194ea6ee2e90ba11e34bf91205d42f088d7fb709551c9e.jpg) +(d) lgt. label +Figure 3. Differences between pseudo-labels in different spaces. + +![](images/3311a85ac4b5d87c7ad39fbef4c91b440d5bb9a920abb2bace77c69987926f12.jpg) + +![](images/de9a6e7cf32de80262c4629e83a94e3f231d7f7965cfc96a2b7ec931a5a5022f.jpg) +(e) rep. mask + +![](images/7fbcb16361f71a321b1f1851c62aef74216a3d53756d95c792d51d6eff4f5f3d.jpg) + +![](images/0c9a2c8b867d5390141cce1649bf2e365f208bc51a70272403e5b3aa16c997ef.jpg) +(f) rep. label + +representation space (rep.), and 3) from the mix pseudolabeling strategy (mix). The pseudo-labels are sampled with corresponding sampling strategies in Sec. 3.4. Tab. 5 illustrates the IoU of pseudo-labels for each class on PASCAL VOC 2012 with 92 labeled images. The results clearly indicate that employing pseudo-labels from the representation space enhances the accuracy of the final pseudo-labels in most classes. This improvement is particularly evident in classes that are originally under-performing, such as the IoU improvement of $11.42\%$ for the chair class and $11.58\%$ for the sofa class. + +Meanwhile, we also visualize the pseudo-labels obtained from logit space (lgt.) and representation space (rep.) in Fig. 3. Fig. 3 (c) and (e) show the masks for pseudolabels produced by the sampling strategies. In particular, the white color represents the valid pixels used during unlabeled learning, while the black color indicates the discarded pixels. Fig. 3 (d) and (f) are the pseudo-labels we obtained from two spaces. The figure clearly illustrates the differences between pseudo-labels produced by different spaces. For example, the parts of the instance edge in pseudo-labels are usually discarded since they are challenging for learning in logit space. However, pseudo-labels from representation space will easily tackle this problem (first row). The pseudo-labels from representation space are more inaccurate in some complex scenes, which be resolved by combining pseudo-labels from logit space (second row). + +We mainly attribute the differences in pseudo-label to the differing concentrations of learning in the two spaces. Specifically, learning in the logit space primarily emphasizes the most discriminative part of features, whereas that in the representation space treats each part of features equally. As a result, learning in the logit space may overlook minor feature differences, leading to sub-optimal performance in predicting instance edges and distinguishing between similar classes (e.g., chair and sofa). Conversely, learning in the representation space produces balanced performance across all image parts and classes. However, this can lead to erroneous predictions for classes with high intra-class variance (e.g., background). By leveraging pseudo-labels from the two spaces, we capitalize on the strengths of the learning in each space and enhance the knowledge exchange between the two spaces. + +Table 4. Results on ADE20K and COCO-Stuff 10K with four different label rates and single data splits. + +
ADE20KCOCO-Stuff 10K
Method1/161/81/41/21/161/81/41/2
Cutmix28.9131.8934.6239.1523.0926.1429.8830.91
Baseline31.1133.2436.4639.9126.5727.9128.9731.08
CSS (crs.)32.0133.9237.8439.9727.0628.8530.5931.73
CSS (mix)32.5534.2137.0140.8527.4629.5129.9831.91
+ +Table 5. The quality of pseudo-labels from different pseudo-labeling strategies. The pseudo-labels are sampled by sampling strategies. + +
sourcebackaero.bicy.birdboatbott.buscarcatchaircow
lgt.96.7895.1477.4793.3881.3787.5496.7695.4894.474.0992.15
rep.90.7196.5061.4275.7553.3054.6584.4680.5591.1128.0388.16
mix96.66↓0.1298.12↑2.9882.76↑5.2994.47↑1.0985.12↑3.7589.94↑2.4097.03↑0.2795.98↑0.5094.65↑0.1819.01↑14.9294.81↑2.66
sourcetabeldoghorsemotorpers.plantsheepsofatraintvmIoU
lgt.58.3494.1293.0191.3793.6850.3391.1018.1686.6564.8978.87
rep.52.2386.9773.2288.7091.7556.1366.7835.2588.5162.6171.57
mix64.35↑6.0194.33↑0.2193.65↑0.6491.50↑0.1393.01↓0.6755.64↑5.3191.56↑0.4629.74↑11.5893.54↑6.8969.48↑4.5982.17↑3.33
+ +Table 6. Results on pseudo-labels from different sources on two different label rates. + +
source92 labels183 labels
logit space67.1170.32
representation space64.2067.52
mix pseudo-labeling68.41↑1.3072.74↑2.42
cross pseudo-labeling67.85↑0.7471.98↑1.66
+ +Results of different strategies To investigate the involvement of different pseudo-labeling strategies, we conduct the experiments as follows: 1) Using pseudo-labels from logit space. 2) Using pseudo-labels from representation space. 3) Using mix pseudo-labeling strategy. 4) Using cross pseudolabeling strategy. Tab. 6 shows the effectiveness of our proposed strategy in two different label rates. The results show that the performances of experiments with collaborative pseudo-labeling strategies are better than the ones whose pseudo-labels come from a single space with two different label rates, which proves the effectiveness of our proposed collaboration between the two spaces. It is worth noting that even though the quality of pseudo-labels from representation space is lower than that from logit space, the performance of the model is also boosted by using the cross pseudo-labeling strategy to maintain consistency between the predictions in two spaces. In addition, our method with the mix pseudo-labeling strategy outperforms that with the cross pseudo-labeling strategy. + +# 5.2. Effectiveness of the Indicator + +Limitation of merely using confidence. To explain the limitation of merely using confidence for learning in the logit and representation spaces, we conduct experiments to show the relations between confidence and similarity. The similarity is the cosine similarity between representations and prototypes, which directly shows the confusion level between the representation and the prototype of its class. + +![](images/b808f15724ab1f45e36ba846846ae5ac6fdbccf2c1e21314c20fd4b0b040d92f.jpg) +(a) +Figure 4. Relations between similarity and confidence. + +![](images/d4e6d66393eb772fd8063bec986ad111d15dd70005f4e8f98ccee0b69d3f013b.jpg) +(b) + +Fig. 4 (a) shows the comparison between the confidence of each prediction and the corresponding similarity. We use the class person for demonstrating in our experiments. It shows clearly that even though fixing the confidence into a small range (from 0.8 to 0.81 in our settings), the similarity varies. Meanwhile, in Fig. 4 (b), different color bars stand for different intervals of confidence, and the lines denote the mean similarity between the prototype and each representation whose corresponding confidence is in the current interval. Fig. 4 (b) illustrates that the mean similarity of the class fluctuates when its interval of confidence rises. Both two figures imply that confidence is not able to represent the confusion level between representations and prototypes since there are no direct and close relations between confidence and similarity. + +Fig. 5 visualizes the similarity and confidence of an image in both logit (lgt.) and representation (rep.) spaces, indicating the varying levels of confusion in the same region when learning in different spaces. + +We also attribute it to the different concentrations of learning two spaces, i.e., the confusing region in one space can be more readily addressed in the other space. + +Thus, it is inappropriate to choose confidence as the indicator to involve representation learning, e.g., sampling more hard samples with threshold and indicator. In contrast, our indicator directly employs similarity between the represen + +![](images/afbb859f0b9726ddd8dd67ba3373e08a7863d8d794adf298f134bddb32e33570.jpg) +high +(a) Image +Figure 5. Visualization of the confusing part in different spaces. + +![](images/7c004d2fc7084f03c4164fb131f750be95def112c593c4cadd15172f5de04802.jpg) +(b) GT + +![](images/6e11740b2e78d6466f069457b1c094909b6eee02dbd9ecf3470bfe3b2de4de31.jpg) +(c) lgt. space + +![](images/692662af87d13f12ef88a56656a8888fa45fb2d2e7820fe9a4ac44e5f98eb755.jpg) +(c) rep. space + +tation and prototype of its class, which directly reflects the confusion level in representation learning. It is more accurate to use similarity as the indicator to sample hard and critical samples in representation learning. + +Results of different indicators. Tab. 7 shows the impact of using different indicators. We conduct experiments on two different label rates (92 and 183) and three different indicators for pseudo-labels: only confidence (conf.), only similarity (smlr.), and confidence for learning logit space while similarity for learning representation space (mix). We use two different approaches to obtain pseudo-labels: from logit space only (seg label) and from mix pseudo-labeling strategy (mix label). It is clear that using both confidence and similarity to involve the learning in their own spaces obtains the best performance. + +Table 7. Results on indicators from different spaces on two different label rates. + +
92 labels183 labels
sourceseg labelmix labelseg labelmix label
conf.67.1168.3370.3271.92
smlr.66.1666.8968.9069.25
mix67.80↑0.6968.41↑1.3071.50↑1.1872.74↑2.42
+ +# 5.3. Ablation study of Components + +In this section, we conduct experiments to introduce our components in CSS step by step, with results shown in Tab. 8. Our baseline is the conventional contrastive-based S4, achieving mIoU of $67.11\%$ on 92 labels and $70.32\%$ on 183 labels. Mix and cross means the pseudo-labels are from the mix pseudo-labeling strategy and cross pseudo-labeling strategy while the indicator is still the confidence in two spaces. Ind means we use the different indicators in different spaces while the pseudo-labels are from logit space. The last two rows represent our proposed two pseudo-labeling strategies with indicators from two spaces. + +# 5.4. Qualitative Results + +Fig. 6 shows the qualitative results of different methods on PASCAL VOC 2012 with 92 labeled images. Baseline means the conventional contrastive-based method. Compared with the original self-training methods (CutMix), + +Table 8. Ablation study on different components of our CSS. + +
component92 labels183 labels
baseline67.1170.32
mix68.33↑1.2271.92↑1.60
cross67.21↑0.1070.87↑0.55
ind67.80↑0.6971.50↑1.18
mix + ind68.41↑1.3072.74↑2.42
cross + ind67.85↑0.7471.98↑1.66
+ +thanks to introducing pixel-wise contrastive learning, the baseline, and our method perform better in some ambiguous regions. Furthermore, benefiting from the supervision of two spaces and different indicators in different spaces, our method outperforms the baseline. + +![](images/0087841697414687131256fde037c71be6289bf332eac50c0f368691b3531be0.jpg) +images. Yellow boxes highlight the main differences. + +# 6. Conclusion + +In this paper, we propose two collaborative pseudolabeling strategies to take full use of the semantic information in the representation space and enhance the knowledge exchange between the logit and representation spaces. Moreover, we employ a new indicator for the learning process in the representation space. Extensive experiments demonstrate that our pseudo-labeling strategies obtain more reliable supervision during unlabeled training and our indicator helps the model to concentrate on more critical parts during representation learning. + +Future work: In this paper, we employ pseudo-labeling strategies to utilize the semantic information in both logit and representation spaces. In the future, we will investigate more powerful strategies to enhance the knowledge exchange between two spaces. + +# References + +[1] Inigo Alonso, Alberto Sabater, David Ferstl, Luis Montesano, and Ana C. Murillo. Semi-supervised semantic segmentation with pixel-level contrastive learning from a class-wise memory bank. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8219-8228, October 2021. +[2] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. +[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. +[4] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. +[5] Mu Chen, Zhedong Zheng, Yi Yang, and Tat-Seng Chua. Pipa: Pixel-and patch-wise self-supervised learning for domain adaptive semantic segmentation. arXiv preprint arXiv:2211.07609, 2022. +[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. +[7] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021. +[8] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020. +[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. +[11] Nanqing Dong and Eric P Xing. Few-shot semantic segmentation with prototype learning. In BMVC, volume 3, 2018. +[12] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. +[13] Jiashuo Fan, Bin Gao, Huan Jin, and Lihui Jiang. Ucc: Uncertainty guided cross-head co-training for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition (CVPR), pages 9947-9956, June 2022. +[14] Zhengyang Feng, Qianyu Zhou, Qiqi Gu, Xin Tan, Guangliang Cheng, Xuequan Lu, Jianping Shi, and Lizhuang Ma. Dmt: Dynamic mutual training for semi-supervised learning. Pattern Recognition, page 108777, 2022. +[15] Geoff French, Timo Aila, Samuli Laine, Michal Mackiewicz, and Graham Finlayson. Semi-supervised semantic segmentation needs strong, high-dimensional perturbations. In In 31th British Machine Vision Conference, 2020. +[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. +[17] Dayan Guan, Jiaxing Huang, Aoran Xiao, and Shijian Lu. Unbiased subclass regularization for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9968-9978, 2022. +[18] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In 2011 international conference on computer vision, pages 991-998. IEEE, 2011. +[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[21] Ruifei He, Jihan Yang, and Xiaojuan Qi. Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6930-6940, October 2021. +[22] Judy Hoffman, Dequan Wang, Fisher Yu, and Trevor Darrell. Fcts in the wild: Pixel-level adversarial and constraint-based adaptation. arXiv preprint arXiv:1612.02649, 2016. +[23] Hanzhe Hu, Jinshi Cui, and Liwei Wang. Region-aware contrastive learning for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16291-16301, 2021. +[24] Hanzhe Hu, Fangyun Wei, Han Hu, Qiwei Ye, Jinshi Cui, and Liwei Wang. Semi-supervised semantic segmentation via adaptive equalization learning. Advances in Neural Information Processing Systems, 34:22106-22118, 2021. +[25] Wei-Chih Hung, Yi-Hsuan Tsai, Yan-Ting Liou, Yen-Yu Lin, and Ming-Hsuan Yang. Adversarial learning for semi-supervised semantic segmentation. arXiv preprint arXiv:1802.07934, 2018. +[26] Zhengkai Jiang, Yuxi Li, Ceyuan Yang, Peng Gao, Yabiao Wang, Ying Tai, and Chengjie Wang. Prototypical contrast adaptation for domain adaptive semantic segmentation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 36-54. Springer, 2022. + +[27] Tarun Kalluri, Girish Varma, Manmohan Chandraker, and CV Jawahar. Universal semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5259-5270, 2019. +[28] Rihuan Ke, Angelica I Aviles-Rivero, Saurabh Pandey, Saikumar Reddy, and Carola-Bibiane Schonlieb. A three-stage self-training framework for semi-supervised semantic segmentation. IEEE Transactions on Image Processing, 31:1805-1815, 2022. +[29] Donghyeon Kwon and Suha Kwak. Semi-supervised semantic segmentation with error localization network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9957-9967, 2022. +[30] Daiqing Li, Junlin Yang, Karsten Kreis, Antonio Torralba, and Sanja Fidler. Semantic segmentation with generative models: Semi-supervised learning and strong out-of-domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8300-8311, 2021. +[31] Jinlu Liu, Liang Song, and Yongqiang Qin. Prototype rectification for few-shot learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16, pages 741–756. Springer, 2020. +[32] Shikun Liu, Shuaifeng Zhi, Edward Johns, and Andrew Davison. Bootstrapping semantic segmentation with regional contrast. In International Conference on Learning Representations, 2022. +[33] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4258-4267, 2022. +[34] Binjie Mao, Xinbang Zhang, Lingfeng Wang, Qian Zhang, Shiming Xiang, and Chunhong Pan. Learning from the target: Dual prototype network for few shot semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1953-1961, 2022. +[35] Sudhanshu Mittal, Maxim Tatarchenko, and Thomas Brox. Semi-supervised semantic segmentation with high-and low-level consistency. IEEE transactions on pattern analysis and machine intelligence, 43(4):1369-1379, 2019. +[36] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979-1993, 2018. +[37] Viktor Olsson, Wilhelm Tranheden, Juliano Pinto, and Lennart Svensson. Classmix: Segmentation-based data augmentation for semi-supervised learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1369-1378, 2021. +[38] Yassine Ouali, Celine Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross-consistency training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. + +[39] Yingwei Pan, Ting Yao, Yehao Li, Yu Wang, Chong-Wah Ngo, and Tao Mei. Transferrable prototypical networks for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2239-2247, 2019. +[40] Jizong Peng, Guillermo Estrada, Marco Pedersoli, and Christian Desrosiers. Deep co-training for semi-supervised image segmentation. Pattern Recognition, 107:107269, 2020. +[41] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. +[42] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. +[43] Korawat Tanwisuth, Xinjie Fan, Huangjie Zheng, Shujian Zhang, Hao Zhang, Bo Chen, and Mingyuan Zhou. A prototype-oriented framework for unsupervised domain adaptation. Advances in Neural Information Processing Systems, 34:17194-17208, 2021. +[44] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. +[45] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7303-7313, 2021. +[46] Yuchao Wang, Haochen Wang, Yujun Shen, Jingjing Fei, Wei Li, Guoqiang Jin, Liwei Wu, Rui Zhao, and Xinyi Le. Semi-supervised semantic segmentation using unreliable pseudo labels. In Proceedings of the IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[47] Chen Wei, Kihyuk Sohn, Clayton Mellina, Alan Yuille, and Fan Yang. Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10857-10866, June 2021. +[48] Haoyu Xie, Changqi Wang, Mingkai Zheng, Minjing Dong, Shan You, and Chang Xu. Boosting semi-supervised semantic segmentation with probabilistic representations. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2938-2946, 2023. +[49] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020. +[50] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021. + +[51] Hai-Ming Xu, Lingqiao Liu, Qiuchen Bian, and Zhen Yang. Semi-supervised semantic segmentation with prototype-based consistency regularization. Advances in Neural Information Processing Systems, 2022. +[52] Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. St++: Make self-training work better for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4268-4277, 2022. +[53] Jianlong Yuan, Yifan Liu, Chunhua Shen, Zhibin Wang, and Hao Li. A simple baseline for semi-supervised semantic segmentation with strong data augmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8229-8238, 2021. +[54] Xiangyu Yue, Zangwei Zheng, Shanghang Zhang, Yang Gao, Trevor Darrell, Kurt Keutzer, and Alberto Sangiovanni Vincentelli. Prototypical cross-domain self-supervised learning for few-shot unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13834-13844, 2021. +[55] Xiangyun Zhao, Raviteja Vemulapalli, Philip Andrew Mansfield, Boqing Gong, Bradley Green, Lior Shapira, and Ying Wu. Contrastive learning for label efficient semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10623-10633, October 2021. +[56] Xu Zheng, Yunhao Luo, Hao Wang, Chong Fu, and Lin Wang. Transformer-cnn cohort: Semi-supervised semantic segmentation by the best of both students, 2022. +[57] Xu Zheng, Jinjing Zhu, Yexin Liu, Zidong Cao, Chong Fu, and Lin Wang. Both style and distortion matter: Dualpath unsupervised domain adaptation for panoramic semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1285-1295, June 2023. +[58] Yuanyi Zhong, Bodi Yuan, Hong Wu, Zhiqiang Yuan, Jian Peng, and Yu-Xiong Wang. Pixel contrastive-consistent semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7273-7282, 2021. +[59] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. +[60] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A prototype view. In CVPR, 2022. +[61] Yanning Zhou, Hang Xu, Wei Zhang, Bin Gao, and Pheng-Ann Heng. C3-semiseg: Contrastive semi-supervised segmentation via cross-set learning and dynamic classbalancing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7036-7045, 2021. \ No newline at end of file diff --git a/2023/Space Engage_ Collaborative Space Supervision for Contrastive-Based Semi-Supervised Semantic Segmentation/images.zip b/2023/Space Engage_ Collaborative Space Supervision for Contrastive-Based Semi-Supervised Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dde9c9d61b68174a68c17e346fca891eb293c05f --- /dev/null +++ b/2023/Space Engage_ Collaborative Space Supervision for Contrastive-Based Semi-Supervised Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a2a87f8289674bcd159b3bcff4c558bbd925e5295231859ed1d7446db120523 +size 584335 diff --git a/2023/Space Engage_ Collaborative Space Supervision for Contrastive-Based Semi-Supervised Semantic Segmentation/layout.json b/2023/Space Engage_ Collaborative Space Supervision for Contrastive-Based Semi-Supervised Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..565a9bc7147f646b642c3318b458a9c9cebaab2e --- /dev/null +++ b/2023/Space Engage_ Collaborative Space Supervision for Contrastive-Based Semi-Supervised Semantic Segmentation/layout.json @@ -0,0 +1,11838 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 103, + 514, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 103, + 514, + 140 + ], + "spans": [ + { + "bbox": [ + 80, + 103, + 514, + 140 + ], + "type": "text", + "content": "Space Engage: Collaborative Space Supervision for Contrastive-based Semi-Supervised Semantic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 160, + 471, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 160, + 471, + 176 + ], + "spans": [ + { + "bbox": [ + 121, + 160, + 471, + 176 + ], + "type": "text", + "content": "Changqi Wang" + }, + { + "bbox": [ + 121, + 160, + 471, + 176 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 121, + 160, + 471, + 176 + ], + "type": "text", + "content": ", Haoyu Xie" + }, + { + "bbox": [ + 121, + 160, + 471, + 176 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 121, + 160, + 471, + 176 + ], + "type": "text", + "content": ", Yuhui Yuan" + }, + { + "bbox": [ + 121, + 160, + 471, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 121, + 160, + 471, + 176 + ], + "type": "text", + "content": ", Chong Fu" + }, + { + "bbox": [ + 121, + 160, + 471, + 176 + ], + "type": "inline_equation", + "content": "^{1,4\\dagger}" + }, + { + "bbox": [ + 121, + 160, + 471, + 176 + ], + "type": "text", + "content": ", Xiangyu Yue" + }, + { + "bbox": [ + 121, + 160, + 471, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 180, + 515, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 180, + 515, + 209 + ], + "spans": [ + { + "bbox": [ + 77, + 180, + 515, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 180, + 515, + 209 + ], + "type": "text", + "content": " School of Computer Science and Engineering, Northeastern University, Shenyang, China \n" + }, + { + "bbox": [ + 77, + 180, + 515, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 180, + 515, + 209 + ], + "type": "text", + "content": " Microsoft Research Asia " + }, + { + "bbox": [ + 77, + 180, + 515, + 209 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 180, + 515, + 209 + ], + "type": "text", + "content": " The Chinese University of Hong Kong" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 60, + 209, + 533, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 209, + 533, + 224 + ], + "spans": [ + { + "bbox": [ + 60, + 209, + 533, + 224 + ], + "type": "text", + "content": "4 Key Laboratory of Intelligent Computing in Medical Image, Ministry of Education, NEU, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 232, + 493, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 232, + 493, + 243 + ], + "spans": [ + { + "bbox": [ + 100, + 232, + 493, + 243 + ], + "type": "text", + "content": "2101668@stu.neu.edu.cn, xiehaoyu@stumail.neu.edu.cn, yuyua@microsoft.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 174, + 246, + 419, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 246, + 419, + 257 + ], + "spans": [ + { + "bbox": [ + 174, + 246, + 419, + 257 + ], + "type": "text", + "content": "fuchong@mail.neu.edu.cn, xyyue@ie.cuhk.edu.hk" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 285, + 192, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 285, + 192, + 298 + ], + "spans": [ + { + "bbox": [ + 143, + 285, + 192, + 298 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 309, + 290, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 309, + 290, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 290, + 574 + ], + "type": "text", + "content": "Semi-Supervised Semantic Segmentation (S4) aims to train a segmentation model with limited labeled images and a substantial volume of unlabeled images. To improve the robustness of representations, powerful methods introduce a pixel-wise contrastive learning approach in latent space (i.e., representation space) that aggregates the representations to their prototypes in a fully supervised manner. However, previous contrastive-based S4 methods merely rely on the supervision from the model's output (logits) in logit space during unlabeled training. In contrast, we utilize the outputs in both logit space and representation space to obtain supervision in a collaborative way. The supervision from two spaces plays two roles: 1) reduces the risk of over-fitting to incorrect semantic information in logits with the help of representations; 2) enhances the knowledge exchange between the two spaces. Furthermore, unlike previous approaches, we use the similarity between representations and prototypes as a new indicator to tilt training those under-performing representations and achieve a more efficient contrastive learning process. Results on two public benchmarks demonstrate the competitive performance of our method compared with state-of-the-art methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 594, + 128, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 128, + 607 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 128, + 607 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 615, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 615, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 615, + 287, + 687 + ], + "type": "text", + "content": "Semantic segmentation is a fundamental task in computer vision, aiming to classify each pixel in an image. Significant progress [22, 4] has been made in training on high-quality labeled images using segmentation models composed of an encoder and a segmentation head. However, annotating images is expensive and time-consuming. Semi" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 309, + 285, + 544, + 373 + ], + "blocks": [ + { + "bbox": [ + 309, + 285, + 544, + 373 + ], + "lines": [ + { + "bbox": [ + 309, + 285, + 544, + 373 + ], + "spans": [ + { + "bbox": [ + 309, + 285, + 544, + 373 + ], + "type": "image", + "image_path": "59bfb879373b542a433ac5a4aaa2599ca463101edce0c030da9bf2c5369d8d07.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 376, + 543, + 479 + ], + "blocks": [ + { + "bbox": [ + 309, + 376, + 543, + 479 + ], + "lines": [ + { + "bbox": [ + 309, + 376, + 543, + 479 + ], + "spans": [ + { + "bbox": [ + 309, + 376, + 543, + 479 + ], + "type": "image", + "image_path": "eeb11bfac1f0b1f95d9c58b267eceba51bb7932527e4ba689ef3e2719777a707.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 485, + 547, + 541 + ], + "lines": [ + { + "bbox": [ + 305, + 485, + 547, + 541 + ], + "spans": [ + { + "bbox": [ + 305, + 485, + 547, + 541 + ], + "type": "text", + "content": "Figure 1. We enhance the knowledge exchange between the logit and representation spaces. Orange and blue represent different classes. Top: Existing contrastive-based S4 methods overlook the semantic information in representation space. Bottom: Our method uses dual-space collaborative supervision." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 567, + 546, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 567, + 546, + 602 + ], + "spans": [ + { + "bbox": [ + 305, + 567, + 546, + 602 + ], + "type": "text", + "content": "supervised Semantic Segmentation (S4) alleviates the thirst for annotation by leveraging unlabeled images to train segmentation models." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Most existing works learn from unlabeled images via self-training [44, 37, 15] or consistency regularization [40, 38, 13] strategies, both of which retrain the model with its predictions on unlabeled images. Recently, great success has been achieved by introducing pixel-wise contrastive learning to semantic segmentation, which endows the model with a stronger features-extracting ability by accessing a more discriminative representation space. Specifically, these methods [45, 23] project each pixel to repre" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "text", + "content": "*equal contribution" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 59, + 703, + 133, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 133, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 133, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 59, + 703, + 133, + 713 + ], + "type": "text", + "content": " corresponding author" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "931" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 262 + ], + "type": "text", + "content": "sensation space as a representation and regularize it in a fully supervised manner, i.e., aggregating the representations with the same class and separating them with different classes. In semi-supervised settings, due to limited labels, most methods [1, 32, 46] obtain supervision from the model's output logits in logit space during the unlabeled training process. However, recent contrastive-based semantic segmentation methods [1, 32, 46, 45] mainly focus on the learning process in logit space while only taking that in representation space as an auxiliary task. The unidirectional supervision makes training dominated by the predicted logits, leading to the neglect of information in the representation space. We argue that this kind of single-space supervision may incorrectly provide semantic guidance to representation learning and fails to facilitate knowledge exchange between the two spaces (see Sec. 5.1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 264, + 286, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 286, + 515 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 286, + 515 + ], + "type": "text", + "content": "In this work, we extend the single-space supervision to a dual-space supervision for contrastive-based S4 and propose Collaborative Space Supervision (CSS). Our key insight is to: i) utilize the semantic information in representations to obtain more reliable guidance during unlabeled training, and to enhance knowledge exchange between two spaces; ii) provide a more accurate reference for the model's performance on predicting each representation to tilt training those under-performing representations. To achieve objective i), we obtain dense semantic predictions by retrieving the nearest class prototype for each representation in the representation space and then engage with predictions from the logits for collaborative supervision of the model. For objective ii), we measure the similarity between the representations and prototypes and use the similarity after a normalization operation as the indicator for guiding the learning process in the representation space. Unlike previous works that utilize confidence as the indicator to involve representation learning, the similarity directly reflects the confusion level between representations and prototypes, resulting in more efficient representation learning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 516, + 286, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 516, + 286, + 599 + ], + "spans": [ + { + "bbox": [ + 49, + 516, + 286, + 599 + ], + "type": "text", + "content": "To summarize, our main contributions are three-fold: 1) The dual-space collaboration for contrastive-based S4, enhances the knowledge exchange between the logit and representation spaces. 2) Utilizing similarity to provide a more accurate reference for the model's performance in representation learning. 3) Extensive experiments on two S4 benchmarks demonstrate the effectiveness of our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 613, + 137, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 613, + 137, + 625 + ], + "spans": [ + { + "bbox": [ + 49, + 613, + 137, + 625 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 634, + 258, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 634, + 258, + 647 + ], + "spans": [ + { + "bbox": [ + 49, + 634, + 258, + 647 + ], + "type": "text", + "content": "2.1. Semi-supervised Semantic Segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 653, + 286, + 712 + ], + "type": "text", + "content": "The aim of S4 is to train a segmentation model with the semi-supervised setting (i.e., a few labeled images and a large number of unlabeled images) to classify each pixel in an entire image. The critical issue of S4 is how to leverage unlabeled images to train the model. Some methods [25," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 307, + 72, + 544, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 544, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 544, + 251 + ], + "type": "text", + "content": "27, 30, 35] based on GANs [16], adversarial training [36], and consistency regularization paradigm [38, 13, 40, 7, 56]. Meanwhile, self-training [28, 44, 49, 61, 53] is also a striking paradigm, which always generates pseudo-labels from model and retransmits the model with the combined supervision of human annotations and pseudo-labels. One essential issue of self-training is the accuracy of pseudo-labels. Some methods [29, 33, 14, 42, 52] try to polish pseudo-labels and provide reliable guidance. Some methods [21, 47, 24, 17] focus on the class-imbalance problems in the dataset and try to alleviate the negative effect from class-biased pseudolabels generated by the model pre-trained on imbalanced labeled images. We build our framework based on the self-training and additionally explore semantic information among different images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 260, + 477, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 260, + 477, + 273 + ], + "spans": [ + { + "bbox": [ + 308, + 260, + 477, + 273 + ], + "type": "text", + "content": "2.2. Pixel-wise Contrastive Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 279, + 544, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 279, + 544, + 506 + ], + "spans": [ + { + "bbox": [ + 307, + 279, + 544, + 506 + ], + "type": "text", + "content": "Pixel-wise contrastive learning explores semantic relations not only in the individual image but also among different images. Different from instance-wise contrastive learning [19, 6, 3], pixel-wise contrastive learning [50, 55, 5, 58] project each pixel to the representation in representation space with the cooperation of encoder and representation head. Representations are then aggregated in their prototypes and are separated from each other in different classes. In semi-supervised settings, most methods [1, 32, 46, 48] use pseudo-labels based on logits to provide semantic information contrastive learning process during training on unlabeled images. Meanwhile, the confidence of logit is used as an indicator to involve the contrastive learning process, e.g., [32] uses the hard representations whose corresponding logit confidence is lower than a threshold to contrast for effective training. As opposed to the above methods, we use collaborative space supervision for contrastive learning on unlabeled images and use a new indicator to involve the contrastive learning progress." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 515, + 450, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 450, + 528 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 450, + 528 + ], + "type": "text", + "content": "2.3. Prototype-based Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 534, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 544, + 712 + ], + "type": "text", + "content": "Prototype-based learning has been widely studied in few-shot learning [41, 11, 31, 34] and unsupervised domain adaption [54, 26, 43, 39, 57]. Recently, it is restudied in semantic segmentation as known as a non-parametric prototype-based classifier [60]. Concretely, the classes in the dataset are presented by a set of non-learnable prototypes, and the dense semantic predictions are thus achieved by assigning the output features to its most similar prototype. Under semi-supervised settings, some methods [51] maintain the consistency between predictions from a linear predictor and a prototype-based predictor. The two predictors are followed by the encoder and project the features to logit space and representation space, respectively. In this work, we combine the semantic information in the logit and representation spaces to provide supervision in a collabora" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 750, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 757 + ], + "type": "text", + "content": "932" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 217, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 217, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 217, + 84 + ], + "type": "text", + "content": "tive way during semi-supervised learning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 97, + 129, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 97, + 129, + 110 + ], + "spans": [ + { + "bbox": [ + 47, + 97, + 129, + 110 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "text", + "content": "In the S4 task, we have a small labeled set " + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_l = \\{(\\pmb{x}_i^l,\\pmb{y}_i^l)\\}_{i=1}^{N_l}" + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "text", + "content": " and a large unlabeled set " + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u = \\{\\pmb{x}_i^u\\}_{i=1}^{N_u}" + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^l,\\pmb{x}_i^u\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "inline_equation", + "content": "H,W" + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "text", + "content": " denote the height, and the width, respectively. And ground truth " + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "inline_equation", + "content": "\\pmb{y}_i^l\\in \\{0,1\\}^{H\\times W\\times |C|}" + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "text", + "content": " with the set of class " + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "text", + "content": ". The goal is to boost model performance with " + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u" + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "text", + "content": ". The base model consists of an encoder " + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "text", + "content": " and a segmentation head " + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "text", + "content": ", which projects features to the logit space " + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H\\times W\\times |C|}" + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "text", + "content": ". We adopt Self-Training (ST) and pixel-wise contrastive learning to our framework, as described in Sec. 3.1. The supervision for " + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u" + }, + { + "bbox": [ + 47, + 118, + 289, + 262 + ], + "type": "text", + "content": " is produced by the collaboration between the logit and representation spaces, as described in Sec. 3.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 270, + 257, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 257, + 284 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 257, + 284 + ], + "type": "text", + "content": "3.1. ST and Pixel-wise Contrastive Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 289, + 289, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 289, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 289, + 471 + ], + "type": "text", + "content": "The main idea of self-training is to pre-train a model on labeled images and use it to produce pseudo-labels as supervision for unlabeled images. A typical framework is the teacher-student framework [44], which consists of a student model and a teacher model. Both the student model and the teacher model are constructed by an encoder and a segmentation head. Parameters of the student model are optimized via Stochastic Gradient Decent (SGD) while parameters of the teacher model are updated by the Exponential Moving Average (EMA) of student model parameters. We denote the encoder and the segmentation head in the student model by " + }, + { + "bbox": [ + 46, + 289, + 289, + 471 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 46, + 289, + 289, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 289, + 289, + 471 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 46, + 289, + 289, + 471 + ], + "type": "text", + "content": " while denoting those in the teacher model by " + }, + { + "bbox": [ + 46, + 289, + 289, + 471 + ], + "type": "inline_equation", + "content": "f^{\\prime}(\\cdot)" + }, + { + "bbox": [ + 46, + 289, + 289, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 289, + 289, + 471 + ], + "type": "inline_equation", + "content": "g^{\\prime}(\\cdot)" + }, + { + "bbox": [ + 46, + 289, + 289, + 471 + ], + "type": "text", + "content": ". The pseudo-labels " + }, + { + "bbox": [ + 46, + 289, + 289, + 471 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_i^{u,lgt}" + }, + { + "bbox": [ + 46, + 289, + 289, + 471 + ], + "type": "text", + "content": " are produced based on the teacher model's output logits " + }, + { + "bbox": [ + 46, + 289, + 289, + 471 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}_i^u = g'(f'(x_i^u))" + }, + { + "bbox": [ + 46, + 289, + 289, + 471 + ], + "type": "text", + "content": " in logit space, formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 480, + 287, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 480, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 287, + 499 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {y}} _ {i} ^ {u, l g t} = \\mathbf {1} _ {c} (\\underset {c} {\\arg \\max } \\{\\hat {\\boldsymbol {p}} _ {i, c} ^ {u} \\} _ {c \\in C}), \\tag {1}", + "image_path": "0d05d1efa1d0a763bb3fc535c91e236bde4eefdabe5e308aaf070843b0b1658a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 510, + 259, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 259, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 259, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 510, + 259, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_c(\\cdot)" + }, + { + "bbox": [ + 47, + 510, + 259, + 521 + ], + "type": "text", + "content": " denotes the one-hot encoding of class " + }, + { + "bbox": [ + 47, + 510, + 259, + 521 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 510, + 259, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": "In order to enhance the ability of the model itself to extract features, recent works [32, 46, 1] additionally employ pixel-wise contrastive learning and introduce a representation head to both teacher and student models. We denote the representation head in the student model as " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " and that in the teacher model as " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "h^{\\prime}(\\cdot)" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": ". The pixel " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " of the class " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " are projected as representations " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "z_{ci}" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " in representation space by the cooperating of " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "z_{ci} = h(f(x_i))" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": ". And the representation " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "z_{ci}" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " is then aggregated to its class centroid (prototype) while separated from representations in different classes " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "z_{\\tilde{c} i}" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " (negatives). The semantic guidance for contrastive learning is from the combination of ground truth " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "y_{i}^{l}" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " and pseudo-labels " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\hat{y}_i^{u,lgt}" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " in logit space. Moreover, in order to emphasize the reliable and crucial aspects during unlabeled and contrastive learning, a sampling strategy is adopted to select valid pixels " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " according to their corre" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "content": "sponding confidence, i.e. the student model's output logits " + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "content": " after a Softmax operation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 96, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 547, + 228 + ], + "type": "text", + "content": "Discussion. In recent works [46, 1, 32], the supervision of unlabeled images is derived solely from the logit space. This overlooks the potential benefits of the supervision from the representation space, leading to two potential limitations: 1) the pseudo-labels " + }, + { + "bbox": [ + 304, + 96, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\hat{y}_i^{u,lgt}" + }, + { + "bbox": [ + 304, + 96, + 547, + 228 + ], + "type": "text", + "content": " obtained from the logit space may contain noise and miss the opportunity to be corrected by semantic information from the representation space; 2) since the confidence from logit space is used as the indicator " + }, + { + "bbox": [ + 304, + 96, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\hat{j}_i" + }, + { + "bbox": [ + 304, + 96, + 547, + 228 + ], + "type": "text", + "content": " for the sampling strategy, learning in the representation space may not be critical enough due to the different confusing parts between the two spaces." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 228, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 300 + ], + "type": "text", + "content": "To mitigate these limitations, we produce pseudo-labels from the representation space and combine them with pseudo-labels from the logit space to provide higher-quality supervision during unlabeled training. Meanwhile, we obtain a new indicator from the representation space for the more effective sampling strategy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 306, + 514, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 306, + 514, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 306, + 514, + 319 + ], + "type": "text", + "content": "3.2. Supervision from Representation Space" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 324, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 545, + 384 + ], + "type": "text", + "content": "In this section, we detail the approach to obtain the pseudo-labels from the representation space. Meanwhile, we simultaneously access a new indicator for the sampling strategy in representation spaces, which provides a critical reference in the contrastive learning process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 384, + 545, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 444 + ], + "type": "text", + "content": "Specifically, we first build a set of prototypes for each class and obtain the pseudo-labels by retrieving the nearest prototype for each representation. We calculate the centroid of all representations in the current class " + }, + { + "bbox": [ + 304, + 384, + 545, + 444 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 384, + 545, + 444 + ], + "type": "text", + "content": " as the prototype " + }, + { + "bbox": [ + 304, + 384, + 545, + 444 + ], + "type": "inline_equation", + "content": "\\rho_{c}" + }, + { + "bbox": [ + 304, + 384, + 545, + 444 + ], + "type": "text", + "content": ", which is formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 393, + 451, + 545, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 451, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 393, + 451, + 545, + 482 + ], + "type": "interline_equation", + "content": "\\rho_ {c} = \\frac {1}{N _ {c}} \\sum_ {i} ^ {N _ {c}} z _ {i} ^ {\\prime}, \\tag {2}", + "image_path": "65b892aafffdd2fd5204ae944570ef04788adffa588261b5be3573d8ac7ca185.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "text", + "content": " is the total number of representations of current class " + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "inline_equation", + "content": "\\pmb{z}_i^{\\prime}" + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "text", + "content": " is the representation projected by the cooperation of the " + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "inline_equation", + "content": "f^{\\prime}(\\cdot)" + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "inline_equation", + "content": "h^\\prime (\\cdot)" + }, + { + "bbox": [ + 304, + 487, + 545, + 546 + ], + "type": "text", + "content": ". Meanwhile, to include more representation information, we update the prototype along the sequential iterations with EMA as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 357, + 555, + 545, + 568 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 555, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 357, + 555, + 545, + 568 + ], + "type": "interline_equation", + "content": "\\hat {\\rho} _ {c} (t) = \\alpha \\hat {\\rho} _ {c} (t - 1) + (1 - \\alpha) \\rho_ {c} (t), \\tag {3}", + "image_path": "3e54e032b1a6e5c38497992f854ff76f5f3228ae6ffba3123cd458dd55b40645.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 574, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 574, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 574, + 545, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 574, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\hat{\\rho}_c(t),\\hat{\\rho}_c(t - 1)" + }, + { + "bbox": [ + 304, + 574, + 545, + 635 + ], + "type": "text", + "content": " mean the current " + }, + { + "bbox": [ + 304, + 574, + 545, + 635 + ], + "type": "inline_equation", + "content": "t^{th}" + }, + { + "bbox": [ + 304, + 574, + 545, + 635 + ], + "type": "text", + "content": " prototype and last " + }, + { + "bbox": [ + 304, + 574, + 545, + 635 + ], + "type": "inline_equation", + "content": "(t - 1)^{th}" + }, + { + "bbox": [ + 304, + 574, + 545, + 635 + ], + "type": "text", + "content": " prototype in iterations, " + }, + { + "bbox": [ + 304, + 574, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\rho_{c}(t)" + }, + { + "bbox": [ + 304, + 574, + 545, + 635 + ], + "type": "text", + "content": " means the prototype calculated by Eq.2 in current iteration, and " + }, + { + "bbox": [ + 304, + 574, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 574, + 545, + 635 + ], + "type": "text", + "content": " is a hyper-parameter that controls the updating speed. Thus, the pseudo-label from the representation space is achieved by:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 640, + 545, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 640, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 315, + 640, + 545, + 658 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {y}} _ {i} ^ {u, r e p} = \\mathbf {1} _ {\\hat {c}} (\\hat {c}), w i t h \\hat {c} = \\underset {c} {\\arg \\max } \\left\\{\\operatorname {s i m} \\left(\\boldsymbol {z} _ {i} ^ {\\prime}, \\hat {\\boldsymbol {\\rho}} _ {c} (t)\\right) \\right\\} _ {c \\in C}, \\tag {4}", + "image_path": "956521ee982d32c39969961c875abe626922ca56fe679cc1876c40ae83f6460f.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 501, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 501, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 501, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 665, + 501, + 677 + ], + "type": "inline_equation", + "content": "sim(\\cdot)" + }, + { + "bbox": [ + 304, + 665, + 501, + 677 + ], + "type": "text", + "content": " is defined as the cosine similarity." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "As for the indicator for the sampling strategy in the representation space, we use the Softmax function on the similarity among the representation and all prototypes, which is" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "933" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 71, + 518, + 264 + ], + "blocks": [ + { + "bbox": [ + 75, + 71, + 518, + 264 + ], + "lines": [ + { + "bbox": [ + 75, + 71, + 518, + 264 + ], + "spans": [ + { + "bbox": [ + 75, + 71, + 518, + 264 + ], + "type": "image", + "image_path": "9460f85ac9f86af3b5a8ad3d7cde59fdaa955bb3d1ce27cd70bd3f30f15d9484.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 268, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 268, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 547, + 304 + ], + "type": "text", + "content": "Figure 2. Overview of our framework. Our training pipeline consists of learning in two spaces: logit space and representation space. The pseudo-labels " + }, + { + "bbox": [ + 46, + 268, + 547, + 304 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_i^u" + }, + { + "bbox": [ + 46, + 268, + 547, + 304 + ], + "type": "text", + "content": " during unlabeled training are produced by the collaboration of two spaces with mix pseudo-labeling strategy (1) or cross pseudo-labeling strategy (2). The indicator for representation learning is produced by similarity " + }, + { + "bbox": [ + 46, + 268, + 547, + 304 + ], + "type": "inline_equation", + "content": "(s_1, s_2, \\text{and} s_3)" + }, + { + "bbox": [ + 46, + 268, + 547, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 322, + 99, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 322, + 99, + 333 + ], + "spans": [ + { + "bbox": [ + 47, + 322, + 99, + 333 + ], + "type": "text", + "content": "followed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 339, + 287, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 339, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 63, + 339, + 287, + 366 + ], + "type": "interline_equation", + "content": "\\hat {j} _ {i} ^ {u, r e p} = \\frac {e ^ {s i m \\left(\\boldsymbol {z} _ {\\boldsymbol {c} i} , \\hat {\\rho} _ {\\boldsymbol {c}} (t)\\right) / \\tau}}{e ^ {s i m \\left(\\boldsymbol {z} _ {\\boldsymbol {c} i} , \\hat {\\rho} _ {\\boldsymbol {c}} (t)\\right) / \\tau} + \\sum_ {\\tilde {c} \\in \\tilde {C}} e ^ {s i m \\left(\\boldsymbol {z} _ {\\boldsymbol {c} i} , \\hat {\\rho} _ {\\tilde {c}} (t)\\right) / \\tau}}, \\tag {5}", + "image_path": "e8b7609d71c8b29d57c44c48792f893b701e47f619290b681380b3d4d8930582.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 373, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 459 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 373, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\hat{\\rho}_{\\tilde{c}}(t)" + }, + { + "bbox": [ + 46, + 373, + 287, + 459 + ], + "type": "text", + "content": " means the prototype with different classes from " + }, + { + "bbox": [ + 46, + 373, + 287, + 459 + ], + "type": "inline_equation", + "content": "z_{ci}" + }, + { + "bbox": [ + 46, + 373, + 287, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 373, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 373, + 287, + 459 + ], + "type": "text", + "content": " is a hyper-parameter. Different from using confidence from logit space as the indicator to involve representation learning [32, 46], the Softmax similarity directly helps the model to discover the confusion between representations and their prototypes and focus on them during the subsequent training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 464, + 234, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 234, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 234, + 478 + ], + "type": "text", + "content": "3.3. Collaboration Between Two Spaces" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 483, + 287, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 520 + ], + "type": "text", + "content": "With the pseudo-labels in two spaces, we propose two pseudo-labeling strategies to strengthen the collaboration between two spaces and obtain more reliable pseudo-labels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 526, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 58, + 526, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 526, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 58, + 526, + 287, + 612 + ], + "type": "text", + "content": "- Mix pseudo-labeling. To mitigate the negative effects of inherent noise from both two spaces during pseudolabeling, we adopt the mix pseudo-labeling strategy that only considers the mutually agreeable pseudolabels between the two spaces. Specifically, we define the set of final pseudo-labels as " + }, + { + "bbox": [ + 58, + 526, + 287, + 612 + ], + "type": "inline_equation", + "content": "\\hat{Y}^u = \\hat{Y}^{u,lgt} \\cap \\hat{Y}^{u,rep}" + }, + { + "bbox": [ + 58, + 526, + 287, + 612 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 58, + 526, + 287, + 612 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_i^{u,lgt} \\in \\hat{Y}^{u,lgt}" + }, + { + "bbox": [ + 58, + 526, + 287, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 58, + 526, + 287, + 612 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_i^{u,rep} \\in \\hat{Y}^{u,rep}" + }, + { + "bbox": [ + 58, + 526, + 287, + 612 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 618, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 618, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 618, + 288, + 713 + ], + "type": "text", + "content": "- Cross pseudo-labeling. Inspired by recent researches [7, 38] that maintain consistency among the predictions of the same image across different models or decoders in different views, we propose a cross pseudolabeling strategy that leverages pseudo-labels from one space to supervise the other. Specifically, we use pseudo-labels " + }, + { + "bbox": [ + 58, + 618, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_i^{u,rep}" + }, + { + "bbox": [ + 58, + 618, + 288, + 713 + ], + "type": "text", + "content": " to supervise the logit space, and vice versa." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 322, + 545, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 545, + 489 + ], + "type": "text", + "content": "The strengths of using pseudo-labels from two spaces in the collaborative way are twofold: 1) obtaining more reliable supervision during unlabeled training, and 2) enabling the strengths of learning in different spaces to complement each other. Since the learning in different feature spaces concentrates on different parts of features, i.e., the logit space mainly focuses on the most discriminative part of features while the representation space treats all parts of features equally, the performance of pseudo-labels from two spaces varies across different classes or regions of images. Therefore, our collaborative pseudo-labeling strategies exchange knowledge between two spaces and provide higher-quality supervision during unlabeled training. The experimental proof is in Sec. 5.1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 491, + 546, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 546, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 546, + 612 + ], + "type": "text", + "content": "As for indicators, we use confidence as the indicator " + }, + { + "bbox": [ + 304, + 491, + 546, + 612 + ], + "type": "inline_equation", + "content": "\\hat{j}_i^{u,lgt}" + }, + { + "bbox": [ + 304, + 491, + 546, + 612 + ], + "type": "text", + "content": " for learning logit space and Softmax similarity as the indicator " + }, + { + "bbox": [ + 304, + 491, + 546, + 612 + ], + "type": "inline_equation", + "content": "\\hat{j}_i^{u,rep}" + }, + { + "bbox": [ + 304, + 491, + 546, + 612 + ], + "type": "text", + "content": " for learning representation space. We argue that the confusing parts of learning in both two spaces are varied due to the different parts of features being concentrated in each space. Therefore, this strategy allows the learning in different spaces to focus on their own confusing parts, which can be more effective than using a single indicator when mining confusing parts for both spaces in the training process. The experimental proof is in Sec. 5.2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 621, + 417, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 417, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 417, + 634 + ], + "type": "text", + "content": "3.4. Training Objective" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "text", + "content": "With the indicators " + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\hat{j}_i^{u,lgt}" + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\hat{j}_i^{u,rep}" + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "text", + "content": ", we adopt some threshold sampling strategies. In logit space, we set a threshold " + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\delta_u" + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "text", + "content": " during unlabeled learning and logits " + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\hat{p}_i^u" + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "text", + "content": " whose indicator " + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\hat{j}_i^{u,lgt}" + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "text", + "content": " is higher than " + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\delta_u" + }, + { + "bbox": [ + 304, + 639, + 545, + 715 + ], + "type": "text", + "content": " will be regarded as the valid logits in logit space. In representation space, our sample strategy can be divided into three parts: 1) Valid Sam" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "934" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "pling Strategy. Similar to the sampling strategy in logit space, a threshold " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "\\delta_w" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": " is used to sample representations whose indicator " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "\\hat{j}_i^{u,rep}" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": " is higher than " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "\\delta_w" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": ". 2) Hard Sampling Strategy. We adopt the hard sampling strategy for tilting to train those confusing representations. Specifically, we set a threshold " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "\\delta_s" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": " to sample representations whose indicator " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "\\hat{j}_i^{u,rep}" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": " is lower than " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "\\delta_s" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": ". 3) Negative Sampling Strategy. We sample negatives according to the similarity between the prototype of current class " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": " and other prototypes. Concretely, the negatives are more likely to be sampled if its prototype is more similar to the prototype of the current class." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "text", + "content": "Cooperated with the ground truth " + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\pmb{y}_i^l" + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "text", + "content": ", pseudo-labels " + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\pmb{y}_i^u" + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "text", + "content": " produced from two spaces in a collaborative way, and different sampling strategies in two spaces, the total learning object is composed with a supervised loss " + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "text", + "content": ", an unsupervised loss " + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "text", + "content": ", and a contrastive loss " + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 124, + 288, + 287, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 288, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 124, + 288, + 287, + 300 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {s} + \\mathcal {L} _ {u} + \\lambda_ {c} \\mathcal {L} _ {c}, \\tag {6}", + "image_path": "34495858b408cf619b19826f7156b26c9fcae77bc59dd91255d75728197050eb.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 311, + 289, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 289, + 359 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 289, + 359 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 311, + 289, + 359 + ], + "type": "inline_equation", + "content": "\\lambda_{c}" + }, + { + "bbox": [ + 47, + 311, + 289, + 359 + ], + "type": "text", + "content": " is used to tune the contribution between logit space and representation space. Specifically, " + }, + { + "bbox": [ + 47, + 311, + 289, + 359 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 47, + 311, + 289, + 359 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 311, + 289, + 359 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 47, + 311, + 289, + 359 + ], + "type": "text", + "content": " are constructed by the Cross-Entropy (CE) " + }, + { + "bbox": [ + 47, + 311, + 289, + 359 + ], + "type": "inline_equation", + "content": "\\ell_{ce}" + }, + { + "bbox": [ + 47, + 311, + 289, + 359 + ], + "type": "text", + "content": " and can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 366, + 287, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 366, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 114, + 366, + 287, + 396 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\frac {1}{| \\mathcal {B} _ {l} |} \\sum_ {\\boldsymbol {x} _ {i} ^ {l} \\in \\mathcal {B} _ {l}} \\ell_ {c e} \\left(\\boldsymbol {p} _ {i} ^ {l}, \\boldsymbol {y} _ {i} ^ {l}\\right), \\tag {7}", + "image_path": "e445c901ce191d33fe5e99eb2ca348781cb7590f220d362b8a062ac8634c48e2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 407, + 287, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 407, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 110, + 407, + 287, + 437 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {u} = \\frac {1}{| \\hat {\\mathcal {B}} _ {u} |} \\sum_ {\\boldsymbol {x} _ {i} ^ {u} \\in \\hat {\\mathcal {B}} _ {u}} \\ell_ {c e} \\left(\\boldsymbol {p} _ {i} ^ {u}, \\hat {\\boldsymbol {y}} _ {i} ^ {u}\\right), \\tag {8}", + "image_path": "bbe28d26fa9e55a345f5f7a62d855f845449a47342be0754d36fb2e2748ae681.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 444, + 287, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 287, + 492 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 444, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_l" + }, + { + "bbox": [ + 47, + 444, + 287, + 492 + ], + "type": "text", + "content": " denotes the labeled images in a mini-batch and " + }, + { + "bbox": [ + 47, + 444, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{B}}_u" + }, + { + "bbox": [ + 47, + 444, + 287, + 492 + ], + "type": "text", + "content": " is the subsets that sampled from unlabeled images in a mini-batch according to the sampling strategy. Meanwhile, the contrastive loss " + }, + { + "bbox": [ + 47, + 444, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 47, + 444, + 287, + 492 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 500, + 287, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 287, + 533 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = - \\frac {1}{| C | \\times | \\hat {\\mathcal {Z}} _ {c} |} \\sum_ {c \\in C} \\sum_ {\\boldsymbol {z} _ {c i} \\in \\hat {\\mathcal {Z}} _ {c}} \\tag {9}", + "image_path": "17ec9f379deca4db8146e9c4991a070f459af65c490c7cdb416611234f856c41.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 531, + 266, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 531, + 266, + 558 + ], + "spans": [ + { + "bbox": [ + 76, + 531, + 266, + 558 + ], + "type": "interline_equation", + "content": "l o g [ \\frac {e ^ {s (\\pmb {z} _ {c i} , \\hat {\\pmb {\\rho}} _ {c}) / \\tau}}{e ^ {s (\\pmb {z} _ {c i} , \\hat {\\pmb {\\rho}} _ {c} (t))) / \\tau} + \\sum_ {\\tilde {e} \\in \\tilde {C}} \\sum_ {\\pmb {z} _ {\\tilde {e} i} \\in \\hat {\\pmb {Z}} _ {\\tilde {c}}} e ^ {s (\\pmb {z} _ {c i} , \\pmb {z} _ {\\tilde {c} i}) / \\tau} ],", + "image_path": "22f68b2dfddec8a3ee3714c00b6c8359f27d2b3472b5aa0da7ad1e6a69897370.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Z}}_c" + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "text", + "content": " is the subset sampled from the set of the representations which belong to class " + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "text", + "content": " according to the sampling strategy, " + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Z}}_{\\tilde{c}}" + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "text", + "content": " is the subset sampled from the set of the representations which bot belong to class " + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\tilde{C}" + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "text", + "content": " denotes the subset of " + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "text", + "content": " with class " + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "text", + "content": " removed, and the supervision information comes from the final pseudo-labels " + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_i^u" + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "text", + "content": " after the pseudo-labeling strategies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "The whole framework is shown in Fig. 2. All the pseudocode of producing pseudo-labels and their indicators from the logit and representation spaces is shown in Algorithm 1, and the pseudo-code of the mix pseudo-labeling strategy is shown in Algorithm 2. The pseudo-code is PyTorch-like." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "Algorithm 1 Pseudo-code of producing pseudo-labels and indicators" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 95, + 545, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 95, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 305, + 95, + 545, + 118 + ], + "type": "text", + "content": "Network: Teacher's encoder " + }, + { + "bbox": [ + 305, + 95, + 545, + 118 + ], + "type": "inline_equation", + "content": "f'" + }, + { + "bbox": [ + 305, + 95, + 545, + 118 + ], + "type": "text", + "content": ", segmentation head " + }, + { + "bbox": [ + 305, + 95, + 545, + 118 + ], + "type": "inline_equation", + "content": "g'" + }, + { + "bbox": [ + 305, + 95, + 545, + 118 + ], + "type": "text", + "content": ", representation head " + }, + { + "bbox": [ + 305, + 95, + 545, + 118 + ], + "type": "inline_equation", + "content": "h'" + }, + { + "bbox": [ + 305, + 95, + 545, + 118 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 120, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 120, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 120, + 545, + 144 + ], + "type": "text", + "content": "Input: Unlabeled mini-batch " + }, + { + "bbox": [ + 305, + 120, + 545, + 144 + ], + "type": "inline_equation", + "content": "B_{u}" + }, + { + "bbox": [ + 305, + 120, + 545, + 144 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 305, + 120, + 545, + 144 + ], + "type": "inline_equation", + "content": "X^{u}" + }, + { + "bbox": [ + 305, + 120, + 545, + 144 + ], + "type": "text", + "content": ", the set of prototypes " + }, + { + "bbox": [ + 305, + 120, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\rho}_c(t)\\}_{c\\in C}" + }, + { + "bbox": [ + 305, + 120, + 545, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 144, + 545, + 285 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 312, + 144, + 393, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 144, + 393, + 155 + ], + "spans": [ + { + "bbox": [ + 312, + 144, + 393, + 155 + ], + "type": "text", + "content": "1: for " + }, + { + "bbox": [ + 312, + 144, + 393, + 155 + ], + "type": "inline_equation", + "content": "X_{u} \\in B_{u}" + }, + { + "bbox": [ + 312, + 144, + 393, + 155 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 156, + 492, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 156, + 492, + 167 + ], + "spans": [ + { + "bbox": [ + 312, + 156, + 492, + 167 + ], + "type": "text", + "content": "2: # Predict logits from unlabeled images." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 167, + 422, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 167, + 422, + 179 + ], + "spans": [ + { + "bbox": [ + 312, + 167, + 422, + 179 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 312, + 167, + 422, + 179 + ], + "type": "inline_equation", + "content": "\\hat{P}^{u,lgt}\\gets g'(f'(X^u))" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 180, + 545, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 180, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 312, + 180, + 545, + 202 + ], + "type": "text", + "content": "4: # Produce labels and indicators based on logits with Eq. 1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 203, + 483, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 203, + 483, + 216 + ], + "spans": [ + { + "bbox": [ + 312, + 203, + 483, + 216 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 312, + 203, + 483, + 216 + ], + "type": "inline_equation", + "content": "\\hat{Y}^{u,lgt},\\hat{J}^{u,lgt}\\gets lgt\\text{-} pseudo(\\tilde{P}^{u,lgt})" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 216, + 530, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 216, + 530, + 227 + ], + "spans": [ + { + "bbox": [ + 312, + 216, + 530, + 227 + ], + "type": "text", + "content": "6: # Predict representations from unlabeled images." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 227, + 425, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 227, + 425, + 239 + ], + "spans": [ + { + "bbox": [ + 312, + 227, + 425, + 239 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 312, + 227, + 425, + 239 + ], + "type": "inline_equation", + "content": "\\hat{R}^{u,rep}\\gets h'(f'(X^u))" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 239, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 239, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 312, + 239, + 545, + 262 + ], + "type": "text", + "content": "8: # Produce labels and indicators based on representations with Eq. 4 and Eq. 5." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 262, + 543, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 262, + 543, + 276 + ], + "spans": [ + { + "bbox": [ + 312, + 262, + 543, + 276 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 312, + 262, + 543, + 276 + ], + "type": "inline_equation", + "content": "\\hat{Y}^{u,rep},\\hat{J}^{u,rep}\\gets rep\\text{-} pseudo(\\hat{R}^{u,rep},\\{\\hat{\\rho}_c(t)\\}_{c\\in C})" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 276, + 358, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 276, + 358, + 285 + ], + "spans": [ + { + "bbox": [ + 312, + 276, + 358, + 285 + ], + "type": "text", + "content": "10: end for" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 302, + 544, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 302, + 544, + 315 + ], + "spans": [ + { + "bbox": [ + 306, + 302, + 544, + 315 + ], + "type": "text", + "content": "Algorithm 2 Pseudo-code of mix pseudo-labeling strategy" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 317, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 317, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 317, + 545, + 342 + ], + "type": "text", + "content": "Input: Pseudo-labels from logit space " + }, + { + "bbox": [ + 306, + 317, + 545, + 342 + ], + "type": "inline_equation", + "content": "\\hat{Y}^{u,lgt}" + }, + { + "bbox": [ + 306, + 317, + 545, + 342 + ], + "type": "text", + "content": " and from representation space " + }, + { + "bbox": [ + 306, + 317, + 545, + 342 + ], + "type": "inline_equation", + "content": "\\hat{Y}^{u,rep}" + }, + { + "bbox": [ + 306, + 317, + 545, + 342 + ], + "type": "text", + "content": ". Indicators " + }, + { + "bbox": [ + 306, + 317, + 545, + 342 + ], + "type": "inline_equation", + "content": "\\hat{J}^{u,lgt}" + }, + { + "bbox": [ + 306, + 317, + 545, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 317, + 545, + 342 + ], + "type": "inline_equation", + "content": "\\hat{J}^{u,rep}" + }, + { + "bbox": [ + 306, + 317, + 545, + 342 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 342, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 342, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 545, + 365 + ], + "type": "text", + "content": "Notation: Threshold " + }, + { + "bbox": [ + 306, + 342, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\delta_{u}" + }, + { + "bbox": [ + 306, + 342, + 545, + 365 + ], + "type": "text", + "content": " for sampling strategy in " + }, + { + "bbox": [ + 306, + 342, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 306, + 342, + 545, + 365 + ], + "type": "text", + "content": ", weak threshold " + }, + { + "bbox": [ + 306, + 342, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\delta_w" + }, + { + "bbox": [ + 306, + 342, + 545, + 365 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 366, + 545, + 489 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 312, + 366, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 366, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 312, + 366, + 545, + 390 + ], + "type": "text", + "content": "1: # Sampling valid pseudo-labels according to the sampling strategy in logit space." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 391, + 488, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 391, + 488, + 405 + ], + "spans": [ + { + "bbox": [ + 312, + 391, + 488, + 405 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 312, + 391, + 488, + 405 + ], + "type": "inline_equation", + "content": "\\hat{Y}_{val}^{u,lgt}\\gets" + }, + { + "bbox": [ + 312, + 391, + 488, + 405 + ], + "type": "text", + "content": " sample_lgt( " + }, + { + "bbox": [ + 312, + 391, + 488, + 405 + ], + "type": "inline_equation", + "content": "\\hat{Y}^{u,lgt},\\hat{J}^{u,lgt},\\delta_u)" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 405, + 545, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 405, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 312, + 405, + 545, + 427 + ], + "type": "text", + "content": "3: # Sampling valid pseudo-labels according to the valid sampling strategy in logit space." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 427, + 498, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 427, + 498, + 441 + ], + "spans": [ + { + "bbox": [ + 312, + 427, + 498, + 441 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 312, + 427, + 498, + 441 + ], + "type": "inline_equation", + "content": "\\hat{Y}_{val}^{u,rep}\\gets" + }, + { + "bbox": [ + 312, + 427, + 498, + 441 + ], + "type": "text", + "content": " sample_rep( " + }, + { + "bbox": [ + 312, + 427, + 498, + 441 + ], + "type": "inline_equation", + "content": "\\hat{Y}^{u,rep},\\hat{J}^{u,rep},\\delta_w)" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 441, + 495, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 441, + 495, + 452 + ], + "spans": [ + { + "bbox": [ + 312, + 441, + 495, + 452 + ], + "type": "text", + "content": "5: # Produce the mask for mix pseudo-labels." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 452, + 441, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 452, + 441, + 465 + ], + "spans": [ + { + "bbox": [ + 312, + 452, + 441, + 465 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 312, + 452, + 441, + 465 + ], + "type": "inline_equation", + "content": "mask_{y} \\gets \\hat{Y}^{u,rep}." + }, + { + "bbox": [ + 312, + 452, + 441, + 465 + ], + "type": "text", + "content": " eq(" + }, + { + "bbox": [ + 312, + 452, + 441, + 465 + ], + "type": "inline_equation", + "content": "\\hat{Y}^{u,lgt}" + }, + { + "bbox": [ + 312, + 452, + 441, + 465 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 312, + 465, + 533, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 465, + 533, + 475 + ], + "spans": [ + { + "bbox": [ + 312, + 465, + 533, + 475 + ], + "type": "text", + "content": "7: # Produce mix pseudo-labels according to the mask." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 312, + 475, + 477, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 475, + 477, + 489 + ], + "spans": [ + { + "bbox": [ + 312, + 475, + 477, + 489 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 312, + 475, + 477, + 489 + ], + "type": "inline_equation", + "content": "\\hat{Y}^u\\gets" + }, + { + "bbox": [ + 312, + 475, + 477, + 489 + ], + "type": "text", + "content": " mask_pseudo " + }, + { + "bbox": [ + 312, + 475, + 477, + 489 + ], + "type": "inline_equation", + "content": "(\\hat{Y}^{u,lgt},\\text{mask}_y)" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 514, + 382, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 382, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 382, + 526 + ], + "type": "text", + "content": "3.5. Discussions" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 305, + 533, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 533, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 305, + 533, + 545, + 556 + ], + "type": "text", + "content": "We discuss the relations between our framework and other most-related S4 frameworks." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 557, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 546, + 715 + ], + "type": "text", + "content": "Compare with PCR [51]. The key insight of PCR and our method lies in the promotion of consistency between the output in the logit and representation spaces. However, PCR exclusively relies on pseudo-labels derived from logit space as the guidance in the logit and representation spaces during the unsupervised training process, thereby neglecting the exploitation of the extensive semantic information inherent in representations and prototypes. In contrast, our method adopts a collaborative way by combining pseudo-labels obtained from both the logit and representation spaces, which improves the quality of pseudo-labels and enhances the knowledge exchange between the two spaces. In addition, since these two approaches are orthogonal, one can employ" + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "935" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "our pseudo-labeling and indicator strategies within the PCR framework, or alternatively incorporate multiple prototypes and consistency loss into our approach. Such combinations may yield further improvements in performance, however, also increase the computational complexity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 289 + ], + "type": "text", + "content": "Compare with CPS [7]. CPS contains two segmentation models initialized differently. It leverages pseudo-labels generated by one model to supervise the other, thereby maintaining the consistency between the output of the two models in the logit space. In a similar vein to CPS, our cross pseudo-labeling strategy in Sec. 3.3 is also motivated by the objective of maintaining consistency during unsupervised training. However, our method is distinct from CPS in that we focus on preserving consistency between the logit space and the representation space. This distinction confers a distinct advantage in terms of memory efficiency, as we solely introduce a MLP as the representation head, rather than introducing an additional segmentation model as in CPS." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 289, + 288, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 288, + 421 + ], + "type": "text", + "content": "Compare with other contrastive based S4 works [32, 1, 46, 23]. We adhere to prior works to build our prototype, describe in Eq. 2. Different from them, we update our prototypes iteratively, similar to [60]. Our sampling strategies are based on the threshold, similar to [32, 23, 45]. However, it is worth noting that our indicators which serve as the basis for comparison with the threshold, are derived from both the logit and representation spaces. This stands in contrast to prior approaches where indicators solely originate from the logit space. The efficacy of this modification is substantiated and extensively discussed in Sec. 5.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 434, + 128, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 128, + 447 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 128, + 447 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 454, + 97, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 97, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 97, + 467 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 473, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 288, + 689 + ], + "type": "text", + "content": "Datasets. We conduct experiments on PASCAL VOC 2012 dataset [12], Cityscapes dataset [9], ADE20K dataset [59], and COCO-Stuff 10K dataset [2] to validate the effectiveness of our proposed method. The original PASCAL VOC 2012 dataset contains 1,464 labeled images in train set and 1,449 images for validation in val set. Following [7, 37], we additionally introduce 9,118 images from SBD [18] as training images. Since the labels in SBD are coarsely annotated, following [46], we use both classic VOC train set (1,464 candidate labeled images) and blender VOC train set (10,582 candidate labeled images). Cityscapes dataset is a dataset for urban scene understanding, which contains 2,975 images in train set and 500 images in val set. ADE20K has 20,210 and 2,000 images in train and val set, with 150 classes in total. COCO-Stuff 10K has 9,000 images in train set and 1,000 images in val set, with 181 classes. The approach of preprocessing labels is followed by MMSegmentation [8]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Network structure. We use Deeplabv3+ [4] with ResNet-101 [20] pre-trained on ImageNet [10] as our network struc" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "ture. The segmentation and representation head are composed of Conv-BN-ReLU-Conv." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 96, + 546, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 546, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 546, + 337 + ], + "type": "text", + "content": "Implementation details. For training on PASCAL VOC 2012 dataset and COCO-Stuff 10K dataset, we set the learning rate as 0.0064, weight decay as 0.0005, crop size as 512 " + }, + { + "bbox": [ + 304, + 96, + 546, + 337 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 304, + 96, + 546, + 337 + ], + "type": "text", + "content": " 512, batch size as 16, and a total of 40,000 iterations. For training on the Cityscapes dataset, we set the learning rate as 0.0038, weight decay as 0.0005, crop size as " + }, + { + "bbox": [ + 304, + 96, + 546, + 337 + ], + "type": "inline_equation", + "content": "768 \\times" + }, + { + "bbox": [ + 304, + 96, + 546, + 337 + ], + "type": "text", + "content": " 768, batch size as 8, and a total of 80,000 iterations. For training on ADE20K dataset, we set the learning rate as 0.0064, weight decay as 0.0005, crop size as " + }, + { + "bbox": [ + 304, + 96, + 546, + 337 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 96, + 546, + 337 + ], + "type": "text", + "content": ", batch size as 16, and a total of 80,000 iterations. We use the poly scheduling to decay the learning rate during the training process: " + }, + { + "bbox": [ + 304, + 96, + 546, + 337 + ], + "type": "inline_equation", + "content": "lr = lr_{base} \\times (1 - \\frac{epoch}{total\\_epoch})^{0.9}" + }, + { + "bbox": [ + 304, + 96, + 546, + 337 + ], + "type": "text", + "content": ". We use the mean of Intersection over Union (mIoU) as the metric in evaluation. We use the sliding window strategy to evaluate the performance of our method on the Cityscapes dataset, following [7]. In addition, when adopting our cross-labeling strategy, due to the requirement of a set of high-quality prototypes when classifying each representation, we first solely use the supervision from logit space for 20 epochs to initialize the prototypes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 346, + 495, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 495, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 495, + 358 + ], + "type": "text", + "content": "4.2. Comparison with Existing Methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "text", + "content": "In this subsection, we first reproduce three baselines: MT [44], CutMix [15], and a same contrastive-based framework with us but with only logit space pseudo-labels and indicators (Baseline) on classic VOC train set. Meanwhile, we make the comparison of our method with mix (CSS (mix)) and cross (CSS (crs.)) pseudo-labeling strategy on blender VOC train set and Cityscapes train set with following recent SOTA S4 methods: CCT [38], CPS [7], U²PL [46], ST++ [52], PRCL [48], PCR [51], and PSMT [33]. Since the data split will dramatically affect the performance in S4, i.e., choosing labeled images plays an important role in the final results, we conduct experiments with three different data splits and report the mean value and standard deviation (blue color numbers). Since the mix pseudo-labeling strategy has better performance, we only use CSS (mix) when compared with SOTAs. Meanwhile, we use ResNet-101 with deep stem blocks as our network structure when compared with SOTAs. Since there is no uniform data split, we use the data splits in U²PL [46]. We use the OHEM loss when training Cityscapes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 664 + ], + "type": "text", + "content": "Results on PASCAL VOC 2012. Tab. 1 shows the comparison with our baselines on Classic PASCAL VOC 2012 set. Our method consistently outperforms baselines with an acceptable standard deviation on all label rates. Tab. 2 shows the comparison with the SOTAs on PASCAL VOC 2012." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 689 + ], + "type": "text", + "content": "Results on Cityscapes. Tab. 3 shows the performance of our method on Cityscapes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "type": "text", + "content": "Results on ADE20K and COCO-Stuff 10K. Tab. 4 shows the results of our method on ADE20K and COCO-Stuff." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "936" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 113, + 286, + 205 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 287, + 112 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 287, + 112 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 287, + 112 + ], + "type": "text", + "content": "Table 1. Results on classic VOC train set with four different label rate. Labeled data splits are from the original VOC train set. All approaches are reproduced with three runs and three different data splits." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 113, + 286, + 205 + ], + "lines": [ + { + "bbox": [ + 49, + 113, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 113, + 286, + 205 + ], + "type": "table", + "html": "
Pascal VOC 2012 (Classic)
Method92183366732
Sup.51.57±3.5854.69±2.4464.86±1.0470.77±0.76
MT58.92±2.9961.63±1.7666.79±0.5371.58±0.51
CutMix65.82±3.6067.91±1.4772.53±0.5074.08±0.49
Baseline66.91±4.2170.32±1.8673.97±1.0476.49±0.58
CSS(crs.)67.03±4.5871.41±1.6774.47±1.0877.08±0.37
CSS(mix)68.09±4.8971.93±1.8874.91±1.1277.57±0.73
", + "image_path": "bd9a2c5ea5cf427244589964771a74592fcaf2ff1121de0bb81f410ba58da557.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 247, + 285, + 358 + ], + "blocks": [ + { + "bbox": [ + 47, + 214, + 286, + 247 + ], + "lines": [ + { + "bbox": [ + 47, + 214, + 286, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 214, + 286, + 247 + ], + "type": "text", + "content": "Table 2. Results on blender VOC train set. All the results from the recent papers [46, 52, 13, 29, 33, 51]. Labeled data is from the augmented VOC train set and the data splits are from [46, 33]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 247, + 285, + 358 + ], + "lines": [ + { + "bbox": [ + 49, + 247, + 285, + 358 + ], + "spans": [ + { + "bbox": [ + 49, + 247, + 285, + 358 + ], + "type": "table", + "html": "
Pascal VOC 2012 (Blender)
Method662132326465291
CCT [38]71.8673.6876.5177.40
CPS [7]74.4876.4477.6878.64
U2PL [46]77.2179.0179.3080.50
ST++ [52]74.7077.9077.90-
PRCL [48]76.9678.1679.0279.59
PCR [51]78.6080.7180.7880.91
PSMT [33]75.5078.2078.7279.76
CSS (mix)78.7379.5480.8281.06
", + "image_path": "f12f4e0946f0045f47e675aba0f182275f9550f7b8d6dc81af541f4daaf59611.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 49, + 413, + 285, + 503 + ], + "blocks": [ + { + "bbox": [ + 46, + 369, + 287, + 413 + ], + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 413 + ], + "type": "text", + "content": "Table 3. Results on Cityscapes. The model is trained on the Cityscapes train set, which consists of 2,975 samples in total, and tested on Cityscapes val set. And all the results from the recent papers [46, 51, 33]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 413, + 285, + 503 + ], + "lines": [ + { + "bbox": [ + 49, + 413, + 285, + 503 + ], + "spans": [ + { + "bbox": [ + 49, + 413, + 285, + 503 + ], + "type": "table", + "html": "
Cityscapes
Method1863727441488
CCT [38]69.3274.1275.9978.10
CPS [7]69.7874.3174.5876.82
U2PL [46]70.3074.3776.4779.05
PCR [51]73.4176.3178.4079.11
PSMT [33]-76.8977.6079.09
CSS (mix)74.0276.9377.9479.62
", + "image_path": "e48f517d8521fc2662cf1f45eaf7a4bb3d4985a65647d5cddde92dbb9cd217aa.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 523, + 138, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 523, + 138, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 523, + 138, + 537 + ], + "type": "text", + "content": "5. Ablative Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 544, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 639 + ], + "type": "text", + "content": "The main contribution of our work lies in 1) collaborative pseudo-labeling strategies and 2) a new indicator for representation learning. To further prove the effectiveness of our proposed method, we conduct ablative studies on these two points. We choose Deeplabv3+ with ResNet-101 pre-trained on ImageNet as our backbone and leverage 92 labeled images and 183 labeled images in PASCAL VOC 2012. The other settings are the same as those in Sec. 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 647, + 285, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 285, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 285, + 659 + ], + "type": "text", + "content": "5.1. Effectiveness of Collaborative pseudo-labeling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Quality of pseudo-labels. To illustrate the superiority of using pseudo-labels from two spaces in a collaborative way as supervision, we conduct experiments to show the quality of pseudo-labels obtained 1) from logit space (lgt.), 2) from" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 347, + 109 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 347, + 109 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 347, + 109 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 347, + 109 + ], + "type": "image", + "image_path": "9ec7b34a9c8edda15bbb118fc8aad3b89f26be65e4f81c35fb50a8784b65bfc6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 347, + 72, + 386, + 148 + ], + "blocks": [ + { + "bbox": [ + 347, + 72, + 386, + 148 + ], + "lines": [ + { + "bbox": [ + 347, + 72, + 386, + 148 + ], + "spans": [ + { + "bbox": [ + 347, + 72, + 386, + 148 + ], + "type": "image", + "image_path": "885b6ec9ed1b4342c72e6b0a35d62753823bf6a702037312c0b75bcb4e800fde.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 150, + 376, + 159 + ], + "lines": [ + { + "bbox": [ + 347, + 150, + 376, + 159 + ], + "spans": [ + { + "bbox": [ + 347, + 150, + 376, + 159 + ], + "type": "text", + "content": "(b) GT" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 386, + 72, + 424, + 109 + ], + "blocks": [ + { + "bbox": [ + 386, + 72, + 424, + 109 + ], + "lines": [ + { + "bbox": [ + 386, + 72, + 424, + 109 + ], + "spans": [ + { + "bbox": [ + 386, + 72, + 424, + 109 + ], + "type": "image", + "image_path": "f53cf14542852989825603154b7a11bfa60e5a33b958812402bd582f13415e77.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 425, + 72, + 464, + 109 + ], + "blocks": [ + { + "bbox": [ + 425, + 72, + 464, + 109 + ], + "lines": [ + { + "bbox": [ + 425, + 72, + 464, + 109 + ], + "spans": [ + { + "bbox": [ + 425, + 72, + 464, + 109 + ], + "type": "image", + "image_path": "3bd1a658daae03d39c76792d3e531523ef9aba32cfb9563c71d41f96e2723fcd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 425, + 110, + 464, + 148 + ], + "blocks": [ + { + "bbox": [ + 425, + 110, + 464, + 148 + ], + "lines": [ + { + "bbox": [ + 425, + 110, + 464, + 148 + ], + "spans": [ + { + "bbox": [ + 425, + 110, + 464, + 148 + ], + "type": "image", + "image_path": "2101bf8846b2951f47194ea6ee2e90ba11e34bf91205d42f088d7fb709551c9e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 150, + 462, + 158 + ], + "lines": [ + { + "bbox": [ + 425, + 150, + 462, + 158 + ], + "spans": [ + { + "bbox": [ + 425, + 150, + 462, + 158 + ], + "type": "text", + "content": "(d) lgt. label" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 164, + 538, + 175 + ], + "lines": [ + { + "bbox": [ + 306, + 164, + 538, + 175 + ], + "spans": [ + { + "bbox": [ + 306, + 164, + 538, + 175 + ], + "type": "text", + "content": "Figure 3. Differences between pseudo-labels in different spaces." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 466, + 72, + 503, + 109 + ], + "blocks": [ + { + "bbox": [ + 466, + 72, + 503, + 109 + ], + "lines": [ + { + "bbox": [ + 466, + 72, + 503, + 109 + ], + "spans": [ + { + "bbox": [ + 466, + 72, + 503, + 109 + ], + "type": "image", + "image_path": "3311a85ac4b5d87c7ad39fbef4c91b440d5bb9a920abb2bace77c69987926f12.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 466, + 110, + 504, + 148 + ], + "blocks": [ + { + "bbox": [ + 466, + 110, + 504, + 148 + ], + "lines": [ + { + "bbox": [ + 466, + 110, + 504, + 148 + ], + "spans": [ + { + "bbox": [ + 466, + 110, + 504, + 148 + ], + "type": "image", + "image_path": "de9a6e7cf32de80262c4629e83a94e3f231d7f7965cfc96a2b7ec931a5a5022f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 150, + 504, + 158 + ], + "lines": [ + { + "bbox": [ + 466, + 150, + 504, + 158 + ], + "spans": [ + { + "bbox": [ + 466, + 150, + 504, + 158 + ], + "type": "text", + "content": "(e) rep. mask" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 506, + 72, + 544, + 109 + ], + "blocks": [ + { + "bbox": [ + 506, + 72, + 544, + 109 + ], + "lines": [ + { + "bbox": [ + 506, + 72, + 544, + 109 + ], + "spans": [ + { + "bbox": [ + 506, + 72, + 544, + 109 + ], + "type": "image", + "image_path": "7fbcb16361f71a321b1f1851c62aef74216a3d53756d95c792d51d6eff4f5f3d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 506, + 110, + 544, + 148 + ], + "blocks": [ + { + "bbox": [ + 506, + 110, + 544, + 148 + ], + "lines": [ + { + "bbox": [ + 506, + 110, + 544, + 148 + ], + "spans": [ + { + "bbox": [ + 506, + 110, + 544, + 148 + ], + "type": "image", + "image_path": "0c9a2c8b867d5390141cce1649bf2e365f208bc51a70272403e5b3aa16c997ef.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 150, + 542, + 158 + ], + "lines": [ + { + "bbox": [ + 507, + 150, + 542, + 158 + ], + "spans": [ + { + "bbox": [ + 507, + 150, + 542, + 158 + ], + "type": "text", + "content": "(f) rep. label" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 198, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 327 + ], + "type": "text", + "content": "representation space (rep.), and 3) from the mix pseudolabeling strategy (mix). The pseudo-labels are sampled with corresponding sampling strategies in Sec. 3.4. Tab. 5 illustrates the IoU of pseudo-labels for each class on PASCAL VOC 2012 with 92 labeled images. The results clearly indicate that employing pseudo-labels from the representation space enhances the accuracy of the final pseudo-labels in most classes. This improvement is particularly evident in classes that are originally under-performing, such as the IoU improvement of " + }, + { + "bbox": [ + 304, + 198, + 545, + 327 + ], + "type": "inline_equation", + "content": "11.42\\%" + }, + { + "bbox": [ + 304, + 198, + 545, + 327 + ], + "type": "text", + "content": " for the chair class and " + }, + { + "bbox": [ + 304, + 198, + 545, + 327 + ], + "type": "inline_equation", + "content": "11.58\\%" + }, + { + "bbox": [ + 304, + 198, + 545, + 327 + ], + "type": "text", + "content": " for the sofa class." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 330, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 545, + 521 + ], + "type": "text", + "content": "Meanwhile, we also visualize the pseudo-labels obtained from logit space (lgt.) and representation space (rep.) in Fig. 3. Fig. 3 (c) and (e) show the masks for pseudolabels produced by the sampling strategies. In particular, the white color represents the valid pixels used during unlabeled learning, while the black color indicates the discarded pixels. Fig. 3 (d) and (f) are the pseudo-labels we obtained from two spaces. The figure clearly illustrates the differences between pseudo-labels produced by different spaces. For example, the parts of the instance edge in pseudo-labels are usually discarded since they are challenging for learning in logit space. However, pseudo-labels from representation space will easily tackle this problem (first row). The pseudo-labels from representation space are more inaccurate in some complex scenes, which be resolved by combining pseudo-labels from logit space (second row)." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "We mainly attribute the differences in pseudo-label to the differing concentrations of learning in the two spaces. Specifically, learning in the logit space primarily emphasizes the most discriminative part of features, whereas that in the representation space treats each part of features equally. As a result, learning in the logit space may overlook minor feature differences, leading to sub-optimal performance in predicting instance edges and distinguishing between similar classes (e.g., chair and sofa). Conversely, learning in the representation space produces balanced performance across all image parts and classes. However, this can lead to erroneous predictions for classes with high intra-class variance (e.g., background). By leveraging pseudo-labels from the two spaces, we capitalize on the strengths of the learning in each space and enhance the knowledge exchange between the two spaces." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "937" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 80, + 545, + 148 + ], + "blocks": [ + { + "bbox": [ + 108, + 69, + 484, + 79 + ], + "lines": [ + { + "bbox": [ + 108, + 69, + 484, + 79 + ], + "spans": [ + { + "bbox": [ + 108, + 69, + 484, + 79 + ], + "type": "text", + "content": "Table 4. Results on ADE20K and COCO-Stuff 10K with four different label rates and single data splits." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 80, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 49, + 80, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 49, + 80, + 545, + 148 + ], + "type": "table", + "html": "
ADE20KCOCO-Stuff 10K
Method1/161/81/41/21/161/81/41/2
Cutmix28.9131.8934.6239.1523.0926.1429.8830.91
Baseline31.1133.2436.4639.9126.5727.9128.9731.08
CSS (crs.)32.0133.9237.8439.9727.0628.8530.5931.73
CSS (mix)32.5534.2137.0140.8527.4629.5129.9831.91
", + "image_path": "230773c88f7d9bc98d41e4e8bafbf593a794fa97ce9c78ee832021623c76db00.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 169, + 545, + 261 + ], + "blocks": [ + { + "bbox": [ + 47, + 159, + 531, + 169 + ], + "lines": [ + { + "bbox": [ + 47, + 159, + 531, + 169 + ], + "spans": [ + { + "bbox": [ + 47, + 159, + 531, + 169 + ], + "type": "text", + "content": "Table 5. The quality of pseudo-labels from different pseudo-labeling strategies. The pseudo-labels are sampled by sampling strategies." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 169, + 545, + 261 + ], + "lines": [ + { + "bbox": [ + 49, + 169, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 49, + 169, + 545, + 261 + ], + "type": "table", + "html": "
sourcebackaero.bicy.birdboatbott.buscarcatchaircow
lgt.96.7895.1477.4793.3881.3787.5496.7695.4894.474.0992.15
rep.90.7196.5061.4275.7553.3054.6584.4680.5591.1128.0388.16
mix96.66↓0.1298.12↑2.9882.76↑5.2994.47↑1.0985.12↑3.7589.94↑2.4097.03↑0.2795.98↑0.5094.65↑0.1819.01↑14.9294.81↑2.66
sourcetabeldoghorsemotorpers.plantsheepsofatraintvmIoU
lgt.58.3494.1293.0191.3793.6850.3391.1018.1686.6564.8978.87
rep.52.2386.9773.2288.7091.7556.1366.7835.2588.5162.6171.57
mix64.35↑6.0194.33↑0.2193.65↑0.6491.50↑0.1393.01↓0.6755.64↑5.3191.56↑0.4629.74↑11.5893.54↑6.8969.48↑4.5982.17↑3.33
", + "image_path": "65b884998334dfe4b377e7bb07c9be9e444b9401c620a45999ce72efde4772b2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 299, + 286, + 357 + ], + "blocks": [ + { + "bbox": [ + 47, + 279, + 286, + 299 + ], + "lines": [ + { + "bbox": [ + 47, + 279, + 286, + 299 + ], + "spans": [ + { + "bbox": [ + 47, + 279, + 286, + 299 + ], + "type": "text", + "content": "Table 6. Results on pseudo-labels from different sources on two different label rates." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 299, + 286, + 357 + ], + "lines": [ + { + "bbox": [ + 48, + 299, + 286, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 286, + 357 + ], + "type": "table", + "html": "
source92 labels183 labels
logit space67.1170.32
representation space64.2067.52
mix pseudo-labeling68.41↑1.3072.74↑2.42
cross pseudo-labeling67.85↑0.7471.98↑1.66
", + "image_path": "cc1af88b0a1528668fa93525f0174368487d95d601cc419cc2e4a68e11c8e349.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 376, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 287, + 604 + ], + "type": "text", + "content": "Results of different strategies To investigate the involvement of different pseudo-labeling strategies, we conduct the experiments as follows: 1) Using pseudo-labels from logit space. 2) Using pseudo-labels from representation space. 3) Using mix pseudo-labeling strategy. 4) Using cross pseudolabeling strategy. Tab. 6 shows the effectiveness of our proposed strategy in two different label rates. The results show that the performances of experiments with collaborative pseudo-labeling strategies are better than the ones whose pseudo-labels come from a single space with two different label rates, which proves the effectiveness of our proposed collaboration between the two spaces. It is worth noting that even though the quality of pseudo-labels from representation space is lower than that from logit space, the performance of the model is also boosted by using the cross pseudo-labeling strategy to maintain consistency between the predictions in two spaces. In addition, our method with the mix pseudo-labeling strategy outperforms that with the cross pseudo-labeling strategy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 611, + 206, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 206, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 206, + 623 + ], + "type": "text", + "content": "5.2. Effectiveness of the Indicator" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Limitation of merely using confidence. To explain the limitation of merely using confidence for learning in the logit and representation spaces, we conduct experiments to show the relations between confidence and similarity. The similarity is the cosine similarity between representations and prototypes, which directly shows the confusion level between the representation and the prototype of its class." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 283, + 415, + 354 + ], + "blocks": [ + { + "bbox": [ + 307, + 283, + 415, + 354 + ], + "lines": [ + { + "bbox": [ + 307, + 283, + 415, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 415, + 354 + ], + "type": "image", + "image_path": "b808f15724ab1f45e36ba846846ae5ac6fdbccf2c1e21314c20fd4b0b040d92f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 354, + 368, + 361 + ], + "lines": [ + { + "bbox": [ + 358, + 354, + 368, + 361 + ], + "spans": [ + { + "bbox": [ + 358, + 354, + 368, + 361 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 324, + 368, + 523, + 379 + ], + "lines": [ + { + "bbox": [ + 324, + 368, + 523, + 379 + ], + "spans": [ + { + "bbox": [ + 324, + 368, + 523, + 379 + ], + "type": "text", + "content": "Figure 4. Relations between similarity and confidence." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 416, + 284, + 544, + 354 + ], + "blocks": [ + { + "bbox": [ + 416, + 284, + 544, + 354 + ], + "lines": [ + { + "bbox": [ + 416, + 284, + 544, + 354 + ], + "spans": [ + { + "bbox": [ + 416, + 284, + 544, + 354 + ], + "type": "image", + "image_path": "d4e6d66393eb772fd8063bec986ad111d15dd70005f4e8f98ccee0b69d3f013b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 354, + 479, + 363 + ], + "lines": [ + { + "bbox": [ + 469, + 354, + 479, + 363 + ], + "spans": [ + { + "bbox": [ + 469, + 354, + 479, + 363 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 400, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 545, + 579 + ], + "type": "text", + "content": "Fig. 4 (a) shows the comparison between the confidence of each prediction and the corresponding similarity. We use the class person for demonstrating in our experiments. It shows clearly that even though fixing the confidence into a small range (from 0.8 to 0.81 in our settings), the similarity varies. Meanwhile, in Fig. 4 (b), different color bars stand for different intervals of confidence, and the lines denote the mean similarity between the prototype and each representation whose corresponding confidence is in the current interval. Fig. 4 (b) illustrates that the mean similarity of the class fluctuates when its interval of confidence rises. Both two figures imply that confidence is not able to represent the confusion level between representations and prototypes since there are no direct and close relations between confidence and similarity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 580, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 628 + ], + "type": "text", + "content": "Fig. 5 visualizes the similarity and confidence of an image in both logit (lgt.) and representation (rep.) spaces, indicating the varying levels of confusion in the same region when learning in different spaces." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 629, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 664 + ], + "type": "text", + "content": "We also attribute it to the different concentrations of learning two spaces, i.e., the confusing region in one space can be more readily addressed in the other space." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Thus, it is inappropriate to choose confidence as the indicator to involve representation learning, e.g., sampling more hard samples with threshold and indicator. In contrast, our indicator directly employs similarity between the represen" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "938" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 89, + 107, + 148 + ], + "blocks": [ + { + "bbox": [ + 51, + 75, + 67, + 82 + ], + "lines": [ + { + "bbox": [ + 51, + 75, + 67, + 82 + ], + "spans": [ + { + "bbox": [ + 51, + 75, + 67, + 82 + ], + "type": "text", + "content": "high" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 89, + 107, + 148 + ], + "lines": [ + { + "bbox": [ + 49, + 89, + 107, + 148 + ], + "spans": [ + { + "bbox": [ + 49, + 89, + 107, + 148 + ], + "type": "image", + "image_path": "afbb859f0b9726ddd8dd67ba3373e08a7863d8d794adf298f134bddb32e33570.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 152, + 94, + 160 + ], + "lines": [ + { + "bbox": [ + 63, + 152, + 94, + 160 + ], + "spans": [ + { + "bbox": [ + 63, + 152, + 94, + 160 + ], + "type": "text", + "content": "(a) Image" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 170, + 279, + 182 + ], + "lines": [ + { + "bbox": [ + 47, + 170, + 279, + 182 + ], + "spans": [ + { + "bbox": [ + 47, + 170, + 279, + 182 + ], + "type": "text", + "content": "Figure 5. Visualization of the confusing part in different spaces." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 109, + 91, + 166, + 148 + ], + "blocks": [ + { + "bbox": [ + 109, + 91, + 166, + 148 + ], + "lines": [ + { + "bbox": [ + 109, + 91, + 166, + 148 + ], + "spans": [ + { + "bbox": [ + 109, + 91, + 166, + 148 + ], + "type": "image", + "image_path": "7c004d2fc7084f03c4164fb131f750be95def112c593c4cadd15172f5de04802.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 152, + 148, + 160 + ], + "lines": [ + { + "bbox": [ + 126, + 152, + 148, + 160 + ], + "spans": [ + { + "bbox": [ + 126, + 152, + 148, + 160 + ], + "type": "text", + "content": "(b) GT" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 167, + 91, + 224, + 148 + ], + "blocks": [ + { + "bbox": [ + 167, + 91, + 224, + 148 + ], + "lines": [ + { + "bbox": [ + 167, + 91, + 224, + 148 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 224, + 148 + ], + "type": "image", + "image_path": "6e11740b2e78d6466f069457b1c094909b6eee02dbd9ecf3470bfe3b2de4de31.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 152, + 217, + 161 + ], + "lines": [ + { + "bbox": [ + 176, + 152, + 217, + 161 + ], + "spans": [ + { + "bbox": [ + 176, + 152, + 217, + 161 + ], + "type": "text", + "content": "(c) lgt. space" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 225, + 91, + 285, + 148 + ], + "blocks": [ + { + "bbox": [ + 225, + 91, + 285, + 148 + ], + "lines": [ + { + "bbox": [ + 225, + 91, + 285, + 148 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 285, + 148 + ], + "type": "image", + "image_path": "692662af87d13f12ef88a56656a8888fa45fb2d2e7820fe9a4ac44e5f98eb755.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 152, + 277, + 161 + ], + "lines": [ + { + "bbox": [ + 233, + 152, + 277, + 161 + ], + "spans": [ + { + "bbox": [ + 233, + 152, + 277, + 161 + ], + "type": "text", + "content": "(c) rep. space" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 201, + 286, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 286, + 249 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 286, + 249 + ], + "type": "text", + "content": "tation and prototype of its class, which directly reflects the confusion level in representation learning. It is more accurate to use similarity as the indicator to sample hard and critical samples in representation learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 249, + 286, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 249, + 286, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 286, + 381 + ], + "type": "text", + "content": "Results of different indicators. Tab. 7 shows the impact of using different indicators. We conduct experiments on two different label rates (92 and 183) and three different indicators for pseudo-labels: only confidence (conf.), only similarity (smlr.), and confidence for learning logit space while similarity for learning representation space (mix). We use two different approaches to obtain pseudo-labels: from logit space only (seg label) and from mix pseudo-labeling strategy (mix label). It is clear that using both confidence and similarity to involve the learning in their own spaces obtains the best performance." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 48, + 410, + 286, + 468 + ], + "blocks": [ + { + "bbox": [ + 47, + 390, + 286, + 410 + ], + "lines": [ + { + "bbox": [ + 47, + 390, + 286, + 410 + ], + "spans": [ + { + "bbox": [ + 47, + 390, + 286, + 410 + ], + "type": "text", + "content": "Table 7. Results on indicators from different spaces on two different label rates." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 410, + 286, + 468 + ], + "lines": [ + { + "bbox": [ + 48, + 410, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 286, + 468 + ], + "type": "table", + "html": "
92 labels183 labels
sourceseg labelmix labelseg labelmix label
conf.67.1168.3370.3271.92
smlr.66.1666.8968.9069.25
mix67.80↑0.6968.41↑1.3071.50↑1.1872.74↑2.42
", + "image_path": "04b71f1e4c0077e41db1af48f3d8f649b8a5e862424f976d0715f10b250eb1a8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 489, + 212, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 212, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 212, + 502 + ], + "type": "text", + "content": "5.3. Ablation study of Components" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 508, + 286, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 286, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 286, + 639 + ], + "type": "text", + "content": "In this section, we conduct experiments to introduce our components in CSS step by step, with results shown in Tab. 8. Our baseline is the conventional contrastive-based S4, achieving mIoU of " + }, + { + "bbox": [ + 46, + 508, + 286, + 639 + ], + "type": "inline_equation", + "content": "67.11\\%" + }, + { + "bbox": [ + 46, + 508, + 286, + 639 + ], + "type": "text", + "content": " on 92 labels and " + }, + { + "bbox": [ + 46, + 508, + 286, + 639 + ], + "type": "inline_equation", + "content": "70.32\\%" + }, + { + "bbox": [ + 46, + 508, + 286, + 639 + ], + "type": "text", + "content": " on 183 labels. Mix and cross means the pseudo-labels are from the mix pseudo-labeling strategy and cross pseudo-labeling strategy while the indicator is still the confidence in two spaces. Ind means we use the different indicators in different spaces while the pseudo-labels are from logit space. The last two rows represent our proposed two pseudo-labeling strategies with indicators from two spaces." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 647, + 159, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 159, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 159, + 659 + ], + "type": "text", + "content": "5.4. Qualitative Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Fig. 6 shows the qualitative results of different methods on PASCAL VOC 2012 with 92 labeled images. Baseline means the conventional contrastive-based method. Compared with the original self-training methods (CutMix)," + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 308, + 79, + 544, + 159 + ], + "blocks": [ + { + "bbox": [ + 306, + 69, + 527, + 79 + ], + "lines": [ + { + "bbox": [ + 306, + 69, + 527, + 79 + ], + "spans": [ + { + "bbox": [ + 306, + 69, + 527, + 79 + ], + "type": "text", + "content": "Table 8. Ablation study on different components of our CSS." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 79, + 544, + 159 + ], + "lines": [ + { + "bbox": [ + 308, + 79, + 544, + 159 + ], + "spans": [ + { + "bbox": [ + 308, + 79, + 544, + 159 + ], + "type": "table", + "html": "
component92 labels183 labels
baseline67.1170.32
mix68.33↑1.2271.92↑1.60
cross67.21↑0.1070.87↑0.55
ind67.80↑0.6971.50↑1.18
mix + ind68.41↑1.3072.74↑2.42
cross + ind67.85↑0.7471.98↑1.66
", + "image_path": "47d7024d5d72288596e8438e9812b8c51186e8af4abaa69f8eeab86b859fd826.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 181, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 181, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 545, + 240 + ], + "type": "text", + "content": "thanks to introducing pixel-wise contrastive learning, the baseline, and our method perform better in some ambiguous regions. Furthermore, benefiting from the supervision of two spaces and different indicators in different spaces, our method outperforms the baseline." + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 308, + 251, + 544, + 417 + ], + "blocks": [ + { + "bbox": [ + 308, + 251, + 544, + 417 + ], + "lines": [ + { + "bbox": [ + 308, + 251, + 544, + 417 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 544, + 417 + ], + "type": "image", + "image_path": "0087841697414687131256fde037c71be6289bf332eac50c0f368691b3531be0.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 430, + 545, + 442 + ], + "lines": [ + { + "bbox": [ + 306, + 430, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 545, + 442 + ], + "type": "text", + "content": "images. Yellow boxes highlight the main differences." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 464, + 378, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 378, + 477 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 378, + 477 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 485, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 604 + ], + "type": "text", + "content": "In this paper, we propose two collaborative pseudolabeling strategies to take full use of the semantic information in the representation space and enhance the knowledge exchange between the logit and representation spaces. Moreover, we employ a new indicator for the learning process in the representation space. Extensive experiments demonstrate that our pseudo-labeling strategies obtain more reliable supervision during unlabeled training and our indicator helps the model to concentrate on more critical parts during representation learning." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "text", + "content": "Future work: In this paper, we employ pseudo-labeling strategies to utilize the semantic information in both logit and representation spaces. In the future, we will investigate more powerful strategies to enhance the knowledge exchange between two spaces." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "939" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Inigo Alonso, Alberto Sabater, David Ferstl, Luis Montesano, and Ana C. Murillo. Semi-supervised semantic segmentation with pixel-level contrastive learning from a class-wise memory bank. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8219-8228, October 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "type": "text", + "content": "[2] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 287, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 287, + 256 + ], + "type": "text", + "content": "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "type": "text", + "content": "[4] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 313, + 287, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 287, + 356 + ], + "type": "text", + "content": "[5] Mu Chen, Zhedong Zheng, Yi Yang, and Tat-Seng Chua. Pipa: Pixel-and patch-wise self-supervised learning for domain adaptive semantic segmentation. arXiv preprint arXiv:2211.07609, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 402, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 287, + 456 + ], + "type": "text", + "content": "[7] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 458, + 286, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 286, + 501 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 286, + 501 + ], + "type": "text", + "content": "[8] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 502, + 286, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 286, + 568 + ], + "type": "text", + "content": "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 286, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 635 + ], + "type": "text", + "content": "[11] Nanqing Dong and Eric P Xing. Few-shot semantic segmentation with prototype learning. In BMVC, volume 3, 2018." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "text", + "content": "[12] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[13] Jiashuo Fan, Bin Gao, Huan Jin, and Lihui Jiang. Ucc: Uncertainty guided cross-head co-training for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), pages 9947-9956, June 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "type": "text", + "content": "[14] Zhengyang Feng, Qianyu Zhou, Qiqi Gu, Xin Tan, Guangliang Cheng, Xuequan Lu, Jianping Shi, and Lizhuang Ma. Dmt: Dynamic mutual training for semi-supervised learning. Pattern Recognition, page 108777, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "text", + "content": "[15] Geoff French, Timo Aila, Samuli Laine, Michal Mackiewicz, and Graham Finlayson. Semi-supervised semantic segmentation needs strong, high-dimensional perturbations. In In 31th British Machine Vision Conference, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "text", + "content": "[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "text", + "content": "[17] Dayan Guan, Jiaxing Huang, Aoran Xiao, and Shijian Lu. Unbiased subclass regularization for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9968-9978, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 283, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 327 + ], + "type": "text", + "content": "[18] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In 2011 international conference on computer vision, pages 991-998. IEEE, 2011." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "text", + "content": "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 427, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 545, + 480 + ], + "type": "text", + "content": "[21] Ruifei He, Jihan Yang, and Xiaojuan Qi. Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6930-6940, October 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "text", + "content": "[22] Judy Hoffman, Dequan Wang, Fisher Yu, and Trevor Darrell. Fcts in the wild: Pixel-level adversarial and constraint-based adaptation. arXiv preprint arXiv:1612.02649, 2016." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "text", + "content": "[23] Hanzhe Hu, Jinshi Cui, and Liwei Wang. Region-aware contrastive learning for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16291-16301, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "text", + "content": "[24] Hanzhe Hu, Fangyun Wei, Han Hu, Qiwei Ye, Jinshi Cui, and Liwei Wang. Semi-supervised semantic segmentation via adaptive equalization learning. Advances in Neural Information Processing Systems, 34:22106-22118, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 604, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 646 + ], + "type": "text", + "content": "[25] Wei-Chih Hung, Yi-Hsuan Tsai, Yan-Ting Liou, Yen-Yu Lin, and Ming-Hsuan Yang. Adversarial learning for semi-supervised semantic segmentation. arXiv preprint arXiv:1802.07934, 2018." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 647, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 713 + ], + "type": "text", + "content": "[26] Zhengkai Jiang, Yuxi Li, Ceyuan Yang, Peng Gao, Yabiao Wang, Ying Tai, and Chengjie Wang. Prototypical contrast adaptation for domain adaptive semantic segmentation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 36-54. Springer, 2022." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "940" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[27] Tarun Kalluri, Girish Varma, Manmohan Chandraker, and CV Jawahar. Universal semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5259-5270, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "type": "text", + "content": "[28] Rihuan Ke, Angelica I Aviles-Rivero, Saurabh Pandey, Saikumar Reddy, and Carola-Bibiane Schonlieb. A three-stage self-training framework for semi-supervised semantic segmentation. IEEE Transactions on Image Processing, 31:1805-1815, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "type": "text", + "content": "[29] Donghyeon Kwon and Suha Kwak. Semi-supervised semantic segmentation with error localization network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9957-9967, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 281 + ], + "type": "text", + "content": "[30] Daiqing Li, Junlin Yang, Karsten Kreis, Antonio Torralba, and Sanja Fidler. Semantic segmentation with generative models: Semi-supervised learning and strong out-of-domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8300-8311, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 283, + 288, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 288, + 336 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 288, + 336 + ], + "type": "text", + "content": "[31] Jinlu Liu, Liang Song, and Yongqiang Qin. Prototype rectification for few-shot learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16, pages 741–756. Springer, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "type": "text", + "content": "[32] Shikun Liu, Shuaifeng Zhi, Edward Johns, and Andrew Davison. Bootstrapping semantic segmentation with regional contrast. In International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 383, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 288, + 446 + ], + "type": "text", + "content": "[33] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4258-4267, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 449, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 503 + ], + "type": "text", + "content": "[34] Binjie Mao, Xinbang Zhang, Lingfeng Wang, Qian Zhang, Shiming Xiang, and Chunhong Pan. Learning from the target: Dual prototype network for few shot semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1953-1961, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 504, + 288, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 288, + 547 + ], + "type": "text", + "content": "[35] Sudhanshu Mittal, Maxim Tatarchenko, and Thomas Brox. Semi-supervised semantic segmentation with high-and low-level consistency. IEEE transactions on pattern analysis and machine intelligence, 43(4):1369-1379, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 548, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 601 + ], + "type": "text", + "content": "[36] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979-1993, 2018." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 603, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 288, + 658 + ], + "type": "text", + "content": "[37] Viktor Olsson, Wilhelm Tranheden, Juliano Pinto, and Lennart Svensson. Classmix: Segmentation-based data augmentation for semi-supervised learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1369-1378, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[38] Yassine Ouali, Celine Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross-consistency training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "text", + "content": "[39] Yingwei Pan, Ting Yao, Yehao Li, Yu Wang, Chong-Wah Ngo, and Tao Mei. Transferrable prototypical networks for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2239-2247, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 546, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 161 + ], + "type": "text", + "content": "[40] Jizong Peng, Guillermo Estrada, Marco Pedersoli, and Christian Desrosiers. Deep co-training for semi-supervised image segmentation. Pattern Recognition, 107:107269, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 162, + 546, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 546, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 546, + 194 + ], + "type": "text", + "content": "[41] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 195, + 546, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 546, + 258 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 546, + 258 + ], + "type": "text", + "content": "[42] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 261, + 546, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 546, + 315 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 546, + 315 + ], + "type": "text", + "content": "[43] Korawat Tanwisuth, Xinjie Fan, Huangjie Zheng, Shujian Zhang, Hao Zhang, Bo Chen, and Mingyuan Zhou. A prototype-oriented framework for unsupervised domain adaptation. Advances in Neural Information Processing Systems, 34:17194-17208, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 316, + 546, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 546, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 546, + 360 + ], + "type": "text", + "content": "[44] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 361, + 546, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 546, + 415 + ], + "type": "text", + "content": "[45] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7303-7313, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 416, + 546, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 546, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 546, + 480 + ], + "type": "text", + "content": "[46] Yuchao Wang, Haochen Wang, Yujun Shen, Jingjing Fei, Wei Li, Guoqiang Jin, Liwei Wu, Rui Zhao, and Xinyi Le. Semi-supervised semantic segmentation using unreliable pseudo labels. In Proceedings of the IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 482, + 546, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 546, + 546 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 546, + 546 + ], + "type": "text", + "content": "[47] Chen Wei, Kihyuk Sohn, Clayton Mellina, Alan Yuille, and Fan Yang. Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10857-10866, June 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 548, + 546, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 546, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 546, + 602 + ], + "type": "text", + "content": "[48] Haoyu Xie, Changqi Wang, Mingkai Zheng, Minjing Dong, Shan You, and Chang Xu. Boosting semi-supervised semantic segmentation with probabilistic representations. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2938-2946, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 603, + 546, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 546, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 546, + 657 + ], + "type": "text", + "content": "[49] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 658, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 546, + 712 + ], + "type": "text", + "content": "[50] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "941" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 664 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[51] Hai-Ming Xu, Lingqiao Liu, Qiuchen Bian, and Zhen Yang. Semi-supervised semantic segmentation with prototype-based consistency regularization. Advances in Neural Information Processing Systems, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "text", + "content": "[52] Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. St++: Make self-training work better for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4268-4277, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "type": "text", + "content": "[53] Jianlong Yuan, Yifan Liu, Chunhua Shen, Zhibin Wang, and Hao Li. A simple baseline for semi-supervised semantic segmentation with strong data augmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8229-8238, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 295 + ], + "type": "text", + "content": "[54] Xiangyu Yue, Zangwei Zheng, Shanghang Zhang, Yang Gao, Trevor Darrell, Kurt Keutzer, and Alberto Sangiovanni Vincentelli. Prototypical cross-domain self-supervised learning for few-shot unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13834-13844, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 297, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 361 + ], + "type": "text", + "content": "[55] Xiangyun Zhao, Raviteja Vemulapalli, Philip Andrew Mansfield, Boqing Gong, Bradley Green, Lior Shapira, and Ying Wu. Contrastive learning for label efficient semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10623-10633, October 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 363, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 363, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 363, + 287, + 396 + ], + "type": "text", + "content": "[56] Xu Zheng, Yunhao Luo, Hao Wang, Chong Fu, and Lin Wang. Transformer-cnn cohort: Semi-supervised semantic segmentation by the best of both students, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 397, + 287, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 397, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 48, + 397, + 287, + 462 + ], + "type": "text", + "content": "[57] Xu Zheng, Jinjing Zhu, Yexin Liu, Zidong Cao, Chong Fu, and Lin Wang. Both style and distortion matter: Dualpath unsupervised domain adaptation for panoramic semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1285-1295, June 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 464, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 464, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 464, + 287, + 518 + ], + "type": "text", + "content": "[58] Yuanyi Zhong, Bodi Yuan, Hong Wu, Zhiqiang Yuan, Jian Peng, and Yu-Xiong Wang. Pixel contrastive-consistent semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7273-7282, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 520, + 287, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 573 + ], + "type": "text", + "content": "[59] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 575, + 287, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 575, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 48, + 575, + 287, + 608 + ], + "type": "text", + "content": "[60] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A prototype view. In CVPR, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 609, + 287, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 609, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 48, + 609, + 287, + 664 + ], + "type": "text", + "content": "[61] Yanning Zhou, Hang Xu, Wei Zhang, Bin Gao, and Pheng-Ann Heng. C3-semiseg: Contrastive semi-supervised segmentation via cross-set learning and dynamic classbalancing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7036-7045, 2021." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "942" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Space-time Prompting for Video Class-incremental Learning/23f1a364-1b5a-487e-9a38-2d1ef0112aa7_content_list.json b/2023/Space-time Prompting for Video Class-incremental Learning/23f1a364-1b5a-487e-9a38-2d1ef0112aa7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a76ed0dc51d3ca9680859532daaf4d4af953fa2f --- /dev/null +++ b/2023/Space-time Prompting for Video Class-incremental Learning/23f1a364-1b5a-487e-9a38-2d1ef0112aa7_content_list.json @@ -0,0 +1,1726 @@ +[ + { + "type": "text", + "text": "Space-time Prompting for Video Class-incremental Learning", + "text_level": 1, + "bbox": [ + 178, + 130, + 794, + 155 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yixuan Pei $^{1}$ Zhiwu Qing $^{2}$ Shiwei Zhang $^{3*}$ Xiang Wang $^{2}$ Yingya Zhang $^{3}$ Deli Zhao $^{3}$ Xueming Qian $^{1,4*}$", + "bbox": [ + 232, + 179, + 736, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Xi'an Jiaotong University", + "bbox": [ + 377, + 233, + 591, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Huazhong University of Science and Technology $^{3}$ Alibaba Group, Shaanxi Yulan Jiuzhou Intelligent Optoelectronic Technology Co., Ltd.", + "bbox": [ + 210, + 251, + 756, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{peiiyixuan@stu, qianxm@mail}.xjtu.edu.cn {qzw,wxiang}@hust.edu.cn", + "bbox": [ + 192, + 289, + 777, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhangjin.zsw, yingya.zyy}@alibaba-inc.com zhaodeli@gmail.com", + "bbox": [ + 212, + 306, + 756, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 356, + 313, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, prompt-based learning has made impressive progress on image class-incremental learning, but it still lacks sufficient exploration in the video domain. In this paper, we will fill this gap by learning multiple prompts based on a powerful image-language pre-trained model, i.e., CLIP, making it fit for video class-incremental learning (VCIL). For this purpose, we present a space-time prompting approach (ST-Prompt) which contains two kinds of prompts, i.e., task-specific prompts and task-agnostic prompts. The task-specific prompts are to address the catastrophic forgetting problem by learning multi-grained prompts, i.e., spatial prompts, temporal prompts and comprehensive prompts, for accurate task identification. The task-agnostic prompts maintain a globally-shared prompt pool, which can empower the pre-trained image models with temporal perception abilities by exchanging contexts between frames. By this means, ST-Prompt can transfer the plentiful knowledge in the image-language pre-trained models to the VCIL task with only a tiny set of prompts to be optimized. To evaluate ST-Prompt, we conduct extensive experiments on three standard benchmarks. The results show that ST-Prompt can significantly surpass the state-of-the-art VCIL methods, especially it gains $9.06\\%$ on HMDB51 dataset under the $1 \\times 25$ stage setting.", + "bbox": [ + 75, + 388, + 473, + 751 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 777, + 209, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video understanding [49, 11, 1, 32] has achieved outstanding performance with the quick development of deep neural networks. However, training such a model heavily depends on the currently available labeled data. When different classification tasks are presented sequentially in prac", + "bbox": [ + 75, + 803, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video Dataset $V^k$", + "bbox": [ + 647, + 358, + 736, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5d816b0a227aea24bbdb09c06be3f4bfee0cae40ffe7f107cbb89183079091af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 373, + 589, + 425 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a3cee9f31a3ea15b508da7c6b0ebee738560c92788704146eb797ad583e95652.jpg", + "image_caption": [ + "···" + ], + "image_footnote": [], + "bbox": [ + 601, + 373, + 668, + 425 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/334d3c7ca92d3366b026260b7325fe423ec8dc67eb5a44fe87353873c515fe90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 373, + 792, + 425 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ac72593e7509c1b7e6f17659d9517856af9dbf2e6a00755d24843001203db69.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 373, + 870, + 425 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stored Videos[35]", + "bbox": [ + 529, + 443, + 620, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/af8a7bf617c85b5a197f881858441ea6e808ae59079a44f9a34fd81fe6340f04.jpg", + "image_caption": [ + "Memory: 585.70M Accuracy: $50.29\\%$" + ], + "image_footnote": [], + "bbox": [ + 506, + 455, + 573, + 508 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/996c38aa2bca57398f23ea538e0c80b152592940d4214de3b899efa7387355be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 455, + 643, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stored Frames[36]", + "bbox": [ + 666, + 443, + 763, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b33c9ffaebaec423ace6f283ab4876d5fa3e1399743f2194d87b17ac597ec1da.jpg", + "image_caption": [ + "Memory: 585.70M Accuracy: $51.43\\%$" + ], + "image_footnote": [], + "bbox": [ + 656, + 460, + 715, + 506 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d1109a9feafa61fa3ff70c0e94f409a269186470ae36c99010545ac9324b93dd.jpg", + "image_caption": [ + "Figure 1. An comparison between existing VCIL methods based on exemplar or frame replay and our ST-Prompt. Compared to the rehearsal-based approaches, our ST-prompt can better solve the problem of catastrophic forgetting when only a small amount of memory is consumed to store a tiny set of prompts." + ], + "image_footnote": [], + "bbox": [ + 717, + 460, + 774, + 507 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ST-Prompts (Ours).", + "bbox": [ + 792, + 443, + 885, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b4cfe96281ae8d0e6a6ef6dde0c6536e11bb471c1d381395662c2539ff752915.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 460, + 813, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7cee4d031d908eba99da31649bf9b77e77a359d8fa7af93858889c8268fd2d76.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 838, + 460, + 867, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f9a668d115558f627839feed4518143f8a62ea0dde28dd68c1c55dc226cc0702.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 799, + 487, + 813, + 497 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/320bcc749e7a10ac9ba647a91a32872d45339871609a726063ee3e8ed24485cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 838, + 487, + 867, + 497 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Memory: 2.98M", + "bbox": [ + 797, + 508, + 867, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accuracy: 55.13", + "bbox": [ + 792, + 522, + 867, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tical applications, a crucial problem is to maintain the accuracy of the model as data arrives over time. Under this circumstance, due to the limitations of memory and privacy, the model can just access the videos of the current task and few even no examples of historical tasks, which requires that the model should adapt to the current task without catastrophic forgetting [34]. The machine learning paradigm to handle the above challenges is called Video Class-Incremental Learning (VCIL) [35, 36, 61].", + "bbox": [ + 496, + 623, + 893, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most existing class-incremental learning approaches have achieved remarkable performance through a rehearsal buffer of the old tasks in both the image [40, 45, 59, 4] and video [36, 47, 36] fields. However, these methods suffer from suboptimal performance with smaller buffer sizes and severe catastrophic forgetting once the rehearsal buffer is not accessed, hence leading to limited applications. To solve this problem, several prompt-based learning methods in image domain [54, 53, 51] are proposed to train only", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 887, + 222, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11932", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a tiny set of parameters, i.e., prompts, to effectively instruct a pre-trained model to learn incremental tasks without buffering past examples. Typically, L2P [54] learns a single prompt pool to attach complementary prompts to the pre-trained backbone. Nevertheless, this new paradigm has shown great potential for images, but it is still underexplored in the video domain. In this paper, we will concentrate on exploring this paradigm on VCIL to bridge this gap. Directly applying these existing methods for the VCIL task tends to be inadequate. The core reason is that the prompt strategy for images cannot model temporal variations and dynamics across spatial features. Meanwhile, video data possesses a more complicated semantic structure than static images, so a single prompt pool like L2P may be insufficient to learn the abundant semantics. Therefore, we think a properly redesigned prompt framework will contribute to improving the performance of the VCIL task.", + "bbox": [ + 75, + 90, + 472, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the above observations, we propose ST-Prompt, a space-time prompting method for video class incremental learning. Specifically, we design ST-Prompt with two components, i.e., task-specific prompts and task-agnostic prompts. The task-specific prompts are tapped by multi-grained prompts, i.e., spatial prompts, temporal prompts and comprehensive prompts, which can better discover the complicated spatio-temporal semantics in videos to address the catastrophic forgetting problem. The task-agnostic prompts learn a complementary prompt for each frame and then gather them back for all frames; hence we can equip the image pre-trained model with temporal modeling capacities without any edition of model structure. Compared with the task-specific prompts, the task-agnostic prompts are task-independent and thus share a generic and globally-shared prompt pool for all tasks. By this means, ST-Prompt has two advantages: 1) it can simultaneously take advantage of the knowledge priors in the image-language pre-trained model and learn the spatiotemporal information, which makes the model more suitable for VCIL; 2) it can markedly reduce the memory cost by just saving a set of prompt parameters without preserving redundant videos anymore, as shown in Figure 1. In the experimental stage, we quantificationally and qualitatively evaluate ST-Prompt on three standard benchmarks. The results show that ST-Prompt can achieve significant gains over the current top-performing approaches, which demonstrates the effectiveness of our method.", + "bbox": [ + 75, + 347, + 472, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the main contributions of this paper are as follows:", + "bbox": [ + 76, + 770, + 470, + 800 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We first explore the great potential of image-language pre-trained model in video incremental tasks and achieve significant performance improvement by utilizing its sufficient semantic information through prompt-based methods;", + "- We proposed ST-Prompt, containing task-specific" + ], + "bbox": [ + 94, + 806, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "prompts and task-agnostic prompts, which can significantly reduce the memory cost and forgetting rate while modeling temporal variations and dynamics across spatial features without modifying the model structure;", + "bbox": [ + 529, + 90, + 890, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- The proposed ST-Prompt achieves remarkable performance on multiple standard action recognition incremental benchmarks over state-of-the-art methods.", + "bbox": [ + 517, + 179, + 890, + 222 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 244, + 640, + 260 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we will introduce the methods related to ST-Prompt, including the works for image and video class-incremental learning, and prompt-based learning.", + "bbox": [ + 498, + 272, + 890, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Class-Incremental Learning", + "text_level": 1, + "bbox": [ + 498, + 334, + 805, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, class-incremental learning has been widely studied and numerous methods have been proposed in the image domain [52, 56, 10, 43]. These methods can be roughly partitioned into three categories: regularization-based methods, exemplar-based methods and architecture-based methods. Regularization-based methods [23, 60, 21] constrain the optimization of network parameters to preserve the important information of learned tasks, in which the most popular technique utilized is knowledge distillation [29, 12, 14]. The exemplar-based methods [40, 57, 45] typically store a small set of representative exemplars [59, 50] or generated synthetic examples [22, 55] from previous classes to prevent catastrophic forgetting, which is always combined with regularization of a distillation loss to encourage knowledge retention further. The architecture-based methods [26, 42, 48] retain the knowledge of old tasks by designing specific components in the architecture [41, 64] or expanding current feature extractor for new data[58, 38].", + "bbox": [ + 496, + 359, + 890, + 631 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Video Class-Incremental Learning", + "text_level": 1, + "bbox": [ + 500, + 648, + 800, + 665 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the video domain, studies on this field are still scarce. The existing solutions are always the combination of regularization-based and exemplar-based methods and focus on how to formulate temporal information better. For example, TCD [35] computes time-channel importance for weighted distillation, [61] decomposes spatiotemporal knowledge transfer for stronger constraint and vCLIMB [47] focuses on temporal-consistency regularization of untrimmed video. Both TCD [35] and vCLIMB [47] have claimed that more stored examples in memory can effectively encourage incremental performance. Specifically, FrameMaker [36] tends to design a memory-efficient video incremental learning method by learning a condensed frame for each representative video of old classes. However, it still demands to preserve some old frames.", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11933", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/342b993ae8f3893291718be0a838aea9be2e10da319578900b771c2cb5effa43.jpg", + "image_caption": [ + "Figure 2. The overall framework of the proposed ST-Prompt, which mainly contains two kinds of prompts, task-specific prompts and task-agnostic prompts. For each input video, we first extract multi-grained features with the image encoder of a CLIP model, then select corresponding prompts from predefined prompt pools. Finally, the prompts are prepended to the video and text embedding feature respectively to instruct the model to learn incremental tasks." + ], + "image_footnote": [], + "bbox": [ + 78, + 85, + 893, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Prompt-based Learning", + "text_level": 1, + "bbox": [ + 76, + 426, + 299, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, with the rise of large pre-trained models [39, 3, 18, 28], the general pre-trained models are utilized to adapt to multiple down-sampling tasks by finetuning or prompting. The concept of prompt was first proposed in the field of natural language processing [31, 5, 37], and later it was more widely used in language[25, 27, 16] and visual[19, 2, 20] tasks to modify the input texts or images, such that the language or image model obtains additional information about a specific task. Some methods consider multi-modality information simultaneously, and design prompts for the vision-language models[63, 62, 17].", + "bbox": [ + 75, + 470, + 472, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Meanwhile, the emergence of large pre-trained models and prompt-based learning also shows excellent potential in incremental learning. L2P [54], DualPrompt [53] and S- iPrompts [51] exploit prompting method to image incremental learning and achieve good performance without saving old exemplars. Especially, S-Prompting utilizes both image and text prompts together with a pretrained vision-language model and gains outstanding performance in domain-incremental learning. In this paper, we focus on the properties of video, adapting large image pre-trained models to the video domain for video class-incremental learning by space-time prompting. We propose to decompose the complex video features into multiple space-time granularities for accurate task identification while exchanging the temporal information between frames by task-agnostic prompts.", + "bbox": [ + 75, + 659, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 426, + 591, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we will first introduce the basic setting of the video class-incremental learning task. Then we discuss the proposed ST-Prompt in detail, as shown in Figure 2. In the end, we present the training and inference procedure.", + "bbox": [ + 498, + 452, + 893, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 500, + 527, + 702, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In video class-incremental learning, different tasks $\\{\\mathcal{T}^1,\\mathcal{T}^2,\\dots ,\\mathcal{T}^K\\}$ arrive in a sequence and each task $\\mathcal{T}^k$ corresponds to its specific dataset $\\mathcal{D}^k = \\{(V_i^k,y_i^k),y_i^k\\in L^k\\}$ , which contains the input video example $V_{i}^{k}$ and its label $y_{i}^{k}$ . The labels in task $k$ belong to a predefined label set $L^k$ and do not appear in previous tasks. The goal of VCIL is to train a single model that has the ability to classify videos of all seen classes correctly. We focus on the more challenging problem of exemplar-free video class-incremental learning, i.e., no exemplar of old classes can be used in the current task, in which the core is to avoid forgetting historical information, termed catastrophic forgetting.", + "bbox": [ + 498, + 551, + 890, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we follow the prompt-based incremental learning paradigm in the image domain [54, 51] to alleviate the forgetting problem, which is composed of two stages: $i$ based on the pre-trained vision-language model like CLIP, they optimize a tiny set of learnable visual and text prompts for each incremental task independently (i.e. task-specific prompts) to learn new knowledge and retain the historical information; $ii$ during inference, due to the lack of task identity of the test example, they select the proper visual and text prompts from the prompt pool by prompt matching, and then preprocess them to the visual embedding $V_{ie}^{k}$ and the class", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11934", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "embedding $W_{we}^{k}$ respectively. The final prediction is determined by the similarities between the visual features and the text features of class names. However, different from static images, videos contain more complicated spatio-temporal clues, and it is intractable to capture these semantic clues well in the video tasks via a single kind of prompt. Meanwhile, the image pre-trained model lacks temporal modeling capability for videos, which requires us to make up for this defect during the prompting process. Overall, the essential point for VCIL is to design a reasonable prompting method to recall the spatio-temporal information in the video task well.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Space-time prompting", + "text_level": 1, + "bbox": [ + 76, + 281, + 284, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed space-time prompting (ST-Prompt) contains two components, i.e., task-specific prompts and task-agnostic prompts.", + "bbox": [ + 75, + 304, + 468, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task-specific Prompts. To prevent catastrophic forgetting, we train task-specific prompts for each incremental task $\\mathcal{T}^k$ with the pre-trained model (the image encoder $f_{v}$ and text encoder $f_{l}$ ) frozen. Specifically, we devote to designing multi-grained prompts to utilize the temporal and spatial variations of different videos for more accurate task identification. Formally, for each task $\\mathcal{T}^k$ , there are three different granularities of task-specific prompt pools as Eq. 1. For brevity, we omit the symbol $k$ in this section. If there is no additional explanation, it indicates the situation in incremental task $\\mathcal{T}^k$ .", + "bbox": [ + 75, + 356, + 472, + 521 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} = \\left\\{\\mathbf {P} _ {s}, \\mathbf {P} _ {t}, \\mathbf {P} _ {c} \\right\\}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 529, + 230, + 546 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {m} = \\left\\{\\left\\{P _ {m 1} ^ {v}, P _ {m 1} ^ {l} \\right\\}, \\dots , \\left\\{P _ {m N} ^ {v}, P _ {m N _ {m}} ^ {l} \\right\\} \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 547, + 468, + 566 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {m n} ^ {v} \\in \\mathbb {R} ^ {L ^ {v} \\times D ^ {v}}, P _ {m n} ^ {l} \\in \\mathbb {R} ^ {L ^ {l} \\times D ^ {l}}, m \\in \\{s, t, c \\},\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 569, + 423, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{P}_s$ , $\\mathbf{P}_t$ and $\\mathbf{P}_c$ are the spatial, temporal and comprehensive task-specific prompt pools. Each element in prompt pools is organized as the formulation of a visual-text prompt pair, where $P_{mn}^v$ and $P_{mn}^l$ are the commensurable visual and text prompt respectively. Additionally, $N_m$ is the numbers of prompts in each prompt pool. $L^v$ and $L^l$ are the length of the visual and text prompt, and $D^v$ and $D^l$ are their embedding dimension.", + "bbox": [ + 75, + 598, + 468, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The selection of prompts from these prompt pools is realized through feature matching as Figure 3(a). We apply K-Means algorithm to store the centroids of video features and search for the nearest centroid of the given video sample to identify its relevant prompt. The matching of three kinds of prompts corresponds to features in three different granularities of video $V$ , i.e. spatial feature $F_{s}(V)$ , temporal feature $F_{t}(V)$ and comprehensive feature $F_{c}(V)$ . Specifically, the spatial and temporal features are calculated by pooling the temporal and spatial dimensions separately, while the comprehensive feature is the average of [CLS] token output. After the feature matching, the selected task-specific", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6e5cda47a219c9d6fcf769e039cbc1774d3a5db56ddeb8ebecbae6944eb3b10a.jpg", + "image_caption": [ + "Figure 3. The details of prompt matching methods in prompt pools. (a) We select prompts corresponding to the nearest centroid with query feature for the task-specific prompts; (b) We utilize the similarities between the frame feature and centroids as the attention map for the task-agnostic prompts." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 890, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "prompts are organized together as:", + "bbox": [ + 498, + 347, + 730, + 361 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP ^ {v} = f \\left(P _ {s *} ^ {v}, P _ {t *} ^ {v}, P _ {c *} ^ {v}\\right), P ^ {l} = f \\left(P _ {s *} ^ {l}, P _ {t *} ^ {l}, P _ {c *} ^ {l}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 369, + 890, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(P_{s*}^v, P_{s*}^l)$ , $(P_{t*}^v, P_{t*}^l)$ and $(P_{c*}^v, P_{c*}^l)$ are the prompts selected from three task-specific prompt pools. $f$ is the organization function, which can be stacking, addition, multiplication, or other operations. In this paper, we use addition for its better performance.", + "bbox": [ + 496, + 397, + 890, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the visual and text prompts are prepended to the visual and text embedding respectively, and the whole input of the visual and text encoder is of the format:", + "bbox": [ + 496, + 474, + 890, + 518 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {p} = \\left[ P ^ {v}; V _ {e} \\right], W _ {w p} = \\left[ P ^ {l}; W _ {w e} \\right]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 529, + 890, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where ; indicates the impending operation, $V_{e}$ is the video embedding feature of video $V$ and $W_{we}$ is the word embedding feature of w-th class name in task $\\mathcal{T}^k$", + "bbox": [ + 496, + 556, + 890, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additionally, when we utilize the visual pre-trained models without a language branch, such as ViT with ImageNet pre-trained weights, the classification head replaces the text encoder and the prompts in each prompt pool just contain the visual prompt $P_{mn}^{v}$ . The other definition and utilization of the task-specific prompts are similar to the above.", + "bbox": [ + 496, + 602, + 890, + 693 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task-agnostic prompts. When tackling video incremental tasks, the image pre-trained model processes each video frame separately and cannot model the temporal variations and relations between different frames. Although we utilize different granularities of prompts for task identification, it is not enough to make the model understand the temporal information of videos well. To enhance the temporal modeling ability and adaptability of image pre-trained models, we propose task-agnostic prompts to highlight the different information contained in frames. Specifically, we establish an additional temporal-modeling prompt pool as $\\mathbf{P_{a}} = \\{P_{a1},\\dots,P_{aN_a}\\}$ , where $P_{an} \\in \\mathbb{R}^{L^a \\times D^v}$ is the prompt and $N_{a}$ is the total number of prompts. Compared", + "bbox": [ + 496, + 704, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11935", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with the task-specific prompts, temporal modeling is a more general concept for videos, which is task-agnostic and all prompts from the whole pool can be selected and attached without task limitation.", + "bbox": [ + 75, + 90, + 468, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the selection of task-agnostic prompts, we also store the centroids of the frame features during training by KMeans algorithm. Differently, in order to utilize the information in the temporal-modeling prompt pool comprehensively, the final prompt according to each frame is not selected by feature matching simply but is calculated by frame-level attention as Figure 3(b). For each frame $I_{j}$ in video $V$ , the attention map for all candidate prompts $S_{j} = \\{S_{j1}, S_{j2}, \\dots, S_{jNa}\\}$ is generated according to the similarity between the frame feature $f_{v}(I_{j})$ and the centroids. Specifically, we utilize the cosine similarity to calculate $S_{jn}$ in this paper. And then the task-agnostic prompt corresponding to the frame $I_{j}$ is calculated as follows:", + "bbox": [ + 75, + 151, + 470, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {j} = \\sum_ {n = 1} ^ {N _ {a}} \\frac {e ^ {S _ {j n} / \\tau}}{\\sum_ {n = 1} ^ {N _ {a}} e ^ {S _ {j n} / \\tau}} P _ {a n} \\in \\mathbb {R} ^ {D ^ {v}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 361, + 468, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ is the temperature parameter. Finally, the prompts of different frames are stacked together to adapt to the video embedding as $V_{p} = [P_{1};P_{2};\\ldots ;P_{T};V_{e}]$ . In this way, each frame will obtain the context information of others. Although they go forward through an image pre-trained model separately, the information exchange of different frames realizes the temporal modeling.", + "bbox": [ + 75, + 412, + 468, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training and Inference", + "text_level": 1, + "bbox": [ + 76, + 529, + 290, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, only the prompts are trainable, while the pre-trained model is fixed. In incremental step $\\mathcal{T}^k$ , the training of prompts in task-specific prompt pools is restrained so that only the prompts corresponding to task $\\mathcal{T}^k$ can be selected and utilized. The task-agnostic prompt pool does not have this constraint because temporal modeling is a more general concept independent of task identification.", + "bbox": [ + 75, + 551, + 468, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference, for each new video, we first extract multi-grained features by the pre-trained image encoder $f_{v}$ , and then select task-specific and task-agnostic prompts from prompt pools through feature matching or frame-level attention. Finally, the selected prompts are prepended to the video and text embedding features respectively. The classification result is according to the similarity output.", + "bbox": [ + 75, + 657, + 468, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When ST-Prompt is utilized based on a one-branch visual model, in order to separate different tasks better, the classification head is also split according to the task identity and every incremental session trains its own classifier. Differently, without the text branch, the classifier cannot utilize the multi-grained task-specific prompts flexibly and can only be selected by the video-level comprehensive feature. The classifier selection corresponds to the task that the selected comprehensive task-specific prompt belongs to.", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 90, + 624, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we compare our results with baselines and state-of-the-art VCIL methods. We also show ablation studies on the effect of different components and analyze our results in different practical settings.", + "bbox": [ + 498, + 114, + 890, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment setup", + "text_level": 1, + "bbox": [ + 498, + 188, + 671, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on UCF101 [44], HMDB51 [24] and Something-Something V2 [13], which are the standard action recognition datasets. We follow the video frequently-used class-incremental benchmarks presented in TCD [35]. For UCF101, the model is trained on 51 classes first, and the other 50 classes are divided into 5, 10 and 25 tasks. For HMDB51, we train the base model using videos from 26 classes, and the rest of the classes are separated into 5 or 25 groups. For Something-Something V2, we first train 84 classes in the initial stage and generate the incremental groups of 10 and 5 classes.", + "bbox": [ + 498, + 212, + 890, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Protocol. After each incremental step, we evaluate the model on all seen classes and the class-incremental methods are finally assessed by the average accuracy of all tasks.", + "bbox": [ + 498, + 402, + 890, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Our method is based on CLIP ViT-B/16 [39], a visual-language model pre-trained with numerous image-text pairs. While the pre-trained model is frozen, the visual and textual prompts are optimized using a batch size of 64 and an initial learning rate of 0.001 with cosine descending for 50 epochs in each task. We carefully choose compared methods in the same environment for a fair comparison. However, most existing class-incremental learning methods are not adaptive to visual-language models. Therefore, we also apply our approach on ImageNet pre-trained ViT-B/16 [8] and reproduce other class-incremental learning methods on the same backbone.", + "bbox": [ + 498, + 486, + 890, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Main Results.", + "text_level": 1, + "bbox": [ + 500, + 680, + 640, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our proposed ST-Prompt with existing state-of-the-art class-incremental learning approaches under multiple challenging settings on three datasets in Table 1 and Table 2. For the approaches in the image domain, we reproduce them by treating the video as a whole as that in the image field. From the tables, we can draw the following conclusions. First, ST-Prompt notably outperforms existing rehearsal-free methods in VCIL, containing two recent prompt-based approaches, L2P and S-iPrompts in the image domain. With ImageNet pre-trained ViT, ST-Prompt respectively surpasses S-iPrompts by around $4.22\\%$ , $4.98\\%$ and $7.27\\%$ on HMDB51, UCF101 and Something-something v2, respectively. This indicates that", + "bbox": [ + 498, + 704, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11936", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b1f0220845eecae49381fafd91339b1ac5fe6849cf2252afe024ee947598656f.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Comparison with the state-of-the-art approaches over class-incremental action recognition performance on UCF101 and HMDB51. Compared methods are grouped based on different rehearsal buffer sizes. † indicates the methods designed for video." + ], + "table_body": "
MethodsBuffer SizeUCF101HMDB51
10 × 5 stages5 × 10 stages2 × 25 stages5 × 5 stages1 × 25 stages
ImageNet pre-trained ViT-B/16
iCaRL [40]70.5869.5167.2843.9037.15
UCIR [15]77.5574.5971.7748.2039.42
PODNet [9]76.5076.1773.8348.3843.35
Co2L [6]10/class78.5477.1375.5949.7644.15
TCD†[35]78.1376.8775.7450.2944.04
FrameMaker†[36]79.3779.5579.3251.4346.37
L2P [54]81.2480.0978.5849.9845.87
iCaRL [40]66.5867.7064.2741.0836.13
UCIR [15]75.5572.7866.7645.3838.40
PODNet [9]74.5074.3672.8245.5642.33
Co2L [6]5/class77.5776.3274.5847.1043.56
TCD†[35]76.1375.0673.7348.3843.02
FrameMaker†[36]78.3777.7476.3150.5845.65
L2P [54]79.9377.7876.2747.6742.56
Finetuning24.8313.616.2418.304.89
Finetuning-frozen67.1065.717.3936.644.95
LwFMC [29]45.7328.0116.4327.0117.21
LwM [7]0/class46.9828.7016.9227.9417.85
L2P [54]76.7275.1173.2645.0039.89
S-iPrompts [51]77.1677.8377.9950.6751.45
ST-Prompt†83.7481.5382.6655.1355.43
CLIP ViT-B/16
Zero shot [39]69.6869.6169.8740.1940.42
S-liPrompts [51]0/class80.6080.2780.4353.1153.89
ST-Prompt†84.7585.5485.6760.1460.54
", + "bbox": [ + 99, + 88, + 870, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "our design of multi-grained task-specific prompts and task-agnostic prompts plays an essential role in video feature formulation and temporal perception. The more tremendous increase on Ssv2 also claims the effectiveness of ST-Prompt for the motion-sensitive dataset. Second, compared with rehearsal-based methods, our ST-Prompt also has a significant advantage. ST-Prompt achieves better performance than the top existing approach, FrameMaker, by approximately $6.38\\%$ and $3.23\\%$ on HMDB51 and UCF101, separately, although FrameMaker stores additional 80 examples for each class. Third, using the pre-trained vision-language model, CLIP, as the backbone for VCIL, the performance of ST-Prompt improves further by about $5.06\\%$ , $2.67\\%$ and $3.49\\%$ respectively. The reason is that CLIP provides more abundant prior knowledge about semantic information, which is vital for subsequent prompting.", + "bbox": [ + 76, + 575, + 472, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study.", + "text_level": 1, + "bbox": [ + 76, + 830, + 235, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we present ablation studies on the properties and effectiveness of our core designs. If not specified, the ablation studies are performed on HMDB51 with 5 steps", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and UCF101 with 10 steps.", + "bbox": [ + 500, + 575, + 683, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Task-specific Prompts. In Table 3, we first explore how to organize different kinds of task-specific prompts together in Eq. 2. The results show that the addition operation achieves better performance than stacking and multiplication, whose reason may be that different prompts can modulate each other better without preponding too many parameters to the embedding feature in this way.", + "bbox": [ + 496, + 612, + 890, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, to prove the effectiveness of different parts of our task-specific prompts, we remove each kind of prompts respectively under the combination method of addition and show their results in Table 4. From the results, we can find that: first, when three kinds of prompts are utilized separately, the comprehensive task-specific prompts achieve the best performance, which demonstrates that they depict the video feature more integrally; second, when the comprehensive prompts are combined with the temporal or spatial prompts, the performance increases by $1.9\\%$ or $1.54\\%$ on HMDB51, and $1.26\\%$ or $1.34\\%$ on UCF101, respectively. Differentiated promotions in different datasets imply that", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11937", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d5cd6f276249f060690e8bd5c87ba9594213176f031b55265d898f85c0c19e8f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsBuffer Size10 × 9 Stages5 × 18 Stages
ImageNet pre-trained ViT-B/16
iCaRL [40]20.4116.62
UCIR [15]24.3219.31
PODNet [9]5/class27.6320.14
\\( \\mathrm {TCD}^{\\dagger } \\)[35]29.3224.69
FrameMaker \\( {}^{ \\dagger } \\)[36]31.4126.57
L2P [54]26.0221.33
Finetuning11.046.23
Finetuning-frozen7.995.97
LwFMC [29]0/class15.2111.41
L2P [54]23.2518.49
S-iPrompts [51]30.5826.33
ST-Prompt36.8431.60
CLIP ViT-B/16
Zero-shot3.563.57
S-liPrompts [51]0/class33.6930.84
\\( \\mathrm {ST}-\\mathrm {Prompt}^{\\dagger } \\)39.9835.44
", + "bbox": [ + 89, + 88, + 459, + 387 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a9fb069e41ce1e6dc67c39ccc9886fd43e581b88020c6279127092472160faaf.jpg", + "table_caption": [ + "Table 2. Comparison with the top approaches on Something-something v2.† indicates the methods designed for video." + ], + "table_footnote": [], + "table_body": "
OperationStackingMultiplicationAddition
HMDB5155.2454.9656.02
UCF10182.8282.4583.58
", + "bbox": [ + 94, + 431, + 450, + 491 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5475ff1e83a0d3f5bfade717b1f836004ec71c2d2f90a8abed915dba9c6913bb.jpg", + "table_caption": [ + "Table 3. Ablation for the ensemble ways of task-specific prompts." + ], + "table_footnote": [], + "table_body": "
STPTTPCTPHMDB51UCF101
✓✘✘44.6272.45
✘✓✘45.4673.84
✘✘✓53.1180.27
✓✓✘49.3476.93
✓✘✓54.6581.61
✘✓✓55.0181.53
✓✓✓56.0283.58
", + "bbox": [ + 117, + 520, + 426, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "temporal information is more critical for HMDB51, while the spatial feature for UCF101; third, utilizing all three kinds of task-specific prompts can further improve the final accuracy, which proves the combination of multi-grained prompts contribute to more correct task identification.", + "bbox": [ + 75, + 715, + 467, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Further, we visualize some clustering results of different granularities of the video feature in Figure 4. The comprehensive feature judges video task-id from the overall perspective, and the clustered videos usually belong to the same spatial and temporal clusters, such as \"Smile\". The temporal and spatial prompts implement the local information for more complicated video task matching. Videos", + "bbox": [ + 75, + 795, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6bbcaca1bfa2650568295d9f135162146d957851d51392fe678b78eb9e678c56.jpg", + "table_caption": [ + "Table 4. Ablations for task-specific prompts on HMDB51 with 5 steps and UCF101 with 10 steps. STP, TTP and CTP represent spatial, temporal and comprehensive prompts respectively." + ], + "table_footnote": [], + "table_body": "
MethodTAPHMDB51UCF101
TSPX56.0283.58
TSP+LEX56.0683.55
TSP+SPX57.0483.87
TSP+CPX58.3084.05
TSP+PE [46]X57.9684.04
TSP+SH [33]X58.2784.25
TSP+TS [30]X58.7484.34
ST-Prompt✓60.1485.54
", + "bbox": [ + 535, + 88, + 856, + 238 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/41cb8bee27e330aa0dd951c46640f374cb531df507cca9902f9b963e5118f6f6.jpg", + "table_caption": [ + "Table 5. Comparison with some baseline methods about task-agnostic prompts (TAP) on HMDB51 with 5 steps and UCF101 with 10 steps." + ], + "table_footnote": [], + "table_body": "
MethodTAPHMDB51UCF101
L2P [54]✘45.0075.11
✓46.7676.07
S-iPrompts [51]✘50.6777.83
✓54.4479.74
S-liPrompts [51]✘53.1180.27
✓55.7781.82
ST-Prompt✘56.0283.58
✓60.1485.54
", + "bbox": [ + 524, + 291, + 867, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Ablations for task-agnostic prompts (TAP) on HMDB51 with 5 steps and UCF101 with 10 steps.", + "bbox": [ + 500, + 441, + 890, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with the same motion but different spatial features are gathered together according to their temporal features, such as \"Dive\" in different scenarios, while spatial features distinguishes actions with mixed temporal information but obvious spatial similarities.", + "bbox": [ + 498, + 497, + 890, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Task-agnostic Prompts. We compare our task-agnostic prompts for temporal modeling with a set of baselines in the proposed framework in Table 5. \"LE\" represents prompt length expansion, which excludes the performance improvement caused by increasing the number of parameters. And the results also prove that learning task-agnostic prompts about spatial (\"SP\") or comprehensive (\"CP\") information are not as effective as temporal modeling. Then we compare the effect of our temporal-modeling prompts with some other parameter-free temporal modeling methods, such as positional encoding (\"PE\"), temporal shuffle (\"SH\") and temporal shift (\"TS\"). We can observe that utilizing task-agnostic prompts to exchange the context of different frames can gain the best performance of $60.14\\%$ . Further, we validate the effectiveness of task-agnostic prompts by adding them to other prompt-based methods in the image domain, as shown in Table 6, which indicates that it is a general approach to improve the ability of temporal perception for image pre-trained models.", + "bbox": [ + 496, + 597, + 890, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, We depict the change of category-wise ac", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11938", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5cb50e22ede36c74cddc46d7bb6fa99d51c879ac857750a220e1ec6df7bcc529.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsBackboneStored ExemplarsMemory CostAverage Accuracy
TCD [35]ImageNet Pre-trained ViT-B/1610 examples/class585.70M50.29
FrameMaker [36]ImageNet Pre-trained ViT-B/1680 examples/class585.70M51.43
ST-PromptImageNet Pre-trained ViT-B/1602.98M55.13
ST-PromptCLIP ViT-B/1604.47M60.14
", + "bbox": [ + 109, + 90, + 859, + 172 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. The comparison of the overhead of our ST-Prompt and other VCIL methods on HMDB51 with 5 steps.", + "bbox": [ + 151, + 172, + 810, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/11e09fd6ce770f778beed8c27aaa2ffa91b4efe92cf2a1119b44bafd785b6ac9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 213, + 467, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2439f4e3c67b916c502da9dec6123db3c19bbdcbe88fc049d9a40a9f9c3cf846.jpg", + "image_caption": [ + "Figure 4. The visualization of multi-grained video feature clustering results.", + "Figure 5. The change of category-wise accuracy after utilizing task-agnostic prompts on HMDB51 with 5 steps." + ], + "image_footnote": [], + "bbox": [ + 81, + 367, + 464, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "curacy after utilizing task-agnostic prompts in Figure 5. The significant increase in the accuracy of some motion-sensitive classes, such as \"Dive\" and \"Somersault\", proves the role of our task-agnostic prompts in improving the temporal modeling capability of models.", + "bbox": [ + 75, + 553, + 467, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Temperature $\\tau$ . The temperature $\\tau$ in Eq. 4 decides the smoothness of the attention weights. Too small $\\tau$ makes the final prompt of each frame lose the understanding of global information, and too large $\\tau$ introduces noise. From Figure 6(a), we observe that $\\tau = 0.3$ guides the best performance of final average accuracy.", + "bbox": [ + 75, + 642, + 467, + 732 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Length of Prompts. The length of different kinds of prompts $L^v, L^t, L^a$ decides the capacity of each learnable prompt. Figure 6(b) illustrates how the average accuracy changes when the length of different prompts varies. It is evident that a too-small visual or text prompt length always negatively affects results, while an oversized prompt may cause underfitting. We set the prompt length as 5 for both visual and text task-specific prompts according to the curves. In addition, we can discover that the final accuracy is not sensitive to the length of task-agnostic prompts", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/24dabd0d866899086d9cd3837bb159abf92529377d3e6ab08733721e1400a451.jpg", + "image_caption": [ + "Figure 6. The effects of some hyperparameters in ST-Prompt." + ], + "image_footnote": [], + "bbox": [ + 506, + 217, + 893, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and even decreases during the length is larger. Hence, we choose the length of the task-agnostic prompts as 1.", + "bbox": [ + 498, + 386, + 890, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Number of Prompts. Figure 6(c) presents how the number of different prompts $N_{m}(m \\in \\{s, t, c, a\\})$ in each incremental task influences the performance of our approach, where the number in the base task is calculated according to the proportion of the number of classes. Three different task-specific prompts and task-agnostic prompts are explored separately. In general, the more prompts per task, i.e., the more pre-defined centroids per task during clustering, the higher accuracy gained. Especially, the comprehensive prompt is more sensitive to the cluster number. For simplicity, the number of different prompts is set as 5.", + "bbox": [ + 496, + 431, + 890, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Overhead Analysis", + "text_level": 1, + "bbox": [ + 500, + 608, + 679, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare the memory cost and average accuracy of ST-Prompt with other VCIL methods in Table 7. Our ST-Prompt utilizes only a set of prompts and prompt keys in each task, while previous rehearsal-based methods store a number of video examples or frames. Therefore, ST-Prompt has significant memory advantages over previous VCIL methods. Our ST-Prompt achieves performance increasing by around $8.71\\%$ with only $0.76\\%$ memory cost of other methods.", + "bbox": [ + 496, + 633, + 890, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussions", + "text_level": 1, + "bbox": [ + 500, + 784, + 619, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation. Compared with the previous video incremental learning approaches, ST-Prompt achieves a significant performance boost without any additional memory cost for video example saving. Nevertheless, ST-Prompt relies on the large pre-trained model, which may fail when the knowledge contained in it is not sufficient. Meanwhile, our", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11939", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ST-Prompt prepends prompts to the embedding feature and has not discussed the phenomenon that prompts are prefixed to the hidden layers.", + "bbox": [ + 76, + 90, + 472, + 137 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Conclusion. In this paper, we propose ST-Prompt, a novel space-time prompting framework for video class incremental learning. It explores how to address the VCIL problem in a rehearsal-free way by learning multiple prompts based on a powerful image-language pre-trained model. ST-Prompt mainly consists of two components, task-specific prompts and task-agnostic prompts. The task-specific prompts tackle the catastrophic forgetting problem by learning multi-grained prompts for task identification. The task-agnostic prompts implement the temporal modeling ability of the image pre-trained model. In this way, ST-Prompt achieves a new state-of-the-art performance, which has already surpassed the methods with a memory buffer. We hope this simple and practical framework can inspire attention to prompting-based methods in VCIL.", + "bbox": [ + 76, + 147, + 472, + 375 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported in part by the NSFC under Grant 62272380 and 62103317, the Science and Technology Program of Xi'an, China under Grant 21RGZN0017, and by Alibaba Group through Alibaba Innovative Research Program.", + "bbox": [ + 76, + 396, + 472, + 472 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 487, + 174, + 503 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 1", + "[2] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Visual prompting: Modifying pixel space to adapt pre-trained models. arXiv preprint arXiv:2203.17274, 2022. 3", + "[3] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022. 3", + "[4] Zalán Borsos, Mojmir Mutny, and Andreas Krause. Coresetts via bilevel optimization for continual learning and streaming. Advances in Neural Information Processing Systems, 2020. 1", + "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 3", + "[6] Hyuntak Cha, Jaeho Lee, and Jinwoo Shin. Co2l: Contrastive continual learning. In ICCV, 2021. 6", + "[7] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In CVPR, 2019. 6", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5" + ], + "bbox": [ + 86, + 512, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, 2020. 6, 7", + "[10] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In CVPR, 2022. 2", + "[11] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 1", + "[12] Tao Feng, Mang Wang, and Hangjie Yuan. Overcoming catastrophic forgetting in incremental object detection via elastic response distillation. In CVPR, 2022. 2", + "[13] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In ICCV, 2017. 5", + "[14] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Lifelong learning via progressive distillation and retrospection. In ECCV, 2018. 2", + "[15] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, 2019. 6, 7", + "[16] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. 3", + "[17] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 3", + "[18] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3", + "[19] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 3", + "[20] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, 2022. 3", + "[21] Minsoo Kang, Jaeyoo Park, and Bohyung Han. Class incremental learning by knowledge distillation with adaptive feature consolidation. In CVPR, 2022. 2", + "[22] Ronald Kemker and Christopher Kanan. Fearnet: Brain-inspired model for incremental learning. In ICLR, 2018. 2", + "[23] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 2017. 2", + "[24] Hildegard Kuehne, Hueihan Jhuang, Estibaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, 2011. 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11940", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021. 3", + "[26] Xilai Li, Yingbo Zhou, Tianfu Wu, Richard Socher, and Caiming Xiong. Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In ICML, 2019. 2", + "[27] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In ACL-IJCNLP, 2021. 3", + "[28] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In ICLR, 2022. 3", + "[29] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 2017. 2, 6, 7", + "[30] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In ICCV, 2019. 7", + "[31] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Comput. Surv., 2022. 3", + "[32] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, 2022. 1", + "[33] Pingchuan Ma, Yao Zhou, Yu Lu, and Wei Zhang. Learning efficient video representation with video shuffle networks. arXiv preprint arXiv:1911.11319, 2019. 7", + "[34] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation. 1989. 1", + "[35] Jaeyoo Park, Minsoo Kang, and Bohyung Han. Class incremental learning for action recognition in videos. In ICCV, 2021. 1, 2, 5, 6, 7, 8", + "[36] Yixuan Pei, Zhiwu Qing, Jun CEN, Xiang Wang, Shiwei Zhang, Yaxiong Wang, Mingqian Tang, Nong Sang, and Xueming Qian. Learning a condensed frame for memory-efficient video class-incremental learning. In NeurIPS, 2022. 1, 2, 6, 7, 8", + "[37] Fabio Petroni, Tim Rocktäschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and Sebastian Riedel. Language models as knowledge bases? In EMNLP, 2019. 3", + "[38] Qi Qin, Wenpeng Hu, Han Peng, Dongyan Zhao, and Bing Liu. Bns: Building network structures dynamically for continual learning. Advances in Neural Information Processing Systems, 2021. 2", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 5, 6", + "[40] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017. 1, 2, 6, 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 2", + "[42] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In ICML, 2018. 2", + "[43] Yujun Shi, Kuangqi Zhou, Jian Liang, Zihang Jiang, Jiashi Feng, Philip HS Torr, Song Bai, and Vincent YF Tan. Mimicking the oracle: An initial phase decorrelation approach for class incremental learning. In CVPR, 2022. 2", + "[44] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5", + "[45] Rishabh Tiwari, Krishnateja Killamsetty, Rishabh Iyer, and Pradeep Shenoy. Gcr: Gradient coreset based replay buffer selection for continual learning. In CVPR, 2022. 1, 2", + "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 2017. 7", + "[47] Andrés Villa, Kumail Alhamoud, Victor Escorcia, Fabian Caba, Juan León Alcazar, and Bernard Ghanem. vclimb: A novel video class incremental learning benchmark. In CVPR, 2022. 1, 2", + "[48] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In ECCV, 2022. 2", + "[49] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In ECCV, 2016. 1", + "[50] Liyuan Wang, Xingxing Zhang, Kuo Yang, Longhui Yu, Chongxuan Li, Lanqing Hong, Shifeng Zhang, Zhenguo Li, Yi Zhong, and Jun Zhu. Memory replay with data compression for continual learning. In ICLR, 2022. 2", + "[51] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. In NeurIPS, 2022. 1, 3, 6, 7", + "[52] Zhen Wang, Liu Liu, Yiqun Duan, Yajing Kong, and Dacheng Tao. Continual learning with lifelong vision transformer. In CVPR, 2022. 2", + "[53] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In ECCV, 2022. 1, 3", + "[54] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In CVPR, 2022. 1, 2, 3, 6, 7", + "[55] Chenshen Wu, Luis Herranz, Xialei Liu, Joost van de Weijer, Bogdan Raducanu, et al. Memory replay gans: Learning to generate new categories without forgetting. In NeurIPS, 2018. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11941", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Tz-Ying Wu, Gurumurthy Swaminathan, Zhizhong Li, Avinash Ravichandran, Nuno Vasconcelos, Rahul Bhotika, and Stefano Soatto. Class-incremental learning with strong pre-trained models. In CVPR, 2022. 2", + "[57] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, 2019. 2", + "[58] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, 2021. 2", + "[59] Jaehong Yoon, Divyam Madaan, Eunho Yang, and Sung Ju Hwang. Online coreset selection for rehearsal-based continual learning. arXiv preprint arXiv:2106.01085, 2021. 1, 2", + "[60] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, 2017. 2", + "[61] Hanbin Zhao, Xin Qin, Shihao Su, Yongjian Fu, Zibo Lin, and Xi Li. When video classification meets incremental classes. In ACM MM, 2021. 1, 2", + "[62] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 3", + "[63] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 2022. 3", + "[64] Kai Zhu, Wei Zhai, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Self-sustaining representation expansion for nonexemplar class-incremental learning. In CVPR, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 489 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11942", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Space-time Prompting for Video Class-incremental Learning/23f1a364-1b5a-487e-9a38-2d1ef0112aa7_model.json b/2023/Space-time Prompting for Video Class-incremental Learning/23f1a364-1b5a-487e-9a38-2d1ef0112aa7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d84d4d081b9af027658c32fdb863bb6bd077145e --- /dev/null +++ b/2023/Space-time Prompting for Video Class-incremental Learning/23f1a364-1b5a-487e-9a38-2d1ef0112aa7_model.json @@ -0,0 +1,2444 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.179, + 0.131, + 0.795, + 0.156 + ], + "angle": 0, + "content": "Space-time Prompting for Video Class-incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.18, + 0.737, + 0.22 + ], + "angle": 0, + "content": "Yixuan Pei\\(^{1}\\) Zhiwu Qing\\(^{2}\\) Shiwei Zhang\\(^{3*}\\) Xiang Wang\\(^{2}\\) Yingya Zhang\\(^{3}\\) Deli Zhao\\(^{3}\\) Xueming Qian\\(^{1,4*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.234, + 0.593, + 0.252 + ], + "angle": 0, + "content": "\\(^{1}\\)Xi'an Jiaotong University" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.252, + 0.758, + 0.288 + ], + "angle": 0, + "content": "\\(^{2}\\)Huazhong University of Science and Technology \\(^{3}\\)Alibaba Group, Shaanxi Yulan Jiuzhou Intelligent Optoelectronic Technology Co., Ltd." + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.29, + 0.778, + 0.305 + ], + "angle": 0, + "content": "{peiiyixuan@stu, qianxm@mail}.xjtu.edu.cn {qzw,wxiang}@hust.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.758, + 0.322 + ], + "angle": 0, + "content": "{zhangjin.zsw, yingya.zyy}@alibaba-inc.com zhaodeli@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.357, + 0.314, + 0.373 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.474, + 0.752 + ], + "angle": 0, + "content": "Recently, prompt-based learning has made impressive progress on image class-incremental learning, but it still lacks sufficient exploration in the video domain. In this paper, we will fill this gap by learning multiple prompts based on a powerful image-language pre-trained model, i.e., CLIP, making it fit for video class-incremental learning (VCIL). For this purpose, we present a space-time prompting approach (ST-Prompt) which contains two kinds of prompts, i.e., task-specific prompts and task-agnostic prompts. The task-specific prompts are to address the catastrophic forgetting problem by learning multi-grained prompts, i.e., spatial prompts, temporal prompts and comprehensive prompts, for accurate task identification. The task-agnostic prompts maintain a globally-shared prompt pool, which can empower the pre-trained image models with temporal perception abilities by exchanging contexts between frames. By this means, ST-Prompt can transfer the plentiful knowledge in the image-language pre-trained models to the VCIL task with only a tiny set of prompts to be optimized. To evaluate ST-Prompt, we conduct extensive experiments on three standard benchmarks. The results show that ST-Prompt can significantly surpass the state-of-the-art VCIL methods, especially it gains \\(9.06\\%\\) on HMDB51 dataset under the \\(1 \\times 25\\) stage setting." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.779, + 0.21, + 0.794 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.804, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Video understanding [49, 11, 1, 32] has achieved outstanding performance with the quick development of deep neural networks. However, training such a model heavily depends on the currently available labeled data. When different classification tasks are presented sequentially in prac" + }, + { + "type": "text", + "bbox": [ + 0.648, + 0.359, + 0.738, + 0.371 + ], + "angle": 0, + "content": "Video Dataset \\(V^k\\)" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.374, + 0.59, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.374, + 0.669, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.4, + 0.709, + 0.408 + ], + "angle": 0, + "content": "···" + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.374, + 0.793, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.374, + 0.872, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.444, + 0.621, + 0.456 + ], + "angle": 0, + "content": "Stored Videos[35]" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.457, + 0.574, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.514, + 0.623, + 0.535 + ], + "angle": 0, + "content": "Memory: 585.70M Accuracy: \\(50.29\\%\\)" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.457, + 0.645, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.668, + 0.444, + 0.764, + 0.456 + ], + "angle": 0, + "content": "Stored Frames[36]" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.461, + 0.716, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.672, + 0.514, + 0.758, + 0.534 + ], + "angle": 0, + "content": "Memory: 585.70M Accuracy: \\(51.43\\%\\)" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.461, + 0.775, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.793, + 0.444, + 0.886, + 0.456 + ], + "angle": 0, + "content": "ST-Prompts (Ours)." + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.462, + 0.814, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.839, + 0.462, + 0.868, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.8, + 0.488, + 0.814, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.839, + 0.488, + 0.868, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.799, + 0.51, + 0.868, + 0.522 + ], + "angle": 0, + "content": "Memory: 2.98M" + }, + { + "type": "text", + "bbox": [ + 0.794, + 0.523, + 0.868, + 0.533 + ], + "angle": 0, + "content": "Accuracy: 55.13" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.542, + 0.895, + 0.613 + ], + "angle": 0, + "content": "Figure 1. An comparison between existing VCIL methods based on exemplar or frame replay and our ST-Prompt. Compared to the rehearsal-based approaches, our ST-prompt can better solve the problem of catastrophic forgetting when only a small amount of memory is consumed to store a tiny set of prompts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.895, + 0.761 + ], + "angle": 0, + "content": "tical applications, a crucial problem is to maintain the accuracy of the model as data arrives over time. Under this circumstance, due to the limitations of memory and privacy, the model can just access the videos of the current task and few even no examples of historical tasks, which requires that the model should adapt to the current task without catastrophic forgetting [34]. The machine learning paradigm to handle the above challenges is called Video Class-Incremental Learning (VCIL) [35, 36, 61]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Most existing class-incremental learning approaches have achieved remarkable performance through a rehearsal buffer of the old tasks in both the image [40, 45, 59, 4] and video [36, 47, 36] fields. However, these methods suffer from suboptimal performance with smaller buffer sizes and severe catastrophic forgetting once the rehearsal buffer is not accessed, hence leading to limited applications. To solve this problem, several prompt-based learning methods in image domain [54, 53, 51] are proposed to train only" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.223, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11932" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.348 + ], + "angle": 0, + "content": "a tiny set of parameters, i.e., prompts, to effectively instruct a pre-trained model to learn incremental tasks without buffering past examples. Typically, L2P [54] learns a single prompt pool to attach complementary prompts to the pre-trained backbone. Nevertheless, this new paradigm has shown great potential for images, but it is still underexplored in the video domain. In this paper, we will concentrate on exploring this paradigm on VCIL to bridge this gap. Directly applying these existing methods for the VCIL task tends to be inadequate. The core reason is that the prompt strategy for images cannot model temporal variations and dynamics across spatial features. Meanwhile, video data possesses a more complicated semantic structure than static images, so a single prompt pool like L2P may be insufficient to learn the abundant semantics. Therefore, we think a properly redesigned prompt framework will contribute to improving the performance of the VCIL task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.473, + 0.77 + ], + "angle": 0, + "content": "Motivated by the above observations, we propose ST-Prompt, a space-time prompting method for video class incremental learning. Specifically, we design ST-Prompt with two components, i.e., task-specific prompts and task-agnostic prompts. The task-specific prompts are tapped by multi-grained prompts, i.e., spatial prompts, temporal prompts and comprehensive prompts, which can better discover the complicated spatio-temporal semantics in videos to address the catastrophic forgetting problem. The task-agnostic prompts learn a complementary prompt for each frame and then gather them back for all frames; hence we can equip the image pre-trained model with temporal modeling capacities without any edition of model structure. Compared with the task-specific prompts, the task-agnostic prompts are task-independent and thus share a generic and globally-shared prompt pool for all tasks. By this means, ST-Prompt has two advantages: 1) it can simultaneously take advantage of the knowledge priors in the image-language pre-trained model and learn the spatiotemporal information, which makes the model more suitable for VCIL; 2) it can markedly reduce the memory cost by just saving a set of prompt parameters without preserving redundant videos anymore, as shown in Figure 1. In the experimental stage, we quantificationally and qualitatively evaluate ST-Prompt on three standard benchmarks. The results show that ST-Prompt can achieve significant gains over the current top-performing approaches, which demonstrates the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.771, + 0.471, + 0.801 + ], + "angle": 0, + "content": "In summary, the main contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.808, + 0.468, + 0.88 + ], + "angle": 0, + "content": "- We first explore the great potential of image-language pre-trained model in video incremental tasks and achieve significant performance improvement by utilizing its sufficient semantic information through prompt-based methods;" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.886, + 0.468, + 0.899 + ], + "angle": 0, + "content": "- We proposed ST-Prompt, containing task-specific" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.808, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "prompts and task-agnostic prompts, which can significantly reduce the memory cost and forgetting rate while modeling temporal variations and dynamics across spatial features without modifying the model structure;" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.18, + 0.892, + 0.223 + ], + "angle": 0, + "content": "- The proposed ST-Prompt achieves remarkable performance on multiple standard action recognition incremental benchmarks over state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.245, + 0.642, + 0.261 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.273, + 0.892, + 0.32 + ], + "angle": 0, + "content": "In this section, we will introduce the methods related to ST-Prompt, including the works for image and video class-incremental learning, and prompt-based learning." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.335, + 0.806, + 0.352 + ], + "angle": 0, + "content": "2.1. Image Class-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.361, + 0.892, + 0.632 + ], + "angle": 0, + "content": "In recent years, class-incremental learning has been widely studied and numerous methods have been proposed in the image domain [52, 56, 10, 43]. These methods can be roughly partitioned into three categories: regularization-based methods, exemplar-based methods and architecture-based methods. Regularization-based methods [23, 60, 21] constrain the optimization of network parameters to preserve the important information of learned tasks, in which the most popular technique utilized is knowledge distillation [29, 12, 14]. The exemplar-based methods [40, 57, 45] typically store a small set of representative exemplars [59, 50] or generated synthetic examples [22, 55] from previous classes to prevent catastrophic forgetting, which is always combined with regularization of a distillation loss to encourage knowledge retention further. The architecture-based methods [26, 42, 48] retain the knowledge of old tasks by designing specific components in the architecture [41, 64] or expanding current feature extractor for new data[58, 38]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.802, + 0.666 + ], + "angle": 0, + "content": "2.2. Video Class-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the video domain, studies on this field are still scarce. The existing solutions are always the combination of regularization-based and exemplar-based methods and focus on how to formulate temporal information better. For example, TCD [35] computes time-channel importance for weighted distillation, [61] decomposes spatiotemporal knowledge transfer for stronger constraint and vCLIMB [47] focuses on temporal-consistency regularization of untrimmed video. Both TCD [35] and vCLIMB [47] have claimed that more stored examples in memory can effectively encourage incremental performance. Specifically, FrameMaker [36] tends to design a memory-efficient video incremental learning method by learning a condensed frame for each representative video of old classes. However, it still demands to preserve some old frames." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "11933" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.087, + 0.895, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.362, + 0.895, + 0.417 + ], + "angle": 0, + "content": "Figure 2. The overall framework of the proposed ST-Prompt, which mainly contains two kinds of prompts, task-specific prompts and task-agnostic prompts. For each input video, we first extract multi-grained features with the image encoder of a CLIP model, then select corresponding prompts from predefined prompt pools. Finally, the prompts are prepended to the video and text embedding feature respectively to instruct the model to learn incremental tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.428, + 0.3, + 0.446 + ], + "angle": 0, + "content": "2.3. Prompt-based Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.473, + 0.641 + ], + "angle": 0, + "content": "Recently, with the rise of large pre-trained models [39, 3, 18, 28], the general pre-trained models are utilized to adapt to multiple down-sampling tasks by finetuning or prompting. The concept of prompt was first proposed in the field of natural language processing [31, 5, 37], and later it was more widely used in language[25, 27, 16] and visual[19, 2, 20] tasks to modify the input texts or images, such that the language or image model obtains additional information about a specific task. Some methods consider multi-modality information simultaneously, and design prompts for the vision-language models[63, 62, 17]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Meanwhile, the emergence of large pre-trained models and prompt-based learning also shows excellent potential in incremental learning. L2P [54], DualPrompt [53] and S- iPrompts [51] exploit prompting method to image incremental learning and achieve good performance without saving old exemplars. Especially, S-Prompting utilizes both image and text prompts together with a pretrained vision-language model and gains outstanding performance in domain-incremental learning. In this paper, we focus on the properties of video, adapting large image pre-trained models to the video domain for video class-incremental learning by space-time prompting. We propose to decompose the complex video features into multiple space-time granularities for accurate task identification while exchanging the temporal information between frames by task-agnostic prompts." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.427, + 0.593, + 0.442 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.453, + 0.895, + 0.515 + ], + "angle": 0, + "content": "In this section, we will first introduce the basic setting of the video class-incremental learning task. Then we discuss the proposed ST-Prompt in detail, as shown in Figure 2. In the end, we present the training and inference procedure." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.528, + 0.704, + 0.542 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.552, + 0.892, + 0.734 + ], + "angle": 0, + "content": "In video class-incremental learning, different tasks \\(\\{\\mathcal{T}^1,\\mathcal{T}^2,\\dots ,\\mathcal{T}^K\\}\\) arrive in a sequence and each task \\(\\mathcal{T}^k\\) corresponds to its specific dataset \\(\\mathcal{D}^k = \\{(V_i^k,y_i^k),y_i^k\\in L^k\\}\\), which contains the input video example \\(V_{i}^{k}\\) and its label \\(y_{i}^{k}\\). The labels in task \\(k\\) belong to a predefined label set \\(L^k\\) and do not appear in previous tasks. The goal of VCIL is to train a single model that has the ability to classify videos of all seen classes correctly. We focus on the more challenging problem of exemplar-free video class-incremental learning, i.e., no exemplar of old classes can be used in the current task, in which the core is to avoid forgetting historical information, termed catastrophic forgetting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this paper, we follow the prompt-based incremental learning paradigm in the image domain [54, 51] to alleviate the forgetting problem, which is composed of two stages: \\(i\\) based on the pre-trained vision-language model like CLIP, they optimize a tiny set of learnable visual and text prompts for each incremental task independently (i.e. task-specific prompts) to learn new knowledge and retain the historical information; \\(ii\\) during inference, due to the lack of task identity of the test example, they select the proper visual and text prompts from the prompt pool by prompt matching, and then preprocess them to the visual embedding \\(V_{ie}^{k}\\) and the class" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11934" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.273 + ], + "angle": 0, + "content": "embedding \\( W_{we}^{k} \\) respectively. The final prediction is determined by the similarities between the visual features and the text features of class names. However, different from static images, videos contain more complicated spatio-temporal clues, and it is intractable to capture these semantic clues well in the video tasks via a single kind of prompt. Meanwhile, the image pre-trained model lacks temporal modeling capability for videos, which requires us to make up for this defect during the prompting process. Overall, the essential point for VCIL is to design a reasonable prompting method to recall the spatio-temporal information in the video task well." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.282, + 0.285, + 0.298 + ], + "angle": 0, + "content": "3.2. Space-time prompting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.47, + 0.351 + ], + "angle": 0, + "content": "The proposed space-time prompting (ST-Prompt) contains two components, i.e., task-specific prompts and task-agnostic prompts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.473, + 0.522 + ], + "angle": 0, + "content": "Task-specific Prompts. To prevent catastrophic forgetting, we train task-specific prompts for each incremental task \\(\\mathcal{T}^k\\) with the pre-trained model (the image encoder \\(f_{v}\\) and text encoder \\(f_{l}\\)) frozen. Specifically, we devote to designing multi-grained prompts to utilize the temporal and spatial variations of different videos for more accurate task identification. Formally, for each task \\(\\mathcal{T}^k\\), there are three different granularities of task-specific prompt pools as Eq. 1. For brevity, we omit the symbol \\(k\\) in this section. If there is no additional explanation, it indicates the situation in incremental task \\(\\mathcal{T}^k\\)." + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.53, + 0.232, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} = \\left\\{\\mathbf {P} _ {s}, \\mathbf {P} _ {t}, \\mathbf {P} _ {c} \\right\\}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.549, + 0.469, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {m} = \\left\\{\\left\\{P _ {m 1} ^ {v}, P _ {m 1} ^ {l} \\right\\}, \\dots , \\left\\{P _ {m N} ^ {v}, P _ {m N _ {m}} ^ {l} \\right\\} \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.57, + 0.424, + 0.591 + ], + "angle": 0, + "content": "\\[\nP _ {m n} ^ {v} \\in \\mathbb {R} ^ {L ^ {v} \\times D ^ {v}}, P _ {m n} ^ {l} \\in \\mathbb {R} ^ {L ^ {l} \\times D ^ {l}}, m \\in \\{s, t, c \\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.719 + ], + "angle": 0, + "content": "where \\(\\mathbf{P}_s\\), \\(\\mathbf{P}_t\\) and \\(\\mathbf{P}_c\\) are the spatial, temporal and comprehensive task-specific prompt pools. Each element in prompt pools is organized as the formulation of a visual-text prompt pair, where \\(P_{mn}^v\\) and \\(P_{mn}^l\\) are the commensurable visual and text prompt respectively. Additionally, \\(N_m\\) is the numbers of prompts in each prompt pool. \\(L^v\\) and \\(L^l\\) are the length of the visual and text prompt, and \\(D^v\\) and \\(D^l\\) are their embedding dimension." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The selection of prompts from these prompt pools is realized through feature matching as Figure 3(a). We apply K-Means algorithm to store the centroids of video features and search for the nearest centroid of the given video sample to identify its relevant prompt. The matching of three kinds of prompts corresponds to features in three different granularities of video \\( V \\), i.e. spatial feature \\( F_{s}(V) \\), temporal feature \\( F_{t}(V) \\) and comprehensive feature \\( F_{c}(V) \\). Specifically, the spatial and temporal features are calculated by pooling the temporal and spatial dimensions separately, while the comprehensive feature is the average of [CLS] token output. After the feature matching, the selected task-specific" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.891, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.264, + 0.892, + 0.335 + ], + "angle": 0, + "content": "Figure 3. The details of prompt matching methods in prompt pools. (a) We select prompts corresponding to the nearest centroid with query feature for the task-specific prompts; (b) We utilize the similarities between the frame feature and centroids as the attention map for the task-agnostic prompts." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.348, + 0.731, + 0.362 + ], + "angle": 0, + "content": "prompts are organized together as:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.371, + 0.892, + 0.39 + ], + "angle": 0, + "content": "\\[\nP ^ {v} = f \\left(P _ {s *} ^ {v}, P _ {t *} ^ {v}, P _ {c *} ^ {v}\\right), P ^ {l} = f \\left(P _ {s *} ^ {l}, P _ {t *} ^ {l}, P _ {c *} ^ {l}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.892, + 0.474 + ], + "angle": 0, + "content": "where \\((P_{s*}^v, P_{s*}^l)\\), \\((P_{t*}^v, P_{t*}^l)\\) and \\((P_{c*}^v, P_{c*}^l)\\) are the prompts selected from three task-specific prompt pools. \\(f\\) is the organization function, which can be stacking, addition, multiplication, or other operations. In this paper, we use addition for its better performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.892, + 0.52 + ], + "angle": 0, + "content": "Finally, the visual and text prompts are prepended to the visual and text embedding respectively, and the whole input of the visual and text encoder is of the format:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.53, + 0.892, + 0.549 + ], + "angle": 0, + "content": "\\[\nV _ {p} = \\left[ P ^ {v}; V _ {e} \\right], W _ {w p} = \\left[ P ^ {l}; W _ {w e} \\right]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.558, + 0.892, + 0.602 + ], + "angle": 0, + "content": "where ; indicates the impending operation, \\(V_{e}\\) is the video embedding feature of video \\(V\\) and \\(W_{we}\\) is the word embedding feature of w-th class name in task \\(\\mathcal{T}^k\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.892, + 0.694 + ], + "angle": 0, + "content": "Additionally, when we utilize the visual pre-trained models without a language branch, such as ViT with ImageNet pre-trained weights, the classification head replaces the text encoder and the prompts in each prompt pool just contain the visual prompt \\( P_{mn}^{v} \\). The other definition and utilization of the task-specific prompts are similar to the above." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Task-agnostic prompts. When tackling video incremental tasks, the image pre-trained model processes each video frame separately and cannot model the temporal variations and relations between different frames. Although we utilize different granularities of prompts for task identification, it is not enough to make the model understand the temporal information of videos well. To enhance the temporal modeling ability and adaptability of image pre-trained models, we propose task-agnostic prompts to highlight the different information contained in frames. Specifically, we establish an additional temporal-modeling prompt pool as \\(\\mathbf{P_{a}} = \\{P_{a1},\\dots,P_{aN_a}\\}\\), where \\(P_{an} \\in \\mathbb{R}^{L^a \\times D^v}\\) is the prompt and \\(N_{a}\\) is the total number of prompts. Compared" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11935" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.151 + ], + "angle": 0, + "content": "with the task-specific prompts, temporal modeling is a more general concept for videos, which is task-agnostic and all prompts from the whole pool can be selected and attached without task limitation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.471, + 0.35 + ], + "angle": 0, + "content": "For the selection of task-agnostic prompts, we also store the centroids of the frame features during training by KMeans algorithm. Differently, in order to utilize the information in the temporal-modeling prompt pool comprehensively, the final prompt according to each frame is not selected by feature matching simply but is calculated by frame-level attention as Figure 3(b). For each frame \\( I_{j} \\) in video \\( V \\), the attention map for all candidate prompts \\( S_{j} = \\{S_{j1}, S_{j2}, \\dots, S_{jNa}\\} \\) is generated according to the similarity between the frame feature \\( f_{v}(I_{j}) \\) and the centroids. Specifically, we utilize the cosine similarity to calculate \\( S_{jn} \\) in this paper. And then the task-agnostic prompt corresponding to the frame \\( I_{j} \\) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.362, + 0.47, + 0.403 + ], + "angle": 0, + "content": "\\[\nP _ {j} = \\sum_ {n = 1} ^ {N _ {a}} \\frac {e ^ {S _ {j n} / \\tau}}{\\sum_ {n = 1} ^ {N _ {a}} e ^ {S _ {j n} / \\tau}} P _ {a n} \\in \\mathbb {R} ^ {D ^ {v}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.469, + 0.52 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the temperature parameter. Finally, the prompts of different frames are stacked together to adapt to the video embedding as \\(V_{p} = [P_{1};P_{2};\\ldots ;P_{T};V_{e}]\\). In this way, each frame will obtain the context information of others. Although they go forward through an image pre-trained model separately, the information exchange of different frames realizes the temporal modeling." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.53, + 0.292, + 0.545 + ], + "angle": 0, + "content": "3.3. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.469, + 0.658 + ], + "angle": 0, + "content": "During training, only the prompts are trainable, while the pre-trained model is fixed. In incremental step \\(\\mathcal{T}^k\\), the training of prompts in task-specific prompt pools is restrained so that only the prompts corresponding to task \\(\\mathcal{T}^k\\) can be selected and utilized. The task-agnostic prompt pool does not have this constraint because temporal modeling is a more general concept independent of task identification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.469, + 0.764 + ], + "angle": 0, + "content": "During inference, for each new video, we first extract multi-grained features by the pre-trained image encoder \\( f_{v} \\), and then select task-specific and task-agnostic prompts from prompt pools through feature matching or frame-level attention. Finally, the selected prompts are prepended to the video and text embedding features respectively. The classification result is according to the similarity output." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.902 + ], + "angle": 0, + "content": "When ST-Prompt is utilized based on a one-branch visual model, in order to separate different tasks better, the classification head is also split according to the task identity and every incremental session trains its own classifier. Differently, without the text branch, the classifier cannot utilize the multi-grained task-specific prompts flexibly and can only be selected by the video-level comprehensive feature. The classifier selection corresponds to the task that the selected comprehensive task-specific prompt belongs to." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.625, + 0.108 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.116, + 0.892, + 0.178 + ], + "angle": 0, + "content": "In this section, we compare our results with baselines and state-of-the-art VCIL methods. We also show ablation studies on the effect of different components and analyze our results in different practical settings." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.189, + 0.673, + 0.206 + ], + "angle": 0, + "content": "4.1. Experiment setup" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.213, + 0.892, + 0.38 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on UCF101 [44], HMDB51 [24] and Something-Something V2 [13], which are the standard action recognition datasets. We follow the video frequently-used class-incremental benchmarks presented in TCD [35]. For UCF101, the model is trained on 51 classes first, and the other 50 classes are divided into 5, 10 and 25 tasks. For HMDB51, we train the base model using videos from 26 classes, and the rest of the classes are separated into 5 or 25 groups. For Something-Something V2, we first train 84 classes in the initial stage and generate the incremental groups of 10 and 5 classes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.403, + 0.892, + 0.464 + ], + "angle": 0, + "content": "Evaluation Protocol. After each incremental step, we evaluate the model on all seen classes and the class-incremental methods are finally assessed by the average accuracy of all tasks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.487, + 0.892, + 0.669 + ], + "angle": 0, + "content": "Implementation Details. Our method is based on CLIP ViT-B/16 [39], a visual-language model pre-trained with numerous image-text pairs. While the pre-trained model is frozen, the visual and textual prompts are optimized using a batch size of 64 and an initial learning rate of 0.001 with cosine descending for 50 epochs in each task. We carefully choose compared methods in the same environment for a fair comparison. However, most existing class-incremental learning methods are not adaptive to visual-language models. Therefore, we also apply our approach on ImageNet pre-trained ViT-B/16 [8] and reproduce other class-incremental learning methods on the same backbone." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.642, + 0.696 + ], + "angle": 0, + "content": "4.2. Main Results." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We compare our proposed ST-Prompt with existing state-of-the-art class-incremental learning approaches under multiple challenging settings on three datasets in Table 1 and Table 2. For the approaches in the image domain, we reproduce them by treating the video as a whole as that in the image field. From the tables, we can draw the following conclusions. First, ST-Prompt notably outperforms existing rehearsal-free methods in VCIL, containing two recent prompt-based approaches, L2P and S-iPrompts in the image domain. With ImageNet pre-trained ViT, ST-Prompt respectively surpasses S-iPrompts by around \\(4.22\\%\\), \\(4.98\\%\\) and \\(7.27\\%\\) on HMDB51, UCF101 and Something-something v2, respectively. This indicates that" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11936" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.871, + 0.53 + ], + "angle": 0, + "content": "
MethodsBuffer SizeUCF101HMDB51
10 × 5 stages5 × 10 stages2 × 25 stages5 × 5 stages1 × 25 stages
ImageNet pre-trained ViT-B/16
iCaRL [40]70.5869.5167.2843.9037.15
UCIR [15]77.5574.5971.7748.2039.42
PODNet [9]76.5076.1773.8348.3843.35
Co2L [6]10/class78.5477.1375.5949.7644.15
TCD†[35]78.1376.8775.7450.2944.04
FrameMaker†[36]79.3779.5579.3251.4346.37
L2P [54]81.2480.0978.5849.9845.87
iCaRL [40]66.5867.7064.2741.0836.13
UCIR [15]75.5572.7866.7645.3838.40
PODNet [9]74.5074.3672.8245.5642.33
Co2L [6]5/class77.5776.3274.5847.1043.56
TCD†[35]76.1375.0673.7348.3843.02
FrameMaker†[36]78.3777.7476.3150.5845.65
L2P [54]79.9377.7876.2747.6742.56
Finetuning24.8313.616.2418.304.89
Finetuning-frozen67.1065.717.3936.644.95
LwFMC [29]45.7328.0116.4327.0117.21
LwM [7]0/class46.9828.7016.9227.9417.85
L2P [54]76.7275.1173.2645.0039.89
S-iPrompts [51]77.1677.8377.9950.6751.45
ST-Prompt†83.7481.5382.6655.1355.43
CLIP ViT-B/16
Zero shot [39]69.6869.6169.8740.1940.42
S-liPrompts [51]0/class80.6080.2780.4353.1153.89
ST-Prompt†84.7585.5485.6760.1460.54
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.531, + 0.892, + 0.561 + ], + "angle": 0, + "content": "Table 1. Comparison with the state-of-the-art approaches over class-incremental action recognition performance on UCF101 and HMDB51. Compared methods are grouped based on different rehearsal buffer sizes. † indicates the methods designed for video." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.577, + 0.473, + 0.82 + ], + "angle": 0, + "content": "our design of multi-grained task-specific prompts and task-agnostic prompts plays an essential role in video feature formulation and temporal perception. The more tremendous increase on Ssv2 also claims the effectiveness of ST-Prompt for the motion-sensitive dataset. Second, compared with rehearsal-based methods, our ST-Prompt also has a significant advantage. ST-Prompt achieves better performance than the top existing approach, FrameMaker, by approximately \\(6.38\\%\\) and \\(3.23\\%\\) on HMDB51 and UCF101, separately, although FrameMaker stores additional 80 examples for each class. Third, using the pre-trained vision-language model, CLIP, as the backbone for VCIL, the performance of ST-Prompt improves further by about \\(5.06\\%\\), \\(2.67\\%\\) and \\(3.49\\%\\) respectively. The reason is that CLIP provides more abundant prior knowledge about semantic information, which is vital for subsequent prompting." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.831, + 0.236, + 0.848 + ], + "angle": 0, + "content": "4.3. Ablation Study." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this section, we present ablation studies on the properties and effectiveness of our core designs. If not specified, the ablation studies are performed on HMDB51 with 5 steps" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.577, + 0.684, + 0.593 + ], + "angle": 0, + "content": "and UCF101 with 10 steps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Task-specific Prompts. In Table 3, we first explore how to organize different kinds of task-specific prompts together in Eq. 2. The results show that the addition operation achieves better performance than stacking and multiplication, whose reason may be that different prompts can modulate each other better without preponding too many parameters to the embedding feature in this way." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Then, to prove the effectiveness of different parts of our task-specific prompts, we remove each kind of prompts respectively under the combination method of addition and show their results in Table 4. From the results, we can find that: first, when three kinds of prompts are utilized separately, the comprehensive task-specific prompts achieve the best performance, which demonstrates that they depict the video feature more integrally; second, when the comprehensive prompts are combined with the temporal or spatial prompts, the performance increases by \\(1.9\\%\\) or \\(1.54\\%\\) on HMDB51, and \\(1.26\\%\\) or \\(1.34\\%\\) on UCF101, respectively. Differentiated promotions in different datasets imply that" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11937" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.089, + 0.46, + 0.388 + ], + "angle": 0, + "content": "
MethodsBuffer Size10 × 9 Stages5 × 18 Stages
ImageNet pre-trained ViT-B/16
iCaRL [40]20.4116.62
UCIR [15]24.3219.31
PODNet [9]5/class27.6320.14
\\( \\mathrm {TCD}^{\\dagger } \\)[35]29.3224.69
FrameMaker \\( {}^{ \\dagger } \\)[36]31.4126.57
L2P [54]26.0221.33
Finetuning11.046.23
Finetuning-frozen7.995.97
LwFMC [29]0/class15.2111.41
L2P [54]23.2518.49
S-iPrompts [51]30.5826.33
ST-Prompt36.8431.60
CLIP ViT-B/16
Zero-shot3.563.57
S-liPrompts [51]0/class33.6930.84
\\( \\mathrm {ST}-\\mathrm {Prompt}^{\\dagger } \\)39.9835.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.388, + 0.468, + 0.415 + ], + "angle": 0, + "content": "Table 2. Comparison with the top approaches on Something-something v2.† indicates the methods designed for video." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.432, + 0.452, + 0.492 + ], + "angle": 0, + "content": "
OperationStackingMultiplicationAddition
HMDB5155.2454.9656.02
UCF10182.8282.4583.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.493, + 0.468, + 0.508 + ], + "angle": 0, + "content": "Table 3. Ablation for the ensemble ways of task-specific prompts." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.521, + 0.428, + 0.657 + ], + "angle": 0, + "content": "
STPTTPCTPHMDB51UCF101
✓✘✘44.6272.45
✘✓✘45.4673.84
✘✘✓53.1180.27
✓✓✘49.3476.93
✓✘✓54.6581.61
✘✓✓55.0181.53
✓✓✓56.0283.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.657, + 0.47, + 0.699 + ], + "angle": 0, + "content": "Table 4. Ablations for task-specific prompts on HMDB51 with 5 steps and UCF101 with 10 steps. STP, TTP and CTP represent spatial, temporal and comprehensive prompts respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.468, + 0.793 + ], + "angle": 0, + "content": "temporal information is more critical for HMDB51, while the spatial feature for UCF101; third, utilizing all three kinds of task-specific prompts can further improve the final accuracy, which proves the combination of multi-grained prompts contribute to more correct task identification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Further, we visualize some clustering results of different granularities of the video feature in Figure 4. The comprehensive feature judges video task-id from the overall perspective, and the clustered videos usually belong to the same spatial and temporal clusters, such as \"Smile\". The temporal and spatial prompts implement the local information for more complicated video task matching. Videos" + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.089, + 0.857, + 0.239 + ], + "angle": 0, + "content": "
MethodTAPHMDB51UCF101
TSPX56.0283.58
TSP+LEX56.0683.55
TSP+SPX57.0483.87
TSP+CPX58.3084.05
TSP+PE [46]X57.9684.04
TSP+SH [33]X58.2784.25
TSP+TS [30]X58.7484.34
ST-Prompt✓60.1485.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.24, + 0.892, + 0.281 + ], + "angle": 0, + "content": "Table 5. Comparison with some baseline methods about task-agnostic prompts (TAP) on HMDB51 with 5 steps and UCF101 with 10 steps." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.292, + 0.868, + 0.442 + ], + "angle": 0, + "content": "
MethodTAPHMDB51UCF101
L2P [54]✘45.0075.11
✓46.7676.07
S-iPrompts [51]✘50.6777.83
✓54.4479.74
S-liPrompts [51]✘53.1180.27
✓55.7781.82
ST-Prompt✘56.0283.58
✓60.1485.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.443, + 0.891, + 0.471 + ], + "angle": 0, + "content": "Table 6. Ablations for task-agnostic prompts (TAP) on HMDB51 with 5 steps and UCF101 with 10 steps." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.498, + 0.892, + 0.573 + ], + "angle": 0, + "content": "with the same motion but different spatial features are gathered together according to their temporal features, such as \"Dive\" in different scenarios, while spatial features distinguishes actions with mixed temporal information but obvious spatial similarities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.884 + ], + "angle": 0, + "content": "Task-agnostic Prompts. We compare our task-agnostic prompts for temporal modeling with a set of baselines in the proposed framework in Table 5. \"LE\" represents prompt length expansion, which excludes the performance improvement caused by increasing the number of parameters. And the results also prove that learning task-agnostic prompts about spatial (\"SP\") or comprehensive (\"CP\") information are not as effective as temporal modeling. Then we compare the effect of our temporal-modeling prompts with some other parameter-free temporal modeling methods, such as positional encoding (\"PE\"), temporal shuffle (\"SH\") and temporal shift (\"TS\"). We can observe that utilizing task-agnostic prompts to exchange the context of different frames can gain the best performance of \\(60.14\\%\\). Further, we validate the effectiveness of task-agnostic prompts by adding them to other prompt-based methods in the image domain, as shown in Table 6, which indicates that it is a general approach to improve the ability of temporal perception for image pre-trained models." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In addition, We depict the change of category-wise ac" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11938" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.091, + 0.861, + 0.174 + ], + "angle": 0, + "content": "
MethodsBackboneStored ExemplarsMemory CostAverage Accuracy
TCD [35]ImageNet Pre-trained ViT-B/1610 examples/class585.70M50.29
FrameMaker [36]ImageNet Pre-trained ViT-B/1680 examples/class585.70M51.43
ST-PromptImageNet Pre-trained ViT-B/1602.98M55.13
ST-PromptCLIP ViT-B/1604.47M60.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.153, + 0.173, + 0.812, + 0.187 + ], + "angle": 0, + "content": "Table 7. The comparison of the overhead of our ST-Prompt and other VCIL methods on HMDB51 with 5 steps." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.214, + 0.468, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.34, + 0.468, + 0.367 + ], + "angle": 0, + "content": "Figure 4. The visualization of multi-grained video feature clustering results." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.368, + 0.465, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.5, + 0.468, + 0.528 + ], + "angle": 0, + "content": "Figure 5. The change of category-wise accuracy after utilizing task-agnostic prompts on HMDB51 with 5 steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.468, + 0.63 + ], + "angle": 0, + "content": "curacy after utilizing task-agnostic prompts in Figure 5. The significant increase in the accuracy of some motion-sensitive classes, such as \"Dive\" and \"Somersault\", proves the role of our task-agnostic prompts in improving the temporal modeling capability of models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.468, + 0.733 + ], + "angle": 0, + "content": "The Temperature \\(\\tau\\). The temperature \\(\\tau\\) in Eq. 4 decides the smoothness of the attention weights. Too small \\(\\tau\\) makes the final prompt of each frame lose the understanding of global information, and too large \\(\\tau\\) introduces noise. From Figure 6(a), we observe that \\(\\tau = 0.3\\) guides the best performance of final average accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The Length of Prompts. The length of different kinds of prompts \\( L^v, L^t, L^a \\) decides the capacity of each learnable prompt. Figure 6(b) illustrates how the average accuracy changes when the length of different prompts varies. It is evident that a too-small visual or text prompt length always negatively affects results, while an oversized prompt may cause underfitting. We set the prompt length as 5 for both visual and text task-specific prompts according to the curves. In addition, we can discover that the final accuracy is not sensitive to the length of task-agnostic prompts" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.218, + 0.895, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.347, + 0.865, + 0.361 + ], + "angle": 0, + "content": "Figure 6. The effects of some hyperparameters in ST-Prompt." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.387, + 0.892, + 0.418 + ], + "angle": 0, + "content": "and even decreases during the length is larger. Hence, we choose the length of the task-agnostic prompts as 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.892, + 0.598 + ], + "angle": 0, + "content": "The Number of Prompts. Figure 6(c) presents how the number of different prompts \\( N_{m}(m \\in \\{s, t, c, a\\}) \\) in each incremental task influences the performance of our approach, where the number in the base task is calculated according to the proportion of the number of classes. Three different task-specific prompts and task-agnostic prompts are explored separately. In general, the more prompts per task, i.e., the more pre-defined centroids per task during clustering, the higher accuracy gained. Especially, the comprehensive prompt is more sensitive to the cluster number. For simplicity, the number of different prompts is set as 5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.609, + 0.681, + 0.625 + ], + "angle": 0, + "content": "4.4. Overhead Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.634, + 0.892, + 0.768 + ], + "angle": 0, + "content": "We compare the memory cost and average accuracy of ST-Prompt with other VCIL methods in Table 7. Our ST-Prompt utilizes only a set of prompts and prompt keys in each task, while previous rehearsal-based methods store a number of video examples or frames. Therefore, ST-Prompt has significant memory advantages over previous VCIL methods. Our ST-Prompt achieves performance increasing by around \\(8.71\\%\\) with only \\(0.76\\%\\) memory cost of other methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.62, + 0.8 + ], + "angle": 0, + "content": "5. Discussions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Limitation. Compared with the previous video incremental learning approaches, ST-Prompt achieves a significant performance boost without any additional memory cost for video example saving. Nevertheless, ST-Prompt relies on the large pre-trained model, which may fail when the knowledge contained in it is not sufficient. Meanwhile, our" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11939" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.138 + ], + "angle": 0, + "content": "ST-Prompt prepends prompts to the embedding feature and has not discussed the phenomenon that prompts are prefixed to the hidden layers." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.148, + 0.473, + 0.376 + ], + "angle": 0, + "content": "Conclusion. In this paper, we propose ST-Prompt, a novel space-time prompting framework for video class incremental learning. It explores how to address the VCIL problem in a rehearsal-free way by learning multiple prompts based on a powerful image-language pre-trained model. ST-Prompt mainly consists of two components, task-specific prompts and task-agnostic prompts. The task-specific prompts tackle the catastrophic forgetting problem by learning multi-grained prompts for task identification. The task-agnostic prompts implement the temporal modeling ability of the image pre-trained model. In this way, ST-Prompt achieves a new state-of-the-art performance, which has already surpassed the methods with a memory buffer. We hope this simple and practical framework can inspire attention to prompting-based methods in VCIL." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.397, + 0.473, + 0.473 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported in part by the NSFC under Grant 62272380 and 62103317, the Science and Technology Program of Xi'an, China under Grant 21RGZN0017, and by Alibaba Group through Alibaba Innovative Research Program." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.488, + 0.175, + 0.504 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.513, + 0.47, + 0.554 + ], + "angle": 0, + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.557, + 0.47, + 0.611 + ], + "angle": 0, + "content": "[2] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Visual prompting: Modifying pixel space to adapt pre-trained models. arXiv preprint arXiv:2203.17274, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.614, + 0.47, + 0.642 + ], + "angle": 0, + "content": "[3] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.644, + 0.469, + 0.697 + ], + "angle": 0, + "content": "[4] Zalán Borsos, Mojmir Mutny, and Andreas Krause. Coresetts via bilevel optimization for continual learning and streaming. Advances in Neural Information Processing Systems, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.701, + 0.47, + 0.757 + ], + "angle": 0, + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.758, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[6] Hyuntak Cha, Jaeho Lee, and Jinwoo Shin. Co2l: Contrastive continual learning. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.789, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[7] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In CVPR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.513, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[9] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[10] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[11] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.238, + 0.892, + 0.279 + ], + "angle": 0, + "content": "[12] Tao Feng, Mang Wang, and Hangjie Yuan. Overcoming catastrophic forgetting in incremental object detection via elastic response distillation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.892, + 0.363 + ], + "angle": 0, + "content": "[13] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In ICCV, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.366, + 0.892, + 0.408 + ], + "angle": 0, + "content": "[14] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Lifelong learning via progressive distillation and retrospection. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.41, + 0.892, + 0.452 + ], + "angle": 0, + "content": "[15] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.454, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[16] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[17] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.555, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[18] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.613, + 0.892, + 0.654 + ], + "angle": 0, + "content": "[19] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.656, + 0.892, + 0.698 + ], + "angle": 0, + "content": "[20] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.7, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[21] Minsoo Kang, Jaeyoo Park, and Bohyung Han. Class incremental learning by knowledge distillation with adaptive feature consolidation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[22] Ronald Kemker and Christopher Kanan. Fearnet: Brain-inspired model for incremental learning. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[23] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Hildegard Kuehne, Hueihan Jhuang, Estibaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, 2011. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "11940" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[25] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[26] Xilai Li, Yingbo Zhou, Tianfu Wu, Richard Socher, and Caiming Xiong. Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In ICML, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[27] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In ACL-IJCNLP, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.469, + 0.29 + ], + "angle": 0, + "content": "[28] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.292, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[29] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 2017. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.335, + 0.469, + 0.362 + ], + "angle": 0, + "content": "[30] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In ICCV, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.364, + 0.469, + 0.418 + ], + "angle": 0, + "content": "[31] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Comput. Surv., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.421, + 0.469, + 0.46 + ], + "angle": 0, + "content": "[32] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.463, + 0.469, + 0.504 + ], + "angle": 0, + "content": "[33] Pingchuan Ma, Yao Zhou, Yu Lu, and Wei Zhang. Learning efficient video representation with video shuffle networks. arXiv preprint arXiv:1911.11319, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.506, + 0.469, + 0.547 + ], + "angle": 0, + "content": "[34] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation. 1989. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.549, + 0.469, + 0.589 + ], + "angle": 0, + "content": "[35] Jaeyoo Park, Minsoo Kang, and Bohyung Han. Class incremental learning for action recognition in videos. In ICCV, 2021. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.591, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[36] Yixuan Pei, Zhiwu Qing, Jun CEN, Xiang Wang, Shiwei Zhang, Yaxiong Wang, Mingqian Tang, Nong Sang, and Xueming Qian. Learning a condensed frame for memory-efficient video class-incremental learning. In NeurIPS, 2022. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.662, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[37] Fabio Petroni, Tim Rocktäschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and Sebastian Riedel. Language models as knowledge bases? In EMNLP, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.719, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[38] Qi Qin, Wenpeng Hu, Han Peng, Dongyan Zhao, and Bing Liu. Bns: Building network structures dynamically for continual learning. Advances in Neural Information Processing Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.776, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[40] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017. 1, 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[41] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[42] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In ICML, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.209, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[43] Yujun Shi, Kuangqi Zhou, Jian Liang, Zihang Jiang, Jiashi Feng, Philip HS Torr, Song Bai, and Vincent YF Tan. Mimicking the oracle: An initial phase decorrelation approach for class incremental learning. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.267, + 0.892, + 0.307 + ], + "angle": 0, + "content": "[44] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.31, + 0.892, + 0.351 + ], + "angle": 0, + "content": "[45] Rishabh Tiwari, Krishnateja Killamsetty, Rishabh Iyer, and Pradeep Shenoy. Gcr: Gradient coreset based replay buffer selection for continual learning. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.354, + 0.892, + 0.409 + ], + "angle": 0, + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.412, + 0.892, + 0.465 + ], + "angle": 0, + "content": "[47] Andrés Villa, Kumail Alhamoud, Victor Escorcia, Fabian Caba, Juan León Alcazar, and Bernard Ghanem. vclimb: A novel video class incremental learning benchmark. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[48] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.513, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[49] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In ECCV, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.571, + 0.892, + 0.626 + ], + "angle": 0, + "content": "[50] Liyuan Wang, Xingxing Zhang, Kuo Yang, Longhui Yu, Chongxuan Li, Lanqing Hong, Shifeng Zhang, Zhenguo Li, Yi Zhong, and Jun Zhu. Memory replay with data compression for continual learning. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.629, + 0.892, + 0.67 + ], + "angle": 0, + "content": "[51] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. In NeurIPS, 2022. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.673, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[52] Zhen Wang, Liu Liu, Yiqun Duan, Yajing Kong, and Dacheng Tao. Continual learning with lifelong vision transformer. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.716, + 0.892, + 0.784 + ], + "angle": 0, + "content": "[53] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In ECCV, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[54] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In CVPR, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[55] Chenshen Wu, Luis Herranz, Xialei Liu, Joost van de Weijer, Bogdan Raducanu, et al. Memory replay gans: Learning to generate new categories without forgetting. In NeurIPS, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11941" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[56] Tz-Ying Wu, Gurumurthy Swaminathan, Zhizhong Li, Avinash Ravichandran, Nuno Vasconcelos, Rahul Bhotika, and Stefano Soatto. Class-incremental learning with strong pre-trained models. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[57] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[58] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.469, + 0.276 + ], + "angle": 0, + "content": "[59] Jaehong Yoon, Divyam Madaan, Eunho Yang, and Sung Ju Hwang. Online coreset selection for rehearsal-based continual learning. arXiv preprint arXiv:2106.01085, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.317 + ], + "angle": 0, + "content": "[60] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[61] Hanbin Zhao, Xin Qin, Shihao Su, Yongjian Fu, Zibo Lin, and Xi Li. When video classification meets incremental classes. In ACM MM, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[62] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[63] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.49 + ], + "angle": 0, + "content": "[64] Kai Zhu, Wei Zhai, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Self-sustaining representation expansion for nonexemplar class-incremental learning. In CVPR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11942" + } + ] +] \ No newline at end of file diff --git a/2023/Space-time Prompting for Video Class-incremental Learning/23f1a364-1b5a-487e-9a38-2d1ef0112aa7_origin.pdf b/2023/Space-time Prompting for Video Class-incremental Learning/23f1a364-1b5a-487e-9a38-2d1ef0112aa7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..82aff74f83578f1e83c6533af5cb9a263e13768c --- /dev/null +++ b/2023/Space-time Prompting for Video Class-incremental Learning/23f1a364-1b5a-487e-9a38-2d1ef0112aa7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa9b59251379be74172d4717fdd85655682d2938b4dafdb70677bb1dc3e34435 +size 1867489 diff --git a/2023/Space-time Prompting for Video Class-incremental Learning/full.md b/2023/Space-time Prompting for Video Class-incremental Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5c7c3e7a2bd180c2f3daaa7cc1d82fd162a3313a --- /dev/null +++ b/2023/Space-time Prompting for Video Class-incremental Learning/full.md @@ -0,0 +1,342 @@ +# Space-time Prompting for Video Class-incremental Learning + +Yixuan Pei $^{1}$ Zhiwu Qing $^{2}$ Shiwei Zhang $^{3*}$ Xiang Wang $^{2}$ Yingya Zhang $^{3}$ Deli Zhao $^{3}$ Xueming Qian $^{1,4*}$ + +$^{1}$ Xi'an Jiaotong University + +$^{2}$ Huazhong University of Science and Technology $^{3}$ Alibaba Group, Shaanxi Yulan Jiuzhou Intelligent Optoelectronic Technology Co., Ltd. + +{peiiyixuan@stu, qianxm@mail}.xjtu.edu.cn {qzw,wxiang}@hust.edu.cn + +{zhangjin.zsw, yingya.zyy}@alibaba-inc.com zhaodeli@gmail.com + +# Abstract + +Recently, prompt-based learning has made impressive progress on image class-incremental learning, but it still lacks sufficient exploration in the video domain. In this paper, we will fill this gap by learning multiple prompts based on a powerful image-language pre-trained model, i.e., CLIP, making it fit for video class-incremental learning (VCIL). For this purpose, we present a space-time prompting approach (ST-Prompt) which contains two kinds of prompts, i.e., task-specific prompts and task-agnostic prompts. The task-specific prompts are to address the catastrophic forgetting problem by learning multi-grained prompts, i.e., spatial prompts, temporal prompts and comprehensive prompts, for accurate task identification. The task-agnostic prompts maintain a globally-shared prompt pool, which can empower the pre-trained image models with temporal perception abilities by exchanging contexts between frames. By this means, ST-Prompt can transfer the plentiful knowledge in the image-language pre-trained models to the VCIL task with only a tiny set of prompts to be optimized. To evaluate ST-Prompt, we conduct extensive experiments on three standard benchmarks. The results show that ST-Prompt can significantly surpass the state-of-the-art VCIL methods, especially it gains $9.06\%$ on HMDB51 dataset under the $1 \times 25$ stage setting. + +# 1. Introduction + +Video understanding [49, 11, 1, 32] has achieved outstanding performance with the quick development of deep neural networks. However, training such a model heavily depends on the currently available labeled data. When different classification tasks are presented sequentially in prac + +Video Dataset $V^k$ + +![](images/5d816b0a227aea24bbdb09c06be3f4bfee0cae40ffe7f107cbb89183079091af.jpg) + +![](images/a3cee9f31a3ea15b508da7c6b0ebee738560c92788704146eb797ad583e95652.jpg) +··· + +![](images/334d3c7ca92d3366b026260b7325fe423ec8dc67eb5a44fe87353873c515fe90.jpg) + +![](images/5ac72593e7509c1b7e6f17659d9517856af9dbf2e6a00755d24843001203db69.jpg) + +Stored Videos[35] + +![](images/af8a7bf617c85b5a197f881858441ea6e808ae59079a44f9a34fd81fe6340f04.jpg) +Memory: 585.70M Accuracy: $50.29\%$ + +![](images/996c38aa2bca57398f23ea538e0c80b152592940d4214de3b899efa7387355be.jpg) + +Stored Frames[36] + +![](images/b33c9ffaebaec423ace6f283ab4876d5fa3e1399743f2194d87b17ac597ec1da.jpg) +Memory: 585.70M Accuracy: $51.43\%$ + +![](images/d1109a9feafa61fa3ff70c0e94f409a269186470ae36c99010545ac9324b93dd.jpg) +Figure 1. An comparison between existing VCIL methods based on exemplar or frame replay and our ST-Prompt. Compared to the rehearsal-based approaches, our ST-prompt can better solve the problem of catastrophic forgetting when only a small amount of memory is consumed to store a tiny set of prompts. + +ST-Prompts (Ours). + +![](images/b4cfe96281ae8d0e6a6ef6dde0c6536e11bb471c1d381395662c2539ff752915.jpg) + +![](images/7cee4d031d908eba99da31649bf9b77e77a359d8fa7af93858889c8268fd2d76.jpg) + +![](images/f9a668d115558f627839feed4518143f8a62ea0dde28dd68c1c55dc226cc0702.jpg) + +![](images/320bcc749e7a10ac9ba647a91a32872d45339871609a726063ee3e8ed24485cf.jpg) + +Memory: 2.98M + +Accuracy: 55.13 + +tical applications, a crucial problem is to maintain the accuracy of the model as data arrives over time. Under this circumstance, due to the limitations of memory and privacy, the model can just access the videos of the current task and few even no examples of historical tasks, which requires that the model should adapt to the current task without catastrophic forgetting [34]. The machine learning paradigm to handle the above challenges is called Video Class-Incremental Learning (VCIL) [35, 36, 61]. + +Most existing class-incremental learning approaches have achieved remarkable performance through a rehearsal buffer of the old tasks in both the image [40, 45, 59, 4] and video [36, 47, 36] fields. However, these methods suffer from suboptimal performance with smaller buffer sizes and severe catastrophic forgetting once the rehearsal buffer is not accessed, hence leading to limited applications. To solve this problem, several prompt-based learning methods in image domain [54, 53, 51] are proposed to train only + +a tiny set of parameters, i.e., prompts, to effectively instruct a pre-trained model to learn incremental tasks without buffering past examples. Typically, L2P [54] learns a single prompt pool to attach complementary prompts to the pre-trained backbone. Nevertheless, this new paradigm has shown great potential for images, but it is still underexplored in the video domain. In this paper, we will concentrate on exploring this paradigm on VCIL to bridge this gap. Directly applying these existing methods for the VCIL task tends to be inadequate. The core reason is that the prompt strategy for images cannot model temporal variations and dynamics across spatial features. Meanwhile, video data possesses a more complicated semantic structure than static images, so a single prompt pool like L2P may be insufficient to learn the abundant semantics. Therefore, we think a properly redesigned prompt framework will contribute to improving the performance of the VCIL task. + +Motivated by the above observations, we propose ST-Prompt, a space-time prompting method for video class incremental learning. Specifically, we design ST-Prompt with two components, i.e., task-specific prompts and task-agnostic prompts. The task-specific prompts are tapped by multi-grained prompts, i.e., spatial prompts, temporal prompts and comprehensive prompts, which can better discover the complicated spatio-temporal semantics in videos to address the catastrophic forgetting problem. The task-agnostic prompts learn a complementary prompt for each frame and then gather them back for all frames; hence we can equip the image pre-trained model with temporal modeling capacities without any edition of model structure. Compared with the task-specific prompts, the task-agnostic prompts are task-independent and thus share a generic and globally-shared prompt pool for all tasks. By this means, ST-Prompt has two advantages: 1) it can simultaneously take advantage of the knowledge priors in the image-language pre-trained model and learn the spatiotemporal information, which makes the model more suitable for VCIL; 2) it can markedly reduce the memory cost by just saving a set of prompt parameters without preserving redundant videos anymore, as shown in Figure 1. In the experimental stage, we quantificationally and qualitatively evaluate ST-Prompt on three standard benchmarks. The results show that ST-Prompt can achieve significant gains over the current top-performing approaches, which demonstrates the effectiveness of our method. + +In summary, the main contributions of this paper are as follows: + +- We first explore the great potential of image-language pre-trained model in video incremental tasks and achieve significant performance improvement by utilizing its sufficient semantic information through prompt-based methods; +- We proposed ST-Prompt, containing task-specific + +prompts and task-agnostic prompts, which can significantly reduce the memory cost and forgetting rate while modeling temporal variations and dynamics across spatial features without modifying the model structure; + +- The proposed ST-Prompt achieves remarkable performance on multiple standard action recognition incremental benchmarks over state-of-the-art methods. + +# 2. Related Work + +In this section, we will introduce the methods related to ST-Prompt, including the works for image and video class-incremental learning, and prompt-based learning. + +# 2.1. Image Class-Incremental Learning + +In recent years, class-incremental learning has been widely studied and numerous methods have been proposed in the image domain [52, 56, 10, 43]. These methods can be roughly partitioned into three categories: regularization-based methods, exemplar-based methods and architecture-based methods. Regularization-based methods [23, 60, 21] constrain the optimization of network parameters to preserve the important information of learned tasks, in which the most popular technique utilized is knowledge distillation [29, 12, 14]. The exemplar-based methods [40, 57, 45] typically store a small set of representative exemplars [59, 50] or generated synthetic examples [22, 55] from previous classes to prevent catastrophic forgetting, which is always combined with regularization of a distillation loss to encourage knowledge retention further. The architecture-based methods [26, 42, 48] retain the knowledge of old tasks by designing specific components in the architecture [41, 64] or expanding current feature extractor for new data[58, 38]. + +# 2.2. Video Class-Incremental Learning + +In the video domain, studies on this field are still scarce. The existing solutions are always the combination of regularization-based and exemplar-based methods and focus on how to formulate temporal information better. For example, TCD [35] computes time-channel importance for weighted distillation, [61] decomposes spatiotemporal knowledge transfer for stronger constraint and vCLIMB [47] focuses on temporal-consistency regularization of untrimmed video. Both TCD [35] and vCLIMB [47] have claimed that more stored examples in memory can effectively encourage incremental performance. Specifically, FrameMaker [36] tends to design a memory-efficient video incremental learning method by learning a condensed frame for each representative video of old classes. However, it still demands to preserve some old frames. + +![](images/342b993ae8f3893291718be0a838aea9be2e10da319578900b771c2cb5effa43.jpg) +Figure 2. The overall framework of the proposed ST-Prompt, which mainly contains two kinds of prompts, task-specific prompts and task-agnostic prompts. For each input video, we first extract multi-grained features with the image encoder of a CLIP model, then select corresponding prompts from predefined prompt pools. Finally, the prompts are prepended to the video and text embedding feature respectively to instruct the model to learn incremental tasks. + +# 2.3. Prompt-based Learning + +Recently, with the rise of large pre-trained models [39, 3, 18, 28], the general pre-trained models are utilized to adapt to multiple down-sampling tasks by finetuning or prompting. The concept of prompt was first proposed in the field of natural language processing [31, 5, 37], and later it was more widely used in language[25, 27, 16] and visual[19, 2, 20] tasks to modify the input texts or images, such that the language or image model obtains additional information about a specific task. Some methods consider multi-modality information simultaneously, and design prompts for the vision-language models[63, 62, 17]. + +Meanwhile, the emergence of large pre-trained models and prompt-based learning also shows excellent potential in incremental learning. L2P [54], DualPrompt [53] and S- iPrompts [51] exploit prompting method to image incremental learning and achieve good performance without saving old exemplars. Especially, S-Prompting utilizes both image and text prompts together with a pretrained vision-language model and gains outstanding performance in domain-incremental learning. In this paper, we focus on the properties of video, adapting large image pre-trained models to the video domain for video class-incremental learning by space-time prompting. We propose to decompose the complex video features into multiple space-time granularities for accurate task identification while exchanging the temporal information between frames by task-agnostic prompts. + +# 3. Method + +In this section, we will first introduce the basic setting of the video class-incremental learning task. Then we discuss the proposed ST-Prompt in detail, as shown in Figure 2. In the end, we present the training and inference procedure. + +# 3.1. Problem Formulation + +In video class-incremental learning, different tasks $\{\mathcal{T}^1,\mathcal{T}^2,\dots ,\mathcal{T}^K\}$ arrive in a sequence and each task $\mathcal{T}^k$ corresponds to its specific dataset $\mathcal{D}^k = \{(V_i^k,y_i^k),y_i^k\in L^k\}$ , which contains the input video example $V_{i}^{k}$ and its label $y_{i}^{k}$ . The labels in task $k$ belong to a predefined label set $L^k$ and do not appear in previous tasks. The goal of VCIL is to train a single model that has the ability to classify videos of all seen classes correctly. We focus on the more challenging problem of exemplar-free video class-incremental learning, i.e., no exemplar of old classes can be used in the current task, in which the core is to avoid forgetting historical information, termed catastrophic forgetting. + +In this paper, we follow the prompt-based incremental learning paradigm in the image domain [54, 51] to alleviate the forgetting problem, which is composed of two stages: $i$ based on the pre-trained vision-language model like CLIP, they optimize a tiny set of learnable visual and text prompts for each incremental task independently (i.e. task-specific prompts) to learn new knowledge and retain the historical information; $ii$ during inference, due to the lack of task identity of the test example, they select the proper visual and text prompts from the prompt pool by prompt matching, and then preprocess them to the visual embedding $V_{ie}^{k}$ and the class + +embedding $W_{we}^{k}$ respectively. The final prediction is determined by the similarities between the visual features and the text features of class names. However, different from static images, videos contain more complicated spatio-temporal clues, and it is intractable to capture these semantic clues well in the video tasks via a single kind of prompt. Meanwhile, the image pre-trained model lacks temporal modeling capability for videos, which requires us to make up for this defect during the prompting process. Overall, the essential point for VCIL is to design a reasonable prompting method to recall the spatio-temporal information in the video task well. + +# 3.2. Space-time prompting + +The proposed space-time prompting (ST-Prompt) contains two components, i.e., task-specific prompts and task-agnostic prompts. + +Task-specific Prompts. To prevent catastrophic forgetting, we train task-specific prompts for each incremental task $\mathcal{T}^k$ with the pre-trained model (the image encoder $f_{v}$ and text encoder $f_{l}$ ) frozen. Specifically, we devote to designing multi-grained prompts to utilize the temporal and spatial variations of different videos for more accurate task identification. Formally, for each task $\mathcal{T}^k$ , there are three different granularities of task-specific prompt pools as Eq. 1. For brevity, we omit the symbol $k$ in this section. If there is no additional explanation, it indicates the situation in incremental task $\mathcal{T}^k$ . + +$$ +\mathbf {P} = \left\{\mathbf {P} _ {s}, \mathbf {P} _ {t}, \mathbf {P} _ {c} \right\} +$$ + +$$ +\mathbf {P} _ {m} = \left\{\left\{P _ {m 1} ^ {v}, P _ {m 1} ^ {l} \right\}, \dots , \left\{P _ {m N} ^ {v}, P _ {m N _ {m}} ^ {l} \right\} \right\}, \tag {1} +$$ + +$$ +P _ {m n} ^ {v} \in \mathbb {R} ^ {L ^ {v} \times D ^ {v}}, P _ {m n} ^ {l} \in \mathbb {R} ^ {L ^ {l} \times D ^ {l}}, m \in \{s, t, c \}, +$$ + +where $\mathbf{P}_s$ , $\mathbf{P}_t$ and $\mathbf{P}_c$ are the spatial, temporal and comprehensive task-specific prompt pools. Each element in prompt pools is organized as the formulation of a visual-text prompt pair, where $P_{mn}^v$ and $P_{mn}^l$ are the commensurable visual and text prompt respectively. Additionally, $N_m$ is the numbers of prompts in each prompt pool. $L^v$ and $L^l$ are the length of the visual and text prompt, and $D^v$ and $D^l$ are their embedding dimension. + +The selection of prompts from these prompt pools is realized through feature matching as Figure 3(a). We apply K-Means algorithm to store the centroids of video features and search for the nearest centroid of the given video sample to identify its relevant prompt. The matching of three kinds of prompts corresponds to features in three different granularities of video $V$ , i.e. spatial feature $F_{s}(V)$ , temporal feature $F_{t}(V)$ and comprehensive feature $F_{c}(V)$ . Specifically, the spatial and temporal features are calculated by pooling the temporal and spatial dimensions separately, while the comprehensive feature is the average of [CLS] token output. After the feature matching, the selected task-specific + +![](images/6e5cda47a219c9d6fcf769e039cbc1774d3a5db56ddeb8ebecbae6944eb3b10a.jpg) +Figure 3. The details of prompt matching methods in prompt pools. (a) We select prompts corresponding to the nearest centroid with query feature for the task-specific prompts; (b) We utilize the similarities between the frame feature and centroids as the attention map for the task-agnostic prompts. + +prompts are organized together as: + +$$ +P ^ {v} = f \left(P _ {s *} ^ {v}, P _ {t *} ^ {v}, P _ {c *} ^ {v}\right), P ^ {l} = f \left(P _ {s *} ^ {l}, P _ {t *} ^ {l}, P _ {c *} ^ {l}\right) \tag {2} +$$ + +where $(P_{s*}^v, P_{s*}^l)$ , $(P_{t*}^v, P_{t*}^l)$ and $(P_{c*}^v, P_{c*}^l)$ are the prompts selected from three task-specific prompt pools. $f$ is the organization function, which can be stacking, addition, multiplication, or other operations. In this paper, we use addition for its better performance. + +Finally, the visual and text prompts are prepended to the visual and text embedding respectively, and the whole input of the visual and text encoder is of the format: + +$$ +V _ {p} = \left[ P ^ {v}; V _ {e} \right], W _ {w p} = \left[ P ^ {l}; W _ {w e} \right]. \tag {3} +$$ + +where ; indicates the impending operation, $V_{e}$ is the video embedding feature of video $V$ and $W_{we}$ is the word embedding feature of w-th class name in task $\mathcal{T}^k$ + +Additionally, when we utilize the visual pre-trained models without a language branch, such as ViT with ImageNet pre-trained weights, the classification head replaces the text encoder and the prompts in each prompt pool just contain the visual prompt $P_{mn}^{v}$ . The other definition and utilization of the task-specific prompts are similar to the above. + +Task-agnostic prompts. When tackling video incremental tasks, the image pre-trained model processes each video frame separately and cannot model the temporal variations and relations between different frames. Although we utilize different granularities of prompts for task identification, it is not enough to make the model understand the temporal information of videos well. To enhance the temporal modeling ability and adaptability of image pre-trained models, we propose task-agnostic prompts to highlight the different information contained in frames. Specifically, we establish an additional temporal-modeling prompt pool as $\mathbf{P_{a}} = \{P_{a1},\dots,P_{aN_a}\}$ , where $P_{an} \in \mathbb{R}^{L^a \times D^v}$ is the prompt and $N_{a}$ is the total number of prompts. Compared + +with the task-specific prompts, temporal modeling is a more general concept for videos, which is task-agnostic and all prompts from the whole pool can be selected and attached without task limitation. + +For the selection of task-agnostic prompts, we also store the centroids of the frame features during training by KMeans algorithm. Differently, in order to utilize the information in the temporal-modeling prompt pool comprehensively, the final prompt according to each frame is not selected by feature matching simply but is calculated by frame-level attention as Figure 3(b). For each frame $I_{j}$ in video $V$ , the attention map for all candidate prompts $S_{j} = \{S_{j1}, S_{j2}, \dots, S_{jNa}\}$ is generated according to the similarity between the frame feature $f_{v}(I_{j})$ and the centroids. Specifically, we utilize the cosine similarity to calculate $S_{jn}$ in this paper. And then the task-agnostic prompt corresponding to the frame $I_{j}$ is calculated as follows: + +$$ +P _ {j} = \sum_ {n = 1} ^ {N _ {a}} \frac {e ^ {S _ {j n} / \tau}}{\sum_ {n = 1} ^ {N _ {a}} e ^ {S _ {j n} / \tau}} P _ {a n} \in \mathbb {R} ^ {D ^ {v}} \tag {4} +$$ + +where $\tau$ is the temperature parameter. Finally, the prompts of different frames are stacked together to adapt to the video embedding as $V_{p} = [P_{1};P_{2};\ldots ;P_{T};V_{e}]$ . In this way, each frame will obtain the context information of others. Although they go forward through an image pre-trained model separately, the information exchange of different frames realizes the temporal modeling. + +# 3.3. Training and Inference + +During training, only the prompts are trainable, while the pre-trained model is fixed. In incremental step $\mathcal{T}^k$ , the training of prompts in task-specific prompt pools is restrained so that only the prompts corresponding to task $\mathcal{T}^k$ can be selected and utilized. The task-agnostic prompt pool does not have this constraint because temporal modeling is a more general concept independent of task identification. + +During inference, for each new video, we first extract multi-grained features by the pre-trained image encoder $f_{v}$ , and then select task-specific and task-agnostic prompts from prompt pools through feature matching or frame-level attention. Finally, the selected prompts are prepended to the video and text embedding features respectively. The classification result is according to the similarity output. + +When ST-Prompt is utilized based on a one-branch visual model, in order to separate different tasks better, the classification head is also split according to the task identity and every incremental session trains its own classifier. Differently, without the text branch, the classifier cannot utilize the multi-grained task-specific prompts flexibly and can only be selected by the video-level comprehensive feature. The classifier selection corresponds to the task that the selected comprehensive task-specific prompt belongs to. + +# 4. Experiment + +In this section, we compare our results with baselines and state-of-the-art VCIL methods. We also show ablation studies on the effect of different components and analyze our results in different practical settings. + +# 4.1. Experiment setup + +Datasets. We evaluate our method on UCF101 [44], HMDB51 [24] and Something-Something V2 [13], which are the standard action recognition datasets. We follow the video frequently-used class-incremental benchmarks presented in TCD [35]. For UCF101, the model is trained on 51 classes first, and the other 50 classes are divided into 5, 10 and 25 tasks. For HMDB51, we train the base model using videos from 26 classes, and the rest of the classes are separated into 5 or 25 groups. For Something-Something V2, we first train 84 classes in the initial stage and generate the incremental groups of 10 and 5 classes. + +Evaluation Protocol. After each incremental step, we evaluate the model on all seen classes and the class-incremental methods are finally assessed by the average accuracy of all tasks. + +Implementation Details. Our method is based on CLIP ViT-B/16 [39], a visual-language model pre-trained with numerous image-text pairs. While the pre-trained model is frozen, the visual and textual prompts are optimized using a batch size of 64 and an initial learning rate of 0.001 with cosine descending for 50 epochs in each task. We carefully choose compared methods in the same environment for a fair comparison. However, most existing class-incremental learning methods are not adaptive to visual-language models. Therefore, we also apply our approach on ImageNet pre-trained ViT-B/16 [8] and reproduce other class-incremental learning methods on the same backbone. + +# 4.2. Main Results. + +We compare our proposed ST-Prompt with existing state-of-the-art class-incremental learning approaches under multiple challenging settings on three datasets in Table 1 and Table 2. For the approaches in the image domain, we reproduce them by treating the video as a whole as that in the image field. From the tables, we can draw the following conclusions. First, ST-Prompt notably outperforms existing rehearsal-free methods in VCIL, containing two recent prompt-based approaches, L2P and S-iPrompts in the image domain. With ImageNet pre-trained ViT, ST-Prompt respectively surpasses S-iPrompts by around $4.22\%$ , $4.98\%$ and $7.27\%$ on HMDB51, UCF101 and Something-something v2, respectively. This indicates that + +
MethodsBuffer SizeUCF101HMDB51
10 × 5 stages5 × 10 stages2 × 25 stages5 × 5 stages1 × 25 stages
ImageNet pre-trained ViT-B/16
iCaRL [40]70.5869.5167.2843.9037.15
UCIR [15]77.5574.5971.7748.2039.42
PODNet [9]76.5076.1773.8348.3843.35
Co2L [6]10/class78.5477.1375.5949.7644.15
TCD†[35]78.1376.8775.7450.2944.04
FrameMaker†[36]79.3779.5579.3251.4346.37
L2P [54]81.2480.0978.5849.9845.87
iCaRL [40]66.5867.7064.2741.0836.13
UCIR [15]75.5572.7866.7645.3838.40
PODNet [9]74.5074.3672.8245.5642.33
Co2L [6]5/class77.5776.3274.5847.1043.56
TCD†[35]76.1375.0673.7348.3843.02
FrameMaker†[36]78.3777.7476.3150.5845.65
L2P [54]79.9377.7876.2747.6742.56
Finetuning24.8313.616.2418.304.89
Finetuning-frozen67.1065.717.3936.644.95
LwFMC [29]45.7328.0116.4327.0117.21
LwM [7]0/class46.9828.7016.9227.9417.85
L2P [54]76.7275.1173.2645.0039.89
S-iPrompts [51]77.1677.8377.9950.6751.45
ST-Prompt†83.7481.5382.6655.1355.43
CLIP ViT-B/16
Zero shot [39]69.6869.6169.8740.1940.42
S-liPrompts [51]0/class80.6080.2780.4353.1153.89
ST-Prompt†84.7585.5485.6760.1460.54
+ +Table 1. Comparison with the state-of-the-art approaches over class-incremental action recognition performance on UCF101 and HMDB51. Compared methods are grouped based on different rehearsal buffer sizes. † indicates the methods designed for video. + +our design of multi-grained task-specific prompts and task-agnostic prompts plays an essential role in video feature formulation and temporal perception. The more tremendous increase on Ssv2 also claims the effectiveness of ST-Prompt for the motion-sensitive dataset. Second, compared with rehearsal-based methods, our ST-Prompt also has a significant advantage. ST-Prompt achieves better performance than the top existing approach, FrameMaker, by approximately $6.38\%$ and $3.23\%$ on HMDB51 and UCF101, separately, although FrameMaker stores additional 80 examples for each class. Third, using the pre-trained vision-language model, CLIP, as the backbone for VCIL, the performance of ST-Prompt improves further by about $5.06\%$ , $2.67\%$ and $3.49\%$ respectively. The reason is that CLIP provides more abundant prior knowledge about semantic information, which is vital for subsequent prompting. + +# 4.3. Ablation Study. + +In this section, we present ablation studies on the properties and effectiveness of our core designs. If not specified, the ablation studies are performed on HMDB51 with 5 steps + +and UCF101 with 10 steps. + +Task-specific Prompts. In Table 3, we first explore how to organize different kinds of task-specific prompts together in Eq. 2. The results show that the addition operation achieves better performance than stacking and multiplication, whose reason may be that different prompts can modulate each other better without preponding too many parameters to the embedding feature in this way. + +Then, to prove the effectiveness of different parts of our task-specific prompts, we remove each kind of prompts respectively under the combination method of addition and show their results in Table 4. From the results, we can find that: first, when three kinds of prompts are utilized separately, the comprehensive task-specific prompts achieve the best performance, which demonstrates that they depict the video feature more integrally; second, when the comprehensive prompts are combined with the temporal or spatial prompts, the performance increases by $1.9\%$ or $1.54\%$ on HMDB51, and $1.26\%$ or $1.34\%$ on UCF101, respectively. Differentiated promotions in different datasets imply that + +
MethodsBuffer Size10 × 9 Stages5 × 18 Stages
ImageNet pre-trained ViT-B/16
iCaRL [40]20.4116.62
UCIR [15]24.3219.31
PODNet [9]5/class27.6320.14
\( \mathrm {TCD}^{\dagger } \)[35]29.3224.69
FrameMaker \( {}^{ \dagger } \)[36]31.4126.57
L2P [54]26.0221.33
Finetuning11.046.23
Finetuning-frozen7.995.97
LwFMC [29]0/class15.2111.41
L2P [54]23.2518.49
S-iPrompts [51]30.5826.33
ST-Prompt36.8431.60
CLIP ViT-B/16
Zero-shot3.563.57
S-liPrompts [51]0/class33.6930.84
\( \mathrm {ST}-\mathrm {Prompt}^{\dagger } \)39.9835.44
+ +Table 2. Comparison with the top approaches on Something-something v2.† indicates the methods designed for video. + +
OperationStackingMultiplicationAddition
HMDB5155.2454.9656.02
UCF10182.8282.4583.58
+ +Table 3. Ablation for the ensemble ways of task-specific prompts. + +
STPTTPCTPHMDB51UCF101
✓✘✘44.6272.45
✘✓✘45.4673.84
✘✘✓53.1180.27
✓✓✘49.3476.93
✓✘✓54.6581.61
✘✓✓55.0181.53
✓✓✓56.0283.58
+ +temporal information is more critical for HMDB51, while the spatial feature for UCF101; third, utilizing all three kinds of task-specific prompts can further improve the final accuracy, which proves the combination of multi-grained prompts contribute to more correct task identification. + +Further, we visualize some clustering results of different granularities of the video feature in Figure 4. The comprehensive feature judges video task-id from the overall perspective, and the clustered videos usually belong to the same spatial and temporal clusters, such as "Smile". The temporal and spatial prompts implement the local information for more complicated video task matching. Videos + +Table 4. Ablations for task-specific prompts on HMDB51 with 5 steps and UCF101 with 10 steps. STP, TTP and CTP represent spatial, temporal and comprehensive prompts respectively. + +
MethodTAPHMDB51UCF101
TSPX56.0283.58
TSP+LEX56.0683.55
TSP+SPX57.0483.87
TSP+CPX58.3084.05
TSP+PE [46]X57.9684.04
TSP+SH [33]X58.2784.25
TSP+TS [30]X58.7484.34
ST-Prompt✓60.1485.54
+ +Table 5. Comparison with some baseline methods about task-agnostic prompts (TAP) on HMDB51 with 5 steps and UCF101 with 10 steps. + +
MethodTAPHMDB51UCF101
L2P [54]✘45.0075.11
✓46.7676.07
S-iPrompts [51]✘50.6777.83
✓54.4479.74
S-liPrompts [51]✘53.1180.27
✓55.7781.82
ST-Prompt✘56.0283.58
✓60.1485.54
+ +Table 6. Ablations for task-agnostic prompts (TAP) on HMDB51 with 5 steps and UCF101 with 10 steps. + +with the same motion but different spatial features are gathered together according to their temporal features, such as "Dive" in different scenarios, while spatial features distinguishes actions with mixed temporal information but obvious spatial similarities. + +Task-agnostic Prompts. We compare our task-agnostic prompts for temporal modeling with a set of baselines in the proposed framework in Table 5. "LE" represents prompt length expansion, which excludes the performance improvement caused by increasing the number of parameters. And the results also prove that learning task-agnostic prompts about spatial ("SP") or comprehensive ("CP") information are not as effective as temporal modeling. Then we compare the effect of our temporal-modeling prompts with some other parameter-free temporal modeling methods, such as positional encoding ("PE"), temporal shuffle ("SH") and temporal shift ("TS"). We can observe that utilizing task-agnostic prompts to exchange the context of different frames can gain the best performance of $60.14\%$ . Further, we validate the effectiveness of task-agnostic prompts by adding them to other prompt-based methods in the image domain, as shown in Table 6, which indicates that it is a general approach to improve the ability of temporal perception for image pre-trained models. + +In addition, We depict the change of category-wise ac + +
MethodsBackboneStored ExemplarsMemory CostAverage Accuracy
TCD [35]ImageNet Pre-trained ViT-B/1610 examples/class585.70M50.29
FrameMaker [36]ImageNet Pre-trained ViT-B/1680 examples/class585.70M51.43
ST-PromptImageNet Pre-trained ViT-B/1602.98M55.13
ST-PromptCLIP ViT-B/1604.47M60.14
+ +Table 7. The comparison of the overhead of our ST-Prompt and other VCIL methods on HMDB51 with 5 steps. + +![](images/11e09fd6ce770f778beed8c27aaa2ffa91b4efe92cf2a1119b44bafd785b6ac9.jpg) + +![](images/2439f4e3c67b916c502da9dec6123db3c19bbdcbe88fc049d9a40a9f9c3cf846.jpg) +Figure 4. The visualization of multi-grained video feature clustering results. +Figure 5. The change of category-wise accuracy after utilizing task-agnostic prompts on HMDB51 with 5 steps. + +curacy after utilizing task-agnostic prompts in Figure 5. The significant increase in the accuracy of some motion-sensitive classes, such as "Dive" and "Somersault", proves the role of our task-agnostic prompts in improving the temporal modeling capability of models. + +The Temperature $\tau$ . The temperature $\tau$ in Eq. 4 decides the smoothness of the attention weights. Too small $\tau$ makes the final prompt of each frame lose the understanding of global information, and too large $\tau$ introduces noise. From Figure 6(a), we observe that $\tau = 0.3$ guides the best performance of final average accuracy. + +The Length of Prompts. The length of different kinds of prompts $L^v, L^t, L^a$ decides the capacity of each learnable prompt. Figure 6(b) illustrates how the average accuracy changes when the length of different prompts varies. It is evident that a too-small visual or text prompt length always negatively affects results, while an oversized prompt may cause underfitting. We set the prompt length as 5 for both visual and text task-specific prompts according to the curves. In addition, we can discover that the final accuracy is not sensitive to the length of task-agnostic prompts + +![](images/24dabd0d866899086d9cd3837bb159abf92529377d3e6ab08733721e1400a451.jpg) +Figure 6. The effects of some hyperparameters in ST-Prompt. + +and even decreases during the length is larger. Hence, we choose the length of the task-agnostic prompts as 1. + +The Number of Prompts. Figure 6(c) presents how the number of different prompts $N_{m}(m \in \{s, t, c, a\})$ in each incremental task influences the performance of our approach, where the number in the base task is calculated according to the proportion of the number of classes. Three different task-specific prompts and task-agnostic prompts are explored separately. In general, the more prompts per task, i.e., the more pre-defined centroids per task during clustering, the higher accuracy gained. Especially, the comprehensive prompt is more sensitive to the cluster number. For simplicity, the number of different prompts is set as 5. + +# 4.4. Overhead Analysis + +We compare the memory cost and average accuracy of ST-Prompt with other VCIL methods in Table 7. Our ST-Prompt utilizes only a set of prompts and prompt keys in each task, while previous rehearsal-based methods store a number of video examples or frames. Therefore, ST-Prompt has significant memory advantages over previous VCIL methods. Our ST-Prompt achieves performance increasing by around $8.71\%$ with only $0.76\%$ memory cost of other methods. + +# 5. Discussions + +Limitation. Compared with the previous video incremental learning approaches, ST-Prompt achieves a significant performance boost without any additional memory cost for video example saving. Nevertheless, ST-Prompt relies on the large pre-trained model, which may fail when the knowledge contained in it is not sufficient. Meanwhile, our + +ST-Prompt prepends prompts to the embedding feature and has not discussed the phenomenon that prompts are prefixed to the hidden layers. + +Conclusion. In this paper, we propose ST-Prompt, a novel space-time prompting framework for video class incremental learning. It explores how to address the VCIL problem in a rehearsal-free way by learning multiple prompts based on a powerful image-language pre-trained model. ST-Prompt mainly consists of two components, task-specific prompts and task-agnostic prompts. The task-specific prompts tackle the catastrophic forgetting problem by learning multi-grained prompts for task identification. The task-agnostic prompts implement the temporal modeling ability of the image pre-trained model. In this way, ST-Prompt achieves a new state-of-the-art performance, which has already surpassed the methods with a memory buffer. We hope this simple and practical framework can inspire attention to prompting-based methods in VCIL. + +Acknowledgments. This work was supported in part by the NSFC under Grant 62272380 and 62103317, the Science and Technology Program of Xi'an, China under Grant 21RGZN0017, and by Alibaba Group through Alibaba Innovative Research Program. + +# References + +[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 1 +[2] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Visual prompting: Modifying pixel space to adapt pre-trained models. arXiv preprint arXiv:2203.17274, 2022. 3 +[3] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022. 3 +[4] Zalán Borsos, Mojmir Mutny, and Andreas Krause. Coresetts via bilevel optimization for continual learning and streaming. Advances in Neural Information Processing Systems, 2020. 1 +[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 3 +[6] Hyuntak Cha, Jaeho Lee, and Jinwoo Shin. Co2l: Contrastive continual learning. In ICCV, 2021. 6 +[7] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In CVPR, 2019. 6 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5 + +[9] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, 2020. 6, 7 +[10] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In CVPR, 2022. 2 +[11] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 1 +[12] Tao Feng, Mang Wang, and Hangjie Yuan. Overcoming catastrophic forgetting in incremental object detection via elastic response distillation. In CVPR, 2022. 2 +[13] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The" something something" video database for learning and evaluating visual common sense. In ICCV, 2017. 5 +[14] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Lifelong learning via progressive distillation and retrospection. In ECCV, 2018. 2 +[15] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, 2019. 6, 7 +[16] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. 3 +[17] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 3 +[18] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3 +[19] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 3 +[20] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, 2022. 3 +[21] Minsoo Kang, Jaeyoo Park, and Bohyung Han. Class incremental learning by knowledge distillation with adaptive feature consolidation. In CVPR, 2022. 2 +[22] Ronald Kemker and Christopher Kanan. Fearnet: Brain-inspired model for incremental learning. In ICLR, 2018. 2 +[23] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 2017. 2 +[24] Hildegard Kuehne, Hueihan Jhuang, Estibaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, 2011. 5 + +[25] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021. 3 +[26] Xilai Li, Yingbo Zhou, Tianfu Wu, Richard Socher, and Caiming Xiong. Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In ICML, 2019. 2 +[27] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In ACL-IJCNLP, 2021. 3 +[28] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In ICLR, 2022. 3 +[29] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 2017. 2, 6, 7 +[30] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In ICCV, 2019. 7 +[31] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Comput. Surv., 2022. 3 +[32] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, 2022. 1 +[33] Pingchuan Ma, Yao Zhou, Yu Lu, and Wei Zhang. Learning efficient video representation with video shuffle networks. arXiv preprint arXiv:1911.11319, 2019. 7 +[34] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation. 1989. 1 +[35] Jaeyoo Park, Minsoo Kang, and Bohyung Han. Class incremental learning for action recognition in videos. In ICCV, 2021. 1, 2, 5, 6, 7, 8 +[36] Yixuan Pei, Zhiwu Qing, Jun CEN, Xiang Wang, Shiwei Zhang, Yaxiong Wang, Mingqian Tang, Nong Sang, and Xueming Qian. Learning a condensed frame for memory-efficient video class-incremental learning. In NeurIPS, 2022. 1, 2, 6, 7, 8 +[37] Fabio Petroni, Tim Rocktäschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and Sebastian Riedel. Language models as knowledge bases? In EMNLP, 2019. 3 +[38] Qi Qin, Wenpeng Hu, Han Peng, Dongyan Zhao, and Bing Liu. Bns: Building network structures dynamically for continual learning. Advances in Neural Information Processing Systems, 2021. 2 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 5, 6 +[40] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017. 1, 2, 6, 7 + +[41] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 2 +[42] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In ICML, 2018. 2 +[43] Yujun Shi, Kuangqi Zhou, Jian Liang, Zihang Jiang, Jiashi Feng, Philip HS Torr, Song Bai, and Vincent YF Tan. Mimicking the oracle: An initial phase decorrelation approach for class incremental learning. In CVPR, 2022. 2 +[44] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5 +[45] Rishabh Tiwari, Krishnateja Killamsetty, Rishabh Iyer, and Pradeep Shenoy. Gcr: Gradient coreset based replay buffer selection for continual learning. In CVPR, 2022. 1, 2 +[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 2017. 7 +[47] Andrés Villa, Kumail Alhamoud, Victor Escorcia, Fabian Caba, Juan León Alcazar, and Bernard Ghanem. vclimb: A novel video class incremental learning benchmark. In CVPR, 2022. 1, 2 +[48] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In ECCV, 2022. 2 +[49] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In ECCV, 2016. 1 +[50] Liyuan Wang, Xingxing Zhang, Kuo Yang, Longhui Yu, Chongxuan Li, Lanqing Hong, Shifeng Zhang, Zhenguo Li, Yi Zhong, and Jun Zhu. Memory replay with data compression for continual learning. In ICLR, 2022. 2 +[51] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. In NeurIPS, 2022. 1, 3, 6, 7 +[52] Zhen Wang, Liu Liu, Yiqun Duan, Yajing Kong, and Dacheng Tao. Continual learning with lifelong vision transformer. In CVPR, 2022. 2 +[53] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In ECCV, 2022. 1, 3 +[54] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In CVPR, 2022. 1, 2, 3, 6, 7 +[55] Chenshen Wu, Luis Herranz, Xialei Liu, Joost van de Weijer, Bogdan Raducanu, et al. Memory replay gans: Learning to generate new categories without forgetting. In NeurIPS, 2018. 2 + +[56] Tz-Ying Wu, Gurumurthy Swaminathan, Zhizhong Li, Avinash Ravichandran, Nuno Vasconcelos, Rahul Bhotika, and Stefano Soatto. Class-incremental learning with strong pre-trained models. In CVPR, 2022. 2 +[57] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, 2019. 2 +[58] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, 2021. 2 +[59] Jaehong Yoon, Divyam Madaan, Eunho Yang, and Sung Ju Hwang. Online coreset selection for rehearsal-based continual learning. arXiv preprint arXiv:2106.01085, 2021. 1, 2 +[60] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, 2017. 2 +[61] Hanbin Zhao, Xin Qin, Shihao Su, Yongjian Fu, Zibo Lin, and Xi Li. When video classification meets incremental classes. In ACM MM, 2021. 1, 2 +[62] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 3 +[63] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 2022. 3 +[64] Kai Zhu, Wei Zhai, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Self-sustaining representation expansion for nonexemplar class-incremental learning. In CVPR, 2022. 2 \ No newline at end of file diff --git a/2023/Space-time Prompting for Video Class-incremental Learning/images.zip b/2023/Space-time Prompting for Video Class-incremental Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b51f5ea78dd8a006178bb49a564a8d8c7dc665e3 --- /dev/null +++ b/2023/Space-time Prompting for Video Class-incremental Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ed336f451f98f17b3568a67c592d2fd26135031f7c11046c600e143f8ffc116 +size 603837 diff --git a/2023/Space-time Prompting for Video Class-incremental Learning/layout.json b/2023/Space-time Prompting for Video Class-incremental Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..87726d1e0bacef636abc7a94d051f8ed6b1add6f --- /dev/null +++ b/2023/Space-time Prompting for Video Class-incremental Learning/layout.json @@ -0,0 +1,9324 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 109, + 103, + 486, + 123 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 103, + 486, + 123 + ], + "spans": [ + { + "bbox": [ + 109, + 103, + 486, + 123 + ], + "type": "text", + "content": "Space-time Prompting for Video Class-incremental Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 142, + 451, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 142, + 451, + 174 + ], + "spans": [ + { + "bbox": [ + 142, + 142, + 451, + 174 + ], + "type": "text", + "content": "Yixuan Pei" + }, + { + "bbox": [ + 142, + 142, + 451, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 142, + 451, + 174 + ], + "type": "text", + "content": " Zhiwu Qing" + }, + { + "bbox": [ + 142, + 142, + 451, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 142, + 451, + 174 + ], + "type": "text", + "content": " Shiwei Zhang" + }, + { + "bbox": [ + 142, + 142, + 451, + 174 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 142, + 142, + 451, + 174 + ], + "type": "text", + "content": " Xiang Wang" + }, + { + "bbox": [ + 142, + 142, + 451, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 142, + 451, + 174 + ], + "type": "text", + "content": " Yingya Zhang" + }, + { + "bbox": [ + 142, + 142, + 451, + 174 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 142, + 142, + 451, + 174 + ], + "type": "text", + "content": " Deli Zhao" + }, + { + "bbox": [ + 142, + 142, + 451, + 174 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 142, + 142, + 451, + 174 + ], + "type": "text", + "content": " Xueming Qian" + }, + { + "bbox": [ + 142, + 142, + 451, + 174 + ], + "type": "inline_equation", + "content": "^{1,4*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 231, + 185, + 362, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 185, + 362, + 199 + ], + "spans": [ + { + "bbox": [ + 231, + 185, + 362, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 231, + 185, + 362, + 199 + ], + "type": "text", + "content": "Xi'an Jiaotong University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 199, + 463, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 199, + 463, + 228 + ], + "spans": [ + { + "bbox": [ + 129, + 199, + 463, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 129, + 199, + 463, + 228 + ], + "type": "text", + "content": "Huazhong University of Science and Technology " + }, + { + "bbox": [ + 129, + 199, + 463, + 228 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 129, + 199, + 463, + 228 + ], + "type": "text", + "content": "Alibaba Group, Shaanxi Yulan Jiuzhou Intelligent Optoelectronic Technology Co., Ltd." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 118, + 229, + 476, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 229, + 476, + 241 + ], + "spans": [ + { + "bbox": [ + 118, + 229, + 476, + 241 + ], + "type": "text", + "content": "{peiiyixuan@stu, qianxm@mail}.xjtu.edu.cn {qzw,wxiang}@hust.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 243, + 463, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 463, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 463, + 255 + ], + "type": "text", + "content": "{zhangjin.zsw, yingya.zyy}@alibaba-inc.com zhaodeli@gmail.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "spans": [ + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 308, + 290, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 290, + 595 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 290, + 595 + ], + "type": "text", + "content": "Recently, prompt-based learning has made impressive progress on image class-incremental learning, but it still lacks sufficient exploration in the video domain. In this paper, we will fill this gap by learning multiple prompts based on a powerful image-language pre-trained model, i.e., CLIP, making it fit for video class-incremental learning (VCIL). For this purpose, we present a space-time prompting approach (ST-Prompt) which contains two kinds of prompts, i.e., task-specific prompts and task-agnostic prompts. The task-specific prompts are to address the catastrophic forgetting problem by learning multi-grained prompts, i.e., spatial prompts, temporal prompts and comprehensive prompts, for accurate task identification. The task-agnostic prompts maintain a globally-shared prompt pool, which can empower the pre-trained image models with temporal perception abilities by exchanging contexts between frames. By this means, ST-Prompt can transfer the plentiful knowledge in the image-language pre-trained models to the VCIL task with only a tiny set of prompts to be optimized. To evaluate ST-Prompt, we conduct extensive experiments on three standard benchmarks. The results show that ST-Prompt can significantly surpass the state-of-the-art VCIL methods, especially it gains " + }, + { + "bbox": [ + 46, + 308, + 290, + 595 + ], + "type": "inline_equation", + "content": "9.06\\%" + }, + { + "bbox": [ + 46, + 308, + 290, + 595 + ], + "type": "text", + "content": " on HMDB51 dataset under the " + }, + { + "bbox": [ + 46, + 308, + 290, + 595 + ], + "type": "inline_equation", + "content": "1 \\times 25" + }, + { + "bbox": [ + 46, + 308, + 290, + 595 + ], + "type": "text", + "content": " stage setting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 636, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 636, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 636, + 287, + 696 + ], + "type": "text", + "content": "Video understanding [49, 11, 1, 32] has achieved outstanding performance with the quick development of deep neural networks. However, training such a model heavily depends on the currently available labeled data. When different classification tasks are presented sequentially in prac" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 396, + 284, + 451, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 284, + 451, + 293 + ], + "spans": [ + { + "bbox": [ + 396, + 284, + 451, + 293 + ], + "type": "text", + "content": "Video Dataset " + }, + { + "bbox": [ + 396, + 284, + 451, + 293 + ], + "type": "inline_equation", + "content": "V^k" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 318, + 296, + 361, + 337 + ], + "blocks": [ + { + "bbox": [ + 318, + 296, + 361, + 337 + ], + "lines": [ + { + "bbox": [ + 318, + 296, + 361, + 337 + ], + "spans": [ + { + "bbox": [ + 318, + 296, + 361, + 337 + ], + "type": "image", + "image_path": "5d816b0a227aea24bbdb09c06be3f4bfee0cae40ffe7f107cbb89183079091af.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 368, + 296, + 409, + 337 + ], + "blocks": [ + { + "bbox": [ + 368, + 296, + 409, + 337 + ], + "lines": [ + { + "bbox": [ + 368, + 296, + 409, + 337 + ], + "spans": [ + { + "bbox": [ + 368, + 296, + 409, + 337 + ], + "type": "image", + "image_path": "a3cee9f31a3ea15b508da7c6b0ebee738560c92788704146eb797ad583e95652.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 316, + 433, + 323 + ], + "lines": [ + { + "bbox": [ + 419, + 316, + 433, + 323 + ], + "spans": [ + { + "bbox": [ + 419, + 316, + 433, + 323 + ], + "type": "text", + "content": "···" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 444, + 296, + 485, + 337 + ], + "blocks": [ + { + "bbox": [ + 444, + 296, + 485, + 337 + ], + "lines": [ + { + "bbox": [ + 444, + 296, + 485, + 337 + ], + "spans": [ + { + "bbox": [ + 444, + 296, + 485, + 337 + ], + "type": "image", + "image_path": "334d3c7ca92d3366b026260b7325fe423ec8dc67eb5a44fe87353873c515fe90.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 492, + 296, + 533, + 337 + ], + "blocks": [ + { + "bbox": [ + 492, + 296, + 533, + 337 + ], + "lines": [ + { + "bbox": [ + 492, + 296, + 533, + 337 + ], + "spans": [ + { + "bbox": [ + 492, + 296, + 533, + 337 + ], + "type": "image", + "image_path": "5ac72593e7509c1b7e6f17659d9517856af9dbf2e6a00755d24843001203db69.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 324, + 351, + 380, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 351, + 380, + 361 + ], + "spans": [ + { + "bbox": [ + 324, + 351, + 380, + 361 + ], + "type": "text", + "content": "Stored Videos[35]" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 310, + 361, + 351, + 403 + ], + "blocks": [ + { + "bbox": [ + 310, + 361, + 351, + 403 + ], + "lines": [ + { + "bbox": [ + 310, + 361, + 351, + 403 + ], + "spans": [ + { + "bbox": [ + 310, + 361, + 351, + 403 + ], + "type": "image", + "image_path": "af8a7bf617c85b5a197f881858441ea6e808ae59079a44f9a34fd81fe6340f04.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 407, + 381, + 423 + ], + "lines": [ + { + "bbox": [ + 328, + 407, + 381, + 423 + ], + "spans": [ + { + "bbox": [ + 328, + 407, + 381, + 423 + ], + "type": "text", + "content": "Memory: 585.70M Accuracy: " + }, + { + "bbox": [ + 328, + 407, + 381, + 423 + ], + "type": "inline_equation", + "content": "50.29\\%" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 353, + 361, + 394, + 403 + ], + "blocks": [ + { + "bbox": [ + 353, + 361, + 394, + 403 + ], + "lines": [ + { + "bbox": [ + 353, + 361, + 394, + 403 + ], + "spans": [ + { + "bbox": [ + 353, + 361, + 394, + 403 + ], + "type": "image", + "image_path": "996c38aa2bca57398f23ea538e0c80b152592940d4214de3b899efa7387355be.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 408, + 351, + 467, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 351, + 467, + 361 + ], + "spans": [ + { + "bbox": [ + 408, + 351, + 467, + 361 + ], + "type": "text", + "content": "Stored Frames[36]" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 402, + 365, + 438, + 401 + ], + "blocks": [ + { + "bbox": [ + 402, + 365, + 438, + 401 + ], + "lines": [ + { + "bbox": [ + 402, + 365, + 438, + 401 + ], + "spans": [ + { + "bbox": [ + 402, + 365, + 438, + 401 + ], + "type": "image", + "image_path": "b33c9ffaebaec423ace6f283ab4876d5fa3e1399743f2194d87b17ac597ec1da.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 407, + 463, + 422 + ], + "lines": [ + { + "bbox": [ + 411, + 407, + 463, + 422 + ], + "spans": [ + { + "bbox": [ + 411, + 407, + 463, + 422 + ], + "type": "text", + "content": "Memory: 585.70M Accuracy: " + }, + { + "bbox": [ + 411, + 407, + 463, + 422 + ], + "type": "inline_equation", + "content": "51.43\\%" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 439, + 365, + 474, + 402 + ], + "blocks": [ + { + "bbox": [ + 439, + 365, + 474, + 402 + ], + "lines": [ + { + "bbox": [ + 439, + 365, + 474, + 402 + ], + "spans": [ + { + "bbox": [ + 439, + 365, + 474, + 402 + ], + "type": "image", + "image_path": "d1109a9feafa61fa3ff70c0e94f409a269186470ae36c99010545ac9324b93dd.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 429, + 547, + 485 + ], + "lines": [ + { + "bbox": [ + 304, + 429, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 547, + 485 + ], + "type": "text", + "content": "Figure 1. An comparison between existing VCIL methods based on exemplar or frame replay and our ST-Prompt. Compared to the rehearsal-based approaches, our ST-prompt can better solve the problem of catastrophic forgetting when only a small amount of memory is consumed to store a tiny set of prompts." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 485, + 351, + 542, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 485, + 351, + 542, + 361 + ], + "spans": [ + { + "bbox": [ + 485, + 351, + 542, + 361 + ], + "type": "text", + "content": "ST-Prompts (Ours)." + } + ] + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 488, + 365, + 498, + 373 + ], + "blocks": [ + { + "bbox": [ + 488, + 365, + 498, + 373 + ], + "lines": [ + { + "bbox": [ + 488, + 365, + 498, + 373 + ], + "spans": [ + { + "bbox": [ + 488, + 365, + 498, + 373 + ], + "type": "image", + "image_path": "b4cfe96281ae8d0e6a6ef6dde0c6536e11bb471c1d381395662c2539ff752915.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 513, + 365, + 531, + 373 + ], + "blocks": [ + { + "bbox": [ + 513, + 365, + 531, + 373 + ], + "lines": [ + { + "bbox": [ + 513, + 365, + 531, + 373 + ], + "spans": [ + { + "bbox": [ + 513, + 365, + 531, + 373 + ], + "type": "image", + "image_path": "7cee4d031d908eba99da31649bf9b77e77a359d8fa7af93858889c8268fd2d76.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 489, + 386, + 498, + 394 + ], + "blocks": [ + { + "bbox": [ + 489, + 386, + 498, + 394 + ], + "lines": [ + { + "bbox": [ + 489, + 386, + 498, + 394 + ], + "spans": [ + { + "bbox": [ + 489, + 386, + 498, + 394 + ], + "type": "image", + "image_path": "f9a668d115558f627839feed4518143f8a62ea0dde28dd68c1c55dc226cc0702.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 513, + 386, + 531, + 394 + ], + "blocks": [ + { + "bbox": [ + 513, + 386, + 531, + 394 + ], + "lines": [ + { + "bbox": [ + 513, + 386, + 531, + 394 + ], + "spans": [ + { + "bbox": [ + 513, + 386, + 531, + 394 + ], + "type": "image", + "image_path": "320bcc749e7a10ac9ba647a91a32872d45339871609a726063ee3e8ed24485cf.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 488, + 403, + 531, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 488, + 403, + 531, + 413 + ], + "spans": [ + { + "bbox": [ + 488, + 403, + 531, + 413 + ], + "type": "text", + "content": "Memory: 2.98M" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 485, + 414, + 531, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 485, + 414, + 531, + 422 + ], + "spans": [ + { + "bbox": [ + 485, + 414, + 531, + 422 + ], + "type": "text", + "content": "Accuracy: 55.13" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 494, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 547, + 602 + ], + "type": "text", + "content": "tical applications, a crucial problem is to maintain the accuracy of the model as data arrives over time. Under this circumstance, due to the limitations of memory and privacy, the model can just access the videos of the current task and few even no examples of historical tasks, which requires that the model should adapt to the current task without catastrophic forgetting [34]. The machine learning paradigm to handle the above challenges is called Video Class-Incremental Learning (VCIL) [35, 36, 61]." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "Most existing class-incremental learning approaches have achieved remarkable performance through a rehearsal buffer of the old tasks in both the image [40, 45, 59, 4] and video [36, 47, 36] fields. However, these methods suffer from suboptimal performance with smaller buffer sizes and severe catastrophic forgetting once the rehearsal buffer is not accessed, hence leading to limited applications. To solve this problem, several prompt-based learning methods in image domain [54, 53, 51] are proposed to train only" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 136, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 136, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 136, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11932" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": "a tiny set of parameters, i.e., prompts, to effectively instruct a pre-trained model to learn incremental tasks without buffering past examples. Typically, L2P [54] learns a single prompt pool to attach complementary prompts to the pre-trained backbone. Nevertheless, this new paradigm has shown great potential for images, but it is still underexplored in the video domain. In this paper, we will concentrate on exploring this paradigm on VCIL to bridge this gap. Directly applying these existing methods for the VCIL task tends to be inadequate. The core reason is that the prompt strategy for images cannot model temporal variations and dynamics across spatial features. Meanwhile, video data possesses a more complicated semantic structure than static images, so a single prompt pool like L2P may be insufficient to learn the abundant semantics. Therefore, we think a properly redesigned prompt framework will contribute to improving the performance of the VCIL task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 275, + 289, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 289, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 289, + 609 + ], + "type": "text", + "content": "Motivated by the above observations, we propose ST-Prompt, a space-time prompting method for video class incremental learning. Specifically, we design ST-Prompt with two components, i.e., task-specific prompts and task-agnostic prompts. The task-specific prompts are tapped by multi-grained prompts, i.e., spatial prompts, temporal prompts and comprehensive prompts, which can better discover the complicated spatio-temporal semantics in videos to address the catastrophic forgetting problem. The task-agnostic prompts learn a complementary prompt for each frame and then gather them back for all frames; hence we can equip the image pre-trained model with temporal modeling capacities without any edition of model structure. Compared with the task-specific prompts, the task-agnostic prompts are task-independent and thus share a generic and globally-shared prompt pool for all tasks. By this means, ST-Prompt has two advantages: 1) it can simultaneously take advantage of the knowledge priors in the image-language pre-trained model and learn the spatiotemporal information, which makes the model more suitable for VCIL; 2) it can markedly reduce the memory cost by just saving a set of prompt parameters without preserving redundant videos anymore, as shown in Figure 1. In the experimental stage, we quantificationally and qualitatively evaluate ST-Prompt on three standard benchmarks. The results show that ST-Prompt can achieve significant gains over the current top-performing approaches, which demonstrates the effectiveness of our method." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 610, + 288, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 288, + 634 + ], + "type": "text", + "content": "In summary, the main contributions of this paper are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 639, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 58, + 639, + 286, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 639, + 286, + 696 + ], + "spans": [ + { + "bbox": [ + 58, + 639, + 286, + 696 + ], + "type": "text", + "content": "- We first explore the great potential of image-language pre-trained model in video incremental tasks and achieve significant performance improvement by utilizing its sufficient semantic information through prompt-based methods;" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 701, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 701, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 701, + 286, + 712 + ], + "type": "text", + "content": "- We proposed ST-Prompt, containing task-specific" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 324, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 545, + 131 + ], + "type": "text", + "content": "prompts and task-agnostic prompts, which can significantly reduce the memory cost and forgetting rate while modeling temporal variations and dynamics across spatial features without modifying the model structure;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 142, + 545, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 142, + 545, + 176 + ], + "spans": [ + { + "bbox": [ + 317, + 142, + 545, + 176 + ], + "type": "text", + "content": "- The proposed ST-Prompt achieves remarkable performance on multiple standard action recognition incremental benchmarks over state-of-the-art methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 194, + 392, + 206 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 194, + 392, + 206 + ], + "spans": [ + { + "bbox": [ + 306, + 194, + 392, + 206 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 216, + 545, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 216, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 305, + 216, + 545, + 253 + ], + "type": "text", + "content": "In this section, we will introduce the methods related to ST-Prompt, including the works for image and video class-incremental learning, and prompt-based learning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 265, + 493, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 265, + 493, + 278 + ], + "spans": [ + { + "bbox": [ + 305, + 265, + 493, + 278 + ], + "type": "text", + "content": "2.1. Image Class-Incremental Learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 285, + 545, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 500 + ], + "type": "text", + "content": "In recent years, class-incremental learning has been widely studied and numerous methods have been proposed in the image domain [52, 56, 10, 43]. These methods can be roughly partitioned into three categories: regularization-based methods, exemplar-based methods and architecture-based methods. Regularization-based methods [23, 60, 21] constrain the optimization of network parameters to preserve the important information of learned tasks, in which the most popular technique utilized is knowledge distillation [29, 12, 14]. The exemplar-based methods [40, 57, 45] typically store a small set of representative exemplars [59, 50] or generated synthetic examples [22, 55] from previous classes to prevent catastrophic forgetting, which is always combined with regularization of a distillation loss to encourage knowledge retention further. The architecture-based methods [26, 42, 48] retain the knowledge of old tasks by designing specific components in the architecture [41, 64] or expanding current feature extractor for new data[58, 38]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 514, + 490, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 490, + 527 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 490, + 527 + ], + "type": "text", + "content": "2.2. Video Class-Incremental Learning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "In the video domain, studies on this field are still scarce. The existing solutions are always the combination of regularization-based and exemplar-based methods and focus on how to formulate temporal information better. For example, TCD [35] computes time-channel importance for weighted distillation, [61] decomposes spatiotemporal knowledge transfer for stronger constraint and vCLIMB [47] focuses on temporal-consistency regularization of untrimmed video. Both TCD [35] and vCLIMB [47] have claimed that more stored examples in memory can effectively encourage incremental performance. Specifically, FrameMaker [36] tends to design a memory-efficient video incremental learning method by learning a condensed frame for each representative video of old classes. However, it still demands to preserve some old frames." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11933" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 68, + 547, + 285 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 547, + 285 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 547, + 285 + ], + "type": "image", + "image_path": "342b993ae8f3893291718be0a838aea9be2e10da319578900b771c2cb5effa43.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 286, + 547, + 330 + ], + "lines": [ + { + "bbox": [ + 46, + 286, + 547, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 547, + 330 + ], + "type": "text", + "content": "Figure 2. The overall framework of the proposed ST-Prompt, which mainly contains two kinds of prompts, task-specific prompts and task-agnostic prompts. For each input video, we first extract multi-grained features with the image encoder of a CLIP model, then select corresponding prompts from predefined prompt pools. Finally, the prompts are prepended to the video and text embedding feature respectively to instruct the model to learn incremental tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 338, + 183, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 183, + 353 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 183, + 353 + ], + "type": "text", + "content": "2.3. Prompt-based Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 373, + 289, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 289, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 289, + 507 + ], + "type": "text", + "content": "Recently, with the rise of large pre-trained models [39, 3, 18, 28], the general pre-trained models are utilized to adapt to multiple down-sampling tasks by finetuning or prompting. The concept of prompt was first proposed in the field of natural language processing [31, 5, 37], and later it was more widely used in language[25, 27, 16] and visual[19, 2, 20] tasks to modify the input texts or images, such that the language or image model obtains additional information about a specific task. Some methods consider multi-modality information simultaneously, and design prompts for the vision-language models[63, 62, 17]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": "Meanwhile, the emergence of large pre-trained models and prompt-based learning also shows excellent potential in incremental learning. L2P [54], DualPrompt [53] and S- iPrompts [51] exploit prompting method to image incremental learning and achieve good performance without saving old exemplars. Especially, S-Prompting utilizes both image and text prompts together with a pretrained vision-language model and gains outstanding performance in domain-incremental learning. In this paper, we focus on the properties of video, adapting large image pre-trained models to the video domain for video class-incremental learning by space-time prompting. We propose to decompose the complex video features into multiple space-time granularities for accurate task identification while exchanging the temporal information between frames by task-agnostic prompts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 338, + 362, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 338, + 362, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 362, + 350 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 358, + 547, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 358, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 358, + 547, + 407 + ], + "type": "text", + "content": "In this section, we will first introduce the basic setting of the video class-incremental learning task. Then we discuss the proposed ST-Prompt in detail, as shown in Figure 2. In the end, we present the training and inference procedure." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 418, + 430, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 430, + 429 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 430, + 429 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 437, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 545, + 581 + ], + "type": "text", + "content": "In video class-incremental learning, different tasks " + }, + { + "bbox": [ + 305, + 437, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}^1,\\mathcal{T}^2,\\dots ,\\mathcal{T}^K\\}" + }, + { + "bbox": [ + 305, + 437, + 545, + 581 + ], + "type": "text", + "content": " arrive in a sequence and each task " + }, + { + "bbox": [ + 305, + 437, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^k" + }, + { + "bbox": [ + 305, + 437, + 545, + 581 + ], + "type": "text", + "content": " corresponds to its specific dataset " + }, + { + "bbox": [ + 305, + 437, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^k = \\{(V_i^k,y_i^k),y_i^k\\in L^k\\}" + }, + { + "bbox": [ + 305, + 437, + 545, + 581 + ], + "type": "text", + "content": ", which contains the input video example " + }, + { + "bbox": [ + 305, + 437, + 545, + 581 + ], + "type": "inline_equation", + "content": "V_{i}^{k}" + }, + { + "bbox": [ + 305, + 437, + 545, + 581 + ], + "type": "text", + "content": " and its label " + }, + { + "bbox": [ + 305, + 437, + 545, + 581 + ], + "type": "inline_equation", + "content": "y_{i}^{k}" + }, + { + "bbox": [ + 305, + 437, + 545, + 581 + ], + "type": "text", + "content": ". The labels in task " + }, + { + "bbox": [ + 305, + 437, + 545, + 581 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 437, + 545, + 581 + ], + "type": "text", + "content": " belong to a predefined label set " + }, + { + "bbox": [ + 305, + 437, + 545, + 581 + ], + "type": "inline_equation", + "content": "L^k" + }, + { + "bbox": [ + 305, + 437, + 545, + 581 + ], + "type": "text", + "content": " and do not appear in previous tasks. The goal of VCIL is to train a single model that has the ability to classify videos of all seen classes correctly. We focus on the more challenging problem of exemplar-free video class-incremental learning, i.e., no exemplar of old classes can be used in the current task, in which the core is to avoid forgetting historical information, termed catastrophic forgetting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we follow the prompt-based incremental learning paradigm in the image domain [54, 51] to alleviate the forgetting problem, which is composed of two stages: " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " based on the pre-trained vision-language model like CLIP, they optimize a tiny set of learnable visual and text prompts for each incremental task independently (i.e. task-specific prompts) to learn new knowledge and retain the historical information; " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "ii" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " during inference, due to the lack of task identity of the test example, they select the proper visual and text prompts from the prompt pool by prompt matching, and then preprocess them to the visual embedding " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "V_{ie}^{k}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " and the class" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11934" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "embedding " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "W_{we}^{k}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " respectively. The final prediction is determined by the similarities between the visual features and the text features of class names. However, different from static images, videos contain more complicated spatio-temporal clues, and it is intractable to capture these semantic clues well in the video tasks via a single kind of prompt. Meanwhile, the image pre-trained model lacks temporal modeling capability for videos, which requires us to make up for this defect during the prompting process. Overall, the essential point for VCIL is to design a reasonable prompting method to recall the spatio-temporal information in the video task well." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 223, + 174, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 223, + 174, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 174, + 236 + ], + "type": "text", + "content": "3.2. Space-time prompting" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 241, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 287, + 277 + ], + "type": "text", + "content": "The proposed space-time prompting (ST-Prompt) contains two components, i.e., task-specific prompts and task-agnostic prompts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 282, + 289, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 289, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 289, + 413 + ], + "type": "text", + "content": "Task-specific Prompts. To prevent catastrophic forgetting, we train task-specific prompts for each incremental task " + }, + { + "bbox": [ + 46, + 282, + 289, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^k" + }, + { + "bbox": [ + 46, + 282, + 289, + 413 + ], + "type": "text", + "content": " with the pre-trained model (the image encoder " + }, + { + "bbox": [ + 46, + 282, + 289, + 413 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 46, + 282, + 289, + 413 + ], + "type": "text", + "content": " and text encoder " + }, + { + "bbox": [ + 46, + 282, + 289, + 413 + ], + "type": "inline_equation", + "content": "f_{l}" + }, + { + "bbox": [ + 46, + 282, + 289, + 413 + ], + "type": "text", + "content": ") frozen. Specifically, we devote to designing multi-grained prompts to utilize the temporal and spatial variations of different videos for more accurate task identification. Formally, for each task " + }, + { + "bbox": [ + 46, + 282, + 289, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^k" + }, + { + "bbox": [ + 46, + 282, + 289, + 413 + ], + "type": "text", + "content": ", there are three different granularities of task-specific prompt pools as Eq. 1. For brevity, we omit the symbol " + }, + { + "bbox": [ + 46, + 282, + 289, + 413 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 282, + 289, + 413 + ], + "type": "text", + "content": " in this section. If there is no additional explanation, it indicates the situation in incremental task " + }, + { + "bbox": [ + 46, + 282, + 289, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^k" + }, + { + "bbox": [ + 46, + 282, + 289, + 413 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 419, + 141, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 419, + 141, + 433 + ], + "spans": [ + { + "bbox": [ + 63, + 419, + 141, + 433 + ], + "type": "interline_equation", + "content": "\\mathbf {P} = \\left\\{\\mathbf {P} _ {s}, \\mathbf {P} _ {t}, \\mathbf {P} _ {c} \\right\\}", + "image_path": "c9a589b49e5629040c710808fd58c3725bfc6ada66aa0e05705e24c8cea64dd4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 434, + 287, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 434, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 63, + 434, + 287, + 449 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {m} = \\left\\{\\left\\{P _ {m 1} ^ {v}, P _ {m 1} ^ {l} \\right\\}, \\dots , \\left\\{P _ {m N} ^ {v}, P _ {m N _ {m}} ^ {l} \\right\\} \\right\\}, \\tag {1}", + "image_path": "48a5a0d70903e9b4ae98736c996c97f3dcd0961934dc7d9cd405c4cb851dd229.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 451, + 259, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 451, + 259, + 468 + ], + "spans": [ + { + "bbox": [ + 63, + 451, + 259, + 468 + ], + "type": "interline_equation", + "content": "P _ {m n} ^ {v} \\in \\mathbb {R} ^ {L ^ {v} \\times D ^ {v}}, P _ {m n} ^ {l} \\in \\mathbb {R} ^ {L ^ {l} \\times D ^ {l}}, m \\in \\{s, t, c \\},", + "image_path": "aceb552af47028ae2e36c45371787fdea252507a6db0d2f6738dc3af84d9a8c9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_s" + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_t" + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_c" + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "text", + "content": " are the spatial, temporal and comprehensive task-specific prompt pools. Each element in prompt pools is organized as the formulation of a visual-text prompt pair, where " + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "inline_equation", + "content": "P_{mn}^v" + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "inline_equation", + "content": "P_{mn}^l" + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "text", + "content": " are the commensurable visual and text prompt respectively. Additionally, " + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "inline_equation", + "content": "N_m" + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "text", + "content": " is the numbers of prompts in each prompt pool. " + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "inline_equation", + "content": "L^v" + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "inline_equation", + "content": "L^l" + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "text", + "content": " are the length of the visual and text prompt, and " + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "inline_equation", + "content": "D^v" + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "inline_equation", + "content": "D^l" + }, + { + "bbox": [ + 46, + 474, + 287, + 569 + ], + "type": "text", + "content": " are their embedding dimension." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "The selection of prompts from these prompt pools is realized through feature matching as Figure 3(a). We apply K-Means algorithm to store the centroids of video features and search for the nearest centroid of the given video sample to identify its relevant prompt. The matching of three kinds of prompts corresponds to features in three different granularities of video " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ", i.e. spatial feature " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "F_{s}(V)" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ", temporal feature " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "F_{t}(V)" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " and comprehensive feature " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "F_{c}(V)" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ". Specifically, the spatial and temporal features are calculated by pooling the temporal and spatial dimensions separately, while the comprehensive feature is the average of [CLS] token output. After the feature matching, the selected task-specific" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 545, + 205 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 545, + 205 + ], + "type": "image", + "image_path": "6e5cda47a219c9d6fcf769e039cbc1774d3a5db56ddeb8ebecbae6944eb3b10a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 209, + 545, + 265 + ], + "lines": [ + { + "bbox": [ + 305, + 209, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 305, + 209, + 545, + 265 + ], + "type": "text", + "content": "Figure 3. The details of prompt matching methods in prompt pools. (a) We select prompts corresponding to the nearest centroid with query feature for the task-specific prompts; (b) We utilize the similarities between the frame feature and centroids as the attention map for the task-agnostic prompts." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 275, + 447, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 275, + 447, + 286 + ], + "spans": [ + { + "bbox": [ + 305, + 275, + 447, + 286 + ], + "type": "text", + "content": "prompts are organized together as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 331, + 293, + 545, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 293, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 331, + 293, + 545, + 308 + ], + "type": "interline_equation", + "content": "P ^ {v} = f \\left(P _ {s *} ^ {v}, P _ {t *} ^ {v}, P _ {c *} ^ {v}\\right), P ^ {l} = f \\left(P _ {s *} ^ {l}, P _ {t *} ^ {l}, P _ {c *} ^ {l}\\right) \\tag {2}", + "image_path": "aaac128bcb09235708ebcb2c6edc5bff4ae7852dc0b4b2879792aad096dabb42.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 315, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 545, + 375 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 315, + 545, + 375 + ], + "type": "inline_equation", + "content": "(P_{s*}^v, P_{s*}^l)" + }, + { + "bbox": [ + 304, + 315, + 545, + 375 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 315, + 545, + 375 + ], + "type": "inline_equation", + "content": "(P_{t*}^v, P_{t*}^l)" + }, + { + "bbox": [ + 304, + 315, + 545, + 375 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 315, + 545, + 375 + ], + "type": "inline_equation", + "content": "(P_{c*}^v, P_{c*}^l)" + }, + { + "bbox": [ + 304, + 315, + 545, + 375 + ], + "type": "text", + "content": " are the prompts selected from three task-specific prompt pools. " + }, + { + "bbox": [ + 304, + 315, + 545, + 375 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 315, + 545, + 375 + ], + "type": "text", + "content": " is the organization function, which can be stacking, addition, multiplication, or other operations. In this paper, we use addition for its better performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 376, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 411 + ], + "type": "text", + "content": "Finally, the visual and text prompts are prepended to the visual and text embedding respectively, and the whole input of the visual and text encoder is of the format:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 355, + 419, + 545, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 419, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 355, + 419, + 545, + 434 + ], + "type": "interline_equation", + "content": "V _ {p} = \\left[ P ^ {v}; V _ {e} \\right], W _ {w p} = \\left[ P ^ {l}; W _ {w e} \\right]. \\tag {3}", + "image_path": "0f7943aa96ec6e550e71f44263befbf574fc89e3bec97b3134effdd0b76922f6.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 441, + 545, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 476 + ], + "type": "text", + "content": "where ; indicates the impending operation, " + }, + { + "bbox": [ + 304, + 441, + 545, + 476 + ], + "type": "inline_equation", + "content": "V_{e}" + }, + { + "bbox": [ + 304, + 441, + 545, + 476 + ], + "type": "text", + "content": " is the video embedding feature of video " + }, + { + "bbox": [ + 304, + 441, + 545, + 476 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 441, + 545, + 476 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 441, + 545, + 476 + ], + "type": "inline_equation", + "content": "W_{we}" + }, + { + "bbox": [ + 304, + 441, + 545, + 476 + ], + "type": "text", + "content": " is the word embedding feature of w-th class name in task " + }, + { + "bbox": [ + 304, + 441, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^k" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 477, + 545, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 545, + 549 + ], + "type": "text", + "content": "Additionally, when we utilize the visual pre-trained models without a language branch, such as ViT with ImageNet pre-trained weights, the classification head replaces the text encoder and the prompts in each prompt pool just contain the visual prompt " + }, + { + "bbox": [ + 304, + 477, + 545, + 549 + ], + "type": "inline_equation", + "content": "P_{mn}^{v}" + }, + { + "bbox": [ + 304, + 477, + 545, + 549 + ], + "type": "text", + "content": ". The other definition and utilization of the task-specific prompts are similar to the above." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "content": "Task-agnostic prompts. When tackling video incremental tasks, the image pre-trained model processes each video frame separately and cannot model the temporal variations and relations between different frames. Although we utilize different granularities of prompts for task identification, it is not enough to make the model understand the temporal information of videos well. To enhance the temporal modeling ability and adaptability of image pre-trained models, we propose task-agnostic prompts to highlight the different information contained in frames. Specifically, we establish an additional temporal-modeling prompt pool as " + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P_{a}} = \\{P_{a1},\\dots,P_{aN_a}\\}" + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "inline_equation", + "content": "P_{an} \\in \\mathbb{R}^{L^a \\times D^v}" + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "content": " is the prompt and " + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "inline_equation", + "content": "N_{a}" + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "content": " is the total number of prompts. Compared" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11935" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": "with the task-specific prompts, temporal modeling is a more general concept for videos, which is task-agnostic and all prompts from the whole pool can be selected and attached without task limitation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 120, + 288, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 288, + 277 + ], + "type": "text", + "content": "For the selection of task-agnostic prompts, we also store the centroids of the frame features during training by KMeans algorithm. Differently, in order to utilize the information in the temporal-modeling prompt pool comprehensively, the final prompt according to each frame is not selected by feature matching simply but is calculated by frame-level attention as Figure 3(b). For each frame " + }, + { + "bbox": [ + 46, + 120, + 288, + 277 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 46, + 120, + 288, + 277 + ], + "type": "text", + "content": " in video " + }, + { + "bbox": [ + 46, + 120, + 288, + 277 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 120, + 288, + 277 + ], + "type": "text", + "content": ", the attention map for all candidate prompts " + }, + { + "bbox": [ + 46, + 120, + 288, + 277 + ], + "type": "inline_equation", + "content": "S_{j} = \\{S_{j1}, S_{j2}, \\dots, S_{jNa}\\}" + }, + { + "bbox": [ + 46, + 120, + 288, + 277 + ], + "type": "text", + "content": " is generated according to the similarity between the frame feature " + }, + { + "bbox": [ + 46, + 120, + 288, + 277 + ], + "type": "inline_equation", + "content": "f_{v}(I_{j})" + }, + { + "bbox": [ + 46, + 120, + 288, + 277 + ], + "type": "text", + "content": " and the centroids. Specifically, we utilize the cosine similarity to calculate " + }, + { + "bbox": [ + 46, + 120, + 288, + 277 + ], + "type": "inline_equation", + "content": "S_{jn}" + }, + { + "bbox": [ + 46, + 120, + 288, + 277 + ], + "type": "text", + "content": " in this paper. And then the task-agnostic prompt corresponding to the frame " + }, + { + "bbox": [ + 46, + 120, + 288, + 277 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 46, + 120, + 288, + 277 + ], + "type": "text", + "content": " is calculated as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 93, + 286, + 287, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 286, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 93, + 286, + 287, + 319 + ], + "type": "interline_equation", + "content": "P _ {j} = \\sum_ {n = 1} ^ {N _ {a}} \\frac {e ^ {S _ {j n} / \\tau}}{\\sum_ {n = 1} ^ {N _ {a}} e ^ {S _ {j n} / \\tau}} P _ {a n} \\in \\mathbb {R} ^ {D ^ {v}} \\tag {4}", + "image_path": "1ad475902c22c0ae95f06cf3372fce7610505574979f2f98eb2e8472067aee96.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "text", + "content": " is the temperature parameter. Finally, the prompts of different frames are stacked together to adapt to the video embedding as " + }, + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "inline_equation", + "content": "V_{p} = [P_{1};P_{2};\\ldots ;P_{T};V_{e}]" + }, + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "text", + "content": ". In this way, each frame will obtain the context information of others. Although they go forward through an image pre-trained model separately, the information exchange of different frames realizes the temporal modeling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 419, + 178, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 178, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 178, + 431 + ], + "type": "text", + "content": "3.3. Training and Inference" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "text", + "content": "During training, only the prompts are trainable, while the pre-trained model is fixed. In incremental step " + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^k" + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "text", + "content": ", the training of prompts in task-specific prompt pools is restrained so that only the prompts corresponding to task " + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^k" + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "text", + "content": " can be selected and utilized. The task-agnostic prompt pool does not have this constraint because temporal modeling is a more general concept independent of task identification." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "content": "During inference, for each new video, we first extract multi-grained features by the pre-trained image encoder " + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "content": ", and then select task-specific and task-agnostic prompts from prompt pools through feature matching or frame-level attention. Finally, the selected prompts are prepended to the video and text embedding features respectively. The classification result is according to the similarity output." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "When ST-Prompt is utilized based on a one-branch visual model, in order to separate different tasks better, the classification head is also split according to the task identity and every incremental session trains its own classifier. Differently, without the text branch, the classifier cannot utilize the multi-grained task-specific prompts flexibly and can only be selected by the video-level comprehensive feature. The classifier selection corresponds to the task that the selected comprehensive task-specific prompt belongs to." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 382, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 382, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 382, + 85 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 91, + 545, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 545, + 140 + ], + "type": "text", + "content": "In this section, we compare our results with baselines and state-of-the-art VCIL methods. We also show ablation studies on the effect of different components and analyze our results in different practical settings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 149, + 411, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 149, + 411, + 163 + ], + "spans": [ + { + "bbox": [ + 305, + 149, + 411, + 163 + ], + "type": "text", + "content": "4.1. Experiment setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 168, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 168, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 168, + 545, + 300 + ], + "type": "text", + "content": "Datasets. We evaluate our method on UCF101 [44], HMDB51 [24] and Something-Something V2 [13], which are the standard action recognition datasets. We follow the video frequently-used class-incremental benchmarks presented in TCD [35]. For UCF101, the model is trained on 51 classes first, and the other 50 classes are divided into 5, 10 and 25 tasks. For HMDB51, we train the base model using videos from 26 classes, and the rest of the classes are separated into 5 or 25 groups. For Something-Something V2, we first train 84 classes in the initial stage and generate the incremental groups of 10 and 5 classes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 319, + 545, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 545, + 367 + ], + "type": "text", + "content": "Evaluation Protocol. After each incremental step, we evaluate the model on all seen classes and the class-incremental methods are finally assessed by the average accuracy of all tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 385, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 385, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 305, + 385, + 545, + 529 + ], + "type": "text", + "content": "Implementation Details. Our method is based on CLIP ViT-B/16 [39], a visual-language model pre-trained with numerous image-text pairs. While the pre-trained model is frozen, the visual and textual prompts are optimized using a batch size of 64 and an initial learning rate of 0.001 with cosine descending for 50 epochs in each task. We carefully choose compared methods in the same environment for a fair comparison. However, most existing class-incremental learning methods are not adaptive to visual-language models. Therefore, we also apply our approach on ImageNet pre-trained ViT-B/16 [8] and reproduce other class-incremental learning methods on the same backbone." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 539, + 392, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 392, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 392, + 551 + ], + "type": "text", + "content": "4.2. Main Results." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 558, + 545, + 713 + ], + "type": "text", + "content": "We compare our proposed ST-Prompt with existing state-of-the-art class-incremental learning approaches under multiple challenging settings on three datasets in Table 1 and Table 2. For the approaches in the image domain, we reproduce them by treating the video as a whole as that in the image field. From the tables, we can draw the following conclusions. First, ST-Prompt notably outperforms existing rehearsal-free methods in VCIL, containing two recent prompt-based approaches, L2P and S-iPrompts in the image domain. With ImageNet pre-trained ViT, ST-Prompt respectively surpasses S-iPrompts by around " + }, + { + "bbox": [ + 305, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "4.22\\%" + }, + { + "bbox": [ + 305, + 558, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "4.98\\%" + }, + { + "bbox": [ + 305, + 558, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "7.27\\%" + }, + { + "bbox": [ + 305, + 558, + 545, + 713 + ], + "type": "text", + "content": " on HMDB51, UCF101 and Something-something v2, respectively. This indicates that" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11936" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 533, + 419 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 533, + 419 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 533, + 419 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 533, + 419 + ], + "type": "table", + "html": "
MethodsBuffer SizeUCF101HMDB51
10 × 5 stages5 × 10 stages2 × 25 stages5 × 5 stages1 × 25 stages
ImageNet pre-trained ViT-B/16
iCaRL [40]70.5869.5167.2843.9037.15
UCIR [15]77.5574.5971.7748.2039.42
PODNet [9]76.5076.1773.8348.3843.35
Co2L [6]10/class78.5477.1375.5949.7644.15
TCD†[35]78.1376.8775.7450.2944.04
FrameMaker†[36]79.3779.5579.3251.4346.37
L2P [54]81.2480.0978.5849.9845.87
iCaRL [40]66.5867.7064.2741.0836.13
UCIR [15]75.5572.7866.7645.3838.40
PODNet [9]74.5074.3672.8245.5642.33
Co2L [6]5/class77.5776.3274.5847.1043.56
TCD†[35]76.1375.0673.7348.3843.02
FrameMaker†[36]78.3777.7476.3150.5845.65
L2P [54]79.9377.7876.2747.6742.56
Finetuning24.8313.616.2418.304.89
Finetuning-frozen67.1065.717.3936.644.95
LwFMC [29]45.7328.0116.4327.0117.21
LwM [7]0/class46.9828.7016.9227.9417.85
L2P [54]76.7275.1173.2645.0039.89
S-iPrompts [51]77.1677.8377.9950.6751.45
ST-Prompt†83.7481.5382.6655.1355.43
CLIP ViT-B/16
Zero shot [39]69.6869.6169.8740.1940.42
S-liPrompts [51]0/class80.6080.2780.4353.1153.89
ST-Prompt†84.7585.5485.6760.1460.54
", + "image_path": "b1f0220845eecae49381fafd91339b1ac5fe6849cf2252afe024ee947598656f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 420, + 545, + 444 + ], + "lines": [ + { + "bbox": [ + 47, + 420, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 545, + 444 + ], + "type": "text", + "content": "Table 1. Comparison with the state-of-the-art approaches over class-incremental action recognition performance on UCF101 and HMDB51. Compared methods are grouped based on different rehearsal buffer sizes. † indicates the methods designed for video." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 456, + 289, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 289, + 649 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 289, + 649 + ], + "type": "text", + "content": "our design of multi-grained task-specific prompts and task-agnostic prompts plays an essential role in video feature formulation and temporal perception. The more tremendous increase on Ssv2 also claims the effectiveness of ST-Prompt for the motion-sensitive dataset. Second, compared with rehearsal-based methods, our ST-Prompt also has a significant advantage. ST-Prompt achieves better performance than the top existing approach, FrameMaker, by approximately " + }, + { + "bbox": [ + 47, + 456, + 289, + 649 + ], + "type": "inline_equation", + "content": "6.38\\%" + }, + { + "bbox": [ + 47, + 456, + 289, + 649 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 456, + 289, + 649 + ], + "type": "inline_equation", + "content": "3.23\\%" + }, + { + "bbox": [ + 47, + 456, + 289, + 649 + ], + "type": "text", + "content": " on HMDB51 and UCF101, separately, although FrameMaker stores additional 80 examples for each class. Third, using the pre-trained vision-language model, CLIP, as the backbone for VCIL, the performance of ST-Prompt improves further by about " + }, + { + "bbox": [ + 47, + 456, + 289, + 649 + ], + "type": "inline_equation", + "content": "5.06\\%" + }, + { + "bbox": [ + 47, + 456, + 289, + 649 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 456, + 289, + 649 + ], + "type": "inline_equation", + "content": "2.67\\%" + }, + { + "bbox": [ + 47, + 456, + 289, + 649 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 456, + 289, + 649 + ], + "type": "inline_equation", + "content": "3.49\\%" + }, + { + "bbox": [ + 47, + 456, + 289, + 649 + ], + "type": "text", + "content": " respectively. The reason is that CLIP provides more abundant prior knowledge about semantic information, which is vital for subsequent prompting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 658, + 144, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 144, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 144, + 671 + ], + "type": "text", + "content": "4.3. Ablation Study." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "In this section, we present ablation studies on the properties and effectiveness of our core designs. If not specified, the ablation studies are performed on HMDB51 with 5 steps" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 456, + 418, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 418, + 469 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 418, + 469 + ], + "type": "text", + "content": "and UCF101 with 10 steps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 485, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 568 + ], + "type": "text", + "content": "Task-specific Prompts. In Table 3, we first explore how to organize different kinds of task-specific prompts together in Eq. 2. The results show that the addition operation achieves better performance than stacking and multiplication, whose reason may be that different prompts can modulate each other better without preponding too many parameters to the embedding feature in this way." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "Then, to prove the effectiveness of different parts of our task-specific prompts, we remove each kind of prompts respectively under the combination method of addition and show their results in Table 4. From the results, we can find that: first, when three kinds of prompts are utilized separately, the comprehensive task-specific prompts achieve the best performance, which demonstrates that they depict the video feature more integrally; second, when the comprehensive prompts are combined with the temporal or spatial prompts, the performance increases by " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "1.54\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " on HMDB51, and " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "1.26\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "1.34\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " on UCF101, respectively. Differentiated promotions in different datasets imply that" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11937" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 281, + 307 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 281, + 307 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 281, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 281, + 307 + ], + "type": "table", + "html": "
MethodsBuffer Size10 × 9 Stages5 × 18 Stages
ImageNet pre-trained ViT-B/16
iCaRL [40]20.4116.62
UCIR [15]24.3219.31
PODNet [9]5/class27.6320.14
\\( \\mathrm {TCD}^{\\dagger } \\)[35]29.3224.69
FrameMaker \\( {}^{ \\dagger } \\)[36]31.4126.57
L2P [54]26.0221.33
Finetuning11.046.23
Finetuning-frozen7.995.97
LwFMC [29]0/class15.2111.41
L2P [54]23.2518.49
S-iPrompts [51]30.5826.33
ST-Prompt36.8431.60
CLIP ViT-B/16
Zero-shot3.563.57
S-liPrompts [51]0/class33.6930.84
\\( \\mathrm {ST}-\\mathrm {Prompt}^{\\dagger } \\)39.9835.44
", + "image_path": "d5cd6f276249f060690e8bd5c87ba9594213176f031b55265d898f85c0c19e8f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 342, + 276, + 389 + ], + "blocks": [ + { + "bbox": [ + 47, + 307, + 286, + 328 + ], + "lines": [ + { + "bbox": [ + 47, + 307, + 286, + 328 + ], + "spans": [ + { + "bbox": [ + 47, + 307, + 286, + 328 + ], + "type": "text", + "content": "Table 2. Comparison with the top approaches on Something-something v2.† indicates the methods designed for video." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 342, + 276, + 389 + ], + "lines": [ + { + "bbox": [ + 58, + 342, + 276, + 389 + ], + "spans": [ + { + "bbox": [ + 58, + 342, + 276, + 389 + ], + "type": "table", + "html": "
OperationStackingMultiplicationAddition
HMDB5155.2454.9656.02
UCF10182.8282.4583.58
", + "image_path": "a9fb069e41ce1e6dc67c39ccc9886fd43e581b88020c6279127092472160faaf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 72, + 412, + 261, + 520 + ], + "blocks": [ + { + "bbox": [ + 47, + 390, + 286, + 402 + ], + "lines": [ + { + "bbox": [ + 47, + 390, + 286, + 402 + ], + "spans": [ + { + "bbox": [ + 47, + 390, + 286, + 402 + ], + "type": "text", + "content": "Table 3. Ablation for the ensemble ways of task-specific prompts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 412, + 261, + 520 + ], + "lines": [ + { + "bbox": [ + 72, + 412, + 261, + 520 + ], + "spans": [ + { + "bbox": [ + 72, + 412, + 261, + 520 + ], + "type": "table", + "html": "
STPTTPCTPHMDB51UCF101
✓✘✘44.6272.45
✘✓✘45.4673.84
✘✘✓53.1180.27
✓✓✘49.3476.93
✓✘✓54.6581.61
✘✓✓55.0181.53
✓✓✓56.0283.58
", + "image_path": "5475ff1e83a0d3f5bfade717b1f836004ec71c2d2f90a8abed915dba9c6913bb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 567, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 286, + 628 + ], + "type": "text", + "content": "temporal information is more critical for HMDB51, while the spatial feature for UCF101; third, utilizing all three kinds of task-specific prompts can further improve the final accuracy, which proves the combination of multi-grained prompts contribute to more correct task identification." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "content": "Further, we visualize some clustering results of different granularities of the video feature in Figure 4. The comprehensive feature judges video task-id from the overall perspective, and the clustered videos usually belong to the same spatial and temporal clusters, such as \"Smile\". The temporal and spatial prompts implement the local information for more complicated video task matching. Videos" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 328, + 70, + 524, + 189 + ], + "blocks": [ + { + "bbox": [ + 46, + 520, + 287, + 553 + ], + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 553 + ], + "type": "text", + "content": "Table 4. Ablations for task-specific prompts on HMDB51 with 5 steps and UCF101 with 10 steps. STP, TTP and CTP represent spatial, temporal and comprehensive prompts respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 70, + 524, + 189 + ], + "lines": [ + { + "bbox": [ + 328, + 70, + 524, + 189 + ], + "spans": [ + { + "bbox": [ + 328, + 70, + 524, + 189 + ], + "type": "table", + "html": "
MethodTAPHMDB51UCF101
TSPX56.0283.58
TSP+LEX56.0683.55
TSP+SPX57.0483.87
TSP+CPX58.3084.05
TSP+PE [46]X57.9684.04
TSP+SH [33]X58.2784.25
TSP+TS [30]X58.7484.34
ST-Prompt✓60.1485.54
", + "image_path": "6bbcaca1bfa2650568295d9f135162146d957851d51392fe678b78eb9e678c56.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 321, + 231, + 531, + 350 + ], + "blocks": [ + { + "bbox": [ + 306, + 190, + 545, + 222 + ], + "lines": [ + { + "bbox": [ + 306, + 190, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 306, + 190, + 545, + 222 + ], + "type": "text", + "content": "Table 5. Comparison with some baseline methods about task-agnostic prompts (TAP) on HMDB51 with 5 steps and UCF101 with 10 steps." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 231, + 531, + 350 + ], + "lines": [ + { + "bbox": [ + 321, + 231, + 531, + 350 + ], + "spans": [ + { + "bbox": [ + 321, + 231, + 531, + 350 + ], + "type": "table", + "html": "
MethodTAPHMDB51UCF101
L2P [54]✘45.0075.11
✓46.7676.07
S-iPrompts [51]✘50.6777.83
✓54.4479.74
S-liPrompts [51]✘53.1180.27
✓55.7781.82
ST-Prompt✘56.0283.58
✓60.1485.54
", + "image_path": "41cb8bee27e330aa0dd951c46640f374cb531df507cca9902f9b963e5118f6f6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 350, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 306, + 350, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 350, + 545, + 373 + ], + "type": "text", + "content": "Table 6. Ablations for task-agnostic prompts (TAP) on HMDB51 with 5 steps and UCF101 with 10 steps." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 394, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 394, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 394, + 545, + 453 + ], + "type": "text", + "content": "with the same motion but different spatial features are gathered together according to their temporal features, such as \"Dive\" in different scenarios, while spatial features distinguishes actions with mixed temporal information but obvious spatial similarities." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 473, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 700 + ], + "type": "text", + "content": "Task-agnostic Prompts. We compare our task-agnostic prompts for temporal modeling with a set of baselines in the proposed framework in Table 5. \"LE\" represents prompt length expansion, which excludes the performance improvement caused by increasing the number of parameters. And the results also prove that learning task-agnostic prompts about spatial (\"SP\") or comprehensive (\"CP\") information are not as effective as temporal modeling. Then we compare the effect of our temporal-modeling prompts with some other parameter-free temporal modeling methods, such as positional encoding (\"PE\"), temporal shuffle (\"SH\") and temporal shift (\"TS\"). We can observe that utilizing task-agnostic prompts to exchange the context of different frames can gain the best performance of " + }, + { + "bbox": [ + 304, + 473, + 545, + 700 + ], + "type": "inline_equation", + "content": "60.14\\%" + }, + { + "bbox": [ + 304, + 473, + 545, + 700 + ], + "type": "text", + "content": ". Further, we validate the effectiveness of task-agnostic prompts by adding them to other prompt-based methods in the image domain, as shown in Table 6, which indicates that it is a general approach to improve the ability of temporal perception for image pre-trained models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "In addition, We depict the change of category-wise ac" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11938" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 72, + 526, + 137 + ], + "blocks": [ + { + "bbox": [ + 67, + 72, + 526, + 137 + ], + "lines": [ + { + "bbox": [ + 67, + 72, + 526, + 137 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 526, + 137 + ], + "type": "table", + "html": "
MethodsBackboneStored ExemplarsMemory CostAverage Accuracy
TCD [35]ImageNet Pre-trained ViT-B/1610 examples/class585.70M50.29
FrameMaker [36]ImageNet Pre-trained ViT-B/1680 examples/class585.70M51.43
ST-PromptImageNet Pre-trained ViT-B/1602.98M55.13
ST-PromptCLIP ViT-B/1604.47M60.14
", + "image_path": "5cb50e22ede36c74cddc46d7bb6fa99d51c879ac857750a220e1ec6df7bcc529.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 93, + 137, + 496, + 148 + ], + "lines": [ + { + "bbox": [ + 93, + 137, + 496, + 148 + ], + "spans": [ + { + "bbox": [ + 93, + 137, + 496, + 148 + ], + "type": "text", + "content": "Table 7. The comparison of the overhead of our ST-Prompt and other VCIL methods on HMDB51 with 5 steps." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 169, + 286, + 264 + ], + "blocks": [ + { + "bbox": [ + 50, + 169, + 286, + 264 + ], + "lines": [ + { + "bbox": [ + 50, + 169, + 286, + 264 + ], + "spans": [ + { + "bbox": [ + 50, + 169, + 286, + 264 + ], + "type": "image", + "image_path": "11e09fd6ce770f778beed8c27aaa2ffa91b4efe92cf2a1119b44bafd785b6ac9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 291, + 284, + 393 + ], + "blocks": [ + { + "bbox": [ + 47, + 269, + 286, + 290 + ], + "lines": [ + { + "bbox": [ + 47, + 269, + 286, + 290 + ], + "spans": [ + { + "bbox": [ + 47, + 269, + 286, + 290 + ], + "type": "text", + "content": "Figure 4. The visualization of multi-grained video feature clustering results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 291, + 284, + 393 + ], + "lines": [ + { + "bbox": [ + 50, + 291, + 284, + 393 + ], + "spans": [ + { + "bbox": [ + 50, + 291, + 284, + 393 + ], + "type": "image", + "image_path": "2439f4e3c67b916c502da9dec6123db3c19bbdcbe88fc049d9a40a9f9c3cf846.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 396, + 286, + 418 + ], + "lines": [ + { + "bbox": [ + 47, + 396, + 286, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 286, + 418 + ], + "type": "text", + "content": "Figure 5. The change of category-wise accuracy after utilizing task-agnostic prompts on HMDB51 with 5 steps." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 438, + 286, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 286, + 498 + ], + "type": "text", + "content": "curacy after utilizing task-agnostic prompts in Figure 5. The significant increase in the accuracy of some motion-sensitive classes, such as \"Dive\" and \"Somersault\", proves the role of our task-agnostic prompts in improving the temporal modeling capability of models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 509, + 286, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 286, + 580 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 286, + 580 + ], + "type": "text", + "content": "The Temperature " + }, + { + "bbox": [ + 46, + 509, + 286, + 580 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 509, + 286, + 580 + ], + "type": "text", + "content": ". The temperature " + }, + { + "bbox": [ + 46, + 509, + 286, + 580 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 509, + 286, + 580 + ], + "type": "text", + "content": " in Eq. 4 decides the smoothness of the attention weights. Too small " + }, + { + "bbox": [ + 46, + 509, + 286, + 580 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 509, + 286, + 580 + ], + "type": "text", + "content": " makes the final prompt of each frame lose the understanding of global information, and too large " + }, + { + "bbox": [ + 46, + 509, + 286, + 580 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 509, + 286, + 580 + ], + "type": "text", + "content": " introduces noise. From Figure 6(a), we observe that " + }, + { + "bbox": [ + 46, + 509, + 286, + 580 + ], + "type": "inline_equation", + "content": "\\tau = 0.3" + }, + { + "bbox": [ + 46, + 509, + 286, + 580 + ], + "type": "text", + "content": " guides the best performance of final average accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "The Length of Prompts. The length of different kinds of prompts " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "L^v, L^t, L^a" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " decides the capacity of each learnable prompt. Figure 6(b) illustrates how the average accuracy changes when the length of different prompts varies. It is evident that a too-small visual or text prompt length always negatively affects results, while an oversized prompt may cause underfitting. We set the prompt length as 5 for both visual and text task-specific prompts according to the curves. In addition, we can discover that the final accuracy is not sensitive to the length of task-agnostic prompts" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 172, + 547, + 273 + ], + "blocks": [ + { + "bbox": [ + 310, + 172, + 547, + 273 + ], + "lines": [ + { + "bbox": [ + 310, + 172, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 310, + 172, + 547, + 273 + ], + "type": "image", + "image_path": "24dabd0d866899086d9cd3837bb159abf92529377d3e6ab08733721e1400a451.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 274, + 529, + 285 + ], + "lines": [ + { + "bbox": [ + 306, + 274, + 529, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 529, + 285 + ], + "type": "text", + "content": "Figure 6. The effects of some hyperparameters in ST-Prompt." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 306, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 306, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 306, + 545, + 331 + ], + "type": "text", + "content": "and even decreases during the length is larger. Hence, we choose the length of the task-agnostic prompts as 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 342, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 473 + ], + "type": "text", + "content": "The Number of Prompts. Figure 6(c) presents how the number of different prompts " + }, + { + "bbox": [ + 304, + 342, + 545, + 473 + ], + "type": "inline_equation", + "content": "N_{m}(m \\in \\{s, t, c, a\\})" + }, + { + "bbox": [ + 304, + 342, + 545, + 473 + ], + "type": "text", + "content": " in each incremental task influences the performance of our approach, where the number in the base task is calculated according to the proportion of the number of classes. Three different task-specific prompts and task-agnostic prompts are explored separately. In general, the more prompts per task, i.e., the more pre-defined centroids per task during clustering, the higher accuracy gained. Especially, the comprehensive prompt is more sensitive to the cluster number. For simplicity, the number of different prompts is set as 5." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 482, + 416, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 482, + 416, + 495 + ], + "spans": [ + { + "bbox": [ + 306, + 482, + 416, + 495 + ], + "type": "text", + "content": "4.4. Overhead Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 502, + 545, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 545, + 608 + ], + "type": "text", + "content": "We compare the memory cost and average accuracy of ST-Prompt with other VCIL methods in Table 7. Our ST-Prompt utilizes only a set of prompts and prompt keys in each task, while previous rehearsal-based methods store a number of video examples or frames. Therefore, ST-Prompt has significant memory advantages over previous VCIL methods. Our ST-Prompt achieves performance increasing by around " + }, + { + "bbox": [ + 304, + 502, + 545, + 608 + ], + "type": "inline_equation", + "content": "8.71\\%" + }, + { + "bbox": [ + 304, + 502, + 545, + 608 + ], + "type": "text", + "content": " with only " + }, + { + "bbox": [ + 304, + 502, + 545, + 608 + ], + "type": "inline_equation", + "content": "0.76\\%" + }, + { + "bbox": [ + 304, + 502, + 545, + 608 + ], + "type": "text", + "content": " memory cost of other methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 621, + 379, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 379, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 379, + 633 + ], + "type": "text", + "content": "5. Discussions" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Limitation. Compared with the previous video incremental learning approaches, ST-Prompt achieves a significant performance boost without any additional memory cost for video example saving. Nevertheless, ST-Prompt relies on the large pre-trained model, which may fail when the knowledge contained in it is not sufficient. Meanwhile, our" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11939" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 109 + ], + "type": "text", + "content": "ST-Prompt prepends prompts to the embedding feature and has not discussed the phenomenon that prompts are prefixed to the hidden layers." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 117, + 289, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 117, + 289, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 117, + 289, + 297 + ], + "type": "text", + "content": "Conclusion. In this paper, we propose ST-Prompt, a novel space-time prompting framework for video class incremental learning. It explores how to address the VCIL problem in a rehearsal-free way by learning multiple prompts based on a powerful image-language pre-trained model. ST-Prompt mainly consists of two components, task-specific prompts and task-agnostic prompts. The task-specific prompts tackle the catastrophic forgetting problem by learning multi-grained prompts for task identification. The task-agnostic prompts implement the temporal modeling ability of the image pre-trained model. In this way, ST-Prompt achieves a new state-of-the-art performance, which has already surpassed the methods with a memory buffer. We hope this simple and practical framework can inspire attention to prompting-based methods in VCIL." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 314, + 289, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 314, + 289, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 289, + 374 + ], + "type": "text", + "content": "Acknowledgments. This work was supported in part by the NSFC under Grant 62272380 and 62103317, the Science and Technology Program of Xi'an, China under Grant 21RGZN0017, and by Alibaba Group through Alibaba Innovative Research Program." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 386, + 107, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 386, + 107, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 386, + 107, + 399 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 406, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 406, + 287, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 406, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 53, + 406, + 287, + 438 + ], + "type": "text", + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 441, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 441, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 53, + 441, + 287, + 483 + ], + "type": "text", + "content": "[2] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Visual prompting: Modifying pixel space to adapt pre-trained models. arXiv preprint arXiv:2203.17274, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 486, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 486, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 486, + 287, + 508 + ], + "type": "text", + "content": "[3] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 510, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 287, + 552 + ], + "type": "text", + "content": "[4] Zalán Borsos, Mojmir Mutny, and Andreas Krause. Coresetts via bilevel optimization for continual learning and streaming. Advances in Neural Information Processing Systems, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 555, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 555, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 53, + 555, + 287, + 599 + ], + "type": "text", + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 600, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 600, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 53, + 600, + 287, + 622 + ], + "type": "text", + "content": "[6] Hyuntak Cha, Jaeho Lee, and Jinwoo Shin. Co2l: Contrastive continual learning. In ICCV, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 624, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 624, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 53, + 624, + 287, + 657 + ], + "type": "text", + "content": "[7] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In CVPR, 2019. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 658, + 287, + 713 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 115 + ], + "type": "text", + "content": "[9] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, 2020. 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 152 + ], + "type": "text", + "content": "[10] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In CVPR, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 153, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 545, + 186 + ], + "type": "text", + "content": "[11] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, 2019. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 188, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 188, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 188, + 545, + 220 + ], + "type": "text", + "content": "[12] Tao Feng, Mang Wang, and Hangjie Yuan. Overcoming catastrophic forgetting in incremental object detection via elastic response distillation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 222, + 545, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 545, + 287 + ], + "type": "text", + "content": "[13] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In ICCV, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 289, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 289, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 289, + 545, + 323 + ], + "type": "text", + "content": "[14] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Lifelong learning via progressive distillation and retrospection. In ECCV, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 324, + 545, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 324, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 324, + 545, + 357 + ], + "type": "text", + "content": "[15] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, 2019. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 359, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 359, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 359, + 545, + 403 + ], + "type": "text", + "content": "[16] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "type": "text", + "content": "[17] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 439, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 439, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 439, + 545, + 483 + ], + "type": "text", + "content": "[18] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 485, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 485, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 307, + 485, + 545, + 517 + ], + "type": "text", + "content": "[19] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 519, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 519, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 519, + 545, + 552 + ], + "type": "text", + "content": "[20] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 554, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 554, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 554, + 545, + 586 + ], + "type": "text", + "content": "[21] Minsoo Kang, Jaeyoo Park, and Bohyung Han. Class incremental learning by knowledge distillation with adaptive feature consolidation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 589, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 589, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 589, + 545, + 611 + ], + "type": "text", + "content": "[22] Ronald Kemker and Christopher Kanan. Fearnet: Brain-inspired model for incremental learning. In ICLR, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 613, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 678 + ], + "type": "text", + "content": "[23] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 2017. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[24] Hildegard Kuehne, Hueihan Jhuang, Estibaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, 2011. 5" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11940" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[25] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "type": "text", + "content": "[26] Xilai Li, Yingbo Zhou, Tianfu Wu, Richard Socher, and Caiming Xiong. Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In ICML, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 287, + 184 + ], + "type": "text", + "content": "[27] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In ACL-IJCNLP, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 287, + 229 + ], + "type": "text", + "content": "[28] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In ICLR, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 231, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 231, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 49, + 231, + 287, + 263 + ], + "type": "text", + "content": "[29] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 2017. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 265, + 287, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 265, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 49, + 265, + 287, + 286 + ], + "type": "text", + "content": "[30] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In ICCV, 2019. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 288, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 288, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 49, + 288, + 287, + 331 + ], + "type": "text", + "content": "[31] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Comput. Surv., 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 333, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 333, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 49, + 333, + 287, + 364 + ], + "type": "text", + "content": "[32] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 366, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 366, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 49, + 366, + 287, + 399 + ], + "type": "text", + "content": "[33] Pingchuan Ma, Yao Zhou, Yu Lu, and Wei Zhang. Learning efficient video representation with video shuffle networks. arXiv preprint arXiv:1911.11319, 2019. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 400, + 287, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 400, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 49, + 400, + 287, + 433 + ], + "type": "text", + "content": "[34] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation. 1989. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 434, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 434, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 49, + 434, + 287, + 466 + ], + "type": "text", + "content": "[35] Jaeyoo Park, Minsoo Kang, and Bohyung Han. Class incremental learning for action recognition in videos. In ICCV, 2021. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 468, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 468, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 49, + 468, + 287, + 521 + ], + "type": "text", + "content": "[36] Yixuan Pei, Zhiwu Qing, Jun CEN, Xiang Wang, Shiwei Zhang, Yaxiong Wang, Mingqian Tang, Nong Sang, and Xueming Qian. Learning a condensed frame for memory-efficient video class-incremental learning. In NeurIPS, 2022. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 524, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 524, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 49, + 524, + 287, + 567 + ], + "type": "text", + "content": "[37] Fabio Petroni, Tim Rocktäschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and Sebastian Riedel. Language models as knowledge bases? In EMNLP, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 569, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 569, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 49, + 569, + 287, + 612 + ], + "type": "text", + "content": "[38] Qi Qin, Wenpeng Hu, Han Peng, Dongyan Zhao, and Bing Liu. Bns: Building network structures dynamically for continual learning. Advances in Neural Information Processing Systems, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 614, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 614, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 49, + 614, + 287, + 667 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "text", + "content": "[40] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017. 1, 2, 6, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[41] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "type": "text", + "content": "[42] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In ICML, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 165, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 165, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 308, + 165, + 545, + 208 + ], + "type": "text", + "content": "[43] Yujun Shi, Kuangqi Zhou, Jian Liang, Zihang Jiang, Jiashi Feng, Philip HS Torr, Song Bai, and Vincent YF Tan. Mimicking the oracle: An initial phase decorrelation approach for class incremental learning. In CVPR, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 211, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 211, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 308, + 211, + 545, + 243 + ], + "type": "text", + "content": "[44] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 245, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 245, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 308, + 245, + 545, + 277 + ], + "type": "text", + "content": "[45] Rishabh Tiwari, Krishnateja Killamsetty, Rishabh Iyer, and Pradeep Shenoy. Gcr: Gradient coreset based replay buffer selection for continual learning. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 280, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 280, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 308, + 280, + 545, + 323 + ], + "type": "text", + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 2017. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 326, + 545, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 326, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 308, + 326, + 545, + 368 + ], + "type": "text", + "content": "[47] Andrés Villa, Kumail Alhamoud, Victor Escorcia, Fabian Caba, Juan León Alcazar, and Bernard Ghanem. vclimb: A novel video class incremental learning benchmark. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 372, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 403 + ], + "type": "text", + "content": "[48] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In ECCV, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 406, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 406, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 406, + 545, + 449 + ], + "type": "text", + "content": "[49] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In ECCV, 2016. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 452, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 452, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 308, + 452, + 545, + 495 + ], + "type": "text", + "content": "[50] Liyuan Wang, Xingxing Zhang, Kuo Yang, Longhui Yu, Chongxuan Li, Lanqing Hong, Shifeng Zhang, Zhenguo Li, Yi Zhong, and Jun Zhu. Memory replay with data compression for continual learning. In ICLR, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 498, + 545, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 498, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 308, + 498, + 545, + 530 + ], + "type": "text", + "content": "[51] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. In NeurIPS, 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 533, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 545, + 564 + ], + "type": "text", + "content": "[52] Zhen Wang, Liu Liu, Yiqun Duan, Yajing Kong, and Dacheng Tao. Continual learning with lifelong vision transformer. In CVPR, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 567, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 620 + ], + "type": "text", + "content": "[53] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In ECCV, 2022. 1, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "type": "text", + "content": "[54] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In CVPR, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[55] Chenshen Wu, Luis Herranz, Xialei Liu, Joost van de Weijer, Bogdan Raducanu, et al. Memory replay gans: Learning to generate new categories without forgetting. In NeurIPS, 2018. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11941" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 388 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[56] Tz-Ying Wu, Gurumurthy Swaminathan, Zhizhong Li, Avinash Ravichandran, Nuno Vasconcelos, Rahul Bhotika, and Stefano Soatto. Class-incremental learning with strong pre-trained models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "text", + "content": "[57] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "text", + "content": "[58] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 218 + ], + "type": "text", + "content": "[59] Jaehong Yoon, Divyam Madaan, Eunho Yang, and Sung Ju Hwang. Online coreset selection for rehearsal-based continual learning. arXiv preprint arXiv:2106.01085, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 251 + ], + "type": "text", + "content": "[60] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 254, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 285 + ], + "type": "text", + "content": "[61] Hanbin Zhao, Xin Qin, Shihao Su, Yongjian Fu, Zibo Lin, and Xi Li. When video classification meets incremental classes. In ACM MM, 2021. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "type": "text", + "content": "[62] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 321, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 354 + ], + "type": "text", + "content": "[63] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 355, + 287, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 388 + ], + "type": "text", + "content": "[64] Kai Zhu, Wei Zhai, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Self-sustaining representation expansion for nonexemplar class-incremental learning. In CVPR, 2022. 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11942" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SpaceEvo_ Hardware-Friendly Search Space Design for Efficient INT8 Inference/7d8d9b36-b64d-45d0-8aa5-d17e8abbee45_content_list.json b/2023/SpaceEvo_ Hardware-Friendly Search Space Design for Efficient INT8 Inference/7d8d9b36-b64d-45d0-8aa5-d17e8abbee45_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bc932a8dc19fd39c01ad55b8acb3c45f9b9a0fa8 --- /dev/null +++ b/2023/SpaceEvo_ Hardware-Friendly Search Space Design for Efficient INT8 Inference/7d8d9b36-b64d-45d0-8aa5-d17e8abbee45_content_list.json @@ -0,0 +1,1608 @@ +[ + { + "type": "text", + "text": "SpaceEvo: Hardware-Friendly Search Space Design for Efficient INT8 Inference", + "text_level": 1, + "bbox": [ + 78, + 130, + 890, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xudong Wang\\*1,2§ Li Lyna Zhang\\*2‡ Jiahang Xu² Quanlu Zhang² Yujing Wang³ Yuqing Yang² Ningxin Zheng² Ting Cao² Mao Yang²", + "bbox": [ + 119, + 191, + 805, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Jiao Tong University $^{2}$ Microsoft Research $^{3}$ Microsoft", + "bbox": [ + 230, + 238, + 733, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xudongwang176@gmail.com; {Izhani, jiahangxu, quzha, yujwang, yuqyang, nizhen, ticao, maoyang}@microsoft.com", + "bbox": [ + 91, + 257, + 854, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/microsoft/Moonlit/tree/main/SpaceEvo", + "bbox": [ + 212, + 273, + 754, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 334, + 313, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The combination of Neural Architecture Search (NAS) and quantization has proven successful in automatically designing low-FLOPs INT8 quantized neural networks (QNN). However, directly applying NAS to design accurate QNN models that achieve low latency on real-world devices leads to inferior performance. In this work, we identify that the poor INT8 latency is due to the quantization-unfriendly issue: the operator and configuration (e.g., channel width) choices in prior art search spaces lead to diverse quantization efficiency and can slow down the INT8 inference speed. To address this challenge, we propose SpaceEvo, an automatic method for designing a dedicated, quantization-friendly search space for each target hardware. The key idea of SpaceEvo is to automatically search hardware-preferred operators and configurations to construct the search space, guided by a metric called $Q-T$ score to quantify how quantization-friendly a candidate search space is. We further train a quantized-for-all supernet over our discovered search space, enabling the searched models to be directly deployed without extra retraining or quantization. Our discovered models, SEQnet, establish new SOTA INT8 quantized accuracy under various latency constraints, achieving up to $10.1\\%$ accuracy improvement on ImageNet than prior art CNNs under the same latency. Extensive experiments on real devices show that SpaceEvo consistently outperforms manually-designed search spaces with up to $2.5\\times$ faster speed while achieving the same accuracy.", + "bbox": [ + 75, + 366, + 473, + 775 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 786, + 209, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "INT8 Quantization[27, 19, 10, 3] is a widely used technique for deploying DNNs on edge devices by reducing $4 \\times$ in model size and memory cost for full-precision (FP32)", + "bbox": [ + 75, + 806, + 468, + 853 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3366466d8acddb126a6fe904cf34998e83103147b276e6c158fcdf74470fdc1a.jpg", + "image_caption": [ + "Figure 1. INT8 quantized model error distributions (EDF) of different NAS search spaces. We propose to search quantization-friendly search space for each hardware, which yields significantly better INT8 model populations than SOTA search spaces." + ], + "image_footnote": [], + "bbox": [ + 509, + 330, + 883, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "models. However, prior art DNN models achieve only marginal speedup from INT8 quantization (in Fig. 2(a)), the still high latency after quantization making them difficult to deploy in latency-critical scenarios. Designing models that achieve high accuracy and low latency after quantization becomes the important but challenging problem.", + "bbox": [ + 496, + 606, + 892, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Architecture Search (NAS) is a powerful tool for automating efficient quantized model design [36, 39, 12, 37, 6]. Recently, OQAT [32] and BatchQuant [1] achieve remarkable search efficiency and accuracy by adopting a two-stage paradigm. The first stage trains a weight-shared quantized supernet assembling all candidate architectures in the search space. This allows all the sub-networks (subnets) to simultaneously reach comparable quantized accuracy as when trained from scratch individually. The second stage uses typical search algorithms to find subnets with best quantized accuracy under different FLOPs constraints. This approach avoids the need to retrain each subnet for accuracy evaluation, greatly improving the search efficiency.", + "bbox": [ + 496, + 703, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 93, + 861, + 202, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{\\S}$ Work was done during the internship at Microsoft Research", + "bbox": [ + 94, + 875, + 416, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author (lzhani@microsoft.com)", + "bbox": [ + 94, + 887, + 351, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5819", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Though promising in optimizing model FLOPs, we identify a significant challenge when directly applying two-stage NAS to low quantized latency scenarios. This is due to the quantization-unfriendly search space issue, where prior art search spaces can unexpectedly impede INT8 latency. Consequently, since INT8 quantization yields only a marginal speedup, we are forced to search for small-sized models to fulfill latency criteria, which can unfortunately restrict NAS to find better quantized models for edge devices. Then, a question naturally arises: Can we design a quantization-friendly search space, allowing NAS to discover larger and superior models that meet the low INT8 latency requirements?", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We start by conducting an in-depth study to understand the factors that determine INT8 quantized latency and how they affect search space design. Our study shows: (1) both operator type and configurations (e.g., channel width) greatly impact the INT8 latency; Improper selections can slow down the INT8 latency. For instance, Squeeze-and-Excitation (SE) [16] and Hardswish [14] are widely-used operators in current search spaces as it improves accuracy with little latency introduced. However, their INT8 inference speeds are slower than FP32 inference on Intel CPU (Fig. 3(a)), because the extra costs (e.g., data transformation between INT32 and INT8) introduced by quantization outweigh the latency reduction by INT8 computation. (2) The quantization efficiency varies across different devices, and the preferred operator types can be contradictory.", + "bbox": [ + 75, + 290, + 472, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The above study motivates us to design specialized quantization-friendly search spaces for each hardware, rather than relying on a single, large search space as seen in SPOS [12] for all hardware, which provides different operator options per layer. This is because two-stage NAS requires the search space to adhere to a specific condition for training the supernet, wherein each layer must utilize a fixed operator. Our study indicates significant variations in optimal operators across different hardware. Thus, customizing the search space for each hardware is crucial for optimal results. However, designing such specialized quantization-friendly search spaces for various edge devices presents a significant challenge, requiring expertise in both AI and hardware domains, as well as many trial and error attempts to optimize accuracy and INT8 latency for each hardware.", + "bbox": [ + 75, + 518, + 472, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose SpaceEvo, a novel method for automatically designing specialized quantization-friendly search space for each hardware. The search space is comprised of hardware-preferred operators and configurations, enabling the search of larger and better models with low INT8 latency. With the discovered search space, we leverage two-stage quantization NAS to train a quantized-for-all supernet, and utilize evolution search [4] to find best quantized models under various INT8 latency constraints. Our approach addresses three key challenges: (1) What is the", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "definition of a quantization-friendly search space in terms of both quantized accuracy and latency? (2) How to automatically design a search space without human expertise? (3) How to handle with the prohibitive cost caused by quality evaluation of a candidate search space?", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the first challenge, we propose a latency-aware $Q-T$ score to quantify the effectiveness of a candidate search space, which measures the INT8 accuracy-latency quality of top-tier subnets in a search space. The behind intuition is that the goal of NAS is to search top subnets with better accuracy-latency tradeoffs.", + "bbox": [ + 496, + 167, + 890, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Then, we introduce an evolutionary-based search algorithm that can effectively search a quantization-friendly search space with highest $Q-T$ score. Searching a search space involves discovering a collection of model population that contains billions of models, which is challenging and easily introduce complexity. To address this challenge, we propose to factorize and encode a search space into a sequence of elastic stages, which have flexible operator types and configurations. Through this design, the task of search space design is then simplified to find a search space with the optimal elastic stages, so that existing search algorithms can be easily applied. Specifically, we design a stage-wise hyperspace to include many candidate search spaces and leverage aging evolution [30] to perform random mutations of elastic stages for search space evolution. The evolution is guided by maximizing the $Q-T$ score.", + "bbox": [ + 496, + 258, + 892, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, estimating the quality score (Q-T score) of a search space involves a costly training process for evaluating the accuracy of sub-networks, which presents a significant obstacle for our evolutionary algorithm. Naively adopting a two-stage NAS approach, training a supernet for each candidate search space [8, 7], is prohibitively expensive, taking of thousands GPU hours. To address this issue, we draw inspiration from block-wise knowledge distillation [26, 22] and propose a block-wise search space quantization scheme. This scheme trains each elastic stage separately and rapidly estimates a model's quantized accuracy by summing block-level loss with a quantized accuracy lookup table, as shown in Fig. 5. This significantly reduces the training and evaluation costs, while providing effective accuracy rankings among search spaces. We summarize our contributions as follows:", + "bbox": [ + 496, + 502, + 892, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We study the INT8 quantization efficiency on real-world edge devices and find that the choices of operator types and configurations in a quantized model can significantly impact the INT8 latency, leaving a huge room for design optimization of quantized models.", + "bbox": [ + 500, + 750, + 890, + 825 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose SpaceEvo to automatically design a hardware-dedicated quantization-friendly search space and leverage two-stage quantization NAS to produce superior INT8 models under various latency constraints.", + "bbox": [ + 500, + 840, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5820", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present three innovative techniques that enable the first-ever efficient and cost-effective evolution search to explore a search space comprising billions of models.", + "- Extensive experiments on two real-world edge devices and ImageNet demonstrate that our automatically designed search spaces significantly surpass manually-designed search spaces. Our discovered models, SEQnet, establish the new state-of-the-art INT8 quantized accuracy-latency tradeoffs. For instance, SEQnet@cpuA4, achieves $80.0\\%$ accuracy on ImageNet, which is 3.3ms faster with $1.8\\%$ higher accuracy than FBNetV3-A. Moreover, SpaceEvo delivers superior tiny models, achieving up to $10.1\\%$ accuracy improvement over ShuffleNetV2x0.5 (41M FLOPs, 4.3ms)." + ], + "bbox": [ + 76, + 90, + 470, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 323, + 227, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Quantization has been widely used for efficiency in deployment. Extensive efforts can be classified into posttraining quantization (PTQ) [27, 2] and quantization-aware training (QAT) [19, 24, 10, 3]. QAT generally outperforms PTQ in quantizing compact DNNs to 8bit and very low-bit (2, 3, 4bit) by finetuning the quantized weights. Despite their success, traditional quantization methods focus on minimizing accuracy loss for a given pre-trained model, but ignore the real-world inference efficiency.", + "bbox": [ + 75, + 349, + 468, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NAS for Quantization. Early works [36, 39, 12, 37, 6] formulates mixed-precision problem into NAS to search layer bit-width with a given architecture. Recently, [32, 1] train a quantized-for-all supernet to search both architecture and bit-width. The searched models can be directly evaluated with comparable accuracy to train-from-scratch. However, little attention is paid on optimizing quantized latency on real-world devices. Through searching quantization-friendly search space, our discovered quantized models can achieve both high accuracy and low latency.", + "bbox": [ + 75, + 484, + 468, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Search Space Design. Starting from [5], the manually-designed MBConv-based space becomes the dominant in most NAS works [5, 4, 41, 35]. RegNet [29] is the first to present standard guidelines to optimize a search space by each dimension. Recently, [17, 28, 23, 40, 8, 7] propose to shrink to a better compact search space by either pruning unimportant operators or configurations. However, these works focus on optimizing the accuracy and little attention is paid on quantization-friendly search space design. Our work is the first lightweight solution towards this direction.", + "bbox": [ + 75, + 636, + 470, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. On-device Quantization Efficiency Analysis", + "text_level": 1, + "bbox": [ + 76, + 800, + 465, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To understand what factors lead to quantization-unfriendly issue, we conduct a comprehensive study on two widely-used edge devices equipped with high-performance inference engine: an Intel CPU device supported with VNNI instructions and onnxruntime [25] (Intel CPU) and", + "bbox": [ + 76, + 825, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1d5f9faa311863c173d7770af5f512e6f891b78bcc75e182cc871cd9dcc78b94.jpg", + "image_caption": [ + "(a) FLOPs and INT8 latency on Intel CPU", + "Figure 2. INT8 latency and speedups (annotated) for SOTA models. FLOPs and FP32 latency are not good indicators of INT8 latency; Compact models have very limited INT8 speedup $(\\sim 1.5\\times)$ ." + ], + "image_footnote": [], + "bbox": [ + 501, + 87, + 705, + 213 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ef11762f49ee40899870cb25b2f045762028736767160477801850f6ba33f6e5.jpg", + "image_caption": [ + "(b) FP32 and INT8 latency on Pixel 4" + ], + "image_footnote": [], + "bbox": [ + 710, + 88, + 893, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a Pixel 4 phone CPU with TFLite 2.7 [11] (Pixel 4). Note that we follow existing practices [42, 34, 38] to measure the latency. We reveal key observations as follows:", + "bbox": [ + 498, + 291, + 890, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Observation 1: FP32 latency and FLOPs are not good indicators of INT8 latency.", + "bbox": [ + 498, + 339, + 890, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To deploy on edge devices, a common belief is that a compact model with low FLOPs or FP32 latency is preferred than a larger model. However, Fig. 2 shows that neither of them is a good indicator of INT8 latency. In Fig. 2(b), a very large model (ResNet18) can be even faster than a compact model (EfficientNet-B0) after quantization. Moreover, the recent SOTA compact models searched by OFA [4] and AttentiveNAS [35] all have marginal INT8 speedups, suggesting that optimizing FLOPs and FP32 latency can not lead to lower INT8 latency.", + "bbox": [ + 496, + 369, + 892, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Observation 2: The choices of operators' types and configurations greatly impact the INT8 latency.", + "bbox": [ + 498, + 522, + 890, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The prior art search spaces adopted in recent two-stage NAS works are MobileNetV2 or MobileNetV3 based chain-structures, where each search space comprises a sequence of blocks (stages). The block type is fixed to the MBCnv and is allowed to search from a handcraft range of hyperparameter configurations including kernel size, expansion ratio, channel width and depth, which are designed with human wisdom. For instance, many works [5, 35] observe that edge-regime CNNs prefer deeper depths and narrower channels, and manually set small channel numbers but large depths in the search space.", + "bbox": [ + 496, + 553, + 893, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, we find that many block type and configuration choices in current search spaces unexpectedly slow down the INT8 latency. We first study the operator type impact in Fig. 3(a). SE and Hardswish are lightweight operators in edge-regime search spaces, but their INT8 inference becomes slower on Intel CPU. Compared to Conv, DWConv can greatly reduce the FLOPs, but it benefits less from INT8 quantization. The root cause is that quantization introduces extra cost, such as (1) data transformation between INT32 and INT8 [34], and (2) additional computation caused by scaling fators and zero points [20]. If the operator has low data-reuse-rate, such as the activation functions (Hswish),", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5821", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/6e1f344dca7d439b905864784d5eb23394e3c8c19543b4ef83f98a3dbff6c374.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
OperatorIntel CPUPixel 4
Conv2.6x2.5x
DWConv1.2x2.0x
SE0.7x1.4x
Hardswish0.7x0.5x
Swish0.8x2.1x
", + "bbox": [ + 78, + 90, + 251, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/09f114e9a3ff2e742520a7fb1111b0ddeec7ceaac879c5098f9cdd8e005736f5.jpg", + "image_caption": [ + "(a) Avg. INT8 latency speedup (relative to FP32) of major operators in MobileNetV3.", + "(b) Conv1x1 speedups", + "Figure 3. (a) The choice of operator type leads to significantly different quantized speedup. (b) Conv1x1 speedups with various channel numbers. Config: $HW = 28$ , $C_{out} = 4 \\times C_{in}$ (expand=4)." + ], + "image_footnote": [], + "bbox": [ + 254, + 88, + 472, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the extra cost may outweigh the latency reduction by the low-bit computation. For high data-reuse operators (Conv), this cost is amortized and thus achieve large speedup [34].", + "bbox": [ + 75, + 284, + 468, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides the operator type, the configuration choices also determine the quantization efficiency. Fig. 3(b) shows the speedups of $\\mathrm{Conv}1\\times 1$ under various channel widths. Results show that small channel widths in OFA search space cannot benefit well from quantization. This is because the additional quantization cost has a large impact when the channel width is small, limiting the latency acceleration. In contrast, SpaceEvo can automatically design a search space with larger channel widths for better efficiency.", + "bbox": [ + 75, + 332, + 468, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Observation 3: Quantization-friendly settings are diverse and contradictory across devices.", + "bbox": [ + 75, + 470, + 468, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Fig. 3, we also observe that the quantization-friendly operators are different and can be contradictory on diverse devices. For instance, Swish achieves a $2 \\times$ speedup on Pixel 4, but it is a quantization-unfriendly operator on CPU with a $0.8 \\times$ slowdown. The reason is that quantization speedups are highly dependent on the inference engines and hardware [34, 21]. Intel VNNI supports the VPDPBWSD hardware instruction [18], which fuses three instructions into one to speedup INT8 computation. Without VNNI, INT8 hardly gains speedup on Intel CPUs. Moreover, the implementations in inference engines have to fully utilize hardware instructions for latency reduction. For example, onnxruntime does not implement a quantization kernel for Hardswish. Even on a VNNI-supported CPU, the use of Hardswish in a quantized model slows down the latency.", + "bbox": [ + 75, + 503, + 468, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The need for hardware specialized search space. The above observations suggest that there is no single structure (block types in a model) that is optimal for quantization on all hardware. This poses a challenge for the two-stage NAS paradigm, as the supernet training requires all models in the search space to share an isomorphic structure. To address this issue, our work proposes to design a specialized quantization-friendly search space for each hardware. Each search space is tailored to the unique characteristics of the hardware and includes an optimal structure with elastic depths, widths, and kernel sizes.", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 500, + 89, + 633, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. The Core Design Concept", + "text_level": 1, + "bbox": [ + 500, + 114, + 732, + 131 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we present our methodology for automatically designing a specialized quantization-friendly search space for any target hardware. Different from architecture search, where the goal is to find the single best model from the space, we aim to discover a model population that contains billions of accurate and INT8 latency-friendly architectures. We draw inspiration from the neural architecture search process and propose to use an evolutionary search algorithm to explore such a quantization-friendly model population. To achieve this, we introduce SpaceEvo, which is built on the following three techniques.", + "bbox": [ + 496, + 133, + 890, + 299 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we need a metric that quantifies how quantization-friendly a candidate search space is. We define a Q-T score that is efficiently measured by top-tier subnets' INT8 accuracy-latency (Sec. 4.2).", + "bbox": [ + 496, + 299, + 890, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Second, existing evolutionary search algorithms are designed for searching a single model architecture rather than a large search space encompassing billions of architectures. we propose a novel approach that we call the \"elastic stage.\" By factorizing the search space into a sequence of elastic stages ((Sec. 4.3), we enable traditional aging evolution methods, such as the aging evolution [30], to be directly applied to search the space (Sec. 4.4).", + "bbox": [ + 496, + 359, + 890, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Third, searching a search space with a maximum Q-T score can be prohibitively costly since the corresponding supernet must be trained from scratch for accuracy evaluation. We propose a block-wise search space quantization scheme to significantly reduce the training cost(Sec. 4.5).", + "bbox": [ + 496, + 481, + 890, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Search Space Quality Score", + "text_level": 1, + "bbox": [ + 498, + 568, + 746, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Latency-aware space quality of Q-T score. Before space search, we need a score to quantify how quantization-friendly a search space is, which serves as the search objective. Since our ultimate goal is to search the best quantized models from the searched space, we treat a space with good quality if its top-tier subnets achieve optimal quantized accuracy under latency constraints $T$ . Due to the fact that real-world applications usually have different deployment requirements, we use multiple INT8 latency constraints to measure a space's quality. For search space $\\mathcal{A}$ and a set of quantized latency constraints $T_{1,\\dots,n}$ , we treat every constraint equally important, and define Q-T score as the sum of each constraint: $\\mathcal{Q}(\\mathcal{A},T_1,\\dots,n) = \\mathcal{Q}(\\mathcal{A},T_1) + \\mathcal{Q}(\\mathcal{A},T_2) + \\dots$ , $\\mathcal{Q}(\\mathcal{A},T_n)$ , $\\mathcal{Q}(\\mathcal{A},T_i)$ is defined as:", + "bbox": [ + 496, + 585, + 890, + 799 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Q} (\\mathcal {A}, T _ {i}) = \\mathbb {E} _ {\\alpha \\in \\mathcal {A}, L A T (\\alpha) \\leq T _ {i}} \\left[ A c c _ {i n t 8} (\\alpha) \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 810, + 890, + 828 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ denotes a top-tier (best searched) subnet in $\\mathcal{A}$ and $Acc_{int8}(\\alpha)$ is its top-1 quantized accuracy evaluated on ImageNet validation set, $LAT(\\alpha)$ predicts the quantized latency on target device.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5822", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fe5f2339213858253dd54127a61124c59344e72f0e77b9f0955cff1d0685ac57.jpg", + "image_caption": [ + "(a) space search overview" + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 299, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a041b75c3603dccee23bc368b1dc4b48bb885ec8431fa871d738b5bdd8db5faa.jpg", + "image_caption": [ + "(b) stage-wise hyperspace", + "Figure 4. (a) We simplify space search into model search process; (b) Illustration of our hyperspace. A sampled search space is encoded by a sequential elastic stages. Contents in blue are searched: an elastic stage can search its block type and channel number list." + ], + "image_footnote": [], + "bbox": [ + 307, + 88, + 893, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, it's non-trivial to obtain the top-tier subnets from a candidate search, as if often involves an expensive full architecture search process. We adopt a zero-cost policy. Specifically, we randomly sample 5k subnets and select top 20 that under the latency constraints as the top-tier models to approximate the expectation term. The top 20 subnets are rapidly identified through the use of an accuracy lookup-table and a quantized latency predictor (Sec. 4.5).", + "bbox": [ + 75, + 271, + 468, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Elastic Stage and Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 398, + 415, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We observe that existing two-stage NAS adopt a chain-structured search space [4, 41, 35], which can be factorized as a sequence of STEM, HEAD and $N$ searchable stages. Each stage defines a range of configurations $c$ (e.g., kernel size, channel width, depth) for a specific block type $b$ , and allows NAS to find the optimal architecture settings.", + "bbox": [ + 75, + 416, + 468, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Elastic stage. Without loss of generality, we define a stage structure in a search space as elastic stage $E_{b,c}$ , which has elastic configurations $c$ for a fixed block $b$ . Suppose a search space $\\mathcal{A}$ has $N$ stages, it can be modularized as:", + "bbox": [ + 75, + 507, + 468, + 568 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} = S T E M \\circ E _ {b, c} ^ {1}, \\dots \\circ E _ {b, c} ^ {N} \\circ H E A D \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 573, + 468, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For instance, the search space in OQAT [32] and BatchQuant [1] can be factorized as 6 elastic stages and STEM (first Conv) and a classification head. Each elastic stage represents a set of configuration choices for the MBv3 block. Through the definition of elastic stage, we can simply use Eq. (2) to denote the contents of a model population.", + "bbox": [ + 75, + 597, + 468, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem definition. Operator type $b$ and configuration $c$ are two crucial objectives when searching quantization-friendly search space. Through the definition of elastic stage, the task of space search can be simplified to find a search space with optimal elastic stages, which has a similar goal with neural architecture search. We formulate our problem as:", + "bbox": [ + 75, + 690, + 468, + 781 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} \\left(E _ {b, c} ^ {1} \\circ E _ {b, c} ^ {2} \\circ \\dots \\circ E _ {b, c} ^ {N}\\right) ^ {*} = \\underset {E _ {b, c} ^ {i} \\in \\mathcal {H} ^ {i}} {\\arg \\max } \\mathcal {Q} \\left(\\mathcal {A} \\left(E _ {b, c} ^ {1} \\circ E _ {b, c} ^ {2} \\circ \\dots \\circ E _ {b, c} ^ {N}\\right), T\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 799, + 486, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{A}(\\cdot)$ denotes the search space, and $E_{b,c}^{i}$ is the $i^{th}$ elastic stage of $\\mathcal{A}(\\cdot)$ . $\\mathcal{H}$ denotes the hyperspace which covers all possible search spaces. $\\mathcal{Q}$ is the Q-T score for measuring a search space's quality.", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 4(a), given the constraints $T$ (i.e., a set of targeted quantized latency), SpaceEvo aims to find the optimal elastic stages $(E_{b,c}^{1} \\circ E_{b,c}^{2} \\circ \\dots \\circ E_{b,c}^{N})^{*}$ from the $1^{st}$ to $N^{th}$ stage for $\\mathcal{A}^*$ that has the maximum Q-T score: the top-tier quantized models can achieve best accuracy under the constraint $T$ . The constraint $T$ can be the quantized latency under any bits. In this work, we focus on the latency of INT8 quantized models, because INT8 quantization is widely supported by real-world edge devices and can achieve real speedup. Other lower bits typically require specific hardware (e.g., FPGA) to get the latency.", + "bbox": [ + 496, + 271, + 890, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Searching the Search Space", + "text_level": 1, + "bbox": [ + 498, + 450, + 746, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We now describe our evolutionary search algorithm that solves the problem in Eq. (3).", + "bbox": [ + 496, + 468, + 890, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hyperspace design. Analogous to NAS, hyperspace $\\mathcal{H}$ defines which search space a search algorithm might discover. Defining a hyperspace to cover many candidate search spaces for space search is a second-order problem for NAS, which can easily introduce high complexity. Fortunately, we can easily construct a large hyperspace through search space modularization in Eq. (2).", + "bbox": [ + 496, + 500, + 890, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We construct a large hyperspace in Fig. 4, in which a search space can be encoded by $N = 6$ sequential elastic stages along with STEM and HEAD. We search the following two dimensions for an elastic stage:", + "bbox": [ + 496, + 604, + 890, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Block (operator) type $b$ : MBv1 [15], MBv2 [31], MBv3 [14], residual bottleneck [13], residual bottleneck with SE, FusedMB [33] and FusedMB with SE. Conv is the major operator in residual bottleneck and FusedMB, thus they are quantization-friendly blocks; the efficiency of MB blocks relies on the device. For example, DWConv and SE are less quantization-efficient on Intel CPU.", + "bbox": [ + 500, + 676, + 890, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Output channel width list $\\text{cout}$ . In Sec. 3, we observe that quantized models can better utilize hardware under a larger channel number setting. However, directly increasing the channel numbers will also lead to longer latency. Therefore, we search the optimal stage-wise channel width list $\\text{cout}$ : $\\{w_{\\min}^*, \\dots, w_{\\max}^*\\}$ , which provides better channel width choices for final INT8 model search.", + "bbox": [ + 500, + 794, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5823", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, as described in Fig. 4(b), we define a wide range of $[w_{min}, w_{max}]$ by enlarging the channel widths in existing search spaces, and allow each elastic stage to choose a subset of $cout$ from $[w_{min}, w_{max}]$ .", + "bbox": [ + 88, + 90, + 468, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Besides channel widths, other configuration dimensions (e.g., kernel size) also impact a model's quantized latency. However, searching all dimensions leads to a large amount of choices in one stage, which exponentially enlarges the hyperspace size. Fortunately, other dimensions usually have a small space (e.g., kernel size selects from $\\{3,5,7\\}$ ). It's easy to find the optimal value for a model in the final NAS process. Therefore, we follow existing practices to configure the choices of kernel size, depth, and expand ratios. Our final searched INT8 model architectures SEQnet suggest that the optimal quantization-friendly kernel sizes and expand ratios are chosen. For example, kernel size of $3\\times 3$ brings more INT8 latency speedups for DWConv, and it is the dominate kernel size choice in DWConv related blocks.", + "bbox": [ + 75, + 156, + 468, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Suppose that a stage has $m$ choices of channel widths, there would be 7 (operator types) $\\times m$ candidates for each stage. In total, for a typical search space with $N = 6$ stages, the hyperspace has $\\sim 10^9$ candidate search spaces, which is extremely large and poses challenges for efficient search.", + "bbox": [ + 75, + 368, + 468, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evolutionary space search. The structure of hyperspace is similar to a typical model search space in NAS [12], so we can easily apply existing NAS search algorithms. Taking this advantage, we leverage aging evolution [30] to search the large hyperspace. We first randomly initialize a population of $P$ search spaces, where each sampled space is encoded as $(E_{b,c}^{1} \\circ E_{b,c}^{2} \\circ \\ldots \\circ E_{b,c}^{N})$ . Each individual is rapidly evaluated with Q-T score. After this, evolution improves the initial population in mutation iterations. At each iteration, we sample $S$ random candidates from the population and select the one with highest score as the parent. Then we alternately mutate the parent for block type and widths to generate two children search spaces. For instance, suppose the $i^{th}$ stage $E_{b,c}^{i}$ is selected for mutation, we first randomly modify its block type and produce $E_{b^*,c}^{i}$ for child 1, then we mutate the widths and produce $E_{b,c^*}^{i}$ for child 2. We evaluate their Q-T scores and add them to current population. The oldest two are removed for next iteration. After all iterations finish, we collect all the sampled space and select the one with best score as the final search space.", + "bbox": [ + 75, + 446, + 470, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Efficient Search Space Quality Evaluation", + "text_level": 1, + "bbox": [ + 76, + 762, + 434, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We now address the efficiency challenge caused by Q-T score evaluation. The most accurate evaluation is to get accuracy by training a supernet (search space) from scratch and measure latency on target device. However, it's impractical to conduct large-scale search due to the prohibitive cost. For example, it costs more than 10 days to train a supernet on 8 V100 GPUs [41]. To reduce the cost, we build an accurate INT8 latency predictor by nn-Meter [42], then", + "bbox": [ + 75, + 780, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8b4d198bbb2577c9cb031571f045c5bc0c5107162a63744ba63b29e70dcc5923.jpg", + "image_caption": [ + "Figure 5. We adopt block-wise knowledge distillation to reduce search space quality evaluation cost. We add two linear transformation layers (Conv $1 \\times 1$ ) to match teacher's feature map widths." + ], + "image_footnote": [], + "bbox": [ + 501, + 104, + 691, + 199 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d36163e3de82bb8f2f8d8bffeacfa366c0fe5d67270e5847b16c73eba7f92a57.jpg", + "table_caption": [ + "Step1: Quantize elastic stages via BKD Step2: Build quantized accuracy LUT" + ], + "table_footnote": [ + "k: kernel size, e: expand ratio, c: width" + ], + "table_body": "
Elastic stageModel block encoding (when depth=1)NSR loss
EMBv2,[32,48]k3e3c48-56x560.39
EMBv3,[32,48]k5e6c32-40x400.37
", + "bbox": [ + 697, + 104, + 890, + 174 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we introduce block-wise quantization scheme.", + "bbox": [ + 496, + 263, + 805, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Block-wise knowledge distillation (BKD) is firstly proposed in DNA [22] and then further improved in DONNA [26]. It originally uses block-wise representation of existing models (teacher) to supervise a corresponding student model block. This technique can provide a relative accuracy ranking of all possible models without requiring them to be trained from scratch. In our work, we extend BKD to supervise the training of all elastic stages (each contains a large amount of blocks).", + "bbox": [ + 496, + 277, + 890, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 5 illustrates the BKD process. In the first step, we use EfficientNet-B5 as the teacher, and separately train each elastic stage to mimic the behavior of corresponding teacher block by minimizing the NSR loss [26] between their output feature maps. Specifically, the $i^{th}$ stage receives the output of $(i - 1)^{th}$ teacher block as the input and is optimized to predict the output of $i^{th}$ teacher block with NSR loss. Since an elastic stage contains many blocks with different channel widths, we add two learnable linear transformation layers at the input and output for each elastic stage to match teacher's feature map shape. Moreover, we adopt sandwich rule [41] to sample four paths to improve the training efficiency. Each elastic stage is firstly trained for 5 epochs and then performed 1 epoch LSQ+ [3] for INT8 quantization.", + "bbox": [ + 496, + 414, + 892, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Accuracy lookup table. In the second step, we construct a INT8 accuracy lookup table to reduce the evaluation cost. Specifically, we evaluate all possible blocks in each elastic stage and record their NSR losses on the validation set in the lookup table. The quantized loss of a model is estimated by summing up the NSR loss of all its blocks by rapidly looking up each elastic stages from the table. We inverse the measured loss to approximate the quantized accuracy for Q-T score evaluation.", + "bbox": [ + 496, + 625, + 890, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our work, the BKD and lookup table construction can be sped up in a parallel way and finished in 1 day, which amounts a one-time cost before aging evolution search.", + "bbox": [ + 496, + 761, + 890, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Evaluation", + "text_level": 1, + "bbox": [ + 500, + 820, + 616, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Setup. We evaluate our method on ImageNet-1k dataset [9] and two popular edge devices. The INT8 latency constraints are \\{8, 10, 15, 20, 25\\} ms for Intel CPU, and \\{15, 20, 25, 30, 35\\} ms for Pixel4. For each device, we search 5k search", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5824", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e0c9fd3e30b55f048ff1e0edc3feb90f788e8f0c42dd5a885663cfb30176019f.jpg", + "image_caption": [ + "Figure 6. Best searched INT8 models with comparison to state-of-the-art NAS search spaces. Our searched spaces are proven to be the most quantization-friendly for the target device." + ], + "image_footnote": [], + "bbox": [ + 78, + 85, + 472, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "spaces in total and return the one with highest Q-T score. The population size $P$ is 500 and sample size $S$ is 125.", + "bbox": [ + 75, + 300, + 467, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Once SpaceEvo discovers a quantization-friendly search space for the target device, we train a quantized-for-all supernet. We start by pretraining a full-precision supernet without quantizers on ImageNet for 360 epochs. We adopt the sandwich rule and inplace distillation in BigNAS [41]. Then, we perform quantization-aware training (QAT) on the trained supernet for 50 epochs, which follows the same training protocol (i.e., sandwich rule and inplace distillation). We use $\\mathrm{LSQ + }$ as the QAT algorithm for better quantized accuracy. To derive INT8 model for deployment, we use the evolutionary search in OFA [4] to search $5\\mathrm{k}$ models for various given INT8 latency constraints. We list out detailed training settings in supplementary materials. In the following, we refer to the two searched spaces as SpaceEvo@CPU and SpaceEvo@Pixel4, the searched model families are SEQnet@cpu and SEQnet@pixel4.", + "bbox": [ + 75, + 330, + 468, + 573 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. The Effectiveness of SpaceEvo", + "text_level": 1, + "bbox": [ + 76, + 580, + 346, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with SOTA search spaces. To demonstrate the high-performance of our searched spaces, we compare with prior art manually-designed search spaces including: (1) MobileNetV3 search space that is adopted in two-stage quantization NAS OQAT [32] and BatchQuant [1]; (2) ProxylessNAS and AttentiveNAS search spaces that achieve superior performance on mobile devices; and (3) ResNet50 search space proposed by OFA that is a handcraft quantization-friendly space on our two devices. For fair comparison, we use one supernet training and QAT receipt for all search spaces. We conduct evolutionary search to compare the best-searched models from each search space. We use the random seed of 0. For all experiments, search space is the only difference.", + "bbox": [ + 75, + 598, + 467, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 6 compares the best searched INT8 models from different search spaces. SpaceEvo@CPU and SpaceEvo@Pixel4 consistently deliver superior quantized models than state-of-the-art search spaces. Under the same latency, the best quantized models from SpaceEvo@cpu significantly surpass the existing state-of-the-art search", + "bbox": [ + 75, + 810, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/22e93fac7a91fbf3ff786cef6341a5323d607b9c8cb05c3ec4191f7576693b4c.jpg", + "image_caption": [ + "Figure 7. Search space design under diverse INT8 latency constraints. SpaceEvo (6-25 ms) delivers superior tiny INT8 models." + ], + "image_footnote": [], + "bbox": [ + 535, + 85, + 851, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f428a6bf1ddef5c56eb20aee1981392e7bed3e642e14bd69ee4b1db1ffc7c59.jpg", + "image_caption": [ + "Figure 8. Search cost measured on 8 Nvidia V100 GPUs." + ], + "image_footnote": [], + "bbox": [ + 500, + 273, + 890, + 337 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "spaces with $+0.7\\%$ to $+3.8\\%$ ( $+0.4\\%$ to $+3.2\\%$ on Pixel4) higher accuracy. Moreover, our search space is the only one that is able to deliver superior quantized models under both extremely low (only $\\sim 5\\mathrm{ms}$ ) and large latency constraints.", + "bbox": [ + 496, + 367, + 890, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SpaceEvo under diverse latency constraints. We extensively study the effectiveness of SpaceEvo under different latency constraints. Specifically, we perform space search under two tight constraints of $\\{10, 15, 20, 25, 30\\}\\mathrm{ms}$ and $\\{6, 10, 15, 20, 25\\}\\mathrm{ms}$ on Pixel4. The results are shown in Fig. 7. Our proposed method can handle the diverse latency requirements and produce high-quality spaces. As expected, the searched spaces under $10 - 30\\mathrm{ms}$ and $6 - 25\\mathrm{ms}$ have much more low-latency quantized models.", + "bbox": [ + 496, + 429, + 892, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further verify the effectiveness of these low-latency models, we compare with existing SOTA tiny models. Significantly, even under the extremely low latency constraints of 6-25 ms, our searched space delivers very competitive tiny quantized models. Compared to the smallest model ShuffleNetV2x0.5, we can achieve $+10.1\\%$ higher accuracy under the same latency of 4.3 ms.", + "bbox": [ + 496, + 566, + 890, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Search cost. As depicted in Fig. 8, our algorithm, SpaceEvo, is designed to be lightweight and suitable for real-world usage, requiring only 25 GPU hours to search a space of 5000 iterations. This remarkable speed is mainly due to our block-wise search space quantization scheme, which significantly reduces the cost of search space quality evaluation. In comparison, Fig. 8 demonstrates that training each search space from scratch without this scheme would consume an impractical 1200k GPU hours.", + "bbox": [ + 496, + 672, + 890, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. The Effectiveness of Discovered INT8 Models", + "text_level": 1, + "bbox": [ + 500, + 821, + 883, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we demonstrate that our searched spaces deliver state-of-the-art quantized models. We compare with two strong baselines: (1) prior art manually-designed and NAS-searched models; and (2) quantization-aware NAS. For", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5825", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/92ba970c0b26a37c35a9d9bc659cf5c7452ae61cea13b36bc648b303520236c6.jpg", + "table_caption": [ + "Table 1. ImageNet results compared with SOTA quantized models on two devices. *: latency compared to FP32 inference." + ], + "table_footnote": [], + "table_body": "
(a) Results on the Intel VNNI CPU with onnxruntime
ModelAcc% INT8CPU latencyFP32 FLOPs
INT8speedup*
MobileNetV3Small66.34.4 ms1.1×67.456M
SEQnet@cpu-A074.74.4 ms2.0×74.8163M
MobileNetV271.47.3 ms2.2×72.0300M
ProxylessNAS-R74.68.8 ms1.8×74.6320M
OQAT-8bit74.89.8 ms1.8×75.2214M
MobileNetV3Large74.510.3 ms1.5×75.2219M
OFA (#25)75.611.2 ms1.5×76.4230M
SEQnet@cpu-A177.48.8 ms2.4×77.5358M
APQ-8bit73.615.0 ms1.5×73.6297M
AttentiveNAS-A076.115.1 ms1.4×77.3203M
OQAT-8bit76.314.9 ms1.7×76.7316M
EfficientNet-B076.718.1 ms1.6×77.3390M
SEQnet@cpu-A278.514.1 ms2.4×78.8638M
APQ-8bit74.920.0 ms1.5×75.0393M
OQAT-8bit76.919.5 ms1.6×77.3405M
AttentiveNAS-A177.222.4 ms1.4×78.4279M
AttentiveNAS-A277.522.5 ms1.3×78.8317M
SEQnet@cpu-A379.518.9 ms2.6×79.6981M
FBNetV2-L175.825.0 ms1.2×77.2325M
FBNetV3-A78.227.7 ms1.3×79.1357M
SEQnet@cpu-A480.024.4 ms2.4×80.11267M
(b) Results on the Google Pixel 4 with TFLite
ModelAcc% INT8Pixel4 Latency|Acc% FP32FLOPs
INT8speedup*
MobileNetV3Small66.36.4 ms1.3×67.456M
SEQnet@pixel4-A073.65.9 ms2.1×73.7107M
MobileNetV271.416.5 ms1.9×72.0300M
ProxylessNAS-R74.618.4 ms1.8×74.6320M
MobileNetV3Large74.515.7 ms1.5×75.2219M
APQ-8bit74.614.9 ms2.0×74.4340M
OFA (#25)75.614.8 ms1.7×76.4230M
OQAT-8bit75.815.2 ms1.9×76.2287M
AttentiveNAS-A076.115.2 ms2.0×77.3203M
SEQnet@pixel4-A177.614.7 ms2.2×77.7274M
APQ-8bit75.120.0 ms1.9×75.1398M
OQAT-8bit76.520.4 ms1.8×76.8347M
AttentiveNAS-A177.221.1 ms2.0×78.4279M
AttentiveNAS-A277.522.7 ms2.0×78.8317M
SEQnet@pixel4-A278.319.4 ms2.3×78.4402M
FBNetV2-L175.826.7 ms1.5×77.2325M
OQAT-8bit77.029.9 ms1.7×77.2443M
FBNetV3-A78.230.5 ms1.5×79.1357M
SEQnet@pixel4-A379.530.8 ms2.1×79.5591M
EfficientNet-B076.736.4 ms1.7×77.3390M
SEQnet@pixel4-A479.935.5 ms2.2×80.0738M
", + "bbox": [ + 76, + 114, + 467, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "baseline (1), we collect official pre-trained FP32 checkpoints and conduct LSQ+ QAT to get the quantized accuracy. The hyperparameter settings follow the original LSQ+ paper, except that we set a larger epoch of 10 to achieve better accuracy. The latency numbers are measured on our devices. For (2), we compare with strong baselines including APQ [37] and OQAT [32]. Specifically, we limit APQ to search for the fixed 8bit (INT8) models. Since OQAT has no 8bit supernet checkpoint, we follow the official source code and conduct supernet QAT for 50 epochs. The final INT8 models are searched under the same INT8 latency predictors for fair comparison.", + "bbox": [ + 75, + 698, + 468, + 878 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. Table 1 summarizes comparison results. Re", + "bbox": [ + 76, + 885, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/336c065d35c53d8fbd1aa4a5c588f7fb4dbd983be893ef492602f998a15cf25b.jpg", + "image_caption": [ + "Figure 9. Q-T score effectiveness (Kendall's $\\tau$ ) on ranking search space quality. We achieve a high space ranking correlation." + ], + "image_footnote": [], + "bbox": [ + 537, + 87, + 851, + 231 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/93c76b942af3b77b0a068b85057dea1cff9908612d848c4259b6bf8d346ec8e4.jpg", + "table_caption": [ + "Table 2. Different space search methods and their best resulting quantized models on Pixel4. Baseline is a SOTA mobile-friendly AttentiveNAS space. *: the search dimension use the same settings in AttentiveNAS." + ], + "table_footnote": [], + "table_body": "
MethodOpWidthBest quantized models
10ms15ms20ms30ms36ms
Baseline--76.677.279.079.5
SpaceEvo-opsearchfix*75.076.677.878.678.8
SpaceEvo-widthfix*search75.477.478.079.179.5
SpaceEvosearchsearch75.777.678.379.579.9
", + "bbox": [ + 501, + 325, + 888, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "markably, our searched model family, SEQnet significantly outperform SOTA efficient models and quantization-aware NAS searched models, with higher INT8 quantized accuracy, lower INT8 latency and better speedups. Without finetuning, our tiny models SEQnet@cpu-A0 and SEQnet@pixel4-A0 achieve $74.7\\%$ and $73.6\\%$ top1 accuracy on ImageNet, which is $8.4\\%$ and $7.3\\%$ higher than MobileNetV3-Small (56M FLOPs) while maintaining the same level quantized latency. For larger models, SEQnet@cpu-A4 $(80.0\\%)$ outperforms FBNetV3-A with $1.8\\%$ higher accuracy while runs 3.3ms faster. In particular, to achieve the same level accuracy (i.e., around $77.2\\%$ ), AttentiveNAS-A1 has 22.4ms latency while SEQnet@cpu-A1 $(77.4\\%)$ only needs 8.8 ms $(2.6\\times$ faster). More importantly, our searched models can better utilize the INT8 hardware optimizations: the latency speedups compared to full-precision inference are all larger than $2\\times$ , and this leaves room to search large-size models with higher accuracy.", + "bbox": [ + 496, + 407, + 890, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 698, + 653, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Q-T score effectiveness. Q-T score is crucial as it guides the space evolution process. To evaluate its effectiveness, we randomly sample 30 search spaces, and measure the rank correlation (Kendall's $\\tau$ ) between their Q-T score and their actual Pareto-frontier models' accuracies. Specifically, we use Intel CPU as the test device and set a same latency constraints of $\\{8, 10, 15, 20, 25\\}$ ms. For each sampled space, we train it from scratch for 50 epochs, and conduct evolutionary search to get the Pareto-frontier models' accuracies. As shown in Fig. 9, the Kendall's $\\tau$ between the Q-T score and the actual Pareto-frontier models' accuracies is 0.8, which indicates a very high rank correlation.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5826", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study on two search dimensions. In Sec. 3, we conclude that operator type and configuration are two key factors impacting INT8 latency, which serves as the two search objectives of SpaceEvo. To verify the effectiveness, we create two strong baselines based on the SOTA edge-regime AttentiveNAS search space: (1) SpaceEvo-op: we fix each elastic stage's width to AttentiveNAS space, then allow each elastic stage to search for the optimal operator; and (2) SpaceEvo-width: we fix all elastic stages' block types to AttentiveNAS space, then search for the optimal width. Table 2 reports the space comparison between different search methods on the Pixel4. By searching both operator type and width, SpaceEvo finds the optimal search space where its best searched quantized models achieve the highest accuracy under all latency constraints. Moreover, even searching for one dimension, SpaceEvo-op and SpaceEvo-width outperform the manually-designed AttentiveNAS space under small latency constraints.", + "bbox": [ + 76, + 90, + 472, + 362 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Search space design implications. We now summarize our learned experience and implications for designing quantization-friendly search spaces. We notice that the searched spaces show different preferences when targeting different devices: (i) All stages should use much wider channel widths compared to existing manually-designed spaces on the cpu, while only early stages prefer wider channels on Pixel 4. (ii) Since SE and Swish are INT8 latency-friendly on mobile phones, so our auto-generated search spaces for Pixel4 have many MBv3 stages. On Intel CPU, INT8 quantization slows down SE, Hardwish, and Swish, making FusedMB and MBv2 the priority for search spaces, with only the last two stages using MBv3. The details are provided in supplementary.", + "bbox": [ + 76, + 363, + 472, + 575 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 588, + 194, + 603 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we introduced SpaceEvo, the first to automatically design a quantization-friendly space for target device, which delivers superior INT8 quantized models with SOTA efficiency on real-world edge devices. Extensive experiments on two popular devices demonstrate its effectiveness compared to prior art manual-designed search spaces. We plan to apply SpaceEvo to other hardware efficiency such as energy-efficient search space design in the future.", + "bbox": [ + 76, + 607, + 468, + 729 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 742, + 173, + 757 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Haoping Bai, Meng Cao, Ping Huang, and Jiulong Shan. Batchquant: Quantized-for-all architecture search with robust quantizer. In NeurIPS, 2021. 1, 3, 5, 7", + "[2] Ron Banner, Yury Nahshan, and Daniel Soudry. Post training 4-bit quantization of convolutional networks for rapid-deployment. In Advances in Neural Information Processing Systems, 2019. 3", + "[3] Yash Bhalgat, Jinwon Lee, Markus Nagel, Tijmen Blankevoort, and Nojun Kwak. Lsq+: Improving low-bit" + ], + "bbox": [ + 84, + 766, + 467, + 893 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "quantization through learnable offsets and better initialization. In CVPRW, 2020. 1, 3, 6", + "[4] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. In ICLR, 2020. 2, 3, 5, 7", + "[5] Han Cai, Ligeng Zhu, and Song Han. ProxylessNAS: Direct neural architecture search on target task and hardware. In International Conference on Learning Representations, 2019. 3", + "[6] Zhaowei Cai and Nuno Vasconcelos. Rethinking differentiable search for mixed-precision neural networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2346–2355. Computer Vision Foundation / IEEE, 2020. 1, 3", + "[7] Minghao Chen, Kan Wu, Bolin Ni, Houwen Peng, Bei Liu, Jianlong Fu, Hongyang Chao, and Haibin Ling. Searching the search space of vision transformers. In NeurIPS, 2021. 2, 3", + "[8] Yuanzheng Ci, Chen Lin, Ming Sun, Boyu Chen, Hongwen Zhang, and Wanli Ouyang. Evolving search space for neural architecture search. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6639-6649, 2021. 2, 3", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248-255. IEEE Computer Society, 2009. 6", + "[10] Steven K. Esser, Jeffrey L. McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S. Modha. Learned step size quantization. In ICLR, 2020. 1, 3", + "[11] Google. Tensorflow lite, 2022. 3", + "[12] Zichao Guo, Xiangyu Zhang, Haoyuan Mu, Wen Heng, Zechun Liu, Yichen Wei, and Jian Sun. Single path one-shot neural architecture search with uniform sampling. In ECCV, 2020. 1, 2, 3, 6", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5", + "[14] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, Quoc V. Le, and Hartwig Adam. Searching for mobilenetv3. In ICCV, 2019. 2, 5", + "[15] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. CoRR, abs/1704.04861, 2017. 5", + "[16] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 2", + "[17] Yiming Hu, Yuding Liang, Zichao Guo, Ruosi Wan, Xiangyu Zhang, Yichen Wei, Qingyi Gu, and Jian Sun. Angle-based search space shrinking for neural architecture search. In ECCV. 3" + ], + "bbox": [ + 501, + 92, + 890, + 897 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5827", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Intel. Intel 64 and ia-32 architectures software developer's manual, 2022. 4", + "[19] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 1, 3", + "[20] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2704-2713, 2018. 3", + "[21] Sumin Kim, Gunju Park, and Youngmin Yi. Performance evaluation of int8 quantized inference on mobile gpus. IEEE Access, 9:164245-164255, 2021. 4", + "[22] Changlin Li, Jiefeng Peng, Liuchun Yuan, Guangrun Wang, Xiaodan Liang, Liang Lin, and Xiaojun Chang. Blockwisely supervised neural architecture search with knowledge distillation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, 2020. 2, 6", + "[23] Xiang Li, Chen Lin, Chuming Li, Ming Sun, Wei Wu, Junjie Yan, and Wanli Ouyang. Improving one-shot nas by suppressing the posterior fading. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[24] Christos Louizos, Mathias Reisser, Tijmen Blankevoort, Efstratos Gavves, and Max Welling. Relaxed quantization for discretized neural networks. In ICLR, 2019. 3", + "[25] Microsoft. onnxruntime, 2022. 3", + "[26] Bert Moons, Parham Noorzad, Andrii Skliar, Giovanni Mariani, Dushyant Mehta, Chris Lott, and Tijmen Blankevoort. Distilling optimal neural networks: Rapid search in diverse spaces. In ICCV, 2021. 2, 6", + "[27] Markus Nagel, Mart van Baalen, Tijmen Blankevoort, and Max Welling. Data-free quantization through weight equalization and bias correction. In ICCV, 2019. 1, 3", + "[28] Asaf Noy, Niv Nayman, Tal Ridnik, Nadav Zamir, Sivan Doveh, Itamar Friedman, Raja Giryes, and Lihi Zelnik. Asap: Architecture search, anneal and prune. In Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, pages 493-503, 2020. 3", + "[29] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In CVPR, 2020. 3", + "[30] Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image classifier architecture search. In AAAI, 2019. 2, 4, 6", + "[31] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5" + ], + "bbox": [ + 78, + 90, + 468, + 840 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Mingzhu Shen, Feng Liang, Ruihao Gong, Yuhang Li, Chuming Li, Chen Lin, Fengwei Yu, Junjie Yan, and Wanlin Ouyang. Once quantization-aware training: High performance extremely low-bit architecture search. In ICCV, 2021. 1, 3, 5, 7, 8", + "[33] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In Proceedings of the 38th International Conference on Machine Learning, pages 10096-10106. PMLR, 2021. 5", + "[34] Xiaohu Tang, Shihao Han, Li Lyna Zhang, Ting Cao, and Yunxin Liu. To bridge neural network design and real-world performance: A behaviour study for neural networks. In A. Smola, A. Dimakis, and I. Stoica, editors, Proceedings of Machine Learning and Systems, volume 3, pages 21-37, 2021. 3, 4", + "[35] Dilin Wang, Meng Li, Chengyue Gong, and Vikas Chandra. Attentivenas: Improving neural architecture search via attentive sampling. In Conference on Computer Vision and Pattern Recognition, 2021. 3, 5", + "[36] Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, and Song Han. Haq: Hardware-aware automated quantization with mixed precision. In CVPR, 2019. 1, 3", + "[37] Tianzhe Wang, Kuan Wang, Han Cai, Ji Lin, Zhijian Liu, and Song Han. Apq: Joint search for network architecture, pruning and quantization policy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 1, 3, 8", + "[38] Xudong Wang, Li Lyna Zhang, Yang Wang, and Mao Yang. Towards efficient vision transformer inference: A first study of transformers on mobile devices. In Proceedings of the 23rd Annual International Workshop on Mobile Computing Systems and Applications, HotMobile '22, page 1-7. Association for Computing Machinery, 2022. 3", + "[39] Bichen Wu, Yanghan Wang, Peizhao Zhang, Yuandong Tian, Peter Vajda, and Kurt Keutzer. Mixed precision quantization of convnets via differentiable neural architecture search. 2019. 1, 3", + "[40] Xin Xia, Xuefeng Xiao, Xing Wang, and Min Zheng. Progressive automatic design of search space for one-shot neural architecture search. In WACV, 2022. 3", + "[41] Jiahui Yu, Pengchong Jin, Hanxiao Liu, Gabriel Bender, Pieter-Jan Kindermans, Mingxing Tan, Thomas Huang, Xiaodan Song, Ruoming Pang, and Quoc Le. Bignas: Scaling up neural architecture search with big single-stage models. In ECCV, 2020. 3, 5, 6, 7", + "[42] Li Lyna Zhang, Shihao Han, Jianyu Wei, Ningxin Zheng, Ting Cao, Yuqing Yang, and Yunxin Liu. nn-meter: Towards accurate latency prediction of deep-learning model inference on diverse edge devices. In Proceedings of the 19th Annual International Conference on Mobile Systems, Applications, and Services, page 81-93, New York, NY, USA, 2021. ACM. 3, 6" + ], + "bbox": [ + 501, + 92, + 890, + 820 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5828", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/SpaceEvo_ Hardware-Friendly Search Space Design for Efficient INT8 Inference/7d8d9b36-b64d-45d0-8aa5-d17e8abbee45_model.json b/2023/SpaceEvo_ Hardware-Friendly Search Space Design for Efficient INT8 Inference/7d8d9b36-b64d-45d0-8aa5-d17e8abbee45_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0a927c250f42516f3cc868d5e7aa4eefba54a01c --- /dev/null +++ b/2023/SpaceEvo_ Hardware-Friendly Search Space Design for Efficient INT8 Inference/7d8d9b36-b64d-45d0-8aa5-d17e8abbee45_model.json @@ -0,0 +1,2200 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.131, + 0.892, + 0.153 + ], + "angle": 0, + "content": "SpaceEvo: Hardware-Friendly Search Space Design for Efficient INT8 Inference" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.192, + 0.806, + 0.231 + ], + "angle": 0, + "content": "Xudong Wang\\*1,2§ Li Lyna Zhang\\*2‡ Jiahang Xu² Quanlu Zhang² Yujing Wang³ Yuqing Yang² Ningxin Zheng² Ting Cao² Mao Yang²" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.239, + 0.735, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)Shanghai Jiao Tong University \\(^{2}\\)Microsoft Research \\(^{3}\\)Microsoft" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.258, + 0.856, + 0.275 + ], + "angle": 0, + "content": "xudongwang176@gmail.com; {Izhani, jiahangxu, quzha, yujwang, yuqyang, nizhen, ticao, maoyang}@microsoft.com" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.275, + 0.756, + 0.291 + ], + "angle": 0, + "content": "https://github.com/microsoft/Moonlit/tree/main/SpaceEvo" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.335, + 0.314, + 0.351 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.474, + 0.776 + ], + "angle": 0, + "content": "The combination of Neural Architecture Search (NAS) and quantization has proven successful in automatically designing low-FLOPs INT8 quantized neural networks (QNN). However, directly applying NAS to design accurate QNN models that achieve low latency on real-world devices leads to inferior performance. In this work, we identify that the poor INT8 latency is due to the quantization-unfriendly issue: the operator and configuration (e.g., channel width) choices in prior art search spaces lead to diverse quantization efficiency and can slow down the INT8 inference speed. To address this challenge, we propose SpaceEvo, an automatic method for designing a dedicated, quantization-friendly search space for each target hardware. The key idea of SpaceEvo is to automatically search hardware-preferred operators and configurations to construct the search space, guided by a metric called \\( Q-T \\) score to quantify how quantization-friendly a candidate search space is. We further train a quantized-for-all supernet over our discovered search space, enabling the searched models to be directly deployed without extra retraining or quantization. Our discovered models, SEQnet, establish new SOTA INT8 quantized accuracy under various latency constraints, achieving up to \\( 10.1\\% \\) accuracy improvement on ImageNet than prior art CNNs under the same latency. Extensive experiments on real devices show that SpaceEvo consistently outperforms manually-designed search spaces with up to \\( 2.5\\times \\) faster speed while achieving the same accuracy." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.787, + 0.21, + 0.803 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.807, + 0.47, + 0.854 + ], + "angle": 0, + "content": "INT8 Quantization[27, 19, 10, 3] is a widely used technique for deploying DNNs on edge devices by reducing \\(4 \\times\\) in model size and memory cost for full-precision (FP32)" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.332, + 0.885, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.536, + 0.894, + 0.593 + ], + "angle": 0, + "content": "Figure 1. INT8 quantized model error distributions (EDF) of different NAS search spaces. We propose to search quantization-friendly search space for each hardware, which yields significantly better INT8 model populations than SOTA search spaces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.893, + 0.699 + ], + "angle": 0, + "content": "models. However, prior art DNN models achieve only marginal speedup from INT8 quantization (in Fig. 2(a)), the still high latency after quantization making them difficult to deploy in latency-critical scenarios. Designing models that achieve high accuracy and low latency after quantization becomes the important but challenging problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Neural Architecture Search (NAS) is a powerful tool for automating efficient quantized model design [36, 39, 12, 37, 6]. Recently, OQAT [32] and BatchQuant [1] achieve remarkable search efficiency and accuracy by adopting a two-stage paradigm. The first stage trains a weight-shared quantized supernet assembling all candidate architectures in the search space. This allows all the sub-networks (subnets) to simultaneously reach comparable quantized accuracy as when trained from scratch individually. The second stage uses typical search algorithms to find subnets with best quantized accuracy under different FLOPs constraints. This approach avoids the need to retrain each subnet for accuracy evaluation, greatly improving the search efficiency." + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.862, + 0.204, + 0.876 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.876, + 0.417, + 0.888 + ], + "angle": 0, + "content": "\\(^{\\S}\\)Work was done during the internship at Microsoft Research" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.352, + 0.9 + ], + "angle": 0, + "content": "†Corresponding author (lzhani@microsoft.com)" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.862, + 0.417, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5819" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "Though promising in optimizing model FLOPs, we identify a significant challenge when directly applying two-stage NAS to low quantized latency scenarios. This is due to the quantization-unfriendly search space issue, where prior art search spaces can unexpectedly impede INT8 latency. Consequently, since INT8 quantization yields only a marginal speedup, we are forced to search for small-sized models to fulfill latency criteria, which can unfortunately restrict NAS to find better quantized models for edge devices. Then, a question naturally arises: Can we design a quantization-friendly search space, allowing NAS to discover larger and superior models that meet the low INT8 latency requirements?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.473, + 0.518 + ], + "angle": 0, + "content": "We start by conducting an in-depth study to understand the factors that determine INT8 quantized latency and how they affect search space design. Our study shows: (1) both operator type and configurations (e.g., channel width) greatly impact the INT8 latency; Improper selections can slow down the INT8 latency. For instance, Squeeze-and-Excitation (SE) [16] and Hardswish [14] are widely-used operators in current search spaces as it improves accuracy with little latency introduced. However, their INT8 inference speeds are slower than FP32 inference on Intel CPU (Fig. 3(a)), because the extra costs (e.g., data transformation between INT32 and INT8) introduced by quantization outweigh the latency reduction by INT8 computation. (2) The quantization efficiency varies across different devices, and the preferred operator types can be contradictory." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.473, + 0.748 + ], + "angle": 0, + "content": "The above study motivates us to design specialized quantization-friendly search spaces for each hardware, rather than relying on a single, large search space as seen in SPOS [12] for all hardware, which provides different operator options per layer. This is because two-stage NAS requires the search space to adhere to a specific condition for training the supernet, wherein each layer must utilize a fixed operator. Our study indicates significant variations in optimal operators across different hardware. Thus, customizing the search space for each hardware is crucial for optimal results. However, designing such specialized quantization-friendly search spaces for various edge devices presents a significant challenge, requiring expertise in both AI and hardware domains, as well as many trial and error attempts to optimize accuracy and INT8 latency for each hardware." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose SpaceEvo, a novel method for automatically designing specialized quantization-friendly search space for each hardware. The search space is comprised of hardware-preferred operators and configurations, enabling the search of larger and better models with low INT8 latency. With the discovered search space, we leverage two-stage quantization NAS to train a quantized-for-all supernet, and utilize evolution search [4] to find best quantized models under various INT8 latency constraints. Our approach addresses three key challenges: (1) What is the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "definition of a quantization-friendly search space in terms of both quantized accuracy and latency? (2) How to automatically design a search space without human expertise? (3) How to handle with the prohibitive cost caused by quality evaluation of a candidate search space?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.169, + 0.892, + 0.259 + ], + "angle": 0, + "content": "To address the first challenge, we propose a latency-aware \\(Q-T\\) score to quantify the effectiveness of a candidate search space, which measures the INT8 accuracy-latency quality of top-tier subnets in a search space. The behind intuition is that the goal of NAS is to search top subnets with better accuracy-latency tradeoffs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.26, + 0.893, + 0.501 + ], + "angle": 0, + "content": "Then, we introduce an evolutionary-based search algorithm that can effectively search a quantization-friendly search space with highest \\( Q-T \\) score. Searching a search space involves discovering a collection of model population that contains billions of models, which is challenging and easily introduce complexity. To address this challenge, we propose to factorize and encode a search space into a sequence of elastic stages, which have flexible operator types and configurations. Through this design, the task of search space design is then simplified to find a search space with the optimal elastic stages, so that existing search algorithms can be easily applied. Specifically, we design a stage-wise hyperspace to include many candidate search spaces and leverage aging evolution [30] to perform random mutations of elastic stages for search space evolution. The evolution is guided by maximizing the \\( Q-T \\) score." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.893, + 0.743 + ], + "angle": 0, + "content": "Finally, estimating the quality score (Q-T score) of a search space involves a costly training process for evaluating the accuracy of sub-networks, which presents a significant obstacle for our evolutionary algorithm. Naively adopting a two-stage NAS approach, training a supernet for each candidate search space [8, 7], is prohibitively expensive, taking of thousands GPU hours. To address this issue, we draw inspiration from block-wise knowledge distillation [26, 22] and propose a block-wise search space quantization scheme. This scheme trains each elastic stage separately and rapidly estimates a model's quantized accuracy by summing block-level loss with a quantized accuracy lookup table, as shown in Fig. 5. This significantly reduces the training and evaluation costs, while providing effective accuracy rankings among search spaces. We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.751, + 0.892, + 0.827 + ], + "angle": 0, + "content": "- We study the INT8 quantization efficiency on real-world edge devices and find that the choices of operator types and configurations in a quantized model can significantly impact the INT8 latency, leaving a huge room for design optimization of quantized models." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "- We propose SpaceEvo to automatically design a hardware-dedicated quantization-friendly search space and leverage two-stage quantization NAS to produce superior INT8 models under various latency constraints." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5820" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "- We present three innovative techniques that enable the first-ever efficient and cost-effective evolution search to explore a search space comprising billions of models." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.146, + 0.471, + 0.312 + ], + "angle": 0, + "content": "- Extensive experiments on two real-world edge devices and ImageNet demonstrate that our automatically designed search spaces significantly surpass manually-designed search spaces. Our discovered models, SEQnet, establish the new state-of-the-art INT8 quantized accuracy-latency tradeoffs. For instance, SEQnet@cpuA4, achieves \\(80.0\\%\\) accuracy on ImageNet, which is 3.3ms faster with \\(1.8\\%\\) higher accuracy than FBNetV3-A. Moreover, SpaceEvo delivers superior tiny models, achieving up to \\(10.1\\%\\) accuracy improvement over ShuffleNetV2x0.5 (41M FLOPs, 4.3ms)." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.092, + 0.471, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.324, + 0.228, + 0.34 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.47, + 0.485 + ], + "angle": 0, + "content": "Quantization has been widely used for efficiency in deployment. Extensive efforts can be classified into posttraining quantization (PTQ) [27, 2] and quantization-aware training (QAT) [19, 24, 10, 3]. QAT generally outperforms PTQ in quantizing compact DNNs to 8bit and very low-bit (2, 3, 4bit) by finetuning the quantized weights. Despite their success, traditional quantization methods focus on minimizing accuracy loss for a given pre-trained model, but ignore the real-world inference efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.47, + 0.637 + ], + "angle": 0, + "content": "NAS for Quantization. Early works [36, 39, 12, 37, 6] formulates mixed-precision problem into NAS to search layer bit-width with a given architecture. Recently, [32, 1] train a quantized-for-all supernet to search both architecture and bit-width. The searched models can be directly evaluated with comparable accuracy to train-from-scratch. However, little attention is paid on optimizing quantized latency on real-world devices. Through searching quantization-friendly search space, our discovered quantized models can achieve both high accuracy and low latency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.637, + 0.471, + 0.789 + ], + "angle": 0, + "content": "Search Space Design. Starting from [5], the manually-designed MBConv-based space becomes the dominant in most NAS works [5, 4, 41, 35]. RegNet [29] is the first to present standard guidelines to optimize a search space by each dimension. Recently, [17, 28, 23, 40, 8, 7] propose to shrink to a better compact search space by either pruning unimportant operators or configurations. However, these works focus on optimizing the accuracy and little attention is paid on quantization-friendly search space design. Our work is the first lightweight solution towards this direction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.466, + 0.818 + ], + "angle": 0, + "content": "3. On-device Quantization Efficiency Analysis" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To understand what factors lead to quantization-unfriendly issue, we conduct a comprehensive study on two widely-used edge devices equipped with high-performance inference engine: an Intel CPU device supported with VNNI instructions and onnxruntime [25] (Intel CPU) and" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.088, + 0.707, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.214, + 0.699, + 0.222 + ], + "angle": 0, + "content": "(a) FLOPs and INT8 latency on Intel CPU" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.089, + 0.895, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.212, + 0.875, + 0.222 + ], + "angle": 0, + "content": "(b) FP32 and INT8 latency on Pixel 4" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.235, + 0.894, + 0.278 + ], + "angle": 0, + "content": "Figure 2. INT8 latency and speedups (annotated) for SOTA models. FLOPs and FP32 latency are not good indicators of INT8 latency; Compact models have very limited INT8 speedup \\((\\sim 1.5\\times)\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.292, + 0.892, + 0.337 + ], + "angle": 0, + "content": "a Pixel 4 phone CPU with TFLite 2.7 [11] (Pixel 4). Note that we follow existing practices [42, 34, 38] to measure the latency. We reveal key observations as follows:" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.34, + 0.892, + 0.37 + ], + "angle": 0, + "content": "Observation 1: FP32 latency and FLOPs are not good indicators of INT8 latency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.893, + 0.522 + ], + "angle": 0, + "content": "To deploy on edge devices, a common belief is that a compact model with low FLOPs or FP32 latency is preferred than a larger model. However, Fig. 2 shows that neither of them is a good indicator of INT8 latency. In Fig. 2(b), a very large model (ResNet18) can be even faster than a compact model (EfficientNet-B0) after quantization. Moreover, the recent SOTA compact models searched by OFA [4] and AttentiveNAS [35] all have marginal INT8 speedups, suggesting that optimizing FLOPs and FP32 latency can not lead to lower INT8 latency." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.523, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Observation 2: The choices of operators' types and configurations greatly impact the INT8 latency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.894, + 0.719 + ], + "angle": 0, + "content": "The prior art search spaces adopted in recent two-stage NAS works are MobileNetV2 or MobileNetV3 based chain-structures, where each search space comprises a sequence of blocks (stages). The block type is fixed to the MBCnv and is allowed to search from a handcraft range of hyperparameter configurations including kernel size, expansion ratio, channel width and depth, which are designed with human wisdom. For instance, many works [5, 35] observe that edge-regime CNNs prefer deeper depths and narrower channels, and manually set small channel numbers but large depths in the search space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.894, + 0.901 + ], + "angle": 0, + "content": "However, we find that many block type and configuration choices in current search spaces unexpectedly slow down the INT8 latency. We first study the operator type impact in Fig. 3(a). SE and Hardswish are lightweight operators in edge-regime search spaces, but their INT8 inference becomes slower on Intel CPU. Compared to Conv, DWConv can greatly reduce the FLOPs, but it benefits less from INT8 quantization. The root cause is that quantization introduces extra cost, such as (1) data transformation between INT32 and INT8 [34], and (2) additional computation caused by scaling fators and zero points [20]. If the operator has low data-reuse-rate, such as the activation functions (Hswish)," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5821" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.092, + 0.253, + 0.182 + ], + "angle": 0, + "content": "
OperatorIntel CPUPixel 4
Conv2.6x2.5x
DWConv1.2x2.0x
SE0.7x1.4x
Hardswish0.7x0.5x
Swish0.8x2.1x
" + }, + { + "type": "image_caption", + "bbox": [ + 0.079, + 0.184, + 0.241, + 0.216 + ], + "angle": 0, + "content": "(a) Avg. INT8 latency speedup (relative to FP32) of major operators in MobileNetV3." + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.089, + 0.473, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.201, + 0.431, + 0.213 + ], + "angle": 0, + "content": "(b) Conv1x1 speedups" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.228, + 0.47, + 0.27 + ], + "angle": 0, + "content": "Figure 3. (a) The choice of operator type leads to significantly different quantized speedup. (b) Conv1x1 speedups with various channel numbers. Config: \\( HW = 28 \\), \\( C_{out} = 4 \\times C_{in} \\) (expand=4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.285, + 0.469, + 0.331 + ], + "angle": 0, + "content": "the extra cost may outweigh the latency reduction by the low-bit computation. For high data-reuse operators (Conv), this cost is amortized and thus achieve large speedup [34]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.469, + 0.468 + ], + "angle": 0, + "content": "Besides the operator type, the configuration choices also determine the quantization efficiency. Fig. 3(b) shows the speedups of \\(\\mathrm{Conv}1\\times 1\\) under various channel widths. Results show that small channel widths in OFA search space cannot benefit well from quantization. This is because the additional quantization cost has a large impact when the channel width is small, limiting the latency acceleration. In contrast, SpaceEvo can automatically design a search space with larger channel widths for better efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.471, + 0.469, + 0.501 + ], + "angle": 0, + "content": "Observation 3: Quantization-friendly settings are diverse and contradictory across devices." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.731 + ], + "angle": 0, + "content": "In Fig. 3, we also observe that the quantization-friendly operators are different and can be contradictory on diverse devices. For instance, Swish achieves a \\(2 \\times\\) speedup on Pixel 4, but it is a quantization-unfriendly operator on CPU with a \\(0.8 \\times\\) slowdown. The reason is that quantization speedups are highly dependent on the inference engines and hardware [34, 21]. Intel VNNI supports the VPDPBWSD hardware instruction [18], which fuses three instructions into one to speedup INT8 computation. Without VNNI, INT8 hardly gains speedup on Intel CPUs. Moreover, the implementations in inference engines have to fully utilize hardware instructions for latency reduction. For example, onnxruntime does not implement a quantization kernel for Hardswish. Even on a VNNI-supported CPU, the use of Hardswish in a quantized model slows down the latency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The need for hardware specialized search space. The above observations suggest that there is no single structure (block types in a model) that is optimal for quantization on all hardware. This poses a challenge for the two-stage NAS paradigm, as the supernet training requires all models in the search space to share an isomorphic structure. To address this issue, our work proposes to design a specialized quantization-friendly search space for each hardware. Each search space is tailored to the unique characteristics of the hardware and includes an optimal structure with elastic depths, widths, and kernel sizes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.635, + 0.108 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.733, + 0.132 + ], + "angle": 0, + "content": "4.1. The Core Design Concept" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.134, + 0.892, + 0.3 + ], + "angle": 0, + "content": "In this section, we present our methodology for automatically designing a specialized quantization-friendly search space for any target hardware. Different from architecture search, where the goal is to find the single best model from the space, we aim to discover a model population that contains billions of accurate and INT8 latency-friendly architectures. We draw inspiration from the neural architecture search process and propose to use an evolutionary search algorithm to explore such a quantization-friendly model population. To achieve this, we introduce SpaceEvo, which is built on the following three techniques." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.892, + 0.36 + ], + "angle": 0, + "content": "First, we need a metric that quantifies how quantization-friendly a candidate search space is. We define a Q-T score that is efficiently measured by top-tier subnets' INT8 accuracy-latency (Sec. 4.2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.361, + 0.892, + 0.482 + ], + "angle": 0, + "content": "Second, existing evolutionary search algorithms are designed for searching a single model architecture rather than a large search space encompassing billions of architectures. we propose a novel approach that we call the \"elastic stage.\" By factorizing the search space into a sequence of elastic stages ((Sec. 4.3), we enable traditional aging evolution methods, such as the aging evolution [30], to be directly applied to search the space (Sec. 4.4)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.892, + 0.558 + ], + "angle": 0, + "content": "Third, searching a search space with a maximum Q-T score can be prohibitively costly since the corresponding supernet must be trained from scratch for accuracy evaluation. We propose a block-wise search space quantization scheme to significantly reduce the training cost(Sec. 4.5)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.569, + 0.748, + 0.585 + ], + "angle": 0, + "content": "4.2. Search Space Quality Score" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.892, + 0.8 + ], + "angle": 0, + "content": "Latency-aware space quality of Q-T score. Before space search, we need a score to quantify how quantization-friendly a search space is, which serves as the search objective. Since our ultimate goal is to search the best quantized models from the searched space, we treat a space with good quality if its top-tier subnets achieve optimal quantized accuracy under latency constraints \\( T \\). Due to the fact that real-world applications usually have different deployment requirements, we use multiple INT8 latency constraints to measure a space's quality. For search space \\( \\mathcal{A} \\) and a set of quantized latency constraints \\( T_{1,\\dots,n} \\), we treat every constraint equally important, and define Q-T score as the sum of each constraint: \\( \\mathcal{Q}(\\mathcal{A},T_1,\\dots,n) = \\mathcal{Q}(\\mathcal{A},T_1) + \\mathcal{Q}(\\mathcal{A},T_2) + \\dots \\), \\( \\mathcal{Q}(\\mathcal{A},T_n) \\), \\( \\mathcal{Q}(\\mathcal{A},T_i) \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.811, + 0.891, + 0.829 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Q} (\\mathcal {A}, T _ {i}) = \\mathbb {E} _ {\\alpha \\in \\mathcal {A}, L A T (\\alpha) \\leq T _ {i}} \\left[ A c c _ {i n t 8} (\\alpha) \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\alpha\\) denotes a top-tier (best searched) subnet in \\(\\mathcal{A}\\) and \\(Acc_{int8}(\\alpha)\\) is its top-1 quantized accuracy evaluated on ImageNet validation set, \\(LAT(\\alpha)\\) predicts the quantized latency on target device." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5822" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.3, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.198, + 0.286, + 0.212 + ], + "angle": 0, + "content": "(a) space search overview" + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.089, + 0.895, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.199, + 0.696, + 0.212 + ], + "angle": 0, + "content": "(b) stage-wise hyperspace" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.218, + 0.893, + 0.247 + ], + "angle": 0, + "content": "Figure 4. (a) We simplify space search into model search process; (b) Illustration of our hyperspace. A sampled search space is encoded by a sequential elastic stages. Contents in blue are searched: an elastic stage can search its block type and channel number list." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.272, + 0.469, + 0.392 + ], + "angle": 0, + "content": "However, it's non-trivial to obtain the top-tier subnets from a candidate search, as if often involves an expensive full architecture search process. We adopt a zero-cost policy. Specifically, we randomly sample 5k subnets and select top 20 that under the latency constraints as the top-tier models to approximate the expectation term. The top 20 subnets are rapidly identified through the use of an accuracy lookup-table and a quantized latency predictor (Sec. 4.5)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.4, + 0.416, + 0.415 + ], + "angle": 0, + "content": "4.3. Elastic Stage and Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.469, + 0.508 + ], + "angle": 0, + "content": "We observe that existing two-stage NAS adopt a chain-structured search space [4, 41, 35], which can be factorized as a sequence of STEM, HEAD and \\(N\\) searchable stages. Each stage defines a range of configurations \\(c\\) (e.g., kernel size, channel width, depth) for a specific block type \\(b\\), and allows NAS to find the optimal architecture settings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.469, + 0.569 + ], + "angle": 0, + "content": "Elastic stage. Without loss of generality, we define a stage structure in a search space as elastic stage \\( E_{b,c} \\), which has elastic configurations \\( c \\) for a fixed block \\( b \\). Suppose a search space \\( \\mathcal{A} \\) has \\( N \\) stages, it can be modularized as:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.574, + 0.469, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} = S T E M \\circ E _ {b, c} ^ {1}, \\dots \\circ E _ {b, c} ^ {N} \\circ H E A D \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.469, + 0.689 + ], + "angle": 0, + "content": "For instance, the search space in OQAT [32] and BatchQuant [1] can be factorized as 6 elastic stages and STEM (first Conv) and a classification head. Each elastic stage represents a set of configuration choices for the MBv3 block. Through the definition of elastic stage, we can simply use Eq. (2) to denote the contents of a model population." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.691, + 0.469, + 0.782 + ], + "angle": 0, + "content": "Problem definition. Operator type \\( b \\) and configuration \\( c \\) are two crucial objectives when searching quantization-friendly search space. Through the definition of elastic stage, the task of space search can be simplified to find a search space with optimal elastic stages, which has a similar goal with neural architecture search. We formulate our problem as:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.8, + 0.488, + 0.839 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} \\left(E _ {b, c} ^ {1} \\circ E _ {b, c} ^ {2} \\circ \\dots \\circ E _ {b, c} ^ {N}\\right) ^ {*} = \\underset {E _ {b, c} ^ {i} \\in \\mathcal {H} ^ {i}} {\\arg \\max } \\mathcal {Q} \\left(\\mathcal {A} \\left(E _ {b, c} ^ {1} \\circ E _ {b, c} ^ {2} \\circ \\dots \\circ E _ {b, c} ^ {N}\\right), T\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}(\\cdot)\\) denotes the search space, and \\(E_{b,c}^{i}\\) is the \\(i^{th}\\) elastic stage of \\(\\mathcal{A}(\\cdot)\\). \\(\\mathcal{H}\\) denotes the hyperspace which covers all possible search spaces. \\(\\mathcal{Q}\\) is the Q-T score for measuring a search space's quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.272, + 0.892, + 0.44 + ], + "angle": 0, + "content": "As illustrated in Fig. 4(a), given the constraints \\( T \\) (i.e., a set of targeted quantized latency), SpaceEvo aims to find the optimal elastic stages \\( (E_{b,c}^{1} \\circ E_{b,c}^{2} \\circ \\dots \\circ E_{b,c}^{N})^{*} \\) from the \\( 1^{st} \\) to \\( N^{th} \\) stage for \\( \\mathcal{A}^* \\) that has the maximum Q-T score: the top-tier quantized models can achieve best accuracy under the constraint \\( T \\). The constraint \\( T \\) can be the quantized latency under any bits. In this work, we focus on the latency of INT8 quantized models, because INT8 quantization is widely supported by real-world edge devices and can achieve real speedup. Other lower bits typically require specific hardware (e.g., FPGA) to get the latency." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.451, + 0.747, + 0.467 + ], + "angle": 0, + "content": "4.4. Searching the Search Space" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.469, + 0.892, + 0.499 + ], + "angle": 0, + "content": "We now describe our evolutionary search algorithm that solves the problem in Eq. (3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.892, + 0.605 + ], + "angle": 0, + "content": "Hyperspace design. Analogous to NAS, hyperspace \\(\\mathcal{H}\\) defines which search space a search algorithm might discover. Defining a hyperspace to cover many candidate search spaces for space search is a second-order problem for NAS, which can easily introduce high complexity. Fortunately, we can easily construct a large hyperspace through search space modularization in Eq. (2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.892, + 0.666 + ], + "angle": 0, + "content": "We construct a large hyperspace in Fig. 4, in which a search space can be encoded by \\( N = 6 \\) sequential elastic stages along with STEM and HEAD. We search the following two dimensions for an elastic stage:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.677, + 0.892, + 0.783 + ], + "angle": 0, + "content": "- Block (operator) type \\( b \\): MBv1 [15], MBv2 [31], MBv3 [14], residual bottleneck [13], residual bottleneck with SE, FusedMB [33] and FusedMB with SE. Conv is the major operator in residual bottleneck and FusedMB, thus they are quantization-friendly blocks; the efficiency of MB blocks relies on the device. For example, DWConv and SE are less quantization-efficient on Intel CPU." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.795, + 0.892, + 0.9 + ], + "angle": 0, + "content": "- Output channel width list \\( \\text{cout} \\). In Sec. 3, we observe that quantized models can better utilize hardware under a larger channel number setting. However, directly increasing the channel numbers will also lead to longer latency. Therefore, we search the optimal stage-wise channel width list \\( \\text{cout} \\): \\( \\{w_{\\min}^*, \\dots, w_{\\max}^*\\} \\), which provides better channel width choices for final INT8 model search." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "5823" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.091, + 0.47, + 0.153 + ], + "angle": 0, + "content": "Specifically, as described in Fig. 4(b), we define a wide range of \\([w_{min}, w_{max}]\\) by enlarging the channel widths in existing search spaces, and allow each elastic stage to choose a subset of \\(cout\\) from \\([w_{min}, w_{max}]\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.157, + 0.47, + 0.367 + ], + "angle": 0, + "content": "Besides channel widths, other configuration dimensions (e.g., kernel size) also impact a model's quantized latency. However, searching all dimensions leads to a large amount of choices in one stage, which exponentially enlarges the hyperspace size. Fortunately, other dimensions usually have a small space (e.g., kernel size selects from \\(\\{3,5,7\\}\\)). It's easy to find the optimal value for a model in the final NAS process. Therefore, we follow existing practices to configure the choices of kernel size, depth, and expand ratios. Our final searched INT8 model architectures SEQnet suggest that the optimal quantization-friendly kernel sizes and expand ratios are chosen. For example, kernel size of \\(3\\times 3\\) brings more INT8 latency speedups for DWConv, and it is the dominate kernel size choice in DWConv related blocks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.47, + 0.446 + ], + "angle": 0, + "content": "Suppose that a stage has \\( m \\) choices of channel widths, there would be 7 (operator types) \\( \\times m \\) candidates for each stage. In total, for a typical search space with \\( N = 6 \\) stages, the hyperspace has \\( \\sim 10^9 \\) candidate search spaces, which is extremely large and poses challenges for efficient search." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.471, + 0.753 + ], + "angle": 0, + "content": "Evolutionary space search. The structure of hyperspace is similar to a typical model search space in NAS [12], so we can easily apply existing NAS search algorithms. Taking this advantage, we leverage aging evolution [30] to search the large hyperspace. We first randomly initialize a population of \\( P \\) search spaces, where each sampled space is encoded as \\( (E_{b,c}^{1} \\circ E_{b,c}^{2} \\circ \\ldots \\circ E_{b,c}^{N}) \\). Each individual is rapidly evaluated with Q-T score. After this, evolution improves the initial population in mutation iterations. At each iteration, we sample \\( S \\) random candidates from the population and select the one with highest score as the parent. Then we alternately mutate the parent for block type and widths to generate two children search spaces. For instance, suppose the \\( i^{th} \\) stage \\( E_{b,c}^{i} \\) is selected for mutation, we first randomly modify its block type and produce \\( E_{b^*,c}^{i} \\) for child 1, then we mutate the widths and produce \\( E_{b,c^*}^{i} \\) for child 2. We evaluate their Q-T scores and add them to current population. The oldest two are removed for next iteration. After all iterations finish, we collect all the sampled space and select the one with best score as the final search space." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.763, + 0.436, + 0.779 + ], + "angle": 0, + "content": "4.5. Efficient Search Space Quality Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We now address the efficiency challenge caused by Q-T score evaluation. The most accurate evaluation is to get accuracy by training a supernet (search space) from scratch and measure latency on target device. However, it's impractical to conduct large-scale search due to the prohibitive cost. For example, it costs more than 10 days to train a supernet on 8 V100 GPUs [41]. To reduce the cost, we build an accurate INT8 latency predictor by nn-Meter [42], then" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.088, + 0.892, + 0.101 + ], + "angle": 0, + "content": "Step1: Quantize elastic stages via BKD Step2: Build quantized accuracy LUT" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.105, + 0.692, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.699, + 0.106, + 0.892, + 0.175 + ], + "angle": 0, + "content": "
Elastic stageModel block encoding (when depth=1)NSR loss
EMBv2,[32,48]k3e3c48-56x560.39
EMBv3,[32,48]k5e6c32-40x400.37
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.699, + 0.177, + 0.886, + 0.187 + ], + "angle": 0, + "content": "k: kernel size, e: expand ratio, c: width" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.206, + 0.892, + 0.248 + ], + "angle": 0, + "content": "Figure 5. We adopt block-wise knowledge distillation to reduce search space quality evaluation cost. We add two linear transformation layers (Conv \\(1 \\times 1\\)) to match teacher's feature map widths." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.264, + 0.807, + 0.277 + ], + "angle": 0, + "content": "we introduce block-wise quantization scheme." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.278, + 0.892, + 0.414 + ], + "angle": 0, + "content": "Block-wise knowledge distillation (BKD) is firstly proposed in DNA [22] and then further improved in DONNA [26]. It originally uses block-wise representation of existing models (teacher) to supervise a corresponding student model block. This technique can provide a relative accuracy ranking of all possible models without requiring them to be trained from scratch. In our work, we extend BKD to supervise the training of all elastic stages (each contains a large amount of blocks)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.893, + 0.626 + ], + "angle": 0, + "content": "Fig. 5 illustrates the BKD process. In the first step, we use EfficientNet-B5 as the teacher, and separately train each elastic stage to mimic the behavior of corresponding teacher block by minimizing the NSR loss [26] between their output feature maps. Specifically, the \\( i^{th} \\) stage receives the output of \\( (i - 1)^{th} \\) teacher block as the input and is optimized to predict the output of \\( i^{th} \\) teacher block with NSR loss. Since an elastic stage contains many blocks with different channel widths, we add two learnable linear transformation layers at the input and output for each elastic stage to match teacher's feature map shape. Moreover, we adopt sandwich rule [41] to sample four paths to improve the training efficiency. Each elastic stage is firstly trained for 5 epochs and then performed 1 epoch LSQ+ [3] for INT8 quantization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.761 + ], + "angle": 0, + "content": "Accuracy lookup table. In the second step, we construct a INT8 accuracy lookup table to reduce the evaluation cost. Specifically, we evaluate all possible blocks in each elastic stage and record their NSR losses on the validation set in the lookup table. The quantized loss of a model is estimated by summing up the NSR loss of all its blocks by rapidly looking up each elastic stages from the table. We inverse the measured loss to approximate the quantized accuracy for Q-T score evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.762, + 0.892, + 0.807 + ], + "angle": 0, + "content": "In our work, the BKD and lookup table construction can be sped up in a parallel way and finished in 1 day, which amounts a one-time cost before aging evolution search." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.821, + 0.617, + 0.836 + ], + "angle": 0, + "content": "5. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Setup. We evaluate our method on ImageNet-1k dataset [9] and two popular edge devices. The INT8 latency constraints are \\{8, 10, 15, 20, 25\\} ms for Intel CPU, and \\{15, 20, 25, 30, 35\\} ms for Pixel4. For each device, we search 5k search" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5824" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.087, + 0.473, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.245, + 0.47, + 0.287 + ], + "angle": 0, + "content": "Figure 6. Best searched INT8 models with comparison to state-of-the-art NAS search spaces. Our searched spaces are proven to be the most quantization-friendly for the target device." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.468, + 0.331 + ], + "angle": 0, + "content": "spaces in total and return the one with highest Q-T score. The population size \\( P \\) is 500 and sample size \\( S \\) is 125." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.47, + 0.574 + ], + "angle": 0, + "content": "Once SpaceEvo discovers a quantization-friendly search space for the target device, we train a quantized-for-all supernet. We start by pretraining a full-precision supernet without quantizers on ImageNet for 360 epochs. We adopt the sandwich rule and inplace distillation in BigNAS [41]. Then, we perform quantization-aware training (QAT) on the trained supernet for 50 epochs, which follows the same training protocol (i.e., sandwich rule and inplace distillation). We use \\(\\mathrm{LSQ + }\\) as the QAT algorithm for better quantized accuracy. To derive INT8 model for deployment, we use the evolutionary search in OFA [4] to search \\(5\\mathrm{k}\\) models for various given INT8 latency constraints. We list out detailed training settings in supplementary materials. In the following, we refer to the two searched spaces as SpaceEvo@CPU and SpaceEvo@Pixel4, the searched model families are SEQnet@cpu and SEQnet@pixel4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.582, + 0.347, + 0.597 + ], + "angle": 0, + "content": "5.1. The Effectiveness of SpaceEvo" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.468, + 0.81 + ], + "angle": 0, + "content": "Comparison with SOTA search spaces. To demonstrate the high-performance of our searched spaces, we compare with prior art manually-designed search spaces including: (1) MobileNetV3 search space that is adopted in two-stage quantization NAS OQAT [32] and BatchQuant [1]; (2) ProxylessNAS and AttentiveNAS search spaces that achieve superior performance on mobile devices; and (3) ResNet50 search space proposed by OFA that is a handcraft quantization-friendly space on our two devices. For fair comparison, we use one supernet training and QAT receipt for all search spaces. We conduct evolutionary search to compare the best-searched models from each search space. We use the random seed of 0. For all experiments, search space is the only difference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Fig. 6 compares the best searched INT8 models from different search spaces. SpaceEvo@CPU and SpaceEvo@Pixel4 consistently deliver superior quantized models than state-of-the-art search spaces. Under the same latency, the best quantized models from SpaceEvo@cpu significantly surpass the existing state-of-the-art search" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.087, + 0.852, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.235, + 0.892, + 0.263 + ], + "angle": 0, + "content": "Figure 7. Search space design under diverse INT8 latency constraints. SpaceEvo (6-25 ms) delivers superior tiny INT8 models." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.274, + 0.892, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.34, + 0.862, + 0.353 + ], + "angle": 0, + "content": "Figure 8. Search cost measured on 8 Nvidia V100 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.892, + 0.429 + ], + "angle": 0, + "content": "spaces with \\(+0.7\\%\\) to \\(+3.8\\%\\) (\\(+0.4\\%\\) to \\(+3.2\\%\\) on Pixel4) higher accuracy. Moreover, our search space is the only one that is able to deliver superior quantized models under both extremely low (only \\(\\sim 5\\mathrm{ms}\\)) and large latency constraints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.893, + 0.566 + ], + "angle": 0, + "content": "SpaceEvo under diverse latency constraints. We extensively study the effectiveness of SpaceEvo under different latency constraints. Specifically, we perform space search under two tight constraints of \\(\\{10, 15, 20, 25, 30\\}\\mathrm{ms}\\) and \\(\\{6, 10, 15, 20, 25\\}\\mathrm{ms}\\) on Pixel4. The results are shown in Fig. 7. Our proposed method can handle the diverse latency requirements and produce high-quality spaces. As expected, the searched spaces under \\(10 - 30\\mathrm{ms}\\) and \\(6 - 25\\mathrm{ms}\\) have much more low-latency quantized models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.673 + ], + "angle": 0, + "content": "To further verify the effectiveness of these low-latency models, we compare with existing SOTA tiny models. Significantly, even under the extremely low latency constraints of 6-25 ms, our searched space delivers very competitive tiny quantized models. Compared to the smallest model ShuffleNetV2x0.5, we can achieve \\(+10.1\\%\\) higher accuracy under the same latency of 4.3 ms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Search cost. As depicted in Fig. 8, our algorithm, SpaceEvo, is designed to be lightweight and suitable for real-world usage, requiring only 25 GPU hours to search a space of 5000 iterations. This remarkable speed is mainly due to our block-wise search space quantization scheme, which significantly reduces the cost of search space quality evaluation. In comparison, Fig. 8 demonstrates that training each search space from scratch without this scheme would consume an impractical 1200k GPU hours." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.822, + 0.884, + 0.837 + ], + "angle": 0, + "content": "5.2. The Effectiveness of Discovered INT8 Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we demonstrate that our searched spaces deliver state-of-the-art quantized models. We compare with two strong baselines: (1) prior art manually-designed and NAS-searched models; and (2) quantization-aware NAS. For" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5825" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.088, + 0.47, + 0.114 + ], + "angle": 0, + "content": "Table 1. ImageNet results compared with SOTA quantized models on two devices. *: latency compared to FP32 inference." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.115, + 0.468, + 0.685 + ], + "angle": 0, + "content": "
(a) Results on the Intel VNNI CPU with onnxruntime
ModelAcc% INT8CPU latencyFP32 FLOPs
INT8speedup*
MobileNetV3Small66.34.4 ms1.1×67.456M
SEQnet@cpu-A074.74.4 ms2.0×74.8163M
MobileNetV271.47.3 ms2.2×72.0300M
ProxylessNAS-R74.68.8 ms1.8×74.6320M
OQAT-8bit74.89.8 ms1.8×75.2214M
MobileNetV3Large74.510.3 ms1.5×75.2219M
OFA (#25)75.611.2 ms1.5×76.4230M
SEQnet@cpu-A177.48.8 ms2.4×77.5358M
APQ-8bit73.615.0 ms1.5×73.6297M
AttentiveNAS-A076.115.1 ms1.4×77.3203M
OQAT-8bit76.314.9 ms1.7×76.7316M
EfficientNet-B076.718.1 ms1.6×77.3390M
SEQnet@cpu-A278.514.1 ms2.4×78.8638M
APQ-8bit74.920.0 ms1.5×75.0393M
OQAT-8bit76.919.5 ms1.6×77.3405M
AttentiveNAS-A177.222.4 ms1.4×78.4279M
AttentiveNAS-A277.522.5 ms1.3×78.8317M
SEQnet@cpu-A379.518.9 ms2.6×79.6981M
FBNetV2-L175.825.0 ms1.2×77.2325M
FBNetV3-A78.227.7 ms1.3×79.1357M
SEQnet@cpu-A480.024.4 ms2.4×80.11267M
(b) Results on the Google Pixel 4 with TFLite
ModelAcc% INT8Pixel4 Latency|Acc% FP32FLOPs
INT8speedup*
MobileNetV3Small66.36.4 ms1.3×67.456M
SEQnet@pixel4-A073.65.9 ms2.1×73.7107M
MobileNetV271.416.5 ms1.9×72.0300M
ProxylessNAS-R74.618.4 ms1.8×74.6320M
MobileNetV3Large74.515.7 ms1.5×75.2219M
APQ-8bit74.614.9 ms2.0×74.4340M
OFA (#25)75.614.8 ms1.7×76.4230M
OQAT-8bit75.815.2 ms1.9×76.2287M
AttentiveNAS-A076.115.2 ms2.0×77.3203M
SEQnet@pixel4-A177.614.7 ms2.2×77.7274M
APQ-8bit75.120.0 ms1.9×75.1398M
OQAT-8bit76.520.4 ms1.8×76.8347M
AttentiveNAS-A177.221.1 ms2.0×78.4279M
AttentiveNAS-A277.522.7 ms2.0×78.8317M
SEQnet@pixel4-A278.319.4 ms2.3×78.4402M
FBNetV2-L175.826.7 ms1.5×77.2325M
OQAT-8bit77.029.9 ms1.7×77.2443M
FBNetV3-A78.230.5 ms1.5×79.1357M
SEQnet@pixel4-A379.530.8 ms2.1×79.5591M
EfficientNet-B076.736.4 ms1.7×77.3390M
SEQnet@pixel4-A479.935.5 ms2.2×80.0738M
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.47, + 0.88 + ], + "angle": 0, + "content": "baseline (1), we collect official pre-trained FP32 checkpoints and conduct LSQ+ QAT to get the quantized accuracy. The hyperparameter settings follow the original LSQ+ paper, except that we set a larger epoch of 10 to achieve better accuracy. The latency numbers are measured on our devices. For (2), we compare with strong baselines including APQ [37] and OQAT [32]. Specifically, we limit APQ to search for the fixed 8bit (INT8) models. Since OQAT has no 8bit supernet checkpoint, we follow the official source code and conduct supernet QAT for 50 epochs. The final INT8 models are searched under the same INT8 latency predictors for fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Results. Table 1 summarizes comparison results. Re" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.088, + 0.852, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.233, + 0.892, + 0.261 + ], + "angle": 0, + "content": "Figure 9. Q-T score effectiveness (Kendall's \\(\\tau\\)) on ranking search space quality. We achieve a high space ranking correlation." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.272, + 0.892, + 0.326 + ], + "angle": 0, + "content": "Table 2. Different space search methods and their best resulting quantized models on Pixel4. Baseline is a SOTA mobile-friendly AttentiveNAS space. *: the search dimension use the same settings in AttentiveNAS." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.326, + 0.89, + 0.394 + ], + "angle": 0, + "content": "
MethodOpWidthBest quantized models
10ms15ms20ms30ms36ms
Baseline--76.677.279.079.5
SpaceEvo-opsearchfix*75.076.677.878.678.8
SpaceEvo-widthfix*search75.477.478.079.179.5
SpaceEvosearchsearch75.777.678.379.579.9
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.68 + ], + "angle": 0, + "content": "markably, our searched model family, SEQnet significantly outperform SOTA efficient models and quantization-aware NAS searched models, with higher INT8 quantized accuracy, lower INT8 latency and better speedups. Without finetuning, our tiny models SEQnet@cpu-A0 and SEQnet@pixel4-A0 achieve \\(74.7\\%\\) and \\(73.6\\%\\) top1 accuracy on ImageNet, which is \\(8.4\\%\\) and \\(7.3\\%\\) higher than MobileNetV3-Small (56M FLOPs) while maintaining the same level quantized latency. For larger models, SEQnet@cpu-A4 \\((80.0\\%)\\) outperforms FBNetV3-A with \\(1.8\\%\\) higher accuracy while runs 3.3ms faster. In particular, to achieve the same level accuracy (i.e., around \\(77.2\\%\\)), AttentiveNAS-A1 has 22.4ms latency while SEQnet@cpu-A1 \\((77.4\\%)\\) only needs 8.8 ms \\((2.6\\times\\) faster). More importantly, our searched models can better utilize the INT8 hardware optimizations: the latency speedups compared to full-precision inference are all larger than \\(2\\times\\), and this leaves room to search large-size models with higher accuracy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.699, + 0.655, + 0.715 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Q-T score effectiveness. Q-T score is crucial as it guides the space evolution process. To evaluate its effectiveness, we randomly sample 30 search spaces, and measure the rank correlation (Kendall's \\(\\tau\\)) between their Q-T score and their actual Pareto-frontier models' accuracies. Specifically, we use Intel CPU as the test device and set a same latency constraints of \\(\\{8, 10, 15, 20, 25\\}\\) ms. For each sampled space, we train it from scratch for 50 epochs, and conduct evolutionary search to get the Pareto-frontier models' accuracies. As shown in Fig. 9, the Kendall's \\(\\tau\\) between the Q-T score and the actual Pareto-frontier models' accuracies is 0.8, which indicates a very high rank correlation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5826" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.363 + ], + "angle": 0, + "content": "Ablation study on two search dimensions. In Sec. 3, we conclude that operator type and configuration are two key factors impacting INT8 latency, which serves as the two search objectives of SpaceEvo. To verify the effectiveness, we create two strong baselines based on the SOTA edge-regime AttentiveNAS search space: (1) SpaceEvo-op: we fix each elastic stage's width to AttentiveNAS space, then allow each elastic stage to search for the optimal operator; and (2) SpaceEvo-width: we fix all elastic stages' block types to AttentiveNAS space, then search for the optimal width. Table 2 reports the space comparison between different search methods on the Pixel4. By searching both operator type and width, SpaceEvo finds the optimal search space where its best searched quantized models achieve the highest accuracy under all latency constraints. Moreover, even searching for one dimension, SpaceEvo-op and SpaceEvo-width outperform the manually-designed AttentiveNAS space under small latency constraints." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.364, + 0.473, + 0.576 + ], + "angle": 0, + "content": "Search space design implications. We now summarize our learned experience and implications for designing quantization-friendly search spaces. We notice that the searched spaces show different preferences when targeting different devices: (i) All stages should use much wider channel widths compared to existing manually-designed spaces on the cpu, while only early stages prefer wider channels on Pixel 4. (ii) Since SE and Swish are INT8 latency-friendly on mobile phones, so our auto-generated search spaces for Pixel4 have many MBv3 stages. On Intel CPU, INT8 quantization slows down SE, Hardwish, and Swish, making FusedMB and MBv2 the priority for search spaces, with only the last two stages using MBv3. The details are provided in supplementary." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.589, + 0.196, + 0.604 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.608, + 0.47, + 0.73 + ], + "angle": 0, + "content": "In this paper, we introduced SpaceEvo, the first to automatically design a quantization-friendly space for target device, which delivers superior INT8 quantized models with SOTA efficiency on real-world edge devices. Extensive experiments on two popular devices demonstrate its effectiveness compared to prior art manual-designed search spaces. We plan to apply SpaceEvo to other hardware efficiency such as energy-efficient search space design in the future." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.743, + 0.174, + 0.758 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.767, + 0.468, + 0.809 + ], + "angle": 0, + "content": "[1] Haoping Bai, Meng Cao, Ping Huang, and Jiulong Shan. Batchquant: Quantized-for-all architecture search with robust quantizer. In NeurIPS, 2021. 1, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.811, + 0.468, + 0.865 + ], + "angle": 0, + "content": "[2] Ron Banner, Yury Nahshan, and Daniel Soudry. Post training 4-bit quantization of convolutional networks for rapid-deployment. In Advances in Neural Information Processing Systems, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.867, + 0.468, + 0.895 + ], + "angle": 0, + "content": "[3] Yash Bhalgat, Jinwon Lee, Markus Nagel, Tijmen Blankevoort, and Nojun Kwak. Lsq+: Improving low-bit" + }, + { + "type": "list", + "bbox": [ + 0.085, + 0.767, + 0.468, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "quantization through learnable offsets and better initialization. In CVPRW, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[4] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. In ICLR, 2020. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.165, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[5] Han Cai, Ligeng Zhu, and Song Han. ProxylessNAS: Direct neural architecture search on target task and hardware. In International Conference on Learning Representations, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.221, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[6] Zhaowei Cai and Nuno Vasconcelos. Rethinking differentiable search for mixed-precision neural networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2346–2355. Computer Vision Foundation / IEEE, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.291, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[7] Minghao Chen, Kan Wu, Bolin Ni, Houwen Peng, Bei Liu, Jianlong Fu, Hongyang Chao, and Haibin Ling. Searching the search space of vision transformers. In NeurIPS, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.347, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[8] Yuanzheng Ci, Chen Lin, Ming Sun, Boyu Chen, Hongwen Zhang, and Wanli Ouyang. Evolving search space for neural architecture search. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6639-6649, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.418, + 0.892, + 0.501 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248-255. IEEE Computer Society, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.503, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[10] Steven K. Esser, Jeffrey L. McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S. Modha. Learned step size quantization. In ICLR, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.73, + 0.559 + ], + "angle": 0, + "content": "[11] Google. Tensorflow lite, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.561, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[12] Zichao Guo, Xiangyu Zhang, Haoyuan Mu, Wen Heng, Zechun Liu, Yichen Wei, and Jian Sun. Single path one-shot neural architecture search with uniform sampling. In ECCV, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.617, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.674, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[14] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, Quoc V. Le, and Hartwig Adam. Searching for mobilenetv3. In ICCV, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.731, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[15] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. CoRR, abs/1704.04861, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.801, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[16] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.844, + 0.892, + 0.898 + ], + "angle": 0, + "content": "[17] Yiming Hu, Yuding Liang, Zichao Guo, Ruosi Wan, Xiangyu Zhang, Yichen Wei, Qingyi Gu, and Jian Sun. Angle-based search space shrinking for neural architecture search. In ECCV. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5827" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.119 + ], + "angle": 0, + "content": "[18] Intel. Intel 64 and ia-32 architectures software developer's manual, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.203 + ], + "angle": 0, + "content": "[19] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.469, + 0.286 + ], + "angle": 0, + "content": "[20] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2704-2713, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.287, + 0.469, + 0.327 + ], + "angle": 0, + "content": "[21] Sumin Kim, Gunju Park, and Youngmin Yi. Performance evaluation of int8 quantized inference on mobile gpus. IEEE Access, 9:164245-164255, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.328, + 0.469, + 0.397 + ], + "angle": 0, + "content": "[22] Changlin Li, Jiefeng Peng, Liuchun Yuan, Guangrun Wang, Xiaodan Liang, Liang Lin, and Xiaojun Chang. Blockwisely supervised neural architecture search with knowledge distillation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.398, + 0.469, + 0.464 + ], + "angle": 0, + "content": "[23] Xiang Li, Chen Lin, Chuming Li, Ming Sun, Wei Wu, Junjie Yan, and Wanli Ouyang. Improving one-shot nas by suppressing the posterior fading. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.467, + 0.469, + 0.508 + ], + "angle": 0, + "content": "[24] Christos Louizos, Mathias Reisser, Tijmen Blankevoort, Efstratos Gavves, and Max Welling. Relaxed quantization for discretized neural networks. In ICLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.509, + 0.31, + 0.521 + ], + "angle": 0, + "content": "[25] Microsoft. onnxruntime, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.522, + 0.469, + 0.577 + ], + "angle": 0, + "content": "[26] Bert Moons, Parham Noorzad, Andrii Skliar, Giovanni Mariani, Dushyant Mehta, Chris Lott, and Tijmen Blankevoort. Distilling optimal neural networks: Rapid search in diverse spaces. In ICCV, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.578, + 0.469, + 0.618 + ], + "angle": 0, + "content": "[27] Markus Nagel, Mart van Baalen, Tijmen Blankevoort, and Max Welling. Data-free quantization through weight equalization and bias correction. In ICCV, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.688 + ], + "angle": 0, + "content": "[28] Asaf Noy, Niv Nayman, Tal Ridnik, Nadav Zamir, Sivan Doveh, Itamar Friedman, Raja Giryes, and Lihi Zelnik. Asap: Architecture search, anneal and prune. In Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, pages 493-503, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.73 + ], + "angle": 0, + "content": "[29] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[30] Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image classifier architecture search. In AAAI, 2019. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.469, + 0.841 + ], + "angle": 0, + "content": "[31] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.841 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.158 + ], + "angle": 0, + "content": "[32] Mingzhu Shen, Feng Liang, Ruihao Gong, Yuhang Li, Chuming Li, Chen Lin, Fengwei Yu, Junjie Yan, and Wanlin Ouyang. Once quantization-aware training: High performance extremely low-bit architecture search. In ICCV, 2021. 1, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.161, + 0.892, + 0.215 + ], + "angle": 0, + "content": "[33] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In Proceedings of the 38th International Conference on Machine Learning, pages 10096-10106. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.299 + ], + "angle": 0, + "content": "[34] Xiaohu Tang, Shihao Han, Li Lyna Zhang, Ting Cao, and Yunxin Liu. To bridge neural network design and real-world performance: A behaviour study for neural networks. In A. Smola, A. Dimakis, and I. Stoica, editors, Proceedings of Machine Learning and Systems, volume 3, pages 21-37, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.301, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[35] Dilin Wang, Meng Li, Chengyue Gong, and Vikas Chandra. Attentivenas: Improving neural architecture search via attentive sampling. In Conference on Computer Vision and Pattern Recognition, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.359, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[36] Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, and Song Han. Haq: Hardware-aware automated quantization with mixed precision. In CVPR, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.401, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[37] Tianzhe Wang, Kuan Wang, Han Cai, Ji Lin, Zhijian Liu, and Song Han. Apq: Joint search for network architecture, pruning and quantization policy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.472, + 0.892, + 0.554 + ], + "angle": 0, + "content": "[38] Xudong Wang, Li Lyna Zhang, Yang Wang, and Mao Yang. Towards efficient vision transformer inference: A first study of transformers on mobile devices. In Proceedings of the 23rd Annual International Workshop on Mobile Computing Systems and Applications, HotMobile '22, page 1-7. Association for Computing Machinery, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.556, + 0.892, + 0.61 + ], + "angle": 0, + "content": "[39] Bichen Wu, Yanghan Wang, Peizhao Zhang, Yuandong Tian, Peter Vajda, and Kurt Keutzer. Mixed precision quantization of convnets via differentiable neural architecture search. 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.612, + 0.892, + 0.653 + ], + "angle": 0, + "content": "[40] Xin Xia, Xuefeng Xiao, Xing Wang, and Min Zheng. Progressive automatic design of search space for one-shot neural architecture search. In WACV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.655, + 0.892, + 0.723 + ], + "angle": 0, + "content": "[41] Jiahui Yu, Pengchong Jin, Hanxiao Liu, Gabriel Bender, Pieter-Jan Kindermans, Mingxing Tan, Thomas Huang, Xiaodan Song, Ruoming Pang, and Quoc Le. Bignas: Scaling up neural architecture search with big single-stage models. In ECCV, 2020. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.726, + 0.892, + 0.821 + ], + "angle": 0, + "content": "[42] Li Lyna Zhang, Shihao Han, Jianyu Wei, Ningxin Zheng, Ting Cao, Yuqing Yang, and Yunxin Liu. nn-meter: Towards accurate latency prediction of deep-learning model inference on diverse edge devices. In Proceedings of the 19th Annual International Conference on Mobile Systems, Applications, and Services, page 81-93, New York, NY, USA, 2021. ACM. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.821 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5828" + } + ] +] \ No newline at end of file diff --git a/2023/SpaceEvo_ Hardware-Friendly Search Space Design for Efficient INT8 Inference/7d8d9b36-b64d-45d0-8aa5-d17e8abbee45_origin.pdf b/2023/SpaceEvo_ Hardware-Friendly Search Space Design for Efficient INT8 Inference/7d8d9b36-b64d-45d0-8aa5-d17e8abbee45_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1c4283c05221f73c5609fe135e9f707326292a64 --- /dev/null +++ b/2023/SpaceEvo_ Hardware-Friendly Search Space Design for Efficient INT8 Inference/7d8d9b36-b64d-45d0-8aa5-d17e8abbee45_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d5aad61e938b6d093ad554271c3045c6ee195bb3834e789b8d240e5a15cf917 +size 5743065 diff --git a/2023/SpaceEvo_ Hardware-Friendly Search Space Design for Efficient INT8 Inference/full.md b/2023/SpaceEvo_ Hardware-Friendly Search Space Design for Efficient INT8 Inference/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bb3ecfd656161aeab1174d6c8131e45e1b9bc23c --- /dev/null +++ b/2023/SpaceEvo_ Hardware-Friendly Search Space Design for Efficient INT8 Inference/full.md @@ -0,0 +1,304 @@ +# SpaceEvo: Hardware-Friendly Search Space Design for Efficient INT8 Inference + +Xudong Wang\*1,2§ Li Lyna Zhang\*2‡ Jiahang Xu² Quanlu Zhang² Yujing Wang³ Yuqing Yang² Ningxin Zheng² Ting Cao² Mao Yang² + +$^{1}$ Shanghai Jiao Tong University $^{2}$ Microsoft Research $^{3}$ Microsoft + +xudongwang176@gmail.com; {Izhani, jiahangxu, quzha, yujwang, yuqyang, nizhen, ticao, maoyang}@microsoft.com + +https://github.com/microsoft/Moonlit/tree/main/SpaceEvo + +# Abstract + +The combination of Neural Architecture Search (NAS) and quantization has proven successful in automatically designing low-FLOPs INT8 quantized neural networks (QNN). However, directly applying NAS to design accurate QNN models that achieve low latency on real-world devices leads to inferior performance. In this work, we identify that the poor INT8 latency is due to the quantization-unfriendly issue: the operator and configuration (e.g., channel width) choices in prior art search spaces lead to diverse quantization efficiency and can slow down the INT8 inference speed. To address this challenge, we propose SpaceEvo, an automatic method for designing a dedicated, quantization-friendly search space for each target hardware. The key idea of SpaceEvo is to automatically search hardware-preferred operators and configurations to construct the search space, guided by a metric called $Q-T$ score to quantify how quantization-friendly a candidate search space is. We further train a quantized-for-all supernet over our discovered search space, enabling the searched models to be directly deployed without extra retraining or quantization. Our discovered models, SEQnet, establish new SOTA INT8 quantized accuracy under various latency constraints, achieving up to $10.1\%$ accuracy improvement on ImageNet than prior art CNNs under the same latency. Extensive experiments on real devices show that SpaceEvo consistently outperforms manually-designed search spaces with up to $2.5\times$ faster speed while achieving the same accuracy. + +# 1. Introduction + +INT8 Quantization[27, 19, 10, 3] is a widely used technique for deploying DNNs on edge devices by reducing $4 \times$ in model size and memory cost for full-precision (FP32) + +![](images/3366466d8acddb126a6fe904cf34998e83103147b276e6c158fcdf74470fdc1a.jpg) +Figure 1. INT8 quantized model error distributions (EDF) of different NAS search spaces. We propose to search quantization-friendly search space for each hardware, which yields significantly better INT8 model populations than SOTA search spaces. + +models. However, prior art DNN models achieve only marginal speedup from INT8 quantization (in Fig. 2(a)), the still high latency after quantization making them difficult to deploy in latency-critical scenarios. Designing models that achieve high accuracy and low latency after quantization becomes the important but challenging problem. + +Neural Architecture Search (NAS) is a powerful tool for automating efficient quantized model design [36, 39, 12, 37, 6]. Recently, OQAT [32] and BatchQuant [1] achieve remarkable search efficiency and accuracy by adopting a two-stage paradigm. The first stage trains a weight-shared quantized supernet assembling all candidate architectures in the search space. This allows all the sub-networks (subnets) to simultaneously reach comparable quantized accuracy as when trained from scratch individually. The second stage uses typical search algorithms to find subnets with best quantized accuracy under different FLOPs constraints. This approach avoids the need to retrain each subnet for accuracy evaluation, greatly improving the search efficiency. + +Though promising in optimizing model FLOPs, we identify a significant challenge when directly applying two-stage NAS to low quantized latency scenarios. This is due to the quantization-unfriendly search space issue, where prior art search spaces can unexpectedly impede INT8 latency. Consequently, since INT8 quantization yields only a marginal speedup, we are forced to search for small-sized models to fulfill latency criteria, which can unfortunately restrict NAS to find better quantized models for edge devices. Then, a question naturally arises: Can we design a quantization-friendly search space, allowing NAS to discover larger and superior models that meet the low INT8 latency requirements? + +We start by conducting an in-depth study to understand the factors that determine INT8 quantized latency and how they affect search space design. Our study shows: (1) both operator type and configurations (e.g., channel width) greatly impact the INT8 latency; Improper selections can slow down the INT8 latency. For instance, Squeeze-and-Excitation (SE) [16] and Hardswish [14] are widely-used operators in current search spaces as it improves accuracy with little latency introduced. However, their INT8 inference speeds are slower than FP32 inference on Intel CPU (Fig. 3(a)), because the extra costs (e.g., data transformation between INT32 and INT8) introduced by quantization outweigh the latency reduction by INT8 computation. (2) The quantization efficiency varies across different devices, and the preferred operator types can be contradictory. + +The above study motivates us to design specialized quantization-friendly search spaces for each hardware, rather than relying on a single, large search space as seen in SPOS [12] for all hardware, which provides different operator options per layer. This is because two-stage NAS requires the search space to adhere to a specific condition for training the supernet, wherein each layer must utilize a fixed operator. Our study indicates significant variations in optimal operators across different hardware. Thus, customizing the search space for each hardware is crucial for optimal results. However, designing such specialized quantization-friendly search spaces for various edge devices presents a significant challenge, requiring expertise in both AI and hardware domains, as well as many trial and error attempts to optimize accuracy and INT8 latency for each hardware. + +In this paper, we propose SpaceEvo, a novel method for automatically designing specialized quantization-friendly search space for each hardware. The search space is comprised of hardware-preferred operators and configurations, enabling the search of larger and better models with low INT8 latency. With the discovered search space, we leverage two-stage quantization NAS to train a quantized-for-all supernet, and utilize evolution search [4] to find best quantized models under various INT8 latency constraints. Our approach addresses three key challenges: (1) What is the + +definition of a quantization-friendly search space in terms of both quantized accuracy and latency? (2) How to automatically design a search space without human expertise? (3) How to handle with the prohibitive cost caused by quality evaluation of a candidate search space? + +To address the first challenge, we propose a latency-aware $Q-T$ score to quantify the effectiveness of a candidate search space, which measures the INT8 accuracy-latency quality of top-tier subnets in a search space. The behind intuition is that the goal of NAS is to search top subnets with better accuracy-latency tradeoffs. + +Then, we introduce an evolutionary-based search algorithm that can effectively search a quantization-friendly search space with highest $Q-T$ score. Searching a search space involves discovering a collection of model population that contains billions of models, which is challenging and easily introduce complexity. To address this challenge, we propose to factorize and encode a search space into a sequence of elastic stages, which have flexible operator types and configurations. Through this design, the task of search space design is then simplified to find a search space with the optimal elastic stages, so that existing search algorithms can be easily applied. Specifically, we design a stage-wise hyperspace to include many candidate search spaces and leverage aging evolution [30] to perform random mutations of elastic stages for search space evolution. The evolution is guided by maximizing the $Q-T$ score. + +Finally, estimating the quality score (Q-T score) of a search space involves a costly training process for evaluating the accuracy of sub-networks, which presents a significant obstacle for our evolutionary algorithm. Naively adopting a two-stage NAS approach, training a supernet for each candidate search space [8, 7], is prohibitively expensive, taking of thousands GPU hours. To address this issue, we draw inspiration from block-wise knowledge distillation [26, 22] and propose a block-wise search space quantization scheme. This scheme trains each elastic stage separately and rapidly estimates a model's quantized accuracy by summing block-level loss with a quantized accuracy lookup table, as shown in Fig. 5. This significantly reduces the training and evaluation costs, while providing effective accuracy rankings among search spaces. We summarize our contributions as follows: + +- We study the INT8 quantization efficiency on real-world edge devices and find that the choices of operator types and configurations in a quantized model can significantly impact the INT8 latency, leaving a huge room for design optimization of quantized models. + +- We propose SpaceEvo to automatically design a hardware-dedicated quantization-friendly search space and leverage two-stage quantization NAS to produce superior INT8 models under various latency constraints. + +- We present three innovative techniques that enable the first-ever efficient and cost-effective evolution search to explore a search space comprising billions of models. +- Extensive experiments on two real-world edge devices and ImageNet demonstrate that our automatically designed search spaces significantly surpass manually-designed search spaces. Our discovered models, SEQnet, establish the new state-of-the-art INT8 quantized accuracy-latency tradeoffs. For instance, SEQnet@cpuA4, achieves $80.0\%$ accuracy on ImageNet, which is 3.3ms faster with $1.8\%$ higher accuracy than FBNetV3-A. Moreover, SpaceEvo delivers superior tiny models, achieving up to $10.1\%$ accuracy improvement over ShuffleNetV2x0.5 (41M FLOPs, 4.3ms). + +# 2. Related Works + +Quantization has been widely used for efficiency in deployment. Extensive efforts can be classified into posttraining quantization (PTQ) [27, 2] and quantization-aware training (QAT) [19, 24, 10, 3]. QAT generally outperforms PTQ in quantizing compact DNNs to 8bit and very low-bit (2, 3, 4bit) by finetuning the quantized weights. Despite their success, traditional quantization methods focus on minimizing accuracy loss for a given pre-trained model, but ignore the real-world inference efficiency. + +NAS for Quantization. Early works [36, 39, 12, 37, 6] formulates mixed-precision problem into NAS to search layer bit-width with a given architecture. Recently, [32, 1] train a quantized-for-all supernet to search both architecture and bit-width. The searched models can be directly evaluated with comparable accuracy to train-from-scratch. However, little attention is paid on optimizing quantized latency on real-world devices. Through searching quantization-friendly search space, our discovered quantized models can achieve both high accuracy and low latency. + +Search Space Design. Starting from [5], the manually-designed MBConv-based space becomes the dominant in most NAS works [5, 4, 41, 35]. RegNet [29] is the first to present standard guidelines to optimize a search space by each dimension. Recently, [17, 28, 23, 40, 8, 7] propose to shrink to a better compact search space by either pruning unimportant operators or configurations. However, these works focus on optimizing the accuracy and little attention is paid on quantization-friendly search space design. Our work is the first lightweight solution towards this direction. + +# 3. On-device Quantization Efficiency Analysis + +To understand what factors lead to quantization-unfriendly issue, we conduct a comprehensive study on two widely-used edge devices equipped with high-performance inference engine: an Intel CPU device supported with VNNI instructions and onnxruntime [25] (Intel CPU) and + +![](images/1d5f9faa311863c173d7770af5f512e6f891b78bcc75e182cc871cd9dcc78b94.jpg) +(a) FLOPs and INT8 latency on Intel CPU +Figure 2. INT8 latency and speedups (annotated) for SOTA models. FLOPs and FP32 latency are not good indicators of INT8 latency; Compact models have very limited INT8 speedup $(\sim 1.5\times)$ . + +![](images/ef11762f49ee40899870cb25b2f045762028736767160477801850f6ba33f6e5.jpg) +(b) FP32 and INT8 latency on Pixel 4 + +a Pixel 4 phone CPU with TFLite 2.7 [11] (Pixel 4). Note that we follow existing practices [42, 34, 38] to measure the latency. We reveal key observations as follows: + +Observation 1: FP32 latency and FLOPs are not good indicators of INT8 latency. + +To deploy on edge devices, a common belief is that a compact model with low FLOPs or FP32 latency is preferred than a larger model. However, Fig. 2 shows that neither of them is a good indicator of INT8 latency. In Fig. 2(b), a very large model (ResNet18) can be even faster than a compact model (EfficientNet-B0) after quantization. Moreover, the recent SOTA compact models searched by OFA [4] and AttentiveNAS [35] all have marginal INT8 speedups, suggesting that optimizing FLOPs and FP32 latency can not lead to lower INT8 latency. + +Observation 2: The choices of operators' types and configurations greatly impact the INT8 latency. + +The prior art search spaces adopted in recent two-stage NAS works are MobileNetV2 or MobileNetV3 based chain-structures, where each search space comprises a sequence of blocks (stages). The block type is fixed to the MBCnv and is allowed to search from a handcraft range of hyperparameter configurations including kernel size, expansion ratio, channel width and depth, which are designed with human wisdom. For instance, many works [5, 35] observe that edge-regime CNNs prefer deeper depths and narrower channels, and manually set small channel numbers but large depths in the search space. + +However, we find that many block type and configuration choices in current search spaces unexpectedly slow down the INT8 latency. We first study the operator type impact in Fig. 3(a). SE and Hardswish are lightweight operators in edge-regime search spaces, but their INT8 inference becomes slower on Intel CPU. Compared to Conv, DWConv can greatly reduce the FLOPs, but it benefits less from INT8 quantization. The root cause is that quantization introduces extra cost, such as (1) data transformation between INT32 and INT8 [34], and (2) additional computation caused by scaling fators and zero points [20]. If the operator has low data-reuse-rate, such as the activation functions (Hswish), + +
OperatorIntel CPUPixel 4
Conv2.6x2.5x
DWConv1.2x2.0x
SE0.7x1.4x
Hardswish0.7x0.5x
Swish0.8x2.1x
+ +![](images/09f114e9a3ff2e742520a7fb1111b0ddeec7ceaac879c5098f9cdd8e005736f5.jpg) +(a) Avg. INT8 latency speedup (relative to FP32) of major operators in MobileNetV3. +(b) Conv1x1 speedups +Figure 3. (a) The choice of operator type leads to significantly different quantized speedup. (b) Conv1x1 speedups with various channel numbers. Config: $HW = 28$ , $C_{out} = 4 \times C_{in}$ (expand=4). + +the extra cost may outweigh the latency reduction by the low-bit computation. For high data-reuse operators (Conv), this cost is amortized and thus achieve large speedup [34]. + +Besides the operator type, the configuration choices also determine the quantization efficiency. Fig. 3(b) shows the speedups of $\mathrm{Conv}1\times 1$ under various channel widths. Results show that small channel widths in OFA search space cannot benefit well from quantization. This is because the additional quantization cost has a large impact when the channel width is small, limiting the latency acceleration. In contrast, SpaceEvo can automatically design a search space with larger channel widths for better efficiency. + +Observation 3: Quantization-friendly settings are diverse and contradictory across devices. + +In Fig. 3, we also observe that the quantization-friendly operators are different and can be contradictory on diverse devices. For instance, Swish achieves a $2 \times$ speedup on Pixel 4, but it is a quantization-unfriendly operator on CPU with a $0.8 \times$ slowdown. The reason is that quantization speedups are highly dependent on the inference engines and hardware [34, 21]. Intel VNNI supports the VPDPBWSD hardware instruction [18], which fuses three instructions into one to speedup INT8 computation. Without VNNI, INT8 hardly gains speedup on Intel CPUs. Moreover, the implementations in inference engines have to fully utilize hardware instructions for latency reduction. For example, onnxruntime does not implement a quantization kernel for Hardswish. Even on a VNNI-supported CPU, the use of Hardswish in a quantized model slows down the latency. + +The need for hardware specialized search space. The above observations suggest that there is no single structure (block types in a model) that is optimal for quantization on all hardware. This poses a challenge for the two-stage NAS paradigm, as the supernet training requires all models in the search space to share an isomorphic structure. To address this issue, our work proposes to design a specialized quantization-friendly search space for each hardware. Each search space is tailored to the unique characteristics of the hardware and includes an optimal structure with elastic depths, widths, and kernel sizes. + +# 4. Methodology + +# 4.1. The Core Design Concept + +In this section, we present our methodology for automatically designing a specialized quantization-friendly search space for any target hardware. Different from architecture search, where the goal is to find the single best model from the space, we aim to discover a model population that contains billions of accurate and INT8 latency-friendly architectures. We draw inspiration from the neural architecture search process and propose to use an evolutionary search algorithm to explore such a quantization-friendly model population. To achieve this, we introduce SpaceEvo, which is built on the following three techniques. + +First, we need a metric that quantifies how quantization-friendly a candidate search space is. We define a Q-T score that is efficiently measured by top-tier subnets' INT8 accuracy-latency (Sec. 4.2). + +Second, existing evolutionary search algorithms are designed for searching a single model architecture rather than a large search space encompassing billions of architectures. we propose a novel approach that we call the "elastic stage." By factorizing the search space into a sequence of elastic stages ((Sec. 4.3), we enable traditional aging evolution methods, such as the aging evolution [30], to be directly applied to search the space (Sec. 4.4). + +Third, searching a search space with a maximum Q-T score can be prohibitively costly since the corresponding supernet must be trained from scratch for accuracy evaluation. We propose a block-wise search space quantization scheme to significantly reduce the training cost(Sec. 4.5). + +# 4.2. Search Space Quality Score + +Latency-aware space quality of Q-T score. Before space search, we need a score to quantify how quantization-friendly a search space is, which serves as the search objective. Since our ultimate goal is to search the best quantized models from the searched space, we treat a space with good quality if its top-tier subnets achieve optimal quantized accuracy under latency constraints $T$ . Due to the fact that real-world applications usually have different deployment requirements, we use multiple INT8 latency constraints to measure a space's quality. For search space $\mathcal{A}$ and a set of quantized latency constraints $T_{1,\dots,n}$ , we treat every constraint equally important, and define Q-T score as the sum of each constraint: $\mathcal{Q}(\mathcal{A},T_1,\dots,n) = \mathcal{Q}(\mathcal{A},T_1) + \mathcal{Q}(\mathcal{A},T_2) + \dots$ , $\mathcal{Q}(\mathcal{A},T_n)$ , $\mathcal{Q}(\mathcal{A},T_i)$ is defined as: + +$$ +\mathcal {Q} (\mathcal {A}, T _ {i}) = \mathbb {E} _ {\alpha \in \mathcal {A}, L A T (\alpha) \leq T _ {i}} \left[ A c c _ {i n t 8} (\alpha) \right] \tag {1} +$$ + +where $\alpha$ denotes a top-tier (best searched) subnet in $\mathcal{A}$ and $Acc_{int8}(\alpha)$ is its top-1 quantized accuracy evaluated on ImageNet validation set, $LAT(\alpha)$ predicts the quantized latency on target device. + +![](images/fe5f2339213858253dd54127a61124c59344e72f0e77b9f0955cff1d0685ac57.jpg) +(a) space search overview + +![](images/a041b75c3603dccee23bc368b1dc4b48bb885ec8431fa871d738b5bdd8db5faa.jpg) +(b) stage-wise hyperspace +Figure 4. (a) We simplify space search into model search process; (b) Illustration of our hyperspace. A sampled search space is encoded by a sequential elastic stages. Contents in blue are searched: an elastic stage can search its block type and channel number list. + +However, it's non-trivial to obtain the top-tier subnets from a candidate search, as if often involves an expensive full architecture search process. We adopt a zero-cost policy. Specifically, we randomly sample 5k subnets and select top 20 that under the latency constraints as the top-tier models to approximate the expectation term. The top 20 subnets are rapidly identified through the use of an accuracy lookup-table and a quantized latency predictor (Sec. 4.5). + +# 4.3. Elastic Stage and Problem Formulation + +We observe that existing two-stage NAS adopt a chain-structured search space [4, 41, 35], which can be factorized as a sequence of STEM, HEAD and $N$ searchable stages. Each stage defines a range of configurations $c$ (e.g., kernel size, channel width, depth) for a specific block type $b$ , and allows NAS to find the optimal architecture settings. + +Elastic stage. Without loss of generality, we define a stage structure in a search space as elastic stage $E_{b,c}$ , which has elastic configurations $c$ for a fixed block $b$ . Suppose a search space $\mathcal{A}$ has $N$ stages, it can be modularized as: + +$$ +\mathcal {A} = S T E M \circ E _ {b, c} ^ {1}, \dots \circ E _ {b, c} ^ {N} \circ H E A D \tag {2} +$$ + +For instance, the search space in OQAT [32] and BatchQuant [1] can be factorized as 6 elastic stages and STEM (first Conv) and a classification head. Each elastic stage represents a set of configuration choices for the MBv3 block. Through the definition of elastic stage, we can simply use Eq. (2) to denote the contents of a model population. + +Problem definition. Operator type $b$ and configuration $c$ are two crucial objectives when searching quantization-friendly search space. Through the definition of elastic stage, the task of space search can be simplified to find a search space with optimal elastic stages, which has a similar goal with neural architecture search. We formulate our problem as: + +$$ +\mathcal {A} \left(E _ {b, c} ^ {1} \circ E _ {b, c} ^ {2} \circ \dots \circ E _ {b, c} ^ {N}\right) ^ {*} = \underset {E _ {b, c} ^ {i} \in \mathcal {H} ^ {i}} {\arg \max } \mathcal {Q} \left(\mathcal {A} \left(E _ {b, c} ^ {1} \circ E _ {b, c} ^ {2} \circ \dots \circ E _ {b, c} ^ {N}\right), T\right) \tag {3} +$$ + +where $\mathcal{A}(\cdot)$ denotes the search space, and $E_{b,c}^{i}$ is the $i^{th}$ elastic stage of $\mathcal{A}(\cdot)$ . $\mathcal{H}$ denotes the hyperspace which covers all possible search spaces. $\mathcal{Q}$ is the Q-T score for measuring a search space's quality. + +As illustrated in Fig. 4(a), given the constraints $T$ (i.e., a set of targeted quantized latency), SpaceEvo aims to find the optimal elastic stages $(E_{b,c}^{1} \circ E_{b,c}^{2} \circ \dots \circ E_{b,c}^{N})^{*}$ from the $1^{st}$ to $N^{th}$ stage for $\mathcal{A}^*$ that has the maximum Q-T score: the top-tier quantized models can achieve best accuracy under the constraint $T$ . The constraint $T$ can be the quantized latency under any bits. In this work, we focus on the latency of INT8 quantized models, because INT8 quantization is widely supported by real-world edge devices and can achieve real speedup. Other lower bits typically require specific hardware (e.g., FPGA) to get the latency. + +# 4.4. Searching the Search Space + +We now describe our evolutionary search algorithm that solves the problem in Eq. (3). + +Hyperspace design. Analogous to NAS, hyperspace $\mathcal{H}$ defines which search space a search algorithm might discover. Defining a hyperspace to cover many candidate search spaces for space search is a second-order problem for NAS, which can easily introduce high complexity. Fortunately, we can easily construct a large hyperspace through search space modularization in Eq. (2). + +We construct a large hyperspace in Fig. 4, in which a search space can be encoded by $N = 6$ sequential elastic stages along with STEM and HEAD. We search the following two dimensions for an elastic stage: + +- Block (operator) type $b$ : MBv1 [15], MBv2 [31], MBv3 [14], residual bottleneck [13], residual bottleneck with SE, FusedMB [33] and FusedMB with SE. Conv is the major operator in residual bottleneck and FusedMB, thus they are quantization-friendly blocks; the efficiency of MB blocks relies on the device. For example, DWConv and SE are less quantization-efficient on Intel CPU. + +- Output channel width list $\text{cout}$ . In Sec. 3, we observe that quantized models can better utilize hardware under a larger channel number setting. However, directly increasing the channel numbers will also lead to longer latency. Therefore, we search the optimal stage-wise channel width list $\text{cout}$ : $\{w_{\min}^*, \dots, w_{\max}^*\}$ , which provides better channel width choices for final INT8 model search. + +Specifically, as described in Fig. 4(b), we define a wide range of $[w_{min}, w_{max}]$ by enlarging the channel widths in existing search spaces, and allow each elastic stage to choose a subset of $cout$ from $[w_{min}, w_{max}]$ . + +Besides channel widths, other configuration dimensions (e.g., kernel size) also impact a model's quantized latency. However, searching all dimensions leads to a large amount of choices in one stage, which exponentially enlarges the hyperspace size. Fortunately, other dimensions usually have a small space (e.g., kernel size selects from $\{3,5,7\}$ ). It's easy to find the optimal value for a model in the final NAS process. Therefore, we follow existing practices to configure the choices of kernel size, depth, and expand ratios. Our final searched INT8 model architectures SEQnet suggest that the optimal quantization-friendly kernel sizes and expand ratios are chosen. For example, kernel size of $3\times 3$ brings more INT8 latency speedups for DWConv, and it is the dominate kernel size choice in DWConv related blocks. + +Suppose that a stage has $m$ choices of channel widths, there would be 7 (operator types) $\times m$ candidates for each stage. In total, for a typical search space with $N = 6$ stages, the hyperspace has $\sim 10^9$ candidate search spaces, which is extremely large and poses challenges for efficient search. + +Evolutionary space search. The structure of hyperspace is similar to a typical model search space in NAS [12], so we can easily apply existing NAS search algorithms. Taking this advantage, we leverage aging evolution [30] to search the large hyperspace. We first randomly initialize a population of $P$ search spaces, where each sampled space is encoded as $(E_{b,c}^{1} \circ E_{b,c}^{2} \circ \ldots \circ E_{b,c}^{N})$ . Each individual is rapidly evaluated with Q-T score. After this, evolution improves the initial population in mutation iterations. At each iteration, we sample $S$ random candidates from the population and select the one with highest score as the parent. Then we alternately mutate the parent for block type and widths to generate two children search spaces. For instance, suppose the $i^{th}$ stage $E_{b,c}^{i}$ is selected for mutation, we first randomly modify its block type and produce $E_{b^*,c}^{i}$ for child 1, then we mutate the widths and produce $E_{b,c^*}^{i}$ for child 2. We evaluate their Q-T scores and add them to current population. The oldest two are removed for next iteration. After all iterations finish, we collect all the sampled space and select the one with best score as the final search space. + +# 4.5. Efficient Search Space Quality Evaluation + +We now address the efficiency challenge caused by Q-T score evaluation. The most accurate evaluation is to get accuracy by training a supernet (search space) from scratch and measure latency on target device. However, it's impractical to conduct large-scale search due to the prohibitive cost. For example, it costs more than 10 days to train a supernet on 8 V100 GPUs [41]. To reduce the cost, we build an accurate INT8 latency predictor by nn-Meter [42], then + +![](images/8b4d198bbb2577c9cb031571f045c5bc0c5107162a63744ba63b29e70dcc5923.jpg) +Figure 5. We adopt block-wise knowledge distillation to reduce search space quality evaluation cost. We add two linear transformation layers (Conv $1 \times 1$ ) to match teacher's feature map widths. + +Step1: Quantize elastic stages via BKD Step2: Build quantized accuracy LUT + +
Elastic stageModel block encoding (when depth=1)NSR loss
EMBv2,[32,48]k3e3c48-56x560.39
EMBv3,[32,48]k5e6c32-40x400.37
+ +k: kernel size, e: expand ratio, c: width + +we introduce block-wise quantization scheme. + +Block-wise knowledge distillation (BKD) is firstly proposed in DNA [22] and then further improved in DONNA [26]. It originally uses block-wise representation of existing models (teacher) to supervise a corresponding student model block. This technique can provide a relative accuracy ranking of all possible models without requiring them to be trained from scratch. In our work, we extend BKD to supervise the training of all elastic stages (each contains a large amount of blocks). + +Fig. 5 illustrates the BKD process. In the first step, we use EfficientNet-B5 as the teacher, and separately train each elastic stage to mimic the behavior of corresponding teacher block by minimizing the NSR loss [26] between their output feature maps. Specifically, the $i^{th}$ stage receives the output of $(i - 1)^{th}$ teacher block as the input and is optimized to predict the output of $i^{th}$ teacher block with NSR loss. Since an elastic stage contains many blocks with different channel widths, we add two learnable linear transformation layers at the input and output for each elastic stage to match teacher's feature map shape. Moreover, we adopt sandwich rule [41] to sample four paths to improve the training efficiency. Each elastic stage is firstly trained for 5 epochs and then performed 1 epoch LSQ+ [3] for INT8 quantization. + +Accuracy lookup table. In the second step, we construct a INT8 accuracy lookup table to reduce the evaluation cost. Specifically, we evaluate all possible blocks in each elastic stage and record their NSR losses on the validation set in the lookup table. The quantized loss of a model is estimated by summing up the NSR loss of all its blocks by rapidly looking up each elastic stages from the table. We inverse the measured loss to approximate the quantized accuracy for Q-T score evaluation. + +In our work, the BKD and lookup table construction can be sped up in a parallel way and finished in 1 day, which amounts a one-time cost before aging evolution search. + +# 5. Evaluation + +Setup. We evaluate our method on ImageNet-1k dataset [9] and two popular edge devices. The INT8 latency constraints are \{8, 10, 15, 20, 25\} ms for Intel CPU, and \{15, 20, 25, 30, 35\} ms for Pixel4. For each device, we search 5k search + +![](images/e0c9fd3e30b55f048ff1e0edc3feb90f788e8f0c42dd5a885663cfb30176019f.jpg) +Figure 6. Best searched INT8 models with comparison to state-of-the-art NAS search spaces. Our searched spaces are proven to be the most quantization-friendly for the target device. + +spaces in total and return the one with highest Q-T score. The population size $P$ is 500 and sample size $S$ is 125. + +Once SpaceEvo discovers a quantization-friendly search space for the target device, we train a quantized-for-all supernet. We start by pretraining a full-precision supernet without quantizers on ImageNet for 360 epochs. We adopt the sandwich rule and inplace distillation in BigNAS [41]. Then, we perform quantization-aware training (QAT) on the trained supernet for 50 epochs, which follows the same training protocol (i.e., sandwich rule and inplace distillation). We use $\mathrm{LSQ + }$ as the QAT algorithm for better quantized accuracy. To derive INT8 model for deployment, we use the evolutionary search in OFA [4] to search $5\mathrm{k}$ models for various given INT8 latency constraints. We list out detailed training settings in supplementary materials. In the following, we refer to the two searched spaces as SpaceEvo@CPU and SpaceEvo@Pixel4, the searched model families are SEQnet@cpu and SEQnet@pixel4. + +# 5.1. The Effectiveness of SpaceEvo + +Comparison with SOTA search spaces. To demonstrate the high-performance of our searched spaces, we compare with prior art manually-designed search spaces including: (1) MobileNetV3 search space that is adopted in two-stage quantization NAS OQAT [32] and BatchQuant [1]; (2) ProxylessNAS and AttentiveNAS search spaces that achieve superior performance on mobile devices; and (3) ResNet50 search space proposed by OFA that is a handcraft quantization-friendly space on our two devices. For fair comparison, we use one supernet training and QAT receipt for all search spaces. We conduct evolutionary search to compare the best-searched models from each search space. We use the random seed of 0. For all experiments, search space is the only difference. + +Fig. 6 compares the best searched INT8 models from different search spaces. SpaceEvo@CPU and SpaceEvo@Pixel4 consistently deliver superior quantized models than state-of-the-art search spaces. Under the same latency, the best quantized models from SpaceEvo@cpu significantly surpass the existing state-of-the-art search + +![](images/22e93fac7a91fbf3ff786cef6341a5323d607b9c8cb05c3ec4191f7576693b4c.jpg) +Figure 7. Search space design under diverse INT8 latency constraints. SpaceEvo (6-25 ms) delivers superior tiny INT8 models. + +![](images/1f428a6bf1ddef5c56eb20aee1981392e7bed3e642e14bd69ee4b1db1ffc7c59.jpg) +Figure 8. Search cost measured on 8 Nvidia V100 GPUs. + +spaces with $+0.7\%$ to $+3.8\%$ ( $+0.4\%$ to $+3.2\%$ on Pixel4) higher accuracy. Moreover, our search space is the only one that is able to deliver superior quantized models under both extremely low (only $\sim 5\mathrm{ms}$ ) and large latency constraints. + +SpaceEvo under diverse latency constraints. We extensively study the effectiveness of SpaceEvo under different latency constraints. Specifically, we perform space search under two tight constraints of $\{10, 15, 20, 25, 30\}\mathrm{ms}$ and $\{6, 10, 15, 20, 25\}\mathrm{ms}$ on Pixel4. The results are shown in Fig. 7. Our proposed method can handle the diverse latency requirements and produce high-quality spaces. As expected, the searched spaces under $10 - 30\mathrm{ms}$ and $6 - 25\mathrm{ms}$ have much more low-latency quantized models. + +To further verify the effectiveness of these low-latency models, we compare with existing SOTA tiny models. Significantly, even under the extremely low latency constraints of 6-25 ms, our searched space delivers very competitive tiny quantized models. Compared to the smallest model ShuffleNetV2x0.5, we can achieve $+10.1\%$ higher accuracy under the same latency of 4.3 ms. + +Search cost. As depicted in Fig. 8, our algorithm, SpaceEvo, is designed to be lightweight and suitable for real-world usage, requiring only 25 GPU hours to search a space of 5000 iterations. This remarkable speed is mainly due to our block-wise search space quantization scheme, which significantly reduces the cost of search space quality evaluation. In comparison, Fig. 8 demonstrates that training each search space from scratch without this scheme would consume an impractical 1200k GPU hours. + +# 5.2. The Effectiveness of Discovered INT8 Models + +In this section, we demonstrate that our searched spaces deliver state-of-the-art quantized models. We compare with two strong baselines: (1) prior art manually-designed and NAS-searched models; and (2) quantization-aware NAS. For + +Table 1. ImageNet results compared with SOTA quantized models on two devices. *: latency compared to FP32 inference. + +
(a) Results on the Intel VNNI CPU with onnxruntime
ModelAcc% INT8CPU latencyFP32 FLOPs
INT8speedup*
MobileNetV3Small66.34.4 ms1.1×67.456M
SEQnet@cpu-A074.74.4 ms2.0×74.8163M
MobileNetV271.47.3 ms2.2×72.0300M
ProxylessNAS-R74.68.8 ms1.8×74.6320M
OQAT-8bit74.89.8 ms1.8×75.2214M
MobileNetV3Large74.510.3 ms1.5×75.2219M
OFA (#25)75.611.2 ms1.5×76.4230M
SEQnet@cpu-A177.48.8 ms2.4×77.5358M
APQ-8bit73.615.0 ms1.5×73.6297M
AttentiveNAS-A076.115.1 ms1.4×77.3203M
OQAT-8bit76.314.9 ms1.7×76.7316M
EfficientNet-B076.718.1 ms1.6×77.3390M
SEQnet@cpu-A278.514.1 ms2.4×78.8638M
APQ-8bit74.920.0 ms1.5×75.0393M
OQAT-8bit76.919.5 ms1.6×77.3405M
AttentiveNAS-A177.222.4 ms1.4×78.4279M
AttentiveNAS-A277.522.5 ms1.3×78.8317M
SEQnet@cpu-A379.518.9 ms2.6×79.6981M
FBNetV2-L175.825.0 ms1.2×77.2325M
FBNetV3-A78.227.7 ms1.3×79.1357M
SEQnet@cpu-A480.024.4 ms2.4×80.11267M
(b) Results on the Google Pixel 4 with TFLite
ModelAcc% INT8Pixel4 Latency|Acc% FP32FLOPs
INT8speedup*
MobileNetV3Small66.36.4 ms1.3×67.456M
SEQnet@pixel4-A073.65.9 ms2.1×73.7107M
MobileNetV271.416.5 ms1.9×72.0300M
ProxylessNAS-R74.618.4 ms1.8×74.6320M
MobileNetV3Large74.515.7 ms1.5×75.2219M
APQ-8bit74.614.9 ms2.0×74.4340M
OFA (#25)75.614.8 ms1.7×76.4230M
OQAT-8bit75.815.2 ms1.9×76.2287M
AttentiveNAS-A076.115.2 ms2.0×77.3203M
SEQnet@pixel4-A177.614.7 ms2.2×77.7274M
APQ-8bit75.120.0 ms1.9×75.1398M
OQAT-8bit76.520.4 ms1.8×76.8347M
AttentiveNAS-A177.221.1 ms2.0×78.4279M
AttentiveNAS-A277.522.7 ms2.0×78.8317M
SEQnet@pixel4-A278.319.4 ms2.3×78.4402M
FBNetV2-L175.826.7 ms1.5×77.2325M
OQAT-8bit77.029.9 ms1.7×77.2443M
FBNetV3-A78.230.5 ms1.5×79.1357M
SEQnet@pixel4-A379.530.8 ms2.1×79.5591M
EfficientNet-B076.736.4 ms1.7×77.3390M
SEQnet@pixel4-A479.935.5 ms2.2×80.0738M
+ +baseline (1), we collect official pre-trained FP32 checkpoints and conduct LSQ+ QAT to get the quantized accuracy. The hyperparameter settings follow the original LSQ+ paper, except that we set a larger epoch of 10 to achieve better accuracy. The latency numbers are measured on our devices. For (2), we compare with strong baselines including APQ [37] and OQAT [32]. Specifically, we limit APQ to search for the fixed 8bit (INT8) models. Since OQAT has no 8bit supernet checkpoint, we follow the official source code and conduct supernet QAT for 50 epochs. The final INT8 models are searched under the same INT8 latency predictors for fair comparison. + +Results. Table 1 summarizes comparison results. Re + +![](images/336c065d35c53d8fbd1aa4a5c588f7fb4dbd983be893ef492602f998a15cf25b.jpg) +Figure 9. Q-T score effectiveness (Kendall's $\tau$ ) on ranking search space quality. We achieve a high space ranking correlation. + +Table 2. Different space search methods and their best resulting quantized models on Pixel4. Baseline is a SOTA mobile-friendly AttentiveNAS space. *: the search dimension use the same settings in AttentiveNAS. + +
MethodOpWidthBest quantized models
10ms15ms20ms30ms36ms
Baseline--76.677.279.079.5
SpaceEvo-opsearchfix*75.076.677.878.678.8
SpaceEvo-widthfix*search75.477.478.079.179.5
SpaceEvosearchsearch75.777.678.379.579.9
+ +markably, our searched model family, SEQnet significantly outperform SOTA efficient models and quantization-aware NAS searched models, with higher INT8 quantized accuracy, lower INT8 latency and better speedups. Without finetuning, our tiny models SEQnet@cpu-A0 and SEQnet@pixel4-A0 achieve $74.7\%$ and $73.6\%$ top1 accuracy on ImageNet, which is $8.4\%$ and $7.3\%$ higher than MobileNetV3-Small (56M FLOPs) while maintaining the same level quantized latency. For larger models, SEQnet@cpu-A4 $(80.0\%)$ outperforms FBNetV3-A with $1.8\%$ higher accuracy while runs 3.3ms faster. In particular, to achieve the same level accuracy (i.e., around $77.2\%$ ), AttentiveNAS-A1 has 22.4ms latency while SEQnet@cpu-A1 $(77.4\%)$ only needs 8.8 ms $(2.6\times$ faster). More importantly, our searched models can better utilize the INT8 hardware optimizations: the latency speedups compared to full-precision inference are all larger than $2\times$ , and this leaves room to search large-size models with higher accuracy. + +# 5.3. Ablation Study + +Q-T score effectiveness. Q-T score is crucial as it guides the space evolution process. To evaluate its effectiveness, we randomly sample 30 search spaces, and measure the rank correlation (Kendall's $\tau$ ) between their Q-T score and their actual Pareto-frontier models' accuracies. Specifically, we use Intel CPU as the test device and set a same latency constraints of $\{8, 10, 15, 20, 25\}$ ms. For each sampled space, we train it from scratch for 50 epochs, and conduct evolutionary search to get the Pareto-frontier models' accuracies. As shown in Fig. 9, the Kendall's $\tau$ between the Q-T score and the actual Pareto-frontier models' accuracies is 0.8, which indicates a very high rank correlation. + +Ablation study on two search dimensions. In Sec. 3, we conclude that operator type and configuration are two key factors impacting INT8 latency, which serves as the two search objectives of SpaceEvo. To verify the effectiveness, we create two strong baselines based on the SOTA edge-regime AttentiveNAS search space: (1) SpaceEvo-op: we fix each elastic stage's width to AttentiveNAS space, then allow each elastic stage to search for the optimal operator; and (2) SpaceEvo-width: we fix all elastic stages' block types to AttentiveNAS space, then search for the optimal width. Table 2 reports the space comparison between different search methods on the Pixel4. By searching both operator type and width, SpaceEvo finds the optimal search space where its best searched quantized models achieve the highest accuracy under all latency constraints. Moreover, even searching for one dimension, SpaceEvo-op and SpaceEvo-width outperform the manually-designed AttentiveNAS space under small latency constraints. + +Search space design implications. We now summarize our learned experience and implications for designing quantization-friendly search spaces. We notice that the searched spaces show different preferences when targeting different devices: (i) All stages should use much wider channel widths compared to existing manually-designed spaces on the cpu, while only early stages prefer wider channels on Pixel 4. (ii) Since SE and Swish are INT8 latency-friendly on mobile phones, so our auto-generated search spaces for Pixel4 have many MBv3 stages. On Intel CPU, INT8 quantization slows down SE, Hardwish, and Swish, making FusedMB and MBv2 the priority for search spaces, with only the last two stages using MBv3. The details are provided in supplementary. + +# 6. Conclusion + +In this paper, we introduced SpaceEvo, the first to automatically design a quantization-friendly space for target device, which delivers superior INT8 quantized models with SOTA efficiency on real-world edge devices. Extensive experiments on two popular devices demonstrate its effectiveness compared to prior art manual-designed search spaces. We plan to apply SpaceEvo to other hardware efficiency such as energy-efficient search space design in the future. + +# References + +[1] Haoping Bai, Meng Cao, Ping Huang, and Jiulong Shan. Batchquant: Quantized-for-all architecture search with robust quantizer. In NeurIPS, 2021. 1, 3, 5, 7 +[2] Ron Banner, Yury Nahshan, and Daniel Soudry. Post training 4-bit quantization of convolutional networks for rapid-deployment. In Advances in Neural Information Processing Systems, 2019. 3 +[3] Yash Bhalgat, Jinwon Lee, Markus Nagel, Tijmen Blankevoort, and Nojun Kwak. Lsq+: Improving low-bit + +quantization through learnable offsets and better initialization. In CVPRW, 2020. 1, 3, 6 +[4] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. In ICLR, 2020. 2, 3, 5, 7 +[5] Han Cai, Ligeng Zhu, and Song Han. ProxylessNAS: Direct neural architecture search on target task and hardware. In International Conference on Learning Representations, 2019. 3 +[6] Zhaowei Cai and Nuno Vasconcelos. Rethinking differentiable search for mixed-precision neural networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2346–2355. Computer Vision Foundation / IEEE, 2020. 1, 3 +[7] Minghao Chen, Kan Wu, Bolin Ni, Houwen Peng, Bei Liu, Jianlong Fu, Hongyang Chao, and Haibin Ling. Searching the search space of vision transformers. In NeurIPS, 2021. 2, 3 +[8] Yuanzheng Ci, Chen Lin, Ming Sun, Boyu Chen, Hongwen Zhang, and Wanli Ouyang. Evolving search space for neural architecture search. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6639-6649, 2021. 2, 3 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248-255. IEEE Computer Society, 2009. 6 +[10] Steven K. Esser, Jeffrey L. McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S. Modha. Learned step size quantization. In ICLR, 2020. 1, 3 +[11] Google. Tensorflow lite, 2022. 3 +[12] Zichao Guo, Xiangyu Zhang, Haoyuan Mu, Wen Heng, Zechun Liu, Yichen Wei, and Jian Sun. Single path one-shot neural architecture search with uniform sampling. In ECCV, 2020. 1, 2, 3, 6 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5 +[14] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, Quoc V. Le, and Hartwig Adam. Searching for mobilenetv3. In ICCV, 2019. 2, 5 +[15] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. CoRR, abs/1704.04861, 2017. 5 +[16] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 2 +[17] Yiming Hu, Yuding Liang, Zichao Guo, Ruosi Wan, Xiangyu Zhang, Yichen Wei, Qingyi Gu, and Jian Sun. Angle-based search space shrinking for neural architecture search. In ECCV. 3 + +[18] Intel. Intel 64 and ia-32 architectures software developer's manual, 2022. 4 +[19] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 1, 3 +[20] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2704-2713, 2018. 3 +[21] Sumin Kim, Gunju Park, and Youngmin Yi. Performance evaluation of int8 quantized inference on mobile gpus. IEEE Access, 9:164245-164255, 2021. 4 +[22] Changlin Li, Jiefeng Peng, Liuchun Yuan, Guangrun Wang, Xiaodan Liang, Liang Lin, and Xiaojun Chang. Blockwisely supervised neural architecture search with knowledge distillation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, 2020. 2, 6 +[23] Xiang Li, Chen Lin, Chuming Li, Ming Sun, Wei Wu, Junjie Yan, and Wanli Ouyang. Improving one-shot nas by suppressing the posterior fading. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[24] Christos Louizos, Mathias Reisser, Tijmen Blankevoort, Efstratos Gavves, and Max Welling. Relaxed quantization for discretized neural networks. In ICLR, 2019. 3 +[25] Microsoft. onnxruntime, 2022. 3 +[26] Bert Moons, Parham Noorzad, Andrii Skliar, Giovanni Mariani, Dushyant Mehta, Chris Lott, and Tijmen Blankevoort. Distilling optimal neural networks: Rapid search in diverse spaces. In ICCV, 2021. 2, 6 +[27] Markus Nagel, Mart van Baalen, Tijmen Blankevoort, and Max Welling. Data-free quantization through weight equalization and bias correction. In ICCV, 2019. 1, 3 +[28] Asaf Noy, Niv Nayman, Tal Ridnik, Nadav Zamir, Sivan Doveh, Itamar Friedman, Raja Giryes, and Lihi Zelnik. Asap: Architecture search, anneal and prune. In Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, pages 493-503, 2020. 3 +[29] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In CVPR, 2020. 3 +[30] Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image classifier architecture search. In AAAI, 2019. 2, 4, 6 +[31] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5 + +[32] Mingzhu Shen, Feng Liang, Ruihao Gong, Yuhang Li, Chuming Li, Chen Lin, Fengwei Yu, Junjie Yan, and Wanlin Ouyang. Once quantization-aware training: High performance extremely low-bit architecture search. In ICCV, 2021. 1, 3, 5, 7, 8 +[33] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In Proceedings of the 38th International Conference on Machine Learning, pages 10096-10106. PMLR, 2021. 5 +[34] Xiaohu Tang, Shihao Han, Li Lyna Zhang, Ting Cao, and Yunxin Liu. To bridge neural network design and real-world performance: A behaviour study for neural networks. In A. Smola, A. Dimakis, and I. Stoica, editors, Proceedings of Machine Learning and Systems, volume 3, pages 21-37, 2021. 3, 4 +[35] Dilin Wang, Meng Li, Chengyue Gong, and Vikas Chandra. Attentivenas: Improving neural architecture search via attentive sampling. In Conference on Computer Vision and Pattern Recognition, 2021. 3, 5 +[36] Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, and Song Han. Haq: Hardware-aware automated quantization with mixed precision. In CVPR, 2019. 1, 3 +[37] Tianzhe Wang, Kuan Wang, Han Cai, Ji Lin, Zhijian Liu, and Song Han. Apq: Joint search for network architecture, pruning and quantization policy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 1, 3, 8 +[38] Xudong Wang, Li Lyna Zhang, Yang Wang, and Mao Yang. Towards efficient vision transformer inference: A first study of transformers on mobile devices. In Proceedings of the 23rd Annual International Workshop on Mobile Computing Systems and Applications, HotMobile '22, page 1-7. Association for Computing Machinery, 2022. 3 +[39] Bichen Wu, Yanghan Wang, Peizhao Zhang, Yuandong Tian, Peter Vajda, and Kurt Keutzer. Mixed precision quantization of convnets via differentiable neural architecture search. 2019. 1, 3 +[40] Xin Xia, Xuefeng Xiao, Xing Wang, and Min Zheng. Progressive automatic design of search space for one-shot neural architecture search. In WACV, 2022. 3 +[41] Jiahui Yu, Pengchong Jin, Hanxiao Liu, Gabriel Bender, Pieter-Jan Kindermans, Mingxing Tan, Thomas Huang, Xiaodan Song, Ruoming Pang, and Quoc Le. Bignas: Scaling up neural architecture search with big single-stage models. In ECCV, 2020. 3, 5, 6, 7 +[42] Li Lyna Zhang, Shihao Han, Jianyu Wei, Ningxin Zheng, Ting Cao, Yuqing Yang, and Yunxin Liu. nn-meter: Towards accurate latency prediction of deep-learning model inference on diverse edge devices. In Proceedings of the 19th Annual International Conference on Mobile Systems, Applications, and Services, page 81-93, New York, NY, USA, 2021. ACM. 3, 6 \ No newline at end of file diff --git a/2023/SpaceEvo_ Hardware-Friendly Search Space Design for Efficient INT8 Inference/images.zip b/2023/SpaceEvo_ Hardware-Friendly Search Space Design for Efficient INT8 Inference/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0bff946741e5f365b84e49967a36c62673229a94 --- /dev/null +++ b/2023/SpaceEvo_ Hardware-Friendly Search Space Design for Efficient INT8 Inference/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b9b95591dfce85d1526d10081e53b0fc9b6b8541a89b6a042dfa734913c1ac6 +size 563843 diff --git a/2023/SpaceEvo_ Hardware-Friendly Search Space Design for Efficient INT8 Inference/layout.json b/2023/SpaceEvo_ Hardware-Friendly Search Space Design for Efficient INT8 Inference/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7882167ff38b212bc3b741a3f5e5c40b63304881 --- /dev/null +++ b/2023/SpaceEvo_ Hardware-Friendly Search Space Design for Efficient INT8 Inference/layout.json @@ -0,0 +1,8829 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 103, + 545, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 103, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 545, + 121 + ], + "type": "text", + "content": "SpaceEvo: Hardware-Friendly Search Space Design for Efficient INT8 Inference" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 152, + 493, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 152, + 493, + 182 + ], + "spans": [ + { + "bbox": [ + 73, + 152, + 493, + 182 + ], + "type": "text", + "content": "Xudong Wang\\*1,2§ Li Lyna Zhang\\*2‡ Jiahang Xu² Quanlu Zhang² Yujing Wang³ Yuqing Yang² Ningxin Zheng² Ting Cao² Mao Yang²" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 141, + 189, + 449, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 189, + 449, + 204 + ], + "spans": [ + { + "bbox": [ + 141, + 189, + 449, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 189, + 449, + 204 + ], + "type": "text", + "content": "Shanghai Jiao Tong University " + }, + { + "bbox": [ + 141, + 189, + 449, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 141, + 189, + 449, + 204 + ], + "type": "text", + "content": "Microsoft Research " + }, + { + "bbox": [ + 141, + 189, + 449, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 141, + 189, + 449, + 204 + ], + "type": "text", + "content": "Microsoft" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 204, + 523, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 204, + 523, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 204, + 523, + 217 + ], + "type": "text", + "content": "xudongwang176@gmail.com; {Izhani, jiahangxu, quzha, yujwang, yuqyang, nizhen, ticao, maoyang}@microsoft.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 217, + 462, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 462, + 230 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 462, + 230 + ], + "type": "text", + "content": "https://github.com/microsoft/Moonlit/tree/main/SpaceEvo" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 265, + 192, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 265, + 192, + 277 + ], + "spans": [ + { + "bbox": [ + 143, + 265, + 192, + 277 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 290, + 290, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 290, + 614 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 290, + 614 + ], + "type": "text", + "content": "The combination of Neural Architecture Search (NAS) and quantization has proven successful in automatically designing low-FLOPs INT8 quantized neural networks (QNN). However, directly applying NAS to design accurate QNN models that achieve low latency on real-world devices leads to inferior performance. In this work, we identify that the poor INT8 latency is due to the quantization-unfriendly issue: the operator and configuration (e.g., channel width) choices in prior art search spaces lead to diverse quantization efficiency and can slow down the INT8 inference speed. To address this challenge, we propose SpaceEvo, an automatic method for designing a dedicated, quantization-friendly search space for each target hardware. The key idea of SpaceEvo is to automatically search hardware-preferred operators and configurations to construct the search space, guided by a metric called " + }, + { + "bbox": [ + 46, + 290, + 290, + 614 + ], + "type": "inline_equation", + "content": "Q-T" + }, + { + "bbox": [ + 46, + 290, + 290, + 614 + ], + "type": "text", + "content": " score to quantify how quantization-friendly a candidate search space is. We further train a quantized-for-all supernet over our discovered search space, enabling the searched models to be directly deployed without extra retraining or quantization. Our discovered models, SEQnet, establish new SOTA INT8 quantized accuracy under various latency constraints, achieving up to " + }, + { + "bbox": [ + 46, + 290, + 290, + 614 + ], + "type": "inline_equation", + "content": "10.1\\%" + }, + { + "bbox": [ + 46, + 290, + 290, + 614 + ], + "type": "text", + "content": " accuracy improvement on ImageNet than prior art CNNs under the same latency. Extensive experiments on real devices show that SpaceEvo consistently outperforms manually-designed search spaces with up to " + }, + { + "bbox": [ + 46, + 290, + 290, + 614 + ], + "type": "inline_equation", + "content": "2.5\\times" + }, + { + "bbox": [ + 46, + 290, + 290, + 614 + ], + "type": "text", + "content": " faster speed while achieving the same accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 639, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 639, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 639, + 287, + 676 + ], + "type": "text", + "content": "INT8 Quantization[27, 19, 10, 3] is a widely used technique for deploying DNNs on edge devices by reducing " + }, + { + "bbox": [ + 46, + 639, + 287, + 676 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 46, + 639, + 287, + 676 + ], + "type": "text", + "content": " in model size and memory cost for full-precision (FP32)" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 262, + 541, + 418 + ], + "blocks": [ + { + "bbox": [ + 312, + 262, + 541, + 418 + ], + "lines": [ + { + "bbox": [ + 312, + 262, + 541, + 418 + ], + "spans": [ + { + "bbox": [ + 312, + 262, + 541, + 418 + ], + "type": "image", + "image_path": "3366466d8acddb126a6fe904cf34998e83103147b276e6c158fcdf74470fdc1a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 424, + 547, + 469 + ], + "lines": [ + { + "bbox": [ + 305, + 424, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 305, + 424, + 547, + 469 + ], + "type": "text", + "content": "Figure 1. INT8 quantized model error distributions (EDF) of different NAS search spaces. We propose to search quantization-friendly search space for each hardware, which yields significantly better INT8 model populations than SOTA search spaces." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 480, + 546, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 546, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 546, + 553 + ], + "type": "text", + "content": "models. However, prior art DNN models achieve only marginal speedup from INT8 quantization (in Fig. 2(a)), the still high latency after quantization making them difficult to deploy in latency-critical scenarios. Designing models that achieve high accuracy and low latency after quantization becomes the important but challenging problem." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "Neural Architecture Search (NAS) is a powerful tool for automating efficient quantized model design [36, 39, 12, 37, 6]. Recently, OQAT [32] and BatchQuant [1] achieve remarkable search efficiency and accuracy by adopting a two-stage paradigm. The first stage trains a weight-shared quantized supernet assembling all candidate architectures in the search space. This allows all the sub-networks (subnets) to simultaneously reach comparable quantized accuracy as when trained from scratch individually. The second stage uses typical search algorithms to find subnets with best quantized accuracy under different FLOPs constraints. This approach avoids the need to retrain each subnet for accuracy evaluation, greatly improving the search efficiency." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 682, + 124, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 682, + 124, + 693 + ], + "spans": [ + { + "bbox": [ + 57, + 682, + 124, + 693 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 58, + 693, + 255, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 255, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 255, + 703 + ], + "type": "inline_equation", + "content": "^{\\S}" + }, + { + "bbox": [ + 58, + 693, + 255, + 703 + ], + "type": "text", + "content": "Work was done during the internship at Microsoft Research" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 58, + 703, + 215, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 215, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 215, + 712 + ], + "type": "text", + "content": "†Corresponding author (lzhani@microsoft.com)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5819" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "Though promising in optimizing model FLOPs, we identify a significant challenge when directly applying two-stage NAS to low quantized latency scenarios. This is due to the quantization-unfriendly search space issue, where prior art search spaces can unexpectedly impede INT8 latency. Consequently, since INT8 quantization yields only a marginal speedup, we are forced to search for small-sized models to fulfill latency criteria, which can unfortunately restrict NAS to find better quantized models for edge devices. Then, a question naturally arises: Can we design a quantization-friendly search space, allowing NAS to discover larger and superior models that meet the low INT8 latency requirements?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 230, + 289, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 230, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 289, + 410 + ], + "type": "text", + "content": "We start by conducting an in-depth study to understand the factors that determine INT8 quantized latency and how they affect search space design. Our study shows: (1) both operator type and configurations (e.g., channel width) greatly impact the INT8 latency; Improper selections can slow down the INT8 latency. For instance, Squeeze-and-Excitation (SE) [16] and Hardswish [14] are widely-used operators in current search spaces as it improves accuracy with little latency introduced. However, their INT8 inference speeds are slower than FP32 inference on Intel CPU (Fig. 3(a)), because the extra costs (e.g., data transformation between INT32 and INT8) introduced by quantization outweigh the latency reduction by INT8 computation. (2) The quantization efficiency varies across different devices, and the preferred operator types can be contradictory." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 411, + 289, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 289, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 289, + 592 + ], + "type": "text", + "content": "The above study motivates us to design specialized quantization-friendly search spaces for each hardware, rather than relying on a single, large search space as seen in SPOS [12] for all hardware, which provides different operator options per layer. This is because two-stage NAS requires the search space to adhere to a specific condition for training the supernet, wherein each layer must utilize a fixed operator. Our study indicates significant variations in optimal operators across different hardware. Thus, customizing the search space for each hardware is crucial for optimal results. However, designing such specialized quantization-friendly search spaces for various edge devices presents a significant challenge, requiring expertise in both AI and hardware domains, as well as many trial and error attempts to optimize accuracy and INT8 latency for each hardware." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "In this paper, we propose SpaceEvo, a novel method for automatically designing specialized quantization-friendly search space for each hardware. The search space is comprised of hardware-preferred operators and configurations, enabling the search of larger and better models with low INT8 latency. With the discovered search space, we leverage two-stage quantization NAS to train a quantized-for-all supernet, and utilize evolution search [4] to find best quantized models under various INT8 latency constraints. Our approach addresses three key challenges: (1) What is the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "definition of a quantization-friendly search space in terms of both quantized accuracy and latency? (2) How to automatically design a search space without human expertise? (3) How to handle with the prohibitive cost caused by quality evaluation of a candidate search space?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 133, + 545, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 205 + ], + "type": "text", + "content": "To address the first challenge, we propose a latency-aware " + }, + { + "bbox": [ + 304, + 133, + 545, + 205 + ], + "type": "inline_equation", + "content": "Q-T" + }, + { + "bbox": [ + 304, + 133, + 545, + 205 + ], + "type": "text", + "content": " score to quantify the effectiveness of a candidate search space, which measures the INT8 accuracy-latency quality of top-tier subnets in a search space. The behind intuition is that the goal of NAS is to search top subnets with better accuracy-latency tradeoffs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "text", + "content": "Then, we introduce an evolutionary-based search algorithm that can effectively search a quantization-friendly search space with highest " + }, + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "inline_equation", + "content": "Q-T" + }, + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "text", + "content": " score. Searching a search space involves discovering a collection of model population that contains billions of models, which is challenging and easily introduce complexity. To address this challenge, we propose to factorize and encode a search space into a sequence of elastic stages, which have flexible operator types and configurations. Through this design, the task of search space design is then simplified to find a search space with the optimal elastic stages, so that existing search algorithms can be easily applied. Specifically, we design a stage-wise hyperspace to include many candidate search spaces and leverage aging evolution [30] to perform random mutations of elastic stages for search space evolution. The evolution is guided by maximizing the " + }, + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "inline_equation", + "content": "Q-T" + }, + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "text", + "content": " score." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 398, + 546, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 546, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 546, + 588 + ], + "type": "text", + "content": "Finally, estimating the quality score (Q-T score) of a search space involves a costly training process for evaluating the accuracy of sub-networks, which presents a significant obstacle for our evolutionary algorithm. Naively adopting a two-stage NAS approach, training a supernet for each candidate search space [8, 7], is prohibitively expensive, taking of thousands GPU hours. To address this issue, we draw inspiration from block-wise knowledge distillation [26, 22] and propose a block-wise search space quantization scheme. This scheme trains each elastic stage separately and rapidly estimates a model's quantized accuracy by summing block-level loss with a quantized accuracy lookup table, as shown in Fig. 5. This significantly reduces the training and evaluation costs, while providing effective accuracy rankings among search spaces. We summarize our contributions as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 594, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 594, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 306, + 594, + 545, + 654 + ], + "type": "text", + "content": "- We study the INT8 quantization efficiency on real-world edge devices and find that the choices of operator types and configurations in a quantized model can significantly impact the INT8 latency, leaving a huge room for design optimization of quantized models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 666, + 545, + 713 + ], + "type": "text", + "content": "- We propose SpaceEvo to automatically design a hardware-dedicated quantization-friendly search space and leverage two-stage quantization NAS to produce superior INT8 models under various latency constraints." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5820" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 247 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "- We present three innovative techniques that enable the first-ever efficient and cost-effective evolution search to explore a search space comprising billions of models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 115, + 288, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 115, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 115, + 288, + 247 + ], + "type": "text", + "content": "- Extensive experiments on two real-world edge devices and ImageNet demonstrate that our automatically designed search spaces significantly surpass manually-designed search spaces. Our discovered models, SEQnet, establish the new state-of-the-art INT8 quantized accuracy-latency tradeoffs. For instance, SEQnet@cpuA4, achieves " + }, + { + "bbox": [ + 47, + 115, + 288, + 247 + ], + "type": "inline_equation", + "content": "80.0\\%" + }, + { + "bbox": [ + 47, + 115, + 288, + 247 + ], + "type": "text", + "content": " accuracy on ImageNet, which is 3.3ms faster with " + }, + { + "bbox": [ + 47, + 115, + 288, + 247 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 47, + 115, + 288, + 247 + ], + "type": "text", + "content": " higher accuracy than FBNetV3-A. Moreover, SpaceEvo delivers superior tiny models, achieving up to " + }, + { + "bbox": [ + 47, + 115, + 288, + 247 + ], + "type": "inline_equation", + "content": "10.1\\%" + }, + { + "bbox": [ + 47, + 115, + 288, + 247 + ], + "type": "text", + "content": " accuracy improvement over ShuffleNetV2x0.5 (41M FLOPs, 4.3ms)." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 256, + 139, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 256, + 139, + 269 + ], + "spans": [ + { + "bbox": [ + 47, + 256, + 139, + 269 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 277, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 384 + ], + "type": "text", + "content": "Quantization has been widely used for efficiency in deployment. Extensive efforts can be classified into posttraining quantization (PTQ) [27, 2] and quantization-aware training (QAT) [19, 24, 10, 3]. QAT generally outperforms PTQ in quantizing compact DNNs to 8bit and very low-bit (2, 3, 4bit) by finetuning the quantized weights. Despite their success, traditional quantization methods focus on minimizing accuracy loss for a given pre-trained model, but ignore the real-world inference efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 384, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 504 + ], + "type": "text", + "content": "NAS for Quantization. Early works [36, 39, 12, 37, 6] formulates mixed-precision problem into NAS to search layer bit-width with a given architecture. Recently, [32, 1] train a quantized-for-all supernet to search both architecture and bit-width. The searched models can be directly evaluated with comparable accuracy to train-from-scratch. However, little attention is paid on optimizing quantized latency on real-world devices. Through searching quantization-friendly search space, our discovered quantized models can achieve both high accuracy and low latency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 504, + 288, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 504, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 288, + 624 + ], + "type": "text", + "content": "Search Space Design. Starting from [5], the manually-designed MBConv-based space becomes the dominant in most NAS works [5, 4, 41, 35]. RegNet [29] is the first to present standard guidelines to optimize a search space by each dimension. Recently, [17, 28, 23, 40, 8, 7] propose to shrink to a better compact search space by either pruning unimportant operators or configurations. However, these works focus on optimizing the accuracy and little attention is paid on quantization-friendly search space design. Our work is the first lightweight solution towards this direction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 634, + 285, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 285, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 285, + 647 + ], + "type": "text", + "content": "3. On-device Quantization Efficiency Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 654, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 715 + ], + "type": "text", + "content": "To understand what factors lead to quantization-unfriendly issue, we conduct a comprehensive study on two widely-used edge devices equipped with high-performance inference engine: an Intel CPU device supported with VNNI instructions and onnxruntime [25] (Intel CPU) and" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 69, + 432, + 169 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 432, + 169 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 432, + 169 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 432, + 169 + ], + "type": "image", + "image_path": "1d5f9faa311863c173d7770af5f512e6f891b78bcc75e182cc871cd9dcc78b94.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 169, + 427, + 175 + ], + "lines": [ + { + "bbox": [ + 320, + 169, + 427, + 175 + ], + "spans": [ + { + "bbox": [ + 320, + 169, + 427, + 175 + ], + "type": "text", + "content": "(a) FLOPs and INT8 latency on Intel CPU" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 186, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 305, + 186, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 305, + 186, + 547, + 220 + ], + "type": "text", + "content": "Figure 2. INT8 latency and speedups (annotated) for SOTA models. FLOPs and FP32 latency are not good indicators of INT8 latency; Compact models have very limited INT8 speedup " + }, + { + "bbox": [ + 305, + 186, + 547, + 220 + ], + "type": "inline_equation", + "content": "(\\sim 1.5\\times)" + }, + { + "bbox": [ + 305, + 186, + 547, + 220 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 435, + 70, + 547, + 167 + ], + "blocks": [ + { + "bbox": [ + 435, + 70, + 547, + 167 + ], + "lines": [ + { + "bbox": [ + 435, + 70, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 435, + 70, + 547, + 167 + ], + "type": "image", + "image_path": "ef11762f49ee40899870cb25b2f045762028736767160477801850f6ba33f6e5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 167, + 535, + 175 + ], + "lines": [ + { + "bbox": [ + 440, + 167, + 535, + 175 + ], + "spans": [ + { + "bbox": [ + 440, + 167, + 535, + 175 + ], + "type": "text", + "content": "(b) FP32 and INT8 latency on Pixel 4" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 231, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 231, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 305, + 231, + 545, + 266 + ], + "type": "text", + "content": "a Pixel 4 phone CPU with TFLite 2.7 [11] (Pixel 4). Note that we follow existing practices [42, 34, 38] to measure the latency. We reveal key observations as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 269, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 269, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 269, + 545, + 293 + ], + "type": "text", + "content": "Observation 1: FP32 latency and FLOPs are not good indicators of INT8 latency." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 293, + 546, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 546, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 546, + 413 + ], + "type": "text", + "content": "To deploy on edge devices, a common belief is that a compact model with low FLOPs or FP32 latency is preferred than a larger model. However, Fig. 2 shows that neither of them is a good indicator of INT8 latency. In Fig. 2(b), a very large model (ResNet18) can be even faster than a compact model (EfficientNet-B0) after quantization. Moreover, the recent SOTA compact models searched by OFA [4] and AttentiveNAS [35] all have marginal INT8 speedups, suggesting that optimizing FLOPs and FP32 latency can not lead to lower INT8 latency." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 414, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 414, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 414, + 545, + 437 + ], + "type": "text", + "content": "Observation 2: The choices of operators' types and configurations greatly impact the INT8 latency." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 438, + 547, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 547, + 569 + ], + "type": "text", + "content": "The prior art search spaces adopted in recent two-stage NAS works are MobileNetV2 or MobileNetV3 based chain-structures, where each search space comprises a sequence of blocks (stages). The block type is fixed to the MBCnv and is allowed to search from a handcraft range of hyperparameter configurations including kernel size, expansion ratio, channel width and depth, which are designed with human wisdom. For instance, many works [5, 35] observe that edge-regime CNNs prefer deeper depths and narrower channels, and manually set small channel numbers but large depths in the search space." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "However, we find that many block type and configuration choices in current search spaces unexpectedly slow down the INT8 latency. We first study the operator type impact in Fig. 3(a). SE and Hardswish are lightweight operators in edge-regime search spaces, but their INT8 inference becomes slower on Intel CPU. Compared to Conv, DWConv can greatly reduce the FLOPs, but it benefits less from INT8 quantization. The root cause is that quantization introduces extra cost, such as (1) data transformation between INT32 and INT8 [34], and (2) additional computation caused by scaling fators and zero points [20]. If the operator has low data-reuse-rate, such as the activation functions (Hswish)," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5821" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 72, + 154, + 144 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 154, + 144 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 154, + 144 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 154, + 144 + ], + "type": "table", + "html": "
OperatorIntel CPUPixel 4
Conv2.6x2.5x
DWConv1.2x2.0x
SE0.7x1.4x
Hardswish0.7x0.5x
Swish0.8x2.1x
", + "image_path": "6e1f344dca7d439b905864784d5eb23394e3c8c19543b4ef83f98a3dbff6c374.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 156, + 70, + 289, + 150 + ], + "blocks": [ + { + "bbox": [ + 48, + 145, + 147, + 171 + ], + "lines": [ + { + "bbox": [ + 48, + 145, + 147, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 145, + 147, + 171 + ], + "type": "text", + "content": "(a) Avg. INT8 latency speedup (relative to FP32) of major operators in MobileNetV3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 70, + 289, + 150 + ], + "lines": [ + { + "bbox": [ + 156, + 70, + 289, + 150 + ], + "spans": [ + { + "bbox": [ + 156, + 70, + 289, + 150 + ], + "type": "image", + "image_path": "09f114e9a3ff2e742520a7fb1111b0ddeec7ceaac879c5098f9cdd8e005736f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 159, + 263, + 168 + ], + "lines": [ + { + "bbox": [ + 187, + 159, + 263, + 168 + ], + "spans": [ + { + "bbox": [ + 187, + 159, + 263, + 168 + ], + "type": "text", + "content": "(b) Conv1x1 speedups" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 180, + 287, + 213 + ], + "lines": [ + { + "bbox": [ + 46, + 180, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 287, + 213 + ], + "type": "text", + "content": "Figure 3. (a) The choice of operator type leads to significantly different quantized speedup. (b) Conv1x1 speedups with various channel numbers. Config: " + }, + { + "bbox": [ + 46, + 180, + 287, + 213 + ], + "type": "inline_equation", + "content": "HW = 28" + }, + { + "bbox": [ + 46, + 180, + 287, + 213 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 180, + 287, + 213 + ], + "type": "inline_equation", + "content": "C_{out} = 4 \\times C_{in}" + }, + { + "bbox": [ + 46, + 180, + 287, + 213 + ], + "type": "text", + "content": " (expand=4)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 225, + 287, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 225, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 287, + 262 + ], + "type": "text", + "content": "the extra cost may outweigh the latency reduction by the low-bit computation. For high data-reuse operators (Conv), this cost is amortized and thus achieve large speedup [34]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 263, + 287, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 287, + 370 + ], + "type": "text", + "content": "Besides the operator type, the configuration choices also determine the quantization efficiency. Fig. 3(b) shows the speedups of " + }, + { + "bbox": [ + 46, + 263, + 287, + 370 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}1\\times 1" + }, + { + "bbox": [ + 46, + 263, + 287, + 370 + ], + "type": "text", + "content": " under various channel widths. Results show that small channel widths in OFA search space cannot benefit well from quantization. This is because the additional quantization cost has a large impact when the channel width is small, limiting the latency acceleration. In contrast, SpaceEvo can automatically design a search space with larger channel widths for better efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 373, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 396 + ], + "type": "text", + "content": "Observation 3: Quantization-friendly settings are diverse and contradictory across devices." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 399, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 578 + ], + "type": "text", + "content": "In Fig. 3, we also observe that the quantization-friendly operators are different and can be contradictory on diverse devices. For instance, Swish achieves a " + }, + { + "bbox": [ + 46, + 399, + 287, + 578 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 46, + 399, + 287, + 578 + ], + "type": "text", + "content": " speedup on Pixel 4, but it is a quantization-unfriendly operator on CPU with a " + }, + { + "bbox": [ + 46, + 399, + 287, + 578 + ], + "type": "inline_equation", + "content": "0.8 \\times" + }, + { + "bbox": [ + 46, + 399, + 287, + 578 + ], + "type": "text", + "content": " slowdown. The reason is that quantization speedups are highly dependent on the inference engines and hardware [34, 21]. Intel VNNI supports the VPDPBWSD hardware instruction [18], which fuses three instructions into one to speedup INT8 computation. Without VNNI, INT8 hardly gains speedup on Intel CPUs. Moreover, the implementations in inference engines have to fully utilize hardware instructions for latency reduction. For example, onnxruntime does not implement a quantization kernel for Hardswish. Even on a VNNI-supported CPU, the use of Hardswish in a quantized model slows down the latency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "The need for hardware specialized search space. The above observations suggest that there is no single structure (block types in a model) that is optimal for quantization on all hardware. This poses a challenge for the two-stage NAS paradigm, as the supernet training requires all models in the search space to share an isomorphic structure. To address this issue, our work proposes to design a specialized quantization-friendly search space for each hardware. Each search space is tailored to the unique characteristics of the hardware and includes an optimal structure with elastic depths, widths, and kernel sizes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 71, + 388, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 388, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 388, + 85 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 91, + 448, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 448, + 104 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 448, + 104 + ], + "type": "text", + "content": "4.1. The Core Design Concept" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 106, + 545, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 106, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 304, + 106, + 545, + 237 + ], + "type": "text", + "content": "In this section, we present our methodology for automatically designing a specialized quantization-friendly search space for any target hardware. Different from architecture search, where the goal is to find the single best model from the space, we aim to discover a model population that contains billions of accurate and INT8 latency-friendly architectures. We draw inspiration from the neural architecture search process and propose to use an evolutionary search algorithm to explore such a quantization-friendly model population. To achieve this, we introduce SpaceEvo, which is built on the following three techniques." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 237, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 545, + 285 + ], + "type": "text", + "content": "First, we need a metric that quantifies how quantization-friendly a candidate search space is. We define a Q-T score that is efficiently measured by top-tier subnets' INT8 accuracy-latency (Sec. 4.2)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 285, + 545, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 381 + ], + "type": "text", + "content": "Second, existing evolutionary search algorithms are designed for searching a single model architecture rather than a large search space encompassing billions of architectures. we propose a novel approach that we call the \"elastic stage.\" By factorizing the search space into a sequence of elastic stages ((Sec. 4.3), we enable traditional aging evolution methods, such as the aging evolution [30], to be directly applied to search the space (Sec. 4.4)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 381, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 545, + 441 + ], + "type": "text", + "content": "Third, searching a search space with a maximum Q-T score can be prohibitively costly since the corresponding supernet must be trained from scratch for accuracy evaluation. We propose a block-wise search space quantization scheme to significantly reduce the training cost(Sec. 4.5)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 450, + 457, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 450, + 457, + 463 + ], + "spans": [ + { + "bbox": [ + 305, + 450, + 457, + 463 + ], + "type": "text", + "content": "4.2. Search Space Quality Score" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 464, + 545, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 633 + ], + "type": "text", + "content": "Latency-aware space quality of Q-T score. Before space search, we need a score to quantify how quantization-friendly a search space is, which serves as the search objective. Since our ultimate goal is to search the best quantized models from the searched space, we treat a space with good quality if its top-tier subnets achieve optimal quantized accuracy under latency constraints " + }, + { + "bbox": [ + 304, + 464, + 545, + 633 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 464, + 545, + 633 + ], + "type": "text", + "content": ". Due to the fact that real-world applications usually have different deployment requirements, we use multiple INT8 latency constraints to measure a space's quality. For search space " + }, + { + "bbox": [ + 304, + 464, + 545, + 633 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 464, + 545, + 633 + ], + "type": "text", + "content": " and a set of quantized latency constraints " + }, + { + "bbox": [ + 304, + 464, + 545, + 633 + ], + "type": "inline_equation", + "content": "T_{1,\\dots,n}" + }, + { + "bbox": [ + 304, + 464, + 545, + 633 + ], + "type": "text", + "content": ", we treat every constraint equally important, and define Q-T score as the sum of each constraint: " + }, + { + "bbox": [ + 304, + 464, + 545, + 633 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}(\\mathcal{A},T_1,\\dots,n) = \\mathcal{Q}(\\mathcal{A},T_1) + \\mathcal{Q}(\\mathcal{A},T_2) + \\dots" + }, + { + "bbox": [ + 304, + 464, + 545, + 633 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 464, + 545, + 633 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}(\\mathcal{A},T_n)" + }, + { + "bbox": [ + 304, + 464, + 545, + 633 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 464, + 545, + 633 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}(\\mathcal{A},T_i)" + }, + { + "bbox": [ + 304, + 464, + 545, + 633 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 339, + 642, + 545, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 642, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 339, + 642, + 545, + 656 + ], + "type": "interline_equation", + "content": "\\mathcal {Q} (\\mathcal {A}, T _ {i}) = \\mathbb {E} _ {\\alpha \\in \\mathcal {A}, L A T (\\alpha) \\leq T _ {i}} \\left[ A c c _ {i n t 8} (\\alpha) \\right] \\tag {1}", + "image_path": "507d2c8f5c2ebdfe45e22758bbea208fee9b4ee0424952c0a05e305db3573705.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " denotes a top-tier (best searched) subnet in " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "Acc_{int8}(\\alpha)" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " is its top-1 quantized accuracy evaluated on ImageNet validation set, " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "LAT(\\alpha)" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " predicts the quantized latency on target device." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5822" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 183, + 152 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 183, + 152 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 183, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 183, + 152 + ], + "type": "image", + "image_path": "fe5f2339213858253dd54127a61124c59344e72f0e77b9f0955cff1d0685ac57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 156, + 175, + 167 + ], + "lines": [ + { + "bbox": [ + 72, + 156, + 175, + 167 + ], + "spans": [ + { + "bbox": [ + 72, + 156, + 175, + 167 + ], + "type": "text", + "content": "(a) space search overview" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 188, + 70, + 547, + 157 + ], + "blocks": [ + { + "bbox": [ + 188, + 70, + 547, + 157 + ], + "lines": [ + { + "bbox": [ + 188, + 70, + 547, + 157 + ], + "spans": [ + { + "bbox": [ + 188, + 70, + 547, + 157 + ], + "type": "image", + "image_path": "a041b75c3603dccee23bc368b1dc4b48bb885ec8431fa871d738b5bdd8db5faa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 157, + 425, + 167 + ], + "lines": [ + { + "bbox": [ + 323, + 157, + 425, + 167 + ], + "spans": [ + { + "bbox": [ + 323, + 157, + 425, + 167 + ], + "type": "text", + "content": "(b) stage-wise hyperspace" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 172, + 546, + 195 + ], + "lines": [ + { + "bbox": [ + 46, + 172, + 546, + 195 + ], + "spans": [ + { + "bbox": [ + 46, + 172, + 546, + 195 + ], + "type": "text", + "content": "Figure 4. (a) We simplify space search into model search process; (b) Illustration of our hyperspace. A sampled search space is encoded by a sequential elastic stages. Contents in blue are searched: an elastic stage can search its block type and channel number list." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 215, + 287, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 215, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 287, + 310 + ], + "type": "text", + "content": "However, it's non-trivial to obtain the top-tier subnets from a candidate search, as if often involves an expensive full architecture search process. We adopt a zero-cost policy. Specifically, we randomly sample 5k subnets and select top 20 that under the latency constraints as the top-tier models to approximate the expectation term. The top 20 subnets are rapidly identified through the use of an accuracy lookup-table and a quantized latency predictor (Sec. 4.5)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 316, + 254, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 254, + 328 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 254, + 328 + ], + "type": "text", + "content": "4.3. Elastic Stage and Problem Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 330, + 287, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 287, + 402 + ], + "type": "text", + "content": "We observe that existing two-stage NAS adopt a chain-structured search space [4, 41, 35], which can be factorized as a sequence of STEM, HEAD and " + }, + { + "bbox": [ + 46, + 330, + 287, + 402 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 330, + 287, + 402 + ], + "type": "text", + "content": " searchable stages. Each stage defines a range of configurations " + }, + { + "bbox": [ + 46, + 330, + 287, + 402 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 330, + 287, + 402 + ], + "type": "text", + "content": " (e.g., kernel size, channel width, depth) for a specific block type " + }, + { + "bbox": [ + 46, + 330, + 287, + 402 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 330, + 287, + 402 + ], + "type": "text", + "content": ", and allows NAS to find the optimal architecture settings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 402, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 450 + ], + "type": "text", + "content": "Elastic stage. Without loss of generality, we define a stage structure in a search space as elastic stage " + }, + { + "bbox": [ + 46, + 402, + 287, + 450 + ], + "type": "inline_equation", + "content": "E_{b,c}" + }, + { + "bbox": [ + 46, + 402, + 287, + 450 + ], + "type": "text", + "content": ", which has elastic configurations " + }, + { + "bbox": [ + 46, + 402, + 287, + 450 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 402, + 287, + 450 + ], + "type": "text", + "content": " for a fixed block " + }, + { + "bbox": [ + 46, + 402, + 287, + 450 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 402, + 287, + 450 + ], + "type": "text", + "content": ". Suppose a search space " + }, + { + "bbox": [ + 46, + 402, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 46, + 402, + 287, + 450 + ], + "type": "text", + "content": " has " + }, + { + "bbox": [ + 46, + 402, + 287, + 450 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 402, + 287, + 450 + ], + "type": "text", + "content": " stages, it can be modularized as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 83, + 454, + 287, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 454, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 83, + 454, + 287, + 469 + ], + "type": "interline_equation", + "content": "\\mathcal {A} = S T E M \\circ E _ {b, c} ^ {1}, \\dots \\circ E _ {b, c} ^ {N} \\circ H E A D \\tag {2}", + "image_path": "cb6c2a86e9b3894a21dfde2aff3f0d84c2b05fca579eda757fd62383b361ebaa.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 473, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 545 + ], + "type": "text", + "content": "For instance, the search space in OQAT [32] and BatchQuant [1] can be factorized as 6 elastic stages and STEM (first Conv) and a classification head. Each elastic stage represents a set of configuration choices for the MBv3 block. Through the definition of elastic stage, we can simply use Eq. (2) to denote the contents of a model population." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 547, + 287, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 547, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 547, + 287, + 619 + ], + "type": "text", + "content": "Problem definition. Operator type " + }, + { + "bbox": [ + 46, + 547, + 287, + 619 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 547, + 287, + 619 + ], + "type": "text", + "content": " and configuration " + }, + { + "bbox": [ + 46, + 547, + 287, + 619 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 547, + 287, + 619 + ], + "type": "text", + "content": " are two crucial objectives when searching quantization-friendly search space. Through the definition of elastic stage, the task of space search can be simplified to find a search space with optimal elastic stages, which has a similar goal with neural architecture search. We formulate our problem as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 633, + 298, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 298, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 298, + 664 + ], + "type": "interline_equation", + "content": "\\mathcal {A} \\left(E _ {b, c} ^ {1} \\circ E _ {b, c} ^ {2} \\circ \\dots \\circ E _ {b, c} ^ {N}\\right) ^ {*} = \\underset {E _ {b, c} ^ {i} \\in \\mathcal {H} ^ {i}} {\\arg \\max } \\mathcal {Q} \\left(\\mathcal {A} \\left(E _ {b, c} ^ {1} \\circ E _ {b, c} ^ {2} \\circ \\dots \\circ E _ {b, c} ^ {N}\\right), T\\right) \\tag {3}", + "image_path": "6832ca8a2bc0227dcf46d1fe72403f7566be182f1badbcb40f6bfe0ab7dbdb2c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\cdot)" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " denotes the search space, and " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "E_{b,c}^{i}" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " elastic stage of " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\cdot)" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " denotes the hyperspace which covers all possible search spaces. " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " is the Q-T score for measuring a search space's quality." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 215, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 215, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 215, + 545, + 348 + ], + "type": "text", + "content": "As illustrated in Fig. 4(a), given the constraints " + }, + { + "bbox": [ + 304, + 215, + 545, + 348 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 215, + 545, + 348 + ], + "type": "text", + "content": " (i.e., a set of targeted quantized latency), SpaceEvo aims to find the optimal elastic stages " + }, + { + "bbox": [ + 304, + 215, + 545, + 348 + ], + "type": "inline_equation", + "content": "(E_{b,c}^{1} \\circ E_{b,c}^{2} \\circ \\dots \\circ E_{b,c}^{N})^{*}" + }, + { + "bbox": [ + 304, + 215, + 545, + 348 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 304, + 215, + 545, + 348 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 304, + 215, + 545, + 348 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 215, + 545, + 348 + ], + "type": "inline_equation", + "content": "N^{th}" + }, + { + "bbox": [ + 304, + 215, + 545, + 348 + ], + "type": "text", + "content": " stage for " + }, + { + "bbox": [ + 304, + 215, + 545, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^*" + }, + { + "bbox": [ + 304, + 215, + 545, + 348 + ], + "type": "text", + "content": " that has the maximum Q-T score: the top-tier quantized models can achieve best accuracy under the constraint " + }, + { + "bbox": [ + 304, + 215, + 545, + 348 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 215, + 545, + 348 + ], + "type": "text", + "content": ". The constraint " + }, + { + "bbox": [ + 304, + 215, + 545, + 348 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 215, + 545, + 348 + ], + "type": "text", + "content": " can be the quantized latency under any bits. In this work, we focus on the latency of INT8 quantized models, because INT8 quantization is widely supported by real-world edge devices and can achieve real speedup. Other lower bits typically require specific hardware (e.g., FPGA) to get the latency." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 357, + 457, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 357, + 457, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 357, + 457, + 369 + ], + "type": "text", + "content": "4.4. Searching the Search Space" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 371, + 545, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 545, + 395 + ], + "type": "text", + "content": "We now describe our evolutionary search algorithm that solves the problem in Eq. (3)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 396, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 479 + ], + "type": "text", + "content": "Hyperspace design. Analogous to NAS, hyperspace " + }, + { + "bbox": [ + 304, + 396, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 304, + 396, + 545, + 479 + ], + "type": "text", + "content": " defines which search space a search algorithm might discover. Defining a hyperspace to cover many candidate search spaces for space search is a second-order problem for NAS, which can easily introduce high complexity. Fortunately, we can easily construct a large hyperspace through search space modularization in Eq. (2)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 479, + 545, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 527 + ], + "type": "text", + "content": "We construct a large hyperspace in Fig. 4, in which a search space can be encoded by " + }, + { + "bbox": [ + 304, + 479, + 545, + 527 + ], + "type": "inline_equation", + "content": "N = 6" + }, + { + "bbox": [ + 304, + 479, + 545, + 527 + ], + "type": "text", + "content": " sequential elastic stages along with STEM and HEAD. We search the following two dimensions for an elastic stage:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 536, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 545, + 620 + ], + "type": "text", + "content": "- Block (operator) type " + }, + { + "bbox": [ + 306, + 536, + 545, + 620 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 306, + 536, + 545, + 620 + ], + "type": "text", + "content": ": MBv1 [15], MBv2 [31], MBv3 [14], residual bottleneck [13], residual bottleneck with SE, FusedMB [33] and FusedMB with SE. Conv is the major operator in residual bottleneck and FusedMB, thus they are quantization-friendly blocks; the efficiency of MB blocks relies on the device. For example, DWConv and SE are less quantization-efficient on Intel CPU." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 629, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 629, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 629, + 545, + 712 + ], + "type": "text", + "content": "- Output channel width list " + }, + { + "bbox": [ + 306, + 629, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\text{cout}" + }, + { + "bbox": [ + 306, + 629, + 545, + 712 + ], + "type": "text", + "content": ". In Sec. 3, we observe that quantized models can better utilize hardware under a larger channel number setting. However, directly increasing the channel numbers will also lead to longer latency. Therefore, we search the optimal stage-wise channel width list " + }, + { + "bbox": [ + 306, + 629, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\text{cout}" + }, + { + "bbox": [ + 306, + 629, + 545, + 712 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 306, + 629, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\{w_{\\min}^*, \\dots, w_{\\max}^*\\}" + }, + { + "bbox": [ + 306, + 629, + 545, + 712 + ], + "type": "text", + "content": ", which provides better channel width choices for final INT8 model search." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5823" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 287, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 287, + 121 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 287, + 121 + ], + "type": "text", + "content": "Specifically, as described in Fig. 4(b), we define a wide range of " + }, + { + "bbox": [ + 54, + 72, + 287, + 121 + ], + "type": "inline_equation", + "content": "[w_{min}, w_{max}]" + }, + { + "bbox": [ + 54, + 72, + 287, + 121 + ], + "type": "text", + "content": " by enlarging the channel widths in existing search spaces, and allow each elastic stage to choose a subset of " + }, + { + "bbox": [ + 54, + 72, + 287, + 121 + ], + "type": "inline_equation", + "content": "cout" + }, + { + "bbox": [ + 54, + 72, + 287, + 121 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 54, + 72, + 287, + 121 + ], + "type": "inline_equation", + "content": "[w_{min}, w_{max}]" + }, + { + "bbox": [ + 54, + 72, + 287, + 121 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 124, + 287, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 124, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 46, + 124, + 287, + 290 + ], + "type": "text", + "content": "Besides channel widths, other configuration dimensions (e.g., kernel size) also impact a model's quantized latency. However, searching all dimensions leads to a large amount of choices in one stage, which exponentially enlarges the hyperspace size. Fortunately, other dimensions usually have a small space (e.g., kernel size selects from " + }, + { + "bbox": [ + 46, + 124, + 287, + 290 + ], + "type": "inline_equation", + "content": "\\{3,5,7\\}" + }, + { + "bbox": [ + 46, + 124, + 287, + 290 + ], + "type": "text", + "content": "). It's easy to find the optimal value for a model in the final NAS process. Therefore, we follow existing practices to configure the choices of kernel size, depth, and expand ratios. Our final searched INT8 model architectures SEQnet suggest that the optimal quantization-friendly kernel sizes and expand ratios are chosen. For example, kernel size of " + }, + { + "bbox": [ + 46, + 124, + 287, + 290 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 46, + 124, + 287, + 290 + ], + "type": "text", + "content": " brings more INT8 latency speedups for DWConv, and it is the dominate kernel size choice in DWConv related blocks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 292, + 287, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 287, + 353 + ], + "type": "text", + "content": "Suppose that a stage has " + }, + { + "bbox": [ + 46, + 292, + 287, + 353 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 292, + 287, + 353 + ], + "type": "text", + "content": " choices of channel widths, there would be 7 (operator types) " + }, + { + "bbox": [ + 46, + 292, + 287, + 353 + ], + "type": "inline_equation", + "content": "\\times m" + }, + { + "bbox": [ + 46, + 292, + 287, + 353 + ], + "type": "text", + "content": " candidates for each stage. In total, for a typical search space with " + }, + { + "bbox": [ + 46, + 292, + 287, + 353 + ], + "type": "inline_equation", + "content": "N = 6" + }, + { + "bbox": [ + 46, + 292, + 287, + 353 + ], + "type": "text", + "content": " stages, the hyperspace has " + }, + { + "bbox": [ + 46, + 292, + 287, + 353 + ], + "type": "inline_equation", + "content": "\\sim 10^9" + }, + { + "bbox": [ + 46, + 292, + 287, + 353 + ], + "type": "text", + "content": " candidate search spaces, which is extremely large and poses challenges for efficient search." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 354, + 288, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 288, + 596 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 288, + 596 + ], + "type": "text", + "content": "Evolutionary space search. The structure of hyperspace is similar to a typical model search space in NAS [12], so we can easily apply existing NAS search algorithms. Taking this advantage, we leverage aging evolution [30] to search the large hyperspace. We first randomly initialize a population of " + }, + { + "bbox": [ + 46, + 354, + 288, + 596 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 354, + 288, + 596 + ], + "type": "text", + "content": " search spaces, where each sampled space is encoded as " + }, + { + "bbox": [ + 46, + 354, + 288, + 596 + ], + "type": "inline_equation", + "content": "(E_{b,c}^{1} \\circ E_{b,c}^{2} \\circ \\ldots \\circ E_{b,c}^{N})" + }, + { + "bbox": [ + 46, + 354, + 288, + 596 + ], + "type": "text", + "content": ". Each individual is rapidly evaluated with Q-T score. After this, evolution improves the initial population in mutation iterations. At each iteration, we sample " + }, + { + "bbox": [ + 46, + 354, + 288, + 596 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 354, + 288, + 596 + ], + "type": "text", + "content": " random candidates from the population and select the one with highest score as the parent. Then we alternately mutate the parent for block type and widths to generate two children search spaces. For instance, suppose the " + }, + { + "bbox": [ + 46, + 354, + 288, + 596 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 354, + 288, + 596 + ], + "type": "text", + "content": " stage " + }, + { + "bbox": [ + 46, + 354, + 288, + 596 + ], + "type": "inline_equation", + "content": "E_{b,c}^{i}" + }, + { + "bbox": [ + 46, + 354, + 288, + 596 + ], + "type": "text", + "content": " is selected for mutation, we first randomly modify its block type and produce " + }, + { + "bbox": [ + 46, + 354, + 288, + 596 + ], + "type": "inline_equation", + "content": "E_{b^*,c}^{i}" + }, + { + "bbox": [ + 46, + 354, + 288, + 596 + ], + "type": "text", + "content": " for child 1, then we mutate the widths and produce " + }, + { + "bbox": [ + 46, + 354, + 288, + 596 + ], + "type": "inline_equation", + "content": "E_{b,c^*}^{i}" + }, + { + "bbox": [ + 46, + 354, + 288, + 596 + ], + "type": "text", + "content": " for child 2. We evaluate their Q-T scores and add them to current population. The oldest two are removed for next iteration. After all iterations finish, we collect all the sampled space and select the one with best score as the final search space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 604, + 266, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 266, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 266, + 616 + ], + "type": "text", + "content": "4.5. Efficient Search Space Quality Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": "We now address the efficiency challenge caused by Q-T score evaluation. The most accurate evaluation is to get accuracy by training a supernet (search space) from scratch and measure latency on target device. However, it's impractical to conduct large-scale search due to the prohibitive cost. For example, it costs more than 10 days to train a supernet on 8 V100 GPUs [41]. To reduce the cost, we build an accurate INT8 latency predictor by nn-Meter [42], then" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 83, + 423, + 158 + ], + "blocks": [ + { + "bbox": [ + 307, + 83, + 423, + 158 + ], + "lines": [ + { + "bbox": [ + 307, + 83, + 423, + 158 + ], + "spans": [ + { + "bbox": [ + 307, + 83, + 423, + 158 + ], + "type": "image", + "image_path": "8b4d198bbb2577c9cb031571f045c5bc0c5107162a63744ba63b29e70dcc5923.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 163, + 545, + 196 + ], + "lines": [ + { + "bbox": [ + 304, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 545, + 196 + ], + "type": "text", + "content": "Figure 5. We adopt block-wise knowledge distillation to reduce search space quality evaluation cost. We add two linear transformation layers (Conv " + }, + { + "bbox": [ + 304, + 163, + 545, + 196 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 163, + 545, + 196 + ], + "type": "text", + "content": ") to match teacher's feature map widths." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 427, + 83, + 545, + 138 + ], + "blocks": [ + { + "bbox": [ + 305, + 69, + 545, + 79 + ], + "lines": [ + { + "bbox": [ + 305, + 69, + 545, + 79 + ], + "spans": [ + { + "bbox": [ + 305, + 69, + 545, + 79 + ], + "type": "text", + "content": "Step1: Quantize elastic stages via BKD Step2: Build quantized accuracy LUT" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 427, + 83, + 545, + 138 + ], + "lines": [ + { + "bbox": [ + 427, + 83, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 427, + 83, + 545, + 138 + ], + "type": "table", + "html": "
Elastic stageModel block encoding (when depth=1)NSR loss
EMBv2,[32,48]k3e3c48-56x560.39
EMBv3,[32,48]k5e6c32-40x400.37
", + "image_path": "d36163e3de82bb8f2f8d8bffeacfa366c0fe5d67270e5847b16c73eba7f92a57.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 427, + 140, + 542, + 148 + ], + "lines": [ + { + "bbox": [ + 427, + 140, + 542, + 148 + ], + "spans": [ + { + "bbox": [ + 427, + 140, + 542, + 148 + ], + "type": "text", + "content": "k: kernel size, e: expand ratio, c: width" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 209, + 493, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 493, + 219 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 493, + 219 + ], + "type": "text", + "content": "we introduce block-wise quantization scheme." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 220, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 327 + ], + "type": "text", + "content": "Block-wise knowledge distillation (BKD) is firstly proposed in DNA [22] and then further improved in DONNA [26]. It originally uses block-wise representation of existing models (teacher) to supervise a corresponding student model block. This technique can provide a relative accuracy ranking of all possible models without requiring them to be trained from scratch. In our work, we extend BKD to supervise the training of all elastic stages (each contains a large amount of blocks)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 328, + 546, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 546, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 546, + 495 + ], + "type": "text", + "content": "Fig. 5 illustrates the BKD process. In the first step, we use EfficientNet-B5 as the teacher, and separately train each elastic stage to mimic the behavior of corresponding teacher block by minimizing the NSR loss [26] between their output feature maps. Specifically, the " + }, + { + "bbox": [ + 304, + 328, + 546, + 495 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 328, + 546, + 495 + ], + "type": "text", + "content": " stage receives the output of " + }, + { + "bbox": [ + 304, + 328, + 546, + 495 + ], + "type": "inline_equation", + "content": "(i - 1)^{th}" + }, + { + "bbox": [ + 304, + 328, + 546, + 495 + ], + "type": "text", + "content": " teacher block as the input and is optimized to predict the output of " + }, + { + "bbox": [ + 304, + 328, + 546, + 495 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 328, + 546, + 495 + ], + "type": "text", + "content": " teacher block with NSR loss. Since an elastic stage contains many blocks with different channel widths, we add two learnable linear transformation layers at the input and output for each elastic stage to match teacher's feature map shape. Moreover, we adopt sandwich rule [41] to sample four paths to improve the training efficiency. Each elastic stage is firstly trained for 5 epochs and then performed 1 epoch LSQ+ [3] for INT8 quantization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "text", + "content": "Accuracy lookup table. In the second step, we construct a INT8 accuracy lookup table to reduce the evaluation cost. Specifically, we evaluate all possible blocks in each elastic stage and record their NSR losses on the validation set in the lookup table. The quantized loss of a model is estimated by summing up the NSR loss of all its blocks by rapidly looking up each elastic stages from the table. We inverse the measured loss to approximate the quantized accuracy for Q-T score evaluation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 603, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 603, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 603, + 545, + 639 + ], + "type": "text", + "content": "In our work, the BKD and lookup table construction can be sped up in a parallel way and finished in 1 day, which amounts a one-time cost before aging evolution search." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 650, + 377, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 650, + 377, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 650, + 377, + 662 + ], + "type": "text", + "content": "5. Evaluation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Setup. We evaluate our method on ImageNet-1k dataset [9] and two popular edge devices. The INT8 latency constraints are \\{8, 10, 15, 20, 25\\} ms for Intel CPU, and \\{15, 20, 25, 30, 35\\} ms for Pixel4. For each device, we search 5k search" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5824" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 68, + 289, + 192 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 289, + 192 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 289, + 192 + ], + "type": "image", + "image_path": "e0c9fd3e30b55f048ff1e0edc3feb90f788e8f0c42dd5a885663cfb30176019f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 194, + 287, + 227 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 287, + 227 + ], + "type": "text", + "content": "Figure 6. Best searched INT8 models with comparison to state-of-the-art NAS search spaces. Our searched spaces are proven to be the most quantization-friendly for the target device." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 238, + 286, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 286, + 262 + ], + "type": "text", + "content": "spaces in total and return the one with highest Q-T score. The population size " + }, + { + "bbox": [ + 46, + 238, + 286, + 262 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 238, + 286, + 262 + ], + "type": "text", + "content": " is 500 and sample size " + }, + { + "bbox": [ + 46, + 238, + 286, + 262 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 238, + 286, + 262 + ], + "type": "text", + "content": " is 125." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 262, + 287, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 454 + ], + "type": "text", + "content": "Once SpaceEvo discovers a quantization-friendly search space for the target device, we train a quantized-for-all supernet. We start by pretraining a full-precision supernet without quantizers on ImageNet for 360 epochs. We adopt the sandwich rule and inplace distillation in BigNAS [41]. Then, we perform quantization-aware training (QAT) on the trained supernet for 50 epochs, which follows the same training protocol (i.e., sandwich rule and inplace distillation). We use " + }, + { + "bbox": [ + 46, + 262, + 287, + 454 + ], + "type": "inline_equation", + "content": "\\mathrm{LSQ + }" + }, + { + "bbox": [ + 46, + 262, + 287, + 454 + ], + "type": "text", + "content": " as the QAT algorithm for better quantized accuracy. To derive INT8 model for deployment, we use the evolutionary search in OFA [4] to search " + }, + { + "bbox": [ + 46, + 262, + 287, + 454 + ], + "type": "inline_equation", + "content": "5\\mathrm{k}" + }, + { + "bbox": [ + 46, + 262, + 287, + 454 + ], + "type": "text", + "content": " models for various given INT8 latency constraints. We list out detailed training settings in supplementary materials. In the following, we refer to the two searched spaces as SpaceEvo@CPU and SpaceEvo@Pixel4, the searched model families are SEQnet@cpu and SEQnet@pixel4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 460, + 212, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 212, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 212, + 472 + ], + "type": "text", + "content": "5.1. The Effectiveness of SpaceEvo" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 474, + 286, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 286, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 286, + 641 + ], + "type": "text", + "content": "Comparison with SOTA search spaces. To demonstrate the high-performance of our searched spaces, we compare with prior art manually-designed search spaces including: (1) MobileNetV3 search space that is adopted in two-stage quantization NAS OQAT [32] and BatchQuant [1]; (2) ProxylessNAS and AttentiveNAS search spaces that achieve superior performance on mobile devices; and (3) ResNet50 search space proposed by OFA that is a handcraft quantization-friendly space on our two devices. For fair comparison, we use one supernet training and QAT receipt for all search spaces. We conduct evolutionary search to compare the best-searched models from each search space. We use the random seed of 0. For all experiments, search space is the only difference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "content": "Fig. 6 compares the best searched INT8 models from different search spaces. SpaceEvo@CPU and SpaceEvo@Pixel4 consistently deliver superior quantized models than state-of-the-art search spaces. Under the same latency, the best quantized models from SpaceEvo@cpu significantly surpass the existing state-of-the-art search" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 328, + 68, + 521, + 186 + ], + "blocks": [ + { + "bbox": [ + 328, + 68, + 521, + 186 + ], + "lines": [ + { + "bbox": [ + 328, + 68, + 521, + 186 + ], + "spans": [ + { + "bbox": [ + 328, + 68, + 521, + 186 + ], + "type": "image", + "image_path": "22e93fac7a91fbf3ff786cef6341a5323d607b9c8cb05c3ec4191f7576693b4c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 186, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 306, + 186, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 306, + 186, + 545, + 208 + ], + "type": "text", + "content": "Figure 7. Search space design under diverse INT8 latency constraints. SpaceEvo (6-25 ms) delivers superior tiny INT8 models." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 217, + 545, + 267 + ], + "blocks": [ + { + "bbox": [ + 306, + 217, + 545, + 267 + ], + "lines": [ + { + "bbox": [ + 306, + 217, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 306, + 217, + 545, + 267 + ], + "type": "image", + "image_path": "1f428a6bf1ddef5c56eb20aee1981392e7bed3e642e14bd69ee4b1db1ffc7c59.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 269, + 527, + 279 + ], + "lines": [ + { + "bbox": [ + 320, + 269, + 527, + 279 + ], + "spans": [ + { + "bbox": [ + 320, + 269, + 527, + 279 + ], + "type": "text", + "content": "Figure 8. Search cost measured on 8 Nvidia V100 GPUs." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 291, + 545, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 545, + 339 + ], + "type": "text", + "content": "spaces with " + }, + { + "bbox": [ + 304, + 291, + 545, + 339 + ], + "type": "inline_equation", + "content": "+0.7\\%" + }, + { + "bbox": [ + 304, + 291, + 545, + 339 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 291, + 545, + 339 + ], + "type": "inline_equation", + "content": "+3.8\\%" + }, + { + "bbox": [ + 304, + 291, + 545, + 339 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 291, + 545, + 339 + ], + "type": "inline_equation", + "content": "+0.4\\%" + }, + { + "bbox": [ + 304, + 291, + 545, + 339 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 291, + 545, + 339 + ], + "type": "inline_equation", + "content": "+3.2\\%" + }, + { + "bbox": [ + 304, + 291, + 545, + 339 + ], + "type": "text", + "content": " on Pixel4) higher accuracy. Moreover, our search space is the only one that is able to deliver superior quantized models under both extremely low (only " + }, + { + "bbox": [ + 304, + 291, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\sim 5\\mathrm{ms}" + }, + { + "bbox": [ + 304, + 291, + 545, + 339 + ], + "type": "text", + "content": ") and large latency constraints." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 340, + 546, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 546, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 546, + 448 + ], + "type": "text", + "content": "SpaceEvo under diverse latency constraints. We extensively study the effectiveness of SpaceEvo under different latency constraints. Specifically, we perform space search under two tight constraints of " + }, + { + "bbox": [ + 304, + 340, + 546, + 448 + ], + "type": "inline_equation", + "content": "\\{10, 15, 20, 25, 30\\}\\mathrm{ms}" + }, + { + "bbox": [ + 304, + 340, + 546, + 448 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 340, + 546, + 448 + ], + "type": "inline_equation", + "content": "\\{6, 10, 15, 20, 25\\}\\mathrm{ms}" + }, + { + "bbox": [ + 304, + 340, + 546, + 448 + ], + "type": "text", + "content": " on Pixel4. The results are shown in Fig. 7. Our proposed method can handle the diverse latency requirements and produce high-quality spaces. As expected, the searched spaces under " + }, + { + "bbox": [ + 304, + 340, + 546, + 448 + ], + "type": "inline_equation", + "content": "10 - 30\\mathrm{ms}" + }, + { + "bbox": [ + 304, + 340, + 546, + 448 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 340, + 546, + 448 + ], + "type": "inline_equation", + "content": "6 - 25\\mathrm{ms}" + }, + { + "bbox": [ + 304, + 340, + 546, + 448 + ], + "type": "text", + "content": " have much more low-latency quantized models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "content": "To further verify the effectiveness of these low-latency models, we compare with existing SOTA tiny models. Significantly, even under the extremely low latency constraints of 6-25 ms, our searched space delivers very competitive tiny quantized models. Compared to the smallest model ShuffleNetV2x0.5, we can achieve " + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "inline_equation", + "content": "+10.1\\%" + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "content": " higher accuracy under the same latency of 4.3 ms." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "text", + "content": "Search cost. As depicted in Fig. 8, our algorithm, SpaceEvo, is designed to be lightweight and suitable for real-world usage, requiring only 25 GPU hours to search a space of 5000 iterations. This remarkable speed is mainly due to our block-wise search space quantization scheme, which significantly reduces the cost of search space quality evaluation. In comparison, Fig. 8 demonstrates that training each search space from scratch without this scheme would consume an impractical 1200k GPU hours." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 651, + 541, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 541, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 541, + 662 + ], + "type": "text", + "content": "5.2. The Effectiveness of Discovered INT8 Models" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "In this section, we demonstrate that our searched spaces deliver state-of-the-art quantized models. We compare with two strong baselines: (1) prior art manually-designed and NAS-searched models; and (2) quantization-aware NAS. For" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5825" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 91, + 286, + 542 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 287, + 90 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 287, + 90 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 287, + 90 + ], + "type": "text", + "content": "Table 1. ImageNet results compared with SOTA quantized models on two devices. *: latency compared to FP32 inference." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 91, + 286, + 542 + ], + "lines": [ + { + "bbox": [ + 47, + 91, + 286, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 286, + 542 + ], + "type": "table", + "html": "
(a) Results on the Intel VNNI CPU with onnxruntime
ModelAcc% INT8CPU latencyFP32 FLOPs
INT8speedup*
MobileNetV3Small66.34.4 ms1.1×67.456M
SEQnet@cpu-A074.74.4 ms2.0×74.8163M
MobileNetV271.47.3 ms2.2×72.0300M
ProxylessNAS-R74.68.8 ms1.8×74.6320M
OQAT-8bit74.89.8 ms1.8×75.2214M
MobileNetV3Large74.510.3 ms1.5×75.2219M
OFA (#25)75.611.2 ms1.5×76.4230M
SEQnet@cpu-A177.48.8 ms2.4×77.5358M
APQ-8bit73.615.0 ms1.5×73.6297M
AttentiveNAS-A076.115.1 ms1.4×77.3203M
OQAT-8bit76.314.9 ms1.7×76.7316M
EfficientNet-B076.718.1 ms1.6×77.3390M
SEQnet@cpu-A278.514.1 ms2.4×78.8638M
APQ-8bit74.920.0 ms1.5×75.0393M
OQAT-8bit76.919.5 ms1.6×77.3405M
AttentiveNAS-A177.222.4 ms1.4×78.4279M
AttentiveNAS-A277.522.5 ms1.3×78.8317M
SEQnet@cpu-A379.518.9 ms2.6×79.6981M
FBNetV2-L175.825.0 ms1.2×77.2325M
FBNetV3-A78.227.7 ms1.3×79.1357M
SEQnet@cpu-A480.024.4 ms2.4×80.11267M
(b) Results on the Google Pixel 4 with TFLite
ModelAcc% INT8Pixel4 Latency|Acc% FP32FLOPs
INT8speedup*
MobileNetV3Small66.36.4 ms1.3×67.456M
SEQnet@pixel4-A073.65.9 ms2.1×73.7107M
MobileNetV271.416.5 ms1.9×72.0300M
ProxylessNAS-R74.618.4 ms1.8×74.6320M
MobileNetV3Large74.515.7 ms1.5×75.2219M
APQ-8bit74.614.9 ms2.0×74.4340M
OFA (#25)75.614.8 ms1.7×76.4230M
OQAT-8bit75.815.2 ms1.9×76.2287M
AttentiveNAS-A076.115.2 ms2.0×77.3203M
SEQnet@pixel4-A177.614.7 ms2.2×77.7274M
APQ-8bit75.120.0 ms1.9×75.1398M
OQAT-8bit76.520.4 ms1.8×76.8347M
AttentiveNAS-A177.221.1 ms2.0×78.4279M
AttentiveNAS-A277.522.7 ms2.0×78.8317M
SEQnet@pixel4-A278.319.4 ms2.3×78.4402M
FBNetV2-L175.826.7 ms1.5×77.2325M
OQAT-8bit77.029.9 ms1.7×77.2443M
FBNetV3-A78.230.5 ms1.5×79.1357M
SEQnet@pixel4-A379.530.8 ms2.1×79.5591M
EfficientNet-B076.736.4 ms1.7×77.3390M
SEQnet@pixel4-A479.935.5 ms2.2×80.0738M
", + "image_path": "92ba970c0b26a37c35a9d9bc659cf5c7452ae61cea13b36bc648b303520236c6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 553, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 287, + 696 + ], + "type": "text", + "content": "baseline (1), we collect official pre-trained FP32 checkpoints and conduct LSQ+ QAT to get the quantized accuracy. The hyperparameter settings follow the original LSQ+ paper, except that we set a larger epoch of 10 to achieve better accuracy. The latency numbers are measured on our devices. For (2), we compare with strong baselines including APQ [37] and OQAT [32]. Specifically, we limit APQ to search for the fixed 8bit (INT8) models. Since OQAT has no 8bit supernet checkpoint, we follow the official source code and conduct supernet QAT for 50 epochs. The final INT8 models are searched under the same INT8 latency predictors for fair comparison." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "content": "Results. Table 1 summarizes comparison results. Re" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 329, + 69, + 521, + 183 + ], + "blocks": [ + { + "bbox": [ + 329, + 69, + 521, + 183 + ], + "lines": [ + { + "bbox": [ + 329, + 69, + 521, + 183 + ], + "spans": [ + { + "bbox": [ + 329, + 69, + 521, + 183 + ], + "type": "image", + "image_path": "336c065d35c53d8fbd1aa4a5c588f7fb4dbd983be893ef492602f998a15cf25b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 184, + 545, + 206 + ], + "lines": [ + { + "bbox": [ + 306, + 184, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 306, + 184, + 545, + 206 + ], + "type": "text", + "content": "Figure 9. Q-T score effectiveness (Kendall's " + }, + { + "bbox": [ + 306, + 184, + 545, + 206 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 306, + 184, + 545, + 206 + ], + "type": "text", + "content": ") on ranking search space quality. We achieve a high space ranking correlation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 258, + 544, + 312 + ], + "blocks": [ + { + "bbox": [ + 306, + 215, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 306, + 215, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 306, + 215, + 545, + 258 + ], + "type": "text", + "content": "Table 2. Different space search methods and their best resulting quantized models on Pixel4. Baseline is a SOTA mobile-friendly AttentiveNAS space. *: the search dimension use the same settings in AttentiveNAS." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 258, + 544, + 312 + ], + "lines": [ + { + "bbox": [ + 307, + 258, + 544, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 258, + 544, + 312 + ], + "type": "table", + "html": "
MethodOpWidthBest quantized models
10ms15ms20ms30ms36ms
Baseline--76.677.279.079.5
SpaceEvo-opsearchfix*75.076.677.878.678.8
SpaceEvo-widthfix*search75.477.478.079.179.5
SpaceEvosearchsearch75.777.678.379.579.9
", + "image_path": "93c76b942af3b77b0a068b85057dea1cff9908612d848c4259b6bf8d346ec8e4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "text", + "content": "markably, our searched model family, SEQnet significantly outperform SOTA efficient models and quantization-aware NAS searched models, with higher INT8 quantized accuracy, lower INT8 latency and better speedups. Without finetuning, our tiny models SEQnet@cpu-A0 and SEQnet@pixel4-A0 achieve " + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "inline_equation", + "content": "74.7\\%" + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "inline_equation", + "content": "73.6\\%" + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "text", + "content": " top1 accuracy on ImageNet, which is " + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "inline_equation", + "content": "8.4\\%" + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "inline_equation", + "content": "7.3\\%" + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "text", + "content": " higher than MobileNetV3-Small (56M FLOPs) while maintaining the same level quantized latency. For larger models, SEQnet@cpu-A4 " + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "inline_equation", + "content": "(80.0\\%)" + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "text", + "content": " outperforms FBNetV3-A with " + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "text", + "content": " higher accuracy while runs 3.3ms faster. In particular, to achieve the same level accuracy (i.e., around " + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "inline_equation", + "content": "77.2\\%" + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "text", + "content": "), AttentiveNAS-A1 has 22.4ms latency while SEQnet@cpu-A1 " + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "inline_equation", + "content": "(77.4\\%)" + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "text", + "content": " only needs 8.8 ms " + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "inline_equation", + "content": "(2.6\\times" + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "text", + "content": " faster). More importantly, our searched models can better utilize the INT8 hardware optimizations: the latency speedups compared to full-precision inference are all larger than " + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 304, + 323, + 545, + 538 + ], + "type": "text", + "content": ", and this leaves room to search large-size models with higher accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 553, + 400, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 553, + 400, + 566 + ], + "spans": [ + { + "bbox": [ + 306, + 553, + 400, + 566 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Q-T score effectiveness. Q-T score is crucial as it guides the space evolution process. To evaluate its effectiveness, we randomly sample 30 search spaces, and measure the rank correlation (Kendall's " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ") between their Q-T score and their actual Pareto-frontier models' accuracies. Specifically, we use Intel CPU as the test device and set a same latency constraints of " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{8, 10, 15, 20, 25\\}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " ms. For each sampled space, we train it from scratch for 50 epochs, and conduct evolutionary search to get the Pareto-frontier models' accuracies. As shown in Fig. 9, the Kendall's " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " between the Q-T score and the actual Pareto-frontier models' accuracies is 0.8, which indicates a very high rank correlation." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5826" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 287 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 287 + ], + "type": "text", + "content": "Ablation study on two search dimensions. In Sec. 3, we conclude that operator type and configuration are two key factors impacting INT8 latency, which serves as the two search objectives of SpaceEvo. To verify the effectiveness, we create two strong baselines based on the SOTA edge-regime AttentiveNAS search space: (1) SpaceEvo-op: we fix each elastic stage's width to AttentiveNAS space, then allow each elastic stage to search for the optimal operator; and (2) SpaceEvo-width: we fix all elastic stages' block types to AttentiveNAS space, then search for the optimal width. Table 2 reports the space comparison between different search methods on the Pixel4. By searching both operator type and width, SpaceEvo finds the optimal search space where its best searched quantized models achieve the highest accuracy under all latency constraints. Moreover, even searching for one dimension, SpaceEvo-op and SpaceEvo-width outperform the manually-designed AttentiveNAS space under small latency constraints." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 288, + 289, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 288, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 288, + 289, + 456 + ], + "type": "text", + "content": "Search space design implications. We now summarize our learned experience and implications for designing quantization-friendly search spaces. We notice that the searched spaces show different preferences when targeting different devices: (i) All stages should use much wider channel widths compared to existing manually-designed spaces on the cpu, while only early stages prefer wider channels on Pixel 4. (ii) Since SE and Swish are INT8 latency-friendly on mobile phones, so our auto-generated search spaces for Pixel4 have many MBv3 stages. On Intel CPU, INT8 quantization slows down SE, Hardwish, and Swish, making FusedMB and MBv2 the priority for search spaces, with only the last two stages using MBv3. The details are provided in supplementary." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 466, + 119, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 119, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 119, + 478 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 481, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 481, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 481, + 287, + 578 + ], + "type": "text", + "content": "In this paper, we introduced SpaceEvo, the first to automatically design a quantization-friendly space for target device, which delivers superior INT8 quantized models with SOTA efficiency on real-world edge devices. Extensive experiments on two popular devices demonstrate its effectiveness compared to prior art manual-designed search spaces. We plan to apply SpaceEvo to other hardware efficiency such as energy-efficient search space design in the future." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 588, + 106, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 106, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 106, + 600 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 607, + 286, + 708 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 52, + 607, + 286, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 607, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 52, + 607, + 286, + 640 + ], + "type": "text", + "content": "[1] Haoping Bai, Meng Cao, Ping Huang, and Jiulong Shan. Batchquant: Quantized-for-all architecture search with robust quantizer. In NeurIPS, 2021. 1, 3, 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 642, + 286, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 642, + 286, + 685 + ], + "spans": [ + { + "bbox": [ + 52, + 642, + 286, + 685 + ], + "type": "text", + "content": "[2] Ron Banner, Yury Nahshan, and Daniel Soudry. Post training 4-bit quantization of convolutional networks for rapid-deployment. In Advances in Neural Information Processing Systems, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 686, + 286, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 686, + 286, + 708 + ], + "spans": [ + { + "bbox": [ + 53, + 686, + 286, + 708 + ], + "type": "text", + "content": "[3] Yash Bhalgat, Jinwon Lee, Markus Nagel, Tijmen Blankevoort, and Nojun Kwak. Lsq+: Improving low-bit" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 711 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "quantization through learnable offsets and better initialization. In CVPRW, 2020. 1, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 545, + 129 + ], + "type": "text", + "content": "[4] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. In ICLR, 2020. 2, 3, 5, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 130, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 130, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 312, + 130, + 545, + 173 + ], + "type": "text", + "content": "[5] Han Cai, Ligeng Zhu, and Song Han. ProxylessNAS: Direct neural architecture search on target task and hardware. In International Conference on Learning Representations, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 175, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 175, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 312, + 175, + 545, + 228 + ], + "type": "text", + "content": "[6] Zhaowei Cai and Nuno Vasconcelos. Rethinking differentiable search for mixed-precision neural networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2346–2355. Computer Vision Foundation / IEEE, 2020. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 230, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 230, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 312, + 230, + 545, + 273 + ], + "type": "text", + "content": "[7] Minghao Chen, Kan Wu, Bolin Ni, Houwen Peng, Bei Liu, Jianlong Fu, Hongyang Chao, and Haibin Ling. Searching the search space of vision transformers. In NeurIPS, 2021. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 274, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 274, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 312, + 274, + 545, + 329 + ], + "type": "text", + "content": "[8] Yuanzheng Ci, Chen Lin, Ming Sun, Boyu Chen, Hongwen Zhang, and Wanli Ouyang. Evolving search space for neural architecture search. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6639-6649, 2021. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 331, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 331, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 312, + 331, + 545, + 396 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248-255. IEEE Computer Society, 2009. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 398, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 398, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 398, + 545, + 430 + ], + "type": "text", + "content": "[10] Steven K. Esser, Jeffrey L. McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S. Modha. Learned step size quantization. In ICLR, 2020. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 432, + 446, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 446, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 446, + 442 + ], + "type": "text", + "content": "[11] Google. Tensorflow lite, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 444, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 444, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 444, + 545, + 487 + ], + "type": "text", + "content": "[12] Zichao Guo, Xiangyu Zhang, Haoyuan Mu, Wen Heng, Zechun Liu, Yichen Wei, and Jian Sun. Single path one-shot neural architecture search with uniform sampling. In ECCV, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 488, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 488, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 545, + 532 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 533, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 545, + 577 + ], + "type": "text", + "content": "[14] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, Quoc V. Le, and Hartwig Adam. Searching for mobilenetv3. In ICCV, 2019. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 578, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 545, + 632 + ], + "type": "text", + "content": "[15] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. CoRR, abs/1704.04861, 2017. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 634, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 634, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 308, + 634, + 545, + 666 + ], + "type": "text", + "content": "[16] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 668, + 545, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 668, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 308, + 668, + 545, + 711 + ], + "type": "text", + "content": "[17] Yiming Hu, Yuding Liang, Zichao Guo, Ruosi Wan, Xiangyu Zhang, Yichen Wei, Qingyi Gu, and Jian Sun. Angle-based search space shrinking for neural architecture search. In ECCV. 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5827" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 666 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 94 + ], + "type": "text", + "content": "[18] Intel. Intel 64 and ia-32 architectures software developer's manual, 2022. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 160 + ], + "type": "text", + "content": "[19] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 287, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 287, + 226 + ], + "type": "text", + "content": "[20] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2704-2713, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 227, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 227, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 48, + 227, + 287, + 258 + ], + "type": "text", + "content": "[21] Sumin Kim, Gunju Park, and Youngmin Yi. Performance evaluation of int8 quantized inference on mobile gpus. IEEE Access, 9:164245-164255, 2021. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 259, + 287, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 259, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 48, + 259, + 287, + 314 + ], + "type": "text", + "content": "[22] Changlin Li, Jiefeng Peng, Liuchun Yuan, Guangrun Wang, Xiaodan Liang, Liang Lin, and Xiaojun Chang. Blockwisely supervised neural architecture search with knowledge distillation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, 2020. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 315, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 315, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 48, + 315, + 287, + 367 + ], + "type": "text", + "content": "[23] Xiang Li, Chen Lin, Chuming Li, Ming Sun, Wei Wu, Junjie Yan, and Wanli Ouyang. Improving one-shot nas by suppressing the posterior fading. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 369, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 369, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 48, + 369, + 287, + 402 + ], + "type": "text", + "content": "[24] Christos Louizos, Mathias Reisser, Tijmen Blankevoort, Efstratos Gavves, and Max Welling. Relaxed quantization for discretized neural networks. In ICLR, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 403, + 189, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 403, + 189, + 412 + ], + "spans": [ + { + "bbox": [ + 48, + 403, + 189, + 412 + ], + "type": "text", + "content": "[25] Microsoft. onnxruntime, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 413, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 413, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 48, + 413, + 287, + 456 + ], + "type": "text", + "content": "[26] Bert Moons, Parham Noorzad, Andrii Skliar, Giovanni Mariani, Dushyant Mehta, Chris Lott, and Tijmen Blankevoort. Distilling optimal neural networks: Rapid search in diverse spaces. In ICCV, 2021. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 457, + 287, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 457, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 48, + 457, + 287, + 489 + ], + "type": "text", + "content": "[27] Markus Nagel, Mart van Baalen, Tijmen Blankevoort, and Max Welling. Data-free quantization through weight equalization and bias correction. In ICCV, 2019. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "text", + "content": "[28] Asaf Noy, Niv Nayman, Tal Ridnik, Nadav Zamir, Sivan Doveh, Itamar Friedman, Raja Giryes, and Lihi Zelnik. Asap: Architecture search, anneal and prune. In Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, pages 493-503, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "type": "text", + "content": "[29] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In CVPR, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "type": "text", + "content": "[30] Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image classifier architecture search. In AAAI, 2019. 2, 4, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 611, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 287, + 666 + ], + "type": "text", + "content": "[31] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 650 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 125 + ], + "type": "text", + "content": "[32] Mingzhu Shen, Feng Liang, Ruihao Gong, Yuhang Li, Chuming Li, Chen Lin, Fengwei Yu, Junjie Yan, and Wanlin Ouyang. Once quantization-aware training: High performance extremely low-bit architecture search. In ICCV, 2021. 1, 3, 5, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 127, + 545, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 127, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 307, + 127, + 545, + 170 + ], + "type": "text", + "content": "[33] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In Proceedings of the 38th International Conference on Machine Learning, pages 10096-10106. PMLR, 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 172, + 545, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 236 + ], + "type": "text", + "content": "[34] Xiaohu Tang, Shihao Han, Li Lyna Zhang, Ting Cao, and Yunxin Liu. To bridge neural network design and real-world performance: A behaviour study for neural networks. In A. Smola, A. Dimakis, and I. Stoica, editors, Proceedings of Machine Learning and Systems, volume 3, pages 21-37, 2021. 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 238, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 238, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 238, + 545, + 282 + ], + "type": "text", + "content": "[35] Dilin Wang, Meng Li, Chengyue Gong, and Vikas Chandra. Attentivenas: Improving neural architecture search via attentive sampling. In Conference on Computer Vision and Pattern Recognition, 2021. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 284, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 284, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 284, + 545, + 316 + ], + "type": "text", + "content": "[36] Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, and Song Han. Haq: Hardware-aware automated quantization with mixed precision. In CVPR, 2019. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 317, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 317, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 307, + 317, + 545, + 371 + ], + "type": "text", + "content": "[37] Tianzhe Wang, Kuan Wang, Han Cai, Ji Lin, Zhijian Liu, and Song Han. Apq: Joint search for network architecture, pruning and quantization policy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 1, 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 373, + 545, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 373, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 307, + 373, + 545, + 438 + ], + "type": "text", + "content": "[38] Xudong Wang, Li Lyna Zhang, Yang Wang, and Mao Yang. Towards efficient vision transformer inference: A first study of transformers on mobile devices. In Proceedings of the 23rd Annual International Workshop on Mobile Computing Systems and Applications, HotMobile '22, page 1-7. Association for Computing Machinery, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 440, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 440, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 440, + 545, + 483 + ], + "type": "text", + "content": "[39] Bichen Wu, Yanghan Wang, Peizhao Zhang, Yuandong Tian, Peter Vajda, and Kurt Keutzer. Mixed precision quantization of convnets via differentiable neural architecture search. 2019. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 484, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 484, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 307, + 484, + 545, + 517 + ], + "type": "text", + "content": "[40] Xin Xia, Xuefeng Xiao, Xing Wang, and Min Zheng. Progressive automatic design of search space for one-shot neural architecture search. In WACV, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 518, + 545, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 518, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 307, + 518, + 545, + 572 + ], + "type": "text", + "content": "[41] Jiahui Yu, Pengchong Jin, Hanxiao Liu, Gabriel Bender, Pieter-Jan Kindermans, Mingxing Tan, Thomas Huang, Xiaodan Song, Ruoming Pang, and Quoc Le. Bignas: Scaling up neural architecture search with big single-stage models. In ECCV, 2020. 3, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 574, + 545, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 574, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 307, + 574, + 545, + 650 + ], + "type": "text", + "content": "[42] Li Lyna Zhang, Shihao Han, Jianyu Wei, Ningxin Zheng, Ting Cao, Yuqing Yang, and Yunxin Liu. nn-meter: Towards accurate latency prediction of deep-learning model inference on diverse edge devices. In Proceedings of the 19th Annual International Conference on Mobile Systems, Applications, and Services, page 81-93, New York, NY, USA, 2021. ACM. 3, 6" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5828" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction/11a08db4-fb69-460a-a705-16e05f3f48ac_content_list.json b/2023/Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction/11a08db4-fb69-460a-a705-16e05f3f48ac_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..83d1f3eb361bc0b4aeb8a4cb1616a7a5ae689f31 --- /dev/null +++ b/2023/Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction/11a08db4-fb69-460a-a705-16e05f3f48ac_content_list.json @@ -0,0 +1,1741 @@ +[ + { + "type": "text", + "text": "Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction", + "text_level": 1, + "bbox": [ + 98, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jaesung Choe1,2 Christopher Choy1 Jaesik Park3 In So Kweon2 Anima Anandkumar1,4", + "bbox": [ + 122, + 191, + 844, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ NVIDIA $^{2}$ KAIST $^{3}$ Seoul National University $^{4}$ Caltech", + "bbox": [ + 250, + 222, + 720, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose an algorithm, 4DRegSDF, for the spacetime surface regularization to improve the fidelity of neural rendering and reconstruction in dynamic scenes. The key idea is to impose local rigidity on the deformable Signed Distance Function (SDF) for temporal coherency. Our approach works by (1) sampling points on the deformed surface by taking gradient steps toward the steepest direction along SDF, (2) extracting differential surface geometry, such as tangent plane or curvature, at each sample, and (3) adjusting the local rigidity at different timestamps. This enables our dynamic surface regularization to align 4D spacetime geometry via 3D canonical space more accurately. Experiments demonstrate that our 4DRegSDF achieves state-of-the-art performance in both reconstruction and rendering quality over synthetic and real-world datasets. https://4dregsdf.github.io/", + "bbox": [ + 75, + 318, + 473, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 580, + 209, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing 4D dynamic scenes from natural videos is one of the fundamental problems of vision tasks that are widely used in various fields, such as robotics, autonomous driving, virtual reality, and many more [6, 8, 14, 24, 32]. It requires an understanding of not just the geometry of scenes but also the texture, illumination, material properties, and refraction for photo-realistic rendering. However, because of the difficulty in reconstructing static 3D geometry alone, learning the dynamics jointly has been a challenge that most conventional methods decomposed the task into multiple stages of geometry reconstruction, motion representation, texture mapping, and illumination estimation [6, 15, 33, 49, 68]. Such sequential pipelines that first reconstruct geometry and learn motion separately are not robust to many failure modes of each stage, resulting in inaccurate reconstruction, which is one of the major bottlenecks for photo-realistic and geometry-oriented spacetime scene understanding.", + "bbox": [ + 75, + 606, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f2f082e520d86b5597eb0dd5ae4d27cee65dc41cde50e5739e1eff907f3a5cc.jpg", + "image_caption": [ + "Rendered color", + "Rendered normal" + ], + "image_footnote": [], + "bbox": [ + 501, + 286, + 696, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5b74dbd736e4a9097e03999c9d3fcfa1efda96139b06ec647c9a105bf73dbbe4.jpg", + "image_caption": [ + "Rendered color" + ], + "image_footnote": [], + "bbox": [ + 596, + 286, + 720, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1916d5d3d8ef87533c53c8a0f7ea53f0123cf419a3e3aa240910836db787914b.jpg", + "image_caption": [ + "Rendered normal" + ], + "image_footnote": [], + "bbox": [ + 699, + 286, + 892, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cb403aabc47411415bb512dca8c076d96f4b6ab198fe50bc37dd7e091b504c02.jpg", + "image_caption": [ + "Spacetime NeRF", + "Figure 1. We present an implicit surface regularization method that fully utilizes the properties of SDF for robust shape reconstruction and novel-view rendering in dynamic scenes." + ], + "image_footnote": [], + "bbox": [ + 500, + 426, + 689, + 502 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4618d5fdd3eb67f0eac60499df6af2d6063e1c99ce9833fe774d5034ce5d5c00.jpg", + "image_caption": [ + "4DRegSDF (ours)" + ], + "image_footnote": [], + "bbox": [ + 699, + 426, + 890, + 502 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent advances in neural implicit representations, however, made accurate 3D reconstruction with differentiable volumetric rendering possible [40,41,46,50]. Mildenhall et al. [41] use sinusoidal encoding on the position as an input to multi-layered perceptrons (MLPs) to learn appearance and geometry of the target static scene. Specifically for photo-realistic rendering, neural radiance fields [4, 41, 54, 61, 63, 71, 79] mostly use density and color as the main output. However, the density representation can result in artifacts such as floaters and blurry boundaries since the density representation has high degrees of freedom [67, 70]. For accurate 3D reconstruction, the Signed Distance Function (SDF) started to gain traction for its high fidelity 3D surface reconstruction [26, 47, 70, 76, 81]. One of the key component for SDF learning is the Eikonol loss [25] which regularizes the SDF to have valid gradient. This acts as a strong regularization not just on the surface but on the entirety of the empty space.", + "bbox": [ + 496, + 579, + 892, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, the 4D surface reconstruction from monocular RGB videos – which we reconstruct a dynamic scene – is still a highly ill-posed problem where we have to recon", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Work done during the internship at NVIDIA.", + "bbox": [ + 101, + 886, + 339, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17871", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "struct geometry and the dynamics of the scene simultaneously. Recent studies proposed to decompose the 4D space into a static geometry/texture (canonical model) and deformation to embed physical constraints into the 4D representation [51, 54]. The decomposition of the geometry could regularize the 4D space, but many papers use additional cues such as flow, or depth prediction [16, 21, 31, 34, 35, 69, 72, 78]. Still, the reconstruction is an ill-posed problem and it often fails for two reasons: (1) learning deformation is sensitive to topological variation and illumination change that often requires energy minimization terms for stable training. [28, 51, 60, 66]; (2) using the density representation – which has high degrees of freedom – could result in overfitting, floaters, and blurry reconstructions [29, 37, 44] as shown in Fig. 1. Thus, regularizing the spacetime domain is one of the most crucial parts of the 4D scene understanding in terms of geometry and graphics.", + "bbox": [ + 75, + 90, + 472, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose an algorithm for spacetime surface regularization that improves the fidelity of neural rendering and reconstruction in dynamic scenes. The key idea is to enforce the local rigidity on the deformable Signed Distance Function (SDF) for the 4D geometry representation. This allows us to (1) regularize the reconstruction using more explicit surface representation instead of fuzzy density measure [41]; (2) sample geometry efficiently for surface regularizations. Specifically, previous studies [37,47,70] show that Signed Distance Function can reconstruct geometries accurately with fewer images since learning the distance to the nearest surface to represent a surface is strong regularizations compared to the density function. In addition, the signed distance encodes not just distance but direction to the closest surface from gradients, thus allowing efficient surface sampling. Then, the sampled points are used to minimize the deformation energy and align 4D spacetime geometry via 3D canonical space.", + "bbox": [ + 75, + 351, + 472, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We use three datasets to evaluate our method: one synthetic dataset from Pumarola et al. [54], and two real datasets: Park et al. [52] and Gao et al. [22]. Our method achieves state-of-the-art performance on both geometry representation and appearance as visualized in Fig. 1. We summarize our contributions as follows:", + "bbox": [ + 76, + 627, + 470, + 717 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Extends the energy minimization in conventional 4D surface reconstruction studies into learning neural dynamic scene reconstruction.", + "- Regularize geometric properties of the 4D surface by the total variation of the curvature (TVC) and the absolute curvature of the SDF.", + "- Robust performance of rendering and reconstruction under the dynamic few-shot setup compared to recent implicit geometric regularization studies [20, 44]." + ], + "bbox": [ + 94, + 729, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 500, + 90, + 643, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conventional reconstruction and rendering. Video often contains time-varying properties, such as illumination change, geometry deformation, topology changes, and rigid or deformable motion. Such time-varying elements result in multi-view inconsistency among different frames and increase the level of difficulties of the video rendering task. For this issue, previous studies often require exhaustive multi-camera setup [5, 13, 48, 82] or synchronized multiple videos [2, 3]. None-learning based methods [1, 28, 60, 62] commonly assume the coarse geometry from RGB-D priors to introduce the energy regularization for the realistic geometry and photometric description of the target dynamic scenes. For instance, terzopoulos et al. [66] use elastic energy to deform the surface, and Slavcheva et al. [60] propose three typical energy minimization to prevent uncontrolled deformations. These ideas are related to the local rigidity to reconstruct plausible and geometricallymeaningful surface [1, 62] along the spatio-temporal domain.", + "bbox": [ + 496, + 114, + 890, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural implicit reconstruction and rendering. Recently, neural scene representation [9, 23, 40, 43, 45, 50, 53, 64] gets noticeable attention with their promising quality of implicit shape representations, such as occupancy or signed distance function. Not just geometry, the studies [36, 46, 58, 59] handle both geometry and appearance from multi-view images by neural implicit representation. In terms of the novelview synthesis task, Mildenhall et al. [41] demonstrates the impressive quality of rendered images. This paper takes a 5D radiance field (3D location and 2D viewing direction) to infer color and density. Then, it accumulates per-point inferences along each ray to determine pixel color. Such formulation makes it possible to train MLPs for learning geometry and appearance simultaneously. Based on this design, the series of following papers address the remaining problems of the pioneering paper [41], such as generalization [7, 57, 71], few-shot images [29, 44, 80], or surface reconstruction by SDF [70, 76, 77]. Nonetheless, these methods mostly focus on targeting static scenes without consideration of the moving objects or temporal change.", + "bbox": [ + 496, + 409, + 890, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural video rendering. More recently, several spacetime view synthesis studies [21,34,51,51,54,55,72] adopt neural implicit representations to render novel view synthesis from a video. Typically, these studies can be divided into two dominant strategies: deformable NeRFs [19,35,51,54] and prior-based video NeRFs [21, 34, 72]. Concurrent works [19, 35] propose fast deformable neural rendering methods by using voxel representations. Typically without relying on off-the-shelf information, such as monocular depth maps [34, 78], surface normal maps [81], or optical flow [35], we aim to render and reconstruct scenes purely from posed images.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17872", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Regularization for neural implicit representation. Relying on the nature of the geometry, a series of papers [29, 44, 55] present the effectiveness of the regularization in learning neural fields. Under the few-shot setup, Niemeyer et al. [44] regularize the unobserved viewpoints via normalizing flow models. Kim et al. [29] apply entropy loss to alleviate reconstruction inconsistency. Fu et al. [20] apply multi-view consistency for the static surface reconstruction. In light of the growing popularity of spatio-temporal surface reconstruction and neural rendering, this paper proposes a novel approach to regulating the spacetime surface through the use of 3D canonical space. We have observed that the direct application of the SDF to this deformation field can result in low-fidelity reconstruction and rendering due to ill-posed problems in dynamic scenes. To overcome this issue, we highlight the importance of our regularization to achieve precise reconstruction and rendering.", + "bbox": [ + 75, + 90, + 472, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 364, + 202, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural radiance field [41] allows fully differentiable learning of geometry and appearance by learning color and density from RGB images. Recent studies [42, 70] on surface reconstruction utilize Signed Distance Function (SDF) to represent scene geometry since The SDF representation has specific favorable characteristics that give it an edge for this particular objective.", + "bbox": [ + 75, + 390, + 468, + 497 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Differential surface geometry in SDF. The Signed Distance Function (SDF) stands as a predominant representation for implicit geometry. We can derive mathematical attributes on a crisp surface from SDF that could be challenging to derive from density-based representations: surface normal vector $\\mathbf{n} = \\frac{\\partial s}{\\partial\\mathbf{p}}$ , tangent plane projection matrix $P = I - \\mathbf{n}^{\\top}\\mathbf{n}$ , and surface curvature through the Hessian matrix $H = \\frac{\\partial^2s}{\\partial\\mathbf{p}^2}$ [73] Note that $I$ denotes the identity matrix, $s$ the SDF value, and $\\mathbf{p}$ a 3D coordinate. We can derive these mathematical attributes because SDF represents the distance to the closest surface at every point. We can force this property for implicit geometry by using the Eikonal loss [25], which forces the norm of gradients of SDF $(\\| \\mathbf{n}\\| _2 - 1)^2$ to be 1, and the loss has been widely used as an important regularization for 3D reconstruction methods [20,25,70,76].", + "bbox": [ + 75, + 503, + 472, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NeuS [70] uses the same volumetric rendering equation in NeRF [41]. However, the opaqueness $\\alpha$ is derived from the SDF value $s$ as $\\alpha = 1 - \\exp \\left(-\\int_{t_{i + 1}}^{t_i}\\rho (t)dt\\right)$ where $\\rho$ is the density function $\\rho = \\max \\left(\\frac{s - \\Delta s}{s + \\Delta s},0\\right)$ and $\\Delta s$ is the displacement of the SDF. Following NeRF [41], NeuS utilizes the volumetric rendering as:", + "bbox": [ + 75, + 758, + 468, + 854 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {C} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i} \\quad \\text {s . t .} \\quad T _ {i} = \\exp \\left(- \\sum_ {j = 1} ^ {i - 1} \\sigma_ {i} \\delta_ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 859, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{c}_i$ denotes the color of the $i$ -th sample point $\\mathbf{p}_i$ , $\\alpha_i$ is the opaqueness over distance $\\delta_i$ on the ray $\\mathbf{r} = \\{\\mathbf{p}_i\\}$ . $T_i$ is the cumulative transmittance along the ray. Based on this formulation, NeuS [70] is trained purely from posed images to describe the static scene surface using the SDF.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deformable NeRF [51, 54] is one way of novel-view synthesis methods in the 4D dynamic scenes. The idea is to decompose the 4D deformed space $[p; t]$ into the 3D canonical geometry $\\mathbf{p}_c$ and the temporal deformation $\\Delta \\mathbf{p}$ as:", + "bbox": [ + 496, + 172, + 890, + 234 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {c} = \\operatorname {D e f o r m} ([ \\mathbf {p}; t ]) = \\mathbf {p} + \\Delta \\mathbf {p}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 244, + 890, + 261 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Delta \\mathbf{p}(= h([p; t]))$ is the amount of space deformation, a function $h(\\cdot)$ represents a deformation field, and the deformation field transports a point $[\\mathbf{p}; t]$ to the canonical coordinate $\\mathbf{p}_c = [x_c, y_c, z_c]$ . Although progress has been made in the field of dynamic neural rendering, we address the importance of the spacetime surface regularization for learning appearance, geometry, and motion in dynamic scenes as visualized in Fig. 1.", + "bbox": [ + 496, + 268, + 890, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 402, + 589, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section presents our framework, 4DRegSDF. Our network represents the 4D spacetime domain as a combination of the static geometry in Signed Distance Function and the deformation field. Given a set of posed frames from a monocular RGB video, our network is trained to encode a dynamic scene (Sec. 4.1) and apply our spacetime surface regularization (Sec. 4.2). In contrast to previous video neural field papers that use priors such as depths or flows [21, 34, 35], our method relies only on posed image frames from a video as input to the system making the system simple and efficient.", + "bbox": [ + 496, + 429, + 890, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Deformable Signed Distance Function", + "text_level": 1, + "bbox": [ + 500, + 604, + 826, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a point $\\mathbf{p}$ at time $t$ within the deformed space and its viewing direction $\\mathbf{d}$ , our model $f$ predicts the color $\\mathbf{c}$ , the SDF value $s$ , and the deformation delta $\\Delta \\mathbf{p}$ . This can be expressed as:", + "bbox": [ + 496, + 627, + 890, + 686 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n[ s; \\Delta \\mathbf {p}; \\mathbf {c} ] = f ([ \\mathbf {p}; t ], \\mathbf {d}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 696, + 890, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As visualized in Fig. 3, we use the same volumetric rendering on the SDF (Eq. 1) [70] to get color value per each ray $\\mathbf{r}$ , and minimize the photometric loss $\\mathcal{L}_{\\mathrm{c}}$ as,", + "bbox": [ + 496, + 723, + 890, + 768 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c}} = \\left| C (\\mathbf {r}) - \\hat {C} (\\mathbf {r}) \\right| _ {1}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 771, + 890, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $C$ and $\\hat{C}$ are ground truth color and predicted color, respectively, along a query ray $\\mathbf{r}$ .", + "bbox": [ + 496, + 806, + 890, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. Spacetime surface regularization", + "text_level": 1, + "bbox": [ + 500, + 847, + 787, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reconstructing dynamic surfaces from videos presents challenges due to its inherent complexity, often leading to", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17873", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f6bf193fcb8a1a2a0af4d6acc03a83fea53a2228916f472de9c910deb58b3438.jpg", + "image_caption": [ + "Figure 2. Surface sampling for our spacetime surface regularization in three stages." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 315, + 214 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f34181a1533192b5e82bacae1d730e52e9892f601ebfc62bf1c6854b46b0f7a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 88, + 599, + 215 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/115b02b2fc9b314b2b6a8feeb9b8e40ee0ca95513cb2185f162d8914c5b0321a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 85, + 875, + 215 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8d07f33533eac7936dd5b99e7b9f3aabcc5a0c6fb7c07cb390cb99904a84369a.jpg", + "image_caption": [ + "Figure 3. Ray sampling for neural rendering [70]." + ], + "image_footnote": [], + "bbox": [ + 93, + 262, + 450, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "visual artifacts as shown in Fig. 6. Accordingly, designing a robust regularization is one of the critical problems for successful 4D reconstruction [1,51,60,62,66,73].", + "bbox": [ + 75, + 420, + 468, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we adopt three regularizations to enhance the learning of deformable Signed Distance Function from videos: 1. total variation of curvature, 2. the absolute curvature space-time, and 3. the Eikonal loss [25]. The first regularization aims to limit the amount of change in curvature induced by motion and the second regularization flattens out the 4D reconstruction and limits unnecessary kinks. The last regularization [25] enforces the gradient of SDF to be valid. First two regularizations are related to the properties of surface and can be computed only on the surface. We elaborate on the sampling strategy and how to augment the samples in continuous space and time, and lastly present the regularizations. Overall pipeline is visualized in Fig. 2.", + "bbox": [ + 75, + 467, + 468, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spacetime surface sampling. The most commonly used sampling in neural fields is ray casting where you select points along a ray for rendering. Many regularizations proposed for NeRF also use samples along a ray [20,29,37,44]. However, surface constraints we adopt require samples on the surface of objects no along rays. For this, we use the definition of the SDF for surface sampling.", + "bbox": [ + 75, + 672, + 468, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The SDF is defined as the (signed) distance to the closest surface from the query point. In static scenes, the surface normal indicates the direction of the closest surface [12, 39, 73]. However, in 4D dynamic scenes, time is another component to determine the closest surface. In our setup, our network $f(\\cdot)$ is trained to infer the SDF value $s$ from a query $[\\mathbf{p}; t]$ . Accordingly, the gradient step towards the closest surface naturally involves and spatial update as", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "Algorithm 1 Spacetime surface regularization \nInput: Randomly initialized 4D point $[\\mathbf{p}_i; t_i]$ . \nOutput: Total regularization $\\mathcal{L}_e$ . \n1: function 4DRegSDF([p_i, t_i]) \n2: [p, t] ← [p_i, t_i] \n3: for iterations do \n4: [p, t] ← StepGrad([p; t]) ▷ Eq. 5 \n5: end for \n6: pc ← Deform([p; t]) ▷ Eq. 2 \n7: t' ← Random(0, 1) ▷ Time augmentation \n8: p' ← InvDeform([pc; t']) ▷ Eq. 6 \n9: $\\mathcal{L}_e \\gets \\text{ComputeLoss}([\\mathbf{p}; t], [\\mathbf{p}'; t'])$ ▷ Eq. 9 \n10: return $\\mathcal{L}_e$ \n11: end function", + "bbox": [ + 500, + 263, + 890, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "well as the time update. Given randomly initialized point $p_i$ at the time $t_i$ , each query point travels along the spacetime SDF 'hills' to be located at the isosurface as:", + "bbox": [ + 498, + 513, + 890, + 558 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n[ \\mathbf {p}; t ] = \\operatorname {S t e p G r a d} ([ \\mathbf {p} _ {i}; t _ {i} ]) = [ \\mathbf {p} _ {i}; t _ {i} ] - s \\cdot \\frac {\\partial s}{\\partial [ \\mathbf {p} _ {i} ; t _ {i} ]}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 565, + 890, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\frac{\\partial s}{\\partial[\\mathbf{p};t]}$ is the gradient of the SDF value $s$ in terms of $[\\mathbf{p};t]$ . By doing so, we can regularize the 4D surface in a wide-spectrum time domain including unobserved timestamps within training samples. Thus, by iteratively conducting this operation above, we move the initial 4D point $[\\mathbf{p}_i; t_i]$ to $[\\mathbf{p}; t]$ closer to the isosurface having smaller SDF value, as described in L. 2-5 of Alg. 1.", + "bbox": [ + 496, + 608, + 890, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal augmentation. Given the surface samples $[\\mathbf{p};t]$ from Eq. 5, we enforce the local rigidity to the surface samples for temporal coherency. First, we augment the time component $t$ into randomly selected $t^{\\prime}$ as visualized in Fig. 2-(b). Specifically, we use the inverse-deformation network $g$ [35,45,74] which is formulated as:", + "bbox": [ + 496, + 724, + 890, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} ^ {\\prime} = \\operatorname {I n v D e f o r m} \\left(\\left[ \\mathbf {p} _ {c}; t ^ {\\prime} \\right]\\right) = g \\left(\\mathbf {p} _ {c}, t ^ {\\prime}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 827, + 890, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_c$ indicates the points at the canonical space from $[\\mathbf{p};t]$ (Eq. 3). $\\mathbf{p}'$ is a inverse-deformed point from $\\mathbf{p}_c$ at the augmented time $t'$ as stated in L. 6-8 of Alg. 1. Through our", + "bbox": [ + 498, + 854, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17874", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "temporal augmentation stage, we get a pair of two surface samples that correspond to each other, $([\\mathbf{p};t],[\\mathbf{p}';t'])$ . This pair will be used for evaluating spatio-temporal constraints the next section.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Regularization on the spacetime curvature. Inspired by the conventional 4D reconstruction studies [1,60,62,66], we regularize geometric properties of dynamic surfaces: the total variation of curvature (TVC) and the absolute curvature of SDF surface on continuous time. The TVC measures the change of the curvature at different time steps and by minimizing the change, we assumes that surface should deform as little as possible while satisfying video observations. On the contrary, the absolute curvature measures curvature values and thus regularizes the complexity of the reconstruction and movement by minimizing the distortion in space. For the TVC, we adopt the stretching loss [73] where it minimizes the difference of inner products of two tangent vectors at two different time steps: $v_{1}, v_{2}$ at $t$ and $v_{1}', v_{2}'$ at $t'$ . The tangent vector is a remainder of a vector on surface $v = P\\mathbf{p}$ where $P(= I - \\mathbf{nn}^{\\top})$ is the tangential projection matrix and $\\mathbf{n}(= \\frac{\\partial s}{\\partial\\mathbf{p}})$ is the surface normal vector at a point $\\mathbf{p}$ . Let $J_{1}, J_{2}$ be the Jacobians transforming $v_{1}, v_{2}$ to $v_{1}', v_{2}'$ respectively. Then, we compute the TVC as:", + "bbox": [ + 75, + 160, + 472, + 449 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left| v _ {1} ^ {\\top} v _ {2} - v _ {1} ^ {\\prime \\top} v _ {2} \\right| = \\left| \\mathbf {p} _ {1} ^ {\\top} P _ {1} ^ {\\top} P _ {2} \\mathbf {p} _ {2} - \\mathbf {p} _ {1} ^ {\\top} P _ {1} ^ {\\top} J _ {1} ^ {\\top} J _ {2} P _ {2} \\mathbf {p} _ {2} \\right|, (7) \\\\ \\propto \\left| P _ {1} ^ {T} \\left(I - J _ {1} ^ {T} J _ {2}\\right) P _ {2} \\right|, (8) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 460, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The absolute curvature of the 4D SDF is the L1 norm of the Hessian matrix $H = \\frac{\\partial^2 s}{\\partial \\mathbf{p}^2}$ . By minimizing the absolute curvature, the reconstruction removes unnecessary curvatures or kinks in the reconstructions making the surface as smooth as possible. The total regularization is:", + "bbox": [ + 75, + 517, + 468, + 594 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {e} (\\mathbf {p}, \\hat {\\mathbf {p}}) = \\| P ^ {\\top} (I - J _ {g} J _ {f}) P \\| _ {2} + \\| H \\| _ {1}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 606, + 468, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $J_{g}$ and $J_{f}$ are the Jacobians that transformed points via the deformation network $f(\\cdot)$ and the inverse deformation network $g(\\cdot)$ , respectively.", + "bbox": [ + 75, + 640, + 468, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We summarize our spacetime surface regularization in Algorithm 1. We achieve high-quality reconstruction results that are geometrically simple while they match video observations through neural rendering and minimizing the photometric loss. Finally, we train our network by combining the proposed losses as below,", + "bbox": [ + 75, + 686, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {c c} \\mathcal {L} _ {c c} + \\lambda_ {\\mathrm {e}} \\mathcal {L} _ {\\mathrm {e}} + \\lambda_ {\\mathrm {e i k}} \\mathcal {L} _ {\\mathrm {e i k}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 792, + 468, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{c} = 1$ , $\\lambda_{cc} = 0.01$ , $\\lambda_{e} = 0.001$ , and $\\lambda_{\\mathrm{eik}} = 0.01$ . $\\mathcal{L}_{\\mathrm{eik}}$ represents the Eikonal loss [25] following NeuS [70]. Also, we enforce cycle consistency loss in learning forward/inverse deformation fields by $\\mathcal{L}_{cc} = \\| \\mathbf{p} - g(\\mathbf{p}_c,t)\\| _1$ [35,45,74].", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 90, + 633, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Novel-view synthesis", + "text_level": 1, + "bbox": [ + 500, + 114, + 694, + 131 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a set of posed images, the novel-view synthesis aims to render images from viewpoints that are not in the part of the provided posed images. Since our 4DRegSDF also follows the neural rendering schemes, we compare the quality of rendered images with the recent spacetime NeRF papers on novel-view synthesis evaluation.", + "bbox": [ + 496, + 138, + 890, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We utilize the Dynamic NeRF Synthetic dataset provided by D-NeRF [54] and the real-world scenes from HyperNeRF dataset [52]. These datasets offer natural monocular videos with dynamic frames and continuous camera motion. We follow the official train/test split for training and evaluation over ours and related studies. Following the conventions [41, 51, 54], we evaluate the rendering performance through PSNR, SSIM, and LPIPS.", + "bbox": [ + 496, + 237, + 890, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison. We compare our method with a few state-of-the-art 4D NeRF papers, D-NeRF [54], TiNeuVox [19], which also do not rely on the prior information [21,34,35]. Also, we extend NeuS [70] for the 4D dynamic reconstruction task by adopting a deformation network from D-NeRF [54]. We denote this deformable NeuS as 'NeuS+D' and use it One of the concurrent works, DeVRF [35], also proposes a deformation-based dynamic neural rendering task. However, the paper requires additional multi-view static scenes for initialization so we did not include it in our experiment.", + "bbox": [ + 496, + 366, + 890, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When compared to existing research, our method achieves state-of-the-art performance among the recent spacetime NeRF methods as shown in Table 1 and Table 2. Many of these studies rely on view-dependent surface representations and the density function [41]. In contrast, our method utilizes the SDF representation, allowing us to more accurately represent geometry. This distinction influences the rendering quality. Notably, in terms of the qualitative reconstruction results, the previous state-of-the-art method struggles to accurately capture surface details as illustrated in Fig. 4 and Fig. 5. We suspect this limitation is rooted in the inherent issues of the density representation. As a result, many of the earlier spacetime NeRF studies place a heavy emphasis on the rendering quality. However, we claim that by imposing scene geometry constraints through our spacetime surface regularization, we can further improve the quality of the geometry under the precise geometry presentation, namely SDF.", + "bbox": [ + 496, + 532, + 890, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Scene reconstruction", + "text_level": 1, + "bbox": [ + 500, + 816, + 696, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. Dycheck dataset [22] is one of the few datasets that provide ground truth geometry information for dynamic scenes. Yet, the ground truth depth maps are only provided for the training data split and we could not evalu", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17875", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/85bea15b6b1d6b300a3a8c905b01a9065d6ac9e1929a336c648a9a351b984aed.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dynamic Synthetic NeRF dataset [54]
Bouncing ballsHell warriorHookJumping jacksLegoMutantStand upT-rex
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
NeRF [41]20.260.9113.510.8116.650.8418.280.8820.300.7920.310.9118.190.8924.490.93
D-NeRF [54]38.930.9825.020.9529.250.9632.800.9821.640.8331.290.9632.790.9831.750.97
TiNeuVox [19]40.730.9928.170.9731.450.9734.230.9825.020.9233.610.9835.430.9932.700.98
TorchNGP [65]40.440.9924.540.9431.460.9731.450.9525.070.9335.860.9935.290.9931.350.97
NeuS+D [70]36.980.9623.770.9428.090.9428.280.9324.170.9030.170.9430.990.9524.940.93
Ours40.890.9925.720.9532.200.9833.630.9925.100.9431.830.9633.600.9933.640.99
", + "bbox": [ + 81, + 88, + 893, + 183 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3649dbfbc24fc21efff7fa28e8c05e8f77b13e054ffbfce14dc83192fd5ab478.jpg", + "table_caption": [ + "Table 1. Rendering evaluation in the D-NeRF synthetic dataset. We use PSNR (↑) and SSIM (↑) for the metric." + ], + "table_footnote": [], + "table_body": "
HyperNeRF dataset [52]
Broom3D PrinterChickenPeel-bananaMEAN
PSNRMS-SSIMPSNRMS-SSIMPSNRMS-SSIMPSNRMS-SSIMPSNRMS-SSIM
NeRF [41]19.90.65320.70.78019.90.77720.00.76920.10.745
NV [36]17.70.62316.20.66517.60.61515.90.38016.90.571
Nerfies [51]19.20.56720.60.83026.70.94322.40.87222.20.803
TorchNGP [65]19.80.57121.10.84127.60.94621.10.85123.50.815
TiNeuVox [19]21.50.68622.80.84128.30.94724.40.87324.30.837
NeuS+D [70]21.90.68922.10.83925.90.93324.60.87523.60.834
Ours23.30.70223.90.85929.80.96525.40.88125.30.851
", + "bbox": [ + 147, + 212, + 823, + 325 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f1cc03519a5e90789c5e6be7c6e43bb3f49acf991e8416ab629e155bf0dd4220.jpg", + "table_caption": [ + "Table 2. Quantitative evaluation in the HyperNeRF dataset [52]." + ], + "table_footnote": [], + "table_body": "
Dycheck dataset [22]
CreeperHaru-sitPaper-windmillSriracha-tree
RMSE↓CD↓F1↑PSNR↑RMSE↓CD↓F1↑PSNR↑RMSE↓CD↓F1↑PSNR↑RMSE↓CD↓F1↑PSNR↑
TorchNGP [65]0.3400.05087.7817.320.5020.08176.3327.590.3110.05781.9823.30.3110.05281.4427.9
TiNeuVox [19]0.3450.04888.4216.690.6340.08075.1826.110.3130.05382.2924.960.3250.05781.5030.57
NeuS+D [70]0.3500.07174.1817.050.4910.11860.1327.880.2950.03891.3222.520.2830.04585.2326.83
Ours0.3230.06482.3617.570.4940.07778.5328.290.3100.05183.2424.260.2900.04292.2530.66
", + "bbox": [ + 80, + 352, + 893, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Evaluation in the Dycheck dataset [22] Note that CD (↓) means Chamfer Distance and F1 (↑) is F1-score.", + "bbox": [ + 140, + 431, + 820, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ate geometry reconstruction quality on the test split. Accordingly, we subsample the official train data to form a new train/test split. The new split consists of $75\\%$ training data and $25\\%$ test data per scene. For evaluation metrics, we follow the three geometric measures, RMSE, Chamfer Distance, and F1-score. RMSE is a widely-adopted metric for depthmap prediction evaluation [10, 17, 27, 56]. Chamfer distance [11, 18] measures the distance between the ground truth point cloud and the prediction. Meanwhile, the F1-score measures the 3D volume occupancy between the ground truth and the predictions [11, 30, 75].", + "bbox": [ + 75, + 462, + 468, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison. To evaluate geometric reconstruction quality, we evaluate on both view specific depth map prediction task and the 3D geometry reconstruction task. For depth map prediction task, we extract the depth map at the zero-crossing of SDF. For density-based NeRFs [19], we render depth maps by computing the expected depth. For the final metric, measure the RMSE in the neural volume rendering schemes [41, 70]. For 3D reconstruction evaluation, we follow the surface extraction pipeline provided by the official D-NeRF implementation for density-based methods [19, 54]. In our case, we extract mesh at the zero-crossings of SDF by using marching-cube algorithm [38] for 3D geometric evaluation. For evaluation metrics, we utilize the Chamfer Distance (CD ↓), F1-score (F1 ↑), and RMSE ↓.", + "bbox": [ + 75, + 640, + 468, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on this formulation, we calculate the 2D/3D geometric performance by ours and recent studies [19]. As", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "described in Table 3, our method still consistently outperforms the quality of 4D surface reconstruction tasks. We will state the detailed ablation study in the following section, however, in short, the dominant performance improvements are driven by the plausible surface representation by Signed Distance Function and the proposed surface regularization schemes. Finally, our 4DRegSDF successfully renders and reconstructs the 4D dynamic scenes thanks to the plausible geometry representation and learning schemes", + "bbox": [ + 496, + 462, + 890, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Ablation studies", + "text_level": 1, + "bbox": [ + 500, + 616, + 660, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we provide additional experiments to analyze each module of the 4DRegSDF. Note that additional ablation results are included in the supplementary material.", + "bbox": [ + 496, + 642, + 890, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of surface regularization. Surface regularization is the key to making 4D surfaces plausible shapes while learning deformation. When comparing with Fig. 6-(b) and (c), we found that $\\mathcal{L}_e$ affects the structure of the reconstructed table. It implies that our surface regularization takes a role in shaping plausible structures by aligning geometry between surfaces at different times.", + "bbox": [ + 496, + 696, + 890, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with different regularization schemes. We compare our surface regularization with previous regularization schemes, e.g. RegNeRF [44] and GeoNeuS [20], as in Table 5. Based on these results, we claim that our spacetime surface regularization is particularly promising compared to previous regularization strategies. Also, we pro", + "bbox": [ + 496, + 809, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17876", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ab5472bfe0c6ecd21415f8034d40141266e2ded41ef4381fb984417a345b6670.jpg", + "image_caption": [ + "No ground truth mesh" + ], + "image_footnote": [], + "bbox": [ + 81, + 99, + 658, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e4141a6296b9b8d91d8bee3c75c51be3537a251b90a938ae805b8af000d415b8.jpg", + "image_caption": [ + "TiNeuVox" + ], + "image_footnote": [], + "bbox": [ + 86, + 236, + 336, + 358 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e49c82ab67e45f1862ff172791628da47cc2d0c55b4aec9b9f811590af923da6.jpg", + "image_caption": [ + "4DRegSDF (ours)" + ], + "image_footnote": [], + "bbox": [ + 367, + 236, + 612, + 356 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8a7800dbc871e2302dd7558f46b2929d24a7aa1516c60d18f27e0147c4bde085.jpg", + "image_caption": [ + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 633, + 234, + 880, + 356 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc47b97938256ed6546117487c00ed6c5361f5dfc73f43f068df333a3f970ffc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 431, + 305, + 588 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/34221338b772edfbd17852710e2303653c0d292ec7dc2b42fa3b4b469f789bb4.jpg", + "image_caption": [ + "TiNeuVox", + "Figure 5. Qualitative results in HyperNeRF dataset [52]. We compare previous state-of-the-art method, TiNeuVox [19], NeuS [70] with deformation [54] and our 4DRegSDF. Left: rendered color images, right: rendered surface normal maps." + ], + "image_footnote": [], + "bbox": [ + 78, + 590, + 305, + 744 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e187cccd275bf3392be5cc5eae0489de5ce4b2237989aa0d60c99cc8aed3b408.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 431, + 540, + 588 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/48f037968eb65f45d15b702c358d3cbe19b0afd46f5759cbdc061fb7ac43a0c4.jpg", + "image_caption": [ + "NeuS + Deformation" + ], + "image_footnote": [], + "bbox": [ + 310, + 590, + 540, + 744 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a360fbe372e5aadff61eedc2646c56c3fb83cbb012b82972e11d7a493e38a795.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 431, + 769, + 588 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2237fa65faf0d54bd8d570093cfb88d461c44918c86630c731100bf4ce66c5af.jpg", + "image_caption": [ + "4DRegSDF (ours)" + ], + "image_footnote": [], + "bbox": [ + 545, + 590, + 769, + 744 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83b96ce0f298a6c21545436ad2bb2bcdc626510e573e78b8fc2bf971900ed3be.jpg", + "image_caption": [ + "Figure 4. Reconstruction results (mesh) with rendered images from our 4DRegSDF and previous state-of-the-art method, TiNeuVox [19]." + ], + "image_footnote": [], + "bbox": [ + 772, + 431, + 887, + 588 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c312ad9de4bacf42acc50785624e4bc67b38c2bb238f5f1ce22c39a41328dc15.jpg", + "image_caption": [ + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 772, + 590, + 887, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vide details of our surface regularization in Table 6. Note that we calculate avg. PSNR from HyperNeRF dataset [52], and avg. RMSE from DyCheck dataset [22].", + "bbox": [ + 75, + 816, + 468, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Deformation network design. Unlike deformable NeRF methods [51, 54] that sequentially process deformation and", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rendering (Eq. 2), our 4DRegSDF proposes to jointly learn deformation and SDF through a unified network $f$ (Eq. 3) as stated in Sec. 4.1. It turns out that our network design shows the better quality of surface normal maps as visualized in Fig. 6-(a) and (c). Such architecture design can", + "bbox": [ + 496, + 816, + 893, + 893 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17877", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6487d9e4362303f41d268f54cbe03f8643f510ea7ed12e20ede89646df7e6eec.jpg", + "image_caption": [ + "(a) NeuS + Deformation" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 343, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/482fc6ba28a449987ec0f8c54cc7278413e362e3e1f1958aa246165b843acba7.jpg", + "image_caption": [ + "(b) 4DRegSDF w/o $\\mathcal{L}_e$", + "Figure 6. Qualitative comparison on (a) NeuS [70] combined with a deformable network [54], (b) ours without surface regularization, and our final 4DRegSDF method. Left: rendered color image, right: rendered surface normal map." + ], + "image_footnote": [], + "bbox": [ + 354, + 88, + 617, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d1a1d8cf7cd9768f999b83423c37f8a3f7d3279b1866b6c6af1667f19a0f4e2.jpg", + "image_caption": [ + "(c) 4DRegSDF with $\\mathcal{L}_e$ (ours)" + ], + "image_footnote": [], + "bbox": [ + 629, + 88, + 890, + 268 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a010613ef315b8c78c1d8a4adc4241908007e83fe4cace68c4a303b683335923.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Ratio of trained data
100% (Tables 2,3)25%
PSNRMS-SSIMCDPSNRMS-SSIMCD
TiNeuVox [19]22.80.8410.05718.10.7020.085
NeuS + D [70]22.10.8390.04519.40.7800.064
Ours w/o Reg.22.40.8390.04218.10.7700.060
Ours23.90.8590.04221.10.8520.058
", + "bbox": [ + 80, + 344, + 467, + 424 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/96b77bd1bb8cc027f38d268e9ddd5c15b82ed49ee731aef770593a36d37a0b9a.jpg", + "table_caption": [ + "Table 4. Ablation study on few-shot frames. We choose the two scenes, 3D printer and Sriracha-tree, for the photometric and geometry evaluation. Ratio indicates the percentage of the data that we used for training the methods." + ], + "table_footnote": [], + "table_body": "
NetworkPreserve (✓)Evaluation
RegNeRFGeoNeuSOursPSNR ↑RMSE ↓
✓24.10.414
4DRegSDF✓24.70.384
✓25.30.354
", + "bbox": [ + 80, + 496, + 467, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "be one choice that reduces deformation ambiguity by learning jointly with SDF within a unified network $f$ and having cyclic deformation constraints $(\\mathcal{L}_{\\mathrm{cc}})$ .", + "bbox": [ + 75, + 623, + 468, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sparse view. To further verify the benefit of our surface regularization schemes, we conduct additional experiments regarding the sparse input views [29, 44]. For experiments, we use '3D printer' and 'Sriracha-tree' for evaluation. As stated in Table 4, when the training data is reduced by $25\\%$ , our method shows less performance drop when using a regularization scheme, showing less than $10\\%$ performance drop. Such results consistently support the claim that regularization techniques are effective for learning radiance fields under the given sparse views [29, 44].", + "bbox": [ + 75, + 676, + 468, + 828 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 844, + 194, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce using geometric spacetime regularizations for neural rendering and surface reconstruction in dynamic", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6bb7478cab5f810a76c358b7c1e13ad60d9c396e53ad12e15dbbe39e49c84f4f.jpg", + "table_caption": [ + "Table 5. Comparison of various regularization methods. RegNeRF [44] regularizes unobserved viewpoints and GeoNeuS [20] applies multi-view consistency as constraint." + ], + "table_footnote": [], + "table_body": "
NetworkPreserve (✓)Evaluation
TVCAbs4D samplePSNR ↑RMSE ↓
4DRegSDF✓✓24.40.392
✓✓24.90.379
✓✓✓25.30.354
", + "bbox": [ + 503, + 344, + 888, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation study on regularization. We decompose $\\mathcal{L}_e$ (Eq. 9) into total variation of curvature (TVC) and absolute curvature of SDF (ABS). 4D sample indicates Eq. 5.", + "bbox": [ + 498, + 414, + 890, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "scenes from a monocular video. While previous dynamic NeRF methods have primarily targeted rendering quality, our approach places greater emphasis on providing accurate geometry representation via surface regularization. Our method has demonstrated promising performance in both dynamic novel-view synthesis and 4D surface reconstruction tasks. However, a notable constraint of this work is it limited capability in processing extended videos and we will extend this work for long format videos.", + "bbox": [ + 496, + 474, + 890, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by IITP grants (No.2021-0-02068: AI Innovation Hub and No.2023R1A1C200781211) funded by the Korea government (MSIT).", + "bbox": [ + 498, + 617, + 890, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 691, + 596, + 707 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Marc Alexa, Daniel Cohen-Or, and David Levin. As-rigid-as-possible shape interpolation. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 157–164, 2000.", + "[2] Luca Ballan, Gabriel J Brostow, Jens Puwein, and Marc Pollefeys. Unstructured video-based rendering: Interactive exploration of casually captured videos. In ACM SIGGRAPH 2010 papers, pages 1-11. 2010.", + "[3] Aayush Bansal, Minh Vo, Yaser Sheikh, Deva Ramanan, and Srinivasa Narasimhan. 4d visualization of dynamic events from unconstrained multi-view videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5366-5375, 2020." + ], + "bbox": [ + 506, + 715, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17878", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[4] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021.", + "[5] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4):86-1, 2020.", + "[6] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph rendering. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 425-432, 2001.", + "[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021.", + "[8] Shenchang Eric Chen and Lance Williams. View interpolation for image synthesis. In Proceedings of the 20th annual conference on Computer graphics and interactive techniques, pages 279-288, 1993.", + "[9] Jaesung Choe, Sunghoon Im, Francois Rameau, Minjun Kang, and In So Kweon. Volumefusion: Deep depth fusion for 3d scene reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16086-16095, 2021.", + "[10] Jaesung Choe, Kyungdon Joo, Tooba Imtiaz, and In So Kweon. Volumetric propagation network: Stereo-lidar fusion for long-range depth estimation. IEEE Robotics and Automation Letters, 6(3):4672-4679, 2021.", + "[11] Jaesung Choe, Byeongin Joung, Francois Rameau, Jaesik Park, and In So Kweon. Deep point cloud reconstruction. arXiv preprint arXiv:2111.11704, 2021.", + "[12] Gene Chou, Ilya Chugunov, and Felix Heide. Gensdf: Two-stage learning of generalizable signed distance functions. Advances in Neural Information Processing Systems, 35:24905-24919, 2022.", + "[13] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015.", + "[14] Paul Debevec, Yizhou Yu, and George Borshukov. Efficient view-dependent image-based rendering with projective texture-mapping. In Eurographics Workshop on Rendering Techniques, pages 105-116. Springer, 1998.", + "[15] Manfredo P Do Carmo. Differential geometry of curves and surfaces: revised and updated second edition. Courier Dover Publications, 2016.", + "[16] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021." + ], + "bbox": [ + 78, + 90, + 470, + 910 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Arda Duzceker, Silvano Galliani, Christoph Vogel, Pablo Speciale, Mihai Dusmanu, and Marc Pollefeys. Deepvideovs: Multi-view stereo on video with recurrent spatiotemporal fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15324-15333, 2021.", + "[18] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017.", + "[19] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. arXiv preprint arXiv:2205.15285, 2022.", + "[20] Qiancheng Fu, Qingshan Xu, Yew-Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. arXiv preprint arXiv:2205.15848, 2022.", + "[21] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021.", + "[22] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. arXiv preprint arXiv:2210.13445, 2022.", + "[23] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4857-4866, 2020.", + "[24] Steven J Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 43-54, 1996.", + "[25] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020.", + "[26] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3d scene reconstruction with the manhattan-world assumption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5511-5520, 2022.", + "[27] Sunghoon Im, Hae-Gon Jeon, Stephen Lin, and In So Kweon. Dpsnet: End-to-end deep plane sweep stereo. arXiv preprint arXiv:1905.00538, 2019.", + "[28] Olga Karpenko, John F Hughes, and Ramesh Raskar. Freeform sketching with variational implicit surfaces. In Computer Graphics Forum, volume 21, pages 585-594. Wiley Online Library, 2002.", + "[29] Mijeong Kim, Seonguk Seo, and Bohyung Han. Infonerf: Ray entropy minimization for few-shot neural volume rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12912-12921, 2022.", + "[30] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017." + ], + "bbox": [ + 501, + 92, + 890, + 910 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17879", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021.", + "[32] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 31-42, 1996.", + "[33] Tzu-Mao Li, Miika Aittala, Frédo Durand, and Jaakko Lehtinen. Differentiable monte carlo ray tracing through edge sampling. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018.", + "[34] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021.", + "[35] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. arXiv preprint arXiv:2205.15723, 2022.", + "[36] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019.", + "[37] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. arXiv preprint arXiv:2206.05737, 2022.", + "[38] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987.", + "[39] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. arXiv preprint arXiv:2011.13495, 2020.", + "[40] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019.", + "[41] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020.", + "[42] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022.", + "[43] Zak Murez, Tarrence van As, James Bartolozzi, Ayan Sinha, Vijay Badrinarayanan, and Andrew Rabinovich. Atlas: End-to-end 3d scene reconstruction from posed images. In European conference on computer vision, pages 414-431. Springer, 2020.", + "[44] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Reg-" + ], + "bbox": [ + 78, + 92, + 468, + 883 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "nerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022.", + "[45] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Occupancy flow: 4d reconstruction by learning particle dynamics. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5379-5389, 2019.", + "[46] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020.", + "[47] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021.", + "[48] Sergio Orts-Escolano, Christoph Rhemann, Sean Fanello, Wayne Chang, Adarsh Kowdle, Yury Degtyarev, David Kim, Philip L Davidson, Sameh Khamis, Mingsong Dou, et al. Holoportation: Virtual 3d teleportation in real-time. In Proceedings of the 29th annual symposium on user interface software and technology, pages 741-754, 2016.", + "[49] Stanley Osher, Ronald Fedkiw, and K Piechor. Level set methods and dynamic implicit surfaces. Appl. Mech. Rev., 57(3):B15-B15, 2004.", + "[50] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019.", + "[51] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021.", + "[52] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), dec 2021.", + "[53] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020.", + "[54] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021.", + "[55] Yi-Ling Qiao, Alexander Gao, and Ming Lin. Neuphysics: Editable neural geometry and physics from monocular videos. In Advances in Neural Information Processing Systems, 2022." + ], + "bbox": [ + 503, + 92, + 890, + 877 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17880", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 2020.", + "[57] Yufan Ren, Tong Zhang, Marc Pollefeys, Sabine Susstrunk, and Fangjinhua Wang. Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16685-16695, 2023.", + "[58] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2304-2314, 2019.", + "[59] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019.", + "[60] Miroslava Slavcheva, Maximilian Baust, Daniel Cremers, and Slobodan Ilic. Killingfusion: Non-rigid 3d reconstruction without correspondences. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1386-1395, 2017.", + "[61] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. arXiv preprint arXiv:2210.15947, 2022.", + "[62] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, volume 4, pages 109–116, 2007.", + "[63] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. arXiv preprint arXiv:2111.11215, 2021.", + "[64] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021.", + "[65] Jiaxiang Tang. Torch-ngp: a pytorch implementation of instant-ngp, 2022. https://github.com/ashawkey/torch-ngp.", + "[66] Demetri Terzopoulos, John Platt, Alan Barr, and Kurt Fleischer. Elastically deformable models. In Proceedings of the 14th annual conference on Computer graphics and interactive techniques, pages 205-214, 1987.", + "[67] Itsuki Ueda, Yoshihiro Fukuhara, Hirokatsu Kataoka, Hiroaki Aizawa, Hidehiko Shishido, and Itaru Kitahara. Neural density-distance fields. arXiv preprint arXiv:2207.14455, 2022.", + "[68] Michael Waechter, Nils Moehrle, and Michael Goesele. Let there be color! large-scale texturing of 3d reconstructions. In European conference on computer vision, pages 836-850. Springer, 2014.", + "[69] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994, 2021." + ], + "bbox": [ + 78, + 90, + 470, + 906 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[70] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021.", + "[71] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021.", + "[72] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021.", + "[73] Guandao Yang, Serge Belongie, Bharath Hariharan, and Vladlen Koltun. Geometry processing with neural fields. Advances in Neural Information Processing Systems, 34:22483-22497, 2021.", + "[74] Gengshan Yang, Minh Vo, Neverova Natalia, Deva Ramanan, Vedaldi Andrea, and Joo Hanbyul. Banmo: Building animatable 3d neural models from many casual videos. In CVPR, 2022.", + "[75] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018.", + "[76] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. NeurIPS, 34:4805-4815, 2021.", + "[77] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv preprint arXiv:2302.14859, 2023.", + "[78] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5336-5345, 2020.", + "[79] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. arXiv preprint arXiv:2112.05131, 2021.", + "[80] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021.", + "[81] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022.", + "[82] C Lawrence Zitnick, Sing Bing Kang, Matthew Uytendaele, Simon Winder, and Richard Szeliski. High-quality video view interpolation using a layered representation. ACM transactions on graphics (TOG), 23(3):600-608, 2004." + ], + "bbox": [ + 501, + 92, + 890, + 882 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17881", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction/11a08db4-fb69-460a-a705-16e05f3f48ac_model.json b/2023/Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction/11a08db4-fb69-460a-a705-16e05f3f48ac_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f99bde4f62e6bcb558e47e5db89d266ced0f2c20 --- /dev/null +++ b/2023/Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction/11a08db4-fb69-460a-a705-16e05f3f48ac_model.json @@ -0,0 +1,2763 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.131, + 0.872, + 0.153 + ], + "angle": 0, + "content": "Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.193, + 0.846, + 0.213 + ], + "angle": 0, + "content": "Jaesung Choe1,2 Christopher Choy1 Jaesik Park3 In So Kweon2 Anima Anandkumar1,4" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.223, + 0.721, + 0.243 + ], + "angle": 0, + "content": "\\(^{1}\\)NVIDIA \\(^{2}\\)KAIST \\(^{3}\\)Seoul National University \\(^{4}\\)Caltech" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.562 + ], + "angle": 0, + "content": "We propose an algorithm, 4DRegSDF, for the spacetime surface regularization to improve the fidelity of neural rendering and reconstruction in dynamic scenes. The key idea is to impose local rigidity on the deformable Signed Distance Function (SDF) for temporal coherency. Our approach works by (1) sampling points on the deformed surface by taking gradient steps toward the steepest direction along SDF, (2) extracting differential surface geometry, such as tangent plane or curvature, at each sample, and (3) adjusting the local rigidity at different timestamps. This enables our dynamic surface regularization to align 4D spacetime geometry via 3D canonical space more accurately. Experiments demonstrate that our 4DRegSDF achieves state-of-the-art performance in both reconstruction and rendering quality over synthetic and real-world datasets. https://4dregsdf.github.io/" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.581, + 0.21, + 0.597 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.607, + 0.471, + 0.879 + ], + "angle": 0, + "content": "Reconstructing 4D dynamic scenes from natural videos is one of the fundamental problems of vision tasks that are widely used in various fields, such as robotics, autonomous driving, virtual reality, and many more [6, 8, 14, 24, 32]. It requires an understanding of not just the geometry of scenes but also the texture, illumination, material properties, and refraction for photo-realistic rendering. However, because of the difficulty in reconstructing static 3D geometry alone, learning the dynamics jointly has been a challenge that most conventional methods decomposed the task into multiple stages of geometry reconstruction, motion representation, texture mapping, and illumination estimation [6, 15, 33, 49, 68]. Such sequential pipelines that first reconstruct geometry and learn motion separately are not robust to many failure modes of each stage, resulting in inaccurate reconstruction, which is one of the major bottlenecks for photo-realistic and geometry-oriented spacetime scene understanding." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.287, + 0.697, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.418, + 0.587, + 0.426 + ], + "angle": 0, + "content": "Rendered color" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.287, + 0.721, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.418, + 0.686, + 0.426 + ], + "angle": 0, + "content": "Rendered normal" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.287, + 0.893, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.418, + 0.783, + 0.426 + ], + "angle": 0, + "content": "Rendered color" + }, + { + "type": "image_caption", + "bbox": [ + 0.803, + 0.418, + 0.882, + 0.426 + ], + "angle": 0, + "content": "Rendered normal" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.428, + 0.691, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.503, + 0.638, + 0.514 + ], + "angle": 0, + "content": "Spacetime NeRF" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.428, + 0.892, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.503, + 0.833, + 0.514 + ], + "angle": 0, + "content": "4DRegSDF (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.519, + 0.894, + 0.562 + ], + "angle": 0, + "content": "Figure 1. We present an implicit surface regularization method that fully utilizes the properties of SDF for robust shape reconstruction and novel-view rendering in dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.893, + 0.852 + ], + "angle": 0, + "content": "The recent advances in neural implicit representations, however, made accurate 3D reconstruction with differentiable volumetric rendering possible [40,41,46,50]. Mildenhall et al. [41] use sinusoidal encoding on the position as an input to multi-layered perceptrons (MLPs) to learn appearance and geometry of the target static scene. Specifically for photo-realistic rendering, neural radiance fields [4, 41, 54, 61, 63, 71, 79] mostly use density and color as the main output. However, the density representation can result in artifacts such as floaters and blurry boundaries since the density representation has high degrees of freedom [67, 70]. For accurate 3D reconstruction, the Signed Distance Function (SDF) started to gain traction for its high fidelity 3D surface reconstruction [26, 47, 70, 76, 81]. One of the key component for SDF learning is the Eikonol loss [25] which regularizes the SDF to have valid gradient. This acts as a strong regularization not just on the surface but on the entirety of the empty space." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "However, the 4D surface reconstruction from monocular RGB videos – which we reconstruct a dynamic scene – is still a highly ill-posed problem where we have to recon" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.887, + 0.341, + 0.9 + ], + "angle": 0, + "content": "Work done during the internship at NVIDIA." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17871" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.35 + ], + "angle": 0, + "content": "struct geometry and the dynamics of the scene simultaneously. Recent studies proposed to decompose the 4D space into a static geometry/texture (canonical model) and deformation to embed physical constraints into the 4D representation [51, 54]. The decomposition of the geometry could regularize the 4D space, but many papers use additional cues such as flow, or depth prediction [16, 21, 31, 34, 35, 69, 72, 78]. Still, the reconstruction is an ill-posed problem and it often fails for two reasons: (1) learning deformation is sensitive to topological variation and illumination change that often requires energy minimization terms for stable training. [28, 51, 60, 66]; (2) using the density representation – which has high degrees of freedom – could result in overfitting, floaters, and blurry reconstructions [29, 37, 44] as shown in Fig. 1. Thus, regularizing the spacetime domain is one of the most crucial parts of the 4D scene understanding in terms of geometry and graphics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.473, + 0.625 + ], + "angle": 0, + "content": "In this paper, we propose an algorithm for spacetime surface regularization that improves the fidelity of neural rendering and reconstruction in dynamic scenes. The key idea is to enforce the local rigidity on the deformable Signed Distance Function (SDF) for the 4D geometry representation. This allows us to (1) regularize the reconstruction using more explicit surface representation instead of fuzzy density measure [41]; (2) sample geometry efficiently for surface regularizations. Specifically, previous studies [37,47,70] show that Signed Distance Function can reconstruct geometries accurately with fewer images since learning the distance to the nearest surface to represent a surface is strong regularizations compared to the density function. In addition, the signed distance encodes not just distance but direction to the closest surface from gradients, thus allowing efficient surface sampling. Then, the sampled points are used to minimize the deformation energy and align 4D spacetime geometry via 3D canonical space." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.628, + 0.471, + 0.718 + ], + "angle": 0, + "content": "We use three datasets to evaluate our method: one synthetic dataset from Pumarola et al. [54], and two real datasets: Park et al. [52] and Gao et al. [22]. Our method achieves state-of-the-art performance on both geometry representation and appearance as visualized in Fig. 1. We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.731, + 0.468, + 0.775 + ], + "angle": 0, + "content": "- Extends the energy minimization in conventional 4D surface reconstruction studies into learning neural dynamic scene reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.793, + 0.468, + 0.838 + ], + "angle": 0, + "content": "- Regularize geometric properties of the 4D surface by the total variation of the curvature (TVC) and the absolute curvature of the SDF." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- Robust performance of rendering and reconstruction under the dynamic few-shot setup compared to recent implicit geometric regularization studies [20, 44]." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.731, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.091, + 0.645, + 0.105 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.403 + ], + "angle": 0, + "content": "Conventional reconstruction and rendering. Video often contains time-varying properties, such as illumination change, geometry deformation, topology changes, and rigid or deformable motion. Such time-varying elements result in multi-view inconsistency among different frames and increase the level of difficulties of the video rendering task. For this issue, previous studies often require exhaustive multi-camera setup [5, 13, 48, 82] or synchronized multiple videos [2, 3]. None-learning based methods [1, 28, 60, 62] commonly assume the coarse geometry from RGB-D priors to introduce the energy regularization for the realistic geometry and photometric description of the target dynamic scenes. For instance, terzopoulos et al. [66] use elastic energy to deform the surface, and Slavcheva et al. [60] propose three typical energy minimization to prevent uncontrolled deformations. These ideas are related to the local rigidity to reconstruct plausible and geometricallymeaningful surface [1, 62] along the spatio-temporal domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.892, + 0.713 + ], + "angle": 0, + "content": "Neural implicit reconstruction and rendering. Recently, neural scene representation [9, 23, 40, 43, 45, 50, 53, 64] gets noticeable attention with their promising quality of implicit shape representations, such as occupancy or signed distance function. Not just geometry, the studies [36, 46, 58, 59] handle both geometry and appearance from multi-view images by neural implicit representation. In terms of the novelview synthesis task, Mildenhall et al. [41] demonstrates the impressive quality of rendered images. This paper takes a 5D radiance field (3D location and 2D viewing direction) to infer color and density. Then, it accumulates per-point inferences along each ray to determine pixel color. Such formulation makes it possible to train MLPs for learning geometry and appearance simultaneously. Based on this design, the series of following papers address the remaining problems of the pioneering paper [41], such as generalization [7, 57, 71], few-shot images [29, 44, 80], or surface reconstruction by SDF [70, 76, 77]. Nonetheless, these methods mostly focus on targeting static scenes without consideration of the moving objects or temporal change." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Neural video rendering. More recently, several spacetime view synthesis studies [21,34,51,51,54,55,72] adopt neural implicit representations to render novel view synthesis from a video. Typically, these studies can be divided into two dominant strategies: deformable NeRFs [19,35,51,54] and prior-based video NeRFs [21, 34, 72]. Concurrent works [19, 35] propose fast deformable neural rendering methods by using voxel representations. Typically without relying on off-the-shelf information, such as monocular depth maps [34, 78], surface normal maps [81], or optical flow [35], we aim to render and reconstruct scenes purely from posed images." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17872" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.351 + ], + "angle": 0, + "content": "Regularization for neural implicit representation. Relying on the nature of the geometry, a series of papers [29, 44, 55] present the effectiveness of the regularization in learning neural fields. Under the few-shot setup, Niemeyer et al. [44] regularize the unobserved viewpoints via normalizing flow models. Kim et al. [29] apply entropy loss to alleviate reconstruction inconsistency. Fu et al. [20] apply multi-view consistency for the static surface reconstruction. In light of the growing popularity of spatio-temporal surface reconstruction and neural rendering, this paper proposes a novel approach to regulating the spacetime surface through the use of 3D canonical space. We have observed that the direct application of the SDF to this deformation field can result in low-fidelity reconstruction and rendering due to ill-posed problems in dynamic scenes. To overcome this issue, we highlight the importance of our regularization to achieve precise reconstruction and rendering." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.365, + 0.204, + 0.383 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.391, + 0.47, + 0.498 + ], + "angle": 0, + "content": "Neural radiance field [41] allows fully differentiable learning of geometry and appearance by learning color and density from RGB images. Recent studies [42, 70] on surface reconstruction utilize Signed Distance Function (SDF) to represent scene geometry since The SDF representation has specific favorable characteristics that give it an edge for this particular objective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.473, + 0.751 + ], + "angle": 0, + "content": "Differential surface geometry in SDF. The Signed Distance Function (SDF) stands as a predominant representation for implicit geometry. We can derive mathematical attributes on a crisp surface from SDF that could be challenging to derive from density-based representations: surface normal vector \\(\\mathbf{n} = \\frac{\\partial s}{\\partial\\mathbf{p}}\\), tangent plane projection matrix \\(P = I - \\mathbf{n}^{\\top}\\mathbf{n}\\), and surface curvature through the Hessian matrix \\(H = \\frac{\\partial^2s}{\\partial\\mathbf{p}^2}\\) [73] Note that \\(I\\) denotes the identity matrix, \\(s\\) the SDF value, and \\(\\mathbf{p}\\) a 3D coordinate. We can derive these mathematical attributes because SDF represents the distance to the closest surface at every point. We can force this property for implicit geometry by using the Eikonal loss [25], which forces the norm of gradients of SDF \\((\\| \\mathbf{n}\\| _2 - 1)^2\\) to be 1, and the loss has been widely used as an important regularization for 3D reconstruction methods [20,25,70,76]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.759, + 0.47, + 0.856 + ], + "angle": 0, + "content": "NeuS [70] uses the same volumetric rendering equation in NeRF [41]. However, the opaqueness \\(\\alpha\\) is derived from the SDF value \\(s\\) as \\(\\alpha = 1 - \\exp \\left(-\\int_{t_{i + 1}}^{t_i}\\rho (t)dt\\right)\\) where \\(\\rho\\) is the density function \\(\\rho = \\max \\left(\\frac{s - \\Delta s}{s + \\Delta s},0\\right)\\) and \\(\\Delta s\\) is the displacement of the SDF. Following NeRF [41], NeuS utilizes the volumetric rendering as:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.861, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\hat {C} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i} \\quad \\text {s . t .} \\quad T _ {i} = \\exp \\left(- \\sum_ {j = 1} ^ {i - 1} \\sigma_ {i} \\delta_ {i}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}_i\\) denotes the color of the \\(i\\)-th sample point \\(\\mathbf{p}_i\\), \\(\\alpha_i\\) is the opaqueness over distance \\(\\delta_i\\) on the ray \\(\\mathbf{r} = \\{\\mathbf{p}_i\\}\\). \\(T_i\\) is the cumulative transmittance along the ray. Based on this formulation, NeuS [70] is trained purely from posed images to describe the static scene surface using the SDF." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.174, + 0.892, + 0.235 + ], + "angle": 0, + "content": "Deformable NeRF [51, 54] is one way of novel-view synthesis methods in the 4D dynamic scenes. The idea is to decompose the 4D deformed space \\([p; t]\\) into the 3D canonical geometry \\(\\mathbf{p}_c\\) and the temporal deformation \\(\\Delta \\mathbf{p}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.245, + 0.892, + 0.262 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {c} = \\operatorname {D e f o r m} ([ \\mathbf {p}; t ]) = \\mathbf {p} + \\Delta \\mathbf {p}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.892, + 0.392 + ], + "angle": 0, + "content": "where \\(\\Delta \\mathbf{p}(= h([p; t]))\\) is the amount of space deformation, a function \\(h(\\cdot)\\) represents a deformation field, and the deformation field transports a point \\([\\mathbf{p}; t]\\) to the canonical coordinate \\(\\mathbf{p}_c = [x_c, y_c, z_c]\\). Although progress has been made in the field of dynamic neural rendering, we address the importance of the spacetime surface regularization for learning appearance, geometry, and motion in dynamic scenes as visualized in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.404, + 0.591, + 0.42 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.596 + ], + "angle": 0, + "content": "This section presents our framework, 4DRegSDF. Our network represents the 4D spacetime domain as a combination of the static geometry in Signed Distance Function and the deformation field. Given a set of posed frames from a monocular RGB video, our network is trained to encode a dynamic scene (Sec. 4.1) and apply our spacetime surface regularization (Sec. 4.2). In contrast to previous video neural field papers that use priors such as depths or flows [21, 34, 35], our method relies only on posed image frames from a video as input to the system making the system simple and efficient." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.827, + 0.621 + ], + "angle": 0, + "content": "4.1. Deformable Signed Distance Function" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.688 + ], + "angle": 0, + "content": "Given a point \\(\\mathbf{p}\\) at time \\(t\\) within the deformed space and its viewing direction \\(\\mathbf{d}\\), our model \\(f\\) predicts the color \\(\\mathbf{c}\\), the SDF value \\(s\\), and the deformation delta \\(\\Delta \\mathbf{p}\\). This can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.698, + 0.891, + 0.715 + ], + "angle": 0, + "content": "\\[\n[ s; \\Delta \\mathbf {p}; \\mathbf {c} ] = f ([ \\mathbf {p}; t ], \\mathbf {d}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.724, + 0.891, + 0.769 + ], + "angle": 0, + "content": "As visualized in Fig. 3, we use the same volumetric rendering on the SDF (Eq. 1) [70] to get color value per each ray \\(\\mathbf{r}\\), and minimize the photometric loss \\(\\mathcal{L}_{\\mathrm{c}}\\) as," + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.772, + 0.891, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c}} = \\left| C (\\mathbf {r}) - \\hat {C} (\\mathbf {r}) \\right| _ {1}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.808, + 0.891, + 0.84 + ], + "angle": 0, + "content": "where \\(C\\) and \\(\\hat{C}\\) are ground truth color and predicted color, respectively, along a query ray \\(\\mathbf{r}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.788, + 0.864 + ], + "angle": 0, + "content": "4.2. Spacetime surface regularization" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Reconstructing dynamic surfaces from videos presents challenges due to its inherent complexity, often leading to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17873" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.316, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.089, + 0.6, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.087, + 0.877, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.226, + 0.731, + 0.24 + ], + "angle": 0, + "content": "Figure 2. Surface sampling for our spacetime surface regularization in three stages." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.263, + 0.451, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.392, + 0.422, + 0.407 + ], + "angle": 0, + "content": "Figure 3. Ray sampling for neural rendering [70]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.47, + 0.466 + ], + "angle": 0, + "content": "visual artifacts as shown in Fig. 6. Accordingly, designing a robust regularization is one of the critical problems for successful 4D reconstruction [1,51,60,62,66,73]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.468, + 0.47, + 0.664 + ], + "angle": 0, + "content": "In this section, we adopt three regularizations to enhance the learning of deformable Signed Distance Function from videos: 1. total variation of curvature, 2. the absolute curvature space-time, and 3. the Eikonal loss [25]. The first regularization aims to limit the amount of change in curvature induced by motion and the second regularization flattens out the 4D reconstruction and limits unnecessary kinks. The last regularization [25] enforces the gradient of SDF to be valid. First two regularizations are related to the properties of surface and can be computed only on the surface. We elaborate on the sampling strategy and how to augment the samples in continuous space and time, and lastly present the regularizations. Overall pipeline is visualized in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.779 + ], + "angle": 0, + "content": "Spacetime surface sampling. The most commonly used sampling in neural fields is ray casting where you select points along a ray for rendering. Many regularizations proposed for NeRF also use samples along a ray [20,29,37,44]. However, surface constraints we adopt require samples on the surface of objects no along rays. For this, we use the definition of the SDF for surface sampling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The SDF is defined as the (signed) distance to the closest surface from the query point. In static scenes, the surface normal indicates the direction of the closest surface [12, 39, 73]. However, in 4D dynamic scenes, time is another component to determine the closest surface. In our setup, our network \\( f(\\cdot) \\) is trained to infer the SDF value \\( s \\) from a query \\( [\\mathbf{p}; t] \\). Accordingly, the gradient step towards the closest surface naturally involves and spatial update as" + }, + { + "type": "algorithm", + "bbox": [ + 0.501, + 0.264, + 0.892, + 0.488 + ], + "angle": 0, + "content": "Algorithm 1 Spacetime surface regularization \nInput: Randomly initialized 4D point \\([\\mathbf{p}_i; t_i]\\). \nOutput: Total regularization \\(\\mathcal{L}_e\\). \n1: function 4DRegSDF([p_i, t_i]) \n2: [p, t] ← [p_i, t_i] \n3: for iterations do \n4: [p, t] ← StepGrad([p; t]) ▷ Eq. 5 \n5: end for \n6: pc ← Deform([p; t]) ▷ Eq. 2 \n7: t' ← Random(0, 1) ▷ Time augmentation \n8: p' ← InvDeform([pc; t']) ▷ Eq. 6 \n9: \\(\\mathcal{L}_e \\gets \\text{ComputeLoss}([\\mathbf{p}; t], [\\mathbf{p}'; t'])\\) ▷ Eq. 9 \n10: return \\(\\mathcal{L}_e\\) \n11: end function" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.514, + 0.892, + 0.559 + ], + "angle": 0, + "content": "well as the time update. Given randomly initialized point \\( p_i \\) at the time \\( t_i \\), each query point travels along the spacetime SDF 'hills' to be located at the isosurface as:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.566, + 0.892, + 0.601 + ], + "angle": 0, + "content": "\\[\n[ \\mathbf {p}; t ] = \\operatorname {S t e p G r a d} ([ \\mathbf {p} _ {i}; t _ {i} ]) = [ \\mathbf {p} _ {i}; t _ {i} ] - s \\cdot \\frac {\\partial s}{\\partial [ \\mathbf {p} _ {i} ; t _ {i} ]}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.892, + 0.719 + ], + "angle": 0, + "content": "where \\(\\frac{\\partial s}{\\partial[\\mathbf{p};t]}\\) is the gradient of the SDF value \\(s\\) in terms of \\([\\mathbf{p};t]\\). By doing so, we can regularize the 4D surface in a wide-spectrum time domain including unobserved timestamps within training samples. Thus, by iteratively conducting this operation above, we move the initial 4D point \\([\\mathbf{p}_i; t_i]\\) to \\([\\mathbf{p}; t]\\) closer to the isosurface having smaller SDF value, as described in L. 2-5 of Alg. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.726, + 0.892, + 0.817 + ], + "angle": 0, + "content": "Temporal augmentation. Given the surface samples \\([\\mathbf{p};t]\\) from Eq. 5, we enforce the local rigidity to the surface samples for temporal coherency. First, we augment the time component \\(t\\) into randomly selected \\(t^{\\prime}\\) as visualized in Fig. 2-(b). Specifically, we use the inverse-deformation network \\(g\\) [35,45,74] which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.828, + 0.892, + 0.846 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} ^ {\\prime} = \\operatorname {I n v D e f o r m} \\left(\\left[ \\mathbf {p} _ {c}; t ^ {\\prime} \\right]\\right) = g \\left(\\mathbf {p} _ {c}, t ^ {\\prime}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_c\\) indicates the points at the canonical space from \\([\\mathbf{p};t]\\) (Eq. 3). \\(\\mathbf{p}'\\) is a inverse-deformed point from \\(\\mathbf{p}_c\\) at the augmented time \\(t'\\) as stated in L. 6-8 of Alg. 1. Through our" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17874" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "temporal augmentation stage, we get a pair of two surface samples that correspond to each other, \\(([\\mathbf{p};t],[\\mathbf{p}';t'])\\). This pair will be used for evaluating spatio-temporal constraints the next section." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.161, + 0.473, + 0.45 + ], + "angle": 0, + "content": "Regularization on the spacetime curvature. Inspired by the conventional 4D reconstruction studies [1,60,62,66], we regularize geometric properties of dynamic surfaces: the total variation of curvature (TVC) and the absolute curvature of SDF surface on continuous time. The TVC measures the change of the curvature at different time steps and by minimizing the change, we assumes that surface should deform as little as possible while satisfying video observations. On the contrary, the absolute curvature measures curvature values and thus regularizes the complexity of the reconstruction and movement by minimizing the distortion in space. For the TVC, we adopt the stretching loss [73] where it minimizes the difference of inner products of two tangent vectors at two different time steps: \\( v_{1}, v_{2} \\) at \\( t \\) and \\( v_{1}', v_{2}' \\) at \\( t' \\). The tangent vector is a remainder of a vector on surface \\( v = P\\mathbf{p} \\) where \\( P(= I - \\mathbf{nn}^{\\top}) \\) is the tangential projection matrix and \\( \\mathbf{n}(= \\frac{\\partial s}{\\partial\\mathbf{p}}) \\) is the surface normal vector at a point \\( \\mathbf{p} \\). Let \\( J_{1}, J_{2} \\) be the Jacobians transforming \\( v_{1}, v_{2} \\) to \\( v_{1}', v_{2}' \\) respectively. Then, we compute the TVC as:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.461, + 0.47, + 0.501 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left| v _ {1} ^ {\\top} v _ {2} - v _ {1} ^ {\\prime \\top} v _ {2} \\right| = \\left| \\mathbf {p} _ {1} ^ {\\top} P _ {1} ^ {\\top} P _ {2} \\mathbf {p} _ {2} - \\mathbf {p} _ {1} ^ {\\top} P _ {1} ^ {\\top} J _ {1} ^ {\\top} J _ {2} P _ {2} \\mathbf {p} _ {2} \\right|, (7) \\\\ \\propto \\left| P _ {1} ^ {T} \\left(I - J _ {1} ^ {T} J _ {2}\\right) P _ {2} \\right|, (8) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.47, + 0.595 + ], + "angle": 0, + "content": "The absolute curvature of the 4D SDF is the L1 norm of the Hessian matrix \\(H = \\frac{\\partial^2 s}{\\partial \\mathbf{p}^2}\\). By minimizing the absolute curvature, the reconstruction removes unnecessary curvatures or kinks in the reconstructions making the surface as smooth as possible. The total regularization is:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.607, + 0.469, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {e} (\\mathbf {p}, \\hat {\\mathbf {p}}) = \\| P ^ {\\top} (I - J _ {g} J _ {f}) P \\| _ {2} + \\| H \\| _ {1}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.469, + 0.687 + ], + "angle": 0, + "content": "where \\(J_{g}\\) and \\(J_{f}\\) are the Jacobians that transformed points via the deformation network \\(f(\\cdot)\\) and the inverse deformation network \\(g(\\cdot)\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.469, + 0.779 + ], + "angle": 0, + "content": "We summarize our spacetime surface regularization in Algorithm 1. We achieve high-quality reconstruction results that are geometrically simple while they match video observations through neural rendering and minimizing the photometric loss. Finally, we train our network by combining the proposed losses as below," + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.794, + 0.469, + 0.811 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {c c} \\mathcal {L} _ {c c} + \\lambda_ {\\mathrm {e}} \\mathcal {L} _ {\\mathrm {e}} + \\lambda_ {\\mathrm {e i k}} \\mathcal {L} _ {\\mathrm {e i k}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "where \\(\\lambda_{c} = 1\\), \\(\\lambda_{cc} = 0.01\\), \\(\\lambda_{e} = 0.001\\), and \\(\\lambda_{\\mathrm{eik}} = 0.01\\). \\(\\mathcal{L}_{\\mathrm{eik}}\\) represents the Eikonal loss [25] following NeuS [70]. Also, we enforce cycle consistency loss in learning forward/inverse deformation fields by \\(\\mathcal{L}_{cc} = \\| \\mathbf{p} - g(\\mathbf{p}_c,t)\\| _1\\) [35,45,74]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.634, + 0.108 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.116, + 0.695, + 0.132 + ], + "angle": 0, + "content": "5.1. Novel-view synthesis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.139, + 0.892, + 0.231 + ], + "angle": 0, + "content": "Given a set of posed images, the novel-view synthesis aims to render images from viewpoints that are not in the part of the provided posed images. Since our 4DRegSDF also follows the neural rendering schemes, we compare the quality of rendered images with the recent spacetime NeRF papers on novel-view synthesis evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.238, + 0.892, + 0.359 + ], + "angle": 0, + "content": "Dataset. We utilize the Dynamic NeRF Synthetic dataset provided by D-NeRF [54] and the real-world scenes from HyperNeRF dataset [52]. These datasets offer natural monocular videos with dynamic frames and continuous camera motion. We follow the official train/test split for training and evaluation over ours and related studies. Following the conventions [41, 51, 54], we evaluate the rendering performance through PSNR, SSIM, and LPIPS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Comparison. We compare our method with a few state-of-the-art 4D NeRF papers, D-NeRF [54], TiNeuVox [19], which also do not rely on the prior information [21,34,35]. Also, we extend NeuS [70] for the 4D dynamic reconstruction task by adopting a deformation network from D-NeRF [54]. We denote this deformable NeuS as 'NeuS+D' and use it One of the concurrent works, DeVRF [35], also proposes a deformation-based dynamic neural rendering task. However, the paper requires additional multi-view static scenes for initialization so we did not include it in our experiment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.806 + ], + "angle": 0, + "content": "When compared to existing research, our method achieves state-of-the-art performance among the recent spacetime NeRF methods as shown in Table 1 and Table 2. Many of these studies rely on view-dependent surface representations and the density function [41]. In contrast, our method utilizes the SDF representation, allowing us to more accurately represent geometry. This distinction influences the rendering quality. Notably, in terms of the qualitative reconstruction results, the previous state-of-the-art method struggles to accurately capture surface details as illustrated in Fig. 4 and Fig. 5. We suspect this limitation is rooted in the inherent issues of the density representation. As a result, many of the earlier spacetime NeRF studies place a heavy emphasis on the rendering quality. However, we claim that by imposing scene geometry constraints through our spacetime surface regularization, we can further improve the quality of the geometry under the precise geometry presentation, namely SDF." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.697, + 0.832 + ], + "angle": 0, + "content": "5.2. Scene reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Dataset. Dycheck dataset [22] is one of the few datasets that provide ground truth geometry information for dynamic scenes. Yet, the ground truth depth maps are only provided for the training data split and we could not evalu" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17875" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.894, + 0.184 + ], + "angle": 0, + "content": "
Dynamic Synthetic NeRF dataset [54]
Bouncing ballsHell warriorHookJumping jacksLegoMutantStand upT-rex
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
NeRF [41]20.260.9113.510.8116.650.8418.280.8820.300.7920.310.9118.190.8924.490.93
D-NeRF [54]38.930.9825.020.9529.250.9632.800.9821.640.8331.290.9632.790.9831.750.97
TiNeuVox [19]40.730.9928.170.9731.450.9734.230.9825.020.9233.610.9835.430.9932.700.98
TorchNGP [65]40.440.9924.540.9431.460.9731.450.9525.070.9335.860.9935.290.9931.350.97
NeuS+D [70]36.980.9623.770.9428.090.9428.280.9324.170.9030.170.9430.990.9524.940.93
Ours40.890.9925.720.9532.200.9833.630.9925.100.9431.830.9633.600.9933.640.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.156, + 0.186, + 0.813, + 0.2 + ], + "angle": 0, + "content": "Table 1. Rendering evaluation in the D-NeRF synthetic dataset. We use PSNR (↑) and SSIM (↑) for the metric." + }, + { + "type": "table", + "bbox": [ + 0.148, + 0.213, + 0.825, + 0.326 + ], + "angle": 0, + "content": "
HyperNeRF dataset [52]
Broom3D PrinterChickenPeel-bananaMEAN
PSNRMS-SSIMPSNRMS-SSIMPSNRMS-SSIMPSNRMS-SSIMPSNRMS-SSIM
NeRF [41]19.90.65320.70.78019.90.77720.00.76920.10.745
NV [36]17.70.62316.20.66517.60.61515.90.38016.90.571
Nerfies [51]19.20.56720.60.83026.70.94322.40.87222.20.803
TorchNGP [65]19.80.57121.10.84127.60.94621.10.85123.50.815
TiNeuVox [19]21.50.68622.80.84128.30.94724.40.87324.30.837
NeuS+D [70]21.90.68922.10.83925.90.93324.60.87523.60.834
Ours23.30.70223.90.85929.80.96525.40.88125.30.851
" + }, + { + "type": "table_caption", + "bbox": [ + 0.29, + 0.327, + 0.674, + 0.341 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation in the HyperNeRF dataset [52]." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.353, + 0.894, + 0.43 + ], + "angle": 0, + "content": "
Dycheck dataset [22]
CreeperHaru-sitPaper-windmillSriracha-tree
RMSE↓CD↓F1↑PSNR↑RMSE↓CD↓F1↑PSNR↑RMSE↓CD↓F1↑PSNR↑RMSE↓CD↓F1↑PSNR↑
TorchNGP [65]0.3400.05087.7817.320.5020.08176.3327.590.3110.05781.9823.30.3110.05281.4427.9
TiNeuVox [19]0.3450.04888.4216.690.6340.08075.1826.110.3130.05382.2924.960.3250.05781.5030.57
NeuS+D [70]0.3500.07174.1817.050.4910.11860.1327.880.2950.03891.3222.520.2830.04585.2326.83
Ours0.3230.06482.3617.570.4940.07778.5328.290.3100.05183.2424.260.2900.04292.2530.66
" + }, + { + "type": "table_caption", + "bbox": [ + 0.142, + 0.432, + 0.821, + 0.445 + ], + "angle": 0, + "content": "Table 3. Evaluation in the Dycheck dataset [22] Note that CD (↓) means Chamfer Distance and F1 (↑) is F1-score." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.47, + 0.63 + ], + "angle": 0, + "content": "ate geometry reconstruction quality on the test split. Accordingly, we subsample the official train data to form a new train/test split. The new split consists of \\(75\\%\\) training data and \\(25\\%\\) test data per scene. For evaluation metrics, we follow the three geometric measures, RMSE, Chamfer Distance, and F1-score. RMSE is a widely-adopted metric for depthmap prediction evaluation [10, 17, 27, 56]. Chamfer distance [11, 18] measures the distance between the ground truth point cloud and the prediction. Meanwhile, the F1-score measures the 3D volume occupancy between the ground truth and the predictions [11, 30, 75]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Comparison. To evaluate geometric reconstruction quality, we evaluate on both view specific depth map prediction task and the 3D geometry reconstruction task. For depth map prediction task, we extract the depth map at the zero-crossing of SDF. For density-based NeRFs [19], we render depth maps by computing the expected depth. For the final metric, measure the RMSE in the neural volume rendering schemes [41, 70]. For 3D reconstruction evaluation, we follow the surface extraction pipeline provided by the official D-NeRF implementation for density-based methods [19, 54]. In our case, we extract mesh at the zero-crossings of SDF by using marching-cube algorithm [38] for 3D geometric evaluation. For evaluation metrics, we utilize the Chamfer Distance (CD ↓), F1-score (F1 ↑), and RMSE ↓." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Based on this formulation, we calculate the 2D/3D geometric performance by ours and recent studies [19]. As" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.6 + ], + "angle": 0, + "content": "described in Table 3, our method still consistently outperforms the quality of 4D surface reconstruction tasks. We will state the detailed ablation study in the following section, however, in short, the dominant performance improvements are driven by the plausible surface representation by Signed Distance Function and the proposed surface regularization schemes. Finally, our 4DRegSDF successfully renders and reconstructs the 4D dynamic scenes thanks to the plausible geometry representation and learning schemes" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.617, + 0.661, + 0.632 + ], + "angle": 0, + "content": "6. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.688 + ], + "angle": 0, + "content": "In this section, we provide additional experiments to analyze each module of the 4DRegSDF. Note that additional ablation results are included in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.802 + ], + "angle": 0, + "content": "Effectiveness of surface regularization. Surface regularization is the key to making 4D surfaces plausible shapes while learning deformation. When comparing with Fig. 6-(b) and (c), we found that \\(\\mathcal{L}_e\\) affects the structure of the reconstructed table. It implies that our surface regularization takes a role in shaping plausible structures by aligning geometry between surfaces at different times." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Comparison with different regularization schemes. We compare our surface regularization with previous regularization schemes, e.g. RegNeRF [44] and GeoNeuS [20], as in Table 5. Based on these results, we claim that our spacetime surface regularization is particularly promising compared to previous regularization strategies. Also, we pro" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17876" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.101, + 0.66, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.167, + 0.825, + 0.181 + ], + "angle": 0, + "content": "No ground truth mesh" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.237, + 0.337, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.359, + 0.25, + 0.37 + ], + "angle": 0, + "content": "TiNeuVox" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.237, + 0.614, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.441, + 0.359, + 0.55, + 0.372 + ], + "angle": 0, + "content": "4DRegSDF (ours)" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.236, + 0.882, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.72, + 0.359, + 0.798, + 0.37 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.39, + 0.888, + 0.406 + ], + "angle": 0, + "content": "Figure 4. Reconstruction results (mesh) with rendered images from our 4DRegSDF and previous state-of-the-art method, TiNeuVox [19]." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.433, + 0.306, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.591, + 0.306, + 0.746 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.747, + 0.221, + 0.757 + ], + "angle": 0, + "content": "TiNeuVox" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.433, + 0.541, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.591, + 0.541, + 0.746 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.747, + 0.482, + 0.757 + ], + "angle": 0, + "content": "NeuS + Deformation" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.433, + 0.77, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.591, + 0.77, + 0.746 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.747, + 0.706, + 0.759 + ], + "angle": 0, + "content": "4DRegSDF (ours)" + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.433, + 0.888, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.591, + 0.888, + 0.746 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.747, + 0.868, + 0.757 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.765, + 0.895, + 0.794 + ], + "angle": 0, + "content": "Figure 5. Qualitative results in HyperNeRF dataset [52]. We compare previous state-of-the-art method, TiNeuVox [19], NeuS [70] with deformation [54] and our 4DRegSDF. Left: rendered color images, right: rendered surface normal maps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.818, + 0.47, + 0.864 + ], + "angle": 0, + "content": "vide details of our surface regularization in Table 6. Note that we calculate avg. PSNR from HyperNeRF dataset [52], and avg. RMSE from DyCheck dataset [22]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Deformation network design. Unlike deformable NeRF methods [51, 54] that sequentially process deformation and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.818, + 0.895, + 0.895 + ], + "angle": 0, + "content": "rendering (Eq. 2), our 4DRegSDF proposes to jointly learn deformation and SDF through a unified network \\( f \\) (Eq. 3) as stated in Sec. 4.1. It turns out that our network design shows the better quality of surface normal maps as visualized in Fig. 6-(a) and (c). Such architecture design can" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17877" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.344, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.272, + 0.285, + 0.285 + ], + "angle": 0, + "content": "(a) NeuS + Deformation" + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.089, + 0.618, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.272, + 0.555, + 0.286 + ], + "angle": 0, + "content": "(b) 4DRegSDF w/o \\(\\mathcal{L}_e\\)" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.089, + 0.892, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.272, + 0.852, + 0.286 + ], + "angle": 0, + "content": "(c) 4DRegSDF with \\(\\mathcal{L}_e\\) (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.295, + 0.893, + 0.324 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison on (a) NeuS [70] combined with a deformable network [54], (b) ours without surface regularization, and our final 4DRegSDF method. Left: rendered color image, right: rendered surface normal map." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.345, + 0.468, + 0.425 + ], + "angle": 0, + "content": "
Ratio of trained data
100% (Tables 2,3)25%
PSNRMS-SSIMCDPSNRMS-SSIMCD
TiNeuVox [19]22.80.8410.05718.10.7020.085
NeuS + D [70]22.10.8390.04519.40.7800.064
Ours w/o Reg.22.40.8390.04218.10.7700.060
Ours23.90.8590.04221.10.8520.058
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.428, + 0.47, + 0.484 + ], + "angle": 0, + "content": "Table 4. Ablation study on few-shot frames. We choose the two scenes, 3D printer and Sriracha-tree, for the photometric and geometry evaluation. Ratio indicates the percentage of the data that we used for training the methods." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.497, + 0.468, + 0.559 + ], + "angle": 0, + "content": "
NetworkPreserve (✓)Evaluation
RegNeRFGeoNeuSOursPSNR ↑RMSE ↓
✓24.10.414
4DRegSDF✓24.70.384
✓25.30.354
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.563, + 0.47, + 0.606 + ], + "angle": 0, + "content": "Table 5. Comparison of various regularization methods. RegNeRF [44] regularizes unobserved viewpoints and GeoNeuS [20] applies multi-view consistency as constraint." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.469, + 0.671 + ], + "angle": 0, + "content": "be one choice that reduces deformation ambiguity by learning jointly with SDF within a unified network \\( f \\) and having cyclic deformation constraints \\( (\\mathcal{L}_{\\mathrm{cc}}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.678, + 0.47, + 0.829 + ], + "angle": 0, + "content": "Sparse view. To further verify the benefit of our surface regularization schemes, we conduct additional experiments regarding the sparse input views [29, 44]. For experiments, we use '3D printer' and 'Sriracha-tree' for evaluation. As stated in Table 4, when the training data is reduced by \\(25\\%\\), our method shows less performance drop when using a regularization scheme, showing less than \\(10\\%\\) performance drop. Such results consistently support the claim that regularization techniques are effective for learning radiance fields under the given sparse views [29, 44]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.196, + 0.86 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We introduce using geometric spacetime regularizations for neural rendering and surface reconstruction in dynamic" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.345, + 0.89, + 0.412 + ], + "angle": 0, + "content": "
NetworkPreserve (✓)Evaluation
TVCAbs4D samplePSNR ↑RMSE ↓
4DRegSDF✓✓24.40.392
✓✓24.90.379
✓✓✓25.30.354
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.415, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Table 6. Ablation study on regularization. We decompose \\(\\mathcal{L}_e\\) (Eq. 9) into total variation of curvature (TVC) and absolute curvature of SDF (ABS). 4D sample indicates Eq. 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.892, + 0.611 + ], + "angle": 0, + "content": "scenes from a monocular video. While previous dynamic NeRF methods have primarily targeted rendering quality, our approach places greater emphasis on providing accurate geometry representation via surface regularization. Our method has demonstrated promising performance in both dynamic novel-view synthesis and 4D surface reconstruction tasks. However, a notable constraint of this work is it limited capability in processing extended videos and we will extend this work for long format videos." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.618, + 0.892, + 0.678 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by IITP grants (No.2021-0-02068: AI Innovation Hub and No.2023R1A1C200781211) funded by the Korea government (MSIT)." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.693, + 0.597, + 0.708 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.717, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[1] Marc Alexa, Daniel Cohen-Or, and David Levin. As-rigid-as-possible shape interpolation. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 157–164, 2000." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.775, + 0.894, + 0.83 + ], + "angle": 0, + "content": "[2] Luca Ballan, Gabriel J Brostow, Jens Puwein, and Marc Pollefeys. Unstructured video-based rendering: Interactive exploration of casually captured videos. In ACM SIGGRAPH 2010 papers, pages 1-11. 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.832, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[3] Aayush Bansal, Minh Vo, Yaser Sheikh, Deva Ramanan, and Srinivasa Narasimhan. 4d visualization of dynamic events from unconstrained multi-view videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5366-5375, 2020." + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.717, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17878" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[4] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.178, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[5] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4):86-1, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.249, + 0.471, + 0.316 + ], + "angle": 0, + "content": "[6] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph rendering. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 425-432, 2001." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.318, + 0.471, + 0.387 + ], + "angle": 0, + "content": "[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.389, + 0.471, + 0.444 + ], + "angle": 0, + "content": "[8] Shenchang Eric Chen and Lance Williams. View interpolation for image synthesis. In Proceedings of the 20th annual conference on Computer graphics and interactive techniques, pages 279-288, 1993." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.446, + 0.471, + 0.514 + ], + "angle": 0, + "content": "[9] Jaesung Choe, Sunghoon Im, Francois Rameau, Minjun Kang, and In So Kweon. Volumefusion: Deep depth fusion for 3d scene reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16086-16095, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.516, + 0.471, + 0.571 + ], + "angle": 0, + "content": "[10] Jaesung Choe, Kyungdon Joo, Tooba Imtiaz, and In So Kweon. Volumetric propagation network: Stereo-lidar fusion for long-range depth estimation. IEEE Robotics and Automation Letters, 6(3):4672-4679, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.573, + 0.471, + 0.613 + ], + "angle": 0, + "content": "[11] Jaesung Choe, Byeongin Joung, Francois Rameau, Jaesik Park, and In So Kweon. Deep point cloud reconstruction. arXiv preprint arXiv:2111.11704, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.616, + 0.471, + 0.669 + ], + "angle": 0, + "content": "[12] Gene Chou, Ilya Chugunov, and Felix Heide. Gensdf: Two-stage learning of generalizable signed distance functions. Advances in Neural Information Processing Systems, 35:24905-24919, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.672, + 0.471, + 0.739 + ], + "angle": 0, + "content": "[13] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.743, + 0.471, + 0.797 + ], + "angle": 0, + "content": "[14] Paul Debevec, Yizhou Yu, and George Borshukov. Efficient view-dependent image-based rendering with projective texture-mapping. In Eurographics Workshop on Rendering Techniques, pages 105-116. Springer, 1998." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.799, + 0.471, + 0.839 + ], + "angle": 0, + "content": "[15] Manfredo P Do Carmo. Differential geometry of curves and surfaces: revised and updated second edition. Courier Dover Publications, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.842, + 0.471, + 0.911 + ], + "angle": 0, + "content": "[16] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.471, + 0.911 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[17] Arda Duzceker, Silvano Galliani, Christoph Vogel, Pablo Speciale, Mihai Dusmanu, and Marc Pollefeys. Deepvideovs: Multi-view stereo on video with recurrent spatiotemporal fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15324-15333, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[18] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[19] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. arXiv preprint arXiv:2205.15285, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[20] Qiancheng Fu, Qingshan Xu, Yew-Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. arXiv preprint arXiv:2205.15848, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[21] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[22] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. arXiv preprint arXiv:2210.13445, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.514 + ], + "angle": 0, + "content": "[23] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4857-4866, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.517, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[24] Steven J Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 43-54, 1996." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.574, + 0.892, + 0.614 + ], + "angle": 0, + "content": "[25] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.616, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[26] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3d scene reconstruction with the manhattan-world assumption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5511-5520, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.687, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[27] Sunghoon Im, Hae-Gon Jeon, Stephen Lin, and In So Kweon. Dpsnet: End-to-end deep plane sweep stereo. arXiv preprint arXiv:1905.00538, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.73, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[28] Olga Karpenko, John F Hughes, and Ramesh Raskar. Freeform sketching with variational implicit surfaces. In Computer Graphics Forum, volume 21, pages 585-594. Wiley Online Library, 2002." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.787, + 0.892, + 0.853 + ], + "angle": 0, + "content": "[29] Mijeong Kim, Seonguk Seo, and Bohyung Han. Infonerf: Ray entropy minimization for few-shot neural volume rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12912-12921, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.856, + 0.892, + 0.911 + ], + "angle": 0, + "content": "[30] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.911 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17879" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[31] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.469, + 0.189 + ], + "angle": 0, + "content": "[32] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 31-42, 1996." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.191, + 0.469, + 0.244 + ], + "angle": 0, + "content": "[33] Tzu-Mao Li, Miika Aittala, Frédo Durand, and Jaakko Lehtinen. Differentiable monte carlo ray tracing through edge sampling. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.246, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[34] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.316, + 0.469, + 0.383 + ], + "angle": 0, + "content": "[35] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. arXiv preprint arXiv:2205.15723, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.385, + 0.469, + 0.439 + ], + "angle": 0, + "content": "[36] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.441, + 0.469, + 0.495 + ], + "angle": 0, + "content": "[37] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. arXiv preprint arXiv:2206.05737, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.497, + 0.469, + 0.537 + ], + "angle": 0, + "content": "[38] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.539, + 0.469, + 0.592 + ], + "angle": 0, + "content": "[39] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. arXiv preprint arXiv:2011.13495, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.594, + 0.469, + 0.662 + ], + "angle": 0, + "content": "[40] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.664, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[41] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.734, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[42] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.789, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[43] Zak Murez, Tarrence van As, James Bartolozzi, Ayan Sinha, Vijay Badrinarayanan, and Andrew Rabinovich. Atlas: End-to-end 3d scene reconstruction from posed images. In European conference on computer vision, pages 414-431. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.859, + 0.469, + 0.885 + ], + "angle": 0, + "content": "[44] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Reg-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.885 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.144 + ], + "angle": 0, + "content": "nerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.147, + 0.892, + 0.215 + ], + "angle": 0, + "content": "[45] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Occupancy flow: 4d reconstruction by learning particle dynamics. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5379-5389, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.218, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[46] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[47] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[48] Sergio Orts-Escolano, Christoph Rhemann, Sean Fanello, Wayne Chang, Adarsh Kowdle, Yury Degtyarev, David Kim, Philip L Davidson, Sameh Khamis, Mingsong Dou, et al. Holoportation: Virtual 3d teleportation in real-time. In Proceedings of the 29th annual symposium on user interface software and technology, pages 741-754, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.443, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[49] Stanley Osher, Ronald Fedkiw, and K Piechor. Level set methods and dynamic implicit surfaces. Appl. Mech. Rev., 57(3):B15-B15, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.486, + 0.892, + 0.554 + ], + "angle": 0, + "content": "[50] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.556, + 0.892, + 0.624 + ], + "angle": 0, + "content": "[51] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.627, + 0.892, + 0.695 + ], + "angle": 0, + "content": "[52] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), dec 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.697, + 0.892, + 0.751 + ], + "angle": 0, + "content": "[53] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.754, + 0.892, + 0.821 + ], + "angle": 0, + "content": "[54] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.824, + 0.892, + 0.878 + ], + "angle": 0, + "content": "[55] Yi-Ling Qiao, Alexander Gao, and Ming Lin. Neuphysics: Editable neural geometry and physics from monocular videos. In Advances in Neural Information Processing Systems, 2022." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.878 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17880" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[56] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.471, + 0.244 + ], + "angle": 0, + "content": "[57] Yufan Ren, Tong Zhang, Marc Pollefeys, Sabine Susstrunk, and Fangjinhua Wang. Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16685-16695, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.245, + 0.471, + 0.314 + ], + "angle": 0, + "content": "[58] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2304-2314, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.315, + 0.471, + 0.383 + ], + "angle": 0, + "content": "[59] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.384, + 0.471, + 0.452 + ], + "angle": 0, + "content": "[60] Miroslava Slavcheva, Maximilian Baust, Daniel Cremers, and Slobodan Ilic. Killingfusion: Non-rigid 3d reconstruction without correspondences. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1386-1395, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.453, + 0.471, + 0.521 + ], + "angle": 0, + "content": "[61] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. arXiv preprint arXiv:2210.15947, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.522, + 0.471, + 0.563 + ], + "angle": 0, + "content": "[62] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, volume 4, pages 109–116, 2007." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.564, + 0.471, + 0.605 + ], + "angle": 0, + "content": "[63] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. arXiv preprint arXiv:2111.11215, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.606, + 0.471, + 0.673 + ], + "angle": 0, + "content": "[64] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.674, + 0.471, + 0.702 + ], + "angle": 0, + "content": "[65] Jiaxiang Tang. Torch-ngp: a pytorch implementation of instant-ngp, 2022. https://github.com/ashawkey/torch-ngp." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.471, + 0.757 + ], + "angle": 0, + "content": "[66] Demetri Terzopoulos, John Platt, Alan Barr, and Kurt Fleischer. Elastically deformable models. In Proceedings of the 14th annual conference on Computer graphics and interactive techniques, pages 205-214, 1987." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.471, + 0.811 + ], + "angle": 0, + "content": "[67] Itsuki Ueda, Yoshihiro Fukuhara, Hirokatsu Kataoka, Hiroaki Aizawa, Hidehiko Shishido, and Itaru Kitahara. Neural density-distance fields. arXiv preprint arXiv:2207.14455, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.812, + 0.471, + 0.867 + ], + "angle": 0, + "content": "[68] Michael Waechter, Nils Moehrle, and Michael Goesele. Let there be color! large-scale texturing of 3d reconstructions. In European conference on computer vision, pages 836-850. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.867, + 0.471, + 0.907 + ], + "angle": 0, + "content": "[69] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.907 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[70] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[71] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[72] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.892, + 0.358 + ], + "angle": 0, + "content": "[73] Guandao Yang, Serge Belongie, Bharath Hariharan, and Vladlen Koltun. Geometry processing with neural fields. Advances in Neural Information Processing Systems, 34:22483-22497, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.36, + 0.892, + 0.415 + ], + "angle": 0, + "content": "[74] Gengshan Yang, Minh Vo, Neverova Natalia, Deva Ramanan, Vedaldi Andrea, and Joo Hanbyul. Banmo: Building animatable 3d neural models from many casual videos. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.417, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[75] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.474, + 0.892, + 0.514 + ], + "angle": 0, + "content": "[76] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. NeurIPS, 34:4805-4815, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.516, + 0.892, + 0.584 + ], + "angle": 0, + "content": "[77] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv preprint arXiv:2302.14859, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.586, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[78] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5336-5345, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.712 + ], + "angle": 0, + "content": "[79] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. arXiv preprint arXiv:2112.05131, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.714, + 0.892, + 0.769 + ], + "angle": 0, + "content": "[80] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.77, + 0.892, + 0.826 + ], + "angle": 0, + "content": "[81] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.827, + 0.892, + 0.883 + ], + "angle": 0, + "content": "[82] C Lawrence Zitnick, Sing Bing Kang, Matthew Uytendaele, Simon Winder, and Richard Szeliski. High-quality video view interpolation using a layered representation. ACM transactions on graphics (TOG), 23(3):600-608, 2004." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.883 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "17881" + } + ] +] \ No newline at end of file diff --git a/2023/Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction/11a08db4-fb69-460a-a705-16e05f3f48ac_origin.pdf b/2023/Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction/11a08db4-fb69-460a-a705-16e05f3f48ac_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5e71dcfae99de28d53f0b281e2b3f5846e9dd98e --- /dev/null +++ b/2023/Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction/11a08db4-fb69-460a-a705-16e05f3f48ac_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a5a26b708a4dd59ca4392f3a43d6605e6263ee23cc80034b2cd8b0d9d0767ed +size 6299704 diff --git a/2023/Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction/full.md b/2023/Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4304cc1e3d58ca64643b180dd019687dce48d319 --- /dev/null +++ b/2023/Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction/full.md @@ -0,0 +1,383 @@ +# Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction + +Jaesung Choe1,2 Christopher Choy1 Jaesik Park3 In So Kweon2 Anima Anandkumar1,4 + +$^{1}$ NVIDIA $^{2}$ KAIST $^{3}$ Seoul National University $^{4}$ Caltech + +# Abstract + +We propose an algorithm, 4DRegSDF, for the spacetime surface regularization to improve the fidelity of neural rendering and reconstruction in dynamic scenes. The key idea is to impose local rigidity on the deformable Signed Distance Function (SDF) for temporal coherency. Our approach works by (1) sampling points on the deformed surface by taking gradient steps toward the steepest direction along SDF, (2) extracting differential surface geometry, such as tangent plane or curvature, at each sample, and (3) adjusting the local rigidity at different timestamps. This enables our dynamic surface regularization to align 4D spacetime geometry via 3D canonical space more accurately. Experiments demonstrate that our 4DRegSDF achieves state-of-the-art performance in both reconstruction and rendering quality over synthetic and real-world datasets. https://4dregsdf.github.io/ + +# 1. Introduction + +Reconstructing 4D dynamic scenes from natural videos is one of the fundamental problems of vision tasks that are widely used in various fields, such as robotics, autonomous driving, virtual reality, and many more [6, 8, 14, 24, 32]. It requires an understanding of not just the geometry of scenes but also the texture, illumination, material properties, and refraction for photo-realistic rendering. However, because of the difficulty in reconstructing static 3D geometry alone, learning the dynamics jointly has been a challenge that most conventional methods decomposed the task into multiple stages of geometry reconstruction, motion representation, texture mapping, and illumination estimation [6, 15, 33, 49, 68]. Such sequential pipelines that first reconstruct geometry and learn motion separately are not robust to many failure modes of each stage, resulting in inaccurate reconstruction, which is one of the major bottlenecks for photo-realistic and geometry-oriented spacetime scene understanding. + +![](images/0f2f082e520d86b5597eb0dd5ae4d27cee65dc41cde50e5739e1eff907f3a5cc.jpg) +Rendered color +Rendered normal + +![](images/5b74dbd736e4a9097e03999c9d3fcfa1efda96139b06ec647c9a105bf73dbbe4.jpg) +Rendered color + +![](images/1916d5d3d8ef87533c53c8a0f7ea53f0123cf419a3e3aa240910836db787914b.jpg) +Rendered normal + +![](images/cb403aabc47411415bb512dca8c076d96f4b6ab198fe50bc37dd7e091b504c02.jpg) +Spacetime NeRF +Figure 1. We present an implicit surface regularization method that fully utilizes the properties of SDF for robust shape reconstruction and novel-view rendering in dynamic scenes. + +![](images/4618d5fdd3eb67f0eac60499df6af2d6063e1c99ce9833fe774d5034ce5d5c00.jpg) +4DRegSDF (ours) + +The recent advances in neural implicit representations, however, made accurate 3D reconstruction with differentiable volumetric rendering possible [40,41,46,50]. Mildenhall et al. [41] use sinusoidal encoding on the position as an input to multi-layered perceptrons (MLPs) to learn appearance and geometry of the target static scene. Specifically for photo-realistic rendering, neural radiance fields [4, 41, 54, 61, 63, 71, 79] mostly use density and color as the main output. However, the density representation can result in artifacts such as floaters and blurry boundaries since the density representation has high degrees of freedom [67, 70]. For accurate 3D reconstruction, the Signed Distance Function (SDF) started to gain traction for its high fidelity 3D surface reconstruction [26, 47, 70, 76, 81]. One of the key component for SDF learning is the Eikonol loss [25] which regularizes the SDF to have valid gradient. This acts as a strong regularization not just on the surface but on the entirety of the empty space. + +However, the 4D surface reconstruction from monocular RGB videos – which we reconstruct a dynamic scene – is still a highly ill-posed problem where we have to recon + +struct geometry and the dynamics of the scene simultaneously. Recent studies proposed to decompose the 4D space into a static geometry/texture (canonical model) and deformation to embed physical constraints into the 4D representation [51, 54]. The decomposition of the geometry could regularize the 4D space, but many papers use additional cues such as flow, or depth prediction [16, 21, 31, 34, 35, 69, 72, 78]. Still, the reconstruction is an ill-posed problem and it often fails for two reasons: (1) learning deformation is sensitive to topological variation and illumination change that often requires energy minimization terms for stable training. [28, 51, 60, 66]; (2) using the density representation – which has high degrees of freedom – could result in overfitting, floaters, and blurry reconstructions [29, 37, 44] as shown in Fig. 1. Thus, regularizing the spacetime domain is one of the most crucial parts of the 4D scene understanding in terms of geometry and graphics. + +In this paper, we propose an algorithm for spacetime surface regularization that improves the fidelity of neural rendering and reconstruction in dynamic scenes. The key idea is to enforce the local rigidity on the deformable Signed Distance Function (SDF) for the 4D geometry representation. This allows us to (1) regularize the reconstruction using more explicit surface representation instead of fuzzy density measure [41]; (2) sample geometry efficiently for surface regularizations. Specifically, previous studies [37,47,70] show that Signed Distance Function can reconstruct geometries accurately with fewer images since learning the distance to the nearest surface to represent a surface is strong regularizations compared to the density function. In addition, the signed distance encodes not just distance but direction to the closest surface from gradients, thus allowing efficient surface sampling. Then, the sampled points are used to minimize the deformation energy and align 4D spacetime geometry via 3D canonical space. + +We use three datasets to evaluate our method: one synthetic dataset from Pumarola et al. [54], and two real datasets: Park et al. [52] and Gao et al. [22]. Our method achieves state-of-the-art performance on both geometry representation and appearance as visualized in Fig. 1. We summarize our contributions as follows: + +- Extends the energy minimization in conventional 4D surface reconstruction studies into learning neural dynamic scene reconstruction. +- Regularize geometric properties of the 4D surface by the total variation of the curvature (TVC) and the absolute curvature of the SDF. +- Robust performance of rendering and reconstruction under the dynamic few-shot setup compared to recent implicit geometric regularization studies [20, 44]. + +# 2. Related works + +Conventional reconstruction and rendering. Video often contains time-varying properties, such as illumination change, geometry deformation, topology changes, and rigid or deformable motion. Such time-varying elements result in multi-view inconsistency among different frames and increase the level of difficulties of the video rendering task. For this issue, previous studies often require exhaustive multi-camera setup [5, 13, 48, 82] or synchronized multiple videos [2, 3]. None-learning based methods [1, 28, 60, 62] commonly assume the coarse geometry from RGB-D priors to introduce the energy regularization for the realistic geometry and photometric description of the target dynamic scenes. For instance, terzopoulos et al. [66] use elastic energy to deform the surface, and Slavcheva et al. [60] propose three typical energy minimization to prevent uncontrolled deformations. These ideas are related to the local rigidity to reconstruct plausible and geometricallymeaningful surface [1, 62] along the spatio-temporal domain. + +Neural implicit reconstruction and rendering. Recently, neural scene representation [9, 23, 40, 43, 45, 50, 53, 64] gets noticeable attention with their promising quality of implicit shape representations, such as occupancy or signed distance function. Not just geometry, the studies [36, 46, 58, 59] handle both geometry and appearance from multi-view images by neural implicit representation. In terms of the novelview synthesis task, Mildenhall et al. [41] demonstrates the impressive quality of rendered images. This paper takes a 5D radiance field (3D location and 2D viewing direction) to infer color and density. Then, it accumulates per-point inferences along each ray to determine pixel color. Such formulation makes it possible to train MLPs for learning geometry and appearance simultaneously. Based on this design, the series of following papers address the remaining problems of the pioneering paper [41], such as generalization [7, 57, 71], few-shot images [29, 44, 80], or surface reconstruction by SDF [70, 76, 77]. Nonetheless, these methods mostly focus on targeting static scenes without consideration of the moving objects or temporal change. + +Neural video rendering. More recently, several spacetime view synthesis studies [21,34,51,51,54,55,72] adopt neural implicit representations to render novel view synthesis from a video. Typically, these studies can be divided into two dominant strategies: deformable NeRFs [19,35,51,54] and prior-based video NeRFs [21, 34, 72]. Concurrent works [19, 35] propose fast deformable neural rendering methods by using voxel representations. Typically without relying on off-the-shelf information, such as monocular depth maps [34, 78], surface normal maps [81], or optical flow [35], we aim to render and reconstruct scenes purely from posed images. + +Regularization for neural implicit representation. Relying on the nature of the geometry, a series of papers [29, 44, 55] present the effectiveness of the regularization in learning neural fields. Under the few-shot setup, Niemeyer et al. [44] regularize the unobserved viewpoints via normalizing flow models. Kim et al. [29] apply entropy loss to alleviate reconstruction inconsistency. Fu et al. [20] apply multi-view consistency for the static surface reconstruction. In light of the growing popularity of spatio-temporal surface reconstruction and neural rendering, this paper proposes a novel approach to regulating the spacetime surface through the use of 3D canonical space. We have observed that the direct application of the SDF to this deformation field can result in low-fidelity reconstruction and rendering due to ill-posed problems in dynamic scenes. To overcome this issue, we highlight the importance of our regularization to achieve precise reconstruction and rendering. + +# 3. Preliminary + +Neural radiance field [41] allows fully differentiable learning of geometry and appearance by learning color and density from RGB images. Recent studies [42, 70] on surface reconstruction utilize Signed Distance Function (SDF) to represent scene geometry since The SDF representation has specific favorable characteristics that give it an edge for this particular objective. + +Differential surface geometry in SDF. The Signed Distance Function (SDF) stands as a predominant representation for implicit geometry. We can derive mathematical attributes on a crisp surface from SDF that could be challenging to derive from density-based representations: surface normal vector $\mathbf{n} = \frac{\partial s}{\partial\mathbf{p}}$ , tangent plane projection matrix $P = I - \mathbf{n}^{\top}\mathbf{n}$ , and surface curvature through the Hessian matrix $H = \frac{\partial^2s}{\partial\mathbf{p}^2}$ [73] Note that $I$ denotes the identity matrix, $s$ the SDF value, and $\mathbf{p}$ a 3D coordinate. We can derive these mathematical attributes because SDF represents the distance to the closest surface at every point. We can force this property for implicit geometry by using the Eikonal loss [25], which forces the norm of gradients of SDF $(\| \mathbf{n}\| _2 - 1)^2$ to be 1, and the loss has been widely used as an important regularization for 3D reconstruction methods [20,25,70,76]. + +NeuS [70] uses the same volumetric rendering equation in NeRF [41]. However, the opaqueness $\alpha$ is derived from the SDF value $s$ as $\alpha = 1 - \exp \left(-\int_{t_{i + 1}}^{t_i}\rho (t)dt\right)$ where $\rho$ is the density function $\rho = \max \left(\frac{s - \Delta s}{s + \Delta s},0\right)$ and $\Delta s$ is the displacement of the SDF. Following NeRF [41], NeuS utilizes the volumetric rendering as: + +$$ +\hat {C} (\mathbf {r}) = \sum_ {i = 1} ^ {N} T _ {i} \alpha_ {i} \mathbf {c} _ {i} \quad \text {s . t .} \quad T _ {i} = \exp \left(- \sum_ {j = 1} ^ {i - 1} \sigma_ {i} \delta_ {i}\right), \tag {1} +$$ + +where $\mathbf{c}_i$ denotes the color of the $i$ -th sample point $\mathbf{p}_i$ , $\alpha_i$ is the opaqueness over distance $\delta_i$ on the ray $\mathbf{r} = \{\mathbf{p}_i\}$ . $T_i$ is the cumulative transmittance along the ray. Based on this formulation, NeuS [70] is trained purely from posed images to describe the static scene surface using the SDF. + +Deformable NeRF [51, 54] is one way of novel-view synthesis methods in the 4D dynamic scenes. The idea is to decompose the 4D deformed space $[p; t]$ into the 3D canonical geometry $\mathbf{p}_c$ and the temporal deformation $\Delta \mathbf{p}$ as: + +$$ +\mathbf {p} _ {c} = \operatorname {D e f o r m} ([ \mathbf {p}; t ]) = \mathbf {p} + \Delta \mathbf {p}, \tag {2} +$$ + +where $\Delta \mathbf{p}(= h([p; t]))$ is the amount of space deformation, a function $h(\cdot)$ represents a deformation field, and the deformation field transports a point $[\mathbf{p}; t]$ to the canonical coordinate $\mathbf{p}_c = [x_c, y_c, z_c]$ . Although progress has been made in the field of dynamic neural rendering, we address the importance of the spacetime surface regularization for learning appearance, geometry, and motion in dynamic scenes as visualized in Fig. 1. + +# 4. Method + +This section presents our framework, 4DRegSDF. Our network represents the 4D spacetime domain as a combination of the static geometry in Signed Distance Function and the deformation field. Given a set of posed frames from a monocular RGB video, our network is trained to encode a dynamic scene (Sec. 4.1) and apply our spacetime surface regularization (Sec. 4.2). In contrast to previous video neural field papers that use priors such as depths or flows [21, 34, 35], our method relies only on posed image frames from a video as input to the system making the system simple and efficient. + +# 4.1. Deformable Signed Distance Function + +Given a point $\mathbf{p}$ at time $t$ within the deformed space and its viewing direction $\mathbf{d}$ , our model $f$ predicts the color $\mathbf{c}$ , the SDF value $s$ , and the deformation delta $\Delta \mathbf{p}$ . This can be expressed as: + +$$ +[ s; \Delta \mathbf {p}; \mathbf {c} ] = f ([ \mathbf {p}; t ], \mathbf {d}), \tag {3} +$$ + +As visualized in Fig. 3, we use the same volumetric rendering on the SDF (Eq. 1) [70] to get color value per each ray $\mathbf{r}$ , and minimize the photometric loss $\mathcal{L}_{\mathrm{c}}$ as, + +$$ +\mathcal {L} _ {\mathrm {c}} = \left| C (\mathbf {r}) - \hat {C} (\mathbf {r}) \right| _ {1}, \tag {4} +$$ + +where $C$ and $\hat{C}$ are ground truth color and predicted color, respectively, along a query ray $\mathbf{r}$ . + +# 4.2. Spacetime surface regularization + +Reconstructing dynamic surfaces from videos presents challenges due to its inherent complexity, often leading to + +![](images/f6bf193fcb8a1a2a0af4d6acc03a83fea53a2228916f472de9c910deb58b3438.jpg) +Figure 2. Surface sampling for our spacetime surface regularization in three stages. + +![](images/f34181a1533192b5e82bacae1d730e52e9892f601ebfc62bf1c6854b46b0f7a1.jpg) + +![](images/115b02b2fc9b314b2b6a8feeb9b8e40ee0ca95513cb2185f162d8914c5b0321a.jpg) + +![](images/8d07f33533eac7936dd5b99e7b9f3aabcc5a0c6fb7c07cb390cb99904a84369a.jpg) +Figure 3. Ray sampling for neural rendering [70]. + +visual artifacts as shown in Fig. 6. Accordingly, designing a robust regularization is one of the critical problems for successful 4D reconstruction [1,51,60,62,66,73]. + +In this section, we adopt three regularizations to enhance the learning of deformable Signed Distance Function from videos: 1. total variation of curvature, 2. the absolute curvature space-time, and 3. the Eikonal loss [25]. The first regularization aims to limit the amount of change in curvature induced by motion and the second regularization flattens out the 4D reconstruction and limits unnecessary kinks. The last regularization [25] enforces the gradient of SDF to be valid. First two regularizations are related to the properties of surface and can be computed only on the surface. We elaborate on the sampling strategy and how to augment the samples in continuous space and time, and lastly present the regularizations. Overall pipeline is visualized in Fig. 2. + +Spacetime surface sampling. The most commonly used sampling in neural fields is ray casting where you select points along a ray for rendering. Many regularizations proposed for NeRF also use samples along a ray [20,29,37,44]. However, surface constraints we adopt require samples on the surface of objects no along rays. For this, we use the definition of the SDF for surface sampling. + +The SDF is defined as the (signed) distance to the closest surface from the query point. In static scenes, the surface normal indicates the direction of the closest surface [12, 39, 73]. However, in 4D dynamic scenes, time is another component to determine the closest surface. In our setup, our network $f(\cdot)$ is trained to infer the SDF value $s$ from a query $[\mathbf{p}; t]$ . Accordingly, the gradient step towards the closest surface naturally involves and spatial update as + +Algorithm 1 Spacetime surface regularization +Input: Randomly initialized 4D point $[\mathbf{p}_i; t_i]$ . +Output: Total regularization $\mathcal{L}_e$ . +1: function 4DRegSDF([p_i, t_i]) +2: [p, t] ← [p_i, t_i] +3: for iterations do +4: [p, t] ← StepGrad([p; t]) ▷ Eq. 5 +5: end for +6: pc ← Deform([p; t]) ▷ Eq. 2 +7: t' ← Random(0, 1) ▷ Time augmentation +8: p' ← InvDeform([pc; t']) ▷ Eq. 6 +9: $\mathcal{L}_e \gets \text{ComputeLoss}([\mathbf{p}; t], [\mathbf{p}'; t'])$ ▷ Eq. 9 +10: return $\mathcal{L}_e$ +11: end function + +well as the time update. Given randomly initialized point $p_i$ at the time $t_i$ , each query point travels along the spacetime SDF 'hills' to be located at the isosurface as: + +$$ +[ \mathbf {p}; t ] = \operatorname {S t e p G r a d} ([ \mathbf {p} _ {i}; t _ {i} ]) = [ \mathbf {p} _ {i}; t _ {i} ] - s \cdot \frac {\partial s}{\partial [ \mathbf {p} _ {i} ; t _ {i} ]}, \tag {5} +$$ + +where $\frac{\partial s}{\partial[\mathbf{p};t]}$ is the gradient of the SDF value $s$ in terms of $[\mathbf{p};t]$ . By doing so, we can regularize the 4D surface in a wide-spectrum time domain including unobserved timestamps within training samples. Thus, by iteratively conducting this operation above, we move the initial 4D point $[\mathbf{p}_i; t_i]$ to $[\mathbf{p}; t]$ closer to the isosurface having smaller SDF value, as described in L. 2-5 of Alg. 1. + +Temporal augmentation. Given the surface samples $[\mathbf{p};t]$ from Eq. 5, we enforce the local rigidity to the surface samples for temporal coherency. First, we augment the time component $t$ into randomly selected $t^{\prime}$ as visualized in Fig. 2-(b). Specifically, we use the inverse-deformation network $g$ [35,45,74] which is formulated as: + +$$ +\mathbf {p} ^ {\prime} = \operatorname {I n v D e f o r m} \left(\left[ \mathbf {p} _ {c}; t ^ {\prime} \right]\right) = g \left(\mathbf {p} _ {c}, t ^ {\prime}\right), \tag {6} +$$ + +where $\mathbf{p}_c$ indicates the points at the canonical space from $[\mathbf{p};t]$ (Eq. 3). $\mathbf{p}'$ is a inverse-deformed point from $\mathbf{p}_c$ at the augmented time $t'$ as stated in L. 6-8 of Alg. 1. Through our + +temporal augmentation stage, we get a pair of two surface samples that correspond to each other, $([\mathbf{p};t],[\mathbf{p}';t'])$ . This pair will be used for evaluating spatio-temporal constraints the next section. + +Regularization on the spacetime curvature. Inspired by the conventional 4D reconstruction studies [1,60,62,66], we regularize geometric properties of dynamic surfaces: the total variation of curvature (TVC) and the absolute curvature of SDF surface on continuous time. The TVC measures the change of the curvature at different time steps and by minimizing the change, we assumes that surface should deform as little as possible while satisfying video observations. On the contrary, the absolute curvature measures curvature values and thus regularizes the complexity of the reconstruction and movement by minimizing the distortion in space. For the TVC, we adopt the stretching loss [73] where it minimizes the difference of inner products of two tangent vectors at two different time steps: $v_{1}, v_{2}$ at $t$ and $v_{1}', v_{2}'$ at $t'$ . The tangent vector is a remainder of a vector on surface $v = P\mathbf{p}$ where $P(= I - \mathbf{nn}^{\top})$ is the tangential projection matrix and $\mathbf{n}(= \frac{\partial s}{\partial\mathbf{p}})$ is the surface normal vector at a point $\mathbf{p}$ . Let $J_{1}, J_{2}$ be the Jacobians transforming $v_{1}, v_{2}$ to $v_{1}', v_{2}'$ respectively. Then, we compute the TVC as: + +$$ +\begin{array}{l} \left| v _ {1} ^ {\top} v _ {2} - v _ {1} ^ {\prime \top} v _ {2} \right| = \left| \mathbf {p} _ {1} ^ {\top} P _ {1} ^ {\top} P _ {2} \mathbf {p} _ {2} - \mathbf {p} _ {1} ^ {\top} P _ {1} ^ {\top} J _ {1} ^ {\top} J _ {2} P _ {2} \mathbf {p} _ {2} \right|, (7) \\ \propto \left| P _ {1} ^ {T} \left(I - J _ {1} ^ {T} J _ {2}\right) P _ {2} \right|, (8) \\ \end{array} +$$ + +The absolute curvature of the 4D SDF is the L1 norm of the Hessian matrix $H = \frac{\partial^2 s}{\partial \mathbf{p}^2}$ . By minimizing the absolute curvature, the reconstruction removes unnecessary curvatures or kinks in the reconstructions making the surface as smooth as possible. The total regularization is: + +$$ +\mathcal {L} _ {e} (\mathbf {p}, \hat {\mathbf {p}}) = \| P ^ {\top} (I - J _ {g} J _ {f}) P \| _ {2} + \| H \| _ {1}, \tag {9} +$$ + +where $J_{g}$ and $J_{f}$ are the Jacobians that transformed points via the deformation network $f(\cdot)$ and the inverse deformation network $g(\cdot)$ , respectively. + +We summarize our spacetime surface regularization in Algorithm 1. We achieve high-quality reconstruction results that are geometrically simple while they match video observations through neural rendering and minimizing the photometric loss. Finally, we train our network by combining the proposed losses as below, + +$$ +\mathcal {L} = \lambda_ {c} \mathcal {L} _ {c} + \lambda_ {c c} \mathcal {L} _ {c c} + \lambda_ {\mathrm {e}} \mathcal {L} _ {\mathrm {e}} + \lambda_ {\mathrm {e i k}} \mathcal {L} _ {\mathrm {e i k}}. \tag {10} +$$ + +where $\lambda_{c} = 1$ , $\lambda_{cc} = 0.01$ , $\lambda_{e} = 0.001$ , and $\lambda_{\mathrm{eik}} = 0.01$ . $\mathcal{L}_{\mathrm{eik}}$ represents the Eikonal loss [25] following NeuS [70]. Also, we enforce cycle consistency loss in learning forward/inverse deformation fields by $\mathcal{L}_{cc} = \| \mathbf{p} - g(\mathbf{p}_c,t)\| _1$ [35,45,74]. + +# 5. Experiments + +# 5.1. Novel-view synthesis + +Given a set of posed images, the novel-view synthesis aims to render images from viewpoints that are not in the part of the provided posed images. Since our 4DRegSDF also follows the neural rendering schemes, we compare the quality of rendered images with the recent spacetime NeRF papers on novel-view synthesis evaluation. + +Dataset. We utilize the Dynamic NeRF Synthetic dataset provided by D-NeRF [54] and the real-world scenes from HyperNeRF dataset [52]. These datasets offer natural monocular videos with dynamic frames and continuous camera motion. We follow the official train/test split for training and evaluation over ours and related studies. Following the conventions [41, 51, 54], we evaluate the rendering performance through PSNR, SSIM, and LPIPS. + +Comparison. We compare our method with a few state-of-the-art 4D NeRF papers, D-NeRF [54], TiNeuVox [19], which also do not rely on the prior information [21,34,35]. Also, we extend NeuS [70] for the 4D dynamic reconstruction task by adopting a deformation network from D-NeRF [54]. We denote this deformable NeuS as 'NeuS+D' and use it One of the concurrent works, DeVRF [35], also proposes a deformation-based dynamic neural rendering task. However, the paper requires additional multi-view static scenes for initialization so we did not include it in our experiment. + +When compared to existing research, our method achieves state-of-the-art performance among the recent spacetime NeRF methods as shown in Table 1 and Table 2. Many of these studies rely on view-dependent surface representations and the density function [41]. In contrast, our method utilizes the SDF representation, allowing us to more accurately represent geometry. This distinction influences the rendering quality. Notably, in terms of the qualitative reconstruction results, the previous state-of-the-art method struggles to accurately capture surface details as illustrated in Fig. 4 and Fig. 5. We suspect this limitation is rooted in the inherent issues of the density representation. As a result, many of the earlier spacetime NeRF studies place a heavy emphasis on the rendering quality. However, we claim that by imposing scene geometry constraints through our spacetime surface regularization, we can further improve the quality of the geometry under the precise geometry presentation, namely SDF. + +# 5.2. Scene reconstruction + +Dataset. Dycheck dataset [22] is one of the few datasets that provide ground truth geometry information for dynamic scenes. Yet, the ground truth depth maps are only provided for the training data split and we could not evalu + +
Dynamic Synthetic NeRF dataset [54]
Bouncing ballsHell warriorHookJumping jacksLegoMutantStand upT-rex
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
NeRF [41]20.260.9113.510.8116.650.8418.280.8820.300.7920.310.9118.190.8924.490.93
D-NeRF [54]38.930.9825.020.9529.250.9632.800.9821.640.8331.290.9632.790.9831.750.97
TiNeuVox [19]40.730.9928.170.9731.450.9734.230.9825.020.9233.610.9835.430.9932.700.98
TorchNGP [65]40.440.9924.540.9431.460.9731.450.9525.070.9335.860.9935.290.9931.350.97
NeuS+D [70]36.980.9623.770.9428.090.9428.280.9324.170.9030.170.9430.990.9524.940.93
Ours40.890.9925.720.9532.200.9833.630.9925.100.9431.830.9633.600.9933.640.99
+ +Table 1. Rendering evaluation in the D-NeRF synthetic dataset. We use PSNR (↑) and SSIM (↑) for the metric. + +
HyperNeRF dataset [52]
Broom3D PrinterChickenPeel-bananaMEAN
PSNRMS-SSIMPSNRMS-SSIMPSNRMS-SSIMPSNRMS-SSIMPSNRMS-SSIM
NeRF [41]19.90.65320.70.78019.90.77720.00.76920.10.745
NV [36]17.70.62316.20.66517.60.61515.90.38016.90.571
Nerfies [51]19.20.56720.60.83026.70.94322.40.87222.20.803
TorchNGP [65]19.80.57121.10.84127.60.94621.10.85123.50.815
TiNeuVox [19]21.50.68622.80.84128.30.94724.40.87324.30.837
NeuS+D [70]21.90.68922.10.83925.90.93324.60.87523.60.834
Ours23.30.70223.90.85929.80.96525.40.88125.30.851
+ +Table 2. Quantitative evaluation in the HyperNeRF dataset [52]. + +
Dycheck dataset [22]
CreeperHaru-sitPaper-windmillSriracha-tree
RMSE↓CD↓F1↑PSNR↑RMSE↓CD↓F1↑PSNR↑RMSE↓CD↓F1↑PSNR↑RMSE↓CD↓F1↑PSNR↑
TorchNGP [65]0.3400.05087.7817.320.5020.08176.3327.590.3110.05781.9823.30.3110.05281.4427.9
TiNeuVox [19]0.3450.04888.4216.690.6340.08075.1826.110.3130.05382.2924.960.3250.05781.5030.57
NeuS+D [70]0.3500.07174.1817.050.4910.11860.1327.880.2950.03891.3222.520.2830.04585.2326.83
Ours0.3230.06482.3617.570.4940.07778.5328.290.3100.05183.2424.260.2900.04292.2530.66
+ +Table 3. Evaluation in the Dycheck dataset [22] Note that CD (↓) means Chamfer Distance and F1 (↑) is F1-score. + +ate geometry reconstruction quality on the test split. Accordingly, we subsample the official train data to form a new train/test split. The new split consists of $75\%$ training data and $25\%$ test data per scene. For evaluation metrics, we follow the three geometric measures, RMSE, Chamfer Distance, and F1-score. RMSE is a widely-adopted metric for depthmap prediction evaluation [10, 17, 27, 56]. Chamfer distance [11, 18] measures the distance between the ground truth point cloud and the prediction. Meanwhile, the F1-score measures the 3D volume occupancy between the ground truth and the predictions [11, 30, 75]. + +Comparison. To evaluate geometric reconstruction quality, we evaluate on both view specific depth map prediction task and the 3D geometry reconstruction task. For depth map prediction task, we extract the depth map at the zero-crossing of SDF. For density-based NeRFs [19], we render depth maps by computing the expected depth. For the final metric, measure the RMSE in the neural volume rendering schemes [41, 70]. For 3D reconstruction evaluation, we follow the surface extraction pipeline provided by the official D-NeRF implementation for density-based methods [19, 54]. In our case, we extract mesh at the zero-crossings of SDF by using marching-cube algorithm [38] for 3D geometric evaluation. For evaluation metrics, we utilize the Chamfer Distance (CD ↓), F1-score (F1 ↑), and RMSE ↓. + +Based on this formulation, we calculate the 2D/3D geometric performance by ours and recent studies [19]. As + +described in Table 3, our method still consistently outperforms the quality of 4D surface reconstruction tasks. We will state the detailed ablation study in the following section, however, in short, the dominant performance improvements are driven by the plausible surface representation by Signed Distance Function and the proposed surface regularization schemes. Finally, our 4DRegSDF successfully renders and reconstructs the 4D dynamic scenes thanks to the plausible geometry representation and learning schemes + +# 6. Ablation studies + +In this section, we provide additional experiments to analyze each module of the 4DRegSDF. Note that additional ablation results are included in the supplementary material. + +Effectiveness of surface regularization. Surface regularization is the key to making 4D surfaces plausible shapes while learning deformation. When comparing with Fig. 6-(b) and (c), we found that $\mathcal{L}_e$ affects the structure of the reconstructed table. It implies that our surface regularization takes a role in shaping plausible structures by aligning geometry between surfaces at different times. + +Comparison with different regularization schemes. We compare our surface regularization with previous regularization schemes, e.g. RegNeRF [44] and GeoNeuS [20], as in Table 5. Based on these results, we claim that our spacetime surface regularization is particularly promising compared to previous regularization strategies. Also, we pro + +![](images/ab5472bfe0c6ecd21415f8034d40141266e2ded41ef4381fb984417a345b6670.jpg) +No ground truth mesh + +![](images/e4141a6296b9b8d91d8bee3c75c51be3537a251b90a938ae805b8af000d415b8.jpg) +TiNeuVox + +![](images/e49c82ab67e45f1862ff172791628da47cc2d0c55b4aec9b9f811590af923da6.jpg) +4DRegSDF (ours) + +![](images/8a7800dbc871e2302dd7558f46b2929d24a7aa1516c60d18f27e0147c4bde085.jpg) +Ground truth + +![](images/fc47b97938256ed6546117487c00ed6c5361f5dfc73f43f068df333a3f970ffc.jpg) + +![](images/34221338b772edfbd17852710e2303653c0d292ec7dc2b42fa3b4b469f789bb4.jpg) +TiNeuVox +Figure 5. Qualitative results in HyperNeRF dataset [52]. We compare previous state-of-the-art method, TiNeuVox [19], NeuS [70] with deformation [54] and our 4DRegSDF. Left: rendered color images, right: rendered surface normal maps. + +![](images/e187cccd275bf3392be5cc5eae0489de5ce4b2237989aa0d60c99cc8aed3b408.jpg) + +![](images/48f037968eb65f45d15b702c358d3cbe19b0afd46f5759cbdc061fb7ac43a0c4.jpg) +NeuS + Deformation + +![](images/a360fbe372e5aadff61eedc2646c56c3fb83cbb012b82972e11d7a493e38a795.jpg) + +![](images/2237fa65faf0d54bd8d570093cfb88d461c44918c86630c731100bf4ce66c5af.jpg) +4DRegSDF (ours) + +![](images/83b96ce0f298a6c21545436ad2bb2bcdc626510e573e78b8fc2bf971900ed3be.jpg) +Figure 4. Reconstruction results (mesh) with rendered images from our 4DRegSDF and previous state-of-the-art method, TiNeuVox [19]. + +![](images/c312ad9de4bacf42acc50785624e4bc67b38c2bb238f5f1ce22c39a41328dc15.jpg) +Ground truth + +vide details of our surface regularization in Table 6. Note that we calculate avg. PSNR from HyperNeRF dataset [52], and avg. RMSE from DyCheck dataset [22]. + +Deformation network design. Unlike deformable NeRF methods [51, 54] that sequentially process deformation and + +rendering (Eq. 2), our 4DRegSDF proposes to jointly learn deformation and SDF through a unified network $f$ (Eq. 3) as stated in Sec. 4.1. It turns out that our network design shows the better quality of surface normal maps as visualized in Fig. 6-(a) and (c). Such architecture design can + +![](images/6487d9e4362303f41d268f54cbe03f8643f510ea7ed12e20ede89646df7e6eec.jpg) +(a) NeuS + Deformation + +![](images/482fc6ba28a449987ec0f8c54cc7278413e362e3e1f1958aa246165b843acba7.jpg) +(b) 4DRegSDF w/o $\mathcal{L}_e$ +Figure 6. Qualitative comparison on (a) NeuS [70] combined with a deformable network [54], (b) ours without surface regularization, and our final 4DRegSDF method. Left: rendered color image, right: rendered surface normal map. + +![](images/5d1a1d8cf7cd9768f999b83423c37f8a3f7d3279b1866b6c6af1667f19a0f4e2.jpg) +(c) 4DRegSDF with $\mathcal{L}_e$ (ours) + +
Ratio of trained data
100% (Tables 2,3)25%
PSNRMS-SSIMCDPSNRMS-SSIMCD
TiNeuVox [19]22.80.8410.05718.10.7020.085
NeuS + D [70]22.10.8390.04519.40.7800.064
Ours w/o Reg.22.40.8390.04218.10.7700.060
Ours23.90.8590.04221.10.8520.058
+ +Table 4. Ablation study on few-shot frames. We choose the two scenes, 3D printer and Sriracha-tree, for the photometric and geometry evaluation. Ratio indicates the percentage of the data that we used for training the methods. + +
NetworkPreserve (✓)Evaluation
RegNeRFGeoNeuSOursPSNR ↑RMSE ↓
✓24.10.414
4DRegSDF✓24.70.384
✓25.30.354
+ +be one choice that reduces deformation ambiguity by learning jointly with SDF within a unified network $f$ and having cyclic deformation constraints $(\mathcal{L}_{\mathrm{cc}})$ . + +Sparse view. To further verify the benefit of our surface regularization schemes, we conduct additional experiments regarding the sparse input views [29, 44]. For experiments, we use '3D printer' and 'Sriracha-tree' for evaluation. As stated in Table 4, when the training data is reduced by $25\%$ , our method shows less performance drop when using a regularization scheme, showing less than $10\%$ performance drop. Such results consistently support the claim that regularization techniques are effective for learning radiance fields under the given sparse views [29, 44]. + +# 7. Conclusion + +We introduce using geometric spacetime regularizations for neural rendering and surface reconstruction in dynamic + +Table 5. Comparison of various regularization methods. RegNeRF [44] regularizes unobserved viewpoints and GeoNeuS [20] applies multi-view consistency as constraint. + +
NetworkPreserve (✓)Evaluation
TVCAbs4D samplePSNR ↑RMSE ↓
4DRegSDF✓✓24.40.392
✓✓24.90.379
✓✓✓25.30.354
+ +Table 6. Ablation study on regularization. We decompose $\mathcal{L}_e$ (Eq. 9) into total variation of curvature (TVC) and absolute curvature of SDF (ABS). 4D sample indicates Eq. 5. + +scenes from a monocular video. While previous dynamic NeRF methods have primarily targeted rendering quality, our approach places greater emphasis on providing accurate geometry representation via surface regularization. Our method has demonstrated promising performance in both dynamic novel-view synthesis and 4D surface reconstruction tasks. However, a notable constraint of this work is it limited capability in processing extended videos and we will extend this work for long format videos. + +Acknowledgement. This work was supported by IITP grants (No.2021-0-02068: AI Innovation Hub and No.2023R1A1C200781211) funded by the Korea government (MSIT). + +# References + +[1] Marc Alexa, Daniel Cohen-Or, and David Levin. As-rigid-as-possible shape interpolation. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 157–164, 2000. +[2] Luca Ballan, Gabriel J Brostow, Jens Puwein, and Marc Pollefeys. Unstructured video-based rendering: Interactive exploration of casually captured videos. In ACM SIGGRAPH 2010 papers, pages 1-11. 2010. +[3] Aayush Bansal, Minh Vo, Yaser Sheikh, Deva Ramanan, and Srinivasa Narasimhan. 4d visualization of dynamic events from unconstrained multi-view videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5366-5375, 2020. + +[4] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. +[5] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4):86-1, 2020. +[6] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph rendering. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 425-432, 2001. +[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. +[8] Shenchang Eric Chen and Lance Williams. View interpolation for image synthesis. In Proceedings of the 20th annual conference on Computer graphics and interactive techniques, pages 279-288, 1993. +[9] Jaesung Choe, Sunghoon Im, Francois Rameau, Minjun Kang, and In So Kweon. Volumefusion: Deep depth fusion for 3d scene reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16086-16095, 2021. +[10] Jaesung Choe, Kyungdon Joo, Tooba Imtiaz, and In So Kweon. Volumetric propagation network: Stereo-lidar fusion for long-range depth estimation. IEEE Robotics and Automation Letters, 6(3):4672-4679, 2021. +[11] Jaesung Choe, Byeongin Joung, Francois Rameau, Jaesik Park, and In So Kweon. Deep point cloud reconstruction. arXiv preprint arXiv:2111.11704, 2021. +[12] Gene Chou, Ilya Chugunov, and Felix Heide. Gensdf: Two-stage learning of generalizable signed distance functions. Advances in Neural Information Processing Systems, 35:24905-24919, 2022. +[13] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015. +[14] Paul Debevec, Yizhou Yu, and George Borshukov. Efficient view-dependent image-based rendering with projective texture-mapping. In Eurographics Workshop on Rendering Techniques, pages 105-116. Springer, 1998. +[15] Manfredo P Do Carmo. Differential geometry of curves and surfaces: revised and updated second edition. Courier Dover Publications, 2016. +[16] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. + +[17] Arda Duzceker, Silvano Galliani, Christoph Vogel, Pablo Speciale, Mihai Dusmanu, and Marc Pollefeys. Deepvideovs: Multi-view stereo on video with recurrent spatiotemporal fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15324-15333, 2021. +[18] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. +[19] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. arXiv preprint arXiv:2205.15285, 2022. +[20] Qiancheng Fu, Qingshan Xu, Yew-Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. arXiv preprint arXiv:2205.15848, 2022. +[21] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. +[22] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. arXiv preprint arXiv:2210.13445, 2022. +[23] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4857-4866, 2020. +[24] Steven J Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 43-54, 1996. +[25] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. +[26] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3d scene reconstruction with the manhattan-world assumption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5511-5520, 2022. +[27] Sunghoon Im, Hae-Gon Jeon, Stephen Lin, and In So Kweon. Dpsnet: End-to-end deep plane sweep stereo. arXiv preprint arXiv:1905.00538, 2019. +[28] Olga Karpenko, John F Hughes, and Ramesh Raskar. Freeform sketching with variational implicit surfaces. In Computer Graphics Forum, volume 21, pages 585-594. Wiley Online Library, 2002. +[29] Mijeong Kim, Seonguk Seo, and Bohyung Han. Infonerf: Ray entropy minimization for few-shot neural volume rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12912-12921, 2022. +[30] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. + +[31] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. +[32] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 31-42, 1996. +[33] Tzu-Mao Li, Miika Aittala, Frédo Durand, and Jaakko Lehtinen. Differentiable monte carlo ray tracing through edge sampling. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018. +[34] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. +[35] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. arXiv preprint arXiv:2205.15723, 2022. +[36] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. +[37] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. arXiv preprint arXiv:2206.05737, 2022. +[38] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. +[39] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. arXiv preprint arXiv:2011.13495, 2020. +[40] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. +[41] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. +[42] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. +[43] Zak Murez, Tarrence van As, James Bartolozzi, Ayan Sinha, Vijay Badrinarayanan, and Andrew Rabinovich. Atlas: End-to-end 3d scene reconstruction from posed images. In European conference on computer vision, pages 414-431. Springer, 2020. +[44] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Reg- + +nerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. +[45] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Occupancy flow: 4d reconstruction by learning particle dynamics. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5379-5389, 2019. +[46] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. +[47] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. +[48] Sergio Orts-Escolano, Christoph Rhemann, Sean Fanello, Wayne Chang, Adarsh Kowdle, Yury Degtyarev, David Kim, Philip L Davidson, Sameh Khamis, Mingsong Dou, et al. Holoportation: Virtual 3d teleportation in real-time. In Proceedings of the 29th annual symposium on user interface software and technology, pages 741-754, 2016. +[49] Stanley Osher, Ronald Fedkiw, and K Piechor. Level set methods and dynamic implicit surfaces. Appl. Mech. Rev., 57(3):B15-B15, 2004. +[50] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019. +[51] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. +[52] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), dec 2021. +[53] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. +[54] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. +[55] Yi-Ling Qiao, Alexander Gao, and Ming Lin. Neuphysics: Editable neural geometry and physics from monocular videos. In Advances in Neural Information Processing Systems, 2022. + +[56] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 2020. +[57] Yufan Ren, Tong Zhang, Marc Pollefeys, Sabine Susstrunk, and Fangjinhua Wang. Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16685-16695, 2023. +[58] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2304-2314, 2019. +[59] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019. +[60] Miroslava Slavcheva, Maximilian Baust, Daniel Cremers, and Slobodan Ilic. Killingfusion: Non-rigid 3d reconstruction without correspondences. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1386-1395, 2017. +[61] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. arXiv preprint arXiv:2210.15947, 2022. +[62] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, volume 4, pages 109–116, 2007. +[63] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. arXiv preprint arXiv:2111.11215, 2021. +[64] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. +[65] Jiaxiang Tang. Torch-ngp: a pytorch implementation of instant-ngp, 2022. https://github.com/ashawkey/torch-ngp. +[66] Demetri Terzopoulos, John Platt, Alan Barr, and Kurt Fleischer. Elastically deformable models. In Proceedings of the 14th annual conference on Computer graphics and interactive techniques, pages 205-214, 1987. +[67] Itsuki Ueda, Yoshihiro Fukuhara, Hirokatsu Kataoka, Hiroaki Aizawa, Hidehiko Shishido, and Itaru Kitahara. Neural density-distance fields. arXiv preprint arXiv:2207.14455, 2022. +[68] Michael Waechter, Nils Moehrle, and Michael Goesele. Let there be color! large-scale texturing of 3d reconstructions. In European conference on computer vision, pages 836-850. Springer, 2014. +[69] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994, 2021. + +[70] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. +[71] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. +[72] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. +[73] Guandao Yang, Serge Belongie, Bharath Hariharan, and Vladlen Koltun. Geometry processing with neural fields. Advances in Neural Information Processing Systems, 34:22483-22497, 2021. +[74] Gengshan Yang, Minh Vo, Neverova Natalia, Deva Ramanan, Vedaldi Andrea, and Joo Hanbyul. Banmo: Building animatable 3d neural models from many casual videos. In CVPR, 2022. +[75] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. +[76] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. NeurIPS, 34:4805-4815, 2021. +[77] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv preprint arXiv:2302.14859, 2023. +[78] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5336-5345, 2020. +[79] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. arXiv preprint arXiv:2112.05131, 2021. +[80] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. +[81] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022. +[82] C Lawrence Zitnick, Sing Bing Kang, Matthew Uytendaele, Simon Winder, and Richard Szeliski. High-quality video view interpolation using a layered representation. ACM transactions on graphics (TOG), 23(3):600-608, 2004. \ No newline at end of file diff --git a/2023/Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction/images.zip b/2023/Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a6b40fce5660deba9400ef73531592153b035643 --- /dev/null +++ b/2023/Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea4f243329c43b80a12b64a2ff7923d001376c8168a15bb00d40ea0396ed44c0 +size 798103 diff --git a/2023/Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction/layout.json b/2023/Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cb9a278d3d6871e0219a21d9ac85abb33dcef2cc --- /dev/null +++ b/2023/Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction/layout.json @@ -0,0 +1,10560 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "type": "text", + "content": "Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 152, + 517, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 152, + 517, + 168 + ], + "spans": [ + { + "bbox": [ + 75, + 152, + 517, + 168 + ], + "type": "text", + "content": "Jaesung Choe1,2 Christopher Choy1 Jaesik Park3 In So Kweon2 Anima Anandkumar1,4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 153, + 176, + 441, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 176, + 441, + 192 + ], + "spans": [ + { + "bbox": [ + 153, + 176, + 441, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 176, + 441, + 192 + ], + "type": "text", + "content": "NVIDIA " + }, + { + "bbox": [ + 153, + 176, + 441, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 176, + 441, + 192 + ], + "type": "text", + "content": "KAIST " + }, + { + "bbox": [ + 153, + 176, + 441, + 192 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 153, + 176, + 441, + 192 + ], + "type": "text", + "content": "Seoul National University " + }, + { + "bbox": [ + 153, + 176, + 441, + 192 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 153, + 176, + 441, + 192 + ], + "type": "text", + "content": "Caltech" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 252, + 290, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 445 + ], + "type": "text", + "content": "We propose an algorithm, 4DRegSDF, for the spacetime surface regularization to improve the fidelity of neural rendering and reconstruction in dynamic scenes. The key idea is to impose local rigidity on the deformable Signed Distance Function (SDF) for temporal coherency. Our approach works by (1) sampling points on the deformed surface by taking gradient steps toward the steepest direction along SDF, (2) extracting differential surface geometry, such as tangent plane or curvature, at each sample, and (3) adjusting the local rigidity at different timestamps. This enables our dynamic surface regularization to align 4D spacetime geometry via 3D canonical space more accurately. Experiments demonstrate that our 4DRegSDF achieves state-of-the-art performance in both reconstruction and rendering quality over synthetic and real-world datasets. https://4dregsdf.github.io/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 460, + 128, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 128, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 128, + 472 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 480, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 480, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 288, + 696 + ], + "type": "text", + "content": "Reconstructing 4D dynamic scenes from natural videos is one of the fundamental problems of vision tasks that are widely used in various fields, such as robotics, autonomous driving, virtual reality, and many more [6, 8, 14, 24, 32]. It requires an understanding of not just the geometry of scenes but also the texture, illumination, material properties, and refraction for photo-realistic rendering. However, because of the difficulty in reconstructing static 3D geometry alone, learning the dynamics jointly has been a challenge that most conventional methods decomposed the task into multiple stages of geometry reconstruction, motion representation, texture mapping, and illumination estimation [6, 15, 33, 49, 68]. Such sequential pipelines that first reconstruct geometry and learn motion separately are not robust to many failure modes of each stage, resulting in inaccurate reconstruction, which is one of the major bottlenecks for photo-realistic and geometry-oriented spacetime scene understanding." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 227, + 426, + 331 + ], + "blocks": [ + { + "bbox": [ + 307, + 227, + 426, + 331 + ], + "lines": [ + { + "bbox": [ + 307, + 227, + 426, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 227, + 426, + 331 + ], + "type": "image", + "image_path": "0f2f082e520d86b5597eb0dd5ae4d27cee65dc41cde50e5739e1eff907f3a5cc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 331, + 359, + 337 + ], + "lines": [ + { + "bbox": [ + 316, + 331, + 359, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 359, + 337 + ], + "type": "text", + "content": "Rendered color" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 371, + 331, + 419, + 337 + ], + "lines": [ + { + "bbox": [ + 371, + 331, + 419, + 337 + ], + "spans": [ + { + "bbox": [ + 371, + 331, + 419, + 337 + ], + "type": "text", + "content": "Rendered normal" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 365, + 227, + 441, + 331 + ], + "blocks": [ + { + "bbox": [ + 365, + 227, + 441, + 331 + ], + "lines": [ + { + "bbox": [ + 365, + 227, + 441, + 331 + ], + "spans": [ + { + "bbox": [ + 365, + 227, + 441, + 331 + ], + "type": "image", + "image_path": "5b74dbd736e4a9097e03999c9d3fcfa1efda96139b06ec647c9a105bf73dbbe4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 331, + 479, + 337 + ], + "lines": [ + { + "bbox": [ + 436, + 331, + 479, + 337 + ], + "spans": [ + { + "bbox": [ + 436, + 331, + 479, + 337 + ], + "type": "text", + "content": "Rendered color" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 428, + 227, + 546, + 331 + ], + "blocks": [ + { + "bbox": [ + 428, + 227, + 546, + 331 + ], + "lines": [ + { + "bbox": [ + 428, + 227, + 546, + 331 + ], + "spans": [ + { + "bbox": [ + 428, + 227, + 546, + 331 + ], + "type": "image", + "image_path": "1916d5d3d8ef87533c53c8a0f7ea53f0123cf419a3e3aa240910836db787914b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 491, + 331, + 539, + 337 + ], + "lines": [ + { + "bbox": [ + 491, + 331, + 539, + 337 + ], + "spans": [ + { + "bbox": [ + 491, + 331, + 539, + 337 + ], + "type": "text", + "content": "Rendered normal" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 306, + 338, + 422, + 398 + ], + "blocks": [ + { + "bbox": [ + 306, + 338, + 422, + 398 + ], + "lines": [ + { + "bbox": [ + 306, + 338, + 422, + 398 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 422, + 398 + ], + "type": "image", + "image_path": "cb403aabc47411415bb512dca8c076d96f4b6ab198fe50bc37dd7e091b504c02.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 398, + 390, + 407 + ], + "lines": [ + { + "bbox": [ + 343, + 398, + 390, + 407 + ], + "spans": [ + { + "bbox": [ + 343, + 398, + 390, + 407 + ], + "type": "text", + "content": "Spacetime NeRF" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 411, + 547, + 445 + ], + "lines": [ + { + "bbox": [ + 305, + 411, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 547, + 445 + ], + "type": "text", + "content": "Figure 1. We present an implicit surface regularization method that fully utilizes the properties of SDF for robust shape reconstruction and novel-view rendering in dynamic scenes." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 428, + 338, + 545, + 398 + ], + "blocks": [ + { + "bbox": [ + 428, + 338, + 545, + 398 + ], + "lines": [ + { + "bbox": [ + 428, + 338, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 428, + 338, + 545, + 398 + ], + "type": "image", + "image_path": "4618d5fdd3eb67f0eac60499df6af2d6063e1c99ce9833fe774d5034ce5d5c00.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 458, + 398, + 509, + 407 + ], + "lines": [ + { + "bbox": [ + 458, + 398, + 509, + 407 + ], + "spans": [ + { + "bbox": [ + 458, + 398, + 509, + 407 + ], + "type": "text", + "content": "4DRegSDF (ours)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 459, + 546, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 546, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 546, + 674 + ], + "type": "text", + "content": "The recent advances in neural implicit representations, however, made accurate 3D reconstruction with differentiable volumetric rendering possible [40,41,46,50]. Mildenhall et al. [41] use sinusoidal encoding on the position as an input to multi-layered perceptrons (MLPs) to learn appearance and geometry of the target static scene. Specifically for photo-realistic rendering, neural radiance fields [4, 41, 54, 61, 63, 71, 79] mostly use density and color as the main output. However, the density representation can result in artifacts such as floaters and blurry boundaries since the density representation has high degrees of freedom [67, 70]. For accurate 3D reconstruction, the Signed Distance Function (SDF) started to gain traction for its high fidelity 3D surface reconstruction [26, 47, 70, 76, 81]. One of the key component for SDF learning is the Eikonol loss [25] which regularizes the SDF to have valid gradient. This acts as a strong regularization not just on the surface but on the entirety of the empty space." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "However, the 4D surface reconstruction from monocular RGB videos – which we reconstruct a dynamic scene – is still a highly ill-posed problem where we have to recon" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 702, + 208, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 702, + 208, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 702, + 208, + 712 + ], + "type": "text", + "content": "Work done during the internship at NVIDIA." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17871" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "content": "struct geometry and the dynamics of the scene simultaneously. Recent studies proposed to decompose the 4D space into a static geometry/texture (canonical model) and deformation to embed physical constraints into the 4D representation [51, 54]. The decomposition of the geometry could regularize the 4D space, but many papers use additional cues such as flow, or depth prediction [16, 21, 31, 34, 35, 69, 72, 78]. Still, the reconstruction is an ill-posed problem and it often fails for two reasons: (1) learning deformation is sensitive to topological variation and illumination change that often requires energy minimization terms for stable training. [28, 51, 60, 66]; (2) using the density representation – which has high degrees of freedom – could result in overfitting, floaters, and blurry reconstructions [29, 37, 44] as shown in Fig. 1. Thus, regularizing the spacetime domain is one of the most crucial parts of the 4D scene understanding in terms of geometry and graphics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 278, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 289, + 495 + ], + "type": "text", + "content": "In this paper, we propose an algorithm for spacetime surface regularization that improves the fidelity of neural rendering and reconstruction in dynamic scenes. The key idea is to enforce the local rigidity on the deformable Signed Distance Function (SDF) for the 4D geometry representation. This allows us to (1) regularize the reconstruction using more explicit surface representation instead of fuzzy density measure [41]; (2) sample geometry efficiently for surface regularizations. Specifically, previous studies [37,47,70] show that Signed Distance Function can reconstruct geometries accurately with fewer images since learning the distance to the nearest surface to represent a surface is strong regularizations compared to the density function. In addition, the signed distance encodes not just distance but direction to the closest surface from gradients, thus allowing efficient surface sampling. Then, the sampled points are used to minimize the deformation energy and align 4D spacetime geometry via 3D canonical space." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 497, + 288, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 288, + 568 + ], + "type": "text", + "content": "We use three datasets to evaluate our method: one synthetic dataset from Pumarola et al. [54], and two real datasets: Park et al. [52] and Gao et al. [22]. Our method achieves state-of-the-art performance on both geometry representation and appearance as visualized in Fig. 1. We summarize our contributions as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 578, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 58, + 578, + 286, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 578, + 286, + 613 + ], + "spans": [ + { + "bbox": [ + 58, + 578, + 286, + 613 + ], + "type": "text", + "content": "- Extends the energy minimization in conventional 4D surface reconstruction studies into learning neural dynamic scene reconstruction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 628, + 286, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 628, + 286, + 663 + ], + "spans": [ + { + "bbox": [ + 58, + 628, + 286, + 663 + ], + "type": "text", + "content": "- Regularize geometric properties of the 4D surface by the total variation of the curvature (TVC) and the absolute curvature of the SDF." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 677, + 286, + 713 + ], + "type": "text", + "content": "- Robust performance of rendering and reconstruction under the dynamic few-shot setup compared to recent implicit geometric regularization studies [20, 44]." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 72, + 394, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 394, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 394, + 83 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 91, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 319 + ], + "type": "text", + "content": "Conventional reconstruction and rendering. Video often contains time-varying properties, such as illumination change, geometry deformation, topology changes, and rigid or deformable motion. Such time-varying elements result in multi-view inconsistency among different frames and increase the level of difficulties of the video rendering task. For this issue, previous studies often require exhaustive multi-camera setup [5, 13, 48, 82] or synchronized multiple videos [2, 3]. None-learning based methods [1, 28, 60, 62] commonly assume the coarse geometry from RGB-D priors to introduce the energy regularization for the realistic geometry and photometric description of the target dynamic scenes. For instance, terzopoulos et al. [66] use elastic energy to deform the surface, and Slavcheva et al. [60] propose three typical energy minimization to prevent uncontrolled deformations. These ideas are related to the local rigidity to reconstruct plausible and geometricallymeaningful surface [1, 62] along the spatio-temporal domain." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 324, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 545, + 564 + ], + "type": "text", + "content": "Neural implicit reconstruction and rendering. Recently, neural scene representation [9, 23, 40, 43, 45, 50, 53, 64] gets noticeable attention with their promising quality of implicit shape representations, such as occupancy or signed distance function. Not just geometry, the studies [36, 46, 58, 59] handle both geometry and appearance from multi-view images by neural implicit representation. In terms of the novelview synthesis task, Mildenhall et al. [41] demonstrates the impressive quality of rendered images. This paper takes a 5D radiance field (3D location and 2D viewing direction) to infer color and density. Then, it accumulates per-point inferences along each ray to determine pixel color. Such formulation makes it possible to train MLPs for learning geometry and appearance simultaneously. Based on this design, the series of following papers address the remaining problems of the pioneering paper [41], such as generalization [7, 57, 71], few-shot images [29, 44, 80], or surface reconstruction by SDF [70, 76, 77]. Nonetheless, these methods mostly focus on targeting static scenes without consideration of the moving objects or temporal change." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Neural video rendering. More recently, several spacetime view synthesis studies [21,34,51,51,54,55,72] adopt neural implicit representations to render novel view synthesis from a video. Typically, these studies can be divided into two dominant strategies: deformable NeRFs [19,35,51,54] and prior-based video NeRFs [21, 34, 72]. Concurrent works [19, 35] propose fast deformable neural rendering methods by using voxel representations. Typically without relying on off-the-shelf information, such as monocular depth maps [34, 78], surface normal maps [81], or optical flow [35], we aim to render and reconstruct scenes purely from posed images." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17872" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "content": "Regularization for neural implicit representation. Relying on the nature of the geometry, a series of papers [29, 44, 55] present the effectiveness of the regularization in learning neural fields. Under the few-shot setup, Niemeyer et al. [44] regularize the unobserved viewpoints via normalizing flow models. Kim et al. [29] apply entropy loss to alleviate reconstruction inconsistency. Fu et al. [20] apply multi-view consistency for the static surface reconstruction. In light of the growing popularity of spatio-temporal surface reconstruction and neural rendering, this paper proposes a novel approach to regulating the spacetime surface through the use of 3D canonical space. We have observed that the direct application of the SDF to this deformation field can result in low-fidelity reconstruction and rendering due to ill-posed problems in dynamic scenes. To overcome this issue, we highlight the importance of our regularization to achieve precise reconstruction and rendering." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 289, + 124, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 124, + 303 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 124, + 303 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 309, + 287, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 309, + 287, + 394 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 287, + 394 + ], + "type": "text", + "content": "Neural radiance field [41] allows fully differentiable learning of geometry and appearance by learning color and density from RGB images. Recent studies [42, 70] on surface reconstruction utilize Signed Distance Function (SDF) to represent scene geometry since The SDF representation has specific favorable characteristics that give it an edge for this particular objective." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 399, + 289, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 289, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 289, + 594 + ], + "type": "text", + "content": "Differential surface geometry in SDF. The Signed Distance Function (SDF) stands as a predominant representation for implicit geometry. We can derive mathematical attributes on a crisp surface from SDF that could be challenging to derive from density-based representations: surface normal vector " + }, + { + "bbox": [ + 46, + 399, + 289, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{n} = \\frac{\\partial s}{\\partial\\mathbf{p}}" + }, + { + "bbox": [ + 46, + 399, + 289, + 594 + ], + "type": "text", + "content": ", tangent plane projection matrix " + }, + { + "bbox": [ + 46, + 399, + 289, + 594 + ], + "type": "inline_equation", + "content": "P = I - \\mathbf{n}^{\\top}\\mathbf{n}" + }, + { + "bbox": [ + 46, + 399, + 289, + 594 + ], + "type": "text", + "content": ", and surface curvature through the Hessian matrix " + }, + { + "bbox": [ + 46, + 399, + 289, + 594 + ], + "type": "inline_equation", + "content": "H = \\frac{\\partial^2s}{\\partial\\mathbf{p}^2}" + }, + { + "bbox": [ + 46, + 399, + 289, + 594 + ], + "type": "text", + "content": " [73] Note that " + }, + { + "bbox": [ + 46, + 399, + 289, + 594 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 399, + 289, + 594 + ], + "type": "text", + "content": " denotes the identity matrix, " + }, + { + "bbox": [ + 46, + 399, + 289, + 594 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 399, + 289, + 594 + ], + "type": "text", + "content": " the SDF value, and " + }, + { + "bbox": [ + 46, + 399, + 289, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 399, + 289, + 594 + ], + "type": "text", + "content": " a 3D coordinate. We can derive these mathematical attributes because SDF represents the distance to the closest surface at every point. We can force this property for implicit geometry by using the Eikonal loss [25], which forces the norm of gradients of SDF " + }, + { + "bbox": [ + 46, + 399, + 289, + 594 + ], + "type": "inline_equation", + "content": "(\\| \\mathbf{n}\\| _2 - 1)^2" + }, + { + "bbox": [ + 46, + 399, + 289, + 594 + ], + "type": "text", + "content": " to be 1, and the loss has been widely used as an important regularization for 3D reconstruction methods [20,25,70,76]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 601, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 287, + 677 + ], + "type": "text", + "content": "NeuS [70] uses the same volumetric rendering equation in NeRF [41]. However, the opaqueness " + }, + { + "bbox": [ + 46, + 601, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 601, + 287, + 677 + ], + "type": "text", + "content": " is derived from the SDF value " + }, + { + "bbox": [ + 46, + 601, + 287, + 677 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 601, + 287, + 677 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 601, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\alpha = 1 - \\exp \\left(-\\int_{t_{i + 1}}^{t_i}\\rho (t)dt\\right)" + }, + { + "bbox": [ + 46, + 601, + 287, + 677 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 601, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 601, + 287, + 677 + ], + "type": "text", + "content": " is the density function " + }, + { + "bbox": [ + 46, + 601, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\rho = \\max \\left(\\frac{s - \\Delta s}{s + \\Delta s},0\\right)" + }, + { + "bbox": [ + 46, + 601, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 601, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\Delta s" + }, + { + "bbox": [ + 46, + 601, + 287, + 677 + ], + "type": "text", + "content": " is the displacement of the SDF. Following NeRF [41], NeuS utilizes the volumetric rendering as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 681, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 681, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 70, + 681, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\hat {C} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i} \\quad \\text {s . t .} \\quad T _ {i} = \\exp \\left(- \\sum_ {j = 1} ^ {i - 1} \\sigma_ {i} \\delta_ {i}\\right), \\tag {1}", + "image_path": "b02031195283bc5940fdfbf9e483d8ec3e54f38226579f4a7adb256322f7561e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " denotes the color of the " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "-th sample point " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\alpha_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " is the opaqueness over distance " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\delta_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " on the ray " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{r} = \\{\\mathbf{p}_i\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "T_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " is the cumulative transmittance along the ray. Based on this formulation, NeuS [70] is trained purely from posed images to describe the static scene surface using the SDF." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 137, + 545, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 137, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 304, + 137, + 545, + 186 + ], + "type": "text", + "content": "Deformable NeRF [51, 54] is one way of novel-view synthesis methods in the 4D dynamic scenes. The idea is to decompose the 4D deformed space " + }, + { + "bbox": [ + 304, + 137, + 545, + 186 + ], + "type": "inline_equation", + "content": "[p; t]" + }, + { + "bbox": [ + 304, + 137, + 545, + 186 + ], + "type": "text", + "content": " into the 3D canonical geometry " + }, + { + "bbox": [ + 304, + 137, + 545, + 186 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_c" + }, + { + "bbox": [ + 304, + 137, + 545, + 186 + ], + "type": "text", + "content": " and the temporal deformation " + }, + { + "bbox": [ + 304, + 137, + 545, + 186 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{p}" + }, + { + "bbox": [ + 304, + 137, + 545, + 186 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 358, + 194, + 545, + 207 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 194, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 358, + 194, + 545, + 207 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {c} = \\operatorname {D e f o r m} ([ \\mathbf {p}; t ]) = \\mathbf {p} + \\Delta \\mathbf {p}, \\tag {2}", + "image_path": "27dee5bddca0be04b4c1a4c23e7ac2bade8701ed1bb2b3df47921d28c91865d1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 213, + 545, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 545, + 310 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 213, + 545, + 310 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{p}(= h([p; t]))" + }, + { + "bbox": [ + 304, + 213, + 545, + 310 + ], + "type": "text", + "content": " is the amount of space deformation, a function " + }, + { + "bbox": [ + 304, + 213, + 545, + 310 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 304, + 213, + 545, + 310 + ], + "type": "text", + "content": " represents a deformation field, and the deformation field transports a point " + }, + { + "bbox": [ + 304, + 213, + 545, + 310 + ], + "type": "inline_equation", + "content": "[\\mathbf{p}; t]" + }, + { + "bbox": [ + 304, + 213, + 545, + 310 + ], + "type": "text", + "content": " to the canonical coordinate " + }, + { + "bbox": [ + 304, + 213, + 545, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_c = [x_c, y_c, z_c]" + }, + { + "bbox": [ + 304, + 213, + 545, + 310 + ], + "type": "text", + "content": ". Although progress has been made in the field of dynamic neural rendering, we address the importance of the spacetime surface regularization for learning appearance, geometry, and motion in dynamic scenes as visualized in Fig. 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 319, + 361, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 319, + 361, + 332 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 361, + 332 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 340, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 472 + ], + "type": "text", + "content": "This section presents our framework, 4DRegSDF. Our network represents the 4D spacetime domain as a combination of the static geometry in Signed Distance Function and the deformation field. Given a set of posed frames from a monocular RGB video, our network is trained to encode a dynamic scene (Sec. 4.1) and apply our spacetime surface regularization (Sec. 4.2). In contrast to previous video neural field papers that use priors such as depths or flows [21, 34, 35], our method relies only on posed image frames from a video as input to the system making the system simple and efficient." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 479, + 506, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 506, + 491 + ], + "type": "text", + "content": "4.1. Deformable Signed Distance Function" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "text", + "content": "Given a point " + }, + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "text", + "content": " within the deformed space and its viewing direction " + }, + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "text", + "content": ", our model " + }, + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "text", + "content": " predicts the color " + }, + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "text", + "content": ", the SDF value " + }, + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "text", + "content": ", and the deformation delta " + }, + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{p}" + }, + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "text", + "content": ". This can be expressed as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 375, + 552, + 545, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 552, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 375, + 552, + 545, + 566 + ], + "type": "interline_equation", + "content": "[ s; \\Delta \\mathbf {p}; \\mathbf {c} ] = f ([ \\mathbf {p}; t ], \\mathbf {d}), \\tag {3}", + "image_path": "d94eb3194989fc5f480615dd093cedf01823eef47cdf3366fd0afe7e4eee3bbe.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 573, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 573, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 573, + 545, + 609 + ], + "type": "text", + "content": "As visualized in Fig. 3, we use the same volumetric rendering on the SDF (Eq. 1) [70] to get color value per each ray " + }, + { + "bbox": [ + 304, + 573, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 304, + 573, + 545, + 609 + ], + "type": "text", + "content": ", and minimize the photometric loss " + }, + { + "bbox": [ + 304, + 573, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{c}}" + }, + { + "bbox": [ + 304, + 573, + 545, + 609 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 378, + 611, + 545, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 611, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 378, + 611, + 545, + 632 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c}} = \\left| C (\\mathbf {r}) - \\hat {C} (\\mathbf {r}) \\right| _ {1}, \\tag {4}", + "image_path": "54fd8f8ee5f27f6c93e7d176801c257eb7190a640aae82aa2e30625a6d8c099d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 639, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 639, + 545, + 665 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 639, + 545, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 639, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 304, + 639, + 545, + 665 + ], + "type": "text", + "content": " are ground truth color and predicted color, respectively, along a query ray " + }, + { + "bbox": [ + 304, + 639, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 304, + 639, + 545, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 671, + 482, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 482, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 482, + 684 + ], + "type": "text", + "content": "4.2. Spacetime surface regularization" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Reconstructing dynamic surfaces from videos presents challenges due to its inherent complexity, often leading to" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17873" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 193, + 170 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 193, + 170 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 193, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 193, + 170 + ], + "type": "image", + "image_path": "f6bf193fcb8a1a2a0af4d6acc03a83fea53a2228916f472de9c910deb58b3438.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 178, + 447, + 190 + ], + "lines": [ + { + "bbox": [ + 145, + 178, + 447, + 190 + ], + "spans": [ + { + "bbox": [ + 145, + 178, + 447, + 190 + ], + "type": "text", + "content": "Figure 2. Surface sampling for our spacetime surface regularization in three stages." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 201, + 70, + 367, + 171 + ], + "blocks": [ + { + "bbox": [ + 201, + 70, + 367, + 171 + ], + "lines": [ + { + "bbox": [ + 201, + 70, + 367, + 171 + ], + "spans": [ + { + "bbox": [ + 201, + 70, + 367, + 171 + ], + "type": "image", + "image_path": "f34181a1533192b5e82bacae1d730e52e9892f601ebfc62bf1c6854b46b0f7a1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 391, + 68, + 536, + 171 + ], + "blocks": [ + { + "bbox": [ + 391, + 68, + 536, + 171 + ], + "lines": [ + { + "bbox": [ + 391, + 68, + 536, + 171 + ], + "spans": [ + { + "bbox": [ + 391, + 68, + 536, + 171 + ], + "type": "image", + "image_path": "115b02b2fc9b314b2b6a8feeb9b8e40ee0ca95513cb2185f162d8914c5b0321a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 208, + 276, + 302 + ], + "blocks": [ + { + "bbox": [ + 57, + 208, + 276, + 302 + ], + "lines": [ + { + "bbox": [ + 57, + 208, + 276, + 302 + ], + "spans": [ + { + "bbox": [ + 57, + 208, + 276, + 302 + ], + "type": "image", + "image_path": "8d07f33533eac7936dd5b99e7b9f3aabcc5a0c6fb7c07cb390cb99904a84369a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 310, + 258, + 322 + ], + "lines": [ + { + "bbox": [ + 76, + 310, + 258, + 322 + ], + "spans": [ + { + "bbox": [ + 76, + 310, + 258, + 322 + ], + "type": "text", + "content": "Figure 3. Ray sampling for neural rendering [70]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 333, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 287, + 369 + ], + "type": "text", + "content": "visual artifacts as shown in Fig. 6. Accordingly, designing a robust regularization is one of the critical problems for successful 4D reconstruction [1,51,60,62,66,73]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 370, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 370, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 370, + 287, + 525 + ], + "type": "text", + "content": "In this section, we adopt three regularizations to enhance the learning of deformable Signed Distance Function from videos: 1. total variation of curvature, 2. the absolute curvature space-time, and 3. the Eikonal loss [25]. The first regularization aims to limit the amount of change in curvature induced by motion and the second regularization flattens out the 4D reconstruction and limits unnecessary kinks. The last regularization [25] enforces the gradient of SDF to be valid. First two regularizations are related to the properties of surface and can be computed only on the surface. We elaborate on the sampling strategy and how to augment the samples in continuous space and time, and lastly present the regularizations. Overall pipeline is visualized in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "content": "Spacetime surface sampling. The most commonly used sampling in neural fields is ray casting where you select points along a ray for rendering. Many regularizations proposed for NeRF also use samples along a ray [20,29,37,44]. However, surface constraints we adopt require samples on the surface of objects no along rays. For this, we use the definition of the SDF for surface sampling." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "The SDF is defined as the (signed) distance to the closest surface from the query point. In static scenes, the surface normal indicates the direction of the closest surface [12, 39, 73]. However, in 4D dynamic scenes, time is another component to determine the closest surface. In our setup, our network " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " is trained to infer the SDF value " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " from a query " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "[\\mathbf{p}; t]" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": ". Accordingly, the gradient step towards the closest surface naturally involves and spatial update as" + } + ] + } + ], + "index": 9 + }, + { + "type": "code", + "bbox": [ + 306, + 209, + 545, + 386 + ], + "blocks": [ + { + "bbox": [ + 306, + 209, + 545, + 386 + ], + "lines": [ + { + "bbox": [ + 306, + 209, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 306, + 209, + 545, + 386 + ], + "type": "text", + "content": "Algorithm 1 Spacetime surface regularization \nInput: Randomly initialized 4D point " + }, + { + "bbox": [ + 306, + 209, + 545, + 386 + ], + "type": "inline_equation", + "content": "[\\mathbf{p}_i; t_i]" + }, + { + "bbox": [ + 306, + 209, + 545, + 386 + ], + "type": "text", + "content": ". \nOutput: Total regularization " + }, + { + "bbox": [ + 306, + 209, + 545, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 306, + 209, + 545, + 386 + ], + "type": "text", + "content": ". \n1: function 4DRegSDF([p_i, t_i]) \n2: [p, t] ← [p_i, t_i] \n3: for iterations do \n4: [p, t] ← StepGrad([p; t]) ▷ Eq. 5 \n5: end for \n6: pc ← Deform([p; t]) ▷ Eq. 2 \n7: t' ← Random(0, 1) ▷ Time augmentation \n8: p' ← InvDeform([pc; t']) ▷ Eq. 6 \n9: " + }, + { + "bbox": [ + 306, + 209, + 545, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e \\gets \\text{ComputeLoss}([\\mathbf{p}; t], [\\mathbf{p}'; t'])" + }, + { + "bbox": [ + 306, + 209, + 545, + 386 + ], + "type": "text", + "content": " ▷ Eq. 9 \n10: return " + }, + { + "bbox": [ + 306, + 209, + 545, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 306, + 209, + 545, + 386 + ], + "type": "text", + "content": " \n11: end function" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_body" + } + ], + "index": 10, + "sub_type": "algorithm" + }, + { + "bbox": [ + 305, + 407, + 545, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 545, + 442 + ], + "type": "text", + "content": "well as the time update. Given randomly initialized point " + }, + { + "bbox": [ + 305, + 407, + 545, + 442 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 305, + 407, + 545, + 442 + ], + "type": "text", + "content": " at the time " + }, + { + "bbox": [ + 305, + 407, + 545, + 442 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 305, + 407, + 545, + 442 + ], + "type": "text", + "content": ", each query point travels along the spacetime SDF 'hills' to be located at the isosurface as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 448, + 545, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 448, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 315, + 448, + 545, + 475 + ], + "type": "interline_equation", + "content": "[ \\mathbf {p}; t ] = \\operatorname {S t e p G r a d} ([ \\mathbf {p} _ {i}; t _ {i} ]) = [ \\mathbf {p} _ {i}; t _ {i} ] - s \\cdot \\frac {\\partial s}{\\partial [ \\mathbf {p} _ {i} ; t _ {i} ]}, \\tag {5}", + "image_path": "aefc08d71fdd027057998170a1846f0e129c16181ae10d47c94255325b81356c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 482, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 545, + 569 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 482, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\frac{\\partial s}{\\partial[\\mathbf{p};t]}" + }, + { + "bbox": [ + 304, + 482, + 545, + 569 + ], + "type": "text", + "content": " is the gradient of the SDF value " + }, + { + "bbox": [ + 304, + 482, + 545, + 569 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 482, + 545, + 569 + ], + "type": "text", + "content": " in terms of " + }, + { + "bbox": [ + 304, + 482, + 545, + 569 + ], + "type": "inline_equation", + "content": "[\\mathbf{p};t]" + }, + { + "bbox": [ + 304, + 482, + 545, + 569 + ], + "type": "text", + "content": ". By doing so, we can regularize the 4D surface in a wide-spectrum time domain including unobserved timestamps within training samples. Thus, by iteratively conducting this operation above, we move the initial 4D point " + }, + { + "bbox": [ + 304, + 482, + 545, + 569 + ], + "type": "inline_equation", + "content": "[\\mathbf{p}_i; t_i]" + }, + { + "bbox": [ + 304, + 482, + 545, + 569 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 482, + 545, + 569 + ], + "type": "inline_equation", + "content": "[\\mathbf{p}; t]" + }, + { + "bbox": [ + 304, + 482, + 545, + 569 + ], + "type": "text", + "content": " closer to the isosurface having smaller SDF value, as described in L. 2-5 of Alg. 1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 574, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 574, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 574, + 545, + 647 + ], + "type": "text", + "content": "Temporal augmentation. Given the surface samples " + }, + { + "bbox": [ + 304, + 574, + 545, + 647 + ], + "type": "inline_equation", + "content": "[\\mathbf{p};t]" + }, + { + "bbox": [ + 304, + 574, + 545, + 647 + ], + "type": "text", + "content": " from Eq. 5, we enforce the local rigidity to the surface samples for temporal coherency. First, we augment the time component " + }, + { + "bbox": [ + 304, + 574, + 545, + 647 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 574, + 545, + 647 + ], + "type": "text", + "content": " into randomly selected " + }, + { + "bbox": [ + 304, + 574, + 545, + 647 + ], + "type": "inline_equation", + "content": "t^{\\prime}" + }, + { + "bbox": [ + 304, + 574, + 545, + 647 + ], + "type": "text", + "content": " as visualized in Fig. 2-(b). Specifically, we use the inverse-deformation network " + }, + { + "bbox": [ + 304, + 574, + 545, + 647 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 574, + 545, + 647 + ], + "type": "text", + "content": " [35,45,74] which is formulated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 348, + 655, + 545, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 655, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 348, + 655, + 545, + 670 + ], + "type": "interline_equation", + "content": "\\mathbf {p} ^ {\\prime} = \\operatorname {I n v D e f o r m} \\left(\\left[ \\mathbf {p} _ {c}; t ^ {\\prime} \\right]\\right) = g \\left(\\mathbf {p} _ {c}, t ^ {\\prime}\\right), \\tag {6}", + "image_path": "718a228aa97b80451c03890a1edc61b6df7cf87d3e2a7753022cc21da0a029e7.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_c" + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": " indicates the points at the canonical space from " + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "[\\mathbf{p};t]" + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": " (Eq. 3). " + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}'" + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": " is a inverse-deformed point from " + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_c" + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": " at the augmented time " + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": " as stated in L. 6-8 of Alg. 1. Through our" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17874" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "temporal augmentation stage, we get a pair of two surface samples that correspond to each other, " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "([\\mathbf{p};t],[\\mathbf{p}';t'])" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": ". This pair will be used for evaluating spatio-temporal constraints the next section." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "spans": [ + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "text", + "content": "Regularization on the spacetime curvature. Inspired by the conventional 4D reconstruction studies [1,60,62,66], we regularize geometric properties of dynamic surfaces: the total variation of curvature (TVC) and the absolute curvature of SDF surface on continuous time. The TVC measures the change of the curvature at different time steps and by minimizing the change, we assumes that surface should deform as little as possible while satisfying video observations. On the contrary, the absolute curvature measures curvature values and thus regularizes the complexity of the reconstruction and movement by minimizing the distortion in space. For the TVC, we adopt the stretching loss [73] where it minimizes the difference of inner products of two tangent vectors at two different time steps: " + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "inline_equation", + "content": "v_{1}, v_{2}" + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "inline_equation", + "content": "v_{1}', v_{2}'" + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "text", + "content": ". The tangent vector is a remainder of a vector on surface " + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "inline_equation", + "content": "v = P\\mathbf{p}" + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "inline_equation", + "content": "P(= I - \\mathbf{nn}^{\\top})" + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "text", + "content": " is the tangential projection matrix and " + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{n}(= \\frac{\\partial s}{\\partial\\mathbf{p}})" + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "text", + "content": " is the surface normal vector at a point " + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "inline_equation", + "content": "J_{1}, J_{2}" + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "text", + "content": " be the Jacobians transforming " + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "inline_equation", + "content": "v_{1}, v_{2}" + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "inline_equation", + "content": "v_{1}', v_{2}'" + }, + { + "bbox": [ + 46, + 127, + 289, + 356 + ], + "type": "text", + "content": " respectively. Then, we compute the TVC as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 365, + 287, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 365, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 61, + 365, + 287, + 396 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left| v _ {1} ^ {\\top} v _ {2} - v _ {1} ^ {\\prime \\top} v _ {2} \\right| = \\left| \\mathbf {p} _ {1} ^ {\\top} P _ {1} ^ {\\top} P _ {2} \\mathbf {p} _ {2} - \\mathbf {p} _ {1} ^ {\\top} P _ {1} ^ {\\top} J _ {1} ^ {\\top} J _ {2} P _ {2} \\mathbf {p} _ {2} \\right|, (7) \\\\ \\propto \\left| P _ {1} ^ {T} \\left(I - J _ {1} ^ {T} J _ {2}\\right) P _ {2} \\right|, (8) \\\\ \\end{array}", + "image_path": "090600a7bb785273c6a6459573f00e592dd726d97274bd80beffd2d7a7d840f9.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 410, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 287, + 471 + ], + "type": "text", + "content": "The absolute curvature of the 4D SDF is the L1 norm of the Hessian matrix " + }, + { + "bbox": [ + 46, + 410, + 287, + 471 + ], + "type": "inline_equation", + "content": "H = \\frac{\\partial^2 s}{\\partial \\mathbf{p}^2}" + }, + { + "bbox": [ + 46, + 410, + 287, + 471 + ], + "type": "text", + "content": ". By minimizing the absolute curvature, the reconstruction removes unnecessary curvatures or kinks in the reconstructions making the surface as smooth as possible. The total regularization is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 480, + 287, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 480, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 80, + 480, + 287, + 496 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {e} (\\mathbf {p}, \\hat {\\mathbf {p}}) = \\| P ^ {\\top} (I - J _ {g} J _ {f}) P \\| _ {2} + \\| H \\| _ {1}, \\tag {9}", + "image_path": "04fcfac4bdebf837608baf20fd98721d10df80533bc402652fb36a6b1bd840fe.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 507, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 287, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 507, + 287, + 544 + ], + "type": "inline_equation", + "content": "J_{g}" + }, + { + "bbox": [ + 46, + 507, + 287, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 507, + 287, + 544 + ], + "type": "inline_equation", + "content": "J_{f}" + }, + { + "bbox": [ + 46, + 507, + 287, + 544 + ], + "type": "text", + "content": " are the Jacobians that transformed points via the deformation network " + }, + { + "bbox": [ + 46, + 507, + 287, + 544 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 46, + 507, + 287, + 544 + ], + "type": "text", + "content": " and the inverse deformation network " + }, + { + "bbox": [ + 46, + 507, + 287, + 544 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 46, + 507, + 287, + 544 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "content": "We summarize our spacetime surface regularization in Algorithm 1. We achieve high-quality reconstruction results that are geometrically simple while they match video observations through neural rendering and minimizing the photometric loss. Finally, we train our network by combining the proposed losses as below," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 85, + 628, + 287, + 642 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 628, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 85, + 628, + 287, + 642 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {c c} \\mathcal {L} _ {c c} + \\lambda_ {\\mathrm {e}} \\mathcal {L} _ {\\mathrm {e}} + \\lambda_ {\\mathrm {e i k}} \\mathcal {L} _ {\\mathrm {e i k}}. \\tag {10}", + "image_path": "aa12602d69a6b265f0e8b60e61996e9241e7cc0f748383a7ffe715c76b90f4ef.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{c} = 1" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{cc} = 0.01" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{e} = 0.001" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{eik}} = 0.01" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{eik}}" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " represents the Eikonal loss [25] following NeuS [70]. Also, we enforce cycle consistency loss in learning forward/inverse deformation fields by " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cc} = \\| \\mathbf{p} - g(\\mathbf{p}_c,t)\\| _1" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " [35,45,74]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 91, + 425, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 425, + 104 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 425, + 104 + ], + "type": "text", + "content": "5.1. Novel-view synthesis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 110, + 545, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 545, + 182 + ], + "type": "text", + "content": "Given a set of posed images, the novel-view synthesis aims to render images from viewpoints that are not in the part of the provided posed images. Since our 4DRegSDF also follows the neural rendering schemes, we compare the quality of rendered images with the recent spacetime NeRF papers on novel-view synthesis evaluation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 188, + 545, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 188, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 304, + 188, + 545, + 284 + ], + "type": "text", + "content": "Dataset. We utilize the Dynamic NeRF Synthetic dataset provided by D-NeRF [54] and the real-world scenes from HyperNeRF dataset [52]. These datasets offer natural monocular videos with dynamic frames and continuous camera motion. We follow the official train/test split for training and evaluation over ours and related studies. Following the conventions [41, 51, 54], we evaluate the rendering performance through PSNR, SSIM, and LPIPS." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 290, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 422 + ], + "type": "text", + "content": "Comparison. We compare our method with a few state-of-the-art 4D NeRF papers, D-NeRF [54], TiNeuVox [19], which also do not rely on the prior information [21,34,35]. Also, we extend NeuS [70] for the 4D dynamic reconstruction task by adopting a deformation network from D-NeRF [54]. We denote this deformable NeuS as 'NeuS+D' and use it One of the concurrent works, DeVRF [35], also proposes a deformation-based dynamic neural rendering task. However, the paper requires additional multi-view static scenes for initialization so we did not include it in our experiment." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 422, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 638 + ], + "type": "text", + "content": "When compared to existing research, our method achieves state-of-the-art performance among the recent spacetime NeRF methods as shown in Table 1 and Table 2. Many of these studies rely on view-dependent surface representations and the density function [41]. In contrast, our method utilizes the SDF representation, allowing us to more accurately represent geometry. This distinction influences the rendering quality. Notably, in terms of the qualitative reconstruction results, the previous state-of-the-art method struggles to accurately capture surface details as illustrated in Fig. 4 and Fig. 5. We suspect this limitation is rooted in the inherent issues of the density representation. As a result, many of the earlier spacetime NeRF studies place a heavy emphasis on the rendering quality. However, we claim that by imposing scene geometry constraints through our spacetime surface regularization, we can further improve the quality of the geometry under the precise geometry presentation, namely SDF." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 647, + 426, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 426, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 426, + 658 + ], + "type": "text", + "content": "5.2. Scene reconstruction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Dataset. Dycheck dataset [22] is one of the few datasets that provide ground truth geometry information for dynamic scenes. Yet, the ground truth depth maps are only provided for the training data split and we could not evalu" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17875" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 547, + 145 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 145 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 145 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 145 + ], + "type": "table", + "html": "
Dynamic Synthetic NeRF dataset [54]
Bouncing ballsHell warriorHookJumping jacksLegoMutantStand upT-rex
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
NeRF [41]20.260.9113.510.8116.650.8418.280.8820.300.7920.310.9118.190.8924.490.93
D-NeRF [54]38.930.9825.020.9529.250.9632.800.9821.640.8331.290.9632.790.9831.750.97
TiNeuVox [19]40.730.9928.170.9731.450.9734.230.9825.020.9233.610.9835.430.9932.700.98
TorchNGP [65]40.440.9924.540.9431.460.9731.450.9525.070.9335.860.9935.290.9931.350.97
NeuS+D [70]36.980.9623.770.9428.090.9428.280.9324.170.9030.170.9430.990.9524.940.93
Ours40.890.9925.720.9532.200.9833.630.9925.100.9431.830.9633.600.9933.640.99
", + "image_path": "85bea15b6b1d6b300a3a8c905b01a9065d6ac9e1929a336c648a9a351b984aed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 90, + 168, + 504, + 258 + ], + "blocks": [ + { + "bbox": [ + 95, + 147, + 497, + 158 + ], + "lines": [ + { + "bbox": [ + 95, + 147, + 497, + 158 + ], + "spans": [ + { + "bbox": [ + 95, + 147, + 497, + 158 + ], + "type": "text", + "content": "Table 1. Rendering evaluation in the D-NeRF synthetic dataset. We use PSNR (↑) and SSIM (↑) for the metric." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 90, + 168, + 504, + 258 + ], + "lines": [ + { + "bbox": [ + 90, + 168, + 504, + 258 + ], + "spans": [ + { + "bbox": [ + 90, + 168, + 504, + 258 + ], + "type": "table", + "html": "
HyperNeRF dataset [52]
Broom3D PrinterChickenPeel-bananaMEAN
PSNRMS-SSIMPSNRMS-SSIMPSNRMS-SSIMPSNRMS-SSIMPSNRMS-SSIM
NeRF [41]19.90.65320.70.78019.90.77720.00.76920.10.745
NV [36]17.70.62316.20.66517.60.61515.90.38016.90.571
Nerfies [51]19.20.56720.60.83026.70.94322.40.87222.20.803
TorchNGP [65]19.80.57121.10.84127.60.94621.10.85123.50.815
TiNeuVox [19]21.50.68622.80.84128.30.94724.40.87324.30.837
NeuS+D [70]21.90.68922.10.83925.90.93324.60.87523.60.834
Ours23.30.70223.90.85929.80.96525.40.88125.30.851
", + "image_path": "3649dbfbc24fc21efff7fa28e8c05e8f77b13e054ffbfce14dc83192fd5ab478.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 49, + 279, + 547, + 340 + ], + "blocks": [ + { + "bbox": [ + 177, + 258, + 412, + 270 + ], + "lines": [ + { + "bbox": [ + 177, + 258, + 412, + 270 + ], + "spans": [ + { + "bbox": [ + 177, + 258, + 412, + 270 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation in the HyperNeRF dataset [52]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 279, + 547, + 340 + ], + "lines": [ + { + "bbox": [ + 49, + 279, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 49, + 279, + 547, + 340 + ], + "type": "table", + "html": "
Dycheck dataset [22]
CreeperHaru-sitPaper-windmillSriracha-tree
RMSE↓CD↓F1↑PSNR↑RMSE↓CD↓F1↑PSNR↑RMSE↓CD↓F1↑PSNR↑RMSE↓CD↓F1↑PSNR↑
TorchNGP [65]0.3400.05087.7817.320.5020.08176.3327.590.3110.05781.9823.30.3110.05281.4427.9
TiNeuVox [19]0.3450.04888.4216.690.6340.08075.1826.110.3130.05382.2924.960.3250.05781.5030.57
NeuS+D [70]0.3500.07174.1817.050.4910.11860.1327.880.2950.03891.3222.520.2830.04585.2326.83
Ours0.3230.06482.3617.570.4940.07778.5328.290.3100.05183.2424.260.2900.04292.2530.66
", + "image_path": "f1cc03519a5e90789c5e6be7c6e43bb3f49acf991e8416ab629e155bf0dd4220.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 342, + 502, + 352 + ], + "lines": [ + { + "bbox": [ + 86, + 342, + 502, + 352 + ], + "spans": [ + { + "bbox": [ + 86, + 342, + 502, + 352 + ], + "type": "text", + "content": "Table 3. Evaluation in the Dycheck dataset [22] Note that CD (↓) means Chamfer Distance and F1 (↑) is F1-score." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 366, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 498 + ], + "type": "text", + "content": "ate geometry reconstruction quality on the test split. Accordingly, we subsample the official train data to form a new train/test split. The new split consists of " + }, + { + "bbox": [ + 46, + 366, + 287, + 498 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 366, + 287, + 498 + ], + "type": "text", + "content": " training data and " + }, + { + "bbox": [ + 46, + 366, + 287, + 498 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 366, + 287, + 498 + ], + "type": "text", + "content": " test data per scene. For evaluation metrics, we follow the three geometric measures, RMSE, Chamfer Distance, and F1-score. RMSE is a widely-adopted metric for depthmap prediction evaluation [10, 17, 27, 56]. Chamfer distance [11, 18] measures the distance between the ground truth point cloud and the prediction. Meanwhile, the F1-score measures the 3D volume occupancy between the ground truth and the predictions [11, 30, 75]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 507, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 287, + 685 + ], + "type": "text", + "content": "Comparison. To evaluate geometric reconstruction quality, we evaluate on both view specific depth map prediction task and the 3D geometry reconstruction task. For depth map prediction task, we extract the depth map at the zero-crossing of SDF. For density-based NeRFs [19], we render depth maps by computing the expected depth. For the final metric, measure the RMSE in the neural volume rendering schemes [41, 70]. For 3D reconstruction evaluation, we follow the surface extraction pipeline provided by the official D-NeRF implementation for density-based methods [19, 54]. In our case, we extract mesh at the zero-crossings of SDF by using marching-cube algorithm [38] for 3D geometric evaluation. For evaluation metrics, we utilize the Chamfer Distance (CD ↓), F1-score (F1 ↑), and RMSE ↓." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Based on this formulation, we calculate the 2D/3D geometric performance by ours and recent studies [19]. As" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 366, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 475 + ], + "type": "text", + "content": "described in Table 3, our method still consistently outperforms the quality of 4D surface reconstruction tasks. We will state the detailed ablation study in the following section, however, in short, the dominant performance improvements are driven by the plausible surface representation by Signed Distance Function and the proposed surface regularization schemes. Finally, our 4DRegSDF successfully renders and reconstructs the 4D dynamic scenes thanks to the plausible geometry representation and learning schemes" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 488, + 404, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 488, + 404, + 500 + ], + "spans": [ + { + "bbox": [ + 306, + 488, + 404, + 500 + ], + "type": "text", + "content": "6. Ablation studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 509, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 544 + ], + "type": "text", + "content": "In this section, we provide additional experiments to analyze each module of the 4DRegSDF. Note that additional ablation results are included in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 552, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 635 + ], + "type": "text", + "content": "Effectiveness of surface regularization. Surface regularization is the key to making 4D surfaces plausible shapes while learning deformation. When comparing with Fig. 6-(b) and (c), we found that " + }, + { + "bbox": [ + 304, + 552, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 304, + 552, + 545, + 635 + ], + "type": "text", + "content": " affects the structure of the reconstructed table. It implies that our surface regularization takes a role in shaping plausible structures by aligning geometry between surfaces at different times." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": "Comparison with different regularization schemes. We compare our surface regularization with previous regularization schemes, e.g. RegNeRF [44] and GeoNeuS [20], as in Table 5. Based on these results, we claim that our spacetime surface regularization is particularly promising compared to previous regularization strategies. Also, we pro" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "17876" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 79, + 403, + 186 + ], + "blocks": [ + { + "bbox": [ + 50, + 79, + 403, + 186 + ], + "lines": [ + { + "bbox": [ + 50, + 79, + 403, + 186 + ], + "spans": [ + { + "bbox": [ + 50, + 79, + 403, + 186 + ], + "type": "image", + "image_path": "ab5472bfe0c6ecd21415f8034d40141266e2ded41ef4381fb984417a345b6670.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 132, + 504, + 143 + ], + "lines": [ + { + "bbox": [ + 423, + 132, + 504, + 143 + ], + "spans": [ + { + "bbox": [ + 423, + 132, + 504, + 143 + ], + "type": "text", + "content": "No ground truth mesh" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 187, + 206, + 284 + ], + "blocks": [ + { + "bbox": [ + 53, + 187, + 206, + 284 + ], + "lines": [ + { + "bbox": [ + 53, + 187, + 206, + 284 + ], + "spans": [ + { + "bbox": [ + 53, + 187, + 206, + 284 + ], + "type": "image", + "image_path": "e4141a6296b9b8d91d8bee3c75c51be3537a251b90a938ae805b8af000d415b8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 284, + 153, + 293 + ], + "lines": [ + { + "bbox": [ + 113, + 284, + 153, + 293 + ], + "spans": [ + { + "bbox": [ + 113, + 284, + 153, + 293 + ], + "type": "text", + "content": "TiNeuVox" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 187, + 375, + 282 + ], + "blocks": [ + { + "bbox": [ + 225, + 187, + 375, + 282 + ], + "lines": [ + { + "bbox": [ + 225, + 187, + 375, + 282 + ], + "spans": [ + { + "bbox": [ + 225, + 187, + 375, + 282 + ], + "type": "image", + "image_path": "e49c82ab67e45f1862ff172791628da47cc2d0c55b4aec9b9f811590af923da6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 284, + 336, + 294 + ], + "lines": [ + { + "bbox": [ + 269, + 284, + 336, + 294 + ], + "spans": [ + { + "bbox": [ + 269, + 284, + 336, + 294 + ], + "type": "text", + "content": "4DRegSDF (ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 388, + 186, + 539, + 282 + ], + "blocks": [ + { + "bbox": [ + 388, + 186, + 539, + 282 + ], + "lines": [ + { + "bbox": [ + 388, + 186, + 539, + 282 + ], + "spans": [ + { + "bbox": [ + 388, + 186, + 539, + 282 + ], + "type": "image", + "image_path": "8a7800dbc871e2302dd7558f46b2929d24a7aa1516c60d18f27e0147c4bde085.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 284, + 488, + 293 + ], + "lines": [ + { + "bbox": [ + 440, + 284, + 488, + 293 + ], + "spans": [ + { + "bbox": [ + 440, + 284, + 488, + 293 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 48, + 342, + 187, + 466 + ], + "blocks": [ + { + "bbox": [ + 48, + 342, + 187, + 466 + ], + "lines": [ + { + "bbox": [ + 48, + 342, + 187, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 187, + 466 + ], + "type": "image", + "image_path": "fc47b97938256ed6546117487c00ed6c5361f5dfc73f43f068df333a3f970ffc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 48, + 468, + 187, + 590 + ], + "blocks": [ + { + "bbox": [ + 48, + 468, + 187, + 590 + ], + "lines": [ + { + "bbox": [ + 48, + 468, + 187, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 187, + 590 + ], + "type": "image", + "image_path": "34221338b772edfbd17852710e2303653c0d292ec7dc2b42fa3b4b469f789bb4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 591, + 135, + 599 + ], + "lines": [ + { + "bbox": [ + 100, + 591, + 135, + 599 + ], + "spans": [ + { + "bbox": [ + 100, + 591, + 135, + 599 + ], + "type": "text", + "content": "TiNeuVox" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 605, + 547, + 628 + ], + "lines": [ + { + "bbox": [ + 46, + 605, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 547, + 628 + ], + "type": "text", + "content": "Figure 5. Qualitative results in HyperNeRF dataset [52]. We compare previous state-of-the-art method, TiNeuVox [19], NeuS [70] with deformation [54] and our 4DRegSDF. Left: rendered color images, right: rendered surface normal maps." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 190, + 342, + 331, + 466 + ], + "blocks": [ + { + "bbox": [ + 190, + 342, + 331, + 466 + ], + "lines": [ + { + "bbox": [ + 190, + 342, + 331, + 466 + ], + "spans": [ + { + "bbox": [ + 190, + 342, + 331, + 466 + ], + "type": "image", + "image_path": "e187cccd275bf3392be5cc5eae0489de5ce4b2237989aa0d60c99cc8aed3b408.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 190, + 468, + 331, + 590 + ], + "blocks": [ + { + "bbox": [ + 190, + 468, + 331, + 590 + ], + "lines": [ + { + "bbox": [ + 190, + 468, + 331, + 590 + ], + "spans": [ + { + "bbox": [ + 190, + 468, + 331, + 590 + ], + "type": "image", + "image_path": "48f037968eb65f45d15b702c358d3cbe19b0afd46f5759cbdc061fb7ac43a0c4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 591, + 294, + 599 + ], + "lines": [ + { + "bbox": [ + 227, + 591, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 227, + 591, + 294, + 599 + ], + "type": "text", + "content": "NeuS + Deformation" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 334, + 342, + 471, + 466 + ], + "blocks": [ + { + "bbox": [ + 334, + 342, + 471, + 466 + ], + "lines": [ + { + "bbox": [ + 334, + 342, + 471, + 466 + ], + "spans": [ + { + "bbox": [ + 334, + 342, + 471, + 466 + ], + "type": "image", + "image_path": "a360fbe372e5aadff61eedc2646c56c3fb83cbb012b82972e11d7a493e38a795.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 334, + 468, + 471, + 590 + ], + "blocks": [ + { + "bbox": [ + 334, + 468, + 471, + 590 + ], + "lines": [ + { + "bbox": [ + 334, + 468, + 471, + 590 + ], + "spans": [ + { + "bbox": [ + 334, + 468, + 471, + 590 + ], + "type": "image", + "image_path": "2237fa65faf0d54bd8d570093cfb88d461c44918c86630c731100bf4ce66c5af.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 591, + 432, + 601 + ], + "lines": [ + { + "bbox": [ + 373, + 591, + 432, + 601 + ], + "spans": [ + { + "bbox": [ + 373, + 591, + 432, + 601 + ], + "type": "text", + "content": "4DRegSDF (ours)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 473, + 342, + 543, + 466 + ], + "blocks": [ + { + "bbox": [ + 47, + 308, + 543, + 321 + ], + "lines": [ + { + "bbox": [ + 47, + 308, + 543, + 321 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 543, + 321 + ], + "type": "text", + "content": "Figure 4. Reconstruction results (mesh) with rendered images from our 4DRegSDF and previous state-of-the-art method, TiNeuVox [19]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 473, + 342, + 543, + 466 + ], + "lines": [ + { + "bbox": [ + 473, + 342, + 543, + 466 + ], + "spans": [ + { + "bbox": [ + 473, + 342, + 543, + 466 + ], + "type": "image", + "image_path": "83b96ce0f298a6c21545436ad2bb2bcdc626510e573e78b8fc2bf971900ed3be.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 473, + 468, + 543, + 590 + ], + "blocks": [ + { + "bbox": [ + 473, + 468, + 543, + 590 + ], + "lines": [ + { + "bbox": [ + 473, + 468, + 543, + 590 + ], + "spans": [ + { + "bbox": [ + 473, + 468, + 543, + 590 + ], + "type": "image", + "image_path": "c312ad9de4bacf42acc50785624e4bc67b38c2bb238f5f1ce22c39a41328dc15.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 591, + 531, + 599 + ], + "lines": [ + { + "bbox": [ + 488, + 591, + 531, + 599 + ], + "spans": [ + { + "bbox": [ + 488, + 591, + 531, + 599 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 647, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 647, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 647, + 287, + 684 + ], + "type": "text", + "content": "vide details of our surface regularization in Table 6. Note that we calculate avg. PSNR from HyperNeRF dataset [52], and avg. RMSE from DyCheck dataset [22]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Deformation network design. Unlike deformable NeRF methods [51, 54] that sequentially process deformation and" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 647, + 547, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 547, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 547, + 708 + ], + "type": "text", + "content": "rendering (Eq. 2), our 4DRegSDF proposes to jointly learn deformation and SDF through a unified network " + }, + { + "bbox": [ + 304, + 647, + 547, + 708 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 647, + 547, + 708 + ], + "type": "text", + "content": " (Eq. 3) as stated in Sec. 4.1. It turns out that our network design shows the better quality of surface normal maps as visualized in Fig. 6-(a) and (c). Such architecture design can" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "17877" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 210, + 213 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 210, + 213 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 210, + 213 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 210, + 213 + ], + "type": "image", + "image_path": "6487d9e4362303f41d268f54cbe03f8643f510ea7ed12e20ede89646df7e6eec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 215, + 174, + 225 + ], + "lines": [ + { + "bbox": [ + 83, + 215, + 174, + 225 + ], + "spans": [ + { + "bbox": [ + 83, + 215, + 174, + 225 + ], + "type": "text", + "content": "(a) NeuS + Deformation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 217, + 70, + 378, + 213 + ], + "blocks": [ + { + "bbox": [ + 217, + 70, + 378, + 213 + ], + "lines": [ + { + "bbox": [ + 217, + 70, + 378, + 213 + ], + "spans": [ + { + "bbox": [ + 217, + 70, + 378, + 213 + ], + "type": "image", + "image_path": "482fc6ba28a449987ec0f8c54cc7278413e362e3e1f1958aa246165b843acba7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 215, + 339, + 226 + ], + "lines": [ + { + "bbox": [ + 253, + 215, + 339, + 226 + ], + "spans": [ + { + "bbox": [ + 253, + 215, + 339, + 226 + ], + "type": "text", + "content": "(b) 4DRegSDF w/o " + }, + { + "bbox": [ + 253, + 215, + 339, + 226 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 233, + 546, + 256 + ], + "lines": [ + { + "bbox": [ + 46, + 233, + 546, + 256 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 546, + 256 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison on (a) NeuS [70] combined with a deformable network [54], (b) ours without surface regularization, and our final 4DRegSDF method. Left: rendered color image, right: rendered surface normal map." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 385, + 70, + 545, + 213 + ], + "blocks": [ + { + "bbox": [ + 385, + 70, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 385, + 70, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 385, + 70, + 545, + 213 + ], + "type": "image", + "image_path": "5d1a1d8cf7cd9768f999b83423c37f8a3f7d3279b1866b6c6af1667f19a0f4e2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 215, + 521, + 226 + ], + "lines": [ + { + "bbox": [ + 409, + 215, + 521, + 226 + ], + "spans": [ + { + "bbox": [ + 409, + 215, + 521, + 226 + ], + "type": "text", + "content": "(c) 4DRegSDF with " + }, + { + "bbox": [ + 409, + 215, + 521, + 226 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 409, + 215, + 521, + 226 + ], + "type": "text", + "content": " (ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 49, + 273, + 286, + 336 + ], + "blocks": [ + { + "bbox": [ + 49, + 273, + 286, + 336 + ], + "lines": [ + { + "bbox": [ + 49, + 273, + 286, + 336 + ], + "spans": [ + { + "bbox": [ + 49, + 273, + 286, + 336 + ], + "type": "table", + "html": "
Ratio of trained data
100% (Tables 2,3)25%
PSNRMS-SSIMCDPSNRMS-SSIMCD
TiNeuVox [19]22.80.8410.05718.10.7020.085
NeuS + D [70]22.10.8390.04519.40.7800.064
Ours w/o Reg.22.40.8390.04218.10.7700.060
Ours23.90.8590.04221.10.8520.058
", + "image_path": "a010613ef315b8c78c1d8a4adc4241908007e83fe4cace68c4a303b683335923.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 49, + 393, + 286, + 442 + ], + "blocks": [ + { + "bbox": [ + 47, + 338, + 287, + 383 + ], + "lines": [ + { + "bbox": [ + 47, + 338, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 287, + 383 + ], + "type": "text", + "content": "Table 4. Ablation study on few-shot frames. We choose the two scenes, 3D printer and Sriracha-tree, for the photometric and geometry evaluation. Ratio indicates the percentage of the data that we used for training the methods." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 393, + 286, + 442 + ], + "lines": [ + { + "bbox": [ + 49, + 393, + 286, + 442 + ], + "spans": [ + { + "bbox": [ + 49, + 393, + 286, + 442 + ], + "type": "table", + "html": "
NetworkPreserve (✓)Evaluation
RegNeRFGeoNeuSOursPSNR ↑RMSE ↓
✓24.10.414
4DRegSDF✓24.70.384
✓25.30.354
", + "image_path": "96b77bd1bb8cc027f38d268e9ddd5c15b82ed49ee731aef770593a36d37a0b9a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 494, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 287, + 531 + ], + "type": "text", + "content": "be one choice that reduces deformation ambiguity by learning jointly with SDF within a unified network " + }, + { + "bbox": [ + 46, + 494, + 287, + 531 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 494, + 287, + 531 + ], + "type": "text", + "content": " and having cyclic deformation constraints " + }, + { + "bbox": [ + 46, + 494, + 287, + 531 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{cc}})" + }, + { + "bbox": [ + 46, + 494, + 287, + 531 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 536, + 287, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 656 + ], + "type": "text", + "content": "Sparse view. To further verify the benefit of our surface regularization schemes, we conduct additional experiments regarding the sparse input views [29, 44]. For experiments, we use '3D printer' and 'Sriracha-tree' for evaluation. As stated in Table 4, when the training data is reduced by " + }, + { + "bbox": [ + 46, + 536, + 287, + 656 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 536, + 287, + 656 + ], + "type": "text", + "content": ", our method shows less performance drop when using a regularization scheme, showing less than " + }, + { + "bbox": [ + 46, + 536, + 287, + 656 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 536, + 287, + 656 + ], + "type": "text", + "content": " performance drop. Such results consistently support the claim that regularization techniques are effective for learning radiance fields under the given sparse views [29, 44]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "We introduce using geometric spacetime regularizations for neural rendering and surface reconstruction in dynamic" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 308, + 273, + 544, + 326 + ], + "blocks": [ + { + "bbox": [ + 47, + 445, + 287, + 479 + ], + "lines": [ + { + "bbox": [ + 47, + 445, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 287, + 479 + ], + "type": "text", + "content": "Table 5. Comparison of various regularization methods. RegNeRF [44] regularizes unobserved viewpoints and GeoNeuS [20] applies multi-view consistency as constraint." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 273, + 544, + 326 + ], + "lines": [ + { + "bbox": [ + 308, + 273, + 544, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 273, + 544, + 326 + ], + "type": "table", + "html": "
NetworkPreserve (✓)Evaluation
TVCAbs4D samplePSNR ↑RMSE ↓
4DRegSDF✓✓24.40.392
✓✓24.90.379
✓✓✓25.30.354
", + "image_path": "6bb7478cab5f810a76c358b7c1e13ad60d9c396e53ad12e15dbbe39e49c84f4f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 328, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 305, + 328, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 328, + 545, + 361 + ], + "type": "text", + "content": "Table 6. Ablation study on regularization. We decompose " + }, + { + "bbox": [ + 305, + 328, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 305, + 328, + 545, + 361 + ], + "type": "text", + "content": " (Eq. 9) into total variation of curvature (TVC) and absolute curvature of SDF (ABS). 4D sample indicates Eq. 5." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 376, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 483 + ], + "type": "text", + "content": "scenes from a monocular video. While previous dynamic NeRF methods have primarily targeted rendering quality, our approach places greater emphasis on providing accurate geometry representation via surface regularization. Our method has demonstrated promising performance in both dynamic novel-view synthesis and 4D surface reconstruction tasks. However, a notable constraint of this work is it limited capability in processing extended videos and we will extend this work for long format videos." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 489, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 489, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 305, + 489, + 545, + 536 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by IITP grants (No.2021-0-02068: AI Innovation Hub and No.2023R1A1C200781211) funded by the Korea government (MSIT)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 548, + 365, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 365, + 560 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 365, + 560 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 310, + 567, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 310, + 567, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 567, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 310, + 567, + 545, + 612 + ], + "type": "text", + "content": "[1] Marc Alexa, Daniel Cohen-Or, and David Levin. As-rigid-as-possible shape interpolation. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 157–164, 2000." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 311, + 613, + 547, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 613, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 311, + 613, + 547, + 657 + ], + "type": "text", + "content": "[2] Luca Ballan, Gabriel J Brostow, Jens Puwein, and Marc Pollefeys. Unstructured video-based rendering: Interactive exploration of casually captured videos. In ACM SIGGRAPH 2010 papers, pages 1-11. 2010." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 311, + 658, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 658, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 311, + 658, + 547, + 713 + ], + "type": "text", + "content": "[3] Aayush Bansal, Minh Vo, Yaser Sheikh, Deva Ramanan, and Srinivasa Narasimhan. 4d visualization of dynamic events from unconstrained multi-view videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5366-5375, 2020." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17878" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 721 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 137 + ], + "type": "text", + "content": "[4] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 140, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 140, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 53, + 140, + 288, + 195 + ], + "type": "text", + "content": "[5] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4):86-1, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 197, + 288, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 197, + 288, + 250 + ], + "spans": [ + { + "bbox": [ + 53, + 197, + 288, + 250 + ], + "type": "text", + "content": "[6] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph rendering. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 425-432, 2001." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 251, + 288, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 306 + ], + "type": "text", + "content": "[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 308, + 288, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 308, + 288, + 351 + ], + "spans": [ + { + "bbox": [ + 53, + 308, + 288, + 351 + ], + "type": "text", + "content": "[8] Shenchang Eric Chen and Lance Williams. View interpolation for image synthesis. In Proceedings of the 20th annual conference on Computer graphics and interactive techniques, pages 279-288, 1993." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 353, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 353, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 353, + 288, + 407 + ], + "type": "text", + "content": "[9] Jaesung Choe, Sunghoon Im, Francois Rameau, Minjun Kang, and In So Kweon. Volumefusion: Deep depth fusion for 3d scene reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16086-16095, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 408, + 288, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 408, + 288, + 452 + ], + "spans": [ + { + "bbox": [ + 48, + 408, + 288, + 452 + ], + "type": "text", + "content": "[10] Jaesung Choe, Kyungdon Joo, Tooba Imtiaz, and In So Kweon. Volumetric propagation network: Stereo-lidar fusion for long-range depth estimation. IEEE Robotics and Automation Letters, 6(3):4672-4679, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 453, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 288, + 485 + ], + "type": "text", + "content": "[11] Jaesung Choe, Byeongin Joung, Francois Rameau, Jaesik Park, and In So Kweon. Deep point cloud reconstruction. arXiv preprint arXiv:2111.11704, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 487, + 288, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 288, + 529 + ], + "type": "text", + "content": "[12] Gene Chou, Ilya Chugunov, and Felix Heide. Gensdf: Two-stage learning of generalizable signed distance functions. Advances in Neural Information Processing Systems, 35:24905-24919, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 532, + 288, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 288, + 585 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 288, + 585 + ], + "type": "text", + "content": "[13] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 588, + 288, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 288, + 631 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 288, + 631 + ], + "type": "text", + "content": "[14] Paul Debevec, Yizhou Yu, and George Borshukov. Efficient view-dependent image-based rendering with projective texture-mapping. In Eurographics Workshop on Rendering Techniques, pages 105-116. Springer, 1998." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 632, + 288, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 632, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 48, + 632, + 288, + 664 + ], + "type": "text", + "content": "[15] Manfredo P Do Carmo. Differential geometry of curves and surfaces: revised and updated second edition. Courier Dover Publications, 2016." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 666, + 288, + 721 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 666, + 288, + 721 + ], + "spans": [ + { + "bbox": [ + 48, + 666, + 288, + 721 + ], + "type": "text", + "content": "[16] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 721 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[17] Arda Duzceker, Silvano Galliani, Christoph Vogel, Pablo Speciale, Mihai Dusmanu, and Marc Pollefeys. Deepvideovs: Multi-view stereo on video with recurrent spatiotemporal fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15324-15333, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 183 + ], + "type": "text", + "content": "[18] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "type": "text", + "content": "[19] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. arXiv preprint arXiv:2205.15285, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 230, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 273 + ], + "type": "text", + "content": "[20] Qiancheng Fu, Qingshan Xu, Yew-Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. arXiv preprint arXiv:2205.15848, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 275, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 545, + 319 + ], + "type": "text", + "content": "[21] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 320, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 352 + ], + "type": "text", + "content": "[22] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. arXiv preprint arXiv:2210.13445, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 354, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 407 + ], + "type": "text", + "content": "[23] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4857-4866, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 409, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 409, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 409, + 545, + 453 + ], + "type": "text", + "content": "[24] Steven J Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 43-54, 1996." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 454, + 545, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 454, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 307, + 454, + 545, + 486 + ], + "type": "text", + "content": "[25] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 487, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 545, + 542 + ], + "type": "text", + "content": "[26] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3d scene reconstruction with the manhattan-world assumption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5511-5520, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 544, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 576 + ], + "type": "text", + "content": "[27] Sunghoon Im, Hae-Gon Jeon, Stephen Lin, and In So Kweon. Dpsnet: End-to-end deep plane sweep stereo. arXiv preprint arXiv:1905.00538, 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 578, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 621 + ], + "type": "text", + "content": "[28] Olga Karpenko, John F Hughes, and Ramesh Raskar. Freeform sketching with variational implicit surfaces. In Computer Graphics Forum, volume 21, pages 585-594. Wiley Online Library, 2002." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 623, + 545, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 623, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 307, + 623, + 545, + 675 + ], + "type": "text", + "content": "[29] Mijeong Kim, Seonguk Seo, and Bohyung Han. Infonerf: Ray entropy minimization for few-shot neural volume rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12912-12921, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 677, + 545, + 721 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 677, + 545, + 721 + ], + "spans": [ + { + "bbox": [ + 307, + 677, + 545, + 721 + ], + "type": "text", + "content": "[30] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17879" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 700 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 287, + 116 + ], + "type": "text", + "content": "[31] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "type": "text", + "content": "[32] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 31-42, 1996." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 151, + 287, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 151, + 287, + 193 + ], + "spans": [ + { + "bbox": [ + 49, + 151, + 287, + 193 + ], + "type": "text", + "content": "[33] Tzu-Mao Li, Miika Aittala, Frédo Durand, and Jaakko Lehtinen. Differentiable monte carlo ray tracing through edge sampling. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 194, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 194, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 49, + 194, + 287, + 248 + ], + "type": "text", + "content": "[34] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 250, + 287, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 250, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 49, + 250, + 287, + 303 + ], + "type": "text", + "content": "[35] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. arXiv preprint arXiv:2205.15723, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 304, + 287, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 304, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 49, + 304, + 287, + 347 + ], + "type": "text", + "content": "[36] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 349, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 349, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 49, + 349, + 287, + 392 + ], + "type": "text", + "content": "[37] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. arXiv preprint arXiv:2206.05737, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 393, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 393, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 49, + 393, + 287, + 425 + ], + "type": "text", + "content": "[38] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 426, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 426, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 49, + 426, + 287, + 468 + ], + "type": "text", + "content": "[39] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. arXiv preprint arXiv:2011.13495, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 470, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 470, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 49, + 470, + 287, + 524 + ], + "type": "text", + "content": "[40] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 525, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 525, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 49, + 525, + 287, + 579 + ], + "type": "text", + "content": "[41] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 581, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 581, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 49, + 581, + 287, + 623 + ], + "type": "text", + "content": "[42] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 287, + 678 + ], + "type": "text", + "content": "[43] Zak Murez, Tarrence van As, James Bartolozzi, Ayan Sinha, Vijay Badrinarayanan, and Andrew Rabinovich. Atlas: End-to-end 3d scene reconstruction from posed images. In European conference on computer vision, pages 414-431. Springer, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 680, + 287, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 680, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 49, + 680, + 287, + 700 + ], + "type": "text", + "content": "[44] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Reg-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 695 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 114 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 114 + ], + "type": "text", + "content": "nerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 116, + 545, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 116, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 308, + 116, + 545, + 170 + ], + "type": "text", + "content": "[45] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Occupancy flow: 4d reconstruction by learning particle dynamics. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5379-5389, 2019." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 172, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 172, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 172, + 545, + 226 + ], + "type": "text", + "content": "[46] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "text", + "content": "[47] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 284, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 349 + ], + "type": "text", + "content": "[48] Sergio Orts-Escolano, Christoph Rhemann, Sean Fanello, Wayne Chang, Adarsh Kowdle, Yury Degtyarev, David Kim, Philip L Davidson, Sameh Khamis, Mingsong Dou, et al. Holoportation: Virtual 3d teleportation in real-time. In Proceedings of the 29th annual symposium on user interface software and technology, pages 741-754, 2016." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 350, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 382 + ], + "type": "text", + "content": "[49] Stanley Osher, Ronald Fedkiw, and K Piechor. Level set methods and dynamic implicit surfaces. Appl. Mech. Rev., 57(3):B15-B15, 2004." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 384, + 545, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 384, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 545, + 438 + ], + "type": "text", + "content": "[50] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 440, + 545, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 440, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 308, + 440, + 545, + 494 + ], + "type": "text", + "content": "[51] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 496, + 545, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 496, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 308, + 496, + 545, + 550 + ], + "type": "text", + "content": "[52] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), dec 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 552, + 545, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 552, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 308, + 552, + 545, + 594 + ], + "type": "text", + "content": "[53] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 597, + 545, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 597, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 308, + 597, + 545, + 650 + ], + "type": "text", + "content": "[54] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 652, + 545, + 695 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 652, + 545, + 695 + ], + "spans": [ + { + "bbox": [ + 308, + 652, + 545, + 695 + ], + "type": "text", + "content": "[55] Yi-Ling Qiao, Alexander Gao, and Ming Lin. Neuphysics: Editable neural geometry and physics from monocular videos. In Advances in Neural Information Processing Systems, 2022." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "17880" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 718 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[56] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 193 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 193 + ], + "type": "text", + "content": "[57] Yufan Ren, Tong Zhang, Marc Pollefeys, Sabine Susstrunk, and Fangjinhua Wang. Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16685-16695, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 194, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 288, + 248 + ], + "type": "text", + "content": "[58] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2304-2314, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 249, + 288, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 249, + 288, + 303 + ], + "spans": [ + { + "bbox": [ + 48, + 249, + 288, + 303 + ], + "type": "text", + "content": "[59] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 304, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 288, + 357 + ], + "type": "text", + "content": "[60] Miroslava Slavcheva, Maximilian Baust, Daniel Cremers, and Slobodan Ilic. Killingfusion: Non-rigid 3d reconstruction without correspondences. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1386-1395, 2017." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 358, + 288, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 358, + 288, + 412 + ], + "spans": [ + { + "bbox": [ + 48, + 358, + 288, + 412 + ], + "type": "text", + "content": "[61] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. arXiv preprint arXiv:2210.15947, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 413, + 288, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 413, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 48, + 413, + 288, + 445 + ], + "type": "text", + "content": "[62] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, volume 4, pages 109–116, 2007." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 446, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 446, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 446, + 288, + 479 + ], + "type": "text", + "content": "[63] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. arXiv preprint arXiv:2111.11215, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 479, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 288, + 533 + ], + "type": "text", + "content": "[64] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 533, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 288, + 555 + ], + "type": "text", + "content": "[65] Jiaxiang Tang. Torch-ngp: a pytorch implementation of instant-ngp, 2022. https://github.com/ashawkey/torch-ngp." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 288, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 288, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 288, + 599 + ], + "type": "text", + "content": "[66] Demetri Terzopoulos, John Platt, Alan Barr, and Kurt Fleischer. Elastically deformable models. In Proceedings of the 14th annual conference on Computer graphics and interactive techniques, pages 205-214, 1987." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 600, + 288, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 288, + 642 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 288, + 642 + ], + "type": "text", + "content": "[67] Itsuki Ueda, Yoshihiro Fukuhara, Hirokatsu Kataoka, Hiroaki Aizawa, Hidehiko Shishido, and Itaru Kitahara. Neural density-distance fields. arXiv preprint arXiv:2207.14455, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 643, + 288, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 643, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 48, + 643, + 288, + 686 + ], + "type": "text", + "content": "[68] Michael Waechter, Nils Moehrle, and Michael Goesele. Let there be color! large-scale texturing of 3d reconstructions. In European conference on computer vision, pages 836-850. Springer, 2014." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 686, + 288, + 718 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 686, + 288, + 718 + ], + "spans": [ + { + "bbox": [ + 48, + 686, + 288, + 718 + ], + "type": "text", + "content": "[69] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994, 2021." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 699 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[70] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 183 + ], + "type": "text", + "content": "[71] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 184, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 545, + 238 + ], + "type": "text", + "content": "[72] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 240, + 545, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 545, + 283 + ], + "type": "text", + "content": "[73] Guandao Yang, Serge Belongie, Bharath Hariharan, and Vladlen Koltun. Geometry processing with neural fields. Advances in Neural Information Processing Systems, 34:22483-22497, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 285, + 545, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 545, + 328 + ], + "type": "text", + "content": "[74] Gengshan Yang, Minh Vo, Neverova Natalia, Deva Ramanan, Vedaldi Andrea, and Joo Hanbyul. Banmo: Building animatable 3d neural models from many casual videos. In CVPR, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 330, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 330, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 330, + 545, + 373 + ], + "type": "text", + "content": "[75] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 375, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 375, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 375, + 545, + 407 + ], + "type": "text", + "content": "[76] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. NeurIPS, 34:4805-4815, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 408, + 545, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 408, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 307, + 408, + 545, + 462 + ], + "type": "text", + "content": "[77] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv preprint arXiv:2302.14859, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 464, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 520 + ], + "type": "text", + "content": "[78] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5336-5345, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 520, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 563 + ], + "type": "text", + "content": "[79] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. arXiv preprint arXiv:2112.05131, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 565, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 565, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 565, + 545, + 609 + ], + "type": "text", + "content": "[80] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 609, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 609, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 307, + 609, + 545, + 654 + ], + "type": "text", + "content": "[81] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 654, + 545, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 654, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 307, + 654, + 545, + 699 + ], + "type": "text", + "content": "[82] C Lawrence Zitnick, Sing Bing Kang, Matthew Uytendaele, Simon Winder, and Richard Szeliski. High-quality video view interpolation using a layered representation. ACM transactions on graphics (TOG), 23(3):600-608, 2004." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17881" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Sparse Instance Conditioned Multimodal Trajectory Prediction/5a87e2e4-dac0-41f3-9966-85985369c6c6_content_list.json b/2023/Sparse Instance Conditioned Multimodal Trajectory Prediction/5a87e2e4-dac0-41f3-9966-85985369c6c6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..eb747335b5f964899f9c39f193068d042c624947 --- /dev/null +++ b/2023/Sparse Instance Conditioned Multimodal Trajectory Prediction/5a87e2e4-dac0-41f3-9966-85985369c6c6_content_list.json @@ -0,0 +1,1576 @@ +[ + { + "type": "text", + "text": "Sparse Instance Conditioned Multimodal Trajectory Prediction", + "text_level": 1, + "bbox": [ + 163, + 130, + 805, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yonghao Dong $^{1}$ Le Wang $^{1*}$ Sanping Zhou $^{1}$ Gang Hua $^{2}$ $^{1}$ National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University \n $^{2}$ Wormpex AI Research", + "bbox": [ + 169, + 179, + 800, + 271 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 303, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pedestrian trajectory prediction is critical in many vision tasks but challenging due to the multimodality of the future trajectory. Most existing methods predict multimodal trajectories conditioned by goals (future endpoints) or instances (all future points). However, goal-conditioned methods ignore the intermediate process and instance-conditioned methods ignore the stochasticity of pedestrian motions. In this paper, we propose a simple yet effective Sparse Instance Conditioned Network (SICNet), which gives a balanced solution between goal-conditioned and instance-conditioned methods. Specifically, SICNet learns comprehensive sparse instances, i.e., representative points of the future trajectory, through a mask generated by a long short-term memory encoder and uses the memory mechanism to store and retrieve such sparse instances. Hence SICNet can decode the observed trajectory into the future prediction conditioned on the stored sparse instance. Moreover, we design a memory refinement module that refines the retrieved sparse instances from the memory to reduce memory recall errors. Extensive experiments on ETH-UCY and SDD datasets show that our method outperforms existing state-of-the-art methods. In addition, ablation studies demonstrate the superiority of our method compared with goal-conditioned and instance-conditioned approaches.", + "bbox": [ + 75, + 335, + 473, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 729, + 209, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pedestrian trajectory prediction aims to predict pedestrians' future paths given their observed trajectories. It plays an important role in autonomous driving [22, 43, 5], human motion prediction [6, 12, 19], video surveillance [57, 48, 18, 32], and visual recognition [38, 42, 46]. Despite the recent advancement [30, 49, 8, 36, 10], pedestrian trajectory prediction is still challenging due to the multimodality of future trajectories. As the pedestrian's motions are stochastic and", + "bbox": [ + 75, + 753, + 470, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9834723deedee81acb7629c918849e3d01939dfaf75a99252920265d774667d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 305, + 694, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b3ca369f619b2a0ce5c7a819e86e2d306f7f93b4e53e00beb0c6bf8133c923a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 305, + 887, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17c4f188ba87645800b3b44f78a1d83ce123d79c50c449ab186f87822314067a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 382, + 694, + 453 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5785870e0f39c99ece345d3ce4ddc16782130a70c1e0e6c583caad5bdbe02b76.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 381, + 888, + 453 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/af08258e5f443f5d0219df92a8af7aea13bb1ae842c81123e9122deb83e0886e.jpg", + "image_caption": [ + "Figure 1. Illustration of different multimodal prediction methods. (a) Goals, instances and sparse-instances prepared for the prediction. (b) Goal-conditioned methods. (c) Instance-conditioned methods. (d) Sparse-instance-conditioned methods. The sparse-instance-conditioned method is more accurate in in-between states than the goal-conditioned method and more flexible than the instance-conditioned method, thus making the better prediction." + ], + "image_footnote": [], + "bbox": [ + 503, + 455, + 888, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "indeterminate [29], multiple future paths are possible given their current states.", + "bbox": [ + 498, + 598, + 890, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To handle multimodality, many methods [27, 49, 55] transform the multimodal trajectories into multimodal goals. They first predict various goals as modality representations and then sample multiple goals to generate diverse future trajectories. However, they only focus on goals while neglecting the intermediate movement process. As illustrated in Figure 1(b), although with accurate goals, the intermediate process could be far from the ground truth. Thus, the multimodal goals are not equal to the multimodal future trajectories. It is necessary to consider the in-between states to comprehensively describe the multimodality of trajectory. Moreover, some other work [27, 28] predicts multimodal trajectories conditioned by instances. However, the instance-conditioned methods will lead to motion stochasticity loss and modality redundancy, resulting in performance degradation and computational resource waste. As shown in Figure 1(c), due to the reduced randomness of the instance-conditioned approach, the prediction retains the same bias", + "bbox": [ + 496, + 628, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 107, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9763", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "when the instance deviates from the ground truth.", + "bbox": [ + 78, + 90, + 403, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To close this gap, we present a Sparse Instance Conditioned Network (SICNet). In our model, a sparse instance, which can give a comprehensive and adaptive description of future modality, is constructed by element-wise multiplication between the future trajectory and a learned mask. Hence, SICNet can predict the future trajectory more precisely guided by the sparse instance. To obtain the related sparse instances during inference, we adopt the memory mechanism to store non-redundant sparse instances corresponding to all observed trajectories. Therefore, SICNet can recall its modality's most related sparse instance and jointly generate the prediction during inference. Moreover, when the memory meets an unfamiliar observed trajectory, i.e., a trajectory quite different from stored items, the memory mechanism may recall false experiences. Hence we propose a memory refinement module with a refinement loss to reduce such recall errors by bridging the gap between the recall and the ground truth. For multimodal trajectories generation, we can retrieve a certain number of the most likely sparse instances from the external memory and then generate corresponding multimodal trajectories. In addition, to further remove the redundancy from the retrieved sparse instances, we use the cluster mechanism as a post-process to obtain succinct multimodal proposals for the prediction instead of the sparse instance during inference.", + "bbox": [ + 80, + 107, + 468, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments and ablation studies on two popular pedestrian trajectory prediction datasets, i.e., ETH-UCY [31, 21] and Stanford Drone Dataset (SDD) [33], demonstrate the superiority of our method over prior state-of-the-art methods. In summary, our contributions are four-fold: (1) We give a balanced solution, i.e., a novel sparse instance, between the goal and instance to guide the multimodal trajectory prediction. (2) We propose a simple yet effective sparse instance conditioned network for the multimodal trajectory prediction. (3) We propose a memory refinement module for the external memory to reduce recall errors when meeting unfamiliar observed trajectories. (4) Extensive experiments and ablation studies demonstrate the superiority and flexibility of our method compared with the goal-conditioned and instance-conditioned approaches.", + "bbox": [ + 80, + 484, + 468, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 80, + 723, + 223, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal Pedestrian Trajectory Prediction. Pedestrian trajectory prediction [2, 23, 3, 39, 45] aims to predict pedestrians' future trajectories based on their observed trajectories. There are some early studies on the deterministic trajectory prediction [1, 44, 14, 50]. Due to the strong randomness and uncertainty of pedestrian motions, there is no single correct future trajectory prediction. Motivated by this, Social GAN [13] proposes the concept of multimodal pedestrian trajectory prediction and emphasizes its importance.", + "bbox": [ + 80, + 750, + 468, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Earlier work generates multimodal trajectories through", + "bbox": [ + 99, + 886, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "latent variables. Social GAN [13] proposes a Generative Adversarial Network based on a social pooling mechanism to generate a multimodal trajectory distribution. SGCN [37] generates multimodal trajectories by modeling the future predictions as a bi-variate Gaussian distribution. The CVAE-based methods [15, 35, 8] generate multimodal trajectories by sampling latent variables from a learned latent space. However, the latent variable conditioned method is persecuted by the problem of model collapse, thus resulting in performance degradation. Moreover, the latent variable conditioned methods also suffer in interpretability.", + "bbox": [ + 501, + 90, + 890, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the model collapse and interpretability issues, PECNet [27] proposes a goal-conditioned approach, which transforms multimodal trajectories into multimodal goals and then samples multiple goals to generate corresponding diverse predictions. After that, amounts of goal-conditioned studies [55, 8, 49] emerge and show superior performance. However, goal-conditioned methods represent different modalities with different possible goals while neglecting the intermediate states, thus suffering in performance improvement. In addition, PCCSNet[27] and MANTRA [28] propose instance-conditioned approaches, which use all points of the future instance to describe each modality precisely. However, it could lead to stochasticity loss and modality redundancy, resulting in performance degradation and computational resource waste. Moreover, Y-net [26] use goals and waypoints to make the prediction. However, the waypoints' number and time step locations are manually determined in Y-net [26], which may decrease the prediction accuracy by manually choosing errors.", + "bbox": [ + 503, + 260, + 890, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, we give a balanced way, i.e., a novel sparse instance, between the goal and instance to guide the multimodal trajectory prediction. The sparse instance is constructed by the learned representative points of the instance, which can improve modality representation accuracy compared with goal-conditioned methods and preserve stochasticity compared with instance-conditioned methods.", + "bbox": [ + 503, + 550, + 890, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Memory Network. The memory-augmented network [51, 11, 25, 24, 53] uses an external memory module to store critical information and has been widely applied in many areas like detection [7, 4], tracking [52, 9], segmentation [47], image generation [17], and video summarization [20]. Memory GAN [17] proposed a novel end-to-end unsupervised GAN network with a learnable memory network that effectively learns a highly multimodal latent space without suffering from structural discontinuity and forgetting problems. OGEMN [7] proposes the first object-guided external memory network for the online video object detection. STMTrack [52] proposes a novel tracking framework built on a space-time memory network, which can fully use historical information related to the target for better adapting to appearance variations during tracking. RMNet [47] memorizes the target object regions, effectively alleviating", + "bbox": [ + 503, + 659, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9764", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b8c818bfbe20c19976c1a850a0c79282ac3897387b75bebbae45b9b12fbbbb25.jpg", + "image_caption": [ + "Figure 2. The framework of our proposed SICNet. SICNet adopts a two-stage training strategy. The sparse instance representation module generates the sparse instances by a learned mask and then trains an external memory to store and retrieve them. The memory refinement module refines the retrieved sparse instances from memory to reduce the searching errors. We cluster the refined sparse instances into multimodal proposals to remove redundancy further. The cluster only exists in inference to save training costs." + ], + "image_footnote": [], + "bbox": [ + 80, + 90, + 890, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "objects' ambiguity and reducing computational complexity.", + "bbox": [ + 75, + 349, + 468, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Two previous memory-based work, MANTRA [28] and MemoNet [49], also use the memory mechanism in multimodal trajectory prediction. However, the difference are mainly in three aspects: (1) The contribution of our proposed SICNet focuses on a novel representation of the trajectory modality, i.e., a learned sparse instance, while MANTRA [28] uses the whole instance and MemoNet [49] uses the goal to represent modalities of multimodal trajectories; (2) We propose a memory refinement module to reduce recall errors when meeting unfamiliar observed trajectories during inference, while MANTRA [28] and MemoNet [49] ignore such circumstances. (3) To demonstrate the superiority of our proposed sparse instance, we do not use extra social interaction and map information, while MANTRA [28] use extra map information and MemoNet [49] use social interaction information. Despite this, we also achieve the best performance compared with the above two methods, which indicates the superiority of our method.", + "bbox": [ + 75, + 364, + 472, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Method", + "text_level": 1, + "bbox": [ + 76, + 648, + 207, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As aforementioned, traditional goal-conditioned methods ignore the intermediate process and instance-conditioned methods ignore the stochasticity of pedestrian motions, which leads to performance degradation. In this section, we present a sparse-instance-conditioned network (SICNet) to close this gap. Our SICNet adopts a two-stage training strategy. In the first stage, we learn the sparse instance, which is used to reconstruct the future trajectory jointly with the observed trajectory. Meanwhile, an external memory is trained to store and retrieve the sparse instance. In the second stage, we refine the retrieved sparse instances from the external memory through a refinement module with a refinement loss to reduce the recall errors. In addition, we use the cluster mechanism to remove redundancy of the retrieved sparse instances further during inference.", + "bbox": [ + 75, + 674, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 498, + 349, + 702, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The objective of pedestrian trajectory prediction is to predict possible future trajectory coordinates based on the observed trajectories. For each target pedestrian, given an observed trajectory $X = \\{(x_{t},y_{t})\\}_{t = 1}^{T_{obs}}$ , where $(x_{t},y_{t})$ is the 2D coordinate at time $t$ , we aim to predict the future trajectory coordinates $Y = \\{(x_{t},y_{t})\\}_{t = T_{obs} + 1}^{T_{pred}}$ . Note that multimodal pedestrian trajectory prediction requires the model to predict $K$ future trajectories $\\{Y_i\\}_{i = 1}^K$ to account for multimodality, while only one future trajectory (ground truth) is provided for training in the dataset.", + "bbox": [ + 496, + 372, + 893, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Sparse Instance Representation", + "text_level": 1, + "bbox": [ + 498, + 534, + 779, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the first training stage, we generate the sparse instance features to represent future modalities and store them by memory mechanism. As shown in Figure 3, we first train the encoder-decoder with a reconstruction loss in this stage. Then we train the memory module to store the sparse instances generated during reconstruction.", + "bbox": [ + 496, + 556, + 893, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature Encoder. As the trajectory is a temporal sequence, we use two LSTM [56] blocks to build the past encoder $f_{p}(\\cdot)$ and the future encoder $f_{y}(\\cdot)$ , respectively. The hidden states of the two encoders are used as their corresponding features, as follows:", + "bbox": [ + 496, + 648, + 893, + 723 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {F} _ {p} = f _ {p} (X), \\\\ \\mathbf {F} _ {y} = f _ {y} (Y), \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 733, + 890, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_p\\in \\mathbb{R}^{1\\times d}$ is the past feature extracted from the observed trajectory $X$ , and $\\mathbf{F}_y\\in \\mathbb{R}^{1\\times d}$ is the future feature extracted from the future trajectory $Y$ . $d$ is the dimension of hidden state in LSTM.", + "bbox": [ + 496, + 777, + 893, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sparse Instance. When inputting an observed trajectory, we want to obtain the corresponding sparse instance of its future modality so that we can construct future trajectories jointly using the observed trajectory and the sparse instance.", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9765", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/09536bbf7de6c330f9f998b4fd8a62fb6f6238b120d410df38d96efc40496b60.jpg", + "image_caption": [ + "Figure 3. The first stage of the training process. In this stage, we train the sparse instance by reconstructing the future trajectory. Then we train an external memory to store and retrieve the learned sparse instances." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 436, + 162 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/501526ea587d50b87d4759eb5aa21ac6e2d4da47701750d0f9974eed9fa328c8.jpg", + "image_caption": [ + "Figure 4. The second stage of the training process. In this stage, we refine the retrieved sparse instances from memory through the memory refinement module to reduce the recall errors." + ], + "image_footnote": [], + "bbox": [ + 78, + 234, + 467, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hence, we first learn an instance mask from the past feature $\\mathbf{F}_p$ through an MLP layer, as follows:", + "bbox": [ + 76, + 382, + 468, + 412 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {F} _ {m} = \\phi (\\mathbf {F} _ {p}, \\theta_ {M}), \\\\ \\begin{array}{l} M = \\mathbb {I} (\\delta (\\mathbf {F} _ {m}) \\leq \\mu)), \\end{array} \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 422, + 468, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi (\\cdot ,\\cdot)$ denotes the MLP layer. $\\theta_{M}$ is a learnable parameter. $\\mathbf{F}_m$ is the learned feature which is used to construct the mask. $M$ denotes the instance mask. $\\delta (\\cdot)$ is sigmoid function to map $\\mathbf{F}_m$ to the mask value [0, 1]. $\\mathbb{I}$ is an indicator function which equals 0 if the inequality holds, otherwise 1. $\\mu$ is the mask threshold. Compared with directly masking instances, masking instances in the high-level feature dimension is more efficient and effective. Hence we generate the sparse instance $\\mathbf{F}_{ins}$ in the feature level dimension, as:", + "bbox": [ + 75, + 469, + 472, + 604 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {i n s} = \\mathbf {F} _ {y} \\odot M, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 618, + 468, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ denotes the element-wise multiplication.", + "bbox": [ + 76, + 647, + 406, + 661 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature Decoder. The sparse instance feature should be representative enough for the future trajectory. In addition, our goal is to predict the future trajectory based on the observed trajectory conditioned on the sparse instance. Therefore, we use an LSTM [56] decoder to reconstruct the future trajectory with the combination of the past feature and the sparse instance feature, as follows:", + "bbox": [ + 75, + 662, + 470, + 767 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY _ {r} = f _ {d} \\left(\\mathbf {F} _ {p} \\oplus \\mathbf {F} _ {i n s}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 781, + 468, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{d}(\\cdot)$ is the decoder implemented by an LSTM block, $\\oplus$ is a concatenation operation, and $Y_{r}$ is the reconstructed future trajectory. The trajectory $\\ell_{2}$ -norm loss is used as the reconstruction loss, as:", + "bbox": [ + 76, + 809, + 470, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {1} = \\left\\| Y - Y _ {r} \\right\\| _ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 882, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $||\\cdot ||_2$ denotes the $\\ell_2$ -norm distance.", + "bbox": [ + 498, + 90, + 782, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Memory Architecture. The key idea of the memory module is to use the encoding of the observed trajectory as a memory key to retrieve possible sparse instances and jointly generate the multimodal predictions. To store the sparse instances generated during the reconstruction process, we first randomly initializes two memory banks, i.e., the past memory bank $\\mathbf{M}_p \\in \\mathbb{R}^{m \\times d}$ , and the instance memory bank $\\mathbf{M}_i \\in \\mathbb{R}^{m \\times d}$ to store the past and the sparse instance features obtained from the reconstruction process, respectively. $m$ denotes the storage size of the memory bank, and $d$ is the dimension of the memory item. To combine the past and sparse instance features, the items in these two memory banks are aligned. Specifically, for the $j$ -th past feature in $\\mathbf{M}_p[j]$ , there is a unique sparse instance feature in $\\mathbf{M}_i[j]$ corresponding to it.", + "bbox": [ + 496, + 106, + 890, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Memory Update. Due to the abundant similar trajectories in the dataset, it brings high redundancy to store all of them in the memory bank. Hence we need to remove modality redundancy by memory updating process. Moreover, traditional memory mechanism [16] updates the memory bank with the stored items' labels, while such labels are not provided in trajectory prediction to distinguish trajectories with different modalities. Therefore, we design an unsupervised memory update strategy based on the distance between the stored and current items, as shown in Figure 5. Inspired by [16], an additional mark memory bank $\\mathbf{A} \\in \\mathbb{R}^{m \\times 1}$ tracks the age of items stored in memory without being used. All items in $\\mathbf{A}$ are initialized to 1 and aligned to their corresponding items in $\\mathbf{M}_p$ and $\\mathbf{M}_i$ . Given a pair of inputs $(\\mathbf{F}_p, \\mathbf{F}_{ins})$ , we first calculate the feature similarities between $\\mathbf{F}_p$ and each item in $\\mathbf{M}_i$ . Then, the $k$ -th $(k \\in \\{1, \\dots, m\\})$ item $\\mathbf{M}_p[k]$ with the highest similarity is used to retrieve the corresponding sparse instance feature, i.e., $\\mathbf{M}_i[k]$ , through the one-to-one architecture of memory banks.", + "bbox": [ + 496, + 332, + 893, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the distance $\\mathcal{D}$ between $\\mathbf{F}_{ins}$ and $\\mathbf{M}_i[k]$ is defined as:", + "bbox": [ + 498, + 619, + 890, + 648 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} = \\left\\| \\mathbf {M} _ {i} [ k ] - \\mathbf {F} _ {i n s} \\right\\| _ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 662, + 890, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We compare $\\mathcal{D}$ with a threshold $\\xi$ to judge whether $F_{ins}$ and $\\mathbf{M}_i[k]$ have the same modality. When $\\mathcal{D} > \\xi$ , which means that $F_{ins}$ and $\\mathbf{M}_i[k]$ have different modalities. Hence we update $\\mathbf{M}_p$ , $\\mathbf{M}_i$ with the oldest age (i.e., corresponding $\\mathbf{A}$ with the maximum value) by $\\mathbf{F}_p$ , $\\mathbf{F}_{ins}$ , respectively, as follows:", + "bbox": [ + 496, + 691, + 890, + 781 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {i} [ s ] \\leftarrow \\mathbf {F} _ {i n s}, \\quad \\mathbf {M} _ {p} [ s ] \\leftarrow \\mathbf {F} _ {p}, \\quad \\mathbf {A} [ s ] \\leftarrow 0, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 795, + 890, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s$ is the index with the oldest age. Since the $s$ -th item in the past memory bank has been updated, the $s$ -th item in $\\mathbf{A}$ is set to 0 to prevent the newly added item from overwriting it. When $\\mathcal{D} \\leq \\xi$ , which means that $F_{ins}$ and $\\mathbf{M}_i[k]$ have the same modality. Hence we only update $\\mathbf{M}_p[k]$ to approach", + "bbox": [ + 496, + 824, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9766", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/460dfc9a0acd3bd37e6e9b7b9afbb113be3dbbe29d38c070e4db4d9db6aa5b2d.jpg", + "image_caption": [ + "Figure 5. An illustration of the memory training process in SICNet. Red and purple lines denote the memory writing process." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 467, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbf{F}_p$ by taking the normalized average of $\\mathbf{M}_p[k]$ and $\\mathbf{F}_p$ , as:", + "bbox": [ + 76, + 287, + 464, + 306 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {p} [ k ] \\leftarrow \\frac {\\mathbf {F} _ {p} + \\mathbf {M} _ {p} [ k ]}{\\| \\mathbf {F} _ {p} + \\mathbf {M} _ {p} [ k ] \\|}, \\quad \\mathbf {A} [ k ] \\leftarrow 0. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 315, + 470, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus, we can retrieve the sparse instance feature $\\mathbf{M}_i[k]$ easily, when we receive the past feature similar to $\\mathbf{F}_p$ . Since the $k$ -th item in the past memory bank has been updated, the $k$ -th item in $\\mathbf{A}$ is set to 0 to prevent the newly added item from covering it. All values in $\\mathbf{A}$ will add the same constant after an epoch.", + "bbox": [ + 76, + 359, + 472, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Memory Refinement", + "text_level": 1, + "bbox": [ + 76, + 460, + 272, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The input observed trajectory encoding could be unfamiliar to the well-trained memory, which could be caused by: (1) modality loss due to removing redundancy during the memory update process; (2) the input modality does not exist in the train set. Such unfamiliarity could lead to the recalled sparse instance not matching the input trajectory. To reduce such recall errors, we design a memory refinement module with a refinement loss in the second training stage, as shown in Figure 4.", + "bbox": [ + 75, + 484, + 468, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first consider the difference between the past feature $\\mathbf{F}_p$ and its corresponding closest past memory item $\\mathbf{M}_p[k]$ with the same modality. To avoid extra optimization similar to Eq. (8), we do not optimize $\\mathbf{F}_p$ to approach $\\mathbf{M}_p[k]$ again but encode the difference between $\\mathbf{F}_p$ and $\\mathbf{M}_p[k]$ as an additional deviation feature $\\mathbf{F}_v$ , which is formulated as:", + "bbox": [ + 76, + 621, + 468, + 712 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {v} = \\phi \\left(\\mathbf {F} _ {p} \\oplus \\mathbf {M} _ {p} [ k ], \\theta_ {V}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 724, + 468, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta_V$ is a learnable parameter, and $\\oplus$ is the concatenation operation. Then, the retrieved sparse instance feature is refined as $\\mathbf{F}_r$ , as follows:", + "bbox": [ + 76, + 753, + 470, + 799 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {r} = \\psi \\left(\\mathbf {F} _ {v} \\oplus \\mathbf {M} _ {i} [ k ], \\theta_ {R}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 811, + 468, + 828 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\psi (\\cdot ,\\cdot)$ is a linear projection with a learnable parameter $\\theta_{R}$ , $\\mathbf{F}_r$ is the refined sparse instance feature. In this case, the model can give a reasonable sparse instance considering the difference between the input observed feature and the stored", + "bbox": [ + 76, + 839, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "observed feature in the memory bank. The loss function of the memory refinement module is:", + "bbox": [ + 498, + 90, + 893, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {2} = \\left\\| \\mathbf {F} _ {i n s} - \\mathbf {F} _ {r} \\right\\| _ {2}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 133, + 893, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the past feature $\\mathbf{F}_{\\mathbf{p}}$ is concatenated with the refinement feature $\\mathbf{F}_r$ to predict the future trajectory. Concretely, we use the pre-trained decoder $f_{d}(\\cdot)$ in the first stage to achieve the prediction, as follows:", + "bbox": [ + 498, + 161, + 893, + 222 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {Y} = f _ {d} \\left(\\mathbf {F} _ {p} \\oplus \\mathbf {F} _ {r}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 233, + 892, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{Y}$ is the predicted future trajectory.", + "bbox": [ + 498, + 263, + 777, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Clustering. We adopt the clustering mechanism during inference to further remove the modality redundancy of the retrieved sparse instances. By retrieving $C$ ( $C >> k$ ) sparse instances corresponding to the past trajectory from the well-trained memory, we can generate $C$ refined sparse instances through the refinement network. Then we cluster the $C$ refined sparse instances into $K$ multimodal proposals. The clustered multimodal proposals will guide the multimodal trajectory prediction instead of sparse instances in inference. Ablation studies verify the effectiveness of this post-process.", + "bbox": [ + 496, + 279, + 893, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Model Training & Inference", + "text_level": 1, + "bbox": [ + 500, + 438, + 753, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. SICNet is trained by a two-stage strategy. In the first stage, we train the encoders $f_{p}(\\cdot)$ , $f_{y}(\\cdot)$ and the decoder $f_{d}(\\cdot)$ in the sparse instance representation module with the reconstruction loss $\\mathcal{L}_1$ . Then we use the memory mechanism to store the sparse instances generated during the reconstruction process. In the second stage, we fix the parameters of $f_{p}(\\cdot)$ , $f_{y}(\\cdot)$ and $f_{d}(\\cdot)$ , and train the memory refinement module with the refinement loss $\\mathcal{L}_2$ .", + "bbox": [ + 496, + 460, + 893, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference. We aim to predict $K$ trajectories to cover the multimodality of future trajectories. During inference, we select the top- $C$ past memory items by feature similarity and then retrieve corresponding $C$ sparse instance items from the instance memory bank, where $C >> K$ is to improve the prediction diversity. Then we obtained $C$ refinement sparse instances through the memory refinement module. Subsequently, the $C$ sparse instance items are clustered into $K$ multimodal proposals. Finally, we feed the combination of the past feature and multimodal proposals into the decoder to generate $K$ predicted future trajectories.", + "bbox": [ + 496, + 583, + 893, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments and Discussions", + "text_level": 1, + "bbox": [ + 498, + 762, + 767, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 500, + 786, + 700, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Datasets. We evaluate our proposed method on two benchmark datasets, i.e., ETH-UCY [31, 21] and Stanford Drone Dataset (SDD) [33]. ETH-UCY is the most commonly used pedestrian trajectory prediction dataset captured in the bird's eye view. It contains five subsets, where the ETH [31] includes ETH and HOTEL subsets, and the", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9767", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/280f2c4e25c1cc2a539c41ec821687ae8238491c091ee7d60f7d6f3b5d4e5a00.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVenue/YearInputETHHOTELUNIVZARA1ZARA2AVG
Social LSTM [1]CVPR2016P+S1.09/2.350.79/1.760.67/1.400.47/1.000.56/1.170.72/1.54
Social GAN [13]CVPR2018P+S0.87/1.620.67/1.370.76/1.520.35/0.680.42/0.840.61/1.21
SoPhie [34]CVPR2019P+S0.70/1.430.76/1.670.54/1.240.30/0.630.38/0.780.51/1.15
MANTRA [28]CVPR2020P+S0.48/0.880.17/0.330.37/0.810.27/0.580.30/0.670.32/0.65
PECNet [27]ECCV2020P+S0.54/0.870.18/0.240.35/0.600.22/0.390.17/0.300.29/0.48
SGCN [37]CVPR2021P+S0.63/1.030.32/0.550.37/0.700.29/0.530.25/0.450.37/0.65
AgentFormer [54]ICCV2021P+S0.45/0.750.14/0.220.25/0.450.18/0.300.14/0.240.23/0.39
PCCSNet [41]ICCV2021P0.28/0.540.11/0.190.29/0.600.21/0.440.15/0.340.21/0.42
CAGN [8]AAAI2022P+S0.41/0.650.13/0.230.32/0.540.21/0.380.16/0.330.25/0.43
SIT [36]AAAI2022P+S0.39/0.620.14/0.220.27/0.470.19/0.330.16/0.290.23/0.38
MemoNet [49]CVPR2022P+S0.40/0.610.11/0.170.24/0.430.18/0.320.14/0.240.21/0.35
Ours-P0.27/0.450.11/0.160.26/0.460.19/0.330.13/0.260.19/0.33
", + "bbox": [ + 80, + 88, + 893, + 301 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6f8846fd4791cee1c4f8e677af96434bb27cb6ed5525e811994d3c02b2cca5d4.jpg", + "table_caption": [ + "Table 1. Comparison with state-of-the-art methods on ETH-UCY in ADE/FDE. P and S indicate inputting the observed history points and social interaction information, respectively. All methods input the observed 8 time steps and output the predicted 12 time steps. Bold indicates the best performance. Underline indicates the second best performance. The Lower the better." + ], + "table_footnote": [], + "table_body": "
MethodVenue/YearInputADE/FDE
Social LSTM [1]CVPR2016P+S31.19/56.97
Social GAN [13]CVPR2018P+S27.23/41.44
SoPhie [34]CVPR2019P+S16.27/29.38
MANTRA [28]CVPR2020P+S8.96/17.76
PECNet [27]ECCV2020P+S9.96/15.88
PCCSNet [41]ICCV2021P8.62/16.16
CAGN [8]AAAI2022P+S9.42/15.93
SIT [36]AAAI2022P+S9.13/15.42
MemoNet [49]CVPR2022P+S8.56/12.66
Ours-P8.44/13.65
", + "bbox": [ + 83, + 377, + 464, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison with state-of-the-art methods on SDD in ADE/FDE. P and S indicate inputting the observed history points and social interaction information, respectively. All methods input the observed 8 time steps and output the predicted 12 time steps. The Lower the better.", + "bbox": [ + 75, + 575, + 470, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "UCY [21] includes UNIV, ZARA1, and ZARA2 subsets. On ETH-UCY, we use the leave-out-one method for training on four subsets and testing on the other one following previous efforts [40, 54, 36]. SDD [33] is a large pedestrian trajectory prediction dataset also captured in bird's eye view. It contains 20 scenes of more than 10,000 trajectories collected from college campuses and is much larger than ETH-UCY. On SDD, we use a prior train-test split for training and testing according to previous methods [13, 27, 36]. We observe the historical trajectories of 8 time steps and predict the subsequent 12 time steps on both ETH-UCY and SDD.", + "bbox": [ + 75, + 671, + 470, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following previous methods [49, 36, 8, 41, 27], we employ two commonly used metrics, i.e., Average Displacement Error (ADE) and Final Displacement Error (FDE), to evaluate the trajectory prediction perfor", + "bbox": [ + 76, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mance. ADE measures the average $\\ell_2$ -norm distance between all points of the ground truth and predicted trajectory. FDE measures the $\\ell_2$ -norm distance between the destination points of the ground truth and predicted trajectory. We use the best-of-20 metrics following previous methods [49, 41].", + "bbox": [ + 496, + 378, + 893, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. In our experiments, the two encoders are all 3-layer BiLSTM [56] architecture with hidden sizes of 48. The decoder is 3-layer BiLSTM [56] architecture with hidden size of 96. The dimensions of $\\theta_{M}$ , $\\theta_{V}$ and $\\theta_{R}$ are set to 48, 48 and 96, respectively. The parameter $\\mu$ of mask threshold is 0.2. The parameter $\\xi$ in memory training is 0.0001. The Adam optimizer is used to train our model by 100 epochs with a learning rate of 0.0005, decaying by 0.5 with an interval of 50.", + "bbox": [ + 496, + 455, + 893, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Analysis", + "text_level": 1, + "bbox": [ + 500, + 603, + 702, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On ETH-UCY. As shown in Table 1, we compare our method with eleven state-of-the-art methods on the ETH-UCY dataset in past six years. Among them, MemoNet [49], PECNet [27], and CAGN [8] are goal-conditioned methods. MANTRA [28] and PCCSNet [41] are instance-conditioned methods. SoPhie [34], SGCN [37] and AgentFormer [54] generate multimodal trajectories by latent variables. The experiment results show that generating multimodal predictions by goals or instance is better than by latent variables.", + "bbox": [ + 496, + 627, + 893, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, the results indicate that our method significantly outperforms all the competing methods on both ETH and UCY. Compared with the previous best method MemoNet [49], our method further improves the performance on ADE and FDE. Meanwhile, Compared with the other instance-conditioned methods, PCCSNet [41], and MANTRA [28], our method also achieve the best performance. The underlying reason could be that our method captures the effective intermediate process compared with", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9768", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4333ccbba8d0781e15db4445fdac2c27065f182524668094317a001115436095.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodK=15K=10K=5
PECNet [27]10.66/17.8511.79/21.0014.49/28.08
PCCSNet [41]9.27/17.6210.36/20.1212.57/24.74
CAGN [8]10.01/17.4911.06/19.8913.85/26.87
MemoNet [49]9.62/15.3611.26/19.6414.43/27.45
Ours9.13/15.3110.27/18.1312.82/23.96
", + "bbox": [ + 81, + 89, + 467, + 200 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2a8859e49e5a7712d918f8ba4ae00a84cb04a713319fb6d484f23d8b4e5fc011.jpg", + "table_caption": [ + "Table 3. Experiments on different best-of- $K$ predictions on SDD in ADE/FDE. The lower the better." + ], + "table_footnote": [], + "table_body": "
MethodSIREFCLADE/FDE
Baseline0.46/0.94
(1)✓✓0.25/0.50
(2)✓✓0.27/0.56
(3)✓✓0.22/0.39
(4)✓✓✓0.19/0.33
", + "bbox": [ + 91, + 250, + 455, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the goal-conditioned methods and is more flexible compared with the instance-conditioned methods.", + "bbox": [ + 75, + 422, + 468, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To show the superiority of our method more clearly, we do not use extra information, such as social interaction and map information, as previous approaches. Nevertheless, our method still achieves the best performance, clearly demonstrating the superiority of our method using sparse instance.", + "bbox": [ + 75, + 453, + 470, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On SDD. We compare our method with nine state-of-the-art methods on the SDD dataset in Table 2. For ADE, our method outperforms all the competing methods. For FDE, our method also achieves comparable performance with the best method MemoNet [49] and exceeds all other methods. The results manifest that the sparse instance representation can comprehensively describe the trajectory multimodality rather than goal-conditioned methods and thus achieves competitive performance. We speculate that our method underperforms MemoNet [49] on FDE because the sparse instance benefits more on the whole trajectory prediction (i.e., ADE) than goal prediction (i.e., FDE).", + "bbox": [ + 75, + 530, + 472, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 724, + 232, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different Best-of- $K$ Predictions. Previous methods commonly use best-of- $K$ (usually $K = 20$ ) as the quantified metric of multimodal trajectory prediction. To further validate the adaptability and effectiveness of our proposed sparse instance representation, we conduct an experiment on various best-of- $K$ predictions with $K = 5, 10, 15$ on SDD. We compare our method with the top three performance goal-conditioned methods, i.e., MemoNet [49], CAGN [8] and PECNet [27] and the best performance instance-conditioned method PCCSNet [41]. As shown in Table 3, our method", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/74dc841b0ef707b621c76fddddbb7afd64ac4003629c9049eaba658c13e8de42.jpg", + "table_caption": [ + "Table 4. Ablation study of each component of our method on ETH-UCY dataset in ADE/FDE. The lower the better." + ], + "table_footnote": [], + "table_body": "
μ00.10.20.30.40.5
ADE0.210.210.190.200.200.21
FDE0.340.330.330.330.340.37
μ0.60.70.80.91.0-
ADE0.250.280.260.240.21-
FDE0.450.540.480.430.38-
", + "bbox": [ + 514, + 88, + 880, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation study of different mask threshold values on the ETH-UCY dataset. The lower the better.", + "bbox": [ + 498, + 220, + 890, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/835407cf1dfcec615275cc63c3172e8a9470844281c907fd888c92dea4850361.jpg", + "image_caption": [ + "Figure 6. Experiments of the sparse instance analysis on the ETH-UCY dataset. We replace our proposed sparse instance with the goal and instance for the multimodal prediction." + ], + "image_footnote": [], + "bbox": [ + 501, + 261, + 888, + 404 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "achieves significant performance on all experimental settings. Interestingly, our method surpasses the previous best method MemoNet [49] when $K = 5, 10, 15$ , despite underperforming MemoNet [49] when $K = 20$ in Table 2. It shows the adaptability of our method of using the sparse instance, which also performs significantly with a small $K$ .", + "bbox": [ + 496, + 474, + 893, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Contribution of Each Component. We divide our proposed SICNet into three components, i.e., the sparse instance representation (SI), the memory refinement module (REF), and the clustering post-process (CL). Ablation experiments of different combinations of the three components are evaluated on the ETH-UCY dataset. As shown in Table 4, all of the three components contribute to the performance improvement, which demonstrates their effectiveness.", + "bbox": [ + 496, + 566, + 893, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of Mask Threshold. As shown in Table 5, we conduct experiments for different mask threshold values $\\mu$ on the ETH-UCY dataset. It indicates that the performance will decrease when the mask threshold value $\\mu$ is too small or too large. When $\\mu = 0$ , the mask is useless and the whole instance is reserved, thus causing the stochasticity loss. When $\\mu = 1$ , it denotes the whole instance is masked, thus causing no guidance information for the multimodal prediction. When $\\mu = 0.2$ , it achieves the best performance.", + "bbox": [ + 496, + 688, + 893, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of Sparse Instance. As shown in Table 6, we replace the sparse instance in SICNet with goal-conditioned and instance-conditioned representations on the ETH-UCY dataset. Experiment results show that using the sparse instance achieves the best performance, which indicates the", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9769", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/114a206eefaa53bceee93d73a4d0d771baaf1ca2d815ddfd9df79b89a5f6c7c9.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison of our method with the goal-conditioned method PECNet [27] on the SDD dataset. Given the observed trajectories, we illustrate the ground truth paths and predicted trajectories by SICNet and the goal-conditioned method PECNet [27] for four different scenes. It shows our prediction is closer to the ground truth than PECNet [27], especially for the intermediate process." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/44482aaf343e512b106d1580fa14c76cf147044e997bd3d6084dbae7d2ec4298.jpg", + "image_caption": [ + "Figure 8. Qualitative comparison of our method with the instance-conditioned method PCCSNet [41] on the ETH-UCY dataset. Given the observed trajectories, we illustrate the ground truth paths and predicted trajectories by SICNet and the instance-conditioned method PCCSNet [41] for four different scenes. It shows our prediction is closer to the ground truth than PCCSNet [41]." + ], + "image_footnote": [], + "bbox": [ + 81, + 321, + 890, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "superiority of our method compared with goal-conditioned and instance-conditioned approaches.", + "bbox": [ + 75, + 564, + 470, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 76, + 604, + 269, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare the visualized trajectories with the well-accepted goal-conditioned method PECNet [27] on SDD in Figure 7, and the instance-conditioned method PCCSNet [41] on ETH-UCY in Figure 8. As shown in Figure 7, trajectories in Coupa show that our method achieves more precise trajectory predictions than PECNet. Trajectories in Little and Quad show that both PECNet and our method can predict destinations close to the ground truth, while our method can predict more precise intermediate trajectories due to our proposed sparse instance representation. Trajectories in Hyang show that our method can also give better predictions, although the ground truth trajectories are irregular. The results verify the superiority of our method compared with goal-conditioned methods. Moreover, as shown in Figure 8, the visualization results show that our prediction is closer to the ground truth than PCCSNet [41], which indicates the superiority of our SICNet compared with the instance-conditioned approaches.", + "bbox": [ + 75, + 628, + 473, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 563, + 619, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present a simple yet effective sparse-instance-conditioned network for pedestrian multimodal trajectory prediction, which leverages sparse instances to guide multimodal prediction. Moreover, we propose a memory refinement module to reduce recall errors when meeting unfamiliar trajectories during inference. Extensive experiments show that our method performs better than previous methods, and ablation studies demonstrate that our sparse-instance-conditioned method can generate multimodal predictions more accurately and adaptively compared with goal-conditioned and instance-conditioned approaches.", + "bbox": [ + 496, + 592, + 895, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 781, + 660, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported partly by National Key R&D Program of China under Grant 2021YFB1714700, NSFC under Grants 62088102 and 62106192, Natural Science Foundation of Shaanxi Province under Grant 2022JC-41, and Fundamental Research Funds for the Central Universities under Grant XTR042021005.", + "bbox": [ + 496, + 810, + 895, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9770", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social LSTM: Human trajectory prediction in crowded spaces. In CVPR, pages 961–971, 2016.", + "[2] Inhwan Bae, Jin-Hwi Park, and Hae-Gon Jeon. Nonprobability sampling network for stochastic human trajectory prediction. In CVPR, pages 6477-6487, 2022.", + "[3] Guangyi Chen, Junlong Li, Jiwen Lu, and Jie Zhou. Human trajectory prediction via counterfactual analysis. In ICCV, pages 9824-9833, 2021.", + "[4] Xinlei Chen and Abhinav Gupta. Spatial memory for context reasoning in object detection. In ICCV, pages 4086-4096, 2017.", + "[5] Chiho Choi, Joon Hee Choi, Jiachen Li, and Srikanth Malla. Shared cross-modal trajectory prediction for autonomous driving. In CVPR, pages 244-253, 2021.", + "[6] Lingwei Dang, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. MSR-GCN: Multi-scale residual graph convolution networks for human motion prediction. In ICCV, pages 11447-11456, 2021.", + "[7] Hanming Deng, Yang Hua, Tao Song, Zongpu Zhang, Zhengui Xue, Ruhui Ma, Neil Robertson, and Haibing Guan. Object guided external memory network for video object detection. In ICCV, pages 6677-6686, 2019.", + "[8] Jinghai Duan, Le Wang, Chengjiang Long, Sanping Zhou, Fang Zheng, Liushuai Shi, and Gang Hua. Complementary attention gated network for pedestrian trajectory prediction. In AAAI, pages 542-550, 2022.", + "[9] Zhihong Fu, Qingjie Liu, Zehua Fu, and Yunhong Wang. STMTrack: Template-free visual tracking with space-time memory networks. In CVPR, pages 13769-13778, 2021.", + "[10] Harshayu Girase, Haiming Gang, Srikanth Malla, Jiachen Li, Akira Kanehara, Karttikeya Mangalam, and Chiho Choi. LOKI: Long term and key intentions for trajectory prediction. In ICCV, pages 9803-9812, 2021.", + "[11] Dong Gong, Frederic Z. Zhang, Javen Qinfeng Shi, and Anton van den Hengel. Memory-augmented dynamic neural relational inference. In ICCV, pages 11843-11852, 2021.", + "[12] Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory prediction via motion indeterminacy diffusion. In CVPR, pages 17113-17122, 2022.", + "[13] Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi. Social GAN: socially acceptable trajectories with generative adversarial networks. In CVPR, pages 2255–2264, 2018.", + "[14] Adam Houenou, Philippe Bonnifait, Véronique Cherfaoui, and Wen Yao. Vehicle trajectory prediction based on motion model and maneuver recognition. In IROS, pages 4363-4369, 2013.", + "[15] Boris Ivanovic and Marco Pavone. The trajectory: Probabilistic multi-agent trajectory modeling with dynamic spatiotemporal graphs. In ICCV, pages 2375-2384, 2019.", + "[16] Lukasz Kaiser, Ofir Nachum, Aurko Roy, and Samy Bengio. Learning to remember rare events. In ICLR, 2017." + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Youngjin Kim, Minjung Kim, and Gunhee Kim. Memorization precedes generation: Learning unsupervised gans with memory networks. In ICLR, 2018.", + "[18] Sumith Kulal, Jiayuan Mao, Alex Aiken, and Jiajun Wu. Programmatic concept learning for human motion description and synthesis. In CVPR, pages 13843-13852, 2022.", + "[19] Mihee Lee, Samuel S. Sohn, Seonghyeon Moon, Sejong Yoon, Mubbasir Kapadia, and Vladimir Pavlovic. MUSE-VAE: Multi-scale vae for environment-aware long term trajectory prediction. In CVPR, pages 2221–2230, 2022.", + "[20] Sangho Lee, Jinyoung Sung, Youngjae Yu, and Gunhee Kim. A memory network approach for story-based temporal summarization of $360^{\\circ}$ videos. In CVPR, pages 1410-1419, 2018.", + "[21] Alon Lerner, Yiorgos Chrysanthou, and Dani Lischinski. Crowds by example. Comput. Graph. Forum, 26(3):655-664, 2007.", + "[22] Peizhao Li, Pu Wang, Karl Berntorp, and Hongfu Liu. Exploiting temporal relations on radar perception for autonomous driving. In CVPR, pages 17071-17080, 2022.", + "[23] Shijie Li, Yanying Zhou, Jinhui Yi, and Juergen Gall. Spatial-temporal consistency network for low-latency trajectory forecasting. In ICCV, pages 1940-1949, 2021.", + "[24] Jialun Liu, Yifan Sun, Feng Zhu, Hongbin Pei, Yi Yang, and Wenhui Li. Learning memory-augmented unidirectional metrics for cross-modality person re-identification. In CVPR, pages 19366-19375, 2022.", + "[25] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In ICCV, pages 13588-13597, 2021.", + "[26] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In ICCV, pages 15233-15242, 2021.", + "[27] Karttikeya Mangalam, Harshayu Girase, Shreyas Agarwal, Kuan-Hui Lee, Ehsan Adeli, Jitendra Malik, and Adrien Gaidon. It is not the journey but the destination: Endpoint conditioned trajectory prediction. In ECCV, pages 759-776, 2020.", + "[28] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. *Mantra: Memory augmented networks for multiple trajectory prediction*. In *CVPR*, pages 7143–7152, 2020.", + "[29] Abduallah Mohamed, Kun Qian, Mohamed Elhoseiny, and Christian Claudel. Social-STGCNN: A social spatio-temporal graph convolutional neural network for human trajectory prediction. In CVPR, pages 14424-14432, 2020.", + "[30] Alessio Monti, Angelo Porrello, Simone Calderara, Pasquale Coscia, Lamberto Ballan, and Rita Cucchiara. How many observations are enough? knowledge distillation for trajectory forecasting. In CVPR, pages 6553-6562, 2022.", + "[31] Stefano Pellegrini, Andreas Ess, Konrad Schindler, and Luc Van Gool. You'll never walk alone: Modeling social behavior for multi-target tracking. In ICCV, pages 261-268, 2009.", + "[32] Zheng Qin, Sanping Zhou, Le Wang, Jinghai Duan, Gang Hua, and Wei Tang. Motiontrack: Learning robust short-term and long-term motions for multi-object tracking. In CVPR, pages 17939-17948, 2023." + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9771", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Alexandre Robicquet, Amir Sadeghian, Alexandre Alahi, and Silvio Savarese. Learning Social Etiquette: Human trajectory understanding in crowded scenes. In ECCV, pages 549-565, 2016.", + "[34] Amir Sadeghian, Vineet Kosaraju, Ali Sadeghian, Noriaki Hirose, Hamid Rezatofighi, and Silvio Savarese. SoPhie: An attentive gan for predicting paths compliant to social and physical constraints. In CVPR, pages 1349-1358, 2019.", + "[35] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory: Dynamically-feasible trajectory forecasting with heterogeneous data. In ECCV, pages 683-700, 2020.", + "[36] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Fang Zheng, Nanning Zheng, and Gang Hua. Social interpretable tree for pedestrian trajectory prediction. In AAAI, pages 2235-2243, 2022.", + "[37] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Mo Zhou, Zhenxing Niu, and Gang Hua. SGCN: Sparse graph convolution network for pedestrian trajectory prediction. In CVPR, pages 8990-8999, 2021.", + "[38] Aravind Srinivas, Tsung-Yi Lin, Niki Parmar, Jonathon Shlens, Pieter Abbeel, and Ashish Vaswani. Bottleneck transformers for visual recognition. In CVPR, pages 16519-16529, 2021.", + "[39] Hao Sun, Zhiqun Zhao, and Zhihai He. Reciprocal learning networks for human trajectory prediction. In CVPR, pages 7416-7425, 2020.", + "[40] Jianhua Sun, Qinhong Jiang, and Cewu Lu. Recursive social behavior graph for trajectory prediction. In CVPR, pages 660-669, 2020.", + "[41] Jianhua Sun, Yuxuan Li, Hao-Shu Fang, and Cewu Lu. Three steps to multimodal trajectory prediction: Modality clustering, classification and synthesis. In ICCV, pages 13250-13259, 2021.", + "[42] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, Wenyu Liu, and Bin Xiao. Deep high-resolution representation learning for visual recognition. IEEE T-PAMI, 43(10):3349-3364, 2020.", + "[43] Jingke Wang, Tengju Ye, Ziqing Gu, and Junbo Chen. LTP: Lane-based trajectory prediction for autonomous driving. In CVPR, pages 17134-17142, 2022.", + "[44] Jürgen Wiest, Matthias Höffken, Ulrich Kreßel, and Klaus Dietmayer. Probabilistic trajectory prediction with gaussian mixture models. In IV, pages 141-146, 2012.", + "[45] Yuxuan Wu, Le Wang, Sanping Zhou, Jinghai Duan, Gang Hua, and Wei Tang. Multi-stream representation learning for pedestrian trajectory prediction. In AAAI, pages 2875-2882, 2023.", + "[46] Zifeng Wu, Chunhua Shen, and Anton Van Den Hengel. Wider or deeper: Revisiting the resnet model for visual recognition. PR, 90:119-133, 2019.", + "[47] Haozhe Xie, Hongxun Yao, Shangchen Zhou, Shengping Zhang, and Wenxiu Sun. Efficient regional memory network for video object segmentation. In CVPR, pages 1286-1295, 2021." + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Bo Xu, Han Huang, Cheng Lu, Ziwen Li, and Yandong Guo. Virtual multi-modality self-supervised foreground matting for human-object interaction. In ICCV, pages 428-437, 2021.", + "[49] Chenxin Xu, Weibo Mao, Wenjun Zhang, and Siheng Chen. Remember intentions: Retrospective-memory-based trajectory prediction. In CVPR, pages 6488-6497, 2022.", + "[50] Hao Xue, Du Q. Huynh, and Mark Reynolds. SS-LSTM: A hierarchical LSTM model for pedestrian trajectory prediction. In WACV, pages 1186-1194, 2018.", + "[51] Gang Yang, Man Zhou, Keyu Yan, Aiping Liu, Xueyang Fu, and Fan Wang. Memory-augmented deep conditional unfolding network for pan-sharpening. In CVPR, pages 1788-1797, 2022.", + "[52] Tianyu Yang and Antoni B. Chan. Visual tracking via dynamic memory networks. IEEE T-PAMI, 43(1):360-374, 2021.", + "[53] Jiyang Yu, Jingen Liu, Liefeng Bo, and Tao Mei. Memory-augmented non-local attention for video super-resolution. In CVPR, pages 17834-17843, 2022.", + "[54] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris Kitani. Agent-Former: Agent-aware transformers for socio-temporal multiagent forecasting. In ICCV, pages 9793–9803, 2021.", + "[55] He Zhao and Richard P. Wildes. Where are you heading? dynamic trajectory prediction with expert goal examples. In ICCV, pages 7609-7618, 2021.", + "[56] Zheng Zhao, Weihai Chen, Xingming Wu, Peter CY Chen, and Jingmeng Liu. LSTM network: a deep learning approach for short-term traffic forecast. IET Intell. Transp. Syst., 11(2):68-75, 2017.", + "[57] Desen Zhou, Zhichao Liu, Jian Wang, Leshan Wang, Tao Hu, Errui Ding, and Jingdong Wang. Human-object interaction detection via disentangled transformer. In CVPR, pages 19568-19577, 2022." + ], + "bbox": [ + 501, + 90, + 893, + 559 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9772", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Sparse Instance Conditioned Multimodal Trajectory Prediction/5a87e2e4-dac0-41f3-9966-85985369c6c6_model.json b/2023/Sparse Instance Conditioned Multimodal Trajectory Prediction/5a87e2e4-dac0-41f3-9966-85985369c6c6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4b05a7b22b9dd0cc12f01b9255936e2a5c7755b6 --- /dev/null +++ b/2023/Sparse Instance Conditioned Multimodal Trajectory Prediction/5a87e2e4-dac0-41f3-9966-85985369c6c6_model.json @@ -0,0 +1,2189 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.165, + 0.131, + 0.806, + 0.154 + ], + "angle": 0, + "content": "Sparse Instance Conditioned Multimodal Trajectory Prediction" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.18, + 0.802, + 0.272 + ], + "angle": 0, + "content": "Yonghao Dong\\(^{1}\\) Le Wang\\(^{1*}\\) Sanping Zhou\\(^{1}\\) Gang Hua\\(^{2}\\) \n\\(^{1}\\)National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University \n\\(^{2}\\)Wormpex AI Research" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.304, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.474, + 0.701 + ], + "angle": 0, + "content": "Pedestrian trajectory prediction is critical in many vision tasks but challenging due to the multimodality of the future trajectory. Most existing methods predict multimodal trajectories conditioned by goals (future endpoints) or instances (all future points). However, goal-conditioned methods ignore the intermediate process and instance-conditioned methods ignore the stochasticity of pedestrian motions. In this paper, we propose a simple yet effective Sparse Instance Conditioned Network (SICNet), which gives a balanced solution between goal-conditioned and instance-conditioned methods. Specifically, SICNet learns comprehensive sparse instances, i.e., representative points of the future trajectory, through a mask generated by a long short-term memory encoder and uses the memory mechanism to store and retrieve such sparse instances. Hence SICNet can decode the observed trajectory into the future prediction conditioned on the stored sparse instance. Moreover, we design a memory refinement module that refines the retrieved sparse instances from the memory to reduce memory recall errors. Extensive experiments on ETH-UCY and SDD datasets show that our method outperforms existing state-of-the-art methods. In addition, ablation studies demonstrate the superiority of our method compared with goal-conditioned and instance-conditioned approaches." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.73, + 0.21, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.755, + 0.472, + 0.877 + ], + "angle": 0, + "content": "Pedestrian trajectory prediction aims to predict pedestrians' future paths given their observed trajectories. It plays an important role in autonomous driving [22, 43, 5], human motion prediction [6, 12, 19], video surveillance [57, 48, 18, 32], and visual recognition [38, 42, 46]. Despite the recent advancement [30, 49, 8, 36, 10], pedestrian trajectory prediction is still challenging due to the multimodality of future trajectories. As the pedestrian's motions are stochastic and" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.306, + 0.695, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.306, + 0.888, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.383, + 0.695, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.382, + 0.889, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.457, + 0.889, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.476, + 0.895, + 0.574 + ], + "angle": 0, + "content": "Figure 1. Illustration of different multimodal prediction methods. (a) Goals, instances and sparse-instances prepared for the prediction. (b) Goal-conditioned methods. (c) Instance-conditioned methods. (d) Sparse-instance-conditioned methods. The sparse-instance-conditioned method is more accurate in in-between states than the goal-conditioned method and more flexible than the instance-conditioned method, thus making the better prediction." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.599, + 0.892, + 0.628 + ], + "angle": 0, + "content": "indeterminate [29], multiple future paths are possible given their current states." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.895, + 0.901 + ], + "angle": 0, + "content": "To handle multimodality, many methods [27, 49, 55] transform the multimodal trajectories into multimodal goals. They first predict various goals as modality representations and then sample multiple goals to generate diverse future trajectories. However, they only focus on goals while neglecting the intermediate movement process. As illustrated in Figure 1(b), although with accurate goals, the intermediate process could be far from the ground truth. Thus, the multimodal goals are not equal to the multimodal future trajectories. It is necessary to consider the in-between states to comprehensively describe the multimodality of trajectory. Moreover, some other work [27, 28] predicts multimodal trajectories conditioned by instances. However, the instance-conditioned methods will lead to motion stochasticity loss and modality redundancy, resulting in performance degradation and computational resource waste. As shown in Figure 1(c), due to the reduced randomness of the instance-conditioned approach, the prediction retains the same bias" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9763" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.404, + 0.106 + ], + "angle": 0, + "content": "when the instance deviates from the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.108, + 0.47, + 0.483 + ], + "angle": 0, + "content": "To close this gap, we present a Sparse Instance Conditioned Network (SICNet). In our model, a sparse instance, which can give a comprehensive and adaptive description of future modality, is constructed by element-wise multiplication between the future trajectory and a learned mask. Hence, SICNet can predict the future trajectory more precisely guided by the sparse instance. To obtain the related sparse instances during inference, we adopt the memory mechanism to store non-redundant sparse instances corresponding to all observed trajectories. Therefore, SICNet can recall its modality's most related sparse instance and jointly generate the prediction during inference. Moreover, when the memory meets an unfamiliar observed trajectory, i.e., a trajectory quite different from stored items, the memory mechanism may recall false experiences. Hence we propose a memory refinement module with a refinement loss to reduce such recall errors by bridging the gap between the recall and the ground truth. For multimodal trajectories generation, we can retrieve a certain number of the most likely sparse instances from the external memory and then generate corresponding multimodal trajectories. In addition, to further remove the redundancy from the retrieved sparse instances, we use the cluster mechanism as a post-process to obtain succinct multimodal proposals for the prediction instead of the sparse instance during inference." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.485, + 0.47, + 0.71 + ], + "angle": 0, + "content": "Extensive experiments and ablation studies on two popular pedestrian trajectory prediction datasets, i.e., ETH-UCY [31, 21] and Stanford Drone Dataset (SDD) [33], demonstrate the superiority of our method over prior state-of-the-art methods. In summary, our contributions are four-fold: (1) We give a balanced solution, i.e., a novel sparse instance, between the goal and instance to guide the multimodal trajectory prediction. (2) We propose a simple yet effective sparse instance conditioned network for the multimodal trajectory prediction. (3) We propose a memory refinement module for the external memory to reduce recall errors when meeting unfamiliar observed trajectories. (4) Extensive experiments and ablation studies demonstrate the superiority and flexibility of our method compared with the goal-conditioned and instance-conditioned approaches." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.724, + 0.225, + 0.74 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.75, + 0.47, + 0.885 + ], + "angle": 0, + "content": "Multimodal Pedestrian Trajectory Prediction. Pedestrian trajectory prediction [2, 23, 3, 39, 45] aims to predict pedestrians' future trajectories based on their observed trajectories. There are some early studies on the deterministic trajectory prediction [1, 44, 14, 50]. Due to the strong randomness and uncertainty of pedestrian motions, there is no single correct future trajectory prediction. Motivated by this, Social GAN [13] proposes the concept of multimodal pedestrian trajectory prediction and emphasizes its importance." + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.887, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Earlier work generates multimodal trajectories through" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.257 + ], + "angle": 0, + "content": "latent variables. Social GAN [13] proposes a Generative Adversarial Network based on a social pooling mechanism to generate a multimodal trajectory distribution. SGCN [37] generates multimodal trajectories by modeling the future predictions as a bi-variate Gaussian distribution. The CVAE-based methods [15, 35, 8] generate multimodal trajectories by sampling latent variables from a learned latent space. However, the latent variable conditioned method is persecuted by the problem of model collapse, thus resulting in performance degradation. Moreover, the latent variable conditioned methods also suffer in interpretability." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.547 + ], + "angle": 0, + "content": "To address the model collapse and interpretability issues, PECNet [27] proposes a goal-conditioned approach, which transforms multimodal trajectories into multimodal goals and then samples multiple goals to generate corresponding diverse predictions. After that, amounts of goal-conditioned studies [55, 8, 49] emerge and show superior performance. However, goal-conditioned methods represent different modalities with different possible goals while neglecting the intermediate states, thus suffering in performance improvement. In addition, PCCSNet[27] and MANTRA [28] propose instance-conditioned approaches, which use all points of the future instance to describe each modality precisely. However, it could lead to stochasticity loss and modality redundancy, resulting in performance degradation and computational resource waste. Moreover, Y-net [26] use goals and waypoints to make the prediction. However, the waypoints' number and time step locations are manually determined in Y-net [26], which may decrease the prediction accuracy by manually choosing errors." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.551, + 0.892, + 0.655 + ], + "angle": 0, + "content": "In contrast, we give a balanced way, i.e., a novel sparse instance, between the goal and instance to guide the multimodal trajectory prediction. The sparse instance is constructed by the learned representative points of the instance, which can improve modality representation accuracy compared with goal-conditioned methods and preserve stochasticity compared with instance-conditioned methods." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.66, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Memory Network. The memory-augmented network [51, 11, 25, 24, 53] uses an external memory module to store critical information and has been widely applied in many areas like detection [7, 4], tracking [52, 9], segmentation [47], image generation [17], and video summarization [20]. Memory GAN [17] proposed a novel end-to-end unsupervised GAN network with a learnable memory network that effectively learns a highly multimodal latent space without suffering from structural discontinuity and forgetting problems. OGEMN [7] proposes the first object-guided external memory network for the online video object detection. STMTrack [52] proposes a novel tracking framework built on a space-time memory network, which can fully use historical information related to the target for better adapting to appearance variations during tracking. RMNet [47] memorizes the target object regions, effectively alleviating" + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9764" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.092, + 0.891, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.27, + 0.895, + 0.327 + ], + "angle": 0, + "content": "Figure 2. The framework of our proposed SICNet. SICNet adopts a two-stage training strategy. The sparse instance representation module generates the sparse instances by a learned mask and then trains an external memory to store and retrieve them. The memory refinement module refines the retrieved sparse instances from memory to reduce the searching errors. We cluster the refined sparse instances into multimodal proposals to remove redundancy further. The cluster only exists in inference to save training costs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.47, + 0.366 + ], + "angle": 0, + "content": "objects' ambiguity and reducing computational complexity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.473, + 0.639 + ], + "angle": 0, + "content": "Two previous memory-based work, MANTRA [28] and MemoNet [49], also use the memory mechanism in multimodal trajectory prediction. However, the difference are mainly in three aspects: (1) The contribution of our proposed SICNet focuses on a novel representation of the trajectory modality, i.e., a learned sparse instance, while MANTRA [28] uses the whole instance and MemoNet [49] uses the goal to represent modalities of multimodal trajectories; (2) We propose a memory refinement module to reduce recall errors when meeting unfamiliar observed trajectories during inference, while MANTRA [28] and MemoNet [49] ignore such circumstances. (3) To demonstrate the superiority of our proposed sparse instance, we do not use extra social interaction and map information, while MANTRA [28] use extra map information and MemoNet [49] use social interaction information. Despite this, we also achieve the best performance compared with the above two methods, which indicates the superiority of our method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.649, + 0.208, + 0.665 + ], + "angle": 0, + "content": "3. Our Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.902 + ], + "angle": 0, + "content": "As aforementioned, traditional goal-conditioned methods ignore the intermediate process and instance-conditioned methods ignore the stochasticity of pedestrian motions, which leads to performance degradation. In this section, we present a sparse-instance-conditioned network (SICNet) to close this gap. Our SICNet adopts a two-stage training strategy. In the first stage, we learn the sparse instance, which is used to reconstruct the future trajectory jointly with the observed trajectory. Meanwhile, an external memory is trained to store and retrieve the sparse instance. In the second stage, we refine the retrieved sparse instances from the external memory through a refinement module with a refinement loss to reduce the recall errors. In addition, we use the cluster mechanism to remove redundancy of the retrieved sparse instances further during inference." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.35, + 0.704, + 0.365 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.895, + 0.527 + ], + "angle": 0, + "content": "The objective of pedestrian trajectory prediction is to predict possible future trajectory coordinates based on the observed trajectories. For each target pedestrian, given an observed trajectory \\( X = \\{(x_{t},y_{t})\\}_{t = 1}^{T_{obs}} \\), where \\( (x_{t},y_{t}) \\) is the 2D coordinate at time \\( t \\), we aim to predict the future trajectory coordinates \\( Y = \\{(x_{t},y_{t})\\}_{t = T_{obs} + 1}^{T_{pred}} \\). Note that multimodal pedestrian trajectory prediction requires the model to predict \\( K \\) future trajectories \\( \\{Y_i\\}_{i = 1}^K \\) to account for multimodality, while only one future trajectory (ground truth) is provided for training in the dataset." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.535, + 0.78, + 0.552 + ], + "angle": 0, + "content": "3.2. Sparse Instance Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.558, + 0.895, + 0.649 + ], + "angle": 0, + "content": "In the first training stage, we generate the sparse instance features to represent future modalities and store them by memory mechanism. As shown in Figure 3, we first train the encoder-decoder with a reconstruction loss in this stage. Then we train the memory module to store the sparse instances generated during reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.649, + 0.895, + 0.724 + ], + "angle": 0, + "content": "Feature Encoder. As the trajectory is a temporal sequence, we use two LSTM [56] blocks to build the past encoder \\( f_{p}(\\cdot) \\) and the future encoder \\( f_{y}(\\cdot) \\), respectively. The hidden states of the two encoders are used as their corresponding features, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.734, + 0.892, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {F} _ {p} = f _ {p} (X), \\\\ \\mathbf {F} _ {y} = f _ {y} (Y), \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.894, + 0.84 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_p\\in \\mathbb{R}^{1\\times d}\\) is the past feature extracted from the observed trajectory \\(X\\), and \\(\\mathbf{F}_y\\in \\mathbb{R}^{1\\times d}\\) is the future feature extracted from the future trajectory \\(Y\\). \\(d\\) is the dimension of hidden state in LSTM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Sparse Instance. When inputting an observed trajectory, we want to obtain the corresponding sparse instance of its future modality so that we can construct future trajectories jointly using the observed trajectory and the sparse instance." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9765" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.437, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.167, + 0.473, + 0.222 + ], + "angle": 0, + "content": "Figure 3. The first stage of the training process. In this stage, we train the sparse instance by reconstructing the future trajectory. Then we train an external memory to store and retrieve the learned sparse instances." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.235, + 0.468, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.317, + 0.471, + 0.358 + ], + "angle": 0, + "content": "Figure 4. The second stage of the training process. In this stage, we refine the retrieved sparse instances from memory through the memory refinement module to reduce the recall errors." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.383, + 0.47, + 0.414 + ], + "angle": 0, + "content": "Hence, we first learn an instance mask from the past feature \\(\\mathbf{F}_p\\) through an MLP layer, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.424, + 0.47, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {F} _ {m} = \\phi (\\mathbf {F} _ {p}, \\theta_ {M}), \\\\ \\begin{array}{l} M = \\mathbb {I} (\\delta (\\mathbf {F} _ {m}) \\leq \\mu)), \\end{array} \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.473, + 0.606 + ], + "angle": 0, + "content": "where \\(\\phi (\\cdot ,\\cdot)\\) denotes the MLP layer. \\(\\theta_{M}\\) is a learnable parameter. \\(\\mathbf{F}_m\\) is the learned feature which is used to construct the mask. \\(M\\) denotes the instance mask. \\(\\delta (\\cdot)\\) is sigmoid function to map \\(\\mathbf{F}_m\\) to the mask value [0, 1]. \\(\\mathbb{I}\\) is an indicator function which equals 0 if the inequality holds, otherwise 1. \\(\\mu\\) is the mask threshold. Compared with directly masking instances, masking instances in the high-level feature dimension is more efficient and effective. Hence we generate the sparse instance \\(\\mathbf{F}_{ins}\\) in the feature level dimension, as:" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.619, + 0.47, + 0.635 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {i n s} = \\mathbf {F} _ {y} \\odot M, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.648, + 0.407, + 0.662 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes the element-wise multiplication." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.663, + 0.471, + 0.768 + ], + "angle": 0, + "content": "Feature Decoder. The sparse instance feature should be representative enough for the future trajectory. In addition, our goal is to predict the future trajectory based on the observed trajectory conditioned on the sparse instance. Therefore, we use an LSTM [56] decoder to reconstruct the future trajectory with the combination of the past feature and the sparse instance feature, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.782, + 0.47, + 0.798 + ], + "angle": 0, + "content": "\\[\nY _ {r} = f _ {d} \\left(\\mathbf {F} _ {p} \\oplus \\mathbf {F} _ {i n s}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.471, + 0.87 + ], + "angle": 0, + "content": "where \\( f_{d}(\\cdot) \\) is the decoder implemented by an LSTM block, \\( \\oplus \\) is a concatenation operation, and \\( Y_{r} \\) is the reconstructed future trajectory. The trajectory \\( \\ell_{2} \\)-norm loss is used as the reconstruction loss, as:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.883, + 0.47, + 0.9 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {1} = \\left\\| Y - Y _ {r} \\right\\| _ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.783, + 0.107 + ], + "angle": 0, + "content": "where \\(||\\cdot ||_2\\) denotes the \\(\\ell_2\\)-norm distance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.107, + 0.892, + 0.333 + ], + "angle": 0, + "content": "Memory Architecture. The key idea of the memory module is to use the encoding of the observed trajectory as a memory key to retrieve possible sparse instances and jointly generate the multimodal predictions. To store the sparse instances generated during the reconstruction process, we first randomly initializes two memory banks, i.e., the past memory bank \\(\\mathbf{M}_p \\in \\mathbb{R}^{m \\times d}\\), and the instance memory bank \\(\\mathbf{M}_i \\in \\mathbb{R}^{m \\times d}\\) to store the past and the sparse instance features obtained from the reconstruction process, respectively. \\(m\\) denotes the storage size of the memory bank, and \\(d\\) is the dimension of the memory item. To combine the past and sparse instance features, the items in these two memory banks are aligned. Specifically, for the \\(j\\)-th past feature in \\(\\mathbf{M}_p[j]\\), there is a unique sparse instance feature in \\(\\mathbf{M}_i[j]\\) corresponding to it." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.895, + 0.62 + ], + "angle": 0, + "content": "Memory Update. Due to the abundant similar trajectories in the dataset, it brings high redundancy to store all of them in the memory bank. Hence we need to remove modality redundancy by memory updating process. Moreover, traditional memory mechanism [16] updates the memory bank with the stored items' labels, while such labels are not provided in trajectory prediction to distinguish trajectories with different modalities. Therefore, we design an unsupervised memory update strategy based on the distance between the stored and current items, as shown in Figure 5. Inspired by [16], an additional mark memory bank \\(\\mathbf{A} \\in \\mathbb{R}^{m \\times 1}\\) tracks the age of items stored in memory without being used. All items in \\(\\mathbf{A}\\) are initialized to 1 and aligned to their corresponding items in \\(\\mathbf{M}_p\\) and \\(\\mathbf{M}_i\\). Given a pair of inputs \\((\\mathbf{F}_p, \\mathbf{F}_{ins})\\), we first calculate the feature similarities between \\(\\mathbf{F}_p\\) and each item in \\(\\mathbf{M}_i\\). Then, the \\(k\\)-th \\((k \\in \\{1, \\dots, m\\})\\) item \\(\\mathbf{M}_p[k]\\) with the highest similarity is used to retrieve the corresponding sparse instance feature, i.e., \\(\\mathbf{M}_i[k]\\), through the one-to-one architecture of memory banks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.62, + 0.892, + 0.649 + ], + "angle": 0, + "content": "Specifically, the distance \\(\\mathcal{D}\\) between \\(\\mathbf{F}_{ins}\\) and \\(\\mathbf{M}_i[k]\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.663, + 0.892, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} = \\left\\| \\mathbf {M} _ {i} [ k ] - \\mathbf {F} _ {i n s} \\right\\| _ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.693, + 0.892, + 0.782 + ], + "angle": 0, + "content": "We compare \\(\\mathcal{D}\\) with a threshold \\(\\xi\\) to judge whether \\(F_{ins}\\) and \\(\\mathbf{M}_i[k]\\) have the same modality. When \\(\\mathcal{D} > \\xi\\), which means that \\(F_{ins}\\) and \\(\\mathbf{M}_i[k]\\) have different modalities. Hence we update \\(\\mathbf{M}_p\\), \\(\\mathbf{M}_i\\) with the oldest age (i.e., corresponding \\(\\mathbf{A}\\) with the maximum value) by \\(\\mathbf{F}_p\\), \\(\\mathbf{F}_{ins}\\), respectively, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.796, + 0.892, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {i} [ s ] \\leftarrow \\mathbf {F} _ {i n s}, \\quad \\mathbf {M} _ {p} [ s ] \\leftarrow \\mathbf {F} _ {p}, \\quad \\mathbf {A} [ s ] \\leftarrow 0, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(s\\) is the index with the oldest age. Since the \\(s\\)-th item in the past memory bank has been updated, the \\(s\\)-th item in \\(\\mathbf{A}\\) is set to 0 to prevent the newly added item from overwriting it. When \\(\\mathcal{D} \\leq \\xi\\), which means that \\(F_{ins}\\) and \\(\\mathbf{M}_i[k]\\) have the same modality. Hence we only update \\(\\mathbf{M}_p[k]\\) to approach" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9766" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.089, + 0.468, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.235, + 0.472, + 0.265 + ], + "angle": 0, + "content": "Figure 5. An illustration of the memory training process in SICNet. Red and purple lines denote the memory writing process." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.289, + 0.465, + 0.307 + ], + "angle": 0, + "content": "\\(\\mathbf{F}_p\\) by taking the normalized average of \\(\\mathbf{M}_p[k]\\) and \\(\\mathbf{F}_p\\), as:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.316, + 0.471, + 0.351 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {p} [ k ] \\leftarrow \\frac {\\mathbf {F} _ {p} + \\mathbf {M} _ {p} [ k ]}{\\| \\mathbf {F} _ {p} + \\mathbf {M} _ {p} [ k ] \\|}, \\quad \\mathbf {A} [ k ] \\leftarrow 0. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.361, + 0.473, + 0.455 + ], + "angle": 0, + "content": "Thus, we can retrieve the sparse instance feature \\(\\mathbf{M}_i[k]\\) easily, when we receive the past feature similar to \\(\\mathbf{F}_p\\). Since the \\(k\\)-th item in the past memory bank has been updated, the \\(k\\)-th item in \\(\\mathbf{A}\\) is set to 0 to prevent the newly added item from covering it. All values in \\(\\mathbf{A}\\) will add the same constant after an epoch." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.462, + 0.274, + 0.479 + ], + "angle": 0, + "content": "3.3. Memory Refinement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.47, + 0.621 + ], + "angle": 0, + "content": "The input observed trajectory encoding could be unfamiliar to the well-trained memory, which could be caused by: (1) modality loss due to removing redundancy during the memory update process; (2) the input modality does not exist in the train set. Such unfamiliarity could lead to the recalled sparse instance not matching the input trajectory. To reduce such recall errors, we design a memory refinement module with a refinement loss in the second training stage, as shown in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.622, + 0.47, + 0.713 + ], + "angle": 0, + "content": "We first consider the difference between the past feature \\(\\mathbf{F}_p\\) and its corresponding closest past memory item \\(\\mathbf{M}_p[k]\\) with the same modality. To avoid extra optimization similar to Eq. (8), we do not optimize \\(\\mathbf{F}_p\\) to approach \\(\\mathbf{M}_p[k]\\) again but encode the difference between \\(\\mathbf{F}_p\\) and \\(\\mathbf{M}_p[k]\\) as an additional deviation feature \\(\\mathbf{F}_v\\), which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.725, + 0.47, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {v} = \\phi \\left(\\mathbf {F} _ {p} \\oplus \\mathbf {M} _ {p} [ k ], \\theta_ {V}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.754, + 0.471, + 0.8 + ], + "angle": 0, + "content": "where \\(\\theta_V\\) is a learnable parameter, and \\(\\oplus\\) is the concatenation operation. Then, the retrieved sparse instance feature is refined as \\(\\mathbf{F}_r\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.812, + 0.47, + 0.829 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {r} = \\psi \\left(\\mathbf {F} _ {v} \\oplus \\mathbf {M} _ {i} [ k ], \\theta_ {R}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "where \\(\\psi (\\cdot ,\\cdot)\\) is a linear projection with a learnable parameter \\(\\theta_{R}\\), \\(\\mathbf{F}_r\\) is the refined sparse instance feature. In this case, the model can give a reasonable sparse instance considering the difference between the input observed feature and the stored" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.894, + 0.122 + ], + "angle": 0, + "content": "observed feature in the memory bank. The loss function of the memory refinement module is:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.134, + 0.894, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {2} = \\left\\| \\mathbf {F} _ {i n s} - \\mathbf {F} _ {r} \\right\\| _ {2}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.162, + 0.895, + 0.223 + ], + "angle": 0, + "content": "Finally, the past feature \\(\\mathbf{F}_{\\mathbf{p}}\\) is concatenated with the refinement feature \\(\\mathbf{F}_r\\) to predict the future trajectory. Concretely, we use the pre-trained decoder \\(f_{d}(\\cdot)\\) in the first stage to achieve the prediction, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.234, + 0.893, + 0.253 + ], + "angle": 0, + "content": "\\[\n\\hat {Y} = f _ {d} \\left(\\mathbf {F} _ {p} \\oplus \\mathbf {F} _ {r}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.264, + 0.779, + 0.28 + ], + "angle": 0, + "content": "where \\(\\hat{Y}\\) is the predicted future trajectory." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.28, + 0.895, + 0.432 + ], + "angle": 0, + "content": "Clustering. We adopt the clustering mechanism during inference to further remove the modality redundancy of the retrieved sparse instances. By retrieving \\( C \\) (\\( C >> k \\)) sparse instances corresponding to the past trajectory from the well-trained memory, we can generate \\( C \\) refined sparse instances through the refinement network. Then we cluster the \\( C \\) refined sparse instances into \\( K \\) multimodal proposals. The clustered multimodal proposals will guide the multimodal trajectory prediction instead of sparse instances in inference. Ablation studies verify the effectiveness of this post-process." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.439, + 0.754, + 0.456 + ], + "angle": 0, + "content": "3.4. Model Training & Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.894, + 0.583 + ], + "angle": 0, + "content": "Training. SICNet is trained by a two-stage strategy. In the first stage, we train the encoders \\( f_{p}(\\cdot) \\), \\( f_{y}(\\cdot) \\) and the decoder \\( f_{d}(\\cdot) \\) in the sparse instance representation module with the reconstruction loss \\( \\mathcal{L}_1 \\). Then we use the memory mechanism to store the sparse instances generated during the reconstruction process. In the second stage, we fix the parameters of \\( f_{p}(\\cdot) \\), \\( f_{y}(\\cdot) \\) and \\( f_{d}(\\cdot) \\), and train the memory refinement module with the refinement loss \\( \\mathcal{L}_2 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.751 + ], + "angle": 0, + "content": "Inference. We aim to predict \\( K \\) trajectories to cover the multimodality of future trajectories. During inference, we select the top- \\( C \\) past memory items by feature similarity and then retrieve corresponding \\( C \\) sparse instance items from the instance memory bank, where \\( C >> K \\) is to improve the prediction diversity. Then we obtained \\( C \\) refinement sparse instances through the memory refinement module. Subsequently, the \\( C \\) sparse instance items are clustered into \\( K \\) multimodal proposals. Finally, we feed the combination of the past feature and multimodal proposals into the decoder to generate \\( K \\) predicted future trajectories." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.763, + 0.769, + 0.78 + ], + "angle": 0, + "content": "4. Experiments and Discussions" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.702, + 0.804 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Evaluation Datasets. We evaluate our proposed method on two benchmark datasets, i.e., ETH-UCY [31, 21] and Stanford Drone Dataset (SDD) [33]. ETH-UCY is the most commonly used pedestrian trajectory prediction dataset captured in the bird's eye view. It contains five subsets, where the ETH [31] includes ETH and HOTEL subsets, and the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9767" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.894, + 0.303 + ], + "angle": 0, + "content": "
MethodVenue/YearInputETHHOTELUNIVZARA1ZARA2AVG
Social LSTM [1]CVPR2016P+S1.09/2.350.79/1.760.67/1.400.47/1.000.56/1.170.72/1.54
Social GAN [13]CVPR2018P+S0.87/1.620.67/1.370.76/1.520.35/0.680.42/0.840.61/1.21
SoPhie [34]CVPR2019P+S0.70/1.430.76/1.670.54/1.240.30/0.630.38/0.780.51/1.15
MANTRA [28]CVPR2020P+S0.48/0.880.17/0.330.37/0.810.27/0.580.30/0.670.32/0.65
PECNet [27]ECCV2020P+S0.54/0.870.18/0.240.35/0.600.22/0.390.17/0.300.29/0.48
SGCN [37]CVPR2021P+S0.63/1.030.32/0.550.37/0.700.29/0.530.25/0.450.37/0.65
AgentFormer [54]ICCV2021P+S0.45/0.750.14/0.220.25/0.450.18/0.300.14/0.240.23/0.39
PCCSNet [41]ICCV2021P0.28/0.540.11/0.190.29/0.600.21/0.440.15/0.340.21/0.42
CAGN [8]AAAI2022P+S0.41/0.650.13/0.230.32/0.540.21/0.380.16/0.330.25/0.43
SIT [36]AAAI2022P+S0.39/0.620.14/0.220.27/0.470.19/0.330.16/0.290.23/0.38
MemoNet [49]CVPR2022P+S0.40/0.610.11/0.170.24/0.430.18/0.320.14/0.240.21/0.35
Ours-P0.27/0.450.11/0.160.26/0.460.19/0.330.13/0.260.19/0.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.312, + 0.893, + 0.356 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-art methods on ETH-UCY in ADE/FDE. P and S indicate inputting the observed history points and social interaction information, respectively. All methods input the observed 8 time steps and output the predicted 12 time steps. Bold indicates the best performance. Underline indicates the second best performance. The Lower the better." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.378, + 0.465, + 0.568 + ], + "angle": 0, + "content": "
MethodVenue/YearInputADE/FDE
Social LSTM [1]CVPR2016P+S31.19/56.97
Social GAN [13]CVPR2018P+S27.23/41.44
SoPhie [34]CVPR2019P+S16.27/29.38
MANTRA [28]CVPR2020P+S8.96/17.76
PECNet [27]ECCV2020P+S9.96/15.88
PCCSNet [41]ICCV2021P8.62/16.16
CAGN [8]AAAI2022P+S9.42/15.93
SIT [36]AAAI2022P+S9.13/15.42
MemoNet [49]CVPR2022P+S8.56/12.66
Ours-P8.44/13.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.577, + 0.472, + 0.649 + ], + "angle": 0, + "content": "Table 2. Comparison with state-of-the-art methods on SDD in ADE/FDE. P and S indicate inputting the observed history points and social interaction information, respectively. All methods input the observed 8 time steps and output the predicted 12 time steps. The Lower the better." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.472, + 0.839 + ], + "angle": 0, + "content": "UCY [21] includes UNIV, ZARA1, and ZARA2 subsets. On ETH-UCY, we use the leave-out-one method for training on four subsets and testing on the other one following previous efforts [40, 54, 36]. SDD [33] is a large pedestrian trajectory prediction dataset also captured in bird's eye view. It contains 20 scenes of more than 10,000 trajectories collected from college campuses and is much larger than ETH-UCY. On SDD, we use a prior train-test split for training and testing according to previous methods [13, 27, 36]. We observe the historical trajectories of 8 time steps and predict the subsequent 12 time steps on both ETH-UCY and SDD." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Evaluation Metrics. Following previous methods [49, 36, 8, 41, 27], we employ two commonly used metrics, i.e., Average Displacement Error (ADE) and Final Displacement Error (FDE), to evaluate the trajectory prediction perfor" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.895, + 0.456 + ], + "angle": 0, + "content": "mance. ADE measures the average \\(\\ell_2\\)-norm distance between all points of the ground truth and predicted trajectory. FDE measures the \\(\\ell_2\\)-norm distance between the destination points of the ground truth and predicted trajectory. We use the best-of-20 metrics following previous methods [49, 41]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.895, + 0.593 + ], + "angle": 0, + "content": "Implementation Details. In our experiments, the two encoders are all 3-layer BiLSTM [56] architecture with hidden sizes of 48. The decoder is 3-layer BiLSTM [56] architecture with hidden size of 96. The dimensions of \\(\\theta_{M}\\), \\(\\theta_{V}\\) and \\(\\theta_{R}\\) are set to 48, 48 and 96, respectively. The parameter \\(\\mu\\) of mask threshold is 0.2. The parameter \\(\\xi\\) in memory training is 0.0001. The Adam optimizer is used to train our model by 100 epochs with a learning rate of 0.0005, decaying by 0.5 with an interval of 50." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.704, + 0.621 + ], + "angle": 0, + "content": "4.2. Quantitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.895, + 0.765 + ], + "angle": 0, + "content": "On ETH-UCY. As shown in Table 1, we compare our method with eleven state-of-the-art methods on the ETH-UCY dataset in past six years. Among them, MemoNet [49], PECNet [27], and CAGN [8] are goal-conditioned methods. MANTRA [28] and PCCSNet [41] are instance-conditioned methods. SoPhie [34], SGCN [37] and AgentFormer [54] generate multimodal trajectories by latent variables. The experiment results show that generating multimodal predictions by goals or instance is better than by latent variables." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Moreover, the results indicate that our method significantly outperforms all the competing methods on both ETH and UCY. Compared with the previous best method MemoNet [49], our method further improves the performance on ADE and FDE. Meanwhile, Compared with the other instance-conditioned methods, PCCSNet [41], and MANTRA [28], our method also achieve the best performance. The underlying reason could be that our method captures the effective intermediate process compared with" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9768" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.201 + ], + "angle": 0, + "content": "
MethodK=15K=10K=5
PECNet [27]10.66/17.8511.79/21.0014.49/28.08
PCCSNet [41]9.27/17.6210.36/20.1212.57/24.74
CAGN [8]10.01/17.4911.06/19.8913.85/26.87
MemoNet [49]9.62/15.3611.26/19.6414.43/27.45
Ours9.13/15.3110.27/18.1312.82/23.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.21, + 0.47, + 0.239 + ], + "angle": 0, + "content": "Table 3. Experiments on different best-of- \\( K \\) predictions on SDD in ADE/FDE. The lower the better." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.25, + 0.457, + 0.361 + ], + "angle": 0, + "content": "
MethodSIREFCLADE/FDE
Baseline0.46/0.94
(1)✓✓0.25/0.50
(2)✓✓0.27/0.56
(3)✓✓0.22/0.39
(4)✓✓✓0.19/0.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.369, + 0.472, + 0.398 + ], + "angle": 0, + "content": "Table 4. Ablation study of each component of our method on ETH-UCY dataset in ADE/FDE. The lower the better." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.423, + 0.47, + 0.452 + ], + "angle": 0, + "content": "the goal-conditioned methods and is more flexible compared with the instance-conditioned methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.454, + 0.472, + 0.531 + ], + "angle": 0, + "content": "To show the superiority of our method more clearly, we do not use extra information, such as social interaction and map information, as previous approaches. Nevertheless, our method still achieves the best performance, clearly demonstrating the superiority of our method using sparse instance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.473, + 0.713 + ], + "angle": 0, + "content": "On SDD. We compare our method with nine state-of-the-art methods on the SDD dataset in Table 2. For ADE, our method outperforms all the competing methods. For FDE, our method also achieves comparable performance with the best method MemoNet [49] and exceeds all other methods. The results manifest that the sparse instance representation can comprehensively describe the trajectory multimodality rather than goal-conditioned methods and thus achieves competitive performance. We speculate that our method underperforms MemoNet [49] on FDE because the sparse instance benefits more on the whole trajectory prediction (i.e., ADE) than goal prediction (i.e., FDE)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.233, + 0.743 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Different Best-of- \\(K\\) Predictions. Previous methods commonly use best-of- \\(K\\) (usually \\(K = 20\\)) as the quantified metric of multimodal trajectory prediction. To further validate the adaptability and effectiveness of our proposed sparse instance representation, we conduct an experiment on various best-of- \\(K\\) predictions with \\(K = 5, 10, 15\\) on SDD. We compare our method with the top three performance goal-conditioned methods, i.e., MemoNet [49], CAGN [8] and PECNet [27] and the best performance instance-conditioned method PCCSNet [41]. As shown in Table 3, our method" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.089, + 0.882, + 0.212 + ], + "angle": 0, + "content": "
μ00.10.20.30.40.5
ADE0.210.210.190.200.200.21
FDE0.340.330.330.330.340.37
μ0.60.70.80.91.0-
ADE0.250.280.260.240.21-
FDE0.450.540.480.430.38-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.221, + 0.892, + 0.248 + ], + "angle": 0, + "content": "Table 5. Ablation study of different mask threshold values on the ETH-UCY dataset. The lower the better." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.262, + 0.89, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.408, + 0.895, + 0.451 + ], + "angle": 0, + "content": "Figure 6. Experiments of the sparse instance analysis on the ETH-UCY dataset. We replace our proposed sparse instance with the goal and instance for the multimodal prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.895, + 0.568 + ], + "angle": 0, + "content": "achieves significant performance on all experimental settings. Interestingly, our method surpasses the previous best method MemoNet [49] when \\( K = 5, 10, 15 \\), despite underperforming MemoNet [49] when \\( K = 20 \\) in Table 2. It shows the adaptability of our method of using the sparse instance, which also performs significantly with a small \\( K \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.895, + 0.688 + ], + "angle": 0, + "content": "Contribution of Each Component. We divide our proposed SICNet into three components, i.e., the sparse instance representation (SI), the memory refinement module (REF), and the clustering post-process (CL). Ablation experiments of different combinations of the three components are evaluated on the ETH-UCY dataset. As shown in Table 4, all of the three components contribute to the performance improvement, which demonstrates their effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.825 + ], + "angle": 0, + "content": "Analysis of Mask Threshold. As shown in Table 5, we conduct experiments for different mask threshold values \\(\\mu\\) on the ETH-UCY dataset. It indicates that the performance will decrease when the mask threshold value \\(\\mu\\) is too small or too large. When \\(\\mu = 0\\), the mask is useless and the whole instance is reserved, thus causing the stochasticity loss. When \\(\\mu = 1\\), it denotes the whole instance is masked, thus causing no guidance information for the multimodal prediction. When \\(\\mu = 0.2\\), it achieves the best performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Analysis of Sparse Instance. As shown in Table 6, we replace the sparse instance in SICNet with goal-conditioned and instance-conditioned representations on the ETH-UCY dataset. Experiment results show that using the sparse instance achieves the best performance, which indicates the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9769" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.895, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.269, + 0.895, + 0.312 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison of our method with the goal-conditioned method PECNet [27] on the SDD dataset. Given the observed trajectories, we illustrate the ground truth paths and predicted trajectories by SICNet and the goal-conditioned method PECNet [27] for four different scenes. It shows our prediction is closer to the ground truth than PECNet [27], especially for the intermediate process." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.322, + 0.892, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.498, + 0.895, + 0.541 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison of our method with the instance-conditioned method PCCSNet [41] on the ETH-UCY dataset. Given the observed trajectories, we illustrate the ground truth paths and predicted trajectories by SICNet and the instance-conditioned method PCCSNet [41] for four different scenes. It shows our prediction is closer to the ground truth than PCCSNet [41]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.471, + 0.597 + ], + "angle": 0, + "content": "superiority of our method compared with goal-conditioned and instance-conditioned approaches." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.27, + 0.623 + ], + "angle": 0, + "content": "4.4. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.474, + 0.903 + ], + "angle": 0, + "content": "We compare the visualized trajectories with the well-accepted goal-conditioned method PECNet [27] on SDD in Figure 7, and the instance-conditioned method PCCSNet [41] on ETH-UCY in Figure 8. As shown in Figure 7, trajectories in Coupa show that our method achieves more precise trajectory predictions than PECNet. Trajectories in Little and Quad show that both PECNet and our method can predict destinations close to the ground truth, while our method can predict more precise intermediate trajectories due to our proposed sparse instance representation. Trajectories in Hyang show that our method can also give better predictions, although the ground truth trajectories are irregular. The results verify the superiority of our method compared with goal-conditioned methods. Moreover, as shown in Figure 8, the visualization results show that our prediction is closer to the ground truth than PCCSNet [41], which indicates the superiority of our SICNet compared with the instance-conditioned approaches." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.564, + 0.62, + 0.58 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.897, + 0.76 + ], + "angle": 0, + "content": "In this paper, we present a simple yet effective sparse-instance-conditioned network for pedestrian multimodal trajectory prediction, which leverages sparse instances to guide multimodal prediction. Moreover, we propose a memory refinement module to reduce recall errors when meeting unfamiliar trajectories during inference. Extensive experiments show that our method performs better than previous methods, and ablation studies demonstrate that our sparse-instance-conditioned method can generate multimodal predictions more accurately and adaptively compared with goal-conditioned and instance-conditioned approaches." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.782, + 0.661, + 0.8 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.896, + 0.9 + ], + "angle": 0, + "content": "This work was supported partly by National Key R&D Program of China under Grant 2021YFB1714700, NSFC under Grants 62088102 and 62106192, Natural Science Foundation of Shaanxi Province under Grant 2022JC-41, and Fundamental Research Funds for the Central Universities under Grant XTR042021005." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9770" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social LSTM: Human trajectory prediction in crowded spaces. In CVPR, pages 961–971, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.214 + ], + "angle": 0, + "content": "[2] Inhwan Bae, Jin-Hwi Park, and Hae-Gon Jeon. Nonprobability sampling network for stochastic human trajectory prediction. In CVPR, pages 6477-6487, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[3] Guangyi Chen, Junlong Li, Jiwen Lu, and Jie Zhou. Human trajectory prediction via counterfactual analysis. In ICCV, pages 9824-9833, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.472, + 0.299 + ], + "angle": 0, + "content": "[4] Xinlei Chen and Abhinav Gupta. Spatial memory for context reasoning in object detection. In ICCV, pages 4086-4096, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.472, + 0.343 + ], + "angle": 0, + "content": "[5] Chiho Choi, Joon Hee Choi, Jiachen Li, and Srikanth Malla. Shared cross-modal trajectory prediction for autonomous driving. In CVPR, pages 244-253, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.346, + 0.472, + 0.4 + ], + "angle": 0, + "content": "[6] Lingwei Dang, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. MSR-GCN: Multi-scale residual graph convolution networks for human motion prediction. In ICCV, pages 11447-11456, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.472, + 0.457 + ], + "angle": 0, + "content": "[7] Hanming Deng, Yang Hua, Tao Song, Zongpu Zhang, Zhengui Xue, Ruhui Ma, Neil Robertson, and Haibing Guan. Object guided external memory network for video object detection. In ICCV, pages 6677-6686, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.46, + 0.472, + 0.514 + ], + "angle": 0, + "content": "[8] Jinghai Duan, Le Wang, Chengjiang Long, Sanping Zhou, Fang Zheng, Liushuai Shi, and Gang Hua. Complementary attention gated network for pedestrian trajectory prediction. In AAAI, pages 542-550, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.472, + 0.558 + ], + "angle": 0, + "content": "[9] Zhihong Fu, Qingjie Liu, Zehua Fu, and Yunhong Wang. STMTrack: Template-free visual tracking with space-time memory networks. In CVPR, pages 13769-13778, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.56, + 0.472, + 0.614 + ], + "angle": 0, + "content": "[10] Harshayu Girase, Haiming Gang, Srikanth Malla, Jiachen Li, Akira Kanehara, Karttikeya Mangalam, and Chiho Choi. LOKI: Long term and key intentions for trajectory prediction. In ICCV, pages 9803-9812, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.472, + 0.657 + ], + "angle": 0, + "content": "[11] Dong Gong, Frederic Z. Zhang, Javen Qinfeng Shi, and Anton van den Hengel. Memory-augmented dynamic neural relational inference. In ICCV, pages 11843-11852, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.472, + 0.714 + ], + "angle": 0, + "content": "[12] Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory prediction via motion indeterminacy diffusion. In CVPR, pages 17113-17122, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.472, + 0.77 + ], + "angle": 0, + "content": "[13] Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi. Social GAN: socially acceptable trajectories with generative adversarial networks. In CVPR, pages 2255–2264, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.472, + 0.827 + ], + "angle": 0, + "content": "[14] Adam Houenou, Philippe Bonnifait, Véronique Cherfaoui, and Wen Yao. Vehicle trajectory prediction based on motion model and maneuver recognition. In IROS, pages 4363-4369, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[15] Boris Ivanovic and Marco Pavone. The trajectory: Probabilistic multi-agent trajectory modeling with dynamic spatiotemporal graphs. In ICCV, pages 2375-2384, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[16] Lukasz Kaiser, Ofir Nachum, Aurko Roy, and Samy Bengio. Learning to remember rare events. In ICLR, 2017." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "[17] Youngjin Kim, Minjung Kim, and Gunhee Kim. Memorization precedes generation: Learning unsupervised gans with memory networks. In ICLR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[18] Sumith Kulal, Jiayuan Mao, Alex Aiken, and Jiajun Wu. Programmatic concept learning for human motion description and synthesis. In CVPR, pages 13843-13852, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[19] Mihee Lee, Samuel S. Sohn, Seonghyeon Moon, Sejong Yoon, Mubbasir Kapadia, and Vladimir Pavlovic. MUSE-VAE: Multi-scale vae for environment-aware long term trajectory prediction. In CVPR, pages 2221–2230, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.894, + 0.273 + ], + "angle": 0, + "content": "[20] Sangho Lee, Jinyoung Sung, Youngjae Yu, and Gunhee Kim. A memory network approach for story-based temporal summarization of \\(360^{\\circ}\\) videos. In CVPR, pages 1410-1419, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.275, + 0.894, + 0.314 + ], + "angle": 0, + "content": "[21] Alon Lerner, Yiorgos Chrysanthou, and Dani Lischinski. Crowds by example. Comput. Graph. Forum, 26(3):655-664, 2007." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.894, + 0.357 + ], + "angle": 0, + "content": "[22] Peizhao Li, Pu Wang, Karl Berntorp, and Hongfu Liu. Exploiting temporal relations on radar perception for autonomous driving. In CVPR, pages 17071-17080, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.894, + 0.399 + ], + "angle": 0, + "content": "[23] Shijie Li, Yanying Zhou, Jinhui Yi, and Juergen Gall. Spatial-temporal consistency network for low-latency trajectory forecasting. In ICCV, pages 1940-1949, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.401, + 0.894, + 0.455 + ], + "angle": 0, + "content": "[24] Jialun Liu, Yifan Sun, Feng Zhu, Hongbin Pei, Yi Yang, and Wenhui Li. Learning memory-augmented unidirectional metrics for cross-modality person re-identification. In CVPR, pages 19366-19375, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.457, + 0.894, + 0.511 + ], + "angle": 0, + "content": "[25] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In ICCV, pages 13588-13597, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.894, + 0.565 + ], + "angle": 0, + "content": "[26] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In ICCV, pages 15233-15242, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.894, + 0.635 + ], + "angle": 0, + "content": "[27] Karttikeya Mangalam, Harshayu Girase, Shreyas Agarwal, Kuan-Hui Lee, Ehsan Adeli, Jitendra Malik, and Adrien Gaidon. It is not the journey but the destination: Endpoint conditioned trajectory prediction. In ECCV, pages 759-776, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.894, + 0.691 + ], + "angle": 0, + "content": "[28] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. *Mantra: Memory augmented networks for multiple trajectory prediction*. In *CVPR*, pages 7143–7152, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.894, + 0.748 + ], + "angle": 0, + "content": "[29] Abduallah Mohamed, Kun Qian, Mohamed Elhoseiny, and Christian Claudel. Social-STGCNN: A social spatio-temporal graph convolutional neural network for human trajectory prediction. In CVPR, pages 14424-14432, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.894, + 0.803 + ], + "angle": 0, + "content": "[30] Alessio Monti, Angelo Porrello, Simone Calderara, Pasquale Coscia, Lamberto Ballan, and Rita Cucchiara. How many observations are enough? knowledge distillation for trajectory forecasting. In CVPR, pages 6553-6562, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.805, + 0.894, + 0.845 + ], + "angle": 0, + "content": "[31] Stefano Pellegrini, Andreas Ess, Konrad Schindler, and Luc Van Gool. You'll never walk alone: Modeling social behavior for multi-target tracking. In ICCV, pages 261-268, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.847, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[32] Zheng Qin, Sanping Zhou, Le Wang, Jinghai Duan, Gang Hua, and Wei Tang. Motiontrack: Learning robust short-term and long-term motions for multi-object tracking. In CVPR, pages 17939-17948, 2023." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9771" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[33] Alexandre Robicquet, Amir Sadeghian, Alexandre Alahi, and Silvio Savarese. Learning Social Etiquette: Human trajectory understanding in crowded scenes. In ECCV, pages 549-565, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.472, + 0.207 + ], + "angle": 0, + "content": "[34] Amir Sadeghian, Vineet Kosaraju, Ali Sadeghian, Noriaki Hirose, Hamid Rezatofighi, and Silvio Savarese. SoPhie: An attentive gan for predicting paths compliant to social and physical constraints. In CVPR, pages 1349-1358, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.209, + 0.472, + 0.263 + ], + "angle": 0, + "content": "[35] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory: Dynamically-feasible trajectory forecasting with heterogeneous data. In ECCV, pages 683-700, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.266, + 0.472, + 0.322 + ], + "angle": 0, + "content": "[36] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Fang Zheng, Nanning Zheng, and Gang Hua. Social interpretable tree for pedestrian trajectory prediction. In AAAI, pages 2235-2243, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.324, + 0.47, + 0.379 + ], + "angle": 0, + "content": "[37] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Mo Zhou, Zhenxing Niu, and Gang Hua. SGCN: Sparse graph convolution network for pedestrian trajectory prediction. In CVPR, pages 8990-8999, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.381, + 0.472, + 0.436 + ], + "angle": 0, + "content": "[38] Aravind Srinivas, Tsung-Yi Lin, Niki Parmar, Jonathon Shlens, Pieter Abbeel, and Ashish Vaswani. Bottleneck transformers for visual recognition. In CVPR, pages 16519-16529, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.439, + 0.47, + 0.481 + ], + "angle": 0, + "content": "[39] Hao Sun, Zhiqun Zhao, and Zhihai He. Reciprocal learning networks for human trajectory prediction. In CVPR, pages 7416-7425, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.47, + 0.525 + ], + "angle": 0, + "content": "[40] Jianhua Sun, Qinhong Jiang, and Cewu Lu. Recursive social behavior graph for trajectory prediction. In CVPR, pages 660-669, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.527, + 0.472, + 0.58 + ], + "angle": 0, + "content": "[41] Jianhua Sun, Yuxuan Li, Hao-Shu Fang, and Cewu Lu. Three steps to multimodal trajectory prediction: Modality clustering, classification and synthesis. In ICCV, pages 13250-13259, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.584, + 0.472, + 0.655 + ], + "angle": 0, + "content": "[42] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, Wenyu Liu, and Bin Xiao. Deep high-resolution representation learning for visual recognition. IEEE T-PAMI, 43(10):3349-3364, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.656, + 0.472, + 0.698 + ], + "angle": 0, + "content": "[43] Jingke Wang, Tengju Ye, Ziqing Gu, and Junbo Chen. LTP: Lane-based trajectory prediction for autonomous driving. In CVPR, pages 17134-17142, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.47, + 0.742 + ], + "angle": 0, + "content": "[44] Jürgen Wiest, Matthias Höffken, Ulrich Kreßel, and Klaus Dietmayer. Probabilistic trajectory prediction with gaussian mixture models. In IV, pages 141-146, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.47, + 0.798 + ], + "angle": 0, + "content": "[45] Yuxuan Wu, Le Wang, Sanping Zhou, Jinghai Duan, Gang Hua, and Wei Tang. Multi-stream representation learning for pedestrian trajectory prediction. In AAAI, pages 2875-2882, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[46] Zifeng Wu, Chunhua Shen, and Anton Van Den Hengel. Wider or deeper: Revisiting the resnet model for visual recognition. PR, 90:119-133, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[47] Haozhe Xie, Hongxun Yao, Shangchen Zhou, Shengping Zhang, and Wenxiu Sun. Efficient regional memory network for video object segmentation. In CVPR, pages 1286-1295, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.135 + ], + "angle": 0, + "content": "[48] Bo Xu, Han Huang, Cheng Lu, Ziwen Li, and Yandong Guo. Virtual multi-modality self-supervised foreground matting for human-object interaction. In ICCV, pages 428-437, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[49] Chenxin Xu, Weibo Mao, Wenjun Zhang, and Siheng Chen. Remember intentions: Retrospective-memory-based trajectory prediction. In CVPR, pages 6488-6497, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.894, + 0.22 + ], + "angle": 0, + "content": "[50] Hao Xue, Du Q. Huynh, and Mark Reynolds. SS-LSTM: A hierarchical LSTM model for pedestrian trajectory prediction. In WACV, pages 1186-1194, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.222, + 0.894, + 0.275 + ], + "angle": 0, + "content": "[51] Gang Yang, Man Zhou, Keyu Yan, Aiping Liu, Xueyang Fu, and Fan Wang. Memory-augmented deep conditional unfolding network for pan-sharpening. In CVPR, pages 1788-1797, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.894, + 0.318 + ], + "angle": 0, + "content": "[52] Tianyu Yang and Antoni B. Chan. Visual tracking via dynamic memory networks. IEEE T-PAMI, 43(1):360-374, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.894, + 0.362 + ], + "angle": 0, + "content": "[53] Jiyang Yu, Jingen Liu, Liefeng Bo, and Tao Mei. Memory-augmented non-local attention for video super-resolution. In CVPR, pages 17834-17843, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.364, + 0.894, + 0.405 + ], + "angle": 0, + "content": "[54] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris Kitani. Agent-Former: Agent-aware transformers for socio-temporal multiagent forecasting. In ICCV, pages 9793–9803, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.407, + 0.894, + 0.448 + ], + "angle": 0, + "content": "[55] He Zhao and Richard P. Wildes. Where are you heading? dynamic trajectory prediction with expert goal examples. In ICCV, pages 7609-7618, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.894, + 0.504 + ], + "angle": 0, + "content": "[56] Zheng Zhao, Weihai Chen, Xingming Wu, Peter CY Chen, and Jingmeng Liu. LSTM network: a deep learning approach for short-term traffic forecast. IET Intell. Transp. Syst., 11(2):68-75, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.894, + 0.56 + ], + "angle": 0, + "content": "[57] Desen Zhou, Zhichao Liu, Jian Wang, Leshan Wang, Tao Hu, Errui Ding, and Jingdong Wang. Human-object interaction detection via disentangled transformer. In CVPR, pages 19568-19577, 2022." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9772" + } + ] +] \ No newline at end of file diff --git a/2023/Sparse Instance Conditioned Multimodal Trajectory Prediction/5a87e2e4-dac0-41f3-9966-85985369c6c6_origin.pdf b/2023/Sparse Instance Conditioned Multimodal Trajectory Prediction/5a87e2e4-dac0-41f3-9966-85985369c6c6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..356798b67733897428018b0b2958250b149f7e47 --- /dev/null +++ b/2023/Sparse Instance Conditioned Multimodal Trajectory Prediction/5a87e2e4-dac0-41f3-9966-85985369c6c6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3510f3206ad7ac4e7807b080578efee819976bf42837c554df8c5d71359c9bbd +size 520543 diff --git a/2023/Sparse Instance Conditioned Multimodal Trajectory Prediction/full.md b/2023/Sparse Instance Conditioned Multimodal Trajectory Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dcd9758585d23787129d1482623ae23bf4e9c6ed --- /dev/null +++ b/2023/Sparse Instance Conditioned Multimodal Trajectory Prediction/full.md @@ -0,0 +1,323 @@ +# Sparse Instance Conditioned Multimodal Trajectory Prediction + +Yonghao Dong $^{1}$ Le Wang $^{1*}$ Sanping Zhou $^{1}$ Gang Hua $^{2}$ $^{1}$ National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University + $^{2}$ Wormpex AI Research + +# Abstract + +Pedestrian trajectory prediction is critical in many vision tasks but challenging due to the multimodality of the future trajectory. Most existing methods predict multimodal trajectories conditioned by goals (future endpoints) or instances (all future points). However, goal-conditioned methods ignore the intermediate process and instance-conditioned methods ignore the stochasticity of pedestrian motions. In this paper, we propose a simple yet effective Sparse Instance Conditioned Network (SICNet), which gives a balanced solution between goal-conditioned and instance-conditioned methods. Specifically, SICNet learns comprehensive sparse instances, i.e., representative points of the future trajectory, through a mask generated by a long short-term memory encoder and uses the memory mechanism to store and retrieve such sparse instances. Hence SICNet can decode the observed trajectory into the future prediction conditioned on the stored sparse instance. Moreover, we design a memory refinement module that refines the retrieved sparse instances from the memory to reduce memory recall errors. Extensive experiments on ETH-UCY and SDD datasets show that our method outperforms existing state-of-the-art methods. In addition, ablation studies demonstrate the superiority of our method compared with goal-conditioned and instance-conditioned approaches. + +# 1. Introduction + +Pedestrian trajectory prediction aims to predict pedestrians' future paths given their observed trajectories. It plays an important role in autonomous driving [22, 43, 5], human motion prediction [6, 12, 19], video surveillance [57, 48, 18, 32], and visual recognition [38, 42, 46]. Despite the recent advancement [30, 49, 8, 36, 10], pedestrian trajectory prediction is still challenging due to the multimodality of future trajectories. As the pedestrian's motions are stochastic and + +![](images/9834723deedee81acb7629c918849e3d01939dfaf75a99252920265d774667d7.jpg) + +![](images/b3ca369f619b2a0ce5c7a819e86e2d306f7f93b4e53e00beb0c6bf8133c923a9.jpg) + +![](images/17c4f188ba87645800b3b44f78a1d83ce123d79c50c449ab186f87822314067a.jpg) + +![](images/5785870e0f39c99ece345d3ce4ddc16782130a70c1e0e6c583caad5bdbe02b76.jpg) + +![](images/af08258e5f443f5d0219df92a8af7aea13bb1ae842c81123e9122deb83e0886e.jpg) +Figure 1. Illustration of different multimodal prediction methods. (a) Goals, instances and sparse-instances prepared for the prediction. (b) Goal-conditioned methods. (c) Instance-conditioned methods. (d) Sparse-instance-conditioned methods. The sparse-instance-conditioned method is more accurate in in-between states than the goal-conditioned method and more flexible than the instance-conditioned method, thus making the better prediction. + +indeterminate [29], multiple future paths are possible given their current states. + +To handle multimodality, many methods [27, 49, 55] transform the multimodal trajectories into multimodal goals. They first predict various goals as modality representations and then sample multiple goals to generate diverse future trajectories. However, they only focus on goals while neglecting the intermediate movement process. As illustrated in Figure 1(b), although with accurate goals, the intermediate process could be far from the ground truth. Thus, the multimodal goals are not equal to the multimodal future trajectories. It is necessary to consider the in-between states to comprehensively describe the multimodality of trajectory. Moreover, some other work [27, 28] predicts multimodal trajectories conditioned by instances. However, the instance-conditioned methods will lead to motion stochasticity loss and modality redundancy, resulting in performance degradation and computational resource waste. As shown in Figure 1(c), due to the reduced randomness of the instance-conditioned approach, the prediction retains the same bias + +when the instance deviates from the ground truth. + +To close this gap, we present a Sparse Instance Conditioned Network (SICNet). In our model, a sparse instance, which can give a comprehensive and adaptive description of future modality, is constructed by element-wise multiplication between the future trajectory and a learned mask. Hence, SICNet can predict the future trajectory more precisely guided by the sparse instance. To obtain the related sparse instances during inference, we adopt the memory mechanism to store non-redundant sparse instances corresponding to all observed trajectories. Therefore, SICNet can recall its modality's most related sparse instance and jointly generate the prediction during inference. Moreover, when the memory meets an unfamiliar observed trajectory, i.e., a trajectory quite different from stored items, the memory mechanism may recall false experiences. Hence we propose a memory refinement module with a refinement loss to reduce such recall errors by bridging the gap between the recall and the ground truth. For multimodal trajectories generation, we can retrieve a certain number of the most likely sparse instances from the external memory and then generate corresponding multimodal trajectories. In addition, to further remove the redundancy from the retrieved sparse instances, we use the cluster mechanism as a post-process to obtain succinct multimodal proposals for the prediction instead of the sparse instance during inference. + +Extensive experiments and ablation studies on two popular pedestrian trajectory prediction datasets, i.e., ETH-UCY [31, 21] and Stanford Drone Dataset (SDD) [33], demonstrate the superiority of our method over prior state-of-the-art methods. In summary, our contributions are four-fold: (1) We give a balanced solution, i.e., a novel sparse instance, between the goal and instance to guide the multimodal trajectory prediction. (2) We propose a simple yet effective sparse instance conditioned network for the multimodal trajectory prediction. (3) We propose a memory refinement module for the external memory to reduce recall errors when meeting unfamiliar observed trajectories. (4) Extensive experiments and ablation studies demonstrate the superiority and flexibility of our method compared with the goal-conditioned and instance-conditioned approaches. + +# 2. Related Works + +Multimodal Pedestrian Trajectory Prediction. Pedestrian trajectory prediction [2, 23, 3, 39, 45] aims to predict pedestrians' future trajectories based on their observed trajectories. There are some early studies on the deterministic trajectory prediction [1, 44, 14, 50]. Due to the strong randomness and uncertainty of pedestrian motions, there is no single correct future trajectory prediction. Motivated by this, Social GAN [13] proposes the concept of multimodal pedestrian trajectory prediction and emphasizes its importance. + +Earlier work generates multimodal trajectories through + +latent variables. Social GAN [13] proposes a Generative Adversarial Network based on a social pooling mechanism to generate a multimodal trajectory distribution. SGCN [37] generates multimodal trajectories by modeling the future predictions as a bi-variate Gaussian distribution. The CVAE-based methods [15, 35, 8] generate multimodal trajectories by sampling latent variables from a learned latent space. However, the latent variable conditioned method is persecuted by the problem of model collapse, thus resulting in performance degradation. Moreover, the latent variable conditioned methods also suffer in interpretability. + +To address the model collapse and interpretability issues, PECNet [27] proposes a goal-conditioned approach, which transforms multimodal trajectories into multimodal goals and then samples multiple goals to generate corresponding diverse predictions. After that, amounts of goal-conditioned studies [55, 8, 49] emerge and show superior performance. However, goal-conditioned methods represent different modalities with different possible goals while neglecting the intermediate states, thus suffering in performance improvement. In addition, PCCSNet[27] and MANTRA [28] propose instance-conditioned approaches, which use all points of the future instance to describe each modality precisely. However, it could lead to stochasticity loss and modality redundancy, resulting in performance degradation and computational resource waste. Moreover, Y-net [26] use goals and waypoints to make the prediction. However, the waypoints' number and time step locations are manually determined in Y-net [26], which may decrease the prediction accuracy by manually choosing errors. + +In contrast, we give a balanced way, i.e., a novel sparse instance, between the goal and instance to guide the multimodal trajectory prediction. The sparse instance is constructed by the learned representative points of the instance, which can improve modality representation accuracy compared with goal-conditioned methods and preserve stochasticity compared with instance-conditioned methods. + +Memory Network. The memory-augmented network [51, 11, 25, 24, 53] uses an external memory module to store critical information and has been widely applied in many areas like detection [7, 4], tracking [52, 9], segmentation [47], image generation [17], and video summarization [20]. Memory GAN [17] proposed a novel end-to-end unsupervised GAN network with a learnable memory network that effectively learns a highly multimodal latent space without suffering from structural discontinuity and forgetting problems. OGEMN [7] proposes the first object-guided external memory network for the online video object detection. STMTrack [52] proposes a novel tracking framework built on a space-time memory network, which can fully use historical information related to the target for better adapting to appearance variations during tracking. RMNet [47] memorizes the target object regions, effectively alleviating + +![](images/b8c818bfbe20c19976c1a850a0c79282ac3897387b75bebbae45b9b12fbbbb25.jpg) +Figure 2. The framework of our proposed SICNet. SICNet adopts a two-stage training strategy. The sparse instance representation module generates the sparse instances by a learned mask and then trains an external memory to store and retrieve them. The memory refinement module refines the retrieved sparse instances from memory to reduce the searching errors. We cluster the refined sparse instances into multimodal proposals to remove redundancy further. The cluster only exists in inference to save training costs. + +objects' ambiguity and reducing computational complexity. + +Two previous memory-based work, MANTRA [28] and MemoNet [49], also use the memory mechanism in multimodal trajectory prediction. However, the difference are mainly in three aspects: (1) The contribution of our proposed SICNet focuses on a novel representation of the trajectory modality, i.e., a learned sparse instance, while MANTRA [28] uses the whole instance and MemoNet [49] uses the goal to represent modalities of multimodal trajectories; (2) We propose a memory refinement module to reduce recall errors when meeting unfamiliar observed trajectories during inference, while MANTRA [28] and MemoNet [49] ignore such circumstances. (3) To demonstrate the superiority of our proposed sparse instance, we do not use extra social interaction and map information, while MANTRA [28] use extra map information and MemoNet [49] use social interaction information. Despite this, we also achieve the best performance compared with the above two methods, which indicates the superiority of our method. + +# 3. Our Method + +As aforementioned, traditional goal-conditioned methods ignore the intermediate process and instance-conditioned methods ignore the stochasticity of pedestrian motions, which leads to performance degradation. In this section, we present a sparse-instance-conditioned network (SICNet) to close this gap. Our SICNet adopts a two-stage training strategy. In the first stage, we learn the sparse instance, which is used to reconstruct the future trajectory jointly with the observed trajectory. Meanwhile, an external memory is trained to store and retrieve the sparse instance. In the second stage, we refine the retrieved sparse instances from the external memory through a refinement module with a refinement loss to reduce the recall errors. In addition, we use the cluster mechanism to remove redundancy of the retrieved sparse instances further during inference. + +# 3.1. Problem Formulation + +The objective of pedestrian trajectory prediction is to predict possible future trajectory coordinates based on the observed trajectories. For each target pedestrian, given an observed trajectory $X = \{(x_{t},y_{t})\}_{t = 1}^{T_{obs}}$ , where $(x_{t},y_{t})$ is the 2D coordinate at time $t$ , we aim to predict the future trajectory coordinates $Y = \{(x_{t},y_{t})\}_{t = T_{obs} + 1}^{T_{pred}}$ . Note that multimodal pedestrian trajectory prediction requires the model to predict $K$ future trajectories $\{Y_i\}_{i = 1}^K$ to account for multimodality, while only one future trajectory (ground truth) is provided for training in the dataset. + +# 3.2. Sparse Instance Representation + +In the first training stage, we generate the sparse instance features to represent future modalities and store them by memory mechanism. As shown in Figure 3, we first train the encoder-decoder with a reconstruction loss in this stage. Then we train the memory module to store the sparse instances generated during reconstruction. + +Feature Encoder. As the trajectory is a temporal sequence, we use two LSTM [56] blocks to build the past encoder $f_{p}(\cdot)$ and the future encoder $f_{y}(\cdot)$ , respectively. The hidden states of the two encoders are used as their corresponding features, as follows: + +$$ +\begin{array}{l} \mathbf {F} _ {p} = f _ {p} (X), \\ \mathbf {F} _ {y} = f _ {y} (Y), \tag {1} \\ \end{array} +$$ + +where $\mathbf{F}_p\in \mathbb{R}^{1\times d}$ is the past feature extracted from the observed trajectory $X$ , and $\mathbf{F}_y\in \mathbb{R}^{1\times d}$ is the future feature extracted from the future trajectory $Y$ . $d$ is the dimension of hidden state in LSTM. + +Sparse Instance. When inputting an observed trajectory, we want to obtain the corresponding sparse instance of its future modality so that we can construct future trajectories jointly using the observed trajectory and the sparse instance. + +![](images/09536bbf7de6c330f9f998b4fd8a62fb6f6238b120d410df38d96efc40496b60.jpg) +Figure 3. The first stage of the training process. In this stage, we train the sparse instance by reconstructing the future trajectory. Then we train an external memory to store and retrieve the learned sparse instances. + +![](images/501526ea587d50b87d4759eb5aa21ac6e2d4da47701750d0f9974eed9fa328c8.jpg) +Figure 4. The second stage of the training process. In this stage, we refine the retrieved sparse instances from memory through the memory refinement module to reduce the recall errors. + +Hence, we first learn an instance mask from the past feature $\mathbf{F}_p$ through an MLP layer, as follows: + +$$ +\begin{array}{l} \mathbf {F} _ {m} = \phi (\mathbf {F} _ {p}, \theta_ {M}), \\ \begin{array}{l} M = \mathbb {I} (\delta (\mathbf {F} _ {m}) \leq \mu)), \end{array} \tag {2} \\ \end{array} +$$ + +where $\phi (\cdot ,\cdot)$ denotes the MLP layer. $\theta_{M}$ is a learnable parameter. $\mathbf{F}_m$ is the learned feature which is used to construct the mask. $M$ denotes the instance mask. $\delta (\cdot)$ is sigmoid function to map $\mathbf{F}_m$ to the mask value [0, 1]. $\mathbb{I}$ is an indicator function which equals 0 if the inequality holds, otherwise 1. $\mu$ is the mask threshold. Compared with directly masking instances, masking instances in the high-level feature dimension is more efficient and effective. Hence we generate the sparse instance $\mathbf{F}_{ins}$ in the feature level dimension, as: + +$$ +\mathbf {F} _ {i n s} = \mathbf {F} _ {y} \odot M, \tag {3} +$$ + +where $\odot$ denotes the element-wise multiplication. + +Feature Decoder. The sparse instance feature should be representative enough for the future trajectory. In addition, our goal is to predict the future trajectory based on the observed trajectory conditioned on the sparse instance. Therefore, we use an LSTM [56] decoder to reconstruct the future trajectory with the combination of the past feature and the sparse instance feature, as follows: + +$$ +Y _ {r} = f _ {d} \left(\mathbf {F} _ {p} \oplus \mathbf {F} _ {i n s}\right), \tag {4} +$$ + +where $f_{d}(\cdot)$ is the decoder implemented by an LSTM block, $\oplus$ is a concatenation operation, and $Y_{r}$ is the reconstructed future trajectory. The trajectory $\ell_{2}$ -norm loss is used as the reconstruction loss, as: + +$$ +\mathcal {L} _ {1} = \left\| Y - Y _ {r} \right\| _ {2}, \tag {5} +$$ + +where $||\cdot ||_2$ denotes the $\ell_2$ -norm distance. + +Memory Architecture. The key idea of the memory module is to use the encoding of the observed trajectory as a memory key to retrieve possible sparse instances and jointly generate the multimodal predictions. To store the sparse instances generated during the reconstruction process, we first randomly initializes two memory banks, i.e., the past memory bank $\mathbf{M}_p \in \mathbb{R}^{m \times d}$ , and the instance memory bank $\mathbf{M}_i \in \mathbb{R}^{m \times d}$ to store the past and the sparse instance features obtained from the reconstruction process, respectively. $m$ denotes the storage size of the memory bank, and $d$ is the dimension of the memory item. To combine the past and sparse instance features, the items in these two memory banks are aligned. Specifically, for the $j$ -th past feature in $\mathbf{M}_p[j]$ , there is a unique sparse instance feature in $\mathbf{M}_i[j]$ corresponding to it. + +Memory Update. Due to the abundant similar trajectories in the dataset, it brings high redundancy to store all of them in the memory bank. Hence we need to remove modality redundancy by memory updating process. Moreover, traditional memory mechanism [16] updates the memory bank with the stored items' labels, while such labels are not provided in trajectory prediction to distinguish trajectories with different modalities. Therefore, we design an unsupervised memory update strategy based on the distance between the stored and current items, as shown in Figure 5. Inspired by [16], an additional mark memory bank $\mathbf{A} \in \mathbb{R}^{m \times 1}$ tracks the age of items stored in memory without being used. All items in $\mathbf{A}$ are initialized to 1 and aligned to their corresponding items in $\mathbf{M}_p$ and $\mathbf{M}_i$ . Given a pair of inputs $(\mathbf{F}_p, \mathbf{F}_{ins})$ , we first calculate the feature similarities between $\mathbf{F}_p$ and each item in $\mathbf{M}_i$ . Then, the $k$ -th $(k \in \{1, \dots, m\})$ item $\mathbf{M}_p[k]$ with the highest similarity is used to retrieve the corresponding sparse instance feature, i.e., $\mathbf{M}_i[k]$ , through the one-to-one architecture of memory banks. + +Specifically, the distance $\mathcal{D}$ between $\mathbf{F}_{ins}$ and $\mathbf{M}_i[k]$ is defined as: + +$$ +\mathcal {D} = \left\| \mathbf {M} _ {i} [ k ] - \mathbf {F} _ {i n s} \right\| _ {2}. \tag {6} +$$ + +We compare $\mathcal{D}$ with a threshold $\xi$ to judge whether $F_{ins}$ and $\mathbf{M}_i[k]$ have the same modality. When $\mathcal{D} > \xi$ , which means that $F_{ins}$ and $\mathbf{M}_i[k]$ have different modalities. Hence we update $\mathbf{M}_p$ , $\mathbf{M}_i$ with the oldest age (i.e., corresponding $\mathbf{A}$ with the maximum value) by $\mathbf{F}_p$ , $\mathbf{F}_{ins}$ , respectively, as follows: + +$$ +\mathbf {M} _ {i} [ s ] \leftarrow \mathbf {F} _ {i n s}, \quad \mathbf {M} _ {p} [ s ] \leftarrow \mathbf {F} _ {p}, \quad \mathbf {A} [ s ] \leftarrow 0, \tag {7} +$$ + +where $s$ is the index with the oldest age. Since the $s$ -th item in the past memory bank has been updated, the $s$ -th item in $\mathbf{A}$ is set to 0 to prevent the newly added item from overwriting it. When $\mathcal{D} \leq \xi$ , which means that $F_{ins}$ and $\mathbf{M}_i[k]$ have the same modality. Hence we only update $\mathbf{M}_p[k]$ to approach + +![](images/460dfc9a0acd3bd37e6e9b7b9afbb113be3dbbe29d38c070e4db4d9db6aa5b2d.jpg) +Figure 5. An illustration of the memory training process in SICNet. Red and purple lines denote the memory writing process. + +$\mathbf{F}_p$ by taking the normalized average of $\mathbf{M}_p[k]$ and $\mathbf{F}_p$ , as: + +$$ +\mathbf {M} _ {p} [ k ] \leftarrow \frac {\mathbf {F} _ {p} + \mathbf {M} _ {p} [ k ]}{\| \mathbf {F} _ {p} + \mathbf {M} _ {p} [ k ] \|}, \quad \mathbf {A} [ k ] \leftarrow 0. \tag {8} +$$ + +Thus, we can retrieve the sparse instance feature $\mathbf{M}_i[k]$ easily, when we receive the past feature similar to $\mathbf{F}_p$ . Since the $k$ -th item in the past memory bank has been updated, the $k$ -th item in $\mathbf{A}$ is set to 0 to prevent the newly added item from covering it. All values in $\mathbf{A}$ will add the same constant after an epoch. + +# 3.3. Memory Refinement + +The input observed trajectory encoding could be unfamiliar to the well-trained memory, which could be caused by: (1) modality loss due to removing redundancy during the memory update process; (2) the input modality does not exist in the train set. Such unfamiliarity could lead to the recalled sparse instance not matching the input trajectory. To reduce such recall errors, we design a memory refinement module with a refinement loss in the second training stage, as shown in Figure 4. + +We first consider the difference between the past feature $\mathbf{F}_p$ and its corresponding closest past memory item $\mathbf{M}_p[k]$ with the same modality. To avoid extra optimization similar to Eq. (8), we do not optimize $\mathbf{F}_p$ to approach $\mathbf{M}_p[k]$ again but encode the difference between $\mathbf{F}_p$ and $\mathbf{M}_p[k]$ as an additional deviation feature $\mathbf{F}_v$ , which is formulated as: + +$$ +\mathbf {F} _ {v} = \phi \left(\mathbf {F} _ {p} \oplus \mathbf {M} _ {p} [ k ], \theta_ {V}\right), \tag {9} +$$ + +where $\theta_V$ is a learnable parameter, and $\oplus$ is the concatenation operation. Then, the retrieved sparse instance feature is refined as $\mathbf{F}_r$ , as follows: + +$$ +\mathbf {F} _ {r} = \psi \left(\mathbf {F} _ {v} \oplus \mathbf {M} _ {i} [ k ], \theta_ {R}\right), \tag {10} +$$ + +where $\psi (\cdot ,\cdot)$ is a linear projection with a learnable parameter $\theta_{R}$ , $\mathbf{F}_r$ is the refined sparse instance feature. In this case, the model can give a reasonable sparse instance considering the difference between the input observed feature and the stored + +observed feature in the memory bank. The loss function of the memory refinement module is: + +$$ +\mathcal {L} _ {2} = \left\| \mathbf {F} _ {i n s} - \mathbf {F} _ {r} \right\| _ {2}. \tag {11} +$$ + +Finally, the past feature $\mathbf{F}_{\mathbf{p}}$ is concatenated with the refinement feature $\mathbf{F}_r$ to predict the future trajectory. Concretely, we use the pre-trained decoder $f_{d}(\cdot)$ in the first stage to achieve the prediction, as follows: + +$$ +\hat {Y} = f _ {d} \left(\mathbf {F} _ {p} \oplus \mathbf {F} _ {r}\right), \tag {12} +$$ + +where $\hat{Y}$ is the predicted future trajectory. + +Clustering. We adopt the clustering mechanism during inference to further remove the modality redundancy of the retrieved sparse instances. By retrieving $C$ ( $C >> k$ ) sparse instances corresponding to the past trajectory from the well-trained memory, we can generate $C$ refined sparse instances through the refinement network. Then we cluster the $C$ refined sparse instances into $K$ multimodal proposals. The clustered multimodal proposals will guide the multimodal trajectory prediction instead of sparse instances in inference. Ablation studies verify the effectiveness of this post-process. + +# 3.4. Model Training & Inference + +Training. SICNet is trained by a two-stage strategy. In the first stage, we train the encoders $f_{p}(\cdot)$ , $f_{y}(\cdot)$ and the decoder $f_{d}(\cdot)$ in the sparse instance representation module with the reconstruction loss $\mathcal{L}_1$ . Then we use the memory mechanism to store the sparse instances generated during the reconstruction process. In the second stage, we fix the parameters of $f_{p}(\cdot)$ , $f_{y}(\cdot)$ and $f_{d}(\cdot)$ , and train the memory refinement module with the refinement loss $\mathcal{L}_2$ . + +Inference. We aim to predict $K$ trajectories to cover the multimodality of future trajectories. During inference, we select the top- $C$ past memory items by feature similarity and then retrieve corresponding $C$ sparse instance items from the instance memory bank, where $C >> K$ is to improve the prediction diversity. Then we obtained $C$ refinement sparse instances through the memory refinement module. Subsequently, the $C$ sparse instance items are clustered into $K$ multimodal proposals. Finally, we feed the combination of the past feature and multimodal proposals into the decoder to generate $K$ predicted future trajectories. + +# 4. Experiments and Discussions + +# 4.1. Experimental Setting + +Evaluation Datasets. We evaluate our proposed method on two benchmark datasets, i.e., ETH-UCY [31, 21] and Stanford Drone Dataset (SDD) [33]. ETH-UCY is the most commonly used pedestrian trajectory prediction dataset captured in the bird's eye view. It contains five subsets, where the ETH [31] includes ETH and HOTEL subsets, and the + +
MethodVenue/YearInputETHHOTELUNIVZARA1ZARA2AVG
Social LSTM [1]CVPR2016P+S1.09/2.350.79/1.760.67/1.400.47/1.000.56/1.170.72/1.54
Social GAN [13]CVPR2018P+S0.87/1.620.67/1.370.76/1.520.35/0.680.42/0.840.61/1.21
SoPhie [34]CVPR2019P+S0.70/1.430.76/1.670.54/1.240.30/0.630.38/0.780.51/1.15
MANTRA [28]CVPR2020P+S0.48/0.880.17/0.330.37/0.810.27/0.580.30/0.670.32/0.65
PECNet [27]ECCV2020P+S0.54/0.870.18/0.240.35/0.600.22/0.390.17/0.300.29/0.48
SGCN [37]CVPR2021P+S0.63/1.030.32/0.550.37/0.700.29/0.530.25/0.450.37/0.65
AgentFormer [54]ICCV2021P+S0.45/0.750.14/0.220.25/0.450.18/0.300.14/0.240.23/0.39
PCCSNet [41]ICCV2021P0.28/0.540.11/0.190.29/0.600.21/0.440.15/0.340.21/0.42
CAGN [8]AAAI2022P+S0.41/0.650.13/0.230.32/0.540.21/0.380.16/0.330.25/0.43
SIT [36]AAAI2022P+S0.39/0.620.14/0.220.27/0.470.19/0.330.16/0.290.23/0.38
MemoNet [49]CVPR2022P+S0.40/0.610.11/0.170.24/0.430.18/0.320.14/0.240.21/0.35
Ours-P0.27/0.450.11/0.160.26/0.460.19/0.330.13/0.260.19/0.33
+ +Table 1. Comparison with state-of-the-art methods on ETH-UCY in ADE/FDE. P and S indicate inputting the observed history points and social interaction information, respectively. All methods input the observed 8 time steps and output the predicted 12 time steps. Bold indicates the best performance. Underline indicates the second best performance. The Lower the better. + +
MethodVenue/YearInputADE/FDE
Social LSTM [1]CVPR2016P+S31.19/56.97
Social GAN [13]CVPR2018P+S27.23/41.44
SoPhie [34]CVPR2019P+S16.27/29.38
MANTRA [28]CVPR2020P+S8.96/17.76
PECNet [27]ECCV2020P+S9.96/15.88
PCCSNet [41]ICCV2021P8.62/16.16
CAGN [8]AAAI2022P+S9.42/15.93
SIT [36]AAAI2022P+S9.13/15.42
MemoNet [49]CVPR2022P+S8.56/12.66
Ours-P8.44/13.65
+ +Table 2. Comparison with state-of-the-art methods on SDD in ADE/FDE. P and S indicate inputting the observed history points and social interaction information, respectively. All methods input the observed 8 time steps and output the predicted 12 time steps. The Lower the better. + +UCY [21] includes UNIV, ZARA1, and ZARA2 subsets. On ETH-UCY, we use the leave-out-one method for training on four subsets and testing on the other one following previous efforts [40, 54, 36]. SDD [33] is a large pedestrian trajectory prediction dataset also captured in bird's eye view. It contains 20 scenes of more than 10,000 trajectories collected from college campuses and is much larger than ETH-UCY. On SDD, we use a prior train-test split for training and testing according to previous methods [13, 27, 36]. We observe the historical trajectories of 8 time steps and predict the subsequent 12 time steps on both ETH-UCY and SDD. + +Evaluation Metrics. Following previous methods [49, 36, 8, 41, 27], we employ two commonly used metrics, i.e., Average Displacement Error (ADE) and Final Displacement Error (FDE), to evaluate the trajectory prediction perfor + +mance. ADE measures the average $\ell_2$ -norm distance between all points of the ground truth and predicted trajectory. FDE measures the $\ell_2$ -norm distance between the destination points of the ground truth and predicted trajectory. We use the best-of-20 metrics following previous methods [49, 41]. + +Implementation Details. In our experiments, the two encoders are all 3-layer BiLSTM [56] architecture with hidden sizes of 48. The decoder is 3-layer BiLSTM [56] architecture with hidden size of 96. The dimensions of $\theta_{M}$ , $\theta_{V}$ and $\theta_{R}$ are set to 48, 48 and 96, respectively. The parameter $\mu$ of mask threshold is 0.2. The parameter $\xi$ in memory training is 0.0001. The Adam optimizer is used to train our model by 100 epochs with a learning rate of 0.0005, decaying by 0.5 with an interval of 50. + +# 4.2. Quantitative Analysis + +On ETH-UCY. As shown in Table 1, we compare our method with eleven state-of-the-art methods on the ETH-UCY dataset in past six years. Among them, MemoNet [49], PECNet [27], and CAGN [8] are goal-conditioned methods. MANTRA [28] and PCCSNet [41] are instance-conditioned methods. SoPhie [34], SGCN [37] and AgentFormer [54] generate multimodal trajectories by latent variables. The experiment results show that generating multimodal predictions by goals or instance is better than by latent variables. + +Moreover, the results indicate that our method significantly outperforms all the competing methods on both ETH and UCY. Compared with the previous best method MemoNet [49], our method further improves the performance on ADE and FDE. Meanwhile, Compared with the other instance-conditioned methods, PCCSNet [41], and MANTRA [28], our method also achieve the best performance. The underlying reason could be that our method captures the effective intermediate process compared with + +
MethodK=15K=10K=5
PECNet [27]10.66/17.8511.79/21.0014.49/28.08
PCCSNet [41]9.27/17.6210.36/20.1212.57/24.74
CAGN [8]10.01/17.4911.06/19.8913.85/26.87
MemoNet [49]9.62/15.3611.26/19.6414.43/27.45
Ours9.13/15.3110.27/18.1312.82/23.96
+ +Table 3. Experiments on different best-of- $K$ predictions on SDD in ADE/FDE. The lower the better. + +
MethodSIREFCLADE/FDE
Baseline0.46/0.94
(1)✓✓0.25/0.50
(2)✓✓0.27/0.56
(3)✓✓0.22/0.39
(4)✓✓✓0.19/0.33
+ +the goal-conditioned methods and is more flexible compared with the instance-conditioned methods. + +To show the superiority of our method more clearly, we do not use extra information, such as social interaction and map information, as previous approaches. Nevertheless, our method still achieves the best performance, clearly demonstrating the superiority of our method using sparse instance. + +On SDD. We compare our method with nine state-of-the-art methods on the SDD dataset in Table 2. For ADE, our method outperforms all the competing methods. For FDE, our method also achieves comparable performance with the best method MemoNet [49] and exceeds all other methods. The results manifest that the sparse instance representation can comprehensively describe the trajectory multimodality rather than goal-conditioned methods and thus achieves competitive performance. We speculate that our method underperforms MemoNet [49] on FDE because the sparse instance benefits more on the whole trajectory prediction (i.e., ADE) than goal prediction (i.e., FDE). + +# 4.3. Ablation Study + +Different Best-of- $K$ Predictions. Previous methods commonly use best-of- $K$ (usually $K = 20$ ) as the quantified metric of multimodal trajectory prediction. To further validate the adaptability and effectiveness of our proposed sparse instance representation, we conduct an experiment on various best-of- $K$ predictions with $K = 5, 10, 15$ on SDD. We compare our method with the top three performance goal-conditioned methods, i.e., MemoNet [49], CAGN [8] and PECNet [27] and the best performance instance-conditioned method PCCSNet [41]. As shown in Table 3, our method + +Table 4. Ablation study of each component of our method on ETH-UCY dataset in ADE/FDE. The lower the better. + +
μ00.10.20.30.40.5
ADE0.210.210.190.200.200.21
FDE0.340.330.330.330.340.37
μ0.60.70.80.91.0-
ADE0.250.280.260.240.21-
FDE0.450.540.480.430.38-
+ +Table 5. Ablation study of different mask threshold values on the ETH-UCY dataset. The lower the better. + +![](images/835407cf1dfcec615275cc63c3172e8a9470844281c907fd888c92dea4850361.jpg) +Figure 6. Experiments of the sparse instance analysis on the ETH-UCY dataset. We replace our proposed sparse instance with the goal and instance for the multimodal prediction. + +achieves significant performance on all experimental settings. Interestingly, our method surpasses the previous best method MemoNet [49] when $K = 5, 10, 15$ , despite underperforming MemoNet [49] when $K = 20$ in Table 2. It shows the adaptability of our method of using the sparse instance, which also performs significantly with a small $K$ . + +Contribution of Each Component. We divide our proposed SICNet into three components, i.e., the sparse instance representation (SI), the memory refinement module (REF), and the clustering post-process (CL). Ablation experiments of different combinations of the three components are evaluated on the ETH-UCY dataset. As shown in Table 4, all of the three components contribute to the performance improvement, which demonstrates their effectiveness. + +Analysis of Mask Threshold. As shown in Table 5, we conduct experiments for different mask threshold values $\mu$ on the ETH-UCY dataset. It indicates that the performance will decrease when the mask threshold value $\mu$ is too small or too large. When $\mu = 0$ , the mask is useless and the whole instance is reserved, thus causing the stochasticity loss. When $\mu = 1$ , it denotes the whole instance is masked, thus causing no guidance information for the multimodal prediction. When $\mu = 0.2$ , it achieves the best performance. + +Analysis of Sparse Instance. As shown in Table 6, we replace the sparse instance in SICNet with goal-conditioned and instance-conditioned representations on the ETH-UCY dataset. Experiment results show that using the sparse instance achieves the best performance, which indicates the + +![](images/114a206eefaa53bceee93d73a4d0d771baaf1ca2d815ddfd9df79b89a5f6c7c9.jpg) +Figure 7. Qualitative comparison of our method with the goal-conditioned method PECNet [27] on the SDD dataset. Given the observed trajectories, we illustrate the ground truth paths and predicted trajectories by SICNet and the goal-conditioned method PECNet [27] for four different scenes. It shows our prediction is closer to the ground truth than PECNet [27], especially for the intermediate process. + +![](images/44482aaf343e512b106d1580fa14c76cf147044e997bd3d6084dbae7d2ec4298.jpg) +Figure 8. Qualitative comparison of our method with the instance-conditioned method PCCSNet [41] on the ETH-UCY dataset. Given the observed trajectories, we illustrate the ground truth paths and predicted trajectories by SICNet and the instance-conditioned method PCCSNet [41] for four different scenes. It shows our prediction is closer to the ground truth than PCCSNet [41]. + +superiority of our method compared with goal-conditioned and instance-conditioned approaches. + +# 4.4. Qualitative Analysis + +We compare the visualized trajectories with the well-accepted goal-conditioned method PECNet [27] on SDD in Figure 7, and the instance-conditioned method PCCSNet [41] on ETH-UCY in Figure 8. As shown in Figure 7, trajectories in Coupa show that our method achieves more precise trajectory predictions than PECNet. Trajectories in Little and Quad show that both PECNet and our method can predict destinations close to the ground truth, while our method can predict more precise intermediate trajectories due to our proposed sparse instance representation. Trajectories in Hyang show that our method can also give better predictions, although the ground truth trajectories are irregular. The results verify the superiority of our method compared with goal-conditioned methods. Moreover, as shown in Figure 8, the visualization results show that our prediction is closer to the ground truth than PCCSNet [41], which indicates the superiority of our SICNet compared with the instance-conditioned approaches. + +# 5. Conclusion + +In this paper, we present a simple yet effective sparse-instance-conditioned network for pedestrian multimodal trajectory prediction, which leverages sparse instances to guide multimodal prediction. Moreover, we propose a memory refinement module to reduce recall errors when meeting unfamiliar trajectories during inference. Extensive experiments show that our method performs better than previous methods, and ablation studies demonstrate that our sparse-instance-conditioned method can generate multimodal predictions more accurately and adaptively compared with goal-conditioned and instance-conditioned approaches. + +# Acknowledgement + +This work was supported partly by National Key R&D Program of China under Grant 2021YFB1714700, NSFC under Grants 62088102 and 62106192, Natural Science Foundation of Shaanxi Province under Grant 2022JC-41, and Fundamental Research Funds for the Central Universities under Grant XTR042021005. + +# References + +[1] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social LSTM: Human trajectory prediction in crowded spaces. In CVPR, pages 961–971, 2016. +[2] Inhwan Bae, Jin-Hwi Park, and Hae-Gon Jeon. Nonprobability sampling network for stochastic human trajectory prediction. In CVPR, pages 6477-6487, 2022. +[3] Guangyi Chen, Junlong Li, Jiwen Lu, and Jie Zhou. Human trajectory prediction via counterfactual analysis. In ICCV, pages 9824-9833, 2021. +[4] Xinlei Chen and Abhinav Gupta. Spatial memory for context reasoning in object detection. In ICCV, pages 4086-4096, 2017. +[5] Chiho Choi, Joon Hee Choi, Jiachen Li, and Srikanth Malla. Shared cross-modal trajectory prediction for autonomous driving. In CVPR, pages 244-253, 2021. +[6] Lingwei Dang, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. MSR-GCN: Multi-scale residual graph convolution networks for human motion prediction. In ICCV, pages 11447-11456, 2021. +[7] Hanming Deng, Yang Hua, Tao Song, Zongpu Zhang, Zhengui Xue, Ruhui Ma, Neil Robertson, and Haibing Guan. Object guided external memory network for video object detection. In ICCV, pages 6677-6686, 2019. +[8] Jinghai Duan, Le Wang, Chengjiang Long, Sanping Zhou, Fang Zheng, Liushuai Shi, and Gang Hua. Complementary attention gated network for pedestrian trajectory prediction. In AAAI, pages 542-550, 2022. +[9] Zhihong Fu, Qingjie Liu, Zehua Fu, and Yunhong Wang. STMTrack: Template-free visual tracking with space-time memory networks. In CVPR, pages 13769-13778, 2021. +[10] Harshayu Girase, Haiming Gang, Srikanth Malla, Jiachen Li, Akira Kanehara, Karttikeya Mangalam, and Chiho Choi. LOKI: Long term and key intentions for trajectory prediction. In ICCV, pages 9803-9812, 2021. +[11] Dong Gong, Frederic Z. Zhang, Javen Qinfeng Shi, and Anton van den Hengel. Memory-augmented dynamic neural relational inference. In ICCV, pages 11843-11852, 2021. +[12] Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory prediction via motion indeterminacy diffusion. In CVPR, pages 17113-17122, 2022. +[13] Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi. Social GAN: socially acceptable trajectories with generative adversarial networks. In CVPR, pages 2255–2264, 2018. +[14] Adam Houenou, Philippe Bonnifait, Véronique Cherfaoui, and Wen Yao. Vehicle trajectory prediction based on motion model and maneuver recognition. In IROS, pages 4363-4369, 2013. +[15] Boris Ivanovic and Marco Pavone. The trajectory: Probabilistic multi-agent trajectory modeling with dynamic spatiotemporal graphs. In ICCV, pages 2375-2384, 2019. +[16] Lukasz Kaiser, Ofir Nachum, Aurko Roy, and Samy Bengio. Learning to remember rare events. In ICLR, 2017. + +[17] Youngjin Kim, Minjung Kim, and Gunhee Kim. Memorization precedes generation: Learning unsupervised gans with memory networks. In ICLR, 2018. +[18] Sumith Kulal, Jiayuan Mao, Alex Aiken, and Jiajun Wu. Programmatic concept learning for human motion description and synthesis. In CVPR, pages 13843-13852, 2022. +[19] Mihee Lee, Samuel S. Sohn, Seonghyeon Moon, Sejong Yoon, Mubbasir Kapadia, and Vladimir Pavlovic. MUSE-VAE: Multi-scale vae for environment-aware long term trajectory prediction. In CVPR, pages 2221–2230, 2022. +[20] Sangho Lee, Jinyoung Sung, Youngjae Yu, and Gunhee Kim. A memory network approach for story-based temporal summarization of $360^{\circ}$ videos. In CVPR, pages 1410-1419, 2018. +[21] Alon Lerner, Yiorgos Chrysanthou, and Dani Lischinski. Crowds by example. Comput. Graph. Forum, 26(3):655-664, 2007. +[22] Peizhao Li, Pu Wang, Karl Berntorp, and Hongfu Liu. Exploiting temporal relations on radar perception for autonomous driving. In CVPR, pages 17071-17080, 2022. +[23] Shijie Li, Yanying Zhou, Jinhui Yi, and Juergen Gall. Spatial-temporal consistency network for low-latency trajectory forecasting. In ICCV, pages 1940-1949, 2021. +[24] Jialun Liu, Yifan Sun, Feng Zhu, Hongbin Pei, Yi Yang, and Wenhui Li. Learning memory-augmented unidirectional metrics for cross-modality person re-identification. In CVPR, pages 19366-19375, 2022. +[25] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In ICCV, pages 13588-13597, 2021. +[26] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In ICCV, pages 15233-15242, 2021. +[27] Karttikeya Mangalam, Harshayu Girase, Shreyas Agarwal, Kuan-Hui Lee, Ehsan Adeli, Jitendra Malik, and Adrien Gaidon. It is not the journey but the destination: Endpoint conditioned trajectory prediction. In ECCV, pages 759-776, 2020. +[28] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. *Mantra: Memory augmented networks for multiple trajectory prediction*. In *CVPR*, pages 7143–7152, 2020. +[29] Abduallah Mohamed, Kun Qian, Mohamed Elhoseiny, and Christian Claudel. Social-STGCNN: A social spatio-temporal graph convolutional neural network for human trajectory prediction. In CVPR, pages 14424-14432, 2020. +[30] Alessio Monti, Angelo Porrello, Simone Calderara, Pasquale Coscia, Lamberto Ballan, and Rita Cucchiara. How many observations are enough? knowledge distillation for trajectory forecasting. In CVPR, pages 6553-6562, 2022. +[31] Stefano Pellegrini, Andreas Ess, Konrad Schindler, and Luc Van Gool. You'll never walk alone: Modeling social behavior for multi-target tracking. In ICCV, pages 261-268, 2009. +[32] Zheng Qin, Sanping Zhou, Le Wang, Jinghai Duan, Gang Hua, and Wei Tang. Motiontrack: Learning robust short-term and long-term motions for multi-object tracking. In CVPR, pages 17939-17948, 2023. + +[33] Alexandre Robicquet, Amir Sadeghian, Alexandre Alahi, and Silvio Savarese. Learning Social Etiquette: Human trajectory understanding in crowded scenes. In ECCV, pages 549-565, 2016. +[34] Amir Sadeghian, Vineet Kosaraju, Ali Sadeghian, Noriaki Hirose, Hamid Rezatofighi, and Silvio Savarese. SoPhie: An attentive gan for predicting paths compliant to social and physical constraints. In CVPR, pages 1349-1358, 2019. +[35] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory: Dynamically-feasible trajectory forecasting with heterogeneous data. In ECCV, pages 683-700, 2020. +[36] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Fang Zheng, Nanning Zheng, and Gang Hua. Social interpretable tree for pedestrian trajectory prediction. In AAAI, pages 2235-2243, 2022. +[37] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Mo Zhou, Zhenxing Niu, and Gang Hua. SGCN: Sparse graph convolution network for pedestrian trajectory prediction. In CVPR, pages 8990-8999, 2021. +[38] Aravind Srinivas, Tsung-Yi Lin, Niki Parmar, Jonathon Shlens, Pieter Abbeel, and Ashish Vaswani. Bottleneck transformers for visual recognition. In CVPR, pages 16519-16529, 2021. +[39] Hao Sun, Zhiqun Zhao, and Zhihai He. Reciprocal learning networks for human trajectory prediction. In CVPR, pages 7416-7425, 2020. +[40] Jianhua Sun, Qinhong Jiang, and Cewu Lu. Recursive social behavior graph for trajectory prediction. In CVPR, pages 660-669, 2020. +[41] Jianhua Sun, Yuxuan Li, Hao-Shu Fang, and Cewu Lu. Three steps to multimodal trajectory prediction: Modality clustering, classification and synthesis. In ICCV, pages 13250-13259, 2021. +[42] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, Wenyu Liu, and Bin Xiao. Deep high-resolution representation learning for visual recognition. IEEE T-PAMI, 43(10):3349-3364, 2020. +[43] Jingke Wang, Tengju Ye, Ziqing Gu, and Junbo Chen. LTP: Lane-based trajectory prediction for autonomous driving. In CVPR, pages 17134-17142, 2022. +[44] Jürgen Wiest, Matthias Höffken, Ulrich Kreßel, and Klaus Dietmayer. Probabilistic trajectory prediction with gaussian mixture models. In IV, pages 141-146, 2012. +[45] Yuxuan Wu, Le Wang, Sanping Zhou, Jinghai Duan, Gang Hua, and Wei Tang. Multi-stream representation learning for pedestrian trajectory prediction. In AAAI, pages 2875-2882, 2023. +[46] Zifeng Wu, Chunhua Shen, and Anton Van Den Hengel. Wider or deeper: Revisiting the resnet model for visual recognition. PR, 90:119-133, 2019. +[47] Haozhe Xie, Hongxun Yao, Shangchen Zhou, Shengping Zhang, and Wenxiu Sun. Efficient regional memory network for video object segmentation. In CVPR, pages 1286-1295, 2021. + +[48] Bo Xu, Han Huang, Cheng Lu, Ziwen Li, and Yandong Guo. Virtual multi-modality self-supervised foreground matting for human-object interaction. In ICCV, pages 428-437, 2021. +[49] Chenxin Xu, Weibo Mao, Wenjun Zhang, and Siheng Chen. Remember intentions: Retrospective-memory-based trajectory prediction. In CVPR, pages 6488-6497, 2022. +[50] Hao Xue, Du Q. Huynh, and Mark Reynolds. SS-LSTM: A hierarchical LSTM model for pedestrian trajectory prediction. In WACV, pages 1186-1194, 2018. +[51] Gang Yang, Man Zhou, Keyu Yan, Aiping Liu, Xueyang Fu, and Fan Wang. Memory-augmented deep conditional unfolding network for pan-sharpening. In CVPR, pages 1788-1797, 2022. +[52] Tianyu Yang and Antoni B. Chan. Visual tracking via dynamic memory networks. IEEE T-PAMI, 43(1):360-374, 2021. +[53] Jiyang Yu, Jingen Liu, Liefeng Bo, and Tao Mei. Memory-augmented non-local attention for video super-resolution. In CVPR, pages 17834-17843, 2022. +[54] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris Kitani. Agent-Former: Agent-aware transformers for socio-temporal multiagent forecasting. In ICCV, pages 9793–9803, 2021. +[55] He Zhao and Richard P. Wildes. Where are you heading? dynamic trajectory prediction with expert goal examples. In ICCV, pages 7609-7618, 2021. +[56] Zheng Zhao, Weihai Chen, Xingming Wu, Peter CY Chen, and Jingmeng Liu. LSTM network: a deep learning approach for short-term traffic forecast. IET Intell. Transp. Syst., 11(2):68-75, 2017. +[57] Desen Zhou, Zhichao Liu, Jian Wang, Leshan Wang, Tao Hu, Errui Ding, and Jingdong Wang. Human-object interaction detection via disentangled transformer. In CVPR, pages 19568-19577, 2022. \ No newline at end of file diff --git a/2023/Sparse Instance Conditioned Multimodal Trajectory Prediction/images.zip b/2023/Sparse Instance Conditioned Multimodal Trajectory Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cab787ac8cd0ec2bc9e0a2a41427c41d4a481a99 --- /dev/null +++ b/2023/Sparse Instance Conditioned Multimodal Trajectory Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7dc9c6c842874d3c1b4b34a76bdb69a1509f5ab5f922d3560cefe8a37d79070b +size 612780 diff --git a/2023/Sparse Instance Conditioned Multimodal Trajectory Prediction/layout.json b/2023/Sparse Instance Conditioned Multimodal Trajectory Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..864ed52fe8692af7ea45806666e8264b0c3f26f0 --- /dev/null +++ b/2023/Sparse Instance Conditioned Multimodal Trajectory Prediction/layout.json @@ -0,0 +1,9662 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 103, + 493, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 103, + 493, + 121 + ], + "spans": [ + { + "bbox": [ + 100, + 103, + 493, + 121 + ], + "type": "text", + "content": "Sparse Instance Conditioned Multimodal Trajectory Prediction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 142, + 490, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 142, + 490, + 215 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 490, + 215 + ], + "type": "text", + "content": "Yonghao Dong" + }, + { + "bbox": [ + 104, + 142, + 490, + 215 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 142, + 490, + 215 + ], + "type": "text", + "content": " Le Wang" + }, + { + "bbox": [ + 104, + 142, + 490, + 215 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 104, + 142, + 490, + 215 + ], + "type": "text", + "content": " Sanping Zhou" + }, + { + "bbox": [ + 104, + 142, + 490, + 215 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 142, + 490, + 215 + ], + "type": "text", + "content": " Gang Hua" + }, + { + "bbox": [ + 104, + 142, + 490, + 215 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 142, + 490, + 215 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 142, + 490, + 215 + ], + "type": "text", + "content": "National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University \n" + }, + { + "bbox": [ + 104, + 142, + 490, + 215 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 142, + 490, + 215 + ], + "type": "text", + "content": "Wormpex AI Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 240, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 240, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 240, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 266, + 290, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 555 + ], + "type": "text", + "content": "Pedestrian trajectory prediction is critical in many vision tasks but challenging due to the multimodality of the future trajectory. Most existing methods predict multimodal trajectories conditioned by goals (future endpoints) or instances (all future points). However, goal-conditioned methods ignore the intermediate process and instance-conditioned methods ignore the stochasticity of pedestrian motions. In this paper, we propose a simple yet effective Sparse Instance Conditioned Network (SICNet), which gives a balanced solution between goal-conditioned and instance-conditioned methods. Specifically, SICNet learns comprehensive sparse instances, i.e., representative points of the future trajectory, through a mask generated by a long short-term memory encoder and uses the memory mechanism to store and retrieve such sparse instances. Hence SICNet can decode the observed trajectory into the future prediction conditioned on the stored sparse instance. Moreover, we design a memory refinement module that refines the retrieved sparse instances from the memory to reduce memory recall errors. Extensive experiments on ETH-UCY and SDD datasets show that our method outperforms existing state-of-the-art methods. In addition, ablation studies demonstrate the superiority of our method compared with goal-conditioned and instance-conditioned approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 597, + 288, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 597, + 288, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 288, + 694 + ], + "type": "text", + "content": "Pedestrian trajectory prediction aims to predict pedestrians' future paths given their observed trajectories. It plays an important role in autonomous driving [22, 43, 5], human motion prediction [6, 12, 19], video surveillance [57, 48, 18, 32], and visual recognition [38, 42, 46]. Despite the recent advancement [30, 49, 8, 36, 10], pedestrian trajectory prediction is still challenging due to the multimodality of future trajectories. As the pedestrian's motions are stochastic and" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 242, + 425, + 300 + ], + "blocks": [ + { + "bbox": [ + 310, + 242, + 425, + 300 + ], + "lines": [ + { + "bbox": [ + 310, + 242, + 425, + 300 + ], + "spans": [ + { + "bbox": [ + 310, + 242, + 425, + 300 + ], + "type": "image", + "image_path": "9834723deedee81acb7629c918849e3d01939dfaf75a99252920265d774667d7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 428, + 242, + 543, + 300 + ], + "blocks": [ + { + "bbox": [ + 428, + 242, + 543, + 300 + ], + "lines": [ + { + "bbox": [ + 428, + 242, + 543, + 300 + ], + "spans": [ + { + "bbox": [ + 428, + 242, + 543, + 300 + ], + "type": "image", + "image_path": "b3ca369f619b2a0ce5c7a819e86e2d306f7f93b4e53e00beb0c6bf8133c923a9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 303, + 425, + 359 + ], + "blocks": [ + { + "bbox": [ + 310, + 303, + 425, + 359 + ], + "lines": [ + { + "bbox": [ + 310, + 303, + 425, + 359 + ], + "spans": [ + { + "bbox": [ + 310, + 303, + 425, + 359 + ], + "type": "image", + "image_path": "17c4f188ba87645800b3b44f78a1d83ce123d79c50c449ab186f87822314067a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 428, + 302, + 544, + 359 + ], + "blocks": [ + { + "bbox": [ + 428, + 302, + 544, + 359 + ], + "lines": [ + { + "bbox": [ + 428, + 302, + 544, + 359 + ], + "spans": [ + { + "bbox": [ + 428, + 302, + 544, + 359 + ], + "type": "image", + "image_path": "5785870e0f39c99ece345d3ce4ddc16782130a70c1e0e6c583caad5bdbe02b76.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 308, + 361, + 544, + 373 + ], + "blocks": [ + { + "bbox": [ + 308, + 361, + 544, + 373 + ], + "lines": [ + { + "bbox": [ + 308, + 361, + 544, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 544, + 373 + ], + "type": "image", + "image_path": "af08258e5f443f5d0219df92a8af7aea13bb1ae842c81123e9122deb83e0886e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 376, + 547, + 454 + ], + "lines": [ + { + "bbox": [ + 305, + 376, + 547, + 454 + ], + "spans": [ + { + "bbox": [ + 305, + 376, + 547, + 454 + ], + "type": "text", + "content": "Figure 1. Illustration of different multimodal prediction methods. (a) Goals, instances and sparse-instances prepared for the prediction. (b) Goal-conditioned methods. (c) Instance-conditioned methods. (d) Sparse-instance-conditioned methods. The sparse-instance-conditioned method is more accurate in in-between states than the goal-conditioned method and more flexible than the instance-conditioned method, thus making the better prediction." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 474, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 474, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 305, + 474, + 545, + 497 + ], + "type": "text", + "content": "indeterminate [29], multiple future paths are possible given their current states." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": "To handle multimodality, many methods [27, 49, 55] transform the multimodal trajectories into multimodal goals. They first predict various goals as modality representations and then sample multiple goals to generate diverse future trajectories. However, they only focus on goals while neglecting the intermediate movement process. As illustrated in Figure 1(b), although with accurate goals, the intermediate process could be far from the ground truth. Thus, the multimodal goals are not equal to the multimodal future trajectories. It is necessary to consider the in-between states to comprehensively describe the multimodality of trajectory. Moreover, some other work [27, 28] predicts multimodal trajectories conditioned by instances. However, the instance-conditioned methods will lead to motion stochasticity loss and modality redundancy, resulting in performance degradation and computational resource waste. As shown in Figure 1(c), due to the reduced randomness of the instance-conditioned approach, the prediction retains the same bias" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9763" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 247, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 247, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 247, + 83 + ], + "type": "text", + "content": "when the instance deviates from the ground truth." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 85, + 287, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 85, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 49, + 85, + 287, + 382 + ], + "type": "text", + "content": "To close this gap, we present a Sparse Instance Conditioned Network (SICNet). In our model, a sparse instance, which can give a comprehensive and adaptive description of future modality, is constructed by element-wise multiplication between the future trajectory and a learned mask. Hence, SICNet can predict the future trajectory more precisely guided by the sparse instance. To obtain the related sparse instances during inference, we adopt the memory mechanism to store non-redundant sparse instances corresponding to all observed trajectories. Therefore, SICNet can recall its modality's most related sparse instance and jointly generate the prediction during inference. Moreover, when the memory meets an unfamiliar observed trajectory, i.e., a trajectory quite different from stored items, the memory mechanism may recall false experiences. Hence we propose a memory refinement module with a refinement loss to reduce such recall errors by bridging the gap between the recall and the ground truth. For multimodal trajectories generation, we can retrieve a certain number of the most likely sparse instances from the external memory and then generate corresponding multimodal trajectories. In addition, to further remove the redundancy from the retrieved sparse instances, we use the cluster mechanism as a post-process to obtain succinct multimodal proposals for the prediction instead of the sparse instance during inference." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 384, + 287, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 384, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 49, + 384, + 287, + 562 + ], + "type": "text", + "content": "Extensive experiments and ablation studies on two popular pedestrian trajectory prediction datasets, i.e., ETH-UCY [31, 21] and Stanford Drone Dataset (SDD) [33], demonstrate the superiority of our method over prior state-of-the-art methods. In summary, our contributions are four-fold: (1) We give a balanced solution, i.e., a novel sparse instance, between the goal and instance to guide the multimodal trajectory prediction. (2) We propose a simple yet effective sparse instance conditioned network for the multimodal trajectory prediction. (3) We propose a memory refinement module for the external memory to reduce recall errors when meeting unfamiliar observed trajectories. (4) Extensive experiments and ablation studies demonstrate the superiority and flexibility of our method compared with the goal-conditioned and instance-conditioned approaches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 573, + 137, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 573, + 137, + 586 + ], + "spans": [ + { + "bbox": [ + 49, + 573, + 137, + 586 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 594, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 594, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 49, + 594, + 287, + 700 + ], + "type": "text", + "content": "Multimodal Pedestrian Trajectory Prediction. Pedestrian trajectory prediction [2, 23, 3, 39, 45] aims to predict pedestrians' future trajectories based on their observed trajectories. There are some early studies on the deterministic trajectory prediction [1, 44, 14, 50]. Due to the strong randomness and uncertainty of pedestrian motions, there is no single correct future trajectory prediction. Motivated by this, Social GAN [13] proposes the concept of multimodal pedestrian trajectory prediction and emphasizes its importance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 702, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 286, + 712 + ], + "type": "text", + "content": "Earlier work generates multimodal trajectories through" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 307, + 72, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 203 + ], + "type": "text", + "content": "latent variables. Social GAN [13] proposes a Generative Adversarial Network based on a social pooling mechanism to generate a multimodal trajectory distribution. SGCN [37] generates multimodal trajectories by modeling the future predictions as a bi-variate Gaussian distribution. The CVAE-based methods [15, 35, 8] generate multimodal trajectories by sampling latent variables from a learned latent space. However, the latent variable conditioned method is persecuted by the problem of model collapse, thus resulting in performance degradation. Moreover, the latent variable conditioned methods also suffer in interpretability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 206, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 433 + ], + "type": "text", + "content": "To address the model collapse and interpretability issues, PECNet [27] proposes a goal-conditioned approach, which transforms multimodal trajectories into multimodal goals and then samples multiple goals to generate corresponding diverse predictions. After that, amounts of goal-conditioned studies [55, 8, 49] emerge and show superior performance. However, goal-conditioned methods represent different modalities with different possible goals while neglecting the intermediate states, thus suffering in performance improvement. In addition, PCCSNet[27] and MANTRA [28] propose instance-conditioned approaches, which use all points of the future instance to describe each modality precisely. However, it could lead to stochasticity loss and modality redundancy, resulting in performance degradation and computational resource waste. Moreover, Y-net [26] use goals and waypoints to make the prediction. However, the waypoints' number and time step locations are manually determined in Y-net [26], which may decrease the prediction accuracy by manually choosing errors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 436, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 436, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 308, + 436, + 545, + 518 + ], + "type": "text", + "content": "In contrast, we give a balanced way, i.e., a novel sparse instance, between the goal and instance to guide the multimodal trajectory prediction. The sparse instance is constructed by the learned representative points of the instance, which can improve modality representation accuracy compared with goal-conditioned methods and preserve stochasticity compared with instance-conditioned methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 522, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 545, + 712 + ], + "type": "text", + "content": "Memory Network. The memory-augmented network [51, 11, 25, 24, 53] uses an external memory module to store critical information and has been widely applied in many areas like detection [7, 4], tracking [52, 9], segmentation [47], image generation [17], and video summarization [20]. Memory GAN [17] proposed a novel end-to-end unsupervised GAN network with a learnable memory network that effectively learns a highly multimodal latent space without suffering from structural discontinuity and forgetting problems. OGEMN [7] proposes the first object-guided external memory network for the online video object detection. STMTrack [52] proposes a novel tracking framework built on a space-time memory network, which can fully use historical information related to the target for better adapting to appearance variations during tracking. RMNet [47] memorizes the target object regions, effectively alleviating" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "9764" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 72, + 545, + 209 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 545, + 209 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 545, + 209 + ], + "type": "image", + "image_path": "b8c818bfbe20c19976c1a850a0c79282ac3897387b75bebbae45b9b12fbbbb25.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 547, + 258 + ], + "type": "text", + "content": "Figure 2. The framework of our proposed SICNet. SICNet adopts a two-stage training strategy. The sparse instance representation module generates the sparse instances by a learned mask and then trains an external memory to store and retrieve them. The memory refinement module refines the retrieved sparse instances from memory to reduce the searching errors. We cluster the refined sparse instances into multimodal proposals to remove redundancy further. The cluster only exists in inference to save training costs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 277, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 289 + ], + "type": "text", + "content": "objects' ambiguity and reducing computational complexity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 289, + 289, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 289, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 289, + 506 + ], + "type": "text", + "content": "Two previous memory-based work, MANTRA [28] and MemoNet [49], also use the memory mechanism in multimodal trajectory prediction. However, the difference are mainly in three aspects: (1) The contribution of our proposed SICNet focuses on a novel representation of the trajectory modality, i.e., a learned sparse instance, while MANTRA [28] uses the whole instance and MemoNet [49] uses the goal to represent modalities of multimodal trajectories; (2) We propose a memory refinement module to reduce recall errors when meeting unfamiliar observed trajectories during inference, while MANTRA [28] and MemoNet [49] ignore such circumstances. (3) To demonstrate the superiority of our proposed sparse instance, we do not use extra social interaction and map information, while MANTRA [28] use extra map information and MemoNet [49] use social interaction information. Despite this, we also achieve the best performance compared with the above two methods, which indicates the superiority of our method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 514, + 127, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 127, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 127, + 526 + ], + "type": "text", + "content": "3. Our Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "content": "As aforementioned, traditional goal-conditioned methods ignore the intermediate process and instance-conditioned methods ignore the stochasticity of pedestrian motions, which leads to performance degradation. In this section, we present a sparse-instance-conditioned network (SICNet) to close this gap. Our SICNet adopts a two-stage training strategy. In the first stage, we learn the sparse instance, which is used to reconstruct the future trajectory jointly with the observed trajectory. Meanwhile, an external memory is trained to store and retrieve the sparse instance. In the second stage, we refine the retrieved sparse instances from the external memory through a refinement module with a refinement loss to reduce the recall errors. In addition, we use the cluster mechanism to remove redundancy of the retrieved sparse instances further during inference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 277, + 430, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 277, + 430, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 430, + 289 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 295, + 547, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 547, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 547, + 417 + ], + "type": "text", + "content": "The objective of pedestrian trajectory prediction is to predict possible future trajectory coordinates based on the observed trajectories. For each target pedestrian, given an observed trajectory " + }, + { + "bbox": [ + 304, + 295, + 547, + 417 + ], + "type": "inline_equation", + "content": "X = \\{(x_{t},y_{t})\\}_{t = 1}^{T_{obs}}" + }, + { + "bbox": [ + 304, + 295, + 547, + 417 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 295, + 547, + 417 + ], + "type": "inline_equation", + "content": "(x_{t},y_{t})" + }, + { + "bbox": [ + 304, + 295, + 547, + 417 + ], + "type": "text", + "content": " is the 2D coordinate at time " + }, + { + "bbox": [ + 304, + 295, + 547, + 417 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 295, + 547, + 417 + ], + "type": "text", + "content": ", we aim to predict the future trajectory coordinates " + }, + { + "bbox": [ + 304, + 295, + 547, + 417 + ], + "type": "inline_equation", + "content": "Y = \\{(x_{t},y_{t})\\}_{t = T_{obs} + 1}^{T_{pred}}" + }, + { + "bbox": [ + 304, + 295, + 547, + 417 + ], + "type": "text", + "content": ". Note that multimodal pedestrian trajectory prediction requires the model to predict " + }, + { + "bbox": [ + 304, + 295, + 547, + 417 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 295, + 547, + 417 + ], + "type": "text", + "content": " future trajectories " + }, + { + "bbox": [ + 304, + 295, + 547, + 417 + ], + "type": "inline_equation", + "content": "\\{Y_i\\}_{i = 1}^K" + }, + { + "bbox": [ + 304, + 295, + 547, + 417 + ], + "type": "text", + "content": " to account for multimodality, while only one future trajectory (ground truth) is provided for training in the dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 423, + 477, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 423, + 477, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 423, + 477, + 437 + ], + "type": "text", + "content": "3.2. Sparse Instance Representation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "type": "text", + "content": "In the first training stage, we generate the sparse instance features to represent future modalities and store them by memory mechanism. As shown in Figure 3, we first train the encoder-decoder with a reconstruction loss in this stage. Then we train the memory module to store the sparse instances generated during reconstruction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 514, + 547, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 547, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 547, + 573 + ], + "type": "text", + "content": "Feature Encoder. As the trajectory is a temporal sequence, we use two LSTM [56] blocks to build the past encoder " + }, + { + "bbox": [ + 304, + 514, + 547, + 573 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 304, + 514, + 547, + 573 + ], + "type": "text", + "content": " and the future encoder " + }, + { + "bbox": [ + 304, + 514, + 547, + 573 + ], + "type": "inline_equation", + "content": "f_{y}(\\cdot)" + }, + { + "bbox": [ + 304, + 514, + 547, + 573 + ], + "type": "text", + "content": ", respectively. The hidden states of the two encoders are used as their corresponding features, as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 396, + 581, + 545, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 581, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 396, + 581, + 545, + 609 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {F} _ {p} = f _ {p} (X), \\\\ \\mathbf {F} _ {y} = f _ {y} (Y), \\tag {1} \\\\ \\end{array}", + "image_path": "9ac42cc7577722b7829f46f63675f7b4abbd7d1d44b91ab779380bb9d1e9386f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 616, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 547, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 616, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p\\in \\mathbb{R}^{1\\times d}" + }, + { + "bbox": [ + 304, + 616, + 547, + 665 + ], + "type": "text", + "content": " is the past feature extracted from the observed trajectory " + }, + { + "bbox": [ + 304, + 616, + 547, + 665 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 616, + 547, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 616, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_y\\in \\mathbb{R}^{1\\times d}" + }, + { + "bbox": [ + 304, + 616, + 547, + 665 + ], + "type": "text", + "content": " is the future feature extracted from the future trajectory " + }, + { + "bbox": [ + 304, + 616, + 547, + 665 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 616, + 547, + 665 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 616, + 547, + 665 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 616, + 547, + 665 + ], + "type": "text", + "content": " is the dimension of hidden state in LSTM." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "Sparse Instance. When inputting an observed trajectory, we want to obtain the corresponding sparse instance of its future modality so that we can construct future trajectories jointly using the observed trajectory and the sparse instance." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9765" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 267, + 129 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 267, + 129 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 267, + 129 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 267, + 129 + ], + "type": "image", + "image_path": "09536bbf7de6c330f9f998b4fd8a62fb6f6238b120d410df38d96efc40496b60.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 132, + 289, + 175 + ], + "lines": [ + { + "bbox": [ + 46, + 132, + 289, + 175 + ], + "spans": [ + { + "bbox": [ + 46, + 132, + 289, + 175 + ], + "type": "text", + "content": "Figure 3. The first stage of the training process. In this stage, we train the sparse instance by reconstructing the future trajectory. Then we train an external memory to store and retrieve the learned sparse instances." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 186, + 286, + 247 + ], + "blocks": [ + { + "bbox": [ + 48, + 186, + 286, + 247 + ], + "lines": [ + { + "bbox": [ + 48, + 186, + 286, + 247 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 286, + 247 + ], + "type": "image", + "image_path": "501526ea587d50b87d4759eb5aa21ac6e2d4da47701750d0f9974eed9fa328c8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 288, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 288, + 283 + ], + "type": "text", + "content": "Figure 4. The second stage of the training process. In this stage, we refine the retrieved sparse instances from memory through the memory refinement module to reduce the recall errors." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 303, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 303, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 303, + 287, + 327 + ], + "type": "text", + "content": "Hence, we first learn an instance mask from the past feature " + }, + { + "bbox": [ + 47, + 303, + 287, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p" + }, + { + "bbox": [ + 47, + 303, + 287, + 327 + ], + "type": "text", + "content": " through an MLP layer, as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 118, + 335, + 287, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 335, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 118, + 335, + 287, + 364 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {F} _ {m} = \\phi (\\mathbf {F} _ {p}, \\theta_ {M}), \\\\ \\begin{array}{l} M = \\mathbb {I} (\\delta (\\mathbf {F} _ {m}) \\leq \\mu)), \\end{array} \\tag {2} \\\\ \\end{array}", + "image_path": "4eb17193b9e345af1e81ee0ce9cf608fcb155afa6a4cfd00f45269d34aa0343b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "type": "text", + "content": " denotes the MLP layer. " + }, + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "type": "inline_equation", + "content": "\\theta_{M}" + }, + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "type": "text", + "content": " is a learnable parameter. " + }, + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_m" + }, + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "type": "text", + "content": " is the learned feature which is used to construct the mask. " + }, + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "type": "text", + "content": " denotes the instance mask. " + }, + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "type": "inline_equation", + "content": "\\delta (\\cdot)" + }, + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "type": "text", + "content": " is sigmoid function to map " + }, + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_m" + }, + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "type": "text", + "content": " to the mask value [0, 1]. " + }, + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "type": "text", + "content": " is an indicator function which equals 0 if the inequality holds, otherwise 1. " + }, + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "type": "text", + "content": " is the mask threshold. Compared with directly masking instances, masking instances in the high-level feature dimension is more efficient and effective. Hence we generate the sparse instance " + }, + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{ins}" + }, + { + "bbox": [ + 46, + 372, + 289, + 479 + ], + "type": "text", + "content": " in the feature level dimension, as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 490, + 287, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 287, + 502 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {i n s} = \\mathbf {F} _ {y} \\odot M, \\tag {3}", + "image_path": "72a71827dce5a17453ccdaf056567a05b60d10f66a244ea760396f045ce3e081.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 513, + 249, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 249, + 524 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 249, + 524 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 513, + 249, + 524 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 47, + 513, + 249, + 524 + ], + "type": "text", + "content": " denotes the element-wise multiplication." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 525, + 288, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 288, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 288, + 608 + ], + "type": "text", + "content": "Feature Decoder. The sparse instance feature should be representative enough for the future trajectory. In addition, our goal is to predict the future trajectory based on the observed trajectory conditioned on the sparse instance. Therefore, we use an LSTM [56] decoder to reconstruct the future trajectory with the combination of the past feature and the sparse instance feature, as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 121, + 619, + 287, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 619, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 121, + 619, + 287, + 632 + ], + "type": "interline_equation", + "content": "Y _ {r} = f _ {d} \\left(\\mathbf {F} _ {p} \\oplus \\mathbf {F} _ {i n s}\\right), \\tag {4}", + "image_path": "f69f84b10b2a1ca836d7c32187e6bf03cba8599601cd09531ec439eb333da440.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 641, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 288, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 641, + 288, + 689 + ], + "type": "inline_equation", + "content": "f_{d}(\\cdot)" + }, + { + "bbox": [ + 47, + 641, + 288, + 689 + ], + "type": "text", + "content": " is the decoder implemented by an LSTM block, " + }, + { + "bbox": [ + 47, + 641, + 288, + 689 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 47, + 641, + 288, + 689 + ], + "type": "text", + "content": " is a concatenation operation, and " + }, + { + "bbox": [ + 47, + 641, + 288, + 689 + ], + "type": "inline_equation", + "content": "Y_{r}" + }, + { + "bbox": [ + 47, + 641, + 288, + 689 + ], + "type": "text", + "content": " is the reconstructed future trajectory. The trajectory " + }, + { + "bbox": [ + 47, + 641, + 288, + 689 + ], + "type": "inline_equation", + "content": "\\ell_{2}" + }, + { + "bbox": [ + 47, + 641, + 288, + 689 + ], + "type": "text", + "content": "-norm loss is used as the reconstruction loss, as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 129, + 699, + 287, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 699, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 129, + 699, + 287, + 712 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {1} = \\left\\| Y - Y _ {r} \\right\\| _ {2}, \\tag {5}", + "image_path": "00e294a07cefc16e993548af6dadc43e295d78b6a853e2e62d23adda208563d5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 72, + 479, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 479, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 479, + 84 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 72, + 479, + 84 + ], + "type": "inline_equation", + "content": "||\\cdot ||_2" + }, + { + "bbox": [ + 305, + 72, + 479, + 84 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 305, + 72, + 479, + 84 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 305, + 72, + 479, + 84 + ], + "type": "text", + "content": "-norm distance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 84, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 84, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 84, + 545, + 263 + ], + "type": "text", + "content": "Memory Architecture. The key idea of the memory module is to use the encoding of the observed trajectory as a memory key to retrieve possible sparse instances and jointly generate the multimodal predictions. To store the sparse instances generated during the reconstruction process, we first randomly initializes two memory banks, i.e., the past memory bank " + }, + { + "bbox": [ + 304, + 84, + 545, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_p \\in \\mathbb{R}^{m \\times d}" + }, + { + "bbox": [ + 304, + 84, + 545, + 263 + ], + "type": "text", + "content": ", and the instance memory bank " + }, + { + "bbox": [ + 304, + 84, + 545, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i \\in \\mathbb{R}^{m \\times d}" + }, + { + "bbox": [ + 304, + 84, + 545, + 263 + ], + "type": "text", + "content": " to store the past and the sparse instance features obtained from the reconstruction process, respectively. " + }, + { + "bbox": [ + 304, + 84, + 545, + 263 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 84, + 545, + 263 + ], + "type": "text", + "content": " denotes the storage size of the memory bank, and " + }, + { + "bbox": [ + 304, + 84, + 545, + 263 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 84, + 545, + 263 + ], + "type": "text", + "content": " is the dimension of the memory item. To combine the past and sparse instance features, the items in these two memory banks are aligned. Specifically, for the " + }, + { + "bbox": [ + 304, + 84, + 545, + 263 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 84, + 545, + 263 + ], + "type": "text", + "content": "-th past feature in " + }, + { + "bbox": [ + 304, + 84, + 545, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_p[j]" + }, + { + "bbox": [ + 304, + 84, + 545, + 263 + ], + "type": "text", + "content": ", there is a unique sparse instance feature in " + }, + { + "bbox": [ + 304, + 84, + 545, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i[j]" + }, + { + "bbox": [ + 304, + 84, + 545, + 263 + ], + "type": "text", + "content": " corresponding to it." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "text", + "content": "Memory Update. Due to the abundant similar trajectories in the dataset, it brings high redundancy to store all of them in the memory bank. Hence we need to remove modality redundancy by memory updating process. Moreover, traditional memory mechanism [16] updates the memory bank with the stored items' labels, while such labels are not provided in trajectory prediction to distinguish trajectories with different modalities. Therefore, we design an unsupervised memory update strategy based on the distance between the stored and current items, as shown in Figure 5. Inspired by [16], an additional mark memory bank " + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{m \\times 1}" + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "text", + "content": " tracks the age of items stored in memory without being used. All items in " + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "text", + "content": " are initialized to 1 and aligned to their corresponding items in " + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_p" + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i" + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "text", + "content": ". Given a pair of inputs " + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "inline_equation", + "content": "(\\mathbf{F}_p, \\mathbf{F}_{ins})" + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "text", + "content": ", we first calculate the feature similarities between " + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p" + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "text", + "content": " and each item in " + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i" + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "text", + "content": ". Then, the " + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "text", + "content": "-th " + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "inline_equation", + "content": "(k \\in \\{1, \\dots, m\\})" + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "text", + "content": " item " + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_p[k]" + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "text", + "content": " with the highest similarity is used to retrieve the corresponding sparse instance feature, i.e., " + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i[k]" + }, + { + "bbox": [ + 304, + 263, + 547, + 491 + ], + "type": "text", + "content": ", through the one-to-one architecture of memory banks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 491, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 491, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 305, + 491, + 545, + 514 + ], + "type": "text", + "content": "Specifically, the distance " + }, + { + "bbox": [ + 305, + 491, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 305, + 491, + 545, + 514 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 305, + 491, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{ins}" + }, + { + "bbox": [ + 305, + 491, + 545, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 491, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i[k]" + }, + { + "bbox": [ + 305, + 491, + 545, + 514 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 376, + 525, + 545, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 525, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 376, + 525, + 545, + 538 + ], + "type": "interline_equation", + "content": "\\mathcal {D} = \\left\\| \\mathbf {M} _ {i} [ k ] - \\mathbf {F} _ {i n s} \\right\\| _ {2}. \\tag {6}", + "image_path": "92d5b21023185dfe3829c7dcd31876f2f7b0ba7ac7038aeeec4c1f29cad50ed9.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "content": "We compare " + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "content": " with a threshold " + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "content": " to judge whether " + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "inline_equation", + "content": "F_{ins}" + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i[k]" + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "content": " have the same modality. When " + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{D} > \\xi" + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "content": ", which means that " + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "inline_equation", + "content": "F_{ins}" + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i[k]" + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "content": " have different modalities. Hence we update " + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_p" + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i" + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "content": " with the oldest age (i.e., corresponding " + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "content": " with the maximum value) by " + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p" + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{ins}" + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "content": ", respectively, as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 332, + 630, + 545, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 630, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 332, + 630, + 545, + 643 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {i} [ s ] \\leftarrow \\mathbf {F} _ {i n s}, \\quad \\mathbf {M} _ {p} [ s ] \\leftarrow \\mathbf {F} _ {p}, \\quad \\mathbf {A} [ s ] \\leftarrow 0, \\tag {7}", + "image_path": "d42c1fefd632794a328fd82bb5307ca8e9393256ff0aece4b16868500871ade0.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "content": " is the index with the oldest age. Since the " + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "content": "-th item in the past memory bank has been updated, the " + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "content": "-th item in " + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "content": " is set to 0 to prevent the newly added item from overwriting it. When " + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D} \\leq \\xi" + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "content": ", which means that " + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "inline_equation", + "content": "F_{ins}" + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i[k]" + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "content": " have the same modality. Hence we only update " + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_p[k]" + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "content": " to approach" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9766" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 286, + 183 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 286, + 183 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 286, + 183 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 286, + 183 + ], + "type": "image", + "image_path": "460dfc9a0acd3bd37e6e9b7b9afbb113be3dbbe29d38c070e4db4d9db6aa5b2d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 186, + 288, + 209 + ], + "lines": [ + { + "bbox": [ + 47, + 186, + 288, + 209 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 288, + 209 + ], + "type": "text", + "content": "Figure 5. An illustration of the memory training process in SICNet. Red and purple lines denote the memory writing process." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 228, + 284, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 284, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 284, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p" + }, + { + "bbox": [ + 47, + 228, + 284, + 243 + ], + "type": "text", + "content": " by taking the normalized average of " + }, + { + "bbox": [ + 47, + 228, + 284, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_p[k]" + }, + { + "bbox": [ + 47, + 228, + 284, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 228, + 284, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p" + }, + { + "bbox": [ + 47, + 228, + 284, + 243 + ], + "type": "text", + "content": ", as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 250, + 288, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 250, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 83, + 250, + 288, + 277 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {p} [ k ] \\leftarrow \\frac {\\mathbf {F} _ {p} + \\mathbf {M} _ {p} [ k ]}{\\| \\mathbf {F} _ {p} + \\mathbf {M} _ {p} [ k ] \\|}, \\quad \\mathbf {A} [ k ] \\leftarrow 0. \\tag {8}", + "image_path": "22cebd98a67ebb8156f331b2f88f34c3ce47501a14683f1676c03ec9528d6c61.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 285, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 289, + 360 + ], + "type": "text", + "content": "Thus, we can retrieve the sparse instance feature " + }, + { + "bbox": [ + 47, + 285, + 289, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i[k]" + }, + { + "bbox": [ + 47, + 285, + 289, + 360 + ], + "type": "text", + "content": " easily, when we receive the past feature similar to " + }, + { + "bbox": [ + 47, + 285, + 289, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p" + }, + { + "bbox": [ + 47, + 285, + 289, + 360 + ], + "type": "text", + "content": ". Since the " + }, + { + "bbox": [ + 47, + 285, + 289, + 360 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 285, + 289, + 360 + ], + "type": "text", + "content": "-th item in the past memory bank has been updated, the " + }, + { + "bbox": [ + 47, + 285, + 289, + 360 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 285, + 289, + 360 + ], + "type": "text", + "content": "-th item in " + }, + { + "bbox": [ + 47, + 285, + 289, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 47, + 285, + 289, + 360 + ], + "type": "text", + "content": " is set to 0 to prevent the newly added item from covering it. All values in " + }, + { + "bbox": [ + 47, + 285, + 289, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 47, + 285, + 289, + 360 + ], + "type": "text", + "content": " will add the same constant after an epoch." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 365, + 167, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 167, + 379 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 167, + 379 + ], + "type": "text", + "content": "3.3. Memory Refinement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 384, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 491 + ], + "type": "text", + "content": "The input observed trajectory encoding could be unfamiliar to the well-trained memory, which could be caused by: (1) modality loss due to removing redundancy during the memory update process; (2) the input modality does not exist in the train set. Such unfamiliarity could lead to the recalled sparse instance not matching the input trajectory. To reduce such recall errors, we design a memory refinement module with a refinement loss in the second training stage, as shown in Figure 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "text", + "content": "We first consider the difference between the past feature " + }, + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p" + }, + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "text", + "content": " and its corresponding closest past memory item " + }, + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_p[k]" + }, + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "text", + "content": " with the same modality. To avoid extra optimization similar to Eq. (8), we do not optimize " + }, + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p" + }, + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "text", + "content": " to approach " + }, + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_p[k]" + }, + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "text", + "content": " again but encode the difference between " + }, + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p" + }, + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_p[k]" + }, + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "text", + "content": " as an additional deviation feature " + }, + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_v" + }, + { + "bbox": [ + 47, + 492, + 287, + 564 + ], + "type": "text", + "content": ", which is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 111, + 574, + 287, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 574, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 111, + 574, + 287, + 588 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {v} = \\phi \\left(\\mathbf {F} _ {p} \\oplus \\mathbf {M} _ {p} [ k ], \\theta_ {V}\\right), \\tag {9}", + "image_path": "470883805bc4cd3c6e7b4c8e904cce0490956e1e1715398f841ea1a7cc83b19b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 597, + 288, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 288, + 633 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 597, + 288, + 633 + ], + "type": "inline_equation", + "content": "\\theta_V" + }, + { + "bbox": [ + 47, + 597, + 288, + 633 + ], + "type": "text", + "content": " is a learnable parameter, and " + }, + { + "bbox": [ + 47, + 597, + 288, + 633 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 47, + 597, + 288, + 633 + ], + "type": "text", + "content": " is the concatenation operation. Then, the retrieved sparse instance feature is refined as " + }, + { + "bbox": [ + 47, + 597, + 288, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_r" + }, + { + "bbox": [ + 47, + 597, + 288, + 633 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 111, + 643, + 287, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 643, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 111, + 643, + 287, + 656 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {r} = \\psi \\left(\\mathbf {F} _ {v} \\oplus \\mathbf {M} _ {i} [ k ], \\theta_ {R}\\right), \\tag {10}", + "image_path": "1209d7061961ef4c175054f6484c6bf9b83a932df07cbd72794bf01834a38486.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\psi (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "text", + "content": " is a linear projection with a learnable parameter " + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{R}" + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_r" + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "text", + "content": " is the refined sparse instance feature. In this case, the model can give a reasonable sparse instance considering the difference between the input observed feature and the stored" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 547, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 547, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 547, + 96 + ], + "type": "text", + "content": "observed feature in the memory bank. The loss function of the memory refinement module is:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 381, + 106, + 547, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 106, + 547, + 119 + ], + "spans": [ + { + "bbox": [ + 381, + 106, + 547, + 119 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {2} = \\left\\| \\mathbf {F} _ {i n s} - \\mathbf {F} _ {r} \\right\\| _ {2}. \\tag {11}", + "image_path": "175c60a9ecd7789d318ae379aa58b9c4cdf7c4e114b332eb11e141a45d285d5b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 128, + 547, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 128, + 547, + 176 + ], + "spans": [ + { + "bbox": [ + 305, + 128, + 547, + 176 + ], + "type": "text", + "content": "Finally, the past feature " + }, + { + "bbox": [ + 305, + 128, + 547, + 176 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{p}}" + }, + { + "bbox": [ + 305, + 128, + 547, + 176 + ], + "type": "text", + "content": " is concatenated with the refinement feature " + }, + { + "bbox": [ + 305, + 128, + 547, + 176 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_r" + }, + { + "bbox": [ + 305, + 128, + 547, + 176 + ], + "type": "text", + "content": " to predict the future trajectory. Concretely, we use the pre-trained decoder " + }, + { + "bbox": [ + 305, + 128, + 547, + 176 + ], + "type": "inline_equation", + "content": "f_{d}(\\cdot)" + }, + { + "bbox": [ + 305, + 128, + 547, + 176 + ], + "type": "text", + "content": " in the first stage to achieve the prediction, as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 385, + 185, + 546, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 185, + 546, + 200 + ], + "spans": [ + { + "bbox": [ + 385, + 185, + 546, + 200 + ], + "type": "interline_equation", + "content": "\\hat {Y} = f _ {d} \\left(\\mathbf {F} _ {p} \\oplus \\mathbf {F} _ {r}\\right), \\tag {12}", + "image_path": "ca8a0bb17e225a4cb0824df112f082f66b752a6e21a6fe2c36fe1bfde02c537a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 209, + 476, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 209, + 476, + 221 + ], + "spans": [ + { + "bbox": [ + 305, + 209, + 476, + 221 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 209, + 476, + 221 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 305, + 209, + 476, + 221 + ], + "type": "text", + "content": " is the predicted future trajectory." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 221, + 547, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 221, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 547, + 342 + ], + "type": "text", + "content": "Clustering. We adopt the clustering mechanism during inference to further remove the modality redundancy of the retrieved sparse instances. By retrieving " + }, + { + "bbox": [ + 304, + 221, + 547, + 342 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 221, + 547, + 342 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 221, + 547, + 342 + ], + "type": "inline_equation", + "content": "C >> k" + }, + { + "bbox": [ + 304, + 221, + 547, + 342 + ], + "type": "text", + "content": ") sparse instances corresponding to the past trajectory from the well-trained memory, we can generate " + }, + { + "bbox": [ + 304, + 221, + 547, + 342 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 221, + 547, + 342 + ], + "type": "text", + "content": " refined sparse instances through the refinement network. Then we cluster the " + }, + { + "bbox": [ + 304, + 221, + 547, + 342 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 221, + 547, + 342 + ], + "type": "text", + "content": " refined sparse instances into " + }, + { + "bbox": [ + 304, + 221, + 547, + 342 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 221, + 547, + 342 + ], + "type": "text", + "content": " multimodal proposals. The clustered multimodal proposals will guide the multimodal trajectory prediction instead of sparse instances in inference. Ablation studies verify the effectiveness of this post-process." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 347, + 461, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 347, + 461, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 347, + 461, + 361 + ], + "type": "text", + "content": "3.4. Model Training & Inference" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 365, + 547, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 547, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 547, + 461 + ], + "type": "text", + "content": "Training. SICNet is trained by a two-stage strategy. In the first stage, we train the encoders " + }, + { + "bbox": [ + 304, + 365, + 547, + 461 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 304, + 365, + 547, + 461 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 365, + 547, + 461 + ], + "type": "inline_equation", + "content": "f_{y}(\\cdot)" + }, + { + "bbox": [ + 304, + 365, + 547, + 461 + ], + "type": "text", + "content": " and the decoder " + }, + { + "bbox": [ + 304, + 365, + 547, + 461 + ], + "type": "inline_equation", + "content": "f_{d}(\\cdot)" + }, + { + "bbox": [ + 304, + 365, + 547, + 461 + ], + "type": "text", + "content": " in the sparse instance representation module with the reconstruction loss " + }, + { + "bbox": [ + 304, + 365, + 547, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 304, + 365, + 547, + 461 + ], + "type": "text", + "content": ". Then we use the memory mechanism to store the sparse instances generated during the reconstruction process. In the second stage, we fix the parameters of " + }, + { + "bbox": [ + 304, + 365, + 547, + 461 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 304, + 365, + 547, + 461 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 365, + 547, + 461 + ], + "type": "inline_equation", + "content": "f_{y}(\\cdot)" + }, + { + "bbox": [ + 304, + 365, + 547, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 365, + 547, + 461 + ], + "type": "inline_equation", + "content": "f_{d}(\\cdot)" + }, + { + "bbox": [ + 304, + 365, + 547, + 461 + ], + "type": "text", + "content": ", and train the memory refinement module with the refinement loss " + }, + { + "bbox": [ + 304, + 365, + 547, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 304, + 365, + 547, + 461 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 462, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 594 + ], + "type": "text", + "content": "Inference. We aim to predict " + }, + { + "bbox": [ + 304, + 462, + 547, + 594 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 462, + 547, + 594 + ], + "type": "text", + "content": " trajectories to cover the multimodality of future trajectories. During inference, we select the top- " + }, + { + "bbox": [ + 304, + 462, + 547, + 594 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 462, + 547, + 594 + ], + "type": "text", + "content": " past memory items by feature similarity and then retrieve corresponding " + }, + { + "bbox": [ + 304, + 462, + 547, + 594 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 462, + 547, + 594 + ], + "type": "text", + "content": " sparse instance items from the instance memory bank, where " + }, + { + "bbox": [ + 304, + 462, + 547, + 594 + ], + "type": "inline_equation", + "content": "C >> K" + }, + { + "bbox": [ + 304, + 462, + 547, + 594 + ], + "type": "text", + "content": " is to improve the prediction diversity. Then we obtained " + }, + { + "bbox": [ + 304, + 462, + 547, + 594 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 462, + 547, + 594 + ], + "type": "text", + "content": " refinement sparse instances through the memory refinement module. Subsequently, the " + }, + { + "bbox": [ + 304, + 462, + 547, + 594 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 462, + 547, + 594 + ], + "type": "text", + "content": " sparse instance items are clustered into " + }, + { + "bbox": [ + 304, + 462, + 547, + 594 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 462, + 547, + 594 + ], + "type": "text", + "content": " multimodal proposals. Finally, we feed the combination of the past feature and multimodal proposals into the decoder to generate " + }, + { + "bbox": [ + 304, + 462, + 547, + 594 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 462, + 547, + 594 + ], + "type": "text", + "content": " predicted future trajectories." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 604, + 470, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 604, + 470, + 617 + ], + "spans": [ + { + "bbox": [ + 305, + 604, + 470, + 617 + ], + "type": "text", + "content": "4. Experiments and Discussions" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 623, + 429, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 429, + 636 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 429, + 636 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "Evaluation Datasets. We evaluate our proposed method on two benchmark datasets, i.e., ETH-UCY [31, 21] and Stanford Drone Dataset (SDD) [33]. ETH-UCY is the most commonly used pedestrian trajectory prediction dataset captured in the bird's eye view. It contains five subsets, where the ETH [31] includes ETH and HOTEL subsets, and the" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9767" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 547, + 239 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 547, + 239 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 547, + 239 + ], + "type": "table", + "html": "
MethodVenue/YearInputETHHOTELUNIVZARA1ZARA2AVG
Social LSTM [1]CVPR2016P+S1.09/2.350.79/1.760.67/1.400.47/1.000.56/1.170.72/1.54
Social GAN [13]CVPR2018P+S0.87/1.620.67/1.370.76/1.520.35/0.680.42/0.840.61/1.21
SoPhie [34]CVPR2019P+S0.70/1.430.76/1.670.54/1.240.30/0.630.38/0.780.51/1.15
MANTRA [28]CVPR2020P+S0.48/0.880.17/0.330.37/0.810.27/0.580.30/0.670.32/0.65
PECNet [27]ECCV2020P+S0.54/0.870.18/0.240.35/0.600.22/0.390.17/0.300.29/0.48
SGCN [37]CVPR2021P+S0.63/1.030.32/0.550.37/0.700.29/0.530.25/0.450.37/0.65
AgentFormer [54]ICCV2021P+S0.45/0.750.14/0.220.25/0.450.18/0.300.14/0.240.23/0.39
PCCSNet [41]ICCV2021P0.28/0.540.11/0.190.29/0.600.21/0.440.15/0.340.21/0.42
CAGN [8]AAAI2022P+S0.41/0.650.13/0.230.32/0.540.21/0.380.16/0.330.25/0.43
SIT [36]AAAI2022P+S0.39/0.620.14/0.220.27/0.470.19/0.330.16/0.290.23/0.38
MemoNet [49]CVPR2022P+S0.40/0.610.11/0.170.24/0.430.18/0.320.14/0.240.21/0.35
Ours-P0.27/0.450.11/0.160.26/0.460.19/0.330.13/0.260.19/0.33
", + "image_path": "280f2c4e25c1cc2a539c41ec821687ae8238491c091ee7d60f7d6f3b5d4e5a00.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 51, + 299, + 284, + 449 + ], + "blocks": [ + { + "bbox": [ + 47, + 247, + 546, + 281 + ], + "lines": [ + { + "bbox": [ + 47, + 247, + 546, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 546, + 281 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-art methods on ETH-UCY in ADE/FDE. P and S indicate inputting the observed history points and social interaction information, respectively. All methods input the observed 8 time steps and output the predicted 12 time steps. Bold indicates the best performance. Underline indicates the second best performance. The Lower the better." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 299, + 284, + 449 + ], + "lines": [ + { + "bbox": [ + 51, + 299, + 284, + 449 + ], + "spans": [ + { + "bbox": [ + 51, + 299, + 284, + 449 + ], + "type": "table", + "html": "
MethodVenue/YearInputADE/FDE
Social LSTM [1]CVPR2016P+S31.19/56.97
Social GAN [13]CVPR2018P+S27.23/41.44
SoPhie [34]CVPR2019P+S16.27/29.38
MANTRA [28]CVPR2020P+S8.96/17.76
PECNet [27]ECCV2020P+S9.96/15.88
PCCSNet [41]ICCV2021P8.62/16.16
CAGN [8]AAAI2022P+S9.42/15.93
SIT [36]AAAI2022P+S9.13/15.42
MemoNet [49]CVPR2022P+S8.56/12.66
Ours-P8.44/13.65
", + "image_path": "6f8846fd4791cee1c4f8e677af96434bb27cb6ed5525e811994d3c02b2cca5d4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 456, + 288, + 514 + ], + "lines": [ + { + "bbox": [ + 46, + 456, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 288, + 514 + ], + "type": "text", + "content": "Table 2. Comparison with state-of-the-art methods on SDD in ADE/FDE. P and S indicate inputting the observed history points and social interaction information, respectively. All methods input the observed 8 time steps and output the predicted 12 time steps. The Lower the better." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 532, + 288, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 288, + 664 + ], + "type": "text", + "content": "UCY [21] includes UNIV, ZARA1, and ZARA2 subsets. On ETH-UCY, we use the leave-out-one method for training on four subsets and testing on the other one following previous efforts [40, 54, 36]. SDD [33] is a large pedestrian trajectory prediction dataset also captured in bird's eye view. It contains 20 scenes of more than 10,000 trajectories collected from college campuses and is much larger than ETH-UCY. On SDD, we use a prior train-test split for training and testing according to previous methods [13, 27, 36]. We observe the historical trajectories of 8 time steps and predict the subsequent 12 time steps on both ETH-UCY and SDD." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "type": "text", + "content": "Evaluation Metrics. Following previous methods [49, 36, 8, 41, 27], we employ two commonly used metrics, i.e., Average Displacement Error (ADE) and Final Displacement Error (FDE), to evaluate the trajectory prediction perfor" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 300, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 547, + 361 + ], + "type": "text", + "content": "mance. ADE measures the average " + }, + { + "bbox": [ + 304, + 300, + 547, + 361 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 300, + 547, + 361 + ], + "type": "text", + "content": "-norm distance between all points of the ground truth and predicted trajectory. FDE measures the " + }, + { + "bbox": [ + 304, + 300, + 547, + 361 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 300, + 547, + 361 + ], + "type": "text", + "content": "-norm distance between the destination points of the ground truth and predicted trajectory. We use the best-of-20 metrics following previous methods [49, 41]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 361, + 547, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 547, + 469 + ], + "type": "text", + "content": "Implementation Details. In our experiments, the two encoders are all 3-layer BiLSTM [56] architecture with hidden sizes of 48. The decoder is 3-layer BiLSTM [56] architecture with hidden size of 96. The dimensions of " + }, + { + "bbox": [ + 304, + 361, + 547, + 469 + ], + "type": "inline_equation", + "content": "\\theta_{M}" + }, + { + "bbox": [ + 304, + 361, + 547, + 469 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 361, + 547, + 469 + ], + "type": "inline_equation", + "content": "\\theta_{V}" + }, + { + "bbox": [ + 304, + 361, + 547, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 361, + 547, + 469 + ], + "type": "inline_equation", + "content": "\\theta_{R}" + }, + { + "bbox": [ + 304, + 361, + 547, + 469 + ], + "type": "text", + "content": " are set to 48, 48 and 96, respectively. The parameter " + }, + { + "bbox": [ + 304, + 361, + 547, + 469 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 361, + 547, + 469 + ], + "type": "text", + "content": " of mask threshold is 0.2. The parameter " + }, + { + "bbox": [ + 304, + 361, + 547, + 469 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 304, + 361, + 547, + 469 + ], + "type": "text", + "content": " in memory training is 0.0001. The Adam optimizer is used to train our model by 100 epochs with a learning rate of 0.0005, decaying by 0.5 with an interval of 50." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 478, + 430, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 430, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 430, + 491 + ], + "type": "text", + "content": "4.2. Quantitative Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 497, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 547, + 605 + ], + "type": "text", + "content": "On ETH-UCY. As shown in Table 1, we compare our method with eleven state-of-the-art methods on the ETH-UCY dataset in past six years. Among them, MemoNet [49], PECNet [27], and CAGN [8] are goal-conditioned methods. MANTRA [28] and PCCSNet [41] are instance-conditioned methods. SoPhie [34], SGCN [37] and AgentFormer [54] generate multimodal trajectories by latent variables. The experiment results show that generating multimodal predictions by goals or instance is better than by latent variables." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "Moreover, the results indicate that our method significantly outperforms all the competing methods on both ETH and UCY. Compared with the previous best method MemoNet [49], our method further improves the performance on ADE and FDE. Meanwhile, Compared with the other instance-conditioned methods, PCCSNet [41], and MANTRA [28], our method also achieve the best performance. The underlying reason could be that our method captures the effective intermediate process compared with" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9768" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 286, + 159 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 159 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 159 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 159 + ], + "type": "table", + "html": "
MethodK=15K=10K=5
PECNet [27]10.66/17.8511.79/21.0014.49/28.08
PCCSNet [41]9.27/17.6210.36/20.1212.57/24.74
CAGN [8]10.01/17.4911.06/19.8913.85/26.87
MemoNet [49]9.62/15.3611.26/19.6414.43/27.45
Ours9.13/15.3110.27/18.1312.82/23.96
", + "image_path": "4333ccbba8d0781e15db4445fdac2c27065f182524668094317a001115436095.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 198, + 279, + 285 + ], + "blocks": [ + { + "bbox": [ + 47, + 166, + 287, + 189 + ], + "lines": [ + { + "bbox": [ + 47, + 166, + 287, + 189 + ], + "spans": [ + { + "bbox": [ + 47, + 166, + 287, + 189 + ], + "type": "text", + "content": "Table 3. Experiments on different best-of- " + }, + { + "bbox": [ + 47, + 166, + 287, + 189 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 166, + 287, + 189 + ], + "type": "text", + "content": " predictions on SDD in ADE/FDE. The lower the better." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 198, + 279, + 285 + ], + "lines": [ + { + "bbox": [ + 56, + 198, + 279, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 279, + 285 + ], + "type": "table", + "html": "
MethodSIREFCLADE/FDE
Baseline0.46/0.94
(1)✓✓0.25/0.50
(2)✓✓0.27/0.56
(3)✓✓0.22/0.39
(4)✓✓✓0.19/0.33
", + "image_path": "2a8859e49e5a7712d918f8ba4ae00a84cb04a713319fb6d484f23d8b4e5fc011.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 335, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 357 + ], + "type": "text", + "content": "the goal-conditioned methods and is more flexible compared with the instance-conditioned methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 359, + 288, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 359, + 288, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 288, + 420 + ], + "type": "text", + "content": "To show the superiority of our method more clearly, we do not use extra information, such as social interaction and map information, as previous approaches. Nevertheless, our method still achieves the best performance, clearly demonstrating the superiority of our method using sparse instance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 420, + 289, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 289, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 289, + 564 + ], + "type": "text", + "content": "On SDD. We compare our method with nine state-of-the-art methods on the SDD dataset in Table 2. For ADE, our method outperforms all the competing methods. For FDE, our method also achieves comparable performance with the best method MemoNet [49] and exceeds all other methods. The results manifest that the sparse instance representation can comprehensively describe the trajectory multimodality rather than goal-conditioned methods and thus achieves competitive performance. We speculate that our method underperforms MemoNet [49] on FDE because the sparse instance benefits more on the whole trajectory prediction (i.e., ADE) than goal prediction (i.e., FDE)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 574, + 142, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 142, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 142, + 588 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Different Best-of- " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " Predictions. Previous methods commonly use best-of- " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " (usually " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "K = 20" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": ") as the quantified metric of multimodal trajectory prediction. To further validate the adaptability and effectiveness of our proposed sparse instance representation, we conduct an experiment on various best-of- " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " predictions with " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "K = 5, 10, 15" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " on SDD. We compare our method with the top three performance goal-conditioned methods, i.e., MemoNet [49], CAGN [8] and PECNet [27] and the best performance instance-conditioned method PCCSNet [41]. As shown in Table 3, our method" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 539, + 167 + ], + "blocks": [ + { + "bbox": [ + 47, + 292, + 288, + 315 + ], + "lines": [ + { + "bbox": [ + 47, + 292, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 47, + 292, + 288, + 315 + ], + "type": "text", + "content": "Table 4. Ablation study of each component of our method on ETH-UCY dataset in ADE/FDE. The lower the better." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 70, + 539, + 167 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 539, + 167 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 539, + 167 + ], + "type": "table", + "html": "
μ00.10.20.30.40.5
ADE0.210.210.190.200.200.21
FDE0.340.330.330.330.340.37
μ0.60.70.80.91.0-
ADE0.250.280.260.240.21-
FDE0.450.540.480.430.38-
", + "image_path": "74dc841b0ef707b621c76fddddbb7afd64ac4003629c9049eaba658c13e8de42.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 175, + 545, + 196 + ], + "lines": [ + { + "bbox": [ + 305, + 175, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 305, + 175, + 545, + 196 + ], + "type": "text", + "content": "Table 5. Ablation study of different mask threshold values on the ETH-UCY dataset. The lower the better." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 307, + 207, + 544, + 320 + ], + "blocks": [ + { + "bbox": [ + 307, + 207, + 544, + 320 + ], + "lines": [ + { + "bbox": [ + 307, + 207, + 544, + 320 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 544, + 320 + ], + "type": "image", + "image_path": "835407cf1dfcec615275cc63c3172e8a9470844281c907fd888c92dea4850361.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 323, + 547, + 357 + ], + "lines": [ + { + "bbox": [ + 305, + 323, + 547, + 357 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 547, + 357 + ], + "type": "text", + "content": "Figure 6. Experiments of the sparse instance analysis on the ETH-UCY dataset. We replace our proposed sparse instance with the goal and instance for the multimodal prediction." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 376, + 547, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 547, + 449 + ], + "type": "text", + "content": "achieves significant performance on all experimental settings. Interestingly, our method surpasses the previous best method MemoNet [49] when " + }, + { + "bbox": [ + 304, + 376, + 547, + 449 + ], + "type": "inline_equation", + "content": "K = 5, 10, 15" + }, + { + "bbox": [ + 304, + 376, + 547, + 449 + ], + "type": "text", + "content": ", despite underperforming MemoNet [49] when " + }, + { + "bbox": [ + 304, + 376, + 547, + 449 + ], + "type": "inline_equation", + "content": "K = 20" + }, + { + "bbox": [ + 304, + 376, + 547, + 449 + ], + "type": "text", + "content": " in Table 2. It shows the adaptability of our method of using the sparse instance, which also performs significantly with a small " + }, + { + "bbox": [ + 304, + 376, + 547, + 449 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 376, + 547, + 449 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 449, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 544 + ], + "type": "text", + "content": "Contribution of Each Component. We divide our proposed SICNet into three components, i.e., the sparse instance representation (SI), the memory refinement module (REF), and the clustering post-process (CL). Ablation experiments of different combinations of the three components are evaluated on the ETH-UCY dataset. As shown in Table 4, all of the three components contribute to the performance improvement, which demonstrates their effectiveness." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 545, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 653 + ], + "type": "text", + "content": "Analysis of Mask Threshold. As shown in Table 5, we conduct experiments for different mask threshold values " + }, + { + "bbox": [ + 304, + 545, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 545, + 547, + 653 + ], + "type": "text", + "content": " on the ETH-UCY dataset. It indicates that the performance will decrease when the mask threshold value " + }, + { + "bbox": [ + 304, + 545, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 545, + 547, + 653 + ], + "type": "text", + "content": " is too small or too large. When " + }, + { + "bbox": [ + 304, + 545, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\mu = 0" + }, + { + "bbox": [ + 304, + 545, + 547, + 653 + ], + "type": "text", + "content": ", the mask is useless and the whole instance is reserved, thus causing the stochasticity loss. When " + }, + { + "bbox": [ + 304, + 545, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\mu = 1" + }, + { + "bbox": [ + 304, + 545, + 547, + 653 + ], + "type": "text", + "content": ", it denotes the whole instance is masked, thus causing no guidance information for the multimodal prediction. When " + }, + { + "bbox": [ + 304, + 545, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\mu = 0.2" + }, + { + "bbox": [ + 304, + 545, + 547, + 653 + ], + "type": "text", + "content": ", it achieves the best performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "Analysis of Sparse Instance. As shown in Table 6, we replace the sparse instance in SICNet with goal-conditioned and instance-conditioned representations on the ETH-UCY dataset. Experiment results show that using the sparse instance achieves the best performance, which indicates the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9769" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 210 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 210 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 210 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 210 + ], + "type": "image", + "image_path": "114a206eefaa53bceee93d73a4d0d771baaf1ca2d815ddfd9df79b89a5f6c7c9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison of our method with the goal-conditioned method PECNet [27] on the SDD dataset. Given the observed trajectories, we illustrate the ground truth paths and predicted trajectories by SICNet and the goal-conditioned method PECNet [27] for four different scenes. It shows our prediction is closer to the ground truth than PECNet [27], especially for the intermediate process." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 255, + 545, + 392 + ], + "blocks": [ + { + "bbox": [ + 50, + 255, + 545, + 392 + ], + "lines": [ + { + "bbox": [ + 50, + 255, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 50, + 255, + 545, + 392 + ], + "type": "image", + "image_path": "44482aaf343e512b106d1580fa14c76cf147044e997bd3d6084dbae7d2ec4298.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 394, + 547, + 428 + ], + "lines": [ + { + "bbox": [ + 46, + 394, + 547, + 428 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 547, + 428 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison of our method with the instance-conditioned method PCCSNet [41] on the ETH-UCY dataset. Given the observed trajectories, we illustrate the ground truth paths and predicted trajectories by SICNet and the instance-conditioned method PCCSNet [41] for four different scenes. It shows our prediction is closer to the ground truth than PCCSNet [41]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 447, + 288, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 288, + 472 + ], + "type": "text", + "content": "superiority of our method compared with goal-conditioned and instance-conditioned approaches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 479, + 165, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 165, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 165, + 493 + ], + "type": "text", + "content": "4.4. Qualitative Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 498, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 290, + 715 + ], + "type": "text", + "content": "We compare the visualized trajectories with the well-accepted goal-conditioned method PECNet [27] on SDD in Figure 7, and the instance-conditioned method PCCSNet [41] on ETH-UCY in Figure 8. As shown in Figure 7, trajectories in Coupa show that our method achieves more precise trajectory predictions than PECNet. Trajectories in Little and Quad show that both PECNet and our method can predict destinations close to the ground truth, while our method can predict more precise intermediate trajectories due to our proposed sparse instance representation. Trajectories in Hyang show that our method can also give better predictions, although the ground truth trajectories are irregular. The results verify the superiority of our method compared with goal-conditioned methods. Moreover, as shown in Figure 8, the visualization results show that our prediction is closer to the ground truth than PCCSNet [41], which indicates the superiority of our SICNet compared with the instance-conditioned approaches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 446, + 379, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 446, + 379, + 459 + ], + "spans": [ + { + "bbox": [ + 306, + 446, + 379, + 459 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 469, + 548, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 548, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 548, + 601 + ], + "type": "text", + "content": "In this paper, we present a simple yet effective sparse-instance-conditioned network for pedestrian multimodal trajectory prediction, which leverages sparse instances to guide multimodal prediction. Moreover, we propose a memory refinement module to reduce recall errors when meeting unfamiliar trajectories during inference. Extensive experiments show that our method performs better than previous methods, and ablation studies demonstrate that our sparse-instance-conditioned method can generate multimodal predictions more accurately and adaptively compared with goal-conditioned and instance-conditioned approaches." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 619, + 404, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 619, + 404, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 619, + 404, + 633 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 548, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 548, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 548, + 712 + ], + "type": "text", + "content": "This work was supported partly by National Key R&D Program of China under Grant 2021YFB1714700, NSFC under Grants 62088102 and 62106192, Natural Science Foundation of Shaanxi Province under Grant 2022JC-41, and Fundamental Research Funds for the Central Universities under Grant XTR042021005." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9770" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social LSTM: Human trajectory prediction in crowded spaces. In CVPR, pages 961–971, 2016." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "type": "text", + "content": "[2] Inhwan Bae, Jin-Hwi Park, and Hae-Gon Jeon. Nonprobability sampling network for stochastic human trajectory prediction. In CVPR, pages 6477-6487, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 288, + 203 + ], + "type": "text", + "content": "[3] Guangyi Chen, Junlong Li, Jiwen Lu, and Jie Zhou. Human trajectory prediction via counterfactual analysis. In ICCV, pages 9824-9833, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 236 + ], + "type": "text", + "content": "[4] Xinlei Chen and Abhinav Gupta. Spatial memory for context reasoning in object detection. In ICCV, pages 4086-4096, 2017." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 239, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 271 + ], + "type": "text", + "content": "[5] Chiho Choi, Joon Hee Choi, Jiachen Li, and Srikanth Malla. Shared cross-modal trajectory prediction for autonomous driving. In CVPR, pages 244-253, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 274, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 288, + 316 + ], + "type": "text", + "content": "[6] Lingwei Dang, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. MSR-GCN: Multi-scale residual graph convolution networks for human motion prediction. In ICCV, pages 11447-11456, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "type": "text", + "content": "[7] Hanming Deng, Yang Hua, Tao Song, Zongpu Zhang, Zhengui Xue, Ruhui Ma, Neil Robertson, and Haibing Guan. Object guided external memory network for video object detection. In ICCV, pages 6677-6686, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 364, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 364, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 364, + 288, + 407 + ], + "type": "text", + "content": "[8] Jinghai Duan, Le Wang, Chengjiang Long, Sanping Zhou, Fang Zheng, Liushuai Shi, and Gang Hua. Complementary attention gated network for pedestrian trajectory prediction. In AAAI, pages 542-550, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 409, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 288, + 441 + ], + "type": "text", + "content": "[9] Zhihong Fu, Qingjie Liu, Zehua Fu, and Yunhong Wang. STMTrack: Template-free visual tracking with space-time memory networks. In CVPR, pages 13769-13778, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 443, + 288, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 443, + 288, + 486 + ], + "spans": [ + { + "bbox": [ + 48, + 443, + 288, + 486 + ], + "type": "text", + "content": "[10] Harshayu Girase, Haiming Gang, Srikanth Malla, Jiachen Li, Akira Kanehara, Karttikeya Mangalam, and Chiho Choi. LOKI: Long term and key intentions for trajectory prediction. In ICCV, pages 9803-9812, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 487, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 288, + 520 + ], + "type": "text", + "content": "[11] Dong Gong, Frederic Z. Zhang, Javen Qinfeng Shi, and Anton van den Hengel. Memory-augmented dynamic neural relational inference. In ICCV, pages 11843-11852, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 522, + 288, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 565 + ], + "type": "text", + "content": "[12] Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory prediction via motion indeterminacy diffusion. In CVPR, pages 17113-17122, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 567, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 288, + 609 + ], + "type": "text", + "content": "[13] Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi. Social GAN: socially acceptable trajectories with generative adversarial networks. In CVPR, pages 2255–2264, 2018." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 612, + 288, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 288, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 288, + 654 + ], + "type": "text", + "content": "[14] Adam Houenou, Philippe Bonnifait, Véronique Cherfaoui, and Wen Yao. Vehicle trajectory prediction based on motion model and maneuver recognition. In IROS, pages 4363-4369, 2013." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 657, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 288, + 689 + ], + "type": "text", + "content": "[15] Boris Ivanovic and Marco Pavone. The trajectory: Probabilistic multi-agent trajectory modeling with dynamic spatiotemporal graphs. In ICCV, pages 2375-2384, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[16] Lukasz Kaiser, Ofir Nachum, Aurko Roy, and Samy Bengio. Learning to remember rare events. In ICLR, 2017." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[17] Youngjin Kim, Minjung Kim, and Gunhee Kim. Memorization precedes generation: Learning unsupervised gans with memory networks. In ICLR, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 139 + ], + "type": "text", + "content": "[18] Sumith Kulal, Jiayuan Mao, Alex Aiken, and Jiajun Wu. Programmatic concept learning for human motion description and synthesis. In CVPR, pages 13843-13852, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 140, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 547, + 183 + ], + "type": "text", + "content": "[19] Mihee Lee, Samuel S. Sohn, Seonghyeon Moon, Sejong Yoon, Mubbasir Kapadia, and Vladimir Pavlovic. MUSE-VAE: Multi-scale vae for environment-aware long term trajectory prediction. In CVPR, pages 2221–2230, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 185, + 547, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 547, + 216 + ], + "type": "text", + "content": "[20] Sangho Lee, Jinyoung Sung, Youngjae Yu, and Gunhee Kim. A memory network approach for story-based temporal summarization of " + }, + { + "bbox": [ + 308, + 185, + 547, + 216 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 308, + 185, + 547, + 216 + ], + "type": "text", + "content": " videos. In CVPR, pages 1410-1419, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 217, + 547, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 547, + 248 + ], + "type": "text", + "content": "[21] Alon Lerner, Yiorgos Chrysanthou, and Dani Lischinski. Crowds by example. Comput. Graph. Forum, 26(3):655-664, 2007." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 251, + 547, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 547, + 282 + ], + "type": "text", + "content": "[22] Peizhao Li, Pu Wang, Karl Berntorp, and Hongfu Liu. Exploiting temporal relations on radar perception for autonomous driving. In CVPR, pages 17071-17080, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 284, + 547, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 547, + 316 + ], + "type": "text", + "content": "[23] Shijie Li, Yanying Zhou, Jinhui Yi, and Juergen Gall. Spatial-temporal consistency network for low-latency trajectory forecasting. In ICCV, pages 1940-1949, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 317, + 547, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 317, + 547, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 317, + 547, + 360 + ], + "type": "text", + "content": "[24] Jialun Liu, Yifan Sun, Feng Zhu, Hongbin Pei, Yi Yang, and Wenhui Li. Learning memory-augmented unidirectional metrics for cross-modality person re-identification. In CVPR, pages 19366-19375, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 361, + 547, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 547, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 547, + 404 + ], + "type": "text", + "content": "[25] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In ICCV, pages 13588-13597, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 405, + 547, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 547, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 547, + 447 + ], + "type": "text", + "content": "[26] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In ICCV, pages 15233-15242, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 449, + 547, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 547, + 502 + ], + "type": "text", + "content": "[27] Karttikeya Mangalam, Harshayu Girase, Shreyas Agarwal, Kuan-Hui Lee, Ehsan Adeli, Jitendra Malik, and Adrien Gaidon. It is not the journey but the destination: Endpoint conditioned trajectory prediction. In ECCV, pages 759-776, 2020." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 504, + 547, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 547, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 547, + 547 + ], + "type": "text", + "content": "[28] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. *Mantra: Memory augmented networks for multiple trajectory prediction*. In *CVPR*, pages 7143–7152, 2020." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 548, + 547, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 547, + 592 + ], + "type": "text", + "content": "[29] Abduallah Mohamed, Kun Qian, Mohamed Elhoseiny, and Christian Claudel. Social-STGCNN: A social spatio-temporal graph convolutional neural network for human trajectory prediction. In CVPR, pages 14424-14432, 2020." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 593, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 547, + 635 + ], + "type": "text", + "content": "[30] Alessio Monti, Angelo Porrello, Simone Calderara, Pasquale Coscia, Lamberto Ballan, and Rita Cucchiara. How many observations are enough? knowledge distillation for trajectory forecasting. In CVPR, pages 6553-6562, 2022." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 637, + 547, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 637, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 637, + 547, + 669 + ], + "type": "text", + "content": "[31] Stefano Pellegrini, Andreas Ess, Konrad Schindler, and Luc Van Gool. You'll never walk alone: Modeling social behavior for multi-target tracking. In ICCV, pages 261-268, 2009." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "type": "text", + "content": "[32] Zheng Qin, Sanping Zhou, Le Wang, Jinghai Duan, Gang Hua, and Wei Tang. Motiontrack: Learning robust short-term and long-term motions for multi-object tracking. In CVPR, pages 17939-17948, 2023." + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "9771" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[33] Alexandre Robicquet, Amir Sadeghian, Alexandre Alahi, and Silvio Savarese. Learning Social Etiquette: Human trajectory understanding in crowded scenes. In ECCV, pages 549-565, 2016." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 163 + ], + "type": "text", + "content": "[34] Amir Sadeghian, Vineet Kosaraju, Ali Sadeghian, Noriaki Hirose, Hamid Rezatofighi, and Silvio Savarese. SoPhie: An attentive gan for predicting paths compliant to social and physical constraints. In CVPR, pages 1349-1358, 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 165, + 288, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 165, + 288, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 288, + 208 + ], + "type": "text", + "content": "[35] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory: Dynamically-feasible trajectory forecasting with heterogeneous data. In ECCV, pages 683-700, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 210, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 210, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 288, + 255 + ], + "type": "text", + "content": "[36] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Fang Zheng, Nanning Zheng, and Gang Hua. Social interpretable tree for pedestrian trajectory prediction. In AAAI, pages 2235-2243, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 256, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 256, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 256, + 287, + 300 + ], + "type": "text", + "content": "[37] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Mo Zhou, Zhenxing Niu, and Gang Hua. SGCN: Sparse graph convolution network for pedestrian trajectory prediction. In CVPR, pages 8990-8999, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 301, + 288, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 301, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 48, + 301, + 288, + 345 + ], + "type": "text", + "content": "[38] Aravind Srinivas, Tsung-Yi Lin, Niki Parmar, Jonathon Shlens, Pieter Abbeel, and Ashish Vaswani. Bottleneck transformers for visual recognition. In CVPR, pages 16519-16529, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 347, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 347, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 347, + 287, + 380 + ], + "type": "text", + "content": "[39] Hao Sun, Zhiqun Zhao, and Zhihai He. Reciprocal learning networks for human trajectory prediction. In CVPR, pages 7416-7425, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 382, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 415 + ], + "type": "text", + "content": "[40] Jianhua Sun, Qinhong Jiang, and Cewu Lu. Recursive social behavior graph for trajectory prediction. In CVPR, pages 660-669, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 417, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 417, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 417, + 288, + 459 + ], + "type": "text", + "content": "[41] Jianhua Sun, Yuxuan Li, Hao-Shu Fang, and Cewu Lu. Three steps to multimodal trajectory prediction: Modality clustering, classification and synthesis. In ICCV, pages 13250-13259, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 462, + 288, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 462, + 288, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 462, + 288, + 518 + ], + "type": "text", + "content": "[42] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, Wenyu Liu, and Bin Xiao. Deep high-resolution representation learning for visual recognition. IEEE T-PAMI, 43(10):3349-3364, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 519, + 288, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 519, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 519, + 288, + 552 + ], + "type": "text", + "content": "[43] Jingke Wang, Tengju Ye, Ziqing Gu, and Junbo Chen. LTP: Lane-based trajectory prediction for autonomous driving. In CVPR, pages 17134-17142, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 554, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 554, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 554, + 287, + 587 + ], + "type": "text", + "content": "[44] Jürgen Wiest, Matthias Höffken, Ulrich Kreßel, and Klaus Dietmayer. Probabilistic trajectory prediction with gaussian mixture models. In IV, pages 141-146, 2012." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 632 + ], + "type": "text", + "content": "[45] Yuxuan Wu, Le Wang, Sanping Zhou, Jinghai Duan, Gang Hua, and Wei Tang. Multi-stream representation learning for pedestrian trajectory prediction. In AAAI, pages 2875-2882, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 667 + ], + "type": "text", + "content": "[46] Zifeng Wu, Chunhua Shen, and Anton Van Den Hengel. Wider or deeper: Revisiting the resnet model for visual recognition. PR, 90:119-133, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[47] Haozhe Xie, Hongxun Yao, Shangchen Zhou, Shengping Zhang, and Wenxiu Sun. Efficient regional memory network for video object segmentation. In CVPR, pages 1286-1295, 2021." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 443 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 72, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 547, + 106 + ], + "type": "text", + "content": "[48] Bo Xu, Han Huang, Cheng Lu, Ziwen Li, and Yandong Guo. Virtual multi-modality self-supervised foreground matting for human-object interaction. In ICCV, pages 428-437, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 140 + ], + "type": "text", + "content": "[49] Chenxin Xu, Weibo Mao, Wenjun Zhang, and Siheng Chen. Remember intentions: Retrospective-memory-based trajectory prediction. In CVPR, pages 6488-6497, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 141, + 547, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 547, + 174 + ], + "type": "text", + "content": "[50] Hao Xue, Du Q. Huynh, and Mark Reynolds. SS-LSTM: A hierarchical LSTM model for pedestrian trajectory prediction. In WACV, pages 1186-1194, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 175, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 547, + 217 + ], + "type": "text", + "content": "[51] Gang Yang, Man Zhou, Keyu Yan, Aiping Liu, Xueyang Fu, and Fan Wang. Memory-augmented deep conditional unfolding network for pan-sharpening. In CVPR, pages 1788-1797, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 220, + 547, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 547, + 251 + ], + "type": "text", + "content": "[52] Tianyu Yang and Antoni B. Chan. Visual tracking via dynamic memory networks. IEEE T-PAMI, 43(1):360-374, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 254, + 547, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 547, + 286 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 547, + 286 + ], + "type": "text", + "content": "[53] Jiyang Yu, Jingen Liu, Liefeng Bo, and Tao Mei. Memory-augmented non-local attention for video super-resolution. In CVPR, pages 17834-17843, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 288, + 547, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 288, + 547, + 320 + ], + "spans": [ + { + "bbox": [ + 308, + 288, + 547, + 320 + ], + "type": "text", + "content": "[54] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris Kitani. Agent-Former: Agent-aware transformers for socio-temporal multiagent forecasting. In ICCV, pages 9793–9803, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 322, + 547, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 322, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 322, + 547, + 354 + ], + "type": "text", + "content": "[55] He Zhao and Richard P. Wildes. Where are you heading? dynamic trajectory prediction with expert goal examples. In ICCV, pages 7609-7618, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 355, + 547, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 547, + 399 + ], + "type": "text", + "content": "[56] Zheng Zhao, Weihai Chen, Xingming Wu, Peter CY Chen, and Jingmeng Liu. LSTM network: a deep learning approach for short-term traffic forecast. IET Intell. Transp. Syst., 11(2):68-75, 2017." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 399, + 547, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 547, + 443 + ], + "type": "text", + "content": "[57] Desen Zhou, Zhichao Liu, Jian Wang, Leshan Wang, Tao Hu, Errui Ding, and Jingdong Wang. Human-object interaction detection via disentangled transformer. In CVPR, pages 19568-19577, 2022." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9772" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file